diff --git a/.gitattributes b/.gitattributes index ca0507bf4dda538c045ef14fcba3c9d8186bdc46..9f8449cb4caef6de2033b0c109e2e288c09451d3 100644 --- a/.gitattributes +++ b/.gitattributes @@ -424,3 +424,66 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text 2025/Chain[[:space:]]of[[:space:]]Semantics[[:space:]]Programming[[:space:]]in[[:space:]]3D[[:space:]]Gaussian[[:space:]]Splatting[[:space:]]Representation[[:space:]]for[[:space:]]3D[[:space:]]Vision[[:space:]]Grounding/0f72fb5d-a469-4fd4-b3fb-2e6e19e60074_origin.pdf filter=lfs diff=lfs merge=lfs -text 2025/ChainHOI_[[:space:]]Joint-based[[:space:]]Kinematic[[:space:]]Chain[[:space:]]Modeling[[:space:]]for[[:space:]]Human-Object[[:space:]]Interaction[[:space:]]Generation/4550042c-6e98-4dbc-b87f-acb949bc7a09_origin.pdf filter=lfs diff=lfs merge=lfs -text 2025/Change3D_[[:space:]]Revisiting[[:space:]]Change[[:space:]]Detection[[:space:]]and[[:space:]]Captioning[[:space:]]from[[:space:]]A[[:space:]]Video[[:space:]]Modeling[[:space:]]Perspective/bf102372-b0ac-450f-98e7-1e9b37a55194_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Channel[[:space:]]Consistency[[:space:]]Prior[[:space:]]and[[:space:]]Self-Reconstruction[[:space:]]Strategy[[:space:]]Based[[:space:]]Unsupervised[[:space:]]Image[[:space:]]Deraining/1cd88811-9c34-4363-ae8a-2a52dd5a816f_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Channel-wise[[:space:]]Noise[[:space:]]Scheduled[[:space:]]Diffusion[[:space:]]for[[:space:]]Inverse[[:space:]]Rendering[[:space:]]in[[:space:]]Indoor[[:space:]]Scenes/e287505c-edc5-4f3b-9021-f507eb27467f_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Chapter-Llama_[[:space:]]Efficient[[:space:]]Chaptering[[:space:]]in[[:space:]]Hour-Long[[:space:]]Videos[[:space:]]with[[:space:]]LLMs/c5a48e4c-7b19-4801-9ffe-45905e053adb_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Charm_[[:space:]]The[[:space:]]Missing[[:space:]]Piece[[:space:]]in[[:space:]]ViT[[:space:]]Fine-Tuning[[:space:]]for[[:space:]]Image[[:space:]]Aesthetic[[:space:]]Assessment/0ec61778-fb6e-453d-bfdb-3d054143034b_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Chat-based[[:space:]]Person[[:space:]]Retrieval[[:space:]]via[[:space:]]Dialogue-Refined[[:space:]]Cross-Modal[[:space:]]Alignment/6896dcb7-a954-4e52-90ab-4e44608cdd32_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Chat2SVG_[[:space:]]Vector[[:space:]]Graphics[[:space:]]Generation[[:space:]]with[[:space:]]Large[[:space:]]Language[[:space:]]Models[[:space:]]and[[:space:]]Image[[:space:]]Diffusion[[:space:]]Models/03e11a55-7ca3-4174-a887-3307219befcc_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/ChatGarment_[[:space:]]Garment[[:space:]]Estimation,[[:space:]]Generation[[:space:]]and[[:space:]]Editing[[:space:]]via[[:space:]]Large[[:space:]]Language[[:space:]]Models/425b09e1-924e-4a43-b277-eb8a41387d21_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/ChatGen_[[:space:]]Automatic[[:space:]]Text-to-Image[[:space:]]Generation[[:space:]]From[[:space:]]FreeStyle[[:space:]]Chatting/0c8e9c4a-231c-4c6d-af84-cd3eb4170579_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/ChatHuman_[[:space:]]Chatting[[:space:]]about[[:space:]]3D[[:space:]]Humans[[:space:]]with[[:space:]]Tools/a1962ae6-92af-4828-9f1c-48e34ea37fec_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/CheXWorld_[[:space:]]Exploring[[:space:]]Image[[:space:]]World[[:space:]]Modeling[[:space:]]for[[:space:]]Radiograph[[:space:]]Representation[[:space:]]Learning/f79a692c-48ff-4b64-8cb4-11cdf9303afb_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/CheXwhatsApp_[[:space:]]A[[:space:]]Dataset[[:space:]]for[[:space:]]Exploring[[:space:]]Challenges[[:space:]]in[[:space:]]the[[:space:]]Diagnosis[[:space:]]of[[:space:]]Chest[[:space:]]X-rays[[:space:]]through[[:space:]]Mobile[[:space:]]Devices/8cf11147-8ad4-4f89-a3ee-6b5492c8beb0_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Cheb-GR_[[:space:]]Rethinking[[:space:]]K-nearest[[:space:]]Neighbor[[:space:]]Search[[:space:]]in[[:space:]]Re-ranking[[:space:]]for[[:space:]]Person[[:space:]]Re-identification/9e933435-e0fc-412e-b745-089375b60f23_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Chebyshev[[:space:]]Attention[[:space:]]Depth[[:space:]]Permutation[[:space:]]Texture[[:space:]]Network[[:space:]]with[[:space:]]Latent[[:space:]]Texture[[:space:]]Attribute[[:space:]]Loss/44aa142c-41b8-482e-a039-e4a4fc955e21_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/CheckManual_[[:space:]]A[[:space:]]New[[:space:]]Challenge[[:space:]]and[[:space:]]Benchmark[[:space:]]for[[:space:]]Manual-based[[:space:]]Appliance[[:space:]]Manipulation/c1aa26f1-f8e2-45ab-92dc-640acaf534e5_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/CholecTrack20_[[:space:]]A[[:space:]]Multi-Perspective[[:space:]]Tracking[[:space:]]Dataset[[:space:]]for[[:space:]]Surgical[[:space:]]Tools/3671c43b-afb0-4272-8573-cad641013bba_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Circumventing[[:space:]]Shortcuts[[:space:]]in[[:space:]]Audio-visual[[:space:]]Deepfake[[:space:]]Detection[[:space:]]Datasets[[:space:]]with[[:space:]]Unsupervised[[:space:]]Learning/371069ba-8973-4b9c-bb98-ecda46029053_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/CityWalker_[[:space:]]Learning[[:space:]]Embodied[[:space:]]Urban[[:space:]]Navigation[[:space:]]from[[:space:]]Web-Scale[[:space:]]Videos/a9edaa72-0d38-4c4e-a25a-9480eb60d753_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Classic[[:space:]]Video[[:space:]]Denoising[[:space:]]in[[:space:]]a[[:space:]]Machine[[:space:]]Learning[[:space:]]World_[[:space:]]Robust,[[:space:]]Fast,[[:space:]]and[[:space:]]Controllable/079f9f38-09f8-4442-836f-ae05472182ae_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Classifier-Free[[:space:]]Guidance[[:space:]]Inside[[:space:]]the[[:space:]]Attraction[[:space:]]Basin[[:space:]]May[[:space:]]Cause[[:space:]]Memorization/a1a30dde-51b7-469a-a404-0c4a54031538_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Classifier-guided[[:space:]]CLIP[[:space:]]Distillation[[:space:]]for[[:space:]]Unsupervised[[:space:]]Multi-label[[:space:]]Classification/325c3ac3-6589-4c70-b2e1-2f7ac7877553_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Classifier-to-Bias_[[:space:]]Toward[[:space:]]Unsupervised[[:space:]]Automatic[[:space:]]Bias[[:space:]]Detection[[:space:]]for[[:space:]]Visual[[:space:]]Classifiers/303d2f58-e46e-4b7a-806e-0d4e64d53df2_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/CleanDIFT_[[:space:]]Diffusion[[:space:]]Features[[:space:]]without[[:space:]]Noise/17f652ff-055a-4bf1-a266-9144370345a9_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/ClearSight_[[:space:]]Visual[[:space:]]Signal[[:space:]]Enhancement[[:space:]]for[[:space:]]Object[[:space:]]Hallucination[[:space:]]Mitigation[[:space:]]in[[:space:]]Multimodal[[:space:]]Large[[:space:]]Language[[:space:]]Models/cc596546-7afd-468d-96ec-564a706a95d3_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/ClimbingCap_[[:space:]]Multi-Modal[[:space:]]Dataset[[:space:]]and[[:space:]]Method[[:space:]]for[[:space:]]Rock[[:space:]]Climbing[[:space:]]in[[:space:]]World[[:space:]]Coordinate/56189f24-e199-449b-838e-e830c93475ae_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Closed-Loop[[:space:]]Supervised[[:space:]]Fine-Tuning[[:space:]]of[[:space:]]Tokenized[[:space:]]Traffic[[:space:]]Models/2dc3d2cc-243e-4ca0-9955-c4c727b1c79c_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Closest[[:space:]]Neighbors[[:space:]]are[[:space:]]Harmful[[:space:]]for[[:space:]]Lightweight[[:space:]]Masked[[:space:]]Auto-encoders/ce070f35-5bfa-41af-9465-b1f84c742005_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Co-op_[[:space:]]Correspondence-based[[:space:]]Novel[[:space:]]Object[[:space:]]Pose[[:space:]]Estimation/5a137c7c-5457-4cd6-bb8c-541c7e90dee7_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/CoA_[[:space:]]Towards[[:space:]]Real[[:space:]]Image[[:space:]]Dehazing[[:space:]]via[[:space:]]Compression-and-Adaptation/06a0b6a5-a641-41a6-99e5-3f3c8a8c1a70_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/CoCoGaussian_[[:space:]]Leveraging[[:space:]]Circle[[:space:]]of[[:space:]]Confusion[[:space:]]for[[:space:]]Gaussian[[:space:]]Splatting[[:space:]]from[[:space:]]Defocused[[:space:]]Images/c9f7dcb9-6d18-430a-9426-253bcaf437ef_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/CoE_[[:space:]]Chain-of-Explanation[[:space:]]via[[:space:]]Automatic[[:space:]]Visual[[:space:]]Concept[[:space:]]Circuit[[:space:]]Description[[:space:]]and[[:space:]]Polysemanticity[[:space:]]Quantification/ed717be1-08cc-4195-89dc-b7ee977099ae_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/CoLLM_[[:space:]]A[[:space:]]Large[[:space:]]Language[[:space:]]Model[[:space:]]for[[:space:]]Composed[[:space:]]Image[[:space:]]Retrieval/02729911-7a78-4d55-af78-f50b5e45a135_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/CoMBO_[[:space:]]Conflict[[:space:]]Mitigation[[:space:]]via[[:space:]]Branched[[:space:]]Optimization[[:space:]]for[[:space:]]Class[[:space:]]Incremental[[:space:]]Segmentation/a4944725-d45e-4b4f-a404-3fd5f2ee3b6a_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/CoMM_[[:space:]]A[[:space:]]Coherent[[:space:]]Interleaved[[:space:]]Image-Text[[:space:]]Dataset[[:space:]]for[[:space:]]Multimodal[[:space:]]Understanding[[:space:]]and[[:space:]]Generation/4d3dc3b4-7f37-464f-b503-0ded8f6008de_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/CoMapGS_[[:space:]]Covisibility[[:space:]]Map-based[[:space:]]Gaussian[[:space:]]Splatting[[:space:]]for[[:space:]]Sparse[[:space:]]Novel[[:space:]]View[[:space:]]Synthesis/025e1d28-cc10-46ad-8c05-b01e6a4dff3b_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/CoMatcher_[[:space:]]Multi-View[[:space:]]Collaborative[[:space:]]Feature[[:space:]]Matching/b3062684-7794-47f2-b9d1-53d25767d94f_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/CoSDH_[[:space:]]Communication-Efficient[[:space:]]Collaborative[[:space:]]Perception[[:space:]]via[[:space:]]Supply-Demand[[:space:]]Awareness[[:space:]]and[[:space:]]Intermediate-Late[[:space:]]Hybridization/82cb35c7-62f2-4e6d-b4eb-be441dfd4b29_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/CoSER_[[:space:]]Towards[[:space:]]Consistent[[:space:]]Dense[[:space:]]Multiview[[:space:]]Text-to-Image[[:space:]]Generator[[:space:]]for[[:space:]]3D[[:space:]]Creation/911def72-ec15-4b90-8a72-6946e2b20044_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/CoSpace_[[:space:]]Benchmarking[[:space:]]Continuous[[:space:]]Space[[:space:]]Perception[[:space:]]Ability[[:space:]]for[[:space:]]Vision-Language[[:space:]]Models/45f50ce2-59a4-414e-86e0-abbb59b51daa_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/CoT-VLA_[[:space:]]Visual[[:space:]]Chain-of-Thought[[:space:]]Reasoning[[:space:]]for[[:space:]]Vision-Language-Action[[:space:]]Models/4f2ea357-dbc3-435d-b0c3-c4ec69dfb496_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Coarse[[:space:]]Correspondences[[:space:]]Boost[[:space:]]Spatial-Temporal[[:space:]]Reasoning[[:space:]]in[[:space:]]Multimodal[[:space:]]Language[[:space:]]Model/53ff4ccd-60e6-479a-bd87-54a994f71bf8_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/CocoER_[[:space:]]Aligning[[:space:]]Multi-Level[[:space:]]Feature[[:space:]]by[[:space:]]Competition[[:space:]]and[[:space:]]Coordination[[:space:]]for[[:space:]]Emotion[[:space:]]Recognition/91abd7f6-c62e-4947-ae7d-c03321b50ab1_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Code-as-Monitor_[[:space:]]Constraint-aware[[:space:]]Visual[[:space:]]Programming[[:space:]]for[[:space:]]Reactive[[:space:]]and[[:space:]]Proactive[[:space:]]Robotic[[:space:]]Failure[[:space:]]Detection/3086dac5-625c-4ed2-b32d-12ee6b15f252_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Coeff-Tuning_[[:space:]]A[[:space:]]Graph[[:space:]]Filter[[:space:]]Subspace[[:space:]]View[[:space:]]for[[:space:]]Tuning[[:space:]]Attention-Based[[:space:]]Large[[:space:]]Models/e7a23930-dc20-4352-a4dc-e2a2616c610f_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Coherent[[:space:]]3D[[:space:]]Portrait[[:space:]]Video[[:space:]]Reconstruction[[:space:]]via[[:space:]]Triplane[[:space:]]Fusion/2088b6b2-40e1-4341-a0c5-c5303c3acd65_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/ColabSfM_[[:space:]]Collaborative[[:space:]]Structure-from-Motion[[:space:]]by[[:space:]]Point[[:space:]]Cloud[[:space:]]Registration/52ce366e-9afd-441c-9c7e-e1d8e8bedeb7_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Collaborative[[:space:]]Decoding[[:space:]]Makes[[:space:]]Visual[[:space:]]Auto-Regressive[[:space:]]Modeling[[:space:]]Efficient/9a935c24-e29c-43fe-a54b-aa3325d4f954_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Collaborative[[:space:]]Tree[[:space:]]Search[[:space:]]for[[:space:]]Enhancing[[:space:]]Embodied[[:space:]]Multi-Agent[[:space:]]Collaboration/509a3bd5-5db7-45e0-bb5b-ba942b10cc92_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Color[[:space:]]Alignment[[:space:]]in[[:space:]]Diffusion/4b8bcc3c-a5c5-4a38-89da-b85960cd1b00_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/ComRoPE_[[:space:]]Scalable[[:space:]]and[[:space:]]Robust[[:space:]]Rotary[[:space:]]Position[[:space:]]Embedding[[:space:]]Parameterized[[:space:]]by[[:space:]]Trainable[[:space:]]Commuting[[:space:]]Angle[[:space:]]Matrices/287888de-80a8-43c2-8d0f-f61f04d80e4e_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/ComfyBench_[[:space:]]Benchmarking[[:space:]]LLM-based[[:space:]]Agents[[:space:]]in[[:space:]]ComfyUI[[:space:]]for[[:space:]]Autonomously[[:space:]]Designing[[:space:]]Collaborative[[:space:]]AI[[:space:]]Systems/9eeb1c75-c4c6-45dd-9896-e14c13fb270d_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Common3D_[[:space:]]Self-Supervised[[:space:]]Learning[[:space:]]of[[:space:]]3D[[:space:]]Morphable[[:space:]]Models[[:space:]]for[[:space:]]Common[[:space:]]Objects[[:space:]]in[[:space:]]Neural[[:space:]]Feature[[:space:]]Space/ecdc4e97-8d79-41e2-abfb-5c0cebf6aa2a_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Commonsense[[:space:]]Video[[:space:]]Question[[:space:]]Answering[[:space:]]through[[:space:]]Video-Grounded[[:space:]]Entailment[[:space:]]Tree[[:space:]]Reasoning/e1080574-1050-4c6e-8509-597d8efaff00_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Community[[:space:]]Forensics_[[:space:]]Using[[:space:]]Thousands[[:space:]]of[[:space:]]Generators[[:space:]]to[[:space:]]Train[[:space:]]Fake[[:space:]]Image[[:space:]]Detectors/f9d7ae16-50eb-4d82-828a-d6dd35d593a9_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/CompGS_[[:space:]]Unleashing[[:space:]]2D[[:space:]]Compositionality[[:space:]]for[[:space:]]Compositional[[:space:]]Text-to-3D[[:space:]]via[[:space:]]Dynamically[[:space:]]Optimizing[[:space:]]3D[[:space:]]Gaussians/d9350216-461c-42c5-bc6e-3da77f71b98b_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Compass[[:space:]]Control_[[:space:]]Multi[[:space:]]Object[[:space:]]Orientation[[:space:]]Control[[:space:]]for[[:space:]]Text-to-Image[[:space:]]Generation/9022e55b-ce42-4998-ae5e-db01a18e1b74_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Complementary[[:space:]]Advantages_[[:space:]]Exploiting[[:space:]]Cross-Field[[:space:]]Frequency[[:space:]]Correlation[[:space:]]for[[:space:]]NIR-Assisted[[:space:]]Image[[:space:]]Denoising/c9ee5d0d-57af-4b5c-aacf-8ca5596906ca_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Completion[[:space:]]as[[:space:]]Enhancement_[[:space:]]A[[:space:]]Degradation-Aware[[:space:]]Selective[[:space:]]Image[[:space:]]Guided[[:space:]]Network[[:space:]]for[[:space:]]Depth[[:space:]]Completion/5b8a9777-b777-4fb5-9ba2-85ebeb5a648a_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Complexity[[:space:]]Experts[[:space:]]are[[:space:]]Task-Discriminative[[:space:]]Learners[[:space:]]for[[:space:]]Any[[:space:]]Image[[:space:]]Restoration/811da5ba-4d1d-41b9-ac9b-e649712870d3_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Composing[[:space:]]Parts[[:space:]]for[[:space:]]Expressive[[:space:]]Object[[:space:]]Generation/edecee1f-cf3a-417c-a3d7-2ea937e0cfb0_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Compositional[[:space:]]Caching[[:space:]]for[[:space:]]Training-free[[:space:]]Open-vocabulary[[:space:]]Attribute[[:space:]]Detection/5997b1e0-b306-408f-91ce-2ab833ab2009_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Compositional[[:space:]]Targeted[[:space:]]Multi-Label[[:space:]]Universal[[:space:]]Perturbations/616cefba-8863-49bf-987c-2e525bd606b6_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Comprehensive[[:space:]]Information[[:space:]]Bottleneck[[:space:]]for[[:space:]]Unveiling[[:space:]]Universal[[:space:]]Attribution[[:space:]]to[[:space:]]Interpret[[:space:]]Vision[[:space:]]Transformers/abcc8427-3423-466d-afdc-b93969a43358_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Comprehensive[[:space:]]Relighting_[[:space:]]Generalizable[[:space:]]and[[:space:]]Consistent[[:space:]]Monocular[[:space:]]Human[[:space:]]Relighting[[:space:]]and[[:space:]]Harmonization/743d9653-d63f-4670-9d09-7ac226ec0ec0_origin.pdf filter=lfs diff=lfs merge=lfs -text diff --git a/2025/Channel Consistency Prior and Self-Reconstruction Strategy Based Unsupervised Image Deraining/1cd88811-9c34-4363-ae8a-2a52dd5a816f_content_list.json b/2025/Channel Consistency Prior and Self-Reconstruction Strategy Based Unsupervised Image Deraining/1cd88811-9c34-4363-ae8a-2a52dd5a816f_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..89673a83ca5c464e17bb927deed64a6077cb7512 --- /dev/null +++ b/2025/Channel Consistency Prior and Self-Reconstruction Strategy Based Unsupervised Image Deraining/1cd88811-9c34-4363-ae8a-2a52dd5a816f_content_list.json @@ -0,0 +1,1562 @@ +[ + { + "type": "text", + "text": "Channel Consistency Prior and Self-Reconstruction Strategy Based Unsupervised Image Deraining", + "text_level": 1, + "bbox": [ + 91, + 128, + 906, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Guanglu Dong $^{1}$ Tianheng Zheng $^{1}$ Yuanzhouhan Cao $^{2}$ Linbo Qing $^{1}$ Chao Ren $^{1*}$ $^{1}$ Sichuan University, Chengdu, China \n $^{2}$ Beijing Jiaotong University, Beijing, China \n{dongguanglu, zhengtianheng}@stu.scu.edu.cn, yzhcao@bjtu.edu.cn \n{qing.lb, chaoren}@scu.edu.cn", + "bbox": [ + 111, + 202, + 890, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 308, + 326, + 324 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, deep image deraining models based on paired datasets have made a series of remarkable progress. However, they cannot be well applied in real-world applications due to the difficulty of obtaining real paired datasets and the poor generalization performance. In this paper, we propose a novel Channel Consistency Prior and Self-Reconstruction Strategy Based Unsupervised Image Deraining framework, CSUD, to tackle the aforementioned challenges. During training with unpaired data, CSUD is capable of generating high-quality pseudo clean and rainy image pairs which are used to enhance the performance of deraining network. Specifically, to preserve more image background details while transferring rain streaks from rainy images to the unpaired clean images, we propose a novel Channel Consistency Loss (CCLoss) by introducing the Channel Consistency Prior (CCP) of rain streaks into training process, thereby ensuring that the generated pseudo rainy images closely resemble the real ones. Furthermore, we propose a novel Self-Reconstruction (SR) strategy to alleviate the redundant information transfer problem of the generator, further improving the deraining performance and the generalization capability of our method. Extensive experiments on multiple synthetic and real-world datasets demonstrate that the deraining performance of CSUD surpasses other state-of-the-art unsupervised methods and CSUD exhibits superior generalization capability. Code is available at https://github.com/GuangluDong0728/CSUD.", + "bbox": [ + 88, + 340, + 485, + 750 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 777, + 220, + 792 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Image deraining is an important task in low-level computer vision [7, 10, 25, 29, 31, 49]. It aims to reconstruct a clean image from its rainy version, which has a profound impact on subsequent high-level tasks such as object detection, recognition, and semantic segmentation", + "bbox": [ + 89, + 801, + 483, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/91b3619f60909abcf4ab764b2b6370bcc5d2ec68f025545256d4d7d1fcc1fb6f.jpg", + "image_caption": [ + "Figure 1. Deraining results on the real rainy images captured by ourselves in real-world scenarios. Compared with the supervised method Restormer [49] and the unsupervised method DerainCycleGAN [40], our CSUD exhibits extremely strong generalization capability in the real world and achieves the best visual results." + ], + "image_footnote": [], + "bbox": [ + 514, + 306, + 906, + 542 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "[24, 26, 33, 41]. To mitigate the negative impact of rain, numerous deep learning based supervised image deraining methods [7, 8, 11, 13, 19, 21, 32, 36, 47, 52] have been proposed and they achieve remarkable progress on synthetic datasets. However, the performance of these methods is largely constrained by the difficulty of capturing a large number of high-quality paired datasets when applied to real-world scenarios. Additionally, attempting to directly apply supervised models trained on synthetic datasets to real-world scenarios is also difficult because of the huge domain gap between synthetic and real-world rain [45] and the overfitting issue of supervised methods [9].", + "bbox": [ + 511, + 643, + 908, + 824 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Compared to supervised methods, unsupervised deep learning methods [3, 14, 40, 42, 46] for image deraining tasks generally achieve end-to-end image deraining without paired clean-rainy images during training. This allows unsupervised methods to effectively handle complex real-world", + "bbox": [ + 511, + 825, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author", + "bbox": [ + 109, + 887, + 233, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "7469", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "scenarios, especially when data collection costs are high. However, due to the high difficulty and challenges of unsupervised deraining methods in training process and the lack of prior knowledge [40], there is still very little relevant research in this field. Additionally, the deraining performance of existing unsupervised methods always lags far behind that of supervised methods, and they face the same generalization issues as supervised methods.", + "bbox": [ + 89, + 90, + 485, + 210 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "As shown in Fig. 1, even the SOTA supervised and unsupervised deraining models [40, 49] encounter substantial generalization issues and struggle to achieve satisfactory results when dealing with rainy images captured in real world, and sometimes they even fail to remove any rain streaks due to the more complex rain distributions in real world. Based on the above analysis, how to design an effective unsupervised deraining framework to avoid the need for real paired data and how to improve the generalization performance of deraining methods have become key research issues.", + "bbox": [ + 89, + 215, + 485, + 366 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we propose a novel channel consistency prior and self-reconstruction strategy based unsupervised image deraining framework which only needs unpaired data to train the network and can be easily generalized to real-world scenarios. We first conduct detailed analysis of the channel consistency prior (CCP) of rain streaks, and then propose a novel channel consistency loss to constrain the generator to synthesize high-quality pseudo rainy images. In addition, we utilize the proposed self-reconstruction (SR) strategy to further improve the performance of the generator and enhance the generalization capability of the derainer. As shown in Figure 1, our CSUD achieves the most promising performance on the real-world captured images compared to both supervised and unsupervised deraining methods. The main contributions of this work are summarized below:", + "bbox": [ + 89, + 369, + 483, + 594 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a novel channel consistency prior and self-reconstruction strategy based unsupervised image deraining framework which is able to generate high-quality pseudo paired data for training. This framework can effectively address the issues of insufficient paired data in real world and significantly improve the generalization capability of the network.", + "- We propose a novel channel consistency loss (CCLoss) according to the channel consistency prior (CCP) of rain streaks to enhance the performance of the generator, preserving more background details while generating pseudo rainy images. To the best of our knowledge, we are the first to incorporate CCP into the training process of unsupervised deraining framework.", + "- We propose a self-reconstruction (SR) strategy to solve the redundant information transfer problem of the generator. SR strategy can facilitate the accuracy of rain streaks transfer in the generator and enhance the deraining performance and the generalization capability of the derainer without altering the existing network structure." + ], + "bbox": [ + 89, + 598, + 483, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- Extensive experiments on numerous benchmark datasets demonstrate that our proposed unsupervised framework can effectively tackle the deraining tasks on both synthetic and real-world rainy images with extremely superior generalization capability in real-world scenarios.", + "bbox": [ + 513, + 90, + 908, + 167 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 513, + 181, + 663, + 198 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Deep Learning Based Single Image Deraining", + "text_level": 1, + "bbox": [ + 511, + 208, + 898, + 224 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Supervised Single Image Deraining. Recently, with the development of deep learning, numerous excellent single image deraining [4, 5, 7, 8, 13, 21, 36, 47, 52] models have emerged. Fu et al. [7] first employ the CNN-based DerainNet to extract and remove the rain layer. Lin et al. [21] propose a generative network equipped with the pixel-wise status estimation and information decoupling for deraining. Lin et al. [23] propose a unified network for low-light-rainy image restoration to jointly handle low-light enhancement and deraining tasks. In addition to the aforementioned works specifically targeting deraining, some networks designed for image restoration [1, 16, 27, 29, 35, 38, 48, 49] have also shown excellent performance in deraining tasks. However, the above fully supervised methods require a large number of high-quality paired training samples, which is almost impossible in real-world scenarios. Simultaneously, due to the overfitting issue, their generalization capability is limited.", + "bbox": [ + 511, + 229, + 906, + 487 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Unsupervised Single Image Deraining. To address the lack of paired datasets in the real world, some outstanding unsupervised deraining methods [3, 14, 40, 42, 46] have been proposed. Wei et al. [40] propose an unsupervised method for rain removal and generation by utilizing the popular CycleGAN [53] framework. Then, through introducing a dual contrastive learning manner to the CycleGAN framework, Chen et al. [3] develop an unpaired adversarial framework which effectively explores mutual properties of the unpaired rainy-clean samples. Ye et al. [46] introduce a novel non-local contrastive learning based unsupervised image deraining method to better distinguish rain streaks from clean images. However, the deraining performance of these unsupervised methods is far inferior to that of supervised methods, especially on real-world scenarios. Compared with previous methods, our proposed CSUD can greatly enhance the unsupervised deraining performance and achieve competitive results compared with some supervised methods.", + "bbox": [ + 511, + 488, + 908, + 760 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Generalization Problem in Image Deraining", + "text_level": 1, + "bbox": [ + 511, + 772, + 885, + 787 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Deep image deraining models often fail to obtain satisfactory performance or even cannot remove any rain streaks when they are applied to real-world scenarios. Generalization problem of deraining models is an urgent issue to be solved. Gu et al. [9] conduct thorough analysis of the generalization problem in image deraining and propose to achieve better generalization by simplifying the complexity of train-", + "bbox": [ + 511, + 795, + 908, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "7470", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/81aba27eb357ed84b6880f74088d87ca8afe2c9ffe2db10b415cffc043c55093.jpg", + "image_caption": [ + "Figure 2. Framework of the proposed CSUD. The red arrows represent the baseline, purple arrows represent the 3 extra constraints, green arrows represent the self-reconstruction (SR) strategy, and blue arrows represent the CCLoss. The 4 input arrows of RIEM represent 4 separate runs of the generator, each time guided by a different image. Both RIEM and CFEM receive only 1 input image at a time." + ], + "image_footnote": [], + "bbox": [ + 91, + 89, + 906, + 378 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ing images. Some semi-supervised learning frameworks [39, 44] are proposed to analyze the residual difference between domains and enhance the deraining capability in real world. In addition, a series of unsupervised deraining methods [3, 40, 46] have also demonstrated a certain degree of generalization capability in real-world scenarios. However, they have not conducted detailed analysis on how to enhance the generalization capability of unsupervised methods. Inspired by [9], we propose a self-reconstruction strategy to make both generator and derainer pay more attention to the reconstruction of image background, which significantly improves the generalization performance of our method.", + "bbox": [ + 88, + 450, + 485, + 633 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 89, + 651, + 225, + 669 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Overall Framework of Proposed Method", + "text_level": 1, + "bbox": [ + 89, + 678, + 439, + 694 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To address the issue of insufficient paired data and the generalization problem in deraining applications, we propose a novel channel consistency prior and self-reconstruction strategy based unsupervised image deraining framework, CSUD. As illustrated in Fig. 2, our CSUD consists of a derainer $Der$ , a rainy image generator $G$ , and a Patch-GAN [12] discriminator $D$ . For $G$ , it consists of a clean feature extraction module (CFEM), a rain information extraction module (RIEM) and 6 residual blocks. Detailed network structures are detailed in the Appendix.", + "bbox": [ + 89, + 702, + 485, + 853 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In generator $G$ , we effectively utilize unpaired clean and rainy images to synthesize pseudo rainy images. As shown by the red arrows in Fig. 2, we design a GAN-based baseline", + "bbox": [ + 89, + 854, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "that transitions from \" $x \\in X \\to x_{s1} \\in S \\to x_{der} \\in N$ \", where $X$ , $S$ and $N$ denote clean images, synthetic rainy images and derained images respectively. $G$ transforms clean image $x$ to synthetic rainy image $x_{s1}$ according to the guide of real rainy image $y$ . In order to better extract rain information and transfer it to the clean images, $y$ and $x$ are initially processed by the RIEM and the CFEM respectively to obtain rainy features and clean features which are concatenated for subsequent process. For the real rainy image $y$ and the synthesized rainy image $x_{s1}$ , the corresponding derained clean images $y_{der}$ and $x_{der}$ are obtained through Der.", + "bbox": [ + 511, + 450, + 906, + 617 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The discriminator $D$ is trained to distinguish between synthetic rainy images $x_{s1} \\in S$ and real rainy images $y \\in Y$ , and the adversarial loss $L_{adv1}$ of the baseline is constrained between them. Furthermore, motivated by existing methods [22, 53], as indicated by the purple arrows in Fig. 2, we add another three processes, $G(Der(x_{s1}), x_{s1}) \\to x_{s2}$ , $G(Der(y), y) \\to y_{s1}$ , and $G(Der(y), x_{s1}) \\to y_{s2}$ , to the framework to enhance training stability. Then, additional three adversarial losses are introduced to CSUD:", + "bbox": [ + 511, + 617, + 908, + 752 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nL _ {G A N} = L _ {a d v 1} + L _ {a d v 2} + L _ {a d v 3} + L _ {a d v 4}. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 562, + 762, + 906, + 779 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Channel Consistency Loss for Generator", + "text_level": 1, + "bbox": [ + 511, + 787, + 862, + 803 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Channel Consistency Prior (CCP) of Rain Streaks. In RGB rainy images, most rain streaks tend to be consistent across the R, G, and B channels, meaning the pixel values of most rain streaks in the three channels are highly similar. As shown in the second, third, and fourth columns of Fig. 3 (a), we separate both synthetic and real-world rainy images", + "bbox": [ + 511, + 809, + 908, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "7471", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/e4c62cb2741d1f3dc28cd2a762e51ef3355402c6d05a920b289d8b2531e9b54e.jpg", + "image_caption": [ + "(a) Visualization of channel consistency prior in synthetic, real-world, and nighttime real-world rainy images" + ], + "image_footnote": [], + "bbox": [ + 91, + 85, + 903, + 359 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/491b20f20ed41772d50a06a006d3b60238cd9753a4e6ce065f6e7a2f8075a132.jpg", + "image_caption": [ + "(b) Cosine similarities of cycle subtractions among R, G, and B channels between clean and rainy images", + "Figure 3. (a) Visualization of channel consistency prior in synthetic and real-world rainy images. The first column presents the rainy RGB images; the second, third and fourth columns present their R, G, B channels, respectively; the fifth, sixth, and seventh columns present the cycle subtractions of R, G, and B channels, respectively. (b) Cosine similarities of the cycle subtractions of R, G, and B channels between clean and rainy images in the 5 synthetic and real-world datasets. More visualizations and analysis are presented in the Appendix." + ], + "image_footnote": [], + "bbox": [ + 91, + 382, + 901, + 474 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "into R, G, and B channels and visualized them. It can be seen clearly that while there are significant differences in the background parts among the three channels, the rain streaks tend to be consistent. We propose to term this phenomenon as \"Channel Consistency Prior (CCP)\" of rainy streaks.", + "bbox": [ + 88, + 578, + 482, + 654 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Theoretically, according to CCP, the cycle subtractions $((r)_R - (r)_G)$ , $((r)_G - (r)_B)$ , $((r)_B - (r)_R)$ of R, G, and B channels in rainy images will almost contain no rain streaks and should be close to $((n)_R - (n)_G)$ , $((n)_G - (n)_B)$ , $((n)_B - (n)_R)$ in clean images, where $r$ and $n$ represent the rainy and clean image, $(\\cdot)_R$ , $(\\cdot)_G$ , and $(\\cdot)_B$ denote R, G, and B channels respectively. In the fifth, sixth, and seventh columns of Fig. 3 (a), it can be observed that the cycle subtractions of R, G, and B channels nearly don't contain any rain streaks, qualitatively confirming the correctness of CCP. Furthermore, as shown in Fig. 3 (b), we quantitatively present the cosine similarities of the cycle subtractions of R, G, and B channels between clean and rainy images in both synthetic and real-world benchmark datasets [18, 20, 36, 43]. It is evident that the cosine similarities are exceptionally high, which can further demonstrate our hypothesis.", + "bbox": [ + 89, + 659, + 485, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Channel Consistency Loss. Based on the above analysis of CCP, the cycle subtractions of R, G, and B channels between input clean image and output pseudo rainy image of the generator $G$ should tend to be consistent. To this end, we propose a novel Channel Consistency Loss (CCLoss) to preserve more color and texture details of the background while generating pseudo rainy images:", + "bbox": [ + 511, + 578, + 906, + 684 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} L _ {C C} = \\left\\| \\left(\\left(x _ {s 1}\\right) _ {R} - \\left(x _ {s 1}\\right) _ {G}\\right), \\left(x _ {R} - x _ {G}\\right) \\right\\| _ {1} \\\\ + \\left\\| \\left(\\left(x _ {s 1}\\right) _ {G} - \\left(x _ {s 1}\\right) _ {B}\\right), \\left(x _ {G} - x _ {B}\\right) \\right\\| _ {1} \\tag {2} \\\\ + \\left\\| \\left(\\left(x _ {s 1}\\right) _ {B} - \\left(x _ {s 1}\\right) _ {R}\\right), \\left(x _ {B} - x _ {R}\\right) \\right\\| _ {1}, \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 552, + 689, + 906, + 744 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $(x_{s1})_R, (x_{s1})_G, (x_{s1})_B, x_R, x_G, x_B$ denotes R, G, B color channels of $x_{s1}$ and $x, ||\\cdot||_1$ represents the $L_1$ norm and $L_{CC}$ represents the proposed CCLoss, respectively. By introducing CCP into our unsupervised deraining framework, $G$ will generate more accurate synthesized rainy images, which will further improve the performance of the derainer.", + "bbox": [ + 511, + 750, + 908, + 842 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Self-Reconstruction Strategy", + "text_level": 1, + "bbox": [ + 511, + 848, + 769, + 864 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Redundant Information Transfer Problem of the Generator. Although adversarial losses can effectively constrain the", + "bbox": [ + 511, + 869, + 908, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "7472", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/7caee91250ebf886ef717848f36f2812849980b90b8299834d7a96c24fd3f045.jpg", + "image_caption": [ + "Figure 4. Visualization of redundant information transfer problem." + ], + "image_footnote": [], + "bbox": [ + 93, + 87, + 486, + 239 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "training process of generator $G$ , the data synthesis process of $G$ is highly susceptible to interference from the redundant information in rainy images, especially those edge textures that are very similar to rain streaks. As shown in Fig. 4, the edge textures of some background objects in rainy images, such as the outline of the woman's hair, the leg of rhino, etc., are mistakenly transferred to the clean images by $G$ . As a result, $G$ will generate inaccurate pseudo rainy images, interfering with the learning process of $Der$ .", + "bbox": [ + 88, + 277, + 483, + 412 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Self-Reconstruction Strategy. To solve the redundant information transfer problem, we propose a novel Self-Reconstruction (SR) strategy, which enables $G$ to extract more accurate rain streak information and synthesizes more realistic pseudo rainy images. The core idea of SR strategy is to encourage $G$ to better distinguish rain streaks from background in the rainy images, ensuring that only rain streak information is extracted for subsequent generation of pseudo rainy images, while redundant information of background is disregarded. As shown by the green arrows in Fig. 2, when the input clean image $x$ itself, the derained image $y_{der}$ are input with clean image $x$ into $G$ , $G$ learns to reconstruct the input clean image $x$ itself without any additional redundant information from $x$ and $y_{der}$ , which means that the self-reconstructed images $x_{s1}$ and $x_{s2}$ should be close to $x$ . Thus, we can utilize the following self-reconstruction loss (SRLoss) as additional training constraint of $G$ :", + "bbox": [ + 89, + 414, + 483, + 672 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {S R - G} = \\left\\| G (x, x), x \\right\\| _ {1} + \\left\\| G (x, y _ {d e r}), x \\right\\| _ {1}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 130, + 681, + 483, + 699 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $L_{SR - G}$ represents the SRLoss for $G$ . SR strategy allows $G$ to pay more attention to reconstructing the background of clean images while learning the transfer of rain streaks, significantly improving the accuracy of generating pseudo rainy images. Under the constraint of $L_{SR - G}$ , $G$ will obtain the ability to generate clean images themselves guided by images without rain streaks. Therefore, the cleaner the derained images $y_{der}$ and $x_{der}$ restored by $Der$ are, the closer the self-reconstructed images $x_{s2}$ and $x_{s3}$ are to the original clean ones $x$ . Then, the following SRLoss can be utilized to further improve the performance of $Der$ :", + "bbox": [ + 89, + 707, + 483, + 875 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {S R - D e r} = \\left\\| G (x, x _ {d e r}), x \\right\\| _ {1} + \\left\\| G (x, y _ {d e r}), x \\right\\| _ {1}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 104, + 883, + 483, + 902 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/c16c3a781b06600c2afbcb800eb781ba306111a89fd058d7ada04e9aab946cb7.jpg", + "image_caption": [ + "Figure 5. The implementation principles of SR strategy." + ], + "image_footnote": [], + "bbox": [ + 516, + 88, + 905, + 244 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $L_{SR - Der}$ represents the SRLoss for $Der$ . Under the constraint of $L_{SR - Der}$ , $Der$ will also focus more on reconstructing the background content while learning to derain, which greatly enhances the generalization capability of $Der$ .", + "bbox": [ + 511, + 291, + 906, + 351 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As shown in the third and fourth columns of Figure 4, SR strategy effectively alleviates the redundant information transfer problem. Through SR strategy, $G$ can achieve robust rain streak transfer with more background detail fidelity, greatly enhancing the deraining performance of Der.", + "bbox": [ + 511, + 352, + 906, + 429 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Loss Function", + "text_level": 1, + "bbox": [ + 511, + 439, + 658, + 454 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Due to the high difficulty and instability of training a GAN-based unsupervised framework, in addition to adopting the L1 loss for $Der$ , we also incorporate SSIM loss and perceptual loss[15] to enhance the performance of $Der$ :", + "bbox": [ + 511, + 460, + 906, + 523 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} L _ {D e r} = \\left\\| x, x _ {d e r} \\right\\| _ {1} + \\lambda_ {1} L _ {S S I M} (x, x _ {d e r}) \\tag {5} \\\\ + \\lambda_ {2} | | \\phi_ {i} (x), \\phi_ {i} (x _ {d e r}) | | _ {1} + \\lambda_ {3} L _ {S R - D e r}, \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 552, + 535, + 905, + 571 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $||\\cdot ||_1$ represents the $L_{1}$ norm, and $L_{SSIM}(\\cdot)$ represents the SSIM loss, $\\phi_i(\\cdot)$ is the layers of pre-trained VGG-19 [34], $\\lambda_1, \\lambda_2, \\lambda_3$ represent the hyperparameters for SSIM loss, perceptual loss and SRLoss for derainer respectively.", + "bbox": [ + 511, + 583, + 906, + 643 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Finally, we obtain the total loss to train our framework:", + "bbox": [ + 531, + 645, + 895, + 660 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {t o t a l} = \\min _ {G} \\max _ {D} L _ {G A N} + L _ {D e r} + \\alpha_ {1} L _ {C C} + \\alpha_ {2} L _ {S R - G}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 513, + 672, + 906, + 705 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\alpha_{1},\\alpha_{2}$ represent the hyperparameters for CCLoss and SRLoss for generator respectively.", + "bbox": [ + 511, + 707, + 905, + 738 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 753, + 643, + 770 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To demonstrate the effectiveness of our CSUD, following [18, 40], we conduct experiments from the following two aspects: (1) unsupervised deraining performance, (2) generalization performance. In experiment (1), we train independent models for different datasets respectively, including both synthetic and real-world datasets. In experiment (2), we only train our model on the synthetic dataset Rain100L [43], and then test on various real-world and nighttime datasets.", + "bbox": [ + 511, + 779, + 908, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "7473", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/a53ce33e3dabe10f4ae742cf45db599b43c94d051e0ed2d52e1a46d8b882ca41.jpg", + "table_caption": [ + "Table 1. Quantitative deraining performance comparisons of different methods. Red and blue colors indicate the 1st and 2nd ranks among the unsupervised methods, respectively. S and U indicate supervised method and unsupervised method, respectively." + ], + "table_footnote": [], + "table_body": "
DatasetsRain100L [43]Rain100H [43]Rain12 [20]Rain800 [17]RealRain1K-L [18]RealRain1K-H [18]
MetricsPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIM
DerainNet [7] (S)27.030.88414.920.592//22.770.81027.090.92522.880.889
DDN [8] (S)32.380.92624.640.84934.040.93321.160.73231.180.91729.170.878
SPA-Net [36] (S)31.950.91926.070.857//24.370.86130.430.94725.760.910
MPRNet [48] (S)34.950.95928.530.87236.840.96428.680.87536.290.97234.740.964
NAFNet [2] (S)37.000.97829.660.90034.810.94328.720.88638.800.98636.110.976
Restormer [49] (S)37.570.97429.460.889//30.330.90540.900.98539.570.981
PromptIR [29] (S)38.340.98328.690.87735.090.94529.070.88736.990.97333.610.953
NeRD-Rain-S [5] (S)42.000.99032.860.93235.390.94228.820.87838.640.97936.690.970
CycleGAN [53] (U)24.610.83420.590.70421.560.84523.950.81920.190.82017.530.762
DerainCycleGAN [40] (U)31.490.93621.130.71033.520.94024.320.84228.160.90124.780.844
NLCL [46] (U)20.500.71922.310.72822.680.73518.940.67023.060.83220.640.719
DCD-GAN [3] (U)31.820.94122.470.75331.560.92425.610.81330.490.93927.820.895
CSUD (Ours)33.280.95424.420.80834.560.95123.580.79532.710.95929.670.928
", + "bbox": [ + 135, + 128, + 864, + 323 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f9de6d25327ccdd6df61d7a9a744fa66437396fe6b1ee7919f4dfff4e9fca88d.jpg", + "image_caption": [ + "Inupt" + ], + "image_footnote": [], + "bbox": [ + 93, + 333, + 194, + 435 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1dac05848eea7e9aca11190c9dd56fe020b9580b74d69830629b4a75156140e5.jpg", + "image_caption": [ + "GT" + ], + "image_footnote": [], + "bbox": [ + 194, + 334, + 294, + 435 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/9daef77910392ccbfcf76ea0e4bb9b001cfca1e5b7295a56fdbf8c7714ffe83e.jpg", + "image_caption": [ + "MPRNet", + "Figure 6. Qualitative deraining performance comparisons on Rain100L [43] and Rain12 [20] datasets. Our CSUD achieves competitive visual results comparable to the SOTA supervised method NeRD-Rain-S [5]." + ], + "image_footnote": [], + "bbox": [ + 295, + 334, + 395, + 435 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/63968dd4f3ae89da832eabfd8b853722d62f175023e1820d335aa2004f8dafd3.jpg", + "image_caption": [ + "NeRD-Rain-S" + ], + "image_footnote": [], + "bbox": [ + 397, + 334, + 496, + 435 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ca3143bc1f9c89048715d6ecc1958e38d7e10371e8f130088a13968a56f55ae1.jpg", + "image_caption": [ + "DerainCycleGAN" + ], + "image_footnote": [], + "bbox": [ + 496, + 334, + 598, + 435 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/3873f182fc80c2779c88202905478da033376e8f7d52bfe58fbeb3630bb1b205.jpg", + "image_caption": [ + "NLCL" + ], + "image_footnote": [], + "bbox": [ + 599, + 334, + 700, + 435 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/979e7b42e4db60f7988a61748a50e6a04153c2010f31f91ab4711fa7cece0281.jpg", + "image_caption": [ + "DCD-GAN" + ], + "image_footnote": [], + "bbox": [ + 700, + 334, + 802, + 435 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/6f7559186be8aced74c8b3a604de130df151533a002ff09894026a86008bc00b.jpg", + "image_caption": [ + "CSUD (Ours)" + ], + "image_footnote": [], + "bbox": [ + 803, + 334, + 903, + 435 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets and Metrics. We utilize 10 challenging benchmark datasets in our study, including synthetic datasets: Rain100L [43], Rain100H [43], Rain12 [20], Rain800 [17], real-world datasets: RealRain1K-L [18], RealRain1K-H [18], SPA-data [36], RainDS [30], Internet-Data [39], and a nighttime dataset which is based on GTAV-NightRain dataset [50] and enlarged the rain density by [51], we name this nighttime dataset as Night-Rain in the following experiments. As for RainDS, we only use the real rain streaks subset for testing. Following existing methods [3, 40, 49], we use PSNR and SSIM [37] to evaluate different methods.", + "bbox": [ + 88, + 508, + 483, + 674 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details. As for derainer, we employ the CNN-based simple image restoration baseline NAFNet [2]. Our framework is implemented by PyTorch [28] with a GeForce RTX 3090 GPU. We adopt the Adam optimizer [6] to train our network for 200 epochs with the initial learning rate of $1e^{-4}$ , followed by another 100 epochs with a learning rate of $1e^{-5}$ . All training images are randomly cropped to $256 \\times 256$ patches in an unpaired manner. The hyperparameters $\\lambda_{1}, \\lambda_{2}$ , and $\\lambda_{3}$ in Eq. 5 are set to 1, 0.2 and 0.5, while $\\alpha_{1}$ and $\\alpha_{2}$ in Eq. 6 are set to 10 and 5, respectively.", + "bbox": [ + 88, + 675, + 483, + 825 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Unsupervised Deraining Results", + "text_level": 1, + "bbox": [ + 89, + 833, + 375, + 849 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We compare our method with unsupervised methods from the most recent years, including [3, 40, 46, 53]. Noting that the related methods for comparison are limited, since only a few", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "unsupervised deraining models have been proposed in the field. We also list some SOTA supervised methods, including [2, 5, 7, 8, 29, 36, 48, 49]. Tab. 1 presents the quantitative deraining results of different methods. It is evident that our CSUD significantly outperforms other unsupervised methods on most datasets. Compared to DCD-GAN[3], our CSUD respectively achieves 1.46dB, 3.00dB, and 2.22dB improvement in PSNR on Rain100L, Rain12 and RealRain1K-L datasets. Additionally, our CSUD even achieves competitive results comparable to several supervised methods. As shown in Fig. 6, compared to other unsupervised methods, our CSUD achieves better deraining results while preserving more color and texture details of image background.", + "bbox": [ + 511, + 508, + 908, + 705 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Generalization Results", + "text_level": 1, + "bbox": [ + 511, + 712, + 723, + 727 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To validate the generalization capability of CSUD, we compare the results on 5 real-world datasets[18, 30, 36, 39] and a nighttime dataset [51] with SOTA supervised and unsupervised methods, including [2, 3, 5, 7, 29, 40, 46, 48, 49]. Noting that there are no ground truth images in the InternetData [39] dataset, thereby only visual results are presented. All methods are trained on synthetic datasets and tested on the unseen real-world and nighttime datasets. A prevailing perspective is that supervised methods usually outperform unsupervised methods. But as shown in Tab. 2, our CSUD almost comprehensively surpasses supervised models on", + "bbox": [ + 511, + 734, + 908, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "7474", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/1eec57afa8d029abcfbd907d5e7b00c23dba0d5e60884060118b60784da07444.jpg", + "table_caption": [ + "Table 2. Quantitative generalization performance comparisons of different deraining methods. Red and blue colors indicate the 1st and 2nd ranks among the unsupervised methods. S, U indicate supervised and unsupervised methods, respectively." + ], + "table_footnote": [], + "table_body": "
DatasetsRealRain1K-L [18]RealRain1K-H [18]SPA-data [36]RainDS [30]Night-Rain [51]
MetricsPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIM
DerainNet [7] (S)25.400.84721.910.74230.780.91221.350.591//
MPRNet [48] (S)26.750.89923.580.84033.000.94223.240.65626.130.817
NAFNet [2] (S)28.110.91524.360.84833.460.94122.480.62526.150.816
Restormer [49] (S)26.290.88323.130.85132.620.93622.960.64825.850.808
PromptIR [29] (S)28.140.91524.280.85133.500.94022.480.62526.370.815
NeRD-Rain-S [5] (S)27.640.90423.960.83133.030.93322.370.62226.380.810
DerainCycleGAN [40] (U)28.250.89924.200.82133.530.93718.750.46025.850.802
NLCL [46] (U)20.180.67419.810.62918.760.72717.000.47722.010.475
DCD-GAN [3] (U)27.490.91223.860.84531.550.93022.130.61025.540.798
CSUD (Ours)29.210.92825.450.87833.570.93922.540.61325.940.805
", + "bbox": [ + 174, + 128, + 823, + 280 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f1dade8c00373d7dcb13ed0742c934d8288abc0a7dd1f2b90eb42f9fcc244d68.jpg", + "image_caption": [ + "Figure 7. Qualitative generalization results on RealRain1K-L [18] and RealRain1K-H [18] real-world datasets. Our CSUD achieves best visual results among both SOTA supervised and unsupervised methods in real-world scenes." + ], + "image_footnote": [], + "bbox": [ + 91, + 290, + 906, + 450 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/9f857987c8fa188a993dddfba19768f8faadfd7501798df20864f426f8083155.jpg", + "table_caption": [ + "Table 3. Effect of CCLoss and SR strategy on CSUD." + ], + "table_footnote": [], + "table_body": "
CCSRRain100L[43]Rain12[20]RealRain1K-L[18]
PSNRSSIMPSNRSSIMPSNRSSIM
XX30.660.92531.600.93130.420.932
✓X31.240.93432.620.94831.140.943
X✓32.690.94933.430.93931.860.950
✓✓33.280.95434.560.95132.710.959
", + "bbox": [ + 101, + 534, + 473, + 619 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "the first 4 real-world datasets, while remaining competitive on the nighttime dataset. This strongly demonstrates the excellent generalization ability of our method.", + "bbox": [ + 89, + 643, + 482, + 686 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As shown in Fig. 7 and Fig. 8, the performance of some most advanced supervised methods like Restormer [49] and NeRD-Rain-S [5] is extremely poor when directly applied to real-world scenarios. In contrast, our CSUD achieves better visual results in real world. It is worth mentioning again that, as shown in Fig. 1 in Sec. 1, our method also shows extremely superior deraining performance on the rainy images captured by ourselves in real scenarios, which further demonstrates our excellent generalization capability.", + "bbox": [ + 89, + 688, + 483, + 824 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Studies", + "text_level": 1, + "bbox": [ + 89, + 833, + 254, + 848 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effect of CCLoss and SR Strategy on Deraining Performance. To verify the effectiveness of CCLoss and SR strategy, we conduct ablation experiments on Rain100L [43],", + "bbox": [ + 89, + 854, + 485, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Rain12 [20] and RealRain1K-L [18] datasets. As shown in Table 3, compared to the baseline, the effect on PSNR is improved by 1.38dB, 1.02dB, and 0.72dB among the above three datasets with CCLoss, respectively. And the effect is improved by 1.38dB, 1.83dB, and 1.44dB with SR strategy, respectively. Furthermore, when both CCLoss and SR strategy are introduced into the baseline, the PSNR on the three datasets is increased by 2.62dB, 2.96dB, and 2.29dB respectively. The above ablation results strongly verify the superiority and importance of CCLoss and SR strategy.", + "bbox": [ + 511, + 510, + 906, + 662 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effect of CSUD framework on Generalization Performance. To validate the effectiveness of our CSUD in improving generalization performance and the universality and robustness of our method, we also conduct experiments incorporating CSUD framework on another 2 different baselines: NeRD-Rain-S [5] and Restormer [49]. Fig. 9 shows the PSNR improvements of different CSUD based models compared to their supervised version. Notably, we train all the above models on Rain100L and test them on the 3 real-world datasets. This strongly demonstrates the effectiveness of CSUD in improving generalization performance.", + "bbox": [ + 511, + 667, + 908, + 834 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effect of Different Training Datasets on Generalization Performance. To further validate the generalization performance of CSUD, we employ 4 different CSUD models which are respectively trained on 4 synthetic and real-", + "bbox": [ + 511, + 839, + 908, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7475", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9e2ad534a36fae4ed759ec9e803bc5e689719c4047cc14cc0afe3695a67f4b50.jpg", + "image_caption": [ + "Input" + ], + "image_footnote": [], + "bbox": [ + 93, + 88, + 194, + 220 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/089d8648dcf54bd2f09675466293f2985658341b3980468cce420ba34a9b0241.jpg", + "image_caption": [ + "Restormer" + ], + "image_footnote": [], + "bbox": [ + 194, + 88, + 295, + 220 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e864e12231b930dc1815ec087f306d4e93ebca13d00514a5e2ab66f587f5a422.jpg", + "image_caption": [ + "PromptIR" + ], + "image_footnote": [], + "bbox": [ + 295, + 88, + 395, + 220 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/910033d295110dce184f4f188b6e95efcf811c662b69612b8c00793f803be3df.jpg", + "image_caption": [ + "NeRD-Rain-S" + ], + "image_footnote": [], + "bbox": [ + 397, + 88, + 496, + 220 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/cf0cd317f077a293d450fa3141c3a3d2c265d39d32271e057a3f556baeab7812.jpg", + "image_caption": [ + "DerainCycleGAN", + "Figure 8. Qualitative generalization results on Internet-Data [39] real-world dataset." + ], + "image_footnote": [], + "bbox": [ + 496, + 88, + 599, + 220 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5615b8809e10dc6283fdabb209b5a1375f16766b7c6e5aee3a0f498e83678005.jpg", + "image_caption": [ + "NLCL" + ], + "image_footnote": [], + "bbox": [ + 599, + 88, + 700, + 220 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3959e7251f64ee540ee6f0d02185a4b4ceba7ae0799c008e5034e13d8c14c8d1.jpg", + "image_caption": [ + "DCD-GAN" + ], + "image_footnote": [], + "bbox": [ + 702, + 88, + 803, + 220 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5b94aed754c140f6a5fb3a8fc45189ebc7fb99c68f79021e2e30ffb9c7bd2da0.jpg", + "image_caption": [ + "CSUD (Ours)" + ], + "image_footnote": [], + "bbox": [ + 803, + 88, + 903, + 220 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2cc05fb52f79d70e476717d8cdc46fdeb784ecd1ce2bdb221f6327e6b63d08cf.jpg", + "image_caption": [ + "Figure 9. PSNR improvement of different baselines with CSUD. We test the models' generalization performance on 3 real-world datasets out of domain compared with their supervised version." + ], + "image_footnote": [], + "bbox": [ + 94, + 276, + 485, + 393 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/9b02bb7c24d93ddff5eacfe2b908bb21c2df8a7babf5b4678f259936a069423c.jpg", + "table_caption": [ + "Table 4. Effect of different training datasets on generalization performance. We test 4 different CSUD models which are trained on 4 different datasets on 3 real-world datasets out of domain." + ], + "table_footnote": [], + "table_body": "
Train datasetsSPA-data [36]Night-Rain [51]RainDS [30]
PSNRSSIMPSNRSSIMPSNRSSIM
Rain100L [43]33.570.93925.940.80522.540.613
Rain100H [43]29.610.88924.530.70221.850.600
RealRain1K-L [18]32.950.93728.140.83522.900.630
RealRain1K-H [18]29.250.89726.310.78722.300.564
", + "bbox": [ + 109, + 510, + 470, + 588 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "world datasets with different rain distributions, and test them on 3 real-world datasets. As shown in Tab. 4, in most cases, CSUD can achieve strong generalization performance, demonstrating the robustness of our approach. Due to the significant differences between rain streaks in Rain100H and real world, its generalization capability is the worst. Additionally, the generalization performance when training on Rain100L is comparable to that when training on real-world datasets, demonstrating our strong generalization ability.", + "bbox": [ + 88, + 613, + 482, + 748 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of the Weight Coefficients of Each Loss Function. We conduct extensive ablation experiments on the weight coefficients of perceptual loss in Eq. (5), $L_{CC}$ and $L_{SR - G}$ in Eq. (6). The results are shown in Tab. 5. Since $L_{SR - Der}$ and $L_{SR - G}$ are highly correlated, to ensure that the losses are on the same scale, the coefficient for $L_{SR - Der}$ is set to 0.1 times that of $L_{SR - G}$ , and $L_{SR}$ in Tab. 5 indicates $L_{SR - G}$ . Based on the experimental results, aside from our selected optimal combination, the deraining and generalization performance of our method remain at a high level", + "bbox": [ + 89, + 750, + 485, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "under most other weight combinations, demonstrating the robustness of our CSUD framework.", + "bbox": [ + 511, + 282, + 906, + 310 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "More detailed explanations of our method, experimental details and analysis can be found in the Appendix.", + "bbox": [ + 511, + 311, + 906, + 343 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/1cb299f40cde955e648dd468d023e8008f22ff2f9408b791c611c2dd9a17015a.jpg", + "table_caption": [ + "Table 5. Ablation experiments on the weight coefficients of each loss function. All models in the table are trained on Rain100L." + ], + "table_footnote": [], + "table_body": "
Lper(λ2)LSR(α2)LCC(α1)Rain100L [43]SPA-data [36]RealRain1K-L [18]
PSNRSSIMPSNRSSIMPSNRSSIM
0.25.00.130.690.93232.130.92928.210.915
0.25.01.033.150.95133.500.93829.750.928
0.21.01032.840.94633.490.93529.100.918
0.2101032.900.95133.400.93329.060.920
05.01032.180.93932.690.93628.240.916
1.05.01032.720.94733.050.93028.700.918
0.25.01033.280.95433.570.93929.210.928
", + "bbox": [ + 517, + 393, + 903, + 488 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 515, + 633, + 531 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we propose a novel unsupervised framework for image deraining to address the lack of paired data and the poor generalization in real-world scenarios. We first propose a channel consistency loss by introducing the channel consistency prior into the optimization process of the generator, ensuring that more background details are preserved while generating pseudo rainy images. In addition, a self-reconstruction strategy is proposed to alleviate the redundant information transfer problem and further improve the deraining performance and the generalization capability of our method. Extensive experiments on multiple synthetic and real-world datasets verify that our method achieves excellent deraining performance with strong generalization capability across various real-world rainy scenarios.", + "bbox": [ + 511, + 541, + 908, + 752 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement. This work was supported by the National Natural Science Foundation of China under Grant 62171304 and partly by the Natural Science Foundation of Sichuan Province under Grant 2024NSFSC1423, the TCL Science and Technology Innovation Fund under grant 25JZH008, the Young Faculty Technology Innovation Capacity Enhancement Program of Sichuan University under Grant 2024SCUQJTX025, and the Beijing Natural Science Foundation L221011.", + "bbox": [ + 511, + 763, + 910, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "7476", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Liangyu Chen, Xin Lu, Jie Zhang, Xiaojie Chu, and Chengpeng Chen. Hinet: Half instance normalization network for image restoration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 182-192, 2021. 2", + "[2] Liangyu Chen, Xiaojie Chu, Xiangyu Zhang, and Jian Sun. Simple baselines for image restoration. In European conference on computer vision, pages 17-33. Springer, 2022. 6, 7", + "[3] Xiang Chen, Jinshan Pan, Kui Jiang, Yufeng Li, Yufeng Huang, Caihua Kong, Longgang Dai, and Zhentao Fan. Unpaired deep image deraining using dual contrastive learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2017-2026, 2022. 1, 2, 3, 6, 7", + "[4] Xiang Chen, Hao Li, Mingqiang Li, and Jinshan Pan. Learning a sparse transformer network for effective image deraining. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5896-5905, 2023. 2", + "[5] Xiang Chen, Jinshan Pan, and Jiangxin Dong. Bidirectional multi-scale implicit neural representations for image deraining. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25627-25636, 2024. 2, 6, 7", + "[6] P Kingma Diederik. Adam: A method for stochastic optimization. (No Title), 2014. 6", + "[7] Xueyang Fu, Jiabin Huang, Xinghao Ding, Yinghao Liao, and John Paisley. Clearing the skies: A deep network architecture for single-image rain removal. IEEE Transactions on Image Processing, 26(6):2944-2956, 2017. 1, 2, 6, 7", + "[8] Xueyang Fu, Jiabin Huang, Delu Zeng, Yue Huang, Xinghao Ding, and John Paisley. Removing rain from single images via a deep detail network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3855-3863, 2017. 1, 2, 6", + "[9] Jinjin Gu, Xianzheng Ma, Xiangtao Kong, Yu Qiao, and Chao Dong. Networks are slacking off: Understanding generalization problem in image deraining. Advances in Neural Information Processing Systems, 36, 2023. 1, 2, 3", + "[10] Kaiming He, Jian Sun, and Xiaou Tang. Single image haze removal using dark channel prior. IEEE transactions on pattern analysis and machine intelligence, 33(12):2341-2353, 2010. 1", + "[11] Xiaowei Hu, Chi-Wing Fu, Lei Zhu, and Pheng-Ann Heng. Depth-attentional features for single-image rain removal. In Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition, pages 8022–8031, 2019. 1", + "[12] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1125-1134, 2017. 3", + "[13] Kui Jiang, Zhongyuan Wang, Peng Yi, Chen Chen, Baojin Huang, Yimin Luo, Jiayi Ma, and Junjun Jiang. Multi-scale progressive fusion network for single image deraining. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8346-8355, 2020. 1, 2" + ], + "bbox": [ + 93, + 114, + 485, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Xin Jin, Zhibo Chen, Jianxin Lin, Zhikai Chen, and Wei Zhou. Unsupervised single image deraining with self-supervised constraints. In 2019 IEEE International Conference on Image Processing (ICIP), pages 2761-2765. IEEE, 2019. 1, 2", + "[15] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14, pages 694-711. Springer, 2016. 5", + "[16] Boyun Li, Xiao Liu, Peng Hu, Zhongqin Wu, Jiancheng Lv, and Xi Peng. All-in-one image restoration for unknown corruption. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 17452-17462, 2022. 2", + "[17] Siyuan Li, Iago Bruno Araujo, Wenqi Ren, Zhangyang Wang, Eric K Tokuda, Roberto Hirata Junior, Roberto Cesar-Junior, Jiawan Zhang, Xiaojie Guo, and Xiaochun Cao. Single image deraining: A comprehensive benchmark analysis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3838-3847, 2019. 6", + "[18] Wei Li, Qiming Zhang, Jing Zhang, Zhen Huang, Xinmei Tian, and Dacheng Tao. Toward real-world single image deraining: A new benchmark and beyond. arXiv preprint arXiv:2206.05514, 2022. 4, 5, 6, 7, 8", + "[19] Xia Li, Jianlong Wu, Zhouchen Lin, Hong Liu, and Hongbin Zha. Recurrent squeeze-and-excitation context aggregation net for single image deraining. In Proceedings of the European conference on computer vision (ECCV), pages 254-269, 2018. 1", + "[20] Yu Li, Robby T Tan, Xiaojie Guo, Jiangbo Lu, and Michael S Brown. Rain streak removal using layer priors. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2736-2744, 2016. 4, 6, 7", + "[21] Di Lin, Xin Wang, Jia Shen, Renjie Zhang, Ruonan Liu, Miao-hui Wang, Wuyuan Xie, Qing Guo, and Ping Li. Generative status estimation and information decoupling for image rain removal. Advances in Neural Information Processing Systems, 35:4612-4625, 2022. 1, 2", + "[22] Xin Lin, Chao Ren, Xiao Liu, Jie Huang, and Yinjie Lei. Unsupervised image denoising in real-world scenarios via self-collaboration parallel generative adversarial branches. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12642-12652, 2023. 3", + "[23] Xin Lin, Jingtong Yue, Sixian Ding, Chao Ren, Chun-Le Guo, and Chongyi Li. Unlocking low-light-rainy image restoration by pairwise degradation feature vector guidance. arXiv preprint arXiv:2305.03997, 2023. 2", + "[24] Li Liu, Wanli Ouyang, Xiaogang Wang, Paul Fieguth, Jie Chen, Xinwang Liu, and Matti Pietikainen. Deep learning for generic object detection: A survey. International journal of computer vision, 128:261-318, 2020. 1", + "[25] Xiao Liu, Xiangyu Liao, Xiuya Shi, Linbo Qing, and Chao Ren. Efficient information modulation network for image super-resolution. In ECAI 2023, pages 1544-1551. IOS Press, 2023. 1", + "[26] Xiao Liu, Xiuya Shi, Lufei Chen, Linbo Qing, and Chao Ren. Efficient parallel multi-scale detail and semantic encoding network for lightweight semantic segmentation. In Proceedings" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "7477", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "of the 31st ACM International Conference on Multimedia, pages 2544-2552, 2023. 1", + "[27] Dongwon Park, Byung Hyun Lee, and Se Young Chun. All-in-one image restoration for unknown degradations using adaptive discriminative filters for specific degradations. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5815-5824. IEEE, 2023. 2", + "[28] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019. 6", + "[29] Vaishnav Potlapalli, Syed Waqas Zamir, Salman H Khan, and Fahad Shahbaz Khan. Prompt: Prompting for all-in-one image restoration. Advances in Neural Information Processing Systems, 36, 2023. 1, 2, 6, 7", + "[30] Ruijie Quan, Xin Yu, Yuanzhi Liang, and Yi Yang. Removing raindrops and rain streaks in one go. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9147-9156, 2021. 6, 7, 8", + "[31] Chao Ren, Yizhong Pan, and Jie Huang. Enhanced latent space blind model for real image denoising via alternative optimization. Advances in Neural Information Processing Systems, 35:38386-38399, 2022. 1", + "[32] Dongwei Ren, Wangmeng Zuo, Qinghua Hu, Pengfei Zhu, and Deyu Meng. Progressive image deraining networks: A better and simpler baseline. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3937-3946, 2019. 1", + "[33] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. IEEE transactions on pattern analysis and machine intelligence, 39(6):1137-1149, 2016. 1", + "[34] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 5", + "[35] Cong Wang, Jinshan Pan, Wei Wang, Jiangxin Dong, Mengzhu Wang, Yakun Ju, and Junyang Chen. Promptstorer: A prompting image restoration method with degradation perception. Advances in Neural Information Processing Systems, 36:8898-8912, 2023. 2", + "[36] Tianyu Wang, Xin Yang, Ke Xu, Shaozhe Chen, Qiang Zhang, and Rynson WH Lau. Spatial attentive single-image deraining with a high quality real rain dataset. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12270-12279, 2019. 1, 2, 4, 6, 7, 8", + "[37] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 6", + "[38] Zhendong Wang, Xiaodong Cun, Jianmin Bao, Wengang Zhou, Jianzhuang Liu, and Houqiang Li. Uformer: A general u-shaped transformer for image restoration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 17683-17693, 2022. 2", + "[39] Wei Wei, Deyu Meng, Qian Zhao, Zongben Xu, and Ying Wu. Semi-supervised transfer learning for image rain removal. In" + ], + "bbox": [ + 91, + 92, + 483, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3877-3886, 2019. 3, 6, 8", + "[40] Yanyan Wei, Zhao Zhang, Yang Wang, Mingliang Xu, Yi Yang, Shuicheng Yan, and Meng Wang. Deraincyclegan: Rain attentive cyclegan for single image deraining and rainmaking. IEEE Transactions on Image Processing, 30:4788-4801, 2021. 1, 2, 3, 5, 6, 7", + "[41] Zbigniew Wojna, Vittorio Ferrari, Sergio Guadarrama, Nathan Silberman, Liang-Chieh Chen, Alireza Fathi, and Jasper Uijlings. The devil is in the decoder: Classification, regression and gans. International Journal of Computer Vision, 127: 1694-1706, 2019. 1", + "[42] Wei Xu, Xinyuan Chen, Haoming Guo, Xiaolin Huang, and Wei Liu. Unsupervised image restoration with quality-task-perception loss. IEEE Transactions on Circuits and Systems for Video Technology, 32(9):5736-5747, 2022. 1, 2", + "[43] Wenhan Yang, Robby T Tan, Jiashi Feng, Zongming Guo, Shuicheng Yan, and Jiaying Liu. Joint rain detection and removal from a single image with contextualized deep networks. IEEE transactions on pattern analysis and machine intelligence, 42(6):1377-1393, 2019. 4, 5, 6, 7, 8", + "[44] Rajeev Yasarla, Vishwanath A Sindagi, and Vishal M Patel. Syn2real transfer learning for image deraining using gaussian processes. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2726-2736, 2020. 3", + "[45] Yuntong Ye, Yi Chang, Hanyu Zhou, and Luxin Yan. Closing the loop: Joint rain generation and removal via disentangled image translation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2053-2062, 2021. 1", + "[46] Yuntong Ye, Changfeng Yu, Yi Chang, Lin Zhu, Xi-Le Zhao, Luxin Yan, and Yonghong Tian. Unsupervised deraining: Where contrastive learning meets self-similarity. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5821-5830, 2022. 1, 2, 3, 6, 7", + "[47] Qiaosi Yi, Juncheng Li, Qinyan Dai, Faming Fang, Guixu Zhang, and Tieyong Zeng. Structure-preserving deraining with residue channel prior guidance. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4238-4247, 2021. 1, 2", + "[48] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, Ming-Hsuan Yang, and Ling Shao. Multi-stage progressive image restoration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14821-14831, 2021. 2, 6, 7", + "[49] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5728-5739, 2022. 1, 2, 6, 7", + "[50] Fan Zhang, Shaodi You, Yu Li, and Ying Fu. GTav-nightrain: Photometric realistic large-scale dataset for night-time rain streak removal. arXiv preprint arXiv:2210.04708, 2022. 6", + "[51] Fan Zhang, Shaodi You, Yu Li, and Ying Fu. Learning rain location prior for nighttime deraining. In Proceedings of" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "7478", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "the IEEE/CVF International Conference on Computer Vision, pages 13148-13157, 2023. 6, 7, 8", + "[52] He Zhang and Vishal M Patel. Density-aware single image de-raining using a multi-stream dense network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 695-704, 2018. 1, 2", + "[53] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In Proceedings of the IEEE international conference on computer vision, pages 2223-2232, 2017. 2, 3, 6" + ], + "bbox": [ + 91, + 90, + 483, + 246 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "7479", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Channel Consistency Prior and Self-Reconstruction Strategy Based Unsupervised Image Deraining/1cd88811-9c34-4363-ae8a-2a52dd5a816f_model.json b/2025/Channel Consistency Prior and Self-Reconstruction Strategy Based Unsupervised Image Deraining/1cd88811-9c34-4363-ae8a-2a52dd5a816f_model.json new file mode 100644 index 0000000000000000000000000000000000000000..46aa338a658d8a263a260e30a40a6b271535d0c4 --- /dev/null +++ b/2025/Channel Consistency Prior and Self-Reconstruction Strategy Based Unsupervised Image Deraining/1cd88811-9c34-4363-ae8a-2a52dd5a816f_model.json @@ -0,0 +1,2367 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.13, + 0.908, + 0.178 + ], + "angle": 0, + "content": "Channel Consistency Prior and Self-Reconstruction Strategy Based Unsupervised Image Deraining" + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.203, + 0.892, + 0.275 + ], + "angle": 0, + "content": "Guanglu Dong\\(^{1}\\) Tianheng Zheng\\(^{1}\\) Yuanzhouhan Cao\\(^{2}\\) Linbo Qing\\(^{1}\\) Chao Ren\\(^{1*}\\) \n\\(^{1}\\)Sichuan University, Chengdu, China \n\\(^{2}\\)Beijing Jiaotong University, Beijing, China \n{dongguanglu, zhengtianheng}@stu.scu.edu.cn, yzhcao@bjtu.edu.cn \n{qing.lb, chaoren}@scu.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.309, + 0.327, + 0.325 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.341, + 0.486, + 0.751 + ], + "angle": 0, + "content": "Recently, deep image deraining models based on paired datasets have made a series of remarkable progress. However, they cannot be well applied in real-world applications due to the difficulty of obtaining real paired datasets and the poor generalization performance. In this paper, we propose a novel Channel Consistency Prior and Self-Reconstruction Strategy Based Unsupervised Image Deraining framework, CSUD, to tackle the aforementioned challenges. During training with unpaired data, CSUD is capable of generating high-quality pseudo clean and rainy image pairs which are used to enhance the performance of deraining network. Specifically, to preserve more image background details while transferring rain streaks from rainy images to the unpaired clean images, we propose a novel Channel Consistency Loss (CCLoss) by introducing the Channel Consistency Prior (CCP) of rain streaks into training process, thereby ensuring that the generated pseudo rainy images closely resemble the real ones. Furthermore, we propose a novel Self-Reconstruction (SR) strategy to alleviate the redundant information transfer problem of the generator, further improving the deraining performance and the generalization capability of our method. Extensive experiments on multiple synthetic and real-world datasets demonstrate that the deraining performance of CSUD surpasses other state-of-the-art unsupervised methods and CSUD exhibits superior generalization capability. Code is available at https://github.com/GuangluDong0728/CSUD." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.778, + 0.222, + 0.793 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.803, + 0.485, + 0.879 + ], + "angle": 0, + "content": "Image deraining is an important task in low-level computer vision [7, 10, 25, 29, 31, 49]. It aims to reconstruct a clean image from its rainy version, which has a profound impact on subsequent high-level tasks such as object detection, recognition, and semantic segmentation" + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.308, + 0.908, + 0.544 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.547, + 0.909, + 0.618 + ], + "angle": 0, + "content": "Figure 1. Deraining results on the real rainy images captured by ourselves in real-world scenarios. Compared with the supervised method Restormer [49] and the unsupervised method DerainCycleGAN [40], our CSUD exhibits extremely strong generalization capability in the real world and achieves the best visual results." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.644, + 0.909, + 0.825 + ], + "angle": 0, + "content": "[24, 26, 33, 41]. To mitigate the negative impact of rain, numerous deep learning based supervised image deraining methods [7, 8, 11, 13, 19, 21, 32, 36, 47, 52] have been proposed and they achieve remarkable progress on synthetic datasets. However, the performance of these methods is largely constrained by the difficulty of capturing a large number of high-quality paired datasets when applied to real-world scenarios. Additionally, attempting to directly apply supervised models trained on synthetic datasets to real-world scenarios is also difficult because of the huge domain gap between synthetic and real-world rain [45] and the overfitting issue of supervised methods [9]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Compared to supervised methods, unsupervised deep learning methods [3, 14, 40, 42, 46] for image deraining tasks generally achieve end-to-end image deraining without paired clean-rainy images during training. This allows unsupervised methods to effectively handle complex real-world" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.235, + 0.9 + ], + "angle": 0, + "content": "*Corresponding author" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7469" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.212 + ], + "angle": 0, + "content": "scenarios, especially when data collection costs are high. However, due to the high difficulty and challenges of unsupervised deraining methods in training process and the lack of prior knowledge [40], there is still very little relevant research in this field. Additionally, the deraining performance of existing unsupervised methods always lags far behind that of supervised methods, and they face the same generalization issues as supervised methods." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.216, + 0.486, + 0.367 + ], + "angle": 0, + "content": "As shown in Fig. 1, even the SOTA supervised and unsupervised deraining models [40, 49] encounter substantial generalization issues and struggle to achieve satisfactory results when dealing with rainy images captured in real world, and sometimes they even fail to remove any rain streaks due to the more complex rain distributions in real world. Based on the above analysis, how to design an effective unsupervised deraining framework to avoid the need for real paired data and how to improve the generalization performance of deraining methods have become key research issues." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.37, + 0.485, + 0.595 + ], + "angle": 0, + "content": "In this paper, we propose a novel channel consistency prior and self-reconstruction strategy based unsupervised image deraining framework which only needs unpaired data to train the network and can be easily generalized to real-world scenarios. We first conduct detailed analysis of the channel consistency prior (CCP) of rain streaks, and then propose a novel channel consistency loss to constrain the generator to synthesize high-quality pseudo rainy images. In addition, we utilize the proposed self-reconstruction (SR) strategy to further improve the performance of the generator and enhance the generalization capability of the derainer. As shown in Figure 1, our CSUD achieves the most promising performance on the real-world captured images compared to both supervised and unsupervised deraining methods. The main contributions of this work are summarized below:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.599, + 0.485, + 0.704 + ], + "angle": 0, + "content": "- We propose a novel channel consistency prior and self-reconstruction strategy based unsupervised image deraining framework which is able to generate high-quality pseudo paired data for training. This framework can effectively address the issues of insufficient paired data in real world and significantly improve the generalization capability of the network." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.705, + 0.485, + 0.81 + ], + "angle": 0, + "content": "- We propose a novel channel consistency loss (CCLoss) according to the channel consistency prior (CCP) of rain streaks to enhance the performance of the generator, preserving more background details while generating pseudo rainy images. To the best of our knowledge, we are the first to incorporate CCP into the training process of unsupervised deraining framework." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.811, + 0.485, + 0.901 + ], + "angle": 0, + "content": "- We propose a self-reconstruction (SR) strategy to solve the redundant information transfer problem of the generator. SR strategy can facilitate the accuracy of rain streaks transfer in the generator and enhance the deraining performance and the generalization capability of the derainer without altering the existing network structure." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.599, + 0.485, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.092, + 0.909, + 0.168 + ], + "angle": 0, + "content": "- Extensive experiments on numerous benchmark datasets demonstrate that our proposed unsupervised framework can effectively tackle the deraining tasks on both synthetic and real-world rainy images with extremely superior generalization capability in real-world scenarios." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.183, + 0.665, + 0.199 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.209, + 0.9, + 0.226 + ], + "angle": 0, + "content": "2.1. Deep Learning Based Single Image Deraining" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.231, + 0.908, + 0.488 + ], + "angle": 0, + "content": "Supervised Single Image Deraining. Recently, with the development of deep learning, numerous excellent single image deraining [4, 5, 7, 8, 13, 21, 36, 47, 52] models have emerged. Fu et al. [7] first employ the CNN-based DerainNet to extract and remove the rain layer. Lin et al. [21] propose a generative network equipped with the pixel-wise status estimation and information decoupling for deraining. Lin et al. [23] propose a unified network for low-light-rainy image restoration to jointly handle low-light enhancement and deraining tasks. In addition to the aforementioned works specifically targeting deraining, some networks designed for image restoration [1, 16, 27, 29, 35, 38, 48, 49] have also shown excellent performance in deraining tasks. However, the above fully supervised methods require a large number of high-quality paired training samples, which is almost impossible in real-world scenarios. Simultaneously, due to the overfitting issue, their generalization capability is limited." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.489, + 0.909, + 0.761 + ], + "angle": 0, + "content": "Unsupervised Single Image Deraining. To address the lack of paired datasets in the real world, some outstanding unsupervised deraining methods [3, 14, 40, 42, 46] have been proposed. Wei et al. [40] propose an unsupervised method for rain removal and generation by utilizing the popular CycleGAN [53] framework. Then, through introducing a dual contrastive learning manner to the CycleGAN framework, Chen et al. [3] develop an unpaired adversarial framework which effectively explores mutual properties of the unpaired rainy-clean samples. Ye et al. [46] introduce a novel non-local contrastive learning based unsupervised image deraining method to better distinguish rain streaks from clean images. However, the deraining performance of these unsupervised methods is far inferior to that of supervised methods, especially on real-world scenarios. Compared with previous methods, our proposed CSUD can greatly enhance the unsupervised deraining performance and achieve competitive results compared with some supervised methods." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.773, + 0.887, + 0.789 + ], + "angle": 0, + "content": "2.2. Generalization Problem in Image Deraining" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Deep image deraining models often fail to obtain satisfactory performance or even cannot remove any rain streaks when they are applied to real-world scenarios. Generalization problem of deraining models is an urgent issue to be solved. Gu et al. [9] conduct thorough analysis of the generalization problem in image deraining and propose to achieve better generalization by simplifying the complexity of train-" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7470" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.093, + 0.09, + 0.907, + 0.38 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.383, + 0.907, + 0.425 + ], + "angle": 0, + "content": "Figure 2. Framework of the proposed CSUD. The red arrows represent the baseline, purple arrows represent the 3 extra constraints, green arrows represent the self-reconstruction (SR) strategy, and blue arrows represent the CCLoss. The 4 input arrows of RIEM represent 4 separate runs of the generator, each time guided by a different image. Both RIEM and CFEM receive only 1 input image at a time." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.451, + 0.486, + 0.634 + ], + "angle": 0, + "content": "ing images. Some semi-supervised learning frameworks [39, 44] are proposed to analyze the residual difference between domains and enhance the deraining capability in real world. In addition, a series of unsupervised deraining methods [3, 40, 46] have also demonstrated a certain degree of generalization capability in real-world scenarios. However, they have not conducted detailed analysis on how to enhance the generalization capability of unsupervised methods. Inspired by [9], we propose a self-reconstruction strategy to make both generator and derainer pay more attention to the reconstruction of image background, which significantly improves the generalization performance of our method." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.652, + 0.226, + 0.67 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.679, + 0.441, + 0.695 + ], + "angle": 0, + "content": "3.1. Overall Framework of Proposed Method" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.703, + 0.486, + 0.854 + ], + "angle": 0, + "content": "To address the issue of insufficient paired data and the generalization problem in deraining applications, we propose a novel channel consistency prior and self-reconstruction strategy based unsupervised image deraining framework, CSUD. As illustrated in Fig. 2, our CSUD consists of a derainer \\( Der \\), a rainy image generator \\( G \\), and a Patch-GAN [12] discriminator \\( D \\). For \\( G \\), it consists of a clean feature extraction module (CFEM), a rain information extraction module (RIEM) and 6 residual blocks. Detailed network structures are detailed in the Appendix." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.485, + 0.902 + ], + "angle": 0, + "content": "In generator \\(G\\), we effectively utilize unpaired clean and rainy images to synthesize pseudo rainy images. As shown by the red arrows in Fig. 2, we design a GAN-based baseline" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.451, + 0.907, + 0.618 + ], + "angle": 0, + "content": "that transitions from \"\\(x \\in X \\to x_{s1} \\in S \\to x_{der} \\in N\\)\", where \\(X\\), \\(S\\) and \\(N\\) denote clean images, synthetic rainy images and derained images respectively. \\(G\\) transforms clean image \\(x\\) to synthetic rainy image \\(x_{s1}\\) according to the guide of real rainy image \\(y\\). In order to better extract rain information and transfer it to the clean images, \\(y\\) and \\(x\\) are initially processed by the RIEM and the CFEM respectively to obtain rainy features and clean features which are concatenated for subsequent process. For the real rainy image \\(y\\) and the synthesized rainy image \\(x_{s1}\\), the corresponding derained clean images \\(y_{der}\\) and \\(x_{der}\\) are obtained through Der." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.618, + 0.909, + 0.753 + ], + "angle": 0, + "content": "The discriminator \\(D\\) is trained to distinguish between synthetic rainy images \\(x_{s1} \\in S\\) and real rainy images \\(y \\in Y\\), and the adversarial loss \\(L_{adv1}\\) of the baseline is constrained between them. Furthermore, motivated by existing methods [22, 53], as indicated by the purple arrows in Fig. 2, we add another three processes, \\(G(Der(x_{s1}), x_{s1}) \\to x_{s2}\\), \\(G(Der(y), y) \\to y_{s1}\\), and \\(G(Der(y), x_{s1}) \\to y_{s2}\\), to the framework to enhance training stability. Then, additional three adversarial losses are introduced to CSUD:" + }, + { + "type": "equation", + "bbox": [ + 0.563, + 0.763, + 0.907, + 0.78 + ], + "angle": 0, + "content": "\\[\nL _ {G A N} = L _ {a d v 1} + L _ {a d v 2} + L _ {a d v 3} + L _ {a d v 4}. \\tag {1}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.789, + 0.863, + 0.804 + ], + "angle": 0, + "content": "3.2. Channel Consistency Loss for Generator" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Channel Consistency Prior (CCP) of Rain Streaks. In RGB rainy images, most rain streaks tend to be consistent across the R, G, and B channels, meaning the pixel values of most rain streaks in the three channels are highly similar. As shown in the second, third, and fourth columns of Fig. 3 (a), we separate both synthetic and real-world rainy images" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "7471" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.086, + 0.905, + 0.361 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.194, + 0.366, + 0.807, + 0.38 + ], + "angle": 0, + "content": "(a) Visualization of channel consistency prior in synthetic, real-world, and nighttime real-world rainy images" + }, + { + "type": "image", + "bbox": [ + 0.092, + 0.383, + 0.903, + 0.476 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.202, + 0.48, + 0.797, + 0.493 + ], + "angle": 0, + "content": "(b) Cosine similarities of cycle subtractions among R, G, and B channels between clean and rainy images" + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.496, + 0.907, + 0.553 + ], + "angle": 0, + "content": "Figure 3. (a) Visualization of channel consistency prior in synthetic and real-world rainy images. The first column presents the rainy RGB images; the second, third and fourth columns present their R, G, B channels, respectively; the fifth, sixth, and seventh columns present the cycle subtractions of R, G, and B channels, respectively. (b) Cosine similarities of the cycle subtractions of R, G, and B channels between clean and rainy images in the 5 synthetic and real-world datasets. More visualizations and analysis are presented in the Appendix." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.579, + 0.483, + 0.655 + ], + "angle": 0, + "content": "into R, G, and B channels and visualized them. It can be seen clearly that while there are significant differences in the background parts among the three channels, the rain streaks tend to be consistent. We propose to term this phenomenon as \"Channel Consistency Prior (CCP)\" of rainy streaks." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.66, + 0.486, + 0.902 + ], + "angle": 0, + "content": "Theoretically, according to CCP, the cycle subtractions \\(((r)_R - (r)_G)\\), \\(((r)_G - (r)_B)\\), \\(((r)_B - (r)_R)\\) of R, G, and B channels in rainy images will almost contain no rain streaks and should be close to \\(((n)_R - (n)_G)\\), \\(((n)_G - (n)_B)\\), \\(((n)_B - (n)_R)\\) in clean images, where \\(r\\) and \\(n\\) represent the rainy and clean image, \\((\\cdot)_R\\), \\((\\cdot)_G\\), and \\((\\cdot)_B\\) denote R, G, and B channels respectively. In the fifth, sixth, and seventh columns of Fig. 3 (a), it can be observed that the cycle subtractions of R, G, and B channels nearly don't contain any rain streaks, qualitatively confirming the correctness of CCP. Furthermore, as shown in Fig. 3 (b), we quantitatively present the cosine similarities of the cycle subtractions of R, G, and B channels between clean and rainy images in both synthetic and real-world benchmark datasets [18, 20, 36, 43]. It is evident that the cosine similarities are exceptionally high, which can further demonstrate our hypothesis." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.579, + 0.907, + 0.685 + ], + "angle": 0, + "content": "Channel Consistency Loss. Based on the above analysis of CCP, the cycle subtractions of R, G, and B channels between input clean image and output pseudo rainy image of the generator \\(G\\) should tend to be consistent. To this end, we propose a novel Channel Consistency Loss (CCLoss) to preserve more color and texture details of the background while generating pseudo rainy images:" + }, + { + "type": "equation", + "bbox": [ + 0.553, + 0.69, + 0.907, + 0.745 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} L _ {C C} = \\left\\| \\left(\\left(x _ {s 1}\\right) _ {R} - \\left(x _ {s 1}\\right) _ {G}\\right), \\left(x _ {R} - x _ {G}\\right) \\right\\| _ {1} \\\\ + \\left\\| \\left(\\left(x _ {s 1}\\right) _ {G} - \\left(x _ {s 1}\\right) _ {B}\\right), \\left(x _ {G} - x _ {B}\\right) \\right\\| _ {1} \\tag {2} \\\\ + \\left\\| \\left(\\left(x _ {s 1}\\right) _ {B} - \\left(x _ {s 1}\\right) _ {R}\\right), \\left(x _ {B} - x _ {R}\\right) \\right\\| _ {1}, \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.751, + 0.909, + 0.843 + ], + "angle": 0, + "content": "where \\((x_{s1})_R, (x_{s1})_G, (x_{s1})_B, x_R, x_G, x_B\\) denotes R, G, B color channels of \\(x_{s1}\\) and \\(x, ||\\cdot||_1\\) represents the \\(L_1\\) norm and \\(L_{CC}\\) represents the proposed CCLoss, respectively. By introducing CCP into our unsupervised deraining framework, \\(G\\) will generate more accurate synthesized rainy images, which will further improve the performance of the derainer." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.849, + 0.771, + 0.866 + ], + "angle": 0, + "content": "3.3. Self-Reconstruction Strategy" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Redundant Information Transfer Problem of the Generator. Although adversarial losses can effectively constrain the" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7472" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.088, + 0.487, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.091, + 0.24, + 0.482, + 0.254 + ], + "angle": 0, + "content": "Figure 4. Visualization of redundant information transfer problem." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.279, + 0.484, + 0.414 + ], + "angle": 0, + "content": "training process of generator \\( G \\), the data synthesis process of \\( G \\) is highly susceptible to interference from the redundant information in rainy images, especially those edge textures that are very similar to rain streaks. As shown in Fig. 4, the edge textures of some background objects in rainy images, such as the outline of the woman's hair, the leg of rhino, etc., are mistakenly transferred to the clean images by \\( G \\). As a result, \\( G \\) will generate inaccurate pseudo rainy images, interfering with the learning process of \\( Der \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.415, + 0.484, + 0.673 + ], + "angle": 0, + "content": "Self-Reconstruction Strategy. To solve the redundant information transfer problem, we propose a novel Self-Reconstruction (SR) strategy, which enables \\( G \\) to extract more accurate rain streak information and synthesizes more realistic pseudo rainy images. The core idea of SR strategy is to encourage \\( G \\) to better distinguish rain streaks from background in the rainy images, ensuring that only rain streak information is extracted for subsequent generation of pseudo rainy images, while redundant information of background is disregarded. As shown by the green arrows in Fig. 2, when the input clean image \\( x \\) itself, the derained image \\( y_{der} \\) are input with clean image \\( x \\) into \\( G \\), \\( G \\) learns to reconstruct the input clean image \\( x \\) itself without any additional redundant information from \\( x \\) and \\( y_{der} \\), which means that the self-reconstructed images \\( x_{s1} \\) and \\( x_{s2} \\) should be close to \\( x \\). Thus, we can utilize the following self-reconstruction loss (SRLoss) as additional training constraint of \\( G \\):" + }, + { + "type": "equation", + "bbox": [ + 0.132, + 0.682, + 0.484, + 0.7 + ], + "angle": 0, + "content": "\\[\nL _ {S R - G} = \\left\\| G (x, x), x \\right\\| _ {1} + \\left\\| G (x, y _ {d e r}), x \\right\\| _ {1}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.708, + 0.484, + 0.875 + ], + "angle": 0, + "content": "where \\( L_{SR - G} \\) represents the SRLoss for \\( G \\). SR strategy allows \\( G \\) to pay more attention to reconstructing the background of clean images while learning the transfer of rain streaks, significantly improving the accuracy of generating pseudo rainy images. Under the constraint of \\( L_{SR - G} \\), \\( G \\) will obtain the ability to generate clean images themselves guided by images without rain streaks. Therefore, the cleaner the derained images \\( y_{der} \\) and \\( x_{der} \\) restored by \\( Der \\) are, the closer the self-reconstructed images \\( x_{s2} \\) and \\( x_{s3} \\) are to the original clean ones \\( x \\). Then, the following SRLoss can be utilized to further improve the performance of \\( Der \\):" + }, + { + "type": "equation", + "bbox": [ + 0.106, + 0.885, + 0.484, + 0.903 + ], + "angle": 0, + "content": "\\[\nL _ {S R - D e r} = \\left\\| G (x, x _ {d e r}), x \\right\\| _ {1} + \\left\\| G (x, y _ {d e r}), x \\right\\| _ {1}, \\tag {4}\n\\]" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.089, + 0.906, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.542, + 0.25, + 0.875, + 0.265 + ], + "angle": 0, + "content": "Figure 5. The implementation principles of SR strategy." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.292, + 0.907, + 0.352 + ], + "angle": 0, + "content": "where \\( L_{SR - Der} \\) represents the SRLoss for \\( Der \\). Under the constraint of \\( L_{SR - Der} \\), \\( Der \\) will also focus more on reconstructing the background content while learning to derain, which greatly enhances the generalization capability of \\( Der \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.353, + 0.907, + 0.43 + ], + "angle": 0, + "content": "As shown in the third and fourth columns of Figure 4, SR strategy effectively alleviates the redundant information transfer problem. Through SR strategy, \\( G \\) can achieve robust rain streak transfer with more background detail fidelity, greatly enhancing the deraining performance of Der." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.44, + 0.659, + 0.455 + ], + "angle": 0, + "content": "3.4. Loss Function" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.462, + 0.908, + 0.524 + ], + "angle": 0, + "content": "Due to the high difficulty and instability of training a GAN-based unsupervised framework, in addition to adopting the L1 loss for \\( Der \\), we also incorporate SSIM loss and perceptual loss[15] to enhance the performance of \\( Der \\):" + }, + { + "type": "equation", + "bbox": [ + 0.553, + 0.536, + 0.906, + 0.572 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} L _ {D e r} = \\left\\| x, x _ {d e r} \\right\\| _ {1} + \\lambda_ {1} L _ {S S I M} (x, x _ {d e r}) \\tag {5} \\\\ + \\lambda_ {2} | | \\phi_ {i} (x), \\phi_ {i} (x _ {d e r}) | | _ {1} + \\lambda_ {3} L _ {S R - D e r}, \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.584, + 0.907, + 0.645 + ], + "angle": 0, + "content": "where \\(||\\cdot ||_1\\) represents the \\(L_{1}\\) norm, and \\(L_{SSIM}(\\cdot)\\) represents the SSIM loss, \\(\\phi_i(\\cdot)\\) is the layers of pre-trained VGG-19 [34], \\(\\lambda_1, \\lambda_2, \\lambda_3\\) represent the hyperparameters for SSIM loss, perceptual loss and SRLoss for derainer respectively." + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.646, + 0.897, + 0.661 + ], + "angle": 0, + "content": "Finally, we obtain the total loss to train our framework:" + }, + { + "type": "equation", + "bbox": [ + 0.514, + 0.674, + 0.908, + 0.707 + ], + "angle": 0, + "content": "\\[\nL _ {t o t a l} = \\min _ {G} \\max _ {D} L _ {G A N} + L _ {D e r} + \\alpha_ {1} L _ {C C} + \\alpha_ {2} L _ {S R - G}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.708, + 0.906, + 0.739 + ], + "angle": 0, + "content": "where \\(\\alpha_{1},\\alpha_{2}\\) represent the hyperparameters for CCLoss and SRLoss for generator respectively." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.754, + 0.645, + 0.771 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.909, + 0.901 + ], + "angle": 0, + "content": "To demonstrate the effectiveness of our CSUD, following [18, 40], we conduct experiments from the following two aspects: (1) unsupervised deraining performance, (2) generalization performance. In experiment (1), we train independent models for different datasets respectively, including both synthetic and real-world datasets. In experiment (2), we only train our model on the synthetic dataset Rain100L [43], and then test on various real-world and nighttime datasets." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7473" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.09, + 0.908, + 0.119 + ], + "angle": 0, + "content": "Table 1. Quantitative deraining performance comparisons of different methods. Red and blue colors indicate the 1st and 2nd ranks among the unsupervised methods, respectively. S and U indicate supervised method and unsupervised method, respectively." + }, + { + "type": "table", + "bbox": [ + 0.137, + 0.129, + 0.865, + 0.324 + ], + "angle": 0, + "content": "
DatasetsRain100L [43]Rain100H [43]Rain12 [20]Rain800 [17]RealRain1K-L [18]RealRain1K-H [18]
MetricsPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIM
DerainNet [7] (S)27.030.88414.920.592//22.770.81027.090.92522.880.889
DDN [8] (S)32.380.92624.640.84934.040.93321.160.73231.180.91729.170.878
SPA-Net [36] (S)31.950.91926.070.857//24.370.86130.430.94725.760.910
MPRNet [48] (S)34.950.95928.530.87236.840.96428.680.87536.290.97234.740.964
NAFNet [2] (S)37.000.97829.660.90034.810.94328.720.88638.800.98636.110.976
Restormer [49] (S)37.570.97429.460.889//30.330.90540.900.98539.570.981
PromptIR [29] (S)38.340.98328.690.87735.090.94529.070.88736.990.97333.610.953
NeRD-Rain-S [5] (S)42.000.99032.860.93235.390.94228.820.87838.640.97936.690.970
CycleGAN [53] (U)24.610.83420.590.70421.560.84523.950.81920.190.82017.530.762
DerainCycleGAN [40] (U)31.490.93621.130.71033.520.94024.320.84228.160.90124.780.844
NLCL [46] (U)20.500.71922.310.72822.680.73518.940.67023.060.83220.640.719
DCD-GAN [3] (U)31.820.94122.470.75331.560.92425.610.81330.490.93927.820.895
CSUD (Ours)33.280.95424.420.80834.560.95123.580.79532.710.95929.670.928
" + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.334, + 0.195, + 0.436 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.129, + 0.438, + 0.16, + 0.449 + ], + "angle": 0, + "content": "Inupt" + }, + { + "type": "image", + "bbox": [ + 0.196, + 0.335, + 0.295, + 0.436 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.236, + 0.438, + 0.255, + 0.448 + ], + "angle": 0, + "content": "GT" + }, + { + "type": "image", + "bbox": [ + 0.296, + 0.335, + 0.397, + 0.436 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.324, + 0.438, + 0.371, + 0.448 + ], + "angle": 0, + "content": "MPRNet" + }, + { + "type": "image", + "bbox": [ + 0.398, + 0.335, + 0.497, + 0.436 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.412, + 0.438, + 0.488, + 0.448 + ], + "angle": 0, + "content": "NeRD-Rain-S" + }, + { + "type": "image", + "bbox": [ + 0.498, + 0.335, + 0.599, + 0.436 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.501, + 0.438, + 0.597, + 0.449 + ], + "angle": 0, + "content": "DerainCycleGAN" + }, + { + "type": "image", + "bbox": [ + 0.6, + 0.335, + 0.701, + 0.436 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.633, + 0.438, + 0.669, + 0.448 + ], + "angle": 0, + "content": "NLCL" + }, + { + "type": "image", + "bbox": [ + 0.702, + 0.335, + 0.803, + 0.436 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.724, + 0.438, + 0.784, + 0.448 + ], + "angle": 0, + "content": "DCD-GAN" + }, + { + "type": "image", + "bbox": [ + 0.804, + 0.335, + 0.904, + 0.436 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.815, + 0.438, + 0.894, + 0.449 + ], + "angle": 0, + "content": "CSUD (Ours)" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.455, + 0.908, + 0.484 + ], + "angle": 0, + "content": "Figure 6. Qualitative deraining performance comparisons on Rain100L [43] and Rain12 [20] datasets. Our CSUD achieves competitive visual results comparable to the SOTA supervised method NeRD-Rain-S [5]." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.51, + 0.484, + 0.675 + ], + "angle": 0, + "content": "Datasets and Metrics. We utilize 10 challenging benchmark datasets in our study, including synthetic datasets: Rain100L [43], Rain100H [43], Rain12 [20], Rain800 [17], real-world datasets: RealRain1K-L [18], RealRain1K-H [18], SPA-data [36], RainDS [30], Internet-Data [39], and a nighttime dataset which is based on GTAV-NightRain dataset [50] and enlarged the rain density by [51], we name this nighttime dataset as Night-Rain in the following experiments. As for RainDS, we only use the real rain streaks subset for testing. Following existing methods [3, 40, 49], we use PSNR and SSIM [37] to evaluate different methods." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.676, + 0.484, + 0.827 + ], + "angle": 0, + "content": "Implementation Details. As for derainer, we employ the CNN-based simple image restoration baseline NAFNet [2]. Our framework is implemented by PyTorch [28] with a GeForce RTX 3090 GPU. We adopt the Adam optimizer [6] to train our network for 200 epochs with the initial learning rate of \\(1e^{-4}\\), followed by another 100 epochs with a learning rate of \\(1e^{-5}\\). All training images are randomly cropped to \\(256 \\times 256\\) patches in an unpaired manner. The hyperparameters \\(\\lambda_{1}, \\lambda_{2}\\), and \\(\\lambda_{3}\\) in Eq. 5 are set to 1, 0.2 and 0.5, while \\(\\alpha_{1}\\) and \\(\\alpha_{2}\\) in Eq. 6 are set to 10 and 5, respectively." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.834, + 0.376, + 0.851 + ], + "angle": 0, + "content": "4.1. Unsupervised Deraining Results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "We compare our method with unsupervised methods from the most recent years, including [3, 40, 46, 53]. Noting that the related methods for comparison are limited, since only a few" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.51, + 0.909, + 0.707 + ], + "angle": 0, + "content": "unsupervised deraining models have been proposed in the field. We also list some SOTA supervised methods, including [2, 5, 7, 8, 29, 36, 48, 49]. Tab. 1 presents the quantitative deraining results of different methods. It is evident that our CSUD significantly outperforms other unsupervised methods on most datasets. Compared to DCD-GAN[3], our CSUD respectively achieves 1.46dB, 3.00dB, and 2.22dB improvement in PSNR on Rain100L, Rain12 and RealRain1K-L datasets. Additionally, our CSUD even achieves competitive results comparable to several supervised methods. As shown in Fig. 6, compared to other unsupervised methods, our CSUD achieves better deraining results while preserving more color and texture details of image background." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.713, + 0.725, + 0.728 + ], + "angle": 0, + "content": "4.2. Generalization Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.909, + 0.901 + ], + "angle": 0, + "content": "To validate the generalization capability of CSUD, we compare the results on 5 real-world datasets[18, 30, 36, 39] and a nighttime dataset [51] with SOTA supervised and unsupervised methods, including [2, 3, 5, 7, 29, 40, 46, 48, 49]. Noting that there are no ground truth images in the InternetData [39] dataset, thereby only visual results are presented. All methods are trained on synthetic datasets and tested on the unseen real-world and nighttime datasets. A prevailing perspective is that supervised methods usually outperform unsupervised methods. But as shown in Tab. 2, our CSUD almost comprehensively surpasses supervised models on" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7474" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.908, + 0.119 + ], + "angle": 0, + "content": "Table 2. Quantitative generalization performance comparisons of different deraining methods. Red and blue colors indicate the 1st and 2nd ranks among the unsupervised methods. S, U indicate supervised and unsupervised methods, respectively." + }, + { + "type": "table", + "bbox": [ + 0.175, + 0.129, + 0.825, + 0.281 + ], + "angle": 0, + "content": "
DatasetsRealRain1K-L [18]RealRain1K-H [18]SPA-data [36]RainDS [30]Night-Rain [51]
MetricsPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIM
DerainNet [7] (S)25.400.84721.910.74230.780.91221.350.591//
MPRNet [48] (S)26.750.89923.580.84033.000.94223.240.65626.130.817
NAFNet [2] (S)28.110.91524.360.84833.460.94122.480.62526.150.816
Restormer [49] (S)26.290.88323.130.85132.620.93622.960.64825.850.808
PromptIR [29] (S)28.140.91524.280.85133.500.94022.480.62526.370.815
NeRD-Rain-S [5] (S)27.640.90423.960.83133.030.93322.370.62226.380.810
DerainCycleGAN [40] (U)28.250.89924.200.82133.530.93718.750.46025.850.802
NLCL [46] (U)20.180.67419.810.62918.760.72717.000.47722.010.475
DCD-GAN [3] (U)27.490.91223.860.84531.550.93022.130.61025.540.798
CSUD (Ours)29.210.92825.450.87833.570.93922.540.61325.940.805
" + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.291, + 0.907, + 0.452 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.456, + 0.908, + 0.484 + ], + "angle": 0, + "content": "Figure 7. Qualitative generalization results on RealRain1K-L [18] and RealRain1K-H [18] real-world datasets. Our CSUD achieves best visual results among both SOTA supervised and unsupervised methods in real-world scenes." + }, + { + "type": "table_caption", + "bbox": [ + 0.127, + 0.509, + 0.446, + 0.523 + ], + "angle": 0, + "content": "Table 3. Effect of CCLoss and SR strategy on CSUD." + }, + { + "type": "table", + "bbox": [ + 0.102, + 0.535, + 0.475, + 0.62 + ], + "angle": 0, + "content": "
CCSRRain100L[43]Rain12[20]RealRain1K-L[18]
PSNRSSIMPSNRSSIMPSNRSSIM
XX30.660.92531.600.93130.420.932
✓X31.240.93432.620.94831.140.943
X✓32.690.94933.430.93931.860.950
✓✓33.280.95434.560.95132.710.959
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.644, + 0.483, + 0.688 + ], + "angle": 0, + "content": "the first 4 real-world datasets, while remaining competitive on the nighttime dataset. This strongly demonstrates the excellent generalization ability of our method." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.484, + 0.825 + ], + "angle": 0, + "content": "As shown in Fig. 7 and Fig. 8, the performance of some most advanced supervised methods like Restormer [49] and NeRD-Rain-S [5] is extremely poor when directly applied to real-world scenarios. In contrast, our CSUD achieves better visual results in real world. It is worth mentioning again that, as shown in Fig. 1 in Sec. 1, our method also shows extremely superior deraining performance on the rainy images captured by ourselves in real scenarios, which further demonstrates our excellent generalization capability." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.834, + 0.256, + 0.849 + ], + "angle": 0, + "content": "4.3. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.486, + 0.901 + ], + "angle": 0, + "content": "Effect of CCLoss and SR Strategy on Deraining Performance. To verify the effectiveness of CCLoss and SR strategy, we conduct ablation experiments on Rain100L [43]," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.511, + 0.907, + 0.663 + ], + "angle": 0, + "content": "Rain12 [20] and RealRain1K-L [18] datasets. As shown in Table 3, compared to the baseline, the effect on PSNR is improved by 1.38dB, 1.02dB, and 0.72dB among the above three datasets with CCLoss, respectively. And the effect is improved by 1.38dB, 1.83dB, and 1.44dB with SR strategy, respectively. Furthermore, when both CCLoss and SR strategy are introduced into the baseline, the PSNR on the three datasets is increased by 2.62dB, 2.96dB, and 2.29dB respectively. The above ablation results strongly verify the superiority and importance of CCLoss and SR strategy." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.668, + 0.909, + 0.835 + ], + "angle": 0, + "content": "Effect of CSUD framework on Generalization Performance. To validate the effectiveness of our CSUD in improving generalization performance and the universality and robustness of our method, we also conduct experiments incorporating CSUD framework on another 2 different baselines: NeRD-Rain-S [5] and Restormer [49]. Fig. 9 shows the PSNR improvements of different CSUD based models compared to their supervised version. Notably, we train all the above models on Rain100L and test them on the 3 real-world datasets. This strongly demonstrates the effectiveness of CSUD in improving generalization performance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.91, + 0.901 + ], + "angle": 0, + "content": "Effect of Different Training Datasets on Generalization Performance. To further validate the generalization performance of CSUD, we employ 4 different CSUD models which are respectively trained on 4 synthetic and real-" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7475" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.089, + 0.196, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.131, + 0.226, + 0.159, + 0.236 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image", + "bbox": [ + 0.196, + 0.089, + 0.296, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.219, + 0.225, + 0.272, + 0.235 + ], + "angle": 0, + "content": "Restormer" + }, + { + "type": "image", + "bbox": [ + 0.296, + 0.089, + 0.397, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.321, + 0.225, + 0.371, + 0.236 + ], + "angle": 0, + "content": "PromptIR" + }, + { + "type": "image", + "bbox": [ + 0.398, + 0.089, + 0.498, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.413, + 0.225, + 0.484, + 0.236 + ], + "angle": 0, + "content": "NeRD-Rain-S" + }, + { + "type": "image", + "bbox": [ + 0.498, + 0.089, + 0.6, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.506, + 0.225, + 0.596, + 0.236 + ], + "angle": 0, + "content": "DerainCycleGAN" + }, + { + "type": "image", + "bbox": [ + 0.601, + 0.089, + 0.702, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.634, + 0.225, + 0.668, + 0.235 + ], + "angle": 0, + "content": "NLCL" + }, + { + "type": "image", + "bbox": [ + 0.703, + 0.089, + 0.804, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.724, + 0.225, + 0.782, + 0.235 + ], + "angle": 0, + "content": "DCD-GAN" + }, + { + "type": "image", + "bbox": [ + 0.805, + 0.089, + 0.905, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.819, + 0.225, + 0.888, + 0.236 + ], + "angle": 0, + "content": "CSUD (Ours)" + }, + { + "type": "image_caption", + "bbox": [ + 0.25, + 0.242, + 0.746, + 0.255 + ], + "angle": 0, + "content": "Figure 8. Qualitative generalization results on Internet-Data [39] real-world dataset." + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.277, + 0.486, + 0.395 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.4, + 0.484, + 0.443 + ], + "angle": 0, + "content": "Figure 9. PSNR improvement of different baselines with CSUD. We test the models' generalization performance on 3 real-world datasets out of domain compared with their supervised version." + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.458, + 0.484, + 0.501 + ], + "angle": 0, + "content": "Table 4. Effect of different training datasets on generalization performance. We test 4 different CSUD models which are trained on 4 different datasets on 3 real-world datasets out of domain." + }, + { + "type": "table", + "bbox": [ + 0.11, + 0.511, + 0.472, + 0.589 + ], + "angle": 0, + "content": "
Train datasetsSPA-data [36]Night-Rain [51]RainDS [30]
PSNRSSIMPSNRSSIMPSNRSSIM
Rain100L [43]33.570.93925.940.80522.540.613
Rain100H [43]29.610.88924.530.70221.850.600
RealRain1K-L [18]32.950.93728.140.83522.900.630
RealRain1K-H [18]29.250.89726.310.78722.300.564
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.614, + 0.483, + 0.749 + ], + "angle": 0, + "content": "world datasets with different rain distributions, and test them on 3 real-world datasets. As shown in Tab. 4, in most cases, CSUD can achieve strong generalization performance, demonstrating the robustness of our approach. Due to the significant differences between rain streaks in Rain100H and real world, its generalization capability is the worst. Additionally, the generalization performance when training on Rain100L is comparable to that when training on real-world datasets, demonstrating our strong generalization ability." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.486, + 0.902 + ], + "angle": 0, + "content": "Effect of the Weight Coefficients of Each Loss Function. We conduct extensive ablation experiments on the weight coefficients of perceptual loss in Eq. (5), \\( L_{CC} \\) and \\( L_{SR - G} \\) in Eq. (6). The results are shown in Tab. 5. Since \\( L_{SR - Der} \\) and \\( L_{SR - G} \\) are highly correlated, to ensure that the losses are on the same scale, the coefficient for \\( L_{SR - Der} \\) is set to 0.1 times that of \\( L_{SR - G} \\), and \\( L_{SR} \\) in Tab. 5 indicates \\( L_{SR - G} \\). Based on the experimental results, aside from our selected optimal combination, the deraining and generalization performance of our method remain at a high level" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.283, + 0.907, + 0.311 + ], + "angle": 0, + "content": "under most other weight combinations, demonstrating the robustness of our CSUD framework." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.313, + 0.908, + 0.344 + ], + "angle": 0, + "content": "More detailed explanations of our method, experimental details and analysis can be found in the Appendix." + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.355, + 0.906, + 0.383 + ], + "angle": 0, + "content": "Table 5. Ablation experiments on the weight coefficients of each loss function. All models in the table are trained on Rain100L." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.395, + 0.905, + 0.489 + ], + "angle": 0, + "content": "
Lper(λ2)LSR(α2)LCC(α1)Rain100L [43]SPA-data [36]RealRain1K-L [18]
PSNRSSIMPSNRSSIMPSNRSSIM
0.25.00.130.690.93232.130.92928.210.915
0.25.01.033.150.95133.500.93829.750.928
0.21.01032.840.94633.490.93529.100.918
0.2101032.900.95133.400.93329.060.920
05.01032.180.93932.690.93628.240.916
1.05.01032.720.94733.050.93028.700.918
0.25.01033.280.95433.570.93929.210.928
" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.516, + 0.634, + 0.532 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.542, + 0.909, + 0.753 + ], + "angle": 0, + "content": "In this paper, we propose a novel unsupervised framework for image deraining to address the lack of paired data and the poor generalization in real-world scenarios. We first propose a channel consistency loss by introducing the channel consistency prior into the optimization process of the generator, ensuring that more background details are preserved while generating pseudo rainy images. In addition, a self-reconstruction strategy is proposed to alleviate the redundant information transfer problem and further improve the deraining performance and the generalization capability of our method. Extensive experiments on multiple synthetic and real-world datasets verify that our method achieves excellent deraining performance with strong generalization capability across various real-world rainy scenarios." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.911, + 0.9 + ], + "angle": 0, + "content": "Acknowledgement. This work was supported by the National Natural Science Foundation of China under Grant 62171304 and partly by the Natural Science Foundation of Sichuan Province under Grant 2024NSFSC1423, the TCL Science and Technology Innovation Fund under grant 25JZH008, the Young Faculty Technology Innovation Capacity Enhancement Program of Sichuan University under Grant 2024SCUQJTX025, and the Beijing Natural Science Foundation L221011." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7476" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.486, + 0.184 + ], + "angle": 0, + "content": "[1] Liangyu Chen, Xin Lu, Jie Zhang, Xiaojie Chu, and Chengpeng Chen. Hinet: Half instance normalization network for image restoration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 182-192, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.185, + 0.485, + 0.239 + ], + "angle": 0, + "content": "[2] Liangyu Chen, Xiaojie Chu, Xiangyu Zhang, and Jian Sun. Simple baselines for image restoration. In European conference on computer vision, pages 17-33. Springer, 2022. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.241, + 0.484, + 0.324 + ], + "angle": 0, + "content": "[3] Xiang Chen, Jinshan Pan, Kui Jiang, Yufeng Li, Yufeng Huang, Caihua Kong, Longgang Dai, and Zhentao Fan. Unpaired deep image deraining using dual contrastive learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2017-2026, 2022. 1, 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.326, + 0.484, + 0.38 + ], + "angle": 0, + "content": "[4] Xiang Chen, Hao Li, Mingqiang Li, and Jinshan Pan. Learning a sparse transformer network for effective image deraining. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5896-5905, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.382, + 0.484, + 0.45 + ], + "angle": 0, + "content": "[5] Xiang Chen, Jinshan Pan, and Jiangxin Dong. Bidirectional multi-scale implicit neural representations for image deraining. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25627-25636, 2024. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.452, + 0.484, + 0.479 + ], + "angle": 0, + "content": "[6] P Kingma Diederik. Adam: A method for stochastic optimization. (No Title), 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.481, + 0.484, + 0.535 + ], + "angle": 0, + "content": "[7] Xueyang Fu, Jiabin Huang, Xinghao Ding, Yinghao Liao, and John Paisley. Clearing the skies: A deep network architecture for single-image rain removal. IEEE Transactions on Image Processing, 26(6):2944-2956, 2017. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.537, + 0.484, + 0.605 + ], + "angle": 0, + "content": "[8] Xueyang Fu, Jiabin Huang, Delu Zeng, Yue Huang, Xinghao Ding, and John Paisley. Removing rain from single images via a deep detail network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3855-3863, 2017. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.607, + 0.484, + 0.661 + ], + "angle": 0, + "content": "[9] Jinjin Gu, Xianzheng Ma, Xiangtao Kong, Yu Qiao, and Chao Dong. Networks are slacking off: Understanding generalization problem in image deraining. Advances in Neural Information Processing Systems, 36, 2023. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.663, + 0.484, + 0.717 + ], + "angle": 0, + "content": "[10] Kaiming He, Jian Sun, and Xiaou Tang. Single image haze removal using dark channel prior. IEEE transactions on pattern analysis and machine intelligence, 33(12):2341-2353, 2010. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.719, + 0.484, + 0.774 + ], + "angle": 0, + "content": "[11] Xiaowei Hu, Chi-Wing Fu, Lei Zhu, and Pheng-Ann Heng. Depth-attentional features for single-image rain removal. In Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition, pages 8022–8031, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.776, + 0.484, + 0.831 + ], + "angle": 0, + "content": "[12] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1125-1134, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.832, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[13] Kui Jiang, Zhongyuan Wang, Peng Yi, Chen Chen, Baojin Huang, Yimin Luo, Jiayi Ma, and Junjun Jiang. Multi-scale progressive fusion network for single image deraining. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8346-8355, 2020. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.486, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.147 + ], + "angle": 0, + "content": "[14] Xin Jin, Zhibo Chen, Jianxin Lin, Zhikai Chen, and Wei Zhou. Unsupervised single image deraining with self-supervised constraints. In 2019 IEEE International Conference on Image Processing (ICIP), pages 2761-2765. IEEE, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.149, + 0.907, + 0.218 + ], + "angle": 0, + "content": "[15] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14, pages 694-711. Springer, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.219, + 0.907, + 0.285 + ], + "angle": 0, + "content": "[16] Boyun Li, Xiao Liu, Peng Hu, Zhongqin Wu, Jiancheng Lv, and Xi Peng. All-in-one image restoration for unknown corruption. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 17452-17462, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.288, + 0.907, + 0.37 + ], + "angle": 0, + "content": "[17] Siyuan Li, Iago Bruno Araujo, Wenqi Ren, Zhangyang Wang, Eric K Tokuda, Roberto Hirata Junior, Roberto Cesar-Junior, Jiawan Zhang, Xiaojie Guo, and Xiaochun Cao. Single image deraining: A comprehensive benchmark analysis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3838-3847, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.372, + 0.906, + 0.426 + ], + "angle": 0, + "content": "[18] Wei Li, Qiming Zhang, Jing Zhang, Zhen Huang, Xinmei Tian, and Dacheng Tao. Toward real-world single image deraining: A new benchmark and beyond. arXiv preprint arXiv:2206.05514, 2022. 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.428, + 0.907, + 0.495 + ], + "angle": 0, + "content": "[19] Xia Li, Jianlong Wu, Zhouchen Lin, Hong Liu, and Hongbin Zha. Recurrent squeeze-and-excitation context aggregation net for single image deraining. In Proceedings of the European conference on computer vision (ECCV), pages 254-269, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.497, + 0.907, + 0.552 + ], + "angle": 0, + "content": "[20] Yu Li, Robby T Tan, Xiaojie Guo, Jiangbo Lu, and Michael S Brown. Rain streak removal using layer priors. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2736-2744, 2016. 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.553, + 0.907, + 0.62 + ], + "angle": 0, + "content": "[21] Di Lin, Xin Wang, Jia Shen, Renjie Zhang, Ruonan Liu, Miao-hui Wang, Wuyuan Xie, Qing Guo, and Ping Li. Generative status estimation and information decoupling for image rain removal. Advances in Neural Information Processing Systems, 35:4612-4625, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.623, + 0.907, + 0.691 + ], + "angle": 0, + "content": "[22] Xin Lin, Chao Ren, Xiao Liu, Jie Huang, and Yinjie Lei. Unsupervised image denoising in real-world scenarios via self-collaboration parallel generative adversarial branches. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12642-12652, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.692, + 0.907, + 0.747 + ], + "angle": 0, + "content": "[23] Xin Lin, Jingtong Yue, Sixian Ding, Chao Ren, Chun-Le Guo, and Chongyi Li. Unlocking low-light-rainy image restoration by pairwise degradation feature vector guidance. arXiv preprint arXiv:2305.03997, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.748, + 0.907, + 0.802 + ], + "angle": 0, + "content": "[24] Li Liu, Wanli Ouyang, Xiaogang Wang, Paul Fieguth, Jie Chen, Xinwang Liu, and Matti Pietikainen. Deep learning for generic object detection: A survey. International journal of computer vision, 128:261-318, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.804, + 0.907, + 0.858 + ], + "angle": 0, + "content": "[25] Xiao Liu, Xiangyu Liao, Xiuya Shi, Linbo Qing, and Chao Ren. Efficient information modulation network for image super-resolution. In ECAI 2023, pages 1544-1551. IOS Press, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.907, + 0.901 + ], + "angle": 0, + "content": "[26] Xiao Liu, Xiuya Shi, Lufei Chen, Linbo Qing, and Chao Ren. Efficient parallel multi-scale detail and semantic encoding network for lightweight semantic segmentation. In Proceedings" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7477" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.126, + 0.093, + 0.484, + 0.12 + ], + "angle": 0, + "content": "of the 31st ACM International Conference on Multimedia, pages 2544-2552, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.121, + 0.484, + 0.189 + ], + "angle": 0, + "content": "[27] Dongwon Park, Byung Hyun Lee, and Se Young Chun. All-in-one image restoration for unknown degradations using adaptive discriminative filters for specific degradations. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5815-5824. IEEE, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.19, + 0.484, + 0.272 + ], + "angle": 0, + "content": "[28] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.274, + 0.483, + 0.328 + ], + "angle": 0, + "content": "[29] Vaishnav Potlapalli, Syed Waqas Zamir, Salman H Khan, and Fahad Shahbaz Khan. Prompt: Prompting for all-in-one image restoration. Advances in Neural Information Processing Systems, 36, 2023. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.33, + 0.484, + 0.385 + ], + "angle": 0, + "content": "[30] Ruijie Quan, Xin Yu, Yuanzhi Liang, and Yi Yang. Removing raindrops and rain streaks in one go. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9147-9156, 2021. 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.386, + 0.483, + 0.44 + ], + "angle": 0, + "content": "[31] Chao Ren, Yizhong Pan, and Jie Huang. Enhanced latent space blind model for real image denoising via alternative optimization. Advances in Neural Information Processing Systems, 35:38386-38399, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.442, + 0.483, + 0.509 + ], + "angle": 0, + "content": "[32] Dongwei Ren, Wangmeng Zuo, Qinghua Hu, Pengfei Zhu, and Deyu Meng. Progressive image deraining networks: A better and simpler baseline. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3937-3946, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.511, + 0.483, + 0.566 + ], + "angle": 0, + "content": "[33] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. IEEE transactions on pattern analysis and machine intelligence, 39(6):1137-1149, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.567, + 0.483, + 0.608 + ], + "angle": 0, + "content": "[34] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.609, + 0.484, + 0.677 + ], + "angle": 0, + "content": "[35] Cong Wang, Jinshan Pan, Wei Wang, Jiangxin Dong, Mengzhu Wang, Yakun Ju, and Junyang Chen. Promptstorer: A prompting image restoration method with degradation perception. Advances in Neural Information Processing Systems, 36:8898-8912, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.679, + 0.484, + 0.747 + ], + "angle": 0, + "content": "[36] Tianyu Wang, Xin Yang, Ke Xu, Shaozhe Chen, Qiang Zhang, and Rynson WH Lau. Spatial attentive single-image deraining with a high quality real rain dataset. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12270-12279, 2019. 1, 2, 4, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.748, + 0.483, + 0.802 + ], + "angle": 0, + "content": "[37] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.804, + 0.483, + 0.872 + ], + "angle": 0, + "content": "[38] Zhendong Wang, Xiaodong Cun, Jianmin Bao, Wengang Zhou, Jianzhuang Liu, and Houqiang Li. Uformer: A general u-shaped transformer for image restoration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 17683-17693, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[39] Wei Wei, Deyu Meng, Qian Zhao, Zongben Xu, and Ying Wu. Semi-supervised transfer learning for image rain removal. In" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3877-3886, 2019. 3, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.907, + 0.19 + ], + "angle": 0, + "content": "[40] Yanyan Wei, Zhao Zhang, Yang Wang, Mingliang Xu, Yi Yang, Shuicheng Yan, and Meng Wang. Deraincyclegan: Rain attentive cyclegan for single image deraining and rainmaking. IEEE Transactions on Image Processing, 30:4788-4801, 2021. 1, 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.193, + 0.907, + 0.261 + ], + "angle": 0, + "content": "[41] Zbigniew Wojna, Vittorio Ferrari, Sergio Guadarrama, Nathan Silberman, Liang-Chieh Chen, Alireza Fathi, and Jasper Uijlings. The devil is in the decoder: Classification, regression and gans. International Journal of Computer Vision, 127: 1694-1706, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.264, + 0.907, + 0.319 + ], + "angle": 0, + "content": "[42] Wei Xu, Xinyuan Chen, Haoming Guo, Xiaolin Huang, and Wei Liu. Unsupervised image restoration with quality-task-perception loss. IEEE Transactions on Circuits and Systems for Video Technology, 32(9):5736-5747, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.321, + 0.907, + 0.389 + ], + "angle": 0, + "content": "[43] Wenhan Yang, Robby T Tan, Jiashi Feng, Zongming Guo, Shuicheng Yan, and Jiaying Liu. Joint rain detection and removal from a single image with contextualized deep networks. IEEE transactions on pattern analysis and machine intelligence, 42(6):1377-1393, 2019. 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.391, + 0.907, + 0.459 + ], + "angle": 0, + "content": "[44] Rajeev Yasarla, Vishwanath A Sindagi, and Vishal M Patel. Syn2real transfer learning for image deraining using gaussian processes. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2726-2736, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.462, + 0.907, + 0.53 + ], + "angle": 0, + "content": "[45] Yuntong Ye, Yi Chang, Hanyu Zhou, and Luxin Yan. Closing the loop: Joint rain generation and removal via disentangled image translation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2053-2062, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.533, + 0.907, + 0.603 + ], + "angle": 0, + "content": "[46] Yuntong Ye, Changfeng Yu, Yi Chang, Lin Zhu, Xi-Le Zhao, Luxin Yan, and Yonghong Tian. Unsupervised deraining: Where contrastive learning meets self-similarity. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5821-5830, 2022. 1, 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.604, + 0.907, + 0.673 + ], + "angle": 0, + "content": "[47] Qiaosi Yi, Juncheng Li, Qinyan Dai, Faming Fang, Guixu Zhang, and Tieyong Zeng. Structure-preserving deraining with residue channel prior guidance. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4238-4247, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.675, + 0.907, + 0.743 + ], + "angle": 0, + "content": "[48] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, Ming-Hsuan Yang, and Ling Shao. Multi-stage progressive image restoration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14821-14831, 2021. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.745, + 0.907, + 0.828 + ], + "angle": 0, + "content": "[49] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5728-5739, 2022. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.83, + 0.907, + 0.872 + ], + "angle": 0, + "content": "[50] Fan Zhang, Shaodi You, Yu Li, and Ying Fu. GTav-nightrain: Photometric realistic large-scale dataset for night-time rain streak removal. arXiv preprint arXiv:2210.04708, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.907, + 0.901 + ], + "angle": 0, + "content": "[51] Fan Zhang, Shaodi You, Yu Li, and Ying Fu. Learning rain location prior for nighttime deraining. In Proceedings of" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7478" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.484, + 0.12 + ], + "angle": 0, + "content": "the IEEE/CVF International Conference on Computer Vision, pages 13148-13157, 2023. 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.484, + 0.177 + ], + "angle": 0, + "content": "[52] He Zhang and Vishal M Patel. Density-aware single image de-raining using a multi-stream dense network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 695-704, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.179, + 0.484, + 0.247 + ], + "angle": 0, + "content": "[53] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In Proceedings of the IEEE international conference on computer vision, pages 2223-2232, 2017. 2, 3, 6" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.247 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "7479" + } + ] +] \ No newline at end of file diff --git a/2025/Channel Consistency Prior and Self-Reconstruction Strategy Based Unsupervised Image Deraining/1cd88811-9c34-4363-ae8a-2a52dd5a816f_origin.pdf b/2025/Channel Consistency Prior and Self-Reconstruction Strategy Based Unsupervised Image Deraining/1cd88811-9c34-4363-ae8a-2a52dd5a816f_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..91e0ae5d0bc985e21b477c5d067c14a3fdb109e4 --- /dev/null +++ b/2025/Channel Consistency Prior and Self-Reconstruction Strategy Based Unsupervised Image Deraining/1cd88811-9c34-4363-ae8a-2a52dd5a816f_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4e0ed81b3e641ef121af55754a3784d02c4ab5368413abc04fbf1205277eb538 +size 8880021 diff --git a/2025/Channel Consistency Prior and Self-Reconstruction Strategy Based Unsupervised Image Deraining/full.md b/2025/Channel Consistency Prior and Self-Reconstruction Strategy Based Unsupervised Image Deraining/full.md new file mode 100644 index 0000000000000000000000000000000000000000..bc116842855d24cf49142b230a15d3a29b1123a2 --- /dev/null +++ b/2025/Channel Consistency Prior and Self-Reconstruction Strategy Based Unsupervised Image Deraining/full.md @@ -0,0 +1,320 @@ +# Channel Consistency Prior and Self-Reconstruction Strategy Based Unsupervised Image Deraining + +Guanglu Dong $^{1}$ Tianheng Zheng $^{1}$ Yuanzhouhan Cao $^{2}$ Linbo Qing $^{1}$ Chao Ren $^{1*}$ $^{1}$ Sichuan University, Chengdu, China + $^{2}$ Beijing Jiaotong University, Beijing, China +{dongguanglu, zhengtianheng}@stu.scu.edu.cn, yzhcao@bjtu.edu.cn +{qing.lb, chaoren}@scu.edu.cn + +# Abstract + +Recently, deep image deraining models based on paired datasets have made a series of remarkable progress. However, they cannot be well applied in real-world applications due to the difficulty of obtaining real paired datasets and the poor generalization performance. In this paper, we propose a novel Channel Consistency Prior and Self-Reconstruction Strategy Based Unsupervised Image Deraining framework, CSUD, to tackle the aforementioned challenges. During training with unpaired data, CSUD is capable of generating high-quality pseudo clean and rainy image pairs which are used to enhance the performance of deraining network. Specifically, to preserve more image background details while transferring rain streaks from rainy images to the unpaired clean images, we propose a novel Channel Consistency Loss (CCLoss) by introducing the Channel Consistency Prior (CCP) of rain streaks into training process, thereby ensuring that the generated pseudo rainy images closely resemble the real ones. Furthermore, we propose a novel Self-Reconstruction (SR) strategy to alleviate the redundant information transfer problem of the generator, further improving the deraining performance and the generalization capability of our method. Extensive experiments on multiple synthetic and real-world datasets demonstrate that the deraining performance of CSUD surpasses other state-of-the-art unsupervised methods and CSUD exhibits superior generalization capability. Code is available at https://github.com/GuangluDong0728/CSUD. + +# 1. Introduction + +Image deraining is an important task in low-level computer vision [7, 10, 25, 29, 31, 49]. It aims to reconstruct a clean image from its rainy version, which has a profound impact on subsequent high-level tasks such as object detection, recognition, and semantic segmentation + +![](images/91b3619f60909abcf4ab764b2b6370bcc5d2ec68f025545256d4d7d1fcc1fb6f.jpg) +Figure 1. Deraining results on the real rainy images captured by ourselves in real-world scenarios. Compared with the supervised method Restormer [49] and the unsupervised method DerainCycleGAN [40], our CSUD exhibits extremely strong generalization capability in the real world and achieves the best visual results. + +[24, 26, 33, 41]. To mitigate the negative impact of rain, numerous deep learning based supervised image deraining methods [7, 8, 11, 13, 19, 21, 32, 36, 47, 52] have been proposed and they achieve remarkable progress on synthetic datasets. However, the performance of these methods is largely constrained by the difficulty of capturing a large number of high-quality paired datasets when applied to real-world scenarios. Additionally, attempting to directly apply supervised models trained on synthetic datasets to real-world scenarios is also difficult because of the huge domain gap between synthetic and real-world rain [45] and the overfitting issue of supervised methods [9]. + +Compared to supervised methods, unsupervised deep learning methods [3, 14, 40, 42, 46] for image deraining tasks generally achieve end-to-end image deraining without paired clean-rainy images during training. This allows unsupervised methods to effectively handle complex real-world + +scenarios, especially when data collection costs are high. However, due to the high difficulty and challenges of unsupervised deraining methods in training process and the lack of prior knowledge [40], there is still very little relevant research in this field. Additionally, the deraining performance of existing unsupervised methods always lags far behind that of supervised methods, and they face the same generalization issues as supervised methods. + +As shown in Fig. 1, even the SOTA supervised and unsupervised deraining models [40, 49] encounter substantial generalization issues and struggle to achieve satisfactory results when dealing with rainy images captured in real world, and sometimes they even fail to remove any rain streaks due to the more complex rain distributions in real world. Based on the above analysis, how to design an effective unsupervised deraining framework to avoid the need for real paired data and how to improve the generalization performance of deraining methods have become key research issues. + +In this paper, we propose a novel channel consistency prior and self-reconstruction strategy based unsupervised image deraining framework which only needs unpaired data to train the network and can be easily generalized to real-world scenarios. We first conduct detailed analysis of the channel consistency prior (CCP) of rain streaks, and then propose a novel channel consistency loss to constrain the generator to synthesize high-quality pseudo rainy images. In addition, we utilize the proposed self-reconstruction (SR) strategy to further improve the performance of the generator and enhance the generalization capability of the derainer. As shown in Figure 1, our CSUD achieves the most promising performance on the real-world captured images compared to both supervised and unsupervised deraining methods. The main contributions of this work are summarized below: + +- We propose a novel channel consistency prior and self-reconstruction strategy based unsupervised image deraining framework which is able to generate high-quality pseudo paired data for training. This framework can effectively address the issues of insufficient paired data in real world and significantly improve the generalization capability of the network. +- We propose a novel channel consistency loss (CCLoss) according to the channel consistency prior (CCP) of rain streaks to enhance the performance of the generator, preserving more background details while generating pseudo rainy images. To the best of our knowledge, we are the first to incorporate CCP into the training process of unsupervised deraining framework. +- We propose a self-reconstruction (SR) strategy to solve the redundant information transfer problem of the generator. SR strategy can facilitate the accuracy of rain streaks transfer in the generator and enhance the deraining performance and the generalization capability of the derainer without altering the existing network structure. + +- Extensive experiments on numerous benchmark datasets demonstrate that our proposed unsupervised framework can effectively tackle the deraining tasks on both synthetic and real-world rainy images with extremely superior generalization capability in real-world scenarios. + +# 2. Related Works + +# 2.1. Deep Learning Based Single Image Deraining + +Supervised Single Image Deraining. Recently, with the development of deep learning, numerous excellent single image deraining [4, 5, 7, 8, 13, 21, 36, 47, 52] models have emerged. Fu et al. [7] first employ the CNN-based DerainNet to extract and remove the rain layer. Lin et al. [21] propose a generative network equipped with the pixel-wise status estimation and information decoupling for deraining. Lin et al. [23] propose a unified network for low-light-rainy image restoration to jointly handle low-light enhancement and deraining tasks. In addition to the aforementioned works specifically targeting deraining, some networks designed for image restoration [1, 16, 27, 29, 35, 38, 48, 49] have also shown excellent performance in deraining tasks. However, the above fully supervised methods require a large number of high-quality paired training samples, which is almost impossible in real-world scenarios. Simultaneously, due to the overfitting issue, their generalization capability is limited. + +Unsupervised Single Image Deraining. To address the lack of paired datasets in the real world, some outstanding unsupervised deraining methods [3, 14, 40, 42, 46] have been proposed. Wei et al. [40] propose an unsupervised method for rain removal and generation by utilizing the popular CycleGAN [53] framework. Then, through introducing a dual contrastive learning manner to the CycleGAN framework, Chen et al. [3] develop an unpaired adversarial framework which effectively explores mutual properties of the unpaired rainy-clean samples. Ye et al. [46] introduce a novel non-local contrastive learning based unsupervised image deraining method to better distinguish rain streaks from clean images. However, the deraining performance of these unsupervised methods is far inferior to that of supervised methods, especially on real-world scenarios. Compared with previous methods, our proposed CSUD can greatly enhance the unsupervised deraining performance and achieve competitive results compared with some supervised methods. + +# 2.2. Generalization Problem in Image Deraining + +Deep image deraining models often fail to obtain satisfactory performance or even cannot remove any rain streaks when they are applied to real-world scenarios. Generalization problem of deraining models is an urgent issue to be solved. Gu et al. [9] conduct thorough analysis of the generalization problem in image deraining and propose to achieve better generalization by simplifying the complexity of train- + +![](images/81aba27eb357ed84b6880f74088d87ca8afe2c9ffe2db10b415cffc043c55093.jpg) +Figure 2. Framework of the proposed CSUD. The red arrows represent the baseline, purple arrows represent the 3 extra constraints, green arrows represent the self-reconstruction (SR) strategy, and blue arrows represent the CCLoss. The 4 input arrows of RIEM represent 4 separate runs of the generator, each time guided by a different image. Both RIEM and CFEM receive only 1 input image at a time. + +ing images. Some semi-supervised learning frameworks [39, 44] are proposed to analyze the residual difference between domains and enhance the deraining capability in real world. In addition, a series of unsupervised deraining methods [3, 40, 46] have also demonstrated a certain degree of generalization capability in real-world scenarios. However, they have not conducted detailed analysis on how to enhance the generalization capability of unsupervised methods. Inspired by [9], we propose a self-reconstruction strategy to make both generator and derainer pay more attention to the reconstruction of image background, which significantly improves the generalization performance of our method. + +# 3. Methodology + +# 3.1. Overall Framework of Proposed Method + +To address the issue of insufficient paired data and the generalization problem in deraining applications, we propose a novel channel consistency prior and self-reconstruction strategy based unsupervised image deraining framework, CSUD. As illustrated in Fig. 2, our CSUD consists of a derainer $Der$ , a rainy image generator $G$ , and a Patch-GAN [12] discriminator $D$ . For $G$ , it consists of a clean feature extraction module (CFEM), a rain information extraction module (RIEM) and 6 residual blocks. Detailed network structures are detailed in the Appendix. + +In generator $G$ , we effectively utilize unpaired clean and rainy images to synthesize pseudo rainy images. As shown by the red arrows in Fig. 2, we design a GAN-based baseline + +that transitions from " $x \in X \to x_{s1} \in S \to x_{der} \in N$ ", where $X$ , $S$ and $N$ denote clean images, synthetic rainy images and derained images respectively. $G$ transforms clean image $x$ to synthetic rainy image $x_{s1}$ according to the guide of real rainy image $y$ . In order to better extract rain information and transfer it to the clean images, $y$ and $x$ are initially processed by the RIEM and the CFEM respectively to obtain rainy features and clean features which are concatenated for subsequent process. For the real rainy image $y$ and the synthesized rainy image $x_{s1}$ , the corresponding derained clean images $y_{der}$ and $x_{der}$ are obtained through Der. + +The discriminator $D$ is trained to distinguish between synthetic rainy images $x_{s1} \in S$ and real rainy images $y \in Y$ , and the adversarial loss $L_{adv1}$ of the baseline is constrained between them. Furthermore, motivated by existing methods [22, 53], as indicated by the purple arrows in Fig. 2, we add another three processes, $G(Der(x_{s1}), x_{s1}) \to x_{s2}$ , $G(Der(y), y) \to y_{s1}$ , and $G(Der(y), x_{s1}) \to y_{s2}$ , to the framework to enhance training stability. Then, additional three adversarial losses are introduced to CSUD: + +$$ +L _ {G A N} = L _ {a d v 1} + L _ {a d v 2} + L _ {a d v 3} + L _ {a d v 4}. \tag {1} +$$ + +# 3.2. Channel Consistency Loss for Generator + +Channel Consistency Prior (CCP) of Rain Streaks. In RGB rainy images, most rain streaks tend to be consistent across the R, G, and B channels, meaning the pixel values of most rain streaks in the three channels are highly similar. As shown in the second, third, and fourth columns of Fig. 3 (a), we separate both synthetic and real-world rainy images + +![](images/e4c62cb2741d1f3dc28cd2a762e51ef3355402c6d05a920b289d8b2531e9b54e.jpg) +(a) Visualization of channel consistency prior in synthetic, real-world, and nighttime real-world rainy images + +![](images/491b20f20ed41772d50a06a006d3b60238cd9753a4e6ce065f6e7a2f8075a132.jpg) +(b) Cosine similarities of cycle subtractions among R, G, and B channels between clean and rainy images +Figure 3. (a) Visualization of channel consistency prior in synthetic and real-world rainy images. The first column presents the rainy RGB images; the second, third and fourth columns present their R, G, B channels, respectively; the fifth, sixth, and seventh columns present the cycle subtractions of R, G, and B channels, respectively. (b) Cosine similarities of the cycle subtractions of R, G, and B channels between clean and rainy images in the 5 synthetic and real-world datasets. More visualizations and analysis are presented in the Appendix. + +into R, G, and B channels and visualized them. It can be seen clearly that while there are significant differences in the background parts among the three channels, the rain streaks tend to be consistent. We propose to term this phenomenon as "Channel Consistency Prior (CCP)" of rainy streaks. + +Theoretically, according to CCP, the cycle subtractions $((r)_R - (r)_G)$ , $((r)_G - (r)_B)$ , $((r)_B - (r)_R)$ of R, G, and B channels in rainy images will almost contain no rain streaks and should be close to $((n)_R - (n)_G)$ , $((n)_G - (n)_B)$ , $((n)_B - (n)_R)$ in clean images, where $r$ and $n$ represent the rainy and clean image, $(\cdot)_R$ , $(\cdot)_G$ , and $(\cdot)_B$ denote R, G, and B channels respectively. In the fifth, sixth, and seventh columns of Fig. 3 (a), it can be observed that the cycle subtractions of R, G, and B channels nearly don't contain any rain streaks, qualitatively confirming the correctness of CCP. Furthermore, as shown in Fig. 3 (b), we quantitatively present the cosine similarities of the cycle subtractions of R, G, and B channels between clean and rainy images in both synthetic and real-world benchmark datasets [18, 20, 36, 43]. It is evident that the cosine similarities are exceptionally high, which can further demonstrate our hypothesis. + +Channel Consistency Loss. Based on the above analysis of CCP, the cycle subtractions of R, G, and B channels between input clean image and output pseudo rainy image of the generator $G$ should tend to be consistent. To this end, we propose a novel Channel Consistency Loss (CCLoss) to preserve more color and texture details of the background while generating pseudo rainy images: + +$$ +\begin{array}{l} L _ {C C} = \left\| \left(\left(x _ {s 1}\right) _ {R} - \left(x _ {s 1}\right) _ {G}\right), \left(x _ {R} - x _ {G}\right) \right\| _ {1} \\ + \left\| \left(\left(x _ {s 1}\right) _ {G} - \left(x _ {s 1}\right) _ {B}\right), \left(x _ {G} - x _ {B}\right) \right\| _ {1} \tag {2} \\ + \left\| \left(\left(x _ {s 1}\right) _ {B} - \left(x _ {s 1}\right) _ {R}\right), \left(x _ {B} - x _ {R}\right) \right\| _ {1}, \\ \end{array} +$$ + +where $(x_{s1})_R, (x_{s1})_G, (x_{s1})_B, x_R, x_G, x_B$ denotes R, G, B color channels of $x_{s1}$ and $x, ||\cdot||_1$ represents the $L_1$ norm and $L_{CC}$ represents the proposed CCLoss, respectively. By introducing CCP into our unsupervised deraining framework, $G$ will generate more accurate synthesized rainy images, which will further improve the performance of the derainer. + +# 3.3. Self-Reconstruction Strategy + +Redundant Information Transfer Problem of the Generator. Although adversarial losses can effectively constrain the + +![](images/7caee91250ebf886ef717848f36f2812849980b90b8299834d7a96c24fd3f045.jpg) +Figure 4. Visualization of redundant information transfer problem. + +training process of generator $G$ , the data synthesis process of $G$ is highly susceptible to interference from the redundant information in rainy images, especially those edge textures that are very similar to rain streaks. As shown in Fig. 4, the edge textures of some background objects in rainy images, such as the outline of the woman's hair, the leg of rhino, etc., are mistakenly transferred to the clean images by $G$ . As a result, $G$ will generate inaccurate pseudo rainy images, interfering with the learning process of $Der$ . + +Self-Reconstruction Strategy. To solve the redundant information transfer problem, we propose a novel Self-Reconstruction (SR) strategy, which enables $G$ to extract more accurate rain streak information and synthesizes more realistic pseudo rainy images. The core idea of SR strategy is to encourage $G$ to better distinguish rain streaks from background in the rainy images, ensuring that only rain streak information is extracted for subsequent generation of pseudo rainy images, while redundant information of background is disregarded. As shown by the green arrows in Fig. 2, when the input clean image $x$ itself, the derained image $y_{der}$ are input with clean image $x$ into $G$ , $G$ learns to reconstruct the input clean image $x$ itself without any additional redundant information from $x$ and $y_{der}$ , which means that the self-reconstructed images $x_{s1}$ and $x_{s2}$ should be close to $x$ . Thus, we can utilize the following self-reconstruction loss (SRLoss) as additional training constraint of $G$ : + +$$ +L _ {S R - G} = \left\| G (x, x), x \right\| _ {1} + \left\| G (x, y _ {d e r}), x \right\| _ {1}, \tag {3} +$$ + +where $L_{SR - G}$ represents the SRLoss for $G$ . SR strategy allows $G$ to pay more attention to reconstructing the background of clean images while learning the transfer of rain streaks, significantly improving the accuracy of generating pseudo rainy images. Under the constraint of $L_{SR - G}$ , $G$ will obtain the ability to generate clean images themselves guided by images without rain streaks. Therefore, the cleaner the derained images $y_{der}$ and $x_{der}$ restored by $Der$ are, the closer the self-reconstructed images $x_{s2}$ and $x_{s3}$ are to the original clean ones $x$ . Then, the following SRLoss can be utilized to further improve the performance of $Der$ : + +$$ +L _ {S R - D e r} = \left\| G (x, x _ {d e r}), x \right\| _ {1} + \left\| G (x, y _ {d e r}), x \right\| _ {1}, \tag {4} +$$ + +![](images/c16c3a781b06600c2afbcb800eb781ba306111a89fd058d7ada04e9aab946cb7.jpg) +Figure 5. The implementation principles of SR strategy. + +where $L_{SR - Der}$ represents the SRLoss for $Der$ . Under the constraint of $L_{SR - Der}$ , $Der$ will also focus more on reconstructing the background content while learning to derain, which greatly enhances the generalization capability of $Der$ . + +As shown in the third and fourth columns of Figure 4, SR strategy effectively alleviates the redundant information transfer problem. Through SR strategy, $G$ can achieve robust rain streak transfer with more background detail fidelity, greatly enhancing the deraining performance of Der. + +# 3.4. Loss Function + +Due to the high difficulty and instability of training a GAN-based unsupervised framework, in addition to adopting the L1 loss for $Der$ , we also incorporate SSIM loss and perceptual loss[15] to enhance the performance of $Der$ : + +$$ +\begin{array}{l} L _ {D e r} = \left\| x, x _ {d e r} \right\| _ {1} + \lambda_ {1} L _ {S S I M} (x, x _ {d e r}) \tag {5} \\ + \lambda_ {2} | | \phi_ {i} (x), \phi_ {i} (x _ {d e r}) | | _ {1} + \lambda_ {3} L _ {S R - D e r}, \\ \end{array} +$$ + +where $||\cdot ||_1$ represents the $L_{1}$ norm, and $L_{SSIM}(\cdot)$ represents the SSIM loss, $\phi_i(\cdot)$ is the layers of pre-trained VGG-19 [34], $\lambda_1, \lambda_2, \lambda_3$ represent the hyperparameters for SSIM loss, perceptual loss and SRLoss for derainer respectively. + +Finally, we obtain the total loss to train our framework: + +$$ +L _ {t o t a l} = \min _ {G} \max _ {D} L _ {G A N} + L _ {D e r} + \alpha_ {1} L _ {C C} + \alpha_ {2} L _ {S R - G}, \tag {6} +$$ + +where $\alpha_{1},\alpha_{2}$ represent the hyperparameters for CCLoss and SRLoss for generator respectively. + +# 4. Experiments + +To demonstrate the effectiveness of our CSUD, following [18, 40], we conduct experiments from the following two aspects: (1) unsupervised deraining performance, (2) generalization performance. In experiment (1), we train independent models for different datasets respectively, including both synthetic and real-world datasets. In experiment (2), we only train our model on the synthetic dataset Rain100L [43], and then test on various real-world and nighttime datasets. + +Table 1. Quantitative deraining performance comparisons of different methods. Red and blue colors indicate the 1st and 2nd ranks among the unsupervised methods, respectively. S and U indicate supervised method and unsupervised method, respectively. + +
DatasetsRain100L [43]Rain100H [43]Rain12 [20]Rain800 [17]RealRain1K-L [18]RealRain1K-H [18]
MetricsPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIM
DerainNet [7] (S)27.030.88414.920.592//22.770.81027.090.92522.880.889
DDN [8] (S)32.380.92624.640.84934.040.93321.160.73231.180.91729.170.878
SPA-Net [36] (S)31.950.91926.070.857//24.370.86130.430.94725.760.910
MPRNet [48] (S)34.950.95928.530.87236.840.96428.680.87536.290.97234.740.964
NAFNet [2] (S)37.000.97829.660.90034.810.94328.720.88638.800.98636.110.976
Restormer [49] (S)37.570.97429.460.889//30.330.90540.900.98539.570.981
PromptIR [29] (S)38.340.98328.690.87735.090.94529.070.88736.990.97333.610.953
NeRD-Rain-S [5] (S)42.000.99032.860.93235.390.94228.820.87838.640.97936.690.970
CycleGAN [53] (U)24.610.83420.590.70421.560.84523.950.81920.190.82017.530.762
DerainCycleGAN [40] (U)31.490.93621.130.71033.520.94024.320.84228.160.90124.780.844
NLCL [46] (U)20.500.71922.310.72822.680.73518.940.67023.060.83220.640.719
DCD-GAN [3] (U)31.820.94122.470.75331.560.92425.610.81330.490.93927.820.895
CSUD (Ours)33.280.95424.420.80834.560.95123.580.79532.710.95929.670.928
+ +![](images/f9de6d25327ccdd6df61d7a9a744fa66437396fe6b1ee7919f4dfff4e9fca88d.jpg) +Inupt + +![](images/1dac05848eea7e9aca11190c9dd56fe020b9580b74d69830629b4a75156140e5.jpg) +GT + +![](images/9daef77910392ccbfcf76ea0e4bb9b001cfca1e5b7295a56fdbf8c7714ffe83e.jpg) +MPRNet +Figure 6. Qualitative deraining performance comparisons on Rain100L [43] and Rain12 [20] datasets. Our CSUD achieves competitive visual results comparable to the SOTA supervised method NeRD-Rain-S [5]. + +![](images/63968dd4f3ae89da832eabfd8b853722d62f175023e1820d335aa2004f8dafd3.jpg) +NeRD-Rain-S + +![](images/ca3143bc1f9c89048715d6ecc1958e38d7e10371e8f130088a13968a56f55ae1.jpg) +DerainCycleGAN + +![](images/3873f182fc80c2779c88202905478da033376e8f7d52bfe58fbeb3630bb1b205.jpg) +NLCL + +![](images/979e7b42e4db60f7988a61748a50e6a04153c2010f31f91ab4711fa7cece0281.jpg) +DCD-GAN + +![](images/6f7559186be8aced74c8b3a604de130df151533a002ff09894026a86008bc00b.jpg) +CSUD (Ours) + +Datasets and Metrics. We utilize 10 challenging benchmark datasets in our study, including synthetic datasets: Rain100L [43], Rain100H [43], Rain12 [20], Rain800 [17], real-world datasets: RealRain1K-L [18], RealRain1K-H [18], SPA-data [36], RainDS [30], Internet-Data [39], and a nighttime dataset which is based on GTAV-NightRain dataset [50] and enlarged the rain density by [51], we name this nighttime dataset as Night-Rain in the following experiments. As for RainDS, we only use the real rain streaks subset for testing. Following existing methods [3, 40, 49], we use PSNR and SSIM [37] to evaluate different methods. + +Implementation Details. As for derainer, we employ the CNN-based simple image restoration baseline NAFNet [2]. Our framework is implemented by PyTorch [28] with a GeForce RTX 3090 GPU. We adopt the Adam optimizer [6] to train our network for 200 epochs with the initial learning rate of $1e^{-4}$ , followed by another 100 epochs with a learning rate of $1e^{-5}$ . All training images are randomly cropped to $256 \times 256$ patches in an unpaired manner. The hyperparameters $\lambda_{1}, \lambda_{2}$ , and $\lambda_{3}$ in Eq. 5 are set to 1, 0.2 and 0.5, while $\alpha_{1}$ and $\alpha_{2}$ in Eq. 6 are set to 10 and 5, respectively. + +# 4.1. Unsupervised Deraining Results + +We compare our method with unsupervised methods from the most recent years, including [3, 40, 46, 53]. Noting that the related methods for comparison are limited, since only a few + +unsupervised deraining models have been proposed in the field. We also list some SOTA supervised methods, including [2, 5, 7, 8, 29, 36, 48, 49]. Tab. 1 presents the quantitative deraining results of different methods. It is evident that our CSUD significantly outperforms other unsupervised methods on most datasets. Compared to DCD-GAN[3], our CSUD respectively achieves 1.46dB, 3.00dB, and 2.22dB improvement in PSNR on Rain100L, Rain12 and RealRain1K-L datasets. Additionally, our CSUD even achieves competitive results comparable to several supervised methods. As shown in Fig. 6, compared to other unsupervised methods, our CSUD achieves better deraining results while preserving more color and texture details of image background. + +# 4.2. Generalization Results + +To validate the generalization capability of CSUD, we compare the results on 5 real-world datasets[18, 30, 36, 39] and a nighttime dataset [51] with SOTA supervised and unsupervised methods, including [2, 3, 5, 7, 29, 40, 46, 48, 49]. Noting that there are no ground truth images in the InternetData [39] dataset, thereby only visual results are presented. All methods are trained on synthetic datasets and tested on the unseen real-world and nighttime datasets. A prevailing perspective is that supervised methods usually outperform unsupervised methods. But as shown in Tab. 2, our CSUD almost comprehensively surpasses supervised models on + +Table 2. Quantitative generalization performance comparisons of different deraining methods. Red and blue colors indicate the 1st and 2nd ranks among the unsupervised methods. S, U indicate supervised and unsupervised methods, respectively. + +
DatasetsRealRain1K-L [18]RealRain1K-H [18]SPA-data [36]RainDS [30]Night-Rain [51]
MetricsPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIM
DerainNet [7] (S)25.400.84721.910.74230.780.91221.350.591//
MPRNet [48] (S)26.750.89923.580.84033.000.94223.240.65626.130.817
NAFNet [2] (S)28.110.91524.360.84833.460.94122.480.62526.150.816
Restormer [49] (S)26.290.88323.130.85132.620.93622.960.64825.850.808
PromptIR [29] (S)28.140.91524.280.85133.500.94022.480.62526.370.815
NeRD-Rain-S [5] (S)27.640.90423.960.83133.030.93322.370.62226.380.810
DerainCycleGAN [40] (U)28.250.89924.200.82133.530.93718.750.46025.850.802
NLCL [46] (U)20.180.67419.810.62918.760.72717.000.47722.010.475
DCD-GAN [3] (U)27.490.91223.860.84531.550.93022.130.61025.540.798
CSUD (Ours)29.210.92825.450.87833.570.93922.540.61325.940.805
+ +![](images/f1dade8c00373d7dcb13ed0742c934d8288abc0a7dd1f2b90eb42f9fcc244d68.jpg) +Figure 7. Qualitative generalization results on RealRain1K-L [18] and RealRain1K-H [18] real-world datasets. Our CSUD achieves best visual results among both SOTA supervised and unsupervised methods in real-world scenes. + +Table 3. Effect of CCLoss and SR strategy on CSUD. + +
CCSRRain100L[43]Rain12[20]RealRain1K-L[18]
PSNRSSIMPSNRSSIMPSNRSSIM
XX30.660.92531.600.93130.420.932
✓X31.240.93432.620.94831.140.943
X✓32.690.94933.430.93931.860.950
✓✓33.280.95434.560.95132.710.959
+ +the first 4 real-world datasets, while remaining competitive on the nighttime dataset. This strongly demonstrates the excellent generalization ability of our method. + +As shown in Fig. 7 and Fig. 8, the performance of some most advanced supervised methods like Restormer [49] and NeRD-Rain-S [5] is extremely poor when directly applied to real-world scenarios. In contrast, our CSUD achieves better visual results in real world. It is worth mentioning again that, as shown in Fig. 1 in Sec. 1, our method also shows extremely superior deraining performance on the rainy images captured by ourselves in real scenarios, which further demonstrates our excellent generalization capability. + +# 4.3. Ablation Studies + +Effect of CCLoss and SR Strategy on Deraining Performance. To verify the effectiveness of CCLoss and SR strategy, we conduct ablation experiments on Rain100L [43], + +Rain12 [20] and RealRain1K-L [18] datasets. As shown in Table 3, compared to the baseline, the effect on PSNR is improved by 1.38dB, 1.02dB, and 0.72dB among the above three datasets with CCLoss, respectively. And the effect is improved by 1.38dB, 1.83dB, and 1.44dB with SR strategy, respectively. Furthermore, when both CCLoss and SR strategy are introduced into the baseline, the PSNR on the three datasets is increased by 2.62dB, 2.96dB, and 2.29dB respectively. The above ablation results strongly verify the superiority and importance of CCLoss and SR strategy. + +Effect of CSUD framework on Generalization Performance. To validate the effectiveness of our CSUD in improving generalization performance and the universality and robustness of our method, we also conduct experiments incorporating CSUD framework on another 2 different baselines: NeRD-Rain-S [5] and Restormer [49]. Fig. 9 shows the PSNR improvements of different CSUD based models compared to their supervised version. Notably, we train all the above models on Rain100L and test them on the 3 real-world datasets. This strongly demonstrates the effectiveness of CSUD in improving generalization performance. + +Effect of Different Training Datasets on Generalization Performance. To further validate the generalization performance of CSUD, we employ 4 different CSUD models which are respectively trained on 4 synthetic and real- + +![](images/9e2ad534a36fae4ed759ec9e803bc5e689719c4047cc14cc0afe3695a67f4b50.jpg) +Input + +![](images/089d8648dcf54bd2f09675466293f2985658341b3980468cce420ba34a9b0241.jpg) +Restormer + +![](images/e864e12231b930dc1815ec087f306d4e93ebca13d00514a5e2ab66f587f5a422.jpg) +PromptIR + +![](images/910033d295110dce184f4f188b6e95efcf811c662b69612b8c00793f803be3df.jpg) +NeRD-Rain-S + +![](images/cf0cd317f077a293d450fa3141c3a3d2c265d39d32271e057a3f556baeab7812.jpg) +DerainCycleGAN +Figure 8. Qualitative generalization results on Internet-Data [39] real-world dataset. + +![](images/5615b8809e10dc6283fdabb209b5a1375f16766b7c6e5aee3a0f498e83678005.jpg) +NLCL + +![](images/3959e7251f64ee540ee6f0d02185a4b4ceba7ae0799c008e5034e13d8c14c8d1.jpg) +DCD-GAN + +![](images/5b94aed754c140f6a5fb3a8fc45189ebc7fb99c68f79021e2e30ffb9c7bd2da0.jpg) +CSUD (Ours) + +![](images/2cc05fb52f79d70e476717d8cdc46fdeb784ecd1ce2bdb221f6327e6b63d08cf.jpg) +Figure 9. PSNR improvement of different baselines with CSUD. We test the models' generalization performance on 3 real-world datasets out of domain compared with their supervised version. + +Table 4. Effect of different training datasets on generalization performance. We test 4 different CSUD models which are trained on 4 different datasets on 3 real-world datasets out of domain. + +
Train datasetsSPA-data [36]Night-Rain [51]RainDS [30]
PSNRSSIMPSNRSSIMPSNRSSIM
Rain100L [43]33.570.93925.940.80522.540.613
Rain100H [43]29.610.88924.530.70221.850.600
RealRain1K-L [18]32.950.93728.140.83522.900.630
RealRain1K-H [18]29.250.89726.310.78722.300.564
+ +world datasets with different rain distributions, and test them on 3 real-world datasets. As shown in Tab. 4, in most cases, CSUD can achieve strong generalization performance, demonstrating the robustness of our approach. Due to the significant differences between rain streaks in Rain100H and real world, its generalization capability is the worst. Additionally, the generalization performance when training on Rain100L is comparable to that when training on real-world datasets, demonstrating our strong generalization ability. + +Effect of the Weight Coefficients of Each Loss Function. We conduct extensive ablation experiments on the weight coefficients of perceptual loss in Eq. (5), $L_{CC}$ and $L_{SR - G}$ in Eq. (6). The results are shown in Tab. 5. Since $L_{SR - Der}$ and $L_{SR - G}$ are highly correlated, to ensure that the losses are on the same scale, the coefficient for $L_{SR - Der}$ is set to 0.1 times that of $L_{SR - G}$ , and $L_{SR}$ in Tab. 5 indicates $L_{SR - G}$ . Based on the experimental results, aside from our selected optimal combination, the deraining and generalization performance of our method remain at a high level + +under most other weight combinations, demonstrating the robustness of our CSUD framework. + +More detailed explanations of our method, experimental details and analysis can be found in the Appendix. + +Table 5. Ablation experiments on the weight coefficients of each loss function. All models in the table are trained on Rain100L. + +
Lper(λ2)LSR(α2)LCC(α1)Rain100L [43]SPA-data [36]RealRain1K-L [18]
PSNRSSIMPSNRSSIMPSNRSSIM
0.25.00.130.690.93232.130.92928.210.915
0.25.01.033.150.95133.500.93829.750.928
0.21.01032.840.94633.490.93529.100.918
0.2101032.900.95133.400.93329.060.920
05.01032.180.93932.690.93628.240.916
1.05.01032.720.94733.050.93028.700.918
0.25.01033.280.95433.570.93929.210.928
+ +# 5. Conclusion + +In this paper, we propose a novel unsupervised framework for image deraining to address the lack of paired data and the poor generalization in real-world scenarios. We first propose a channel consistency loss by introducing the channel consistency prior into the optimization process of the generator, ensuring that more background details are preserved while generating pseudo rainy images. In addition, a self-reconstruction strategy is proposed to alleviate the redundant information transfer problem and further improve the deraining performance and the generalization capability of our method. Extensive experiments on multiple synthetic and real-world datasets verify that our method achieves excellent deraining performance with strong generalization capability across various real-world rainy scenarios. + +Acknowledgement. This work was supported by the National Natural Science Foundation of China under Grant 62171304 and partly by the Natural Science Foundation of Sichuan Province under Grant 2024NSFSC1423, the TCL Science and Technology Innovation Fund under grant 25JZH008, the Young Faculty Technology Innovation Capacity Enhancement Program of Sichuan University under Grant 2024SCUQJTX025, and the Beijing Natural Science Foundation L221011. + +# References + +[1] Liangyu Chen, Xin Lu, Jie Zhang, Xiaojie Chu, and Chengpeng Chen. Hinet: Half instance normalization network for image restoration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 182-192, 2021. 2 +[2] Liangyu Chen, Xiaojie Chu, Xiangyu Zhang, and Jian Sun. Simple baselines for image restoration. In European conference on computer vision, pages 17-33. Springer, 2022. 6, 7 +[3] Xiang Chen, Jinshan Pan, Kui Jiang, Yufeng Li, Yufeng Huang, Caihua Kong, Longgang Dai, and Zhentao Fan. Unpaired deep image deraining using dual contrastive learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2017-2026, 2022. 1, 2, 3, 6, 7 +[4] Xiang Chen, Hao Li, Mingqiang Li, and Jinshan Pan. Learning a sparse transformer network for effective image deraining. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5896-5905, 2023. 2 +[5] Xiang Chen, Jinshan Pan, and Jiangxin Dong. Bidirectional multi-scale implicit neural representations for image deraining. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25627-25636, 2024. 2, 6, 7 +[6] P Kingma Diederik. Adam: A method for stochastic optimization. (No Title), 2014. 6 +[7] Xueyang Fu, Jiabin Huang, Xinghao Ding, Yinghao Liao, and John Paisley. Clearing the skies: A deep network architecture for single-image rain removal. IEEE Transactions on Image Processing, 26(6):2944-2956, 2017. 1, 2, 6, 7 +[8] Xueyang Fu, Jiabin Huang, Delu Zeng, Yue Huang, Xinghao Ding, and John Paisley. Removing rain from single images via a deep detail network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3855-3863, 2017. 1, 2, 6 +[9] Jinjin Gu, Xianzheng Ma, Xiangtao Kong, Yu Qiao, and Chao Dong. Networks are slacking off: Understanding generalization problem in image deraining. Advances in Neural Information Processing Systems, 36, 2023. 1, 2, 3 +[10] Kaiming He, Jian Sun, and Xiaou Tang. Single image haze removal using dark channel prior. IEEE transactions on pattern analysis and machine intelligence, 33(12):2341-2353, 2010. 1 +[11] Xiaowei Hu, Chi-Wing Fu, Lei Zhu, and Pheng-Ann Heng. Depth-attentional features for single-image rain removal. In Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition, pages 8022–8031, 2019. 1 +[12] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1125-1134, 2017. 3 +[13] Kui Jiang, Zhongyuan Wang, Peng Yi, Chen Chen, Baojin Huang, Yimin Luo, Jiayi Ma, and Junjun Jiang. Multi-scale progressive fusion network for single image deraining. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8346-8355, 2020. 1, 2 + +[14] Xin Jin, Zhibo Chen, Jianxin Lin, Zhikai Chen, and Wei Zhou. Unsupervised single image deraining with self-supervised constraints. In 2019 IEEE International Conference on Image Processing (ICIP), pages 2761-2765. IEEE, 2019. 1, 2 +[15] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14, pages 694-711. Springer, 2016. 5 +[16] Boyun Li, Xiao Liu, Peng Hu, Zhongqin Wu, Jiancheng Lv, and Xi Peng. All-in-one image restoration for unknown corruption. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 17452-17462, 2022. 2 +[17] Siyuan Li, Iago Bruno Araujo, Wenqi Ren, Zhangyang Wang, Eric K Tokuda, Roberto Hirata Junior, Roberto Cesar-Junior, Jiawan Zhang, Xiaojie Guo, and Xiaochun Cao. Single image deraining: A comprehensive benchmark analysis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3838-3847, 2019. 6 +[18] Wei Li, Qiming Zhang, Jing Zhang, Zhen Huang, Xinmei Tian, and Dacheng Tao. Toward real-world single image deraining: A new benchmark and beyond. arXiv preprint arXiv:2206.05514, 2022. 4, 5, 6, 7, 8 +[19] Xia Li, Jianlong Wu, Zhouchen Lin, Hong Liu, and Hongbin Zha. Recurrent squeeze-and-excitation context aggregation net for single image deraining. In Proceedings of the European conference on computer vision (ECCV), pages 254-269, 2018. 1 +[20] Yu Li, Robby T Tan, Xiaojie Guo, Jiangbo Lu, and Michael S Brown. Rain streak removal using layer priors. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2736-2744, 2016. 4, 6, 7 +[21] Di Lin, Xin Wang, Jia Shen, Renjie Zhang, Ruonan Liu, Miao-hui Wang, Wuyuan Xie, Qing Guo, and Ping Li. Generative status estimation and information decoupling for image rain removal. Advances in Neural Information Processing Systems, 35:4612-4625, 2022. 1, 2 +[22] Xin Lin, Chao Ren, Xiao Liu, Jie Huang, and Yinjie Lei. Unsupervised image denoising in real-world scenarios via self-collaboration parallel generative adversarial branches. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12642-12652, 2023. 3 +[23] Xin Lin, Jingtong Yue, Sixian Ding, Chao Ren, Chun-Le Guo, and Chongyi Li. Unlocking low-light-rainy image restoration by pairwise degradation feature vector guidance. arXiv preprint arXiv:2305.03997, 2023. 2 +[24] Li Liu, Wanli Ouyang, Xiaogang Wang, Paul Fieguth, Jie Chen, Xinwang Liu, and Matti Pietikainen. Deep learning for generic object detection: A survey. International journal of computer vision, 128:261-318, 2020. 1 +[25] Xiao Liu, Xiangyu Liao, Xiuya Shi, Linbo Qing, and Chao Ren. Efficient information modulation network for image super-resolution. In ECAI 2023, pages 1544-1551. IOS Press, 2023. 1 +[26] Xiao Liu, Xiuya Shi, Lufei Chen, Linbo Qing, and Chao Ren. Efficient parallel multi-scale detail and semantic encoding network for lightweight semantic segmentation. In Proceedings + +of the 31st ACM International Conference on Multimedia, pages 2544-2552, 2023. 1 +[27] Dongwon Park, Byung Hyun Lee, and Se Young Chun. All-in-one image restoration for unknown degradations using adaptive discriminative filters for specific degradations. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5815-5824. IEEE, 2023. 2 +[28] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019. 6 +[29] Vaishnav Potlapalli, Syed Waqas Zamir, Salman H Khan, and Fahad Shahbaz Khan. Prompt: Prompting for all-in-one image restoration. Advances in Neural Information Processing Systems, 36, 2023. 1, 2, 6, 7 +[30] Ruijie Quan, Xin Yu, Yuanzhi Liang, and Yi Yang. Removing raindrops and rain streaks in one go. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9147-9156, 2021. 6, 7, 8 +[31] Chao Ren, Yizhong Pan, and Jie Huang. Enhanced latent space blind model for real image denoising via alternative optimization. Advances in Neural Information Processing Systems, 35:38386-38399, 2022. 1 +[32] Dongwei Ren, Wangmeng Zuo, Qinghua Hu, Pengfei Zhu, and Deyu Meng. Progressive image deraining networks: A better and simpler baseline. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3937-3946, 2019. 1 +[33] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. IEEE transactions on pattern analysis and machine intelligence, 39(6):1137-1149, 2016. 1 +[34] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 5 +[35] Cong Wang, Jinshan Pan, Wei Wang, Jiangxin Dong, Mengzhu Wang, Yakun Ju, and Junyang Chen. Promptstorer: A prompting image restoration method with degradation perception. Advances in Neural Information Processing Systems, 36:8898-8912, 2023. 2 +[36] Tianyu Wang, Xin Yang, Ke Xu, Shaozhe Chen, Qiang Zhang, and Rynson WH Lau. Spatial attentive single-image deraining with a high quality real rain dataset. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12270-12279, 2019. 1, 2, 4, 6, 7, 8 +[37] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 6 +[38] Zhendong Wang, Xiaodong Cun, Jianmin Bao, Wengang Zhou, Jianzhuang Liu, and Houqiang Li. Uformer: A general u-shaped transformer for image restoration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 17683-17693, 2022. 2 +[39] Wei Wei, Deyu Meng, Qian Zhao, Zongben Xu, and Ying Wu. Semi-supervised transfer learning for image rain removal. In + +Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3877-3886, 2019. 3, 6, 8 +[40] Yanyan Wei, Zhao Zhang, Yang Wang, Mingliang Xu, Yi Yang, Shuicheng Yan, and Meng Wang. Deraincyclegan: Rain attentive cyclegan for single image deraining and rainmaking. IEEE Transactions on Image Processing, 30:4788-4801, 2021. 1, 2, 3, 5, 6, 7 +[41] Zbigniew Wojna, Vittorio Ferrari, Sergio Guadarrama, Nathan Silberman, Liang-Chieh Chen, Alireza Fathi, and Jasper Uijlings. The devil is in the decoder: Classification, regression and gans. International Journal of Computer Vision, 127: 1694-1706, 2019. 1 +[42] Wei Xu, Xinyuan Chen, Haoming Guo, Xiaolin Huang, and Wei Liu. Unsupervised image restoration with quality-task-perception loss. IEEE Transactions on Circuits and Systems for Video Technology, 32(9):5736-5747, 2022. 1, 2 +[43] Wenhan Yang, Robby T Tan, Jiashi Feng, Zongming Guo, Shuicheng Yan, and Jiaying Liu. Joint rain detection and removal from a single image with contextualized deep networks. IEEE transactions on pattern analysis and machine intelligence, 42(6):1377-1393, 2019. 4, 5, 6, 7, 8 +[44] Rajeev Yasarla, Vishwanath A Sindagi, and Vishal M Patel. Syn2real transfer learning for image deraining using gaussian processes. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2726-2736, 2020. 3 +[45] Yuntong Ye, Yi Chang, Hanyu Zhou, and Luxin Yan. Closing the loop: Joint rain generation and removal via disentangled image translation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2053-2062, 2021. 1 +[46] Yuntong Ye, Changfeng Yu, Yi Chang, Lin Zhu, Xi-Le Zhao, Luxin Yan, and Yonghong Tian. Unsupervised deraining: Where contrastive learning meets self-similarity. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5821-5830, 2022. 1, 2, 3, 6, 7 +[47] Qiaosi Yi, Juncheng Li, Qinyan Dai, Faming Fang, Guixu Zhang, and Tieyong Zeng. Structure-preserving deraining with residue channel prior guidance. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4238-4247, 2021. 1, 2 +[48] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, Ming-Hsuan Yang, and Ling Shao. Multi-stage progressive image restoration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14821-14831, 2021. 2, 6, 7 +[49] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5728-5739, 2022. 1, 2, 6, 7 +[50] Fan Zhang, Shaodi You, Yu Li, and Ying Fu. GTav-nightrain: Photometric realistic large-scale dataset for night-time rain streak removal. arXiv preprint arXiv:2210.04708, 2022. 6 +[51] Fan Zhang, Shaodi You, Yu Li, and Ying Fu. Learning rain location prior for nighttime deraining. In Proceedings of + +the IEEE/CVF International Conference on Computer Vision, pages 13148-13157, 2023. 6, 7, 8 +[52] He Zhang and Vishal M Patel. Density-aware single image de-raining using a multi-stream dense network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 695-704, 2018. 1, 2 +[53] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In Proceedings of the IEEE international conference on computer vision, pages 2223-2232, 2017. 2, 3, 6 \ No newline at end of file diff --git a/2025/Channel Consistency Prior and Self-Reconstruction Strategy Based Unsupervised Image Deraining/images.zip b/2025/Channel Consistency Prior and Self-Reconstruction Strategy Based Unsupervised Image Deraining/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..8d0bc3ed7fcd2cae229978dd70dbbf0aa79ad590 --- /dev/null +++ b/2025/Channel Consistency Prior and Self-Reconstruction Strategy Based Unsupervised Image Deraining/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0d1000ec3eabb0845b45b79a00b12c91eef1659cc0b32f402e194ed931bee3c5 +size 1032163 diff --git a/2025/Channel Consistency Prior and Self-Reconstruction Strategy Based Unsupervised Image Deraining/layout.json b/2025/Channel Consistency Prior and Self-Reconstruction Strategy Based Unsupervised Image Deraining/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..4cc3ee35424ae6dd05312c7960322c84e0cd5dd0 --- /dev/null +++ b/2025/Channel Consistency Prior and Self-Reconstruction Strategy Based Unsupervised Image Deraining/layout.json @@ -0,0 +1,9611 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 56, + 102, + 555, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 102, + 555, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 102, + 555, + 140 + ], + "type": "text", + "content": "Channel Consistency Prior and Self-Reconstruction Strategy Based Unsupervised Image Deraining" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 68, + 160, + 545, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 160, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 68, + 160, + 545, + 217 + ], + "type": "text", + "content": "Guanglu Dong" + }, + { + "bbox": [ + 68, + 160, + 545, + 217 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 68, + 160, + 545, + 217 + ], + "type": "text", + "content": " Tianheng Zheng" + }, + { + "bbox": [ + 68, + 160, + 545, + 217 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 68, + 160, + 545, + 217 + ], + "type": "text", + "content": " Yuanzhouhan Cao" + }, + { + "bbox": [ + 68, + 160, + 545, + 217 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 68, + 160, + 545, + 217 + ], + "type": "text", + "content": " Linbo Qing" + }, + { + "bbox": [ + 68, + 160, + 545, + 217 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 68, + 160, + 545, + 217 + ], + "type": "text", + "content": " Chao Ren" + }, + { + "bbox": [ + 68, + 160, + 545, + 217 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 68, + 160, + 545, + 217 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 68, + 160, + 545, + 217 + ], + "type": "text", + "content": "Sichuan University, Chengdu, China \n" + }, + { + "bbox": [ + 68, + 160, + 545, + 217 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 68, + 160, + 545, + 217 + ], + "type": "text", + "content": "Beijing Jiaotong University, Beijing, China \n{dongguanglu, zhengtianheng}@stu.scu.edu.cn, yzhcao@bjtu.edu.cn \n{qing.lb, chaoren}@scu.edu.cn" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 151, + 244, + 200, + 257 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 244, + 200, + 257 + ], + "spans": [ + { + "bbox": [ + 151, + 244, + 200, + 257 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 270, + 297, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 270, + 297, + 594 + ], + "spans": [ + { + "bbox": [ + 54, + 270, + 297, + 594 + ], + "type": "text", + "content": "Recently, deep image deraining models based on paired datasets have made a series of remarkable progress. However, they cannot be well applied in real-world applications due to the difficulty of obtaining real paired datasets and the poor generalization performance. In this paper, we propose a novel Channel Consistency Prior and Self-Reconstruction Strategy Based Unsupervised Image Deraining framework, CSUD, to tackle the aforementioned challenges. During training with unpaired data, CSUD is capable of generating high-quality pseudo clean and rainy image pairs which are used to enhance the performance of deraining network. Specifically, to preserve more image background details while transferring rain streaks from rainy images to the unpaired clean images, we propose a novel Channel Consistency Loss (CCLoss) by introducing the Channel Consistency Prior (CCP) of rain streaks into training process, thereby ensuring that the generated pseudo rainy images closely resemble the real ones. Furthermore, we propose a novel Self-Reconstruction (SR) strategy to alleviate the redundant information transfer problem of the generator, further improving the deraining performance and the generalization capability of our method. Extensive experiments on multiple synthetic and real-world datasets demonstrate that the deraining performance of CSUD surpasses other state-of-the-art unsupervised methods and CSUD exhibits superior generalization capability. Code is available at https://github.com/GuangluDong0728/CSUD." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 616, + 135, + 628 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 616, + 135, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 616, + 135, + 628 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 635, + 296, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 635, + 296, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 635, + 296, + 696 + ], + "type": "text", + "content": "Image deraining is an important task in low-level computer vision [7, 10, 25, 29, 31, 49]. It aims to reconstruct a clean image from its rainy version, which has a profound impact on subsequent high-level tasks such as object detection, recognition, and semantic segmentation" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 315, + 243, + 555, + 430 + ], + "blocks": [ + { + "bbox": [ + 315, + 243, + 555, + 430 + ], + "lines": [ + { + "bbox": [ + 315, + 243, + 555, + 430 + ], + "spans": [ + { + "bbox": [ + 315, + 243, + 555, + 430 + ], + "type": "image", + "image_path": "91b3619f60909abcf4ab764b2b6370bcc5d2ec68f025545256d4d7d1fcc1fb6f.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 433, + 556, + 489 + ], + "lines": [ + { + "bbox": [ + 313, + 433, + 556, + 489 + ], + "spans": [ + { + "bbox": [ + 313, + 433, + 556, + 489 + ], + "type": "text", + "content": "Figure 1. Deraining results on the real rainy images captured by ourselves in real-world scenarios. Compared with the supervised method Restormer [49] and the unsupervised method DerainCycleGAN [40], our CSUD exhibits extremely strong generalization capability in the real world and achieves the best visual results." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 510, + 556, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 556, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 556, + 653 + ], + "type": "text", + "content": "[24, 26, 33, 41]. To mitigate the negative impact of rain, numerous deep learning based supervised image deraining methods [7, 8, 11, 13, 19, 21, 32, 36, 47, 52] have been proposed and they achieve remarkable progress on synthetic datasets. However, the performance of these methods is largely constrained by the difficulty of capturing a large number of high-quality paired datasets when applied to real-world scenarios. Additionally, attempting to directly apply supervised models trained on synthetic datasets to real-world scenarios is also difficult because of the huge domain gap between synthetic and real-world rain [45] and the overfitting issue of supervised methods [9]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "text", + "content": "Compared to supervised methods, unsupervised deep learning methods [3, 14, 40, 42, 46] for image deraining tasks generally achieve end-to-end image deraining without paired clean-rainy images during training. This allows unsupervised methods to effectively handle complex real-world" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 703, + 143, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 143, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 143, + 712 + ], + "type": "text", + "content": "*Corresponding author" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7469" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 167 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 167 + ], + "type": "text", + "content": "scenarios, especially when data collection costs are high. However, due to the high difficulty and challenges of unsupervised deraining methods in training process and the lack of prior knowledge [40], there is still very little relevant research in this field. Additionally, the deraining performance of existing unsupervised methods always lags far behind that of supervised methods, and they face the same generalization issues as supervised methods." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 171, + 297, + 290 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 171, + 297, + 290 + ], + "spans": [ + { + "bbox": [ + 55, + 171, + 297, + 290 + ], + "type": "text", + "content": "As shown in Fig. 1, even the SOTA supervised and unsupervised deraining models [40, 49] encounter substantial generalization issues and struggle to achieve satisfactory results when dealing with rainy images captured in real world, and sometimes they even fail to remove any rain streaks due to the more complex rain distributions in real world. Based on the above analysis, how to design an effective unsupervised deraining framework to avoid the need for real paired data and how to improve the generalization performance of deraining methods have become key research issues." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 293, + 296, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 293, + 296, + 471 + ], + "spans": [ + { + "bbox": [ + 55, + 293, + 296, + 471 + ], + "type": "text", + "content": "In this paper, we propose a novel channel consistency prior and self-reconstruction strategy based unsupervised image deraining framework which only needs unpaired data to train the network and can be easily generalized to real-world scenarios. We first conduct detailed analysis of the channel consistency prior (CCP) of rain streaks, and then propose a novel channel consistency loss to constrain the generator to synthesize high-quality pseudo rainy images. In addition, we utilize the proposed self-reconstruction (SR) strategy to further improve the performance of the generator and enhance the generalization capability of the derainer. As shown in Figure 1, our CSUD achieves the most promising performance on the real-world captured images compared to both supervised and unsupervised deraining methods. The main contributions of this work are summarized below:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 474, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 55, + 474, + 296, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 474, + 296, + 557 + ], + "spans": [ + { + "bbox": [ + 55, + 474, + 296, + 557 + ], + "type": "text", + "content": "- We propose a novel channel consistency prior and self-reconstruction strategy based unsupervised image deraining framework which is able to generate high-quality pseudo paired data for training. This framework can effectively address the issues of insufficient paired data in real world and significantly improve the generalization capability of the network." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 558, + 296, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 296, + 641 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 296, + 641 + ], + "type": "text", + "content": "- We propose a novel channel consistency loss (CCLoss) according to the channel consistency prior (CCP) of rain streaks to enhance the performance of the generator, preserving more background details while generating pseudo rainy images. To the best of our knowledge, we are the first to incorporate CCP into the training process of unsupervised deraining framework." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "type": "text", + "content": "- We propose a self-reconstruction (SR) strategy to solve the redundant information transfer problem of the generator. SR strategy can facilitate the accuracy of rain streaks transfer in the generator and enhance the deraining performance and the generalization capability of the derainer without altering the existing network structure." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 72, + 556, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 556, + 133 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 556, + 133 + ], + "type": "text", + "content": "- Extensive experiments on numerous benchmark datasets demonstrate that our proposed unsupervised framework can effectively tackle the deraining tasks on both synthetic and real-world rainy images with extremely superior generalization capability in real-world scenarios." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 144, + 406, + 157 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 144, + 406, + 157 + ], + "spans": [ + { + "bbox": [ + 314, + 144, + 406, + 157 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 165, + 550, + 178 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 165, + 550, + 178 + ], + "spans": [ + { + "bbox": [ + 313, + 165, + 550, + 178 + ], + "type": "text", + "content": "2.1. Deep Learning Based Single Image Deraining" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 182, + 555, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 182, + 555, + 386 + ], + "spans": [ + { + "bbox": [ + 313, + 182, + 555, + 386 + ], + "type": "text", + "content": "Supervised Single Image Deraining. Recently, with the development of deep learning, numerous excellent single image deraining [4, 5, 7, 8, 13, 21, 36, 47, 52] models have emerged. Fu et al. [7] first employ the CNN-based DerainNet to extract and remove the rain layer. Lin et al. [21] propose a generative network equipped with the pixel-wise status estimation and information decoupling for deraining. Lin et al. [23] propose a unified network for low-light-rainy image restoration to jointly handle low-light enhancement and deraining tasks. In addition to the aforementioned works specifically targeting deraining, some networks designed for image restoration [1, 16, 27, 29, 35, 38, 48, 49] have also shown excellent performance in deraining tasks. However, the above fully supervised methods require a large number of high-quality paired training samples, which is almost impossible in real-world scenarios. Simultaneously, due to the overfitting issue, their generalization capability is limited." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 387, + 556, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 387, + 556, + 602 + ], + "spans": [ + { + "bbox": [ + 313, + 387, + 556, + 602 + ], + "type": "text", + "content": "Unsupervised Single Image Deraining. To address the lack of paired datasets in the real world, some outstanding unsupervised deraining methods [3, 14, 40, 42, 46] have been proposed. Wei et al. [40] propose an unsupervised method for rain removal and generation by utilizing the popular CycleGAN [53] framework. Then, through introducing a dual contrastive learning manner to the CycleGAN framework, Chen et al. [3] develop an unpaired adversarial framework which effectively explores mutual properties of the unpaired rainy-clean samples. Ye et al. [46] introduce a novel non-local contrastive learning based unsupervised image deraining method to better distinguish rain streaks from clean images. However, the deraining performance of these unsupervised methods is far inferior to that of supervised methods, especially on real-world scenarios. Compared with previous methods, our proposed CSUD can greatly enhance the unsupervised deraining performance and achieve competitive results compared with some supervised methods." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 612, + 542, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 612, + 542, + 624 + ], + "spans": [ + { + "bbox": [ + 313, + 612, + 542, + 624 + ], + "type": "text", + "content": "2.2. Generalization Problem in Image Deraining" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 630, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 556, + 713 + ], + "type": "text", + "content": "Deep image deraining models often fail to obtain satisfactory performance or even cannot remove any rain streaks when they are applied to real-world scenarios. Generalization problem of deraining models is an urgent issue to be solved. Gu et al. [9] conduct thorough analysis of the generalization problem in image deraining and propose to achieve better generalization by simplifying the complexity of train-" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7470" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 71, + 555, + 300 + ], + "blocks": [ + { + "bbox": [ + 56, + 71, + 555, + 300 + ], + "lines": [ + { + "bbox": [ + 56, + 71, + 555, + 300 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 555, + 300 + ], + "type": "image", + "image_path": "81aba27eb357ed84b6880f74088d87ca8afe2c9ffe2db10b415cffc043c55093.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 303, + 555, + 336 + ], + "lines": [ + { + "bbox": [ + 55, + 303, + 555, + 336 + ], + "spans": [ + { + "bbox": [ + 55, + 303, + 555, + 336 + ], + "type": "text", + "content": "Figure 2. Framework of the proposed CSUD. The red arrows represent the baseline, purple arrows represent the 3 extra constraints, green arrows represent the self-reconstruction (SR) strategy, and blue arrows represent the CCLoss. The 4 input arrows of RIEM represent 4 separate runs of the generator, each time guided by a different image. Both RIEM and CFEM receive only 1 input image at a time." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 357, + 297, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 357, + 297, + 502 + ], + "spans": [ + { + "bbox": [ + 54, + 357, + 297, + 502 + ], + "type": "text", + "content": "ing images. Some semi-supervised learning frameworks [39, 44] are proposed to analyze the residual difference between domains and enhance the deraining capability in real world. In addition, a series of unsupervised deraining methods [3, 40, 46] have also demonstrated a certain degree of generalization capability in real-world scenarios. However, they have not conducted detailed analysis on how to enhance the generalization capability of unsupervised methods. Inspired by [9], we propose a self-reconstruction strategy to make both generator and derainer pay more attention to the reconstruction of image background, which significantly improves the generalization performance of our method." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 516, + 138, + 530 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 516, + 138, + 530 + ], + "spans": [ + { + "bbox": [ + 55, + 516, + 138, + 530 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 537, + 269, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 537, + 269, + 550 + ], + "spans": [ + { + "bbox": [ + 55, + 537, + 269, + 550 + ], + "type": "text", + "content": "3.1. Overall Framework of Proposed Method" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 556, + 297, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 556, + 297, + 676 + ], + "spans": [ + { + "bbox": [ + 55, + 556, + 297, + 676 + ], + "type": "text", + "content": "To address the issue of insufficient paired data and the generalization problem in deraining applications, we propose a novel channel consistency prior and self-reconstruction strategy based unsupervised image deraining framework, CSUD. As illustrated in Fig. 2, our CSUD consists of a derainer " + }, + { + "bbox": [ + 55, + 556, + 297, + 676 + ], + "type": "inline_equation", + "content": "Der" + }, + { + "bbox": [ + 55, + 556, + 297, + 676 + ], + "type": "text", + "content": ", a rainy image generator " + }, + { + "bbox": [ + 55, + 556, + 297, + 676 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 55, + 556, + 297, + 676 + ], + "type": "text", + "content": ", and a Patch-GAN [12] discriminator " + }, + { + "bbox": [ + 55, + 556, + 297, + 676 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 55, + 556, + 297, + 676 + ], + "type": "text", + "content": ". For " + }, + { + "bbox": [ + 55, + 556, + 297, + 676 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 55, + 556, + 297, + 676 + ], + "type": "text", + "content": ", it consists of a clean feature extraction module (CFEM), a rain information extraction module (RIEM) and 6 residual blocks. Detailed network structures are detailed in the Appendix." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": "In generator " + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": ", we effectively utilize unpaired clean and rainy images to synthesize pseudo rainy images. As shown by the red arrows in Fig. 2, we design a GAN-based baseline" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 357, + 555, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 357, + 555, + 489 + ], + "spans": [ + { + "bbox": [ + 313, + 357, + 555, + 489 + ], + "type": "text", + "content": "that transitions from \"" + }, + { + "bbox": [ + 313, + 357, + 555, + 489 + ], + "type": "inline_equation", + "content": "x \\in X \\to x_{s1} \\in S \\to x_{der} \\in N" + }, + { + "bbox": [ + 313, + 357, + 555, + 489 + ], + "type": "text", + "content": "\", where " + }, + { + "bbox": [ + 313, + 357, + 555, + 489 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 313, + 357, + 555, + 489 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 357, + 555, + 489 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 313, + 357, + 555, + 489 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 357, + 555, + 489 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 357, + 555, + 489 + ], + "type": "text", + "content": " denote clean images, synthetic rainy images and derained images respectively. " + }, + { + "bbox": [ + 313, + 357, + 555, + 489 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 313, + 357, + 555, + 489 + ], + "type": "text", + "content": " transforms clean image " + }, + { + "bbox": [ + 313, + 357, + 555, + 489 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 357, + 555, + 489 + ], + "type": "text", + "content": " to synthetic rainy image " + }, + { + "bbox": [ + 313, + 357, + 555, + 489 + ], + "type": "inline_equation", + "content": "x_{s1}" + }, + { + "bbox": [ + 313, + 357, + 555, + 489 + ], + "type": "text", + "content": " according to the guide of real rainy image " + }, + { + "bbox": [ + 313, + 357, + 555, + 489 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 313, + 357, + 555, + 489 + ], + "type": "text", + "content": ". In order to better extract rain information and transfer it to the clean images, " + }, + { + "bbox": [ + 313, + 357, + 555, + 489 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 313, + 357, + 555, + 489 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 357, + 555, + 489 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 357, + 555, + 489 + ], + "type": "text", + "content": " are initially processed by the RIEM and the CFEM respectively to obtain rainy features and clean features which are concatenated for subsequent process. For the real rainy image " + }, + { + "bbox": [ + 313, + 357, + 555, + 489 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 313, + 357, + 555, + 489 + ], + "type": "text", + "content": " and the synthesized rainy image " + }, + { + "bbox": [ + 313, + 357, + 555, + 489 + ], + "type": "inline_equation", + "content": "x_{s1}" + }, + { + "bbox": [ + 313, + 357, + 555, + 489 + ], + "type": "text", + "content": ", the corresponding derained clean images " + }, + { + "bbox": [ + 313, + 357, + 555, + 489 + ], + "type": "inline_equation", + "content": "y_{der}" + }, + { + "bbox": [ + 313, + 357, + 555, + 489 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 357, + 555, + 489 + ], + "type": "inline_equation", + "content": "x_{der}" + }, + { + "bbox": [ + 313, + 357, + 555, + 489 + ], + "type": "text", + "content": " are obtained through Der." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 489, + 556, + 596 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 489, + 556, + 596 + ], + "spans": [ + { + "bbox": [ + 313, + 489, + 556, + 596 + ], + "type": "text", + "content": "The discriminator " + }, + { + "bbox": [ + 313, + 489, + 556, + 596 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 313, + 489, + 556, + 596 + ], + "type": "text", + "content": " is trained to distinguish between synthetic rainy images " + }, + { + "bbox": [ + 313, + 489, + 556, + 596 + ], + "type": "inline_equation", + "content": "x_{s1} \\in S" + }, + { + "bbox": [ + 313, + 489, + 556, + 596 + ], + "type": "text", + "content": " and real rainy images " + }, + { + "bbox": [ + 313, + 489, + 556, + 596 + ], + "type": "inline_equation", + "content": "y \\in Y" + }, + { + "bbox": [ + 313, + 489, + 556, + 596 + ], + "type": "text", + "content": ", and the adversarial loss " + }, + { + "bbox": [ + 313, + 489, + 556, + 596 + ], + "type": "inline_equation", + "content": "L_{adv1}" + }, + { + "bbox": [ + 313, + 489, + 556, + 596 + ], + "type": "text", + "content": " of the baseline is constrained between them. Furthermore, motivated by existing methods [22, 53], as indicated by the purple arrows in Fig. 2, we add another three processes, " + }, + { + "bbox": [ + 313, + 489, + 556, + 596 + ], + "type": "inline_equation", + "content": "G(Der(x_{s1}), x_{s1}) \\to x_{s2}" + }, + { + "bbox": [ + 313, + 489, + 556, + 596 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 489, + 556, + 596 + ], + "type": "inline_equation", + "content": "G(Der(y), y) \\to y_{s1}" + }, + { + "bbox": [ + 313, + 489, + 556, + 596 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 489, + 556, + 596 + ], + "type": "inline_equation", + "content": "G(Der(y), x_{s1}) \\to y_{s2}" + }, + { + "bbox": [ + 313, + 489, + 556, + 596 + ], + "type": "text", + "content": ", to the framework to enhance training stability. Then, additional three adversarial losses are introduced to CSUD:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 344, + 604, + 555, + 617 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 344, + 604, + 555, + 617 + ], + "spans": [ + { + "bbox": [ + 344, + 604, + 555, + 617 + ], + "type": "interline_equation", + "content": "L _ {G A N} = L _ {a d v 1} + L _ {a d v 2} + L _ {a d v 3} + L _ {a d v 4}. \\tag {1}", + "image_path": "ec2f1ad3d40ce6a5d7ad0ccdae91a33f370cac065d08bba2eceb02f347af9180.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 624, + 528, + 636 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 624, + 528, + 636 + ], + "spans": [ + { + "bbox": [ + 313, + 624, + 528, + 636 + ], + "type": "text", + "content": "3.2. Channel Consistency Loss for Generator" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 641, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 556, + 714 + ], + "type": "text", + "content": "Channel Consistency Prior (CCP) of Rain Streaks. In RGB rainy images, most rain streaks tend to be consistent across the R, G, and B channels, meaning the pixel values of most rain streaks in the three channels are highly similar. As shown in the second, third, and fourth columns of Fig. 3 (a), we separate both synthetic and real-world rainy images" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "7471" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 68, + 553, + 285 + ], + "blocks": [ + { + "bbox": [ + 56, + 68, + 553, + 285 + ], + "lines": [ + { + "bbox": [ + 56, + 68, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 56, + 68, + 553, + 285 + ], + "type": "image", + "image_path": "e4c62cb2741d1f3dc28cd2a762e51ef3355402c6d05a920b289d8b2531e9b54e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 118, + 289, + 493, + 300 + ], + "lines": [ + { + "bbox": [ + 118, + 289, + 493, + 300 + ], + "spans": [ + { + "bbox": [ + 118, + 289, + 493, + 300 + ], + "type": "text", + "content": "(a) Visualization of channel consistency prior in synthetic, real-world, and nighttime real-world rainy images" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 56, + 303, + 552, + 376 + ], + "blocks": [ + { + "bbox": [ + 56, + 303, + 552, + 376 + ], + "lines": [ + { + "bbox": [ + 56, + 303, + 552, + 376 + ], + "spans": [ + { + "bbox": [ + 56, + 303, + 552, + 376 + ], + "type": "image", + "image_path": "491b20f20ed41772d50a06a006d3b60238cd9753a4e6ce065f6e7a2f8075a132.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 123, + 380, + 487, + 390 + ], + "lines": [ + { + "bbox": [ + 123, + 380, + 487, + 390 + ], + "spans": [ + { + "bbox": [ + 123, + 380, + 487, + 390 + ], + "type": "text", + "content": "(b) Cosine similarities of cycle subtractions among R, G, and B channels between clean and rainy images" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 54, + 392, + 555, + 437 + ], + "lines": [ + { + "bbox": [ + 54, + 392, + 555, + 437 + ], + "spans": [ + { + "bbox": [ + 54, + 392, + 555, + 437 + ], + "type": "text", + "content": "Figure 3. (a) Visualization of channel consistency prior in synthetic and real-world rainy images. The first column presents the rainy RGB images; the second, third and fourth columns present their R, G, B channels, respectively; the fifth, sixth, and seventh columns present the cycle subtractions of R, G, and B channels, respectively. (b) Cosine similarities of the cycle subtractions of R, G, and B channels between clean and rainy images in the 5 synthetic and real-world datasets. More visualizations and analysis are presented in the Appendix." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 458, + 295, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 458, + 295, + 518 + ], + "spans": [ + { + "bbox": [ + 54, + 458, + 295, + 518 + ], + "type": "text", + "content": "into R, G, and B channels and visualized them. It can be seen clearly that while there are significant differences in the background parts among the three channels, the rain streaks tend to be consistent. We propose to term this phenomenon as \"Channel Consistency Prior (CCP)\" of rainy streaks." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 522, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 522, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 522, + 297, + 714 + ], + "type": "text", + "content": "Theoretically, according to CCP, the cycle subtractions " + }, + { + "bbox": [ + 55, + 522, + 297, + 714 + ], + "type": "inline_equation", + "content": "((r)_R - (r)_G)" + }, + { + "bbox": [ + 55, + 522, + 297, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 522, + 297, + 714 + ], + "type": "inline_equation", + "content": "((r)_G - (r)_B)" + }, + { + "bbox": [ + 55, + 522, + 297, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 522, + 297, + 714 + ], + "type": "inline_equation", + "content": "((r)_B - (r)_R)" + }, + { + "bbox": [ + 55, + 522, + 297, + 714 + ], + "type": "text", + "content": " of R, G, and B channels in rainy images will almost contain no rain streaks and should be close to " + }, + { + "bbox": [ + 55, + 522, + 297, + 714 + ], + "type": "inline_equation", + "content": "((n)_R - (n)_G)" + }, + { + "bbox": [ + 55, + 522, + 297, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 522, + 297, + 714 + ], + "type": "inline_equation", + "content": "((n)_G - (n)_B)" + }, + { + "bbox": [ + 55, + 522, + 297, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 522, + 297, + 714 + ], + "type": "inline_equation", + "content": "((n)_B - (n)_R)" + }, + { + "bbox": [ + 55, + 522, + 297, + 714 + ], + "type": "text", + "content": " in clean images, where " + }, + { + "bbox": [ + 55, + 522, + 297, + 714 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 55, + 522, + 297, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 522, + 297, + 714 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 522, + 297, + 714 + ], + "type": "text", + "content": " represent the rainy and clean image, " + }, + { + "bbox": [ + 55, + 522, + 297, + 714 + ], + "type": "inline_equation", + "content": "(\\cdot)_R" + }, + { + "bbox": [ + 55, + 522, + 297, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 522, + 297, + 714 + ], + "type": "inline_equation", + "content": "(\\cdot)_G" + }, + { + "bbox": [ + 55, + 522, + 297, + 714 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 522, + 297, + 714 + ], + "type": "inline_equation", + "content": "(\\cdot)_B" + }, + { + "bbox": [ + 55, + 522, + 297, + 714 + ], + "type": "text", + "content": " denote R, G, and B channels respectively. In the fifth, sixth, and seventh columns of Fig. 3 (a), it can be observed that the cycle subtractions of R, G, and B channels nearly don't contain any rain streaks, qualitatively confirming the correctness of CCP. Furthermore, as shown in Fig. 3 (b), we quantitatively present the cosine similarities of the cycle subtractions of R, G, and B channels between clean and rainy images in both synthetic and real-world benchmark datasets [18, 20, 36, 43]. It is evident that the cosine similarities are exceptionally high, which can further demonstrate our hypothesis." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 458, + 555, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 458, + 555, + 542 + ], + "spans": [ + { + "bbox": [ + 313, + 458, + 555, + 542 + ], + "type": "text", + "content": "Channel Consistency Loss. Based on the above analysis of CCP, the cycle subtractions of R, G, and B channels between input clean image and output pseudo rainy image of the generator " + }, + { + "bbox": [ + 313, + 458, + 555, + 542 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 313, + 458, + 555, + 542 + ], + "type": "text", + "content": " should tend to be consistent. To this end, we propose a novel Channel Consistency Loss (CCLoss) to preserve more color and texture details of the background while generating pseudo rainy images:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 338, + 546, + 555, + 590 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 338, + 546, + 555, + 590 + ], + "spans": [ + { + "bbox": [ + 338, + 546, + 555, + 590 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} L _ {C C} = \\left\\| \\left(\\left(x _ {s 1}\\right) _ {R} - \\left(x _ {s 1}\\right) _ {G}\\right), \\left(x _ {R} - x _ {G}\\right) \\right\\| _ {1} \\\\ + \\left\\| \\left(\\left(x _ {s 1}\\right) _ {G} - \\left(x _ {s 1}\\right) _ {B}\\right), \\left(x _ {G} - x _ {B}\\right) \\right\\| _ {1} \\tag {2} \\\\ + \\left\\| \\left(\\left(x _ {s 1}\\right) _ {B} - \\left(x _ {s 1}\\right) _ {R}\\right), \\left(x _ {B} - x _ {R}\\right) \\right\\| _ {1}, \\\\ \\end{array}", + "image_path": "a2502c90854b63abb5a0f159e39aa219e277299cf7db6d3c9347a886e6c357dc.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 594, + 556, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 556, + 667 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 556, + 667 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 594, + 556, + 667 + ], + "type": "inline_equation", + "content": "(x_{s1})_R, (x_{s1})_G, (x_{s1})_B, x_R, x_G, x_B" + }, + { + "bbox": [ + 313, + 594, + 556, + 667 + ], + "type": "text", + "content": " denotes R, G, B color channels of " + }, + { + "bbox": [ + 313, + 594, + 556, + 667 + ], + "type": "inline_equation", + "content": "x_{s1}" + }, + { + "bbox": [ + 313, + 594, + 556, + 667 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 594, + 556, + 667 + ], + "type": "inline_equation", + "content": "x, ||\\cdot||_1" + }, + { + "bbox": [ + 313, + 594, + 556, + 667 + ], + "type": "text", + "content": " represents the " + }, + { + "bbox": [ + 313, + 594, + 556, + 667 + ], + "type": "inline_equation", + "content": "L_1" + }, + { + "bbox": [ + 313, + 594, + 556, + 667 + ], + "type": "text", + "content": " norm and " + }, + { + "bbox": [ + 313, + 594, + 556, + 667 + ], + "type": "inline_equation", + "content": "L_{CC}" + }, + { + "bbox": [ + 313, + 594, + 556, + 667 + ], + "type": "text", + "content": " represents the proposed CCLoss, respectively. By introducing CCP into our unsupervised deraining framework, " + }, + { + "bbox": [ + 313, + 594, + 556, + 667 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 313, + 594, + 556, + 667 + ], + "type": "text", + "content": " will generate more accurate synthesized rainy images, which will further improve the performance of the derainer." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 672, + 471, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 672, + 471, + 685 + ], + "spans": [ + { + "bbox": [ + 313, + 672, + 471, + 685 + ], + "type": "text", + "content": "3.3. Self-Reconstruction Strategy" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 689, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 556, + 714 + ], + "type": "text", + "content": "Redundant Information Transfer Problem of the Generator. Although adversarial losses can effectively constrain the" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7472" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 69, + 298, + 190 + ], + "blocks": [ + { + "bbox": [ + 57, + 69, + 298, + 190 + ], + "lines": [ + { + "bbox": [ + 57, + 69, + 298, + 190 + ], + "spans": [ + { + "bbox": [ + 57, + 69, + 298, + 190 + ], + "type": "image", + "image_path": "7caee91250ebf886ef717848f36f2812849980b90b8299834d7a96c24fd3f045.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 190, + 294, + 201 + ], + "lines": [ + { + "bbox": [ + 55, + 190, + 294, + 201 + ], + "spans": [ + { + "bbox": [ + 55, + 190, + 294, + 201 + ], + "type": "text", + "content": "Figure 4. Visualization of redundant information transfer problem." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 220, + 296, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 220, + 296, + 327 + ], + "spans": [ + { + "bbox": [ + 54, + 220, + 296, + 327 + ], + "type": "text", + "content": "training process of generator " + }, + { + "bbox": [ + 54, + 220, + 296, + 327 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 54, + 220, + 296, + 327 + ], + "type": "text", + "content": ", the data synthesis process of " + }, + { + "bbox": [ + 54, + 220, + 296, + 327 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 54, + 220, + 296, + 327 + ], + "type": "text", + "content": " is highly susceptible to interference from the redundant information in rainy images, especially those edge textures that are very similar to rain streaks. As shown in Fig. 4, the edge textures of some background objects in rainy images, such as the outline of the woman's hair, the leg of rhino, etc., are mistakenly transferred to the clean images by " + }, + { + "bbox": [ + 54, + 220, + 296, + 327 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 54, + 220, + 296, + 327 + ], + "type": "text", + "content": ". As a result, " + }, + { + "bbox": [ + 54, + 220, + 296, + 327 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 54, + 220, + 296, + 327 + ], + "type": "text", + "content": " will generate inaccurate pseudo rainy images, interfering with the learning process of " + }, + { + "bbox": [ + 54, + 220, + 296, + 327 + ], + "type": "inline_equation", + "content": "Der" + }, + { + "bbox": [ + 54, + 220, + 296, + 327 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 328, + 296, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 328, + 296, + 533 + ], + "spans": [ + { + "bbox": [ + 55, + 328, + 296, + 533 + ], + "type": "text", + "content": "Self-Reconstruction Strategy. To solve the redundant information transfer problem, we propose a novel Self-Reconstruction (SR) strategy, which enables " + }, + { + "bbox": [ + 55, + 328, + 296, + 533 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 55, + 328, + 296, + 533 + ], + "type": "text", + "content": " to extract more accurate rain streak information and synthesizes more realistic pseudo rainy images. The core idea of SR strategy is to encourage " + }, + { + "bbox": [ + 55, + 328, + 296, + 533 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 55, + 328, + 296, + 533 + ], + "type": "text", + "content": " to better distinguish rain streaks from background in the rainy images, ensuring that only rain streak information is extracted for subsequent generation of pseudo rainy images, while redundant information of background is disregarded. As shown by the green arrows in Fig. 2, when the input clean image " + }, + { + "bbox": [ + 55, + 328, + 296, + 533 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 328, + 296, + 533 + ], + "type": "text", + "content": " itself, the derained image " + }, + { + "bbox": [ + 55, + 328, + 296, + 533 + ], + "type": "inline_equation", + "content": "y_{der}" + }, + { + "bbox": [ + 55, + 328, + 296, + 533 + ], + "type": "text", + "content": " are input with clean image " + }, + { + "bbox": [ + 55, + 328, + 296, + 533 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 328, + 296, + 533 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 55, + 328, + 296, + 533 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 55, + 328, + 296, + 533 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 328, + 296, + 533 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 55, + 328, + 296, + 533 + ], + "type": "text", + "content": " learns to reconstruct the input clean image " + }, + { + "bbox": [ + 55, + 328, + 296, + 533 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 328, + 296, + 533 + ], + "type": "text", + "content": " itself without any additional redundant information from " + }, + { + "bbox": [ + 55, + 328, + 296, + 533 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 328, + 296, + 533 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 328, + 296, + 533 + ], + "type": "inline_equation", + "content": "y_{der}" + }, + { + "bbox": [ + 55, + 328, + 296, + 533 + ], + "type": "text", + "content": ", which means that the self-reconstructed images " + }, + { + "bbox": [ + 55, + 328, + 296, + 533 + ], + "type": "inline_equation", + "content": "x_{s1}" + }, + { + "bbox": [ + 55, + 328, + 296, + 533 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 328, + 296, + 533 + ], + "type": "inline_equation", + "content": "x_{s2}" + }, + { + "bbox": [ + 55, + 328, + 296, + 533 + ], + "type": "text", + "content": " should be close to " + }, + { + "bbox": [ + 55, + 328, + 296, + 533 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 328, + 296, + 533 + ], + "type": "text", + "content": ". Thus, we can utilize the following self-reconstruction loss (SRLoss) as additional training constraint of " + }, + { + "bbox": [ + 55, + 328, + 296, + 533 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 55, + 328, + 296, + 533 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 80, + 540, + 296, + 554 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 540, + 296, + 554 + ], + "spans": [ + { + "bbox": [ + 80, + 540, + 296, + 554 + ], + "type": "interline_equation", + "content": "L _ {S R - G} = \\left\\| G (x, x), x \\right\\| _ {1} + \\left\\| G (x, y _ {d e r}), x \\right\\| _ {1}, \\tag {3}", + "image_path": "9fae11eb3585d56de3588cd165abb622444cc65cf7d7d4d38ec35638aa319637.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 560, + 296, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 560, + 296, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 560, + 296, + 693 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 560, + 296, + 693 + ], + "type": "inline_equation", + "content": "L_{SR - G}" + }, + { + "bbox": [ + 55, + 560, + 296, + 693 + ], + "type": "text", + "content": " represents the SRLoss for " + }, + { + "bbox": [ + 55, + 560, + 296, + 693 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 55, + 560, + 296, + 693 + ], + "type": "text", + "content": ". SR strategy allows " + }, + { + "bbox": [ + 55, + 560, + 296, + 693 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 55, + 560, + 296, + 693 + ], + "type": "text", + "content": " to pay more attention to reconstructing the background of clean images while learning the transfer of rain streaks, significantly improving the accuracy of generating pseudo rainy images. Under the constraint of " + }, + { + "bbox": [ + 55, + 560, + 296, + 693 + ], + "type": "inline_equation", + "content": "L_{SR - G}" + }, + { + "bbox": [ + 55, + 560, + 296, + 693 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 560, + 296, + 693 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 55, + 560, + 296, + 693 + ], + "type": "text", + "content": " will obtain the ability to generate clean images themselves guided by images without rain streaks. Therefore, the cleaner the derained images " + }, + { + "bbox": [ + 55, + 560, + 296, + 693 + ], + "type": "inline_equation", + "content": "y_{der}" + }, + { + "bbox": [ + 55, + 560, + 296, + 693 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 560, + 296, + 693 + ], + "type": "inline_equation", + "content": "x_{der}" + }, + { + "bbox": [ + 55, + 560, + 296, + 693 + ], + "type": "text", + "content": " restored by " + }, + { + "bbox": [ + 55, + 560, + 296, + 693 + ], + "type": "inline_equation", + "content": "Der" + }, + { + "bbox": [ + 55, + 560, + 296, + 693 + ], + "type": "text", + "content": " are, the closer the self-reconstructed images " + }, + { + "bbox": [ + 55, + 560, + 296, + 693 + ], + "type": "inline_equation", + "content": "x_{s2}" + }, + { + "bbox": [ + 55, + 560, + 296, + 693 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 560, + 296, + 693 + ], + "type": "inline_equation", + "content": "x_{s3}" + }, + { + "bbox": [ + 55, + 560, + 296, + 693 + ], + "type": "text", + "content": " are to the original clean ones " + }, + { + "bbox": [ + 55, + 560, + 296, + 693 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 560, + 296, + 693 + ], + "type": "text", + "content": ". Then, the following SRLoss can be utilized to further improve the performance of " + }, + { + "bbox": [ + 55, + 560, + 296, + 693 + ], + "type": "inline_equation", + "content": "Der" + }, + { + "bbox": [ + 55, + 560, + 296, + 693 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 64, + 700, + 296, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 700, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 64, + 700, + 296, + 715 + ], + "type": "interline_equation", + "content": "L _ {S R - D e r} = \\left\\| G (x, x _ {d e r}), x \\right\\| _ {1} + \\left\\| G (x, y _ {d e r}), x \\right\\| _ {1}, \\tag {4}", + "image_path": "1022bb06c1bb34588cf6cd807e81a9e7576785e0d7282df05670d30e62e4b885.jpg" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 316, + 70, + 554, + 194 + ], + "blocks": [ + { + "bbox": [ + 316, + 70, + 554, + 194 + ], + "lines": [ + { + "bbox": [ + 316, + 70, + 554, + 194 + ], + "spans": [ + { + "bbox": [ + 316, + 70, + 554, + 194 + ], + "type": "image", + "image_path": "c16c3a781b06600c2afbcb800eb781ba306111a89fd058d7ada04e9aab946cb7.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 331, + 198, + 535, + 209 + ], + "lines": [ + { + "bbox": [ + 331, + 198, + 535, + 209 + ], + "spans": [ + { + "bbox": [ + 331, + 198, + 535, + 209 + ], + "type": "text", + "content": "Figure 5. The implementation principles of SR strategy." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 231, + 555, + 278 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 231, + 555, + 278 + ], + "spans": [ + { + "bbox": [ + 313, + 231, + 555, + 278 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 231, + 555, + 278 + ], + "type": "inline_equation", + "content": "L_{SR - Der}" + }, + { + "bbox": [ + 313, + 231, + 555, + 278 + ], + "type": "text", + "content": " represents the SRLoss for " + }, + { + "bbox": [ + 313, + 231, + 555, + 278 + ], + "type": "inline_equation", + "content": "Der" + }, + { + "bbox": [ + 313, + 231, + 555, + 278 + ], + "type": "text", + "content": ". Under the constraint of " + }, + { + "bbox": [ + 313, + 231, + 555, + 278 + ], + "type": "inline_equation", + "content": "L_{SR - Der}" + }, + { + "bbox": [ + 313, + 231, + 555, + 278 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 231, + 555, + 278 + ], + "type": "inline_equation", + "content": "Der" + }, + { + "bbox": [ + 313, + 231, + 555, + 278 + ], + "type": "text", + "content": " will also focus more on reconstructing the background content while learning to derain, which greatly enhances the generalization capability of " + }, + { + "bbox": [ + 313, + 231, + 555, + 278 + ], + "type": "inline_equation", + "content": "Der" + }, + { + "bbox": [ + 313, + 231, + 555, + 278 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 279, + 555, + 340 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 279, + 555, + 340 + ], + "spans": [ + { + "bbox": [ + 313, + 279, + 555, + 340 + ], + "type": "text", + "content": "As shown in the third and fourth columns of Figure 4, SR strategy effectively alleviates the redundant information transfer problem. Through SR strategy, " + }, + { + "bbox": [ + 313, + 279, + 555, + 340 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 313, + 279, + 555, + 340 + ], + "type": "text", + "content": " can achieve robust rain streak transfer with more background detail fidelity, greatly enhancing the deraining performance of Der." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 348, + 403, + 360 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 348, + 403, + 360 + ], + "spans": [ + { + "bbox": [ + 313, + 348, + 403, + 360 + ], + "type": "text", + "content": "3.4. Loss Function" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 365, + 555, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 365, + 555, + 415 + ], + "spans": [ + { + "bbox": [ + 313, + 365, + 555, + 415 + ], + "type": "text", + "content": "Due to the high difficulty and instability of training a GAN-based unsupervised framework, in addition to adopting the L1 loss for " + }, + { + "bbox": [ + 313, + 365, + 555, + 415 + ], + "type": "inline_equation", + "content": "Der" + }, + { + "bbox": [ + 313, + 365, + 555, + 415 + ], + "type": "text", + "content": ", we also incorporate SSIM loss and perceptual loss[15] to enhance the performance of " + }, + { + "bbox": [ + 313, + 365, + 555, + 415 + ], + "type": "inline_equation", + "content": "Der" + }, + { + "bbox": [ + 313, + 365, + 555, + 415 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 338, + 424, + 554, + 453 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 338, + 424, + 554, + 453 + ], + "spans": [ + { + "bbox": [ + 338, + 424, + 554, + 453 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} L _ {D e r} = \\left\\| x, x _ {d e r} \\right\\| _ {1} + \\lambda_ {1} L _ {S S I M} (x, x _ {d e r}) \\tag {5} \\\\ + \\lambda_ {2} | | \\phi_ {i} (x), \\phi_ {i} (x _ {d e r}) | | _ {1} + \\lambda_ {3} L _ {S R - D e r}, \\\\ \\end{array}", + "image_path": "aac3430d5d4e6b5496876908f2c29993a4a63f89e8ab1bc573b81c9d60eb1ee3.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 462, + 555, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 462, + 555, + 510 + ], + "spans": [ + { + "bbox": [ + 313, + 462, + 555, + 510 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 462, + 555, + 510 + ], + "type": "inline_equation", + "content": "||\\cdot ||_1" + }, + { + "bbox": [ + 313, + 462, + 555, + 510 + ], + "type": "text", + "content": " represents the " + }, + { + "bbox": [ + 313, + 462, + 555, + 510 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 313, + 462, + 555, + 510 + ], + "type": "text", + "content": " norm, and " + }, + { + "bbox": [ + 313, + 462, + 555, + 510 + ], + "type": "inline_equation", + "content": "L_{SSIM}(\\cdot)" + }, + { + "bbox": [ + 313, + 462, + 555, + 510 + ], + "type": "text", + "content": " represents the SSIM loss, " + }, + { + "bbox": [ + 313, + 462, + 555, + 510 + ], + "type": "inline_equation", + "content": "\\phi_i(\\cdot)" + }, + { + "bbox": [ + 313, + 462, + 555, + 510 + ], + "type": "text", + "content": " is the layers of pre-trained VGG-19 [34], " + }, + { + "bbox": [ + 313, + 462, + 555, + 510 + ], + "type": "inline_equation", + "content": "\\lambda_1, \\lambda_2, \\lambda_3" + }, + { + "bbox": [ + 313, + 462, + 555, + 510 + ], + "type": "text", + "content": " represent the hyperparameters for SSIM loss, perceptual loss and SRLoss for derainer respectively." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 325, + 511, + 548, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 511, + 548, + 523 + ], + "spans": [ + { + "bbox": [ + 325, + 511, + 548, + 523 + ], + "type": "text", + "content": "Finally, we obtain the total loss to train our framework:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 314, + 533, + 555, + 559 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 533, + 555, + 559 + ], + "spans": [ + { + "bbox": [ + 314, + 533, + 555, + 559 + ], + "type": "interline_equation", + "content": "L _ {t o t a l} = \\min _ {G} \\max _ {D} L _ {G A N} + L _ {D e r} + \\alpha_ {1} L _ {C C} + \\alpha_ {2} L _ {S R - G}, \\tag {6}", + "image_path": "f75fe63d88941c5d0064bc0c1a7a9fc6cddc6395d64b862ad6d60e1c59490622.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 560, + 554, + 585 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 560, + 554, + 585 + ], + "spans": [ + { + "bbox": [ + 313, + 560, + 554, + 585 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 560, + 554, + 585 + ], + "type": "inline_equation", + "content": "\\alpha_{1},\\alpha_{2}" + }, + { + "bbox": [ + 313, + 560, + 554, + 585 + ], + "type": "text", + "content": " represent the hyperparameters for CCLoss and SRLoss for generator respectively." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 597, + 394, + 610 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 597, + 394, + 610 + ], + "spans": [ + { + "bbox": [ + 313, + 597, + 394, + 610 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 617, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 556, + 713 + ], + "type": "text", + "content": "To demonstrate the effectiveness of our CSUD, following [18, 40], we conduct experiments from the following two aspects: (1) unsupervised deraining performance, (2) generalization performance. In experiment (1), we train independent models for different datasets respectively, including both synthetic and real-world datasets. In experiment (2), we only train our model on the synthetic dataset Rain100L [43], and then test on various real-world and nighttime datasets." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7473" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 83, + 102, + 529, + 256 + ], + "blocks": [ + { + "bbox": [ + 55, + 71, + 555, + 94 + ], + "lines": [ + { + "bbox": [ + 55, + 71, + 555, + 94 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 555, + 94 + ], + "type": "text", + "content": "Table 1. Quantitative deraining performance comparisons of different methods. Red and blue colors indicate the 1st and 2nd ranks among the unsupervised methods, respectively. S and U indicate supervised method and unsupervised method, respectively." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 83, + 102, + 529, + 256 + ], + "lines": [ + { + "bbox": [ + 83, + 102, + 529, + 256 + ], + "spans": [ + { + "bbox": [ + 83, + 102, + 529, + 256 + ], + "type": "table", + "html": "
DatasetsRain100L [43]Rain100H [43]Rain12 [20]Rain800 [17]RealRain1K-L [18]RealRain1K-H [18]
MetricsPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIM
DerainNet [7] (S)27.030.88414.920.592//22.770.81027.090.92522.880.889
DDN [8] (S)32.380.92624.640.84934.040.93321.160.73231.180.91729.170.878
SPA-Net [36] (S)31.950.91926.070.857//24.370.86130.430.94725.760.910
MPRNet [48] (S)34.950.95928.530.87236.840.96428.680.87536.290.97234.740.964
NAFNet [2] (S)37.000.97829.660.90034.810.94328.720.88638.800.98636.110.976
Restormer [49] (S)37.570.97429.460.889//30.330.90540.900.98539.570.981
PromptIR [29] (S)38.340.98328.690.87735.090.94529.070.88736.990.97333.610.953
NeRD-Rain-S [5] (S)42.000.99032.860.93235.390.94228.820.87838.640.97936.690.970
CycleGAN [53] (U)24.610.83420.590.70421.560.84523.950.81920.190.82017.530.762
DerainCycleGAN [40] (U)31.490.93621.130.71033.520.94024.320.84228.160.90124.780.844
NLCL [46] (U)20.500.71922.310.72822.680.73518.940.67023.060.83220.640.719
DCD-GAN [3] (U)31.820.94122.470.75331.560.92425.610.81330.490.93927.820.895
CSUD (Ours)33.280.95424.420.80834.560.95123.580.79532.710.95929.670.928
", + "image_path": "a53ce33e3dabe10f4ae742cf45db599b43c94d051e0ed2d52e1a46d8b882ca41.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 57, + 264, + 119, + 345 + ], + "blocks": [ + { + "bbox": [ + 57, + 264, + 119, + 345 + ], + "lines": [ + { + "bbox": [ + 57, + 264, + 119, + 345 + ], + "spans": [ + { + "bbox": [ + 57, + 264, + 119, + 345 + ], + "type": "image", + "image_path": "f9de6d25327ccdd6df61d7a9a744fa66437396fe6b1ee7919f4dfff4e9fca88d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 78, + 346, + 97, + 355 + ], + "lines": [ + { + "bbox": [ + 78, + 346, + 97, + 355 + ], + "spans": [ + { + "bbox": [ + 78, + 346, + 97, + 355 + ], + "type": "text", + "content": "Inupt" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 119, + 265, + 180, + 345 + ], + "blocks": [ + { + "bbox": [ + 119, + 265, + 180, + 345 + ], + "lines": [ + { + "bbox": [ + 119, + 265, + 180, + 345 + ], + "spans": [ + { + "bbox": [ + 119, + 265, + 180, + 345 + ], + "type": "image", + "image_path": "1dac05848eea7e9aca11190c9dd56fe020b9580b74d69830629b4a75156140e5.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 144, + 346, + 156, + 354 + ], + "lines": [ + { + "bbox": [ + 144, + 346, + 156, + 354 + ], + "spans": [ + { + "bbox": [ + 144, + 346, + 156, + 354 + ], + "type": "text", + "content": "GT" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 181, + 265, + 242, + 345 + ], + "blocks": [ + { + "bbox": [ + 181, + 265, + 242, + 345 + ], + "lines": [ + { + "bbox": [ + 181, + 265, + 242, + 345 + ], + "spans": [ + { + "bbox": [ + 181, + 265, + 242, + 345 + ], + "type": "image", + "image_path": "9daef77910392ccbfcf76ea0e4bb9b001cfca1e5b7295a56fdbf8c7714ffe83e.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 198, + 346, + 227, + 354 + ], + "lines": [ + { + "bbox": [ + 198, + 346, + 227, + 354 + ], + "spans": [ + { + "bbox": [ + 198, + 346, + 227, + 354 + ], + "type": "text", + "content": "MPRNet" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 360, + 555, + 383 + ], + "lines": [ + { + "bbox": [ + 55, + 360, + 555, + 383 + ], + "spans": [ + { + "bbox": [ + 55, + 360, + 555, + 383 + ], + "type": "text", + "content": "Figure 6. Qualitative deraining performance comparisons on Rain100L [43] and Rain12 [20] datasets. Our CSUD achieves competitive visual results comparable to the SOTA supervised method NeRD-Rain-S [5]." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 243, + 265, + 304, + 345 + ], + "blocks": [ + { + "bbox": [ + 243, + 265, + 304, + 345 + ], + "lines": [ + { + "bbox": [ + 243, + 265, + 304, + 345 + ], + "spans": [ + { + "bbox": [ + 243, + 265, + 304, + 345 + ], + "type": "image", + "image_path": "63968dd4f3ae89da832eabfd8b853722d62f175023e1820d335aa2004f8dafd3.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 252, + 346, + 298, + 354 + ], + "lines": [ + { + "bbox": [ + 252, + 346, + 298, + 354 + ], + "spans": [ + { + "bbox": [ + 252, + 346, + 298, + 354 + ], + "type": "text", + "content": "NeRD-Rain-S" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 304, + 265, + 366, + 345 + ], + "blocks": [ + { + "bbox": [ + 304, + 265, + 366, + 345 + ], + "lines": [ + { + "bbox": [ + 304, + 265, + 366, + 345 + ], + "spans": [ + { + "bbox": [ + 304, + 265, + 366, + 345 + ], + "type": "image", + "image_path": "ca3143bc1f9c89048715d6ecc1958e38d7e10371e8f130088a13968a56f55ae1.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 346, + 365, + 355 + ], + "lines": [ + { + "bbox": [ + 306, + 346, + 365, + 355 + ], + "spans": [ + { + "bbox": [ + 306, + 346, + 365, + 355 + ], + "type": "text", + "content": "DerainCycleGAN" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 367, + 265, + 429, + 345 + ], + "blocks": [ + { + "bbox": [ + 367, + 265, + 429, + 345 + ], + "lines": [ + { + "bbox": [ + 367, + 265, + 429, + 345 + ], + "spans": [ + { + "bbox": [ + 367, + 265, + 429, + 345 + ], + "type": "image", + "image_path": "3873f182fc80c2779c88202905478da033376e8f7d52bfe58fbeb3630bb1b205.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 387, + 346, + 409, + 354 + ], + "lines": [ + { + "bbox": [ + 387, + 346, + 409, + 354 + ], + "spans": [ + { + "bbox": [ + 387, + 346, + 409, + 354 + ], + "type": "text", + "content": "NLCL" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 429, + 265, + 491, + 345 + ], + "blocks": [ + { + "bbox": [ + 429, + 265, + 491, + 345 + ], + "lines": [ + { + "bbox": [ + 429, + 265, + 491, + 345 + ], + "spans": [ + { + "bbox": [ + 429, + 265, + 491, + 345 + ], + "type": "image", + "image_path": "979e7b42e4db60f7988a61748a50e6a04153c2010f31f91ab4711fa7cece0281.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 443, + 346, + 479, + 354 + ], + "lines": [ + { + "bbox": [ + 443, + 346, + 479, + 354 + ], + "spans": [ + { + "bbox": [ + 443, + 346, + 479, + 354 + ], + "type": "text", + "content": "DCD-GAN" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 492, + 265, + 553, + 345 + ], + "blocks": [ + { + "bbox": [ + 492, + 265, + 553, + 345 + ], + "lines": [ + { + "bbox": [ + 492, + 265, + 553, + 345 + ], + "spans": [ + { + "bbox": [ + 492, + 265, + 553, + 345 + ], + "type": "image", + "image_path": "6f7559186be8aced74c8b3a604de130df151533a002ff09894026a86008bc00b.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 498, + 346, + 547, + 355 + ], + "lines": [ + { + "bbox": [ + 498, + 346, + 547, + 355 + ], + "spans": [ + { + "bbox": [ + 498, + 346, + 547, + 355 + ], + "type": "text", + "content": "CSUD (Ours)" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "bbox": [ + 54, + 403, + 296, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 403, + 296, + 534 + ], + "spans": [ + { + "bbox": [ + 54, + 403, + 296, + 534 + ], + "type": "text", + "content": "Datasets and Metrics. We utilize 10 challenging benchmark datasets in our study, including synthetic datasets: Rain100L [43], Rain100H [43], Rain12 [20], Rain800 [17], real-world datasets: RealRain1K-L [18], RealRain1K-H [18], SPA-data [36], RainDS [30], Internet-Data [39], and a nighttime dataset which is based on GTAV-NightRain dataset [50] and enlarged the rain density by [51], we name this nighttime dataset as Night-Rain in the following experiments. As for RainDS, we only use the real rain streaks subset for testing. Following existing methods [3, 40, 49], we use PSNR and SSIM [37] to evaluate different methods." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 54, + 535, + 296, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 535, + 296, + 654 + ], + "spans": [ + { + "bbox": [ + 54, + 535, + 296, + 654 + ], + "type": "text", + "content": "Implementation Details. As for derainer, we employ the CNN-based simple image restoration baseline NAFNet [2]. Our framework is implemented by PyTorch [28] with a GeForce RTX 3090 GPU. We adopt the Adam optimizer [6] to train our network for 200 epochs with the initial learning rate of " + }, + { + "bbox": [ + 54, + 535, + 296, + 654 + ], + "type": "inline_equation", + "content": "1e^{-4}" + }, + { + "bbox": [ + 54, + 535, + 296, + 654 + ], + "type": "text", + "content": ", followed by another 100 epochs with a learning rate of " + }, + { + "bbox": [ + 54, + 535, + 296, + 654 + ], + "type": "inline_equation", + "content": "1e^{-5}" + }, + { + "bbox": [ + 54, + 535, + 296, + 654 + ], + "type": "text", + "content": ". All training images are randomly cropped to " + }, + { + "bbox": [ + 54, + 535, + 296, + 654 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 54, + 535, + 296, + 654 + ], + "type": "text", + "content": " patches in an unpaired manner. The hyperparameters " + }, + { + "bbox": [ + 54, + 535, + 296, + 654 + ], + "type": "inline_equation", + "content": "\\lambda_{1}, \\lambda_{2}" + }, + { + "bbox": [ + 54, + 535, + 296, + 654 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 54, + 535, + 296, + 654 + ], + "type": "inline_equation", + "content": "\\lambda_{3}" + }, + { + "bbox": [ + 54, + 535, + 296, + 654 + ], + "type": "text", + "content": " in Eq. 5 are set to 1, 0.2 and 0.5, while " + }, + { + "bbox": [ + 54, + 535, + 296, + 654 + ], + "type": "inline_equation", + "content": "\\alpha_{1}" + }, + { + "bbox": [ + 54, + 535, + 296, + 654 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 535, + 296, + 654 + ], + "type": "inline_equation", + "content": "\\alpha_{2}" + }, + { + "bbox": [ + 54, + 535, + 296, + 654 + ], + "type": "text", + "content": " in Eq. 6 are set to 10 and 5, respectively." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 55, + 660, + 230, + 673 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 660, + 230, + 673 + ], + "spans": [ + { + "bbox": [ + 55, + 660, + 230, + 673 + ], + "type": "text", + "content": "4.1. Unsupervised Deraining Results" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "We compare our method with unsupervised methods from the most recent years, including [3, 40, 46, 53]. Noting that the related methods for comparison are limited, since only a few" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 403, + 556, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 403, + 556, + 559 + ], + "spans": [ + { + "bbox": [ + 313, + 403, + 556, + 559 + ], + "type": "text", + "content": "unsupervised deraining models have been proposed in the field. We also list some SOTA supervised methods, including [2, 5, 7, 8, 29, 36, 48, 49]. Tab. 1 presents the quantitative deraining results of different methods. It is evident that our CSUD significantly outperforms other unsupervised methods on most datasets. Compared to DCD-GAN[3], our CSUD respectively achieves 1.46dB, 3.00dB, and 2.22dB improvement in PSNR on Rain100L, Rain12 and RealRain1K-L datasets. Additionally, our CSUD even achieves competitive results comparable to several supervised methods. As shown in Fig. 6, compared to other unsupervised methods, our CSUD achieves better deraining results while preserving more color and texture details of image background." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 564, + 443, + 576 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 564, + 443, + 576 + ], + "spans": [ + { + "bbox": [ + 313, + 564, + 443, + 576 + ], + "type": "text", + "content": "4.2. Generalization Results" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 582, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 556, + 713 + ], + "type": "text", + "content": "To validate the generalization capability of CSUD, we compare the results on 5 real-world datasets[18, 30, 36, 39] and a nighttime dataset [51] with SOTA supervised and unsupervised methods, including [2, 3, 5, 7, 29, 40, 46, 48, 49]. Noting that there are no ground truth images in the InternetData [39] dataset, thereby only visual results are presented. All methods are trained on synthetic datasets and tested on the unseen real-world and nighttime datasets. A prevailing perspective is that supervised methods usually outperform unsupervised methods. But as shown in Tab. 2, our CSUD almost comprehensively surpasses supervised models on" + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7474" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 107, + 102, + 504, + 222 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "text", + "content": "Table 2. Quantitative generalization performance comparisons of different deraining methods. Red and blue colors indicate the 1st and 2nd ranks among the unsupervised methods. S, U indicate supervised and unsupervised methods, respectively." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 107, + 102, + 504, + 222 + ], + "lines": [ + { + "bbox": [ + 107, + 102, + 504, + 222 + ], + "spans": [ + { + "bbox": [ + 107, + 102, + 504, + 222 + ], + "type": "table", + "html": "
DatasetsRealRain1K-L [18]RealRain1K-H [18]SPA-data [36]RainDS [30]Night-Rain [51]
MetricsPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIM
DerainNet [7] (S)25.400.84721.910.74230.780.91221.350.591//
MPRNet [48] (S)26.750.89923.580.84033.000.94223.240.65626.130.817
NAFNet [2] (S)28.110.91524.360.84833.460.94122.480.62526.150.816
Restormer [49] (S)26.290.88323.130.85132.620.93622.960.64825.850.808
PromptIR [29] (S)28.140.91524.280.85133.500.94022.480.62526.370.815
NeRD-Rain-S [5] (S)27.640.90423.960.83133.030.93322.370.62226.380.810
DerainCycleGAN [40] (U)28.250.89924.200.82133.530.93718.750.46025.850.802
NLCL [46] (U)20.180.67419.810.62918.760.72717.000.47722.010.475
DCD-GAN [3] (U)27.490.91223.860.84531.550.93022.130.61025.540.798
CSUD (Ours)29.210.92825.450.87833.570.93922.540.61325.940.805
", + "image_path": "1eec57afa8d029abcfbd907d5e7b00c23dba0d5e60884060118b60784da07444.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 56, + 230, + 555, + 357 + ], + "blocks": [ + { + "bbox": [ + 56, + 230, + 555, + 357 + ], + "lines": [ + { + "bbox": [ + 56, + 230, + 555, + 357 + ], + "spans": [ + { + "bbox": [ + 56, + 230, + 555, + 357 + ], + "type": "image", + "image_path": "f1dade8c00373d7dcb13ed0742c934d8288abc0a7dd1f2b90eb42f9fcc244d68.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 361, + 555, + 383 + ], + "lines": [ + { + "bbox": [ + 55, + 361, + 555, + 383 + ], + "spans": [ + { + "bbox": [ + 55, + 361, + 555, + 383 + ], + "type": "text", + "content": "Figure 7. Qualitative generalization results on RealRain1K-L [18] and RealRain1K-H [18] real-world datasets. Our CSUD achieves best visual results among both SOTA supervised and unsupervised methods in real-world scenes." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 62, + 423, + 290, + 491 + ], + "blocks": [ + { + "bbox": [ + 77, + 403, + 272, + 414 + ], + "lines": [ + { + "bbox": [ + 77, + 403, + 272, + 414 + ], + "spans": [ + { + "bbox": [ + 77, + 403, + 272, + 414 + ], + "type": "text", + "content": "Table 3. Effect of CCLoss and SR strategy on CSUD." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 62, + 423, + 290, + 491 + ], + "lines": [ + { + "bbox": [ + 62, + 423, + 290, + 491 + ], + "spans": [ + { + "bbox": [ + 62, + 423, + 290, + 491 + ], + "type": "table", + "html": "
CCSRRain100L[43]Rain12[20]RealRain1K-L[18]
PSNRSSIMPSNRSSIMPSNRSSIM
XX30.660.92531.600.93130.420.932
✓X31.240.93432.620.94831.140.943
X✓32.690.94933.430.93931.860.950
✓✓33.280.95434.560.95132.710.959
", + "image_path": "9f857987c8fa188a993dddfba19768f8faadfd7501798df20864f426f8083155.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 510, + 295, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 295, + 544 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 295, + 544 + ], + "type": "text", + "content": "the first 4 real-world datasets, while remaining competitive on the nighttime dataset. This strongly demonstrates the excellent generalization ability of our method." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 545, + 296, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 296, + 653 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 296, + 653 + ], + "type": "text", + "content": "As shown in Fig. 7 and Fig. 8, the performance of some most advanced supervised methods like Restormer [49] and NeRD-Rain-S [5] is extremely poor when directly applied to real-world scenarios. In contrast, our CSUD achieves better visual results in real world. It is worth mentioning again that, as shown in Fig. 1 in Sec. 1, our method also shows extremely superior deraining performance on the rainy images captured by ourselves in real scenarios, which further demonstrates our excellent generalization capability." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 660, + 156, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 660, + 156, + 672 + ], + "spans": [ + { + "bbox": [ + 55, + 660, + 156, + 672 + ], + "type": "text", + "content": "4.3. Ablation Studies" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 677, + 297, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 297, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 297, + 713 + ], + "type": "text", + "content": "Effect of CCLoss and SR Strategy on Deraining Performance. To verify the effectiveness of CCLoss and SR strategy, we conduct ablation experiments on Rain100L [43]," + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 404, + 555, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 404, + 555, + 525 + ], + "spans": [ + { + "bbox": [ + 313, + 404, + 555, + 525 + ], + "type": "text", + "content": "Rain12 [20] and RealRain1K-L [18] datasets. As shown in Table 3, compared to the baseline, the effect on PSNR is improved by 1.38dB, 1.02dB, and 0.72dB among the above three datasets with CCLoss, respectively. And the effect is improved by 1.38dB, 1.83dB, and 1.44dB with SR strategy, respectively. Furthermore, when both CCLoss and SR strategy are introduced into the baseline, the PSNR on the three datasets is increased by 2.62dB, 2.96dB, and 2.29dB respectively. The above ablation results strongly verify the superiority and importance of CCLoss and SR strategy." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 529, + 556, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 529, + 556, + 661 + ], + "spans": [ + { + "bbox": [ + 313, + 529, + 556, + 661 + ], + "type": "text", + "content": "Effect of CSUD framework on Generalization Performance. To validate the effectiveness of our CSUD in improving generalization performance and the universality and robustness of our method, we also conduct experiments incorporating CSUD framework on another 2 different baselines: NeRD-Rain-S [5] and Restormer [49]. Fig. 9 shows the PSNR improvements of different CSUD based models compared to their supervised version. Notably, we train all the above models on Rain100L and test them on the 3 real-world datasets. This strongly demonstrates the effectiveness of CSUD in improving generalization performance." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 665, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 556, + 713 + ], + "type": "text", + "content": "Effect of Different Training Datasets on Generalization Performance. To further validate the generalization performance of CSUD, we employ 4 different CSUD models which are respectively trained on 4 synthetic and real-" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7475" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 70, + 119, + 175 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 119, + 175 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 119, + 175 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 119, + 175 + ], + "type": "image", + "image_path": "9e2ad534a36fae4ed759ec9e803bc5e689719c4047cc14cc0afe3695a67f4b50.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 80, + 178, + 97, + 186 + ], + "lines": [ + { + "bbox": [ + 80, + 178, + 97, + 186 + ], + "spans": [ + { + "bbox": [ + 80, + 178, + 97, + 186 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 119, + 70, + 181, + 175 + ], + "blocks": [ + { + "bbox": [ + 119, + 70, + 181, + 175 + ], + "lines": [ + { + "bbox": [ + 119, + 70, + 181, + 175 + ], + "spans": [ + { + "bbox": [ + 119, + 70, + 181, + 175 + ], + "type": "image", + "image_path": "089d8648dcf54bd2f09675466293f2985658341b3980468cce420ba34a9b0241.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 134, + 178, + 166, + 186 + ], + "lines": [ + { + "bbox": [ + 134, + 178, + 166, + 186 + ], + "spans": [ + { + "bbox": [ + 134, + 178, + 166, + 186 + ], + "type": "text", + "content": "Restormer" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 181, + 70, + 242, + 175 + ], + "blocks": [ + { + "bbox": [ + 181, + 70, + 242, + 175 + ], + "lines": [ + { + "bbox": [ + 181, + 70, + 242, + 175 + ], + "spans": [ + { + "bbox": [ + 181, + 70, + 242, + 175 + ], + "type": "image", + "image_path": "e864e12231b930dc1815ec087f306d4e93ebca13d00514a5e2ab66f587f5a422.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 196, + 178, + 227, + 186 + ], + "lines": [ + { + "bbox": [ + 196, + 178, + 227, + 186 + ], + "spans": [ + { + "bbox": [ + 196, + 178, + 227, + 186 + ], + "type": "text", + "content": "PromptIR" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 243, + 70, + 304, + 175 + ], + "blocks": [ + { + "bbox": [ + 243, + 70, + 304, + 175 + ], + "lines": [ + { + "bbox": [ + 243, + 70, + 304, + 175 + ], + "spans": [ + { + "bbox": [ + 243, + 70, + 304, + 175 + ], + "type": "image", + "image_path": "910033d295110dce184f4f188b6e95efcf811c662b69612b8c00793f803be3df.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 252, + 178, + 296, + 186 + ], + "lines": [ + { + "bbox": [ + 252, + 178, + 296, + 186 + ], + "spans": [ + { + "bbox": [ + 252, + 178, + 296, + 186 + ], + "type": "text", + "content": "NeRD-Rain-S" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 304, + 70, + 367, + 175 + ], + "blocks": [ + { + "bbox": [ + 304, + 70, + 367, + 175 + ], + "lines": [ + { + "bbox": [ + 304, + 70, + 367, + 175 + ], + "spans": [ + { + "bbox": [ + 304, + 70, + 367, + 175 + ], + "type": "image", + "image_path": "cf0cd317f077a293d450fa3141c3a3d2c265d39d32271e057a3f556baeab7812.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 309, + 178, + 364, + 186 + ], + "lines": [ + { + "bbox": [ + 309, + 178, + 364, + 186 + ], + "spans": [ + { + "bbox": [ + 309, + 178, + 364, + 186 + ], + "type": "text", + "content": "DerainCycleGAN" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 153, + 191, + 456, + 201 + ], + "lines": [ + { + "bbox": [ + 153, + 191, + 456, + 201 + ], + "spans": [ + { + "bbox": [ + 153, + 191, + 456, + 201 + ], + "type": "text", + "content": "Figure 8. Qualitative generalization results on Internet-Data [39] real-world dataset." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 367, + 70, + 429, + 175 + ], + "blocks": [ + { + "bbox": [ + 367, + 70, + 429, + 175 + ], + "lines": [ + { + "bbox": [ + 367, + 70, + 429, + 175 + ], + "spans": [ + { + "bbox": [ + 367, + 70, + 429, + 175 + ], + "type": "image", + "image_path": "5615b8809e10dc6283fdabb209b5a1375f16766b7c6e5aee3a0f498e83678005.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 388, + 178, + 408, + 186 + ], + "lines": [ + { + "bbox": [ + 388, + 178, + 408, + 186 + ], + "spans": [ + { + "bbox": [ + 388, + 178, + 408, + 186 + ], + "type": "text", + "content": "NLCL" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 430, + 70, + 492, + 175 + ], + "blocks": [ + { + "bbox": [ + 430, + 70, + 492, + 175 + ], + "lines": [ + { + "bbox": [ + 430, + 70, + 492, + 175 + ], + "spans": [ + { + "bbox": [ + 430, + 70, + 492, + 175 + ], + "type": "image", + "image_path": "3959e7251f64ee540ee6f0d02185a4b4ceba7ae0799c008e5034e13d8c14c8d1.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 443, + 178, + 478, + 186 + ], + "lines": [ + { + "bbox": [ + 443, + 178, + 478, + 186 + ], + "spans": [ + { + "bbox": [ + 443, + 178, + 478, + 186 + ], + "type": "text", + "content": "DCD-GAN" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 492, + 70, + 553, + 175 + ], + "blocks": [ + { + "bbox": [ + 492, + 70, + 553, + 175 + ], + "lines": [ + { + "bbox": [ + 492, + 70, + 553, + 175 + ], + "spans": [ + { + "bbox": [ + 492, + 70, + 553, + 175 + ], + "type": "image", + "image_path": "5b94aed754c140f6a5fb3a8fc45189ebc7fb99c68f79021e2e30ffb9c7bd2da0.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 501, + 178, + 543, + 186 + ], + "lines": [ + { + "bbox": [ + 501, + 178, + 543, + 186 + ], + "spans": [ + { + "bbox": [ + 501, + 178, + 543, + 186 + ], + "type": "text", + "content": "CSUD (Ours)" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 58, + 219, + 297, + 312 + ], + "blocks": [ + { + "bbox": [ + 58, + 219, + 297, + 312 + ], + "lines": [ + { + "bbox": [ + 58, + 219, + 297, + 312 + ], + "spans": [ + { + "bbox": [ + 58, + 219, + 297, + 312 + ], + "type": "image", + "image_path": "2cc05fb52f79d70e476717d8cdc46fdeb784ecd1ce2bdb221f6327e6b63d08cf.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 316, + 296, + 350 + ], + "lines": [ + { + "bbox": [ + 55, + 316, + 296, + 350 + ], + "spans": [ + { + "bbox": [ + 55, + 316, + 296, + 350 + ], + "type": "text", + "content": "Figure 9. PSNR improvement of different baselines with CSUD. We test the models' generalization performance on 3 real-world datasets out of domain compared with their supervised version." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "table", + "bbox": [ + 67, + 404, + 288, + 466 + ], + "blocks": [ + { + "bbox": [ + 55, + 362, + 296, + 396 + ], + "lines": [ + { + "bbox": [ + 55, + 362, + 296, + 396 + ], + "spans": [ + { + "bbox": [ + 55, + 362, + 296, + 396 + ], + "type": "text", + "content": "Table 4. Effect of different training datasets on generalization performance. We test 4 different CSUD models which are trained on 4 different datasets on 3 real-world datasets out of domain." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 67, + 404, + 288, + 466 + ], + "lines": [ + { + "bbox": [ + 67, + 404, + 288, + 466 + ], + "spans": [ + { + "bbox": [ + 67, + 404, + 288, + 466 + ], + "type": "table", + "html": "
Train datasetsSPA-data [36]Night-Rain [51]RainDS [30]
PSNRSSIMPSNRSSIMPSNRSSIM
Rain100L [43]33.570.93925.940.80522.540.613
Rain100H [43]29.610.88924.530.70221.850.600
RealRain1K-L [18]32.950.93728.140.83522.900.630
RealRain1K-H [18]29.250.89726.310.78722.300.564
", + "image_path": "9b02bb7c24d93ddff5eacfe2b908bb21c2df8a7babf5b4678f259936a069423c.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "table_body" + } + ], + "index": 20 + }, + { + "bbox": [ + 54, + 486, + 295, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 486, + 295, + 593 + ], + "spans": [ + { + "bbox": [ + 54, + 486, + 295, + 593 + ], + "type": "text", + "content": "world datasets with different rain distributions, and test them on 3 real-world datasets. As shown in Tab. 4, in most cases, CSUD can achieve strong generalization performance, demonstrating the robustness of our approach. Due to the significant differences between rain streaks in Rain100H and real world, its generalization capability is the worst. Additionally, the generalization performance when training on Rain100L is comparable to that when training on real-world datasets, demonstrating our strong generalization ability." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 55, + 594, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 297, + 714 + ], + "type": "text", + "content": "Effect of the Weight Coefficients of Each Loss Function. We conduct extensive ablation experiments on the weight coefficients of perceptual loss in Eq. (5), " + }, + { + "bbox": [ + 55, + 594, + 297, + 714 + ], + "type": "inline_equation", + "content": "L_{CC}" + }, + { + "bbox": [ + 55, + 594, + 297, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 594, + 297, + 714 + ], + "type": "inline_equation", + "content": "L_{SR - G}" + }, + { + "bbox": [ + 55, + 594, + 297, + 714 + ], + "type": "text", + "content": " in Eq. (6). The results are shown in Tab. 5. Since " + }, + { + "bbox": [ + 55, + 594, + 297, + 714 + ], + "type": "inline_equation", + "content": "L_{SR - Der}" + }, + { + "bbox": [ + 55, + 594, + 297, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 594, + 297, + 714 + ], + "type": "inline_equation", + "content": "L_{SR - G}" + }, + { + "bbox": [ + 55, + 594, + 297, + 714 + ], + "type": "text", + "content": " are highly correlated, to ensure that the losses are on the same scale, the coefficient for " + }, + { + "bbox": [ + 55, + 594, + 297, + 714 + ], + "type": "inline_equation", + "content": "L_{SR - Der}" + }, + { + "bbox": [ + 55, + 594, + 297, + 714 + ], + "type": "text", + "content": " is set to 0.1 times that of " + }, + { + "bbox": [ + 55, + 594, + 297, + 714 + ], + "type": "inline_equation", + "content": "L_{SR - G}" + }, + { + "bbox": [ + 55, + 594, + 297, + 714 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 594, + 297, + 714 + ], + "type": "inline_equation", + "content": "L_{SR}" + }, + { + "bbox": [ + 55, + 594, + 297, + 714 + ], + "type": "text", + "content": " in Tab. 5 indicates " + }, + { + "bbox": [ + 55, + 594, + 297, + 714 + ], + "type": "inline_equation", + "content": "L_{SR - G}" + }, + { + "bbox": [ + 55, + 594, + 297, + 714 + ], + "type": "text", + "content": ". Based on the experimental results, aside from our selected optimal combination, the deraining and generalization performance of our method remain at a high level" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 224, + 555, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 224, + 555, + 246 + ], + "spans": [ + { + "bbox": [ + 313, + 224, + 555, + 246 + ], + "type": "text", + "content": "under most other weight combinations, demonstrating the robustness of our CSUD framework." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 247, + 555, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 247, + 555, + 272 + ], + "spans": [ + { + "bbox": [ + 313, + 247, + 555, + 272 + ], + "type": "text", + "content": "More detailed explanations of our method, experimental details and analysis can be found in the Appendix." + } + ] + } + ], + "index": 24 + }, + { + "type": "table", + "bbox": [ + 317, + 312, + 553, + 387 + ], + "blocks": [ + { + "bbox": [ + 313, + 281, + 554, + 303 + ], + "lines": [ + { + "bbox": [ + 313, + 281, + 554, + 303 + ], + "spans": [ + { + "bbox": [ + 313, + 281, + 554, + 303 + ], + "type": "text", + "content": "Table 5. Ablation experiments on the weight coefficients of each loss function. All models in the table are trained on Rain100L." + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 312, + 553, + 387 + ], + "lines": [ + { + "bbox": [ + 317, + 312, + 553, + 387 + ], + "spans": [ + { + "bbox": [ + 317, + 312, + 553, + 387 + ], + "type": "table", + "html": "
Lper(λ2)LSR(α2)LCC(α1)Rain100L [43]SPA-data [36]RealRain1K-L [18]
PSNRSSIMPSNRSSIMPSNRSSIM
0.25.00.130.690.93232.130.92928.210.915
0.25.01.033.150.95133.500.93829.750.928
0.21.01032.840.94633.490.93529.100.918
0.2101032.900.95133.400.93329.060.920
05.01032.180.93932.690.93628.240.916
1.05.01032.720.94733.050.93028.700.918
0.25.01033.280.95433.570.93929.210.928
", + "image_path": "1cb299f40cde955e648dd468d023e8008f22ff2f9408b791c611c2dd9a17015a.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "table_body" + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 408, + 388, + 421 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 408, + 388, + 421 + ], + "spans": [ + { + "bbox": [ + 313, + 408, + 388, + 421 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 313, + 429, + 556, + 596 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 429, + 556, + 596 + ], + "spans": [ + { + "bbox": [ + 313, + 429, + 556, + 596 + ], + "type": "text", + "content": "In this paper, we propose a novel unsupervised framework for image deraining to address the lack of paired data and the poor generalization in real-world scenarios. We first propose a channel consistency loss by introducing the channel consistency prior into the optimization process of the generator, ensuring that more background details are preserved while generating pseudo rainy images. In addition, a self-reconstruction strategy is proposed to alleviate the redundant information transfer problem and further improve the deraining performance and the generalization capability of our method. Extensive experiments on multiple synthetic and real-world datasets verify that our method achieves excellent deraining performance with strong generalization capability across various real-world rainy scenarios." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 313, + 605, + 557, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 557, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 557, + 712 + ], + "type": "text", + "content": "Acknowledgement. This work was supported by the National Natural Science Foundation of China under Grant 62171304 and partly by the Natural Science Foundation of Sichuan Province under Grant 2024NSFSC1423, the TCL Science and Technology Innovation Fund under grant 25JZH008, the Young Faculty Technology Innovation Capacity Enhancement Program of Sichuan University under Grant 2024SCUQJTX025, and the Beijing Natural Science Foundation L221011." + } + ] + } + ], + "index": 29 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7476" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 297, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 91, + 297, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 297, + 145 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 297, + 145 + ], + "type": "text", + "content": "[1] Liangyu Chen, Xin Lu, Jie Zhang, Xiaojie Chu, and Chengpeng Chen. Hinet: Half instance normalization network for image restoration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 182-192, 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 146, + 296, + 189 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 146, + 296, + 189 + ], + "spans": [ + { + "bbox": [ + 61, + 146, + 296, + 189 + ], + "type": "text", + "content": "[2] Liangyu Chen, Xiaojie Chu, Xiangyu Zhang, and Jian Sun. Simple baselines for image restoration. In European conference on computer vision, pages 17-33. Springer, 2022. 6, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 190, + 296, + 256 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 190, + 296, + 256 + ], + "spans": [ + { + "bbox": [ + 62, + 190, + 296, + 256 + ], + "type": "text", + "content": "[3] Xiang Chen, Jinshan Pan, Kui Jiang, Yufeng Li, Yufeng Huang, Caihua Kong, Longgang Dai, and Zhentao Fan. Unpaired deep image deraining using dual contrastive learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2017-2026, 2022. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 258, + 296, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 258, + 296, + 300 + ], + "spans": [ + { + "bbox": [ + 62, + 258, + 296, + 300 + ], + "type": "text", + "content": "[4] Xiang Chen, Hao Li, Mingqiang Li, and Jinshan Pan. Learning a sparse transformer network for effective image deraining. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5896-5905, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 302, + 296, + 356 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 302, + 296, + 356 + ], + "spans": [ + { + "bbox": [ + 62, + 302, + 296, + 356 + ], + "type": "text", + "content": "[5] Xiang Chen, Jinshan Pan, and Jiangxin Dong. Bidirectional multi-scale implicit neural representations for image deraining. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25627-25636, 2024. 2, 6, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 357, + 296, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 357, + 296, + 379 + ], + "spans": [ + { + "bbox": [ + 62, + 357, + 296, + 379 + ], + "type": "text", + "content": "[6] P Kingma Diederik. Adam: A method for stochastic optimization. (No Title), 2014. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 380, + 296, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 380, + 296, + 423 + ], + "spans": [ + { + "bbox": [ + 62, + 380, + 296, + 423 + ], + "type": "text", + "content": "[7] Xueyang Fu, Jiabin Huang, Xinghao Ding, Yinghao Liao, and John Paisley. Clearing the skies: A deep network architecture for single-image rain removal. IEEE Transactions on Image Processing, 26(6):2944-2956, 2017. 1, 2, 6, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 425, + 296, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 425, + 296, + 479 + ], + "spans": [ + { + "bbox": [ + 62, + 425, + 296, + 479 + ], + "type": "text", + "content": "[8] Xueyang Fu, Jiabin Huang, Delu Zeng, Yue Huang, Xinghao Ding, and John Paisley. Removing rain from single images via a deep detail network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3855-3863, 2017. 1, 2, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 480, + 296, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 480, + 296, + 523 + ], + "spans": [ + { + "bbox": [ + 62, + 480, + 296, + 523 + ], + "type": "text", + "content": "[9] Jinjin Gu, Xianzheng Ma, Xiangtao Kong, Yu Qiao, and Chao Dong. Networks are slacking off: Understanding generalization problem in image deraining. Advances in Neural Information Processing Systems, 36, 2023. 1, 2, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 525, + 296, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 525, + 296, + 567 + ], + "spans": [ + { + "bbox": [ + 57, + 525, + 296, + 567 + ], + "type": "text", + "content": "[10] Kaiming He, Jian Sun, and Xiaou Tang. Single image haze removal using dark channel prior. IEEE transactions on pattern analysis and machine intelligence, 33(12):2341-2353, 2010. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 569, + 296, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 569, + 296, + 613 + ], + "spans": [ + { + "bbox": [ + 57, + 569, + 296, + 613 + ], + "type": "text", + "content": "[11] Xiaowei Hu, Chi-Wing Fu, Lei Zhu, and Pheng-Ann Heng. Depth-attentional features for single-image rain removal. In Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition, pages 8022–8031, 2019. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 614, + 296, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 614, + 296, + 658 + ], + "spans": [ + { + "bbox": [ + 57, + 614, + 296, + 658 + ], + "type": "text", + "content": "[12] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1125-1134, 2017. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 658, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 658, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 658, + 296, + 713 + ], + "type": "text", + "content": "[13] Kui Jiang, Zhongyuan Wang, Peng Yi, Chen Chen, Baojin Huang, Yimin Luo, Jiayi Ma, and Junjun Jiang. Multi-scale progressive fusion network for single image deraining. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8346-8355, 2020. 1, 2" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 316, + 73, + 555, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 555, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 555, + 116 + ], + "type": "text", + "content": "[14] Xin Jin, Zhibo Chen, Jianxin Lin, Zhikai Chen, and Wei Zhou. Unsupervised single image deraining with self-supervised constraints. In 2019 IEEE International Conference on Image Processing (ICIP), pages 2761-2765. IEEE, 2019. 1, 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 118, + 555, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 118, + 555, + 172 + ], + "spans": [ + { + "bbox": [ + 317, + 118, + 555, + 172 + ], + "type": "text", + "content": "[15] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14, pages 694-711. Springer, 2016. 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 173, + 555, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 173, + 555, + 225 + ], + "spans": [ + { + "bbox": [ + 317, + 173, + 555, + 225 + ], + "type": "text", + "content": "[16] Boyun Li, Xiao Liu, Peng Hu, Zhongqin Wu, Jiancheng Lv, and Xi Peng. All-in-one image restoration for unknown corruption. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 17452-17462, 2022. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 228, + 555, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 228, + 555, + 293 + ], + "spans": [ + { + "bbox": [ + 316, + 228, + 555, + 293 + ], + "type": "text", + "content": "[17] Siyuan Li, Iago Bruno Araujo, Wenqi Ren, Zhangyang Wang, Eric K Tokuda, Roberto Hirata Junior, Roberto Cesar-Junior, Jiawan Zhang, Xiaojie Guo, and Xiaochun Cao. Single image deraining: A comprehensive benchmark analysis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3838-3847, 2019. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 294, + 554, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 294, + 554, + 337 + ], + "spans": [ + { + "bbox": [ + 316, + 294, + 554, + 337 + ], + "type": "text", + "content": "[18] Wei Li, Qiming Zhang, Jing Zhang, Zhen Huang, Xinmei Tian, and Dacheng Tao. Toward real-world single image deraining: A new benchmark and beyond. arXiv preprint arXiv:2206.05514, 2022. 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 338, + 555, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 338, + 555, + 392 + ], + "spans": [ + { + "bbox": [ + 317, + 338, + 555, + 392 + ], + "type": "text", + "content": "[19] Xia Li, Jianlong Wu, Zhouchen Lin, Hong Liu, and Hongbin Zha. Recurrent squeeze-and-excitation context aggregation net for single image deraining. In Proceedings of the European conference on computer vision (ECCV), pages 254-269, 2018. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 393, + 555, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 393, + 555, + 437 + ], + "spans": [ + { + "bbox": [ + 317, + 393, + 555, + 437 + ], + "type": "text", + "content": "[20] Yu Li, Robby T Tan, Xiaojie Guo, Jiangbo Lu, and Michael S Brown. Rain streak removal using layer priors. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2736-2744, 2016. 4, 6, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 437, + 555, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 437, + 555, + 491 + ], + "spans": [ + { + "bbox": [ + 317, + 437, + 555, + 491 + ], + "type": "text", + "content": "[21] Di Lin, Xin Wang, Jia Shen, Renjie Zhang, Ruonan Liu, Miao-hui Wang, Wuyuan Xie, Qing Guo, and Ping Li. Generative status estimation and information decoupling for image rain removal. Advances in Neural Information Processing Systems, 35:4612-4625, 2022. 1, 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 493, + 555, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 493, + 555, + 547 + ], + "spans": [ + { + "bbox": [ + 317, + 493, + 555, + 547 + ], + "type": "text", + "content": "[22] Xin Lin, Chao Ren, Xiao Liu, Jie Huang, and Yinjie Lei. Unsupervised image denoising in real-world scenarios via self-collaboration parallel generative adversarial branches. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12642-12652, 2023. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 548, + 555, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 548, + 555, + 591 + ], + "spans": [ + { + "bbox": [ + 317, + 548, + 555, + 591 + ], + "type": "text", + "content": "[23] Xin Lin, Jingtong Yue, Sixian Ding, Chao Ren, Chun-Le Guo, and Chongyi Li. Unlocking low-light-rainy image restoration by pairwise degradation feature vector guidance. arXiv preprint arXiv:2305.03997, 2023. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 592, + 555, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 592, + 555, + 635 + ], + "spans": [ + { + "bbox": [ + 317, + 592, + 555, + 635 + ], + "type": "text", + "content": "[24] Li Liu, Wanli Ouyang, Xiaogang Wang, Paul Fieguth, Jie Chen, Xinwang Liu, and Matti Pietikainen. Deep learning for generic object detection: A survey. International journal of computer vision, 128:261-318, 2020. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 636, + 555, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 636, + 555, + 679 + ], + "spans": [ + { + "bbox": [ + 317, + 636, + 555, + 679 + ], + "type": "text", + "content": "[25] Xiao Liu, Xiangyu Liao, Xiuya Shi, Linbo Qing, and Chao Ren. Efficient information modulation network for image super-resolution. In ECAI 2023, pages 1544-1551. IOS Press, 2023. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 681, + 555, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 555, + 713 + ], + "type": "text", + "content": "[26] Xiao Liu, Xiuya Shi, Lufei Chen, Linbo Qing, and Chao Ren. Efficient parallel multi-scale detail and semantic encoding network for lightweight semantic segmentation. In Proceedings" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7477" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 77, + 73, + 296, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 73, + 296, + 95 + ], + "spans": [ + { + "bbox": [ + 77, + 73, + 296, + 95 + ], + "type": "text", + "content": "of the 31st ACM International Conference on Multimedia, pages 2544-2552, 2023. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 95, + 296, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 95, + 296, + 149 + ], + "spans": [ + { + "bbox": [ + 56, + 95, + 296, + 149 + ], + "type": "text", + "content": "[27] Dongwon Park, Byung Hyun Lee, and Se Young Chun. All-in-one image restoration for unknown degradations using adaptive discriminative filters for specific degradations. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5815-5824. IEEE, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 150, + 296, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 150, + 296, + 215 + ], + "spans": [ + { + "bbox": [ + 56, + 150, + 296, + 215 + ], + "type": "text", + "content": "[28] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 217, + 295, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 217, + 295, + 259 + ], + "spans": [ + { + "bbox": [ + 56, + 217, + 295, + 259 + ], + "type": "text", + "content": "[29] Vaishnav Potlapalli, Syed Waqas Zamir, Salman H Khan, and Fahad Shahbaz Khan. Prompt: Prompting for all-in-one image restoration. Advances in Neural Information Processing Systems, 36, 2023. 1, 2, 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 261, + 296, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 261, + 296, + 304 + ], + "spans": [ + { + "bbox": [ + 56, + 261, + 296, + 304 + ], + "type": "text", + "content": "[30] Ruijie Quan, Xin Yu, Yuanzhi Liang, and Yi Yang. Removing raindrops and rain streaks in one go. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9147-9156, 2021. 6, 7, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 305, + 295, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 305, + 295, + 348 + ], + "spans": [ + { + "bbox": [ + 56, + 305, + 295, + 348 + ], + "type": "text", + "content": "[31] Chao Ren, Yizhong Pan, and Jie Huang. Enhanced latent space blind model for real image denoising via alternative optimization. Advances in Neural Information Processing Systems, 35:38386-38399, 2022. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 350, + 295, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 350, + 295, + 403 + ], + "spans": [ + { + "bbox": [ + 56, + 350, + 295, + 403 + ], + "type": "text", + "content": "[32] Dongwei Ren, Wangmeng Zuo, Qinghua Hu, Pengfei Zhu, and Deyu Meng. Progressive image deraining networks: A better and simpler baseline. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3937-3946, 2019. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 404, + 295, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 404, + 295, + 448 + ], + "spans": [ + { + "bbox": [ + 56, + 404, + 295, + 448 + ], + "type": "text", + "content": "[33] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. IEEE transactions on pattern analysis and machine intelligence, 39(6):1137-1149, 2016. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 449, + 295, + 481 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 449, + 295, + 481 + ], + "spans": [ + { + "bbox": [ + 56, + 449, + 295, + 481 + ], + "type": "text", + "content": "[34] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 482, + 296, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 482, + 296, + 536 + ], + "spans": [ + { + "bbox": [ + 56, + 482, + 296, + 536 + ], + "type": "text", + "content": "[35] Cong Wang, Jinshan Pan, Wei Wang, Jiangxin Dong, Mengzhu Wang, Yakun Ju, and Junyang Chen. Promptstorer: A prompting image restoration method with degradation perception. Advances in Neural Information Processing Systems, 36:8898-8912, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 537, + 296, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 537, + 296, + 591 + ], + "spans": [ + { + "bbox": [ + 56, + 537, + 296, + 591 + ], + "type": "text", + "content": "[36] Tianyu Wang, Xin Yang, Ke Xu, Shaozhe Chen, Qiang Zhang, and Rynson WH Lau. Spatial attentive single-image deraining with a high quality real rain dataset. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12270-12279, 2019. 1, 2, 4, 6, 7, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 592, + 295, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 592, + 295, + 635 + ], + "spans": [ + { + "bbox": [ + 56, + 592, + 295, + 635 + ], + "type": "text", + "content": "[37] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 636, + 295, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 636, + 295, + 690 + ], + "spans": [ + { + "bbox": [ + 56, + 636, + 295, + 690 + ], + "type": "text", + "content": "[38] Zhendong Wang, Xiaodong Cun, Jianmin Bao, Wengang Zhou, Jianzhuang Liu, and Houqiang Li. Uformer: A general u-shaped transformer for image restoration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 17683-17693, 2022. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 691, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 295, + 713 + ], + "type": "text", + "content": "[39] Wei Wei, Deyu Meng, Qian Zhao, Zongben Xu, and Ying Wu. Semi-supervised transfer learning for image rain removal. In" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "text", + "content": "Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3877-3886, 2019. 3, 6, 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 96, + 555, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 555, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 555, + 150 + ], + "type": "text", + "content": "[40] Yanyan Wei, Zhao Zhang, Yang Wang, Mingliang Xu, Yi Yang, Shuicheng Yan, and Meng Wang. Deraincyclegan: Rain attentive cyclegan for single image deraining and rainmaking. IEEE Transactions on Image Processing, 30:4788-4801, 2021. 1, 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 152, + 555, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 555, + 206 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 555, + 206 + ], + "type": "text", + "content": "[41] Zbigniew Wojna, Vittorio Ferrari, Sergio Guadarrama, Nathan Silberman, Liang-Chieh Chen, Alireza Fathi, and Jasper Uijlings. The devil is in the decoder: Classification, regression and gans. International Journal of Computer Vision, 127: 1694-1706, 2019. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 209, + 555, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 209, + 555, + 252 + ], + "spans": [ + { + "bbox": [ + 316, + 209, + 555, + 252 + ], + "type": "text", + "content": "[42] Wei Xu, Xinyuan Chen, Haoming Guo, Xiaolin Huang, and Wei Liu. Unsupervised image restoration with quality-task-perception loss. IEEE Transactions on Circuits and Systems for Video Technology, 32(9):5736-5747, 2022. 1, 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 254, + 555, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 254, + 555, + 308 + ], + "spans": [ + { + "bbox": [ + 316, + 254, + 555, + 308 + ], + "type": "text", + "content": "[43] Wenhan Yang, Robby T Tan, Jiashi Feng, Zongming Guo, Shuicheng Yan, and Jiaying Liu. Joint rain detection and removal from a single image with contextualized deep networks. IEEE transactions on pattern analysis and machine intelligence, 42(6):1377-1393, 2019. 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 309, + 555, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 309, + 555, + 363 + ], + "spans": [ + { + "bbox": [ + 316, + 309, + 555, + 363 + ], + "type": "text", + "content": "[44] Rajeev Yasarla, Vishwanath A Sindagi, and Vishal M Patel. Syn2real transfer learning for image deraining using gaussian processes. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2726-2736, 2020. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 365, + 555, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 365, + 555, + 419 + ], + "spans": [ + { + "bbox": [ + 316, + 365, + 555, + 419 + ], + "type": "text", + "content": "[45] Yuntong Ye, Yi Chang, Hanyu Zhou, and Luxin Yan. Closing the loop: Joint rain generation and removal via disentangled image translation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2053-2062, 2021. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 422, + 555, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 422, + 555, + 477 + ], + "spans": [ + { + "bbox": [ + 316, + 422, + 555, + 477 + ], + "type": "text", + "content": "[46] Yuntong Ye, Changfeng Yu, Yi Chang, Lin Zhu, Xi-Le Zhao, Luxin Yan, and Yonghong Tian. Unsupervised deraining: Where contrastive learning meets self-similarity. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5821-5830, 2022. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 478, + 555, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 478, + 555, + 533 + ], + "spans": [ + { + "bbox": [ + 316, + 478, + 555, + 533 + ], + "type": "text", + "content": "[47] Qiaosi Yi, Juncheng Li, Qinyan Dai, Faming Fang, Guixu Zhang, and Tieyong Zeng. Structure-preserving deraining with residue channel prior guidance. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4238-4247, 2021. 1, 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 534, + 555, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 534, + 555, + 588 + ], + "spans": [ + { + "bbox": [ + 316, + 534, + 555, + 588 + ], + "type": "text", + "content": "[48] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, Ming-Hsuan Yang, and Ling Shao. Multi-stage progressive image restoration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14821-14831, 2021. 2, 6, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 590, + 555, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 590, + 555, + 655 + ], + "spans": [ + { + "bbox": [ + 316, + 590, + 555, + 655 + ], + "type": "text", + "content": "[49] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5728-5739, 2022. 1, 2, 6, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 657, + 555, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 657, + 555, + 690 + ], + "spans": [ + { + "bbox": [ + 316, + 657, + 555, + 690 + ], + "type": "text", + "content": "[50] Fan Zhang, Shaodi You, Yu Li, and Ying Fu. GTav-nightrain: Photometric realistic large-scale dataset for night-time rain streak removal. arXiv preprint arXiv:2210.04708, 2022. 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 691, + 555, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 555, + 713 + ], + "type": "text", + "content": "[51] Fan Zhang, Shaodi You, Yu Li, and Ying Fu. Learning rain location prior for nighttime deraining. In Proceedings of" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7478" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 195 + ], + "type": "list", + "angle": 0, + "index": 3, + "blocks": [ + { + "bbox": [ + 76, + 72, + 296, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 296, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 296, + 95 + ], + "type": "text", + "content": "the IEEE/CVF International Conference on Computer Vision, pages 13148-13157, 2023. 6, 7, 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 296, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 296, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 296, + 140 + ], + "type": "text", + "content": "[52] He Zhang and Vishal M Patel. Density-aware single image de-raining using a multi-stream dense network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 695-704, 2018. 1, 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 141, + 296, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 141, + 296, + 195 + ], + "spans": [ + { + "bbox": [ + 56, + 141, + 296, + 195 + ], + "type": "text", + "content": "[53] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In Proceedings of the IEEE international conference on computer vision, pages 2223-2232, 2017. 2, 3, 6" + } + ] + } + ], + "index": 2 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "7479" + } + ] + } + ], + "index": 4 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Channel-wise Noise Scheduled Diffusion for Inverse Rendering in Indoor Scenes/e287505c-edc5-4f3b-9021-f507eb27467f_content_list.json b/2025/Channel-wise Noise Scheduled Diffusion for Inverse Rendering in Indoor Scenes/e287505c-edc5-4f3b-9021-f507eb27467f_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..07cb0c75291b6e663be4a11f71606ab203ee33b0 --- /dev/null +++ b/2025/Channel-wise Noise Scheduled Diffusion for Inverse Rendering in Indoor Scenes/e287505c-edc5-4f3b-9021-f507eb27467f_content_list.json @@ -0,0 +1,1533 @@ +[ + { + "type": "text", + "text": "Channel-wise Noise Scheduled Diffusion for Inverse Rendering in Indoor Scenes", + "text_level": 1, + "bbox": [ + 94, + 130, + 901, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "JunYong Choi $^{1,2}$ Min-cheol Sagong $^{1}$ SeokYeong Lee $^{1,2}$ Seung-Won Jung $^{2}$ Ig-Jae Kim $^{1,3,4}$ Junghyun Cho $^{1,3,4}$", + "bbox": [ + 181, + 178, + 813, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Korea Institute of Science and Technology(KIST) $^{2}$ Korea University $^{3}$ AI-Robotics, KIST School, University of Science and Technology $^{4}$ Yonsei-KIST Convergence Research Institute", + "bbox": [ + 217, + 222, + 779, + 276 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{happily, mcsagong, shapin94, drjay, jhcho}@kist.re.kr swjung83@korea.ac.kr", + "bbox": [ + 176, + 279, + 815, + 292 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7377b628915045238014c3364bbc99a164411f2b4f5b225be147cbd081143e76.jpg", + "image_caption": [ + "Figure 1. Diffusion-Based Inverse Rendering. We present a diffusion-based inverse rendering framework that addresses the two competing goals of accuracy and diversity in inverse rendering by using two distinct models, each dedicated to achieving one of the objectives. (a) Our first model predicts a single accurate solution, while (b) the second model presents diverse possible solutions. (c) This enables practical applications such as object insertion and material editing (e.g., increasing the roughness of the brown table)." + ], + "image_footnote": [], + "bbox": [ + 94, + 319, + 901, + 508 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 585, + 326, + 601 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We propose a diffusion-based inverse rendering framework that decomposes a single RGB image into geometry, material, and lighting. Inverse rendering is inherently ill-posed, making it difficult to predict a single accurate solution. To address this challenge, recent generative model-based methods aim to present a range of possible solutions. However, finding a single accurate solution and generating diverse solutions can be conflicting. In this paper, we propose a channel-wise noise scheduling approach that allows a single diffusion model architecture to achieve two conflicting objectives. The resulting two diffusion models, trained with different channel-wise noise schedules, can predict a single highly accurate solution and present multiple possible solutions. The experimental results demonstrate the superiority of our two models in terms of both diversity and accuracy, which translates to enhanced performance in downstream applications such as object insertion and material editing.", + "bbox": [ + 88, + 643, + 485, + 902 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 514, + 633, + 643, + 648 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Single-image inverse rendering is a long-standing challenge in computer vision and computer graphics that involves the decomposition of an RGB image into its constituent elements of material, lighting, and geometry. This technology enables a wide range of applications, including object insertion, material editing, and relighting. Although inverse rendering has a clear solution, interpreting a single observed radiance can be complex, as it may result from various combinations of material, geometry, and lighting. For instance, a high radiance might be due to the material being white, intense lighting, or a pronounced specular component caused by the Fresnel effect. This inherent ill-posedness of inverse rendering introduces ambiguity and difficulty, particularly when working with a single-image. Therefore, it is important not only to predict an optimal solution, but also to propose multiple potential solutions. However, as discussed in", + "bbox": [ + 511, + 657, + 906, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "5773", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "other studies [15, 41], enforcing only the reconstruction loss cannot improve the sample diversity, leading to conflicts between the objectives.", + "bbox": [ + 89, + 90, + 480, + 136 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Learning-based methods [4, 5, 19, 47, 48] leveraging large-scale synthetic datasets have demonstrated excellent performance. Strong supervised signals from reconstruction losses on well-aligned ground truths have enabled these methods to effectively decompose scenes. However, because these methods are based on deterministic approaches, they do not account for the solution space of possible combinations. Recently, inverse rendering studies based on Latent Diffusion Models (LDM) [28, 31] have emerged [14, 22, 43]. These probabilistic methods, which rely on priors learned from large image-text datasets, not only provide multiple solutions but also exhibit excellent predictive capability. However, they often overlook the trade-off between sample diversity and prediction accuracy and they struggle to account for the dependencies between modalities. Additionally, they rely on the pretrained three-channel autoencoder of LDM, which is effective at generating image-based data such as RGB images but falls short in handling high-dimensional data like per-pixel lighting, which is crucial for applications such as photorealistic object insertion and material editing.", + "bbox": [ + 91, + 138, + 482, + 455 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "While probabilistic methods are effective for capturing multiple potential solutions in ill-posed problems like inverse rendering, they can be sub-optimal for predicting a definitive ground truth. We found that diffusion models exhibit a trade-off between diversity and accuracy depending on the entire timestep $T$ . Motivated by these observations, we propose two conditional diffusion models, each focusing on either diversity or accuracy. Our diffusion models predict multiple modalities, such as a depth map, normal map, diffuse albedo map, roughness map, and per-pixel environment map, simultaneously from a single RGB image. The first proposed model, the Progressive Diffusion Model (PDM), focuses on diversity using a large $T$ for training. The second model, the Switchable Diffusion Model (SDM), is trained with a smaller $T$ to prioritize accuracy. To improve the quality of inverse rendering, a channel-wise noise scheduling approach is introduced to adjust the transition between modalities during generation.", + "bbox": [ + 89, + 458, + 482, + 729 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our PDM and SDM operate in a low-resolution pixel space rather than a latent space, allowing them to handle per-pixel environment maps. To mitigate the high computational costs associated with handling high-dimensional data directly, we introduce a method for encoding these maps into compact feature vectors. Additionally, we present a simple RGB-guided super-resolution model for up-sampling outputs.", + "bbox": [ + 89, + 732, + 482, + 852 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We evaluated the performance by comparing it with other inverse rendering baselines on synthetic [4] and real-world [9, 37, 40] datasets. The experimental results showed", + "bbox": [ + 89, + 854, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "that SDM outperforms in images with fewer ambiguous regions, where accurate predictions are more feasible, while PDM excels in complex scenes with numerous ambiguous regions requiring diverse predictions, indicating the complementary nature of SDM and PDM. In summary, our main contributions are as follows:", + "bbox": [ + 511, + 90, + 903, + 180 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. We propose a novel diffusion-based inverse rendering framework that incorporates per-pixel environment maps while considering the correlations between modalities.", + "2. We propose channel-wise noise scheduling, which improves generation quality by adjusting the modality generation transition in multi-modality diffusion models.", + "3. We demonstrate superior performance over existing baselines on both real and synthetic datasets, highlighting the importance of considering both diversity and accuracy in inverse rendering." + ], + "bbox": [ + 511, + 183, + 903, + 333 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 513, + 345, + 661, + 361 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Inverse rendering. Inverse rendering involves the extraction of depth, normal, SVBRDF, and lighting information from RGB images. Initial research concentrated primarily on specific areas, such as shape-from-shading [13, 25, 45] and intrinsic image decomposition [1, 10, 16, 21]. Advances in deep learning have facilitated the study of complex properties, such as SVBRDF [6, 17, 18, 33] and lighting [8, 9, 36, 39]. Li et al. [19, 20] introduced OpenRooms and a spherical-Gaussian lighting model. IRISformer [48] leverages long-range attention through a dense vision transformer. Zhu et al. [47] introduced the refined InteriorVerse dataset and incorporated Monte Carlo ray tracing into the inverse rendering. Choi et al. [4, 5] extended this approach to a multi-view setup. However, all of these methods rely on deterministic models, limiting their ability to propose alternative solutions. Our work diverges by proposing PDM and SDM, capable of generating multiple solutions and surpassing deterministic models in flexibility and accuracy.", + "bbox": [ + 511, + 372, + 903, + 643 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Diffusion models. Diffusion models [7, 11, 35], which have recently emerged as powerful tools for image generation, are now being applied to various tasks. Several studies have trained diffusion models directly on target datasets to generate materials [38], distant lighting [23, 27], and human face reflectance [26]. Recently, research has begun exploring inverse rendering by utilizing the strong priors of pretrained LDMs. Kocsis et al. [14] fine-tuned an LDM on synthetic datasets, enabling the generation of multiple plausible material maps. RGB $\\leftrightarrow$ X [43] demonstrated that LDMs can be used for both inverse rendering and rendering. Luo et al. [22] trained ControlNet [44] to predict multiple intrinsic modalities. However, since these methods predict each modality independently, they fail to address inter-modality dependencies and struggle with environment maps, limiting realistic lighting inference and applications. In contrast, our framework accounts for these dependencies, enhancing", + "bbox": [ + 511, + 643, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "5774", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/5947288bcd84badd1106de1e993bcd22acee452e3816230a4ceefbc08002b84e.jpg", + "image_caption": [ + "Figure 2. Entire pipeline. Our channel-wise noise scheduling assists the DM's inference by adjusting the transitions between modalities based on the timestep. Since the DM operates at a low resolution, modalities are up-sampled through our SRM." + ], + "image_footnote": [], + "bbox": [ + 91, + 85, + 903, + 234 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "sample fidelity, while per-pixel lighting enables realistic, high-fidelity object insertion.", + "bbox": [ + 88, + 300, + 480, + 333 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 343, + 183, + 359 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given an input RGB image, denoted as $\\mathrm{I} \\in \\mathbb{R}^{3 \\times H \\times W}$ , our goal is to obtain the following outputs: a normal map $\\mathrm{N} \\in \\mathbb{R}^{3 \\times H \\times W}$ , a depth map $\\mathrm{D} \\in \\mathbb{R}^{H \\times W}$ , a diffuse albedo map $\\mathrm{A} \\in \\mathbb{R}^{3 \\times H \\times W}$ , a roughness map $\\mathrm{R} \\in \\mathbb{R}^{H \\times W}$ , and a per-pixel environment map, represented as $\\mathrm{E} \\in \\mathbb{R}^{3 \\times C \\times H \\times W}$ where $H$ and $W$ represent the image dimensions, and $C$ denotes the angular resolution of the environment map. First, we introduce the synthetic dataset used in our experiments. Second, we introduce a method to encode per-pixel lighting into a neural feature vector. Next, we introduce the conditional diffusion model that serves as the foundation of our approach, along with two channel-wise noise scheduling techniques that allow us to customize the model for our desired application. Since the diffusion model operates in the low-resolution pixel space, we introduce a RGB-guided super-resolution model for upsampling. The entire pipeline is shown in Fig. 2.", + "bbox": [ + 89, + 368, + 483, + 626 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Dataset", + "text_level": 1, + "bbox": [ + 89, + 636, + 187, + 648 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To train our conditional diffusion model, we need well-aligned ground truth scene information (N, D, A, R, E) pairs for RGB images. For this purpose, two synthetic datasets, OpenRooms [20] and InteriorVerse [47], are suitable; however, the lighting part of InteriorVerse is currently not available. Additionally, OpenRooms often contains overly simplistic images (such as views of walls or incorrectly placed light sources), which hinder the diffusion model from learning the data distribution. Therefore, we used the OpenRooms FF dataset [4], which involves filtering out inappropriate images from OpenRooms.", + "bbox": [ + 89, + 657, + 482, + 824 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Implicit Lighting Representation", + "text_level": 1, + "bbox": [ + 89, + 832, + 380, + 849 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As a method for handling high-dimensional data such as per-pixel environment maps, a recent method called MAIR++ [5] introduced the implicit lighting representation", + "bbox": [ + 89, + 854, + 483, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "(ILR), which represents environment maps as neural feature vectors. MAIR++ processes the RGB, geometry, and material using a CNN to obtain the ILR, which is subsequently processed by three different decoders to obtain an environment map, shading map, and specular radiance. This approach can render images in a realistic and computationally efficient way. We are motivated to use the ILR as our lighting representation. However, since the previous ILR [5] is not inferred from the environment map, it neither accurately reproduces the ground truth lighting nor effectively captures the detailed high-dynamic range (HDR) radiance values. In contrast, we re-design the ILR as an MLP-based autoencoder that takes the environment map as input.", + "bbox": [ + 511, + 300, + 906, + 496 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Since the HDR radiance values of the environment map cover a very wide range, using them directly as input causes the network to diverge easily. Therefore, we transform the input into the log space using $\\log 1p$ [29] and add batch normalization to the encoder MLP. These two simple techniques allow the network to handle HDR lighting more effectively, leading to significantly improved performance. Mathematically, the encoder is defined as follows:", + "bbox": [ + 511, + 498, + 908, + 619 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {f} = \\tanh (\\operatorname {E n c o d e r} (\\log 1 p (\\bar {\\mathrm {E}}))), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 602, + 640, + 906, + 659 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\bar{\\cdot}$ represents the ground truth, and $f\\in \\mathbb{R}^{96}$ is the ILR neural feature vector. To enable realistic rendering through $f$ , the three decoders are applied:", + "bbox": [ + 511, + 669, + 906, + 717 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {E} = \\exp \\operatorname {m 1} (\\operatorname {D e c o d e r} _ {E} (\\boldsymbol {f}))\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 727, + 756, + 743 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {S} = \\exp \\operatorname {m 1} (\\operatorname {D e c o d e r} _ {S} (\\boldsymbol {f})), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 568, + 746, + 903, + 762 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {I} _ {s} = \\exp \\operatorname {m 1} (\\operatorname {D e c o d e r} _ {I} (\\boldsymbol {f}, \\bar {\\mathrm {R}}, \\gamma (\\boldsymbol {r}), \\bar {\\mathrm {N}} \\cdot \\boldsymbol {v})).\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 763, + 851, + 781 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathsf{expm1}$ [29] is the inverse of $\\log 1p$ , $S \\in \\mathbb{R}^{3 \\times H \\times W}$ denotes the shading map, $I_s \\in \\mathbb{R}^{3 \\times H \\times W}$ represents the specular image, $\\gamma$ refers to the frequency positional encoding, and $\\boldsymbol{r}$ and $\\boldsymbol{v}$ indicate the reflection direction and the viewing direction, respectively. Through these three decoders, $\\boldsymbol{f}$ can not only reconstruct the environment map but also be used to render arbitrary views.", + "bbox": [ + 511, + 792, + 906, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "5775", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/b03deee3066ac285ec0abd47e63d15c86ed50386fe7cece914580869cf5ad2b7.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
TNLDLALRLELIL
MSE ↓var ↑MSE ↓var ↑MSE ↓var ↑MSE ↓var ↑MSE ↓var ↑MSE ↓var ↑
10002.2360.1830.0650.0200.4180.0345.5180.10612.6679.610.1790.042
1002.2380.1820.0640.0190.4130.0345.3540.09412.6677.620.1770.041
102.0260.1380.0590.0130.3740.0245.4400.06612.1362.380.1620.03
11.8170.00.0610.00.3340.05.5450.011.370.060.1370.0
", + "bbox": [ + 99, + 88, + 898, + 176 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Table 1. Ablation of training $T$ . The best values in each column are highlighted in bold, and the worst values are underlined. In most cases, as the value of $T$ decreases, prediction accuracy improves, but the diversity of the samples is reduced. To evaluate under general conditions, we fixed $\\tau$ to 1.0 for all settings, and the DDIM [34] steps were selected as 2, 2, 2, and 1, respectively.", + "bbox": [ + 88, + 186, + 906, + 231 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Diffusion-based Inverse Rendering", + "text_level": 1, + "bbox": [ + 89, + 255, + 393, + 272 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Since the introduction of LDM [28, 31], many diffusion-based methods [14, 22, 43] have used pretrained LDMs that have strong priors for images. However, these methods do not provide meaningful priors for high-dimensional data such as environment maps. Therefore, we adopt a cascaded structure [12] that first generates $\\mathbf{z}_0 = (\\mathrm{N}_L,\\mathrm{D}_L,\\mathrm{A}_L,\\mathrm{R}_L,\\boldsymbol {f}_L)$ and then increases the resolution, where the subscript $L$ denotes a low-resolution image of size $H_{L}\\times W_{L}$ . During inference, our conditional diffusion model (DM) takes $\\mathrm{I}_L$ as a condition and iteratively denoises the Gaussian noise image $\\mathbf{z}_{T - 1}\\sim \\mathcal{N}(0,1)$ to produce $\\mathbf{z}_0$ . During training, we predict the velocity $\\mathbf{v}_t$ [32], and the loss function $L_{\\mathrm{DM}}$ is given by:", + "bbox": [ + 91, + 277, + 483, + 474 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\mathrm {D M}} = \\left\\| \\mathbf {v} _ {t} - \\mathrm {D M} \\left(t, \\mathbf {z} _ {t}, \\mathrm {I} _ {L}, \\operatorname {C L I P} \\left(\\mathrm {I} _ {L}\\right)\\right) \\right\\| _ {2} ^ {2}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 145, + 487, + 482, + 506 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {v} _ {t} = \\sqrt {\\bar {\\alpha} _ {t}} \\mathbf {z} _ {T - 1} - \\sqrt {1 - \\bar {\\alpha} _ {t}} \\mathbf {z} _ {0}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 183, + 517, + 482, + 534 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $T$ is the entire timestep, $t \\in [0, T - 1]$ is uniformly sampled during training, $\\bar{\\alpha}_t$ is the noise scheduler, $\\mathbf{z}_t = \\sqrt{\\bar{\\alpha}_t} \\mathbf{z}_0 + \\sqrt{1 - \\bar{\\alpha}_t} \\mathbf{z}_{T-1}$ , and CLIP is the pretrained image encoder [30]. $\\mathrm{I}_L$ is concatenated with $\\mathbf{z}_t$ , and the CLIP embedding is used as a cross-attention condition.", + "bbox": [ + 89, + 537, + 482, + 613 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Channel-wise noise scheduling. Since we generate multiple interdependent modalities simultaneously, we hypothesize that there is an optimal transition to generate these modalities in inverse rendering. For example, according to previous works [18, 42, 46], geometry typically has the least uncertainty and supports the prediction of material and lighting, so generating geometry first could be advantageous. On the other hand, since lighting has the highest uncertainty, generating it last could be beneficial. To validate this hypothesis, we divide $\\mathbf{z}_0$ into three categories based on the modality attributes: geometry $(\\mathrm{N}_L,\\mathrm{D}_L)$ , material $(\\mathrm{A}_L,\\mathrm{R}_L)$ , and lighting $(f_{L})$ , and accordingly apply different noise scheduling to the three parts. Specifically, following Chen et al. [2], we select the cosine scheduler as the base noise scheduler and parameterize it as follows:", + "bbox": [ + 89, + 614, + 482, + 840 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} v _ {s} = \\cos (s \\cdot 2 \\pi) ^ {2 \\tau}, v _ {b} = \\cos (b \\cdot 2 \\pi) ^ {2 \\tau} \\\\ \\alpha_ {t} = \\cos \\left(\\left(\\left(b - s\\right) \\cdot t + s\\right) \\cdot 2 \\pi\\right) ^ {2 \\tau}, \\bar {\\alpha} _ {t} = \\frac {v _ {b} - \\alpha_ {t}}{v _ {b} - v _ {s}}. \\tag {5} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 109, + 853, + 482, + 905 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Throughout the paper, we fix $s = 0.008$ and $b = 1$ and control the scheduler using only $\\tau$ . When $\\tau = 1$ , $\\bar{\\alpha}_t$ is equivalent to the cosine scheduler. The behavior of the noise scheduler as a function of $\\tau$ is shown in Fig. 3.", + "bbox": [ + 511, + 256, + 906, + 316 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/98d4a95e6e7b27792db66e13204d66f1c7569a2af432ef5974ed70517122b0c3.jpg", + "image_caption": [ + "Figure 3. Parameterizing noise scheduler. A larger $\\tau$ value results in a slower generation of the modality." + ], + "image_footnote": [], + "bbox": [ + 514, + 328, + 903, + 454 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Fig. 3 shows that as the value of $\\tau$ increases, the noise increases more gradually. To determine the optimal modality generation order using $\\tau$ , we conducted experiments by applying different values of $\\tau$ to geometry, material, and lighting, meaning that different noise levels are applied separately to each channel of the data during the denoising process of the diffusion model. The ablation study in Tab. 7 confirms that generating in the order of geometry, material, and lighting is the most effective, aligning with our intuition. As a result, through our channel-wise noise scheduling, it focuses on specific modalities at different timesteps, enabling more effective predictions of other modalities in subsequent timesteps.", + "bbox": [ + 511, + 508, + 906, + 703 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In general, $T$ in diffusion models is set to a large number, such as 1000. This helps approximate the data distribution and increases the diversity of the generated data [24]. However, we note that such diffusion models are challenging in predicting a single definitive ground truth. To this end, we conducted an ablation study on $T$ values by training the DM with $T = 1000, 100, 10$ , and 1. For each setting, we performed sampling 20 times in the test set to measure the MSE and the variance (var) of the samples. Tab. 1 shows that reducing $T$ improves accuracy at the cost of diversity in most cases. In the extreme case of $T = 1$ , the method becomes deterministic.", + "bbox": [ + 511, + 704, + 908, + 883 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To explore models focused on diversity and accuracy, we", + "bbox": [ + 532, + 885, + 906, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "5776", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "develop two types of diffusion models: one with $T = 1000$ and the other with $T = 1$ . Following the ablation study in Tab. 7, we select $\\tau = (0.9, 1.2, 1.5)$ when $T = 1000$ to prioritize diversity. Due to its gradual generation process, we name this model the Progressive Diffusion Model (PDM). However, this gradual generation process cannot be applied to the diffusion model with $T = 1$ . Moreover, $T = 1$ limits the ability of the diffusion model to exploit its full potential over multiple timesteps [3]. To this end, we define our new noise scheduler, as shown in Tab. 2.", + "bbox": [ + 89, + 90, + 483, + 241 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/05fe74ef7c818670f5cb3079446dbb7b26e89289abb0b2e725a1228c7aaa5738.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
tt mod 3 = 0t mod 3 = 1t mod 3 = 2T-1
\\( \\bar{a}_{t}^{G} \\)1100
\\( \\bar{a}_{t}^{M} \\)1010
\\( \\bar{a}_{t}^{L} \\)0110
", + "bbox": [ + 102, + 253, + 470, + 306 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "A value of 0 implies that pure Gaussian noise is added, while a value of 1 means that no noise is added. With our new noise scheduler, $T$ is redefined as $3n + 1$ where $n > 0$ . After the initial prediction at $t = T - 1$ , the noise is switched per channel, providing conditions for the next diffusion step. Taking into account this feature of the switchable noise scheduling, we name this model Switchable Diffusion Model (SDM). The SDM is trained to predict the modality with an SNR of 0 at each step, allowing for more refined predictions by utilizing the previous step's predictions.", + "bbox": [ + 89, + 364, + 483, + 517 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. RGB-Guided Super Resolution", + "text_level": 1, + "bbox": [ + 89, + 526, + 366, + 542 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We use a UNet as the RGB-guided super-resolution model, consisting of a shared encoder and five separate decoders. The encoder processes the RGB image into a dense feature map, which, along with the rendered diffuse $\\mathrm{I}_d$ and specular $\\mathrm{I}_s$ images from $\\mathbf{z}_0$ , is fed into each decoder. Our RGB-guided Super-Resolution Model (SRM) is as follows:", + "bbox": [ + 89, + 547, + 483, + 640 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {N}, \\mathrm {D}, \\mathrm {A}, \\mathrm {R}, \\boldsymbol {f} = \\operatorname {S R M} \\left(\\mathbf {z} _ {0}, \\mathrm {I}, \\mathrm {I} _ {d}, \\mathrm {I} _ {s}\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 648, + 482, + 664 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Evaluation", + "text_level": 1, + "bbox": [ + 89, + 678, + 207, + 694 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we evaluate our method on both synthetic and real-world datasets. We first evaluate on a synthetic dataset, followed by inverse rendering and object insertion on a real-world spatially varying lighting dataset. Furthermore, we evaluated our method on the sub-tasks of inverse rendering, namely intrinsic decomposition and geometry prediction. More details on training, model architectures, and additional experimental results are provided in the Supplementary Materials.", + "bbox": [ + 89, + 704, + 482, + 839 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation settings. We used $C = 128$ , $H = 240$ , $W = 320$ , $H_{L} = 60$ , and $W_{L} = 80$ . For SDM, we set $T = 4$ . In all experiments, PDM generated 10 samples and the mean and variance of these samples were evaluated.", + "bbox": [ + 89, + 839, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Baselines. We compared our method with state-of-the-art inverse rendering frameworks capable of per-pixel lighting inference, specifically those by Li et al. [19] and Zhu et al. [47]. Both methods were trained on the OpenRooms FF [4] dataset. However, since the code for Zhu's lighting model has not been released, we used their pretrained lighting model, which was trained on the InteriorVerse dataset.", + "bbox": [ + 511, + 90, + 905, + 196 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/c53b69649ece6843a6ea562e8e28547d1e8a0cc5e3b784f70ccc5ff248faa6f8.jpg", + "image_caption": [ + "Figure 4. Qualitative synthetic evaluation. Only our method successfully decomposes specular radiance into material and lighting." + ], + "image_footnote": [], + "bbox": [ + 516, + 210, + 903, + 354 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Evaluation on Synthetic Scenes", + "text_level": 1, + "bbox": [ + 511, + 424, + 790, + 439 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Tab. 3 includes quantitative prediction results on the synthetic [4] test dataset. Although our target is single-view, we include results from the multi-view baselines MAIR [4] and MAIR++ [5]. Our SDM, which prioritizes prediction accuracy, demonstrated superior performance compared to existing single-image baselines in all areas except roughness, and our PDM also demonstrated respectable performance comparable to the baselines. Notably, our lighting accuracy was comparable to that of MAIR, and the albedo accuracy outperformed even the multi-view methods. Additionally, it can be observed that both MAIR++ and our method, which are capable of ILR-based neural rendering, exhibit significantly superior re-rendering (I) performance. Figure 4 presents an illustrative example of this fact. Even in images with strong specular radiance, our SDM and PDM successfully decomposed and re-rendered the image in contrast to the failure of the baseline methods.", + "bbox": [ + 511, + 446, + 906, + 702 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/fd55a6fd3f6bacf37eea4c71f7c5abdb1ecd9aab50cc5cf3ce9eb949e31ca568.jpg", + "table_caption": [ + "Table 2. SDM noise scheduler. Here, $\\bar{\\alpha}_t^G$ , $\\bar{\\alpha}_t^M$ , and $\\bar{\\alpha}_t^L$ are the schedulers for geometry, material, and lighting, respectively." + ], + "table_footnote": [], + "table_body": "
MSE(×10-2)N ↓D ↓A ↓R ↓E ↓I ↓
Li et al. [19]3.7820.1350.8696.49615.490.491
Zhu et al. [47]1.7380.0620.5284.208*31.15*3.565
SDM1.5020.0390.3494.96913.350.156
PDM(Mean)1.8310.0470.3944.81914.080.150
PDM(Best)1.7820.0440.3814.64413.750.146
MAIR [4]1.0100.1400.6393.45713.790.784
MAIR++ [5]1.0100.0070.4782.49711.610.058
", + "bbox": [ + 517, + 715, + 903, + 813 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 3. Quantitative synthetic evaluation. Bold text indicates the best performance among single-view methods, while underlined text represents the second-best. * denotes models trained on the InteriorVerse dataset.", + "bbox": [ + 511, + 824, + 905, + 878 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5777", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/52170fac97f4b9667a11d36f51110a77e1637bc6bd17f40d9dc8936eb6617719.jpg", + "image_caption": [ + "Input Image" + ], + "image_footnote": [], + "bbox": [ + 91, + 87, + 248, + 169 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/34e2a9088aa5e6a47585e3a46c69df49b0eda539371e05038eccf396e906c58a.jpg", + "image_caption": [ + "N" + ], + "image_footnote": [], + "bbox": [ + 243, + 88, + 429, + 253 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b120857f00227edf41eb9d88dd7888366ca5d6cd79c9dcd1a9faa766cbcd520a.jpg", + "image_caption": [ + "A" + ], + "image_footnote": [], + "bbox": [ + 429, + 88, + 586, + 253 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c94dcc6c1a73d9888ed29b4e3044b29f32168c0f8995a7ea927caaeebc566e42.jpg", + "image_caption": [ + "Id" + ], + "image_footnote": [], + "bbox": [ + 586, + 88, + 745, + 253 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e743fb28423b44050dcfd35343fd145f0e4c2b5ce49893f412ff7a17cb791b83.jpg", + "image_caption": [ + "I" + ], + "image_footnote": [], + "bbox": [ + 745, + 88, + 905, + 253 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f9c98b41c00b6411f5738a01c3cb455eee3017fbe5df5595772a64d789f2416b.jpg", + "image_caption": [ + "Figure 5. Qualitative real-world evaluation. Our method successfully recognizes shadows under the desk, achieving plausible geometry prediction and realistic re-rendering results.", + "Li et al." + ], + "image_footnote": [], + "bbox": [ + 91, + 320, + 295, + 526 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1bfaa68d1c4334484df848afbddcce15da9448ec4a0055e34dfea64971f4e484.jpg", + "image_caption": [ + "Zhu et al." + ], + "image_footnote": [], + "bbox": [ + 295, + 320, + 498, + 526 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/59d547382ef8f11af670ea66072873a7c00d6fbe9a6e2a24d52a71cde35bc479.jpg", + "image_caption": [ + "Ours(SDM)", + "Figure 6. Qualitative object insertion evaluation. We accurately predicted the direction of light and shadows even in real-world images. Additionally, our rendered objects were closer to the ground truth in terms of hue and brightness compared to the baselines." + ], + "image_footnote": [], + "bbox": [ + 500, + 320, + 702, + 527 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e225683f39f3c94a1947dcefa2d163877e8f7d76c1d811baf96694a8d972d6d0.jpg", + "image_caption": [ + "RGB $\\leftrightarrow$ X" + ], + "image_footnote": [], + "bbox": [ + 704, + 320, + 905, + 526 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Evaluation on Real-world Scenes", + "text_level": 1, + "bbox": [ + 89, + 604, + 380, + 619 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Qualitative evaluation. We evaluated the generalization capability of our method on real-world images from Garon et al. [9]. As shown in Fig. 5, both SDM and PDM successfully separated the albedo even in the presence of strong shadows under the desk. Li et al. [19] partially removed the shadows but failed to predict the geometry, while Zhu et al. [47] accurately predicted the normal but failed to estimate the albedo.", + "bbox": [ + 89, + 628, + 483, + 748 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/ddf81cd8ff4827dccbb0732c42d8ca2819af04103830f7b88510bfa051ad2ec3.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Li et al. [19]Zhu et al. [47]SDMGround Truth
0.1780.1530.3220.348
", + "bbox": [ + 114, + 762, + 460, + 792 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 4. AMT user study on object insertion. Our method was selected by nearly as many users as the ground truth.", + "bbox": [ + 89, + 804, + 482, + 833 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Lighting Evaluation. We also evaluated lighting performance through object insertion, using the object insertion tool from Li et al. [19]. For fairness, we excluded PDM", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "and focused on evaluating only SDM. As shown in Fig. 6, our method accurately predicts strong directional lighting, as well as the surrounding light's tone and shadows, resulting in the most realistic object insertions. In contrast, Zhu et al. [47] failed to predict the correct light direction, while Li et al. [19] struggles with realistic rendering due to a lack of high-frequency details in the environment map. The diffusion-based method RGB $\\leftrightarrow$ X [43] relies on inpainting for object insertion, which compromises the fidelity of the inserted objects and leads to visible artifacts. Additionally, since it cannot handle environment maps, it struggled to accurately predict the direction of light. For a more quantitative evaluation, we conducted an AMT user study on object insertion, including the ground truth. The results in Tab. 4 confirmed the superiority of our method. Users were asked: \"Of the four images, please select the one that shows white rabbits blending naturally into the surrounding light.\" Our method was selected by users nearly as frequently as the ground truth.", + "bbox": [ + 511, + 604, + 906, + 893 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "5778", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1bab5064a7f194e7f390e7ae32edf7cb716177c71e87514ee173925c3ea94c30.jpg", + "image_caption": [ + "Figure 7. Intrinsic decomposition. Our method successfully decomposes the input image into albedo and shading. RGB $\\leftrightarrow$ X [43], leveraging its strong prior, achieves successful albedo prediction but struggles with realistic rendering, resulting in excessive brightness in the left sample and incorrect specular highlights in the right sample." + ], + "image_footnote": [], + "bbox": [ + 94, + 87, + 504, + 238 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c0dee3c1dfe2482606782c0c655f53590dd6bb0432acacbfc2fb17c4a525cb33.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 506, + 87, + 905, + 239 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7ad47f7574d1da95d022fcf6feb415f3b3d70a58102718caebca2fecb497fd8f.jpg", + "image_caption": [ + "Figure 8. Geometry prediction. Our method achieves the most accurate prediction of the ceiling structure." + ], + "image_footnote": [], + "bbox": [ + 94, + 314, + 480, + 494 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/a00aaf44cf2954677bea0c4ece36e461ab089f9e7373df5fa3f386730a531335.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodA ↓Hue ↓Texture ↓WHDR ↓S ↓I ↓
Li et al. [19]0.9245.7810.4730.3011.1620.021
Zhu et al. [47]1.6854.6380.4720.3011.4410.028
SDM0.7055.0560.4590.2931.1810.006
PDM(Mean)0.7575.2600.4590.3061.1560.005
PDM(Best)0.6234.5270.4560.2781.1130.005
RGB↔X [43](Mean)0.6023.6900.4480.1681.2610.055
RGB↔X [43](Best)0.3313.0910.4410.1461.1960.052
", + "bbox": [ + 94, + 547, + 478, + 635 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Evaluation on Sub-tasks", + "text_level": 1, + "bbox": [ + 89, + 695, + 313, + 709 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Intrinsic decomposition. We evaluated our method on intrinsic decomposition using the recently introduced MAW [40] dataset, which provides more refined annotations for 888 images, allowing more quantitative comparisons than the WHDR metric used in the IIW [1] dataset. Tab. 5 presents comparisons for MSE for A, S, and I, along with metrics for hue, texture, and WHDR for A. The hue metric evaluates the chromaticity component, while the texture metric is based on LPIPS perceptual similarity. Our method generally outperformed existing deterministic methods [19, 47]. RGB $\\leftrightarrow$ X [43], which learns strong priors from diverse large datasets, excels in albedo infer", + "bbox": [ + 89, + 719, + 483, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ence; however, due to its reliance on diffusion-based rendering, it reconstructed the input images with the lowest accuracy. In contrast, our method, trained only on OpenRooms FF, achieved the highest performance in shading estimation and most faithfully reproduced input images, demonstrating high fidelity. This can also be observed in Fig. 7. In the left sample, RGB $\\leftrightarrow$ X [43] fails to predict the shading intensity as it does not account for HDR lighting. Similarly, in the right sample, RGB $\\leftrightarrow$ X [43] introduces specular highlights on the desk that are absent in the input image, resulting in a loss of fidelity. In contrast, our method predicts spatially consistent albedo and accurately estimates brightness, enabling a more realistic reproduction of the input image.", + "bbox": [ + 511, + 316, + 906, + 513 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Geometry prediction. Tab. 6 includes the geometry prediction evaluation on the DIODE [37] indoor dataset. Similarly to the intrinsic decomposition task, both our PDM and SDM demonstrate superior or comparable performance to the baselines. Fig. 8 provides qualitative examples, demonstrating that our method exhibits strong generalization capability in geometry prediction.", + "bbox": [ + 511, + 515, + 908, + 621 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/7f390b0700e25580a6ff93f41cbb9a3ccb36b7f78da7bd3bce13a9a12253a47f.jpg", + "table_caption": [ + "Table 5. Intrinsic decomposition evaluation on MAW [40]. Bold text indicates the best performance." + ], + "table_footnote": [], + "table_body": "
MetricLi et al. [19]Zhu et al. [47]SDMPDM(Mean)PDM(Best)
N ↓0.1960.1710.1470.1520.144
D ↓0.0590.0570.0570.0550.050
I ↓0.0150.0130.0030.0020.002
", + "bbox": [ + 517, + 633, + 901, + 681 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 6. Geometry prediction evaluation on DIODE [37]. Bold text indicates the best performance.", + "bbox": [ + 511, + 684, + 906, + 712 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Modality Dependencies", + "text_level": 1, + "bbox": [ + 511, + 741, + 728, + 758 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To generate a reasonable solution for inverse rendering, it is crucial to account for the dependencies between modalities. For instance, if a material is predicted to be darker, the lighting should be predicted brighter to compensate for the input image. However, existing diffusion-based methods [14, 22, 43] infer each modality independently, making it difficult to capture such dependencies. In contrast, our PDM generates all modalities simultaneously, enabling it to account for these relationships. Fig. 9 illustrates this with", + "bbox": [ + 511, + 763, + 908, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "5779", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/91cf3562a8c816a8758fe3db0188570cfeea05733ccf70ee790f2a8079f7db5e.jpg", + "image_caption": [ + "Figure 9. Modality dependencies. Two samples of N, D, A, S, and R predicted by PDM." + ], + "image_footnote": [], + "bbox": [ + 93, + 87, + 482, + 238 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "two different samples generated by PDM. For the center table in the top image, the first sample predicts the normal pointing upward with bright shading, resulting in a darker albedo. Conversely, the second sample predicts the normal facing the camera, producing darker shading and a brighter albedo. Additionally, for the area under the cabinet in the bottom image, the first sample interprets it as a shadow and predicts a bright albedo, while the second sample treats it as part of the albedo and generates bright shading.", + "bbox": [ + 88, + 301, + 483, + 439 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Ablations", + "text_level": 1, + "bbox": [ + 89, + 448, + 202, + 462 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Modality generation order. To explore the optimal modality generation order, we conducted an ablation study on the modality generation sequence. There are no restrictions on $\\tau$ ; however, we selected three values: $\\tau = 0.9, 1.2,$ and 1.5. These values were applied separately to $\\tau_{G}$ (geometry), $\\tau_{M}$ (material), and $\\tau_{L}$ (lighting), and the models were trained independently. The performance was then evaluated on the test dataset. Tab. 7 shows that this channel-wise noise scheduling can significantly impact the inverse rendering quality. In particular, $\\tau_{G}, \\tau_{M}, \\tau_{L} = (0.9, 1.2, 1.5)$ produces the best quality, aligning with our intuition. In contrast, generating in reverse order results in a detrimental effect. Fig. 10 provides examples of these two cases. The reverse order (L-M-G) exhibits artifacts in areas with strong shadows, whereas our method produces plausible results.", + "bbox": [ + 89, + 470, + 483, + 698 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/fd15152b5dc1a362d2549bb22ce1d2541b45b2c1917e19ce5383ae2ec6f68fd2.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
τGτMτLNL↓DL↓AL↓RL↓EL↓IL↓
1.01.01.02.1400.0620.4035.75412.190.163
0.91.21.52.0240.0590.3765.17611.940.159
0.91.51.22.1040.0600.3905.72312.150.165
1.20.91.52.0730.0600.3885.39412.200.161
1.21.50.92.2020.0640.4115.88212.330.167
1.50.91.22.2110.0640.4055.66412.360.160
1.51.20.92.2450.0660.4145.98612.530.166
", + "bbox": [ + 94, + 709, + 478, + 814 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Iterative Prediction of SDM. We conducted an ablation study to evaluate the benefits of iterative prediction for", + "bbox": [ + 89, + 869, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/86f3fd7445c5b2da785fd0a4f9e4f9bd156c25fb5bd13c4a3828960cbb251844.jpg", + "image_caption": [ + "Figure 10. Comparison of modality generation order. Generating in the order of geometry, material, and lighting helps effectively infer shadows." + ], + "image_footnote": [], + "bbox": [ + 516, + 88, + 905, + 212 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/0c2795bc08dbb2f6b4587649dba7ad6c388e7c8b9e5811efe9f6472f284a9a3b.jpg", + "table_caption": [ + "Table 7. Ablation of the generation order. The best values are highlighted in bold, while the worst values are underlined." + ], + "table_footnote": [], + "table_body": "
TN ↓D ↓A ↓R ↓E ↓I ↓
11.7330.0600.3185.83711.240.132
41.5780.0520.2885.74310.960.137
71.6300.0500.2925.75910.970.139
", + "bbox": [ + 537, + 279, + 880, + 338 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 8. Ablation of SDM iterative prediction. Bold text indicates the best performance.", + "bbox": [ + 511, + 348, + 905, + 377 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "SDM. In Tab. 8, the improvements at $T = 4$ and $T = 7$ over $T = 1$ indicate that iterative prediction utilizes previously predicted modalities to inform subsequent predictions. However, the performance drop at $T = 7$ compared to $T = 4$ suggests that excessive iterations do not always yield better results.", + "bbox": [ + 511, + 404, + 906, + 494 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 511, + 633, + 526 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We tackled inverse rendering with two opposing goals: attaining diversity and ensuring accuracy. To address these objectives, we proposed a diffusion-based framework, in which channel-wise noise scheduling allowed each diffusion model to better understand the correlations between different modalities.", + "bbox": [ + 511, + 537, + 905, + 626 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations and future work. Our approach has the limitation of requiring two separate models; thus, integrating them into a unified model would enhance usability. We explored re-rendering error as an alternative metric to overcome the constraint of depending only on the predicted variance of PDM for estimating inverse rendering uncertainty, but it could be imprecise. Further research on the handling of uncertainty could improve the complementarity of SDM and PDM. Moreover, our method refrains from exploiting the priors of pretrained LDMs. Using those priors to process per-pixel environment maps might enhance the performance and expand the applications of LDMs.", + "bbox": [ + 511, + 628, + 905, + 808 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements. This work was supported by Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government(MSIT)(No.RS-2023-00227592, Development of 3D Object Identification Technology Robust to Viewpoint Changes)", + "bbox": [ + 511, + 810, + 905, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "5780", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Sean Bell, Kavita Bala, and Noah Snively. Intrinsic images in the wild. ACM Transactions on Graphics (TOG), 33(4): 1-12, 2014. 2, 7", + "[2] Ting Chen. On the importance of noise scheduling for diffusion models. arXiv preprint arXiv:2301.10972, 2023. 4", + "[3] Jooyoung Choi, Jungbeom Lee, Chaehun Shin, Sungwon Kim, Hyunwoo Kim, and Sungroh Yoon. Perception prioritized training of diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11472-11481, 2022. 5", + "[4] JunYong Choi, SeokYeong Lee, Haesol Park, Seung-Won Jung, Ig-Jae Kim, and Junghyun Cho. Mair: multi-view attention inverse rendering with 3d spatially-varying lighting estimation. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8392-8401. IEEE, 2023. 2, 3, 5", + "[5] JunYong Choi, SeokYeong Lee, Haesol Park, Seung-Won Jung, Ig-Jae Kim, and Junghyun Cho. Mair++: Improving multi-view attention inverse rendering with implicit lighting representation. arXiv preprint arXiv:2408.06707, 2024. 2, 3, 5", + "[6] Valentin Deschaintre, Miika Aittala, Fredo Durand, George Drettakis, and Adrien Bousseau. Single-image svbrdf capture with a rendering-aware deep network. ACM Transactions on Graphics (ToG), 37(4):1-15, 2018. 2", + "[7] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 2", + "[8] Marc-Andre Gardner, Yannick Hold-Geoffroy, Kalyan Sunkavalli, Christian Gagne, and Jean-François Lalonde. Deep parametric indoor lighting estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7175–7183, 2019. 2", + "[9] Mathieu Garon, Kalyan Sunkavalli, Sunil Hadap, Nathan Carr, and Jean-François Lalonde. Fast spatially-varying indoor lighting estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6908-6917, 2019. 2, 6", + "[10] Roger Grosse, Micah K Johnson, Edward H Adelson, and William T Freeman. Ground truth dataset and baseline evaluations for intrinsic image algorithms. In 2009 IEEE 12th International Conference on Computer Vision, pages 2335-2342. IEEE, 2009. 2", + "[11] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2", + "[12] Jonathan Ho, Chitwan Sahara, William Chan, David J Fleet, Mohammad Norouzi, and Tim Salimans. Cascaded diffusion models for high fidelity image generation. Journal of Machine Learning Research, 23(47):1-33, 2022. 4", + "[13] Berthold KP Horn and Michael J Brooks. Shape from shading. MIT press, 1989. 2", + "[14] Peter Kocsis, Vincent Sitzmann, and Matthias Nießner. Intrinsic image diffusion for single-view material estimation. arXiv preprint arXiv:2312.12274, 2023. 2, 4, 7" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Soochan Lee, Junsoo Ha, and Gunhee Kim. Harmonizing maximum likelihood with gans for multimodal conditional generation. arXiv preprint arXiv:1902.09225, 2019. 2", + "[16] Zhengqi Li and Noah Snavely. Cginintrinsics: Better intrinsic image decomposition through physically-based rendering. In Proceedings of the European conference on computer vision (ECCV), pages 371-387, 2018. 2", + "[17] Zhengqin Li, Kalyan Sunkavalli, and Manmohan Chandraker. Materials for masses: Svbrdf acquisition with a single mobile phone image. In Proceedings of the European conference on computer vision (ECCV), pages 72-87, 2018. 2", + "[18] Zhengqin Li, Zexiang Xu, Ravi Ramamoorthi, Kalyan Sunkavalli, and Manmohan Chandraker. Learning to reconstruct shape and spatially-varying reflectance from a single image. ACM Transactions on Graphics (TOG), 37(6):1-11, 2018. 2, 4", + "[19] Zhengqin Li, Mohammad Shafiei, Ravi Ramamoorthi, Kalyan Sunkavalli, and Manmohan Chandraker. Inverse rendering for complex indoor scenes: Shape, spatially-varying lighting and svbrdf from a single image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2475-2484, 2020. 2, 5, 6, 7", + "[20] Zhengqin Li, Ting-Wei Yu, Shen Sang, Sarah Wang, Meng Song, Yuhan Liu, Yu-Ying Yeh, Rui Zhu, Nitesh Gundavarapu, Jia Shi, et al. Openrooms: An open framework for photorealistic indoor scene datasets. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7190–7199, 2021. 2, 3", + "[21] Jundan Luo, Zhaoyang Huang, Yijin Li, Xiaowei Zhou, Guofeng Zhang, and Hujun Bao. Niid-net: Adapting surface normal knowledge for intrinsic image decomposition in indoor scenes. IEEE Transactions on Visualization and Computer Graphics, 26(12):3434-3445, 2020. 2", + "[22] Jundan Luo, Duygu Ceylan, Jae Shin Yoon, Nanxuan Zhao, Julien Philip, Anna Frühstück, Wenbin Li, Christian Richardt, and Tuanfeng Wang. Intrinsicdiffusion: Joint intrinsic layers from latent diffusion models. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 2, 4, 7", + "[23] Linjie Lyu, Ayush Tewari, Marc Habermann, Shunsuke Saito, Michael Zollhöfer, Thomas Leimkuhler, and Christian Theobalt. Diffusion posterior illumination for ambiguity-aware inverse rendering. ACM Transactions on Graphics (TOG), 42(6):1-14, 2023. 2", + "[24] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International conference on machine learning, pages 8162-8171. PMLR, 2021. 4", + "[25] Geoffrey Oxholm and Ko Nishino. Shape and reflectance from natural illumination. In European Conference on Computer Vision, pages 528-541. Springer, 2012. 2", + "[26] Foivos Paraperas Papantoniou, Alexandros Lattas, Stylianos Moschoglou, and Stefanos Zafeiriou. Relightify: Relightable 3d faces from a single image via diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8806-8817, 2023. 2" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "5781", + "bbox": [ + 482, + 945, + 513, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[27] Pakkapon Phongthawee, Worameth Chinchuthakun, Nontaphat Sinsunthithet, Varun Jampani, Amit Raj, Pramook Khungurn, and Supasorn Suwajanakorn. Diffusionlight: Light probes for free by painting a chrome ball. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 98-108, 2024. 2", + "[28] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 2, 4", + "[29] William H. Press, Saul A. Teukolsky, William T. Vetterling, and Brian P. Flannery. Numerical Recipes 3rd Edition: The Art of Scientific Computing. Cambridge University Press, 2007. 3", + "[30] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 4", + "[31] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2, 4", + "[32] Tim Salimans and Jonathan Ho. Progressive distillation for fast sampling of diffusion models. arXiv preprint arXiv:2202.00512, 2022. 4", + "[33] Shen Sang and Manmohan Chandraker. Single-shot neural relighting and svbrdf estimation. In European Conference on Computer Vision, pages 85-101. Springer, 2020. 2", + "[34] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 4", + "[35] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. Advances in neural information processing systems, 32, 2019. 2", + "[36] Pratul P Srinivasan, Ben Mildenhall, Matthew Tancik, Jonathan T Barron, Richard Tucker, and Noah Snively. Lighthouse: Predicting lighting volumes for spatially-coherent illumination. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8080-8089, 2020. 2", + "[37] Igor Vasiljevic, Nick Kolkin, Shanyi Zhang, Ruotian Luo, Haochen Wang, Falcon Z Dai, Andrea F Daniele, Mohammadreza Mostajabi, Steven Basart, Matthew R Walter, et al. Diode: A dense indoor and outdoor depth dataset. arXiv preprint arXiv:1908.00463, 2019. 2, 7", + "[38] Giuseppe Vecchio, Renato Sortino, Simone Palazzo, and Concetto Spampinato. Matfuse: controllable material generation with diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4429-4438, 2024. 2", + "[39] Zian Wang, Jonah Philion, Sanja Fidler, and Jan Kautz. Learning indoor inverse rendering with 3d spatially-varying lighting. In Proceedings of the IEEE/CVF International" + ], + "bbox": [ + 91, + 92, + 480, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Conference on Computer Vision, pages 12538-12547, 2021. 2", + "[40] Jiaye Wu, Sanjoy Chowdhury, Hariharmano Shanmugaraja, David Jacobs, and Soumyadip Sengupta. Measured albedo in the wild: Filling the gap in intrinsics evaluation. In 2023 IEEE International Conference on Computational Photography (ICCP), pages 1-12. IEEE, 2023. 2, 7", + "[41] Jiahui Yu, Zhe Lin, Jimei Yang, Xiaohui Shen, Xin Lu, and Thomas S Huang. Free-form image inpainting with gated convolution. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4471-4480, 2019. 2", + "[42] Ye Yu and William AP Smith. Inverserendernet: Learning single image inverse rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3155-3164, 2019. 4", + "[43] Zheng Zeng, Valentin Deschaintre, Iliyan Georgiev, Yannick Hold-Geoffroy, Yiwei Hu, Fujun Luan, Ling-Qi Yan, and Miloš Hasan. Rgb $\\leftrightarrow$ x: Image decomposition and synthesis using material-and lighting-aware diffusion models. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 2, 4, 6, 7", + "[44] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 2", + "[45] Ruo Zhang, Ping-Sing Tsai, James Edwin Cryer, and Mubarak Shah. Shape-from-shading: a survey. IEEE transactions on pattern analysis and machine intelligence, 21(8): 690-706, 1999. 2", + "[46] Xiuming Zhang, Pratul P Srinivasan, Boyang Deng, Paul Debevec, William T Freeman, and Jonathan T Barron. Nerfactor: Neural factorization of shape and reflectance under an unknown illumination. ACM Transactions on Graphics (TOG), 40(6):1-18, 2021. 4", + "[47] Jingsen Zhu, Fujun Luan, Yuchi Huo, Zihao Lin, Zhihua Zhong, Dianbing Xi, Rui Wang, Hujun Bao, Jiaxiang Zheng, and Rui Tang. Learning-based inverse rendering of complex indoor scenes with differentiable monte carlo raytracing. In SIGGRAPH Asia 2022 Conference Papers. ACM, 2022. 2, 3, 5, 6, 7", + "[48] Rui Zhu, Zhengqin Li, Janarbek Matai, Fatih Porikli, and Manmohan Chandraker. Irisformer: Dense vision transformers for single-image inverse rendering in indoor scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2822-2831, 2022. 2" + ], + "bbox": [ + 516, + 92, + 906, + 726 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "5782", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Channel-wise Noise Scheduled Diffusion for Inverse Rendering in Indoor Scenes/e287505c-edc5-4f3b-9021-f507eb27467f_model.json b/2025/Channel-wise Noise Scheduled Diffusion for Inverse Rendering in Indoor Scenes/e287505c-edc5-4f3b-9021-f507eb27467f_model.json new file mode 100644 index 0000000000000000000000000000000000000000..f568a32acc71a5ba1e81a73811770f94cbaf2fc7 --- /dev/null +++ b/2025/Channel-wise Noise Scheduled Diffusion for Inverse Rendering in Indoor Scenes/e287505c-edc5-4f3b-9021-f507eb27467f_model.json @@ -0,0 +1,2178 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.096, + 0.131, + 0.903, + 0.153 + ], + "angle": 0, + "content": "Channel-wise Noise Scheduled Diffusion for Inverse Rendering in Indoor Scenes" + }, + { + "type": "text", + "bbox": [ + 0.183, + 0.179, + 0.815, + 0.218 + ], + "angle": 0, + "content": "JunYong Choi\\(^{1,2}\\) Min-cheol Sagong\\(^{1}\\) SeokYeong Lee\\(^{1,2}\\) Seung-Won Jung\\(^{2}\\) Ig-Jae Kim\\(^{1,3,4}\\) Junghyun Cho\\(^{1,3,4}\\)" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.223, + 0.78, + 0.277 + ], + "angle": 0, + "content": "\\(^{1}\\)Korea Institute of Science and Technology(KIST) \\(^{2}\\)Korea University \\(^{3}\\)AI-Robotics, KIST School, University of Science and Technology \\(^{4}\\)Yonsei-KIST Convergence Research Institute" + }, + { + "type": "text", + "bbox": [ + 0.178, + 0.28, + 0.816, + 0.294 + ], + "angle": 0, + "content": "{happily, mcsagong, shapin94, drjay, jhcho}@kist.re.kr swjung83@korea.ac.kr" + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.32, + 0.903, + 0.51 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.518, + 0.907, + 0.575 + ], + "angle": 0, + "content": "Figure 1. Diffusion-Based Inverse Rendering. We present a diffusion-based inverse rendering framework that addresses the two competing goals of accuracy and diversity in inverse rendering by using two distinct models, each dedicated to achieving one of the objectives. (a) Our first model predicts a single accurate solution, while (b) the second model presents diverse possible solutions. (c) This enables practical applications such as object insertion and material editing (e.g., increasing the roughness of the brown table)." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.587, + 0.327, + 0.602 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.644, + 0.486, + 0.903 + ], + "angle": 0, + "content": "We propose a diffusion-based inverse rendering framework that decomposes a single RGB image into geometry, material, and lighting. Inverse rendering is inherently ill-posed, making it difficult to predict a single accurate solution. To address this challenge, recent generative model-based methods aim to present a range of possible solutions. However, finding a single accurate solution and generating diverse solutions can be conflicting. In this paper, we propose a channel-wise noise scheduling approach that allows a single diffusion model architecture to achieve two conflicting objectives. The resulting two diffusion models, trained with different channel-wise noise schedules, can predict a single highly accurate solution and present multiple possible solutions. The experimental results demonstrate the superiority of our two models in terms of both diversity and accuracy, which translates to enhanced performance in downstream applications such as object insertion and material editing." + }, + { + "type": "title", + "bbox": [ + 0.515, + 0.634, + 0.645, + 0.649 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.659, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Single-image inverse rendering is a long-standing challenge in computer vision and computer graphics that involves the decomposition of an RGB image into its constituent elements of material, lighting, and geometry. This technology enables a wide range of applications, including object insertion, material editing, and relighting. Although inverse rendering has a clear solution, interpreting a single observed radiance can be complex, as it may result from various combinations of material, geometry, and lighting. For instance, a high radiance might be due to the material being white, intense lighting, or a pronounced specular component caused by the Fresnel effect. This inherent ill-posedness of inverse rendering introduces ambiguity and difficulty, particularly when working with a single-image. Therefore, it is important not only to predict an optimal solution, but also to propose multiple potential solutions. However, as discussed in" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5773" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.137 + ], + "angle": 0, + "content": "other studies [15, 41], enforcing only the reconstruction loss cannot improve the sample diversity, leading to conflicts between the objectives." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.139, + 0.483, + 0.456 + ], + "angle": 0, + "content": "Learning-based methods [4, 5, 19, 47, 48] leveraging large-scale synthetic datasets have demonstrated excellent performance. Strong supervised signals from reconstruction losses on well-aligned ground truths have enabled these methods to effectively decompose scenes. However, because these methods are based on deterministic approaches, they do not account for the solution space of possible combinations. Recently, inverse rendering studies based on Latent Diffusion Models (LDM) [28, 31] have emerged [14, 22, 43]. These probabilistic methods, which rely on priors learned from large image-text datasets, not only provide multiple solutions but also exhibit excellent predictive capability. However, they often overlook the trade-off between sample diversity and prediction accuracy and they struggle to account for the dependencies between modalities. Additionally, they rely on the pretrained three-channel autoencoder of LDM, which is effective at generating image-based data such as RGB images but falls short in handling high-dimensional data like per-pixel lighting, which is crucial for applications such as photorealistic object insertion and material editing." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.459, + 0.483, + 0.73 + ], + "angle": 0, + "content": "While probabilistic methods are effective for capturing multiple potential solutions in ill-posed problems like inverse rendering, they can be sub-optimal for predicting a definitive ground truth. We found that diffusion models exhibit a trade-off between diversity and accuracy depending on the entire timestep \\( T \\). Motivated by these observations, we propose two conditional diffusion models, each focusing on either diversity or accuracy. Our diffusion models predict multiple modalities, such as a depth map, normal map, diffuse albedo map, roughness map, and per-pixel environment map, simultaneously from a single RGB image. The first proposed model, the Progressive Diffusion Model (PDM), focuses on diversity using a large \\( T \\) for training. The second model, the Switchable Diffusion Model (SDM), is trained with a smaller \\( T \\) to prioritize accuracy. To improve the quality of inverse rendering, a channel-wise noise scheduling approach is introduced to adjust the transition between modalities during generation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.733, + 0.483, + 0.853 + ], + "angle": 0, + "content": "Our PDM and SDM operate in a low-resolution pixel space rather than a latent space, allowing them to handle per-pixel environment maps. To mitigate the high computational costs associated with handling high-dimensional data directly, we introduce a method for encoding these maps into compact feature vectors. Additionally, we present a simple RGB-guided super-resolution model for up-sampling outputs." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.483, + 0.901 + ], + "angle": 0, + "content": "We evaluated the performance by comparing it with other inverse rendering baselines on synthetic [4] and real-world [9, 37, 40] datasets. The experimental results showed" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.181 + ], + "angle": 0, + "content": "that SDM outperforms in images with fewer ambiguous regions, where accurate predictions are more feasible, while PDM excels in complex scenes with numerous ambiguous regions requiring diverse predictions, indicating the complementary nature of SDM and PDM. In summary, our main contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.184, + 0.905, + 0.228 + ], + "angle": 0, + "content": "1. We propose a novel diffusion-based inverse rendering framework that incorporates per-pixel environment maps while considering the correlations between modalities." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.23, + 0.904, + 0.273 + ], + "angle": 0, + "content": "2. We propose channel-wise noise scheduling, which improves generation quality by adjusting the modality generation transition in multi-modality diffusion models." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.275, + 0.905, + 0.334 + ], + "angle": 0, + "content": "3. We demonstrate superior performance over existing baselines on both real and synthetic datasets, highlighting the importance of considering both diversity and accuracy in inverse rendering." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.184, + 0.905, + 0.334 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.347, + 0.663, + 0.362 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.373, + 0.905, + 0.644 + ], + "angle": 0, + "content": "Inverse rendering. Inverse rendering involves the extraction of depth, normal, SVBRDF, and lighting information from RGB images. Initial research concentrated primarily on specific areas, such as shape-from-shading [13, 25, 45] and intrinsic image decomposition [1, 10, 16, 21]. Advances in deep learning have facilitated the study of complex properties, such as SVBRDF [6, 17, 18, 33] and lighting [8, 9, 36, 39]. Li et al. [19, 20] introduced OpenRooms and a spherical-Gaussian lighting model. IRISformer [48] leverages long-range attention through a dense vision transformer. Zhu et al. [47] introduced the refined InteriorVerse dataset and incorporated Monte Carlo ray tracing into the inverse rendering. Choi et al. [4, 5] extended this approach to a multi-view setup. However, all of these methods rely on deterministic models, limiting their ability to propose alternative solutions. Our work diverges by proposing PDM and SDM, capable of generating multiple solutions and surpassing deterministic models in flexibility and accuracy." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.645, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Diffusion models. Diffusion models [7, 11, 35], which have recently emerged as powerful tools for image generation, are now being applied to various tasks. Several studies have trained diffusion models directly on target datasets to generate materials [38], distant lighting [23, 27], and human face reflectance [26]. Recently, research has begun exploring inverse rendering by utilizing the strong priors of pretrained LDMs. Kocsis et al. [14] fine-tuned an LDM on synthetic datasets, enabling the generation of multiple plausible material maps. RGB \\(\\leftrightarrow\\) X [43] demonstrated that LDMs can be used for both inverse rendering and rendering. Luo et al. [22] trained ControlNet [44] to predict multiple intrinsic modalities. However, since these methods predict each modality independently, they fail to address inter-modality dependencies and struggle with environment maps, limiting realistic lighting inference and applications. In contrast, our framework accounts for these dependencies, enhancing" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5774" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.087, + 0.905, + 0.235 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.245, + 0.908, + 0.276 + ], + "angle": 0, + "content": "Figure 2. Entire pipeline. Our channel-wise noise scheduling assists the DM's inference by adjusting the transitions between modalities based on the timestep. Since the DM operates at a low resolution, modalities are up-sampled through our SRM." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.301, + 0.482, + 0.334 + ], + "angle": 0, + "content": "sample fidelity, while per-pixel lighting enables realistic, high-fidelity object insertion." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.344, + 0.184, + 0.36 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.369, + 0.485, + 0.627 + ], + "angle": 0, + "content": "Given an input RGB image, denoted as \\(\\mathrm{I} \\in \\mathbb{R}^{3 \\times H \\times W}\\), our goal is to obtain the following outputs: a normal map \\(\\mathrm{N} \\in \\mathbb{R}^{3 \\times H \\times W}\\), a depth map \\(\\mathrm{D} \\in \\mathbb{R}^{H \\times W}\\), a diffuse albedo map \\(\\mathrm{A} \\in \\mathbb{R}^{3 \\times H \\times W}\\), a roughness map \\(\\mathrm{R} \\in \\mathbb{R}^{H \\times W}\\), and a per-pixel environment map, represented as \\(\\mathrm{E} \\in \\mathbb{R}^{3 \\times C \\times H \\times W}\\) where \\(H\\) and \\(W\\) represent the image dimensions, and \\(C\\) denotes the angular resolution of the environment map. First, we introduce the synthetic dataset used in our experiments. Second, we introduce a method to encode per-pixel lighting into a neural feature vector. Next, we introduce the conditional diffusion model that serves as the foundation of our approach, along with two channel-wise noise scheduling techniques that allow us to customize the model for our desired application. Since the diffusion model operates in the low-resolution pixel space, we introduce a RGB-guided super-resolution model for upsampling. The entire pipeline is shown in Fig. 2." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.637, + 0.188, + 0.65 + ], + "angle": 0, + "content": "3.1. Dataset" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.658, + 0.483, + 0.825 + ], + "angle": 0, + "content": "To train our conditional diffusion model, we need well-aligned ground truth scene information (N, D, A, R, E) pairs for RGB images. For this purpose, two synthetic datasets, OpenRooms [20] and InteriorVerse [47], are suitable; however, the lighting part of InteriorVerse is currently not available. Additionally, OpenRooms often contains overly simplistic images (such as views of walls or incorrectly placed light sources), which hinder the diffusion model from learning the data distribution. Therefore, we used the OpenRooms FF dataset [4], which involves filtering out inappropriate images from OpenRooms." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.833, + 0.381, + 0.851 + ], + "angle": 0, + "content": "3.2. Implicit Lighting Representation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.903 + ], + "angle": 0, + "content": "As a method for handling high-dimensional data such as per-pixel environment maps, a recent method called MAIR++ [5] introduced the implicit lighting representation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.301, + 0.907, + 0.497 + ], + "angle": 0, + "content": "(ILR), which represents environment maps as neural feature vectors. MAIR++ processes the RGB, geometry, and material using a CNN to obtain the ILR, which is subsequently processed by three different decoders to obtain an environment map, shading map, and specular radiance. This approach can render images in a realistic and computationally efficient way. We are motivated to use the ILR as our lighting representation. However, since the previous ILR [5] is not inferred from the environment map, it neither accurately reproduces the ground truth lighting nor effectively captures the detailed high-dynamic range (HDR) radiance values. In contrast, we re-design the ILR as an MLP-based autoencoder that takes the environment map as input." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.499, + 0.909, + 0.621 + ], + "angle": 0, + "content": "Since the HDR radiance values of the environment map cover a very wide range, using them directly as input causes the network to diverge easily. Therefore, we transform the input into the log space using \\(\\log 1p\\) [29] and add batch normalization to the encoder MLP. These two simple techniques allow the network to handle HDR lighting more effectively, leading to significantly improved performance. Mathematically, the encoder is defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.604, + 0.641, + 0.907, + 0.66 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {f} = \\tanh (\\operatorname {E n c o d e r} (\\log 1 p (\\bar {\\mathrm {E}}))), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.67, + 0.908, + 0.718 + ], + "angle": 0, + "content": "where \\(\\bar{\\cdot}\\) represents the ground truth, and \\(f\\in \\mathbb{R}^{96}\\) is the ILR neural feature vector. To enable realistic rendering through \\(f\\), the three decoders are applied:" + }, + { + "type": "equation", + "bbox": [ + 0.567, + 0.728, + 0.758, + 0.744 + ], + "angle": 0, + "content": "\\[\n\\mathrm {E} = \\exp \\operatorname {m 1} (\\operatorname {D e c o d e r} _ {E} (\\boldsymbol {f}))\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.57, + 0.747, + 0.905, + 0.763 + ], + "angle": 0, + "content": "\\[\n\\mathrm {S} = \\exp \\operatorname {m 1} (\\operatorname {D e c o d e r} _ {S} (\\boldsymbol {f})), \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.567, + 0.765, + 0.852, + 0.782 + ], + "angle": 0, + "content": "\\[\n\\mathrm {I} _ {s} = \\exp \\operatorname {m 1} (\\operatorname {D e c o d e r} _ {I} (\\boldsymbol {f}, \\bar {\\mathrm {R}}, \\gamma (\\boldsymbol {r}), \\bar {\\mathrm {N}} \\cdot \\boldsymbol {v})).\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.794, + 0.907, + 0.902 + ], + "angle": 0, + "content": "where \\(\\mathsf{expm1}\\) [29] is the inverse of \\(\\log 1p\\), \\(S \\in \\mathbb{R}^{3 \\times H \\times W}\\) denotes the shading map, \\(I_s \\in \\mathbb{R}^{3 \\times H \\times W}\\) represents the specular image, \\(\\gamma\\) refers to the frequency positional encoding, and \\(\\boldsymbol{r}\\) and \\(\\boldsymbol{v}\\) indicate the reflection direction and the viewing direction, respectively. Through these three decoders, \\(\\boldsymbol{f}\\) can not only reconstruct the environment map but also be used to render arbitrary views." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5775" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.101, + 0.089, + 0.899, + 0.178 + ], + "angle": 0, + "content": "
TNLDLALRLELIL
MSE ↓var ↑MSE ↓var ↑MSE ↓var ↑MSE ↓var ↑MSE ↓var ↑MSE ↓var ↑
10002.2360.1830.0650.0200.4180.0345.5180.10612.6679.610.1790.042
1002.2380.1820.0640.0190.4130.0345.3540.09412.6677.620.1770.041
102.0260.1380.0590.0130.3740.0245.4400.06612.1362.380.1620.03
11.8170.00.0610.00.3340.05.5450.011.370.060.1370.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.089, + 0.188, + 0.908, + 0.232 + ], + "angle": 0, + "content": "Table 1. Ablation of training \\( T \\). The best values in each column are highlighted in bold, and the worst values are underlined. In most cases, as the value of \\( T \\) decreases, prediction accuracy improves, but the diversity of the samples is reduced. To evaluate under general conditions, we fixed \\( \\tau \\) to 1.0 for all settings, and the DDIM [34] steps were selected as 2, 2, 2, and 1, respectively." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.256, + 0.395, + 0.273 + ], + "angle": 0, + "content": "3.3. Diffusion-based Inverse Rendering" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.278, + 0.485, + 0.475 + ], + "angle": 0, + "content": "Since the introduction of LDM [28, 31], many diffusion-based methods [14, 22, 43] have used pretrained LDMs that have strong priors for images. However, these methods do not provide meaningful priors for high-dimensional data such as environment maps. Therefore, we adopt a cascaded structure [12] that first generates \\(\\mathbf{z}_0 = (\\mathrm{N}_L,\\mathrm{D}_L,\\mathrm{A}_L,\\mathrm{R}_L,\\boldsymbol {f}_L)\\) and then increases the resolution, where the subscript \\(L\\) denotes a low-resolution image of size \\(H_{L}\\times W_{L}\\). During inference, our conditional diffusion model (DM) takes \\(\\mathrm{I}_L\\) as a condition and iteratively denoises the Gaussian noise image \\(\\mathbf{z}_{T - 1}\\sim \\mathcal{N}(0,1)\\) to produce \\(\\mathbf{z}_0\\). During training, we predict the velocity \\(\\mathbf{v}_t\\) [32], and the loss function \\(L_{\\mathrm{DM}}\\) is given by:" + }, + { + "type": "equation", + "bbox": [ + 0.146, + 0.488, + 0.483, + 0.507 + ], + "angle": 0, + "content": "\\[\nL _ {\\mathrm {D M}} = \\left\\| \\mathbf {v} _ {t} - \\mathrm {D M} \\left(t, \\mathbf {z} _ {t}, \\mathrm {I} _ {L}, \\operatorname {C L I P} \\left(\\mathrm {I} _ {L}\\right)\\right) \\right\\| _ {2} ^ {2}, \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.184, + 0.518, + 0.483, + 0.535 + ], + "angle": 0, + "content": "\\[\n\\mathbf {v} _ {t} = \\sqrt {\\bar {\\alpha} _ {t}} \\mathbf {z} _ {T - 1} - \\sqrt {1 - \\bar {\\alpha} _ {t}} \\mathbf {z} _ {0}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.539, + 0.483, + 0.614 + ], + "angle": 0, + "content": "where \\(T\\) is the entire timestep, \\(t \\in [0, T - 1]\\) is uniformly sampled during training, \\(\\bar{\\alpha}_t\\) is the noise scheduler, \\(\\mathbf{z}_t = \\sqrt{\\bar{\\alpha}_t} \\mathbf{z}_0 + \\sqrt{1 - \\bar{\\alpha}_t} \\mathbf{z}_{T-1}\\), and CLIP is the pretrained image encoder [30]. \\(\\mathrm{I}_L\\) is concatenated with \\(\\mathbf{z}_t\\), and the CLIP embedding is used as a cross-attention condition." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.615, + 0.483, + 0.842 + ], + "angle": 0, + "content": "Channel-wise noise scheduling. Since we generate multiple interdependent modalities simultaneously, we hypothesize that there is an optimal transition to generate these modalities in inverse rendering. For example, according to previous works [18, 42, 46], geometry typically has the least uncertainty and supports the prediction of material and lighting, so generating geometry first could be advantageous. On the other hand, since lighting has the highest uncertainty, generating it last could be beneficial. To validate this hypothesis, we divide \\(\\mathbf{z}_0\\) into three categories based on the modality attributes: geometry \\((\\mathrm{N}_L,\\mathrm{D}_L)\\), material \\((\\mathrm{A}_L,\\mathrm{R}_L)\\), and lighting \\((f_{L})\\), and accordingly apply different noise scheduling to the three parts. Specifically, following Chen et al. [2], we select the cosine scheduler as the base noise scheduler and parameterize it as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.111, + 0.854, + 0.483, + 0.906 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} v _ {s} = \\cos (s \\cdot 2 \\pi) ^ {2 \\tau}, v _ {b} = \\cos (b \\cdot 2 \\pi) ^ {2 \\tau} \\\\ \\alpha_ {t} = \\cos \\left(\\left(\\left(b - s\\right) \\cdot t + s\\right) \\cdot 2 \\pi\\right) ^ {2 \\tau}, \\bar {\\alpha} _ {t} = \\frac {v _ {b} - \\alpha_ {t}}{v _ {b} - v _ {s}}. \\tag {5} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.257, + 0.907, + 0.318 + ], + "angle": 0, + "content": "Throughout the paper, we fix \\( s = 0.008 \\) and \\( b = 1 \\) and control the scheduler using only \\( \\tau \\). When \\( \\tau = 1 \\), \\( \\bar{\\alpha}_t \\) is equivalent to the cosine scheduler. The behavior of the noise scheduler as a function of \\( \\tau \\) is shown in Fig. 3." + }, + { + "type": "image", + "bbox": [ + 0.515, + 0.329, + 0.905, + 0.455 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.464, + 0.907, + 0.493 + ], + "angle": 0, + "content": "Figure 3. Parameterizing noise scheduler. A larger \\(\\tau\\) value results in a slower generation of the modality." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.509, + 0.907, + 0.704 + ], + "angle": 0, + "content": "Fig. 3 shows that as the value of \\(\\tau\\) increases, the noise increases more gradually. To determine the optimal modality generation order using \\(\\tau\\), we conducted experiments by applying different values of \\(\\tau\\) to geometry, material, and lighting, meaning that different noise levels are applied separately to each channel of the data during the denoising process of the diffusion model. The ablation study in Tab. 7 confirms that generating in the order of geometry, material, and lighting is the most effective, aligning with our intuition. As a result, through our channel-wise noise scheduling, it focuses on specific modalities at different timesteps, enabling more effective predictions of other modalities in subsequent timesteps." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.909, + 0.885 + ], + "angle": 0, + "content": "In general, \\( T \\) in diffusion models is set to a large number, such as 1000. This helps approximate the data distribution and increases the diversity of the generated data [24]. However, we note that such diffusion models are challenging in predicting a single definitive ground truth. To this end, we conducted an ablation study on \\( T \\) values by training the DM with \\( T = 1000, 100, 10 \\), and 1. For each setting, we performed sampling 20 times in the test set to measure the MSE and the variance (var) of the samples. Tab. 1 shows that reducing \\( T \\) improves accuracy at the cost of diversity in most cases. In the extreme case of \\( T = 1 \\), the method becomes deterministic." + }, + { + "type": "text", + "bbox": [ + 0.534, + 0.886, + 0.907, + 0.902 + ], + "angle": 0, + "content": "To explore models focused on diversity and accuracy, we" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5776" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.242 + ], + "angle": 0, + "content": "develop two types of diffusion models: one with \\( T = 1000 \\) and the other with \\( T = 1 \\). Following the ablation study in Tab. 7, we select \\( \\tau = (0.9, 1.2, 1.5) \\) when \\( T = 1000 \\) to prioritize diversity. Due to its gradual generation process, we name this model the Progressive Diffusion Model (PDM). However, this gradual generation process cannot be applied to the diffusion model with \\( T = 1 \\). Moreover, \\( T = 1 \\) limits the ability of the diffusion model to exploit its full potential over multiple timesteps [3]. To this end, we define our new noise scheduler, as shown in Tab. 2." + }, + { + "type": "table", + "bbox": [ + 0.104, + 0.255, + 0.472, + 0.308 + ], + "angle": 0, + "content": "
tt mod 3 = 0t mod 3 = 1t mod 3 = 2T-1
\\( \\bar{a}_{t}^{G} \\)1100
\\( \\bar{a}_{t}^{M} \\)1010
\\( \\bar{a}_{t}^{L} \\)0110
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.319, + 0.483, + 0.348 + ], + "angle": 0, + "content": "Table 2. SDM noise scheduler. Here, \\(\\bar{\\alpha}_t^G\\), \\(\\bar{\\alpha}_t^M\\), and \\(\\bar{\\alpha}_t^L\\) are the schedulers for geometry, material, and lighting, respectively." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.366, + 0.484, + 0.518 + ], + "angle": 0, + "content": "A value of 0 implies that pure Gaussian noise is added, while a value of 1 means that no noise is added. With our new noise scheduler, \\( T \\) is redefined as \\( 3n + 1 \\) where \\( n > 0 \\). After the initial prediction at \\( t = T - 1 \\), the noise is switched per channel, providing conditions for the next diffusion step. Taking into account this feature of the switchable noise scheduling, we name this model Switchable Diffusion Model (SDM). The SDM is trained to predict the modality with an SNR of 0 at each step, allowing for more refined predictions by utilizing the previous step's predictions." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.527, + 0.367, + 0.543 + ], + "angle": 0, + "content": "3.4. RGB-Guided Super Resolution" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.549, + 0.484, + 0.641 + ], + "angle": 0, + "content": "We use a UNet as the RGB-guided super-resolution model, consisting of a shared encoder and five separate decoders. The encoder processes the RGB image into a dense feature map, which, along with the rendered diffuse \\(\\mathrm{I}_d\\) and specular \\(\\mathrm{I}_s\\) images from \\(\\mathbf{z}_0\\), is fed into each decoder. Our RGB-guided Super-Resolution Model (SRM) is as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.168, + 0.649, + 0.483, + 0.665 + ], + "angle": 0, + "content": "\\[\n\\mathrm {N}, \\mathrm {D}, \\mathrm {A}, \\mathrm {R}, \\boldsymbol {f} = \\operatorname {S R M} \\left(\\mathbf {z} _ {0}, \\mathrm {I}, \\mathrm {I} _ {d}, \\mathrm {I} _ {s}\\right), \\tag {6}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.679, + 0.208, + 0.695 + ], + "angle": 0, + "content": "4. Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.483, + 0.84 + ], + "angle": 0, + "content": "In this section, we evaluate our method on both synthetic and real-world datasets. We first evaluate on a synthetic dataset, followed by inverse rendering and object insertion on a real-world spatially varying lighting dataset. Furthermore, we evaluated our method on the sub-tasks of inverse rendering, namely intrinsic decomposition and geometry prediction. More details on training, model architectures, and additional experimental results are provided in the Supplementary Materials." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Evaluation settings. We used \\( C = 128 \\), \\( H = 240 \\), \\( W = 320 \\), \\( H_{L} = 60 \\), and \\( W_{L} = 80 \\). For SDM, we set \\( T = 4 \\). In all experiments, PDM generated 10 samples and the mean and variance of these samples were evaluated." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.198 + ], + "angle": 0, + "content": "Baselines. We compared our method with state-of-the-art inverse rendering frameworks capable of per-pixel lighting inference, specifically those by Li et al. [19] and Zhu et al. [47]. Both methods were trained on the OpenRooms FF [4] dataset. However, since the code for Zhu's lighting model has not been released, we used their pretrained lighting model, which was trained on the InteriorVerse dataset." + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.211, + 0.905, + 0.356 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.366, + 0.906, + 0.395 + ], + "angle": 0, + "content": "Figure 4. Qualitative synthetic evaluation. Only our method successfully decomposes specular radiance into material and lighting." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.425, + 0.791, + 0.44 + ], + "angle": 0, + "content": "4.1. Evaluation on Synthetic Scenes" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.447, + 0.907, + 0.703 + ], + "angle": 0, + "content": "Tab. 3 includes quantitative prediction results on the synthetic [4] test dataset. Although our target is single-view, we include results from the multi-view baselines MAIR [4] and MAIR++ [5]. Our SDM, which prioritizes prediction accuracy, demonstrated superior performance compared to existing single-image baselines in all areas except roughness, and our PDM also demonstrated respectable performance comparable to the baselines. Notably, our lighting accuracy was comparable to that of MAIR, and the albedo accuracy outperformed even the multi-view methods. Additionally, it can be observed that both MAIR++ and our method, which are capable of ILR-based neural rendering, exhibit significantly superior re-rendering (I) performance. Figure 4 presents an illustrative example of this fact. Even in images with strong specular radiance, our SDM and PDM successfully decomposed and re-rendered the image in contrast to the failure of the baseline methods." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.716, + 0.905, + 0.814 + ], + "angle": 0, + "content": "
MSE(×10-2)N ↓D ↓A ↓R ↓E ↓I ↓
Li et al. [19]3.7820.1350.8696.49615.490.491
Zhu et al. [47]1.7380.0620.5284.208*31.15*3.565
SDM1.5020.0390.3494.96913.350.156
PDM(Mean)1.8310.0470.3944.81914.080.150
PDM(Best)1.7820.0440.3814.64413.750.146
MAIR [4]1.0100.1400.6393.45713.790.784
MAIR++ [5]1.0100.0070.4782.49711.610.058
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.825, + 0.906, + 0.88 + ], + "angle": 0, + "content": "Table 3. Quantitative synthetic evaluation. Bold text indicates the best performance among single-view methods, while underlined text represents the second-best. * denotes models trained on the InteriorVerse dataset." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5777" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.088, + 0.249, + 0.17 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.135, + 0.171, + 0.207, + 0.182 + ], + "angle": 0, + "content": "Input Image" + }, + { + "type": "image", + "bbox": [ + 0.244, + 0.089, + 0.43, + 0.254 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.301, + 0.255, + 0.314, + 0.265 + ], + "angle": 0, + "content": "N" + }, + { + "type": "image", + "bbox": [ + 0.431, + 0.089, + 0.588, + 0.255 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.461, + 0.256, + 0.473, + 0.265 + ], + "angle": 0, + "content": "A" + }, + { + "type": "image", + "bbox": [ + 0.587, + 0.089, + 0.746, + 0.254 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.617, + 0.255, + 0.631, + 0.267 + ], + "angle": 0, + "content": "Id" + }, + { + "type": "image", + "bbox": [ + 0.746, + 0.089, + 0.906, + 0.254 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.779, + 0.255, + 0.787, + 0.265 + ], + "angle": 0, + "content": "I" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.279, + 0.908, + 0.308 + ], + "angle": 0, + "content": "Figure 5. Qualitative real-world evaluation. Our method successfully recognizes shadows under the desk, achieving plausible geometry prediction and realistic re-rendering results." + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.321, + 0.296, + 0.527 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.176, + 0.528, + 0.212, + 0.535 + ], + "angle": 0, + "content": "Li et al." + }, + { + "type": "image", + "bbox": [ + 0.297, + 0.321, + 0.499, + 0.527 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.375, + 0.528, + 0.42, + 0.536 + ], + "angle": 0, + "content": "Zhu et al." + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.321, + 0.703, + 0.528 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.576, + 0.528, + 0.626, + 0.536 + ], + "angle": 0, + "content": "Ours(SDM)" + }, + { + "type": "image", + "bbox": [ + 0.705, + 0.321, + 0.906, + 0.527 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.787, + 0.528, + 0.823, + 0.535 + ], + "angle": 0, + "content": "RGB \\(\\leftrightarrow\\) X" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.551, + 0.908, + 0.581 + ], + "angle": 0, + "content": "Figure 6. Qualitative object insertion evaluation. We accurately predicted the direction of light and shadows even in real-world images. Additionally, our rendered objects were closer to the ground truth in terms of hue and brightness compared to the baselines." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.606, + 0.381, + 0.621 + ], + "angle": 0, + "content": "4.2. Evaluation on Real-world Scenes" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.629, + 0.484, + 0.749 + ], + "angle": 0, + "content": "Qualitative evaluation. We evaluated the generalization capability of our method on real-world images from Garon et al. [9]. As shown in Fig. 5, both SDM and PDM successfully separated the albedo even in the presence of strong shadows under the desk. Li et al. [19] partially removed the shadows but failed to predict the geometry, while Zhu et al. [47] accurately predicted the normal but failed to estimate the albedo." + }, + { + "type": "table", + "bbox": [ + 0.116, + 0.763, + 0.462, + 0.794 + ], + "angle": 0, + "content": "
Li et al. [19]Zhu et al. [47]SDMGround Truth
0.1780.1530.3220.348
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.805, + 0.483, + 0.834 + ], + "angle": 0, + "content": "Table 4. AMT user study on object insertion. Our method was selected by nearly as many users as the ground truth." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Lighting Evaluation. We also evaluated lighting performance through object insertion, using the object insertion tool from Li et al. [19]. For fairness, we excluded PDM" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.606, + 0.907, + 0.895 + ], + "angle": 0, + "content": "and focused on evaluating only SDM. As shown in Fig. 6, our method accurately predicts strong directional lighting, as well as the surrounding light's tone and shadows, resulting in the most realistic object insertions. In contrast, Zhu et al. [47] failed to predict the correct light direction, while Li et al. [19] struggles with realistic rendering due to a lack of high-frequency details in the environment map. The diffusion-based method RGB \\(\\leftrightarrow\\) X [43] relies on inpainting for object insertion, which compromises the fidelity of the inserted objects and leads to visible artifacts. Additionally, since it cannot handle environment maps, it struggled to accurately predict the direction of light. For a more quantitative evaluation, we conducted an AMT user study on object insertion, including the ground truth. The results in Tab. 4 confirmed the superiority of our method. Users were asked: \"Of the four images, please select the one that shows white rabbits blending naturally into the surrounding light.\" Our method was selected by users nearly as frequently as the ground truth." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5778" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.088, + 0.506, + 0.239 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.088, + 0.906, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.25, + 0.908, + 0.293 + ], + "angle": 0, + "content": "Figure 7. Intrinsic decomposition. Our method successfully decomposes the input image into albedo and shading. RGB \\(\\leftrightarrow\\) X [43], leveraging its strong prior, achieves successful albedo prediction but struggles with realistic rendering, resulting in excessive brightness in the left sample and incorrect specular highlights in the right sample." + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.315, + 0.482, + 0.495 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.503, + 0.483, + 0.533 + ], + "angle": 0, + "content": "Figure 8. Geometry prediction. Our method achieves the most accurate prediction of the ceiling structure." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.548, + 0.48, + 0.636 + ], + "angle": 0, + "content": "
MethodA ↓Hue ↓Texture ↓WHDR ↓S ↓I ↓
Li et al. [19]0.9245.7810.4730.3011.1620.021
Zhu et al. [47]1.6854.6380.4720.3011.4410.028
SDM0.7055.0560.4590.2931.1810.006
PDM(Mean)0.7575.2600.4590.3061.1560.005
PDM(Best)0.6234.5270.4560.2781.1130.005
RGB↔X [43](Mean)0.6023.6900.4480.1681.2610.055
RGB↔X [43](Best)0.3313.0910.4410.1461.1960.052
" + }, + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.639, + 0.483, + 0.668 + ], + "angle": 0, + "content": "Table 5. Intrinsic decomposition evaluation on MAW [40]. Bold text indicates the best performance." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.696, + 0.315, + 0.71 + ], + "angle": 0, + "content": "4.3. Evaluation on Sub-tasks" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Intrinsic decomposition. We evaluated our method on intrinsic decomposition using the recently introduced MAW [40] dataset, which provides more refined annotations for 888 images, allowing more quantitative comparisons than the WHDR metric used in the IIW [1] dataset. Tab. 5 presents comparisons for MSE for A, S, and I, along with metrics for hue, texture, and WHDR for A. The hue metric evaluates the chromaticity component, while the texture metric is based on LPIPS perceptual similarity. Our method generally outperformed existing deterministic methods [19, 47]. RGB \\(\\leftrightarrow\\) X [43], which learns strong priors from diverse large datasets, excels in albedo infer" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.318, + 0.907, + 0.515 + ], + "angle": 0, + "content": "ence; however, due to its reliance on diffusion-based rendering, it reconstructed the input images with the lowest accuracy. In contrast, our method, trained only on OpenRooms FF, achieved the highest performance in shading estimation and most faithfully reproduced input images, demonstrating high fidelity. This can also be observed in Fig. 7. In the left sample, RGB \\(\\leftrightarrow\\) X [43] fails to predict the shading intensity as it does not account for HDR lighting. Similarly, in the right sample, RGB \\(\\leftrightarrow\\) X [43] introduces specular highlights on the desk that are absent in the input image, resulting in a loss of fidelity. In contrast, our method predicts spatially consistent albedo and accurately estimates brightness, enabling a more realistic reproduction of the input image." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.516, + 0.909, + 0.622 + ], + "angle": 0, + "content": "Geometry prediction. Tab. 6 includes the geometry prediction evaluation on the DIODE [37] indoor dataset. Similarly to the intrinsic decomposition task, both our PDM and SDM demonstrate superior or comparable performance to the baselines. Fig. 8 provides qualitative examples, demonstrating that our method exhibits strong generalization capability in geometry prediction." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.635, + 0.903, + 0.682 + ], + "angle": 0, + "content": "
MetricLi et al. [19]Zhu et al. [47]SDMPDM(Mean)PDM(Best)
N ↓0.1960.1710.1470.1520.144
D ↓0.0590.0570.0570.0550.050
I ↓0.0150.0130.0030.0020.002
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.685, + 0.907, + 0.713 + ], + "angle": 0, + "content": "Table 6. Geometry prediction evaluation on DIODE [37]. Bold text indicates the best performance." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.742, + 0.73, + 0.759 + ], + "angle": 0, + "content": "4.4. Modality Dependencies" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.909, + 0.901 + ], + "angle": 0, + "content": "To generate a reasonable solution for inverse rendering, it is crucial to account for the dependencies between modalities. For instance, if a material is predicted to be darker, the lighting should be predicted brighter to compensate for the input image. However, existing diffusion-based methods [14, 22, 43] infer each modality independently, making it difficult to capture such dependencies. In contrast, our PDM generates all modalities simultaneously, enabling it to account for these relationships. Fig. 9 illustrates this with" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5779" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.088, + 0.483, + 0.239 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.247, + 0.483, + 0.275 + ], + "angle": 0, + "content": "Figure 9. Modality dependencies. Two samples of N, D, A, S, and R predicted by PDM." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.303, + 0.484, + 0.44 + ], + "angle": 0, + "content": "two different samples generated by PDM. For the center table in the top image, the first sample predicts the normal pointing upward with bright shading, resulting in a darker albedo. Conversely, the second sample predicts the normal facing the camera, producing darker shading and a brighter albedo. Additionally, for the area under the cabinet in the bottom image, the first sample interprets it as a shadow and predicts a bright albedo, while the second sample treats it as part of the albedo and generates bright shading." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.449, + 0.203, + 0.463 + ], + "angle": 0, + "content": "4.5. Ablations" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.471, + 0.484, + 0.699 + ], + "angle": 0, + "content": "Modality generation order. To explore the optimal modality generation order, we conducted an ablation study on the modality generation sequence. There are no restrictions on \\(\\tau\\); however, we selected three values: \\(\\tau = 0.9, 1.2,\\) and 1.5. These values were applied separately to \\(\\tau_{G}\\) (geometry), \\(\\tau_{M}\\) (material), and \\(\\tau_{L}\\) (lighting), and the models were trained independently. The performance was then evaluated on the test dataset. Tab. 7 shows that this channel-wise noise scheduling can significantly impact the inverse rendering quality. In particular, \\(\\tau_{G}, \\tau_{M}, \\tau_{L} = (0.9, 1.2, 1.5)\\) produces the best quality, aligning with our intuition. In contrast, generating in reverse order results in a detrimental effect. Fig. 10 provides examples of these two cases. The reverse order (L-M-G) exhibits artifacts in areas with strong shadows, whereas our method produces plausible results." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.71, + 0.48, + 0.815 + ], + "angle": 0, + "content": "
τGτMτLNL↓DL↓AL↓RL↓EL↓IL↓
1.01.01.02.1400.0620.4035.75412.190.163
0.91.21.52.0240.0590.3765.17611.940.159
0.91.51.22.1040.0600.3905.72312.150.165
1.20.91.52.0730.0600.3885.39412.200.161
1.21.50.92.2020.0640.4115.88212.330.167
1.50.91.22.2110.0640.4055.66412.360.160
1.51.20.92.2450.0660.4145.98612.530.166
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.825, + 0.483, + 0.854 + ], + "angle": 0, + "content": "Table 7. Ablation of the generation order. The best values are highlighted in bold, while the worst values are underlined." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Iterative Prediction of SDM. We conducted an ablation study to evaluate the benefits of iterative prediction for" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.089, + 0.906, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.224, + 0.906, + 0.266 + ], + "angle": 0, + "content": "Figure 10. Comparison of modality generation order. Generating in the order of geometry, material, and lighting helps effectively infer shadows." + }, + { + "type": "table", + "bbox": [ + 0.538, + 0.28, + 0.882, + 0.339 + ], + "angle": 0, + "content": "
TN ↓D ↓A ↓R ↓E ↓I ↓
11.7330.0600.3185.83711.240.132
41.5780.0520.2885.74310.960.137
71.6300.0500.2925.75910.970.139
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.349, + 0.906, + 0.378 + ], + "angle": 0, + "content": "Table 8. Ablation of SDM iterative prediction. Bold text indicates the best performance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.405, + 0.907, + 0.496 + ], + "angle": 0, + "content": "SDM. In Tab. 8, the improvements at \\( T = 4 \\) and \\( T = 7 \\) over \\( T = 1 \\) indicate that iterative prediction utilizes previously predicted modalities to inform subsequent predictions. However, the performance drop at \\( T = 7 \\) compared to \\( T = 4 \\) suggests that excessive iterations do not always yield better results." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.512, + 0.634, + 0.527 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.538, + 0.906, + 0.627 + ], + "angle": 0, + "content": "We tackled inverse rendering with two opposing goals: attaining diversity and ensuring accuracy. To address these objectives, we proposed a diffusion-based framework, in which channel-wise noise scheduling allowed each diffusion model to better understand the correlations between different modalities." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.629, + 0.906, + 0.809 + ], + "angle": 0, + "content": "Limitations and future work. Our approach has the limitation of requiring two separate models; thus, integrating them into a unified model would enhance usability. We explored re-rendering error as an alternative metric to overcome the constraint of depending only on the predicted variance of PDM for estimating inverse rendering uncertainty, but it could be imprecise. Further research on the handling of uncertainty could improve the complementarity of SDM and PDM. Moreover, our method refrains from exploiting the priors of pretrained LDMs. Using those priors to process per-pixel environment maps might enhance the performance and expand the applications of LDMs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Acknowledgements. This work was supported by Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government(MSIT)(No.RS-2023-00227592, Development of 3D Object Identification Technology Robust to Viewpoint Changes)" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5780" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.156 + ], + "angle": 0, + "content": "[1] Sean Bell, Kavita Bala, and Noah Snively. Intrinsic images in the wild. ACM Transactions on Graphics (TOG), 33(4): 1-12, 2014. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.159, + 0.484, + 0.186 + ], + "angle": 0, + "content": "[2] Ting Chen. On the importance of noise scheduling for diffusion models. arXiv preprint arXiv:2301.10972, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.189, + 0.484, + 0.257 + ], + "angle": 0, + "content": "[3] Jooyoung Choi, Jungbeom Lee, Chaehun Shin, Sungwon Kim, Hyunwoo Kim, and Sungroh Yoon. Perception prioritized training of diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11472-11481, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.259, + 0.483, + 0.342 + ], + "angle": 0, + "content": "[4] JunYong Choi, SeokYeong Lee, Haesol Park, Seung-Won Jung, Ig-Jae Kim, and Junghyun Cho. Mair: multi-view attention inverse rendering with 3d spatially-varying lighting estimation. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8392-8401. IEEE, 2023. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.344, + 0.483, + 0.413 + ], + "angle": 0, + "content": "[5] JunYong Choi, SeokYeong Lee, Haesol Park, Seung-Won Jung, Ig-Jae Kim, and Junghyun Cho. Mair++: Improving multi-view attention inverse rendering with implicit lighting representation. arXiv preprint arXiv:2408.06707, 2024. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.415, + 0.483, + 0.47 + ], + "angle": 0, + "content": "[6] Valentin Deschaintre, Miika Aittala, Fredo Durand, George Drettakis, and Adrien Bousseau. Single-image svbrdf capture with a rendering-aware deep network. ACM Transactions on Graphics (ToG), 37(4):1-15, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.472, + 0.483, + 0.514 + ], + "angle": 0, + "content": "[7] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.516, + 0.483, + 0.585 + ], + "angle": 0, + "content": "[8] Marc-Andre Gardner, Yannick Hold-Geoffroy, Kalyan Sunkavalli, Christian Gagne, and Jean-François Lalonde. Deep parametric indoor lighting estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7175–7183, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.587, + 0.483, + 0.656 + ], + "angle": 0, + "content": "[9] Mathieu Garon, Kalyan Sunkavalli, Sunil Hadap, Nathan Carr, and Jean-François Lalonde. Fast spatially-varying indoor lighting estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6908-6917, 2019. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.658, + 0.483, + 0.727 + ], + "angle": 0, + "content": "[10] Roger Grosse, Micah K Johnson, Edward H Adelson, and William T Freeman. Ground truth dataset and baseline evaluations for intrinsic image algorithms. In 2009 IEEE 12th International Conference on Computer Vision, pages 2335-2342. IEEE, 2009. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.729, + 0.483, + 0.771 + ], + "angle": 0, + "content": "[11] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.773, + 0.484, + 0.828 + ], + "angle": 0, + "content": "[12] Jonathan Ho, Chitwan Sahara, William Chan, David J Fleet, Mohammad Norouzi, and Tim Salimans. Cascaded diffusion models for high fidelity image generation. Journal of Machine Learning Research, 23(47):1-33, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.83, + 0.483, + 0.858 + ], + "angle": 0, + "content": "[13] Berthold KP Horn and Michael J Brooks. Shape from shading. MIT press, 1989. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.86, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[14] Peter Kocsis, Vincent Sitzmann, and Matthias Nießner. Intrinsic image diffusion for single-view material estimation. arXiv preprint arXiv:2312.12274, 2023. 2, 4, 7" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.135 + ], + "angle": 0, + "content": "[15] Soochan Lee, Junsoo Ha, and Gunhee Kim. Harmonizing maximum likelihood with gans for multimodal conditional generation. arXiv preprint arXiv:1902.09225, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.137, + 0.905, + 0.192 + ], + "angle": 0, + "content": "[16] Zhengqi Li and Noah Snavely. Cginintrinsics: Better intrinsic image decomposition through physically-based rendering. In Proceedings of the European conference on computer vision (ECCV), pages 371-387, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.194, + 0.905, + 0.261 + ], + "angle": 0, + "content": "[17] Zhengqin Li, Kalyan Sunkavalli, and Manmohan Chandraker. Materials for masses: Svbrdf acquisition with a single mobile phone image. In Proceedings of the European conference on computer vision (ECCV), pages 72-87, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.264, + 0.905, + 0.332 + ], + "angle": 0, + "content": "[18] Zhengqin Li, Zexiang Xu, Ravi Ramamoorthi, Kalyan Sunkavalli, and Manmohan Chandraker. Learning to reconstruct shape and spatially-varying reflectance from a single image. ACM Transactions on Graphics (TOG), 37(6):1-11, 2018. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.334, + 0.907, + 0.418 + ], + "angle": 0, + "content": "[19] Zhengqin Li, Mohammad Shafiei, Ravi Ramamoorthi, Kalyan Sunkavalli, and Manmohan Chandraker. Inverse rendering for complex indoor scenes: Shape, spatially-varying lighting and svbrdf from a single image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2475-2484, 2020. 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.42, + 0.907, + 0.503 + ], + "angle": 0, + "content": "[20] Zhengqin Li, Ting-Wei Yu, Shen Sang, Sarah Wang, Meng Song, Yuhan Liu, Yu-Ying Yeh, Rui Zhu, Nitesh Gundavarapu, Jia Shi, et al. Openrooms: An open framework for photorealistic indoor scene datasets. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7190–7199, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.505, + 0.905, + 0.574 + ], + "angle": 0, + "content": "[21] Jundan Luo, Zhaoyang Huang, Yijin Li, Xiaowei Zhou, Guofeng Zhang, and Hujun Bao. Niid-net: Adapting surface normal knowledge for intrinsic image decomposition in indoor scenes. IEEE Transactions on Visualization and Computer Graphics, 26(12):3434-3445, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.576, + 0.905, + 0.657 + ], + "angle": 0, + "content": "[22] Jundan Luo, Duygu Ceylan, Jae Shin Yoon, Nanxuan Zhao, Julien Philip, Anna Frühstück, Wenbin Li, Christian Richardt, and Tuanfeng Wang. Intrinsicdiffusion: Joint intrinsic layers from latent diffusion models. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 2, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.66, + 0.905, + 0.729 + ], + "angle": 0, + "content": "[23] Linjie Lyu, Ayush Tewari, Marc Habermann, Shunsuke Saito, Michael Zollhöfer, Thomas Leimkuhler, and Christian Theobalt. Diffusion posterior illumination for ambiguity-aware inverse rendering. ACM Transactions on Graphics (TOG), 42(6):1-14, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.731, + 0.905, + 0.786 + ], + "angle": 0, + "content": "[24] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International conference on machine learning, pages 8162-8171. PMLR, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.788, + 0.905, + 0.83 + ], + "angle": 0, + "content": "[25] Geoffrey Oxholm and Ko Nishino. Shape and reflectance from natural illumination. In European Conference on Computer Vision, pages 528-541. Springer, 2012. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.832, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[26] Foivos Paraperas Papantoniou, Alexandros Lattas, Stylianos Moschoglou, and Stefanos Zafeiriou. Relightify: Relightable 3d faces from a single image via diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8806-8817, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.514, + 0.957 + ], + "angle": 0, + "content": "5781" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.176 + ], + "angle": 0, + "content": "[27] Pakkapon Phongthawee, Worameth Chinchuthakun, Nontaphat Sinsunthithet, Varun Jampani, Amit Raj, Pramook Khungurn, and Supasorn Suwajanakorn. Diffusionlight: Light probes for free by painting a chrome ball. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 98-108, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.178, + 0.482, + 0.247 + ], + "angle": 0, + "content": "[28] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.249, + 0.482, + 0.303 + ], + "angle": 0, + "content": "[29] William H. Press, Saul A. Teukolsky, William T. Vetterling, and Brian P. Flannery. Numerical Recipes 3rd Edition: The Art of Scientific Computing. Cambridge University Press, 2007. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.305, + 0.482, + 0.387 + ], + "angle": 0, + "content": "[30] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.39, + 0.482, + 0.459 + ], + "angle": 0, + "content": "[31] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.461, + 0.482, + 0.502 + ], + "angle": 0, + "content": "[32] Tim Salimans and Jonathan Ho. Progressive distillation for fast sampling of diffusion models. arXiv preprint arXiv:2202.00512, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.504, + 0.482, + 0.546 + ], + "angle": 0, + "content": "[33] Shen Sang and Manmohan Chandraker. Single-shot neural relighting and svbrdf estimation. In European Conference on Computer Vision, pages 85-101. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.548, + 0.482, + 0.588 + ], + "angle": 0, + "content": "[34] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.59, + 0.482, + 0.631 + ], + "angle": 0, + "content": "[35] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. Advances in neural information processing systems, 32, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.633, + 0.482, + 0.716 + ], + "angle": 0, + "content": "[36] Pratul P Srinivasan, Ben Mildenhall, Matthew Tancik, Jonathan T Barron, Richard Tucker, and Noah Snively. Lighthouse: Predicting lighting volumes for spatially-coherent illumination. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8080-8089, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.718, + 0.482, + 0.787 + ], + "angle": 0, + "content": "[37] Igor Vasiljevic, Nick Kolkin, Shanyi Zhang, Ruotian Luo, Haochen Wang, Falcon Z Dai, Andrea F Daniele, Mohammadreza Mostajabi, Steven Basart, Matthew R Walter, et al. Diode: A dense indoor and outdoor depth dataset. arXiv preprint arXiv:1908.00463, 2019. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.789, + 0.482, + 0.857 + ], + "angle": 0, + "content": "[38] Giuseppe Vecchio, Renato Sortino, Simone Palazzo, and Concetto Spampinato. Matfuse: controllable material generation with diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4429-4438, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.859, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[39] Zian Wang, Jonah Philion, Sanja Fidler, and Jan Kautz. Learning indoor inverse rendering with 3d spatially-varying lighting. In Proceedings of the IEEE/CVF International" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.093, + 0.905, + 0.119 + ], + "angle": 0, + "content": "Conference on Computer Vision, pages 12538-12547, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[40] Jiaye Wu, Sanjoy Chowdhury, Hariharmano Shanmugaraja, David Jacobs, and Soumyadip Sengupta. Measured albedo in the wild: Filling the gap in intrinsics evaluation. In 2023 IEEE International Conference on Computational Photography (ICCP), pages 1-12. IEEE, 2023. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.192, + 0.905, + 0.247 + ], + "angle": 0, + "content": "[41] Jiahui Yu, Zhe Lin, Jimei Yang, Xiaohui Shen, Xin Lu, and Thomas S Huang. Free-form image inpainting with gated convolution. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4471-4480, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.249, + 0.907, + 0.304 + ], + "angle": 0, + "content": "[42] Ye Yu and William AP Smith. Inverserendernet: Learning single image inverse rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3155-3164, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.305, + 0.905, + 0.387 + ], + "angle": 0, + "content": "[43] Zheng Zeng, Valentin Deschaintre, Iliyan Georgiev, Yannick Hold-Geoffroy, Yiwei Hu, Fujun Luan, Ling-Qi Yan, and Miloš Hasan. Rgb \\(\\leftrightarrow\\) x: Image decomposition and synthesis using material-and lighting-aware diffusion models. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 2, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.39, + 0.905, + 0.445 + ], + "angle": 0, + "content": "[44] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.447, + 0.905, + 0.5 + ], + "angle": 0, + "content": "[45] Ruo Zhang, Ping-Sing Tsai, James Edwin Cryer, and Mubarak Shah. Shape-from-shading: a survey. IEEE transactions on pattern analysis and machine intelligence, 21(8): 690-706, 1999. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.503, + 0.905, + 0.572 + ], + "angle": 0, + "content": "[46] Xiuming Zhang, Pratul P Srinivasan, Boyang Deng, Paul Debevec, William T Freeman, and Jonathan T Barron. Nerfactor: Neural factorization of shape and reflectance under an unknown illumination. ACM Transactions on Graphics (TOG), 40(6):1-18, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.574, + 0.905, + 0.655 + ], + "angle": 0, + "content": "[47] Jingsen Zhu, Fujun Luan, Yuchi Huo, Zihao Lin, Zhihua Zhong, Dianbing Xi, Rui Wang, Hujun Bao, Jiaxiang Zheng, and Rui Tang. Learning-based inverse rendering of complex indoor scenes with differentiable monte carlo raytracing. In SIGGRAPH Asia 2022 Conference Papers. ACM, 2022. 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.657, + 0.905, + 0.727 + ], + "angle": 0, + "content": "[48] Rui Zhu, Zhengqin Li, Janarbek Matai, Fatih Porikli, and Manmohan Chandraker. Irisformer: Dense vision transformers for single-image inverse rendering in indoor scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2822-2831, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.727 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "5782" + } + ] +] \ No newline at end of file diff --git a/2025/Channel-wise Noise Scheduled Diffusion for Inverse Rendering in Indoor Scenes/e287505c-edc5-4f3b-9021-f507eb27467f_origin.pdf b/2025/Channel-wise Noise Scheduled Diffusion for Inverse Rendering in Indoor Scenes/e287505c-edc5-4f3b-9021-f507eb27467f_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..098aa187348c1819256e348ad6b33292213e88b4 --- /dev/null +++ b/2025/Channel-wise Noise Scheduled Diffusion for Inverse Rendering in Indoor Scenes/e287505c-edc5-4f3b-9021-f507eb27467f_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:feb845b82e8723d16c565a45d5887a6b4d996a2b5f331b726b50a1b92464449e +size 6003401 diff --git a/2025/Channel-wise Noise Scheduled Diffusion for Inverse Rendering in Indoor Scenes/full.md b/2025/Channel-wise Noise Scheduled Diffusion for Inverse Rendering in Indoor Scenes/full.md new file mode 100644 index 0000000000000000000000000000000000000000..deff2ce66ce07130394bef376ce1acda3838f8f3 --- /dev/null +++ b/2025/Channel-wise Noise Scheduled Diffusion for Inverse Rendering in Indoor Scenes/full.md @@ -0,0 +1,307 @@ +# Channel-wise Noise Scheduled Diffusion for Inverse Rendering in Indoor Scenes + +JunYong Choi $^{1,2}$ Min-cheol Sagong $^{1}$ SeokYeong Lee $^{1,2}$ Seung-Won Jung $^{2}$ Ig-Jae Kim $^{1,3,4}$ Junghyun Cho $^{1,3,4}$ + +$^{1}$ Korea Institute of Science and Technology(KIST) $^{2}$ Korea University $^{3}$ AI-Robotics, KIST School, University of Science and Technology $^{4}$ Yonsei-KIST Convergence Research Institute + +{happily, mcsagong, shapin94, drjay, jhcho}@kist.re.kr swjung83@korea.ac.kr + +![](images/7377b628915045238014c3364bbc99a164411f2b4f5b225be147cbd081143e76.jpg) +Figure 1. Diffusion-Based Inverse Rendering. We present a diffusion-based inverse rendering framework that addresses the two competing goals of accuracy and diversity in inverse rendering by using two distinct models, each dedicated to achieving one of the objectives. (a) Our first model predicts a single accurate solution, while (b) the second model presents diverse possible solutions. (c) This enables practical applications such as object insertion and material editing (e.g., increasing the roughness of the brown table). + +# Abstract + +We propose a diffusion-based inverse rendering framework that decomposes a single RGB image into geometry, material, and lighting. Inverse rendering is inherently ill-posed, making it difficult to predict a single accurate solution. To address this challenge, recent generative model-based methods aim to present a range of possible solutions. However, finding a single accurate solution and generating diverse solutions can be conflicting. In this paper, we propose a channel-wise noise scheduling approach that allows a single diffusion model architecture to achieve two conflicting objectives. The resulting two diffusion models, trained with different channel-wise noise schedules, can predict a single highly accurate solution and present multiple possible solutions. The experimental results demonstrate the superiority of our two models in terms of both diversity and accuracy, which translates to enhanced performance in downstream applications such as object insertion and material editing. + +# 1. Introduction + +Single-image inverse rendering is a long-standing challenge in computer vision and computer graphics that involves the decomposition of an RGB image into its constituent elements of material, lighting, and geometry. This technology enables a wide range of applications, including object insertion, material editing, and relighting. Although inverse rendering has a clear solution, interpreting a single observed radiance can be complex, as it may result from various combinations of material, geometry, and lighting. For instance, a high radiance might be due to the material being white, intense lighting, or a pronounced specular component caused by the Fresnel effect. This inherent ill-posedness of inverse rendering introduces ambiguity and difficulty, particularly when working with a single-image. Therefore, it is important not only to predict an optimal solution, but also to propose multiple potential solutions. However, as discussed in + +other studies [15, 41], enforcing only the reconstruction loss cannot improve the sample diversity, leading to conflicts between the objectives. + +Learning-based methods [4, 5, 19, 47, 48] leveraging large-scale synthetic datasets have demonstrated excellent performance. Strong supervised signals from reconstruction losses on well-aligned ground truths have enabled these methods to effectively decompose scenes. However, because these methods are based on deterministic approaches, they do not account for the solution space of possible combinations. Recently, inverse rendering studies based on Latent Diffusion Models (LDM) [28, 31] have emerged [14, 22, 43]. These probabilistic methods, which rely on priors learned from large image-text datasets, not only provide multiple solutions but also exhibit excellent predictive capability. However, they often overlook the trade-off between sample diversity and prediction accuracy and they struggle to account for the dependencies between modalities. Additionally, they rely on the pretrained three-channel autoencoder of LDM, which is effective at generating image-based data such as RGB images but falls short in handling high-dimensional data like per-pixel lighting, which is crucial for applications such as photorealistic object insertion and material editing. + +While probabilistic methods are effective for capturing multiple potential solutions in ill-posed problems like inverse rendering, they can be sub-optimal for predicting a definitive ground truth. We found that diffusion models exhibit a trade-off between diversity and accuracy depending on the entire timestep $T$ . Motivated by these observations, we propose two conditional diffusion models, each focusing on either diversity or accuracy. Our diffusion models predict multiple modalities, such as a depth map, normal map, diffuse albedo map, roughness map, and per-pixel environment map, simultaneously from a single RGB image. The first proposed model, the Progressive Diffusion Model (PDM), focuses on diversity using a large $T$ for training. The second model, the Switchable Diffusion Model (SDM), is trained with a smaller $T$ to prioritize accuracy. To improve the quality of inverse rendering, a channel-wise noise scheduling approach is introduced to adjust the transition between modalities during generation. + +Our PDM and SDM operate in a low-resolution pixel space rather than a latent space, allowing them to handle per-pixel environment maps. To mitigate the high computational costs associated with handling high-dimensional data directly, we introduce a method for encoding these maps into compact feature vectors. Additionally, we present a simple RGB-guided super-resolution model for up-sampling outputs. + +We evaluated the performance by comparing it with other inverse rendering baselines on synthetic [4] and real-world [9, 37, 40] datasets. The experimental results showed + +that SDM outperforms in images with fewer ambiguous regions, where accurate predictions are more feasible, while PDM excels in complex scenes with numerous ambiguous regions requiring diverse predictions, indicating the complementary nature of SDM and PDM. In summary, our main contributions are as follows: + +1. We propose a novel diffusion-based inverse rendering framework that incorporates per-pixel environment maps while considering the correlations between modalities. +2. We propose channel-wise noise scheduling, which improves generation quality by adjusting the modality generation transition in multi-modality diffusion models. +3. We demonstrate superior performance over existing baselines on both real and synthetic datasets, highlighting the importance of considering both diversity and accuracy in inverse rendering. + +# 2. Related Works + +Inverse rendering. Inverse rendering involves the extraction of depth, normal, SVBRDF, and lighting information from RGB images. Initial research concentrated primarily on specific areas, such as shape-from-shading [13, 25, 45] and intrinsic image decomposition [1, 10, 16, 21]. Advances in deep learning have facilitated the study of complex properties, such as SVBRDF [6, 17, 18, 33] and lighting [8, 9, 36, 39]. Li et al. [19, 20] introduced OpenRooms and a spherical-Gaussian lighting model. IRISformer [48] leverages long-range attention through a dense vision transformer. Zhu et al. [47] introduced the refined InteriorVerse dataset and incorporated Monte Carlo ray tracing into the inverse rendering. Choi et al. [4, 5] extended this approach to a multi-view setup. However, all of these methods rely on deterministic models, limiting their ability to propose alternative solutions. Our work diverges by proposing PDM and SDM, capable of generating multiple solutions and surpassing deterministic models in flexibility and accuracy. + +Diffusion models. Diffusion models [7, 11, 35], which have recently emerged as powerful tools for image generation, are now being applied to various tasks. Several studies have trained diffusion models directly on target datasets to generate materials [38], distant lighting [23, 27], and human face reflectance [26]. Recently, research has begun exploring inverse rendering by utilizing the strong priors of pretrained LDMs. Kocsis et al. [14] fine-tuned an LDM on synthetic datasets, enabling the generation of multiple plausible material maps. RGB $\leftrightarrow$ X [43] demonstrated that LDMs can be used for both inverse rendering and rendering. Luo et al. [22] trained ControlNet [44] to predict multiple intrinsic modalities. However, since these methods predict each modality independently, they fail to address inter-modality dependencies and struggle with environment maps, limiting realistic lighting inference and applications. In contrast, our framework accounts for these dependencies, enhancing + +![](images/5947288bcd84badd1106de1e993bcd22acee452e3816230a4ceefbc08002b84e.jpg) +Figure 2. Entire pipeline. Our channel-wise noise scheduling assists the DM's inference by adjusting the transitions between modalities based on the timestep. Since the DM operates at a low resolution, modalities are up-sampled through our SRM. + +sample fidelity, while per-pixel lighting enables realistic, high-fidelity object insertion. + +# 3. Method + +Given an input RGB image, denoted as $\mathrm{I} \in \mathbb{R}^{3 \times H \times W}$ , our goal is to obtain the following outputs: a normal map $\mathrm{N} \in \mathbb{R}^{3 \times H \times W}$ , a depth map $\mathrm{D} \in \mathbb{R}^{H \times W}$ , a diffuse albedo map $\mathrm{A} \in \mathbb{R}^{3 \times H \times W}$ , a roughness map $\mathrm{R} \in \mathbb{R}^{H \times W}$ , and a per-pixel environment map, represented as $\mathrm{E} \in \mathbb{R}^{3 \times C \times H \times W}$ where $H$ and $W$ represent the image dimensions, and $C$ denotes the angular resolution of the environment map. First, we introduce the synthetic dataset used in our experiments. Second, we introduce a method to encode per-pixel lighting into a neural feature vector. Next, we introduce the conditional diffusion model that serves as the foundation of our approach, along with two channel-wise noise scheduling techniques that allow us to customize the model for our desired application. Since the diffusion model operates in the low-resolution pixel space, we introduce a RGB-guided super-resolution model for upsampling. The entire pipeline is shown in Fig. 2. + +# 3.1. Dataset + +To train our conditional diffusion model, we need well-aligned ground truth scene information (N, D, A, R, E) pairs for RGB images. For this purpose, two synthetic datasets, OpenRooms [20] and InteriorVerse [47], are suitable; however, the lighting part of InteriorVerse is currently not available. Additionally, OpenRooms often contains overly simplistic images (such as views of walls or incorrectly placed light sources), which hinder the diffusion model from learning the data distribution. Therefore, we used the OpenRooms FF dataset [4], which involves filtering out inappropriate images from OpenRooms. + +# 3.2. Implicit Lighting Representation + +As a method for handling high-dimensional data such as per-pixel environment maps, a recent method called MAIR++ [5] introduced the implicit lighting representation + +(ILR), which represents environment maps as neural feature vectors. MAIR++ processes the RGB, geometry, and material using a CNN to obtain the ILR, which is subsequently processed by three different decoders to obtain an environment map, shading map, and specular radiance. This approach can render images in a realistic and computationally efficient way. We are motivated to use the ILR as our lighting representation. However, since the previous ILR [5] is not inferred from the environment map, it neither accurately reproduces the ground truth lighting nor effectively captures the detailed high-dynamic range (HDR) radiance values. In contrast, we re-design the ILR as an MLP-based autoencoder that takes the environment map as input. + +Since the HDR radiance values of the environment map cover a very wide range, using them directly as input causes the network to diverge easily. Therefore, we transform the input into the log space using $\log 1p$ [29] and add batch normalization to the encoder MLP. These two simple techniques allow the network to handle HDR lighting more effectively, leading to significantly improved performance. Mathematically, the encoder is defined as follows: + +$$ +\boldsymbol {f} = \tanh (\operatorname {E n c o d e r} (\log 1 p (\bar {\mathrm {E}}))), \tag {1} +$$ + +where $\bar{\cdot}$ represents the ground truth, and $f\in \mathbb{R}^{96}$ is the ILR neural feature vector. To enable realistic rendering through $f$ , the three decoders are applied: + +$$ +\mathrm {E} = \exp \operatorname {m 1} (\operatorname {D e c o d e r} _ {E} (\boldsymbol {f})) +$$ + +$$ +\mathrm {S} = \exp \operatorname {m 1} (\operatorname {D e c o d e r} _ {S} (\boldsymbol {f})), \tag {2} +$$ + +$$ +\mathrm {I} _ {s} = \exp \operatorname {m 1} (\operatorname {D e c o d e r} _ {I} (\boldsymbol {f}, \bar {\mathrm {R}}, \gamma (\boldsymbol {r}), \bar {\mathrm {N}} \cdot \boldsymbol {v})). +$$ + +where $\mathsf{expm1}$ [29] is the inverse of $\log 1p$ , $S \in \mathbb{R}^{3 \times H \times W}$ denotes the shading map, $I_s \in \mathbb{R}^{3 \times H \times W}$ represents the specular image, $\gamma$ refers to the frequency positional encoding, and $\boldsymbol{r}$ and $\boldsymbol{v}$ indicate the reflection direction and the viewing direction, respectively. Through these three decoders, $\boldsymbol{f}$ can not only reconstruct the environment map but also be used to render arbitrary views. + +
TNLDLALRLELIL
MSE ↓var ↑MSE ↓var ↑MSE ↓var ↑MSE ↓var ↑MSE ↓var ↑MSE ↓var ↑
10002.2360.1830.0650.0200.4180.0345.5180.10612.6679.610.1790.042
1002.2380.1820.0640.0190.4130.0345.3540.09412.6677.620.1770.041
102.0260.1380.0590.0130.3740.0245.4400.06612.1362.380.1620.03
11.8170.00.0610.00.3340.05.5450.011.370.060.1370.0
+ +Table 1. Ablation of training $T$ . The best values in each column are highlighted in bold, and the worst values are underlined. In most cases, as the value of $T$ decreases, prediction accuracy improves, but the diversity of the samples is reduced. To evaluate under general conditions, we fixed $\tau$ to 1.0 for all settings, and the DDIM [34] steps were selected as 2, 2, 2, and 1, respectively. + +# 3.3. Diffusion-based Inverse Rendering + +Since the introduction of LDM [28, 31], many diffusion-based methods [14, 22, 43] have used pretrained LDMs that have strong priors for images. However, these methods do not provide meaningful priors for high-dimensional data such as environment maps. Therefore, we adopt a cascaded structure [12] that first generates $\mathbf{z}_0 = (\mathrm{N}_L,\mathrm{D}_L,\mathrm{A}_L,\mathrm{R}_L,\boldsymbol {f}_L)$ and then increases the resolution, where the subscript $L$ denotes a low-resolution image of size $H_{L}\times W_{L}$ . During inference, our conditional diffusion model (DM) takes $\mathrm{I}_L$ as a condition and iteratively denoises the Gaussian noise image $\mathbf{z}_{T - 1}\sim \mathcal{N}(0,1)$ to produce $\mathbf{z}_0$ . During training, we predict the velocity $\mathbf{v}_t$ [32], and the loss function $L_{\mathrm{DM}}$ is given by: + +$$ +L _ {\mathrm {D M}} = \left\| \mathbf {v} _ {t} - \mathrm {D M} \left(t, \mathbf {z} _ {t}, \mathrm {I} _ {L}, \operatorname {C L I P} \left(\mathrm {I} _ {L}\right)\right) \right\| _ {2} ^ {2}, \tag {3} +$$ + +$$ +\mathbf {v} _ {t} = \sqrt {\bar {\alpha} _ {t}} \mathbf {z} _ {T - 1} - \sqrt {1 - \bar {\alpha} _ {t}} \mathbf {z} _ {0}, \tag {4} +$$ + +where $T$ is the entire timestep, $t \in [0, T - 1]$ is uniformly sampled during training, $\bar{\alpha}_t$ is the noise scheduler, $\mathbf{z}_t = \sqrt{\bar{\alpha}_t} \mathbf{z}_0 + \sqrt{1 - \bar{\alpha}_t} \mathbf{z}_{T-1}$ , and CLIP is the pretrained image encoder [30]. $\mathrm{I}_L$ is concatenated with $\mathbf{z}_t$ , and the CLIP embedding is used as a cross-attention condition. + +Channel-wise noise scheduling. Since we generate multiple interdependent modalities simultaneously, we hypothesize that there is an optimal transition to generate these modalities in inverse rendering. For example, according to previous works [18, 42, 46], geometry typically has the least uncertainty and supports the prediction of material and lighting, so generating geometry first could be advantageous. On the other hand, since lighting has the highest uncertainty, generating it last could be beneficial. To validate this hypothesis, we divide $\mathbf{z}_0$ into three categories based on the modality attributes: geometry $(\mathrm{N}_L,\mathrm{D}_L)$ , material $(\mathrm{A}_L,\mathrm{R}_L)$ , and lighting $(f_{L})$ , and accordingly apply different noise scheduling to the three parts. Specifically, following Chen et al. [2], we select the cosine scheduler as the base noise scheduler and parameterize it as follows: + +$$ +\begin{array}{l} v _ {s} = \cos (s \cdot 2 \pi) ^ {2 \tau}, v _ {b} = \cos (b \cdot 2 \pi) ^ {2 \tau} \\ \alpha_ {t} = \cos \left(\left(\left(b - s\right) \cdot t + s\right) \cdot 2 \pi\right) ^ {2 \tau}, \bar {\alpha} _ {t} = \frac {v _ {b} - \alpha_ {t}}{v _ {b} - v _ {s}}. \tag {5} \\ \end{array} +$$ + +Throughout the paper, we fix $s = 0.008$ and $b = 1$ and control the scheduler using only $\tau$ . When $\tau = 1$ , $\bar{\alpha}_t$ is equivalent to the cosine scheduler. The behavior of the noise scheduler as a function of $\tau$ is shown in Fig. 3. + +![](images/98d4a95e6e7b27792db66e13204d66f1c7569a2af432ef5974ed70517122b0c3.jpg) +Figure 3. Parameterizing noise scheduler. A larger $\tau$ value results in a slower generation of the modality. + +Fig. 3 shows that as the value of $\tau$ increases, the noise increases more gradually. To determine the optimal modality generation order using $\tau$ , we conducted experiments by applying different values of $\tau$ to geometry, material, and lighting, meaning that different noise levels are applied separately to each channel of the data during the denoising process of the diffusion model. The ablation study in Tab. 7 confirms that generating in the order of geometry, material, and lighting is the most effective, aligning with our intuition. As a result, through our channel-wise noise scheduling, it focuses on specific modalities at different timesteps, enabling more effective predictions of other modalities in subsequent timesteps. + +In general, $T$ in diffusion models is set to a large number, such as 1000. This helps approximate the data distribution and increases the diversity of the generated data [24]. However, we note that such diffusion models are challenging in predicting a single definitive ground truth. To this end, we conducted an ablation study on $T$ values by training the DM with $T = 1000, 100, 10$ , and 1. For each setting, we performed sampling 20 times in the test set to measure the MSE and the variance (var) of the samples. Tab. 1 shows that reducing $T$ improves accuracy at the cost of diversity in most cases. In the extreme case of $T = 1$ , the method becomes deterministic. + +To explore models focused on diversity and accuracy, we + +develop two types of diffusion models: one with $T = 1000$ and the other with $T = 1$ . Following the ablation study in Tab. 7, we select $\tau = (0.9, 1.2, 1.5)$ when $T = 1000$ to prioritize diversity. Due to its gradual generation process, we name this model the Progressive Diffusion Model (PDM). However, this gradual generation process cannot be applied to the diffusion model with $T = 1$ . Moreover, $T = 1$ limits the ability of the diffusion model to exploit its full potential over multiple timesteps [3]. To this end, we define our new noise scheduler, as shown in Tab. 2. + +
tt mod 3 = 0t mod 3 = 1t mod 3 = 2T-1
\( \bar{a}_{t}^{G} \)1100
\( \bar{a}_{t}^{M} \)1010
\( \bar{a}_{t}^{L} \)0110
+ +A value of 0 implies that pure Gaussian noise is added, while a value of 1 means that no noise is added. With our new noise scheduler, $T$ is redefined as $3n + 1$ where $n > 0$ . After the initial prediction at $t = T - 1$ , the noise is switched per channel, providing conditions for the next diffusion step. Taking into account this feature of the switchable noise scheduling, we name this model Switchable Diffusion Model (SDM). The SDM is trained to predict the modality with an SNR of 0 at each step, allowing for more refined predictions by utilizing the previous step's predictions. + +# 3.4. RGB-Guided Super Resolution + +We use a UNet as the RGB-guided super-resolution model, consisting of a shared encoder and five separate decoders. The encoder processes the RGB image into a dense feature map, which, along with the rendered diffuse $\mathrm{I}_d$ and specular $\mathrm{I}_s$ images from $\mathbf{z}_0$ , is fed into each decoder. Our RGB-guided Super-Resolution Model (SRM) is as follows: + +$$ +\mathrm {N}, \mathrm {D}, \mathrm {A}, \mathrm {R}, \boldsymbol {f} = \operatorname {S R M} \left(\mathbf {z} _ {0}, \mathrm {I}, \mathrm {I} _ {d}, \mathrm {I} _ {s}\right), \tag {6} +$$ + +# 4. Evaluation + +In this section, we evaluate our method on both synthetic and real-world datasets. We first evaluate on a synthetic dataset, followed by inverse rendering and object insertion on a real-world spatially varying lighting dataset. Furthermore, we evaluated our method on the sub-tasks of inverse rendering, namely intrinsic decomposition and geometry prediction. More details on training, model architectures, and additional experimental results are provided in the Supplementary Materials. + +Evaluation settings. We used $C = 128$ , $H = 240$ , $W = 320$ , $H_{L} = 60$ , and $W_{L} = 80$ . For SDM, we set $T = 4$ . In all experiments, PDM generated 10 samples and the mean and variance of these samples were evaluated. + +Baselines. We compared our method with state-of-the-art inverse rendering frameworks capable of per-pixel lighting inference, specifically those by Li et al. [19] and Zhu et al. [47]. Both methods were trained on the OpenRooms FF [4] dataset. However, since the code for Zhu's lighting model has not been released, we used their pretrained lighting model, which was trained on the InteriorVerse dataset. + +![](images/c53b69649ece6843a6ea562e8e28547d1e8a0cc5e3b784f70ccc5ff248faa6f8.jpg) +Figure 4. Qualitative synthetic evaluation. Only our method successfully decomposes specular radiance into material and lighting. + +# 4.1. Evaluation on Synthetic Scenes + +Tab. 3 includes quantitative prediction results on the synthetic [4] test dataset. Although our target is single-view, we include results from the multi-view baselines MAIR [4] and MAIR++ [5]. Our SDM, which prioritizes prediction accuracy, demonstrated superior performance compared to existing single-image baselines in all areas except roughness, and our PDM also demonstrated respectable performance comparable to the baselines. Notably, our lighting accuracy was comparable to that of MAIR, and the albedo accuracy outperformed even the multi-view methods. Additionally, it can be observed that both MAIR++ and our method, which are capable of ILR-based neural rendering, exhibit significantly superior re-rendering (I) performance. Figure 4 presents an illustrative example of this fact. Even in images with strong specular radiance, our SDM and PDM successfully decomposed and re-rendered the image in contrast to the failure of the baseline methods. + +Table 2. SDM noise scheduler. Here, $\bar{\alpha}_t^G$ , $\bar{\alpha}_t^M$ , and $\bar{\alpha}_t^L$ are the schedulers for geometry, material, and lighting, respectively. + +
MSE(×10-2)N ↓D ↓A ↓R ↓E ↓I ↓
Li et al. [19]3.7820.1350.8696.49615.490.491
Zhu et al. [47]1.7380.0620.5284.208*31.15*3.565
SDM1.5020.0390.3494.96913.350.156
PDM(Mean)1.8310.0470.3944.81914.080.150
PDM(Best)1.7820.0440.3814.64413.750.146
MAIR [4]1.0100.1400.6393.45713.790.784
MAIR++ [5]1.0100.0070.4782.49711.610.058
+ +Table 3. Quantitative synthetic evaluation. Bold text indicates the best performance among single-view methods, while underlined text represents the second-best. * denotes models trained on the InteriorVerse dataset. + +![](images/52170fac97f4b9667a11d36f51110a77e1637bc6bd17f40d9dc8936eb6617719.jpg) +Input Image + +![](images/34e2a9088aa5e6a47585e3a46c69df49b0eda539371e05038eccf396e906c58a.jpg) +N + +![](images/b120857f00227edf41eb9d88dd7888366ca5d6cd79c9dcd1a9faa766cbcd520a.jpg) +A + +![](images/c94dcc6c1a73d9888ed29b4e3044b29f32168c0f8995a7ea927caaeebc566e42.jpg) +Id + +![](images/e743fb28423b44050dcfd35343fd145f0e4c2b5ce49893f412ff7a17cb791b83.jpg) +I + +![](images/f9c98b41c00b6411f5738a01c3cb455eee3017fbe5df5595772a64d789f2416b.jpg) +Figure 5. Qualitative real-world evaluation. Our method successfully recognizes shadows under the desk, achieving plausible geometry prediction and realistic re-rendering results. +Li et al. + +![](images/1bfaa68d1c4334484df848afbddcce15da9448ec4a0055e34dfea64971f4e484.jpg) +Zhu et al. + +![](images/59d547382ef8f11af670ea66072873a7c00d6fbe9a6e2a24d52a71cde35bc479.jpg) +Ours(SDM) +Figure 6. Qualitative object insertion evaluation. We accurately predicted the direction of light and shadows even in real-world images. Additionally, our rendered objects were closer to the ground truth in terms of hue and brightness compared to the baselines. + +![](images/e225683f39f3c94a1947dcefa2d163877e8f7d76c1d811baf96694a8d972d6d0.jpg) +RGB $\leftrightarrow$ X + +# 4.2. Evaluation on Real-world Scenes + +Qualitative evaluation. We evaluated the generalization capability of our method on real-world images from Garon et al. [9]. As shown in Fig. 5, both SDM and PDM successfully separated the albedo even in the presence of strong shadows under the desk. Li et al. [19] partially removed the shadows but failed to predict the geometry, while Zhu et al. [47] accurately predicted the normal but failed to estimate the albedo. + +
Li et al. [19]Zhu et al. [47]SDMGround Truth
0.1780.1530.3220.348
+ +Table 4. AMT user study on object insertion. Our method was selected by nearly as many users as the ground truth. + +Lighting Evaluation. We also evaluated lighting performance through object insertion, using the object insertion tool from Li et al. [19]. For fairness, we excluded PDM + +and focused on evaluating only SDM. As shown in Fig. 6, our method accurately predicts strong directional lighting, as well as the surrounding light's tone and shadows, resulting in the most realistic object insertions. In contrast, Zhu et al. [47] failed to predict the correct light direction, while Li et al. [19] struggles with realistic rendering due to a lack of high-frequency details in the environment map. The diffusion-based method RGB $\leftrightarrow$ X [43] relies on inpainting for object insertion, which compromises the fidelity of the inserted objects and leads to visible artifacts. Additionally, since it cannot handle environment maps, it struggled to accurately predict the direction of light. For a more quantitative evaluation, we conducted an AMT user study on object insertion, including the ground truth. The results in Tab. 4 confirmed the superiority of our method. Users were asked: "Of the four images, please select the one that shows white rabbits blending naturally into the surrounding light." Our method was selected by users nearly as frequently as the ground truth. + +![](images/1bab5064a7f194e7f390e7ae32edf7cb716177c71e87514ee173925c3ea94c30.jpg) +Figure 7. Intrinsic decomposition. Our method successfully decomposes the input image into albedo and shading. RGB $\leftrightarrow$ X [43], leveraging its strong prior, achieves successful albedo prediction but struggles with realistic rendering, resulting in excessive brightness in the left sample and incorrect specular highlights in the right sample. + +![](images/c0dee3c1dfe2482606782c0c655f53590dd6bb0432acacbfc2fb17c4a525cb33.jpg) + +![](images/7ad47f7574d1da95d022fcf6feb415f3b3d70a58102718caebca2fecb497fd8f.jpg) +Figure 8. Geometry prediction. Our method achieves the most accurate prediction of the ceiling structure. + +
MethodA ↓Hue ↓Texture ↓WHDR ↓S ↓I ↓
Li et al. [19]0.9245.7810.4730.3011.1620.021
Zhu et al. [47]1.6854.6380.4720.3011.4410.028
SDM0.7055.0560.4590.2931.1810.006
PDM(Mean)0.7575.2600.4590.3061.1560.005
PDM(Best)0.6234.5270.4560.2781.1130.005
RGB↔X [43](Mean)0.6023.6900.4480.1681.2610.055
RGB↔X [43](Best)0.3313.0910.4410.1461.1960.052
+ +# 4.3. Evaluation on Sub-tasks + +Intrinsic decomposition. We evaluated our method on intrinsic decomposition using the recently introduced MAW [40] dataset, which provides more refined annotations for 888 images, allowing more quantitative comparisons than the WHDR metric used in the IIW [1] dataset. Tab. 5 presents comparisons for MSE for A, S, and I, along with metrics for hue, texture, and WHDR for A. The hue metric evaluates the chromaticity component, while the texture metric is based on LPIPS perceptual similarity. Our method generally outperformed existing deterministic methods [19, 47]. RGB $\leftrightarrow$ X [43], which learns strong priors from diverse large datasets, excels in albedo infer + +ence; however, due to its reliance on diffusion-based rendering, it reconstructed the input images with the lowest accuracy. In contrast, our method, trained only on OpenRooms FF, achieved the highest performance in shading estimation and most faithfully reproduced input images, demonstrating high fidelity. This can also be observed in Fig. 7. In the left sample, RGB $\leftrightarrow$ X [43] fails to predict the shading intensity as it does not account for HDR lighting. Similarly, in the right sample, RGB $\leftrightarrow$ X [43] introduces specular highlights on the desk that are absent in the input image, resulting in a loss of fidelity. In contrast, our method predicts spatially consistent albedo and accurately estimates brightness, enabling a more realistic reproduction of the input image. + +Geometry prediction. Tab. 6 includes the geometry prediction evaluation on the DIODE [37] indoor dataset. Similarly to the intrinsic decomposition task, both our PDM and SDM demonstrate superior or comparable performance to the baselines. Fig. 8 provides qualitative examples, demonstrating that our method exhibits strong generalization capability in geometry prediction. + +Table 5. Intrinsic decomposition evaluation on MAW [40]. Bold text indicates the best performance. + +
MetricLi et al. [19]Zhu et al. [47]SDMPDM(Mean)PDM(Best)
N ↓0.1960.1710.1470.1520.144
D ↓0.0590.0570.0570.0550.050
I ↓0.0150.0130.0030.0020.002
+ +Table 6. Geometry prediction evaluation on DIODE [37]. Bold text indicates the best performance. + +# 4.4. Modality Dependencies + +To generate a reasonable solution for inverse rendering, it is crucial to account for the dependencies between modalities. For instance, if a material is predicted to be darker, the lighting should be predicted brighter to compensate for the input image. However, existing diffusion-based methods [14, 22, 43] infer each modality independently, making it difficult to capture such dependencies. In contrast, our PDM generates all modalities simultaneously, enabling it to account for these relationships. Fig. 9 illustrates this with + +![](images/91cf3562a8c816a8758fe3db0188570cfeea05733ccf70ee790f2a8079f7db5e.jpg) +Figure 9. Modality dependencies. Two samples of N, D, A, S, and R predicted by PDM. + +two different samples generated by PDM. For the center table in the top image, the first sample predicts the normal pointing upward with bright shading, resulting in a darker albedo. Conversely, the second sample predicts the normal facing the camera, producing darker shading and a brighter albedo. Additionally, for the area under the cabinet in the bottom image, the first sample interprets it as a shadow and predicts a bright albedo, while the second sample treats it as part of the albedo and generates bright shading. + +# 4.5. Ablations + +Modality generation order. To explore the optimal modality generation order, we conducted an ablation study on the modality generation sequence. There are no restrictions on $\tau$ ; however, we selected three values: $\tau = 0.9, 1.2,$ and 1.5. These values were applied separately to $\tau_{G}$ (geometry), $\tau_{M}$ (material), and $\tau_{L}$ (lighting), and the models were trained independently. The performance was then evaluated on the test dataset. Tab. 7 shows that this channel-wise noise scheduling can significantly impact the inverse rendering quality. In particular, $\tau_{G}, \tau_{M}, \tau_{L} = (0.9, 1.2, 1.5)$ produces the best quality, aligning with our intuition. In contrast, generating in reverse order results in a detrimental effect. Fig. 10 provides examples of these two cases. The reverse order (L-M-G) exhibits artifacts in areas with strong shadows, whereas our method produces plausible results. + +
τGτMτLNL↓DL↓AL↓RL↓EL↓IL↓
1.01.01.02.1400.0620.4035.75412.190.163
0.91.21.52.0240.0590.3765.17611.940.159
0.91.51.22.1040.0600.3905.72312.150.165
1.20.91.52.0730.0600.3885.39412.200.161
1.21.50.92.2020.0640.4115.88212.330.167
1.50.91.22.2110.0640.4055.66412.360.160
1.51.20.92.2450.0660.4145.98612.530.166
+ +Iterative Prediction of SDM. We conducted an ablation study to evaluate the benefits of iterative prediction for + +![](images/86f3fd7445c5b2da785fd0a4f9e4f9bd156c25fb5bd13c4a3828960cbb251844.jpg) +Figure 10. Comparison of modality generation order. Generating in the order of geometry, material, and lighting helps effectively infer shadows. + +Table 7. Ablation of the generation order. The best values are highlighted in bold, while the worst values are underlined. + +
TN ↓D ↓A ↓R ↓E ↓I ↓
11.7330.0600.3185.83711.240.132
41.5780.0520.2885.74310.960.137
71.6300.0500.2925.75910.970.139
+ +Table 8. Ablation of SDM iterative prediction. Bold text indicates the best performance. + +SDM. In Tab. 8, the improvements at $T = 4$ and $T = 7$ over $T = 1$ indicate that iterative prediction utilizes previously predicted modalities to inform subsequent predictions. However, the performance drop at $T = 7$ compared to $T = 4$ suggests that excessive iterations do not always yield better results. + +# 5. Conclusion + +We tackled inverse rendering with two opposing goals: attaining diversity and ensuring accuracy. To address these objectives, we proposed a diffusion-based framework, in which channel-wise noise scheduling allowed each diffusion model to better understand the correlations between different modalities. + +Limitations and future work. Our approach has the limitation of requiring two separate models; thus, integrating them into a unified model would enhance usability. We explored re-rendering error as an alternative metric to overcome the constraint of depending only on the predicted variance of PDM for estimating inverse rendering uncertainty, but it could be imprecise. Further research on the handling of uncertainty could improve the complementarity of SDM and PDM. Moreover, our method refrains from exploiting the priors of pretrained LDMs. Using those priors to process per-pixel environment maps might enhance the performance and expand the applications of LDMs. + +Acknowledgements. This work was supported by Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government(MSIT)(No.RS-2023-00227592, Development of 3D Object Identification Technology Robust to Viewpoint Changes) + +# References + +[1] Sean Bell, Kavita Bala, and Noah Snively. Intrinsic images in the wild. ACM Transactions on Graphics (TOG), 33(4): 1-12, 2014. 2, 7 +[2] Ting Chen. On the importance of noise scheduling for diffusion models. arXiv preprint arXiv:2301.10972, 2023. 4 +[3] Jooyoung Choi, Jungbeom Lee, Chaehun Shin, Sungwon Kim, Hyunwoo Kim, and Sungroh Yoon. Perception prioritized training of diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11472-11481, 2022. 5 +[4] JunYong Choi, SeokYeong Lee, Haesol Park, Seung-Won Jung, Ig-Jae Kim, and Junghyun Cho. Mair: multi-view attention inverse rendering with 3d spatially-varying lighting estimation. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8392-8401. IEEE, 2023. 2, 3, 5 +[5] JunYong Choi, SeokYeong Lee, Haesol Park, Seung-Won Jung, Ig-Jae Kim, and Junghyun Cho. Mair++: Improving multi-view attention inverse rendering with implicit lighting representation. arXiv preprint arXiv:2408.06707, 2024. 2, 3, 5 +[6] Valentin Deschaintre, Miika Aittala, Fredo Durand, George Drettakis, and Adrien Bousseau. Single-image svbrdf capture with a rendering-aware deep network. ACM Transactions on Graphics (ToG), 37(4):1-15, 2018. 2 +[7] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 2 +[8] Marc-Andre Gardner, Yannick Hold-Geoffroy, Kalyan Sunkavalli, Christian Gagne, and Jean-François Lalonde. Deep parametric indoor lighting estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7175–7183, 2019. 2 +[9] Mathieu Garon, Kalyan Sunkavalli, Sunil Hadap, Nathan Carr, and Jean-François Lalonde. Fast spatially-varying indoor lighting estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6908-6917, 2019. 2, 6 +[10] Roger Grosse, Micah K Johnson, Edward H Adelson, and William T Freeman. Ground truth dataset and baseline evaluations for intrinsic image algorithms. In 2009 IEEE 12th International Conference on Computer Vision, pages 2335-2342. IEEE, 2009. 2 +[11] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2 +[12] Jonathan Ho, Chitwan Sahara, William Chan, David J Fleet, Mohammad Norouzi, and Tim Salimans. Cascaded diffusion models for high fidelity image generation. Journal of Machine Learning Research, 23(47):1-33, 2022. 4 +[13] Berthold KP Horn and Michael J Brooks. Shape from shading. MIT press, 1989. 2 +[14] Peter Kocsis, Vincent Sitzmann, and Matthias Nießner. Intrinsic image diffusion for single-view material estimation. arXiv preprint arXiv:2312.12274, 2023. 2, 4, 7 + +[15] Soochan Lee, Junsoo Ha, and Gunhee Kim. Harmonizing maximum likelihood with gans for multimodal conditional generation. arXiv preprint arXiv:1902.09225, 2019. 2 +[16] Zhengqi Li and Noah Snavely. Cginintrinsics: Better intrinsic image decomposition through physically-based rendering. In Proceedings of the European conference on computer vision (ECCV), pages 371-387, 2018. 2 +[17] Zhengqin Li, Kalyan Sunkavalli, and Manmohan Chandraker. Materials for masses: Svbrdf acquisition with a single mobile phone image. In Proceedings of the European conference on computer vision (ECCV), pages 72-87, 2018. 2 +[18] Zhengqin Li, Zexiang Xu, Ravi Ramamoorthi, Kalyan Sunkavalli, and Manmohan Chandraker. Learning to reconstruct shape and spatially-varying reflectance from a single image. ACM Transactions on Graphics (TOG), 37(6):1-11, 2018. 2, 4 +[19] Zhengqin Li, Mohammad Shafiei, Ravi Ramamoorthi, Kalyan Sunkavalli, and Manmohan Chandraker. Inverse rendering for complex indoor scenes: Shape, spatially-varying lighting and svbrdf from a single image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2475-2484, 2020. 2, 5, 6, 7 +[20] Zhengqin Li, Ting-Wei Yu, Shen Sang, Sarah Wang, Meng Song, Yuhan Liu, Yu-Ying Yeh, Rui Zhu, Nitesh Gundavarapu, Jia Shi, et al. Openrooms: An open framework for photorealistic indoor scene datasets. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7190–7199, 2021. 2, 3 +[21] Jundan Luo, Zhaoyang Huang, Yijin Li, Xiaowei Zhou, Guofeng Zhang, and Hujun Bao. Niid-net: Adapting surface normal knowledge for intrinsic image decomposition in indoor scenes. IEEE Transactions on Visualization and Computer Graphics, 26(12):3434-3445, 2020. 2 +[22] Jundan Luo, Duygu Ceylan, Jae Shin Yoon, Nanxuan Zhao, Julien Philip, Anna Frühstück, Wenbin Li, Christian Richardt, and Tuanfeng Wang. Intrinsicdiffusion: Joint intrinsic layers from latent diffusion models. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 2, 4, 7 +[23] Linjie Lyu, Ayush Tewari, Marc Habermann, Shunsuke Saito, Michael Zollhöfer, Thomas Leimkuhler, and Christian Theobalt. Diffusion posterior illumination for ambiguity-aware inverse rendering. ACM Transactions on Graphics (TOG), 42(6):1-14, 2023. 2 +[24] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International conference on machine learning, pages 8162-8171. PMLR, 2021. 4 +[25] Geoffrey Oxholm and Ko Nishino. Shape and reflectance from natural illumination. In European Conference on Computer Vision, pages 528-541. Springer, 2012. 2 +[26] Foivos Paraperas Papantoniou, Alexandros Lattas, Stylianos Moschoglou, and Stefanos Zafeiriou. Relightify: Relightable 3d faces from a single image via diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8806-8817, 2023. 2 + +[27] Pakkapon Phongthawee, Worameth Chinchuthakun, Nontaphat Sinsunthithet, Varun Jampani, Amit Raj, Pramook Khungurn, and Supasorn Suwajanakorn. Diffusionlight: Light probes for free by painting a chrome ball. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 98-108, 2024. 2 +[28] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 2, 4 +[29] William H. Press, Saul A. Teukolsky, William T. Vetterling, and Brian P. Flannery. Numerical Recipes 3rd Edition: The Art of Scientific Computing. Cambridge University Press, 2007. 3 +[30] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 4 +[31] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2, 4 +[32] Tim Salimans and Jonathan Ho. Progressive distillation for fast sampling of diffusion models. arXiv preprint arXiv:2202.00512, 2022. 4 +[33] Shen Sang and Manmohan Chandraker. Single-shot neural relighting and svbrdf estimation. In European Conference on Computer Vision, pages 85-101. Springer, 2020. 2 +[34] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 4 +[35] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. Advances in neural information processing systems, 32, 2019. 2 +[36] Pratul P Srinivasan, Ben Mildenhall, Matthew Tancik, Jonathan T Barron, Richard Tucker, and Noah Snively. Lighthouse: Predicting lighting volumes for spatially-coherent illumination. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8080-8089, 2020. 2 +[37] Igor Vasiljevic, Nick Kolkin, Shanyi Zhang, Ruotian Luo, Haochen Wang, Falcon Z Dai, Andrea F Daniele, Mohammadreza Mostajabi, Steven Basart, Matthew R Walter, et al. Diode: A dense indoor and outdoor depth dataset. arXiv preprint arXiv:1908.00463, 2019. 2, 7 +[38] Giuseppe Vecchio, Renato Sortino, Simone Palazzo, and Concetto Spampinato. Matfuse: controllable material generation with diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4429-4438, 2024. 2 +[39] Zian Wang, Jonah Philion, Sanja Fidler, and Jan Kautz. Learning indoor inverse rendering with 3d spatially-varying lighting. In Proceedings of the IEEE/CVF International + +Conference on Computer Vision, pages 12538-12547, 2021. 2 +[40] Jiaye Wu, Sanjoy Chowdhury, Hariharmano Shanmugaraja, David Jacobs, and Soumyadip Sengupta. Measured albedo in the wild: Filling the gap in intrinsics evaluation. In 2023 IEEE International Conference on Computational Photography (ICCP), pages 1-12. IEEE, 2023. 2, 7 +[41] Jiahui Yu, Zhe Lin, Jimei Yang, Xiaohui Shen, Xin Lu, and Thomas S Huang. Free-form image inpainting with gated convolution. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4471-4480, 2019. 2 +[42] Ye Yu and William AP Smith. Inverserendernet: Learning single image inverse rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3155-3164, 2019. 4 +[43] Zheng Zeng, Valentin Deschaintre, Iliyan Georgiev, Yannick Hold-Geoffroy, Yiwei Hu, Fujun Luan, Ling-Qi Yan, and Miloš Hasan. Rgb $\leftrightarrow$ x: Image decomposition and synthesis using material-and lighting-aware diffusion models. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 2, 4, 6, 7 +[44] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 2 +[45] Ruo Zhang, Ping-Sing Tsai, James Edwin Cryer, and Mubarak Shah. Shape-from-shading: a survey. IEEE transactions on pattern analysis and machine intelligence, 21(8): 690-706, 1999. 2 +[46] Xiuming Zhang, Pratul P Srinivasan, Boyang Deng, Paul Debevec, William T Freeman, and Jonathan T Barron. Nerfactor: Neural factorization of shape and reflectance under an unknown illumination. ACM Transactions on Graphics (TOG), 40(6):1-18, 2021. 4 +[47] Jingsen Zhu, Fujun Luan, Yuchi Huo, Zihao Lin, Zhihua Zhong, Dianbing Xi, Rui Wang, Hujun Bao, Jiaxiang Zheng, and Rui Tang. Learning-based inverse rendering of complex indoor scenes with differentiable monte carlo raytracing. In SIGGRAPH Asia 2022 Conference Papers. ACM, 2022. 2, 3, 5, 6, 7 +[48] Rui Zhu, Zhengqin Li, Janarbek Matai, Fatih Porikli, and Manmohan Chandraker. Irisformer: Dense vision transformers for single-image inverse rendering in indoor scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2822-2831, 2022. 2 \ No newline at end of file diff --git a/2025/Channel-wise Noise Scheduled Diffusion for Inverse Rendering in Indoor Scenes/images.zip b/2025/Channel-wise Noise Scheduled Diffusion for Inverse Rendering in Indoor Scenes/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..578977b0fcc81c281c2b92c29d04ef015d75b412 --- /dev/null +++ b/2025/Channel-wise Noise Scheduled Diffusion for Inverse Rendering in Indoor Scenes/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a9fe5e3798d22de5c54dd975e29608fb7edf8de50ea0bfbd2dd27b1e1e79f338 +size 801917 diff --git a/2025/Channel-wise Noise Scheduled Diffusion for Inverse Rendering in Indoor Scenes/layout.json b/2025/Channel-wise Noise Scheduled Diffusion for Inverse Rendering in Indoor Scenes/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..e947af2bcf9754f9744a26ff3bd6425bed46d0d5 --- /dev/null +++ b/2025/Channel-wise Noise Scheduled Diffusion for Inverse Rendering in Indoor Scenes/layout.json @@ -0,0 +1,9017 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 58, + 103, + 552, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 103, + 552, + 121 + ], + "spans": [ + { + "bbox": [ + 58, + 103, + 552, + 121 + ], + "type": "text", + "content": "Channel-wise Noise Scheduled Diffusion for Inverse Rendering in Indoor Scenes" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 111, + 141, + 498, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 141, + 498, + 172 + ], + "spans": [ + { + "bbox": [ + 111, + 141, + 498, + 172 + ], + "type": "text", + "content": "JunYong Choi" + }, + { + "bbox": [ + 111, + 141, + 498, + 172 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 111, + 141, + 498, + 172 + ], + "type": "text", + "content": " Min-cheol Sagong" + }, + { + "bbox": [ + 111, + 141, + 498, + 172 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 111, + 141, + 498, + 172 + ], + "type": "text", + "content": " SeokYeong Lee" + }, + { + "bbox": [ + 111, + 141, + 498, + 172 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 111, + 141, + 498, + 172 + ], + "type": "text", + "content": " Seung-Won Jung" + }, + { + "bbox": [ + 111, + 141, + 498, + 172 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 111, + 141, + 498, + 172 + ], + "type": "text", + "content": " Ig-Jae Kim" + }, + { + "bbox": [ + 111, + 141, + 498, + 172 + ], + "type": "inline_equation", + "content": "^{1,3,4}" + }, + { + "bbox": [ + 111, + 141, + 498, + 172 + ], + "type": "text", + "content": " Junghyun Cho" + }, + { + "bbox": [ + 111, + 141, + 498, + 172 + ], + "type": "inline_equation", + "content": "^{1,3,4}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 176, + 477, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 176, + 477, + 219 + ], + "spans": [ + { + "bbox": [ + 133, + 176, + 477, + 219 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 133, + 176, + 477, + 219 + ], + "type": "text", + "content": "Korea Institute of Science and Technology(KIST) " + }, + { + "bbox": [ + 133, + 176, + 477, + 219 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 133, + 176, + 477, + 219 + ], + "type": "text", + "content": "Korea University " + }, + { + "bbox": [ + 133, + 176, + 477, + 219 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 133, + 176, + 477, + 219 + ], + "type": "text", + "content": "AI-Robotics, KIST School, University of Science and Technology " + }, + { + "bbox": [ + 133, + 176, + 477, + 219 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 133, + 176, + 477, + 219 + ], + "type": "text", + "content": "Yonsei-KIST Convergence Research Institute" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 108, + 221, + 499, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 221, + 499, + 232 + ], + "spans": [ + { + "bbox": [ + 108, + 221, + 499, + 232 + ], + "type": "text", + "content": "{happily, mcsagong, shapin94, drjay, jhcho}@kist.re.kr swjung83@korea.ac.kr" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 58, + 253, + 552, + 403 + ], + "blocks": [ + { + "bbox": [ + 58, + 253, + 552, + 403 + ], + "lines": [ + { + "bbox": [ + 58, + 253, + 552, + 403 + ], + "spans": [ + { + "bbox": [ + 58, + 253, + 552, + 403 + ], + "type": "image", + "image_path": "7377b628915045238014c3364bbc99a164411f2b4f5b225be147cbd081143e76.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 410, + 555, + 455 + ], + "lines": [ + { + "bbox": [ + 55, + 410, + 555, + 455 + ], + "spans": [ + { + "bbox": [ + 55, + 410, + 555, + 455 + ], + "type": "text", + "content": "Figure 1. Diffusion-Based Inverse Rendering. We present a diffusion-based inverse rendering framework that addresses the two competing goals of accuracy and diversity in inverse rendering by using two distinct models, each dedicated to achieving one of the objectives. (a) Our first model predicts a single accurate solution, while (b) the second model presents diverse possible solutions. (c) This enables practical applications such as object insertion and material editing (e.g., increasing the roughness of the brown table)." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 151, + 464, + 200, + 476 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 464, + 200, + 476 + ], + "spans": [ + { + "bbox": [ + 151, + 464, + 200, + 476 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 54, + 510, + 297, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 510, + 297, + 715 + ], + "spans": [ + { + "bbox": [ + 54, + 510, + 297, + 715 + ], + "type": "text", + "content": "We propose a diffusion-based inverse rendering framework that decomposes a single RGB image into geometry, material, and lighting. Inverse rendering is inherently ill-posed, making it difficult to predict a single accurate solution. To address this challenge, recent generative model-based methods aim to present a range of possible solutions. However, finding a single accurate solution and generating diverse solutions can be conflicting. In this paper, we propose a channel-wise noise scheduling approach that allows a single diffusion model architecture to achieve two conflicting objectives. The resulting two diffusion models, trained with different channel-wise noise schedules, can predict a single highly accurate solution and present multiple possible solutions. The experimental results demonstrate the superiority of our two models in terms of both diversity and accuracy, which translates to enhanced performance in downstream applications such as object insertion and material editing." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 315, + 502, + 394, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 502, + 394, + 514 + ], + "spans": [ + { + "bbox": [ + 315, + 502, + 394, + 514 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 521, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 555, + 715 + ], + "type": "text", + "content": "Single-image inverse rendering is a long-standing challenge in computer vision and computer graphics that involves the decomposition of an RGB image into its constituent elements of material, lighting, and geometry. This technology enables a wide range of applications, including object insertion, material editing, and relighting. Although inverse rendering has a clear solution, interpreting a single observed radiance can be complex, as it may result from various combinations of material, geometry, and lighting. For instance, a high radiance might be due to the material being white, intense lighting, or a pronounced specular component caused by the Fresnel effect. This inherent ill-posedness of inverse rendering introduces ambiguity and difficulty, particularly when working with a single-image. Therefore, it is important not only to predict an optimal solution, but also to propose multiple potential solutions. However, as discussed in" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "5773" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "type": "text", + "content": "other studies [15, 41], enforcing only the reconstruction loss cannot improve the sample diversity, leading to conflicts between the objectives." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 110, + 295, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 110, + 295, + 361 + ], + "spans": [ + { + "bbox": [ + 56, + 110, + 295, + 361 + ], + "type": "text", + "content": "Learning-based methods [4, 5, 19, 47, 48] leveraging large-scale synthetic datasets have demonstrated excellent performance. Strong supervised signals from reconstruction losses on well-aligned ground truths have enabled these methods to effectively decompose scenes. However, because these methods are based on deterministic approaches, they do not account for the solution space of possible combinations. Recently, inverse rendering studies based on Latent Diffusion Models (LDM) [28, 31] have emerged [14, 22, 43]. These probabilistic methods, which rely on priors learned from large image-text datasets, not only provide multiple solutions but also exhibit excellent predictive capability. However, they often overlook the trade-off between sample diversity and prediction accuracy and they struggle to account for the dependencies between modalities. Additionally, they rely on the pretrained three-channel autoencoder of LDM, which is effective at generating image-based data such as RGB images but falls short in handling high-dimensional data like per-pixel lighting, which is crucial for applications such as photorealistic object insertion and material editing." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 363, + 295, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 363, + 295, + 578 + ], + "spans": [ + { + "bbox": [ + 55, + 363, + 295, + 578 + ], + "type": "text", + "content": "While probabilistic methods are effective for capturing multiple potential solutions in ill-posed problems like inverse rendering, they can be sub-optimal for predicting a definitive ground truth. We found that diffusion models exhibit a trade-off between diversity and accuracy depending on the entire timestep " + }, + { + "bbox": [ + 55, + 363, + 295, + 578 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 363, + 295, + 578 + ], + "type": "text", + "content": ". Motivated by these observations, we propose two conditional diffusion models, each focusing on either diversity or accuracy. Our diffusion models predict multiple modalities, such as a depth map, normal map, diffuse albedo map, roughness map, and per-pixel environment map, simultaneously from a single RGB image. The first proposed model, the Progressive Diffusion Model (PDM), focuses on diversity using a large " + }, + { + "bbox": [ + 55, + 363, + 295, + 578 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 363, + 295, + 578 + ], + "type": "text", + "content": " for training. The second model, the Switchable Diffusion Model (SDM), is trained with a smaller " + }, + { + "bbox": [ + 55, + 363, + 295, + 578 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 363, + 295, + 578 + ], + "type": "text", + "content": " to prioritize accuracy. To improve the quality of inverse rendering, a channel-wise noise scheduling approach is introduced to adjust the transition between modalities during generation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 580, + 295, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 580, + 295, + 675 + ], + "spans": [ + { + "bbox": [ + 55, + 580, + 295, + 675 + ], + "type": "text", + "content": "Our PDM and SDM operate in a low-resolution pixel space rather than a latent space, allowing them to handle per-pixel environment maps. To mitigate the high computational costs associated with handling high-dimensional data directly, we introduce a method for encoding these maps into compact feature vectors. Additionally, we present a simple RGB-guided super-resolution model for up-sampling outputs." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "content": "We evaluated the performance by comparing it with other inverse rendering baselines on synthetic [4] and real-world [9, 37, 40] datasets. The experimental results showed" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 553, + 143 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 143 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 143 + ], + "type": "text", + "content": "that SDM outperforms in images with fewer ambiguous regions, where accurate predictions are more feasible, while PDM excels in complex scenes with numerous ambiguous regions requiring diverse predictions, indicating the complementary nature of SDM and PDM. In summary, our main contributions are as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 145, + 553, + 264 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 315, + 145, + 553, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 145, + 553, + 180 + ], + "spans": [ + { + "bbox": [ + 315, + 145, + 553, + 180 + ], + "type": "text", + "content": "1. We propose a novel diffusion-based inverse rendering framework that incorporates per-pixel environment maps while considering the correlations between modalities." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 182, + 553, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 182, + 553, + 216 + ], + "spans": [ + { + "bbox": [ + 313, + 182, + 553, + 216 + ], + "type": "text", + "content": "2. We propose channel-wise noise scheduling, which improves generation quality by adjusting the modality generation transition in multi-modality diffusion models." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 217, + 553, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 217, + 553, + 264 + ], + "spans": [ + { + "bbox": [ + 313, + 217, + 553, + 264 + ], + "type": "text", + "content": "3. We demonstrate superior performance over existing baselines on both real and synthetic datasets, highlighting the importance of considering both diversity and accuracy in inverse rendering." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 274, + 405, + 286 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 274, + 405, + 286 + ], + "spans": [ + { + "bbox": [ + 314, + 274, + 405, + 286 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 295, + 553, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 295, + 553, + 510 + ], + "spans": [ + { + "bbox": [ + 313, + 295, + 553, + 510 + ], + "type": "text", + "content": "Inverse rendering. Inverse rendering involves the extraction of depth, normal, SVBRDF, and lighting information from RGB images. Initial research concentrated primarily on specific areas, such as shape-from-shading [13, 25, 45] and intrinsic image decomposition [1, 10, 16, 21]. Advances in deep learning have facilitated the study of complex properties, such as SVBRDF [6, 17, 18, 33] and lighting [8, 9, 36, 39]. Li et al. [19, 20] introduced OpenRooms and a spherical-Gaussian lighting model. IRISformer [48] leverages long-range attention through a dense vision transformer. Zhu et al. [47] introduced the refined InteriorVerse dataset and incorporated Monte Carlo ray tracing into the inverse rendering. Choi et al. [4, 5] extended this approach to a multi-view setup. However, all of these methods rely on deterministic models, limiting their ability to propose alternative solutions. Our work diverges by proposing PDM and SDM, capable of generating multiple solutions and surpassing deterministic models in flexibility and accuracy." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 510, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 553, + 713 + ], + "type": "text", + "content": "Diffusion models. Diffusion models [7, 11, 35], which have recently emerged as powerful tools for image generation, are now being applied to various tasks. Several studies have trained diffusion models directly on target datasets to generate materials [38], distant lighting [23, 27], and human face reflectance [26]. Recently, research has begun exploring inverse rendering by utilizing the strong priors of pretrained LDMs. Kocsis et al. [14] fine-tuned an LDM on synthetic datasets, enabling the generation of multiple plausible material maps. RGB " + }, + { + "bbox": [ + 313, + 510, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\leftrightarrow" + }, + { + "bbox": [ + 313, + 510, + 553, + 713 + ], + "type": "text", + "content": " X [43] demonstrated that LDMs can be used for both inverse rendering and rendering. Luo et al. [22] trained ControlNet [44] to predict multiple intrinsic modalities. However, since these methods predict each modality independently, they fail to address inter-modality dependencies and struggle with environment maps, limiting realistic lighting inference and applications. In contrast, our framework accounts for these dependencies, enhancing" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "5774" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 68, + 553, + 186 + ], + "blocks": [ + { + "bbox": [ + 56, + 68, + 553, + 186 + ], + "lines": [ + { + "bbox": [ + 56, + 68, + 553, + 186 + ], + "spans": [ + { + "bbox": [ + 56, + 68, + 553, + 186 + ], + "type": "image", + "image_path": "5947288bcd84badd1106de1e993bcd22acee452e3816230a4ceefbc08002b84e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 194, + 555, + 218 + ], + "lines": [ + { + "bbox": [ + 54, + 194, + 555, + 218 + ], + "spans": [ + { + "bbox": [ + 54, + 194, + 555, + 218 + ], + "type": "text", + "content": "Figure 2. Entire pipeline. Our channel-wise noise scheduling assists the DM's inference by adjusting the transitions between modalities based on the timestep. Since the DM operates at a low resolution, modalities are up-sampled through our SRM." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 238, + 294, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 238, + 294, + 264 + ], + "spans": [ + { + "bbox": [ + 54, + 238, + 294, + 264 + ], + "type": "text", + "content": "sample fidelity, while per-pixel lighting enables realistic, high-fidelity object insertion." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 272, + 112, + 285 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 272, + 112, + 285 + ], + "spans": [ + { + "bbox": [ + 55, + 272, + 112, + 285 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 292, + 296, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 292, + 296, + 496 + ], + "spans": [ + { + "bbox": [ + 55, + 292, + 296, + 496 + ], + "type": "text", + "content": "Given an input RGB image, denoted as " + }, + { + "bbox": [ + 55, + 292, + 296, + 496 + ], + "type": "inline_equation", + "content": "\\mathrm{I} \\in \\mathbb{R}^{3 \\times H \\times W}" + }, + { + "bbox": [ + 55, + 292, + 296, + 496 + ], + "type": "text", + "content": ", our goal is to obtain the following outputs: a normal map " + }, + { + "bbox": [ + 55, + 292, + 296, + 496 + ], + "type": "inline_equation", + "content": "\\mathrm{N} \\in \\mathbb{R}^{3 \\times H \\times W}" + }, + { + "bbox": [ + 55, + 292, + 296, + 496 + ], + "type": "text", + "content": ", a depth map " + }, + { + "bbox": [ + 55, + 292, + 296, + 496 + ], + "type": "inline_equation", + "content": "\\mathrm{D} \\in \\mathbb{R}^{H \\times W}" + }, + { + "bbox": [ + 55, + 292, + 296, + 496 + ], + "type": "text", + "content": ", a diffuse albedo map " + }, + { + "bbox": [ + 55, + 292, + 296, + 496 + ], + "type": "inline_equation", + "content": "\\mathrm{A} \\in \\mathbb{R}^{3 \\times H \\times W}" + }, + { + "bbox": [ + 55, + 292, + 296, + 496 + ], + "type": "text", + "content": ", a roughness map " + }, + { + "bbox": [ + 55, + 292, + 296, + 496 + ], + "type": "inline_equation", + "content": "\\mathrm{R} \\in \\mathbb{R}^{H \\times W}" + }, + { + "bbox": [ + 55, + 292, + 296, + 496 + ], + "type": "text", + "content": ", and a per-pixel environment map, represented as " + }, + { + "bbox": [ + 55, + 292, + 296, + 496 + ], + "type": "inline_equation", + "content": "\\mathrm{E} \\in \\mathbb{R}^{3 \\times C \\times H \\times W}" + }, + { + "bbox": [ + 55, + 292, + 296, + 496 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 55, + 292, + 296, + 496 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 55, + 292, + 296, + 496 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 292, + 296, + 496 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 55, + 292, + 296, + 496 + ], + "type": "text", + "content": " represent the image dimensions, and " + }, + { + "bbox": [ + 55, + 292, + 296, + 496 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 55, + 292, + 296, + 496 + ], + "type": "text", + "content": " denotes the angular resolution of the environment map. First, we introduce the synthetic dataset used in our experiments. Second, we introduce a method to encode per-pixel lighting into a neural feature vector. Next, we introduce the conditional diffusion model that serves as the foundation of our approach, along with two channel-wise noise scheduling techniques that allow us to customize the model for our desired application. Since the diffusion model operates in the low-resolution pixel space, we introduce a RGB-guided super-resolution model for upsampling. The entire pipeline is shown in Fig. 2." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 504, + 115, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 504, + 115, + 514 + ], + "spans": [ + { + "bbox": [ + 55, + 504, + 115, + 514 + ], + "type": "text", + "content": "3.1. Dataset" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 521, + 295, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 295, + 653 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 295, + 653 + ], + "type": "text", + "content": "To train our conditional diffusion model, we need well-aligned ground truth scene information (N, D, A, R, E) pairs for RGB images. For this purpose, two synthetic datasets, OpenRooms [20] and InteriorVerse [47], are suitable; however, the lighting part of InteriorVerse is currently not available. Additionally, OpenRooms often contains overly simplistic images (such as views of walls or incorrectly placed light sources), which hinder the diffusion model from learning the data distribution. Therefore, we used the OpenRooms FF dataset [4], which involves filtering out inappropriate images from OpenRooms." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 659, + 233, + 673 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 659, + 233, + 673 + ], + "spans": [ + { + "bbox": [ + 55, + 659, + 233, + 673 + ], + "type": "text", + "content": "3.2. Implicit Lighting Representation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": "As a method for handling high-dimensional data such as per-pixel environment maps, a recent method called MAIR++ [5] introduced the implicit lighting representation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 238, + 555, + 393 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 238, + 555, + 393 + ], + "spans": [ + { + "bbox": [ + 313, + 238, + 555, + 393 + ], + "type": "text", + "content": "(ILR), which represents environment maps as neural feature vectors. MAIR++ processes the RGB, geometry, and material using a CNN to obtain the ILR, which is subsequently processed by three different decoders to obtain an environment map, shading map, and specular radiance. This approach can render images in a realistic and computationally efficient way. We are motivated to use the ILR as our lighting representation. However, since the previous ILR [5] is not inferred from the environment map, it neither accurately reproduces the ground truth lighting nor effectively captures the detailed high-dynamic range (HDR) radiance values. In contrast, we re-design the ILR as an MLP-based autoencoder that takes the environment map as input." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 395, + 556, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 395, + 556, + 491 + ], + "spans": [ + { + "bbox": [ + 313, + 395, + 556, + 491 + ], + "type": "text", + "content": "Since the HDR radiance values of the environment map cover a very wide range, using them directly as input causes the network to diverge easily. Therefore, we transform the input into the log space using " + }, + { + "bbox": [ + 313, + 395, + 556, + 491 + ], + "type": "inline_equation", + "content": "\\log 1p" + }, + { + "bbox": [ + 313, + 395, + 556, + 491 + ], + "type": "text", + "content": " [29] and add batch normalization to the encoder MLP. These two simple techniques allow the network to handle HDR lighting more effectively, leading to significantly improved performance. Mathematically, the encoder is defined as follows:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 369, + 507, + 555, + 522 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 369, + 507, + 555, + 522 + ], + "spans": [ + { + "bbox": [ + 369, + 507, + 555, + 522 + ], + "type": "interline_equation", + "content": "\\boldsymbol {f} = \\tanh (\\operatorname {E n c o d e r} (\\log 1 p (\\bar {\\mathrm {E}}))), \\tag {1}", + "image_path": "e1d78e63c5b54e8563d92f87ae0732eb7260d18342657cd534baa69ebd7cea0b.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 530, + 555, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 530, + 555, + 568 + ], + "spans": [ + { + "bbox": [ + 313, + 530, + 555, + 568 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 530, + 555, + 568 + ], + "type": "inline_equation", + "content": "\\bar{\\cdot}" + }, + { + "bbox": [ + 313, + 530, + 555, + 568 + ], + "type": "text", + "content": " represents the ground truth, and " + }, + { + "bbox": [ + 313, + 530, + 555, + 568 + ], + "type": "inline_equation", + "content": "f\\in \\mathbb{R}^{96}" + }, + { + "bbox": [ + 313, + 530, + 555, + 568 + ], + "type": "text", + "content": " is the ILR neural feature vector. To enable realistic rendering through " + }, + { + "bbox": [ + 313, + 530, + 555, + 568 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 313, + 530, + 555, + 568 + ], + "type": "text", + "content": ", the three decoders are applied:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 347, + 576, + 463, + 589 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 576, + 463, + 589 + ], + "spans": [ + { + "bbox": [ + 347, + 576, + 463, + 589 + ], + "type": "interline_equation", + "content": "\\mathrm {E} = \\exp \\operatorname {m 1} (\\operatorname {D e c o d e r} _ {E} (\\boldsymbol {f}))", + "image_path": "5f57a1cbfa77fde1344fecb1b634496f6b80577727b7b41f0d9380c256771dc1.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 348, + 591, + 553, + 604 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 348, + 591, + 553, + 604 + ], + "spans": [ + { + "bbox": [ + 348, + 591, + 553, + 604 + ], + "type": "interline_equation", + "content": "\\mathrm {S} = \\exp \\operatorname {m 1} (\\operatorname {D e c o d e r} _ {S} (\\boldsymbol {f})), \\tag {2}", + "image_path": "d6cd01a5aac1435af4be9ce8e649dd5092198039bb8e3ff450e620e8b0f38692.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 347, + 605, + 521, + 619 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 605, + 521, + 619 + ], + "spans": [ + { + "bbox": [ + 347, + 605, + 521, + 619 + ], + "type": "interline_equation", + "content": "\\mathrm {I} _ {s} = \\exp \\operatorname {m 1} (\\operatorname {D e c o d e r} _ {I} (\\boldsymbol {f}, \\bar {\\mathrm {R}}, \\gamma (\\boldsymbol {r}), \\bar {\\mathrm {N}} \\cdot \\boldsymbol {v})).", + "image_path": "76d831edb54696a82f554c9872f0da80a379e80e1028bb3c7213880e9af71301.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 628, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 628, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 628, + 555, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 628, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\mathsf{expm1}" + }, + { + "bbox": [ + 313, + 628, + 555, + 714 + ], + "type": "text", + "content": " [29] is the inverse of " + }, + { + "bbox": [ + 313, + 628, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\log 1p" + }, + { + "bbox": [ + 313, + 628, + 555, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 628, + 555, + 714 + ], + "type": "inline_equation", + "content": "S \\in \\mathbb{R}^{3 \\times H \\times W}" + }, + { + "bbox": [ + 313, + 628, + 555, + 714 + ], + "type": "text", + "content": " denotes the shading map, " + }, + { + "bbox": [ + 313, + 628, + 555, + 714 + ], + "type": "inline_equation", + "content": "I_s \\in \\mathbb{R}^{3 \\times H \\times W}" + }, + { + "bbox": [ + 313, + 628, + 555, + 714 + ], + "type": "text", + "content": " represents the specular image, " + }, + { + "bbox": [ + 313, + 628, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 313, + 628, + 555, + 714 + ], + "type": "text", + "content": " refers to the frequency positional encoding, and " + }, + { + "bbox": [ + 313, + 628, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\boldsymbol{r}" + }, + { + "bbox": [ + 313, + 628, + 555, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 628, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\boldsymbol{v}" + }, + { + "bbox": [ + 313, + 628, + 555, + 714 + ], + "type": "text", + "content": " indicate the reflection direction and the viewing direction, respectively. Through these three decoders, " + }, + { + "bbox": [ + 313, + 628, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\boldsymbol{f}" + }, + { + "bbox": [ + 313, + 628, + 555, + 714 + ], + "type": "text", + "content": " can not only reconstruct the environment map but also be used to render arbitrary views." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5775" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 61, + 70, + 550, + 140 + ], + "blocks": [ + { + "bbox": [ + 61, + 70, + 550, + 140 + ], + "lines": [ + { + "bbox": [ + 61, + 70, + 550, + 140 + ], + "spans": [ + { + "bbox": [ + 61, + 70, + 550, + 140 + ], + "type": "table", + "html": "
TNLDLALRLELIL
MSE ↓var ↑MSE ↓var ↑MSE ↓var ↑MSE ↓var ↑MSE ↓var ↑MSE ↓var ↑
10002.2360.1830.0650.0200.4180.0345.5180.10612.6679.610.1790.042
1002.2380.1820.0640.0190.4130.0345.3540.09412.6677.620.1770.041
102.0260.1380.0590.0130.3740.0245.4400.06612.1362.380.1620.03
11.8170.00.0610.00.3340.05.5450.011.370.060.1370.0
", + "image_path": "b03deee3066ac285ec0abd47e63d15c86ed50386fe7cece914580869cf5ad2b7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 148, + 555, + 183 + ], + "lines": [ + { + "bbox": [ + 54, + 148, + 555, + 183 + ], + "spans": [ + { + "bbox": [ + 54, + 148, + 555, + 183 + ], + "type": "text", + "content": "Table 1. Ablation of training " + }, + { + "bbox": [ + 54, + 148, + 555, + 183 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 54, + 148, + 555, + 183 + ], + "type": "text", + "content": ". The best values in each column are highlighted in bold, and the worst values are underlined. In most cases, as the value of " + }, + { + "bbox": [ + 54, + 148, + 555, + 183 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 54, + 148, + 555, + 183 + ], + "type": "text", + "content": " decreases, prediction accuracy improves, but the diversity of the samples is reduced. To evaluate under general conditions, we fixed " + }, + { + "bbox": [ + 54, + 148, + 555, + 183 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 54, + 148, + 555, + 183 + ], + "type": "text", + "content": " to 1.0 for all settings, and the DDIM [34] steps were selected as 2, 2, 2, and 1, respectively." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 202, + 241, + 216 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 202, + 241, + 216 + ], + "spans": [ + { + "bbox": [ + 55, + 202, + 241, + 216 + ], + "type": "text", + "content": "3.3. Diffusion-based Inverse Rendering" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 220, + 296, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 220, + 296, + 376 + ], + "spans": [ + { + "bbox": [ + 56, + 220, + 296, + 376 + ], + "type": "text", + "content": "Since the introduction of LDM [28, 31], many diffusion-based methods [14, 22, 43] have used pretrained LDMs that have strong priors for images. However, these methods do not provide meaningful priors for high-dimensional data such as environment maps. Therefore, we adopt a cascaded structure [12] that first generates " + }, + { + "bbox": [ + 56, + 220, + 296, + 376 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_0 = (\\mathrm{N}_L,\\mathrm{D}_L,\\mathrm{A}_L,\\mathrm{R}_L,\\boldsymbol {f}_L)" + }, + { + "bbox": [ + 56, + 220, + 296, + 376 + ], + "type": "text", + "content": " and then increases the resolution, where the subscript " + }, + { + "bbox": [ + 56, + 220, + 296, + 376 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 56, + 220, + 296, + 376 + ], + "type": "text", + "content": " denotes a low-resolution image of size " + }, + { + "bbox": [ + 56, + 220, + 296, + 376 + ], + "type": "inline_equation", + "content": "H_{L}\\times W_{L}" + }, + { + "bbox": [ + 56, + 220, + 296, + 376 + ], + "type": "text", + "content": ". During inference, our conditional diffusion model (DM) takes " + }, + { + "bbox": [ + 56, + 220, + 296, + 376 + ], + "type": "inline_equation", + "content": "\\mathrm{I}_L" + }, + { + "bbox": [ + 56, + 220, + 296, + 376 + ], + "type": "text", + "content": " as a condition and iteratively denoises the Gaussian noise image " + }, + { + "bbox": [ + 56, + 220, + 296, + 376 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{T - 1}\\sim \\mathcal{N}(0,1)" + }, + { + "bbox": [ + 56, + 220, + 296, + 376 + ], + "type": "text", + "content": " to produce " + }, + { + "bbox": [ + 56, + 220, + 296, + 376 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_0" + }, + { + "bbox": [ + 56, + 220, + 296, + 376 + ], + "type": "text", + "content": ". During training, we predict the velocity " + }, + { + "bbox": [ + 56, + 220, + 296, + 376 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_t" + }, + { + "bbox": [ + 56, + 220, + 296, + 376 + ], + "type": "text", + "content": " [32], and the loss function " + }, + { + "bbox": [ + 56, + 220, + 296, + 376 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{DM}}" + }, + { + "bbox": [ + 56, + 220, + 296, + 376 + ], + "type": "text", + "content": " is given by:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 89, + 386, + 295, + 401 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 386, + 295, + 401 + ], + "spans": [ + { + "bbox": [ + 89, + 386, + 295, + 401 + ], + "type": "interline_equation", + "content": "L _ {\\mathrm {D M}} = \\left\\| \\mathbf {v} _ {t} - \\mathrm {D M} \\left(t, \\mathbf {z} _ {t}, \\mathrm {I} _ {L}, \\operatorname {C L I P} \\left(\\mathrm {I} _ {L}\\right)\\right) \\right\\| _ {2} ^ {2}, \\tag {3}", + "image_path": "de1791208b275904b454e68e25d30c9c93706e22fd812d733aa52fa64e59de90.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 112, + 410, + 295, + 423 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 112, + 410, + 295, + 423 + ], + "spans": [ + { + "bbox": [ + 112, + 410, + 295, + 423 + ], + "type": "interline_equation", + "content": "\\mathbf {v} _ {t} = \\sqrt {\\bar {\\alpha} _ {t}} \\mathbf {z} _ {T - 1} - \\sqrt {1 - \\bar {\\alpha} _ {t}} \\mathbf {z} _ {0}, \\tag {4}", + "image_path": "a3982129aa9b500a2b078badbd3f7a98c56c10735568b9a5a9bcdab2d7b14fbd.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 426, + 295, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 426, + 295, + 486 + ], + "spans": [ + { + "bbox": [ + 55, + 426, + 295, + 486 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 426, + 295, + 486 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 426, + 295, + 486 + ], + "type": "text", + "content": " is the entire timestep, " + }, + { + "bbox": [ + 55, + 426, + 295, + 486 + ], + "type": "inline_equation", + "content": "t \\in [0, T - 1]" + }, + { + "bbox": [ + 55, + 426, + 295, + 486 + ], + "type": "text", + "content": " is uniformly sampled during training, " + }, + { + "bbox": [ + 55, + 426, + 295, + 486 + ], + "type": "inline_equation", + "content": "\\bar{\\alpha}_t" + }, + { + "bbox": [ + 55, + 426, + 295, + 486 + ], + "type": "text", + "content": " is the noise scheduler, " + }, + { + "bbox": [ + 55, + 426, + 295, + 486 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_t = \\sqrt{\\bar{\\alpha}_t} \\mathbf{z}_0 + \\sqrt{1 - \\bar{\\alpha}_t} \\mathbf{z}_{T-1}" + }, + { + "bbox": [ + 55, + 426, + 295, + 486 + ], + "type": "text", + "content": ", and CLIP is the pretrained image encoder [30]. " + }, + { + "bbox": [ + 55, + 426, + 295, + 486 + ], + "type": "inline_equation", + "content": "\\mathrm{I}_L" + }, + { + "bbox": [ + 55, + 426, + 295, + 486 + ], + "type": "text", + "content": " is concatenated with " + }, + { + "bbox": [ + 55, + 426, + 295, + 486 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_t" + }, + { + "bbox": [ + 55, + 426, + 295, + 486 + ], + "type": "text", + "content": ", and the CLIP embedding is used as a cross-attention condition." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 487, + 295, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 487, + 295, + 666 + ], + "spans": [ + { + "bbox": [ + 55, + 487, + 295, + 666 + ], + "type": "text", + "content": "Channel-wise noise scheduling. Since we generate multiple interdependent modalities simultaneously, we hypothesize that there is an optimal transition to generate these modalities in inverse rendering. For example, according to previous works [18, 42, 46], geometry typically has the least uncertainty and supports the prediction of material and lighting, so generating geometry first could be advantageous. On the other hand, since lighting has the highest uncertainty, generating it last could be beneficial. To validate this hypothesis, we divide " + }, + { + "bbox": [ + 55, + 487, + 295, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_0" + }, + { + "bbox": [ + 55, + 487, + 295, + 666 + ], + "type": "text", + "content": " into three categories based on the modality attributes: geometry " + }, + { + "bbox": [ + 55, + 487, + 295, + 666 + ], + "type": "inline_equation", + "content": "(\\mathrm{N}_L,\\mathrm{D}_L)" + }, + { + "bbox": [ + 55, + 487, + 295, + 666 + ], + "type": "text", + "content": ", material " + }, + { + "bbox": [ + 55, + 487, + 295, + 666 + ], + "type": "inline_equation", + "content": "(\\mathrm{A}_L,\\mathrm{R}_L)" + }, + { + "bbox": [ + 55, + 487, + 295, + 666 + ], + "type": "text", + "content": ", and lighting " + }, + { + "bbox": [ + 55, + 487, + 295, + 666 + ], + "type": "inline_equation", + "content": "(f_{L})" + }, + { + "bbox": [ + 55, + 487, + 295, + 666 + ], + "type": "text", + "content": ", and accordingly apply different noise scheduling to the three parts. Specifically, following Chen et al. [2], we select the cosine scheduler as the base noise scheduler and parameterize it as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 67, + 676, + 295, + 717 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 676, + 295, + 717 + ], + "spans": [ + { + "bbox": [ + 67, + 676, + 295, + 717 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} v _ {s} = \\cos (s \\cdot 2 \\pi) ^ {2 \\tau}, v _ {b} = \\cos (b \\cdot 2 \\pi) ^ {2 \\tau} \\\\ \\alpha_ {t} = \\cos \\left(\\left(\\left(b - s\\right) \\cdot t + s\\right) \\cdot 2 \\pi\\right) ^ {2 \\tau}, \\bar {\\alpha} _ {t} = \\frac {v _ {b} - \\alpha_ {t}}{v _ {b} - v _ {s}}. \\tag {5} \\\\ \\end{array}", + "image_path": "08f1528b42110282d57e20afaac07bdf2724f62aef3727689ec0f2822c06e706.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 203, + 555, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 203, + 555, + 251 + ], + "spans": [ + { + "bbox": [ + 313, + 203, + 555, + 251 + ], + "type": "text", + "content": "Throughout the paper, we fix " + }, + { + "bbox": [ + 313, + 203, + 555, + 251 + ], + "type": "inline_equation", + "content": "s = 0.008" + }, + { + "bbox": [ + 313, + 203, + 555, + 251 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 203, + 555, + 251 + ], + "type": "inline_equation", + "content": "b = 1" + }, + { + "bbox": [ + 313, + 203, + 555, + 251 + ], + "type": "text", + "content": " and control the scheduler using only " + }, + { + "bbox": [ + 313, + 203, + 555, + 251 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 203, + 555, + 251 + ], + "type": "text", + "content": ". When " + }, + { + "bbox": [ + 313, + 203, + 555, + 251 + ], + "type": "inline_equation", + "content": "\\tau = 1" + }, + { + "bbox": [ + 313, + 203, + 555, + 251 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 203, + 555, + 251 + ], + "type": "inline_equation", + "content": "\\bar{\\alpha}_t" + }, + { + "bbox": [ + 313, + 203, + 555, + 251 + ], + "type": "text", + "content": " is equivalent to the cosine scheduler. The behavior of the noise scheduler as a function of " + }, + { + "bbox": [ + 313, + 203, + 555, + 251 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 203, + 555, + 251 + ], + "type": "text", + "content": " is shown in Fig. 3." + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 315, + 260, + 553, + 360 + ], + "blocks": [ + { + "bbox": [ + 315, + 260, + 553, + 360 + ], + "lines": [ + { + "bbox": [ + 315, + 260, + 553, + 360 + ], + "spans": [ + { + "bbox": [ + 315, + 260, + 553, + 360 + ], + "type": "image", + "image_path": "98d4a95e6e7b27792db66e13204d66f1c7569a2af432ef5974ed70517122b0c3.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 367, + 555, + 390 + ], + "lines": [ + { + "bbox": [ + 313, + 367, + 555, + 390 + ], + "spans": [ + { + "bbox": [ + 313, + 367, + 555, + 390 + ], + "type": "text", + "content": "Figure 3. Parameterizing noise scheduler. A larger " + }, + { + "bbox": [ + 313, + 367, + 555, + 390 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 367, + 555, + 390 + ], + "type": "text", + "content": " value results in a slower generation of the modality." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 403, + 555, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 403, + 555, + 557 + ], + "spans": [ + { + "bbox": [ + 313, + 403, + 555, + 557 + ], + "type": "text", + "content": "Fig. 3 shows that as the value of " + }, + { + "bbox": [ + 313, + 403, + 555, + 557 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 403, + 555, + 557 + ], + "type": "text", + "content": " increases, the noise increases more gradually. To determine the optimal modality generation order using " + }, + { + "bbox": [ + 313, + 403, + 555, + 557 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 403, + 555, + 557 + ], + "type": "text", + "content": ", we conducted experiments by applying different values of " + }, + { + "bbox": [ + 313, + 403, + 555, + 557 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 403, + 555, + 557 + ], + "type": "text", + "content": " to geometry, material, and lighting, meaning that different noise levels are applied separately to each channel of the data during the denoising process of the diffusion model. The ablation study in Tab. 7 confirms that generating in the order of geometry, material, and lighting is the most effective, aligning with our intuition. As a result, through our channel-wise noise scheduling, it focuses on specific modalities at different timesteps, enabling more effective predictions of other modalities in subsequent timesteps." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 558, + 556, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 556, + 700 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 556, + 700 + ], + "type": "text", + "content": "In general, " + }, + { + "bbox": [ + 313, + 558, + 556, + 700 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 558, + 556, + 700 + ], + "type": "text", + "content": " in diffusion models is set to a large number, such as 1000. This helps approximate the data distribution and increases the diversity of the generated data [24]. However, we note that such diffusion models are challenging in predicting a single definitive ground truth. To this end, we conducted an ablation study on " + }, + { + "bbox": [ + 313, + 558, + 556, + 700 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 558, + 556, + 700 + ], + "type": "text", + "content": " values by training the DM with " + }, + { + "bbox": [ + 313, + 558, + 556, + 700 + ], + "type": "inline_equation", + "content": "T = 1000, 100, 10" + }, + { + "bbox": [ + 313, + 558, + 556, + 700 + ], + "type": "text", + "content": ", and 1. For each setting, we performed sampling 20 times in the test set to measure the MSE and the variance (var) of the samples. Tab. 1 shows that reducing " + }, + { + "bbox": [ + 313, + 558, + 556, + 700 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 558, + 556, + 700 + ], + "type": "text", + "content": " improves accuracy at the cost of diversity in most cases. In the extreme case of " + }, + { + "bbox": [ + 313, + 558, + 556, + 700 + ], + "type": "inline_equation", + "content": "T = 1" + }, + { + "bbox": [ + 313, + 558, + 556, + 700 + ], + "type": "text", + "content": ", the method becomes deterministic." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 326, + 701, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 701, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 326, + 701, + 555, + 714 + ], + "type": "text", + "content": "To explore models focused on diversity and accuracy, we" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5776" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 191 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 191 + ], + "type": "text", + "content": "develop two types of diffusion models: one with " + }, + { + "bbox": [ + 55, + 72, + 296, + 191 + ], + "type": "inline_equation", + "content": "T = 1000" + }, + { + "bbox": [ + 55, + 72, + 296, + 191 + ], + "type": "text", + "content": " and the other with " + }, + { + "bbox": [ + 55, + 72, + 296, + 191 + ], + "type": "inline_equation", + "content": "T = 1" + }, + { + "bbox": [ + 55, + 72, + 296, + 191 + ], + "type": "text", + "content": ". Following the ablation study in Tab. 7, we select " + }, + { + "bbox": [ + 55, + 72, + 296, + 191 + ], + "type": "inline_equation", + "content": "\\tau = (0.9, 1.2, 1.5)" + }, + { + "bbox": [ + 55, + 72, + 296, + 191 + ], + "type": "text", + "content": " when " + }, + { + "bbox": [ + 55, + 72, + 296, + 191 + ], + "type": "inline_equation", + "content": "T = 1000" + }, + { + "bbox": [ + 55, + 72, + 296, + 191 + ], + "type": "text", + "content": " to prioritize diversity. Due to its gradual generation process, we name this model the Progressive Diffusion Model (PDM). However, this gradual generation process cannot be applied to the diffusion model with " + }, + { + "bbox": [ + 55, + 72, + 296, + 191 + ], + "type": "inline_equation", + "content": "T = 1" + }, + { + "bbox": [ + 55, + 72, + 296, + 191 + ], + "type": "text", + "content": ". Moreover, " + }, + { + "bbox": [ + 55, + 72, + 296, + 191 + ], + "type": "inline_equation", + "content": "T = 1" + }, + { + "bbox": [ + 55, + 72, + 296, + 191 + ], + "type": "text", + "content": " limits the ability of the diffusion model to exploit its full potential over multiple timesteps [3]. To this end, we define our new noise scheduler, as shown in Tab. 2." + } + ] + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 63, + 201, + 288, + 243 + ], + "blocks": [ + { + "bbox": [ + 63, + 201, + 288, + 243 + ], + "lines": [ + { + "bbox": [ + 63, + 201, + 288, + 243 + ], + "spans": [ + { + "bbox": [ + 63, + 201, + 288, + 243 + ], + "type": "table", + "html": "
tt mod 3 = 0t mod 3 = 1t mod 3 = 2T-1
\\( \\bar{a}_{t}^{G} \\)1100
\\( \\bar{a}_{t}^{M} \\)1010
\\( \\bar{a}_{t}^{L} \\)0110
", + "image_path": "05fe74ef7c818670f5cb3079446dbb7b26e89289abb0b2e725a1228c7aaa5738.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 289, + 296, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 289, + 296, + 410 + ], + "spans": [ + { + "bbox": [ + 55, + 289, + 296, + 410 + ], + "type": "text", + "content": "A value of 0 implies that pure Gaussian noise is added, while a value of 1 means that no noise is added. With our new noise scheduler, " + }, + { + "bbox": [ + 55, + 289, + 296, + 410 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 289, + 296, + 410 + ], + "type": "text", + "content": " is redefined as " + }, + { + "bbox": [ + 55, + 289, + 296, + 410 + ], + "type": "inline_equation", + "content": "3n + 1" + }, + { + "bbox": [ + 55, + 289, + 296, + 410 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 55, + 289, + 296, + 410 + ], + "type": "inline_equation", + "content": "n > 0" + }, + { + "bbox": [ + 55, + 289, + 296, + 410 + ], + "type": "text", + "content": ". After the initial prediction at " + }, + { + "bbox": [ + 55, + 289, + 296, + 410 + ], + "type": "inline_equation", + "content": "t = T - 1" + }, + { + "bbox": [ + 55, + 289, + 296, + 410 + ], + "type": "text", + "content": ", the noise is switched per channel, providing conditions for the next diffusion step. Taking into account this feature of the switchable noise scheduling, we name this model Switchable Diffusion Model (SDM). The SDM is trained to predict the modality with an SNR of 0 at each step, allowing for more refined predictions by utilizing the previous step's predictions." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 417, + 224, + 430 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 417, + 224, + 430 + ], + "spans": [ + { + "bbox": [ + 55, + 417, + 224, + 430 + ], + "type": "text", + "content": "3.4. RGB-Guided Super Resolution" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 434, + 296, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 434, + 296, + 507 + ], + "spans": [ + { + "bbox": [ + 55, + 434, + 296, + 507 + ], + "type": "text", + "content": "We use a UNet as the RGB-guided super-resolution model, consisting of a shared encoder and five separate decoders. The encoder processes the RGB image into a dense feature map, which, along with the rendered diffuse " + }, + { + "bbox": [ + 55, + 434, + 296, + 507 + ], + "type": "inline_equation", + "content": "\\mathrm{I}_d" + }, + { + "bbox": [ + 55, + 434, + 296, + 507 + ], + "type": "text", + "content": " and specular " + }, + { + "bbox": [ + 55, + 434, + 296, + 507 + ], + "type": "inline_equation", + "content": "\\mathrm{I}_s" + }, + { + "bbox": [ + 55, + 434, + 296, + 507 + ], + "type": "text", + "content": " images from " + }, + { + "bbox": [ + 55, + 434, + 296, + 507 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_0" + }, + { + "bbox": [ + 55, + 434, + 296, + 507 + ], + "type": "text", + "content": ", is fed into each decoder. Our RGB-guided Super-Resolution Model (SRM) is as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 102, + 514, + 295, + 526 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 514, + 295, + 526 + ], + "spans": [ + { + "bbox": [ + 102, + 514, + 295, + 526 + ], + "type": "interline_equation", + "content": "\\mathrm {N}, \\mathrm {D}, \\mathrm {A}, \\mathrm {R}, \\boldsymbol {f} = \\operatorname {S R M} \\left(\\mathbf {z} _ {0}, \\mathrm {I}, \\mathrm {I} _ {d}, \\mathrm {I} _ {s}\\right), \\tag {6}", + "image_path": "e079b0377904e4eee0b404d46cde9a6f3e589f32caa00d07040e4af2d4da4d5e.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 537, + 127, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 537, + 127, + 550 + ], + "spans": [ + { + "bbox": [ + 55, + 537, + 127, + 550 + ], + "type": "text", + "content": "4. Evaluation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 558, + 295, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 295, + 665 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 295, + 665 + ], + "type": "text", + "content": "In this section, we evaluate our method on both synthetic and real-world datasets. We first evaluate on a synthetic dataset, followed by inverse rendering and object insertion on a real-world spatially varying lighting dataset. Furthermore, we evaluated our method on the sub-tasks of inverse rendering, namely intrinsic decomposition and geometry prediction. More details on training, model architectures, and additional experimental results are provided in the Supplementary Materials." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "content": "Evaluation settings. We used " + }, + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "inline_equation", + "content": "C = 128" + }, + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "inline_equation", + "content": "H = 240" + }, + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "inline_equation", + "content": "W = 320" + }, + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "inline_equation", + "content": "H_{L} = 60" + }, + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "inline_equation", + "content": "W_{L} = 80" + }, + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "content": ". For SDM, we set " + }, + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "inline_equation", + "content": "T = 4" + }, + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "content": ". In all experiments, PDM generated 10 samples and the mean and variance of these samples were evaluated." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 554, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 156 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 156 + ], + "type": "text", + "content": "Baselines. We compared our method with state-of-the-art inverse rendering frameworks capable of per-pixel lighting inference, specifically those by Li et al. [19] and Zhu et al. [47]. Both methods were trained on the OpenRooms FF [4] dataset. However, since the code for Zhu's lighting model has not been released, we used their pretrained lighting model, which was trained on the InteriorVerse dataset." + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 316, + 167, + 553, + 281 + ], + "blocks": [ + { + "bbox": [ + 316, + 167, + 553, + 281 + ], + "lines": [ + { + "bbox": [ + 316, + 167, + 553, + 281 + ], + "spans": [ + { + "bbox": [ + 316, + 167, + 553, + 281 + ], + "type": "image", + "image_path": "c53b69649ece6843a6ea562e8e28547d1e8a0cc5e3b784f70ccc5ff248faa6f8.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 289, + 554, + 312 + ], + "lines": [ + { + "bbox": [ + 313, + 289, + 554, + 312 + ], + "spans": [ + { + "bbox": [ + 313, + 289, + 554, + 312 + ], + "type": "text", + "content": "Figure 4. Qualitative synthetic evaluation. Only our method successfully decomposes specular radiance into material and lighting." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 336, + 484, + 348 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 336, + 484, + 348 + ], + "spans": [ + { + "bbox": [ + 313, + 336, + 484, + 348 + ], + "type": "text", + "content": "4.1. Evaluation on Synthetic Scenes" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 354, + 555, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 354, + 555, + 556 + ], + "spans": [ + { + "bbox": [ + 313, + 354, + 555, + 556 + ], + "type": "text", + "content": "Tab. 3 includes quantitative prediction results on the synthetic [4] test dataset. Although our target is single-view, we include results from the multi-view baselines MAIR [4] and MAIR++ [5]. Our SDM, which prioritizes prediction accuracy, demonstrated superior performance compared to existing single-image baselines in all areas except roughness, and our PDM also demonstrated respectable performance comparable to the baselines. Notably, our lighting accuracy was comparable to that of MAIR, and the albedo accuracy outperformed even the multi-view methods. Additionally, it can be observed that both MAIR++ and our method, which are capable of ILR-based neural rendering, exhibit significantly superior re-rendering (I) performance. Figure 4 presents an illustrative example of this fact. Even in images with strong specular radiance, our SDM and PDM successfully decomposed and re-rendered the image in contrast to the failure of the baseline methods." + } + ] + } + ], + "index": 14 + }, + { + "type": "table", + "bbox": [ + 317, + 567, + 553, + 644 + ], + "blocks": [ + { + "bbox": [ + 55, + 252, + 295, + 275 + ], + "lines": [ + { + "bbox": [ + 55, + 252, + 295, + 275 + ], + "spans": [ + { + "bbox": [ + 55, + 252, + 295, + 275 + ], + "type": "text", + "content": "Table 2. SDM noise scheduler. Here, " + }, + { + "bbox": [ + 55, + 252, + 295, + 275 + ], + "type": "inline_equation", + "content": "\\bar{\\alpha}_t^G" + }, + { + "bbox": [ + 55, + 252, + 295, + 275 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 252, + 295, + 275 + ], + "type": "inline_equation", + "content": "\\bar{\\alpha}_t^M" + }, + { + "bbox": [ + 55, + 252, + 295, + 275 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 252, + 295, + 275 + ], + "type": "inline_equation", + "content": "\\bar{\\alpha}_t^L" + }, + { + "bbox": [ + 55, + 252, + 295, + 275 + ], + "type": "text", + "content": " are the schedulers for geometry, material, and lighting, respectively." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 567, + 553, + 644 + ], + "lines": [ + { + "bbox": [ + 317, + 567, + 553, + 644 + ], + "spans": [ + { + "bbox": [ + 317, + 567, + 553, + 644 + ], + "type": "table", + "html": "
MSE(×10-2)N ↓D ↓A ↓R ↓E ↓I ↓
Li et al. [19]3.7820.1350.8696.49615.490.491
Zhu et al. [47]1.7380.0620.5284.208*31.15*3.565
SDM1.5020.0390.3494.96913.350.156
PDM(Mean)1.8310.0470.3944.81914.080.150
PDM(Best)1.7820.0440.3814.64413.750.146
MAIR [4]1.0100.1400.6393.45713.790.784
MAIR++ [5]1.0100.0070.4782.49711.610.058
", + "image_path": "fd55a6fd3f6bacf37eea4c71f7c5abdb1ecd9aab50cc5cf3ce9eb949e31ca568.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_body" + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 653, + 554, + 696 + ], + "lines": [ + { + "bbox": [ + 313, + 653, + 554, + 696 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 554, + 696 + ], + "type": "text", + "content": "Table 3. Quantitative synthetic evaluation. Bold text indicates the best performance among single-view methods, while underlined text represents the second-best. * denotes models trained on the InteriorVerse dataset." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "5777" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 69, + 152, + 134 + ], + "blocks": [ + { + "bbox": [ + 56, + 69, + 152, + 134 + ], + "lines": [ + { + "bbox": [ + 56, + 69, + 152, + 134 + ], + "spans": [ + { + "bbox": [ + 56, + 69, + 152, + 134 + ], + "type": "image", + "image_path": "52170fac97f4b9667a11d36f51110a77e1637bc6bd17f40d9dc8936eb6617719.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 82, + 135, + 126, + 144 + ], + "lines": [ + { + "bbox": [ + 82, + 135, + 126, + 144 + ], + "spans": [ + { + "bbox": [ + 82, + 135, + 126, + 144 + ], + "type": "text", + "content": "Input Image" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 149, + 70, + 263, + 201 + ], + "blocks": [ + { + "bbox": [ + 149, + 70, + 263, + 201 + ], + "lines": [ + { + "bbox": [ + 149, + 70, + 263, + 201 + ], + "spans": [ + { + "bbox": [ + 149, + 70, + 263, + 201 + ], + "type": "image", + "image_path": "34e2a9088aa5e6a47585e3a46c69df49b0eda539371e05038eccf396e906c58a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 184, + 201, + 192, + 209 + ], + "lines": [ + { + "bbox": [ + 184, + 201, + 192, + 209 + ], + "spans": [ + { + "bbox": [ + 184, + 201, + 192, + 209 + ], + "type": "text", + "content": "N" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 263, + 70, + 359, + 201 + ], + "blocks": [ + { + "bbox": [ + 263, + 70, + 359, + 201 + ], + "lines": [ + { + "bbox": [ + 263, + 70, + 359, + 201 + ], + "spans": [ + { + "bbox": [ + 263, + 70, + 359, + 201 + ], + "type": "image", + "image_path": "b120857f00227edf41eb9d88dd7888366ca5d6cd79c9dcd1a9faa766cbcd520a.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 282, + 202, + 289, + 209 + ], + "lines": [ + { + "bbox": [ + 282, + 202, + 289, + 209 + ], + "spans": [ + { + "bbox": [ + 282, + 202, + 289, + 209 + ], + "type": "text", + "content": "A" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 359, + 70, + 456, + 201 + ], + "blocks": [ + { + "bbox": [ + 359, + 70, + 456, + 201 + ], + "lines": [ + { + "bbox": [ + 359, + 70, + 456, + 201 + ], + "spans": [ + { + "bbox": [ + 359, + 70, + 456, + 201 + ], + "type": "image", + "image_path": "c94dcc6c1a73d9888ed29b4e3044b29f32168c0f8995a7ea927caaeebc566e42.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 377, + 201, + 386, + 211 + ], + "lines": [ + { + "bbox": [ + 377, + 201, + 386, + 211 + ], + "spans": [ + { + "bbox": [ + 377, + 201, + 386, + 211 + ], + "type": "text", + "content": "Id" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 456, + 70, + 554, + 201 + ], + "blocks": [ + { + "bbox": [ + 456, + 70, + 554, + 201 + ], + "lines": [ + { + "bbox": [ + 456, + 70, + 554, + 201 + ], + "spans": [ + { + "bbox": [ + 456, + 70, + 554, + 201 + ], + "type": "image", + "image_path": "e743fb28423b44050dcfd35343fd145f0e4c2b5ce49893f412ff7a17cb791b83.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 476, + 201, + 481, + 209 + ], + "lines": [ + { + "bbox": [ + 476, + 201, + 481, + 209 + ], + "spans": [ + { + "bbox": [ + 476, + 201, + 481, + 209 + ], + "type": "text", + "content": "I" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 56, + 254, + 181, + 417 + ], + "blocks": [ + { + "bbox": [ + 55, + 220, + 555, + 243 + ], + "lines": [ + { + "bbox": [ + 55, + 220, + 555, + 243 + ], + "spans": [ + { + "bbox": [ + 55, + 220, + 555, + 243 + ], + "type": "text", + "content": "Figure 5. Qualitative real-world evaluation. Our method successfully recognizes shadows under the desk, achieving plausible geometry prediction and realistic re-rendering results." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 56, + 254, + 181, + 417 + ], + "lines": [ + { + "bbox": [ + 56, + 254, + 181, + 417 + ], + "spans": [ + { + "bbox": [ + 56, + 254, + 181, + 417 + ], + "type": "image", + "image_path": "f9c98b41c00b6411f5738a01c3cb455eee3017fbe5df5595772a64d789f2416b.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 107, + 418, + 129, + 423 + ], + "lines": [ + { + "bbox": [ + 107, + 418, + 129, + 423 + ], + "spans": [ + { + "bbox": [ + 107, + 418, + 129, + 423 + ], + "type": "text", + "content": "Li et al." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 181, + 254, + 305, + 417 + ], + "blocks": [ + { + "bbox": [ + 181, + 254, + 305, + 417 + ], + "lines": [ + { + "bbox": [ + 181, + 254, + 305, + 417 + ], + "spans": [ + { + "bbox": [ + 181, + 254, + 305, + 417 + ], + "type": "image", + "image_path": "1bfaa68d1c4334484df848afbddcce15da9448ec4a0055e34dfea64971f4e484.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 229, + 418, + 257, + 424 + ], + "lines": [ + { + "bbox": [ + 229, + 418, + 257, + 424 + ], + "spans": [ + { + "bbox": [ + 229, + 418, + 257, + 424 + ], + "type": "text", + "content": "Zhu et al." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 306, + 254, + 430, + 418 + ], + "blocks": [ + { + "bbox": [ + 306, + 254, + 430, + 418 + ], + "lines": [ + { + "bbox": [ + 306, + 254, + 430, + 418 + ], + "spans": [ + { + "bbox": [ + 306, + 254, + 430, + 418 + ], + "type": "image", + "image_path": "59d547382ef8f11af670ea66072873a7c00d6fbe9a6e2a24d52a71cde35bc479.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 352, + 418, + 383, + 424 + ], + "lines": [ + { + "bbox": [ + 352, + 418, + 383, + 424 + ], + "spans": [ + { + "bbox": [ + 352, + 418, + 383, + 424 + ], + "type": "text", + "content": "Ours(SDM)" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 436, + 555, + 460 + ], + "lines": [ + { + "bbox": [ + 55, + 436, + 555, + 460 + ], + "spans": [ + { + "bbox": [ + 55, + 436, + 555, + 460 + ], + "type": "text", + "content": "Figure 6. Qualitative object insertion evaluation. We accurately predicted the direction of light and shadows even in real-world images. Additionally, our rendered objects were closer to the ground truth in terms of hue and brightness compared to the baselines." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 431, + 254, + 554, + 417 + ], + "blocks": [ + { + "bbox": [ + 431, + 254, + 554, + 417 + ], + "lines": [ + { + "bbox": [ + 431, + 254, + 554, + 417 + ], + "spans": [ + { + "bbox": [ + 431, + 254, + 554, + 417 + ], + "type": "image", + "image_path": "e225683f39f3c94a1947dcefa2d163877e8f7d76c1d811baf96694a8d972d6d0.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 481, + 418, + 503, + 423 + ], + "lines": [ + { + "bbox": [ + 481, + 418, + 503, + 423 + ], + "spans": [ + { + "bbox": [ + 481, + 418, + 503, + 423 + ], + "type": "text", + "content": "RGB " + }, + { + "bbox": [ + 481, + 418, + 503, + 423 + ], + "type": "inline_equation", + "content": "\\leftrightarrow" + }, + { + "bbox": [ + 481, + 418, + 503, + 423 + ], + "type": "text", + "content": " X" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "bbox": [ + 55, + 479, + 233, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 479, + 233, + 491 + ], + "spans": [ + { + "bbox": [ + 55, + 479, + 233, + 491 + ], + "type": "text", + "content": "4.2. Evaluation on Real-world Scenes" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 55, + 498, + 296, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 296, + 593 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 296, + 593 + ], + "type": "text", + "content": "Qualitative evaluation. We evaluated the generalization capability of our method on real-world images from Garon et al. [9]. As shown in Fig. 5, both SDM and PDM successfully separated the albedo even in the presence of strong shadows under the desk. Li et al. [19] partially removed the shadows but failed to predict the geometry, while Zhu et al. [47] accurately predicted the normal but failed to estimate the albedo." + } + ] + } + ], + "index": 21 + }, + { + "type": "table", + "bbox": [ + 70, + 604, + 282, + 628 + ], + "blocks": [ + { + "bbox": [ + 70, + 604, + 282, + 628 + ], + "lines": [ + { + "bbox": [ + 70, + 604, + 282, + 628 + ], + "spans": [ + { + "bbox": [ + 70, + 604, + 282, + 628 + ], + "type": "table", + "html": "
Li et al. [19]Zhu et al. [47]SDMGround Truth
0.1780.1530.3220.348
", + "image_path": "ddf81cd8ff4827dccbb0732c42d8ca2819af04103830f7b88510bfa051ad2ec3.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "table_body" + } + ], + "index": 22 + }, + { + "bbox": [ + 55, + 637, + 295, + 660 + ], + "lines": [ + { + "bbox": [ + 55, + 637, + 295, + 660 + ], + "spans": [ + { + "bbox": [ + 55, + 637, + 295, + 660 + ], + "type": "text", + "content": "Table 4. AMT user study on object insertion. Our method was selected by nearly as many users as the ground truth." + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "Lighting Evaluation. We also evaluated lighting performance through object insertion, using the object insertion tool from Li et al. [19]. For fairness, we excluded PDM" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 479, + 555, + 708 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 479, + 555, + 708 + ], + "spans": [ + { + "bbox": [ + 313, + 479, + 555, + 708 + ], + "type": "text", + "content": "and focused on evaluating only SDM. As shown in Fig. 6, our method accurately predicts strong directional lighting, as well as the surrounding light's tone and shadows, resulting in the most realistic object insertions. In contrast, Zhu et al. [47] failed to predict the correct light direction, while Li et al. [19] struggles with realistic rendering due to a lack of high-frequency details in the environment map. The diffusion-based method RGB " + }, + { + "bbox": [ + 313, + 479, + 555, + 708 + ], + "type": "inline_equation", + "content": "\\leftrightarrow" + }, + { + "bbox": [ + 313, + 479, + 555, + 708 + ], + "type": "text", + "content": " X [43] relies on inpainting for object insertion, which compromises the fidelity of the inserted objects and leads to visible artifacts. Additionally, since it cannot handle environment maps, it struggled to accurately predict the direction of light. For a more quantitative evaluation, we conducted an AMT user study on object insertion, including the ground truth. The results in Tab. 4 confirmed the superiority of our method. Users were asked: \"Of the four images, please select the one that shows white rabbits blending naturally into the surrounding light.\" Our method was selected by users nearly as frequently as the ground truth." + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "5778" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 69, + 309, + 189 + ], + "blocks": [ + { + "bbox": [ + 58, + 69, + 309, + 189 + ], + "lines": [ + { + "bbox": [ + 58, + 69, + 309, + 189 + ], + "spans": [ + { + "bbox": [ + 58, + 69, + 309, + 189 + ], + "type": "image", + "image_path": "1bab5064a7f194e7f390e7ae32edf7cb716177c71e87514ee173925c3ea94c30.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 198, + 555, + 232 + ], + "lines": [ + { + "bbox": [ + 55, + 198, + 555, + 232 + ], + "spans": [ + { + "bbox": [ + 55, + 198, + 555, + 232 + ], + "type": "text", + "content": "Figure 7. Intrinsic decomposition. Our method successfully decomposes the input image into albedo and shading. RGB " + }, + { + "bbox": [ + 55, + 198, + 555, + 232 + ], + "type": "inline_equation", + "content": "\\leftrightarrow" + }, + { + "bbox": [ + 55, + 198, + 555, + 232 + ], + "type": "text", + "content": " X [43], leveraging its strong prior, achieves successful albedo prediction but struggles with realistic rendering, resulting in excessive brightness in the left sample and incorrect specular highlights in the right sample." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 310, + 69, + 554, + 190 + ], + "blocks": [ + { + "bbox": [ + 310, + 69, + 554, + 190 + ], + "lines": [ + { + "bbox": [ + 310, + 69, + 554, + 190 + ], + "spans": [ + { + "bbox": [ + 310, + 69, + 554, + 190 + ], + "type": "image", + "image_path": "c0dee3c1dfe2482606782c0c655f53590dd6bb0432acacbfc2fb17c4a525cb33.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 58, + 249, + 294, + 392 + ], + "blocks": [ + { + "bbox": [ + 58, + 249, + 294, + 392 + ], + "lines": [ + { + "bbox": [ + 58, + 249, + 294, + 392 + ], + "spans": [ + { + "bbox": [ + 58, + 249, + 294, + 392 + ], + "type": "image", + "image_path": "7ad47f7574d1da95d022fcf6feb415f3b3d70a58102718caebca2fecb497fd8f.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 398, + 295, + 422 + ], + "lines": [ + { + "bbox": [ + 55, + 398, + 295, + 422 + ], + "spans": [ + { + "bbox": [ + 55, + 398, + 295, + 422 + ], + "type": "text", + "content": "Figure 8. Geometry prediction. Our method achieves the most accurate prediction of the ceiling structure." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 58, + 434, + 293, + 503 + ], + "blocks": [ + { + "bbox": [ + 58, + 434, + 293, + 503 + ], + "lines": [ + { + "bbox": [ + 58, + 434, + 293, + 503 + ], + "spans": [ + { + "bbox": [ + 58, + 434, + 293, + 503 + ], + "type": "table", + "html": "
MethodA ↓Hue ↓Texture ↓WHDR ↓S ↓I ↓
Li et al. [19]0.9245.7810.4730.3011.1620.021
Zhu et al. [47]1.6854.6380.4720.3011.4410.028
SDM0.7055.0560.4590.2931.1810.006
PDM(Mean)0.7575.2600.4590.3061.1560.005
PDM(Best)0.6234.5270.4560.2781.1130.005
RGB↔X [43](Mean)0.6023.6900.4480.1681.2610.055
RGB↔X [43](Best)0.3313.0910.4410.1461.1960.052
", + "image_path": "a00aaf44cf2954677bea0c4ece36e461ab089f9e7373df5fa3f386730a531335.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 551, + 192, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 551, + 192, + 562 + ], + "spans": [ + { + "bbox": [ + 55, + 551, + 192, + 562 + ], + "type": "text", + "content": "4.3. Evaluation on Sub-tasks" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "content": "Intrinsic decomposition. We evaluated our method on intrinsic decomposition using the recently introduced MAW [40] dataset, which provides more refined annotations for 888 images, allowing more quantitative comparisons than the WHDR metric used in the IIW [1] dataset. Tab. 5 presents comparisons for MSE for A, S, and I, along with metrics for hue, texture, and WHDR for A. The hue metric evaluates the chromaticity component, while the texture metric is based on LPIPS perceptual similarity. Our method generally outperformed existing deterministic methods [19, 47]. RGB " + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\leftrightarrow" + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "content": " X [43], which learns strong priors from diverse large datasets, excels in albedo infer" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 251, + 555, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 251, + 555, + 407 + ], + "spans": [ + { + "bbox": [ + 313, + 251, + 555, + 407 + ], + "type": "text", + "content": "ence; however, due to its reliance on diffusion-based rendering, it reconstructed the input images with the lowest accuracy. In contrast, our method, trained only on OpenRooms FF, achieved the highest performance in shading estimation and most faithfully reproduced input images, demonstrating high fidelity. This can also be observed in Fig. 7. In the left sample, RGB " + }, + { + "bbox": [ + 313, + 251, + 555, + 407 + ], + "type": "inline_equation", + "content": "\\leftrightarrow" + }, + { + "bbox": [ + 313, + 251, + 555, + 407 + ], + "type": "text", + "content": " X [43] fails to predict the shading intensity as it does not account for HDR lighting. Similarly, in the right sample, RGB " + }, + { + "bbox": [ + 313, + 251, + 555, + 407 + ], + "type": "inline_equation", + "content": "\\leftrightarrow" + }, + { + "bbox": [ + 313, + 251, + 555, + 407 + ], + "type": "text", + "content": " X [43] introduces specular highlights on the desk that are absent in the input image, resulting in a loss of fidelity. In contrast, our method predicts spatially consistent albedo and accurately estimates brightness, enabling a more realistic reproduction of the input image." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 408, + 556, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 408, + 556, + 492 + ], + "spans": [ + { + "bbox": [ + 313, + 408, + 556, + 492 + ], + "type": "text", + "content": "Geometry prediction. Tab. 6 includes the geometry prediction evaluation on the DIODE [37] indoor dataset. Similarly to the intrinsic decomposition task, both our PDM and SDM demonstrate superior or comparable performance to the baselines. Fig. 8 provides qualitative examples, demonstrating that our method exhibits strong generalization capability in geometry prediction." + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 317, + 502, + 552, + 540 + ], + "blocks": [ + { + "bbox": [ + 55, + 506, + 295, + 529 + ], + "lines": [ + { + "bbox": [ + 55, + 506, + 295, + 529 + ], + "spans": [ + { + "bbox": [ + 55, + 506, + 295, + 529 + ], + "type": "text", + "content": "Table 5. Intrinsic decomposition evaluation on MAW [40]. Bold text indicates the best performance." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 502, + 552, + 540 + ], + "lines": [ + { + "bbox": [ + 317, + 502, + 552, + 540 + ], + "spans": [ + { + "bbox": [ + 317, + 502, + 552, + 540 + ], + "type": "table", + "html": "
MetricLi et al. [19]Zhu et al. [47]SDMPDM(Mean)PDM(Best)
N ↓0.1960.1710.1470.1520.144
D ↓0.0590.0570.0570.0550.050
I ↓0.0150.0130.0030.0020.002
", + "image_path": "7f390b0700e25580a6ff93f41cbb9a3ccb36b7f78da7bd3bce13a9a12253a47f.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 542, + 555, + 564 + ], + "lines": [ + { + "bbox": [ + 313, + 542, + 555, + 564 + ], + "spans": [ + { + "bbox": [ + 313, + 542, + 555, + 564 + ], + "type": "text", + "content": "Table 6. Geometry prediction evaluation on DIODE [37]. Bold text indicates the best performance." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 587, + 446, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 587, + 446, + 601 + ], + "spans": [ + { + "bbox": [ + 313, + 587, + 446, + 601 + ], + "type": "text", + "content": "4.4. Modality Dependencies" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 605, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 556, + 713 + ], + "type": "text", + "content": "To generate a reasonable solution for inverse rendering, it is crucial to account for the dependencies between modalities. For instance, if a material is predicted to be darker, the lighting should be predicted brighter to compensate for the input image. However, existing diffusion-based methods [14, 22, 43] infer each modality independently, making it difficult to capture such dependencies. In contrast, our PDM generates all modalities simultaneously, enabling it to account for these relationships. Fig. 9 illustrates this with" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "5779" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 69, + 295, + 189 + ], + "blocks": [ + { + "bbox": [ + 57, + 69, + 295, + 189 + ], + "lines": [ + { + "bbox": [ + 57, + 69, + 295, + 189 + ], + "spans": [ + { + "bbox": [ + 57, + 69, + 295, + 189 + ], + "type": "image", + "image_path": "91cf3562a8c816a8758fe3db0188570cfeea05733ccf70ee790f2a8079f7db5e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 195, + 295, + 217 + ], + "lines": [ + { + "bbox": [ + 55, + 195, + 295, + 217 + ], + "spans": [ + { + "bbox": [ + 55, + 195, + 295, + 217 + ], + "type": "text", + "content": "Figure 9. Modality dependencies. Two samples of N, D, A, S, and R predicted by PDM." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 239, + 296, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 239, + 296, + 348 + ], + "spans": [ + { + "bbox": [ + 54, + 239, + 296, + 348 + ], + "type": "text", + "content": "two different samples generated by PDM. For the center table in the top image, the first sample predicts the normal pointing upward with bright shading, resulting in a darker albedo. Conversely, the second sample predicts the normal facing the camera, producing darker shading and a brighter albedo. Additionally, for the area under the cabinet in the bottom image, the first sample interprets it as a shadow and predicts a bright albedo, while the second sample treats it as part of the albedo and generates bright shading." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 355, + 124, + 366 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 355, + 124, + 366 + ], + "spans": [ + { + "bbox": [ + 55, + 355, + 124, + 366 + ], + "type": "text", + "content": "4.5. Ablations" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 373, + 296, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 373, + 296, + 553 + ], + "spans": [ + { + "bbox": [ + 55, + 373, + 296, + 553 + ], + "type": "text", + "content": "Modality generation order. To explore the optimal modality generation order, we conducted an ablation study on the modality generation sequence. There are no restrictions on " + }, + { + "bbox": [ + 55, + 373, + 296, + 553 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 55, + 373, + 296, + 553 + ], + "type": "text", + "content": "; however, we selected three values: " + }, + { + "bbox": [ + 55, + 373, + 296, + 553 + ], + "type": "inline_equation", + "content": "\\tau = 0.9, 1.2," + }, + { + "bbox": [ + 55, + 373, + 296, + 553 + ], + "type": "text", + "content": " and 1.5. These values were applied separately to " + }, + { + "bbox": [ + 55, + 373, + 296, + 553 + ], + "type": "inline_equation", + "content": "\\tau_{G}" + }, + { + "bbox": [ + 55, + 373, + 296, + 553 + ], + "type": "text", + "content": " (geometry), " + }, + { + "bbox": [ + 55, + 373, + 296, + 553 + ], + "type": "inline_equation", + "content": "\\tau_{M}" + }, + { + "bbox": [ + 55, + 373, + 296, + 553 + ], + "type": "text", + "content": " (material), and " + }, + { + "bbox": [ + 55, + 373, + 296, + 553 + ], + "type": "inline_equation", + "content": "\\tau_{L}" + }, + { + "bbox": [ + 55, + 373, + 296, + 553 + ], + "type": "text", + "content": " (lighting), and the models were trained independently. The performance was then evaluated on the test dataset. Tab. 7 shows that this channel-wise noise scheduling can significantly impact the inverse rendering quality. In particular, " + }, + { + "bbox": [ + 55, + 373, + 296, + 553 + ], + "type": "inline_equation", + "content": "\\tau_{G}, \\tau_{M}, \\tau_{L} = (0.9, 1.2, 1.5)" + }, + { + "bbox": [ + 55, + 373, + 296, + 553 + ], + "type": "text", + "content": " produces the best quality, aligning with our intuition. In contrast, generating in reverse order results in a detrimental effect. Fig. 10 provides examples of these two cases. The reverse order (L-M-G) exhibits artifacts in areas with strong shadows, whereas our method produces plausible results." + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 58, + 562, + 293, + 645 + ], + "blocks": [ + { + "bbox": [ + 58, + 562, + 293, + 645 + ], + "lines": [ + { + "bbox": [ + 58, + 562, + 293, + 645 + ], + "spans": [ + { + "bbox": [ + 58, + 562, + 293, + 645 + ], + "type": "table", + "html": "
τGτMτLNL↓DL↓AL↓RL↓EL↓IL↓
1.01.01.02.1400.0620.4035.75412.190.163
0.91.21.52.0240.0590.3765.17611.940.159
0.91.51.22.1040.0600.3905.72312.150.165
1.20.91.52.0730.0600.3885.39412.200.161
1.21.50.92.2020.0640.4115.88212.330.167
1.50.91.22.2110.0640.4055.66412.360.160
1.51.20.92.2450.0660.4145.98612.530.166
", + "image_path": "fd15152b5dc1a362d2549bb22ce1d2541b45b2c1917e19ce5383ae2ec6f68fd2.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": "Iterative Prediction of SDM. We conducted an ablation study to evaluate the benefits of iterative prediction for" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 316, + 70, + 554, + 168 + ], + "blocks": [ + { + "bbox": [ + 316, + 70, + 554, + 168 + ], + "lines": [ + { + "bbox": [ + 316, + 70, + 554, + 168 + ], + "spans": [ + { + "bbox": [ + 316, + 70, + 554, + 168 + ], + "type": "image", + "image_path": "86f3fd7445c5b2da785fd0a4f9e4f9bd156c25fb5bd13c4a3828960cbb251844.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 177, + 554, + 210 + ], + "lines": [ + { + "bbox": [ + 313, + 177, + 554, + 210 + ], + "spans": [ + { + "bbox": [ + 313, + 177, + 554, + 210 + ], + "type": "text", + "content": "Figure 10. Comparison of modality generation order. Generating in the order of geometry, material, and lighting helps effectively infer shadows." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 329, + 221, + 539, + 268 + ], + "blocks": [ + { + "bbox": [ + 55, + 653, + 295, + 676 + ], + "lines": [ + { + "bbox": [ + 55, + 653, + 295, + 676 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 295, + 676 + ], + "type": "text", + "content": "Table 7. Ablation of the generation order. The best values are highlighted in bold, while the worst values are underlined." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 329, + 221, + 539, + 268 + ], + "lines": [ + { + "bbox": [ + 329, + 221, + 539, + 268 + ], + "spans": [ + { + "bbox": [ + 329, + 221, + 539, + 268 + ], + "type": "table", + "html": "
TN ↓D ↓A ↓R ↓E ↓I ↓
11.7330.0600.3185.83711.240.132
41.5780.0520.2885.74310.960.137
71.6300.0500.2925.75910.970.139
", + "image_path": "0c2795bc08dbb2f6b4587649dba7ad6c388e7c8b9e5811efe9f6472f284a9a3b.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 276, + 554, + 299 + ], + "lines": [ + { + "bbox": [ + 313, + 276, + 554, + 299 + ], + "spans": [ + { + "bbox": [ + 313, + 276, + 554, + 299 + ], + "type": "text", + "content": "Table 8. Ablation of SDM iterative prediction. Bold text indicates the best performance." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 320, + 555, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 320, + 555, + 392 + ], + "spans": [ + { + "bbox": [ + 313, + 320, + 555, + 392 + ], + "type": "text", + "content": "SDM. In Tab. 8, the improvements at " + }, + { + "bbox": [ + 313, + 320, + 555, + 392 + ], + "type": "inline_equation", + "content": "T = 4" + }, + { + "bbox": [ + 313, + 320, + 555, + 392 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 320, + 555, + 392 + ], + "type": "inline_equation", + "content": "T = 7" + }, + { + "bbox": [ + 313, + 320, + 555, + 392 + ], + "type": "text", + "content": " over " + }, + { + "bbox": [ + 313, + 320, + 555, + 392 + ], + "type": "inline_equation", + "content": "T = 1" + }, + { + "bbox": [ + 313, + 320, + 555, + 392 + ], + "type": "text", + "content": " indicate that iterative prediction utilizes previously predicted modalities to inform subsequent predictions. However, the performance drop at " + }, + { + "bbox": [ + 313, + 320, + 555, + 392 + ], + "type": "inline_equation", + "content": "T = 7" + }, + { + "bbox": [ + 313, + 320, + 555, + 392 + ], + "type": "text", + "content": " compared to " + }, + { + "bbox": [ + 313, + 320, + 555, + 392 + ], + "type": "inline_equation", + "content": "T = 4" + }, + { + "bbox": [ + 313, + 320, + 555, + 392 + ], + "type": "text", + "content": " suggests that excessive iterations do not always yield better results." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 405, + 388, + 417 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 405, + 388, + 417 + ], + "spans": [ + { + "bbox": [ + 314, + 405, + 388, + 417 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 426, + 554, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 554, + 496 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 554, + 496 + ], + "type": "text", + "content": "We tackled inverse rendering with two opposing goals: attaining diversity and ensuring accuracy. To address these objectives, we proposed a diffusion-based framework, in which channel-wise noise scheduling allowed each diffusion model to better understand the correlations between different modalities." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 498, + 554, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 554, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 554, + 640 + ], + "type": "text", + "content": "Limitations and future work. Our approach has the limitation of requiring two separate models; thus, integrating them into a unified model would enhance usability. We explored re-rendering error as an alternative metric to overcome the constraint of depending only on the predicted variance of PDM for estimating inverse rendering uncertainty, but it could be imprecise. Further research on the handling of uncertainty could improve the complementarity of SDM and PDM. Moreover, our method refrains from exploiting the priors of pretrained LDMs. Using those priors to process per-pixel environment maps might enhance the performance and expand the applications of LDMs." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 642, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 554, + 713 + ], + "type": "text", + "content": "Acknowledgements. This work was supported by Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government(MSIT)(No.RS-2023-00227592, Development of 3D Object Identification Technology Robust to Viewpoint Changes)" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "5780" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 123 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 123 + ], + "type": "text", + "content": "[1] Sean Bell, Kavita Bala, and Noah Snively. Intrinsic images in the wild. ACM Transactions on Graphics (TOG), 33(4): 1-12, 2014. 2, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 125, + 296, + 147 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 125, + 296, + 147 + ], + "spans": [ + { + "bbox": [ + 61, + 125, + 296, + 147 + ], + "type": "text", + "content": "[2] Ting Chen. On the importance of noise scheduling for diffusion models. arXiv preprint arXiv:2301.10972, 2023. 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 149, + 296, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 149, + 296, + 203 + ], + "spans": [ + { + "bbox": [ + 62, + 149, + 296, + 203 + ], + "type": "text", + "content": "[3] Jooyoung Choi, Jungbeom Lee, Chaehun Shin, Sungwon Kim, Hyunwoo Kim, and Sungroh Yoon. Perception prioritized training of diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11472-11481, 2022. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 205, + 295, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 205, + 295, + 270 + ], + "spans": [ + { + "bbox": [ + 62, + 205, + 295, + 270 + ], + "type": "text", + "content": "[4] JunYong Choi, SeokYeong Lee, Haesol Park, Seung-Won Jung, Ig-Jae Kim, and Junghyun Cho. Mair: multi-view attention inverse rendering with 3d spatially-varying lighting estimation. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8392-8401. IEEE, 2023. 2, 3, 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 272, + 295, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 272, + 295, + 327 + ], + "spans": [ + { + "bbox": [ + 62, + 272, + 295, + 327 + ], + "type": "text", + "content": "[5] JunYong Choi, SeokYeong Lee, Haesol Park, Seung-Won Jung, Ig-Jae Kim, and Junghyun Cho. Mair++: Improving multi-view attention inverse rendering with implicit lighting representation. arXiv preprint arXiv:2408.06707, 2024. 2, 3, 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 328, + 295, + 372 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 328, + 295, + 372 + ], + "spans": [ + { + "bbox": [ + 62, + 328, + 295, + 372 + ], + "type": "text", + "content": "[6] Valentin Deschaintre, Miika Aittala, Fredo Durand, George Drettakis, and Adrien Bousseau. Single-image svbrdf capture with a rendering-aware deep network. ACM Transactions on Graphics (ToG), 37(4):1-15, 2018. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 373, + 295, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 373, + 295, + 407 + ], + "spans": [ + { + "bbox": [ + 62, + 373, + 295, + 407 + ], + "type": "text", + "content": "[7] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 408, + 295, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 408, + 295, + 463 + ], + "spans": [ + { + "bbox": [ + 62, + 408, + 295, + 463 + ], + "type": "text", + "content": "[8] Marc-Andre Gardner, Yannick Hold-Geoffroy, Kalyan Sunkavalli, Christian Gagne, and Jean-François Lalonde. Deep parametric indoor lighting estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7175–7183, 2019. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 464, + 295, + 519 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 464, + 295, + 519 + ], + "spans": [ + { + "bbox": [ + 62, + 464, + 295, + 519 + ], + "type": "text", + "content": "[9] Mathieu Garon, Kalyan Sunkavalli, Sunil Hadap, Nathan Carr, and Jean-François Lalonde. Fast spatially-varying indoor lighting estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6908-6917, 2019. 2, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 521, + 295, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 521, + 295, + 575 + ], + "spans": [ + { + "bbox": [ + 57, + 521, + 295, + 575 + ], + "type": "text", + "content": "[10] Roger Grosse, Micah K Johnson, Edward H Adelson, and William T Freeman. Ground truth dataset and baseline evaluations for intrinsic image algorithms. In 2009 IEEE 12th International Conference on Computer Vision, pages 2335-2342. IEEE, 2009. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 577, + 295, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 577, + 295, + 610 + ], + "spans": [ + { + "bbox": [ + 57, + 577, + 295, + 610 + ], + "type": "text", + "content": "[11] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 612, + 296, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 612, + 296, + 655 + ], + "spans": [ + { + "bbox": [ + 57, + 612, + 296, + 655 + ], + "type": "text", + "content": "[12] Jonathan Ho, Chitwan Sahara, William Chan, David J Fleet, Mohammad Norouzi, and Tim Salimans. Cascaded diffusion models for high fidelity image generation. Journal of Machine Learning Research, 23(47):1-33, 2022. 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 657, + 295, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 657, + 295, + 679 + ], + "spans": [ + { + "bbox": [ + 57, + 657, + 295, + 679 + ], + "type": "text", + "content": "[13] Berthold KP Horn and Michael J Brooks. Shape from shading. MIT press, 1989. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 681, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 681, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 681, + 295, + 713 + ], + "type": "text", + "content": "[14] Peter Kocsis, Vincent Sitzmann, and Matthias Nießner. Intrinsic image diffusion for single-view material estimation. arXiv preprint arXiv:2312.12274, 2023. 2, 4, 7" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "text", + "content": "[15] Soochan Lee, Junsoo Ha, and Gunhee Kim. Harmonizing maximum likelihood with gans for multimodal conditional generation. arXiv preprint arXiv:1902.09225, 2019. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 108, + 553, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 108, + 553, + 152 + ], + "spans": [ + { + "bbox": [ + 317, + 108, + 553, + 152 + ], + "type": "text", + "content": "[16] Zhengqi Li and Noah Snavely. Cginintrinsics: Better intrinsic image decomposition through physically-based rendering. In Proceedings of the European conference on computer vision (ECCV), pages 371-387, 2018. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 153, + 553, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 153, + 553, + 206 + ], + "spans": [ + { + "bbox": [ + 317, + 153, + 553, + 206 + ], + "type": "text", + "content": "[17] Zhengqin Li, Kalyan Sunkavalli, and Manmohan Chandraker. Materials for masses: Svbrdf acquisition with a single mobile phone image. In Proceedings of the European conference on computer vision (ECCV), pages 72-87, 2018. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 209, + 553, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 209, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 316, + 209, + 553, + 262 + ], + "type": "text", + "content": "[18] Zhengqin Li, Zexiang Xu, Ravi Ramamoorthi, Kalyan Sunkavalli, and Manmohan Chandraker. Learning to reconstruct shape and spatially-varying reflectance from a single image. ACM Transactions on Graphics (TOG), 37(6):1-11, 2018. 2, 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 264, + 555, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 264, + 555, + 331 + ], + "spans": [ + { + "bbox": [ + 316, + 264, + 555, + 331 + ], + "type": "text", + "content": "[19] Zhengqin Li, Mohammad Shafiei, Ravi Ramamoorthi, Kalyan Sunkavalli, and Manmohan Chandraker. Inverse rendering for complex indoor scenes: Shape, spatially-varying lighting and svbrdf from a single image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2475-2484, 2020. 2, 5, 6, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 332, + 555, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 332, + 555, + 398 + ], + "spans": [ + { + "bbox": [ + 316, + 332, + 555, + 398 + ], + "type": "text", + "content": "[20] Zhengqin Li, Ting-Wei Yu, Shen Sang, Sarah Wang, Meng Song, Yuhan Liu, Yu-Ying Yeh, Rui Zhu, Nitesh Gundavarapu, Jia Shi, et al. Openrooms: An open framework for photorealistic indoor scene datasets. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7190–7199, 2021. 2, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 399, + 553, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 399, + 553, + 454 + ], + "spans": [ + { + "bbox": [ + 316, + 399, + 553, + 454 + ], + "type": "text", + "content": "[21] Jundan Luo, Zhaoyang Huang, Yijin Li, Xiaowei Zhou, Guofeng Zhang, and Hujun Bao. Niid-net: Adapting surface normal knowledge for intrinsic image decomposition in indoor scenes. IEEE Transactions on Visualization and Computer Graphics, 26(12):3434-3445, 2020. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 456, + 553, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 456, + 553, + 520 + ], + "spans": [ + { + "bbox": [ + 317, + 456, + 553, + 520 + ], + "type": "text", + "content": "[22] Jundan Luo, Duygu Ceylan, Jae Shin Yoon, Nanxuan Zhao, Julien Philip, Anna Frühstück, Wenbin Li, Christian Richardt, and Tuanfeng Wang. Intrinsicdiffusion: Joint intrinsic layers from latent diffusion models. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 2, 4, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 522, + 553, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 522, + 553, + 577 + ], + "spans": [ + { + "bbox": [ + 317, + 522, + 553, + 577 + ], + "type": "text", + "content": "[23] Linjie Lyu, Ayush Tewari, Marc Habermann, Shunsuke Saito, Michael Zollhöfer, Thomas Leimkuhler, and Christian Theobalt. Diffusion posterior illumination for ambiguity-aware inverse rendering. ACM Transactions on Graphics (TOG), 42(6):1-14, 2023. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 578, + 553, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 578, + 553, + 622 + ], + "spans": [ + { + "bbox": [ + 316, + 578, + 553, + 622 + ], + "type": "text", + "content": "[24] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International conference on machine learning, pages 8162-8171. PMLR, 2021. 4" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 624, + 553, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 553, + 657 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 553, + 657 + ], + "type": "text", + "content": "[25] Geoffrey Oxholm and Ko Nishino. Shape and reflectance from natural illumination. In European Conference on Computer Vision, pages 528-541. Springer, 2012. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 658, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 553, + 713 + ], + "type": "text", + "content": "[26] Foivos Paraperas Papantoniou, Alexandros Lattas, Stylianos Moschoglou, and Stefanos Zafeiriou. Relightify: Relightable 3d faces from a single image via diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8806-8817, 2023. 2" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "text", + "content": "5781" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 139 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 139 + ], + "type": "text", + "content": "[27] Pakkapon Phongthawee, Worameth Chinchuthakun, Nontaphat Sinsunthithet, Varun Jampani, Amit Raj, Pramook Khungurn, and Supasorn Suwajanakorn. Diffusionlight: Light probes for free by painting a chrome ball. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 98-108, 2024. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 140, + 294, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 140, + 294, + 195 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 294, + 195 + ], + "type": "text", + "content": "[28] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 2, 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 197, + 294, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 197, + 294, + 239 + ], + "spans": [ + { + "bbox": [ + 56, + 197, + 294, + 239 + ], + "type": "text", + "content": "[29] William H. Press, Saul A. Teukolsky, William T. Vetterling, and Brian P. Flannery. Numerical Recipes 3rd Edition: The Art of Scientific Computing. Cambridge University Press, 2007. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 241, + 294, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 241, + 294, + 306 + ], + "spans": [ + { + "bbox": [ + 56, + 241, + 294, + 306 + ], + "type": "text", + "content": "[30] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 308, + 294, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 308, + 294, + 363 + ], + "spans": [ + { + "bbox": [ + 56, + 308, + 294, + 363 + ], + "type": "text", + "content": "[31] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2, 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 365, + 294, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 365, + 294, + 397 + ], + "spans": [ + { + "bbox": [ + 56, + 365, + 294, + 397 + ], + "type": "text", + "content": "[32] Tim Salimans and Jonathan Ho. Progressive distillation for fast sampling of diffusion models. arXiv preprint arXiv:2202.00512, 2022. 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 399, + 294, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 399, + 294, + 432 + ], + "spans": [ + { + "bbox": [ + 56, + 399, + 294, + 432 + ], + "type": "text", + "content": "[33] Shen Sang and Manmohan Chandraker. Single-shot neural relighting and svbrdf estimation. In European Conference on Computer Vision, pages 85-101. Springer, 2020. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 434, + 294, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 434, + 294, + 465 + ], + "spans": [ + { + "bbox": [ + 56, + 434, + 294, + 465 + ], + "type": "text", + "content": "[34] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 467, + 294, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 467, + 294, + 499 + ], + "spans": [ + { + "bbox": [ + 56, + 467, + 294, + 499 + ], + "type": "text", + "content": "[35] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. Advances in neural information processing systems, 32, 2019. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 501, + 294, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 501, + 294, + 567 + ], + "spans": [ + { + "bbox": [ + 56, + 501, + 294, + 567 + ], + "type": "text", + "content": "[36] Pratul P Srinivasan, Ben Mildenhall, Matthew Tancik, Jonathan T Barron, Richard Tucker, and Noah Snively. Lighthouse: Predicting lighting volumes for spatially-coherent illumination. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8080-8089, 2020. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 568, + 294, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 568, + 294, + 623 + ], + "spans": [ + { + "bbox": [ + 56, + 568, + 294, + 623 + ], + "type": "text", + "content": "[37] Igor Vasiljevic, Nick Kolkin, Shanyi Zhang, Ruotian Luo, Haochen Wang, Falcon Z Dai, Andrea F Daniele, Mohammadreza Mostajabi, Steven Basart, Matthew R Walter, et al. Diode: A dense indoor and outdoor depth dataset. arXiv preprint arXiv:1908.00463, 2019. 2, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 624, + 294, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 294, + 678 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 294, + 678 + ], + "type": "text", + "content": "[38] Giuseppe Vecchio, Renato Sortino, Simone Palazzo, and Concetto Spampinato. Matfuse: controllable material generation with diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4429-4438, 2024. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 680, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 680, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 680, + 294, + 713 + ], + "type": "text", + "content": "[39] Zian Wang, Jonah Philion, Sanja Fidler, and Jan Kautz. Learning indoor inverse rendering with 3d spatially-varying lighting. In Proceedings of the IEEE/CVF International" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 575 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "type": "text", + "content": "Conference on Computer Vision, pages 12538-12547, 2021. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 96, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 150 + ], + "type": "text", + "content": "[40] Jiaye Wu, Sanjoy Chowdhury, Hariharmano Shanmugaraja, David Jacobs, and Soumyadip Sengupta. Measured albedo in the wild: Filling the gap in intrinsics evaluation. In 2023 IEEE International Conference on Computational Photography (ICCP), pages 1-12. IEEE, 2023. 2, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 152, + 553, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 553, + 195 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 553, + 195 + ], + "type": "text", + "content": "[41] Jiahui Yu, Zhe Lin, Jimei Yang, Xiaohui Shen, Xin Lu, and Thomas S Huang. Free-form image inpainting with gated convolution. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4471-4480, 2019. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 197, + 555, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 197, + 555, + 240 + ], + "spans": [ + { + "bbox": [ + 316, + 197, + 555, + 240 + ], + "type": "text", + "content": "[42] Ye Yu and William AP Smith. Inverserendernet: Learning single image inverse rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3155-3164, 2019. 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 241, + 553, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 241, + 553, + 306 + ], + "spans": [ + { + "bbox": [ + 316, + 241, + 553, + 306 + ], + "type": "text", + "content": "[43] Zheng Zeng, Valentin Deschaintre, Iliyan Georgiev, Yannick Hold-Geoffroy, Yiwei Hu, Fujun Luan, Ling-Qi Yan, and Miloš Hasan. Rgb " + }, + { + "bbox": [ + 316, + 241, + 553, + 306 + ], + "type": "inline_equation", + "content": "\\leftrightarrow" + }, + { + "bbox": [ + 316, + 241, + 553, + 306 + ], + "type": "text", + "content": " x: Image decomposition and synthesis using material-and lighting-aware diffusion models. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 2, 4, 6, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 308, + 553, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 308, + 553, + 352 + ], + "spans": [ + { + "bbox": [ + 316, + 308, + 553, + 352 + ], + "type": "text", + "content": "[44] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 354, + 553, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 354, + 553, + 396 + ], + "spans": [ + { + "bbox": [ + 316, + 354, + 553, + 396 + ], + "type": "text", + "content": "[45] Ruo Zhang, Ping-Sing Tsai, James Edwin Cryer, and Mubarak Shah. Shape-from-shading: a survey. IEEE transactions on pattern analysis and machine intelligence, 21(8): 690-706, 1999. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 398, + 553, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 398, + 553, + 453 + ], + "spans": [ + { + "bbox": [ + 316, + 398, + 553, + 453 + ], + "type": "text", + "content": "[46] Xiuming Zhang, Pratul P Srinivasan, Boyang Deng, Paul Debevec, William T Freeman, and Jonathan T Barron. Nerfactor: Neural factorization of shape and reflectance under an unknown illumination. ACM Transactions on Graphics (TOG), 40(6):1-18, 2021. 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 454, + 553, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 454, + 553, + 518 + ], + "spans": [ + { + "bbox": [ + 316, + 454, + 553, + 518 + ], + "type": "text", + "content": "[47] Jingsen Zhu, Fujun Luan, Yuchi Huo, Zihao Lin, Zhihua Zhong, Dianbing Xi, Rui Wang, Hujun Bao, Jiaxiang Zheng, and Rui Tang. Learning-based inverse rendering of complex indoor scenes with differentiable monte carlo raytracing. In SIGGRAPH Asia 2022 Conference Papers. ACM, 2022. 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 520, + 553, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 520, + 553, + 575 + ], + "spans": [ + { + "bbox": [ + 316, + 520, + 553, + 575 + ], + "type": "text", + "content": "[48] Rui Zhu, Zhengqin Li, Janarbek Matai, Fatih Porikli, and Manmohan Chandraker. Irisformer: Dense vision transformers for single-image inverse rendering in indoor scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2822-2831, 2022. 2" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "5782" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Chapter-Llama_ Efficient Chaptering in Hour-Long Videos with LLMs/c5a48e4c-7b19-4801-9ffe-45905e053adb_content_list.json b/2025/Chapter-Llama_ Efficient Chaptering in Hour-Long Videos with LLMs/c5a48e4c-7b19-4801-9ffe-45905e053adb_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..b06fb9ef908fe5da30e1d83289e17b10994430dd --- /dev/null +++ b/2025/Chapter-Llama_ Efficient Chaptering in Hour-Long Videos with LLMs/c5a48e4c-7b19-4801-9ffe-45905e053adb_content_list.json @@ -0,0 +1,1351 @@ +[ + { + "type": "text", + "text": "Chapter-Llama: Efficient Chaptering in Hour-Long Videos with LLMs", + "text_level": 1, + "bbox": [ + 156, + 128, + 841, + 151 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Lucas Ventura1,2", + "bbox": [ + 225, + 178, + 354, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Antoine Yang3", + "bbox": [ + 379, + 179, + 493, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Cordelia Schmid2", + "bbox": [ + 517, + 179, + 653, + 195 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Gül Varol", + "bbox": [ + 679, + 179, + 761, + 195 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ LIGM, École des Ponts, IP Paris, Univ Gustave Eiffel, CNRS", + "bbox": [ + 321, + 207, + 674, + 222 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2Inria, École normale supérieure, CNRS, PSL Research University 3Google DeepMind", + "bbox": [ + 246, + 224, + 750, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "https://imagine.enpc.fr/~lucas.ventura/chapter-llama/", + "bbox": [ + 269, + 243, + 720, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 250, + 292, + 325, + 308 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We address the task of video chaptering, i.e., partitioning a long video timeline into semantic units and generating corresponding chapter titles. While relatively underexplored, automatic chaptering has the potential to enable efficient navigation and content retrieval in long-form videos. In this paper, we achieve strong chaptering performance on hour-long videos by efficiently addressing the problem in the text domain with our 'Chapter-Llama' framework. Specifically, we leverage a pretrained large language model (LLM) with large context window, and feed as input (i) speech transcripts and (ii) captions describing video frames, along with their respective timestamps. Given the inefficiency of exhaustively captioning all frames, we propose a lightweight speech-guided frame selection strategy based on speech transcript content, and experimentally demonstrate remarkable advantages. We train the LLM to output timestamps for the chapter boundaries, as well as free-form chapter titles. This simple yet powerful approach scales to processing one-hour long videos in a single forward pass. Our results demonstrate substantial improvements (e.g., 45.3 vs $26.7\\%$ F1 score) over the state of the art on the recent VidChapters-7M benchmark. To promote further research, we release our code and models at our project page.", + "bbox": [ + 88, + 325, + 485, + 650 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 681, + 215, + 696 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "According to a study by [46], the video durations uploaded to the popular online video sharing platform YouTube have increased steadily over the years. Videos have become longer since the first video upload in 2005 [19, 47]. In 2020, $25\\%$ of videos were estimated to be longer than 15 minutes, $5\\%$ more than 3 hours [46]. Long-form videos such as news, sports, educational, and vlog streams can often span extensive durations and cover multiple topics [97]. Finding specific content within increased video duration and volume makes efficient content navigation more important than ever.", + "bbox": [ + 89, + 707, + 483, + 854 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "However, much of the traditional video analysis research has focused on processing short videos of a few seconds [4, 15, 34, 55, 63, 67, 74, 78, 85, 87, 98, 110]. At the same", + "bbox": [ + 89, + 856, + 483, + 900 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/44bed3c22be373cb339c9910f1657007af24ef60b8b4665f684bfc137a87e205.jpg", + "image_caption": [ + "Figure 1. Chapter-Llama: Our method generates automatic video chapters for hour-long videos by training a large language model (LLM) to predict chapter boundaries and titles. The LLM processes transcribed speech (ASR) and descriptive captions of key frames, which are sampled based on ASR content. This text-based approach, equipped with speech-based frame selection, enables efficient processing of long-form content." + ], + "image_footnote": [], + "bbox": [ + 517, + 292, + 901, + 496 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "time, the definition of long videos has changed within the past decade. Early works claimed processing 100 frames (i.e., a few seconds) to be long [61, 93] as opposed to ingesting up to 16 frames [83, 92]. With the introduction of datasets containing 1-5 minute videos [29, 37, 44, 56, 82, 126], several minutes were considered very long. Studying hour-long videos has only recently seen an interest in the context of movie description [31], video captioning [40], or grounding [32, 84]. Very recently, the work of [109] collected the VidChapters-7M dataset with videos spanning from minutes to hours, along with their user-defined video chapters, and proposed the video chapter generation task, automatically dividing a video into thematic sections (i.e., chapters) with descriptive concise chapter titles. Video chaptering, if achieved successfully, can offer a compelling solution to long content indexing, bypassing the current need for time-consuming manual annotation by video owners [109].", + "bbox": [ + 511, + 619, + 908, + 854 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we address the challenge of automatic video chaptering with a simple yet effective framework designed to handle hour-long videos. Existing work for chaptering [109]", + "bbox": [ + 511, + 856, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "18947", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "relics on a dense video captioning model Vid2Seq [110], which combines multimodal inputs from video frames and ASR-based speech transcriptions. However, Vid2Seq operates on a fixed number of equally sampled frames (i.e., 100 frames), potentially missing important visual information. Furthermore, their approach based on transformer architecture uses video frame features directly, which requires learning a mapping from the visual modality to the textual modality. In contrast, our method is designed to address these limitations by (i) dynamically sampling keyframes from the video based on the speech content, and (ii) designing a purely text-based model leveraging image captioning to convert RGB frames into text.", + "bbox": [ + 89, + 90, + 480, + 265 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our approach leverages a pretrained LLM, which we finetune specifically for the video chaptering task to predict jointly the chapter boundary timestamps and chapter titles, both in text form. The appeal of our model lies in processing only textual data as input, allowing us effectively leverage the long-context understanding capabilities of the LLM to scale to long videos. In particular, we incorporate speech transcriptions from automatic speech recognition (ASR) and automatic frame captions. Captioning has been used for video understanding as an intermediate representation in recent works, but in the context of retrieval or question answering (QA) for shorter videos (maximum 3 minutes) [58, 95, 116, 121]. In longer videos, since captioning every frame is computationally prohibitive, we employ a speech-based frame selection strategy that scales efficiently while preserving important content. Similar in spirit to [43], we primarily use audio to determine keyframes, specifically bootstrapping with an LLM trained only with the speech inputs. However, even when transforming a video into text, LLMs have a limited context window, allowing a maximum number of tokens as input in a single forward pass. To mitigate context window limitations for very long video inputs, we simply perform an iterative prediction, sequentially processing the video, where each iteration typically operates on a window length of about an hour duration. We evaluate our approach on 'short' (0-15 min), 'medium' (15-30 min), and 'long' (30-60 min) videos from the VidChapters-7M dataset [109], demonstrating significant improvements over the state of the art across multiple metrics, including temporal boundary accuracy and semantic relevance of chapter titles. Our experiments show that finetuning the LLM, our speech-based frame selection strategy, and the integration of modalities from both speech and captions are crucial for achieving high-quality video chaptering results.", + "bbox": [ + 91, + 268, + 485, + 738 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions are the following: (i) We introduce Chapter-Llama: our framework leverages a pretrained LLM and finetunes for the underexplored task of video chaptering by transforming the video input into text form through ASR and captioning. (ii) We scale efficiently to hour-long videos by incorporating a speech-based frame sampling strategy, captioning only a subset of the video frames. (iii) Our simple and effective approach outperforms the state of the art on the recent VidChapters-7M benchmark by a large margin (e.g., 45.3 vs 26.7 F1 score). These results are complemented by a comprehensive set of experiments analyzing our components.", + "bbox": [ + 89, + 739, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 89, + 648, + 106 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We provide an overview of video tasks related to video chaptering, such as temporal segmentation and captioning, along with a discussion on works focusing on long-form and LLM-based video understanding.", + "bbox": [ + 511, + 114, + 906, + 174 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Temporal video segmentation. While video chaptering is a new task [109], there is a rich literature on methods focused on temporally segmenting a video in various forms. One task is shot detection [72, 76, 81], where any visual changes (e.g., shifting between two cameras) would require a temporal boundary, not necessarily modeling semantic shifts. Video scene segmentation, often studied on movies [38], is primarily focusing on grouping scenes with similar content [13, 14, 38, 39, 59, 65, 66, 71, 75, 77, 102, 111]. Another line of work considers boundary detection for temporal action segmentation [7, 23, 26, 48, 113], or localization [18, 54, 118, 120]. Unlike chaptering with free-form text, action segmentation assigns a label from a predefined set of categories, and typically defines short atomic actions as the unit. In contrast to these tasks, chapter boundaries can take various different forms depending on the type and the granularity of the video (e.g., each exercise within sports video, each slide within a lecture, each step in instructional video, each topic in a podcast video). Shot, scene, or action boundaries therefore may or may not correspond to complex chapter boundary definitions. Moreover, these tasks are mostly tackled with vision-only inputs [81, 113, 120], without leveraging speech. While text and audio segmentation have also been tackled separately [28, 73], video chaptering is based on both audio and vision inputs [109].", + "bbox": [ + 511, + 186, + 908, + 537 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Video captioning. Generating chapter titles [109] is relevant to the task of captioning that seeks to describe the video content with text. There is a large literature on single video captioning [16, 50, 78, 80], often focusing on short video clips. Typical datasets for training such as MSR-VTT [107], WebVid [5], HowTo100M [57], Video-CC [60] include captions of videos spanning a few seconds (5-15sec on average). In generic event boundary captioning [100], event intervals are similarly short, in the order of 2 seconds. On the other hand, video summarization methods operate on longer videos; however, their goal is to reduce the entire video into a single summary description [1, 2, 33, 40, 51, 117, 123, 124], not necessarily with a temporal segmentation component. Dense video captioning [37, 44, 99, 110, 127, 128] is the closest to video chaptering in terms of problem formulation, aiming to both temporally localize and caption different events. Indeed, prior work on video chaptering trains the dense captioning method of Vid2Seq [110] on the VidChapters-7M dataset [109], but relies on a fixed number of equally sampled frames. In this paper, we leverage some of the annotations of this dataset to train an LLM-based chaptering model substantially outperforming previous methods [109, 110].", + "bbox": [ + 511, + 550, + 908, + 858 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Long-form video understanding. The definition of long videos has evolved with the release of various datasets spanning", + "bbox": [ + 511, + 869, + 906, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "18948", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "seconds [106, 108], a few minutes [22, 29, 56, 86], 10-30 minutes [2, 125], or one hour [24, 40, 84, 104, 109]. MLVU [125] introduces a benchmark for evaluating multiple long video understanding tasks such as summarization and QA; however, the data is not suitable for chaptering due to lack of annotations. Video-MME [24] also contains hour-long videos for QA. MAD [31, 84] provides audio description for long movies, but each description spans a few seconds and the sparse coverage over the video is different than contiguous chapters. Recently, Ego4D-HCap [40] was proposed for hierarchical video summarization. However, this dataset involves dense captioning with visual inputs only, while we focus on video chaptering with visual and speech inputs. To the best of our knowledge, VidChapters7M [109] is the only open-sourced dataset for training and evaluating chapter generation, which we employ in this paper. Nonpublic related datasets include NewsNet [104] which includes hierarchical temporal segmentation annotations, the TV news chaptering dataset used in [30], and the ChapterGen dataset [10].", + "bbox": [ + 89, + 90, + 485, + 354 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Increased video lengths led to a range of works focusing on efficient temporal modeling strategies. A common technique to deal with longer videos is to use pre-extracted visual features [31, 84, 115]. For end-to-end learning with transformers, several works explored factorized spatio-temporal attention [3, 5, 8]. Others have looked at various ways to incorporate a memory mechanisms [42, 103], blockwise attention [52, 53], or captioning frames to exploit LLMs [101, 121]. Given the redundancy in consecutive video frames, frame selection methods were explored in the context of short video captioning and action recognition [17, 105], as well as 'long' video QA in 3-minute durations [64, 88, 114]. Most common approach with current large video models is to perform sparse sampling with equal spacing [12, 45, 110]. SCSampler [43] exploits the low-dimensional audio modality to efficiently select salient video clips for action recognition. In our method, we also leverage audio, but in the form of ASR, and run the costly frame captioning step only on keyframes on locations predicted by a speech-based frame selection module.", + "bbox": [ + 89, + 354, + 486, + 633 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "LLM use in video understanding. LLMs such as GPT [9, 68], Llama [20, 90, 91], and Gemini [27, 89], have been leveraged in different ways for improving video understanding. A popular approach is to train 'bridge' modules between pretrained visual backbones [69] and LLMs to build vision-language models (VLMs) that can ingest videos (e.g., Video-Llama [122], Video-LLaVa [49]). Other works have employed LLMs for automatic construction of video datasets [2, 40, 80, 96], tool use [58], storing memory in video QA [42], and temporal localization [36]. Similar to us, VideoTree [101] and VideoAgent [21] caption keyframes before passing them to an LLM together with a question for answer generation, addressing the limitations of [121] which performs a similar methodology without keyframe selection on shorter videos. In this study, we find that captioning alone is not sufficient, and needs to be complemented with ASR for competitive chaptering performance. Close to us, [2] exploits ASR on long videos and summarizes", + "bbox": [ + 89, + 651, + 486, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "them with LLMs to generate pseudo-labels for video summarization training. In our work, we leverage LLMs, specifically finetuning a Llama model [20] for chaptering by prompting with speech transcription and frame captions. We show that finetuning is essential for adapting to the task so that the LLM picks up relevant content within the large context input [79].", + "bbox": [ + 511, + 90, + 908, + 181 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Chapter-Llama: LLM-based Video Chaptering", + "text_level": 1, + "bbox": [ + 511, + 191, + 906, + 209 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We provide an overview of our video chaptering framework, referred to as Chapter-Llama, in Fig. 2. Given video frames and speech transcripts, we aim at predicting relevant chapter boundaries and titles. For this, we first select video frames to process with a speech-based frame selection module. Then we use an off-the-shelf visual captioner to map the selected frames in the text space. We feed the resulting captions, along with speech transcripts, to the LLM which outputs the chapter boundaries and titles jointly as a single sequence of tokens. Finally, we devise an iterative prediction procedure in case the input text sequence is too long to handle for the LLM. We next describe in more detail each component.", + "bbox": [ + 511, + 215, + 908, + 393 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Task formulation. Video chaptering [109] aims at segmenting a video into semantically meaningful chapters, and generating a title for each segment. The chapters are contiguous, with no gaps between them, and together span the entire video duration from start to end. Formally, given video frames $V = (v_{1}, v_{2}, \\dots, v_{N})$ and temporally-aligned speech transcripts $S = (s_{1}, s_{2}, \\dots, s_{M})$ , where each speech transcript contains an utterance and its associated start and end timestamps, the task is to output a sequence of chapters $C = (c_{1}, c_{2}, \\dots, c_{L})$ , where each chapter $c_{i}$ is a tuple $(b_{i}, t_{i})$ containing a start timestamp $b_{i}$ and a descriptive title $t_{i}$ . The end time of chapter $i$ is implicitly defined by the start time of the subsequent chapter $b_{i+1}$ , or total video duration if $i = L$ .", + "bbox": [ + 511, + 402, + 908, + 594 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Speech-based frame selection. Video chaptering involves processing hour-long videos. Therefore, densely sampling frames is computationally intractable due to numerous inference passes through a vision model (e.g., a visual captioner) and exceeding standard LLM context lengths. Upon inspection of our data, we found that while the speech transcription has 257 tokens per minute on average, a caption is 66 tokens long on average hence captions would take 3,960 tokens per minute when sampling a video at 1 FPS. To address these challenges, we employ a frame selection strategy.", + "bbox": [ + 511, + 604, + 908, + 750 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Specifically, we use speech transcripts to guide which video frames to process for the vision model. This is done by first training a speech-only variant of our LLM to predict a sequence of chapter boundaries $\\{\\hat{b}_1,\\hat{b}_2,\\dots,\\hat{b}_K\\}$ from speech transcripts $S$ only. For each predicted boundary $\\hat{b}_i$ , we sample a frame $v_{i}$ from the video at that timestamp. Note that this variant is cheaper compared to the full model as it only needs ASR transcription from the audio stream, without requiring any processing of the RGB stream (i.e., captioning). We then process the video frames only at the time locations predicted", + "bbox": [ + 511, + 750, + 910, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "18949", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/d1b29773e7e14bc67b1a6e1d1b5d7b7983d9bbb7886a695c0b73e34943d3f3be.jpg", + "image_caption": [ + "Figure 2. Method overview: Our Chapter-Llama framework first selects video frames to process using speech information. Then we use a visual captioner to map the selected frames in the text space. We feed the resulting captions, along with speech transcripts, to the LLM which outputs the chapter boundaries and titles jointly as a single sequence of tokens." + ], + "image_footnote": [], + "bbox": [ + 96, + 85, + 898, + 310 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "by this model. The visual information thus complements the previous 'blind' predictions from the narrations, and allows us to refine the predictions. This results in a video representations $V_{\\text{sampled}} = (v_1, v_2, \\dots, v_K)$ where $K << N$ . For the videos that lack speech entirely (e.g., about 3% of the videos in [109]), we sample frames at 10-second intervals, with an upper bound of 100 frames to maintain computational practicality.", + "bbox": [ + 89, + 376, + 483, + 479 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Mapping video to text with timestamps. To leverage the knowledge of a pretrained LLM, we map all our inputs to text. This includes: (1) speech transcriptions $S = (s_{1}, s_{2}, \\dots, s_{M})$ from the audio modality, and (2) caption descriptions $V_{\\text{caption}} = (d_{1}, d_{2}, \\dots, d_{K})$ from the visual modality. In detail, for speech transcriptions, we use ASR outputs provided by [109], obtained using the Whisper-Large-V2 [70] model through the WhisperX [6] implementation. For captioning, we employ MiniCPM-V [112] as an image captioner, applied independently on the selected video frames, i.e., $d_{i} = \\text{Captioner}(v_{i})$ .", + "bbox": [ + 89, + 508, + 483, + 656 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As we aim at predicting relevant chapter boundaries, we provide temporal information to the LLM. For both modalities, we prepend the timestamp information formatted as \"HH:MM:SS\" to encode the location at which the speech or caption is obtained.", + "bbox": [ + 89, + 659, + 483, + 718 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Captions naturally come from a single point in time. Speech segments cover intervals, but their duration is typically very short (3-4 seconds). We therefore simply use the start time of each transcribed speech interval. We interleave the speech and caption inputs based on their timestamps in a sorted order. We add a modality-specific prefix to each timestamp to denote which modality the information is extracted from (i.e., ASR for speech transcripts, Caption for captions).", + "bbox": [ + 89, + 720, + 483, + 838 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We prepend the text combining speech transcripts and captions with a fixed prompt that provides task instructions (see sup. mat. for the exact wording). This prompt occupies approximately 90 tokens and is independent of video length.", + "bbox": [ + 89, + 840, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Language model. We derive our framework by making use of a powerful pretrained LLM. Specifically, we employ the recent Llama-3.1-8B-Instruct [20] model and further finetune on chapter annotations using the LoRA technique [35]. Given the input structure previously described, the LLM is trained to output chapters, where each chapter consists of a timestamp in HH:MM:SS format followed by a free-form chapter title. We treat both the timestamps and titles simply as text tokens and apply the standard cross-entropy loss over the original vocabulary of the pretrained LLM. We apply teacher forcing during training and decode tokens autoregressively at inference. Note that the final model (taking both speech and captions as input) is trained independently from the speech-only version of our model used for frame selection, but these two models share the same backbone, and only differ in their LoRA parameters (13MB each). Across all experiments, we finetune models for a single epoch and use the same hyperparameters. We provide these hyperparameters, along with implementation details in Appendix A, and provide experiments with several Llama variants in Appendix C.", + "bbox": [ + 511, + 376, + 908, + 655 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Iterative prediction for long videos. The inputs may exceed the context window limitation of the LLM, especially in the case of long videos. For example, on an A6000 GPU, the Llama-3.1-8B-Instruct [20] model can process videos up to around 15k tokens during training, which corresponds to 50 minutes of video content on average, and 25k tokens during inference, which corresponds to 80 minutes of video content on average. To address this issue, during training, we select videos that have less than 15k tokens. Since there are videos up to 1 hour long in the training set that satisfy this constraint, and since we do not need the entire training dataset to achieve good performance, this token limitation does not hinder our training. During evaluation, we predict chapters for each chunk sequentially, such that the start of a chunk is the end of the previous chunk. Finally, we merge the predictions from all chunks to obtain chapter boundaries for the complete video. We provide more details in Appendix A.4.", + "bbox": [ + 511, + 666, + 908, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "18950", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 90, + 217, + 107 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we start by describing the data and evaluation metrics used in our experiments (Sec. 4.1). Next, we compare our results with the state of the art (Sec. 4.2), and then provide a series of ablations in our framework (Sec. 4.3). Finally, we investigate the impact of testing with very long videos exceeding our context window limitations (Sec. 4.4).", + "bbox": [ + 89, + 114, + 483, + 203 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Data and evaluation", + "text_level": 1, + "bbox": [ + 89, + 215, + 282, + 229 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Data. We train and evaluate on the recently released VidChapters-7M [109] dataset that includes user-annotated chaptered videos sourced from YouTube. Speech transcripts are obtained using Whisper [70] as the ASR method. In the original release, there is a total of 817k videos, spanning 8M chapters, with 2.4 minutes per chapter and 5.4 words per chapter title, totaling to 23 minutes and 8.3 chapters per video on average. Data is split into 801k training, 8.2k validation, and 8.2k test videos. To measure performance at different video lengths, we define three categories depending on video duration: 'short' (0-15min), 'medium' (15-30min), and 'long' (30-60min) videos. In this work, we use a subset of the training data as we observe increasing the training set brings diminishing returns at the cost of extended training times (see Fig. 4). Specifically, we use about 20k training videos (10k short videos used for the speech-based frame selection model and another 10k videos evenly split across short, medium and long durations for the final model. For state-of-the-art comparisons (Sec. 4.2), we employ the full official test set, which also contains videos without any speech (2.5% of the videos), and videos longer than 60 minutes (e.g., there are few videos that last about 12 hours). In ablations (Sec. 4.3), both for faster experimentation, and to limit the use of the test set during experimentation, we train on a randomly sampled subset of 1k videos (evenly split between short, medium, and long) and report results on a randomly sampled subset of 300 validation videos (100 from each duration) that have at least one speech utterance.", + "bbox": [ + 91, + 239, + 483, + 619 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation metrics. We primarily monitor temporal segmentation metrics to evaluate our chapter boundary detections. In particular, we employ tIoU and F1 scores. For tIoU (temporal Intersection over Union), we first compute the optimal matching between predicted and ground truth segments by greedily selecting pairs with highest IoU scores. The tIoU score is then calculated as the mean IoU across all matched pairs, multiplied by 100 to obtain a percentage. For F1 score, we first compute precision and recall at different IoU thresholds (ranging from 0.5 to 0.95 with a step of 0.05). At each threshold, a prediction is considered correct if it has IoU above the threshold with a ground truth segment. The precision is the ratio of correct predictions to total predictions, while recall is the ratio of matched ground truth segments to total ground truth segments. The F1 score is then computed as the harmonic mean of precision and recall. The final F1 metric is the average across all thresholds, multiplied by 100 to obtain a percentage. Note that [109] uses recall and precision metrics in two ways: (1)", + "bbox": [ + 89, + 636, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "by considering timestamps within 3 or 5 second thresholds as matches, and (2) by considering segments with IoU above 0.5 or 0.7 as matches. While these metrics provide point estimates at specific thresholds, we find that IoU and F1 scores offer several advantages: they evaluate performance continuously across multiple thresholds, are more interpretable, and provide a more comprehensive evaluation of the model. For completeness, we also report the metrics used in [109] in Appendix C.", + "bbox": [ + 511, + 90, + 906, + 208 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For chapter title evaluation, we follow [109] and report SODA (S) [25] and CIDEr (C) [94], which measure the quality of the titles for the predicted segments that match to the ground segments (see [109] for details).", + "bbox": [ + 511, + 209, + 906, + 267 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Comparison with the state of the art", + "text_level": 1, + "bbox": [ + 511, + 277, + 828, + 292 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In Tab. 1, we report the performance of our model on the full VidChapters-7M test set [109] ('All' columns), and compare to the state of the art reported in [109], which uses Vid2Seq [110]. Moreover, we evaluate four proprietary models using our speech-based frame selection and captioning in a zero-shot manner.", + "bbox": [ + 511, + 300, + 908, + 373 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We observe that our finetuned Chapter-Llama achieves substantial performance improvements across all metrics and video duration categories. (e.g., 45.3 vs 26.7 F1 and 19.3 vs 11.6 SODA compared to Vid2Seq). Notably, our improvement over Vid2Seq is more important for medium and long videos compared to short videos. Note that our final approach was trained using the subset of data detailed in the previous section, specifically 20k videos, which constitutes only $2.5\\%$ of the total available training data. In contrast, the baseline Vid2Seq model [110] was trained on a considerably larger dataset, utilizing both HowTo100M [57] and the entire VidChapters-7M training set.", + "bbox": [ + 511, + 375, + 908, + 535 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Additionally, we report performances of our model without training on any chapter annotations (i.e., both the speech-based frame selector and the LLM are not finetuned, and run with the same prompt as in the finetuned setting). We see that our zero-shot method also achieves competitive performance (e.g., 29.5 F1), whereas Vid2Seq only trained on HowTo100M does not generalize (3.0 F1).", + "bbox": [ + 511, + 536, + 908, + 637 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Finally, when zero-shot evaluating the proprietary models, GPT4-o [62] and Gemini variants [27], with our speech-based frame selection and captioning inputs, we observe competitive performances (e.g., 42.2 F1 with Gemini-1.5-Pro); however, our Chapter-Llama still surpasses on all metrics. Note that, due to API costs of the proprietary models, we performed their evaluation on a random $10\\%$ subset of the test set; however, we verified that the scores are similar between $10\\%$ and $100\\%$ of the test set when evaluating with Chapter-Llama.", + "bbox": [ + 511, + 638, + 908, + 771 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Qualitative comparison. In Fig. 3, we provide qualitative examples comparing our method against Vid2Seq [109, 110] and our zero-shot baseline. Our predictions align well with the ground truth chapters, accurately capturing both the temporal boundaries and generating relevant titles. In contrast, Vid2Seq segments tend to be less accurate, and we also observe that it often produces repetitive titles (bottom example). The zero-shot Chapter-Llama baseline tends to generate relatively longer and", + "bbox": [ + 511, + 782, + 908, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "18951", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/999eb42551b65875519085b63c1a803fe9942893b8b6146575f1458687b57c3d.jpg", + "table_caption": [], + "table_footnote": [ + "Table 1. Comparison to the state of the art on VidChapters-7M test set: We split the table into (bottom) the comparison between Chapter-Llama and the state-of-the-art method Vid2Seq [110], and (top) the evaluation of proprietary models. Chapter-Llama significantly outperforms Vid2Seq trained and reported by [109] (45.3 vs 26.7 F1). Our method also achieves strong performance in zero-shot mode - without finetuning (Ft.) on any chapter annotation (29.5 F1). Furthermore, we report performance of proprietary models in such zero-shot setting, using our speech-based frame selection and captioning, and observe inferior results than Chapter-Llama (42.2 F1 with Gemini-1.5-Pro). Note that we use the full official 8.1k test set videos ('All'), unlike in the remaining experiments that report on the validation subset. We also report the performance breakdown into short (4891), medium (1736), and long (892) test videos. Our model was trained on 10k videos balanced across short, medium and long durations. $\\dagger$ denotes evaluation on a random $10\\%$ subset of the test set due to API costs of proprietary models." + ], + "table_body": "
BackboneFrame selectionFt.ShortMediumLongAll
F1tIoUSCF1tIoUSCF1tIoUSCF1tIoUSC
GPT-4o-mini [62]†OursX32.164.57.242.430.562.36.130.628.061.06.027.331.263.66.837.8
GPT-4o [62]†OursX37.768.08.453.838.168.88.151.436.566.26.634.837.668.08.151.0
Gemini-2.0-Flash [27]†OursX39.969.212.072.843.871.411.270.334.966.29.051.640.269.311.469.7
Gemini-1.5-Pro [27]†OursX41.770.611.765.343.871.811.261.441.370.610.155.342.270.911.463.2
Vid2Seq [109, 110]EquidistantX2.528.60.30.33.229.70.30.44.632.00.30.53.029.30.30.4
Llama 3.1-8BOursX29.963.47.134.530.662.75.428.126.659.33.618.929.562.56.230.7
Vid2Seq [109, 110]Equidistant✓33.463.715.274.919.053.37.531.916.750.85.928.426.758.611.655.8
Llama 3.1-8B (Chapter-Llama)Ours✓45.572.220.2103.546.772.318.898.741.369.215.891.245.371.819.3100.9
", + "bbox": [ + 94, + 114, + 898, + 258 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/0f877bd2a95cfe6c084d86dd98b61f7f56dd78f44b04f115579661268a192f75.jpg", + "image_caption": [ + "Figure 3. Qualitative results: We display two examples and compare our Chapter-Llama results against the ground truth (GT), as well as the zero-shot (ZS) and Vid2Seq (VS) baselines. For each example, we show the corresponding SODA (S) and CIDEr (C) scores. Our method overall shows the highest similarity with the GT, while Vid2Seq can suffer from repeated chapter titles, and zero-shot generations tend to over-segment." + ], + "image_footnote": [], + "bbox": [ + 112, + 419, + 879, + 830 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "18952", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/df6aac37afe671ccead58ee5c9d1075ac0a4b154b50a7193862c719feec7187b.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModalitiesSegmentationTitles
SpeechCaptionsF1tIoUSC
Zero-shotX✓12.648.61.96.4
✓X22.757.34.419.7
✓✓29.963.06.933.7
FinetunedX✓39.167.75.920.2
✓X38.568.113.967.3
✓✓42.670.616.482.4
", + "bbox": [ + 112, + 88, + 455, + 220 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "verbose chapter titles and often generates chapters that appear to be continuations of previous chapters rather than distinct segments, while also exhibiting over-segmentation issues. We provide more examples in Appendix E.", + "bbox": [ + 89, + 320, + 483, + 380 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation studies", + "text_level": 1, + "bbox": [ + 89, + 387, + 251, + 402 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In the following, we experiment with (i) the contribution of speech and caption modalities, along with the effect of LLM finetuning, (ii) the effect of our frame selection method for captioning, (iii) the amount of training data, and (iv) the use of frame embeddings instead of captions. As mentioned above, we use 1k training and 300 validation videos for these ablations.", + "bbox": [ + 89, + 409, + 483, + 500 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Modalities and LLM finetuning. In Tab. 2, we ablate the impact of finetuning the LLM and the contribution of each of the speech and caption modalities. In the top block, we run our baselines in zero-shot setting as introduced in the previous section. The speech-only baseline outperforms the captions-only baseline by a large margin in the zero-shot setting. This suggests that speech contains more relevant information for chaptering, as was previously observed by [109].", + "bbox": [ + 89, + 508, + 483, + 626 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As shown in the bottom block of Tab. 2, we observe large performance improvements when finetuning the LLM, as opposed to zero-shot. We hypothesize that zero-shot prompting with a long multi-modal text, potentially containing redundant and irrelevant information, may overwhelm the LLM [79, 101]. We obtain our best model by combining the two modalities, which performs better than the individual speech-only or caption-only models. This demonstrates the multi-modal capabilities of our model.", + "bbox": [ + 89, + 626, + 483, + 758 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Speech-based frame selection. In Tab. 3, we examine a number of strategies to sample frames at which we extract captions. In addition to previously described metrics, for each of the frame sampling approaches, we report the average number of captions per video and the average number of text tokens per minute. For reference, we also report an off-the-shelf shot detection [11] and Vid2Seq [109, 110].", + "bbox": [ + 89, + 768, + 482, + 871 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We compare our speech-based frame selection strategy to various baselines. We experiment with sampling (i) uniformly", + "bbox": [ + 89, + 871, + 483, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/a65838918ec9a355117d23ded7dc9da7391704a721123c41b31820413e9b07a7.jpg", + "table_caption": [ + "Table 2. Contribution of different modalities and finetuning: Finetuning the LLM with 1k videos largely improves chaptering performance on 300 validation videos, see bottom block vs top block. In the finetuned setting, we further demonstrate the advantages of combining both modalities, i.e., transcribed speech from ASR and automatic captions extracted from video frames." + ], + "table_footnote": [], + "table_body": "
MethodFrame selection for captionsaverage #frames#tokens per min.SegmentationTitles SC
F1tIoU
BASELINES
Shot detection [11]n/a49.4n/a6.237.6--
Vid2Seq [109, 110]100 equidistant100.0128.625.457.811.255.0
CHAPTER-LLAMA VARIANTS
SpeechCaption
✓✘n/an/a248.638.568.113.967.3
100 equidistant100.0449.121.053.88.436.0
Every 10 sec.83.1280.312.845.94.313.0
✘✓Shot boundaries49.4193.216.250.73.912.4
10 equidistant10.041.811.046.43.69.0
Speech-based10.336.239.167.75.920.2
100 equidistant100.0746.239.267.416.183.8
Every 10 sec.83.1570.141.069.315.477.3
✓✓Shot boundaries40.4481.740.669.115.879.3
10 equidistant10.0326.140.167.915.877.5
Speech-based10.3320.442.670.616.482.4
", + "bbox": [ + 517, + 89, + 903, + 303 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Frame selection strategies for captioning: We evaluate different approaches for selecting frames to extract captions from, comparing our speech-based selection method against baselines. The table shows results for models trained on 1k videos and evaluated on 300 validation videos. We experiment with using speech only, captions only, and both modalities (bottom section). For caption extraction, we compare our speech-based approach to other alternatives such as equidistant sampling (100 or 10 frames), uniformly sampling every 10 seconds, or sampling at shot boundaries using [11]. Our speech-based frame selection achieves the best overall performance (F1: 42.6, tIoU: 70.6) while requiring significantly fewer number of frames on average (10.3) compared to other sampling approaches. The tokens-per-minute statistic shows the total input length including both speech transcriptions and captions, excluding the fixed prompt template.", + "bbox": [ + 511, + 308, + 906, + 497 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "100 frames as in Vid2Seq, (ii) every 10 seconds, (iii) at shot boundaries detected by an off-the-shelf shot detector [11], (iv) 10 equidistant frames to be similar to our speech-based locations (i.e., 10.0 vs 10.3 number of frames on average), and (v) sampling at frames predicted as chapter boundaries by our LLM that inputs only speech. In all cases, we limit the maximum number of frames to 100 as in [109, 110] to handle extreme durations.", + "bbox": [ + 511, + 513, + 906, + 632 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In both caption-only and caption+speech settings, our speech-based frame selection approach achieves better segmentation results than the more frame-expensive baselines '100 equidistant', 'every 10 sec', and 'shot boundaries', while using much less frames, and also improves over the '10 equidistant' baseline which uses a similar number of frames. This demonstrates the effectiveness of our speech-based frame selection strategy.", + "bbox": [ + 511, + 633, + 906, + 736 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For reference, we also report positive comparison against shot detection and Vid2Seq [109, 110]. Note Vid2Seq has less tokens per min. compared to our 100 equidistant variants, because Vid2Seq uses a different timestamp tokenizer in the input.", + "bbox": [ + 511, + 737, + 906, + 796 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Amount of training data. Given the large-scale nature of the VidChapters-7M training set, we investigate how much chapter data is needed for LoRA finetuning the LLM. We plot the performance against the number of training videos in Fig. 4. We start by the zero-shot baseline as the first data point, and report our method with 1k, 5k, 7k, and 10k videos, split evenly", + "bbox": [ + 511, + 811, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "18953", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7e58313b19ae66019b928c53d56a372c0f5a15c1b0f4c22319e6125742aa64fe.jpg", + "image_caption": [ + "Figure 4. Amount of training data: Our experiments show a substantial improvement when moving from zero-shot to training with 1k videos. Beyond 1k videos, performance continues to improve but at a much slower rate, motivating our choice of using only 10k training videos for our final LLM." + ], + "image_footnote": [], + "bbox": [ + 89, + 87, + 478, + 217 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "between three durations. We see that after increasing above several thousand training videos starts to bring diminishing returns. We therefore keep 10k training videos for our final LLM, which makes our approach highly efficient to train (40min on 4 H100 GPUs). Note that here we focus on the chaptering LLM and always use frame sampling locations from a speech-based module trained on 10k separate videos.", + "bbox": [ + 89, + 306, + 483, + 409 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Frame embeddings vs captions. In Tab. 4, we investigate whether raw visual embeddings could serve as an alternative to textual captions. To this end, we experiment with replacing the captions with frame embeddings. Specifically, for each frame, we extract the 1152-dimensional output embedding corresponding to the [CLS] token from a frozen SigLIP model [119], and feed through a 2-layer MLP mapping network. We initialize the MLP weights from MANTIS [41] and train jointly with the LLM during finetuning. The results with 'Speech+Embeddings' are better than 'Speech' alone (38.5 vs 40.4 F1), but worse than 'Speech+Captions' (42.6 vs 40.4 F1). The performance gap between 'Speech+Embeddings' and 'Speech+Captions' may be due to the richer information provided by captions, which use multiple tokens per frame, directly in text form, compared to the single [CLS] token frame embedding, requiring a mapping network to be ingested by an LLM. Finally, while combining all modalities achieves the best performance (44.4 F1), we exclude frame embeddings from our final model due to practical considerations, e.g., they add complexity, increase processing time by 2.5x, and require 3000x more storage space.", + "bbox": [ + 91, + 425, + 483, + 719 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Iterative prediction on longer videos", + "text_level": 1, + "bbox": [ + 89, + 729, + 405, + 744 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In our ablation studies, our experimental setting considered training and evaluating with videos that fit within the LLM context window. In Tab. 5, we evaluate the benefit of our iterative prediction procedure for handling videos that exceed the LLM context window. For this, we identify videos in the validation set whose inputs exceed the LLM inference context window ( $>35k$ tokens), resulting in 110 videos. On this challenging subset, we find that our iterative prediction procedure improves chaptering results compared to the baseline that only runs the LLM once by cropping the input to the first", + "bbox": [ + 89, + 753, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/ee3499cfab2b54a432978aee1ccea9ef6acc47e7276e5ab27d6dc471502f9109.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModalitiesSegmentationTitles
SpeechEmbeddingsCaptionsF1tIoUSC
✓--38.568.113.967.3
-✓-38.466.53.47.3
--✓39.167.75.920.2
✓✓-40.468.215.374.9
✓-✓42.670.616.482.4
✓✓✓44.471.516.384.2
", + "bbox": [ + 517, + 88, + 897, + 226 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/5aa4096cd947c2cb8adb4fd7a3831eeaa843ebf8743d6f7f53cd244a383982eb.jpg", + "table_caption": [ + "Table 4. Frame embeddings vs captions: We compare using frame captions versus visual features from a frozen SigLIP model projected through a learned 2-layer MLP mapping network ('Embeddings'). While the 'Speech+Embeddings' combination performs better than speech alone (40.4 vs 38.5 F1), it underperforms compared to the 'Speech+Captions' combination (42.6 vs 40.4 F1). All models are trained with 1k videos and evaluated on 300 videos." + ], + "table_footnote": [], + "table_body": "
Window# tok.avg # iter.Subset exceeding 35k tokens
F1tIoUSC
First10k113.150.54.031.2
15k116.654.95.443.3
20k118.756.76.647.5
Iterative10k8.518.557.16.925.1
15k5.423.660.18.735.2
20k4.125.361.410.344.0
", + "bbox": [ + 535, + 334, + 880, + 463 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. Iterative prediction: Our iterative prediction procedure improves chaptering results on the subset of 110 videos which exceed 35k tokens compared to the baseline that only runs the LLM once (by only taking the first window, and discarding the rest of the input sequence), across various context windows. As we increase the context window in the iterative prediction, the performance gradually improves and the average number of iterations decreases. The model is trained with 1k videos.", + "bbox": [ + 511, + 467, + 905, + 574 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "input window, across various context windows (10k, 15k, and 20k). We refer to Appendix B for details on the video lengths and statistics of videos that exceed the LLM context window.", + "bbox": [ + 511, + 593, + 903, + 638 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusions", + "text_level": 1, + "bbox": [ + 511, + 654, + 633, + 669 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We presented Chapter-Llama, an approach that leverages LLMs for hour-long video chaptering by mapping video to text using speech transcripts and efficiently captioning video frames sampled with a speech-based frame selector. Our results on VidChapters-7M consequently improved the state of the art by a large margin. We experimentally demonstrated the benefits of our components through an extensive ablation study. One limitation of our approach is that it relies on the accuracy of the ASR and the visual captioner. Future work can explore hierarchical chaptering with several granularities and consider the audio modality beyond speech. We also note that the LLM, the visual captioner, and speech transcription models are trained on large Web datasets, which can contain biases that can lead to inaccurate chaptering, especially for videos depicting underrepresented topics.", + "bbox": [ + 511, + 680, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "18954", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements. This work was granted access to the HPC resources of IDRIS under the allocation 2024-AD011014696 made by GENCI. This work was funded in part by the ANR project CorVis ANR-21-CE23-0003-01, a research gift from Google, the French government under management of Agence Nationale de la Recherche as part of the \"France 2030\" program, reference ANR-23-IACL-0008 (PR[AI]RIE-PSAI projet), and the ANR project VideoPredict ANR-21-FAI1-0002-01. Cordelia Schmid would like to acknowledge the support by the Körber European Science Prize. The authors would also like to thank Guillaume Astruc, Nikos Athanasiou, Hyolim Kang, and Nicolas Dufour for their feedback.", + "bbox": [ + 89, + 90, + 485, + 267 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 282, + 183, + 299 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Soheyla Amirian, Khaled Rasheed, Thiab R Taha, and Hamid R Arabnia. Automatic generation of descriptive titles for video clips using deep learning. In Advances in Artificial Intelligence and Applied Cognitive Computing: Proceedings from ICAI'20 and ACC'20, 2021. 2", + "[2] D. Argaw, S. Yoon, F. Heilbron, H. Deilamsalehy, T. Bui, Z. Wang, F. Dernoncourt, and J. Chung. Scaling up video summarization pretraining with large language models. In CVPR, 2024. 2, 3", + "[3] Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lucic, and Cordelia Schmid. ViViT: A video vision transformer. In ICCV, 2021. 3", + "[4] Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio Neural machine translation by jointly learning to align and translate. ICLR, 2015. 1", + "[5] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In ICCV, 2021. 2, 3", + "[6] Max Bain, Jaesung Huh, Tengda Han, and Andrew Zisserman. WhisperX: Time-accurate speech transcription of long-form audio. In Interspeech, 2023. 4", + "[7] Nadine Behrmann, S Alireza Golestaneh, Zico Kolter, Jürgen Gall, and Mehdi Noroozi. Unified fully and timestamp supervised temporal action segmentation via sequence to sequence translation. In ECCV, 2022. 2", + "[8] Gedas Bertasius, Heng Wang, and Lorenzo Torresani. Is space-time attention all you need for video understanding? In ICML, 2021. 3", + "[9] Tom Brown et al. Language models are few-shot learners. In NeurIPS, 2020. 3", + "[10] Xiao Cao, Zitan Chen, Canyu Le, and Lei Meng. Multi-modal video chapter generation. In BMVC, 2022. 3", + "[11] Brandon Castellano. Pyscenedetect: Intelligent scene cut detection and video splitting tool. https://pyscenedetect.readthedocs.io/en/latest/, 2018.7", + "[12] Houlun Chen, Xin Wang, Hong Chen, Zihan Song, Jia Jia, and Wenwu Zhu. Grounding-prompter: Prompting llm with multimodal information for temporal sentence grounding in long videos. arXiv:2312.17117, 2023. 3", + "[13] Shixing Chen, Xiaohan Nie, David Fan, Dongqing Zhang, Vimal Bhat, and Raffay Hamid. Shot contrastive self-supervised learning for scene boundary detection. In CVPR, 2021. 2" + ], + "bbox": [ + 99, + 306, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Shixing Chen, Chun-Hao Liu, Xiang Hao, Xiaohan Nie, Maxim Arap, and Raffay Hamid. Movies2Scenes: Using movie metadata to learn scene representation. In CVPR, 2023. 2", + "[15] Sihan Chen, Xingjian He, Longteng Guo, Xinxin Zhu, Weining Wang, Jinhui Tang, and Jing Liu. VALOR: Vision-audiolanguage omni-perception pretraining model and dataset. TPAMI, 2024. 1", + "[16] Sihan Chen, Handong Li, Qunbo Wang, Zijia Zhao, Mingzhen Sun, Xinxin Zhu, and Jing Liu. VAST: A vision-audio-subtitle-text omni-modality foundation model and dataset. NeurIPS, 2024. 2", + "[17] Yangyu Chen, Shuhui Wang, Weigang Zhang, and Qingming Huang. Less is more: Picking informative frames for video captioning. In ECCV, 2018. 3", + "[18] Feng Cheng and Gedas Bertasius. TALLformer: Temporal action localization with long-memory transformer. In ECCV, 2022. 2", + "[19] Xu Cheng, Cameron Dale, and Jiangchuan Liu. Statistics and social network of youtube videos. In 16th International Workshop on Quality of Service, 2008. 1", + "[20] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The Llama 3 herd of models. arXiv:2407.21783, 2024. 3, 4", + "[21] Yue Fan, Xiaojian Ma, Rujie Wu, Yuntao Du, Jiaqi Li, Zhi Gao, and Qing Li. VideoAgent: A memory-augmented multimodal agent for video understanding. In ECCV, 2024. 3", + "[22] Xinyu Fang, Kangrui Mao, Haodong Duan, Xiangyu Zhao, Yining Li, Dahua Lin, and Kai Chen. MMBench-video: A long-form multi-shot benchmark for holistic video understanding. In NeurIPS Datasets and Benchmarks, 2024. 3", + "[23] Yazan Abu Farha and Jurgen Gall. MS-TCN: Multi-stage temporal convolutional network for action segmentation. In CVPR, 2019. 2", + "[24] Chaoyou Fu, Yuhan Dai, Yondong Luo, Lei Li, Shuhuai Ren, Renrui Zhang, Zihan Wang, Chenyu Zhou, Yunhang Shen, Mengdan Zhang, et al. Video-MME: The first-ever comprehensive evaluation benchmark of multi-modal LLMs in video analysis. arXiv:2405.21075, 2024. 3", + "[25] Soichiro Fujita, Tsutomu Hirao, Hidetaka Kamigaito, Manabu Okumura, and Masaaki Nagata. SODA: Story oriented dense video captioning evaluation framework. In ECCV, 2020. 5", + "[26] Shang-Hua Gao, Qi Han, Zhong-Yu Li, Pai Peng, Liang Wang, and Ming-Ming Cheng. Global2Local: Efficient structure search for video action segmentation. In CVPR, 2021. 2", + "[27] Gemini Team et al. Gemini: A family of highly capable multimodal models. arXiv:2312.11805, 2024. 3, 5, 6", + "[28] Azin Ghazimatin, Ekaterina Garmash, Gustavo Penha, Kristen Sheets, Martin Achenbach, Oguz Semerci, Remi Galvez, Marcus Tannenberg, Sahitya Mantravadi, Divya Narayanan, Ofeliya Kalaydzhyan, Douglas Cole, Ben Carterette, Ann Clifton, Paul N. Bennett, Claudia Hauff, and Mounia Lalmas. PODTILE: Facilitating podcast episode browsing with auto-generated chapters. In Proceedings of the 33rd ACM International Conference on Information and Knowledge Management, 2024. 2", + "[29] Kristen Grauman et al. Ego4D: Around the world in 3,000 hours of egocentric video. In CVPR, 2022. 1, 3", + "[30] Khalil Guetari, Yannis Tevisen, and Frédéric Petitpont. Multimodal chaptering for long-form TV newscast video. arXiv:2406.17590, 2024. 3" + ], + "bbox": [ + 522, + 92, + 906, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "18955", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[31] Tengda Han, Max Bain, Arsha Nagrani, Gül Varol, Weidi Xie, and Andrew Zisserman. AutoAD: Movie description in context. In CVPR, 2023. 1, 3", + "[32] Tanveer Hannan, Md Mohaiminul Islam, Thomas Seidl, and Gedas Bertasius. RGNet: A unified retrieval and grounding network for long videos. In ECCV, 2023. 1", + "[33] Bo He, Jun Wang, Jielin Qiu, Trung Bui, Abhinav Shrivastava, and Zhaowen Wang. Align and attend: Multimodal summarization with dual contrastive losses. In CVPR, 2023. 2", + "[34] Chiori Hori, Takaaki Hori, Teng-Yok Lee, Ziming Zhang, Bret Harsham, John R Hershey, Tim K Marks, and Kazuhiko Sumi. Attention-based multimodal fusion for video description. In ICCV, 2017. 1", + "[35] Edward J Hu, yelong shen, Phillip Wallis, Zeyuan Allen-Zhu, Yanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In ICLR, 2022. 4", + "[36] De-An Huang, Shijia Liao, Subhashree Radhakrishnan, Hongxu Yin, Pavlo Molchanov, Zhiding Yu, and Jan Kautz. Lita: Language instructed temporal-localization assistant. In ECCV, 2024. 3", + "[37] Gabriel Huang, Bo Pang, Zhenhai Zhu, Clara Rivera, and Radu Soricut. Multimodal pretraining for dense video captioning. In AACL-IJCNLP, 2020. 1, 2", + "[38] Qingqiu Huang, Yu Xiong, Anyi Rao, Jiaze Wang, and Dahua Lin. Movienet: A holistic dataset for movie understanding. In ECCV, 2020. 2", + "[39] Md Mohaiminul Islam, Mahmudul Hasan, Kishan Shamsundar Athrey, Tony Braskich, and Gedas Bertasius. Efficient movie scene detection using state-space transformers. In CVPR, 2023. 2", + "[40] Md Mohaiminul Islam, Ngan Ho, Xitong Yang, Tushar Nagarajan, Lorenzo Torresani, and Gedas Bertasius. Video ReCap: Recursive captioning of hour-long videos. CVPR, 2024. 1, 2, 3", + "[41] Dongfu Jiang, Xuan He, Huaye Zeng, Cong Wei, Max Ku, Qian Liu, and Wenhu Chen. Mantis: Interleaved multi-image instruction tuning. TMLR, 2024. 8", + "[42] Kumara Kahatapitiya, Kanchana Ranasinghe, Jongwoo Park, and Michael S Ryoo. Language repository for long video understanding. arXiv:2403.14622, 2024. 3", + "[43] Bruno Korbar, Du Tran, and Lorenzo Torresani. SCSampler: Sampling salient clips from video for efficient action recognition. In ICCV, 2019. 2, 3", + "[44] Ranjay Krishna, Kenji Hata, Frederic Ren, Li Fei-Fei, and Juan Carlos Niebles. Dense-captioning events in videos. In ICCV, 2017. 1, 2", + "[45] Jie Lei, Linjie Li, Luowei Zhou, Zhe Gan, Tamara L Berg, Mohit Bansal, and Jingjing Liu. Less is more: ClipBERT for video-and-language learning via sparse sampling. In CVPR, 2021. 3", + "[46] Feng Li, Jae Won Chung, and Mark Claypool. Three-year trends in YouTube video content and encoding. Proceedings of the 18th International Conference on Signal Processing and Multimedia Applications (SIGMAP), 2021. 1", + "[47] Mingzhe Li, Mark Claypool, Robert Kinicki, and James Nichols. Characteristics of streaming media stored on the web. ACM Trans. Internet Technol., 2005. 1", + "[48] Zhe Li, Yazan Abu Farha, and Jurgen Gall. Temporal action segmentation from timestamp supervision. In CVPR, 2021. 2", + "[49] Bin Lin, Bin Zhu, Yang Ye, Munan Ning, Peng Jin, and Li Yuan. Video-LLaVA: Learning united visual representation by alignment before projection. arXiv:2311.10122, 2023. 3" + ], + "bbox": [ + 99, + 92, + 485, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[50] Kevin Lin, Linjie Li, Chung-Ching Lin, Faisal Ahmed, Zhe Gan, Zicheng Liu, Yumao Lu, and Lijuan Wang. SwinBERT: End-to-end transformers with sparse attention for video captioning. In CVPR, 2022. 2", + "[51] Kevin Qinghong Lin, Pengchuan Zhang, Joya Chen, Shraman Pramanick, Difei Gao, Alex Jinpeng Wang, Rui Yan, and Mike Zheng Shou. UniVTG: Towards unified video-language temporal grounding. In ICCV, 2023. 2", + "[52] Hao Liu, Matei Zaharia, and Pieter Abbeel. Ring attention with blockwise transformers for near-infinite context. arXiv:2310.01889, 2023. 3", + "[53] Hao Liu, Wilson Yan, Matei Zaharia, and Pieter Abbeel. World model on million-length video and language with blockwise ringattention. arXiv:2402.08268, 2024. 3", + "[54] Xiaolong Liu, Qimeng Wang, Yao Hu, Xu Tang, Shiwei Zhang, Song Bai, and Xiang Bai. End-to-end temporal action detection with transformer. In IEEE Transactions on Image Processing, 2022. 2", + "[55] Huaishao Luo, Lei Ji, Botian Shi, Haoyang Huang, Nan Duan, Tianrui Li, Jason Li, Taroon Bharti, and Ming Zhou. UniVL: A unified video and language pre-training model for multimodal understanding and generation. arXiv:2002.06353, 2020. 1", + "[56] Karttikeya Mangalam, Raiymbek Akshulakov, and Jitendra Malik. EgoSchema: A diagnostic benchmark for very long-form video language understanding. In NeurIPS Track on Datasets and Benchmarks, 2023. 1, 3", + "[57] Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic. HowTo100M: Learning a text-video embedding by watching hundred million narrated video clips. In ICCV, 2019. 2, 5", + "[58] Juhong Min, Shyamal Buch, Arsha Nagrani, Minsu Cho, and Cordelia Schmid. MoReVQA: Exploring modular reasoning models for video question answering. In CVPR, 2024. 2, 3", + "[59] Jonghwan Mun, Minchul Shin, Gunsoo Han, Sangho Lee, Seongsu Ha, Joonseok Lee, and Eun-Sol Kim. BaSSL: Boundary-aware self-supervised learning for video scene segmentation. In ACCV, 2022. 2", + "[60] Arsha Nagrani, Paul Hongsuck Seo, Bryan Seybold, Anja Hauth, Santiago Manen, Chen Sun, and Cordelia Schmid. Learning audio-video modalities from image captions. In ECCV, 2022. 2", + "[61] Joe Yue-Hei Ng, Matthew Hausknecht, Sudheendra Vijayanarasimhan, Oriol Vinyals, Rajat Monga, and George Toderici. Beyond short snippets: Deep networks for video classification. In CVPR, 2015. 1", + "[62] OpenAI. Gpt-4o system card. arXiv:2410.21276, 2024. 5, 6", + "[63] Yingwei Pan, Ting Yao, Houqiang Li, and Tao Mei. Video captioning with transferred semantic attributes. In CVPR, 2017. 1", + "[64] Pinelopi Papalampidi, Skanda Koppula, Shreya Pathak, Justin Chiu, Joe Heyward, Viorica Patraucean, Jiajun Shen, Antoine Miech, Andrew Zisserman, and Aida Nematzdeh. A simple recipe for contrastively pre-training video-first encoders beyond 16 frames. In CVPR, 2024. 3", + "[65] Alejandro Pardo, Fabian Caba Heilbron, Juan León Alcázar, Ali Thabet, and Bernard Ghanem. Moviecuts: A new dataset and benchmark for cut type recognition. In ECCV, 2022. 2", + "[66] Jinwoo Park, Jungeun Kim, Jaegwang Seok, Sukhyun Lee, and Junyeong Kim. Contrasting multi-modal similarity framework for video scene segmentation. IEEE Access, 2024. 2" + ], + "bbox": [ + 522, + 92, + 906, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "18956", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[67] Wenjie Pei, Jiyuan Zhang, Xiangrong Wang, Lei Ke, Xiaoyong Shen, and Yu-Wing Tai. Memory-attended recurrent network for video captioning. In CVPR, 2019. 1", + "[68] Alec Radford, Karthik Narasimhan, Tim Salimans, and Ilya Sutskever. Improving language understanding by generative pre-training. Preprint, 2018. 3", + "[69] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 3", + "[70] Alec Radford, Jong Wook Kim, Tao Xu, Greg Brockman, Christine McLeavey, and Ilya Sutskever. Robust speech recognition via large-scale weak supervision. In ICML, 2023. 4, 5", + "[71] Anyi Rao, Linning Xu, Yu Xiong, Guodong Xu, Qingqiu Huang, Bolei Zhou, and Dahua Lin. A local-to-global approach to multi-modal movie scene segmentation. In CVPR, 2020. 2", + "[72] Zeeshan Rasheed and Mubarak Shah. Scene detection in hollywood movies and tv shows. In CVPR, 2003. 2", + "[73] Fabian Retkowski and Alexander Waibel. From text segmentation to smart chaptering: A novel benchmark for structuring video transcriptions. arXiv:2402.17633, 2024. 2", + "[74] Marcus Rohrbach, Wei Qiu, Ivan Titov, Stefan Thater, Manfred Pinkal, and Bernt Schiele. Translating video content to natural language descriptions. In ICCV, 2013. 1", + "[75] Daniel Rotman, Dror Porat, and Gal Ashour. Robust video scene detection using multimodal fusion of optimally grouped features. IEEE 19th International Workshop on Multimedia Signal Processing (MMSP), 2017. 2", + "[76] Yong Rui, Thomas S Huang, and Sharad Mehrotra. Exploring video structure beyond the shots. In IEEE International Conference on Multimedia Computing and Systems, 1998. 2", + "[77] Najmeh Sadoughi, Xinyu Li, Avijit Vajpayee, David Fan, Bing Shuai, Hector Santos-Villalobos, Vimal Bhat, and Rohith MV. MEGA: Multimodal alignment aggregation and distillation for cinematic video segmentation. In ICCV, 2023. 2", + "[78] Paul Hongsuck Seo, Arsha Nagrani, Anurag Arnab, and Cordelia Schmid. End-to-end generative pretraining for multimodal video captioning. In CVPR, 2022. 1, 2", + "[79] Freda Shi, Xinyun Chen, Kanishka Misra, Nathan Scales, David Dohan, Ed H. Chi, Nathanael Schärli, and Denny Zhou. Large language models can be easily distracted by irrelevant context. In ICML, 2023. 3, 7", + "[80] Nina Shvetsova, Anna Kukleva, Xudong Hong, Christian Rupprecht, Bernt Schiele, and Hilde Kuehne. HowToCaption: Prompting llms to transform video annotations at scale. In ECCV, 2024. 2, 3", + "[81] Panagiotis Sidiropoulos, Vasileios Mezaris, Ioannis Kompat-siaris, Hugo Meinedo, Miguel Bugalho, and Isabel Trancoso. Temporal video segmentation to scenes using high-level audiovisual features. IEEE TCSVT, 2011. 2", + "[82] Gunnar A. Sigurdsson, Gül Varol, Xiaolong Wang, Ivan Laptev, Ali Farhadi, and Abhinav Gupta. Hollywood in homes: Crowdsourcing data collection for activity understanding. In ECCV, 2016. 1", + "[83] Karen Simonyan and Andrew Zisserman. Two-stream convolutional networks for action recognition in videos. In ICLR, 2014. 1", + "[84] Mattia Soldan, Alejandro Pardo, Juan León Alcazar, Fabian Caba, Chen Zhao, Silvio Giancola, and Bernard Ghanem." + ], + "bbox": [ + 99, + 92, + 482, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "MAD: A scalable dataset for language grounding in videos from movie audio descriptions. In CVPR, 2022. 1, 3", + "[85] Chen Sun, Austin Myers, Carl Vondrick, Kevin Murphy, and Cordelia Schmid. VideoBERT: A joint model for video and language representation learning. In ICCV, 2019. 1", + "[86] Yuchong Sun, Hongwei Xue, Ruihua Song, Bei Liu, Huan Yang, and Jianlong Fu. Long-form video-language pre-training with multimodal temporal contrastive learning. In NeurIPS, 2022. 3", + "[87] Ilya Sutskever, Oriol Vinyals, and Quoc V Le. Sequence to sequence learning with neural networks. NeurIPS, 2014. 1", + "[88] Reuben Tan, Xineng Sun, Ping Hu, Jui hsien Wang, Hanieh Deilamsalehy, Bryan A. Plummer, Bryan Russell, and Kate Saenko. Koala: Key frame-conditioned long video-LLM. In CVPR, 2024. 3", + "[89] Gemini Team, Petko Georgiev, Ving Ian Lei, Ryan Burnell, Libin Bai, Anmol Gulati, Garrett Tanzer, Damien Vincent, Zhufeng Pan, Shibo Wang, et al. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context. arXiv:2403.05530, 2024. 3", + "[90] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. LLaMA: Open and efficient foundation language models. arXiv:2302.13971, 2023. 3", + "[91] Hugo Touvron et al. LLaMA 2: Open foundation and fine-tuned chat models. arXiv:2307.09288, 2023. 3", + "[92] Du Tran, Lubomir Bourdev, Rob Fergus, Lorenzo Torresani, and Manohar Paluri. Learning spatiotemporal features with 3d convolutional networks. In ICCV, 2015. 1", + "[93] Gül Varol, Ivan Laptev, and Cordelia Schmid. Long-term temporal convolutions for action recognition. TPAMI, 2018. 1", + "[94] Ramakrishna Vedantam, C Lawrence Zitnick, and Devi Parikh CIDER: Consensus-based image description evaluation. In CVPR, 2015. 5", + "[95] Lucas Ventura, Cordelia Schmid, and Gül Varol. Learning text-to-video retrieval from image captioning. *IJCV*, 2024. 2", + "[96] Lucas Ventura, Antoine Yang, Cordelia Schmid, and Güll Varol. CoVR-2: Automatic data construction for composed video retrieval. TPAMI, 2024. 3", + "[97] Caroline Violot, Tugrulcan Elmas, Igor Bilogrevic, and Mathias Humbert. Shorts vs. regular videos on YouTube: A comparative analysis of user engagement and content creation trends. In Proceedings of the 16th ACM Web Science Conference, Association for Computing Machinery, 2024. 1", + "[98] Junke Wang, Dongdong Chen, Zuxuan Wu, Chong Luo, Luowei Zhou, Yucheng Zhao, Yujia Xie, Ce Liu, Yu-Gang Jiang, and Lu Yuan. Omnivl: One foundation model for image-language and video-language tasks. NeurIPS, 2022. 1", + "[99] Teng Wang, Ruimao Zhang, Zhichao Lu, Feng Zheng, Ran Cheng, and Ping Luo. End-to-end dense video captioning with parallel decoding. In ICCV, 2021. 2", + "[100] Yuxuan Wang, Difei Gao, Licheng Yu, Weixian Lei, Matt Feiszli, and Mike Zheng Shou. GEB+: A benchmark for generic event boundary captioning, grounding and retrieval. In ECCV, 2022. 2", + "[101] Ziyang Wang, Shoubin Yu, Elias Stengel-Eskin, Jaehong Yoon, Feng Cheng, Gedas Bertasius, and Mohit Bansal. VideoTree: Adaptive tree-based video representation for LLM reasoning on long videos. arXiv:2405.19209, 2024. 3, 7" + ], + "bbox": [ + 516, + 92, + 906, + 898 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "18957", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[102] Xi Wei, Zhangxiang Shi, Tianzhu Zhang, Xiaoyuan Yu, and Lei Xiao. Multimodal high-order relation transformer for scene boundary detection. In ICCV, 2023. 2", + "[103] Chao-Yuan Wu, Yanghao Li, Karttikeya Mangalam, Haoqi Fan, Bo Xiong, Jitendra Malik, and Christoph Feichtenhofer. MeMViT: Memory-augmented multiscale vision transformer for efficient long-term video recognition. In CVPR, 2022. 3", + "[104] Haoqian Wu, Keyu Chen, Haozhe Liu, Mingchen Zhuge, Bing Li, Ruizhi Qiao, Xiujun Shu, Bei Gan, Liangsheng Xu, Bo Ren, Mengmeng Xu, Wentian Zhang, Raghavendra Ramachandra, Chia-Wen Lin, and Bernard Ghanem. NewsNet: A novel dataset for hierarchical temporal segmentation. In CVPR, 2023. 3", + "[105] Zuxuan Wu, Caiming Xiong, Chih-Yao Ma, Richard Socher, and Larry S. Davis. AdaFrame: Adaptive frame selection for fast video recognition. In CVPR, 2019. 3", + "[106] Junbin Xiao, Xindi Shang, Angela Yao, and Tat-Seng Chua. NExT-QA: Next phase of question-answering to explaining temporal actions. In CVPR, 2021. 3", + "[107] Jun Xu, Tao Mei, Ting Yao, and Yong Rui. MSR-VTT: A large video description dataset for bridging video and language. In CVPR, 2016. 2", + "[108] Hongwei Xue, Tiankai Hang, Yanhong Zeng, Yuchong Sun, Bei Liu, Huan Yang, Jianlong Fu, and Baining Guo. Advancing high-resolution video-language representation with large-scale video transcriptions. In CVPR, 2022. 3", + "[109] Antoine Yang, Arsha Nagrani, Ivan Laptev, Josef Sivic, and Cordelia Schmid. VidChapters-7M: Video chapters at scale. In NeurIPS Track on Datasets and Benchmarks, 2023. 1, 2, 3, 4, 5, 6, 7", + "[110] Antoine Yang, Arsha Nagrani, Paul Hongsuck Seo, Antoine Miech, Jordi Pont-Tuset, Ivan Laptev, Josef Sivic, and Cordelia Schmid. Vid2Seq: Large-scale pretraining of a visual language model for dense video captioning. In CVPR, 2023. 1, 2, 3, 5, 6, 7", + "[111] Yang Yang, Yurui Huang, Weili Guo, Baohua Xu, and Dingyin Xia. Towards global video scene segmentation with context-aware transformer. AAAI, 2023. 2", + "[112] Yuan Yao, Tianyu Yu, Ao Zhang, Chongyi Wang, Junbo Cui, Hongji Zhu, Tianchi Cai, Haoyu Li, Weilin Zhao, Zhihui He, Qianyu Chen, Huarong Zhou, Zhensheng Zou, Haoye Zhang, Shengding Hu, Zhi Zheng, Jie Zhou, Jie Cai, Xu Han, Guoyang Zeng, Dahai Li, Zhiyuan Liu, and Maosong Sun. MiniCPM-V: A GPT-4V level MLLM on your phone. arXiv:2408.01800, 2024. 4", + "[113] Fangqiu Yi, Hongyu Wen, and Tingting Jiang. ASformer: Transformer for action segmentation. In BMVC, 2021. 2", + "[114] Shoubin Yu, Jaemin Cho, Prateek Yadav, and Mohit Bansal. Self-chained image-language model for video localization and question answering. In NeurIPS, 2023. 3", + "[115] Rowan Zellers, Ximing Lu, Jack Hessel, Youngjae Yu, Jae Sung Park, Jize Cao, Ali Farhadi, and Yejin Choi. MERLOT: Multimodal neural script knowledge models. In NeurIPS, 2021. 3", + "[116] Andy Zeng, Maria Attarian, Brian Ichter, Krzysztof Choromanski, Adrian Wong, Stefan Welker, Federico Tombari, Aveek Purohit, Michael Ryoo, Vikas Sindhwani, et al. Socratic models: Composing zero-shot multimodal reasoning with language. In ICLR, 2023. 2", + "[117] Kuo-Hao Zeng, Tseng-Hung Chen, Juan Carlos Niebles, and Min Sun. Title generation for user generated videos. In ECCV, 2016. 2" + ], + "bbox": [ + 89, + 90, + 482, + 898 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[118] Runhao Zeng, Wenbing Huang, Mingkui Tan, Yu Rong, Peilin Zhao, Junzhou Huang, and Chuang Gan. Graph convolutional networks for temporal action localization. In CVPR, 2019. 2", + "[119] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. ICCV, 2023. 8", + "[120] Chenlin Zhang, Jianxin Wu, and Yin Li. ActionFormer: Localizing moments of actions with transformers. In ECCV, 2022. 2", + "[121] Ce Zhang, Taixi Lu, Md Mohaiminul Islam, Ziyang Wang, Shoubin Yu, Mohit Bansal, and Gedas Bertasius. A simple LLM framework for long-range video question-answering. In EMNLP, 2024. 2, 3", + "[122] Hang Zhang, Xin Li, and Lidong Bing. Video-LLaMA: An instruction-tuned audio-visual language model for video understanding. In EMNLP Demo, 2023. 3", + "[123] Shengyu Zhang, Ziqi Tan, Zhou Zhao, Jin Yu, Kun Kuang, Tan Jiang, Jingren Zhou, Hongxia Yang, and Fei Wu. Comprehensive information integration modeling framework for video titling. In ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, 2020. 2", + "[124] Cairong Zhao, Chutian Wang, Zifan Song, Guosheng Hu, Hao-nan Chen, and Xiaofan Zhai. Cap2Sum: Learning to summarize videos by generating captions. arXiv:2408.12800, 2024. 2", + "[125] Junjie Zhou, Yan Shu, Bo Zhao, Boya Wu, Shitao Xiao, Xi Yang, Yongping Xiong, Bo Zhang, Tiejun Huang, and Zheng Liu. MLVU: A comprehensive benchmark for multi-task long video understanding. arXiv:2406.04264, 2024. 3", + "[126] Luowei Zhou, Xu Chenliang, and Jason J. Corso. Towards automatic learning of procedures from web instructional videos. In AAAI, 2018. 1", + "[127] Luowei Zhou, Yingbo Zhou, Jason J Corso, Richard Socher, and Caiming Xiong. End-to-end dense video captioning with masked transformer. In CVPR, 2018. 2", + "[128] Xingyi Zhou, Anurag Arnab, Shyamal Buch, Shen Yan, Austin Myers, Xuehan Xiong, Arsha Nagrani, and Cordelia Schmid. Streaming dense video captioning. In CVPR, 2024. 2" + ], + "bbox": [ + 516, + 90, + 906, + 588 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "18958", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2025/Chapter-Llama_ Efficient Chaptering in Hour-Long Videos with LLMs/c5a48e4c-7b19-4801-9ffe-45905e053adb_model.json b/2025/Chapter-Llama_ Efficient Chaptering in Hour-Long Videos with LLMs/c5a48e4c-7b19-4801-9ffe-45905e053adb_model.json new file mode 100644 index 0000000000000000000000000000000000000000..6c9b3ca5f1e6c5486e437b1e1089e3a7597e2c83 --- /dev/null +++ b/2025/Chapter-Llama_ Efficient Chaptering in Hour-Long Videos with LLMs/c5a48e4c-7b19-4801-9ffe-45905e053adb_model.json @@ -0,0 +1,2677 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.157, + 0.13, + 0.842, + 0.152 + ], + "angle": 0, + "content": "Chapter-Llama: Efficient Chaptering in Hour-Long Videos with LLMs" + }, + { + "type": "text", + "bbox": [ + 0.227, + 0.179, + 0.356, + 0.198 + ], + "angle": 0, + "content": "Lucas Ventura1,2" + }, + { + "type": "text", + "bbox": [ + 0.38, + 0.18, + 0.495, + 0.198 + ], + "angle": 0, + "content": "Antoine Yang3" + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.18, + 0.655, + 0.196 + ], + "angle": 0, + "content": "Cordelia Schmid2" + }, + { + "type": "text", + "bbox": [ + 0.68, + 0.18, + 0.762, + 0.196 + ], + "angle": 0, + "content": "Gül Varol" + }, + { + "type": "text", + "bbox": [ + 0.323, + 0.208, + 0.676, + 0.223 + ], + "angle": 0, + "content": "\\(^{1}\\)LIGM, École des Ponts, IP Paris, Univ Gustave Eiffel, CNRS" + }, + { + "type": "text", + "bbox": [ + 0.248, + 0.226, + 0.75, + 0.241 + ], + "angle": 0, + "content": "2Inria, École normale supérieure, CNRS, PSL Research University 3Google DeepMind" + }, + { + "type": "text", + "bbox": [ + 0.27, + 0.244, + 0.721, + 0.258 + ], + "angle": 0, + "content": "https://imagine.enpc.fr/~lucas.ventura/chapter-llama/" + }, + { + "type": "title", + "bbox": [ + 0.25, + 0.293, + 0.326, + 0.309 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.326, + 0.486, + 0.651 + ], + "angle": 0, + "content": "We address the task of video chaptering, i.e., partitioning a long video timeline into semantic units and generating corresponding chapter titles. While relatively underexplored, automatic chaptering has the potential to enable efficient navigation and content retrieval in long-form videos. In this paper, we achieve strong chaptering performance on hour-long videos by efficiently addressing the problem in the text domain with our 'Chapter-Llama' framework. Specifically, we leverage a pretrained large language model (LLM) with large context window, and feed as input (i) speech transcripts and (ii) captions describing video frames, along with their respective timestamps. Given the inefficiency of exhaustively captioning all frames, we propose a lightweight speech-guided frame selection strategy based on speech transcript content, and experimentally demonstrate remarkable advantages. We train the LLM to output timestamps for the chapter boundaries, as well as free-form chapter titles. This simple yet powerful approach scales to processing one-hour long videos in a single forward pass. Our results demonstrate substantial improvements (e.g., 45.3 vs \\(26.7\\%\\) F1 score) over the state of the art on the recent VidChapters-7M benchmark. To promote further research, we release our code and models at our project page." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.682, + 0.216, + 0.697 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.708, + 0.484, + 0.856 + ], + "angle": 0, + "content": "According to a study by [46], the video durations uploaded to the popular online video sharing platform YouTube have increased steadily over the years. Videos have become longer since the first video upload in 2005 [19, 47]. In 2020, \\(25\\%\\) of videos were estimated to be longer than 15 minutes, \\(5\\%\\) more than 3 hours [46]. Long-form videos such as news, sports, educational, and vlog streams can often span extensive durations and cover multiple topics [97]. Finding specific content within increased video duration and volume makes efficient content navigation more important than ever." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.857, + 0.484, + 0.901 + ], + "angle": 0, + "content": "However, much of the traditional video analysis research has focused on processing short videos of a few seconds [4, 15, 34, 55, 63, 67, 74, 78, 85, 87, 98, 110]. At the same" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.293, + 0.903, + 0.497 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.502, + 0.907, + 0.597 + ], + "angle": 0, + "content": "Figure 1. Chapter-Llama: Our method generates automatic video chapters for hour-long videos by training a large language model (LLM) to predict chapter boundaries and titles. The LLM processes transcribed speech (ASR) and descriptive captions of key frames, which are sampled based on ASR content. This text-based approach, equipped with speech-based frame selection, enables efficient processing of long-form content." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.62, + 0.909, + 0.856 + ], + "angle": 0, + "content": "time, the definition of long videos has changed within the past decade. Early works claimed processing 100 frames (i.e., a few seconds) to be long [61, 93] as opposed to ingesting up to 16 frames [83, 92]. With the introduction of datasets containing 1-5 minute videos [29, 37, 44, 56, 82, 126], several minutes were considered very long. Studying hour-long videos has only recently seen an interest in the context of movie description [31], video captioning [40], or grounding [32, 84]. Very recently, the work of [109] collected the VidChapters-7M dataset with videos spanning from minutes to hours, along with their user-defined video chapters, and proposed the video chapter generation task, automatically dividing a video into thematic sections (i.e., chapters) with descriptive concise chapter titles. Video chaptering, if achieved successfully, can offer a compelling solution to long content indexing, bypassing the current need for time-consuming manual annotation by video owners [109]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.857, + 0.909, + 0.903 + ], + "angle": 0, + "content": "In this paper, we address the challenge of automatic video chaptering with a simple yet effective framework designed to handle hour-long videos. Existing work for chaptering [109]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18947" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.266 + ], + "angle": 0, + "content": "relics on a dense video captioning model Vid2Seq [110], which combines multimodal inputs from video frames and ASR-based speech transcriptions. However, Vid2Seq operates on a fixed number of equally sampled frames (i.e., 100 frames), potentially missing important visual information. Furthermore, their approach based on transformer architecture uses video frame features directly, which requires learning a mapping from the visual modality to the textual modality. In contrast, our method is designed to address these limitations by (i) dynamically sampling keyframes from the video based on the speech content, and (ii) designing a purely text-based model leveraging image captioning to convert RGB frames into text." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.269, + 0.486, + 0.739 + ], + "angle": 0, + "content": "Our approach leverages a pretrained LLM, which we finetune specifically for the video chaptering task to predict jointly the chapter boundary timestamps and chapter titles, both in text form. The appeal of our model lies in processing only textual data as input, allowing us effectively leverage the long-context understanding capabilities of the LLM to scale to long videos. In particular, we incorporate speech transcriptions from automatic speech recognition (ASR) and automatic frame captions. Captioning has been used for video understanding as an intermediate representation in recent works, but in the context of retrieval or question answering (QA) for shorter videos (maximum 3 minutes) [58, 95, 116, 121]. In longer videos, since captioning every frame is computationally prohibitive, we employ a speech-based frame selection strategy that scales efficiently while preserving important content. Similar in spirit to [43], we primarily use audio to determine keyframes, specifically bootstrapping with an LLM trained only with the speech inputs. However, even when transforming a video into text, LLMs have a limited context window, allowing a maximum number of tokens as input in a single forward pass. To mitigate context window limitations for very long video inputs, we simply perform an iterative prediction, sequentially processing the video, where each iteration typically operates on a window length of about an hour duration. We evaluate our approach on 'short' (0-15 min), 'medium' (15-30 min), and 'long' (30-60 min) videos from the VidChapters-7M dataset [109], demonstrating significant improvements over the state of the art across multiple metrics, including temporal boundary accuracy and semantic relevance of chapter titles. Our experiments show that finetuning the LLM, our speech-based frame selection strategy, and the integration of modalities from both speech and captions are crucial for achieving high-quality video chaptering results." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.74, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Our contributions are the following: (i) We introduce Chapter-Llama: our framework leverages a pretrained LLM and finetunes for the underexplored task of video chaptering by transforming the video input into text form through ASR and captioning. (ii) We scale efficiently to hour-long videos by incorporating a speech-based frame sampling strategy, captioning only a subset of the video frames. (iii) Our simple and effective approach outperforms the state of the art on the recent VidChapters-7M benchmark by a large margin (e.g., 45.3 vs 26.7 F1 score). These results are complemented by a comprehensive set of experiments analyzing our components." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.09, + 0.65, + 0.107 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.115, + 0.907, + 0.175 + ], + "angle": 0, + "content": "We provide an overview of video tasks related to video chaptering, such as temporal segmentation and captioning, along with a discussion on works focusing on long-form and LLM-based video understanding." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.187, + 0.909, + 0.538 + ], + "angle": 0, + "content": "Temporal video segmentation. While video chaptering is a new task [109], there is a rich literature on methods focused on temporally segmenting a video in various forms. One task is shot detection [72, 76, 81], where any visual changes (e.g., shifting between two cameras) would require a temporal boundary, not necessarily modeling semantic shifts. Video scene segmentation, often studied on movies [38], is primarily focusing on grouping scenes with similar content [13, 14, 38, 39, 59, 65, 66, 71, 75, 77, 102, 111]. Another line of work considers boundary detection for temporal action segmentation [7, 23, 26, 48, 113], or localization [18, 54, 118, 120]. Unlike chaptering with free-form text, action segmentation assigns a label from a predefined set of categories, and typically defines short atomic actions as the unit. In contrast to these tasks, chapter boundaries can take various different forms depending on the type and the granularity of the video (e.g., each exercise within sports video, each slide within a lecture, each step in instructional video, each topic in a podcast video). Shot, scene, or action boundaries therefore may or may not correspond to complex chapter boundary definitions. Moreover, these tasks are mostly tackled with vision-only inputs [81, 113, 120], without leveraging speech. While text and audio segmentation have also been tackled separately [28, 73], video chaptering is based on both audio and vision inputs [109]." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.551, + 0.909, + 0.859 + ], + "angle": 0, + "content": "Video captioning. Generating chapter titles [109] is relevant to the task of captioning that seeks to describe the video content with text. There is a large literature on single video captioning [16, 50, 78, 80], often focusing on short video clips. Typical datasets for training such as MSR-VTT [107], WebVid [5], HowTo100M [57], Video-CC [60] include captions of videos spanning a few seconds (5-15sec on average). In generic event boundary captioning [100], event intervals are similarly short, in the order of 2 seconds. On the other hand, video summarization methods operate on longer videos; however, their goal is to reduce the entire video into a single summary description [1, 2, 33, 40, 51, 117, 123, 124], not necessarily with a temporal segmentation component. Dense video captioning [37, 44, 99, 110, 127, 128] is the closest to video chaptering in terms of problem formulation, aiming to both temporally localize and caption different events. Indeed, prior work on video chaptering trains the dense captioning method of Vid2Seq [110] on the VidChapters-7M dataset [109], but relies on a fixed number of equally sampled frames. In this paper, we leverage some of the annotations of this dataset to train an LLM-based chaptering model substantially outperforming previous methods [109, 110]." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.908, + 0.903 + ], + "angle": 0, + "content": "Long-form video understanding. The definition of long videos has evolved with the release of various datasets spanning" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18948" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.091, + 0.486, + 0.355 + ], + "angle": 0, + "content": "seconds [106, 108], a few minutes [22, 29, 56, 86], 10-30 minutes [2, 125], or one hour [24, 40, 84, 104, 109]. MLVU [125] introduces a benchmark for evaluating multiple long video understanding tasks such as summarization and QA; however, the data is not suitable for chaptering due to lack of annotations. Video-MME [24] also contains hour-long videos for QA. MAD [31, 84] provides audio description for long movies, but each description spans a few seconds and the sparse coverage over the video is different than contiguous chapters. Recently, Ego4D-HCap [40] was proposed for hierarchical video summarization. However, this dataset involves dense captioning with visual inputs only, while we focus on video chaptering with visual and speech inputs. To the best of our knowledge, VidChapters7M [109] is the only open-sourced dataset for training and evaluating chapter generation, which we employ in this paper. Nonpublic related datasets include NewsNet [104] which includes hierarchical temporal segmentation annotations, the TV news chaptering dataset used in [30], and the ChapterGen dataset [10]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.356, + 0.487, + 0.635 + ], + "angle": 0, + "content": "Increased video lengths led to a range of works focusing on efficient temporal modeling strategies. A common technique to deal with longer videos is to use pre-extracted visual features [31, 84, 115]. For end-to-end learning with transformers, several works explored factorized spatio-temporal attention [3, 5, 8]. Others have looked at various ways to incorporate a memory mechanisms [42, 103], blockwise attention [52, 53], or captioning frames to exploit LLMs [101, 121]. Given the redundancy in consecutive video frames, frame selection methods were explored in the context of short video captioning and action recognition [17, 105], as well as 'long' video QA in 3-minute durations [64, 88, 114]. Most common approach with current large video models is to perform sparse sampling with equal spacing [12, 45, 110]. SCSampler [43] exploits the low-dimensional audio modality to efficiently select salient video clips for action recognition. In our method, we also leverage audio, but in the form of ASR, and run the costly frame captioning step only on keyframes on locations predicted by a speech-based frame selection module." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.652, + 0.487, + 0.902 + ], + "angle": 0, + "content": "LLM use in video understanding. LLMs such as GPT [9, 68], Llama [20, 90, 91], and Gemini [27, 89], have been leveraged in different ways for improving video understanding. A popular approach is to train 'bridge' modules between pretrained visual backbones [69] and LLMs to build vision-language models (VLMs) that can ingest videos (e.g., Video-Llama [122], Video-LLaVa [49]). Other works have employed LLMs for automatic construction of video datasets [2, 40, 80, 96], tool use [58], storing memory in video QA [42], and temporal localization [36]. Similar to us, VideoTree [101] and VideoAgent [21] caption keyframes before passing them to an LLM together with a question for answer generation, addressing the limitations of [121] which performs a similar methodology without keyframe selection on shorter videos. In this study, we find that captioning alone is not sufficient, and needs to be complemented with ASR for competitive chaptering performance. Close to us, [2] exploits ASR on long videos and summarizes" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.909, + 0.182 + ], + "angle": 0, + "content": "them with LLMs to generate pseudo-labels for video summarization training. In our work, we leverage LLMs, specifically finetuning a Llama model [20] for chaptering by prompting with speech transcription and frame captions. We show that finetuning is essential for adapting to the task so that the LLM picks up relevant content within the large context input [79]." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.192, + 0.907, + 0.21 + ], + "angle": 0, + "content": "3. Chapter-Llama: LLM-based Video Chaptering" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.217, + 0.909, + 0.394 + ], + "angle": 0, + "content": "We provide an overview of our video chaptering framework, referred to as Chapter-Llama, in Fig. 2. Given video frames and speech transcripts, we aim at predicting relevant chapter boundaries and titles. For this, we first select video frames to process with a speech-based frame selection module. Then we use an off-the-shelf visual captioner to map the selected frames in the text space. We feed the resulting captions, along with speech transcripts, to the LLM which outputs the chapter boundaries and titles jointly as a single sequence of tokens. Finally, we devise an iterative prediction procedure in case the input text sequence is too long to handle for the LLM. We next describe in more detail each component." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.403, + 0.909, + 0.595 + ], + "angle": 0, + "content": "Task formulation. Video chaptering [109] aims at segmenting a video into semantically meaningful chapters, and generating a title for each segment. The chapters are contiguous, with no gaps between them, and together span the entire video duration from start to end. Formally, given video frames \\( V = (v_{1}, v_{2}, \\dots, v_{N}) \\) and temporally-aligned speech transcripts \\( S = (s_{1}, s_{2}, \\dots, s_{M}) \\), where each speech transcript contains an utterance and its associated start and end timestamps, the task is to output a sequence of chapters \\( C = (c_{1}, c_{2}, \\dots, c_{L}) \\), where each chapter \\( c_{i} \\) is a tuple \\( (b_{i}, t_{i}) \\) containing a start timestamp \\( b_{i} \\) and a descriptive title \\( t_{i} \\). The end time of chapter \\( i \\) is implicitly defined by the start time of the subsequent chapter \\( b_{i+1} \\), or total video duration if \\( i = L \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.605, + 0.909, + 0.751 + ], + "angle": 0, + "content": "Speech-based frame selection. Video chaptering involves processing hour-long videos. Therefore, densely sampling frames is computationally intractable due to numerous inference passes through a vision model (e.g., a visual captioner) and exceeding standard LLM context lengths. Upon inspection of our data, we found that while the speech transcription has 257 tokens per minute on average, a caption is 66 tokens long on average hence captions would take 3,960 tokens per minute when sampling a video at 1 FPS. To address these challenges, we employ a frame selection strategy." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.751, + 0.911, + 0.903 + ], + "angle": 0, + "content": "Specifically, we use speech transcripts to guide which video frames to process for the vision model. This is done by first training a speech-only variant of our LLM to predict a sequence of chapter boundaries \\(\\{\\hat{b}_1,\\hat{b}_2,\\dots,\\hat{b}_K\\}\\) from speech transcripts \\(S\\) only. For each predicted boundary \\(\\hat{b}_i\\), we sample a frame \\(v_{i}\\) from the video at that timestamp. Note that this variant is cheaper compared to the full model as it only needs ASR transcription from the audio stream, without requiring any processing of the RGB stream (i.e., captioning). We then process the video frames only at the time locations predicted" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18949" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.087, + 0.9, + 0.311 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.316, + 0.908, + 0.358 + ], + "angle": 0, + "content": "Figure 2. Method overview: Our Chapter-Llama framework first selects video frames to process using speech information. Then we use a visual captioner to map the selected frames in the text space. We feed the resulting captions, along with speech transcripts, to the LLM which outputs the chapter boundaries and titles jointly as a single sequence of tokens." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.377, + 0.484, + 0.48 + ], + "angle": 0, + "content": "by this model. The visual information thus complements the previous 'blind' predictions from the narrations, and allows us to refine the predictions. This results in a video representations \\( V_{\\text{sampled}} = (v_1, v_2, \\dots, v_K) \\) where \\( K << N \\). For the videos that lack speech entirely (e.g., about 3% of the videos in [109]), we sample frames at 10-second intervals, with an upper bound of 100 frames to maintain computational practicality." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.51, + 0.484, + 0.657 + ], + "angle": 0, + "content": "Mapping video to text with timestamps. To leverage the knowledge of a pretrained LLM, we map all our inputs to text. This includes: (1) speech transcriptions \\( S = (s_{1}, s_{2}, \\dots, s_{M}) \\) from the audio modality, and (2) caption descriptions \\( V_{\\text{caption}} = (d_{1}, d_{2}, \\dots, d_{K}) \\) from the visual modality. In detail, for speech transcriptions, we use ASR outputs provided by [109], obtained using the Whisper-Large-V2 [70] model through the WhisperX [6] implementation. For captioning, we employ MiniCPM-V [112] as an image captioner, applied independently on the selected video frames, i.e., \\( d_{i} = \\text{Captioner}(v_{i}) \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.66, + 0.485, + 0.719 + ], + "angle": 0, + "content": "As we aim at predicting relevant chapter boundaries, we provide temporal information to the LLM. For both modalities, we prepend the timestamp information formatted as \"HH:MM:SS\" to encode the location at which the speech or caption is obtained." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.722, + 0.484, + 0.839 + ], + "angle": 0, + "content": "Captions naturally come from a single point in time. Speech segments cover intervals, but their duration is typically very short (3-4 seconds). We therefore simply use the start time of each transcribed speech interval. We interleave the speech and caption inputs based on their timestamps in a sorted order. We add a modality-specific prefix to each timestamp to denote which modality the information is extracted from (i.e., ASR for speech transcripts, Caption for captions)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.842, + 0.484, + 0.902 + ], + "angle": 0, + "content": "We prepend the text combining speech transcripts and captions with a fixed prompt that provides task instructions (see sup. mat. for the exact wording). This prompt occupies approximately 90 tokens and is independent of video length." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.377, + 0.909, + 0.656 + ], + "angle": 0, + "content": "Language model. We derive our framework by making use of a powerful pretrained LLM. Specifically, we employ the recent Llama-3.1-8B-Instruct [20] model and further finetune on chapter annotations using the LoRA technique [35]. Given the input structure previously described, the LLM is trained to output chapters, where each chapter consists of a timestamp in HH:MM:SS format followed by a free-form chapter title. We treat both the timestamps and titles simply as text tokens and apply the standard cross-entropy loss over the original vocabulary of the pretrained LLM. We apply teacher forcing during training and decode tokens autoregressively at inference. Note that the final model (taking both speech and captions as input) is trained independently from the speech-only version of our model used for frame selection, but these two models share the same backbone, and only differ in their LoRA parameters (13MB each). Across all experiments, we finetune models for a single epoch and use the same hyperparameters. We provide these hyperparameters, along with implementation details in Appendix A, and provide experiments with several Llama variants in Appendix C." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.667, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Iterative prediction for long videos. The inputs may exceed the context window limitation of the LLM, especially in the case of long videos. For example, on an A6000 GPU, the Llama-3.1-8B-Instruct [20] model can process videos up to around 15k tokens during training, which corresponds to 50 minutes of video content on average, and 25k tokens during inference, which corresponds to 80 minutes of video content on average. To address this issue, during training, we select videos that have less than 15k tokens. Since there are videos up to 1 hour long in the training set that satisfy this constraint, and since we do not need the entire training dataset to achieve good performance, this token limitation does not hinder our training. During evaluation, we predict chapters for each chunk sequentially, such that the start of a chunk is the end of the previous chunk. Finally, we merge the predictions from all chunks to obtain chapter boundaries for the complete video. We provide more details in Appendix A.4." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18950" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.218, + 0.108 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.116, + 0.485, + 0.204 + ], + "angle": 0, + "content": "In this section, we start by describing the data and evaluation metrics used in our experiments (Sec. 4.1). Next, we compare our results with the state of the art (Sec. 4.2), and then provide a series of ablations in our framework (Sec. 4.3). Finally, we investigate the impact of testing with very long videos exceeding our context window limitations (Sec. 4.4)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.216, + 0.283, + 0.231 + ], + "angle": 0, + "content": "4.1. Data and evaluation" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.24, + 0.485, + 0.621 + ], + "angle": 0, + "content": "Data. We train and evaluate on the recently released VidChapters-7M [109] dataset that includes user-annotated chaptered videos sourced from YouTube. Speech transcripts are obtained using Whisper [70] as the ASR method. In the original release, there is a total of 817k videos, spanning 8M chapters, with 2.4 minutes per chapter and 5.4 words per chapter title, totaling to 23 minutes and 8.3 chapters per video on average. Data is split into 801k training, 8.2k validation, and 8.2k test videos. To measure performance at different video lengths, we define three categories depending on video duration: 'short' (0-15min), 'medium' (15-30min), and 'long' (30-60min) videos. In this work, we use a subset of the training data as we observe increasing the training set brings diminishing returns at the cost of extended training times (see Fig. 4). Specifically, we use about 20k training videos (10k short videos used for the speech-based frame selection model and another 10k videos evenly split across short, medium and long durations for the final model. For state-of-the-art comparisons (Sec. 4.2), we employ the full official test set, which also contains videos without any speech (2.5% of the videos), and videos longer than 60 minutes (e.g., there are few videos that last about 12 hours). In ablations (Sec. 4.3), both for faster experimentation, and to limit the use of the test set during experimentation, we train on a randomly sampled subset of 1k videos (evenly split between short, medium, and long) and report results on a randomly sampled subset of 300 validation videos (100 from each duration) that have at least one speech utterance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.637, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Evaluation metrics. We primarily monitor temporal segmentation metrics to evaluate our chapter boundary detections. In particular, we employ tIoU and F1 scores. For tIoU (temporal Intersection over Union), we first compute the optimal matching between predicted and ground truth segments by greedily selecting pairs with highest IoU scores. The tIoU score is then calculated as the mean IoU across all matched pairs, multiplied by 100 to obtain a percentage. For F1 score, we first compute precision and recall at different IoU thresholds (ranging from 0.5 to 0.95 with a step of 0.05). At each threshold, a prediction is considered correct if it has IoU above the threshold with a ground truth segment. The precision is the ratio of correct predictions to total predictions, while recall is the ratio of matched ground truth segments to total ground truth segments. The F1 score is then computed as the harmonic mean of precision and recall. The final F1 metric is the average across all thresholds, multiplied by 100 to obtain a percentage. Note that [109] uses recall and precision metrics in two ways: (1)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.209 + ], + "angle": 0, + "content": "by considering timestamps within 3 or 5 second thresholds as matches, and (2) by considering segments with IoU above 0.5 or 0.7 as matches. While these metrics provide point estimates at specific thresholds, we find that IoU and F1 scores offer several advantages: they evaluate performance continuously across multiple thresholds, are more interpretable, and provide a more comprehensive evaluation of the model. For completeness, we also report the metrics used in [109] in Appendix C." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.21, + 0.907, + 0.268 + ], + "angle": 0, + "content": "For chapter title evaluation, we follow [109] and report SODA (S) [25] and CIDEr (C) [94], which measure the quality of the titles for the predicted segments that match to the ground segments (see [109] for details)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.278, + 0.829, + 0.294 + ], + "angle": 0, + "content": "4.2. Comparison with the state of the art" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.301, + 0.909, + 0.374 + ], + "angle": 0, + "content": "In Tab. 1, we report the performance of our model on the full VidChapters-7M test set [109] ('All' columns), and compare to the state of the art reported in [109], which uses Vid2Seq [110]. Moreover, we evaluate four proprietary models using our speech-based frame selection and captioning in a zero-shot manner." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.375, + 0.909, + 0.536 + ], + "angle": 0, + "content": "We observe that our finetuned Chapter-Llama achieves substantial performance improvements across all metrics and video duration categories. (e.g., 45.3 vs 26.7 F1 and 19.3 vs 11.6 SODA compared to Vid2Seq). Notably, our improvement over Vid2Seq is more important for medium and long videos compared to short videos. Note that our final approach was trained using the subset of data detailed in the previous section, specifically 20k videos, which constitutes only \\(2.5\\%\\) of the total available training data. In contrast, the baseline Vid2Seq model [110] was trained on a considerably larger dataset, utilizing both HowTo100M [57] and the entire VidChapters-7M training set." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.537, + 0.909, + 0.638 + ], + "angle": 0, + "content": "Additionally, we report performances of our model without training on any chapter annotations (i.e., both the speech-based frame selector and the LLM are not finetuned, and run with the same prompt as in the finetuned setting). We see that our zero-shot method also achieves competitive performance (e.g., 29.5 F1), whereas Vid2Seq only trained on HowTo100M does not generalize (3.0 F1)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.64, + 0.909, + 0.772 + ], + "angle": 0, + "content": "Finally, when zero-shot evaluating the proprietary models, GPT4-o [62] and Gemini variants [27], with our speech-based frame selection and captioning inputs, we observe competitive performances (e.g., 42.2 F1 with Gemini-1.5-Pro); however, our Chapter-Llama still surpasses on all metrics. Note that, due to API costs of the proprietary models, we performed their evaluation on a random \\(10\\%\\) subset of the test set; however, we verified that the scores are similar between \\(10\\%\\) and \\(100\\%\\) of the test set when evaluating with Chapter-Llama." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.784, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Qualitative comparison. In Fig. 3, we provide qualitative examples comparing our method against Vid2Seq [109, 110] and our zero-shot baseline. Our predictions align well with the ground truth chapters, accurately capturing both the temporal boundaries and generating relevant titles. In contrast, Vid2Seq segments tend to be less accurate, and we also observe that it often produces repetitive titles (bottom example). The zero-shot Chapter-Llama baseline tends to generate relatively longer and" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "18951" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.116, + 0.899, + 0.26 + ], + "angle": 0, + "content": "
BackboneFrame selectionFt.ShortMediumLongAll
F1tIoUSCF1tIoUSCF1tIoUSCF1tIoUSC
GPT-4o-mini [62]†OursX32.164.57.242.430.562.36.130.628.061.06.027.331.263.66.837.8
GPT-4o [62]†OursX37.768.08.453.838.168.88.151.436.566.26.634.837.668.08.151.0
Gemini-2.0-Flash [27]†OursX39.969.212.072.843.871.411.270.334.966.29.051.640.269.311.469.7
Gemini-1.5-Pro [27]†OursX41.770.611.765.343.871.811.261.441.370.610.155.342.270.911.463.2
Vid2Seq [109, 110]EquidistantX2.528.60.30.33.229.70.30.44.632.00.30.53.029.30.30.4
Llama 3.1-8BOursX29.963.47.134.530.662.75.428.126.659.33.618.929.562.56.230.7
Vid2Seq [109, 110]Equidistant✓33.463.715.274.919.053.37.531.916.750.85.928.426.758.611.655.8
Llama 3.1-8B (Chapter-Llama)Ours✓45.572.220.2103.546.772.318.898.741.369.215.891.245.371.819.3100.9
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.09, + 0.263, + 0.907, + 0.373 + ], + "angle": 0, + "content": "Table 1. Comparison to the state of the art on VidChapters-7M test set: We split the table into (bottom) the comparison between Chapter-Llama and the state-of-the-art method Vid2Seq [110], and (top) the evaluation of proprietary models. Chapter-Llama significantly outperforms Vid2Seq trained and reported by [109] (45.3 vs 26.7 F1). Our method also achieves strong performance in zero-shot mode - without finetuning (Ft.) on any chapter annotation (29.5 F1). Furthermore, we report performance of proprietary models in such zero-shot setting, using our speech-based frame selection and captioning, and observe inferior results than Chapter-Llama (42.2 F1 with Gemini-1.5-Pro). Note that we use the full official 8.1k test set videos ('All'), unlike in the remaining experiments that report on the validation subset. We also report the performance breakdown into short (4891), medium (1736), and long (892) test videos. Our model was trained on 10k videos balanced across short, medium and long durations. \\(\\dagger\\) denotes evaluation on a random \\(10\\%\\) subset of the test set due to API costs of proprietary models." + }, + { + "type": "image", + "bbox": [ + 0.114, + 0.42, + 0.88, + 0.832 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.837, + 0.907, + 0.879 + ], + "angle": 0, + "content": "Figure 3. Qualitative results: We display two examples and compare our Chapter-Llama results against the ground truth (GT), as well as the zero-shot (ZS) and Vid2Seq (VS) baselines. For each example, we show the corresponding SODA (S) and CIDEr (C) scores. Our method overall shows the highest similarity with the GT, while Vid2Seq can suffer from repeated chapter titles, and zero-shot generations tend to over-segment." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18952" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.114, + 0.089, + 0.457, + 0.221 + ], + "angle": 0, + "content": "
ModalitiesSegmentationTitles
SpeechCaptionsF1tIoUSC
Zero-shotX✓12.648.61.96.4
✓X22.757.34.419.7
✓✓29.963.06.933.7
FinetunedX✓39.167.75.920.2
✓X38.568.113.967.3
✓✓42.670.616.482.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.226, + 0.485, + 0.307 + ], + "angle": 0, + "content": "Table 2. Contribution of different modalities and finetuning: Finetuning the LLM with 1k videos largely improves chaptering performance on 300 validation videos, see bottom block vs top block. In the finetuned setting, we further demonstrate the advantages of combining both modalities, i.e., transcribed speech from ASR and automatic captions extracted from video frames." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.321, + 0.484, + 0.381 + ], + "angle": 0, + "content": "verbose chapter titles and often generates chapters that appear to be continuations of previous chapters rather than distinct segments, while also exhibiting over-segmentation issues. We provide more examples in Appendix E." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.388, + 0.253, + 0.403 + ], + "angle": 0, + "content": "4.3. Ablation studies" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.41, + 0.484, + 0.5 + ], + "angle": 0, + "content": "In the following, we experiment with (i) the contribution of speech and caption modalities, along with the effect of LLM finetuning, (ii) the effect of our frame selection method for captioning, (iii) the amount of training data, and (iv) the use of frame embeddings instead of captions. As mentioned above, we use 1k training and 300 validation videos for these ablations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.51, + 0.484, + 0.627 + ], + "angle": 0, + "content": "Modalities and LLM finetuning. In Tab. 2, we ablate the impact of finetuning the LLM and the contribution of each of the speech and caption modalities. In the top block, we run our baselines in zero-shot setting as introduced in the previous section. The speech-only baseline outperforms the captions-only baseline by a large margin in the zero-shot setting. This suggests that speech contains more relevant information for chaptering, as was previously observed by [109]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.627, + 0.484, + 0.759 + ], + "angle": 0, + "content": "As shown in the bottom block of Tab. 2, we observe large performance improvements when finetuning the LLM, as opposed to zero-shot. We hypothesize that zero-shot prompting with a long multi-modal text, potentially containing redundant and irrelevant information, may overwhelm the LLM [79, 101]. We obtain our best model by combining the two modalities, which performs better than the individual speech-only or caption-only models. This demonstrates the multi-modal capabilities of our model." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.769, + 0.483, + 0.872 + ], + "angle": 0, + "content": "Speech-based frame selection. In Tab. 3, we examine a number of strategies to sample frames at which we extract captions. In addition to previously described metrics, for each of the frame sampling approaches, we report the average number of captions per video and the average number of text tokens per minute. For reference, we also report an off-the-shelf shot detection [11] and Vid2Seq [109, 110]." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.872, + 0.484, + 0.902 + ], + "angle": 0, + "content": "We compare our speech-based frame selection strategy to various baselines. We experiment with sampling (i) uniformly" + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.09, + 0.905, + 0.304 + ], + "angle": 0, + "content": "
MethodFrame selection for captionsaverage #frames#tokens per min.SegmentationTitles SC
F1tIoU
BASELINES
Shot detection [11]n/a49.4n/a6.237.6--
Vid2Seq [109, 110]100 equidistant100.0128.625.457.811.255.0
CHAPTER-LLAMA VARIANTS
SpeechCaption
✓✘n/an/a248.638.568.113.967.3
100 equidistant100.0449.121.053.88.436.0
Every 10 sec.83.1280.312.845.94.313.0
✘✓Shot boundaries49.4193.216.250.73.912.4
10 equidistant10.041.811.046.43.69.0
Speech-based10.336.239.167.75.920.2
100 equidistant100.0746.239.267.416.183.8
Every 10 sec.83.1570.141.069.315.477.3
✓✓Shot boundaries40.4481.740.669.115.879.3
10 equidistant10.0326.140.167.915.877.5
Speech-based10.3320.442.670.616.482.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.309, + 0.908, + 0.498 + ], + "angle": 0, + "content": "Table 3. Frame selection strategies for captioning: We evaluate different approaches for selecting frames to extract captions from, comparing our speech-based selection method against baselines. The table shows results for models trained on 1k videos and evaluated on 300 validation videos. We experiment with using speech only, captions only, and both modalities (bottom section). For caption extraction, we compare our speech-based approach to other alternatives such as equidistant sampling (100 or 10 frames), uniformly sampling every 10 seconds, or sampling at shot boundaries using [11]. Our speech-based frame selection achieves the best overall performance (F1: 42.6, tIoU: 70.6) while requiring significantly fewer number of frames on average (10.3) compared to other sampling approaches. The tokens-per-minute statistic shows the total input length including both speech transcriptions and captions, excluding the fixed prompt template." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.515, + 0.907, + 0.633 + ], + "angle": 0, + "content": "100 frames as in Vid2Seq, (ii) every 10 seconds, (iii) at shot boundaries detected by an off-the-shelf shot detector [11], (iv) 10 equidistant frames to be similar to our speech-based locations (i.e., 10.0 vs 10.3 number of frames on average), and (v) sampling at frames predicted as chapter boundaries by our LLM that inputs only speech. In all cases, we limit the maximum number of frames to 100 as in [109, 110] to handle extreme durations." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.634, + 0.907, + 0.737 + ], + "angle": 0, + "content": "In both caption-only and caption+speech settings, our speech-based frame selection approach achieves better segmentation results than the more frame-expensive baselines '100 equidistant', 'every 10 sec', and 'shot boundaries', while using much less frames, and also improves over the '10 equidistant' baseline which uses a similar number of frames. This demonstrates the effectiveness of our speech-based frame selection strategy." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.738, + 0.907, + 0.797 + ], + "angle": 0, + "content": "For reference, we also report positive comparison against shot detection and Vid2Seq [109, 110]. Note Vid2Seq has less tokens per min. compared to our 100 equidistant variants, because Vid2Seq uses a different timestamp tokenizer in the input." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.813, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Amount of training data. Given the large-scale nature of the VidChapters-7M training set, we investigate how much chapter data is needed for LoRA finetuning the LLM. We plot the performance against the number of training videos in Fig. 4. We start by the zero-shot baseline as the first data point, and report our method with 1k, 5k, 7k, and 10k videos, split evenly" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "18953" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.091, + 0.088, + 0.48, + 0.218 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.221, + 0.483, + 0.288 + ], + "angle": 0, + "content": "Figure 4. Amount of training data: Our experiments show a substantial improvement when moving from zero-shot to training with 1k videos. Beyond 1k videos, performance continues to improve but at a much slower rate, motivating our choice of using only 10k training videos for our final LLM." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.307, + 0.484, + 0.41 + ], + "angle": 0, + "content": "between three durations. We see that after increasing above several thousand training videos starts to bring diminishing returns. We therefore keep 10k training videos for our final LLM, which makes our approach highly efficient to train (40min on 4 H100 GPUs). Note that here we focus on the chaptering LLM and always use frame sampling locations from a speech-based module trained on 10k separate videos." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.426, + 0.485, + 0.72 + ], + "angle": 0, + "content": "Frame embeddings vs captions. In Tab. 4, we investigate whether raw visual embeddings could serve as an alternative to textual captions. To this end, we experiment with replacing the captions with frame embeddings. Specifically, for each frame, we extract the 1152-dimensional output embedding corresponding to the [CLS] token from a frozen SigLIP model [119], and feed through a 2-layer MLP mapping network. We initialize the MLP weights from MANTIS [41] and train jointly with the LLM during finetuning. The results with 'Speech+Embeddings' are better than 'Speech' alone (38.5 vs 40.4 F1), but worse than 'Speech+Captions' (42.6 vs 40.4 F1). The performance gap between 'Speech+Embeddings' and 'Speech+Captions' may be due to the richer information provided by captions, which use multiple tokens per frame, directly in text form, compared to the single [CLS] token frame embedding, requiring a mapping network to be ingested by an LLM. Finally, while combining all modalities achieves the best performance (44.4 F1), we exclude frame embeddings from our final model due to practical considerations, e.g., they add complexity, increase processing time by 2.5x, and require 3000x more storage space." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.731, + 0.406, + 0.746 + ], + "angle": 0, + "content": "4.4. Iterative prediction on longer videos" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.754, + 0.483, + 0.901 + ], + "angle": 0, + "content": "In our ablation studies, our experimental setting considered training and evaluating with videos that fit within the LLM context window. In Tab. 5, we evaluate the benefit of our iterative prediction procedure for handling videos that exceed the LLM context window. For this, we identify videos in the validation set whose inputs exceed the LLM inference context window (\\(>35k\\) tokens), resulting in 110 videos. On this challenging subset, we find that our iterative prediction procedure improves chaptering results compared to the baseline that only runs the LLM once by cropping the input to the first" + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.089, + 0.898, + 0.227 + ], + "angle": 0, + "content": "
ModalitiesSegmentationTitles
SpeechEmbeddingsCaptionsF1tIoUSC
✓--38.568.113.967.3
-✓-38.466.53.47.3
--✓39.167.75.920.2
✓✓-40.468.215.374.9
✓-✓42.670.616.482.4
✓✓✓44.471.516.384.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.231, + 0.905, + 0.326 + ], + "angle": 0, + "content": "Table 4. Frame embeddings vs captions: We compare using frame captions versus visual features from a frozen SigLIP model projected through a learned 2-layer MLP mapping network ('Embeddings'). While the 'Speech+Embeddings' combination performs better than speech alone (40.4 vs 38.5 F1), it underperforms compared to the 'Speech+Captions' combination (42.6 vs 40.4 F1). All models are trained with 1k videos and evaluated on 300 videos." + }, + { + "type": "table", + "bbox": [ + 0.536, + 0.335, + 0.882, + 0.464 + ], + "angle": 0, + "content": "
Window# tok.avg # iter.Subset exceeding 35k tokens
F1tIoUSC
First10k113.150.54.031.2
15k116.654.95.443.3
20k118.756.76.647.5
Iterative10k8.518.557.16.925.1
15k5.423.660.18.735.2
20k4.125.361.410.344.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.468, + 0.906, + 0.575 + ], + "angle": 0, + "content": "Table 5. Iterative prediction: Our iterative prediction procedure improves chaptering results on the subset of 110 videos which exceed 35k tokens compared to the baseline that only runs the LLM once (by only taking the first window, and discarding the rest of the input sequence), across various context windows. As we increase the context window in the iterative prediction, the performance gradually improves and the average number of iterations decreases. The model is trained with 1k videos." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.594, + 0.905, + 0.639 + ], + "angle": 0, + "content": "input window, across various context windows (10k, 15k, and 20k). We refer to Appendix B for details on the video lengths and statistics of videos that exceed the LLM context window." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.655, + 0.634, + 0.67 + ], + "angle": 0, + "content": "5. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.681, + 0.907, + 0.901 + ], + "angle": 0, + "content": "We presented Chapter-Llama, an approach that leverages LLMs for hour-long video chaptering by mapping video to text using speech transcripts and efficiently captioning video frames sampled with a speech-based frame selector. Our results on VidChapters-7M consequently improved the state of the art by a large margin. We experimentally demonstrated the benefits of our components through an extensive ablation study. One limitation of our approach is that it relies on the accuracy of the ASR and the visual captioner. Future work can explore hierarchical chaptering with several granularities and consider the audio modality beyond speech. We also note that the LLM, the visual captioner, and speech transcription models are trained on large Web datasets, which can contain biases that can lead to inaccurate chaptering, especially for videos depicting underrepresented topics." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18954" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.268 + ], + "angle": 0, + "content": "Acknowledgements. This work was granted access to the HPC resources of IDRIS under the allocation 2024-AD011014696 made by GENCI. This work was funded in part by the ANR project CorVis ANR-21-CE23-0003-01, a research gift from Google, the French government under management of Agence Nationale de la Recherche as part of the \"France 2030\" program, reference ANR-23-IACL-0008 (PR[AI]RIE-PSAI projet), and the ANR project VideoPredict ANR-21-FAI1-0002-01. Cordelia Schmid would like to acknowledge the support by the Körber European Science Prize. The authors would also like to thank Guillaume Astruc, Nikos Athanasiou, Hyolim Kang, and Nicolas Dufour for their feedback." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.284, + 0.184, + 0.3 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.107, + 0.308, + 0.482, + 0.376 + ], + "angle": 0, + "content": "[1] Soheyla Amirian, Khaled Rasheed, Thiab R Taha, and Hamid R Arabnia. Automatic generation of descriptive titles for video clips using deep learning. In Advances in Artificial Intelligence and Applied Cognitive Computing: Proceedings from ICAI'20 and ACC'20, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.107, + 0.379, + 0.482, + 0.432 + ], + "angle": 0, + "content": "[2] D. Argaw, S. Yoon, F. Heilbron, H. Deilamsalehy, T. Bui, Z. Wang, F. Dernoncourt, and J. Chung. Scaling up video summarization pretraining with large language models. In CVPR, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.107, + 0.435, + 0.482, + 0.474 + ], + "angle": 0, + "content": "[3] Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lucic, and Cordelia Schmid. ViViT: A video vision transformer. In ICCV, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.107, + 0.477, + 0.482, + 0.517 + ], + "angle": 0, + "content": "[4] Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio Neural machine translation by jointly learning to align and translate. ICLR, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.107, + 0.519, + 0.482, + 0.559 + ], + "angle": 0, + "content": "[5] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In ICCV, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.107, + 0.562, + 0.482, + 0.602 + ], + "angle": 0, + "content": "[6] Max Bain, Jaesung Huh, Tengda Han, and Andrew Zisserman. WhisperX: Time-accurate speech transcription of long-form audio. In Interspeech, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.107, + 0.605, + 0.482, + 0.658 + ], + "angle": 0, + "content": "[7] Nadine Behrmann, S Alireza Golestaneh, Zico Kolter, Jürgen Gall, and Mehdi Noroozi. Unified fully and timestamp supervised temporal action segmentation via sequence to sequence translation. In ECCV, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.107, + 0.661, + 0.482, + 0.701 + ], + "angle": 0, + "content": "[8] Gedas Bertasius, Heng Wang, and Lorenzo Torresani. Is space-time attention all you need for video understanding? In ICML, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.107, + 0.704, + 0.482, + 0.73 + ], + "angle": 0, + "content": "[9] Tom Brown et al. Language models are few-shot learners. In NeurIPS, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.733, + 0.482, + 0.759 + ], + "angle": 0, + "content": "[10] Xiao Cao, Zitan Chen, Canyu Le, and Lei Meng. Multi-modal video chapter generation. In BMVC, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.762, + 0.485, + 0.801 + ], + "angle": 0, + "content": "[11] Brandon Castellano. Pyscenedetect: Intelligent scene cut detection and video splitting tool. https://pyscenedetect.readthedocs.io/en/latest/, 2018.7" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.805, + 0.482, + 0.857 + ], + "angle": 0, + "content": "[12] Houlun Chen, Xin Wang, Hong Chen, Zihan Song, Jia Jia, and Wenwu Zhu. Grounding-prompter: Prompting llm with multimodal information for temporal sentence grounding in long videos. arXiv:2312.17117, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.861, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[13] Shixing Chen, Xiaohan Nie, David Fan, Dongqing Zhang, Vimal Bhat, and Raffay Hamid. Shot contrastive self-supervised learning for scene boundary detection. In CVPR, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.1, + 0.308, + 0.485, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.093, + 0.905, + 0.133 + ], + "angle": 0, + "content": "[14] Shixing Chen, Chun-Hao Liu, Xiang Hao, Xiaohan Nie, Maxim Arap, and Raffay Hamid. Movies2Scenes: Using movie metadata to learn scene representation. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.134, + 0.907, + 0.186 + ], + "angle": 0, + "content": "[15] Sihan Chen, Xingjian He, Longteng Guo, Xinxin Zhu, Weining Wang, Jinhui Tang, and Jing Liu. VALOR: Vision-audiolanguage omni-perception pretraining model and dataset. TPAMI, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.188, + 0.907, + 0.24 + ], + "angle": 0, + "content": "[16] Sihan Chen, Handong Li, Qunbo Wang, Zijia Zhao, Mingzhen Sun, Xinxin Zhu, and Jing Liu. VAST: A vision-audio-subtitle-text omni-modality foundation model and dataset. NeurIPS, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.242, + 0.905, + 0.281 + ], + "angle": 0, + "content": "[17] Yangyu Chen, Shuhui Wang, Weigang Zhang, and Qingming Huang. Less is more: Picking informative frames for video captioning. In ECCV, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.282, + 0.905, + 0.32 + ], + "angle": 0, + "content": "[18] Feng Cheng and Gedas Bertasius. TALLformer: Temporal action localization with long-memory transformer. In ECCV, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.322, + 0.905, + 0.362 + ], + "angle": 0, + "content": "[19] Xu Cheng, Cameron Dale, and Jiangchuan Liu. Statistics and social network of youtube videos. In 16th International Workshop on Quality of Service, 2008. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.363, + 0.905, + 0.415 + ], + "angle": 0, + "content": "[20] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The Llama 3 herd of models. arXiv:2407.21783, 2024. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.417, + 0.905, + 0.456 + ], + "angle": 0, + "content": "[21] Yue Fan, Xiaojian Ma, Rujie Wu, Yuntao Du, Jiaqi Li, Zhi Gao, and Qing Li. VideoAgent: A memory-augmented multimodal agent for video understanding. In ECCV, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.457, + 0.907, + 0.509 + ], + "angle": 0, + "content": "[22] Xinyu Fang, Kangrui Mao, Haodong Duan, Xiangyu Zhao, Yining Li, Dahua Lin, and Kai Chen. MMBench-video: A long-form multi-shot benchmark for holistic video understanding. In NeurIPS Datasets and Benchmarks, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.511, + 0.905, + 0.55 + ], + "angle": 0, + "content": "[23] Yazan Abu Farha and Jurgen Gall. MS-TCN: Multi-stage temporal convolutional network for action segmentation. In CVPR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.551, + 0.905, + 0.617 + ], + "angle": 0, + "content": "[24] Chaoyou Fu, Yuhan Dai, Yondong Luo, Lei Li, Shuhuai Ren, Renrui Zhang, Zihan Wang, Chenyu Zhou, Yunhang Shen, Mengdan Zhang, et al. Video-MME: The first-ever comprehensive evaluation benchmark of multi-modal LLMs in video analysis. arXiv:2405.21075, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.619, + 0.905, + 0.658 + ], + "angle": 0, + "content": "[25] Soichiro Fujita, Tsutomu Hirao, Hidetaka Kamigaito, Manabu Okumura, and Masaaki Nagata. SODA: Story oriented dense video captioning evaluation framework. In ECCV, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.66, + 0.905, + 0.698 + ], + "angle": 0, + "content": "[26] Shang-Hua Gao, Qi Han, Zhong-Yu Li, Pai Peng, Liang Wang, and Ming-Ming Cheng. Global2Local: Efficient structure search for video action segmentation. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.7, + 0.905, + 0.725 + ], + "angle": 0, + "content": "[27] Gemini Team et al. Gemini: A family of highly capable multimodal models. arXiv:2312.11805, 2024. 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.727, + 0.907, + 0.833 + ], + "angle": 0, + "content": "[28] Azin Ghazimatin, Ekaterina Garmash, Gustavo Penha, Kristen Sheets, Martin Achenbach, Oguz Semerci, Remi Galvez, Marcus Tannenberg, Sahitya Mantravadi, Divya Narayanan, Ofeliya Kalaydzhyan, Douglas Cole, Ben Carterette, Ann Clifton, Paul N. Bennett, Claudia Hauff, and Mounia Lalmas. PODTILE: Facilitating podcast episode browsing with auto-generated chapters. In Proceedings of the 33rd ACM International Conference on Information and Knowledge Management, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.835, + 0.905, + 0.859 + ], + "angle": 0, + "content": "[29] Kristen Grauman et al. Ego4D: Around the world in 3,000 hours of egocentric video. In CVPR, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.861, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[30] Khalil Guetari, Yannis Tevisen, and Frédéric Petitpont. Multimodal chaptering for long-form TV newscast video. arXiv:2406.17590, 2024. 3" + }, + { + "type": "list", + "bbox": [ + 0.523, + 0.093, + 0.907, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18955" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.093, + 0.486, + 0.133 + ], + "angle": 0, + "content": "[31] Tengda Han, Max Bain, Arsha Nagrani, Gül Varol, Weidi Xie, and Andrew Zisserman. AutoAD: Movie description in context. In CVPR, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.134, + 0.483, + 0.173 + ], + "angle": 0, + "content": "[32] Tanveer Hannan, Md Mohaiminul Islam, Thomas Seidl, and Gedas Bertasius. RGNet: A unified retrieval and grounding network for long videos. In ECCV, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.175, + 0.483, + 0.213 + ], + "angle": 0, + "content": "[33] Bo He, Jun Wang, Jielin Qiu, Trung Bui, Abhinav Shrivastava, and Zhaowen Wang. Align and attend: Multimodal summarization with dual contrastive losses. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.215, + 0.483, + 0.267 + ], + "angle": 0, + "content": "[34] Chiori Hori, Takaaki Hori, Teng-Yok Lee, Ziming Zhang, Bret Harsham, John R Hershey, Tim K Marks, and Kazuhiko Sumi. Attention-based multimodal fusion for video description. In ICCV, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.269, + 0.483, + 0.309 + ], + "angle": 0, + "content": "[35] Edward J Hu, yelong shen, Phillip Wallis, Zeyuan Allen-Zhu, Yanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In ICLR, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.31, + 0.483, + 0.36 + ], + "angle": 0, + "content": "[36] De-An Huang, Shijia Liao, Subhashree Radhakrishnan, Hongxu Yin, Pavlo Molchanov, Zhiding Yu, and Jan Kautz. Lita: Language instructed temporal-localization assistant. In ECCV, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.362, + 0.482, + 0.402 + ], + "angle": 0, + "content": "[37] Gabriel Huang, Bo Pang, Zhenhai Zhu, Clara Rivera, and Radu Soricut. Multimodal pretraining for dense video captioning. In AACL-IJCNLP, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.403, + 0.482, + 0.442 + ], + "angle": 0, + "content": "[38] Qingqiu Huang, Yu Xiong, Anyi Rao, Jiaze Wang, and Dahua Lin. Movienet: A holistic dataset for movie understanding. In ECCV, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.444, + 0.483, + 0.495 + ], + "angle": 0, + "content": "[39] Md Mohaiminul Islam, Mahmudul Hasan, Kishan Shamsundar Athrey, Tony Braskich, and Gedas Bertasius. Efficient movie scene detection using state-space transformers. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.497, + 0.483, + 0.537 + ], + "angle": 0, + "content": "[40] Md Mohaiminul Islam, Ngan Ho, Xitong Yang, Tushar Nagarajan, Lorenzo Torresani, and Gedas Bertasius. Video ReCap: Recursive captioning of hour-long videos. CVPR, 2024. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.538, + 0.483, + 0.577 + ], + "angle": 0, + "content": "[41] Dongfu Jiang, Xuan He, Huaye Zeng, Cong Wei, Max Ku, Qian Liu, and Wenhu Chen. Mantis: Interleaved multi-image instruction tuning. TMLR, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.579, + 0.483, + 0.618 + ], + "angle": 0, + "content": "[42] Kumara Kahatapitiya, Kanchana Ranasinghe, Jongwoo Park, and Michael S Ryoo. Language repository for long video understanding. arXiv:2403.14622, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.62, + 0.483, + 0.657 + ], + "angle": 0, + "content": "[43] Bruno Korbar, Du Tran, and Lorenzo Torresani. SCSampler: Sampling salient clips from video for efficient action recognition. In ICCV, 2019. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.66, + 0.483, + 0.698 + ], + "angle": 0, + "content": "[44] Ranjay Krishna, Kenji Hata, Frederic Ren, Li Fei-Fei, and Juan Carlos Niebles. Dense-captioning events in videos. In ICCV, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.7, + 0.483, + 0.739 + ], + "angle": 0, + "content": "[45] Jie Lei, Linjie Li, Luowei Zhou, Zhe Gan, Tamara L Berg, Mohit Bansal, and Jingjing Liu. Less is more: ClipBERT for video-and-language learning via sparse sampling. In CVPR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.741, + 0.483, + 0.792 + ], + "angle": 0, + "content": "[46] Feng Li, Jae Won Chung, and Mark Claypool. Three-year trends in YouTube video content and encoding. Proceedings of the 18th International Conference on Signal Processing and Multimedia Applications (SIGMAP), 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.794, + 0.483, + 0.833 + ], + "angle": 0, + "content": "[47] Mingzhe Li, Mark Claypool, Robert Kinicki, and James Nichols. Characteristics of streaming media stored on the web. ACM Trans. Internet Technol., 2005. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.835, + 0.483, + 0.86 + ], + "angle": 0, + "content": "[48] Zhe Li, Yazan Abu Farha, and Jurgen Gall. Temporal action segmentation from timestamp supervision. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.862, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[49] Bin Lin, Bin Zhu, Yang Ye, Munan Ning, Peng Jin, and Li Yuan. Video-LLaVA: Learning united visual representation by alignment before projection. arXiv:2311.10122, 2023. 3" + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.093, + 0.486, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.093, + 0.906, + 0.146 + ], + "angle": 0, + "content": "[50] Kevin Lin, Linjie Li, Chung-Ching Lin, Faisal Ahmed, Zhe Gan, Zicheng Liu, Yumao Lu, and Lijuan Wang. SwinBERT: End-to-end transformers with sparse attention for video captioning. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.149, + 0.906, + 0.203 + ], + "angle": 0, + "content": "[51] Kevin Qinghong Lin, Pengchuan Zhang, Joya Chen, Shraman Pramanick, Difei Gao, Alex Jinpeng Wang, Rui Yan, and Mike Zheng Shou. UniVTG: Towards unified video-language temporal grounding. In ICCV, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.205, + 0.907, + 0.243 + ], + "angle": 0, + "content": "[52] Hao Liu, Matei Zaharia, and Pieter Abbeel. Ring attention with blockwise transformers for near-infinite context. arXiv:2310.01889, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.246, + 0.906, + 0.286 + ], + "angle": 0, + "content": "[53] Hao Liu, Wilson Yan, Matei Zaharia, and Pieter Abbeel. World model on million-length video and language with blockwise ringattention. arXiv:2402.08268, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.288, + 0.907, + 0.342 + ], + "angle": 0, + "content": "[54] Xiaolong Liu, Qimeng Wang, Yao Hu, Xu Tang, Shiwei Zhang, Song Bai, and Xiang Bai. End-to-end temporal action detection with transformer. In IEEE Transactions on Image Processing, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.344, + 0.906, + 0.398 + ], + "angle": 0, + "content": "[55] Huaishao Luo, Lei Ji, Botian Shi, Haoyang Huang, Nan Duan, Tianrui Li, Jason Li, Taroon Bharti, and Ming Zhou. UniVL: A unified video and language pre-training model for multimodal understanding and generation. arXiv:2002.06353, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.4, + 0.906, + 0.453 + ], + "angle": 0, + "content": "[56] Karttikeya Mangalam, Raiymbek Akshulakov, and Jitendra Malik. EgoSchema: A diagnostic benchmark for very long-form video language understanding. In NeurIPS Track on Datasets and Benchmarks, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.455, + 0.906, + 0.508 + ], + "angle": 0, + "content": "[57] Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic. HowTo100M: Learning a text-video embedding by watching hundred million narrated video clips. In ICCV, 2019. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.511, + 0.906, + 0.551 + ], + "angle": 0, + "content": "[58] Juhong Min, Shyamal Buch, Arsha Nagrani, Minsu Cho, and Cordelia Schmid. MoReVQA: Exploring modular reasoning models for video question answering. In CVPR, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.553, + 0.906, + 0.606 + ], + "angle": 0, + "content": "[59] Jonghwan Mun, Minchul Shin, Gunsoo Han, Sangho Lee, Seongsu Ha, Joonseok Lee, and Eun-Sol Kim. BaSSL: Boundary-aware self-supervised learning for video scene segmentation. In ACCV, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.608, + 0.906, + 0.648 + ], + "angle": 0, + "content": "[60] Arsha Nagrani, Paul Hongsuck Seo, Bryan Seybold, Anja Hauth, Santiago Manen, Chen Sun, and Cordelia Schmid. Learning audio-video modalities from image captions. In ECCV, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.65, + 0.906, + 0.703 + ], + "angle": 0, + "content": "[61] Joe Yue-Hei Ng, Matthew Hausknecht, Sudheendra Vijayanarasimhan, Oriol Vinyals, Rajat Monga, and George Toderici. Beyond short snippets: Deep networks for video classification. In CVPR, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.706, + 0.895, + 0.719 + ], + "angle": 0, + "content": "[62] OpenAI. Gpt-4o system card. arXiv:2410.21276, 2024. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.722, + 0.906, + 0.748 + ], + "angle": 0, + "content": "[63] Yingwei Pan, Ting Yao, Houqiang Li, and Tao Mei. Video captioning with transferred semantic attributes. In CVPR, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.75, + 0.906, + 0.815 + ], + "angle": 0, + "content": "[64] Pinelopi Papalampidi, Skanda Koppula, Shreya Pathak, Justin Chiu, Joe Heyward, Viorica Patraucean, Jiajun Shen, Antoine Miech, Andrew Zisserman, and Aida Nematzdeh. A simple recipe for contrastively pre-training video-first encoders beyond 16 frames. In CVPR, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.818, + 0.906, + 0.858 + ], + "angle": 0, + "content": "[65] Alejandro Pardo, Fabian Caba Heilbron, Juan León Alcázar, Ali Thabet, and Bernard Ghanem. Moviecuts: A new dataset and benchmark for cut type recognition. In ECCV, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.861, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[66] Jinwoo Park, Jungeun Kim, Jaegwang Seok, Sukhyun Lee, and Junyeong Kim. Contrasting multi-modal similarity framework for video scene segmentation. IEEE Access, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.524, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18956" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.093, + 0.482, + 0.133 + ], + "angle": 0, + "content": "[67] Wenjie Pei, Jiyuan Zhang, Xiangrong Wang, Lei Ke, Xiaoyong Shen, and Yu-Wing Tai. Memory-attended recurrent network for video captioning. In CVPR, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.134, + 0.482, + 0.174 + ], + "angle": 0, + "content": "[68] Alec Radford, Karthik Narasimhan, Tim Salimans, and Ilya Sutskever. Improving language understanding by generative pre-training. Preprint, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.176, + 0.483, + 0.228 + ], + "angle": 0, + "content": "[69] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.23, + 0.483, + 0.27 + ], + "angle": 0, + "content": "[70] Alec Radford, Jong Wook Kim, Tao Xu, Greg Brockman, Christine McLeavey, and Ilya Sutskever. Robust speech recognition via large-scale weak supervision. In ICML, 2023. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.271, + 0.482, + 0.311 + ], + "angle": 0, + "content": "[71] Anyi Rao, Linning Xu, Yu Xiong, Guodong Xu, Qingqiu Huang, Bolei Zhou, and Dahua Lin. A local-to-global approach to multi-modal movie scene segmentation. In CVPR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.312, + 0.482, + 0.339 + ], + "angle": 0, + "content": "[72] Zeeshan Rasheed and Mubarak Shah. Scene detection in hollywood movies and tv shows. In CVPR, 2003. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.341, + 0.483, + 0.38 + ], + "angle": 0, + "content": "[73] Fabian Retkowski and Alexander Waibel. From text segmentation to smart chaptering: A novel benchmark for structuring video transcriptions. arXiv:2402.17633, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.381, + 0.482, + 0.422 + ], + "angle": 0, + "content": "[74] Marcus Rohrbach, Wei Qiu, Ivan Titov, Stefan Thater, Manfred Pinkal, and Bernt Schiele. Translating video content to natural language descriptions. In ICCV, 2013. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.423, + 0.482, + 0.476 + ], + "angle": 0, + "content": "[75] Daniel Rotman, Dror Porat, and Gal Ashour. Robust video scene detection using multimodal fusion of optimally grouped features. IEEE 19th International Workshop on Multimedia Signal Processing (MMSP), 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.478, + 0.482, + 0.518 + ], + "angle": 0, + "content": "[76] Yong Rui, Thomas S Huang, and Sharad Mehrotra. Exploring video structure beyond the shots. In IEEE International Conference on Multimedia Computing and Systems, 1998. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.519, + 0.483, + 0.572 + ], + "angle": 0, + "content": "[77] Najmeh Sadoughi, Xinyu Li, Avijit Vajpayee, David Fan, Bing Shuai, Hector Santos-Villalobos, Vimal Bhat, and Rohith MV. MEGA: Multimodal alignment aggregation and distillation for cinematic video segmentation. In ICCV, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.573, + 0.482, + 0.613 + ], + "angle": 0, + "content": "[78] Paul Hongsuck Seo, Arsha Nagrani, Anurag Arnab, and Cordelia Schmid. End-to-end generative pretraining for multimodal video captioning. In CVPR, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.614, + 0.483, + 0.667 + ], + "angle": 0, + "content": "[79] Freda Shi, Xinyun Chen, Kanishka Misra, Nathan Scales, David Dohan, Ed H. Chi, Nathanael Schärli, and Denny Zhou. Large language models can be easily distracted by irrelevant context. In ICML, 2023. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.668, + 0.483, + 0.722 + ], + "angle": 0, + "content": "[80] Nina Shvetsova, Anna Kukleva, Xudong Hong, Christian Rupprecht, Bernt Schiele, and Hilde Kuehne. HowToCaption: Prompting llms to transform video annotations at scale. In ECCV, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.724, + 0.483, + 0.776 + ], + "angle": 0, + "content": "[81] Panagiotis Sidiropoulos, Vasileios Mezaris, Ioannis Kompat-siaris, Hugo Meinedo, Miguel Bugalho, and Isabel Trancoso. Temporal video segmentation to scenes using high-level audiovisual features. IEEE TCSVT, 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.778, + 0.483, + 0.831 + ], + "angle": 0, + "content": "[82] Gunnar A. Sigurdsson, Gül Varol, Xiaolong Wang, Ivan Laptev, Ali Farhadi, and Abhinav Gupta. Hollywood in homes: Crowdsourcing data collection for activity understanding. In ECCV, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.833, + 0.483, + 0.872 + ], + "angle": 0, + "content": "[83] Karen Simonyan and Andrew Zisserman. Two-stream convolutional networks for action recognition in videos. In ICLR, 2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.873, + 0.483, + 0.9 + ], + "angle": 0, + "content": "[84] Mattia Soldan, Alejandro Pardo, Juan León Alcazar, Fabian Caba, Chen Zhao, Silvio Giancola, and Bernard Ghanem." + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.093, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.553, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "MAD: A scalable dataset for language grounding in videos from movie audio descriptions. In CVPR, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.122, + 0.905, + 0.162 + ], + "angle": 0, + "content": "[85] Chen Sun, Austin Myers, Carl Vondrick, Kevin Murphy, and Cordelia Schmid. VideoBERT: A joint model for video and language representation learning. In ICCV, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.164, + 0.905, + 0.204 + ], + "angle": 0, + "content": "[86] Yuchong Sun, Hongwei Xue, Ruihua Song, Bei Liu, Huan Yang, and Jianlong Fu. Long-form video-language pre-training with multimodal temporal contrastive learning. In NeurIPS, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.205, + 0.905, + 0.233 + ], + "angle": 0, + "content": "[87] Ilya Sutskever, Oriol Vinyals, and Quoc V Le. Sequence to sequence learning with neural networks. NeurIPS, 2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.234, + 0.905, + 0.287 + ], + "angle": 0, + "content": "[88] Reuben Tan, Xineng Sun, Ping Hu, Jui hsien Wang, Hanieh Deilamsalehy, Bryan A. Plummer, Bryan Russell, and Kate Saenko. Koala: Key frame-conditioned long video-LLM. In CVPR, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.289, + 0.907, + 0.355 + ], + "angle": 0, + "content": "[89] Gemini Team, Petko Georgiev, Ving Ian Lei, Ryan Burnell, Libin Bai, Anmol Gulati, Garrett Tanzer, Damien Vincent, Zhufeng Pan, Shibo Wang, et al. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context. arXiv:2403.05530, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.358, + 0.907, + 0.424 + ], + "angle": 0, + "content": "[90] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. LLaMA: Open and efficient foundation language models. arXiv:2302.13971, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.426, + 0.905, + 0.453 + ], + "angle": 0, + "content": "[91] Hugo Touvron et al. LLaMA 2: Open foundation and fine-tuned chat models. arXiv:2307.09288, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.455, + 0.905, + 0.495 + ], + "angle": 0, + "content": "[92] Du Tran, Lubomir Bourdev, Rob Fergus, Lorenzo Torresani, and Manohar Paluri. Learning spatiotemporal features with 3d convolutional networks. In ICCV, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.497, + 0.905, + 0.525 + ], + "angle": 0, + "content": "[93] Gül Varol, Ivan Laptev, and Cordelia Schmid. Long-term temporal convolutions for action recognition. TPAMI, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.527, + 0.905, + 0.566 + ], + "angle": 0, + "content": "[94] Ramakrishna Vedantam, C Lawrence Zitnick, and Devi Parikh CIDER: Consensus-based image description evaluation. In CVPR, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.568, + 0.905, + 0.595 + ], + "angle": 0, + "content": "[95] Lucas Ventura, Cordelia Schmid, and Gül Varol. Learning text-to-video retrieval from image captioning. *IJCV*, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.597, + 0.905, + 0.637 + ], + "angle": 0, + "content": "[96] Lucas Ventura, Antoine Yang, Cordelia Schmid, and Güll Varol. CoVR-2: Automatic data construction for composed video retrieval. TPAMI, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.639, + 0.905, + 0.706 + ], + "angle": 0, + "content": "[97] Caroline Violot, Tugrulcan Elmas, Igor Bilogrevic, and Mathias Humbert. Shorts vs. regular videos on YouTube: A comparative analysis of user engagement and content creation trends. In Proceedings of the 16th ACM Web Science Conference, Association for Computing Machinery, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.708, + 0.905, + 0.762 + ], + "angle": 0, + "content": "[98] Junke Wang, Dongdong Chen, Zuxuan Wu, Chong Luo, Luowei Zhou, Yucheng Zhao, Yujia Xie, Ce Liu, Yu-Gang Jiang, and Lu Yuan. Omnivl: One foundation model for image-language and video-language tasks. NeurIPS, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.763, + 0.905, + 0.803 + ], + "angle": 0, + "content": "[99] Teng Wang, Ruimao Zhang, Zhichao Lu, Feng Zheng, Ran Cheng, and Ping Luo. End-to-end dense video captioning with parallel decoding. In ICCV, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.805, + 0.905, + 0.845 + ], + "angle": 0, + "content": "[100] Yuxuan Wang, Difei Gao, Licheng Yu, Weixian Lei, Matt Feiszli, and Mike Zheng Shou. GEB+: A benchmark for generic event boundary captioning, grounding and retrieval. In ECCV, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.847, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[101] Ziyang Wang, Shoubin Yu, Elias Stengel-Eskin, Jaehong Yoon, Feng Cheng, Gedas Bertasius, and Mohit Bansal. VideoTree: Adaptive tree-based video representation for LLM reasoning on long videos. arXiv:2405.19209, 2024. 3, 7" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18957" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.092, + 0.482, + 0.133 + ], + "angle": 0, + "content": "[102] Xi Wei, Zhangxiang Shi, Tianzhu Zhang, Xiaoyuan Yu, and Lei Xiao. Multimodal high-order relation transformer for scene boundary detection. In ICCV, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.091, + 0.134, + 0.483, + 0.188 + ], + "angle": 0, + "content": "[103] Chao-Yuan Wu, Yanghao Li, Karttikeya Mangalam, Haoqi Fan, Bo Xiong, Jitendra Malik, and Christoph Feichtenhofer. MeMViT: Memory-augmented multiscale vision transformer for efficient long-term video recognition. In CVPR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.189, + 0.483, + 0.256 + ], + "angle": 0, + "content": "[104] Haoqian Wu, Keyu Chen, Haozhe Liu, Mingchen Zhuge, Bing Li, Ruizhi Qiao, Xiujun Shu, Bei Gan, Liangsheng Xu, Bo Ren, Mengmeng Xu, Wentian Zhang, Raghavendra Ramachandra, Chia-Wen Lin, and Bernard Ghanem. NewsNet: A novel dataset for hierarchical temporal segmentation. In CVPR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.257, + 0.483, + 0.298 + ], + "angle": 0, + "content": "[105] Zuxuan Wu, Caiming Xiong, Chih-Yao Ma, Richard Socher, and Larry S. Davis. AdaFrame: Adaptive frame selection for fast video recognition. In CVPR, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.299, + 0.483, + 0.339 + ], + "angle": 0, + "content": "[106] Junbin Xiao, Xindi Shang, Angela Yao, and Tat-Seng Chua. NExT-QA: Next phase of question-answering to explaining temporal actions. In CVPR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.34, + 0.482, + 0.38 + ], + "angle": 0, + "content": "[107] Jun Xu, Tao Mei, Ting Yao, and Yong Rui. MSR-VTT: A large video description dataset for bridging video and language. In CVPR, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.381, + 0.483, + 0.435 + ], + "angle": 0, + "content": "[108] Hongwei Xue, Tiankai Hang, Yanhong Zeng, Yuchong Sun, Bei Liu, Huan Yang, Jianlong Fu, and Baining Guo. Advancing high-resolution video-language representation with large-scale video transcriptions. In CVPR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.436, + 0.483, + 0.489 + ], + "angle": 0, + "content": "[109] Antoine Yang, Arsha Nagrani, Ivan Laptev, Josef Sivic, and Cordelia Schmid. VidChapters-7M: Video chapters at scale. In NeurIPS Track on Datasets and Benchmarks, 2023. 1, 2, 3, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.49, + 0.483, + 0.544 + ], + "angle": 0, + "content": "[110] Antoine Yang, Arsha Nagrani, Paul Hongsuck Seo, Antoine Miech, Jordi Pont-Tuset, Ivan Laptev, Josef Sivic, and Cordelia Schmid. Vid2Seq: Large-scale pretraining of a visual language model for dense video captioning. In CVPR, 2023. 1, 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.546, + 0.483, + 0.585 + ], + "angle": 0, + "content": "[111] Yang Yang, Yurui Huang, Weili Guo, Baohua Xu, and Dingyin Xia. Towards global video scene segmentation with context-aware transformer. AAAI, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.587, + 0.483, + 0.679 + ], + "angle": 0, + "content": "[112] Yuan Yao, Tianyu Yu, Ao Zhang, Chongyi Wang, Junbo Cui, Hongji Zhu, Tianchi Cai, Haoyu Li, Weilin Zhao, Zhihui He, Qianyu Chen, Huarong Zhou, Zhensheng Zou, Haoye Zhang, Shengding Hu, Zhi Zheng, Jie Zhou, Jie Cai, Xu Han, Guoyang Zeng, Dahai Li, Zhiyuan Liu, and Maosong Sun. MiniCPM-V: A GPT-4V level MLLM on your phone. arXiv:2408.01800, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.682, + 0.483, + 0.708 + ], + "angle": 0, + "content": "[113] Fangqiu Yi, Hongyu Wen, and Tingting Jiang. ASformer: Transformer for action segmentation. In BMVC, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.71, + 0.483, + 0.75 + ], + "angle": 0, + "content": "[114] Shoubin Yu, Jaemin Cho, Prateek Yadav, and Mohit Bansal. Self-chained image-language model for video localization and question answering. In NeurIPS, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.751, + 0.483, + 0.791 + ], + "angle": 0, + "content": "[115] Rowan Zellers, Ximing Lu, Jack Hessel, Youngjae Yu, Jae Sung Park, Jize Cao, Ali Farhadi, and Yejin Choi. MERLOT: Multimodal neural script knowledge models. In NeurIPS, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.792, + 0.483, + 0.858 + ], + "angle": 0, + "content": "[116] Andy Zeng, Maria Attarian, Brian Ichter, Krzysztof Choromanski, Adrian Wong, Stefan Welker, Federico Tombari, Aveek Purohit, Michael Ryoo, Vikas Sindhwani, et al. Socratic models: Composing zero-shot multimodal reasoning with language. In ICLR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.86, + 0.483, + 0.899 + ], + "angle": 0, + "content": "[117] Kuo-Hao Zeng, Tseng-Hung Chen, Juan Carlos Niebles, and Min Sun. Title generation for user generated videos. In ECCV, 2016. 2" + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.092, + 0.483, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.133 + ], + "angle": 0, + "content": "[118] Runhao Zeng, Wenbing Huang, Mingkui Tan, Yu Rong, Peilin Zhao, Junzhou Huang, and Chuang Gan. Graph convolutional networks for temporal action localization. In CVPR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.134, + 0.907, + 0.173 + ], + "angle": 0, + "content": "[119] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. ICCV, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.176, + 0.907, + 0.203 + ], + "angle": 0, + "content": "[120] Chenlin Zhang, Jianxin Wu, and Yin Li. ActionFormer: Localizing moments of actions with transformers. In ECCV, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.204, + 0.906, + 0.257 + ], + "angle": 0, + "content": "[121] Ce Zhang, Taixi Lu, Md Mohaiminul Islam, Ziyang Wang, Shoubin Yu, Mohit Bansal, and Gedas Bertasius. A simple LLM framework for long-range video question-answering. In EMNLP, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.259, + 0.906, + 0.299 + ], + "angle": 0, + "content": "[122] Hang Zhang, Xin Li, and Lidong Bing. Video-LLaMA: An instruction-tuned audio-visual language model for video understanding. In EMNLP Demo, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.3, + 0.906, + 0.368 + ], + "angle": 0, + "content": "[123] Shengyu Zhang, Ziqi Tan, Zhou Zhao, Jin Yu, Kun Kuang, Tan Jiang, Jingren Zhou, Hongxia Yang, and Fei Wu. Comprehensive information integration modeling framework for video titling. In ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.369, + 0.906, + 0.41 + ], + "angle": 0, + "content": "[124] Cairong Zhao, Chutian Wang, Zifan Song, Guosheng Hu, Hao-nan Chen, and Xiaofan Zhai. Cap2Sum: Learning to summarize videos by generating captions. arXiv:2408.12800, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.411, + 0.905, + 0.464 + ], + "angle": 0, + "content": "[125] Junjie Zhou, Yan Shu, Bo Zhao, Boya Wu, Shitao Xiao, Xi Yang, Yongping Xiong, Bo Zhang, Tiejun Huang, and Zheng Liu. MLVU: A comprehensive benchmark for multi-task long video understanding. arXiv:2406.04264, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.466, + 0.906, + 0.505 + ], + "angle": 0, + "content": "[126] Luowei Zhou, Xu Chenliang, and Jason J. Corso. Towards automatic learning of procedures from web instructional videos. In AAAI, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.507, + 0.906, + 0.546 + ], + "angle": 0, + "content": "[127] Luowei Zhou, Yingbo Zhou, Jason J Corso, Richard Socher, and Caiming Xiong. End-to-end dense video captioning with masked transformer. In CVPR, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.548, + 0.906, + 0.589 + ], + "angle": 0, + "content": "[128] Xingyi Zhou, Anurag Arnab, Shyamal Buch, Shen Yan, Austin Myers, Xuehan Xiong, Arsha Nagrani, and Cordelia Schmid. Streaming dense video captioning. In CVPR, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.907, + 0.589 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "18958" + } + ] +] \ No newline at end of file diff --git a/2025/Chapter-Llama_ Efficient Chaptering in Hour-Long Videos with LLMs/c5a48e4c-7b19-4801-9ffe-45905e053adb_origin.pdf b/2025/Chapter-Llama_ Efficient Chaptering in Hour-Long Videos with LLMs/c5a48e4c-7b19-4801-9ffe-45905e053adb_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..248617a29727381538ad5ba04b63fc85be010766 --- /dev/null +++ b/2025/Chapter-Llama_ Efficient Chaptering in Hour-Long Videos with LLMs/c5a48e4c-7b19-4801-9ffe-45905e053adb_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fb3c1db3c714c95350d380aa3d67e578e2bb255f11f4ea4ea327abcfc74cd958 +size 6613013 diff --git a/2025/Chapter-Llama_ Efficient Chaptering in Hour-Long Videos with LLMs/full.md b/2025/Chapter-Llama_ Efficient Chaptering in Hour-Long Videos with LLMs/full.md new file mode 100644 index 0000000000000000000000000000000000000000..b4e264676e82f224061256d2d09978a4777e48a1 --- /dev/null +++ b/2025/Chapter-Llama_ Efficient Chaptering in Hour-Long Videos with LLMs/full.md @@ -0,0 +1,312 @@ +# Chapter-Llama: Efficient Chaptering in Hour-Long Videos with LLMs + +Lucas Ventura1,2 + +Antoine Yang3 + +Cordelia Schmid2 + +Gül Varol + +$^{1}$ LIGM, École des Ponts, IP Paris, Univ Gustave Eiffel, CNRS + +2Inria, École normale supérieure, CNRS, PSL Research University 3Google DeepMind + +https://imagine.enpc.fr/~lucas.ventura/chapter-llama/ + +# Abstract + +We address the task of video chaptering, i.e., partitioning a long video timeline into semantic units and generating corresponding chapter titles. While relatively underexplored, automatic chaptering has the potential to enable efficient navigation and content retrieval in long-form videos. In this paper, we achieve strong chaptering performance on hour-long videos by efficiently addressing the problem in the text domain with our 'Chapter-Llama' framework. Specifically, we leverage a pretrained large language model (LLM) with large context window, and feed as input (i) speech transcripts and (ii) captions describing video frames, along with their respective timestamps. Given the inefficiency of exhaustively captioning all frames, we propose a lightweight speech-guided frame selection strategy based on speech transcript content, and experimentally demonstrate remarkable advantages. We train the LLM to output timestamps for the chapter boundaries, as well as free-form chapter titles. This simple yet powerful approach scales to processing one-hour long videos in a single forward pass. Our results demonstrate substantial improvements (e.g., 45.3 vs $26.7\%$ F1 score) over the state of the art on the recent VidChapters-7M benchmark. To promote further research, we release our code and models at our project page. + +# 1. Introduction + +According to a study by [46], the video durations uploaded to the popular online video sharing platform YouTube have increased steadily over the years. Videos have become longer since the first video upload in 2005 [19, 47]. In 2020, $25\%$ of videos were estimated to be longer than 15 minutes, $5\%$ more than 3 hours [46]. Long-form videos such as news, sports, educational, and vlog streams can often span extensive durations and cover multiple topics [97]. Finding specific content within increased video duration and volume makes efficient content navigation more important than ever. + +However, much of the traditional video analysis research has focused on processing short videos of a few seconds [4, 15, 34, 55, 63, 67, 74, 78, 85, 87, 98, 110]. At the same + +![](images/44bed3c22be373cb339c9910f1657007af24ef60b8b4665f684bfc137a87e205.jpg) +Figure 1. Chapter-Llama: Our method generates automatic video chapters for hour-long videos by training a large language model (LLM) to predict chapter boundaries and titles. The LLM processes transcribed speech (ASR) and descriptive captions of key frames, which are sampled based on ASR content. This text-based approach, equipped with speech-based frame selection, enables efficient processing of long-form content. + +time, the definition of long videos has changed within the past decade. Early works claimed processing 100 frames (i.e., a few seconds) to be long [61, 93] as opposed to ingesting up to 16 frames [83, 92]. With the introduction of datasets containing 1-5 minute videos [29, 37, 44, 56, 82, 126], several minutes were considered very long. Studying hour-long videos has only recently seen an interest in the context of movie description [31], video captioning [40], or grounding [32, 84]. Very recently, the work of [109] collected the VidChapters-7M dataset with videos spanning from minutes to hours, along with their user-defined video chapters, and proposed the video chapter generation task, automatically dividing a video into thematic sections (i.e., chapters) with descriptive concise chapter titles. Video chaptering, if achieved successfully, can offer a compelling solution to long content indexing, bypassing the current need for time-consuming manual annotation by video owners [109]. + +In this paper, we address the challenge of automatic video chaptering with a simple yet effective framework designed to handle hour-long videos. Existing work for chaptering [109] + +relics on a dense video captioning model Vid2Seq [110], which combines multimodal inputs from video frames and ASR-based speech transcriptions. However, Vid2Seq operates on a fixed number of equally sampled frames (i.e., 100 frames), potentially missing important visual information. Furthermore, their approach based on transformer architecture uses video frame features directly, which requires learning a mapping from the visual modality to the textual modality. In contrast, our method is designed to address these limitations by (i) dynamically sampling keyframes from the video based on the speech content, and (ii) designing a purely text-based model leveraging image captioning to convert RGB frames into text. + +Our approach leverages a pretrained LLM, which we finetune specifically for the video chaptering task to predict jointly the chapter boundary timestamps and chapter titles, both in text form. The appeal of our model lies in processing only textual data as input, allowing us effectively leverage the long-context understanding capabilities of the LLM to scale to long videos. In particular, we incorporate speech transcriptions from automatic speech recognition (ASR) and automatic frame captions. Captioning has been used for video understanding as an intermediate representation in recent works, but in the context of retrieval or question answering (QA) for shorter videos (maximum 3 minutes) [58, 95, 116, 121]. In longer videos, since captioning every frame is computationally prohibitive, we employ a speech-based frame selection strategy that scales efficiently while preserving important content. Similar in spirit to [43], we primarily use audio to determine keyframes, specifically bootstrapping with an LLM trained only with the speech inputs. However, even when transforming a video into text, LLMs have a limited context window, allowing a maximum number of tokens as input in a single forward pass. To mitigate context window limitations for very long video inputs, we simply perform an iterative prediction, sequentially processing the video, where each iteration typically operates on a window length of about an hour duration. We evaluate our approach on 'short' (0-15 min), 'medium' (15-30 min), and 'long' (30-60 min) videos from the VidChapters-7M dataset [109], demonstrating significant improvements over the state of the art across multiple metrics, including temporal boundary accuracy and semantic relevance of chapter titles. Our experiments show that finetuning the LLM, our speech-based frame selection strategy, and the integration of modalities from both speech and captions are crucial for achieving high-quality video chaptering results. + +Our contributions are the following: (i) We introduce Chapter-Llama: our framework leverages a pretrained LLM and finetunes for the underexplored task of video chaptering by transforming the video input into text form through ASR and captioning. (ii) We scale efficiently to hour-long videos by incorporating a speech-based frame sampling strategy, captioning only a subset of the video frames. (iii) Our simple and effective approach outperforms the state of the art on the recent VidChapters-7M benchmark by a large margin (e.g., 45.3 vs 26.7 F1 score). These results are complemented by a comprehensive set of experiments analyzing our components. + +# 2. Related Work + +We provide an overview of video tasks related to video chaptering, such as temporal segmentation and captioning, along with a discussion on works focusing on long-form and LLM-based video understanding. + +Temporal video segmentation. While video chaptering is a new task [109], there is a rich literature on methods focused on temporally segmenting a video in various forms. One task is shot detection [72, 76, 81], where any visual changes (e.g., shifting between two cameras) would require a temporal boundary, not necessarily modeling semantic shifts. Video scene segmentation, often studied on movies [38], is primarily focusing on grouping scenes with similar content [13, 14, 38, 39, 59, 65, 66, 71, 75, 77, 102, 111]. Another line of work considers boundary detection for temporal action segmentation [7, 23, 26, 48, 113], or localization [18, 54, 118, 120]. Unlike chaptering with free-form text, action segmentation assigns a label from a predefined set of categories, and typically defines short atomic actions as the unit. In contrast to these tasks, chapter boundaries can take various different forms depending on the type and the granularity of the video (e.g., each exercise within sports video, each slide within a lecture, each step in instructional video, each topic in a podcast video). Shot, scene, or action boundaries therefore may or may not correspond to complex chapter boundary definitions. Moreover, these tasks are mostly tackled with vision-only inputs [81, 113, 120], without leveraging speech. While text and audio segmentation have also been tackled separately [28, 73], video chaptering is based on both audio and vision inputs [109]. + +Video captioning. Generating chapter titles [109] is relevant to the task of captioning that seeks to describe the video content with text. There is a large literature on single video captioning [16, 50, 78, 80], often focusing on short video clips. Typical datasets for training such as MSR-VTT [107], WebVid [5], HowTo100M [57], Video-CC [60] include captions of videos spanning a few seconds (5-15sec on average). In generic event boundary captioning [100], event intervals are similarly short, in the order of 2 seconds. On the other hand, video summarization methods operate on longer videos; however, their goal is to reduce the entire video into a single summary description [1, 2, 33, 40, 51, 117, 123, 124], not necessarily with a temporal segmentation component. Dense video captioning [37, 44, 99, 110, 127, 128] is the closest to video chaptering in terms of problem formulation, aiming to both temporally localize and caption different events. Indeed, prior work on video chaptering trains the dense captioning method of Vid2Seq [110] on the VidChapters-7M dataset [109], but relies on a fixed number of equally sampled frames. In this paper, we leverage some of the annotations of this dataset to train an LLM-based chaptering model substantially outperforming previous methods [109, 110]. + +Long-form video understanding. The definition of long videos has evolved with the release of various datasets spanning + +seconds [106, 108], a few minutes [22, 29, 56, 86], 10-30 minutes [2, 125], or one hour [24, 40, 84, 104, 109]. MLVU [125] introduces a benchmark for evaluating multiple long video understanding tasks such as summarization and QA; however, the data is not suitable for chaptering due to lack of annotations. Video-MME [24] also contains hour-long videos for QA. MAD [31, 84] provides audio description for long movies, but each description spans a few seconds and the sparse coverage over the video is different than contiguous chapters. Recently, Ego4D-HCap [40] was proposed for hierarchical video summarization. However, this dataset involves dense captioning with visual inputs only, while we focus on video chaptering with visual and speech inputs. To the best of our knowledge, VidChapters7M [109] is the only open-sourced dataset for training and evaluating chapter generation, which we employ in this paper. Nonpublic related datasets include NewsNet [104] which includes hierarchical temporal segmentation annotations, the TV news chaptering dataset used in [30], and the ChapterGen dataset [10]. + +Increased video lengths led to a range of works focusing on efficient temporal modeling strategies. A common technique to deal with longer videos is to use pre-extracted visual features [31, 84, 115]. For end-to-end learning with transformers, several works explored factorized spatio-temporal attention [3, 5, 8]. Others have looked at various ways to incorporate a memory mechanisms [42, 103], blockwise attention [52, 53], or captioning frames to exploit LLMs [101, 121]. Given the redundancy in consecutive video frames, frame selection methods were explored in the context of short video captioning and action recognition [17, 105], as well as 'long' video QA in 3-minute durations [64, 88, 114]. Most common approach with current large video models is to perform sparse sampling with equal spacing [12, 45, 110]. SCSampler [43] exploits the low-dimensional audio modality to efficiently select salient video clips for action recognition. In our method, we also leverage audio, but in the form of ASR, and run the costly frame captioning step only on keyframes on locations predicted by a speech-based frame selection module. + +LLM use in video understanding. LLMs such as GPT [9, 68], Llama [20, 90, 91], and Gemini [27, 89], have been leveraged in different ways for improving video understanding. A popular approach is to train 'bridge' modules between pretrained visual backbones [69] and LLMs to build vision-language models (VLMs) that can ingest videos (e.g., Video-Llama [122], Video-LLaVa [49]). Other works have employed LLMs for automatic construction of video datasets [2, 40, 80, 96], tool use [58], storing memory in video QA [42], and temporal localization [36]. Similar to us, VideoTree [101] and VideoAgent [21] caption keyframes before passing them to an LLM together with a question for answer generation, addressing the limitations of [121] which performs a similar methodology without keyframe selection on shorter videos. In this study, we find that captioning alone is not sufficient, and needs to be complemented with ASR for competitive chaptering performance. Close to us, [2] exploits ASR on long videos and summarizes + +them with LLMs to generate pseudo-labels for video summarization training. In our work, we leverage LLMs, specifically finetuning a Llama model [20] for chaptering by prompting with speech transcription and frame captions. We show that finetuning is essential for adapting to the task so that the LLM picks up relevant content within the large context input [79]. + +# 3. Chapter-Llama: LLM-based Video Chaptering + +We provide an overview of our video chaptering framework, referred to as Chapter-Llama, in Fig. 2. Given video frames and speech transcripts, we aim at predicting relevant chapter boundaries and titles. For this, we first select video frames to process with a speech-based frame selection module. Then we use an off-the-shelf visual captioner to map the selected frames in the text space. We feed the resulting captions, along with speech transcripts, to the LLM which outputs the chapter boundaries and titles jointly as a single sequence of tokens. Finally, we devise an iterative prediction procedure in case the input text sequence is too long to handle for the LLM. We next describe in more detail each component. + +Task formulation. Video chaptering [109] aims at segmenting a video into semantically meaningful chapters, and generating a title for each segment. The chapters are contiguous, with no gaps between them, and together span the entire video duration from start to end. Formally, given video frames $V = (v_{1}, v_{2}, \dots, v_{N})$ and temporally-aligned speech transcripts $S = (s_{1}, s_{2}, \dots, s_{M})$ , where each speech transcript contains an utterance and its associated start and end timestamps, the task is to output a sequence of chapters $C = (c_{1}, c_{2}, \dots, c_{L})$ , where each chapter $c_{i}$ is a tuple $(b_{i}, t_{i})$ containing a start timestamp $b_{i}$ and a descriptive title $t_{i}$ . The end time of chapter $i$ is implicitly defined by the start time of the subsequent chapter $b_{i+1}$ , or total video duration if $i = L$ . + +Speech-based frame selection. Video chaptering involves processing hour-long videos. Therefore, densely sampling frames is computationally intractable due to numerous inference passes through a vision model (e.g., a visual captioner) and exceeding standard LLM context lengths. Upon inspection of our data, we found that while the speech transcription has 257 tokens per minute on average, a caption is 66 tokens long on average hence captions would take 3,960 tokens per minute when sampling a video at 1 FPS. To address these challenges, we employ a frame selection strategy. + +Specifically, we use speech transcripts to guide which video frames to process for the vision model. This is done by first training a speech-only variant of our LLM to predict a sequence of chapter boundaries $\{\hat{b}_1,\hat{b}_2,\dots,\hat{b}_K\}$ from speech transcripts $S$ only. For each predicted boundary $\hat{b}_i$ , we sample a frame $v_{i}$ from the video at that timestamp. Note that this variant is cheaper compared to the full model as it only needs ASR transcription from the audio stream, without requiring any processing of the RGB stream (i.e., captioning). We then process the video frames only at the time locations predicted + +![](images/d1b29773e7e14bc67b1a6e1d1b5d7b7983d9bbb7886a695c0b73e34943d3f3be.jpg) +Figure 2. Method overview: Our Chapter-Llama framework first selects video frames to process using speech information. Then we use a visual captioner to map the selected frames in the text space. We feed the resulting captions, along with speech transcripts, to the LLM which outputs the chapter boundaries and titles jointly as a single sequence of tokens. + +by this model. The visual information thus complements the previous 'blind' predictions from the narrations, and allows us to refine the predictions. This results in a video representations $V_{\text{sampled}} = (v_1, v_2, \dots, v_K)$ where $K << N$ . For the videos that lack speech entirely (e.g., about 3% of the videos in [109]), we sample frames at 10-second intervals, with an upper bound of 100 frames to maintain computational practicality. + +Mapping video to text with timestamps. To leverage the knowledge of a pretrained LLM, we map all our inputs to text. This includes: (1) speech transcriptions $S = (s_{1}, s_{2}, \dots, s_{M})$ from the audio modality, and (2) caption descriptions $V_{\text{caption}} = (d_{1}, d_{2}, \dots, d_{K})$ from the visual modality. In detail, for speech transcriptions, we use ASR outputs provided by [109], obtained using the Whisper-Large-V2 [70] model through the WhisperX [6] implementation. For captioning, we employ MiniCPM-V [112] as an image captioner, applied independently on the selected video frames, i.e., $d_{i} = \text{Captioner}(v_{i})$ . + +As we aim at predicting relevant chapter boundaries, we provide temporal information to the LLM. For both modalities, we prepend the timestamp information formatted as "HH:MM:SS" to encode the location at which the speech or caption is obtained. + +Captions naturally come from a single point in time. Speech segments cover intervals, but their duration is typically very short (3-4 seconds). We therefore simply use the start time of each transcribed speech interval. We interleave the speech and caption inputs based on their timestamps in a sorted order. We add a modality-specific prefix to each timestamp to denote which modality the information is extracted from (i.e., ASR for speech transcripts, Caption for captions). + +We prepend the text combining speech transcripts and captions with a fixed prompt that provides task instructions (see sup. mat. for the exact wording). This prompt occupies approximately 90 tokens and is independent of video length. + +Language model. We derive our framework by making use of a powerful pretrained LLM. Specifically, we employ the recent Llama-3.1-8B-Instruct [20] model and further finetune on chapter annotations using the LoRA technique [35]. Given the input structure previously described, the LLM is trained to output chapters, where each chapter consists of a timestamp in HH:MM:SS format followed by a free-form chapter title. We treat both the timestamps and titles simply as text tokens and apply the standard cross-entropy loss over the original vocabulary of the pretrained LLM. We apply teacher forcing during training and decode tokens autoregressively at inference. Note that the final model (taking both speech and captions as input) is trained independently from the speech-only version of our model used for frame selection, but these two models share the same backbone, and only differ in their LoRA parameters (13MB each). Across all experiments, we finetune models for a single epoch and use the same hyperparameters. We provide these hyperparameters, along with implementation details in Appendix A, and provide experiments with several Llama variants in Appendix C. + +Iterative prediction for long videos. The inputs may exceed the context window limitation of the LLM, especially in the case of long videos. For example, on an A6000 GPU, the Llama-3.1-8B-Instruct [20] model can process videos up to around 15k tokens during training, which corresponds to 50 minutes of video content on average, and 25k tokens during inference, which corresponds to 80 minutes of video content on average. To address this issue, during training, we select videos that have less than 15k tokens. Since there are videos up to 1 hour long in the training set that satisfy this constraint, and since we do not need the entire training dataset to achieve good performance, this token limitation does not hinder our training. During evaluation, we predict chapters for each chunk sequentially, such that the start of a chunk is the end of the previous chunk. Finally, we merge the predictions from all chunks to obtain chapter boundaries for the complete video. We provide more details in Appendix A.4. + +# 4. Experiments + +In this section, we start by describing the data and evaluation metrics used in our experiments (Sec. 4.1). Next, we compare our results with the state of the art (Sec. 4.2), and then provide a series of ablations in our framework (Sec. 4.3). Finally, we investigate the impact of testing with very long videos exceeding our context window limitations (Sec. 4.4). + +# 4.1. Data and evaluation + +Data. We train and evaluate on the recently released VidChapters-7M [109] dataset that includes user-annotated chaptered videos sourced from YouTube. Speech transcripts are obtained using Whisper [70] as the ASR method. In the original release, there is a total of 817k videos, spanning 8M chapters, with 2.4 minutes per chapter and 5.4 words per chapter title, totaling to 23 minutes and 8.3 chapters per video on average. Data is split into 801k training, 8.2k validation, and 8.2k test videos. To measure performance at different video lengths, we define three categories depending on video duration: 'short' (0-15min), 'medium' (15-30min), and 'long' (30-60min) videos. In this work, we use a subset of the training data as we observe increasing the training set brings diminishing returns at the cost of extended training times (see Fig. 4). Specifically, we use about 20k training videos (10k short videos used for the speech-based frame selection model and another 10k videos evenly split across short, medium and long durations for the final model. For state-of-the-art comparisons (Sec. 4.2), we employ the full official test set, which also contains videos without any speech (2.5% of the videos), and videos longer than 60 minutes (e.g., there are few videos that last about 12 hours). In ablations (Sec. 4.3), both for faster experimentation, and to limit the use of the test set during experimentation, we train on a randomly sampled subset of 1k videos (evenly split between short, medium, and long) and report results on a randomly sampled subset of 300 validation videos (100 from each duration) that have at least one speech utterance. + +Evaluation metrics. We primarily monitor temporal segmentation metrics to evaluate our chapter boundary detections. In particular, we employ tIoU and F1 scores. For tIoU (temporal Intersection over Union), we first compute the optimal matching between predicted and ground truth segments by greedily selecting pairs with highest IoU scores. The tIoU score is then calculated as the mean IoU across all matched pairs, multiplied by 100 to obtain a percentage. For F1 score, we first compute precision and recall at different IoU thresholds (ranging from 0.5 to 0.95 with a step of 0.05). At each threshold, a prediction is considered correct if it has IoU above the threshold with a ground truth segment. The precision is the ratio of correct predictions to total predictions, while recall is the ratio of matched ground truth segments to total ground truth segments. The F1 score is then computed as the harmonic mean of precision and recall. The final F1 metric is the average across all thresholds, multiplied by 100 to obtain a percentage. Note that [109] uses recall and precision metrics in two ways: (1) + +by considering timestamps within 3 or 5 second thresholds as matches, and (2) by considering segments with IoU above 0.5 or 0.7 as matches. While these metrics provide point estimates at specific thresholds, we find that IoU and F1 scores offer several advantages: they evaluate performance continuously across multiple thresholds, are more interpretable, and provide a more comprehensive evaluation of the model. For completeness, we also report the metrics used in [109] in Appendix C. + +For chapter title evaluation, we follow [109] and report SODA (S) [25] and CIDEr (C) [94], which measure the quality of the titles for the predicted segments that match to the ground segments (see [109] for details). + +# 4.2. Comparison with the state of the art + +In Tab. 1, we report the performance of our model on the full VidChapters-7M test set [109] ('All' columns), and compare to the state of the art reported in [109], which uses Vid2Seq [110]. Moreover, we evaluate four proprietary models using our speech-based frame selection and captioning in a zero-shot manner. + +We observe that our finetuned Chapter-Llama achieves substantial performance improvements across all metrics and video duration categories. (e.g., 45.3 vs 26.7 F1 and 19.3 vs 11.6 SODA compared to Vid2Seq). Notably, our improvement over Vid2Seq is more important for medium and long videos compared to short videos. Note that our final approach was trained using the subset of data detailed in the previous section, specifically 20k videos, which constitutes only $2.5\%$ of the total available training data. In contrast, the baseline Vid2Seq model [110] was trained on a considerably larger dataset, utilizing both HowTo100M [57] and the entire VidChapters-7M training set. + +Additionally, we report performances of our model without training on any chapter annotations (i.e., both the speech-based frame selector and the LLM are not finetuned, and run with the same prompt as in the finetuned setting). We see that our zero-shot method also achieves competitive performance (e.g., 29.5 F1), whereas Vid2Seq only trained on HowTo100M does not generalize (3.0 F1). + +Finally, when zero-shot evaluating the proprietary models, GPT4-o [62] and Gemini variants [27], with our speech-based frame selection and captioning inputs, we observe competitive performances (e.g., 42.2 F1 with Gemini-1.5-Pro); however, our Chapter-Llama still surpasses on all metrics. Note that, due to API costs of the proprietary models, we performed their evaluation on a random $10\%$ subset of the test set; however, we verified that the scores are similar between $10\%$ and $100\%$ of the test set when evaluating with Chapter-Llama. + +Qualitative comparison. In Fig. 3, we provide qualitative examples comparing our method against Vid2Seq [109, 110] and our zero-shot baseline. Our predictions align well with the ground truth chapters, accurately capturing both the temporal boundaries and generating relevant titles. In contrast, Vid2Seq segments tend to be less accurate, and we also observe that it often produces repetitive titles (bottom example). The zero-shot Chapter-Llama baseline tends to generate relatively longer and + +
BackboneFrame selectionFt.ShortMediumLongAll
F1tIoUSCF1tIoUSCF1tIoUSCF1tIoUSC
GPT-4o-mini [62]†OursX32.164.57.242.430.562.36.130.628.061.06.027.331.263.66.837.8
GPT-4o [62]†OursX37.768.08.453.838.168.88.151.436.566.26.634.837.668.08.151.0
Gemini-2.0-Flash [27]†OursX39.969.212.072.843.871.411.270.334.966.29.051.640.269.311.469.7
Gemini-1.5-Pro [27]†OursX41.770.611.765.343.871.811.261.441.370.610.155.342.270.911.463.2
Vid2Seq [109, 110]EquidistantX2.528.60.30.33.229.70.30.44.632.00.30.53.029.30.30.4
Llama 3.1-8BOursX29.963.47.134.530.662.75.428.126.659.33.618.929.562.56.230.7
Vid2Seq [109, 110]Equidistant✓33.463.715.274.919.053.37.531.916.750.85.928.426.758.611.655.8
Llama 3.1-8B (Chapter-Llama)Ours✓45.572.220.2103.546.772.318.898.741.369.215.891.245.371.819.3100.9
+ +Table 1. Comparison to the state of the art on VidChapters-7M test set: We split the table into (bottom) the comparison between Chapter-Llama and the state-of-the-art method Vid2Seq [110], and (top) the evaluation of proprietary models. Chapter-Llama significantly outperforms Vid2Seq trained and reported by [109] (45.3 vs 26.7 F1). Our method also achieves strong performance in zero-shot mode - without finetuning (Ft.) on any chapter annotation (29.5 F1). Furthermore, we report performance of proprietary models in such zero-shot setting, using our speech-based frame selection and captioning, and observe inferior results than Chapter-Llama (42.2 F1 with Gemini-1.5-Pro). Note that we use the full official 8.1k test set videos ('All'), unlike in the remaining experiments that report on the validation subset. We also report the performance breakdown into short (4891), medium (1736), and long (892) test videos. Our model was trained on 10k videos balanced across short, medium and long durations. $\dagger$ denotes evaluation on a random $10\%$ subset of the test set due to API costs of proprietary models. + +![](images/0f877bd2a95cfe6c084d86dd98b61f7f56dd78f44b04f115579661268a192f75.jpg) +Figure 3. Qualitative results: We display two examples and compare our Chapter-Llama results against the ground truth (GT), as well as the zero-shot (ZS) and Vid2Seq (VS) baselines. For each example, we show the corresponding SODA (S) and CIDEr (C) scores. Our method overall shows the highest similarity with the GT, while Vid2Seq can suffer from repeated chapter titles, and zero-shot generations tend to over-segment. + +
ModalitiesSegmentationTitles
SpeechCaptionsF1tIoUSC
Zero-shotX✓12.648.61.96.4
✓X22.757.34.419.7
✓✓29.963.06.933.7
FinetunedX✓39.167.75.920.2
✓X38.568.113.967.3
✓✓42.670.616.482.4
+ +verbose chapter titles and often generates chapters that appear to be continuations of previous chapters rather than distinct segments, while also exhibiting over-segmentation issues. We provide more examples in Appendix E. + +# 4.3. Ablation studies + +In the following, we experiment with (i) the contribution of speech and caption modalities, along with the effect of LLM finetuning, (ii) the effect of our frame selection method for captioning, (iii) the amount of training data, and (iv) the use of frame embeddings instead of captions. As mentioned above, we use 1k training and 300 validation videos for these ablations. + +Modalities and LLM finetuning. In Tab. 2, we ablate the impact of finetuning the LLM and the contribution of each of the speech and caption modalities. In the top block, we run our baselines in zero-shot setting as introduced in the previous section. The speech-only baseline outperforms the captions-only baseline by a large margin in the zero-shot setting. This suggests that speech contains more relevant information for chaptering, as was previously observed by [109]. + +As shown in the bottom block of Tab. 2, we observe large performance improvements when finetuning the LLM, as opposed to zero-shot. We hypothesize that zero-shot prompting with a long multi-modal text, potentially containing redundant and irrelevant information, may overwhelm the LLM [79, 101]. We obtain our best model by combining the two modalities, which performs better than the individual speech-only or caption-only models. This demonstrates the multi-modal capabilities of our model. + +Speech-based frame selection. In Tab. 3, we examine a number of strategies to sample frames at which we extract captions. In addition to previously described metrics, for each of the frame sampling approaches, we report the average number of captions per video and the average number of text tokens per minute. For reference, we also report an off-the-shelf shot detection [11] and Vid2Seq [109, 110]. + +We compare our speech-based frame selection strategy to various baselines. We experiment with sampling (i) uniformly + +Table 2. Contribution of different modalities and finetuning: Finetuning the LLM with 1k videos largely improves chaptering performance on 300 validation videos, see bottom block vs top block. In the finetuned setting, we further demonstrate the advantages of combining both modalities, i.e., transcribed speech from ASR and automatic captions extracted from video frames. + +
MethodFrame selection for captionsaverage #frames#tokens per min.SegmentationTitles SC
F1tIoU
BASELINES
Shot detection [11]n/a49.4n/a6.237.6--
Vid2Seq [109, 110]100 equidistant100.0128.625.457.811.255.0
CHAPTER-LLAMA VARIANTS
SpeechCaption
✓✘n/an/a248.638.568.113.967.3
100 equidistant100.0449.121.053.88.436.0
Every 10 sec.83.1280.312.845.94.313.0
✘✓Shot boundaries49.4193.216.250.73.912.4
10 equidistant10.041.811.046.43.69.0
Speech-based10.336.239.167.75.920.2
100 equidistant100.0746.239.267.416.183.8
Every 10 sec.83.1570.141.069.315.477.3
✓✓Shot boundaries40.4481.740.669.115.879.3
10 equidistant10.0326.140.167.915.877.5
Speech-based10.3320.442.670.616.482.4
+ +Table 3. Frame selection strategies for captioning: We evaluate different approaches for selecting frames to extract captions from, comparing our speech-based selection method against baselines. The table shows results for models trained on 1k videos and evaluated on 300 validation videos. We experiment with using speech only, captions only, and both modalities (bottom section). For caption extraction, we compare our speech-based approach to other alternatives such as equidistant sampling (100 or 10 frames), uniformly sampling every 10 seconds, or sampling at shot boundaries using [11]. Our speech-based frame selection achieves the best overall performance (F1: 42.6, tIoU: 70.6) while requiring significantly fewer number of frames on average (10.3) compared to other sampling approaches. The tokens-per-minute statistic shows the total input length including both speech transcriptions and captions, excluding the fixed prompt template. + +100 frames as in Vid2Seq, (ii) every 10 seconds, (iii) at shot boundaries detected by an off-the-shelf shot detector [11], (iv) 10 equidistant frames to be similar to our speech-based locations (i.e., 10.0 vs 10.3 number of frames on average), and (v) sampling at frames predicted as chapter boundaries by our LLM that inputs only speech. In all cases, we limit the maximum number of frames to 100 as in [109, 110] to handle extreme durations. + +In both caption-only and caption+speech settings, our speech-based frame selection approach achieves better segmentation results than the more frame-expensive baselines '100 equidistant', 'every 10 sec', and 'shot boundaries', while using much less frames, and also improves over the '10 equidistant' baseline which uses a similar number of frames. This demonstrates the effectiveness of our speech-based frame selection strategy. + +For reference, we also report positive comparison against shot detection and Vid2Seq [109, 110]. Note Vid2Seq has less tokens per min. compared to our 100 equidistant variants, because Vid2Seq uses a different timestamp tokenizer in the input. + +Amount of training data. Given the large-scale nature of the VidChapters-7M training set, we investigate how much chapter data is needed for LoRA finetuning the LLM. We plot the performance against the number of training videos in Fig. 4. We start by the zero-shot baseline as the first data point, and report our method with 1k, 5k, 7k, and 10k videos, split evenly + +![](images/7e58313b19ae66019b928c53d56a372c0f5a15c1b0f4c22319e6125742aa64fe.jpg) +Figure 4. Amount of training data: Our experiments show a substantial improvement when moving from zero-shot to training with 1k videos. Beyond 1k videos, performance continues to improve but at a much slower rate, motivating our choice of using only 10k training videos for our final LLM. + +between three durations. We see that after increasing above several thousand training videos starts to bring diminishing returns. We therefore keep 10k training videos for our final LLM, which makes our approach highly efficient to train (40min on 4 H100 GPUs). Note that here we focus on the chaptering LLM and always use frame sampling locations from a speech-based module trained on 10k separate videos. + +Frame embeddings vs captions. In Tab. 4, we investigate whether raw visual embeddings could serve as an alternative to textual captions. To this end, we experiment with replacing the captions with frame embeddings. Specifically, for each frame, we extract the 1152-dimensional output embedding corresponding to the [CLS] token from a frozen SigLIP model [119], and feed through a 2-layer MLP mapping network. We initialize the MLP weights from MANTIS [41] and train jointly with the LLM during finetuning. The results with 'Speech+Embeddings' are better than 'Speech' alone (38.5 vs 40.4 F1), but worse than 'Speech+Captions' (42.6 vs 40.4 F1). The performance gap between 'Speech+Embeddings' and 'Speech+Captions' may be due to the richer information provided by captions, which use multiple tokens per frame, directly in text form, compared to the single [CLS] token frame embedding, requiring a mapping network to be ingested by an LLM. Finally, while combining all modalities achieves the best performance (44.4 F1), we exclude frame embeddings from our final model due to practical considerations, e.g., they add complexity, increase processing time by 2.5x, and require 3000x more storage space. + +# 4.4. Iterative prediction on longer videos + +In our ablation studies, our experimental setting considered training and evaluating with videos that fit within the LLM context window. In Tab. 5, we evaluate the benefit of our iterative prediction procedure for handling videos that exceed the LLM context window. For this, we identify videos in the validation set whose inputs exceed the LLM inference context window ( $>35k$ tokens), resulting in 110 videos. On this challenging subset, we find that our iterative prediction procedure improves chaptering results compared to the baseline that only runs the LLM once by cropping the input to the first + +
ModalitiesSegmentationTitles
SpeechEmbeddingsCaptionsF1tIoUSC
✓--38.568.113.967.3
-✓-38.466.53.47.3
--✓39.167.75.920.2
✓✓-40.468.215.374.9
✓-✓42.670.616.482.4
✓✓✓44.471.516.384.2
+ +Table 4. Frame embeddings vs captions: We compare using frame captions versus visual features from a frozen SigLIP model projected through a learned 2-layer MLP mapping network ('Embeddings'). While the 'Speech+Embeddings' combination performs better than speech alone (40.4 vs 38.5 F1), it underperforms compared to the 'Speech+Captions' combination (42.6 vs 40.4 F1). All models are trained with 1k videos and evaluated on 300 videos. + +
Window# tok.avg # iter.Subset exceeding 35k tokens
F1tIoUSC
First10k113.150.54.031.2
15k116.654.95.443.3
20k118.756.76.647.5
Iterative10k8.518.557.16.925.1
15k5.423.660.18.735.2
20k4.125.361.410.344.0
+ +Table 5. Iterative prediction: Our iterative prediction procedure improves chaptering results on the subset of 110 videos which exceed 35k tokens compared to the baseline that only runs the LLM once (by only taking the first window, and discarding the rest of the input sequence), across various context windows. As we increase the context window in the iterative prediction, the performance gradually improves and the average number of iterations decreases. The model is trained with 1k videos. + +input window, across various context windows (10k, 15k, and 20k). We refer to Appendix B for details on the video lengths and statistics of videos that exceed the LLM context window. + +# 5. Conclusions + +We presented Chapter-Llama, an approach that leverages LLMs for hour-long video chaptering by mapping video to text using speech transcripts and efficiently captioning video frames sampled with a speech-based frame selector. Our results on VidChapters-7M consequently improved the state of the art by a large margin. We experimentally demonstrated the benefits of our components through an extensive ablation study. One limitation of our approach is that it relies on the accuracy of the ASR and the visual captioner. Future work can explore hierarchical chaptering with several granularities and consider the audio modality beyond speech. We also note that the LLM, the visual captioner, and speech transcription models are trained on large Web datasets, which can contain biases that can lead to inaccurate chaptering, especially for videos depicting underrepresented topics. + +Acknowledgements. This work was granted access to the HPC resources of IDRIS under the allocation 2024-AD011014696 made by GENCI. This work was funded in part by the ANR project CorVis ANR-21-CE23-0003-01, a research gift from Google, the French government under management of Agence Nationale de la Recherche as part of the "France 2030" program, reference ANR-23-IACL-0008 (PR[AI]RIE-PSAI projet), and the ANR project VideoPredict ANR-21-FAI1-0002-01. Cordelia Schmid would like to acknowledge the support by the Körber European Science Prize. The authors would also like to thank Guillaume Astruc, Nikos Athanasiou, Hyolim Kang, and Nicolas Dufour for their feedback. + +# References + +[1] Soheyla Amirian, Khaled Rasheed, Thiab R Taha, and Hamid R Arabnia. Automatic generation of descriptive titles for video clips using deep learning. In Advances in Artificial Intelligence and Applied Cognitive Computing: Proceedings from ICAI'20 and ACC'20, 2021. 2 +[2] D. Argaw, S. Yoon, F. Heilbron, H. Deilamsalehy, T. Bui, Z. Wang, F. Dernoncourt, and J. Chung. Scaling up video summarization pretraining with large language models. In CVPR, 2024. 2, 3 +[3] Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lucic, and Cordelia Schmid. ViViT: A video vision transformer. In ICCV, 2021. 3 +[4] Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio Neural machine translation by jointly learning to align and translate. ICLR, 2015. 1 +[5] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In ICCV, 2021. 2, 3 +[6] Max Bain, Jaesung Huh, Tengda Han, and Andrew Zisserman. WhisperX: Time-accurate speech transcription of long-form audio. In Interspeech, 2023. 4 +[7] Nadine Behrmann, S Alireza Golestaneh, Zico Kolter, Jürgen Gall, and Mehdi Noroozi. Unified fully and timestamp supervised temporal action segmentation via sequence to sequence translation. In ECCV, 2022. 2 +[8] Gedas Bertasius, Heng Wang, and Lorenzo Torresani. Is space-time attention all you need for video understanding? In ICML, 2021. 3 +[9] Tom Brown et al. Language models are few-shot learners. In NeurIPS, 2020. 3 +[10] Xiao Cao, Zitan Chen, Canyu Le, and Lei Meng. Multi-modal video chapter generation. In BMVC, 2022. 3 +[11] Brandon Castellano. Pyscenedetect: Intelligent scene cut detection and video splitting tool. https://pyscenedetect.readthedocs.io/en/latest/, 2018.7 +[12] Houlun Chen, Xin Wang, Hong Chen, Zihan Song, Jia Jia, and Wenwu Zhu. Grounding-prompter: Prompting llm with multimodal information for temporal sentence grounding in long videos. arXiv:2312.17117, 2023. 3 +[13] Shixing Chen, Xiaohan Nie, David Fan, Dongqing Zhang, Vimal Bhat, and Raffay Hamid. Shot contrastive self-supervised learning for scene boundary detection. In CVPR, 2021. 2 + +[14] Shixing Chen, Chun-Hao Liu, Xiang Hao, Xiaohan Nie, Maxim Arap, and Raffay Hamid. Movies2Scenes: Using movie metadata to learn scene representation. In CVPR, 2023. 2 +[15] Sihan Chen, Xingjian He, Longteng Guo, Xinxin Zhu, Weining Wang, Jinhui Tang, and Jing Liu. VALOR: Vision-audiolanguage omni-perception pretraining model and dataset. TPAMI, 2024. 1 +[16] Sihan Chen, Handong Li, Qunbo Wang, Zijia Zhao, Mingzhen Sun, Xinxin Zhu, and Jing Liu. VAST: A vision-audio-subtitle-text omni-modality foundation model and dataset. NeurIPS, 2024. 2 +[17] Yangyu Chen, Shuhui Wang, Weigang Zhang, and Qingming Huang. Less is more: Picking informative frames for video captioning. In ECCV, 2018. 3 +[18] Feng Cheng and Gedas Bertasius. TALLformer: Temporal action localization with long-memory transformer. In ECCV, 2022. 2 +[19] Xu Cheng, Cameron Dale, and Jiangchuan Liu. Statistics and social network of youtube videos. In 16th International Workshop on Quality of Service, 2008. 1 +[20] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The Llama 3 herd of models. arXiv:2407.21783, 2024. 3, 4 +[21] Yue Fan, Xiaojian Ma, Rujie Wu, Yuntao Du, Jiaqi Li, Zhi Gao, and Qing Li. VideoAgent: A memory-augmented multimodal agent for video understanding. In ECCV, 2024. 3 +[22] Xinyu Fang, Kangrui Mao, Haodong Duan, Xiangyu Zhao, Yining Li, Dahua Lin, and Kai Chen. MMBench-video: A long-form multi-shot benchmark for holistic video understanding. In NeurIPS Datasets and Benchmarks, 2024. 3 +[23] Yazan Abu Farha and Jurgen Gall. MS-TCN: Multi-stage temporal convolutional network for action segmentation. In CVPR, 2019. 2 +[24] Chaoyou Fu, Yuhan Dai, Yondong Luo, Lei Li, Shuhuai Ren, Renrui Zhang, Zihan Wang, Chenyu Zhou, Yunhang Shen, Mengdan Zhang, et al. Video-MME: The first-ever comprehensive evaluation benchmark of multi-modal LLMs in video analysis. arXiv:2405.21075, 2024. 3 +[25] Soichiro Fujita, Tsutomu Hirao, Hidetaka Kamigaito, Manabu Okumura, and Masaaki Nagata. SODA: Story oriented dense video captioning evaluation framework. In ECCV, 2020. 5 +[26] Shang-Hua Gao, Qi Han, Zhong-Yu Li, Pai Peng, Liang Wang, and Ming-Ming Cheng. Global2Local: Efficient structure search for video action segmentation. In CVPR, 2021. 2 +[27] Gemini Team et al. Gemini: A family of highly capable multimodal models. arXiv:2312.11805, 2024. 3, 5, 6 +[28] Azin Ghazimatin, Ekaterina Garmash, Gustavo Penha, Kristen Sheets, Martin Achenbach, Oguz Semerci, Remi Galvez, Marcus Tannenberg, Sahitya Mantravadi, Divya Narayanan, Ofeliya Kalaydzhyan, Douglas Cole, Ben Carterette, Ann Clifton, Paul N. Bennett, Claudia Hauff, and Mounia Lalmas. PODTILE: Facilitating podcast episode browsing with auto-generated chapters. In Proceedings of the 33rd ACM International Conference on Information and Knowledge Management, 2024. 2 +[29] Kristen Grauman et al. Ego4D: Around the world in 3,000 hours of egocentric video. In CVPR, 2022. 1, 3 +[30] Khalil Guetari, Yannis Tevisen, and Frédéric Petitpont. Multimodal chaptering for long-form TV newscast video. arXiv:2406.17590, 2024. 3 + +[31] Tengda Han, Max Bain, Arsha Nagrani, Gül Varol, Weidi Xie, and Andrew Zisserman. AutoAD: Movie description in context. In CVPR, 2023. 1, 3 +[32] Tanveer Hannan, Md Mohaiminul Islam, Thomas Seidl, and Gedas Bertasius. RGNet: A unified retrieval and grounding network for long videos. In ECCV, 2023. 1 +[33] Bo He, Jun Wang, Jielin Qiu, Trung Bui, Abhinav Shrivastava, and Zhaowen Wang. Align and attend: Multimodal summarization with dual contrastive losses. In CVPR, 2023. 2 +[34] Chiori Hori, Takaaki Hori, Teng-Yok Lee, Ziming Zhang, Bret Harsham, John R Hershey, Tim K Marks, and Kazuhiko Sumi. Attention-based multimodal fusion for video description. In ICCV, 2017. 1 +[35] Edward J Hu, yelong shen, Phillip Wallis, Zeyuan Allen-Zhu, Yanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In ICLR, 2022. 4 +[36] De-An Huang, Shijia Liao, Subhashree Radhakrishnan, Hongxu Yin, Pavlo Molchanov, Zhiding Yu, and Jan Kautz. Lita: Language instructed temporal-localization assistant. In ECCV, 2024. 3 +[37] Gabriel Huang, Bo Pang, Zhenhai Zhu, Clara Rivera, and Radu Soricut. Multimodal pretraining for dense video captioning. In AACL-IJCNLP, 2020. 1, 2 +[38] Qingqiu Huang, Yu Xiong, Anyi Rao, Jiaze Wang, and Dahua Lin. Movienet: A holistic dataset for movie understanding. In ECCV, 2020. 2 +[39] Md Mohaiminul Islam, Mahmudul Hasan, Kishan Shamsundar Athrey, Tony Braskich, and Gedas Bertasius. Efficient movie scene detection using state-space transformers. In CVPR, 2023. 2 +[40] Md Mohaiminul Islam, Ngan Ho, Xitong Yang, Tushar Nagarajan, Lorenzo Torresani, and Gedas Bertasius. Video ReCap: Recursive captioning of hour-long videos. CVPR, 2024. 1, 2, 3 +[41] Dongfu Jiang, Xuan He, Huaye Zeng, Cong Wei, Max Ku, Qian Liu, and Wenhu Chen. Mantis: Interleaved multi-image instruction tuning. TMLR, 2024. 8 +[42] Kumara Kahatapitiya, Kanchana Ranasinghe, Jongwoo Park, and Michael S Ryoo. Language repository for long video understanding. arXiv:2403.14622, 2024. 3 +[43] Bruno Korbar, Du Tran, and Lorenzo Torresani. SCSampler: Sampling salient clips from video for efficient action recognition. In ICCV, 2019. 2, 3 +[44] Ranjay Krishna, Kenji Hata, Frederic Ren, Li Fei-Fei, and Juan Carlos Niebles. Dense-captioning events in videos. In ICCV, 2017. 1, 2 +[45] Jie Lei, Linjie Li, Luowei Zhou, Zhe Gan, Tamara L Berg, Mohit Bansal, and Jingjing Liu. Less is more: ClipBERT for video-and-language learning via sparse sampling. In CVPR, 2021. 3 +[46] Feng Li, Jae Won Chung, and Mark Claypool. Three-year trends in YouTube video content and encoding. Proceedings of the 18th International Conference on Signal Processing and Multimedia Applications (SIGMAP), 2021. 1 +[47] Mingzhe Li, Mark Claypool, Robert Kinicki, and James Nichols. Characteristics of streaming media stored on the web. ACM Trans. Internet Technol., 2005. 1 +[48] Zhe Li, Yazan Abu Farha, and Jurgen Gall. Temporal action segmentation from timestamp supervision. In CVPR, 2021. 2 +[49] Bin Lin, Bin Zhu, Yang Ye, Munan Ning, Peng Jin, and Li Yuan. Video-LLaVA: Learning united visual representation by alignment before projection. arXiv:2311.10122, 2023. 3 + +[50] Kevin Lin, Linjie Li, Chung-Ching Lin, Faisal Ahmed, Zhe Gan, Zicheng Liu, Yumao Lu, and Lijuan Wang. SwinBERT: End-to-end transformers with sparse attention for video captioning. In CVPR, 2022. 2 +[51] Kevin Qinghong Lin, Pengchuan Zhang, Joya Chen, Shraman Pramanick, Difei Gao, Alex Jinpeng Wang, Rui Yan, and Mike Zheng Shou. UniVTG: Towards unified video-language temporal grounding. In ICCV, 2023. 2 +[52] Hao Liu, Matei Zaharia, and Pieter Abbeel. Ring attention with blockwise transformers for near-infinite context. arXiv:2310.01889, 2023. 3 +[53] Hao Liu, Wilson Yan, Matei Zaharia, and Pieter Abbeel. World model on million-length video and language with blockwise ringattention. arXiv:2402.08268, 2024. 3 +[54] Xiaolong Liu, Qimeng Wang, Yao Hu, Xu Tang, Shiwei Zhang, Song Bai, and Xiang Bai. End-to-end temporal action detection with transformer. In IEEE Transactions on Image Processing, 2022. 2 +[55] Huaishao Luo, Lei Ji, Botian Shi, Haoyang Huang, Nan Duan, Tianrui Li, Jason Li, Taroon Bharti, and Ming Zhou. UniVL: A unified video and language pre-training model for multimodal understanding and generation. arXiv:2002.06353, 2020. 1 +[56] Karttikeya Mangalam, Raiymbek Akshulakov, and Jitendra Malik. EgoSchema: A diagnostic benchmark for very long-form video language understanding. In NeurIPS Track on Datasets and Benchmarks, 2023. 1, 3 +[57] Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic. HowTo100M: Learning a text-video embedding by watching hundred million narrated video clips. In ICCV, 2019. 2, 5 +[58] Juhong Min, Shyamal Buch, Arsha Nagrani, Minsu Cho, and Cordelia Schmid. MoReVQA: Exploring modular reasoning models for video question answering. In CVPR, 2024. 2, 3 +[59] Jonghwan Mun, Minchul Shin, Gunsoo Han, Sangho Lee, Seongsu Ha, Joonseok Lee, and Eun-Sol Kim. BaSSL: Boundary-aware self-supervised learning for video scene segmentation. In ACCV, 2022. 2 +[60] Arsha Nagrani, Paul Hongsuck Seo, Bryan Seybold, Anja Hauth, Santiago Manen, Chen Sun, and Cordelia Schmid. Learning audio-video modalities from image captions. In ECCV, 2022. 2 +[61] Joe Yue-Hei Ng, Matthew Hausknecht, Sudheendra Vijayanarasimhan, Oriol Vinyals, Rajat Monga, and George Toderici. Beyond short snippets: Deep networks for video classification. In CVPR, 2015. 1 +[62] OpenAI. Gpt-4o system card. arXiv:2410.21276, 2024. 5, 6 +[63] Yingwei Pan, Ting Yao, Houqiang Li, and Tao Mei. Video captioning with transferred semantic attributes. In CVPR, 2017. 1 +[64] Pinelopi Papalampidi, Skanda Koppula, Shreya Pathak, Justin Chiu, Joe Heyward, Viorica Patraucean, Jiajun Shen, Antoine Miech, Andrew Zisserman, and Aida Nematzdeh. A simple recipe for contrastively pre-training video-first encoders beyond 16 frames. In CVPR, 2024. 3 +[65] Alejandro Pardo, Fabian Caba Heilbron, Juan León Alcázar, Ali Thabet, and Bernard Ghanem. Moviecuts: A new dataset and benchmark for cut type recognition. In ECCV, 2022. 2 +[66] Jinwoo Park, Jungeun Kim, Jaegwang Seok, Sukhyun Lee, and Junyeong Kim. Contrasting multi-modal similarity framework for video scene segmentation. IEEE Access, 2024. 2 + +[67] Wenjie Pei, Jiyuan Zhang, Xiangrong Wang, Lei Ke, Xiaoyong Shen, and Yu-Wing Tai. Memory-attended recurrent network for video captioning. In CVPR, 2019. 1 +[68] Alec Radford, Karthik Narasimhan, Tim Salimans, and Ilya Sutskever. Improving language understanding by generative pre-training. Preprint, 2018. 3 +[69] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 3 +[70] Alec Radford, Jong Wook Kim, Tao Xu, Greg Brockman, Christine McLeavey, and Ilya Sutskever. Robust speech recognition via large-scale weak supervision. In ICML, 2023. 4, 5 +[71] Anyi Rao, Linning Xu, Yu Xiong, Guodong Xu, Qingqiu Huang, Bolei Zhou, and Dahua Lin. A local-to-global approach to multi-modal movie scene segmentation. In CVPR, 2020. 2 +[72] Zeeshan Rasheed and Mubarak Shah. Scene detection in hollywood movies and tv shows. In CVPR, 2003. 2 +[73] Fabian Retkowski and Alexander Waibel. From text segmentation to smart chaptering: A novel benchmark for structuring video transcriptions. arXiv:2402.17633, 2024. 2 +[74] Marcus Rohrbach, Wei Qiu, Ivan Titov, Stefan Thater, Manfred Pinkal, and Bernt Schiele. Translating video content to natural language descriptions. In ICCV, 2013. 1 +[75] Daniel Rotman, Dror Porat, and Gal Ashour. Robust video scene detection using multimodal fusion of optimally grouped features. IEEE 19th International Workshop on Multimedia Signal Processing (MMSP), 2017. 2 +[76] Yong Rui, Thomas S Huang, and Sharad Mehrotra. Exploring video structure beyond the shots. In IEEE International Conference on Multimedia Computing and Systems, 1998. 2 +[77] Najmeh Sadoughi, Xinyu Li, Avijit Vajpayee, David Fan, Bing Shuai, Hector Santos-Villalobos, Vimal Bhat, and Rohith MV. MEGA: Multimodal alignment aggregation and distillation for cinematic video segmentation. In ICCV, 2023. 2 +[78] Paul Hongsuck Seo, Arsha Nagrani, Anurag Arnab, and Cordelia Schmid. End-to-end generative pretraining for multimodal video captioning. In CVPR, 2022. 1, 2 +[79] Freda Shi, Xinyun Chen, Kanishka Misra, Nathan Scales, David Dohan, Ed H. Chi, Nathanael Schärli, and Denny Zhou. Large language models can be easily distracted by irrelevant context. In ICML, 2023. 3, 7 +[80] Nina Shvetsova, Anna Kukleva, Xudong Hong, Christian Rupprecht, Bernt Schiele, and Hilde Kuehne. HowToCaption: Prompting llms to transform video annotations at scale. In ECCV, 2024. 2, 3 +[81] Panagiotis Sidiropoulos, Vasileios Mezaris, Ioannis Kompat-siaris, Hugo Meinedo, Miguel Bugalho, and Isabel Trancoso. Temporal video segmentation to scenes using high-level audiovisual features. IEEE TCSVT, 2011. 2 +[82] Gunnar A. Sigurdsson, Gül Varol, Xiaolong Wang, Ivan Laptev, Ali Farhadi, and Abhinav Gupta. Hollywood in homes: Crowdsourcing data collection for activity understanding. In ECCV, 2016. 1 +[83] Karen Simonyan and Andrew Zisserman. Two-stream convolutional networks for action recognition in videos. In ICLR, 2014. 1 +[84] Mattia Soldan, Alejandro Pardo, Juan León Alcazar, Fabian Caba, Chen Zhao, Silvio Giancola, and Bernard Ghanem. + +MAD: A scalable dataset for language grounding in videos from movie audio descriptions. In CVPR, 2022. 1, 3 +[85] Chen Sun, Austin Myers, Carl Vondrick, Kevin Murphy, and Cordelia Schmid. VideoBERT: A joint model for video and language representation learning. In ICCV, 2019. 1 +[86] Yuchong Sun, Hongwei Xue, Ruihua Song, Bei Liu, Huan Yang, and Jianlong Fu. Long-form video-language pre-training with multimodal temporal contrastive learning. In NeurIPS, 2022. 3 +[87] Ilya Sutskever, Oriol Vinyals, and Quoc V Le. Sequence to sequence learning with neural networks. NeurIPS, 2014. 1 +[88] Reuben Tan, Xineng Sun, Ping Hu, Jui hsien Wang, Hanieh Deilamsalehy, Bryan A. Plummer, Bryan Russell, and Kate Saenko. Koala: Key frame-conditioned long video-LLM. In CVPR, 2024. 3 +[89] Gemini Team, Petko Georgiev, Ving Ian Lei, Ryan Burnell, Libin Bai, Anmol Gulati, Garrett Tanzer, Damien Vincent, Zhufeng Pan, Shibo Wang, et al. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context. arXiv:2403.05530, 2024. 3 +[90] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. LLaMA: Open and efficient foundation language models. arXiv:2302.13971, 2023. 3 +[91] Hugo Touvron et al. LLaMA 2: Open foundation and fine-tuned chat models. arXiv:2307.09288, 2023. 3 +[92] Du Tran, Lubomir Bourdev, Rob Fergus, Lorenzo Torresani, and Manohar Paluri. Learning spatiotemporal features with 3d convolutional networks. In ICCV, 2015. 1 +[93] Gül Varol, Ivan Laptev, and Cordelia Schmid. Long-term temporal convolutions for action recognition. TPAMI, 2018. 1 +[94] Ramakrishna Vedantam, C Lawrence Zitnick, and Devi Parikh CIDER: Consensus-based image description evaluation. In CVPR, 2015. 5 +[95] Lucas Ventura, Cordelia Schmid, and Gül Varol. Learning text-to-video retrieval from image captioning. *IJCV*, 2024. 2 +[96] Lucas Ventura, Antoine Yang, Cordelia Schmid, and Güll Varol. CoVR-2: Automatic data construction for composed video retrieval. TPAMI, 2024. 3 +[97] Caroline Violot, Tugrulcan Elmas, Igor Bilogrevic, and Mathias Humbert. Shorts vs. regular videos on YouTube: A comparative analysis of user engagement and content creation trends. In Proceedings of the 16th ACM Web Science Conference, Association for Computing Machinery, 2024. 1 +[98] Junke Wang, Dongdong Chen, Zuxuan Wu, Chong Luo, Luowei Zhou, Yucheng Zhao, Yujia Xie, Ce Liu, Yu-Gang Jiang, and Lu Yuan. Omnivl: One foundation model for image-language and video-language tasks. NeurIPS, 2022. 1 +[99] Teng Wang, Ruimao Zhang, Zhichao Lu, Feng Zheng, Ran Cheng, and Ping Luo. End-to-end dense video captioning with parallel decoding. In ICCV, 2021. 2 +[100] Yuxuan Wang, Difei Gao, Licheng Yu, Weixian Lei, Matt Feiszli, and Mike Zheng Shou. GEB+: A benchmark for generic event boundary captioning, grounding and retrieval. In ECCV, 2022. 2 +[101] Ziyang Wang, Shoubin Yu, Elias Stengel-Eskin, Jaehong Yoon, Feng Cheng, Gedas Bertasius, and Mohit Bansal. VideoTree: Adaptive tree-based video representation for LLM reasoning on long videos. arXiv:2405.19209, 2024. 3, 7 + +[102] Xi Wei, Zhangxiang Shi, Tianzhu Zhang, Xiaoyuan Yu, and Lei Xiao. Multimodal high-order relation transformer for scene boundary detection. In ICCV, 2023. 2 +[103] Chao-Yuan Wu, Yanghao Li, Karttikeya Mangalam, Haoqi Fan, Bo Xiong, Jitendra Malik, and Christoph Feichtenhofer. MeMViT: Memory-augmented multiscale vision transformer for efficient long-term video recognition. In CVPR, 2022. 3 +[104] Haoqian Wu, Keyu Chen, Haozhe Liu, Mingchen Zhuge, Bing Li, Ruizhi Qiao, Xiujun Shu, Bei Gan, Liangsheng Xu, Bo Ren, Mengmeng Xu, Wentian Zhang, Raghavendra Ramachandra, Chia-Wen Lin, and Bernard Ghanem. NewsNet: A novel dataset for hierarchical temporal segmentation. In CVPR, 2023. 3 +[105] Zuxuan Wu, Caiming Xiong, Chih-Yao Ma, Richard Socher, and Larry S. Davis. AdaFrame: Adaptive frame selection for fast video recognition. In CVPR, 2019. 3 +[106] Junbin Xiao, Xindi Shang, Angela Yao, and Tat-Seng Chua. NExT-QA: Next phase of question-answering to explaining temporal actions. In CVPR, 2021. 3 +[107] Jun Xu, Tao Mei, Ting Yao, and Yong Rui. MSR-VTT: A large video description dataset for bridging video and language. In CVPR, 2016. 2 +[108] Hongwei Xue, Tiankai Hang, Yanhong Zeng, Yuchong Sun, Bei Liu, Huan Yang, Jianlong Fu, and Baining Guo. Advancing high-resolution video-language representation with large-scale video transcriptions. In CVPR, 2022. 3 +[109] Antoine Yang, Arsha Nagrani, Ivan Laptev, Josef Sivic, and Cordelia Schmid. VidChapters-7M: Video chapters at scale. In NeurIPS Track on Datasets and Benchmarks, 2023. 1, 2, 3, 4, 5, 6, 7 +[110] Antoine Yang, Arsha Nagrani, Paul Hongsuck Seo, Antoine Miech, Jordi Pont-Tuset, Ivan Laptev, Josef Sivic, and Cordelia Schmid. Vid2Seq: Large-scale pretraining of a visual language model for dense video captioning. In CVPR, 2023. 1, 2, 3, 5, 6, 7 +[111] Yang Yang, Yurui Huang, Weili Guo, Baohua Xu, and Dingyin Xia. Towards global video scene segmentation with context-aware transformer. AAAI, 2023. 2 +[112] Yuan Yao, Tianyu Yu, Ao Zhang, Chongyi Wang, Junbo Cui, Hongji Zhu, Tianchi Cai, Haoyu Li, Weilin Zhao, Zhihui He, Qianyu Chen, Huarong Zhou, Zhensheng Zou, Haoye Zhang, Shengding Hu, Zhi Zheng, Jie Zhou, Jie Cai, Xu Han, Guoyang Zeng, Dahai Li, Zhiyuan Liu, and Maosong Sun. MiniCPM-V: A GPT-4V level MLLM on your phone. arXiv:2408.01800, 2024. 4 +[113] Fangqiu Yi, Hongyu Wen, and Tingting Jiang. ASformer: Transformer for action segmentation. In BMVC, 2021. 2 +[114] Shoubin Yu, Jaemin Cho, Prateek Yadav, and Mohit Bansal. Self-chained image-language model for video localization and question answering. In NeurIPS, 2023. 3 +[115] Rowan Zellers, Ximing Lu, Jack Hessel, Youngjae Yu, Jae Sung Park, Jize Cao, Ali Farhadi, and Yejin Choi. MERLOT: Multimodal neural script knowledge models. In NeurIPS, 2021. 3 +[116] Andy Zeng, Maria Attarian, Brian Ichter, Krzysztof Choromanski, Adrian Wong, Stefan Welker, Federico Tombari, Aveek Purohit, Michael Ryoo, Vikas Sindhwani, et al. Socratic models: Composing zero-shot multimodal reasoning with language. In ICLR, 2023. 2 +[117] Kuo-Hao Zeng, Tseng-Hung Chen, Juan Carlos Niebles, and Min Sun. Title generation for user generated videos. In ECCV, 2016. 2 + +[118] Runhao Zeng, Wenbing Huang, Mingkui Tan, Yu Rong, Peilin Zhao, Junzhou Huang, and Chuang Gan. Graph convolutional networks for temporal action localization. In CVPR, 2019. 2 +[119] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. ICCV, 2023. 8 +[120] Chenlin Zhang, Jianxin Wu, and Yin Li. ActionFormer: Localizing moments of actions with transformers. In ECCV, 2022. 2 +[121] Ce Zhang, Taixi Lu, Md Mohaiminul Islam, Ziyang Wang, Shoubin Yu, Mohit Bansal, and Gedas Bertasius. A simple LLM framework for long-range video question-answering. In EMNLP, 2024. 2, 3 +[122] Hang Zhang, Xin Li, and Lidong Bing. Video-LLaMA: An instruction-tuned audio-visual language model for video understanding. In EMNLP Demo, 2023. 3 +[123] Shengyu Zhang, Ziqi Tan, Zhou Zhao, Jin Yu, Kun Kuang, Tan Jiang, Jingren Zhou, Hongxia Yang, and Fei Wu. Comprehensive information integration modeling framework for video titling. In ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, 2020. 2 +[124] Cairong Zhao, Chutian Wang, Zifan Song, Guosheng Hu, Hao-nan Chen, and Xiaofan Zhai. Cap2Sum: Learning to summarize videos by generating captions. arXiv:2408.12800, 2024. 2 +[125] Junjie Zhou, Yan Shu, Bo Zhao, Boya Wu, Shitao Xiao, Xi Yang, Yongping Xiong, Bo Zhang, Tiejun Huang, and Zheng Liu. MLVU: A comprehensive benchmark for multi-task long video understanding. arXiv:2406.04264, 2024. 3 +[126] Luowei Zhou, Xu Chenliang, and Jason J. Corso. Towards automatic learning of procedures from web instructional videos. In AAAI, 2018. 1 +[127] Luowei Zhou, Yingbo Zhou, Jason J Corso, Richard Socher, and Caiming Xiong. End-to-end dense video captioning with masked transformer. In CVPR, 2018. 2 +[128] Xingyi Zhou, Anurag Arnab, Shyamal Buch, Shen Yan, Austin Myers, Xuehan Xiong, Arsha Nagrani, and Cordelia Schmid. Streaming dense video captioning. In CVPR, 2024. 2 \ No newline at end of file diff --git a/2025/Chapter-Llama_ Efficient Chaptering in Hour-Long Videos with LLMs/images.zip b/2025/Chapter-Llama_ Efficient Chaptering in Hour-Long Videos with LLMs/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..9bce066a204e59cb92badfa5ecd4ff98694b11f9 --- /dev/null +++ b/2025/Chapter-Llama_ Efficient Chaptering in Hour-Long Videos with LLMs/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c9a16f7e08a75213ef91d809ee8b834a08e5e837e5322f2ede677a55e700763e +size 570978 diff --git a/2025/Chapter-Llama_ Efficient Chaptering in Hour-Long Videos with LLMs/layout.json b/2025/Chapter-Llama_ Efficient Chaptering in Hour-Long Videos with LLMs/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..9488af944e463c74d6fe89e2ba56fcc107f869e1 --- /dev/null +++ b/2025/Chapter-Llama_ Efficient Chaptering in Hour-Long Videos with LLMs/layout.json @@ -0,0 +1,8642 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 96, + 102, + 515, + 120 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 102, + 515, + 120 + ], + "spans": [ + { + "bbox": [ + 96, + 102, + 515, + 120 + ], + "type": "text", + "content": "Chapter-Llama: Efficient Chaptering in Hour-Long Videos with LLMs" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 141, + 217, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 141, + 217, + 156 + ], + "spans": [ + { + "bbox": [ + 138, + 141, + 217, + 156 + ], + "type": "text", + "content": "Lucas Ventura1,2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 232, + 142, + 302, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 232, + 142, + 302, + 156 + ], + "spans": [ + { + "bbox": [ + 232, + 142, + 302, + 156 + ], + "type": "text", + "content": "Antoine Yang3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 317, + 142, + 400, + 155 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 142, + 400, + 155 + ], + "spans": [ + { + "bbox": [ + 317, + 142, + 400, + 155 + ], + "type": "text", + "content": "Cordelia Schmid2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 416, + 142, + 466, + 155 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 416, + 142, + 466, + 155 + ], + "spans": [ + { + "bbox": [ + 416, + 142, + 466, + 155 + ], + "type": "text", + "content": "Gül Varol" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 197, + 164, + 413, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 197, + 164, + 413, + 176 + ], + "spans": [ + { + "bbox": [ + 197, + 164, + 413, + 176 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 197, + 164, + 413, + 176 + ], + "type": "text", + "content": "LIGM, École des Ponts, IP Paris, Univ Gustave Eiffel, CNRS" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 151, + 178, + 459, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 178, + 459, + 190 + ], + "spans": [ + { + "bbox": [ + 151, + 178, + 459, + 190 + ], + "type": "text", + "content": "2Inria, École normale supérieure, CNRS, PSL Research University 3Google DeepMind" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 165, + 193, + 441, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 193, + 441, + 204 + ], + "spans": [ + { + "bbox": [ + 165, + 193, + 441, + 204 + ], + "type": "text", + "content": "https://imagine.enpc.fr/~lucas.ventura/chapter-llama/" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 153, + 232, + 199, + 244 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 153, + 232, + 199, + 244 + ], + "spans": [ + { + "bbox": [ + 153, + 232, + 199, + 244 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 54, + 258, + 297, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 258, + 297, + 515 + ], + "spans": [ + { + "bbox": [ + 54, + 258, + 297, + 515 + ], + "type": "text", + "content": "We address the task of video chaptering, i.e., partitioning a long video timeline into semantic units and generating corresponding chapter titles. While relatively underexplored, automatic chaptering has the potential to enable efficient navigation and content retrieval in long-form videos. In this paper, we achieve strong chaptering performance on hour-long videos by efficiently addressing the problem in the text domain with our 'Chapter-Llama' framework. Specifically, we leverage a pretrained large language model (LLM) with large context window, and feed as input (i) speech transcripts and (ii) captions describing video frames, along with their respective timestamps. Given the inefficiency of exhaustively captioning all frames, we propose a lightweight speech-guided frame selection strategy based on speech transcript content, and experimentally demonstrate remarkable advantages. We train the LLM to output timestamps for the chapter boundaries, as well as free-form chapter titles. This simple yet powerful approach scales to processing one-hour long videos in a single forward pass. Our results demonstrate substantial improvements (e.g., 45.3 vs " + }, + { + "bbox": [ + 54, + 258, + 297, + 515 + ], + "type": "inline_equation", + "content": "26.7\\%" + }, + { + "bbox": [ + 54, + 258, + 297, + 515 + ], + "type": "text", + "content": " F1 score) over the state of the art on the recent VidChapters-7M benchmark. To promote further research, we release our code and models at our project page." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 540, + 132, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 540, + 132, + 552 + ], + "spans": [ + { + "bbox": [ + 55, + 540, + 132, + 552 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 560, + 296, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 560, + 296, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 560, + 296, + 677 + ], + "type": "text", + "content": "According to a study by [46], the video durations uploaded to the popular online video sharing platform YouTube have increased steadily over the years. Videos have become longer since the first video upload in 2005 [19, 47]. In 2020, " + }, + { + "bbox": [ + 55, + 560, + 296, + 677 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 55, + 560, + 296, + 677 + ], + "type": "text", + "content": " of videos were estimated to be longer than 15 minutes, " + }, + { + "bbox": [ + 55, + 560, + 296, + 677 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 55, + 560, + 296, + 677 + ], + "type": "text", + "content": " more than 3 hours [46]. Long-form videos such as news, sports, educational, and vlog streams can often span extensive durations and cover multiple topics [97]. Finding specific content within increased video duration and volume makes efficient content navigation more important than ever." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 678, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 678, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 678, + 296, + 713 + ], + "type": "text", + "content": "However, much of the traditional video analysis research has focused on processing short videos of a few seconds [4, 15, 34, 55, 63, 67, 74, 78, 85, 87, 98, 110]. At the same" + } + ] + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 317, + 232, + 552, + 393 + ], + "blocks": [ + { + "bbox": [ + 317, + 232, + 552, + 393 + ], + "lines": [ + { + "bbox": [ + 317, + 232, + 552, + 393 + ], + "spans": [ + { + "bbox": [ + 317, + 232, + 552, + 393 + ], + "type": "image", + "image_path": "44bed3c22be373cb339c9910f1657007af24ef60b8b4665f684bfc137a87e205.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 397, + 555, + 472 + ], + "lines": [ + { + "bbox": [ + 313, + 397, + 555, + 472 + ], + "spans": [ + { + "bbox": [ + 313, + 397, + 555, + 472 + ], + "type": "text", + "content": "Figure 1. Chapter-Llama: Our method generates automatic video chapters for hour-long videos by training a large language model (LLM) to predict chapter boundaries and titles. The LLM processes transcribed speech (ASR) and descriptive captions of key frames, which are sampled based on ASR content. This text-based approach, equipped with speech-based frame selection, enables efficient processing of long-form content." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 491, + 556, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 491, + 556, + 677 + ], + "spans": [ + { + "bbox": [ + 313, + 491, + 556, + 677 + ], + "type": "text", + "content": "time, the definition of long videos has changed within the past decade. Early works claimed processing 100 frames (i.e., a few seconds) to be long [61, 93] as opposed to ingesting up to 16 frames [83, 92]. With the introduction of datasets containing 1-5 minute videos [29, 37, 44, 56, 82, 126], several minutes were considered very long. Studying hour-long videos has only recently seen an interest in the context of movie description [31], video captioning [40], or grounding [32, 84]. Very recently, the work of [109] collected the VidChapters-7M dataset with videos spanning from minutes to hours, along with their user-defined video chapters, and proposed the video chapter generation task, automatically dividing a video into thematic sections (i.e., chapters) with descriptive concise chapter titles. Video chaptering, if achieved successfully, can offer a compelling solution to long content indexing, bypassing the current need for time-consuming manual annotation by video owners [109]." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 678, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 678, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 678, + 556, + 715 + ], + "type": "text", + "content": "In this paper, we address the challenge of automatic video chaptering with a simple yet effective framework designed to handle hour-long videos. Existing work for chaptering [109]" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18947" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 210 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 210 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 210 + ], + "type": "text", + "content": "relics on a dense video captioning model Vid2Seq [110], which combines multimodal inputs from video frames and ASR-based speech transcriptions. However, Vid2Seq operates on a fixed number of equally sampled frames (i.e., 100 frames), potentially missing important visual information. Furthermore, their approach based on transformer architecture uses video frame features directly, which requires learning a mapping from the visual modality to the textual modality. In contrast, our method is designed to address these limitations by (i) dynamically sampling keyframes from the video based on the speech content, and (ii) designing a purely text-based model leveraging image captioning to convert RGB frames into text." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 213, + 297, + 585 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 213, + 297, + 585 + ], + "spans": [ + { + "bbox": [ + 56, + 213, + 297, + 585 + ], + "type": "text", + "content": "Our approach leverages a pretrained LLM, which we finetune specifically for the video chaptering task to predict jointly the chapter boundary timestamps and chapter titles, both in text form. The appeal of our model lies in processing only textual data as input, allowing us effectively leverage the long-context understanding capabilities of the LLM to scale to long videos. In particular, we incorporate speech transcriptions from automatic speech recognition (ASR) and automatic frame captions. Captioning has been used for video understanding as an intermediate representation in recent works, but in the context of retrieval or question answering (QA) for shorter videos (maximum 3 minutes) [58, 95, 116, 121]. In longer videos, since captioning every frame is computationally prohibitive, we employ a speech-based frame selection strategy that scales efficiently while preserving important content. Similar in spirit to [43], we primarily use audio to determine keyframes, specifically bootstrapping with an LLM trained only with the speech inputs. However, even when transforming a video into text, LLMs have a limited context window, allowing a maximum number of tokens as input in a single forward pass. To mitigate context window limitations for very long video inputs, we simply perform an iterative prediction, sequentially processing the video, where each iteration typically operates on a window length of about an hour duration. We evaluate our approach on 'short' (0-15 min), 'medium' (15-30 min), and 'long' (30-60 min) videos from the VidChapters-7M dataset [109], demonstrating significant improvements over the state of the art across multiple metrics, including temporal boundary accuracy and semantic relevance of chapter titles. Our experiments show that finetuning the LLM, our speech-based frame selection strategy, and the integration of modalities from both speech and captions are crucial for achieving high-quality video chaptering results." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 586, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 586, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 586, + 295, + 713 + ], + "type": "text", + "content": "Our contributions are the following: (i) We introduce Chapter-Llama: our framework leverages a pretrained LLM and finetunes for the underexplored task of video chaptering by transforming the video input into text form through ASR and captioning. (ii) We scale efficiently to hour-long videos by incorporating a speech-based frame sampling strategy, captioning only a subset of the video frames. (iii) Our simple and effective approach outperforms the state of the art on the recent VidChapters-7M benchmark by a large margin (e.g., 45.3 vs 26.7 F1 score). These results are complemented by a comprehensive set of experiments analyzing our components." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 314, + 71, + 397, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 71, + 397, + 84 + ], + "spans": [ + { + "bbox": [ + 314, + 71, + 397, + 84 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 91, + 555, + 138 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 91, + 555, + 138 + ], + "spans": [ + { + "bbox": [ + 313, + 91, + 555, + 138 + ], + "type": "text", + "content": "We provide an overview of video tasks related to video chaptering, such as temporal segmentation and captioning, along with a discussion on works focusing on long-form and LLM-based video understanding." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 148, + 556, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 148, + 556, + 426 + ], + "spans": [ + { + "bbox": [ + 313, + 148, + 556, + 426 + ], + "type": "text", + "content": "Temporal video segmentation. While video chaptering is a new task [109], there is a rich literature on methods focused on temporally segmenting a video in various forms. One task is shot detection [72, 76, 81], where any visual changes (e.g., shifting between two cameras) would require a temporal boundary, not necessarily modeling semantic shifts. Video scene segmentation, often studied on movies [38], is primarily focusing on grouping scenes with similar content [13, 14, 38, 39, 59, 65, 66, 71, 75, 77, 102, 111]. Another line of work considers boundary detection for temporal action segmentation [7, 23, 26, 48, 113], or localization [18, 54, 118, 120]. Unlike chaptering with free-form text, action segmentation assigns a label from a predefined set of categories, and typically defines short atomic actions as the unit. In contrast to these tasks, chapter boundaries can take various different forms depending on the type and the granularity of the video (e.g., each exercise within sports video, each slide within a lecture, each step in instructional video, each topic in a podcast video). Shot, scene, or action boundaries therefore may or may not correspond to complex chapter boundary definitions. Moreover, these tasks are mostly tackled with vision-only inputs [81, 113, 120], without leveraging speech. While text and audio segmentation have also been tackled separately [28, 73], video chaptering is based on both audio and vision inputs [109]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 436, + 556, + 680 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 436, + 556, + 680 + ], + "spans": [ + { + "bbox": [ + 313, + 436, + 556, + 680 + ], + "type": "text", + "content": "Video captioning. Generating chapter titles [109] is relevant to the task of captioning that seeks to describe the video content with text. There is a large literature on single video captioning [16, 50, 78, 80], often focusing on short video clips. Typical datasets for training such as MSR-VTT [107], WebVid [5], HowTo100M [57], Video-CC [60] include captions of videos spanning a few seconds (5-15sec on average). In generic event boundary captioning [100], event intervals are similarly short, in the order of 2 seconds. On the other hand, video summarization methods operate on longer videos; however, their goal is to reduce the entire video into a single summary description [1, 2, 33, 40, 51, 117, 123, 124], not necessarily with a temporal segmentation component. Dense video captioning [37, 44, 99, 110, 127, 128] is the closest to video chaptering in terms of problem formulation, aiming to both temporally localize and caption different events. Indeed, prior work on video chaptering trains the dense captioning method of Vid2Seq [110] on the VidChapters-7M dataset [109], but relies on a fixed number of equally sampled frames. In this paper, we leverage some of the annotations of this dataset to train an LLM-based chaptering model substantially outperforming previous methods [109, 110]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "content": "Long-form video understanding. The definition of long videos has evolved with the release of various datasets spanning" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "18948" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 281 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 281 + ], + "type": "text", + "content": "seconds [106, 108], a few minutes [22, 29, 56, 86], 10-30 minutes [2, 125], or one hour [24, 40, 84, 104, 109]. MLVU [125] introduces a benchmark for evaluating multiple long video understanding tasks such as summarization and QA; however, the data is not suitable for chaptering due to lack of annotations. Video-MME [24] also contains hour-long videos for QA. MAD [31, 84] provides audio description for long movies, but each description spans a few seconds and the sparse coverage over the video is different than contiguous chapters. Recently, Ego4D-HCap [40] was proposed for hierarchical video summarization. However, this dataset involves dense captioning with visual inputs only, while we focus on video chaptering with visual and speech inputs. To the best of our knowledge, VidChapters7M [109] is the only open-sourced dataset for training and evaluating chapter generation, which we employ in this paper. Nonpublic related datasets include NewsNet [104] which includes hierarchical temporal segmentation annotations, the TV news chaptering dataset used in [30], and the ChapterGen dataset [10]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 281, + 298, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 281, + 298, + 502 + ], + "spans": [ + { + "bbox": [ + 55, + 281, + 298, + 502 + ], + "type": "text", + "content": "Increased video lengths led to a range of works focusing on efficient temporal modeling strategies. A common technique to deal with longer videos is to use pre-extracted visual features [31, 84, 115]. For end-to-end learning with transformers, several works explored factorized spatio-temporal attention [3, 5, 8]. Others have looked at various ways to incorporate a memory mechanisms [42, 103], blockwise attention [52, 53], or captioning frames to exploit LLMs [101, 121]. Given the redundancy in consecutive video frames, frame selection methods were explored in the context of short video captioning and action recognition [17, 105], as well as 'long' video QA in 3-minute durations [64, 88, 114]. Most common approach with current large video models is to perform sparse sampling with equal spacing [12, 45, 110]. SCSampler [43] exploits the low-dimensional audio modality to efficiently select salient video clips for action recognition. In our method, we also leverage audio, but in the form of ASR, and run the costly frame captioning step only on keyframes on locations predicted by a speech-based frame selection module." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 516, + 298, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 516, + 298, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 516, + 298, + 714 + ], + "type": "text", + "content": "LLM use in video understanding. LLMs such as GPT [9, 68], Llama [20, 90, 91], and Gemini [27, 89], have been leveraged in different ways for improving video understanding. A popular approach is to train 'bridge' modules between pretrained visual backbones [69] and LLMs to build vision-language models (VLMs) that can ingest videos (e.g., Video-Llama [122], Video-LLaVa [49]). Other works have employed LLMs for automatic construction of video datasets [2, 40, 80, 96], tool use [58], storing memory in video QA [42], and temporal localization [36]. Similar to us, VideoTree [101] and VideoAgent [21] caption keyframes before passing them to an LLM together with a question for answer generation, addressing the limitations of [121] which performs a similar methodology without keyframe selection on shorter videos. In this study, we find that captioning alone is not sufficient, and needs to be complemented with ASR for competitive chaptering performance. Close to us, [2] exploits ASR on long videos and summarizes" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 313, + 72, + 556, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 556, + 144 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 556, + 144 + ], + "type": "text", + "content": "them with LLMs to generate pseudo-labels for video summarization training. In our work, we leverage LLMs, specifically finetuning a Llama model [20] for chaptering by prompting with speech transcription and frame captions. We show that finetuning is essential for adapting to the task so that the LLM picks up relevant content within the large context input [79]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 152, + 555, + 166 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 152, + 555, + 166 + ], + "spans": [ + { + "bbox": [ + 313, + 152, + 555, + 166 + ], + "type": "text", + "content": "3. Chapter-Llama: LLM-based Video Chaptering" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 171, + 556, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 171, + 556, + 312 + ], + "spans": [ + { + "bbox": [ + 313, + 171, + 556, + 312 + ], + "type": "text", + "content": "We provide an overview of our video chaptering framework, referred to as Chapter-Llama, in Fig. 2. Given video frames and speech transcripts, we aim at predicting relevant chapter boundaries and titles. For this, we first select video frames to process with a speech-based frame selection module. Then we use an off-the-shelf visual captioner to map the selected frames in the text space. We feed the resulting captions, along with speech transcripts, to the LLM which outputs the chapter boundaries and titles jointly as a single sequence of tokens. Finally, we devise an iterative prediction procedure in case the input text sequence is too long to handle for the LLM. We next describe in more detail each component." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 319, + 556, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 319, + 556, + 471 + ], + "spans": [ + { + "bbox": [ + 313, + 319, + 556, + 471 + ], + "type": "text", + "content": "Task formulation. Video chaptering [109] aims at segmenting a video into semantically meaningful chapters, and generating a title for each segment. The chapters are contiguous, with no gaps between them, and together span the entire video duration from start to end. Formally, given video frames " + }, + { + "bbox": [ + 313, + 319, + 556, + 471 + ], + "type": "inline_equation", + "content": "V = (v_{1}, v_{2}, \\dots, v_{N})" + }, + { + "bbox": [ + 313, + 319, + 556, + 471 + ], + "type": "text", + "content": " and temporally-aligned speech transcripts " + }, + { + "bbox": [ + 313, + 319, + 556, + 471 + ], + "type": "inline_equation", + "content": "S = (s_{1}, s_{2}, \\dots, s_{M})" + }, + { + "bbox": [ + 313, + 319, + 556, + 471 + ], + "type": "text", + "content": ", where each speech transcript contains an utterance and its associated start and end timestamps, the task is to output a sequence of chapters " + }, + { + "bbox": [ + 313, + 319, + 556, + 471 + ], + "type": "inline_equation", + "content": "C = (c_{1}, c_{2}, \\dots, c_{L})" + }, + { + "bbox": [ + 313, + 319, + 556, + 471 + ], + "type": "text", + "content": ", where each chapter " + }, + { + "bbox": [ + 313, + 319, + 556, + 471 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 313, + 319, + 556, + 471 + ], + "type": "text", + "content": " is a tuple " + }, + { + "bbox": [ + 313, + 319, + 556, + 471 + ], + "type": "inline_equation", + "content": "(b_{i}, t_{i})" + }, + { + "bbox": [ + 313, + 319, + 556, + 471 + ], + "type": "text", + "content": " containing a start timestamp " + }, + { + "bbox": [ + 313, + 319, + 556, + 471 + ], + "type": "inline_equation", + "content": "b_{i}" + }, + { + "bbox": [ + 313, + 319, + 556, + 471 + ], + "type": "text", + "content": " and a descriptive title " + }, + { + "bbox": [ + 313, + 319, + 556, + 471 + ], + "type": "inline_equation", + "content": "t_{i}" + }, + { + "bbox": [ + 313, + 319, + 556, + 471 + ], + "type": "text", + "content": ". The end time of chapter " + }, + { + "bbox": [ + 313, + 319, + 556, + 471 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 319, + 556, + 471 + ], + "type": "text", + "content": " is implicitly defined by the start time of the subsequent chapter " + }, + { + "bbox": [ + 313, + 319, + 556, + 471 + ], + "type": "inline_equation", + "content": "b_{i+1}" + }, + { + "bbox": [ + 313, + 319, + 556, + 471 + ], + "type": "text", + "content": ", or total video duration if " + }, + { + "bbox": [ + 313, + 319, + 556, + 471 + ], + "type": "inline_equation", + "content": "i = L" + }, + { + "bbox": [ + 313, + 319, + 556, + 471 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 479, + 556, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 479, + 556, + 594 + ], + "spans": [ + { + "bbox": [ + 313, + 479, + 556, + 594 + ], + "type": "text", + "content": "Speech-based frame selection. Video chaptering involves processing hour-long videos. Therefore, densely sampling frames is computationally intractable due to numerous inference passes through a vision model (e.g., a visual captioner) and exceeding standard LLM context lengths. Upon inspection of our data, we found that while the speech transcription has 257 tokens per minute on average, a caption is 66 tokens long on average hence captions would take 3,960 tokens per minute when sampling a video at 1 FPS. To address these challenges, we employ a frame selection strategy." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 594, + 557, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 557, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 557, + 715 + ], + "type": "text", + "content": "Specifically, we use speech transcripts to guide which video frames to process for the vision model. This is done by first training a speech-only variant of our LLM to predict a sequence of chapter boundaries " + }, + { + "bbox": [ + 313, + 594, + 557, + 715 + ], + "type": "inline_equation", + "content": "\\{\\hat{b}_1,\\hat{b}_2,\\dots,\\hat{b}_K\\}" + }, + { + "bbox": [ + 313, + 594, + 557, + 715 + ], + "type": "text", + "content": " from speech transcripts " + }, + { + "bbox": [ + 313, + 594, + 557, + 715 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 313, + 594, + 557, + 715 + ], + "type": "text", + "content": " only. For each predicted boundary " + }, + { + "bbox": [ + 313, + 594, + 557, + 715 + ], + "type": "inline_equation", + "content": "\\hat{b}_i" + }, + { + "bbox": [ + 313, + 594, + 557, + 715 + ], + "type": "text", + "content": ", we sample a frame " + }, + { + "bbox": [ + 313, + 594, + 557, + 715 + ], + "type": "inline_equation", + "content": "v_{i}" + }, + { + "bbox": [ + 313, + 594, + 557, + 715 + ], + "type": "text", + "content": " from the video at that timestamp. Note that this variant is cheaper compared to the full model as it only needs ASR transcription from the audio stream, without requiring any processing of the RGB stream (i.e., captioning). We then process the video frames only at the time locations predicted" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18949" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 68, + 550, + 246 + ], + "blocks": [ + { + "bbox": [ + 59, + 68, + 550, + 246 + ], + "lines": [ + { + "bbox": [ + 59, + 68, + 550, + 246 + ], + "spans": [ + { + "bbox": [ + 59, + 68, + 550, + 246 + ], + "type": "image", + "image_path": "d1b29773e7e14bc67b1a6e1d1b5d7b7983d9bbb7886a695c0b73e34943d3f3be.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 250, + 555, + 283 + ], + "lines": [ + { + "bbox": [ + 55, + 250, + 555, + 283 + ], + "spans": [ + { + "bbox": [ + 55, + 250, + 555, + 283 + ], + "type": "text", + "content": "Figure 2. Method overview: Our Chapter-Llama framework first selects video frames to process using speech information. Then we use a visual captioner to map the selected frames in the text space. We feed the resulting captions, along with speech transcripts, to the LLM which outputs the chapter boundaries and titles jointly as a single sequence of tokens." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 298, + 296, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 298, + 296, + 380 + ], + "spans": [ + { + "bbox": [ + 55, + 298, + 296, + 380 + ], + "type": "text", + "content": "by this model. The visual information thus complements the previous 'blind' predictions from the narrations, and allows us to refine the predictions. This results in a video representations " + }, + { + "bbox": [ + 55, + 298, + 296, + 380 + ], + "type": "inline_equation", + "content": "V_{\\text{sampled}} = (v_1, v_2, \\dots, v_K)" + }, + { + "bbox": [ + 55, + 298, + 296, + 380 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 55, + 298, + 296, + 380 + ], + "type": "inline_equation", + "content": "K << N" + }, + { + "bbox": [ + 55, + 298, + 296, + 380 + ], + "type": "text", + "content": ". For the videos that lack speech entirely (e.g., about 3% of the videos in [109]), we sample frames at 10-second intervals, with an upper bound of 100 frames to maintain computational practicality." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 403, + 296, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 403, + 296, + 520 + ], + "spans": [ + { + "bbox": [ + 55, + 403, + 296, + 520 + ], + "type": "text", + "content": "Mapping video to text with timestamps. To leverage the knowledge of a pretrained LLM, we map all our inputs to text. This includes: (1) speech transcriptions " + }, + { + "bbox": [ + 55, + 403, + 296, + 520 + ], + "type": "inline_equation", + "content": "S = (s_{1}, s_{2}, \\dots, s_{M})" + }, + { + "bbox": [ + 55, + 403, + 296, + 520 + ], + "type": "text", + "content": " from the audio modality, and (2) caption descriptions " + }, + { + "bbox": [ + 55, + 403, + 296, + 520 + ], + "type": "inline_equation", + "content": "V_{\\text{caption}} = (d_{1}, d_{2}, \\dots, d_{K})" + }, + { + "bbox": [ + 55, + 403, + 296, + 520 + ], + "type": "text", + "content": " from the visual modality. In detail, for speech transcriptions, we use ASR outputs provided by [109], obtained using the Whisper-Large-V2 [70] model through the WhisperX [6] implementation. For captioning, we employ MiniCPM-V [112] as an image captioner, applied independently on the selected video frames, i.e., " + }, + { + "bbox": [ + 55, + 403, + 296, + 520 + ], + "type": "inline_equation", + "content": "d_{i} = \\text{Captioner}(v_{i})" + }, + { + "bbox": [ + 55, + 403, + 296, + 520 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 522, + 296, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 522, + 296, + 569 + ], + "spans": [ + { + "bbox": [ + 55, + 522, + 296, + 569 + ], + "type": "text", + "content": "As we aim at predicting relevant chapter boundaries, we provide temporal information to the LLM. For both modalities, we prepend the timestamp information formatted as \"HH:MM:SS\" to encode the location at which the speech or caption is obtained." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 571, + 296, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 571, + 296, + 664 + ], + "spans": [ + { + "bbox": [ + 55, + 571, + 296, + 664 + ], + "type": "text", + "content": "Captions naturally come from a single point in time. Speech segments cover intervals, but their duration is typically very short (3-4 seconds). We therefore simply use the start time of each transcribed speech interval. We interleave the speech and caption inputs based on their timestamps in a sorted order. We add a modality-specific prefix to each timestamp to denote which modality the information is extracted from (i.e., ASR for speech transcripts, Caption for captions)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 666, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 666, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 666, + 296, + 714 + ], + "type": "text", + "content": "We prepend the text combining speech transcripts and captions with a fixed prompt that provides task instructions (see sup. mat. for the exact wording). This prompt occupies approximately 90 tokens and is independent of video length." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 298, + 556, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 298, + 556, + 519 + ], + "spans": [ + { + "bbox": [ + 313, + 298, + 556, + 519 + ], + "type": "text", + "content": "Language model. We derive our framework by making use of a powerful pretrained LLM. Specifically, we employ the recent Llama-3.1-8B-Instruct [20] model and further finetune on chapter annotations using the LoRA technique [35]. Given the input structure previously described, the LLM is trained to output chapters, where each chapter consists of a timestamp in HH:MM:SS format followed by a free-form chapter title. We treat both the timestamps and titles simply as text tokens and apply the standard cross-entropy loss over the original vocabulary of the pretrained LLM. We apply teacher forcing during training and decode tokens autoregressively at inference. Note that the final model (taking both speech and captions as input) is trained independently from the speech-only version of our model used for frame selection, but these two models share the same backbone, and only differ in their LoRA parameters (13MB each). Across all experiments, we finetune models for a single epoch and use the same hyperparameters. We provide these hyperparameters, along with implementation details in Appendix A, and provide experiments with several Llama variants in Appendix C." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 528, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 528, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 528, + 556, + 713 + ], + "type": "text", + "content": "Iterative prediction for long videos. The inputs may exceed the context window limitation of the LLM, especially in the case of long videos. For example, on an A6000 GPU, the Llama-3.1-8B-Instruct [20] model can process videos up to around 15k tokens during training, which corresponds to 50 minutes of video content on average, and 25k tokens during inference, which corresponds to 80 minutes of video content on average. To address this issue, during training, we select videos that have less than 15k tokens. Since there are videos up to 1 hour long in the training set that satisfy this constraint, and since we do not need the entire training dataset to achieve good performance, this token limitation does not hinder our training. During evaluation, we predict chapters for each chunk sequentially, such that the start of a chunk is the end of the previous chunk. Finally, we merge the predictions from all chunks to obtain chapter boundaries for the complete video. We provide more details in Appendix A.4." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18950" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 133, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 133, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 133, + 85 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 161 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 161 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 161 + ], + "type": "text", + "content": "In this section, we start by describing the data and evaluation metrics used in our experiments (Sec. 4.1). Next, we compare our results with the state of the art (Sec. 4.2), and then provide a series of ablations in our framework (Sec. 4.3). Finally, we investigate the impact of testing with very long videos exceeding our context window limitations (Sec. 4.4)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 171, + 173, + 182 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 171, + 173, + 182 + ], + "spans": [ + { + "bbox": [ + 55, + 171, + 173, + 182 + ], + "type": "text", + "content": "4.1. Data and evaluation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 190, + 296, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 190, + 296, + 491 + ], + "spans": [ + { + "bbox": [ + 56, + 190, + 296, + 491 + ], + "type": "text", + "content": "Data. We train and evaluate on the recently released VidChapters-7M [109] dataset that includes user-annotated chaptered videos sourced from YouTube. Speech transcripts are obtained using Whisper [70] as the ASR method. In the original release, there is a total of 817k videos, spanning 8M chapters, with 2.4 minutes per chapter and 5.4 words per chapter title, totaling to 23 minutes and 8.3 chapters per video on average. Data is split into 801k training, 8.2k validation, and 8.2k test videos. To measure performance at different video lengths, we define three categories depending on video duration: 'short' (0-15min), 'medium' (15-30min), and 'long' (30-60min) videos. In this work, we use a subset of the training data as we observe increasing the training set brings diminishing returns at the cost of extended training times (see Fig. 4). Specifically, we use about 20k training videos (10k short videos used for the speech-based frame selection model and another 10k videos evenly split across short, medium and long durations for the final model. For state-of-the-art comparisons (Sec. 4.2), we employ the full official test set, which also contains videos without any speech (2.5% of the videos), and videos longer than 60 minutes (e.g., there are few videos that last about 12 hours). In ablations (Sec. 4.3), both for faster experimentation, and to limit the use of the test set during experimentation, we train on a randomly sampled subset of 1k videos (evenly split between short, medium, and long) and report results on a randomly sampled subset of 300 validation videos (100 from each duration) that have at least one speech utterance." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 504, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 504, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 504, + 296, + 714 + ], + "type": "text", + "content": "Evaluation metrics. We primarily monitor temporal segmentation metrics to evaluate our chapter boundary detections. In particular, we employ tIoU and F1 scores. For tIoU (temporal Intersection over Union), we first compute the optimal matching between predicted and ground truth segments by greedily selecting pairs with highest IoU scores. The tIoU score is then calculated as the mean IoU across all matched pairs, multiplied by 100 to obtain a percentage. For F1 score, we first compute precision and recall at different IoU thresholds (ranging from 0.5 to 0.95 with a step of 0.05). At each threshold, a prediction is considered correct if it has IoU above the threshold with a ground truth segment. The precision is the ratio of correct predictions to total predictions, while recall is the ratio of matched ground truth segments to total ground truth segments. The F1 score is then computed as the harmonic mean of precision and recall. The final F1 metric is the average across all thresholds, multiplied by 100 to obtain a percentage. Note that [109] uses recall and precision metrics in two ways: (1)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 555, + 165 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 165 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 165 + ], + "type": "text", + "content": "by considering timestamps within 3 or 5 second thresholds as matches, and (2) by considering segments with IoU above 0.5 or 0.7 as matches. While these metrics provide point estimates at specific thresholds, we find that IoU and F1 scores offer several advantages: they evaluate performance continuously across multiple thresholds, are more interpretable, and provide a more comprehensive evaluation of the model. For completeness, we also report the metrics used in [109] in Appendix C." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 166, + 555, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 166, + 555, + 212 + ], + "spans": [ + { + "bbox": [ + 313, + 166, + 555, + 212 + ], + "type": "text", + "content": "For chapter title evaluation, we follow [109] and report SODA (S) [25] and CIDEr (C) [94], which measure the quality of the titles for the predicted segments that match to the ground segments (see [109] for details)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 220, + 507, + 232 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 220, + 507, + 232 + ], + "spans": [ + { + "bbox": [ + 313, + 220, + 507, + 232 + ], + "type": "text", + "content": "4.2. Comparison with the state of the art" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 238, + 556, + 296 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 238, + 556, + 296 + ], + "spans": [ + { + "bbox": [ + 313, + 238, + 556, + 296 + ], + "type": "text", + "content": "In Tab. 1, we report the performance of our model on the full VidChapters-7M test set [109] ('All' columns), and compare to the state of the art reported in [109], which uses Vid2Seq [110]. Moreover, we evaluate four proprietary models using our speech-based frame selection and captioning in a zero-shot manner." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 297, + 556, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 297, + 556, + 424 + ], + "spans": [ + { + "bbox": [ + 313, + 297, + 556, + 424 + ], + "type": "text", + "content": "We observe that our finetuned Chapter-Llama achieves substantial performance improvements across all metrics and video duration categories. (e.g., 45.3 vs 26.7 F1 and 19.3 vs 11.6 SODA compared to Vid2Seq). Notably, our improvement over Vid2Seq is more important for medium and long videos compared to short videos. Note that our final approach was trained using the subset of data detailed in the previous section, specifically 20k videos, which constitutes only " + }, + { + "bbox": [ + 313, + 297, + 556, + 424 + ], + "type": "inline_equation", + "content": "2.5\\%" + }, + { + "bbox": [ + 313, + 297, + 556, + 424 + ], + "type": "text", + "content": " of the total available training data. In contrast, the baseline Vid2Seq model [110] was trained on a considerably larger dataset, utilizing both HowTo100M [57] and the entire VidChapters-7M training set." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 425, + 556, + 505 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 425, + 556, + 505 + ], + "spans": [ + { + "bbox": [ + 313, + 425, + 556, + 505 + ], + "type": "text", + "content": "Additionally, we report performances of our model without training on any chapter annotations (i.e., both the speech-based frame selector and the LLM are not finetuned, and run with the same prompt as in the finetuned setting). We see that our zero-shot method also achieves competitive performance (e.g., 29.5 F1), whereas Vid2Seq only trained on HowTo100M does not generalize (3.0 F1)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 506, + 556, + 611 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 506, + 556, + 611 + ], + "spans": [ + { + "bbox": [ + 313, + 506, + 556, + 611 + ], + "type": "text", + "content": "Finally, when zero-shot evaluating the proprietary models, GPT4-o [62] and Gemini variants [27], with our speech-based frame selection and captioning inputs, we observe competitive performances (e.g., 42.2 F1 with Gemini-1.5-Pro); however, our Chapter-Llama still surpasses on all metrics. Note that, due to API costs of the proprietary models, we performed their evaluation on a random " + }, + { + "bbox": [ + 313, + 506, + 556, + 611 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 313, + 506, + 556, + 611 + ], + "type": "text", + "content": " subset of the test set; however, we verified that the scores are similar between " + }, + { + "bbox": [ + 313, + 506, + 556, + 611 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 313, + 506, + 556, + 611 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 506, + 556, + 611 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 313, + 506, + 556, + 611 + ], + "type": "text", + "content": " of the test set when evaluating with Chapter-Llama." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 620, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 620, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 620, + 556, + 715 + ], + "type": "text", + "content": "Qualitative comparison. In Fig. 3, we provide qualitative examples comparing our method against Vid2Seq [109, 110] and our zero-shot baseline. Our predictions align well with the ground truth chapters, accurately capturing both the temporal boundaries and generating relevant titles. In contrast, Vid2Seq segments tend to be less accurate, and we also observe that it often produces repetitive titles (bottom example). The zero-shot Chapter-Llama baseline tends to generate relatively longer and" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18951" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 91, + 550, + 205 + ], + "blocks": [ + { + "bbox": [ + 58, + 91, + 550, + 205 + ], + "lines": [ + { + "bbox": [ + 58, + 91, + 550, + 205 + ], + "spans": [ + { + "bbox": [ + 58, + 91, + 550, + 205 + ], + "type": "table", + "html": "
BackboneFrame selectionFt.ShortMediumLongAll
F1tIoUSCF1tIoUSCF1tIoUSCF1tIoUSC
GPT-4o-mini [62]†OursX32.164.57.242.430.562.36.130.628.061.06.027.331.263.66.837.8
GPT-4o [62]†OursX37.768.08.453.838.168.88.151.436.566.26.634.837.668.08.151.0
Gemini-2.0-Flash [27]†OursX39.969.212.072.843.871.411.270.334.966.29.051.640.269.311.469.7
Gemini-1.5-Pro [27]†OursX41.770.611.765.343.871.811.261.441.370.610.155.342.270.911.463.2
Vid2Seq [109, 110]EquidistantX2.528.60.30.33.229.70.30.44.632.00.30.53.029.30.30.4
Llama 3.1-8BOursX29.963.47.134.530.662.75.428.126.659.33.618.929.562.56.230.7
Vid2Seq [109, 110]Equidistant✓33.463.715.274.919.053.37.531.916.750.85.928.426.758.611.655.8
Llama 3.1-8B (Chapter-Llama)Ours✓45.572.220.2103.546.772.318.898.741.369.215.891.245.371.819.3100.9
", + "image_path": "999eb42551b65875519085b63c1a803fe9942893b8b6146575f1458687b57c3d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 55, + 208, + 555, + 295 + ], + "lines": [ + { + "bbox": [ + 55, + 208, + 555, + 295 + ], + "spans": [ + { + "bbox": [ + 55, + 208, + 555, + 295 + ], + "type": "text", + "content": "Table 1. Comparison to the state of the art on VidChapters-7M test set: We split the table into (bottom) the comparison between Chapter-Llama and the state-of-the-art method Vid2Seq [110], and (top) the evaluation of proprietary models. Chapter-Llama significantly outperforms Vid2Seq trained and reported by [109] (45.3 vs 26.7 F1). Our method also achieves strong performance in zero-shot mode - without finetuning (Ft.) on any chapter annotation (29.5 F1). Furthermore, we report performance of proprietary models in such zero-shot setting, using our speech-based frame selection and captioning, and observe inferior results than Chapter-Llama (42.2 F1 with Gemini-1.5-Pro). Note that we use the full official 8.1k test set videos ('All'), unlike in the remaining experiments that report on the validation subset. We also report the performance breakdown into short (4891), medium (1736), and long (892) test videos. Our model was trained on 10k videos balanced across short, medium and long durations. " + }, + { + "bbox": [ + 55, + 208, + 555, + 295 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 55, + 208, + 555, + 295 + ], + "type": "text", + "content": " denotes evaluation on a random " + }, + { + "bbox": [ + 55, + 208, + 555, + 295 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 55, + 208, + 555, + 295 + ], + "type": "text", + "content": " subset of the test set due to API costs of proprietary models." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 69, + 332, + 538, + 658 + ], + "blocks": [ + { + "bbox": [ + 69, + 332, + 538, + 658 + ], + "lines": [ + { + "bbox": [ + 69, + 332, + 538, + 658 + ], + "spans": [ + { + "bbox": [ + 69, + 332, + 538, + 658 + ], + "type": "image", + "image_path": "0f877bd2a95cfe6c084d86dd98b61f7f56dd78f44b04f115579661268a192f75.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 662, + 555, + 696 + ], + "lines": [ + { + "bbox": [ + 55, + 662, + 555, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 662, + 555, + 696 + ], + "type": "text", + "content": "Figure 3. Qualitative results: We display two examples and compare our Chapter-Llama results against the ground truth (GT), as well as the zero-shot (ZS) and Vid2Seq (VS) baselines. For each example, we show the corresponding SODA (S) and CIDEr (C) scores. Our method overall shows the highest similarity with the GT, while Vid2Seq can suffer from repeated chapter titles, and zero-shot generations tend to over-segment." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "18952" + } + ] + } + ], + "index": 4 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 69, + 70, + 279, + 175 + ], + "blocks": [ + { + "bbox": [ + 69, + 70, + 279, + 175 + ], + "lines": [ + { + "bbox": [ + 69, + 70, + 279, + 175 + ], + "spans": [ + { + "bbox": [ + 69, + 70, + 279, + 175 + ], + "type": "table", + "html": "
ModalitiesSegmentationTitles
SpeechCaptionsF1tIoUSC
Zero-shotX✓12.648.61.96.4
✓X22.757.34.419.7
✓✓29.963.06.933.7
FinetunedX✓39.167.75.920.2
✓X38.568.113.967.3
✓✓42.670.616.482.4
", + "image_path": "df6aac37afe671ccead58ee5c9d1075ac0a4b154b50a7193862c719feec7187b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 254, + 296, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 254, + 296, + 301 + ], + "spans": [ + { + "bbox": [ + 55, + 254, + 296, + 301 + ], + "type": "text", + "content": "verbose chapter titles and often generates chapters that appear to be continuations of previous chapters rather than distinct segments, while also exhibiting over-segmentation issues. We provide more examples in Appendix E." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 307, + 154, + 319 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 307, + 154, + 319 + ], + "spans": [ + { + "bbox": [ + 55, + 307, + 154, + 319 + ], + "type": "text", + "content": "4.3. Ablation studies" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 324, + 296, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 324, + 296, + 396 + ], + "spans": [ + { + "bbox": [ + 55, + 324, + 296, + 396 + ], + "type": "text", + "content": "In the following, we experiment with (i) the contribution of speech and caption modalities, along with the effect of LLM finetuning, (ii) the effect of our frame selection method for captioning, (iii) the amount of training data, and (iv) the use of frame embeddings instead of captions. As mentioned above, we use 1k training and 300 validation videos for these ablations." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 403, + 296, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 403, + 296, + 496 + ], + "spans": [ + { + "bbox": [ + 55, + 403, + 296, + 496 + ], + "type": "text", + "content": "Modalities and LLM finetuning. In Tab. 2, we ablate the impact of finetuning the LLM and the contribution of each of the speech and caption modalities. In the top block, we run our baselines in zero-shot setting as introduced in the previous section. The speech-only baseline outperforms the captions-only baseline by a large margin in the zero-shot setting. This suggests that speech contains more relevant information for chaptering, as was previously observed by [109]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 496, + 296, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 496, + 296, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 496, + 296, + 601 + ], + "type": "text", + "content": "As shown in the bottom block of Tab. 2, we observe large performance improvements when finetuning the LLM, as opposed to zero-shot. We hypothesize that zero-shot prompting with a long multi-modal text, potentially containing redundant and irrelevant information, may overwhelm the LLM [79, 101]. We obtain our best model by combining the two modalities, which performs better than the individual speech-only or caption-only models. This demonstrates the multi-modal capabilities of our model." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 609, + 295, + 690 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 609, + 295, + 690 + ], + "spans": [ + { + "bbox": [ + 55, + 609, + 295, + 690 + ], + "type": "text", + "content": "Speech-based frame selection. In Tab. 3, we examine a number of strategies to sample frames at which we extract captions. In addition to previously described metrics, for each of the frame sampling approaches, we report the average number of captions per video and the average number of text tokens per minute. For reference, we also report an off-the-shelf shot detection [11] and Vid2Seq [109, 110]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 690, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 690, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 690, + 296, + 714 + ], + "type": "text", + "content": "We compare our speech-based frame selection strategy to various baselines. We experiment with sampling (i) uniformly" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 317, + 71, + 553, + 240 + ], + "blocks": [ + { + "bbox": [ + 55, + 178, + 296, + 243 + ], + "lines": [ + { + "bbox": [ + 55, + 178, + 296, + 243 + ], + "spans": [ + { + "bbox": [ + 55, + 178, + 296, + 243 + ], + "type": "text", + "content": "Table 2. Contribution of different modalities and finetuning: Finetuning the LLM with 1k videos largely improves chaptering performance on 300 validation videos, see bottom block vs top block. In the finetuned setting, we further demonstrate the advantages of combining both modalities, i.e., transcribed speech from ASR and automatic captions extracted from video frames." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 71, + 553, + 240 + ], + "lines": [ + { + "bbox": [ + 317, + 71, + 553, + 240 + ], + "spans": [ + { + "bbox": [ + 317, + 71, + 553, + 240 + ], + "type": "table", + "html": "
MethodFrame selection for captionsaverage #frames#tokens per min.SegmentationTitles SC
F1tIoU
BASELINES
Shot detection [11]n/a49.4n/a6.237.6--
Vid2Seq [109, 110]100 equidistant100.0128.625.457.811.255.0
CHAPTER-LLAMA VARIANTS
SpeechCaption
✓✘n/an/a248.638.568.113.967.3
100 equidistant100.0449.121.053.88.436.0
Every 10 sec.83.1280.312.845.94.313.0
✘✓Shot boundaries49.4193.216.250.73.912.4
10 equidistant10.041.811.046.43.69.0
Speech-based10.336.239.167.75.920.2
100 equidistant100.0746.239.267.416.183.8
Every 10 sec.83.1570.141.069.315.477.3
✓✓Shot boundaries40.4481.740.669.115.879.3
10 equidistant10.0326.140.167.915.877.5
Speech-based10.3320.442.670.616.482.4
", + "image_path": "a65838918ec9a355117d23ded7dc9da7391704a721123c41b31820413e9b07a7.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 244, + 555, + 394 + ], + "lines": [ + { + "bbox": [ + 313, + 244, + 555, + 394 + ], + "spans": [ + { + "bbox": [ + 313, + 244, + 555, + 394 + ], + "type": "text", + "content": "Table 3. Frame selection strategies for captioning: We evaluate different approaches for selecting frames to extract captions from, comparing our speech-based selection method against baselines. The table shows results for models trained on 1k videos and evaluated on 300 validation videos. We experiment with using speech only, captions only, and both modalities (bottom section). For caption extraction, we compare our speech-based approach to other alternatives such as equidistant sampling (100 or 10 frames), uniformly sampling every 10 seconds, or sampling at shot boundaries using [11]. Our speech-based frame selection achieves the best overall performance (F1: 42.6, tIoU: 70.6) while requiring significantly fewer number of frames on average (10.3) compared to other sampling approaches. The tokens-per-minute statistic shows the total input length including both speech transcriptions and captions, excluding the fixed prompt template." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 407, + 555, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 407, + 555, + 501 + ], + "spans": [ + { + "bbox": [ + 313, + 407, + 555, + 501 + ], + "type": "text", + "content": "100 frames as in Vid2Seq, (ii) every 10 seconds, (iii) at shot boundaries detected by an off-the-shelf shot detector [11], (iv) 10 equidistant frames to be similar to our speech-based locations (i.e., 10.0 vs 10.3 number of frames on average), and (v) sampling at frames predicted as chapter boundaries by our LLM that inputs only speech. In all cases, we limit the maximum number of frames to 100 as in [109, 110] to handle extreme durations." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 502, + 555, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 502, + 555, + 583 + ], + "spans": [ + { + "bbox": [ + 313, + 502, + 555, + 583 + ], + "type": "text", + "content": "In both caption-only and caption+speech settings, our speech-based frame selection approach achieves better segmentation results than the more frame-expensive baselines '100 equidistant', 'every 10 sec', and 'shot boundaries', while using much less frames, and also improves over the '10 equidistant' baseline which uses a similar number of frames. This demonstrates the effectiveness of our speech-based frame selection strategy." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 584, + 555, + 631 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 584, + 555, + 631 + ], + "spans": [ + { + "bbox": [ + 313, + 584, + 555, + 631 + ], + "type": "text", + "content": "For reference, we also report positive comparison against shot detection and Vid2Seq [109, 110]. Note Vid2Seq has less tokens per min. compared to our 100 equidistant variants, because Vid2Seq uses a different timestamp tokenizer in the input." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 643, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 643, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 643, + 556, + 715 + ], + "type": "text", + "content": "Amount of training data. Given the large-scale nature of the VidChapters-7M training set, we investigate how much chapter data is needed for LoRA finetuning the LLM. We plot the performance against the number of training videos in Fig. 4. We start by the zero-shot baseline as the first data point, and report our method with 1k, 5k, 7k, and 10k videos, split evenly" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18953" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 55, + 69, + 293, + 172 + ], + "blocks": [ + { + "bbox": [ + 55, + 69, + 293, + 172 + ], + "lines": [ + { + "bbox": [ + 55, + 69, + 293, + 172 + ], + "spans": [ + { + "bbox": [ + 55, + 69, + 293, + 172 + ], + "type": "image", + "image_path": "7e58313b19ae66019b928c53d56a372c0f5a15c1b0f4c22319e6125742aa64fe.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 175, + 295, + 228 + ], + "lines": [ + { + "bbox": [ + 55, + 175, + 295, + 228 + ], + "spans": [ + { + "bbox": [ + 55, + 175, + 295, + 228 + ], + "type": "text", + "content": "Figure 4. Amount of training data: Our experiments show a substantial improvement when moving from zero-shot to training with 1k videos. Beyond 1k videos, performance continues to improve but at a much slower rate, motivating our choice of using only 10k training videos for our final LLM." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 243, + 296, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 243, + 296, + 324 + ], + "spans": [ + { + "bbox": [ + 55, + 243, + 296, + 324 + ], + "type": "text", + "content": "between three durations. We see that after increasing above several thousand training videos starts to bring diminishing returns. We therefore keep 10k training videos for our final LLM, which makes our approach highly efficient to train (40min on 4 H100 GPUs). Note that here we focus on the chaptering LLM and always use frame sampling locations from a speech-based module trained on 10k separate videos." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 337, + 296, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 337, + 296, + 570 + ], + "spans": [ + { + "bbox": [ + 56, + 337, + 296, + 570 + ], + "type": "text", + "content": "Frame embeddings vs captions. In Tab. 4, we investigate whether raw visual embeddings could serve as an alternative to textual captions. To this end, we experiment with replacing the captions with frame embeddings. Specifically, for each frame, we extract the 1152-dimensional output embedding corresponding to the [CLS] token from a frozen SigLIP model [119], and feed through a 2-layer MLP mapping network. We initialize the MLP weights from MANTIS [41] and train jointly with the LLM during finetuning. The results with 'Speech+Embeddings' are better than 'Speech' alone (38.5 vs 40.4 F1), but worse than 'Speech+Captions' (42.6 vs 40.4 F1). The performance gap between 'Speech+Embeddings' and 'Speech+Captions' may be due to the richer information provided by captions, which use multiple tokens per frame, directly in text form, compared to the single [CLS] token frame embedding, requiring a mapping network to be ingested by an LLM. Finally, while combining all modalities achieves the best performance (44.4 F1), we exclude frame embeddings from our final model due to practical considerations, e.g., they add complexity, increase processing time by 2.5x, and require 3000x more storage space." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 578, + 248, + 590 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 578, + 248, + 590 + ], + "spans": [ + { + "bbox": [ + 55, + 578, + 248, + 590 + ], + "type": "text", + "content": "4.4. Iterative prediction on longer videos" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 597, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 597, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 597, + 295, + 713 + ], + "type": "text", + "content": "In our ablation studies, our experimental setting considered training and evaluating with videos that fit within the LLM context window. In Tab. 5, we evaluate the benefit of our iterative prediction procedure for handling videos that exceed the LLM context window. For this, we identify videos in the validation set whose inputs exceed the LLM inference context window (" + }, + { + "bbox": [ + 55, + 597, + 295, + 713 + ], + "type": "inline_equation", + "content": ">35k" + }, + { + "bbox": [ + 55, + 597, + 295, + 713 + ], + "type": "text", + "content": " tokens), resulting in 110 videos. On this challenging subset, we find that our iterative prediction procedure improves chaptering results compared to the baseline that only runs the LLM once by cropping the input to the first" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 317, + 70, + 549, + 179 + ], + "blocks": [ + { + "bbox": [ + 317, + 70, + 549, + 179 + ], + "lines": [ + { + "bbox": [ + 317, + 70, + 549, + 179 + ], + "spans": [ + { + "bbox": [ + 317, + 70, + 549, + 179 + ], + "type": "table", + "html": "
ModalitiesSegmentationTitles
SpeechEmbeddingsCaptionsF1tIoUSC
✓--38.568.113.967.3
-✓-38.466.53.47.3
--✓39.167.75.920.2
✓✓-40.468.215.374.9
✓-✓42.670.616.482.4
✓✓✓44.471.516.384.2
", + "image_path": "ee3499cfab2b54a432978aee1ccea9ef6acc47e7276e5ab27d6dc471502f9109.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 328, + 265, + 539, + 367 + ], + "blocks": [ + { + "bbox": [ + 313, + 182, + 553, + 258 + ], + "lines": [ + { + "bbox": [ + 313, + 182, + 553, + 258 + ], + "spans": [ + { + "bbox": [ + 313, + 182, + 553, + 258 + ], + "type": "text", + "content": "Table 4. Frame embeddings vs captions: We compare using frame captions versus visual features from a frozen SigLIP model projected through a learned 2-layer MLP mapping network ('Embeddings'). While the 'Speech+Embeddings' combination performs better than speech alone (40.4 vs 38.5 F1), it underperforms compared to the 'Speech+Captions' combination (42.6 vs 40.4 F1). All models are trained with 1k videos and evaluated on 300 videos." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 328, + 265, + 539, + 367 + ], + "lines": [ + { + "bbox": [ + 328, + 265, + 539, + 367 + ], + "spans": [ + { + "bbox": [ + 328, + 265, + 539, + 367 + ], + "type": "table", + "html": "
Window# tok.avg # iter.Subset exceeding 35k tokens
F1tIoUSC
First10k113.150.54.031.2
15k116.654.95.443.3
20k118.756.76.647.5
Iterative10k8.518.557.16.925.1
15k5.423.660.18.735.2
20k4.125.361.410.344.0
", + "image_path": "5aa4096cd947c2cb8adb4fd7a3831eeaa843ebf8743d6f7f53cd244a383982eb.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 370, + 554, + 455 + ], + "lines": [ + { + "bbox": [ + 313, + 370, + 554, + 455 + ], + "spans": [ + { + "bbox": [ + 313, + 370, + 554, + 455 + ], + "type": "text", + "content": "Table 5. Iterative prediction: Our iterative prediction procedure improves chaptering results on the subset of 110 videos which exceed 35k tokens compared to the baseline that only runs the LLM once (by only taking the first window, and discarding the rest of the input sequence), across various context windows. As we increase the context window in the iterative prediction, the performance gradually improves and the average number of iterations decreases. The model is trained with 1k videos." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 470, + 553, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 470, + 553, + 506 + ], + "spans": [ + { + "bbox": [ + 313, + 470, + 553, + 506 + ], + "type": "text", + "content": "input window, across various context windows (10k, 15k, and 20k). We refer to Appendix B for details on the video lengths and statistics of videos that exceed the LLM context window." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 518, + 388, + 530 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 518, + 388, + 530 + ], + "spans": [ + { + "bbox": [ + 313, + 518, + 388, + 530 + ], + "type": "text", + "content": "5. Conclusions" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 539, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 539, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 539, + 555, + 713 + ], + "type": "text", + "content": "We presented Chapter-Llama, an approach that leverages LLMs for hour-long video chaptering by mapping video to text using speech transcripts and efficiently captioning video frames sampled with a speech-based frame selector. Our results on VidChapters-7M consequently improved the state of the art by a large margin. We experimentally demonstrated the benefits of our components through an extensive ablation study. One limitation of our approach is that it relies on the accuracy of the ASR and the visual captioner. Future work can explore hierarchical chaptering with several granularities and consider the audio modality beyond speech. We also note that the LLM, the visual captioner, and speech transcription models are trained on large Web datasets, which can contain biases that can lead to inaccurate chaptering, especially for videos depicting underrepresented topics." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18954" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 212 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 212 + ], + "type": "text", + "content": "Acknowledgements. This work was granted access to the HPC resources of IDRIS under the allocation 2024-AD011014696 made by GENCI. This work was funded in part by the ANR project CorVis ANR-21-CE23-0003-01, a research gift from Google, the French government under management of Agence Nationale de la Recherche as part of the \"France 2030\" program, reference ANR-23-IACL-0008 (PR[AI]RIE-PSAI projet), and the ANR project VideoPredict ANR-21-FAI1-0002-01. Cordelia Schmid would like to acknowledge the support by the Körber European Science Prize. The authors would also like to thank Guillaume Astruc, Nikos Athanasiou, Hyolim Kang, and Nicolas Dufour for their feedback." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 224, + 112, + 237 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 224, + 112, + 237 + ], + "spans": [ + { + "bbox": [ + 56, + 224, + 112, + 237 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 243, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 65, + 243, + 294, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 243, + 294, + 297 + ], + "spans": [ + { + "bbox": [ + 65, + 243, + 294, + 297 + ], + "type": "text", + "content": "[1] Soheyla Amirian, Khaled Rasheed, Thiab R Taha, and Hamid R Arabnia. Automatic generation of descriptive titles for video clips using deep learning. In Advances in Artificial Intelligence and Applied Cognitive Computing: Proceedings from ICAI'20 and ACC'20, 2021. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 65, + 300, + 294, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 300, + 294, + 342 + ], + "spans": [ + { + "bbox": [ + 65, + 300, + 294, + 342 + ], + "type": "text", + "content": "[2] D. Argaw, S. Yoon, F. Heilbron, H. Deilamsalehy, T. Bui, Z. Wang, F. Dernoncourt, and J. Chung. Scaling up video summarization pretraining with large language models. In CVPR, 2024. 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 65, + 344, + 294, + 375 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 344, + 294, + 375 + ], + "spans": [ + { + "bbox": [ + 65, + 344, + 294, + 375 + ], + "type": "text", + "content": "[3] Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lucic, and Cordelia Schmid. ViViT: A video vision transformer. In ICCV, 2021. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 65, + 377, + 294, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 377, + 294, + 409 + ], + "spans": [ + { + "bbox": [ + 65, + 377, + 294, + 409 + ], + "type": "text", + "content": "[4] Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio Neural machine translation by jointly learning to align and translate. ICLR, 2015. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 65, + 411, + 294, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 411, + 294, + 442 + ], + "spans": [ + { + "bbox": [ + 65, + 411, + 294, + 442 + ], + "type": "text", + "content": "[5] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In ICCV, 2021. 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 65, + 445, + 294, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 445, + 294, + 476 + ], + "spans": [ + { + "bbox": [ + 65, + 445, + 294, + 476 + ], + "type": "text", + "content": "[6] Max Bain, Jaesung Huh, Tengda Han, and Andrew Zisserman. WhisperX: Time-accurate speech transcription of long-form audio. In Interspeech, 2023. 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 65, + 479, + 294, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 479, + 294, + 521 + ], + "spans": [ + { + "bbox": [ + 65, + 479, + 294, + 521 + ], + "type": "text", + "content": "[7] Nadine Behrmann, S Alireza Golestaneh, Zico Kolter, Jürgen Gall, and Mehdi Noroozi. Unified fully and timestamp supervised temporal action segmentation via sequence to sequence translation. In ECCV, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 65, + 523, + 294, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 523, + 294, + 555 + ], + "spans": [ + { + "bbox": [ + 65, + 523, + 294, + 555 + ], + "type": "text", + "content": "[8] Gedas Bertasius, Heng Wang, and Lorenzo Torresani. Is space-time attention all you need for video understanding? In ICML, 2021. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 65, + 557, + 294, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 557, + 294, + 578 + ], + "spans": [ + { + "bbox": [ + 65, + 557, + 294, + 578 + ], + "type": "text", + "content": "[9] Tom Brown et al. Language models are few-shot learners. In NeurIPS, 2020. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 61, + 580, + 294, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 580, + 294, + 601 + ], + "spans": [ + { + "bbox": [ + 61, + 580, + 294, + 601 + ], + "type": "text", + "content": "[10] Xiao Cao, Zitan Chen, Canyu Le, and Lei Meng. Multi-modal video chapter generation. In BMVC, 2022. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 61, + 603, + 296, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 603, + 296, + 634 + ], + "spans": [ + { + "bbox": [ + 61, + 603, + 296, + 634 + ], + "type": "text", + "content": "[11] Brandon Castellano. Pyscenedetect: Intelligent scene cut detection and video splitting tool. https://pyscenedetect.readthedocs.io/en/latest/, 2018.7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 61, + 637, + 294, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 637, + 294, + 678 + ], + "spans": [ + { + "bbox": [ + 61, + 637, + 294, + 678 + ], + "type": "text", + "content": "[12] Houlun Chen, Xin Wang, Hong Chen, Zihan Song, Jia Jia, and Wenwu Zhu. Grounding-prompter: Prompting llm with multimodal information for temporal sentence grounding in long videos. arXiv:2312.17117, 2023. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 61, + 681, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 681, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 61, + 681, + 294, + 713 + ], + "type": "text", + "content": "[13] Shixing Chen, Xiaohan Nie, David Fan, Dongqing Zhang, Vimal Bhat, and Raffay Hamid. Shot contrastive self-supervised learning for scene boundary detection. In CVPR, 2021. 2" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 320, + 73, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 320, + 73, + 553, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 73, + 553, + 105 + ], + "spans": [ + { + "bbox": [ + 320, + 73, + 553, + 105 + ], + "type": "text", + "content": "[14] Shixing Chen, Chun-Hao Liu, Xiang Hao, Xiaohan Nie, Maxim Arap, and Raffay Hamid. Movies2Scenes: Using movie metadata to learn scene representation. In CVPR, 2023. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 320, + 106, + 555, + 147 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 106, + 555, + 147 + ], + "spans": [ + { + "bbox": [ + 320, + 106, + 555, + 147 + ], + "type": "text", + "content": "[15] Sihan Chen, Xingjian He, Longteng Guo, Xinxin Zhu, Weining Wang, Jinhui Tang, and Jing Liu. VALOR: Vision-audiolanguage omni-perception pretraining model and dataset. TPAMI, 2024. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 320, + 148, + 555, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 148, + 555, + 190 + ], + "spans": [ + { + "bbox": [ + 320, + 148, + 555, + 190 + ], + "type": "text", + "content": "[16] Sihan Chen, Handong Li, Qunbo Wang, Zijia Zhao, Mingzhen Sun, Xinxin Zhu, and Jing Liu. VAST: A vision-audio-subtitle-text omni-modality foundation model and dataset. NeurIPS, 2024. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 320, + 191, + 553, + 222 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 191, + 553, + 222 + ], + "spans": [ + { + "bbox": [ + 320, + 191, + 553, + 222 + ], + "type": "text", + "content": "[17] Yangyu Chen, Shuhui Wang, Weigang Zhang, and Qingming Huang. Less is more: Picking informative frames for video captioning. In ECCV, 2018. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 320, + 223, + 553, + 253 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 223, + 553, + 253 + ], + "spans": [ + { + "bbox": [ + 320, + 223, + 553, + 253 + ], + "type": "text", + "content": "[18] Feng Cheng and Gedas Bertasius. TALLformer: Temporal action localization with long-memory transformer. In ECCV, 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 320, + 255, + 553, + 286 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 255, + 553, + 286 + ], + "spans": [ + { + "bbox": [ + 320, + 255, + 553, + 286 + ], + "type": "text", + "content": "[19] Xu Cheng, Cameron Dale, and Jiangchuan Liu. Statistics and social network of youtube videos. In 16th International Workshop on Quality of Service, 2008. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 320, + 287, + 553, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 287, + 553, + 328 + ], + "spans": [ + { + "bbox": [ + 320, + 287, + 553, + 328 + ], + "type": "text", + "content": "[20] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The Llama 3 herd of models. arXiv:2407.21783, 2024. 3, 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 320, + 330, + 553, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 330, + 553, + 361 + ], + "spans": [ + { + "bbox": [ + 320, + 330, + 553, + 361 + ], + "type": "text", + "content": "[21] Yue Fan, Xiaojian Ma, Rujie Wu, Yuntao Du, Jiaqi Li, Zhi Gao, and Qing Li. VideoAgent: A memory-augmented multimodal agent for video understanding. In ECCV, 2024. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 320, + 361, + 555, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 361, + 555, + 403 + ], + "spans": [ + { + "bbox": [ + 320, + 361, + 555, + 403 + ], + "type": "text", + "content": "[22] Xinyu Fang, Kangrui Mao, Haodong Duan, Xiangyu Zhao, Yining Li, Dahua Lin, and Kai Chen. MMBench-video: A long-form multi-shot benchmark for holistic video understanding. In NeurIPS Datasets and Benchmarks, 2024. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 320, + 404, + 553, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 404, + 553, + 435 + ], + "spans": [ + { + "bbox": [ + 320, + 404, + 553, + 435 + ], + "type": "text", + "content": "[23] Yazan Abu Farha and Jurgen Gall. MS-TCN: Multi-stage temporal convolutional network for action segmentation. In CVPR, 2019. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 320, + 436, + 553, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 436, + 553, + 488 + ], + "spans": [ + { + "bbox": [ + 320, + 436, + 553, + 488 + ], + "type": "text", + "content": "[24] Chaoyou Fu, Yuhan Dai, Yondong Luo, Lei Li, Shuhuai Ren, Renrui Zhang, Zihan Wang, Chenyu Zhou, Yunhang Shen, Mengdan Zhang, et al. Video-MME: The first-ever comprehensive evaluation benchmark of multi-modal LLMs in video analysis. arXiv:2405.21075, 2024. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 320, + 490, + 553, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 490, + 553, + 521 + ], + "spans": [ + { + "bbox": [ + 320, + 490, + 553, + 521 + ], + "type": "text", + "content": "[25] Soichiro Fujita, Tsutomu Hirao, Hidetaka Kamigaito, Manabu Okumura, and Masaaki Nagata. SODA: Story oriented dense video captioning evaluation framework. In ECCV, 2020. 5" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 320, + 522, + 553, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 522, + 553, + 552 + ], + "spans": [ + { + "bbox": [ + 320, + 522, + 553, + 552 + ], + "type": "text", + "content": "[26] Shang-Hua Gao, Qi Han, Zhong-Yu Li, Pai Peng, Liang Wang, and Ming-Ming Cheng. Global2Local: Efficient structure search for video action segmentation. In CVPR, 2021. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 320, + 554, + 553, + 574 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 554, + 553, + 574 + ], + "spans": [ + { + "bbox": [ + 320, + 554, + 553, + 574 + ], + "type": "text", + "content": "[27] Gemini Team et al. Gemini: A family of highly capable multimodal models. arXiv:2312.11805, 2024. 3, 5, 6" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 320, + 575, + 555, + 659 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 575, + 555, + 659 + ], + "spans": [ + { + "bbox": [ + 320, + 575, + 555, + 659 + ], + "type": "text", + "content": "[28] Azin Ghazimatin, Ekaterina Garmash, Gustavo Penha, Kristen Sheets, Martin Achenbach, Oguz Semerci, Remi Galvez, Marcus Tannenberg, Sahitya Mantravadi, Divya Narayanan, Ofeliya Kalaydzhyan, Douglas Cole, Ben Carterette, Ann Clifton, Paul N. Bennett, Claudia Hauff, and Mounia Lalmas. PODTILE: Facilitating podcast episode browsing with auto-generated chapters. In Proceedings of the 33rd ACM International Conference on Information and Knowledge Management, 2024. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 320, + 661, + 553, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 661, + 553, + 680 + ], + "spans": [ + { + "bbox": [ + 320, + 661, + 553, + 680 + ], + "type": "text", + "content": "[29] Kristen Grauman et al. Ego4D: Around the world in 3,000 hours of egocentric video. In CVPR, 2022. 1, 3" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 320, + 681, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 681, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 320, + 681, + 553, + 712 + ], + "type": "text", + "content": "[30] Khalil Guetari, Yannis Tevisen, and Frédéric Petitpont. Multimodal chaptering for long-form TV newscast video. arXiv:2406.17590, 2024. 3" + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "18955" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 61, + 73, + 297, + 713 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 61, + 73, + 297, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 73, + 297, + 105 + ], + "spans": [ + { + "bbox": [ + 61, + 73, + 297, + 105 + ], + "type": "text", + "content": "[31] Tengda Han, Max Bain, Arsha Nagrani, Gül Varol, Weidi Xie, and Andrew Zisserman. AutoAD: Movie description in context. In CVPR, 2023. 1, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 106, + 295, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 106, + 295, + 137 + ], + "spans": [ + { + "bbox": [ + 61, + 106, + 295, + 137 + ], + "type": "text", + "content": "[32] Tanveer Hannan, Md Mohaiminul Islam, Thomas Seidl, and Gedas Bertasius. RGNet: A unified retrieval and grounding network for long videos. In ECCV, 2023. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 138, + 295, + 168 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 138, + 295, + 168 + ], + "spans": [ + { + "bbox": [ + 62, + 138, + 295, + 168 + ], + "type": "text", + "content": "[33] Bo He, Jun Wang, Jielin Qiu, Trung Bui, Abhinav Shrivastava, and Zhaowen Wang. Align and attend: Multimodal summarization with dual contrastive losses. In CVPR, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 170, + 295, + 211 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 170, + 295, + 211 + ], + "spans": [ + { + "bbox": [ + 62, + 170, + 295, + 211 + ], + "type": "text", + "content": "[34] Chiori Hori, Takaaki Hori, Teng-Yok Lee, Ziming Zhang, Bret Harsham, John R Hershey, Tim K Marks, and Kazuhiko Sumi. Attention-based multimodal fusion for video description. In ICCV, 2017. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 213, + 295, + 244 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 213, + 295, + 244 + ], + "spans": [ + { + "bbox": [ + 62, + 213, + 295, + 244 + ], + "type": "text", + "content": "[35] Edward J Hu, yelong shen, Phillip Wallis, Zeyuan Allen-Zhu, Yanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In ICLR, 2022. 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 245, + 295, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 245, + 295, + 285 + ], + "spans": [ + { + "bbox": [ + 62, + 245, + 295, + 285 + ], + "type": "text", + "content": "[36] De-An Huang, Shijia Liao, Subhashree Radhakrishnan, Hongxu Yin, Pavlo Molchanov, Zhiding Yu, and Jan Kautz. Lita: Language instructed temporal-localization assistant. In ECCV, 2024. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 286, + 294, + 318 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 286, + 294, + 318 + ], + "spans": [ + { + "bbox": [ + 62, + 286, + 294, + 318 + ], + "type": "text", + "content": "[37] Gabriel Huang, Bo Pang, Zhenhai Zhu, Clara Rivera, and Radu Soricut. Multimodal pretraining for dense video captioning. In AACL-IJCNLP, 2020. 1, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 319, + 294, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 319, + 294, + 350 + ], + "spans": [ + { + "bbox": [ + 62, + 319, + 294, + 350 + ], + "type": "text", + "content": "[38] Qingqiu Huang, Yu Xiong, Anyi Rao, Jiaze Wang, and Dahua Lin. Movienet: A holistic dataset for movie understanding. In ECCV, 2020. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 351, + 295, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 351, + 295, + 392 + ], + "spans": [ + { + "bbox": [ + 62, + 351, + 295, + 392 + ], + "type": "text", + "content": "[39] Md Mohaiminul Islam, Mahmudul Hasan, Kishan Shamsundar Athrey, Tony Braskich, and Gedas Bertasius. Efficient movie scene detection using state-space transformers. In CVPR, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 393, + 295, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 393, + 295, + 425 + ], + "spans": [ + { + "bbox": [ + 62, + 393, + 295, + 425 + ], + "type": "text", + "content": "[40] Md Mohaiminul Islam, Ngan Ho, Xitong Yang, Tushar Nagarajan, Lorenzo Torresani, and Gedas Bertasius. Video ReCap: Recursive captioning of hour-long videos. CVPR, 2024. 1, 2, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 426, + 295, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 426, + 295, + 456 + ], + "spans": [ + { + "bbox": [ + 62, + 426, + 295, + 456 + ], + "type": "text", + "content": "[41] Dongfu Jiang, Xuan He, Huaye Zeng, Cong Wei, Max Ku, Qian Liu, and Wenhu Chen. Mantis: Interleaved multi-image instruction tuning. TMLR, 2024. 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 458, + 295, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 458, + 295, + 489 + ], + "spans": [ + { + "bbox": [ + 62, + 458, + 295, + 489 + ], + "type": "text", + "content": "[42] Kumara Kahatapitiya, Kanchana Ranasinghe, Jongwoo Park, and Michael S Ryoo. Language repository for long video understanding. arXiv:2403.14622, 2024. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 62, + 491, + 295, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 491, + 295, + 520 + ], + "spans": [ + { + "bbox": [ + 62, + 491, + 295, + 520 + ], + "type": "text", + "content": "[43] Bruno Korbar, Du Tran, and Lorenzo Torresani. SCSampler: Sampling salient clips from video for efficient action recognition. In ICCV, 2019. 2, 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 62, + 522, + 295, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 522, + 295, + 552 + ], + "spans": [ + { + "bbox": [ + 62, + 522, + 295, + 552 + ], + "type": "text", + "content": "[44] Ranjay Krishna, Kenji Hata, Frederic Ren, Li Fei-Fei, and Juan Carlos Niebles. Dense-captioning events in videos. In ICCV, 2017. 1, 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 62, + 554, + 295, + 585 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 554, + 295, + 585 + ], + "spans": [ + { + "bbox": [ + 62, + 554, + 295, + 585 + ], + "type": "text", + "content": "[45] Jie Lei, Linjie Li, Luowei Zhou, Zhe Gan, Tamara L Berg, Mohit Bansal, and Jingjing Liu. Less is more: ClipBERT for video-and-language learning via sparse sampling. In CVPR, 2021. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 62, + 586, + 295, + 627 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 586, + 295, + 627 + ], + "spans": [ + { + "bbox": [ + 62, + 586, + 295, + 627 + ], + "type": "text", + "content": "[46] Feng Li, Jae Won Chung, and Mark Claypool. Three-year trends in YouTube video content and encoding. Proceedings of the 18th International Conference on Signal Processing and Multimedia Applications (SIGMAP), 2021. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 62, + 628, + 295, + 659 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 628, + 295, + 659 + ], + "spans": [ + { + "bbox": [ + 62, + 628, + 295, + 659 + ], + "type": "text", + "content": "[47] Mingzhe Li, Mark Claypool, Robert Kinicki, and James Nichols. Characteristics of streaming media stored on the web. ACM Trans. Internet Technol., 2005. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 62, + 661, + 295, + 681 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 661, + 295, + 681 + ], + "spans": [ + { + "bbox": [ + 62, + 661, + 295, + 681 + ], + "type": "text", + "content": "[48] Zhe Li, Yazan Abu Farha, and Jurgen Gall. Temporal action segmentation from timestamp supervision. In CVPR, 2021. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 62, + 682, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 682, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 62, + 682, + 295, + 713 + ], + "type": "text", + "content": "[49] Bin Lin, Bin Zhu, Yang Ye, Munan Ning, Peng Jin, and Li Yuan. Video-LLaVA: Learning united visual representation by alignment before projection. arXiv:2311.10122, 2023. 3" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 320, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 37, + "blocks": [ + { + "bbox": [ + 320, + 73, + 554, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 73, + 554, + 115 + ], + "spans": [ + { + "bbox": [ + 320, + 73, + 554, + 115 + ], + "type": "text", + "content": "[50] Kevin Lin, Linjie Li, Chung-Ching Lin, Faisal Ahmed, Zhe Gan, Zicheng Liu, Yumao Lu, and Lijuan Wang. SwinBERT: End-to-end transformers with sparse attention for video captioning. In CVPR, 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 320, + 118, + 554, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 118, + 554, + 160 + ], + "spans": [ + { + "bbox": [ + 320, + 118, + 554, + 160 + ], + "type": "text", + "content": "[51] Kevin Qinghong Lin, Pengchuan Zhang, Joya Chen, Shraman Pramanick, Difei Gao, Alex Jinpeng Wang, Rui Yan, and Mike Zheng Shou. UniVTG: Towards unified video-language temporal grounding. In ICCV, 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 320, + 162, + 555, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 162, + 555, + 192 + ], + "spans": [ + { + "bbox": [ + 320, + 162, + 555, + 192 + ], + "type": "text", + "content": "[52] Hao Liu, Matei Zaharia, and Pieter Abbeel. Ring attention with blockwise transformers for near-infinite context. arXiv:2310.01889, 2023. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 320, + 194, + 554, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 194, + 554, + 226 + ], + "spans": [ + { + "bbox": [ + 320, + 194, + 554, + 226 + ], + "type": "text", + "content": "[53] Hao Liu, Wilson Yan, Matei Zaharia, and Pieter Abbeel. World model on million-length video and language with blockwise ringattention. arXiv:2402.08268, 2024. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 320, + 228, + 555, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 228, + 555, + 270 + ], + "spans": [ + { + "bbox": [ + 320, + 228, + 555, + 270 + ], + "type": "text", + "content": "[54] Xiaolong Liu, Qimeng Wang, Yao Hu, Xu Tang, Shiwei Zhang, Song Bai, and Xiang Bai. End-to-end temporal action detection with transformer. In IEEE Transactions on Image Processing, 2022. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 320, + 272, + 554, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 272, + 554, + 315 + ], + "spans": [ + { + "bbox": [ + 320, + 272, + 554, + 315 + ], + "type": "text", + "content": "[55] Huaishao Luo, Lei Ji, Botian Shi, Haoyang Huang, Nan Duan, Tianrui Li, Jason Li, Taroon Bharti, and Ming Zhou. UniVL: A unified video and language pre-training model for multimodal understanding and generation. arXiv:2002.06353, 2020. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 320, + 316, + 554, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 316, + 554, + 358 + ], + "spans": [ + { + "bbox": [ + 320, + 316, + 554, + 358 + ], + "type": "text", + "content": "[56] Karttikeya Mangalam, Raiymbek Akshulakov, and Jitendra Malik. EgoSchema: A diagnostic benchmark for very long-form video language understanding. In NeurIPS Track on Datasets and Benchmarks, 2023. 1, 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 320, + 360, + 554, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 360, + 554, + 402 + ], + "spans": [ + { + "bbox": [ + 320, + 360, + 554, + 402 + ], + "type": "text", + "content": "[57] Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic. HowTo100M: Learning a text-video embedding by watching hundred million narrated video clips. In ICCV, 2019. 2, 5" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 320, + 404, + 554, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 404, + 554, + 436 + ], + "spans": [ + { + "bbox": [ + 320, + 404, + 554, + 436 + ], + "type": "text", + "content": "[58] Juhong Min, Shyamal Buch, Arsha Nagrani, Minsu Cho, and Cordelia Schmid. MoReVQA: Exploring modular reasoning models for video question answering. In CVPR, 2024. 2, 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 320, + 437, + 554, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 437, + 554, + 479 + ], + "spans": [ + { + "bbox": [ + 320, + 437, + 554, + 479 + ], + "type": "text", + "content": "[59] Jonghwan Mun, Minchul Shin, Gunsoo Han, Sangho Lee, Seongsu Ha, Joonseok Lee, and Eun-Sol Kim. BaSSL: Boundary-aware self-supervised learning for video scene segmentation. In ACCV, 2022. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 320, + 481, + 554, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 481, + 554, + 513 + ], + "spans": [ + { + "bbox": [ + 320, + 481, + 554, + 513 + ], + "type": "text", + "content": "[60] Arsha Nagrani, Paul Hongsuck Seo, Bryan Seybold, Anja Hauth, Santiago Manen, Chen Sun, and Cordelia Schmid. Learning audio-video modalities from image captions. In ECCV, 2022. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 320, + 514, + 554, + 556 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 514, + 554, + 556 + ], + "spans": [ + { + "bbox": [ + 320, + 514, + 554, + 556 + ], + "type": "text", + "content": "[61] Joe Yue-Hei Ng, Matthew Hausknecht, Sudheendra Vijayanarasimhan, Oriol Vinyals, Rajat Monga, and George Toderici. Beyond short snippets: Deep networks for video classification. In CVPR, 2015. 1" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 320, + 559, + 547, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 559, + 547, + 569 + ], + "spans": [ + { + "bbox": [ + 320, + 559, + 547, + 569 + ], + "type": "text", + "content": "[62] OpenAI. Gpt-4o system card. arXiv:2410.21276, 2024. 5, 6" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 320, + 571, + 554, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 571, + 554, + 592 + ], + "spans": [ + { + "bbox": [ + 320, + 571, + 554, + 592 + ], + "type": "text", + "content": "[63] Yingwei Pan, Ting Yao, Houqiang Li, and Tao Mei. Video captioning with transferred semantic attributes. In CVPR, 2017. 1" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 320, + 594, + 554, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 594, + 554, + 645 + ], + "spans": [ + { + "bbox": [ + 320, + 594, + 554, + 645 + ], + "type": "text", + "content": "[64] Pinelopi Papalampidi, Skanda Koppula, Shreya Pathak, Justin Chiu, Joe Heyward, Viorica Patraucean, Jiajun Shen, Antoine Miech, Andrew Zisserman, and Aida Nematzdeh. A simple recipe for contrastively pre-training video-first encoders beyond 16 frames. In CVPR, 2024. 3" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 320, + 647, + 554, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 647, + 554, + 679 + ], + "spans": [ + { + "bbox": [ + 320, + 647, + 554, + 679 + ], + "type": "text", + "content": "[65] Alejandro Pardo, Fabian Caba Heilbron, Juan León Alcázar, Ali Thabet, and Bernard Ghanem. Moviecuts: A new dataset and benchmark for cut type recognition. In ECCV, 2022. 2" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 320, + 681, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 681, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 320, + 681, + 554, + 713 + ], + "type": "text", + "content": "[66] Jinwoo Park, Jungeun Kim, Jaegwang Seok, Sukhyun Lee, and Junyeong Kim. Contrasting multi-modal similarity framework for video scene segmentation. IEEE Access, 2024. 2" + } + ] + } + ], + "index": 36 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "18956" + } + ] + } + ], + "index": 38 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 61, + 73, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 61, + 73, + 294, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 73, + 294, + 105 + ], + "spans": [ + { + "bbox": [ + 61, + 73, + 294, + 105 + ], + "type": "text", + "content": "[67] Wenjie Pei, Jiyuan Zhang, Xiangrong Wang, Lei Ke, Xiaoyong Shen, and Yu-Wing Tai. Memory-attended recurrent network for video captioning. In CVPR, 2019. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 106, + 294, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 106, + 294, + 137 + ], + "spans": [ + { + "bbox": [ + 61, + 106, + 294, + 137 + ], + "type": "text", + "content": "[68] Alec Radford, Karthik Narasimhan, Tim Salimans, and Ilya Sutskever. Improving language understanding by generative pre-training. Preprint, 2018. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 139, + 295, + 180 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 139, + 295, + 180 + ], + "spans": [ + { + "bbox": [ + 62, + 139, + 295, + 180 + ], + "type": "text", + "content": "[69] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 182, + 295, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 182, + 295, + 213 + ], + "spans": [ + { + "bbox": [ + 62, + 182, + 295, + 213 + ], + "type": "text", + "content": "[70] Alec Radford, Jong Wook Kim, Tao Xu, Greg Brockman, Christine McLeavey, and Ilya Sutskever. Robust speech recognition via large-scale weak supervision. In ICML, 2023. 4, 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 214, + 294, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 214, + 294, + 246 + ], + "spans": [ + { + "bbox": [ + 62, + 214, + 294, + 246 + ], + "type": "text", + "content": "[71] Anyi Rao, Linning Xu, Yu Xiong, Guodong Xu, Qingqiu Huang, Bolei Zhou, and Dahua Lin. A local-to-global approach to multi-modal movie scene segmentation. In CVPR, 2020. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 247, + 294, + 268 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 247, + 294, + 268 + ], + "spans": [ + { + "bbox": [ + 62, + 247, + 294, + 268 + ], + "type": "text", + "content": "[72] Zeeshan Rasheed and Mubarak Shah. Scene detection in hollywood movies and tv shows. In CVPR, 2003. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 270, + 295, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 270, + 295, + 300 + ], + "spans": [ + { + "bbox": [ + 62, + 270, + 295, + 300 + ], + "type": "text", + "content": "[73] Fabian Retkowski and Alexander Waibel. From text segmentation to smart chaptering: A novel benchmark for structuring video transcriptions. arXiv:2402.17633, 2024. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 301, + 294, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 301, + 294, + 334 + ], + "spans": [ + { + "bbox": [ + 62, + 301, + 294, + 334 + ], + "type": "text", + "content": "[74] Marcus Rohrbach, Wei Qiu, Ivan Titov, Stefan Thater, Manfred Pinkal, and Bernt Schiele. Translating video content to natural language descriptions. In ICCV, 2013. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 335, + 294, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 335, + 294, + 376 + ], + "spans": [ + { + "bbox": [ + 62, + 335, + 294, + 376 + ], + "type": "text", + "content": "[75] Daniel Rotman, Dror Porat, and Gal Ashour. Robust video scene detection using multimodal fusion of optimally grouped features. IEEE 19th International Workshop on Multimedia Signal Processing (MMSP), 2017. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 378, + 294, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 378, + 294, + 410 + ], + "spans": [ + { + "bbox": [ + 62, + 378, + 294, + 410 + ], + "type": "text", + "content": "[76] Yong Rui, Thomas S Huang, and Sharad Mehrotra. Exploring video structure beyond the shots. In IEEE International Conference on Multimedia Computing and Systems, 1998. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 411, + 295, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 411, + 295, + 453 + ], + "spans": [ + { + "bbox": [ + 62, + 411, + 295, + 453 + ], + "type": "text", + "content": "[77] Najmeh Sadoughi, Xinyu Li, Avijit Vajpayee, David Fan, Bing Shuai, Hector Santos-Villalobos, Vimal Bhat, and Rohith MV. MEGA: Multimodal alignment aggregation and distillation for cinematic video segmentation. In ICCV, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 453, + 294, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 453, + 294, + 485 + ], + "spans": [ + { + "bbox": [ + 62, + 453, + 294, + 485 + ], + "type": "text", + "content": "[78] Paul Hongsuck Seo, Arsha Nagrani, Anurag Arnab, and Cordelia Schmid. End-to-end generative pretraining for multimodal video captioning. In CVPR, 2022. 1, 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 62, + 486, + 295, + 528 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 486, + 295, + 528 + ], + "spans": [ + { + "bbox": [ + 62, + 486, + 295, + 528 + ], + "type": "text", + "content": "[79] Freda Shi, Xinyun Chen, Kanishka Misra, Nathan Scales, David Dohan, Ed H. Chi, Nathanael Schärli, and Denny Zhou. Large language models can be easily distracted by irrelevant context. In ICML, 2023. 3, 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 62, + 529, + 295, + 571 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 529, + 295, + 571 + ], + "spans": [ + { + "bbox": [ + 62, + 529, + 295, + 571 + ], + "type": "text", + "content": "[80] Nina Shvetsova, Anna Kukleva, Xudong Hong, Christian Rupprecht, Bernt Schiele, and Hilde Kuehne. HowToCaption: Prompting llms to transform video annotations at scale. In ECCV, 2024. 2, 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 62, + 573, + 295, + 614 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 573, + 295, + 614 + ], + "spans": [ + { + "bbox": [ + 62, + 573, + 295, + 614 + ], + "type": "text", + "content": "[81] Panagiotis Sidiropoulos, Vasileios Mezaris, Ioannis Kompat-siaris, Hugo Meinedo, Miguel Bugalho, and Isabel Trancoso. Temporal video segmentation to scenes using high-level audiovisual features. IEEE TCSVT, 2011. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 62, + 616, + 295, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 616, + 295, + 658 + ], + "spans": [ + { + "bbox": [ + 62, + 616, + 295, + 658 + ], + "type": "text", + "content": "[82] Gunnar A. Sigurdsson, Gül Varol, Xiaolong Wang, Ivan Laptev, Ali Farhadi, and Abhinav Gupta. Hollywood in homes: Crowdsourcing data collection for activity understanding. In ECCV, 2016. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 62, + 659, + 295, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 659, + 295, + 690 + ], + "spans": [ + { + "bbox": [ + 62, + 659, + 295, + 690 + ], + "type": "text", + "content": "[83] Karen Simonyan and Andrew Zisserman. Two-stream convolutional networks for action recognition in videos. In ICLR, 2014. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 62, + 691, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 691, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 62, + 691, + 295, + 712 + ], + "type": "text", + "content": "[84] Mattia Soldan, Alejandro Pardo, Juan León Alcazar, Fabian Caba, Chen Zhao, Silvio Giancola, and Bernard Ghanem." + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 37, + "blocks": [ + { + "bbox": [ + 338, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 338, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 338, + 73, + 553, + 95 + ], + "type": "text", + "content": "MAD: A scalable dataset for language grounding in videos from movie audio descriptions. In CVPR, 2022. 1, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 320, + 96, + 553, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 96, + 553, + 128 + ], + "spans": [ + { + "bbox": [ + 320, + 96, + 553, + 128 + ], + "type": "text", + "content": "[85] Chen Sun, Austin Myers, Carl Vondrick, Kevin Murphy, and Cordelia Schmid. VideoBERT: A joint model for video and language representation learning. In ICCV, 2019. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 320, + 129, + 553, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 129, + 553, + 161 + ], + "spans": [ + { + "bbox": [ + 320, + 129, + 553, + 161 + ], + "type": "text", + "content": "[86] Yuchong Sun, Hongwei Xue, Ruihua Song, Bei Liu, Huan Yang, and Jianlong Fu. Long-form video-language pre-training with multimodal temporal contrastive learning. In NeurIPS, 2022. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 320, + 162, + 553, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 162, + 553, + 184 + ], + "spans": [ + { + "bbox": [ + 320, + 162, + 553, + 184 + ], + "type": "text", + "content": "[87] Ilya Sutskever, Oriol Vinyals, and Quoc V Le. Sequence to sequence learning with neural networks. NeurIPS, 2014. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 320, + 185, + 553, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 185, + 553, + 227 + ], + "spans": [ + { + "bbox": [ + 320, + 185, + 553, + 227 + ], + "type": "text", + "content": "[88] Reuben Tan, Xineng Sun, Ping Hu, Jui hsien Wang, Hanieh Deilamsalehy, Bryan A. Plummer, Bryan Russell, and Kate Saenko. Koala: Key frame-conditioned long video-LLM. In CVPR, 2024. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 320, + 228, + 555, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 228, + 555, + 281 + ], + "spans": [ + { + "bbox": [ + 320, + 228, + 555, + 281 + ], + "type": "text", + "content": "[89] Gemini Team, Petko Georgiev, Ving Ian Lei, Ryan Burnell, Libin Bai, Anmol Gulati, Garrett Tanzer, Damien Vincent, Zhufeng Pan, Shibo Wang, et al. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context. arXiv:2403.05530, 2024. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 320, + 283, + 555, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 283, + 555, + 335 + ], + "spans": [ + { + "bbox": [ + 320, + 283, + 555, + 335 + ], + "type": "text", + "content": "[90] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. LLaMA: Open and efficient foundation language models. arXiv:2302.13971, 2023. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 320, + 337, + 553, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 337, + 553, + 358 + ], + "spans": [ + { + "bbox": [ + 320, + 337, + 553, + 358 + ], + "type": "text", + "content": "[91] Hugo Touvron et al. LLaMA 2: Open foundation and fine-tuned chat models. arXiv:2307.09288, 2023. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 320, + 360, + 553, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 360, + 553, + 392 + ], + "spans": [ + { + "bbox": [ + 320, + 360, + 553, + 392 + ], + "type": "text", + "content": "[92] Du Tran, Lubomir Bourdev, Rob Fergus, Lorenzo Torresani, and Manohar Paluri. Learning spatiotemporal features with 3d convolutional networks. In ICCV, 2015. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 320, + 393, + 553, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 393, + 553, + 415 + ], + "spans": [ + { + "bbox": [ + 320, + 393, + 553, + 415 + ], + "type": "text", + "content": "[93] Gül Varol, Ivan Laptev, and Cordelia Schmid. Long-term temporal convolutions for action recognition. TPAMI, 2018. 1" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 320, + 417, + 553, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 417, + 553, + 448 + ], + "spans": [ + { + "bbox": [ + 320, + 417, + 553, + 448 + ], + "type": "text", + "content": "[94] Ramakrishna Vedantam, C Lawrence Zitnick, and Devi Parikh CIDER: Consensus-based image description evaluation. In CVPR, 2015. 5" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 320, + 449, + 553, + 471 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 449, + 553, + 471 + ], + "spans": [ + { + "bbox": [ + 320, + 449, + 553, + 471 + ], + "type": "text", + "content": "[95] Lucas Ventura, Cordelia Schmid, and Gül Varol. Learning text-to-video retrieval from image captioning. *IJCV*, 2024. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 320, + 472, + 553, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 472, + 553, + 504 + ], + "spans": [ + { + "bbox": [ + 320, + 472, + 553, + 504 + ], + "type": "text", + "content": "[96] Lucas Ventura, Antoine Yang, Cordelia Schmid, and Güll Varol. CoVR-2: Automatic data construction for composed video retrieval. TPAMI, 2024. 3" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 320, + 506, + 553, + 559 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 506, + 553, + 559 + ], + "spans": [ + { + "bbox": [ + 320, + 506, + 553, + 559 + ], + "type": "text", + "content": "[97] Caroline Violot, Tugrulcan Elmas, Igor Bilogrevic, and Mathias Humbert. Shorts vs. regular videos on YouTube: A comparative analysis of user engagement and content creation trends. In Proceedings of the 16th ACM Web Science Conference, Association for Computing Machinery, 2024. 1" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 320, + 560, + 553, + 603 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 560, + 553, + 603 + ], + "spans": [ + { + "bbox": [ + 320, + 560, + 553, + 603 + ], + "type": "text", + "content": "[98] Junke Wang, Dongdong Chen, Zuxuan Wu, Chong Luo, Luowei Zhou, Yucheng Zhao, Yujia Xie, Ce Liu, Yu-Gang Jiang, and Lu Yuan. Omnivl: One foundation model for image-language and video-language tasks. NeurIPS, 2022. 1" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 320, + 604, + 553, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 604, + 553, + 635 + ], + "spans": [ + { + "bbox": [ + 320, + 604, + 553, + 635 + ], + "type": "text", + "content": "[99] Teng Wang, Ruimao Zhang, Zhichao Lu, Feng Zheng, Ran Cheng, and Ping Luo. End-to-end dense video captioning with parallel decoding. In ICCV, 2021. 2" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 316, + 637, + 553, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 637, + 553, + 669 + ], + "spans": [ + { + "bbox": [ + 316, + 637, + 553, + 669 + ], + "type": "text", + "content": "[100] Yuxuan Wang, Difei Gao, Licheng Yu, Weixian Lei, Matt Feiszli, and Mike Zheng Shou. GEB+: A benchmark for generic event boundary captioning, grounding and retrieval. In ECCV, 2022. 2" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "text", + "content": "[101] Ziyang Wang, Shoubin Yu, Elias Stengel-Eskin, Jaehong Yoon, Feng Cheng, Gedas Bertasius, and Mohit Bansal. VideoTree: Adaptive tree-based video representation for LLM reasoning on long videos. arXiv:2405.19209, 2024. 3, 7" + } + ] + } + ], + "index": 36 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18957" + } + ] + } + ], + "index": 38 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 105 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 105 + ], + "type": "text", + "content": "[102] Xi Wei, Zhangxiang Shi, Tianzhu Zhang, Xiaoyuan Yu, and Lei Xiao. Multimodal high-order relation transformer for scene boundary detection. In ICCV, 2023. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 106, + 295, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 106, + 295, + 148 + ], + "spans": [ + { + "bbox": [ + 55, + 106, + 295, + 148 + ], + "type": "text", + "content": "[103] Chao-Yuan Wu, Yanghao Li, Karttikeya Mangalam, Haoqi Fan, Bo Xiong, Jitendra Malik, and Christoph Feichtenhofer. MeMViT: Memory-augmented multiscale vision transformer for efficient long-term video recognition. In CVPR, 2022. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 149, + 295, + 202 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 149, + 295, + 202 + ], + "spans": [ + { + "bbox": [ + 57, + 149, + 295, + 202 + ], + "type": "text", + "content": "[104] Haoqian Wu, Keyu Chen, Haozhe Liu, Mingchen Zhuge, Bing Li, Ruizhi Qiao, Xiujun Shu, Bei Gan, Liangsheng Xu, Bo Ren, Mengmeng Xu, Wentian Zhang, Raghavendra Ramachandra, Chia-Wen Lin, and Bernard Ghanem. NewsNet: A novel dataset for hierarchical temporal segmentation. In CVPR, 2023. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 203, + 295, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 203, + 295, + 236 + ], + "spans": [ + { + "bbox": [ + 57, + 203, + 295, + 236 + ], + "type": "text", + "content": "[105] Zuxuan Wu, Caiming Xiong, Chih-Yao Ma, Richard Socher, and Larry S. Davis. AdaFrame: Adaptive frame selection for fast video recognition. In CVPR, 2019. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 57, + 236, + 295, + 268 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 236, + 295, + 268 + ], + "spans": [ + { + "bbox": [ + 57, + 236, + 295, + 268 + ], + "type": "text", + "content": "[106] Junbin Xiao, Xindi Shang, Angela Yao, and Tat-Seng Chua. NExT-QA: Next phase of question-answering to explaining temporal actions. In CVPR, 2021. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 57, + 269, + 294, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 269, + 294, + 300 + ], + "spans": [ + { + "bbox": [ + 57, + 269, + 294, + 300 + ], + "type": "text", + "content": "[107] Jun Xu, Tao Mei, Ting Yao, and Yong Rui. MSR-VTT: A large video description dataset for bridging video and language. In CVPR, 2016. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 57, + 301, + 295, + 344 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 301, + 295, + 344 + ], + "spans": [ + { + "bbox": [ + 57, + 301, + 295, + 344 + ], + "type": "text", + "content": "[108] Hongwei Xue, Tiankai Hang, Yanhong Zeng, Yuchong Sun, Bei Liu, Huan Yang, Jianlong Fu, and Baining Guo. Advancing high-resolution video-language representation with large-scale video transcriptions. In CVPR, 2022. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 57, + 345, + 295, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 345, + 295, + 387 + ], + "spans": [ + { + "bbox": [ + 57, + 345, + 295, + 387 + ], + "type": "text", + "content": "[109] Antoine Yang, Arsha Nagrani, Ivan Laptev, Josef Sivic, and Cordelia Schmid. VidChapters-7M: Video chapters at scale. In NeurIPS Track on Datasets and Benchmarks, 2023. 1, 2, 3, 4, 5, 6, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 57, + 388, + 295, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 388, + 295, + 430 + ], + "spans": [ + { + "bbox": [ + 57, + 388, + 295, + 430 + ], + "type": "text", + "content": "[110] Antoine Yang, Arsha Nagrani, Paul Hongsuck Seo, Antoine Miech, Jordi Pont-Tuset, Ivan Laptev, Josef Sivic, and Cordelia Schmid. Vid2Seq: Large-scale pretraining of a visual language model for dense video captioning. In CVPR, 2023. 1, 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 57, + 432, + 295, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 432, + 295, + 463 + ], + "spans": [ + { + "bbox": [ + 57, + 432, + 295, + 463 + ], + "type": "text", + "content": "[111] Yang Yang, Yurui Huang, Weili Guo, Baohua Xu, and Dingyin Xia. Towards global video scene segmentation with context-aware transformer. AAAI, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 464, + 295, + 537 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 464, + 295, + 537 + ], + "spans": [ + { + "bbox": [ + 57, + 464, + 295, + 537 + ], + "type": "text", + "content": "[112] Yuan Yao, Tianyu Yu, Ao Zhang, Chongyi Wang, Junbo Cui, Hongji Zhu, Tianchi Cai, Haoyu Li, Weilin Zhao, Zhihui He, Qianyu Chen, Huarong Zhou, Zhensheng Zou, Haoye Zhang, Shengding Hu, Zhi Zheng, Jie Zhou, Jie Cai, Xu Han, Guoyang Zeng, Dahai Li, Zhiyuan Liu, and Maosong Sun. MiniCPM-V: A GPT-4V level MLLM on your phone. arXiv:2408.01800, 2024. 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 540, + 295, + 560 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 540, + 295, + 560 + ], + "spans": [ + { + "bbox": [ + 57, + 540, + 295, + 560 + ], + "type": "text", + "content": "[113] Fangqiu Yi, Hongyu Wen, and Tingting Jiang. ASformer: Transformer for action segmentation. In BMVC, 2021. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 562, + 295, + 594 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 562, + 295, + 594 + ], + "spans": [ + { + "bbox": [ + 57, + 562, + 295, + 594 + ], + "type": "text", + "content": "[114] Shoubin Yu, Jaemin Cho, Prateek Yadav, and Mohit Bansal. Self-chained image-language model for video localization and question answering. In NeurIPS, 2023. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 594, + 295, + 626 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 594, + 295, + 626 + ], + "spans": [ + { + "bbox": [ + 57, + 594, + 295, + 626 + ], + "type": "text", + "content": "[115] Rowan Zellers, Ximing Lu, Jack Hessel, Youngjae Yu, Jae Sung Park, Jize Cao, Ali Farhadi, and Yejin Choi. MERLOT: Multimodal neural script knowledge models. In NeurIPS, 2021. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 627, + 295, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 627, + 295, + 679 + ], + "spans": [ + { + "bbox": [ + 57, + 627, + 295, + 679 + ], + "type": "text", + "content": "[116] Andy Zeng, Maria Attarian, Brian Ichter, Krzysztof Choromanski, Adrian Wong, Stefan Welker, Federico Tombari, Aveek Purohit, Michael Ryoo, Vikas Sindhwani, et al. Socratic models: Composing zero-shot multimodal reasoning with language. In ICLR, 2023. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 681, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 681, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 681, + 295, + 712 + ], + "type": "text", + "content": "[117] Kuo-Hao Zeng, Tseng-Hung Chen, Juan Carlos Niebles, and Min Sun. Title generation for user generated videos. In ECCV, 2016. 2" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 555, + 466 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 105 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 105 + ], + "type": "text", + "content": "[118] Runhao Zeng, Wenbing Huang, Mingkui Tan, Yu Rong, Peilin Zhao, Junzhou Huang, and Chuang Gan. Graph convolutional networks for temporal action localization. In CVPR, 2019. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 106, + 555, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 106, + 555, + 137 + ], + "spans": [ + { + "bbox": [ + 317, + 106, + 555, + 137 + ], + "type": "text", + "content": "[119] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. ICCV, 2023. 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 139, + 555, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 139, + 555, + 160 + ], + "spans": [ + { + "bbox": [ + 316, + 139, + 555, + 160 + ], + "type": "text", + "content": "[120] Chenlin Zhang, Jianxin Wu, and Yin Li. ActionFormer: Localizing moments of actions with transformers. In ECCV, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 161, + 554, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 161, + 554, + 203 + ], + "spans": [ + { + "bbox": [ + 317, + 161, + 554, + 203 + ], + "type": "text", + "content": "[121] Ce Zhang, Taixi Lu, Md Mohaiminul Islam, Ziyang Wang, Shoubin Yu, Mohit Bansal, and Gedas Bertasius. A simple LLM framework for long-range video question-answering. In EMNLP, 2024. 2, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 205, + 554, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 205, + 554, + 236 + ], + "spans": [ + { + "bbox": [ + 316, + 205, + 554, + 236 + ], + "type": "text", + "content": "[122] Hang Zhang, Xin Li, and Lidong Bing. Video-LLaMA: An instruction-tuned audio-visual language model for video understanding. In EMNLP Demo, 2023. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 237, + 554, + 291 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 237, + 554, + 291 + ], + "spans": [ + { + "bbox": [ + 316, + 237, + 554, + 291 + ], + "type": "text", + "content": "[123] Shengyu Zhang, Ziqi Tan, Zhou Zhao, Jin Yu, Kun Kuang, Tan Jiang, Jingren Zhou, Hongxia Yang, and Fei Wu. Comprehensive information integration modeling framework for video titling. In ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, 2020. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 292, + 554, + 324 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 292, + 554, + 324 + ], + "spans": [ + { + "bbox": [ + 316, + 292, + 554, + 324 + ], + "type": "text", + "content": "[124] Cairong Zhao, Chutian Wang, Zifan Song, Guosheng Hu, Hao-nan Chen, and Xiaofan Zhai. Cap2Sum: Learning to summarize videos by generating captions. arXiv:2408.12800, 2024. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 325, + 553, + 367 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 325, + 553, + 367 + ], + "spans": [ + { + "bbox": [ + 316, + 325, + 553, + 367 + ], + "type": "text", + "content": "[125] Junjie Zhou, Yan Shu, Bo Zhao, Boya Wu, Shitao Xiao, Xi Yang, Yongping Xiong, Bo Zhang, Tiejun Huang, and Zheng Liu. MLVU: A comprehensive benchmark for multi-task long video understanding. arXiv:2406.04264, 2024. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 369, + 554, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 369, + 554, + 399 + ], + "spans": [ + { + "bbox": [ + 316, + 369, + 554, + 399 + ], + "type": "text", + "content": "[126] Luowei Zhou, Xu Chenliang, and Jason J. Corso. Towards automatic learning of procedures from web instructional videos. In AAAI, 2018. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 401, + 554, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 401, + 554, + 432 + ], + "spans": [ + { + "bbox": [ + 316, + 401, + 554, + 432 + ], + "type": "text", + "content": "[127] Luowei Zhou, Yingbo Zhou, Jason J Corso, Richard Socher, and Caiming Xiong. End-to-end dense video captioning with masked transformer. In CVPR, 2018. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 434, + 554, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 434, + 554, + 466 + ], + "spans": [ + { + "bbox": [ + 316, + 434, + 554, + 466 + ], + "type": "text", + "content": "[128] Xingyi Zhou, Anurag Arnab, Shyamal Buch, Shen Yan, Austin Myers, Xuehan Xiong, Arsha Nagrani, and Cordelia Schmid. Streaming dense video captioning. In CVPR, 2024. 2" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18958" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Charm_ The Missing Piece in ViT Fine-Tuning for Image Aesthetic Assessment/0ec61778-fb6e-453d-bfdb-3d054143034b_content_list.json b/2025/Charm_ The Missing Piece in ViT Fine-Tuning for Image Aesthetic Assessment/0ec61778-fb6e-453d-bfdb-3d054143034b_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..0bfdd7a08efe4718c5fa17343b703340cda9ece4 --- /dev/null +++ b/2025/Charm_ The Missing Piece in ViT Fine-Tuning for Image Aesthetic Assessment/0ec61778-fb6e-453d-bfdb-3d054143034b_content_list.json @@ -0,0 +1,1491 @@ +[ + { + "type": "text", + "text": "Charm: The Missing Piece in ViT fine-tuning for Image Aesthetic Assessment", + "text_level": 1, + "bbox": [ + 107, + 130, + 888, + 151 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Fatemeh Behrad, Tinne Tuytelaars, Johan Wagemans \nKU Leuven University, Belgium", + "text_level": 1, + "bbox": [ + 287, + 181, + 709, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{fatemeh.behrad, tinne.tuytelaars, johan.wagemans}@kuleuven.be", + "bbox": [ + 222, + 219, + 769, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 268, + 326, + 282 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The capacity of Vision transformers (ViTs) to handle variable-sized inputs is often constrained by computational complexity and batch processing limitations. Consequently, ViTs are typically trained on small, fixed-size images obtained through downscaling or cropping. While reducing computational burden, these methods result in significant information loss, negatively affecting tasks like image aesthetic assessment. We introduce Charm, a novel tokenization approach that preserves Composition, High-resolution, Aspect Ratio, and Multi-scale information simultaneously. Charm prioritizes high-resolution details in specific regions while downscaling others, enabling shorter fixed-size input sequences for ViTs while incorporating essential information. Charm is designed to be compatible with pre-trained ViTs and their learned positional embeddings. By providing multiscale input and introducing variety to input tokens, Charm improves ViT performance and generalizability for image aesthetic assessment. We avoid cropping or changing the aspect ratio to further preserve information. Extensive experiments demonstrate significant performance improvements on various image aesthetic and quality assessment datasets (up to $8.1\\%$ ) using a lightweight ViT backbone. Code and pre-trained models are available at https://github.com/FBehrad/Charm.", + "bbox": [ + 89, + 300, + 483, + 662 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 691, + 220, + 708 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Humans are inherently drawn to aesthetically pleasing images, making image aesthetics a vital consideration in fields like photography, advertising, e-commerce, and social media. Image aesthetic assessment (IAA) aims to develop computational models that can accurately predict human aesthetic preferences for images.", + "bbox": [ + 89, + 718, + 482, + 809 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Deep learning approaches applied to IAA have evolved significantly, with the Vision Transformer (ViT) architecture [9] demonstrating promising results [16, 21, 22]. ViTs process images as sequences of non-overlapping patches, allowing them to handle images of various sizes. However, in practice, this flexibility is limited for several rea", + "bbox": [ + 89, + 810, + 483, + 901 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9d6cb4720a139a08e737ee1fe9145a5835116f767d3a1b74c7867565304b0b9c.jpg", + "image_caption": [ + "Figure 1. Traditional image preprocessing methods (right) often distort image aesthetics by rescaling and cropping images to a fixed size. To preserve critical aesthetic information, our approach (left) selectively maintains some regions in their original resolution while downscaling the others." + ], + "image_footnote": [], + "bbox": [ + 524, + 267, + 906, + 573 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "sons. First, position embeddings of pre-trained ViTs are learned for a specific, fixed input size. Additionally, batch processing requires all sequences in a batch to be the same length, which is challenging with images of different dimensions. Grouping images by size or padding can help. However, grouping images may affect the learning process and convergence, as it makes batches less representative of the entire distribution, while padding can lead to inefficient resource utilization.", + "bbox": [ + 511, + 685, + 906, + 820 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "This brings us to another major issue when processing high-resolution images, which is the excessive computation and memory requirements. This can be particularly challenging when dealing with limited computational resources. As a result, a standard approach is to resize and", + "bbox": [ + 511, + 825, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "7815", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "downscale images to a smaller, fixed size to manage these constraints more effectively. This approach works for tasks where downscaling or cropping does not affect labels. However, in IAA, these operations can change aesthetic scores by altering the image composition and aspect ratio and losing high-resolution details (Figure 1). To unlock the full potential of ViTs in IAA, a new approach is needed to preserve critical aesthetic information while remaining computationally practical.", + "bbox": [ + 89, + 90, + 480, + 227 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Several techniques, such as graph attention networks [11], adaptive fractional dilated convolution networks [6], and multi-scale ViT [21], have been proposed to preserve image aspect ratio for IAA. Others, such as native resolution ViT [8], extend beyond IAA, demonstrating the benefits of preserving aspect ratio in other tasks. However, to remain practical, existing approaches still require downscaling the entire image, which limits their effectiveness. Fixed-size inputs can also limit the generalization of IAA models across various image sizes, which can be addressed by multi-scale inputs [18, 21, 23, 47].", + "bbox": [ + 89, + 229, + 480, + 396 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "One possible solution to preserve high-resolution information is a learnable resizer that reduces image dimensions while minimizing information loss during downsampling [7, 42]. While effective for other tasks, these methods can introduce distortions that alter the aesthetic quality of images. Another approach is to prioritize high-resolution information in some regions while downscaling less critical areas. This solution is inspired by the human brain, which processes visual information efficiently by preserving both low- and high-resolution details across the cortical hierarchy, maintaining spatial relationships, and utilizing selective attention and eye movements [12, 17]. Mixed-resolution [36] and mixed-scale tokenization [13] are steps in this direction. However, these methods produce identical representations for different resolutions of the same image because they first downscale all images to a fixed, low resolution for batch processing. While this approach is sufficient for image classification, it is inadequate for IAA, where resolution influences the aesthetic score.", + "bbox": [ + 89, + 398, + 482, + 685 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contribution. Existing approaches often prioritize a single factor crucial for aesthetics, neglecting others. Preserving composition, high resolution, aspect ratio, and multi-scale information simultaneously remains a challenge. We introduce Charm to fill this gap. Charm avoids fully downscaling images and maintains some areas in their original resolution (Figure 1). Our goal is to create a balance between computational efficiency and network performance. Charm effectively captures aspect ratio, composition, and multiscale information. Our framework-agnostic tokenization strategy avoids extensive pre-training by not modifying the pre-trained ViT architecture or its learned position embeddings. To demonstrate the broader applicability of our approach, we evaluate it on both IAA and image", + "bbox": [ + 89, + 689, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "quality assessment (IQA) tasks.", + "bbox": [ + 511, + 90, + 723, + 106 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 119, + 653, + 135 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Image aesthetic assessment Existing deep learning-based IAA methods often require images to be resized or cropped to a relatively small fixed size, leading to loss of aspect ratio, composition, and high-resolution detail. Using a batch size of 1 avoids the need for image downscaling or cropping [34]. However, this can be computationally inefficient for large-scale training. Graph-based representations, such as those proposed by Ghosal et al. [11], can be memory-intensive for large images. Similarly, adaptive fractional dilated convolutions, introduced by Chen et al. [6], can be computationally expensive and require careful batch processing. A more recent strategy for preserving high resolution and aspect ratio is the use of learnable resizers. These methods employ auxiliary networks to reduce the dimensionality of input images while minimizing information loss during downsampling [7, 42, 45, 46]. While effective for image classification, these approaches introduce distortions that negatively impact image aesthetics.", + "bbox": [ + 511, + 143, + 903, + 417 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Another challenge associated with fixed-size inputs is a decrease in model generalization across various image sizes. While this challenge can be mitigated by supplying multi-scale inputs to the network [18, 21, 23, 47], existing approaches introduce non-negligible computational costs.", + "bbox": [ + 511, + 417, + 903, + 492 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Vision transformers Among existing neural networks, ViTs offer the advantage of processing images in their original size. However, they are still constrained by the need for fixed-size inputs due to batch processing. NaViT [8] addresses this problem by packing patches from different images into a fixed-size sequence. However, this approach can be computationally expensive as it can result in excessively long input sequences for high-resolution images.", + "bbox": [ + 511, + 493, + 903, + 613 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The computational costs of ViTs are influenced by the number of input tokens they process, so downscaling images can reduce the number of tokens and improve efficiency. Token dropping, either randomized or structured [27], and token merging [3] are techniques to improve ViT efficiency. Some methods, like token fusion [25], combine these approaches to achieve even better results.", + "bbox": [ + 511, + 614, + 903, + 718 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Previous studies have shown that incorporating multiscale features can enhance ViT performance. MUSIQ [21], for instance, utilizes patches from different scales to capture multi-scale information. While effective, this approach modifies the first layers of the ViT and introduces new position embeddings, which necessitates extensive training for each ViT backbone. FlexiViT [2], a more recent approach, extracts multiscale features using random patch sizes in each training epoch.", + "bbox": [ + 511, + 719, + 903, + 854 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Closer to our approach are mixed-resolution [36] and mixed-scale tokenization [13]. These methods first downscale images to a fixed size, then extract two-scale patches,", + "bbox": [ + 511, + 854, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "7816", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "preserving key regions in higher resolution while downscaling others. However, this initial downscaling discards high-frequency details and limits patches to two fixed resolutions. In contrast, Charm retains high-frequency information from the original image, learning from patches at varying resolutions (see Appendix 6 for details). Mixed-resolution ViT [36] further limits the input format, requiring images to be square. Additionally, it replaces learned position embeddings with sinusoidal ones, necessitating additional pre-training of each ViT backbone for compatibility.", + "bbox": [ + 89, + 90, + 480, + 241 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Existing methods fail to simultaneously preserve multiscale information, aspect ratio, composition, and high-resolution details. While all of these factors are crucial for accurate IAA, current methods typically prioritize computational efficiency at the expense of preserving these critical aspects. In contrast, our method successfully preserves all these factors while remaining computationally practical.", + "bbox": [ + 89, + 242, + 480, + 347 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Position embedding Transformers lack inherent mechanisms to capture the order of elements within an input sequence [44]. This stems from their reliance on self-attention, which is permutation-invariant – meaning the order of elements in the input does not affect the final output [1]. As a result, they require position embeddings to encode the sequence order. These position embeddings can be a deterministic function of position, typically using sinusoidal functions [44] or learned representations. Learnable position embeddings have achieved comparable performance to deterministic approaches [9]. However, both deterministic and learnable position embeddings are pre-trained for a specific sequence length.", + "bbox": [ + 89, + 347, + 482, + 542 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Relative position embeddings offer a more flexible approach by capturing the relative distance between elements [38]. This approach is well-suited for variable-length inputs. However, it necessitates significant modifications to the transformer attention mechanism. Recent methods like hash-based spatial embedding [21] and factorized position embedding [8] support variable aspect ratios and input resolutions without modifying attention modules. However, they cannot replace the position embedding of a pre-trained transformer without extensive retraining of the backbone on large-scale datasets. In contrast, our method adapts the position embedding of a pre-trained ViT model, requiring no additional pre-training and enabling efficient fine-tuning on downstream tasks with limited data.", + "bbox": [ + 89, + 544, + 482, + 753 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Unlock the potentials of ViTs in IAA", + "text_level": 1, + "bbox": [ + 89, + 768, + 419, + 786 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we introduce Charm, a novel tokenization approach designed to enhance the performance of standard ViTs in image aesthetic and quality assessment tasks. Figure 2 provides an overview of our proposed approach.", + "bbox": [ + 89, + 794, + 480, + 854 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We initially partition the input image into fixed-size patches, preserving a selection of patches at their original resolution while downscaling others. This approach enables", + "bbox": [ + 89, + 854, + 482, + 900 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/70910cb10c8517a4894a13e7f87ca44c2e3a2e4cb9c58c816255eea38be65c48.jpg", + "image_caption": [ + "Figure 2. Our approach involves partitioning the input image into fixed-size patches and selectively preserving a subset at its original resolution while downscaling others. Position and scale embeddings are then prepared to encode spatial relationships across different scales. The remaining steps follow the standard ViT. + indicates element-wise addition." + ], + "image_footnote": [], + "bbox": [ + 555, + 88, + 867, + 364 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "the model to capture both fine-grained details and broader contextual information. Subsequently, position embeddings are prepared to encode the spatial relationships between patches across different resolutions and within the same resolution. Following previous studies [21], we also incorporate scale embeddings to aid the model in distinguishing information from different scales.", + "bbox": [ + 511, + 476, + 903, + 580 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "After encoding the input patches into a sequence of tokens using the patch encoding module, we add position and scale embeddings. A learnable \"classification token\" (CLS) is prepended, and all tokens are processed through the transformer encoder. The final output is obtained by feeding the CLS token state at the output of the transformer through a fully connected layer. Our approach is compatible with various ViT architectures, as we do not modify the core ViT network structure.", + "bbox": [ + 511, + 584, + 905, + 717 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminaries", + "text_level": 1, + "bbox": [ + 511, + 734, + 653, + 750 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "An image $x \\in \\mathbb{R}^{h \\times w \\times c}$ , where $(h, w, c)$ represents the height, width, and number of channels, respectively, undergoes tokenization in a standard ViT. This process divides the image into a sequence of $s$ patches $x_i \\in \\mathbb{R}^{p \\times p \\times c}$ , where $i \\in (1, \\ldots, s)$ and $p$ represents the patch size. The sequence length $s = \\left\\lfloor \\frac{h}{p} \\right\\rfloor \\cdot \\left\\lfloor \\frac{w}{p} \\right\\rfloor$ determines the number of patches after tokenization, controlling the ViT's computational complexity. Given that the patch size $p$ remains fixed for a given ViT, the number of patches depends on the in-", + "bbox": [ + 511, + 757, + 906, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "7817", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "put image resolution. Consequently, high-resolution images increase computational complexity.", + "bbox": [ + 89, + 90, + 485, + 122 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Charm tokenization", + "text_level": 1, + "bbox": [ + 89, + 128, + 282, + 143 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.1. Tokenization", + "text_level": 1, + "bbox": [ + 89, + 152, + 227, + 165 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Batch processing necessitates a consistent spatial resolution within each minibatch, leading to a fixed input length $(l)$ for the ViT. Processing high-resolution images requires a large $l$ , which increases memory consumption and computational demands during training. Our method addresses this by strategically reducing $l$ while preserving aesthetic-related information (Figure 2). Given a fixed patch size $(p)$ , we can consider two scenarios. For low-resolution images $(s \\leq l)$ , we follow the standard ViT approach, tokenizing the entire image with the patch size of $p$ . For high-resolution images $(s > l)$ , we keep some parts in high resolution while downscaling other areas.", + "bbox": [ + 89, + 170, + 483, + 349 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "When dealing with 2 scales, we initially tokenize the image using a patch size of $p' = p \\times n$ , where $n \\geq 2$ . Subsequently, we select high-resolution patches and further tokenize them using a patch size of $p$ while the remaining patches are downscaled to $p$ . The downscaling factor for both height and width is $f = 1/n$ . Finally, we arrange all high-resolution patches at the beginning, followed by low-resolution ones to form the final input for the transformer encoder. This ordering is feasible due to the permutation invariance of ViTs [1], allowing us to avoid unnecessary implementation loops. To ensure a consistent input length ( $l$ ), padding or random token dropping is employed as needed. Additional details on the 3-scale version are available in the Appendix 7.", + "bbox": [ + 89, + 352, + 483, + 563 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.2. Patch Importance", + "text_level": 1, + "bbox": [ + 89, + 569, + 263, + 585 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Several strategies can be employed to determine which patches to retain at high resolution. The main goal is to preserve visually interesting areas and sharp details in high resolution. To identify these regions, we use frequency, gradient, entropy, and saliency maps. We compare them with a random selection of patches in each training epoch. More details on patch selection strategies, including illustrative examples, can be found in Figure 8 of Appendix 8.", + "bbox": [ + 89, + 588, + 483, + 709 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.3. Position Embedding", + "text_level": 1, + "bbox": [ + 89, + 715, + 277, + 729 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ViTs lack inherent knowledge of spatial relationships between patches. Position embeddings address this by injecting learned position information into each patch embedding, allowing the self-attention mechanism to capture how different parts of the image relate. Replacing the standard position embeddings of ViTs with custom ones [36] can be expensive, requiring extensive pre-training on large datasets. It also introduces compatibility issues when used with different pre-trained ViTs. To address these issues, our approach adapts the pre-trained position embeddings from the standard ViT models.", + "bbox": [ + 89, + 734, + 483, + 898 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For low-resolution images, we follow the standard ViT approach of resizing pre-trained position embeddings via bilinear interpolation. However, for high-resolution images, we create two position embeddings using bilinear interpolation based on the different patch sizes $(p$ and $p^{\\prime})$ employed during tokenization (Figure 2). One position embedding retains the spatial relationships between low-resolution patches, while the other captures the finer spatial relationships within the high-resolution patches. These position embeddings, high and low resolution, are compatible because they are derived from the same pre-trained embedding. Similar to standard ViTs, each token from the output of the patch encoding module is then summed with its corresponding position embedding.", + "bbox": [ + 511, + 90, + 903, + 301 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Unlike existing methods [13, 36], our approach does not have a fixed image size, so position embeddings vary for each image. To avoid recalculating position embeddings in each epoch, we prepare them in the data loaders, keeping them frozen during fine-tuning. This aligns with the standard practice of freezing the initial layers (including position embeddings) and fine-tuning only the final layers.", + "bbox": [ + 511, + 303, + 903, + 409 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.4. Scale Embedding", + "text_level": 1, + "bbox": [ + 511, + 419, + 681, + 434 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our tokenization approach results in multiscale tokens. Following previous studies [21], to help ViT distinguish between patches from different scales and capture their relationships, we incorporate learnable scale embeddings. These scale embeddings are randomly initialized and added element-wise to all patch embeddings (Figure 2).", + "bbox": [ + 511, + 438, + 903, + 530 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Fine-tuning", + "text_level": 1, + "bbox": [ + 511, + 541, + 640, + 556 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our method requires no additional pre-training. We process images in their original resolution without resizing or random cropping during fine-tuning. We avoid normalizing the images as normalization can introduce distortions that affect the overall aesthetic/quality score (Table 9 in Appendix 9). Our data augmentation is limited to random horizontal flipping and random rotations (90, 180, 270 degrees) with a $50\\%$ probability, as they do not result in loss of composition or aspect ratio. While these methods might change the aesthetics for humans, they have been shown to improve the performance of ViTs (Appendix 10). During evaluation, data augmentation is completely omitted. We employ the L1 loss function for predicting the overall aesthetic/quality score and the Earth Mover's Distance (EMD) loss [37] for predicting the aesthetic score distribution.", + "bbox": [ + 511, + 564, + 906, + 790 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 806, + 645, + 823 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Datasets", + "text_level": 1, + "bbox": [ + 511, + 832, + 614, + 847 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We evaluate our models on various IAA datasets, including AVA [35], PARA [48], TAD66k [14], AADB [26], BAID [49], and two IQA datasets, SPAQ [10] and KonIQ10k [19].", + "bbox": [ + 511, + 854, + 903, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "7818", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/de01c09f0bdd248741726e0cb4a83fbb33c7de8e365abd857bbfbf066e5376e8.jpg", + "image_caption": [ + "Figure 3. An image with its aesthetic score distribution and mean aesthetic score, from the AVA dataset." + ], + "image_footnote": [], + "bbox": [ + 94, + 87, + 480, + 207 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We only use the images and their corresponding ratings, excluding any additional information provided in the datasets. The AVA and PARA datasets associate each image with an aesthetic score distribution, while other datasets associate each image with an overall aesthetic/quality score (Figure 3). As the SPAQ and KonIQ10k datasets lack an official train-test split, we repeat the experiments for 5 different splits and report the median performance (as done in [41]).", + "bbox": [ + 88, + 260, + 482, + 381 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The AVA dataset remains prominent due to its large number of images, allowing for effective training of large models. Research utilizing smaller datasets is relatively limited, and most existing approaches leverage the additional information provided in these datasets. However, considering various datasets with different resolution ranges allows us to validate the efficiency of our approach in various scenarios. For more details on the resolution distribution of each dataset, see Figure 13 in Appendix 11.", + "bbox": [ + 88, + 382, + 482, + 518 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Implementation details", + "text_level": 1, + "bbox": [ + 89, + 529, + 305, + 544 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Saliency maps for patch selection are generated using the SAM-HQ model [24] with the text prompt of \"salient object\" (Appendix 8). For gradient, frequency, and entropy calculations, we use the Sobel filter, fast Fourier transform, and Shannon entropy, respectively.", + "bbox": [ + 88, + 551, + 482, + 627 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In all experiments, we fine-tune the entire ViT models, while keeping their position embeddings frozen. Also, we use scale embeddings alongside position embeddings to enable Charm to capture relationships between tokens both within and across scales (see Appendix 12). The patch size $(p)$ and patch stride are chosen to match the filter size and stride of the patch encoding modules within these models (16 in ViT-small and 14 in Dinov2-small and Dinov2-large), ensuring compatibility and replicating the models' internal tokenization process.", + "bbox": [ + 88, + 628, + 482, + 779 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The maximum number of tokens $(l)$ is set to 512 for AVA, 768 for TAD66k, and 1024 for other datasets. These values are chosen based on the image resolution in the datasets (Appendix 13). For all experiments, we use a factor size of 0.5.", + "bbox": [ + 89, + 779, + 482, + 854 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The AdamW optimizer [33] is employed for fine-tuning. We use an initial learning rate of 1e-5 for all datasets except BAID (1e-6). We gradually reduce the learning rate dur", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/8c86b8054f7412a05cdd9bac202b68e8275ce6ebc568d074c8e8cd3310b42eee.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetCharmPLCCSRCCACC
IAA
AVA-0.7100.7060.802
✓0.779(↑ 6.9%)0.777(↑ 7.1%)0.826(↑ 2.4%)
AADB-0.6950.6820.754
✓0.767(↑ 7.2%)0.754(↑ 7.2%)0.767(↑ 1.3%)
TAD66k-0.4290.4010.646
✓0.488(↑ 5.9%)0.458(↑ 5.7%)0.794 (↑ 14.8%)
PARA-0.9040.8550.863
✓0.938(↑ 3.4%)0.905(↑ 5%)0.892(↑ 2.9%)
BAID-0.4280.3420.750
✓0.439(↑ 1.1%)0.368(↑ 2.6%)0.763(↑ 1.3%)
IQA
SPAQ-0.9110.9070.907
✓0.919(↑ 0.8%)0.915(↑ 0.8%)0.917(↑ 1.0%)
KonIQ10k-0.8960.8680.938
✓0.944(↑ 4.8%)0.93(↑ 6.2%)0.954(↑ 1.6%)
", + "bbox": [ + 517, + 87, + 903, + 429 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1. Performance improvement across different IAA and IQA datasets by replacing the standard tokenization of Dinov2-small with Charm (with a random patch selection strategy).", + "bbox": [ + 511, + 439, + 906, + 482 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "ing training using the cosine annealing learning rate scheduler. The batch size is 64 for AVA, 32 for TAD66k, and 16 for the remaining datasets. Early stopping is also implemented to prevent overfitting. Fine-tuning is conducted on a single NVIDIA RTX A4500 GPU. Model performance is evaluated using three key metrics: Spearman rank-ordered correlation coefficient (SRCC), Pearson linear correlation coefficient (PLCC), and classification accuracy (ACC).", + "bbox": [ + 511, + 500, + 906, + 621 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Comparing with existing methods", + "text_level": 1, + "bbox": [ + 511, + 633, + 808, + 648 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We benchmark Charm with various ViT backbones on different IAA and IQA datasets. We first fine-tune models on the datasets to create a baseline. We then replace their original patch encoding modules with two-scale Charm tokenization and fine-tune the entire models. We also considered a three-scale version of Charm but did not observe additional performance gains (Appendix 7). Therefore, we focus on the two-scale version throughout the paper.", + "bbox": [ + 511, + 657, + 906, + 777 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The results (Table 1 and Table 13 in Appendix 14) demonstrate that by simply replacing the patch encoding module of existing ViT models with Charm, we can achieve significant performance improvements, up to $7.5\\%$ in PLCC, $8.1\\%$ in SRCC, and $14.8\\%$ in classification accuracy across various datasets. Among datasets, PARA and SPAQ datasets include extra high-resolution images (e.g., 2k by 3k pixels). Our experiments (Appendix 15) reveal", + "bbox": [ + 511, + 779, + 908, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "7819", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/0a48bbd50870ed4473c3bdc5e0f4ee7efb2285272ae3e4890dcd02421f79454e.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelARHRMSPLCCSRCCACC
Ghosal et al.[11]✓--0.7640.762-
Chen et al. [6]✓--0.6710.6490.832
Dinov2-Small (+ Padding) *✓✓-0.7090.7030.801
MUSIQ [21]✓-✓0.7380.7260.815
FlexiViT [2]*--✓0.7370.7350.812
Swin [32]*--✓0.7480.7510.816
Dinov2-Small (+ Muller [43]) *-✓-0.6820.6750.794
ViT-small (+ Charm) *✓✓✓0.7620.7600.827
Dinov2-small (+ Charm) *✓✓✓0.7790.7770.826
Dinov2-large (+ Charm) *✓✓✓0.7830.7810.828
", + "bbox": [ + 250, + 87, + 746, + 244 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Results on the AVA dataset. Bold and underlined numbers indicate the best and the second-best results, respectively. AR, HR, and MS represent preserving aspect ratio, high-resolution details, and multi-scale information, respectively. * indicates models that we have fine-tuned on the AVA dataset. Appendix 14 shows the performance of the standard ViT-small, Dinov2-small, and Dinov2-large.", + "bbox": [ + 89, + 255, + 906, + 299 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "a threshold for performance improvement by preserving high-resolution information. Beyond this threshold, further performance gains are limited. Therefore, to decrease computational costs, we downscale images in PARA and SPAQ to a maximum edge of 1024.", + "bbox": [ + 88, + 324, + 480, + 400 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As shown in Table 2, our approach outperforms existing methods in terms of preserving resolution, aspect ratio, and multiscale information. We compare Charm with the simple approach of adding padding to ensure consistent input shapes within each batch while preserving the aspect ratio. However, this padding proves detrimental to IAA and significantly increases computational costs (Appendix 16). Also, we observe that Muller [43], a recent approach focusing on preserving high-resolution information, can introduce noise and distortions to the image, negatively affecting aesthetics (Appendix 16).", + "bbox": [ + 88, + 400, + 482, + 566 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Recent studies have demonstrated the effectiveness of incorporating additional modalities, such as attributes ([5, 20, 28]) or text captions ([29, 30, 39, 50]), to improve IAA model performance (Table 15 in Appendix 17). Other studies have concentrated on other research directions [15, 40], like addressing the long tails of distributions in IAA datasets [31]. While the main contributions of these studies are orthogonal to Charm's, integrating Charm with them further enhances their performance. For example, incorporating Charm into ELTA [31] with a Dinov2-large backbone improves its PLCC, SRCC, and ACC on the AVA dataset from 0.742, 0.742, and 0.811 to 0.787, 0.786, and 0.829, respectively. Also, our approach demonstrates strong performance in IQA, with a slight performance gap of less than $1\\%$ in SPAQ and $0.4\\%$ in KonIQ10k compared to state-of-the-art methods [4, 21, 41] (Table 16 in Appendix 17).", + "bbox": [ + 89, + 566, + 482, + 809 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Model size is another crucial factor, as larger models often excel in complex tasks like IAA. In this paper, we demonstrate that Charm is highly effective even with lightweight ViTs, which can be used for applications with limited computational resources. As shown in Figure 4, Dinov2-small + Charm achieves comparable performance", + "bbox": [ + 89, + 810, + 482, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "to state-of-the-art models while being significantly smaller. Nevertheless, Charm can serve as a plug-in for larger, more expensive models if higher performance is desired and computational power allows. For instance, plugging Charm into ViT-base instead of Dinov2-small results in PLCC of 0.515, SRCC of 0.467, and ACC of 0.779 after only 9 epochs on the BAID dataset.", + "bbox": [ + 511, + 324, + 906, + 430 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Ablation study", + "text_level": 1, + "bbox": [ + 511, + 445, + 663, + 462 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4.1. The importance of preserving aspect ratio and high-resolution details", + "text_level": 1, + "bbox": [ + 511, + 468, + 905, + 500 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Current approaches usually distort images by resizing them to a square shape, disregarding the significance of maintaining their original size and aspect ratio. We argue that such preprocessing can be detrimental to IAA. To validate this claim, we compare the performance of Dinov2-small using different tokenization approaches. Table 3 presents the results of these experiments. Except for the standard tokenization (described in Appendix 9), all other settings involve multiscale inputs and maintain composition by avoiding image cropping.", + "bbox": [ + 511, + 503, + 905, + 655 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We initially downscale images to $256 \\times 256$ and then preserve some areas at that resolution while downscaling others by a factor of 0.5 (MS in Table 3). While this approach has been effective in image classification ([13] and [36]), it decreases performance in our IAA task. This is because the initial downscaling to a fixed size results in significant information loss critical for IAA, and further downscaling worsens the issue.", + "bbox": [ + 511, + 656, + 905, + 777 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We then experiment with keeping the image in its original resolution while changing the aspect ratio by resizing the image based on the maximum edge to create a square image (HR + MS in Table 3). This results in a nearly $3\\%$ increase in PLCC and SRCC compared to the standard approach. Finally, we use Charm, which preserves both aspect ratio and high-resolution details, resulting in the best performance. This analysis highlights the importance of preserv", + "bbox": [ + 511, + 779, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "7820", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5921f7f7008a27a222885dcd325854054845763dd141457a698dc98513d1ba9a.jpg", + "image_caption": [ + "Figure 4. Comparison of state-of-the-art IAA models' performance on the AVA dataset vs. their number of parameters. Our approach achieves comparable performance without using additional modalities and with significantly fewer parameters. To account for the significant differences in the number of parameters, we use a logarithmic scale on the x-axis. For a detailed comparison of Dinov2-small + Charm with state-of-the-art models on all datasets, see Appendix 17." + ], + "image_footnote": [], + "bbox": [ + 93, + 88, + 480, + 276 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/4c5af7659f6068b611642bdbdf6c3961a80e259980a332bb88891b338728651f.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
TokenizationPLCCSRCCACC
Standard0.7340.7320.808
MS0.7100.7050.801
HR + MS0.7630.7600.820
Charm (AR + HR + MS)0.7790.7770.826
", + "bbox": [ + 119, + 412, + 455, + 488 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ing both aspect ratio and original resolution in IAA tasks.", + "bbox": [ + 89, + 570, + 470, + 585 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We also examined the impact of different numbers of tokens across tokenization approaches in Table 3 but found no differences in the results (see Appendix 18).", + "bbox": [ + 89, + 585, + 482, + 631 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4.2. Comparison of patch selection strategies", + "text_level": 1, + "bbox": [ + 89, + 638, + 416, + 654 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To evaluate the importance of patches, we fine-tune Dinov2-small on the AADB dataset with different patch selection strategies. Table 4 demonstrates that all strategies outperform the standard ViT approach [9]. This suggests that factors beyond preserving high-resolution details also contribute to performance improvements (Section 4.4.1).", + "bbox": [ + 89, + 657, + 482, + 748 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Among all methods, the random selection of patches in each epoch stands out, likely for two reasons. First, ViTs are prone to overfitting, especially when fine-tuned on small datasets, a common challenge in IAA. Random patch selection in each epoch addresses this by introducing more variety to the ViT input and acting as a form of data augmentation. Second, random sampling can better capture the subjective and often unpredictable nature of beauty, providing a more representative view of the entire image. Following random selection, frequency, gradient, and entropy-based", + "bbox": [ + 88, + 750, + 482, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/47fa52318501d89690f165b66dcf8bc1c43f0f8eb5c12bacec3b212fee325d80.jpg", + "table_caption": [ + "Table 3. Performance of Dinov2-small on the AVA dataset with four different tokenization strategies. MS, HR, and AR represent preserving multiscale information, high-resolution details, and aspect ratio, respectively." + ], + "table_footnote": [], + "table_body": "
Patch selectionPLCCSRCCACC
Random0.7670.7540.767
Entropy0.7260.7140.761
Frequency0.7560.7470.761
Gradient0.7340.7210.766
Saliency0.7510.7380.756
Standard0.6950.6820.754
", + "bbox": [ + 571, + 87, + 849, + 189 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. Performance of Dinov2-small on the AADB dataset using Charm with different patch selection strategies. Bold and underlined values highlight the best and second-best results.", + "bbox": [ + 511, + 199, + 905, + 242 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/11376fa752a04693d6692ea20f5fbbad7a86a080f479408b30da4065fc7d45cc.jpg", + "image_caption": [ + "Figure 5. The epoch at which each model achieves its highest validation performance across different datasets. Charm generally leads to faster convergence. The average number of epochs over 5 runs is reported for SPAQ and KonIQ10k." + ], + "image_footnote": [], + "bbox": [ + 514, + 256, + 901, + 385 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "methods stand out. These methods still have a certain level of randomness, as their fully deterministic version results in highly correlated selected patches (Appendix 8).", + "bbox": [ + 511, + 465, + 905, + 510 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ViTs, originally trained for object classification, have an inherent bias towards salient objects. However, IAA requires attention to both foreground and background [15], which explains why the saliency-based approach performs less effectively. For a more comprehensive analysis of patch selection strategies please refer to Appendix 8.", + "bbox": [ + 511, + 510, + 905, + 602 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4.3. The effect of Charm on convergence speed", + "text_level": 1, + "bbox": [ + 511, + 611, + 854, + 626 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We fine-tune all models until convergence, meaning their performance on the validation set stabilizes. Figure 5 shows the epoch at which each model achieves its best validation performance. As shown in this figure, Charm consistently accelerates the convergence of Dinov2-small across different datasets. Notably, the most significant reduction (50%) in training epochs occurs on AADB, our smallest dataset, likely due to Charm's ability to mitigate overfitting through random patch selection (see Appendix 8). On our largest datasets, AVA and TaD66k, both standard and Charm tokenization converge at similar epochs.", + "bbox": [ + 511, + 630, + 905, + 796 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4.4. Computational costs", + "text_level": 1, + "bbox": [ + 511, + 805, + 702, + 820 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Figure 6 compares the runtime of Charm using different patch selection strategies with the standard approach on a 12th Gen Intel(R) Core(TM) i9-12900K CPU. Table 3 demonstrates that multiscale inputs can negatively impact performance for very small images. In this case, we use a", + "bbox": [ + 511, + 825, + 905, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7821", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/507c4a082204bde3c10408a461700aa528a3708f8eb28c929075bbd86e4f19b1.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelInput sizeCharm#tokensmsGMACsMB
Dinov2224 x 224-2565.76.11202.9
-small-207032.884.012091.8
640 x 640✓2-scale:5127.3 (↓ 77.7%)13.46 (↓ 84%)346.0 (↓ 83.5%)
✓3-scale:7009.3 (↓ 71.6%)19.60 (↓ 76.7%)494.3 (↓ 76.4%)
", + "bbox": [ + 184, + 87, + 812, + 161 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. Dinov2-small inference cost breakdown for processing one single image: number of tokens (#tokens) based on varying input sizes, runtime in milliseconds (ms), Giga multiply accumulation (GMACs), and GPU memory in Megabytes (MB). Bold and underlined values highlight the most and second-most computationally efficient configurations. Percentages indicate the reduction in computational cost compared to processing the image in its original size. Similar results for ViT-small are presented in Table 17 in the Appendix.", + "bbox": [ + 89, + 172, + 906, + 229 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/64a2e6d2dd79a19fbd594a407610eff45754782f191d9deb597270e756b15a3e.jpg", + "image_caption": [ + "Figure 6. Comparison of Charm runtime using different patch selection strategies with the standard approach (Appendix 9). The x-axis shows the total number of pixels." + ], + "image_footnote": [], + "bbox": [ + 114, + 253, + 460, + 445 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "single-scale tokenization approach (Section 3.2.1). Figure 6 illustrates the increase in the computational cost of Charm with larger image sizes. To reduce the overhead of saliency mask generation (16 seconds per image), we employ precomputed masks. The frequency-based method provides a cost-effective alternative to on-the-fly saliency detection, identifying regions comparable to or even superior to those selected by SAM-HQ. Random patch selection emerges as the most effective strategy, offering the best performance with the lowest computational cost.", + "bbox": [ + 88, + 518, + 482, + 670 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The computational cost of ViTs is influenced by the number of input tokens. The standard method involves downscaling and cropping images to $224 \\times 224$ , resulting in a very low number of tokens and lower costs. However, processing images at their original size can significantly increase computational costs. Our approach not only improves performance but also significantly reduces the computational costs of processing images in the original resolution (Table 5). We observe a decrease of over $84\\%$ in GMACs, $77\\%$ in runtime, $83\\%$ in GPU memory consumption, and $75\\%$ in the number of tokens for Dinov2-small when processing one $640 \\times 640$ image using a 2-scale Charm. Even the 3-scale version, which is more costly, has significantly lower costs compared to processing images at their original size. While Charm may incur slightly", + "bbox": [ + 89, + 674, + 482, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "higher computational costs compared to the standard approach of downscaling and cropping, the substantial performance gains make it a worthwhile trade-off.", + "bbox": [ + 511, + 255, + 906, + 300 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4.5. Downscaling factor $(f)$", + "text_level": 1, + "bbox": [ + 511, + 310, + 720, + 325 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our method initially tokenizes images with an initial patch size $(p')$ . Subsequently, selected patches are further tokenized using a smaller patch size $(p)$ , while the remaining patches are downscaled to $p$ . We experiment with different initial patch sizes $p' = (28, 42, 56)$ , corresponding to downscaling factors $(f)$ of 0.5, 0.66, and 0.75. The best performance is achieved with a factor of 0.5 (Appendix 19). A larger factor negatively impacts performance, as it leads to more significant downscaling of unselected patches, resulting in information loss for IAA.", + "bbox": [ + 511, + 329, + 906, + 481 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 494, + 633, + 511 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Processing images in their original resolution can be computationally expensive, requiring significant memory and processing power. Also, it is challenging to process images of varying aspect ratios and resolutions due to batch processing constraints. Charm effectively balances computational cost with the preservation of composition, high-resolution details, aspect ratio, and multi-scale information, which are all crucial for achieving optimal performance in IAA. By introducing variability into the input, our method enhances the generalizability of ViTs in IAA. Charm is compatible with any ViT architecture (except Swin [32], see Appendix 20) and does not require additional pre-training of ViTs. Our method consistently outperforms existing approaches that address similar research questions, achieving promising results across diverse IAA and IQA datasets.", + "bbox": [ + 511, + 521, + 906, + 748 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "While our approach effectively preserves critical aesthetic-related factors, there's a potential for information loss due to selective downscaling. To further improve performance, future work could explore more sophisticated patch selection strategies, incorporate other modalities, and integrate our method with existing IAA-specific networks.", + "bbox": [ + 511, + 750, + 908, + 840 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements: Funded by the European Union (ERC AdG, GRAPPA, 101053925, awarded to Johan Wagemans) and the Research Foundation-Flanders (FWO, 1159925N, awarded to Fatemeh Behrad).", + "bbox": [ + 511, + 840, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "7822", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Irwan Bello, Barret Zoph, Ashish Vaswani, Jonathon Shlens, and Quoc V Le. Attention augmented convolutional networks. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3286-3295, 2019. 3, 4", + "[2] Lucas Beyer, Pavel Izmailov, Alexander Kolesnikov, Mathilde Caron, Simon Kornblith, Xiaohua Zhai, Matthias Minderer, Michael Tschannen, Ibrahim Alabdulmohsin, and Filip Pavetic. Flexivit: One model for all patch sizes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14496-14506, 2023. 2, 6", + "[3] Daniel Bolya, Cheng-Yang Fu, Xiaoliang Dai, Peizhao Zhang, Christoph Feichtenhofer, and Judy Hoffman. Token merging: Your vit but faster. arXiv preprint arXiv:2210.09461, 2022. 2", + "[4] Chaofeng Chen, Jiadi Mo, Jingwen Hou, Haoning Wu, Liang Liao, Wenxiu Sun, Qiong Yan, and Weisi Lin. Topiq: A top-down approach from semantics to distortions for image quality assessment. IEEE Transactions on Image Processing, 2024. 6", + "[5] Hangwei Chen, Feng Shao, Baoyang Mu, and Qiuping Jiang. Image aesthetics assessment with emotion-aware multibranch network. IEEE Transactions on Instrumentation and Measurement, 73:1-15, 2024. 6", + "[6] Qiuyu Chen, Wei Zhang, Ning Zhou, Peng Lei, Yi Xu, Yu Zheng, and Jianping Fan. Adaptive fractional dilated convolution network for image aesthetics assessment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14114-14123, 2020. 2, 6", + "[7] Bowen Cheng, Bin Xiao, Jingdong Wang, Honghui Shi, Thomas S Huang, and Lei Zhang. Higherhrnet: Scale-aware representation learning for bottom-up human pose estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5386-5395, 2020. 2", + "[8] Mostafa Dehghani, Basil Mustafa, Josip Djolonga, Jonathan Heek, Matthias Minderer, Mathilde Caron, Andreas Steiner, Joan Puigcerver, Robert Geirhos, Ibrahim M Alabdulmohsin, et al. Patch n'pack: Navit, a vision transformer for any aspect ratio and resolution. Advances in Neural Information Processing Systems, 36, 2024. 2, 3", + "[9] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 1, 3, 7", + "[10] Yuming Fang, Hanwei Zhu, Yan Zeng, Kede Ma, and Zhou Wang. Perceptual quality assessment of smartphone photography. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3677-3686, 2020. 4", + "[11] Koustav Ghosal and Aljosa Smolic. Image aesthetics assessment using graph attention network. In 2022 26th International Conference on Pattern Recognition (ICPR), pages 3160-3167. IEEE, 2022. 2, 6" + ], + "bbox": [ + 93, + 114, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[12] Charles D Gilbert and Wu Li. Top-down influences on visual processing. Nature Reviews Neuroscience, 14(5):350-363, 2013. 2", + "[13] Jakob Drachmann Havtorn, Amélie Royer, Tijmen Blankevoort, and Babak Ehteshami Bejnordi. Msvit: Dynamic mixed-scale tokenization for vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 838-848, 2023. 2, 4, 6", + "[14] Shuai He, Yongchang Zhang, Rui Xie, Dongxiang Jiang, and Anlong Ming. Rethinking image aesthetics assessment: Models, datasets and benchmarks. In *IJCAI*, pages 942–948, 2022. 4", + "[15] Shuai He, Anlong Ming, Shuntian Zheng, Haobin Zhong, and Huadong Ma. Eat: An enhancer for aesthetics-oriented transformers. In Proceedings of the 31st ACM International Conference on Multimedia, pages 1023-1032, 2023. 6, 7", + "[16] Simon Hentschel, Konstantin Kobs, and Andreas Hotho. Clip knows image aesthetics. Frontiers in Artificial Intelligence, 5:976235, 2022. 1", + "[17] Shaul Hochstein and Merav Ahissar. View from the top: Hierarchies and reverse hierarchies in the visual system. Neuron, 36(5):791-804, 2002. 2", + "[18] Vlad Hosu, Bastian Goldlucke, and Dietmar Saupe. Effective aesthetics prediction with multi-level spatially pooled features. In proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9375-9383, 2019. 2", + "[19] Vlad Hosu, Hanhe Lin, Tamas Sziranyi, and Dietmar Saupe. Koniq-10k: An ecologically valid database for deep learning of blind image quality assessment. IEEE Transactions on Image Processing, 29:4041–4056, 2020. 4", + "[20] Yipo Huang, Leida Li, Pengfei Chen, Jinjian Wu, Yuzhe Yang, Yaqian Li, and Guangming Shi. Coarse-to-fine image aesthetics assessment with dynamic attribute selection. IEEE Transactions on Multimedia, 26:9316-9329, 2024. 6", + "[21] Junjie Ke, Qifei Wang, Yilin Wang, Peyman Milanfar, and Feng Yang. Musiq: Multi-scale image quality transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5148-5157, 2021. 1, 2, 3, 4, 6", + "[22] Junjie Ke, Keren Ye, Jiahui Yu, Yonghui Wu, Peyman Milanfar, and Feng Yang. Vila: Learning image aesthetics from user comments with vision-language pretraining. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10041-10051, 2023. 1", + "[23] Junjie Ke, Tianhao Zhang, Yilin Wang, Peyman Milanfar, and Feng Yang. Mret: Multi-resolution transformer for video quality assessment. Frontiers in Signal Processing, 3:1137006, 2023. 2", + "[24] Lei Ke, Mingqiao Ye, Martin Danelljan, Yu-Wing Tai, Chi-Keung Tang, Fisher Yu, et al. Segment anything in high quality. Advances in Neural Information Processing Systems, 36, 2024. 5", + "[25] Minchul Kim, Shangqian Gao, Yen-Chang Hsu, Yilin Shen, and Hongxia Jin. Token fusion: Bridging the gap between token pruning and token merging. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1383-1392, 2024. 2" + ], + "bbox": [ + 93, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "7823", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[26] Shu Kong, Xiaohui Shen, Zhe Lin, Radomir Mech, and Charless Fowlkes. Photo aesthetics ranking network with attributes and content adaptation. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part I 14, pages 662-679. Springer, 2016. 4", + "[27] Jin Li, Yaoming Wang, Xiaopeng Zhang, Bowen Shi, Dongsheng Jiang, Chenglin Li, Wenrui Dai, Hongkai Xiong, and Qi Tian. Ailurus: a scalable vit framework for dense prediction. Advances in Neural Information Processing Systems, 36, 2024. 2", + "[28] Leida Li, Yipo Huang, Jinjian Wu, Yuzhe Yang, Yaqian Li, Yandong Guo, and Guangming Shi. Theme-aware visual attribute reasoning for image aesthetics assessment. IEEE Transactions on Circuits and Systems for Video Technology, 33(9):4798-4811, 2023. 6", + "[29] Leida Li, Tong Zhu, Pengfei Chen, Yuzhe Yang, Yaqian Li, and Weisi Lin. Image aesthetics assessment with attribute-assisted multimodal memory network. IEEE Transactions on Circuits and Systems for Video Technology, 33(12):7413-7424, 2023. 6", + "[30] Leida Li, Xiangfei Sheng, Pengfei Chen, Jinjian Wu, and Weisheng Dong. Towards explainable image aesthetics assessment with attribute-oriented critiques generation. IEEE Transactions on Circuits and Systems for Video Technology, 2024. 6", + "[31] Limin Liu, Shuai He, Anlong Ming, Rui Xie, and Huadong Ma. Elta: An enhancer against long-tail for aesthetic-oriented models. In *Forty-first International Conference on Machine Learning*. 6", + "[32] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 6, 8", + "[33] Ilya Loshchilov and Frank Hutter. Fixing weight decay regularization in adam. 2018. 5", + "[34] Long Mai, Hailin Jin, and Feng Liu. Composition-preserving deep photo aesthetics assessment. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 497-506, 2016. 2", + "[35] Naila Murray, Luca Marchesotti, and Florent Perronnin. Ava: A large-scale database for aesthetic visual analysis. In 2012 IEEE Conference on Computer Vision and Pattern Recognition, pages 2408-2415, 2012. 4", + "[36] Tomer Ronen, Omer Levy, and Avram Golbert. Vision transformers with mixed-resolution tokenization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4613-4622, 2023. 2, 3, 4, 6", + "[37] Yossi Rubner, Carlo Tomasi, and Leonidas J Guibas. A metric for distributions with applications to image databases. In Sixth international conference on computer vision (IEEE Cat. No. 98CH36271), pages 59-66. IEEE, 1998. 4", + "[38] Peter Shaw, Jakob Uszkoreit, and Ashish Vaswani. Self-attention with relative position representations. arXiv preprint arXiv:1803.02155, 2018. 3", + "[39] Xiangfei Sheng, Leida Li, Pengfei Chen, Jinjian Wu, Weisheng Dong, Yuzhe Yang, Liwu Xu, Yaqian Li, and" + ], + "bbox": [ + 91, + 90, + 485, + 902 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Guangming Shi. Aesclip: Multi-attribute contrastive learning for image aesthetics assessment. In Proceedings of the 31st ACM International Conference on Multimedia, pages 1117-1126, 2023. 6", + "[40] Tengfei Shi, Chenglizhao Chen, Zhenyu Wu, Aimin Hao, and Yuming Fang. Improving image aesthetic assessment via multiple image joint learning. ACM Transactions on Multimedia Computing, Communications and Applications, 2024. 6", + "[41] Nyeong-Ho Shin, Seon-Ho Lee, and Chang-Su Kim. Blind image quality assessment based on geometric order learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12799-12808, 2024. 5, 6", + "[42] Hossein Talebi and Peyman Milanfar. Learning to resize images for computer vision tasks. In Proceedings of the IEEE/CVF International Conference on Computer vision, pages 497-506, 2021. 2", + "[43] Zhengzhong Tu, Peyman Milanfar, and Hossein Talebi. Muller: Multilayer laplacian resizer for vision. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6877-6887, 2023. 6", + "[44] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 3", + "[45] Jingdong Wang, Ke Sun, Tianheng Cheng, Borui Jiang, Chaorui Deng, Yang Zhao, Dong Liu, Yadong Mu, Mingkui Tan, Xinggang Wang, et al. Deep high-resolution representation learning for visual recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(10):3349-3364, 2020. 2", + "[46] Jian Wang, Xiang Long, Guowei Chen, Zewu Wu, Zeyu Chen, and Errui Ding. U-hrnet: delving into improving semantic representation of high resolution network for dense prediction. arXiv preprint arXiv:2210.07140, 2022. 2", + "[47] Oliver Wiedemann, Vlad Hosu, Shaolin Su, and Dietmar Saupe. Konx: cross-resolution image quality assessment. Quality and User Experience, 8(1):8, 2023. 2", + "[48] Yuzhe Yang, Liwu Xu, Leida Li, Nan Qie, Yaqian Li, Peng Zhang, and Yandong Guo. Personalized image aesthetics assessment with rich attributes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19861-19869, 2022. 4", + "[49] Ran Yi, Haoyuan Tian, Zhihao Gu, Yu-Kun Lai, and Paul L Rosin. Towards artistic image aesthetics assessment: a large-scale dataset and a new method. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22388-22397, 2023. 4", + "[50] Tong Zhu, Leida Li, Pengfei Chen, Jinjian Wu, Yuzhe Yang, and Yaqian Li. Emotion-aware hierarchical interaction network for multimodal image aesthetics assessment. Pattern Recognition, 154:110584, 2024. 6" + ], + "bbox": [ + 516, + 90, + 905, + 839 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "7824", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Charm_ The Missing Piece in ViT Fine-Tuning for Image Aesthetic Assessment/0ec61778-fb6e-453d-bfdb-3d054143034b_model.json b/2025/Charm_ The Missing Piece in ViT Fine-Tuning for Image Aesthetic Assessment/0ec61778-fb6e-453d-bfdb-3d054143034b_model.json new file mode 100644 index 0000000000000000000000000000000000000000..22e5a86b6c246667d8d371f378f676f29360862f --- /dev/null +++ b/2025/Charm_ The Missing Piece in ViT Fine-Tuning for Image Aesthetic Assessment/0ec61778-fb6e-453d-bfdb-3d054143034b_model.json @@ -0,0 +1,2024 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.109, + 0.131, + 0.89, + 0.152 + ], + "angle": 0, + "content": "Charm: The Missing Piece in ViT fine-tuning for Image Aesthetic Assessment" + }, + { + "type": "title", + "bbox": [ + 0.288, + 0.182, + 0.71, + 0.218 + ], + "angle": 0, + "content": "Fatemeh Behrad, Tinne Tuytelaars, Johan Wagemans \nKU Leuven University, Belgium" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.22, + 0.77, + 0.234 + ], + "angle": 0, + "content": "{fatemeh.behrad, tinne.tuytelaars, johan.wagemans}@kuleuven.be" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.269, + 0.327, + 0.284 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.301, + 0.485, + 0.664 + ], + "angle": 0, + "content": "The capacity of Vision transformers (ViTs) to handle variable-sized inputs is often constrained by computational complexity and batch processing limitations. Consequently, ViTs are typically trained on small, fixed-size images obtained through downscaling or cropping. While reducing computational burden, these methods result in significant information loss, negatively affecting tasks like image aesthetic assessment. We introduce Charm, a novel tokenization approach that preserves Composition, High-resolution, Aspect Ratio, and Multi-scale information simultaneously. Charm prioritizes high-resolution details in specific regions while downscaling others, enabling shorter fixed-size input sequences for ViTs while incorporating essential information. Charm is designed to be compatible with pre-trained ViTs and their learned positional embeddings. By providing multiscale input and introducing variety to input tokens, Charm improves ViT performance and generalizability for image aesthetic assessment. We avoid cropping or changing the aspect ratio to further preserve information. Extensive experiments demonstrate significant performance improvements on various image aesthetic and quality assessment datasets (up to \\(8.1\\%\\) ) using a lightweight ViT backbone. Code and pre-trained models are available at https://github.com/FBehrad/Charm." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.693, + 0.222, + 0.709 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.719, + 0.483, + 0.81 + ], + "angle": 0, + "content": "Humans are inherently drawn to aesthetically pleasing images, making image aesthetics a vital consideration in fields like photography, advertising, e-commerce, and social media. Image aesthetic assessment (IAA) aims to develop computational models that can accurately predict human aesthetic preferences for images." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Deep learning approaches applied to IAA have evolved significantly, with the Vision Transformer (ViT) architecture [9] demonstrating promising results [16, 21, 22]. ViTs process images as sequences of non-overlapping patches, allowing them to handle images of various sizes. However, in practice, this flexibility is limited for several rea" + }, + { + "type": "image", + "bbox": [ + 0.525, + 0.268, + 0.907, + 0.574 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.583, + 0.907, + 0.653 + ], + "angle": 0, + "content": "Figure 1. Traditional image preprocessing methods (right) often distort image aesthetics by rescaling and cropping images to a fixed size. To preserve critical aesthetic information, our approach (left) selectively maintains some regions in their original resolution while downscaling the others." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.686, + 0.907, + 0.821 + ], + "angle": 0, + "content": "sons. First, position embeddings of pre-trained ViTs are learned for a specific, fixed input size. Additionally, batch processing requires all sequences in a batch to be the same length, which is challenging with images of different dimensions. Grouping images by size or padding can help. However, grouping images may affect the learning process and convergence, as it makes batches less representative of the entire distribution, while padding can lead to inefficient resource utilization." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.909, + 0.903 + ], + "angle": 0, + "content": "This brings us to another major issue when processing high-resolution images, which is the excessive computation and memory requirements. This can be particularly challenging when dealing with limited computational resources. As a result, a standard approach is to resize and" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7815" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.228 + ], + "angle": 0, + "content": "downscale images to a smaller, fixed size to manage these constraints more effectively. This approach works for tasks where downscaling or cropping does not affect labels. However, in IAA, these operations can change aesthetic scores by altering the image composition and aspect ratio and losing high-resolution details (Figure 1). To unlock the full potential of ViTs in IAA, a new approach is needed to preserve critical aesthetic information while remaining computationally practical." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.231, + 0.482, + 0.397 + ], + "angle": 0, + "content": "Several techniques, such as graph attention networks [11], adaptive fractional dilated convolution networks [6], and multi-scale ViT [21], have been proposed to preserve image aspect ratio for IAA. Others, such as native resolution ViT [8], extend beyond IAA, demonstrating the benefits of preserving aspect ratio in other tasks. However, to remain practical, existing approaches still require downscaling the entire image, which limits their effectiveness. Fixed-size inputs can also limit the generalization of IAA models across various image sizes, which can be addressed by multi-scale inputs [18, 21, 23, 47]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.4, + 0.483, + 0.686 + ], + "angle": 0, + "content": "One possible solution to preserve high-resolution information is a learnable resizer that reduces image dimensions while minimizing information loss during downsampling [7, 42]. While effective for other tasks, these methods can introduce distortions that alter the aesthetic quality of images. Another approach is to prioritize high-resolution information in some regions while downscaling less critical areas. This solution is inspired by the human brain, which processes visual information efficiently by preserving both low- and high-resolution details across the cortical hierarchy, maintaining spatial relationships, and utilizing selective attention and eye movements [12, 17]. Mixed-resolution [36] and mixed-scale tokenization [13] are steps in this direction. However, these methods produce identical representations for different resolutions of the same image because they first downscale all images to a fixed, low resolution for batch processing. While this approach is sufficient for image classification, it is inadequate for IAA, where resolution influences the aesthetic score." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.69, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Our contribution. Existing approaches often prioritize a single factor crucial for aesthetics, neglecting others. Preserving composition, high resolution, aspect ratio, and multi-scale information simultaneously remains a challenge. We introduce Charm to fill this gap. Charm avoids fully downscaling images and maintains some areas in their original resolution (Figure 1). Our goal is to create a balance between computational efficiency and network performance. Charm effectively captures aspect ratio, composition, and multiscale information. Our framework-agnostic tokenization strategy avoids extensive pre-training by not modifying the pre-trained ViT architecture or its learned position embeddings. To demonstrate the broader applicability of our approach, we evaluate it on both IAA and image" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.725, + 0.107 + ], + "angle": 0, + "content": "quality assessment (IQA) tasks." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.12, + 0.655, + 0.136 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.145, + 0.905, + 0.418 + ], + "angle": 0, + "content": "Image aesthetic assessment Existing deep learning-based IAA methods often require images to be resized or cropped to a relatively small fixed size, leading to loss of aspect ratio, composition, and high-resolution detail. Using a batch size of 1 avoids the need for image downscaling or cropping [34]. However, this can be computationally inefficient for large-scale training. Graph-based representations, such as those proposed by Ghosal et al. [11], can be memory-intensive for large images. Similarly, adaptive fractional dilated convolutions, introduced by Chen et al. [6], can be computationally expensive and require careful batch processing. A more recent strategy for preserving high resolution and aspect ratio is the use of learnable resizers. These methods employ auxiliary networks to reduce the dimensionality of input images while minimizing information loss during downsampling [7, 42, 45, 46]. While effective for image classification, these approaches introduce distortions that negatively impact image aesthetics." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.418, + 0.905, + 0.493 + ], + "angle": 0, + "content": "Another challenge associated with fixed-size inputs is a decrease in model generalization across various image sizes. While this challenge can be mitigated by supplying multi-scale inputs to the network [18, 21, 23, 47], existing approaches introduce non-negligible computational costs." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.494, + 0.905, + 0.614 + ], + "angle": 0, + "content": "Vision transformers Among existing neural networks, ViTs offer the advantage of processing images in their original size. However, they are still constrained by the need for fixed-size inputs due to batch processing. NaViT [8] addresses this problem by packing patches from different images into a fixed-size sequence. However, this approach can be computationally expensive as it can result in excessively long input sequences for high-resolution images." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.615, + 0.905, + 0.719 + ], + "angle": 0, + "content": "The computational costs of ViTs are influenced by the number of input tokens they process, so downscaling images can reduce the number of tokens and improve efficiency. Token dropping, either randomized or structured [27], and token merging [3] are techniques to improve ViT efficiency. Some methods, like token fusion [25], combine these approaches to achieve even better results." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.72, + 0.905, + 0.856 + ], + "angle": 0, + "content": "Previous studies have shown that incorporating multiscale features can enhance ViT performance. MUSIQ [21], for instance, utilizes patches from different scales to capture multi-scale information. While effective, this approach modifies the first layers of the ViT and introduces new position embeddings, which necessitates extensive training for each ViT backbone. FlexiViT [2], a more recent approach, extracts multiscale features using random patch sizes in each training epoch." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.856, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Closer to our approach are mixed-resolution [36] and mixed-scale tokenization [13]. These methods first downscale images to a fixed size, then extract two-scale patches," + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7816" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.242 + ], + "angle": 0, + "content": "preserving key regions in higher resolution while downscaling others. However, this initial downscaling discards high-frequency details and limits patches to two fixed resolutions. In contrast, Charm retains high-frequency information from the original image, learning from patches at varying resolutions (see Appendix 6 for details). Mixed-resolution ViT [36] further limits the input format, requiring images to be square. Additionally, it replaces learned position embeddings with sinusoidal ones, necessitating additional pre-training of each ViT backbone for compatibility." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.243, + 0.482, + 0.348 + ], + "angle": 0, + "content": "Existing methods fail to simultaneously preserve multiscale information, aspect ratio, composition, and high-resolution details. While all of these factors are crucial for accurate IAA, current methods typically prioritize computational efficiency at the expense of preserving these critical aspects. In contrast, our method successfully preserves all these factors while remaining computationally practical." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.348, + 0.483, + 0.544 + ], + "angle": 0, + "content": "Position embedding Transformers lack inherent mechanisms to capture the order of elements within an input sequence [44]. This stems from their reliance on self-attention, which is permutation-invariant – meaning the order of elements in the input does not affect the final output [1]. As a result, they require position embeddings to encode the sequence order. These position embeddings can be a deterministic function of position, typically using sinusoidal functions [44] or learned representations. Learnable position embeddings have achieved comparable performance to deterministic approaches [9]. However, both deterministic and learnable position embeddings are pre-trained for a specific sequence length." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.545, + 0.483, + 0.755 + ], + "angle": 0, + "content": "Relative position embeddings offer a more flexible approach by capturing the relative distance between elements [38]. This approach is well-suited for variable-length inputs. However, it necessitates significant modifications to the transformer attention mechanism. Recent methods like hash-based spatial embedding [21] and factorized position embedding [8] support variable aspect ratios and input resolutions without modifying attention modules. However, they cannot replace the position embedding of a pre-trained transformer without extensive retraining of the backbone on large-scale datasets. In contrast, our method adapts the position embedding of a pre-trained ViT model, requiring no additional pre-training and enabling efficient fine-tuning on downstream tasks with limited data." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.77, + 0.421, + 0.787 + ], + "angle": 0, + "content": "3. Unlock the potentials of ViTs in IAA" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.482, + 0.856 + ], + "angle": 0, + "content": "In this section, we introduce Charm, a novel tokenization approach designed to enhance the performance of standard ViTs in image aesthetic and quality assessment tasks. Figure 2 provides an overview of our proposed approach." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.483, + 0.901 + ], + "angle": 0, + "content": "We initially partition the input image into fixed-size patches, preserving a selection of patches at their original resolution while downscaling others. This approach enables" + }, + { + "type": "image", + "bbox": [ + 0.557, + 0.089, + 0.868, + 0.365 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.376, + 0.907, + 0.459 + ], + "angle": 0, + "content": "Figure 2. Our approach involves partitioning the input image into fixed-size patches and selectively preserving a subset at its original resolution while downscaling others. Position and scale embeddings are then prepared to encode spatial relationships across different scales. The remaining steps follow the standard ViT. + indicates element-wise addition." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.477, + 0.905, + 0.582 + ], + "angle": 0, + "content": "the model to capture both fine-grained details and broader contextual information. Subsequently, position embeddings are prepared to encode the spatial relationships between patches across different resolutions and within the same resolution. Following previous studies [21], we also incorporate scale embeddings to aid the model in distinguishing information from different scales." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.585, + 0.906, + 0.718 + ], + "angle": 0, + "content": "After encoding the input patches into a sequence of tokens using the patch encoding module, we add position and scale embeddings. A learnable \"classification token\" (CLS) is prepended, and all tokens are processed through the transformer encoder. The final output is obtained by feeding the CLS token state at the output of the transformer through a fully connected layer. Our approach is compatible with various ViT architectures, as we do not modify the core ViT network structure." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.735, + 0.655, + 0.75 + ], + "angle": 0, + "content": "3.1. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.758, + 0.907, + 0.901 + ], + "angle": 0, + "content": "An image \\(x \\in \\mathbb{R}^{h \\times w \\times c}\\), where \\((h, w, c)\\) represents the height, width, and number of channels, respectively, undergoes tokenization in a standard ViT. This process divides the image into a sequence of \\(s\\) patches \\(x_i \\in \\mathbb{R}^{p \\times p \\times c}\\), where \\(i \\in (1, \\ldots, s)\\) and \\(p\\) represents the patch size. The sequence length \\(s = \\left\\lfloor \\frac{h}{p} \\right\\rfloor \\cdot \\left\\lfloor \\frac{w}{p} \\right\\rfloor\\) determines the number of patches after tokenization, controlling the ViT's computational complexity. Given that the patch size \\(p\\) remains fixed for a given ViT, the number of patches depends on the in-" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7817" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.123 + ], + "angle": 0, + "content": "put image resolution. Consequently, high-resolution images increase computational complexity." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.13, + 0.283, + 0.145 + ], + "angle": 0, + "content": "3.2. Charm tokenization" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.153, + 0.228, + 0.166 + ], + "angle": 0, + "content": "3.2.1. Tokenization" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.171, + 0.485, + 0.351 + ], + "angle": 0, + "content": "Batch processing necessitates a consistent spatial resolution within each minibatch, leading to a fixed input length \\((l)\\) for the ViT. Processing high-resolution images requires a large \\(l\\), which increases memory consumption and computational demands during training. Our method addresses this by strategically reducing \\(l\\) while preserving aesthetic-related information (Figure 2). Given a fixed patch size \\((p)\\), we can consider two scenarios. For low-resolution images \\((s \\leq l)\\), we follow the standard ViT approach, tokenizing the entire image with the patch size of \\(p\\). For high-resolution images \\((s > l)\\), we keep some parts in high resolution while downscaling other areas." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.353, + 0.485, + 0.564 + ], + "angle": 0, + "content": "When dealing with 2 scales, we initially tokenize the image using a patch size of \\( p' = p \\times n \\), where \\( n \\geq 2 \\). Subsequently, we select high-resolution patches and further tokenize them using a patch size of \\( p \\) while the remaining patches are downscaled to \\( p \\). The downscaling factor for both height and width is \\( f = 1/n \\). Finally, we arrange all high-resolution patches at the beginning, followed by low-resolution ones to form the final input for the transformer encoder. This ordering is feasible due to the permutation invariance of ViTs [1], allowing us to avoid unnecessary implementation loops. To ensure a consistent input length (\\( l \\)), padding or random token dropping is employed as needed. Additional details on the 3-scale version are available in the Appendix 7." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.57, + 0.264, + 0.586 + ], + "angle": 0, + "content": "3.2.2. Patch Importance" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.589, + 0.485, + 0.71 + ], + "angle": 0, + "content": "Several strategies can be employed to determine which patches to retain at high resolution. The main goal is to preserve visually interesting areas and sharp details in high resolution. To identify these regions, we use frequency, gradient, entropy, and saliency maps. We compare them with a random selection of patches in each training epoch. More details on patch selection strategies, including illustrative examples, can be found in Figure 8 of Appendix 8." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.716, + 0.279, + 0.731 + ], + "angle": 0, + "content": "3.2.3. Position Embedding" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.485, + 0.9 + ], + "angle": 0, + "content": "ViTs lack inherent knowledge of spatial relationships between patches. Position embeddings address this by injecting learned position information into each patch embedding, allowing the self-attention mechanism to capture how different parts of the image relate. Replacing the standard position embeddings of ViTs with custom ones [36] can be expensive, requiring extensive pre-training on large datasets. It also introduces compatibility issues when used with different pre-trained ViTs. To address these issues, our approach adapts the pre-trained position embeddings from the standard ViT models." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.303 + ], + "angle": 0, + "content": "For low-resolution images, we follow the standard ViT approach of resizing pre-trained position embeddings via bilinear interpolation. However, for high-resolution images, we create two position embeddings using bilinear interpolation based on the different patch sizes \\((p\\) and \\(p^{\\prime})\\) employed during tokenization (Figure 2). One position embedding retains the spatial relationships between low-resolution patches, while the other captures the finer spatial relationships within the high-resolution patches. These position embeddings, high and low resolution, are compatible because they are derived from the same pre-trained embedding. Similar to standard ViTs, each token from the output of the patch encoding module is then summed with its corresponding position embedding." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.304, + 0.905, + 0.41 + ], + "angle": 0, + "content": "Unlike existing methods [13, 36], our approach does not have a fixed image size, so position embeddings vary for each image. To avoid recalculating position embeddings in each epoch, we prepare them in the data loaders, keeping them frozen during fine-tuning. This aligns with the standard practice of freezing the initial layers (including position embeddings) and fine-tuning only the final layers." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.42, + 0.682, + 0.435 + ], + "angle": 0, + "content": "3.2.4. Scale Embedding" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.439, + 0.905, + 0.531 + ], + "angle": 0, + "content": "Our tokenization approach results in multiscale tokens. Following previous studies [21], to help ViT distinguish between patches from different scales and capture their relationships, we incorporate learnable scale embeddings. These scale embeddings are randomly initialized and added element-wise to all patch embeddings (Figure 2)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.542, + 0.642, + 0.558 + ], + "angle": 0, + "content": "3.3. Fine-tuning" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.565, + 0.907, + 0.791 + ], + "angle": 0, + "content": "Our method requires no additional pre-training. We process images in their original resolution without resizing or random cropping during fine-tuning. We avoid normalizing the images as normalization can introduce distortions that affect the overall aesthetic/quality score (Table 9 in Appendix 9). Our data augmentation is limited to random horizontal flipping and random rotations (90, 180, 270 degrees) with a \\(50\\%\\) probability, as they do not result in loss of composition or aspect ratio. While these methods might change the aesthetics for humans, they have been shown to improve the performance of ViTs (Appendix 10). During evaluation, data augmentation is completely omitted. We employ the L1 loss function for predicting the overall aesthetic/quality score and the Earth Mover's Distance (EMD) loss [37] for predicting the aesthetic score distribution." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.807, + 0.646, + 0.824 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.833, + 0.616, + 0.848 + ], + "angle": 0, + "content": "4.1. Datasets" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.905, + 0.903 + ], + "angle": 0, + "content": "We evaluate our models on various IAA datasets, including AVA [35], PARA [48], TAD66k [14], AADB [26], BAID [49], and two IQA datasets, SPAQ [10] and KonIQ10k [19]." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7818" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.088, + 0.482, + 0.208 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.218, + 0.483, + 0.247 + ], + "angle": 0, + "content": "Figure 3. An image with its aesthetic score distribution and mean aesthetic score, from the AVA dataset." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.261, + 0.483, + 0.382 + ], + "angle": 0, + "content": "We only use the images and their corresponding ratings, excluding any additional information provided in the datasets. The AVA and PARA datasets associate each image with an aesthetic score distribution, while other datasets associate each image with an overall aesthetic/quality score (Figure 3). As the SPAQ and KonIQ10k datasets lack an official train-test split, we repeat the experiments for 5 different splits and report the median performance (as done in [41])." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.383, + 0.483, + 0.519 + ], + "angle": 0, + "content": "The AVA dataset remains prominent due to its large number of images, allowing for effective training of large models. Research utilizing smaller datasets is relatively limited, and most existing approaches leverage the additional information provided in these datasets. However, considering various datasets with different resolution ranges allows us to validate the efficiency of our approach in various scenarios. For more details on the resolution distribution of each dataset, see Figure 13 in Appendix 11." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.53, + 0.307, + 0.545 + ], + "angle": 0, + "content": "4.2. Implementation details" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.552, + 0.483, + 0.628 + ], + "angle": 0, + "content": "Saliency maps for patch selection are generated using the SAM-HQ model [24] with the text prompt of \"salient object\" (Appendix 8). For gradient, frequency, and entropy calculations, we use the Sobel filter, fast Fourier transform, and Shannon entropy, respectively." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.629, + 0.483, + 0.78 + ], + "angle": 0, + "content": "In all experiments, we fine-tune the entire ViT models, while keeping their position embeddings frozen. Also, we use scale embeddings alongside position embeddings to enable Charm to capture relationships between tokens both within and across scales (see Appendix 12). The patch size \\((p)\\) and patch stride are chosen to match the filter size and stride of the patch encoding modules within these models (16 in ViT-small and 14 in Dinov2-small and Dinov2-large), ensuring compatibility and replicating the models' internal tokenization process." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.483, + 0.855 + ], + "angle": 0, + "content": "The maximum number of tokens \\((l)\\) is set to 512 for AVA, 768 for TAD66k, and 1024 for other datasets. These values are chosen based on the image resolution in the datasets (Appendix 13). For all experiments, we use a factor size of 0.5." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "The AdamW optimizer [33] is employed for fine-tuning. We use an initial learning rate of 1e-5 for all datasets except BAID (1e-6). We gradually reduce the learning rate dur" + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.088, + 0.905, + 0.43 + ], + "angle": 0, + "content": "
DatasetCharmPLCCSRCCACC
IAA
AVA-0.7100.7060.802
✓0.779(↑ 6.9%)0.777(↑ 7.1%)0.826(↑ 2.4%)
AADB-0.6950.6820.754
✓0.767(↑ 7.2%)0.754(↑ 7.2%)0.767(↑ 1.3%)
TAD66k-0.4290.4010.646
✓0.488(↑ 5.9%)0.458(↑ 5.7%)0.794 (↑ 14.8%)
PARA-0.9040.8550.863
✓0.938(↑ 3.4%)0.905(↑ 5%)0.892(↑ 2.9%)
BAID-0.4280.3420.750
✓0.439(↑ 1.1%)0.368(↑ 2.6%)0.763(↑ 1.3%)
IQA
SPAQ-0.9110.9070.907
✓0.919(↑ 0.8%)0.915(↑ 0.8%)0.917(↑ 1.0%)
KonIQ10k-0.8960.8680.938
✓0.944(↑ 4.8%)0.93(↑ 6.2%)0.954(↑ 1.6%)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.44, + 0.907, + 0.483 + ], + "angle": 0, + "content": "Table 1. Performance improvement across different IAA and IQA datasets by replacing the standard tokenization of Dinov2-small with Charm (with a random patch selection strategy)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.5, + 0.907, + 0.622 + ], + "angle": 0, + "content": "ing training using the cosine annealing learning rate scheduler. The batch size is 64 for AVA, 32 for TAD66k, and 16 for the remaining datasets. Early stopping is also implemented to prevent overfitting. Fine-tuning is conducted on a single NVIDIA RTX A4500 GPU. Model performance is evaluated using three key metrics: Spearman rank-ordered correlation coefficient (SRCC), Pearson linear correlation coefficient (PLCC), and classification accuracy (ACC)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.635, + 0.81, + 0.65 + ], + "angle": 0, + "content": "4.3. Comparing with existing methods" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.658, + 0.907, + 0.779 + ], + "angle": 0, + "content": "We benchmark Charm with various ViT backbones on different IAA and IQA datasets. We first fine-tune models on the datasets to create a baseline. We then replace their original patch encoding modules with two-scale Charm tokenization and fine-tune the entire models. We also considered a three-scale version of Charm but did not observe additional performance gains (Appendix 7). Therefore, we focus on the two-scale version throughout the paper." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.909, + 0.902 + ], + "angle": 0, + "content": "The results (Table 1 and Table 13 in Appendix 14) demonstrate that by simply replacing the patch encoding module of existing ViT models with Charm, we can achieve significant performance improvements, up to \\(7.5\\%\\) in PLCC, \\(8.1\\%\\) in SRCC, and \\(14.8\\%\\) in classification accuracy across various datasets. Among datasets, PARA and SPAQ datasets include extra high-resolution images (e.g., 2k by 3k pixels). Our experiments (Appendix 15) reveal" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7819" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.251, + 0.088, + 0.747, + 0.246 + ], + "angle": 0, + "content": "
ModelARHRMSPLCCSRCCACC
Ghosal et al.[11]✓--0.7640.762-
Chen et al. [6]✓--0.6710.6490.832
Dinov2-Small (+ Padding) *✓✓-0.7090.7030.801
MUSIQ [21]✓-✓0.7380.7260.815
FlexiViT [2]*--✓0.7370.7350.812
Swin [32]*--✓0.7480.7510.816
Dinov2-Small (+ Muller [43]) *-✓-0.6820.6750.794
ViT-small (+ Charm) *✓✓✓0.7620.7600.827
Dinov2-small (+ Charm) *✓✓✓0.7790.7770.826
Dinov2-large (+ Charm) *✓✓✓0.7830.7810.828
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.256, + 0.907, + 0.3 + ], + "angle": 0, + "content": "Table 2. Results on the AVA dataset. Bold and underlined numbers indicate the best and the second-best results, respectively. AR, HR, and MS represent preserving aspect ratio, high-resolution details, and multi-scale information, respectively. * indicates models that we have fine-tuned on the AVA dataset. Appendix 14 shows the performance of the standard ViT-small, Dinov2-small, and Dinov2-large." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.325, + 0.482, + 0.401 + ], + "angle": 0, + "content": "a threshold for performance improvement by preserving high-resolution information. Beyond this threshold, further performance gains are limited. Therefore, to decrease computational costs, we downscale images in PARA and SPAQ to a maximum edge of 1024." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.401, + 0.483, + 0.568 + ], + "angle": 0, + "content": "As shown in Table 2, our approach outperforms existing methods in terms of preserving resolution, aspect ratio, and multiscale information. We compare Charm with the simple approach of adding padding to ensure consistent input shapes within each batch while preserving the aspect ratio. However, this padding proves detrimental to IAA and significantly increases computational costs (Appendix 16). Also, we observe that Muller [43], a recent approach focusing on preserving high-resolution information, can introduce noise and distortions to the image, negatively affecting aesthetics (Appendix 16)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.568, + 0.483, + 0.81 + ], + "angle": 0, + "content": "Recent studies have demonstrated the effectiveness of incorporating additional modalities, such as attributes ([5, 20, 28]) or text captions ([29, 30, 39, 50]), to improve IAA model performance (Table 15 in Appendix 17). Other studies have concentrated on other research directions [15, 40], like addressing the long tails of distributions in IAA datasets [31]. While the main contributions of these studies are orthogonal to Charm's, integrating Charm with them further enhances their performance. For example, incorporating Charm into ELTA [31] with a Dinov2-large backbone improves its PLCC, SRCC, and ACC on the AVA dataset from 0.742, 0.742, and 0.811 to 0.787, 0.786, and 0.829, respectively. Also, our approach demonstrates strong performance in IQA, with a slight performance gap of less than \\(1\\%\\) in SPAQ and \\(0.4\\%\\) in KonIQ10k compared to state-of-the-art methods [4, 21, 41] (Table 16 in Appendix 17)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Model size is another crucial factor, as larger models often excel in complex tasks like IAA. In this paper, we demonstrate that Charm is highly effective even with lightweight ViTs, which can be used for applications with limited computational resources. As shown in Figure 4, Dinov2-small + Charm achieves comparable performance" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.325, + 0.907, + 0.431 + ], + "angle": 0, + "content": "to state-of-the-art models while being significantly smaller. Nevertheless, Charm can serve as a plug-in for larger, more expensive models if higher performance is desired and computational power allows. For instance, plugging Charm into ViT-base instead of Dinov2-small results in PLCC of 0.515, SRCC of 0.467, and ACC of 0.779 after only 9 epochs on the BAID dataset." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.446, + 0.665, + 0.463 + ], + "angle": 0, + "content": "4.4. Ablation study" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.469, + 0.906, + 0.501 + ], + "angle": 0, + "content": "4.4.1. The importance of preserving aspect ratio and high-resolution details" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.505, + 0.906, + 0.656 + ], + "angle": 0, + "content": "Current approaches usually distort images by resizing them to a square shape, disregarding the significance of maintaining their original size and aspect ratio. We argue that such preprocessing can be detrimental to IAA. To validate this claim, we compare the performance of Dinov2-small using different tokenization approaches. Table 3 presents the results of these experiments. Except for the standard tokenization (described in Appendix 9), all other settings involve multiscale inputs and maintain composition by avoiding image cropping." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.657, + 0.906, + 0.779 + ], + "angle": 0, + "content": "We initially downscale images to \\(256 \\times 256\\) and then preserve some areas at that resolution while downscaling others by a factor of 0.5 (MS in Table 3). While this approach has been effective in image classification ([13] and [36]), it decreases performance in our IAA task. This is because the initial downscaling to a fixed size results in significant information loss critical for IAA, and further downscaling worsens the issue." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.909, + 0.903 + ], + "angle": 0, + "content": "We then experiment with keeping the image in its original resolution while changing the aspect ratio by resizing the image based on the maximum edge to create a square image (HR + MS in Table 3). This results in a nearly \\(3\\%\\) increase in PLCC and SRCC compared to the standard approach. Finally, we use Charm, which preserves both aspect ratio and high-resolution details, resulting in the best performance. This analysis highlights the importance of preserv" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.956 + ], + "angle": 0, + "content": "7820" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.089, + 0.482, + 0.277 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.289, + 0.485, + 0.401 + ], + "angle": 0, + "content": "Figure 4. Comparison of state-of-the-art IAA models' performance on the AVA dataset vs. their number of parameters. Our approach achieves comparable performance without using additional modalities and with significantly fewer parameters. To account for the significant differences in the number of parameters, we use a logarithmic scale on the x-axis. For a detailed comparison of Dinov2-small + Charm with state-of-the-art models on all datasets, see Appendix 17." + }, + { + "type": "table", + "bbox": [ + 0.12, + 0.414, + 0.456, + 0.489 + ], + "angle": 0, + "content": "
TokenizationPLCCSRCCACC
Standard0.7340.7320.808
MS0.7100.7050.801
HR + MS0.7630.7600.820
Charm (AR + HR + MS)0.7790.7770.826
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.498, + 0.483, + 0.555 + ], + "angle": 0, + "content": "Table 3. Performance of Dinov2-small on the AVA dataset with four different tokenization strategies. MS, HR, and AR represent preserving multiscale information, high-resolution details, and aspect ratio, respectively." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.571, + 0.471, + 0.586 + ], + "angle": 0, + "content": "ing both aspect ratio and original resolution in IAA tasks." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.586, + 0.483, + 0.632 + ], + "angle": 0, + "content": "We also examined the impact of different numbers of tokens across tokenization approaches in Table 3 but found no differences in the results (see Appendix 18)." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.64, + 0.418, + 0.655 + ], + "angle": 0, + "content": "4.4.2. Comparison of patch selection strategies" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.659, + 0.483, + 0.749 + ], + "angle": 0, + "content": "To evaluate the importance of patches, we fine-tune Dinov2-small on the AADB dataset with different patch selection strategies. Table 4 demonstrates that all strategies outperform the standard ViT approach [9]. This suggests that factors beyond preserving high-resolution details also contribute to performance improvements (Section 4.4.1)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.75, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Among all methods, the random selection of patches in each epoch stands out, likely for two reasons. First, ViTs are prone to overfitting, especially when fine-tuned on small datasets, a common challenge in IAA. Random patch selection in each epoch addresses this by introducing more variety to the ViT input and acting as a form of data augmentation. Second, random sampling can better capture the subjective and often unpredictable nature of beauty, providing a more representative view of the entire image. Following random selection, frequency, gradient, and entropy-based" + }, + { + "type": "table", + "bbox": [ + 0.572, + 0.088, + 0.851, + 0.19 + ], + "angle": 0, + "content": "
Patch selectionPLCCSRCCACC
Random0.7670.7540.767
Entropy0.7260.7140.761
Frequency0.7560.7470.761
Gradient0.7340.7210.766
Saliency0.7510.7380.756
Standard0.6950.6820.754
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.2, + 0.906, + 0.243 + ], + "angle": 0, + "content": "Table 4. Performance of Dinov2-small on the AADB dataset using Charm with different patch selection strategies. Bold and underlined values highlight the best and second-best results." + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.257, + 0.903, + 0.386 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.394, + 0.906, + 0.452 + ], + "angle": 0, + "content": "Figure 5. The epoch at which each model achieves its highest validation performance across different datasets. Charm generally leads to faster convergence. The average number of epochs over 5 runs is reported for SPAQ and KonIQ10k." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.466, + 0.906, + 0.511 + ], + "angle": 0, + "content": "methods stand out. These methods still have a certain level of randomness, as their fully deterministic version results in highly correlated selected patches (Appendix 8)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.511, + 0.906, + 0.603 + ], + "angle": 0, + "content": "ViTs, originally trained for object classification, have an inherent bias towards salient objects. However, IAA requires attention to both foreground and background [15], which explains why the saliency-based approach performs less effectively. For a more comprehensive analysis of patch selection strategies please refer to Appendix 8." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.612, + 0.855, + 0.627 + ], + "angle": 0, + "content": "4.4.3. The effect of Charm on convergence speed" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.631, + 0.906, + 0.797 + ], + "angle": 0, + "content": "We fine-tune all models until convergence, meaning their performance on the validation set stabilizes. Figure 5 shows the epoch at which each model achieves its best validation performance. As shown in this figure, Charm consistently accelerates the convergence of Dinov2-small across different datasets. Notably, the most significant reduction (50%) in training epochs occurs on AADB, our smallest dataset, likely due to Charm's ability to mitigate overfitting through random patch selection (see Appendix 8). On our largest datasets, AVA and TaD66k, both standard and Charm tokenization converge at similar epochs." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.806, + 0.704, + 0.821 + ], + "angle": 0, + "content": "4.4.4. Computational costs" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Figure 6 compares the runtime of Charm using different patch selection strategies with the standard approach on a 12th Gen Intel(R) Core(TM) i9-12900K CPU. Table 3 demonstrates that multiscale inputs can negatively impact performance for very small images. In this case, we use a" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "7821" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.185, + 0.088, + 0.813, + 0.162 + ], + "angle": 0, + "content": "
ModelInput sizeCharm#tokensmsGMACsMB
Dinov2224 x 224-2565.76.11202.9
-small-207032.884.012091.8
640 x 640✓2-scale:5127.3 (↓ 77.7%)13.46 (↓ 84%)346.0 (↓ 83.5%)
✓3-scale:7009.3 (↓ 71.6%)19.60 (↓ 76.7%)494.3 (↓ 76.4%)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.173, + 0.907, + 0.23 + ], + "angle": 0, + "content": "Table 5. Dinov2-small inference cost breakdown for processing one single image: number of tokens (#tokens) based on varying input sizes, runtime in milliseconds (ms), Giga multiply accumulation (GMACs), and GPU memory in Megabytes (MB). Bold and underlined values highlight the most and second-most computationally efficient configurations. Percentages indicate the reduction in computational cost compared to processing the image in its original size. Similar results for ViT-small are presented in Table 17 in the Appendix." + }, + { + "type": "image", + "bbox": [ + 0.116, + 0.254, + 0.462, + 0.446 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.455, + 0.483, + 0.497 + ], + "angle": 0, + "content": "Figure 6. Comparison of Charm runtime using different patch selection strategies with the standard approach (Appendix 9). The x-axis shows the total number of pixels." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.519, + 0.483, + 0.671 + ], + "angle": 0, + "content": "single-scale tokenization approach (Section 3.2.1). Figure 6 illustrates the increase in the computational cost of Charm with larger image sizes. To reduce the overhead of saliency mask generation (16 seconds per image), we employ precomputed masks. The frequency-based method provides a cost-effective alternative to on-the-fly saliency detection, identifying regions comparable to or even superior to those selected by SAM-HQ. Random patch selection emerges as the most effective strategy, offering the best performance with the lowest computational cost." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.675, + 0.483, + 0.903 + ], + "angle": 0, + "content": "The computational cost of ViTs is influenced by the number of input tokens. The standard method involves downscaling and cropping images to \\(224 \\times 224\\), resulting in a very low number of tokens and lower costs. However, processing images at their original size can significantly increase computational costs. Our approach not only improves performance but also significantly reduces the computational costs of processing images in the original resolution (Table 5). We observe a decrease of over \\(84\\%\\) in GMACs, \\(77\\%\\) in runtime, \\(83\\%\\) in GPU memory consumption, and \\(75\\%\\) in the number of tokens for Dinov2-small when processing one \\(640 \\times 640\\) image using a 2-scale Charm. Even the 3-scale version, which is more costly, has significantly lower costs compared to processing images at their original size. While Charm may incur slightly" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.256, + 0.907, + 0.301 + ], + "angle": 0, + "content": "higher computational costs compared to the standard approach of downscaling and cropping, the substantial performance gains make it a worthwhile trade-off." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.311, + 0.722, + 0.327 + ], + "angle": 0, + "content": "4.4.5. Downscaling factor \\((f)\\)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.33, + 0.907, + 0.482 + ], + "angle": 0, + "content": "Our method initially tokenizes images with an initial patch size \\((p')\\). Subsequently, selected patches are further tokenized using a smaller patch size \\((p)\\), while the remaining patches are downscaled to \\(p\\). We experiment with different initial patch sizes \\(p' = (28, 42, 56)\\), corresponding to downscaling factors \\((f)\\) of 0.5, 0.66, and 0.75. The best performance is achieved with a factor of 0.5 (Appendix 19). A larger factor negatively impacts performance, as it leads to more significant downscaling of unselected patches, resulting in information loss for IAA." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.496, + 0.634, + 0.512 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.522, + 0.907, + 0.749 + ], + "angle": 0, + "content": "Processing images in their original resolution can be computationally expensive, requiring significant memory and processing power. Also, it is challenging to process images of varying aspect ratios and resolutions due to batch processing constraints. Charm effectively balances computational cost with the preservation of composition, high-resolution details, aspect ratio, and multi-scale information, which are all crucial for achieving optimal performance in IAA. By introducing variability into the input, our method enhances the generalizability of ViTs in IAA. Charm is compatible with any ViT architecture (except Swin [32], see Appendix 20) and does not require additional pre-training of ViTs. Our method consistently outperforms existing approaches that address similar research questions, achieving promising results across diverse IAA and IQA datasets." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.909, + 0.841 + ], + "angle": 0, + "content": "While our approach effectively preserves critical aesthetic-related factors, there's a potential for information loss due to selective downscaling. To further improve performance, future work could explore more sophisticated patch selection strategies, incorporate other modalities, and integrate our method with existing IAA-specific networks." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.841, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Acknowledgements: Funded by the European Union (ERC AdG, GRAPPA, 101053925, awarded to Johan Wagemans) and the Research Foundation-Flanders (FWO, 1159925N, awarded to Fatemeh Behrad)." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7822" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.171 + ], + "angle": 0, + "content": "[1] Irwan Bello, Barret Zoph, Ashish Vaswani, Jonathon Shlens, and Quoc V Le. Attention augmented convolutional networks. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3286-3295, 2019. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.173, + 0.484, + 0.268 + ], + "angle": 0, + "content": "[2] Lucas Beyer, Pavel Izmailov, Alexander Kolesnikov, Mathilde Caron, Simon Kornblith, Xiaohua Zhai, Matthias Minderer, Michael Tschannen, Ibrahim Alabdulmohsin, and Filip Pavetic. Flexivit: One model for all patch sizes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14496-14506, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.271, + 0.483, + 0.325 + ], + "angle": 0, + "content": "[3] Daniel Bolya, Cheng-Yang Fu, Xiaoliang Dai, Peizhao Zhang, Christoph Feichtenhofer, and Judy Hoffman. Token merging: Your vit but faster. arXiv preprint arXiv:2210.09461, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.327, + 0.483, + 0.394 + ], + "angle": 0, + "content": "[4] Chaofeng Chen, Jiadi Mo, Jingwen Hou, Haoning Wu, Liang Liao, Wenxiu Sun, Qiong Yan, and Weisi Lin. Topiq: A top-down approach from semantics to distortions for image quality assessment. IEEE Transactions on Image Processing, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.397, + 0.483, + 0.451 + ], + "angle": 0, + "content": "[5] Hangwei Chen, Feng Shao, Baoyang Mu, and Qiuping Jiang. Image aesthetics assessment with emotion-aware multibranch network. IEEE Transactions on Instrumentation and Measurement, 73:1-15, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.453, + 0.483, + 0.522 + ], + "angle": 0, + "content": "[6] Qiuyu Chen, Wei Zhang, Ning Zhou, Peng Lei, Yi Xu, Yu Zheng, and Jianping Fan. Adaptive fractional dilated convolution network for image aesthetics assessment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14114-14123, 2020. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.524, + 0.483, + 0.605 + ], + "angle": 0, + "content": "[7] Bowen Cheng, Bin Xiao, Jingdong Wang, Honghui Shi, Thomas S Huang, and Lei Zhang. Higherhrnet: Scale-aware representation learning for bottom-up human pose estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5386-5395, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.607, + 0.483, + 0.689 + ], + "angle": 0, + "content": "[8] Mostafa Dehghani, Basil Mustafa, Josip Djolonga, Jonathan Heek, Matthias Minderer, Mathilde Caron, Andreas Steiner, Joan Puigcerver, Robert Geirhos, Ibrahim M Alabdulmohsin, et al. Patch n'pack: Navit, a vision transformer for any aspect ratio and resolution. Advances in Neural Information Processing Systems, 36, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.691, + 0.483, + 0.773 + ], + "angle": 0, + "content": "[9] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 1, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.776, + 0.483, + 0.843 + ], + "angle": 0, + "content": "[10] Yuming Fang, Hanwei Zhu, Yan Zeng, Kede Ma, and Zhou Wang. Perceptual quality assessment of smartphone photography. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3677-3686, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.483, + 0.9 + ], + "angle": 0, + "content": "[11] Koustav Ghosal and Aljosa Smolic. Image aesthetics assessment using graph attention network. In 2022 26th International Conference on Pattern Recognition (ICPR), pages 3160-3167. IEEE, 2022. 2, 6" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.133 + ], + "angle": 0, + "content": "[12] Charles D Gilbert and Wu Li. Top-down influences on visual processing. Nature Reviews Neuroscience, 14(5):350-363, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.136, + 0.905, + 0.205 + ], + "angle": 0, + "content": "[13] Jakob Drachmann Havtorn, Amélie Royer, Tijmen Blankevoort, and Babak Ehteshami Bejnordi. Msvit: Dynamic mixed-scale tokenization for vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 838-848, 2023. 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.208, + 0.905, + 0.261 + ], + "angle": 0, + "content": "[14] Shuai He, Yongchang Zhang, Rui Xie, Dongxiang Jiang, and Anlong Ming. Rethinking image aesthetics assessment: Models, datasets and benchmarks. In *IJCAI*, pages 942–948, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.264, + 0.905, + 0.319 + ], + "angle": 0, + "content": "[15] Shuai He, Anlong Ming, Shuntian Zheng, Haobin Zhong, and Huadong Ma. Eat: An enhancer for aesthetics-oriented transformers. In Proceedings of the 31st ACM International Conference on Multimedia, pages 1023-1032, 2023. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.321, + 0.905, + 0.36 + ], + "angle": 0, + "content": "[16] Simon Hentschel, Konstantin Kobs, and Andreas Hotho. Clip knows image aesthetics. Frontiers in Artificial Intelligence, 5:976235, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.363, + 0.904, + 0.404 + ], + "angle": 0, + "content": "[17] Shaul Hochstein and Merav Ahissar. View from the top: Hierarchies and reverse hierarchies in the visual system. Neuron, 36(5):791-804, 2002. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.407, + 0.905, + 0.474 + ], + "angle": 0, + "content": "[18] Vlad Hosu, Bastian Goldlucke, and Dietmar Saupe. Effective aesthetics prediction with multi-level spatially pooled features. In proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9375-9383, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.477, + 0.905, + 0.532 + ], + "angle": 0, + "content": "[19] Vlad Hosu, Hanhe Lin, Tamas Sziranyi, and Dietmar Saupe. Koniq-10k: An ecologically valid database for deep learning of blind image quality assessment. IEEE Transactions on Image Processing, 29:4041–4056, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.534, + 0.905, + 0.588 + ], + "angle": 0, + "content": "[20] Yipo Huang, Leida Li, Pengfei Chen, Jinjian Wu, Yuzhe Yang, Yaqian Li, and Guangming Shi. Coarse-to-fine image aesthetics assessment with dynamic attribute selection. IEEE Transactions on Multimedia, 26:9316-9329, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.591, + 0.905, + 0.646 + ], + "angle": 0, + "content": "[21] Junjie Ke, Qifei Wang, Yilin Wang, Peyman Milanfar, and Feng Yang. Musiq: Multi-scale image quality transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5148-5157, 2021. 1, 2, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.649, + 0.905, + 0.716 + ], + "angle": 0, + "content": "[22] Junjie Ke, Keren Ye, Jiahui Yu, Yonghui Wu, Peyman Milanfar, and Feng Yang. Vila: Learning image aesthetics from user comments with vision-language pretraining. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10041-10051, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.719, + 0.905, + 0.772 + ], + "angle": 0, + "content": "[23] Junjie Ke, Tianhao Zhang, Yilin Wang, Peyman Milanfar, and Feng Yang. Mret: Multi-resolution transformer for video quality assessment. Frontiers in Signal Processing, 3:1137006, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.775, + 0.905, + 0.829 + ], + "angle": 0, + "content": "[24] Lei Ke, Mingqiao Ye, Martin Danelljan, Yu-Wing Tai, Chi-Keung Tang, Fisher Yu, et al. Segment anything in high quality. Advances in Neural Information Processing Systems, 36, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.832, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[25] Minchul Kim, Shangqian Gao, Yen-Chang Hsu, Yilin Shen, and Hongxia Jin. Token fusion: Bridging the gap between token pruning and token merging. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1383-1392, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7823" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.176 + ], + "angle": 0, + "content": "[26] Shu Kong, Xiaohui Shen, Zhe Lin, Radomir Mech, and Charless Fowlkes. Photo aesthetics ranking network with attributes and content adaptation. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part I 14, pages 662-679. Springer, 2016. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.177, + 0.484, + 0.245 + ], + "angle": 0, + "content": "[27] Jin Li, Yaoming Wang, Xiaopeng Zhang, Bowen Shi, Dongsheng Jiang, Chenglin Li, Wenrui Dai, Hongkai Xiong, and Qi Tian. Ailurus: a scalable vit framework for dense prediction. Advances in Neural Information Processing Systems, 36, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.246, + 0.484, + 0.315 + ], + "angle": 0, + "content": "[28] Leida Li, Yipo Huang, Jinjian Wu, Yuzhe Yang, Yaqian Li, Yandong Guo, and Guangming Shi. Theme-aware visual attribute reasoning for image aesthetics assessment. IEEE Transactions on Circuits and Systems for Video Technology, 33(9):4798-4811, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.316, + 0.484, + 0.383 + ], + "angle": 0, + "content": "[29] Leida Li, Tong Zhu, Pengfei Chen, Yuzhe Yang, Yaqian Li, and Weisi Lin. Image aesthetics assessment with attribute-assisted multimodal memory network. IEEE Transactions on Circuits and Systems for Video Technology, 33(12):7413-7424, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.385, + 0.484, + 0.453 + ], + "angle": 0, + "content": "[30] Leida Li, Xiangfei Sheng, Pengfei Chen, Jinjian Wu, and Weisheng Dong. Towards explainable image aesthetics assessment with attribute-oriented critiques generation. IEEE Transactions on Circuits and Systems for Video Technology, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.455, + 0.484, + 0.51 + ], + "angle": 0, + "content": "[31] Limin Liu, Shuai He, Anlong Ming, Rui Xie, and Huadong Ma. Elta: An enhancer against long-tail for aesthetic-oriented models. In *Forty-first International Conference on Machine Learning*. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.511, + 0.484, + 0.58 + ], + "angle": 0, + "content": "[32] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.58, + 0.484, + 0.607 + ], + "angle": 0, + "content": "[33] Ilya Loshchilov and Frank Hutter. Fixing weight decay regularization in adam. 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.608, + 0.484, + 0.664 + ], + "angle": 0, + "content": "[34] Long Mai, Hailin Jin, and Feng Liu. Composition-preserving deep photo aesthetics assessment. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 497-506, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.664, + 0.484, + 0.72 + ], + "angle": 0, + "content": "[35] Naila Murray, Luca Marchesotti, and Florent Perronnin. Ava: A large-scale database for aesthetic visual analysis. In 2012 IEEE Conference on Computer Vision and Pattern Recognition, pages 2408-2415, 2012. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.721, + 0.484, + 0.775 + ], + "angle": 0, + "content": "[36] Tomer Ronen, Omer Levy, and Avram Golbert. Vision transformers with mixed-resolution tokenization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4613-4622, 2023. 2, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.776, + 0.484, + 0.831 + ], + "angle": 0, + "content": "[37] Yossi Rubner, Carlo Tomasi, and Leonidas J Guibas. A metric for distributions with applications to image databases. In Sixth international conference on computer vision (IEEE Cat. No. 98CH36271), pages 59-66. IEEE, 1998. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.484, + 0.873 + ], + "angle": 0, + "content": "[38] Peter Shaw, Jakob Uszkoreit, and Ashish Vaswani. Self-attention with relative position representations. arXiv preprint arXiv:1803.02155, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.484, + 0.903 + ], + "angle": 0, + "content": "[39] Xiangfei Sheng, Leida Li, Pengfei Chen, Jinjian Wu, Weisheng Dong, Yuzhe Yang, Liwu Xu, Yaqian Li, and" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.092, + 0.906, + 0.147 + ], + "angle": 0, + "content": "Guangming Shi. Aesclip: Multi-attribute contrastive learning for image aesthetics assessment. In Proceedings of the 31st ACM International Conference on Multimedia, pages 1117-1126, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.149, + 0.906, + 0.218 + ], + "angle": 0, + "content": "[40] Tengfei Shi, Chenglizhao Chen, Zhenyu Wu, Aimin Hao, and Yuming Fang. Improving image aesthetic assessment via multiple image joint learning. ACM Transactions on Multimedia Computing, Communications and Applications, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.22, + 0.906, + 0.288 + ], + "angle": 0, + "content": "[41] Nyeong-Ho Shin, Seon-Ho Lee, and Chang-Su Kim. Blind image quality assessment based on geometric order learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12799-12808, 2024. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.29, + 0.906, + 0.346 + ], + "angle": 0, + "content": "[42] Hossein Talebi and Peyman Milanfar. Learning to resize images for computer vision tasks. In Proceedings of the IEEE/CVF International Conference on Computer vision, pages 497-506, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.346, + 0.906, + 0.403 + ], + "angle": 0, + "content": "[43] Zhengzhong Tu, Peyman Milanfar, and Hossein Talebi. Muller: Multilayer laplacian resizer for vision. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6877-6887, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.403, + 0.906, + 0.459 + ], + "angle": 0, + "content": "[44] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.46, + 0.906, + 0.543 + ], + "angle": 0, + "content": "[45] Jingdong Wang, Ke Sun, Tianheng Cheng, Borui Jiang, Chaorui Deng, Yang Zhao, Dong Liu, Yadong Mu, Mingkui Tan, Xinggang Wang, et al. Deep high-resolution representation learning for visual recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(10):3349-3364, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.544, + 0.906, + 0.6 + ], + "angle": 0, + "content": "[46] Jian Wang, Xiang Long, Guowei Chen, Zewu Wu, Zeyu Chen, and Errui Ding. U-hrnet: delving into improving semantic representation of high resolution network for dense prediction. arXiv preprint arXiv:2210.07140, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.601, + 0.906, + 0.643 + ], + "angle": 0, + "content": "[47] Oliver Wiedemann, Vlad Hosu, Shaolin Su, and Dietmar Saupe. Konx: cross-resolution image quality assessment. Quality and User Experience, 8(1):8, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.644, + 0.906, + 0.713 + ], + "angle": 0, + "content": "[48] Yuzhe Yang, Liwu Xu, Leida Li, Nan Qie, Yaqian Li, Peng Zhang, and Yandong Guo. Personalized image aesthetics assessment with rich attributes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19861-19869, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.714, + 0.906, + 0.784 + ], + "angle": 0, + "content": "[49] Ran Yi, Haoyuan Tian, Zhihao Gu, Yu-Kun Lai, and Paul L Rosin. Towards artistic image aesthetics assessment: a large-scale dataset and a new method. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22388-22397, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.785, + 0.906, + 0.84 + ], + "angle": 0, + "content": "[50] Tong Zhu, Leida Li, Pengfei Chen, Jinjian Wu, Yuzhe Yang, and Yaqian Li. Emotion-aware hierarchical interaction network for multimodal image aesthetics assessment. Pattern Recognition, 154:110584, 2024. 6" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.906, + 0.84 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "7824" + } + ] +] \ No newline at end of file diff --git a/2025/Charm_ The Missing Piece in ViT Fine-Tuning for Image Aesthetic Assessment/0ec61778-fb6e-453d-bfdb-3d054143034b_origin.pdf b/2025/Charm_ The Missing Piece in ViT Fine-Tuning for Image Aesthetic Assessment/0ec61778-fb6e-453d-bfdb-3d054143034b_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..ec0d207aa47063161bcf114ea8e1f3fdecc824b5 --- /dev/null +++ b/2025/Charm_ The Missing Piece in ViT Fine-Tuning for Image Aesthetic Assessment/0ec61778-fb6e-453d-bfdb-3d054143034b_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:374ac9528e2040cba89bc2bc353eb0717c8b4d0523f2216e884da51b37928709 +size 2666175 diff --git a/2025/Charm_ The Missing Piece in ViT Fine-Tuning for Image Aesthetic Assessment/full.md b/2025/Charm_ The Missing Piece in ViT Fine-Tuning for Image Aesthetic Assessment/full.md new file mode 100644 index 0000000000000000000000000000000000000000..7961f2c6f126a5c661e34f3d8c860ed2acb5de3a --- /dev/null +++ b/2025/Charm_ The Missing Piece in ViT Fine-Tuning for Image Aesthetic Assessment/full.md @@ -0,0 +1,279 @@ +# Charm: The Missing Piece in ViT fine-tuning for Image Aesthetic Assessment + +# Fatemeh Behrad, Tinne Tuytelaars, Johan Wagemans +KU Leuven University, Belgium + +{fatemeh.behrad, tinne.tuytelaars, johan.wagemans}@kuleuven.be + +# Abstract + +The capacity of Vision transformers (ViTs) to handle variable-sized inputs is often constrained by computational complexity and batch processing limitations. Consequently, ViTs are typically trained on small, fixed-size images obtained through downscaling or cropping. While reducing computational burden, these methods result in significant information loss, negatively affecting tasks like image aesthetic assessment. We introduce Charm, a novel tokenization approach that preserves Composition, High-resolution, Aspect Ratio, and Multi-scale information simultaneously. Charm prioritizes high-resolution details in specific regions while downscaling others, enabling shorter fixed-size input sequences for ViTs while incorporating essential information. Charm is designed to be compatible with pre-trained ViTs and their learned positional embeddings. By providing multiscale input and introducing variety to input tokens, Charm improves ViT performance and generalizability for image aesthetic assessment. We avoid cropping or changing the aspect ratio to further preserve information. Extensive experiments demonstrate significant performance improvements on various image aesthetic and quality assessment datasets (up to $8.1\%$ ) using a lightweight ViT backbone. Code and pre-trained models are available at https://github.com/FBehrad/Charm. + +# 1. Introduction + +Humans are inherently drawn to aesthetically pleasing images, making image aesthetics a vital consideration in fields like photography, advertising, e-commerce, and social media. Image aesthetic assessment (IAA) aims to develop computational models that can accurately predict human aesthetic preferences for images. + +Deep learning approaches applied to IAA have evolved significantly, with the Vision Transformer (ViT) architecture [9] demonstrating promising results [16, 21, 22]. ViTs process images as sequences of non-overlapping patches, allowing them to handle images of various sizes. However, in practice, this flexibility is limited for several rea + +![](images/9d6cb4720a139a08e737ee1fe9145a5835116f767d3a1b74c7867565304b0b9c.jpg) +Figure 1. Traditional image preprocessing methods (right) often distort image aesthetics by rescaling and cropping images to a fixed size. To preserve critical aesthetic information, our approach (left) selectively maintains some regions in their original resolution while downscaling the others. + +sons. First, position embeddings of pre-trained ViTs are learned for a specific, fixed input size. Additionally, batch processing requires all sequences in a batch to be the same length, which is challenging with images of different dimensions. Grouping images by size or padding can help. However, grouping images may affect the learning process and convergence, as it makes batches less representative of the entire distribution, while padding can lead to inefficient resource utilization. + +This brings us to another major issue when processing high-resolution images, which is the excessive computation and memory requirements. This can be particularly challenging when dealing with limited computational resources. As a result, a standard approach is to resize and + +downscale images to a smaller, fixed size to manage these constraints more effectively. This approach works for tasks where downscaling or cropping does not affect labels. However, in IAA, these operations can change aesthetic scores by altering the image composition and aspect ratio and losing high-resolution details (Figure 1). To unlock the full potential of ViTs in IAA, a new approach is needed to preserve critical aesthetic information while remaining computationally practical. + +Several techniques, such as graph attention networks [11], adaptive fractional dilated convolution networks [6], and multi-scale ViT [21], have been proposed to preserve image aspect ratio for IAA. Others, such as native resolution ViT [8], extend beyond IAA, demonstrating the benefits of preserving aspect ratio in other tasks. However, to remain practical, existing approaches still require downscaling the entire image, which limits their effectiveness. Fixed-size inputs can also limit the generalization of IAA models across various image sizes, which can be addressed by multi-scale inputs [18, 21, 23, 47]. + +One possible solution to preserve high-resolution information is a learnable resizer that reduces image dimensions while minimizing information loss during downsampling [7, 42]. While effective for other tasks, these methods can introduce distortions that alter the aesthetic quality of images. Another approach is to prioritize high-resolution information in some regions while downscaling less critical areas. This solution is inspired by the human brain, which processes visual information efficiently by preserving both low- and high-resolution details across the cortical hierarchy, maintaining spatial relationships, and utilizing selective attention and eye movements [12, 17]. Mixed-resolution [36] and mixed-scale tokenization [13] are steps in this direction. However, these methods produce identical representations for different resolutions of the same image because they first downscale all images to a fixed, low resolution for batch processing. While this approach is sufficient for image classification, it is inadequate for IAA, where resolution influences the aesthetic score. + +Our contribution. Existing approaches often prioritize a single factor crucial for aesthetics, neglecting others. Preserving composition, high resolution, aspect ratio, and multi-scale information simultaneously remains a challenge. We introduce Charm to fill this gap. Charm avoids fully downscaling images and maintains some areas in their original resolution (Figure 1). Our goal is to create a balance between computational efficiency and network performance. Charm effectively captures aspect ratio, composition, and multiscale information. Our framework-agnostic tokenization strategy avoids extensive pre-training by not modifying the pre-trained ViT architecture or its learned position embeddings. To demonstrate the broader applicability of our approach, we evaluate it on both IAA and image + +quality assessment (IQA) tasks. + +# 2. Related Work + +Image aesthetic assessment Existing deep learning-based IAA methods often require images to be resized or cropped to a relatively small fixed size, leading to loss of aspect ratio, composition, and high-resolution detail. Using a batch size of 1 avoids the need for image downscaling or cropping [34]. However, this can be computationally inefficient for large-scale training. Graph-based representations, such as those proposed by Ghosal et al. [11], can be memory-intensive for large images. Similarly, adaptive fractional dilated convolutions, introduced by Chen et al. [6], can be computationally expensive and require careful batch processing. A more recent strategy for preserving high resolution and aspect ratio is the use of learnable resizers. These methods employ auxiliary networks to reduce the dimensionality of input images while minimizing information loss during downsampling [7, 42, 45, 46]. While effective for image classification, these approaches introduce distortions that negatively impact image aesthetics. + +Another challenge associated with fixed-size inputs is a decrease in model generalization across various image sizes. While this challenge can be mitigated by supplying multi-scale inputs to the network [18, 21, 23, 47], existing approaches introduce non-negligible computational costs. + +Vision transformers Among existing neural networks, ViTs offer the advantage of processing images in their original size. However, they are still constrained by the need for fixed-size inputs due to batch processing. NaViT [8] addresses this problem by packing patches from different images into a fixed-size sequence. However, this approach can be computationally expensive as it can result in excessively long input sequences for high-resolution images. + +The computational costs of ViTs are influenced by the number of input tokens they process, so downscaling images can reduce the number of tokens and improve efficiency. Token dropping, either randomized or structured [27], and token merging [3] are techniques to improve ViT efficiency. Some methods, like token fusion [25], combine these approaches to achieve even better results. + +Previous studies have shown that incorporating multiscale features can enhance ViT performance. MUSIQ [21], for instance, utilizes patches from different scales to capture multi-scale information. While effective, this approach modifies the first layers of the ViT and introduces new position embeddings, which necessitates extensive training for each ViT backbone. FlexiViT [2], a more recent approach, extracts multiscale features using random patch sizes in each training epoch. + +Closer to our approach are mixed-resolution [36] and mixed-scale tokenization [13]. These methods first downscale images to a fixed size, then extract two-scale patches, + +preserving key regions in higher resolution while downscaling others. However, this initial downscaling discards high-frequency details and limits patches to two fixed resolutions. In contrast, Charm retains high-frequency information from the original image, learning from patches at varying resolutions (see Appendix 6 for details). Mixed-resolution ViT [36] further limits the input format, requiring images to be square. Additionally, it replaces learned position embeddings with sinusoidal ones, necessitating additional pre-training of each ViT backbone for compatibility. + +Existing methods fail to simultaneously preserve multiscale information, aspect ratio, composition, and high-resolution details. While all of these factors are crucial for accurate IAA, current methods typically prioritize computational efficiency at the expense of preserving these critical aspects. In contrast, our method successfully preserves all these factors while remaining computationally practical. + +Position embedding Transformers lack inherent mechanisms to capture the order of elements within an input sequence [44]. This stems from their reliance on self-attention, which is permutation-invariant – meaning the order of elements in the input does not affect the final output [1]. As a result, they require position embeddings to encode the sequence order. These position embeddings can be a deterministic function of position, typically using sinusoidal functions [44] or learned representations. Learnable position embeddings have achieved comparable performance to deterministic approaches [9]. However, both deterministic and learnable position embeddings are pre-trained for a specific sequence length. + +Relative position embeddings offer a more flexible approach by capturing the relative distance between elements [38]. This approach is well-suited for variable-length inputs. However, it necessitates significant modifications to the transformer attention mechanism. Recent methods like hash-based spatial embedding [21] and factorized position embedding [8] support variable aspect ratios and input resolutions without modifying attention modules. However, they cannot replace the position embedding of a pre-trained transformer without extensive retraining of the backbone on large-scale datasets. In contrast, our method adapts the position embedding of a pre-trained ViT model, requiring no additional pre-training and enabling efficient fine-tuning on downstream tasks with limited data. + +# 3. Unlock the potentials of ViTs in IAA + +In this section, we introduce Charm, a novel tokenization approach designed to enhance the performance of standard ViTs in image aesthetic and quality assessment tasks. Figure 2 provides an overview of our proposed approach. + +We initially partition the input image into fixed-size patches, preserving a selection of patches at their original resolution while downscaling others. This approach enables + +![](images/70910cb10c8517a4894a13e7f87ca44c2e3a2e4cb9c58c816255eea38be65c48.jpg) +Figure 2. Our approach involves partitioning the input image into fixed-size patches and selectively preserving a subset at its original resolution while downscaling others. Position and scale embeddings are then prepared to encode spatial relationships across different scales. The remaining steps follow the standard ViT. + indicates element-wise addition. + +the model to capture both fine-grained details and broader contextual information. Subsequently, position embeddings are prepared to encode the spatial relationships between patches across different resolutions and within the same resolution. Following previous studies [21], we also incorporate scale embeddings to aid the model in distinguishing information from different scales. + +After encoding the input patches into a sequence of tokens using the patch encoding module, we add position and scale embeddings. A learnable "classification token" (CLS) is prepended, and all tokens are processed through the transformer encoder. The final output is obtained by feeding the CLS token state at the output of the transformer through a fully connected layer. Our approach is compatible with various ViT architectures, as we do not modify the core ViT network structure. + +# 3.1. Preliminaries + +An image $x \in \mathbb{R}^{h \times w \times c}$ , where $(h, w, c)$ represents the height, width, and number of channels, respectively, undergoes tokenization in a standard ViT. This process divides the image into a sequence of $s$ patches $x_i \in \mathbb{R}^{p \times p \times c}$ , where $i \in (1, \ldots, s)$ and $p$ represents the patch size. The sequence length $s = \left\lfloor \frac{h}{p} \right\rfloor \cdot \left\lfloor \frac{w}{p} \right\rfloor$ determines the number of patches after tokenization, controlling the ViT's computational complexity. Given that the patch size $p$ remains fixed for a given ViT, the number of patches depends on the in- + +put image resolution. Consequently, high-resolution images increase computational complexity. + +# 3.2. Charm tokenization + +# 3.2.1. Tokenization + +Batch processing necessitates a consistent spatial resolution within each minibatch, leading to a fixed input length $(l)$ for the ViT. Processing high-resolution images requires a large $l$ , which increases memory consumption and computational demands during training. Our method addresses this by strategically reducing $l$ while preserving aesthetic-related information (Figure 2). Given a fixed patch size $(p)$ , we can consider two scenarios. For low-resolution images $(s \leq l)$ , we follow the standard ViT approach, tokenizing the entire image with the patch size of $p$ . For high-resolution images $(s > l)$ , we keep some parts in high resolution while downscaling other areas. + +When dealing with 2 scales, we initially tokenize the image using a patch size of $p' = p \times n$ , where $n \geq 2$ . Subsequently, we select high-resolution patches and further tokenize them using a patch size of $p$ while the remaining patches are downscaled to $p$ . The downscaling factor for both height and width is $f = 1/n$ . Finally, we arrange all high-resolution patches at the beginning, followed by low-resolution ones to form the final input for the transformer encoder. This ordering is feasible due to the permutation invariance of ViTs [1], allowing us to avoid unnecessary implementation loops. To ensure a consistent input length ( $l$ ), padding or random token dropping is employed as needed. Additional details on the 3-scale version are available in the Appendix 7. + +# 3.2.2. Patch Importance + +Several strategies can be employed to determine which patches to retain at high resolution. The main goal is to preserve visually interesting areas and sharp details in high resolution. To identify these regions, we use frequency, gradient, entropy, and saliency maps. We compare them with a random selection of patches in each training epoch. More details on patch selection strategies, including illustrative examples, can be found in Figure 8 of Appendix 8. + +# 3.2.3. Position Embedding + +ViTs lack inherent knowledge of spatial relationships between patches. Position embeddings address this by injecting learned position information into each patch embedding, allowing the self-attention mechanism to capture how different parts of the image relate. Replacing the standard position embeddings of ViTs with custom ones [36] can be expensive, requiring extensive pre-training on large datasets. It also introduces compatibility issues when used with different pre-trained ViTs. To address these issues, our approach adapts the pre-trained position embeddings from the standard ViT models. + +For low-resolution images, we follow the standard ViT approach of resizing pre-trained position embeddings via bilinear interpolation. However, for high-resolution images, we create two position embeddings using bilinear interpolation based on the different patch sizes $(p$ and $p^{\prime})$ employed during tokenization (Figure 2). One position embedding retains the spatial relationships between low-resolution patches, while the other captures the finer spatial relationships within the high-resolution patches. These position embeddings, high and low resolution, are compatible because they are derived from the same pre-trained embedding. Similar to standard ViTs, each token from the output of the patch encoding module is then summed with its corresponding position embedding. + +Unlike existing methods [13, 36], our approach does not have a fixed image size, so position embeddings vary for each image. To avoid recalculating position embeddings in each epoch, we prepare them in the data loaders, keeping them frozen during fine-tuning. This aligns with the standard practice of freezing the initial layers (including position embeddings) and fine-tuning only the final layers. + +# 3.2.4. Scale Embedding + +Our tokenization approach results in multiscale tokens. Following previous studies [21], to help ViT distinguish between patches from different scales and capture their relationships, we incorporate learnable scale embeddings. These scale embeddings are randomly initialized and added element-wise to all patch embeddings (Figure 2). + +# 3.3. Fine-tuning + +Our method requires no additional pre-training. We process images in their original resolution without resizing or random cropping during fine-tuning. We avoid normalizing the images as normalization can introduce distortions that affect the overall aesthetic/quality score (Table 9 in Appendix 9). Our data augmentation is limited to random horizontal flipping and random rotations (90, 180, 270 degrees) with a $50\%$ probability, as they do not result in loss of composition or aspect ratio. While these methods might change the aesthetics for humans, they have been shown to improve the performance of ViTs (Appendix 10). During evaluation, data augmentation is completely omitted. We employ the L1 loss function for predicting the overall aesthetic/quality score and the Earth Mover's Distance (EMD) loss [37] for predicting the aesthetic score distribution. + +# 4. Experiments + +# 4.1. Datasets + +We evaluate our models on various IAA datasets, including AVA [35], PARA [48], TAD66k [14], AADB [26], BAID [49], and two IQA datasets, SPAQ [10] and KonIQ10k [19]. + +![](images/de01c09f0bdd248741726e0cb4a83fbb33c7de8e365abd857bbfbf066e5376e8.jpg) +Figure 3. An image with its aesthetic score distribution and mean aesthetic score, from the AVA dataset. + +We only use the images and their corresponding ratings, excluding any additional information provided in the datasets. The AVA and PARA datasets associate each image with an aesthetic score distribution, while other datasets associate each image with an overall aesthetic/quality score (Figure 3). As the SPAQ and KonIQ10k datasets lack an official train-test split, we repeat the experiments for 5 different splits and report the median performance (as done in [41]). + +The AVA dataset remains prominent due to its large number of images, allowing for effective training of large models. Research utilizing smaller datasets is relatively limited, and most existing approaches leverage the additional information provided in these datasets. However, considering various datasets with different resolution ranges allows us to validate the efficiency of our approach in various scenarios. For more details on the resolution distribution of each dataset, see Figure 13 in Appendix 11. + +# 4.2. Implementation details + +Saliency maps for patch selection are generated using the SAM-HQ model [24] with the text prompt of "salient object" (Appendix 8). For gradient, frequency, and entropy calculations, we use the Sobel filter, fast Fourier transform, and Shannon entropy, respectively. + +In all experiments, we fine-tune the entire ViT models, while keeping their position embeddings frozen. Also, we use scale embeddings alongside position embeddings to enable Charm to capture relationships between tokens both within and across scales (see Appendix 12). The patch size $(p)$ and patch stride are chosen to match the filter size and stride of the patch encoding modules within these models (16 in ViT-small and 14 in Dinov2-small and Dinov2-large), ensuring compatibility and replicating the models' internal tokenization process. + +The maximum number of tokens $(l)$ is set to 512 for AVA, 768 for TAD66k, and 1024 for other datasets. These values are chosen based on the image resolution in the datasets (Appendix 13). For all experiments, we use a factor size of 0.5. + +The AdamW optimizer [33] is employed for fine-tuning. We use an initial learning rate of 1e-5 for all datasets except BAID (1e-6). We gradually reduce the learning rate dur + +
DatasetCharmPLCCSRCCACC
IAA
AVA-0.7100.7060.802
✓0.779(↑ 6.9%)0.777(↑ 7.1%)0.826(↑ 2.4%)
AADB-0.6950.6820.754
✓0.767(↑ 7.2%)0.754(↑ 7.2%)0.767(↑ 1.3%)
TAD66k-0.4290.4010.646
✓0.488(↑ 5.9%)0.458(↑ 5.7%)0.794 (↑ 14.8%)
PARA-0.9040.8550.863
✓0.938(↑ 3.4%)0.905(↑ 5%)0.892(↑ 2.9%)
BAID-0.4280.3420.750
✓0.439(↑ 1.1%)0.368(↑ 2.6%)0.763(↑ 1.3%)
IQA
SPAQ-0.9110.9070.907
✓0.919(↑ 0.8%)0.915(↑ 0.8%)0.917(↑ 1.0%)
KonIQ10k-0.8960.8680.938
✓0.944(↑ 4.8%)0.93(↑ 6.2%)0.954(↑ 1.6%)
+ +Table 1. Performance improvement across different IAA and IQA datasets by replacing the standard tokenization of Dinov2-small with Charm (with a random patch selection strategy). + +ing training using the cosine annealing learning rate scheduler. The batch size is 64 for AVA, 32 for TAD66k, and 16 for the remaining datasets. Early stopping is also implemented to prevent overfitting. Fine-tuning is conducted on a single NVIDIA RTX A4500 GPU. Model performance is evaluated using three key metrics: Spearman rank-ordered correlation coefficient (SRCC), Pearson linear correlation coefficient (PLCC), and classification accuracy (ACC). + +# 4.3. Comparing with existing methods + +We benchmark Charm with various ViT backbones on different IAA and IQA datasets. We first fine-tune models on the datasets to create a baseline. We then replace their original patch encoding modules with two-scale Charm tokenization and fine-tune the entire models. We also considered a three-scale version of Charm but did not observe additional performance gains (Appendix 7). Therefore, we focus on the two-scale version throughout the paper. + +The results (Table 1 and Table 13 in Appendix 14) demonstrate that by simply replacing the patch encoding module of existing ViT models with Charm, we can achieve significant performance improvements, up to $7.5\%$ in PLCC, $8.1\%$ in SRCC, and $14.8\%$ in classification accuracy across various datasets. Among datasets, PARA and SPAQ datasets include extra high-resolution images (e.g., 2k by 3k pixels). Our experiments (Appendix 15) reveal + +
ModelARHRMSPLCCSRCCACC
Ghosal et al.[11]✓--0.7640.762-
Chen et al. [6]✓--0.6710.6490.832
Dinov2-Small (+ Padding) *✓✓-0.7090.7030.801
MUSIQ [21]✓-✓0.7380.7260.815
FlexiViT [2]*--✓0.7370.7350.812
Swin [32]*--✓0.7480.7510.816
Dinov2-Small (+ Muller [43]) *-✓-0.6820.6750.794
ViT-small (+ Charm) *✓✓✓0.7620.7600.827
Dinov2-small (+ Charm) *✓✓✓0.7790.7770.826
Dinov2-large (+ Charm) *✓✓✓0.7830.7810.828
+ +Table 2. Results on the AVA dataset. Bold and underlined numbers indicate the best and the second-best results, respectively. AR, HR, and MS represent preserving aspect ratio, high-resolution details, and multi-scale information, respectively. * indicates models that we have fine-tuned on the AVA dataset. Appendix 14 shows the performance of the standard ViT-small, Dinov2-small, and Dinov2-large. + +a threshold for performance improvement by preserving high-resolution information. Beyond this threshold, further performance gains are limited. Therefore, to decrease computational costs, we downscale images in PARA and SPAQ to a maximum edge of 1024. + +As shown in Table 2, our approach outperforms existing methods in terms of preserving resolution, aspect ratio, and multiscale information. We compare Charm with the simple approach of adding padding to ensure consistent input shapes within each batch while preserving the aspect ratio. However, this padding proves detrimental to IAA and significantly increases computational costs (Appendix 16). Also, we observe that Muller [43], a recent approach focusing on preserving high-resolution information, can introduce noise and distortions to the image, negatively affecting aesthetics (Appendix 16). + +Recent studies have demonstrated the effectiveness of incorporating additional modalities, such as attributes ([5, 20, 28]) or text captions ([29, 30, 39, 50]), to improve IAA model performance (Table 15 in Appendix 17). Other studies have concentrated on other research directions [15, 40], like addressing the long tails of distributions in IAA datasets [31]. While the main contributions of these studies are orthogonal to Charm's, integrating Charm with them further enhances their performance. For example, incorporating Charm into ELTA [31] with a Dinov2-large backbone improves its PLCC, SRCC, and ACC on the AVA dataset from 0.742, 0.742, and 0.811 to 0.787, 0.786, and 0.829, respectively. Also, our approach demonstrates strong performance in IQA, with a slight performance gap of less than $1\%$ in SPAQ and $0.4\%$ in KonIQ10k compared to state-of-the-art methods [4, 21, 41] (Table 16 in Appendix 17). + +Model size is another crucial factor, as larger models often excel in complex tasks like IAA. In this paper, we demonstrate that Charm is highly effective even with lightweight ViTs, which can be used for applications with limited computational resources. As shown in Figure 4, Dinov2-small + Charm achieves comparable performance + +to state-of-the-art models while being significantly smaller. Nevertheless, Charm can serve as a plug-in for larger, more expensive models if higher performance is desired and computational power allows. For instance, plugging Charm into ViT-base instead of Dinov2-small results in PLCC of 0.515, SRCC of 0.467, and ACC of 0.779 after only 9 epochs on the BAID dataset. + +# 4.4. Ablation study + +# 4.4.1. The importance of preserving aspect ratio and high-resolution details + +Current approaches usually distort images by resizing them to a square shape, disregarding the significance of maintaining their original size and aspect ratio. We argue that such preprocessing can be detrimental to IAA. To validate this claim, we compare the performance of Dinov2-small using different tokenization approaches. Table 3 presents the results of these experiments. Except for the standard tokenization (described in Appendix 9), all other settings involve multiscale inputs and maintain composition by avoiding image cropping. + +We initially downscale images to $256 \times 256$ and then preserve some areas at that resolution while downscaling others by a factor of 0.5 (MS in Table 3). While this approach has been effective in image classification ([13] and [36]), it decreases performance in our IAA task. This is because the initial downscaling to a fixed size results in significant information loss critical for IAA, and further downscaling worsens the issue. + +We then experiment with keeping the image in its original resolution while changing the aspect ratio by resizing the image based on the maximum edge to create a square image (HR + MS in Table 3). This results in a nearly $3\%$ increase in PLCC and SRCC compared to the standard approach. Finally, we use Charm, which preserves both aspect ratio and high-resolution details, resulting in the best performance. This analysis highlights the importance of preserv + +![](images/5921f7f7008a27a222885dcd325854054845763dd141457a698dc98513d1ba9a.jpg) +Figure 4. Comparison of state-of-the-art IAA models' performance on the AVA dataset vs. their number of parameters. Our approach achieves comparable performance without using additional modalities and with significantly fewer parameters. To account for the significant differences in the number of parameters, we use a logarithmic scale on the x-axis. For a detailed comparison of Dinov2-small + Charm with state-of-the-art models on all datasets, see Appendix 17. + +
TokenizationPLCCSRCCACC
Standard0.7340.7320.808
MS0.7100.7050.801
HR + MS0.7630.7600.820
Charm (AR + HR + MS)0.7790.7770.826
+ +ing both aspect ratio and original resolution in IAA tasks. + +We also examined the impact of different numbers of tokens across tokenization approaches in Table 3 but found no differences in the results (see Appendix 18). + +# 4.4.2. Comparison of patch selection strategies + +To evaluate the importance of patches, we fine-tune Dinov2-small on the AADB dataset with different patch selection strategies. Table 4 demonstrates that all strategies outperform the standard ViT approach [9]. This suggests that factors beyond preserving high-resolution details also contribute to performance improvements (Section 4.4.1). + +Among all methods, the random selection of patches in each epoch stands out, likely for two reasons. First, ViTs are prone to overfitting, especially when fine-tuned on small datasets, a common challenge in IAA. Random patch selection in each epoch addresses this by introducing more variety to the ViT input and acting as a form of data augmentation. Second, random sampling can better capture the subjective and often unpredictable nature of beauty, providing a more representative view of the entire image. Following random selection, frequency, gradient, and entropy-based + +Table 3. Performance of Dinov2-small on the AVA dataset with four different tokenization strategies. MS, HR, and AR represent preserving multiscale information, high-resolution details, and aspect ratio, respectively. + +
Patch selectionPLCCSRCCACC
Random0.7670.7540.767
Entropy0.7260.7140.761
Frequency0.7560.7470.761
Gradient0.7340.7210.766
Saliency0.7510.7380.756
Standard0.6950.6820.754
+ +Table 4. Performance of Dinov2-small on the AADB dataset using Charm with different patch selection strategies. Bold and underlined values highlight the best and second-best results. + +![](images/11376fa752a04693d6692ea20f5fbbad7a86a080f479408b30da4065fc7d45cc.jpg) +Figure 5. The epoch at which each model achieves its highest validation performance across different datasets. Charm generally leads to faster convergence. The average number of epochs over 5 runs is reported for SPAQ and KonIQ10k. + +methods stand out. These methods still have a certain level of randomness, as their fully deterministic version results in highly correlated selected patches (Appendix 8). + +ViTs, originally trained for object classification, have an inherent bias towards salient objects. However, IAA requires attention to both foreground and background [15], which explains why the saliency-based approach performs less effectively. For a more comprehensive analysis of patch selection strategies please refer to Appendix 8. + +# 4.4.3. The effect of Charm on convergence speed + +We fine-tune all models until convergence, meaning their performance on the validation set stabilizes. Figure 5 shows the epoch at which each model achieves its best validation performance. As shown in this figure, Charm consistently accelerates the convergence of Dinov2-small across different datasets. Notably, the most significant reduction (50%) in training epochs occurs on AADB, our smallest dataset, likely due to Charm's ability to mitigate overfitting through random patch selection (see Appendix 8). On our largest datasets, AVA and TaD66k, both standard and Charm tokenization converge at similar epochs. + +# 4.4.4. Computational costs + +Figure 6 compares the runtime of Charm using different patch selection strategies with the standard approach on a 12th Gen Intel(R) Core(TM) i9-12900K CPU. Table 3 demonstrates that multiscale inputs can negatively impact performance for very small images. In this case, we use a + +
ModelInput sizeCharm#tokensmsGMACsMB
Dinov2224 x 224-2565.76.11202.9
-small-207032.884.012091.8
640 x 640✓2-scale:5127.3 (↓ 77.7%)13.46 (↓ 84%)346.0 (↓ 83.5%)
✓3-scale:7009.3 (↓ 71.6%)19.60 (↓ 76.7%)494.3 (↓ 76.4%)
+ +Table 5. Dinov2-small inference cost breakdown for processing one single image: number of tokens (#tokens) based on varying input sizes, runtime in milliseconds (ms), Giga multiply accumulation (GMACs), and GPU memory in Megabytes (MB). Bold and underlined values highlight the most and second-most computationally efficient configurations. Percentages indicate the reduction in computational cost compared to processing the image in its original size. Similar results for ViT-small are presented in Table 17 in the Appendix. + +![](images/64a2e6d2dd79a19fbd594a407610eff45754782f191d9deb597270e756b15a3e.jpg) +Figure 6. Comparison of Charm runtime using different patch selection strategies with the standard approach (Appendix 9). The x-axis shows the total number of pixels. + +single-scale tokenization approach (Section 3.2.1). Figure 6 illustrates the increase in the computational cost of Charm with larger image sizes. To reduce the overhead of saliency mask generation (16 seconds per image), we employ precomputed masks. The frequency-based method provides a cost-effective alternative to on-the-fly saliency detection, identifying regions comparable to or even superior to those selected by SAM-HQ. Random patch selection emerges as the most effective strategy, offering the best performance with the lowest computational cost. + +The computational cost of ViTs is influenced by the number of input tokens. The standard method involves downscaling and cropping images to $224 \times 224$ , resulting in a very low number of tokens and lower costs. However, processing images at their original size can significantly increase computational costs. Our approach not only improves performance but also significantly reduces the computational costs of processing images in the original resolution (Table 5). We observe a decrease of over $84\%$ in GMACs, $77\%$ in runtime, $83\%$ in GPU memory consumption, and $75\%$ in the number of tokens for Dinov2-small when processing one $640 \times 640$ image using a 2-scale Charm. Even the 3-scale version, which is more costly, has significantly lower costs compared to processing images at their original size. While Charm may incur slightly + +higher computational costs compared to the standard approach of downscaling and cropping, the substantial performance gains make it a worthwhile trade-off. + +# 4.4.5. Downscaling factor $(f)$ + +Our method initially tokenizes images with an initial patch size $(p')$ . Subsequently, selected patches are further tokenized using a smaller patch size $(p)$ , while the remaining patches are downscaled to $p$ . We experiment with different initial patch sizes $p' = (28, 42, 56)$ , corresponding to downscaling factors $(f)$ of 0.5, 0.66, and 0.75. The best performance is achieved with a factor of 0.5 (Appendix 19). A larger factor negatively impacts performance, as it leads to more significant downscaling of unselected patches, resulting in information loss for IAA. + +# 5. Conclusion + +Processing images in their original resolution can be computationally expensive, requiring significant memory and processing power. Also, it is challenging to process images of varying aspect ratios and resolutions due to batch processing constraints. Charm effectively balances computational cost with the preservation of composition, high-resolution details, aspect ratio, and multi-scale information, which are all crucial for achieving optimal performance in IAA. By introducing variability into the input, our method enhances the generalizability of ViTs in IAA. Charm is compatible with any ViT architecture (except Swin [32], see Appendix 20) and does not require additional pre-training of ViTs. Our method consistently outperforms existing approaches that address similar research questions, achieving promising results across diverse IAA and IQA datasets. + +While our approach effectively preserves critical aesthetic-related factors, there's a potential for information loss due to selective downscaling. To further improve performance, future work could explore more sophisticated patch selection strategies, incorporate other modalities, and integrate our method with existing IAA-specific networks. + +Acknowledgements: Funded by the European Union (ERC AdG, GRAPPA, 101053925, awarded to Johan Wagemans) and the Research Foundation-Flanders (FWO, 1159925N, awarded to Fatemeh Behrad). + +# References + +[1] Irwan Bello, Barret Zoph, Ashish Vaswani, Jonathon Shlens, and Quoc V Le. Attention augmented convolutional networks. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3286-3295, 2019. 3, 4 +[2] Lucas Beyer, Pavel Izmailov, Alexander Kolesnikov, Mathilde Caron, Simon Kornblith, Xiaohua Zhai, Matthias Minderer, Michael Tschannen, Ibrahim Alabdulmohsin, and Filip Pavetic. Flexivit: One model for all patch sizes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14496-14506, 2023. 2, 6 +[3] Daniel Bolya, Cheng-Yang Fu, Xiaoliang Dai, Peizhao Zhang, Christoph Feichtenhofer, and Judy Hoffman. Token merging: Your vit but faster. arXiv preprint arXiv:2210.09461, 2022. 2 +[4] Chaofeng Chen, Jiadi Mo, Jingwen Hou, Haoning Wu, Liang Liao, Wenxiu Sun, Qiong Yan, and Weisi Lin. Topiq: A top-down approach from semantics to distortions for image quality assessment. IEEE Transactions on Image Processing, 2024. 6 +[5] Hangwei Chen, Feng Shao, Baoyang Mu, and Qiuping Jiang. Image aesthetics assessment with emotion-aware multibranch network. IEEE Transactions on Instrumentation and Measurement, 73:1-15, 2024. 6 +[6] Qiuyu Chen, Wei Zhang, Ning Zhou, Peng Lei, Yi Xu, Yu Zheng, and Jianping Fan. Adaptive fractional dilated convolution network for image aesthetics assessment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14114-14123, 2020. 2, 6 +[7] Bowen Cheng, Bin Xiao, Jingdong Wang, Honghui Shi, Thomas S Huang, and Lei Zhang. Higherhrnet: Scale-aware representation learning for bottom-up human pose estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5386-5395, 2020. 2 +[8] Mostafa Dehghani, Basil Mustafa, Josip Djolonga, Jonathan Heek, Matthias Minderer, Mathilde Caron, Andreas Steiner, Joan Puigcerver, Robert Geirhos, Ibrahim M Alabdulmohsin, et al. Patch n'pack: Navit, a vision transformer for any aspect ratio and resolution. Advances in Neural Information Processing Systems, 36, 2024. 2, 3 +[9] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 1, 3, 7 +[10] Yuming Fang, Hanwei Zhu, Yan Zeng, Kede Ma, and Zhou Wang. Perceptual quality assessment of smartphone photography. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3677-3686, 2020. 4 +[11] Koustav Ghosal and Aljosa Smolic. Image aesthetics assessment using graph attention network. In 2022 26th International Conference on Pattern Recognition (ICPR), pages 3160-3167. IEEE, 2022. 2, 6 + +[12] Charles D Gilbert and Wu Li. Top-down influences on visual processing. Nature Reviews Neuroscience, 14(5):350-363, 2013. 2 +[13] Jakob Drachmann Havtorn, Amélie Royer, Tijmen Blankevoort, and Babak Ehteshami Bejnordi. Msvit: Dynamic mixed-scale tokenization for vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 838-848, 2023. 2, 4, 6 +[14] Shuai He, Yongchang Zhang, Rui Xie, Dongxiang Jiang, and Anlong Ming. Rethinking image aesthetics assessment: Models, datasets and benchmarks. In *IJCAI*, pages 942–948, 2022. 4 +[15] Shuai He, Anlong Ming, Shuntian Zheng, Haobin Zhong, and Huadong Ma. Eat: An enhancer for aesthetics-oriented transformers. In Proceedings of the 31st ACM International Conference on Multimedia, pages 1023-1032, 2023. 6, 7 +[16] Simon Hentschel, Konstantin Kobs, and Andreas Hotho. Clip knows image aesthetics. Frontiers in Artificial Intelligence, 5:976235, 2022. 1 +[17] Shaul Hochstein and Merav Ahissar. View from the top: Hierarchies and reverse hierarchies in the visual system. Neuron, 36(5):791-804, 2002. 2 +[18] Vlad Hosu, Bastian Goldlucke, and Dietmar Saupe. Effective aesthetics prediction with multi-level spatially pooled features. In proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9375-9383, 2019. 2 +[19] Vlad Hosu, Hanhe Lin, Tamas Sziranyi, and Dietmar Saupe. Koniq-10k: An ecologically valid database for deep learning of blind image quality assessment. IEEE Transactions on Image Processing, 29:4041–4056, 2020. 4 +[20] Yipo Huang, Leida Li, Pengfei Chen, Jinjian Wu, Yuzhe Yang, Yaqian Li, and Guangming Shi. Coarse-to-fine image aesthetics assessment with dynamic attribute selection. IEEE Transactions on Multimedia, 26:9316-9329, 2024. 6 +[21] Junjie Ke, Qifei Wang, Yilin Wang, Peyman Milanfar, and Feng Yang. Musiq: Multi-scale image quality transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5148-5157, 2021. 1, 2, 3, 4, 6 +[22] Junjie Ke, Keren Ye, Jiahui Yu, Yonghui Wu, Peyman Milanfar, and Feng Yang. Vila: Learning image aesthetics from user comments with vision-language pretraining. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10041-10051, 2023. 1 +[23] Junjie Ke, Tianhao Zhang, Yilin Wang, Peyman Milanfar, and Feng Yang. Mret: Multi-resolution transformer for video quality assessment. Frontiers in Signal Processing, 3:1137006, 2023. 2 +[24] Lei Ke, Mingqiao Ye, Martin Danelljan, Yu-Wing Tai, Chi-Keung Tang, Fisher Yu, et al. Segment anything in high quality. Advances in Neural Information Processing Systems, 36, 2024. 5 +[25] Minchul Kim, Shangqian Gao, Yen-Chang Hsu, Yilin Shen, and Hongxia Jin. Token fusion: Bridging the gap between token pruning and token merging. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1383-1392, 2024. 2 + +[26] Shu Kong, Xiaohui Shen, Zhe Lin, Radomir Mech, and Charless Fowlkes. Photo aesthetics ranking network with attributes and content adaptation. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part I 14, pages 662-679. Springer, 2016. 4 +[27] Jin Li, Yaoming Wang, Xiaopeng Zhang, Bowen Shi, Dongsheng Jiang, Chenglin Li, Wenrui Dai, Hongkai Xiong, and Qi Tian. Ailurus: a scalable vit framework for dense prediction. Advances in Neural Information Processing Systems, 36, 2024. 2 +[28] Leida Li, Yipo Huang, Jinjian Wu, Yuzhe Yang, Yaqian Li, Yandong Guo, and Guangming Shi. Theme-aware visual attribute reasoning for image aesthetics assessment. IEEE Transactions on Circuits and Systems for Video Technology, 33(9):4798-4811, 2023. 6 +[29] Leida Li, Tong Zhu, Pengfei Chen, Yuzhe Yang, Yaqian Li, and Weisi Lin. Image aesthetics assessment with attribute-assisted multimodal memory network. IEEE Transactions on Circuits and Systems for Video Technology, 33(12):7413-7424, 2023. 6 +[30] Leida Li, Xiangfei Sheng, Pengfei Chen, Jinjian Wu, and Weisheng Dong. Towards explainable image aesthetics assessment with attribute-oriented critiques generation. IEEE Transactions on Circuits and Systems for Video Technology, 2024. 6 +[31] Limin Liu, Shuai He, Anlong Ming, Rui Xie, and Huadong Ma. Elta: An enhancer against long-tail for aesthetic-oriented models. In *Forty-first International Conference on Machine Learning*. 6 +[32] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 6, 8 +[33] Ilya Loshchilov and Frank Hutter. Fixing weight decay regularization in adam. 2018. 5 +[34] Long Mai, Hailin Jin, and Feng Liu. Composition-preserving deep photo aesthetics assessment. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 497-506, 2016. 2 +[35] Naila Murray, Luca Marchesotti, and Florent Perronnin. Ava: A large-scale database for aesthetic visual analysis. In 2012 IEEE Conference on Computer Vision and Pattern Recognition, pages 2408-2415, 2012. 4 +[36] Tomer Ronen, Omer Levy, and Avram Golbert. Vision transformers with mixed-resolution tokenization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4613-4622, 2023. 2, 3, 4, 6 +[37] Yossi Rubner, Carlo Tomasi, and Leonidas J Guibas. A metric for distributions with applications to image databases. In Sixth international conference on computer vision (IEEE Cat. No. 98CH36271), pages 59-66. IEEE, 1998. 4 +[38] Peter Shaw, Jakob Uszkoreit, and Ashish Vaswani. Self-attention with relative position representations. arXiv preprint arXiv:1803.02155, 2018. 3 +[39] Xiangfei Sheng, Leida Li, Pengfei Chen, Jinjian Wu, Weisheng Dong, Yuzhe Yang, Liwu Xu, Yaqian Li, and + +Guangming Shi. Aesclip: Multi-attribute contrastive learning for image aesthetics assessment. In Proceedings of the 31st ACM International Conference on Multimedia, pages 1117-1126, 2023. 6 +[40] Tengfei Shi, Chenglizhao Chen, Zhenyu Wu, Aimin Hao, and Yuming Fang. Improving image aesthetic assessment via multiple image joint learning. ACM Transactions on Multimedia Computing, Communications and Applications, 2024. 6 +[41] Nyeong-Ho Shin, Seon-Ho Lee, and Chang-Su Kim. Blind image quality assessment based on geometric order learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12799-12808, 2024. 5, 6 +[42] Hossein Talebi and Peyman Milanfar. Learning to resize images for computer vision tasks. In Proceedings of the IEEE/CVF International Conference on Computer vision, pages 497-506, 2021. 2 +[43] Zhengzhong Tu, Peyman Milanfar, and Hossein Talebi. Muller: Multilayer laplacian resizer for vision. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6877-6887, 2023. 6 +[44] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 3 +[45] Jingdong Wang, Ke Sun, Tianheng Cheng, Borui Jiang, Chaorui Deng, Yang Zhao, Dong Liu, Yadong Mu, Mingkui Tan, Xinggang Wang, et al. Deep high-resolution representation learning for visual recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(10):3349-3364, 2020. 2 +[46] Jian Wang, Xiang Long, Guowei Chen, Zewu Wu, Zeyu Chen, and Errui Ding. U-hrnet: delving into improving semantic representation of high resolution network for dense prediction. arXiv preprint arXiv:2210.07140, 2022. 2 +[47] Oliver Wiedemann, Vlad Hosu, Shaolin Su, and Dietmar Saupe. Konx: cross-resolution image quality assessment. Quality and User Experience, 8(1):8, 2023. 2 +[48] Yuzhe Yang, Liwu Xu, Leida Li, Nan Qie, Yaqian Li, Peng Zhang, and Yandong Guo. Personalized image aesthetics assessment with rich attributes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19861-19869, 2022. 4 +[49] Ran Yi, Haoyuan Tian, Zhihao Gu, Yu-Kun Lai, and Paul L Rosin. Towards artistic image aesthetics assessment: a large-scale dataset and a new method. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22388-22397, 2023. 4 +[50] Tong Zhu, Leida Li, Pengfei Chen, Jinjian Wu, Yuzhe Yang, and Yaqian Li. Emotion-aware hierarchical interaction network for multimodal image aesthetics assessment. Pattern Recognition, 154:110584, 2024. 6 \ No newline at end of file diff --git a/2025/Charm_ The Missing Piece in ViT Fine-Tuning for Image Aesthetic Assessment/images.zip b/2025/Charm_ The Missing Piece in ViT Fine-Tuning for Image Aesthetic Assessment/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..9ba5096e6ea8417c06175c49dfad637ea7fb5062 --- /dev/null +++ b/2025/Charm_ The Missing Piece in ViT Fine-Tuning for Image Aesthetic Assessment/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9e2334db099474bf0ff59291b82a48c54e4c37a8ae7b50503a84d66d30f5e56d +size 385877 diff --git a/2025/Charm_ The Missing Piece in ViT Fine-Tuning for Image Aesthetic Assessment/layout.json b/2025/Charm_ The Missing Piece in ViT Fine-Tuning for Image Aesthetic Assessment/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..fddb024fd4a1007814a52f2dab157d4cb91bba00 --- /dev/null +++ b/2025/Charm_ The Missing Piece in ViT Fine-Tuning for Image Aesthetic Assessment/layout.json @@ -0,0 +1,7093 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 66, + 103, + 544, + 120 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 103, + 544, + 120 + ], + "spans": [ + { + "bbox": [ + 66, + 103, + 544, + 120 + ], + "type": "text", + "content": "Charm: The Missing Piece in ViT fine-tuning for Image Aesthetic Assessment" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 176, + 144, + 434, + 172 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 176, + 144, + 434, + 172 + ], + "spans": [ + { + "bbox": [ + 176, + 144, + 434, + 172 + ], + "type": "text", + "content": "Fatemeh Behrad, Tinne Tuytelaars, Johan Wagemans \nKU Leuven University, Belgium" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 136, + 174, + 471, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 174, + 471, + 185 + ], + "spans": [ + { + "bbox": [ + 136, + 174, + 471, + 185 + ], + "type": "text", + "content": "{fatemeh.behrad, tinne.tuytelaars, johan.wagemans}@kuleuven.be" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 213, + 200, + 224 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 213, + 200, + 224 + ], + "spans": [ + { + "bbox": [ + 151, + 213, + 200, + 224 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 238, + 296, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 238, + 296, + 525 + ], + "spans": [ + { + "bbox": [ + 55, + 238, + 296, + 525 + ], + "type": "text", + "content": "The capacity of Vision transformers (ViTs) to handle variable-sized inputs is often constrained by computational complexity and batch processing limitations. Consequently, ViTs are typically trained on small, fixed-size images obtained through downscaling or cropping. While reducing computational burden, these methods result in significant information loss, negatively affecting tasks like image aesthetic assessment. We introduce Charm, a novel tokenization approach that preserves Composition, High-resolution, Aspect Ratio, and Multi-scale information simultaneously. Charm prioritizes high-resolution details in specific regions while downscaling others, enabling shorter fixed-size input sequences for ViTs while incorporating essential information. Charm is designed to be compatible with pre-trained ViTs and their learned positional embeddings. By providing multiscale input and introducing variety to input tokens, Charm improves ViT performance and generalizability for image aesthetic assessment. We avoid cropping or changing the aspect ratio to further preserve information. Extensive experiments demonstrate significant performance improvements on various image aesthetic and quality assessment datasets (up to " + }, + { + "bbox": [ + 55, + 238, + 296, + 525 + ], + "type": "inline_equation", + "content": "8.1\\%" + }, + { + "bbox": [ + 55, + 238, + 296, + 525 + ], + "type": "text", + "content": " ) using a lightweight ViT backbone. Code and pre-trained models are available at https://github.com/FBehrad/Charm." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 548, + 135, + 561 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 548, + 135, + 561 + ], + "spans": [ + { + "bbox": [ + 55, + 548, + 135, + 561 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 569, + 295, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 569, + 295, + 641 + ], + "spans": [ + { + "bbox": [ + 55, + 569, + 295, + 641 + ], + "type": "text", + "content": "Humans are inherently drawn to aesthetically pleasing images, making image aesthetics a vital consideration in fields like photography, advertising, e-commerce, and social media. Image aesthetic assessment (IAA) aims to develop computational models that can accurately predict human aesthetic preferences for images." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": "Deep learning approaches applied to IAA have evolved significantly, with the Vision Transformer (ViT) architecture [9] demonstrating promising results [16, 21, 22]. ViTs process images as sequences of non-overlapping patches, allowing them to handle images of various sizes. However, in practice, this flexibility is limited for several rea" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 321, + 212, + 555, + 454 + ], + "blocks": [ + { + "bbox": [ + 321, + 212, + 555, + 454 + ], + "lines": [ + { + "bbox": [ + 321, + 212, + 555, + 454 + ], + "spans": [ + { + "bbox": [ + 321, + 212, + 555, + 454 + ], + "type": "image", + "image_path": "9d6cb4720a139a08e737ee1fe9145a5835116f767d3a1b74c7867565304b0b9c.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 461, + 555, + 517 + ], + "lines": [ + { + "bbox": [ + 313, + 461, + 555, + 517 + ], + "spans": [ + { + "bbox": [ + 313, + 461, + 555, + 517 + ], + "type": "text", + "content": "Figure 1. Traditional image preprocessing methods (right) often distort image aesthetics by rescaling and cropping images to a fixed size. To preserve critical aesthetic information, our approach (left) selectively maintains some regions in their original resolution while downscaling the others." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 543, + 555, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 543, + 555, + 650 + ], + "spans": [ + { + "bbox": [ + 313, + 543, + 555, + 650 + ], + "type": "text", + "content": "sons. First, position embeddings of pre-trained ViTs are learned for a specific, fixed input size. Additionally, batch processing requires all sequences in a batch to be the same length, which is challenging with images of different dimensions. Grouping images by size or padding can help. However, grouping images may affect the learning process and convergence, as it makes batches less representative of the entire distribution, while padding can lead to inefficient resource utilization." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "text", + "content": "This brings us to another major issue when processing high-resolution images, which is the excessive computation and memory requirements. This can be particularly challenging when dealing with limited computational resources. As a result, a standard approach is to resize and" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7815" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 180 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 180 + ], + "type": "text", + "content": "downscale images to a smaller, fixed size to manage these constraints more effectively. This approach works for tasks where downscaling or cropping does not affect labels. However, in IAA, these operations can change aesthetic scores by altering the image composition and aspect ratio and losing high-resolution details (Figure 1). To unlock the full potential of ViTs in IAA, a new approach is needed to preserve critical aesthetic information while remaining computationally practical." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 182, + 294, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 182, + 294, + 314 + ], + "spans": [ + { + "bbox": [ + 55, + 182, + 294, + 314 + ], + "type": "text", + "content": "Several techniques, such as graph attention networks [11], adaptive fractional dilated convolution networks [6], and multi-scale ViT [21], have been proposed to preserve image aspect ratio for IAA. Others, such as native resolution ViT [8], extend beyond IAA, demonstrating the benefits of preserving aspect ratio in other tasks. However, to remain practical, existing approaches still require downscaling the entire image, which limits their effectiveness. Fixed-size inputs can also limit the generalization of IAA models across various image sizes, which can be addressed by multi-scale inputs [18, 21, 23, 47]." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 316, + 295, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 316, + 295, + 543 + ], + "spans": [ + { + "bbox": [ + 55, + 316, + 295, + 543 + ], + "type": "text", + "content": "One possible solution to preserve high-resolution information is a learnable resizer that reduces image dimensions while minimizing information loss during downsampling [7, 42]. While effective for other tasks, these methods can introduce distortions that alter the aesthetic quality of images. Another approach is to prioritize high-resolution information in some regions while downscaling less critical areas. This solution is inspired by the human brain, which processes visual information efficiently by preserving both low- and high-resolution details across the cortical hierarchy, maintaining spatial relationships, and utilizing selective attention and eye movements [12, 17]. Mixed-resolution [36] and mixed-scale tokenization [13] are steps in this direction. However, these methods produce identical representations for different resolutions of the same image because they first downscale all images to a fixed, low resolution for batch processing. While this approach is sufficient for image classification, it is inadequate for IAA, where resolution influences the aesthetic score." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 546, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 546, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 546, + 295, + 713 + ], + "type": "text", + "content": "Our contribution. Existing approaches often prioritize a single factor crucial for aesthetics, neglecting others. Preserving composition, high resolution, aspect ratio, and multi-scale information simultaneously remains a challenge. We introduce Charm to fill this gap. Charm avoids fully downscaling images and maintains some areas in their original resolution (Figure 1). Our goal is to create a balance between computational efficiency and network performance. Charm effectively captures aspect ratio, composition, and multiscale information. Our framework-agnostic tokenization strategy avoids extensive pre-training by not modifying the pre-trained ViT architecture or its learned position embeddings. To demonstrate the broader applicability of our approach, we evaluate it on both IAA and image" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 443, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 443, + 84 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 443, + 84 + ], + "type": "text", + "content": "quality assessment (IQA) tasks." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 314, + 95, + 400, + 107 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 95, + 400, + 107 + ], + "spans": [ + { + "bbox": [ + 314, + 95, + 400, + 107 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 114, + 553, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 114, + 553, + 331 + ], + "spans": [ + { + "bbox": [ + 313, + 114, + 553, + 331 + ], + "type": "text", + "content": "Image aesthetic assessment Existing deep learning-based IAA methods often require images to be resized or cropped to a relatively small fixed size, leading to loss of aspect ratio, composition, and high-resolution detail. Using a batch size of 1 avoids the need for image downscaling or cropping [34]. However, this can be computationally inefficient for large-scale training. Graph-based representations, such as those proposed by Ghosal et al. [11], can be memory-intensive for large images. Similarly, adaptive fractional dilated convolutions, introduced by Chen et al. [6], can be computationally expensive and require careful batch processing. A more recent strategy for preserving high resolution and aspect ratio is the use of learnable resizers. These methods employ auxiliary networks to reduce the dimensionality of input images while minimizing information loss during downsampling [7, 42, 45, 46]. While effective for image classification, these approaches introduce distortions that negatively impact image aesthetics." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 331, + 553, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 331, + 553, + 390 + ], + "spans": [ + { + "bbox": [ + 313, + 331, + 553, + 390 + ], + "type": "text", + "content": "Another challenge associated with fixed-size inputs is a decrease in model generalization across various image sizes. While this challenge can be mitigated by supplying multi-scale inputs to the network [18, 21, 23, 47], existing approaches introduce non-negligible computational costs." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 391, + 553, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 391, + 553, + 486 + ], + "spans": [ + { + "bbox": [ + 313, + 391, + 553, + 486 + ], + "type": "text", + "content": "Vision transformers Among existing neural networks, ViTs offer the advantage of processing images in their original size. However, they are still constrained by the need for fixed-size inputs due to batch processing. NaViT [8] addresses this problem by packing patches from different images into a fixed-size sequence. However, this approach can be computationally expensive as it can result in excessively long input sequences for high-resolution images." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 487, + 553, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 487, + 553, + 569 + ], + "spans": [ + { + "bbox": [ + 313, + 487, + 553, + 569 + ], + "type": "text", + "content": "The computational costs of ViTs are influenced by the number of input tokens they process, so downscaling images can reduce the number of tokens and improve efficiency. Token dropping, either randomized or structured [27], and token merging [3] are techniques to improve ViT efficiency. Some methods, like token fusion [25], combine these approaches to achieve even better results." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 570, + 553, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 553, + 677 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 553, + 677 + ], + "type": "text", + "content": "Previous studies have shown that incorporating multiscale features can enhance ViT performance. MUSIQ [21], for instance, utilizes patches from different scales to capture multi-scale information. While effective, this approach modifies the first layers of the ViT and introduces new position embeddings, which necessitates extensive training for each ViT backbone. FlexiViT [2], a more recent approach, extracts multiscale features using random patch sizes in each training epoch." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "content": "Closer to our approach are mixed-resolution [36] and mixed-scale tokenization [13]. These methods first downscale images to a fixed size, then extract two-scale patches," + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7816" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "type": "text", + "content": "preserving key regions in higher resolution while downscaling others. However, this initial downscaling discards high-frequency details and limits patches to two fixed resolutions. In contrast, Charm retains high-frequency information from the original image, learning from patches at varying resolutions (see Appendix 6 for details). Mixed-resolution ViT [36] further limits the input format, requiring images to be square. Additionally, it replaces learned position embeddings with sinusoidal ones, necessitating additional pre-training of each ViT backbone for compatibility." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 192, + 294, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 192, + 294, + 275 + ], + "spans": [ + { + "bbox": [ + 55, + 192, + 294, + 275 + ], + "type": "text", + "content": "Existing methods fail to simultaneously preserve multiscale information, aspect ratio, composition, and high-resolution details. While all of these factors are crucial for accurate IAA, current methods typically prioritize computational efficiency at the expense of preserving these critical aspects. In contrast, our method successfully preserves all these factors while remaining computationally practical." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 275, + 295, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 275, + 295, + 430 + ], + "spans": [ + { + "bbox": [ + 55, + 275, + 295, + 430 + ], + "type": "text", + "content": "Position embedding Transformers lack inherent mechanisms to capture the order of elements within an input sequence [44]. This stems from their reliance on self-attention, which is permutation-invariant – meaning the order of elements in the input does not affect the final output [1]. As a result, they require position embeddings to encode the sequence order. These position embeddings can be a deterministic function of position, typically using sinusoidal functions [44] or learned representations. Learnable position embeddings have achieved comparable performance to deterministic approaches [9]. However, both deterministic and learnable position embeddings are pre-trained for a specific sequence length." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 431, + 295, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 431, + 295, + 597 + ], + "spans": [ + { + "bbox": [ + 55, + 431, + 295, + 597 + ], + "type": "text", + "content": "Relative position embeddings offer a more flexible approach by capturing the relative distance between elements [38]. This approach is well-suited for variable-length inputs. However, it necessitates significant modifications to the transformer attention mechanism. Recent methods like hash-based spatial embedding [21] and factorized position embedding [8] support variable aspect ratios and input resolutions without modifying attention modules. However, they cannot replace the position embedding of a pre-trained transformer without extensive retraining of the backbone on large-scale datasets. In contrast, our method adapts the position embedding of a pre-trained ViT model, requiring no additional pre-training and enabling efficient fine-tuning on downstream tasks with limited data." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 609, + 257, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 609, + 257, + 623 + ], + "spans": [ + { + "bbox": [ + 55, + 609, + 257, + 623 + ], + "type": "text", + "content": "3. Unlock the potentials of ViTs in IAA" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 629, + 294, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 294, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 294, + 677 + ], + "type": "text", + "content": "In this section, we introduce Charm, a novel tokenization approach designed to enhance the performance of standard ViTs in image aesthetic and quality assessment tasks. Figure 2 provides an overview of our proposed approach." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "content": "We initially partition the input image into fixed-size patches, preserving a selection of patches at their original resolution while downscaling others. This approach enables" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 340, + 70, + 531, + 289 + ], + "blocks": [ + { + "bbox": [ + 340, + 70, + 531, + 289 + ], + "lines": [ + { + "bbox": [ + 340, + 70, + 531, + 289 + ], + "spans": [ + { + "bbox": [ + 340, + 70, + 531, + 289 + ], + "type": "image", + "image_path": "70910cb10c8517a4894a13e7f87ca44c2e3a2e4cb9c58c816255eea38be65c48.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 297, + 555, + 363 + ], + "lines": [ + { + "bbox": [ + 313, + 297, + 555, + 363 + ], + "spans": [ + { + "bbox": [ + 313, + 297, + 555, + 363 + ], + "type": "text", + "content": "Figure 2. Our approach involves partitioning the input image into fixed-size patches and selectively preserving a subset at its original resolution while downscaling others. Position and scale embeddings are then prepared to encode spatial relationships across different scales. The remaining steps follow the standard ViT. + indicates element-wise addition." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 377, + 553, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 377, + 553, + 460 + ], + "spans": [ + { + "bbox": [ + 313, + 377, + 553, + 460 + ], + "type": "text", + "content": "the model to capture both fine-grained details and broader contextual information. Subsequently, position embeddings are prepared to encode the spatial relationships between patches across different resolutions and within the same resolution. Following previous studies [21], we also incorporate scale embeddings to aid the model in distinguishing information from different scales." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 463, + 554, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 463, + 554, + 568 + ], + "spans": [ + { + "bbox": [ + 313, + 463, + 554, + 568 + ], + "type": "text", + "content": "After encoding the input patches into a sequence of tokens using the patch encoding module, we add position and scale embeddings. A learnable \"classification token\" (CLS) is prepended, and all tokens are processed through the transformer encoder. The final output is obtained by feeding the CLS token state at the output of the transformer through a fully connected layer. Our approach is compatible with various ViT architectures, as we do not modify the core ViT network structure." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 582, + 400, + 594 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 400, + 594 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 400, + 594 + ], + "type": "text", + "content": "3.1. Preliminaries" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 600, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 600, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 600, + 555, + 713 + ], + "type": "text", + "content": "An image " + }, + { + "bbox": [ + 313, + 600, + 555, + 713 + ], + "type": "inline_equation", + "content": "x \\in \\mathbb{R}^{h \\times w \\times c}" + }, + { + "bbox": [ + 313, + 600, + 555, + 713 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 600, + 555, + 713 + ], + "type": "inline_equation", + "content": "(h, w, c)" + }, + { + "bbox": [ + 313, + 600, + 555, + 713 + ], + "type": "text", + "content": " represents the height, width, and number of channels, respectively, undergoes tokenization in a standard ViT. This process divides the image into a sequence of " + }, + { + "bbox": [ + 313, + 600, + 555, + 713 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 313, + 600, + 555, + 713 + ], + "type": "text", + "content": " patches " + }, + { + "bbox": [ + 313, + 600, + 555, + 713 + ], + "type": "inline_equation", + "content": "x_i \\in \\mathbb{R}^{p \\times p \\times c}" + }, + { + "bbox": [ + 313, + 600, + 555, + 713 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 600, + 555, + 713 + ], + "type": "inline_equation", + "content": "i \\in (1, \\ldots, s)" + }, + { + "bbox": [ + 313, + 600, + 555, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 600, + 555, + 713 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 313, + 600, + 555, + 713 + ], + "type": "text", + "content": " represents the patch size. The sequence length " + }, + { + "bbox": [ + 313, + 600, + 555, + 713 + ], + "type": "inline_equation", + "content": "s = \\left\\lfloor \\frac{h}{p} \\right\\rfloor \\cdot \\left\\lfloor \\frac{w}{p} \\right\\rfloor" + }, + { + "bbox": [ + 313, + 600, + 555, + 713 + ], + "type": "text", + "content": " determines the number of patches after tokenization, controlling the ViT's computational complexity. Given that the patch size " + }, + { + "bbox": [ + 313, + 600, + 555, + 713 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 313, + 600, + 555, + 713 + ], + "type": "text", + "content": " remains fixed for a given ViT, the number of patches depends on the in-" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7817" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 97 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 97 + ], + "type": "text", + "content": "put image resolution. Consequently, high-resolution images increase computational complexity." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 102, + 173, + 114 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 102, + 173, + 114 + ], + "spans": [ + { + "bbox": [ + 55, + 102, + 173, + 114 + ], + "type": "text", + "content": "3.2. Charm tokenization" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 121, + 139, + 131 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 121, + 139, + 131 + ], + "spans": [ + { + "bbox": [ + 55, + 121, + 139, + 131 + ], + "type": "text", + "content": "3.2.1. Tokenization" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 135, + 296, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 135, + 296, + 277 + ], + "spans": [ + { + "bbox": [ + 55, + 135, + 296, + 277 + ], + "type": "text", + "content": "Batch processing necessitates a consistent spatial resolution within each minibatch, leading to a fixed input length " + }, + { + "bbox": [ + 55, + 135, + 296, + 277 + ], + "type": "inline_equation", + "content": "(l)" + }, + { + "bbox": [ + 55, + 135, + 296, + 277 + ], + "type": "text", + "content": " for the ViT. Processing high-resolution images requires a large " + }, + { + "bbox": [ + 55, + 135, + 296, + 277 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 55, + 135, + 296, + 277 + ], + "type": "text", + "content": ", which increases memory consumption and computational demands during training. Our method addresses this by strategically reducing " + }, + { + "bbox": [ + 55, + 135, + 296, + 277 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 55, + 135, + 296, + 277 + ], + "type": "text", + "content": " while preserving aesthetic-related information (Figure 2). Given a fixed patch size " + }, + { + "bbox": [ + 55, + 135, + 296, + 277 + ], + "type": "inline_equation", + "content": "(p)" + }, + { + "bbox": [ + 55, + 135, + 296, + 277 + ], + "type": "text", + "content": ", we can consider two scenarios. For low-resolution images " + }, + { + "bbox": [ + 55, + 135, + 296, + 277 + ], + "type": "inline_equation", + "content": "(s \\leq l)" + }, + { + "bbox": [ + 55, + 135, + 296, + 277 + ], + "type": "text", + "content": ", we follow the standard ViT approach, tokenizing the entire image with the patch size of " + }, + { + "bbox": [ + 55, + 135, + 296, + 277 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 55, + 135, + 296, + 277 + ], + "type": "text", + "content": ". For high-resolution images " + }, + { + "bbox": [ + 55, + 135, + 296, + 277 + ], + "type": "inline_equation", + "content": "(s > l)" + }, + { + "bbox": [ + 55, + 135, + 296, + 277 + ], + "type": "text", + "content": ", we keep some parts in high resolution while downscaling other areas." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 279, + 296, + 446 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 279, + 296, + 446 + ], + "spans": [ + { + "bbox": [ + 55, + 279, + 296, + 446 + ], + "type": "text", + "content": "When dealing with 2 scales, we initially tokenize the image using a patch size of " + }, + { + "bbox": [ + 55, + 279, + 296, + 446 + ], + "type": "inline_equation", + "content": "p' = p \\times n" + }, + { + "bbox": [ + 55, + 279, + 296, + 446 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 279, + 296, + 446 + ], + "type": "inline_equation", + "content": "n \\geq 2" + }, + { + "bbox": [ + 55, + 279, + 296, + 446 + ], + "type": "text", + "content": ". Subsequently, we select high-resolution patches and further tokenize them using a patch size of " + }, + { + "bbox": [ + 55, + 279, + 296, + 446 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 55, + 279, + 296, + 446 + ], + "type": "text", + "content": " while the remaining patches are downscaled to " + }, + { + "bbox": [ + 55, + 279, + 296, + 446 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 55, + 279, + 296, + 446 + ], + "type": "text", + "content": ". The downscaling factor for both height and width is " + }, + { + "bbox": [ + 55, + 279, + 296, + 446 + ], + "type": "inline_equation", + "content": "f = 1/n" + }, + { + "bbox": [ + 55, + 279, + 296, + 446 + ], + "type": "text", + "content": ". Finally, we arrange all high-resolution patches at the beginning, followed by low-resolution ones to form the final input for the transformer encoder. This ordering is feasible due to the permutation invariance of ViTs [1], allowing us to avoid unnecessary implementation loops. To ensure a consistent input length (" + }, + { + "bbox": [ + 55, + 279, + 296, + 446 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 55, + 279, + 296, + 446 + ], + "type": "text", + "content": "), padding or random token dropping is employed as needed. Additional details on the 3-scale version are available in the Appendix 7." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 451, + 161, + 464 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 451, + 161, + 464 + ], + "spans": [ + { + "bbox": [ + 55, + 451, + 161, + 464 + ], + "type": "text", + "content": "3.2.2. Patch Importance" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 466, + 296, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 466, + 296, + 562 + ], + "spans": [ + { + "bbox": [ + 55, + 466, + 296, + 562 + ], + "type": "text", + "content": "Several strategies can be employed to determine which patches to retain at high resolution. The main goal is to preserve visually interesting areas and sharp details in high resolution. To identify these regions, we use frequency, gradient, entropy, and saliency maps. We compare them with a random selection of patches in each training epoch. More details on patch selection strategies, including illustrative examples, can be found in Figure 8 of Appendix 8." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 567, + 170, + 578 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 567, + 170, + 578 + ], + "spans": [ + { + "bbox": [ + 55, + 567, + 170, + 578 + ], + "type": "text", + "content": "3.2.3. Position Embedding" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 582, + 296, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 296, + 712 + ], + "type": "text", + "content": "ViTs lack inherent knowledge of spatial relationships between patches. Position embeddings address this by injecting learned position information into each patch embedding, allowing the self-attention mechanism to capture how different parts of the image relate. Replacing the standard position embeddings of ViTs with custom ones [36] can be expensive, requiring extensive pre-training on large datasets. It also introduces compatibility issues when used with different pre-trained ViTs. To address these issues, our approach adapts the pre-trained position embeddings from the standard ViT models." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 553, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 239 + ], + "type": "text", + "content": "For low-resolution images, we follow the standard ViT approach of resizing pre-trained position embeddings via bilinear interpolation. However, for high-resolution images, we create two position embeddings using bilinear interpolation based on the different patch sizes " + }, + { + "bbox": [ + 313, + 72, + 553, + 239 + ], + "type": "inline_equation", + "content": "(p" + }, + { + "bbox": [ + 313, + 72, + 553, + 239 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 72, + 553, + 239 + ], + "type": "inline_equation", + "content": "p^{\\prime})" + }, + { + "bbox": [ + 313, + 72, + 553, + 239 + ], + "type": "text", + "content": " employed during tokenization (Figure 2). One position embedding retains the spatial relationships between low-resolution patches, while the other captures the finer spatial relationships within the high-resolution patches. These position embeddings, high and low resolution, are compatible because they are derived from the same pre-trained embedding. Similar to standard ViTs, each token from the output of the patch encoding module is then summed with its corresponding position embedding." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 240, + 553, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 240, + 553, + 324 + ], + "spans": [ + { + "bbox": [ + 313, + 240, + 553, + 324 + ], + "type": "text", + "content": "Unlike existing methods [13, 36], our approach does not have a fixed image size, so position embeddings vary for each image. To avoid recalculating position embeddings in each epoch, we prepare them in the data loaders, keeping them frozen during fine-tuning. This aligns with the standard practice of freezing the initial layers (including position embeddings) and fine-tuning only the final layers." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 332, + 417, + 344 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 332, + 417, + 344 + ], + "spans": [ + { + "bbox": [ + 313, + 332, + 417, + 344 + ], + "type": "text", + "content": "3.2.4. Scale Embedding" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 347, + 553, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 347, + 553, + 420 + ], + "spans": [ + { + "bbox": [ + 313, + 347, + 553, + 420 + ], + "type": "text", + "content": "Our tokenization approach results in multiscale tokens. Following previous studies [21], to help ViT distinguish between patches from different scales and capture their relationships, we incorporate learnable scale embeddings. These scale embeddings are randomly initialized and added element-wise to all patch embeddings (Figure 2)." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 429, + 392, + 441 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 429, + 392, + 441 + ], + "spans": [ + { + "bbox": [ + 313, + 429, + 392, + 441 + ], + "type": "text", + "content": "3.3. Fine-tuning" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 447, + 555, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 447, + 555, + 626 + ], + "spans": [ + { + "bbox": [ + 313, + 447, + 555, + 626 + ], + "type": "text", + "content": "Our method requires no additional pre-training. We process images in their original resolution without resizing or random cropping during fine-tuning. We avoid normalizing the images as normalization can introduce distortions that affect the overall aesthetic/quality score (Table 9 in Appendix 9). Our data augmentation is limited to random horizontal flipping and random rotations (90, 180, 270 degrees) with a " + }, + { + "bbox": [ + 313, + 447, + 555, + 626 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 313, + 447, + 555, + 626 + ], + "type": "text", + "content": " probability, as they do not result in loss of composition or aspect ratio. While these methods might change the aesthetics for humans, they have been shown to improve the performance of ViTs (Appendix 10). During evaluation, data augmentation is completely omitted. We employ the L1 loss function for predicting the overall aesthetic/quality score and the Earth Mover's Distance (EMD) loss [37] for predicting the aesthetic score distribution." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 639, + 395, + 652 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 639, + 395, + 652 + ], + "spans": [ + { + "bbox": [ + 313, + 639, + 395, + 652 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 659, + 376, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 659, + 376, + 671 + ], + "spans": [ + { + "bbox": [ + 313, + 659, + 376, + 671 + ], + "type": "text", + "content": "4.1. Datasets" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 677, + 553, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 553, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 553, + 715 + ], + "type": "text", + "content": "We evaluate our models on various IAA datasets, including AVA [35], PARA [48], TAD66k [14], AADB [26], BAID [49], and two IQA datasets, SPAQ [10] and KonIQ10k [19]." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7818" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 69, + 294, + 164 + ], + "blocks": [ + { + "bbox": [ + 58, + 69, + 294, + 164 + ], + "lines": [ + { + "bbox": [ + 58, + 69, + 294, + 164 + ], + "spans": [ + { + "bbox": [ + 58, + 69, + 294, + 164 + ], + "type": "image", + "image_path": "de01c09f0bdd248741726e0cb4a83fbb33c7de8e365abd857bbfbf066e5376e8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 172, + 295, + 195 + ], + "lines": [ + { + "bbox": [ + 55, + 172, + 295, + 195 + ], + "spans": [ + { + "bbox": [ + 55, + 172, + 295, + 195 + ], + "type": "text", + "content": "Figure 3. An image with its aesthetic score distribution and mean aesthetic score, from the AVA dataset." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 206, + 295, + 302 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 206, + 295, + 302 + ], + "spans": [ + { + "bbox": [ + 54, + 206, + 295, + 302 + ], + "type": "text", + "content": "We only use the images and their corresponding ratings, excluding any additional information provided in the datasets. The AVA and PARA datasets associate each image with an aesthetic score distribution, while other datasets associate each image with an overall aesthetic/quality score (Figure 3). As the SPAQ and KonIQ10k datasets lack an official train-test split, we repeat the experiments for 5 different splits and report the median performance (as done in [41])." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 303, + 295, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 303, + 295, + 411 + ], + "spans": [ + { + "bbox": [ + 54, + 303, + 295, + 411 + ], + "type": "text", + "content": "The AVA dataset remains prominent due to its large number of images, allowing for effective training of large models. Research utilizing smaller datasets is relatively limited, and most existing approaches leverage the additional information provided in these datasets. However, considering various datasets with different resolution ranges allows us to validate the efficiency of our approach in various scenarios. For more details on the resolution distribution of each dataset, see Figure 13 in Appendix 11." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 419, + 187, + 431 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 419, + 187, + 431 + ], + "spans": [ + { + "bbox": [ + 55, + 419, + 187, + 431 + ], + "type": "text", + "content": "4.2. Implementation details" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 437, + 295, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 437, + 295, + 497 + ], + "spans": [ + { + "bbox": [ + 54, + 437, + 295, + 497 + ], + "type": "text", + "content": "Saliency maps for patch selection are generated using the SAM-HQ model [24] with the text prompt of \"salient object\" (Appendix 8). For gradient, frequency, and entropy calculations, we use the Sobel filter, fast Fourier transform, and Shannon entropy, respectively." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 498, + 295, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 498, + 295, + 617 + ], + "spans": [ + { + "bbox": [ + 54, + 498, + 295, + 617 + ], + "type": "text", + "content": "In all experiments, we fine-tune the entire ViT models, while keeping their position embeddings frozen. Also, we use scale embeddings alongside position embeddings to enable Charm to capture relationships between tokens both within and across scales (see Appendix 12). The patch size " + }, + { + "bbox": [ + 54, + 498, + 295, + 617 + ], + "type": "inline_equation", + "content": "(p)" + }, + { + "bbox": [ + 54, + 498, + 295, + 617 + ], + "type": "text", + "content": " and patch stride are chosen to match the filter size and stride of the patch encoding modules within these models (16 in ViT-small and 14 in Dinov2-small and Dinov2-large), ensuring compatibility and replicating the models' internal tokenization process." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 617, + 295, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 295, + 677 + ], + "type": "text", + "content": "The maximum number of tokens " + }, + { + "bbox": [ + 55, + 617, + 295, + 677 + ], + "type": "inline_equation", + "content": "(l)" + }, + { + "bbox": [ + 55, + 617, + 295, + 677 + ], + "type": "text", + "content": " is set to 512 for AVA, 768 for TAD66k, and 1024 for other datasets. These values are chosen based on the image resolution in the datasets (Appendix 13). For all experiments, we use a factor size of 0.5." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "The AdamW optimizer [33] is employed for fine-tuning. We use an initial learning rate of 1e-5 for all datasets except BAID (1e-6). We gradually reduce the learning rate dur" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 317, + 69, + 553, + 340 + ], + "blocks": [ + { + "bbox": [ + 317, + 69, + 553, + 340 + ], + "lines": [ + { + "bbox": [ + 317, + 69, + 553, + 340 + ], + "spans": [ + { + "bbox": [ + 317, + 69, + 553, + 340 + ], + "type": "table", + "html": "
DatasetCharmPLCCSRCCACC
IAA
AVA-0.7100.7060.802
✓0.779(↑ 6.9%)0.777(↑ 7.1%)0.826(↑ 2.4%)
AADB-0.6950.6820.754
✓0.767(↑ 7.2%)0.754(↑ 7.2%)0.767(↑ 1.3%)
TAD66k-0.4290.4010.646
✓0.488(↑ 5.9%)0.458(↑ 5.7%)0.794 (↑ 14.8%)
PARA-0.9040.8550.863
✓0.938(↑ 3.4%)0.905(↑ 5%)0.892(↑ 2.9%)
BAID-0.4280.3420.750
✓0.439(↑ 1.1%)0.368(↑ 2.6%)0.763(↑ 1.3%)
IQA
SPAQ-0.9110.9070.907
✓0.919(↑ 0.8%)0.915(↑ 0.8%)0.917(↑ 1.0%)
KonIQ10k-0.8960.8680.938
✓0.944(↑ 4.8%)0.93(↑ 6.2%)0.954(↑ 1.6%)
", + "image_path": "8c86b8054f7412a05cdd9bac202b68e8275ce6ebc568d074c8e8cd3310b42eee.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 348, + 555, + 382 + ], + "lines": [ + { + "bbox": [ + 313, + 348, + 555, + 382 + ], + "spans": [ + { + "bbox": [ + 313, + 348, + 555, + 382 + ], + "type": "text", + "content": "Table 1. Performance improvement across different IAA and IQA datasets by replacing the standard tokenization of Dinov2-small with Charm (with a random patch selection strategy)." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 396, + 555, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 396, + 555, + 492 + ], + "spans": [ + { + "bbox": [ + 313, + 396, + 555, + 492 + ], + "type": "text", + "content": "ing training using the cosine annealing learning rate scheduler. The batch size is 64 for AVA, 32 for TAD66k, and 16 for the remaining datasets. Early stopping is also implemented to prevent overfitting. Fine-tuning is conducted on a single NVIDIA RTX A4500 GPU. Model performance is evaluated using three key metrics: Spearman rank-ordered correlation coefficient (SRCC), Pearson linear correlation coefficient (PLCC), and classification accuracy (ACC)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 502, + 495, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 502, + 495, + 514 + ], + "spans": [ + { + "bbox": [ + 313, + 502, + 495, + 514 + ], + "type": "text", + "content": "4.3. Comparing with existing methods" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 521, + 555, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 555, + 616 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 555, + 616 + ], + "type": "text", + "content": "We benchmark Charm with various ViT backbones on different IAA and IQA datasets. We first fine-tune models on the datasets to create a baseline. We then replace their original patch encoding modules with two-scale Charm tokenization and fine-tune the entire models. We also considered a three-scale version of Charm but did not observe additional performance gains (Appendix 7). Therefore, we focus on the two-scale version throughout the paper." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "type": "text", + "content": "The results (Table 1 and Table 13 in Appendix 14) demonstrate that by simply replacing the patch encoding module of existing ViT models with Charm, we can achieve significant performance improvements, up to " + }, + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "type": "inline_equation", + "content": "7.5\\%" + }, + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "type": "text", + "content": " in PLCC, " + }, + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "type": "inline_equation", + "content": "8.1\\%" + }, + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "type": "text", + "content": " in SRCC, and " + }, + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "type": "inline_equation", + "content": "14.8\\%" + }, + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "type": "text", + "content": " in classification accuracy across various datasets. Among datasets, PARA and SPAQ datasets include extra high-resolution images (e.g., 2k by 3k pixels). Our experiments (Appendix 15) reveal" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7819" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 153, + 69, + 457, + 194 + ], + "blocks": [ + { + "bbox": [ + 153, + 69, + 457, + 194 + ], + "lines": [ + { + "bbox": [ + 153, + 69, + 457, + 194 + ], + "spans": [ + { + "bbox": [ + 153, + 69, + 457, + 194 + ], + "type": "table", + "html": "
ModelARHRMSPLCCSRCCACC
Ghosal et al.[11]✓--0.7640.762-
Chen et al. [6]✓--0.6710.6490.832
Dinov2-Small (+ Padding) *✓✓-0.7090.7030.801
MUSIQ [21]✓-✓0.7380.7260.815
FlexiViT [2]*--✓0.7370.7350.812
Swin [32]*--✓0.7480.7510.816
Dinov2-Small (+ Muller [43]) *-✓-0.6820.6750.794
ViT-small (+ Charm) *✓✓✓0.7620.7600.827
Dinov2-small (+ Charm) *✓✓✓0.7790.7770.826
Dinov2-large (+ Charm) *✓✓✓0.7830.7810.828
", + "image_path": "0a48bbd50870ed4473c3bdc5e0f4ee7efb2285272ae3e4890dcd02421f79454e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 202, + 555, + 237 + ], + "lines": [ + { + "bbox": [ + 55, + 202, + 555, + 237 + ], + "spans": [ + { + "bbox": [ + 55, + 202, + 555, + 237 + ], + "type": "text", + "content": "Table 2. Results on the AVA dataset. Bold and underlined numbers indicate the best and the second-best results, respectively. AR, HR, and MS represent preserving aspect ratio, high-resolution details, and multi-scale information, respectively. * indicates models that we have fine-tuned on the AVA dataset. Appendix 14 shows the performance of the standard ViT-small, Dinov2-small, and Dinov2-large." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 257, + 294, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 257, + 294, + 317 + ], + "spans": [ + { + "bbox": [ + 54, + 257, + 294, + 317 + ], + "type": "text", + "content": "a threshold for performance improvement by preserving high-resolution information. Beyond this threshold, further performance gains are limited. Therefore, to decrease computational costs, we downscale images in PARA and SPAQ to a maximum edge of 1024." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 317, + 295, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 317, + 295, + 449 + ], + "spans": [ + { + "bbox": [ + 54, + 317, + 295, + 449 + ], + "type": "text", + "content": "As shown in Table 2, our approach outperforms existing methods in terms of preserving resolution, aspect ratio, and multiscale information. We compare Charm with the simple approach of adding padding to ensure consistent input shapes within each batch while preserving the aspect ratio. However, this padding proves detrimental to IAA and significantly increases computational costs (Appendix 16). Also, we observe that Muller [43], a recent approach focusing on preserving high-resolution information, can introduce noise and distortions to the image, negatively affecting aesthetics (Appendix 16)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 449, + 295, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 449, + 295, + 641 + ], + "spans": [ + { + "bbox": [ + 55, + 449, + 295, + 641 + ], + "type": "text", + "content": "Recent studies have demonstrated the effectiveness of incorporating additional modalities, such as attributes ([5, 20, 28]) or text captions ([29, 30, 39, 50]), to improve IAA model performance (Table 15 in Appendix 17). Other studies have concentrated on other research directions [15, 40], like addressing the long tails of distributions in IAA datasets [31]. While the main contributions of these studies are orthogonal to Charm's, integrating Charm with them further enhances their performance. For example, incorporating Charm into ELTA [31] with a Dinov2-large backbone improves its PLCC, SRCC, and ACC on the AVA dataset from 0.742, 0.742, and 0.811 to 0.787, 0.786, and 0.829, respectively. Also, our approach demonstrates strong performance in IQA, with a slight performance gap of less than " + }, + { + "bbox": [ + 55, + 449, + 295, + 641 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 55, + 449, + 295, + 641 + ], + "type": "text", + "content": " in SPAQ and " + }, + { + "bbox": [ + 55, + 449, + 295, + 641 + ], + "type": "inline_equation", + "content": "0.4\\%" + }, + { + "bbox": [ + 55, + 449, + 295, + 641 + ], + "type": "text", + "content": " in KonIQ10k compared to state-of-the-art methods [4, 21, 41] (Table 16 in Appendix 17)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "type": "text", + "content": "Model size is another crucial factor, as larger models often excel in complex tasks like IAA. In this paper, we demonstrate that Charm is highly effective even with lightweight ViTs, which can be used for applications with limited computational resources. As shown in Figure 4, Dinov2-small + Charm achieves comparable performance" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 257, + 555, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 257, + 555, + 341 + ], + "spans": [ + { + "bbox": [ + 313, + 257, + 555, + 341 + ], + "type": "text", + "content": "to state-of-the-art models while being significantly smaller. Nevertheless, Charm can serve as a plug-in for larger, more expensive models if higher performance is desired and computational power allows. For instance, plugging Charm into ViT-base instead of Dinov2-small results in PLCC of 0.515, SRCC of 0.467, and ACC of 0.779 after only 9 epochs on the BAID dataset." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 353, + 406, + 366 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 353, + 406, + 366 + ], + "spans": [ + { + "bbox": [ + 313, + 353, + 406, + 366 + ], + "type": "text", + "content": "4.4. Ablation study" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 371, + 554, + 396 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 371, + 554, + 396 + ], + "spans": [ + { + "bbox": [ + 313, + 371, + 554, + 396 + ], + "type": "text", + "content": "4.4.1. The importance of preserving aspect ratio and high-resolution details" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 399, + 554, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 399, + 554, + 519 + ], + "spans": [ + { + "bbox": [ + 313, + 399, + 554, + 519 + ], + "type": "text", + "content": "Current approaches usually distort images by resizing them to a square shape, disregarding the significance of maintaining their original size and aspect ratio. We argue that such preprocessing can be detrimental to IAA. To validate this claim, we compare the performance of Dinov2-small using different tokenization approaches. Table 3 presents the results of these experiments. Except for the standard tokenization (described in Appendix 9), all other settings involve multiscale inputs and maintain composition by avoiding image cropping." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 520, + 554, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 520, + 554, + 616 + ], + "spans": [ + { + "bbox": [ + 313, + 520, + 554, + 616 + ], + "type": "text", + "content": "We initially downscale images to " + }, + { + "bbox": [ + 313, + 520, + 554, + 616 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 313, + 520, + 554, + 616 + ], + "type": "text", + "content": " and then preserve some areas at that resolution while downscaling others by a factor of 0.5 (MS in Table 3). While this approach has been effective in image classification ([13] and [36]), it decreases performance in our IAA task. This is because the initial downscaling to a fixed size results in significant information loss critical for IAA, and further downscaling worsens the issue." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "text", + "content": "We then experiment with keeping the image in its original resolution while changing the aspect ratio by resizing the image based on the maximum edge to create a square image (HR + MS in Table 3). This results in a nearly " + }, + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "inline_equation", + "content": "3\\%" + }, + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "text", + "content": " increase in PLCC and SRCC compared to the standard approach. Finally, we use Charm, which preserves both aspect ratio and high-resolution details, resulting in the best performance. This analysis highlights the importance of preserv" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "7820" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 70, + 294, + 219 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 294, + 219 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 294, + 219 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 294, + 219 + ], + "type": "image", + "image_path": "5921f7f7008a27a222885dcd325854054845763dd141457a698dc98513d1ba9a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 228, + 296, + 317 + ], + "lines": [ + { + "bbox": [ + 55, + 228, + 296, + 317 + ], + "spans": [ + { + "bbox": [ + 55, + 228, + 296, + 317 + ], + "type": "text", + "content": "Figure 4. Comparison of state-of-the-art IAA models' performance on the AVA dataset vs. their number of parameters. Our approach achieves comparable performance without using additional modalities and with significantly fewer parameters. To account for the significant differences in the number of parameters, we use a logarithmic scale on the x-axis. For a detailed comparison of Dinov2-small + Charm with state-of-the-art models on all datasets, see Appendix 17." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 73, + 327, + 279, + 387 + ], + "blocks": [ + { + "bbox": [ + 73, + 327, + 279, + 387 + ], + "lines": [ + { + "bbox": [ + 73, + 327, + 279, + 387 + ], + "spans": [ + { + "bbox": [ + 73, + 327, + 279, + 387 + ], + "type": "table", + "html": "
TokenizationPLCCSRCCACC
Standard0.7340.7320.808
MS0.7100.7050.801
HR + MS0.7630.7600.820
Charm (AR + HR + MS)0.7790.7770.826
", + "image_path": "4c5af7659f6068b611642bdbdf6c3961a80e259980a332bb88891b338728651f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 452, + 288, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 452, + 288, + 464 + ], + "spans": [ + { + "bbox": [ + 55, + 452, + 288, + 464 + ], + "type": "text", + "content": "ing both aspect ratio and original resolution in IAA tasks." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 464, + 295, + 500 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 464, + 295, + 500 + ], + "spans": [ + { + "bbox": [ + 55, + 464, + 295, + 500 + ], + "type": "text", + "content": "We also examined the impact of different numbers of tokens across tokenization approaches in Table 3 but found no differences in the results (see Appendix 18)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 506, + 255, + 518 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 506, + 255, + 518 + ], + "spans": [ + { + "bbox": [ + 55, + 506, + 255, + 518 + ], + "type": "text", + "content": "4.4.2. Comparison of patch selection strategies" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 521, + 295, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 295, + 593 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 295, + 593 + ], + "type": "text", + "content": "To evaluate the importance of patches, we fine-tune Dinov2-small on the AADB dataset with different patch selection strategies. Table 4 demonstrates that all strategies outperform the standard ViT approach [9]. This suggests that factors beyond preserving high-resolution details also contribute to performance improvements (Section 4.4.1)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 594, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 594, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 594, + 295, + 714 + ], + "type": "text", + "content": "Among all methods, the random selection of patches in each epoch stands out, likely for two reasons. First, ViTs are prone to overfitting, especially when fine-tuned on small datasets, a common challenge in IAA. Random patch selection in each epoch addresses this by introducing more variety to the ViT input and acting as a form of data augmentation. Second, random sampling can better capture the subjective and often unpredictable nature of beauty, providing a more representative view of the entire image. Following random selection, frequency, gradient, and entropy-based" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 350, + 69, + 520, + 150 + ], + "blocks": [ + { + "bbox": [ + 55, + 394, + 295, + 439 + ], + "lines": [ + { + "bbox": [ + 55, + 394, + 295, + 439 + ], + "spans": [ + { + "bbox": [ + 55, + 394, + 295, + 439 + ], + "type": "text", + "content": "Table 3. Performance of Dinov2-small on the AVA dataset with four different tokenization strategies. MS, HR, and AR represent preserving multiscale information, high-resolution details, and aspect ratio, respectively." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 350, + 69, + 520, + 150 + ], + "lines": [ + { + "bbox": [ + 350, + 69, + 520, + 150 + ], + "spans": [ + { + "bbox": [ + 350, + 69, + 520, + 150 + ], + "type": "table", + "html": "
Patch selectionPLCCSRCCACC
Random0.7670.7540.767
Entropy0.7260.7140.761
Frequency0.7560.7470.761
Gradient0.7340.7210.766
Saliency0.7510.7380.756
Standard0.6950.6820.754
", + "image_path": "47fa52318501d89690f165b66dcf8bc1c43f0f8eb5c12bacec3b212fee325d80.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 158, + 554, + 192 + ], + "lines": [ + { + "bbox": [ + 313, + 158, + 554, + 192 + ], + "spans": [ + { + "bbox": [ + 313, + 158, + 554, + 192 + ], + "type": "text", + "content": "Table 4. Performance of Dinov2-small on the AADB dataset using Charm with different patch selection strategies. Bold and underlined values highlight the best and second-best results." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 315, + 203, + 552, + 305 + ], + "blocks": [ + { + "bbox": [ + 315, + 203, + 552, + 305 + ], + "lines": [ + { + "bbox": [ + 315, + 203, + 552, + 305 + ], + "spans": [ + { + "bbox": [ + 315, + 203, + 552, + 305 + ], + "type": "image", + "image_path": "11376fa752a04693d6692ea20f5fbbad7a86a080f479408b30da4065fc7d45cc.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 312, + 554, + 357 + ], + "lines": [ + { + "bbox": [ + 313, + 312, + 554, + 357 + ], + "spans": [ + { + "bbox": [ + 313, + 312, + 554, + 357 + ], + "type": "text", + "content": "Figure 5. The epoch at which each model achieves its highest validation performance across different datasets. Charm generally leads to faster convergence. The average number of epochs over 5 runs is reported for SPAQ and KonIQ10k." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 369, + 554, + 404 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 369, + 554, + 404 + ], + "spans": [ + { + "bbox": [ + 313, + 369, + 554, + 404 + ], + "type": "text", + "content": "methods stand out. These methods still have a certain level of randomness, as their fully deterministic version results in highly correlated selected patches (Appendix 8)." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 404, + 554, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 404, + 554, + 477 + ], + "spans": [ + { + "bbox": [ + 313, + 404, + 554, + 477 + ], + "type": "text", + "content": "ViTs, originally trained for object classification, have an inherent bias towards salient objects. However, IAA requires attention to both foreground and background [15], which explains why the saliency-based approach performs less effectively. For a more comprehensive analysis of patch selection strategies please refer to Appendix 8." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 484, + 523, + 496 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 484, + 523, + 496 + ], + "spans": [ + { + "bbox": [ + 313, + 484, + 523, + 496 + ], + "type": "text", + "content": "4.4.3. The effect of Charm on convergence speed" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 499, + 554, + 631 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 499, + 554, + 631 + ], + "spans": [ + { + "bbox": [ + 313, + 499, + 554, + 631 + ], + "type": "text", + "content": "We fine-tune all models until convergence, meaning their performance on the validation set stabilizes. Figure 5 shows the epoch at which each model achieves its best validation performance. As shown in this figure, Charm consistently accelerates the convergence of Dinov2-small across different datasets. Notably, the most significant reduction (50%) in training epochs occurs on AADB, our smallest dataset, likely due to Charm's ability to mitigate overfitting through random patch selection (see Appendix 8). On our largest datasets, AVA and TaD66k, both standard and Charm tokenization converge at similar epochs." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 638, + 430, + 650 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 638, + 430, + 650 + ], + "spans": [ + { + "bbox": [ + 313, + 638, + 430, + 650 + ], + "type": "text", + "content": "4.4.4. Computational costs" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "text", + "content": "Figure 6 compares the runtime of Charm using different patch selection strategies with the standard approach on a 12th Gen Intel(R) Core(TM) i9-12900K CPU. Table 3 demonstrates that multiscale inputs can negatively impact performance for very small images. In this case, we use a" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "7821" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 113, + 69, + 497, + 128 + ], + "blocks": [ + { + "bbox": [ + 113, + 69, + 497, + 128 + ], + "lines": [ + { + "bbox": [ + 113, + 69, + 497, + 128 + ], + "spans": [ + { + "bbox": [ + 113, + 69, + 497, + 128 + ], + "type": "table", + "html": "
ModelInput sizeCharm#tokensmsGMACsMB
Dinov2224 x 224-2565.76.11202.9
-small-207032.884.012091.8
640 x 640✓2-scale:5127.3 (↓ 77.7%)13.46 (↓ 84%)346.0 (↓ 83.5%)
✓3-scale:7009.3 (↓ 71.6%)19.60 (↓ 76.7%)494.3 (↓ 76.4%)
", + "image_path": "507c4a082204bde3c10408a461700aa528a3708f8eb28c929075bbd86e4f19b1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 137, + 555, + 182 + ], + "lines": [ + { + "bbox": [ + 55, + 137, + 555, + 182 + ], + "spans": [ + { + "bbox": [ + 55, + 137, + 555, + 182 + ], + "type": "text", + "content": "Table 5. Dinov2-small inference cost breakdown for processing one single image: number of tokens (#tokens) based on varying input sizes, runtime in milliseconds (ms), Giga multiply accumulation (GMACs), and GPU memory in Megabytes (MB). Bold and underlined values highlight the most and second-most computationally efficient configurations. Percentages indicate the reduction in computational cost compared to processing the image in its original size. Similar results for ViT-small are presented in Table 17 in the Appendix." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 70, + 201, + 282, + 353 + ], + "blocks": [ + { + "bbox": [ + 70, + 201, + 282, + 353 + ], + "lines": [ + { + "bbox": [ + 70, + 201, + 282, + 353 + ], + "spans": [ + { + "bbox": [ + 70, + 201, + 282, + 353 + ], + "type": "image", + "image_path": "64a2e6d2dd79a19fbd594a407610eff45754782f191d9deb597270e756b15a3e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 360, + 295, + 393 + ], + "lines": [ + { + "bbox": [ + 55, + 360, + 295, + 393 + ], + "spans": [ + { + "bbox": [ + 55, + 360, + 295, + 393 + ], + "type": "text", + "content": "Figure 6. Comparison of Charm runtime using different patch selection strategies with the standard approach (Appendix 9). The x-axis shows the total number of pixels." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 411, + 295, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 411, + 295, + 531 + ], + "spans": [ + { + "bbox": [ + 54, + 411, + 295, + 531 + ], + "type": "text", + "content": "single-scale tokenization approach (Section 3.2.1). Figure 6 illustrates the increase in the computational cost of Charm with larger image sizes. To reduce the overhead of saliency mask generation (16 seconds per image), we employ precomputed masks. The frequency-based method provides a cost-effective alternative to on-the-fly saliency detection, identifying regions comparable to or even superior to those selected by SAM-HQ. Random patch selection emerges as the most effective strategy, offering the best performance with the lowest computational cost." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 534, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 295, + 715 + ], + "type": "text", + "content": "The computational cost of ViTs is influenced by the number of input tokens. The standard method involves downscaling and cropping images to " + }, + { + "bbox": [ + 55, + 534, + 295, + 715 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 55, + 534, + 295, + 715 + ], + "type": "text", + "content": ", resulting in a very low number of tokens and lower costs. However, processing images at their original size can significantly increase computational costs. Our approach not only improves performance but also significantly reduces the computational costs of processing images in the original resolution (Table 5). We observe a decrease of over " + }, + { + "bbox": [ + 55, + 534, + 295, + 715 + ], + "type": "inline_equation", + "content": "84\\%" + }, + { + "bbox": [ + 55, + 534, + 295, + 715 + ], + "type": "text", + "content": " in GMACs, " + }, + { + "bbox": [ + 55, + 534, + 295, + 715 + ], + "type": "inline_equation", + "content": "77\\%" + }, + { + "bbox": [ + 55, + 534, + 295, + 715 + ], + "type": "text", + "content": " in runtime, " + }, + { + "bbox": [ + 55, + 534, + 295, + 715 + ], + "type": "inline_equation", + "content": "83\\%" + }, + { + "bbox": [ + 55, + 534, + 295, + 715 + ], + "type": "text", + "content": " in GPU memory consumption, and " + }, + { + "bbox": [ + 55, + 534, + 295, + 715 + ], + "type": "inline_equation", + "content": "75\\%" + }, + { + "bbox": [ + 55, + 534, + 295, + 715 + ], + "type": "text", + "content": " in the number of tokens for Dinov2-small when processing one " + }, + { + "bbox": [ + 55, + 534, + 295, + 715 + ], + "type": "inline_equation", + "content": "640 \\times 640" + }, + { + "bbox": [ + 55, + 534, + 295, + 715 + ], + "type": "text", + "content": " image using a 2-scale Charm. Even the 3-scale version, which is more costly, has significantly lower costs compared to processing images at their original size. While Charm may incur slightly" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 202, + 555, + 238 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 202, + 555, + 238 + ], + "spans": [ + { + "bbox": [ + 313, + 202, + 555, + 238 + ], + "type": "text", + "content": "higher computational costs compared to the standard approach of downscaling and cropping, the substantial performance gains make it a worthwhile trade-off." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 246, + 441, + 258 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 246, + 441, + 258 + ], + "spans": [ + { + "bbox": [ + 313, + 246, + 441, + 258 + ], + "type": "text", + "content": "4.4.5. Downscaling factor " + }, + { + "bbox": [ + 313, + 246, + 441, + 258 + ], + "type": "inline_equation", + "content": "(f)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 261, + 555, + 381 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 261, + 555, + 381 + ], + "spans": [ + { + "bbox": [ + 313, + 261, + 555, + 381 + ], + "type": "text", + "content": "Our method initially tokenizes images with an initial patch size " + }, + { + "bbox": [ + 313, + 261, + 555, + 381 + ], + "type": "inline_equation", + "content": "(p')" + }, + { + "bbox": [ + 313, + 261, + 555, + 381 + ], + "type": "text", + "content": ". Subsequently, selected patches are further tokenized using a smaller patch size " + }, + { + "bbox": [ + 313, + 261, + 555, + 381 + ], + "type": "inline_equation", + "content": "(p)" + }, + { + "bbox": [ + 313, + 261, + 555, + 381 + ], + "type": "text", + "content": ", while the remaining patches are downscaled to " + }, + { + "bbox": [ + 313, + 261, + 555, + 381 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 313, + 261, + 555, + 381 + ], + "type": "text", + "content": ". We experiment with different initial patch sizes " + }, + { + "bbox": [ + 313, + 261, + 555, + 381 + ], + "type": "inline_equation", + "content": "p' = (28, 42, 56)" + }, + { + "bbox": [ + 313, + 261, + 555, + 381 + ], + "type": "text", + "content": ", corresponding to downscaling factors " + }, + { + "bbox": [ + 313, + 261, + 555, + 381 + ], + "type": "inline_equation", + "content": "(f)" + }, + { + "bbox": [ + 313, + 261, + 555, + 381 + ], + "type": "text", + "content": " of 0.5, 0.66, and 0.75. The best performance is achieved with a factor of 0.5 (Appendix 19). A larger factor negatively impacts performance, as it leads to more significant downscaling of unselected patches, resulting in information loss for IAA." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 392, + 388, + 405 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 392, + 388, + 405 + ], + "spans": [ + { + "bbox": [ + 313, + 392, + 388, + 405 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 413, + 555, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 413, + 555, + 593 + ], + "spans": [ + { + "bbox": [ + 313, + 413, + 555, + 593 + ], + "type": "text", + "content": "Processing images in their original resolution can be computationally expensive, requiring significant memory and processing power. Also, it is challenging to process images of varying aspect ratios and resolutions due to batch processing constraints. Charm effectively balances computational cost with the preservation of composition, high-resolution details, aspect ratio, and multi-scale information, which are all crucial for achieving optimal performance in IAA. By introducing variability into the input, our method enhances the generalizability of ViTs in IAA. Charm is compatible with any ViT architecture (except Swin [32], see Appendix 20) and does not require additional pre-training of ViTs. Our method consistently outperforms existing approaches that address similar research questions, achieving promising results across diverse IAA and IQA datasets." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 594, + 556, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 556, + 666 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 556, + 666 + ], + "type": "text", + "content": "While our approach effectively preserves critical aesthetic-related factors, there's a potential for information loss due to selective downscaling. To further improve performance, future work could explore more sophisticated patch selection strategies, incorporate other modalities, and integrate our method with existing IAA-specific networks." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 666, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 666, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 666, + 555, + 713 + ], + "type": "text", + "content": "Acknowledgements: Funded by the European Union (ERC AdG, GRAPPA, 101053925, awarded to Johan Wagemans) and the Research Foundation-Flanders (FWO, 1159925N, awarded to Fatemeh Behrad)." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7822" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "type": "text", + "content": "[1] Irwan Bello, Barret Zoph, Ashish Vaswani, Jonathon Shlens, and Quoc V Le. Attention augmented convolutional networks. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3286-3295, 2019. 3, 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 137, + 296, + 212 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 137, + 296, + 212 + ], + "spans": [ + { + "bbox": [ + 61, + 137, + 296, + 212 + ], + "type": "text", + "content": "[2] Lucas Beyer, Pavel Izmailov, Alexander Kolesnikov, Mathilde Caron, Simon Kornblith, Xiaohua Zhai, Matthias Minderer, Michael Tschannen, Ibrahim Alabdulmohsin, and Filip Pavetic. Flexivit: One model for all patch sizes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14496-14506, 2023. 2, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 214, + 295, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 214, + 295, + 257 + ], + "spans": [ + { + "bbox": [ + 62, + 214, + 295, + 257 + ], + "type": "text", + "content": "[3] Daniel Bolya, Cheng-Yang Fu, Xiaoliang Dai, Peizhao Zhang, Christoph Feichtenhofer, and Judy Hoffman. Token merging: Your vit but faster. arXiv preprint arXiv:2210.09461, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 258, + 295, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 258, + 295, + 312 + ], + "spans": [ + { + "bbox": [ + 62, + 258, + 295, + 312 + ], + "type": "text", + "content": "[4] Chaofeng Chen, Jiadi Mo, Jingwen Hou, Haoning Wu, Liang Liao, Wenxiu Sun, Qiong Yan, and Weisi Lin. Topiq: A top-down approach from semantics to distortions for image quality assessment. IEEE Transactions on Image Processing, 2024. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 314, + 295, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 314, + 295, + 357 + ], + "spans": [ + { + "bbox": [ + 62, + 314, + 295, + 357 + ], + "type": "text", + "content": "[5] Hangwei Chen, Feng Shao, Baoyang Mu, and Qiuping Jiang. Image aesthetics assessment with emotion-aware multibranch network. IEEE Transactions on Instrumentation and Measurement, 73:1-15, 2024. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 358, + 295, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 358, + 295, + 413 + ], + "spans": [ + { + "bbox": [ + 62, + 358, + 295, + 413 + ], + "type": "text", + "content": "[6] Qiuyu Chen, Wei Zhang, Ning Zhou, Peng Lei, Yi Xu, Yu Zheng, and Jianping Fan. Adaptive fractional dilated convolution network for image aesthetics assessment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14114-14123, 2020. 2, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 415, + 295, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 415, + 295, + 479 + ], + "spans": [ + { + "bbox": [ + 62, + 415, + 295, + 479 + ], + "type": "text", + "content": "[7] Bowen Cheng, Bin Xiao, Jingdong Wang, Honghui Shi, Thomas S Huang, and Lei Zhang. Higherhrnet: Scale-aware representation learning for bottom-up human pose estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5386-5395, 2020. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 480, + 295, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 480, + 295, + 545 + ], + "spans": [ + { + "bbox": [ + 62, + 480, + 295, + 545 + ], + "type": "text", + "content": "[8] Mostafa Dehghani, Basil Mustafa, Josip Djolonga, Jonathan Heek, Matthias Minderer, Mathilde Caron, Andreas Steiner, Joan Puigcerver, Robert Geirhos, Ibrahim M Alabdulmohsin, et al. Patch n'pack: Navit, a vision transformer for any aspect ratio and resolution. Advances in Neural Information Processing Systems, 36, 2024. 2, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 547, + 295, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 547, + 295, + 612 + ], + "spans": [ + { + "bbox": [ + 62, + 547, + 295, + 612 + ], + "type": "text", + "content": "[9] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 1, 3, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 614, + 295, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 614, + 295, + 667 + ], + "spans": [ + { + "bbox": [ + 57, + 614, + 295, + 667 + ], + "type": "text", + "content": "[10] Yuming Fang, Hanwei Zhu, Yan Zeng, Kede Ma, and Zhou Wang. Perceptual quality assessment of smartphone photography. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3677-3686, 2020. 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 670, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 295, + 712 + ], + "type": "text", + "content": "[11] Koustav Ghosal and Aljosa Smolic. Image aesthetics assessment using graph attention network. In 2022 26th International Conference on Pattern Recognition (ICPR), pages 3160-3167. IEEE, 2022. 2, 6" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 57, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 105 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 105 + ], + "type": "text", + "content": "[12] Charles D Gilbert and Wu Li. Top-down influences on visual processing. Nature Reviews Neuroscience, 14(5):350-363, 2013. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 317, + 107, + 553, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 107, + 553, + 162 + ], + "spans": [ + { + "bbox": [ + 317, + 107, + 553, + 162 + ], + "type": "text", + "content": "[13] Jakob Drachmann Havtorn, Amélie Royer, Tijmen Blankevoort, and Babak Ehteshami Bejnordi. Msvit: Dynamic mixed-scale tokenization for vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 838-848, 2023. 2, 4, 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 164, + 553, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 164, + 553, + 206 + ], + "spans": [ + { + "bbox": [ + 317, + 164, + 553, + 206 + ], + "type": "text", + "content": "[14] Shuai He, Yongchang Zhang, Rui Xie, Dongxiang Jiang, and Anlong Ming. Rethinking image aesthetics assessment: Models, datasets and benchmarks. In *IJCAI*, pages 942–948, 2022. 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 209, + 553, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 209, + 553, + 252 + ], + "spans": [ + { + "bbox": [ + 316, + 209, + 553, + 252 + ], + "type": "text", + "content": "[15] Shuai He, Anlong Ming, Shuntian Zheng, Haobin Zhong, and Huadong Ma. Eat: An enhancer for aesthetics-oriented transformers. In Proceedings of the 31st ACM International Conference on Multimedia, pages 1023-1032, 2023. 6, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 254, + 553, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 254, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 316, + 254, + 553, + 285 + ], + "type": "text", + "content": "[16] Simon Hentschel, Konstantin Kobs, and Andreas Hotho. Clip knows image aesthetics. Frontiers in Artificial Intelligence, 5:976235, 2022. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 287, + 553, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 287, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 316, + 287, + 553, + 319 + ], + "type": "text", + "content": "[17] Shaul Hochstein and Merav Ahissar. View from the top: Hierarchies and reverse hierarchies in the visual system. Neuron, 36(5):791-804, 2002. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 322, + 553, + 375 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 322, + 553, + 375 + ], + "spans": [ + { + "bbox": [ + 317, + 322, + 553, + 375 + ], + "type": "text", + "content": "[18] Vlad Hosu, Bastian Goldlucke, and Dietmar Saupe. Effective aesthetics prediction with multi-level spatially pooled features. In proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9375-9383, 2019. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 377, + 553, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 377, + 553, + 421 + ], + "spans": [ + { + "bbox": [ + 317, + 377, + 553, + 421 + ], + "type": "text", + "content": "[19] Vlad Hosu, Hanhe Lin, Tamas Sziranyi, and Dietmar Saupe. Koniq-10k: An ecologically valid database for deep learning of blind image quality assessment. IEEE Transactions on Image Processing, 29:4041–4056, 2020. 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 422, + 553, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 422, + 553, + 465 + ], + "spans": [ + { + "bbox": [ + 317, + 422, + 553, + 465 + ], + "type": "text", + "content": "[20] Yipo Huang, Leida Li, Pengfei Chen, Jinjian Wu, Yuzhe Yang, Yaqian Li, and Guangming Shi. Coarse-to-fine image aesthetics assessment with dynamic attribute selection. IEEE Transactions on Multimedia, 26:9316-9329, 2024. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 468, + 553, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 468, + 553, + 511 + ], + "spans": [ + { + "bbox": [ + 317, + 468, + 553, + 511 + ], + "type": "text", + "content": "[21] Junjie Ke, Qifei Wang, Yilin Wang, Peyman Milanfar, and Feng Yang. Musiq: Multi-scale image quality transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5148-5157, 2021. 1, 2, 3, 4, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 514, + 553, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 514, + 553, + 567 + ], + "spans": [ + { + "bbox": [ + 317, + 514, + 553, + 567 + ], + "type": "text", + "content": "[22] Junjie Ke, Keren Ye, Jiahui Yu, Yonghui Wu, Peyman Milanfar, and Feng Yang. Vila: Learning image aesthetics from user comments with vision-language pretraining. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10041-10051, 2023. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 569, + 553, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 569, + 553, + 611 + ], + "spans": [ + { + "bbox": [ + 317, + 569, + 553, + 611 + ], + "type": "text", + "content": "[23] Junjie Ke, Tianhao Zhang, Yilin Wang, Peyman Milanfar, and Feng Yang. Mret: Multi-resolution transformer for video quality assessment. Frontiers in Signal Processing, 3:1137006, 2023. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 613, + 553, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 613, + 553, + 656 + ], + "spans": [ + { + "bbox": [ + 317, + 613, + 553, + 656 + ], + "type": "text", + "content": "[24] Lei Ke, Mingqiao Ye, Martin Danelljan, Yu-Wing Tai, Chi-Keung Tang, Fisher Yu, et al. Segment anything in high quality. Advances in Neural Information Processing Systems, 36, 2024. 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 658, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 658, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 658, + 553, + 713 + ], + "type": "text", + "content": "[25] Minchul Kim, Shangqian Gao, Yen-Chang Hsu, Yilin Shen, and Hongxia Jin. Token fusion: Bridging the gap between token pruning and token merging. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1383-1392, 2024. 2" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7823" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 715 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 297, + 139 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 297, + 139 + ], + "type": "text", + "content": "[26] Shu Kong, Xiaohui Shen, Zhe Lin, Radomir Mech, and Charless Fowlkes. Photo aesthetics ranking network with attributes and content adaptation. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part I 14, pages 662-679. Springer, 2016. 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 140, + 296, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 140, + 296, + 194 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 296, + 194 + ], + "type": "text", + "content": "[27] Jin Li, Yaoming Wang, Xiaopeng Zhang, Bowen Shi, Dongsheng Jiang, Chenglin Li, Wenrui Dai, Hongkai Xiong, and Qi Tian. Ailurus: a scalable vit framework for dense prediction. Advances in Neural Information Processing Systems, 36, 2024. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 194, + 296, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 194, + 296, + 249 + ], + "spans": [ + { + "bbox": [ + 56, + 194, + 296, + 249 + ], + "type": "text", + "content": "[28] Leida Li, Yipo Huang, Jinjian Wu, Yuzhe Yang, Yaqian Li, Yandong Guo, and Guangming Shi. Theme-aware visual attribute reasoning for image aesthetics assessment. IEEE Transactions on Circuits and Systems for Video Technology, 33(9):4798-4811, 2023. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 250, + 296, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 250, + 296, + 303 + ], + "spans": [ + { + "bbox": [ + 56, + 250, + 296, + 303 + ], + "type": "text", + "content": "[29] Leida Li, Tong Zhu, Pengfei Chen, Yuzhe Yang, Yaqian Li, and Weisi Lin. Image aesthetics assessment with attribute-assisted multimodal memory network. IEEE Transactions on Circuits and Systems for Video Technology, 33(12):7413-7424, 2023. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 304, + 296, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 304, + 296, + 358 + ], + "spans": [ + { + "bbox": [ + 56, + 304, + 296, + 358 + ], + "type": "text", + "content": "[30] Leida Li, Xiangfei Sheng, Pengfei Chen, Jinjian Wu, and Weisheng Dong. Towards explainable image aesthetics assessment with attribute-oriented critiques generation. IEEE Transactions on Circuits and Systems for Video Technology, 2024. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 360, + 296, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 360, + 296, + 403 + ], + "spans": [ + { + "bbox": [ + 56, + 360, + 296, + 403 + ], + "type": "text", + "content": "[31] Limin Liu, Shuai He, Anlong Ming, Rui Xie, and Huadong Ma. Elta: An enhancer against long-tail for aesthetic-oriented models. In *Forty-first International Conference on Machine Learning*. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 404, + 296, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 404, + 296, + 459 + ], + "spans": [ + { + "bbox": [ + 56, + 404, + 296, + 459 + ], + "type": "text", + "content": "[32] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 6, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 459, + 296, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 459, + 296, + 480 + ], + "spans": [ + { + "bbox": [ + 56, + 459, + 296, + 480 + ], + "type": "text", + "content": "[33] Ilya Loshchilov and Frank Hutter. Fixing weight decay regularization in adam. 2018. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 481, + 296, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 481, + 296, + 525 + ], + "spans": [ + { + "bbox": [ + 56, + 481, + 296, + 525 + ], + "type": "text", + "content": "[34] Long Mai, Hailin Jin, and Feng Liu. Composition-preserving deep photo aesthetics assessment. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 497-506, 2016. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 525, + 296, + 570 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 525, + 296, + 570 + ], + "spans": [ + { + "bbox": [ + 56, + 525, + 296, + 570 + ], + "type": "text", + "content": "[35] Naila Murray, Luca Marchesotti, and Florent Perronnin. Ava: A large-scale database for aesthetic visual analysis. In 2012 IEEE Conference on Computer Vision and Pattern Recognition, pages 2408-2415, 2012. 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 571, + 296, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 571, + 296, + 613 + ], + "spans": [ + { + "bbox": [ + 56, + 571, + 296, + 613 + ], + "type": "text", + "content": "[36] Tomer Ronen, Omer Levy, and Avram Golbert. Vision transformers with mixed-resolution tokenization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4613-4622, 2023. 2, 3, 4, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 614, + 296, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 614, + 296, + 658 + ], + "spans": [ + { + "bbox": [ + 56, + 614, + 296, + 658 + ], + "type": "text", + "content": "[37] Yossi Rubner, Carlo Tomasi, and Leonidas J Guibas. A metric for distributions with applications to image databases. In Sixth international conference on computer vision (IEEE Cat. No. 98CH36271), pages 59-66. IEEE, 1998. 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 658, + 296, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 296, + 691 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 296, + 691 + ], + "type": "text", + "content": "[38] Peter Shaw, Jakob Uszkoreit, and Ashish Vaswani. Self-attention with relative position representations. arXiv preprint arXiv:1803.02155, 2018. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 691, + 296, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 296, + 715 + ], + "type": "text", + "content": "[39] Xiangfei Sheng, Leida Li, Pengfei Chen, Jinjian Wu, Weisheng Dong, Yuzhe Yang, Liwu Xu, Yaqian Li, and" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 554, + 665 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 333, + 72, + 554, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 72, + 554, + 116 + ], + "spans": [ + { + "bbox": [ + 333, + 72, + 554, + 116 + ], + "type": "text", + "content": "Guangming Shi. Aesclip: Multi-attribute contrastive learning for image aesthetics assessment. In Proceedings of the 31st ACM International Conference on Multimedia, pages 1117-1126, 2023. 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 118, + 554, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 554, + 172 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 554, + 172 + ], + "type": "text", + "content": "[40] Tengfei Shi, Chenglizhao Chen, Zhenyu Wu, Aimin Hao, and Yuming Fang. Improving image aesthetic assessment via multiple image joint learning. ACM Transactions on Multimedia Computing, Communications and Applications, 2024. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 174, + 554, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 174, + 554, + 228 + ], + "spans": [ + { + "bbox": [ + 316, + 174, + 554, + 228 + ], + "type": "text", + "content": "[41] Nyeong-Ho Shin, Seon-Ho Lee, and Chang-Su Kim. Blind image quality assessment based on geometric order learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12799-12808, 2024. 5, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 229, + 554, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 229, + 554, + 274 + ], + "spans": [ + { + "bbox": [ + 316, + 229, + 554, + 274 + ], + "type": "text", + "content": "[42] Hossein Talebi and Peyman Milanfar. Learning to resize images for computer vision tasks. In Proceedings of the IEEE/CVF International Conference on Computer vision, pages 497-506, 2021. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 274, + 554, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 274, + 554, + 319 + ], + "spans": [ + { + "bbox": [ + 316, + 274, + 554, + 319 + ], + "type": "text", + "content": "[43] Zhengzhong Tu, Peyman Milanfar, and Hossein Talebi. Muller: Multilayer laplacian resizer for vision. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6877-6887, 2023. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 319, + 554, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 319, + 554, + 363 + ], + "spans": [ + { + "bbox": [ + 316, + 319, + 554, + 363 + ], + "type": "text", + "content": "[44] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 364, + 554, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 364, + 554, + 430 + ], + "spans": [ + { + "bbox": [ + 316, + 364, + 554, + 430 + ], + "type": "text", + "content": "[45] Jingdong Wang, Ke Sun, Tianheng Cheng, Borui Jiang, Chaorui Deng, Yang Zhao, Dong Liu, Yadong Mu, Mingkui Tan, Xinggang Wang, et al. Deep high-resolution representation learning for visual recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(10):3349-3364, 2020. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 430, + 554, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 430, + 554, + 475 + ], + "spans": [ + { + "bbox": [ + 316, + 430, + 554, + 475 + ], + "type": "text", + "content": "[46] Jian Wang, Xiang Long, Guowei Chen, Zewu Wu, Zeyu Chen, and Errui Ding. U-hrnet: delving into improving semantic representation of high resolution network for dense prediction. arXiv preprint arXiv:2210.07140, 2022. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 475, + 554, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 475, + 554, + 509 + ], + "spans": [ + { + "bbox": [ + 316, + 475, + 554, + 509 + ], + "type": "text", + "content": "[47] Oliver Wiedemann, Vlad Hosu, Shaolin Su, and Dietmar Saupe. Konx: cross-resolution image quality assessment. Quality and User Experience, 8(1):8, 2023. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 510, + 554, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 510, + 554, + 564 + ], + "spans": [ + { + "bbox": [ + 316, + 510, + 554, + 564 + ], + "type": "text", + "content": "[48] Yuzhe Yang, Liwu Xu, Leida Li, Nan Qie, Yaqian Li, Peng Zhang, and Yandong Guo. Personalized image aesthetics assessment with rich attributes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19861-19869, 2022. 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 565, + 554, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 565, + 554, + 620 + ], + "spans": [ + { + "bbox": [ + 316, + 565, + 554, + 620 + ], + "type": "text", + "content": "[49] Ran Yi, Haoyuan Tian, Zhihao Gu, Yu-Kun Lai, and Paul L Rosin. Towards artistic image aesthetics assessment: a large-scale dataset and a new method. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22388-22397, 2023. 4" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 621, + 554, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 621, + 554, + 665 + ], + "spans": [ + { + "bbox": [ + 316, + 621, + 554, + 665 + ], + "type": "text", + "content": "[50] Tong Zhu, Leida Li, Pengfei Chen, Jinjian Wu, Yuzhe Yang, and Yaqian Li. Emotion-aware hierarchical interaction network for multimodal image aesthetics assessment. Pattern Recognition, 154:110584, 2024. 6" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "7824" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Chat-based Person Retrieval via Dialogue-Refined Cross-Modal Alignment/6896dcb7-a954-4e52-90ab-4e44608cdd32_content_list.json b/2025/Chat-based Person Retrieval via Dialogue-Refined Cross-Modal Alignment/6896dcb7-a954-4e52-90ab-4e44608cdd32_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..0bba8fa5bc1920d1b3066fb2d028c873253f2e91 --- /dev/null +++ b/2025/Chat-based Person Retrieval via Dialogue-Refined Cross-Modal Alignment/6896dcb7-a954-4e52-90ab-4e44608cdd32_content_list.json @@ -0,0 +1,1737 @@ +[ + { + "type": "text", + "text": "Chat-based Person Retrieval via Dialogue-Refined Cross-Modal Alignment", + "text_level": 1, + "bbox": [ + 120, + 130, + 875, + 154 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yang Bai $^{1,2}$ , Yucheng Ji $^{3}$ , Min Cao $^{3}$ , Jinqiao Wang $^{2,4}$ , Mang Ye $^{1*}$", + "bbox": [ + 225, + 179, + 776, + 199 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ National Engineering Research Center for Multimedia Software, School of Computer Science, Wuhan University, Wuhan, China", + "bbox": [ + 238, + 199, + 756, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Wuhan AI Research, Wuhan, China", + "bbox": [ + 352, + 233, + 645, + 250 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{3}$ School of Computer Science and Technology, Soochow University, Suzhou, China \n $^{4}$ Institute of Automation, Chinese Academy of Science, Beijing, China", + "bbox": [ + 169, + 250, + 826, + 287 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{ybai.ai, yemang}@whu.edu.cn", + "bbox": [ + 374, + 289, + 622, + 304 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "https://github.com/Flame-Chasers/DiaNA", + "bbox": [ + 326, + 306, + 665, + 320 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 356, + 326, + 372 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Traditional text-based person retrieval (TPR) relies on a single-shot text as query to retrieve the target person, assuming that the query completely captures the user's search intent. However, in real-world scenarios, it can be challenging to ensure the information completeness of such single-shot text. To address this limitation, we propose chat-based person retrieval (ChatPR), a new paradigm that takes an interactive dialogue as query to perform the person retrieval, engaging the user in conversational context to progressively refine the query for accurate person retrieval. The primary challenge in ChatPR is the lack of available dialogue-image paired data. To overcome this challenge, we establish ChatPedes, the first dataset designed for ChatPR, which is constructed by leveraging large language models to automate the question generation and simulate user responses. Additionally, to bridge the modality gap between dialogues and images, we propose a dialogue-refined cross-modal alignment (DiaNA) framework, which leverages two adaptive attribute refiner modules to bottleneck the conversational and visual information for fine-grained cross-modal alignment. Moreover, we propose a dialogue-specific data augmentation strategy, random round retaining, to further enhance the model's generalization ability across varying dialogue lengths. Extensive experiments demonstrate that DiaNA significantly outperforms existing TPR approaches, highlighting the effectiveness of conversational interactions for person retrieval.", + "bbox": [ + 89, + 388, + 483, + 796 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 823, + 220, + 838 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Given the paramount importance of ensuring public safety through intelligent video surveillance, there is a substan", + "bbox": [ + 89, + 848, + 482, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ff8ea91dc79fec4497d59b56db88a76873a99574fa8e2b24f370176ad75dcadb.jpg", + "image_caption": [ + "(a) Illustration of text-based person retrieval (TPR)." + ], + "image_footnote": [], + "bbox": [ + 517, + 354, + 906, + 425 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/19454dbdae73835ec660da26a9342485a666e4116a94ab6c26d698924212da10.jpg", + "image_caption": [ + "(b) Illustration of chat-based person retrieval (ChatPR).", + "Figure 1. Comparison of TPR and ChatPR. (a) TPR relies on a single-shot text as the query, which fails to completely capture users' search intent due to the absence of several details. (b) ChatPR, in contrast, engages users in interactive conversations, prompting them to offer additional clues and clarify the negative intents. This interaction allows a progressive refinement of the query, thereby enhancing the retrieval accuracy. The correct image is marked with green rectangles." + ], + "image_footnote": [], + "bbox": [ + 517, + 448, + 906, + 625 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tial need to develop systems capable of accurately retrieving a person of interest from a large image gallery. In this context, text-based person retrieval (TPR) [23, 39, 53] employs free-form natural language as the query to perform the retrieval, effectively addressing the limitation that visual queries, such as person images (referred to as person re-identification [34, 51, 54, 56]), are not always readily available in real-world scenarios. Consequently, TPR has", + "bbox": [ + 511, + 779, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding Author: Mang Ye", + "bbox": [ + 109, + 887, + 292, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "3952", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "gained ever-rising attention in recent years [24, 42, 50, 58].", + "bbox": [ + 89, + 90, + 480, + 107 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Existing methods in TPR typically rely on a single-shot text query, assuming that the query comprehensively captures the user's search intent. However, in practical applications, it can be challenging to ensure the completeness of information in a single-shot text query for person retrieval. As illustrated in Fig. 1a, the user provides an incomplete single-shot text query with the absence of several details (e.g., \"bag\"), leading to an unsatisfactory retrieval. Alternatively, engaging in multiple rounds of iterative interaction using a question-answer format allows for the conveyance of more detailed information. Recognizing this limitation of single-shot text query in TPR, we introduce a new retrieval paradigm: chat-based person retrieval (ChatPR), which takes interactive dialogues as the query to facilitate a more nuanced and comprehensive interpretation of users' search intent for accurate person retrieval. As shown in Fig. 1b, ChatPR engages users in conversational interaction, allowing the retrieval system to progressively refine the query by actively soliciting additional information, thereby achieving more accurate retrieval.", + "bbox": [ + 89, + 109, + 482, + 411 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ChatPR presents three advantages over traditional TPR. Firstly, it aligns more closely with real-world applications, in which multiple rounds of iterative interaction through a question-answer format provide a more natural way to query the target person. Secondly, the question-answer format actively prompts users for additional information through conversational interactions, allowing the query to encompass more nuanced details. Lastly, it can explicitly capture the user's negative retrieval intents, which are often overlooked in single-shot text queries but are crucial for filtering out irrelevant targets and narrowing the search scope. These unique features highlight the necessity of ChatPR to leverage interactive dialogues for accurate person retrieval.", + "bbox": [ + 89, + 414, + 482, + 611 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, when addressing the ChatPR task, we are faced with two main challenges: ① Lack of dialogue-image paired data: While the research community has witnessed an increase in the availability of text-image paired datasets for TPR [12, 30, 61, 62], there remains an absence of dialogue-image paired dataset. This data deficiency poses a significant barrier to the research and development of ChatPR. ② Complexity of cross-modal alignment: In contrast to the single-shot text, dialogue data presents a more complex structure and longer context. This poses a significant challenge for the traditional TPR models in understanding conversational dialogues, since their text encoder, such as BERT [11] or CLIP [44], is predominantly pretrained on single-text corpora, thereby posing an obstacle to the subsequent cross-modal alignment. Additionally, dialogues tend to have sparser information density, necessitating multiple rounds of interaction to identify specific attributes, which further impedes the cross-modal alignment between dialogue and images.", + "bbox": [ + 89, + 613, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In response to the first challenge, considering that collecting such data via human crowd-sourcing is not only time-consuming but also lacks standardized and well-defined guidelines [35], we propose leveraging the advanced instruction-following capabilities of large language models (LLMs) [8, 14] to convert existing image-text pairs into structured dialogue-image formats. While LLM-based dialogue generation is highly efficient, the resulting data inevitably introduces noise. To address this, we introduce a diverse set of specialized evaluators, composed of LLMs with meticulously crafted instructions, to assess the relevance and quality of each dialogue round. These individual evaluation results are then integrated to facilitate data cleaning. As a result, we construct ChatPedes, the first dataset specifically tailored for ChatPR. This dataset offers a robust benchmark for advancing person retrieval within real-world conversational contexts.", + "bbox": [ + 511, + 90, + 903, + 347 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address the second challenge, we propose a dialogue-refined cross-modal alignment (DiaNA) framework, which consists of dual encoders to comprehensively understand dialogues and images, respectively. Given the natural affinity of LLMs in handling dialogue, we adopt the open-source Llama 3 [14] as the dialogue encoder to effectively comprehend the multi-round interactions. Additionally, DiaNA incorporates two adaptive attribute refiners on the dialogue and image branches, which leverage attribute queries to respectively bottleneck the conversational and visual information. This design establishes a fine-grained cross-modal alignment, thereby enhancing the effectiveness of DiaNA. Furthermore, we propose a dialogue-specific data augmentation strategy, random round retaining (R3), to simulate the variability in dialogue lengths in real-world interactions, where complete information is not always available. This strategy further enhances DiaNA's generalization ability across varying lengths of dialogue context.", + "bbox": [ + 511, + 349, + 903, + 622 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Overall, the primary contributions of this work can be summarized as follows:", + "bbox": [ + 511, + 625, + 903, + 652 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Recognizing the limitations of single-shot text query in text-based person retrieval, we introduce a new retrieval paradigm: chat-based person retrieval (ChatPR), which engages users in interactive conversations to progressively refine their queries for accurate person retrieval.", + "- To fill in the data lack, we establish the first dialogue-image paired dataset, ChatPedes, offering the community a robust benchmark for ChatPR.", + "- To address the challenges in cross-modal alignment, we propose a dialogue-refined cross-modal alignment (Di-ana) framework to enable fine-grained dialogue-image alignment, and introduce a random round retaining strategy to further boost the model's generalization ability.", + "- Extensive experiments demonstrate the effectiveness of DiaNA in handling multi-round dialogues to improve person retrieval performance over conversational contexts." + ], + "bbox": [ + 513, + 656, + 903, + 897 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "3953", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 91, + 89, + 230, + 104 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1. Text-based Person Retrieval", + "text_level": 1, + "bbox": [ + 89, + 118, + 341, + 133 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Text-based person retrieval [39] aims to find the specific person image based on a textual description. Since Li et al. [30] introduced this task, the community has witnessed its growing flourish in recent years. The studies initially depend on simple global alignment [17, 60], and then gradually evolve to multi-granularity correspondences [7, 19, 45]. More recently, a growing number of works [1, 2, 5, 18, 24, 42] have pivoted towards integrating visual-language pretraining models [28, 29, 44] to leverage general cross-modal knowledge for this fine-grained retrieval. Among them, IRRA [24] incorporates a multi-modal interaction encoder on CLIP [44] to facilitate implicit cross-modal relation reasoning. Furthermore, TBPS-CLIP [5] provides a comprehensive empirical study on the application of CLIP in this field from the perspectives of data augmentation, loss function and training tricks.", + "bbox": [ + 89, + 145, + 483, + 387 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Despite these advancements, existing methods implicitly assume that the single text query fully encapsulates the user's search intent. However, in real-world scenarios, user-provided queries often lack sufficient details and complete information, necessitating iterative refinement for satisfactory retrieval. To address this limitation, this paper introduces a novel chat-based person retrieval paradigm, which engages users in interactive dialogues to progressively refine their queries for more accurate and effective retrieval.", + "bbox": [ + 89, + 392, + 483, + 527 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Interactive Cross-Modal Retrieval", + "text_level": 1, + "bbox": [ + 89, + 551, + 388, + 566 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "There has been a significant progress in interactive cross-modal retrieval systems, drawing inspiration from visual dialogue [10, 25]. These systems engage users in a dynamic feedback loop towards the desired retrieval target. Typically, user feedback can be categorized into two types: relevance feedback and difference feedback. Relevance feedback [48] involves users assigning scores to the retrieved results based on their relevance to the query, allowing the system to re-rank its retrieval list accordingly. In contrast, difference feedback [4, 33] requires users to specify the distinctions between the target image and the reference image through pre-defined attribute tags [22, 49] or open-form descriptions [21]. Moreover, recent methods [26, 27] bridge the feedback with the query through conversational questioning, enabling the users to be pro-actively asked according to the dialogue history in each search attempt.", + "bbox": [ + 89, + 578, + 482, + 820 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Different from the aforementioned works, this paper focuses on the specific field of ChatPR, a more fine-grained retrieval task, where users need to provide more comprehensive and discriminative queries to retrieve the target pedestrian images from a vast image gallery.", + "bbox": [ + 89, + 825, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Dialogue Construction Instruction", + "text_level": 1, + "bbox": [ + 604, + 95, + 815, + 108 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Task Definition", + "text_level": 1, + "bbox": [ + 563, + 119, + 650, + 131 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Act as an Assistant and a User to convert the giving captions of the pedestrian into a question-answer format. You need to integrate the information from the both captions. The Assistant's goal is to reveal as many details as possible ... details not mentioned in the captions.", + "bbox": [ + 535, + 133, + 883, + 178 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Workflow", + "text_level": 1, + "bbox": [ + 568, + 190, + 625, + 200 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. Understand the two sentences in the captions.", + "2. Firstly, the Assistant asks the User to describe the person's features.", + "3. The User describes the person's features based on the captions." + ], + "bbox": [ + 537, + 207, + 870, + 239 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "6. Don't provide any question or answer that has already been mentioned in previous dialogue history.", + "bbox": [ + 537, + 250, + 880, + 272 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Examples", + "text_level": 1, + "bbox": [ + 573, + 284, + 625, + 296 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Captions: [\"This man has dark hair and wears a black hoodie with white pattern outside and dark pants ..., \"He's wearing a black hooded sweatshirt with a white lining ...\"", + "Output:" + ], + "bbox": [ + 540, + 301, + 885, + 345 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "Assistant: Please describe the person.", + "User: The person is wearing a black hoodie, green pants, a pair of red and white shoes.", + "Assistant: What's the person's gender?", + "User: He is male.", + "." + ], + "bbox": [ + 553, + 345, + 885, + 406 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Figure 2. The instruction for dialogue construction, which consists of three parts: task definition, workflow and examples. This design aims to activate LLMs' instruction-following, chain-of-thought and in-context capabilities, respectively.", + "bbox": [ + 511, + 424, + 906, + 481 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Large Language Models", + "text_level": 1, + "bbox": [ + 511, + 491, + 736, + 507 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In recent years, the field of artificial intelligence has witnessed a significant leap forward with the advent of large language models (LLMs) [6, 32, 55, 59]. Among the most notable contributions to this domain are the generative pretrained transformer (GPT) series [3, 40, 41, 43]. These models are trained on massive corpora, allowing them to capture a wide range of linguistic patterns and world knowledge. Moreover, the large-scale parameters endow LLMs with emergent abilities [47] that are not present in smaller models [16], such as instruction-following [38] and in-context [13] learning. These capabilities allow LLMs to accurately interpret and execute complex human instructions, as well as understand and utilize context from a few examples to perform tasks effectively.", + "bbox": [ + 511, + 513, + 906, + 724 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this work, we utilize LLMs to automate question generation and simulate user responses for the construction of ChatPedes. By providing carefully crafted instructions and a few examples, we fully exploit their instruction-following and in-context learning capabilities to create the first benchmark for ChatPR.", + "bbox": [ + 511, + 724, + 906, + 814 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Benchmark", + "text_level": 1, + "bbox": [ + 511, + 830, + 635, + 844 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recognizing the limitations of single-shot text queries with incomplete descriptions in TPR, we propose the ChatPR task, which enables the progressive query refine", + "bbox": [ + 511, + 854, + 905, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3954", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/c1b19b32370dfabdefb46f6d1794262cf737b8d94285fe483686481cb0859860.jpg", + "image_caption": [ + "Figure 3. Category distribution of dialogue data from Llama 3 [14] and InternVL [8], evaluated by Qwen 2.5 [46]. Over one-quarter of the data is not matched, highlighting the necessity of data cleaning." + ], + "image_footnote": [], + "bbox": [ + 112, + 88, + 464, + 180 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ment through conversational interactions for accurate person retrieval. However, the primary challenge lies in the lack of dialogue-image paired data. To address this, we construct the ChatPedes dataset through two key phases: dialogue construction and data cleaning.", + "bbox": [ + 89, + 255, + 483, + 333 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Dialogue Construction", + "text_level": 1, + "bbox": [ + 89, + 345, + 302, + 362 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In recent years, the research community has witnessed an increase in the availability of text-image paired datasets for TPR [12, 30, 61, 62]. Typically, to create high-quality datasets, a large workforce is employed to manually annotate the images with textual descriptions. However, when it comes to annotating with conversational dialogues via human crowd-sourcing for ChatPR, two major challenges arise: ① Increased difficulty of dialogue annotation: Compared to single-shot text, dialogue data presents a more complex structure and longer context, significantly increasing the manual annotation workload. ② Less well-defined guidelines: Unlike textual annotation tasks with clear instructions, formulating appropriate conversations for dialogue annotation lacks well-defined guidelines [35], complicating the annotation process. Consequently, these challenges make it difficult to employ human crowd-sourcing for creating dialogue-image paired datasets.", + "bbox": [ + 88, + 368, + 482, + 626 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Inspired by the success of LLMs in text annotation tasks [20, 50], we leverage their advanced instruction-following capabilities to automate dialogue annotations, based on the well-known text-image paired CUHK-PEDES dataset [30] in TPR. As illustrated in Fig. 4, dialogue construction aims to construct on-going conversations through the interaction between the LLM-based AI agent and the user. During the interaction, the AI agent formulates follow-up questions based on previous dialogue history. The user, also simulated by the LLM, responds with corresponding answers based on the characteristics of the target person, drawing from the existing annotated captions.", + "bbox": [ + 89, + 627, + 482, + 808 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In general, it is expected that the dialogue data produced by the LLM-based AI agent contains comprehensive and valuable information, which is advantageous for subsequent model training. Consequently, we empirically design the instructions shown in Fig. 2. Structurally, the instruction consists of three parts: task definition, workflow, and a", + "bbox": [ + 89, + 810, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/348c5e9be210b13d612537298e44ba155f5044fa38e6b4aef3f3420f18b9a374.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetsIDsImagesTextsDlg.s.RoundsWords
maxminavgmaxminavg
CUHK-PEDES [30]13,00340,20680,440----961523.5
ICFG-PEDES [12]4,10254,52254,522----83937.2
RSTPReid [61]4,10120,50541,010----701126.5
UFine6926 [62]6,92626,20652,412----2183080.8
ChatPedes (Ours)12,00337,128-74,2561717.738812114.4
", + "bbox": [ + 516, + 88, + 903, + 188 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Table 1. Comparison between our established ChatPedes and the widely-used TPR benchmarks. \"Dlgs.\" refers to dialogues. ChatPedes uses the multi-round dialogues as queries, demonstrating a longer context with conversational interactions.", + "bbox": [ + 511, + 191, + 906, + 247 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "few manually designed examples, aiming to activate LLMs' instruction-following, chain-of-thought and in-context capabilities to construct high-quality dialogue data. In the dialogue content, the first question provided by the AI agent typically inquires about a holistic description of the target person. Furthermore, to enhance data diversity, we leverage two publicly available LLMs, Llama 3 [14] and InternVL [8], to generate two distinct dialogues for each person image. The experiments detailed in Sec. 5.3.2 substantiate the benefits of this data diversification strategy.", + "bbox": [ + 511, + 255, + 906, + 409 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Data Cleaning", + "text_level": 1, + "bbox": [ + 513, + 414, + 661, + 431 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Although LLMs facilitate the dialogue data collection, the generated data inevitably contains noise, which would significantly impair the optimization of the retrieval process [57]. To perform the data cleaning, we first define the following four classes for each dialogue round:", + "bbox": [ + 511, + 436, + 906, + 513 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Matched: The answer either aligns with the information present in the annotated captions or accurately reflects the absence of such details not mentioned in the captions.", + "- Contradictory: The answer contradicts the information present in the captions.", + "- Hallucinatory: The answer contains fabricated information that is not present in the captions.", + "- Repeated: The question or answer contains information that has been mentioned in earlier dialogue rounds." + ], + "bbox": [ + 511, + 512, + 905, + 647 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Subsequently, we employ LLMs, including Qwen 2.5 [46], Llama 3 [14] and InternVL [8], as specialized evaluators through the crafted instructions (detailed in Appendix B.1) to classify each dialogue round into the four predefined categories. As illustrated in Fig. 3, the category distribution estimated by Qwen 2.5 reveals that over one-quarter of the dialogue data is classified as unmatched, highlighting the necessity for data cleaning. To address this, a vote-based ensemble strategy is adopted to perform the data cleaning. Specifically, we calculate the matching score $s$ for each dialogue round, which represents the proportion of the answers labeled as \"Matched\" by the LLMs:", + "bbox": [ + 511, + 648, + 906, + 829 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ns = \\frac {1}{N _ {q}} \\sum_ {i = 1} ^ {N _ {q}} \\mathbb {1} \\left(a _ {i} = \\text {M a t c h e d}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 604, + 835, + 906, + 878 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbb{1}(\\cdot)$ is identity function, $N_{q}$ denotes the number of", + "bbox": [ + 511, + 885, + 906, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "3955", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/7bad6ec9e6d5ee641dbb85082db89546f8c76b47b8dca0e7aed473262a4c260d.jpg", + "image_caption": [ + "Figure 4. The overview of ChatPR. Dialogue Construction aims to construct an on-going conversation, in which an AI agent formulates follow-up questions based on previous dialogue history, prompting the user to provide additional information about the target person to be searched. This iterative process forms a comprehensive Dialogue Query for the target person. DiaNA aims to bridge the modality gap between the dialogue and the image, which incorporates two adaptive attribute refiner (AAR) modules by leveraging learnable attribute queries to extract key information for fine-grained cross-modal alignment. During training, random round retaining (R3) strategy simulates the incomplete query in reality to enhance the model's generalization ability." + ], + "image_footnote": [], + "bbox": [ + 93, + 87, + 906, + 349 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "the adopted evaluators, and $a_{i}$ represents the $i$ -th evaluator's answer. If the score $s$ is less than half of the number of evaluators $N_{q}$ , this round of conversation is regarded as data noise and then discarded from the dialogue. Detailed statistics of data cleaning are provided in Appendix C.", + "bbox": [ + 89, + 449, + 482, + 523 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Finally, we establish ChatPedes, a dialogue-image paired dataset designed for ChatPR. As shown in Tab. 1, ChatPedes comprises 37,128 images from 12,003 identities, with each image accompanied by two generated dialogues. Consistent with the data split in CUHK-PEDES [30], we divide 34,054 image from 11,003 identities into the training set, and 3,074 image from 1,000 identities into the test set. In addition, ChatPedes is characterized by multi-round dialogues and longer context, exhibiting the longest average length of 114.4 words per dialogue among all benchmarks. Visualization examples are presented in Appendix E.", + "bbox": [ + 89, + 523, + 482, + 691 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Method", + "text_level": 1, + "bbox": [ + 89, + 702, + 181, + 717 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Dialogue-Refined Cross-Modal Alignment", + "text_level": 1, + "bbox": [ + 89, + 727, + 449, + 743 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "ChatPR engages users in interactive dialogues to progressively refine their queries, where they are pro-actively questioned and prompted to provide more detailed information for more accurate person retrieval. To bridge the modality gap between dialogues and images, we propose a dialogue-refined cross-modal alignment (DiaNA) framework, as illustrated in Fig. 4, which aims to learn a dialogue-image shared latent space, pulling the pairs with consistent semantics closer together, while pushing those with different semantics further apart. DiaNA is composed of dual", + "bbox": [ + 89, + 750, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "encoders to comprehend dialogue queries and person images, respectively, along with two adaptive attribute refiners to bottleneck the conversational and visual information for fine-grained cross-modal alignment.", + "bbox": [ + 511, + 449, + 906, + 510 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1.1. Dual Encoders", + "text_level": 1, + "bbox": [ + 511, + 515, + 661, + 527 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In contrast to the single-shot text query in TPR, dialogue data in ChatPR presents a more complex structure and longer context, rendering traditional text processing models, such as BERT [11] and CLIP [44] (text encoder), less suitable for this task. Given the proficiency of LLMs in handling dialogue data, we adopt the publicly accessible Llama 3 [14] to comprehend the dialogues within our framework.", + "bbox": [ + 511, + 534, + 905, + 638 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Formally, we denote a dialogue as $\\mathbf{D} = \\{(Q_i,A_i)\\}_{i = 1}^N$ , where $Q_{i}$ and $A_{i}$ represent the $i$ -th round of question and answer in the dialogue with a total of $N$ rounds. We construct the instruction-format input, as depicted in Fig. 5. Following Vicuna [9], we prepend a system message $X_{\\mathrm{system - message}}$ (detailed in Appendix B.2) to activate the instruction-following capability of the LLM, enabling a more comprehensive understanding of the dialogue data. Finally, the entire input is fed into Llama 3, yielding an embedding sequence $\\mathcal{W} = \\{w_{bos},w_1,\\dots ,w_{eos}\\}$ , where $w_{eos}$ captures the full semantics of the dialogue due to the decoder-only architecture of Llama 3. Subsequently, the global embedding $w_{eos}$ is mapped into the shared space:", + "bbox": [ + 511, + 638, + 906, + 837 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nw = \\Phi_ {D} \\left(w _ {e o s}\\right) \\in \\mathbb {R} ^ {d}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 633, + 849, + 903, + 867 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\Phi_D$ is a projector composed of a linear layer, $d$ is the dimension of the space.", + "bbox": [ + 511, + 869, + 905, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "3956", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/b707f2ce3ebe60a9367343ea56fbe0d52479ab6451938eaa147ef40cfc26e4e9.jpg", + "image_caption": [ + "Figure 5. The instruction-formatted input sequence. A system message is prepended to prompt Llama 3 in capturing the semantics of the dialogue context." + ], + "image_footnote": [], + "bbox": [ + 112, + 88, + 460, + 161 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For the visual branch, given an image $\\mathbf{I}$ , we employ Swin Transformer [36, 37] to encode the image into a visual sequence $\\mathcal{V} = \\{v_{1}, v_{2}, \\dots, v_{h \\times w}\\}$ , where $h$ and $w$ represent the height and width of the feature map at the last stage, respectively. Subsequently, a global average pooling layer is adopted to integrate the overall semantics, yielding a global visual embedding $v_{g} = \\frac{1}{h \\times w} \\sum_{i=1}^{h \\times w} v_{i}$ . Similarly, we introduce a projector to map $v_{g}$ into the shared space:", + "bbox": [ + 89, + 215, + 483, + 339 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nv = \\Phi_ {I} (v _ {g}) \\in \\mathbb {R} ^ {d}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 222, + 345, + 482, + 364 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1.2. Adaptive Attribute Refiner", + "text_level": 1, + "bbox": [ + 89, + 372, + 326, + 386 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Unlike the single-shot text where content is generally concentrated, dialogue data typically exhibits sparser information density, requiring multiple rounds of interaction to ascertain specific attributes. This sparsity poses a significant challenge in cross-modal alignment, necessitating information refinement to effectively bridge the modality gap between dialogues and images. To address this, we propose two adaptive attribute refiner (AAR) modules on the dialogue and the image encoders to learn the attribute representations for fine-grained cross-modal alignment.", + "bbox": [ + 89, + 390, + 483, + 541 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Specifically, we first introduce a dialogue AAR upon the dialogue encoder. Considering that different persons have varying attributes, we introduce a set of learnable attribute queries $\\mathcal{Q} = \\{q_1, q_2, \\dots, q_K\\}$ to focus on the distinct attributes of each individual, where $K$ denotes the number of attribute queries. By interacting with the dialogue embeddings $\\mathcal{W}$ , these attribute queries adaptively extract the person-specific attribute representations from the sparse dialogue content. As illustrated in Fig. 4, the dialogue AAR applies a computationally efficient structure with a single multi-head cross attention (MCA) followed by 2-layer transformer blocks. The interaction between the attribute queries and the dialogue embeddings is performed through the cross-attention mechanism:", + "bbox": [ + 89, + 542, + 483, + 752 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathcal {W}} = \\operatorname {T r a n s f o r m e r} (\\operatorname {M C A} (\\mathcal {Q}, \\mathcal {W}, \\mathcal {W})), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 153, + 760, + 482, + 777 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\hat{\\mathcal{W}} = \\{\\hat{w}_1,\\hat{w}_2,\\dots ,\\hat{w}_K\\}$ indicates the adaptively refined dialogue attribute embeddings.", + "bbox": [ + 89, + 785, + 482, + 815 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In parallel, we symmetrically introduce a visual AAR to extract the visual attributes from the image, mitigating the impacts of the inherent redundancy and background noise in visual information:", + "bbox": [ + 89, + 816, + 483, + 875 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathcal {V}} = \\operatorname {T r a n s f o r m e r} (\\operatorname {M C A} (\\mathcal {Q}, \\mathcal {V}, \\mathcal {V})), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 163, + 883, + 482, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\hat{\\mathcal{V}} = \\{\\hat{v}_1,\\hat{v}_2,\\dots ,\\hat{v}_K\\}$ denotes the visual attribute embeddings. Furthermore, we leverage a parameter-sharing mechanism between the two AAR modules, which not only reduces the parameter amount but also mitigates their modality gap [15] between dialogues and images.", + "bbox": [ + 511, + 90, + 905, + 167 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Optimization Objectives", + "text_level": 1, + "bbox": [ + 511, + 176, + 736, + 193 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Incorporating the dual encoders and the AAR modules, we construct DiaNA, which is then optimized through both global (coarse-grained) and local (fine-grained) alignment. Specifically, we use the cosine similarity to measure the global alignment between the dialogue and the image:", + "bbox": [ + 511, + 199, + 905, + 276 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\ns _ {g} (\\mathbf {D}, \\mathbf {I}) = \\frac {w ^ {\\mathrm {T}} v}{\\| w \\| \\| v \\|}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 635, + 286, + 905, + 321 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Furthermore, the average cosine distance among the refined attribute embeddings is utilized to measure the local semantics consistency between the dialogue and the image:", + "bbox": [ + 511, + 332, + 905, + 378 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\ns _ {l} (\\mathbf {D}, \\mathbf {I}) = \\frac {1}{K} \\sum_ {i = 1} ^ {K} \\frac {\\hat {w} _ {i} ^ {\\mathrm {T}} \\hat {v} _ {i}}{\\| \\hat {w} _ {i} \\| \\| \\hat {v} _ {i} \\|.} \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 607, + 390, + 905, + 431 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "According to the global and local similarities between the dialogue and the image, we construct the overall optimization object by employing the normalized image-text contrastive (N-ITC) loss [5] and the similarity distribution matching (SDM) loss [24]:", + "bbox": [ + 511, + 443, + 905, + 518 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\underbrace {\\mathcal {L} _ {g} ^ {n i t c} + \\mathcal {L} _ {g} ^ {s d m}} _ {\\text {g l o b a l a l i n g m e n t}} + \\underbrace {\\mathcal {L} _ {l} ^ {n i t c} + \\mathcal {L} _ {l} ^ {s d m}} _ {\\text {l o c a l a l i n g m e n t}}. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 584, + 529, + 903, + 568 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Dialogue-specific Data Augmentation", + "text_level": 1, + "bbox": [ + 511, + 592, + 834, + 608 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In real-world scenarios, complete dialogues that capture users' full search intent are not always available, since obtaining these complete dialogues typically requires numerous rounds of interaction, which can be not user-friendly. To enhance the model's generalization ability over dialogues of varying rounds, we propose a dialogue-specific data augmentation strategy, random round retaining (R3), which simulates the inherent variability of the dialogue in practice.", + "bbox": [ + 511, + 614, + 905, + 734 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Specifically, at each training iteration, for a complete dialogue $\\mathbf{D} = \\{(Q_i,A_i)\\}_{i = 1}^N$ , only the initial random $i$ rounds are retained and fed into the model:", + "bbox": [ + 511, + 736, + 905, + 780 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {D}} = \\left\\{\\left(Q _ {i}, A _ {i}\\right) \\right\\} _ {i = 1} ^ {n}, \\quad n \\sim \\operatorname {r a n d i n t} (1, N), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 792, + 903, + 811 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\operatorname{randint}(1, N)$ denotes a discrete uniform distribution with equal probability for each integer in range $[1, N]$ . The retained $\\hat{\\mathbf{D}}$ reflects the partial availability of information in practical scenarios, effectively enhancing the model's generalization ability over varying dialogue lengths.", + "bbox": [ + 511, + 823, + 905, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "3957", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/65282b51bc2420e24e1c9ff4717dcdb53efa8bfebf9c1610592e7bb4ad13b81f.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodRefPret. Dataw/o Data Cleaningw/ Data Cleaning
R-1R-5R-10mAPR-1R-5R-10mAP
IRRA [24]CVPR'23WIT46.0865.6073.0242.9048.2667.9675.6544.64
TBPS-CLIP [5]AAAI'24WIT47.7466.4173.7843.8050.2169.2776.4345.44
RDE [42]CVPR'24WIT50.2068.2575.1345.9452.7770.6177.4747.98
APTM [52]MM'23MALS39.6256.9964.2336.5741.1557.9165.1638.00
AUL [31]AAAI'24MALS40.0557.0164.4137.0242.0159.2166.0138.56
DiaNA (Ours)-MALS73.7289.6694.2465.4075.6790.6894.7166.89
", + "bbox": [ + 142, + 88, + 864, + 222 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Comparison with state-of-the-art methods on ChatPedes. \"Pret. Data\" denotes pretraining data. All reported results of other methods are reproduced using their publicly available code.", + "bbox": [ + 89, + 224, + 906, + 253 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 89, + 263, + 223, + 280 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we conduct extensive experiments to verify the effectiveness of our method. The implementation details are provided in Appendix A.", + "bbox": [ + 89, + 287, + 483, + 335 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.1. Evaluation Protocol", + "text_level": 1, + "bbox": [ + 89, + 345, + 279, + 359 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We adopt the widely-used Rank-K (R-K for short, $\\mathrm{K} = 1,5,10$ ) metric in our experiments, which denotes the percentage of successful retrieval that the ground-truth image is found within the top K ranked results. In addition, we also employ the mean average precision (mAP) as an auxiliary metric to evaluate the overall retrieval performance.", + "bbox": [ + 89, + 368, + 483, + 460 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Comparison with SOTA Methods", + "text_level": 1, + "bbox": [ + 89, + 469, + 382, + 484 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We compare our method against five state-of-the-art (SOTA) baselines: IRRA [24], TBPS-CLIP [5], RDE [42], APTM [52] and AUL [31]. Among these, IRRA, TBPS-CLIP and RDE are built on the vision-language foundation model CLIP [44], which is pretrained on a large-scale dataset WIT with 400 million image-text paired data collected from the Internet. APTM and AUL are pretrained on the synthetic text-image paired dataset MALS [52]. It is worth noting that these baselines are originally developed for the TPR task. We reproduce their performance on ChatPedes dataset using their open-source codes, treating the multi-turn dialogue data as a single long-form text input. As shown in Tab. 7, regardless of whether data cleaning is applied, our method outperforms these SOTA baselines by a large margin, while these baselines demonstrate a suboptimal retrieval accuracy, indicating their limitations in effectively handling the complex structure and longer contextual dependencies of dialogue data.", + "bbox": [ + 89, + 492, + 482, + 763 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In addition, we further conduct extensive comparative experiments over the dialogue round, simulating real-world scenarios where users are pro-actively questioned to progressively refine their queries through conversational interactions. This process is exemplified in Appendix E.1. As shown in Fig. 6, it can be observed that: ① As the dialogue progresses, the retrieval performance improves across all methods. However, the performance of the baselines becomes saturated after the initial few rounds of interaction,", + "bbox": [ + 89, + 765, + 483, + 900 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/faf14a2c0db406e51a0611d2a6217a13c035ac18c6ca0e6791b2c644c3696b3b.jpg", + "image_caption": [ + "Figure 6. Performance comparison over dialogue round." + ], + "image_footnote": [], + "bbox": [ + 516, + 263, + 906, + 402 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "which we attribute to their limitations in processing longer and more complex dialogues. In contrast, our method presents a significantly pronounced improvement as the dialogue evolves on account of our carefully crafted design for handling dialogue data. This comprehensive comparison demonstrates the efficacy of our method and showcases its superior capability in handling the intricacies of dialogue-image alignment. Additionally, we also notice that our method achieves competitive performance at the first round of dialogue. This round typically involves a holistic description of the target person, functioning similarly to the single-shot text in the traditional TPR task. The experimental comparison also indicates the promising generalization capabilities of our method in TPR. Further detailed experimental results on TPR benchmarks are provided in Appendix D.", + "bbox": [ + 511, + 434, + 906, + 662 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Ablation Study", + "text_level": 1, + "bbox": [ + 511, + 675, + 666, + 691 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3.1. Effectiveness of Data Cleaning", + "text_level": 1, + "bbox": [ + 511, + 699, + 772, + 714 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "During dialogue construction, we leverage LLMs to generate dialogue data. However, a major limitation of this automatic generation through LLMs lies in the inevitable introduction of noise, which can be classified into three categories: Contradictory, Hallucinatory and Repeated, as delineated in Sec. 3.2. To evaluate the impact of the noise, we conduct comparative experiments with and without data cleaning. Tab. 7 presents the performance comparison with the five SOTA baselines, where all approaches exhibit a significant improvement. In particular, RDE achieves a notable increase of $2.57\\%$ at R-1, demonstrating the effectiveness of data cleaning and the importance of high-quality data.", + "bbox": [ + 511, + 719, + 906, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "3958", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/47737baa3e5c81569ec91f094b58f092de790fae81df169c3afa9ef67ba03c66.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Data DiversityR-1R-5R-10mAP
Llama 3InternVL
X✓71.0188.4093.1863.08
✓X71.6288.5893.5963.73
✓✓75.6790.6894.7166.89
", + "bbox": [ + 102, + 88, + 459, + 172 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0abb2854e2375e367e1571ad2a47b4ebdf4ff8b69757ac381ff10e92e72f7f80.jpg", + "image_caption": [ + "Figure 7. Performance comparison across a varying number of attribute queries $K$ in AAR modules. Notably, $K = 0$ indicates that the AAR is not used." + ], + "image_footnote": [], + "bbox": [ + 120, + 217, + 454, + 353 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.3.2. Effectiveness of Data Diversity", + "text_level": 1, + "bbox": [ + 89, + 410, + 349, + 426 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We employ Llama 3 [14] and InternVL [8] as different annotators for labeling diverse dialogue data when constructing ChatPedes. Tab. 3 presents the contribution of data diversity to the retrieval performance. It can be observed that: ① Relying on a single LLM yields a suboptimal outcome, with $71.01\\%$ for InternVL and $71.62\\%$ for Llama 3 at R-1. Additionally, Llama 3 slightly outperforms InternVL due to its more advanced capability in dialogue generation. ② The combination of Llama 3 and InternVL leads to a substantial increase across all metrics, demonstrating the importance of data diversity. By combining annotations generated by different LLMs, the established ChatPedes dataset effectively captures a wider variety of dialogue contexts and nuances, which in turn bolsters the model's retrieval capabilities.", + "bbox": [ + 89, + 430, + 483, + 643 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.3.3. Effectiveness of Adaptive Attribute Refiner", + "text_level": 1, + "bbox": [ + 89, + 654, + 436, + 669 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Considering the sparse information density in dialogues and the inherent redundancy in images, we propose two adaptive attribute refiner (AAR) modules, which employ a set of learnable attribute-specific queries to adaptively extract the relevant attributes from the dialogue and the image. Fig. 7 illustrates the performance variation over the number of attribute queries $K$ , where, notably, $K = 0$ denotes AAR is discarded. We observe that: 1 Only relying on global alignment without AAR results in a suboptimal retrieval performance, underscoring the importance of key information extraction. 2 When introducing AAR, the performance exhibits slight fluctuations, reaching the peak of $75.67\\%$ on R-1 and $66.89\\%$ on mAP at $K = 16$ . Overall, the performance remains relatively stable, indicating the model's robustness when equipped with AAR.", + "bbox": [ + 89, + 672, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/f235b15f610fc83c69cfd403e69607b7dc696961444a551e5376e21a59419b82.jpg", + "table_caption": [ + "Table 3. Ablation study on data diversity. We employ Llama 3 [14] and InternVL [8] to construct the ChatPedes benchmark." + ], + "table_footnote": [], + "table_body": "
MethodMetricDialogue Round
12468all
DiaNA(w/o R3)26.5034.8953.0766.4073.1675.18
DiaNA(w/ R3)R-129.2537.6455.6067.9973.6775.67
Δ+2.75+2.75+2.53+1.59+0.51+0.49
DiaNA(w/o R3)26.3333.6749.0659.6764.6466.19
DiaNA(w/ R3)mAP28.7836.3751.0560.8865.3266.89
Δ+2.45+2.70+1.99+1.21+0.68+0.70
", + "bbox": [ + 524, + 89, + 893, + 205 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. Comparison of random round retaining (R3) over dialogue round. \"all\" refers to the complete dialogue.", + "bbox": [ + 511, + 208, + 903, + 236 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.3.4. Effectiveness of Random Round Retaining", + "text_level": 1, + "bbox": [ + 511, + 250, + 856, + 265 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The proposed random round retaining (R3) strategy selectively retains a random number of initial rounds of the dialogue during training, simulating the users' incomplete search intent in real-world interactions. As shown in Tab. 4, we observe that: $\\mathbf{1}$ R3 yields performance gains across all dialogue rounds, with gains from $2.75\\%$ in single-round interactions to $0.49\\%$ in complete conversations at R-1, demonstrating its robustness and versatility. $\\mathbf{2}$ In particular, R3 brings more substantial improvement at the initial few rounds of dialogues, suggesting that R3 is particularly effective when dialogue information is limited. This is especially crucial in practical scenarios where users expect to quickly retrieve the target person with minimal interaction. Overall, R3 empowers users to obtain accurate retrieval results even in the early stages of interaction, thereby addressing the real-world demands for both efficiency and precision.", + "bbox": [ + 511, + 268, + 906, + 512 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 526, + 633, + 542 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Recognizing the inherent limitations of traditional TPR where single-shot text queries fail to fully encapsulate users' search intent, we introduce ChatPR, a pioneering paradigm for person retrieval that engages users in interactive dialogues to progressively refine their search queries. For this, we build the first benchmark ChatPedes through LLMs to automate question generation and simulate users' responses. To tackle this task, we propose the DiaNA framework that leverages two adaptive attribute refiners to bottleneck the conversational and visual information for fine-grained cross-modal alignment. Additionally, a novel data augmentation strategy of random round retaining is proposed to mimic the incomplete queries in real-world scenarios, remarkably enhancing the model's generalization ability across varying dialogue lengths. Extensive experiments significantly demonstrate the importance of conversational interaction in person retrieval and the superior effectiveness of our method in ChatPR.", + "bbox": [ + 511, + 551, + 906, + 823 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement. This work is supported by the National Science Foundation of China under Grants (62176188, 62361166629, 62476188). The numerical calculations are supported by the supercomputing system in the Supercomputing Center of Wuhan University.", + "bbox": [ + 511, + 825, + 905, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "3959", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Yang Bai, Min Cao, Daming Gao, Ziqiang Cao, Chen Chen, Zhenfeng Fan, Liqiang Nie, and Min Zhang. Rasa: Relation and sensitivity aware representation learning for text-based person search. In Proceedings of the Thirty-Second International Joint Conference on Artificial Intelligence, pages 555–563, 2023. 3", + "[2] Yang Bai, Jingyao Wang, Min Cao, Chen Chen, Ziqiang Cao, Liqiang Nie, and Min Zhang. Text-based person search without parallel image-text data. In Proceedings of the 31st ACM International Conference on Multimedia, pages 757-767, 2023. 3", + "[3] Tom B Brown. Language models are few-shot learners. arXiv preprint arXiv:2005.14165, 2020. 3", + "[4] Guanyu Cai, Jun Zhang, Xinyang Jiang, Yifei Gong, Lianghua He, Fufu Yu, Pai Peng, Xiaowei Guo, Feiyue Huang, and Xing Sun. Ask&confirm: active detail enriching for cross-modal retrieval with partial query. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1835-1844, 2021. 3", + "[5] Min Cao, Yang Bai, Ziyin Zeng, Mang Ye, and Min Zhang. An empirical study of clip for text-based person search. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 465-473, 2024. 3, 6, 7", + "[6] Yupeng Chang, Xu Wang, Jindong Wang, Yuan Wu, Linyi Yang, Kaijie Zhu, Hao Chen, Xiaoyuan Yi, Cunxiang Wang, Yidong Wang, et al. A survey on evaluation of large language models. ACM Transactions on Intelligent Systems and Technology, 15(3):1-45, 2024. 3", + "[7] Yuhao Chen, Guoqing Zhang, Yujiang Lu, Zhenxing Wang, and Yuhui Zheng. Tipcb: A simple but effective part-based convolutional baseline for text-based person search. Neurocomputing, 494:171-181, 2022. 3", + "[8] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24185-24198, 2024. 2, 4, 8", + "[9] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhang-hao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yong-hao Zhuang, Joseph E. Gonzalez, Ion Stoica, and Eric P. Xing. Vicuna: An open-source chatbot impressing gpt-4 with $90\\%$ chatgpt quality. https://lmsys.org/blog/2023-03-30-vicuna, 2023.5", + "[10] Abhishek Das, Satwik Kottur, Khushi Gupta, Avi Singh, Deshraj Yadav, Jose M. F. Moura, Devi Parikh, and Dhruv Batra. Visual dialog. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 326-335, 2017. 3", + "[11] Jacob Devlin. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 2, 5", + "[12] Zefeng Ding, Changxing Ding, Zhiyin Shao, and Dacheng Tao. Semantically self-aligned network for text-to-image part-aware person re-identification. arXiv preprint arXiv:2107.12666, 2021. 2, 4" + ], + "bbox": [ + 93, + 114, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Qingxiu Dong, Lei Li, Damai Dai, Ce Zheng, Zhiyong Wu, Baobao Chang, Xu Sun, Jingjing Xu, and Zhifang Sui. A survey on in-context learning. arXiv preprint arXiv:2301.00234, 2022. 3", + "[14] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 2, 4, 5, 8", + "[15] Sedigheh Eslami and Gerard de Melo. Mitigate the gap: Investigating approaches for improving cross-modal alignment in clip. arXiv preprint arXiv:2406.17639, 2024. 6", + "[16] Yiyang Fang, Wenke Huang, Guancheng Wan, Kehua Su, and Mang Ye. Emoe: Modality-specific enhanced dynamic emotion experts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2025. 3", + "[17] Ammarah Farooq, Muhammad Awais, Fei Yan, Josef Kittler, Ali Akbari, and Syed Safwan Khalid. A convolutional baseline for person re-identification using vision and language descriptions. arXiv preprint arXiv:2003.00808, 2020. 3", + "[18] Takuro Fujii and Shuhei Tarashima. Bilma: Bidirectional local-matching for text-based person re-identification. In Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops, pages 2786-2790, 2023. 3", + "[19] Chenyang Gao, Guanyu Cai, Xinyang Jiang, Feng Zheng, Jun Zhang, Yifei Gong, Fangzhou Lin, Xing Sun, and Xi-ang Bai. Conditional feature learning based transformer for text-based person search. IEEE Transactions on Image Processing, 31:6097–6108, 2022. 3", + "[20] Fabrizio Gilardi, Meysam Alizadeh, and Maël Kubli. Chatgpt outperforms crowd workers for text-annotation tasks. Proceedings of the National Academy of Sciences, 120(30): e2305016120, 2023. 4", + "[21] Xiaoxiao Guo, Hui Wu, Yu Cheng, Steven Rennie, Gerald Tesauro, and Rogerio Feris. Dialog-based interactive image retrieval. In Advances in Neural Information Processing Systems, 2018. 3", + "[22] Yuxin Hou, Eleonora Vig, Michael Donoser, and Loris Bazzani. Learning attribute-driven disentangled representations for interactive fashion retrieval. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12147-12157, 2021. 3", + "[23] S. Irene, A. John Prakash, and V. Rhymend Uthariaraj. Person search over security video surveillance systems using deep learning methods: A review. Image and Vision Computing, 143:104930, 2024. 1", + "[24] Ding Jiang and Mang Ye. Cross-modal implicit relation reasoning and aligning for text-to-image person retrieval. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2787-2797, 2023. 2, 3, 6, 7", + "[25] Gi-Cheon Kang, Sungdong Kim, Jin-Hwa Kim, Donghyun Kwak, and Byoung-Tak Zhang. The dialog must go on: Improving visual dialog via generative self-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6746-6756, 2023. 3", + "[26] Saehyung Lee, Sangwon Yu, Junsung Park, Jihun Yi, and Sungroh Yoon. Interactive text-to-image retrieval with large" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "3960", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "language models: A plug-and-play approach. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics, pages 791-809, 2024. 3", + "[27] Matan Levy, Rami Ben-Ari, Nir Darshan, and Dani Lischinski. Chatting makes perfect: Chat-based image retrieval. In Advances in Neural Information Processing Systems, pages 61437-61449, 2023. 3", + "[28] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi. Align before fuse: Vision and language representation learning with momentum distillation. In Advances in Neural Information Processing Systems, pages 9694-9705, 2021. 3", + "[29] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. BLIP: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In Proceedings of the 39th International Conference on Machine Learning, pages 12888-12900, 2022. 3", + "[30] Shuang Li, Tong Xiao, Hongsheng Li, Bolei Zhou, Dayu Yue, and Xiaogang Wang. Person search with natural language description. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2017. 2, 3, 4, 5", + "[31] Shenshen Li, Chen He, Xing Xu, Fumin Shen, Yang Yang, and Heng Tao Shen. Adaptive uncertainty-based learning for text-based person retrieval. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3172-3180, 2024. 7", + "[32] Jian Liang, Wenke Huang, Guancheng Wan, Qu Yang, and Mang Ye. Lorasculpt: Sculpting lora for harmonizing general and specialized knowledge in multimodal large language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2025. 3", + "[33] Kaiqu Liang and Samuel Albanie. Simple baselines for interactive video retrieval with questions and answers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11091-11101, 2023. 3", + "[34] Fangyi Liu, Mang Ye, and Bo Du. Learning a generalizable re-identification model from unlabelled data with domain-agnostic expert. Visual Intelligence, 2(1):28, 2024. 1", + "[35] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In Advances in Neural Information Processing Systems, pages 34892-34916, 2023. 2, 4", + "[36] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. 6", + "[37] Ze Liu, Han Hu, Yutong Lin, Zhuliang Yao, Zhenda Xie, Yixuan Wei, Jia Ning, Yue Cao, Zheng Zhang, Li Dong, Furu Wei, and Baining Guo. Swin transformer v2: Scaling up capacity and resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12009-12019, 2022. 6", + "[38] Renze Lou, Kai Zhang, and Wenpeng Yin. A comprehensive survey on instruction following. arXiv preprint arXiv:2303.10475, 2023. 3", + "[39] Kai Niu, Yanyi Liu, Yuzhou Long, Yan Huang, Liang Wang, and Yanning Zhang. An overview of text-based person" + ], + "bbox": [ + 91, + 92, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "search: Recent advances and future directions. IEEE Transactions on Circuits and Systems for Video Technology, 34(9): 7803-7819, 2024. 1, 3", + "[40] OpenAI. Gpt-4 technical report, 2023. 3", + "[41] OpenAI. Chatgpt. https://openai.com/blog/chatgpt, 2023.3", + "[42] Yang Qin, Yingke Chen, Dezhong Peng, Xi Peng, Joey Tianyi Zhou, and Peng Hu. Noisy-correspondence learning for text-to-image person re-identification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 27197-27206, 2024. 2, 3, 7", + "[43] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. 3", + "[44] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In Proceedings of the 38th International Conference on Machine Learning, pages 8748-8763, 2021. 2, 3, 5, 7", + "[45] Zhiyin Shao, Xinyu Zhang, Meng Fang, Zhifeng Lin, Jian Wang, and Changxing Ding. Learning granularity-unified representations for text-to-image person re-identification. In Proceedings of the 30th ACM International Conference on Multimedia, pages 5566–5574, 2022. 3", + "[46] Qwen Team. Qwen2.5: A party of foundation models. https://qwenlm.github.io/blog/qwen2.5, 2024.4", + "[47] Jason Wei, Yi Tay, Rishi Bommasani, Colin Raffel, Barret Zoph, Sebastian Borgeaud, Dani Yogatama, Maarten Bosma, Denny Zhou, Donald Metzler, et al. Emergent abilities of large language models. arXiv preprint arXiv:2206.07682, 2022.3", + "[48] Hong Wu, Hanqing Lu, and Songde Ma. Willhunter: interactive image retrieval with multilevel relevance. In Proceedings of the 17th International Conference on Pattern Recognition, pages 1009-1012, 2004. 3", + "[49] Yingjia Xu, Mengxia Wu, Zixin Guo, Min Cao, Mang Ye, and Jorma Laaksonen. Efficient text-to-video retrieval via multi-modal multi-tags derived pre-screening. Visual Intelligence, 3(1):1-13, 2025. 3", + "[50] Jingfeng Yang, Hongye Jin, Ruixiang Tang, Xiaotian Han, Qizhang Feng, Haoming Jiang, Shaochen Zhong, Bing Yin, and Xia Hu. Harnessing the power of llms in practice: A survey on chatgpt and beyond. ACM Transactions on Knowledge Discovery from Data, 18(6):1-32, 2024. 2, 4", + "[51] Jinxi Yang, He Li, Bo Du, and Mang Ye. Cheb-gr: Rethinking k-nearest neighbor search in re-ranking for person re-identification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2025. 1", + "[52] Shuyu Yang, Yinan Zhou, Zhedong Zheng, Yaxiong Wang, Li Zhu, and Yujiao Wu. Towards unified text-based person retrieval: A large-scale multi-attribute and language search benchmark. In Proceedings of the 31st ACM International Conference on Multimedia, pages 4492–4501, 2023. 7" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "3961", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[53] Mang Ye, Jianbing Shen, Gaojie Lin, Tao Xiang, Ling Shao, and Steven C. H. Hoi. Deep learning for person re-identification: A survey and outlook. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(6):2872-2893, 2022. 1", + "[54] Mang Ye, Shuoyi Chen, Chenyue Li, Wei-Shi Zheng, David Crandall, and Bo Du. Transformer for object re-identification: A survey. International Journal of Computer Vision, pages 1–31, 2024. 1", + "[55] Mang Ye, Xuankun Rong, Wenke Huang, Bo Du, Nenghai Yu, and Dacheng Tao. A survey of safety on large vision-language models: Attacks, defenses and evaluations. arXiv preprint arXiv:2502.14881, 2025. 3", + "[56] Mang Ye, Zesen Wu, and Bo Du. Dual-level matching with outlier filtering for unsupervised visible-infrared person re-identification. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025. 1", + "[57] Chiyuan Zhang, Samy Bengio, Moritz Hardt, Benjamin Recht, and Oriol Vinyals. Understanding deep learning (still) requires rethinking generalization. Communications of the ACM, 64(3):107-115, 2021. 4", + "[58] Ying Zhang and Hutchuan Lu. Deep cross-modal projection learning for image-text matching. In Proceedings of the European Conference on Computer Vision, pages 686-701, 2018. 2", + "[59] Wayne Xin Zhao, Kun Zhou, Junyi Li, Tianyi Tang, Xiaolei Wang, Yupeng Hou, Yingqian Min, Beichen Zhang, Junjie Zhang, Zican Dong, et al. A survey of large language models. arXiv preprint arXiv:2303.18223, 2023. 3", + "[60] Zhedong Zheng, Liang Zheng, Michael Garrett, Yi Yang, Mingliang Xu, and Yi-Dong Shen. Dual-path convolutional image-text embeddings with instance loss. ACM Transactions on Multimedia Computing, Communications, and Applications, 16(2):1-23, 2020. 3", + "[61] Aichun Zhu, Zijie Wang, Yifeng Li, Xili Wan, Jing Jin, Tian Wang, Fangqiang Hu, and Gang Hua. Dssl: Deep surroundings-person separation learning for text-based person retrieval. In Proceedings of the 29th ACM International Conference on Multimedia, page 209-217, 2021. 2, 4", + "[62] Jialong Zuo, Hanyu Zhou, Ying Nie, Feng Zhang, Tianyu Guo, Nong Sang, Yunhe Wang, and Changxin Gao. Ufinebench: Towards text-based person retrieval with ultra-fine granularity. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22010-22019, 2024. 2, 4" + ], + "bbox": [ + 91, + 92, + 480, + 724 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "3962", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Chat-based Person Retrieval via Dialogue-Refined Cross-Modal Alignment/6896dcb7-a954-4e52-90ab-4e44608cdd32_model.json b/2025/Chat-based Person Retrieval via Dialogue-Refined Cross-Modal Alignment/6896dcb7-a954-4e52-90ab-4e44608cdd32_model.json new file mode 100644 index 0000000000000000000000000000000000000000..2715fed86d2ca90e290abe53e40f86ec9d92d47b --- /dev/null +++ b/2025/Chat-based Person Retrieval via Dialogue-Refined Cross-Modal Alignment/6896dcb7-a954-4e52-90ab-4e44608cdd32_model.json @@ -0,0 +1,2576 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.122, + 0.131, + 0.877, + 0.155 + ], + "angle": 0, + "content": "Chat-based Person Retrieval via Dialogue-Refined Cross-Modal Alignment" + }, + { + "type": "text", + "bbox": [ + 0.227, + 0.18, + 0.777, + 0.2 + ], + "angle": 0, + "content": "Yang Bai\\(^{1,2}\\), Yucheng Ji\\(^{3}\\), Min Cao\\(^{3}\\), Jinqiao Wang\\(^{2,4}\\), Mang Ye\\(^{1*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.2, + 0.757, + 0.234 + ], + "angle": 0, + "content": "\\(^{1}\\)National Engineering Research Center for Multimedia Software, School of Computer Science, Wuhan University, Wuhan, China" + }, + { + "type": "text", + "bbox": [ + 0.353, + 0.234, + 0.646, + 0.251 + ], + "angle": 0, + "content": "\\(^{2}\\)Wuhan AI Research, Wuhan, China" + }, + { + "type": "text", + "bbox": [ + 0.17, + 0.251, + 0.828, + 0.288 + ], + "angle": 0, + "content": "\\(^{3}\\)School of Computer Science and Technology, Soochow University, Suzhou, China \n\\(^{4}\\)Institute of Automation, Chinese Academy of Science, Beijing, China" + }, + { + "type": "text", + "bbox": [ + 0.375, + 0.29, + 0.623, + 0.305 + ], + "angle": 0, + "content": "{ybai.ai, yemang}@whu.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.328, + 0.308, + 0.666, + 0.321 + ], + "angle": 0, + "content": "https://github.com/Flame-Chasers/DiaNA" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.357, + 0.327, + 0.373 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.389, + 0.485, + 0.797 + ], + "angle": 0, + "content": "Traditional text-based person retrieval (TPR) relies on a single-shot text as query to retrieve the target person, assuming that the query completely captures the user's search intent. However, in real-world scenarios, it can be challenging to ensure the information completeness of such single-shot text. To address this limitation, we propose chat-based person retrieval (ChatPR), a new paradigm that takes an interactive dialogue as query to perform the person retrieval, engaging the user in conversational context to progressively refine the query for accurate person retrieval. The primary challenge in ChatPR is the lack of available dialogue-image paired data. To overcome this challenge, we establish ChatPedes, the first dataset designed for ChatPR, which is constructed by leveraging large language models to automate the question generation and simulate user responses. Additionally, to bridge the modality gap between dialogues and images, we propose a dialogue-refined cross-modal alignment (DiaNA) framework, which leverages two adaptive attribute refiner modules to bottleneck the conversational and visual information for fine-grained cross-modal alignment. Moreover, we propose a dialogue-specific data augmentation strategy, random round retaining, to further enhance the model's generalization ability across varying dialogue lengths. Extensive experiments demonstrate that DiaNA significantly outperforms existing TPR approaches, highlighting the effectiveness of conversational interactions for person retrieval." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.824, + 0.222, + 0.839 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.849, + 0.483, + 0.88 + ], + "angle": 0, + "content": "Given the paramount importance of ensuring public safety through intelligent video surveillance, there is a substan" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.355, + 0.907, + 0.426 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.574, + 0.428, + 0.846, + 0.441 + ], + "angle": 0, + "content": "(a) Illustration of text-based person retrieval (TPR)." + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.449, + 0.907, + 0.626 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.564, + 0.628, + 0.856, + 0.641 + ], + "angle": 0, + "content": "(b) Illustration of chat-based person retrieval (ChatPR)." + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.648, + 0.907, + 0.76 + ], + "angle": 0, + "content": "Figure 1. Comparison of TPR and ChatPR. (a) TPR relies on a single-shot text as the query, which fails to completely capture users' search intent due to the absence of several details. (b) ChatPR, in contrast, engages users in interactive conversations, prompting them to offer additional clues and clarify the negative intents. This interaction allows a progressive refinement of the query, thereby enhancing the retrieval accuracy. The correct image is marked with green rectangles." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.909, + 0.903 + ], + "angle": 0, + "content": "tial need to develop systems capable of accurately retrieving a person of interest from a large image gallery. In this context, text-based person retrieval (TPR) [23, 39, 53] employs free-form natural language as the query to perform the retrieval, effectively addressing the limitation that visual queries, such as person images (referred to as person re-identification [34, 51, 54, 56]), are not always readily available in real-world scenarios. Consequently, TPR has" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.293, + 0.901 + ], + "angle": 0, + "content": "*Corresponding Author: Mang Ye" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3952" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.482, + 0.108 + ], + "angle": 0, + "content": "gained ever-rising attention in recent years [24, 42, 50, 58]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.11, + 0.483, + 0.412 + ], + "angle": 0, + "content": "Existing methods in TPR typically rely on a single-shot text query, assuming that the query comprehensively captures the user's search intent. However, in practical applications, it can be challenging to ensure the completeness of information in a single-shot text query for person retrieval. As illustrated in Fig. 1a, the user provides an incomplete single-shot text query with the absence of several details (e.g., \"bag\"), leading to an unsatisfactory retrieval. Alternatively, engaging in multiple rounds of iterative interaction using a question-answer format allows for the conveyance of more detailed information. Recognizing this limitation of single-shot text query in TPR, we introduce a new retrieval paradigm: chat-based person retrieval (ChatPR), which takes interactive dialogues as the query to facilitate a more nuanced and comprehensive interpretation of users' search intent for accurate person retrieval. As shown in Fig. 1b, ChatPR engages users in conversational interaction, allowing the retrieval system to progressively refine the query by actively soliciting additional information, thereby achieving more accurate retrieval." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.415, + 0.483, + 0.612 + ], + "angle": 0, + "content": "ChatPR presents three advantages over traditional TPR. Firstly, it aligns more closely with real-world applications, in which multiple rounds of iterative interaction through a question-answer format provide a more natural way to query the target person. Secondly, the question-answer format actively prompts users for additional information through conversational interactions, allowing the query to encompass more nuanced details. Lastly, it can explicitly capture the user's negative retrieval intents, which are often overlooked in single-shot text queries but are crucial for filtering out irrelevant targets and narrowing the search scope. These unique features highlight the necessity of ChatPR to leverage interactive dialogues for accurate person retrieval." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.614, + 0.483, + 0.902 + ], + "angle": 0, + "content": "However, when addressing the ChatPR task, we are faced with two main challenges: ① Lack of dialogue-image paired data: While the research community has witnessed an increase in the availability of text-image paired datasets for TPR [12, 30, 61, 62], there remains an absence of dialogue-image paired dataset. This data deficiency poses a significant barrier to the research and development of ChatPR. ② Complexity of cross-modal alignment: In contrast to the single-shot text, dialogue data presents a more complex structure and longer context. This poses a significant challenge for the traditional TPR models in understanding conversational dialogues, since their text encoder, such as BERT [11] or CLIP [44], is predominantly pretrained on single-text corpora, thereby posing an obstacle to the subsequent cross-modal alignment. Additionally, dialogues tend to have sparser information density, necessitating multiple rounds of interaction to identify specific attributes, which further impedes the cross-modal alignment between dialogue and images." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.348 + ], + "angle": 0, + "content": "In response to the first challenge, considering that collecting such data via human crowd-sourcing is not only time-consuming but also lacks standardized and well-defined guidelines [35], we propose leveraging the advanced instruction-following capabilities of large language models (LLMs) [8, 14] to convert existing image-text pairs into structured dialogue-image formats. While LLM-based dialogue generation is highly efficient, the resulting data inevitably introduces noise. To address this, we introduce a diverse set of specialized evaluators, composed of LLMs with meticulously crafted instructions, to assess the relevance and quality of each dialogue round. These individual evaluation results are then integrated to facilitate data cleaning. As a result, we construct ChatPedes, the first dataset specifically tailored for ChatPR. This dataset offers a robust benchmark for advancing person retrieval within real-world conversational contexts." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.35, + 0.905, + 0.623 + ], + "angle": 0, + "content": "To address the second challenge, we propose a dialogue-refined cross-modal alignment (DiaNA) framework, which consists of dual encoders to comprehensively understand dialogues and images, respectively. Given the natural affinity of LLMs in handling dialogue, we adopt the open-source Llama 3 [14] as the dialogue encoder to effectively comprehend the multi-round interactions. Additionally, DiaNA incorporates two adaptive attribute refiners on the dialogue and image branches, which leverage attribute queries to respectively bottleneck the conversational and visual information. This design establishes a fine-grained cross-modal alignment, thereby enhancing the effectiveness of DiaNA. Furthermore, we propose a dialogue-specific data augmentation strategy, random round retaining (R3), to simulate the variability in dialogue lengths in real-world interactions, where complete information is not always available. This strategy further enhances DiaNA's generalization ability across varying lengths of dialogue context." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.625, + 0.905, + 0.654 + ], + "angle": 0, + "content": "Overall, the primary contributions of this work can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.657, + 0.905, + 0.732 + ], + "angle": 0, + "content": "- Recognizing the limitations of single-shot text query in text-based person retrieval, we introduce a new retrieval paradigm: chat-based person retrieval (ChatPR), which engages users in interactive conversations to progressively refine their queries for accurate person retrieval." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.733, + 0.905, + 0.775 + ], + "angle": 0, + "content": "- To fill in the data lack, we establish the first dialogue-image paired dataset, ChatPedes, offering the community a robust benchmark for ChatPR." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.778, + 0.905, + 0.853 + ], + "angle": 0, + "content": "- To address the challenges in cross-modal alignment, we propose a dialogue-refined cross-modal alignment (Di-ana) framework to enable fine-grained dialogue-image alignment, and introduce a random round retaining strategy to further boost the model's generalization ability." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.854, + 0.905, + 0.898 + ], + "angle": 0, + "content": "- Extensive experiments demonstrate the effectiveness of DiaNA in handling multi-round dialogues to improve person retrieval performance over conversational contexts." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.657, + 0.905, + 0.898 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3953" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.09, + 0.232, + 0.105 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.119, + 0.342, + 0.135 + ], + "angle": 0, + "content": "2.1. Text-based Person Retrieval" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.146, + 0.484, + 0.388 + ], + "angle": 0, + "content": "Text-based person retrieval [39] aims to find the specific person image based on a textual description. Since Li et al. [30] introduced this task, the community has witnessed its growing flourish in recent years. The studies initially depend on simple global alignment [17, 60], and then gradually evolve to multi-granularity correspondences [7, 19, 45]. More recently, a growing number of works [1, 2, 5, 18, 24, 42] have pivoted towards integrating visual-language pretraining models [28, 29, 44] to leverage general cross-modal knowledge for this fine-grained retrieval. Among them, IRRA [24] incorporates a multi-modal interaction encoder on CLIP [44] to facilitate implicit cross-modal relation reasoning. Furthermore, TBPS-CLIP [5] provides a comprehensive empirical study on the application of CLIP in this field from the perspectives of data augmentation, loss function and training tricks." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.393, + 0.484, + 0.528 + ], + "angle": 0, + "content": "Despite these advancements, existing methods implicitly assume that the single text query fully encapsulates the user's search intent. However, in real-world scenarios, user-provided queries often lack sufficient details and complete information, necessitating iterative refinement for satisfactory retrieval. To address this limitation, this paper introduces a novel chat-based person retrieval paradigm, which engages users in interactive dialogues to progressively refine their queries for more accurate and effective retrieval." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.552, + 0.39, + 0.567 + ], + "angle": 0, + "content": "2.2. Interactive Cross-Modal Retrieval" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.579, + 0.483, + 0.821 + ], + "angle": 0, + "content": "There has been a significant progress in interactive cross-modal retrieval systems, drawing inspiration from visual dialogue [10, 25]. These systems engage users in a dynamic feedback loop towards the desired retrieval target. Typically, user feedback can be categorized into two types: relevance feedback and difference feedback. Relevance feedback [48] involves users assigning scores to the retrieved results based on their relevance to the query, allowing the system to re-rank its retrieval list accordingly. In contrast, difference feedback [4, 33] requires users to specify the distinctions between the target image and the reference image through pre-defined attribute tags [22, 49] or open-form descriptions [21]. Moreover, recent methods [26, 27] bridge the feedback with the query through conversational questioning, enabling the users to be pro-actively asked according to the dialogue history in each search attempt." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Different from the aforementioned works, this paper focuses on the specific field of ChatPR, a more fine-grained retrieval task, where users need to provide more comprehensive and discriminative queries to retrieve the target pedestrian images from a vast image gallery." + }, + { + "type": "title", + "bbox": [ + 0.606, + 0.096, + 0.816, + 0.109 + ], + "angle": 0, + "content": "Dialogue Construction Instruction" + }, + { + "type": "title", + "bbox": [ + 0.565, + 0.12, + 0.651, + 0.132 + ], + "angle": 0, + "content": "Task Definition" + }, + { + "type": "text", + "bbox": [ + 0.536, + 0.134, + 0.884, + 0.179 + ], + "angle": 0, + "content": "Act as an Assistant and a User to convert the giving captions of the pedestrian into a question-answer format. You need to integrate the information from the both captions. The Assistant's goal is to reveal as many details as possible ... details not mentioned in the captions." + }, + { + "type": "title", + "bbox": [ + 0.569, + 0.191, + 0.627, + 0.202 + ], + "angle": 0, + "content": "Workflow" + }, + { + "type": "text", + "bbox": [ + 0.539, + 0.208, + 0.769, + 0.219 + ], + "angle": 0, + "content": "1. Understand the two sentences in the captions." + }, + { + "type": "text", + "bbox": [ + 0.539, + 0.219, + 0.872, + 0.229 + ], + "angle": 0, + "content": "2. Firstly, the Assistant asks the User to describe the person's features." + }, + { + "type": "text", + "bbox": [ + 0.539, + 0.229, + 0.85, + 0.24 + ], + "angle": 0, + "content": "3. The User describes the person's features based on the captions." + }, + { + "type": "list", + "bbox": [ + 0.539, + 0.208, + 0.872, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.539, + 0.251, + 0.882, + 0.273 + ], + "angle": 0, + "content": "6. Don't provide any question or answer that has already been mentioned in previous dialogue history." + }, + { + "type": "title", + "bbox": [ + 0.574, + 0.285, + 0.627, + 0.297 + ], + "angle": 0, + "content": "Examples" + }, + { + "type": "text", + "bbox": [ + 0.541, + 0.302, + 0.886, + 0.335 + ], + "angle": 0, + "content": "- Captions: [\"This man has dark hair and wears a black hoodie with white pattern outside and dark pants ..., \"He's wearing a black hooded sweatshirt with a white lining ...\"" + }, + { + "type": "text", + "bbox": [ + 0.542, + 0.335, + 0.597, + 0.346 + ], + "angle": 0, + "content": "Output:" + }, + { + "type": "list", + "bbox": [ + 0.541, + 0.302, + 0.886, + 0.346 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.555, + 0.347, + 0.734, + 0.357 + ], + "angle": 0, + "content": "Assistant: Please describe the person." + }, + { + "type": "text", + "bbox": [ + 0.555, + 0.357, + 0.887, + 0.377 + ], + "angle": 0, + "content": "User: The person is wearing a black hoodie, green pants, a pair of red and white shoes." + }, + { + "type": "text", + "bbox": [ + 0.555, + 0.378, + 0.736, + 0.39 + ], + "angle": 0, + "content": "Assistant: What's the person's gender?" + }, + { + "type": "text", + "bbox": [ + 0.555, + 0.39, + 0.64, + 0.4 + ], + "angle": 0, + "content": "User: He is male." + }, + { + "type": "text", + "bbox": [ + 0.555, + 0.4, + 0.565, + 0.407 + ], + "angle": 0, + "content": "." + }, + { + "type": "list", + "bbox": [ + 0.555, + 0.347, + 0.887, + 0.407 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.425, + 0.907, + 0.482 + ], + "angle": 0, + "content": "Figure 2. The instruction for dialogue construction, which consists of three parts: task definition, workflow and examples. This design aims to activate LLMs' instruction-following, chain-of-thought and in-context capabilities, respectively." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.492, + 0.738, + 0.508 + ], + "angle": 0, + "content": "2.3. Large Language Models" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.514, + 0.907, + 0.726 + ], + "angle": 0, + "content": "In recent years, the field of artificial intelligence has witnessed a significant leap forward with the advent of large language models (LLMs) [6, 32, 55, 59]. Among the most notable contributions to this domain are the generative pretrained transformer (GPT) series [3, 40, 41, 43]. These models are trained on massive corpora, allowing them to capture a wide range of linguistic patterns and world knowledge. Moreover, the large-scale parameters endow LLMs with emergent abilities [47] that are not present in smaller models [16], such as instruction-following [38] and in-context [13] learning. These capabilities allow LLMs to accurately interpret and execute complex human instructions, as well as understand and utilize context from a few examples to perform tasks effectively." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.726, + 0.907, + 0.815 + ], + "angle": 0, + "content": "In this work, we utilize LLMs to automate question generation and simulate user responses for the construction of ChatPedes. By providing carefully crafted instructions and a few examples, we fully exploit their instruction-following and in-context learning capabilities to create the first benchmark for ChatPR." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.831, + 0.637, + 0.845 + ], + "angle": 0, + "content": "3. Benchmark" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Recognizing the limitations of single-shot text queries with incomplete descriptions in TPR, we propose the ChatPR task, which enables the progressive query refine" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3954" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.114, + 0.089, + 0.465, + 0.181 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.186, + 0.486, + 0.245 + ], + "angle": 0, + "content": "Figure 3. Category distribution of dialogue data from Llama 3 [14] and InternVL [8], evaluated by Qwen 2.5 [46]. Over one-quarter of the data is not matched, highlighting the necessity of data cleaning." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.256, + 0.485, + 0.334 + ], + "angle": 0, + "content": "ment through conversational interactions for accurate person retrieval. However, the primary challenge lies in the lack of dialogue-image paired data. To address this, we construct the ChatPedes dataset through two key phases: dialogue construction and data cleaning." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.346, + 0.303, + 0.363 + ], + "angle": 0, + "content": "3.1. Dialogue Construction" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.369, + 0.483, + 0.627 + ], + "angle": 0, + "content": "In recent years, the research community has witnessed an increase in the availability of text-image paired datasets for TPR [12, 30, 61, 62]. Typically, to create high-quality datasets, a large workforce is employed to manually annotate the images with textual descriptions. However, when it comes to annotating with conversational dialogues via human crowd-sourcing for ChatPR, two major challenges arise: ① Increased difficulty of dialogue annotation: Compared to single-shot text, dialogue data presents a more complex structure and longer context, significantly increasing the manual annotation workload. ② Less well-defined guidelines: Unlike textual annotation tasks with clear instructions, formulating appropriate conversations for dialogue annotation lacks well-defined guidelines [35], complicating the annotation process. Consequently, these challenges make it difficult to employ human crowd-sourcing for creating dialogue-image paired datasets." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.628, + 0.483, + 0.809 + ], + "angle": 0, + "content": "Inspired by the success of LLMs in text annotation tasks [20, 50], we leverage their advanced instruction-following capabilities to automate dialogue annotations, based on the well-known text-image paired CUHK-PEDES dataset [30] in TPR. As illustrated in Fig. 4, dialogue construction aims to construct on-going conversations through the interaction between the LLM-based AI agent and the user. During the interaction, the AI agent formulates follow-up questions based on previous dialogue history. The user, also simulated by the LLM, responds with corresponding answers based on the characteristics of the target person, drawing from the existing annotated captions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.484, + 0.902 + ], + "angle": 0, + "content": "In general, it is expected that the dialogue data produced by the LLM-based AI agent contains comprehensive and valuable information, which is advantageous for subsequent model training. Consequently, we empirically design the instructions shown in Fig. 2. Structurally, the instruction consists of three parts: task definition, workflow, and a" + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.089, + 0.905, + 0.189 + ], + "angle": 0, + "content": "
DatasetsIDsImagesTextsDlg.s.RoundsWords
maxminavgmaxminavg
CUHK-PEDES [30]13,00340,20680,440----961523.5
ICFG-PEDES [12]4,10254,52254,522----83937.2
RSTPReid [61]4,10120,50541,010----701126.5
UFine6926 [62]6,92626,20652,412----2183080.8
ChatPedes (Ours)12,00337,128-74,2561717.738812114.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.193, + 0.907, + 0.248 + ], + "angle": 0, + "content": "Table 1. Comparison between our established ChatPedes and the widely-used TPR benchmarks. \"Dlgs.\" refers to dialogues. ChatPedes uses the multi-round dialogues as queries, demonstrating a longer context with conversational interactions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.256, + 0.907, + 0.41 + ], + "angle": 0, + "content": "few manually designed examples, aiming to activate LLMs' instruction-following, chain-of-thought and in-context capabilities to construct high-quality dialogue data. In the dialogue content, the first question provided by the AI agent typically inquires about a holistic description of the target person. Furthermore, to enhance data diversity, we leverage two publicly available LLMs, Llama 3 [14] and InternVL [8], to generate two distinct dialogues for each person image. The experiments detailed in Sec. 5.3.2 substantiate the benefits of this data diversification strategy." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.415, + 0.663, + 0.433 + ], + "angle": 0, + "content": "3.2. Data Cleaning" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.437, + 0.907, + 0.514 + ], + "angle": 0, + "content": "Although LLMs facilitate the dialogue data collection, the generated data inevitably contains noise, which would significantly impair the optimization of the retrieval process [57]. To perform the data cleaning, we first define the following four classes for each dialogue round:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.513, + 0.906, + 0.558 + ], + "angle": 0, + "content": "- Matched: The answer either aligns with the information present in the annotated captions or accurately reflects the absence of such details not mentioned in the captions." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.558, + 0.905, + 0.587 + ], + "angle": 0, + "content": "- Contradictory: The answer contradicts the information present in the captions." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.588, + 0.905, + 0.619 + ], + "angle": 0, + "content": "- Hallucinatory: The answer contains fabricated information that is not present in the captions." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.618, + 0.905, + 0.648 + ], + "angle": 0, + "content": "- Repeated: The question or answer contains information that has been mentioned in earlier dialogue rounds." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.513, + 0.906, + 0.648 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.649, + 0.907, + 0.83 + ], + "angle": 0, + "content": "Subsequently, we employ LLMs, including Qwen 2.5 [46], Llama 3 [14] and InternVL [8], as specialized evaluators through the crafted instructions (detailed in Appendix B.1) to classify each dialogue round into the four predefined categories. As illustrated in Fig. 3, the category distribution estimated by Qwen 2.5 reveals that over one-quarter of the dialogue data is classified as unmatched, highlighting the necessity for data cleaning. To address this, a vote-based ensemble strategy is adopted to perform the data cleaning. Specifically, we calculate the matching score \\( s \\) for each dialogue round, which represents the proportion of the answers labeled as \"Matched\" by the LLMs:" + }, + { + "type": "equation", + "bbox": [ + 0.605, + 0.837, + 0.907, + 0.88 + ], + "angle": 0, + "content": "\\[\ns = \\frac {1}{N _ {q}} \\sum_ {i = 1} ^ {N _ {q}} \\mathbb {1} \\left(a _ {i} = \\text {M a t c h e d}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.886, + 0.907, + 0.902 + ], + "angle": 0, + "content": "where \\(\\mathbb{1}(\\cdot)\\) is identity function, \\(N_{q}\\) denotes the number of" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3955" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.088, + 0.907, + 0.35 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.355, + 0.907, + 0.438 + ], + "angle": 0, + "content": "Figure 4. The overview of ChatPR. Dialogue Construction aims to construct an on-going conversation, in which an AI agent formulates follow-up questions based on previous dialogue history, prompting the user to provide additional information about the target person to be searched. This iterative process forms a comprehensive Dialogue Query for the target person. DiaNA aims to bridge the modality gap between the dialogue and the image, which incorporates two adaptive attribute refiner (AAR) modules by leveraging learnable attribute queries to extract key information for fine-grained cross-modal alignment. During training, random round retaining (R3) strategy simulates the incomplete query in reality to enhance the model's generalization ability." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.45, + 0.483, + 0.525 + ], + "angle": 0, + "content": "the adopted evaluators, and \\( a_{i} \\) represents the \\( i \\)-th evaluator's answer. If the score \\( s \\) is less than half of the number of evaluators \\( N_{q} \\), this round of conversation is regarded as data noise and then discarded from the dialogue. Detailed statistics of data cleaning are provided in Appendix C." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.525, + 0.483, + 0.692 + ], + "angle": 0, + "content": "Finally, we establish ChatPedes, a dialogue-image paired dataset designed for ChatPR. As shown in Tab. 1, ChatPedes comprises 37,128 images from 12,003 identities, with each image accompanied by two generated dialogues. Consistent with the data split in CUHK-PEDES [30], we divide 34,054 image from 11,003 identities into the training set, and 3,074 image from 1,000 identities into the test set. In addition, ChatPedes is characterized by multi-round dialogues and longer context, exhibiting the longest average length of 114.4 words per dialogue among all benchmarks. Visualization examples are presented in Appendix E." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.703, + 0.182, + 0.718 + ], + "angle": 0, + "content": "4. Method" + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.728, + 0.45, + 0.744 + ], + "angle": 0, + "content": "4.1. Dialogue-Refined Cross-Modal Alignment" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.484, + 0.902 + ], + "angle": 0, + "content": "ChatPR engages users in interactive dialogues to progressively refine their queries, where they are pro-actively questioned and prompted to provide more detailed information for more accurate person retrieval. To bridge the modality gap between dialogues and images, we propose a dialogue-refined cross-modal alignment (DiaNA) framework, as illustrated in Fig. 4, which aims to learn a dialogue-image shared latent space, pulling the pairs with consistent semantics closer together, while pushing those with different semantics further apart. DiaNA is composed of dual" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.45, + 0.907, + 0.511 + ], + "angle": 0, + "content": "encoders to comprehend dialogue queries and person images, respectively, along with two adaptive attribute refiners to bottleneck the conversational and visual information for fine-grained cross-modal alignment." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.516, + 0.663, + 0.529 + ], + "angle": 0, + "content": "4.1.1. Dual Encoders" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.535, + 0.906, + 0.64 + ], + "angle": 0, + "content": "In contrast to the single-shot text query in TPR, dialogue data in ChatPR presents a more complex structure and longer context, rendering traditional text processing models, such as BERT [11] and CLIP [44] (text encoder), less suitable for this task. Given the proficiency of LLMs in handling dialogue data, we adopt the publicly accessible Llama 3 [14] to comprehend the dialogues within our framework." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.64, + 0.907, + 0.838 + ], + "angle": 0, + "content": "Formally, we denote a dialogue as \\(\\mathbf{D} = \\{(Q_i,A_i)\\}_{i = 1}^N\\), where \\(Q_{i}\\) and \\(A_{i}\\) represent the \\(i\\)-th round of question and answer in the dialogue with a total of \\(N\\) rounds. We construct the instruction-format input, as depicted in Fig. 5. Following Vicuna [9], we prepend a system message \\(X_{\\mathrm{system - message}}\\) (detailed in Appendix B.2) to activate the instruction-following capability of the LLM, enabling a more comprehensive understanding of the dialogue data. Finally, the entire input is fed into Llama 3, yielding an embedding sequence \\(\\mathcal{W} = \\{w_{bos},w_1,\\dots ,w_{eos}\\}\\), where \\(w_{eos}\\) captures the full semantics of the dialogue due to the decoder-only architecture of Llama 3. Subsequently, the global embedding \\(w_{eos}\\) is mapped into the shared space:" + }, + { + "type": "equation", + "bbox": [ + 0.634, + 0.85, + 0.905, + 0.868 + ], + "angle": 0, + "content": "\\[\nw = \\Phi_ {D} \\left(w _ {e o s}\\right) \\in \\mathbb {R} ^ {d}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.906, + 0.901 + ], + "angle": 0, + "content": "where \\(\\Phi_D\\) is a projector composed of a linear layer, \\(d\\) is the dimension of the space." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3956" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.114, + 0.089, + 0.462, + 0.162 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.167, + 0.483, + 0.208 + ], + "angle": 0, + "content": "Figure 5. The instruction-formatted input sequence. A system message is prepended to prompt Llama 3 in capturing the semantics of the dialogue context." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.217, + 0.484, + 0.34 + ], + "angle": 0, + "content": "For the visual branch, given an image \\(\\mathbf{I}\\), we employ Swin Transformer [36, 37] to encode the image into a visual sequence \\(\\mathcal{V} = \\{v_{1}, v_{2}, \\dots, v_{h \\times w}\\}\\), where \\(h\\) and \\(w\\) represent the height and width of the feature map at the last stage, respectively. Subsequently, a global average pooling layer is adopted to integrate the overall semantics, yielding a global visual embedding \\(v_{g} = \\frac{1}{h \\times w} \\sum_{i=1}^{h \\times w} v_{i}\\). Similarly, we introduce a projector to map \\(v_{g}\\) into the shared space:" + }, + { + "type": "equation", + "bbox": [ + 0.223, + 0.346, + 0.483, + 0.365 + ], + "angle": 0, + "content": "\\[\nv = \\Phi_ {I} (v _ {g}) \\in \\mathbb {R} ^ {d}. \\tag {3}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.373, + 0.327, + 0.387 + ], + "angle": 0, + "content": "4.1.2. Adaptive Attribute Refiner" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.391, + 0.484, + 0.542 + ], + "angle": 0, + "content": "Unlike the single-shot text where content is generally concentrated, dialogue data typically exhibits sparser information density, requiring multiple rounds of interaction to ascertain specific attributes. This sparsity poses a significant challenge in cross-modal alignment, necessitating information refinement to effectively bridge the modality gap between dialogues and images. To address this, we propose two adaptive attribute refiner (AAR) modules on the dialogue and the image encoders to learn the attribute representations for fine-grained cross-modal alignment." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.543, + 0.484, + 0.753 + ], + "angle": 0, + "content": "Specifically, we first introduce a dialogue AAR upon the dialogue encoder. Considering that different persons have varying attributes, we introduce a set of learnable attribute queries \\(\\mathcal{Q} = \\{q_1, q_2, \\dots, q_K\\}\\) to focus on the distinct attributes of each individual, where \\(K\\) denotes the number of attribute queries. By interacting with the dialogue embeddings \\(\\mathcal{W}\\), these attribute queries adaptively extract the person-specific attribute representations from the sparse dialogue content. As illustrated in Fig. 4, the dialogue AAR applies a computationally efficient structure with a single multi-head cross attention (MCA) followed by 2-layer transformer blocks. The interaction between the attribute queries and the dialogue embeddings is performed through the cross-attention mechanism:" + }, + { + "type": "equation", + "bbox": [ + 0.155, + 0.761, + 0.483, + 0.779 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathcal {W}} = \\operatorname {T r a n s f o r m e r} (\\operatorname {M C A} (\\mathcal {Q}, \\mathcal {W}, \\mathcal {W})), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.786, + 0.483, + 0.816 + ], + "angle": 0, + "content": "where \\(\\hat{\\mathcal{W}} = \\{\\hat{w}_1,\\hat{w}_2,\\dots ,\\hat{w}_K\\}\\) indicates the adaptively refined dialogue attribute embeddings." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.817, + 0.484, + 0.876 + ], + "angle": 0, + "content": "In parallel, we symmetrically introduce a visual AAR to extract the visual attributes from the image, mitigating the impacts of the inherent redundancy and background noise in visual information:" + }, + { + "type": "equation", + "bbox": [ + 0.165, + 0.884, + 0.483, + 0.902 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathcal {V}} = \\operatorname {T r a n s f o r m e r} (\\operatorname {M C A} (\\mathcal {Q}, \\mathcal {V}, \\mathcal {V})), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.091, + 0.906, + 0.168 + ], + "angle": 0, + "content": "where \\(\\hat{\\mathcal{V}} = \\{\\hat{v}_1,\\hat{v}_2,\\dots ,\\hat{v}_K\\}\\) denotes the visual attribute embeddings. Furthermore, we leverage a parameter-sharing mechanism between the two AAR modules, which not only reduces the parameter amount but also mitigates their modality gap [15] between dialogues and images." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.178, + 0.738, + 0.194 + ], + "angle": 0, + "content": "4.2. Optimization Objectives" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.2, + 0.906, + 0.277 + ], + "angle": 0, + "content": "Incorporating the dual encoders and the AAR modules, we construct DiaNA, which is then optimized through both global (coarse-grained) and local (fine-grained) alignment. Specifically, we use the cosine similarity to measure the global alignment between the dialogue and the image:" + }, + { + "type": "equation", + "bbox": [ + 0.637, + 0.287, + 0.906, + 0.322 + ], + "angle": 0, + "content": "\\[\ns _ {g} (\\mathbf {D}, \\mathbf {I}) = \\frac {w ^ {\\mathrm {T}} v}{\\| w \\| \\| v \\|}. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.333, + 0.906, + 0.379 + ], + "angle": 0, + "content": "Furthermore, the average cosine distance among the refined attribute embeddings is utilized to measure the local semantics consistency between the dialogue and the image:" + }, + { + "type": "equation", + "bbox": [ + 0.609, + 0.391, + 0.906, + 0.432 + ], + "angle": 0, + "content": "\\[\ns _ {l} (\\mathbf {D}, \\mathbf {I}) = \\frac {1}{K} \\sum_ {i = 1} ^ {K} \\frac {\\hat {w} _ {i} ^ {\\mathrm {T}} \\hat {v} _ {i}}{\\| \\hat {w} _ {i} \\| \\| \\hat {v} _ {i} \\|.} \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.444, + 0.906, + 0.52 + ], + "angle": 0, + "content": "According to the global and local similarities between the dialogue and the image, we construct the overall optimization object by employing the normalized image-text contrastive (N-ITC) loss [5] and the similarity distribution matching (SDM) loss [24]:" + }, + { + "type": "equation", + "bbox": [ + 0.585, + 0.53, + 0.905, + 0.569 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\underbrace {\\mathcal {L} _ {g} ^ {n i t c} + \\mathcal {L} _ {g} ^ {s d m}} _ {\\text {g l o b a l a l i n g m e n t}} + \\underbrace {\\mathcal {L} _ {l} ^ {n i t c} + \\mathcal {L} _ {l} ^ {s d m}} _ {\\text {l o c a l a l i n g m e n t}}. \\tag {8}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.593, + 0.836, + 0.609 + ], + "angle": 0, + "content": "4.3. Dialogue-specific Data Augmentation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.616, + 0.906, + 0.736 + ], + "angle": 0, + "content": "In real-world scenarios, complete dialogues that capture users' full search intent are not always available, since obtaining these complete dialogues typically requires numerous rounds of interaction, which can be not user-friendly. To enhance the model's generalization ability over dialogues of varying rounds, we propose a dialogue-specific data augmentation strategy, random round retaining (R3), which simulates the inherent variability of the dialogue in practice." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.737, + 0.906, + 0.781 + ], + "angle": 0, + "content": "Specifically, at each training iteration, for a complete dialogue \\(\\mathbf{D} = \\{(Q_i,A_i)\\}_{i = 1}^N\\), only the initial random \\(i\\) rounds are retained and fed into the model:" + }, + { + "type": "equation", + "bbox": [ + 0.567, + 0.794, + 0.905, + 0.813 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {D}} = \\left\\{\\left(Q _ {i}, A _ {i}\\right) \\right\\} _ {i = 1} ^ {n}, \\quad n \\sim \\operatorname {r a n d i n t} (1, N), \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.824, + 0.906, + 0.901 + ], + "angle": 0, + "content": "where \\(\\operatorname{randint}(1, N)\\) denotes a discrete uniform distribution with equal probability for each integer in range \\([1, N]\\). The retained \\(\\hat{\\mathbf{D}}\\) reflects the partial availability of information in practical scenarios, effectively enhancing the model's generalization ability over varying dialogue lengths." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3957" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.143, + 0.089, + 0.865, + 0.223 + ], + "angle": 0, + "content": "
MethodRefPret. Dataw/o Data Cleaningw/ Data Cleaning
R-1R-5R-10mAPR-1R-5R-10mAP
IRRA [24]CVPR'23WIT46.0865.6073.0242.9048.2667.9675.6544.64
TBPS-CLIP [5]AAAI'24WIT47.7466.4173.7843.8050.2169.2776.4345.44
RDE [42]CVPR'24WIT50.2068.2575.1345.9452.7770.6177.4747.98
APTM [52]MM'23MALS39.6256.9964.2336.5741.1557.9165.1638.00
AUL [31]AAAI'24MALS40.0557.0164.4137.0242.0159.2166.0138.56
DiaNA (Ours)-MALS73.7289.6694.2465.4075.6790.6894.7166.89
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.225, + 0.907, + 0.255 + ], + "angle": 0, + "content": "Table 2. Comparison with state-of-the-art methods on ChatPedes. \"Pret. Data\" denotes pretraining data. All reported results of other methods are reproduced using their publicly available code." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.264, + 0.224, + 0.281 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.289, + 0.484, + 0.336 + ], + "angle": 0, + "content": "In this section, we conduct extensive experiments to verify the effectiveness of our method. The implementation details are provided in Appendix A." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.346, + 0.281, + 0.361 + ], + "angle": 0, + "content": "5.1. Evaluation Protocol" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.369, + 0.484, + 0.462 + ], + "angle": 0, + "content": "We adopt the widely-used Rank-K (R-K for short, \\( \\mathrm{K} = 1,5,10 \\)) metric in our experiments, which denotes the percentage of successful retrieval that the ground-truth image is found within the top K ranked results. In addition, we also employ the mean average precision (mAP) as an auxiliary metric to evaluate the overall retrieval performance." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.47, + 0.383, + 0.485 + ], + "angle": 0, + "content": "5.2. Comparison with SOTA Methods" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.493, + 0.483, + 0.765 + ], + "angle": 0, + "content": "We compare our method against five state-of-the-art (SOTA) baselines: IRRA [24], TBPS-CLIP [5], RDE [42], APTM [52] and AUL [31]. Among these, IRRA, TBPS-CLIP and RDE are built on the vision-language foundation model CLIP [44], which is pretrained on a large-scale dataset WIT with 400 million image-text paired data collected from the Internet. APTM and AUL are pretrained on the synthetic text-image paired dataset MALS [52]. It is worth noting that these baselines are originally developed for the TPR task. We reproduce their performance on ChatPedes dataset using their open-source codes, treating the multi-turn dialogue data as a single long-form text input. As shown in Tab. 7, regardless of whether data cleaning is applied, our method outperforms these SOTA baselines by a large margin, while these baselines demonstrate a suboptimal retrieval accuracy, indicating their limitations in effectively handling the complex structure and longer contextual dependencies of dialogue data." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.766, + 0.484, + 0.901 + ], + "angle": 0, + "content": "In addition, we further conduct extensive comparative experiments over the dialogue round, simulating real-world scenarios where users are pro-actively questioned to progressively refine their queries through conversational interactions. This process is exemplified in Appendix E.1. As shown in Fig. 6, it can be observed that: ① As the dialogue progresses, the retrieval performance improves across all methods. However, the performance of the baselines becomes saturated after the initial few rounds of interaction," + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.264, + 0.907, + 0.403 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.542, + 0.407, + 0.878, + 0.421 + ], + "angle": 0, + "content": "Figure 6. Performance comparison over dialogue round." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.435, + 0.907, + 0.664 + ], + "angle": 0, + "content": "which we attribute to their limitations in processing longer and more complex dialogues. In contrast, our method presents a significantly pronounced improvement as the dialogue evolves on account of our carefully crafted design for handling dialogue data. This comprehensive comparison demonstrates the efficacy of our method and showcases its superior capability in handling the intricacies of dialogue-image alignment. Additionally, we also notice that our method achieves competitive performance at the first round of dialogue. This round typically involves a holistic description of the target person, functioning similarly to the single-shot text in the traditional TPR task. The experimental comparison also indicates the promising generalization capabilities of our method in TPR. Further detailed experimental results on TPR benchmarks are provided in Appendix D." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.676, + 0.667, + 0.692 + ], + "angle": 0, + "content": "5.3. Ablation Study" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.7, + 0.773, + 0.715 + ], + "angle": 0, + "content": "5.3.1. Effectiveness of Data Cleaning" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.907, + 0.903 + ], + "angle": 0, + "content": "During dialogue construction, we leverage LLMs to generate dialogue data. However, a major limitation of this automatic generation through LLMs lies in the inevitable introduction of noise, which can be classified into three categories: Contradictory, Hallucinatory and Repeated, as delineated in Sec. 3.2. To evaluate the impact of the noise, we conduct comparative experiments with and without data cleaning. Tab. 7 presents the performance comparison with the five SOTA baselines, where all approaches exhibit a significant improvement. In particular, RDE achieves a notable increase of \\(2.57\\%\\) at R-1, demonstrating the effectiveness of data cleaning and the importance of high-quality data." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3958" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.104, + 0.089, + 0.46, + 0.174 + ], + "angle": 0, + "content": "
Data DiversityR-1R-5R-10mAP
Llama 3InternVL
X✓71.0188.4093.1863.08
✓X71.6288.5893.5963.73
✓✓75.6790.6894.7166.89
" + }, + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.179, + 0.483, + 0.207 + ], + "angle": 0, + "content": "Table 3. Ablation study on data diversity. We employ Llama 3 [14] and InternVL [8] to construct the ChatPedes benchmark." + }, + { + "type": "image", + "bbox": [ + 0.122, + 0.218, + 0.455, + 0.354 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.359, + 0.485, + 0.4 + ], + "angle": 0, + "content": "Figure 7. Performance comparison across a varying number of attribute queries \\( K \\) in AAR modules. Notably, \\( K = 0 \\) indicates that the AAR is not used." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.411, + 0.351, + 0.427 + ], + "angle": 0, + "content": "5.3.2. Effectiveness of Data Diversity" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.431, + 0.484, + 0.644 + ], + "angle": 0, + "content": "We employ Llama 3 [14] and InternVL [8] as different annotators for labeling diverse dialogue data when constructing ChatPedes. Tab. 3 presents the contribution of data diversity to the retrieval performance. It can be observed that: ① Relying on a single LLM yields a suboptimal outcome, with \\(71.01\\%\\) for InternVL and \\(71.62\\%\\) for Llama 3 at R-1. Additionally, Llama 3 slightly outperforms InternVL due to its more advanced capability in dialogue generation. ② The combination of Llama 3 and InternVL leads to a substantial increase across all metrics, demonstrating the importance of data diversity. By combining annotations generated by different LLMs, the established ChatPedes dataset effectively captures a wider variety of dialogue contexts and nuances, which in turn bolsters the model's retrieval capabilities." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.655, + 0.437, + 0.67 + ], + "angle": 0, + "content": "5.3.3. Effectiveness of Adaptive Attribute Refiner" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.674, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Considering the sparse information density in dialogues and the inherent redundancy in images, we propose two adaptive attribute refiner (AAR) modules, which employ a set of learnable attribute-specific queries to adaptively extract the relevant attributes from the dialogue and the image. Fig. 7 illustrates the performance variation over the number of attribute queries \\( K \\), where, notably, \\( K = 0 \\) denotes AAR is discarded. We observe that: 1 Only relying on global alignment without AAR results in a suboptimal retrieval performance, underscoring the importance of key information extraction. 2 When introducing AAR, the performance exhibits slight fluctuations, reaching the peak of \\( 75.67\\% \\) on R-1 and \\( 66.89\\% \\) on mAP at \\( K = 16 \\). Overall, the performance remains relatively stable, indicating the model's robustness when equipped with AAR." + }, + { + "type": "table", + "bbox": [ + 0.525, + 0.09, + 0.895, + 0.206 + ], + "angle": 0, + "content": "
MethodMetricDialogue Round
12468all
DiaNA(w/o R3)26.5034.8953.0766.4073.1675.18
DiaNA(w/ R3)R-129.2537.6455.6067.9973.6775.67
Δ+2.75+2.75+2.53+1.59+0.51+0.49
DiaNA(w/o R3)26.3333.6749.0659.6764.6466.19
DiaNA(w/ R3)mAP28.7836.3751.0560.8865.3266.89
Δ+2.45+2.70+1.99+1.21+0.68+0.70
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.209, + 0.905, + 0.237 + ], + "angle": 0, + "content": "Table 4. Comparison of random round retaining (R3) over dialogue round. \"all\" refers to the complete dialogue." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.25, + 0.857, + 0.266 + ], + "angle": 0, + "content": "5.3.4. Effectiveness of Random Round Retaining" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.269, + 0.907, + 0.513 + ], + "angle": 0, + "content": "The proposed random round retaining (R3) strategy selectively retains a random number of initial rounds of the dialogue during training, simulating the users' incomplete search intent in real-world interactions. As shown in Tab. 4, we observe that: \\( \\mathbf{1} \\) R3 yields performance gains across all dialogue rounds, with gains from \\( 2.75\\% \\) in single-round interactions to \\( 0.49\\% \\) in complete conversations at R-1, demonstrating its robustness and versatility. \\( \\mathbf{2} \\) In particular, R3 brings more substantial improvement at the initial few rounds of dialogues, suggesting that R3 is particularly effective when dialogue information is limited. This is especially crucial in practical scenarios where users expect to quickly retrieve the target person with minimal interaction. Overall, R3 empowers users to obtain accurate retrieval results even in the early stages of interaction, thereby addressing the real-world demands for both efficiency and precision." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.527, + 0.634, + 0.543 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.553, + 0.907, + 0.824 + ], + "angle": 0, + "content": "Recognizing the inherent limitations of traditional TPR where single-shot text queries fail to fully encapsulate users' search intent, we introduce ChatPR, a pioneering paradigm for person retrieval that engages users in interactive dialogues to progressively refine their search queries. For this, we build the first benchmark ChatPedes through LLMs to automate question generation and simulate users' responses. To tackle this task, we propose the DiaNA framework that leverages two adaptive attribute refiners to bottleneck the conversational and visual information for fine-grained cross-modal alignment. Additionally, a novel data augmentation strategy of random round retaining is proposed to mimic the incomplete queries in real-world scenarios, remarkably enhancing the model's generalization ability across varying dialogue lengths. Extensive experiments significantly demonstrate the importance of conversational interaction in person retrieval and the superior effectiveness of our method in ChatPR." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.906, + 0.903 + ], + "angle": 0, + "content": "Acknowledgement. This work is supported by the National Science Foundation of China under Grants (62176188, 62361166629, 62476188). The numerical calculations are supported by the supercomputing system in the Supercomputing Center of Wuhan University." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3959" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.197 + ], + "angle": 0, + "content": "[1] Yang Bai, Min Cao, Daming Gao, Ziqiang Cao, Chen Chen, Zhenfeng Fan, Liqiang Nie, and Min Zhang. Rasa: Relation and sensitivity aware representation learning for text-based person search. In Proceedings of the Thirty-Second International Joint Conference on Artificial Intelligence, pages 555–563, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.2, + 0.484, + 0.268 + ], + "angle": 0, + "content": "[2] Yang Bai, Jingyao Wang, Min Cao, Chen Chen, Ziqiang Cao, Liqiang Nie, and Min Zhang. Text-based person search without parallel image-text data. In Proceedings of the 31st ACM International Conference on Multimedia, pages 757-767, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.271, + 0.482, + 0.297 + ], + "angle": 0, + "content": "[3] Tom B Brown. Language models are few-shot learners. arXiv preprint arXiv:2005.14165, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.299, + 0.482, + 0.381 + ], + "angle": 0, + "content": "[4] Guanyu Cai, Jun Zhang, Xinyang Jiang, Yifei Gong, Lianghua He, Fufu Yu, Pai Peng, Xiaowei Guo, Feiyue Huang, and Xing Sun. Ask&confirm: active detail enriching for cross-modal retrieval with partial query. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1835-1844, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.383, + 0.482, + 0.437 + ], + "angle": 0, + "content": "[5] Min Cao, Yang Bai, Ziyin Zeng, Mang Ye, and Min Zhang. An empirical study of clip for text-based person search. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 465-473, 2024. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.439, + 0.482, + 0.507 + ], + "angle": 0, + "content": "[6] Yupeng Chang, Xu Wang, Jindong Wang, Yuan Wu, Linyi Yang, Kaijie Zhu, Hao Chen, Xiaoyuan Yi, Cunxiang Wang, Yidong Wang, et al. A survey on evaluation of large language models. ACM Transactions on Intelligent Systems and Technology, 15(3):1-45, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.509, + 0.482, + 0.564 + ], + "angle": 0, + "content": "[7] Yuhao Chen, Guoqing Zhang, Yujiang Lu, Zhenxing Wang, and Yuhui Zheng. Tipcb: A simple but effective part-based convolutional baseline for text-based person search. Neurocomputing, 494:171-181, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.565, + 0.482, + 0.648 + ], + "angle": 0, + "content": "[8] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24185-24198, 2024. 2, 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.65, + 0.482, + 0.731 + ], + "angle": 0, + "content": "[9] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhang-hao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yong-hao Zhuang, Joseph E. Gonzalez, Ion Stoica, and Eric P. Xing. Vicuna: An open-source chatbot impressing gpt-4 with \\(90\\%\\) chatgpt quality. https://lmsys.org/blog/2023-03-30-vicuna, 2023.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.733, + 0.482, + 0.801 + ], + "angle": 0, + "content": "[10] Abhishek Das, Satwik Kottur, Khushi Gupta, Avi Singh, Deshraj Yadav, Jose M. F. Moura, Devi Parikh, and Dhruv Batra. Visual dialog. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 326-335, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.804, + 0.482, + 0.843 + ], + "angle": 0, + "content": "[11] Jacob Devlin. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.095, + 0.846, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[12] Zefeng Ding, Changxing Ding, Zhiyin Shao, and Dacheng Tao. Semantically self-aligned network for text-to-image part-aware person re-identification. arXiv preprint arXiv:2107.12666, 2021. 2, 4" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[13] Qingxiu Dong, Lei Li, Damai Dai, Ce Zheng, Zhiyong Wu, Baobao Chang, Xu Sun, Jingjing Xu, and Zhifang Sui. A survey on in-context learning. arXiv preprint arXiv:2301.00234, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.149, + 0.905, + 0.217 + ], + "angle": 0, + "content": "[14] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 2, 4, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.219, + 0.905, + 0.259 + ], + "angle": 0, + "content": "[15] Sedigheh Eslami and Gerard de Melo. Mitigate the gap: Investigating approaches for improving cross-modal alignment in clip. arXiv preprint arXiv:2406.17639, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.261, + 0.905, + 0.315 + ], + "angle": 0, + "content": "[16] Yiyang Fang, Wenke Huang, Guancheng Wan, Kehua Su, and Mang Ye. Emoe: Modality-specific enhanced dynamic emotion experts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.317, + 0.905, + 0.371 + ], + "angle": 0, + "content": "[17] Ammarah Farooq, Muhammad Awais, Fei Yan, Josef Kittler, Ali Akbari, and Syed Safwan Khalid. A convolutional baseline for person re-identification using vision and language descriptions. arXiv preprint arXiv:2003.00808, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.372, + 0.905, + 0.427 + ], + "angle": 0, + "content": "[18] Takuro Fujii and Shuhei Tarashima. Bilma: Bidirectional local-matching for text-based person re-identification. In Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops, pages 2786-2790, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.429, + 0.905, + 0.496 + ], + "angle": 0, + "content": "[19] Chenyang Gao, Guanyu Cai, Xinyang Jiang, Feng Zheng, Jun Zhang, Yifei Gong, Fangzhou Lin, Xing Sun, and Xi-ang Bai. Conditional feature learning based transformer for text-based person search. IEEE Transactions on Image Processing, 31:6097–6108, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.498, + 0.905, + 0.551 + ], + "angle": 0, + "content": "[20] Fabrizio Gilardi, Meysam Alizadeh, and Maël Kubli. Chatgpt outperforms crowd workers for text-annotation tasks. Proceedings of the National Academy of Sciences, 120(30): e2305016120, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.553, + 0.905, + 0.607 + ], + "angle": 0, + "content": "[21] Xiaoxiao Guo, Hui Wu, Yu Cheng, Steven Rennie, Gerald Tesauro, and Rogerio Feris. Dialog-based interactive image retrieval. In Advances in Neural Information Processing Systems, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.609, + 0.905, + 0.677 + ], + "angle": 0, + "content": "[22] Yuxin Hou, Eleonora Vig, Michael Donoser, and Loris Bazzani. Learning attribute-driven disentangled representations for interactive fashion retrieval. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12147-12157, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.679, + 0.905, + 0.734 + ], + "angle": 0, + "content": "[23] S. Irene, A. John Prakash, and V. Rhymend Uthariaraj. Person search over security video surveillance systems using deep learning methods: A review. Image and Vision Computing, 143:104930, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.735, + 0.905, + 0.801 + ], + "angle": 0, + "content": "[24] Ding Jiang and Mang Ye. Cross-modal implicit relation reasoning and aligning for text-to-image person retrieval. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2787-2797, 2023. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.804, + 0.905, + 0.872 + ], + "angle": 0, + "content": "[25] Gi-Cheon Kang, Sungdong Kim, Jin-Hwa Kim, Donghyun Kwak, and Byoung-Tak Zhang. The dialog must go on: Improving visual dialog via generative self-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6746-6756, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[26] Saehyung Lee, Sangwon Yu, Junsung Park, Jihun Yi, and Sungroh Yoon. Interactive text-to-image retrieval with large" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.956 + ], + "angle": 0, + "content": "3960" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.093, + 0.483, + 0.134 + ], + "angle": 0, + "content": "language models: A plug-and-play approach. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics, pages 791-809, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.483, + 0.19 + ], + "angle": 0, + "content": "[27] Matan Levy, Rami Ben-Ari, Nir Darshan, and Dani Lischinski. Chatting makes perfect: Chat-based image retrieval. In Advances in Neural Information Processing Systems, pages 61437-61449, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.193, + 0.482, + 0.262 + ], + "angle": 0, + "content": "[28] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi. Align before fuse: Vision and language representation learning with momentum distillation. In Advances in Neural Information Processing Systems, pages 9694-9705, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.264, + 0.482, + 0.332 + ], + "angle": 0, + "content": "[29] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. BLIP: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In Proceedings of the 39th International Conference on Machine Learning, pages 12888-12900, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.334, + 0.482, + 0.401 + ], + "angle": 0, + "content": "[30] Shuang Li, Tong Xiao, Hongsheng Li, Bolei Zhou, Dayu Yue, and Xiaogang Wang. Person search with natural language description. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2017. 2, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.405, + 0.482, + 0.461 + ], + "angle": 0, + "content": "[31] Shenshen Li, Chen He, Xing Xu, Fumin Shen, Yang Yang, and Heng Tao Shen. Adaptive uncertainty-based learning for text-based person retrieval. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3172-3180, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.462, + 0.482, + 0.531 + ], + "angle": 0, + "content": "[32] Jian Liang, Wenke Huang, Guancheng Wan, Qu Yang, and Mang Ye. Lorasculpt: Sculpting lora for harmonizing general and specialized knowledge in multimodal large language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.532, + 0.482, + 0.588 + ], + "angle": 0, + "content": "[33] Kaiqu Liang and Samuel Albanie. Simple baselines for interactive video retrieval with questions and answers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11091-11101, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.589, + 0.482, + 0.631 + ], + "angle": 0, + "content": "[34] Fangyi Liu, Mang Ye, and Bo Du. Learning a generalizable re-identification model from unlabelled data with domain-agnostic expert. Visual Intelligence, 2(1):28, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.632, + 0.482, + 0.674 + ], + "angle": 0, + "content": "[35] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In Advances in Neural Information Processing Systems, pages 34892-34916, 2023. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.676, + 0.482, + 0.744 + ], + "angle": 0, + "content": "[36] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.746, + 0.482, + 0.828 + ], + "angle": 0, + "content": "[37] Ze Liu, Han Hu, Yutong Lin, Zhuliang Yao, Zhenda Xie, Yixuan Wei, Jia Ning, Yue Cao, Zheng Zhang, Li Dong, Furu Wei, and Baining Guo. Swin transformer v2: Scaling up capacity and resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12009-12019, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.83, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[38] Renze Lou, Kai Zhang, and Wenpeng Yin. A comprehensive survey on instruction following. arXiv preprint arXiv:2303.10475, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[39] Kai Niu, Yanyi Liu, Yuzhou Long, Yan Huang, Liang Wang, and Yanning Zhang. An overview of text-based person" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "search: Recent advances and future directions. IEEE Transactions on Circuits and Systems for Video Technology, 34(9): 7803-7819, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.137, + 0.79, + 0.151 + ], + "angle": 0, + "content": "[40] OpenAI. Gpt-4 technical report, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.154, + 0.905, + 0.181 + ], + "angle": 0, + "content": "[41] OpenAI. Chatgpt. https://openai.com/blog/chatgpt, 2023.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.184, + 0.905, + 0.253 + ], + "angle": 0, + "content": "[42] Yang Qin, Yingke Chen, Dezhong Peng, Xi Peng, Joey Tianyi Zhou, and Peng Hu. Noisy-correspondence learning for text-to-image person re-identification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 27197-27206, 2024. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.256, + 0.905, + 0.297 + ], + "angle": 0, + "content": "[43] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.3, + 0.905, + 0.396 + ], + "angle": 0, + "content": "[44] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In Proceedings of the 38th International Conference on Machine Learning, pages 8748-8763, 2021. 2, 3, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.399, + 0.905, + 0.468 + ], + "angle": 0, + "content": "[45] Zhiyin Shao, Xinyu Zhang, Meng Fang, Zhifeng Lin, Jian Wang, and Changxing Ding. Learning granularity-unified representations for text-to-image person re-identification. In Proceedings of the 30th ACM International Conference on Multimedia, pages 5566–5574, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.471, + 0.905, + 0.511 + ], + "angle": 0, + "content": "[46] Qwen Team. Qwen2.5: A party of foundation models. https://qwenlm.github.io/blog/qwen2.5, 2024.4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.514, + 0.905, + 0.582 + ], + "angle": 0, + "content": "[47] Jason Wei, Yi Tay, Rishi Bommasani, Colin Raffel, Barret Zoph, Sebastian Borgeaud, Dani Yogatama, Maarten Bosma, Denny Zhou, Donald Metzler, et al. Emergent abilities of large language models. arXiv preprint arXiv:2206.07682, 2022.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.586, + 0.905, + 0.642 + ], + "angle": 0, + "content": "[48] Hong Wu, Hanqing Lu, and Songde Ma. Willhunter: interactive image retrieval with multilevel relevance. In Proceedings of the 17th International Conference on Pattern Recognition, pages 1009-1012, 2004. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.644, + 0.905, + 0.699 + ], + "angle": 0, + "content": "[49] Yingjia Xu, Mengxia Wu, Zixin Guo, Min Cao, Mang Ye, and Jorma Laaksonen. Efficient text-to-video retrieval via multi-modal multi-tags derived pre-screening. Visual Intelligence, 3(1):1-13, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.702, + 0.905, + 0.771 + ], + "angle": 0, + "content": "[50] Jingfeng Yang, Hongye Jin, Ruixiang Tang, Xiaotian Han, Qizhang Feng, Haoming Jiang, Shaochen Zhong, Bing Yin, and Xia Hu. Harnessing the power of llms in practice: A survey on chatgpt and beyond. ACM Transactions on Knowledge Discovery from Data, 18(6):1-32, 2024. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.774, + 0.905, + 0.829 + ], + "angle": 0, + "content": "[51] Jinxi Yang, He Li, Bo Du, and Mang Ye. Cheb-gr: Rethinking k-nearest neighbor search in re-ranking for person re-identification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.832, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[52] Shuyu Yang, Yinan Zhou, Zhedong Zheng, Yaxiong Wang, Li Zhu, and Yujiao Wu. Towards unified text-based person retrieval: A large-scale multi-attribute and language search benchmark. In Proceedings of the 31st ACM International Conference on Multimedia, pages 4492–4501, 2023. 7" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "3961" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.16 + ], + "angle": 0, + "content": "[53] Mang Ye, Jianbing Shen, Gaojie Lin, Tao Xiang, Ling Shao, and Steven C. H. Hoi. Deep learning for person re-identification: A survey and outlook. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(6):2872-2893, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.164, + 0.482, + 0.219 + ], + "angle": 0, + "content": "[54] Mang Ye, Shuoyi Chen, Chenyue Li, Wei-Shi Zheng, David Crandall, and Bo Du. Transformer for object re-identification: A survey. International Journal of Computer Vision, pages 1–31, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.221, + 0.482, + 0.275 + ], + "angle": 0, + "content": "[55] Mang Ye, Xuankun Rong, Wenke Huang, Bo Du, Nenghai Yu, and Dacheng Tao. A survey of safety on large vision-language models: Attacks, defenses and evaluations. arXiv preprint arXiv:2502.14881, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.277, + 0.482, + 0.331 + ], + "angle": 0, + "content": "[56] Mang Ye, Zesen Wu, and Bo Du. Dual-level matching with outlier filtering for unsupervised visible-infrared person re-identification. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.333, + 0.482, + 0.387 + ], + "angle": 0, + "content": "[57] Chiyuan Zhang, Samy Bengio, Moritz Hardt, Benjamin Recht, and Oriol Vinyals. Understanding deep learning (still) requires rethinking generalization. Communications of the ACM, 64(3):107-115, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.391, + 0.482, + 0.443 + ], + "angle": 0, + "content": "[58] Ying Zhang and Hutchuan Lu. Deep cross-modal projection learning for image-text matching. In Proceedings of the European Conference on Computer Vision, pages 686-701, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.447, + 0.482, + 0.5 + ], + "angle": 0, + "content": "[59] Wayne Xin Zhao, Kun Zhou, Junyi Li, Tianyi Tang, Xiaolei Wang, Yupeng Hou, Yingqian Min, Beichen Zhang, Junjie Zhang, Zican Dong, et al. A survey of large language models. arXiv preprint arXiv:2303.18223, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.504, + 0.482, + 0.572 + ], + "angle": 0, + "content": "[60] Zhedong Zheng, Liang Zheng, Michael Garrett, Yi Yang, Mingliang Xu, and Yi-Dong Shen. Dual-path convolutional image-text embeddings with instance loss. ACM Transactions on Multimedia Computing, Communications, and Applications, 16(2):1-23, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.574, + 0.482, + 0.642 + ], + "angle": 0, + "content": "[61] Aichun Zhu, Zijie Wang, Yifeng Li, Xili Wan, Jing Jin, Tian Wang, Fangqiang Hu, and Gang Hua. Dssl: Deep surroundings-person separation learning for text-based person retrieval. In Proceedings of the 29th ACM International Conference on Multimedia, page 209-217, 2021. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.645, + 0.482, + 0.726 + ], + "angle": 0, + "content": "[62] Jialong Zuo, Hanyu Zhou, Ying Nie, Feng Zhang, Tianyu Guo, Nong Sang, Yunhe Wang, and Changxin Gao. Ufinebench: Towards text-based person retrieval with ultra-fine granularity. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22010-22019, 2024. 2, 4" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.726 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3962" + } + ] +] \ No newline at end of file diff --git a/2025/Chat-based Person Retrieval via Dialogue-Refined Cross-Modal Alignment/6896dcb7-a954-4e52-90ab-4e44608cdd32_origin.pdf b/2025/Chat-based Person Retrieval via Dialogue-Refined Cross-Modal Alignment/6896dcb7-a954-4e52-90ab-4e44608cdd32_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..d539f085f2d5de8a0af8ac18d61bcd227c47d769 --- /dev/null +++ b/2025/Chat-based Person Retrieval via Dialogue-Refined Cross-Modal Alignment/6896dcb7-a954-4e52-90ab-4e44608cdd32_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fc7f1312599f32fbbc7e75ba8b3b082bfaa6fe5a2ab85ffa4ee56a918a0c862b +size 1464769 diff --git a/2025/Chat-based Person Retrieval via Dialogue-Refined Cross-Modal Alignment/full.md b/2025/Chat-based Person Retrieval via Dialogue-Refined Cross-Modal Alignment/full.md new file mode 100644 index 0000000000000000000000000000000000000000..269379c1a399e16858932681f7f482e2cfc83f75 --- /dev/null +++ b/2025/Chat-based Person Retrieval via Dialogue-Refined Cross-Modal Alignment/full.md @@ -0,0 +1,354 @@ +# Chat-based Person Retrieval via Dialogue-Refined Cross-Modal Alignment + +Yang Bai $^{1,2}$ , Yucheng Ji $^{3}$ , Min Cao $^{3}$ , Jinqiao Wang $^{2,4}$ , Mang Ye $^{1*}$ + +$^{1}$ National Engineering Research Center for Multimedia Software, School of Computer Science, Wuhan University, Wuhan, China + +$^{2}$ Wuhan AI Research, Wuhan, China + +$^{3}$ School of Computer Science and Technology, Soochow University, Suzhou, China + $^{4}$ Institute of Automation, Chinese Academy of Science, Beijing, China + +{ybai.ai, yemang}@whu.edu.cn + +https://github.com/Flame-Chasers/DiaNA + +# Abstract + +Traditional text-based person retrieval (TPR) relies on a single-shot text as query to retrieve the target person, assuming that the query completely captures the user's search intent. However, in real-world scenarios, it can be challenging to ensure the information completeness of such single-shot text. To address this limitation, we propose chat-based person retrieval (ChatPR), a new paradigm that takes an interactive dialogue as query to perform the person retrieval, engaging the user in conversational context to progressively refine the query for accurate person retrieval. The primary challenge in ChatPR is the lack of available dialogue-image paired data. To overcome this challenge, we establish ChatPedes, the first dataset designed for ChatPR, which is constructed by leveraging large language models to automate the question generation and simulate user responses. Additionally, to bridge the modality gap between dialogues and images, we propose a dialogue-refined cross-modal alignment (DiaNA) framework, which leverages two adaptive attribute refiner modules to bottleneck the conversational and visual information for fine-grained cross-modal alignment. Moreover, we propose a dialogue-specific data augmentation strategy, random round retaining, to further enhance the model's generalization ability across varying dialogue lengths. Extensive experiments demonstrate that DiaNA significantly outperforms existing TPR approaches, highlighting the effectiveness of conversational interactions for person retrieval. + +# 1. Introduction + +Given the paramount importance of ensuring public safety through intelligent video surveillance, there is a substan + +![](images/ff8ea91dc79fec4497d59b56db88a76873a99574fa8e2b24f370176ad75dcadb.jpg) +(a) Illustration of text-based person retrieval (TPR). + +![](images/19454dbdae73835ec660da26a9342485a666e4116a94ab6c26d698924212da10.jpg) +(b) Illustration of chat-based person retrieval (ChatPR). +Figure 1. Comparison of TPR and ChatPR. (a) TPR relies on a single-shot text as the query, which fails to completely capture users' search intent due to the absence of several details. (b) ChatPR, in contrast, engages users in interactive conversations, prompting them to offer additional clues and clarify the negative intents. This interaction allows a progressive refinement of the query, thereby enhancing the retrieval accuracy. The correct image is marked with green rectangles. + +tial need to develop systems capable of accurately retrieving a person of interest from a large image gallery. In this context, text-based person retrieval (TPR) [23, 39, 53] employs free-form natural language as the query to perform the retrieval, effectively addressing the limitation that visual queries, such as person images (referred to as person re-identification [34, 51, 54, 56]), are not always readily available in real-world scenarios. Consequently, TPR has + +gained ever-rising attention in recent years [24, 42, 50, 58]. + +Existing methods in TPR typically rely on a single-shot text query, assuming that the query comprehensively captures the user's search intent. However, in practical applications, it can be challenging to ensure the completeness of information in a single-shot text query for person retrieval. As illustrated in Fig. 1a, the user provides an incomplete single-shot text query with the absence of several details (e.g., "bag"), leading to an unsatisfactory retrieval. Alternatively, engaging in multiple rounds of iterative interaction using a question-answer format allows for the conveyance of more detailed information. Recognizing this limitation of single-shot text query in TPR, we introduce a new retrieval paradigm: chat-based person retrieval (ChatPR), which takes interactive dialogues as the query to facilitate a more nuanced and comprehensive interpretation of users' search intent for accurate person retrieval. As shown in Fig. 1b, ChatPR engages users in conversational interaction, allowing the retrieval system to progressively refine the query by actively soliciting additional information, thereby achieving more accurate retrieval. + +ChatPR presents three advantages over traditional TPR. Firstly, it aligns more closely with real-world applications, in which multiple rounds of iterative interaction through a question-answer format provide a more natural way to query the target person. Secondly, the question-answer format actively prompts users for additional information through conversational interactions, allowing the query to encompass more nuanced details. Lastly, it can explicitly capture the user's negative retrieval intents, which are often overlooked in single-shot text queries but are crucial for filtering out irrelevant targets and narrowing the search scope. These unique features highlight the necessity of ChatPR to leverage interactive dialogues for accurate person retrieval. + +However, when addressing the ChatPR task, we are faced with two main challenges: ① Lack of dialogue-image paired data: While the research community has witnessed an increase in the availability of text-image paired datasets for TPR [12, 30, 61, 62], there remains an absence of dialogue-image paired dataset. This data deficiency poses a significant barrier to the research and development of ChatPR. ② Complexity of cross-modal alignment: In contrast to the single-shot text, dialogue data presents a more complex structure and longer context. This poses a significant challenge for the traditional TPR models in understanding conversational dialogues, since their text encoder, such as BERT [11] or CLIP [44], is predominantly pretrained on single-text corpora, thereby posing an obstacle to the subsequent cross-modal alignment. Additionally, dialogues tend to have sparser information density, necessitating multiple rounds of interaction to identify specific attributes, which further impedes the cross-modal alignment between dialogue and images. + +In response to the first challenge, considering that collecting such data via human crowd-sourcing is not only time-consuming but also lacks standardized and well-defined guidelines [35], we propose leveraging the advanced instruction-following capabilities of large language models (LLMs) [8, 14] to convert existing image-text pairs into structured dialogue-image formats. While LLM-based dialogue generation is highly efficient, the resulting data inevitably introduces noise. To address this, we introduce a diverse set of specialized evaluators, composed of LLMs with meticulously crafted instructions, to assess the relevance and quality of each dialogue round. These individual evaluation results are then integrated to facilitate data cleaning. As a result, we construct ChatPedes, the first dataset specifically tailored for ChatPR. This dataset offers a robust benchmark for advancing person retrieval within real-world conversational contexts. + +To address the second challenge, we propose a dialogue-refined cross-modal alignment (DiaNA) framework, which consists of dual encoders to comprehensively understand dialogues and images, respectively. Given the natural affinity of LLMs in handling dialogue, we adopt the open-source Llama 3 [14] as the dialogue encoder to effectively comprehend the multi-round interactions. Additionally, DiaNA incorporates two adaptive attribute refiners on the dialogue and image branches, which leverage attribute queries to respectively bottleneck the conversational and visual information. This design establishes a fine-grained cross-modal alignment, thereby enhancing the effectiveness of DiaNA. Furthermore, we propose a dialogue-specific data augmentation strategy, random round retaining (R3), to simulate the variability in dialogue lengths in real-world interactions, where complete information is not always available. This strategy further enhances DiaNA's generalization ability across varying lengths of dialogue context. + +Overall, the primary contributions of this work can be summarized as follows: + +- Recognizing the limitations of single-shot text query in text-based person retrieval, we introduce a new retrieval paradigm: chat-based person retrieval (ChatPR), which engages users in interactive conversations to progressively refine their queries for accurate person retrieval. +- To fill in the data lack, we establish the first dialogue-image paired dataset, ChatPedes, offering the community a robust benchmark for ChatPR. +- To address the challenges in cross-modal alignment, we propose a dialogue-refined cross-modal alignment (Di-ana) framework to enable fine-grained dialogue-image alignment, and introduce a random round retaining strategy to further boost the model's generalization ability. +- Extensive experiments demonstrate the effectiveness of DiaNA in handling multi-round dialogues to improve person retrieval performance over conversational contexts. + +# 2. Related Work + +# 2.1. Text-based Person Retrieval + +Text-based person retrieval [39] aims to find the specific person image based on a textual description. Since Li et al. [30] introduced this task, the community has witnessed its growing flourish in recent years. The studies initially depend on simple global alignment [17, 60], and then gradually evolve to multi-granularity correspondences [7, 19, 45]. More recently, a growing number of works [1, 2, 5, 18, 24, 42] have pivoted towards integrating visual-language pretraining models [28, 29, 44] to leverage general cross-modal knowledge for this fine-grained retrieval. Among them, IRRA [24] incorporates a multi-modal interaction encoder on CLIP [44] to facilitate implicit cross-modal relation reasoning. Furthermore, TBPS-CLIP [5] provides a comprehensive empirical study on the application of CLIP in this field from the perspectives of data augmentation, loss function and training tricks. + +Despite these advancements, existing methods implicitly assume that the single text query fully encapsulates the user's search intent. However, in real-world scenarios, user-provided queries often lack sufficient details and complete information, necessitating iterative refinement for satisfactory retrieval. To address this limitation, this paper introduces a novel chat-based person retrieval paradigm, which engages users in interactive dialogues to progressively refine their queries for more accurate and effective retrieval. + +# 2.2. Interactive Cross-Modal Retrieval + +There has been a significant progress in interactive cross-modal retrieval systems, drawing inspiration from visual dialogue [10, 25]. These systems engage users in a dynamic feedback loop towards the desired retrieval target. Typically, user feedback can be categorized into two types: relevance feedback and difference feedback. Relevance feedback [48] involves users assigning scores to the retrieved results based on their relevance to the query, allowing the system to re-rank its retrieval list accordingly. In contrast, difference feedback [4, 33] requires users to specify the distinctions between the target image and the reference image through pre-defined attribute tags [22, 49] or open-form descriptions [21]. Moreover, recent methods [26, 27] bridge the feedback with the query through conversational questioning, enabling the users to be pro-actively asked according to the dialogue history in each search attempt. + +Different from the aforementioned works, this paper focuses on the specific field of ChatPR, a more fine-grained retrieval task, where users need to provide more comprehensive and discriminative queries to retrieve the target pedestrian images from a vast image gallery. + +# Dialogue Construction Instruction + +# Task Definition + +Act as an Assistant and a User to convert the giving captions of the pedestrian into a question-answer format. You need to integrate the information from the both captions. The Assistant's goal is to reveal as many details as possible ... details not mentioned in the captions. + +# Workflow + +1. Understand the two sentences in the captions. +2. Firstly, the Assistant asks the User to describe the person's features. +3. The User describes the person's features based on the captions. + +6. Don't provide any question or answer that has already been mentioned in previous dialogue history. + +# Examples + +- Captions: ["This man has dark hair and wears a black hoodie with white pattern outside and dark pants ..., "He's wearing a black hooded sweatshirt with a white lining ..." +Output: + +Assistant: Please describe the person. +User: The person is wearing a black hoodie, green pants, a pair of red and white shoes. +Assistant: What's the person's gender? +User: He is male. +. + +Figure 2. The instruction for dialogue construction, which consists of three parts: task definition, workflow and examples. This design aims to activate LLMs' instruction-following, chain-of-thought and in-context capabilities, respectively. + +# 2.3. Large Language Models + +In recent years, the field of artificial intelligence has witnessed a significant leap forward with the advent of large language models (LLMs) [6, 32, 55, 59]. Among the most notable contributions to this domain are the generative pretrained transformer (GPT) series [3, 40, 41, 43]. These models are trained on massive corpora, allowing them to capture a wide range of linguistic patterns and world knowledge. Moreover, the large-scale parameters endow LLMs with emergent abilities [47] that are not present in smaller models [16], such as instruction-following [38] and in-context [13] learning. These capabilities allow LLMs to accurately interpret and execute complex human instructions, as well as understand and utilize context from a few examples to perform tasks effectively. + +In this work, we utilize LLMs to automate question generation and simulate user responses for the construction of ChatPedes. By providing carefully crafted instructions and a few examples, we fully exploit their instruction-following and in-context learning capabilities to create the first benchmark for ChatPR. + +# 3. Benchmark + +Recognizing the limitations of single-shot text queries with incomplete descriptions in TPR, we propose the ChatPR task, which enables the progressive query refine + +![](images/c1b19b32370dfabdefb46f6d1794262cf737b8d94285fe483686481cb0859860.jpg) +Figure 3. Category distribution of dialogue data from Llama 3 [14] and InternVL [8], evaluated by Qwen 2.5 [46]. Over one-quarter of the data is not matched, highlighting the necessity of data cleaning. + +ment through conversational interactions for accurate person retrieval. However, the primary challenge lies in the lack of dialogue-image paired data. To address this, we construct the ChatPedes dataset through two key phases: dialogue construction and data cleaning. + +# 3.1. Dialogue Construction + +In recent years, the research community has witnessed an increase in the availability of text-image paired datasets for TPR [12, 30, 61, 62]. Typically, to create high-quality datasets, a large workforce is employed to manually annotate the images with textual descriptions. However, when it comes to annotating with conversational dialogues via human crowd-sourcing for ChatPR, two major challenges arise: ① Increased difficulty of dialogue annotation: Compared to single-shot text, dialogue data presents a more complex structure and longer context, significantly increasing the manual annotation workload. ② Less well-defined guidelines: Unlike textual annotation tasks with clear instructions, formulating appropriate conversations for dialogue annotation lacks well-defined guidelines [35], complicating the annotation process. Consequently, these challenges make it difficult to employ human crowd-sourcing for creating dialogue-image paired datasets. + +Inspired by the success of LLMs in text annotation tasks [20, 50], we leverage their advanced instruction-following capabilities to automate dialogue annotations, based on the well-known text-image paired CUHK-PEDES dataset [30] in TPR. As illustrated in Fig. 4, dialogue construction aims to construct on-going conversations through the interaction between the LLM-based AI agent and the user. During the interaction, the AI agent formulates follow-up questions based on previous dialogue history. The user, also simulated by the LLM, responds with corresponding answers based on the characteristics of the target person, drawing from the existing annotated captions. + +In general, it is expected that the dialogue data produced by the LLM-based AI agent contains comprehensive and valuable information, which is advantageous for subsequent model training. Consequently, we empirically design the instructions shown in Fig. 2. Structurally, the instruction consists of three parts: task definition, workflow, and a + +
DatasetsIDsImagesTextsDlg.s.RoundsWords
maxminavgmaxminavg
CUHK-PEDES [30]13,00340,20680,440----961523.5
ICFG-PEDES [12]4,10254,52254,522----83937.2
RSTPReid [61]4,10120,50541,010----701126.5
UFine6926 [62]6,92626,20652,412----2183080.8
ChatPedes (Ours)12,00337,128-74,2561717.738812114.4
+ +Table 1. Comparison between our established ChatPedes and the widely-used TPR benchmarks. "Dlgs." refers to dialogues. ChatPedes uses the multi-round dialogues as queries, demonstrating a longer context with conversational interactions. + +few manually designed examples, aiming to activate LLMs' instruction-following, chain-of-thought and in-context capabilities to construct high-quality dialogue data. In the dialogue content, the first question provided by the AI agent typically inquires about a holistic description of the target person. Furthermore, to enhance data diversity, we leverage two publicly available LLMs, Llama 3 [14] and InternVL [8], to generate two distinct dialogues for each person image. The experiments detailed in Sec. 5.3.2 substantiate the benefits of this data diversification strategy. + +# 3.2. Data Cleaning + +Although LLMs facilitate the dialogue data collection, the generated data inevitably contains noise, which would significantly impair the optimization of the retrieval process [57]. To perform the data cleaning, we first define the following four classes for each dialogue round: + +- Matched: The answer either aligns with the information present in the annotated captions or accurately reflects the absence of such details not mentioned in the captions. +- Contradictory: The answer contradicts the information present in the captions. +- Hallucinatory: The answer contains fabricated information that is not present in the captions. +- Repeated: The question or answer contains information that has been mentioned in earlier dialogue rounds. + +Subsequently, we employ LLMs, including Qwen 2.5 [46], Llama 3 [14] and InternVL [8], as specialized evaluators through the crafted instructions (detailed in Appendix B.1) to classify each dialogue round into the four predefined categories. As illustrated in Fig. 3, the category distribution estimated by Qwen 2.5 reveals that over one-quarter of the dialogue data is classified as unmatched, highlighting the necessity for data cleaning. To address this, a vote-based ensemble strategy is adopted to perform the data cleaning. Specifically, we calculate the matching score $s$ for each dialogue round, which represents the proportion of the answers labeled as "Matched" by the LLMs: + +$$ +s = \frac {1}{N _ {q}} \sum_ {i = 1} ^ {N _ {q}} \mathbb {1} \left(a _ {i} = \text {M a t c h e d}\right), \tag {1} +$$ + +where $\mathbb{1}(\cdot)$ is identity function, $N_{q}$ denotes the number of + +![](images/7bad6ec9e6d5ee641dbb85082db89546f8c76b47b8dca0e7aed473262a4c260d.jpg) +Figure 4. The overview of ChatPR. Dialogue Construction aims to construct an on-going conversation, in which an AI agent formulates follow-up questions based on previous dialogue history, prompting the user to provide additional information about the target person to be searched. This iterative process forms a comprehensive Dialogue Query for the target person. DiaNA aims to bridge the modality gap between the dialogue and the image, which incorporates two adaptive attribute refiner (AAR) modules by leveraging learnable attribute queries to extract key information for fine-grained cross-modal alignment. During training, random round retaining (R3) strategy simulates the incomplete query in reality to enhance the model's generalization ability. + +the adopted evaluators, and $a_{i}$ represents the $i$ -th evaluator's answer. If the score $s$ is less than half of the number of evaluators $N_{q}$ , this round of conversation is regarded as data noise and then discarded from the dialogue. Detailed statistics of data cleaning are provided in Appendix C. + +Finally, we establish ChatPedes, a dialogue-image paired dataset designed for ChatPR. As shown in Tab. 1, ChatPedes comprises 37,128 images from 12,003 identities, with each image accompanied by two generated dialogues. Consistent with the data split in CUHK-PEDES [30], we divide 34,054 image from 11,003 identities into the training set, and 3,074 image from 1,000 identities into the test set. In addition, ChatPedes is characterized by multi-round dialogues and longer context, exhibiting the longest average length of 114.4 words per dialogue among all benchmarks. Visualization examples are presented in Appendix E. + +# 4. Method + +# 4.1. Dialogue-Refined Cross-Modal Alignment + +ChatPR engages users in interactive dialogues to progressively refine their queries, where they are pro-actively questioned and prompted to provide more detailed information for more accurate person retrieval. To bridge the modality gap between dialogues and images, we propose a dialogue-refined cross-modal alignment (DiaNA) framework, as illustrated in Fig. 4, which aims to learn a dialogue-image shared latent space, pulling the pairs with consistent semantics closer together, while pushing those with different semantics further apart. DiaNA is composed of dual + +encoders to comprehend dialogue queries and person images, respectively, along with two adaptive attribute refiners to bottleneck the conversational and visual information for fine-grained cross-modal alignment. + +# 4.1.1. Dual Encoders + +In contrast to the single-shot text query in TPR, dialogue data in ChatPR presents a more complex structure and longer context, rendering traditional text processing models, such as BERT [11] and CLIP [44] (text encoder), less suitable for this task. Given the proficiency of LLMs in handling dialogue data, we adopt the publicly accessible Llama 3 [14] to comprehend the dialogues within our framework. + +Formally, we denote a dialogue as $\mathbf{D} = \{(Q_i,A_i)\}_{i = 1}^N$ , where $Q_{i}$ and $A_{i}$ represent the $i$ -th round of question and answer in the dialogue with a total of $N$ rounds. We construct the instruction-format input, as depicted in Fig. 5. Following Vicuna [9], we prepend a system message $X_{\mathrm{system - message}}$ (detailed in Appendix B.2) to activate the instruction-following capability of the LLM, enabling a more comprehensive understanding of the dialogue data. Finally, the entire input is fed into Llama 3, yielding an embedding sequence $\mathcal{W} = \{w_{bos},w_1,\dots ,w_{eos}\}$ , where $w_{eos}$ captures the full semantics of the dialogue due to the decoder-only architecture of Llama 3. Subsequently, the global embedding $w_{eos}$ is mapped into the shared space: + +$$ +w = \Phi_ {D} \left(w _ {e o s}\right) \in \mathbb {R} ^ {d}, \tag {2} +$$ + +where $\Phi_D$ is a projector composed of a linear layer, $d$ is the dimension of the space. + +![](images/b707f2ce3ebe60a9367343ea56fbe0d52479ab6451938eaa147ef40cfc26e4e9.jpg) +Figure 5. The instruction-formatted input sequence. A system message is prepended to prompt Llama 3 in capturing the semantics of the dialogue context. + +For the visual branch, given an image $\mathbf{I}$ , we employ Swin Transformer [36, 37] to encode the image into a visual sequence $\mathcal{V} = \{v_{1}, v_{2}, \dots, v_{h \times w}\}$ , where $h$ and $w$ represent the height and width of the feature map at the last stage, respectively. Subsequently, a global average pooling layer is adopted to integrate the overall semantics, yielding a global visual embedding $v_{g} = \frac{1}{h \times w} \sum_{i=1}^{h \times w} v_{i}$ . Similarly, we introduce a projector to map $v_{g}$ into the shared space: + +$$ +v = \Phi_ {I} (v _ {g}) \in \mathbb {R} ^ {d}. \tag {3} +$$ + +# 4.1.2. Adaptive Attribute Refiner + +Unlike the single-shot text where content is generally concentrated, dialogue data typically exhibits sparser information density, requiring multiple rounds of interaction to ascertain specific attributes. This sparsity poses a significant challenge in cross-modal alignment, necessitating information refinement to effectively bridge the modality gap between dialogues and images. To address this, we propose two adaptive attribute refiner (AAR) modules on the dialogue and the image encoders to learn the attribute representations for fine-grained cross-modal alignment. + +Specifically, we first introduce a dialogue AAR upon the dialogue encoder. Considering that different persons have varying attributes, we introduce a set of learnable attribute queries $\mathcal{Q} = \{q_1, q_2, \dots, q_K\}$ to focus on the distinct attributes of each individual, where $K$ denotes the number of attribute queries. By interacting with the dialogue embeddings $\mathcal{W}$ , these attribute queries adaptively extract the person-specific attribute representations from the sparse dialogue content. As illustrated in Fig. 4, the dialogue AAR applies a computationally efficient structure with a single multi-head cross attention (MCA) followed by 2-layer transformer blocks. The interaction between the attribute queries and the dialogue embeddings is performed through the cross-attention mechanism: + +$$ +\hat {\mathcal {W}} = \operatorname {T r a n s f o r m e r} (\operatorname {M C A} (\mathcal {Q}, \mathcal {W}, \mathcal {W})), \tag {4} +$$ + +where $\hat{\mathcal{W}} = \{\hat{w}_1,\hat{w}_2,\dots ,\hat{w}_K\}$ indicates the adaptively refined dialogue attribute embeddings. + +In parallel, we symmetrically introduce a visual AAR to extract the visual attributes from the image, mitigating the impacts of the inherent redundancy and background noise in visual information: + +$$ +\hat {\mathcal {V}} = \operatorname {T r a n s f o r m e r} (\operatorname {M C A} (\mathcal {Q}, \mathcal {V}, \mathcal {V})), \tag {5} +$$ + +where $\hat{\mathcal{V}} = \{\hat{v}_1,\hat{v}_2,\dots ,\hat{v}_K\}$ denotes the visual attribute embeddings. Furthermore, we leverage a parameter-sharing mechanism between the two AAR modules, which not only reduces the parameter amount but also mitigates their modality gap [15] between dialogues and images. + +# 4.2. Optimization Objectives + +Incorporating the dual encoders and the AAR modules, we construct DiaNA, which is then optimized through both global (coarse-grained) and local (fine-grained) alignment. Specifically, we use the cosine similarity to measure the global alignment between the dialogue and the image: + +$$ +s _ {g} (\mathbf {D}, \mathbf {I}) = \frac {w ^ {\mathrm {T}} v}{\| w \| \| v \|}. \tag {6} +$$ + +Furthermore, the average cosine distance among the refined attribute embeddings is utilized to measure the local semantics consistency between the dialogue and the image: + +$$ +s _ {l} (\mathbf {D}, \mathbf {I}) = \frac {1}{K} \sum_ {i = 1} ^ {K} \frac {\hat {w} _ {i} ^ {\mathrm {T}} \hat {v} _ {i}}{\| \hat {w} _ {i} \| \| \hat {v} _ {i} \|.} \tag {7} +$$ + +According to the global and local similarities between the dialogue and the image, we construct the overall optimization object by employing the normalized image-text contrastive (N-ITC) loss [5] and the similarity distribution matching (SDM) loss [24]: + +$$ +\mathcal {L} = \underbrace {\mathcal {L} _ {g} ^ {n i t c} + \mathcal {L} _ {g} ^ {s d m}} _ {\text {g l o b a l a l i n g m e n t}} + \underbrace {\mathcal {L} _ {l} ^ {n i t c} + \mathcal {L} _ {l} ^ {s d m}} _ {\text {l o c a l a l i n g m e n t}}. \tag {8} +$$ + +# 4.3. Dialogue-specific Data Augmentation + +In real-world scenarios, complete dialogues that capture users' full search intent are not always available, since obtaining these complete dialogues typically requires numerous rounds of interaction, which can be not user-friendly. To enhance the model's generalization ability over dialogues of varying rounds, we propose a dialogue-specific data augmentation strategy, random round retaining (R3), which simulates the inherent variability of the dialogue in practice. + +Specifically, at each training iteration, for a complete dialogue $\mathbf{D} = \{(Q_i,A_i)\}_{i = 1}^N$ , only the initial random $i$ rounds are retained and fed into the model: + +$$ +\hat {\mathbf {D}} = \left\{\left(Q _ {i}, A _ {i}\right) \right\} _ {i = 1} ^ {n}, \quad n \sim \operatorname {r a n d i n t} (1, N), \tag {9} +$$ + +where $\operatorname{randint}(1, N)$ denotes a discrete uniform distribution with equal probability for each integer in range $[1, N]$ . The retained $\hat{\mathbf{D}}$ reflects the partial availability of information in practical scenarios, effectively enhancing the model's generalization ability over varying dialogue lengths. + +
MethodRefPret. Dataw/o Data Cleaningw/ Data Cleaning
R-1R-5R-10mAPR-1R-5R-10mAP
IRRA [24]CVPR'23WIT46.0865.6073.0242.9048.2667.9675.6544.64
TBPS-CLIP [5]AAAI'24WIT47.7466.4173.7843.8050.2169.2776.4345.44
RDE [42]CVPR'24WIT50.2068.2575.1345.9452.7770.6177.4747.98
APTM [52]MM'23MALS39.6256.9964.2336.5741.1557.9165.1638.00
AUL [31]AAAI'24MALS40.0557.0164.4137.0242.0159.2166.0138.56
DiaNA (Ours)-MALS73.7289.6694.2465.4075.6790.6894.7166.89
+ +Table 2. Comparison with state-of-the-art methods on ChatPedes. "Pret. Data" denotes pretraining data. All reported results of other methods are reproduced using their publicly available code. + +# 5. Experiments + +In this section, we conduct extensive experiments to verify the effectiveness of our method. The implementation details are provided in Appendix A. + +# 5.1. Evaluation Protocol + +We adopt the widely-used Rank-K (R-K for short, $\mathrm{K} = 1,5,10$ ) metric in our experiments, which denotes the percentage of successful retrieval that the ground-truth image is found within the top K ranked results. In addition, we also employ the mean average precision (mAP) as an auxiliary metric to evaluate the overall retrieval performance. + +# 5.2. Comparison with SOTA Methods + +We compare our method against five state-of-the-art (SOTA) baselines: IRRA [24], TBPS-CLIP [5], RDE [42], APTM [52] and AUL [31]. Among these, IRRA, TBPS-CLIP and RDE are built on the vision-language foundation model CLIP [44], which is pretrained on a large-scale dataset WIT with 400 million image-text paired data collected from the Internet. APTM and AUL are pretrained on the synthetic text-image paired dataset MALS [52]. It is worth noting that these baselines are originally developed for the TPR task. We reproduce their performance on ChatPedes dataset using their open-source codes, treating the multi-turn dialogue data as a single long-form text input. As shown in Tab. 7, regardless of whether data cleaning is applied, our method outperforms these SOTA baselines by a large margin, while these baselines demonstrate a suboptimal retrieval accuracy, indicating their limitations in effectively handling the complex structure and longer contextual dependencies of dialogue data. + +In addition, we further conduct extensive comparative experiments over the dialogue round, simulating real-world scenarios where users are pro-actively questioned to progressively refine their queries through conversational interactions. This process is exemplified in Appendix E.1. As shown in Fig. 6, it can be observed that: ① As the dialogue progresses, the retrieval performance improves across all methods. However, the performance of the baselines becomes saturated after the initial few rounds of interaction, + +![](images/faf14a2c0db406e51a0611d2a6217a13c035ac18c6ca0e6791b2c644c3696b3b.jpg) +Figure 6. Performance comparison over dialogue round. + +which we attribute to their limitations in processing longer and more complex dialogues. In contrast, our method presents a significantly pronounced improvement as the dialogue evolves on account of our carefully crafted design for handling dialogue data. This comprehensive comparison demonstrates the efficacy of our method and showcases its superior capability in handling the intricacies of dialogue-image alignment. Additionally, we also notice that our method achieves competitive performance at the first round of dialogue. This round typically involves a holistic description of the target person, functioning similarly to the single-shot text in the traditional TPR task. The experimental comparison also indicates the promising generalization capabilities of our method in TPR. Further detailed experimental results on TPR benchmarks are provided in Appendix D. + +# 5.3. Ablation Study + +# 5.3.1. Effectiveness of Data Cleaning + +During dialogue construction, we leverage LLMs to generate dialogue data. However, a major limitation of this automatic generation through LLMs lies in the inevitable introduction of noise, which can be classified into three categories: Contradictory, Hallucinatory and Repeated, as delineated in Sec. 3.2. To evaluate the impact of the noise, we conduct comparative experiments with and without data cleaning. Tab. 7 presents the performance comparison with the five SOTA baselines, where all approaches exhibit a significant improvement. In particular, RDE achieves a notable increase of $2.57\%$ at R-1, demonstrating the effectiveness of data cleaning and the importance of high-quality data. + +
Data DiversityR-1R-5R-10mAP
Llama 3InternVL
X✓71.0188.4093.1863.08
✓X71.6288.5893.5963.73
✓✓75.6790.6894.7166.89
+ +![](images/0abb2854e2375e367e1571ad2a47b4ebdf4ff8b69757ac381ff10e92e72f7f80.jpg) +Figure 7. Performance comparison across a varying number of attribute queries $K$ in AAR modules. Notably, $K = 0$ indicates that the AAR is not used. + +# 5.3.2. Effectiveness of Data Diversity + +We employ Llama 3 [14] and InternVL [8] as different annotators for labeling diverse dialogue data when constructing ChatPedes. Tab. 3 presents the contribution of data diversity to the retrieval performance. It can be observed that: ① Relying on a single LLM yields a suboptimal outcome, with $71.01\%$ for InternVL and $71.62\%$ for Llama 3 at R-1. Additionally, Llama 3 slightly outperforms InternVL due to its more advanced capability in dialogue generation. ② The combination of Llama 3 and InternVL leads to a substantial increase across all metrics, demonstrating the importance of data diversity. By combining annotations generated by different LLMs, the established ChatPedes dataset effectively captures a wider variety of dialogue contexts and nuances, which in turn bolsters the model's retrieval capabilities. + +# 5.3.3. Effectiveness of Adaptive Attribute Refiner + +Considering the sparse information density in dialogues and the inherent redundancy in images, we propose two adaptive attribute refiner (AAR) modules, which employ a set of learnable attribute-specific queries to adaptively extract the relevant attributes from the dialogue and the image. Fig. 7 illustrates the performance variation over the number of attribute queries $K$ , where, notably, $K = 0$ denotes AAR is discarded. We observe that: 1 Only relying on global alignment without AAR results in a suboptimal retrieval performance, underscoring the importance of key information extraction. 2 When introducing AAR, the performance exhibits slight fluctuations, reaching the peak of $75.67\%$ on R-1 and $66.89\%$ on mAP at $K = 16$ . Overall, the performance remains relatively stable, indicating the model's robustness when equipped with AAR. + +Table 3. Ablation study on data diversity. We employ Llama 3 [14] and InternVL [8] to construct the ChatPedes benchmark. + +
MethodMetricDialogue Round
12468all
DiaNA(w/o R3)26.5034.8953.0766.4073.1675.18
DiaNA(w/ R3)R-129.2537.6455.6067.9973.6775.67
Δ+2.75+2.75+2.53+1.59+0.51+0.49
DiaNA(w/o R3)26.3333.6749.0659.6764.6466.19
DiaNA(w/ R3)mAP28.7836.3751.0560.8865.3266.89
Δ+2.45+2.70+1.99+1.21+0.68+0.70
+ +Table 4. Comparison of random round retaining (R3) over dialogue round. "all" refers to the complete dialogue. + +# 5.3.4. Effectiveness of Random Round Retaining + +The proposed random round retaining (R3) strategy selectively retains a random number of initial rounds of the dialogue during training, simulating the users' incomplete search intent in real-world interactions. As shown in Tab. 4, we observe that: $\mathbf{1}$ R3 yields performance gains across all dialogue rounds, with gains from $2.75\%$ in single-round interactions to $0.49\%$ in complete conversations at R-1, demonstrating its robustness and versatility. $\mathbf{2}$ In particular, R3 brings more substantial improvement at the initial few rounds of dialogues, suggesting that R3 is particularly effective when dialogue information is limited. This is especially crucial in practical scenarios where users expect to quickly retrieve the target person with minimal interaction. Overall, R3 empowers users to obtain accurate retrieval results even in the early stages of interaction, thereby addressing the real-world demands for both efficiency and precision. + +# 6. Conclusion + +Recognizing the inherent limitations of traditional TPR where single-shot text queries fail to fully encapsulate users' search intent, we introduce ChatPR, a pioneering paradigm for person retrieval that engages users in interactive dialogues to progressively refine their search queries. For this, we build the first benchmark ChatPedes through LLMs to automate question generation and simulate users' responses. To tackle this task, we propose the DiaNA framework that leverages two adaptive attribute refiners to bottleneck the conversational and visual information for fine-grained cross-modal alignment. Additionally, a novel data augmentation strategy of random round retaining is proposed to mimic the incomplete queries in real-world scenarios, remarkably enhancing the model's generalization ability across varying dialogue lengths. Extensive experiments significantly demonstrate the importance of conversational interaction in person retrieval and the superior effectiveness of our method in ChatPR. + +Acknowledgement. This work is supported by the National Science Foundation of China under Grants (62176188, 62361166629, 62476188). The numerical calculations are supported by the supercomputing system in the Supercomputing Center of Wuhan University. + +# References + +[1] Yang Bai, Min Cao, Daming Gao, Ziqiang Cao, Chen Chen, Zhenfeng Fan, Liqiang Nie, and Min Zhang. Rasa: Relation and sensitivity aware representation learning for text-based person search. In Proceedings of the Thirty-Second International Joint Conference on Artificial Intelligence, pages 555–563, 2023. 3 +[2] Yang Bai, Jingyao Wang, Min Cao, Chen Chen, Ziqiang Cao, Liqiang Nie, and Min Zhang. Text-based person search without parallel image-text data. In Proceedings of the 31st ACM International Conference on Multimedia, pages 757-767, 2023. 3 +[3] Tom B Brown. Language models are few-shot learners. arXiv preprint arXiv:2005.14165, 2020. 3 +[4] Guanyu Cai, Jun Zhang, Xinyang Jiang, Yifei Gong, Lianghua He, Fufu Yu, Pai Peng, Xiaowei Guo, Feiyue Huang, and Xing Sun. Ask&confirm: active detail enriching for cross-modal retrieval with partial query. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1835-1844, 2021. 3 +[5] Min Cao, Yang Bai, Ziyin Zeng, Mang Ye, and Min Zhang. An empirical study of clip for text-based person search. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 465-473, 2024. 3, 6, 7 +[6] Yupeng Chang, Xu Wang, Jindong Wang, Yuan Wu, Linyi Yang, Kaijie Zhu, Hao Chen, Xiaoyuan Yi, Cunxiang Wang, Yidong Wang, et al. A survey on evaluation of large language models. ACM Transactions on Intelligent Systems and Technology, 15(3):1-45, 2024. 3 +[7] Yuhao Chen, Guoqing Zhang, Yujiang Lu, Zhenxing Wang, and Yuhui Zheng. Tipcb: A simple but effective part-based convolutional baseline for text-based person search. Neurocomputing, 494:171-181, 2022. 3 +[8] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24185-24198, 2024. 2, 4, 8 +[9] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhang-hao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yong-hao Zhuang, Joseph E. Gonzalez, Ion Stoica, and Eric P. Xing. Vicuna: An open-source chatbot impressing gpt-4 with $90\%$ chatgpt quality. https://lmsys.org/blog/2023-03-30-vicuna, 2023.5 +[10] Abhishek Das, Satwik Kottur, Khushi Gupta, Avi Singh, Deshraj Yadav, Jose M. F. Moura, Devi Parikh, and Dhruv Batra. Visual dialog. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 326-335, 2017. 3 +[11] Jacob Devlin. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 2, 5 +[12] Zefeng Ding, Changxing Ding, Zhiyin Shao, and Dacheng Tao. Semantically self-aligned network for text-to-image part-aware person re-identification. arXiv preprint arXiv:2107.12666, 2021. 2, 4 + +[13] Qingxiu Dong, Lei Li, Damai Dai, Ce Zheng, Zhiyong Wu, Baobao Chang, Xu Sun, Jingjing Xu, and Zhifang Sui. A survey on in-context learning. arXiv preprint arXiv:2301.00234, 2022. 3 +[14] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 2, 4, 5, 8 +[15] Sedigheh Eslami and Gerard de Melo. Mitigate the gap: Investigating approaches for improving cross-modal alignment in clip. arXiv preprint arXiv:2406.17639, 2024. 6 +[16] Yiyang Fang, Wenke Huang, Guancheng Wan, Kehua Su, and Mang Ye. Emoe: Modality-specific enhanced dynamic emotion experts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2025. 3 +[17] Ammarah Farooq, Muhammad Awais, Fei Yan, Josef Kittler, Ali Akbari, and Syed Safwan Khalid. A convolutional baseline for person re-identification using vision and language descriptions. arXiv preprint arXiv:2003.00808, 2020. 3 +[18] Takuro Fujii and Shuhei Tarashima. Bilma: Bidirectional local-matching for text-based person re-identification. In Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops, pages 2786-2790, 2023. 3 +[19] Chenyang Gao, Guanyu Cai, Xinyang Jiang, Feng Zheng, Jun Zhang, Yifei Gong, Fangzhou Lin, Xing Sun, and Xi-ang Bai. Conditional feature learning based transformer for text-based person search. IEEE Transactions on Image Processing, 31:6097–6108, 2022. 3 +[20] Fabrizio Gilardi, Meysam Alizadeh, and Maël Kubli. Chatgpt outperforms crowd workers for text-annotation tasks. Proceedings of the National Academy of Sciences, 120(30): e2305016120, 2023. 4 +[21] Xiaoxiao Guo, Hui Wu, Yu Cheng, Steven Rennie, Gerald Tesauro, and Rogerio Feris. Dialog-based interactive image retrieval. In Advances in Neural Information Processing Systems, 2018. 3 +[22] Yuxin Hou, Eleonora Vig, Michael Donoser, and Loris Bazzani. Learning attribute-driven disentangled representations for interactive fashion retrieval. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12147-12157, 2021. 3 +[23] S. Irene, A. John Prakash, and V. Rhymend Uthariaraj. Person search over security video surveillance systems using deep learning methods: A review. Image and Vision Computing, 143:104930, 2024. 1 +[24] Ding Jiang and Mang Ye. Cross-modal implicit relation reasoning and aligning for text-to-image person retrieval. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2787-2797, 2023. 2, 3, 6, 7 +[25] Gi-Cheon Kang, Sungdong Kim, Jin-Hwa Kim, Donghyun Kwak, and Byoung-Tak Zhang. The dialog must go on: Improving visual dialog via generative self-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6746-6756, 2023. 3 +[26] Saehyung Lee, Sangwon Yu, Junsung Park, Jihun Yi, and Sungroh Yoon. Interactive text-to-image retrieval with large + +language models: A plug-and-play approach. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics, pages 791-809, 2024. 3 +[27] Matan Levy, Rami Ben-Ari, Nir Darshan, and Dani Lischinski. Chatting makes perfect: Chat-based image retrieval. In Advances in Neural Information Processing Systems, pages 61437-61449, 2023. 3 +[28] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi. Align before fuse: Vision and language representation learning with momentum distillation. In Advances in Neural Information Processing Systems, pages 9694-9705, 2021. 3 +[29] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. BLIP: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In Proceedings of the 39th International Conference on Machine Learning, pages 12888-12900, 2022. 3 +[30] Shuang Li, Tong Xiao, Hongsheng Li, Bolei Zhou, Dayu Yue, and Xiaogang Wang. Person search with natural language description. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2017. 2, 3, 4, 5 +[31] Shenshen Li, Chen He, Xing Xu, Fumin Shen, Yang Yang, and Heng Tao Shen. Adaptive uncertainty-based learning for text-based person retrieval. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3172-3180, 2024. 7 +[32] Jian Liang, Wenke Huang, Guancheng Wan, Qu Yang, and Mang Ye. Lorasculpt: Sculpting lora for harmonizing general and specialized knowledge in multimodal large language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2025. 3 +[33] Kaiqu Liang and Samuel Albanie. Simple baselines for interactive video retrieval with questions and answers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11091-11101, 2023. 3 +[34] Fangyi Liu, Mang Ye, and Bo Du. Learning a generalizable re-identification model from unlabelled data with domain-agnostic expert. Visual Intelligence, 2(1):28, 2024. 1 +[35] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In Advances in Neural Information Processing Systems, pages 34892-34916, 2023. 2, 4 +[36] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. 6 +[37] Ze Liu, Han Hu, Yutong Lin, Zhuliang Yao, Zhenda Xie, Yixuan Wei, Jia Ning, Yue Cao, Zheng Zhang, Li Dong, Furu Wei, and Baining Guo. Swin transformer v2: Scaling up capacity and resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12009-12019, 2022. 6 +[38] Renze Lou, Kai Zhang, and Wenpeng Yin. A comprehensive survey on instruction following. arXiv preprint arXiv:2303.10475, 2023. 3 +[39] Kai Niu, Yanyi Liu, Yuzhou Long, Yan Huang, Liang Wang, and Yanning Zhang. An overview of text-based person + +search: Recent advances and future directions. IEEE Transactions on Circuits and Systems for Video Technology, 34(9): 7803-7819, 2024. 1, 3 +[40] OpenAI. Gpt-4 technical report, 2023. 3 +[41] OpenAI. Chatgpt. https://openai.com/blog/chatgpt, 2023.3 +[42] Yang Qin, Yingke Chen, Dezhong Peng, Xi Peng, Joey Tianyi Zhou, and Peng Hu. Noisy-correspondence learning for text-to-image person re-identification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 27197-27206, 2024. 2, 3, 7 +[43] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. 3 +[44] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In Proceedings of the 38th International Conference on Machine Learning, pages 8748-8763, 2021. 2, 3, 5, 7 +[45] Zhiyin Shao, Xinyu Zhang, Meng Fang, Zhifeng Lin, Jian Wang, and Changxing Ding. Learning granularity-unified representations for text-to-image person re-identification. In Proceedings of the 30th ACM International Conference on Multimedia, pages 5566–5574, 2022. 3 +[46] Qwen Team. Qwen2.5: A party of foundation models. https://qwenlm.github.io/blog/qwen2.5, 2024.4 +[47] Jason Wei, Yi Tay, Rishi Bommasani, Colin Raffel, Barret Zoph, Sebastian Borgeaud, Dani Yogatama, Maarten Bosma, Denny Zhou, Donald Metzler, et al. Emergent abilities of large language models. arXiv preprint arXiv:2206.07682, 2022.3 +[48] Hong Wu, Hanqing Lu, and Songde Ma. Willhunter: interactive image retrieval with multilevel relevance. In Proceedings of the 17th International Conference on Pattern Recognition, pages 1009-1012, 2004. 3 +[49] Yingjia Xu, Mengxia Wu, Zixin Guo, Min Cao, Mang Ye, and Jorma Laaksonen. Efficient text-to-video retrieval via multi-modal multi-tags derived pre-screening. Visual Intelligence, 3(1):1-13, 2025. 3 +[50] Jingfeng Yang, Hongye Jin, Ruixiang Tang, Xiaotian Han, Qizhang Feng, Haoming Jiang, Shaochen Zhong, Bing Yin, and Xia Hu. Harnessing the power of llms in practice: A survey on chatgpt and beyond. ACM Transactions on Knowledge Discovery from Data, 18(6):1-32, 2024. 2, 4 +[51] Jinxi Yang, He Li, Bo Du, and Mang Ye. Cheb-gr: Rethinking k-nearest neighbor search in re-ranking for person re-identification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2025. 1 +[52] Shuyu Yang, Yinan Zhou, Zhedong Zheng, Yaxiong Wang, Li Zhu, and Yujiao Wu. Towards unified text-based person retrieval: A large-scale multi-attribute and language search benchmark. In Proceedings of the 31st ACM International Conference on Multimedia, pages 4492–4501, 2023. 7 + +[53] Mang Ye, Jianbing Shen, Gaojie Lin, Tao Xiang, Ling Shao, and Steven C. H. Hoi. Deep learning for person re-identification: A survey and outlook. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(6):2872-2893, 2022. 1 +[54] Mang Ye, Shuoyi Chen, Chenyue Li, Wei-Shi Zheng, David Crandall, and Bo Du. Transformer for object re-identification: A survey. International Journal of Computer Vision, pages 1–31, 2024. 1 +[55] Mang Ye, Xuankun Rong, Wenke Huang, Bo Du, Nenghai Yu, and Dacheng Tao. A survey of safety on large vision-language models: Attacks, defenses and evaluations. arXiv preprint arXiv:2502.14881, 2025. 3 +[56] Mang Ye, Zesen Wu, and Bo Du. Dual-level matching with outlier filtering for unsupervised visible-infrared person re-identification. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025. 1 +[57] Chiyuan Zhang, Samy Bengio, Moritz Hardt, Benjamin Recht, and Oriol Vinyals. Understanding deep learning (still) requires rethinking generalization. Communications of the ACM, 64(3):107-115, 2021. 4 +[58] Ying Zhang and Hutchuan Lu. Deep cross-modal projection learning for image-text matching. In Proceedings of the European Conference on Computer Vision, pages 686-701, 2018. 2 +[59] Wayne Xin Zhao, Kun Zhou, Junyi Li, Tianyi Tang, Xiaolei Wang, Yupeng Hou, Yingqian Min, Beichen Zhang, Junjie Zhang, Zican Dong, et al. A survey of large language models. arXiv preprint arXiv:2303.18223, 2023. 3 +[60] Zhedong Zheng, Liang Zheng, Michael Garrett, Yi Yang, Mingliang Xu, and Yi-Dong Shen. Dual-path convolutional image-text embeddings with instance loss. ACM Transactions on Multimedia Computing, Communications, and Applications, 16(2):1-23, 2020. 3 +[61] Aichun Zhu, Zijie Wang, Yifeng Li, Xili Wan, Jing Jin, Tian Wang, Fangqiang Hu, and Gang Hua. Dssl: Deep surroundings-person separation learning for text-based person retrieval. In Proceedings of the 29th ACM International Conference on Multimedia, page 209-217, 2021. 2, 4 +[62] Jialong Zuo, Hanyu Zhou, Ying Nie, Feng Zhang, Tianyu Guo, Nong Sang, Yunhe Wang, and Changxin Gao. Ufinebench: Towards text-based person retrieval with ultra-fine granularity. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22010-22019, 2024. 2, 4 \ No newline at end of file diff --git a/2025/Chat-based Person Retrieval via Dialogue-Refined Cross-Modal Alignment/images.zip b/2025/Chat-based Person Retrieval via Dialogue-Refined Cross-Modal Alignment/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..5b00bc6d8c975a203cd117e0008a2d400a369fd5 --- /dev/null +++ b/2025/Chat-based Person Retrieval via Dialogue-Refined Cross-Modal Alignment/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c157a2a9e1d060f3494e61d34dc0caa4586a1d4902b8a60f9c77a43a7c06b6bb +size 429838 diff --git a/2025/Chat-based Person Retrieval via Dialogue-Refined Cross-Modal Alignment/layout.json b/2025/Chat-based Person Retrieval via Dialogue-Refined Cross-Modal Alignment/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..33975ff1440b4f644d62f79ad26358d7fa4be50b --- /dev/null +++ b/2025/Chat-based Person Retrieval via Dialogue-Refined Cross-Modal Alignment/layout.json @@ -0,0 +1,8878 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 74, + 103, + 536, + 122 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 103, + 536, + 122 + ], + "spans": [ + { + "bbox": [ + 74, + 103, + 536, + 122 + ], + "type": "text", + "content": "Chat-based Person Retrieval via Dialogue-Refined Cross-Modal Alignment" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 142, + 475, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 142, + 475, + 158 + ], + "spans": [ + { + "bbox": [ + 138, + 142, + 475, + 158 + ], + "type": "text", + "content": "Yang Bai" + }, + { + "bbox": [ + 138, + 142, + 475, + 158 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 138, + 142, + 475, + 158 + ], + "type": "text", + "content": ", Yucheng Ji" + }, + { + "bbox": [ + 138, + 142, + 475, + 158 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 138, + 142, + 475, + 158 + ], + "type": "text", + "content": ", Min Cao" + }, + { + "bbox": [ + 138, + 142, + 475, + 158 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 138, + 142, + 475, + 158 + ], + "type": "text", + "content": ", Jinqiao Wang" + }, + { + "bbox": [ + 138, + 142, + 475, + 158 + ], + "type": "inline_equation", + "content": "^{2,4}" + }, + { + "bbox": [ + 138, + 142, + 475, + 158 + ], + "type": "text", + "content": ", Mang Ye" + }, + { + "bbox": [ + 138, + 142, + 475, + 158 + ], + "type": "inline_equation", + "content": "^{1*}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 146, + 158, + 463, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 158, + 463, + 185 + ], + "spans": [ + { + "bbox": [ + 146, + 158, + 463, + 185 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 146, + 158, + 463, + 185 + ], + "type": "text", + "content": "National Engineering Research Center for Multimedia Software, School of Computer Science, Wuhan University, Wuhan, China" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 216, + 185, + 395, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 216, + 185, + 395, + 198 + ], + "spans": [ + { + "bbox": [ + 216, + 185, + 395, + 198 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 216, + 185, + 395, + 198 + ], + "type": "text", + "content": "Wuhan AI Research, Wuhan, China" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 104, + 198, + 506, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 198, + 506, + 228 + ], + "spans": [ + { + "bbox": [ + 104, + 198, + 506, + 228 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 104, + 198, + 506, + 228 + ], + "type": "text", + "content": "School of Computer Science and Technology, Soochow University, Suzhou, China \n" + }, + { + "bbox": [ + 104, + 198, + 506, + 228 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 104, + 198, + 506, + 228 + ], + "type": "text", + "content": "Institute of Automation, Chinese Academy of Science, Beijing, China" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 229, + 229, + 381, + 241 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 229, + 229, + 381, + 241 + ], + "spans": [ + { + "bbox": [ + 229, + 229, + 381, + 241 + ], + "type": "text", + "content": "{ybai.ai, yemang}@whu.edu.cn" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 200, + 243, + 407, + 254 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 200, + 243, + 407, + 254 + ], + "spans": [ + { + "bbox": [ + 200, + 243, + 407, + 254 + ], + "type": "text", + "content": "https://github.com/Flame-Chasers/DiaNA" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 151, + 282, + 200, + 295 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 282, + 200, + 295 + ], + "spans": [ + { + "bbox": [ + 151, + 282, + 200, + 295 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 308, + 296, + 631 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 308, + 296, + 631 + ], + "spans": [ + { + "bbox": [ + 55, + 308, + 296, + 631 + ], + "type": "text", + "content": "Traditional text-based person retrieval (TPR) relies on a single-shot text as query to retrieve the target person, assuming that the query completely captures the user's search intent. However, in real-world scenarios, it can be challenging to ensure the information completeness of such single-shot text. To address this limitation, we propose chat-based person retrieval (ChatPR), a new paradigm that takes an interactive dialogue as query to perform the person retrieval, engaging the user in conversational context to progressively refine the query for accurate person retrieval. The primary challenge in ChatPR is the lack of available dialogue-image paired data. To overcome this challenge, we establish ChatPedes, the first dataset designed for ChatPR, which is constructed by leveraging large language models to automate the question generation and simulate user responses. Additionally, to bridge the modality gap between dialogues and images, we propose a dialogue-refined cross-modal alignment (DiaNA) framework, which leverages two adaptive attribute refiner modules to bottleneck the conversational and visual information for fine-grained cross-modal alignment. Moreover, we propose a dialogue-specific data augmentation strategy, random round retaining, to further enhance the model's generalization ability across varying dialogue lengths. Extensive experiments demonstrate that DiaNA significantly outperforms existing TPR approaches, highlighting the effectiveness of conversational interactions for person retrieval." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 652, + 135, + 664 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 652, + 135, + 664 + ], + "spans": [ + { + "bbox": [ + 56, + 652, + 135, + 664 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 672, + 295, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 672, + 295, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 672, + 295, + 696 + ], + "type": "text", + "content": "Given the paramount importance of ensuring public safety through intelligent video surveillance, there is a substan" + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 317, + 281, + 555, + 337 + ], + "blocks": [ + { + "bbox": [ + 317, + 281, + 555, + 337 + ], + "lines": [ + { + "bbox": [ + 317, + 281, + 555, + 337 + ], + "spans": [ + { + "bbox": [ + 317, + 281, + 555, + 337 + ], + "type": "image", + "image_path": "ff8ea91dc79fec4497d59b56db88a76873a99574fa8e2b24f370176ad75dcadb.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 351, + 338, + 517, + 349 + ], + "lines": [ + { + "bbox": [ + 351, + 338, + 517, + 349 + ], + "spans": [ + { + "bbox": [ + 351, + 338, + 517, + 349 + ], + "type": "text", + "content": "(a) Illustration of text-based person retrieval (TPR)." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 317, + 355, + 555, + 495 + ], + "blocks": [ + { + "bbox": [ + 317, + 355, + 555, + 495 + ], + "lines": [ + { + "bbox": [ + 317, + 355, + 555, + 495 + ], + "spans": [ + { + "bbox": [ + 317, + 355, + 555, + 495 + ], + "type": "image", + "image_path": "19454dbdae73835ec660da26a9342485a666e4116a94ab6c26d698924212da10.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 345, + 497, + 523, + 507 + ], + "lines": [ + { + "bbox": [ + 345, + 497, + 523, + 507 + ], + "spans": [ + { + "bbox": [ + 345, + 497, + 523, + 507 + ], + "type": "text", + "content": "(b) Illustration of chat-based person retrieval (ChatPR)." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 513, + 555, + 601 + ], + "lines": [ + { + "bbox": [ + 313, + 513, + 555, + 601 + ], + "spans": [ + { + "bbox": [ + 313, + 513, + 555, + 601 + ], + "type": "text", + "content": "Figure 1. Comparison of TPR and ChatPR. (a) TPR relies on a single-shot text as the query, which fails to completely capture users' search intent due to the absence of several details. (b) ChatPR, in contrast, engages users in interactive conversations, prompting them to offer additional clues and clarify the negative intents. This interaction allows a progressive refinement of the query, thereby enhancing the retrieval accuracy. The correct image is marked with green rectangles." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "text", + "content": "tial need to develop systems capable of accurately retrieving a person of interest from a large image gallery. In this context, text-based person retrieval (TPR) [23, 39, 53] employs free-form natural language as the query to perform the retrieval, effectively addressing the limitation that visual queries, such as person images (referred to as person re-identification [34, 51, 54, 56]), are not always readily available in real-world scenarios. Consequently, TPR has" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 703, + 179, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 179, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 179, + 713 + ], + "type": "text", + "content": "*Corresponding Author: Mang Ye" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3952" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 85 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 85 + ], + "type": "text", + "content": "gained ever-rising attention in recent years [24, 42, 50, 58]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 87, + 295, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 87, + 295, + 326 + ], + "spans": [ + { + "bbox": [ + 55, + 87, + 295, + 326 + ], + "type": "text", + "content": "Existing methods in TPR typically rely on a single-shot text query, assuming that the query comprehensively captures the user's search intent. However, in practical applications, it can be challenging to ensure the completeness of information in a single-shot text query for person retrieval. As illustrated in Fig. 1a, the user provides an incomplete single-shot text query with the absence of several details (e.g., \"bag\"), leading to an unsatisfactory retrieval. Alternatively, engaging in multiple rounds of iterative interaction using a question-answer format allows for the conveyance of more detailed information. Recognizing this limitation of single-shot text query in TPR, we introduce a new retrieval paradigm: chat-based person retrieval (ChatPR), which takes interactive dialogues as the query to facilitate a more nuanced and comprehensive interpretation of users' search intent for accurate person retrieval. As shown in Fig. 1b, ChatPR engages users in conversational interaction, allowing the retrieval system to progressively refine the query by actively soliciting additional information, thereby achieving more accurate retrieval." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 328, + 295, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 328, + 295, + 484 + ], + "spans": [ + { + "bbox": [ + 55, + 328, + 295, + 484 + ], + "type": "text", + "content": "ChatPR presents three advantages over traditional TPR. Firstly, it aligns more closely with real-world applications, in which multiple rounds of iterative interaction through a question-answer format provide a more natural way to query the target person. Secondly, the question-answer format actively prompts users for additional information through conversational interactions, allowing the query to encompass more nuanced details. Lastly, it can explicitly capture the user's negative retrieval intents, which are often overlooked in single-shot text queries but are crucial for filtering out irrelevant targets and narrowing the search scope. These unique features highlight the necessity of ChatPR to leverage interactive dialogues for accurate person retrieval." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 486, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 486, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 486, + 295, + 714 + ], + "type": "text", + "content": "However, when addressing the ChatPR task, we are faced with two main challenges: ① Lack of dialogue-image paired data: While the research community has witnessed an increase in the availability of text-image paired datasets for TPR [12, 30, 61, 62], there remains an absence of dialogue-image paired dataset. This data deficiency poses a significant barrier to the research and development of ChatPR. ② Complexity of cross-modal alignment: In contrast to the single-shot text, dialogue data presents a more complex structure and longer context. This poses a significant challenge for the traditional TPR models in understanding conversational dialogues, since their text encoder, such as BERT [11] or CLIP [44], is predominantly pretrained on single-text corpora, thereby posing an obstacle to the subsequent cross-modal alignment. Additionally, dialogues tend to have sparser information density, necessitating multiple rounds of interaction to identify specific attributes, which further impedes the cross-modal alignment between dialogue and images." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 553, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 275 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 275 + ], + "type": "text", + "content": "In response to the first challenge, considering that collecting such data via human crowd-sourcing is not only time-consuming but also lacks standardized and well-defined guidelines [35], we propose leveraging the advanced instruction-following capabilities of large language models (LLMs) [8, 14] to convert existing image-text pairs into structured dialogue-image formats. While LLM-based dialogue generation is highly efficient, the resulting data inevitably introduces noise. To address this, we introduce a diverse set of specialized evaluators, composed of LLMs with meticulously crafted instructions, to assess the relevance and quality of each dialogue round. These individual evaluation results are then integrated to facilitate data cleaning. As a result, we construct ChatPedes, the first dataset specifically tailored for ChatPR. This dataset offers a robust benchmark for advancing person retrieval within real-world conversational contexts." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 277, + 553, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 277, + 553, + 493 + ], + "spans": [ + { + "bbox": [ + 313, + 277, + 553, + 493 + ], + "type": "text", + "content": "To address the second challenge, we propose a dialogue-refined cross-modal alignment (DiaNA) framework, which consists of dual encoders to comprehensively understand dialogues and images, respectively. Given the natural affinity of LLMs in handling dialogue, we adopt the open-source Llama 3 [14] as the dialogue encoder to effectively comprehend the multi-round interactions. Additionally, DiaNA incorporates two adaptive attribute refiners on the dialogue and image branches, which leverage attribute queries to respectively bottleneck the conversational and visual information. This design establishes a fine-grained cross-modal alignment, thereby enhancing the effectiveness of DiaNA. Furthermore, we propose a dialogue-specific data augmentation strategy, random round retaining (R3), to simulate the variability in dialogue lengths in real-world interactions, where complete information is not always available. This strategy further enhances DiaNA's generalization ability across varying lengths of dialogue context." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 495, + 553, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 495, + 553, + 517 + ], + "spans": [ + { + "bbox": [ + 313, + 495, + 553, + 517 + ], + "type": "text", + "content": "Overall, the primary contributions of this work can be summarized as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 520, + 553, + 711 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 314, + 520, + 553, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 520, + 553, + 579 + ], + "spans": [ + { + "bbox": [ + 314, + 520, + 553, + 579 + ], + "type": "text", + "content": "- Recognizing the limitations of single-shot text query in text-based person retrieval, we introduce a new retrieval paradigm: chat-based person retrieval (ChatPR), which engages users in interactive conversations to progressively refine their queries for accurate person retrieval." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 580, + 553, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 580, + 553, + 613 + ], + "spans": [ + { + "bbox": [ + 314, + 580, + 553, + 613 + ], + "type": "text", + "content": "- To fill in the data lack, we establish the first dialogue-image paired dataset, ChatPedes, offering the community a robust benchmark for ChatPR." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 616, + 553, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 616, + 553, + 675 + ], + "spans": [ + { + "bbox": [ + 314, + 616, + 553, + 675 + ], + "type": "text", + "content": "- To address the challenges in cross-modal alignment, we propose a dialogue-refined cross-modal alignment (Di-ana) framework to enable fine-grained dialogue-image alignment, and introduce a random round retaining strategy to further boost the model's generalization ability." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 676, + 553, + 711 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 676, + 553, + 711 + ], + "spans": [ + { + "bbox": [ + 314, + 676, + 553, + 711 + ], + "type": "text", + "content": "- Extensive experiments demonstrate the effectiveness of DiaNA in handling multi-round dialogues to improve person retrieval performance over conversational contexts." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3953" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 141, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 141, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 141, + 83 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 94, + 209, + 106 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 94, + 209, + 106 + ], + "spans": [ + { + "bbox": [ + 55, + 94, + 209, + 106 + ], + "type": "text", + "content": "2.1. Text-based Person Retrieval" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 115, + 296, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 115, + 296, + 307 + ], + "spans": [ + { + "bbox": [ + 55, + 115, + 296, + 307 + ], + "type": "text", + "content": "Text-based person retrieval [39] aims to find the specific person image based on a textual description. Since Li et al. [30] introduced this task, the community has witnessed its growing flourish in recent years. The studies initially depend on simple global alignment [17, 60], and then gradually evolve to multi-granularity correspondences [7, 19, 45]. More recently, a growing number of works [1, 2, 5, 18, 24, 42] have pivoted towards integrating visual-language pretraining models [28, 29, 44] to leverage general cross-modal knowledge for this fine-grained retrieval. Among them, IRRA [24] incorporates a multi-modal interaction encoder on CLIP [44] to facilitate implicit cross-modal relation reasoning. Furthermore, TBPS-CLIP [5] provides a comprehensive empirical study on the application of CLIP in this field from the perspectives of data augmentation, loss function and training tricks." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 311, + 296, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 311, + 296, + 418 + ], + "spans": [ + { + "bbox": [ + 55, + 311, + 296, + 418 + ], + "type": "text", + "content": "Despite these advancements, existing methods implicitly assume that the single text query fully encapsulates the user's search intent. However, in real-world scenarios, user-provided queries often lack sufficient details and complete information, necessitating iterative refinement for satisfactory retrieval. To address this limitation, this paper introduces a novel chat-based person retrieval paradigm, which engages users in interactive dialogues to progressively refine their queries for more accurate and effective retrieval." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 437, + 238, + 449 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 437, + 238, + 449 + ], + "spans": [ + { + "bbox": [ + 55, + 437, + 238, + 449 + ], + "type": "text", + "content": "2.2. Interactive Cross-Modal Retrieval" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 458, + 295, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 458, + 295, + 650 + ], + "spans": [ + { + "bbox": [ + 55, + 458, + 295, + 650 + ], + "type": "text", + "content": "There has been a significant progress in interactive cross-modal retrieval systems, drawing inspiration from visual dialogue [10, 25]. These systems engage users in a dynamic feedback loop towards the desired retrieval target. Typically, user feedback can be categorized into two types: relevance feedback and difference feedback. Relevance feedback [48] involves users assigning scores to the retrieved results based on their relevance to the query, allowing the system to re-rank its retrieval list accordingly. In contrast, difference feedback [4, 33] requires users to specify the distinctions between the target image and the reference image through pre-defined attribute tags [22, 49] or open-form descriptions [21]. Moreover, recent methods [26, 27] bridge the feedback with the query through conversational questioning, enabling the users to be pro-actively asked according to the dialogue history in each search attempt." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": "Different from the aforementioned works, this paper focuses on the specific field of ChatPR, a more fine-grained retrieval task, where users need to provide more comprehensive and discriminative queries to retrieve the target pedestrian images from a vast image gallery." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 370, + 76, + 499, + 86 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 76, + 499, + 86 + ], + "spans": [ + { + "bbox": [ + 370, + 76, + 499, + 86 + ], + "type": "text", + "content": "Dialogue Construction Instruction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 345, + 95, + 398, + 104 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 95, + 398, + 104 + ], + "spans": [ + { + "bbox": [ + 345, + 95, + 398, + 104 + ], + "type": "text", + "content": "Task Definition" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 328, + 106, + 541, + 141 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 106, + 541, + 141 + ], + "spans": [ + { + "bbox": [ + 328, + 106, + 541, + 141 + ], + "type": "text", + "content": "Act as an Assistant and a User to convert the giving captions of the pedestrian into a question-answer format. You need to integrate the information from the both captions. The Assistant's goal is to reveal as many details as possible ... details not mentioned in the captions." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 348, + 151, + 383, + 159 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 348, + 151, + 383, + 159 + ], + "spans": [ + { + "bbox": [ + 348, + 151, + 383, + 159 + ], + "type": "text", + "content": "Workflow" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 329, + 164, + 533, + 190 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 329, + 164, + 470, + 173 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 329, + 164, + 470, + 173 + ], + "spans": [ + { + "bbox": [ + 329, + 164, + 470, + 173 + ], + "type": "text", + "content": "1. Understand the two sentences in the captions." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 329, + 173, + 533, + 181 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 329, + 173, + 533, + 181 + ], + "spans": [ + { + "bbox": [ + 329, + 173, + 533, + 181 + ], + "type": "text", + "content": "2. Firstly, the Assistant asks the User to describe the person's features." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 329, + 181, + 520, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 329, + 181, + 520, + 190 + ], + "spans": [ + { + "bbox": [ + 329, + 181, + 520, + 190 + ], + "type": "text", + "content": "3. The User describes the person's features based on the captions." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 329, + 198, + 539, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 329, + 198, + 539, + 216 + ], + "spans": [ + { + "bbox": [ + 329, + 198, + 539, + 216 + ], + "type": "text", + "content": "6. Don't provide any question or answer that has already been mentioned in previous dialogue history." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 351, + 225, + 383, + 235 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 225, + 383, + 235 + ], + "spans": [ + { + "bbox": [ + 351, + 225, + 383, + 235 + ], + "type": "text", + "content": "Examples" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 331, + 239, + 542, + 274 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 331, + 239, + 542, + 265 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 239, + 542, + 265 + ], + "spans": [ + { + "bbox": [ + 331, + 239, + 542, + 265 + ], + "type": "text", + "content": "- Captions: [\"This man has dark hair and wears a black hoodie with white pattern outside and dark pants ..., \"He's wearing a black hooded sweatshirt with a white lining ...\"" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 331, + 265, + 365, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 265, + 365, + 274 + ], + "spans": [ + { + "bbox": [ + 331, + 265, + 365, + 274 + ], + "type": "text", + "content": "Output:" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 339, + 274, + 542, + 322 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 339, + 274, + 449, + 282 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 274, + 449, + 282 + ], + "spans": [ + { + "bbox": [ + 339, + 274, + 449, + 282 + ], + "type": "text", + "content": "Assistant: Please describe the person." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 339, + 282, + 542, + 298 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 282, + 542, + 298 + ], + "spans": [ + { + "bbox": [ + 339, + 282, + 542, + 298 + ], + "type": "text", + "content": "User: The person is wearing a black hoodie, green pants, a pair of red and white shoes." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 339, + 299, + 450, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 299, + 450, + 308 + ], + "spans": [ + { + "bbox": [ + 339, + 299, + 450, + 308 + ], + "type": "text", + "content": "Assistant: What's the person's gender?" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 339, + 308, + 391, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 308, + 391, + 316 + ], + "spans": [ + { + "bbox": [ + 339, + 308, + 391, + 316 + ], + "type": "text", + "content": "User: He is male." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 339, + 316, + 345, + 322 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 316, + 345, + 322 + ], + "spans": [ + { + "bbox": [ + 339, + 316, + 345, + 322 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 336, + 555, + 381 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 336, + 555, + 381 + ], + "spans": [ + { + "bbox": [ + 313, + 336, + 555, + 381 + ], + "type": "text", + "content": "Figure 2. The instruction for dialogue construction, which consists of three parts: task definition, workflow and examples. This design aims to activate LLMs' instruction-following, chain-of-thought and in-context capabilities, respectively." + } + ] + } + ], + "index": 26, + "type": "text" + }, + { + "bbox": [ + 313, + 389, + 451, + 402 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 389, + 451, + 402 + ], + "spans": [ + { + "bbox": [ + 313, + 389, + 451, + 402 + ], + "type": "text", + "content": "2.3. Large Language Models" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 313, + 407, + 555, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 407, + 555, + 574 + ], + "spans": [ + { + "bbox": [ + 313, + 407, + 555, + 574 + ], + "type": "text", + "content": "In recent years, the field of artificial intelligence has witnessed a significant leap forward with the advent of large language models (LLMs) [6, 32, 55, 59]. Among the most notable contributions to this domain are the generative pretrained transformer (GPT) series [3, 40, 41, 43]. These models are trained on massive corpora, allowing them to capture a wide range of linguistic patterns and world knowledge. Moreover, the large-scale parameters endow LLMs with emergent abilities [47] that are not present in smaller models [16], such as instruction-following [38] and in-context [13] learning. These capabilities allow LLMs to accurately interpret and execute complex human instructions, as well as understand and utilize context from a few examples to perform tasks effectively." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 313, + 574, + 555, + 645 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 574, + 555, + 645 + ], + "spans": [ + { + "bbox": [ + 313, + 574, + 555, + 645 + ], + "type": "text", + "content": "In this work, we utilize LLMs to automate question generation and simulate user responses for the construction of ChatPedes. By providing carefully crafted instructions and a few examples, we fully exploit their instruction-following and in-context learning capabilities to create the first benchmark for ChatPR." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 313, + 658, + 389, + 669 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 658, + 389, + 669 + ], + "spans": [ + { + "bbox": [ + 313, + 658, + 389, + 669 + ], + "type": "text", + "content": "3. Benchmark" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "content": "Recognizing the limitations of single-shot text queries with incomplete descriptions in TPR, we propose the ChatPR task, which enables the progressive query refine" + } + ] + } + ], + "index": 31 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3954" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 69, + 70, + 284, + 143 + ], + "blocks": [ + { + "bbox": [ + 69, + 70, + 284, + 143 + ], + "lines": [ + { + "bbox": [ + 69, + 70, + 284, + 143 + ], + "spans": [ + { + "bbox": [ + 69, + 70, + 284, + 143 + ], + "type": "image", + "image_path": "c1b19b32370dfabdefb46f6d1794262cf737b8d94285fe483686481cb0859860.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 147, + 297, + 194 + ], + "lines": [ + { + "bbox": [ + 55, + 147, + 297, + 194 + ], + "spans": [ + { + "bbox": [ + 55, + 147, + 297, + 194 + ], + "type": "text", + "content": "Figure 3. Category distribution of dialogue data from Llama 3 [14] and InternVL [8], evaluated by Qwen 2.5 [46]. Over one-quarter of the data is not matched, highlighting the necessity of data cleaning." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 202, + 296, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 202, + 296, + 264 + ], + "spans": [ + { + "bbox": [ + 55, + 202, + 296, + 264 + ], + "type": "text", + "content": "ment through conversational interactions for accurate person retrieval. However, the primary challenge lies in the lack of dialogue-image paired data. To address this, we construct the ChatPedes dataset through two key phases: dialogue construction and data cleaning." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 274, + 185, + 287 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 274, + 185, + 287 + ], + "spans": [ + { + "bbox": [ + 55, + 274, + 185, + 287 + ], + "type": "text", + "content": "3.1. Dialogue Construction" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 292, + 295, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 292, + 295, + 496 + ], + "spans": [ + { + "bbox": [ + 54, + 292, + 295, + 496 + ], + "type": "text", + "content": "In recent years, the research community has witnessed an increase in the availability of text-image paired datasets for TPR [12, 30, 61, 62]. Typically, to create high-quality datasets, a large workforce is employed to manually annotate the images with textual descriptions. However, when it comes to annotating with conversational dialogues via human crowd-sourcing for ChatPR, two major challenges arise: ① Increased difficulty of dialogue annotation: Compared to single-shot text, dialogue data presents a more complex structure and longer context, significantly increasing the manual annotation workload. ② Less well-defined guidelines: Unlike textual annotation tasks with clear instructions, formulating appropriate conversations for dialogue annotation lacks well-defined guidelines [35], complicating the annotation process. Consequently, these challenges make it difficult to employ human crowd-sourcing for creating dialogue-image paired datasets." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 497, + 295, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 497, + 295, + 640 + ], + "spans": [ + { + "bbox": [ + 55, + 497, + 295, + 640 + ], + "type": "text", + "content": "Inspired by the success of LLMs in text annotation tasks [20, 50], we leverage their advanced instruction-following capabilities to automate dialogue annotations, based on the well-known text-image paired CUHK-PEDES dataset [30] in TPR. As illustrated in Fig. 4, dialogue construction aims to construct on-going conversations through the interaction between the LLM-based AI agent and the user. During the interaction, the AI agent formulates follow-up questions based on previous dialogue history. The user, also simulated by the LLM, responds with corresponding answers based on the characteristics of the target person, drawing from the existing annotated captions." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": "In general, it is expected that the dialogue data produced by the LLM-based AI agent contains comprehensive and valuable information, which is advantageous for subsequent model training. Consequently, we empirically design the instructions shown in Fig. 2. Structurally, the instruction consists of three parts: task definition, workflow, and a" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 316, + 70, + 553, + 149 + ], + "blocks": [ + { + "bbox": [ + 316, + 70, + 553, + 149 + ], + "lines": [ + { + "bbox": [ + 316, + 70, + 553, + 149 + ], + "spans": [ + { + "bbox": [ + 316, + 70, + 553, + 149 + ], + "type": "table", + "html": "
DatasetsIDsImagesTextsDlg.s.RoundsWords
maxminavgmaxminavg
CUHK-PEDES [30]13,00340,20680,440----961523.5
ICFG-PEDES [12]4,10254,52254,522----83937.2
RSTPReid [61]4,10120,50541,010----701126.5
UFine6926 [62]6,92626,20652,412----2183080.8
ChatPedes (Ours)12,00337,128-74,2561717.738812114.4
", + "image_path": "348c5e9be210b13d612537298e44ba155f5044fa38e6b4aef3f3420f18b9a374.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 152, + 555, + 196 + ], + "lines": [ + { + "bbox": [ + 313, + 152, + 555, + 196 + ], + "spans": [ + { + "bbox": [ + 313, + 152, + 555, + 196 + ], + "type": "text", + "content": "Table 1. Comparison between our established ChatPedes and the widely-used TPR benchmarks. \"Dlgs.\" refers to dialogues. ChatPedes uses the multi-round dialogues as queries, demonstrating a longer context with conversational interactions." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 202, + 555, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 202, + 555, + 324 + ], + "spans": [ + { + "bbox": [ + 313, + 202, + 555, + 324 + ], + "type": "text", + "content": "few manually designed examples, aiming to activate LLMs' instruction-following, chain-of-thought and in-context capabilities to construct high-quality dialogue data. In the dialogue content, the first question provided by the AI agent typically inquires about a holistic description of the target person. Furthermore, to enhance data diversity, we leverage two publicly available LLMs, Llama 3 [14] and InternVL [8], to generate two distinct dialogues for each person image. The experiments detailed in Sec. 5.3.2 substantiate the benefits of this data diversification strategy." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 328, + 405, + 342 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 328, + 405, + 342 + ], + "spans": [ + { + "bbox": [ + 314, + 328, + 405, + 342 + ], + "type": "text", + "content": "3.2. Data Cleaning" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 346, + 555, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 346, + 555, + 407 + ], + "spans": [ + { + "bbox": [ + 313, + 346, + 555, + 407 + ], + "type": "text", + "content": "Although LLMs facilitate the dialogue data collection, the generated data inevitably contains noise, which would significantly impair the optimization of the retrieval process [57]. To perform the data cleaning, we first define the following four classes for each dialogue round:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 406, + 554, + 513 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 314, + 406, + 554, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 406, + 554, + 441 + ], + "spans": [ + { + "bbox": [ + 314, + 406, + 554, + 441 + ], + "type": "text", + "content": "- Matched: The answer either aligns with the information present in the annotated captions or accurately reflects the absence of such details not mentioned in the captions." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 441, + 553, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 441, + 553, + 464 + ], + "spans": [ + { + "bbox": [ + 314, + 441, + 553, + 464 + ], + "type": "text", + "content": "- Contradictory: The answer contradicts the information present in the captions." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 314, + 465, + 553, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 465, + 553, + 490 + ], + "spans": [ + { + "bbox": [ + 314, + 465, + 553, + 490 + ], + "type": "text", + "content": "- Hallucinatory: The answer contains fabricated information that is not present in the captions." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 489, + 553, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 489, + 553, + 513 + ], + "spans": [ + { + "bbox": [ + 313, + 489, + 553, + 513 + ], + "type": "text", + "content": "- Repeated: The question or answer contains information that has been mentioned in earlier dialogue rounds." + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 514, + 555, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 514, + 555, + 657 + ], + "spans": [ + { + "bbox": [ + 313, + 514, + 555, + 657 + ], + "type": "text", + "content": "Subsequently, we employ LLMs, including Qwen 2.5 [46], Llama 3 [14] and InternVL [8], as specialized evaluators through the crafted instructions (detailed in Appendix B.1) to classify each dialogue round into the four predefined categories. As illustrated in Fig. 3, the category distribution estimated by Qwen 2.5 reveals that over one-quarter of the dialogue data is classified as unmatched, highlighting the necessity for data cleaning. To address this, a vote-based ensemble strategy is adopted to perform the data cleaning. Specifically, we calculate the matching score " + }, + { + "bbox": [ + 313, + 514, + 555, + 657 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 313, + 514, + 555, + 657 + ], + "type": "text", + "content": " for each dialogue round, which represents the proportion of the answers labeled as \"Matched\" by the LLMs:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 370, + 662, + 555, + 696 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 662, + 555, + 696 + ], + "spans": [ + { + "bbox": [ + 370, + 662, + 555, + 696 + ], + "type": "interline_equation", + "content": "s = \\frac {1}{N _ {q}} \\sum_ {i = 1} ^ {N _ {q}} \\mathbb {1} \\left(a _ {i} = \\text {M a t c h e d}\\right), \\tag {1}", + "image_path": "51c9204308e3a931e1931acdcd97c9dbb6dfc859b875701b307b08bf1a02546a.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 701, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 701, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 701, + 555, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 701, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\mathbb{1}(\\cdot)" + }, + { + "bbox": [ + 313, + 701, + 555, + 714 + ], + "type": "text", + "content": " is identity function, " + }, + { + "bbox": [ + 313, + 701, + 555, + 714 + ], + "type": "inline_equation", + "content": "N_{q}" + }, + { + "bbox": [ + 313, + 701, + 555, + 714 + ], + "type": "text", + "content": " denotes the number of" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3955" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 69, + 555, + 277 + ], + "blocks": [ + { + "bbox": [ + 57, + 69, + 555, + 277 + ], + "lines": [ + { + "bbox": [ + 57, + 69, + 555, + 277 + ], + "spans": [ + { + "bbox": [ + 57, + 69, + 555, + 277 + ], + "type": "image", + "image_path": "7bad6ec9e6d5ee641dbb85082db89546f8c76b47b8dca0e7aed473262a4c260d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 281, + 555, + 346 + ], + "lines": [ + { + "bbox": [ + 55, + 281, + 555, + 346 + ], + "spans": [ + { + "bbox": [ + 55, + 281, + 555, + 346 + ], + "type": "text", + "content": "Figure 4. The overview of ChatPR. Dialogue Construction aims to construct an on-going conversation, in which an AI agent formulates follow-up questions based on previous dialogue history, prompting the user to provide additional information about the target person to be searched. This iterative process forms a comprehensive Dialogue Query for the target person. DiaNA aims to bridge the modality gap between the dialogue and the image, which incorporates two adaptive attribute refiner (AAR) modules by leveraging learnable attribute queries to extract key information for fine-grained cross-modal alignment. During training, random round retaining (R3) strategy simulates the incomplete query in reality to enhance the model's generalization ability." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 356, + 295, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 356, + 295, + 415 + ], + "spans": [ + { + "bbox": [ + 55, + 356, + 295, + 415 + ], + "type": "text", + "content": "the adopted evaluators, and " + }, + { + "bbox": [ + 55, + 356, + 295, + 415 + ], + "type": "inline_equation", + "content": "a_{i}" + }, + { + "bbox": [ + 55, + 356, + 295, + 415 + ], + "type": "text", + "content": " represents the " + }, + { + "bbox": [ + 55, + 356, + 295, + 415 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 356, + 295, + 415 + ], + "type": "text", + "content": "-th evaluator's answer. If the score " + }, + { + "bbox": [ + 55, + 356, + 295, + 415 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 55, + 356, + 295, + 415 + ], + "type": "text", + "content": " is less than half of the number of evaluators " + }, + { + "bbox": [ + 55, + 356, + 295, + 415 + ], + "type": "inline_equation", + "content": "N_{q}" + }, + { + "bbox": [ + 55, + 356, + 295, + 415 + ], + "type": "text", + "content": ", this round of conversation is regarded as data noise and then discarded from the dialogue. Detailed statistics of data cleaning are provided in Appendix C." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 415, + 295, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 415, + 295, + 548 + ], + "spans": [ + { + "bbox": [ + 55, + 415, + 295, + 548 + ], + "type": "text", + "content": "Finally, we establish ChatPedes, a dialogue-image paired dataset designed for ChatPR. As shown in Tab. 1, ChatPedes comprises 37,128 images from 12,003 identities, with each image accompanied by two generated dialogues. Consistent with the data split in CUHK-PEDES [30], we divide 34,054 image from 11,003 identities into the training set, and 3,074 image from 1,000 identities into the test set. In addition, ChatPedes is characterized by multi-round dialogues and longer context, exhibiting the longest average length of 114.4 words per dialogue among all benchmarks. Visualization examples are presented in Appendix E." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 556, + 111, + 568 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 556, + 111, + 568 + ], + "spans": [ + { + "bbox": [ + 55, + 556, + 111, + 568 + ], + "type": "text", + "content": "4. Method" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 576, + 275, + 589 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 576, + 275, + 589 + ], + "spans": [ + { + "bbox": [ + 55, + 576, + 275, + 589 + ], + "type": "text", + "content": "4.1. Dialogue-Refined Cross-Modal Alignment" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "text", + "content": "ChatPR engages users in interactive dialogues to progressively refine their queries, where they are pro-actively questioned and prompted to provide more detailed information for more accurate person retrieval. To bridge the modality gap between dialogues and images, we propose a dialogue-refined cross-modal alignment (DiaNA) framework, as illustrated in Fig. 4, which aims to learn a dialogue-image shared latent space, pulling the pairs with consistent semantics closer together, while pushing those with different semantics further apart. DiaNA is composed of dual" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 356, + 555, + 404 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 356, + 555, + 404 + ], + "spans": [ + { + "bbox": [ + 313, + 356, + 555, + 404 + ], + "type": "text", + "content": "encoders to comprehend dialogue queries and person images, respectively, along with two adaptive attribute refiners to bottleneck the conversational and visual information for fine-grained cross-modal alignment." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 408, + 405, + 418 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 408, + 405, + 418 + ], + "spans": [ + { + "bbox": [ + 313, + 408, + 405, + 418 + ], + "type": "text", + "content": "4.1.1. Dual Encoders" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 423, + 554, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 423, + 554, + 506 + ], + "spans": [ + { + "bbox": [ + 313, + 423, + 554, + 506 + ], + "type": "text", + "content": "In contrast to the single-shot text query in TPR, dialogue data in ChatPR presents a more complex structure and longer context, rendering traditional text processing models, such as BERT [11] and CLIP [44] (text encoder), less suitable for this task. Given the proficiency of LLMs in handling dialogue data, we adopt the publicly accessible Llama 3 [14] to comprehend the dialogues within our framework." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 506, + 555, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 506, + 555, + 663 + ], + "spans": [ + { + "bbox": [ + 313, + 506, + 555, + 663 + ], + "type": "text", + "content": "Formally, we denote a dialogue as " + }, + { + "bbox": [ + 313, + 506, + 555, + 663 + ], + "type": "inline_equation", + "content": "\\mathbf{D} = \\{(Q_i,A_i)\\}_{i = 1}^N" + }, + { + "bbox": [ + 313, + 506, + 555, + 663 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 506, + 555, + 663 + ], + "type": "inline_equation", + "content": "Q_{i}" + }, + { + "bbox": [ + 313, + 506, + 555, + 663 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 506, + 555, + 663 + ], + "type": "inline_equation", + "content": "A_{i}" + }, + { + "bbox": [ + 313, + 506, + 555, + 663 + ], + "type": "text", + "content": " represent the " + }, + { + "bbox": [ + 313, + 506, + 555, + 663 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 506, + 555, + 663 + ], + "type": "text", + "content": "-th round of question and answer in the dialogue with a total of " + }, + { + "bbox": [ + 313, + 506, + 555, + 663 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 506, + 555, + 663 + ], + "type": "text", + "content": " rounds. We construct the instruction-format input, as depicted in Fig. 5. Following Vicuna [9], we prepend a system message " + }, + { + "bbox": [ + 313, + 506, + 555, + 663 + ], + "type": "inline_equation", + "content": "X_{\\mathrm{system - message}}" + }, + { + "bbox": [ + 313, + 506, + 555, + 663 + ], + "type": "text", + "content": " (detailed in Appendix B.2) to activate the instruction-following capability of the LLM, enabling a more comprehensive understanding of the dialogue data. Finally, the entire input is fed into Llama 3, yielding an embedding sequence " + }, + { + "bbox": [ + 313, + 506, + 555, + 663 + ], + "type": "inline_equation", + "content": "\\mathcal{W} = \\{w_{bos},w_1,\\dots ,w_{eos}\\}" + }, + { + "bbox": [ + 313, + 506, + 555, + 663 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 506, + 555, + 663 + ], + "type": "inline_equation", + "content": "w_{eos}" + }, + { + "bbox": [ + 313, + 506, + 555, + 663 + ], + "type": "text", + "content": " captures the full semantics of the dialogue due to the decoder-only architecture of Llama 3. Subsequently, the global embedding " + }, + { + "bbox": [ + 313, + 506, + 555, + 663 + ], + "type": "inline_equation", + "content": "w_{eos}" + }, + { + "bbox": [ + 313, + 506, + 555, + 663 + ], + "type": "text", + "content": " is mapped into the shared space:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 388, + 673, + 553, + 687 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 388, + 673, + 553, + 687 + ], + "spans": [ + { + "bbox": [ + 388, + 673, + 553, + 687 + ], + "type": "interline_equation", + "content": "w = \\Phi_ {D} \\left(w _ {e o s}\\right) \\in \\mathbb {R} ^ {d}, \\tag {2}", + "image_path": "41db2baee0cb76ba66f153b1cd77a6add4e9a5464352ff793594921d158c4b89.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\Phi_D" + }, + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "content": " is a projector composed of a linear layer, " + }, + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "content": " is the dimension of the space." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3956" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 69, + 70, + 282, + 128 + ], + "blocks": [ + { + "bbox": [ + 69, + 70, + 282, + 128 + ], + "lines": [ + { + "bbox": [ + 69, + 70, + 282, + 128 + ], + "spans": [ + { + "bbox": [ + 69, + 70, + 282, + 128 + ], + "type": "image", + "image_path": "b707f2ce3ebe60a9367343ea56fbe0d52479ab6451938eaa147ef40cfc26e4e9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 132, + 295, + 164 + ], + "lines": [ + { + "bbox": [ + 55, + 132, + 295, + 164 + ], + "spans": [ + { + "bbox": [ + 55, + 132, + 295, + 164 + ], + "type": "text", + "content": "Figure 5. The instruction-formatted input sequence. A system message is prepended to prompt Llama 3 in capturing the semantics of the dialogue context." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 171, + 296, + 269 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 171, + 296, + 269 + ], + "spans": [ + { + "bbox": [ + 55, + 171, + 296, + 269 + ], + "type": "text", + "content": "For the visual branch, given an image " + }, + { + "bbox": [ + 55, + 171, + 296, + 269 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 55, + 171, + 296, + 269 + ], + "type": "text", + "content": ", we employ Swin Transformer [36, 37] to encode the image into a visual sequence " + }, + { + "bbox": [ + 55, + 171, + 296, + 269 + ], + "type": "inline_equation", + "content": "\\mathcal{V} = \\{v_{1}, v_{2}, \\dots, v_{h \\times w}\\}" + }, + { + "bbox": [ + 55, + 171, + 296, + 269 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 171, + 296, + 269 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 55, + 171, + 296, + 269 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 171, + 296, + 269 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 55, + 171, + 296, + 269 + ], + "type": "text", + "content": " represent the height and width of the feature map at the last stage, respectively. Subsequently, a global average pooling layer is adopted to integrate the overall semantics, yielding a global visual embedding " + }, + { + "bbox": [ + 55, + 171, + 296, + 269 + ], + "type": "inline_equation", + "content": "v_{g} = \\frac{1}{h \\times w} \\sum_{i=1}^{h \\times w} v_{i}" + }, + { + "bbox": [ + 55, + 171, + 296, + 269 + ], + "type": "text", + "content": ". Similarly, we introduce a projector to map " + }, + { + "bbox": [ + 55, + 171, + 296, + 269 + ], + "type": "inline_equation", + "content": "v_{g}" + }, + { + "bbox": [ + 55, + 171, + 296, + 269 + ], + "type": "text", + "content": " into the shared space:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 136, + 274, + 295, + 289 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 274, + 295, + 289 + ], + "spans": [ + { + "bbox": [ + 136, + 274, + 295, + 289 + ], + "type": "interline_equation", + "content": "v = \\Phi_ {I} (v _ {g}) \\in \\mathbb {R} ^ {d}. \\tag {3}", + "image_path": "d18ed8cfccf120198540cbdcbcc7e3d661f2beda0cc54e1c2ba2bfa6bf35b798.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 295, + 200, + 306 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 295, + 200, + 306 + ], + "spans": [ + { + "bbox": [ + 55, + 295, + 200, + 306 + ], + "type": "text", + "content": "4.1.2. Adaptive Attribute Refiner" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 309, + 296, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 309, + 296, + 429 + ], + "spans": [ + { + "bbox": [ + 55, + 309, + 296, + 429 + ], + "type": "text", + "content": "Unlike the single-shot text where content is generally concentrated, dialogue data typically exhibits sparser information density, requiring multiple rounds of interaction to ascertain specific attributes. This sparsity poses a significant challenge in cross-modal alignment, necessitating information refinement to effectively bridge the modality gap between dialogues and images. To address this, we propose two adaptive attribute refiner (AAR) modules on the dialogue and the image encoders to learn the attribute representations for fine-grained cross-modal alignment." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 430, + 296, + 596 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 430, + 296, + 596 + ], + "spans": [ + { + "bbox": [ + 55, + 430, + 296, + 596 + ], + "type": "text", + "content": "Specifically, we first introduce a dialogue AAR upon the dialogue encoder. Considering that different persons have varying attributes, we introduce a set of learnable attribute queries " + }, + { + "bbox": [ + 55, + 430, + 296, + 596 + ], + "type": "inline_equation", + "content": "\\mathcal{Q} = \\{q_1, q_2, \\dots, q_K\\}" + }, + { + "bbox": [ + 55, + 430, + 296, + 596 + ], + "type": "text", + "content": " to focus on the distinct attributes of each individual, where " + }, + { + "bbox": [ + 55, + 430, + 296, + 596 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 430, + 296, + 596 + ], + "type": "text", + "content": " denotes the number of attribute queries. By interacting with the dialogue embeddings " + }, + { + "bbox": [ + 55, + 430, + 296, + 596 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 55, + 430, + 296, + 596 + ], + "type": "text", + "content": ", these attribute queries adaptively extract the person-specific attribute representations from the sparse dialogue content. As illustrated in Fig. 4, the dialogue AAR applies a computationally efficient structure with a single multi-head cross attention (MCA) followed by 2-layer transformer blocks. The interaction between the attribute queries and the dialogue embeddings is performed through the cross-attention mechanism:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 94, + 602, + 295, + 616 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 602, + 295, + 616 + ], + "spans": [ + { + "bbox": [ + 94, + 602, + 295, + 616 + ], + "type": "interline_equation", + "content": "\\hat {\\mathcal {W}} = \\operatorname {T r a n s f o r m e r} (\\operatorname {M C A} (\\mathcal {Q}, \\mathcal {W}, \\mathcal {W})), \\tag {4}", + "image_path": "96e916bdf4d779f518aac9e8272fab6d1d750a3b6fb2e5fc253f092fe3d0f05e.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 622, + 295, + 646 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 622, + 295, + 646 + ], + "spans": [ + { + "bbox": [ + 55, + 622, + 295, + 646 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 622, + 295, + 646 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{W}} = \\{\\hat{w}_1,\\hat{w}_2,\\dots ,\\hat{w}_K\\}" + }, + { + "bbox": [ + 55, + 622, + 295, + 646 + ], + "type": "text", + "content": " indicates the adaptively refined dialogue attribute embeddings." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 647, + 296, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 647, + 296, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 647, + 296, + 693 + ], + "type": "text", + "content": "In parallel, we symmetrically introduce a visual AAR to extract the visual attributes from the image, mitigating the impacts of the inherent redundancy and background noise in visual information:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 100, + 700, + 295, + 714 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 700, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 100, + 700, + 295, + 714 + ], + "type": "interline_equation", + "content": "\\hat {\\mathcal {V}} = \\operatorname {T r a n s f o r m e r} (\\operatorname {M C A} (\\mathcal {Q}, \\mathcal {V}, \\mathcal {V})), \\tag {5}", + "image_path": "e849d1fc043dbb928babcf238d6a28429ca1598ead96fe630acbf281eb89feb1.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{V}} = \\{\\hat{v}_1,\\hat{v}_2,\\dots ,\\hat{v}_K\\}" + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "text", + "content": " denotes the visual attribute embeddings. Furthermore, we leverage a parameter-sharing mechanism between the two AAR modules, which not only reduces the parameter amount but also mitigates their modality gap [15] between dialogues and images." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 140, + 451, + 153 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 140, + 451, + 153 + ], + "spans": [ + { + "bbox": [ + 313, + 140, + 451, + 153 + ], + "type": "text", + "content": "4.2. Optimization Objectives" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 158, + 554, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 158, + 554, + 219 + ], + "spans": [ + { + "bbox": [ + 313, + 158, + 554, + 219 + ], + "type": "text", + "content": "Incorporating the dual encoders and the AAR modules, we construct DiaNA, which is then optimized through both global (coarse-grained) and local (fine-grained) alignment. Specifically, we use the cosine similarity to measure the global alignment between the dialogue and the image:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 389, + 227, + 554, + 255 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 389, + 227, + 554, + 255 + ], + "spans": [ + { + "bbox": [ + 389, + 227, + 554, + 255 + ], + "type": "interline_equation", + "content": "s _ {g} (\\mathbf {D}, \\mathbf {I}) = \\frac {w ^ {\\mathrm {T}} v}{\\| w \\| \\| v \\|}. \\tag {6}", + "image_path": "205b3ab6bd2672b7218c14b4171fa166a3a6af0eaa27d0800031277a9f54b0f6.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 263, + 554, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 263, + 554, + 300 + ], + "spans": [ + { + "bbox": [ + 313, + 263, + 554, + 300 + ], + "type": "text", + "content": "Furthermore, the average cosine distance among the refined attribute embeddings is utilized to measure the local semantics consistency between the dialogue and the image:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 372, + 309, + 554, + 342 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 309, + 554, + 342 + ], + "spans": [ + { + "bbox": [ + 372, + 309, + 554, + 342 + ], + "type": "interline_equation", + "content": "s _ {l} (\\mathbf {D}, \\mathbf {I}) = \\frac {1}{K} \\sum_ {i = 1} ^ {K} \\frac {\\hat {w} _ {i} ^ {\\mathrm {T}} \\hat {v} _ {i}}{\\| \\hat {w} _ {i} \\| \\| \\hat {v} _ {i} \\|.} \\tag {7}", + "image_path": "e82039e39752af055748636e70ffb169ac927053475913438d988886117fd75c.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 351, + 554, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 351, + 554, + 411 + ], + "spans": [ + { + "bbox": [ + 313, + 351, + 554, + 411 + ], + "type": "text", + "content": "According to the global and local similarities between the dialogue and the image, we construct the overall optimization object by employing the normalized image-text contrastive (N-ITC) loss [5] and the similarity distribution matching (SDM) loss [24]:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 358, + 419, + 553, + 450 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 419, + 553, + 450 + ], + "spans": [ + { + "bbox": [ + 358, + 419, + 553, + 450 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\underbrace {\\mathcal {L} _ {g} ^ {n i t c} + \\mathcal {L} _ {g} ^ {s d m}} _ {\\text {g l o b a l a l i n g m e n t}} + \\underbrace {\\mathcal {L} _ {l} ^ {n i t c} + \\mathcal {L} _ {l} ^ {s d m}} _ {\\text {l o c a l a l i n g m e n t}}. \\tag {8}", + "image_path": "d2e3b8717c627779d583c67942dfdbd2441ae41b9c8feb0b62c7e5a726b39343.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 469, + 511, + 482 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 469, + 511, + 482 + ], + "spans": [ + { + "bbox": [ + 313, + 469, + 511, + 482 + ], + "type": "text", + "content": "4.3. Dialogue-specific Data Augmentation" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 487, + 554, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 487, + 554, + 582 + ], + "spans": [ + { + "bbox": [ + 313, + 487, + 554, + 582 + ], + "type": "text", + "content": "In real-world scenarios, complete dialogues that capture users' full search intent are not always available, since obtaining these complete dialogues typically requires numerous rounds of interaction, which can be not user-friendly. To enhance the model's generalization ability over dialogues of varying rounds, we propose a dialogue-specific data augmentation strategy, random round retaining (R3), which simulates the inherent variability of the dialogue in practice." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 583, + 554, + 618 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 583, + 554, + 618 + ], + "spans": [ + { + "bbox": [ + 313, + 583, + 554, + 618 + ], + "type": "text", + "content": "Specifically, at each training iteration, for a complete dialogue " + }, + { + "bbox": [ + 313, + 583, + 554, + 618 + ], + "type": "inline_equation", + "content": "\\mathbf{D} = \\{(Q_i,A_i)\\}_{i = 1}^N" + }, + { + "bbox": [ + 313, + 583, + 554, + 618 + ], + "type": "text", + "content": ", only the initial random " + }, + { + "bbox": [ + 313, + 583, + 554, + 618 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 583, + 554, + 618 + ], + "type": "text", + "content": " rounds are retained and fed into the model:" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 347, + 628, + 553, + 643 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 628, + 553, + 643 + ], + "spans": [ + { + "bbox": [ + 347, + 628, + 553, + 643 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {D}} = \\left\\{\\left(Q _ {i}, A _ {i}\\right) \\right\\} _ {i = 1} ^ {n}, \\quad n \\sim \\operatorname {r a n d i n t} (1, N), \\tag {9}", + "image_path": "cc8ecca0b7d1e785d8b87410afbd85b6de741f654e17012ef83e828cccc6f107.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 652, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 652, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 652, + 554, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 652, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\operatorname{randint}(1, N)" + }, + { + "bbox": [ + 313, + 652, + 554, + 713 + ], + "type": "text", + "content": " denotes a discrete uniform distribution with equal probability for each integer in range " + }, + { + "bbox": [ + 313, + 652, + 554, + 713 + ], + "type": "inline_equation", + "content": "[1, N]" + }, + { + "bbox": [ + 313, + 652, + 554, + 713 + ], + "type": "text", + "content": ". The retained " + }, + { + "bbox": [ + 313, + 652, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{D}}" + }, + { + "bbox": [ + 313, + 652, + 554, + 713 + ], + "type": "text", + "content": " reflects the partial availability of information in practical scenarios, effectively enhancing the model's generalization ability over varying dialogue lengths." + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3957" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 87, + 70, + 529, + 176 + ], + "blocks": [ + { + "bbox": [ + 87, + 70, + 529, + 176 + ], + "lines": [ + { + "bbox": [ + 87, + 70, + 529, + 176 + ], + "spans": [ + { + "bbox": [ + 87, + 70, + 529, + 176 + ], + "type": "table", + "html": "
MethodRefPret. Dataw/o Data Cleaningw/ Data Cleaning
R-1R-5R-10mAPR-1R-5R-10mAP
IRRA [24]CVPR'23WIT46.0865.6073.0242.9048.2667.9675.6544.64
TBPS-CLIP [5]AAAI'24WIT47.7466.4173.7843.8050.2169.2776.4345.44
RDE [42]CVPR'24WIT50.2068.2575.1345.9452.7770.6177.4747.98
APTM [52]MM'23MALS39.6256.9964.2336.5741.1557.9165.1638.00
AUL [31]AAAI'24MALS40.0557.0164.4137.0242.0159.2166.0138.56
DiaNA (Ours)-MALS73.7289.6694.2465.4075.6790.6894.7166.89
", + "image_path": "65282b51bc2420e24e1c9ff4717dcdb53efa8bfebf9c1610592e7bb4ad13b81f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 178, + 555, + 201 + ], + "lines": [ + { + "bbox": [ + 55, + 178, + 555, + 201 + ], + "spans": [ + { + "bbox": [ + 55, + 178, + 555, + 201 + ], + "type": "text", + "content": "Table 2. Comparison with state-of-the-art methods on ChatPedes. \"Pret. Data\" denotes pretraining data. All reported results of other methods are reproduced using their publicly available code." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 209, + 137, + 222 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 209, + 137, + 222 + ], + "spans": [ + { + "bbox": [ + 55, + 209, + 137, + 222 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 228, + 296, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 228, + 296, + 266 + ], + "spans": [ + { + "bbox": [ + 55, + 228, + 296, + 266 + ], + "type": "text", + "content": "In this section, we conduct extensive experiments to verify the effectiveness of our method. The implementation details are provided in Appendix A." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 274, + 171, + 285 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 274, + 171, + 285 + ], + "spans": [ + { + "bbox": [ + 55, + 274, + 171, + 285 + ], + "type": "text", + "content": "5.1. Evaluation Protocol" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 292, + 296, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 292, + 296, + 365 + ], + "spans": [ + { + "bbox": [ + 55, + 292, + 296, + 365 + ], + "type": "text", + "content": "We adopt the widely-used Rank-K (R-K for short, " + }, + { + "bbox": [ + 55, + 292, + 296, + 365 + ], + "type": "inline_equation", + "content": "\\mathrm{K} = 1,5,10" + }, + { + "bbox": [ + 55, + 292, + 296, + 365 + ], + "type": "text", + "content": ") metric in our experiments, which denotes the percentage of successful retrieval that the ground-truth image is found within the top K ranked results. In addition, we also employ the mean average precision (mAP) as an auxiliary metric to evaluate the overall retrieval performance." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 372, + 234, + 384 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 372, + 234, + 384 + ], + "spans": [ + { + "bbox": [ + 55, + 372, + 234, + 384 + ], + "type": "text", + "content": "5.2. Comparison with SOTA Methods" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 390, + 295, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 390, + 295, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 390, + 295, + 605 + ], + "type": "text", + "content": "We compare our method against five state-of-the-art (SOTA) baselines: IRRA [24], TBPS-CLIP [5], RDE [42], APTM [52] and AUL [31]. Among these, IRRA, TBPS-CLIP and RDE are built on the vision-language foundation model CLIP [44], which is pretrained on a large-scale dataset WIT with 400 million image-text paired data collected from the Internet. APTM and AUL are pretrained on the synthetic text-image paired dataset MALS [52]. It is worth noting that these baselines are originally developed for the TPR task. We reproduce their performance on ChatPedes dataset using their open-source codes, treating the multi-turn dialogue data as a single long-form text input. As shown in Tab. 7, regardless of whether data cleaning is applied, our method outperforms these SOTA baselines by a large margin, while these baselines demonstrate a suboptimal retrieval accuracy, indicating their limitations in effectively handling the complex structure and longer contextual dependencies of dialogue data." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 606, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 606, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 606, + 296, + 713 + ], + "type": "text", + "content": "In addition, we further conduct extensive comparative experiments over the dialogue round, simulating real-world scenarios where users are pro-actively questioned to progressively refine their queries through conversational interactions. This process is exemplified in Appendix E.1. As shown in Fig. 6, it can be observed that: ① As the dialogue progresses, the retrieval performance improves across all methods. However, the performance of the baselines becomes saturated after the initial few rounds of interaction," + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 316, + 209, + 555, + 319 + ], + "blocks": [ + { + "bbox": [ + 316, + 209, + 555, + 319 + ], + "lines": [ + { + "bbox": [ + 316, + 209, + 555, + 319 + ], + "spans": [ + { + "bbox": [ + 316, + 209, + 555, + 319 + ], + "type": "image", + "image_path": "faf14a2c0db406e51a0611d2a6217a13c035ac18c6ca0e6791b2c644c3696b3b.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 331, + 322, + 537, + 333 + ], + "lines": [ + { + "bbox": [ + 331, + 322, + 537, + 333 + ], + "spans": [ + { + "bbox": [ + 331, + 322, + 537, + 333 + ], + "type": "text", + "content": "Figure 6. Performance comparison over dialogue round." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 344, + 555, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 344, + 555, + 525 + ], + "spans": [ + { + "bbox": [ + 313, + 344, + 555, + 525 + ], + "type": "text", + "content": "which we attribute to their limitations in processing longer and more complex dialogues. In contrast, our method presents a significantly pronounced improvement as the dialogue evolves on account of our carefully crafted design for handling dialogue data. This comprehensive comparison demonstrates the efficacy of our method and showcases its superior capability in handling the intricacies of dialogue-image alignment. Additionally, we also notice that our method achieves competitive performance at the first round of dialogue. This round typically involves a holistic description of the target person, functioning similarly to the single-shot text in the traditional TPR task. The experimental comparison also indicates the promising generalization capabilities of our method in TPR. Further detailed experimental results on TPR benchmarks are provided in Appendix D." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 535, + 408, + 548 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 535, + 408, + 548 + ], + "spans": [ + { + "bbox": [ + 313, + 535, + 408, + 548 + ], + "type": "text", + "content": "5.3. Ablation Study" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 554, + 473, + 566 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 554, + 473, + 566 + ], + "spans": [ + { + "bbox": [ + 313, + 554, + 473, + 566 + ], + "type": "text", + "content": "5.3.1. Effectiveness of Data Cleaning" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 570, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 555, + 715 + ], + "type": "text", + "content": "During dialogue construction, we leverage LLMs to generate dialogue data. However, a major limitation of this automatic generation through LLMs lies in the inevitable introduction of noise, which can be classified into three categories: Contradictory, Hallucinatory and Repeated, as delineated in Sec. 3.2. To evaluate the impact of the noise, we conduct comparative experiments with and without data cleaning. Tab. 7 presents the performance comparison with the five SOTA baselines, where all approaches exhibit a significant improvement. In particular, RDE achieves a notable increase of " + }, + { + "bbox": [ + 313, + 570, + 555, + 715 + ], + "type": "inline_equation", + "content": "2.57\\%" + }, + { + "bbox": [ + 313, + 570, + 555, + 715 + ], + "type": "text", + "content": " at R-1, demonstrating the effectiveness of data cleaning and the importance of high-quality data." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3958" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 63, + 70, + 281, + 137 + ], + "blocks": [ + { + "bbox": [ + 63, + 70, + 281, + 137 + ], + "lines": [ + { + "bbox": [ + 63, + 70, + 281, + 137 + ], + "spans": [ + { + "bbox": [ + 63, + 70, + 281, + 137 + ], + "type": "table", + "html": "
Data DiversityR-1R-5R-10mAP
Llama 3InternVL
X✓71.0188.4093.1863.08
✓X71.6288.5893.5963.73
✓✓75.6790.6894.7166.89
", + "image_path": "47737baa3e5c81569ec91f094b58f092de790fae81df169c3afa9ef67ba03c66.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 74, + 172, + 278, + 280 + ], + "blocks": [ + { + "bbox": [ + 74, + 172, + 278, + 280 + ], + "lines": [ + { + "bbox": [ + 74, + 172, + 278, + 280 + ], + "spans": [ + { + "bbox": [ + 74, + 172, + 278, + 280 + ], + "type": "image", + "image_path": "0abb2854e2375e367e1571ad2a47b4ebdf4ff8b69757ac381ff10e92e72f7f80.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 284, + 296, + 316 + ], + "lines": [ + { + "bbox": [ + 55, + 284, + 296, + 316 + ], + "spans": [ + { + "bbox": [ + 55, + 284, + 296, + 316 + ], + "type": "text", + "content": "Figure 7. Performance comparison across a varying number of attribute queries " + }, + { + "bbox": [ + 55, + 284, + 296, + 316 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 284, + 296, + 316 + ], + "type": "text", + "content": " in AAR modules. Notably, " + }, + { + "bbox": [ + 55, + 284, + 296, + 316 + ], + "type": "inline_equation", + "content": "K = 0" + }, + { + "bbox": [ + 55, + 284, + 296, + 316 + ], + "type": "text", + "content": " indicates that the AAR is not used." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 325, + 214, + 338 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 325, + 214, + 338 + ], + "spans": [ + { + "bbox": [ + 55, + 325, + 214, + 338 + ], + "type": "text", + "content": "5.3.2. Effectiveness of Data Diversity" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 341, + 296, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 341, + 296, + 510 + ], + "spans": [ + { + "bbox": [ + 55, + 341, + 296, + 510 + ], + "type": "text", + "content": "We employ Llama 3 [14] and InternVL [8] as different annotators for labeling diverse dialogue data when constructing ChatPedes. Tab. 3 presents the contribution of data diversity to the retrieval performance. It can be observed that: ① Relying on a single LLM yields a suboptimal outcome, with " + }, + { + "bbox": [ + 55, + 341, + 296, + 510 + ], + "type": "inline_equation", + "content": "71.01\\%" + }, + { + "bbox": [ + 55, + 341, + 296, + 510 + ], + "type": "text", + "content": " for InternVL and " + }, + { + "bbox": [ + 55, + 341, + 296, + 510 + ], + "type": "inline_equation", + "content": "71.62\\%" + }, + { + "bbox": [ + 55, + 341, + 296, + 510 + ], + "type": "text", + "content": " for Llama 3 at R-1. Additionally, Llama 3 slightly outperforms InternVL due to its more advanced capability in dialogue generation. ② The combination of Llama 3 and InternVL leads to a substantial increase across all metrics, demonstrating the importance of data diversity. By combining annotations generated by different LLMs, the established ChatPedes dataset effectively captures a wider variety of dialogue contexts and nuances, which in turn bolsters the model's retrieval capabilities." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 518, + 267, + 530 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 518, + 267, + 530 + ], + "spans": [ + { + "bbox": [ + 55, + 518, + 267, + 530 + ], + "type": "text", + "content": "5.3.3. Effectiveness of Adaptive Attribute Refiner" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 533, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 533, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 533, + 295, + 713 + ], + "type": "text", + "content": "Considering the sparse information density in dialogues and the inherent redundancy in images, we propose two adaptive attribute refiner (AAR) modules, which employ a set of learnable attribute-specific queries to adaptively extract the relevant attributes from the dialogue and the image. Fig. 7 illustrates the performance variation over the number of attribute queries " + }, + { + "bbox": [ + 55, + 533, + 295, + 713 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 533, + 295, + 713 + ], + "type": "text", + "content": ", where, notably, " + }, + { + "bbox": [ + 55, + 533, + 295, + 713 + ], + "type": "inline_equation", + "content": "K = 0" + }, + { + "bbox": [ + 55, + 533, + 295, + 713 + ], + "type": "text", + "content": " denotes AAR is discarded. We observe that: 1 Only relying on global alignment without AAR results in a suboptimal retrieval performance, underscoring the importance of key information extraction. 2 When introducing AAR, the performance exhibits slight fluctuations, reaching the peak of " + }, + { + "bbox": [ + 55, + 533, + 295, + 713 + ], + "type": "inline_equation", + "content": "75.67\\%" + }, + { + "bbox": [ + 55, + 533, + 295, + 713 + ], + "type": "text", + "content": " on R-1 and " + }, + { + "bbox": [ + 55, + 533, + 295, + 713 + ], + "type": "inline_equation", + "content": "66.89\\%" + }, + { + "bbox": [ + 55, + 533, + 295, + 713 + ], + "type": "text", + "content": " on mAP at " + }, + { + "bbox": [ + 55, + 533, + 295, + 713 + ], + "type": "inline_equation", + "content": "K = 16" + }, + { + "bbox": [ + 55, + 533, + 295, + 713 + ], + "type": "text", + "content": ". Overall, the performance remains relatively stable, indicating the model's robustness when equipped with AAR." + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 321, + 71, + 547, + 163 + ], + "blocks": [ + { + "bbox": [ + 55, + 141, + 295, + 163 + ], + "lines": [ + { + "bbox": [ + 55, + 141, + 295, + 163 + ], + "spans": [ + { + "bbox": [ + 55, + 141, + 295, + 163 + ], + "type": "text", + "content": "Table 3. Ablation study on data diversity. We employ Llama 3 [14] and InternVL [8] to construct the ChatPedes benchmark." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 321, + 71, + 547, + 163 + ], + "lines": [ + { + "bbox": [ + 321, + 71, + 547, + 163 + ], + "spans": [ + { + "bbox": [ + 321, + 71, + 547, + 163 + ], + "type": "table", + "html": "
MethodMetricDialogue Round
12468all
DiaNA(w/o R3)26.5034.8953.0766.4073.1675.18
DiaNA(w/ R3)R-129.2537.6455.6067.9973.6775.67
Δ+2.75+2.75+2.53+1.59+0.51+0.49
DiaNA(w/o R3)26.3333.6749.0659.6764.6466.19
DiaNA(w/ R3)mAP28.7836.3751.0560.8865.3266.89
Δ+2.45+2.70+1.99+1.21+0.68+0.70
", + "image_path": "f235b15f610fc83c69cfd403e69607b7dc696961444a551e5376e21a59419b82.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 165, + 553, + 187 + ], + "lines": [ + { + "bbox": [ + 313, + 165, + 553, + 187 + ], + "spans": [ + { + "bbox": [ + 313, + 165, + 553, + 187 + ], + "type": "text", + "content": "Table 4. Comparison of random round retaining (R3) over dialogue round. \"all\" refers to the complete dialogue." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 198, + 524, + 210 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 198, + 524, + 210 + ], + "spans": [ + { + "bbox": [ + 313, + 198, + 524, + 210 + ], + "type": "text", + "content": "5.3.4. Effectiveness of Random Round Retaining" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 213, + 555, + 406 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 213, + 555, + 406 + ], + "spans": [ + { + "bbox": [ + 313, + 213, + 555, + 406 + ], + "type": "text", + "content": "The proposed random round retaining (R3) strategy selectively retains a random number of initial rounds of the dialogue during training, simulating the users' incomplete search intent in real-world interactions. As shown in Tab. 4, we observe that: " + }, + { + "bbox": [ + 313, + 213, + 555, + 406 + ], + "type": "inline_equation", + "content": "\\mathbf{1}" + }, + { + "bbox": [ + 313, + 213, + 555, + 406 + ], + "type": "text", + "content": " R3 yields performance gains across all dialogue rounds, with gains from " + }, + { + "bbox": [ + 313, + 213, + 555, + 406 + ], + "type": "inline_equation", + "content": "2.75\\%" + }, + { + "bbox": [ + 313, + 213, + 555, + 406 + ], + "type": "text", + "content": " in single-round interactions to " + }, + { + "bbox": [ + 313, + 213, + 555, + 406 + ], + "type": "inline_equation", + "content": "0.49\\%" + }, + { + "bbox": [ + 313, + 213, + 555, + 406 + ], + "type": "text", + "content": " in complete conversations at R-1, demonstrating its robustness and versatility. " + }, + { + "bbox": [ + 313, + 213, + 555, + 406 + ], + "type": "inline_equation", + "content": "\\mathbf{2}" + }, + { + "bbox": [ + 313, + 213, + 555, + 406 + ], + "type": "text", + "content": " In particular, R3 brings more substantial improvement at the initial few rounds of dialogues, suggesting that R3 is particularly effective when dialogue information is limited. This is especially crucial in practical scenarios where users expect to quickly retrieve the target person with minimal interaction. Overall, R3 empowers users to obtain accurate retrieval results even in the early stages of interaction, thereby addressing the real-world demands for both efficiency and precision." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 417, + 388, + 430 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 417, + 388, + 430 + ], + "spans": [ + { + "bbox": [ + 313, + 417, + 388, + 430 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 437, + 555, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 555, + 652 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 555, + 652 + ], + "type": "text", + "content": "Recognizing the inherent limitations of traditional TPR where single-shot text queries fail to fully encapsulate users' search intent, we introduce ChatPR, a pioneering paradigm for person retrieval that engages users in interactive dialogues to progressively refine their search queries. For this, we build the first benchmark ChatPedes through LLMs to automate question generation and simulate users' responses. To tackle this task, we propose the DiaNA framework that leverages two adaptive attribute refiners to bottleneck the conversational and visual information for fine-grained cross-modal alignment. Additionally, a novel data augmentation strategy of random round retaining is proposed to mimic the incomplete queries in real-world scenarios, remarkably enhancing the model's generalization ability across varying dialogue lengths. Extensive experiments significantly demonstrate the importance of conversational interaction in person retrieval and the superior effectiveness of our method in ChatPR." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 654, + 554, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 554, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 554, + 715 + ], + "type": "text", + "content": "Acknowledgement. This work is supported by the National Science Foundation of China under Grants (62176188, 62361166629, 62476188). The numerical calculations are supported by the supercomputing system in the Supercomputing Center of Wuhan University." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3959" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 156 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 156 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 156 + ], + "type": "text", + "content": "[1] Yang Bai, Min Cao, Daming Gao, Ziqiang Cao, Chen Chen, Zhenfeng Fan, Liqiang Nie, and Min Zhang. Rasa: Relation and sensitivity aware representation learning for text-based person search. In Proceedings of the Thirty-Second International Joint Conference on Artificial Intelligence, pages 555–563, 2023. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 158, + 296, + 212 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 158, + 296, + 212 + ], + "spans": [ + { + "bbox": [ + 61, + 158, + 296, + 212 + ], + "type": "text", + "content": "[2] Yang Bai, Jingyao Wang, Min Cao, Chen Chen, Ziqiang Cao, Liqiang Nie, and Min Zhang. Text-based person search without parallel image-text data. In Proceedings of the 31st ACM International Conference on Multimedia, pages 757-767, 2023. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 214, + 294, + 235 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 214, + 294, + 235 + ], + "spans": [ + { + "bbox": [ + 62, + 214, + 294, + 235 + ], + "type": "text", + "content": "[3] Tom B Brown. Language models are few-shot learners. arXiv preprint arXiv:2005.14165, 2020. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 236, + 294, + 301 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 236, + 294, + 301 + ], + "spans": [ + { + "bbox": [ + 62, + 236, + 294, + 301 + ], + "type": "text", + "content": "[4] Guanyu Cai, Jun Zhang, Xinyang Jiang, Yifei Gong, Lianghua He, Fufu Yu, Pai Peng, Xiaowei Guo, Feiyue Huang, and Xing Sun. Ask&confirm: active detail enriching for cross-modal retrieval with partial query. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1835-1844, 2021. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 303, + 294, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 303, + 294, + 346 + ], + "spans": [ + { + "bbox": [ + 62, + 303, + 294, + 346 + ], + "type": "text", + "content": "[5] Min Cao, Yang Bai, Ziyin Zeng, Mang Ye, and Min Zhang. An empirical study of clip for text-based person search. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 465-473, 2024. 3, 6, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 347, + 294, + 401 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 347, + 294, + 401 + ], + "spans": [ + { + "bbox": [ + 62, + 347, + 294, + 401 + ], + "type": "text", + "content": "[6] Yupeng Chang, Xu Wang, Jindong Wang, Yuan Wu, Linyi Yang, Kaijie Zhu, Hao Chen, Xiaoyuan Yi, Cunxiang Wang, Yidong Wang, et al. A survey on evaluation of large language models. ACM Transactions on Intelligent Systems and Technology, 15(3):1-45, 2024. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 403, + 294, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 403, + 294, + 446 + ], + "spans": [ + { + "bbox": [ + 62, + 403, + 294, + 446 + ], + "type": "text", + "content": "[7] Yuhao Chen, Guoqing Zhang, Yujiang Lu, Zhenxing Wang, and Yuhui Zheng. Tipcb: A simple but effective part-based convolutional baseline for text-based person search. Neurocomputing, 494:171-181, 2022. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 447, + 294, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 447, + 294, + 513 + ], + "spans": [ + { + "bbox": [ + 62, + 447, + 294, + 513 + ], + "type": "text", + "content": "[8] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24185-24198, 2024. 2, 4, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 514, + 294, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 514, + 294, + 578 + ], + "spans": [ + { + "bbox": [ + 62, + 514, + 294, + 578 + ], + "type": "text", + "content": "[9] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhang-hao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yong-hao Zhuang, Joseph E. Gonzalez, Ion Stoica, and Eric P. Xing. Vicuna: An open-source chatbot impressing gpt-4 with " + }, + { + "bbox": [ + 62, + 514, + 294, + 578 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 62, + 514, + 294, + 578 + ], + "type": "text", + "content": " chatgpt quality. https://lmsys.org/blog/2023-03-30-vicuna, 2023.5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 580, + 294, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 580, + 294, + 634 + ], + "spans": [ + { + "bbox": [ + 57, + 580, + 294, + 634 + ], + "type": "text", + "content": "[10] Abhishek Das, Satwik Kottur, Khushi Gupta, Avi Singh, Deshraj Yadav, Jose M. F. Moura, Devi Parikh, and Dhruv Batra. Visual dialog. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 326-335, 2017. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 636, + 294, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 636, + 294, + 667 + ], + "spans": [ + { + "bbox": [ + 57, + 636, + 294, + 667 + ], + "type": "text", + "content": "[11] Jacob Devlin. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 2, 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 58, + 670, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 670, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 670, + 294, + 712 + ], + "type": "text", + "content": "[12] Zefeng Ding, Changxing Ding, Zhiyin Shao, and Dacheng Tao. Semantically self-aligned network for text-to-image part-aware person re-identification. arXiv preprint arXiv:2107.12666, 2021. 2, 4" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "text", + "content": "[13] Qingxiu Dong, Lei Li, Damai Dai, Ce Zheng, Zhiyong Wu, Baobao Chang, Xu Sun, Jingjing Xu, and Zhifang Sui. A survey on in-context learning. arXiv preprint arXiv:2301.00234, 2022. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 118, + 553, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 118, + 553, + 171 + ], + "spans": [ + { + "bbox": [ + 317, + 118, + 553, + 171 + ], + "type": "text", + "content": "[14] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 2, 4, 5, 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 173, + 553, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 173, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 317, + 173, + 553, + 205 + ], + "type": "text", + "content": "[15] Sedigheh Eslami and Gerard de Melo. Mitigate the gap: Investigating approaches for improving cross-modal alignment in clip. arXiv preprint arXiv:2406.17639, 2024. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 206, + 553, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 206, + 553, + 249 + ], + "spans": [ + { + "bbox": [ + 317, + 206, + 553, + 249 + ], + "type": "text", + "content": "[16] Yiyang Fang, Wenke Huang, Guancheng Wan, Kehua Su, and Mang Ye. Emoe: Modality-specific enhanced dynamic emotion experts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2025. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 251, + 553, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 251, + 553, + 293 + ], + "spans": [ + { + "bbox": [ + 317, + 251, + 553, + 293 + ], + "type": "text", + "content": "[17] Ammarah Farooq, Muhammad Awais, Fei Yan, Josef Kittler, Ali Akbari, and Syed Safwan Khalid. A convolutional baseline for person re-identification using vision and language descriptions. arXiv preprint arXiv:2003.00808, 2020. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 294, + 553, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 294, + 553, + 338 + ], + "spans": [ + { + "bbox": [ + 317, + 294, + 553, + 338 + ], + "type": "text", + "content": "[18] Takuro Fujii and Shuhei Tarashima. Bilma: Bidirectional local-matching for text-based person re-identification. In Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops, pages 2786-2790, 2023. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 339, + 553, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 339, + 553, + 392 + ], + "spans": [ + { + "bbox": [ + 317, + 339, + 553, + 392 + ], + "type": "text", + "content": "[19] Chenyang Gao, Guanyu Cai, Xinyang Jiang, Feng Zheng, Jun Zhang, Yifei Gong, Fangzhou Lin, Xing Sun, and Xi-ang Bai. Conditional feature learning based transformer for text-based person search. IEEE Transactions on Image Processing, 31:6097–6108, 2022. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 394, + 553, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 394, + 553, + 436 + ], + "spans": [ + { + "bbox": [ + 317, + 394, + 553, + 436 + ], + "type": "text", + "content": "[20] Fabrizio Gilardi, Meysam Alizadeh, and Maël Kubli. Chatgpt outperforms crowd workers for text-annotation tasks. Proceedings of the National Academy of Sciences, 120(30): e2305016120, 2023. 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 437, + 553, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 437, + 553, + 480 + ], + "spans": [ + { + "bbox": [ + 317, + 437, + 553, + 480 + ], + "type": "text", + "content": "[21] Xiaoxiao Guo, Hui Wu, Yu Cheng, Steven Rennie, Gerald Tesauro, and Rogerio Feris. Dialog-based interactive image retrieval. In Advances in Neural Information Processing Systems, 2018. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 482, + 553, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 482, + 553, + 536 + ], + "spans": [ + { + "bbox": [ + 317, + 482, + 553, + 536 + ], + "type": "text", + "content": "[22] Yuxin Hou, Eleonora Vig, Michael Donoser, and Loris Bazzani. Learning attribute-driven disentangled representations for interactive fashion retrieval. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12147-12157, 2021. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 537, + 553, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 537, + 553, + 581 + ], + "spans": [ + { + "bbox": [ + 317, + 537, + 553, + 581 + ], + "type": "text", + "content": "[23] S. Irene, A. John Prakash, and V. Rhymend Uthariaraj. Person search over security video surveillance systems using deep learning methods: A review. Image and Vision Computing, 143:104930, 2024. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 582, + 553, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 582, + 553, + 634 + ], + "spans": [ + { + "bbox": [ + 317, + 582, + 553, + 634 + ], + "type": "text", + "content": "[24] Ding Jiang and Mang Ye. Cross-modal implicit relation reasoning and aligning for text-to-image person retrieval. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2787-2797, 2023. 2, 3, 6, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 636, + 553, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 636, + 553, + 690 + ], + "spans": [ + { + "bbox": [ + 317, + 636, + 553, + 690 + ], + "type": "text", + "content": "[25] Gi-Cheon Kang, Sungdong Kim, Jin-Hwa Kim, Donghyun Kwak, and Byoung-Tak Zhang. The dialog must go on: Improving visual dialog via generative self-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6746-6756, 2023. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "text", + "content": "[26] Saehyung Lee, Sangwon Yu, Junsung Park, Jihun Yi, and Sungroh Yoon. Interactive text-to-image retrieval with large" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "3960" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 76, + 73, + 295, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 73, + 295, + 106 + ], + "spans": [ + { + "bbox": [ + 76, + 73, + 295, + 106 + ], + "type": "text", + "content": "language models: A plug-and-play approach. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics, pages 791-809, 2024. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 295, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 295, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 295, + 150 + ], + "type": "text", + "content": "[27] Matan Levy, Rami Ben-Ari, Nir Darshan, and Dani Lischinski. Chatting makes perfect: Chat-based image retrieval. In Advances in Neural Information Processing Systems, pages 61437-61449, 2023. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 294, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 294, + 207 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 294, + 207 + ], + "type": "text", + "content": "[28] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi. Align before fuse: Vision and language representation learning with momentum distillation. In Advances in Neural Information Processing Systems, pages 9694-9705, 2021. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 209, + 294, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 209, + 294, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 209, + 294, + 262 + ], + "type": "text", + "content": "[29] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. BLIP: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In Proceedings of the 39th International Conference on Machine Learning, pages 12888-12900, 2022. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 264, + 294, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 264, + 294, + 317 + ], + "spans": [ + { + "bbox": [ + 56, + 264, + 294, + 317 + ], + "type": "text", + "content": "[30] Shuang Li, Tong Xiao, Hongsheng Li, Bolei Zhou, Dayu Yue, and Xiaogang Wang. Person search with natural language description. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2017. 2, 3, 4, 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 320, + 294, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 320, + 294, + 365 + ], + "spans": [ + { + "bbox": [ + 56, + 320, + 294, + 365 + ], + "type": "text", + "content": "[31] Shenshen Li, Chen He, Xing Xu, Fumin Shen, Yang Yang, and Heng Tao Shen. Adaptive uncertainty-based learning for text-based person retrieval. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3172-3180, 2024. 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 365, + 294, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 365, + 294, + 420 + ], + "spans": [ + { + "bbox": [ + 56, + 365, + 294, + 420 + ], + "type": "text", + "content": "[32] Jian Liang, Wenke Huang, Guancheng Wan, Qu Yang, and Mang Ye. Lorasculpt: Sculpting lora for harmonizing general and specialized knowledge in multimodal large language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2025. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 421, + 294, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 421, + 294, + 465 + ], + "spans": [ + { + "bbox": [ + 56, + 421, + 294, + 465 + ], + "type": "text", + "content": "[33] Kaiqu Liang and Samuel Albanie. Simple baselines for interactive video retrieval with questions and answers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11091-11101, 2023. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 466, + 294, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 466, + 294, + 499 + ], + "spans": [ + { + "bbox": [ + 56, + 466, + 294, + 499 + ], + "type": "text", + "content": "[34] Fangyi Liu, Mang Ye, and Bo Du. Learning a generalizable re-identification model from unlabelled data with domain-agnostic expert. Visual Intelligence, 2(1):28, 2024. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 500, + 294, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 500, + 294, + 533 + ], + "spans": [ + { + "bbox": [ + 56, + 500, + 294, + 533 + ], + "type": "text", + "content": "[35] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In Advances in Neural Information Processing Systems, pages 34892-34916, 2023. 2, 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 535, + 294, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 535, + 294, + 589 + ], + "spans": [ + { + "bbox": [ + 56, + 535, + 294, + 589 + ], + "type": "text", + "content": "[36] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 590, + 294, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 590, + 294, + 655 + ], + "spans": [ + { + "bbox": [ + 56, + 590, + 294, + 655 + ], + "type": "text", + "content": "[37] Ze Liu, Han Hu, Yutong Lin, Zhuliang Yao, Zhenda Xie, Yixuan Wei, Jia Ning, Yue Cao, Zheng Zhang, Li Dong, Furu Wei, and Baining Guo. Swin transformer v2: Scaling up capacity and resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12009-12019, 2022. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 657, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 657, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 657, + 294, + 689 + ], + "type": "text", + "content": "[38] Renze Lou, Kai Zhang, and Wenpeng Yin. A comprehensive survey on instruction following. arXiv preprint arXiv:2303.10475, 2023. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "text", + "content": "[39] Kai Niu, Yanyi Liu, Yuzhou Long, Yan Huang, Liang Wang, and Yanning Zhang. An overview of text-based person" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 106 + ], + "type": "text", + "content": "search: Recent advances and future directions. IEEE Transactions on Circuits and Systems for Video Technology, 34(9): 7803-7819, 2024. 1, 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 108, + 483, + 119 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 108, + 483, + 119 + ], + "spans": [ + { + "bbox": [ + 316, + 108, + 483, + 119 + ], + "type": "text", + "content": "[40] OpenAI. Gpt-4 technical report, 2023. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 121, + 553, + 143 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 121, + 553, + 143 + ], + "spans": [ + { + "bbox": [ + 316, + 121, + 553, + 143 + ], + "type": "text", + "content": "[41] OpenAI. Chatgpt. https://openai.com/blog/chatgpt, 2023.3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 145, + 553, + 200 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 145, + 553, + 200 + ], + "spans": [ + { + "bbox": [ + 316, + 145, + 553, + 200 + ], + "type": "text", + "content": "[42] Yang Qin, Yingke Chen, Dezhong Peng, Xi Peng, Joey Tianyi Zhou, and Peng Hu. Noisy-correspondence learning for text-to-image person re-identification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 27197-27206, 2024. 2, 3, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 202, + 553, + 235 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 202, + 553, + 235 + ], + "spans": [ + { + "bbox": [ + 316, + 202, + 553, + 235 + ], + "type": "text", + "content": "[43] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 237, + 553, + 313 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 237, + 553, + 313 + ], + "spans": [ + { + "bbox": [ + 316, + 237, + 553, + 313 + ], + "type": "text", + "content": "[44] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In Proceedings of the 38th International Conference on Machine Learning, pages 8748-8763, 2021. 2, 3, 5, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 316, + 553, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 316, + 553, + 370 + ], + "spans": [ + { + "bbox": [ + 316, + 316, + 553, + 370 + ], + "type": "text", + "content": "[45] Zhiyin Shao, Xinyu Zhang, Meng Fang, Zhifeng Lin, Jian Wang, and Changxing Ding. Learning granularity-unified representations for text-to-image person re-identification. In Proceedings of the 30th ACM International Conference on Multimedia, pages 5566–5574, 2022. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 373, + 553, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 373, + 553, + 404 + ], + "spans": [ + { + "bbox": [ + 316, + 373, + 553, + 404 + ], + "type": "text", + "content": "[46] Qwen Team. Qwen2.5: A party of foundation models. https://qwenlm.github.io/blog/qwen2.5, 2024.4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 407, + 553, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 407, + 553, + 460 + ], + "spans": [ + { + "bbox": [ + 316, + 407, + 553, + 460 + ], + "type": "text", + "content": "[47] Jason Wei, Yi Tay, Rishi Bommasani, Colin Raffel, Barret Zoph, Sebastian Borgeaud, Dani Yogatama, Maarten Bosma, Denny Zhou, Donald Metzler, et al. Emergent abilities of large language models. arXiv preprint arXiv:2206.07682, 2022.3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 464, + 553, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 464, + 553, + 508 + ], + "spans": [ + { + "bbox": [ + 316, + 464, + 553, + 508 + ], + "type": "text", + "content": "[48] Hong Wu, Hanqing Lu, and Songde Ma. Willhunter: interactive image retrieval with multilevel relevance. In Proceedings of the 17th International Conference on Pattern Recognition, pages 1009-1012, 2004. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 510, + 553, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 510, + 553, + 553 + ], + "spans": [ + { + "bbox": [ + 316, + 510, + 553, + 553 + ], + "type": "text", + "content": "[49] Yingjia Xu, Mengxia Wu, Zixin Guo, Min Cao, Mang Ye, and Jorma Laaksonen. Efficient text-to-video retrieval via multi-modal multi-tags derived pre-screening. Visual Intelligence, 3(1):1-13, 2025. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 555, + 553, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 555, + 553, + 610 + ], + "spans": [ + { + "bbox": [ + 316, + 555, + 553, + 610 + ], + "type": "text", + "content": "[50] Jingfeng Yang, Hongye Jin, Ruixiang Tang, Xiaotian Han, Qizhang Feng, Haoming Jiang, Shaochen Zhong, Bing Yin, and Xia Hu. Harnessing the power of llms in practice: A survey on chatgpt and beyond. ACM Transactions on Knowledge Discovery from Data, 18(6):1-32, 2024. 2, 4" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 613, + 553, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 613, + 553, + 656 + ], + "spans": [ + { + "bbox": [ + 316, + 613, + 553, + 656 + ], + "type": "text", + "content": "[51] Jinxi Yang, He Li, Bo Du, and Mang Ye. Cheb-gr: Rethinking k-nearest neighbor search in re-ranking for person re-identification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2025. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 658, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 553, + 713 + ], + "type": "text", + "content": "[52] Shuyu Yang, Yinan Zhou, Zhedong Zheng, Yaxiong Wang, Li Zhu, and Yujiao Wu. Towards unified text-based person retrieval: A large-scale multi-attribute and language search benchmark. In Proceedings of the 31st ACM International Conference on Multimedia, pages 4492–4501, 2023. 7" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "3961" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 574 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 126 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 126 + ], + "type": "text", + "content": "[53] Mang Ye, Jianbing Shen, Gaojie Lin, Tao Xiang, Ling Shao, and Steven C. H. Hoi. Deep learning for person re-identification: A survey and outlook. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(6):2872-2893, 2022. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 294, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 294, + 173 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 294, + 173 + ], + "type": "text", + "content": "[54] Mang Ye, Shuoyi Chen, Chenyue Li, Wei-Shi Zheng, David Crandall, and Bo Du. Transformer for object re-identification: A survey. International Journal of Computer Vision, pages 1–31, 2024. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 175, + 294, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 294, + 217 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 294, + 217 + ], + "type": "text", + "content": "[55] Mang Ye, Xuankun Rong, Wenke Huang, Bo Du, Nenghai Yu, and Dacheng Tao. A survey of safety on large vision-language models: Attacks, defenses and evaluations. arXiv preprint arXiv:2502.14881, 2025. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 219, + 294, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 219, + 294, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 219, + 294, + 262 + ], + "type": "text", + "content": "[56] Mang Ye, Zesen Wu, and Bo Du. Dual-level matching with outlier filtering for unsupervised visible-infrared person re-identification. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 263, + 294, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 263, + 294, + 306 + ], + "spans": [ + { + "bbox": [ + 56, + 263, + 294, + 306 + ], + "type": "text", + "content": "[57] Chiyuan Zhang, Samy Bengio, Moritz Hardt, Benjamin Recht, and Oriol Vinyals. Understanding deep learning (still) requires rethinking generalization. Communications of the ACM, 64(3):107-115, 2021. 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 309, + 294, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 309, + 294, + 350 + ], + "spans": [ + { + "bbox": [ + 56, + 309, + 294, + 350 + ], + "type": "text", + "content": "[58] Ying Zhang and Hutchuan Lu. Deep cross-modal projection learning for image-text matching. In Proceedings of the European Conference on Computer Vision, pages 686-701, 2018. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 354, + 294, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 354, + 294, + 396 + ], + "spans": [ + { + "bbox": [ + 56, + 354, + 294, + 396 + ], + "type": "text", + "content": "[59] Wayne Xin Zhao, Kun Zhou, Junyi Li, Tianyi Tang, Xiaolei Wang, Yupeng Hou, Yingqian Min, Beichen Zhang, Junjie Zhang, Zican Dong, et al. A survey of large language models. arXiv preprint arXiv:2303.18223, 2023. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 399, + 294, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 399, + 294, + 453 + ], + "spans": [ + { + "bbox": [ + 56, + 399, + 294, + 453 + ], + "type": "text", + "content": "[60] Zhedong Zheng, Liang Zheng, Michael Garrett, Yi Yang, Mingliang Xu, and Yi-Dong Shen. Dual-path convolutional image-text embeddings with instance loss. ACM Transactions on Multimedia Computing, Communications, and Applications, 16(2):1-23, 2020. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 454, + 294, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 454, + 294, + 508 + ], + "spans": [ + { + "bbox": [ + 56, + 454, + 294, + 508 + ], + "type": "text", + "content": "[61] Aichun Zhu, Zijie Wang, Yifeng Li, Xili Wan, Jing Jin, Tian Wang, Fangqiang Hu, and Gang Hua. Dssl: Deep surroundings-person separation learning for text-based person retrieval. In Proceedings of the 29th ACM International Conference on Multimedia, page 209-217, 2021. 2, 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 510, + 294, + 574 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 510, + 294, + 574 + ], + "spans": [ + { + "bbox": [ + 56, + 510, + 294, + 574 + ], + "type": "text", + "content": "[62] Jialong Zuo, Hanyu Zhou, Ying Nie, Feng Zhang, Tianyu Guo, Nong Sang, Yunhe Wang, and Changxin Gao. Ufinebench: Towards text-based person retrieval with ultra-fine granularity. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22010-22019, 2024. 2, 4" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3962" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Chat2SVG_ Vector Graphics Generation with Large Language Models and Image Diffusion Models/03e11a55-7ca3-4174-a887-3307219befcc_content_list.json b/2025/Chat2SVG_ Vector Graphics Generation with Large Language Models and Image Diffusion Models/03e11a55-7ca3-4174-a887-3307219befcc_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..9aad6a3c2863ef3facf858e63ff6cc483b2b604c --- /dev/null +++ b/2025/Chat2SVG_ Vector Graphics Generation with Large Language Models and Image Diffusion Models/03e11a55-7ca3-4174-a887-3307219befcc_content_list.json @@ -0,0 +1,1468 @@ +[ + { + "type": "text", + "text": "Chat2SVG: Vector Graphics Generation with Large Language Models and Image Diffusion Models", + "text_level": 1, + "bbox": [ + 122, + 128, + 875, + 176 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/aa1694faa01481c5c6a00468d7bfb09ff0a914d584a65416ee9bddad38c0d8f3.jpg", + "image_caption": [ + "Ronghuan Wu \nCity University of Hong Kong \nrh.wu@my.cityu.edu.hk", + "Wanchao Su Monash University wanchao.su@monash.edu", + "Jing Liao* \nCity University of Hong Kong \njingliao@cityu.edu.hk", + "Figure 1. SVG examples generated by our Chat2SVG. We highlight some shapes to demonstrate semantic clarity and path quality." + ], + "image_footnote": [], + "bbox": [ + 122, + 262, + 883, + 523 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 553, + 326, + 569 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Scalable Vector Graphics (SVG) has become the de facto standard for vector graphics in digital design, offering resolution independence and precise control over individual elements. Despite their advantages, creating high-quality SVG content remains challenging, as it demands technical expertise with professional editing software and a considerable time investment to craft complex shapes. Recent text-to-SVG generation methods aim to make vector graphics creation more accessible, but they still encounter limitations in shape regularity, generalization ability, and expressiveness. To address these challenges, we introduce Chat2SVG, a hybrid framework that combines the strengths of Large Language Models (LLMs) and image diffusion models for text-to-SVG generation. Our approach first uses an LLM to generate semantically meaningful SVG templates from basic geometric primitives. Guided by image diffusion models, a dual-stage optimization pipeline refines paths in latent space and adjusts point coordinates to enhance geometric complexity. Extensive experiments show", + "bbox": [ + 89, + 585, + 483, + 875 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "that Chat2SVG outperforms existing methods in visual fidelity, path regularity, and semantic alignment. Additionally, our system enables intuitive editing through natural language instructions, making professional vector graphics creation accessible to all users. Our code is available at https://chat2svg.github.io/.", + "bbox": [ + 511, + 555, + 906, + 646 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 676, + 643, + 691 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Scalable Vector Graphics (SVG), a vector image format based on geometric shapes, has become the standard for modern digital design, offering resolution independence and precise control over individual elements. However, creating high-quality SVG content remains challenging for non-expert users, as it requires both expertise with professional design software and considerable time investment to create complex shapes. To make vector graphics creation more accessible, recent research has focused on developing text-to-SVG systems that enable users to express their creative ideas through simple text prompts rather than complex manual editing.", + "bbox": [ + 509, + 703, + 906, + 883 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Existing approaches have explored image-based meth", + "bbox": [ + 532, + 885, + 903, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding Author.", + "bbox": [ + 109, + 887, + 240, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "23690", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ods for text-to-SVG generation, iteratively optimizing a large collection of shape elements (e.g., cubic Bezier curves, typically 100 to 1000) by rendering them into images with differentiable rasterizers [20] and evaluating them using text-image similarity metrics like CLIP loss [28] and Score Distillation Sampling (SDS) loss [27]. While these image-based methods [9, 16, 43, 44, 50] can generate visually impressive SVG through the combination of numerous paths and powerful image models, they face a critical limitation in maintaining the regularity and semantics of shapes. Specifically, semantic components that should be represented as single elements often end up fragmented across multiple overlapping paths. Although these fragmented paths, when viewed collectively, can yield visually appealing outputs, they fundamentally conflict with professional design principles, where each semantic component is intentionally crafted as a single, regularized path.", + "bbox": [ + 89, + 90, + 480, + 348 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Given SVG is defined using Extensible Markup Language (XML), another solution to text-to-SVG synthesis involves language-based methods, which treat SVG scripts as text input. Recent works [35, 41] have proposed specialized tokenization strategies to embed SVG scripts and trained Sequence-To-Sequence (seq2seq) models on domain-specific datasets (e.g., icons and fonts). While these approaches achieve good generation quality within their training domains, they suffer from limited generalization due to the absence of large-scale, general-purpose text-SVG training data. Meanwhile, some works [26, 31] show that while state-of-the-art Large Language Models (LLMs) can generate basic geometric shapes (e.g., circles and rectangles) with layouts matching text prompts, they struggle to produce the complex geometric details required for professional SVG applications. Consequently, the limited generalization and poor expressiveness of language-based methods hinder their adoption for text-to-SVG generation.", + "bbox": [ + 91, + 352, + 483, + 625 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address the aforementioned problems, we propose a novel hybrid framework that leverages the complementary strengths of image-based and language-based methods. Our approach first utilizes an LLM to synthesize SVG templates composed of basic primitives and then optimizes their geometric details guided by image diffusion models. This hybrid framework addresses the key limitations of both paradigms: The LLM-based template generation overcomes the domain-specific generalization constraints of traditional language models while ensuring shape regularity, as each element naturally corresponds to a single semantic component. The subsequent image-based optimization then enhances the expressiveness of these well-structured templates by capturing complex geometric details that LLMs alone struggle to produce. Specifically, to fully exploit LLMs' ability to create reasonable and complex layouts that match text prompts, we design an SVG-oriented prompt with multiple stages, including prompt ex", + "bbox": [ + 89, + 628, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "pansion, SVG script generation, and visual refinement. To refine the generated SVG templates, we implement a dual-stage optimization strategy: (1) We first use an image diffusion model to synthesize detailed image-level targets, then follow [50] to optimize path-level latent vectors with a pretrained SVG VAE, eliminating common issues such as self-intersections and jagged curves; (2) We further refine geometric details by directly optimizing the point coordinates to capture fine-grained visual elements. Comprehensive experiments show that Chat2SVG consistently outperforms existing methods regarding overall visual quality, individual path regularity, and semantic coherence. Furthermore, our system enables intuitive editing through iterative natural language instructions, making vector graphics creation more accessible to non-expert users. Our contributions are summarized as follows:", + "bbox": [ + 511, + 90, + 903, + 330 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Hybrid SVG Generation Framework. We introduce a novel text-to-SVG generation paradigm that combines Large Language Models with image diffusion models to produce high-quality SVG outputs.", + "- SVG-Oriented Prompt Design. We develop a specialized prompt system that directs LLMs to generate SVG templates using basic geometric primitives.", + "- Dual-Stage Optimization. We implement a two-phase optimization process that preserves the semantic meaning of each shape while eliminating artifacts such as self-intersecting and jagged paths.", + "- Iterative Editing. We enable iterative refinement of SVG through natural language instructions, making SVG creation more accessible." + ], + "bbox": [ + 513, + 333, + 903, + 544 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 560, + 653, + 575 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Large Language Models for Design", + "text_level": 1, + "bbox": [ + 511, + 585, + 818, + 603 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "LLMs have emerged as powerful tools for graphics design tasks, demonstrating remarkable capabilities in understanding and reasoning about complex design specifications across diverse domains. Recent works have explored their applications in (1) visual layout and composition, including poster designs [5, 21, 34, 47], 3D scene arrangements [2, 10, 14, 32, 33, 38, 51], and placement of basic shapes for image generation [8, 46]; (2) content creation and manipulation for shapes [11], materials [15], and animation [22, 36]; and (3) design understanding [18, 42]. While these demonstrate LLMs' versatility, their potential for vector graphics remains unexplored. Our work addresses this gap by showing that LLMs can effectively generate structured SVG templates, extending their capabilities to the domain of vector graphics creation.", + "bbox": [ + 511, + 608, + 906, + 835 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Text-Guided Vector Graphics Generation", + "text_level": 1, + "bbox": [ + 511, + 847, + 867, + 863 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Text-to-SVG generation approaches can be roughly categorized into image-based and language-based methods.", + "bbox": [ + 511, + 869, + 903, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "23691", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Image-based methods [9, 16, 43, 44, 50] start with a large collection of randomly initialized shapes, render vector graphics with a differentiable renderer [20], and iteratively optimize path parameters by minimizing text-image similarity losses (e.g., CLIP [28] loss and Score Distillation Sampling [27] loss). Despite their visually pleasant appearance, these methods often generate SVG results containing multiple fragmented paths that lack individual semantic correspondence, which is undesirable for real design scenarios. Language-based methods [3, 35, 41] design specialized tokenization strategies to encode vector graphics into discrete tokens, which are then concatenated with text tokens, and train seq2seq models like autoregressive transformers [37] on domain-specific datasets (e.g., icons and fonts). Although these approaches are conceptually elegant, their generalizability is limited by the scarcity of large-scale, general-purpose vector graphics datasets. Some works [26, 31] have explored LLMs' vector graphics generation capabilities, showing that while LLMs can create reasonable layouts matching text prompts, they struggle to produce complex geometric shapes. Our Chat2SVG combines image-based and language-based methods to address the limitations of each paradigm, producing semantically meaningful SVG where each path corresponds to a distinct visual element and contains refined geometric details.", + "bbox": [ + 89, + 90, + 485, + 470 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Vector Graphics Representation Learning", + "text_level": 1, + "bbox": [ + 89, + 477, + 449, + 494 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Vector graphics representation learning is essential for sketch-based retrieval, reconstruction, and generation tasks. A seminal study, SketchRNN [12], combines a Recurrent Neural Network (RNN) and a VAE to learn vector representations. Lopes et al. [24] trained a VAE to represent image-level font styles, then passed the latent embedding into a decoder for vector font generation. To increase reconstruction quality, later approaches like Sketchformer [30] and DeepSVG [4] adopted Transformer [37] architectures, followed by dual-modality methods [23, 39, 40] that leveraged both vector and image features. However, encoding the entire complex SVG into a single latent embedding often causes detail loss. To address this limitation, Zhang et al. [50] proposed path-level representational learning, creating a smooth latent space for individual paths. We leverage these path-level latent embeddings in our optimization to eliminate self-intersecting and jagged paths.", + "bbox": [ + 89, + 500, + 485, + 758 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 768, + 183, + 786 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our Chat2SVG, as shown in Figure 2, begins with an SVG-oriented prompting approach (Section 3.1) that guides an LLM to create reasonable SVG templates. We then perform a dual-stage optimization (Section 3.2) to enhance the geometric details of SVG templates by optimizing SVG paths in both latent and control point space, guided by image diffusion models.", + "bbox": [ + 89, + 794, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. SVG Template Generation with LLMs", + "text_level": 1, + "bbox": [ + 513, + 90, + 844, + 107 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Prompt Expansion. Users often provide vague and brief prompts to the LLM when describing their desired graphics. Such unstructured input may lead to oversimplified SVG (regarding both the number of elements and structures) that fails to reflect the intended design. Consequently, we propose a three-level prompt expansion strategy that systematically delineates the initial prompts: (1) Scene-level: Starting with an initial prompt, we instruct the LLM to analyze it holistically and identify essential objects that should appear. The LLM then expands the prompt by suggesting complementary objects to enhance scene completeness. (2) Object-level: For each object in the expanded prompt, we guide the LLM to systematically break it down into its components. For instance, when describing a lion, the LLM deconstructs it into distinct parts such as the body, head, mane, legs, eyes, ears, and tail. This decomposition ensures that no critical components are missing during SVG script generation. (3) Layout-level: After object decomposition, we direct the LLM to develop a comprehensive layout plan. This includes determining the position and size of each element on the canvas, selecting appropriate colors for visual harmony, and establishing clear spatial relationships to ensure a cohesive overall composition.", + "bbox": [ + 511, + 112, + 906, + 460 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "SVG Script Generation. After prompt expansion, we obtain a detailed scene description. We then convert this natural language specification into SVG scripts in XML format. Since the LLM has limitations in synthesizing geometrically complex paths, we constrain the shapes to a carefully selected set of basic primitives: rectangles, ellipses, lines, polylines, polygons, and short paths. We set the canvas size to $512 \\times 512$ . Each geometric element is assigned a unique ID and includes semantic annotations in its comments.", + "bbox": [ + 511, + 460, + 908, + 595 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Visual Rectification. Since prompt expansion is purely text-based, even when the generated SVG script accurately follows the detailed prompt, visual inconsistencies (e.g., misaligned components, disproportionate scaling, and incorrect path ordering) can emerge during rendering. For example, in Figure 2, the initial SVG template shows a lion with a misshapen mane and missing facial features. We thus adopt a visual rectification strategy in which we render the SVG and provide the rendered image back to the LLM (with vision capability) for inspection. The LLM analyzes the visual output, identifies any inconsistencies or oddities, and generates corrected SVG code. This visual refinement loop can be performed iteratively. In our experiments, we find that two iterations of refinement are typically sufficient to generate well-structured SVG templates with appropriate spatial layouts.", + "bbox": [ + 511, + 598, + 908, + 839 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Furthermore, to enhance the quality of prompt expansion and SVG script generation, we provide curated in-context examples in the prompts. The complete set of prompts and examples is available in the supplementary material.", + "bbox": [ + 511, + 840, + 908, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "23692", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/fe924cf106c9beec49f8582b30b7f20d22c68f6d792b6640cf4a576f58ece790.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 163, + 89, + 834, + 193 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/f9134506dd5b3e81b330b3ffdd59872c440a6c8ebe9c1fed23ce7918399d1308.jpg", + "image_caption": [ + "Figure 2. The system pipeline of Chat2SVG. Given a text prompt, our system first leverages an LLM to generate an SVG template composed of basic geometric primitives. The rendered template is enhanced through SDEdit [25] with ControlNet [49] to add visual details while preserving the overall composition, yielding a target image. The SVG then undergoes a dual-stage optimization process to match the target image. (1) Primitives are converted to latent embeddings through latent inversion and optimized along with their visual attributes (i.e., filling colors $c_{i}$ , stroke properties $s_{i}$ , and transformation matrices $T_{i}$ ). (2) Point-level optimization is performed to refine the geometric details of SVG paths." + ], + "image_footnote": [], + "bbox": [ + 163, + 195, + 535, + 320 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/80317083313508ff8a164d592182c2dde4f4cbe8a5589373af247225e0bb9622.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 540, + 195, + 834, + 320 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. SVG Optimization Guided by Image Diffusion", + "text_level": 1, + "bbox": [ + 89, + 441, + 480, + 459 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Detail Enhancement. While the primitive shapes in the SVG template have accurate semantic meaning, they lack the geometric details necessary for a professional design. To improve visual expressiveness, we first render the SVG template into a raster image, and then use an image editing method to generate a more detailed version that serves as our optimization target. Specifically, we employ SDEdit [25], which enhances the input image by first adding noise and then progressively denoising it with an image diffusion model to produce outputs with richer details, such as the contour of the lion's mane illustrated in Figure 2. To maintain structural similarity between the enhanced output and the original image, we incorporate a ControlNet (tile version) [49] into our pipeline. By using Gaussian-blurred versions of the initial images as control signals, the ControlNet effectively maintains the overall composition throughout the enhancement process.", + "bbox": [ + 88, + 470, + 482, + 728 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Apart from geometric details, this image-to-image translation process also introduces new semantic parts such as the lion's beard. To faithfully reproduce the target image, we propose using the Segment Anything Model (SAM) [17] to identify these new decorative details. Specifically, we denote the rendered images of the initial SVG template and its diffusion-enhanced counterpart as $I_{\\mathrm{template}}$ and $I_{\\mathrm{target}}$ , respectively. We apply SAM to extract mask sets from both images: $\\{m_i\\}_{i=1}^Q$ for the template image and $\\{m_j\\}_{j=1}^K$ for the target image, where $Q$ and $K$ represent the number of masks in each collection. Our approach is based on the ob", + "bbox": [ + 89, + 734, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "servation that the diffusion process simultaneously generates fine decorative details that are absent in the original template (e.g., the beard of the lion in Figure 2) while preserving the structural integrity of larger components (such as the body). Consequently, we implement an area-based mask-filtering strategy to identify these newly generated elements. For each mask $m_j$ detected in $I_{\\mathrm{target}}$ , we evaluate it against all masks in $\\{m_i\\}_{i=1}^Q$ using Intersection over Union (IoU) metrics [29]. When $m_j$ overlaps with a template mask $m_i$ but their IoU falls below a threshold, we classify $m_j$ as a new decorative component. Alternatively, if their IoU exceeds the threshold, we interpret $m_j$ as a variation of an existing component. For all filtered masks classified as new decorative shapes, we approximate their boundaries with polygons and integrate them into the SVG script.", + "bbox": [ + 511, + 443, + 906, + 672 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We define an SVG script $G$ as a collection of $M$ paths (i.e., shapes), $G = \\{P_i\\}_{i=1}^M$ . Each path $P_i$ consists of a sequence of $N_i$ commands, $P_i = \\{C_i^j\\}_{j=1}^{N_i}$ . A command $C_i^j = (U_i^j, V_i^j)$ is defined by its type $U_i^j \\in \\{\\mathbb{M}, \\mathbb{C}\\}$ and its associated control points $V_i^j$ , where $\\mathbb{M}$ represents Move and $\\mathbb{C}$ represents cubic Bézier curves. For consistency, we convert all other primitive shapes (e.g., rectangles and ellipses) in the SVG template into cubic Bézier curves.", + "bbox": [ + 511, + 676, + 908, + 805 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Previous image-based methods [9, 16, 43, 44] optimize SVG shapes by directly manipulating the control points $V_{i}^{j}$ . Despite convenience, this approach often leads to artifacts such as self-intersecting curves and unnatural deformations. To overcome these limitations, Zhang et al. [50] introduced a path-level SVG VAE. This model's latent space effec", + "bbox": [ + 511, + 809, + 908, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "23693", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "tively captures common shape patterns and geometric constraints, allowing for the optimization of path latent vectors at a higher level to produce smooth outputs.", + "bbox": [ + 89, + 90, + 480, + 136 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Latent Optimization. We leverage this pretrained SVG VAE to conduct the latent optimization. However, their SVG encoder expects a fixed number of commands (10 cubic Bezier curves), while our SVG primitives contain varying (typically fewer) commands. To resolve this mismatch, we develop a latent inversion process that converts our primitives into latent embeddings. We start with a randomly sampled latent vector $z$ and decode it into a command sequence. Then we evenly sample points $X = \\{x_{i}\\}_{i=1}^{N}$ along its contour. Similarly, we sample an equal number of points $Y = \\{y_{j}\\}_{j=1}^{N}$ along the contour of our target primitive shape. Using these point sets, we compute the Earth Mover's Distance (EMD) [1, 7, 48]:", + "bbox": [ + 89, + 136, + 482, + 332 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\ell_ {\\mathrm {E M D}} (X, Y) = \\min _ {\\phi : X \\rightarrow Y} \\sum_ {x \\in X} \\| x - \\phi (x) \\| _ {2}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 143, + 339, + 482, + 369 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\phi$ represents a bijective mapping between the point sets. By minimizing this EMD loss, we gradually optimize the latent embedding $z$ to match the target primitive.", + "bbox": [ + 89, + 377, + 482, + 421 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After obtaining latent vectors $\\{z_i\\}_{i=1}^M$ for all paths, we jointly optimize these vectors along with visual attribute parameters: filling colors $\\{c_i\\}_{i=1}^M$ , stroke properties (color and width) $\\{s_i\\}_{i=1}^M$ , and transformation matrices $\\{T_i\\}_{i=1}^M$ . At each optimization iteration $t$ , we decode the latent vectors into command sequences and apply transformation matrices to the control points. We then combine color and stroke attributes together, and use a differentiable rasterizer [20] to render the complete SVG into an image $I_t$ . Our optimization objective consists of three loss terms: (1) An MSE loss $\\ell_{\\mathrm{MSE}}$ between the rendered image $I_t$ and target image $I_{\\mathrm{target}}$ to ensure visual similarity; (2) A curvature loss that reduces sharp bends and fluctuations along the contour by computing the discrete second derivative:", + "bbox": [ + 89, + 421, + 483, + 633 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\ell_ {\\text {c u r v a t u r e}} = \\frac {\\sum_ {i = 1} ^ {M} \\sum_ {j = 1} ^ {N _ {i} - 2} \\left\\| V _ {i} ^ {j} - 2 V _ {i} ^ {j + 1} + V _ {i} ^ {j + 2} \\right\\| _ {2} ^ {2}}{\\sum_ {i = 1} ^ {M} \\left(N _ {i} - 2\\right)}; \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 114, + 640, + 482, + 710 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "(3) A path-level IoU loss that preserves the overall composition and prevents paths from drifting too far from their initial positions:", + "bbox": [ + 89, + 715, + 482, + 762 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\ell_ {\\mathrm {I o U}} = \\frac {1}{M} \\sum_ {i = 1} ^ {M} \\left(1 - \\frac {\\left| m _ {i} ^ {t} \\cap m _ {i} ^ {0} \\right|}{\\left| m _ {i} ^ {t} \\cup m _ {i} ^ {0} \\right|}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 768, + 482, + 809 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $m_i^t$ and $m_i^0$ denote the binary masks of the $i$ -th path at the current iteration $t$ and its initial state, respectively.", + "bbox": [ + 89, + 816, + 482, + 847 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The final loss function combines these three terms with empirically determined weights:", + "bbox": [ + 89, + 847, + 482, + 877 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\ell_ {\\text {l a t e n t}} = \\ell_ {\\mathrm {M S E}} + \\lambda_ {1} \\ell_ {\\text {c u r v a t u r e}} + \\lambda_ {2} \\ell_ {\\mathrm {I o U}}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 150, + 885, + 482, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda_1 = 5e - 4$ and $\\lambda_{2} = 5e - 6$", + "bbox": [ + 511, + 90, + 763, + 104 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Point Optimization. While latent optimization effectively aligns SVG shapes with their target positions and contours, it has a notable limitation: the resulting shapes tend to be overly smooth, lacking the intricate details that are often essential in professional vector graphics. For example, in Figure 2, the lion's mane after latent optimization does not capture the fine contour present in the target image. Therefore, we add a second optimization stage that directly refines the control points $V_{i}^{j}$ of the paths, similar to previous image-based works [9, 16, 43, 44]. To achieve finer granularity in shape control, we split each cubic Bezier curve at its midpoint, effectively doubling the number of control points in the SVG. This increased control point density allows for more precise shape adjustments. The point optimization stage employs a loss function that combines the MSE loss and the curvature loss:", + "bbox": [ + 511, + 106, + 903, + 347 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\ell_ {\\text {p o i n t}} = \\ell_ {\\mathrm {M S E}} + \\lambda_ {3} \\ell_ {\\text {c u r v a t u r e}}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 607, + 361, + 903, + 378 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The parameter $\\lambda_3$ decreases linearly from $1e - 3$ to $5e - 5$ throughout the optimization process.", + "bbox": [ + 511, + 388, + 903, + 420 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Iterative Editing", + "text_level": 1, + "bbox": [ + 513, + 430, + 676, + 446 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "During SVG script generation (Section 3.1), we instruct the LLM to annotate the semantic label of each path. This design enables users to refine the initial SVG template using natural language, as the LLM can precisely locate and modify specific elements based on semantic understanding.", + "bbox": [ + 511, + 452, + 903, + 527 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our system supports iterative refinement through multiple rounds of natural language instructions. However, when passing the edited SVG template image into SDEdit, image diffusion models may introduce different shape variations and decorative details, which compromise the user's intention to maintain the already optimized shapes. To preserve consistency across editing iterations, we instruct the LLM to output both the modified SVG script and a list of specifically changed paths. This precise tracking enables selective optimization of modified shapes while preserving unaltered elements from previous iterations, maintaining visual coherence throughout the entire refinement process.", + "bbox": [ + 511, + 527, + 905, + 709 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 513, + 724, + 645, + 739 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details. In our experiments, we use claude-3-5-sonnet as our backbone LLM model due to its leading generation capabilities. To evaluate our approach, we create a prompt set by having the LLM generate 125 text prompts across five categories: animals, food, objects, scenes, and novel concept combinations. For each prompt, we generate one template and perform 2 rounds of visual rectification. This process is repeated 5 times, yielding $(1 + 2) \\times 5 = 15$ candidate SVG templates per prompt. These templates are rendered as images, and we follow a", + "bbox": [ + 511, + 750, + 906, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "23694", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/cd01b7e83cd888ef4faaa359d8e3600ab0db1b8221bee6904932994e764ff4a1.jpg", + "image_caption": [ + "Figure 3. Qualitative Comparison. (1) Methods refining open-ended strokes, i.e., CLIPDraw [9] and DiffSketcher [43], often produce distorted and disorganized strokes to approximate objects, presenting messy appearance and poor text alignment. (2) VectorFusion [16] and SVGDreamer [44] produce elements that consist of multiple jagged, irregular, and fragmented shapes, such as the body of the flamingo (first row) and the plane (third row). (3) T2V-NPR [50] attempts to resolve these issues by learning a latent representation of paths and merging fragmented shapes. However, it still cannot guarantee the semantic meanings of the paths, leading to less-semantic paths such as a plane body with surrounding clouds in the third row. In contrast, our method produces SVG with superior text alignment, higher visual quality, and well-structured paths exhibiting geometric regularity and clear semantic definition." + ], + "image_footnote": [], + "bbox": [ + 133, + 85, + 867, + 556 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "standard practice to select the highest quality SVG using the ImageReward [45] metric. Detailed optimization configurations are provided in the supplementary materials.", + "bbox": [ + 89, + 689, + 480, + 736 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines. We compare our approach against five state-of-the-art image-based text-to-SVG generation methods: CLIPDraw [9], DiffSketcher [43], VectorFusion [16], SVGDreamer [44], and T2V-NPR [50]. Current language-based methods [35, 41] are confined to specific categories, with no support for general text-to-SVG generation, so we do not include them in our comparison. For a fair comparison, we set these baseline methods to use a similar number of shapes as our optimized SVG.", + "bbox": [ + 89, + 741, + 482, + 878 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metrics. We evaluate the quality of our gen", + "bbox": [ + 89, + 885, + 482, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "erated SVG across three dimensions: visual fidelity, path regularity, and semantic alignment.", + "bbox": [ + 511, + 689, + 903, + 719 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Image-level Fidelity. We evaluate visual fidelity using a ground-truth dataset of well-designed vector graphics, specifically 52,805 colored SVG files downloaded from SVGRepo1. We compute the Fréchet Inception Distance (FID) [13] between the rendered images of our generated SVG and this professional design collection, using features extracted by the CLIP image encoder [28].", + "- Path-level Regularity. We assess path quality using a transformer-based Auto-Encoder trained on the FIGR-8-SVG [6] dataset, following the DeepSVG [4] architecture" + ], + "bbox": [ + 511, + 720, + 903, + 872 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "1https://www.svgrepo.com", + "bbox": [ + 531, + 887, + 723, + 898 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "23695", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/e71a77f3e2c2286f20798e692568ae2886876fd20455a53b0a16ce4e2607f445.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodImage FID ↓Path FID ↓Text Alignment ↑
CLIPDraw [9]46.7770.130.3048
DiffSketcher [43]44.8966.480.2623
VectorFusion [16]39.5256.790.2982
SVGDreamer [44]35.4847.950.2919
T2V-NPR [50]39.8642.030.3078
Chat2SVG (Ours)33.3139.070.3090
", + "bbox": [ + 122, + 88, + 452, + 200 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 1. Quantitative comparison of text-to-SVG generation methods across image fidelity, vector regularity, and text alignment.", + "bbox": [ + 89, + 209, + 482, + 239 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ture. Our model, trained with reconstruction tasks, encodes each drawing command $C_i^j$ into a latent vector. We represent each path as the mean of its commands' latent embeddings, and calculate the FID between these path representations and the ground truth paths from FIGR-8-SVG. This metric quantifies how closely our generated paths align with professional SVG design patterns.", + "bbox": [ + 102, + 265, + 482, + 369 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "- Text-level Alignment. We assess semantic alignment by computing the CLIP score [28] between the text prompt and the rendered SVG image.", + "bbox": [ + 89, + 371, + 483, + 416 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.1. Comparison with Existing Methods", + "text_level": 1, + "bbox": [ + 89, + 425, + 400, + 441 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Quantitative Comparison. Table 1 shows the evaluation metrics for all baseline methods. Our method achieves the best image FID score, indicating that our generated SVG closely aligns with professional design patterns. Regarding path regularity, our method yields paths that are closest to the professionally designed SVG, as evidenced by the lowest path FID. Furthermore, our Chat2SVG achieves the highest score in text-SVG alignment, validating the significance of LLM-generated SVG templates.", + "bbox": [ + 88, + 446, + 482, + 582 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Qualitative Comparison. In Figure 3, we present a side-by-side comparison between our Chat2SVG and the baselines. Methods that optimize open-ended strokes based on CLIP or Diffusion models (i.e., CLIPDraw [9] and DiffS-ketcher [43]) produce results with messy visual appearances and poor text alignment. These methods sometimes fail to synthesize complete objects, such as the flamingo (first row) and the plane (third row) in Figure 3. In contrast, our method generates SVG results with clear layouts and greater fidelity to the prompts.", + "bbox": [ + 88, + 584, + 482, + 733 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Methods that optimize closed shapes via score distillation of diffusion models (i.e., VectorFusion [16], SVGDreamer [44], and T2V-NPR [50]) can better align with the input text, but still produce fragmented paths with limited semantic meaning. For instance, in the plane example (third row of Figure 3), VectorFusion and SVGDreamer create planes and clouds composed of multiple jagged, irregular, and fragmented shapes, which only appear meaningful when viewed as a whole. This hinders convenient editing by graphic designers. T2V-NPR [50] addresses the issue of jagged paths using a path VAE (although it can be", + "bbox": [ + 88, + 734, + 482, + 901 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/31a541393ca3e17bc2f81e2e8155d6fed320563e6c26162a7339b00768ddd09d.jpg", + "image_caption": [ + "Figure 4. User Study. Our Chat2SVG achieves the highest user selection ratio across all three evaluation criteria." + ], + "image_footnote": [], + "bbox": [ + 553, + 87, + 867, + 265 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "overly smooth at times) and reduces fragmented shapes by merging shapes with similar colors. However, its merging operation ignores the semantic meaning of shapes, resulting in the body of the plane being merged with the surrounding clouds. Additionally, the semantic meaning of the path is sometimes unclear, as exemplified by the cloud in the bottom-right corner, whose contour is actually approximated by the blue background instead of a true cloud shape. In contrast, our Chat2SVG adopts a unique approach by using LLMs to generate paths representing semantic components and performing dual-stage optimization based on image diffusion to enhance path expressiveness, thus ensuring both path regularity and visual fidelity of generated SVG.", + "bbox": [ + 511, + 333, + 906, + 530 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "User Study. To further compare Chat2SVG with baseline methods, we conducted a user study examining the same key aspects: (1) overall visual aesthetics, (2) path regularity, and (3) text-SVG alignment. We randomly sampled 20 prompts and generated SVG outputs using our approach and five baseline methods. For each prompt, participants were shown all generated SVG results along with the corresponding text prompt and highlighted paths. They were then asked to select the highest quality result regarding the three evaluation criteria. We recruited 31 participants (18 male, 13 female) through university mailing lists, with a mean age of 26 years. Among them, 17 participants reported prior experience in graphic design, providing a balanced mix of expert and novice perspectives. Analyzing the average selection ratio across all three metrics (Figure 4), we found that Chat2SVG-generated results were consistently preferred by participants over baseline methods.", + "bbox": [ + 511, + 531, + 908, + 787 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Ablation Study", + "text_level": 1, + "bbox": [ + 511, + 801, + 666, + 816 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We conduct ablation experiments to evaluate the effectiveness of key components in our pipeline (Section 3.1). First, we remove the SVG template generation and use randomly initialized shapes to approximate target images. As shown in the second row of Table 2, both image FID and path FID", + "bbox": [ + 511, + 825, + 905, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "23696", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/13f07943d5304d0b1bee7c5cdd542cec907fd793ef28f52a277d146cb6272fc6.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Chat2SVG VariantImage FID ↓Path FID ↓Text Alignment ↑
Default33.3139.070.3090
No SVG Template47.2170.130.2968
No SVG Optimization33.4536.120.3044
", + "bbox": [ + 112, + 88, + 460, + 162 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 2. Qualitative results of the ablation study.", + "bbox": [ + 140, + 172, + 431, + 186 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/88128b6fc59cd146790d4597fa859c088b39f479d7e423e29cb6749cd828d3b9.jpg", + "image_caption": [ + "Figure 5. Qualitative results of ablation study." + ], + "image_footnote": [], + "bbox": [ + 94, + 200, + 483, + 308 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "values increase substantially. This performance degradation can be attributed to the substantial gap between randomly initialized shapes and target shapes, making optimization more challenging. Meanwhile, in the second column of Figure 5, the dog's legs and ears lack semantic-clear components, highlighting the importance of SVG template generation. Second, we eliminate the dual-stage optimization. As shown in the third row of Table 2, this results in the lowest path FID, as primitives generated by LLMs naturally align with the path regularity presented in the ground-truth dataset. However, the visual outcome in Figure 5 lacks the necessary details, indicating that the overall visual quality suffers without dual-stage optimization. A more detailed ablation study is provided in the supplementary material.", + "bbox": [ + 89, + 359, + 483, + 571 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Iterative Editing Results", + "text_level": 1, + "bbox": [ + 89, + 580, + 313, + 597 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this section, we present the iterative editing results. As demonstrated in Figure 6, users can progressively refine the SVG through multiple rounds of natural language instructions. During editing, we preserve the optimized shapes from previous iterations while only modifying the specified shapes. The results showcase how Chat2SVG can accurately interpret and execute user editing requests, maintaining both semantic coherence and visual quality throughout the modifications. Additional editing examples are provided in the supplementary material.", + "bbox": [ + 89, + 603, + 483, + 755 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 89, + 768, + 209, + 784 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we present Chat2SVG, a novel paradigm for text-to-SVG generation that combines LLMs and image diffusion models. Through our carefully designed SVG template generation and dual-stage SVG optimization pipeline, our method generates high-quality SVG outputs that exhibit strong visual fidelity, path regularity, and text alignment.", + "bbox": [ + 89, + 794, + 483, + 883 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Despite our method's superior results, there are several", + "bbox": [ + 109, + 885, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/62cd837d2b193eecf7c28ab864031b4f738964edd919c617dec2e4e805051e96.jpg", + "image_caption": [ + "Figure 6. Iterative Editing. We perform two rounds of refinement on each SVG template and show the optimized output. This figure shows editing types including deletion, modification, and addition." + ], + "image_footnote": [], + "bbox": [ + 535, + 88, + 883, + 345 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "(1) Layout Precision", + "bbox": [ + 535, + 414, + 635, + 425 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f7ea2084ffb93095b4075d8ab0d9fb1bd80874dbcee42fe466a7dcc3b0f709e9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 535, + 426, + 633, + 465 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/20e7459baa524fb618e1a24a76b23f9ed3bf25bd6115da62d5709794501aae16.jpg", + "image_caption": [ + "SVG Target Template Image", + "Optimized SVG" + ], + "image_footnote": [], + "bbox": [ + 557, + 489, + 609, + 534 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "(2) Path Addition", + "bbox": [ + 651, + 414, + 738, + 425 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/18979784f96b2fb99b328c336ee0d238a340c6532366dc223f1e496b0d2de66a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 640, + 426, + 746, + 468 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d6b078a9667f664bd408f16b9507351cbb3729016cd93fc7b282da93fd66253f.jpg", + "image_caption": [ + "SVG Target Template Image", + "Optimized SVG" + ], + "image_footnote": [], + "bbox": [ + 666, + 489, + 725, + 534 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "(3) Shape Complexity", + "bbox": [ + 764, + 414, + 870, + 425 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/fb5f5c36a016363d6cf9d58799a2c8df46c0fb0803a737a7cf4920860c2da95a.jpg", + "image_caption": [ + "SVG Template" + ], + "image_footnote": [], + "bbox": [ + 759, + 426, + 823, + 469 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/635ee330f3881215cb11e15b074e60f1cca7990a71f9773ccb73424e5c8ba303.jpg", + "image_caption": [ + "Target Image" + ], + "image_footnote": [], + "bbox": [ + 823, + 426, + 883, + 469 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8a028051f52b1a74157cfd3cf71eab3726afcc148799ce13b06765b8ff914ea6.jpg", + "image_caption": [ + "Optimized SVG", + "Figure 7. Limitations of our method include imprecise layout generation, missing important visual elements, and insufficient shape complexity." + ], + "image_footnote": [], + "bbox": [ + 764, + 489, + 828, + 535 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "aspects that could be further improved: (1) LLMs may produce imprecise visual layouts, such as the burger in Figure 7. This can be alleviated by collecting an SVG dataset and extracting the corresponding layouts to fine-tune LLMs. (2) Our approach uses SAM to identify new parts. However, SAM may overlook important regions, like the pink frosting on the donut in Figure 7. Incorporating semantic-aware SAM [19] could potentially produce more accurate results. (3) The pretrained SVG VAE from [50] has a fixed number of commands. This makes it hard to represent intricate shapes. For example, the lion's mane in Figure 7 lacks sufficient detail. Training a new SVG VAE with varying-length commands would help solve this problem. Our work can be extended to generate other types of vectorized data, such as icons and fonts. Additionally, through consistent editing of SVG across multiple frames, coupled with shape interpolation, our Chat2SVG has the potential to facilitate keyframe animation generation.", + "bbox": [ + 511, + 628, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "23697", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 91, + 90, + 250, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 196, + 187, + 213 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Panos Achlioptas, Olga Diamanti, Ioannis Mitliagkas, and Leonidas Guibas. Learning representations and generative models for 3D point clouds. In International Conference on Machine Learning, pages 40-49, 2018. 5", + "[2] Rio Aguina-Kang, Maxim Gumin, Do Heon Han, Stewart Morris, Seung-Jean Yoo, Aditya Ganeshan, R. Kenny Jones, Qiuhong Anna Wei, Kailiang Fu, and Daniel Ritchie. Open-universe indoor scene generation using LLM program synthesis and uncurated object databases. In arXiv preprint arXiv:2403.09675, 2024. 2", + "[3] Jonas Belouadi, Anne Lauscher, and Steffen Eger. AutomataTikZ: Text-guided synthesis of scientific vector graphics with TikZ. In arXiv preprint arXiv:2310.00367, 2023. 3", + "[4] Alexandre Carlier, Martin Danelljan, Alexandre Alahi, and Radu Timofte. DeepSVG: A hierarchical generative network for vector graphics animation. In Advances in Neural Information Processing Systems, pages 16351-16361, 2020. 3, 6", + "[5] Yutao Cheng, Zhao Zhang, Maoke Yang, Hui Nie, Chunyuan Li, Xinglong Wu, and Jie Shao. Graphic design with large multimodal model. In arXiv preprint arXiv:2404.14368, 2024. 2", + "[6] Louis Clouatre and Marc Demers. FIgR: Few-shot image generation with reptile. In arXiv preprint arXiv:1901.02199, 2019. 6", + "[7] Haoqiang Fan, Hao Su, and Leonidas Guibas. A point set generation network for 3D object reconstruction from a single image. In IEEE Conference on Computer Vision and Pattern Recognition, pages 605-613, 2017. 5", + "[8] Weixi Feng, Wanrong Zhu, Tsu-jui Fu, Varun Jampani, Arjun Akula, Xuehai He, Sugato Basu, Xin Wang, and William Wang. LayoutGPT: Compositional visual planning and generation with large language models. In Advances in Neural Information Processing Systems, pages 18225-18250, 2024. 2", + "[9] Kevin Frans, Lisa B. Soros, and Olaf Witkowski. CLIPDraw: Exploring text-to-drawing synthesis through language-image encoders. In Advances in Neural Information Processing Systems, pages 5207-5218, 2021. 2, 3, 4, 5, 6, 7", + "[10] Rao Fu, Zehao Wen, Zichen Liu, and Srinath Sridhar. AnyHome: Open-vocabulary generation of structured and textured 3D homes. In European Conference on Computer Vision, pages 52-70, 2025. 2", + "[11] Aditya Ganeshan, Ryan Y. Huang, Xianghao Xu, R. Kenny Jones, and Daniel Ritchie. ParSEL: Parameterized shape editing with language. In arXiv preprint arXiv:2405.20319, 2024. 2" + ], + "bbox": [ + 93, + 223, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "The work described in this paper was fully supported by a GRF grant from the Research Grants Council (RGC) of the Hong Kong Special Administrative Region, China [Project No. CityU 11216122].", + "[12] David Ha and Douglas Eck. A neural representation of sketch drawings. In arXiv preprint arXiv:1704.03477, 2017. 3", + "[13] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. GANs trained by a two time-scale update rule converge to a local Nash equilibrium. In Advances in Neural Information Processing Systems, pages 6629-6640, 2017. 6", + "[14] Ziniu Hu, Ahmet Iscen, Aashi Jain, Thomas Kipf, Yisong Yue, David A. Ross, Cordelia Schmid, and Alireza Fathi. SceneCraft: An LLM agent for synthesizing 3D scenes as Blender code. In International Conference on Machine Learning, pages 19252-19282, 2024. 2", + "[15] Ian Huang, Guandao Yang, and Leonidas Guibas. Blender: Editing 3D graphics with vision-language models. In arXiv preprint arXiv:2404.17672, 2024. 2", + "[16] Ajay Jain, Amber Xie, and Pieter Abbeel. VectorFusion: Text-to-SVG by abstracting pixel-based diffusion models. In IEEE Conference on Computer Vision and Pattern Recognition, pages 1911-1920, 2022. 2, 3, 4, 5, 6, 7", + "[17] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dólar, and Ross Girshick. Segment anything. In IEEE International Conference on Computer Vision, pages 4015-4026, 2023. 4", + "[18] Peter Kulits, Haiwen Feng, Weiyang Liu, Victoria Abrevaya, and Michael J. Black. Re-thinking inverse graphics with large language models. In arXiv preprint arXiv:2404.15228, 2024. 2", + "[19] Feng Li, Hao Zhang, Peize Sun, Xueyan Zou, Shilong Liu, Jianwei Yang, Chunyuan Li, Lei Zhang, and Jianfeng Gao. Semantic-SAM: Segment and recognize anything at any granularity. In arXiv preprint arXiv:2307.04767, 2023. 8", + "[20] Tzu-Mao Li, Michal Lukáč, Michael Gharbi, and RaganKelley Jonathan. Differentiable vector graphics rasterization for editing and learning. ACM Transactions on Graphics, 39 (6):1-15, 2020. 2, 3, 5", + "[21] Jiawei Lin, Jiaqi Guo, Shizhao Sun, Zijiang Yang, JianGuang Lou, and Dongmei Zhang. LayoutPrompter: Awaken the design ability of large language models. In Advances in Neural Information Processing Systems, pages 43852-43879, 2024. 2", + "[22] Vivian Liu, Rubaiat H. Kazi, Li-Yi Wei, Matthew Fisher, Timothy Langlois, Seth Walker, and Lydia Chilton. LogoMotion: Visually grounded code generation for content-aware animation. In arXiv preprint arXiv:2405.07065, 2024. 2", + "[23] YingTian Liu, Zhifei Zhang, YuanChen Guo, Matthew Fisher, Zhaowen Wang, and SongHai Zhang. DualVector: Unsupervised vector font synthesis with dual-part representation. In IEEE Conference on Computer Vision and Pattern Recognition, pages 14193-14202, 2023. 3", + "[24] Raphael G. Lopes, David Ha, Douglas Eck, and Jonathon Shlens. A learned representation for scalable vector graphics. In IEEE International Conference on Computer Vision, pages 7930-7939, 2019. 3" + ], + "bbox": [ + 93, + 92, + 905, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "23698", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[25] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia-jun Wu, JunYan Zhu, and Stefano Ermon. SDEdit: Guided image synthesis and editing with stochastic differential equations. In arXiv preprint arXiv:2108.01073, 2021. 4", + "[26] OpenAI. GPT-4 technical report. In arXiv preprint arXiv:2303.08774, 2023. 2, 3", + "[27] Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. DreamFusion: Text-to-3D using 2D diffusion. In arXiv preprint arXiv:2209.14988, 2022. 2, 3", + "[28] Alec Radford, Jong-Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763, 2021. 2, 3, 6, 7", + "[29] Hamid Rezatofighi, Nathan Tsoi, JunYoung Gwak, Amir Sadeghian, Ian Reid, and Silvio Savarese. Generalized intersection over union: A metric and a loss for bounding box regression. In IEEE Conference on Computer Vision and Pattern Recognition, pages 658-666, 2019. 4", + "[30] Leo S. F. Ribeiro, Tu Bui, John Collomosse, and Moacir Ponti. Sketchformer: Transformer-based representation for sketched structure. In IEEE Conference on Computer Vision and Pattern Recognition, pages 14153-14162, 2020. 3", + "[31] Pratyusha Sharma, Tamar Rott Shaham, Manel Baradad, Stephanie Fu, Adrian Rodriguez-Munoz, Shivam Duggal, Phillip Isola, and Antonio Torralba. A vision check-up for language models. In IEEE Conference on Computer Vision and Pattern Recognition, pages 14410-14419, 2024. 2, 3", + "[32] Chunyi Sun, Junlin Han, Weijian Deng, Xinlong Wang, Zishan Qin, and Stephen Gould. 3D-GPT: Procedural 3D modeling with large language models. In arXiv preprint arXiv:2310.12945, 2023. 2", + "[33] Hou In Ivan Tam, Hou In Derek Pun, Austin T. Wang, Angel X. Chang, and Manolis Savva. SceneMotifCoder: Example-driven visual program learning for generating 3D object arrangements. In arXiv preprint arXiv:2408.02211, 2024. 2", + "[34] Zecheng Tang, Chenfei Wu, Juntao Li, and Nan Duan. LayoutNUWA: Revealing the hidden layout expertise of large language models. In arXiv preprint arXiv:2309.09506, 2023. 2", + "[35] Zecheng Tang, Chenfei Wu, Zekai Zhang, Mingheng Ni, Shengming Yin, Yu Liu, Zhengyuan Yang, Lijuan Wang, Zicheng Liu, Juntao Li, and Nan Duan. StrokeNUWA: Tokenizing strokes for vector graphic synthesis. In arXiv preprint arXiv:2401.17093, 2024. 2, 3, 6", + "[36] Tiffany Tseng, Ruijia Cheng, and Jeffrey Nichols. Keyframer: Empowering animation design using large language models. In arXiv preprint arXiv:2402.06071, 2024. 2", + "[37] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems, pages 6000–6010, 2017. 3" + ], + "bbox": [ + 91, + 90, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[38] Can Wang, Hongliang Zhong, Menglei Chai, Mingming He, Dongdong Chen, and Jing Liao. Chat2Layout: Interactive 3D furniture layout with a multimodal LLM. In arXiv preprint arXiv:2407.21333, 2024. 2", + "[39] Yizhi Wang and Zhouhui Lian. DeepVecFont: Synthesizing high-quality vector fonts via dual-modality learning. ACM Transactions on Graphics, 40(6):1-15, 2021. 3", + "[40] Yuqing Wang, Yizhi Wang, Longhui Yu, Yuesheng Zhu, and Zhouhui Lian. DeepVecFont-v2: Exploiting Transformers to synthesize vector fonts with higher quality. In IEEE Conference on Computer Vision and Pattern Recognition, pages 18320-18328, 2023. 3", + "[41] Ronghuan Wu, Wanchao Su, Kede Ma, and Jing Liao. IconShop: Text-guided vector icon synthesis with autoregressive Transformers. ACM Transactions on Graphics, 42(6):1-14, 2023. 2, 3, 6", + "[42] Tong Wu, Guandao Yang, Zhibing Li, Kai Zhang, Ziwei Liu, Leonidas Guibas, Dahua Lin, and Gordon Wetzstein. GPT-4V (ision) is a human-aligned evaluator for text-to-3D generation. In IEEE Conference on Computer Vision and Pattern Recognition, pages 22227–22238, 2024. 2", + "[43] Ximing Xing, Chuang Wang, Haitao Zhou, Jing Zhang, Qian Yu, and Dong Xu. DiffSketcher: Text guided vector sketch synthesis through latent diffusion models. In Advances in Neural Information Processing Systems, pages 15869-15889, 2023. 2, 3, 4, 5, 6, 7", + "[44] Ximing Xing, Haitao Zhou, Chuang Wang, Jing Zhang, Dong Xu, and Qian Yu. SVGDreamer: Text guided SVG generation with diffusion model. In IEEE Conference on Computer Vision and Pattern Recognition, pages 4546-4555, 2024. 2, 3, 4, 5, 6, 7", + "[45] Jiazheng Xu, Xiao Liu, Yuchen Wu, Yuxuan Tong, Qinkai Li, Ming Ding, Jie Tang, and Yuxiao Dong. ImageReward: Learning and evaluating human preferences for text-to-image generation. In Advances in Neural Information Processing Systems, pages 15903-15935, 2024. 6", + "[46] Yutaro Yamada, Khyathi Chandu, Yuchen Lin, Jack Hessel, Ilker Yildirim, and Yejin Choi. L3GO: Language agents with chain-of-3D-thoughts for generating unconventional objects. In arXiv preprint arXiv:2402.09052, 2024. 2", + "[47] Tao Yang, Yingmin Luo, Zhongang Qi, Yang Wu, Ying Shan, and Chang Wen Chen. PosterLLaVa: Constructing a unified multi-modal layout generator with LLM. In arXiv preprint arXiv:2406.02884, 2024. 2", + "[48] Lequan Yu, Xianzhi Li, Chi-Wing Fu, Daniel Cohen-Or, and Pheng-Ann Heng. PU-Net: Point cloud upsampling network. In IEEE Conference on Computer Vision and Pattern Recognition, pages 2790–2799, 2018. 5", + "[49] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In IEEE International Conference on Computer Vision, pages 3836-3847, 2023. 4", + "[50] Peiying Zhang, Nanxuan Zhao, and Jing Liao. Text-to-vector generation with neural path representation. ACM Transactions on Graphics, 43(4):1-13, 2024. 2, 3, 4, 6, 7, 8", + "[51] Qihang Zhang, Chaoyang Wang, Aliaksandr Siarohin, Peiye Zhuang, Yinghao Xu, Ceyuan Yang, Dahua Lin, Bolei Zhou," + ], + "bbox": [ + 516, + 90, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "23699", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Sergey Tulyakov, and Hsin-Ying Lee. SceneWiz3D: Towards text-guided 3D scene composition. In arXiv preprint arXiv:2312.08885, 2023. 2", + "bbox": [ + 122, + 90, + 485, + 136 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "23700", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Chat2SVG_ Vector Graphics Generation with Large Language Models and Image Diffusion Models/03e11a55-7ca3-4174-a887-3307219befcc_model.json b/2025/Chat2SVG_ Vector Graphics Generation with Large Language Models and Image Diffusion Models/03e11a55-7ca3-4174-a887-3307219befcc_model.json new file mode 100644 index 0000000000000000000000000000000000000000..a469733fe65b439c6db685ed09d05ad437aaf2c9 --- /dev/null +++ b/2025/Chat2SVG_ Vector Graphics Generation with Large Language Models and Image Diffusion Models/03e11a55-7ca3-4174-a887-3307219befcc_model.json @@ -0,0 +1,2158 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.123, + 0.13, + 0.877, + 0.177 + ], + "angle": 0, + "content": "Chat2SVG: Vector Graphics Generation with Large Language Models and Image Diffusion Models" + }, + { + "type": "image_caption", + "bbox": [ + 0.123, + 0.204, + 0.368, + 0.258 + ], + "angle": 0, + "content": "Ronghuan Wu \nCity University of Hong Kong \nrh.wu@my.cityu.edu.hk" + }, + { + "type": "image_caption", + "bbox": [ + 0.404, + 0.204, + 0.594, + 0.257 + ], + "angle": 0, + "content": "Wanchao Su Monash University wanchao.su@monash.edu" + }, + { + "type": "image_caption", + "bbox": [ + 0.631, + 0.204, + 0.877, + 0.258 + ], + "angle": 0, + "content": "Jing Liao* \nCity University of Hong Kong \njingliao@cityu.edu.hk" + }, + { + "type": "image", + "bbox": [ + 0.124, + 0.263, + 0.885, + 0.525 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.113, + 0.534, + 0.883, + 0.551 + ], + "angle": 0, + "content": "Figure 1. SVG examples generated by our Chat2SVG. We highlight some shapes to demonstrate semantic clarity and path quality." + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.554, + 0.327, + 0.57 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.587, + 0.484, + 0.875 + ], + "angle": 0, + "content": "Scalable Vector Graphics (SVG) has become the de facto standard for vector graphics in digital design, offering resolution independence and precise control over individual elements. Despite their advantages, creating high-quality SVG content remains challenging, as it demands technical expertise with professional editing software and a considerable time investment to craft complex shapes. Recent text-to-SVG generation methods aim to make vector graphics creation more accessible, but they still encounter limitations in shape regularity, generalization ability, and expressiveness. To address these challenges, we introduce Chat2SVG, a hybrid framework that combines the strengths of Large Language Models (LLMs) and image diffusion models for text-to-SVG generation. Our approach first uses an LLM to generate semantically meaningful SVG templates from basic geometric primitives. Guided by image diffusion models, a dual-stage optimization pipeline refines paths in latent space and adjusts point coordinates to enhance geometric complexity. Extensive experiments show" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.556, + 0.907, + 0.647 + ], + "angle": 0, + "content": "that Chat2SVG outperforms existing methods in visual fidelity, path regularity, and semantic alignment. Additionally, our system enables intuitive editing through natural language instructions, making professional vector graphics creation accessible to all users. Our code is available at https://chat2svg.github.io/." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.678, + 0.645, + 0.693 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.704, + 0.907, + 0.885 + ], + "angle": 0, + "content": "Scalable Vector Graphics (SVG), a vector image format based on geometric shapes, has become the standard for modern digital design, offering resolution independence and precise control over individual elements. However, creating high-quality SVG content remains challenging for non-expert users, as it requires both expertise with professional design software and considerable time investment to create complex shapes. To make vector graphics creation more accessible, recent research has focused on developing text-to-SVG systems that enable users to express their creative ideas through simple text prompts rather than complex manual editing." + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.886, + 0.905, + 0.902 + ], + "angle": 0, + "content": "Existing approaches have explored image-based meth" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.241, + 0.901 + ], + "angle": 0, + "content": "*Corresponding Author." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "23690" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.349 + ], + "angle": 0, + "content": "ods for text-to-SVG generation, iteratively optimizing a large collection of shape elements (e.g., cubic Bezier curves, typically 100 to 1000) by rendering them into images with differentiable rasterizers [20] and evaluating them using text-image similarity metrics like CLIP loss [28] and Score Distillation Sampling (SDS) loss [27]. While these image-based methods [9, 16, 43, 44, 50] can generate visually impressive SVG through the combination of numerous paths and powerful image models, they face a critical limitation in maintaining the regularity and semantics of shapes. Specifically, semantic components that should be represented as single elements often end up fragmented across multiple overlapping paths. Although these fragmented paths, when viewed collectively, can yield visually appealing outputs, they fundamentally conflict with professional design principles, where each semantic component is intentionally crafted as a single, regularized path." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.353, + 0.485, + 0.626 + ], + "angle": 0, + "content": "Given SVG is defined using Extensible Markup Language (XML), another solution to text-to-SVG synthesis involves language-based methods, which treat SVG scripts as text input. Recent works [35, 41] have proposed specialized tokenization strategies to embed SVG scripts and trained Sequence-To-Sequence (seq2seq) models on domain-specific datasets (e.g., icons and fonts). While these approaches achieve good generation quality within their training domains, they suffer from limited generalization due to the absence of large-scale, general-purpose text-SVG training data. Meanwhile, some works [26, 31] show that while state-of-the-art Large Language Models (LLMs) can generate basic geometric shapes (e.g., circles and rectangles) with layouts matching text prompts, they struggle to produce the complex geometric details required for professional SVG applications. Consequently, the limited generalization and poor expressiveness of language-based methods hinder their adoption for text-to-SVG generation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.629, + 0.483, + 0.902 + ], + "angle": 0, + "content": "To address the aforementioned problems, we propose a novel hybrid framework that leverages the complementary strengths of image-based and language-based methods. Our approach first utilizes an LLM to synthesize SVG templates composed of basic primitives and then optimizes their geometric details guided by image diffusion models. This hybrid framework addresses the key limitations of both paradigms: The LLM-based template generation overcomes the domain-specific generalization constraints of traditional language models while ensuring shape regularity, as each element naturally corresponds to a single semantic component. The subsequent image-based optimization then enhances the expressiveness of these well-structured templates by capturing complex geometric details that LLMs alone struggle to produce. Specifically, to fully exploit LLMs' ability to create reasonable and complex layouts that match text prompts, we design an SVG-oriented prompt with multiple stages, including prompt ex" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.332 + ], + "angle": 0, + "content": "pansion, SVG script generation, and visual refinement. To refine the generated SVG templates, we implement a dual-stage optimization strategy: (1) We first use an image diffusion model to synthesize detailed image-level targets, then follow [50] to optimize path-level latent vectors with a pretrained SVG VAE, eliminating common issues such as self-intersections and jagged curves; (2) We further refine geometric details by directly optimizing the point coordinates to capture fine-grained visual elements. Comprehensive experiments show that Chat2SVG consistently outperforms existing methods regarding overall visual quality, individual path regularity, and semantic coherence. Furthermore, our system enables intuitive editing through iterative natural language instructions, making vector graphics creation more accessible to non-expert users. Our contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.334, + 0.905, + 0.395 + ], + "angle": 0, + "content": "- Hybrid SVG Generation Framework. We introduce a novel text-to-SVG generation paradigm that combines Large Language Models with image diffusion models to produce high-quality SVG outputs." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.395, + 0.905, + 0.439 + ], + "angle": 0, + "content": "- SVG-Oriented Prompt Design. We develop a specialized prompt system that directs LLMs to generate SVG templates using basic geometric primitives." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.44, + 0.905, + 0.5 + ], + "angle": 0, + "content": "- Dual-Stage Optimization. We implement a two-phase optimization process that preserves the semantic meaning of each shape while eliminating artifacts such as self-intersecting and jagged paths." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.5, + 0.905, + 0.545 + ], + "angle": 0, + "content": "- Iterative Editing. We enable iterative refinement of SVG through natural language instructions, making SVG creation more accessible." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.334, + 0.905, + 0.545 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.561, + 0.655, + 0.577 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.587, + 0.82, + 0.604 + ], + "angle": 0, + "content": "2.1. Large Language Models for Design" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.609, + 0.907, + 0.837 + ], + "angle": 0, + "content": "LLMs have emerged as powerful tools for graphics design tasks, demonstrating remarkable capabilities in understanding and reasoning about complex design specifications across diverse domains. Recent works have explored their applications in (1) visual layout and composition, including poster designs [5, 21, 34, 47], 3D scene arrangements [2, 10, 14, 32, 33, 38, 51], and placement of basic shapes for image generation [8, 46]; (2) content creation and manipulation for shapes [11], materials [15], and animation [22, 36]; and (3) design understanding [18, 42]. While these demonstrate LLMs' versatility, their potential for vector graphics remains unexplored. Our work addresses this gap by showing that LLMs can effectively generate structured SVG templates, extending their capabilities to the domain of vector graphics creation." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.848, + 0.868, + 0.864 + ], + "angle": 0, + "content": "2.2. Text-Guided Vector Graphics Generation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.905, + 0.902 + ], + "angle": 0, + "content": "Text-to-SVG generation approaches can be roughly categorized into image-based and language-based methods." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "23691" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.471 + ], + "angle": 0, + "content": "Image-based methods [9, 16, 43, 44, 50] start with a large collection of randomly initialized shapes, render vector graphics with a differentiable renderer [20], and iteratively optimize path parameters by minimizing text-image similarity losses (e.g., CLIP [28] loss and Score Distillation Sampling [27] loss). Despite their visually pleasant appearance, these methods often generate SVG results containing multiple fragmented paths that lack individual semantic correspondence, which is undesirable for real design scenarios. Language-based methods [3, 35, 41] design specialized tokenization strategies to encode vector graphics into discrete tokens, which are then concatenated with text tokens, and train seq2seq models like autoregressive transformers [37] on domain-specific datasets (e.g., icons and fonts). Although these approaches are conceptually elegant, their generalizability is limited by the scarcity of large-scale, general-purpose vector graphics datasets. Some works [26, 31] have explored LLMs' vector graphics generation capabilities, showing that while LLMs can create reasonable layouts matching text prompts, they struggle to produce complex geometric shapes. Our Chat2SVG combines image-based and language-based methods to address the limitations of each paradigm, producing semantically meaningful SVG where each path corresponds to a distinct visual element and contains refined geometric details." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.478, + 0.45, + 0.495 + ], + "angle": 0, + "content": "2.3. Vector Graphics Representation Learning" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.5, + 0.486, + 0.759 + ], + "angle": 0, + "content": "Vector graphics representation learning is essential for sketch-based retrieval, reconstruction, and generation tasks. A seminal study, SketchRNN [12], combines a Recurrent Neural Network (RNN) and a VAE to learn vector representations. Lopes et al. [24] trained a VAE to represent image-level font styles, then passed the latent embedding into a decoder for vector font generation. To increase reconstruction quality, later approaches like Sketchformer [30] and DeepSVG [4] adopted Transformer [37] architectures, followed by dual-modality methods [23, 39, 40] that leveraged both vector and image features. However, encoding the entire complex SVG into a single latent embedding often causes detail loss. To address this limitation, Zhang et al. [50] proposed path-level representational learning, creating a smooth latent space for individual paths. We leverage these path-level latent embeddings in our optimization to eliminate self-intersecting and jagged paths." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.77, + 0.184, + 0.787 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.485, + 0.901 + ], + "angle": 0, + "content": "Our Chat2SVG, as shown in Figure 2, begins with an SVG-oriented prompting approach (Section 3.1) that guides an LLM to create reasonable SVG templates. We then perform a dual-stage optimization (Section 3.2) to enhance the geometric details of SVG templates by optimizing SVG paths in both latent and control point space, guided by image diffusion models." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.091, + 0.845, + 0.108 + ], + "angle": 0, + "content": "3.1. SVG Template Generation with LLMs" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.113, + 0.907, + 0.461 + ], + "angle": 0, + "content": "Prompt Expansion. Users often provide vague and brief prompts to the LLM when describing their desired graphics. Such unstructured input may lead to oversimplified SVG (regarding both the number of elements and structures) that fails to reflect the intended design. Consequently, we propose a three-level prompt expansion strategy that systematically delineates the initial prompts: (1) Scene-level: Starting with an initial prompt, we instruct the LLM to analyze it holistically and identify essential objects that should appear. The LLM then expands the prompt by suggesting complementary objects to enhance scene completeness. (2) Object-level: For each object in the expanded prompt, we guide the LLM to systematically break it down into its components. For instance, when describing a lion, the LLM deconstructs it into distinct parts such as the body, head, mane, legs, eyes, ears, and tail. This decomposition ensures that no critical components are missing during SVG script generation. (3) Layout-level: After object decomposition, we direct the LLM to develop a comprehensive layout plan. This includes determining the position and size of each element on the canvas, selecting appropriate colors for visual harmony, and establishing clear spatial relationships to ensure a cohesive overall composition." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.462, + 0.909, + 0.597 + ], + "angle": 0, + "content": "SVG Script Generation. After prompt expansion, we obtain a detailed scene description. We then convert this natural language specification into SVG scripts in XML format. Since the LLM has limitations in synthesizing geometrically complex paths, we constrain the shapes to a carefully selected set of basic primitives: rectangles, ellipses, lines, polylines, polygons, and short paths. We set the canvas size to \\(512 \\times 512\\). Each geometric element is assigned a unique ID and includes semantic annotations in its comments." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.599, + 0.909, + 0.84 + ], + "angle": 0, + "content": "Visual Rectification. Since prompt expansion is purely text-based, even when the generated SVG script accurately follows the detailed prompt, visual inconsistencies (e.g., misaligned components, disproportionate scaling, and incorrect path ordering) can emerge during rendering. For example, in Figure 2, the initial SVG template shows a lion with a misshapen mane and missing facial features. We thus adopt a visual rectification strategy in which we render the SVG and provide the rendered image back to the LLM (with vision capability) for inspection. The LLM analyzes the visual output, identifies any inconsistencies or oddities, and generates corrected SVG code. This visual refinement loop can be performed iteratively. In our experiments, we find that two iterations of refinement are typically sufficient to generate well-structured SVG templates with appropriate spatial layouts." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.841, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Furthermore, to enhance the quality of prompt expansion and SVG script generation, we provide curated in-context examples in the prompts. The complete set of prompts and examples is available in the supplementary material." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.521, + 0.957 + ], + "angle": 0, + "content": "23692" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.165, + 0.09, + 0.835, + 0.194 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.164, + 0.196, + 0.536, + 0.321 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.542, + 0.196, + 0.835, + 0.321 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.333, + 0.907, + 0.418 + ], + "angle": 0, + "content": "Figure 2. The system pipeline of Chat2SVG. Given a text prompt, our system first leverages an LLM to generate an SVG template composed of basic geometric primitives. The rendered template is enhanced through SDEdit [25] with ControlNet [49] to add visual details while preserving the overall composition, yielding a target image. The SVG then undergoes a dual-stage optimization process to match the target image. (1) Primitives are converted to latent embeddings through latent inversion and optimized along with their visual attributes (i.e., filling colors \\( c_{i} \\), stroke properties \\( s_{i} \\), and transformation matrices \\( T_{i} \\)). (2) Point-level optimization is performed to refine the geometric details of SVG paths." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.443, + 0.482, + 0.46 + ], + "angle": 0, + "content": "3.2. SVG Optimization Guided by Image Diffusion" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.471, + 0.483, + 0.729 + ], + "angle": 0, + "content": "Detail Enhancement. While the primitive shapes in the SVG template have accurate semantic meaning, they lack the geometric details necessary for a professional design. To improve visual expressiveness, we first render the SVG template into a raster image, and then use an image editing method to generate a more detailed version that serves as our optimization target. Specifically, we employ SDEdit [25], which enhances the input image by first adding noise and then progressively denoising it with an image diffusion model to produce outputs with richer details, such as the contour of the lion's mane illustrated in Figure 2. To maintain structural similarity between the enhanced output and the original image, we incorporate a ControlNet (tile version) [49] into our pipeline. By using Gaussian-blurred versions of the initial images as control signals, the ControlNet effectively maintains the overall composition throughout the enhancement process." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Apart from geometric details, this image-to-image translation process also introduces new semantic parts such as the lion's beard. To faithfully reproduce the target image, we propose using the Segment Anything Model (SAM) [17] to identify these new decorative details. Specifically, we denote the rendered images of the initial SVG template and its diffusion-enhanced counterpart as \\( I_{\\mathrm{template}} \\) and \\( I_{\\mathrm{target}} \\), respectively. We apply SAM to extract mask sets from both images: \\( \\{m_i\\}_{i=1}^Q \\) for the template image and \\( \\{m_j\\}_{j=1}^K \\) for the target image, where \\( Q \\) and \\( K \\) represent the number of masks in each collection. Our approach is based on the ob" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.444, + 0.907, + 0.674 + ], + "angle": 0, + "content": "servation that the diffusion process simultaneously generates fine decorative details that are absent in the original template (e.g., the beard of the lion in Figure 2) while preserving the structural integrity of larger components (such as the body). Consequently, we implement an area-based mask-filtering strategy to identify these newly generated elements. For each mask \\( m_j \\) detected in \\( I_{\\mathrm{target}} \\), we evaluate it against all masks in \\( \\{m_i\\}_{i=1}^Q \\) using Intersection over Union (IoU) metrics [29]. When \\( m_j \\) overlaps with a template mask \\( m_i \\) but their IoU falls below a threshold, we classify \\( m_j \\) as a new decorative component. Alternatively, if their IoU exceeds the threshold, we interpret \\( m_j \\) as a variation of an existing component. For all filtered masks classified as new decorative shapes, we approximate their boundaries with polygons and integrate them into the SVG script." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.678, + 0.909, + 0.806 + ], + "angle": 0, + "content": "We define an SVG script \\( G \\) as a collection of \\( M \\) paths (i.e., shapes), \\( G = \\{P_i\\}_{i=1}^M \\). Each path \\( P_i \\) consists of a sequence of \\( N_i \\) commands, \\( P_i = \\{C_i^j\\}_{j=1}^{N_i} \\). A command \\( C_i^j = (U_i^j, V_i^j) \\) is defined by its type \\( U_i^j \\in \\{\\mathbb{M}, \\mathbb{C}\\} \\) and its associated control points \\( V_i^j \\), where \\( \\mathbb{M} \\) represents Move and \\( \\mathbb{C} \\) represents cubic Bézier curves. For consistency, we convert all other primitive shapes (e.g., rectangles and ellipses) in the SVG template into cubic Bézier curves." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Previous image-based methods [9, 16, 43, 44] optimize SVG shapes by directly manipulating the control points \\( V_{i}^{j} \\). Despite convenience, this approach often leads to artifacts such as self-intersecting curves and unnatural deformations. To overcome these limitations, Zhang et al. [50] introduced a path-level SVG VAE. This model's latent space effec" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23693" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.137 + ], + "angle": 0, + "content": "tively captures common shape patterns and geometric constraints, allowing for the optimization of path latent vectors at a higher level to produce smooth outputs." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.137, + 0.483, + 0.333 + ], + "angle": 0, + "content": "Latent Optimization. We leverage this pretrained SVG VAE to conduct the latent optimization. However, their SVG encoder expects a fixed number of commands (10 cubic Bezier curves), while our SVG primitives contain varying (typically fewer) commands. To resolve this mismatch, we develop a latent inversion process that converts our primitives into latent embeddings. We start with a randomly sampled latent vector \\(z\\) and decode it into a command sequence. Then we evenly sample points \\(X = \\{x_{i}\\}_{i=1}^{N}\\) along its contour. Similarly, we sample an equal number of points \\(Y = \\{y_{j}\\}_{j=1}^{N}\\) along the contour of our target primitive shape. Using these point sets, we compute the Earth Mover's Distance (EMD) [1, 7, 48]:" + }, + { + "type": "equation", + "bbox": [ + 0.145, + 0.34, + 0.483, + 0.371 + ], + "angle": 0, + "content": "\\[\n\\ell_ {\\mathrm {E M D}} (X, Y) = \\min _ {\\phi : X \\rightarrow Y} \\sum_ {x \\in X} \\| x - \\phi (x) \\| _ {2}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.378, + 0.483, + 0.422 + ], + "angle": 0, + "content": "where \\(\\phi\\) represents a bijective mapping between the point sets. By minimizing this EMD loss, we gradually optimize the latent embedding \\(z\\) to match the target primitive." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.422, + 0.484, + 0.635 + ], + "angle": 0, + "content": "After obtaining latent vectors \\(\\{z_i\\}_{i=1}^M\\) for all paths, we jointly optimize these vectors along with visual attribute parameters: filling colors \\(\\{c_i\\}_{i=1}^M\\), stroke properties (color and width) \\(\\{s_i\\}_{i=1}^M\\), and transformation matrices \\(\\{T_i\\}_{i=1}^M\\). At each optimization iteration \\(t\\), we decode the latent vectors into command sequences and apply transformation matrices to the control points. We then combine color and stroke attributes together, and use a differentiable rasterizer [20] to render the complete SVG into an image \\(I_t\\). Our optimization objective consists of three loss terms: (1) An MSE loss \\(\\ell_{\\mathrm{MSE}}\\) between the rendered image \\(I_t\\) and target image \\(I_{\\mathrm{target}}\\) to ensure visual similarity; (2) A curvature loss that reduces sharp bends and fluctuations along the contour by computing the discrete second derivative:" + }, + { + "type": "equation", + "bbox": [ + 0.116, + 0.641, + 0.483, + 0.711 + ], + "angle": 0, + "content": "\\[\n\\ell_ {\\text {c u r v a t u r e}} = \\frac {\\sum_ {i = 1} ^ {M} \\sum_ {j = 1} ^ {N _ {i} - 2} \\left\\| V _ {i} ^ {j} - 2 V _ {i} ^ {j + 1} + V _ {i} ^ {j + 2} \\right\\| _ {2} ^ {2}}{\\sum_ {i = 1} ^ {M} \\left(N _ {i} - 2\\right)}; \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.717, + 0.483, + 0.763 + ], + "angle": 0, + "content": "(3) A path-level IoU loss that preserves the overall composition and prevents paths from drifting too far from their initial positions:" + }, + { + "type": "equation", + "bbox": [ + 0.168, + 0.77, + 0.483, + 0.81 + ], + "angle": 0, + "content": "\\[\n\\ell_ {\\mathrm {I o U}} = \\frac {1}{M} \\sum_ {i = 1} ^ {M} \\left(1 - \\frac {\\left| m _ {i} ^ {t} \\cap m _ {i} ^ {0} \\right|}{\\left| m _ {i} ^ {t} \\cup m _ {i} ^ {0} \\right|}\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.817, + 0.483, + 0.848 + ], + "angle": 0, + "content": "where \\( m_i^t \\) and \\( m_i^0 \\) denote the binary masks of the \\( i \\)-th path at the current iteration \\( t \\) and its initial state, respectively." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.848, + 0.483, + 0.878 + ], + "angle": 0, + "content": "The final loss function combines these three terms with empirically determined weights:" + }, + { + "type": "equation", + "bbox": [ + 0.151, + 0.886, + 0.483, + 0.902 + ], + "angle": 0, + "content": "\\[\n\\ell_ {\\text {l a t e n t}} = \\ell_ {\\mathrm {M S E}} + \\lambda_ {1} \\ell_ {\\text {c u r v a t u r e}} + \\lambda_ {2} \\ell_ {\\mathrm {I o U}}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.764, + 0.106 + ], + "angle": 0, + "content": "where \\(\\lambda_1 = 5e - 4\\) and \\(\\lambda_{2} = 5e - 6\\)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.107, + 0.905, + 0.348 + ], + "angle": 0, + "content": "Point Optimization. While latent optimization effectively aligns SVG shapes with their target positions and contours, it has a notable limitation: the resulting shapes tend to be overly smooth, lacking the intricate details that are often essential in professional vector graphics. For example, in Figure 2, the lion's mane after latent optimization does not capture the fine contour present in the target image. Therefore, we add a second optimization stage that directly refines the control points \\( V_{i}^{j} \\) of the paths, similar to previous image-based works [9, 16, 43, 44]. To achieve finer granularity in shape control, we split each cubic Bezier curve at its midpoint, effectively doubling the number of control points in the SVG. This increased control point density allows for more precise shape adjustments. The point optimization stage employs a loss function that combines the MSE loss and the curvature loss:" + }, + { + "type": "equation", + "bbox": [ + 0.608, + 0.362, + 0.905, + 0.379 + ], + "angle": 0, + "content": "\\[\n\\ell_ {\\text {p o i n t}} = \\ell_ {\\mathrm {M S E}} + \\lambda_ {3} \\ell_ {\\text {c u r v a t u r e}}. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.39, + 0.905, + 0.421 + ], + "angle": 0, + "content": "The parameter \\(\\lambda_3\\) decreases linearly from \\(1e - 3\\) to \\(5e - 5\\) throughout the optimization process." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.431, + 0.677, + 0.447 + ], + "angle": 0, + "content": "3.3. Iterative Editing" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.453, + 0.905, + 0.528 + ], + "angle": 0, + "content": "During SVG script generation (Section 3.1), we instruct the LLM to annotate the semantic label of each path. This design enables users to refine the initial SVG template using natural language, as the LLM can precisely locate and modify specific elements based on semantic understanding." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.529, + 0.906, + 0.71 + ], + "angle": 0, + "content": "Our system supports iterative refinement through multiple rounds of natural language instructions. However, when passing the edited SVG template image into SDEdit, image diffusion models may introduce different shape variations and decorative details, which compromise the user's intention to maintain the already optimized shapes. To preserve consistency across editing iterations, we instruct the LLM to output both the modified SVG script and a list of specifically changed paths. This precise tracking enables selective optimization of modified shapes while preserving unaltered elements from previous iterations, maintaining visual coherence throughout the entire refinement process." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.725, + 0.646, + 0.741 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Implementation Details. In our experiments, we use claude-3-5-sonnet as our backbone LLM model due to its leading generation capabilities. To evaluate our approach, we create a prompt set by having the LLM generate 125 text prompts across five categories: animals, food, objects, scenes, and novel concept combinations. For each prompt, we generate one template and perform 2 rounds of visual rectification. This process is repeated 5 times, yielding \\((1 + 2) \\times 5 = 15\\) candidate SVG templates per prompt. These templates are rendered as images, and we follow a" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23694" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.134, + 0.086, + 0.868, + 0.557 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.566, + 0.908, + 0.665 + ], + "angle": 0, + "content": "Figure 3. Qualitative Comparison. (1) Methods refining open-ended strokes, i.e., CLIPDraw [9] and DiffSketcher [43], often produce distorted and disorganized strokes to approximate objects, presenting messy appearance and poor text alignment. (2) VectorFusion [16] and SVGDreamer [44] produce elements that consist of multiple jagged, irregular, and fragmented shapes, such as the body of the flamingo (first row) and the plane (third row). (3) T2V-NPR [50] attempts to resolve these issues by learning a latent representation of paths and merging fragmented shapes. However, it still cannot guarantee the semantic meanings of the paths, leading to less-semantic paths such as a plane body with surrounding clouds in the third row. In contrast, our method produces SVG with superior text alignment, higher visual quality, and well-structured paths exhibiting geometric regularity and clear semantic definition." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.69, + 0.482, + 0.737 + ], + "angle": 0, + "content": "standard practice to select the highest quality SVG using the ImageReward [45] metric. Detailed optimization configurations are provided in the supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.742, + 0.483, + 0.879 + ], + "angle": 0, + "content": "Baselines. We compare our approach against five state-of-the-art image-based text-to-SVG generation methods: CLIPDraw [9], DiffSketcher [43], VectorFusion [16], SVGDreamer [44], and T2V-NPR [50]. Current language-based methods [35, 41] are confined to specific categories, with no support for general text-to-SVG generation, so we do not include them in our comparison. For a fair comparison, we set these baseline methods to use a similar number of shapes as our optimized SVG." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.886, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Evaluation Metrics. We evaluate the quality of our gen" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.69, + 0.905, + 0.72 + ], + "angle": 0, + "content": "erated SVG across three dimensions: visual fidelity, path regularity, and semantic alignment." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.722, + 0.905, + 0.827 + ], + "angle": 0, + "content": "- Image-level Fidelity. We evaluate visual fidelity using a ground-truth dataset of well-designed vector graphics, specifically 52,805 colored SVG files downloaded from SVGRepo1. We compute the Fréchet Inception Distance (FID) [13] between the rendered images of our generated SVG and this professional design collection, using features extracted by the CLIP image encoder [28]." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.828, + 0.905, + 0.873 + ], + "angle": 0, + "content": "- Path-level Regularity. We assess path quality using a transformer-based Auto-Encoder trained on the FIGR-8-SVG [6] dataset, following the DeepSVG [4] architecture" + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.722, + 0.905, + 0.873 + ], + "angle": 0, + "content": null + }, + { + "type": "page_footnote", + "bbox": [ + 0.532, + 0.888, + 0.724, + 0.9 + ], + "angle": 0, + "content": "1https://www.svgrepo.com" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23695" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.124, + 0.089, + 0.454, + 0.201 + ], + "angle": 0, + "content": "
MethodImage FID ↓Path FID ↓Text Alignment ↑
CLIPDraw [9]46.7770.130.3048
DiffSketcher [43]44.8966.480.2623
VectorFusion [16]39.5256.790.2982
SVGDreamer [44]35.4847.950.2919
T2V-NPR [50]39.8642.030.3078
Chat2SVG (Ours)33.3139.070.3090
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.21, + 0.483, + 0.24 + ], + "angle": 0, + "content": "Table 1. Quantitative comparison of text-to-SVG generation methods across image fidelity, vector regularity, and text alignment." + }, + { + "type": "text", + "bbox": [ + 0.103, + 0.266, + 0.483, + 0.37 + ], + "angle": 0, + "content": "ture. Our model, trained with reconstruction tasks, encodes each drawing command \\( C_i^j \\) into a latent vector. We represent each path as the mean of its commands' latent embeddings, and calculate the FID between these path representations and the ground truth paths from FIGR-8-SVG. This metric quantifies how closely our generated paths align with professional SVG design patterns." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.372, + 0.484, + 0.417 + ], + "angle": 0, + "content": "- Text-level Alignment. We assess semantic alignment by computing the CLIP score [28] between the text prompt and the rendered SVG image." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.426, + 0.401, + 0.442 + ], + "angle": 0, + "content": "4.1. Comparison with Existing Methods" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.448, + 0.483, + 0.583 + ], + "angle": 0, + "content": "Quantitative Comparison. Table 1 shows the evaluation metrics for all baseline methods. Our method achieves the best image FID score, indicating that our generated SVG closely aligns with professional design patterns. Regarding path regularity, our method yields paths that are closest to the professionally designed SVG, as evidenced by the lowest path FID. Furthermore, our Chat2SVG achieves the highest score in text-SVG alignment, validating the significance of LLM-generated SVG templates." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.585, + 0.483, + 0.734 + ], + "angle": 0, + "content": "Qualitative Comparison. In Figure 3, we present a side-by-side comparison between our Chat2SVG and the baselines. Methods that optimize open-ended strokes based on CLIP or Diffusion models (i.e., CLIPDraw [9] and DiffS-ketcher [43]) produce results with messy visual appearances and poor text alignment. These methods sometimes fail to synthesize complete objects, such as the flamingo (first row) and the plane (third row) in Figure 3. In contrast, our method generates SVG results with clear layouts and greater fidelity to the prompts." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.735, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Methods that optimize closed shapes via score distillation of diffusion models (i.e., VectorFusion [16], SVGDreamer [44], and T2V-NPR [50]) can better align with the input text, but still produce fragmented paths with limited semantic meaning. For instance, in the plane example (third row of Figure 3), VectorFusion and SVGDreamer create planes and clouds composed of multiple jagged, irregular, and fragmented shapes, which only appear meaningful when viewed as a whole. This hinders convenient editing by graphic designers. T2V-NPR [50] addresses the issue of jagged paths using a path VAE (although it can be" + }, + { + "type": "image", + "bbox": [ + 0.554, + 0.088, + 0.868, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.276, + 0.907, + 0.305 + ], + "angle": 0, + "content": "Figure 4. User Study. Our Chat2SVG achieves the highest user selection ratio across all three evaluation criteria." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.334, + 0.907, + 0.531 + ], + "angle": 0, + "content": "overly smooth at times) and reduces fragmented shapes by merging shapes with similar colors. However, its merging operation ignores the semantic meaning of shapes, resulting in the body of the plane being merged with the surrounding clouds. Additionally, the semantic meaning of the path is sometimes unclear, as exemplified by the cloud in the bottom-right corner, whose contour is actually approximated by the blue background instead of a true cloud shape. In contrast, our Chat2SVG adopts a unique approach by using LLMs to generate paths representing semantic components and performing dual-stage optimization based on image diffusion to enhance path expressiveness, thus ensuring both path regularity and visual fidelity of generated SVG." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.532, + 0.909, + 0.789 + ], + "angle": 0, + "content": "User Study. To further compare Chat2SVG with baseline methods, we conducted a user study examining the same key aspects: (1) overall visual aesthetics, (2) path regularity, and (3) text-SVG alignment. We randomly sampled 20 prompts and generated SVG outputs using our approach and five baseline methods. For each prompt, participants were shown all generated SVG results along with the corresponding text prompt and highlighted paths. They were then asked to select the highest quality result regarding the three evaluation criteria. We recruited 31 participants (18 male, 13 female) through university mailing lists, with a mean age of 26 years. Among them, 17 participants reported prior experience in graphic design, providing a balanced mix of expert and novice perspectives. Analyzing the average selection ratio across all three metrics (Figure 4), we found that Chat2SVG-generated results were consistently preferred by participants over baseline methods." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.802, + 0.667, + 0.818 + ], + "angle": 0, + "content": "4.2. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.906, + 0.903 + ], + "angle": 0, + "content": "We conduct ablation experiments to evaluate the effectiveness of key components in our pipeline (Section 3.1). First, we remove the SVG template generation and use randomly initialized shapes to approximate target images. As shown in the second row of Table 2, both image FID and path FID" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23696" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.114, + 0.089, + 0.462, + 0.164 + ], + "angle": 0, + "content": "
Chat2SVG VariantImage FID ↓Path FID ↓Text Alignment ↑
Default33.3139.070.3090
No SVG Template47.2170.130.2968
No SVG Optimization33.4536.120.3044
" + }, + { + "type": "table_caption", + "bbox": [ + 0.142, + 0.174, + 0.432, + 0.188 + ], + "angle": 0, + "content": "Table 2. Qualitative results of the ablation study." + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.202, + 0.484, + 0.309 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.15, + 0.32, + 0.424, + 0.334 + ], + "angle": 0, + "content": "Figure 5. Qualitative results of ablation study." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.36, + 0.484, + 0.572 + ], + "angle": 0, + "content": "values increase substantially. This performance degradation can be attributed to the substantial gap between randomly initialized shapes and target shapes, making optimization more challenging. Meanwhile, in the second column of Figure 5, the dog's legs and ears lack semantic-clear components, highlighting the importance of SVG template generation. Second, we eliminate the dual-stage optimization. As shown in the third row of Table 2, this results in the lowest path FID, as primitives generated by LLMs naturally align with the path regularity presented in the ground-truth dataset. However, the visual outcome in Figure 5 lacks the necessary details, indicating that the overall visual quality suffers without dual-stage optimization. A more detailed ablation study is provided in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.582, + 0.315, + 0.598 + ], + "angle": 0, + "content": "4.3. Iterative Editing Results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.604, + 0.484, + 0.756 + ], + "angle": 0, + "content": "In this section, we present the iterative editing results. As demonstrated in Figure 6, users can progressively refine the SVG through multiple rounds of natural language instructions. During editing, we preserve the optimized shapes from previous iterations while only modifying the specified shapes. The results showcase how Chat2SVG can accurately interpret and execute user editing requests, maintaining both semantic coherence and visual quality throughout the modifications. Additional editing examples are provided in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.77, + 0.21, + 0.785 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.484, + 0.885 + ], + "angle": 0, + "content": "In this paper, we present Chat2SVG, a novel paradigm for text-to-SVG generation that combines LLMs and image diffusion models. Through our carefully designed SVG template generation and dual-stage SVG optimization pipeline, our method generates high-quality SVG outputs that exhibit strong visual fidelity, path regularity, and text alignment." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.886, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Despite our method's superior results, there are several" + }, + { + "type": "image", + "bbox": [ + 0.536, + 0.089, + 0.885, + 0.346 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.357, + 0.907, + 0.4 + ], + "angle": 0, + "content": "Figure 6. Iterative Editing. We perform two rounds of refinement on each SVG template and show the optimized output. This figure shows editing types including deletion, modification, and addition." + }, + { + "type": "text", + "bbox": [ + 0.536, + 0.415, + 0.636, + 0.426 + ], + "angle": 0, + "content": "(1) Layout Precision" + }, + { + "type": "image", + "bbox": [ + 0.536, + 0.428, + 0.634, + 0.466 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.536, + 0.47, + 0.627, + 0.49 + ], + "angle": 0, + "content": "SVG Target Template Image" + }, + { + "type": "image", + "bbox": [ + 0.558, + 0.491, + 0.61, + 0.535 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.545, + 0.536, + 0.623, + 0.547 + ], + "angle": 0, + "content": "Optimized SVG" + }, + { + "type": "text", + "bbox": [ + 0.653, + 0.415, + 0.74, + 0.426 + ], + "angle": 0, + "content": "(2) Path Addition" + }, + { + "type": "image", + "bbox": [ + 0.642, + 0.428, + 0.748, + 0.469 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.651, + 0.47, + 0.739, + 0.49 + ], + "angle": 0, + "content": "SVG Target Template Image" + }, + { + "type": "image", + "bbox": [ + 0.667, + 0.491, + 0.726, + 0.535 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.656, + 0.536, + 0.735, + 0.547 + ], + "angle": 0, + "content": "Optimized SVG" + }, + { + "type": "text", + "bbox": [ + 0.765, + 0.415, + 0.872, + 0.426 + ], + "angle": 0, + "content": "(3) Shape Complexity" + }, + { + "type": "image", + "bbox": [ + 0.761, + 0.428, + 0.824, + 0.47 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.767, + 0.47, + 0.814, + 0.49 + ], + "angle": 0, + "content": "SVG Template" + }, + { + "type": "image", + "bbox": [ + 0.825, + 0.428, + 0.885, + 0.47 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.84, + 0.47, + 0.871, + 0.49 + ], + "angle": 0, + "content": "Target Image" + }, + { + "type": "image", + "bbox": [ + 0.766, + 0.491, + 0.829, + 0.536 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.786, + 0.536, + 0.864, + 0.547 + ], + "angle": 0, + "content": "Optimized SVG" + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.558, + 0.907, + 0.601 + ], + "angle": 0, + "content": "Figure 7. Limitations of our method include imprecise layout generation, missing important visual elements, and insufficient shape complexity." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.629, + 0.907, + 0.901 + ], + "angle": 0, + "content": "aspects that could be further improved: (1) LLMs may produce imprecise visual layouts, such as the burger in Figure 7. This can be alleviated by collecting an SVG dataset and extracting the corresponding layouts to fine-tune LLMs. (2) Our approach uses SAM to identify new parts. However, SAM may overlook important regions, like the pink frosting on the donut in Figure 7. Incorporating semantic-aware SAM [19] could potentially produce more accurate results. (3) The pretrained SVG VAE from [50] has a fixed number of commands. This makes it hard to represent intricate shapes. For example, the lion's mane in Figure 7 lacks sufficient detail. Training a new SVG VAE with varying-length commands would help solve this problem. Our work can be extended to generate other types of vectorized data, such as icons and fonts. Additionally, through consistent editing of SVG across multiple frames, coupled with shape interpolation, our Chat2SVG has the potential to facilitate keyframe animation generation." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23697" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.251, + 0.108 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.114, + 0.484, + 0.172 + ], + "angle": 0, + "content": "The work described in this paper was fully supported by a GRF grant from the Research Grants Council (RGC) of the Hong Kong Special Administrative Region, China [Project No. CityU 11216122]." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.198, + 0.188, + 0.214 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.224, + 0.484, + 0.279 + ], + "angle": 0, + "content": "[1] Panos Achlioptas, Olga Diamanti, Ioannis Mitliagkas, and Leonidas Guibas. Learning representations and generative models for 3D point clouds. In International Conference on Machine Learning, pages 40-49, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.281, + 0.484, + 0.363 + ], + "angle": 0, + "content": "[2] Rio Aguina-Kang, Maxim Gumin, Do Heon Han, Stewart Morris, Seung-Jean Yoo, Aditya Ganeshan, R. Kenny Jones, Qiuhong Anna Wei, Kailiang Fu, and Daniel Ritchie. Open-universe indoor scene generation using LLM program synthesis and uncurated object databases. In arXiv preprint arXiv:2403.09675, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.365, + 0.483, + 0.419 + ], + "angle": 0, + "content": "[3] Jonas Belouadi, Anne Lauscher, and Steffen Eger. AutomataTikZ: Text-guided synthesis of scientific vector graphics with TikZ. In arXiv preprint arXiv:2310.00367, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.422, + 0.483, + 0.49 + ], + "angle": 0, + "content": "[4] Alexandre Carlier, Martin Danelljan, Alexandre Alahi, and Radu Timofte. DeepSVG: A hierarchical generative network for vector graphics animation. In Advances in Neural Information Processing Systems, pages 16351-16361, 2020. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.492, + 0.484, + 0.546 + ], + "angle": 0, + "content": "[5] Yutao Cheng, Zhao Zhang, Maoke Yang, Hui Nie, Chunyuan Li, Xinglong Wu, and Jie Shao. Graphic design with large multimodal model. In arXiv preprint arXiv:2404.14368, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.549, + 0.483, + 0.588 + ], + "angle": 0, + "content": "[6] Louis Clouatre and Marc Demers. FIgR: Few-shot image generation with reptile. In arXiv preprint arXiv:1901.02199, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.591, + 0.483, + 0.647 + ], + "angle": 0, + "content": "[7] Haoqiang Fan, Hao Su, and Leonidas Guibas. A point set generation network for 3D object reconstruction from a single image. In IEEE Conference on Computer Vision and Pattern Recognition, pages 605-613, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.649, + 0.483, + 0.729 + ], + "angle": 0, + "content": "[8] Weixi Feng, Wanrong Zhu, Tsu-jui Fu, Varun Jampani, Arjun Akula, Xuehai He, Sugato Basu, Xin Wang, and William Wang. LayoutGPT: Compositional visual planning and generation with large language models. In Advances in Neural Information Processing Systems, pages 18225-18250, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.733, + 0.483, + 0.788 + ], + "angle": 0, + "content": "[9] Kevin Frans, Lisa B. Soros, and Olaf Witkowski. CLIPDraw: Exploring text-to-drawing synthesis through language-image encoders. In Advances in Neural Information Processing Systems, pages 5207-5218, 2021. 2, 3, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.79, + 0.483, + 0.844 + ], + "angle": 0, + "content": "[10] Rao Fu, Zehao Wen, Zichen Liu, and Srinath Sridhar. AnyHome: Open-vocabulary generation of structured and textured 3D homes. In European Conference on Computer Vision, pages 52-70, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.483, + 0.9 + ], + "angle": 0, + "content": "[11] Aditya Ganeshan, Ryan Y. Huang, Xianghao Xu, R. Kenny Jones, and Daniel Ritchie. ParSEL: Parameterized shape editing with language. In arXiv preprint arXiv:2405.20319, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.224, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.133 + ], + "angle": 0, + "content": "[12] David Ha and Douglas Eck. A neural representation of sketch drawings. In arXiv preprint arXiv:1704.03477, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.136, + 0.906, + 0.205 + ], + "angle": 0, + "content": "[13] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. GANs trained by a two time-scale update rule converge to a local Nash equilibrium. In Advances in Neural Information Processing Systems, pages 6629-6640, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.208, + 0.906, + 0.276 + ], + "angle": 0, + "content": "[14] Ziniu Hu, Ahmet Iscen, Aashi Jain, Thomas Kipf, Yisong Yue, David A. Ross, Cordelia Schmid, and Alireza Fathi. SceneCraft: An LLM agent for synthesizing 3D scenes as Blender code. In International Conference on Machine Learning, pages 19252-19282, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.278, + 0.905, + 0.319 + ], + "angle": 0, + "content": "[15] Ian Huang, Guandao Yang, and Leonidas Guibas. Blender: Editing 3D graphics with vision-language models. In arXiv preprint arXiv:2404.17672, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.321, + 0.905, + 0.376 + ], + "angle": 0, + "content": "[16] Ajay Jain, Amber Xie, and Pieter Abbeel. VectorFusion: Text-to-SVG by abstracting pixel-based diffusion models. In IEEE Conference on Computer Vision and Pattern Recognition, pages 1911-1920, 2022. 2, 3, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.378, + 0.905, + 0.447 + ], + "angle": 0, + "content": "[17] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dólar, and Ross Girshick. Segment anything. In IEEE International Conference on Computer Vision, pages 4015-4026, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.449, + 0.905, + 0.502 + ], + "angle": 0, + "content": "[18] Peter Kulits, Haiwen Feng, Weiyang Liu, Victoria Abrevaya, and Michael J. Black. Re-thinking inverse graphics with large language models. In arXiv preprint arXiv:2404.15228, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.505, + 0.905, + 0.572 + ], + "angle": 0, + "content": "[19] Feng Li, Hao Zhang, Peize Sun, Xueyan Zou, Shilong Liu, Jianwei Yang, Chunyuan Li, Lei Zhang, and Jianfeng Gao. Semantic-SAM: Segment and recognize anything at any granularity. In arXiv preprint arXiv:2307.04767, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.576, + 0.905, + 0.631 + ], + "angle": 0, + "content": "[20] Tzu-Mao Li, Michal Lukáč, Michael Gharbi, and RaganKelley Jonathan. Differentiable vector graphics rasterization for editing and learning. ACM Transactions on Graphics, 39 (6):1-15, 2020. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.633, + 0.905, + 0.701 + ], + "angle": 0, + "content": "[21] Jiawei Lin, Jiaqi Guo, Shizhao Sun, Zijiang Yang, JianGuang Lou, and Dongmei Zhang. LayoutPrompter: Awaken the design ability of large language models. In Advances in Neural Information Processing Systems, pages 43852-43879, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.704, + 0.905, + 0.771 + ], + "angle": 0, + "content": "[22] Vivian Liu, Rubaiat H. Kazi, Li-Yi Wei, Matthew Fisher, Timothy Langlois, Seth Walker, and Lydia Chilton. LogoMotion: Visually grounded code generation for content-aware animation. In arXiv preprint arXiv:2405.07065, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.775, + 0.905, + 0.844 + ], + "angle": 0, + "content": "[23] YingTian Liu, Zhifei Zhang, YuanChen Guo, Matthew Fisher, Zhaowen Wang, and SongHai Zhang. DualVector: Unsupervised vector font synthesis with dual-part representation. In IEEE Conference on Computer Vision and Pattern Recognition, pages 14193-14202, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[24] Raphael G. Lopes, David Ha, Douglas Eck, and Jonathon Shlens. A learned representation for scalable vector graphics. In IEEE International Conference on Computer Vision, pages 7930-7939, 2019. 3" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.093, + 0.906, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23698" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.148 + ], + "angle": 0, + "content": "[25] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia-jun Wu, JunYan Zhu, and Stefano Ermon. SDEdit: Guided image synthesis and editing with stochastic differential equations. In arXiv preprint arXiv:2108.01073, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.151, + 0.483, + 0.178 + ], + "angle": 0, + "content": "[26] OpenAI. GPT-4 technical report. In arXiv preprint arXiv:2303.08774, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.181, + 0.483, + 0.223 + ], + "angle": 0, + "content": "[27] Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. DreamFusion: Text-to-3D using 2D diffusion. In arXiv preprint arXiv:2209.14988, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.226, + 0.483, + 0.322 + ], + "angle": 0, + "content": "[28] Alec Radford, Jong-Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763, 2021. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.325, + 0.483, + 0.395 + ], + "angle": 0, + "content": "[29] Hamid Rezatofighi, Nathan Tsoi, JunYoung Gwak, Amir Sadeghian, Ian Reid, and Silvio Savarese. Generalized intersection over union: A metric and a loss for bounding box regression. In IEEE Conference on Computer Vision and Pattern Recognition, pages 658-666, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.397, + 0.483, + 0.453 + ], + "angle": 0, + "content": "[30] Leo S. F. Ribeiro, Tu Bui, John Collomosse, and Moacir Ponti. Sketchformer: Transformer-based representation for sketched structure. In IEEE Conference on Computer Vision and Pattern Recognition, pages 14153-14162, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.455, + 0.483, + 0.525 + ], + "angle": 0, + "content": "[31] Pratyusha Sharma, Tamar Rott Shaham, Manel Baradad, Stephanie Fu, Adrian Rodriguez-Munoz, Shivam Duggal, Phillip Isola, and Antonio Torralba. A vision check-up for language models. In IEEE Conference on Computer Vision and Pattern Recognition, pages 14410-14419, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.527, + 0.483, + 0.582 + ], + "angle": 0, + "content": "[32] Chunyi Sun, Junlin Han, Weijian Deng, Xinlong Wang, Zishan Qin, and Stephen Gould. 3D-GPT: Procedural 3D modeling with large language models. In arXiv preprint arXiv:2310.12945, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.585, + 0.483, + 0.653 + ], + "angle": 0, + "content": "[33] Hou In Ivan Tam, Hou In Derek Pun, Austin T. Wang, Angel X. Chang, and Manolis Savva. SceneMotifCoder: Example-driven visual program learning for generating 3D object arrangements. In arXiv preprint arXiv:2408.02211, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.657, + 0.483, + 0.711 + ], + "angle": 0, + "content": "[34] Zecheng Tang, Chenfei Wu, Juntao Li, and Nan Duan. LayoutNUWA: Revealing the hidden layout expertise of large language models. In arXiv preprint arXiv:2309.09506, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.715, + 0.483, + 0.785 + ], + "angle": 0, + "content": "[35] Zecheng Tang, Chenfei Wu, Zekai Zhang, Mingheng Ni, Shengming Yin, Yu Liu, Zhengyuan Yang, Lijuan Wang, Zicheng Liu, Juntao Li, and Nan Duan. StrokeNUWA: Tokenizing strokes for vector graphic synthesis. In arXiv preprint arXiv:2401.17093, 2024. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.787, + 0.483, + 0.841 + ], + "angle": 0, + "content": "[36] Tiffany Tseng, Ruijia Cheng, and Jeffrey Nichols. Keyframer: Empowering animation design using large language models. In arXiv preprint arXiv:2402.06071, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.845, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[37] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems, pages 6000–6010, 2017. 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.148 + ], + "angle": 0, + "content": "[38] Can Wang, Hongliang Zhong, Menglei Chai, Mingming He, Dongdong Chen, and Jing Liao. Chat2Layout: Interactive 3D furniture layout with a multimodal LLM. In arXiv preprint arXiv:2407.21333, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.15, + 0.905, + 0.191 + ], + "angle": 0, + "content": "[39] Yizhi Wang and Zhouhui Lian. DeepVecFont: Synthesizing high-quality vector fonts via dual-modality learning. ACM Transactions on Graphics, 40(6):1-15, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.193, + 0.905, + 0.261 + ], + "angle": 0, + "content": "[40] Yuqing Wang, Yizhi Wang, Longhui Yu, Yuesheng Zhu, and Zhouhui Lian. DeepVecFont-v2: Exploiting Transformers to synthesize vector fonts with higher quality. In IEEE Conference on Computer Vision and Pattern Recognition, pages 18320-18328, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.263, + 0.905, + 0.318 + ], + "angle": 0, + "content": "[41] Ronghuan Wu, Wanchao Su, Kede Ma, and Jing Liao. IconShop: Text-guided vector icon synthesis with autoregressive Transformers. ACM Transactions on Graphics, 42(6):1-14, 2023. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.32, + 0.905, + 0.389 + ], + "angle": 0, + "content": "[42] Tong Wu, Guandao Yang, Zhibing Li, Kai Zhang, Ziwei Liu, Leonidas Guibas, Dahua Lin, and Gordon Wetzstein. GPT-4V (ision) is a human-aligned evaluator for text-to-3D generation. In IEEE Conference on Computer Vision and Pattern Recognition, pages 22227–22238, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.391, + 0.905, + 0.459 + ], + "angle": 0, + "content": "[43] Ximing Xing, Chuang Wang, Haitao Zhou, Jing Zhang, Qian Yu, and Dong Xu. DiffSketcher: Text guided vector sketch synthesis through latent diffusion models. In Advances in Neural Information Processing Systems, pages 15869-15889, 2023. 2, 3, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.461, + 0.905, + 0.531 + ], + "angle": 0, + "content": "[44] Ximing Xing, Haitao Zhou, Chuang Wang, Jing Zhang, Dong Xu, and Qian Yu. SVGDreamer: Text guided SVG generation with diffusion model. In IEEE Conference on Computer Vision and Pattern Recognition, pages 4546-4555, 2024. 2, 3, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.532, + 0.905, + 0.602 + ], + "angle": 0, + "content": "[45] Jiazheng Xu, Xiao Liu, Yuchen Wu, Yuxuan Tong, Qinkai Li, Ming Ding, Jie Tang, and Yuxiao Dong. ImageReward: Learning and evaluating human preferences for text-to-image generation. In Advances in Neural Information Processing Systems, pages 15903-15935, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.603, + 0.905, + 0.658 + ], + "angle": 0, + "content": "[46] Yutaro Yamada, Khyathi Chandu, Yuchen Lin, Jack Hessel, Ilker Yildirim, and Yejin Choi. L3GO: Language agents with chain-of-3D-thoughts for generating unconventional objects. In arXiv preprint arXiv:2402.09052, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.66, + 0.905, + 0.714 + ], + "angle": 0, + "content": "[47] Tao Yang, Yingmin Luo, Zhongang Qi, Yang Wu, Ying Shan, and Chang Wen Chen. PosterLLaVa: Constructing a unified multi-modal layout generator with LLM. In arXiv preprint arXiv:2406.02884, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.716, + 0.905, + 0.772 + ], + "angle": 0, + "content": "[48] Lequan Yu, Xianzhi Li, Chi-Wing Fu, Daniel Cohen-Or, and Pheng-Ann Heng. PU-Net: Point cloud upsampling network. In IEEE Conference on Computer Vision and Pattern Recognition, pages 2790–2799, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.774, + 0.905, + 0.828 + ], + "angle": 0, + "content": "[49] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In IEEE International Conference on Computer Vision, pages 3836-3847, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.83, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[50] Peiying Zhang, Nanxuan Zhao, and Jing Liao. Text-to-vector generation with neural path representation. ACM Transactions on Graphics, 43(4):1-13, 2024. 2, 3, 4, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[51] Qihang Zhang, Chaoyang Wang, Aliaksandr Siarohin, Peiye Zhuang, Yinghao Xu, Ceyuan Yang, Dahua Lin, Bolei Zhou," + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "23699" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.124, + 0.092, + 0.486, + 0.137 + ], + "angle": 0, + "content": "Sergey Tulyakov, and Hsin-Ying Lee. SceneWiz3D: Towards text-guided 3D scene composition. In arXiv preprint arXiv:2312.08885, 2023. 2" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.521, + 0.957 + ], + "angle": 0, + "content": "23700" + } + ] +] \ No newline at end of file diff --git a/2025/Chat2SVG_ Vector Graphics Generation with Large Language Models and Image Diffusion Models/03e11a55-7ca3-4174-a887-3307219befcc_origin.pdf b/2025/Chat2SVG_ Vector Graphics Generation with Large Language Models and Image Diffusion Models/03e11a55-7ca3-4174-a887-3307219befcc_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..874640ffcbf857b4bf0470003a3ceac497cd3067 --- /dev/null +++ b/2025/Chat2SVG_ Vector Graphics Generation with Large Language Models and Image Diffusion Models/03e11a55-7ca3-4174-a887-3307219befcc_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1cb7616b3e49fdf4dac30b409cbec9130ec2f5f2e4986c6251396eb9b2ac71dc +size 3487865 diff --git a/2025/Chat2SVG_ Vector Graphics Generation with Large Language Models and Image Diffusion Models/full.md b/2025/Chat2SVG_ Vector Graphics Generation with Large Language Models and Image Diffusion Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..867eb9a47c76f3cc1c461f86c381353196293096 --- /dev/null +++ b/2025/Chat2SVG_ Vector Graphics Generation with Large Language Models and Image Diffusion Models/full.md @@ -0,0 +1,290 @@ +# Chat2SVG: Vector Graphics Generation with Large Language Models and Image Diffusion Models + +![](images/aa1694faa01481c5c6a00468d7bfb09ff0a914d584a65416ee9bddad38c0d8f3.jpg) +Ronghuan Wu +City University of Hong Kong +rh.wu@my.cityu.edu.hk +Wanchao Su Monash University wanchao.su@monash.edu +Jing Liao* +City University of Hong Kong +jingliao@cityu.edu.hk +Figure 1. SVG examples generated by our Chat2SVG. We highlight some shapes to demonstrate semantic clarity and path quality. + +# Abstract + +Scalable Vector Graphics (SVG) has become the de facto standard for vector graphics in digital design, offering resolution independence and precise control over individual elements. Despite their advantages, creating high-quality SVG content remains challenging, as it demands technical expertise with professional editing software and a considerable time investment to craft complex shapes. Recent text-to-SVG generation methods aim to make vector graphics creation more accessible, but they still encounter limitations in shape regularity, generalization ability, and expressiveness. To address these challenges, we introduce Chat2SVG, a hybrid framework that combines the strengths of Large Language Models (LLMs) and image diffusion models for text-to-SVG generation. Our approach first uses an LLM to generate semantically meaningful SVG templates from basic geometric primitives. Guided by image diffusion models, a dual-stage optimization pipeline refines paths in latent space and adjusts point coordinates to enhance geometric complexity. Extensive experiments show + +that Chat2SVG outperforms existing methods in visual fidelity, path regularity, and semantic alignment. Additionally, our system enables intuitive editing through natural language instructions, making professional vector graphics creation accessible to all users. Our code is available at https://chat2svg.github.io/. + +# 1. Introduction + +Scalable Vector Graphics (SVG), a vector image format based on geometric shapes, has become the standard for modern digital design, offering resolution independence and precise control over individual elements. However, creating high-quality SVG content remains challenging for non-expert users, as it requires both expertise with professional design software and considerable time investment to create complex shapes. To make vector graphics creation more accessible, recent research has focused on developing text-to-SVG systems that enable users to express their creative ideas through simple text prompts rather than complex manual editing. + +Existing approaches have explored image-based meth + +ods for text-to-SVG generation, iteratively optimizing a large collection of shape elements (e.g., cubic Bezier curves, typically 100 to 1000) by rendering them into images with differentiable rasterizers [20] and evaluating them using text-image similarity metrics like CLIP loss [28] and Score Distillation Sampling (SDS) loss [27]. While these image-based methods [9, 16, 43, 44, 50] can generate visually impressive SVG through the combination of numerous paths and powerful image models, they face a critical limitation in maintaining the regularity and semantics of shapes. Specifically, semantic components that should be represented as single elements often end up fragmented across multiple overlapping paths. Although these fragmented paths, when viewed collectively, can yield visually appealing outputs, they fundamentally conflict with professional design principles, where each semantic component is intentionally crafted as a single, regularized path. + +Given SVG is defined using Extensible Markup Language (XML), another solution to text-to-SVG synthesis involves language-based methods, which treat SVG scripts as text input. Recent works [35, 41] have proposed specialized tokenization strategies to embed SVG scripts and trained Sequence-To-Sequence (seq2seq) models on domain-specific datasets (e.g., icons and fonts). While these approaches achieve good generation quality within their training domains, they suffer from limited generalization due to the absence of large-scale, general-purpose text-SVG training data. Meanwhile, some works [26, 31] show that while state-of-the-art Large Language Models (LLMs) can generate basic geometric shapes (e.g., circles and rectangles) with layouts matching text prompts, they struggle to produce the complex geometric details required for professional SVG applications. Consequently, the limited generalization and poor expressiveness of language-based methods hinder their adoption for text-to-SVG generation. + +To address the aforementioned problems, we propose a novel hybrid framework that leverages the complementary strengths of image-based and language-based methods. Our approach first utilizes an LLM to synthesize SVG templates composed of basic primitives and then optimizes their geometric details guided by image diffusion models. This hybrid framework addresses the key limitations of both paradigms: The LLM-based template generation overcomes the domain-specific generalization constraints of traditional language models while ensuring shape regularity, as each element naturally corresponds to a single semantic component. The subsequent image-based optimization then enhances the expressiveness of these well-structured templates by capturing complex geometric details that LLMs alone struggle to produce. Specifically, to fully exploit LLMs' ability to create reasonable and complex layouts that match text prompts, we design an SVG-oriented prompt with multiple stages, including prompt ex + +pansion, SVG script generation, and visual refinement. To refine the generated SVG templates, we implement a dual-stage optimization strategy: (1) We first use an image diffusion model to synthesize detailed image-level targets, then follow [50] to optimize path-level latent vectors with a pretrained SVG VAE, eliminating common issues such as self-intersections and jagged curves; (2) We further refine geometric details by directly optimizing the point coordinates to capture fine-grained visual elements. Comprehensive experiments show that Chat2SVG consistently outperforms existing methods regarding overall visual quality, individual path regularity, and semantic coherence. Furthermore, our system enables intuitive editing through iterative natural language instructions, making vector graphics creation more accessible to non-expert users. Our contributions are summarized as follows: + +- Hybrid SVG Generation Framework. We introduce a novel text-to-SVG generation paradigm that combines Large Language Models with image diffusion models to produce high-quality SVG outputs. +- SVG-Oriented Prompt Design. We develop a specialized prompt system that directs LLMs to generate SVG templates using basic geometric primitives. +- Dual-Stage Optimization. We implement a two-phase optimization process that preserves the semantic meaning of each shape while eliminating artifacts such as self-intersecting and jagged paths. +- Iterative Editing. We enable iterative refinement of SVG through natural language instructions, making SVG creation more accessible. + +# 2. Related Work + +# 2.1. Large Language Models for Design + +LLMs have emerged as powerful tools for graphics design tasks, demonstrating remarkable capabilities in understanding and reasoning about complex design specifications across diverse domains. Recent works have explored their applications in (1) visual layout and composition, including poster designs [5, 21, 34, 47], 3D scene arrangements [2, 10, 14, 32, 33, 38, 51], and placement of basic shapes for image generation [8, 46]; (2) content creation and manipulation for shapes [11], materials [15], and animation [22, 36]; and (3) design understanding [18, 42]. While these demonstrate LLMs' versatility, their potential for vector graphics remains unexplored. Our work addresses this gap by showing that LLMs can effectively generate structured SVG templates, extending their capabilities to the domain of vector graphics creation. + +# 2.2. Text-Guided Vector Graphics Generation + +Text-to-SVG generation approaches can be roughly categorized into image-based and language-based methods. + +Image-based methods [9, 16, 43, 44, 50] start with a large collection of randomly initialized shapes, render vector graphics with a differentiable renderer [20], and iteratively optimize path parameters by minimizing text-image similarity losses (e.g., CLIP [28] loss and Score Distillation Sampling [27] loss). Despite their visually pleasant appearance, these methods often generate SVG results containing multiple fragmented paths that lack individual semantic correspondence, which is undesirable for real design scenarios. Language-based methods [3, 35, 41] design specialized tokenization strategies to encode vector graphics into discrete tokens, which are then concatenated with text tokens, and train seq2seq models like autoregressive transformers [37] on domain-specific datasets (e.g., icons and fonts). Although these approaches are conceptually elegant, their generalizability is limited by the scarcity of large-scale, general-purpose vector graphics datasets. Some works [26, 31] have explored LLMs' vector graphics generation capabilities, showing that while LLMs can create reasonable layouts matching text prompts, they struggle to produce complex geometric shapes. Our Chat2SVG combines image-based and language-based methods to address the limitations of each paradigm, producing semantically meaningful SVG where each path corresponds to a distinct visual element and contains refined geometric details. + +# 2.3. Vector Graphics Representation Learning + +Vector graphics representation learning is essential for sketch-based retrieval, reconstruction, and generation tasks. A seminal study, SketchRNN [12], combines a Recurrent Neural Network (RNN) and a VAE to learn vector representations. Lopes et al. [24] trained a VAE to represent image-level font styles, then passed the latent embedding into a decoder for vector font generation. To increase reconstruction quality, later approaches like Sketchformer [30] and DeepSVG [4] adopted Transformer [37] architectures, followed by dual-modality methods [23, 39, 40] that leveraged both vector and image features. However, encoding the entire complex SVG into a single latent embedding often causes detail loss. To address this limitation, Zhang et al. [50] proposed path-level representational learning, creating a smooth latent space for individual paths. We leverage these path-level latent embeddings in our optimization to eliminate self-intersecting and jagged paths. + +# 3. Method + +Our Chat2SVG, as shown in Figure 2, begins with an SVG-oriented prompting approach (Section 3.1) that guides an LLM to create reasonable SVG templates. We then perform a dual-stage optimization (Section 3.2) to enhance the geometric details of SVG templates by optimizing SVG paths in both latent and control point space, guided by image diffusion models. + +# 3.1. SVG Template Generation with LLMs + +Prompt Expansion. Users often provide vague and brief prompts to the LLM when describing their desired graphics. Such unstructured input may lead to oversimplified SVG (regarding both the number of elements and structures) that fails to reflect the intended design. Consequently, we propose a three-level prompt expansion strategy that systematically delineates the initial prompts: (1) Scene-level: Starting with an initial prompt, we instruct the LLM to analyze it holistically and identify essential objects that should appear. The LLM then expands the prompt by suggesting complementary objects to enhance scene completeness. (2) Object-level: For each object in the expanded prompt, we guide the LLM to systematically break it down into its components. For instance, when describing a lion, the LLM deconstructs it into distinct parts such as the body, head, mane, legs, eyes, ears, and tail. This decomposition ensures that no critical components are missing during SVG script generation. (3) Layout-level: After object decomposition, we direct the LLM to develop a comprehensive layout plan. This includes determining the position and size of each element on the canvas, selecting appropriate colors for visual harmony, and establishing clear spatial relationships to ensure a cohesive overall composition. + +SVG Script Generation. After prompt expansion, we obtain a detailed scene description. We then convert this natural language specification into SVG scripts in XML format. Since the LLM has limitations in synthesizing geometrically complex paths, we constrain the shapes to a carefully selected set of basic primitives: rectangles, ellipses, lines, polylines, polygons, and short paths. We set the canvas size to $512 \times 512$ . Each geometric element is assigned a unique ID and includes semantic annotations in its comments. + +Visual Rectification. Since prompt expansion is purely text-based, even when the generated SVG script accurately follows the detailed prompt, visual inconsistencies (e.g., misaligned components, disproportionate scaling, and incorrect path ordering) can emerge during rendering. For example, in Figure 2, the initial SVG template shows a lion with a misshapen mane and missing facial features. We thus adopt a visual rectification strategy in which we render the SVG and provide the rendered image back to the LLM (with vision capability) for inspection. The LLM analyzes the visual output, identifies any inconsistencies or oddities, and generates corrected SVG code. This visual refinement loop can be performed iteratively. In our experiments, we find that two iterations of refinement are typically sufficient to generate well-structured SVG templates with appropriate spatial layouts. + +Furthermore, to enhance the quality of prompt expansion and SVG script generation, we provide curated in-context examples in the prompts. The complete set of prompts and examples is available in the supplementary material. + +![](images/fe924cf106c9beec49f8582b30b7f20d22c68f6d792b6640cf4a576f58ece790.jpg) + +![](images/f9134506dd5b3e81b330b3ffdd59872c440a6c8ebe9c1fed23ce7918399d1308.jpg) +Figure 2. The system pipeline of Chat2SVG. Given a text prompt, our system first leverages an LLM to generate an SVG template composed of basic geometric primitives. The rendered template is enhanced through SDEdit [25] with ControlNet [49] to add visual details while preserving the overall composition, yielding a target image. The SVG then undergoes a dual-stage optimization process to match the target image. (1) Primitives are converted to latent embeddings through latent inversion and optimized along with their visual attributes (i.e., filling colors $c_{i}$ , stroke properties $s_{i}$ , and transformation matrices $T_{i}$ ). (2) Point-level optimization is performed to refine the geometric details of SVG paths. + +![](images/80317083313508ff8a164d592182c2dde4f4cbe8a5589373af247225e0bb9622.jpg) + +# 3.2. SVG Optimization Guided by Image Diffusion + +Detail Enhancement. While the primitive shapes in the SVG template have accurate semantic meaning, they lack the geometric details necessary for a professional design. To improve visual expressiveness, we first render the SVG template into a raster image, and then use an image editing method to generate a more detailed version that serves as our optimization target. Specifically, we employ SDEdit [25], which enhances the input image by first adding noise and then progressively denoising it with an image diffusion model to produce outputs with richer details, such as the contour of the lion's mane illustrated in Figure 2. To maintain structural similarity between the enhanced output and the original image, we incorporate a ControlNet (tile version) [49] into our pipeline. By using Gaussian-blurred versions of the initial images as control signals, the ControlNet effectively maintains the overall composition throughout the enhancement process. + +Apart from geometric details, this image-to-image translation process also introduces new semantic parts such as the lion's beard. To faithfully reproduce the target image, we propose using the Segment Anything Model (SAM) [17] to identify these new decorative details. Specifically, we denote the rendered images of the initial SVG template and its diffusion-enhanced counterpart as $I_{\mathrm{template}}$ and $I_{\mathrm{target}}$ , respectively. We apply SAM to extract mask sets from both images: $\{m_i\}_{i=1}^Q$ for the template image and $\{m_j\}_{j=1}^K$ for the target image, where $Q$ and $K$ represent the number of masks in each collection. Our approach is based on the ob + +servation that the diffusion process simultaneously generates fine decorative details that are absent in the original template (e.g., the beard of the lion in Figure 2) while preserving the structural integrity of larger components (such as the body). Consequently, we implement an area-based mask-filtering strategy to identify these newly generated elements. For each mask $m_j$ detected in $I_{\mathrm{target}}$ , we evaluate it against all masks in $\{m_i\}_{i=1}^Q$ using Intersection over Union (IoU) metrics [29]. When $m_j$ overlaps with a template mask $m_i$ but their IoU falls below a threshold, we classify $m_j$ as a new decorative component. Alternatively, if their IoU exceeds the threshold, we interpret $m_j$ as a variation of an existing component. For all filtered masks classified as new decorative shapes, we approximate their boundaries with polygons and integrate them into the SVG script. + +We define an SVG script $G$ as a collection of $M$ paths (i.e., shapes), $G = \{P_i\}_{i=1}^M$ . Each path $P_i$ consists of a sequence of $N_i$ commands, $P_i = \{C_i^j\}_{j=1}^{N_i}$ . A command $C_i^j = (U_i^j, V_i^j)$ is defined by its type $U_i^j \in \{\mathbb{M}, \mathbb{C}\}$ and its associated control points $V_i^j$ , where $\mathbb{M}$ represents Move and $\mathbb{C}$ represents cubic Bézier curves. For consistency, we convert all other primitive shapes (e.g., rectangles and ellipses) in the SVG template into cubic Bézier curves. + +Previous image-based methods [9, 16, 43, 44] optimize SVG shapes by directly manipulating the control points $V_{i}^{j}$ . Despite convenience, this approach often leads to artifacts such as self-intersecting curves and unnatural deformations. To overcome these limitations, Zhang et al. [50] introduced a path-level SVG VAE. This model's latent space effec + +tively captures common shape patterns and geometric constraints, allowing for the optimization of path latent vectors at a higher level to produce smooth outputs. + +Latent Optimization. We leverage this pretrained SVG VAE to conduct the latent optimization. However, their SVG encoder expects a fixed number of commands (10 cubic Bezier curves), while our SVG primitives contain varying (typically fewer) commands. To resolve this mismatch, we develop a latent inversion process that converts our primitives into latent embeddings. We start with a randomly sampled latent vector $z$ and decode it into a command sequence. Then we evenly sample points $X = \{x_{i}\}_{i=1}^{N}$ along its contour. Similarly, we sample an equal number of points $Y = \{y_{j}\}_{j=1}^{N}$ along the contour of our target primitive shape. Using these point sets, we compute the Earth Mover's Distance (EMD) [1, 7, 48]: + +$$ +\ell_ {\mathrm {E M D}} (X, Y) = \min _ {\phi : X \rightarrow Y} \sum_ {x \in X} \| x - \phi (x) \| _ {2}, \tag {1} +$$ + +where $\phi$ represents a bijective mapping between the point sets. By minimizing this EMD loss, we gradually optimize the latent embedding $z$ to match the target primitive. + +After obtaining latent vectors $\{z_i\}_{i=1}^M$ for all paths, we jointly optimize these vectors along with visual attribute parameters: filling colors $\{c_i\}_{i=1}^M$ , stroke properties (color and width) $\{s_i\}_{i=1}^M$ , and transformation matrices $\{T_i\}_{i=1}^M$ . At each optimization iteration $t$ , we decode the latent vectors into command sequences and apply transformation matrices to the control points. We then combine color and stroke attributes together, and use a differentiable rasterizer [20] to render the complete SVG into an image $I_t$ . Our optimization objective consists of three loss terms: (1) An MSE loss $\ell_{\mathrm{MSE}}$ between the rendered image $I_t$ and target image $I_{\mathrm{target}}$ to ensure visual similarity; (2) A curvature loss that reduces sharp bends and fluctuations along the contour by computing the discrete second derivative: + +$$ +\ell_ {\text {c u r v a t u r e}} = \frac {\sum_ {i = 1} ^ {M} \sum_ {j = 1} ^ {N _ {i} - 2} \left\| V _ {i} ^ {j} - 2 V _ {i} ^ {j + 1} + V _ {i} ^ {j + 2} \right\| _ {2} ^ {2}}{\sum_ {i = 1} ^ {M} \left(N _ {i} - 2\right)}; \tag {2} +$$ + +(3) A path-level IoU loss that preserves the overall composition and prevents paths from drifting too far from their initial positions: + +$$ +\ell_ {\mathrm {I o U}} = \frac {1}{M} \sum_ {i = 1} ^ {M} \left(1 - \frac {\left| m _ {i} ^ {t} \cap m _ {i} ^ {0} \right|}{\left| m _ {i} ^ {t} \cup m _ {i} ^ {0} \right|}\right), \tag {3} +$$ + +where $m_i^t$ and $m_i^0$ denote the binary masks of the $i$ -th path at the current iteration $t$ and its initial state, respectively. + +The final loss function combines these three terms with empirically determined weights: + +$$ +\ell_ {\text {l a t e n t}} = \ell_ {\mathrm {M S E}} + \lambda_ {1} \ell_ {\text {c u r v a t u r e}} + \lambda_ {2} \ell_ {\mathrm {I o U}}, \tag {4} +$$ + +where $\lambda_1 = 5e - 4$ and $\lambda_{2} = 5e - 6$ + +Point Optimization. While latent optimization effectively aligns SVG shapes with their target positions and contours, it has a notable limitation: the resulting shapes tend to be overly smooth, lacking the intricate details that are often essential in professional vector graphics. For example, in Figure 2, the lion's mane after latent optimization does not capture the fine contour present in the target image. Therefore, we add a second optimization stage that directly refines the control points $V_{i}^{j}$ of the paths, similar to previous image-based works [9, 16, 43, 44]. To achieve finer granularity in shape control, we split each cubic Bezier curve at its midpoint, effectively doubling the number of control points in the SVG. This increased control point density allows for more precise shape adjustments. The point optimization stage employs a loss function that combines the MSE loss and the curvature loss: + +$$ +\ell_ {\text {p o i n t}} = \ell_ {\mathrm {M S E}} + \lambda_ {3} \ell_ {\text {c u r v a t u r e}}. \tag {5} +$$ + +The parameter $\lambda_3$ decreases linearly from $1e - 3$ to $5e - 5$ throughout the optimization process. + +# 3.3. Iterative Editing + +During SVG script generation (Section 3.1), we instruct the LLM to annotate the semantic label of each path. This design enables users to refine the initial SVG template using natural language, as the LLM can precisely locate and modify specific elements based on semantic understanding. + +Our system supports iterative refinement through multiple rounds of natural language instructions. However, when passing the edited SVG template image into SDEdit, image diffusion models may introduce different shape variations and decorative details, which compromise the user's intention to maintain the already optimized shapes. To preserve consistency across editing iterations, we instruct the LLM to output both the modified SVG script and a list of specifically changed paths. This precise tracking enables selective optimization of modified shapes while preserving unaltered elements from previous iterations, maintaining visual coherence throughout the entire refinement process. + +# 4. Experiments + +Implementation Details. In our experiments, we use claude-3-5-sonnet as our backbone LLM model due to its leading generation capabilities. To evaluate our approach, we create a prompt set by having the LLM generate 125 text prompts across five categories: animals, food, objects, scenes, and novel concept combinations. For each prompt, we generate one template and perform 2 rounds of visual rectification. This process is repeated 5 times, yielding $(1 + 2) \times 5 = 15$ candidate SVG templates per prompt. These templates are rendered as images, and we follow a + +![](images/cd01b7e83cd888ef4faaa359d8e3600ab0db1b8221bee6904932994e764ff4a1.jpg) +Figure 3. Qualitative Comparison. (1) Methods refining open-ended strokes, i.e., CLIPDraw [9] and DiffSketcher [43], often produce distorted and disorganized strokes to approximate objects, presenting messy appearance and poor text alignment. (2) VectorFusion [16] and SVGDreamer [44] produce elements that consist of multiple jagged, irregular, and fragmented shapes, such as the body of the flamingo (first row) and the plane (third row). (3) T2V-NPR [50] attempts to resolve these issues by learning a latent representation of paths and merging fragmented shapes. However, it still cannot guarantee the semantic meanings of the paths, leading to less-semantic paths such as a plane body with surrounding clouds in the third row. In contrast, our method produces SVG with superior text alignment, higher visual quality, and well-structured paths exhibiting geometric regularity and clear semantic definition. + +standard practice to select the highest quality SVG using the ImageReward [45] metric. Detailed optimization configurations are provided in the supplementary materials. + +Baselines. We compare our approach against five state-of-the-art image-based text-to-SVG generation methods: CLIPDraw [9], DiffSketcher [43], VectorFusion [16], SVGDreamer [44], and T2V-NPR [50]. Current language-based methods [35, 41] are confined to specific categories, with no support for general text-to-SVG generation, so we do not include them in our comparison. For a fair comparison, we set these baseline methods to use a similar number of shapes as our optimized SVG. + +Evaluation Metrics. We evaluate the quality of our gen + +erated SVG across three dimensions: visual fidelity, path regularity, and semantic alignment. + +- Image-level Fidelity. We evaluate visual fidelity using a ground-truth dataset of well-designed vector graphics, specifically 52,805 colored SVG files downloaded from SVGRepo1. We compute the Fréchet Inception Distance (FID) [13] between the rendered images of our generated SVG and this professional design collection, using features extracted by the CLIP image encoder [28]. +- Path-level Regularity. We assess path quality using a transformer-based Auto-Encoder trained on the FIGR-8-SVG [6] dataset, following the DeepSVG [4] architecture + +
MethodImage FID ↓Path FID ↓Text Alignment ↑
CLIPDraw [9]46.7770.130.3048
DiffSketcher [43]44.8966.480.2623
VectorFusion [16]39.5256.790.2982
SVGDreamer [44]35.4847.950.2919
T2V-NPR [50]39.8642.030.3078
Chat2SVG (Ours)33.3139.070.3090
+ +Table 1. Quantitative comparison of text-to-SVG generation methods across image fidelity, vector regularity, and text alignment. + +ture. Our model, trained with reconstruction tasks, encodes each drawing command $C_i^j$ into a latent vector. We represent each path as the mean of its commands' latent embeddings, and calculate the FID between these path representations and the ground truth paths from FIGR-8-SVG. This metric quantifies how closely our generated paths align with professional SVG design patterns. + +- Text-level Alignment. We assess semantic alignment by computing the CLIP score [28] between the text prompt and the rendered SVG image. + +# 4.1. Comparison with Existing Methods + +Quantitative Comparison. Table 1 shows the evaluation metrics for all baseline methods. Our method achieves the best image FID score, indicating that our generated SVG closely aligns with professional design patterns. Regarding path regularity, our method yields paths that are closest to the professionally designed SVG, as evidenced by the lowest path FID. Furthermore, our Chat2SVG achieves the highest score in text-SVG alignment, validating the significance of LLM-generated SVG templates. + +Qualitative Comparison. In Figure 3, we present a side-by-side comparison between our Chat2SVG and the baselines. Methods that optimize open-ended strokes based on CLIP or Diffusion models (i.e., CLIPDraw [9] and DiffS-ketcher [43]) produce results with messy visual appearances and poor text alignment. These methods sometimes fail to synthesize complete objects, such as the flamingo (first row) and the plane (third row) in Figure 3. In contrast, our method generates SVG results with clear layouts and greater fidelity to the prompts. + +Methods that optimize closed shapes via score distillation of diffusion models (i.e., VectorFusion [16], SVGDreamer [44], and T2V-NPR [50]) can better align with the input text, but still produce fragmented paths with limited semantic meaning. For instance, in the plane example (third row of Figure 3), VectorFusion and SVGDreamer create planes and clouds composed of multiple jagged, irregular, and fragmented shapes, which only appear meaningful when viewed as a whole. This hinders convenient editing by graphic designers. T2V-NPR [50] addresses the issue of jagged paths using a path VAE (although it can be + +![](images/31a541393ca3e17bc2f81e2e8155d6fed320563e6c26162a7339b00768ddd09d.jpg) +Figure 4. User Study. Our Chat2SVG achieves the highest user selection ratio across all three evaluation criteria. + +overly smooth at times) and reduces fragmented shapes by merging shapes with similar colors. However, its merging operation ignores the semantic meaning of shapes, resulting in the body of the plane being merged with the surrounding clouds. Additionally, the semantic meaning of the path is sometimes unclear, as exemplified by the cloud in the bottom-right corner, whose contour is actually approximated by the blue background instead of a true cloud shape. In contrast, our Chat2SVG adopts a unique approach by using LLMs to generate paths representing semantic components and performing dual-stage optimization based on image diffusion to enhance path expressiveness, thus ensuring both path regularity and visual fidelity of generated SVG. + +User Study. To further compare Chat2SVG with baseline methods, we conducted a user study examining the same key aspects: (1) overall visual aesthetics, (2) path regularity, and (3) text-SVG alignment. We randomly sampled 20 prompts and generated SVG outputs using our approach and five baseline methods. For each prompt, participants were shown all generated SVG results along with the corresponding text prompt and highlighted paths. They were then asked to select the highest quality result regarding the three evaluation criteria. We recruited 31 participants (18 male, 13 female) through university mailing lists, with a mean age of 26 years. Among them, 17 participants reported prior experience in graphic design, providing a balanced mix of expert and novice perspectives. Analyzing the average selection ratio across all three metrics (Figure 4), we found that Chat2SVG-generated results were consistently preferred by participants over baseline methods. + +# 4.2. Ablation Study + +We conduct ablation experiments to evaluate the effectiveness of key components in our pipeline (Section 3.1). First, we remove the SVG template generation and use randomly initialized shapes to approximate target images. As shown in the second row of Table 2, both image FID and path FID + +
Chat2SVG VariantImage FID ↓Path FID ↓Text Alignment ↑
Default33.3139.070.3090
No SVG Template47.2170.130.2968
No SVG Optimization33.4536.120.3044
+ +Table 2. Qualitative results of the ablation study. + +![](images/88128b6fc59cd146790d4597fa859c088b39f479d7e423e29cb6749cd828d3b9.jpg) +Figure 5. Qualitative results of ablation study. + +values increase substantially. This performance degradation can be attributed to the substantial gap between randomly initialized shapes and target shapes, making optimization more challenging. Meanwhile, in the second column of Figure 5, the dog's legs and ears lack semantic-clear components, highlighting the importance of SVG template generation. Second, we eliminate the dual-stage optimization. As shown in the third row of Table 2, this results in the lowest path FID, as primitives generated by LLMs naturally align with the path regularity presented in the ground-truth dataset. However, the visual outcome in Figure 5 lacks the necessary details, indicating that the overall visual quality suffers without dual-stage optimization. A more detailed ablation study is provided in the supplementary material. + +# 4.3. Iterative Editing Results + +In this section, we present the iterative editing results. As demonstrated in Figure 6, users can progressively refine the SVG through multiple rounds of natural language instructions. During editing, we preserve the optimized shapes from previous iterations while only modifying the specified shapes. The results showcase how Chat2SVG can accurately interpret and execute user editing requests, maintaining both semantic coherence and visual quality throughout the modifications. Additional editing examples are provided in the supplementary material. + +# 5. Conclusion + +In this paper, we present Chat2SVG, a novel paradigm for text-to-SVG generation that combines LLMs and image diffusion models. Through our carefully designed SVG template generation and dual-stage SVG optimization pipeline, our method generates high-quality SVG outputs that exhibit strong visual fidelity, path regularity, and text alignment. + +Despite our method's superior results, there are several + +![](images/62cd837d2b193eecf7c28ab864031b4f738964edd919c617dec2e4e805051e96.jpg) +Figure 6. Iterative Editing. We perform two rounds of refinement on each SVG template and show the optimized output. This figure shows editing types including deletion, modification, and addition. + +(1) Layout Precision + +![](images/f7ea2084ffb93095b4075d8ab0d9fb1bd80874dbcee42fe466a7dcc3b0f709e9.jpg) + +![](images/20e7459baa524fb618e1a24a76b23f9ed3bf25bd6115da62d5709794501aae16.jpg) +SVG Target Template Image +Optimized SVG + +(2) Path Addition + +![](images/18979784f96b2fb99b328c336ee0d238a340c6532366dc223f1e496b0d2de66a.jpg) + +![](images/d6b078a9667f664bd408f16b9507351cbb3729016cd93fc7b282da93fd66253f.jpg) +SVG Target Template Image +Optimized SVG + +(3) Shape Complexity + +![](images/fb5f5c36a016363d6cf9d58799a2c8df46c0fb0803a737a7cf4920860c2da95a.jpg) +SVG Template + +![](images/635ee330f3881215cb11e15b074e60f1cca7990a71f9773ccb73424e5c8ba303.jpg) +Target Image + +![](images/8a028051f52b1a74157cfd3cf71eab3726afcc148799ce13b06765b8ff914ea6.jpg) +Optimized SVG +Figure 7. Limitations of our method include imprecise layout generation, missing important visual elements, and insufficient shape complexity. + +aspects that could be further improved: (1) LLMs may produce imprecise visual layouts, such as the burger in Figure 7. This can be alleviated by collecting an SVG dataset and extracting the corresponding layouts to fine-tune LLMs. (2) Our approach uses SAM to identify new parts. However, SAM may overlook important regions, like the pink frosting on the donut in Figure 7. Incorporating semantic-aware SAM [19] could potentially produce more accurate results. (3) The pretrained SVG VAE from [50] has a fixed number of commands. This makes it hard to represent intricate shapes. For example, the lion's mane in Figure 7 lacks sufficient detail. Training a new SVG VAE with varying-length commands would help solve this problem. Our work can be extended to generate other types of vectorized data, such as icons and fonts. Additionally, through consistent editing of SVG across multiple frames, coupled with shape interpolation, our Chat2SVG has the potential to facilitate keyframe animation generation. + +# Acknowledgement + +# References + +[1] Panos Achlioptas, Olga Diamanti, Ioannis Mitliagkas, and Leonidas Guibas. Learning representations and generative models for 3D point clouds. In International Conference on Machine Learning, pages 40-49, 2018. 5 +[2] Rio Aguina-Kang, Maxim Gumin, Do Heon Han, Stewart Morris, Seung-Jean Yoo, Aditya Ganeshan, R. Kenny Jones, Qiuhong Anna Wei, Kailiang Fu, and Daniel Ritchie. Open-universe indoor scene generation using LLM program synthesis and uncurated object databases. In arXiv preprint arXiv:2403.09675, 2024. 2 +[3] Jonas Belouadi, Anne Lauscher, and Steffen Eger. AutomataTikZ: Text-guided synthesis of scientific vector graphics with TikZ. In arXiv preprint arXiv:2310.00367, 2023. 3 +[4] Alexandre Carlier, Martin Danelljan, Alexandre Alahi, and Radu Timofte. DeepSVG: A hierarchical generative network for vector graphics animation. In Advances in Neural Information Processing Systems, pages 16351-16361, 2020. 3, 6 +[5] Yutao Cheng, Zhao Zhang, Maoke Yang, Hui Nie, Chunyuan Li, Xinglong Wu, and Jie Shao. Graphic design with large multimodal model. In arXiv preprint arXiv:2404.14368, 2024. 2 +[6] Louis Clouatre and Marc Demers. FIgR: Few-shot image generation with reptile. In arXiv preprint arXiv:1901.02199, 2019. 6 +[7] Haoqiang Fan, Hao Su, and Leonidas Guibas. A point set generation network for 3D object reconstruction from a single image. In IEEE Conference on Computer Vision and Pattern Recognition, pages 605-613, 2017. 5 +[8] Weixi Feng, Wanrong Zhu, Tsu-jui Fu, Varun Jampani, Arjun Akula, Xuehai He, Sugato Basu, Xin Wang, and William Wang. LayoutGPT: Compositional visual planning and generation with large language models. In Advances in Neural Information Processing Systems, pages 18225-18250, 2024. 2 +[9] Kevin Frans, Lisa B. Soros, and Olaf Witkowski. CLIPDraw: Exploring text-to-drawing synthesis through language-image encoders. In Advances in Neural Information Processing Systems, pages 5207-5218, 2021. 2, 3, 4, 5, 6, 7 +[10] Rao Fu, Zehao Wen, Zichen Liu, and Srinath Sridhar. AnyHome: Open-vocabulary generation of structured and textured 3D homes. In European Conference on Computer Vision, pages 52-70, 2025. 2 +[11] Aditya Ganeshan, Ryan Y. Huang, Xianghao Xu, R. Kenny Jones, and Daniel Ritchie. ParSEL: Parameterized shape editing with language. In arXiv preprint arXiv:2405.20319, 2024. 2 + +The work described in this paper was fully supported by a GRF grant from the Research Grants Council (RGC) of the Hong Kong Special Administrative Region, China [Project No. CityU 11216122]. +[12] David Ha and Douglas Eck. A neural representation of sketch drawings. In arXiv preprint arXiv:1704.03477, 2017. 3 +[13] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. GANs trained by a two time-scale update rule converge to a local Nash equilibrium. In Advances in Neural Information Processing Systems, pages 6629-6640, 2017. 6 +[14] Ziniu Hu, Ahmet Iscen, Aashi Jain, Thomas Kipf, Yisong Yue, David A. Ross, Cordelia Schmid, and Alireza Fathi. SceneCraft: An LLM agent for synthesizing 3D scenes as Blender code. In International Conference on Machine Learning, pages 19252-19282, 2024. 2 +[15] Ian Huang, Guandao Yang, and Leonidas Guibas. Blender: Editing 3D graphics with vision-language models. In arXiv preprint arXiv:2404.17672, 2024. 2 +[16] Ajay Jain, Amber Xie, and Pieter Abbeel. VectorFusion: Text-to-SVG by abstracting pixel-based diffusion models. In IEEE Conference on Computer Vision and Pattern Recognition, pages 1911-1920, 2022. 2, 3, 4, 5, 6, 7 +[17] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dólar, and Ross Girshick. Segment anything. In IEEE International Conference on Computer Vision, pages 4015-4026, 2023. 4 +[18] Peter Kulits, Haiwen Feng, Weiyang Liu, Victoria Abrevaya, and Michael J. Black. Re-thinking inverse graphics with large language models. In arXiv preprint arXiv:2404.15228, 2024. 2 +[19] Feng Li, Hao Zhang, Peize Sun, Xueyan Zou, Shilong Liu, Jianwei Yang, Chunyuan Li, Lei Zhang, and Jianfeng Gao. Semantic-SAM: Segment and recognize anything at any granularity. In arXiv preprint arXiv:2307.04767, 2023. 8 +[20] Tzu-Mao Li, Michal Lukáč, Michael Gharbi, and RaganKelley Jonathan. Differentiable vector graphics rasterization for editing and learning. ACM Transactions on Graphics, 39 (6):1-15, 2020. 2, 3, 5 +[21] Jiawei Lin, Jiaqi Guo, Shizhao Sun, Zijiang Yang, JianGuang Lou, and Dongmei Zhang. LayoutPrompter: Awaken the design ability of large language models. In Advances in Neural Information Processing Systems, pages 43852-43879, 2024. 2 +[22] Vivian Liu, Rubaiat H. Kazi, Li-Yi Wei, Matthew Fisher, Timothy Langlois, Seth Walker, and Lydia Chilton. LogoMotion: Visually grounded code generation for content-aware animation. In arXiv preprint arXiv:2405.07065, 2024. 2 +[23] YingTian Liu, Zhifei Zhang, YuanChen Guo, Matthew Fisher, Zhaowen Wang, and SongHai Zhang. DualVector: Unsupervised vector font synthesis with dual-part representation. In IEEE Conference on Computer Vision and Pattern Recognition, pages 14193-14202, 2023. 3 +[24] Raphael G. Lopes, David Ha, Douglas Eck, and Jonathon Shlens. A learned representation for scalable vector graphics. In IEEE International Conference on Computer Vision, pages 7930-7939, 2019. 3 + +[25] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia-jun Wu, JunYan Zhu, and Stefano Ermon. SDEdit: Guided image synthesis and editing with stochastic differential equations. In arXiv preprint arXiv:2108.01073, 2021. 4 +[26] OpenAI. GPT-4 technical report. In arXiv preprint arXiv:2303.08774, 2023. 2, 3 +[27] Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. DreamFusion: Text-to-3D using 2D diffusion. In arXiv preprint arXiv:2209.14988, 2022. 2, 3 +[28] Alec Radford, Jong-Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763, 2021. 2, 3, 6, 7 +[29] Hamid Rezatofighi, Nathan Tsoi, JunYoung Gwak, Amir Sadeghian, Ian Reid, and Silvio Savarese. Generalized intersection over union: A metric and a loss for bounding box regression. In IEEE Conference on Computer Vision and Pattern Recognition, pages 658-666, 2019. 4 +[30] Leo S. F. Ribeiro, Tu Bui, John Collomosse, and Moacir Ponti. Sketchformer: Transformer-based representation for sketched structure. In IEEE Conference on Computer Vision and Pattern Recognition, pages 14153-14162, 2020. 3 +[31] Pratyusha Sharma, Tamar Rott Shaham, Manel Baradad, Stephanie Fu, Adrian Rodriguez-Munoz, Shivam Duggal, Phillip Isola, and Antonio Torralba. A vision check-up for language models. In IEEE Conference on Computer Vision and Pattern Recognition, pages 14410-14419, 2024. 2, 3 +[32] Chunyi Sun, Junlin Han, Weijian Deng, Xinlong Wang, Zishan Qin, and Stephen Gould. 3D-GPT: Procedural 3D modeling with large language models. In arXiv preprint arXiv:2310.12945, 2023. 2 +[33] Hou In Ivan Tam, Hou In Derek Pun, Austin T. Wang, Angel X. Chang, and Manolis Savva. SceneMotifCoder: Example-driven visual program learning for generating 3D object arrangements. In arXiv preprint arXiv:2408.02211, 2024. 2 +[34] Zecheng Tang, Chenfei Wu, Juntao Li, and Nan Duan. LayoutNUWA: Revealing the hidden layout expertise of large language models. In arXiv preprint arXiv:2309.09506, 2023. 2 +[35] Zecheng Tang, Chenfei Wu, Zekai Zhang, Mingheng Ni, Shengming Yin, Yu Liu, Zhengyuan Yang, Lijuan Wang, Zicheng Liu, Juntao Li, and Nan Duan. StrokeNUWA: Tokenizing strokes for vector graphic synthesis. In arXiv preprint arXiv:2401.17093, 2024. 2, 3, 6 +[36] Tiffany Tseng, Ruijia Cheng, and Jeffrey Nichols. Keyframer: Empowering animation design using large language models. In arXiv preprint arXiv:2402.06071, 2024. 2 +[37] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems, pages 6000–6010, 2017. 3 + +[38] Can Wang, Hongliang Zhong, Menglei Chai, Mingming He, Dongdong Chen, and Jing Liao. Chat2Layout: Interactive 3D furniture layout with a multimodal LLM. In arXiv preprint arXiv:2407.21333, 2024. 2 +[39] Yizhi Wang and Zhouhui Lian. DeepVecFont: Synthesizing high-quality vector fonts via dual-modality learning. ACM Transactions on Graphics, 40(6):1-15, 2021. 3 +[40] Yuqing Wang, Yizhi Wang, Longhui Yu, Yuesheng Zhu, and Zhouhui Lian. DeepVecFont-v2: Exploiting Transformers to synthesize vector fonts with higher quality. In IEEE Conference on Computer Vision and Pattern Recognition, pages 18320-18328, 2023. 3 +[41] Ronghuan Wu, Wanchao Su, Kede Ma, and Jing Liao. IconShop: Text-guided vector icon synthesis with autoregressive Transformers. ACM Transactions on Graphics, 42(6):1-14, 2023. 2, 3, 6 +[42] Tong Wu, Guandao Yang, Zhibing Li, Kai Zhang, Ziwei Liu, Leonidas Guibas, Dahua Lin, and Gordon Wetzstein. GPT-4V (ision) is a human-aligned evaluator for text-to-3D generation. In IEEE Conference on Computer Vision and Pattern Recognition, pages 22227–22238, 2024. 2 +[43] Ximing Xing, Chuang Wang, Haitao Zhou, Jing Zhang, Qian Yu, and Dong Xu. DiffSketcher: Text guided vector sketch synthesis through latent diffusion models. In Advances in Neural Information Processing Systems, pages 15869-15889, 2023. 2, 3, 4, 5, 6, 7 +[44] Ximing Xing, Haitao Zhou, Chuang Wang, Jing Zhang, Dong Xu, and Qian Yu. SVGDreamer: Text guided SVG generation with diffusion model. In IEEE Conference on Computer Vision and Pattern Recognition, pages 4546-4555, 2024. 2, 3, 4, 5, 6, 7 +[45] Jiazheng Xu, Xiao Liu, Yuchen Wu, Yuxuan Tong, Qinkai Li, Ming Ding, Jie Tang, and Yuxiao Dong. ImageReward: Learning and evaluating human preferences for text-to-image generation. In Advances in Neural Information Processing Systems, pages 15903-15935, 2024. 6 +[46] Yutaro Yamada, Khyathi Chandu, Yuchen Lin, Jack Hessel, Ilker Yildirim, and Yejin Choi. L3GO: Language agents with chain-of-3D-thoughts for generating unconventional objects. In arXiv preprint arXiv:2402.09052, 2024. 2 +[47] Tao Yang, Yingmin Luo, Zhongang Qi, Yang Wu, Ying Shan, and Chang Wen Chen. PosterLLaVa: Constructing a unified multi-modal layout generator with LLM. In arXiv preprint arXiv:2406.02884, 2024. 2 +[48] Lequan Yu, Xianzhi Li, Chi-Wing Fu, Daniel Cohen-Or, and Pheng-Ann Heng. PU-Net: Point cloud upsampling network. In IEEE Conference on Computer Vision and Pattern Recognition, pages 2790–2799, 2018. 5 +[49] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In IEEE International Conference on Computer Vision, pages 3836-3847, 2023. 4 +[50] Peiying Zhang, Nanxuan Zhao, and Jing Liao. Text-to-vector generation with neural path representation. ACM Transactions on Graphics, 43(4):1-13, 2024. 2, 3, 4, 6, 7, 8 +[51] Qihang Zhang, Chaoyang Wang, Aliaksandr Siarohin, Peiye Zhuang, Yinghao Xu, Ceyuan Yang, Dahua Lin, Bolei Zhou, + +Sergey Tulyakov, and Hsin-Ying Lee. SceneWiz3D: Towards text-guided 3D scene composition. In arXiv preprint arXiv:2312.08885, 2023. 2 \ No newline at end of file diff --git a/2025/Chat2SVG_ Vector Graphics Generation with Large Language Models and Image Diffusion Models/images.zip b/2025/Chat2SVG_ Vector Graphics Generation with Large Language Models and Image Diffusion Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..138b79bcbb06e49875047502ef7f68200417364f --- /dev/null +++ b/2025/Chat2SVG_ Vector Graphics Generation with Large Language Models and Image Diffusion Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6982c6446984c450d82d6044a32500d6846332a159f896986b4426f72435077c +size 568847 diff --git a/2025/Chat2SVG_ Vector Graphics Generation with Large Language Models and Image Diffusion Models/layout.json b/2025/Chat2SVG_ Vector Graphics Generation with Large Language Models and Image Diffusion Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..085bca2bba33743a2a64fef0fd39f5be2dd80244 --- /dev/null +++ b/2025/Chat2SVG_ Vector Graphics Generation with Large Language Models and Image Diffusion Models/layout.json @@ -0,0 +1,7716 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 75, + 102, + 536, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 102, + 536, + 140 + ], + "spans": [ + { + "bbox": [ + 75, + 102, + 536, + 140 + ], + "type": "text", + "content": "Chat2SVG: Vector Graphics Generation with Large Language Models and Image Diffusion Models" + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 75, + 208, + 541, + 415 + ], + "blocks": [ + { + "bbox": [ + 75, + 161, + 225, + 204 + ], + "lines": [ + { + "bbox": [ + 75, + 161, + 225, + 204 + ], + "spans": [ + { + "bbox": [ + 75, + 161, + 225, + 204 + ], + "type": "text", + "content": "Ronghuan Wu \nCity University of Hong Kong \nrh.wu@my.cityu.edu.hk" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 247, + 161, + 363, + 203 + ], + "lines": [ + { + "bbox": [ + 247, + 161, + 363, + 203 + ], + "spans": [ + { + "bbox": [ + 247, + 161, + 363, + 203 + ], + "type": "text", + "content": "Wanchao Su Monash University wanchao.su@monash.edu" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 386, + 161, + 536, + 204 + ], + "lines": [ + { + "bbox": [ + 386, + 161, + 536, + 204 + ], + "spans": [ + { + "bbox": [ + 386, + 161, + 536, + 204 + ], + "type": "text", + "content": "Jing Liao* \nCity University of Hong Kong \njingliao@cityu.edu.hk" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 75, + 208, + 541, + 415 + ], + "lines": [ + { + "bbox": [ + 75, + 208, + 541, + 415 + ], + "spans": [ + { + "bbox": [ + 75, + 208, + 541, + 415 + ], + "type": "image", + "image_path": "aa1694faa01481c5c6a00468d7bfb09ff0a914d584a65416ee9bddad38c0d8f3.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 69, + 422, + 540, + 436 + ], + "lines": [ + { + "bbox": [ + 69, + 422, + 540, + 436 + ], + "spans": [ + { + "bbox": [ + 69, + 422, + 540, + 436 + ], + "type": "text", + "content": "Figure 1. SVG examples generated by our Chat2SVG. We highlight some shapes to demonstrate semantic clarity and path quality." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 152, + 438, + 200, + 451 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 438, + 200, + 451 + ], + "spans": [ + { + "bbox": [ + 152, + 438, + 200, + 451 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 464, + 296, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 464, + 296, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 464, + 296, + 693 + ], + "type": "text", + "content": "Scalable Vector Graphics (SVG) has become the de facto standard for vector graphics in digital design, offering resolution independence and precise control over individual elements. Despite their advantages, creating high-quality SVG content remains challenging, as it demands technical expertise with professional editing software and a considerable time investment to craft complex shapes. Recent text-to-SVG generation methods aim to make vector graphics creation more accessible, but they still encounter limitations in shape regularity, generalization ability, and expressiveness. To address these challenges, we introduce Chat2SVG, a hybrid framework that combines the strengths of Large Language Models (LLMs) and image diffusion models for text-to-SVG generation. Our approach first uses an LLM to generate semantically meaningful SVG templates from basic geometric primitives. Guided by image diffusion models, a dual-stage optimization pipeline refines paths in latent space and adjusts point coordinates to enhance geometric complexity. Extensive experiments show" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 440, + 555, + 512 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 440, + 555, + 512 + ], + "spans": [ + { + "bbox": [ + 313, + 440, + 555, + 512 + ], + "type": "text", + "content": "that Chat2SVG outperforms existing methods in visual fidelity, path regularity, and semantic alignment. Additionally, our system enables intuitive editing through natural language instructions, making professional vector graphics creation accessible to all users. Our code is available at https://chat2svg.github.io/." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 536, + 394, + 548 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 536, + 394, + 548 + ], + "spans": [ + { + "bbox": [ + 314, + 536, + 394, + 548 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 312, + 557, + 555, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 557, + 555, + 700 + ], + "spans": [ + { + "bbox": [ + 312, + 557, + 555, + 700 + ], + "type": "text", + "content": "Scalable Vector Graphics (SVG), a vector image format based on geometric shapes, has become the standard for modern digital design, offering resolution independence and precise control over individual elements. However, creating high-quality SVG content remains challenging for non-expert users, as it requires both expertise with professional design software and considerable time investment to create complex shapes. To make vector graphics creation more accessible, recent research has focused on developing text-to-SVG systems that enable users to express their creative ideas through simple text prompts rather than complex manual editing." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 326, + 701, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 701, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 326, + 701, + 553, + 714 + ], + "type": "text", + "content": "Existing approaches have explored image-based meth" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 703, + 147, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 147, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 147, + 713 + ], + "type": "text", + "content": "*Corresponding Author." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23690" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 276 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 276 + ], + "type": "text", + "content": "ods for text-to-SVG generation, iteratively optimizing a large collection of shape elements (e.g., cubic Bezier curves, typically 100 to 1000) by rendering them into images with differentiable rasterizers [20] and evaluating them using text-image similarity metrics like CLIP loss [28] and Score Distillation Sampling (SDS) loss [27]. While these image-based methods [9, 16, 43, 44, 50] can generate visually impressive SVG through the combination of numerous paths and powerful image models, they face a critical limitation in maintaining the regularity and semantics of shapes. Specifically, semantic components that should be represented as single elements often end up fragmented across multiple overlapping paths. Although these fragmented paths, when viewed collectively, can yield visually appealing outputs, they fundamentally conflict with professional design principles, where each semantic component is intentionally crafted as a single, regularized path." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 279, + 296, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 279, + 296, + 495 + ], + "spans": [ + { + "bbox": [ + 56, + 279, + 296, + 495 + ], + "type": "text", + "content": "Given SVG is defined using Extensible Markup Language (XML), another solution to text-to-SVG synthesis involves language-based methods, which treat SVG scripts as text input. Recent works [35, 41] have proposed specialized tokenization strategies to embed SVG scripts and trained Sequence-To-Sequence (seq2seq) models on domain-specific datasets (e.g., icons and fonts). While these approaches achieve good generation quality within their training domains, they suffer from limited generalization due to the absence of large-scale, general-purpose text-SVG training data. Meanwhile, some works [26, 31] show that while state-of-the-art Large Language Models (LLMs) can generate basic geometric shapes (e.g., circles and rectangles) with layouts matching text prompts, they struggle to produce the complex geometric details required for professional SVG applications. Consequently, the limited generalization and poor expressiveness of language-based methods hinder their adoption for text-to-SVG generation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 498, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 295, + 714 + ], + "type": "text", + "content": "To address the aforementioned problems, we propose a novel hybrid framework that leverages the complementary strengths of image-based and language-based methods. Our approach first utilizes an LLM to synthesize SVG templates composed of basic primitives and then optimizes their geometric details guided by image diffusion models. This hybrid framework addresses the key limitations of both paradigms: The LLM-based template generation overcomes the domain-specific generalization constraints of traditional language models while ensuring shape regularity, as each element naturally corresponds to a single semantic component. The subsequent image-based optimization then enhances the expressiveness of these well-structured templates by capturing complex geometric details that LLMs alone struggle to produce. Specifically, to fully exploit LLMs' ability to create reasonable and complex layouts that match text prompts, we design an SVG-oriented prompt with multiple stages, including prompt ex" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 313, + 72, + 553, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 262 + ], + "type": "text", + "content": "pansion, SVG script generation, and visual refinement. To refine the generated SVG templates, we implement a dual-stage optimization strategy: (1) We first use an image diffusion model to synthesize detailed image-level targets, then follow [50] to optimize path-level latent vectors with a pretrained SVG VAE, eliminating common issues such as self-intersections and jagged curves; (2) We further refine geometric details by directly optimizing the point coordinates to capture fine-grained visual elements. Comprehensive experiments show that Chat2SVG consistently outperforms existing methods regarding overall visual quality, individual path regularity, and semantic coherence. Furthermore, our system enables intuitive editing through iterative natural language instructions, making vector graphics creation more accessible to non-expert users. Our contributions are summarized as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 314, + 264, + 553, + 431 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 314, + 264, + 553, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 264, + 553, + 312 + ], + "spans": [ + { + "bbox": [ + 314, + 264, + 553, + 312 + ], + "type": "text", + "content": "- Hybrid SVG Generation Framework. We introduce a novel text-to-SVG generation paradigm that combines Large Language Models with image diffusion models to produce high-quality SVG outputs." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 314, + 312, + 553, + 347 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 312, + 553, + 347 + ], + "spans": [ + { + "bbox": [ + 314, + 312, + 553, + 347 + ], + "type": "text", + "content": "- SVG-Oriented Prompt Design. We develop a specialized prompt system that directs LLMs to generate SVG templates using basic geometric primitives." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 348, + 553, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 348, + 553, + 396 + ], + "spans": [ + { + "bbox": [ + 314, + 348, + 553, + 396 + ], + "type": "text", + "content": "- Dual-Stage Optimization. We implement a two-phase optimization process that preserves the semantic meaning of each shape while eliminating artifacts such as self-intersecting and jagged paths." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 396, + 553, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 396, + 553, + 431 + ], + "spans": [ + { + "bbox": [ + 314, + 396, + 553, + 431 + ], + "type": "text", + "content": "- Iterative Editing. We enable iterative refinement of SVG through natural language instructions, making SVG creation more accessible." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 444, + 400, + 456 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 444, + 400, + 456 + ], + "spans": [ + { + "bbox": [ + 314, + 444, + 400, + 456 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 464, + 501, + 478 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 464, + 501, + 478 + ], + "spans": [ + { + "bbox": [ + 313, + 464, + 501, + 478 + ], + "type": "text", + "content": "2.1. Large Language Models for Design" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 482, + 555, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 482, + 555, + 662 + ], + "spans": [ + { + "bbox": [ + 313, + 482, + 555, + 662 + ], + "type": "text", + "content": "LLMs have emerged as powerful tools for graphics design tasks, demonstrating remarkable capabilities in understanding and reasoning about complex design specifications across diverse domains. Recent works have explored their applications in (1) visual layout and composition, including poster designs [5, 21, 34, 47], 3D scene arrangements [2, 10, 14, 32, 33, 38, 51], and placement of basic shapes for image generation [8, 46]; (2) content creation and manipulation for shapes [11], materials [15], and animation [22, 36]; and (3) design understanding [18, 42]. While these demonstrate LLMs' versatility, their potential for vector graphics remains unexplored. Our work addresses this gap by showing that LLMs can effectively generate structured SVG templates, extending their capabilities to the domain of vector graphics creation." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 671, + 531, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 671, + 531, + 684 + ], + "spans": [ + { + "bbox": [ + 313, + 671, + 531, + 684 + ], + "type": "text", + "content": "2.2. Text-Guided Vector Graphics Generation" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "type": "text", + "content": "Text-to-SVG generation approaches can be roughly categorized into image-based and language-based methods." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23691" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 373 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 373 + ], + "type": "text", + "content": "Image-based methods [9, 16, 43, 44, 50] start with a large collection of randomly initialized shapes, render vector graphics with a differentiable renderer [20], and iteratively optimize path parameters by minimizing text-image similarity losses (e.g., CLIP [28] loss and Score Distillation Sampling [27] loss). Despite their visually pleasant appearance, these methods often generate SVG results containing multiple fragmented paths that lack individual semantic correspondence, which is undesirable for real design scenarios. Language-based methods [3, 35, 41] design specialized tokenization strategies to encode vector graphics into discrete tokens, which are then concatenated with text tokens, and train seq2seq models like autoregressive transformers [37] on domain-specific datasets (e.g., icons and fonts). Although these approaches are conceptually elegant, their generalizability is limited by the scarcity of large-scale, general-purpose vector graphics datasets. Some works [26, 31] have explored LLMs' vector graphics generation capabilities, showing that while LLMs can create reasonable layouts matching text prompts, they struggle to produce complex geometric shapes. Our Chat2SVG combines image-based and language-based methods to address the limitations of each paradigm, producing semantically meaningful SVG where each path corresponds to a distinct visual element and contains refined geometric details." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 378, + 275, + 392 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 378, + 275, + 392 + ], + "spans": [ + { + "bbox": [ + 55, + 378, + 275, + 392 + ], + "type": "text", + "content": "2.3. Vector Graphics Representation Learning" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 396, + 297, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 396, + 297, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 396, + 297, + 601 + ], + "type": "text", + "content": "Vector graphics representation learning is essential for sketch-based retrieval, reconstruction, and generation tasks. A seminal study, SketchRNN [12], combines a Recurrent Neural Network (RNN) and a VAE to learn vector representations. Lopes et al. [24] trained a VAE to represent image-level font styles, then passed the latent embedding into a decoder for vector font generation. To increase reconstruction quality, later approaches like Sketchformer [30] and DeepSVG [4] adopted Transformer [37] architectures, followed by dual-modality methods [23, 39, 40] that leveraged both vector and image features. However, encoding the entire complex SVG into a single latent embedding often causes detail loss. To address this limitation, Zhang et al. [50] proposed path-level representational learning, creating a smooth latent space for individual paths. We leverage these path-level latent embeddings in our optimization to eliminate self-intersecting and jagged paths." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 609, + 112, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 609, + 112, + 623 + ], + "spans": [ + { + "bbox": [ + 55, + 609, + 112, + 623 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "content": "Our Chat2SVG, as shown in Figure 2, begins with an SVG-oriented prompting approach (Section 3.1) that guides an LLM to create reasonable SVG templates. We then perform a dual-stage optimization (Section 3.2) to enhance the geometric details of SVG templates by optimizing SVG paths in both latent and control point space, guided by image diffusion models." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 314, + 72, + 517, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 517, + 85 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 517, + 85 + ], + "type": "text", + "content": "3.1. SVG Template Generation with LLMs" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 89, + 555, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 89, + 555, + 365 + ], + "spans": [ + { + "bbox": [ + 313, + 89, + 555, + 365 + ], + "type": "text", + "content": "Prompt Expansion. Users often provide vague and brief prompts to the LLM when describing their desired graphics. Such unstructured input may lead to oversimplified SVG (regarding both the number of elements and structures) that fails to reflect the intended design. Consequently, we propose a three-level prompt expansion strategy that systematically delineates the initial prompts: (1) Scene-level: Starting with an initial prompt, we instruct the LLM to analyze it holistically and identify essential objects that should appear. The LLM then expands the prompt by suggesting complementary objects to enhance scene completeness. (2) Object-level: For each object in the expanded prompt, we guide the LLM to systematically break it down into its components. For instance, when describing a lion, the LLM deconstructs it into distinct parts such as the body, head, mane, legs, eyes, ears, and tail. This decomposition ensures that no critical components are missing during SVG script generation. (3) Layout-level: After object decomposition, we direct the LLM to develop a comprehensive layout plan. This includes determining the position and size of each element on the canvas, selecting appropriate colors for visual harmony, and establishing clear spatial relationships to ensure a cohesive overall composition." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 365, + 556, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 365, + 556, + 472 + ], + "spans": [ + { + "bbox": [ + 313, + 365, + 556, + 472 + ], + "type": "text", + "content": "SVG Script Generation. After prompt expansion, we obtain a detailed scene description. We then convert this natural language specification into SVG scripts in XML format. Since the LLM has limitations in synthesizing geometrically complex paths, we constrain the shapes to a carefully selected set of basic primitives: rectangles, ellipses, lines, polylines, polygons, and short paths. We set the canvas size to " + }, + { + "bbox": [ + 313, + 365, + 556, + 472 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 313, + 365, + 556, + 472 + ], + "type": "text", + "content": ". Each geometric element is assigned a unique ID and includes semantic annotations in its comments." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 474, + 556, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 474, + 556, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 474, + 556, + 665 + ], + "type": "text", + "content": "Visual Rectification. Since prompt expansion is purely text-based, even when the generated SVG script accurately follows the detailed prompt, visual inconsistencies (e.g., misaligned components, disproportionate scaling, and incorrect path ordering) can emerge during rendering. For example, in Figure 2, the initial SVG template shows a lion with a misshapen mane and missing facial features. We thus adopt a visual rectification strategy in which we render the SVG and provide the rendered image back to the LLM (with vision capability) for inspection. The LLM analyzes the visual output, identifies any inconsistencies or oddities, and generates corrected SVG code. This visual refinement loop can be performed iteratively. In our experiments, we find that two iterations of refinement are typically sufficient to generate well-structured SVG templates with appropriate spatial layouts." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 666, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 666, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 666, + 556, + 715 + ], + "type": "text", + "content": "Furthermore, to enhance the quality of prompt expansion and SVG script generation, we provide curated in-context examples in the prompts. The complete set of prompts and examples is available in the supplementary material." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "23692" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 100, + 71, + 511, + 153 + ], + "blocks": [ + { + "bbox": [ + 100, + 71, + 511, + 153 + ], + "lines": [ + { + "bbox": [ + 100, + 71, + 511, + 153 + ], + "spans": [ + { + "bbox": [ + 100, + 71, + 511, + 153 + ], + "type": "image", + "image_path": "fe924cf106c9beec49f8582b30b7f20d22c68f6d792b6640cf4a576f58ece790.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 100, + 155, + 328, + 254 + ], + "blocks": [ + { + "bbox": [ + 100, + 155, + 328, + 254 + ], + "lines": [ + { + "bbox": [ + 100, + 155, + 328, + 254 + ], + "spans": [ + { + "bbox": [ + 100, + 155, + 328, + 254 + ], + "type": "image", + "image_path": "f9134506dd5b3e81b330b3ffdd59872c440a6c8ebe9c1fed23ce7918399d1308.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 263, + 555, + 331 + ], + "lines": [ + { + "bbox": [ + 54, + 263, + 555, + 331 + ], + "spans": [ + { + "bbox": [ + 54, + 263, + 555, + 331 + ], + "type": "text", + "content": "Figure 2. The system pipeline of Chat2SVG. Given a text prompt, our system first leverages an LLM to generate an SVG template composed of basic geometric primitives. The rendered template is enhanced through SDEdit [25] with ControlNet [49] to add visual details while preserving the overall composition, yielding a target image. The SVG then undergoes a dual-stage optimization process to match the target image. (1) Primitives are converted to latent embeddings through latent inversion and optimized along with their visual attributes (i.e., filling colors " + }, + { + "bbox": [ + 54, + 263, + 555, + 331 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 54, + 263, + 555, + 331 + ], + "type": "text", + "content": ", stroke properties " + }, + { + "bbox": [ + 54, + 263, + 555, + 331 + ], + "type": "inline_equation", + "content": "s_{i}" + }, + { + "bbox": [ + 54, + 263, + 555, + 331 + ], + "type": "text", + "content": ", and transformation matrices " + }, + { + "bbox": [ + 54, + 263, + 555, + 331 + ], + "type": "inline_equation", + "content": "T_{i}" + }, + { + "bbox": [ + 54, + 263, + 555, + 331 + ], + "type": "text", + "content": "). (2) Point-level optimization is performed to refine the geometric details of SVG paths." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 331, + 155, + 511, + 254 + ], + "blocks": [ + { + "bbox": [ + 331, + 155, + 511, + 254 + ], + "lines": [ + { + "bbox": [ + 331, + 155, + 511, + 254 + ], + "spans": [ + { + "bbox": [ + 331, + 155, + 511, + 254 + ], + "type": "image", + "image_path": "80317083313508ff8a164d592182c2dde4f4cbe8a5589373af247225e0bb9622.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 350, + 294, + 364 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 350, + 294, + 364 + ], + "spans": [ + { + "bbox": [ + 55, + 350, + 294, + 364 + ], + "type": "text", + "content": "3.2. SVG Optimization Guided by Image Diffusion" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 373, + 295, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 373, + 295, + 577 + ], + "spans": [ + { + "bbox": [ + 54, + 373, + 295, + 577 + ], + "type": "text", + "content": "Detail Enhancement. While the primitive shapes in the SVG template have accurate semantic meaning, they lack the geometric details necessary for a professional design. To improve visual expressiveness, we first render the SVG template into a raster image, and then use an image editing method to generate a more detailed version that serves as our optimization target. Specifically, we employ SDEdit [25], which enhances the input image by first adding noise and then progressively denoising it with an image diffusion model to produce outputs with richer details, such as the contour of the lion's mane illustrated in Figure 2. To maintain structural similarity between the enhanced output and the original image, we incorporate a ControlNet (tile version) [49] into our pipeline. By using Gaussian-blurred versions of the initial images as control signals, the ControlNet effectively maintains the overall composition throughout the enhancement process." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "content": "Apart from geometric details, this image-to-image translation process also introduces new semantic parts such as the lion's beard. To faithfully reproduce the target image, we propose using the Segment Anything Model (SAM) [17] to identify these new decorative details. Specifically, we denote the rendered images of the initial SVG template and its diffusion-enhanced counterpart as " + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{template}}" + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{target}}" + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "content": ", respectively. We apply SAM to extract mask sets from both images: " + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\{m_i\\}_{i=1}^Q" + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "content": " for the template image and " + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\{m_j\\}_{j=1}^K" + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "content": " for the target image, where " + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "content": " represent the number of masks in each collection. Our approach is based on the ob" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 351, + 555, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 351, + 555, + 533 + ], + "spans": [ + { + "bbox": [ + 313, + 351, + 555, + 533 + ], + "type": "text", + "content": "servation that the diffusion process simultaneously generates fine decorative details that are absent in the original template (e.g., the beard of the lion in Figure 2) while preserving the structural integrity of larger components (such as the body). Consequently, we implement an area-based mask-filtering strategy to identify these newly generated elements. For each mask " + }, + { + "bbox": [ + 313, + 351, + 555, + 533 + ], + "type": "inline_equation", + "content": "m_j" + }, + { + "bbox": [ + 313, + 351, + 555, + 533 + ], + "type": "text", + "content": " detected in " + }, + { + "bbox": [ + 313, + 351, + 555, + 533 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{target}}" + }, + { + "bbox": [ + 313, + 351, + 555, + 533 + ], + "type": "text", + "content": ", we evaluate it against all masks in " + }, + { + "bbox": [ + 313, + 351, + 555, + 533 + ], + "type": "inline_equation", + "content": "\\{m_i\\}_{i=1}^Q" + }, + { + "bbox": [ + 313, + 351, + 555, + 533 + ], + "type": "text", + "content": " using Intersection over Union (IoU) metrics [29]. When " + }, + { + "bbox": [ + 313, + 351, + 555, + 533 + ], + "type": "inline_equation", + "content": "m_j" + }, + { + "bbox": [ + 313, + 351, + 555, + 533 + ], + "type": "text", + "content": " overlaps with a template mask " + }, + { + "bbox": [ + 313, + 351, + 555, + 533 + ], + "type": "inline_equation", + "content": "m_i" + }, + { + "bbox": [ + 313, + 351, + 555, + 533 + ], + "type": "text", + "content": " but their IoU falls below a threshold, we classify " + }, + { + "bbox": [ + 313, + 351, + 555, + 533 + ], + "type": "inline_equation", + "content": "m_j" + }, + { + "bbox": [ + 313, + 351, + 555, + 533 + ], + "type": "text", + "content": " as a new decorative component. Alternatively, if their IoU exceeds the threshold, we interpret " + }, + { + "bbox": [ + 313, + 351, + 555, + 533 + ], + "type": "inline_equation", + "content": "m_j" + }, + { + "bbox": [ + 313, + 351, + 555, + 533 + ], + "type": "text", + "content": " as a variation of an existing component. For all filtered masks classified as new decorative shapes, we approximate their boundaries with polygons and integrate them into the SVG script." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 536, + 556, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 536, + 556, + 638 + ], + "spans": [ + { + "bbox": [ + 313, + 536, + 556, + 638 + ], + "type": "text", + "content": "We define an SVG script " + }, + { + "bbox": [ + 313, + 536, + 556, + 638 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 313, + 536, + 556, + 638 + ], + "type": "text", + "content": " as a collection of " + }, + { + "bbox": [ + 313, + 536, + 556, + 638 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 536, + 556, + 638 + ], + "type": "text", + "content": " paths (i.e., shapes), " + }, + { + "bbox": [ + 313, + 536, + 556, + 638 + ], + "type": "inline_equation", + "content": "G = \\{P_i\\}_{i=1}^M" + }, + { + "bbox": [ + 313, + 536, + 556, + 638 + ], + "type": "text", + "content": ". Each path " + }, + { + "bbox": [ + 313, + 536, + 556, + 638 + ], + "type": "inline_equation", + "content": "P_i" + }, + { + "bbox": [ + 313, + 536, + 556, + 638 + ], + "type": "text", + "content": " consists of a sequence of " + }, + { + "bbox": [ + 313, + 536, + 556, + 638 + ], + "type": "inline_equation", + "content": "N_i" + }, + { + "bbox": [ + 313, + 536, + 556, + 638 + ], + "type": "text", + "content": " commands, " + }, + { + "bbox": [ + 313, + 536, + 556, + 638 + ], + "type": "inline_equation", + "content": "P_i = \\{C_i^j\\}_{j=1}^{N_i}" + }, + { + "bbox": [ + 313, + 536, + 556, + 638 + ], + "type": "text", + "content": ". A command " + }, + { + "bbox": [ + 313, + 536, + 556, + 638 + ], + "type": "inline_equation", + "content": "C_i^j = (U_i^j, V_i^j)" + }, + { + "bbox": [ + 313, + 536, + 556, + 638 + ], + "type": "text", + "content": " is defined by its type " + }, + { + "bbox": [ + 313, + 536, + 556, + 638 + ], + "type": "inline_equation", + "content": "U_i^j \\in \\{\\mathbb{M}, \\mathbb{C}\\}" + }, + { + "bbox": [ + 313, + 536, + 556, + 638 + ], + "type": "text", + "content": " and its associated control points " + }, + { + "bbox": [ + 313, + 536, + 556, + 638 + ], + "type": "inline_equation", + "content": "V_i^j" + }, + { + "bbox": [ + 313, + 536, + 556, + 638 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 536, + 556, + 638 + ], + "type": "inline_equation", + "content": "\\mathbb{M}" + }, + { + "bbox": [ + 313, + 536, + 556, + 638 + ], + "type": "text", + "content": " represents Move and " + }, + { + "bbox": [ + 313, + 536, + 556, + 638 + ], + "type": "inline_equation", + "content": "\\mathbb{C}" + }, + { + "bbox": [ + 313, + 536, + 556, + 638 + ], + "type": "text", + "content": " represents cubic Bézier curves. For consistency, we convert all other primitive shapes (e.g., rectangles and ellipses) in the SVG template into cubic Bézier curves." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "text", + "content": "Previous image-based methods [9, 16, 43, 44] optimize SVG shapes by directly manipulating the control points " + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "inline_equation", + "content": "V_{i}^{j}" + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "text", + "content": ". Despite convenience, this approach often leads to artifacts such as self-intersecting curves and unnatural deformations. To overcome these limitations, Zhang et al. [50] introduced a path-level SVG VAE. This model's latent space effec" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23693" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "type": "text", + "content": "tively captures common shape patterns and geometric constraints, allowing for the optimization of path latent vectors at a higher level to produce smooth outputs." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 108, + 295, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 108, + 295, + 263 + ], + "spans": [ + { + "bbox": [ + 55, + 108, + 295, + 263 + ], + "type": "text", + "content": "Latent Optimization. We leverage this pretrained SVG VAE to conduct the latent optimization. However, their SVG encoder expects a fixed number of commands (10 cubic Bezier curves), while our SVG primitives contain varying (typically fewer) commands. To resolve this mismatch, we develop a latent inversion process that converts our primitives into latent embeddings. We start with a randomly sampled latent vector " + }, + { + "bbox": [ + 55, + 108, + 295, + 263 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 55, + 108, + 295, + 263 + ], + "type": "text", + "content": " and decode it into a command sequence. Then we evenly sample points " + }, + { + "bbox": [ + 55, + 108, + 295, + 263 + ], + "type": "inline_equation", + "content": "X = \\{x_{i}\\}_{i=1}^{N}" + }, + { + "bbox": [ + 55, + 108, + 295, + 263 + ], + "type": "text", + "content": " along its contour. Similarly, we sample an equal number of points " + }, + { + "bbox": [ + 55, + 108, + 295, + 263 + ], + "type": "inline_equation", + "content": "Y = \\{y_{j}\\}_{j=1}^{N}" + }, + { + "bbox": [ + 55, + 108, + 295, + 263 + ], + "type": "text", + "content": " along the contour of our target primitive shape. Using these point sets, we compute the Earth Mover's Distance (EMD) [1, 7, 48]:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 88, + 269, + 295, + 293 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 269, + 295, + 293 + ], + "spans": [ + { + "bbox": [ + 88, + 269, + 295, + 293 + ], + "type": "interline_equation", + "content": "\\ell_ {\\mathrm {E M D}} (X, Y) = \\min _ {\\phi : X \\rightarrow Y} \\sum_ {x \\in X} \\| x - \\phi (x) \\| _ {2}, \\tag {1}", + "image_path": "3c029271fe09827d1cb189fd3605393369f8146f3ac5ae5e01be8c48902eb1af.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 299, + 295, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 299, + 295, + 334 + ], + "spans": [ + { + "bbox": [ + 55, + 299, + 295, + 334 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 299, + 295, + 334 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 55, + 299, + 295, + 334 + ], + "type": "text", + "content": " represents a bijective mapping between the point sets. By minimizing this EMD loss, we gradually optimize the latent embedding " + }, + { + "bbox": [ + 55, + 299, + 295, + 334 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 55, + 299, + 295, + 334 + ], + "type": "text", + "content": " to match the target primitive." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 334, + 296, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 334, + 296, + 502 + ], + "spans": [ + { + "bbox": [ + 55, + 334, + 296, + 502 + ], + "type": "text", + "content": "After obtaining latent vectors " + }, + { + "bbox": [ + 55, + 334, + 296, + 502 + ], + "type": "inline_equation", + "content": "\\{z_i\\}_{i=1}^M" + }, + { + "bbox": [ + 55, + 334, + 296, + 502 + ], + "type": "text", + "content": " for all paths, we jointly optimize these vectors along with visual attribute parameters: filling colors " + }, + { + "bbox": [ + 55, + 334, + 296, + 502 + ], + "type": "inline_equation", + "content": "\\{c_i\\}_{i=1}^M" + }, + { + "bbox": [ + 55, + 334, + 296, + 502 + ], + "type": "text", + "content": ", stroke properties (color and width) " + }, + { + "bbox": [ + 55, + 334, + 296, + 502 + ], + "type": "inline_equation", + "content": "\\{s_i\\}_{i=1}^M" + }, + { + "bbox": [ + 55, + 334, + 296, + 502 + ], + "type": "text", + "content": ", and transformation matrices " + }, + { + "bbox": [ + 55, + 334, + 296, + 502 + ], + "type": "inline_equation", + "content": "\\{T_i\\}_{i=1}^M" + }, + { + "bbox": [ + 55, + 334, + 296, + 502 + ], + "type": "text", + "content": ". At each optimization iteration " + }, + { + "bbox": [ + 55, + 334, + 296, + 502 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 334, + 296, + 502 + ], + "type": "text", + "content": ", we decode the latent vectors into command sequences and apply transformation matrices to the control points. We then combine color and stroke attributes together, and use a differentiable rasterizer [20] to render the complete SVG into an image " + }, + { + "bbox": [ + 55, + 334, + 296, + 502 + ], + "type": "inline_equation", + "content": "I_t" + }, + { + "bbox": [ + 55, + 334, + 296, + 502 + ], + "type": "text", + "content": ". Our optimization objective consists of three loss terms: (1) An MSE loss " + }, + { + "bbox": [ + 55, + 334, + 296, + 502 + ], + "type": "inline_equation", + "content": "\\ell_{\\mathrm{MSE}}" + }, + { + "bbox": [ + 55, + 334, + 296, + 502 + ], + "type": "text", + "content": " between the rendered image " + }, + { + "bbox": [ + 55, + 334, + 296, + 502 + ], + "type": "inline_equation", + "content": "I_t" + }, + { + "bbox": [ + 55, + 334, + 296, + 502 + ], + "type": "text", + "content": " and target image " + }, + { + "bbox": [ + 55, + 334, + 296, + 502 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{target}}" + }, + { + "bbox": [ + 55, + 334, + 296, + 502 + ], + "type": "text", + "content": " to ensure visual similarity; (2) A curvature loss that reduces sharp bends and fluctuations along the contour by computing the discrete second derivative:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 70, + 507, + 295, + 563 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 507, + 295, + 563 + ], + "spans": [ + { + "bbox": [ + 70, + 507, + 295, + 563 + ], + "type": "interline_equation", + "content": "\\ell_ {\\text {c u r v a t u r e}} = \\frac {\\sum_ {i = 1} ^ {M} \\sum_ {j = 1} ^ {N _ {i} - 2} \\left\\| V _ {i} ^ {j} - 2 V _ {i} ^ {j + 1} + V _ {i} ^ {j + 2} \\right\\| _ {2} ^ {2}}{\\sum_ {i = 1} ^ {M} \\left(N _ {i} - 2\\right)}; \\tag {2}", + "image_path": "e5a7220c12353fe98671c040570752a7ec94d938457416233f413dfa4d1493b1.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 567, + 295, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 567, + 295, + 604 + ], + "spans": [ + { + "bbox": [ + 55, + 567, + 295, + 604 + ], + "type": "text", + "content": "(3) A path-level IoU loss that preserves the overall composition and prevents paths from drifting too far from their initial positions:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 102, + 609, + 295, + 641 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 609, + 295, + 641 + ], + "spans": [ + { + "bbox": [ + 102, + 609, + 295, + 641 + ], + "type": "interline_equation", + "content": "\\ell_ {\\mathrm {I o U}} = \\frac {1}{M} \\sum_ {i = 1} ^ {M} \\left(1 - \\frac {\\left| m _ {i} ^ {t} \\cap m _ {i} ^ {0} \\right|}{\\left| m _ {i} ^ {t} \\cup m _ {i} ^ {0} \\right|}\\right), \\tag {3}", + "image_path": "2245f3019d63679ed834401f1c1a33a39da31380965535eb157bb7cbfadda9f8.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 647, + 295, + 671 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 647, + 295, + 671 + ], + "spans": [ + { + "bbox": [ + 55, + 647, + 295, + 671 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 647, + 295, + 671 + ], + "type": "inline_equation", + "content": "m_i^t" + }, + { + "bbox": [ + 55, + 647, + 295, + 671 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 647, + 295, + 671 + ], + "type": "inline_equation", + "content": "m_i^0" + }, + { + "bbox": [ + 55, + 647, + 295, + 671 + ], + "type": "text", + "content": " denote the binary masks of the " + }, + { + "bbox": [ + 55, + 647, + 295, + 671 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 647, + 295, + 671 + ], + "type": "text", + "content": "-th path at the current iteration " + }, + { + "bbox": [ + 55, + 647, + 295, + 671 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 647, + 295, + 671 + ], + "type": "text", + "content": " and its initial state, respectively." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 671, + 295, + 695 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 671, + 295, + 695 + ], + "spans": [ + { + "bbox": [ + 55, + 671, + 295, + 695 + ], + "type": "text", + "content": "The final loss function combines these three terms with empirically determined weights:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 92, + 701, + 295, + 714 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 92, + 701, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 92, + 701, + 295, + 714 + ], + "type": "interline_equation", + "content": "\\ell_ {\\text {l a t e n t}} = \\ell_ {\\mathrm {M S E}} + \\lambda_ {1} \\ell_ {\\text {c u r v a t u r e}} + \\lambda_ {2} \\ell_ {\\mathrm {I o U}}, \\tag {4}", + "image_path": "e395ea87f89349860841a3f218ad71d18559bbcab5db553d3b284ffa395d6053.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 72, + 467, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 467, + 83 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 467, + 83 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 72, + 467, + 83 + ], + "type": "inline_equation", + "content": "\\lambda_1 = 5e - 4" + }, + { + "bbox": [ + 313, + 72, + 467, + 83 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 72, + 467, + 83 + ], + "type": "inline_equation", + "content": "\\lambda_{2} = 5e - 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 84, + 553, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 84, + 553, + 275 + ], + "spans": [ + { + "bbox": [ + 313, + 84, + 553, + 275 + ], + "type": "text", + "content": "Point Optimization. While latent optimization effectively aligns SVG shapes with their target positions and contours, it has a notable limitation: the resulting shapes tend to be overly smooth, lacking the intricate details that are often essential in professional vector graphics. For example, in Figure 2, the lion's mane after latent optimization does not capture the fine contour present in the target image. Therefore, we add a second optimization stage that directly refines the control points " + }, + { + "bbox": [ + 313, + 84, + 553, + 275 + ], + "type": "inline_equation", + "content": "V_{i}^{j}" + }, + { + "bbox": [ + 313, + 84, + 553, + 275 + ], + "type": "text", + "content": " of the paths, similar to previous image-based works [9, 16, 43, 44]. To achieve finer granularity in shape control, we split each cubic Bezier curve at its midpoint, effectively doubling the number of control points in the SVG. This increased control point density allows for more precise shape adjustments. The point optimization stage employs a loss function that combines the MSE loss and the curvature loss:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 372, + 286, + 553, + 300 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 286, + 553, + 300 + ], + "spans": [ + { + "bbox": [ + 372, + 286, + 553, + 300 + ], + "type": "interline_equation", + "content": "\\ell_ {\\text {p o i n t}} = \\ell_ {\\mathrm {M S E}} + \\lambda_ {3} \\ell_ {\\text {c u r v a t u r e}}. \\tag {5}", + "image_path": "043a75d50c608853d3a60feb3be81f57182e037833a9cd2df682cf441bf1e5c1.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 308, + 553, + 333 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 308, + 553, + 333 + ], + "spans": [ + { + "bbox": [ + 313, + 308, + 553, + 333 + ], + "type": "text", + "content": "The parameter " + }, + { + "bbox": [ + 313, + 308, + 553, + 333 + ], + "type": "inline_equation", + "content": "\\lambda_3" + }, + { + "bbox": [ + 313, + 308, + 553, + 333 + ], + "type": "text", + "content": " decreases linearly from " + }, + { + "bbox": [ + 313, + 308, + 553, + 333 + ], + "type": "inline_equation", + "content": "1e - 3" + }, + { + "bbox": [ + 313, + 308, + 553, + 333 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 308, + 553, + 333 + ], + "type": "inline_equation", + "content": "5e - 5" + }, + { + "bbox": [ + 313, + 308, + 553, + 333 + ], + "type": "text", + "content": " throughout the optimization process." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 314, + 341, + 414, + 354 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 341, + 414, + 354 + ], + "spans": [ + { + "bbox": [ + 314, + 341, + 414, + 354 + ], + "type": "text", + "content": "3.3. Iterative Editing" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 358, + 553, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 358, + 553, + 418 + ], + "spans": [ + { + "bbox": [ + 313, + 358, + 553, + 418 + ], + "type": "text", + "content": "During SVG script generation (Section 3.1), we instruct the LLM to annotate the semantic label of each path. This design enables users to refine the initial SVG template using natural language, as the LLM can precisely locate and modify specific elements based on semantic understanding." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 418, + 554, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 418, + 554, + 562 + ], + "spans": [ + { + "bbox": [ + 313, + 418, + 554, + 562 + ], + "type": "text", + "content": "Our system supports iterative refinement through multiple rounds of natural language instructions. However, when passing the edited SVG template image into SDEdit, image diffusion models may introduce different shape variations and decorative details, which compromise the user's intention to maintain the already optimized shapes. To preserve consistency across editing iterations, we instruct the LLM to output both the modified SVG script and a list of specifically changed paths. This precise tracking enables selective optimization of modified shapes while preserving unaltered elements from previous iterations, maintaining visual coherence throughout the entire refinement process." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 314, + 574, + 395, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 574, + 395, + 586 + ], + "spans": [ + { + "bbox": [ + 314, + 574, + 395, + 586 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "content": "Implementation Details. In our experiments, we use claude-3-5-sonnet as our backbone LLM model due to its leading generation capabilities. To evaluate our approach, we create a prompt set by having the LLM generate 125 text prompts across five categories: animals, food, objects, scenes, and novel concept combinations. For each prompt, we generate one template and perform 2 rounds of visual rectification. This process is repeated 5 times, yielding " + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "inline_equation", + "content": "(1 + 2) \\times 5 = 15" + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "content": " candidate SVG templates per prompt. These templates are rendered as images, and we follow a" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23694" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 82, + 68, + 531, + 441 + ], + "blocks": [ + { + "bbox": [ + 82, + 68, + 531, + 441 + ], + "lines": [ + { + "bbox": [ + 82, + 68, + 531, + 441 + ], + "spans": [ + { + "bbox": [ + 82, + 68, + 531, + 441 + ], + "type": "image", + "image_path": "cd01b7e83cd888ef4faaa359d8e3600ab0db1b8221bee6904932994e764ff4a1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 448, + 555, + 526 + ], + "lines": [ + { + "bbox": [ + 54, + 448, + 555, + 526 + ], + "spans": [ + { + "bbox": [ + 54, + 448, + 555, + 526 + ], + "type": "text", + "content": "Figure 3. Qualitative Comparison. (1) Methods refining open-ended strokes, i.e., CLIPDraw [9] and DiffSketcher [43], often produce distorted and disorganized strokes to approximate objects, presenting messy appearance and poor text alignment. (2) VectorFusion [16] and SVGDreamer [44] produce elements that consist of multiple jagged, irregular, and fragmented shapes, such as the body of the flamingo (first row) and the plane (third row). (3) T2V-NPR [50] attempts to resolve these issues by learning a latent representation of paths and merging fragmented shapes. However, it still cannot guarantee the semantic meanings of the paths, leading to less-semantic paths such as a plane body with surrounding clouds in the third row. In contrast, our method produces SVG with superior text alignment, higher visual quality, and well-structured paths exhibiting geometric regularity and clear semantic definition." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 546, + 294, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 546, + 294, + 583 + ], + "spans": [ + { + "bbox": [ + 55, + 546, + 294, + 583 + ], + "type": "text", + "content": "standard practice to select the highest quality SVG using the ImageReward [45] metric. Detailed optimization configurations are provided in the supplementary materials." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 587, + 295, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 587, + 295, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 587, + 295, + 696 + ], + "type": "text", + "content": "Baselines. We compare our approach against five state-of-the-art image-based text-to-SVG generation methods: CLIPDraw [9], DiffSketcher [43], VectorFusion [16], SVGDreamer [44], and T2V-NPR [50]. Current language-based methods [35, 41] are confined to specific categories, with no support for general text-to-SVG generation, so we do not include them in our comparison. For a fair comparison, we set these baseline methods to use a similar number of shapes as our optimized SVG." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 701, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 701, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 701, + 295, + 714 + ], + "type": "text", + "content": "Evaluation Metrics. We evaluate the quality of our gen" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 546, + 553, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 546, + 553, + 570 + ], + "spans": [ + { + "bbox": [ + 313, + 546, + 553, + 570 + ], + "type": "text", + "content": "erated SVG across three dimensions: visual fidelity, path regularity, and semantic alignment." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 571, + 553, + 691 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 314, + 571, + 553, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 571, + 553, + 654 + ], + "spans": [ + { + "bbox": [ + 314, + 571, + 553, + 654 + ], + "type": "text", + "content": "- Image-level Fidelity. We evaluate visual fidelity using a ground-truth dataset of well-designed vector graphics, specifically 52,805 colored SVG files downloaded from SVGRepo1. We compute the Fréchet Inception Distance (FID) [13] between the rendered images of our generated SVG and this professional design collection, using features extracted by the CLIP image encoder [28]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 655, + 553, + 691 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 655, + 553, + 691 + ], + "spans": [ + { + "bbox": [ + 313, + 655, + 553, + 691 + ], + "type": "text", + "content": "- Path-level Regularity. We assess path quality using a transformer-based Auto-Encoder trained on the FIGR-8-SVG [6] dataset, following the DeepSVG [4] architecture" + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 325, + 703, + 443, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 703, + 443, + 712 + ], + "spans": [ + { + "bbox": [ + 325, + 703, + 443, + 712 + ], + "type": "text", + "content": "1https://www.svgrepo.com" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23695" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 75, + 70, + 277, + 159 + ], + "blocks": [ + { + "bbox": [ + 75, + 70, + 277, + 159 + ], + "lines": [ + { + "bbox": [ + 75, + 70, + 277, + 159 + ], + "spans": [ + { + "bbox": [ + 75, + 70, + 277, + 159 + ], + "type": "table", + "html": "
MethodImage FID ↓Path FID ↓Text Alignment ↑
CLIPDraw [9]46.7770.130.3048
DiffSketcher [43]44.8966.480.2623
VectorFusion [16]39.5256.790.2982
SVGDreamer [44]35.4847.950.2919
T2V-NPR [50]39.8642.030.3078
Chat2SVG (Ours)33.3139.070.3090
", + "image_path": "e71a77f3e2c2286f20798e692568ae2886876fd20455a53b0a16ce4e2607f445.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 166, + 295, + 190 + ], + "lines": [ + { + "bbox": [ + 55, + 166, + 295, + 190 + ], + "spans": [ + { + "bbox": [ + 55, + 166, + 295, + 190 + ], + "type": "text", + "content": "Table 1. Quantitative comparison of text-to-SVG generation methods across image fidelity, vector regularity, and text alignment." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 63, + 210, + 295, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 210, + 295, + 293 + ], + "spans": [ + { + "bbox": [ + 63, + 210, + 295, + 293 + ], + "type": "text", + "content": "ture. Our model, trained with reconstruction tasks, encodes each drawing command " + }, + { + "bbox": [ + 63, + 210, + 295, + 293 + ], + "type": "inline_equation", + "content": "C_i^j" + }, + { + "bbox": [ + 63, + 210, + 295, + 293 + ], + "type": "text", + "content": " into a latent vector. We represent each path as the mean of its commands' latent embeddings, and calculate the FID between these path representations and the ground truth paths from FIGR-8-SVG. This metric quantifies how closely our generated paths align with professional SVG design patterns." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 294, + 296, + 330 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 294, + 296, + 330 + ], + "spans": [ + { + "bbox": [ + 55, + 294, + 296, + 330 + ], + "type": "text", + "content": "- Text-level Alignment. We assess semantic alignment by computing the CLIP score [28] between the text prompt and the rendered SVG image." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 337, + 245, + 350 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 337, + 245, + 350 + ], + "spans": [ + { + "bbox": [ + 55, + 337, + 245, + 350 + ], + "type": "text", + "content": "4.1. Comparison with Existing Methods" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 354, + 295, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 354, + 295, + 461 + ], + "spans": [ + { + "bbox": [ + 54, + 354, + 295, + 461 + ], + "type": "text", + "content": "Quantitative Comparison. Table 1 shows the evaluation metrics for all baseline methods. Our method achieves the best image FID score, indicating that our generated SVG closely aligns with professional design patterns. Regarding path regularity, our method yields paths that are closest to the professionally designed SVG, as evidenced by the lowest path FID. Furthermore, our Chat2SVG achieves the highest score in text-SVG alignment, validating the significance of LLM-generated SVG templates." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 463, + 295, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 463, + 295, + 581 + ], + "spans": [ + { + "bbox": [ + 54, + 463, + 295, + 581 + ], + "type": "text", + "content": "Qualitative Comparison. In Figure 3, we present a side-by-side comparison between our Chat2SVG and the baselines. Methods that optimize open-ended strokes based on CLIP or Diffusion models (i.e., CLIPDraw [9] and DiffS-ketcher [43]) produce results with messy visual appearances and poor text alignment. These methods sometimes fail to synthesize complete objects, such as the flamingo (first row) and the plane (third row) in Figure 3. In contrast, our method generates SVG results with clear layouts and greater fidelity to the prompts." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 582, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 582, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 582, + 295, + 714 + ], + "type": "text", + "content": "Methods that optimize closed shapes via score distillation of diffusion models (i.e., VectorFusion [16], SVGDreamer [44], and T2V-NPR [50]) can better align with the input text, but still produce fragmented paths with limited semantic meaning. For instance, in the plane example (third row of Figure 3), VectorFusion and SVGDreamer create planes and clouds composed of multiple jagged, irregular, and fragmented shapes, which only appear meaningful when viewed as a whole. This hinders convenient editing by graphic designers. T2V-NPR [50] addresses the issue of jagged paths using a path VAE (although it can be" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 339, + 69, + 531, + 210 + ], + "blocks": [ + { + "bbox": [ + 339, + 69, + 531, + 210 + ], + "lines": [ + { + "bbox": [ + 339, + 69, + 531, + 210 + ], + "spans": [ + { + "bbox": [ + 339, + 69, + 531, + 210 + ], + "type": "image", + "image_path": "31a541393ca3e17bc2f81e2e8155d6fed320563e6c26162a7339b00768ddd09d.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 218, + 555, + 241 + ], + "lines": [ + { + "bbox": [ + 313, + 218, + 555, + 241 + ], + "spans": [ + { + "bbox": [ + 313, + 218, + 555, + 241 + ], + "type": "text", + "content": "Figure 4. User Study. Our Chat2SVG achieves the highest user selection ratio across all three evaluation criteria." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 264, + 555, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 264, + 555, + 420 + ], + "spans": [ + { + "bbox": [ + 313, + 264, + 555, + 420 + ], + "type": "text", + "content": "overly smooth at times) and reduces fragmented shapes by merging shapes with similar colors. However, its merging operation ignores the semantic meaning of shapes, resulting in the body of the plane being merged with the surrounding clouds. Additionally, the semantic meaning of the path is sometimes unclear, as exemplified by the cloud in the bottom-right corner, whose contour is actually approximated by the blue background instead of a true cloud shape. In contrast, our Chat2SVG adopts a unique approach by using LLMs to generate paths representing semantic components and performing dual-stage optimization based on image diffusion to enhance path expressiveness, thus ensuring both path regularity and visual fidelity of generated SVG." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 421, + 556, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 421, + 556, + 624 + ], + "spans": [ + { + "bbox": [ + 313, + 421, + 556, + 624 + ], + "type": "text", + "content": "User Study. To further compare Chat2SVG with baseline methods, we conducted a user study examining the same key aspects: (1) overall visual aesthetics, (2) path regularity, and (3) text-SVG alignment. We randomly sampled 20 prompts and generated SVG outputs using our approach and five baseline methods. For each prompt, participants were shown all generated SVG results along with the corresponding text prompt and highlighted paths. They were then asked to select the highest quality result regarding the three evaluation criteria. We recruited 31 participants (18 male, 13 female) through university mailing lists, with a mean age of 26 years. Among them, 17 participants reported prior experience in graphic design, providing a balanced mix of expert and novice perspectives. Analyzing the average selection ratio across all three metrics (Figure 4), we found that Chat2SVG-generated results were consistently preferred by participants over baseline methods." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 635, + 408, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 635, + 408, + 647 + ], + "spans": [ + { + "bbox": [ + 313, + 635, + 408, + 647 + ], + "type": "text", + "content": "4.2. Ablation Study" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 654, + 554, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 554, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 554, + 715 + ], + "type": "text", + "content": "We conduct ablation experiments to evaluate the effectiveness of key components in our pipeline (Section 3.1). First, we remove the SVG template generation and use randomly initialized shapes to approximate target images. As shown in the second row of Table 2, both image FID and path FID" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23696" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 69, + 70, + 282, + 129 + ], + "blocks": [ + { + "bbox": [ + 69, + 70, + 282, + 129 + ], + "lines": [ + { + "bbox": [ + 69, + 70, + 282, + 129 + ], + "spans": [ + { + "bbox": [ + 69, + 70, + 282, + 129 + ], + "type": "table", + "html": "
Chat2SVG VariantImage FID ↓Path FID ↓Text Alignment ↑
Default33.3139.070.3090
No SVG Template47.2170.130.2968
No SVG Optimization33.4536.120.3044
", + "image_path": "13f07943d5304d0b1bee7c5cdd542cec907fd793ef28f52a277d146cb6272fc6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 86, + 137, + 264, + 148 + ], + "lines": [ + { + "bbox": [ + 86, + 137, + 264, + 148 + ], + "spans": [ + { + "bbox": [ + 86, + 137, + 264, + 148 + ], + "type": "text", + "content": "Table 2. Qualitative results of the ablation study." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 58, + 159, + 296, + 244 + ], + "blocks": [ + { + "bbox": [ + 58, + 159, + 296, + 244 + ], + "lines": [ + { + "bbox": [ + 58, + 159, + 296, + 244 + ], + "spans": [ + { + "bbox": [ + 58, + 159, + 296, + 244 + ], + "type": "image", + "image_path": "88128b6fc59cd146790d4597fa859c088b39f479d7e423e29cb6749cd828d3b9.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 91, + 253, + 259, + 264 + ], + "lines": [ + { + "bbox": [ + 91, + 253, + 259, + 264 + ], + "spans": [ + { + "bbox": [ + 91, + 253, + 259, + 264 + ], + "type": "text", + "content": "Figure 5. Qualitative results of ablation study." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 285, + 296, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 285, + 296, + 453 + ], + "spans": [ + { + "bbox": [ + 55, + 285, + 296, + 453 + ], + "type": "text", + "content": "values increase substantially. This performance degradation can be attributed to the substantial gap between randomly initialized shapes and target shapes, making optimization more challenging. Meanwhile, in the second column of Figure 5, the dog's legs and ears lack semantic-clear components, highlighting the importance of SVG template generation. Second, we eliminate the dual-stage optimization. As shown in the third row of Table 2, this results in the lowest path FID, as primitives generated by LLMs naturally align with the path regularity presented in the ground-truth dataset. However, the visual outcome in Figure 5 lacks the necessary details, indicating that the overall visual quality suffers without dual-stage optimization. A more detailed ablation study is provided in the supplementary material." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 460, + 192, + 473 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 460, + 192, + 473 + ], + "spans": [ + { + "bbox": [ + 55, + 460, + 192, + 473 + ], + "type": "text", + "content": "4.3. Iterative Editing Results" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 478, + 296, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 478, + 296, + 598 + ], + "spans": [ + { + "bbox": [ + 55, + 478, + 296, + 598 + ], + "type": "text", + "content": "In this section, we present the iterative editing results. As demonstrated in Figure 6, users can progressively refine the SVG through multiple rounds of natural language instructions. During editing, we preserve the optimized shapes from previous iterations while only modifying the specified shapes. The results showcase how Chat2SVG can accurately interpret and execute user editing requests, maintaining both semantic coherence and visual quality throughout the modifications. Additional editing examples are provided in the supplementary material." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 609, + 128, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 609, + 128, + 621 + ], + "spans": [ + { + "bbox": [ + 55, + 609, + 128, + 621 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 629, + 296, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 296, + 700 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 296, + 700 + ], + "type": "text", + "content": "In this paper, we present Chat2SVG, a novel paradigm for text-to-SVG generation that combines LLMs and image diffusion models. Through our carefully designed SVG template generation and dual-stage SVG optimization pipeline, our method generates high-quality SVG outputs that exhibit strong visual fidelity, path regularity, and text alignment." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "type": "text", + "content": "Despite our method's superior results, there are several" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 328, + 70, + 541, + 274 + ], + "blocks": [ + { + "bbox": [ + 328, + 70, + 541, + 274 + ], + "lines": [ + { + "bbox": [ + 328, + 70, + 541, + 274 + ], + "spans": [ + { + "bbox": [ + 328, + 70, + 541, + 274 + ], + "type": "image", + "image_path": "62cd837d2b193eecf7c28ab864031b4f738964edd919c617dec2e4e805051e96.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 282, + 555, + 316 + ], + "lines": [ + { + "bbox": [ + 313, + 282, + 555, + 316 + ], + "spans": [ + { + "bbox": [ + 313, + 282, + 555, + 316 + ], + "type": "text", + "content": "Figure 6. Iterative Editing. We perform two rounds of refinement on each SVG template and show the optimized output. This figure shows editing types including deletion, modification, and addition." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 328, + 328, + 389, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 328, + 389, + 337 + ], + "spans": [ + { + "bbox": [ + 328, + 328, + 389, + 337 + ], + "type": "text", + "content": "(1) Layout Precision" + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 328, + 338, + 388, + 369 + ], + "blocks": [ + { + "bbox": [ + 328, + 338, + 388, + 369 + ], + "lines": [ + { + "bbox": [ + 328, + 338, + 388, + 369 + ], + "spans": [ + { + "bbox": [ + 328, + 338, + 388, + 369 + ], + "type": "image", + "image_path": "f7ea2084ffb93095b4075d8ab0d9fb1bd80874dbcee42fe466a7dcc3b0f709e9.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 341, + 388, + 373, + 423 + ], + "blocks": [ + { + "bbox": [ + 328, + 372, + 383, + 388 + ], + "lines": [ + { + "bbox": [ + 328, + 372, + 383, + 388 + ], + "spans": [ + { + "bbox": [ + 328, + 372, + 383, + 388 + ], + "type": "text", + "content": "SVG Target Template Image" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 341, + 388, + 373, + 423 + ], + "lines": [ + { + "bbox": [ + 341, + 388, + 373, + 423 + ], + "spans": [ + { + "bbox": [ + 341, + 388, + 373, + 423 + ], + "type": "image", + "image_path": "20e7459baa524fb618e1a24a76b23f9ed3bf25bd6115da62d5709794501aae16.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 333, + 424, + 381, + 433 + ], + "lines": [ + { + "bbox": [ + 333, + 424, + 381, + 433 + ], + "spans": [ + { + "bbox": [ + 333, + 424, + 381, + 433 + ], + "type": "text", + "content": "Optimized SVG" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 399, + 328, + 452, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 399, + 328, + 452, + 337 + ], + "spans": [ + { + "bbox": [ + 399, + 328, + 452, + 337 + ], + "type": "text", + "content": "(2) Path Addition" + } + ] + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 392, + 338, + 457, + 371 + ], + "blocks": [ + { + "bbox": [ + 392, + 338, + 457, + 371 + ], + "lines": [ + { + "bbox": [ + 392, + 338, + 457, + 371 + ], + "spans": [ + { + "bbox": [ + 392, + 338, + 457, + 371 + ], + "type": "image", + "image_path": "18979784f96b2fb99b328c336ee0d238a340c6532366dc223f1e496b0d2de66a.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 408, + 388, + 444, + 423 + ], + "blocks": [ + { + "bbox": [ + 398, + 372, + 452, + 388 + ], + "lines": [ + { + "bbox": [ + 398, + 372, + 452, + 388 + ], + "spans": [ + { + "bbox": [ + 398, + 372, + 452, + 388 + ], + "type": "text", + "content": "SVG Target Template Image" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 408, + 388, + 444, + 423 + ], + "lines": [ + { + "bbox": [ + 408, + 388, + 444, + 423 + ], + "spans": [ + { + "bbox": [ + 408, + 388, + 444, + 423 + ], + "type": "image", + "image_path": "d6b078a9667f664bd408f16b9507351cbb3729016cd93fc7b282da93fd66253f.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 401, + 424, + 449, + 433 + ], + "lines": [ + { + "bbox": [ + 401, + 424, + 449, + 433 + ], + "spans": [ + { + "bbox": [ + 401, + 424, + 449, + 433 + ], + "type": "text", + "content": "Optimized SVG" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "bbox": [ + 468, + 328, + 533, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 468, + 328, + 533, + 337 + ], + "spans": [ + { + "bbox": [ + 468, + 328, + 533, + 337 + ], + "type": "text", + "content": "(3) Shape Complexity" + } + ] + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 465, + 338, + 504, + 372 + ], + "blocks": [ + { + "bbox": [ + 465, + 338, + 504, + 372 + ], + "lines": [ + { + "bbox": [ + 465, + 338, + 504, + 372 + ], + "spans": [ + { + "bbox": [ + 465, + 338, + 504, + 372 + ], + "type": "image", + "image_path": "fb5f5c36a016363d6cf9d58799a2c8df46c0fb0803a737a7cf4920860c2da95a.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 469, + 372, + 498, + 388 + ], + "lines": [ + { + "bbox": [ + 469, + 372, + 498, + 388 + ], + "spans": [ + { + "bbox": [ + 469, + 372, + 498, + 388 + ], + "type": "text", + "content": "SVG Template" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 504, + 338, + 541, + 372 + ], + "blocks": [ + { + "bbox": [ + 504, + 338, + 541, + 372 + ], + "lines": [ + { + "bbox": [ + 504, + 338, + 541, + 372 + ], + "spans": [ + { + "bbox": [ + 504, + 338, + 541, + 372 + ], + "type": "image", + "image_path": "635ee330f3881215cb11e15b074e60f1cca7990a71f9773ccb73424e5c8ba303.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 514, + 372, + 533, + 388 + ], + "lines": [ + { + "bbox": [ + 514, + 372, + 533, + 388 + ], + "spans": [ + { + "bbox": [ + 514, + 372, + 533, + 388 + ], + "type": "text", + "content": "Target Image" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 468, + 388, + 507, + 424 + ], + "blocks": [ + { + "bbox": [ + 468, + 388, + 507, + 424 + ], + "lines": [ + { + "bbox": [ + 468, + 388, + 507, + 424 + ], + "spans": [ + { + "bbox": [ + 468, + 388, + 507, + 424 + ], + "type": "image", + "image_path": "8a028051f52b1a74157cfd3cf71eab3726afcc148799ce13b06765b8ff914ea6.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 481, + 424, + 528, + 433 + ], + "lines": [ + { + "bbox": [ + 481, + 424, + 528, + 433 + ], + "spans": [ + { + "bbox": [ + 481, + 424, + 528, + 433 + ], + "type": "text", + "content": "Optimized SVG" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 441, + 555, + 475 + ], + "lines": [ + { + "bbox": [ + 313, + 441, + 555, + 475 + ], + "spans": [ + { + "bbox": [ + 313, + 441, + 555, + 475 + ], + "type": "text", + "content": "Figure 7. Limitations of our method include imprecise layout generation, missing important visual elements, and insufficient shape complexity." + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 27 + }, + { + "bbox": [ + 313, + 498, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 555, + 713 + ], + "type": "text", + "content": "aspects that could be further improved: (1) LLMs may produce imprecise visual layouts, such as the burger in Figure 7. This can be alleviated by collecting an SVG dataset and extracting the corresponding layouts to fine-tune LLMs. (2) Our approach uses SAM to identify new parts. However, SAM may overlook important regions, like the pink frosting on the donut in Figure 7. Incorporating semantic-aware SAM [19] could potentially produce more accurate results. (3) The pretrained SVG VAE from [50] has a fixed number of commands. This makes it hard to represent intricate shapes. For example, the lion's mane in Figure 7 lacks sufficient detail. Training a new SVG VAE with varying-length commands would help solve this problem. Our work can be extended to generate other types of vectorized data, such as icons and fonts. Additionally, through consistent editing of SVG across multiple frames, coupled with shape interpolation, our Chat2SVG has the potential to facilitate keyframe animation generation." + } + ] + } + ], + "index": 30 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23697" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 156, + 115, + 169 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 156, + 115, + 169 + ], + "spans": [ + { + "bbox": [ + 56, + 156, + 115, + 169 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 177, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 177, + 296, + 220 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 177, + 296, + 220 + ], + "spans": [ + { + "bbox": [ + 61, + 177, + 296, + 220 + ], + "type": "text", + "content": "[1] Panos Achlioptas, Olga Diamanti, Ioannis Mitliagkas, and Leonidas Guibas. Learning representations and generative models for 3D point clouds. In International Conference on Machine Learning, pages 40-49, 2018. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 222, + 296, + 287 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 222, + 296, + 287 + ], + "spans": [ + { + "bbox": [ + 61, + 222, + 296, + 287 + ], + "type": "text", + "content": "[2] Rio Aguina-Kang, Maxim Gumin, Do Heon Han, Stewart Morris, Seung-Jean Yoo, Aditya Ganeshan, R. Kenny Jones, Qiuhong Anna Wei, Kailiang Fu, and Daniel Ritchie. Open-universe indoor scene generation using LLM program synthesis and uncurated object databases. In arXiv preprint arXiv:2403.09675, 2024. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 289, + 295, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 289, + 295, + 331 + ], + "spans": [ + { + "bbox": [ + 62, + 289, + 295, + 331 + ], + "type": "text", + "content": "[3] Jonas Belouadi, Anne Lauscher, and Steffen Eger. AutomataTikZ: Text-guided synthesis of scientific vector graphics with TikZ. In arXiv preprint arXiv:2310.00367, 2023. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 334, + 295, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 334, + 295, + 388 + ], + "spans": [ + { + "bbox": [ + 62, + 334, + 295, + 388 + ], + "type": "text", + "content": "[4] Alexandre Carlier, Martin Danelljan, Alexandre Alahi, and Radu Timofte. DeepSVG: A hierarchical generative network for vector graphics animation. In Advances in Neural Information Processing Systems, pages 16351-16361, 2020. 3, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 389, + 296, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 389, + 296, + 432 + ], + "spans": [ + { + "bbox": [ + 62, + 389, + 296, + 432 + ], + "type": "text", + "content": "[5] Yutao Cheng, Zhao Zhang, Maoke Yang, Hui Nie, Chunyuan Li, Xinglong Wu, and Jie Shao. Graphic design with large multimodal model. In arXiv preprint arXiv:2404.14368, 2024. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 434, + 295, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 434, + 295, + 465 + ], + "spans": [ + { + "bbox": [ + 62, + 434, + 295, + 465 + ], + "type": "text", + "content": "[6] Louis Clouatre and Marc Demers. FIgR: Few-shot image generation with reptile. In arXiv preprint arXiv:1901.02199, 2019. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 468, + 295, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 468, + 295, + 512 + ], + "spans": [ + { + "bbox": [ + 62, + 468, + 295, + 512 + ], + "type": "text", + "content": "[7] Haoqiang Fan, Hao Su, and Leonidas Guibas. A point set generation network for 3D object reconstruction from a single image. In IEEE Conference on Computer Vision and Pattern Recognition, pages 605-613, 2017. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 514, + 295, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 514, + 295, + 577 + ], + "spans": [ + { + "bbox": [ + 62, + 514, + 295, + 577 + ], + "type": "text", + "content": "[8] Weixi Feng, Wanrong Zhu, Tsu-jui Fu, Varun Jampani, Arjun Akula, Xuehai He, Sugato Basu, Xin Wang, and William Wang. LayoutGPT: Compositional visual planning and generation with large language models. In Advances in Neural Information Processing Systems, pages 18225-18250, 2024. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 580, + 295, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 580, + 295, + 624 + ], + "spans": [ + { + "bbox": [ + 62, + 580, + 295, + 624 + ], + "type": "text", + "content": "[9] Kevin Frans, Lisa B. Soros, and Olaf Witkowski. CLIPDraw: Exploring text-to-drawing synthesis through language-image encoders. In Advances in Neural Information Processing Systems, pages 5207-5218, 2021. 2, 3, 4, 5, 6, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 625, + 295, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 625, + 295, + 668 + ], + "spans": [ + { + "bbox": [ + 57, + 625, + 295, + 668 + ], + "type": "text", + "content": "[10] Rao Fu, Zehao Wen, Zichen Liu, and Srinath Sridhar. AnyHome: Open-vocabulary generation of structured and textured 3D homes. In European Conference on Computer Vision, pages 52-70, 2025. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 670, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 295, + 712 + ], + "type": "text", + "content": "[11] Aditya Ganeshan, Ryan Y. Huang, Xianghao Xu, R. Kenny Jones, and Daniel Ritchie. ParSEL: Parameterized shape editing with language. In arXiv preprint arXiv:2405.20319, 2024. 2" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 57, + 73, + 554, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 55, + 90, + 296, + 136 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 90, + 296, + 136 + ], + "spans": [ + { + "bbox": [ + 55, + 90, + 296, + 136 + ], + "type": "text", + "content": "The work described in this paper was fully supported by a GRF grant from the Research Grants Council (RGC) of the Hong Kong Special Administrative Region, China [Project No. CityU 11216122]." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 316, + 73, + 554, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 105 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 105 + ], + "type": "text", + "content": "[12] David Ha and Douglas Eck. A neural representation of sketch drawings. In arXiv preprint arXiv:1704.03477, 2017. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 107, + 554, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 107, + 554, + 162 + ], + "spans": [ + { + "bbox": [ + 317, + 107, + 554, + 162 + ], + "type": "text", + "content": "[13] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. GANs trained by a two time-scale update rule converge to a local Nash equilibrium. In Advances in Neural Information Processing Systems, pages 6629-6640, 2017. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 164, + 554, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 164, + 554, + 218 + ], + "spans": [ + { + "bbox": [ + 317, + 164, + 554, + 218 + ], + "type": "text", + "content": "[14] Ziniu Hu, Ahmet Iscen, Aashi Jain, Thomas Kipf, Yisong Yue, David A. Ross, Cordelia Schmid, and Alireza Fathi. SceneCraft: An LLM agent for synthesizing 3D scenes as Blender code. In International Conference on Machine Learning, pages 19252-19282, 2024. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 220, + 553, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 220, + 553, + 252 + ], + "spans": [ + { + "bbox": [ + 316, + 220, + 553, + 252 + ], + "type": "text", + "content": "[15] Ian Huang, Guandao Yang, and Leonidas Guibas. Blender: Editing 3D graphics with vision-language models. In arXiv preprint arXiv:2404.17672, 2024. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 254, + 553, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 254, + 553, + 297 + ], + "spans": [ + { + "bbox": [ + 317, + 254, + 553, + 297 + ], + "type": "text", + "content": "[16] Ajay Jain, Amber Xie, and Pieter Abbeel. VectorFusion: Text-to-SVG by abstracting pixel-based diffusion models. In IEEE Conference on Computer Vision and Pattern Recognition, pages 1911-1920, 2022. 2, 3, 4, 5, 6, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 299, + 553, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 299, + 553, + 354 + ], + "spans": [ + { + "bbox": [ + 317, + 299, + 553, + 354 + ], + "type": "text", + "content": "[17] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dólar, and Ross Girshick. Segment anything. In IEEE International Conference on Computer Vision, pages 4015-4026, 2023. 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 355, + 553, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 355, + 553, + 397 + ], + "spans": [ + { + "bbox": [ + 317, + 355, + 553, + 397 + ], + "type": "text", + "content": "[18] Peter Kulits, Haiwen Feng, Weiyang Liu, Victoria Abrevaya, and Michael J. Black. Re-thinking inverse graphics with large language models. In arXiv preprint arXiv:2404.15228, 2024. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 399, + 553, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 399, + 553, + 453 + ], + "spans": [ + { + "bbox": [ + 317, + 399, + 553, + 453 + ], + "type": "text", + "content": "[19] Feng Li, Hao Zhang, Peize Sun, Xueyan Zou, Shilong Liu, Jianwei Yang, Chunyuan Li, Lei Zhang, and Jianfeng Gao. Semantic-SAM: Segment and recognize anything at any granularity. In arXiv preprint arXiv:2307.04767, 2023. 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 456, + 553, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 456, + 553, + 499 + ], + "spans": [ + { + "bbox": [ + 317, + 456, + 553, + 499 + ], + "type": "text", + "content": "[20] Tzu-Mao Li, Michal Lukáč, Michael Gharbi, and RaganKelley Jonathan. Differentiable vector graphics rasterization for editing and learning. ACM Transactions on Graphics, 39 (6):1-15, 2020. 2, 3, 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 501, + 553, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 501, + 553, + 555 + ], + "spans": [ + { + "bbox": [ + 317, + 501, + 553, + 555 + ], + "type": "text", + "content": "[21] Jiawei Lin, Jiaqi Guo, Shizhao Sun, Zijiang Yang, JianGuang Lou, and Dongmei Zhang. LayoutPrompter: Awaken the design ability of large language models. In Advances in Neural Information Processing Systems, pages 43852-43879, 2024. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 557, + 553, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 557, + 553, + 610 + ], + "spans": [ + { + "bbox": [ + 317, + 557, + 553, + 610 + ], + "type": "text", + "content": "[22] Vivian Liu, Rubaiat H. Kazi, Li-Yi Wei, Matthew Fisher, Timothy Langlois, Seth Walker, and Lydia Chilton. LogoMotion: Visually grounded code generation for content-aware animation. In arXiv preprint arXiv:2405.07065, 2024. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 613, + 553, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 613, + 553, + 668 + ], + "spans": [ + { + "bbox": [ + 317, + 613, + 553, + 668 + ], + "type": "text", + "content": "[23] YingTian Liu, Zhifei Zhang, YuanChen Guo, Matthew Fisher, Zhaowen Wang, and SongHai Zhang. DualVector: Unsupervised vector font synthesis with dual-part representation. In IEEE Conference on Computer Vision and Pattern Recognition, pages 14193-14202, 2023. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "text", + "content": "[24] Raphael G. Lopes, David Ha, Douglas Eck, and Jonathon Shlens. A learned representation for scalable vector graphics. In IEEE International Conference on Computer Vision, pages 7930-7939, 2019. 3" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23698" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 117 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 117 + ], + "type": "text", + "content": "[25] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia-jun Wu, JunYan Zhu, and Stefano Ermon. SDEdit: Guided image synthesis and editing with stochastic differential equations. In arXiv preprint arXiv:2108.01073, 2021. 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 119, + 295, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 119, + 295, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 119, + 295, + 140 + ], + "type": "text", + "content": "[26] OpenAI. GPT-4 technical report. In arXiv preprint arXiv:2303.08774, 2023. 2, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 143, + 295, + 176 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 143, + 295, + 176 + ], + "spans": [ + { + "bbox": [ + 56, + 143, + 295, + 176 + ], + "type": "text", + "content": "[27] Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. DreamFusion: Text-to-3D using 2D diffusion. In arXiv preprint arXiv:2209.14988, 2022. 2, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 178, + 295, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 178, + 295, + 255 + ], + "spans": [ + { + "bbox": [ + 56, + 178, + 295, + 255 + ], + "type": "text", + "content": "[28] Alec Radford, Jong-Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763, 2021. 2, 3, 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 257, + 295, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 257, + 295, + 312 + ], + "spans": [ + { + "bbox": [ + 56, + 257, + 295, + 312 + ], + "type": "text", + "content": "[29] Hamid Rezatofighi, Nathan Tsoi, JunYoung Gwak, Amir Sadeghian, Ian Reid, and Silvio Savarese. Generalized intersection over union: A metric and a loss for bounding box regression. In IEEE Conference on Computer Vision and Pattern Recognition, pages 658-666, 2019. 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 314, + 295, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 314, + 295, + 358 + ], + "spans": [ + { + "bbox": [ + 56, + 314, + 295, + 358 + ], + "type": "text", + "content": "[30] Leo S. F. Ribeiro, Tu Bui, John Collomosse, and Moacir Ponti. Sketchformer: Transformer-based representation for sketched structure. In IEEE Conference on Computer Vision and Pattern Recognition, pages 14153-14162, 2020. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 360, + 295, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 360, + 295, + 415 + ], + "spans": [ + { + "bbox": [ + 56, + 360, + 295, + 415 + ], + "type": "text", + "content": "[31] Pratyusha Sharma, Tamar Rott Shaham, Manel Baradad, Stephanie Fu, Adrian Rodriguez-Munoz, Shivam Duggal, Phillip Isola, and Antonio Torralba. A vision check-up for language models. In IEEE Conference on Computer Vision and Pattern Recognition, pages 14410-14419, 2024. 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 417, + 295, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 417, + 295, + 460 + ], + "spans": [ + { + "bbox": [ + 56, + 417, + 295, + 460 + ], + "type": "text", + "content": "[32] Chunyi Sun, Junlin Han, Weijian Deng, Xinlong Wang, Zishan Qin, and Stephen Gould. 3D-GPT: Procedural 3D modeling with large language models. In arXiv preprint arXiv:2310.12945, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 463, + 295, + 517 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 463, + 295, + 517 + ], + "spans": [ + { + "bbox": [ + 56, + 463, + 295, + 517 + ], + "type": "text", + "content": "[33] Hou In Ivan Tam, Hou In Derek Pun, Austin T. Wang, Angel X. Chang, and Manolis Savva. SceneMotifCoder: Example-driven visual program learning for generating 3D object arrangements. In arXiv preprint arXiv:2408.02211, 2024. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 520, + 295, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 520, + 295, + 563 + ], + "spans": [ + { + "bbox": [ + 56, + 520, + 295, + 563 + ], + "type": "text", + "content": "[34] Zecheng Tang, Chenfei Wu, Juntao Li, and Nan Duan. LayoutNUWA: Revealing the hidden layout expertise of large language models. In arXiv preprint arXiv:2309.09506, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 566, + 295, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 566, + 295, + 621 + ], + "spans": [ + { + "bbox": [ + 56, + 566, + 295, + 621 + ], + "type": "text", + "content": "[35] Zecheng Tang, Chenfei Wu, Zekai Zhang, Mingheng Ni, Shengming Yin, Yu Liu, Zhengyuan Yang, Lijuan Wang, Zicheng Liu, Juntao Li, and Nan Duan. StrokeNUWA: Tokenizing strokes for vector graphic synthesis. In arXiv preprint arXiv:2401.17093, 2024. 2, 3, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 623, + 295, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 623, + 295, + 666 + ], + "spans": [ + { + "bbox": [ + 56, + 623, + 295, + 666 + ], + "type": "text", + "content": "[36] Tiffany Tseng, Ruijia Cheng, and Jeffrey Nichols. Keyframer: Empowering animation design using large language models. In arXiv preprint arXiv:2402.06071, 2024. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 669, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 669, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 669, + 295, + 713 + ], + "type": "text", + "content": "[37] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems, pages 6000–6010, 2017. 3" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 117 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 117 + ], + "type": "text", + "content": "[38] Can Wang, Hongliang Zhong, Menglei Chai, Mingming He, Dongdong Chen, and Jing Liao. Chat2Layout: Interactive 3D furniture layout with a multimodal LLM. In arXiv preprint arXiv:2407.21333, 2024. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 118, + 553, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 553, + 151 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 553, + 151 + ], + "type": "text", + "content": "[39] Yizhi Wang and Zhouhui Lian. DeepVecFont: Synthesizing high-quality vector fonts via dual-modality learning. ACM Transactions on Graphics, 40(6):1-15, 2021. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 152, + 553, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 553, + 206 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 553, + 206 + ], + "type": "text", + "content": "[40] Yuqing Wang, Yizhi Wang, Longhui Yu, Yuesheng Zhu, and Zhouhui Lian. DeepVecFont-v2: Exploiting Transformers to synthesize vector fonts with higher quality. In IEEE Conference on Computer Vision and Pattern Recognition, pages 18320-18328, 2023. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 208, + 553, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 208, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 316, + 208, + 553, + 251 + ], + "type": "text", + "content": "[41] Ronghuan Wu, Wanchao Su, Kede Ma, and Jing Liao. IconShop: Text-guided vector icon synthesis with autoregressive Transformers. ACM Transactions on Graphics, 42(6):1-14, 2023. 2, 3, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 253, + 553, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 253, + 553, + 308 + ], + "spans": [ + { + "bbox": [ + 316, + 253, + 553, + 308 + ], + "type": "text", + "content": "[42] Tong Wu, Guandao Yang, Zhibing Li, Kai Zhang, Ziwei Liu, Leonidas Guibas, Dahua Lin, and Gordon Wetzstein. GPT-4V (ision) is a human-aligned evaluator for text-to-3D generation. In IEEE Conference on Computer Vision and Pattern Recognition, pages 22227–22238, 2024. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 309, + 553, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 309, + 553, + 363 + ], + "spans": [ + { + "bbox": [ + 316, + 309, + 553, + 363 + ], + "type": "text", + "content": "[43] Ximing Xing, Chuang Wang, Haitao Zhou, Jing Zhang, Qian Yu, and Dong Xu. DiffSketcher: Text guided vector sketch synthesis through latent diffusion models. In Advances in Neural Information Processing Systems, pages 15869-15889, 2023. 2, 3, 4, 5, 6, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 365, + 553, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 365, + 553, + 420 + ], + "spans": [ + { + "bbox": [ + 316, + 365, + 553, + 420 + ], + "type": "text", + "content": "[44] Ximing Xing, Haitao Zhou, Chuang Wang, Jing Zhang, Dong Xu, and Qian Yu. SVGDreamer: Text guided SVG generation with diffusion model. In IEEE Conference on Computer Vision and Pattern Recognition, pages 4546-4555, 2024. 2, 3, 4, 5, 6, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 421, + 553, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 421, + 553, + 476 + ], + "spans": [ + { + "bbox": [ + 316, + 421, + 553, + 476 + ], + "type": "text", + "content": "[45] Jiazheng Xu, Xiao Liu, Yuchen Wu, Yuxuan Tong, Qinkai Li, Ming Ding, Jie Tang, and Yuxiao Dong. ImageReward: Learning and evaluating human preferences for text-to-image generation. In Advances in Neural Information Processing Systems, pages 15903-15935, 2024. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 477, + 553, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 477, + 553, + 521 + ], + "spans": [ + { + "bbox": [ + 316, + 477, + 553, + 521 + ], + "type": "text", + "content": "[46] Yutaro Yamada, Khyathi Chandu, Yuchen Lin, Jack Hessel, Ilker Yildirim, and Yejin Choi. L3GO: Language agents with chain-of-3D-thoughts for generating unconventional objects. In arXiv preprint arXiv:2402.09052, 2024. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 522, + 553, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 522, + 553, + 565 + ], + "spans": [ + { + "bbox": [ + 316, + 522, + 553, + 565 + ], + "type": "text", + "content": "[47] Tao Yang, Yingmin Luo, Zhongang Qi, Yang Wu, Ying Shan, and Chang Wen Chen. PosterLLaVa: Constructing a unified multi-modal layout generator with LLM. In arXiv preprint arXiv:2406.02884, 2024. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 567, + 553, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 567, + 553, + 611 + ], + "spans": [ + { + "bbox": [ + 316, + 567, + 553, + 611 + ], + "type": "text", + "content": "[48] Lequan Yu, Xianzhi Li, Chi-Wing Fu, Daniel Cohen-Or, and Pheng-Ann Heng. PU-Net: Point cloud upsampling network. In IEEE Conference on Computer Vision and Pattern Recognition, pages 2790–2799, 2018. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 613, + 553, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 613, + 553, + 655 + ], + "spans": [ + { + "bbox": [ + 316, + 613, + 553, + 655 + ], + "type": "text", + "content": "[49] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In IEEE International Conference on Computer Vision, pages 3836-3847, 2023. 4" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 657, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 657, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 657, + 553, + 689 + ], + "type": "text", + "content": "[50] Peiying Zhang, Nanxuan Zhao, and Jing Liao. Text-to-vector generation with neural path representation. ACM Transactions on Graphics, 43(4):1-13, 2024. 2, 3, 4, 6, 7, 8" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "text", + "content": "[51] Qihang Zhang, Chaoyang Wang, Aliaksandr Siarohin, Peiye Zhuang, Yinghao Xu, Ceyuan Yang, Dahua Lin, Bolei Zhou," + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "23699" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 75, + 72, + 297, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 72, + 297, + 108 + ], + "spans": [ + { + "bbox": [ + 75, + 72, + 297, + 108 + ], + "type": "text", + "content": "Sergey Tulyakov, and Hsin-Ying Lee. SceneWiz3D: Towards text-guided 3D scene composition. In arXiv preprint arXiv:2312.08885, 2023. 2" + } + ] + } + ], + "index": 0 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "23700" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/ChatGarment_ Garment Estimation, Generation and Editing via Large Language Models/425b09e1-924e-4a43-b277-eb8a41387d21_content_list.json b/2025/ChatGarment_ Garment Estimation, Generation and Editing via Large Language Models/425b09e1-924e-4a43-b277-eb8a41387d21_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..9d357db1a5e0c8ed81331fb0b1d7ff27d4855a0a --- /dev/null +++ b/2025/ChatGarment_ Garment Estimation, Generation and Editing via Large Language Models/425b09e1-924e-4a43-b277-eb8a41387d21_content_list.json @@ -0,0 +1,1769 @@ +[ + { + "type": "text", + "text": "ChatGarment: Garment Estimation, Generation and Editing via Large Language Models", + "text_level": 1, + "bbox": [ + 191, + 130, + 808, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Siyuan Bian $^{1,2}$ Chenghao Xu $^{1,3}$ Yuliang Xiu $^{1,4}$ Artur Grigorev $^{1,5}$ Zhen Liu $^{1,6}$ Cewu Lu $^{2}$ Michael J. Black $^{1}$ Yao Feng $^{7,8,\\dagger}$", + "bbox": [ + 155, + 200, + 841, + 243 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Max Planck Institute for Intelligent Systems $^{2}$ Shanghai Jiao Tong University", + "bbox": [ + 181, + 246, + 815, + 266 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{3}$ EPFL $^{4}$ Westlake University $^{5}$ ETH Zürich $^{6}$ Mila, University of Montreal $^{7}$ Meshcapade $^{8}$ Stanford University", + "bbox": [ + 181, + 266, + 815, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "chatgarment.github.io", + "bbox": [ + 401, + 305, + 589, + 318 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/abd2410ab08bf1b0d7bd21125ec50641268779933d62491b99f6ccfad435656e.jpg", + "image_caption": [ + "Figure 1. As a multimodal 3D garment creator, ChatGarment understands both images and language. It can estimate complex 3D garments represented as sewing patterns from a single image, which can be easily animated and simulated. It also supports garment editing based on text instructions." + ], + "image_footnote": [], + "bbox": [ + 89, + 348, + 509, + 513 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e47b86281e490117fe4b90dff48e9929045d0f0b0a9a4e34ea853b57c70dc216.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 519, + 348, + 906, + 513 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 561, + 326, + 577 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We introduce ChatGarment, a novel approach that leverages large vision-language models (VLMs) to automate the estimation, generation, and editing of 3D garments from images or text descriptions. Unlike previous methods that struggle in real-world scenarios or lack interactive editing capabilities, ChatGarment can estimate sewing patterns from in-the-wild images or sketches, generate them from text descriptions, and edit garments based on user instructions, all within an interactive dialogue. These sewing patterns can then be draped on a 3D body and animated. This is achieved by finetuning a VLM to directly generate a JSON file that includes both textual descriptions of garment types and styles, as well as continuous numerical attributes. This JSON file is then used to create sewing patterns through a programming parametric model. To support this, we refine the existing programming model, GarmentCode, by expanding its garment type coverage and simplifying its structure for efficient VLM fine-tuning. Additionally, we construct a", + "bbox": [ + 86, + 595, + 485, + 868 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "large-scale dataset of image-to-sewing-pattern and text-to-sewing-pattern pairs through an automated data pipeline. Extensive evaluations demonstrate ChatGarment's ability to accurately reconstruct, generate, and edit garments from multimodal inputs, highlighting its potential to simplify workflows in fashion and gaming applications. Code and data are available at https://chatgarment.github.io/.", + "bbox": [ + 511, + 563, + 908, + 669 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 679, + 643, + 694 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Imagine seeing a photo of someone in a nice looking outfit that you would like for yourself or your 3D avatar. Our goal is to take that photo and convert it into a 2D sewing pattern and then enable the user to use language to edit the pattern, for example, making the sleeves longer or editing the neckline. We then drape the garment on a 3D body and animate it. This is illustrated in Fig. 1. With this approach, given just an image and/or text, we construct sewing patterns for 3D garments that can be immediately used in fashion or gaming applications. This lightweight process is in stark contrast to how sewing patterns and 3D garments are traditionally created by an artist or clothing designer. In contrast to the traditional labor-intensive process, our approach exploits", + "bbox": [ + 509, + 703, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "This work was done while YF was at Meshcapade.", + "bbox": [ + 109, + 885, + 383, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "2924", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "large vision-language models (VLMs) for the first time to democratize the capture and design of clothing.", + "bbox": [ + 89, + 90, + 485, + 122 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recent work explores the generation or estimation of 3D clothing using text and/or images as input [6, 8, 21, 42]. These methods use various 3D representations including 3D meshes, point clouds, or implicit representations like Unsigned Distance Functions (UDFs) and Signed Distance Functions (SDFs). We observe, however, that most clothing is designed and manufactured using 2D patterns, and posit that this is a more natural representation. Sewing patterns can be seamlessly integrated into existing garment design pipelines for animation and manufacturing. Recent approaches [16, 30, 33] use sewing patterns for garment estimation and generation, but mapping images or text to sewing patterns remains challenging due to limited training data, leading to poor generalization across diverse clothing types. These methods also does not support interactive editing, making post-processing labor-intensive for 3D artists. Fortunately, recent work, such as GarmentCode [24, 25], provides a programmatic framework for encoding garment information into a JSON file. This JSON file includes textual descriptions of garment types, styles and numerical attributes, offering a rich representation of sewing patterns. The JSON file is then processed by the GarmentCode framework to produce 2D sewing patterns, which can subsequently be draped on a 3D body. However, to date, there is no automated method to directly extract a GarmentCode representation (JSON file) from an image or text description.", + "bbox": [ + 91, + 125, + 485, + 517 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our observation is that a JSON file resembles natural language, making it suitable for interpretation by large language models (LLMs). At the same time, Vision-LLMs (VLMs) excel at understanding images and garment concepts, such as types of garments and specific features like long sleeves. If we enable the VLM to translate its garment understanding into the structured format of GarmentCode, then it becomes capable of generating sewing patterns. This allows the VLM to take an image as input and output a garment, while also supporting garment generation and editing through text descriptions or interactive dialogue.", + "bbox": [ + 89, + 520, + 485, + 686 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Specifically, we introduce ChatGarment, a method that finetunes a VLM to take text queries, optionally with images, as input and outputs a JSON garment file. This JSON file is then used for GarmentCode generation. The JSON file includes textual descriptions of garment types (e.g., skirts, pants), styles (e.g., collar type), and continuous numerical attributes (e.g., pant length). Inspired by previous work [11, 14, 26], we introduce a new token () to represent these numerical values and train an MLP projection layer to decode them from the token embedding. Those numbers, combined with the VLM's language output, are formatted into the final JSON file. In our approach, we keep the VLM's vision encoder fixed, fine-tune the language model using LoRA [18], and jointly train the MLP projection layer. To", + "bbox": [ + 89, + 688, + 485, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "adapt GarmentCode [24] for more efficient VLM training, we create an improved version that supports a wider range of clothing types (e.g. open-jacket) and removes unnecessary settings to reduce token usage in training. We also develop an automated data construction pipeline that leverages existing tools for garment generation, draping, simulation, rendering, and automatic labeling with GPT-4o. This pipeline enables building a dataset with image-to-GarmentCode and text-to-GarmentCode pairs, including 20K new garments and 1M images with detailed descriptions, supporting both garment creation and editing.", + "bbox": [ + 511, + 90, + 906, + 256 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We evaluate ChatGarment across a diverse set of tasks, including those specifically trained for, and novel applications that demonstrate the generalization capabilities of VLMs. For single-image garment reconstruction, ChatGarment outperforms prior sewing-pattern-specific and LLM-based models on the Dress4D [54] and CloSE [2] evaluation datasets. Additionally, we demonstrate garment editing, generation, and multi-turn dialogue, highlighting our approach's versatility. ChatGarment can flexibly use text and images to create and edit 3D garments that can be readily animated, introducing a new workflow for garment design.", + "bbox": [ + 511, + 257, + 908, + 422 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We summarize our contributions as follows:", + "bbox": [ + 535, + 424, + 823, + 436 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- A unified model capable of estimating, generating and editing sewing patterns from multimodal inputs.", + "- The first approach to leverage VLMs for directly generating JSON files for garment creation.", + "- A refined version of GarmentCode that supports more diverse garment types and is optimized for VLM training.", + "- An automatic data construction pipeline for generating image-to-sewing-pattern and text-to-sewing-pattern data, along with the release of a large-scale dataset." + ], + "bbox": [ + 511, + 439, + 906, + 574 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 588, + 653, + 604 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Garment Creation. Most 3D garments have predefined templates (e.g., sewing patterns) and categories (e.g., skirts, shirts, pants, dresses, etc.), but their non-rigid nature makes them highly deformable with dynamic motions. Efforts to create 3D garments can be grouped into \"template-based\" or \"free-form\" methods, based on their 3D representations.", + "bbox": [ + 511, + 613, + 906, + 703 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "A line of template-based approaches register predefined garment templates to 2D observations, which could be real captures, e.g. from monocular video [5, 21, 36, 42] and images [3, 6, 7, 62], generated data [49], or derived from the \"proxy gradient\" computed from 2D diffusion models [27, 47] via score distillation sampling (SDS) [41]. When the template is classified correctly, the predefined garment template, refined through a coarse-to-fine step, ensures topological stability and correctness. Unfortunately, template-based methods struggle to accurately model garments that deviate from the predefined templates, particularly those with diverse styles, such as long skirts, or garments that undergo topological changes, like opening a jacket.", + "bbox": [ + 511, + 704, + 908, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2925", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address these limitations, free-form representations like (un)signed distance functions [6, 8, 50, 59], occupancy fields [45, 46, 56, 57], tetrahedrons [20, 35, 48], and point clouds [37, 38] are leveraged. However, this flexibility can compromise the inherent structure of garments, potentially leading to non-garment shapes and floating artifacts. Without shape regularization, incomplete observations lead to the reconstructed unseen parts being overly smooth. Even worse, rigging or simulating these free-form garments becomes extremely challenging.", + "bbox": [ + 89, + 90, + 483, + 243 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The sewing pattern is a well-balanced solution – with pre-defined yet flexible templates. Several methods exploit sewing patterns to reconstruct 3D garments. MulayCap [51] estimates the parameters of a sewing pattern from a monocular video. ISP [30] parameterizes sewing patterns as implicit functions in a 2D space. Using ISP as a garment representation, GarmentRecovery [29] trains a feedforward network to infer sewing patterns from in-the-wild photos. Neural-Tailor [23] uses an RNN to produce more complex sewing patterns from point clouds. Wang et al. [53] learn a shared garment shape space that can be inferred from multimodal input such as sketches, SewFormer [33] uses a transformer to infer multi-panel patterns from monocular images, and DressCode [16] estimates quantized sewing patterns with a GPT-based model from text inputs. Despite promising results, these approaches are limited to basic sewing patterns, featuring only front and back panels, or constrained by the limited sewing patterns seen in the training data. Additionally, none of the above works have explored using programming parametric sewing patterns, such as GarmentCode [24, 25] to create more complex garments, either in generation or reconstruction.", + "bbox": [ + 91, + 258, + 486, + 592 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "VLMs for 3D tasks. With their broad understanding of the world, LLMs and VLMs are widely used to reason about 3D. There has been significant work on generating, editing, and analyzing 3D objects and scenes [13, 17, 19, 31, 44, 52, 58, 61]. While these methods show the power of VLMs to reason about objects, they do not address our problem of representing 2D patterns that correspond to 3D shapes – this is a special property of garments. Some attempts have been made to connect sewing patterns with language, such as DressCode [16]. This motivates us to present ChatGarment, which connects a large language model to GarmentCode, sharing a similar approach to recent strategies [11, 32] by finetuning LLMs [18, 34, 43]. Powered by VLMs, ChatGarment significantly improves generalization and robustness in reconstructing out-of-domain images and increases diversity in 3D clothing generation. It also supports sewing pattern editing and multi-turn dialogue, greatly simplifying the 3D garment workflow.", + "bbox": [ + 89, + 628, + 486, + 901 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/6fac4910610603e5339c6775b96168d2d4399033ac95c6da4987b2d00d66b316.jpg", + "image_caption": [ + "Figure 2. Pipeline of ChatGarment. ChatGarment takes text or an image as input and generates a JSON file. The JSON file is decoded into 2D sewing patterns using GarmentCode [24] and then draped onto the human body. The final 3D garments are compatible with simulation software (e.g., MAYA, Blender, Style3D, etc.)." + ], + "image_footnote": [], + "bbox": [ + 517, + 88, + 905, + 186 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 513, + 275, + 604, + 290 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Pipeline", + "text_level": 1, + "bbox": [ + 513, + 299, + 612, + 316 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The goal of ChatGarment is to simplify the workflow of garment creation by automatically estimating, generating, and editing garments from multimodal image and text input. To achieve this, we finetune LLaVA [32], a multimodal large language model $f_{\\phi}$ . It takes an input image $X_{v}$ and text prompts $X_{q}$ , and produces a textual response $Y_{t}$ as $Y_{t} = f_{\\phi}(X_{q},X_{v})$ or $Y_{t} = f_{\\phi}(X_{q})$ in the absence of an image. From this output, a structured JSON file $J_{t}$ is extracted as a subset of $Y_{t}$ , which will be decoded into sewing patterns using the GarmentCode [24] decoder $D_{g}$ . This pipeline is illustrated in Fig. 2.", + "bbox": [ + 511, + 321, + 908, + 489 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "LLM-friendly Rich GarmentCode. We use GarmentCode to convert the JSON file into a 3D garment mesh in the canonical pose. GarmentCode [24] is a programming parametric model that uses a domain-specific language (DSL) for creating 2D sewing patterns, with a hierarchical JSON structure to define garment components and stitching rules. These patterns can be draped onto body models via a warp-based pattern stitching algorithm [24, 39], and further be simulated easily. GarmentCode can model diverse garments with intricate geometric details, including different cuts, frills, and pleats. To better align GarmentCode with our model's language-driven requirements, we develop a refined version, GarmentCodeRC (Richer & Cleaner) with two major improvements (see Fig. 3).", + "bbox": [ + 511, + 492, + 908, + 703 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "First, we modify stitching and panel positioning to better support diverse garment types, such as open-front jackets, high-waist skirts, and fitted pant legs. This allows us to model a broader range of real-world garments and aligns more closely with natural garment descriptions. Second, we simplify GarmentCode's JSON configuration to better suit LLM training. The original configuration is redundant, applying the same settings across all garment types. We optimize this by automatically removing irrelevant settings (e.g., omitting skirt-related parameters for upper-body garments). This adjustment reduces the average language token count from 900 to 350, decreasing ambiguity in LLM training and improving efficiency. Furthermore, to stabilize model train", + "bbox": [ + 511, + 704, + 910, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "2926", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/2120e0a1dba49351e967bf370b866a4c51c61e51ba6e2dd0686596fdd41cdc6a.jpg", + "image_caption": [ + "Figure 3. GarmentCodeRC. Left: new options to model open-front jackets, high-waist skirts, and tight pant legs. Right: simplified JSON configuration for more efficient LLM training." + ], + "image_footnote": [], + "bbox": [ + 93, + 88, + 482, + 263 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ing, we normalize all floating-point values to a $[0,1]$ range. Examples of the GarmentCodeRC JSON configuration are given in Sup. Mat. Throughout this paper, \"GarmentCode\" refers to our enriched and simplified version.", + "bbox": [ + 89, + 329, + 483, + 390 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "LLM for Textual and Numeric Outputs. The output JSON file $J_{t}$ contains both textual and numeric descriptions. Previous research [11, 14, 26] has shown that LLMs struggle in predicting continuous and precise numerical values. To overcome this, following ChatPose [11, 26], we encode these continuous values into language token embeddings and train a projection layer to recover accurate numerical values, such as garment length. In this setup, the language embedding for $Y_{t}$ is represented as $H_{t}$ . From the textual output $Y_{t}$ , upon encountering the start token , we extract the textual content of the JSON file, which spans from to . The language embedding at the end token , where numeric information is encoded, is represented as $H_{t}^{\\mathrm{n}}$ . A projection layer uses this embedding to predict the numerical values as $\\mathbf{n} = g_{\\Theta}(H_{t}^{\\mathrm{n}})$ . Finally, we merge these numerical values with the textual descriptions extracted earlier and format them into our final sewing pattern JSON configuration file $J_{t}$ .", + "bbox": [ + 89, + 393, + 483, + 667 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Automatic Data Construction", + "text_level": 1, + "bbox": [ + 89, + 679, + 354, + 694 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To train ChatGarment, we need image-to-JSON and text-to-JSON paired data. While some work like SewFormer [33] offers image-to-JSON pairs, they fall short in garment variety. For example, they lack tight-sleeve garments and only support a limited variety of skirts. Additionally, text descriptions for garment editing are missing in current datasets. We address these gaps by developing an automated data construction pipeline that integrates various resources.", + "bbox": [ + 89, + 702, + 483, + 823 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Garment Generation, Simulation, and Rendering. Following GarmentCodeData [25], we sample 20,000 garments by randomizing values in sewing pattern configurations. To improve real-world garment relevance, we adjust the sampling ratios and assign lower weights for asymmetric gar", + "bbox": [ + 89, + 824, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/e021b681664b01253c5a846583e40a5ee372ebc51e9f5e8b6b7f910e3f2f5468.jpg", + "image_caption": [ + "Figure 4. Data Construction Pipeline. We generate garments from JSON configurations, simulate them with ContourCraft [15] and render with Blender." + ], + "image_footnote": [], + "bbox": [ + 517, + 90, + 903, + 227 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ments and those with intricate cuffs. To create outfits, we pair top and bottom garments together or use single-piece items like dresses or jumpsuits. The garment set is draped on a SMPL-X [40] neutral model.", + "bbox": [ + 511, + 300, + 903, + 359 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We automate garment simulation using ContourCraft [15], chosen for its stability and scalability, with SMPL-X motion sequences sampled from the BEDLAM dataset [4]. Specific garment vertices are attached to corresponding body vertices using predefined rules to prevent garments from slipping off. For example, the upper edges of pants and skirts are connected to the nearest body vertices, and similar attachments are made for upper-body garment vertices near the shoulders. We filter out simulations with significant interpenetration or improper positioning in two steps. First, GarmentCode performs a sewing-pattern validation check, ensuring the integrity of panel edges and stitching. Second, we verify simulation results by assessing inter-penetration, and identify garments at risk of slipping off by inspecting the proximity of each garment vertex to the nearest body part. Each garment is then rendered in Blender from four random perspectives at elevation angles between 0 and 30 degrees, with garment and body textures sourced from the BEDLAM dataset. These steps are shown in Fig. 4.", + "bbox": [ + 511, + 363, + 908, + 650 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "GPT-4o Labeling. The previous method [16] prompts GPT-4o to generate the general geometric descriptions for garments, such as the garment types, lengths, widths and sleeve lengths. However, these general descriptions are insufficient for garment detail understanding: the finetuned VLM struggles to interpret detailed features of garment components, such as pant cuffs, or execute specific garment edits. To address this, we construct a new dataset with low-level descriptions for individual garment parts. For image-based reconstruction, we instruct GPT-4o to generate descriptions for all visible garment parts in the image. For garment editing, we sample 5,000 garment parts (shirt main body, collar, sleeves, sleeve cuffs, waistband, pant legs, pant cuffs, and skirts), and use GPT-4o to generate descriptive text for each part. Assembling these components yields 20,000 new garments with detailed annotations.", + "bbox": [ + 511, + 657, + 908, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "2927", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For garment editing, we randomly select two garments that differ in certain parts, with the editing prompt being: \"Change the garment sewing pattern by modifying PART-1 to DESC-1, PART-2 to DESC-2, while keeping other parts unchanged.\" PART-1, PART-2, represent part names that differ between the two garments, and DESC-1, DESC-2 are target garment part descriptions.", + "bbox": [ + 89, + 90, + 486, + 227 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We finally build a synthetic dataset of 40K garments and 1 million images with text annotations, supporting garment reconstruction and editing from multimodal inputs. Plus, we use GPT-4o to label part-level details of 38K SHHQ [12] images to make the VLM better understand in-the-wild garment images. Compared to previous garment sewing pattern datasets [33], our dataset offers: 1) Diversity: GarmentCode generates a wider range of real-world garment types; 2) Detailed Descriptions: Garments are labeled with detailed geometric features; 3) Precise Control: Allows for precise, low-level garment manipulation.", + "bbox": [ + 89, + 229, + 486, + 397 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Training", + "text_level": 1, + "bbox": [ + 89, + 411, + 194, + 428 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We use LLaVA [32] as our base VLM and finetune it with LoRA [18], with its parameters denoted as $\\phi_{lora}$ . Additionally, we optimize the LLM token embedding $e_{token}$ , the LLM head layer $\\mathrm{lm\\_head}_{\\Theta'}$ , and the projection layer $g_{\\Theta}$ .", + "bbox": [ + 89, + 436, + 485, + 500 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Loss Function. As described in Sec. 3.1, the numeric embedding is extracted from the token [ENDS] as $H_{t}^{\\mathbf{n}}$ . An MLP projection layer maps the embedding $H_{t}^{\\mathbf{n}}$ to 76 float values $\\mathbf{n}$ in the JSON configuration. Given the ground truth text output $\\hat{Y}_{t}$ and sewing pattern values $\\hat{\\mathbf{n}}$ , we optimize with:", + "bbox": [ + 89, + 506, + 483, + 595 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\operatorname {C E} \\left(\\hat {Y} _ {t}, Y _ {t}\\right) + \\lambda_ {\\mathbf {n}} | (\\hat {\\mathbf {n}} - \\mathbf {n}) \\cdot \\mathbf {m} |, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 161, + 601, + 483, + 619 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where the first term is a cross-entropy loss, and the second term is the L1 difference for numeric values. $\\mathbf{m}$ is a mask that filters out irrelevant numeric values, and $\\lambda_{\\mathbf{n}} = 0.1$ is the weight for the L1 loss. See more training details in Sup. Mat.", + "bbox": [ + 89, + 632, + 486, + 694 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Multi-turn Dialogue with Multimodal Inputs.", + "text_level": 1, + "bbox": [ + 89, + 709, + 475, + 726 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After training, our model can handle multimodal input reconstruction, garment generation, and editing, as reflected in the training data. Additionally, it exhibits zero-shot reasoning for garment sewing patterns within multi-turn dialogues, despite being trained only on single-turn dialogue datasets. This enables users to create a garment from an image or text description and then refine it within the same interaction. A typical workflow might start with the VLM estimating a sewing pattern from an image. After reviewing the result, the user can request specific adjustments, providing a flexible, user-centered design process (see Fig. 5 for an example).", + "bbox": [ + 89, + 734, + 486, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 513, + 90, + 645, + 107 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We conduct experiments on image-based reconstruction (Sec 4.1), garment editing (Sec 4.2) and text-based garment generation (Sec 4.3), along with ablation studies (Sec 4.4).", + "bbox": [ + 511, + 114, + 906, + 161 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Image-based Reconstruction", + "text_level": 1, + "bbox": [ + 511, + 170, + 769, + 186 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "**Benchmarks.** We conduct image-based reconstruction experiments on the CloSe [2] and Dress4D [54] datasets. CloSe is a large-scale scan collection including 3D clothing segmentations across 18 clothing categories. Following CloSe, we use 145 scans with accurate SMPL-X fits as our validation set. As CloSe offers limited instances of loose garments, such as skirts and dresses, we further add samples from Dress4D, which has real-world outfits recorded using a multi-view volumetric capture system. We include 4 loose-fitting outfits and 36 images rendered from them.", + "bbox": [ + 511, + 194, + 908, + 345 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details. Our approach applies the Chain-of-Thought (CoT) [55] method. We first prompt ChatGarment to generate a text description of the outfit in the image. The generated text descriptions are combined with the input image to estimate the final garment JSON configuration. More details are provided in Sup. Mat.", + "bbox": [ + 511, + 349, + 910, + 441 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Baselines. We compare our method with two previous approaches, SewFormer [33] and GarmentRecovery [29]. We use the publicly available code and checkpoints for these methods in our evaluation. Additionally, we introduce several new baseline models for comparison:", + "bbox": [ + 511, + 444, + 908, + 520 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- DressCode. We employ DressCode [16] to generate garments from image descriptions. Specifically, we utilize GPT-4o to generate garment descriptions from input images. These descriptions are processed by DressCode to create sewing patterns.", + "- GPT-4o. We design prompt templates and query GPT-4o using in-context learning (ICL) [10]. We use Garment Generator [22] as the sewing pattern configuration tool, as it has significantly fewer possible configurations and allows for more efficient template design and significantly fewer conversation rounds. We first prompt GPT-4o to choose the most suitable template from a predefined list, then query the parameter value by displaying three garment renderings with uniformly sampled parameter values.", + "- LLaVA*. In this setup, we use the LLaVA model to directly output the numerical values for the sewing pattern instead of using the sewing pattern projection layer. The remaining components are identical to our method.", + "- LLaVA-T. We first employ GPT-4o to generate detailed textual descriptions of the garment, and ChatGarment is prompted to reconstruct the sewing pattern based solely on the text description.", + "- LLaVA-I. We input the image into the model and prompt it to directly output the sewing pattern without referring to text descriptions." + ], + "bbox": [ + 513, + 523, + 908, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "2928", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/3c49984c3ca2d60b788f7fb25bc7e3b38052313ac3b89bf9d4fb0cfb9b53f922.jpg", + "image_caption": [ + "(a) Image-based Sewing Pattern Estimation" + ], + "image_footnote": [], + "bbox": [ + 135, + 88, + 547, + 208 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/20c9eb0f2636559fd7855dffabee2eb9e7a8a17de88d2059668276385a21414e.jpg", + "image_caption": [ + "(c) Text-based Sewing Pattern Generation and Editing", + "Figure 5. ChatGarment's dialog modes. Images and texts are adaptively combined to guide garment generation and editing. Text output between STARTS and ENDS contains information about the JSON configuration, which is then converted to 3D garments (JSON2Garment) for visualization and simulation purposes." + ], + "image_footnote": [], + "bbox": [ + 553, + 88, + 852, + 352 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/813cca7457ecf6aaa37f69bd0b31616fa7f4a747ff650888e2fa027b876fce51.jpg", + "image_caption": [ + "(b) Image-guided Sewing Pattern Generation" + ], + "image_footnote": [], + "bbox": [ + 135, + 232, + 547, + 351 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/5a34cda29fb563e6eea3d040dd77fabed5c8cce55029ad7606da47cf951a0321.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsTarget-PoseA-PoseFailure Rate (↓)
CD (↓)F-Score (↑)CD (↓)F-Score (↑)
Sewformer [33]27.060.5520.660.584.3%
DressCode [16]20.160.6012.120.6111.4%
GarmentRecovery [29]9.690.555.750.64-
GPT-4o11.040.6210.440.650
LLaVA*5.290.726.290.760
LLaVA-T6.170.727.250.740
LLaVA-I4.320.754.930.780
Ours3.120.753.060.780
", + "bbox": [ + 94, + 425, + 483, + 553 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/82f803c1bdb693d4b6139b9ccfb1c2e58716edfe0da2f878a02238369ed0facd.jpg", + "table_caption": [ + "Table 1. Quantitative comparisons of Image-based garment reconstruction on Dress4D. ChatGarment achieves the best performance on all metrics." + ], + "table_footnote": [], + "table_body": "
MethodsCLOSE
CD (↓)F-Score (↑)Failure Rate (↓)
SewFormer [33]9.700.7088.33 %
DressCode [16]15.770.6167.0%
GarmentRecovery [29]2.390.785-
GPT-4o3.520.7550
LLaVA*3.080.7750
LLaVA-T3.760.7790
LLaVA-I4.550.7610
Ours2.940.7900
", + "bbox": [ + 122, + 598, + 452, + 742 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Quantitative comparisons of Image-based reconstruction on CloSe. The garments are deformed to the target pose.", + "bbox": [ + 89, + 746, + 485, + 773 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Since methods use different default human poses, we apply Linear Blend Skinning [1] to re-pose all garments to the ground-truth body shape and pose for fair comparison.", + "bbox": [ + 89, + 789, + 483, + 835 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Results. We employ the mean chamfer distance, the mean F-Score at $\\tau = 0.001$ and the stitching failure rate as quantitative metrics. As Tab. 1 and Tab. 2 show, ChatGarment significantly outperforms existing methods on loose garments", + "bbox": [ + 89, + 839, + 485, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "in Dress4D. On the CloSE dataset, which mostly features tight garments, GarmentRecovery [29] does well by leveraging garment segmentation and normal maps as guidance in an optimization-based framework. However, these cues are insufficient for accurately reconstructing loose garments. Our method has similar performance in this case, but is more general. Also, our algorithm robustly reconstructs garments with no stitching failures. In contrast, DressCode and SewFormer often generate invalid garment panels leading to failed stitching. The primary difference lies in the output format: DressCode and SewFormer predict the panel edges and stitches of the sewing pattern, while our method and the GPT-4o-based approach estimate the JSON configuration. Although direct edge and stitch prediction offers greater flexibility in theory, our experiments show that it does not enhance reconstruction accuracy. Instead, it increases the likelihood of generating invalid sewing patterns when handling out-of-distribution (OOD) data.", + "bbox": [ + 511, + 429, + 906, + 700 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our model demonstrates superior performance compared to LLaVA*, highlighting the effectiveness of the proposed projection layer. Furthermore, our CoT-based two-step reconstruction algorithm outperforms both image-based (LLaVA-I) and text-based (LLaVA-T) reconstruction baselines. The text descriptions detail garment types and styles, offering valuable information for reconstruction, while images reveal intricate details that are hard to describe in words. Thus, combining both text and image inputs is the most effective strategy for image-based garment reconstruction.", + "bbox": [ + 511, + 702, + 908, + 853 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Figure 6 shows our model accurately estimates garment type, style, and details like skirt hems and collar styles. In contrast, other methods often misinterpret or overlook these", + "bbox": [ + 511, + 854, + 906, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "2929", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/bf6460101eca7784f1e6d4f87d92d62a334ea8f67984485cc13b574466e9eca0.jpg", + "image_caption": [ + "Figure 6. Image-based Garment Reconstruction. Unlike SewFormer [33], DressCode [16], GarmentRecovery [29], and the GPT-4o-based method, which often make mistakes or ignores garments. ChatGarment faithfully captures the shape, style and composition of the garments." + ], + "image_footnote": [], + "bbox": [ + 91, + 85, + 903, + 537 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "details. SewFormer occasionally fails to reconstruct garments due to detection or stitching failure, and GarmentRecovery lacks support for whole-body dress reconstruction. Additionally, inaccuracies in 2D garment segmentation [28] can lead to errors in GarmentRecovery reconstruction.", + "bbox": [ + 89, + 587, + 485, + 662 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Garment designers often create image sketches before making sewing patterns. Surprisingly, despite not being trained on garment sketch images, ChatGarment demonstrates strong generalization, accurately predicting garment types and styles from sketches (Fig. 8).", + "bbox": [ + 89, + 664, + 485, + 739 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Garment Editing", + "text_level": 1, + "bbox": [ + 89, + 753, + 261, + 771 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "**Benchmarks.** To evaluate garment editing performance, we construct an extra evaluation dataset using methods outlined in Sec. 3.2. We manually select garment pairs with clear textual descriptions and substantial visual differences. This dataset consists of 135 garment pairs, each accompanied by corresponding images and text descriptions. The evaluation of garment editing involves three components: the source garment, the target garment, and the editing prompts. The", + "bbox": [ + 89, + 780, + 483, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "goal is to derive the sewing pattern for the target garment using the source garment and the editing prompts.", + "bbox": [ + 511, + 587, + 906, + 618 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Implementation Details. For garment editing, we concatenate the source garment JSON file and the editing prompt together and pass it into ChatGarment.", + "bbox": [ + 511, + 622, + 908, + 667 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Baselines. We modify DressCode [16] and SewFormer [33] for comparison with ChatGarment:", + "bbox": [ + 511, + 672, + 906, + 702 + ], + "page_idx": 6 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- DressCode with edited text. We generate a text description of the edited garment and use DressCode to create the corresponding sewing patterns. The description is produced by GPT-4o, which takes as inputs: 1) garment editing instructions and 2) front and back views of the source garment. The generated description follows DressCode's original input format.", + "- SewFormer with edited images. We obtain the front view of the edited garment with TurboEdit [9] to which we pass 1) the front view of the source garment, 2) the text description of the source garment, and 3) the editing instructions, and then use SewFormer to turn the image of the edited garment into the corresponding sewing pattern." + ], + "bbox": [ + 511, + 704, + 908, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "2930", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9ff1a21d768bdfbea49cc33ffc7fdd7b8cadd8b9e2fd791d5626cb2c4771067e.jpg", + "image_caption": [ + "Prompts: Change the shirt body to: [long vertical sections, 'evenly shaped sides', 'mildly loose fit around upper body, 'straight front opening']" + ], + "image_footnote": [], + "bbox": [ + 98, + 88, + 151, + 142 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/19c5bf0ef2f9f16d057286836b1fe4d4c45216d93d4d45573ddab11973464445.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 292, + 88, + 346, + 142 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c9084a328857aca1e3dcb54751bc2af2211e17043d397584fe818f0a22acb1c2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 354, + 89, + 408, + 142 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c565fac7f7a928ebea2c9840217b0f8bcc7804fc257560615d6b0d09d8dc2ae7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 418, + 90, + 470, + 142 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/37afdfb42b8ab401af45ab2fa4688c7756bec37b332d490e7bf5ea88d47c195b.jpg", + "image_caption": [ + "Prompts: Change the sleeves to [short sleeves length], 'loose fit around the arms', 'wide sleeve openings relative to the arm size']" + ], + "image_footnote": [], + "bbox": [ + 98, + 154, + 151, + 208 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/180bd5d407ab0829487cc6b440899d3c2a2a4a8d2d07b9f1dd9e3cd77c93a4ca.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 292, + 154, + 346, + 208 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d8521218d572a4209f659e187388cdb9f7eae3f8a9ccc89cef820ed2b91c2ea8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 354, + 154, + 408, + 208 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/533a58d36cf25e8ea9f26083089489d56b94c13b65609c010a598106f74149cf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 418, + 154, + 470, + 208 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4c2ed7e6fef802bedfa647d2577483b9f594f881657e242f633f0581f272feb5.jpg", + "image_caption": [ + "Prompts: Change the collar to: 'rounded neckline', 'slightly loose fit', 'even curvature throughout']", + "Source" + ], + "image_footnote": [], + "bbox": [ + 98, + 220, + 151, + 273 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0d2f8dd010f2162a2ffc2f052f46ad145b3266f43a0897489b3b78d4827e2936.jpg", + "image_caption": [ + "SewFormer" + ], + "image_footnote": [], + "bbox": [ + 292, + 222, + 346, + 273 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d7e07ac8de50fa2558d416c6654e2e43ad0b05adb548ff3088ad974182bbcbc5.jpg", + "image_caption": [ + "DressCode" + ], + "image_footnote": [], + "bbox": [ + 354, + 222, + 408, + 273 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b74c049ce986a552e2356e8fa67f3c3f38af1ccc7f17641a1c9bbb9e139c3fcd.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 418, + 222, + 470, + 273 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a1d4c71b3bac124afae014ab4b230de9e45737b948fd2a68f9b0ce130b2c703b.jpg", + "image_caption": [ + "Prompts: Change the sleeves to: 'long, full-length sleeves', 'loose fit', 'ends at wrist level', 'wide openings at the wrist']" + ], + "image_footnote": [], + "bbox": [ + 517, + 89, + 571, + 142 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3e06c9b751311b6cfdc479aa01ff0e4c1619d868d25622d7f1c1f24bb76c54fd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 725, + 89, + 777, + 142 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/abea4b22b6cd2727c52dbc058a507361893f96d8c512a90b6eea5f74c270671e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 790, + 90, + 843, + 142 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6d75eb984adc16070c6856410fb3db55e1610999957b8982fbc87709ca0ea804.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 849, + 89, + 901, + 142 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/33fe012a9b1b041659498d6d9aa8210fc18240de2ebeff8bfd25eef6a56d961f.jpg", + "image_caption": [ + "Prompts: Change the pant legs to: [wide-leg silhouette], 'ankle-length hem', 'flowy fit', 'loose and relaxed cut']" + ], + "image_footnote": [], + "bbox": [ + 517, + 154, + 571, + 207 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0f5432f029a6a1ab872524fd68192027da4db7b0dae5db4931aa79ac0d7dc467.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 725, + 154, + 777, + 208 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f06ef61ee1af0a17e79c22b8a48ce9496d8f597492c9fa7517e773f451809aa3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 790, + 154, + 843, + 208 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d9151dea9357da3aeb5c95355a408ea45ae190a09206af7f9aa60d3f59a593f9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 849, + 154, + 901, + 208 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0e616a57aaa7916db9ec0c2e07bdd4e2f2f99c31a0bbdea544148f6c75c239cc.jpg", + "image_caption": [ + "Prompts: Change the shirt body to: ['straight vertical silhouette', 'moderate fit around the torso', 'hemline falls above the knees']", + "Source" + ], + "image_footnote": [], + "bbox": [ + 517, + 220, + 571, + 273 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/439c4f6a13ddb4a2efb711b79c3a4d59966c3bb8879c8b5b3a28c58ab6270b49.jpg", + "image_caption": [ + "SewFormer" + ], + "image_footnote": [], + "bbox": [ + 725, + 222, + 777, + 273 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/882692b5467db619b543d82c1b7c0a3c0d4f6f2fe71441478178609fe77792c4.jpg", + "image_caption": [ + "DressCode" + ], + "image_footnote": [], + "bbox": [ + 790, + 223, + 843, + 273 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/df09e02a735e5c81883df1e6f035b71d84af936fc82b003c61576e6009e9d92a.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 849, + 222, + 901, + 273 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7b3dd8724c3b750637254fc4a1515a12c4f805f3b7365816b61ba294d5846d4e.jpg", + "image_caption": [ + "Figure 7. Results on Garment Editing. We present six examples where the models (SewFormer [33], DressCode [16], and ChatGarment) edit the source garment shown in the source image according to the given editing instructions in the prompt boxes.", + "Input" + ], + "image_footnote": [], + "bbox": [ + 117, + 345, + 202, + 513 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/62a5b265abfc2ad7acde6368371ddcdeb3708f65303741b7d64f963f940b1feb.jpg", + "image_caption": [ + "2D Sewing Pattern" + ], + "image_footnote": [], + "bbox": [ + 210, + 345, + 370, + 513 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b1872f8f762845b0bb375ccc5c754dee17e38382397433a491a15656ac4fa1fa.jpg", + "image_caption": [ + "3D Garment", + "Figure 8. Reconstruction results from garment sketches. The generated garments accurately capture the shape and style of the input sketch, despite our model being trained without sketches." + ], + "image_footnote": [], + "bbox": [ + 372, + 347, + 452, + 516 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Results. We compare our method with SewFormer and Dresscode on Chamfer Distance and F-score in Tab. 3, where our approach demonstrates superior performance. The qualitative results in Fig. 7 show that our method accurately modifies the targeted garment parts according to the input prompt. In contrast, SewFormer and DressCode often fail to precisely follow the prompt instructions and frequently alter other unintended areas of the garment.", + "bbox": [ + 89, + 604, + 483, + 726 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Text-based Generation", + "text_level": 1, + "bbox": [ + 89, + 739, + 302, + 753 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Following DressCode [16], we use GPT-4o to generate 150 highly diverse text labels from 85 selected in-the-wild images featuring various tops, skirts, pants, and dresses. Next, we render the generated 3D garments with default textures and compute the CLIP score using the provided text labels. See quantitative results in Tab. 4. Our model achieves a higher CLIP score compared to DressCode [16], highlighting the superior prompt-following capabilities of our approach. The qualitative results are shown in Sup. Mat.", + "bbox": [ + 89, + 763, + 485, + 902 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/df74d6c2c39c30b0baa674c4bbeb6290ea5e06a40e4caea7da033904a54a50ea.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsCD (↓)F-Score (↑)
SewFormer [33]28.770.548
DressCode [16]10.970.750
Ours2.510.893
", + "bbox": [ + 584, + 340, + 831, + 392 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/fb7247fb472c239030c251f3787dc30725196d914eee8f5164a92bf59483aebf.jpg", + "table_caption": [ + "Table 3. Quantitative comparisons for garment editing. ChatGarment outperforms other two methods by a large margin." + ], + "table_footnote": [], + "table_body": "
MethodsCLIP score (↑)
DressCode [16]22.8
Ours23.7
", + "bbox": [ + 607, + 434, + 810, + 473 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. Quantitative comparisons for text-based garment generation. ChatGarment achieves a higher CLIP score than DressCode, indicating better alignment with the text inputs.", + "bbox": [ + 511, + 478, + 906, + 520 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Ablation", + "text_level": 1, + "bbox": [ + 513, + 534, + 617, + 549 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We evaluate the impact of different aspects of ChatGarment, including multi-modal LLM backbones, and various training datasets. Please refer to Sup. Mat.", + "bbox": [ + 511, + 558, + 906, + 604 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 617, + 633, + 632 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We introduce ChatGarment, a VLM-based model that estimates garments from images and generates or edits garments based on text descriptions. Leveraging the capabilities of LLMs, ChatGarment supports multi-turn dialogue interactions for 3D garment design. It surpasses existing sewing pattern-specific and LLM-based methods in accurately estimating sewing patterns from a single image. While ChatGarment excels in many areas, it may struggle with precise garment editing. For instance, altering the skirt length might unintentionally affect other parts of the garment. This issue could be addressed with more detailed text instructions or a differential CSG optimizer [60] for finer adjustments. Nonetheless, ChatGarment offers the first VLM-based framework for unified garment estimation, generation, and editing. Future directions could include expanding into garment manufacturing design or improving clothing simulation with more fine-grained material attributes.", + "bbox": [ + 511, + 642, + 908, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "2931", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments and Disclosure", + "text_level": 1, + "bbox": [ + 91, + 90, + 377, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We thank Maria Korosteleva for help on GarmentCode, and Giorgio Becherini for advice on data acquisition and processing. Yuliang Xiu is funded by Max Planck Institute for Intelligent Systems, the Research Center for Industries of the Future (RCIF) at Westlake University, and the Westlake Education Foundation.", + "bbox": [ + 89, + 114, + 485, + 204 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "While MJB is a co-founder and Chief Scientist at Meshcapade, his research in this project was performed solely at, and funded solely by, the Max Planck Society.", + "bbox": [ + 89, + 205, + 483, + 250 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 277, + 187, + 292 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Rinet Abdrashitov, Kim Raichstat, Jared Monsen, and David Hill. Robust skin weights transfer via weight inpainting. In SIGGRAPH Asia 2023 Technical Communications, pages 1-4. 2023. 6", + "[2] Dimitrije Antic, Garvita Tiwari, Batuhan Ozcomlekci, Ricardo Marin, and Gerard Pons-Moll. CloSe: A 3D clothing segmentation dataset and model. In International Conference on 3D Vision (3DV), 2024. 2, 5", + "[3] Bharat Lal Bhatnagar, Garvita Tiwari, Christian Theobalt, and Gerard Pons-Moll. Multi-garment net: Learning to dress 3d people from images. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5420-5430, 2019. 2", + "[4] Michael J Black, Priyanka Patel, Joachim Tesch, and Jinlong Yang. Bedlam: A synthetic dataset of bodies exhibiting detailed lifelike animated motion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8726-8737, 2023. 4", + "[5] Andrés Casado-Elvira, Marc Comino Trinidad, and Dan Casas. Pergamo: Personalized 3d garments from monocular video. In Computer Graphics Forum, pages 293–304. Wiley Online Library, 2022. 2", + "[6] Enric Corona, Albert Pumarola, Guillem Alenya, Gerard Pons-Moll, and Francesc Moreno-Noguer. Simplicit: Topology-aware generative model for clothed people. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11875-11885, 2021. 2, 3", + "[7] R Daněřek, Endri Dibra, Cengiz Öztireli, Remo Ziegler, and Markus Gross. Deepgarment: 3d garment shape estimation from a single image. In Computer Graphics Forum, pages 269-280. Wiley Online Library, 2017. 2", + "[8] Luca De Luigi, Ren Li, Benoit Guillard, Mathieu Salzmann, and Pascal Fua. Drapenet: Garment generation and self-supervised draping. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1451-1460, 2023. 2, 3", + "[9] Gilad Deutch, Rinon Gal, Daniel Garibi, Or Patashnik, and Daniel Cohen-Or. Turboedit: Text-based image editing using few-step diffusion models. arXiv preprint arXiv:2408.00735, 2024. 7", + "[10] Qingxiu Dong, Lei Li, Damai Dai, Ce Zheng, Jingyuan Ma, Rui Li, Heming Xia, Jingjing Xu, Zhiyong Wu, Tianyu" + ], + "bbox": [ + 93, + 301, + 485, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Liu, et al. A survey on in-context learning. arXiv preprint arXiv:2301.00234, 2022. 5", + "[11] Yao Feng, Jing Lin, Sai Kumar Dwivedi, Yu Sun, Priyanka Patel, and Michael J. Black. ChatPose: Chatting about 3d human pose. In CVPR, 2024. 2, 3, 4", + "[12] Jianglin Fu, Shikai Li, Yuming Jiang, Kwan-Yee Lin, Chen Qian, Chen-Change Loy, Wayne Wu, and Ziwei Liu. Stylegan-human: A data-centric odyssey of human generation. arXiv preprint, arXiv:2204.11823, 2022. 5", + "[13] Rao Fu, Xiao Zhan, Yiwen Chen, Daniel Ritchie, and Srinath Sridhar. Shapecraft: A recursive text-conditioned 3d shape generation model. Advances in Neural Information Processing Systems, 35:8882-8895, 2022. 3", + "[14] Siavash Golkar, Mariel Pettee, Michael Eickenberg, Alberto Bietti, Miles Cranmer, Geraud Krawezik, Francois Lanusse, Michael McCabe, Ruben Ohana, Liam Parker, et al. xVal: A continuous number encoding for large language models. arXiv preprint arXiv:2310.02989, 2023. 2, 4", + "[15] Artur Grigorev, Giorgio Becherini, Michael Black, Otmar Hilliges, and Bernhard Thomaszewski. Contourcraft: Learning to resolve intersections in neural multi-garment simulations. In ACM SIGGRAPH 2024 Conference Papers, pages 1-10, 2024. 4", + "[16] Kai He, Kaixin Yao, Qixuan Zhang, Jingyi Yu, Lingjie Liu, and Lan Xu. Dresscode: Autoregressively sewing and generating garments from text guidance. ACM Transactions on Graphics (TOG), 43(4):1-13, 2024. 2, 3, 4, 5, 6, 7, 8", + "[17] Yining Hong, Haoyu Zhen, Peihao Chen, Shuhong Zheng, Yilun Du, Zhenfang Chen, and Chuang Gan. 3d-llm: Injecting the 3d world into large language models. Advances in Neural Information Processing Systems, 36:20482-20494, 2023. 3", + "[18] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In ICLR, 2022. 2, 3, 5", + "[19] Ziniu Hu, Ahmet Iscen, Aashi Jain, Thomas Kipf, Yisong Yue, David A Ross, Cordelia Schmid, and Alireza Fathi. Scenecraft: An IIm agent for synthesizing 3d scenes as blender code. In *Forty-first International Conference on Machine Learning*, 2024. 3", + "[20] Yangyi Huang, Hongwei Yi, Yuliang Xiu, Tingting Liao, Ji-axiang Tang, Deng Cai, and Justus Thies. TeCH: Text-guided Reconstruction of Lifelike Clothed Humans. In International Conference on 3D Vision (3DV), 2024. 3", + "[21] Boyi Jiang, Juyong Zhang, Yang Hong, Jinhao Luo, Ligang Liu, and Hujun Bao. Bcnet: Learning body and cloth shape from a single image. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XX 16, pages 18-35. Springer, 2020. 2", + "[22] Maria Korosteleva and Sung-Hee Lee. Generating datasets of 3d garments with sewing patterns. arXiv preprint arXiv:2109.05633, 2021. 5", + "[23] Maria Korosteleva and Sung-Hee Lee. Neuraltailer: Reconstructing sewing pattern structures from 3d point clouds of garments. ACM Transactions on Graphics (TOG), 41(4):1-16, 2022. 3", + "[24] Maria Korosteleva and Olga Sorkine-Hornung. GarmentCode: Programming parametric sewing patterns. ACM Transaction" + ], + "bbox": [ + 516, + 92, + 906, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "2932", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "on Graphics, 42(6), 2023. SIGGRAPH ASIA 2023 issue. 2, 3", + "[25] Maria Korosteleva, Timur Levent Kesdogan, Fabian Kemper, Stephan Wenninger, Jasmin Koller, Yuhan Zhang, Mario Botsch, and Olga Sorkine-Hornung. GarmentCodeData: A dataset of 3D made-to-measure garments with sewing patterns. In Computer Vision – ECCV 2024, 2024. 2, 3, 4", + "[26] Peter Kulits, Haiwen Feng, Weiyang Liu, Victoria Abrevaya, and Michael J Black. Re-thinking inverse graphics with large language models. Transactions on Machine Learning Research, 2024. 2, 4", + "[27] Boqian Li, Xuan Li, Ying Jiang, Tianyi Xie, Feng Gao, Huamin Wang, Yin Yang, and Chenfanfu Jiang. Garmentdreamer: 3dgs guided garment synthesis with diverse geometry and texture details. In 3DV, 2025. 2", + "[28] Peike Li, Yunqiu Xu, Yunchao Wei, and Yi Yang. Self-correction for human parsing. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(6):3260-3271, 2020. 7", + "[29] Ren Li, Corentin Dumery, Benoit Guillard, and Pascal Fua. Garment recovery with shape and deformation priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1586-1595, 2024. 3, 5, 6, 7", + "[30] Ren Li, Benoit Guillard, and Pascal Fua. Isp: Multi-layered garment draping with implicit sewing patterns. Advances in Neural Information Processing Systems, 36, 2024. 2, 3", + "[31] Dingning Liu, Xiaoshui Huang, Yuenan Hou, Zhihui Wang, Zhenfei Yin, Yongshun Gong, Peng Gao, and Wanli Ouyang. Uni3d-llm: Unifying point cloud perception, generation and editing with large language models. arXiv preprint arXiv:2402.03327, 2024. 3", + "[32] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning, 2023. 3, 5", + "[33] Lijuan Liu, Xiangyu Xu, Zhijie Lin, Jiabin Liang, and Shuicheng Yan. Towards garment sewing pattern reconstruction from a single image. ACM Transactions on Graphics (TOG), 42(6):1-15, 2023. 2, 3, 4, 5, 6, 7, 8", + "[34] Weiyang Liu, Zeju Qiu, Yao Feng, Yuliang Xiu, Yuxuan Xue, Longhui Yu, Haiwen Feng, Zhen Liu, Juyeon Heo, Songyou Peng, Yandong Wen, Michael J. Black, Adrian Weller, and Bernhard Scholkopf. Parameter-efficient orthogonal finetuning via butterfly factorization. In ICLR, 2024. 3", + "[35] Zhen Liu, Yao Feng, Yuliang Xiu, Weiyang Liu, Liam Paull, Michael J. Black, and Bernhard Scholkopf. Ghost on the shell: An expressive representation of general 3d shapes. In The Twelfth International Conference on Learning Representations, 2024. 3", + "[36] Zhongjin Luo, Haolin Liu, Chenghong Li, Wanghao Du, Zirong Jin, Yinyu Nie, Weikai Chen, and Xiaoguang Han. Garverselod: High-fidelity 3d garment reconstruction from a single in-the-wild image using a dataset with levels of details. In ACM Transactions on Graphics (TOG), 2024. 2", + "[37] Qianli Ma, Jinlong Yang, Siyu Tang, and Michael J. Black. The power of points for modeling humans in clothing. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 10974-10984, 2021. 3", + "[38] Qianli Ma, Jinlong Yang, Michael J. Black, and Siyu Tang. Neural point-based shape modeling of humans in challenging" + ], + "bbox": [ + 91, + 90, + 483, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "clothing. In International Conference on 3D Vision (3DV), pages 679-689, 2022. 3", + "[39] Miles Macklin. Warp: A high-performance python framework forgpu simulation and graphics. https://github.com/nvidia/warp, 2022. NVIDIA GPU Technology Conference (GTC). 3", + "[40] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed A. A. Osman, Dimitrios Tzionas, and Michael J. Black. Expressive body capture: 3D hands, face, and body from a single image. In CVPR, 2019. 4", + "[41] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. In ICLR, 2023. 2", + "[42] Lingteng Qiu, Guanying Chen, Jiapeng Zhou, Mutian Xu, Junle Wang, and Xiaoguang Han. Rec-mv: Reconstructing 3d dynamic cloth from monocular videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4637-4646, 2023. 2", + "[43] Zeju Qiu, Weiyang Liu, Haiwen Feng, Yuxuan Xue, Yao Feng, Zhen Liu, Dan Zhang, Adrian Weller, and Bernhard Schölkopf. Controlling text-to-image diffusion by orthogonal finetuning. In NeurIPS, 2023. 3", + "[44] Zeju Qiu, Weiyang Liu, Haiwen Feng, Zhen Liu, Tim Z Xiao, Katherine M Collins, Joshua B Tenenbaum, Adrian Weller, Michael J Black, and Bernhard Scholkopf. Can large language models understand symbolic graphics programs? arXiv preprint arXiv:2408.08313, 2024. 3", + "[45] Shunsuke Saito, Zeng Huang, Ryota Natsume, Shigeo Morishima, Hao Li, and Angjoo Kanazawa. PIFu: Pixel-aligned implicit function for high-resolution clothed human digitization. In ICCV, pages 2304-2314, 2019. 3", + "[46] Shunsuke Saito, Tomas Simon, Jason Saragih, and Hanbyul Joo. PIFuHD: Multi-Level Pixel-Aligned Implicit Function for High-Resolution 3D Human Digitization. In CVPR, pages 81–90, 2020. 3", + "[47] Nikolaos Sarafianos, Tuur Stuyck, Xiaoyu Xiang, Yilei Li, Jovan Popovic, and Rakesh Ranjan. Garment3dgen: 3d garment stylization and texture generation. In 3DV, 2025. 2", + "[48] Tianchang Shen, Jun Gao, Kangxue Yin, Ming-Yu Liu, and Sanja Fidler. Deep marching tetrahedra: a hybrid representation for high-resolution 3d shape synthesis. In Advances in Neural Information Processing Systems (NeurIPS), 2021. 3", + "[49] Yu Shen, Junbang Liang, and Ming C. Lin. Gan-based garment generation using sewing pattern images. In ECCV, 2020. 2", + "[50] Astitva Srivastava, Pranav Manu, Amit Raj, Varun Jampani, and Avinash Sharma. Wordrobe: Text-guided generation of textured 3d garments. In ECCV, pages 458-475. Springer, 2025. 3", + "[51] Zhaoqi Su, Weilin Wan, Tao Yu, Lingjie Liu, Lu Fang, Wenping Wang, and Yebin Liu. Mulaycap: Multi-layer human performance capture using a monocular video camera. IEEE Transactions on Visualization and Computer Graphics, 28(4): 1862-1879, 2020. 3", + "[52] Chunyi Sun, Junlin Han, Weijian Deng, Xinlong Wang, Zishan Qin, and Stephen Gould. 3d-gpt: Procedural 3d modeling with large language models. arXiv preprint arXiv:2310.12945, 2023. 3" + ], + "bbox": [ + 516, + 92, + 908, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "2933", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[53] Tuanfeng Y. Wang, Duygu Ceylan, Jovan Popovic, and Niloy J. Mitra. Learning a shared shape space for multimodal garment design. ACM Trans. Graph., 37(6), 2018. 3", + "[54] Wenbo Wang, Hsuan-I Ho, Chen Guo, Boxiang Rong, Artur Grigorev, Jie Song, Juan Jose Zarate, and Otmar Hilliges. 4D-DRESS: A 4D dataset of real-world human clothing with semantic annotations. In IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2024. 2, 5", + "[55] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824-24837, 2022. 5", + "[56] Yuliang Xiu, Jinlong Yang, Dimitrios Tzionas, and Michael J. Black. ICON: Implicit Clothed humans Obtained from Normals. In CVPR, 2022. 3", + "[57] Yuliang Xiu, Jinlong Yang, Xu Cao, Dimitrios Tzionas, and Michael J. Black. ECON: Explicit Clothed humans Optimized via Normal integration. In CVPR, 2023. 3", + "[58] Runsen Xu, Xiaolong Wang, Tai Wang, Yilun Chen, Jiangmiao Pang, and Dahua Lin. Pointllm: Empowering large language models to understand point clouds. In European Conference on Computer Vision, pages 131-147. Springer, 2025. 3", + "[59] Zhengming Yu, Zhiyang Dou, Xiaoxiao Long, Cheng Lin, Zekun Li, Yuan Liu, Norman Müller, Taku Komura, Marc Habermann, Christian Theobalt, et al. Surf-d: Generating high-quality surfaces of arbitrary topologies using diffusion models. In ECCV, pages 419-438. Springer, 2025. 3", + "[60] Haocheng Yuan, Adrien Bousseau, Hao Pan, Chengquan Zhang, Niloy J Mitra, and Changjian Li. Diffcsg: Differentiable csg via rasterization. arXiv preprint arXiv:2409.01421, 2024.8", + "[61] Longwen Zhang, Ziyu Wang, Qixuan Zhang, Qiwei Qiu, Anqi Pang, Haoran Jiang, Wei Yang, Lan Xu, and Jingyi Yu. Clay: A controllable large-scale generative model for creating high-quality 3d assets. ACM Transactions on Graphics (TOG), 43 (4):1-20, 2024. 3", + "[62] Heming Zhu, Lingteng Qiu, Yuda Qiu, and Xiaoguang Han. Registering explicit to implicit: Towards high-fidelity garment mesh reconstruction from single images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3845-3854, 2022. 2" + ], + "bbox": [ + 91, + 92, + 485, + 712 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "2934", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/ChatGarment_ Garment Estimation, Generation and Editing via Large Language Models/425b09e1-924e-4a43-b277-eb8a41387d21_model.json b/2025/ChatGarment_ Garment Estimation, Generation and Editing via Large Language Models/425b09e1-924e-4a43-b277-eb8a41387d21_model.json new file mode 100644 index 0000000000000000000000000000000000000000..2be1d0297bf3380f9fe5a7e5fc13e91c231f51d0 --- /dev/null +++ b/2025/ChatGarment_ Garment Estimation, Generation and Editing via Large Language Models/425b09e1-924e-4a43-b277-eb8a41387d21_model.json @@ -0,0 +1,2708 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.192, + 0.131, + 0.809, + 0.177 + ], + "angle": 0, + "content": "ChatGarment: Garment Estimation, Generation and Editing via Large Language Models" + }, + { + "type": "text", + "bbox": [ + 0.156, + 0.202, + 0.842, + 0.244 + ], + "angle": 0, + "content": "Siyuan Bian\\(^{1,2}\\) Chenghao Xu\\(^{1,3}\\) Yuliang Xiu\\(^{1,4}\\) Artur Grigorev\\(^{1,5}\\) Zhen Liu\\(^{1,6}\\) Cewu Lu\\(^{2}\\) Michael J. Black\\(^{1}\\) Yao Feng\\(^{7,8,\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.182, + 0.247, + 0.816, + 0.267 + ], + "angle": 0, + "content": "\\(^{1}\\)Max Planck Institute for Intelligent Systems \\(^{2}\\)Shanghai Jiao Tong University" + }, + { + "type": "text", + "bbox": [ + 0.182, + 0.267, + 0.816, + 0.303 + ], + "angle": 0, + "content": "\\(^{3}\\)EPFL \\(^{4}\\)Westlake University \\(^{5}\\)ETH Zürich \\(^{6}\\)Mila, University of Montreal \\(^{7}\\)Meshcapade \\(^{8}\\)Stanford University" + }, + { + "type": "text", + "bbox": [ + 0.402, + 0.306, + 0.591, + 0.319 + ], + "angle": 0, + "content": "chatgarment.github.io" + }, + { + "type": "image", + "bbox": [ + 0.091, + 0.349, + 0.511, + 0.514 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.349, + 0.907, + 0.515 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.525, + 0.907, + 0.551 + ], + "angle": 0, + "content": "Figure 1. As a multimodal 3D garment creator, ChatGarment understands both images and language. It can estimate complex 3D garments represented as sewing patterns from a single image, which can be easily animated and simulated. It also supports garment editing based on text instructions." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.562, + 0.327, + 0.578 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.596, + 0.486, + 0.869 + ], + "angle": 0, + "content": "We introduce ChatGarment, a novel approach that leverages large vision-language models (VLMs) to automate the estimation, generation, and editing of 3D garments from images or text descriptions. Unlike previous methods that struggle in real-world scenarios or lack interactive editing capabilities, ChatGarment can estimate sewing patterns from in-the-wild images or sketches, generate them from text descriptions, and edit garments based on user instructions, all within an interactive dialogue. These sewing patterns can then be draped on a 3D body and animated. This is achieved by finetuning a VLM to directly generate a JSON file that includes both textual descriptions of garment types and styles, as well as continuous numerical attributes. This JSON file is then used to create sewing patterns through a programming parametric model. To support this, we refine the existing programming model, GarmentCode, by expanding its garment type coverage and simplifying its structure for efficient VLM fine-tuning. Additionally, we construct a" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.564, + 0.909, + 0.67 + ], + "angle": 0, + "content": "large-scale dataset of image-to-sewing-pattern and text-to-sewing-pattern pairs through an automated data pipeline. Extensive evaluations demonstrate ChatGarment's ability to accurately reconstruct, generate, and edit garments from multimodal inputs, highlighting its potential to simplify workflows in fashion and gaming applications. Code and data are available at https://chatgarment.github.io/." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.68, + 0.645, + 0.695 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.704, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Imagine seeing a photo of someone in a nice looking outfit that you would like for yourself or your 3D avatar. Our goal is to take that photo and convert it into a 2D sewing pattern and then enable the user to use language to edit the pattern, for example, making the sleeves longer or editing the neckline. We then drape the garment on a 3D body and animate it. This is illustrated in Fig. 1. With this approach, given just an image and/or text, we construct sewing patterns for 3D garments that can be immediately used in fashion or gaming applications. This lightweight process is in stark contrast to how sewing patterns and 3D garments are traditionally created by an artist or clothing designer. In contrast to the traditional labor-intensive process, our approach exploits" + }, + { + "type": "page_footnote", + "bbox": [ + 0.111, + 0.886, + 0.384, + 0.901 + ], + "angle": 0, + "content": "This work was done while YF was at Meshcapade." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2924" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.123 + ], + "angle": 0, + "content": "large vision-language models (VLMs) for the first time to democratize the capture and design of clothing." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.126, + 0.486, + 0.518 + ], + "angle": 0, + "content": "Recent work explores the generation or estimation of 3D clothing using text and/or images as input [6, 8, 21, 42]. These methods use various 3D representations including 3D meshes, point clouds, or implicit representations like Unsigned Distance Functions (UDFs) and Signed Distance Functions (SDFs). We observe, however, that most clothing is designed and manufactured using 2D patterns, and posit that this is a more natural representation. Sewing patterns can be seamlessly integrated into existing garment design pipelines for animation and manufacturing. Recent approaches [16, 30, 33] use sewing patterns for garment estimation and generation, but mapping images or text to sewing patterns remains challenging due to limited training data, leading to poor generalization across diverse clothing types. These methods also does not support interactive editing, making post-processing labor-intensive for 3D artists. Fortunately, recent work, such as GarmentCode [24, 25], provides a programmatic framework for encoding garment information into a JSON file. This JSON file includes textual descriptions of garment types, styles and numerical attributes, offering a rich representation of sewing patterns. The JSON file is then processed by the GarmentCode framework to produce 2D sewing patterns, which can subsequently be draped on a 3D body. However, to date, there is no automated method to directly extract a GarmentCode representation (JSON file) from an image or text description." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.521, + 0.486, + 0.688 + ], + "angle": 0, + "content": "Our observation is that a JSON file resembles natural language, making it suitable for interpretation by large language models (LLMs). At the same time, Vision-LLMs (VLMs) excel at understanding images and garment concepts, such as types of garments and specific features like long sleeves. If we enable the VLM to translate its garment understanding into the structured format of GarmentCode, then it becomes capable of generating sewing patterns. This allows the VLM to take an image as input and output a garment, while also supporting garment generation and editing through text descriptions or interactive dialogue." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.486, + 0.902 + ], + "angle": 0, + "content": "Specifically, we introduce ChatGarment, a method that finetunes a VLM to take text queries, optionally with images, as input and outputs a JSON garment file. This JSON file is then used for GarmentCode generation. The JSON file includes textual descriptions of garment types (e.g., skirts, pants), styles (e.g., collar type), and continuous numerical attributes (e.g., pant length). Inspired by previous work [11, 14, 26], we introduce a new token () to represent these numerical values and train an MLP projection layer to decode them from the token embedding. Those numbers, combined with the VLM's language output, are formatted into the final JSON file. In our approach, we keep the VLM's vision encoder fixed, fine-tune the language model using LoRA [18], and jointly train the MLP projection layer. To" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.908, + 0.257 + ], + "angle": 0, + "content": "adapt GarmentCode [24] for more efficient VLM training, we create an improved version that supports a wider range of clothing types (e.g. open-jacket) and removes unnecessary settings to reduce token usage in training. We also develop an automated data construction pipeline that leverages existing tools for garment generation, draping, simulation, rendering, and automatic labeling with GPT-4o. This pipeline enables building a dataset with image-to-GarmentCode and text-to-GarmentCode pairs, including 20K new garments and 1M images with detailed descriptions, supporting both garment creation and editing." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.258, + 0.909, + 0.424 + ], + "angle": 0, + "content": "We evaluate ChatGarment across a diverse set of tasks, including those specifically trained for, and novel applications that demonstrate the generalization capabilities of VLMs. For single-image garment reconstruction, ChatGarment outperforms prior sewing-pattern-specific and LLM-based models on the Dress4D [54] and CloSE [2] evaluation datasets. Additionally, we demonstrate garment editing, generation, and multi-turn dialogue, highlighting our approach's versatility. ChatGarment can flexibly use text and images to create and edit 3D garments that can be readily animated, introducing a new workflow for garment design." + }, + { + "type": "text", + "bbox": [ + 0.536, + 0.425, + 0.825, + 0.438 + ], + "angle": 0, + "content": "We summarize our contributions as follows:" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.44, + 0.905, + 0.469 + ], + "angle": 0, + "content": "- A unified model capable of estimating, generating and editing sewing patterns from multimodal inputs." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.47, + 0.907, + 0.498 + ], + "angle": 0, + "content": "- The first approach to leverage VLMs for directly generating JSON files for garment creation." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.5, + 0.907, + 0.53 + ], + "angle": 0, + "content": "- A refined version of GarmentCode that supports more diverse garment types and is optimized for VLM training." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.53, + 0.907, + 0.575 + ], + "angle": 0, + "content": "- An automatic data construction pipeline for generating image-to-sewing-pattern and text-to-sewing-pattern data, along with the release of a large-scale dataset." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.44, + 0.907, + 0.575 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.589, + 0.655, + 0.605 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.614, + 0.908, + 0.704 + ], + "angle": 0, + "content": "Garment Creation. Most 3D garments have predefined templates (e.g., sewing patterns) and categories (e.g., skirts, shirts, pants, dresses, etc.), but their non-rigid nature makes them highly deformable with dynamic motions. Efforts to create 3D garments can be grouped into \"template-based\" or \"free-form\" methods, based on their 3D representations." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.909, + 0.902 + ], + "angle": 0, + "content": "A line of template-based approaches register predefined garment templates to 2D observations, which could be real captures, e.g. from monocular video [5, 21, 36, 42] and images [3, 6, 7, 62], generated data [49], or derived from the \"proxy gradient\" computed from 2D diffusion models [27, 47] via score distillation sampling (SDS) [41]. When the template is classified correctly, the predefined garment template, refined through a coarse-to-fine step, ensures topological stability and correctness. Unfortunately, template-based methods struggle to accurately model garments that deviate from the predefined templates, particularly those with diverse styles, such as long skirts, or garments that undergo topological changes, like opening a jacket." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2925" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.244 + ], + "angle": 0, + "content": "To address these limitations, free-form representations like (un)signed distance functions [6, 8, 50, 59], occupancy fields [45, 46, 56, 57], tetrahedrons [20, 35, 48], and point clouds [37, 38] are leveraged. However, this flexibility can compromise the inherent structure of garments, potentially leading to non-garment shapes and floating artifacts. Without shape regularization, incomplete observations lead to the reconstructed unseen parts being overly smooth. Even worse, rigging or simulating these free-form garments becomes extremely challenging." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.26, + 0.487, + 0.593 + ], + "angle": 0, + "content": "The sewing pattern is a well-balanced solution – with pre-defined yet flexible templates. Several methods exploit sewing patterns to reconstruct 3D garments. MulayCap [51] estimates the parameters of a sewing pattern from a monocular video. ISP [30] parameterizes sewing patterns as implicit functions in a 2D space. Using ISP as a garment representation, GarmentRecovery [29] trains a feedforward network to infer sewing patterns from in-the-wild photos. Neural-Tailor [23] uses an RNN to produce more complex sewing patterns from point clouds. Wang et al. [53] learn a shared garment shape space that can be inferred from multimodal input such as sketches, SewFormer [33] uses a transformer to infer multi-panel patterns from monocular images, and DressCode [16] estimates quantized sewing patterns with a GPT-based model from text inputs. Despite promising results, these approaches are limited to basic sewing patterns, featuring only front and back panels, or constrained by the limited sewing patterns seen in the training data. Additionally, none of the above works have explored using programming parametric sewing patterns, such as GarmentCode [24, 25] to create more complex garments, either in generation or reconstruction." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.629, + 0.487, + 0.902 + ], + "angle": 0, + "content": "VLMs for 3D tasks. With their broad understanding of the world, LLMs and VLMs are widely used to reason about 3D. There has been significant work on generating, editing, and analyzing 3D objects and scenes [13, 17, 19, 31, 44, 52, 58, 61]. While these methods show the power of VLMs to reason about objects, they do not address our problem of representing 2D patterns that correspond to 3D shapes – this is a special property of garments. Some attempts have been made to connect sewing patterns with language, such as DressCode [16]. This motivates us to present ChatGarment, which connects a large language model to GarmentCode, sharing a similar approach to recent strategies [11, 32] by finetuning LLMs [18, 34, 43]. Powered by VLMs, ChatGarment significantly improves generalization and robustness in reconstructing out-of-domain images and increases diversity in 3D clothing generation. It also supports sewing pattern editing and multi-turn dialogue, greatly simplifying the 3D garment workflow." + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.089, + 0.906, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.196, + 0.908, + 0.267 + ], + "angle": 0, + "content": "Figure 2. Pipeline of ChatGarment. ChatGarment takes text or an image as input and generates a JSON file. The JSON file is decoded into 2D sewing patterns using GarmentCode [24] and then draped onto the human body. The final 3D garments are compatible with simulation software (e.g., MAYA, Blender, Style3D, etc.)." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.276, + 0.605, + 0.291 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.3, + 0.613, + 0.317 + ], + "angle": 0, + "content": "3.1. Pipeline" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.323, + 0.909, + 0.49 + ], + "angle": 0, + "content": "The goal of ChatGarment is to simplify the workflow of garment creation by automatically estimating, generating, and editing garments from multimodal image and text input. To achieve this, we finetune LLaVA [32], a multimodal large language model \\( f_{\\phi} \\). It takes an input image \\( X_{v} \\) and text prompts \\( X_{q} \\), and produces a textual response \\( Y_{t} \\) as \\( Y_{t} = f_{\\phi}(X_{q},X_{v}) \\) or \\( Y_{t} = f_{\\phi}(X_{q}) \\) in the absence of an image. From this output, a structured JSON file \\( J_{t} \\) is extracted as a subset of \\( Y_{t} \\), which will be decoded into sewing patterns using the GarmentCode [24] decoder \\( D_{g} \\). This pipeline is illustrated in Fig. 2." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.493, + 0.909, + 0.704 + ], + "angle": 0, + "content": "LLM-friendly Rich GarmentCode. We use GarmentCode to convert the JSON file into a 3D garment mesh in the canonical pose. GarmentCode [24] is a programming parametric model that uses a domain-specific language (DSL) for creating 2D sewing patterns, with a hierarchical JSON structure to define garment components and stitching rules. These patterns can be draped onto body models via a warp-based pattern stitching algorithm [24, 39], and further be simulated easily. GarmentCode can model diverse garments with intricate geometric details, including different cuts, frills, and pleats. To better align GarmentCode with our model's language-driven requirements, we develop a refined version, GarmentCodeRC (Richer & Cleaner) with two major improvements (see Fig. 3)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.911, + 0.902 + ], + "angle": 0, + "content": "First, we modify stitching and panel positioning to better support diverse garment types, such as open-front jackets, high-waist skirts, and fitted pant legs. This allows us to model a broader range of real-world garments and aligns more closely with natural garment descriptions. Second, we simplify GarmentCode's JSON configuration to better suit LLM training. The original configuration is redundant, applying the same settings across all garment types. We optimize this by automatically removing irrelevant settings (e.g., omitting skirt-related parameters for upper-body garments). This adjustment reduces the average language token count from 900 to 350, decreasing ambiguity in LLM training and improving efficiency. Furthermore, to stabilize model train" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "2926" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.089, + 0.483, + 0.264 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.27, + 0.483, + 0.312 + ], + "angle": 0, + "content": "Figure 3. GarmentCodeRC. Left: new options to model open-front jackets, high-waist skirts, and tight pant legs. Right: simplified JSON configuration for more efficient LLM training." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.33, + 0.485, + 0.391 + ], + "angle": 0, + "content": "ing, we normalize all floating-point values to a \\([0,1]\\) range. Examples of the GarmentCodeRC JSON configuration are given in Sup. Mat. Throughout this paper, \"GarmentCode\" refers to our enriched and simplified version." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.395, + 0.484, + 0.668 + ], + "angle": 0, + "content": "LLM for Textual and Numeric Outputs. The output JSON file \\( J_{t} \\) contains both textual and numeric descriptions. Previous research [11, 14, 26] has shown that LLMs struggle in predicting continuous and precise numerical values. To overcome this, following ChatPose [11, 26], we encode these continuous values into language token embeddings and train a projection layer to recover accurate numerical values, such as garment length. In this setup, the language embedding for \\( Y_{t} \\) is represented as \\( H_{t} \\). From the textual output \\( Y_{t} \\), upon encountering the start token , we extract the textual content of the JSON file, which spans from to . The language embedding at the end token , where numeric information is encoded, is represented as \\( H_{t}^{\\mathrm{n}} \\). A projection layer uses this embedding to predict the numerical values as \\( \\mathbf{n} = g_{\\Theta}(H_{t}^{\\mathrm{n}}) \\). Finally, we merge these numerical values with the textual descriptions extracted earlier and format them into our final sewing pattern JSON configuration file \\( J_{t} \\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.68, + 0.355, + 0.695 + ], + "angle": 0, + "content": "3.2. Automatic Data Construction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.703, + 0.484, + 0.824 + ], + "angle": 0, + "content": "To train ChatGarment, we need image-to-JSON and text-to-JSON paired data. While some work like SewFormer [33] offers image-to-JSON pairs, they fall short in garment variety. For example, they lack tight-sleeve garments and only support a limited variety of skirts. Additionally, text descriptions for garment editing are missing in current datasets. We address these gaps by developing an automated data construction pipeline that integrates various resources." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Garment Generation, Simulation, and Rendering. Following GarmentCodeData [25], we sample 20,000 garments by randomizing values in sewing pattern configurations. To improve real-world garment relevance, we adjust the sampling ratios and assign lower weights for asymmetric gar" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.091, + 0.905, + 0.228 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.237, + 0.907, + 0.278 + ], + "angle": 0, + "content": "Figure 4. Data Construction Pipeline. We generate garments from JSON configurations, simulate them with ContourCraft [15] and render with Blender." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.301, + 0.905, + 0.36 + ], + "angle": 0, + "content": "ments and those with intricate cuffs. To create outfits, we pair top and bottom garments together or use single-piece items like dresses or jumpsuits. The garment set is draped on a SMPL-X [40] neutral model." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.364, + 0.909, + 0.651 + ], + "angle": 0, + "content": "We automate garment simulation using ContourCraft [15], chosen for its stability and scalability, with SMPL-X motion sequences sampled from the BEDLAM dataset [4]. Specific garment vertices are attached to corresponding body vertices using predefined rules to prevent garments from slipping off. For example, the upper edges of pants and skirts are connected to the nearest body vertices, and similar attachments are made for upper-body garment vertices near the shoulders. We filter out simulations with significant interpenetration or improper positioning in two steps. First, GarmentCode performs a sewing-pattern validation check, ensuring the integrity of panel edges and stitching. Second, we verify simulation results by assessing inter-penetration, and identify garments at risk of slipping off by inspecting the proximity of each garment vertex to the nearest body part. Each garment is then rendered in Blender from four random perspectives at elevation angles between 0 and 30 degrees, with garment and body textures sourced from the BEDLAM dataset. These steps are shown in Fig. 4." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.659, + 0.909, + 0.901 + ], + "angle": 0, + "content": "GPT-4o Labeling. The previous method [16] prompts GPT-4o to generate the general geometric descriptions for garments, such as the garment types, lengths, widths and sleeve lengths. However, these general descriptions are insufficient for garment detail understanding: the finetuned VLM struggles to interpret detailed features of garment components, such as pant cuffs, or execute specific garment edits. To address this, we construct a new dataset with low-level descriptions for individual garment parts. For image-based reconstruction, we instruct GPT-4o to generate descriptions for all visible garment parts in the image. For garment editing, we sample 5,000 garment parts (shirt main body, collar, sleeves, sleeve cuffs, waistband, pant legs, pant cuffs, and skirts), and use GPT-4o to generate descriptive text for each part. Assembling these components yields 20,000 new garments with detailed annotations." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2927" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.488, + 0.228 + ], + "angle": 0, + "content": "For garment editing, we randomly select two garments that differ in certain parts, with the editing prompt being: \"Change the garment sewing pattern by modifying PART-1 to DESC-1, PART-2 to DESC-2, while keeping other parts unchanged.\" PART-1, PART-2, represent part names that differ between the two garments, and DESC-1, DESC-2 are target garment part descriptions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.23, + 0.487, + 0.398 + ], + "angle": 0, + "content": "We finally build a synthetic dataset of 40K garments and 1 million images with text annotations, supporting garment reconstruction and editing from multimodal inputs. Plus, we use GPT-4o to label part-level details of 38K SHHQ [12] images to make the VLM better understand in-the-wild garment images. Compared to previous garment sewing pattern datasets [33], our dataset offers: 1) Diversity: GarmentCode generates a wider range of real-world garment types; 2) Detailed Descriptions: Garments are labeled with detailed geometric features; 3) Precise Control: Allows for precise, low-level garment manipulation." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.412, + 0.196, + 0.429 + ], + "angle": 0, + "content": "3.3. Training" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.437, + 0.486, + 0.501 + ], + "angle": 0, + "content": "We use LLaVA [32] as our base VLM and finetune it with LoRA [18], with its parameters denoted as \\(\\phi_{lora}\\). Additionally, we optimize the LLM token embedding \\(e_{token}\\), the LLM head layer \\(\\mathrm{lm\\_head}_{\\Theta'}\\), and the projection layer \\(g_{\\Theta}\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.507, + 0.485, + 0.596 + ], + "angle": 0, + "content": "Loss Function. As described in Sec. 3.1, the numeric embedding is extracted from the token [ENDS] as \\( H_{t}^{\\mathbf{n}} \\). An MLP projection layer maps the embedding \\( H_{t}^{\\mathbf{n}} \\) to 76 float values \\( \\mathbf{n} \\) in the JSON configuration. Given the ground truth text output \\( \\hat{Y}_{t} \\) and sewing pattern values \\( \\hat{\\mathbf{n}} \\), we optimize with:" + }, + { + "type": "equation", + "bbox": [ + 0.163, + 0.602, + 0.484, + 0.621 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\operatorname {C E} \\left(\\hat {Y} _ {t}, Y _ {t}\\right) + \\lambda_ {\\mathbf {n}} | (\\hat {\\mathbf {n}} - \\mathbf {n}) \\cdot \\mathbf {m} |, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.633, + 0.487, + 0.695 + ], + "angle": 0, + "content": "where the first term is a cross-entropy loss, and the second term is the L1 difference for numeric values. \\(\\mathbf{m}\\) is a mask that filters out irrelevant numeric values, and \\(\\lambda_{\\mathbf{n}} = 0.1\\) is the weight for the L1 loss. See more training details in Sup. Mat." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.71, + 0.476, + 0.727 + ], + "angle": 0, + "content": "3.4. Multi-turn Dialogue with Multimodal Inputs." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.487, + 0.903 + ], + "angle": 0, + "content": "After training, our model can handle multimodal input reconstruction, garment generation, and editing, as reflected in the training data. Additionally, it exhibits zero-shot reasoning for garment sewing patterns within multi-turn dialogues, despite being trained only on single-turn dialogue datasets. This enables users to create a garment from an image or text description and then refine it within the same interaction. A typical workflow might start with the VLM estimating a sewing pattern from an image. After reviewing the result, the user can request specific adjustments, providing a flexible, user-centered design process (see Fig. 5 for an example)." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.091, + 0.646, + 0.108 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.116, + 0.907, + 0.162 + ], + "angle": 0, + "content": "We conduct experiments on image-based reconstruction (Sec 4.1), garment editing (Sec 4.2) and text-based garment generation (Sec 4.3), along with ablation studies (Sec 4.4)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.171, + 0.771, + 0.187 + ], + "angle": 0, + "content": "4.1. Image-based Reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.195, + 0.909, + 0.347 + ], + "angle": 0, + "content": "**Benchmarks.** We conduct image-based reconstruction experiments on the CloSe [2] and Dress4D [54] datasets. CloSe is a large-scale scan collection including 3D clothing segmentations across 18 clothing categories. Following CloSe, we use 145 scans with accurate SMPL-X fits as our validation set. As CloSe offers limited instances of loose garments, such as skirts and dresses, we further add samples from Dress4D, which has real-world outfits recorded using a multi-view volumetric capture system. We include 4 loose-fitting outfits and 36 images rendered from them." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.35, + 0.911, + 0.442 + ], + "angle": 0, + "content": "Implementation Details. Our approach applies the Chain-of-Thought (CoT) [55] method. We first prompt ChatGarment to generate a text description of the outfit in the image. The generated text descriptions are combined with the input image to estimate the final garment JSON configuration. More details are provided in Sup. Mat." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.445, + 0.909, + 0.521 + ], + "angle": 0, + "content": "Baselines. We compare our method with two previous approaches, SewFormer [33] and GarmentRecovery [29]. We use the publicly available code and checkpoints for these methods in our evaluation. Additionally, we introduce several new baseline models for comparison:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.524, + 0.909, + 0.599 + ], + "angle": 0, + "content": "- DressCode. We employ DressCode [16] to generate garments from image descriptions. Specifically, we utilize GPT-4o to generate garment descriptions from input images. These descriptions are processed by DressCode to create sewing patterns." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.599, + 0.909, + 0.735 + ], + "angle": 0, + "content": "- GPT-4o. We design prompt templates and query GPT-4o using in-context learning (ICL) [10]. We use Garment Generator [22] as the sewing pattern configuration tool, as it has significantly fewer possible configurations and allows for more efficient template design and significantly fewer conversation rounds. We first prompt GPT-4o to choose the most suitable template from a predefined list, then query the parameter value by displaying three garment renderings with uniformly sampled parameter values." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.735, + 0.905, + 0.795 + ], + "angle": 0, + "content": "- LLaVA*. In this setup, we use the LLaVA model to directly output the numerical values for the sewing pattern instead of using the sewing pattern projection layer. The remaining components are identical to our method." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.796, + 0.905, + 0.855 + ], + "angle": 0, + "content": "- LLaVA-T. We first employ GPT-4o to generate detailed textual descriptions of the garment, and ChatGarment is prompted to reconstruct the sewing pattern based solely on the text description." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.856, + 0.905, + 0.901 + ], + "angle": 0, + "content": "- LLaVA-I. We input the image into the model and prompt it to directly output the sewing pattern without referring to text descriptions." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.524, + 0.909, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2928" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.137, + 0.089, + 0.548, + 0.209 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.223, + 0.215, + 0.46, + 0.228 + ], + "angle": 0, + "content": "(a) Image-based Sewing Pattern Estimation" + }, + { + "type": "image", + "bbox": [ + 0.554, + 0.089, + 0.853, + 0.353 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.555, + 0.356, + 0.848, + 0.37 + ], + "angle": 0, + "content": "(c) Text-based Sewing Pattern Generation and Editing" + }, + { + "type": "image", + "bbox": [ + 0.137, + 0.233, + 0.549, + 0.352 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.218, + 0.356, + 0.465, + 0.37 + ], + "angle": 0, + "content": "(b) Image-guided Sewing Pattern Generation" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.372, + 0.908, + 0.414 + ], + "angle": 0, + "content": "Figure 5. ChatGarment's dialog modes. Images and texts are adaptively combined to guide garment generation and editing. Text output between STARTS and ENDS contains information about the JSON configuration, which is then converted to 3D garments (JSON2Garment) for visualization and simulation purposes." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.426, + 0.485, + 0.554 + ], + "angle": 0, + "content": "
MethodsTarget-PoseA-PoseFailure Rate (↓)
CD (↓)F-Score (↑)CD (↓)F-Score (↑)
Sewformer [33]27.060.5520.660.584.3%
DressCode [16]20.160.6012.120.6111.4%
GarmentRecovery [29]9.690.555.750.64-
GPT-4o11.040.6210.440.650
LLaVA*5.290.726.290.760
LLaVA-T6.170.727.250.740
LLaVA-I4.320.754.930.780
Ours3.120.753.060.780
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.559, + 0.486, + 0.598 + ], + "angle": 0, + "content": "Table 1. Quantitative comparisons of Image-based garment reconstruction on Dress4D. ChatGarment achieves the best performance on all metrics." + }, + { + "type": "table", + "bbox": [ + 0.123, + 0.599, + 0.454, + 0.743 + ], + "angle": 0, + "content": "
MethodsCLOSE
CD (↓)F-Score (↑)Failure Rate (↓)
SewFormer [33]9.700.7088.33 %
DressCode [16]15.770.6167.0%
GarmentRecovery [29]2.390.785-
GPT-4o3.520.7550
LLaVA*3.080.7750
LLaVA-T3.760.7790
LLaVA-I4.550.7610
Ours2.940.7900
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.747, + 0.486, + 0.775 + ], + "angle": 0, + "content": "Table 2. Quantitative comparisons of Image-based reconstruction on CloSe. The garments are deformed to the target pose." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.79, + 0.484, + 0.836 + ], + "angle": 0, + "content": "Since methods use different default human poses, we apply Linear Blend Skinning [1] to re-pose all garments to the ground-truth body shape and pose for fair comparison." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.486, + 0.902 + ], + "angle": 0, + "content": "Results. We employ the mean chamfer distance, the mean F-Score at \\(\\tau = 0.001\\) and the stitching failure rate as quantitative metrics. As Tab. 1 and Tab. 2 show, ChatGarment significantly outperforms existing methods on loose garments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.43, + 0.908, + 0.701 + ], + "angle": 0, + "content": "in Dress4D. On the CloSE dataset, which mostly features tight garments, GarmentRecovery [29] does well by leveraging garment segmentation and normal maps as guidance in an optimization-based framework. However, these cues are insufficient for accurately reconstructing loose garments. Our method has similar performance in this case, but is more general. Also, our algorithm robustly reconstructs garments with no stitching failures. In contrast, DressCode and SewFormer often generate invalid garment panels leading to failed stitching. The primary difference lies in the output format: DressCode and SewFormer predict the panel edges and stitches of the sewing pattern, while our method and the GPT-4o-based approach estimate the JSON configuration. Although direct edge and stitch prediction offers greater flexibility in theory, our experiments show that it does not enhance reconstruction accuracy. Instead, it increases the likelihood of generating invalid sewing patterns when handling out-of-distribution (OOD) data." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.703, + 0.909, + 0.854 + ], + "angle": 0, + "content": "Our model demonstrates superior performance compared to LLaVA*, highlighting the effectiveness of the proposed projection layer. Furthermore, our CoT-based two-step reconstruction algorithm outperforms both image-based (LLaVA-I) and text-based (LLaVA-T) reconstruction baselines. The text descriptions detail garment types and styles, offering valuable information for reconstruction, while images reveal intricate details that are hard to describe in words. Thus, combining both text and image inputs is the most effective strategy for image-based garment reconstruction." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.908, + 0.901 + ], + "angle": 0, + "content": "Figure 6 shows our model accurately estimates garment type, style, and details like skirt hems and collar styles. In contrast, other methods often misinterpret or overlook these" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2929" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.086, + 0.905, + 0.539 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.547, + 0.908, + 0.577 + ], + "angle": 0, + "content": "Figure 6. Image-based Garment Reconstruction. Unlike SewFormer [33], DressCode [16], GarmentRecovery [29], and the GPT-4o-based method, which often make mistakes or ignores garments. ChatGarment faithfully captures the shape, style and composition of the garments." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.588, + 0.486, + 0.663 + ], + "angle": 0, + "content": "details. SewFormer occasionally fails to reconstruct garments due to detection or stitching failure, and GarmentRecovery lacks support for whole-body dress reconstruction. Additionally, inaccuracies in 2D garment segmentation [28] can lead to errors in GarmentRecovery reconstruction." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.665, + 0.486, + 0.741 + ], + "angle": 0, + "content": "Garment designers often create image sketches before making sewing patterns. Surprisingly, despite not being trained on garment sketch images, ChatGarment demonstrates strong generalization, accurately predicting garment types and styles from sketches (Fig. 8)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.755, + 0.262, + 0.772 + ], + "angle": 0, + "content": "4.2. Garment Editing" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.485, + 0.903 + ], + "angle": 0, + "content": "**Benchmarks.** To evaluate garment editing performance, we construct an extra evaluation dataset using methods outlined in Sec. 3.2. We manually select garment pairs with clear textual descriptions and substantial visual differences. This dataset consists of 135 garment pairs, each accompanied by corresponding images and text descriptions. The evaluation of garment editing involves three components: the source garment, the target garment, and the editing prompts. The" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.588, + 0.907, + 0.619 + ], + "angle": 0, + "content": "goal is to derive the sewing pattern for the target garment using the source garment and the editing prompts." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.623, + 0.909, + 0.669 + ], + "angle": 0, + "content": "Implementation Details. For garment editing, we concatenate the source garment JSON file and the editing prompt together and pass it into ChatGarment." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.673, + 0.907, + 0.703 + ], + "angle": 0, + "content": "Baselines. We modify DressCode [16] and SewFormer [33] for comparison with ChatGarment:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.705, + 0.909, + 0.81 + ], + "angle": 0, + "content": "- DressCode with edited text. We generate a text description of the edited garment and use DressCode to create the corresponding sewing patterns. The description is produced by GPT-4o, which takes as inputs: 1) garment editing instructions and 2) front and back views of the source garment. The generated description follows DressCode's original input format." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.811, + 0.909, + 0.902 + ], + "angle": 0, + "content": "- SewFormer with edited images. We obtain the front view of the edited garment with TurboEdit [9] to which we pass 1) the front view of the source garment, 2) the text description of the source garment, and 3) the editing instructions, and then use SewFormer to turn the image of the edited garment into the corresponding sewing pattern." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.705, + 0.909, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2930" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.089, + 0.152, + 0.143 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.153, + 0.095, + 0.289, + 0.143 + ], + "angle": 0, + "content": "Prompts: Change the shirt body to: [long vertical sections, 'evenly shaped sides', 'mildly loose fit around upper body, 'straight front opening']" + }, + { + "type": "image", + "bbox": [ + 0.293, + 0.089, + 0.348, + 0.143 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.355, + 0.09, + 0.409, + 0.143 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.419, + 0.091, + 0.471, + 0.143 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.155, + 0.152, + 0.209 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.153, + 0.158, + 0.285, + 0.208 + ], + "angle": 0, + "content": "Prompts: Change the sleeves to [short sleeves length], 'loose fit around the arms', 'wide sleeve openings relative to the arm size']" + }, + { + "type": "image", + "bbox": [ + 0.293, + 0.155, + 0.348, + 0.209 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.355, + 0.155, + 0.41, + 0.209 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.419, + 0.155, + 0.471, + 0.209 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.222, + 0.152, + 0.275 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.153, + 0.232, + 0.283, + 0.27 + ], + "angle": 0, + "content": "Prompts: Change the collar to: 'rounded neckline', 'slightly loose fit', 'even curvature throughout']" + }, + { + "type": "image", + "bbox": [ + 0.293, + 0.223, + 0.348, + 0.275 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.355, + 0.223, + 0.41, + 0.275 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.419, + 0.223, + 0.471, + 0.275 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.109, + 0.282, + 0.147, + 0.292 + ], + "angle": 0, + "content": "Source" + }, + { + "type": "image_caption", + "bbox": [ + 0.29, + 0.281, + 0.352, + 0.292 + ], + "angle": 0, + "content": "SewFormer" + }, + { + "type": "image_caption", + "bbox": [ + 0.355, + 0.282, + 0.414, + 0.292 + ], + "angle": 0, + "content": "DressCode" + }, + { + "type": "image_caption", + "bbox": [ + 0.431, + 0.282, + 0.459, + 0.292 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.09, + 0.572, + 0.143 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.581, + 0.099, + 0.718, + 0.14 + ], + "angle": 0, + "content": "Prompts: Change the sleeves to: 'long, full-length sleeves', 'loose fit', 'ends at wrist level', 'wide openings at the wrist']" + }, + { + "type": "image", + "bbox": [ + 0.726, + 0.09, + 0.778, + 0.143 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.792, + 0.091, + 0.844, + 0.143 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.85, + 0.09, + 0.902, + 0.143 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.155, + 0.572, + 0.208 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.581, + 0.165, + 0.712, + 0.203 + ], + "angle": 0, + "content": "Prompts: Change the pant legs to: [wide-leg silhouette], 'ankle-length hem', 'flowy fit', 'loose and relaxed cut']" + }, + { + "type": "image", + "bbox": [ + 0.726, + 0.155, + 0.778, + 0.209 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.792, + 0.155, + 0.844, + 0.209 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.85, + 0.155, + 0.902, + 0.209 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.519, + 0.222, + 0.572, + 0.275 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.581, + 0.226, + 0.716, + 0.275 + ], + "angle": 0, + "content": "Prompts: Change the shirt body to: ['straight vertical silhouette', 'moderate fit around the torso', 'hemline falls above the knees']" + }, + { + "type": "image", + "bbox": [ + 0.726, + 0.223, + 0.778, + 0.275 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.792, + 0.224, + 0.844, + 0.275 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.85, + 0.223, + 0.902, + 0.275 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.527, + 0.283, + 0.565, + 0.294 + ], + "angle": 0, + "content": "Source" + }, + { + "type": "image_caption", + "bbox": [ + 0.722, + 0.283, + 0.783, + 0.293 + ], + "angle": 0, + "content": "SewFormer" + }, + { + "type": "image_caption", + "bbox": [ + 0.792, + 0.283, + 0.849, + 0.293 + ], + "angle": 0, + "content": "DressCode" + }, + { + "type": "image_caption", + "bbox": [ + 0.865, + 0.283, + 0.892, + 0.293 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.304, + 0.908, + 0.334 + ], + "angle": 0, + "content": "Figure 7. Results on Garment Editing. We present six examples where the models (SewFormer [33], DressCode [16], and ChatGarment) edit the source garment shown in the source image according to the given editing instructions in the prompt boxes." + }, + { + "type": "image", + "bbox": [ + 0.119, + 0.347, + 0.203, + 0.515 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.143, + 0.521, + 0.179, + 0.535 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image", + "bbox": [ + 0.212, + 0.347, + 0.371, + 0.515 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.23, + 0.521, + 0.351, + 0.535 + ], + "angle": 0, + "content": "2D Sewing Pattern" + }, + { + "type": "image", + "bbox": [ + 0.373, + 0.348, + 0.453, + 0.517 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.372, + 0.521, + 0.453, + 0.534 + ], + "angle": 0, + "content": "3D Garment" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.548, + 0.484, + 0.59 + ], + "angle": 0, + "content": "Figure 8. Reconstruction results from garment sketches. The generated garments accurately capture the shape and style of the input sketch, despite our model being trained without sketches." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.605, + 0.484, + 0.727 + ], + "angle": 0, + "content": "Results. We compare our method with SewFormer and Dresscode on Chamfer Distance and F-score in Tab. 3, where our approach demonstrates superior performance. The qualitative results in Fig. 7 show that our method accurately modifies the targeted garment parts according to the input prompt. In contrast, SewFormer and DressCode often fail to precisely follow the prompt instructions and frequently alter other unintended areas of the garment." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.74, + 0.303, + 0.755 + ], + "angle": 0, + "content": "4.3. Text-based Generation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.486, + 0.903 + ], + "angle": 0, + "content": "Following DressCode [16], we use GPT-4o to generate 150 highly diverse text labels from 85 selected in-the-wild images featuring various tops, skirts, pants, and dresses. Next, we render the generated 3D garments with default textures and compute the CLIP score using the provided text labels. See quantitative results in Tab. 4. Our model achieves a higher CLIP score compared to DressCode [16], highlighting the superior prompt-following capabilities of our approach. The qualitative results are shown in Sup. Mat." + }, + { + "type": "table", + "bbox": [ + 0.586, + 0.341, + 0.832, + 0.393 + ], + "angle": 0, + "content": "
MethodsCD (↓)F-Score (↑)
SewFormer [33]28.770.548
DressCode [16]10.970.750
Ours2.510.893
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.398, + 0.908, + 0.426 + ], + "angle": 0, + "content": "Table 3. Quantitative comparisons for garment editing. ChatGarment outperforms other two methods by a large margin." + }, + { + "type": "table", + "bbox": [ + 0.608, + 0.435, + 0.811, + 0.474 + ], + "angle": 0, + "content": "
MethodsCLIP score (↑)
DressCode [16]22.8
Ours23.7
" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.479, + 0.908, + 0.521 + ], + "angle": 0, + "content": "Table 4. Quantitative comparisons for text-based garment generation. ChatGarment achieves a higher CLIP score than DressCode, indicating better alignment with the text inputs." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.535, + 0.618, + 0.55 + ], + "angle": 0, + "content": "4.4. Ablation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.559, + 0.908, + 0.605 + ], + "angle": 0, + "content": "We evaluate the impact of different aspects of ChatGarment, including multi-modal LLM backbones, and various training datasets. Please refer to Sup. Mat." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.618, + 0.634, + 0.633 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.643, + 0.909, + 0.9 + ], + "angle": 0, + "content": "We introduce ChatGarment, a VLM-based model that estimates garments from images and generates or edits garments based on text descriptions. Leveraging the capabilities of LLMs, ChatGarment supports multi-turn dialogue interactions for 3D garment design. It surpasses existing sewing pattern-specific and LLM-based methods in accurately estimating sewing patterns from a single image. While ChatGarment excels in many areas, it may struggle with precise garment editing. For instance, altering the skirt length might unintentionally affect other parts of the garment. This issue could be addressed with more detailed text instructions or a differential CSG optimizer [60] for finer adjustments. Nonetheless, ChatGarment offers the first VLM-based framework for unified garment estimation, generation, and editing. Future directions could include expanding into garment manufacturing design or improving clothing simulation with more fine-grained material attributes." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "2931" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.379, + 0.108 + ], + "angle": 0, + "content": "Acknowledgments and Disclosure" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.116, + 0.486, + 0.205 + ], + "angle": 0, + "content": "We thank Maria Korosteleva for help on GarmentCode, and Giorgio Becherini for advice on data acquisition and processing. Yuliang Xiu is funded by Max Planck Institute for Intelligent Systems, the Research Center for Industries of the Future (RCIF) at Westlake University, and the Westlake Education Foundation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.206, + 0.485, + 0.25 + ], + "angle": 0, + "content": "While MJB is a co-founder and Chief Scientist at Meshcapade, his research in this project was performed solely at, and funded solely by, the Max Planck Society." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.278, + 0.188, + 0.294 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.303, + 0.486, + 0.358 + ], + "angle": 0, + "content": "[1] Rinet Abdrashitov, Kim Raichstat, Jared Monsen, and David Hill. Robust skin weights transfer via weight inpainting. In SIGGRAPH Asia 2023 Technical Communications, pages 1-4. 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.36, + 0.484, + 0.415 + ], + "angle": 0, + "content": "[2] Dimitrije Antic, Garvita Tiwari, Batuhan Ozcomlekci, Ricardo Marin, and Gerard Pons-Moll. CloSe: A 3D clothing segmentation dataset and model. In International Conference on 3D Vision (3DV), 2024. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.418, + 0.486, + 0.485 + ], + "angle": 0, + "content": "[3] Bharat Lal Bhatnagar, Garvita Tiwari, Christian Theobalt, and Gerard Pons-Moll. Multi-garment net: Learning to dress 3d people from images. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5420-5430, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.488, + 0.485, + 0.558 + ], + "angle": 0, + "content": "[4] Michael J Black, Priyanka Patel, Joachim Tesch, and Jinlong Yang. Bedlam: A synthetic dataset of bodies exhibiting detailed lifelike animated motion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8726-8737, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.56, + 0.485, + 0.615 + ], + "angle": 0, + "content": "[5] Andrés Casado-Elvira, Marc Comino Trinidad, and Dan Casas. Pergamo: Personalized 3d garments from monocular video. In Computer Graphics Forum, pages 293–304. Wiley Online Library, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.617, + 0.486, + 0.686 + ], + "angle": 0, + "content": "[6] Enric Corona, Albert Pumarola, Guillem Alenya, Gerard Pons-Moll, and Francesc Moreno-Noguer. Simplicit: Topology-aware generative model for clothed people. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11875-11885, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.688, + 0.485, + 0.743 + ], + "angle": 0, + "content": "[7] R Daněřek, Endri Dibra, Cengiz Öztireli, Remo Ziegler, and Markus Gross. Deepgarment: 3d garment shape estimation from a single image. In Computer Graphics Forum, pages 269-280. Wiley Online Library, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.746, + 0.485, + 0.813 + ], + "angle": 0, + "content": "[8] Luca De Luigi, Ren Li, Benoit Guillard, Mathieu Salzmann, and Pascal Fua. Drapenet: Garment generation and self-supervised draping. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1451-1460, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.816, + 0.486, + 0.87 + ], + "angle": 0, + "content": "[9] Gilad Deutch, Rinon Gal, Daniel Garibi, Or Patashnik, and Daniel Cohen-Or. Turboedit: Text-based image editing using few-step diffusion models. arXiv preprint arXiv:2408.00735, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.485, + 0.902 + ], + "angle": 0, + "content": "[10] Qingxiu Dong, Lei Li, Damai Dai, Ce Zheng, Jingyuan Ma, Rui Li, Heming Xia, Jingjing Xu, Zhiyong Wu, Tianyu" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.303, + 0.486, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.907, + 0.12 + ], + "angle": 0, + "content": "Liu, et al. A survey on in-context learning. arXiv preprint arXiv:2301.00234, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.907, + 0.162 + ], + "angle": 0, + "content": "[11] Yao Feng, Jing Lin, Sai Kumar Dwivedi, Yu Sun, Priyanka Patel, and Michael J. Black. ChatPose: Chatting about 3d human pose. In CVPR, 2024. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.164, + 0.907, + 0.218 + ], + "angle": 0, + "content": "[12] Jianglin Fu, Shikai Li, Yuming Jiang, Kwan-Yee Lin, Chen Qian, Chen-Change Loy, Wayne Wu, and Ziwei Liu. Stylegan-human: A data-centric odyssey of human generation. arXiv preprint, arXiv:2204.11823, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.22, + 0.907, + 0.274 + ], + "angle": 0, + "content": "[13] Rao Fu, Xiao Zhan, Yiwen Chen, Daniel Ritchie, and Srinath Sridhar. Shapecraft: A recursive text-conditioned 3d shape generation model. Advances in Neural Information Processing Systems, 35:8882-8895, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.275, + 0.907, + 0.343 + ], + "angle": 0, + "content": "[14] Siavash Golkar, Mariel Pettee, Michael Eickenberg, Alberto Bietti, Miles Cranmer, Geraud Krawezik, Francois Lanusse, Michael McCabe, Ruben Ohana, Liam Parker, et al. xVal: A continuous number encoding for large language models. arXiv preprint arXiv:2310.02989, 2023. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.345, + 0.907, + 0.412 + ], + "angle": 0, + "content": "[15] Artur Grigorev, Giorgio Becherini, Michael Black, Otmar Hilliges, and Bernhard Thomaszewski. Contourcraft: Learning to resolve intersections in neural multi-garment simulations. In ACM SIGGRAPH 2024 Conference Papers, pages 1-10, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.414, + 0.907, + 0.468 + ], + "angle": 0, + "content": "[16] Kai He, Kaixin Yao, Qixuan Zhang, Jingyi Yu, Lingjie Liu, and Lan Xu. Dresscode: Autoregressively sewing and generating garments from text guidance. ACM Transactions on Graphics (TOG), 43(4):1-13, 2024. 2, 3, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.47, + 0.907, + 0.525 + ], + "angle": 0, + "content": "[17] Yining Hong, Haoyu Zhen, Peihao Chen, Shuhong Zheng, Yilun Du, Zhenfang Chen, and Chuang Gan. 3d-llm: Injecting the 3d world into large language models. Advances in Neural Information Processing Systems, 36:20482-20494, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.526, + 0.907, + 0.579 + ], + "angle": 0, + "content": "[18] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In ICLR, 2022. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.581, + 0.907, + 0.65 + ], + "angle": 0, + "content": "[19] Ziniu Hu, Ahmet Iscen, Aashi Jain, Thomas Kipf, Yisong Yue, David A Ross, Cordelia Schmid, and Alireza Fathi. Scenecraft: An IIm agent for synthesizing 3d scenes as blender code. In *Forty-first International Conference on Machine Learning*, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.651, + 0.907, + 0.705 + ], + "angle": 0, + "content": "[20] Yangyi Huang, Hongwei Yi, Yuliang Xiu, Tingting Liao, Ji-axiang Tang, Deng Cai, and Justus Thies. TeCH: Text-guided Reconstruction of Lifelike Clothed Humans. In International Conference on 3D Vision (3DV), 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.707, + 0.907, + 0.775 + ], + "angle": 0, + "content": "[21] Boyi Jiang, Juyong Zhang, Yang Hong, Jinhao Luo, Ligang Liu, and Hujun Bao. Bcnet: Learning body and cloth shape from a single image. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XX 16, pages 18-35. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.777, + 0.907, + 0.816 + ], + "angle": 0, + "content": "[22] Maria Korosteleva and Sung-Hee Lee. Generating datasets of 3d garments with sewing patterns. arXiv preprint arXiv:2109.05633, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.818, + 0.907, + 0.871 + ], + "angle": 0, + "content": "[23] Maria Korosteleva and Sung-Hee Lee. Neuraltailer: Reconstructing sewing pattern structures from 3d point clouds of garments. ACM Transactions on Graphics (TOG), 41(4):1-16, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.907, + 0.902 + ], + "angle": 0, + "content": "[24] Maria Korosteleva and Olga Sorkine-Hornung. GarmentCode: Programming parametric sewing patterns. ACM Transaction" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "2932" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.484, + 0.119 + ], + "angle": 0, + "content": "on Graphics, 42(6), 2023. SIGGRAPH ASIA 2023 issue. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.121, + 0.484, + 0.189 + ], + "angle": 0, + "content": "[25] Maria Korosteleva, Timur Levent Kesdogan, Fabian Kemper, Stephan Wenninger, Jasmin Koller, Yuhan Zhang, Mario Botsch, and Olga Sorkine-Hornung. GarmentCodeData: A dataset of 3D made-to-measure garments with sewing patterns. In Computer Vision – ECCV 2024, 2024. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.19, + 0.484, + 0.245 + ], + "angle": 0, + "content": "[26] Peter Kulits, Haiwen Feng, Weiyang Liu, Victoria Abrevaya, and Michael J Black. Re-thinking inverse graphics with large language models. Transactions on Machine Learning Research, 2024. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.246, + 0.484, + 0.301 + ], + "angle": 0, + "content": "[27] Boqian Li, Xuan Li, Ying Jiang, Tianyi Xie, Feng Gao, Huamin Wang, Yin Yang, and Chenfanfu Jiang. Garmentdreamer: 3dgs guided garment synthesis with diverse geometry and texture details. In 3DV, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.302, + 0.484, + 0.356 + ], + "angle": 0, + "content": "[28] Peike Li, Yunqiu Xu, Yunchao Wei, and Yi Yang. Self-correction for human parsing. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(6):3260-3271, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.358, + 0.484, + 0.413 + ], + "angle": 0, + "content": "[29] Ren Li, Corentin Dumery, Benoit Guillard, and Pascal Fua. Garment recovery with shape and deformation priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1586-1595, 2024. 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.414, + 0.484, + 0.455 + ], + "angle": 0, + "content": "[30] Ren Li, Benoit Guillard, and Pascal Fua. Isp: Multi-layered garment draping with implicit sewing patterns. Advances in Neural Information Processing Systems, 36, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.456, + 0.484, + 0.524 + ], + "angle": 0, + "content": "[31] Dingning Liu, Xiaoshui Huang, Yuenan Hou, Zhihui Wang, Zhenfei Yin, Yongshun Gong, Peng Gao, and Wanli Ouyang. Uni3d-llm: Unifying point cloud perception, generation and editing with large language models. arXiv preprint arXiv:2402.03327, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.525, + 0.484, + 0.553 + ], + "angle": 0, + "content": "[32] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning, 2023. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.553, + 0.484, + 0.608 + ], + "angle": 0, + "content": "[33] Lijuan Liu, Xiangyu Xu, Zhijie Lin, Jiabin Liang, and Shuicheng Yan. Towards garment sewing pattern reconstruction from a single image. ACM Transactions on Graphics (TOG), 42(6):1-15, 2023. 2, 3, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.609, + 0.484, + 0.678 + ], + "angle": 0, + "content": "[34] Weiyang Liu, Zeju Qiu, Yao Feng, Yuliang Xiu, Yuxuan Xue, Longhui Yu, Haiwen Feng, Zhen Liu, Juyeon Heo, Songyou Peng, Yandong Wen, Michael J. Black, Adrian Weller, and Bernhard Scholkopf. Parameter-efficient orthogonal finetuning via butterfly factorization. In ICLR, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.679, + 0.484, + 0.747 + ], + "angle": 0, + "content": "[35] Zhen Liu, Yao Feng, Yuliang Xiu, Weiyang Liu, Liam Paull, Michael J. Black, and Bernhard Scholkopf. Ghost on the shell: An expressive representation of general 3d shapes. In The Twelfth International Conference on Learning Representations, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.748, + 0.484, + 0.817 + ], + "angle": 0, + "content": "[36] Zhongjin Luo, Haolin Liu, Chenghong Li, Wanghao Du, Zirong Jin, Yinyu Nie, Weikai Chen, and Xiaoguang Han. Garverselod: High-fidelity 3d garment reconstruction from a single in-the-wild image using a dataset with levels of details. In ACM Transactions on Graphics (TOG), 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.818, + 0.484, + 0.873 + ], + "angle": 0, + "content": "[37] Qianli Ma, Jinlong Yang, Siyu Tang, and Michael J. Black. The power of points for modeling humans in clothing. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 10974-10984, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.874, + 0.484, + 0.902 + ], + "angle": 0, + "content": "[38] Qianli Ma, Jinlong Yang, Michael J. Black, and Siyu Tang. Neural point-based shape modeling of humans in challenging" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.908, + 0.12 + ], + "angle": 0, + "content": "clothing. In International Conference on 3D Vision (3DV), pages 679-689, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.121, + 0.909, + 0.176 + ], + "angle": 0, + "content": "[39] Miles Macklin. Warp: A high-performance python framework forgpu simulation and graphics. https://github.com/nvidia/warp, 2022. NVIDIA GPU Technology Conference (GTC). 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.177, + 0.908, + 0.232 + ], + "angle": 0, + "content": "[40] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed A. A. Osman, Dimitrios Tzionas, and Michael J. Black. Expressive body capture: 3D hands, face, and body from a single image. In CVPR, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.233, + 0.908, + 0.272 + ], + "angle": 0, + "content": "[41] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. In ICLR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.274, + 0.908, + 0.344 + ], + "angle": 0, + "content": "[42] Lingteng Qiu, Guanying Chen, Jiapeng Zhou, Mutian Xu, Junle Wang, and Xiaoguang Han. Rec-mv: Reconstructing 3d dynamic cloth from monocular videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4637-4646, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.344, + 0.906, + 0.399 + ], + "angle": 0, + "content": "[43] Zeju Qiu, Weiyang Liu, Haiwen Feng, Yuxuan Xue, Yao Feng, Zhen Liu, Dan Zhang, Adrian Weller, and Bernhard Schölkopf. Controlling text-to-image diffusion by orthogonal finetuning. In NeurIPS, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.4, + 0.906, + 0.468 + ], + "angle": 0, + "content": "[44] Zeju Qiu, Weiyang Liu, Haiwen Feng, Zhen Liu, Tim Z Xiao, Katherine M Collins, Joshua B Tenenbaum, Adrian Weller, Michael J Black, and Bernhard Scholkopf. Can large language models understand symbolic graphics programs? arXiv preprint arXiv:2408.08313, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.469, + 0.908, + 0.525 + ], + "angle": 0, + "content": "[45] Shunsuke Saito, Zeng Huang, Ryota Natsume, Shigeo Morishima, Hao Li, and Angjoo Kanazawa. PIFu: Pixel-aligned implicit function for high-resolution clothed human digitization. In ICCV, pages 2304-2314, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.525, + 0.906, + 0.579 + ], + "angle": 0, + "content": "[46] Shunsuke Saito, Tomas Simon, Jason Saragih, and Hanbyul Joo. PIFuHD: Multi-Level Pixel-Aligned Implicit Function for High-Resolution 3D Human Digitization. In CVPR, pages 81–90, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.581, + 0.906, + 0.622 + ], + "angle": 0, + "content": "[47] Nikolaos Sarafianos, Tuur Stuyck, Xiaoyu Xiang, Yilei Li, Jovan Popovic, and Rakesh Ranjan. Garment3dgen: 3d garment stylization and texture generation. In 3DV, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.623, + 0.908, + 0.678 + ], + "angle": 0, + "content": "[48] Tianchang Shen, Jun Gao, Kangxue Yin, Ming-Yu Liu, and Sanja Fidler. Deep marching tetrahedra: a hybrid representation for high-resolution 3d shape synthesis. In Advances in Neural Information Processing Systems (NeurIPS), 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.679, + 0.908, + 0.718 + ], + "angle": 0, + "content": "[49] Yu Shen, Junbang Liang, and Ming C. Lin. Gan-based garment generation using sewing pattern images. In ECCV, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.72, + 0.908, + 0.774 + ], + "angle": 0, + "content": "[50] Astitva Srivastava, Pranav Manu, Amit Raj, Varun Jampani, and Avinash Sharma. Wordrobe: Text-guided generation of textured 3d garments. In ECCV, pages 458-475. Springer, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.776, + 0.908, + 0.844 + ], + "angle": 0, + "content": "[51] Zhaoqi Su, Weilin Wan, Tao Yu, Lingjie Liu, Lu Fang, Wenping Wang, and Yebin Liu. Mulaycap: Multi-layer human performance capture using a monocular video camera. IEEE Transactions on Visualization and Computer Graphics, 28(4): 1862-1879, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.845, + 0.908, + 0.9 + ], + "angle": 0, + "content": "[52] Chunyi Sun, Junlin Han, Weijian Deng, Xinlong Wang, Zishan Qin, and Stephen Gould. 3d-gpt: Procedural 3d modeling with large language models. arXiv preprint arXiv:2310.12945, 2023. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.909, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2933" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.486, + 0.146 + ], + "angle": 0, + "content": "[53] Tuanfeng Y. Wang, Duygu Ceylan, Jovan Popovic, and Niloy J. Mitra. Learning a shared shape space for multimodal garment design. ACM Trans. Graph., 37(6), 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.485, + 0.219 + ], + "angle": 0, + "content": "[54] Wenbo Wang, Hsuan-I Ho, Chen Guo, Boxiang Rong, Artur Grigorev, Jie Song, Juan Jose Zarate, and Otmar Hilliges. 4D-DRESS: A 4D dataset of real-world human clothing with semantic annotations. In IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2024. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.22, + 0.485, + 0.288 + ], + "angle": 0, + "content": "[55] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824-24837, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.29, + 0.485, + 0.331 + ], + "angle": 0, + "content": "[56] Yuliang Xiu, Jinlong Yang, Dimitrios Tzionas, and Michael J. Black. ICON: Implicit Clothed humans Obtained from Normals. In CVPR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.333, + 0.482, + 0.374 + ], + "angle": 0, + "content": "[57] Yuliang Xiu, Jinlong Yang, Xu Cao, Dimitrios Tzionas, and Michael J. Black. ECON: Explicit Clothed humans Optimized via Normal integration. In CVPR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.376, + 0.484, + 0.444 + ], + "angle": 0, + "content": "[58] Runsen Xu, Xiaolong Wang, Tai Wang, Yilun Chen, Jiangmiao Pang, and Dahua Lin. Pointllm: Empowering large language models to understand point clouds. In European Conference on Computer Vision, pages 131-147. Springer, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.446, + 0.484, + 0.515 + ], + "angle": 0, + "content": "[59] Zhengming Yu, Zhiyang Dou, Xiaoxiao Long, Cheng Lin, Zekun Li, Yuan Liu, Norman Müller, Taku Komura, Marc Habermann, Christian Theobalt, et al. Surf-d: Generating high-quality surfaces of arbitrary topologies using diffusion models. In ECCV, pages 419-438. Springer, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.517, + 0.484, + 0.571 + ], + "angle": 0, + "content": "[60] Haocheng Yuan, Adrien Bousseau, Hao Pan, Chengquan Zhang, Niloy J Mitra, and Changjian Li. Diffcsg: Differentiable csg via rasterization. arXiv preprint arXiv:2409.01421, 2024.8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.574, + 0.484, + 0.642 + ], + "angle": 0, + "content": "[61] Longwen Zhang, Ziyu Wang, Qixuan Zhang, Qiwei Qiu, Anqi Pang, Haoran Jiang, Wei Yang, Lan Xu, and Jingyi Yu. Clay: A controllable large-scale generative model for creating high-quality 3d assets. ACM Transactions on Graphics (TOG), 43 (4):1-20, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.644, + 0.484, + 0.713 + ], + "angle": 0, + "content": "[62] Heming Zhu, Lingteng Qiu, Yuda Qiu, and Xiaoguang Han. Registering explicit to implicit: Towards high-fidelity garment mesh reconstruction from single images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3845-3854, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.486, + 0.713 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "2934" + } + ] +] \ No newline at end of file diff --git a/2025/ChatGarment_ Garment Estimation, Generation and Editing via Large Language Models/425b09e1-924e-4a43-b277-eb8a41387d21_origin.pdf b/2025/ChatGarment_ Garment Estimation, Generation and Editing via Large Language Models/425b09e1-924e-4a43-b277-eb8a41387d21_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..093234778b2e3fd3b76b3f06789641f56926c25b --- /dev/null +++ b/2025/ChatGarment_ Garment Estimation, Generation and Editing via Large Language Models/425b09e1-924e-4a43-b277-eb8a41387d21_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:db35cc0b5aba3776004346b73213e8ec420b5e3e93a6de3f2bfbe20f9cfa8836 +size 2391850 diff --git a/2025/ChatGarment_ Garment Estimation, Generation and Editing via Large Language Models/full.md b/2025/ChatGarment_ Garment Estimation, Generation and Editing via Large Language Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..f1861afb3ca6a912fcbe3420e88604e3d6ed703a --- /dev/null +++ b/2025/ChatGarment_ Garment Estimation, Generation and Editing via Large Language Models/full.md @@ -0,0 +1,349 @@ +# ChatGarment: Garment Estimation, Generation and Editing via Large Language Models + +Siyuan Bian $^{1,2}$ Chenghao Xu $^{1,3}$ Yuliang Xiu $^{1,4}$ Artur Grigorev $^{1,5}$ Zhen Liu $^{1,6}$ Cewu Lu $^{2}$ Michael J. Black $^{1}$ Yao Feng $^{7,8,\dagger}$ + +$^{1}$ Max Planck Institute for Intelligent Systems $^{2}$ Shanghai Jiao Tong University + +$^{3}$ EPFL $^{4}$ Westlake University $^{5}$ ETH Zürich $^{6}$ Mila, University of Montreal $^{7}$ Meshcapade $^{8}$ Stanford University + +chatgarment.github.io + +![](images/abd2410ab08bf1b0d7bd21125ec50641268779933d62491b99f6ccfad435656e.jpg) +Figure 1. As a multimodal 3D garment creator, ChatGarment understands both images and language. It can estimate complex 3D garments represented as sewing patterns from a single image, which can be easily animated and simulated. It also supports garment editing based on text instructions. + +![](images/e47b86281e490117fe4b90dff48e9929045d0f0b0a9a4e34ea853b57c70dc216.jpg) + +# Abstract + +We introduce ChatGarment, a novel approach that leverages large vision-language models (VLMs) to automate the estimation, generation, and editing of 3D garments from images or text descriptions. Unlike previous methods that struggle in real-world scenarios or lack interactive editing capabilities, ChatGarment can estimate sewing patterns from in-the-wild images or sketches, generate them from text descriptions, and edit garments based on user instructions, all within an interactive dialogue. These sewing patterns can then be draped on a 3D body and animated. This is achieved by finetuning a VLM to directly generate a JSON file that includes both textual descriptions of garment types and styles, as well as continuous numerical attributes. This JSON file is then used to create sewing patterns through a programming parametric model. To support this, we refine the existing programming model, GarmentCode, by expanding its garment type coverage and simplifying its structure for efficient VLM fine-tuning. Additionally, we construct a + +large-scale dataset of image-to-sewing-pattern and text-to-sewing-pattern pairs through an automated data pipeline. Extensive evaluations demonstrate ChatGarment's ability to accurately reconstruct, generate, and edit garments from multimodal inputs, highlighting its potential to simplify workflows in fashion and gaming applications. Code and data are available at https://chatgarment.github.io/. + +# 1. Introduction + +Imagine seeing a photo of someone in a nice looking outfit that you would like for yourself or your 3D avatar. Our goal is to take that photo and convert it into a 2D sewing pattern and then enable the user to use language to edit the pattern, for example, making the sleeves longer or editing the neckline. We then drape the garment on a 3D body and animate it. This is illustrated in Fig. 1. With this approach, given just an image and/or text, we construct sewing patterns for 3D garments that can be immediately used in fashion or gaming applications. This lightweight process is in stark contrast to how sewing patterns and 3D garments are traditionally created by an artist or clothing designer. In contrast to the traditional labor-intensive process, our approach exploits + +large vision-language models (VLMs) for the first time to democratize the capture and design of clothing. + +Recent work explores the generation or estimation of 3D clothing using text and/or images as input [6, 8, 21, 42]. These methods use various 3D representations including 3D meshes, point clouds, or implicit representations like Unsigned Distance Functions (UDFs) and Signed Distance Functions (SDFs). We observe, however, that most clothing is designed and manufactured using 2D patterns, and posit that this is a more natural representation. Sewing patterns can be seamlessly integrated into existing garment design pipelines for animation and manufacturing. Recent approaches [16, 30, 33] use sewing patterns for garment estimation and generation, but mapping images or text to sewing patterns remains challenging due to limited training data, leading to poor generalization across diverse clothing types. These methods also does not support interactive editing, making post-processing labor-intensive for 3D artists. Fortunately, recent work, such as GarmentCode [24, 25], provides a programmatic framework for encoding garment information into a JSON file. This JSON file includes textual descriptions of garment types, styles and numerical attributes, offering a rich representation of sewing patterns. The JSON file is then processed by the GarmentCode framework to produce 2D sewing patterns, which can subsequently be draped on a 3D body. However, to date, there is no automated method to directly extract a GarmentCode representation (JSON file) from an image or text description. + +Our observation is that a JSON file resembles natural language, making it suitable for interpretation by large language models (LLMs). At the same time, Vision-LLMs (VLMs) excel at understanding images and garment concepts, such as types of garments and specific features like long sleeves. If we enable the VLM to translate its garment understanding into the structured format of GarmentCode, then it becomes capable of generating sewing patterns. This allows the VLM to take an image as input and output a garment, while also supporting garment generation and editing through text descriptions or interactive dialogue. + +Specifically, we introduce ChatGarment, a method that finetunes a VLM to take text queries, optionally with images, as input and outputs a JSON garment file. This JSON file is then used for GarmentCode generation. The JSON file includes textual descriptions of garment types (e.g., skirts, pants), styles (e.g., collar type), and continuous numerical attributes (e.g., pant length). Inspired by previous work [11, 14, 26], we introduce a new token () to represent these numerical values and train an MLP projection layer to decode them from the token embedding. Those numbers, combined with the VLM's language output, are formatted into the final JSON file. In our approach, we keep the VLM's vision encoder fixed, fine-tune the language model using LoRA [18], and jointly train the MLP projection layer. To + +adapt GarmentCode [24] for more efficient VLM training, we create an improved version that supports a wider range of clothing types (e.g. open-jacket) and removes unnecessary settings to reduce token usage in training. We also develop an automated data construction pipeline that leverages existing tools for garment generation, draping, simulation, rendering, and automatic labeling with GPT-4o. This pipeline enables building a dataset with image-to-GarmentCode and text-to-GarmentCode pairs, including 20K new garments and 1M images with detailed descriptions, supporting both garment creation and editing. + +We evaluate ChatGarment across a diverse set of tasks, including those specifically trained for, and novel applications that demonstrate the generalization capabilities of VLMs. For single-image garment reconstruction, ChatGarment outperforms prior sewing-pattern-specific and LLM-based models on the Dress4D [54] and CloSE [2] evaluation datasets. Additionally, we demonstrate garment editing, generation, and multi-turn dialogue, highlighting our approach's versatility. ChatGarment can flexibly use text and images to create and edit 3D garments that can be readily animated, introducing a new workflow for garment design. + +We summarize our contributions as follows: + +- A unified model capable of estimating, generating and editing sewing patterns from multimodal inputs. +- The first approach to leverage VLMs for directly generating JSON files for garment creation. +- A refined version of GarmentCode that supports more diverse garment types and is optimized for VLM training. +- An automatic data construction pipeline for generating image-to-sewing-pattern and text-to-sewing-pattern data, along with the release of a large-scale dataset. + +# 2. Related Work + +Garment Creation. Most 3D garments have predefined templates (e.g., sewing patterns) and categories (e.g., skirts, shirts, pants, dresses, etc.), but their non-rigid nature makes them highly deformable with dynamic motions. Efforts to create 3D garments can be grouped into "template-based" or "free-form" methods, based on their 3D representations. + +A line of template-based approaches register predefined garment templates to 2D observations, which could be real captures, e.g. from monocular video [5, 21, 36, 42] and images [3, 6, 7, 62], generated data [49], or derived from the "proxy gradient" computed from 2D diffusion models [27, 47] via score distillation sampling (SDS) [41]. When the template is classified correctly, the predefined garment template, refined through a coarse-to-fine step, ensures topological stability and correctness. Unfortunately, template-based methods struggle to accurately model garments that deviate from the predefined templates, particularly those with diverse styles, such as long skirts, or garments that undergo topological changes, like opening a jacket. + +To address these limitations, free-form representations like (un)signed distance functions [6, 8, 50, 59], occupancy fields [45, 46, 56, 57], tetrahedrons [20, 35, 48], and point clouds [37, 38] are leveraged. However, this flexibility can compromise the inherent structure of garments, potentially leading to non-garment shapes and floating artifacts. Without shape regularization, incomplete observations lead to the reconstructed unseen parts being overly smooth. Even worse, rigging or simulating these free-form garments becomes extremely challenging. + +The sewing pattern is a well-balanced solution – with pre-defined yet flexible templates. Several methods exploit sewing patterns to reconstruct 3D garments. MulayCap [51] estimates the parameters of a sewing pattern from a monocular video. ISP [30] parameterizes sewing patterns as implicit functions in a 2D space. Using ISP as a garment representation, GarmentRecovery [29] trains a feedforward network to infer sewing patterns from in-the-wild photos. Neural-Tailor [23] uses an RNN to produce more complex sewing patterns from point clouds. Wang et al. [53] learn a shared garment shape space that can be inferred from multimodal input such as sketches, SewFormer [33] uses a transformer to infer multi-panel patterns from monocular images, and DressCode [16] estimates quantized sewing patterns with a GPT-based model from text inputs. Despite promising results, these approaches are limited to basic sewing patterns, featuring only front and back panels, or constrained by the limited sewing patterns seen in the training data. Additionally, none of the above works have explored using programming parametric sewing patterns, such as GarmentCode [24, 25] to create more complex garments, either in generation or reconstruction. + +VLMs for 3D tasks. With their broad understanding of the world, LLMs and VLMs are widely used to reason about 3D. There has been significant work on generating, editing, and analyzing 3D objects and scenes [13, 17, 19, 31, 44, 52, 58, 61]. While these methods show the power of VLMs to reason about objects, they do not address our problem of representing 2D patterns that correspond to 3D shapes – this is a special property of garments. Some attempts have been made to connect sewing patterns with language, such as DressCode [16]. This motivates us to present ChatGarment, which connects a large language model to GarmentCode, sharing a similar approach to recent strategies [11, 32] by finetuning LLMs [18, 34, 43]. Powered by VLMs, ChatGarment significantly improves generalization and robustness in reconstructing out-of-domain images and increases diversity in 3D clothing generation. It also supports sewing pattern editing and multi-turn dialogue, greatly simplifying the 3D garment workflow. + +![](images/6fac4910610603e5339c6775b96168d2d4399033ac95c6da4987b2d00d66b316.jpg) +Figure 2. Pipeline of ChatGarment. ChatGarment takes text or an image as input and generates a JSON file. The JSON file is decoded into 2D sewing patterns using GarmentCode [24] and then draped onto the human body. The final 3D garments are compatible with simulation software (e.g., MAYA, Blender, Style3D, etc.). + +# 3. Method + +# 3.1. Pipeline + +The goal of ChatGarment is to simplify the workflow of garment creation by automatically estimating, generating, and editing garments from multimodal image and text input. To achieve this, we finetune LLaVA [32], a multimodal large language model $f_{\phi}$ . It takes an input image $X_{v}$ and text prompts $X_{q}$ , and produces a textual response $Y_{t}$ as $Y_{t} = f_{\phi}(X_{q},X_{v})$ or $Y_{t} = f_{\phi}(X_{q})$ in the absence of an image. From this output, a structured JSON file $J_{t}$ is extracted as a subset of $Y_{t}$ , which will be decoded into sewing patterns using the GarmentCode [24] decoder $D_{g}$ . This pipeline is illustrated in Fig. 2. + +LLM-friendly Rich GarmentCode. We use GarmentCode to convert the JSON file into a 3D garment mesh in the canonical pose. GarmentCode [24] is a programming parametric model that uses a domain-specific language (DSL) for creating 2D sewing patterns, with a hierarchical JSON structure to define garment components and stitching rules. These patterns can be draped onto body models via a warp-based pattern stitching algorithm [24, 39], and further be simulated easily. GarmentCode can model diverse garments with intricate geometric details, including different cuts, frills, and pleats. To better align GarmentCode with our model's language-driven requirements, we develop a refined version, GarmentCodeRC (Richer & Cleaner) with two major improvements (see Fig. 3). + +First, we modify stitching and panel positioning to better support diverse garment types, such as open-front jackets, high-waist skirts, and fitted pant legs. This allows us to model a broader range of real-world garments and aligns more closely with natural garment descriptions. Second, we simplify GarmentCode's JSON configuration to better suit LLM training. The original configuration is redundant, applying the same settings across all garment types. We optimize this by automatically removing irrelevant settings (e.g., omitting skirt-related parameters for upper-body garments). This adjustment reduces the average language token count from 900 to 350, decreasing ambiguity in LLM training and improving efficiency. Furthermore, to stabilize model train + +![](images/2120e0a1dba49351e967bf370b866a4c51c61e51ba6e2dd0686596fdd41cdc6a.jpg) +Figure 3. GarmentCodeRC. Left: new options to model open-front jackets, high-waist skirts, and tight pant legs. Right: simplified JSON configuration for more efficient LLM training. + +ing, we normalize all floating-point values to a $[0,1]$ range. Examples of the GarmentCodeRC JSON configuration are given in Sup. Mat. Throughout this paper, "GarmentCode" refers to our enriched and simplified version. + +LLM for Textual and Numeric Outputs. The output JSON file $J_{t}$ contains both textual and numeric descriptions. Previous research [11, 14, 26] has shown that LLMs struggle in predicting continuous and precise numerical values. To overcome this, following ChatPose [11, 26], we encode these continuous values into language token embeddings and train a projection layer to recover accurate numerical values, such as garment length. In this setup, the language embedding for $Y_{t}$ is represented as $H_{t}$ . From the textual output $Y_{t}$ , upon encountering the start token , we extract the textual content of the JSON file, which spans from to . The language embedding at the end token , where numeric information is encoded, is represented as $H_{t}^{\mathrm{n}}$ . A projection layer uses this embedding to predict the numerical values as $\mathbf{n} = g_{\Theta}(H_{t}^{\mathrm{n}})$ . Finally, we merge these numerical values with the textual descriptions extracted earlier and format them into our final sewing pattern JSON configuration file $J_{t}$ . + +# 3.2. Automatic Data Construction + +To train ChatGarment, we need image-to-JSON and text-to-JSON paired data. While some work like SewFormer [33] offers image-to-JSON pairs, they fall short in garment variety. For example, they lack tight-sleeve garments and only support a limited variety of skirts. Additionally, text descriptions for garment editing are missing in current datasets. We address these gaps by developing an automated data construction pipeline that integrates various resources. + +Garment Generation, Simulation, and Rendering. Following GarmentCodeData [25], we sample 20,000 garments by randomizing values in sewing pattern configurations. To improve real-world garment relevance, we adjust the sampling ratios and assign lower weights for asymmetric gar + +![](images/e021b681664b01253c5a846583e40a5ee372ebc51e9f5e8b6b7f910e3f2f5468.jpg) +Figure 4. Data Construction Pipeline. We generate garments from JSON configurations, simulate them with ContourCraft [15] and render with Blender. + +ments and those with intricate cuffs. To create outfits, we pair top and bottom garments together or use single-piece items like dresses or jumpsuits. The garment set is draped on a SMPL-X [40] neutral model. + +We automate garment simulation using ContourCraft [15], chosen for its stability and scalability, with SMPL-X motion sequences sampled from the BEDLAM dataset [4]. Specific garment vertices are attached to corresponding body vertices using predefined rules to prevent garments from slipping off. For example, the upper edges of pants and skirts are connected to the nearest body vertices, and similar attachments are made for upper-body garment vertices near the shoulders. We filter out simulations with significant interpenetration or improper positioning in two steps. First, GarmentCode performs a sewing-pattern validation check, ensuring the integrity of panel edges and stitching. Second, we verify simulation results by assessing inter-penetration, and identify garments at risk of slipping off by inspecting the proximity of each garment vertex to the nearest body part. Each garment is then rendered in Blender from four random perspectives at elevation angles between 0 and 30 degrees, with garment and body textures sourced from the BEDLAM dataset. These steps are shown in Fig. 4. + +GPT-4o Labeling. The previous method [16] prompts GPT-4o to generate the general geometric descriptions for garments, such as the garment types, lengths, widths and sleeve lengths. However, these general descriptions are insufficient for garment detail understanding: the finetuned VLM struggles to interpret detailed features of garment components, such as pant cuffs, or execute specific garment edits. To address this, we construct a new dataset with low-level descriptions for individual garment parts. For image-based reconstruction, we instruct GPT-4o to generate descriptions for all visible garment parts in the image. For garment editing, we sample 5,000 garment parts (shirt main body, collar, sleeves, sleeve cuffs, waistband, pant legs, pant cuffs, and skirts), and use GPT-4o to generate descriptive text for each part. Assembling these components yields 20,000 new garments with detailed annotations. + +For garment editing, we randomly select two garments that differ in certain parts, with the editing prompt being: "Change the garment sewing pattern by modifying PART-1 to DESC-1, PART-2 to DESC-2, while keeping other parts unchanged." PART-1, PART-2, represent part names that differ between the two garments, and DESC-1, DESC-2 are target garment part descriptions. + +We finally build a synthetic dataset of 40K garments and 1 million images with text annotations, supporting garment reconstruction and editing from multimodal inputs. Plus, we use GPT-4o to label part-level details of 38K SHHQ [12] images to make the VLM better understand in-the-wild garment images. Compared to previous garment sewing pattern datasets [33], our dataset offers: 1) Diversity: GarmentCode generates a wider range of real-world garment types; 2) Detailed Descriptions: Garments are labeled with detailed geometric features; 3) Precise Control: Allows for precise, low-level garment manipulation. + +# 3.3. Training + +We use LLaVA [32] as our base VLM and finetune it with LoRA [18], with its parameters denoted as $\phi_{lora}$ . Additionally, we optimize the LLM token embedding $e_{token}$ , the LLM head layer $\mathrm{lm\_head}_{\Theta'}$ , and the projection layer $g_{\Theta}$ . + +Loss Function. As described in Sec. 3.1, the numeric embedding is extracted from the token [ENDS] as $H_{t}^{\mathbf{n}}$ . An MLP projection layer maps the embedding $H_{t}^{\mathbf{n}}$ to 76 float values $\mathbf{n}$ in the JSON configuration. Given the ground truth text output $\hat{Y}_{t}$ and sewing pattern values $\hat{\mathbf{n}}$ , we optimize with: + +$$ +\mathcal {L} = \operatorname {C E} \left(\hat {Y} _ {t}, Y _ {t}\right) + \lambda_ {\mathbf {n}} | (\hat {\mathbf {n}} - \mathbf {n}) \cdot \mathbf {m} |, \tag {1} +$$ + +where the first term is a cross-entropy loss, and the second term is the L1 difference for numeric values. $\mathbf{m}$ is a mask that filters out irrelevant numeric values, and $\lambda_{\mathbf{n}} = 0.1$ is the weight for the L1 loss. See more training details in Sup. Mat. + +# 3.4. Multi-turn Dialogue with Multimodal Inputs. + +After training, our model can handle multimodal input reconstruction, garment generation, and editing, as reflected in the training data. Additionally, it exhibits zero-shot reasoning for garment sewing patterns within multi-turn dialogues, despite being trained only on single-turn dialogue datasets. This enables users to create a garment from an image or text description and then refine it within the same interaction. A typical workflow might start with the VLM estimating a sewing pattern from an image. After reviewing the result, the user can request specific adjustments, providing a flexible, user-centered design process (see Fig. 5 for an example). + +# 4. Experiments + +We conduct experiments on image-based reconstruction (Sec 4.1), garment editing (Sec 4.2) and text-based garment generation (Sec 4.3), along with ablation studies (Sec 4.4). + +# 4.1. Image-based Reconstruction + +**Benchmarks.** We conduct image-based reconstruction experiments on the CloSe [2] and Dress4D [54] datasets. CloSe is a large-scale scan collection including 3D clothing segmentations across 18 clothing categories. Following CloSe, we use 145 scans with accurate SMPL-X fits as our validation set. As CloSe offers limited instances of loose garments, such as skirts and dresses, we further add samples from Dress4D, which has real-world outfits recorded using a multi-view volumetric capture system. We include 4 loose-fitting outfits and 36 images rendered from them. + +Implementation Details. Our approach applies the Chain-of-Thought (CoT) [55] method. We first prompt ChatGarment to generate a text description of the outfit in the image. The generated text descriptions are combined with the input image to estimate the final garment JSON configuration. More details are provided in Sup. Mat. + +Baselines. We compare our method with two previous approaches, SewFormer [33] and GarmentRecovery [29]. We use the publicly available code and checkpoints for these methods in our evaluation. Additionally, we introduce several new baseline models for comparison: + +- DressCode. We employ DressCode [16] to generate garments from image descriptions. Specifically, we utilize GPT-4o to generate garment descriptions from input images. These descriptions are processed by DressCode to create sewing patterns. +- GPT-4o. We design prompt templates and query GPT-4o using in-context learning (ICL) [10]. We use Garment Generator [22] as the sewing pattern configuration tool, as it has significantly fewer possible configurations and allows for more efficient template design and significantly fewer conversation rounds. We first prompt GPT-4o to choose the most suitable template from a predefined list, then query the parameter value by displaying three garment renderings with uniformly sampled parameter values. +- LLaVA*. In this setup, we use the LLaVA model to directly output the numerical values for the sewing pattern instead of using the sewing pattern projection layer. The remaining components are identical to our method. +- LLaVA-T. We first employ GPT-4o to generate detailed textual descriptions of the garment, and ChatGarment is prompted to reconstruct the sewing pattern based solely on the text description. +- LLaVA-I. We input the image into the model and prompt it to directly output the sewing pattern without referring to text descriptions. + +![](images/3c49984c3ca2d60b788f7fb25bc7e3b38052313ac3b89bf9d4fb0cfb9b53f922.jpg) +(a) Image-based Sewing Pattern Estimation + +![](images/20c9eb0f2636559fd7855dffabee2eb9e7a8a17de88d2059668276385a21414e.jpg) +(c) Text-based Sewing Pattern Generation and Editing +Figure 5. ChatGarment's dialog modes. Images and texts are adaptively combined to guide garment generation and editing. Text output between STARTS and ENDS contains information about the JSON configuration, which is then converted to 3D garments (JSON2Garment) for visualization and simulation purposes. + +![](images/813cca7457ecf6aaa37f69bd0b31616fa7f4a747ff650888e2fa027b876fce51.jpg) +(b) Image-guided Sewing Pattern Generation + +
MethodsTarget-PoseA-PoseFailure Rate (↓)
CD (↓)F-Score (↑)CD (↓)F-Score (↑)
Sewformer [33]27.060.5520.660.584.3%
DressCode [16]20.160.6012.120.6111.4%
GarmentRecovery [29]9.690.555.750.64-
GPT-4o11.040.6210.440.650
LLaVA*5.290.726.290.760
LLaVA-T6.170.727.250.740
LLaVA-I4.320.754.930.780
Ours3.120.753.060.780
+ +Table 1. Quantitative comparisons of Image-based garment reconstruction on Dress4D. ChatGarment achieves the best performance on all metrics. + +
MethodsCLOSE
CD (↓)F-Score (↑)Failure Rate (↓)
SewFormer [33]9.700.7088.33 %
DressCode [16]15.770.6167.0%
GarmentRecovery [29]2.390.785-
GPT-4o3.520.7550
LLaVA*3.080.7750
LLaVA-T3.760.7790
LLaVA-I4.550.7610
Ours2.940.7900
+ +Table 2. Quantitative comparisons of Image-based reconstruction on CloSe. The garments are deformed to the target pose. + +Since methods use different default human poses, we apply Linear Blend Skinning [1] to re-pose all garments to the ground-truth body shape and pose for fair comparison. + +Results. We employ the mean chamfer distance, the mean F-Score at $\tau = 0.001$ and the stitching failure rate as quantitative metrics. As Tab. 1 and Tab. 2 show, ChatGarment significantly outperforms existing methods on loose garments + +in Dress4D. On the CloSE dataset, which mostly features tight garments, GarmentRecovery [29] does well by leveraging garment segmentation and normal maps as guidance in an optimization-based framework. However, these cues are insufficient for accurately reconstructing loose garments. Our method has similar performance in this case, but is more general. Also, our algorithm robustly reconstructs garments with no stitching failures. In contrast, DressCode and SewFormer often generate invalid garment panels leading to failed stitching. The primary difference lies in the output format: DressCode and SewFormer predict the panel edges and stitches of the sewing pattern, while our method and the GPT-4o-based approach estimate the JSON configuration. Although direct edge and stitch prediction offers greater flexibility in theory, our experiments show that it does not enhance reconstruction accuracy. Instead, it increases the likelihood of generating invalid sewing patterns when handling out-of-distribution (OOD) data. + +Our model demonstrates superior performance compared to LLaVA*, highlighting the effectiveness of the proposed projection layer. Furthermore, our CoT-based two-step reconstruction algorithm outperforms both image-based (LLaVA-I) and text-based (LLaVA-T) reconstruction baselines. The text descriptions detail garment types and styles, offering valuable information for reconstruction, while images reveal intricate details that are hard to describe in words. Thus, combining both text and image inputs is the most effective strategy for image-based garment reconstruction. + +Figure 6 shows our model accurately estimates garment type, style, and details like skirt hems and collar styles. In contrast, other methods often misinterpret or overlook these + +![](images/bf6460101eca7784f1e6d4f87d92d62a334ea8f67984485cc13b574466e9eca0.jpg) +Figure 6. Image-based Garment Reconstruction. Unlike SewFormer [33], DressCode [16], GarmentRecovery [29], and the GPT-4o-based method, which often make mistakes or ignores garments. ChatGarment faithfully captures the shape, style and composition of the garments. + +details. SewFormer occasionally fails to reconstruct garments due to detection or stitching failure, and GarmentRecovery lacks support for whole-body dress reconstruction. Additionally, inaccuracies in 2D garment segmentation [28] can lead to errors in GarmentRecovery reconstruction. + +Garment designers often create image sketches before making sewing patterns. Surprisingly, despite not being trained on garment sketch images, ChatGarment demonstrates strong generalization, accurately predicting garment types and styles from sketches (Fig. 8). + +# 4.2. Garment Editing + +**Benchmarks.** To evaluate garment editing performance, we construct an extra evaluation dataset using methods outlined in Sec. 3.2. We manually select garment pairs with clear textual descriptions and substantial visual differences. This dataset consists of 135 garment pairs, each accompanied by corresponding images and text descriptions. The evaluation of garment editing involves three components: the source garment, the target garment, and the editing prompts. The + +goal is to derive the sewing pattern for the target garment using the source garment and the editing prompts. + +Implementation Details. For garment editing, we concatenate the source garment JSON file and the editing prompt together and pass it into ChatGarment. + +Baselines. We modify DressCode [16] and SewFormer [33] for comparison with ChatGarment: + +- DressCode with edited text. We generate a text description of the edited garment and use DressCode to create the corresponding sewing patterns. The description is produced by GPT-4o, which takes as inputs: 1) garment editing instructions and 2) front and back views of the source garment. The generated description follows DressCode's original input format. +- SewFormer with edited images. We obtain the front view of the edited garment with TurboEdit [9] to which we pass 1) the front view of the source garment, 2) the text description of the source garment, and 3) the editing instructions, and then use SewFormer to turn the image of the edited garment into the corresponding sewing pattern. + +![](images/9ff1a21d768bdfbea49cc33ffc7fdd7b8cadd8b9e2fd791d5626cb2c4771067e.jpg) +Prompts: Change the shirt body to: [long vertical sections, 'evenly shaped sides', 'mildly loose fit around upper body, 'straight front opening'] + +![](images/19c5bf0ef2f9f16d057286836b1fe4d4c45216d93d4d45573ddab11973464445.jpg) + +![](images/c9084a328857aca1e3dcb54751bc2af2211e17043d397584fe818f0a22acb1c2.jpg) + +![](images/c565fac7f7a928ebea2c9840217b0f8bcc7804fc257560615d6b0d09d8dc2ae7.jpg) + +![](images/37afdfb42b8ab401af45ab2fa4688c7756bec37b332d490e7bf5ea88d47c195b.jpg) +Prompts: Change the sleeves to [short sleeves length], 'loose fit around the arms', 'wide sleeve openings relative to the arm size'] + +![](images/180bd5d407ab0829487cc6b440899d3c2a2a4a8d2d07b9f1dd9e3cd77c93a4ca.jpg) + +![](images/d8521218d572a4209f659e187388cdb9f7eae3f8a9ccc89cef820ed2b91c2ea8.jpg) + +![](images/533a58d36cf25e8ea9f26083089489d56b94c13b65609c010a598106f74149cf.jpg) + +![](images/4c2ed7e6fef802bedfa647d2577483b9f594f881657e242f633f0581f272feb5.jpg) +Prompts: Change the collar to: 'rounded neckline', 'slightly loose fit', 'even curvature throughout'] +Source + +![](images/0d2f8dd010f2162a2ffc2f052f46ad145b3266f43a0897489b3b78d4827e2936.jpg) +SewFormer + +![](images/d7e07ac8de50fa2558d416c6654e2e43ad0b05adb548ff3088ad974182bbcbc5.jpg) +DressCode + +![](images/b74c049ce986a552e2356e8fa67f3c3f38af1ccc7f17641a1c9bbb9e139c3fcd.jpg) +Ours + +![](images/a1d4c71b3bac124afae014ab4b230de9e45737b948fd2a68f9b0ce130b2c703b.jpg) +Prompts: Change the sleeves to: 'long, full-length sleeves', 'loose fit', 'ends at wrist level', 'wide openings at the wrist'] + +![](images/3e06c9b751311b6cfdc479aa01ff0e4c1619d868d25622d7f1c1f24bb76c54fd.jpg) + +![](images/abea4b22b6cd2727c52dbc058a507361893f96d8c512a90b6eea5f74c270671e.jpg) + +![](images/6d75eb984adc16070c6856410fb3db55e1610999957b8982fbc87709ca0ea804.jpg) + +![](images/33fe012a9b1b041659498d6d9aa8210fc18240de2ebeff8bfd25eef6a56d961f.jpg) +Prompts: Change the pant legs to: [wide-leg silhouette], 'ankle-length hem', 'flowy fit', 'loose and relaxed cut'] + +![](images/0f5432f029a6a1ab872524fd68192027da4db7b0dae5db4931aa79ac0d7dc467.jpg) + +![](images/f06ef61ee1af0a17e79c22b8a48ce9496d8f597492c9fa7517e773f451809aa3.jpg) + +![](images/d9151dea9357da3aeb5c95355a408ea45ae190a09206af7f9aa60d3f59a593f9.jpg) + +![](images/0e616a57aaa7916db9ec0c2e07bdd4e2f2f99c31a0bbdea544148f6c75c239cc.jpg) +Prompts: Change the shirt body to: ['straight vertical silhouette', 'moderate fit around the torso', 'hemline falls above the knees'] +Source + +![](images/439c4f6a13ddb4a2efb711b79c3a4d59966c3bb8879c8b5b3a28c58ab6270b49.jpg) +SewFormer + +![](images/882692b5467db619b543d82c1b7c0a3c0d4f6f2fe71441478178609fe77792c4.jpg) +DressCode + +![](images/df09e02a735e5c81883df1e6f035b71d84af936fc82b003c61576e6009e9d92a.jpg) +Ours + +![](images/7b3dd8724c3b750637254fc4a1515a12c4f805f3b7365816b61ba294d5846d4e.jpg) +Figure 7. Results on Garment Editing. We present six examples where the models (SewFormer [33], DressCode [16], and ChatGarment) edit the source garment shown in the source image according to the given editing instructions in the prompt boxes. +Input + +![](images/62a5b265abfc2ad7acde6368371ddcdeb3708f65303741b7d64f963f940b1feb.jpg) +2D Sewing Pattern + +![](images/b1872f8f762845b0bb375ccc5c754dee17e38382397433a491a15656ac4fa1fa.jpg) +3D Garment +Figure 8. Reconstruction results from garment sketches. The generated garments accurately capture the shape and style of the input sketch, despite our model being trained without sketches. + +Results. We compare our method with SewFormer and Dresscode on Chamfer Distance and F-score in Tab. 3, where our approach demonstrates superior performance. The qualitative results in Fig. 7 show that our method accurately modifies the targeted garment parts according to the input prompt. In contrast, SewFormer and DressCode often fail to precisely follow the prompt instructions and frequently alter other unintended areas of the garment. + +# 4.3. Text-based Generation + +Following DressCode [16], we use GPT-4o to generate 150 highly diverse text labels from 85 selected in-the-wild images featuring various tops, skirts, pants, and dresses. Next, we render the generated 3D garments with default textures and compute the CLIP score using the provided text labels. See quantitative results in Tab. 4. Our model achieves a higher CLIP score compared to DressCode [16], highlighting the superior prompt-following capabilities of our approach. The qualitative results are shown in Sup. Mat. + +
MethodsCD (↓)F-Score (↑)
SewFormer [33]28.770.548
DressCode [16]10.970.750
Ours2.510.893
+ +Table 3. Quantitative comparisons for garment editing. ChatGarment outperforms other two methods by a large margin. + +
MethodsCLIP score (↑)
DressCode [16]22.8
Ours23.7
+ +Table 4. Quantitative comparisons for text-based garment generation. ChatGarment achieves a higher CLIP score than DressCode, indicating better alignment with the text inputs. + +# 4.4. Ablation + +We evaluate the impact of different aspects of ChatGarment, including multi-modal LLM backbones, and various training datasets. Please refer to Sup. Mat. + +# 5. Conclusion + +We introduce ChatGarment, a VLM-based model that estimates garments from images and generates or edits garments based on text descriptions. Leveraging the capabilities of LLMs, ChatGarment supports multi-turn dialogue interactions for 3D garment design. It surpasses existing sewing pattern-specific and LLM-based methods in accurately estimating sewing patterns from a single image. While ChatGarment excels in many areas, it may struggle with precise garment editing. For instance, altering the skirt length might unintentionally affect other parts of the garment. This issue could be addressed with more detailed text instructions or a differential CSG optimizer [60] for finer adjustments. Nonetheless, ChatGarment offers the first VLM-based framework for unified garment estimation, generation, and editing. Future directions could include expanding into garment manufacturing design or improving clothing simulation with more fine-grained material attributes. + +# Acknowledgments and Disclosure + +We thank Maria Korosteleva for help on GarmentCode, and Giorgio Becherini for advice on data acquisition and processing. Yuliang Xiu is funded by Max Planck Institute for Intelligent Systems, the Research Center for Industries of the Future (RCIF) at Westlake University, and the Westlake Education Foundation. + +While MJB is a co-founder and Chief Scientist at Meshcapade, his research in this project was performed solely at, and funded solely by, the Max Planck Society. + +# References + +[1] Rinet Abdrashitov, Kim Raichstat, Jared Monsen, and David Hill. Robust skin weights transfer via weight inpainting. In SIGGRAPH Asia 2023 Technical Communications, pages 1-4. 2023. 6 +[2] Dimitrije Antic, Garvita Tiwari, Batuhan Ozcomlekci, Ricardo Marin, and Gerard Pons-Moll. CloSe: A 3D clothing segmentation dataset and model. In International Conference on 3D Vision (3DV), 2024. 2, 5 +[3] Bharat Lal Bhatnagar, Garvita Tiwari, Christian Theobalt, and Gerard Pons-Moll. Multi-garment net: Learning to dress 3d people from images. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5420-5430, 2019. 2 +[4] Michael J Black, Priyanka Patel, Joachim Tesch, and Jinlong Yang. Bedlam: A synthetic dataset of bodies exhibiting detailed lifelike animated motion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8726-8737, 2023. 4 +[5] Andrés Casado-Elvira, Marc Comino Trinidad, and Dan Casas. Pergamo: Personalized 3d garments from monocular video. In Computer Graphics Forum, pages 293–304. Wiley Online Library, 2022. 2 +[6] Enric Corona, Albert Pumarola, Guillem Alenya, Gerard Pons-Moll, and Francesc Moreno-Noguer. Simplicit: Topology-aware generative model for clothed people. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11875-11885, 2021. 2, 3 +[7] R Daněřek, Endri Dibra, Cengiz Öztireli, Remo Ziegler, and Markus Gross. Deepgarment: 3d garment shape estimation from a single image. In Computer Graphics Forum, pages 269-280. Wiley Online Library, 2017. 2 +[8] Luca De Luigi, Ren Li, Benoit Guillard, Mathieu Salzmann, and Pascal Fua. Drapenet: Garment generation and self-supervised draping. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1451-1460, 2023. 2, 3 +[9] Gilad Deutch, Rinon Gal, Daniel Garibi, Or Patashnik, and Daniel Cohen-Or. Turboedit: Text-based image editing using few-step diffusion models. arXiv preprint arXiv:2408.00735, 2024. 7 +[10] Qingxiu Dong, Lei Li, Damai Dai, Ce Zheng, Jingyuan Ma, Rui Li, Heming Xia, Jingjing Xu, Zhiyong Wu, Tianyu + +Liu, et al. A survey on in-context learning. arXiv preprint arXiv:2301.00234, 2022. 5 +[11] Yao Feng, Jing Lin, Sai Kumar Dwivedi, Yu Sun, Priyanka Patel, and Michael J. Black. ChatPose: Chatting about 3d human pose. In CVPR, 2024. 2, 3, 4 +[12] Jianglin Fu, Shikai Li, Yuming Jiang, Kwan-Yee Lin, Chen Qian, Chen-Change Loy, Wayne Wu, and Ziwei Liu. Stylegan-human: A data-centric odyssey of human generation. arXiv preprint, arXiv:2204.11823, 2022. 5 +[13] Rao Fu, Xiao Zhan, Yiwen Chen, Daniel Ritchie, and Srinath Sridhar. Shapecraft: A recursive text-conditioned 3d shape generation model. Advances in Neural Information Processing Systems, 35:8882-8895, 2022. 3 +[14] Siavash Golkar, Mariel Pettee, Michael Eickenberg, Alberto Bietti, Miles Cranmer, Geraud Krawezik, Francois Lanusse, Michael McCabe, Ruben Ohana, Liam Parker, et al. xVal: A continuous number encoding for large language models. arXiv preprint arXiv:2310.02989, 2023. 2, 4 +[15] Artur Grigorev, Giorgio Becherini, Michael Black, Otmar Hilliges, and Bernhard Thomaszewski. Contourcraft: Learning to resolve intersections in neural multi-garment simulations. In ACM SIGGRAPH 2024 Conference Papers, pages 1-10, 2024. 4 +[16] Kai He, Kaixin Yao, Qixuan Zhang, Jingyi Yu, Lingjie Liu, and Lan Xu. Dresscode: Autoregressively sewing and generating garments from text guidance. ACM Transactions on Graphics (TOG), 43(4):1-13, 2024. 2, 3, 4, 5, 6, 7, 8 +[17] Yining Hong, Haoyu Zhen, Peihao Chen, Shuhong Zheng, Yilun Du, Zhenfang Chen, and Chuang Gan. 3d-llm: Injecting the 3d world into large language models. Advances in Neural Information Processing Systems, 36:20482-20494, 2023. 3 +[18] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In ICLR, 2022. 2, 3, 5 +[19] Ziniu Hu, Ahmet Iscen, Aashi Jain, Thomas Kipf, Yisong Yue, David A Ross, Cordelia Schmid, and Alireza Fathi. Scenecraft: An IIm agent for synthesizing 3d scenes as blender code. In *Forty-first International Conference on Machine Learning*, 2024. 3 +[20] Yangyi Huang, Hongwei Yi, Yuliang Xiu, Tingting Liao, Ji-axiang Tang, Deng Cai, and Justus Thies. TeCH: Text-guided Reconstruction of Lifelike Clothed Humans. In International Conference on 3D Vision (3DV), 2024. 3 +[21] Boyi Jiang, Juyong Zhang, Yang Hong, Jinhao Luo, Ligang Liu, and Hujun Bao. Bcnet: Learning body and cloth shape from a single image. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XX 16, pages 18-35. Springer, 2020. 2 +[22] Maria Korosteleva and Sung-Hee Lee. Generating datasets of 3d garments with sewing patterns. arXiv preprint arXiv:2109.05633, 2021. 5 +[23] Maria Korosteleva and Sung-Hee Lee. Neuraltailer: Reconstructing sewing pattern structures from 3d point clouds of garments. ACM Transactions on Graphics (TOG), 41(4):1-16, 2022. 3 +[24] Maria Korosteleva and Olga Sorkine-Hornung. GarmentCode: Programming parametric sewing patterns. ACM Transaction + +on Graphics, 42(6), 2023. SIGGRAPH ASIA 2023 issue. 2, 3 +[25] Maria Korosteleva, Timur Levent Kesdogan, Fabian Kemper, Stephan Wenninger, Jasmin Koller, Yuhan Zhang, Mario Botsch, and Olga Sorkine-Hornung. GarmentCodeData: A dataset of 3D made-to-measure garments with sewing patterns. In Computer Vision – ECCV 2024, 2024. 2, 3, 4 +[26] Peter Kulits, Haiwen Feng, Weiyang Liu, Victoria Abrevaya, and Michael J Black. Re-thinking inverse graphics with large language models. Transactions on Machine Learning Research, 2024. 2, 4 +[27] Boqian Li, Xuan Li, Ying Jiang, Tianyi Xie, Feng Gao, Huamin Wang, Yin Yang, and Chenfanfu Jiang. Garmentdreamer: 3dgs guided garment synthesis with diverse geometry and texture details. In 3DV, 2025. 2 +[28] Peike Li, Yunqiu Xu, Yunchao Wei, and Yi Yang. Self-correction for human parsing. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(6):3260-3271, 2020. 7 +[29] Ren Li, Corentin Dumery, Benoit Guillard, and Pascal Fua. Garment recovery with shape and deformation priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1586-1595, 2024. 3, 5, 6, 7 +[30] Ren Li, Benoit Guillard, and Pascal Fua. Isp: Multi-layered garment draping with implicit sewing patterns. Advances in Neural Information Processing Systems, 36, 2024. 2, 3 +[31] Dingning Liu, Xiaoshui Huang, Yuenan Hou, Zhihui Wang, Zhenfei Yin, Yongshun Gong, Peng Gao, and Wanli Ouyang. Uni3d-llm: Unifying point cloud perception, generation and editing with large language models. arXiv preprint arXiv:2402.03327, 2024. 3 +[32] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning, 2023. 3, 5 +[33] Lijuan Liu, Xiangyu Xu, Zhijie Lin, Jiabin Liang, and Shuicheng Yan. Towards garment sewing pattern reconstruction from a single image. ACM Transactions on Graphics (TOG), 42(6):1-15, 2023. 2, 3, 4, 5, 6, 7, 8 +[34] Weiyang Liu, Zeju Qiu, Yao Feng, Yuliang Xiu, Yuxuan Xue, Longhui Yu, Haiwen Feng, Zhen Liu, Juyeon Heo, Songyou Peng, Yandong Wen, Michael J. Black, Adrian Weller, and Bernhard Scholkopf. Parameter-efficient orthogonal finetuning via butterfly factorization. In ICLR, 2024. 3 +[35] Zhen Liu, Yao Feng, Yuliang Xiu, Weiyang Liu, Liam Paull, Michael J. Black, and Bernhard Scholkopf. Ghost on the shell: An expressive representation of general 3d shapes. In The Twelfth International Conference on Learning Representations, 2024. 3 +[36] Zhongjin Luo, Haolin Liu, Chenghong Li, Wanghao Du, Zirong Jin, Yinyu Nie, Weikai Chen, and Xiaoguang Han. Garverselod: High-fidelity 3d garment reconstruction from a single in-the-wild image using a dataset with levels of details. In ACM Transactions on Graphics (TOG), 2024. 2 +[37] Qianli Ma, Jinlong Yang, Siyu Tang, and Michael J. Black. The power of points for modeling humans in clothing. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 10974-10984, 2021. 3 +[38] Qianli Ma, Jinlong Yang, Michael J. Black, and Siyu Tang. Neural point-based shape modeling of humans in challenging + +clothing. In International Conference on 3D Vision (3DV), pages 679-689, 2022. 3 +[39] Miles Macklin. Warp: A high-performance python framework forgpu simulation and graphics. https://github.com/nvidia/warp, 2022. NVIDIA GPU Technology Conference (GTC). 3 +[40] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed A. A. Osman, Dimitrios Tzionas, and Michael J. Black. Expressive body capture: 3D hands, face, and body from a single image. In CVPR, 2019. 4 +[41] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. In ICLR, 2023. 2 +[42] Lingteng Qiu, Guanying Chen, Jiapeng Zhou, Mutian Xu, Junle Wang, and Xiaoguang Han. Rec-mv: Reconstructing 3d dynamic cloth from monocular videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4637-4646, 2023. 2 +[43] Zeju Qiu, Weiyang Liu, Haiwen Feng, Yuxuan Xue, Yao Feng, Zhen Liu, Dan Zhang, Adrian Weller, and Bernhard Schölkopf. Controlling text-to-image diffusion by orthogonal finetuning. In NeurIPS, 2023. 3 +[44] Zeju Qiu, Weiyang Liu, Haiwen Feng, Zhen Liu, Tim Z Xiao, Katherine M Collins, Joshua B Tenenbaum, Adrian Weller, Michael J Black, and Bernhard Scholkopf. Can large language models understand symbolic graphics programs? arXiv preprint arXiv:2408.08313, 2024. 3 +[45] Shunsuke Saito, Zeng Huang, Ryota Natsume, Shigeo Morishima, Hao Li, and Angjoo Kanazawa. PIFu: Pixel-aligned implicit function for high-resolution clothed human digitization. In ICCV, pages 2304-2314, 2019. 3 +[46] Shunsuke Saito, Tomas Simon, Jason Saragih, and Hanbyul Joo. PIFuHD: Multi-Level Pixel-Aligned Implicit Function for High-Resolution 3D Human Digitization. In CVPR, pages 81–90, 2020. 3 +[47] Nikolaos Sarafianos, Tuur Stuyck, Xiaoyu Xiang, Yilei Li, Jovan Popovic, and Rakesh Ranjan. Garment3dgen: 3d garment stylization and texture generation. In 3DV, 2025. 2 +[48] Tianchang Shen, Jun Gao, Kangxue Yin, Ming-Yu Liu, and Sanja Fidler. Deep marching tetrahedra: a hybrid representation for high-resolution 3d shape synthesis. In Advances in Neural Information Processing Systems (NeurIPS), 2021. 3 +[49] Yu Shen, Junbang Liang, and Ming C. Lin. Gan-based garment generation using sewing pattern images. In ECCV, 2020. 2 +[50] Astitva Srivastava, Pranav Manu, Amit Raj, Varun Jampani, and Avinash Sharma. Wordrobe: Text-guided generation of textured 3d garments. In ECCV, pages 458-475. Springer, 2025. 3 +[51] Zhaoqi Su, Weilin Wan, Tao Yu, Lingjie Liu, Lu Fang, Wenping Wang, and Yebin Liu. Mulaycap: Multi-layer human performance capture using a monocular video camera. IEEE Transactions on Visualization and Computer Graphics, 28(4): 1862-1879, 2020. 3 +[52] Chunyi Sun, Junlin Han, Weijian Deng, Xinlong Wang, Zishan Qin, and Stephen Gould. 3d-gpt: Procedural 3d modeling with large language models. arXiv preprint arXiv:2310.12945, 2023. 3 + +[53] Tuanfeng Y. Wang, Duygu Ceylan, Jovan Popovic, and Niloy J. Mitra. Learning a shared shape space for multimodal garment design. ACM Trans. Graph., 37(6), 2018. 3 +[54] Wenbo Wang, Hsuan-I Ho, Chen Guo, Boxiang Rong, Artur Grigorev, Jie Song, Juan Jose Zarate, and Otmar Hilliges. 4D-DRESS: A 4D dataset of real-world human clothing with semantic annotations. In IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2024. 2, 5 +[55] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824-24837, 2022. 5 +[56] Yuliang Xiu, Jinlong Yang, Dimitrios Tzionas, and Michael J. Black. ICON: Implicit Clothed humans Obtained from Normals. In CVPR, 2022. 3 +[57] Yuliang Xiu, Jinlong Yang, Xu Cao, Dimitrios Tzionas, and Michael J. Black. ECON: Explicit Clothed humans Optimized via Normal integration. In CVPR, 2023. 3 +[58] Runsen Xu, Xiaolong Wang, Tai Wang, Yilun Chen, Jiangmiao Pang, and Dahua Lin. Pointllm: Empowering large language models to understand point clouds. In European Conference on Computer Vision, pages 131-147. Springer, 2025. 3 +[59] Zhengming Yu, Zhiyang Dou, Xiaoxiao Long, Cheng Lin, Zekun Li, Yuan Liu, Norman Müller, Taku Komura, Marc Habermann, Christian Theobalt, et al. Surf-d: Generating high-quality surfaces of arbitrary topologies using diffusion models. In ECCV, pages 419-438. Springer, 2025. 3 +[60] Haocheng Yuan, Adrien Bousseau, Hao Pan, Chengquan Zhang, Niloy J Mitra, and Changjian Li. Diffcsg: Differentiable csg via rasterization. arXiv preprint arXiv:2409.01421, 2024.8 +[61] Longwen Zhang, Ziyu Wang, Qixuan Zhang, Qiwei Qiu, Anqi Pang, Haoran Jiang, Wei Yang, Lan Xu, and Jingyi Yu. Clay: A controllable large-scale generative model for creating high-quality 3d assets. ACM Transactions on Graphics (TOG), 43 (4):1-20, 2024. 3 +[62] Heming Zhu, Lingteng Qiu, Yuda Qiu, and Xiaoguang Han. Registering explicit to implicit: Towards high-fidelity garment mesh reconstruction from single images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3845-3854, 2022. 2 \ No newline at end of file diff --git a/2025/ChatGarment_ Garment Estimation, Generation and Editing via Large Language Models/images.zip b/2025/ChatGarment_ Garment Estimation, Generation and Editing via Large Language Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..d795ce4f44b923be36bdec4704f1833716228541 --- /dev/null +++ b/2025/ChatGarment_ Garment Estimation, Generation and Editing via Large Language Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:59cc35edf4c8390d10f7e753362e868569fbe8f6f542b697ca792bc0df2064eb +size 502532 diff --git a/2025/ChatGarment_ Garment Estimation, Generation and Editing via Large Language Models/layout.json b/2025/ChatGarment_ Garment Estimation, Generation and Editing via Large Language Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..8c9f796c56e4ff768e05e9a80402e0a393e4dbab --- /dev/null +++ b/2025/ChatGarment_ Garment Estimation, Generation and Editing via Large Language Models/layout.json @@ -0,0 +1,9382 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 117, + 103, + 495, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 117, + 103, + 495, + 140 + ], + "spans": [ + { + "bbox": [ + 117, + 103, + 495, + 140 + ], + "type": "text", + "content": "ChatGarment: Garment Estimation, Generation and Editing via Large Language Models" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 95, + 159, + 515, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 95, + 159, + 515, + 193 + ], + "spans": [ + { + "bbox": [ + 95, + 159, + 515, + 193 + ], + "type": "text", + "content": "Siyuan Bian" + }, + { + "bbox": [ + 95, + 159, + 515, + 193 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 95, + 159, + 515, + 193 + ], + "type": "text", + "content": " Chenghao Xu" + }, + { + "bbox": [ + 95, + 159, + 515, + 193 + ], + "type": "inline_equation", + "content": "^{1,3}" + }, + { + "bbox": [ + 95, + 159, + 515, + 193 + ], + "type": "text", + "content": " Yuliang Xiu" + }, + { + "bbox": [ + 95, + 159, + 515, + 193 + ], + "type": "inline_equation", + "content": "^{1,4}" + }, + { + "bbox": [ + 95, + 159, + 515, + 193 + ], + "type": "text", + "content": " Artur Grigorev" + }, + { + "bbox": [ + 95, + 159, + 515, + 193 + ], + "type": "inline_equation", + "content": "^{1,5}" + }, + { + "bbox": [ + 95, + 159, + 515, + 193 + ], + "type": "text", + "content": " Zhen Liu" + }, + { + "bbox": [ + 95, + 159, + 515, + 193 + ], + "type": "inline_equation", + "content": "^{1,6}" + }, + { + "bbox": [ + 95, + 159, + 515, + 193 + ], + "type": "text", + "content": " Cewu Lu" + }, + { + "bbox": [ + 95, + 159, + 515, + 193 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 95, + 159, + 515, + 193 + ], + "type": "text", + "content": " Michael J. Black" + }, + { + "bbox": [ + 95, + 159, + 515, + 193 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 95, + 159, + 515, + 193 + ], + "type": "text", + "content": " Yao Feng" + }, + { + "bbox": [ + 95, + 159, + 515, + 193 + ], + "type": "inline_equation", + "content": "^{7,8,\\dagger}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 111, + 195, + 499, + 211 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 195, + 499, + 211 + ], + "spans": [ + { + "bbox": [ + 111, + 195, + 499, + 211 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 111, + 195, + 499, + 211 + ], + "type": "text", + "content": "Max Planck Institute for Intelligent Systems " + }, + { + "bbox": [ + 111, + 195, + 499, + 211 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 111, + 195, + 499, + 211 + ], + "type": "text", + "content": "Shanghai Jiao Tong University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 111, + 211, + 499, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 211, + 499, + 239 + ], + "spans": [ + { + "bbox": [ + 111, + 211, + 499, + 239 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 111, + 211, + 499, + 239 + ], + "type": "text", + "content": "EPFL " + }, + { + "bbox": [ + 111, + 211, + 499, + 239 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 111, + 211, + 499, + 239 + ], + "type": "text", + "content": "Westlake University " + }, + { + "bbox": [ + 111, + 211, + 499, + 239 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 111, + 211, + 499, + 239 + ], + "type": "text", + "content": "ETH Zürich " + }, + { + "bbox": [ + 111, + 211, + 499, + 239 + ], + "type": "inline_equation", + "content": "^{6}" + }, + { + "bbox": [ + 111, + 211, + 499, + 239 + ], + "type": "text", + "content": "Mila, University of Montreal " + }, + { + "bbox": [ + 111, + 211, + 499, + 239 + ], + "type": "inline_equation", + "content": "^{7}" + }, + { + "bbox": [ + 111, + 211, + 499, + 239 + ], + "type": "text", + "content": "Meshcapade " + }, + { + "bbox": [ + 111, + 211, + 499, + 239 + ], + "type": "inline_equation", + "content": "^{8}" + }, + { + "bbox": [ + 111, + 211, + 499, + 239 + ], + "type": "text", + "content": "Stanford University" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 246, + 242, + 361, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 246, + 242, + 361, + 252 + ], + "spans": [ + { + "bbox": [ + 246, + 242, + 361, + 252 + ], + "type": "text", + "content": "chatgarment.github.io" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 55, + 276, + 312, + 407 + ], + "blocks": [ + { + "bbox": [ + 55, + 276, + 312, + 407 + ], + "lines": [ + { + "bbox": [ + 55, + 276, + 312, + 407 + ], + "spans": [ + { + "bbox": [ + 55, + 276, + 312, + 407 + ], + "type": "image", + "image_path": "abd2410ab08bf1b0d7bd21125ec50641268779933d62491b99f6ccfad435656e.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 415, + 555, + 436 + ], + "lines": [ + { + "bbox": [ + 55, + 415, + 555, + 436 + ], + "spans": [ + { + "bbox": [ + 55, + 415, + 555, + 436 + ], + "type": "text", + "content": "Figure 1. As a multimodal 3D garment creator, ChatGarment understands both images and language. It can estimate complex 3D garments represented as sewing patterns from a single image, which can be easily animated and simulated. It also supports garment editing based on text instructions." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 318, + 276, + 555, + 407 + ], + "blocks": [ + { + "bbox": [ + 318, + 276, + 555, + 407 + ], + "lines": [ + { + "bbox": [ + 318, + 276, + 555, + 407 + ], + "spans": [ + { + "bbox": [ + 318, + 276, + 555, + 407 + ], + "type": "image", + "image_path": "e47b86281e490117fe4b90dff48e9929045d0f0b0a9a4e34ea853b57c70dc216.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 151, + 445, + 200, + 457 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 445, + 200, + 457 + ], + "spans": [ + { + "bbox": [ + 151, + 445, + 200, + 457 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 472, + 297, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 472, + 297, + 688 + ], + "spans": [ + { + "bbox": [ + 53, + 472, + 297, + 688 + ], + "type": "text", + "content": "We introduce ChatGarment, a novel approach that leverages large vision-language models (VLMs) to automate the estimation, generation, and editing of 3D garments from images or text descriptions. Unlike previous methods that struggle in real-world scenarios or lack interactive editing capabilities, ChatGarment can estimate sewing patterns from in-the-wild images or sketches, generate them from text descriptions, and edit garments based on user instructions, all within an interactive dialogue. These sewing patterns can then be draped on a 3D body and animated. This is achieved by finetuning a VLM to directly generate a JSON file that includes both textual descriptions of garment types and styles, as well as continuous numerical attributes. This JSON file is then used to create sewing patterns through a programming parametric model. To support this, we refine the existing programming model, GarmentCode, by expanding its garment type coverage and simplifying its structure for efficient VLM fine-tuning. Additionally, we construct a" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 446, + 556, + 530 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 446, + 556, + 530 + ], + "spans": [ + { + "bbox": [ + 313, + 446, + 556, + 530 + ], + "type": "text", + "content": "large-scale dataset of image-to-sewing-pattern and text-to-sewing-pattern pairs through an automated data pipeline. Extensive evaluations demonstrate ChatGarment's ability to accurately reconstruct, generate, and edit garments from multimodal inputs, highlighting its potential to simplify workflows in fashion and gaming applications. Code and data are available at https://chatgarment.github.io/." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 538, + 394, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 538, + 394, + 550 + ], + "spans": [ + { + "bbox": [ + 314, + 538, + 394, + 550 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 312, + 557, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 557, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 312, + 557, + 556, + 715 + ], + "type": "text", + "content": "Imagine seeing a photo of someone in a nice looking outfit that you would like for yourself or your 3D avatar. Our goal is to take that photo and convert it into a 2D sewing pattern and then enable the user to use language to edit the pattern, for example, making the sleeves longer or editing the neckline. We then drape the garment on a 3D body and animate it. This is illustrated in Fig. 1. With this approach, given just an image and/or text, we construct sewing patterns for 3D garments that can be immediately used in fashion or gaming applications. This lightweight process is in stark contrast to how sewing patterns and 3D garments are traditionally created by an artist or clothing designer. In contrast to the traditional labor-intensive process, our approach exploits" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 701, + 235, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 701, + 235, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 701, + 235, + 713 + ], + "type": "text", + "content": "This work was done while YF was at Meshcapade." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2924" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 97 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 97 + ], + "type": "text", + "content": "large vision-language models (VLMs) for the first time to democratize the capture and design of clothing." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 99, + 297, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 99, + 297, + 410 + ], + "spans": [ + { + "bbox": [ + 56, + 99, + 297, + 410 + ], + "type": "text", + "content": "Recent work explores the generation or estimation of 3D clothing using text and/or images as input [6, 8, 21, 42]. These methods use various 3D representations including 3D meshes, point clouds, or implicit representations like Unsigned Distance Functions (UDFs) and Signed Distance Functions (SDFs). We observe, however, that most clothing is designed and manufactured using 2D patterns, and posit that this is a more natural representation. Sewing patterns can be seamlessly integrated into existing garment design pipelines for animation and manufacturing. Recent approaches [16, 30, 33] use sewing patterns for garment estimation and generation, but mapping images or text to sewing patterns remains challenging due to limited training data, leading to poor generalization across diverse clothing types. These methods also does not support interactive editing, making post-processing labor-intensive for 3D artists. Fortunately, recent work, such as GarmentCode [24, 25], provides a programmatic framework for encoding garment information into a JSON file. This JSON file includes textual descriptions of garment types, styles and numerical attributes, offering a rich representation of sewing patterns. The JSON file is then processed by the GarmentCode framework to produce 2D sewing patterns, which can subsequently be draped on a 3D body. However, to date, there is no automated method to directly extract a GarmentCode representation (JSON file) from an image or text description." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 412, + 297, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 412, + 297, + 544 + ], + "spans": [ + { + "bbox": [ + 55, + 412, + 297, + 544 + ], + "type": "text", + "content": "Our observation is that a JSON file resembles natural language, making it suitable for interpretation by large language models (LLMs). At the same time, Vision-LLMs (VLMs) excel at understanding images and garment concepts, such as types of garments and specific features like long sleeves. If we enable the VLM to translate its garment understanding into the structured format of GarmentCode, then it becomes capable of generating sewing patterns. This allows the VLM to take an image as input and output a garment, while also supporting garment generation and editing through text descriptions or interactive dialogue." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 545, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 297, + 714 + ], + "type": "text", + "content": "Specifically, we introduce ChatGarment, a method that finetunes a VLM to take text queries, optionally with images, as input and outputs a JSON garment file. This JSON file is then used for GarmentCode generation. The JSON file includes textual descriptions of garment types (e.g., skirts, pants), styles (e.g., collar type), and continuous numerical attributes (e.g., pant length). Inspired by previous work [11, 14, 26], we introduce a new token () to represent these numerical values and train an MLP projection layer to decode them from the token embedding. Those numbers, combined with the VLM's language output, are formatted into the final JSON file. In our approach, we keep the VLM's vision encoder fixed, fine-tune the language model using LoRA [18], and jointly train the MLP projection layer. To" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 555, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 203 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 203 + ], + "type": "text", + "content": "adapt GarmentCode [24] for more efficient VLM training, we create an improved version that supports a wider range of clothing types (e.g. open-jacket) and removes unnecessary settings to reduce token usage in training. We also develop an automated data construction pipeline that leverages existing tools for garment generation, draping, simulation, rendering, and automatic labeling with GPT-4o. This pipeline enables building a dataset with image-to-GarmentCode and text-to-GarmentCode pairs, including 20K new garments and 1M images with detailed descriptions, supporting both garment creation and editing." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 204, + 556, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 204, + 556, + 335 + ], + "spans": [ + { + "bbox": [ + 313, + 204, + 556, + 335 + ], + "type": "text", + "content": "We evaluate ChatGarment across a diverse set of tasks, including those specifically trained for, and novel applications that demonstrate the generalization capabilities of VLMs. For single-image garment reconstruction, ChatGarment outperforms prior sewing-pattern-specific and LLM-based models on the Dress4D [54] and CloSE [2] evaluation datasets. Additionally, we demonstrate garment editing, generation, and multi-turn dialogue, highlighting our approach's versatility. ChatGarment can flexibly use text and images to create and edit 3D garments that can be readily animated, introducing a new workflow for garment design." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 328, + 336, + 504, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 336, + 504, + 346 + ], + "spans": [ + { + "bbox": [ + 328, + 336, + 504, + 346 + ], + "type": "text", + "content": "We summarize our contributions as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 348, + 555, + 455 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 313, + 348, + 553, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 348, + 553, + 371 + ], + "spans": [ + { + "bbox": [ + 313, + 348, + 553, + 371 + ], + "type": "text", + "content": "- A unified model capable of estimating, generating and editing sewing patterns from multimodal inputs." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 372, + 555, + 394 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 372, + 555, + 394 + ], + "spans": [ + { + "bbox": [ + 313, + 372, + 555, + 394 + ], + "type": "text", + "content": "- The first approach to leverage VLMs for directly generating JSON files for garment creation." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 396, + 555, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 396, + 555, + 419 + ], + "spans": [ + { + "bbox": [ + 313, + 396, + 555, + 419 + ], + "type": "text", + "content": "- A refined version of GarmentCode that supports more diverse garment types and is optimized for VLM training." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 419, + 555, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 419, + 555, + 455 + ], + "spans": [ + { + "bbox": [ + 313, + 419, + 555, + 455 + ], + "type": "text", + "content": "- An automatic data construction pipeline for generating image-to-sewing-pattern and text-to-sewing-pattern data, along with the release of a large-scale dataset." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 466, + 400, + 479 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 466, + 400, + 479 + ], + "spans": [ + { + "bbox": [ + 314, + 466, + 400, + 479 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 486, + 555, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 486, + 555, + 557 + ], + "spans": [ + { + "bbox": [ + 313, + 486, + 555, + 557 + ], + "type": "text", + "content": "Garment Creation. Most 3D garments have predefined templates (e.g., sewing patterns) and categories (e.g., skirts, shirts, pants, dresses, etc.), but their non-rigid nature makes them highly deformable with dynamic motions. Efforts to create 3D garments can be grouped into \"template-based\" or \"free-form\" methods, based on their 3D representations." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 558, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 556, + 714 + ], + "type": "text", + "content": "A line of template-based approaches register predefined garment templates to 2D observations, which could be real captures, e.g. from monocular video [5, 21, 36, 42] and images [3, 6, 7, 62], generated data [49], or derived from the \"proxy gradient\" computed from 2D diffusion models [27, 47] via score distillation sampling (SDS) [41]. When the template is classified correctly, the predefined garment template, refined through a coarse-to-fine step, ensures topological stability and correctness. Unfortunately, template-based methods struggle to accurately model garments that deviate from the predefined templates, particularly those with diverse styles, such as long skirts, or garments that undergo topological changes, like opening a jacket." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2925" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 193 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 193 + ], + "type": "text", + "content": "To address these limitations, free-form representations like (un)signed distance functions [6, 8, 50, 59], occupancy fields [45, 46, 56, 57], tetrahedrons [20, 35, 48], and point clouds [37, 38] are leveraged. However, this flexibility can compromise the inherent structure of garments, potentially leading to non-garment shapes and floating artifacts. Without shape regularization, incomplete observations lead to the reconstructed unseen parts being overly smooth. Even worse, rigging or simulating these free-form garments becomes extremely challenging." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 205, + 298, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 205, + 298, + 469 + ], + "spans": [ + { + "bbox": [ + 56, + 205, + 298, + 469 + ], + "type": "text", + "content": "The sewing pattern is a well-balanced solution – with pre-defined yet flexible templates. Several methods exploit sewing patterns to reconstruct 3D garments. MulayCap [51] estimates the parameters of a sewing pattern from a monocular video. ISP [30] parameterizes sewing patterns as implicit functions in a 2D space. Using ISP as a garment representation, GarmentRecovery [29] trains a feedforward network to infer sewing patterns from in-the-wild photos. Neural-Tailor [23] uses an RNN to produce more complex sewing patterns from point clouds. Wang et al. [53] learn a shared garment shape space that can be inferred from multimodal input such as sketches, SewFormer [33] uses a transformer to infer multi-panel patterns from monocular images, and DressCode [16] estimates quantized sewing patterns with a GPT-based model from text inputs. Despite promising results, these approaches are limited to basic sewing patterns, featuring only front and back panels, or constrained by the limited sewing patterns seen in the training data. Additionally, none of the above works have explored using programming parametric sewing patterns, such as GarmentCode [24, 25] to create more complex garments, either in generation or reconstruction." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 498, + 298, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 298, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 298, + 714 + ], + "type": "text", + "content": "VLMs for 3D tasks. With their broad understanding of the world, LLMs and VLMs are widely used to reason about 3D. There has been significant work on generating, editing, and analyzing 3D objects and scenes [13, 17, 19, 31, 44, 52, 58, 61]. While these methods show the power of VLMs to reason about objects, they do not address our problem of representing 2D patterns that correspond to 3D shapes – this is a special property of garments. Some attempts have been made to connect sewing patterns with language, such as DressCode [16]. This motivates us to present ChatGarment, which connects a large language model to GarmentCode, sharing a similar approach to recent strategies [11, 32] by finetuning LLMs [18, 34, 43]. Powered by VLMs, ChatGarment significantly improves generalization and robustness in reconstructing out-of-domain images and increases diversity in 3D clothing generation. It also supports sewing pattern editing and multi-turn dialogue, greatly simplifying the 3D garment workflow." + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 317, + 70, + 554, + 148 + ], + "blocks": [ + { + "bbox": [ + 317, + 70, + 554, + 148 + ], + "lines": [ + { + "bbox": [ + 317, + 70, + 554, + 148 + ], + "spans": [ + { + "bbox": [ + 317, + 70, + 554, + 148 + ], + "type": "image", + "image_path": "6fac4910610603e5339c6775b96168d2d4399033ac95c6da4987b2d00d66b316.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 155, + 555, + 211 + ], + "lines": [ + { + "bbox": [ + 313, + 155, + 555, + 211 + ], + "spans": [ + { + "bbox": [ + 313, + 155, + 555, + 211 + ], + "type": "text", + "content": "Figure 2. Pipeline of ChatGarment. ChatGarment takes text or an image as input and generates a JSON file. The JSON file is decoded into 2D sewing patterns using GarmentCode [24] and then draped onto the human body. The final 3D garments are compatible with simulation software (e.g., MAYA, Blender, Style3D, etc.)." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 314, + 218, + 370, + 230 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 218, + 370, + 230 + ], + "spans": [ + { + "bbox": [ + 314, + 218, + 370, + 230 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 237, + 375, + 251 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 237, + 375, + 251 + ], + "spans": [ + { + "bbox": [ + 314, + 237, + 375, + 251 + ], + "type": "text", + "content": "3.1. Pipeline" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 255, + 556, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 255, + 556, + 388 + ], + "spans": [ + { + "bbox": [ + 313, + 255, + 556, + 388 + ], + "type": "text", + "content": "The goal of ChatGarment is to simplify the workflow of garment creation by automatically estimating, generating, and editing garments from multimodal image and text input. To achieve this, we finetune LLaVA [32], a multimodal large language model " + }, + { + "bbox": [ + 313, + 255, + 556, + 388 + ], + "type": "inline_equation", + "content": "f_{\\phi}" + }, + { + "bbox": [ + 313, + 255, + 556, + 388 + ], + "type": "text", + "content": ". It takes an input image " + }, + { + "bbox": [ + 313, + 255, + 556, + 388 + ], + "type": "inline_equation", + "content": "X_{v}" + }, + { + "bbox": [ + 313, + 255, + 556, + 388 + ], + "type": "text", + "content": " and text prompts " + }, + { + "bbox": [ + 313, + 255, + 556, + 388 + ], + "type": "inline_equation", + "content": "X_{q}" + }, + { + "bbox": [ + 313, + 255, + 556, + 388 + ], + "type": "text", + "content": ", and produces a textual response " + }, + { + "bbox": [ + 313, + 255, + 556, + 388 + ], + "type": "inline_equation", + "content": "Y_{t}" + }, + { + "bbox": [ + 313, + 255, + 556, + 388 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 313, + 255, + 556, + 388 + ], + "type": "inline_equation", + "content": "Y_{t} = f_{\\phi}(X_{q},X_{v})" + }, + { + "bbox": [ + 313, + 255, + 556, + 388 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 313, + 255, + 556, + 388 + ], + "type": "inline_equation", + "content": "Y_{t} = f_{\\phi}(X_{q})" + }, + { + "bbox": [ + 313, + 255, + 556, + 388 + ], + "type": "text", + "content": " in the absence of an image. From this output, a structured JSON file " + }, + { + "bbox": [ + 313, + 255, + 556, + 388 + ], + "type": "inline_equation", + "content": "J_{t}" + }, + { + "bbox": [ + 313, + 255, + 556, + 388 + ], + "type": "text", + "content": " is extracted as a subset of " + }, + { + "bbox": [ + 313, + 255, + 556, + 388 + ], + "type": "inline_equation", + "content": "Y_{t}" + }, + { + "bbox": [ + 313, + 255, + 556, + 388 + ], + "type": "text", + "content": ", which will be decoded into sewing patterns using the GarmentCode [24] decoder " + }, + { + "bbox": [ + 313, + 255, + 556, + 388 + ], + "type": "inline_equation", + "content": "D_{g}" + }, + { + "bbox": [ + 313, + 255, + 556, + 388 + ], + "type": "text", + "content": ". This pipeline is illustrated in Fig. 2." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 390, + 556, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 390, + 556, + 557 + ], + "spans": [ + { + "bbox": [ + 313, + 390, + 556, + 557 + ], + "type": "text", + "content": "LLM-friendly Rich GarmentCode. We use GarmentCode to convert the JSON file into a 3D garment mesh in the canonical pose. GarmentCode [24] is a programming parametric model that uses a domain-specific language (DSL) for creating 2D sewing patterns, with a hierarchical JSON structure to define garment components and stitching rules. These patterns can be draped onto body models via a warp-based pattern stitching algorithm [24, 39], and further be simulated easily. GarmentCode can model diverse garments with intricate geometric details, including different cuts, frills, and pleats. To better align GarmentCode with our model's language-driven requirements, we develop a refined version, GarmentCodeRC (Richer & Cleaner) with two major improvements (see Fig. 3)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 558, + 557, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 557, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 557, + 714 + ], + "type": "text", + "content": "First, we modify stitching and panel positioning to better support diverse garment types, such as open-front jackets, high-waist skirts, and fitted pant legs. This allows us to model a broader range of real-world garments and aligns more closely with natural garment descriptions. Second, we simplify GarmentCode's JSON configuration to better suit LLM training. The original configuration is redundant, applying the same settings across all garment types. We optimize this by automatically removing irrelevant settings (e.g., omitting skirt-related parameters for upper-body garments). This adjustment reduces the average language token count from 900 to 350, decreasing ambiguity in LLM training and improving efficiency. Furthermore, to stabilize model train" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2926" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 70, + 295, + 209 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 295, + 209 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 295, + 209 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 295, + 209 + ], + "type": "image", + "image_path": "2120e0a1dba49351e967bf370b866a4c51c61e51ba6e2dd0686596fdd41cdc6a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 213, + 295, + 247 + ], + "lines": [ + { + "bbox": [ + 55, + 213, + 295, + 247 + ], + "spans": [ + { + "bbox": [ + 55, + 213, + 295, + 247 + ], + "type": "text", + "content": "Figure 3. GarmentCodeRC. Left: new options to model open-front jackets, high-waist skirts, and tight pant legs. Right: simplified JSON configuration for more efficient LLM training." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 261, + 296, + 309 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 261, + 296, + 309 + ], + "spans": [ + { + "bbox": [ + 55, + 261, + 296, + 309 + ], + "type": "text", + "content": "ing, we normalize all floating-point values to a " + }, + { + "bbox": [ + 55, + 261, + 296, + 309 + ], + "type": "inline_equation", + "content": "[0,1]" + }, + { + "bbox": [ + 55, + 261, + 296, + 309 + ], + "type": "text", + "content": " range. Examples of the GarmentCodeRC JSON configuration are given in Sup. Mat. Throughout this paper, \"GarmentCode\" refers to our enriched and simplified version." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 312, + 296, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 312, + 296, + 529 + ], + "spans": [ + { + "bbox": [ + 55, + 312, + 296, + 529 + ], + "type": "text", + "content": "LLM for Textual and Numeric Outputs. The output JSON file " + }, + { + "bbox": [ + 55, + 312, + 296, + 529 + ], + "type": "inline_equation", + "content": "J_{t}" + }, + { + "bbox": [ + 55, + 312, + 296, + 529 + ], + "type": "text", + "content": " contains both textual and numeric descriptions. Previous research [11, 14, 26] has shown that LLMs struggle in predicting continuous and precise numerical values. To overcome this, following ChatPose [11, 26], we encode these continuous values into language token embeddings and train a projection layer to recover accurate numerical values, such as garment length. In this setup, the language embedding for " + }, + { + "bbox": [ + 55, + 312, + 296, + 529 + ], + "type": "inline_equation", + "content": "Y_{t}" + }, + { + "bbox": [ + 55, + 312, + 296, + 529 + ], + "type": "text", + "content": " is represented as " + }, + { + "bbox": [ + 55, + 312, + 296, + 529 + ], + "type": "inline_equation", + "content": "H_{t}" + }, + { + "bbox": [ + 55, + 312, + 296, + 529 + ], + "type": "text", + "content": ". From the textual output " + }, + { + "bbox": [ + 55, + 312, + 296, + 529 + ], + "type": "inline_equation", + "content": "Y_{t}" + }, + { + "bbox": [ + 55, + 312, + 296, + 529 + ], + "type": "text", + "content": ", upon encountering the start token , we extract the textual content of the JSON file, which spans from to . The language embedding at the end token , where numeric information is encoded, is represented as " + }, + { + "bbox": [ + 55, + 312, + 296, + 529 + ], + "type": "inline_equation", + "content": "H_{t}^{\\mathrm{n}}" + }, + { + "bbox": [ + 55, + 312, + 296, + 529 + ], + "type": "text", + "content": ". A projection layer uses this embedding to predict the numerical values as " + }, + { + "bbox": [ + 55, + 312, + 296, + 529 + ], + "type": "inline_equation", + "content": "\\mathbf{n} = g_{\\Theta}(H_{t}^{\\mathrm{n}})" + }, + { + "bbox": [ + 55, + 312, + 296, + 529 + ], + "type": "text", + "content": ". Finally, we merge these numerical values with the textual descriptions extracted earlier and format them into our final sewing pattern JSON configuration file " + }, + { + "bbox": [ + 55, + 312, + 296, + 529 + ], + "type": "inline_equation", + "content": "J_{t}" + }, + { + "bbox": [ + 55, + 312, + 296, + 529 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 538, + 217, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 538, + 217, + 550 + ], + "spans": [ + { + "bbox": [ + 55, + 538, + 217, + 550 + ], + "type": "text", + "content": "3.2. Automatic Data Construction" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 556, + 296, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 556, + 296, + 652 + ], + "spans": [ + { + "bbox": [ + 55, + 556, + 296, + 652 + ], + "type": "text", + "content": "To train ChatGarment, we need image-to-JSON and text-to-JSON paired data. While some work like SewFormer [33] offers image-to-JSON pairs, they fall short in garment variety. For example, they lack tight-sleeve garments and only support a limited variety of skirts. Additionally, text descriptions for garment editing are missing in current datasets. We address these gaps by developing an automated data construction pipeline that integrates various resources." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "content": "Garment Generation, Simulation, and Rendering. Following GarmentCodeData [25], we sample 20,000 garments by randomizing values in sewing pattern configurations. To improve real-world garment relevance, we adjust the sampling ratios and assign lower weights for asymmetric gar" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 317, + 72, + 553, + 180 + ], + "blocks": [ + { + "bbox": [ + 317, + 72, + 553, + 180 + ], + "lines": [ + { + "bbox": [ + 317, + 72, + 553, + 180 + ], + "spans": [ + { + "bbox": [ + 317, + 72, + 553, + 180 + ], + "type": "image", + "image_path": "e021b681664b01253c5a846583e40a5ee372ebc51e9f5e8b6b7f910e3f2f5468.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 187, + 555, + 220 + ], + "lines": [ + { + "bbox": [ + 313, + 187, + 555, + 220 + ], + "spans": [ + { + "bbox": [ + 313, + 187, + 555, + 220 + ], + "type": "text", + "content": "Figure 4. Data Construction Pipeline. We generate garments from JSON configurations, simulate them with ContourCraft [15] and render with Blender." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 238, + 553, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 238, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 313, + 238, + 553, + 285 + ], + "type": "text", + "content": "ments and those with intricate cuffs. To create outfits, we pair top and bottom garments together or use single-piece items like dresses or jumpsuits. The garment set is draped on a SMPL-X [40] neutral model." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 288, + 556, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 288, + 556, + 515 + ], + "spans": [ + { + "bbox": [ + 313, + 288, + 556, + 515 + ], + "type": "text", + "content": "We automate garment simulation using ContourCraft [15], chosen for its stability and scalability, with SMPL-X motion sequences sampled from the BEDLAM dataset [4]. Specific garment vertices are attached to corresponding body vertices using predefined rules to prevent garments from slipping off. For example, the upper edges of pants and skirts are connected to the nearest body vertices, and similar attachments are made for upper-body garment vertices near the shoulders. We filter out simulations with significant interpenetration or improper positioning in two steps. First, GarmentCode performs a sewing-pattern validation check, ensuring the integrity of panel edges and stitching. Second, we verify simulation results by assessing inter-penetration, and identify garments at risk of slipping off by inspecting the proximity of each garment vertex to the nearest body part. Each garment is then rendered in Blender from four random perspectives at elevation angles between 0 and 30 degrees, with garment and body textures sourced from the BEDLAM dataset. These steps are shown in Fig. 4." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 521, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 556, + 713 + ], + "type": "text", + "content": "GPT-4o Labeling. The previous method [16] prompts GPT-4o to generate the general geometric descriptions for garments, such as the garment types, lengths, widths and sleeve lengths. However, these general descriptions are insufficient for garment detail understanding: the finetuned VLM struggles to interpret detailed features of garment components, such as pant cuffs, or execute specific garment edits. To address this, we construct a new dataset with low-level descriptions for individual garment parts. For image-based reconstruction, we instruct GPT-4o to generate descriptions for all visible garment parts in the image. For garment editing, we sample 5,000 garment parts (shirt main body, collar, sleeves, sleeve cuffs, waistband, pant legs, pant cuffs, and skirts), and use GPT-4o to generate descriptive text for each part. Assembling these components yields 20,000 new garments with detailed annotations." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2927" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 298, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 298, + 180 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 298, + 180 + ], + "type": "text", + "content": "For garment editing, we randomly select two garments that differ in certain parts, with the editing prompt being: \"Change the garment sewing pattern by modifying PART-1 to DESC-1, PART-2 to DESC-2, while keeping other parts unchanged.\" PART-1, PART-2, represent part names that differ between the two garments, and DESC-1, DESC-2 are target garment part descriptions." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 182, + 298, + 315 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 182, + 298, + 315 + ], + "spans": [ + { + "bbox": [ + 55, + 182, + 298, + 315 + ], + "type": "text", + "content": "We finally build a synthetic dataset of 40K garments and 1 million images with text annotations, supporting garment reconstruction and editing from multimodal inputs. Plus, we use GPT-4o to label part-level details of 38K SHHQ [12] images to make the VLM better understand in-the-wild garment images. Compared to previous garment sewing pattern datasets [33], our dataset offers: 1) Diversity: GarmentCode generates a wider range of real-world garment types; 2) Detailed Descriptions: Garments are labeled with detailed geometric features; 3) Precise Control: Allows for precise, low-level garment manipulation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 326, + 119, + 339 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 326, + 119, + 339 + ], + "spans": [ + { + "bbox": [ + 55, + 326, + 119, + 339 + ], + "type": "text", + "content": "3.3. Training" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 346, + 297, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 346, + 297, + 396 + ], + "spans": [ + { + "bbox": [ + 55, + 346, + 297, + 396 + ], + "type": "text", + "content": "We use LLaVA [32] as our base VLM and finetune it with LoRA [18], with its parameters denoted as " + }, + { + "bbox": [ + 55, + 346, + 297, + 396 + ], + "type": "inline_equation", + "content": "\\phi_{lora}" + }, + { + "bbox": [ + 55, + 346, + 297, + 396 + ], + "type": "text", + "content": ". Additionally, we optimize the LLM token embedding " + }, + { + "bbox": [ + 55, + 346, + 297, + 396 + ], + "type": "inline_equation", + "content": "e_{token}" + }, + { + "bbox": [ + 55, + 346, + 297, + 396 + ], + "type": "text", + "content": ", the LLM head layer " + }, + { + "bbox": [ + 55, + 346, + 297, + 396 + ], + "type": "inline_equation", + "content": "\\mathrm{lm\\_head}_{\\Theta'}" + }, + { + "bbox": [ + 55, + 346, + 297, + 396 + ], + "type": "text", + "content": ", and the projection layer " + }, + { + "bbox": [ + 55, + 346, + 297, + 396 + ], + "type": "inline_equation", + "content": "g_{\\Theta}" + }, + { + "bbox": [ + 55, + 346, + 297, + 396 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 401, + 296, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 401, + 296, + 472 + ], + "spans": [ + { + "bbox": [ + 55, + 401, + 296, + 472 + ], + "type": "text", + "content": "Loss Function. As described in Sec. 3.1, the numeric embedding is extracted from the token [ENDS] as " + }, + { + "bbox": [ + 55, + 401, + 296, + 472 + ], + "type": "inline_equation", + "content": "H_{t}^{\\mathbf{n}}" + }, + { + "bbox": [ + 55, + 401, + 296, + 472 + ], + "type": "text", + "content": ". An MLP projection layer maps the embedding " + }, + { + "bbox": [ + 55, + 401, + 296, + 472 + ], + "type": "inline_equation", + "content": "H_{t}^{\\mathbf{n}}" + }, + { + "bbox": [ + 55, + 401, + 296, + 472 + ], + "type": "text", + "content": " to 76 float values " + }, + { + "bbox": [ + 55, + 401, + 296, + 472 + ], + "type": "inline_equation", + "content": "\\mathbf{n}" + }, + { + "bbox": [ + 55, + 401, + 296, + 472 + ], + "type": "text", + "content": " in the JSON configuration. Given the ground truth text output " + }, + { + "bbox": [ + 55, + 401, + 296, + 472 + ], + "type": "inline_equation", + "content": "\\hat{Y}_{t}" + }, + { + "bbox": [ + 55, + 401, + 296, + 472 + ], + "type": "text", + "content": " and sewing pattern values " + }, + { + "bbox": [ + 55, + 401, + 296, + 472 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{n}}" + }, + { + "bbox": [ + 55, + 401, + 296, + 472 + ], + "type": "text", + "content": ", we optimize with:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 99, + 476, + 296, + 491 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 476, + 296, + 491 + ], + "spans": [ + { + "bbox": [ + 99, + 476, + 296, + 491 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\operatorname {C E} \\left(\\hat {Y} _ {t}, Y _ {t}\\right) + \\lambda_ {\\mathbf {n}} | (\\hat {\\mathbf {n}} - \\mathbf {n}) \\cdot \\mathbf {m} |, \\tag {1}", + "image_path": "20a5be87cd0f37d571a1273bb63a307916de7d09a2398713aeb83f1b161a39be.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 501, + 298, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 501, + 298, + 550 + ], + "spans": [ + { + "bbox": [ + 55, + 501, + 298, + 550 + ], + "type": "text", + "content": "where the first term is a cross-entropy loss, and the second term is the L1 difference for numeric values. " + }, + { + "bbox": [ + 55, + 501, + 298, + 550 + ], + "type": "inline_equation", + "content": "\\mathbf{m}" + }, + { + "bbox": [ + 55, + 501, + 298, + 550 + ], + "type": "text", + "content": " is a mask that filters out irrelevant numeric values, and " + }, + { + "bbox": [ + 55, + 501, + 298, + 550 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathbf{n}} = 0.1" + }, + { + "bbox": [ + 55, + 501, + 298, + 550 + ], + "type": "text", + "content": " is the weight for the L1 loss. See more training details in Sup. Mat." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 562, + 291, + 575 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 562, + 291, + 575 + ], + "spans": [ + { + "bbox": [ + 55, + 562, + 291, + 575 + ], + "type": "text", + "content": "3.4. Multi-turn Dialogue with Multimodal Inputs." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 582, + 298, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 298, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 298, + 715 + ], + "type": "text", + "content": "After training, our model can handle multimodal input reconstruction, garment generation, and editing, as reflected in the training data. Additionally, it exhibits zero-shot reasoning for garment sewing patterns within multi-turn dialogues, despite being trained only on single-turn dialogue datasets. This enables users to create a garment from an image or text description and then refine it within the same interaction. A typical workflow might start with the VLM estimating a sewing pattern from an image. After reviewing the result, the user can request specific adjustments, providing a flexible, user-centered design process (see Fig. 5 for an example)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 72, + 395, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 395, + 85 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 395, + 85 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 91, + 555, + 128 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 91, + 555, + 128 + ], + "spans": [ + { + "bbox": [ + 313, + 91, + 555, + 128 + ], + "type": "text", + "content": "We conduct experiments on image-based reconstruction (Sec 4.1), garment editing (Sec 4.2) and text-based garment generation (Sec 4.3), along with ablation studies (Sec 4.4)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 135, + 471, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 135, + 471, + 148 + ], + "spans": [ + { + "bbox": [ + 313, + 135, + 471, + 148 + ], + "type": "text", + "content": "4.1. Image-based Reconstruction" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 154, + 556, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 154, + 556, + 274 + ], + "spans": [ + { + "bbox": [ + 313, + 154, + 556, + 274 + ], + "type": "text", + "content": "**Benchmarks.** We conduct image-based reconstruction experiments on the CloSe [2] and Dress4D [54] datasets. CloSe is a large-scale scan collection including 3D clothing segmentations across 18 clothing categories. Following CloSe, we use 145 scans with accurate SMPL-X fits as our validation set. As CloSe offers limited instances of loose garments, such as skirts and dresses, we further add samples from Dress4D, which has real-world outfits recorded using a multi-view volumetric capture system. We include 4 loose-fitting outfits and 36 images rendered from them." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 277, + 557, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 277, + 557, + 350 + ], + "spans": [ + { + "bbox": [ + 313, + 277, + 557, + 350 + ], + "type": "text", + "content": "Implementation Details. Our approach applies the Chain-of-Thought (CoT) [55] method. We first prompt ChatGarment to generate a text description of the outfit in the image. The generated text descriptions are combined with the input image to estimate the final garment JSON configuration. More details are provided in Sup. Mat." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 352, + 556, + 412 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 352, + 556, + 412 + ], + "spans": [ + { + "bbox": [ + 313, + 352, + 556, + 412 + ], + "type": "text", + "content": "Baselines. We compare our method with two previous approaches, SewFormer [33] and GarmentRecovery [29]. We use the publicly available code and checkpoints for these methods in our evaluation. Additionally, we introduce several new baseline models for comparison:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 314, + 415, + 556, + 713 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 314, + 415, + 556, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 415, + 556, + 474 + ], + "spans": [ + { + "bbox": [ + 314, + 415, + 556, + 474 + ], + "type": "text", + "content": "- DressCode. We employ DressCode [16] to generate garments from image descriptions. Specifically, we utilize GPT-4o to generate garment descriptions from input images. These descriptions are processed by DressCode to create sewing patterns." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 315, + 474, + 556, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 474, + 556, + 582 + ], + "spans": [ + { + "bbox": [ + 315, + 474, + 556, + 582 + ], + "type": "text", + "content": "- GPT-4o. We design prompt templates and query GPT-4o using in-context learning (ICL) [10]. We use Garment Generator [22] as the sewing pattern configuration tool, as it has significantly fewer possible configurations and allows for more efficient template design and significantly fewer conversation rounds. We first prompt GPT-4o to choose the most suitable template from a predefined list, then query the parameter value by displaying three garment renderings with uniformly sampled parameter values." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 315, + 582, + 553, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 582, + 553, + 629 + ], + "spans": [ + { + "bbox": [ + 315, + 582, + 553, + 629 + ], + "type": "text", + "content": "- LLaVA*. In this setup, we use the LLaVA model to directly output the numerical values for the sewing pattern instead of using the sewing pattern projection layer. The remaining components are identical to our method." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 315, + 630, + 553, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 630, + 553, + 677 + ], + "spans": [ + { + "bbox": [ + 315, + 630, + 553, + 677 + ], + "type": "text", + "content": "- LLaVA-T. We first employ GPT-4o to generate detailed textual descriptions of the garment, and ChatGarment is prompted to reconstruct the sewing pattern based solely on the text description." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 315, + 677, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 677, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 315, + 677, + 553, + 713 + ], + "type": "text", + "content": "- LLaVA-I. We input the image into the model and prompt it to directly output the sewing pattern without referring to text descriptions." + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "2928" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 83, + 70, + 335, + 165 + ], + "blocks": [ + { + "bbox": [ + 83, + 70, + 335, + 165 + ], + "lines": [ + { + "bbox": [ + 83, + 70, + 335, + 165 + ], + "spans": [ + { + "bbox": [ + 83, + 70, + 335, + 165 + ], + "type": "image", + "image_path": "3c49984c3ca2d60b788f7fb25bc7e3b38052313ac3b89bf9d4fb0cfb9b53f922.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 136, + 170, + 281, + 180 + ], + "lines": [ + { + "bbox": [ + 136, + 170, + 281, + 180 + ], + "spans": [ + { + "bbox": [ + 136, + 170, + 281, + 180 + ], + "type": "text", + "content": "(a) Image-based Sewing Pattern Estimation" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 339, + 70, + 522, + 279 + ], + "blocks": [ + { + "bbox": [ + 339, + 70, + 522, + 279 + ], + "lines": [ + { + "bbox": [ + 339, + 70, + 522, + 279 + ], + "spans": [ + { + "bbox": [ + 339, + 70, + 522, + 279 + ], + "type": "image", + "image_path": "20c9eb0f2636559fd7855dffabee2eb9e7a8a17de88d2059668276385a21414e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 339, + 281, + 518, + 293 + ], + "lines": [ + { + "bbox": [ + 339, + 281, + 518, + 293 + ], + "spans": [ + { + "bbox": [ + 339, + 281, + 518, + 293 + ], + "type": "text", + "content": "(c) Text-based Sewing Pattern Generation and Editing" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 294, + 555, + 327 + ], + "lines": [ + { + "bbox": [ + 55, + 294, + 555, + 327 + ], + "spans": [ + { + "bbox": [ + 55, + 294, + 555, + 327 + ], + "type": "text", + "content": "Figure 5. ChatGarment's dialog modes. Images and texts are adaptively combined to guide garment generation and editing. Text output between STARTS and ENDS contains information about the JSON configuration, which is then converted to 3D garments (JSON2Garment) for visualization and simulation purposes." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 83, + 184, + 335, + 278 + ], + "blocks": [ + { + "bbox": [ + 83, + 184, + 335, + 278 + ], + "lines": [ + { + "bbox": [ + 83, + 184, + 335, + 278 + ], + "spans": [ + { + "bbox": [ + 83, + 184, + 335, + 278 + ], + "type": "image", + "image_path": "813cca7457ecf6aaa37f69bd0b31616fa7f4a747ff650888e2fa027b876fce51.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 133, + 281, + 284, + 293 + ], + "lines": [ + { + "bbox": [ + 133, + 281, + 284, + 293 + ], + "spans": [ + { + "bbox": [ + 133, + 281, + 284, + 293 + ], + "type": "text", + "content": "(b) Image-guided Sewing Pattern Generation" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 58, + 337, + 296, + 438 + ], + "blocks": [ + { + "bbox": [ + 58, + 337, + 296, + 438 + ], + "lines": [ + { + "bbox": [ + 58, + 337, + 296, + 438 + ], + "spans": [ + { + "bbox": [ + 58, + 337, + 296, + 438 + ], + "type": "table", + "html": "
MethodsTarget-PoseA-PoseFailure Rate (↓)
CD (↓)F-Score (↑)CD (↓)F-Score (↑)
Sewformer [33]27.060.5520.660.584.3%
DressCode [16]20.160.6012.120.6111.4%
GarmentRecovery [29]9.690.555.750.64-
GPT-4o11.040.6210.440.650
LLaVA*5.290.726.290.760
LLaVA-T6.170.727.250.740
LLaVA-I4.320.754.930.780
Ours3.120.753.060.780
", + "image_path": "5a34cda29fb563e6eea3d040dd77fabed5c8cce55029ad7606da47cf951a0321.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 75, + 474, + 277, + 588 + ], + "blocks": [ + { + "bbox": [ + 55, + 442, + 297, + 473 + ], + "lines": [ + { + "bbox": [ + 55, + 442, + 297, + 473 + ], + "spans": [ + { + "bbox": [ + 55, + 442, + 297, + 473 + ], + "type": "text", + "content": "Table 1. Quantitative comparisons of Image-based garment reconstruction on Dress4D. ChatGarment achieves the best performance on all metrics." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 75, + 474, + 277, + 588 + ], + "lines": [ + { + "bbox": [ + 75, + 474, + 277, + 588 + ], + "spans": [ + { + "bbox": [ + 75, + 474, + 277, + 588 + ], + "type": "table", + "html": "
MethodsCLOSE
CD (↓)F-Score (↑)Failure Rate (↓)
SewFormer [33]9.700.7088.33 %
DressCode [16]15.770.6167.0%
GarmentRecovery [29]2.390.785-
GPT-4o3.520.7550
LLaVA*3.080.7750
LLaVA-T3.760.7790
LLaVA-I4.550.7610
Ours2.940.7900
", + "image_path": "82f803c1bdb693d4b6139b9ccfb1c2e58716edfe0da2f878a02238369ed0facd.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 591, + 297, + 613 + ], + "lines": [ + { + "bbox": [ + 55, + 591, + 297, + 613 + ], + "spans": [ + { + "bbox": [ + 55, + 591, + 297, + 613 + ], + "type": "text", + "content": "Table 2. Quantitative comparisons of Image-based reconstruction on CloSe. The garments are deformed to the target pose." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 625, + 296, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 625, + 296, + 662 + ], + "spans": [ + { + "bbox": [ + 55, + 625, + 296, + 662 + ], + "type": "text", + "content": "Since methods use different default human poses, we apply Linear Blend Skinning [1] to re-pose all garments to the ground-truth body shape and pose for fair comparison." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 665, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 297, + 714 + ], + "type": "text", + "content": "Results. We employ the mean chamfer distance, the mean F-Score at " + }, + { + "bbox": [ + 55, + 665, + 297, + 714 + ], + "type": "inline_equation", + "content": "\\tau = 0.001" + }, + { + "bbox": [ + 55, + 665, + 297, + 714 + ], + "type": "text", + "content": " and the stitching failure rate as quantitative metrics. As Tab. 1 and Tab. 2 show, ChatGarment significantly outperforms existing methods on loose garments" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 340, + 555, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 340, + 555, + 555 + ], + "spans": [ + { + "bbox": [ + 313, + 340, + 555, + 555 + ], + "type": "text", + "content": "in Dress4D. On the CloSE dataset, which mostly features tight garments, GarmentRecovery [29] does well by leveraging garment segmentation and normal maps as guidance in an optimization-based framework. However, these cues are insufficient for accurately reconstructing loose garments. Our method has similar performance in this case, but is more general. Also, our algorithm robustly reconstructs garments with no stitching failures. In contrast, DressCode and SewFormer often generate invalid garment panels leading to failed stitching. The primary difference lies in the output format: DressCode and SewFormer predict the panel edges and stitches of the sewing pattern, while our method and the GPT-4o-based approach estimate the JSON configuration. Although direct edge and stitch prediction offers greater flexibility in theory, our experiments show that it does not enhance reconstruction accuracy. Instead, it increases the likelihood of generating invalid sewing patterns when handling out-of-distribution (OOD) data." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 556, + 556, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 556, + 556, + 676 + ], + "spans": [ + { + "bbox": [ + 313, + 556, + 556, + 676 + ], + "type": "text", + "content": "Our model demonstrates superior performance compared to LLaVA*, highlighting the effectiveness of the proposed projection layer. Furthermore, our CoT-based two-step reconstruction algorithm outperforms both image-based (LLaVA-I) and text-based (LLaVA-T) reconstruction baselines. The text descriptions detail garment types and styles, offering valuable information for reconstruction, while images reveal intricate details that are hard to describe in words. Thus, combining both text and image inputs is the most effective strategy for image-based garment reconstruction." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "content": "Figure 6 shows our model accurately estimates garment type, style, and details like skirt hems and collar styles. In contrast, other methods often misinterpret or overlook these" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2929" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 68, + 553, + 426 + ], + "blocks": [ + { + "bbox": [ + 56, + 68, + 553, + 426 + ], + "lines": [ + { + "bbox": [ + 56, + 68, + 553, + 426 + ], + "spans": [ + { + "bbox": [ + 56, + 68, + 553, + 426 + ], + "type": "image", + "image_path": "bf6460101eca7784f1e6d4f87d92d62a334ea8f67984485cc13b574466e9eca0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 433, + 555, + 456 + ], + "lines": [ + { + "bbox": [ + 55, + 433, + 555, + 456 + ], + "spans": [ + { + "bbox": [ + 55, + 433, + 555, + 456 + ], + "type": "text", + "content": "Figure 6. Image-based Garment Reconstruction. Unlike SewFormer [33], DressCode [16], GarmentRecovery [29], and the GPT-4o-based method, which often make mistakes or ignores garments. ChatGarment faithfully captures the shape, style and composition of the garments." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 465, + 297, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 465, + 297, + 525 + ], + "spans": [ + { + "bbox": [ + 55, + 465, + 297, + 525 + ], + "type": "text", + "content": "details. SewFormer occasionally fails to reconstruct garments due to detection or stitching failure, and GarmentRecovery lacks support for whole-body dress reconstruction. Additionally, inaccuracies in 2D garment segmentation [28] can lead to errors in GarmentRecovery reconstruction." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 526, + 297, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 526, + 297, + 586 + ], + "spans": [ + { + "bbox": [ + 55, + 526, + 297, + 586 + ], + "type": "text", + "content": "Garment designers often create image sketches before making sewing patterns. Surprisingly, despite not being trained on garment sketch images, ChatGarment demonstrates strong generalization, accurately predicting garment types and styles from sketches (Fig. 8)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 597, + 160, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 597, + 160, + 611 + ], + "spans": [ + { + "bbox": [ + 55, + 597, + 160, + 611 + ], + "type": "text", + "content": "4.2. Garment Editing" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 618, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 296, + 715 + ], + "type": "text", + "content": "**Benchmarks.** To evaluate garment editing performance, we construct an extra evaluation dataset using methods outlined in Sec. 3.2. We manually select garment pairs with clear textual descriptions and substantial visual differences. This dataset consists of 135 garment pairs, each accompanied by corresponding images and text descriptions. The evaluation of garment editing involves three components: the source garment, the target garment, and the editing prompts. The" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 465, + 555, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 465, + 555, + 490 + ], + "spans": [ + { + "bbox": [ + 313, + 465, + 555, + 490 + ], + "type": "text", + "content": "goal is to derive the sewing pattern for the target garment using the source garment and the editing prompts." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 493, + 556, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 493, + 556, + 529 + ], + "spans": [ + { + "bbox": [ + 313, + 493, + 556, + 529 + ], + "type": "text", + "content": "Implementation Details. For garment editing, we concatenate the source garment JSON file and the editing prompt together and pass it into ChatGarment." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 533, + 555, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 555, + 556 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 555, + 556 + ], + "type": "text", + "content": "Baselines. We modify DressCode [16] and SewFormer [33] for comparison with ChatGarment:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 558, + 556, + 714 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 314, + 558, + 556, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 558, + 556, + 641 + ], + "spans": [ + { + "bbox": [ + 314, + 558, + 556, + 641 + ], + "type": "text", + "content": "- DressCode with edited text. We generate a text description of the edited garment and use DressCode to create the corresponding sewing patterns. The description is produced by GPT-4o, which takes as inputs: 1) garment editing instructions and 2) front and back views of the source garment. The generated description follows DressCode's original input format." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "content": "- SewFormer with edited images. We obtain the front view of the edited garment with TurboEdit [9] to which we pass 1) the front view of the source garment, 2) the text description of the source garment, and 3) the editing instructions, and then use SewFormer to turn the image of the edited garment into the corresponding sewing pattern." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2930" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 70, + 93, + 113 + ], + "blocks": [ + { + "bbox": [ + 60, + 70, + 93, + 113 + ], + "lines": [ + { + "bbox": [ + 60, + 70, + 93, + 113 + ], + "spans": [ + { + "bbox": [ + 60, + 70, + 93, + 113 + ], + "type": "image", + "image_path": "9ff1a21d768bdfbea49cc33ffc7fdd7b8cadd8b9e2fd791d5626cb2c4771067e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 93, + 75, + 176, + 113 + ], + "lines": [ + { + "bbox": [ + 93, + 75, + 176, + 113 + ], + "spans": [ + { + "bbox": [ + 93, + 75, + 176, + 113 + ], + "type": "text", + "content": "Prompts: Change the shirt body to: [long vertical sections, 'evenly shaped sides', 'mildly loose fit around upper body, 'straight front opening']" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 179, + 70, + 212, + 113 + ], + "blocks": [ + { + "bbox": [ + 179, + 70, + 212, + 113 + ], + "lines": [ + { + "bbox": [ + 179, + 70, + 212, + 113 + ], + "spans": [ + { + "bbox": [ + 179, + 70, + 212, + 113 + ], + "type": "image", + "image_path": "19c5bf0ef2f9f16d057286836b1fe4d4c45216d93d4d45573ddab11973464445.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 217, + 71, + 250, + 113 + ], + "blocks": [ + { + "bbox": [ + 217, + 71, + 250, + 113 + ], + "lines": [ + { + "bbox": [ + 217, + 71, + 250, + 113 + ], + "spans": [ + { + "bbox": [ + 217, + 71, + 250, + 113 + ], + "type": "image", + "image_path": "c9084a328857aca1e3dcb54751bc2af2211e17043d397584fe818f0a22acb1c2.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 256, + 72, + 288, + 113 + ], + "blocks": [ + { + "bbox": [ + 256, + 72, + 288, + 113 + ], + "lines": [ + { + "bbox": [ + 256, + 72, + 288, + 113 + ], + "spans": [ + { + "bbox": [ + 256, + 72, + 288, + 113 + ], + "type": "image", + "image_path": "c565fac7f7a928ebea2c9840217b0f8bcc7804fc257560615d6b0d09d8dc2ae7.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 60, + 122, + 93, + 165 + ], + "blocks": [ + { + "bbox": [ + 60, + 122, + 93, + 165 + ], + "lines": [ + { + "bbox": [ + 60, + 122, + 93, + 165 + ], + "spans": [ + { + "bbox": [ + 60, + 122, + 93, + 165 + ], + "type": "image", + "image_path": "37afdfb42b8ab401af45ab2fa4688c7756bec37b332d490e7bf5ea88d47c195b.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 93, + 125, + 174, + 164 + ], + "lines": [ + { + "bbox": [ + 93, + 125, + 174, + 164 + ], + "spans": [ + { + "bbox": [ + 93, + 125, + 174, + 164 + ], + "type": "text", + "content": "Prompts: Change the sleeves to [short sleeves length], 'loose fit around the arms', 'wide sleeve openings relative to the arm size']" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 179, + 122, + 212, + 165 + ], + "blocks": [ + { + "bbox": [ + 179, + 122, + 212, + 165 + ], + "lines": [ + { + "bbox": [ + 179, + 122, + 212, + 165 + ], + "spans": [ + { + "bbox": [ + 179, + 122, + 212, + 165 + ], + "type": "image", + "image_path": "180bd5d407ab0829487cc6b440899d3c2a2a4a8d2d07b9f1dd9e3cd77c93a4ca.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 217, + 122, + 250, + 165 + ], + "blocks": [ + { + "bbox": [ + 217, + 122, + 250, + 165 + ], + "lines": [ + { + "bbox": [ + 217, + 122, + 250, + 165 + ], + "spans": [ + { + "bbox": [ + 217, + 122, + 250, + 165 + ], + "type": "image", + "image_path": "d8521218d572a4209f659e187388cdb9f7eae3f8a9ccc89cef820ed2b91c2ea8.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 256, + 122, + 288, + 165 + ], + "blocks": [ + { + "bbox": [ + 256, + 122, + 288, + 165 + ], + "lines": [ + { + "bbox": [ + 256, + 122, + 288, + 165 + ], + "spans": [ + { + "bbox": [ + 256, + 122, + 288, + 165 + ], + "type": "image", + "image_path": "533a58d36cf25e8ea9f26083089489d56b94c13b65609c010a598106f74149cf.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 60, + 175, + 93, + 217 + ], + "blocks": [ + { + "bbox": [ + 60, + 175, + 93, + 217 + ], + "lines": [ + { + "bbox": [ + 60, + 175, + 93, + 217 + ], + "spans": [ + { + "bbox": [ + 60, + 175, + 93, + 217 + ], + "type": "image", + "image_path": "4c2ed7e6fef802bedfa647d2577483b9f594f881657e242f633f0581f272feb5.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 93, + 183, + 173, + 213 + ], + "lines": [ + { + "bbox": [ + 93, + 183, + 173, + 213 + ], + "spans": [ + { + "bbox": [ + 93, + 183, + 173, + 213 + ], + "type": "text", + "content": "Prompts: Change the collar to: 'rounded neckline', 'slightly loose fit', 'even curvature throughout']" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 66, + 223, + 89, + 231 + ], + "lines": [ + { + "bbox": [ + 66, + 223, + 89, + 231 + ], + "spans": [ + { + "bbox": [ + 66, + 223, + 89, + 231 + ], + "type": "text", + "content": "Source" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 179, + 176, + 212, + 217 + ], + "blocks": [ + { + "bbox": [ + 179, + 176, + 212, + 217 + ], + "lines": [ + { + "bbox": [ + 179, + 176, + 212, + 217 + ], + "spans": [ + { + "bbox": [ + 179, + 176, + 212, + 217 + ], + "type": "image", + "image_path": "0d2f8dd010f2162a2ffc2f052f46ad145b3266f43a0897489b3b78d4827e2936.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 177, + 222, + 215, + 231 + ], + "lines": [ + { + "bbox": [ + 177, + 222, + 215, + 231 + ], + "spans": [ + { + "bbox": [ + 177, + 222, + 215, + 231 + ], + "type": "text", + "content": "SewFormer" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 217, + 176, + 250, + 217 + ], + "blocks": [ + { + "bbox": [ + 217, + 176, + 250, + 217 + ], + "lines": [ + { + "bbox": [ + 217, + 176, + 250, + 217 + ], + "spans": [ + { + "bbox": [ + 217, + 176, + 250, + 217 + ], + "type": "image", + "image_path": "d7e07ac8de50fa2558d416c6654e2e43ad0b05adb548ff3088ad974182bbcbc5.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 217, + 223, + 253, + 231 + ], + "lines": [ + { + "bbox": [ + 217, + 223, + 253, + 231 + ], + "spans": [ + { + "bbox": [ + 217, + 223, + 253, + 231 + ], + "type": "text", + "content": "DressCode" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 256, + 176, + 288, + 217 + ], + "blocks": [ + { + "bbox": [ + 256, + 176, + 288, + 217 + ], + "lines": [ + { + "bbox": [ + 256, + 176, + 288, + 217 + ], + "spans": [ + { + "bbox": [ + 256, + 176, + 288, + 217 + ], + "type": "image", + "image_path": "b74c049ce986a552e2356e8fa67f3c3f38af1ccc7f17641a1c9bbb9e139c3fcd.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 263, + 223, + 280, + 231 + ], + "lines": [ + { + "bbox": [ + 263, + 223, + 280, + 231 + ], + "spans": [ + { + "bbox": [ + 263, + 223, + 280, + 231 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 317, + 71, + 350, + 113 + ], + "blocks": [ + { + "bbox": [ + 317, + 71, + 350, + 113 + ], + "lines": [ + { + "bbox": [ + 317, + 71, + 350, + 113 + ], + "spans": [ + { + "bbox": [ + 317, + 71, + 350, + 113 + ], + "type": "image", + "image_path": "a1d4c71b3bac124afae014ab4b230de9e45737b948fd2a68f9b0ce130b2c703b.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 355, + 78, + 439, + 110 + ], + "lines": [ + { + "bbox": [ + 355, + 78, + 439, + 110 + ], + "spans": [ + { + "bbox": [ + 355, + 78, + 439, + 110 + ], + "type": "text", + "content": "Prompts: Change the sleeves to: 'long, full-length sleeves', 'loose fit', 'ends at wrist level', 'wide openings at the wrist']" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 444, + 71, + 476, + 113 + ], + "blocks": [ + { + "bbox": [ + 444, + 71, + 476, + 113 + ], + "lines": [ + { + "bbox": [ + 444, + 71, + 476, + 113 + ], + "spans": [ + { + "bbox": [ + 444, + 71, + 476, + 113 + ], + "type": "image", + "image_path": "3e06c9b751311b6cfdc479aa01ff0e4c1619d868d25622d7f1c1f24bb76c54fd.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 484, + 72, + 516, + 113 + ], + "blocks": [ + { + "bbox": [ + 484, + 72, + 516, + 113 + ], + "lines": [ + { + "bbox": [ + 484, + 72, + 516, + 113 + ], + "spans": [ + { + "bbox": [ + 484, + 72, + 516, + 113 + ], + "type": "image", + "image_path": "abea4b22b6cd2727c52dbc058a507361893f96d8c512a90b6eea5f74c270671e.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 520, + 71, + 552, + 113 + ], + "blocks": [ + { + "bbox": [ + 520, + 71, + 552, + 113 + ], + "lines": [ + { + "bbox": [ + 520, + 71, + 552, + 113 + ], + "spans": [ + { + "bbox": [ + 520, + 71, + 552, + 113 + ], + "type": "image", + "image_path": "6d75eb984adc16070c6856410fb3db55e1610999957b8982fbc87709ca0ea804.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 317, + 122, + 350, + 164 + ], + "blocks": [ + { + "bbox": [ + 317, + 122, + 350, + 164 + ], + "lines": [ + { + "bbox": [ + 317, + 122, + 350, + 164 + ], + "spans": [ + { + "bbox": [ + 317, + 122, + 350, + 164 + ], + "type": "image", + "image_path": "33fe012a9b1b041659498d6d9aa8210fc18240de2ebeff8bfd25eef6a56d961f.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 355, + 130, + 435, + 160 + ], + "lines": [ + { + "bbox": [ + 355, + 130, + 435, + 160 + ], + "spans": [ + { + "bbox": [ + 355, + 130, + 435, + 160 + ], + "type": "text", + "content": "Prompts: Change the pant legs to: [wide-leg silhouette], 'ankle-length hem', 'flowy fit', 'loose and relaxed cut']" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 444, + 122, + 476, + 165 + ], + "blocks": [ + { + "bbox": [ + 444, + 122, + 476, + 165 + ], + "lines": [ + { + "bbox": [ + 444, + 122, + 476, + 165 + ], + "spans": [ + { + "bbox": [ + 444, + 122, + 476, + 165 + ], + "type": "image", + "image_path": "0f5432f029a6a1ab872524fd68192027da4db7b0dae5db4931aa79ac0d7dc467.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 484, + 122, + 516, + 165 + ], + "blocks": [ + { + "bbox": [ + 484, + 122, + 516, + 165 + ], + "lines": [ + { + "bbox": [ + 484, + 122, + 516, + 165 + ], + "spans": [ + { + "bbox": [ + 484, + 122, + 516, + 165 + ], + "type": "image", + "image_path": "f06ef61ee1af0a17e79c22b8a48ce9496d8f597492c9fa7517e773f451809aa3.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 520, + 122, + 552, + 165 + ], + "blocks": [ + { + "bbox": [ + 520, + 122, + 552, + 165 + ], + "lines": [ + { + "bbox": [ + 520, + 122, + 552, + 165 + ], + "spans": [ + { + "bbox": [ + 520, + 122, + 552, + 165 + ], + "type": "image", + "image_path": "d9151dea9357da3aeb5c95355a408ea45ae190a09206af7f9aa60d3f59a593f9.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 317, + 175, + 350, + 217 + ], + "blocks": [ + { + "bbox": [ + 317, + 175, + 350, + 217 + ], + "lines": [ + { + "bbox": [ + 317, + 175, + 350, + 217 + ], + "spans": [ + { + "bbox": [ + 317, + 175, + 350, + 217 + ], + "type": "image", + "image_path": "0e616a57aaa7916db9ec0c2e07bdd4e2f2f99c31a0bbdea544148f6c75c239cc.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 355, + 178, + 438, + 217 + ], + "lines": [ + { + "bbox": [ + 355, + 178, + 438, + 217 + ], + "spans": [ + { + "bbox": [ + 355, + 178, + 438, + 217 + ], + "type": "text", + "content": "Prompts: Change the shirt body to: ['straight vertical silhouette', 'moderate fit around the torso', 'hemline falls above the knees']" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 322, + 224, + 345, + 232 + ], + "lines": [ + { + "bbox": [ + 322, + 224, + 345, + 232 + ], + "spans": [ + { + "bbox": [ + 322, + 224, + 345, + 232 + ], + "type": "text", + "content": "Source" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_caption" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 444, + 176, + 476, + 217 + ], + "blocks": [ + { + "bbox": [ + 444, + 176, + 476, + 217 + ], + "lines": [ + { + "bbox": [ + 444, + 176, + 476, + 217 + ], + "spans": [ + { + "bbox": [ + 444, + 176, + 476, + 217 + ], + "type": "image", + "image_path": "439c4f6a13ddb4a2efb711b79c3a4d59966c3bb8879c8b5b3a28c58ab6270b49.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 441, + 224, + 479, + 232 + ], + "lines": [ + { + "bbox": [ + 441, + 224, + 479, + 232 + ], + "spans": [ + { + "bbox": [ + 441, + 224, + 479, + 232 + ], + "type": "text", + "content": "SewFormer" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_caption" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 484, + 177, + 516, + 217 + ], + "blocks": [ + { + "bbox": [ + 484, + 177, + 516, + 217 + ], + "lines": [ + { + "bbox": [ + 484, + 177, + 516, + 217 + ], + "spans": [ + { + "bbox": [ + 484, + 177, + 516, + 217 + ], + "type": "image", + "image_path": "882692b5467db619b543d82c1b7c0a3c0d4f6f2fe71441478178609fe77792c4.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 484, + 224, + 519, + 232 + ], + "lines": [ + { + "bbox": [ + 484, + 224, + 519, + 232 + ], + "spans": [ + { + "bbox": [ + 484, + 224, + 519, + 232 + ], + "type": "text", + "content": "DressCode" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_caption" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 520, + 176, + 552, + 217 + ], + "blocks": [ + { + "bbox": [ + 520, + 176, + 552, + 217 + ], + "lines": [ + { + "bbox": [ + 520, + 176, + 552, + 217 + ], + "spans": [ + { + "bbox": [ + 520, + 176, + 552, + 217 + ], + "type": "image", + "image_path": "df09e02a735e5c81883df1e6f035b71d84af936fc82b003c61576e6009e9d92a.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 529, + 224, + 545, + 232 + ], + "lines": [ + { + "bbox": [ + 529, + 224, + 545, + 232 + ], + "spans": [ + { + "bbox": [ + 529, + 224, + 545, + 232 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_caption" + } + ], + "index": 33 + }, + { + "type": "image", + "bbox": [ + 72, + 274, + 124, + 407 + ], + "blocks": [ + { + "bbox": [ + 55, + 240, + 555, + 264 + ], + "lines": [ + { + "bbox": [ + 55, + 240, + 555, + 264 + ], + "spans": [ + { + "bbox": [ + 55, + 240, + 555, + 264 + ], + "type": "text", + "content": "Figure 7. Results on Garment Editing. We present six examples where the models (SewFormer [33], DressCode [16], and ChatGarment) edit the source garment shown in the source image according to the given editing instructions in the prompt boxes." + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 72, + 274, + 124, + 407 + ], + "lines": [ + { + "bbox": [ + 72, + 274, + 124, + 407 + ], + "spans": [ + { + "bbox": [ + 72, + 274, + 124, + 407 + ], + "type": "image", + "image_path": "7b3dd8724c3b750637254fc4a1515a12c4f805f3b7365816b61ba294d5846d4e.jpg" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 87, + 412, + 109, + 423 + ], + "lines": [ + { + "bbox": [ + 87, + 412, + 109, + 423 + ], + "spans": [ + { + "bbox": [ + 87, + 412, + 109, + 423 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_caption" + } + ], + "index": 39 + }, + { + "type": "image", + "bbox": [ + 129, + 274, + 227, + 407 + ], + "blocks": [ + { + "bbox": [ + 129, + 274, + 227, + 407 + ], + "lines": [ + { + "bbox": [ + 129, + 274, + 227, + 407 + ], + "spans": [ + { + "bbox": [ + 129, + 274, + 227, + 407 + ], + "type": "image", + "image_path": "62a5b265abfc2ad7acde6368371ddcdeb3708f65303741b7d64f963f940b1feb.jpg" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 140, + 412, + 214, + 423 + ], + "lines": [ + { + "bbox": [ + 140, + 412, + 214, + 423 + ], + "spans": [ + { + "bbox": [ + 140, + 412, + 214, + 423 + ], + "type": "text", + "content": "2D Sewing Pattern" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_caption" + } + ], + "index": 41 + }, + { + "type": "image", + "bbox": [ + 228, + 275, + 277, + 409 + ], + "blocks": [ + { + "bbox": [ + 228, + 275, + 277, + 409 + ], + "lines": [ + { + "bbox": [ + 228, + 275, + 277, + 409 + ], + "spans": [ + { + "bbox": [ + 228, + 275, + 277, + 409 + ], + "type": "image", + "image_path": "b1872f8f762845b0bb375ccc5c754dee17e38382397433a491a15656ac4fa1fa.jpg" + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 227, + 412, + 277, + 422 + ], + "lines": [ + { + "bbox": [ + 227, + 412, + 277, + 422 + ], + "spans": [ + { + "bbox": [ + 227, + 412, + 277, + 422 + ], + "type": "text", + "content": "3D Garment" + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 434, + 296, + 467 + ], + "lines": [ + { + "bbox": [ + 55, + 434, + 296, + 467 + ], + "spans": [ + { + "bbox": [ + 55, + 434, + 296, + 467 + ], + "type": "text", + "content": "Figure 8. Reconstruction results from garment sketches. The generated garments accurately capture the shape and style of the input sketch, despite our model being trained without sketches." + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "image_caption" + } + ], + "index": 43 + }, + { + "bbox": [ + 55, + 479, + 296, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 479, + 296, + 575 + ], + "spans": [ + { + "bbox": [ + 55, + 479, + 296, + 575 + ], + "type": "text", + "content": "Results. We compare our method with SewFormer and Dresscode on Chamfer Distance and F-score in Tab. 3, where our approach demonstrates superior performance. The qualitative results in Fig. 7 show that our method accurately modifies the targeted garment parts according to the input prompt. In contrast, SewFormer and DressCode often fail to precisely follow the prompt instructions and frequently alter other unintended areas of the garment." + } + ] + } + ], + "index": 46 + }, + { + "bbox": [ + 55, + 586, + 185, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 586, + 185, + 597 + ], + "spans": [ + { + "bbox": [ + 55, + 586, + 185, + 597 + ], + "type": "text", + "content": "4.3. Text-based Generation" + } + ] + } + ], + "index": 47 + }, + { + "bbox": [ + 55, + 605, + 297, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 297, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 297, + 715 + ], + "type": "text", + "content": "Following DressCode [16], we use GPT-4o to generate 150 highly diverse text labels from 85 selected in-the-wild images featuring various tops, skirts, pants, and dresses. Next, we render the generated 3D garments with default textures and compute the CLIP score using the provided text labels. See quantitative results in Tab. 4. Our model achieves a higher CLIP score compared to DressCode [16], highlighting the superior prompt-following capabilities of our approach. The qualitative results are shown in Sup. Mat." + } + ] + } + ], + "index": 48 + }, + { + "type": "table", + "bbox": [ + 358, + 270, + 509, + 311 + ], + "blocks": [ + { + "bbox": [ + 358, + 270, + 509, + 311 + ], + "lines": [ + { + "bbox": [ + 358, + 270, + 509, + 311 + ], + "spans": [ + { + "bbox": [ + 358, + 270, + 509, + 311 + ], + "type": "table", + "html": "
MethodsCD (↓)F-Score (↑)
SewFormer [33]28.770.548
DressCode [16]10.970.750
Ours2.510.893
", + "image_path": "df74d6c2c39c30b0baa674c4bbeb6290ea5e06a40e4caea7da033904a54a50ea.jpg" + } + ] + } + ], + "index": 49, + "angle": 0, + "type": "table_body" + } + ], + "index": 49 + }, + { + "type": "table", + "bbox": [ + 372, + 344, + 496, + 375 + ], + "blocks": [ + { + "bbox": [ + 313, + 315, + 555, + 337 + ], + "lines": [ + { + "bbox": [ + 313, + 315, + 555, + 337 + ], + "spans": [ + { + "bbox": [ + 313, + 315, + 555, + 337 + ], + "type": "text", + "content": "Table 3. Quantitative comparisons for garment editing. ChatGarment outperforms other two methods by a large margin." + } + ] + } + ], + "index": 50, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 372, + 344, + 496, + 375 + ], + "lines": [ + { + "bbox": [ + 372, + 344, + 496, + 375 + ], + "spans": [ + { + "bbox": [ + 372, + 344, + 496, + 375 + ], + "type": "table", + "html": "
MethodsCLIP score (↑)
DressCode [16]22.8
Ours23.7
", + "image_path": "fb7247fb472c239030c251f3787dc30725196d914eee8f5164a92bf59483aebf.jpg" + } + ] + } + ], + "index": 51, + "angle": 0, + "type": "table_body" + } + ], + "index": 51 + }, + { + "bbox": [ + 313, + 379, + 555, + 412 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 379, + 555, + 412 + ], + "spans": [ + { + "bbox": [ + 313, + 379, + 555, + 412 + ], + "type": "text", + "content": "Table 4. Quantitative comparisons for text-based garment generation. ChatGarment achieves a higher CLIP score than DressCode, indicating better alignment with the text inputs." + } + ] + } + ], + "index": 52 + }, + { + "bbox": [ + 314, + 423, + 378, + 435 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 423, + 378, + 435 + ], + "spans": [ + { + "bbox": [ + 314, + 423, + 378, + 435 + ], + "type": "text", + "content": "4.4. Ablation" + } + ] + } + ], + "index": 53 + }, + { + "bbox": [ + 313, + 442, + 555, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 442, + 555, + 479 + ], + "spans": [ + { + "bbox": [ + 313, + 442, + 555, + 479 + ], + "type": "text", + "content": "We evaluate the impact of different aspects of ChatGarment, including multi-modal LLM backbones, and various training datasets. Please refer to Sup. Mat." + } + ] + } + ], + "index": 54 + }, + { + "bbox": [ + 313, + 489, + 388, + 501 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 489, + 388, + 501 + ], + "spans": [ + { + "bbox": [ + 313, + 489, + 388, + 501 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 55 + }, + { + "bbox": [ + 313, + 509, + 556, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 509, + 556, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 509, + 556, + 712 + ], + "type": "text", + "content": "We introduce ChatGarment, a VLM-based model that estimates garments from images and generates or edits garments based on text descriptions. Leveraging the capabilities of LLMs, ChatGarment supports multi-turn dialogue interactions for 3D garment design. It surpasses existing sewing pattern-specific and LLM-based methods in accurately estimating sewing patterns from a single image. While ChatGarment excels in many areas, it may struggle with precise garment editing. For instance, altering the skirt length might unintentionally affect other parts of the garment. This issue could be addressed with more detailed text instructions or a differential CSG optimizer [60] for finer adjustments. Nonetheless, ChatGarment offers the first VLM-based framework for unified garment estimation, generation, and editing. Future directions could include expanding into garment manufacturing design or improving clothing simulation with more fine-grained material attributes." + } + ] + } + ], + "index": 56 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "2931" + } + ] + } + ], + "index": 57 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 231, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 231, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 231, + 85 + ], + "type": "text", + "content": "Acknowledgments and Disclosure" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 297, + 162 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 297, + 162 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 297, + 162 + ], + "type": "text", + "content": "We thank Maria Korosteleva for help on GarmentCode, and Giorgio Becherini for advice on data acquisition and processing. Yuliang Xiu is funded by Max Planck Institute for Intelligent Systems, the Research Center for Industries of the Future (RCIF) at Westlake University, and the Westlake Education Foundation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 163, + 296, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 163, + 296, + 198 + ], + "spans": [ + { + "bbox": [ + 55, + 163, + 296, + 198 + ], + "type": "text", + "content": "While MJB is a co-founder and Chief Scientist at Meshcapade, his research in this project was performed solely at, and funded solely by, the Max Planck Society." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 220, + 115, + 232 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 220, + 115, + 232 + ], + "spans": [ + { + "bbox": [ + 56, + 220, + 115, + 232 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 57, + 239, + 297, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 239, + 297, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 239, + 297, + 283 + ], + "spans": [ + { + "bbox": [ + 61, + 239, + 297, + 283 + ], + "type": "text", + "content": "[1] Rinet Abdrashitov, Kim Raichstat, Jared Monsen, and David Hill. Robust skin weights transfer via weight inpainting. In SIGGRAPH Asia 2023 Technical Communications, pages 1-4. 2023. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 285, + 296, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 285, + 296, + 328 + ], + "spans": [ + { + "bbox": [ + 61, + 285, + 296, + 328 + ], + "type": "text", + "content": "[2] Dimitrije Antic, Garvita Tiwari, Batuhan Ozcomlekci, Ricardo Marin, and Gerard Pons-Moll. CloSe: A 3D clothing segmentation dataset and model. In International Conference on 3D Vision (3DV), 2024. 2, 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 61, + 331, + 297, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 331, + 297, + 384 + ], + "spans": [ + { + "bbox": [ + 61, + 331, + 297, + 384 + ], + "type": "text", + "content": "[3] Bharat Lal Bhatnagar, Garvita Tiwari, Christian Theobalt, and Gerard Pons-Moll. Multi-garment net: Learning to dress 3d people from images. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5420-5430, 2019. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 61, + 386, + 296, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 386, + 296, + 441 + ], + "spans": [ + { + "bbox": [ + 61, + 386, + 296, + 441 + ], + "type": "text", + "content": "[4] Michael J Black, Priyanka Patel, Joachim Tesch, and Jinlong Yang. Bedlam: A synthetic dataset of bodies exhibiting detailed lifelike animated motion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8726-8737, 2023. 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 443, + 296, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 443, + 296, + 487 + ], + "spans": [ + { + "bbox": [ + 62, + 443, + 296, + 487 + ], + "type": "text", + "content": "[5] Andrés Casado-Elvira, Marc Comino Trinidad, and Dan Casas. Pergamo: Personalized 3d garments from monocular video. In Computer Graphics Forum, pages 293–304. Wiley Online Library, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 488, + 297, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 488, + 297, + 543 + ], + "spans": [ + { + "bbox": [ + 62, + 488, + 297, + 543 + ], + "type": "text", + "content": "[6] Enric Corona, Albert Pumarola, Guillem Alenya, Gerard Pons-Moll, and Francesc Moreno-Noguer. Simplicit: Topology-aware generative model for clothed people. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11875-11885, 2021. 2, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 544, + 296, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 544, + 296, + 588 + ], + "spans": [ + { + "bbox": [ + 62, + 544, + 296, + 588 + ], + "type": "text", + "content": "[7] R Daněřek, Endri Dibra, Cengiz Öztireli, Remo Ziegler, and Markus Gross. Deepgarment: 3d garment shape estimation from a single image. In Computer Graphics Forum, pages 269-280. Wiley Online Library, 2017. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 590, + 296, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 590, + 296, + 643 + ], + "spans": [ + { + "bbox": [ + 62, + 590, + 296, + 643 + ], + "type": "text", + "content": "[8] Luca De Luigi, Ren Li, Benoit Guillard, Mathieu Salzmann, and Pascal Fua. Drapenet: Garment generation and self-supervised draping. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1451-1460, 2023. 2, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 62, + 646, + 297, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 646, + 297, + 689 + ], + "spans": [ + { + "bbox": [ + 62, + 646, + 297, + 689 + ], + "type": "text", + "content": "[9] Gilad Deutch, Rinon Gal, Daniel Garibi, Or Patashnik, and Daniel Cohen-Or. Turboedit: Text-based image editing using few-step diffusion models. arXiv preprint arXiv:2408.00735, 2024. 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 691, + 296, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 296, + 714 + ], + "type": "text", + "content": "[10] Qingxiu Dong, Lei Li, Damai Dai, Ce Zheng, Jingyuan Ma, Rui Li, Heming Xia, Jingjing Xu, Zhiyong Wu, Tianyu" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 714 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 333, + 73, + 555, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 555, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 555, + 95 + ], + "type": "text", + "content": "Liu, et al. A survey on in-context learning. arXiv preprint arXiv:2301.00234, 2022. 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 96, + 555, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 555, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 555, + 128 + ], + "type": "text", + "content": "[11] Yao Feng, Jing Lin, Sai Kumar Dwivedi, Yu Sun, Priyanka Patel, and Michael J. Black. ChatPose: Chatting about 3d human pose. In CVPR, 2024. 2, 3, 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 129, + 555, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 555, + 172 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 555, + 172 + ], + "type": "text", + "content": "[12] Jianglin Fu, Shikai Li, Yuming Jiang, Kwan-Yee Lin, Chen Qian, Chen-Change Loy, Wayne Wu, and Ziwei Liu. Stylegan-human: A data-centric odyssey of human generation. arXiv preprint, arXiv:2204.11823, 2022. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 174, + 555, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 174, + 555, + 217 + ], + "spans": [ + { + "bbox": [ + 316, + 174, + 555, + 217 + ], + "type": "text", + "content": "[13] Rao Fu, Xiao Zhan, Yiwen Chen, Daniel Ritchie, and Srinath Sridhar. Shapecraft: A recursive text-conditioned 3d shape generation model. Advances in Neural Information Processing Systems, 35:8882-8895, 2022. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 217, + 555, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 217, + 555, + 271 + ], + "spans": [ + { + "bbox": [ + 316, + 217, + 555, + 271 + ], + "type": "text", + "content": "[14] Siavash Golkar, Mariel Pettee, Michael Eickenberg, Alberto Bietti, Miles Cranmer, Geraud Krawezik, Francois Lanusse, Michael McCabe, Ruben Ohana, Liam Parker, et al. xVal: A continuous number encoding for large language models. arXiv preprint arXiv:2310.02989, 2023. 2, 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 273, + 555, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 273, + 555, + 326 + ], + "spans": [ + { + "bbox": [ + 316, + 273, + 555, + 326 + ], + "type": "text", + "content": "[15] Artur Grigorev, Giorgio Becherini, Michael Black, Otmar Hilliges, and Bernhard Thomaszewski. Contourcraft: Learning to resolve intersections in neural multi-garment simulations. In ACM SIGGRAPH 2024 Conference Papers, pages 1-10, 2024. 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 327, + 555, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 327, + 555, + 370 + ], + "spans": [ + { + "bbox": [ + 316, + 327, + 555, + 370 + ], + "type": "text", + "content": "[16] Kai He, Kaixin Yao, Qixuan Zhang, Jingyi Yu, Lingjie Liu, and Lan Xu. Dresscode: Autoregressively sewing and generating garments from text guidance. ACM Transactions on Graphics (TOG), 43(4):1-13, 2024. 2, 3, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 372, + 555, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 372, + 555, + 415 + ], + "spans": [ + { + "bbox": [ + 316, + 372, + 555, + 415 + ], + "type": "text", + "content": "[17] Yining Hong, Haoyu Zhen, Peihao Chen, Shuhong Zheng, Yilun Du, Zhenfang Chen, and Chuang Gan. 3d-llm: Injecting the 3d world into large language models. Advances in Neural Information Processing Systems, 36:20482-20494, 2023. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 416, + 555, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 416, + 555, + 458 + ], + "spans": [ + { + "bbox": [ + 316, + 416, + 555, + 458 + ], + "type": "text", + "content": "[18] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In ICLR, 2022. 2, 3, 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 460, + 555, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 460, + 555, + 514 + ], + "spans": [ + { + "bbox": [ + 316, + 460, + 555, + 514 + ], + "type": "text", + "content": "[19] Ziniu Hu, Ahmet Iscen, Aashi Jain, Thomas Kipf, Yisong Yue, David A Ross, Cordelia Schmid, and Alireza Fathi. Scenecraft: An IIm agent for synthesizing 3d scenes as blender code. In *Forty-first International Conference on Machine Learning*, 2024. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 515, + 555, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 515, + 555, + 558 + ], + "spans": [ + { + "bbox": [ + 316, + 515, + 555, + 558 + ], + "type": "text", + "content": "[20] Yangyi Huang, Hongwei Yi, Yuliang Xiu, Tingting Liao, Ji-axiang Tang, Deng Cai, and Justus Thies. TeCH: Text-guided Reconstruction of Lifelike Clothed Humans. In International Conference on 3D Vision (3DV), 2024. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 559, + 555, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 559, + 555, + 613 + ], + "spans": [ + { + "bbox": [ + 316, + 559, + 555, + 613 + ], + "type": "text", + "content": "[21] Boyi Jiang, Juyong Zhang, Yang Hong, Jinhao Luo, Ligang Liu, and Hujun Bao. Bcnet: Learning body and cloth shape from a single image. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XX 16, pages 18-35. Springer, 2020. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 615, + 555, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 615, + 555, + 646 + ], + "spans": [ + { + "bbox": [ + 316, + 615, + 555, + 646 + ], + "type": "text", + "content": "[22] Maria Korosteleva and Sung-Hee Lee. Generating datasets of 3d garments with sewing patterns. arXiv preprint arXiv:2109.05633, 2021. 5" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 647, + 555, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 555, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 555, + 689 + ], + "type": "text", + "content": "[23] Maria Korosteleva and Sung-Hee Lee. Neuraltailer: Reconstructing sewing pattern structures from 3d point clouds of garments. ACM Transactions on Graphics (TOG), 41(4):1-16, 2022. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 691, + 555, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 555, + 714 + ], + "type": "text", + "content": "[24] Maria Korosteleva and Olga Sorkine-Hornung. GarmentCode: Programming parametric sewing patterns. ACM Transaction" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "2932" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 76, + 72, + 296, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 296, + 94 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 296, + 94 + ], + "type": "text", + "content": "on Graphics, 42(6), 2023. SIGGRAPH ASIA 2023 issue. 2, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 95, + 296, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 95, + 296, + 149 + ], + "spans": [ + { + "bbox": [ + 56, + 95, + 296, + 149 + ], + "type": "text", + "content": "[25] Maria Korosteleva, Timur Levent Kesdogan, Fabian Kemper, Stephan Wenninger, Jasmin Koller, Yuhan Zhang, Mario Botsch, and Olga Sorkine-Hornung. GarmentCodeData: A dataset of 3D made-to-measure garments with sewing patterns. In Computer Vision – ECCV 2024, 2024. 2, 3, 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 150, + 296, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 150, + 296, + 194 + ], + "spans": [ + { + "bbox": [ + 56, + 150, + 296, + 194 + ], + "type": "text", + "content": "[26] Peter Kulits, Haiwen Feng, Weiyang Liu, Victoria Abrevaya, and Michael J Black. Re-thinking inverse graphics with large language models. Transactions on Machine Learning Research, 2024. 2, 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 194, + 296, + 238 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 194, + 296, + 238 + ], + "spans": [ + { + "bbox": [ + 56, + 194, + 296, + 238 + ], + "type": "text", + "content": "[27] Boqian Li, Xuan Li, Ying Jiang, Tianyi Xie, Feng Gao, Huamin Wang, Yin Yang, and Chenfanfu Jiang. Garmentdreamer: 3dgs guided garment synthesis with diverse geometry and texture details. In 3DV, 2025. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 239, + 296, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 239, + 296, + 281 + ], + "spans": [ + { + "bbox": [ + 56, + 239, + 296, + 281 + ], + "type": "text", + "content": "[28] Peike Li, Yunqiu Xu, Yunchao Wei, and Yi Yang. Self-correction for human parsing. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(6):3260-3271, 2020. 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 283, + 296, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 283, + 296, + 327 + ], + "spans": [ + { + "bbox": [ + 56, + 283, + 296, + 327 + ], + "type": "text", + "content": "[29] Ren Li, Corentin Dumery, Benoit Guillard, and Pascal Fua. Garment recovery with shape and deformation priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1586-1595, 2024. 3, 5, 6, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 327, + 296, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 327, + 296, + 360 + ], + "spans": [ + { + "bbox": [ + 56, + 327, + 296, + 360 + ], + "type": "text", + "content": "[30] Ren Li, Benoit Guillard, and Pascal Fua. Isp: Multi-layered garment draping with implicit sewing patterns. Advances in Neural Information Processing Systems, 36, 2024. 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 361, + 296, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 361, + 296, + 415 + ], + "spans": [ + { + "bbox": [ + 56, + 361, + 296, + 415 + ], + "type": "text", + "content": "[31] Dingning Liu, Xiaoshui Huang, Yuenan Hou, Zhihui Wang, Zhenfei Yin, Yongshun Gong, Peng Gao, and Wanli Ouyang. Uni3d-llm: Unifying point cloud perception, generation and editing with large language models. arXiv preprint arXiv:2402.03327, 2024. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 415, + 296, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 415, + 296, + 437 + ], + "spans": [ + { + "bbox": [ + 56, + 415, + 296, + 437 + ], + "type": "text", + "content": "[32] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning, 2023. 3, 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 437, + 296, + 481 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 437, + 296, + 481 + ], + "spans": [ + { + "bbox": [ + 56, + 437, + 296, + 481 + ], + "type": "text", + "content": "[33] Lijuan Liu, Xiangyu Xu, Zhijie Lin, Jiabin Liang, and Shuicheng Yan. Towards garment sewing pattern reconstruction from a single image. ACM Transactions on Graphics (TOG), 42(6):1-15, 2023. 2, 3, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 482, + 296, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 482, + 296, + 536 + ], + "spans": [ + { + "bbox": [ + 56, + 482, + 296, + 536 + ], + "type": "text", + "content": "[34] Weiyang Liu, Zeju Qiu, Yao Feng, Yuliang Xiu, Yuxuan Xue, Longhui Yu, Haiwen Feng, Zhen Liu, Juyeon Heo, Songyou Peng, Yandong Wen, Michael J. Black, Adrian Weller, and Bernhard Scholkopf. Parameter-efficient orthogonal finetuning via butterfly factorization. In ICLR, 2024. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 537, + 296, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 537, + 296, + 591 + ], + "spans": [ + { + "bbox": [ + 56, + 537, + 296, + 591 + ], + "type": "text", + "content": "[35] Zhen Liu, Yao Feng, Yuliang Xiu, Weiyang Liu, Liam Paull, Michael J. Black, and Bernhard Scholkopf. Ghost on the shell: An expressive representation of general 3d shapes. In The Twelfth International Conference on Learning Representations, 2024. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 592, + 296, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 592, + 296, + 647 + ], + "spans": [ + { + "bbox": [ + 56, + 592, + 296, + 647 + ], + "type": "text", + "content": "[36] Zhongjin Luo, Haolin Liu, Chenghong Li, Wanghao Du, Zirong Jin, Yinyu Nie, Weikai Chen, and Xiaoguang Han. Garverselod: High-fidelity 3d garment reconstruction from a single in-the-wild image using a dataset with levels of details. In ACM Transactions on Graphics (TOG), 2024. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 647, + 296, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 296, + 691 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 296, + 691 + ], + "type": "text", + "content": "[37] Qianli Ma, Jinlong Yang, Siyu Tang, and Michael J. Black. The power of points for modeling humans in clothing. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 10974-10984, 2021. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 692, + 296, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 692, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 692, + 296, + 714 + ], + "type": "text", + "content": "[38] Qianli Ma, Jinlong Yang, Michael J. Black, and Siyu Tang. Neural point-based shape modeling of humans in challenging" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 556, + 712 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 333, + 73, + 555, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 555, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 555, + 95 + ], + "type": "text", + "content": "clothing. In International Conference on 3D Vision (3DV), pages 679-689, 2022. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 95, + 556, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 95, + 556, + 139 + ], + "spans": [ + { + "bbox": [ + 316, + 95, + 556, + 139 + ], + "type": "text", + "content": "[39] Miles Macklin. Warp: A high-performance python framework forgpu simulation and graphics. https://github.com/nvidia/warp, 2022. NVIDIA GPU Technology Conference (GTC). 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 140, + 555, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 140, + 555, + 183 + ], + "spans": [ + { + "bbox": [ + 316, + 140, + 555, + 183 + ], + "type": "text", + "content": "[40] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed A. A. Osman, Dimitrios Tzionas, and Michael J. Black. Expressive body capture: 3D hands, face, and body from a single image. In CVPR, 2019. 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 184, + 555, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 184, + 555, + 215 + ], + "spans": [ + { + "bbox": [ + 316, + 184, + 555, + 215 + ], + "type": "text", + "content": "[41] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. In ICLR, 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 217, + 555, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 217, + 555, + 272 + ], + "spans": [ + { + "bbox": [ + 316, + 217, + 555, + 272 + ], + "type": "text", + "content": "[42] Lingteng Qiu, Guanying Chen, Jiapeng Zhou, Mutian Xu, Junle Wang, and Xiaoguang Han. Rec-mv: Reconstructing 3d dynamic cloth from monocular videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4637-4646, 2023. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 272, + 554, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 272, + 554, + 316 + ], + "spans": [ + { + "bbox": [ + 316, + 272, + 554, + 316 + ], + "type": "text", + "content": "[43] Zeju Qiu, Weiyang Liu, Haiwen Feng, Yuxuan Xue, Yao Feng, Zhen Liu, Dan Zhang, Adrian Weller, and Bernhard Schölkopf. Controlling text-to-image diffusion by orthogonal finetuning. In NeurIPS, 2023. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 316, + 554, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 316, + 554, + 370 + ], + "spans": [ + { + "bbox": [ + 316, + 316, + 554, + 370 + ], + "type": "text", + "content": "[44] Zeju Qiu, Weiyang Liu, Haiwen Feng, Zhen Liu, Tim Z Xiao, Katherine M Collins, Joshua B Tenenbaum, Adrian Weller, Michael J Black, and Bernhard Scholkopf. Can large language models understand symbolic graphics programs? arXiv preprint arXiv:2408.08313, 2024. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 371, + 555, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 371, + 555, + 415 + ], + "spans": [ + { + "bbox": [ + 316, + 371, + 555, + 415 + ], + "type": "text", + "content": "[45] Shunsuke Saito, Zeng Huang, Ryota Natsume, Shigeo Morishima, Hao Li, and Angjoo Kanazawa. PIFu: Pixel-aligned implicit function for high-resolution clothed human digitization. In ICCV, pages 2304-2314, 2019. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 415, + 554, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 415, + 554, + 458 + ], + "spans": [ + { + "bbox": [ + 316, + 415, + 554, + 458 + ], + "type": "text", + "content": "[46] Shunsuke Saito, Tomas Simon, Jason Saragih, and Hanbyul Joo. PIFuHD: Multi-Level Pixel-Aligned Implicit Function for High-Resolution 3D Human Digitization. In CVPR, pages 81–90, 2020. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 460, + 554, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 460, + 554, + 492 + ], + "spans": [ + { + "bbox": [ + 316, + 460, + 554, + 492 + ], + "type": "text", + "content": "[47] Nikolaos Sarafianos, Tuur Stuyck, Xiaoyu Xiang, Yilei Li, Jovan Popovic, and Rakesh Ranjan. Garment3dgen: 3d garment stylization and texture generation. In 3DV, 2025. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 493, + 555, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 493, + 555, + 536 + ], + "spans": [ + { + "bbox": [ + 316, + 493, + 555, + 536 + ], + "type": "text", + "content": "[48] Tianchang Shen, Jun Gao, Kangxue Yin, Ming-Yu Liu, and Sanja Fidler. Deep marching tetrahedra: a hybrid representation for high-resolution 3d shape synthesis. In Advances in Neural Information Processing Systems (NeurIPS), 2021. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 537, + 555, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 537, + 555, + 568 + ], + "spans": [ + { + "bbox": [ + 316, + 537, + 555, + 568 + ], + "type": "text", + "content": "[49] Yu Shen, Junbang Liang, and Ming C. Lin. Gan-based garment generation using sewing pattern images. In ECCV, 2020. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 570, + 555, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 570, + 555, + 613 + ], + "spans": [ + { + "bbox": [ + 316, + 570, + 555, + 613 + ], + "type": "text", + "content": "[50] Astitva Srivastava, Pranav Manu, Amit Raj, Varun Jampani, and Avinash Sharma. Wordrobe: Text-guided generation of textured 3d garments. In ECCV, pages 458-475. Springer, 2025. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 614, + 555, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 614, + 555, + 668 + ], + "spans": [ + { + "bbox": [ + 316, + 614, + 555, + 668 + ], + "type": "text", + "content": "[51] Zhaoqi Su, Weilin Wan, Tao Yu, Lingjie Liu, Lu Fang, Wenping Wang, and Yebin Liu. Mulaycap: Multi-layer human performance capture using a monocular video camera. IEEE Transactions on Visualization and Computer Graphics, 28(4): 1862-1879, 2020. 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 669, + 555, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 669, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 669, + 555, + 712 + ], + "type": "text", + "content": "[52] Chunyi Sun, Junlin Han, Weijian Deng, Xinlong Wang, Zishan Qin, and Stephen Gould. 3d-gpt: Procedural 3d modeling with large language models. arXiv preprint arXiv:2310.12945, 2023. 3" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2933" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 297, + 564 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 56, + 73, + 297, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 297, + 115 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 297, + 115 + ], + "type": "text", + "content": "[53] Tuanfeng Y. Wang, Duygu Ceylan, Jovan Popovic, and Niloy J. Mitra. Learning a shared shape space for multimodal garment design. ACM Trans. Graph., 37(6), 2018. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 296, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 296, + 173 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 296, + 173 + ], + "type": "text", + "content": "[54] Wenbo Wang, Hsuan-I Ho, Chen Guo, Boxiang Rong, Artur Grigorev, Jie Song, Juan Jose Zarate, and Otmar Hilliges. 4D-DRESS: A 4D dataset of real-world human clothing with semantic annotations. In IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2024. 2, 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 174, + 296, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 174, + 296, + 228 + ], + "spans": [ + { + "bbox": [ + 56, + 174, + 296, + 228 + ], + "type": "text", + "content": "[55] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824-24837, 2022. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 229, + 296, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 229, + 296, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 229, + 296, + 262 + ], + "type": "text", + "content": "[56] Yuliang Xiu, Jinlong Yang, Dimitrios Tzionas, and Michael J. Black. ICON: Implicit Clothed humans Obtained from Normals. In CVPR, 2022. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 263, + 294, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 263, + 294, + 296 + ], + "spans": [ + { + "bbox": [ + 56, + 263, + 294, + 296 + ], + "type": "text", + "content": "[57] Yuliang Xiu, Jinlong Yang, Xu Cao, Dimitrios Tzionas, and Michael J. Black. ECON: Explicit Clothed humans Optimized via Normal integration. In CVPR, 2023. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 297, + 296, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 297, + 296, + 351 + ], + "spans": [ + { + "bbox": [ + 56, + 297, + 296, + 351 + ], + "type": "text", + "content": "[58] Runsen Xu, Xiaolong Wang, Tai Wang, Yilun Chen, Jiangmiao Pang, and Dahua Lin. Pointllm: Empowering large language models to understand point clouds. In European Conference on Computer Vision, pages 131-147. Springer, 2025. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 353, + 296, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 353, + 296, + 407 + ], + "spans": [ + { + "bbox": [ + 56, + 353, + 296, + 407 + ], + "type": "text", + "content": "[59] Zhengming Yu, Zhiyang Dou, Xiaoxiao Long, Cheng Lin, Zekun Li, Yuan Liu, Norman Müller, Taku Komura, Marc Habermann, Christian Theobalt, et al. Surf-d: Generating high-quality surfaces of arbitrary topologies using diffusion models. In ECCV, pages 419-438. Springer, 2025. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 409, + 296, + 452 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 409, + 296, + 452 + ], + "spans": [ + { + "bbox": [ + 56, + 409, + 296, + 452 + ], + "type": "text", + "content": "[60] Haocheng Yuan, Adrien Bousseau, Hao Pan, Chengquan Zhang, Niloy J Mitra, and Changjian Li. Diffcsg: Differentiable csg via rasterization. arXiv preprint arXiv:2409.01421, 2024.8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 454, + 296, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 454, + 296, + 508 + ], + "spans": [ + { + "bbox": [ + 56, + 454, + 296, + 508 + ], + "type": "text", + "content": "[61] Longwen Zhang, Ziyu Wang, Qixuan Zhang, Qiwei Qiu, Anqi Pang, Haoran Jiang, Wei Yang, Lan Xu, and Jingyi Yu. Clay: A controllable large-scale generative model for creating high-quality 3d assets. ACM Transactions on Graphics (TOG), 43 (4):1-20, 2024. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 510, + 296, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 510, + 296, + 564 + ], + "spans": [ + { + "bbox": [ + 56, + 510, + 296, + 564 + ], + "type": "text", + "content": "[62] Heming Zhu, Lingteng Qiu, Yuda Qiu, and Xiaoguang Han. Registering explicit to implicit: Towards high-fidelity garment mesh reconstruction from single images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3845-3854, 2022. 2" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2934" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/ChatGen_ Automatic Text-to-Image Generation From FreeStyle Chatting/0c8e9c4a-231c-4c6d-af84-cd3eb4170579_content_list.json b/2025/ChatGen_ Automatic Text-to-Image Generation From FreeStyle Chatting/0c8e9c4a-231c-4c6d-af84-cd3eb4170579_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..bef74d955db21142529eb0235e61cbb83513c594 --- /dev/null +++ b/2025/ChatGen_ Automatic Text-to-Image Generation From FreeStyle Chatting/0c8e9c4a-231c-4c6d-af84-cd3eb4170579_content_list.json @@ -0,0 +1,2270 @@ +[ + { + "type": "image", + "img_path": "images/db1b627fef85d9975afd23e19c76597338f4300abef7e78599868821ef8ab03e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 112, + 123, + 150, + 159 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ChatGen: Automatic Text-to-Image Generation From FreeStyle Chatting", + "text_level": 1, + "bbox": [ + 150, + 140, + 885, + 162 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Chengyou Jia $^{1*}$ , Changliang Xia $^{1*}$ , Zhuohang Dang $^{1}$ , Weijia Wu $^{2}$ , Hangwei Qian $^{3}$ , Minnan Luo $^{1\\dagger}$ $^{1}$ School of Computer Science and Technology, MOEKLINNS Lab, Xi'an Jiaotong University \n $^{2}$ National University of Singapore $^{3}$ CFAR, A\\*STAR", + "bbox": [ + 109, + 189, + 893, + 244 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "cp3jia@stu.xjtu.edu.cn, 202066@stu.xjtu.edu.cn, dangzhuohang@stu.xjtu.edu.cn, weijiawu96@gmail.com, qian0045@e.ntu.edu.sg, minnluo@xjtu.edu.cn", + "bbox": [ + 156, + 247, + 836, + 279 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 313, + 326, + 329 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite the significant advancements in text-to-image (T2I) generative models, users often face a trial-and-error challenge in practical scenarios. This challenge arises from the complexity and uncertainty of tedious steps such as crafting suitable prompts, selecting appropriate models, and configuring specific arguments, making users resort to labor-intensive attempts for desired images. This paper proposes Automatic T2I generation, which aims to automate these tedious steps, allowing users to simply describe their needs in a freestyle chatting way. To systematically study this problem, we first introduce ChatGenBench, a novel benchmark designed for Automatic T2I. It features high-quality paired data with diverse freestyle inputs, enabling comprehensive evaluation of automatic T2I models across all steps. Additionally, recognizing Automatic T2I as a complex multi-step reasoning task, we propose ChatGenEvo, a multi-stage evolution strategy that progressively equips models with essential automation skills. Through extensive evaluation across step-wise accuracy and image quality, ChatGen-Evo significantly enhances performance over various baselines. Our evaluation also uncovers valuable insights for advancing automatic T2I. Our data, code, and models will be publicly available.", + "bbox": [ + 88, + 347, + 483, + 694 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 715, + 220, + 729 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In recent years, text-to-image (T2I) generative models have attracted considerable attention [1, 23, 31, 32]. Building on advancements in large-scale T2I models such as DALL-E [31] and Stable Diffusion [32], the open-source community has significantly expanded the capabilities of T2I generation. Researchers fine-tune open-source models on specialized datasets, resulting in a diverse selection of task-specific models available on platforms like Civitai [6] and Hugging", + "bbox": [ + 89, + 741, + 483, + 863 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5cdd819e119516922bb9791598092e026f4c4d5dd6706a31278adeba8876c65c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 535, + 316, + 885, + 526 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/437ff6f487593781e614df16cb07e5a2ac787539257ca89cbe30109cbb0c3f64.jpg", + "image_caption": [ + "Figure 1. Illustration of tedious steps in T2I. Our method can select an appropriate model with suitable prompts and arguments. Note: Arle is a character from the game Genshin Impact." + ], + "image_footnote": [], + "bbox": [ + 535, + 529, + 885, + 638 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Face [8]. This variety provides users with a broader range of options to meet customization needs, facilitating the growing adoption of T2I models in real-world applications.", + "bbox": [ + 511, + 686, + 906, + 733 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "However, the rapid development of T2I models within the open-source community has also introduced significant challenges for users. When non-experts attempt to create images with specific requirements, they often encounter a trial-and-error process involving several tedious steps. As shown in Figure 1, these steps include crafting suitable prompts, selecting appropriate models, and configuring specific model arguments. The complexity and uncertainty of each step turn the process into an arduous journey, resembling \"mice in a maze\". In real-world scenarios, this iterative process consumes substantial time and resources", + "bbox": [ + 509, + 734, + 908, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal Contribution.", + "bbox": [ + 107, + 875, + 222, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author.", + "bbox": [ + 109, + 887, + 235, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "13284", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "as users continuously adjust settings to regenerate images. Therefore, we pose the challenging problem: can we automate these labor-intensive steps in T2I generation, allowing users to simply describe their needs in a chatting style and receive desired images effortlessly?", + "bbox": [ + 89, + 90, + 480, + 167 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Previous attempts, such as BeautifulPrompt [4, 38] for generating high-quality prompts from low-quality ones and DiffAgent [49] for model selection using large language models, have made some progress in addressing these challenges. However, these methods only seek to automate a specific step in Figure 1, lacking comprehensive research on the automation of entire T2I process. Moreover, they fail to support diverse types of freestyle input, leaving them far from real-world scenarios. To bridge these gaps, we aim to develop the model that can accept arbitrary user input, similar to ChatGPT [24], and automatically generate all necessary components for generation, termed as Automatic T2I.", + "bbox": [ + 89, + 169, + 480, + 349 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To this end, we first introduce ChatGenBench, the benchmark specifically designed for this task. ChatGenBench offers a substantial dataset of high-quality paired data from 6,807 customized models. Each data pair comprises a user's freestyle chat input, a refined prompt, an appropriate model, and optimal arguments. This comprehensive step-by-step trail enables step-wise evaluation of automatic T2I models, ensuring both quality assessment of the final image output and precise identification of potential automation bottlenecks. Furthermore, ChatGenBench integrates various types of data, enabling testing with multimodal and historical input to simulate practical scenarios.", + "bbox": [ + 89, + 352, + 482, + 532 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "An intuitive approach to achieve the goal is to build supervised fine-tuning (SFT) data and tune multimodal large language models. However, we argue that automated T2I should be viewed as a complex multi-step reasoning problem. Directly predicting outputs leads the model to focus on simple text mappings, rather than developing the diverse skills needed for robust automation. Inspired by OpenAI's o1 [25], we propose ChatGen-Evo with a multi-stage evolution strategy. This approach provides targeted feedback at each stage, progressively equipping models with essential automation skills. First, we construct SFT data to train the model to generate high-quality prompts from freestyle inputs. Next, we augment the model's vocabulary with specialized ModelTokens, enabling effective model selection without affecting other functions. Finally, we guide the model in configuring arguments based on prompts and models selected in previous stages. By decomposing the task into clear stages, the model gains crucial automation skills, delivering outputs aligned with user expectations.", + "bbox": [ + 89, + 536, + 482, + 823 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We conduct a comprehensive evaluation of various methods on the novel ChatGenBench to study Automatic T2I. ChatGen-Evo significantly outperforms other methods across all metrics, including both step-wise assessments and image quality evaluations. Extensive quantitative and", + "bbox": [ + 89, + 825, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "qualitative results, along with ablation studies, underscore the critical importance of the multi-stage evolution strategy. Additionally, further experiments reveal the impact of different stages on the final T2I output, uncovering valuable insights into the challenges and opportunities in Automatic T2I. We highlight some promising directions, such as exploring the scaling laws of prompt rewriting, improving model selection in few-shot scenarios, and leveraging reasoning approaches to enhance performance. The contributions of this paper are summarized as follows:", + "bbox": [ + 511, + 90, + 903, + 242 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose the novel and challenging problem of Automatic T2I generation, aiming to develop the model that can handle user's freestyle chatting and automatically produce all necessary components for image generation.", + "- We introduce ChatGenBench, a benchmark specifically designed for Automatic T2I. It includes a comprehensive step-by-step trail for step-wise evaluation, supporting multimodal or historical user inputs.", + "- We present ChatGen-Evo, which adopts a multi-stage evolution strategy. By decomposing the task into distinct stages, ChatGen-Evo enables the model to progressively acquire essential Automatic T2I capabilities.", + "- Extensive experiments on ChatGenBench not only demonstrate the superiority of ChatGen-Evo but also provide valuable insights for advancing automatic T2I." + ], + "bbox": [ + 511, + 244, + 903, + 470 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 511, + 486, + 661, + 501 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Text-to-Image Generation", + "text_level": 1, + "bbox": [ + 511, + 511, + 748, + 526 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "With the advancement of diffusion models [13, 27, 34], text-to-image (T2I) generation [1, 23, 31] has demonstrated exceptional capabilities in high-quality image generation and textual description alignment. Large-scale models such as DALL-E [1] enhance text-image alignment by leveraging the joint feature space of CLIP [30]. Moreover, Stable Diffusion [32], a well-established open-source model, has gained substantial attention. Numerous methods have been developed to fine-tune it or design additional modules for specialized tasks, such as customized image generation [17, 18, 33], layout-to-image generation [14, 19, 20, 47] and image edit [2, 16]. These diverse models have significantly expanded the capabilities of T2I. However, they also present significant learning challenges for non-expert users, underscoring the increasing need for automatic T2I generation.", + "bbox": [ + 511, + 534, + 903, + 762 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. LLMs for Text-to-Image Generation", + "text_level": 1, + "bbox": [ + 511, + 771, + 828, + 786 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Large language models (LLMs), like ChatGPT [24], Llama [37], have demonstrated impressive capability in language understanding [3] and problem solving [45]. Recently, LLMs have also been applied to image generation. [7, 9, 20, 36, 43] leverage LLMs to generate detailed layout information from complex prompts, enabling the creation of sub-elements and control over their positioning.", + "bbox": [ + 511, + 794, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "13285", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/47bba521a124c435dac97253b6eec3ba31b070c1ca5af27dc5c30cf25b123934.jpg", + "image_caption": [ + "Figure 2. Illustration of the data collection pipeline." + ], + "image_footnote": [], + "bbox": [ + 133, + 88, + 859, + 208 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "[28, 39, 49] propose employ LLMs for model selection, akin to tool usage [26, 29, 44]. However, the above methods still require specialized prompt inputs and involve complex model usage. Currently, no approach thoroughly considers leveraging LLMs to relieve users from all tedious steps in T2I. Our method aims to address this gap.", + "bbox": [ + 89, + 232, + 483, + 324 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 89, + 338, + 225, + 356 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our goal is to relieve users from tedious steps and automate to produce the desired images from user's freestyle input. To achieve this objective, we first introduce ChatGenBench, a benchmark comprising a large set of user inputs in a chatting style, built upon a foundation of over 6000 personalized models to evaluate automated image generation results. A comprehensive description of data collection, construction, and analysis is provided in Sec. 3.1. Then, we present ChatGen-Evo in Sec. 3.2, which uses a multi-stage evolution strategy to train MLLM for Automatic T2I.", + "bbox": [ + 89, + 364, + 483, + 515 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. ChatGenBench: Benchmarking Automatic T2I", + "text_level": 1, + "bbox": [ + 89, + 527, + 483, + 542 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For clarity, an example of this data is shown on the right side of Figure 2. The following sections detail the data collection and construction process, which primarily involves High-Quality Human Demonstration Collection and using LLM-driven Role-Play to simulate user input.", + "bbox": [ + 89, + 551, + 483, + 628 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1.1 High-Quality Human Demonstration Collection", + "text_level": 1, + "bbox": [ + 89, + 648, + 477, + 662 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Civitai [6] is a vibrant community where users share customized AI models for generating high-quality images. Members contribute demonstrations with detailed prompts, model specifications, and arguments, supported by a feedback system that ensures quality validation. These features make Civitai an ideal platform for collecting raw data.", + "bbox": [ + 89, + 672, + 482, + 763 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Collection: We start by collecting demonstrations based on established evaluation metrics within Civitai, including download counts, upvotes, and other user feedback. These metrics enable us to collect data that has been validated through community engagement. By focusing on these indicators, we identify a subset of high-quality results.", + "bbox": [ + 89, + 763, + 482, + 854 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Filtering: Following the initial collection, we implement a rigorous filtering process to ensure data quality. This involves excluding demonstrations associated with inactive", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "or outdated models, removing duplicates, and filtering out NSFW content. This careful curation refined the dataset to include the most effective demonstrations. Ultimately, this process results in a curated set of 44,881 high-quality human demonstrations across 6,807 unique models.", + "bbox": [ + 511, + 232, + 905, + 309 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1.2 LLM-Driven Role-Play for Chatting Generation", + "text_level": 1, + "bbox": [ + 511, + 332, + 901, + 348 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "While the demonstrations collected from communities include the essential procedural information needed for automation, they lack a critical element: freestyle chatting inputs. Such data is not available on open platforms, which has been a key limitation for previous methods [49]. To address this, we propose an LLM-Driven Role-Play strategy for Chatting Generation. As shown in Figure 2, we predefine over 100 roles from everyday life (e.g., students, doctors, professors) and prompt the LLM to simulate these roles, translating each demonstration into freestyle chatting input with tones and habits of the character. This approach significantly enhances data diversity and creates more lifelike inputs. To further enhance reverse synthesis diversity, we incorporate multiple LLMs, adjust diversity arguments, and use BERTScore [48] to filter out overly similar inputs.", + "bbox": [ + 511, + 357, + 906, + 584 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Additionally, we define three types of freestyle input formats: single-input, consisting of a single chatting-style sentence; multimodal-input, combining a sentence with an image; and history-input, comprising multiple rounds of dialogue history. These formats effectively simulate how users typically inquire about image generation needs, greatly expanding the practical value of automatic T2I. An example of the single-input prompt is shown in the following. Additional prompts and details are provided in the Appendix.", + "bbox": [ + 511, + 585, + 905, + 722 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Prompt: Sing-Input chatting generation", + "text_level": 1, + "bbox": [ + 524, + 736, + 784, + 752 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "System: You are a professional user experience designer who plays various personas to convert complex and professional content for non-professional users. Please merge the following prompt and model information into a single freestyle query. Remove any obvious details that non-professional users would avoid. Make it similar to what non-professional users may write. The converted single-text query should be colloquial and as brief as possible.", + "bbox": [ + 521, + 762, + 895, + 845 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Require: {ROLE}, {PROMPT}, {MODEL}, {Example 1,...,n} \nInput: You are playing the {ROLE}. Please generate a single text query based on the following professional {PROMPT} for the {MODEL}. You can refer to following examples: {Example 1,...,n}.", + "bbox": [ + 522, + 845, + 893, + 895 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "13286", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/54bbbf3262b5169d39b07cb8d9d5874f34e4671d616a15f81af5c69fd4d597ce.jpg", + "table_caption": [ + "Table 1. Comparison of different methods for benchmarking automatic text-to-image generation." + ], + "table_footnote": [], + "table_body": "
TrainSet (# Numbers)TestSet (# Numbers)Based Models?Step-wise Evaluation?Freestyle User-input?Multi-modal Support?History Support?
BeautifulPrompt [4]143K2KSingleX✓XX
DiffChat [38]234K5KSingleX✓X✓
DiffusionGPT [28]--Multi(≈20)X✓XX
DABench [49]40K5KMulti(5K)XXXX
ChatGenBench256K14KMulti(6K)✓✓✓✓
", + "bbox": [ + 91, + 103, + 911, + 228 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/47060c90baf3a52e8e1e13645835c3d9fa4cbccab62e92b963117c05a67b40f7.jpg", + "table_caption": [ + "Table 2. Summary of dataset statistics." + ], + "table_footnote": [], + "table_body": "
DatasetTotalSingleM-ModalHistory
TrainSet256,606147,88869,54839,170
TestSet Init74,36442,83820,21411,312
Benchmark14,56411,0111,6681,132
Supervised10,2408,0091,0991,132
Few-Shot4,3243,002569753
", + "bbox": [ + 91, + 250, + 486, + 371 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1.3 Benchmark Construction", + "text_level": 1, + "bbox": [ + 89, + 376, + 321, + 390 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Following the above steps, we generate 330,970 freestyle inputs from 44,881 demonstrations. Considering the largescale generated data, we don't randomly split a test set for benchmarking. Instead, we carefully select high-quality, non-overlapping samples as the benchmark for evaluation. The selection involves the following steps:", + "bbox": [ + 89, + 400, + 482, + 491 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "TestSet Split: We split the data based on model origin. For data associated with the same model, we use BERTScore to assess the similarity between samples. From this, we select about $20\\%$ most semantically distinct data as initial TestSet in Table 2. This maximizes non-overlap with TrainSet.", + "bbox": [ + 89, + 491, + 483, + 566 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "TestSet Filtering: Building on the initial TestSet, we perform multiple filtering rounds to ensure the final ChatGen-Bench's quality. This involves the following steps:", + "bbox": [ + 89, + 566, + 483, + 613 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Length Filtering: Remove excessively long entries or those with too many dialogue turns for consistency.", + "- Colloquialism Check: Utilize the Spacy [35] module to filter data for alignment with natural, everyday language.", + "- LLM-Based Evaluation: Employ LLM to assess and select data that best matches the chatting tone.", + "- Manual Verification: Manual verification is conducted to eliminate any inappropriate or irrelevant samples." + ], + "bbox": [ + 89, + 613, + 482, + 733 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Details of each step are provided in the Appendix. Through this rigorous process, we refine the initial 74,364 entries to a final set of 14,564 high-quality, freestyle chatting samples, constituting the novel benchmark.", + "bbox": [ + 89, + 734, + 482, + 794 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Setting Division: In practical scenarios, new models often have limited demonstration data, underscoring the importance of evaluations under constrained data. Therefore, we further divide ChatGenBench into two settings—Supervised and Few-shot—based on the availability of samples for each model in the TrainSet. The complete data composition is detailed in Table 2.", + "bbox": [ + 89, + 795, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1.4 Benchmark Analysis", + "text_level": 1, + "bbox": [ + 511, + 233, + 714, + 248 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ChatGenBench offers distinct advantages over previous benchmarks, as summarized in Table 1. On the data level, ChatGenBench includes large-scale high-quality data and a broader range of T2I models. Additionally, human demonstrations provide relative ground truth across each step, allowing for step-wise evaluation that pinpoints potential challenges in automation models. ChatGenBench is also the novel benchmark to support multiple input types, making it more aligned with real-world scenarios.", + "bbox": [ + 511, + 257, + 906, + 393 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. ChatGen: Achieving Automatic T2I", + "text_level": 1, + "bbox": [ + 511, + 404, + 825, + 420 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our goal is to train the model capable of processing freestyle user inputs and generating the necessary components for image generation (prompt, model, and argument), thereby achieving Automatic T2I. In this section, we first introduce a baseline method, ChatGen-Base. We then analyze the limitations of this approach and subsequently propose ChatGen-Evo, which leverages the multi-stage evolution strategy to enhance performance.", + "bbox": [ + 511, + 428, + 906, + 549 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.1 ChatGen-Base with SFT", + "text_level": 1, + "bbox": [ + 511, + 568, + 738, + 583 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We first apply supervised fine-tuning (SFT) to develop the ChatGen-Base model as an intuitive baseline. Given a set of freestyle chatting inputs $c$ (which may include text, images, and historical context) and their corresponding outputs comprising prompt $p$ , model $m$ , and argument $r$ , we use the auto-regressive objective to maximize the following loss:", + "bbox": [ + 511, + 593, + 905, + 696 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {s f t} ^ {\\text {b a s e}} = - \\sum_ {t} \\log P _ {\\pi} (p, m, r \\mid c, * _ {< t}). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 712, + 906, + 742 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ChatGen-Base satisfies the essential requirements for generating outputs, including handling diverse input modalities and producing corresponding responses. However, even after fine-tuning SOTA open-source MLLMs, the model frequently produces unsatisfactory results.", + "bbox": [ + 511, + 750, + 905, + 824 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We attribute this to the complexity of automatic T2I, which requires multi-step reasoning. The quality of generated prompts directly impacts model selection, while model selection serves as a prerequisite for argument configuration. Previous studies [21, 25, 40] have shown that directly", + "bbox": [ + 511, + 825, + 905, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "13287", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/c815dc7614cd5792197e1fa79c8a33e6fa8f6c92bb255f77bd1a850ba35bee55.jpg", + "image_caption": [ + "Figure 3. Illustration of the framework for ChatGen-Base and ChatGen-Evo." + ], + "image_footnote": [], + "bbox": [ + 138, + 90, + 857, + 291 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "predicting answers often fails in multi-step reasoning tasks. Additionally, supervised fine-tuning tends to encourage the model to learn simple text mappings instead of developing skills needed for automation, leading to poor generalization.", + "bbox": [ + 89, + 321, + 483, + 383 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2.2 ChatGen-Evo", + "text_level": 1, + "bbox": [ + 89, + 398, + 241, + 414 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To address the limitations of ChatGen-Base, we propose ChatGen-Evo, which trains MLLM $M$ using a multi-stage evolution strategy. Instead of relying on final outcome supervision as in traditional fine-tuning, the multi-stage evolution strategy in ChatGen-Evo employs stage supervision. By providing more precise feedback at each stage, this approach gradually enables the MLLM to acquire the necessary capabilities for automated T2I. As shown in Figure 3, the training process includes three main stages:", + "bbox": [ + 89, + 422, + 483, + 559 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Stage 1: Prompt Writing via SFT. We first train the MLLM using SFT with pairs of freestyle inputs and high-quality prompts. Different from the objective in ChatGenbase, the current stage focuses on a more specific and simplified task instead:", + "bbox": [ + 89, + 561, + 483, + 638 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {s f t} ^ {\\text {s t a g e 1}} = - \\sum_ {t} \\log P _ {\\pi} \\left(p \\mid c ^ {\\prime}, * _ {< t}\\right). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 163, + 646, + 482, + 676 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here, $c'$ represents the freestyle input with a prompt prefix that clarifies the task being performed. This prefix helps preserve the MLLM's original capabilities, minimizing catastrophic forgetting [22]. Through this stage, the model learns how to rewrite inputs into effective prompts.", + "bbox": [ + 89, + 685, + 482, + 762 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Stage 2: Model Selection via ModelToken. We introduce the ModelToken strategy to equip the model with model selection capabilities without impacting the prompt-writing ability learned in Stage 1. Inspired by token learning [10, 15], ModelToken extends this approach by representing each candidate T2I model as a unique token within the MLLM's vocabulary. Specifically, the model tokens are parameterized as an embedding matrix $W_{\\mathcal{M}} \\in \\mathbb{R}^{|\\mathcal{M}| \\times d}$ and appended to the original word token matrix $W_{\\nu} \\in \\mathbb{R}^{|\\mathcal{V}| \\times d}$ .", + "bbox": [ + 89, + 763, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "ModelToken Training: during training, the user input $c$ and prompt $p$ are concatenated as a prefix, with the special model token $<\\text{Model}_i>$ appended as the ground truth for next-token prediction. The training objective is:", + "bbox": [ + 511, + 321, + 906, + 383 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL \\left(W _ {\\mathcal {M}}\\right) = - \\log P _ {\\pi} \\left(< \\text {M o d e l} _ {-} i > | c, p\\right). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 573, + 393, + 906, + 411 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Unlike typical next-token prediction training, the embedding matrix $W_{\\mathcal{M}}$ represents the only tunable parameters, significantly enhancing training efficiency. With a small parameter size, fewer training samples are required, improving performance in scenarios with limited data.", + "bbox": [ + 511, + 422, + 905, + 498 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Inference for Model Selection: Once the embedding matrix is trained, the inference process concatenates the model token and original word token, forming the new language modeling head of the MLLM. In this way, the MLLM predicts the next token with the following probability:", + "bbox": [ + 511, + 500, + 905, + 575 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nP _ {\\pi} (m \\mid c, p) = \\operatorname {s o f t m a x} \\left(\\left[ W _ {\\nu}; W _ {\\mathcal {M}} \\right] \\cdot h _ {i - 1}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 585, + 903, + 604 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where the operation $[;]$ denotes concatenation, and $h_{i-1} \\in \\mathbb{R}^d$ represents the last hidden state. Once a model token is predicted, the MLLM stops decoding, and the corresponding model $m$ is selected. Additionally, information such as the model's description and demonstrations is loaded for subsequent use. After Stage 2, the model not only retains its prompt-writing skills but also learns to select models.", + "bbox": [ + 511, + 608, + 905, + 715 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Stage 3: Argument Configuration via In-Context Learning. After the above two stages, we have obtained the prompt $p$ and model $m$ from the original user input $c$ . The model now needs to generate the appropriate argument configuration to complete the final generation. Due to the careful design of the previous stages, the model's in-context learning ability is maximally preserved. Therefore, we adopt a training-free approach: we guide the MLLM using in-context demonstrations of $m$ , the user's original input $c$ , and the rewritten prompt $p$ . The MLLM can follow the demonstration pattern to complete the parameter configuration for the current user request:", + "bbox": [ + 511, + 719, + 906, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "13288", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\na = M (c, p, D (m)), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 215, + 90, + 480, + 107 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $D(m)$ represents the set of demonstrations for model $m$ in the training dataset. Thanks to the prior acquisition of the relevant model and prompt in earlier stages, this approach frees up context space, enabling extensive demonstrations. Additionally, the train-free approach avoids interfering with the trained model from the previous two stages.", + "bbox": [ + 89, + 111, + 483, + 203 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiment", + "text_level": 1, + "bbox": [ + 89, + 215, + 215, + 233 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We conduct a comprehensive evaluation of various methods on the novel ChatGenBench. First, in section 4.2, we compare ChatGen-Evo with other baseline models, highlighting the efficacy and efficiency of our multi-stage evolution strategy. Next, in section 4.3, we perform extensive ablation studies, uncovering the impact of each step on the final results and providing valuable insights. Finally, we provide visualizations of the generated images in section 4.4.", + "bbox": [ + 89, + 242, + 483, + 363 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experimental Settings", + "text_level": 1, + "bbox": [ + 89, + 373, + 297, + 388 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Training Setups. We adopt InternVL2 [5] as the base MLLM, fully fine-tuning it for both ChatGen-Base and the first stage of ChatGen-Evo. In the second stage of ChatGen-Evo, all model parameters are frozen except for the ModelToken embeddings. We employ the AdamW optimizer with a learning rate of 4e-5 and a weight decay of 1.0 over 5 epochs, maintaining these settings consistently across training stages. All experiments are conducted on 8 A100 GPUs.", + "bbox": [ + 89, + 396, + 483, + 517 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Metrics for Step-wise Evaluation: Leveraging the comprehensive process data in ChatGenBench, we introduce the step-wise evaluation metrics to assess the distinct abilities of automatic T2I models in key stages:", + "bbox": [ + 89, + 525, + 483, + 585 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Prompt BERTScore: To assess prompt rewriting ability, we use BERTScore [48] to compare predicted prompts with high-quality, human-validated prompts. BERTScore leverages pre-trained contextual embeddings to match words in candidate and reference sentences. The metric ranges from 0 to 1, with 1 indicating highly similar meanings and 0 signifying complete dissimilarity.", + "- Selection Accuracy: We calculate the accuracy of model selection by comparing the predicted T2I model with the human-validated model.", + "- Argument Accuracy: We evaluate argument configuration by calculating the exact match accuracy between the predicted and human-validated arguments. The overall accuracy is obtained by averaging individual arguments." + ], + "bbox": [ + 89, + 588, + 482, + 800 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "It is important to note that obtaining the absolute ground truth across all stages is nearly impossible due to the infinite search space. Therefore, we use human-validated high-quality records as relative ground truth. While these are not absolute, the comprehensive evaluation of the large-scale benchmark is able to reflect the capabilities of automation models, as confirmed by experimental results.", + "bbox": [ + 89, + 801, + 483, + 909 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Image Quality Evaluation: We use HPS v2 [41] and ImageReward [42] metrics to assess the quality of generated images, reflecting alignment with human preferences. Additionally, we employ FID and CLIP Score to evaluate how well the generated images meet user requirements. FID [12] measures the distance between automatically generated images and human-validated high-quality images, while CLIP Score [11] calculates the similarity between the generated images and human-validated prompts. To provide an intuitive and comprehensive measure of image quality, we normalize and combine these four metrics into an aggregated score, Unified Metric [49]. Details of each metric and the calculation of Unified Metric are presented in the Appendix.", + "bbox": [ + 511, + 90, + 905, + 287 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baseline Without Fine-Tuning. We further establish a baseline that uses the default in-context learning capabilities of the MLLM for prompt rewriting, along with a single Stable Diffusion model and a fixed set of default parameters. This baseline helps emphasize the significance of prompt rewriting and multi-model utilization.", + "bbox": [ + 511, + 296, + 905, + 387 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Main Experiment", + "text_level": 1, + "bbox": [ + 511, + 397, + 687, + 412 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2.1 Quantitative Results", + "text_level": 1, + "bbox": [ + 511, + 420, + 710, + 434 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 3 presents the main quantitative results of ChatGenEvo compared to different baselines. Overall, ChatGenEvo significantly outperforms other methods across all metrics, including both step-wise and final image quality evaluations. Specifically, the low performance of the baseline highlights the importance of effective prompt rewriting and multi-model selection, underscoring the necessity of dedicated Automatic T2I methods. Additionally, fine-tuning MLLMs with progressively larger parameter scales yields steady performance improvements. Remarkably, ChatGenEvo achieves performance comparable to ChatGen-Base at 8B, despite utilizing a significantly smaller parameter scale.", + "bbox": [ + 511, + 444, + 905, + 626 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The comparisons across different settings provide further insights. Prompt rewriting ability demonstrates strong generalizability, effectively transferring to rarely seen models. The higher few-shot performance can be attributed to the bias in data distribution between the two settings, as reflected in the \"Baseline\" results, which reveal a difference of approximately 0.3. In contrast, model selection and parameter configuration show a noticeable decline in performance under few-shot conditions, highlighting their reliance on more training samples. Therefore, exploring the scaling laws of prompt rewriting and enhancing model selection in few-shot scenarios are promising directions.", + "bbox": [ + 511, + 627, + 905, + 808 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2.2 Human Evaluation", + "text_level": 1, + "bbox": [ + 511, + 816, + 700, + 830 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We conduct a user study using pairwise comparisons to further evaluate ChatGen-Base(8B) and ChatGen-Evo(2B). Users are presented with two images generated from the same input: one by ChatGen-Base and the other by", + "bbox": [ + 511, + 840, + 905, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "13289", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/829b109738723a5a2f239391dab4f3dae4bc4bbfea5791856b96fadb51dce82b.jpg", + "table_caption": [ + "Table 3. The Step-wise and Final evaluation results of different methods on ChatGenBench." + ], + "table_footnote": [], + "table_body": "
Step-wise EvaluationFinal Evaluation
PromptBertScore ↑SelectionAcc ↑ConfigAcc ↑FIDScore ↓CLIPScore ↑HPSv2 ↑ImageReward ↑UnifiedMetric ↑
SupervisedBaseline0.026--32.764.620.2-34.637.3
ChatGen-Base(2B)0.1840.2060.38421.369.923.52.459.0
ChatGen-Base(4B)0.1970.2300.49020.770.023.41.558.7
ChatGen-Base(8B)0.2080.2640.50920.870.723.94.060.7
ChatGen-Evo (2B)0.2470.3280.53719.172.925.18.965.9
Few-shotBaseline0.055--54.463.420.0-40.229.7
ChatGen-Base(2B)0.2210.1530.34942.869.123.3-4.851.1
ChatGen-Base(4B)0.2360.1710.44841.269.423.4-4.351.9
ChatGen-Base(8B)0.2520.2010.46241.470.623.4-3.152.5
ChatGen-Evo (2B)0.2830.2310.49340.772.525.05.159.2
", + "bbox": [ + 112, + 103, + 877, + 305 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/5a41d55bb7d1fcaa692c5fb4eae5d98a5b968c4e7562cd3e14fa8d6b98280c7d.jpg", + "table_caption": [ + "Table 4. Efficiency comparison. Training efficiency is measured in GPU hours, while inference is expressed as the seconds required to process each data. The training was conducted on 8 A100 GPUs, and inference was performed on a single A100 80GB GPU." + ], + "table_footnote": [], + "table_body": "
MethodParamsTrainingInference
ChatGen-Base(2B)2.21B76h1.1s
ChatGen-Base(8B)8.08B240h2.3s
ChatGen-Evo(2B)2.24B100h1.9s
", + "bbox": [ + 99, + 367, + 470, + 444 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ChatGen-Evo. It is tasked with selecting the image that better matches the image quality and relevance to the given input. We sample 2,000 image pairs for supervised and 1,000 for few-shot. More details are provided in the Appendix.", + "bbox": [ + 89, + 446, + 482, + 508 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As shown in Figure 4, the human evaluation results are consistent with the quantitative metrics, highlighting that ChatGen-Evo outperforms ChatGen-Base in both image quality and relevance. Additionally, in few-shot settings, ChatGen-Evo demonstrates a higher win rate, showcasing the effectiveness of our approach in data-scarce scenarios.", + "bbox": [ + 89, + 508, + 483, + 599 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2.3 Efficiency Comparison", + "text_level": 1, + "bbox": [ + 89, + 607, + 303, + 622 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Similar to existing multi-stage reasoning methods [25, 40, 46], ChatGen-Evo does not offer an efficiency advantage over direct prediction approaches like ChatGen-Base. However, the significant performance gains more than compensate for this drawback. As shown in Table 4, ChatGen-Evo achieves the performance level of ChatGen-Base at 8B parameters using only 2B parameters. Therefore, when com", + "bbox": [ + 89, + 632, + 482, + 737 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/fb50ae8208316390f900da2ba5d44dcd2b2f869a16ea75cfa84f657be5880e8a.jpg", + "image_caption": [ + "Figure 4. User study results of ChatGen-Base and ChatGen-Evo." + ], + "image_footnote": [], + "bbox": [ + 94, + 741, + 478, + 881 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/bb5b16447b2b1ed2577dccf011c37df621f8af42cc5ff99d20167f8165888191.jpg", + "table_caption": [ + "Table 5. Ablation experiments on the supervised setting." + ], + "table_footnote": [], + "table_body": "
StageMethodsStep Score ↑FID Score ↓Image Reward ↑Unified Metric ↑
Prompt WritingBaseline0.02619.3-13.150.8
Chat-Base0.18421.33.159.7
Chat-Evo0.24717.610.566.8
Model SelectionChat-Base0.20620.59.166.0
Chat-Evo0.55318.216.869.7
Argument SettingChat-Base0.38417.910.469.1
Chat-Evo0.87117.117.970.3
", + "bbox": [ + 516, + 325, + 897, + 479 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "paring efficiency under equal performance, ChatGen-Evo maintains a relative advantage over ChatGen-Base.", + "bbox": [ + 511, + 488, + 905, + 518 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Analysis", + "text_level": 1, + "bbox": [ + 511, + 531, + 614, + 547 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3.1 Capability Analysis.", + "text_level": 1, + "bbox": [ + 511, + 555, + 707, + 570 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We conduct ablation experiments on ChatGenBench to evaluate the contribution of individual steps to final performance by providing ground truth for the other steps.", + "bbox": [ + 511, + 580, + 905, + 626 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Prompt Writing: We first compare the prompt writing capabilities of different methods, along with their final performance. In these experiments, each method's predicted prompt was passed along with the correct model and argument. As shown in Table 5, all methods exhibit noticeable gaps compared to human-validated prompts, emphasizing the complexity of prompt rewriting. Comparisons with the results in Table 3 also highlight the further improvements achieved by selecting the correct model and arguments, particularly for the \"Baseline\" methods. Moreover, variations in prompts significantly influence the final results, highlighting its critical role in Automatic T2I.", + "bbox": [ + 511, + 628, + 905, + 808 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Model Selection: Table 5 also shows the impact of model selection where human-validated prompt and argument are provided. With well-crafted prompts, ChatGen-Evo demonstrates a substantial performance boost from $32.8\\%$ to $55.3\\%$ , indicating the strong influence of prompt quality on model selection accuracy. This supports our perspec", + "bbox": [ + 511, + 810, + 905, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "13290", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/85cefe667a402dce8fec1cd3fa25ae01cf17bd463dbeca1f804c429ec53d31e1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 140, + 95, + 166, + 114 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Let's craft a fierce anime warrior maiden, ablaze with power and passion!", + "bbox": [ + 168, + 97, + 334, + 114 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d1b7e35f4720dbfde592aeb9ac30469536e22126a27ab9789939dd7293ca57c4.jpg", + "image_caption": [ + "Default SD", + "ChatGen-Base" + ], + "image_footnote": [], + "bbox": [ + 140, + 116, + 205, + 190 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5b16fcf44ebd45aa84c978e53d7e1e357eccae4a2e2190570f20e3f876ffc7fe.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 207, + 117, + 272, + 190 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3561f8c5ca0c901fe0170e7fd2e59920dd18aa8a121bf0aee12f56541c989756.jpg", + "image_caption": [ + "ChatGen-Evo" + ], + "image_footnote": [], + "bbox": [ + 274, + 117, + 339, + 191 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6b887e4e0a5962454eae8930ee9c586e7439e1492f6a6583134c8e15cf44e381.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 349, + 95, + 369, + 114 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Imagine a fluffy cyberpunk critter in a detailed, artsy, and concept style. Keep the image perfect !", + "bbox": [ + 370, + 97, + 571, + 114 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/14c5d1b4a9d4406a2a256f68da248be33ea593b23667e6e8afba99074835a74d.jpg", + "image_caption": [ + "Default SD" + ], + "image_footnote": [], + "bbox": [ + 349, + 117, + 421, + 191 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/290d606325f79fc3357ac4a46105c4dc430f254157266ed492f69762cd438b33.jpg", + "image_caption": [ + "ChatGen-Base" + ], + "image_footnote": [], + "bbox": [ + 424, + 117, + 475, + 191 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/393eb531804b83a2ebe13550ec5b301f693698fb319682630bf59d312a3907b7.jpg", + "image_caption": [ + "ChatGen-Evo" + ], + "image_footnote": [], + "bbox": [ + 501, + 117, + 575, + 191 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9ff084a696b5424bb23e10dca752975dbbb1895804105ded2f21451ef2e0e129.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 602, + 95, + 625, + 114 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Do you reckon we could mix pirates and punk for a car design? Make Something stands out.", + "bbox": [ + 625, + 97, + 813, + 114 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0969957ea7d22bbc2d6ad6501d3a77884c078a1453da4b24d8959a7b875653c8.jpg", + "image_caption": [ + "Default SD" + ], + "image_footnote": [], + "bbox": [ + 583, + 117, + 656, + 191 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1396a873ef64093583c0666293ebdb0b195c600982175fcb49fef8d93e4f1b45.jpg", + "image_caption": [ + "ChatGen-Base" + ], + "image_footnote": [], + "bbox": [ + 660, + 117, + 756, + 191 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/af069c871ac878976299eb99f1bf14c02057221e57d58e23cb641281321357aa.jpg", + "image_caption": [ + "ChatGen-Evo" + ], + "image_footnote": [], + "bbox": [ + 761, + 117, + 856, + 191 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/cf795807719c2c106a1237bf9c663731e8d7f6a7c45cf9dfd2f86eb9da183d9a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 140, + 213, + 163, + 232 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "I want the final image echoes the reference's shitty, comic-inspired style?", + "bbox": [ + 166, + 214, + 326, + 232 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6d2091d230945f1bec1ad717d37c6d6fef650985b5db25f799525bd4f615ec46.jpg", + "image_caption": [ + "reference image" + ], + "image_footnote": [], + "bbox": [ + 140, + 234, + 200, + 282 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0ed95a1db6d50657d89ebbdf0aca7f177664f23d66b897391f776374c0d53186.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 202, + 234, + 263, + 299 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/38a64f804862b3279d45958a82ed66442eccd9eaacdd70fc7251433881dabfbf.jpg", + "image_caption": [ + "ChatGen-Base", + "ChatGen-Evo" + ], + "image_footnote": [], + "bbox": [ + 264, + 234, + 349, + 300 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1ba0d0a657deb9366dd654ce15de0324a33cf41f53c1e7eb6b345ec2ba5dee00.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 351, + 213, + 372, + 232 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Can we create the image similar to the original, with soft colors and intricate details?", + "bbox": [ + 374, + 214, + 566, + 232 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c486fc45431016f1573560bf876ad90347b5088c0d3682982561124a8485346f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 352, + 234, + 413, + 301 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/06f14b95b13cc32fcc019291d49cdd1466168f8aa38e67496e8e0f1d043856fd.jpg", + "image_caption": [ + "ChatGen-Base" + ], + "image_footnote": [], + "bbox": [ + 424, + 234, + 509, + 301 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5729c35dbffc8d22ac6f74322938166944cff611e72cb2816b3b5c2a92b5eaf2.jpg", + "image_caption": [ + "ChatGen-Evo" + ], + "image_footnote": [], + "bbox": [ + 516, + 234, + 602, + 301 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b13d3281eab8b69afdbc068dcc37f9a8f2cceef5654ad67a8b8cfb203ce1a59b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 620, + 213, + 643, + 232 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Recreate the intricate patterns of the reference, blending realism and dreamlike qualities?", + "bbox": [ + 645, + 215, + 836, + 232 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f4e7b527ed0afaf40dc7d9d3acf697b920fefdc29be7d50983bdd2187a6db0e2.jpg", + "image_caption": [ + "reference image" + ], + "image_footnote": [], + "bbox": [ + 622, + 234, + 681, + 284 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/58a91af8787f1b1903dadad9fea8123decbcd558f6fdea4763c8443f32650fdb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 691, + 234, + 763, + 311 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8caf608f5a4cb0adad4c1891ad62c2d20a894940a0a8e071ae71dd63c88bfee3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 767, + 234, + 844, + 311 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4a783e720cbb2bd08ce6a2275ebcb8d1b2f30f9cb6a2c6e3ab9432248166bb13.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 143, + 323, + 163, + 342 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Hey, you know Flandre Scarlet? Create an image for her!", + "bbox": [ + 163, + 323, + 258, + 349 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/bd70a3ded7549128c9bed057a457e0e4ba9a0038bbcbed63bb561390cf761ff7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 143, + 364, + 163, + 383 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Picture her having a blast in a creek, all cute and nature!", + "bbox": [ + 163, + 363, + 258, + 390 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a84e8b0ec1b78686ecb1b44cc5c820d14374315e5be4b316d764edfe19e541d5.jpg", + "image_caption": [ + "Round 1", + "Figure 5. Examples of images generated by different methods. Three rows represent single, multi-modal and historical inputs, respectively." + ], + "image_footnote": [], + "bbox": [ + 271, + 321, + 357, + 388 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6045c2b6cd86a2f94f4cfde6ce2d3e35c0a84fe8c60e349dd95583b76c58e919.jpg", + "image_caption": [ + "Round 2" + ], + "image_footnote": [], + "bbox": [ + 359, + 321, + 473, + 388 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/17715eb3d1aca0dea58a4544b7068cb3ce25fee9b0911b8f2f72b157621c0f60.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 480, + 321, + 501, + 340 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Creat Van Gogh painting of a human-like possum", + "bbox": [ + 501, + 323, + 602, + 340 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/43ec982290053fb206dab9a247b870ac6b258321432c84ed77c029c40f47cbdd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 480, + 349, + 501, + 359 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "It's holding a fruit basket.", + "bbox": [ + 501, + 351, + 606, + 361 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1b59fc1c3840f32a6404b8c5ca1c9d10874d2badbc841edaace16a5ea8e6951c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 480, + 373, + 501, + 382 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "With vibrant yellows, deep blue colors for fruits!", + "bbox": [ + 501, + 373, + 611, + 390 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d41c3d132477d36efd47e16ca0ab03bb8dca4f28012bd3e2f0929fa462f7b278.jpg", + "image_caption": [ + "Round 1" + ], + "image_footnote": [], + "bbox": [ + 627, + 316, + 699, + 388 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b12bbf252b8ff64368442f406a5e0ab4b6f7673e7c05dd4ae453fe6254374439.jpg", + "image_caption": [ + "Round 2" + ], + "image_footnote": [], + "bbox": [ + 700, + 316, + 772, + 388 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c293acfb79e559807a2082a70c3b07fe1c0338c865dbfbd22ae987fd03a0c253.jpg", + "image_caption": [ + "Round 3" + ], + "image_footnote": [], + "bbox": [ + 769, + 316, + 851, + 388 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "tive that Automatic T2I fundamentally involves multi-step reasoning. In contrast, ChatGen-Base, which directly predicts all results, fails to adapt and thus produces unchanged outputs. Furthermore, these results also emphasize the substantial impact of model selection on overall performance.", + "bbox": [ + 89, + 421, + 482, + 496 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Argument Configuration: When provided with high-quality prompts and appropriate model selection, ChatGenEvo exhibits notable performance improvements. It improves configuration accuracy from $53.7\\%$ to $87.1\\%$ and Unified Score from 65.9 to 70.3. Overall, while argument configuration has a relatively smaller impact compared to previous stages, it remains a crucial component.", + "bbox": [ + 89, + 497, + 482, + 602 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The above findings suggest that outcomes in earlier steps significantly influence predictions in subsequent ones. Therefore, exploring more reasoning methods for advancing automatic T2I represents a promising research direction.", + "bbox": [ + 89, + 603, + 482, + 662 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3.2 Input Type Analysis.", + "text_level": 1, + "bbox": [ + 89, + 678, + 287, + 694 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 6 presents ChatGen-Evo performance across different input types. Multimodal inputs lead to better performance, as images may offer clearer prompt and model identification cues compared to text alone. Additionally, handling historical data remains a significant challenge, with the lowest performance across all metrics. These results point to valuable directions in enhancing history-based prompt generation.", + "bbox": [ + 89, + 702, + 482, + 808 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Visualizations", + "text_level": 1, + "bbox": [ + 89, + 816, + 233, + 830 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Figure 5 presents examples of images generated by different methods. In the first row, ChatGen-Evo understands user requirements and identifies suitable models to generate style-matching images. The second row demonstrates", + "bbox": [ + 89, + 839, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/70d6a33477080d81ee9a99524c5ae0c4a9b07ccc5ece6234c3a0410e38c2092a.jpg", + "table_caption": [ + "Table 6. The evaluation results of different input types." + ], + "table_footnote": [], + "table_body": "
Input TypeStep-wise EvaluationFinal
Prompt BertScore ↑Selection Acc ↑Config Acc ↑Unified Score ↑
Single0.2520.3310.53968.1
M-Modal0.2770.3810.55469.3
History0.1650.2590.50560.1
", + "bbox": [ + 522, + 436, + 890, + 534 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "results based on multi-modal user inputs, where ChatGen-Evo shows a superior understanding of the reference image and preserves more details to generate refined outputs. The third row illustrates ChatGen-Evo's capability in handling historical data, ensuring that each step inherits the previous style while making appropriate modifications.", + "bbox": [ + 511, + 547, + 905, + 638 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 660, + 632, + 676 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our research aims to automate tedious steps in T2I generation, allowing users to simply describe their needs in a freestyle chatting way. We introduce ChatGenBench for benchmarking Automatic T2I task. It includes high-quality paired data with diverse freestyle user inputs, enabling comprehensive evaluation across all steps. Furthermore, we argue that Automatic T2I should be regarded as a multi-step reasoning task. Consequently, we propose ChatGen-Evo, a multi-stage evolution strategy that progressively equips models with essential automation skills. Extensive evaluations not only demonstrate the superiority of ChatGen-Evo but also provide valuable insights for advancing Automatic T2I. We believe this represents a significant step toward the future of automated generative models.", + "bbox": [ + 511, + 689, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "13291", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 91, + 90, + 250, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported by the National Nature Science Foundation of China (No. 62192781, No. 62272374), the Natural Science Foundation of Shaanxi Province (2024JC-JCQN-62), the National Nature Science Foundation of China (No. 62202367, No. 62250009, No. 62137002), the Key Research and Development Project in Shaanxi Province No. 2022GXLH-01-03, Project of China Knowledge Center for Engineering Science and Technology, and Project of Chinese academy of engineering \"The Online and Offline Mixed Educational Service System for 'The Belt and Road' Training in MOOC China\", and the K. C. Wong Education Foundation.", + "bbox": [ + 89, + 114, + 485, + 297 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 309, + 187, + 325 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] James Betker, Gabriel Goh, Li Jing, Tim Brooks, Jianfeng Wang, Linjie Li, Long Ouyang, Juntang Zhuang, Joyce Lee, Yufei Guo, Wesam Manassra, Prafulla Dhariwal, Casey Chu, and Yunxin Jiao. Improving image generation with better captions. 2023. 1, 2", + "[2] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18392-18402, 2023. 2", + "[3] Tom B Brown. Language models are few-shot learners. arXiv preprint arXiv:2005.14165, 2020. 2", + "[4] Tingfeng Cao, Chengyu Wang, Bingyan Liu, Ziheng Wu, Jinhui Zhu, and Jun Huang. Beautifulprompt: Towards automatic prompt engineering for text-to-image synthesis. arXiv preprint arXiv:2311.06752, 2023. 2, 4", + "[5] Zhe Chen, Weiyun Wang, Hao Tian, Shenglong Ye, Zhangwei Gao, Erfei Cui, Wenwen Tong, Kongzhi Hu, Jiapeng Luo, Zheng Ma, et al. How far are we to gpt-4v? closing the gap to commercial multimodal models with open-source suites. arXiv preprint arXiv:2404.16821, 2024. 6", + "[6] Civitai. Civitai. https://civitai.com/, 2022.1,3", + "[7] Runpei Dong, Chunrui Han, Yang Peng, Zekun Qi, Zheng Ge, Jinrong Yang, Liang Zhao, Jianjian Sun, Hongyu Zhou, Haoran Wei, et al. Dreamllm: Synergistic multimodal comprehension and creation. arXiv preprint arXiv:2309.11499, 2023. 2", + "[8] Hugging Face. Hugging face. https://huggingface.co/, 2016.1", + "[9] Yutong Feng, Biao Gong, Di Chen, Yujun Shen, Yu Liu, and Jingren Zhou. Ranni: Taming text-to-image diffusion for accurate instruction following. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4744-4753, 2024. 2", + "[10] Shibo Hao, Tianyang Liu, Zhen Wang, and Zhiting Hu. Toolkengpt: Augmenting frozen language models with massive tools via tool embeddings. Advances in neural information processing systems, 36, 2024. 5", + "[11] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation met" + ], + "bbox": [ + 93, + 335, + 482, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ric for image captioning. In Empirical Methods in Natural Language Processing (EMNLP), 2021. 6", + "[12] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 6", + "[13] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2", + "[14] Chengyou Jia, Minnan Luo, Zhuohang Dang, Guang Dai, Xiaojun Chang, Mengmeng Wang, and Jingdong Wang. Ssmg: Spatial-semantic map guided diffusion model for free-form layout-to-image generation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2480–2488, 2024. 2", + "[15] Chengyou Jia, Minnan Luo, Zhuohang Dang, Qiushi Sun, Fangzhi Xu, Junlin Hu, Tianbao Xie, and Zhiyong Wu. Agentstore: Scalable integration of heterogeneous agents as specialized generalist computer assistant. arXiv preprint arXiv:2410.18603, 2024. 5", + "[16] Bahjat Kawar, Shiran Zada, Oran Lang, Omer Tov, Huiwen Chang, Tali Dekel, Inbar Mosseri, and Michal Irani. Imagc: Text-based real image editing with diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6007-6017, 2023. 2", + "[17] Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization of text-to-image diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1931-1941, 2023. 2", + "[18] Dongxu Li, Junnan Li, and Steven Hoi. Blip-diffusion: Pretrained subject representation for controllable text-to-image generation and editing. Advances in Neural Information Processing Systems, 36, 2024. 2", + "[19] Yuheng Li, Haotian Liu, Qingyang Wu, Fangzhou Mu, Jianwei Yang, Jianfeng Gao, Chunyuan Li, and Yong Jae Lee. Gligen: Open-set grounded text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22511-22521, 2023. 2", + "[20] Long Lian, Boyi Li, Adam Yala, and Trevor Darrell. Llm-grounded diffusion: Enhancing prompt understanding of text-to-image diffusion models with large language models. arXiv preprint arXiv:2305.13655, 2023. 2", + "[21] Hunter Lightman, Vineet Kosaraju, Yura Burda, Harri Edwards, Bowen Baker, Teddy Lee, Jan Leike, John Schulman, Ilya Sutskever, and Karl Cobbe. Let's verify step by step. arXiv preprint arXiv:2305.20050, 2023. 4", + "[22] Yun Luo, Zhen Yang, Fandong Meng, Yafu Li, Jie Zhou, and Yue Zhang. An empirical study of catastrophic forgetting in large language models during continual fine-tuning. arXiv preprint arXiv:2308.08747, 2023. 5", + "[23] Alexander Quinn Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. In International Conference on Machine Learning (ICML), 2022. 1, 2" + ], + "bbox": [ + 516, + 92, + 906, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "13292", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[24] OpenAI. Chatgpt, 2022. Large language model. 2", + "[25] OpenAI. Learning to reason with llms. https://openai.com/index/learning-to-reason-with-llms/, 2024. Accessed: 2024-11-06. 2, 4, 7", + "[26] Shishir G Patil, Tianjun Zhang, Xin Wang, and Joseph E Gonzalez. Gorilla: Large language model connected with massive apis. arXiv preprint arXiv:2305.15334, 2023. 3", + "[27] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4195-4205, 2023. 2", + "[28] Jie Qin, Jie Wu, Weifeng Chen, Yuxi Ren, Huixia Li, Hefeng Wu, Xuefeng Xiao, Rui Wang, and Shilei Wen. Diffusiongpt: Llm-driven text-to-image generation system. arXiv preprint arXiv:2401.10061, 2024. 3, 4", + "[29] Yujia Qin, Shihao Liang, Yining Ye, Kunlun Zhu, Lan Yan, Yaxi Lu, Yankai Lin, Xin Cong, Xiangru Tang, Bill Qian, et al. Toolllm: Facilitating large language models to master 16000+ real-world apis. arXiv preprint arXiv:2307.16789, 2023. 3", + "[30] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2", + "[31] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 1, 2", + "[32] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 2", + "[33] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 22500-22510, 2023. 2", + "[34] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 2", + "[35] SpaCy. spacy: Industrial-strength natural language processing in python. https://spacy.io/, 2015.4", + "[36] Omost Team. Omost github page, 2024. 2", + "[37] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 2", + "[38] Jiapeng Wang, Chengyu Wang, Tingfeng Cao, Jun Huang, and Lianwen Jin. Diffchat: Learning to chat with text-to-image synthesis models for interactive image creation. arXiv preprint arXiv:2403.04997, 2024. 2, 4" + ], + "bbox": [ + 91, + 90, + 483, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[39] Zhenyu Wang, Aoxue Li, Zhenguo Li, and Xihui Liu. Genartist: Multimodal llm as an agent for unified image generation and editing. arXiv preprint arXiv:2407.05600, 2024. 3", + "[40] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824-24837, 2022. 4, 7", + "[41] Xiaoshi Wu, Yiming Hao, Keqiang Sun, Yixiong Chen, Feng Zhu, Rui Zhao, and Hongsheng Li. Human preference score v2: A solid benchmark for evaluating human preferences of text-to-image synthesis. arXiv preprint arXiv:2306.09341, 2023. 6", + "[42] Jiazheng Xu, Xiao Liu, Yuchen Wu, Yuxuan Tong, Qinkai Li, Ming Ding, Jie Tang, and Yuxiao Dong. Imagereward: Learning and evaluating human preferences for text-to-image generation. arXiv preprint arXiv:2304.05977, 2023.6", + "[43] Ling Yang, Zhaochen Yu, Chenlin Meng, Minkai Xu, Stefano Ermon, and CUI Bin. Mastering text-to-image diffusion: Recaptioning, planning, and generating with multimodal llms. In *Forty-first International Conference on Machine Learning*, 2024. 2", + "[44] Rui Yang, Lin Song, Yanwei Li, Sijie Zhao, Yixiao Ge, Xiu Li, and Ying Shan. Gpt4tools: Teaching large language model to use tools via self-instruction. Advances in Neural Information Processing Systems, 36, 2024. 3", + "[45] Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, and Yuan Cao. React: Synergizing reasoning and acting in language models. arXiv preprint arXiv:2210.03629, 2022. 2", + "[46] Shunyu Yao, Dian Yu, Jeffrey Zhao, Izhak Shafran, Tom Griffiths, Yuan Cao, and Karthik Narasimhan. Tree of thoughts: Deliberate problem solving with large language models. Advances in Neural Information Processing Systems, 36, 2024. 7", + "[47] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 2", + "[48] Tianyi Zhang, Varsha Kishore, Felix Wu, Kilian Q Weinberger, and Yoav Artzi. Bertscore: Evaluating text generation with bert. In International Conference on Learning Representations. 3, 6", + "[49] Lirui Zhao, Yue Yang, Kaipeng Zhang, Wenqi Shao, Yuxin Zhang, Yu Qiao, Ping Luo, and Rongrong Ji. Diffagent: Fast and accurate text-to-image api selection with large language model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6390-6399, 2024. 2, 3, 4, 6" + ], + "bbox": [ + 516, + 92, + 903, + 810 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "13293", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/ChatGen_ Automatic Text-to-Image Generation From FreeStyle Chatting/0c8e9c4a-231c-4c6d-af84-cd3eb4170579_model.json b/2025/ChatGen_ Automatic Text-to-Image Generation From FreeStyle Chatting/0c8e9c4a-231c-4c6d-af84-cd3eb4170579_model.json new file mode 100644 index 0000000000000000000000000000000000000000..2f8e32320f98adfed61fcfdd41d947ff2783b01c --- /dev/null +++ b/2025/ChatGen_ Automatic Text-to-Image Generation From FreeStyle Chatting/0c8e9c4a-231c-4c6d-af84-cd3eb4170579_model.json @@ -0,0 +1,3047 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "image", + "bbox": [ + 0.113, + 0.124, + 0.151, + 0.16 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.151, + 0.141, + 0.886, + 0.164 + ], + "angle": 0, + "content": "ChatGen: Automatic Text-to-Image Generation From FreeStyle Chatting" + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.19, + 0.895, + 0.245 + ], + "angle": 0, + "content": "Chengyou Jia\\(^{1*}\\), Changliang Xia\\(^{1*}\\), Zhuohang Dang\\(^{1}\\), Weijia Wu\\(^{2}\\), Hangwei Qian\\(^{3}\\), Minnan Luo\\(^{1\\dagger}\\) \n\\(^{1}\\)School of Computer Science and Technology, MOEKLINNS Lab, Xi'an Jiaotong University \n\\(^{2}\\)National University of Singapore \\(^{3}\\)CFAR, A\\*STAR" + }, + { + "type": "text", + "bbox": [ + 0.157, + 0.248, + 0.837, + 0.28 + ], + "angle": 0, + "content": "cp3jia@stu.xjtu.edu.cn, 202066@stu.xjtu.edu.cn, dangzhuohang@stu.xjtu.edu.cn, weijiawu96@gmail.com, qian0045@e.ntu.edu.sg, minnluo@xjtu.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.314, + 0.327, + 0.33 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.348, + 0.485, + 0.695 + ], + "angle": 0, + "content": "Despite the significant advancements in text-to-image (T2I) generative models, users often face a trial-and-error challenge in practical scenarios. This challenge arises from the complexity and uncertainty of tedious steps such as crafting suitable prompts, selecting appropriate models, and configuring specific arguments, making users resort to labor-intensive attempts for desired images. This paper proposes Automatic T2I generation, which aims to automate these tedious steps, allowing users to simply describe their needs in a freestyle chatting way. To systematically study this problem, we first introduce ChatGenBench, a novel benchmark designed for Automatic T2I. It features high-quality paired data with diverse freestyle inputs, enabling comprehensive evaluation of automatic T2I models across all steps. Additionally, recognizing Automatic T2I as a complex multi-step reasoning task, we propose ChatGenEvo, a multi-stage evolution strategy that progressively equips models with essential automation skills. Through extensive evaluation across step-wise accuracy and image quality, ChatGen-Evo significantly enhances performance over various baselines. Our evaluation also uncovers valuable insights for advancing automatic T2I. Our data, code, and models will be publicly available." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.716, + 0.222, + 0.731 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.742, + 0.484, + 0.864 + ], + "angle": 0, + "content": "In recent years, text-to-image (T2I) generative models have attracted considerable attention [1, 23, 31, 32]. Building on advancements in large-scale T2I models such as DALL-E [31] and Stable Diffusion [32], the open-source community has significantly expanded the capabilities of T2I generation. Researchers fine-tune open-source models on specialized datasets, resulting in a diverse selection of task-specific models available on platforms like Civitai [6] and Hugging" + }, + { + "type": "image", + "bbox": [ + 0.536, + 0.318, + 0.886, + 0.527 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.536, + 0.53, + 0.886, + 0.639 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.643, + 0.907, + 0.685 + ], + "angle": 0, + "content": "Figure 1. Illustration of tedious steps in T2I. Our method can select an appropriate model with suitable prompts and arguments. Note: Arle is a character from the game Genshin Impact." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.687, + 0.907, + 0.734 + ], + "angle": 0, + "content": "Face [8]. This variety provides users with a broader range of options to meet customization needs, facilitating the growing adoption of T2I models in real-world applications." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.735, + 0.909, + 0.901 + ], + "angle": 0, + "content": "However, the rapid development of T2I models within the open-source community has also introduced significant challenges for users. When non-experts attempt to create images with specific requirements, they often encounter a trial-and-error process involving several tedious steps. As shown in Figure 1, these steps include crafting suitable prompts, selecting appropriate models, and configuring specific model arguments. The complexity and uncertainty of each step turn the process into an arduous journey, resembling \"mice in a maze\". In real-world scenarios, this iterative process consumes substantial time and resources" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.875, + 0.223, + 0.888 + ], + "angle": 0, + "content": "*Equal Contribution." + }, + { + "type": "page_footnote", + "bbox": [ + 0.111, + 0.888, + 0.236, + 0.9 + ], + "angle": 0, + "content": "† Corresponding author." + }, + { + "type": "list", + "bbox": [ + 0.109, + 0.875, + 0.236, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13284" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.168 + ], + "angle": 0, + "content": "as users continuously adjust settings to regenerate images. Therefore, we pose the challenging problem: can we automate these labor-intensive steps in T2I generation, allowing users to simply describe their needs in a chatting style and receive desired images effortlessly?" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.17, + 0.482, + 0.351 + ], + "angle": 0, + "content": "Previous attempts, such as BeautifulPrompt [4, 38] for generating high-quality prompts from low-quality ones and DiffAgent [49] for model selection using large language models, have made some progress in addressing these challenges. However, these methods only seek to automate a specific step in Figure 1, lacking comprehensive research on the automation of entire T2I process. Moreover, they fail to support diverse types of freestyle input, leaving them far from real-world scenarios. To bridge these gaps, we aim to develop the model that can accept arbitrary user input, similar to ChatGPT [24], and automatically generate all necessary components for generation, termed as Automatic T2I." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.353, + 0.483, + 0.534 + ], + "angle": 0, + "content": "To this end, we first introduce ChatGenBench, the benchmark specifically designed for this task. ChatGenBench offers a substantial dataset of high-quality paired data from 6,807 customized models. Each data pair comprises a user's freestyle chat input, a refined prompt, an appropriate model, and optimal arguments. This comprehensive step-by-step trail enables step-wise evaluation of automatic T2I models, ensuring both quality assessment of the final image output and precise identification of potential automation bottlenecks. Furthermore, ChatGenBench integrates various types of data, enabling testing with multimodal and historical input to simulate practical scenarios." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.537, + 0.483, + 0.824 + ], + "angle": 0, + "content": "An intuitive approach to achieve the goal is to build supervised fine-tuning (SFT) data and tune multimodal large language models. However, we argue that automated T2I should be viewed as a complex multi-step reasoning problem. Directly predicting outputs leads the model to focus on simple text mappings, rather than developing the diverse skills needed for robust automation. Inspired by OpenAI's o1 [25], we propose ChatGen-Evo with a multi-stage evolution strategy. This approach provides targeted feedback at each stage, progressively equipping models with essential automation skills. First, we construct SFT data to train the model to generate high-quality prompts from freestyle inputs. Next, we augment the model's vocabulary with specialized ModelTokens, enabling effective model selection without affecting other functions. Finally, we guide the model in configuring arguments based on prompts and models selected in previous stages. By decomposing the task into clear stages, the model gains crucial automation skills, delivering outputs aligned with user expectations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.483, + 0.901 + ], + "angle": 0, + "content": "We conduct a comprehensive evaluation of various methods on the novel ChatGenBench to study Automatic T2I. ChatGen-Evo significantly outperforms other methods across all metrics, including both step-wise assessments and image quality evaluations. Extensive quantitative and" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.243 + ], + "angle": 0, + "content": "qualitative results, along with ablation studies, underscore the critical importance of the multi-stage evolution strategy. Additionally, further experiments reveal the impact of different stages on the final T2I output, uncovering valuable insights into the challenges and opportunities in Automatic T2I. We highlight some promising directions, such as exploring the scaling laws of prompt rewriting, improving model selection in few-shot scenarios, and leveraging reasoning approaches to enhance performance. The contributions of this paper are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.246, + 0.905, + 0.306 + ], + "angle": 0, + "content": "- We propose the novel and challenging problem of Automatic T2I generation, aiming to develop the model that can handle user's freestyle chatting and automatically produce all necessary components for image generation." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.307, + 0.905, + 0.367 + ], + "angle": 0, + "content": "- We introduce ChatGenBench, a benchmark specifically designed for Automatic T2I. It includes a comprehensive step-by-step trail for step-wise evaluation, supporting multimodal or historical user inputs." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.367, + 0.905, + 0.426 + ], + "angle": 0, + "content": "- We present ChatGen-Evo, which adopts a multi-stage evolution strategy. By decomposing the task into distinct stages, ChatGen-Evo enables the model to progressively acquire essential Automatic T2I capabilities." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.428, + 0.905, + 0.472 + ], + "angle": 0, + "content": "- Extensive experiments on ChatGenBench not only demonstrate the superiority of ChatGen-Evo but also provide valuable insights for advancing automatic T2I." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.246, + 0.905, + 0.472 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.487, + 0.663, + 0.502 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.512, + 0.749, + 0.527 + ], + "angle": 0, + "content": "2.1. Text-to-Image Generation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.535, + 0.905, + 0.763 + ], + "angle": 0, + "content": "With the advancement of diffusion models [13, 27, 34], text-to-image (T2I) generation [1, 23, 31] has demonstrated exceptional capabilities in high-quality image generation and textual description alignment. Large-scale models such as DALL-E [1] enhance text-image alignment by leveraging the joint feature space of CLIP [30]. Moreover, Stable Diffusion [32], a well-established open-source model, has gained substantial attention. Numerous methods have been developed to fine-tune it or design additional modules for specialized tasks, such as customized image generation [17, 18, 33], layout-to-image generation [14, 19, 20, 47] and image edit [2, 16]. These diverse models have significantly expanded the capabilities of T2I. However, they also present significant learning challenges for non-expert users, underscoring the increasing need for automatic T2I generation." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.772, + 0.829, + 0.787 + ], + "angle": 0, + "content": "2.2. LLMs for Text-to-Image Generation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Large language models (LLMs), like ChatGPT [24], Llama [37], have demonstrated impressive capability in language understanding [3] and problem solving [45]. Recently, LLMs have also been applied to image generation. [7, 9, 20, 36, 43] leverage LLMs to generate detailed layout information from complex prompts, enabling the creation of sub-elements and control over their positioning." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.518, + 0.956 + ], + "angle": 0, + "content": "13285" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.135, + 0.089, + 0.861, + 0.209 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.343, + 0.214, + 0.653, + 0.228 + ], + "angle": 0, + "content": "Figure 2. Illustration of the data collection pipeline." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.233, + 0.484, + 0.325 + ], + "angle": 0, + "content": "[28, 39, 49] propose employ LLMs for model selection, akin to tool usage [26, 29, 44]. However, the above methods still require specialized prompt inputs and involve complex model usage. Currently, no approach thoroughly considers leveraging LLMs to relieve users from all tedious steps in T2I. Our method aims to address this gap." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.339, + 0.226, + 0.357 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.365, + 0.484, + 0.516 + ], + "angle": 0, + "content": "Our goal is to relieve users from tedious steps and automate to produce the desired images from user's freestyle input. To achieve this objective, we first introduce ChatGenBench, a benchmark comprising a large set of user inputs in a chatting style, built upon a foundation of over 6000 personalized models to evaluate automated image generation results. A comprehensive description of data collection, construction, and analysis is provided in Sec. 3.1. Then, we present ChatGen-Evo in Sec. 3.2, which uses a multi-stage evolution strategy to train MLLM for Automatic T2I." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.528, + 0.484, + 0.544 + ], + "angle": 0, + "content": "3.1. ChatGenBench: Benchmarking Automatic T2I" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.552, + 0.484, + 0.629 + ], + "angle": 0, + "content": "For clarity, an example of this data is shown on the right side of Figure 2. The following sections detail the data collection and construction process, which primarily involves High-Quality Human Demonstration Collection and using LLM-driven Role-Play to simulate user input." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.649, + 0.478, + 0.664 + ], + "angle": 0, + "content": "3.1.1 High-Quality Human Demonstration Collection" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.673, + 0.483, + 0.764 + ], + "angle": 0, + "content": "Civitai [6] is a vibrant community where users share customized AI models for generating high-quality images. Members contribute demonstrations with detailed prompts, model specifications, and arguments, supported by a feedback system that ensures quality validation. These features make Civitai an ideal platform for collecting raw data." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.856 + ], + "angle": 0, + "content": "Collection: We start by collecting demonstrations based on established evaluation metrics within Civitai, including download counts, upvotes, and other user feedback. These metrics enable us to collect data that has been validated through community engagement. By focusing on these indicators, we identify a subset of high-quality results." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Filtering: Following the initial collection, we implement a rigorous filtering process to ensure data quality. This involves excluding demonstrations associated with inactive" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.233, + 0.906, + 0.31 + ], + "angle": 0, + "content": "or outdated models, removing duplicates, and filtering out NSFW content. This careful curation refined the dataset to include the most effective demonstrations. Ultimately, this process results in a curated set of 44,881 high-quality human demonstrations across 6,807 unique models." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.333, + 0.903, + 0.349 + ], + "angle": 0, + "content": "3.1.2 LLM-Driven Role-Play for Chatting Generation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.358, + 0.907, + 0.585 + ], + "angle": 0, + "content": "While the demonstrations collected from communities include the essential procedural information needed for automation, they lack a critical element: freestyle chatting inputs. Such data is not available on open platforms, which has been a key limitation for previous methods [49]. To address this, we propose an LLM-Driven Role-Play strategy for Chatting Generation. As shown in Figure 2, we predefine over 100 roles from everyday life (e.g., students, doctors, professors) and prompt the LLM to simulate these roles, translating each demonstration into freestyle chatting input with tones and habits of the character. This approach significantly enhances data diversity and creates more lifelike inputs. To further enhance reverse synthesis diversity, we incorporate multiple LLMs, adjust diversity arguments, and use BERTScore [48] to filter out overly similar inputs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.587, + 0.906, + 0.723 + ], + "angle": 0, + "content": "Additionally, we define three types of freestyle input formats: single-input, consisting of a single chatting-style sentence; multimodal-input, combining a sentence with an image; and history-input, comprising multiple rounds of dialogue history. These formats effectively simulate how users typically inquire about image generation needs, greatly expanding the practical value of automatic T2I. An example of the single-input prompt is shown in the following. Additional prompts and details are provided in the Appendix." + }, + { + "type": "title", + "bbox": [ + 0.525, + 0.737, + 0.785, + 0.753 + ], + "angle": 0, + "content": "Prompt: Sing-Input chatting generation" + }, + { + "type": "text", + "bbox": [ + 0.522, + 0.763, + 0.897, + 0.847 + ], + "angle": 0, + "content": "System: You are a professional user experience designer who plays various personas to convert complex and professional content for non-professional users. Please merge the following prompt and model information into a single freestyle query. Remove any obvious details that non-professional users would avoid. Make it similar to what non-professional users may write. The converted single-text query should be colloquial and as brief as possible." + }, + { + "type": "text", + "bbox": [ + 0.524, + 0.847, + 0.895, + 0.896 + ], + "angle": 0, + "content": "Require: {ROLE}, {PROMPT}, {MODEL}, {Example 1,...,n} \nInput: You are playing the {ROLE}. Please generate a single text query based on the following professional {PROMPT} for the {MODEL}. You can refer to following examples: {Example 1,...,n}." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13286" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.212, + 0.09, + 0.784, + 0.103 + ], + "angle": 0, + "content": "Table 1. Comparison of different methods for benchmarking automatic text-to-image generation." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.104, + 0.912, + 0.229 + ], + "angle": 0, + "content": "
TrainSet (# Numbers)TestSet (# Numbers)Based Models?Step-wise Evaluation?Freestyle User-input?Multi-modal Support?History Support?
BeautifulPrompt [4]143K2KSingleX✓XX
DiffChat [38]234K5KSingleX✓X✓
DiffusionGPT [28]--Multi(≈20)X✓XX
DABench [49]40K5KMulti(5K)XXXX
ChatGenBench256K14KMulti(6K)✓✓✓✓
" + }, + { + "type": "table_caption", + "bbox": [ + 0.171, + 0.233, + 0.402, + 0.247 + ], + "angle": 0, + "content": "Table 2. Summary of dataset statistics." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.25, + 0.487, + 0.372 + ], + "angle": 0, + "content": "
DatasetTotalSingleM-ModalHistory
TrainSet256,606147,88869,54839,170
TestSet Init74,36442,83820,21411,312
Benchmark14,56411,0111,6681,132
Supervised10,2408,0091,0991,132
Few-Shot4,3243,002569753
" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.377, + 0.323, + 0.391 + ], + "angle": 0, + "content": "3.1.3 Benchmark Construction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.401, + 0.483, + 0.492 + ], + "angle": 0, + "content": "Following the above steps, we generate 330,970 freestyle inputs from 44,881 demonstrations. Considering the largescale generated data, we don't randomly split a test set for benchmarking. Instead, we carefully select high-quality, non-overlapping samples as the benchmark for evaluation. The selection involves the following steps:" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.492, + 0.484, + 0.568 + ], + "angle": 0, + "content": "TestSet Split: We split the data based on model origin. For data associated with the same model, we use BERTScore to assess the similarity between samples. From this, we select about \\(20\\%\\) most semantically distinct data as initial TestSet in Table 2. This maximizes non-overlap with TrainSet." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.568, + 0.484, + 0.614 + ], + "angle": 0, + "content": "TestSet Filtering: Building on the initial TestSet, we perform multiple filtering rounds to ensure the final ChatGen-Bench's quality. This involves the following steps:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.614, + 0.483, + 0.644 + ], + "angle": 0, + "content": "- Length Filtering: Remove excessively long entries or those with too many dialogue turns for consistency." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.644, + 0.483, + 0.673 + ], + "angle": 0, + "content": "- Colloquialism Check: Utilize the Spacy [35] module to filter data for alignment with natural, everyday language." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.673, + 0.483, + 0.703 + ], + "angle": 0, + "content": "- LLM-Based Evaluation: Employ LLM to assess and select data that best matches the chatting tone." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.704, + 0.483, + 0.734 + ], + "angle": 0, + "content": "- Manual Verification: Manual verification is conducted to eliminate any inappropriate or irrelevant samples." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.614, + 0.483, + 0.734 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.483, + 0.795 + ], + "angle": 0, + "content": "Details of each step are provided in the Appendix. Through this rigorous process, we refine the initial 74,364 entries to a final set of 14,564 high-quality, freestyle chatting samples, constituting the novel benchmark." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Setting Division: In practical scenarios, new models often have limited demonstration data, underscoring the importance of evaluations under constrained data. Therefore, we further divide ChatGenBench into two settings—Supervised and Few-shot—based on the availability of samples for each model in the TrainSet. The complete data composition is detailed in Table 2." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.234, + 0.715, + 0.249 + ], + "angle": 0, + "content": "3.1.4 Benchmark Analysis" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.258, + 0.907, + 0.395 + ], + "angle": 0, + "content": "ChatGenBench offers distinct advantages over previous benchmarks, as summarized in Table 1. On the data level, ChatGenBench includes large-scale high-quality data and a broader range of T2I models. Additionally, human demonstrations provide relative ground truth across each step, allowing for step-wise evaluation that pinpoints potential challenges in automation models. ChatGenBench is also the novel benchmark to support multiple input types, making it more aligned with real-world scenarios." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.405, + 0.826, + 0.421 + ], + "angle": 0, + "content": "3.2. ChatGen: Achieving Automatic T2I" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.429, + 0.907, + 0.55 + ], + "angle": 0, + "content": "Our goal is to train the model capable of processing freestyle user inputs and generating the necessary components for image generation (prompt, model, and argument), thereby achieving Automatic T2I. In this section, we first introduce a baseline method, ChatGen-Base. We then analyze the limitations of this approach and subsequently propose ChatGen-Evo, which leverages the multi-stage evolution strategy to enhance performance." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.569, + 0.74, + 0.584 + ], + "angle": 0, + "content": "3.2.1 ChatGen-Base with SFT" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.594, + 0.906, + 0.698 + ], + "angle": 0, + "content": "We first apply supervised fine-tuning (SFT) to develop the ChatGen-Base model as an intuitive baseline. Given a set of freestyle chatting inputs \\( c \\) (which may include text, images, and historical context) and their corresponding outputs comprising prompt \\( p \\), model \\( m \\), and argument \\( r \\), we use the auto-regressive objective to maximize the following loss:" + }, + { + "type": "equation", + "bbox": [ + 0.578, + 0.713, + 0.907, + 0.743 + ], + "angle": 0, + "content": "\\[\nL _ {s f t} ^ {\\text {b a s e}} = - \\sum_ {t} \\log P _ {\\pi} (p, m, r \\mid c, * _ {< t}). \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.906, + 0.825 + ], + "angle": 0, + "content": "ChatGen-Base satisfies the essential requirements for generating outputs, including handling diverse input modalities and producing corresponding responses. However, even after fine-tuning SOTA open-source MLLMs, the model frequently produces unsatisfactory results." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.906, + 0.901 + ], + "angle": 0, + "content": "We attribute this to the complexity of automatic T2I, which requires multi-step reasoning. The quality of generated prompts directly impacts model selection, while model selection serves as a prerequisite for argument configuration. Previous studies [21, 25, 40] have shown that directly" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13287" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.139, + 0.091, + 0.859, + 0.292 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.27, + 0.299, + 0.725, + 0.313 + ], + "angle": 0, + "content": "Figure 3. Illustration of the framework for ChatGen-Base and ChatGen-Evo." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.322, + 0.484, + 0.385 + ], + "angle": 0, + "content": "predicting answers often fails in multi-step reasoning tasks. Additionally, supervised fine-tuning tends to encourage the model to learn simple text mappings instead of developing skills needed for automation, leading to poor generalization." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.4, + 0.242, + 0.415 + ], + "angle": 0, + "content": "3.2.2 ChatGen-Evo" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.424, + 0.484, + 0.56 + ], + "angle": 0, + "content": "To address the limitations of ChatGen-Base, we propose ChatGen-Evo, which trains MLLM \\(M\\) using a multi-stage evolution strategy. Instead of relying on final outcome supervision as in traditional fine-tuning, the multi-stage evolution strategy in ChatGen-Evo employs stage supervision. By providing more precise feedback at each stage, this approach gradually enables the MLLM to acquire the necessary capabilities for automated T2I. As shown in Figure 3, the training process includes three main stages:" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.563, + 0.484, + 0.639 + ], + "angle": 0, + "content": "Stage 1: Prompt Writing via SFT. We first train the MLLM using SFT with pairs of freestyle inputs and high-quality prompts. Different from the objective in ChatGenbase, the current stage focuses on a more specific and simplified task instead:" + }, + { + "type": "equation", + "bbox": [ + 0.165, + 0.647, + 0.483, + 0.677 + ], + "angle": 0, + "content": "\\[\nL _ {s f t} ^ {\\text {s t a g e 1}} = - \\sum_ {t} \\log P _ {\\pi} \\left(p \\mid c ^ {\\prime}, * _ {< t}\\right). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.686, + 0.483, + 0.763 + ], + "angle": 0, + "content": "Here, \\( c' \\) represents the freestyle input with a prompt prefix that clarifies the task being performed. This prefix helps preserve the MLLM's original capabilities, minimizing catastrophic forgetting [22]. Through this stage, the model learns how to rewrite inputs into effective prompts." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Stage 2: Model Selection via ModelToken. We introduce the ModelToken strategy to equip the model with model selection capabilities without impacting the prompt-writing ability learned in Stage 1. Inspired by token learning [10, 15], ModelToken extends this approach by representing each candidate T2I model as a unique token within the MLLM's vocabulary. Specifically, the model tokens are parameterized as an embedding matrix \\( W_{\\mathcal{M}} \\in \\mathbb{R}^{|\\mathcal{M}| \\times d} \\) and appended to the original word token matrix \\( W_{\\nu} \\in \\mathbb{R}^{|\\mathcal{V}| \\times d} \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.322, + 0.907, + 0.385 + ], + "angle": 0, + "content": "ModelToken Training: during training, the user input \\( c \\) and prompt \\( p \\) are concatenated as a prefix, with the special model token \\( <\\text{Model}_i> \\) appended as the ground truth for next-token prediction. The training objective is:" + }, + { + "type": "equation", + "bbox": [ + 0.575, + 0.395, + 0.907, + 0.412 + ], + "angle": 0, + "content": "\\[\nL \\left(W _ {\\mathcal {M}}\\right) = - \\log P _ {\\pi} \\left(< \\text {M o d e l} _ {-} i > | c, p\\right). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.424, + 0.906, + 0.499 + ], + "angle": 0, + "content": "Unlike typical next-token prediction training, the embedding matrix \\( W_{\\mathcal{M}} \\) represents the only tunable parameters, significantly enhancing training efficiency. With a small parameter size, fewer training samples are required, improving performance in scenarios with limited data." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.5, + 0.906, + 0.576 + ], + "angle": 0, + "content": "Inference for Model Selection: Once the embedding matrix is trained, the inference process concatenates the model token and original word token, forming the new language modeling head of the MLLM. In this way, the MLLM predicts the next token with the following probability:" + }, + { + "type": "equation", + "bbox": [ + 0.568, + 0.587, + 0.905, + 0.605 + ], + "angle": 0, + "content": "\\[\nP _ {\\pi} (m \\mid c, p) = \\operatorname {s o f t m a x} \\left(\\left[ W _ {\\nu}; W _ {\\mathcal {M}} \\right] \\cdot h _ {i - 1}\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.609, + 0.906, + 0.716 + ], + "angle": 0, + "content": "where the operation \\([;]\\) denotes concatenation, and \\(h_{i-1} \\in \\mathbb{R}^d\\) represents the last hidden state. Once a model token is predicted, the MLLM stops decoding, and the corresponding model \\(m\\) is selected. Additionally, information such as the model's description and demonstrations is loaded for subsequent use. After Stage 2, the model not only retains its prompt-writing skills but also learns to select models." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Stage 3: Argument Configuration via In-Context Learning. After the above two stages, we have obtained the prompt \\( p \\) and model \\( m \\) from the original user input \\( c \\). The model now needs to generate the appropriate argument configuration to complete the final generation. Due to the careful design of the previous stages, the model's in-context learning ability is maximally preserved. Therefore, we adopt a training-free approach: we guide the MLLM using in-context demonstrations of \\( m \\), the user's original input \\( c \\), and the rewritten prompt \\( p \\). The MLLM can follow the demonstration pattern to complete the parameter configuration for the current user request:" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13288" + } + ], + [ + { + "type": "equation", + "bbox": [ + 0.217, + 0.092, + 0.482, + 0.108 + ], + "angle": 0, + "content": "\\[\na = M (c, p, D (m)), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.112, + 0.484, + 0.204 + ], + "angle": 0, + "content": "where \\( D(m) \\) represents the set of demonstrations for model \\( m \\) in the training dataset. Thanks to the prior acquisition of the relevant model and prompt in earlier stages, this approach frees up context space, enabling extensive demonstrations. Additionally, the train-free approach avoids interfering with the trained model from the previous two stages." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.217, + 0.216, + 0.234 + ], + "angle": 0, + "content": "4. Experiment" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.243, + 0.484, + 0.364 + ], + "angle": 0, + "content": "We conduct a comprehensive evaluation of various methods on the novel ChatGenBench. First, in section 4.2, we compare ChatGen-Evo with other baseline models, highlighting the efficacy and efficiency of our multi-stage evolution strategy. Next, in section 4.3, we perform extensive ablation studies, uncovering the impact of each step on the final results and providing valuable insights. Finally, we provide visualizations of the generated images in section 4.4." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.374, + 0.298, + 0.39 + ], + "angle": 0, + "content": "4.1. Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.397, + 0.484, + 0.518 + ], + "angle": 0, + "content": "Training Setups. We adopt InternVL2 [5] as the base MLLM, fully fine-tuning it for both ChatGen-Base and the first stage of ChatGen-Evo. In the second stage of ChatGen-Evo, all model parameters are frozen except for the ModelToken embeddings. We employ the AdamW optimizer with a learning rate of 4e-5 and a weight decay of 1.0 over 5 epochs, maintaining these settings consistently across training stages. All experiments are conducted on 8 A100 GPUs." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.526, + 0.484, + 0.587 + ], + "angle": 0, + "content": "Metrics for Step-wise Evaluation: Leveraging the comprehensive process data in ChatGenBench, we introduce the step-wise evaluation metrics to assess the distinct abilities of automatic T2I models in key stages:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.589, + 0.483, + 0.695 + ], + "angle": 0, + "content": "- Prompt BERTScore: To assess prompt rewriting ability, we use BERTScore [48] to compare predicted prompts with high-quality, human-validated prompts. BERTScore leverages pre-trained contextual embeddings to match words in candidate and reference sentences. The metric ranges from 0 to 1, with 1 indicating highly similar meanings and 0 signifying complete dissimilarity." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.696, + 0.483, + 0.739 + ], + "angle": 0, + "content": "- Selection Accuracy: We calculate the accuracy of model selection by comparing the predicted T2I model with the human-validated model." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.741, + 0.483, + 0.801 + ], + "angle": 0, + "content": "- Argument Accuracy: We evaluate argument configuration by calculating the exact match accuracy between the predicted and human-validated arguments. The overall accuracy is obtained by averaging individual arguments." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.589, + 0.483, + 0.801 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.803, + 0.484, + 0.91 + ], + "angle": 0, + "content": "It is important to note that obtaining the absolute ground truth across all stages is nearly impossible due to the infinite search space. Therefore, we use human-validated high-quality records as relative ground truth. While these are not absolute, the comprehensive evaluation of the large-scale benchmark is able to reflect the capabilities of automation models, as confirmed by experimental results." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.289 + ], + "angle": 0, + "content": "Image Quality Evaluation: We use HPS v2 [41] and ImageReward [42] metrics to assess the quality of generated images, reflecting alignment with human preferences. Additionally, we employ FID and CLIP Score to evaluate how well the generated images meet user requirements. FID [12] measures the distance between automatically generated images and human-validated high-quality images, while CLIP Score [11] calculates the similarity between the generated images and human-validated prompts. To provide an intuitive and comprehensive measure of image quality, we normalize and combine these four metrics into an aggregated score, Unified Metric [49]. Details of each metric and the calculation of Unified Metric are presented in the Appendix." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.297, + 0.906, + 0.388 + ], + "angle": 0, + "content": "Baseline Without Fine-Tuning. We further establish a baseline that uses the default in-context learning capabilities of the MLLM for prompt rewriting, along with a single Stable Diffusion model and a fixed set of default parameters. This baseline helps emphasize the significance of prompt rewriting and multi-model utilization." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.398, + 0.688, + 0.414 + ], + "angle": 0, + "content": "4.2. Main Experiment" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.421, + 0.712, + 0.435 + ], + "angle": 0, + "content": "4.2.1 Quantitative Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.445, + 0.906, + 0.627 + ], + "angle": 0, + "content": "Table 3 presents the main quantitative results of ChatGenEvo compared to different baselines. Overall, ChatGenEvo significantly outperforms other methods across all metrics, including both step-wise and final image quality evaluations. Specifically, the low performance of the baseline highlights the importance of effective prompt rewriting and multi-model selection, underscoring the necessity of dedicated Automatic T2I methods. Additionally, fine-tuning MLLMs with progressively larger parameter scales yields steady performance improvements. Remarkably, ChatGenEvo achieves performance comparable to ChatGen-Base at 8B, despite utilizing a significantly smaller parameter scale." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.628, + 0.906, + 0.809 + ], + "angle": 0, + "content": "The comparisons across different settings provide further insights. Prompt rewriting ability demonstrates strong generalizability, effectively transferring to rarely seen models. The higher few-shot performance can be attributed to the bias in data distribution between the two settings, as reflected in the \"Baseline\" results, which reveal a difference of approximately 0.3. In contrast, model selection and parameter configuration show a noticeable decline in performance under few-shot conditions, highlighting their reliance on more training samples. Therefore, exploring the scaling laws of prompt rewriting and enhancing model selection in few-shot scenarios are promising directions." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.817, + 0.702, + 0.831 + ], + "angle": 0, + "content": "4.2.2 Human Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.841, + 0.906, + 0.902 + ], + "angle": 0, + "content": "We conduct a user study using pairwise comparisons to further evaluate ChatGen-Base(8B) and ChatGen-Evo(2B). Users are presented with two images generated from the same input: one by ChatGen-Base and the other by" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13289" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.226, + 0.09, + 0.768, + 0.103 + ], + "angle": 0, + "content": "Table 3. The Step-wise and Final evaluation results of different methods on ChatGenBench." + }, + { + "type": "table", + "bbox": [ + 0.113, + 0.104, + 0.879, + 0.306 + ], + "angle": 0, + "content": "
Step-wise EvaluationFinal Evaluation
PromptBertScore ↑SelectionAcc ↑ConfigAcc ↑FIDScore ↓CLIPScore ↑HPSv2 ↑ImageReward ↑UnifiedMetric ↑
SupervisedBaseline0.026--32.764.620.2-34.637.3
ChatGen-Base(2B)0.1840.2060.38421.369.923.52.459.0
ChatGen-Base(4B)0.1970.2300.49020.770.023.41.558.7
ChatGen-Base(8B)0.2080.2640.50920.870.723.94.060.7
ChatGen-Evo (2B)0.2470.3280.53719.172.925.18.965.9
Few-shotBaseline0.055--54.463.420.0-40.229.7
ChatGen-Base(2B)0.2210.1530.34942.869.123.3-4.851.1
ChatGen-Base(4B)0.2360.1710.44841.269.423.4-4.351.9
ChatGen-Base(8B)0.2520.2010.46241.470.623.4-3.152.5
ChatGen-Evo (2B)0.2830.2310.49340.772.525.05.159.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.311, + 0.482, + 0.367 + ], + "angle": 0, + "content": "Table 4. Efficiency comparison. Training efficiency is measured in GPU hours, while inference is expressed as the seconds required to process each data. The training was conducted on 8 A100 GPUs, and inference was performed on a single A100 80GB GPU." + }, + { + "type": "table", + "bbox": [ + 0.1, + 0.368, + 0.472, + 0.445 + ], + "angle": 0, + "content": "
MethodParamsTrainingInference
ChatGen-Base(2B)2.21B76h1.1s
ChatGen-Base(8B)8.08B240h2.3s
ChatGen-Evo(2B)2.24B100h1.9s
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.448, + 0.483, + 0.509 + ], + "angle": 0, + "content": "ChatGen-Evo. It is tasked with selecting the image that better matches the image quality and relevance to the given input. We sample 2,000 image pairs for supervised and 1,000 for few-shot. More details are provided in the Appendix." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.51, + 0.484, + 0.6 + ], + "angle": 0, + "content": "As shown in Figure 4, the human evaluation results are consistent with the quantitative metrics, highlighting that ChatGen-Evo outperforms ChatGen-Base in both image quality and relevance. Additionally, in few-shot settings, ChatGen-Evo demonstrates a higher win rate, showcasing the effectiveness of our approach in data-scarce scenarios." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.608, + 0.305, + 0.623 + ], + "angle": 0, + "content": "4.2.3 Efficiency Comparison" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.633, + 0.483, + 0.738 + ], + "angle": 0, + "content": "Similar to existing multi-stage reasoning methods [25, 40, 46], ChatGen-Evo does not offer an efficiency advantage over direct prediction approaches like ChatGen-Base. However, the significant performance gains more than compensate for this drawback. As shown in Table 4, ChatGen-Evo achieves the performance level of ChatGen-Base at 8B parameters using only 2B parameters. Therefore, when com" + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.742, + 0.48, + 0.882 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.094, + 0.883, + 0.479, + 0.897 + ], + "angle": 0, + "content": "Figure 4. User study results of ChatGen-Base and ChatGen-Evo." + }, + { + "type": "table_caption", + "bbox": [ + 0.542, + 0.311, + 0.876, + 0.325 + ], + "angle": 0, + "content": "Table 5. Ablation experiments on the supervised setting." + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.327, + 0.898, + 0.48 + ], + "angle": 0, + "content": "
StageMethodsStep Score ↑FID Score ↓Image Reward ↑Unified Metric ↑
Prompt WritingBaseline0.02619.3-13.150.8
Chat-Base0.18421.33.159.7
Chat-Evo0.24717.610.566.8
Model SelectionChat-Base0.20620.59.166.0
Chat-Evo0.55318.216.869.7
Argument SettingChat-Base0.38417.910.469.1
Chat-Evo0.87117.117.970.3
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.489, + 0.906, + 0.52 + ], + "angle": 0, + "content": "paring efficiency under equal performance, ChatGen-Evo maintains a relative advantage over ChatGen-Base." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.532, + 0.616, + 0.548 + ], + "angle": 0, + "content": "4.3. Analysis" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.556, + 0.708, + 0.571 + ], + "angle": 0, + "content": "4.3.1 Capability Analysis." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.582, + 0.906, + 0.627 + ], + "angle": 0, + "content": "We conduct ablation experiments on ChatGenBench to evaluate the contribution of individual steps to final performance by providing ground truth for the other steps." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.629, + 0.906, + 0.809 + ], + "angle": 0, + "content": "Prompt Writing: We first compare the prompt writing capabilities of different methods, along with their final performance. In these experiments, each method's predicted prompt was passed along with the correct model and argument. As shown in Table 5, all methods exhibit noticeable gaps compared to human-validated prompts, emphasizing the complexity of prompt rewriting. Comparisons with the results in Table 3 also highlight the further improvements achieved by selecting the correct model and arguments, particularly for the \"Baseline\" methods. Moreover, variations in prompts significantly influence the final results, highlighting its critical role in Automatic T2I." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Model Selection: Table 5 also shows the impact of model selection where human-validated prompt and argument are provided. With well-crafted prompts, ChatGen-Evo demonstrates a substantial performance boost from \\(32.8\\%\\) to \\(55.3\\%\\), indicating the strong influence of prompt quality on model selection accuracy. This supports our perspec" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13290" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.141, + 0.097, + 0.168, + 0.116 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.169, + 0.098, + 0.335, + 0.116 + ], + "angle": 0, + "content": "Let's craft a fierce anime warrior maiden, ablaze with power and passion!" + }, + { + "type": "image", + "bbox": [ + 0.141, + 0.117, + 0.207, + 0.191 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.149, + 0.194, + 0.198, + 0.202 + ], + "angle": 0, + "content": "Default SD" + }, + { + "type": "image", + "bbox": [ + 0.209, + 0.118, + 0.273, + 0.191 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.209, + 0.193, + 0.27, + 0.201 + ], + "angle": 0, + "content": "ChatGen-Base" + }, + { + "type": "image", + "bbox": [ + 0.276, + 0.118, + 0.341, + 0.192 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.277, + 0.193, + 0.339, + 0.202 + ], + "angle": 0, + "content": "ChatGen-Evo" + }, + { + "type": "image", + "bbox": [ + 0.35, + 0.097, + 0.37, + 0.116 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.372, + 0.098, + 0.573, + 0.116 + ], + "angle": 0, + "content": "Imagine a fluffy cyberpunk critter in a detailed, artsy, and concept style. Keep the image perfect !" + }, + { + "type": "image", + "bbox": [ + 0.351, + 0.118, + 0.423, + 0.192 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.362, + 0.193, + 0.411, + 0.201 + ], + "angle": 0, + "content": "Default SD" + }, + { + "type": "image", + "bbox": [ + 0.425, + 0.118, + 0.476, + 0.192 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.431, + 0.193, + 0.498, + 0.202 + ], + "angle": 0, + "content": "ChatGen-Base" + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.118, + 0.576, + 0.192 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.505, + 0.193, + 0.568, + 0.202 + ], + "angle": 0, + "content": "ChatGen-Evo" + }, + { + "type": "image", + "bbox": [ + 0.603, + 0.097, + 0.626, + 0.116 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.627, + 0.098, + 0.814, + 0.116 + ], + "angle": 0, + "content": "Do you reckon we could mix pirates and punk for a car design? Make Something stands out." + }, + { + "type": "image", + "bbox": [ + 0.584, + 0.118, + 0.658, + 0.192 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.595, + 0.193, + 0.645, + 0.202 + ], + "angle": 0, + "content": "Default SD" + }, + { + "type": "image", + "bbox": [ + 0.661, + 0.118, + 0.758, + 0.192 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.677, + 0.193, + 0.742, + 0.202 + ], + "angle": 0, + "content": "ChatGen-Base" + }, + { + "type": "image", + "bbox": [ + 0.762, + 0.118, + 0.857, + 0.192 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.778, + 0.193, + 0.839, + 0.202 + ], + "angle": 0, + "content": "ChatGen-Evo" + }, + { + "type": "image", + "bbox": [ + 0.142, + 0.214, + 0.165, + 0.233 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.167, + 0.215, + 0.327, + 0.233 + ], + "angle": 0, + "content": "I want the final image echoes the reference's shitty, comic-inspired style?" + }, + { + "type": "image", + "bbox": [ + 0.141, + 0.235, + 0.201, + 0.283 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.147, + 0.285, + 0.194, + 0.303 + ], + "angle": 0, + "content": "reference image" + }, + { + "type": "image", + "bbox": [ + 0.203, + 0.235, + 0.264, + 0.3 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.201, + 0.302, + 0.265, + 0.309 + ], + "angle": 0, + "content": "ChatGen-Base" + }, + { + "type": "image", + "bbox": [ + 0.265, + 0.235, + 0.35, + 0.301 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.273, + 0.302, + 0.335, + 0.31 + ], + "angle": 0, + "content": "ChatGen-Evo" + }, + { + "type": "image", + "bbox": [ + 0.352, + 0.214, + 0.374, + 0.233 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.375, + 0.215, + 0.568, + 0.233 + ], + "angle": 0, + "content": "Can we create the image similar to the original, with soft colors and intricate details?" + }, + { + "type": "image", + "bbox": [ + 0.353, + 0.235, + 0.415, + 0.302 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.425, + 0.235, + 0.511, + 0.302 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.437, + 0.302, + 0.502, + 0.31 + ], + "angle": 0, + "content": "ChatGen-Base" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.235, + 0.604, + 0.302 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.529, + 0.302, + 0.591, + 0.31 + ], + "angle": 0, + "content": "ChatGen-Evo" + }, + { + "type": "image", + "bbox": [ + 0.622, + 0.214, + 0.644, + 0.233 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.647, + 0.216, + 0.837, + 0.233 + ], + "angle": 0, + "content": "Recreate the intricate patterns of the reference, blending realism and dreamlike qualities?" + }, + { + "type": "image", + "bbox": [ + 0.624, + 0.235, + 0.683, + 0.285 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.629, + 0.286, + 0.677, + 0.305 + ], + "angle": 0, + "content": "reference image" + }, + { + "type": "image", + "bbox": [ + 0.692, + 0.235, + 0.764, + 0.313 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.769, + 0.235, + 0.845, + 0.313 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.144, + 0.324, + 0.165, + 0.343 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.165, + 0.324, + 0.259, + 0.351 + ], + "angle": 0, + "content": "Hey, you know Flandre Scarlet? Create an image for her!" + }, + { + "type": "image", + "bbox": [ + 0.144, + 0.366, + 0.165, + 0.385 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.165, + 0.364, + 0.259, + 0.391 + ], + "angle": 0, + "content": "Picture her having a blast in a creek, all cute and nature!" + }, + { + "type": "image", + "bbox": [ + 0.272, + 0.322, + 0.358, + 0.389 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.297, + 0.39, + 0.335, + 0.397 + ], + "angle": 0, + "content": "Round 1" + }, + { + "type": "image", + "bbox": [ + 0.361, + 0.322, + 0.475, + 0.39 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.399, + 0.39, + 0.437, + 0.397 + ], + "angle": 0, + "content": "Round 2" + }, + { + "type": "image", + "bbox": [ + 0.481, + 0.322, + 0.502, + 0.342 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.503, + 0.324, + 0.604, + 0.342 + ], + "angle": 0, + "content": "Creat Van Gogh painting of a human-like possum" + }, + { + "type": "image", + "bbox": [ + 0.482, + 0.351, + 0.502, + 0.361 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.503, + 0.352, + 0.607, + 0.362 + ], + "angle": 0, + "content": "It's holding a fruit basket." + }, + { + "type": "image", + "bbox": [ + 0.481, + 0.374, + 0.502, + 0.383 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.503, + 0.374, + 0.612, + 0.391 + ], + "angle": 0, + "content": "With vibrant yellows, deep blue colors for fruits!" + }, + { + "type": "image", + "bbox": [ + 0.629, + 0.317, + 0.7, + 0.389 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.645, + 0.39, + 0.683, + 0.397 + ], + "angle": 0, + "content": "Round 1" + }, + { + "type": "image", + "bbox": [ + 0.701, + 0.317, + 0.774, + 0.389 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.723, + 0.39, + 0.76, + 0.397 + ], + "angle": 0, + "content": "Round 2" + }, + { + "type": "image", + "bbox": [ + 0.771, + 0.317, + 0.852, + 0.39 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.796, + 0.39, + 0.834, + 0.397 + ], + "angle": 0, + "content": "Round 3" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.403, + 0.904, + 0.417 + ], + "angle": 0, + "content": "Figure 5. Examples of images generated by different methods. Three rows represent single, multi-modal and historical inputs, respectively." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.422, + 0.483, + 0.497 + ], + "angle": 0, + "content": "tive that Automatic T2I fundamentally involves multi-step reasoning. In contrast, ChatGen-Base, which directly predicts all results, fails to adapt and thus produces unchanged outputs. Furthermore, these results also emphasize the substantial impact of model selection on overall performance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.498, + 0.483, + 0.603 + ], + "angle": 0, + "content": "Argument Configuration: When provided with high-quality prompts and appropriate model selection, ChatGenEvo exhibits notable performance improvements. It improves configuration accuracy from \\(53.7\\%\\) to \\(87.1\\%\\) and Unified Score from 65.9 to 70.3. Overall, while argument configuration has a relatively smaller impact compared to previous stages, it remains a crucial component." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.604, + 0.483, + 0.664 + ], + "angle": 0, + "content": "The above findings suggest that outcomes in earlier steps significantly influence predictions in subsequent ones. Therefore, exploring more reasoning methods for advancing automatic T2I represents a promising research direction." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.679, + 0.289, + 0.695 + ], + "angle": 0, + "content": "4.3.2 Input Type Analysis." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.703, + 0.483, + 0.809 + ], + "angle": 0, + "content": "Table 6 presents ChatGen-Evo performance across different input types. Multimodal inputs lead to better performance, as images may offer clearer prompt and model identification cues compared to text alone. Additionally, handling historical data remains a significant challenge, with the lowest performance across all metrics. These results point to valuable directions in enhancing history-based prompt generation." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.818, + 0.235, + 0.832 + ], + "angle": 0, + "content": "4.4. Visualizations" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Figure 5 presents examples of images generated by different methods. In the first row, ChatGen-Evo understands user requirements and identifies suitable models to generate style-matching images. The second row demonstrates" + }, + { + "type": "table_caption", + "bbox": [ + 0.547, + 0.421, + 0.872, + 0.435 + ], + "angle": 0, + "content": "Table 6. The evaluation results of different input types." + }, + { + "type": "table", + "bbox": [ + 0.523, + 0.437, + 0.892, + 0.535 + ], + "angle": 0, + "content": "
Input TypeStep-wise EvaluationFinal
Prompt BertScore ↑Selection Acc ↑Config Acc ↑Unified Score ↑
Single0.2520.3310.53968.1
M-Modal0.2770.3810.55469.3
History0.1650.2590.50560.1
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.548, + 0.906, + 0.64 + ], + "angle": 0, + "content": "results based on multi-modal user inputs, where ChatGen-Evo shows a superior understanding of the reference image and preserves more details to generate refined outputs. The third row illustrates ChatGen-Evo's capability in handling historical data, ensuring that each step inherits the previous style while making appropriate modifications." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.661, + 0.633, + 0.677 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.69, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Our research aims to automate tedious steps in T2I generation, allowing users to simply describe their needs in a freestyle chatting way. We introduce ChatGenBench for benchmarking Automatic T2I task. It includes high-quality paired data with diverse freestyle user inputs, enabling comprehensive evaluation across all steps. Furthermore, we argue that Automatic T2I should be regarded as a multi-step reasoning task. Consequently, we propose ChatGen-Evo, a multi-stage evolution strategy that progressively equips models with essential automation skills. Extensive evaluations not only demonstrate the superiority of ChatGen-Evo but also provide valuable insights for advancing Automatic T2I. We believe this represents a significant step toward the future of automated generative models." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "13291" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.251, + 0.108 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.486, + 0.298 + ], + "angle": 0, + "content": "This work was supported by the National Nature Science Foundation of China (No. 62192781, No. 62272374), the Natural Science Foundation of Shaanxi Province (2024JC-JCQN-62), the National Nature Science Foundation of China (No. 62202367, No. 62250009, No. 62137002), the Key Research and Development Project in Shaanxi Province No. 2022GXLH-01-03, Project of China Knowledge Center for Engineering Science and Technology, and Project of Chinese academy of engineering \"The Online and Offline Mixed Educational Service System for 'The Belt and Road' Training in MOOC China\", and the K. C. Wong Education Foundation." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.31, + 0.188, + 0.326 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.336, + 0.482, + 0.404 + ], + "angle": 0, + "content": "[1] James Betker, Gabriel Goh, Li Jing, Tim Brooks, Jianfeng Wang, Linjie Li, Long Ouyang, Juntang Zhuang, Joyce Lee, Yufei Guo, Wesam Manassra, Prafulla Dhariwal, Casey Chu, and Yunxin Jiao. Improving image generation with better captions. 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.406, + 0.482, + 0.473 + ], + "angle": 0, + "content": "[2] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18392-18402, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.476, + 0.482, + 0.504 + ], + "angle": 0, + "content": "[3] Tom B Brown. Language models are few-shot learners. arXiv preprint arXiv:2005.14165, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.506, + 0.482, + 0.56 + ], + "angle": 0, + "content": "[4] Tingfeng Cao, Chengyu Wang, Bingyan Liu, Ziheng Wu, Jinhui Zhu, and Jun Huang. Beautifulprompt: Towards automatic prompt engineering for text-to-image synthesis. arXiv preprint arXiv:2311.06752, 2023. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.562, + 0.482, + 0.631 + ], + "angle": 0, + "content": "[5] Zhe Chen, Weiyun Wang, Hao Tian, Shenglong Ye, Zhangwei Gao, Erfei Cui, Wenwen Tong, Kongzhi Hu, Jiapeng Luo, Zheng Ma, et al. How far are we to gpt-4v? closing the gap to commercial multimodal models with open-source suites. arXiv preprint arXiv:2404.16821, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.633, + 0.47, + 0.646 + ], + "angle": 0, + "content": "[6] Civitai. Civitai. https://civitai.com/, 2022.1,3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.648, + 0.482, + 0.715 + ], + "angle": 0, + "content": "[7] Runpei Dong, Chunrui Han, Yang Peng, Zekun Qi, Zheng Ge, Jinrong Yang, Liang Zhao, Jianjian Sun, Hongyu Zhou, Haoran Wei, et al. Dreamllm: Synergistic multimodal comprehension and creation. arXiv preprint arXiv:2309.11499, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.718, + 0.48, + 0.744 + ], + "angle": 0, + "content": "[8] Hugging Face. Hugging face. https://huggingface.co/, 2016.1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.746, + 0.483, + 0.815 + ], + "angle": 0, + "content": "[9] Yutong Feng, Biao Gong, Di Chen, Yujun Shen, Yu Liu, and Jingren Zhou. Ranni: Taming text-to-image diffusion for accurate instruction following. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4744-4753, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.817, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[10] Shibo Hao, Tianyang Liu, Zhen Wang, and Zhiting Hu. Toolkengpt: Augmenting frozen language models with massive tools via tool embeddings. Advances in neural information processing systems, 36, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[11] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation met" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.336, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.906, + 0.12 + ], + "angle": 0, + "content": "ric for image captioning. In Empirical Methods in Natural Language Processing (EMNLP), 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.906, + 0.188 + ], + "angle": 0, + "content": "[12] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.19, + 0.905, + 0.232 + ], + "angle": 0, + "content": "[13] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.234, + 0.907, + 0.314 + ], + "angle": 0, + "content": "[14] Chengyou Jia, Minnan Luo, Zhuohang Dang, Guang Dai, Xiaojun Chang, Mengmeng Wang, and Jingdong Wang. Ssmg: Spatial-semantic map guided diffusion model for free-form layout-to-image generation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2480–2488, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.317, + 0.905, + 0.384 + ], + "angle": 0, + "content": "[15] Chengyou Jia, Minnan Luo, Zhuohang Dang, Qiushi Sun, Fangzhi Xu, Junlin Hu, Tianbao Xie, and Zhiyong Wu. Agentstore: Scalable integration of heterogeneous agents as specialized generalist computer assistant. arXiv preprint arXiv:2410.18603, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.386, + 0.905, + 0.455 + ], + "angle": 0, + "content": "[16] Bahjat Kawar, Shiran Zada, Oran Lang, Omer Tov, Huiwen Chang, Tali Dekel, Inbar Mosseri, and Michal Irani. Imagc: Text-based real image editing with diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6007-6017, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.457, + 0.905, + 0.524 + ], + "angle": 0, + "content": "[17] Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization of text-to-image diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1931-1941, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.526, + 0.905, + 0.58 + ], + "angle": 0, + "content": "[18] Dongxu Li, Junnan Li, and Steven Hoi. Blip-diffusion: Pretrained subject representation for controllable text-to-image generation and editing. Advances in Neural Information Processing Systems, 36, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.582, + 0.905, + 0.649 + ], + "angle": 0, + "content": "[19] Yuheng Li, Haotian Liu, Qingyang Wu, Fangzhou Mu, Jianwei Yang, Jianfeng Gao, Chunyuan Li, and Yong Jae Lee. Gligen: Open-set grounded text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22511-22521, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.651, + 0.905, + 0.705 + ], + "angle": 0, + "content": "[20] Long Lian, Boyi Li, Adam Yala, and Trevor Darrell. Llm-grounded diffusion: Enhancing prompt understanding of text-to-image diffusion models with large language models. arXiv preprint arXiv:2305.13655, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.707, + 0.905, + 0.761 + ], + "angle": 0, + "content": "[21] Hunter Lightman, Vineet Kosaraju, Yura Burda, Harri Edwards, Bowen Baker, Teddy Lee, Jan Leike, John Schulman, Ilya Sutskever, and Karl Cobbe. Let's verify step by step. arXiv preprint arXiv:2305.20050, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.763, + 0.905, + 0.817 + ], + "angle": 0, + "content": "[22] Yun Luo, Zhen Yang, Fandong Meng, Yafu Li, Jie Zhou, and Yue Zhang. An empirical study of catastrophic forgetting in large language models during continual fine-tuning. arXiv preprint arXiv:2308.08747, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.818, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[23] Alexander Quinn Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. In International Conference on Machine Learning (ICML), 2022. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13292" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.092, + 0.424, + 0.108 + ], + "angle": 0, + "content": "[24] OpenAI. Chatgpt, 2022. Large language model. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.109, + 0.482, + 0.15 + ], + "angle": 0, + "content": "[25] OpenAI. Learning to reason with llms. https://openai.com/index/learning-to-reason-with-llms/, 2024. Accessed: 2024-11-06. 2, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.152, + 0.483, + 0.194 + ], + "angle": 0, + "content": "[26] Shishir G Patil, Tianjun Zhang, Xin Wang, and Joseph E Gonzalez. Gorilla: Large language model connected with massive apis. arXiv preprint arXiv:2305.15334, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.196, + 0.483, + 0.25 + ], + "angle": 0, + "content": "[27] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4195-4205, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.254, + 0.483, + 0.31 + ], + "angle": 0, + "content": "[28] Jie Qin, Jie Wu, Weifeng Chen, Yuxi Ren, Huixia Li, Hefeng Wu, Xuefeng Xiao, Rui Wang, and Shilei Wen. Diffusiongpt: Llm-driven text-to-image generation system. arXiv preprint arXiv:2401.10061, 2024. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.312, + 0.483, + 0.38 + ], + "angle": 0, + "content": "[29] Yujia Qin, Shihao Liang, Yining Ye, Kunlun Zhu, Lan Yan, Yaxi Lu, Yankai Lin, Xin Cong, Xiangru Tang, Bill Qian, et al. Toolllm: Facilitating large language models to master 16000+ real-world apis. arXiv preprint arXiv:2307.16789, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.383, + 0.483, + 0.466 + ], + "angle": 0, + "content": "[30] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.469, + 0.482, + 0.524 + ], + "angle": 0, + "content": "[31] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.527, + 0.485, + 0.596 + ], + "angle": 0, + "content": "[32] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.599, + 0.483, + 0.68 + ], + "angle": 0, + "content": "[33] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 22500-22510, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.684, + 0.483, + 0.725 + ], + "angle": 0, + "content": "[34] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.728, + 0.483, + 0.756 + ], + "angle": 0, + "content": "[35] SpaCy. spacy: Industrial-strength natural language processing in python. https://spacy.io/, 2015.4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.758, + 0.378, + 0.772 + ], + "angle": 0, + "content": "[36] Omost Team. Omost github page, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.774, + 0.483, + 0.843 + ], + "angle": 0, + "content": "[37] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[38] Jiapeng Wang, Chengyu Wang, Tingfeng Cao, Jun Huang, and Lianwen Jin. Diffchat: Learning to chat with text-to-image synthesis models for interactive image creation. arXiv preprint arXiv:2403.04997, 2024. 2, 4" + }, + { + "type": "list", + "bbox": [ + 0.092, + 0.092, + 0.485, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.146 + ], + "angle": 0, + "content": "[39] Zhenyu Wang, Aoxue Li, Zhenguo Li, and Xihui Liu. Genartist: Multimodal llm as an agent for unified image generation and editing. arXiv preprint arXiv:2407.05600, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.149, + 0.905, + 0.219 + ], + "angle": 0, + "content": "[40] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824-24837, 2022. 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.221, + 0.905, + 0.288 + ], + "angle": 0, + "content": "[41] Xiaoshi Wu, Yiming Hao, Keqiang Sun, Yixiong Chen, Feng Zhu, Rui Zhao, and Hongsheng Li. Human preference score v2: A solid benchmark for evaluating human preferences of text-to-image synthesis. arXiv preprint arXiv:2306.09341, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.291, + 0.905, + 0.358 + ], + "angle": 0, + "content": "[42] Jiazheng Xu, Xiao Liu, Yuchen Wu, Yuxuan Tong, Qinkai Li, Ming Ding, Jie Tang, and Yuxiao Dong. Imagereward: Learning and evaluating human preferences for text-to-image generation. arXiv preprint arXiv:2304.05977, 2023.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.361, + 0.904, + 0.43 + ], + "angle": 0, + "content": "[43] Ling Yang, Zhaochen Yu, Chenlin Meng, Minkai Xu, Stefano Ermon, and CUI Bin. Mastering text-to-image diffusion: Recaptioning, planning, and generating with multimodal llms. In *Forty-first International Conference on Machine Learning*, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.432, + 0.905, + 0.487 + ], + "angle": 0, + "content": "[44] Rui Yang, Lin Song, Yanwei Li, Sijie Zhao, Yixiao Ge, Xiu Li, and Ying Shan. Gpt4tools: Teaching large language model to use tools via self-instruction. Advances in Neural Information Processing Systems, 36, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.488, + 0.905, + 0.543 + ], + "angle": 0, + "content": "[45] Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, and Yuan Cao. React: Synergizing reasoning and acting in language models. arXiv preprint arXiv:2210.03629, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.545, + 0.905, + 0.613 + ], + "angle": 0, + "content": "[46] Shunyu Yao, Dian Yu, Jeffrey Zhao, Izhak Shafran, Tom Griffiths, Yuan Cao, and Karthik Narasimhan. Tree of thoughts: Deliberate problem solving with large language models. Advances in Neural Information Processing Systems, 36, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.615, + 0.905, + 0.671 + ], + "angle": 0, + "content": "[47] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.672, + 0.905, + 0.727 + ], + "angle": 0, + "content": "[48] Tianyi Zhang, Varsha Kishore, Felix Wu, Kilian Q Weinberger, and Yoav Artzi. Bertscore: Evaluating text generation with bert. In International Conference on Learning Representations. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.728, + 0.905, + 0.811 + ], + "angle": 0, + "content": "[49] Lirui Zhao, Yue Yang, Kaipeng Zhang, Wenqi Shao, Yuxin Zhang, Yu Qiao, Ping Luo, and Rongrong Ji. Diffagent: Fast and accurate text-to-image api selection with large language model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6390-6399, 2024. 2, 3, 4, 6" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.811 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13293" + } + ] +] \ No newline at end of file diff --git a/2025/ChatGen_ Automatic Text-to-Image Generation From FreeStyle Chatting/0c8e9c4a-231c-4c6d-af84-cd3eb4170579_origin.pdf b/2025/ChatGen_ Automatic Text-to-Image Generation From FreeStyle Chatting/0c8e9c4a-231c-4c6d-af84-cd3eb4170579_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..81908f41826afc86b8133162bf9c3e88707367a9 --- /dev/null +++ b/2025/ChatGen_ Automatic Text-to-Image Generation From FreeStyle Chatting/0c8e9c4a-231c-4c6d-af84-cd3eb4170579_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9a91b6ac9a848fd2c44a8994cd0b0274d5bbb789183e3b8ff7a44105c7493cab +size 4962452 diff --git a/2025/ChatGen_ Automatic Text-to-Image Generation From FreeStyle Chatting/full.md b/2025/ChatGen_ Automatic Text-to-Image Generation From FreeStyle Chatting/full.md new file mode 100644 index 0000000000000000000000000000000000000000..714d0e534b8c02874c7f2d77639f2288edb66613 --- /dev/null +++ b/2025/ChatGen_ Automatic Text-to-Image Generation From FreeStyle Chatting/full.md @@ -0,0 +1,438 @@ +![](images/db1b627fef85d9975afd23e19c76597338f4300abef7e78599868821ef8ab03e.jpg) + +# ChatGen: Automatic Text-to-Image Generation From FreeStyle Chatting + +Chengyou Jia $^{1*}$ , Changliang Xia $^{1*}$ , Zhuohang Dang $^{1}$ , Weijia Wu $^{2}$ , Hangwei Qian $^{3}$ , Minnan Luo $^{1\dagger}$ $^{1}$ School of Computer Science and Technology, MOEKLINNS Lab, Xi'an Jiaotong University + $^{2}$ National University of Singapore $^{3}$ CFAR, A\*STAR + +cp3jia@stu.xjtu.edu.cn, 202066@stu.xjtu.edu.cn, dangzhuohang@stu.xjtu.edu.cn, weijiawu96@gmail.com, qian0045@e.ntu.edu.sg, minnluo@xjtu.edu.cn + +# Abstract + +Despite the significant advancements in text-to-image (T2I) generative models, users often face a trial-and-error challenge in practical scenarios. This challenge arises from the complexity and uncertainty of tedious steps such as crafting suitable prompts, selecting appropriate models, and configuring specific arguments, making users resort to labor-intensive attempts for desired images. This paper proposes Automatic T2I generation, which aims to automate these tedious steps, allowing users to simply describe their needs in a freestyle chatting way. To systematically study this problem, we first introduce ChatGenBench, a novel benchmark designed for Automatic T2I. It features high-quality paired data with diverse freestyle inputs, enabling comprehensive evaluation of automatic T2I models across all steps. Additionally, recognizing Automatic T2I as a complex multi-step reasoning task, we propose ChatGenEvo, a multi-stage evolution strategy that progressively equips models with essential automation skills. Through extensive evaluation across step-wise accuracy and image quality, ChatGen-Evo significantly enhances performance over various baselines. Our evaluation also uncovers valuable insights for advancing automatic T2I. Our data, code, and models will be publicly available. + +# 1. Introduction + +In recent years, text-to-image (T2I) generative models have attracted considerable attention [1, 23, 31, 32]. Building on advancements in large-scale T2I models such as DALL-E [31] and Stable Diffusion [32], the open-source community has significantly expanded the capabilities of T2I generation. Researchers fine-tune open-source models on specialized datasets, resulting in a diverse selection of task-specific models available on platforms like Civitai [6] and Hugging + +![](images/5cdd819e119516922bb9791598092e026f4c4d5dd6706a31278adeba8876c65c.jpg) + +![](images/437ff6f487593781e614df16cb07e5a2ac787539257ca89cbe30109cbb0c3f64.jpg) +Figure 1. Illustration of tedious steps in T2I. Our method can select an appropriate model with suitable prompts and arguments. Note: Arle is a character from the game Genshin Impact. + +Face [8]. This variety provides users with a broader range of options to meet customization needs, facilitating the growing adoption of T2I models in real-world applications. + +However, the rapid development of T2I models within the open-source community has also introduced significant challenges for users. When non-experts attempt to create images with specific requirements, they often encounter a trial-and-error process involving several tedious steps. As shown in Figure 1, these steps include crafting suitable prompts, selecting appropriate models, and configuring specific model arguments. The complexity and uncertainty of each step turn the process into an arduous journey, resembling "mice in a maze". In real-world scenarios, this iterative process consumes substantial time and resources + +as users continuously adjust settings to regenerate images. Therefore, we pose the challenging problem: can we automate these labor-intensive steps in T2I generation, allowing users to simply describe their needs in a chatting style and receive desired images effortlessly? + +Previous attempts, such as BeautifulPrompt [4, 38] for generating high-quality prompts from low-quality ones and DiffAgent [49] for model selection using large language models, have made some progress in addressing these challenges. However, these methods only seek to automate a specific step in Figure 1, lacking comprehensive research on the automation of entire T2I process. Moreover, they fail to support diverse types of freestyle input, leaving them far from real-world scenarios. To bridge these gaps, we aim to develop the model that can accept arbitrary user input, similar to ChatGPT [24], and automatically generate all necessary components for generation, termed as Automatic T2I. + +To this end, we first introduce ChatGenBench, the benchmark specifically designed for this task. ChatGenBench offers a substantial dataset of high-quality paired data from 6,807 customized models. Each data pair comprises a user's freestyle chat input, a refined prompt, an appropriate model, and optimal arguments. This comprehensive step-by-step trail enables step-wise evaluation of automatic T2I models, ensuring both quality assessment of the final image output and precise identification of potential automation bottlenecks. Furthermore, ChatGenBench integrates various types of data, enabling testing with multimodal and historical input to simulate practical scenarios. + +An intuitive approach to achieve the goal is to build supervised fine-tuning (SFT) data and tune multimodal large language models. However, we argue that automated T2I should be viewed as a complex multi-step reasoning problem. Directly predicting outputs leads the model to focus on simple text mappings, rather than developing the diverse skills needed for robust automation. Inspired by OpenAI's o1 [25], we propose ChatGen-Evo with a multi-stage evolution strategy. This approach provides targeted feedback at each stage, progressively equipping models with essential automation skills. First, we construct SFT data to train the model to generate high-quality prompts from freestyle inputs. Next, we augment the model's vocabulary with specialized ModelTokens, enabling effective model selection without affecting other functions. Finally, we guide the model in configuring arguments based on prompts and models selected in previous stages. By decomposing the task into clear stages, the model gains crucial automation skills, delivering outputs aligned with user expectations. + +We conduct a comprehensive evaluation of various methods on the novel ChatGenBench to study Automatic T2I. ChatGen-Evo significantly outperforms other methods across all metrics, including both step-wise assessments and image quality evaluations. Extensive quantitative and + +qualitative results, along with ablation studies, underscore the critical importance of the multi-stage evolution strategy. Additionally, further experiments reveal the impact of different stages on the final T2I output, uncovering valuable insights into the challenges and opportunities in Automatic T2I. We highlight some promising directions, such as exploring the scaling laws of prompt rewriting, improving model selection in few-shot scenarios, and leveraging reasoning approaches to enhance performance. The contributions of this paper are summarized as follows: + +- We propose the novel and challenging problem of Automatic T2I generation, aiming to develop the model that can handle user's freestyle chatting and automatically produce all necessary components for image generation. +- We introduce ChatGenBench, a benchmark specifically designed for Automatic T2I. It includes a comprehensive step-by-step trail for step-wise evaluation, supporting multimodal or historical user inputs. +- We present ChatGen-Evo, which adopts a multi-stage evolution strategy. By decomposing the task into distinct stages, ChatGen-Evo enables the model to progressively acquire essential Automatic T2I capabilities. +- Extensive experiments on ChatGenBench not only demonstrate the superiority of ChatGen-Evo but also provide valuable insights for advancing automatic T2I. + +# 2. Related Works + +# 2.1. Text-to-Image Generation + +With the advancement of diffusion models [13, 27, 34], text-to-image (T2I) generation [1, 23, 31] has demonstrated exceptional capabilities in high-quality image generation and textual description alignment. Large-scale models such as DALL-E [1] enhance text-image alignment by leveraging the joint feature space of CLIP [30]. Moreover, Stable Diffusion [32], a well-established open-source model, has gained substantial attention. Numerous methods have been developed to fine-tune it or design additional modules for specialized tasks, such as customized image generation [17, 18, 33], layout-to-image generation [14, 19, 20, 47] and image edit [2, 16]. These diverse models have significantly expanded the capabilities of T2I. However, they also present significant learning challenges for non-expert users, underscoring the increasing need for automatic T2I generation. + +# 2.2. LLMs for Text-to-Image Generation + +Large language models (LLMs), like ChatGPT [24], Llama [37], have demonstrated impressive capability in language understanding [3] and problem solving [45]. Recently, LLMs have also been applied to image generation. [7, 9, 20, 36, 43] leverage LLMs to generate detailed layout information from complex prompts, enabling the creation of sub-elements and control over their positioning. + +![](images/47bba521a124c435dac97253b6eec3ba31b070c1ca5af27dc5c30cf25b123934.jpg) +Figure 2. Illustration of the data collection pipeline. + +[28, 39, 49] propose employ LLMs for model selection, akin to tool usage [26, 29, 44]. However, the above methods still require specialized prompt inputs and involve complex model usage. Currently, no approach thoroughly considers leveraging LLMs to relieve users from all tedious steps in T2I. Our method aims to address this gap. + +# 3. Methodology + +Our goal is to relieve users from tedious steps and automate to produce the desired images from user's freestyle input. To achieve this objective, we first introduce ChatGenBench, a benchmark comprising a large set of user inputs in a chatting style, built upon a foundation of over 6000 personalized models to evaluate automated image generation results. A comprehensive description of data collection, construction, and analysis is provided in Sec. 3.1. Then, we present ChatGen-Evo in Sec. 3.2, which uses a multi-stage evolution strategy to train MLLM for Automatic T2I. + +# 3.1. ChatGenBench: Benchmarking Automatic T2I + +For clarity, an example of this data is shown on the right side of Figure 2. The following sections detail the data collection and construction process, which primarily involves High-Quality Human Demonstration Collection and using LLM-driven Role-Play to simulate user input. + +# 3.1.1 High-Quality Human Demonstration Collection + +Civitai [6] is a vibrant community where users share customized AI models for generating high-quality images. Members contribute demonstrations with detailed prompts, model specifications, and arguments, supported by a feedback system that ensures quality validation. These features make Civitai an ideal platform for collecting raw data. + +Collection: We start by collecting demonstrations based on established evaluation metrics within Civitai, including download counts, upvotes, and other user feedback. These metrics enable us to collect data that has been validated through community engagement. By focusing on these indicators, we identify a subset of high-quality results. + +Filtering: Following the initial collection, we implement a rigorous filtering process to ensure data quality. This involves excluding demonstrations associated with inactive + +or outdated models, removing duplicates, and filtering out NSFW content. This careful curation refined the dataset to include the most effective demonstrations. Ultimately, this process results in a curated set of 44,881 high-quality human demonstrations across 6,807 unique models. + +# 3.1.2 LLM-Driven Role-Play for Chatting Generation + +While the demonstrations collected from communities include the essential procedural information needed for automation, they lack a critical element: freestyle chatting inputs. Such data is not available on open platforms, which has been a key limitation for previous methods [49]. To address this, we propose an LLM-Driven Role-Play strategy for Chatting Generation. As shown in Figure 2, we predefine over 100 roles from everyday life (e.g., students, doctors, professors) and prompt the LLM to simulate these roles, translating each demonstration into freestyle chatting input with tones and habits of the character. This approach significantly enhances data diversity and creates more lifelike inputs. To further enhance reverse synthesis diversity, we incorporate multiple LLMs, adjust diversity arguments, and use BERTScore [48] to filter out overly similar inputs. + +Additionally, we define three types of freestyle input formats: single-input, consisting of a single chatting-style sentence; multimodal-input, combining a sentence with an image; and history-input, comprising multiple rounds of dialogue history. These formats effectively simulate how users typically inquire about image generation needs, greatly expanding the practical value of automatic T2I. An example of the single-input prompt is shown in the following. Additional prompts and details are provided in the Appendix. + +# Prompt: Sing-Input chatting generation + +System: You are a professional user experience designer who plays various personas to convert complex and professional content for non-professional users. Please merge the following prompt and model information into a single freestyle query. Remove any obvious details that non-professional users would avoid. Make it similar to what non-professional users may write. The converted single-text query should be colloquial and as brief as possible. + +Require: {ROLE}, {PROMPT}, {MODEL}, {Example 1,...,n} +Input: You are playing the {ROLE}. Please generate a single text query based on the following professional {PROMPT} for the {MODEL}. You can refer to following examples: {Example 1,...,n}. + +Table 1. Comparison of different methods for benchmarking automatic text-to-image generation. + +
TrainSet (# Numbers)TestSet (# Numbers)Based Models?Step-wise Evaluation?Freestyle User-input?Multi-modal Support?History Support?
BeautifulPrompt [4]143K2KSingleX✓XX
DiffChat [38]234K5KSingleX✓X✓
DiffusionGPT [28]--Multi(≈20)X✓XX
DABench [49]40K5KMulti(5K)XXXX
ChatGenBench256K14KMulti(6K)✓✓✓✓
+ +Table 2. Summary of dataset statistics. + +
DatasetTotalSingleM-ModalHistory
TrainSet256,606147,88869,54839,170
TestSet Init74,36442,83820,21411,312
Benchmark14,56411,0111,6681,132
Supervised10,2408,0091,0991,132
Few-Shot4,3243,002569753
+ +# 3.1.3 Benchmark Construction + +Following the above steps, we generate 330,970 freestyle inputs from 44,881 demonstrations. Considering the largescale generated data, we don't randomly split a test set for benchmarking. Instead, we carefully select high-quality, non-overlapping samples as the benchmark for evaluation. The selection involves the following steps: + +TestSet Split: We split the data based on model origin. For data associated with the same model, we use BERTScore to assess the similarity between samples. From this, we select about $20\%$ most semantically distinct data as initial TestSet in Table 2. This maximizes non-overlap with TrainSet. + +TestSet Filtering: Building on the initial TestSet, we perform multiple filtering rounds to ensure the final ChatGen-Bench's quality. This involves the following steps: + +- Length Filtering: Remove excessively long entries or those with too many dialogue turns for consistency. +- Colloquialism Check: Utilize the Spacy [35] module to filter data for alignment with natural, everyday language. +- LLM-Based Evaluation: Employ LLM to assess and select data that best matches the chatting tone. +- Manual Verification: Manual verification is conducted to eliminate any inappropriate or irrelevant samples. + +Details of each step are provided in the Appendix. Through this rigorous process, we refine the initial 74,364 entries to a final set of 14,564 high-quality, freestyle chatting samples, constituting the novel benchmark. + +Setting Division: In practical scenarios, new models often have limited demonstration data, underscoring the importance of evaluations under constrained data. Therefore, we further divide ChatGenBench into two settings—Supervised and Few-shot—based on the availability of samples for each model in the TrainSet. The complete data composition is detailed in Table 2. + +# 3.1.4 Benchmark Analysis + +ChatGenBench offers distinct advantages over previous benchmarks, as summarized in Table 1. On the data level, ChatGenBench includes large-scale high-quality data and a broader range of T2I models. Additionally, human demonstrations provide relative ground truth across each step, allowing for step-wise evaluation that pinpoints potential challenges in automation models. ChatGenBench is also the novel benchmark to support multiple input types, making it more aligned with real-world scenarios. + +# 3.2. ChatGen: Achieving Automatic T2I + +Our goal is to train the model capable of processing freestyle user inputs and generating the necessary components for image generation (prompt, model, and argument), thereby achieving Automatic T2I. In this section, we first introduce a baseline method, ChatGen-Base. We then analyze the limitations of this approach and subsequently propose ChatGen-Evo, which leverages the multi-stage evolution strategy to enhance performance. + +# 3.2.1 ChatGen-Base with SFT + +We first apply supervised fine-tuning (SFT) to develop the ChatGen-Base model as an intuitive baseline. Given a set of freestyle chatting inputs $c$ (which may include text, images, and historical context) and their corresponding outputs comprising prompt $p$ , model $m$ , and argument $r$ , we use the auto-regressive objective to maximize the following loss: + +$$ +L _ {s f t} ^ {\text {b a s e}} = - \sum_ {t} \log P _ {\pi} (p, m, r \mid c, * _ {< t}). \tag {1} +$$ + +ChatGen-Base satisfies the essential requirements for generating outputs, including handling diverse input modalities and producing corresponding responses. However, even after fine-tuning SOTA open-source MLLMs, the model frequently produces unsatisfactory results. + +We attribute this to the complexity of automatic T2I, which requires multi-step reasoning. The quality of generated prompts directly impacts model selection, while model selection serves as a prerequisite for argument configuration. Previous studies [21, 25, 40] have shown that directly + +![](images/c815dc7614cd5792197e1fa79c8a33e6fa8f6c92bb255f77bd1a850ba35bee55.jpg) +Figure 3. Illustration of the framework for ChatGen-Base and ChatGen-Evo. + +predicting answers often fails in multi-step reasoning tasks. Additionally, supervised fine-tuning tends to encourage the model to learn simple text mappings instead of developing skills needed for automation, leading to poor generalization. + +# 3.2.2 ChatGen-Evo + +To address the limitations of ChatGen-Base, we propose ChatGen-Evo, which trains MLLM $M$ using a multi-stage evolution strategy. Instead of relying on final outcome supervision as in traditional fine-tuning, the multi-stage evolution strategy in ChatGen-Evo employs stage supervision. By providing more precise feedback at each stage, this approach gradually enables the MLLM to acquire the necessary capabilities for automated T2I. As shown in Figure 3, the training process includes three main stages: + +Stage 1: Prompt Writing via SFT. We first train the MLLM using SFT with pairs of freestyle inputs and high-quality prompts. Different from the objective in ChatGenbase, the current stage focuses on a more specific and simplified task instead: + +$$ +L _ {s f t} ^ {\text {s t a g e 1}} = - \sum_ {t} \log P _ {\pi} \left(p \mid c ^ {\prime}, * _ {< t}\right). \tag {2} +$$ + +Here, $c'$ represents the freestyle input with a prompt prefix that clarifies the task being performed. This prefix helps preserve the MLLM's original capabilities, minimizing catastrophic forgetting [22]. Through this stage, the model learns how to rewrite inputs into effective prompts. + +Stage 2: Model Selection via ModelToken. We introduce the ModelToken strategy to equip the model with model selection capabilities without impacting the prompt-writing ability learned in Stage 1. Inspired by token learning [10, 15], ModelToken extends this approach by representing each candidate T2I model as a unique token within the MLLM's vocabulary. Specifically, the model tokens are parameterized as an embedding matrix $W_{\mathcal{M}} \in \mathbb{R}^{|\mathcal{M}| \times d}$ and appended to the original word token matrix $W_{\nu} \in \mathbb{R}^{|\mathcal{V}| \times d}$ . + +ModelToken Training: during training, the user input $c$ and prompt $p$ are concatenated as a prefix, with the special model token $<\text{Model}_i>$ appended as the ground truth for next-token prediction. The training objective is: + +$$ +L \left(W _ {\mathcal {M}}\right) = - \log P _ {\pi} \left(< \text {M o d e l} _ {-} i > | c, p\right). \tag {3} +$$ + +Unlike typical next-token prediction training, the embedding matrix $W_{\mathcal{M}}$ represents the only tunable parameters, significantly enhancing training efficiency. With a small parameter size, fewer training samples are required, improving performance in scenarios with limited data. + +Inference for Model Selection: Once the embedding matrix is trained, the inference process concatenates the model token and original word token, forming the new language modeling head of the MLLM. In this way, the MLLM predicts the next token with the following probability: + +$$ +P _ {\pi} (m \mid c, p) = \operatorname {s o f t m a x} \left(\left[ W _ {\nu}; W _ {\mathcal {M}} \right] \cdot h _ {i - 1}\right), \tag {4} +$$ + +where the operation $[;]$ denotes concatenation, and $h_{i-1} \in \mathbb{R}^d$ represents the last hidden state. Once a model token is predicted, the MLLM stops decoding, and the corresponding model $m$ is selected. Additionally, information such as the model's description and demonstrations is loaded for subsequent use. After Stage 2, the model not only retains its prompt-writing skills but also learns to select models. + +Stage 3: Argument Configuration via In-Context Learning. After the above two stages, we have obtained the prompt $p$ and model $m$ from the original user input $c$ . The model now needs to generate the appropriate argument configuration to complete the final generation. Due to the careful design of the previous stages, the model's in-context learning ability is maximally preserved. Therefore, we adopt a training-free approach: we guide the MLLM using in-context demonstrations of $m$ , the user's original input $c$ , and the rewritten prompt $p$ . The MLLM can follow the demonstration pattern to complete the parameter configuration for the current user request: + +$$ +a = M (c, p, D (m)), \tag {5} +$$ + +where $D(m)$ represents the set of demonstrations for model $m$ in the training dataset. Thanks to the prior acquisition of the relevant model and prompt in earlier stages, this approach frees up context space, enabling extensive demonstrations. Additionally, the train-free approach avoids interfering with the trained model from the previous two stages. + +# 4. Experiment + +We conduct a comprehensive evaluation of various methods on the novel ChatGenBench. First, in section 4.2, we compare ChatGen-Evo with other baseline models, highlighting the efficacy and efficiency of our multi-stage evolution strategy. Next, in section 4.3, we perform extensive ablation studies, uncovering the impact of each step on the final results and providing valuable insights. Finally, we provide visualizations of the generated images in section 4.4. + +# 4.1. Experimental Settings + +Training Setups. We adopt InternVL2 [5] as the base MLLM, fully fine-tuning it for both ChatGen-Base and the first stage of ChatGen-Evo. In the second stage of ChatGen-Evo, all model parameters are frozen except for the ModelToken embeddings. We employ the AdamW optimizer with a learning rate of 4e-5 and a weight decay of 1.0 over 5 epochs, maintaining these settings consistently across training stages. All experiments are conducted on 8 A100 GPUs. + +Metrics for Step-wise Evaluation: Leveraging the comprehensive process data in ChatGenBench, we introduce the step-wise evaluation metrics to assess the distinct abilities of automatic T2I models in key stages: + +- Prompt BERTScore: To assess prompt rewriting ability, we use BERTScore [48] to compare predicted prompts with high-quality, human-validated prompts. BERTScore leverages pre-trained contextual embeddings to match words in candidate and reference sentences. The metric ranges from 0 to 1, with 1 indicating highly similar meanings and 0 signifying complete dissimilarity. +- Selection Accuracy: We calculate the accuracy of model selection by comparing the predicted T2I model with the human-validated model. +- Argument Accuracy: We evaluate argument configuration by calculating the exact match accuracy between the predicted and human-validated arguments. The overall accuracy is obtained by averaging individual arguments. + +It is important to note that obtaining the absolute ground truth across all stages is nearly impossible due to the infinite search space. Therefore, we use human-validated high-quality records as relative ground truth. While these are not absolute, the comprehensive evaluation of the large-scale benchmark is able to reflect the capabilities of automation models, as confirmed by experimental results. + +Image Quality Evaluation: We use HPS v2 [41] and ImageReward [42] metrics to assess the quality of generated images, reflecting alignment with human preferences. Additionally, we employ FID and CLIP Score to evaluate how well the generated images meet user requirements. FID [12] measures the distance between automatically generated images and human-validated high-quality images, while CLIP Score [11] calculates the similarity between the generated images and human-validated prompts. To provide an intuitive and comprehensive measure of image quality, we normalize and combine these four metrics into an aggregated score, Unified Metric [49]. Details of each metric and the calculation of Unified Metric are presented in the Appendix. + +Baseline Without Fine-Tuning. We further establish a baseline that uses the default in-context learning capabilities of the MLLM for prompt rewriting, along with a single Stable Diffusion model and a fixed set of default parameters. This baseline helps emphasize the significance of prompt rewriting and multi-model utilization. + +# 4.2. Main Experiment + +# 4.2.1 Quantitative Results + +Table 3 presents the main quantitative results of ChatGenEvo compared to different baselines. Overall, ChatGenEvo significantly outperforms other methods across all metrics, including both step-wise and final image quality evaluations. Specifically, the low performance of the baseline highlights the importance of effective prompt rewriting and multi-model selection, underscoring the necessity of dedicated Automatic T2I methods. Additionally, fine-tuning MLLMs with progressively larger parameter scales yields steady performance improvements. Remarkably, ChatGenEvo achieves performance comparable to ChatGen-Base at 8B, despite utilizing a significantly smaller parameter scale. + +The comparisons across different settings provide further insights. Prompt rewriting ability demonstrates strong generalizability, effectively transferring to rarely seen models. The higher few-shot performance can be attributed to the bias in data distribution between the two settings, as reflected in the "Baseline" results, which reveal a difference of approximately 0.3. In contrast, model selection and parameter configuration show a noticeable decline in performance under few-shot conditions, highlighting their reliance on more training samples. Therefore, exploring the scaling laws of prompt rewriting and enhancing model selection in few-shot scenarios are promising directions. + +# 4.2.2 Human Evaluation + +We conduct a user study using pairwise comparisons to further evaluate ChatGen-Base(8B) and ChatGen-Evo(2B). Users are presented with two images generated from the same input: one by ChatGen-Base and the other by + +Table 3. The Step-wise and Final evaluation results of different methods on ChatGenBench. + +
Step-wise EvaluationFinal Evaluation
PromptBertScore ↑SelectionAcc ↑ConfigAcc ↑FIDScore ↓CLIPScore ↑HPSv2 ↑ImageReward ↑UnifiedMetric ↑
SupervisedBaseline0.026--32.764.620.2-34.637.3
ChatGen-Base(2B)0.1840.2060.38421.369.923.52.459.0
ChatGen-Base(4B)0.1970.2300.49020.770.023.41.558.7
ChatGen-Base(8B)0.2080.2640.50920.870.723.94.060.7
ChatGen-Evo (2B)0.2470.3280.53719.172.925.18.965.9
Few-shotBaseline0.055--54.463.420.0-40.229.7
ChatGen-Base(2B)0.2210.1530.34942.869.123.3-4.851.1
ChatGen-Base(4B)0.2360.1710.44841.269.423.4-4.351.9
ChatGen-Base(8B)0.2520.2010.46241.470.623.4-3.152.5
ChatGen-Evo (2B)0.2830.2310.49340.772.525.05.159.2
+ +Table 4. Efficiency comparison. Training efficiency is measured in GPU hours, while inference is expressed as the seconds required to process each data. The training was conducted on 8 A100 GPUs, and inference was performed on a single A100 80GB GPU. + +
MethodParamsTrainingInference
ChatGen-Base(2B)2.21B76h1.1s
ChatGen-Base(8B)8.08B240h2.3s
ChatGen-Evo(2B)2.24B100h1.9s
+ +ChatGen-Evo. It is tasked with selecting the image that better matches the image quality and relevance to the given input. We sample 2,000 image pairs for supervised and 1,000 for few-shot. More details are provided in the Appendix. + +As shown in Figure 4, the human evaluation results are consistent with the quantitative metrics, highlighting that ChatGen-Evo outperforms ChatGen-Base in both image quality and relevance. Additionally, in few-shot settings, ChatGen-Evo demonstrates a higher win rate, showcasing the effectiveness of our approach in data-scarce scenarios. + +# 4.2.3 Efficiency Comparison + +Similar to existing multi-stage reasoning methods [25, 40, 46], ChatGen-Evo does not offer an efficiency advantage over direct prediction approaches like ChatGen-Base. However, the significant performance gains more than compensate for this drawback. As shown in Table 4, ChatGen-Evo achieves the performance level of ChatGen-Base at 8B parameters using only 2B parameters. Therefore, when com + +![](images/fb50ae8208316390f900da2ba5d44dcd2b2f869a16ea75cfa84f657be5880e8a.jpg) +Figure 4. User study results of ChatGen-Base and ChatGen-Evo. + +Table 5. Ablation experiments on the supervised setting. + +
StageMethodsStep Score ↑FID Score ↓Image Reward ↑Unified Metric ↑
Prompt WritingBaseline0.02619.3-13.150.8
Chat-Base0.18421.33.159.7
Chat-Evo0.24717.610.566.8
Model SelectionChat-Base0.20620.59.166.0
Chat-Evo0.55318.216.869.7
Argument SettingChat-Base0.38417.910.469.1
Chat-Evo0.87117.117.970.3
+ +paring efficiency under equal performance, ChatGen-Evo maintains a relative advantage over ChatGen-Base. + +# 4.3. Analysis + +# 4.3.1 Capability Analysis. + +We conduct ablation experiments on ChatGenBench to evaluate the contribution of individual steps to final performance by providing ground truth for the other steps. + +Prompt Writing: We first compare the prompt writing capabilities of different methods, along with their final performance. In these experiments, each method's predicted prompt was passed along with the correct model and argument. As shown in Table 5, all methods exhibit noticeable gaps compared to human-validated prompts, emphasizing the complexity of prompt rewriting. Comparisons with the results in Table 3 also highlight the further improvements achieved by selecting the correct model and arguments, particularly for the "Baseline" methods. Moreover, variations in prompts significantly influence the final results, highlighting its critical role in Automatic T2I. + +Model Selection: Table 5 also shows the impact of model selection where human-validated prompt and argument are provided. With well-crafted prompts, ChatGen-Evo demonstrates a substantial performance boost from $32.8\%$ to $55.3\%$ , indicating the strong influence of prompt quality on model selection accuracy. This supports our perspec + +![](images/85cefe667a402dce8fec1cd3fa25ae01cf17bd463dbeca1f804c429ec53d31e1.jpg) + +Let's craft a fierce anime warrior maiden, ablaze with power and passion! + +![](images/d1b7e35f4720dbfde592aeb9ac30469536e22126a27ab9789939dd7293ca57c4.jpg) +Default SD +ChatGen-Base + +![](images/5b16fcf44ebd45aa84c978e53d7e1e357eccae4a2e2190570f20e3f876ffc7fe.jpg) + +![](images/3561f8c5ca0c901fe0170e7fd2e59920dd18aa8a121bf0aee12f56541c989756.jpg) +ChatGen-Evo + +![](images/6b887e4e0a5962454eae8930ee9c586e7439e1492f6a6583134c8e15cf44e381.jpg) + +Imagine a fluffy cyberpunk critter in a detailed, artsy, and concept style. Keep the image perfect ! + +![](images/14c5d1b4a9d4406a2a256f68da248be33ea593b23667e6e8afba99074835a74d.jpg) +Default SD + +![](images/290d606325f79fc3357ac4a46105c4dc430f254157266ed492f69762cd438b33.jpg) +ChatGen-Base + +![](images/393eb531804b83a2ebe13550ec5b301f693698fb319682630bf59d312a3907b7.jpg) +ChatGen-Evo + +![](images/9ff084a696b5424bb23e10dca752975dbbb1895804105ded2f21451ef2e0e129.jpg) + +Do you reckon we could mix pirates and punk for a car design? Make Something stands out. + +![](images/0969957ea7d22bbc2d6ad6501d3a77884c078a1453da4b24d8959a7b875653c8.jpg) +Default SD + +![](images/1396a873ef64093583c0666293ebdb0b195c600982175fcb49fef8d93e4f1b45.jpg) +ChatGen-Base + +![](images/af069c871ac878976299eb99f1bf14c02057221e57d58e23cb641281321357aa.jpg) +ChatGen-Evo + +![](images/cf795807719c2c106a1237bf9c663731e8d7f6a7c45cf9dfd2f86eb9da183d9a.jpg) + +I want the final image echoes the reference's shitty, comic-inspired style? + +![](images/6d2091d230945f1bec1ad717d37c6d6fef650985b5db25f799525bd4f615ec46.jpg) +reference image + +![](images/0ed95a1db6d50657d89ebbdf0aca7f177664f23d66b897391f776374c0d53186.jpg) + +![](images/38a64f804862b3279d45958a82ed66442eccd9eaacdd70fc7251433881dabfbf.jpg) +ChatGen-Base +ChatGen-Evo + +![](images/1ba0d0a657deb9366dd654ce15de0324a33cf41f53c1e7eb6b345ec2ba5dee00.jpg) + +Can we create the image similar to the original, with soft colors and intricate details? + +![](images/c486fc45431016f1573560bf876ad90347b5088c0d3682982561124a8485346f.jpg) + +![](images/06f14b95b13cc32fcc019291d49cdd1466168f8aa38e67496e8e0f1d043856fd.jpg) +ChatGen-Base + +![](images/5729c35dbffc8d22ac6f74322938166944cff611e72cb2816b3b5c2a92b5eaf2.jpg) +ChatGen-Evo + +![](images/b13d3281eab8b69afdbc068dcc37f9a8f2cceef5654ad67a8b8cfb203ce1a59b.jpg) + +Recreate the intricate patterns of the reference, blending realism and dreamlike qualities? + +![](images/f4e7b527ed0afaf40dc7d9d3acf697b920fefdc29be7d50983bdd2187a6db0e2.jpg) +reference image + +![](images/58a91af8787f1b1903dadad9fea8123decbcd558f6fdea4763c8443f32650fdb.jpg) + +![](images/8caf608f5a4cb0adad4c1891ad62c2d20a894940a0a8e071ae71dd63c88bfee3.jpg) + +![](images/4a783e720cbb2bd08ce6a2275ebcb8d1b2f30f9cb6a2c6e3ab9432248166bb13.jpg) + +Hey, you know Flandre Scarlet? Create an image for her! + +![](images/bd70a3ded7549128c9bed057a457e0e4ba9a0038bbcbed63bb561390cf761ff7.jpg) + +Picture her having a blast in a creek, all cute and nature! + +![](images/a84e8b0ec1b78686ecb1b44cc5c820d14374315e5be4b316d764edfe19e541d5.jpg) +Round 1 +Figure 5. Examples of images generated by different methods. Three rows represent single, multi-modal and historical inputs, respectively. + +![](images/6045c2b6cd86a2f94f4cfde6ce2d3e35c0a84fe8c60e349dd95583b76c58e919.jpg) +Round 2 + +![](images/17715eb3d1aca0dea58a4544b7068cb3ce25fee9b0911b8f2f72b157621c0f60.jpg) + +Creat Van Gogh painting of a human-like possum + +![](images/43ec982290053fb206dab9a247b870ac6b258321432c84ed77c029c40f47cbdd.jpg) + +It's holding a fruit basket. + +![](images/1b59fc1c3840f32a6404b8c5ca1c9d10874d2badbc841edaace16a5ea8e6951c.jpg) + +With vibrant yellows, deep blue colors for fruits! + +![](images/d41c3d132477d36efd47e16ca0ab03bb8dca4f28012bd3e2f0929fa462f7b278.jpg) +Round 1 + +![](images/b12bbf252b8ff64368442f406a5e0ab4b6f7673e7c05dd4ae453fe6254374439.jpg) +Round 2 + +![](images/c293acfb79e559807a2082a70c3b07fe1c0338c865dbfbd22ae987fd03a0c253.jpg) +Round 3 + +tive that Automatic T2I fundamentally involves multi-step reasoning. In contrast, ChatGen-Base, which directly predicts all results, fails to adapt and thus produces unchanged outputs. Furthermore, these results also emphasize the substantial impact of model selection on overall performance. + +Argument Configuration: When provided with high-quality prompts and appropriate model selection, ChatGenEvo exhibits notable performance improvements. It improves configuration accuracy from $53.7\%$ to $87.1\%$ and Unified Score from 65.9 to 70.3. Overall, while argument configuration has a relatively smaller impact compared to previous stages, it remains a crucial component. + +The above findings suggest that outcomes in earlier steps significantly influence predictions in subsequent ones. Therefore, exploring more reasoning methods for advancing automatic T2I represents a promising research direction. + +# 4.3.2 Input Type Analysis. + +Table 6 presents ChatGen-Evo performance across different input types. Multimodal inputs lead to better performance, as images may offer clearer prompt and model identification cues compared to text alone. Additionally, handling historical data remains a significant challenge, with the lowest performance across all metrics. These results point to valuable directions in enhancing history-based prompt generation. + +# 4.4. Visualizations + +Figure 5 presents examples of images generated by different methods. In the first row, ChatGen-Evo understands user requirements and identifies suitable models to generate style-matching images. The second row demonstrates + +Table 6. The evaluation results of different input types. + +
Input TypeStep-wise EvaluationFinal
Prompt BertScore ↑Selection Acc ↑Config Acc ↑Unified Score ↑
Single0.2520.3310.53968.1
M-Modal0.2770.3810.55469.3
History0.1650.2590.50560.1
+ +results based on multi-modal user inputs, where ChatGen-Evo shows a superior understanding of the reference image and preserves more details to generate refined outputs. The third row illustrates ChatGen-Evo's capability in handling historical data, ensuring that each step inherits the previous style while making appropriate modifications. + +# 5. Conclusion + +Our research aims to automate tedious steps in T2I generation, allowing users to simply describe their needs in a freestyle chatting way. We introduce ChatGenBench for benchmarking Automatic T2I task. It includes high-quality paired data with diverse freestyle user inputs, enabling comprehensive evaluation across all steps. Furthermore, we argue that Automatic T2I should be regarded as a multi-step reasoning task. Consequently, we propose ChatGen-Evo, a multi-stage evolution strategy that progressively equips models with essential automation skills. Extensive evaluations not only demonstrate the superiority of ChatGen-Evo but also provide valuable insights for advancing Automatic T2I. We believe this represents a significant step toward the future of automated generative models. + +# Acknowledgments + +This work was supported by the National Nature Science Foundation of China (No. 62192781, No. 62272374), the Natural Science Foundation of Shaanxi Province (2024JC-JCQN-62), the National Nature Science Foundation of China (No. 62202367, No. 62250009, No. 62137002), the Key Research and Development Project in Shaanxi Province No. 2022GXLH-01-03, Project of China Knowledge Center for Engineering Science and Technology, and Project of Chinese academy of engineering "The Online and Offline Mixed Educational Service System for 'The Belt and Road' Training in MOOC China", and the K. C. Wong Education Foundation. + +# References + +[1] James Betker, Gabriel Goh, Li Jing, Tim Brooks, Jianfeng Wang, Linjie Li, Long Ouyang, Juntang Zhuang, Joyce Lee, Yufei Guo, Wesam Manassra, Prafulla Dhariwal, Casey Chu, and Yunxin Jiao. Improving image generation with better captions. 2023. 1, 2 +[2] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18392-18402, 2023. 2 +[3] Tom B Brown. Language models are few-shot learners. arXiv preprint arXiv:2005.14165, 2020. 2 +[4] Tingfeng Cao, Chengyu Wang, Bingyan Liu, Ziheng Wu, Jinhui Zhu, and Jun Huang. Beautifulprompt: Towards automatic prompt engineering for text-to-image synthesis. arXiv preprint arXiv:2311.06752, 2023. 2, 4 +[5] Zhe Chen, Weiyun Wang, Hao Tian, Shenglong Ye, Zhangwei Gao, Erfei Cui, Wenwen Tong, Kongzhi Hu, Jiapeng Luo, Zheng Ma, et al. How far are we to gpt-4v? closing the gap to commercial multimodal models with open-source suites. arXiv preprint arXiv:2404.16821, 2024. 6 +[6] Civitai. Civitai. https://civitai.com/, 2022.1,3 +[7] Runpei Dong, Chunrui Han, Yang Peng, Zekun Qi, Zheng Ge, Jinrong Yang, Liang Zhao, Jianjian Sun, Hongyu Zhou, Haoran Wei, et al. Dreamllm: Synergistic multimodal comprehension and creation. arXiv preprint arXiv:2309.11499, 2023. 2 +[8] Hugging Face. Hugging face. https://huggingface.co/, 2016.1 +[9] Yutong Feng, Biao Gong, Di Chen, Yujun Shen, Yu Liu, and Jingren Zhou. Ranni: Taming text-to-image diffusion for accurate instruction following. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4744-4753, 2024. 2 +[10] Shibo Hao, Tianyang Liu, Zhen Wang, and Zhiting Hu. Toolkengpt: Augmenting frozen language models with massive tools via tool embeddings. Advances in neural information processing systems, 36, 2024. 5 +[11] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation met + +ric for image captioning. In Empirical Methods in Natural Language Processing (EMNLP), 2021. 6 +[12] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 6 +[13] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2 +[14] Chengyou Jia, Minnan Luo, Zhuohang Dang, Guang Dai, Xiaojun Chang, Mengmeng Wang, and Jingdong Wang. Ssmg: Spatial-semantic map guided diffusion model for free-form layout-to-image generation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2480–2488, 2024. 2 +[15] Chengyou Jia, Minnan Luo, Zhuohang Dang, Qiushi Sun, Fangzhi Xu, Junlin Hu, Tianbao Xie, and Zhiyong Wu. Agentstore: Scalable integration of heterogeneous agents as specialized generalist computer assistant. arXiv preprint arXiv:2410.18603, 2024. 5 +[16] Bahjat Kawar, Shiran Zada, Oran Lang, Omer Tov, Huiwen Chang, Tali Dekel, Inbar Mosseri, and Michal Irani. Imagc: Text-based real image editing with diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6007-6017, 2023. 2 +[17] Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization of text-to-image diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1931-1941, 2023. 2 +[18] Dongxu Li, Junnan Li, and Steven Hoi. Blip-diffusion: Pretrained subject representation for controllable text-to-image generation and editing. Advances in Neural Information Processing Systems, 36, 2024. 2 +[19] Yuheng Li, Haotian Liu, Qingyang Wu, Fangzhou Mu, Jianwei Yang, Jianfeng Gao, Chunyuan Li, and Yong Jae Lee. Gligen: Open-set grounded text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22511-22521, 2023. 2 +[20] Long Lian, Boyi Li, Adam Yala, and Trevor Darrell. Llm-grounded diffusion: Enhancing prompt understanding of text-to-image diffusion models with large language models. arXiv preprint arXiv:2305.13655, 2023. 2 +[21] Hunter Lightman, Vineet Kosaraju, Yura Burda, Harri Edwards, Bowen Baker, Teddy Lee, Jan Leike, John Schulman, Ilya Sutskever, and Karl Cobbe. Let's verify step by step. arXiv preprint arXiv:2305.20050, 2023. 4 +[22] Yun Luo, Zhen Yang, Fandong Meng, Yafu Li, Jie Zhou, and Yue Zhang. An empirical study of catastrophic forgetting in large language models during continual fine-tuning. arXiv preprint arXiv:2308.08747, 2023. 5 +[23] Alexander Quinn Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. In International Conference on Machine Learning (ICML), 2022. 1, 2 + +[24] OpenAI. Chatgpt, 2022. Large language model. 2 +[25] OpenAI. Learning to reason with llms. https://openai.com/index/learning-to-reason-with-llms/, 2024. Accessed: 2024-11-06. 2, 4, 7 +[26] Shishir G Patil, Tianjun Zhang, Xin Wang, and Joseph E Gonzalez. Gorilla: Large language model connected with massive apis. arXiv preprint arXiv:2305.15334, 2023. 3 +[27] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4195-4205, 2023. 2 +[28] Jie Qin, Jie Wu, Weifeng Chen, Yuxi Ren, Huixia Li, Hefeng Wu, Xuefeng Xiao, Rui Wang, and Shilei Wen. Diffusiongpt: Llm-driven text-to-image generation system. arXiv preprint arXiv:2401.10061, 2024. 3, 4 +[29] Yujia Qin, Shihao Liang, Yining Ye, Kunlun Zhu, Lan Yan, Yaxi Lu, Yankai Lin, Xin Cong, Xiangru Tang, Bill Qian, et al. Toolllm: Facilitating large language models to master 16000+ real-world apis. arXiv preprint arXiv:2307.16789, 2023. 3 +[30] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2 +[31] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 1, 2 +[32] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 2 +[33] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 22500-22510, 2023. 2 +[34] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 2 +[35] SpaCy. spacy: Industrial-strength natural language processing in python. https://spacy.io/, 2015.4 +[36] Omost Team. Omost github page, 2024. 2 +[37] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 2 +[38] Jiapeng Wang, Chengyu Wang, Tingfeng Cao, Jun Huang, and Lianwen Jin. Diffchat: Learning to chat with text-to-image synthesis models for interactive image creation. arXiv preprint arXiv:2403.04997, 2024. 2, 4 + +[39] Zhenyu Wang, Aoxue Li, Zhenguo Li, and Xihui Liu. Genartist: Multimodal llm as an agent for unified image generation and editing. arXiv preprint arXiv:2407.05600, 2024. 3 +[40] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824-24837, 2022. 4, 7 +[41] Xiaoshi Wu, Yiming Hao, Keqiang Sun, Yixiong Chen, Feng Zhu, Rui Zhao, and Hongsheng Li. Human preference score v2: A solid benchmark for evaluating human preferences of text-to-image synthesis. arXiv preprint arXiv:2306.09341, 2023. 6 +[42] Jiazheng Xu, Xiao Liu, Yuchen Wu, Yuxuan Tong, Qinkai Li, Ming Ding, Jie Tang, and Yuxiao Dong. Imagereward: Learning and evaluating human preferences for text-to-image generation. arXiv preprint arXiv:2304.05977, 2023.6 +[43] Ling Yang, Zhaochen Yu, Chenlin Meng, Minkai Xu, Stefano Ermon, and CUI Bin. Mastering text-to-image diffusion: Recaptioning, planning, and generating with multimodal llms. In *Forty-first International Conference on Machine Learning*, 2024. 2 +[44] Rui Yang, Lin Song, Yanwei Li, Sijie Zhao, Yixiao Ge, Xiu Li, and Ying Shan. Gpt4tools: Teaching large language model to use tools via self-instruction. Advances in Neural Information Processing Systems, 36, 2024. 3 +[45] Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, and Yuan Cao. React: Synergizing reasoning and acting in language models. arXiv preprint arXiv:2210.03629, 2022. 2 +[46] Shunyu Yao, Dian Yu, Jeffrey Zhao, Izhak Shafran, Tom Griffiths, Yuan Cao, and Karthik Narasimhan. Tree of thoughts: Deliberate problem solving with large language models. Advances in Neural Information Processing Systems, 36, 2024. 7 +[47] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 2 +[48] Tianyi Zhang, Varsha Kishore, Felix Wu, Kilian Q Weinberger, and Yoav Artzi. Bertscore: Evaluating text generation with bert. In International Conference on Learning Representations. 3, 6 +[49] Lirui Zhao, Yue Yang, Kaipeng Zhang, Wenqi Shao, Yuxin Zhang, Yu Qiao, Ping Luo, and Rongrong Ji. Diffagent: Fast and accurate text-to-image api selection with large language model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6390-6399, 2024. 2, 3, 4, 6 \ No newline at end of file diff --git a/2025/ChatGen_ Automatic Text-to-Image Generation From FreeStyle Chatting/images.zip b/2025/ChatGen_ Automatic Text-to-Image Generation From FreeStyle Chatting/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..8bd914d00d82516514e6639eab032399a888878f --- /dev/null +++ b/2025/ChatGen_ Automatic Text-to-Image Generation From FreeStyle Chatting/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2f37855c9ecff2f22b4a9786784ff9fab09c2aae713e1d7ae656e928c2885331 +size 672983 diff --git a/2025/ChatGen_ Automatic Text-to-Image Generation From FreeStyle Chatting/layout.json b/2025/ChatGen_ Automatic Text-to-Image Generation From FreeStyle Chatting/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..32ef6b811de868eb42845c6bee34a23f8c769d9d --- /dev/null +++ b/2025/ChatGen_ Automatic Text-to-Image Generation From FreeStyle Chatting/layout.json @@ -0,0 +1,10318 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 69, + 98, + 92, + 126 + ], + "blocks": [ + { + "bbox": [ + 69, + 98, + 92, + 126 + ], + "lines": [ + { + "bbox": [ + 69, + 98, + 92, + 126 + ], + "spans": [ + { + "bbox": [ + 69, + 98, + 92, + 126 + ], + "type": "image", + "image_path": "db1b627fef85d9975afd23e19c76597338f4300abef7e78599868821ef8ab03e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 92, + 111, + 542, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 92, + 111, + 542, + 129 + ], + "spans": [ + { + "bbox": [ + 92, + 111, + 542, + 129 + ], + "type": "text", + "content": "ChatGen: Automatic Text-to-Image Generation From FreeStyle Chatting" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 150, + 547, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 150, + 547, + 194 + ], + "spans": [ + { + "bbox": [ + 67, + 150, + 547, + 194 + ], + "type": "text", + "content": "Chengyou Jia" + }, + { + "bbox": [ + 67, + 150, + 547, + 194 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 67, + 150, + 547, + 194 + ], + "type": "text", + "content": ", Changliang Xia" + }, + { + "bbox": [ + 67, + 150, + 547, + 194 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 67, + 150, + 547, + 194 + ], + "type": "text", + "content": ", Zhuohang Dang" + }, + { + "bbox": [ + 67, + 150, + 547, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 67, + 150, + 547, + 194 + ], + "type": "text", + "content": ", Weijia Wu" + }, + { + "bbox": [ + 67, + 150, + 547, + 194 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 67, + 150, + 547, + 194 + ], + "type": "text", + "content": ", Hangwei Qian" + }, + { + "bbox": [ + 67, + 150, + 547, + 194 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 67, + 150, + 547, + 194 + ], + "type": "text", + "content": ", Minnan Luo" + }, + { + "bbox": [ + 67, + 150, + 547, + 194 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 67, + 150, + 547, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 67, + 150, + 547, + 194 + ], + "type": "text", + "content": "School of Computer Science and Technology, MOEKLINNS Lab, Xi'an Jiaotong University \n" + }, + { + "bbox": [ + 67, + 150, + 547, + 194 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 67, + 150, + 547, + 194 + ], + "type": "text", + "content": "National University of Singapore " + }, + { + "bbox": [ + 67, + 150, + 547, + 194 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 67, + 150, + 547, + 194 + ], + "type": "text", + "content": "CFAR, A\\*STAR" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 96, + 196, + 512, + 221 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 196, + 512, + 221 + ], + "spans": [ + { + "bbox": [ + 96, + 196, + 512, + 221 + ], + "type": "text", + "content": "cp3jia@stu.xjtu.edu.cn, 202066@stu.xjtu.edu.cn, dangzhuohang@stu.xjtu.edu.cn, weijiawu96@gmail.com, qian0045@e.ntu.edu.sg, minnluo@xjtu.edu.cn" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 151, + 248, + 200, + 261 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 248, + 200, + 261 + ], + "spans": [ + { + "bbox": [ + 151, + 248, + 200, + 261 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 275, + 296, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 275, + 296, + 550 + ], + "spans": [ + { + "bbox": [ + 54, + 275, + 296, + 550 + ], + "type": "text", + "content": "Despite the significant advancements in text-to-image (T2I) generative models, users often face a trial-and-error challenge in practical scenarios. This challenge arises from the complexity and uncertainty of tedious steps such as crafting suitable prompts, selecting appropriate models, and configuring specific arguments, making users resort to labor-intensive attempts for desired images. This paper proposes Automatic T2I generation, which aims to automate these tedious steps, allowing users to simply describe their needs in a freestyle chatting way. To systematically study this problem, we first introduce ChatGenBench, a novel benchmark designed for Automatic T2I. It features high-quality paired data with diverse freestyle inputs, enabling comprehensive evaluation of automatic T2I models across all steps. Additionally, recognizing Automatic T2I as a complex multi-step reasoning task, we propose ChatGenEvo, a multi-stage evolution strategy that progressively equips models with essential automation skills. Through extensive evaluation across step-wise accuracy and image quality, ChatGen-Evo significantly enhances performance over various baselines. Our evaluation also uncovers valuable insights for advancing automatic T2I. Our data, code, and models will be publicly available." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 567, + 135, + 578 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 567, + 135, + 578 + ], + "spans": [ + { + "bbox": [ + 56, + 567, + 135, + 578 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 587, + 296, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 587, + 296, + 684 + ], + "spans": [ + { + "bbox": [ + 55, + 587, + 296, + 684 + ], + "type": "text", + "content": "In recent years, text-to-image (T2I) generative models have attracted considerable attention [1, 23, 31, 32]. Building on advancements in large-scale T2I models such as DALL-E [31] and Stable Diffusion [32], the open-source community has significantly expanded the capabilities of T2I generation. Researchers fine-tune open-source models on specialized datasets, resulting in a diverse selection of task-specific models available on platforms like Civitai [6] and Hugging" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 328, + 251, + 542, + 417 + ], + "blocks": [ + { + "bbox": [ + 328, + 251, + 542, + 417 + ], + "lines": [ + { + "bbox": [ + 328, + 251, + 542, + 417 + ], + "spans": [ + { + "bbox": [ + 328, + 251, + 542, + 417 + ], + "type": "image", + "image_path": "5cdd819e119516922bb9791598092e026f4c4d5dd6706a31278adeba8876c65c.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 328, + 419, + 542, + 506 + ], + "blocks": [ + { + "bbox": [ + 328, + 419, + 542, + 506 + ], + "lines": [ + { + "bbox": [ + 328, + 419, + 542, + 506 + ], + "spans": [ + { + "bbox": [ + 328, + 419, + 542, + 506 + ], + "type": "image", + "image_path": "437ff6f487593781e614df16cb07e5a2ac787539257ca89cbe30109cbb0c3f64.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 509, + 555, + 542 + ], + "lines": [ + { + "bbox": [ + 313, + 509, + 555, + 542 + ], + "spans": [ + { + "bbox": [ + 313, + 509, + 555, + 542 + ], + "type": "text", + "content": "Figure 1. Illustration of tedious steps in T2I. Our method can select an appropriate model with suitable prompts and arguments. Note: Arle is a character from the game Genshin Impact." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 544, + 555, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 544, + 555, + 581 + ], + "spans": [ + { + "bbox": [ + 313, + 544, + 555, + 581 + ], + "type": "text", + "content": "Face [8]. This variety provides users with a broader range of options to meet customization needs, facilitating the growing adoption of T2I models in real-world applications." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 312, + 582, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 582, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 312, + 582, + 556, + 713 + ], + "type": "text", + "content": "However, the rapid development of T2I models within the open-source community has also introduced significant challenges for users. When non-experts attempt to create images with specific requirements, they often encounter a trial-and-error process involving several tedious steps. As shown in Figure 1, these steps include crafting suitable prompts, selecting appropriate models, and configuring specific model arguments. The complexity and uncertainty of each step turn the process into an arduous journey, resembling \"mice in a maze\". In real-world scenarios, this iterative process consumes substantial time and resources" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 693, + 136, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 693, + 136, + 703 + ], + "spans": [ + { + "bbox": [ + 66, + 693, + 136, + 703 + ], + "type": "text", + "content": "*Equal Contribution." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "type": "text", + "content": "† Corresponding author." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13284" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "type": "text", + "content": "as users continuously adjust settings to regenerate images. Therefore, we pose the challenging problem: can we automate these labor-intensive steps in T2I generation, allowing users to simply describe their needs in a chatting style and receive desired images effortlessly?" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 134, + 294, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 134, + 294, + 277 + ], + "spans": [ + { + "bbox": [ + 55, + 134, + 294, + 277 + ], + "type": "text", + "content": "Previous attempts, such as BeautifulPrompt [4, 38] for generating high-quality prompts from low-quality ones and DiffAgent [49] for model selection using large language models, have made some progress in addressing these challenges. However, these methods only seek to automate a specific step in Figure 1, lacking comprehensive research on the automation of entire T2I process. Moreover, they fail to support diverse types of freestyle input, leaving them far from real-world scenarios. To bridge these gaps, we aim to develop the model that can accept arbitrary user input, similar to ChatGPT [24], and automatically generate all necessary components for generation, termed as Automatic T2I." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 279, + 295, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 279, + 295, + 422 + ], + "spans": [ + { + "bbox": [ + 55, + 279, + 295, + 422 + ], + "type": "text", + "content": "To this end, we first introduce ChatGenBench, the benchmark specifically designed for this task. ChatGenBench offers a substantial dataset of high-quality paired data from 6,807 customized models. Each data pair comprises a user's freestyle chat input, a refined prompt, an appropriate model, and optimal arguments. This comprehensive step-by-step trail enables step-wise evaluation of automatic T2I models, ensuring both quality assessment of the final image output and precise identification of potential automation bottlenecks. Furthermore, ChatGenBench integrates various types of data, enabling testing with multimodal and historical input to simulate practical scenarios." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 425, + 295, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 425, + 295, + 652 + ], + "spans": [ + { + "bbox": [ + 55, + 425, + 295, + 652 + ], + "type": "text", + "content": "An intuitive approach to achieve the goal is to build supervised fine-tuning (SFT) data and tune multimodal large language models. However, we argue that automated T2I should be viewed as a complex multi-step reasoning problem. Directly predicting outputs leads the model to focus on simple text mappings, rather than developing the diverse skills needed for robust automation. Inspired by OpenAI's o1 [25], we propose ChatGen-Evo with a multi-stage evolution strategy. This approach provides targeted feedback at each stage, progressively equipping models with essential automation skills. First, we construct SFT data to train the model to generate high-quality prompts from freestyle inputs. Next, we augment the model's vocabulary with specialized ModelTokens, enabling effective model selection without affecting other functions. Finally, we guide the model in configuring arguments based on prompts and models selected in previous stages. By decomposing the task into clear stages, the model gains crucial automation skills, delivering outputs aligned with user expectations." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 654, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 295, + 713 + ], + "type": "text", + "content": "We conduct a comprehensive evaluation of various methods on the novel ChatGenBench to study Automatic T2I. ChatGen-Evo significantly outperforms other methods across all metrics, including both step-wise assessments and image quality evaluations. Extensive quantitative and" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 553, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 192 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 192 + ], + "type": "text", + "content": "qualitative results, along with ablation studies, underscore the critical importance of the multi-stage evolution strategy. Additionally, further experiments reveal the impact of different stages on the final T2I output, uncovering valuable insights into the challenges and opportunities in Automatic T2I. We highlight some promising directions, such as exploring the scaling laws of prompt rewriting, improving model selection in few-shot scenarios, and leveraging reasoning approaches to enhance performance. The contributions of this paper are summarized as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 194, + 553, + 373 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 313, + 194, + 553, + 242 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 194, + 553, + 242 + ], + "spans": [ + { + "bbox": [ + 313, + 194, + 553, + 242 + ], + "type": "text", + "content": "- We propose the novel and challenging problem of Automatic T2I generation, aiming to develop the model that can handle user's freestyle chatting and automatically produce all necessary components for image generation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 243, + 553, + 290 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 243, + 553, + 290 + ], + "spans": [ + { + "bbox": [ + 313, + 243, + 553, + 290 + ], + "type": "text", + "content": "- We introduce ChatGenBench, a benchmark specifically designed for Automatic T2I. It includes a comprehensive step-by-step trail for step-wise evaluation, supporting multimodal or historical user inputs." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 290, + 553, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 290, + 553, + 337 + ], + "spans": [ + { + "bbox": [ + 313, + 290, + 553, + 337 + ], + "type": "text", + "content": "- We present ChatGen-Evo, which adopts a multi-stage evolution strategy. By decomposing the task into distinct stages, ChatGen-Evo enables the model to progressively acquire essential Automatic T2I capabilities." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 338, + 553, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 338, + 553, + 373 + ], + "spans": [ + { + "bbox": [ + 313, + 338, + 553, + 373 + ], + "type": "text", + "content": "- Extensive experiments on ChatGenBench not only demonstrate the superiority of ChatGen-Evo but also provide valuable insights for advancing automatic T2I." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 385, + 405, + 397 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 385, + 405, + 397 + ], + "spans": [ + { + "bbox": [ + 313, + 385, + 405, + 397 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 405, + 458, + 417 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 405, + 458, + 417 + ], + "spans": [ + { + "bbox": [ + 313, + 405, + 458, + 417 + ], + "type": "text", + "content": "2.1. Text-to-Image Generation" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 423, + 553, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 423, + 553, + 604 + ], + "spans": [ + { + "bbox": [ + 313, + 423, + 553, + 604 + ], + "type": "text", + "content": "With the advancement of diffusion models [13, 27, 34], text-to-image (T2I) generation [1, 23, 31] has demonstrated exceptional capabilities in high-quality image generation and textual description alignment. Large-scale models such as DALL-E [1] enhance text-image alignment by leveraging the joint feature space of CLIP [30]. Moreover, Stable Diffusion [32], a well-established open-source model, has gained substantial attention. Numerous methods have been developed to fine-tune it or design additional modules for specialized tasks, such as customized image generation [17, 18, 33], layout-to-image generation [14, 19, 20, 47] and image edit [2, 16]. These diverse models have significantly expanded the capabilities of T2I. However, they also present significant learning challenges for non-expert users, underscoring the increasing need for automatic T2I generation." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 611, + 507, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 611, + 507, + 623 + ], + "spans": [ + { + "bbox": [ + 313, + 611, + 507, + 623 + ], + "type": "text", + "content": "2.2. LLMs for Text-to-Image Generation" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 629, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 553, + 713 + ], + "type": "text", + "content": "Large language models (LLMs), like ChatGPT [24], Llama [37], have demonstrated impressive capability in language understanding [3] and problem solving [45]. Recently, LLMs have also been applied to image generation. [7, 9, 20, 36, 43] leverage LLMs to generate detailed layout information from complex prompts, enabling the creation of sub-elements and control over their positioning." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "13285" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 82, + 70, + 526, + 165 + ], + "blocks": [ + { + "bbox": [ + 82, + 70, + 526, + 165 + ], + "lines": [ + { + "bbox": [ + 82, + 70, + 526, + 165 + ], + "spans": [ + { + "bbox": [ + 82, + 70, + 526, + 165 + ], + "type": "image", + "image_path": "47bba521a124c435dac97253b6eec3ba31b070c1ca5af27dc5c30cf25b123934.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 209, + 169, + 399, + 180 + ], + "lines": [ + { + "bbox": [ + 209, + 169, + 399, + 180 + ], + "spans": [ + { + "bbox": [ + 209, + 169, + 399, + 180 + ], + "type": "text", + "content": "Figure 2. Illustration of the data collection pipeline." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 184, + 296, + 257 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 184, + 296, + 257 + ], + "spans": [ + { + "bbox": [ + 55, + 184, + 296, + 257 + ], + "type": "text", + "content": "[28, 39, 49] propose employ LLMs for model selection, akin to tool usage [26, 29, 44]. However, the above methods still require specialized prompt inputs and involve complex model usage. Currently, no approach thoroughly considers leveraging LLMs to relieve users from all tedious steps in T2I. Our method aims to address this gap." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 268, + 138, + 282 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 268, + 138, + 282 + ], + "spans": [ + { + "bbox": [ + 55, + 268, + 138, + 282 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 289, + 296, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 289, + 296, + 408 + ], + "spans": [ + { + "bbox": [ + 55, + 289, + 296, + 408 + ], + "type": "text", + "content": "Our goal is to relieve users from tedious steps and automate to produce the desired images from user's freestyle input. To achieve this objective, we first introduce ChatGenBench, a benchmark comprising a large set of user inputs in a chatting style, built upon a foundation of over 6000 personalized models to evaluate automated image generation results. A comprehensive description of data collection, construction, and analysis is provided in Sec. 3.1. Then, we present ChatGen-Evo in Sec. 3.2, which uses a multi-stage evolution strategy to train MLLM for Automatic T2I." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 418, + 296, + 430 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 418, + 296, + 430 + ], + "spans": [ + { + "bbox": [ + 55, + 418, + 296, + 430 + ], + "type": "text", + "content": "3.1. ChatGenBench: Benchmarking Automatic T2I" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 437, + 296, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 437, + 296, + 498 + ], + "spans": [ + { + "bbox": [ + 55, + 437, + 296, + 498 + ], + "type": "text", + "content": "For clarity, an example of this data is shown on the right side of Figure 2. The following sections detail the data collection and construction process, which primarily involves High-Quality Human Demonstration Collection and using LLM-driven Role-Play to simulate user input." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 514, + 292, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 514, + 292, + 525 + ], + "spans": [ + { + "bbox": [ + 55, + 514, + 292, + 525 + ], + "type": "text", + "content": "3.1.1 High-Quality Human Demonstration Collection" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 533, + 295, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 533, + 295, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 533, + 295, + 605 + ], + "type": "text", + "content": "Civitai [6] is a vibrant community where users share customized AI models for generating high-quality images. Members contribute demonstrations with detailed prompts, model specifications, and arguments, supported by a feedback system that ensures quality validation. These features make Civitai an ideal platform for collecting raw data." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 605, + 295, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 677 + ], + "type": "text", + "content": "Collection: We start by collecting demonstrations based on established evaluation metrics within Civitai, including download counts, upvotes, and other user feedback. These metrics enable us to collect data that has been validated through community engagement. By focusing on these indicators, we identify a subset of high-quality results." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "Filtering: Following the initial collection, we implement a rigorous filtering process to ensure data quality. This involves excluding demonstrations associated with inactive" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 184, + 554, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 184, + 554, + 245 + ], + "spans": [ + { + "bbox": [ + 313, + 184, + 554, + 245 + ], + "type": "text", + "content": "or outdated models, removing duplicates, and filtering out NSFW content. This careful curation refined the dataset to include the most effective demonstrations. Ultimately, this process results in a curated set of 44,881 high-quality human demonstrations across 6,807 unique models." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 263, + 552, + 276 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 263, + 552, + 276 + ], + "spans": [ + { + "bbox": [ + 313, + 263, + 552, + 276 + ], + "type": "text", + "content": "3.1.2 LLM-Driven Role-Play for Chatting Generation" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 283, + 555, + 463 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 283, + 555, + 463 + ], + "spans": [ + { + "bbox": [ + 313, + 283, + 555, + 463 + ], + "type": "text", + "content": "While the demonstrations collected from communities include the essential procedural information needed for automation, they lack a critical element: freestyle chatting inputs. Such data is not available on open platforms, which has been a key limitation for previous methods [49]. To address this, we propose an LLM-Driven Role-Play strategy for Chatting Generation. As shown in Figure 2, we predefine over 100 roles from everyday life (e.g., students, doctors, professors) and prompt the LLM to simulate these roles, translating each demonstration into freestyle chatting input with tones and habits of the character. This approach significantly enhances data diversity and creates more lifelike inputs. To further enhance reverse synthesis diversity, we incorporate multiple LLMs, adjust diversity arguments, and use BERTScore [48] to filter out overly similar inputs." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 464, + 554, + 572 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 464, + 554, + 572 + ], + "spans": [ + { + "bbox": [ + 313, + 464, + 554, + 572 + ], + "type": "text", + "content": "Additionally, we define three types of freestyle input formats: single-input, consisting of a single chatting-style sentence; multimodal-input, combining a sentence with an image; and history-input, comprising multiple rounds of dialogue history. These formats effectively simulate how users typically inquire about image generation needs, greatly expanding the practical value of automatic T2I. An example of the single-input prompt is shown in the following. Additional prompts and details are provided in the Appendix." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 321, + 583, + 480, + 596 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 583, + 480, + 596 + ], + "spans": [ + { + "bbox": [ + 321, + 583, + 480, + 596 + ], + "type": "text", + "content": "Prompt: Sing-Input chatting generation" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 319, + 604, + 548, + 670 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 604, + 548, + 670 + ], + "spans": [ + { + "bbox": [ + 319, + 604, + 548, + 670 + ], + "type": "text", + "content": "System: You are a professional user experience designer who plays various personas to convert complex and professional content for non-professional users. Please merge the following prompt and model information into a single freestyle query. Remove any obvious details that non-professional users would avoid. Make it similar to what non-professional users may write. The converted single-text query should be colloquial and as brief as possible." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 320, + 670, + 547, + 709 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 670, + 547, + 709 + ], + "spans": [ + { + "bbox": [ + 320, + 670, + 547, + 709 + ], + "type": "text", + "content": "Require: {ROLE}, {PROMPT}, {MODEL}, {Example 1,...,n} \nInput: You are playing the {ROLE}. Please generate a single text query based on the following professional {PROMPT} for the {MODEL}. You can refer to following examples: {Example 1,...,n}." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "13286" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 82, + 558, + 181 + ], + "blocks": [ + { + "bbox": [ + 129, + 71, + 479, + 81 + ], + "lines": [ + { + "bbox": [ + 129, + 71, + 479, + 81 + ], + "spans": [ + { + "bbox": [ + 129, + 71, + 479, + 81 + ], + "type": "text", + "content": "Table 1. Comparison of different methods for benchmarking automatic text-to-image generation." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 82, + 558, + 181 + ], + "lines": [ + { + "bbox": [ + 56, + 82, + 558, + 181 + ], + "spans": [ + { + "bbox": [ + 56, + 82, + 558, + 181 + ], + "type": "table", + "html": "
TrainSet (# Numbers)TestSet (# Numbers)Based Models?Step-wise Evaluation?Freestyle User-input?Multi-modal Support?History Support?
BeautifulPrompt [4]143K2KSingleX✓XX
DiffChat [38]234K5KSingleX✓X✓
DiffusionGPT [28]--Multi(≈20)X✓XX
DABench [49]40K5KMulti(5K)XXXX
ChatGenBench256K14KMulti(6K)✓✓✓✓
", + "image_path": "54bbbf3262b5169d39b07cb8d9d5874f34e4671d616a15f81af5c69fd4d597ce.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 56, + 198, + 298, + 294 + ], + "blocks": [ + { + "bbox": [ + 104, + 184, + 246, + 195 + ], + "lines": [ + { + "bbox": [ + 104, + 184, + 246, + 195 + ], + "spans": [ + { + "bbox": [ + 104, + 184, + 246, + 195 + ], + "type": "text", + "content": "Table 2. Summary of dataset statistics." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 198, + 298, + 294 + ], + "lines": [ + { + "bbox": [ + 56, + 198, + 298, + 294 + ], + "spans": [ + { + "bbox": [ + 56, + 198, + 298, + 294 + ], + "type": "table", + "html": "
DatasetTotalSingleM-ModalHistory
TrainSet256,606147,88869,54839,170
TestSet Init74,36442,83820,21411,312
Benchmark14,56411,0111,6681,132
Supervised10,2408,0091,0991,132
Few-Shot4,3243,002569753
", + "image_path": "47060c90baf3a52e8e1e13645835c3d9fa4cbccab62e92b963117c05a67b40f7.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 298, + 197, + 309 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 298, + 197, + 309 + ], + "spans": [ + { + "bbox": [ + 55, + 298, + 197, + 309 + ], + "type": "text", + "content": "3.1.3 Benchmark Construction" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 317, + 295, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 317, + 295, + 389 + ], + "spans": [ + { + "bbox": [ + 55, + 317, + 295, + 389 + ], + "type": "text", + "content": "Following the above steps, we generate 330,970 freestyle inputs from 44,881 demonstrations. Considering the largescale generated data, we don't randomly split a test set for benchmarking. Instead, we carefully select high-quality, non-overlapping samples as the benchmark for evaluation. The selection involves the following steps:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 389, + 296, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 389, + 296, + 449 + ], + "spans": [ + { + "bbox": [ + 55, + 389, + 296, + 449 + ], + "type": "text", + "content": "TestSet Split: We split the data based on model origin. For data associated with the same model, we use BERTScore to assess the similarity between samples. From this, we select about " + }, + { + "bbox": [ + 55, + 389, + 296, + 449 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 55, + 389, + 296, + 449 + ], + "type": "text", + "content": " most semantically distinct data as initial TestSet in Table 2. This maximizes non-overlap with TrainSet." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 449, + 296, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 449, + 296, + 486 + ], + "spans": [ + { + "bbox": [ + 55, + 449, + 296, + 486 + ], + "type": "text", + "content": "TestSet Filtering: Building on the initial TestSet, we perform multiple filtering rounds to ensure the final ChatGen-Bench's quality. This involves the following steps:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 486, + 295, + 581 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 55, + 486, + 295, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 486, + 295, + 510 + ], + "spans": [ + { + "bbox": [ + 55, + 486, + 295, + 510 + ], + "type": "text", + "content": "- Length Filtering: Remove excessively long entries or those with too many dialogue turns for consistency." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 510, + 295, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 295, + 533 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 295, + 533 + ], + "type": "text", + "content": "- Colloquialism Check: Utilize the Spacy [35] module to filter data for alignment with natural, everyday language." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 533, + 295, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 533, + 295, + 556 + ], + "spans": [ + { + "bbox": [ + 55, + 533, + 295, + 556 + ], + "type": "text", + "content": "- LLM-Based Evaluation: Employ LLM to assess and select data that best matches the chatting tone." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 557, + 295, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 557, + 295, + 581 + ], + "spans": [ + { + "bbox": [ + 55, + 557, + 295, + 581 + ], + "type": "text", + "content": "- Manual Verification: Manual verification is conducted to eliminate any inappropriate or irrelevant samples." + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 582, + 295, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 295, + 629 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 295, + 629 + ], + "type": "text", + "content": "Details of each step are provided in the Appendix. Through this rigorous process, we refine the initial 74,364 entries to a final set of 14,564 high-quality, freestyle chatting samples, constituting the novel benchmark." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "text", + "content": "Setting Division: In practical scenarios, new models often have limited demonstration data, underscoring the importance of evaluations under constrained data. Therefore, we further divide ChatGenBench into two settings—Supervised and Few-shot—based on the availability of samples for each model in the TrainSet. The complete data composition is detailed in Table 2." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 185, + 437, + 197 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 185, + 437, + 197 + ], + "spans": [ + { + "bbox": [ + 313, + 185, + 437, + 197 + ], + "type": "text", + "content": "3.1.4 Benchmark Analysis" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 204, + 555, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 204, + 555, + 312 + ], + "spans": [ + { + "bbox": [ + 313, + 204, + 555, + 312 + ], + "type": "text", + "content": "ChatGenBench offers distinct advantages over previous benchmarks, as summarized in Table 1. On the data level, ChatGenBench includes large-scale high-quality data and a broader range of T2I models. Additionally, human demonstrations provide relative ground truth across each step, allowing for step-wise evaluation that pinpoints potential challenges in automation models. ChatGenBench is also the novel benchmark to support multiple input types, making it more aligned with real-world scenarios." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 320, + 505, + 333 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 320, + 505, + 333 + ], + "spans": [ + { + "bbox": [ + 313, + 320, + 505, + 333 + ], + "type": "text", + "content": "3.2. ChatGen: Achieving Automatic T2I" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 339, + 555, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 339, + 555, + 435 + ], + "spans": [ + { + "bbox": [ + 313, + 339, + 555, + 435 + ], + "type": "text", + "content": "Our goal is to train the model capable of processing freestyle user inputs and generating the necessary components for image generation (prompt, model, and argument), thereby achieving Automatic T2I. In this section, we first introduce a baseline method, ChatGen-Base. We then analyze the limitations of this approach and subsequently propose ChatGen-Evo, which leverages the multi-stage evolution strategy to enhance performance." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 450, + 452, + 462 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 450, + 452, + 462 + ], + "spans": [ + { + "bbox": [ + 313, + 450, + 452, + 462 + ], + "type": "text", + "content": "3.2.1 ChatGen-Base with SFT" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 470, + 554, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 470, + 554, + 552 + ], + "spans": [ + { + "bbox": [ + 313, + 470, + 554, + 552 + ], + "type": "text", + "content": "We first apply supervised fine-tuning (SFT) to develop the ChatGen-Base model as an intuitive baseline. Given a set of freestyle chatting inputs " + }, + { + "bbox": [ + 313, + 470, + 554, + 552 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 470, + 554, + 552 + ], + "type": "text", + "content": " (which may include text, images, and historical context) and their corresponding outputs comprising prompt " + }, + { + "bbox": [ + 313, + 470, + 554, + 552 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 313, + 470, + 554, + 552 + ], + "type": "text", + "content": ", model " + }, + { + "bbox": [ + 313, + 470, + 554, + 552 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 470, + 554, + 552 + ], + "type": "text", + "content": ", and argument " + }, + { + "bbox": [ + 313, + 470, + 554, + 552 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 313, + 470, + 554, + 552 + ], + "type": "text", + "content": ", we use the auto-regressive objective to maximize the following loss:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 353, + 564, + 555, + 588 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 564, + 555, + 588 + ], + "spans": [ + { + "bbox": [ + 353, + 564, + 555, + 588 + ], + "type": "interline_equation", + "content": "L _ {s f t} ^ {\\text {b a s e}} = - \\sum_ {t} \\log P _ {\\pi} (p, m, r \\mid c, * _ {< t}). \\tag {1}", + "image_path": "6858ce58755351e84ca136c6cb2e73813c77ce7cd5a3105f1d759b3bc633891e.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 594, + 554, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 554, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 554, + 653 + ], + "type": "text", + "content": "ChatGen-Base satisfies the essential requirements for generating outputs, including handling diverse input modalities and producing corresponding responses. However, even after fine-tuning SOTA open-source MLLMs, the model frequently produces unsatisfactory results." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "text", + "content": "We attribute this to the complexity of automatic T2I, which requires multi-step reasoning. The quality of generated prompts directly impacts model selection, while model selection serves as a prerequisite for argument configuration. Previous studies [21, 25, 40] have shown that directly" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13287" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 85, + 72, + 525, + 231 + ], + "blocks": [ + { + "bbox": [ + 85, + 72, + 525, + 231 + ], + "lines": [ + { + "bbox": [ + 85, + 72, + 525, + 231 + ], + "spans": [ + { + "bbox": [ + 85, + 72, + 525, + 231 + ], + "type": "image", + "image_path": "c815dc7614cd5792197e1fa79c8a33e6fa8f6c92bb255f77bd1a850ba35bee55.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 165, + 236, + 443, + 247 + ], + "lines": [ + { + "bbox": [ + 165, + 236, + 443, + 247 + ], + "spans": [ + { + "bbox": [ + 165, + 236, + 443, + 247 + ], + "type": "text", + "content": "Figure 3. Illustration of the framework for ChatGen-Base and ChatGen-Evo." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 255, + 296, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 255, + 296, + 304 + ], + "spans": [ + { + "bbox": [ + 55, + 255, + 296, + 304 + ], + "type": "text", + "content": "predicting answers often fails in multi-step reasoning tasks. Additionally, supervised fine-tuning tends to encourage the model to learn simple text mappings instead of developing skills needed for automation, leading to poor generalization." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 316, + 148, + 328 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 316, + 148, + 328 + ], + "spans": [ + { + "bbox": [ + 55, + 316, + 148, + 328 + ], + "type": "text", + "content": "3.2.2 ChatGen-Evo" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 335, + 296, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 335, + 296, + 443 + ], + "spans": [ + { + "bbox": [ + 55, + 335, + 296, + 443 + ], + "type": "text", + "content": "To address the limitations of ChatGen-Base, we propose ChatGen-Evo, which trains MLLM " + }, + { + "bbox": [ + 55, + 335, + 296, + 443 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 335, + 296, + 443 + ], + "type": "text", + "content": " using a multi-stage evolution strategy. Instead of relying on final outcome supervision as in traditional fine-tuning, the multi-stage evolution strategy in ChatGen-Evo employs stage supervision. By providing more precise feedback at each stage, this approach gradually enables the MLLM to acquire the necessary capabilities for automated T2I. As shown in Figure 3, the training process includes three main stages:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 445, + 296, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 445, + 296, + 506 + ], + "spans": [ + { + "bbox": [ + 55, + 445, + 296, + 506 + ], + "type": "text", + "content": "Stage 1: Prompt Writing via SFT. We first train the MLLM using SFT with pairs of freestyle inputs and high-quality prompts. Different from the objective in ChatGenbase, the current stage focuses on a more specific and simplified task instead:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 100, + 512, + 295, + 536 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 512, + 295, + 536 + ], + "spans": [ + { + "bbox": [ + 100, + 512, + 295, + 536 + ], + "type": "interline_equation", + "content": "L _ {s f t} ^ {\\text {s t a g e 1}} = - \\sum_ {t} \\log P _ {\\pi} \\left(p \\mid c ^ {\\prime}, * _ {< t}\\right). \\tag {2}", + "image_path": "1605e3994a12463a4d47197631d133366e1d14dcb3725e0e7a166bdd052bbe66.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 543, + 295, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 543, + 295, + 604 + ], + "spans": [ + { + "bbox": [ + 55, + 543, + 295, + 604 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 55, + 543, + 295, + 604 + ], + "type": "inline_equation", + "content": "c'" + }, + { + "bbox": [ + 55, + 543, + 295, + 604 + ], + "type": "text", + "content": " represents the freestyle input with a prompt prefix that clarifies the task being performed. This prefix helps preserve the MLLM's original capabilities, minimizing catastrophic forgetting [22]. Through this stage, the model learns how to rewrite inputs into effective prompts." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "text", + "content": "Stage 2: Model Selection via ModelToken. We introduce the ModelToken strategy to equip the model with model selection capabilities without impacting the prompt-writing ability learned in Stage 1. Inspired by token learning [10, 15], ModelToken extends this approach by representing each candidate T2I model as a unique token within the MLLM's vocabulary. Specifically, the model tokens are parameterized as an embedding matrix " + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "inline_equation", + "content": "W_{\\mathcal{M}} \\in \\mathbb{R}^{|\\mathcal{M}| \\times d}" + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "text", + "content": " and appended to the original word token matrix " + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "inline_equation", + "content": "W_{\\nu} \\in \\mathbb{R}^{|\\mathcal{V}| \\times d}" + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 255, + 555, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 255, + 555, + 304 + ], + "spans": [ + { + "bbox": [ + 313, + 255, + 555, + 304 + ], + "type": "text", + "content": "ModelToken Training: during training, the user input " + }, + { + "bbox": [ + 313, + 255, + 555, + 304 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 255, + 555, + 304 + ], + "type": "text", + "content": " and prompt " + }, + { + "bbox": [ + 313, + 255, + 555, + 304 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 313, + 255, + 555, + 304 + ], + "type": "text", + "content": " are concatenated as a prefix, with the special model token " + }, + { + "bbox": [ + 313, + 255, + 555, + 304 + ], + "type": "inline_equation", + "content": "<\\text{Model}_i>" + }, + { + "bbox": [ + 313, + 255, + 555, + 304 + ], + "type": "text", + "content": " appended as the ground truth for next-token prediction. The training objective is:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 351, + 312, + 555, + 326 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 312, + 555, + 326 + ], + "spans": [ + { + "bbox": [ + 351, + 312, + 555, + 326 + ], + "type": "interline_equation", + "content": "L \\left(W _ {\\mathcal {M}}\\right) = - \\log P _ {\\pi} \\left(< \\text {M o d e l} _ {-} i > | c, p\\right). \\tag {3}", + "image_path": "c93fe344d65fa9d076c3c13dd2f891b749157e7ffe198fe5dcfe8655e2bdd386.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 335, + 554, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 335, + 554, + 395 + ], + "spans": [ + { + "bbox": [ + 313, + 335, + 554, + 395 + ], + "type": "text", + "content": "Unlike typical next-token prediction training, the embedding matrix " + }, + { + "bbox": [ + 313, + 335, + 554, + 395 + ], + "type": "inline_equation", + "content": "W_{\\mathcal{M}}" + }, + { + "bbox": [ + 313, + 335, + 554, + 395 + ], + "type": "text", + "content": " represents the only tunable parameters, significantly enhancing training efficiency. With a small parameter size, fewer training samples are required, improving performance in scenarios with limited data." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 396, + 554, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 396, + 554, + 456 + ], + "spans": [ + { + "bbox": [ + 313, + 396, + 554, + 456 + ], + "type": "text", + "content": "Inference for Model Selection: Once the embedding matrix is trained, the inference process concatenates the model token and original word token, forming the new language modeling head of the MLLM. In this way, the MLLM predicts the next token with the following probability:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 347, + 464, + 553, + 479 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 464, + 553, + 479 + ], + "spans": [ + { + "bbox": [ + 347, + 464, + 553, + 479 + ], + "type": "interline_equation", + "content": "P _ {\\pi} (m \\mid c, p) = \\operatorname {s o f t m a x} \\left(\\left[ W _ {\\nu}; W _ {\\mathcal {M}} \\right] \\cdot h _ {i - 1}\\right), \\tag {4}", + "image_path": "4c0b4dd892169d567eef5f3942325321bde137ec0bda831cd73432b57464eeef.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 482, + 554, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 482, + 554, + 567 + ], + "spans": [ + { + "bbox": [ + 313, + 482, + 554, + 567 + ], + "type": "text", + "content": "where the operation " + }, + { + "bbox": [ + 313, + 482, + 554, + 567 + ], + "type": "inline_equation", + "content": "[;]" + }, + { + "bbox": [ + 313, + 482, + 554, + 567 + ], + "type": "text", + "content": " denotes concatenation, and " + }, + { + "bbox": [ + 313, + 482, + 554, + 567 + ], + "type": "inline_equation", + "content": "h_{i-1} \\in \\mathbb{R}^d" + }, + { + "bbox": [ + 313, + 482, + 554, + 567 + ], + "type": "text", + "content": " represents the last hidden state. Once a model token is predicted, the MLLM stops decoding, and the corresponding model " + }, + { + "bbox": [ + 313, + 482, + 554, + 567 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 482, + 554, + 567 + ], + "type": "text", + "content": " is selected. Additionally, information such as the model's description and demonstrations is loaded for subsequent use. After Stage 2, the model not only retains its prompt-writing skills but also learns to select models." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 570, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 555, + 714 + ], + "type": "text", + "content": "Stage 3: Argument Configuration via In-Context Learning. After the above two stages, we have obtained the prompt " + }, + { + "bbox": [ + 313, + 570, + 555, + 714 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 313, + 570, + 555, + 714 + ], + "type": "text", + "content": " and model " + }, + { + "bbox": [ + 313, + 570, + 555, + 714 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 570, + 555, + 714 + ], + "type": "text", + "content": " from the original user input " + }, + { + "bbox": [ + 313, + 570, + 555, + 714 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 570, + 555, + 714 + ], + "type": "text", + "content": ". The model now needs to generate the appropriate argument configuration to complete the final generation. Due to the careful design of the previous stages, the model's in-context learning ability is maximally preserved. Therefore, we adopt a training-free approach: we guide the MLLM using in-context demonstrations of " + }, + { + "bbox": [ + 313, + 570, + 555, + 714 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 570, + 555, + 714 + ], + "type": "text", + "content": ", the user's original input " + }, + { + "bbox": [ + 313, + 570, + 555, + 714 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 570, + 555, + 714 + ], + "type": "text", + "content": ", and the rewritten prompt " + }, + { + "bbox": [ + 313, + 570, + 555, + 714 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 313, + 570, + 555, + 714 + ], + "type": "text", + "content": ". The MLLM can follow the demonstration pattern to complete the parameter configuration for the current user request:" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13288" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 72, + 294, + 85 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 72, + 294, + 85 + ], + "spans": [ + { + "bbox": [ + 132, + 72, + 294, + 85 + ], + "type": "interline_equation", + "content": "a = M (c, p, D (m)), \\tag {5}", + "image_path": "a973914434ed073a8deb2c0b7885056ecc8223f250b9d3739bd402e7503887ba.jpg" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 88, + 296, + 161 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 88, + 296, + 161 + ], + "spans": [ + { + "bbox": [ + 55, + 88, + 296, + 161 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 88, + 296, + 161 + ], + "type": "inline_equation", + "content": "D(m)" + }, + { + "bbox": [ + 55, + 88, + 296, + 161 + ], + "type": "text", + "content": " represents the set of demonstrations for model " + }, + { + "bbox": [ + 55, + 88, + 296, + 161 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 55, + 88, + 296, + 161 + ], + "type": "text", + "content": " in the training dataset. Thanks to the prior acquisition of the relevant model and prompt in earlier stages, this approach frees up context space, enabling extensive demonstrations. Additionally, the train-free approach avoids interfering with the trained model from the previous two stages." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 171, + 132, + 185 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 171, + 132, + 185 + ], + "spans": [ + { + "bbox": [ + 55, + 171, + 132, + 185 + ], + "type": "text", + "content": "4. Experiment" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 192, + 296, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 192, + 296, + 288 + ], + "spans": [ + { + "bbox": [ + 55, + 192, + 296, + 288 + ], + "type": "text", + "content": "We conduct a comprehensive evaluation of various methods on the novel ChatGenBench. First, in section 4.2, we compare ChatGen-Evo with other baseline models, highlighting the efficacy and efficiency of our multi-stage evolution strategy. Next, in section 4.3, we perform extensive ablation studies, uncovering the impact of each step on the final results and providing valuable insights. Finally, we provide visualizations of the generated images in section 4.4." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 296, + 182, + 308 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 296, + 182, + 308 + ], + "spans": [ + { + "bbox": [ + 55, + 296, + 182, + 308 + ], + "type": "text", + "content": "4.1. Experimental Settings" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 314, + 296, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 314, + 296, + 410 + ], + "spans": [ + { + "bbox": [ + 55, + 314, + 296, + 410 + ], + "type": "text", + "content": "Training Setups. We adopt InternVL2 [5] as the base MLLM, fully fine-tuning it for both ChatGen-Base and the first stage of ChatGen-Evo. In the second stage of ChatGen-Evo, all model parameters are frozen except for the ModelToken embeddings. We employ the AdamW optimizer with a learning rate of 4e-5 and a weight decay of 1.0 over 5 epochs, maintaining these settings consistently across training stages. All experiments are conducted on 8 A100 GPUs." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 416, + 296, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 416, + 296, + 464 + ], + "spans": [ + { + "bbox": [ + 55, + 416, + 296, + 464 + ], + "type": "text", + "content": "Metrics for Step-wise Evaluation: Leveraging the comprehensive process data in ChatGenBench, we introduce the step-wise evaluation metrics to assess the distinct abilities of automatic T2I models in key stages:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 466, + 295, + 634 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 55, + 466, + 295, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 466, + 295, + 550 + ], + "spans": [ + { + "bbox": [ + 55, + 466, + 295, + 550 + ], + "type": "text", + "content": "- Prompt BERTScore: To assess prompt rewriting ability, we use BERTScore [48] to compare predicted prompts with high-quality, human-validated prompts. BERTScore leverages pre-trained contextual embeddings to match words in candidate and reference sentences. The metric ranges from 0 to 1, with 1 indicating highly similar meanings and 0 signifying complete dissimilarity." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 551, + 295, + 585 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 551, + 295, + 585 + ], + "spans": [ + { + "bbox": [ + 55, + 551, + 295, + 585 + ], + "type": "text", + "content": "- Selection Accuracy: We calculate the accuracy of model selection by comparing the predicted T2I model with the human-validated model." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 586, + 295, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 586, + 295, + 634 + ], + "spans": [ + { + "bbox": [ + 55, + 586, + 295, + 634 + ], + "type": "text", + "content": "- Argument Accuracy: We evaluate argument configuration by calculating the exact match accuracy between the predicted and human-validated arguments. The overall accuracy is obtained by averaging individual arguments." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 635, + 296, + 720 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 635, + 296, + 720 + ], + "spans": [ + { + "bbox": [ + 55, + 635, + 296, + 720 + ], + "type": "text", + "content": "It is important to note that obtaining the absolute ground truth across all stages is nearly impossible due to the infinite search space. Therefore, we use human-validated high-quality records as relative ground truth. While these are not absolute, the comprehensive evaluation of the large-scale benchmark is able to reflect the capabilities of automation models, as confirmed by experimental results." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 72, + 554, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 228 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 228 + ], + "type": "text", + "content": "Image Quality Evaluation: We use HPS v2 [41] and ImageReward [42] metrics to assess the quality of generated images, reflecting alignment with human preferences. Additionally, we employ FID and CLIP Score to evaluate how well the generated images meet user requirements. FID [12] measures the distance between automatically generated images and human-validated high-quality images, while CLIP Score [11] calculates the similarity between the generated images and human-validated prompts. To provide an intuitive and comprehensive measure of image quality, we normalize and combine these four metrics into an aggregated score, Unified Metric [49]. Details of each metric and the calculation of Unified Metric are presented in the Appendix." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 235, + 554, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 235, + 554, + 307 + ], + "spans": [ + { + "bbox": [ + 313, + 235, + 554, + 307 + ], + "type": "text", + "content": "Baseline Without Fine-Tuning. We further establish a baseline that uses the default in-context learning capabilities of the MLLM for prompt rewriting, along with a single Stable Diffusion model and a fixed set of default parameters. This baseline helps emphasize the significance of prompt rewriting and multi-model utilization." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 315, + 421, + 327 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 315, + 421, + 327 + ], + "spans": [ + { + "bbox": [ + 313, + 315, + 421, + 327 + ], + "type": "text", + "content": "4.2. Main Experiment" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 333, + 435, + 344 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 333, + 435, + 344 + ], + "spans": [ + { + "bbox": [ + 313, + 333, + 435, + 344 + ], + "type": "text", + "content": "4.2.1 Quantitative Results" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 352, + 554, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 352, + 554, + 496 + ], + "spans": [ + { + "bbox": [ + 313, + 352, + 554, + 496 + ], + "type": "text", + "content": "Table 3 presents the main quantitative results of ChatGenEvo compared to different baselines. Overall, ChatGenEvo significantly outperforms other methods across all metrics, including both step-wise and final image quality evaluations. Specifically, the low performance of the baseline highlights the importance of effective prompt rewriting and multi-model selection, underscoring the necessity of dedicated Automatic T2I methods. Additionally, fine-tuning MLLMs with progressively larger parameter scales yields steady performance improvements. Remarkably, ChatGenEvo achieves performance comparable to ChatGen-Base at 8B, despite utilizing a significantly smaller parameter scale." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 497, + 554, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 497, + 554, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 497, + 554, + 640 + ], + "type": "text", + "content": "The comparisons across different settings provide further insights. Prompt rewriting ability demonstrates strong generalizability, effectively transferring to rarely seen models. The higher few-shot performance can be attributed to the bias in data distribution between the two settings, as reflected in the \"Baseline\" results, which reveal a difference of approximately 0.3. In contrast, model selection and parameter configuration show a noticeable decline in performance under few-shot conditions, highlighting their reliance on more training samples. Therefore, exploring the scaling laws of prompt rewriting and enhancing model selection in few-shot scenarios are promising directions." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 647, + 429, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 647, + 429, + 658 + ], + "spans": [ + { + "bbox": [ + 313, + 647, + 429, + 658 + ], + "type": "text", + "content": "4.2.2 Human Evaluation" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 666, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 666, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 666, + 554, + 714 + ], + "type": "text", + "content": "We conduct a user study using pairwise comparisons to further evaluate ChatGen-Base(8B) and ChatGen-Evo(2B). Users are presented with two images generated from the same input: one by ChatGen-Base and the other by" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "13289" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 69, + 82, + 537, + 242 + ], + "blocks": [ + { + "bbox": [ + 138, + 71, + 470, + 81 + ], + "lines": [ + { + "bbox": [ + 138, + 71, + 470, + 81 + ], + "spans": [ + { + "bbox": [ + 138, + 71, + 470, + 81 + ], + "type": "text", + "content": "Table 3. The Step-wise and Final evaluation results of different methods on ChatGenBench." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 69, + 82, + 537, + 242 + ], + "lines": [ + { + "bbox": [ + 69, + 82, + 537, + 242 + ], + "spans": [ + { + "bbox": [ + 69, + 82, + 537, + 242 + ], + "type": "table", + "html": "
Step-wise EvaluationFinal Evaluation
PromptBertScore ↑SelectionAcc ↑ConfigAcc ↑FIDScore ↓CLIPScore ↑HPSv2 ↑ImageReward ↑UnifiedMetric ↑
SupervisedBaseline0.026--32.764.620.2-34.637.3
ChatGen-Base(2B)0.1840.2060.38421.369.923.52.459.0
ChatGen-Base(4B)0.1970.2300.49020.770.023.41.558.7
ChatGen-Base(8B)0.2080.2640.50920.870.723.94.060.7
ChatGen-Evo (2B)0.2470.3280.53719.172.925.18.965.9
Few-shotBaseline0.055--54.463.420.0-40.229.7
ChatGen-Base(2B)0.2210.1530.34942.869.123.3-4.851.1
ChatGen-Base(4B)0.2360.1710.44841.269.423.4-4.351.9
ChatGen-Base(8B)0.2520.2010.46241.470.623.4-3.152.5
ChatGen-Evo (2B)0.2830.2310.49340.772.525.05.159.2
", + "image_path": "829b109738723a5a2f239391dab4f3dae4bc4bbfea5791856b96fadb51dce82b.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 61, + 291, + 288, + 352 + ], + "blocks": [ + { + "bbox": [ + 55, + 246, + 294, + 290 + ], + "lines": [ + { + "bbox": [ + 55, + 246, + 294, + 290 + ], + "spans": [ + { + "bbox": [ + 55, + 246, + 294, + 290 + ], + "type": "text", + "content": "Table 4. Efficiency comparison. Training efficiency is measured in GPU hours, while inference is expressed as the seconds required to process each data. The training was conducted on 8 A100 GPUs, and inference was performed on a single A100 80GB GPU." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 61, + 291, + 288, + 352 + ], + "lines": [ + { + "bbox": [ + 61, + 291, + 288, + 352 + ], + "spans": [ + { + "bbox": [ + 61, + 291, + 288, + 352 + ], + "type": "table", + "html": "
MethodParamsTrainingInference
ChatGen-Base(2B)2.21B76h1.1s
ChatGen-Base(8B)8.08B240h2.3s
ChatGen-Evo(2B)2.24B100h1.9s
", + "image_path": "5a41d55bb7d1fcaa692c5fb4eae5d98a5b968c4e7562cd3e14fa8d6b98280c7d.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 354, + 295, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 354, + 295, + 403 + ], + "spans": [ + { + "bbox": [ + 55, + 354, + 295, + 403 + ], + "type": "text", + "content": "ChatGen-Evo. It is tasked with selecting the image that better matches the image quality and relevance to the given input. We sample 2,000 image pairs for supervised and 1,000 for few-shot. More details are provided in the Appendix." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 403, + 296, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 403, + 296, + 475 + ], + "spans": [ + { + "bbox": [ + 55, + 403, + 296, + 475 + ], + "type": "text", + "content": "As shown in Figure 4, the human evaluation results are consistent with the quantitative metrics, highlighting that ChatGen-Evo outperforms ChatGen-Base in both image quality and relevance. Additionally, in few-shot settings, ChatGen-Evo demonstrates a higher win rate, showcasing the effectiveness of our approach in data-scarce scenarios." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 481, + 186, + 493 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 481, + 186, + 493 + ], + "spans": [ + { + "bbox": [ + 55, + 481, + 186, + 493 + ], + "type": "text", + "content": "4.2.3 Efficiency Comparison" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 501, + 295, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 501, + 295, + 584 + ], + "spans": [ + { + "bbox": [ + 55, + 501, + 295, + 584 + ], + "type": "text", + "content": "Similar to existing multi-stage reasoning methods [25, 40, 46], ChatGen-Evo does not offer an efficiency advantage over direct prediction approaches like ChatGen-Base. However, the significant performance gains more than compensate for this drawback. As shown in Table 4, ChatGen-Evo achieves the performance level of ChatGen-Base at 8B parameters using only 2B parameters. Therefore, when com" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 58, + 587, + 293, + 698 + ], + "blocks": [ + { + "bbox": [ + 58, + 587, + 293, + 698 + ], + "lines": [ + { + "bbox": [ + 58, + 587, + 293, + 698 + ], + "spans": [ + { + "bbox": [ + 58, + 587, + 293, + 698 + ], + "type": "image", + "image_path": "fb50ae8208316390f900da2ba5d44dcd2b2f869a16ea75cfa84f657be5880e8a.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 57, + 699, + 293, + 710 + ], + "lines": [ + { + "bbox": [ + 57, + 699, + 293, + 710 + ], + "spans": [ + { + "bbox": [ + 57, + 699, + 293, + 710 + ], + "type": "text", + "content": "Figure 4. User study results of ChatGen-Base and ChatGen-Evo." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 316, + 258, + 549, + 380 + ], + "blocks": [ + { + "bbox": [ + 331, + 246, + 536, + 257 + ], + "lines": [ + { + "bbox": [ + 331, + 246, + 536, + 257 + ], + "spans": [ + { + "bbox": [ + 331, + 246, + 536, + 257 + ], + "type": "text", + "content": "Table 5. Ablation experiments on the supervised setting." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 316, + 258, + 549, + 380 + ], + "lines": [ + { + "bbox": [ + 316, + 258, + 549, + 380 + ], + "spans": [ + { + "bbox": [ + 316, + 258, + 549, + 380 + ], + "type": "table", + "html": "
StageMethodsStep Score ↑FID Score ↓Image Reward ↑Unified Metric ↑
Prompt WritingBaseline0.02619.3-13.150.8
Chat-Base0.18421.33.159.7
Chat-Evo0.24717.610.566.8
Model SelectionChat-Base0.20620.59.166.0
Chat-Evo0.55318.216.869.7
Argument SettingChat-Base0.38417.910.469.1
Chat-Evo0.87117.117.970.3
", + "image_path": "bb5b16447b2b1ed2577dccf011c37df621f8af42cc5ff99d20167f8165888191.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 387, + 554, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 387, + 554, + 411 + ], + "spans": [ + { + "bbox": [ + 313, + 387, + 554, + 411 + ], + "type": "text", + "content": "paring efficiency under equal performance, ChatGen-Evo maintains a relative advantage over ChatGen-Base." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 421, + 376, + 434 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 421, + 376, + 434 + ], + "spans": [ + { + "bbox": [ + 313, + 421, + 376, + 434 + ], + "type": "text", + "content": "4.3. Analysis" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 440, + 433, + 452 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 440, + 433, + 452 + ], + "spans": [ + { + "bbox": [ + 313, + 440, + 433, + 452 + ], + "type": "text", + "content": "4.3.1 Capability Analysis." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 460, + 554, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 460, + 554, + 496 + ], + "spans": [ + { + "bbox": [ + 313, + 460, + 554, + 496 + ], + "type": "text", + "content": "We conduct ablation experiments on ChatGenBench to evaluate the contribution of individual steps to final performance by providing ground truth for the other steps." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 498, + 554, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 554, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 554, + 640 + ], + "type": "text", + "content": "Prompt Writing: We first compare the prompt writing capabilities of different methods, along with their final performance. In these experiments, each method's predicted prompt was passed along with the correct model and argument. As shown in Table 5, all methods exhibit noticeable gaps compared to human-validated prompts, emphasizing the complexity of prompt rewriting. Comparisons with the results in Table 3 also highlight the further improvements achieved by selecting the correct model and arguments, particularly for the \"Baseline\" methods. Moreover, variations in prompts significantly influence the final results, highlighting its critical role in Automatic T2I." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 642, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 554, + 713 + ], + "type": "text", + "content": "Model Selection: Table 5 also shows the impact of model selection where human-validated prompt and argument are provided. With well-crafted prompts, ChatGen-Evo demonstrates a substantial performance boost from " + }, + { + "bbox": [ + 313, + 642, + 554, + 713 + ], + "type": "inline_equation", + "content": "32.8\\%" + }, + { + "bbox": [ + 313, + 642, + 554, + 713 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 642, + 554, + 713 + ], + "type": "inline_equation", + "content": "55.3\\%" + }, + { + "bbox": [ + 313, + 642, + 554, + 713 + ], + "type": "text", + "content": ", indicating the strong influence of prompt quality on model selection accuracy. This supports our perspec" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13290" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 86, + 76, + 102, + 91 + ], + "blocks": [ + { + "bbox": [ + 86, + 76, + 102, + 91 + ], + "lines": [ + { + "bbox": [ + 86, + 76, + 102, + 91 + ], + "spans": [ + { + "bbox": [ + 86, + 76, + 102, + 91 + ], + "type": "image", + "image_path": "85cefe667a402dce8fec1cd3fa25ae01cf17bd463dbeca1f804c429ec53d31e1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 103, + 77, + 205, + 91 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 103, + 77, + 205, + 91 + ], + "spans": [ + { + "bbox": [ + 103, + 77, + 205, + 91 + ], + "type": "text", + "content": "Let's craft a fierce anime warrior maiden, ablaze with power and passion!" + } + ] + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 86, + 92, + 126, + 151 + ], + "blocks": [ + { + "bbox": [ + 86, + 92, + 126, + 151 + ], + "lines": [ + { + "bbox": [ + 86, + 92, + 126, + 151 + ], + "spans": [ + { + "bbox": [ + 86, + 92, + 126, + 151 + ], + "type": "image", + "image_path": "d1b7e35f4720dbfde592aeb9ac30469536e22126a27ab9789939dd7293ca57c4.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 91, + 153, + 121, + 159 + ], + "lines": [ + { + "bbox": [ + 91, + 153, + 121, + 159 + ], + "spans": [ + { + "bbox": [ + 91, + 153, + 121, + 159 + ], + "type": "text", + "content": "Default SD" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 127, + 152, + 165, + 159 + ], + "lines": [ + { + "bbox": [ + 127, + 152, + 165, + 159 + ], + "spans": [ + { + "bbox": [ + 127, + 152, + 165, + 159 + ], + "type": "text", + "content": "ChatGen-Base" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 127, + 93, + 167, + 151 + ], + "blocks": [ + { + "bbox": [ + 127, + 93, + 167, + 151 + ], + "lines": [ + { + "bbox": [ + 127, + 93, + 167, + 151 + ], + "spans": [ + { + "bbox": [ + 127, + 93, + 167, + 151 + ], + "type": "image", + "image_path": "5b16fcf44ebd45aa84c978e53d7e1e357eccae4a2e2190570f20e3f876ffc7fe.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 168, + 93, + 208, + 152 + ], + "blocks": [ + { + "bbox": [ + 168, + 93, + 208, + 152 + ], + "lines": [ + { + "bbox": [ + 168, + 93, + 208, + 152 + ], + "spans": [ + { + "bbox": [ + 168, + 93, + 208, + 152 + ], + "type": "image", + "image_path": "3561f8c5ca0c901fe0170e7fd2e59920dd18aa8a121bf0aee12f56541c989756.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 169, + 152, + 207, + 159 + ], + "lines": [ + { + "bbox": [ + 169, + 152, + 207, + 159 + ], + "spans": [ + { + "bbox": [ + 169, + 152, + 207, + 159 + ], + "type": "text", + "content": "ChatGen-Evo" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 214, + 76, + 226, + 91 + ], + "blocks": [ + { + "bbox": [ + 214, + 76, + 226, + 91 + ], + "lines": [ + { + "bbox": [ + 214, + 76, + 226, + 91 + ], + "spans": [ + { + "bbox": [ + 214, + 76, + 226, + 91 + ], + "type": "image", + "image_path": "6b887e4e0a5962454eae8930ee9c586e7439e1492f6a6583134c8e15cf44e381.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 227, + 77, + 350, + 91 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 227, + 77, + 350, + 91 + ], + "spans": [ + { + "bbox": [ + 227, + 77, + 350, + 91 + ], + "type": "text", + "content": "Imagine a fluffy cyberpunk critter in a detailed, artsy, and concept style. Keep the image perfect !" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 214, + 93, + 258, + 152 + ], + "blocks": [ + { + "bbox": [ + 214, + 93, + 258, + 152 + ], + "lines": [ + { + "bbox": [ + 214, + 93, + 258, + 152 + ], + "spans": [ + { + "bbox": [ + 214, + 93, + 258, + 152 + ], + "type": "image", + "image_path": "14c5d1b4a9d4406a2a256f68da248be33ea593b23667e6e8afba99074835a74d.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 221, + 152, + 251, + 159 + ], + "lines": [ + { + "bbox": [ + 221, + 152, + 251, + 159 + ], + "spans": [ + { + "bbox": [ + 221, + 152, + 251, + 159 + ], + "type": "text", + "content": "Default SD" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 260, + 93, + 291, + 152 + ], + "blocks": [ + { + "bbox": [ + 260, + 93, + 291, + 152 + ], + "lines": [ + { + "bbox": [ + 260, + 93, + 291, + 152 + ], + "spans": [ + { + "bbox": [ + 260, + 93, + 291, + 152 + ], + "type": "image", + "image_path": "290d606325f79fc3357ac4a46105c4dc430f254157266ed492f69762cd438b33.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 263, + 152, + 304, + 159 + ], + "lines": [ + { + "bbox": [ + 263, + 152, + 304, + 159 + ], + "spans": [ + { + "bbox": [ + 263, + 152, + 304, + 159 + ], + "type": "text", + "content": "ChatGen-Base" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 307, + 93, + 352, + 152 + ], + "blocks": [ + { + "bbox": [ + 307, + 93, + 352, + 152 + ], + "lines": [ + { + "bbox": [ + 307, + 93, + 352, + 152 + ], + "spans": [ + { + "bbox": [ + 307, + 93, + 352, + 152 + ], + "type": "image", + "image_path": "393eb531804b83a2ebe13550ec5b301f693698fb319682630bf59d312a3907b7.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 309, + 152, + 347, + 159 + ], + "lines": [ + { + "bbox": [ + 309, + 152, + 347, + 159 + ], + "spans": [ + { + "bbox": [ + 309, + 152, + 347, + 159 + ], + "type": "text", + "content": "ChatGen-Evo" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 369, + 76, + 383, + 91 + ], + "blocks": [ + { + "bbox": [ + 369, + 76, + 383, + 91 + ], + "lines": [ + { + "bbox": [ + 369, + 76, + 383, + 91 + ], + "spans": [ + { + "bbox": [ + 369, + 76, + 383, + 91 + ], + "type": "image", + "image_path": "9ff084a696b5424bb23e10dca752975dbbb1895804105ded2f21451ef2e0e129.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "bbox": [ + 383, + 77, + 498, + 91 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 383, + 77, + 498, + 91 + ], + "spans": [ + { + "bbox": [ + 383, + 77, + 498, + 91 + ], + "type": "text", + "content": "Do you reckon we could mix pirates and punk for a car design? Make Something stands out." + } + ] + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 357, + 93, + 402, + 152 + ], + "blocks": [ + { + "bbox": [ + 357, + 93, + 402, + 152 + ], + "lines": [ + { + "bbox": [ + 357, + 93, + 402, + 152 + ], + "spans": [ + { + "bbox": [ + 357, + 93, + 402, + 152 + ], + "type": "image", + "image_path": "0969957ea7d22bbc2d6ad6501d3a77884c078a1453da4b24d8959a7b875653c8.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 364, + 152, + 394, + 159 + ], + "lines": [ + { + "bbox": [ + 364, + 152, + 394, + 159 + ], + "spans": [ + { + "bbox": [ + 364, + 152, + 394, + 159 + ], + "type": "text", + "content": "Default SD" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 404, + 93, + 463, + 152 + ], + "blocks": [ + { + "bbox": [ + 404, + 93, + 463, + 152 + ], + "lines": [ + { + "bbox": [ + 404, + 93, + 463, + 152 + ], + "spans": [ + { + "bbox": [ + 404, + 93, + 463, + 152 + ], + "type": "image", + "image_path": "1396a873ef64093583c0666293ebdb0b195c600982175fcb49fef8d93e4f1b45.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 414, + 152, + 454, + 159 + ], + "lines": [ + { + "bbox": [ + 414, + 152, + 454, + 159 + ], + "spans": [ + { + "bbox": [ + 414, + 152, + 454, + 159 + ], + "type": "text", + "content": "ChatGen-Base" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 466, + 93, + 524, + 152 + ], + "blocks": [ + { + "bbox": [ + 466, + 93, + 524, + 152 + ], + "lines": [ + { + "bbox": [ + 466, + 93, + 524, + 152 + ], + "spans": [ + { + "bbox": [ + 466, + 93, + 524, + 152 + ], + "type": "image", + "image_path": "af069c871ac878976299eb99f1bf14c02057221e57d58e23cb641281321357aa.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 476, + 152, + 513, + 159 + ], + "lines": [ + { + "bbox": [ + 476, + 152, + 513, + 159 + ], + "spans": [ + { + "bbox": [ + 476, + 152, + 513, + 159 + ], + "type": "text", + "content": "ChatGen-Evo" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 86, + 169, + 100, + 184 + ], + "blocks": [ + { + "bbox": [ + 86, + 169, + 100, + 184 + ], + "lines": [ + { + "bbox": [ + 86, + 169, + 100, + 184 + ], + "spans": [ + { + "bbox": [ + 86, + 169, + 100, + 184 + ], + "type": "image", + "image_path": "cf795807719c2c106a1237bf9c663731e8d7f6a7c45cf9dfd2f86eb9da183d9a.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "bbox": [ + 102, + 170, + 200, + 184 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 170, + 200, + 184 + ], + "spans": [ + { + "bbox": [ + 102, + 170, + 200, + 184 + ], + "type": "text", + "content": "I want the final image echoes the reference's shitty, comic-inspired style?" + } + ] + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 86, + 186, + 123, + 224 + ], + "blocks": [ + { + "bbox": [ + 86, + 186, + 123, + 224 + ], + "lines": [ + { + "bbox": [ + 86, + 186, + 123, + 224 + ], + "spans": [ + { + "bbox": [ + 86, + 186, + 123, + 224 + ], + "type": "image", + "image_path": "6d2091d230945f1bec1ad717d37c6d6fef650985b5db25f799525bd4f615ec46.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 89, + 225, + 118, + 239 + ], + "lines": [ + { + "bbox": [ + 89, + 225, + 118, + 239 + ], + "spans": [ + { + "bbox": [ + 89, + 225, + 118, + 239 + ], + "type": "text", + "content": "reference image" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 124, + 186, + 161, + 237 + ], + "blocks": [ + { + "bbox": [ + 124, + 186, + 161, + 237 + ], + "lines": [ + { + "bbox": [ + 124, + 186, + 161, + 237 + ], + "spans": [ + { + "bbox": [ + 124, + 186, + 161, + 237 + ], + "type": "image", + "image_path": "0ed95a1db6d50657d89ebbdf0aca7f177664f23d66b897391f776374c0d53186.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 162, + 186, + 214, + 238 + ], + "blocks": [ + { + "bbox": [ + 123, + 239, + 162, + 244 + ], + "lines": [ + { + "bbox": [ + 123, + 239, + 162, + 244 + ], + "spans": [ + { + "bbox": [ + 123, + 239, + 162, + 244 + ], + "type": "text", + "content": "ChatGen-Base" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 162, + 186, + 214, + 238 + ], + "lines": [ + { + "bbox": [ + 162, + 186, + 214, + 238 + ], + "spans": [ + { + "bbox": [ + 162, + 186, + 214, + 238 + ], + "type": "image", + "image_path": "38a64f804862b3279d45958a82ed66442eccd9eaacdd70fc7251433881dabfbf.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 167, + 239, + 205, + 245 + ], + "lines": [ + { + "bbox": [ + 167, + 239, + 205, + 245 + ], + "spans": [ + { + "bbox": [ + 167, + 239, + 205, + 245 + ], + "type": "text", + "content": "ChatGen-Evo" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 215, + 169, + 228, + 184 + ], + "blocks": [ + { + "bbox": [ + 215, + 169, + 228, + 184 + ], + "lines": [ + { + "bbox": [ + 215, + 169, + 228, + 184 + ], + "spans": [ + { + "bbox": [ + 215, + 169, + 228, + 184 + ], + "type": "image", + "image_path": "1ba0d0a657deb9366dd654ce15de0324a33cf41f53c1e7eb6b345ec2ba5dee00.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + } + ], + "index": 32 + }, + { + "bbox": [ + 229, + 170, + 347, + 184 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 229, + 170, + 347, + 184 + ], + "spans": [ + { + "bbox": [ + 229, + 170, + 347, + 184 + ], + "type": "text", + "content": "Can we create the image similar to the original, with soft colors and intricate details?" + } + ] + } + ], + "index": 33 + }, + { + "type": "image", + "bbox": [ + 216, + 186, + 253, + 239 + ], + "blocks": [ + { + "bbox": [ + 216, + 186, + 253, + 239 + ], + "lines": [ + { + "bbox": [ + 216, + 186, + 253, + 239 + ], + "spans": [ + { + "bbox": [ + 216, + 186, + 253, + 239 + ], + "type": "image", + "image_path": "c486fc45431016f1573560bf876ad90347b5088c0d3682982561124a8485346f.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 260, + 186, + 312, + 239 + ], + "blocks": [ + { + "bbox": [ + 260, + 186, + 312, + 239 + ], + "lines": [ + { + "bbox": [ + 260, + 186, + 312, + 239 + ], + "spans": [ + { + "bbox": [ + 260, + 186, + 312, + 239 + ], + "type": "image", + "image_path": "06f14b95b13cc32fcc019291d49cdd1466168f8aa38e67496e8e0f1d043856fd.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 267, + 239, + 307, + 245 + ], + "lines": [ + { + "bbox": [ + 267, + 239, + 307, + 245 + ], + "spans": [ + { + "bbox": [ + 267, + 239, + 307, + 245 + ], + "type": "text", + "content": "ChatGen-Base" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_caption" + } + ], + "index": 35 + }, + { + "type": "image", + "bbox": [ + 316, + 186, + 369, + 239 + ], + "blocks": [ + { + "bbox": [ + 316, + 186, + 369, + 239 + ], + "lines": [ + { + "bbox": [ + 316, + 186, + 369, + 239 + ], + "spans": [ + { + "bbox": [ + 316, + 186, + 369, + 239 + ], + "type": "image", + "image_path": "5729c35dbffc8d22ac6f74322938166944cff611e72cb2816b3b5c2a92b5eaf2.jpg" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 323, + 239, + 361, + 245 + ], + "lines": [ + { + "bbox": [ + 323, + 239, + 361, + 245 + ], + "spans": [ + { + "bbox": [ + 323, + 239, + 361, + 245 + ], + "type": "text", + "content": "ChatGen-Evo" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_caption" + } + ], + "index": 37 + }, + { + "type": "image", + "bbox": [ + 380, + 169, + 394, + 184 + ], + "blocks": [ + { + "bbox": [ + 380, + 169, + 394, + 184 + ], + "lines": [ + { + "bbox": [ + 380, + 169, + 394, + 184 + ], + "spans": [ + { + "bbox": [ + 380, + 169, + 394, + 184 + ], + "type": "image", + "image_path": "b13d3281eab8b69afdbc068dcc37f9a8f2cceef5654ad67a8b8cfb203ce1a59b.jpg" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_body" + } + ], + "index": 39 + }, + { + "bbox": [ + 395, + 171, + 512, + 184 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 395, + 171, + 512, + 184 + ], + "spans": [ + { + "bbox": [ + 395, + 171, + 512, + 184 + ], + "type": "text", + "content": "Recreate the intricate patterns of the reference, blending realism and dreamlike qualities?" + } + ] + } + ], + "index": 40 + }, + { + "type": "image", + "bbox": [ + 381, + 186, + 417, + 225 + ], + "blocks": [ + { + "bbox": [ + 381, + 186, + 417, + 225 + ], + "lines": [ + { + "bbox": [ + 381, + 186, + 417, + 225 + ], + "spans": [ + { + "bbox": [ + 381, + 186, + 417, + 225 + ], + "type": "image", + "image_path": "f4e7b527ed0afaf40dc7d9d3acf697b920fefdc29be7d50983bdd2187a6db0e2.jpg" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 384, + 226, + 414, + 241 + ], + "lines": [ + { + "bbox": [ + 384, + 226, + 414, + 241 + ], + "spans": [ + { + "bbox": [ + 384, + 226, + 414, + 241 + ], + "type": "text", + "content": "reference image" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_caption" + } + ], + "index": 41 + }, + { + "type": "image", + "bbox": [ + 423, + 186, + 467, + 247 + ], + "blocks": [ + { + "bbox": [ + 423, + 186, + 467, + 247 + ], + "lines": [ + { + "bbox": [ + 423, + 186, + 467, + 247 + ], + "spans": [ + { + "bbox": [ + 423, + 186, + 467, + 247 + ], + "type": "image", + "image_path": "58a91af8787f1b1903dadad9fea8123decbcd558f6fdea4763c8443f32650fdb.jpg" + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_body" + } + ], + "index": 43 + }, + { + "type": "image", + "bbox": [ + 470, + 186, + 517, + 247 + ], + "blocks": [ + { + "bbox": [ + 470, + 186, + 517, + 247 + ], + "lines": [ + { + "bbox": [ + 470, + 186, + 517, + 247 + ], + "spans": [ + { + "bbox": [ + 470, + 186, + 517, + 247 + ], + "type": "image", + "image_path": "8caf608f5a4cb0adad4c1891ad62c2d20a894940a0a8e071ae71dd63c88bfee3.jpg" + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "image_body" + } + ], + "index": 44 + }, + { + "type": "image", + "bbox": [ + 88, + 256, + 100, + 271 + ], + "blocks": [ + { + "bbox": [ + 88, + 256, + 100, + 271 + ], + "lines": [ + { + "bbox": [ + 88, + 256, + 100, + 271 + ], + "spans": [ + { + "bbox": [ + 88, + 256, + 100, + 271 + ], + "type": "image", + "image_path": "4a783e720cbb2bd08ce6a2275ebcb8d1b2f30f9cb6a2c6e3ab9432248166bb13.jpg" + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "image_body" + } + ], + "index": 45 + }, + { + "bbox": [ + 100, + 256, + 158, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 256, + 158, + 277 + ], + "spans": [ + { + "bbox": [ + 100, + 256, + 158, + 277 + ], + "type": "text", + "content": "Hey, you know Flandre Scarlet? Create an image for her!" + } + ] + } + ], + "index": 46 + }, + { + "type": "image", + "bbox": [ + 88, + 289, + 100, + 304 + ], + "blocks": [ + { + "bbox": [ + 88, + 289, + 100, + 304 + ], + "lines": [ + { + "bbox": [ + 88, + 289, + 100, + 304 + ], + "spans": [ + { + "bbox": [ + 88, + 289, + 100, + 304 + ], + "type": "image", + "image_path": "bd70a3ded7549128c9bed057a457e0e4ba9a0038bbcbed63bb561390cf761ff7.jpg" + } + ] + } + ], + "index": 47, + "angle": 0, + "type": "image_body" + } + ], + "index": 47 + }, + { + "bbox": [ + 100, + 288, + 158, + 309 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 288, + 158, + 309 + ], + "spans": [ + { + "bbox": [ + 100, + 288, + 158, + 309 + ], + "type": "text", + "content": "Picture her having a blast in a creek, all cute and nature!" + } + ] + } + ], + "index": 48 + }, + { + "type": "image", + "bbox": [ + 166, + 255, + 219, + 308 + ], + "blocks": [ + { + "bbox": [ + 166, + 255, + 219, + 308 + ], + "lines": [ + { + "bbox": [ + 166, + 255, + 219, + 308 + ], + "spans": [ + { + "bbox": [ + 166, + 255, + 219, + 308 + ], + "type": "image", + "image_path": "a84e8b0ec1b78686ecb1b44cc5c820d14374315e5be4b316d764edfe19e541d5.jpg" + } + ] + } + ], + "index": 49, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 181, + 308, + 205, + 314 + ], + "lines": [ + { + "bbox": [ + 181, + 308, + 205, + 314 + ], + "spans": [ + { + "bbox": [ + 181, + 308, + 205, + 314 + ], + "type": "text", + "content": "Round 1" + } + ] + } + ], + "index": 50, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 319, + 553, + 330 + ], + "lines": [ + { + "bbox": [ + 55, + 319, + 553, + 330 + ], + "spans": [ + { + "bbox": [ + 55, + 319, + 553, + 330 + ], + "type": "text", + "content": "Figure 5. Examples of images generated by different methods. Three rows represent single, multi-modal and historical inputs, respectively." + } + ] + } + ], + "index": 65, + "angle": 0, + "type": "image_caption" + } + ], + "index": 49 + }, + { + "type": "image", + "bbox": [ + 220, + 255, + 290, + 308 + ], + "blocks": [ + { + "bbox": [ + 220, + 255, + 290, + 308 + ], + "lines": [ + { + "bbox": [ + 220, + 255, + 290, + 308 + ], + "spans": [ + { + "bbox": [ + 220, + 255, + 290, + 308 + ], + "type": "image", + "image_path": "6045c2b6cd86a2f94f4cfde6ce2d3e35c0a84fe8c60e349dd95583b76c58e919.jpg" + } + ] + } + ], + "index": 51, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 244, + 308, + 267, + 314 + ], + "lines": [ + { + "bbox": [ + 244, + 308, + 267, + 314 + ], + "spans": [ + { + "bbox": [ + 244, + 308, + 267, + 314 + ], + "type": "text", + "content": "Round 2" + } + ] + } + ], + "index": 52, + "angle": 0, + "type": "image_caption" + } + ], + "index": 51 + }, + { + "type": "image", + "bbox": [ + 294, + 255, + 307, + 270 + ], + "blocks": [ + { + "bbox": [ + 294, + 255, + 307, + 270 + ], + "lines": [ + { + "bbox": [ + 294, + 255, + 307, + 270 + ], + "spans": [ + { + "bbox": [ + 294, + 255, + 307, + 270 + ], + "type": "image", + "image_path": "17715eb3d1aca0dea58a4544b7068cb3ce25fee9b0911b8f2f72b157621c0f60.jpg" + } + ] + } + ], + "index": 53, + "angle": 0, + "type": "image_body" + } + ], + "index": 53 + }, + { + "bbox": [ + 307, + 256, + 369, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 256, + 369, + 270 + ], + "spans": [ + { + "bbox": [ + 307, + 256, + 369, + 270 + ], + "type": "text", + "content": "Creat Van Gogh painting of a human-like possum" + } + ] + } + ], + "index": 54 + }, + { + "type": "image", + "bbox": [ + 294, + 277, + 307, + 285 + ], + "blocks": [ + { + "bbox": [ + 294, + 277, + 307, + 285 + ], + "lines": [ + { + "bbox": [ + 294, + 277, + 307, + 285 + ], + "spans": [ + { + "bbox": [ + 294, + 277, + 307, + 285 + ], + "type": "image", + "image_path": "43ec982290053fb206dab9a247b870ac6b258321432c84ed77c029c40f47cbdd.jpg" + } + ] + } + ], + "index": 55, + "angle": 0, + "type": "image_body" + } + ], + "index": 55 + }, + { + "bbox": [ + 307, + 278, + 371, + 286 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 278, + 371, + 286 + ], + "spans": [ + { + "bbox": [ + 307, + 278, + 371, + 286 + ], + "type": "text", + "content": "It's holding a fruit basket." + } + ] + } + ], + "index": 56 + }, + { + "type": "image", + "bbox": [ + 294, + 296, + 307, + 303 + ], + "blocks": [ + { + "bbox": [ + 294, + 296, + 307, + 303 + ], + "lines": [ + { + "bbox": [ + 294, + 296, + 307, + 303 + ], + "spans": [ + { + "bbox": [ + 294, + 296, + 307, + 303 + ], + "type": "image", + "image_path": "1b59fc1c3840f32a6404b8c5ca1c9d10874d2badbc841edaace16a5ea8e6951c.jpg" + } + ] + } + ], + "index": 57, + "angle": 0, + "type": "image_body" + } + ], + "index": 57 + }, + { + "bbox": [ + 307, + 296, + 374, + 309 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 296, + 374, + 309 + ], + "spans": [ + { + "bbox": [ + 307, + 296, + 374, + 309 + ], + "type": "text", + "content": "With vibrant yellows, deep blue colors for fruits!" + } + ] + } + ], + "index": 58 + }, + { + "type": "image", + "bbox": [ + 384, + 251, + 428, + 308 + ], + "blocks": [ + { + "bbox": [ + 384, + 251, + 428, + 308 + ], + "lines": [ + { + "bbox": [ + 384, + 251, + 428, + 308 + ], + "spans": [ + { + "bbox": [ + 384, + 251, + 428, + 308 + ], + "type": "image", + "image_path": "d41c3d132477d36efd47e16ca0ab03bb8dca4f28012bd3e2f0929fa462f7b278.jpg" + } + ] + } + ], + "index": 59, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 394, + 308, + 417, + 314 + ], + "lines": [ + { + "bbox": [ + 394, + 308, + 417, + 314 + ], + "spans": [ + { + "bbox": [ + 394, + 308, + 417, + 314 + ], + "type": "text", + "content": "Round 1" + } + ] + } + ], + "index": 60, + "angle": 0, + "type": "image_caption" + } + ], + "index": 59 + }, + { + "type": "image", + "bbox": [ + 429, + 251, + 473, + 308 + ], + "blocks": [ + { + "bbox": [ + 429, + 251, + 473, + 308 + ], + "lines": [ + { + "bbox": [ + 429, + 251, + 473, + 308 + ], + "spans": [ + { + "bbox": [ + 429, + 251, + 473, + 308 + ], + "type": "image", + "image_path": "b12bbf252b8ff64368442f406a5e0ab4b6f7673e7c05dd4ae453fe6254374439.jpg" + } + ] + } + ], + "index": 61, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 442, + 308, + 465, + 314 + ], + "lines": [ + { + "bbox": [ + 442, + 308, + 465, + 314 + ], + "spans": [ + { + "bbox": [ + 442, + 308, + 465, + 314 + ], + "type": "text", + "content": "Round 2" + } + ] + } + ], + "index": 62, + "angle": 0, + "type": "image_caption" + } + ], + "index": 61 + }, + { + "type": "image", + "bbox": [ + 471, + 251, + 521, + 308 + ], + "blocks": [ + { + "bbox": [ + 471, + 251, + 521, + 308 + ], + "lines": [ + { + "bbox": [ + 471, + 251, + 521, + 308 + ], + "spans": [ + { + "bbox": [ + 471, + 251, + 521, + 308 + ], + "type": "image", + "image_path": "c293acfb79e559807a2082a70c3b07fe1c0338c865dbfbd22ae987fd03a0c253.jpg" + } + ] + } + ], + "index": 63, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 487, + 308, + 510, + 314 + ], + "lines": [ + { + "bbox": [ + 487, + 308, + 510, + 314 + ], + "spans": [ + { + "bbox": [ + 487, + 308, + 510, + 314 + ], + "type": "text", + "content": "Round 3" + } + ] + } + ], + "index": 64, + "angle": 0, + "type": "image_caption" + } + ], + "index": 63 + }, + { + "bbox": [ + 55, + 334, + 295, + 393 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 334, + 295, + 393 + ], + "spans": [ + { + "bbox": [ + 55, + 334, + 295, + 393 + ], + "type": "text", + "content": "tive that Automatic T2I fundamentally involves multi-step reasoning. In contrast, ChatGen-Base, which directly predicts all results, fails to adapt and thus produces unchanged outputs. Furthermore, these results also emphasize the substantial impact of model selection on overall performance." + } + ] + } + ], + "index": 66 + }, + { + "bbox": [ + 55, + 394, + 295, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 394, + 295, + 477 + ], + "spans": [ + { + "bbox": [ + 55, + 394, + 295, + 477 + ], + "type": "text", + "content": "Argument Configuration: When provided with high-quality prompts and appropriate model selection, ChatGenEvo exhibits notable performance improvements. It improves configuration accuracy from " + }, + { + "bbox": [ + 55, + 394, + 295, + 477 + ], + "type": "inline_equation", + "content": "53.7\\%" + }, + { + "bbox": [ + 55, + 394, + 295, + 477 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 394, + 295, + 477 + ], + "type": "inline_equation", + "content": "87.1\\%" + }, + { + "bbox": [ + 55, + 394, + 295, + 477 + ], + "type": "text", + "content": " and Unified Score from 65.9 to 70.3. Overall, while argument configuration has a relatively smaller impact compared to previous stages, it remains a crucial component." + } + ] + } + ], + "index": 67 + }, + { + "bbox": [ + 55, + 478, + 295, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 478, + 295, + 525 + ], + "spans": [ + { + "bbox": [ + 55, + 478, + 295, + 525 + ], + "type": "text", + "content": "The above findings suggest that outcomes in earlier steps significantly influence predictions in subsequent ones. Therefore, exploring more reasoning methods for advancing automatic T2I represents a promising research direction." + } + ] + } + ], + "index": 68 + }, + { + "bbox": [ + 55, + 537, + 176, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 537, + 176, + 550 + ], + "spans": [ + { + "bbox": [ + 55, + 537, + 176, + 550 + ], + "type": "text", + "content": "4.3.2 Input Type Analysis." + } + ] + } + ], + "index": 69 + }, + { + "bbox": [ + 55, + 556, + 295, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 556, + 295, + 640 + ], + "spans": [ + { + "bbox": [ + 55, + 556, + 295, + 640 + ], + "type": "text", + "content": "Table 6 presents ChatGen-Evo performance across different input types. Multimodal inputs lead to better performance, as images may offer clearer prompt and model identification cues compared to text alone. Additionally, handling historical data remains a significant challenge, with the lowest performance across all metrics. These results point to valuable directions in enhancing history-based prompt generation." + } + ] + } + ], + "index": 70 + }, + { + "bbox": [ + 55, + 647, + 143, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 647, + 143, + 658 + ], + "spans": [ + { + "bbox": [ + 55, + 647, + 143, + 658 + ], + "type": "text", + "content": "4.4. Visualizations" + } + ] + } + ], + "index": 71 + }, + { + "bbox": [ + 55, + 665, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 295, + 713 + ], + "type": "text", + "content": "Figure 5 presents examples of images generated by different methods. In the first row, ChatGen-Evo understands user requirements and identifies suitable models to generate style-matching images. The second row demonstrates" + } + ] + } + ], + "index": 72 + }, + { + "type": "table", + "bbox": [ + 320, + 346, + 545, + 423 + ], + "blocks": [ + { + "bbox": [ + 334, + 333, + 533, + 344 + ], + "lines": [ + { + "bbox": [ + 334, + 333, + 533, + 344 + ], + "spans": [ + { + "bbox": [ + 334, + 333, + 533, + 344 + ], + "type": "text", + "content": "Table 6. The evaluation results of different input types." + } + ] + } + ], + "index": 73, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 320, + 346, + 545, + 423 + ], + "lines": [ + { + "bbox": [ + 320, + 346, + 545, + 423 + ], + "spans": [ + { + "bbox": [ + 320, + 346, + 545, + 423 + ], + "type": "table", + "html": "
Input TypeStep-wise EvaluationFinal
Prompt BertScore ↑Selection Acc ↑Config Acc ↑Unified Score ↑
Single0.2520.3310.53968.1
M-Modal0.2770.3810.55469.3
History0.1650.2590.50560.1
", + "image_path": "70d6a33477080d81ee9a99524c5ae0c4a9b07ccc5ece6234c3a0410e38c2092a.jpg" + } + ] + } + ], + "index": 74, + "angle": 0, + "type": "table_body" + } + ], + "index": 74 + }, + { + "bbox": [ + 313, + 434, + 554, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 434, + 554, + 506 + ], + "spans": [ + { + "bbox": [ + 313, + 434, + 554, + 506 + ], + "type": "text", + "content": "results based on multi-modal user inputs, where ChatGen-Evo shows a superior understanding of the reference image and preserves more details to generate refined outputs. The third row illustrates ChatGen-Evo's capability in handling historical data, ensuring that each step inherits the previous style while making appropriate modifications." + } + ] + } + ], + "index": 75 + }, + { + "bbox": [ + 314, + 523, + 387, + 536 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 523, + 387, + 536 + ], + "spans": [ + { + "bbox": [ + 314, + 523, + 387, + 536 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 76 + }, + { + "bbox": [ + 313, + 546, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 546, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 546, + 555, + 713 + ], + "type": "text", + "content": "Our research aims to automate tedious steps in T2I generation, allowing users to simply describe their needs in a freestyle chatting way. We introduce ChatGenBench for benchmarking Automatic T2I task. It includes high-quality paired data with diverse freestyle user inputs, enabling comprehensive evaluation across all steps. Furthermore, we argue that Automatic T2I should be regarded as a multi-step reasoning task. Consequently, we propose ChatGen-Evo, a multi-stage evolution strategy that progressively equips models with essential automation skills. Extensive evaluations not only demonstrate the superiority of ChatGen-Evo but also provide valuable insights for advancing Automatic T2I. We believe this represents a significant step toward the future of automated generative models." + } + ] + } + ], + "index": 77 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13291" + } + ] + } + ], + "index": 78 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 297, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 297, + 236 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 297, + 236 + ], + "type": "text", + "content": "This work was supported by the National Nature Science Foundation of China (No. 62192781, No. 62272374), the Natural Science Foundation of Shaanxi Province (2024JC-JCQN-62), the National Nature Science Foundation of China (No. 62202367, No. 62250009, No. 62137002), the Key Research and Development Project in Shaanxi Province No. 2022GXLH-01-03, Project of China Knowledge Center for Engineering Science and Technology, and Project of Chinese academy of engineering \"The Online and Offline Mixed Educational Service System for 'The Belt and Road' Training in MOOC China\", and the K. C. Wong Education Foundation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 245, + 115, + 258 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 245, + 115, + 258 + ], + "spans": [ + { + "bbox": [ + 56, + 245, + 115, + 258 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 266, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 266, + 294, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 266, + 294, + 319 + ], + "spans": [ + { + "bbox": [ + 61, + 266, + 294, + 319 + ], + "type": "text", + "content": "[1] James Betker, Gabriel Goh, Li Jing, Tim Brooks, Jianfeng Wang, Linjie Li, Long Ouyang, Juntang Zhuang, Joyce Lee, Yufei Guo, Wesam Manassra, Prafulla Dhariwal, Casey Chu, and Yunxin Jiao. Improving image generation with better captions. 2023. 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 321, + 294, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 321, + 294, + 374 + ], + "spans": [ + { + "bbox": [ + 61, + 321, + 294, + 374 + ], + "type": "text", + "content": "[2] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18392-18402, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 376, + 294, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 376, + 294, + 399 + ], + "spans": [ + { + "bbox": [ + 62, + 376, + 294, + 399 + ], + "type": "text", + "content": "[3] Tom B Brown. Language models are few-shot learners. arXiv preprint arXiv:2005.14165, 2020. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 400, + 294, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 400, + 294, + 443 + ], + "spans": [ + { + "bbox": [ + 62, + 400, + 294, + 443 + ], + "type": "text", + "content": "[4] Tingfeng Cao, Chengyu Wang, Bingyan Liu, Ziheng Wu, Jinhui Zhu, and Jun Huang. Beautifulprompt: Towards automatic prompt engineering for text-to-image synthesis. arXiv preprint arXiv:2311.06752, 2023. 2, 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 445, + 294, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 445, + 294, + 499 + ], + "spans": [ + { + "bbox": [ + 62, + 445, + 294, + 499 + ], + "type": "text", + "content": "[5] Zhe Chen, Weiyun Wang, Hao Tian, Shenglong Ye, Zhangwei Gao, Erfei Cui, Wenwen Tong, Kongzhi Hu, Jiapeng Luo, Zheng Ma, et al. How far are we to gpt-4v? closing the gap to commercial multimodal models with open-source suites. arXiv preprint arXiv:2404.16821, 2024. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 501, + 287, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 501, + 287, + 511 + ], + "spans": [ + { + "bbox": [ + 62, + 501, + 287, + 511 + ], + "type": "text", + "content": "[6] Civitai. Civitai. https://civitai.com/, 2022.1,3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 513, + 294, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 513, + 294, + 566 + ], + "spans": [ + { + "bbox": [ + 62, + 513, + 294, + 566 + ], + "type": "text", + "content": "[7] Runpei Dong, Chunrui Han, Yang Peng, Zekun Qi, Zheng Ge, Jinrong Yang, Liang Zhao, Jianjian Sun, Hongyu Zhou, Haoran Wei, et al. Dreamllm: Synergistic multimodal comprehension and creation. arXiv preprint arXiv:2309.11499, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 568, + 293, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 568, + 293, + 589 + ], + "spans": [ + { + "bbox": [ + 62, + 568, + 293, + 589 + ], + "type": "text", + "content": "[8] Hugging Face. Hugging face. https://huggingface.co/, 2016.1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 590, + 295, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 590, + 295, + 645 + ], + "spans": [ + { + "bbox": [ + 62, + 590, + 295, + 645 + ], + "type": "text", + "content": "[9] Yutong Feng, Biao Gong, Di Chen, Yujun Shen, Yu Liu, and Jingren Zhou. Ranni: Taming text-to-image diffusion for accurate instruction following. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4744-4753, 2024. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 647, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 294, + 689 + ], + "type": "text", + "content": "[10] Shibo Hao, Tianyang Liu, Zhen Wang, and Zhiting Hu. Toolkengpt: Augmenting frozen language models with massive tools via tool embeddings. Advances in neural information processing systems, 36, 2024. 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 294, + 713 + ], + "type": "text", + "content": "[11] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation met" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 333, + 73, + 554, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 554, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 554, + 95 + ], + "type": "text", + "content": "ric for image captioning. In Empirical Methods in Natural Language Processing (EMNLP), 2021. 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 96, + 554, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 554, + 148 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 554, + 148 + ], + "type": "text", + "content": "[12] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 150, + 553, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 150, + 553, + 183 + ], + "spans": [ + { + "bbox": [ + 316, + 150, + 553, + 183 + ], + "type": "text", + "content": "[13] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 185, + 555, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 185, + 555, + 248 + ], + "spans": [ + { + "bbox": [ + 317, + 185, + 555, + 248 + ], + "type": "text", + "content": "[14] Chengyou Jia, Minnan Luo, Zhuohang Dang, Guang Dai, Xiaojun Chang, Mengmeng Wang, and Jingdong Wang. Ssmg: Spatial-semantic map guided diffusion model for free-form layout-to-image generation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2480–2488, 2024. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 251, + 553, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 251, + 553, + 304 + ], + "spans": [ + { + "bbox": [ + 316, + 251, + 553, + 304 + ], + "type": "text", + "content": "[15] Chengyou Jia, Minnan Luo, Zhuohang Dang, Qiushi Sun, Fangzhi Xu, Junlin Hu, Tianbao Xie, and Zhiyong Wu. Agentstore: Scalable integration of heterogeneous agents as specialized generalist computer assistant. arXiv preprint arXiv:2410.18603, 2024. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 305, + 553, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 305, + 553, + 360 + ], + "spans": [ + { + "bbox": [ + 316, + 305, + 553, + 360 + ], + "type": "text", + "content": "[16] Bahjat Kawar, Shiran Zada, Oran Lang, Omer Tov, Huiwen Chang, Tali Dekel, Inbar Mosseri, and Michal Irani. Imagc: Text-based real image editing with diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6007-6017, 2023. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 361, + 553, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 361, + 553, + 415 + ], + "spans": [ + { + "bbox": [ + 317, + 361, + 553, + 415 + ], + "type": "text", + "content": "[17] Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization of text-to-image diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1931-1941, 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 416, + 553, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 416, + 553, + 459 + ], + "spans": [ + { + "bbox": [ + 316, + 416, + 553, + 459 + ], + "type": "text", + "content": "[18] Dongxu Li, Junnan Li, and Steven Hoi. Blip-diffusion: Pretrained subject representation for controllable text-to-image generation and editing. Advances in Neural Information Processing Systems, 36, 2024. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 460, + 553, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 460, + 553, + 514 + ], + "spans": [ + { + "bbox": [ + 317, + 460, + 553, + 514 + ], + "type": "text", + "content": "[19] Yuheng Li, Haotian Liu, Qingyang Wu, Fangzhou Mu, Jianwei Yang, Jianfeng Gao, Chunyuan Li, and Yong Jae Lee. Gligen: Open-set grounded text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22511-22521, 2023. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 515, + 553, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 515, + 553, + 558 + ], + "spans": [ + { + "bbox": [ + 317, + 515, + 553, + 558 + ], + "type": "text", + "content": "[20] Long Lian, Boyi Li, Adam Yala, and Trevor Darrell. Llm-grounded diffusion: Enhancing prompt understanding of text-to-image diffusion models with large language models. arXiv preprint arXiv:2305.13655, 2023. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 559, + 553, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 559, + 553, + 602 + ], + "spans": [ + { + "bbox": [ + 317, + 559, + 553, + 602 + ], + "type": "text", + "content": "[21] Hunter Lightman, Vineet Kosaraju, Yura Burda, Harri Edwards, Bowen Baker, Teddy Lee, Jan Leike, John Schulman, Ilya Sutskever, and Karl Cobbe. Let's verify step by step. arXiv preprint arXiv:2305.20050, 2023. 4" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 604, + 553, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 604, + 553, + 647 + ], + "spans": [ + { + "bbox": [ + 316, + 604, + 553, + 647 + ], + "type": "text", + "content": "[22] Yun Luo, Zhen Yang, Fandong Meng, Yafu Li, Jie Zhou, and Yue Zhang. An empirical study of catastrophic forgetting in large language models during continual fine-tuning. arXiv preprint arXiv:2308.08747, 2023. 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 647, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 712 + ], + "type": "text", + "content": "[23] Alexander Quinn Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. In International Conference on Machine Learning (ICML), 2022. 1, 2" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "13292" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 56, + 72, + 259, + 85 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 259, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 259, + 85 + ], + "type": "text", + "content": "[24] OpenAI. Chatgpt, 2022. Large language model. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 86, + 294, + 118 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 86, + 294, + 118 + ], + "spans": [ + { + "bbox": [ + 57, + 86, + 294, + 118 + ], + "type": "text", + "content": "[25] OpenAI. Learning to reason with llms. https://openai.com/index/learning-to-reason-with-llms/, 2024. Accessed: 2024-11-06. 2, 4, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 120, + 295, + 153 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 120, + 295, + 153 + ], + "spans": [ + { + "bbox": [ + 57, + 120, + 295, + 153 + ], + "type": "text", + "content": "[26] Shishir G Patil, Tianjun Zhang, Xin Wang, and Joseph E Gonzalez. Gorilla: Large language model connected with massive apis. arXiv preprint arXiv:2305.15334, 2023. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 155, + 295, + 198 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 155, + 295, + 198 + ], + "spans": [ + { + "bbox": [ + 57, + 155, + 295, + 198 + ], + "type": "text", + "content": "[27] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4195-4205, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 57, + 201, + 295, + 245 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 201, + 295, + 245 + ], + "spans": [ + { + "bbox": [ + 57, + 201, + 295, + 245 + ], + "type": "text", + "content": "[28] Jie Qin, Jie Wu, Weifeng Chen, Yuxi Ren, Huixia Li, Hefeng Wu, Xuefeng Xiao, Rui Wang, and Shilei Wen. Diffusiongpt: Llm-driven text-to-image generation system. arXiv preprint arXiv:2401.10061, 2024. 3, 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 57, + 247, + 295, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 247, + 295, + 300 + ], + "spans": [ + { + "bbox": [ + 57, + 247, + 295, + 300 + ], + "type": "text", + "content": "[29] Yujia Qin, Shihao Liang, Yining Ye, Kunlun Zhu, Lan Yan, Yaxi Lu, Yankai Lin, Xin Cong, Xiangru Tang, Bill Qian, et al. Toolllm: Facilitating large language models to master 16000+ real-world apis. arXiv preprint arXiv:2307.16789, 2023. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 303, + 295, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 303, + 295, + 369 + ], + "spans": [ + { + "bbox": [ + 56, + 303, + 295, + 369 + ], + "type": "text", + "content": "[30] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 57, + 371, + 294, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 371, + 294, + 415 + ], + "spans": [ + { + "bbox": [ + 57, + 371, + 294, + 415 + ], + "type": "text", + "content": "[31] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 417, + 296, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 417, + 296, + 472 + ], + "spans": [ + { + "bbox": [ + 56, + 417, + 296, + 472 + ], + "type": "text", + "content": "[32] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 474, + 295, + 538 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 474, + 295, + 538 + ], + "spans": [ + { + "bbox": [ + 56, + 474, + 295, + 538 + ], + "type": "text", + "content": "[33] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 22500-22510, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 541, + 295, + 574 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 541, + 295, + 574 + ], + "spans": [ + { + "bbox": [ + 57, + 541, + 295, + 574 + ], + "type": "text", + "content": "[34] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 576, + 295, + 598 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 576, + 295, + 598 + ], + "spans": [ + { + "bbox": [ + 57, + 576, + 295, + 598 + ], + "type": "text", + "content": "[35] SpaCy. spacy: Industrial-strength natural language processing in python. https://spacy.io/, 2015.4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 600, + 231, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 600, + 231, + 611 + ], + "spans": [ + { + "bbox": [ + 57, + 600, + 231, + 611 + ], + "type": "text", + "content": "[36] Omost Team. Omost github page, 2024. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 613, + 295, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 613, + 295, + 667 + ], + "spans": [ + { + "bbox": [ + 57, + 613, + 295, + 667 + ], + "type": "text", + "content": "[37] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "type": "text", + "content": "[38] Jiapeng Wang, Chengyu Wang, Tingfeng Cao, Jun Huang, and Lianwen Jin. Diffchat: Learning to chat with text-to-image synthesis models for interactive image creation. arXiv preprint arXiv:2403.04997, 2024. 2, 4" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 642 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 115 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 115 + ], + "type": "text", + "content": "[39] Zhenyu Wang, Aoxue Li, Zhenguo Li, and Xihui Liu. Genartist: Multimodal llm as an agent for unified image generation and editing. arXiv preprint arXiv:2407.05600, 2024. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 118, + 553, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 118, + 553, + 173 + ], + "spans": [ + { + "bbox": [ + 317, + 118, + 553, + 173 + ], + "type": "text", + "content": "[40] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824-24837, 2022. 4, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 175, + 553, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 175, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 317, + 175, + 553, + 228 + ], + "type": "text", + "content": "[41] Xiaoshi Wu, Yiming Hao, Keqiang Sun, Yixiong Chen, Feng Zhu, Rui Zhao, and Hongsheng Li. Human preference score v2: A solid benchmark for evaluating human preferences of text-to-image synthesis. arXiv preprint arXiv:2306.09341, 2023. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 230, + 553, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 230, + 553, + 283 + ], + "spans": [ + { + "bbox": [ + 316, + 230, + 553, + 283 + ], + "type": "text", + "content": "[42] Jiazheng Xu, Xiao Liu, Yuchen Wu, Yuxuan Tong, Qinkai Li, Ming Ding, Jie Tang, and Yuxiao Dong. Imagereward: Learning and evaluating human preferences for text-to-image generation. arXiv preprint arXiv:2304.05977, 2023.6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 285, + 553, + 340 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 285, + 553, + 340 + ], + "spans": [ + { + "bbox": [ + 316, + 285, + 553, + 340 + ], + "type": "text", + "content": "[43] Ling Yang, Zhaochen Yu, Chenlin Meng, Minkai Xu, Stefano Ermon, and CUI Bin. Mastering text-to-image diffusion: Recaptioning, planning, and generating with multimodal llms. In *Forty-first International Conference on Machine Learning*, 2024. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 342, + 553, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 342, + 553, + 385 + ], + "spans": [ + { + "bbox": [ + 316, + 342, + 553, + 385 + ], + "type": "text", + "content": "[44] Rui Yang, Lin Song, Yanwei Li, Sijie Zhao, Yixiao Ge, Xiu Li, and Ying Shan. Gpt4tools: Teaching large language model to use tools via self-instruction. Advances in Neural Information Processing Systems, 36, 2024. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 386, + 553, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 386, + 553, + 430 + ], + "spans": [ + { + "bbox": [ + 316, + 386, + 553, + 430 + ], + "type": "text", + "content": "[45] Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, and Yuan Cao. React: Synergizing reasoning and acting in language models. arXiv preprint arXiv:2210.03629, 2022. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 431, + 553, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 431, + 553, + 485 + ], + "spans": [ + { + "bbox": [ + 316, + 431, + 553, + 485 + ], + "type": "text", + "content": "[46] Shunyu Yao, Dian Yu, Jeffrey Zhao, Izhak Shafran, Tom Griffiths, Yuan Cao, and Karthik Narasimhan. Tree of thoughts: Deliberate problem solving with large language models. Advances in Neural Information Processing Systems, 36, 2024. 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 487, + 553, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 487, + 553, + 531 + ], + "spans": [ + { + "bbox": [ + 316, + 487, + 553, + 531 + ], + "type": "text", + "content": "[47] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 532, + 553, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 532, + 553, + 575 + ], + "spans": [ + { + "bbox": [ + 316, + 532, + 553, + 575 + ], + "type": "text", + "content": "[48] Tianyi Zhang, Varsha Kishore, Felix Wu, Kilian Q Weinberger, and Yoav Artzi. Bertscore: Evaluating text generation with bert. In International Conference on Learning Representations. 3, 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 576, + 553, + 642 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 576, + 553, + 642 + ], + "spans": [ + { + "bbox": [ + 316, + 576, + 553, + 642 + ], + "type": "text", + "content": "[49] Lirui Zhao, Yue Yang, Kaipeng Zhang, Wenqi Shao, Yuxin Zhang, Yu Qiao, Ping Luo, and Rongrong Ji. Diffagent: Fast and accurate text-to-image api selection with large language model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6390-6399, 2024. 2, 3, 4, 6" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "13293" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/ChatHuman_ Chatting about 3D Humans with Tools/a1962ae6-92af-4828-9f1c-48e34ea37fec_content_list.json b/2025/ChatHuman_ Chatting about 3D Humans with Tools/a1962ae6-92af-4828-9f1c-48e34ea37fec_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..a71fe6dfa31ffb37e2a4016a078ab082f92335fa --- /dev/null +++ b/2025/ChatHuman_ Chatting about 3D Humans with Tools/a1962ae6-92af-4828-9f1c-48e34ea37fec_content_list.json @@ -0,0 +1,1560 @@ +[ + { + "type": "text", + "text": "ChatHuman: Chatting about 3D Humans with Tools", + "text_level": 1, + "bbox": [ + 233, + 130, + 764, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jing Lin $^{3,4,\\star}$ Yao Feng $^{2,3,\\star}$ Weiyang Liu $^{1,5}$ Michael J. Black", + "bbox": [ + 233, + 178, + 761, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Max Planck Institute for Intelligent Systems, Tübingen $^{2}$ Stanford University $^{3}$ Meshcapade $^{4}$ Tsinghua University $^{5}$ University of Cambridge", + "bbox": [ + 191, + 198, + 805, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "chathuman.github.io", + "bbox": [ + 410, + 237, + 581, + 250 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/19ac7f316394b3d5dd28cec3f9e9bb30fde837e0cdd86cfd1d4936421a0a83b6.jpg", + "image_caption": [ + "Figure 1. ChatHuman is a LLM-based agent that uses a multimodal LLM to exploit and combine tools, discriminate their results, and integrate the results to solve tasks related to 3D Humans." + ], + "image_footnote": [], + "bbox": [ + 98, + 263, + 903, + 440 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 497, + 326, + 512 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Numerous methods have been proposed to detect, estimate, and analyze properties of people in images, including 3D pose, shape, contact, human-object interaction, and emotion. While widely applicable in vision and other areas, such methods require expert knowledge to select, use, and interpret the results. To address this, we introduce ChatHuman, a language-driven system that integrates the capabilities of specialized methods into a unified framework. ChatHuman functions as an assistant proficient in utilizing, analyzing, and interacting with tools specific to 3D human tasks, adeptly discussing and resolving related challenges. Built on a Large Language Model (LLM) framework, ChatHuman is trained to autonomously select, apply, and interpret a diverse set of tools in response to user inputs. Our approach overcomes significant hurdles in adapting LLMs to 3D human tasks, including the need for domain-specific knowledge and the ability to interpret complex 3D outputs. The innovations of ChatHuman include leveraging academic publications to instruct the LLM on tool usage, employing a retrieval-augmented generation model to create in-context learning examples for managing new tools, and effectively discriminating between and integrating tool results by transforming", + "bbox": [ + 88, + 530, + 485, + 863 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "specialized 3D outputs into comprehensible formats. Experiments demonstrate that ChatHuman surpasses existing models in both tool selection accuracy and overall performance across various 3D human tasks, and it supports interactive chatting with users. ChatHuman represents a significant step toward consolidating diverse analytical methods into a unified, robust system for 3D human tasks. Code and data are available at chathuman.github.io.", + "bbox": [ + 511, + 500, + 908, + 619 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 633, + 645, + 648 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Research on 3D humans has progressed rapidly, resulting in the creation of many tools that can perform tasks like estimating a human's 3D pose from a single image [18, 25, 27, 36, 37], predicting face/body shapes [8, 15], capturing emotions [9, 15], and identifying regions of touch/contact [46, 68], generating human poses from text descriptions [10], and animating human images [84]. Each of these tools, however, focuses on a specific problem, functioning as isolated \"specialists\". Moreover, these separate tools cannot benefit from the expertise of others, and combining them to solve more complex tasks requires significant domain expertise. Ideally, we would have a single model that can adaptively leverage different tools to solve complex 3D human-related problems while offering intuitive user interaction through natural language input. Recent work such as ChatPose [16] has taken initial steps in this direction, unifying", + "bbox": [ + 509, + 657, + 910, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution. This work was done while YF and JL were at Meshcapade.", + "bbox": [ + 89, + 875, + 485, + 901 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "8150", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ing pose generation, estimation, and general understanding within an LLM framework. Unfortunately, ChatPose lacks the accuracy of the best specialist methods.", + "bbox": [ + 89, + 90, + 480, + 136 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address these issues, we build a multi-modal LLM, called ChatHuman, that specializes in using digital human modeling tools, enabling it to autonomously interpret instructions and complete diverse tasks related to 3D humans; see Fig. 1. Specifically, we teach an LLM to use a wide range of specialized human-related models for tasks like 3D pose estimation, emotion recognition, contact reasoning, and more, effectively extending the LLM's capabilities to the domain of 3D humans. This goes beyond providing a natural-language interface to these tools, as the LLM can use its broad understanding of humans to augment tool results or to analyze and integrate their outputs, providing better responses than any single tool alone.", + "bbox": [ + 88, + 138, + 482, + 335 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "With ChatHuman, we introduce a novel approach by fine-tuning an LLM to act as an agent that autonomously calls appropriate tools based on user inputs, completing tasks and enhancing responses with tool-generated results. Similar in spirit, recent works have employed off-the-shelf or fine-tuned LLMs for tasks like basic vision (e.g., Visual ChatGPT [53]), mobile applications (e.g., AppAgent [77]), biology (e.g., AmadeusGPT [78]) and system automation (e.g., GPT4Tools [74]). Our work, however, differs by focusing specifically on the unique challenges of 3D human understanding. This domain requires precise, specialized terminology and a nuanced understanding of 3D-specific tools, which conventional LLMs lack. To teach the network this specialized terminology, we do what we would do as humans – we have the LLM read the papers describing the methods. Even with that knowledge, the LLM needs to understand the task goals, select an appropriate tool or tools, interpret results, and resolve differing results. These skills are all beyond the abilities of general LLMs.", + "bbox": [ + 91, + 335, + 485, + 625 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address these challenges, we design the following training pipeline: 1) We utilize relevant literature about the tools to familiarize the LLM with domain knowledge, helping it know when and how to use these tools; 2) After using a tool, the LLM evaluates the reliability of the outcome using its \"judgment\" and compares different methods to identify the most reliable results; 3) It combines these results with its general knowledge to create response. This pipeline represents several key innovations, laying a foundation for LLMs to effectively handle complex, tool-driven 3D human tasks.", + "bbox": [ + 89, + 626, + 482, + 777 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Retrieval-Augmented Tool Use: Details about tools are typically present in corresponding academic paper. We give the LLM access to these papers and demonstrate that \"reading the paper\" improves tool use performance. We further analyze which paper sections are most effective for instructing tool use. When encountering a new tool, users often turn to the user guide for assistance. We compile the complete documentation for these tools and utilize a paper-based", + "bbox": [ + 89, + 780, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Retrieval-Augmented Generation (RAG) mechanism to improve the LLM's understanding and management of new tools. This means that, although the LLM has not encountered such tools during fine-tuning, it can still effectively use the tools with the help of the paper-based RAG mechanism. In some cases, tasks require combining multiple tools. To address a broader range of tool usage scenarios, we employ a graph-based invocation system, which includes a node for single-tool use, a chain for sequential tool execution, and a DAG for multi-tool combinations as shown in Fig. 3.", + "bbox": [ + 511, + 90, + 906, + 242 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3D Human-Related Tool Result Integration: Analyzing outputs from tools is crucial, as these outputs, such as body meshes, model parameters (e.g., SMPL pose), or motion sequences, are highly varied and complex. To make these results compatible with our LLM analysis system, we convert them into visual formats that the LLM can easily interpret. Guided by Cognitive Load Theory [63], we present these outputs as multiple-choice options, streamlining the selection process and enhancing the LLM's effectiveness in handling 3D human-related tasks. Combined with the LLM's extensive general knowledge, these integrated results enable it to generate sophisticated responses about 3D humans.", + "bbox": [ + 511, + 243, + 908, + 425 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Specifically, ChatHuman consists of a multimodal LLM LLaVA [41], and 26 tools involving 3D Humans and general vision tasks. The LLM is finetuned to use these tools and incorporate their results. User requests can be in the form of text descriptions, images or other 3D information (if applicable), and the model produces text descriptions, images, or other 3D outputs after tool reasoning. Extensive evaluations demonstrate that ChatHuman not only surpasses previous models in tool-use accuracy but also improves performance on various human-related tasks. It achieves this by reasoning about multiple outputs, evaluating their veracity, and combining them with its own knowledge.", + "bbox": [ + 511, + 425, + 908, + 606 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Summarizing, our key contributions include: (1) a framework that leverages LLMs to interact with users and address human-centric tasks using specialist tools; (2) a scientific-paper-based RAG mechanism that ensures precise tool use by comprehending tool descriptions from scholarly articles, enhancing tool applications and interactions; and (3) the integration of tool outcomes with LLMs, enabling the LLM to effectively explain tool results and interact with users. Additionally, the LLM is fine-tuned to distinguish between optimal and suboptimal tool results, improving overall accuracy. ChatHuman achieves superior performance in tool use and human-related tasks compared with other LLM-based methods or task-specific methods. The code, trained models, and datasets are available for research purposes.", + "bbox": [ + 511, + 608, + 908, + 820 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 513, + 829, + 651, + 844 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3D Humans: Many 3D human analysis tasks leverage parametric models like SMPL [43], SMPL-X [48], or GHUM [73] for the body, BFM [49] or FLAME [35] for", + "bbox": [ + 511, + 854, + 906, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "8151", + "bbox": [ + 482, + 944, + 511, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "faces, and MANO [55] for hands. These models enable the representation of the human body, face, and hands as low-dimensional vectors, facilitating subsequent applications in estimation and generation. Estimation of human pose and shape either relies on optimization methods [4, 24] or regression methods [8, 14, 18, 25, 27, 32, 36, 37, 56, 81], which estimate body shape and pose parameters from a single image. Similarly, face reconstruction methods [12, 15, 66] estimate shape and expression parameters of the face model from single images. The analysis of contact, vital for understanding human-environment interaction and social touch, has seen recent attention [19, 46, 68]. Generative modeling techniques such as PoseScript [10] and PoseFix [11] provide methods for synthesizing and correcting 3D human poses based on textual descriptions, while language-to-3D generation methods [5, 20, 82] facilitate 3D avatar creation. Additionally, numerous methods generate human motions [1, 23, 33, 34, 45, 50, 64, 65], and recent language-to-video models are even able to generate humans moving [3, 84].", + "bbox": [ + 89, + 90, + 483, + 378 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "These basic methods excel in their respective scenarios, but are typically treated in isolation. When mature, such tools are often incorporated into software systems for animators that require significant domain knowledge. In contrast, recent generative models provide language interfaces to image, video, and 3D generation tools, making them accessible to novices. Until recently, such language-based control has not been possible for 3D humans. ChatPose [16] makes a step in this direction, unifying pose generation, estimation, and an LLM's general understanding into one model, but remains limited in its task capabilities. In contrast, our model integrates the performance of 26 3D human-related tasks into a single, LLM-based model. ChatHuman enables non-experts to solve real-world tasks by invoking appropriate tools and adding an extra layer of language-driven understanding that effectively leverages the tool outputs.", + "bbox": [ + 89, + 382, + 483, + 625 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Large Language Models and Tool Use: To expand LLM capabilities without expensive retraining, recent work has focused on enabling them to use specialized tools. In this approach, a tool library is constructed and LLMs act as planners to coordinate tool usage. Various tools have been adopted, e.g., vision modules [53, 62, 76], mobile applications [77], community tools [58], special tools [79], and system tools [74]. However, general-purpose LLMs often lack a deep understanding of specific tools, especially those requiring domain knowledge. To address this, recent work [13, 28, 71] proposes to fintune general-purpose LLMs (e.g., LLaMA [67], LLaVA [40, 42]) with domain-specific tool-use data. Some methods use additional tool documentation to improve accuracy [21, 80], while others compose different tools to accomplish complex tasks [29, 57, 72]. Distinct from previous work, ChatHuman focuses on 3D human tasks through language interaction by leveraging off-the-shelf human-related tools.", + "bbox": [ + 89, + 628, + 486, + 900 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/c8a28cec09c4397d56651e279c93958dabb01518558dd93f02c2e68e2b1e75a8.jpg", + "image_caption": [ + "Figure 2. Paper-based Retrieval-Augmented Tool Use. We feed academic papers describing each tool to GPT-4 to build a document for each tool. During inference, given a user query, a relevant sample is retrieved from the documents and provided to the LLM-based agent as an in-context example to improve tool use accuracy." + ], + "image_footnote": [], + "bbox": [ + 521, + 92, + 903, + 200 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Retrieval Augmented Generation: RAG [17, 31, 83] is a technique to enhance generative tasks by retrieving relevant information from external databases, allowing for continual knowledge update. Here, we design a RAG mechanism to facilitate the use of new unseen tools.", + "bbox": [ + 511, + 292, + 908, + 369 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 513, + 382, + 604, + 398 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Overall Pipeline", + "text_level": 1, + "bbox": [ + 513, + 406, + 673, + 422 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ChatHuman consists of a multimodal LLM $f_{\\phi}(\\cdot)$ , along with a set of 3D human-related functions $f_{\\theta_1}(\\cdot), f_{\\theta_2}(\\cdot), \\ldots$ . These functions serve as tools for various tasks, such as 3D human pose estimation, pose generation, and 3D face reconstruction. Our model takes input text queries $X_{q}$ , images $X_{v}$ , and optionally $X_{m}$ representing other 3D human-related modalities (e.g., SMPL parameters for 3D human poses). Then it invokes tools and integrates their results to generate outputs as text $Y_{t}$ , images $Y_{v}$ , or 3D human-related modalities $Y_{m}$ .", + "bbox": [ + 511, + 430, + 908, + 566 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Retrieval-Augmented Tool Usage", + "text_level": 1, + "bbox": [ + 511, + 575, + 803, + 590 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Teaching LLMs to decide when and how to use tools effectively is a significant challenge. A basic approach [53, 74] might involve including tool usage scenarios and input arguments within the LLM prompt, represented as $Y_{\\mathrm{tool}} = f_{\\phi}(X_q,X_t)$ , where $X_{t}$ denotes tool definitions. However, this approach often falls short for specialized tools, especially given the variety of advanced tools for 3D human tasks. Many tools require background knowledge for correct use and have multiple application scenarios. For instance, the HMR tool [18] may be queried with requests like, \"Can you estimate this person's pose?\", \"What are the SMPL parameters?\", or \"Provide the 3D mesh of this person.\" Capturing all possible usage scenarios succinctly in a prompt is difficult, and as tools proliferate, prompt descriptions become unwieldy. To address these challenges, we introduce paper-based Retrieval-Augmented Generation (RAG) [30] and build a tool graph for tool combination. As shown in Fig. 2, we feed academic papers associated with each tool into GPT-4, prompting it to summarize the tool's functions and generate possible user queries for tool activation. These", + "bbox": [ + 511, + 598, + 908, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "8152", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/a1e29eac086005c243e703ba19fe4add3ee45b7dfba3245a57fa16497403ddf5.jpg", + "image_caption": [ + "Figure 3. Examples of tool use graphs. Tool use patterns include: a Node, which uses a single tool; a Chain, which requires sequential tool execution; and a DAG, which combines multiple tools." + ], + "image_footnote": [], + "bbox": [ + 94, + 90, + 480, + 165 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/426956da5ff2035e46c7cfebbe5860d6a16b611375086f24a36f24a46f749f51.jpg", + "image_caption": [ + "Figure 4. Illustration of tool result integration. After tool invocation, results are transformed into VLM-compatible representations. The VLM then incorporates these transformed outputs to generate accurate, informative responses to user queries." + ], + "image_footnote": [], + "bbox": [ + 96, + 224, + 478, + 358 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "papers, with their rich background and detailed instructions, enable the generation of user queries that cover diverse use cases. By combining these queries with each tool's structured arguments, we compile a document of question-answer pairs for each tool's operation. Figure 2 provides an example from one of these documents. These documents serve as an auxiliary knowledge base during inference, from which we retrieve a relevant example $X_{e}$ in response to a user query $X_{q}$ . The retrieval process matches the text embedding of the query with embeddings in the tool documents using a text embedding model [61]. The retrieved sample is then presented to the agent $f_{\\phi}$ as an in-context learning example:", + "bbox": [ + 89, + 445, + 483, + 628 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nX _ {e} = f _ {r} \\left(X _ {q}\\right), Y _ {\\text {t o o l}} = f _ {\\phi} \\left(X _ {q}, X _ {e}, X _ {t}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 151, + 648, + 483, + 666 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $f_{r}$ is the retrieval function, and $Y_{\\mathrm{tool}}$ is a textual description of the tool invocation, specifying tool selection, names, and input arguments for tool calls.", + "bbox": [ + 89, + 684, + 485, + 729 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Graph-based Tool Invocation. Note that the tool use description $Y_{tool}$ varies depending on task settings, as shown for a single tool case in Fig. 2. However, some complex tasks require combining multiple tools. To handle this, we introduce a graph-based mechanism for tool invocation. We construct a tool graph with three structure types: nodes (single tool calls), chains (tool sequences for dependent tasks), and directed acyclic graphs (DAGs) [59] for complex multibranch operations. For each user query, the model predicts an appropriate tool graph and invokes the tools accordingly. Examples of tool graphs are shown in Fig. 3.", + "bbox": [ + 89, + 734, + 485, + 901 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/f00e8d6c697c375a36deb56f780b917385776a2f1bef2eeddaff0cff60e8a3af.jpg", + "image_caption": [ + "Figure 5. Illustration of tool results discrimination. When multiple plausible tools exist for a task, ChatHuman discriminates and chooses the best result as the final response." + ], + "image_footnote": [], + "bbox": [ + 519, + 93, + 898, + 219 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Tool Result Integration", + "text_level": 1, + "bbox": [ + 511, + 285, + 728, + 301 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "After using tools, integrating their results is essential to effectively engage with users and solve problems. However, outputs from different tools vary widely, appearing as language, images, or vectors (like SMPL poses), which can challenge current multimodal LLMs, such as LLaVA [41], that process only text and images. To utilize these varied results and enhance the LLM's understanding of 3D humans, thereby improving its ability to apply world knowledge to problem-solving, we introduce a tool-conditioned transformation, $\\Psi(\\cdot)$ . As shown in Fig. 4, this transformation converts tool outputs $Y_{m}$ into textual or visual formats the LLM can process. For example, we transform the vertex-wise contact label from DECO [68] into body part-level descriptions using SMPL's [43] vertex-to-part mapping dictionary, and render the mesh generated by PoseScipt [10] into an RGB image using rendering techniques. See Sup. Mat. for more details. The transformed results are then merged with the user query as context for response generation:", + "bbox": [ + 511, + 308, + 908, + 580 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nY _ {t} = f _ {\\phi} \\left(X _ {q}, \\Psi \\left(Y _ {m}\\right)\\right). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 633, + 590, + 906, + 609 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In scenarios where multiple tools can address a request (Fig. 5), we present outcomes as multiple-choice questions, prompting the model to select the most relevant answer:", + "bbox": [ + 511, + 618, + 906, + 664 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nY _ {t} = f _ {\\phi} \\left(X _ {q}, \\Psi \\left(Y _ {m 1}\\right), \\Psi \\left(Y _ {m 2}\\right), \\dots\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 674, + 906, + 691 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $Y_{mi}$ denotes the $i$ -th tool result. Since different tools have different failure modes, this process enables ChatHuman to identify the best method case by case, producing more accurate output than any individual method alone.", + "bbox": [ + 511, + 702, + 908, + 762 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Training Data Construction", + "text_level": 1, + "bbox": [ + 511, + 786, + 764, + 801 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Tool Usage Instruction-following Data. To teach the LLM-based agent to correctly use tools, we construct 90K instruction-response pairs about tool usage. Following GPT4Tools [74], we provide GPT-4 [47] with a textual description of COCO training images [39] and a tool-related prompt containing a tool description. To improve efficiency,", + "bbox": [ + 511, + 810, + 908, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "8153", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/04cef13e70446eb9d50cafc9e26de846c05d88e22127698bd1e3ec71a070a82e.jpg", + "image_caption": [ + "Figure 6. Illustration of our instruction-following data construction pipeline. We construct tool usage and feedback data by providing GPT-4 with various tool-related information, image content, and ground truth labels. Gray text shows some example instructions." + ], + "image_footnote": [], + "bbox": [ + 93, + 80, + 898, + 349 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "we first prompt GPT-4 to summarize paper content, re-articulate tool functions, and enumerate 50 potential user queries for tool activation (see Fig. 6(a)).", + "bbox": [ + 89, + 402, + 483, + 446 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Tool Feedback Instruction-following Data. To help the multimodal LLM model discriminate and integrate the tool results, we construct 88K pairs of instruction-following data based on existing datasets 3DPW [70], MOYO [69], PoseScript [10] and SHAPY [8] (see Fig. 6(b)(c)). Please see Sup. Mat. for more details about data construction.", + "bbox": [ + 89, + 446, + 483, + 537 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Model Training", + "text_level": 1, + "bbox": [ + 89, + 546, + 246, + 561 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Once we have data, we use LoRA [22] to finetune the LLM $f_{\\phi}(\\cdot)$ with the cross entropy loss. More specifically, with the ground truth tool invocation labels $\\hat{Y}_{tool}$ and response label $\\hat{Y}_t$ , we optimize the model using the following objective function: $\\mathcal{L} = \\mathbf{CE}(\\hat{Y}_{tool}, Y_{tool}) + \\mathbf{CE}(\\hat{Y}_t, Y_t)$ , where $\\mathbf{CE}$ denotes the cross-entropy loss. See Sup. Mat. for details.", + "bbox": [ + 89, + 569, + 483, + 662 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments and Results", + "text_level": 1, + "bbox": [ + 89, + 671, + 323, + 688 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Implementation Details", + "text_level": 1, + "bbox": [ + 89, + 696, + 308, + 712 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We use LLaVA-1.5 [41] as the VLM backbone, with CLIP [51] for vision encoding and Vicuna [7] for the LLM backbone. For retrieval, we adopt INSTRUCTOR [61] for text embedding and utilize Chroma's vector similarity searching algorithm to identify relevant examples. To preserve the generalization of the pretrained multi-modal LLM, we use LoRA [22] to perform efficient finetuning, with rank 128 and alpha 256. We implement tool utilization with LangChain [6], which enables automatic parsing of tool names and input parameters, followed by tool invocation Optimization uses AdamW [44], with a learning rate of 2e-4 and weight decay of 0. All models are finetuned over 2 epochs with a mix", + "bbox": [ + 89, + 719, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/fb2b54d2dc0029c0be1c5f3448e02ed55275852958ad8cb250d0d57671fc02c2.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
PerceptionReasoningGeneration
Body Pose Estimation [18]Selective Person Pose Detection [18, 41]Text-to-Pose Generation [10]
Body Shape Measurement [2]Specific Person Shape Measurement [2, 41]Speculative Pose Generation [10, 41]
Hand Pose Estimation [38]Targeted Hand Pose Estimation [38, 41]Text-to-Image Generation [54]
Face Reconstruction [15]Described Person Face Reconstruction [15, 41]Text-based Pose Editing [11]
Human Segmentation [26]Described Person Segmentation [26, 41]Remove Someone From The Photo [26, 41, 54]
HOI Detection [68]Selective Person Contact Estimation [41, 68]Replace Someone From The Photo [26, 41, 54]
Pose Description [10]Visual Question Answering [41]Instruct Image Using Text [54]
Image Caption [41]Text-to-Motion Generation [50]
Motion Capture [60]Text-to-Video Generation [50, 54, 84]
Image-to-Video Generation [50, 84]
", + "bbox": [ + 514, + 392, + 903, + 484 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/9f8f939c4bc3bc2810f5ced72bd5f80279d68dce1b696357d34fd5217f1b11ed.jpg", + "table_caption": [ + "Table 1. ChatHuman supports 26 human-related tools, including 9 perception tools, 10 generation tools, and 7 reasoning tools. Tools in grey are unseen tools that are not included in the training data." + ], + "table_footnote": [], + "table_body": "
MethodSeen ToolsUnseen Tools
SRtSRactSRargsSRIoUSRtSRactSRargsSRIoU
GPT4Tools [75]0.6090.5470.5250.5200.5660.6120.5460.5420.5250.573
GPT4Tools-FT [75]0.8250.7100.6870.6900.7410.9040.8070.6900.7470.800
Visual ChatGPT-3.5 [53]0.4980.3190.2370.2510.7910.5070.3140.2260.2930.803
Visual ChatGPT-4 [53]0.8920.8020.7150.7530.7970.9980.9130.8010.8720.907
ChatHuman1.0000.9740.9500.9700.9750.9990.9670.8930.9540.953
", + "bbox": [ + 514, + 541, + 903, + 628 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 2. Tool use accuracy comparison. Successful rate of thought $(\\mathrm{SR}_{\\mathrm{t}})$ , action $(\\mathrm{SR}_{\\mathrm{act}})$ , arguments $(\\mathrm{SR}_{\\mathrm{args}})$ , execution (SR), and IoU are reported.", + "bbox": [ + 511, + 636, + 906, + 678 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "ture of tool usage, tool feedback, and LLaVA multimodal instruction-tuning data, using 8 Nvidia A100-80G GPUs with the DeepSpeed [52] engine. Unless otherwise specified, we use LLaVA-1.5-7B as the base model for the ablation study. We support 26 tools, as listed in Tab. 1.", + "bbox": [ + 511, + 693, + 906, + 768 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Evaluation on Tool Usage", + "text_level": 1, + "bbox": [ + 511, + 773, + 746, + 789 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Tool Usage Benchmark. To evaluate tool usage accuracy, we construct a validation and test set. The validation set has 1000 samples with the same tools as the training set, while the test set includes 689 samples related to 3 tools unseen during training. Split of seen and unseen tools are detailed in Table 1. Similar to our training data construction, we feed a textual description of COCO validation set image, a", + "bbox": [ + 511, + 795, + 906, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "8154", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/48c2839a53c2d970e59df4baa3a9287a51c1c8cb7b72537477c6f0a9dcac77b4.jpg", + "image_caption": [ + "Figure 7. Visualization of Animation Processing. Left: Conversation between the user and ChatHuman. Right: ChatHuman automatically calls tools to solve the task. ChatHuman can handle tasks beyond the capabilities of individual tools." + ], + "image_footnote": [], + "bbox": [ + 94, + 88, + 477, + 205 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "tool description, and some examples summarized from the tool paper into GPT-4 and prompt it to generate instruction-following data. We use the image description captioned by LLaVA [41] instead of the original image captions to ensure differences between training and test sets. Finally, we manually check all question-answering pairs for accuracy.", + "bbox": [ + 88, + 270, + 482, + 361 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines. We compare our method with Visual ChatGPT [53] and GPT4Tools [75] on the proposed evaluation set and report 5 metrics proposed in GPT4Tools [75]. See Sup. Mat. for details of the metrics. For Visual ChatGPT, we experiment with two versions of GPT: \"gpt-3.5-turbo-1106\" and \"gpt-4-turbo-preview\". For GPT4Tools, we adopt the official pretrained 13B model. For a fair comparison, we also finetune GPT4Tools with our training data using the official training code and obtain a variant, GPT4Tools-FT.", + "bbox": [ + 88, + 362, + 483, + 496 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2 shows that the original GPT4Tools struggles on our benchmark due to mismatches between its tools and our human-centric ones. Visual ChatGPT-4 exhibits impressive tool use accuracy, showing its powerful zero-shot ability to follow a standardized format and use tools accurately. ChatHuman consistently outperforms other methods, particularly with tools not seen in training, thanks to our paper-based RAG mechanism.", + "bbox": [ + 88, + 500, + 483, + 619 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Evaluation on 3D Human Related Tasks", + "text_level": 1, + "bbox": [ + 89, + 626, + 433, + 641 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Character Animation. ChatHuman employs tools for text-to-motion and image-to-video generation. We demonstrate how these tools are utilized to interact with users and reason about motions based on conversations in Fig. 7 and Fig. 1. ChatHuman can also tackle tasks that cannot be resolved with a single tool. For instance, text-to-human video generation poses significant challenges due to the complexity of motion. Therefore, another option is to first generate a motion sequence via text-to-motion generation, then apply a video generation model conditioned on this sequence. The internal processing within ChatHuman, detailing how it analyzes and solves tasks, is visualized in Fig. 7. We also compare our text-to-video generation results with those of Pika1. The qualitative comparisons are shown in Fig. 8.", + "bbox": [ + 89, + 650, + 483, + 862 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/c17b93646c4df1699a3baba4f1aeb049e25b41917a4741d70d287b0983075cff.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Method3DPW [70]RPE Benchmark [16]
MPJPE ↓PA-MPJPE ↓MPJRE ↓MPJPE ↓PA-MPJPE ↓MPJRE ↓
SPIN [27]102.962.910.1244.9107.312.4
HMR 2.0 [18]91.058.49.2225.2105.112.1
LLaVA-S [41]440.8205.421.8490.7207.421.1
LLaVA*-S [41]232.1101.112.8---
GPT4-S [47]322.0136.716.0---
LLaVA-P [41]335.2172.316.5391.5191.917.8
GPT4-P [47]396.5203.418.6---
ChatPose [16]163.681.910.4253.6103.811.7
ChatHuman91.358.79.2147.279.110.3
", + "bbox": [ + 516, + 89, + 903, + 214 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 3. Comparison of vanilla human pose estimation and reasoning-based pose estimation on 3DPW [70] and RPE [16]. LLaVA* is fine-tuned with human keypoint data. \"S\" uses multimodal LLMs for keypoint detection and SMPLify [4] for pose optimization. \"P\" utilizes multimodal LLMs for textual pose descriptions, processed by PoseScript [10] to generate poses. MPJPE (in mm), MP-MPJPE (in mm), and MPJRE ( $\\times 100$ ) are reported.", + "bbox": [ + 511, + 217, + 906, + 315 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/03654e6214846602acdd54266ced783c595f6630aecbfad01c8e70afaf5d4632.jpg", + "image_caption": [ + "Figure 8. Left: Comparison to $\\mathrm{Pika}^1$ on text to video generation. Right: Qualitative comparison with ChatPose [16] and HMR 2.0 [18] on reasoning-based human pose estimation task." + ], + "image_footnote": [], + "bbox": [ + 517, + 325, + 903, + 441 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Pose Estimation. Following ChatPose [16], we evaluate the performance of our method on both classical and reasoning-based pose estimation (RPE) tasks. MPJPE, PA-MPJPE, and MPJRE on the 3DPW [70] and RPE [16] benchmarks are reported. For the reasoning-based pose estimation task, ChatHuman first grounds a human based on a textual description and feeds it into the pose estimation tool to get the result. As shown in Table 3, ChatHuman achieves comparable performance to the task-specific models on the classical pose estimation task. For reasoning-based human pose estimation, which involves both reasoning ability and advanced human pose estimation ability, ChatHuman outperforms both task-specific and multi-modal LLM methods by a large margin $(34.6\\% \\downarrow$ in MPVPE). As shown in Fig. 8, only our method achieves a satisfactory result. The multimodal LLM competitor ChatPose finds the correct person but fails to obtain an accurate pose due to its limited perception ability, while the task-specific tool does not match the correct person due to the lack of reasoning ability. This demonstrates the advantage of ChatHuman, which combines task-specific tool use expertise with the general reasoning ability of an LLM.", + "bbox": [ + 511, + 503, + 908, + 820 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Pose Generation. Here we evaluate the pose generation capability of ChatHuman on the classical text-to-pose generation task and the speculative pose generation task (SPG) [16]. Following previous work [10, 16], we report the text-to-pose recall rate $R^{T2P}$ and pose-to-text recall rate $R^{P2T}$", + "bbox": [ + 511, + 825, + 908, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "1We use the demo available at https://pika.art/ (as of May 2025) to obtain the results.", + "bbox": [ + 89, + 875, + 482, + 898 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "8155", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/d3cf2e7d69cf859eff8e4e40dc817c7432c35bef38b6a682ce6b22587e9db5ef.jpg", + "table_caption": [], + "table_footnote": [ + "Table 4. Comparison of classical and speculative pose generation on PoseScript [10] and SPG [16]. \"P\" denotes using LLMs to rephrase textual pose descriptions, which are then processed by PoseScript [10] to generate poses. Top 5, 10, 20 recall rates are reported." + ], + "table_body": "
MethodPoseScript [10]SPG Benchmark [16]
\\( R^{P2T} \\)↑\\( R^{T2P} \\)↑\\( R^{P2T} \\)↑\\( R^{T2P} \\)↑
PoseScript [10]40.452.365.041.454.165.91.53.56.21.42.35.1
ChatPose [16]17.625.335.828.039.054.43.35.58.23.55.811.0
LLaVA-P [41]------2.14.07.12.13.36.1
GPT4-P [47]------2.74.79.22.75.38.2
ChatHuman41.852.665.142.152.366.53.25.09.93.56.510.6
", + "bbox": [ + 93, + 88, + 480, + 172 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "of the retrieval models trained on real poses and evaluated on generated poses. For the SPG task, ChatHuman first rephrases the indirect pose descriptions into explicit ones and adopts PoseScript (journal version) [10] to generate a pose. As shown in Table 4, our method archives comparable performance to the SOTA methods on both benchmarks. In contrast, the previous LLM-based method, ChatPose, performs poorly on the classical pose generation benchmark, while the task-specific model, PoseScript, lags in the SPG benchmark due to limited reasoning ability.", + "bbox": [ + 88, + 243, + 482, + 393 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Body Shape Measurement. We evaluate the body shape measurement accuracy of ChatHuman. We randomly sample 100 images from the HBW validation set [8] and compare our method with a multimodal LLM, LLaVA [41], and a SOTA body shape estimation method, CLIFF-BEDLAM [2]. For LLaVA and ChatHuman, we ask them the same question to inquire about the height, weight, chest, waist, and hip circumferences of a person and then prompt GPT-3.5 to extract the value from the model output. The details of the question and prompt are available in Sup. Mat. CLIFF-BEDLAM predicts the body shape parameters, which are then converted to measurements based on the shape-to-measurement function from SHAPY [8]. Anthropometric measurement errors are reported in Table 5(a). ChatHuman achieves superior performance in most measurements, outperforming the multimodal LLM competitor LLaVA by $42\\%$ and CLIFF-BEDLAM by $15.7\\%$ in average metric accuracy.", + "bbox": [ + 89, + 398, + 482, + 655 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Human-Object Interaction (HoI). We evaluate the human-object interaction understanding ability of ChatHuman on the DECO [68] test set. The ground truth (GT) labels are obtained by converting the vertex-level contact labels into body part-level contact labels with SMPL's vertex-to-part mapping dictionary. Given a human image, we ask the multimodal LLM to detect the body parts contacting objects and prompt GPT-3.5 to extract the body part labels from the answer. Subsequently, we compare the predicted body parts with the GT label and compute the average detection precision, recall rate, and F1 Score. Table 5(b) shows that ChatHuman achieves SOTA precision and F1 score, demonstrating superior human-object interaction understanding ability. Notably, although LLaVA has a high recall rate, its precision and F1 score are rather poor, which means that it tends to predict all the body parts to be in contact with objects.", + "bbox": [ + 89, + 659, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/cd4187b31d0f42024c3a4a71201ada83c787e56b6a7800aeb6998383466528f7.jpg", + "table_caption": [], + "table_footnote": [ + "Table 5. Comparison of body shape and HOI estimation." + ], + "table_body": "
MethodHeight ↓Chest ↓Waist ↓Hip ↓
LLaVA [41]6.716.522.917.6
CLIFF-BEDLAM [2]7.88.613.57.0
ChatHuman6.76.113.06.4
(a) Comparison of body shape measurement. Measurement errors (in cm) on HBW [8] are reported.
", + "bbox": [ + 514, + 89, + 717, + 174 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/4333de6ee1269e1a100485f6381d0119ac836b07550f4369fa2a56bcdc4931fa.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodPrecision↑ Recall↑ F1Score ↑
LLaVA [41]0.260.810.39
GPT-4 [47]0.610.480.49
ChatHuman0.670.670.63
(b) Comparison of HOI estimation. Precision, Recall, and F1 score on DECO [68] are reported.
", + "bbox": [ + 723, + 89, + 900, + 174 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Ablation Study", + "text_level": 1, + "bbox": [ + 511, + 204, + 666, + 220 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Paper-based RAG Mechanism. To improve tool use accuracy, we design a paper-based RAG mechanism. We perform a breakdown ablation to investigate the effect of each component and their interactions. The baseline model, created by removing the RAG operation and trained with instruction-following data without paper content, is compared in Table 6(a). The baseline model's success rate (SR) is 0.96 for seen tools and 0.82 for unseen tools. Adding RAG increases the SR for unseen tools to 0.89, demonstrating its effectiveness in zero-shot settings. Further incorporating articles into training data boosts the performance: the successful rate of arguments $(\\mathrm{SR}_{\\mathrm{args}})$ rises from 0.93 to 0.95 for the seen tools and 0.84 to 0.94 for the unseen tools. This suggests that scholarly articles can help create high-quality instruction-following data and tool documents due to their detailed use instructions and diverse application scenarios. For a detailed analysis of each component's effect on instructing tool usage, please see Sup. Mat.", + "bbox": [ + 511, + 229, + 906, + 500 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Multiple Tools Invocation. One of the advantages of using a VLM as an agent is its powerful generalization capacity. To test the robustness and generalization ability of ChatHuman, we conduct the following ablation study. During training, we only include the tool graphs with no more than three tools, while during evaluation, the user queries might need up to five tools to solve. Table 6(b) depicts the results. As shown, ChatHuman exhibits an excellent robustness in this out-of-domain setting (more than three tools combination) with an action accuracy higher than $90\\%$ .", + "bbox": [ + 511, + 502, + 908, + 654 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Tool Result Integration. We first conduct an ablation to study how the tools can enhance the human understanding capacity of multimodal LLM. The model without tools is our multimodal LLM backbone, LLaVA-1.5-7B [41], and the model with tools is our ChatHuman. The quantitative results are listed in Table 7. When equipped with tools, the HOI contact detection F1 score increases from 0.39 to 0.63 and the average body shape measurement error declines by $38\\%$ . These results demonstrate the effectiveness of tools in enriching the LLM's comprehension of human models and behaviors. Additionally, we study whether ChatHuman can utilize its world knowledge to discriminate and improve the tool performance. We design two discrimination schemes, i.e., selection and modification, and conduct an ablation study on two human-related tasks by comparing ChatHuman with the SOTA task-specific tools. For the selection scheme, we ex", + "bbox": [ + 511, + 657, + 908, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "8156", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/1e6e12559e9ae42a8516043f9ab9450827d2e7d7ef54bb6535f7f36eb76f920f.jpg", + "table_caption": [], + "table_footnote": [ + "(a) Ablation study of paper-based RAG mechanism." + ], + "table_body": "
PaperRAGSeen ToolsUnseen Tools
SRtSRactSRargsSRIoUSRtSRactSRargsSRIoU
××0.9980.9670.9280.9600.9640.9460.8940.7750.8220.872
×✓1.0000.9670.9280.9610.9650.9960.9450.8420.8910.927
✓✓1.0000.9740.9500.9700.9750.9990.9670.8930.9540.953
", + "bbox": [ + 135, + 88, + 537, + 160 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/59c9cd37d004dc0341c7ce6e8f0ae452ed8a572b06e45904f833380c5a8ca8a6.jpg", + "table_caption": [], + "table_footnote": [ + "(b) Ablation study of tool number.", + "Table 6. Ablations related to tool usage. Successful rate of thought, action, arguments, execution, and IoU are reported." + ], + "table_body": "
Tool NumerSRtSRactSRargsSRIoU
20.9970.9600.9430.9280.973
30.9980.9590.9310.9320.974
40.9980.9430.9280.8750.968
50.9960.9290.8990.8470.950
", + "bbox": [ + 593, + 92, + 857, + 160 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/f6024c8c892d5c29b7079fb047c6203a9a686c38257206cbe95f87aae6b3ef06.jpg", + "table_caption": [], + "table_footnote": [ + "(a) HOI Contact Detection.", + "(b) Body Shape Measurement." + ], + "table_body": "
MethodPrecision↑Recall↑F1Score↑MethodHeight↓Chest↓Waist↓Hip↓
w/o Tool0.260.810.39w/o Tool6.716.522.917.6
w/ Tool0.670.670.63w/ Tool6.76.113.06.4
", + "bbox": [ + 91, + 205, + 478, + 246 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/cbae583db8abd3c1c334636552c8c54224cd698ad609d1a3fba89e43cc6bbba4.jpg", + "table_caption": [], + "table_footnote": [ + "(a) Mesh Error (in mm) on MixPose." + ], + "table_body": "
MethodMPJPE ↓PA-MPJPE ↓PA-MPVPE ↓
Tool A126.281.4101.9
Tool B124.084.6104.7
ChatHuman119.678.298.3
", + "bbox": [ + 91, + 294, + 287, + 342 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/8e3783a73a94862f6019a2977a0b3b84ccef7f178846c54ec67e7bc3ff80149e.jpg", + "table_caption": [ + "Table 7. Ablation study on the impact of tool usage for human-object contact detection and body shape estimation." + ], + "table_footnote": [ + "(b) Body Shape Measurement Error (in cm) on HBW [8]." + ], + "table_body": "
MethodHeight ↓Chest ↓Waist ↓Hip ↓
Tool [2]7.88.613.57.0
ChatHuman6.76.113.06.4
", + "bbox": [ + 285, + 292, + 477, + 330 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 8. Study revealing how tool use improves human understanding on pose estimation and body shape measurement tasks.", + "bbox": [ + 91, + 358, + 483, + 386 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5392627be968c7daa2879649bdbc37834a41ec515c7e336dfb41dd9a2384464d.jpg", + "image_caption": [ + "(a) Tool generates unreasonable output", + "Figure 9. Illustration of how ChatHuman discriminates and integrates tool results. The Agent will fix the unreasonable tool result and integrate the reasonable tool result to generate a final response." + ], + "image_footnote": [], + "bbox": [ + 94, + 388, + 279, + 460 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/728061aad1464158687002c1bc86383ed210f88f91d32c7b96b015e60ae2676f.jpg", + "image_caption": [ + "(b) Tool generates reasonable output" + ], + "image_footnote": [], + "bbox": [ + 279, + 388, + 470, + 460 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "experiment with the pose estimation task and select two SOTA methods, HMR 2.0 [18] and CLIFF-SMPLify [4, 36], as our tools to generate two poses of each person. We then prompt the LLM-based agent to discriminate the results and choose the better one as the final response. Different tools excel in different scenarios and, to cover more diverse human poses and camera views, we build a new benchmark MixPose by selecting 100 images with extreme camera views from the MoYo [69] test set, 100 full-body samples and 100 severely-truncated samples from 3DPW [70] test set. Details of the prompt and MixPose benchmark are in Sup. Mat. As shown in Table 8 (a), ChatHuman archives a lower reconstruction error on the MixPose benchmark, validating the agent's effectiveness as a discriminator to improve tool performance. For the modification scheme, we validate on the body shape measurement task. We use CLIFF-BEDLAM [2] as tool and prompt the agent to discriminate and modify the tool result. The result is reported in Table 8 (b), and Fig. 9. The LLM-based agent enhances tool performance by using its general world knowledge to identify and correct unreasonable tool results, such as height and weight in Fig. 9 (a).", + "bbox": [ + 89, + 521, + 483, + 839 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Discussion and Concluding Remarks", + "text_level": 1, + "bbox": [ + 89, + 848, + 421, + 864 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We introduce ChatHuman, an LLM-based model designed to learn the use of tools related to 3D humans and assist users", + "bbox": [ + 89, + 869, + 482, + 898 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/26d1acca3eb15e0015bcb7dce7482816c37cf4da2ce94530c81c44d99a1ffac9.jpg", + "image_caption": [ + "Figure 10. Human interaction can improve the performance and tool usage accuracy of ChatHuman." + ], + "image_footnote": [], + "bbox": [ + 517, + 205, + 905, + 292 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "in solving tasks associated with 3D humans. ChatHuman processes requests from users, analyzes the needs, and utilizes the necessary tools. It then evaluates the tools' outputs to respond to the user's queries.", + "bbox": [ + 511, + 353, + 906, + 414 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations. ChatHuman may fail in certain calling scenarios, particularly when the user request is vague, and subsequent LLM internal analysis cannot rectify an incorrect initial function call. However, further interaction with users can remedy this if they provide additional information. Fig. 10 illustrates an instance of using body estimation and face reconstruction tools for avatar creation. Even with the application and analysis of the tool, outcomes like height estimation may not be entirely precise. One contributing factor is the accuracy of the training data; for instance, most height labels in datasets use the official height of models or celebrities, which may not account for variations like shoe height, such as a 7-inch heel. Incorporating more cues from users, combined with the LLM's knowledge of the world and reasoning capabilities, can enhance result accuracy, as shown in Fig. 10. Our system is currently limited by the academic methods used. Incorporating better academic methods will enhance model performance. Notably, adding new tools requires no additional training, allowing our method to evolve and improve as new techniques are developed.", + "bbox": [ + 511, + 419, + 908, + 720 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Future Work. ChatHuman offers several avenues for future development. In particular, user interaction/dialog offers opportunities to learn from user feedback. This could exploit reinforcement learning to refine the model's understanding.", + "bbox": [ + 511, + 727, + 906, + 787 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements and Disclosures", + "text_level": 1, + "bbox": [ + 511, + 803, + 815, + 819 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We thank Naureen Mahmood, Nicolas Keller and Nicolas Heron for the support of the data collection. While MJB is a co-founder and Chief Scientist at Meshcapade, his research in this project was performed solely at, and funded solely by, the Max Planck Society.", + "bbox": [ + 511, + 825, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8157", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Kfir Aberman, Rundi Wu, Dani Lischinski, Baoquan Chen, and Daniel Cohen-Or. Learning character-agnostic motion for motion retargeting in 2d. arXiv preprint arXiv:1905.01680, 2019.3", + "[2] Michael J Black, Priyanka Patel, Joachim Tesch, and Jinlong Yang. Bedlam: A synthetic dataset of bodies exhibiting detailed lifelike animated motion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8726-8737, 2023. 5, 7, 8", + "[3] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 3", + "[4] Federica Bogo, Angjoo Kanazawa, Christoph Lassner, Peter Gehler, Javier Romero, and Michael J. Black. Keep it SMPL: Automatic estimation of 3D human pose and shape from a single image. In ECCV, 2016. 3, 6, 8", + "[5] Yukang Cao, Yan-Pei Cao, Kai Han, Ying Shan, and KwanYee K Wong. Dreamavatar: Text-and-shape guided 3d human avatar generation via diffusion models. arXiv preprint arXiv:2304.00916, 2023. 3", + "[6] Harrison Chase and LangChain Contributors. Langchain, 2022. 5", + "[7] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhanghao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E. Gonzalez, Ion Stoica, and Eric P. Xing. Vicuna: An open-source chatbot impressing gpt-4 with $90\\%$ * chatgpt quality, 2023. 5", + "[8] Vasileios Choutas, Lea Müller, Chun-Hao P Huang, Siyu Tang, Dimitrios Tzionas, and Michael J Black. Accurate 3d body shape regression using metric and semantic attributes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2718-2728, 2022. 1, 3, 5, 7, 8", + "[9] Radek Daněček, Michael J Black, and Timo Bolkart. Emoca: Emotion driven monocular face capture and animation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20311-20322, 2022. 1", + "[10] Ginger Delmas, Philippe Weinzaepfel, Thomas Lucas, Francesc Moreno-Noguer, and Grégory Rogez. Posescript: 3d human poses from natural language. In ECCV, 2022. 1, 3, 4, 5, 6, 7", + "[11] Ginger Delmas, Philippe Weinzaepfel, Francesc Moreno-Noguer, and Grégory Rogez. Posefix: Correcting 3d human poses with natural language. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15018-15028, 2023. 3, 5", + "[12] Yu Deng, Jiaolong Yang, Sicheng Xu, Dong Chen, Yunde Jia, and Xin Tong. Accurate 3d face reconstruction with weakly-supervised learning: From single image to image set. In IEEE Computer Vision and Pattern Recognition Workshops, 2019. 3", + "[13] Yu Du, Fangyun Wei, and Hongyang Zhang. Anytool: Self-reflective, hierarchical agents for large-scale api calls. arXiv preprint arXiv:2402.04253, 2024. 3" + ], + "bbox": [ + 93, + 114, + 485, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Yao Feng, Vasileios Choutas, Timo Bolkart, Dimitrios Tzionas, and Michael J. Black. Collaborative regression of expressive bodies using moderation. In 3DV, 2021. 3", + "[15] Yao Feng, Haiwen Feng, Michael J Black, and Timo Bolkart. Learning an animatable detailed 3d face model from in-the-wild images. ACM Transactions on Graphics, 40(4):1-13, 2021. 1, 3, 5", + "[16] Yao Feng, Jing Lin, Sai Kumar Dwivedi, Yu Sun, Priyanka Patel, and Michael J. Black. ChatPose: Chatting about 3d human pose. In CVPR, 2024. 1, 3, 6, 7", + "[17] Yunfan Gao, Yun Xiong, Xinyu Gao, Kangxiang Jia, Jinliu Pan, Yuxi Bi, Yi Dai, Jiawei Sun, and Haofen Wang. Retrievalaugmented generation for large language models: A survey. arXiv preprint arXiv:2312.10997, 2023. 3", + "[18] Shubham Goel, Georgios Pavlakos, Jathushan Rajasegaran, Angjoo Kanazawa, and Jitendra Malik. Humans in 4D: Reconstructing and tracking humans with transformers. In ICCV, 2023. 1, 3, 5, 6, 8", + "[19] Sookwan Han and Hanbyul Joo. Chorus: Learning canonicalized 3d human-object spatial relations from unbounded synthesized images. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15835-15846, 2023. 3", + "[20] Fangzhou Hong, Mingyuan Zhang, Liang Pan, Zhongang Cai, Lei Yang, and Ziwei Liu. Avatarclip: Zero-shot text-driven generation and animation of 3d avatars. arXiv preprint arXiv:2205.08535, 2022. 3", + "[21] Cheng-Yu Hsieh, Si-An Chen, Chun-Liang Li, Yasuhisa Fujii, Alexander Ratner, Chen-Yu Lee, Ranjay Krishna, and Tomas Pfister. Tool documentation enables zero-shot tool-usage with large language models. arXiv preprint arXiv:2308.00675, 2023. 3", + "[22] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv:2106.09685, 2021. 5", + "[23] Yifeng Jiang, Jungdam Won, Yuting Ye, and C Karen Liu. Drop: Dynamics responses from human motion prior and projective dynamics. SIGGRAPH Asia, 2023. 3", + "[24] Hanbyul Joo, Natalia Neverova, and Andrea Vedaldi. Exemplar fine-tuning for 3D human pose fitting towards in-the-wild 3D human pose estimation. In 3DV, 2020. 3", + "[25] Angjoo Kanazawa, Michael J. Black, David W. Jacobs, and Jitendra Malik. End-to-end recovery of human shape and pose. In CVPR, 2018. 1, 3", + "[26] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 5", + "[27] Nikos Kolotouros, Georgios Pavlakos, Michael J. Black, and Kostas Daniilidis. Learning to reconstruct 3D human pose and shape via model-fitting in the loop. In ICCV, pages 2252-2261, 2019. 1, 3, 6", + "[28] Yilun Kong, Jingqing Ruan, Yihong Chen, Bin Zhang, Tianpeng Bao, Shiwei Shi, Guoqing Du, Xiaoru Hu, Hangyu Mao, Ziyue Li, et al. Tptu-v2: Boosting task planning and tool" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "8158", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "usage of large language model-based agents in real-world systems. arXiv preprint arXiv:2311.11315, 2023. 3", + "[29] Yilun Kong, Jingqing Ruan, Yihong Chen, Bin Zhang, Tianpeng Bao, Shiwei Shi, Guoqing Du, Xiaoru Hu, Hangyu Mao, Ziyue Li, et al. Tptu-v2: Boosting task planning and tool usage of large language model-based agents in real-world systems. arXiv preprint arXiv:2311.11315, 2023. 3", + "[30] Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Kuttler, Mike Lewis, Wen-tau Yih, Tim Roektaschel, et al. Retrievalaugmented generation for knowledge-intensive nlp tasks. Advances in neural information processing systems, 33:9459-9474, 2020. 3", + "[31] Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Kuttler, Mike Lewis, Wen-tau Yih, Tim Roektaschel, et al. Retrievalaugmented generation for knowledge-intensive nlp tasks. Advances in Neural Information Processing Systems, 33:9459-9474, 2020. 3", + "[32] Jiefeng Li, Chao Xu, Zhicun Chen, Siyuan Bian, Lixin Yang, and Cewu Lu. HybrIK: A hybrid analytical-neural inverse kinematics solution for 3D human pose and shape estimation. In CVPR, 2021. 3", + "[33] Jiaman Li, Jiajun Wu, and C Karen Liu. Object motion guided human motion synthesis. ACM Transactions on Graphics (TOG), 42(6):1-11, 2023. 3", + "[34] Peizhuo Li, Kfir Aberman, Rana Hanocka, Libin Liu, Olga Sorkine-Hornung, and Baoquan Chen. Learning skeletal articulations with neural blend shapes. ACM Transactions on Graphics (TOG), 40(4):1, 2021. 3", + "[35] Tianye Li, Timo Bolkart, Michael. J. Black, Hao Li, and Javier Romero. Learning a model of facial shape and expression from 4D scans. ACM Transactions on Graphics, (Proc. SIGGRAPH Asia), 36(6):194:1-194:17, 2017. 2", + "[36] Zhihao Li, Jianzhuang Liu, Zhensong Zhang, Songcen Xu, and Youliang Yan. CLIFF: Carrying location information in full frames into human pose and shape estimation. In ECCV, 2022. 1, 3, 8", + "[37] Jing Lin, Ailing Zeng, Haoqian Wang, Lei Zhang, and Yu Li. One-stage 3d whole-body mesh recovery with component aware transformer. CVPR, 2023. 1, 3", + "[38] Kevin Lin, Lijuan Wang, and Zicheng Liu. End-to-end human pose and mesh reconstruction with transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1954-1963, 2021. 5", + "[39] Tsung-Yi Lin, Michael Maire, Serge J. Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C. Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 4", + "[40] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. arXiv preprint arXiv:2304.08485, 2023. 3", + "[41] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, 2023. 2, 4, 5, 6, 7", + "[42] Shilong Liu, Hao Cheng, Haotian Liu, Hao Zhang, Feng Li, Tianhe Ren, Xueyan Zou, Jianwei Yang, Hang Su, Jun Zhu, Lei Zhang, Jianfeng Gao, and Chunyuan Li. Llava-plus: Learning to use tools for creating multimodal agents, 2023. 3" + ], + "bbox": [ + 91, + 92, + 483, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[43] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J. Black. SMPL: A skinned multiperson linear model. In ACM TOG, 2015. 2, 4", + "[44] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 5", + "[45] Thomas Lucas*, Fabien Baradel*, Philippe Weinzaepfel, and Grégory Rogez. Posegpt: Quantization-based 3d human motion generation and forecasting. In European Conference on Computer Vision (ECCV), 2022. 3", + "[46] Lea Muller, Ahmed AA Osman, Siyu Tang, Chun-Hao P Huang, and Michael J Black. On self-contact and human pose. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9990-9999, 2021. 1, 3", + "[47] OpenAI. GPT-4 technical report. 2023. 4, 6, 7", + "[48] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed A. A. Osman, Dimitrios Tzionas, and Michael J. Black. Expressive body capture: 3D hands, face, and body from a single image. In CVPR, 2019. 2", + "[49] P. Paysan, R. Knothe, B. Amberg, S. Romdhani, and T. Vetter. A 3D face model for pose and illumination invariant face recognition. Genova, Italy, 2009. IEEE. 2", + "[50] Mathis Petrovich, Michael J. Black, and Gül Varol. TMR: Text-to-motion retrieval using contrastive 3D human motion synthesis. In International Conference on Computer Vision (ICCV), 2023. 3, 5", + "[51] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 5", + "[52] Jeff Rasley, Samyam Rajbhandari, Olatunjri Ruwase, and Yuxiong He. Deepspeed: System optimizations enable training deep learning models with over 100 billion parameters. arXiv preprint arXiv:2002.11681, 2020. 5", + "[53] Jeff Rasley, Samyam Rajbhandari, Olatunj Ruwase, and Yuxiong He. Visual chatgpt: Talking, drawing and editing with visual foundation models. arXiv preprint arXiv:2303.04671, 2023. 2, 3, 5, 6", + "[54] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 5", + "[55] Javier Romero, Dimitrios Tzionas, and Michael J. Black. Embodied hands: Modeling and capturing hands and bodies together. ACM Transactions on Graphics, (Proc. SIGGRAPH Asia), 36(6), 2017. 3", + "[56] Yu Rong, Takaaki Shiratori, and Hanbyul Joo. Frankmocap: A monocular 3d whole-body pose estimation system via regression and integration. In IEEE International Conference on Computer Vision Workshops, 2021. 3", + "[57] Jingqing Ruan, Yihong Chen, Bin Zhang, Zhiwei Xu, Tianpeng Bao, Hangyu Mao, Ziyue Li, Xingyu Zeng, Rui Zhao, et al. Tptu: Task planning and tool usage of large language model-based ai agents. In NeurIPS 2023 Foundation Models for Decision Making Workshop, 2023. 3" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "8159", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[58] Yongliang Shen, Kaitao Song, Xu Tan, Dongsheng Li, Weiming Lu, and Yueting Zhuang. Huggingppt: Solving ai tasks with chatgpt and its friends in huggingface. arXiv preprint arXiv:2303.17580, 2023. 3", + "[59] Yongliang Shen, Kaitao Song, Xu Tan, Wenqi Zhang, Kan Ren, Siyu Yuan, Weiming Lu, Dongsheng Li, and Yueting Zhuang. Taskbench: Benchmarking large language models for task automation. arXiv preprint arXiv:2311.18760, 2023.4", + "[60] Soyong Shin, Juyong Kim, Eni Halilaj, and Michael J. Black. Wham: Reconstructing world-grounded humans with accurate 3d motion. In CVPR, 2024. 5", + "[61] Hongjin Su, Weijia Shi, Jungo Kasai, Yizhong Wang, Yushi Hu, Mari Ostendorf, Wen-tau Yih, Noah A Smith, Luke Zettlemoyer, and Tao Yu. One embedder, any task: Instructionfinetuned text embeddings. arXiv preprint arXiv:2212.09741, 2022. 4.5", + "[62] Dídac Surís, Sachit Menon, and Carl Vondrick. Vipergpt: Visual inference via python execution for reasoning. Proceedings of IEEE International Conference on Computer Vision (ICCV), 2023. 3", + "[63] John Sweller, Paul Ayres, and Slava Kalyuga. Cognitive Load Theory. Springer, New York, NY, 2011. 2", + "[64] Guy Tevet, Brian Gordon, Amir Hertz, Amit H Bermano, and Daniel Cohen-Or. Motionclip: Exposing human motion generation to clip space. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23–27, 2022, Proceedings, Part XXII, pages 358–374. Springer, 2022. 3", + "[65] Guy Tevet, Sigal Raab, Brian Gordon, Yoni Shafir, Daniel Cohen-or, and Amit Haim Bermano. Human motion diffusion model. In The Eleventh International Conference on Learning Representations, 2023. 3", + "[66] Ayush Tewari, Michael Zollhofer, Hyeongwoo Kim, Pablo Garrido, Florian Bernard, Patrick Perez, and Christian Theobalt. Mofa: Model-based deep convolutional face autoencoder for unsupervised monocular reconstruction. In Proceedings of the IEEE international conference on computer vision workshops, pages 1274-1283, 2017. 3", + "[67] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 3", + "[68] Shashank Tripathi, Agniv Chatterjee, Jean-Claude Passy, Hongwei Yi, Dimitrios Tzionas, and Michael J Black. Deco: Dense estimation of 3d human-scene contact in the wild. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8001-8013, 2023. 1, 3, 4, 5, 7", + "[69] Shashank Tripathi, Lea Müller, Chun-Hao P Huang, Omid Taheri, Michael J Black, and Dimitrios Tzionas. 3d human pose estimation via intuitive physics. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4713-4725, 2023. 5, 8", + "[70] Timo von Marcard, Roberto Henschel, Michael J. Black, Bodo Rosenhahn, and Gerard Pons-Moll. Recovering accurate 3D human pose in the wild using IMUs and a moving camera. In ECCV, 2018. 5, 6, 8" + ], + "bbox": [ + 91, + 92, + 483, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[71] Chenyu Wang, Weixin Luo, Qianyu Chen, Haonan Mai, Jindi Guo, Sixun Dong, Xiaohua (Michael) Xuan, Zhengxin Li, Lin Ma, and Shenghua Gao. Mllm-tool: A multimodal large language model for tool agent learning. arXiv preprint arXiv:2401.10727, 2024. 3", + "[72] Qinzhuo Wu, Wei Liu, Jian Luan, and Bin Wang. Toolplanner: A tool augmented llm for multi granularity instructions with path planning and feedback. arXiv preprint arXiv:2409.14826, 2024.3", + "[73] Hongyi Xu, Eduard Gabriel Bazavan, Andrei Zanfir, William T. Freeman, Rahul Sukthankar, and Cristian Sminchisescu. GHUM & GHUML: Generative 3D human shape and articulated pose models. In CVPR, 2020. 2", + "[74] Rui Yang, Lin Song, Yanwei Li, Sijie Zhao, Yixiao Ge, Xiu Li, and Ying Shan. GPT4Tools: Teaching llm to use tools via self-instruction. arXiv preprint arXiv:2305.18752, 2023. 2, 3, 4", + "[75] Rui Yang, Lin Song, Yanwei Li, Sijie Zhao, Yixiao Ge, Xiu Li, and Ying Shan. Gpt4tools: Teaching large language model to use tools via self-instruction, 2023. 5, 6", + "[76] Zhengyuan Yang, Linjie Li, Jianfeng Wang, Kevin Lin, Ehsan Azarnasab, Faisal Ahmed, Zicheng Liu, Ce Liu, Michael Zeng, and Lijuan Wang. Mm-react: Prompting chatgpt for multimodal reasoning and action. arXiv preprint arXiv:2303.11381, 2023. 3", + "[77] Zhao Yang, Jiaxuan Liu, Yucheng Han, Xin Chen, Zebiao Huang, Bin Fu, and Gang Yu. Appagent: Multimodal agents as smartphone users. arXiv preprint arXiv:2312.13771, 2023. 2, 3", + "[78] Shaokai Ye, Jessy Lauer, Mu Zhou, Alexander Mathis, and Mackenzie Mathis. Amadeusgpt: a natural language interface for interactive animal behavioral analysis. Advances in neural information processing systems, 36:6297-6329, 2023. 2", + "[79] Shaokai Ye, Jessy Lauer, Mu Zhou, Alexander Mathis, and Mackenzie Mathis. Amadeusgpt: a natural language interface for interactive animal behavioral analysis. Advances in neural information processing systems, 36, 2024. 3", + "[80] Siyu Yuan, Kaitao Song, Jiangjie Chen, Xu Tan, Yongliang Shen, Ren Kan, Dongsheng Li, and Deqing Yang. Easytool: Enhancing llm-based agents with concise tool instruction. arXiv preprint arXiv:2401.06201, 2024. 3", + "[81] Hongwen Zhang, Yating Tian, Xinchi Zhou, Wanli Ouyang, Yebin Liu, Limin Wang, and Zhenan Sun. Pymaf: 3d human pose and shape regression with pyramidal mesh alignment feedback loop. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11446-11456, 2021. 3", + "[82] Longwen Zhang, Qiwei Qiu, Hongyang Lin, Qixuan Zhang, Cheng Shi, Wei Yang, Ye Shi, Sibei Yang, Lan Xu, and Jingyi Yu. Dreamface: Progressive generation of animatable 3d faces under text guidance, 2023. 3", + "[83] Ruochen Zhao, Hailin Chen, Weishi Wang, Fangkai Jiao, Xuan Long Do, Chengwei Qin, Bosheng Ding, Xiaobao Guo, Minzhi Li, Xingxuan Li, et al. Retrieving multimodal information for augmented generation: A survey. arXiv preprint arXiv:2303.10868, 2023. 3", + "[84] Shenhao Zhu, Junming Leo Chen, Zuozhuo Dai, Yinghui Xu, Xun Cao, Yao Yao, Hao Zhu, and Siyu Zhu. Champ:" + ], + "bbox": [ + 516, + 92, + 906, + 898 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "8160", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Controllable and consistent human image animation with 3d parametric guidance, 2024. 1, 3, 5", + "bbox": [ + 124, + 90, + 482, + 119 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "8161", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2025/ChatHuman_ Chatting about 3D Humans with Tools/a1962ae6-92af-4828-9f1c-48e34ea37fec_model.json b/2025/ChatHuman_ Chatting about 3D Humans with Tools/a1962ae6-92af-4828-9f1c-48e34ea37fec_model.json new file mode 100644 index 0000000000000000000000000000000000000000..4b01723de82e8332bdd23cc0e7388ebbd3d06d6b --- /dev/null +++ b/2025/ChatHuman_ Chatting about 3D Humans with Tools/a1962ae6-92af-4828-9f1c-48e34ea37fec_model.json @@ -0,0 +1,2556 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.234, + 0.131, + 0.766, + 0.154 + ], + "angle": 0, + "content": "ChatHuman: Chatting about 3D Humans with Tools" + }, + { + "type": "text", + "bbox": [ + 0.234, + 0.179, + 0.762, + 0.199 + ], + "angle": 0, + "content": "Jing Lin\\(^{3,4,\\star}\\) Yao Feng\\(^{2,3,\\star}\\) Weiyang Liu\\(^{1,5}\\) Michael J. Black" + }, + { + "type": "text", + "bbox": [ + 0.192, + 0.199, + 0.807, + 0.236 + ], + "angle": 0, + "content": "\\(^{1}\\)Max Planck Institute for Intelligent Systems, Tübingen \\(^{2}\\)Stanford University \\(^{3}\\)Meshcapade \\(^{4}\\)Tsinghua University \\(^{5}\\)University of Cambridge" + }, + { + "type": "text", + "bbox": [ + 0.411, + 0.238, + 0.583, + 0.251 + ], + "angle": 0, + "content": "chathuman.github.io" + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.264, + 0.904, + 0.441 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.457, + 0.908, + 0.484 + ], + "angle": 0, + "content": "Figure 1. ChatHuman is a LLM-based agent that uses a multimodal LLM to exploit and combine tools, discriminate their results, and integrate the results to solve tasks related to 3D Humans." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.498, + 0.327, + 0.513 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.531, + 0.486, + 0.864 + ], + "angle": 0, + "content": "Numerous methods have been proposed to detect, estimate, and analyze properties of people in images, including 3D pose, shape, contact, human-object interaction, and emotion. While widely applicable in vision and other areas, such methods require expert knowledge to select, use, and interpret the results. To address this, we introduce ChatHuman, a language-driven system that integrates the capabilities of specialized methods into a unified framework. ChatHuman functions as an assistant proficient in utilizing, analyzing, and interacting with tools specific to 3D human tasks, adeptly discussing and resolving related challenges. Built on a Large Language Model (LLM) framework, ChatHuman is trained to autonomously select, apply, and interpret a diverse set of tools in response to user inputs. Our approach overcomes significant hurdles in adapting LLMs to 3D human tasks, including the need for domain-specific knowledge and the ability to interpret complex 3D outputs. The innovations of ChatHuman include leveraging academic publications to instruct the LLM on tool usage, employing a retrieval-augmented generation model to create in-context learning examples for managing new tools, and effectively discriminating between and integrating tool results by transforming" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.5, + 0.909, + 0.621 + ], + "angle": 0, + "content": "specialized 3D outputs into comprehensible formats. Experiments demonstrate that ChatHuman surpasses existing models in both tool selection accuracy and overall performance across various 3D human tasks, and it supports interactive chatting with users. ChatHuman represents a significant step toward consolidating diverse analytical methods into a unified, robust system for 3D human tasks. Code and data are available at chathuman.github.io." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.634, + 0.646, + 0.649 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.659, + 0.911, + 0.903 + ], + "angle": 0, + "content": "Research on 3D humans has progressed rapidly, resulting in the creation of many tools that can perform tasks like estimating a human's 3D pose from a single image [18, 25, 27, 36, 37], predicting face/body shapes [8, 15], capturing emotions [9, 15], and identifying regions of touch/contact [46, 68], generating human poses from text descriptions [10], and animating human images [84]. Each of these tools, however, focuses on a specific problem, functioning as isolated \"specialists\". Moreover, these separate tools cannot benefit from the expertise of others, and combining them to solve more complex tasks requires significant domain expertise. Ideally, we would have a single model that can adaptively leverage different tools to solve complex 3D human-related problems while offering intuitive user interaction through natural language input. Recent work such as ChatPose [16] has taken initial steps in this direction, unifying" + }, + { + "type": "page_footnote", + "bbox": [ + 0.091, + 0.875, + 0.486, + 0.902 + ], + "angle": 0, + "content": "*Equal contribution. This work was done while YF and JL were at Meshcapade." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "8150" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.137 + ], + "angle": 0, + "content": "ing pose generation, estimation, and general understanding within an LLM framework. Unfortunately, ChatPose lacks the accuracy of the best specialist methods." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.139, + 0.483, + 0.336 + ], + "angle": 0, + "content": "To address these issues, we build a multi-modal LLM, called ChatHuman, that specializes in using digital human modeling tools, enabling it to autonomously interpret instructions and complete diverse tasks related to 3D humans; see Fig. 1. Specifically, we teach an LLM to use a wide range of specialized human-related models for tasks like 3D pose estimation, emotion recognition, contact reasoning, and more, effectively extending the LLM's capabilities to the domain of 3D humans. This goes beyond providing a natural-language interface to these tools, as the LLM can use its broad understanding of humans to augment tool results or to analyze and integrate their outputs, providing better responses than any single tool alone." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.337, + 0.486, + 0.625 + ], + "angle": 0, + "content": "With ChatHuman, we introduce a novel approach by fine-tuning an LLM to act as an agent that autonomously calls appropriate tools based on user inputs, completing tasks and enhancing responses with tool-generated results. Similar in spirit, recent works have employed off-the-shelf or fine-tuned LLMs for tasks like basic vision (e.g., Visual ChatGPT [53]), mobile applications (e.g., AppAgent [77]), biology (e.g., AmadeusGPT [78]) and system automation (e.g., GPT4Tools [74]). Our work, however, differs by focusing specifically on the unique challenges of 3D human understanding. This domain requires precise, specialized terminology and a nuanced understanding of 3D-specific tools, which conventional LLMs lack. To teach the network this specialized terminology, we do what we would do as humans – we have the LLM read the papers describing the methods. Even with that knowledge, the LLM needs to understand the task goals, select an appropriate tool or tools, interpret results, and resolve differing results. These skills are all beyond the abilities of general LLMs." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.627, + 0.483, + 0.779 + ], + "angle": 0, + "content": "To address these challenges, we design the following training pipeline: 1) We utilize relevant literature about the tools to familiarize the LLM with domain knowledge, helping it know when and how to use these tools; 2) After using a tool, the LLM evaluates the reliability of the outcome using its \"judgment\" and compares different methods to identify the most reliable results; 3) It combines these results with its general knowledge to create response. This pipeline represents several key innovations, laying a foundation for LLMs to effectively handle complex, tool-driven 3D human tasks." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Retrieval-Augmented Tool Use: Details about tools are typically present in corresponding academic paper. We give the LLM access to these papers and demonstrate that \"reading the paper\" improves tool use performance. We further analyze which paper sections are most effective for instructing tool use. When encountering a new tool, users often turn to the user guide for assistance. We compile the complete documentation for these tools and utilize a paper-based" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.908, + 0.243 + ], + "angle": 0, + "content": "Retrieval-Augmented Generation (RAG) mechanism to improve the LLM's understanding and management of new tools. This means that, although the LLM has not encountered such tools during fine-tuning, it can still effectively use the tools with the help of the paper-based RAG mechanism. In some cases, tasks require combining multiple tools. To address a broader range of tool usage scenarios, we employ a graph-based invocation system, which includes a node for single-tool use, a chain for sequential tool execution, and a DAG for multi-tool combinations as shown in Fig. 3." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.244, + 0.909, + 0.426 + ], + "angle": 0, + "content": "3D Human-Related Tool Result Integration: Analyzing outputs from tools is crucial, as these outputs, such as body meshes, model parameters (e.g., SMPL pose), or motion sequences, are highly varied and complex. To make these results compatible with our LLM analysis system, we convert them into visual formats that the LLM can easily interpret. Guided by Cognitive Load Theory [63], we present these outputs as multiple-choice options, streamlining the selection process and enhancing the LLM's effectiveness in handling 3D human-related tasks. Combined with the LLM's extensive general knowledge, these integrated results enable it to generate sophisticated responses about 3D humans." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.426, + 0.909, + 0.607 + ], + "angle": 0, + "content": "Specifically, ChatHuman consists of a multimodal LLM LLaVA [41], and 26 tools involving 3D Humans and general vision tasks. The LLM is finetuned to use these tools and incorporate their results. User requests can be in the form of text descriptions, images or other 3D information (if applicable), and the model produces text descriptions, images, or other 3D outputs after tool reasoning. Extensive evaluations demonstrate that ChatHuman not only surpasses previous models in tool-use accuracy but also improves performance on various human-related tasks. It achieves this by reasoning about multiple outputs, evaluating their veracity, and combining them with its own knowledge." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.609, + 0.909, + 0.821 + ], + "angle": 0, + "content": "Summarizing, our key contributions include: (1) a framework that leverages LLMs to interact with users and address human-centric tasks using specialist tools; (2) a scientific-paper-based RAG mechanism that ensures precise tool use by comprehending tool descriptions from scholarly articles, enhancing tool applications and interactions; and (3) the integration of tool outcomes with LLMs, enabling the LLM to effectively explain tool results and interact with users. Additionally, the LLM is fine-tuned to distinguish between optimal and suboptimal tool results, improving overall accuracy. ChatHuman achieves superior performance in tool use and human-related tasks compared with other LLM-based methods or task-specific methods. The code, trained models, and datasets are available for research purposes." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.83, + 0.652, + 0.845 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.908, + 0.903 + ], + "angle": 0, + "content": "3D Humans: Many 3D human analysis tasks leverage parametric models like SMPL [43], SMPL-X [48], or GHUM [73] for the body, BFM [49] or FLAME [35] for" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.513, + 0.956 + ], + "angle": 0, + "content": "8151" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.091, + 0.485, + 0.38 + ], + "angle": 0, + "content": "faces, and MANO [55] for hands. These models enable the representation of the human body, face, and hands as low-dimensional vectors, facilitating subsequent applications in estimation and generation. Estimation of human pose and shape either relies on optimization methods [4, 24] or regression methods [8, 14, 18, 25, 27, 32, 36, 37, 56, 81], which estimate body shape and pose parameters from a single image. Similarly, face reconstruction methods [12, 15, 66] estimate shape and expression parameters of the face model from single images. The analysis of contact, vital for understanding human-environment interaction and social touch, has seen recent attention [19, 46, 68]. Generative modeling techniques such as PoseScript [10] and PoseFix [11] provide methods for synthesizing and correcting 3D human poses based on textual descriptions, while language-to-3D generation methods [5, 20, 82] facilitate 3D avatar creation. Additionally, numerous methods generate human motions [1, 23, 33, 34, 45, 50, 64, 65], and recent language-to-video models are even able to generate humans moving [3, 84]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.383, + 0.484, + 0.626 + ], + "angle": 0, + "content": "These basic methods excel in their respective scenarios, but are typically treated in isolation. When mature, such tools are often incorporated into software systems for animators that require significant domain knowledge. In contrast, recent generative models provide language interfaces to image, video, and 3D generation tools, making them accessible to novices. Until recently, such language-based control has not been possible for 3D humans. ChatPose [16] makes a step in this direction, unifying pose generation, estimation, and an LLM's general understanding into one model, but remains limited in its task capabilities. In contrast, our model integrates the performance of 26 3D human-related tasks into a single, LLM-based model. ChatHuman enables non-experts to solve real-world tasks by invoking appropriate tools and adding an extra layer of language-driven understanding that effectively leverages the tool outputs." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.629, + 0.487, + 0.901 + ], + "angle": 0, + "content": "Large Language Models and Tool Use: To expand LLM capabilities without expensive retraining, recent work has focused on enabling them to use specialized tools. In this approach, a tool library is constructed and LLMs act as planners to coordinate tool usage. Various tools have been adopted, e.g., vision modules [53, 62, 76], mobile applications [77], community tools [58], special tools [79], and system tools [74]. However, general-purpose LLMs often lack a deep understanding of specific tools, especially those requiring domain knowledge. To address this, recent work [13, 28, 71] proposes to fintune general-purpose LLMs (e.g., LLaMA [67], LLaVA [40, 42]) with domain-specific tool-use data. Some methods use additional tool documentation to improve accuracy [21, 80], while others compose different tools to accomplish complex tasks [29, 57, 72]. Distinct from previous work, ChatHuman focuses on 3D human tasks through language interaction by leveraging off-the-shelf human-related tools." + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.093, + 0.905, + 0.201 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.208, + 0.908, + 0.278 + ], + "angle": 0, + "content": "Figure 2. Paper-based Retrieval-Augmented Tool Use. We feed academic papers describing each tool to GPT-4 to build a document for each tool. During inference, given a user query, a relevant sample is retrieved from the documents and provided to the LLM-based agent as an in-context example to improve tool use accuracy." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.294, + 0.909, + 0.37 + ], + "angle": 0, + "content": "Retrieval Augmented Generation: RAG [17, 31, 83] is a technique to enhance generative tasks by retrieving relevant information from external databases, allowing for continual knowledge update. Here, we design a RAG mechanism to facilitate the use of new unseen tools." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.383, + 0.605, + 0.399 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.407, + 0.674, + 0.424 + ], + "angle": 0, + "content": "3.1. Overall Pipeline" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.431, + 0.909, + 0.568 + ], + "angle": 0, + "content": "ChatHuman consists of a multimodal LLM \\( f_{\\phi}(\\cdot) \\), along with a set of 3D human-related functions \\( f_{\\theta_1}(\\cdot), f_{\\theta_2}(\\cdot), \\ldots \\). These functions serve as tools for various tasks, such as 3D human pose estimation, pose generation, and 3D face reconstruction. Our model takes input text queries \\( X_{q} \\), images \\( X_{v} \\), and optionally \\( X_{m} \\) representing other 3D human-related modalities (e.g., SMPL parameters for 3D human poses). Then it invokes tools and integrates their results to generate outputs as text \\( Y_{t} \\), images \\( Y_{v} \\), or 3D human-related modalities \\( Y_{m} \\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.576, + 0.804, + 0.592 + ], + "angle": 0, + "content": "3.2. Retrieval-Augmented Tool Usage" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.599, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Teaching LLMs to decide when and how to use tools effectively is a significant challenge. A basic approach [53, 74] might involve including tool usage scenarios and input arguments within the LLM prompt, represented as \\( Y_{\\mathrm{tool}} = f_{\\phi}(X_q,X_t) \\), where \\( X_{t} \\) denotes tool definitions. However, this approach often falls short for specialized tools, especially given the variety of advanced tools for 3D human tasks. Many tools require background knowledge for correct use and have multiple application scenarios. For instance, the HMR tool [18] may be queried with requests like, \"Can you estimate this person's pose?\", \"What are the SMPL parameters?\", or \"Provide the 3D mesh of this person.\" Capturing all possible usage scenarios succinctly in a prompt is difficult, and as tools proliferate, prompt descriptions become unwieldy. To address these challenges, we introduce paper-based Retrieval-Augmented Generation (RAG) [30] and build a tool graph for tool combination. As shown in Fig. 2, we feed academic papers associated with each tool into GPT-4, prompting it to summarize the tool's functions and generate possible user queries for tool activation. These" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8152" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.092, + 0.482, + 0.166 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.17, + 0.483, + 0.211 + ], + "angle": 0, + "content": "Figure 3. Examples of tool use graphs. Tool use patterns include: a Node, which uses a single tool; a Chain, which requires sequential tool execution; and a DAG, which combines multiple tools." + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.226, + 0.48, + 0.359 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.364, + 0.485, + 0.421 + ], + "angle": 0, + "content": "Figure 4. Illustration of tool result integration. After tool invocation, results are transformed into VLM-compatible representations. The VLM then incorporates these transformed outputs to generate accurate, informative responses to user queries." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.446, + 0.484, + 0.629 + ], + "angle": 0, + "content": "papers, with their rich background and detailed instructions, enable the generation of user queries that cover diverse use cases. By combining these queries with each tool's structured arguments, we compile a document of question-answer pairs for each tool's operation. Figure 2 provides an example from one of these documents. These documents serve as an auxiliary knowledge base during inference, from which we retrieve a relevant example \\( X_{e} \\) in response to a user query \\( X_{q} \\). The retrieval process matches the text embedding of the query with embeddings in the tool documents using a text embedding model [61]. The retrieved sample is then presented to the agent \\( f_{\\phi} \\) as an in-context learning example:" + }, + { + "type": "equation", + "bbox": [ + 0.152, + 0.649, + 0.484, + 0.667 + ], + "angle": 0, + "content": "\\[\nX _ {e} = f _ {r} \\left(X _ {q}\\right), Y _ {\\text {t o o l}} = f _ {\\phi} \\left(X _ {q}, X _ {e}, X _ {t}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.685, + 0.486, + 0.731 + ], + "angle": 0, + "content": "where \\( f_{r} \\) is the retrieval function, and \\( Y_{\\mathrm{tool}} \\) is a textual description of the tool invocation, specifying tool selection, names, and input arguments for tool calls." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.486, + 0.902 + ], + "angle": 0, + "content": "Graph-based Tool Invocation. Note that the tool use description \\( Y_{tool} \\) varies depending on task settings, as shown for a single tool case in Fig. 2. However, some complex tasks require combining multiple tools. To handle this, we introduce a graph-based mechanism for tool invocation. We construct a tool graph with three structure types: nodes (single tool calls), chains (tool sequences for dependent tasks), and directed acyclic graphs (DAGs) [59] for complex multibranch operations. For each user query, the model predicts an appropriate tool graph and invokes the tools accordingly. Examples of tool graphs are shown in Fig. 3." + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.094, + 0.9, + 0.22 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.228, + 0.908, + 0.27 + ], + "angle": 0, + "content": "Figure 5. Illustration of tool results discrimination. When multiple plausible tools exist for a task, ChatHuman discriminates and chooses the best result as the final response." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.286, + 0.729, + 0.302 + ], + "angle": 0, + "content": "3.3. Tool Result Integration" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.309, + 0.909, + 0.581 + ], + "angle": 0, + "content": "After using tools, integrating their results is essential to effectively engage with users and solve problems. However, outputs from different tools vary widely, appearing as language, images, or vectors (like SMPL poses), which can challenge current multimodal LLMs, such as LLaVA [41], that process only text and images. To utilize these varied results and enhance the LLM's understanding of 3D humans, thereby improving its ability to apply world knowledge to problem-solving, we introduce a tool-conditioned transformation, \\(\\Psi(\\cdot)\\). As shown in Fig. 4, this transformation converts tool outputs \\(Y_{m}\\) into textual or visual formats the LLM can process. For example, we transform the vertex-wise contact label from DECO [68] into body part-level descriptions using SMPL's [43] vertex-to-part mapping dictionary, and render the mesh generated by PoseScipt [10] into an RGB image using rendering techniques. See Sup. Mat. for more details. The transformed results are then merged with the user query as context for response generation:" + }, + { + "type": "equation", + "bbox": [ + 0.634, + 0.592, + 0.907, + 0.61 + ], + "angle": 0, + "content": "\\[\nY _ {t} = f _ {\\phi} \\left(X _ {q}, \\Psi \\left(Y _ {m}\\right)\\right). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.619, + 0.908, + 0.665 + ], + "angle": 0, + "content": "In scenarios where multiple tools can address a request (Fig. 5), we present outcomes as multiple-choice questions, prompting the model to select the most relevant answer:" + }, + { + "type": "equation", + "bbox": [ + 0.59, + 0.675, + 0.907, + 0.693 + ], + "angle": 0, + "content": "\\[\nY _ {t} = f _ {\\phi} \\left(X _ {q}, \\Psi \\left(Y _ {m 1}\\right), \\Psi \\left(Y _ {m 2}\\right), \\dots\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.703, + 0.909, + 0.763 + ], + "angle": 0, + "content": "where \\( Y_{mi} \\) denotes the \\( i \\)-th tool result. Since different tools have different failure modes, this process enables ChatHuman to identify the best method case by case, producing more accurate output than any individual method alone." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.787, + 0.765, + 0.803 + ], + "angle": 0, + "content": "3.4. Training Data Construction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Tool Usage Instruction-following Data. To teach the LLM-based agent to correctly use tools, we construct 90K instruction-response pairs about tool usage. Following GPT4Tools [74], we provide GPT-4 [47] with a textual description of COCO training images [39] and a tool-related prompt containing a tool description. To improve efficiency," + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8153" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.082, + 0.9, + 0.35 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.355, + 0.907, + 0.385 + ], + "angle": 0, + "content": "Figure 6. Illustration of our instruction-following data construction pipeline. We construct tool usage and feedback data by providing GPT-4 with various tool-related information, image content, and ground truth labels. Gray text shows some example instructions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.403, + 0.484, + 0.447 + ], + "angle": 0, + "content": "we first prompt GPT-4 to summarize paper content, re-articulate tool functions, and enumerate 50 potential user queries for tool activation (see Fig. 6(a))." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.448, + 0.484, + 0.538 + ], + "angle": 0, + "content": "Tool Feedback Instruction-following Data. To help the multimodal LLM model discriminate and integrate the tool results, we construct 88K pairs of instruction-following data based on existing datasets 3DPW [70], MOYO [69], PoseScript [10] and SHAPY [8] (see Fig. 6(b)(c)). Please see Sup. Mat. for more details about data construction." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.547, + 0.248, + 0.563 + ], + "angle": 0, + "content": "3.5. Model Training" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.57, + 0.484, + 0.663 + ], + "angle": 0, + "content": "Once we have data, we use LoRA [22] to finetune the LLM \\( f_{\\phi}(\\cdot) \\) with the cross entropy loss. More specifically, with the ground truth tool invocation labels \\( \\hat{Y}_{tool} \\) and response label \\( \\hat{Y}_t \\), we optimize the model using the following objective function: \\( \\mathcal{L} = \\mathbf{CE}(\\hat{Y}_{tool}, Y_{tool}) + \\mathbf{CE}(\\hat{Y}_t, Y_t) \\), where \\( \\mathbf{CE} \\) denotes the cross-entropy loss. See Sup. Mat. for details." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.672, + 0.325, + 0.689 + ], + "angle": 0, + "content": "4. Experiments and Results" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.697, + 0.309, + 0.713 + ], + "angle": 0, + "content": "4.1. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.485, + 0.901 + ], + "angle": 0, + "content": "We use LLaVA-1.5 [41] as the VLM backbone, with CLIP [51] for vision encoding and Vicuna [7] for the LLM backbone. For retrieval, we adopt INSTRUCTOR [61] for text embedding and utilize Chroma's vector similarity searching algorithm to identify relevant examples. To preserve the generalization of the pretrained multi-modal LLM, we use LoRA [22] to perform efficient finetuning, with rank 128 and alpha 256. We implement tool utilization with LangChain [6], which enables automatic parsing of tool names and input parameters, followed by tool invocation Optimization uses AdamW [44], with a learning rate of 2e-4 and weight decay of 0. All models are finetuned over 2 epochs with a mix" + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.393, + 0.905, + 0.485 + ], + "angle": 0, + "content": "
PerceptionReasoningGeneration
Body Pose Estimation [18]Selective Person Pose Detection [18, 41]Text-to-Pose Generation [10]
Body Shape Measurement [2]Specific Person Shape Measurement [2, 41]Speculative Pose Generation [10, 41]
Hand Pose Estimation [38]Targeted Hand Pose Estimation [38, 41]Text-to-Image Generation [54]
Face Reconstruction [15]Described Person Face Reconstruction [15, 41]Text-based Pose Editing [11]
Human Segmentation [26]Described Person Segmentation [26, 41]Remove Someone From The Photo [26, 41, 54]
HOI Detection [68]Selective Person Contact Estimation [41, 68]Replace Someone From The Photo [26, 41, 54]
Pose Description [10]Visual Question Answering [41]Instruct Image Using Text [54]
Image Caption [41]Text-to-Motion Generation [50]
Motion Capture [60]Text-to-Video Generation [50, 54, 84]
Image-to-Video Generation [50, 84]
" + }, + { + "type": "table_caption", + "bbox": [ + 0.512, + 0.49, + 0.907, + 0.532 + ], + "angle": 0, + "content": "Table 1. ChatHuman supports 26 human-related tools, including 9 perception tools, 10 generation tools, and 7 reasoning tools. Tools in grey are unseen tools that are not included in the training data." + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.542, + 0.905, + 0.63 + ], + "angle": 0, + "content": "
MethodSeen ToolsUnseen Tools
SRtSRactSRargsSRIoUSRtSRactSRargsSRIoU
GPT4Tools [75]0.6090.5470.5250.5200.5660.6120.5460.5420.5250.573
GPT4Tools-FT [75]0.8250.7100.6870.6900.7410.9040.8070.6900.7470.800
Visual ChatGPT-3.5 [53]0.4980.3190.2370.2510.7910.5070.3140.2260.2930.803
Visual ChatGPT-4 [53]0.8920.8020.7150.7530.7970.9980.9130.8010.8720.907
ChatHuman1.0000.9740.9500.9700.9750.9990.9670.8930.9540.953
" + }, + { + "type": "table_caption", + "bbox": [ + 0.512, + 0.637, + 0.907, + 0.679 + ], + "angle": 0, + "content": "Table 2. Tool use accuracy comparison. Successful rate of thought \\((\\mathrm{SR}_{\\mathrm{t}})\\), action \\((\\mathrm{SR}_{\\mathrm{act}})\\), arguments \\((\\mathrm{SR}_{\\mathrm{args}})\\), execution (SR), and IoU are reported." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.694, + 0.907, + 0.77 + ], + "angle": 0, + "content": "ture of tool usage, tool feedback, and LLaVA multimodal instruction-tuning data, using 8 Nvidia A100-80G GPUs with the DeepSpeed [52] engine. Unless otherwise specified, we use LLaVA-1.5-7B as the base model for the ablation study. We support 26 tools, as listed in Tab. 1." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.774, + 0.747, + 0.79 + ], + "angle": 0, + "content": "4.2. Evaluation on Tool Usage" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Tool Usage Benchmark. To evaluate tool usage accuracy, we construct a validation and test set. The validation set has 1000 samples with the same tools as the training set, while the test set includes 689 samples related to 3 tools unseen during training. Split of seen and unseen tools are detailed in Table 1. Similar to our training data construction, we feed a textual description of COCO validation set image, a" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8154" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.089, + 0.478, + 0.206 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.209, + 0.484, + 0.264 + ], + "angle": 0, + "content": "Figure 7. Visualization of Animation Processing. Left: Conversation between the user and ChatHuman. Right: ChatHuman automatically calls tools to solve the task. ChatHuman can handle tasks beyond the capabilities of individual tools." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.271, + 0.483, + 0.362 + ], + "angle": 0, + "content": "tool description, and some examples summarized from the tool paper into GPT-4 and prompt it to generate instruction-following data. We use the image description captioned by LLaVA [41] instead of the original image captions to ensure differences between training and test sets. Finally, we manually check all question-answering pairs for accuracy." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.363, + 0.484, + 0.497 + ], + "angle": 0, + "content": "Baselines. We compare our method with Visual ChatGPT [53] and GPT4Tools [75] on the proposed evaluation set and report 5 metrics proposed in GPT4Tools [75]. See Sup. Mat. for details of the metrics. For Visual ChatGPT, we experiment with two versions of GPT: \"gpt-3.5-turbo-1106\" and \"gpt-4-turbo-preview\". For GPT4Tools, we adopt the official pretrained 13B model. For a fair comparison, we also finetune GPT4Tools with our training data using the official training code and obtain a variant, GPT4Tools-FT." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.5, + 0.484, + 0.621 + ], + "angle": 0, + "content": "Table 2 shows that the original GPT4Tools struggles on our benchmark due to mismatches between its tools and our human-centric ones. Visual ChatGPT-4 exhibits impressive tool use accuracy, showing its powerful zero-shot ability to follow a standardized format and use tools accurately. ChatHuman consistently outperforms other methods, particularly with tools not seen in training, thanks to our paper-based RAG mechanism." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.627, + 0.434, + 0.642 + ], + "angle": 0, + "content": "4.3. Evaluation on 3D Human Related Tasks" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.651, + 0.484, + 0.863 + ], + "angle": 0, + "content": "Character Animation. ChatHuman employs tools for text-to-motion and image-to-video generation. We demonstrate how these tools are utilized to interact with users and reason about motions based on conversations in Fig. 7 and Fig. 1. ChatHuman can also tackle tasks that cannot be resolved with a single tool. For instance, text-to-human video generation poses significant challenges due to the complexity of motion. Therefore, another option is to first generate a motion sequence via text-to-motion generation, then apply a video generation model conditioned on this sequence. The internal processing within ChatHuman, detailing how it analyzes and solves tasks, is visualized in Fig. 7. We also compare our text-to-video generation results with those of Pika1. The qualitative comparisons are shown in Fig. 8." + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.09, + 0.905, + 0.215 + ], + "angle": 0, + "content": "
Method3DPW [70]RPE Benchmark [16]
MPJPE ↓PA-MPJPE ↓MPJRE ↓MPJPE ↓PA-MPJPE ↓MPJRE ↓
SPIN [27]102.962.910.1244.9107.312.4
HMR 2.0 [18]91.058.49.2225.2105.112.1
LLaVA-S [41]440.8205.421.8490.7207.421.1
LLaVA*-S [41]232.1101.112.8---
GPT4-S [47]322.0136.716.0---
LLaVA-P [41]335.2172.316.5391.5191.917.8
GPT4-P [47]396.5203.418.6---
ChatPose [16]163.681.910.4253.6103.811.7
ChatHuman91.358.79.2147.279.110.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.512, + 0.218, + 0.907, + 0.316 + ], + "angle": 0, + "content": "Table 3. Comparison of vanilla human pose estimation and reasoning-based pose estimation on 3DPW [70] and RPE [16]. LLaVA* is fine-tuned with human keypoint data. \"S\" uses multimodal LLMs for keypoint detection and SMPLify [4] for pose optimization. \"P\" utilizes multimodal LLMs for textual pose descriptions, processed by PoseScript [10] to generate poses. MPJPE (in mm), MP-MPJPE (in mm), and MPJRE (\\( \\times 100 \\)) are reported." + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.326, + 0.905, + 0.442 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.446, + 0.907, + 0.489 + ], + "angle": 0, + "content": "Figure 8. Left: Comparison to \\(\\mathrm{Pika}^1\\) on text to video generation. Right: Qualitative comparison with ChatPose [16] and HMR 2.0 [18] on reasoning-based human pose estimation task." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.504, + 0.909, + 0.821 + ], + "angle": 0, + "content": "Pose Estimation. Following ChatPose [16], we evaluate the performance of our method on both classical and reasoning-based pose estimation (RPE) tasks. MPJPE, PA-MPJPE, and MPJRE on the 3DPW [70] and RPE [16] benchmarks are reported. For the reasoning-based pose estimation task, ChatHuman first grounds a human based on a textual description and feeds it into the pose estimation tool to get the result. As shown in Table 3, ChatHuman achieves comparable performance to the task-specific models on the classical pose estimation task. For reasoning-based human pose estimation, which involves both reasoning ability and advanced human pose estimation ability, ChatHuman outperforms both task-specific and multi-modal LLM methods by a large margin \\((34.6\\% \\downarrow\\) in MPVPE). As shown in Fig. 8, only our method achieves a satisfactory result. The multimodal LLM competitor ChatPose finds the correct person but fails to obtain an accurate pose due to its limited perception ability, while the task-specific tool does not match the correct person due to the lack of reasoning ability. This demonstrates the advantage of ChatHuman, which combines task-specific tool use expertise with the general reasoning ability of an LLM." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Pose Generation. Here we evaluate the pose generation capability of ChatHuman on the classical text-to-pose generation task and the speculative pose generation task (SPG) [16]. Following previous work [10, 16], we report the text-to-pose recall rate \\( R^{T2P} \\) and pose-to-text recall rate \\( R^{P2T} \\)" + }, + { + "type": "page_footnote", + "bbox": [ + 0.091, + 0.875, + 0.483, + 0.9 + ], + "angle": 0, + "content": "1We use the demo available at https://pika.art/ (as of May 2025) to obtain the results." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8155" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.094, + 0.089, + 0.482, + 0.174 + ], + "angle": 0, + "content": "
MethodPoseScript [10]SPG Benchmark [16]
\\( R^{P2T} \\)↑\\( R^{T2P} \\)↑\\( R^{P2T} \\)↑\\( R^{T2P} \\)↑
PoseScript [10]40.452.365.041.454.165.91.53.56.21.42.35.1
ChatPose [16]17.625.335.828.039.054.43.35.58.23.55.811.0
LLaVA-P [41]------2.14.07.12.13.36.1
GPT4-P [47]------2.74.79.22.75.38.2
ChatHuman41.852.665.142.152.366.53.25.09.93.56.510.6
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.09, + 0.178, + 0.483, + 0.234 + ], + "angle": 0, + "content": "Table 4. Comparison of classical and speculative pose generation on PoseScript [10] and SPG [16]. \"P\" denotes using LLMs to rephrase textual pose descriptions, which are then processed by PoseScript [10] to generate poses. Top 5, 10, 20 recall rates are reported." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.244, + 0.483, + 0.395 + ], + "angle": 0, + "content": "of the retrieval models trained on real poses and evaluated on generated poses. For the SPG task, ChatHuman first rephrases the indirect pose descriptions into explicit ones and adopts PoseScript (journal version) [10] to generate a pose. As shown in Table 4, our method archives comparable performance to the SOTA methods on both benchmarks. In contrast, the previous LLM-based method, ChatPose, performs poorly on the classical pose generation benchmark, while the task-specific model, PoseScript, lags in the SPG benchmark due to limited reasoning ability." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.399, + 0.483, + 0.656 + ], + "angle": 0, + "content": "Body Shape Measurement. We evaluate the body shape measurement accuracy of ChatHuman. We randomly sample 100 images from the HBW validation set [8] and compare our method with a multimodal LLM, LLaVA [41], and a SOTA body shape estimation method, CLIFF-BEDLAM [2]. For LLaVA and ChatHuman, we ask them the same question to inquire about the height, weight, chest, waist, and hip circumferences of a person and then prompt GPT-3.5 to extract the value from the model output. The details of the question and prompt are available in Sup. Mat. CLIFF-BEDLAM predicts the body shape parameters, which are then converted to measurements based on the shape-to-measurement function from SHAPY [8]. Anthropometric measurement errors are reported in Table 5(a). ChatHuman achieves superior performance in most measurements, outperforming the multimodal LLM competitor LLaVA by \\(42\\%\\) and CLIFF-BEDLAM by \\(15.7\\%\\) in average metric accuracy." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.66, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Human-Object Interaction (HoI). We evaluate the human-object interaction understanding ability of ChatHuman on the DECO [68] test set. The ground truth (GT) labels are obtained by converting the vertex-level contact labels into body part-level contact labels with SMPL's vertex-to-part mapping dictionary. Given a human image, we ask the multimodal LLM to detect the body parts contacting objects and prompt GPT-3.5 to extract the body part labels from the answer. Subsequently, we compare the predicted body parts with the GT label and compute the average detection precision, recall rate, and F1 Score. Table 5(b) shows that ChatHuman achieves SOTA precision and F1 score, demonstrating superior human-object interaction understanding ability. Notably, although LLaVA has a high recall rate, its precision and F1 score are rather poor, which means that it tends to predict all the body parts to be in contact with objects." + }, + { + "type": "table", + "bbox": [ + 0.515, + 0.09, + 0.718, + 0.175 + ], + "angle": 0, + "content": "
MethodHeight ↓Chest ↓Waist ↓Hip ↓
LLaVA [41]6.716.522.917.6
CLIFF-BEDLAM [2]7.88.613.57.0
ChatHuman6.76.113.06.4
(a) Comparison of body shape measurement. Measurement errors (in cm) on HBW [8] are reported.
" + }, + { + "type": "table", + "bbox": [ + 0.724, + 0.09, + 0.901, + 0.175 + ], + "angle": 0, + "content": "
MethodPrecision↑ Recall↑ F1Score ↑
LLaVA [41]0.260.810.39
GPT-4 [47]0.610.480.49
ChatHuman0.670.670.63
(b) Comparison of HOI estimation. Precision, Recall, and F1 score on DECO [68] are reported.
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.543, + 0.18, + 0.874, + 0.193 + ], + "angle": 0, + "content": "Table 5. Comparison of body shape and HOI estimation." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.205, + 0.667, + 0.221 + ], + "angle": 0, + "content": "4.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.23, + 0.908, + 0.501 + ], + "angle": 0, + "content": "Paper-based RAG Mechanism. To improve tool use accuracy, we design a paper-based RAG mechanism. We perform a breakdown ablation to investigate the effect of each component and their interactions. The baseline model, created by removing the RAG operation and trained with instruction-following data without paper content, is compared in Table 6(a). The baseline model's success rate (SR) is 0.96 for seen tools and 0.82 for unseen tools. Adding RAG increases the SR for unseen tools to 0.89, demonstrating its effectiveness in zero-shot settings. Further incorporating articles into training data boosts the performance: the successful rate of arguments \\((\\mathrm{SR}_{\\mathrm{args}})\\) rises from 0.93 to 0.95 for the seen tools and 0.84 to 0.94 for the unseen tools. This suggests that scholarly articles can help create high-quality instruction-following data and tool documents due to their detailed use instructions and diverse application scenarios. For a detailed analysis of each component's effect on instructing tool usage, please see Sup. Mat." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.503, + 0.909, + 0.655 + ], + "angle": 0, + "content": "Multiple Tools Invocation. One of the advantages of using a VLM as an agent is its powerful generalization capacity. To test the robustness and generalization ability of ChatHuman, we conduct the following ablation study. During training, we only include the tool graphs with no more than three tools, while during evaluation, the user queries might need up to five tools to solve. Table 6(b) depicts the results. As shown, ChatHuman exhibits an excellent robustness in this out-of-domain setting (more than three tools combination) with an action accuracy higher than \\(90\\%\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.659, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Tool Result Integration. We first conduct an ablation to study how the tools can enhance the human understanding capacity of multimodal LLM. The model without tools is our multimodal LLM backbone, LLaVA-1.5-7B [41], and the model with tools is our ChatHuman. The quantitative results are listed in Table 7. When equipped with tools, the HOI contact detection F1 score increases from 0.39 to 0.63 and the average body shape measurement error declines by \\(38\\%\\). These results demonstrate the effectiveness of tools in enriching the LLM's comprehension of human models and behaviors. Additionally, we study whether ChatHuman can utilize its world knowledge to discriminate and improve the tool performance. We design two discrimination schemes, i.e., selection and modification, and conduct an ablation study on two human-related tasks by comparing ChatHuman with the SOTA task-specific tools. For the selection scheme, we ex" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8156" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.137, + 0.089, + 0.538, + 0.161 + ], + "angle": 0, + "content": "
PaperRAGSeen ToolsUnseen Tools
SRtSRactSRargsSRIoUSRtSRactSRargsSRIoU
××0.9980.9670.9280.9600.9640.9460.8940.7750.8220.872
×✓1.0000.9670.9280.9610.9650.9960.9450.8420.8910.927
✓✓1.0000.9740.9500.9700.9750.9990.9670.8930.9540.953
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.199, + 0.164, + 0.473, + 0.176 + ], + "angle": 0, + "content": "(a) Ablation study of paper-based RAG mechanism." + }, + { + "type": "table", + "bbox": [ + 0.594, + 0.093, + 0.858, + 0.161 + ], + "angle": 0, + "content": "
Tool NumerSRtSRactSRargsSRIoU
20.9970.9600.9430.9280.973
30.9980.9590.9310.9320.974
40.9980.9430.9280.8750.968
50.9960.9290.8990.8470.950
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.634, + 0.164, + 0.816, + 0.176 + ], + "angle": 0, + "content": "(b) Ablation study of tool number." + }, + { + "type": "table_footnote", + "bbox": [ + 0.149, + 0.178, + 0.846, + 0.192 + ], + "angle": 0, + "content": "Table 6. Ablations related to tool usage. Successful rate of thought, action, arguments, execution, and IoU are reported." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.207, + 0.48, + 0.247 + ], + "angle": 0, + "content": "
MethodPrecision↑Recall↑F1Score↑MethodHeight↓Chest↓Waist↓Hip↓
w/o Tool0.260.810.39w/o Tool6.716.522.917.6
w/ Tool0.670.670.63w/ Tool6.76.113.06.4
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.115, + 0.25, + 0.26, + 0.261 + ], + "angle": 0, + "content": "(a) HOI Contact Detection." + }, + { + "type": "table_footnote", + "bbox": [ + 0.3, + 0.25, + 0.462, + 0.262 + ], + "angle": 0, + "content": "(b) Body Shape Measurement." + }, + { + "type": "table_caption", + "bbox": [ + 0.093, + 0.265, + 0.483, + 0.292 + ], + "angle": 0, + "content": "Table 7. Ablation study on the impact of tool usage for human-object contact detection and body shape estimation." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.295, + 0.288, + 0.343 + ], + "angle": 0, + "content": "
MethodMPJPE ↓PA-MPJPE ↓PA-MPVPE ↓
Tool A126.281.4101.9
Tool B124.084.6104.7
ChatHuman119.678.298.3
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.093, + 0.345, + 0.287, + 0.357 + ], + "angle": 0, + "content": "(a) Mesh Error (in mm) on MixPose." + }, + { + "type": "table", + "bbox": [ + 0.287, + 0.294, + 0.478, + 0.332 + ], + "angle": 0, + "content": "
MethodHeight ↓Chest ↓Waist ↓Hip ↓
Tool [2]7.88.613.57.0
ChatHuman6.76.113.06.4
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.294, + 0.334, + 0.478, + 0.357 + ], + "angle": 0, + "content": "(b) Body Shape Measurement Error (in cm) on HBW [8]." + }, + { + "type": "table_caption", + "bbox": [ + 0.093, + 0.359, + 0.484, + 0.387 + ], + "angle": 0, + "content": "Table 8. Study revealing how tool use improves human understanding on pose estimation and body shape measurement tasks." + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.39, + 0.28, + 0.462 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.129, + 0.463, + 0.266, + 0.471 + ], + "angle": 0, + "content": "(a) Tool generates unreasonable output" + }, + { + "type": "image", + "bbox": [ + 0.28, + 0.39, + 0.471, + 0.462 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.313, + 0.463, + 0.442, + 0.471 + ], + "angle": 0, + "content": "(b) Tool generates reasonable output" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.476, + 0.484, + 0.517 + ], + "angle": 0, + "content": "Figure 9. Illustration of how ChatHuman discriminates and integrates tool results. The Agent will fix the unreasonable tool result and integrate the reasonable tool result to generate a final response." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.522, + 0.484, + 0.84 + ], + "angle": 0, + "content": "experiment with the pose estimation task and select two SOTA methods, HMR 2.0 [18] and CLIFF-SMPLify [4, 36], as our tools to generate two poses of each person. We then prompt the LLM-based agent to discriminate the results and choose the better one as the final response. Different tools excel in different scenarios and, to cover more diverse human poses and camera views, we build a new benchmark MixPose by selecting 100 images with extreme camera views from the MoYo [69] test set, 100 full-body samples and 100 severely-truncated samples from 3DPW [70] test set. Details of the prompt and MixPose benchmark are in Sup. Mat. As shown in Table 8 (a), ChatHuman archives a lower reconstruction error on the MixPose benchmark, validating the agent's effectiveness as a discriminator to improve tool performance. For the modification scheme, we validate on the body shape measurement task. We use CLIFF-BEDLAM [2] as tool and prompt the agent to discriminate and modify the tool result. The result is reported in Table 8 (b), and Fig. 9. The LLM-based agent enhances tool performance by using its general world knowledge to identify and correct unreasonable tool results, such as height and weight in Fig. 9 (a)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.849, + 0.423, + 0.866 + ], + "angle": 0, + "content": "5. Discussion and Concluding Remarks" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.9 + ], + "angle": 0, + "content": "We introduce ChatHuman, an LLM-based model designed to learn the use of tools related to 3D humans and assist users" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.207, + 0.906, + 0.294 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.305, + 0.906, + 0.333 + ], + "angle": 0, + "content": "Figure 10. Human interaction can improve the performance and tool usage accuracy of ChatHuman." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.354, + 0.907, + 0.415 + ], + "angle": 0, + "content": "in solving tasks associated with 3D humans. ChatHuman processes requests from users, analyzes the needs, and utilizes the necessary tools. It then evaluates the tools' outputs to respond to the user's queries." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.42, + 0.909, + 0.722 + ], + "angle": 0, + "content": "Limitations. ChatHuman may fail in certain calling scenarios, particularly when the user request is vague, and subsequent LLM internal analysis cannot rectify an incorrect initial function call. However, further interaction with users can remedy this if they provide additional information. Fig. 10 illustrates an instance of using body estimation and face reconstruction tools for avatar creation. Even with the application and analysis of the tool, outcomes like height estimation may not be entirely precise. One contributing factor is the accuracy of the training data; for instance, most height labels in datasets use the official height of models or celebrities, which may not account for variations like shoe height, such as a 7-inch heel. Incorporating more cues from users, combined with the LLM's knowledge of the world and reasoning capabilities, can enhance result accuracy, as shown in Fig. 10. Our system is currently limited by the academic methods used. Incorporating better academic methods will enhance model performance. Notably, adding new tools requires no additional training, allowing our method to evolve and improve as new techniques are developed." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.728, + 0.908, + 0.789 + ], + "angle": 0, + "content": "Future Work. ChatHuman offers several avenues for future development. In particular, user interaction/dialog offers opportunities to learn from user feedback. This could exploit reinforcement learning to refine the model's understanding." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.804, + 0.816, + 0.82 + ], + "angle": 0, + "content": "Acknowledgements and Disclosures" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.907, + 0.901 + ], + "angle": 0, + "content": "We thank Naureen Mahmood, Nicolas Keller and Nicolas Heron for the support of the data collection. While MJB is a co-founder and Chief Scientist at Meshcapade, his research in this project was performed solely at, and funded solely by, the Max Planck Society." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8157" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.486, + 0.169 + ], + "angle": 0, + "content": "[1] Kfir Aberman, Rundi Wu, Dani Lischinski, Baoquan Chen, and Daniel Cohen-Or. Learning character-agnostic motion for motion retargeting in 2d. arXiv preprint arXiv:1905.01680, 2019.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.172, + 0.484, + 0.241 + ], + "angle": 0, + "content": "[2] Michael J Black, Priyanka Patel, Joachim Tesch, and Jinlong Yang. Bedlam: A synthetic dataset of bodies exhibiting detailed lifelike animated motion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8726-8737, 2023. 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.242, + 0.484, + 0.311 + ], + "angle": 0, + "content": "[3] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.312, + 0.484, + 0.367 + ], + "angle": 0, + "content": "[4] Federica Bogo, Angjoo Kanazawa, Christoph Lassner, Peter Gehler, Javier Romero, and Michael J. Black. Keep it SMPL: Automatic estimation of 3D human pose and shape from a single image. In ECCV, 2016. 3, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.368, + 0.484, + 0.422 + ], + "angle": 0, + "content": "[5] Yukang Cao, Yan-Pei Cao, Kai Han, Ying Shan, and KwanYee K Wong. Dreamavatar: Text-and-shape guided 3d human avatar generation via diffusion models. arXiv preprint arXiv:2304.00916, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.424, + 0.484, + 0.451 + ], + "angle": 0, + "content": "[6] Harrison Chase and LangChain Contributors. Langchain, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.453, + 0.484, + 0.522 + ], + "angle": 0, + "content": "[7] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhanghao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E. Gonzalez, Ion Stoica, and Eric P. Xing. Vicuna: An open-source chatbot impressing gpt-4 with \\(90\\%\\) * chatgpt quality, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.523, + 0.484, + 0.605 + ], + "angle": 0, + "content": "[8] Vasileios Choutas, Lea Müller, Chun-Hao P Huang, Siyu Tang, Dimitrios Tzionas, and Michael J Black. Accurate 3d body shape regression using metric and semantic attributes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2718-2728, 2022. 1, 3, 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.607, + 0.484, + 0.662 + ], + "angle": 0, + "content": "[9] Radek Daněček, Michael J Black, and Timo Bolkart. Emoca: Emotion driven monocular face capture and animation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20311-20322, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.663, + 0.484, + 0.717 + ], + "angle": 0, + "content": "[10] Ginger Delmas, Philippe Weinzaepfel, Thomas Lucas, Francesc Moreno-Noguer, and Grégory Rogez. Posescript: 3d human poses from natural language. In ECCV, 2022. 1, 3, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.719, + 0.484, + 0.788 + ], + "angle": 0, + "content": "[11] Ginger Delmas, Philippe Weinzaepfel, Francesc Moreno-Noguer, and Grégory Rogez. Posefix: Correcting 3d human poses with natural language. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15018-15028, 2023. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.79, + 0.484, + 0.856 + ], + "angle": 0, + "content": "[12] Yu Deng, Jiaolong Yang, Sicheng Xu, Dong Chen, Yunde Jia, and Xin Tong. Accurate 3d face reconstruction with weakly-supervised learning: From single image to image set. In IEEE Computer Vision and Pattern Recognition Workshops, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.859, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[13] Yu Du, Fangyun Wei, and Hongyang Zhang. Anytool: Self-reflective, hierarchical agents for large-scale api calls. arXiv preprint arXiv:2402.04253, 2024. 3" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.486, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.135 + ], + "angle": 0, + "content": "[14] Yao Feng, Vasileios Choutas, Timo Bolkart, Dimitrios Tzionas, and Michael J. Black. Collaborative regression of expressive bodies using moderation. In 3DV, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.136, + 0.907, + 0.19 + ], + "angle": 0, + "content": "[15] Yao Feng, Haiwen Feng, Michael J Black, and Timo Bolkart. Learning an animatable detailed 3d face model from in-the-wild images. ACM Transactions on Graphics, 40(4):1-13, 2021. 1, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.192, + 0.906, + 0.234 + ], + "angle": 0, + "content": "[16] Yao Feng, Jing Lin, Sai Kumar Dwivedi, Yu Sun, Priyanka Patel, and Michael J. Black. ChatPose: Chatting about 3d human pose. In CVPR, 2024. 1, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.235, + 0.907, + 0.291 + ], + "angle": 0, + "content": "[17] Yunfan Gao, Yun Xiong, Xinyu Gao, Kangxiang Jia, Jinliu Pan, Yuxi Bi, Yi Dai, Jiawei Sun, and Haofen Wang. Retrievalaugmented generation for large language models: A survey. arXiv preprint arXiv:2312.10997, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.293, + 0.907, + 0.347 + ], + "angle": 0, + "content": "[18] Shubham Goel, Georgios Pavlakos, Jathushan Rajasegaran, Angjoo Kanazawa, and Jitendra Malik. Humans in 4D: Reconstructing and tracking humans with transformers. In ICCV, 2023. 1, 3, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.349, + 0.907, + 0.417 + ], + "angle": 0, + "content": "[19] Sookwan Han and Hanbyul Joo. Chorus: Learning canonicalized 3d human-object spatial relations from unbounded synthesized images. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15835-15846, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.419, + 0.907, + 0.473 + ], + "angle": 0, + "content": "[20] Fangzhou Hong, Mingyuan Zhang, Liang Pan, Zhongang Cai, Lei Yang, and Ziwei Liu. Avatarclip: Zero-shot text-driven generation and animation of 3d avatars. arXiv preprint arXiv:2205.08535, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.476, + 0.906, + 0.544 + ], + "angle": 0, + "content": "[21] Cheng-Yu Hsieh, Si-An Chen, Chun-Liang Li, Yasuhisa Fujii, Alexander Ratner, Chen-Yu Lee, Ranjay Krishna, and Tomas Pfister. Tool documentation enables zero-shot tool-usage with large language models. arXiv preprint arXiv:2308.00675, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.546, + 0.907, + 0.601 + ], + "angle": 0, + "content": "[22] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv:2106.09685, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.603, + 0.907, + 0.645 + ], + "angle": 0, + "content": "[23] Yifeng Jiang, Jungdam Won, Yuting Ye, and C Karen Liu. Drop: Dynamics responses from human motion prior and projective dynamics. SIGGRAPH Asia, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.647, + 0.907, + 0.687 + ], + "angle": 0, + "content": "[24] Hanbyul Joo, Natalia Neverova, and Andrea Vedaldi. Exemplar fine-tuning for 3D human pose fitting towards in-the-wild 3D human pose estimation. In 3DV, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.689, + 0.906, + 0.731 + ], + "angle": 0, + "content": "[25] Angjoo Kanazawa, Michael J. Black, David W. Jacobs, and Jitendra Malik. End-to-end recovery of human shape and pose. In CVPR, 2018. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.732, + 0.907, + 0.802 + ], + "angle": 0, + "content": "[26] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.803, + 0.906, + 0.857 + ], + "angle": 0, + "content": "[27] Nikos Kolotouros, Georgios Pavlakos, Michael J. Black, and Kostas Daniilidis. Learning to reconstruct 3D human pose and shape via model-fitting in the loop. In ICCV, pages 2252-2261, 2019. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.859, + 0.907, + 0.901 + ], + "angle": 0, + "content": "[28] Yilun Kong, Jingqing Ruan, Yihong Chen, Bin Zhang, Tianpeng Bao, Shiwei Shi, Guoqing Du, Xiaoru Hu, Hangyu Mao, Ziyue Li, et al. Tptu-v2: Boosting task planning and tool" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8158" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.126, + 0.093, + 0.482, + 0.12 + ], + "angle": 0, + "content": "usage of large language model-based agents in real-world systems. arXiv preprint arXiv:2311.11315, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.121, + 0.484, + 0.189 + ], + "angle": 0, + "content": "[29] Yilun Kong, Jingqing Ruan, Yihong Chen, Bin Zhang, Tianpeng Bao, Shiwei Shi, Guoqing Du, Xiaoru Hu, Hangyu Mao, Ziyue Li, et al. Tptu-v2: Boosting task planning and tool usage of large language model-based agents in real-world systems. arXiv preprint arXiv:2311.11315, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.19, + 0.485, + 0.272 + ], + "angle": 0, + "content": "[30] Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Kuttler, Mike Lewis, Wen-tau Yih, Tim Roektaschel, et al. Retrievalaugmented generation for knowledge-intensive nlp tasks. Advances in neural information processing systems, 33:9459-9474, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.274, + 0.484, + 0.355 + ], + "angle": 0, + "content": "[31] Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Kuttler, Mike Lewis, Wen-tau Yih, Tim Roektaschel, et al. Retrievalaugmented generation for knowledge-intensive nlp tasks. Advances in Neural Information Processing Systems, 33:9459-9474, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.357, + 0.484, + 0.412 + ], + "angle": 0, + "content": "[32] Jiefeng Li, Chao Xu, Zhicun Chen, Siyuan Bian, Lixin Yang, and Cewu Lu. HybrIK: A hybrid analytical-neural inverse kinematics solution for 3D human pose and shape estimation. In CVPR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.413, + 0.482, + 0.453 + ], + "angle": 0, + "content": "[33] Jiaman Li, Jiajun Wu, and C Karen Liu. Object motion guided human motion synthesis. ACM Transactions on Graphics (TOG), 42(6):1-11, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.455, + 0.482, + 0.509 + ], + "angle": 0, + "content": "[34] Peizhuo Li, Kfir Aberman, Rana Hanocka, Libin Liu, Olga Sorkine-Hornung, and Baoquan Chen. Learning skeletal articulations with neural blend shapes. ACM Transactions on Graphics (TOG), 40(4):1, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.511, + 0.484, + 0.565 + ], + "angle": 0, + "content": "[35] Tianye Li, Timo Bolkart, Michael. J. Black, Hao Li, and Javier Romero. Learning a model of facial shape and expression from 4D scans. ACM Transactions on Graphics, (Proc. SIGGRAPH Asia), 36(6):194:1-194:17, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.567, + 0.484, + 0.62 + ], + "angle": 0, + "content": "[36] Zhihao Li, Jianzhuang Liu, Zhensong Zhang, Songcen Xu, and Youliang Yan. CLIFF: Carrying location information in full frames into human pose and shape estimation. In ECCV, 2022. 1, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.622, + 0.482, + 0.662 + ], + "angle": 0, + "content": "[37] Jing Lin, Ailing Zeng, Haoqian Wang, Lei Zhang, and Yu Li. One-stage 3d whole-body mesh recovery with component aware transformer. CVPR, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.664, + 0.484, + 0.719 + ], + "angle": 0, + "content": "[38] Kevin Lin, Lijuan Wang, and Zicheng Liu. End-to-end human pose and mesh reconstruction with transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1954-1963, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.72, + 0.482, + 0.773 + ], + "angle": 0, + "content": "[39] Tsung-Yi Lin, Michael Maire, Serge J. Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C. Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.776, + 0.484, + 0.815 + ], + "angle": 0, + "content": "[40] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. arXiv preprint arXiv:2304.08485, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.818, + 0.484, + 0.845 + ], + "angle": 0, + "content": "[41] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, 2023. 2, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[42] Shilong Liu, Hao Cheng, Haotian Liu, Hao Zhang, Feng Li, Tianhe Ren, Xueyan Zou, Jianwei Yang, Hang Su, Jun Zhu, Lei Zhang, Jianfeng Gao, and Chunyuan Li. Llava-plus: Learning to use tools for creating multimodal agents, 2023. 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.485, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.134 + ], + "angle": 0, + "content": "[43] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J. Black. SMPL: A skinned multiperson linear model. In ACM TOG, 2015. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.906, + 0.163 + ], + "angle": 0, + "content": "[44] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.165, + 0.906, + 0.22 + ], + "angle": 0, + "content": "[45] Thomas Lucas*, Fabien Baradel*, Philippe Weinzaepfel, and Grégory Rogez. Posegpt: Quantization-based 3d human motion generation and forecasting. In European Conference on Computer Vision (ECCV), 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.222, + 0.907, + 0.277 + ], + "angle": 0, + "content": "[46] Lea Muller, Ahmed AA Osman, Siyu Tang, Chun-Hao P Huang, and Michael J Black. On self-contact and human pose. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9990-9999, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.279, + 0.825, + 0.292 + ], + "angle": 0, + "content": "[47] OpenAI. GPT-4 technical report. 2023. 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.294, + 0.907, + 0.348 + ], + "angle": 0, + "content": "[48] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed A. A. Osman, Dimitrios Tzionas, and Michael J. Black. Expressive body capture: 3D hands, face, and body from a single image. In CVPR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.35, + 0.907, + 0.391 + ], + "angle": 0, + "content": "[49] P. Paysan, R. Knothe, B. Amberg, S. Romdhani, and T. Vetter. A 3D face model for pose and illumination invariant face recognition. Genova, Italy, 2009. IEEE. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.393, + 0.906, + 0.448 + ], + "angle": 0, + "content": "[50] Mathis Petrovich, Michael J. Black, and Gül Varol. TMR: Text-to-motion retrieval using contrastive 3D human motion synthesis. In International Conference on Computer Vision (ICCV), 2023. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.45, + 0.907, + 0.532 + ], + "angle": 0, + "content": "[51] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.534, + 0.907, + 0.589 + ], + "angle": 0, + "content": "[52] Jeff Rasley, Samyam Rajbhandari, Olatunjri Ruwase, and Yuxiong He. Deepspeed: System optimizations enable training deep learning models with over 100 billion parameters. arXiv preprint arXiv:2002.11681, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.591, + 0.907, + 0.645 + ], + "angle": 0, + "content": "[53] Jeff Rasley, Samyam Rajbhandari, Olatunj Ruwase, and Yuxiong He. Visual chatgpt: Talking, drawing and editing with visual foundation models. arXiv preprint arXiv:2303.04671, 2023. 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.648, + 0.907, + 0.716 + ], + "angle": 0, + "content": "[54] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.718, + 0.907, + 0.773 + ], + "angle": 0, + "content": "[55] Javier Romero, Dimitrios Tzionas, and Michael J. Black. Embodied hands: Modeling and capturing hands and bodies together. ACM Transactions on Graphics, (Proc. SIGGRAPH Asia), 36(6), 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.775, + 0.907, + 0.83 + ], + "angle": 0, + "content": "[56] Yu Rong, Takaaki Shiratori, and Hanbyul Joo. Frankmocap: A monocular 3d whole-body pose estimation system via regression and integration. In IEEE International Conference on Computer Vision Workshops, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.832, + 0.907, + 0.901 + ], + "angle": 0, + "content": "[57] Jingqing Ruan, Yihong Chen, Bin Zhang, Zhiwei Xu, Tianpeng Bao, Hangyu Mao, Ziyue Li, Xingyu Zeng, Rui Zhao, et al. Tptu: Task planning and tool usage of large language model-based ai agents. In NeurIPS 2023 Foundation Models for Decision Making Workshop, 2023. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8159" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.484, + 0.146 + ], + "angle": 0, + "content": "[58] Yongliang Shen, Kaitao Song, Xu Tan, Dongsheng Li, Weiming Lu, and Yueting Zhuang. Huggingppt: Solving ai tasks with chatgpt and its friends in huggingface. arXiv preprint arXiv:2303.17580, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.148, + 0.484, + 0.215 + ], + "angle": 0, + "content": "[59] Yongliang Shen, Kaitao Song, Xu Tan, Wenqi Zhang, Kan Ren, Siyu Yuan, Weiming Lu, Dongsheng Li, and Yueting Zhuang. Taskbench: Benchmarking large language models for task automation. arXiv preprint arXiv:2311.18760, 2023.4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.218, + 0.484, + 0.259 + ], + "angle": 0, + "content": "[60] Soyong Shin, Juyong Kim, Eni Halilaj, and Michael J. Black. Wham: Reconstructing world-grounded humans with accurate 3d motion. In CVPR, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.261, + 0.484, + 0.327 + ], + "angle": 0, + "content": "[61] Hongjin Su, Weijia Shi, Jungo Kasai, Yizhong Wang, Yushi Hu, Mari Ostendorf, Wen-tau Yih, Noah A Smith, Luke Zettlemoyer, and Tao Yu. One embedder, any task: Instructionfinetuned text embeddings. arXiv preprint arXiv:2212.09741, 2022. 4.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.33, + 0.484, + 0.384 + ], + "angle": 0, + "content": "[62] Dídac Surís, Sachit Menon, and Carl Vondrick. Vipergpt: Visual inference via python execution for reasoning. Proceedings of IEEE International Conference on Computer Vision (ICCV), 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.386, + 0.484, + 0.412 + ], + "angle": 0, + "content": "[63] John Sweller, Paul Ayres, and Slava Kalyuga. Cognitive Load Theory. Springer, New York, NY, 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.414, + 0.484, + 0.495 + ], + "angle": 0, + "content": "[64] Guy Tevet, Brian Gordon, Amir Hertz, Amit H Bermano, and Daniel Cohen-Or. Motionclip: Exposing human motion generation to clip space. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23–27, 2022, Proceedings, Part XXII, pages 358–374. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.498, + 0.484, + 0.552 + ], + "angle": 0, + "content": "[65] Guy Tevet, Sigal Raab, Brian Gordon, Yoni Shafir, Daniel Cohen-or, and Amit Haim Bermano. Human motion diffusion model. In The Eleventh International Conference on Learning Representations, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.554, + 0.484, + 0.636 + ], + "angle": 0, + "content": "[66] Ayush Tewari, Michael Zollhofer, Hyeongwoo Kim, Pablo Garrido, Florian Bernard, Patrick Perez, and Christian Theobalt. Mofa: Model-based deep convolutional face autoencoder for unsupervised monocular reconstruction. In Proceedings of the IEEE international conference on computer vision workshops, pages 1274-1283, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.637, + 0.484, + 0.704 + ], + "angle": 0, + "content": "[67] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.707, + 0.484, + 0.774 + ], + "angle": 0, + "content": "[68] Shashank Tripathi, Agniv Chatterjee, Jean-Claude Passy, Hongwei Yi, Dimitrios Tzionas, and Michael J Black. Deco: Dense estimation of 3d human-scene contact in the wild. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8001-8013, 2023. 1, 3, 4, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.776, + 0.484, + 0.844 + ], + "angle": 0, + "content": "[69] Shashank Tripathi, Lea Müller, Chun-Hao P Huang, Omid Taheri, Michael J Black, and Dimitrios Tzionas. 3d human pose estimation via intuitive physics. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4713-4725, 2023. 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.484, + 0.899 + ], + "angle": 0, + "content": "[70] Timo von Marcard, Roberto Henschel, Michael J. Black, Bodo Rosenhahn, and Gerard Pons-Moll. Recovering accurate 3D human pose in the wild using IMUs and a moving camera. In ECCV, 2018. 5, 6, 8" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.484, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.16 + ], + "angle": 0, + "content": "[71] Chenyu Wang, Weixin Luo, Qianyu Chen, Haonan Mai, Jindi Guo, Sixun Dong, Xiaohua (Michael) Xuan, Zhengxin Li, Lin Ma, and Shenghua Gao. Mllm-tool: A multimodal large language model for tool agent learning. arXiv preprint arXiv:2401.10727, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.163, + 0.907, + 0.216 + ], + "angle": 0, + "content": "[72] Qinzhuo Wu, Wei Liu, Jian Luan, and Bin Wang. Toolplanner: A tool augmented llm for multi granularity instructions with path planning and feedback. arXiv preprint arXiv:2409.14826, 2024.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.219, + 0.907, + 0.272 + ], + "angle": 0, + "content": "[73] Hongyi Xu, Eduard Gabriel Bazavan, Andrei Zanfir, William T. Freeman, Rahul Sukthankar, and Cristian Sminchisescu. GHUM & GHUML: Generative 3D human shape and articulated pose models. In CVPR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.274, + 0.907, + 0.326 + ], + "angle": 0, + "content": "[74] Rui Yang, Lin Song, Yanwei Li, Sijie Zhao, Yixiao Ge, Xiu Li, and Ying Shan. GPT4Tools: Teaching llm to use tools via self-instruction. arXiv preprint arXiv:2305.18752, 2023. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.33, + 0.906, + 0.37 + ], + "angle": 0, + "content": "[75] Rui Yang, Lin Song, Yanwei Li, Sijie Zhao, Yixiao Ge, Xiu Li, and Ying Shan. Gpt4tools: Teaching large language model to use tools via self-instruction, 2023. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.372, + 0.907, + 0.439 + ], + "angle": 0, + "content": "[76] Zhengyuan Yang, Linjie Li, Jianfeng Wang, Kevin Lin, Ehsan Azarnasab, Faisal Ahmed, Zicheng Liu, Ce Liu, Michael Zeng, and Lijuan Wang. Mm-react: Prompting chatgpt for multimodal reasoning and action. arXiv preprint arXiv:2303.11381, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.442, + 0.907, + 0.495 + ], + "angle": 0, + "content": "[77] Zhao Yang, Jiaxuan Liu, Yucheng Han, Xin Chen, Zebiao Huang, Bin Fu, and Gang Yu. Appagent: Multimodal agents as smartphone users. arXiv preprint arXiv:2312.13771, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.498, + 0.906, + 0.551 + ], + "angle": 0, + "content": "[78] Shaokai Ye, Jessy Lauer, Mu Zhou, Alexander Mathis, and Mackenzie Mathis. Amadeusgpt: a natural language interface for interactive animal behavioral analysis. Advances in neural information processing systems, 36:6297-6329, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.554, + 0.906, + 0.607 + ], + "angle": 0, + "content": "[79] Shaokai Ye, Jessy Lauer, Mu Zhou, Alexander Mathis, and Mackenzie Mathis. Amadeusgpt: a natural language interface for interactive animal behavioral analysis. Advances in neural information processing systems, 36, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.609, + 0.907, + 0.663 + ], + "angle": 0, + "content": "[80] Siyu Yuan, Kaitao Song, Jiangjie Chen, Xu Tan, Yongliang Shen, Ren Kan, Dongsheng Li, and Deqing Yang. Easytool: Enhancing llm-based agents with concise tool instruction. arXiv preprint arXiv:2401.06201, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.665, + 0.907, + 0.746 + ], + "angle": 0, + "content": "[81] Hongwen Zhang, Yating Tian, Xinchi Zhou, Wanli Ouyang, Yebin Liu, Limin Wang, and Zhenan Sun. Pymaf: 3d human pose and shape regression with pyramidal mesh alignment feedback loop. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11446-11456, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.749, + 0.907, + 0.802 + ], + "angle": 0, + "content": "[82] Longwen Zhang, Qiwei Qiu, Hongyang Lin, Qixuan Zhang, Cheng Shi, Wei Yang, Ye Shi, Sibei Yang, Lan Xu, and Jingyi Yu. Dreamface: Progressive generation of animatable 3d faces under text guidance, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.804, + 0.907, + 0.87 + ], + "angle": 0, + "content": "[83] Ruochen Zhao, Hailin Chen, Weishi Wang, Fangkai Jiao, Xuan Long Do, Chengwei Qin, Bosheng Ding, Xiaobao Guo, Minzhi Li, Xingxuan Li, et al. Retrieving multimodal information for augmented generation: A survey. arXiv preprint arXiv:2303.10868, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.907, + 0.9 + ], + "angle": 0, + "content": "[84] Shenhao Zhu, Junming Leo Chen, Zuozhuo Dai, Yinghui Xu, Xun Cao, Yao Yao, Hao Zhu, and Siyu Zhu. Champ:" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.956 + ], + "angle": 0, + "content": "8160" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.125, + 0.092, + 0.483, + 0.121 + ], + "angle": 0, + "content": "Controllable and consistent human image animation with 3d parametric guidance, 2024. 1, 3, 5" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.515, + 0.957 + ], + "angle": 0, + "content": "8161" + } + ] +] \ No newline at end of file diff --git a/2025/ChatHuman_ Chatting about 3D Humans with Tools/a1962ae6-92af-4828-9f1c-48e34ea37fec_origin.pdf b/2025/ChatHuman_ Chatting about 3D Humans with Tools/a1962ae6-92af-4828-9f1c-48e34ea37fec_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..96630c2d09b537d1a2afeac4ca5e1fb5434acc83 --- /dev/null +++ b/2025/ChatHuman_ Chatting about 3D Humans with Tools/a1962ae6-92af-4828-9f1c-48e34ea37fec_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3088eb23307d00d3d1d27fc7ec1bef8bd12a2c14495bf0ddf4f7cd0504db464b +size 4110988 diff --git a/2025/ChatHuman_ Chatting about 3D Humans with Tools/full.md b/2025/ChatHuman_ Chatting about 3D Humans with Tools/full.md new file mode 100644 index 0000000000000000000000000000000000000000..73a05b509baca45ed34979f1ce715ece852c4954 --- /dev/null +++ b/2025/ChatHuman_ Chatting about 3D Humans with Tools/full.md @@ -0,0 +1,328 @@ +# ChatHuman: Chatting about 3D Humans with Tools + +Jing Lin $^{3,4,\star}$ Yao Feng $^{2,3,\star}$ Weiyang Liu $^{1,5}$ Michael J. Black + +$^{1}$ Max Planck Institute for Intelligent Systems, Tübingen $^{2}$ Stanford University $^{3}$ Meshcapade $^{4}$ Tsinghua University $^{5}$ University of Cambridge + +chathuman.github.io + +![](images/19ac7f316394b3d5dd28cec3f9e9bb30fde837e0cdd86cfd1d4936421a0a83b6.jpg) +Figure 1. ChatHuman is a LLM-based agent that uses a multimodal LLM to exploit and combine tools, discriminate their results, and integrate the results to solve tasks related to 3D Humans. + +# Abstract + +Numerous methods have been proposed to detect, estimate, and analyze properties of people in images, including 3D pose, shape, contact, human-object interaction, and emotion. While widely applicable in vision and other areas, such methods require expert knowledge to select, use, and interpret the results. To address this, we introduce ChatHuman, a language-driven system that integrates the capabilities of specialized methods into a unified framework. ChatHuman functions as an assistant proficient in utilizing, analyzing, and interacting with tools specific to 3D human tasks, adeptly discussing and resolving related challenges. Built on a Large Language Model (LLM) framework, ChatHuman is trained to autonomously select, apply, and interpret a diverse set of tools in response to user inputs. Our approach overcomes significant hurdles in adapting LLMs to 3D human tasks, including the need for domain-specific knowledge and the ability to interpret complex 3D outputs. The innovations of ChatHuman include leveraging academic publications to instruct the LLM on tool usage, employing a retrieval-augmented generation model to create in-context learning examples for managing new tools, and effectively discriminating between and integrating tool results by transforming + +specialized 3D outputs into comprehensible formats. Experiments demonstrate that ChatHuman surpasses existing models in both tool selection accuracy and overall performance across various 3D human tasks, and it supports interactive chatting with users. ChatHuman represents a significant step toward consolidating diverse analytical methods into a unified, robust system for 3D human tasks. Code and data are available at chathuman.github.io. + +# 1. Introduction + +Research on 3D humans has progressed rapidly, resulting in the creation of many tools that can perform tasks like estimating a human's 3D pose from a single image [18, 25, 27, 36, 37], predicting face/body shapes [8, 15], capturing emotions [9, 15], and identifying regions of touch/contact [46, 68], generating human poses from text descriptions [10], and animating human images [84]. Each of these tools, however, focuses on a specific problem, functioning as isolated "specialists". Moreover, these separate tools cannot benefit from the expertise of others, and combining them to solve more complex tasks requires significant domain expertise. Ideally, we would have a single model that can adaptively leverage different tools to solve complex 3D human-related problems while offering intuitive user interaction through natural language input. Recent work such as ChatPose [16] has taken initial steps in this direction, unifying + +ing pose generation, estimation, and general understanding within an LLM framework. Unfortunately, ChatPose lacks the accuracy of the best specialist methods. + +To address these issues, we build a multi-modal LLM, called ChatHuman, that specializes in using digital human modeling tools, enabling it to autonomously interpret instructions and complete diverse tasks related to 3D humans; see Fig. 1. Specifically, we teach an LLM to use a wide range of specialized human-related models for tasks like 3D pose estimation, emotion recognition, contact reasoning, and more, effectively extending the LLM's capabilities to the domain of 3D humans. This goes beyond providing a natural-language interface to these tools, as the LLM can use its broad understanding of humans to augment tool results or to analyze and integrate their outputs, providing better responses than any single tool alone. + +With ChatHuman, we introduce a novel approach by fine-tuning an LLM to act as an agent that autonomously calls appropriate tools based on user inputs, completing tasks and enhancing responses with tool-generated results. Similar in spirit, recent works have employed off-the-shelf or fine-tuned LLMs for tasks like basic vision (e.g., Visual ChatGPT [53]), mobile applications (e.g., AppAgent [77]), biology (e.g., AmadeusGPT [78]) and system automation (e.g., GPT4Tools [74]). Our work, however, differs by focusing specifically on the unique challenges of 3D human understanding. This domain requires precise, specialized terminology and a nuanced understanding of 3D-specific tools, which conventional LLMs lack. To teach the network this specialized terminology, we do what we would do as humans – we have the LLM read the papers describing the methods. Even with that knowledge, the LLM needs to understand the task goals, select an appropriate tool or tools, interpret results, and resolve differing results. These skills are all beyond the abilities of general LLMs. + +To address these challenges, we design the following training pipeline: 1) We utilize relevant literature about the tools to familiarize the LLM with domain knowledge, helping it know when and how to use these tools; 2) After using a tool, the LLM evaluates the reliability of the outcome using its "judgment" and compares different methods to identify the most reliable results; 3) It combines these results with its general knowledge to create response. This pipeline represents several key innovations, laying a foundation for LLMs to effectively handle complex, tool-driven 3D human tasks. + +Retrieval-Augmented Tool Use: Details about tools are typically present in corresponding academic paper. We give the LLM access to these papers and demonstrate that "reading the paper" improves tool use performance. We further analyze which paper sections are most effective for instructing tool use. When encountering a new tool, users often turn to the user guide for assistance. We compile the complete documentation for these tools and utilize a paper-based + +Retrieval-Augmented Generation (RAG) mechanism to improve the LLM's understanding and management of new tools. This means that, although the LLM has not encountered such tools during fine-tuning, it can still effectively use the tools with the help of the paper-based RAG mechanism. In some cases, tasks require combining multiple tools. To address a broader range of tool usage scenarios, we employ a graph-based invocation system, which includes a node for single-tool use, a chain for sequential tool execution, and a DAG for multi-tool combinations as shown in Fig. 3. + +3D Human-Related Tool Result Integration: Analyzing outputs from tools is crucial, as these outputs, such as body meshes, model parameters (e.g., SMPL pose), or motion sequences, are highly varied and complex. To make these results compatible with our LLM analysis system, we convert them into visual formats that the LLM can easily interpret. Guided by Cognitive Load Theory [63], we present these outputs as multiple-choice options, streamlining the selection process and enhancing the LLM's effectiveness in handling 3D human-related tasks. Combined with the LLM's extensive general knowledge, these integrated results enable it to generate sophisticated responses about 3D humans. + +Specifically, ChatHuman consists of a multimodal LLM LLaVA [41], and 26 tools involving 3D Humans and general vision tasks. The LLM is finetuned to use these tools and incorporate their results. User requests can be in the form of text descriptions, images or other 3D information (if applicable), and the model produces text descriptions, images, or other 3D outputs after tool reasoning. Extensive evaluations demonstrate that ChatHuman not only surpasses previous models in tool-use accuracy but also improves performance on various human-related tasks. It achieves this by reasoning about multiple outputs, evaluating their veracity, and combining them with its own knowledge. + +Summarizing, our key contributions include: (1) a framework that leverages LLMs to interact with users and address human-centric tasks using specialist tools; (2) a scientific-paper-based RAG mechanism that ensures precise tool use by comprehending tool descriptions from scholarly articles, enhancing tool applications and interactions; and (3) the integration of tool outcomes with LLMs, enabling the LLM to effectively explain tool results and interact with users. Additionally, the LLM is fine-tuned to distinguish between optimal and suboptimal tool results, improving overall accuracy. ChatHuman achieves superior performance in tool use and human-related tasks compared with other LLM-based methods or task-specific methods. The code, trained models, and datasets are available for research purposes. + +# 2. Related work + +3D Humans: Many 3D human analysis tasks leverage parametric models like SMPL [43], SMPL-X [48], or GHUM [73] for the body, BFM [49] or FLAME [35] for + +faces, and MANO [55] for hands. These models enable the representation of the human body, face, and hands as low-dimensional vectors, facilitating subsequent applications in estimation and generation. Estimation of human pose and shape either relies on optimization methods [4, 24] or regression methods [8, 14, 18, 25, 27, 32, 36, 37, 56, 81], which estimate body shape and pose parameters from a single image. Similarly, face reconstruction methods [12, 15, 66] estimate shape and expression parameters of the face model from single images. The analysis of contact, vital for understanding human-environment interaction and social touch, has seen recent attention [19, 46, 68]. Generative modeling techniques such as PoseScript [10] and PoseFix [11] provide methods for synthesizing and correcting 3D human poses based on textual descriptions, while language-to-3D generation methods [5, 20, 82] facilitate 3D avatar creation. Additionally, numerous methods generate human motions [1, 23, 33, 34, 45, 50, 64, 65], and recent language-to-video models are even able to generate humans moving [3, 84]. + +These basic methods excel in their respective scenarios, but are typically treated in isolation. When mature, such tools are often incorporated into software systems for animators that require significant domain knowledge. In contrast, recent generative models provide language interfaces to image, video, and 3D generation tools, making them accessible to novices. Until recently, such language-based control has not been possible for 3D humans. ChatPose [16] makes a step in this direction, unifying pose generation, estimation, and an LLM's general understanding into one model, but remains limited in its task capabilities. In contrast, our model integrates the performance of 26 3D human-related tasks into a single, LLM-based model. ChatHuman enables non-experts to solve real-world tasks by invoking appropriate tools and adding an extra layer of language-driven understanding that effectively leverages the tool outputs. + +Large Language Models and Tool Use: To expand LLM capabilities without expensive retraining, recent work has focused on enabling them to use specialized tools. In this approach, a tool library is constructed and LLMs act as planners to coordinate tool usage. Various tools have been adopted, e.g., vision modules [53, 62, 76], mobile applications [77], community tools [58], special tools [79], and system tools [74]. However, general-purpose LLMs often lack a deep understanding of specific tools, especially those requiring domain knowledge. To address this, recent work [13, 28, 71] proposes to fintune general-purpose LLMs (e.g., LLaMA [67], LLaVA [40, 42]) with domain-specific tool-use data. Some methods use additional tool documentation to improve accuracy [21, 80], while others compose different tools to accomplish complex tasks [29, 57, 72]. Distinct from previous work, ChatHuman focuses on 3D human tasks through language interaction by leveraging off-the-shelf human-related tools. + +![](images/c8a28cec09c4397d56651e279c93958dabb01518558dd93f02c2e68e2b1e75a8.jpg) +Figure 2. Paper-based Retrieval-Augmented Tool Use. We feed academic papers describing each tool to GPT-4 to build a document for each tool. During inference, given a user query, a relevant sample is retrieved from the documents and provided to the LLM-based agent as an in-context example to improve tool use accuracy. + +Retrieval Augmented Generation: RAG [17, 31, 83] is a technique to enhance generative tasks by retrieving relevant information from external databases, allowing for continual knowledge update. Here, we design a RAG mechanism to facilitate the use of new unseen tools. + +# 3. Method + +# 3.1. Overall Pipeline + +ChatHuman consists of a multimodal LLM $f_{\phi}(\cdot)$ , along with a set of 3D human-related functions $f_{\theta_1}(\cdot), f_{\theta_2}(\cdot), \ldots$ . These functions serve as tools for various tasks, such as 3D human pose estimation, pose generation, and 3D face reconstruction. Our model takes input text queries $X_{q}$ , images $X_{v}$ , and optionally $X_{m}$ representing other 3D human-related modalities (e.g., SMPL parameters for 3D human poses). Then it invokes tools and integrates their results to generate outputs as text $Y_{t}$ , images $Y_{v}$ , or 3D human-related modalities $Y_{m}$ . + +# 3.2. Retrieval-Augmented Tool Usage + +Teaching LLMs to decide when and how to use tools effectively is a significant challenge. A basic approach [53, 74] might involve including tool usage scenarios and input arguments within the LLM prompt, represented as $Y_{\mathrm{tool}} = f_{\phi}(X_q,X_t)$ , where $X_{t}$ denotes tool definitions. However, this approach often falls short for specialized tools, especially given the variety of advanced tools for 3D human tasks. Many tools require background knowledge for correct use and have multiple application scenarios. For instance, the HMR tool [18] may be queried with requests like, "Can you estimate this person's pose?", "What are the SMPL parameters?", or "Provide the 3D mesh of this person." Capturing all possible usage scenarios succinctly in a prompt is difficult, and as tools proliferate, prompt descriptions become unwieldy. To address these challenges, we introduce paper-based Retrieval-Augmented Generation (RAG) [30] and build a tool graph for tool combination. As shown in Fig. 2, we feed academic papers associated with each tool into GPT-4, prompting it to summarize the tool's functions and generate possible user queries for tool activation. These + +![](images/a1e29eac086005c243e703ba19fe4add3ee45b7dfba3245a57fa16497403ddf5.jpg) +Figure 3. Examples of tool use graphs. Tool use patterns include: a Node, which uses a single tool; a Chain, which requires sequential tool execution; and a DAG, which combines multiple tools. + +![](images/426956da5ff2035e46c7cfebbe5860d6a16b611375086f24a36f24a46f749f51.jpg) +Figure 4. Illustration of tool result integration. After tool invocation, results are transformed into VLM-compatible representations. The VLM then incorporates these transformed outputs to generate accurate, informative responses to user queries. + +papers, with their rich background and detailed instructions, enable the generation of user queries that cover diverse use cases. By combining these queries with each tool's structured arguments, we compile a document of question-answer pairs for each tool's operation. Figure 2 provides an example from one of these documents. These documents serve as an auxiliary knowledge base during inference, from which we retrieve a relevant example $X_{e}$ in response to a user query $X_{q}$ . The retrieval process matches the text embedding of the query with embeddings in the tool documents using a text embedding model [61]. The retrieved sample is then presented to the agent $f_{\phi}$ as an in-context learning example: + +$$ +X _ {e} = f _ {r} \left(X _ {q}\right), Y _ {\text {t o o l}} = f _ {\phi} \left(X _ {q}, X _ {e}, X _ {t}\right), \tag {1} +$$ + +where $f_{r}$ is the retrieval function, and $Y_{\mathrm{tool}}$ is a textual description of the tool invocation, specifying tool selection, names, and input arguments for tool calls. + +Graph-based Tool Invocation. Note that the tool use description $Y_{tool}$ varies depending on task settings, as shown for a single tool case in Fig. 2. However, some complex tasks require combining multiple tools. To handle this, we introduce a graph-based mechanism for tool invocation. We construct a tool graph with three structure types: nodes (single tool calls), chains (tool sequences for dependent tasks), and directed acyclic graphs (DAGs) [59] for complex multibranch operations. For each user query, the model predicts an appropriate tool graph and invokes the tools accordingly. Examples of tool graphs are shown in Fig. 3. + +![](images/f00e8d6c697c375a36deb56f780b917385776a2f1bef2eeddaff0cff60e8a3af.jpg) +Figure 5. Illustration of tool results discrimination. When multiple plausible tools exist for a task, ChatHuman discriminates and chooses the best result as the final response. + +# 3.3. Tool Result Integration + +After using tools, integrating their results is essential to effectively engage with users and solve problems. However, outputs from different tools vary widely, appearing as language, images, or vectors (like SMPL poses), which can challenge current multimodal LLMs, such as LLaVA [41], that process only text and images. To utilize these varied results and enhance the LLM's understanding of 3D humans, thereby improving its ability to apply world knowledge to problem-solving, we introduce a tool-conditioned transformation, $\Psi(\cdot)$ . As shown in Fig. 4, this transformation converts tool outputs $Y_{m}$ into textual or visual formats the LLM can process. For example, we transform the vertex-wise contact label from DECO [68] into body part-level descriptions using SMPL's [43] vertex-to-part mapping dictionary, and render the mesh generated by PoseScipt [10] into an RGB image using rendering techniques. See Sup. Mat. for more details. The transformed results are then merged with the user query as context for response generation: + +$$ +Y _ {t} = f _ {\phi} \left(X _ {q}, \Psi \left(Y _ {m}\right)\right). \tag {2} +$$ + +In scenarios where multiple tools can address a request (Fig. 5), we present outcomes as multiple-choice questions, prompting the model to select the most relevant answer: + +$$ +Y _ {t} = f _ {\phi} \left(X _ {q}, \Psi \left(Y _ {m 1}\right), \Psi \left(Y _ {m 2}\right), \dots\right), \tag {3} +$$ + +where $Y_{mi}$ denotes the $i$ -th tool result. Since different tools have different failure modes, this process enables ChatHuman to identify the best method case by case, producing more accurate output than any individual method alone. + +# 3.4. Training Data Construction + +Tool Usage Instruction-following Data. To teach the LLM-based agent to correctly use tools, we construct 90K instruction-response pairs about tool usage. Following GPT4Tools [74], we provide GPT-4 [47] with a textual description of COCO training images [39] and a tool-related prompt containing a tool description. To improve efficiency, + +![](images/04cef13e70446eb9d50cafc9e26de846c05d88e22127698bd1e3ec71a070a82e.jpg) +Figure 6. Illustration of our instruction-following data construction pipeline. We construct tool usage and feedback data by providing GPT-4 with various tool-related information, image content, and ground truth labels. Gray text shows some example instructions. + +we first prompt GPT-4 to summarize paper content, re-articulate tool functions, and enumerate 50 potential user queries for tool activation (see Fig. 6(a)). + +Tool Feedback Instruction-following Data. To help the multimodal LLM model discriminate and integrate the tool results, we construct 88K pairs of instruction-following data based on existing datasets 3DPW [70], MOYO [69], PoseScript [10] and SHAPY [8] (see Fig. 6(b)(c)). Please see Sup. Mat. for more details about data construction. + +# 3.5. Model Training + +Once we have data, we use LoRA [22] to finetune the LLM $f_{\phi}(\cdot)$ with the cross entropy loss. More specifically, with the ground truth tool invocation labels $\hat{Y}_{tool}$ and response label $\hat{Y}_t$ , we optimize the model using the following objective function: $\mathcal{L} = \mathbf{CE}(\hat{Y}_{tool}, Y_{tool}) + \mathbf{CE}(\hat{Y}_t, Y_t)$ , where $\mathbf{CE}$ denotes the cross-entropy loss. See Sup. Mat. for details. + +# 4. Experiments and Results + +# 4.1. Implementation Details + +We use LLaVA-1.5 [41] as the VLM backbone, with CLIP [51] for vision encoding and Vicuna [7] for the LLM backbone. For retrieval, we adopt INSTRUCTOR [61] for text embedding and utilize Chroma's vector similarity searching algorithm to identify relevant examples. To preserve the generalization of the pretrained multi-modal LLM, we use LoRA [22] to perform efficient finetuning, with rank 128 and alpha 256. We implement tool utilization with LangChain [6], which enables automatic parsing of tool names and input parameters, followed by tool invocation Optimization uses AdamW [44], with a learning rate of 2e-4 and weight decay of 0. All models are finetuned over 2 epochs with a mix + +
PerceptionReasoningGeneration
Body Pose Estimation [18]Selective Person Pose Detection [18, 41]Text-to-Pose Generation [10]
Body Shape Measurement [2]Specific Person Shape Measurement [2, 41]Speculative Pose Generation [10, 41]
Hand Pose Estimation [38]Targeted Hand Pose Estimation [38, 41]Text-to-Image Generation [54]
Face Reconstruction [15]Described Person Face Reconstruction [15, 41]Text-based Pose Editing [11]
Human Segmentation [26]Described Person Segmentation [26, 41]Remove Someone From The Photo [26, 41, 54]
HOI Detection [68]Selective Person Contact Estimation [41, 68]Replace Someone From The Photo [26, 41, 54]
Pose Description [10]Visual Question Answering [41]Instruct Image Using Text [54]
Image Caption [41]Text-to-Motion Generation [50]
Motion Capture [60]Text-to-Video Generation [50, 54, 84]
Image-to-Video Generation [50, 84]
+ +Table 1. ChatHuman supports 26 human-related tools, including 9 perception tools, 10 generation tools, and 7 reasoning tools. Tools in grey are unseen tools that are not included in the training data. + +
MethodSeen ToolsUnseen Tools
SRtSRactSRargsSRIoUSRtSRactSRargsSRIoU
GPT4Tools [75]0.6090.5470.5250.5200.5660.6120.5460.5420.5250.573
GPT4Tools-FT [75]0.8250.7100.6870.6900.7410.9040.8070.6900.7470.800
Visual ChatGPT-3.5 [53]0.4980.3190.2370.2510.7910.5070.3140.2260.2930.803
Visual ChatGPT-4 [53]0.8920.8020.7150.7530.7970.9980.9130.8010.8720.907
ChatHuman1.0000.9740.9500.9700.9750.9990.9670.8930.9540.953
+ +Table 2. Tool use accuracy comparison. Successful rate of thought $(\mathrm{SR}_{\mathrm{t}})$ , action $(\mathrm{SR}_{\mathrm{act}})$ , arguments $(\mathrm{SR}_{\mathrm{args}})$ , execution (SR), and IoU are reported. + +ture of tool usage, tool feedback, and LLaVA multimodal instruction-tuning data, using 8 Nvidia A100-80G GPUs with the DeepSpeed [52] engine. Unless otherwise specified, we use LLaVA-1.5-7B as the base model for the ablation study. We support 26 tools, as listed in Tab. 1. + +# 4.2. Evaluation on Tool Usage + +Tool Usage Benchmark. To evaluate tool usage accuracy, we construct a validation and test set. The validation set has 1000 samples with the same tools as the training set, while the test set includes 689 samples related to 3 tools unseen during training. Split of seen and unseen tools are detailed in Table 1. Similar to our training data construction, we feed a textual description of COCO validation set image, a + +![](images/48c2839a53c2d970e59df4baa3a9287a51c1c8cb7b72537477c6f0a9dcac77b4.jpg) +Figure 7. Visualization of Animation Processing. Left: Conversation between the user and ChatHuman. Right: ChatHuman automatically calls tools to solve the task. ChatHuman can handle tasks beyond the capabilities of individual tools. + +tool description, and some examples summarized from the tool paper into GPT-4 and prompt it to generate instruction-following data. We use the image description captioned by LLaVA [41] instead of the original image captions to ensure differences between training and test sets. Finally, we manually check all question-answering pairs for accuracy. + +Baselines. We compare our method with Visual ChatGPT [53] and GPT4Tools [75] on the proposed evaluation set and report 5 metrics proposed in GPT4Tools [75]. See Sup. Mat. for details of the metrics. For Visual ChatGPT, we experiment with two versions of GPT: "gpt-3.5-turbo-1106" and "gpt-4-turbo-preview". For GPT4Tools, we adopt the official pretrained 13B model. For a fair comparison, we also finetune GPT4Tools with our training data using the official training code and obtain a variant, GPT4Tools-FT. + +Table 2 shows that the original GPT4Tools struggles on our benchmark due to mismatches between its tools and our human-centric ones. Visual ChatGPT-4 exhibits impressive tool use accuracy, showing its powerful zero-shot ability to follow a standardized format and use tools accurately. ChatHuman consistently outperforms other methods, particularly with tools not seen in training, thanks to our paper-based RAG mechanism. + +# 4.3. Evaluation on 3D Human Related Tasks + +Character Animation. ChatHuman employs tools for text-to-motion and image-to-video generation. We demonstrate how these tools are utilized to interact with users and reason about motions based on conversations in Fig. 7 and Fig. 1. ChatHuman can also tackle tasks that cannot be resolved with a single tool. For instance, text-to-human video generation poses significant challenges due to the complexity of motion. Therefore, another option is to first generate a motion sequence via text-to-motion generation, then apply a video generation model conditioned on this sequence. The internal processing within ChatHuman, detailing how it analyzes and solves tasks, is visualized in Fig. 7. We also compare our text-to-video generation results with those of Pika1. The qualitative comparisons are shown in Fig. 8. + +
Method3DPW [70]RPE Benchmark [16]
MPJPE ↓PA-MPJPE ↓MPJRE ↓MPJPE ↓PA-MPJPE ↓MPJRE ↓
SPIN [27]102.962.910.1244.9107.312.4
HMR 2.0 [18]91.058.49.2225.2105.112.1
LLaVA-S [41]440.8205.421.8490.7207.421.1
LLaVA*-S [41]232.1101.112.8---
GPT4-S [47]322.0136.716.0---
LLaVA-P [41]335.2172.316.5391.5191.917.8
GPT4-P [47]396.5203.418.6---
ChatPose [16]163.681.910.4253.6103.811.7
ChatHuman91.358.79.2147.279.110.3
+ +Table 3. Comparison of vanilla human pose estimation and reasoning-based pose estimation on 3DPW [70] and RPE [16]. LLaVA* is fine-tuned with human keypoint data. "S" uses multimodal LLMs for keypoint detection and SMPLify [4] for pose optimization. "P" utilizes multimodal LLMs for textual pose descriptions, processed by PoseScript [10] to generate poses. MPJPE (in mm), MP-MPJPE (in mm), and MPJRE ( $\times 100$ ) are reported. + +![](images/03654e6214846602acdd54266ced783c595f6630aecbfad01c8e70afaf5d4632.jpg) +Figure 8. Left: Comparison to $\mathrm{Pika}^1$ on text to video generation. Right: Qualitative comparison with ChatPose [16] and HMR 2.0 [18] on reasoning-based human pose estimation task. + +Pose Estimation. Following ChatPose [16], we evaluate the performance of our method on both classical and reasoning-based pose estimation (RPE) tasks. MPJPE, PA-MPJPE, and MPJRE on the 3DPW [70] and RPE [16] benchmarks are reported. For the reasoning-based pose estimation task, ChatHuman first grounds a human based on a textual description and feeds it into the pose estimation tool to get the result. As shown in Table 3, ChatHuman achieves comparable performance to the task-specific models on the classical pose estimation task. For reasoning-based human pose estimation, which involves both reasoning ability and advanced human pose estimation ability, ChatHuman outperforms both task-specific and multi-modal LLM methods by a large margin $(34.6\% \downarrow$ in MPVPE). As shown in Fig. 8, only our method achieves a satisfactory result. The multimodal LLM competitor ChatPose finds the correct person but fails to obtain an accurate pose due to its limited perception ability, while the task-specific tool does not match the correct person due to the lack of reasoning ability. This demonstrates the advantage of ChatHuman, which combines task-specific tool use expertise with the general reasoning ability of an LLM. + +Pose Generation. Here we evaluate the pose generation capability of ChatHuman on the classical text-to-pose generation task and the speculative pose generation task (SPG) [16]. Following previous work [10, 16], we report the text-to-pose recall rate $R^{T2P}$ and pose-to-text recall rate $R^{P2T}$ + +
MethodPoseScript [10]SPG Benchmark [16]
\( R^{P2T} \)↑\( R^{T2P} \)↑\( R^{P2T} \)↑\( R^{T2P} \)↑
PoseScript [10]40.452.365.041.454.165.91.53.56.21.42.35.1
ChatPose [16]17.625.335.828.039.054.43.35.58.23.55.811.0
LLaVA-P [41]------2.14.07.12.13.36.1
GPT4-P [47]------2.74.79.22.75.38.2
ChatHuman41.852.665.142.152.366.53.25.09.93.56.510.6
+ +Table 4. Comparison of classical and speculative pose generation on PoseScript [10] and SPG [16]. "P" denotes using LLMs to rephrase textual pose descriptions, which are then processed by PoseScript [10] to generate poses. Top 5, 10, 20 recall rates are reported. + +of the retrieval models trained on real poses and evaluated on generated poses. For the SPG task, ChatHuman first rephrases the indirect pose descriptions into explicit ones and adopts PoseScript (journal version) [10] to generate a pose. As shown in Table 4, our method archives comparable performance to the SOTA methods on both benchmarks. In contrast, the previous LLM-based method, ChatPose, performs poorly on the classical pose generation benchmark, while the task-specific model, PoseScript, lags in the SPG benchmark due to limited reasoning ability. + +Body Shape Measurement. We evaluate the body shape measurement accuracy of ChatHuman. We randomly sample 100 images from the HBW validation set [8] and compare our method with a multimodal LLM, LLaVA [41], and a SOTA body shape estimation method, CLIFF-BEDLAM [2]. For LLaVA and ChatHuman, we ask them the same question to inquire about the height, weight, chest, waist, and hip circumferences of a person and then prompt GPT-3.5 to extract the value from the model output. The details of the question and prompt are available in Sup. Mat. CLIFF-BEDLAM predicts the body shape parameters, which are then converted to measurements based on the shape-to-measurement function from SHAPY [8]. Anthropometric measurement errors are reported in Table 5(a). ChatHuman achieves superior performance in most measurements, outperforming the multimodal LLM competitor LLaVA by $42\%$ and CLIFF-BEDLAM by $15.7\%$ in average metric accuracy. + +Human-Object Interaction (HoI). We evaluate the human-object interaction understanding ability of ChatHuman on the DECO [68] test set. The ground truth (GT) labels are obtained by converting the vertex-level contact labels into body part-level contact labels with SMPL's vertex-to-part mapping dictionary. Given a human image, we ask the multimodal LLM to detect the body parts contacting objects and prompt GPT-3.5 to extract the body part labels from the answer. Subsequently, we compare the predicted body parts with the GT label and compute the average detection precision, recall rate, and F1 Score. Table 5(b) shows that ChatHuman achieves SOTA precision and F1 score, demonstrating superior human-object interaction understanding ability. Notably, although LLaVA has a high recall rate, its precision and F1 score are rather poor, which means that it tends to predict all the body parts to be in contact with objects. + +
MethodHeight ↓Chest ↓Waist ↓Hip ↓
LLaVA [41]6.716.522.917.6
CLIFF-BEDLAM [2]7.88.613.57.0
ChatHuman6.76.113.06.4
(a) Comparison of body shape measurement. Measurement errors (in cm) on HBW [8] are reported.
+ +Table 5. Comparison of body shape and HOI estimation. + +
MethodPrecision↑ Recall↑ F1Score ↑
LLaVA [41]0.260.810.39
GPT-4 [47]0.610.480.49
ChatHuman0.670.670.63
(b) Comparison of HOI estimation. Precision, Recall, and F1 score on DECO [68] are reported.
+ +# 4.4. Ablation Study + +Paper-based RAG Mechanism. To improve tool use accuracy, we design a paper-based RAG mechanism. We perform a breakdown ablation to investigate the effect of each component and their interactions. The baseline model, created by removing the RAG operation and trained with instruction-following data without paper content, is compared in Table 6(a). The baseline model's success rate (SR) is 0.96 for seen tools and 0.82 for unseen tools. Adding RAG increases the SR for unseen tools to 0.89, demonstrating its effectiveness in zero-shot settings. Further incorporating articles into training data boosts the performance: the successful rate of arguments $(\mathrm{SR}_{\mathrm{args}})$ rises from 0.93 to 0.95 for the seen tools and 0.84 to 0.94 for the unseen tools. This suggests that scholarly articles can help create high-quality instruction-following data and tool documents due to their detailed use instructions and diverse application scenarios. For a detailed analysis of each component's effect on instructing tool usage, please see Sup. Mat. + +Multiple Tools Invocation. One of the advantages of using a VLM as an agent is its powerful generalization capacity. To test the robustness and generalization ability of ChatHuman, we conduct the following ablation study. During training, we only include the tool graphs with no more than three tools, while during evaluation, the user queries might need up to five tools to solve. Table 6(b) depicts the results. As shown, ChatHuman exhibits an excellent robustness in this out-of-domain setting (more than three tools combination) with an action accuracy higher than $90\%$ . + +Tool Result Integration. We first conduct an ablation to study how the tools can enhance the human understanding capacity of multimodal LLM. The model without tools is our multimodal LLM backbone, LLaVA-1.5-7B [41], and the model with tools is our ChatHuman. The quantitative results are listed in Table 7. When equipped with tools, the HOI contact detection F1 score increases from 0.39 to 0.63 and the average body shape measurement error declines by $38\%$ . These results demonstrate the effectiveness of tools in enriching the LLM's comprehension of human models and behaviors. Additionally, we study whether ChatHuman can utilize its world knowledge to discriminate and improve the tool performance. We design two discrimination schemes, i.e., selection and modification, and conduct an ablation study on two human-related tasks by comparing ChatHuman with the SOTA task-specific tools. For the selection scheme, we ex + +
PaperRAGSeen ToolsUnseen Tools
SRtSRactSRargsSRIoUSRtSRactSRargsSRIoU
××0.9980.9670.9280.9600.9640.9460.8940.7750.8220.872
×✓1.0000.9670.9280.9610.9650.9960.9450.8420.8910.927
✓✓1.0000.9740.9500.9700.9750.9990.9670.8930.9540.953
+ +(a) Ablation study of paper-based RAG mechanism. + +
Tool NumerSRtSRactSRargsSRIoU
20.9970.9600.9430.9280.973
30.9980.9590.9310.9320.974
40.9980.9430.9280.8750.968
50.9960.9290.8990.8470.950
+ +(b) Ablation study of tool number. +Table 6. Ablations related to tool usage. Successful rate of thought, action, arguments, execution, and IoU are reported. + +
MethodPrecision↑Recall↑F1Score↑MethodHeight↓Chest↓Waist↓Hip↓
w/o Tool0.260.810.39w/o Tool6.716.522.917.6
w/ Tool0.670.670.63w/ Tool6.76.113.06.4
+ +(a) HOI Contact Detection. +(b) Body Shape Measurement. + +
MethodMPJPE ↓PA-MPJPE ↓PA-MPVPE ↓
Tool A126.281.4101.9
Tool B124.084.6104.7
ChatHuman119.678.298.3
+ +(a) Mesh Error (in mm) on MixPose. + +Table 7. Ablation study on the impact of tool usage for human-object contact detection and body shape estimation. + +
MethodHeight ↓Chest ↓Waist ↓Hip ↓
Tool [2]7.88.613.57.0
ChatHuman6.76.113.06.4
+ +(b) Body Shape Measurement Error (in cm) on HBW [8]. + +Table 8. Study revealing how tool use improves human understanding on pose estimation and body shape measurement tasks. + +![](images/5392627be968c7daa2879649bdbc37834a41ec515c7e336dfb41dd9a2384464d.jpg) +(a) Tool generates unreasonable output +Figure 9. Illustration of how ChatHuman discriminates and integrates tool results. The Agent will fix the unreasonable tool result and integrate the reasonable tool result to generate a final response. + +![](images/728061aad1464158687002c1bc86383ed210f88f91d32c7b96b015e60ae2676f.jpg) +(b) Tool generates reasonable output + +experiment with the pose estimation task and select two SOTA methods, HMR 2.0 [18] and CLIFF-SMPLify [4, 36], as our tools to generate two poses of each person. We then prompt the LLM-based agent to discriminate the results and choose the better one as the final response. Different tools excel in different scenarios and, to cover more diverse human poses and camera views, we build a new benchmark MixPose by selecting 100 images with extreme camera views from the MoYo [69] test set, 100 full-body samples and 100 severely-truncated samples from 3DPW [70] test set. Details of the prompt and MixPose benchmark are in Sup. Mat. As shown in Table 8 (a), ChatHuman archives a lower reconstruction error on the MixPose benchmark, validating the agent's effectiveness as a discriminator to improve tool performance. For the modification scheme, we validate on the body shape measurement task. We use CLIFF-BEDLAM [2] as tool and prompt the agent to discriminate and modify the tool result. The result is reported in Table 8 (b), and Fig. 9. The LLM-based agent enhances tool performance by using its general world knowledge to identify and correct unreasonable tool results, such as height and weight in Fig. 9 (a). + +# 5. Discussion and Concluding Remarks + +We introduce ChatHuman, an LLM-based model designed to learn the use of tools related to 3D humans and assist users + +![](images/26d1acca3eb15e0015bcb7dce7482816c37cf4da2ce94530c81c44d99a1ffac9.jpg) +Figure 10. Human interaction can improve the performance and tool usage accuracy of ChatHuman. + +in solving tasks associated with 3D humans. ChatHuman processes requests from users, analyzes the needs, and utilizes the necessary tools. It then evaluates the tools' outputs to respond to the user's queries. + +Limitations. ChatHuman may fail in certain calling scenarios, particularly when the user request is vague, and subsequent LLM internal analysis cannot rectify an incorrect initial function call. However, further interaction with users can remedy this if they provide additional information. Fig. 10 illustrates an instance of using body estimation and face reconstruction tools for avatar creation. Even with the application and analysis of the tool, outcomes like height estimation may not be entirely precise. One contributing factor is the accuracy of the training data; for instance, most height labels in datasets use the official height of models or celebrities, which may not account for variations like shoe height, such as a 7-inch heel. Incorporating more cues from users, combined with the LLM's knowledge of the world and reasoning capabilities, can enhance result accuracy, as shown in Fig. 10. Our system is currently limited by the academic methods used. Incorporating better academic methods will enhance model performance. Notably, adding new tools requires no additional training, allowing our method to evolve and improve as new techniques are developed. + +Future Work. ChatHuman offers several avenues for future development. In particular, user interaction/dialog offers opportunities to learn from user feedback. This could exploit reinforcement learning to refine the model's understanding. + +# Acknowledgements and Disclosures + +We thank Naureen Mahmood, Nicolas Keller and Nicolas Heron for the support of the data collection. While MJB is a co-founder and Chief Scientist at Meshcapade, his research in this project was performed solely at, and funded solely by, the Max Planck Society. + +# References + +[1] Kfir Aberman, Rundi Wu, Dani Lischinski, Baoquan Chen, and Daniel Cohen-Or. Learning character-agnostic motion for motion retargeting in 2d. arXiv preprint arXiv:1905.01680, 2019.3 +[2] Michael J Black, Priyanka Patel, Joachim Tesch, and Jinlong Yang. Bedlam: A synthetic dataset of bodies exhibiting detailed lifelike animated motion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8726-8737, 2023. 5, 7, 8 +[3] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 3 +[4] Federica Bogo, Angjoo Kanazawa, Christoph Lassner, Peter Gehler, Javier Romero, and Michael J. Black. Keep it SMPL: Automatic estimation of 3D human pose and shape from a single image. In ECCV, 2016. 3, 6, 8 +[5] Yukang Cao, Yan-Pei Cao, Kai Han, Ying Shan, and KwanYee K Wong. Dreamavatar: Text-and-shape guided 3d human avatar generation via diffusion models. arXiv preprint arXiv:2304.00916, 2023. 3 +[6] Harrison Chase and LangChain Contributors. Langchain, 2022. 5 +[7] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhanghao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E. Gonzalez, Ion Stoica, and Eric P. Xing. Vicuna: An open-source chatbot impressing gpt-4 with $90\%$ * chatgpt quality, 2023. 5 +[8] Vasileios Choutas, Lea Müller, Chun-Hao P Huang, Siyu Tang, Dimitrios Tzionas, and Michael J Black. Accurate 3d body shape regression using metric and semantic attributes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2718-2728, 2022. 1, 3, 5, 7, 8 +[9] Radek Daněček, Michael J Black, and Timo Bolkart. Emoca: Emotion driven monocular face capture and animation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20311-20322, 2022. 1 +[10] Ginger Delmas, Philippe Weinzaepfel, Thomas Lucas, Francesc Moreno-Noguer, and Grégory Rogez. Posescript: 3d human poses from natural language. In ECCV, 2022. 1, 3, 4, 5, 6, 7 +[11] Ginger Delmas, Philippe Weinzaepfel, Francesc Moreno-Noguer, and Grégory Rogez. Posefix: Correcting 3d human poses with natural language. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15018-15028, 2023. 3, 5 +[12] Yu Deng, Jiaolong Yang, Sicheng Xu, Dong Chen, Yunde Jia, and Xin Tong. Accurate 3d face reconstruction with weakly-supervised learning: From single image to image set. In IEEE Computer Vision and Pattern Recognition Workshops, 2019. 3 +[13] Yu Du, Fangyun Wei, and Hongyang Zhang. Anytool: Self-reflective, hierarchical agents for large-scale api calls. arXiv preprint arXiv:2402.04253, 2024. 3 + +[14] Yao Feng, Vasileios Choutas, Timo Bolkart, Dimitrios Tzionas, and Michael J. Black. Collaborative regression of expressive bodies using moderation. In 3DV, 2021. 3 +[15] Yao Feng, Haiwen Feng, Michael J Black, and Timo Bolkart. Learning an animatable detailed 3d face model from in-the-wild images. ACM Transactions on Graphics, 40(4):1-13, 2021. 1, 3, 5 +[16] Yao Feng, Jing Lin, Sai Kumar Dwivedi, Yu Sun, Priyanka Patel, and Michael J. Black. ChatPose: Chatting about 3d human pose. In CVPR, 2024. 1, 3, 6, 7 +[17] Yunfan Gao, Yun Xiong, Xinyu Gao, Kangxiang Jia, Jinliu Pan, Yuxi Bi, Yi Dai, Jiawei Sun, and Haofen Wang. Retrievalaugmented generation for large language models: A survey. arXiv preprint arXiv:2312.10997, 2023. 3 +[18] Shubham Goel, Georgios Pavlakos, Jathushan Rajasegaran, Angjoo Kanazawa, and Jitendra Malik. Humans in 4D: Reconstructing and tracking humans with transformers. In ICCV, 2023. 1, 3, 5, 6, 8 +[19] Sookwan Han and Hanbyul Joo. Chorus: Learning canonicalized 3d human-object spatial relations from unbounded synthesized images. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15835-15846, 2023. 3 +[20] Fangzhou Hong, Mingyuan Zhang, Liang Pan, Zhongang Cai, Lei Yang, and Ziwei Liu. Avatarclip: Zero-shot text-driven generation and animation of 3d avatars. arXiv preprint arXiv:2205.08535, 2022. 3 +[21] Cheng-Yu Hsieh, Si-An Chen, Chun-Liang Li, Yasuhisa Fujii, Alexander Ratner, Chen-Yu Lee, Ranjay Krishna, and Tomas Pfister. Tool documentation enables zero-shot tool-usage with large language models. arXiv preprint arXiv:2308.00675, 2023. 3 +[22] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv:2106.09685, 2021. 5 +[23] Yifeng Jiang, Jungdam Won, Yuting Ye, and C Karen Liu. Drop: Dynamics responses from human motion prior and projective dynamics. SIGGRAPH Asia, 2023. 3 +[24] Hanbyul Joo, Natalia Neverova, and Andrea Vedaldi. Exemplar fine-tuning for 3D human pose fitting towards in-the-wild 3D human pose estimation. In 3DV, 2020. 3 +[25] Angjoo Kanazawa, Michael J. Black, David W. Jacobs, and Jitendra Malik. End-to-end recovery of human shape and pose. In CVPR, 2018. 1, 3 +[26] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 5 +[27] Nikos Kolotouros, Georgios Pavlakos, Michael J. Black, and Kostas Daniilidis. Learning to reconstruct 3D human pose and shape via model-fitting in the loop. In ICCV, pages 2252-2261, 2019. 1, 3, 6 +[28] Yilun Kong, Jingqing Ruan, Yihong Chen, Bin Zhang, Tianpeng Bao, Shiwei Shi, Guoqing Du, Xiaoru Hu, Hangyu Mao, Ziyue Li, et al. Tptu-v2: Boosting task planning and tool + +usage of large language model-based agents in real-world systems. arXiv preprint arXiv:2311.11315, 2023. 3 +[29] Yilun Kong, Jingqing Ruan, Yihong Chen, Bin Zhang, Tianpeng Bao, Shiwei Shi, Guoqing Du, Xiaoru Hu, Hangyu Mao, Ziyue Li, et al. Tptu-v2: Boosting task planning and tool usage of large language model-based agents in real-world systems. arXiv preprint arXiv:2311.11315, 2023. 3 +[30] Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Kuttler, Mike Lewis, Wen-tau Yih, Tim Roektaschel, et al. Retrievalaugmented generation for knowledge-intensive nlp tasks. Advances in neural information processing systems, 33:9459-9474, 2020. 3 +[31] Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Kuttler, Mike Lewis, Wen-tau Yih, Tim Roektaschel, et al. Retrievalaugmented generation for knowledge-intensive nlp tasks. Advances in Neural Information Processing Systems, 33:9459-9474, 2020. 3 +[32] Jiefeng Li, Chao Xu, Zhicun Chen, Siyuan Bian, Lixin Yang, and Cewu Lu. HybrIK: A hybrid analytical-neural inverse kinematics solution for 3D human pose and shape estimation. In CVPR, 2021. 3 +[33] Jiaman Li, Jiajun Wu, and C Karen Liu. Object motion guided human motion synthesis. ACM Transactions on Graphics (TOG), 42(6):1-11, 2023. 3 +[34] Peizhuo Li, Kfir Aberman, Rana Hanocka, Libin Liu, Olga Sorkine-Hornung, and Baoquan Chen. Learning skeletal articulations with neural blend shapes. ACM Transactions on Graphics (TOG), 40(4):1, 2021. 3 +[35] Tianye Li, Timo Bolkart, Michael. J. Black, Hao Li, and Javier Romero. Learning a model of facial shape and expression from 4D scans. ACM Transactions on Graphics, (Proc. SIGGRAPH Asia), 36(6):194:1-194:17, 2017. 2 +[36] Zhihao Li, Jianzhuang Liu, Zhensong Zhang, Songcen Xu, and Youliang Yan. CLIFF: Carrying location information in full frames into human pose and shape estimation. In ECCV, 2022. 1, 3, 8 +[37] Jing Lin, Ailing Zeng, Haoqian Wang, Lei Zhang, and Yu Li. One-stage 3d whole-body mesh recovery with component aware transformer. CVPR, 2023. 1, 3 +[38] Kevin Lin, Lijuan Wang, and Zicheng Liu. End-to-end human pose and mesh reconstruction with transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1954-1963, 2021. 5 +[39] Tsung-Yi Lin, Michael Maire, Serge J. Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C. Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 4 +[40] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. arXiv preprint arXiv:2304.08485, 2023. 3 +[41] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, 2023. 2, 4, 5, 6, 7 +[42] Shilong Liu, Hao Cheng, Haotian Liu, Hao Zhang, Feng Li, Tianhe Ren, Xueyan Zou, Jianwei Yang, Hang Su, Jun Zhu, Lei Zhang, Jianfeng Gao, and Chunyuan Li. Llava-plus: Learning to use tools for creating multimodal agents, 2023. 3 + +[43] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J. Black. SMPL: A skinned multiperson linear model. In ACM TOG, 2015. 2, 4 +[44] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 5 +[45] Thomas Lucas*, Fabien Baradel*, Philippe Weinzaepfel, and Grégory Rogez. Posegpt: Quantization-based 3d human motion generation and forecasting. In European Conference on Computer Vision (ECCV), 2022. 3 +[46] Lea Muller, Ahmed AA Osman, Siyu Tang, Chun-Hao P Huang, and Michael J Black. On self-contact and human pose. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9990-9999, 2021. 1, 3 +[47] OpenAI. GPT-4 technical report. 2023. 4, 6, 7 +[48] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed A. A. Osman, Dimitrios Tzionas, and Michael J. Black. Expressive body capture: 3D hands, face, and body from a single image. In CVPR, 2019. 2 +[49] P. Paysan, R. Knothe, B. Amberg, S. Romdhani, and T. Vetter. A 3D face model for pose and illumination invariant face recognition. Genova, Italy, 2009. IEEE. 2 +[50] Mathis Petrovich, Michael J. Black, and Gül Varol. TMR: Text-to-motion retrieval using contrastive 3D human motion synthesis. In International Conference on Computer Vision (ICCV), 2023. 3, 5 +[51] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 5 +[52] Jeff Rasley, Samyam Rajbhandari, Olatunjri Ruwase, and Yuxiong He. Deepspeed: System optimizations enable training deep learning models with over 100 billion parameters. arXiv preprint arXiv:2002.11681, 2020. 5 +[53] Jeff Rasley, Samyam Rajbhandari, Olatunj Ruwase, and Yuxiong He. Visual chatgpt: Talking, drawing and editing with visual foundation models. arXiv preprint arXiv:2303.04671, 2023. 2, 3, 5, 6 +[54] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 5 +[55] Javier Romero, Dimitrios Tzionas, and Michael J. Black. Embodied hands: Modeling and capturing hands and bodies together. ACM Transactions on Graphics, (Proc. SIGGRAPH Asia), 36(6), 2017. 3 +[56] Yu Rong, Takaaki Shiratori, and Hanbyul Joo. Frankmocap: A monocular 3d whole-body pose estimation system via regression and integration. In IEEE International Conference on Computer Vision Workshops, 2021. 3 +[57] Jingqing Ruan, Yihong Chen, Bin Zhang, Zhiwei Xu, Tianpeng Bao, Hangyu Mao, Ziyue Li, Xingyu Zeng, Rui Zhao, et al. Tptu: Task planning and tool usage of large language model-based ai agents. In NeurIPS 2023 Foundation Models for Decision Making Workshop, 2023. 3 + +[58] Yongliang Shen, Kaitao Song, Xu Tan, Dongsheng Li, Weiming Lu, and Yueting Zhuang. Huggingppt: Solving ai tasks with chatgpt and its friends in huggingface. arXiv preprint arXiv:2303.17580, 2023. 3 +[59] Yongliang Shen, Kaitao Song, Xu Tan, Wenqi Zhang, Kan Ren, Siyu Yuan, Weiming Lu, Dongsheng Li, and Yueting Zhuang. Taskbench: Benchmarking large language models for task automation. arXiv preprint arXiv:2311.18760, 2023.4 +[60] Soyong Shin, Juyong Kim, Eni Halilaj, and Michael J. Black. Wham: Reconstructing world-grounded humans with accurate 3d motion. In CVPR, 2024. 5 +[61] Hongjin Su, Weijia Shi, Jungo Kasai, Yizhong Wang, Yushi Hu, Mari Ostendorf, Wen-tau Yih, Noah A Smith, Luke Zettlemoyer, and Tao Yu. One embedder, any task: Instructionfinetuned text embeddings. arXiv preprint arXiv:2212.09741, 2022. 4.5 +[62] Dídac Surís, Sachit Menon, and Carl Vondrick. Vipergpt: Visual inference via python execution for reasoning. Proceedings of IEEE International Conference on Computer Vision (ICCV), 2023. 3 +[63] John Sweller, Paul Ayres, and Slava Kalyuga. Cognitive Load Theory. Springer, New York, NY, 2011. 2 +[64] Guy Tevet, Brian Gordon, Amir Hertz, Amit H Bermano, and Daniel Cohen-Or. Motionclip: Exposing human motion generation to clip space. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23–27, 2022, Proceedings, Part XXII, pages 358–374. Springer, 2022. 3 +[65] Guy Tevet, Sigal Raab, Brian Gordon, Yoni Shafir, Daniel Cohen-or, and Amit Haim Bermano. Human motion diffusion model. In The Eleventh International Conference on Learning Representations, 2023. 3 +[66] Ayush Tewari, Michael Zollhofer, Hyeongwoo Kim, Pablo Garrido, Florian Bernard, Patrick Perez, and Christian Theobalt. Mofa: Model-based deep convolutional face autoencoder for unsupervised monocular reconstruction. In Proceedings of the IEEE international conference on computer vision workshops, pages 1274-1283, 2017. 3 +[67] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 3 +[68] Shashank Tripathi, Agniv Chatterjee, Jean-Claude Passy, Hongwei Yi, Dimitrios Tzionas, and Michael J Black. Deco: Dense estimation of 3d human-scene contact in the wild. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8001-8013, 2023. 1, 3, 4, 5, 7 +[69] Shashank Tripathi, Lea Müller, Chun-Hao P Huang, Omid Taheri, Michael J Black, and Dimitrios Tzionas. 3d human pose estimation via intuitive physics. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4713-4725, 2023. 5, 8 +[70] Timo von Marcard, Roberto Henschel, Michael J. Black, Bodo Rosenhahn, and Gerard Pons-Moll. Recovering accurate 3D human pose in the wild using IMUs and a moving camera. In ECCV, 2018. 5, 6, 8 + +[71] Chenyu Wang, Weixin Luo, Qianyu Chen, Haonan Mai, Jindi Guo, Sixun Dong, Xiaohua (Michael) Xuan, Zhengxin Li, Lin Ma, and Shenghua Gao. Mllm-tool: A multimodal large language model for tool agent learning. arXiv preprint arXiv:2401.10727, 2024. 3 +[72] Qinzhuo Wu, Wei Liu, Jian Luan, and Bin Wang. Toolplanner: A tool augmented llm for multi granularity instructions with path planning and feedback. arXiv preprint arXiv:2409.14826, 2024.3 +[73] Hongyi Xu, Eduard Gabriel Bazavan, Andrei Zanfir, William T. Freeman, Rahul Sukthankar, and Cristian Sminchisescu. GHUM & GHUML: Generative 3D human shape and articulated pose models. In CVPR, 2020. 2 +[74] Rui Yang, Lin Song, Yanwei Li, Sijie Zhao, Yixiao Ge, Xiu Li, and Ying Shan. GPT4Tools: Teaching llm to use tools via self-instruction. arXiv preprint arXiv:2305.18752, 2023. 2, 3, 4 +[75] Rui Yang, Lin Song, Yanwei Li, Sijie Zhao, Yixiao Ge, Xiu Li, and Ying Shan. Gpt4tools: Teaching large language model to use tools via self-instruction, 2023. 5, 6 +[76] Zhengyuan Yang, Linjie Li, Jianfeng Wang, Kevin Lin, Ehsan Azarnasab, Faisal Ahmed, Zicheng Liu, Ce Liu, Michael Zeng, and Lijuan Wang. Mm-react: Prompting chatgpt for multimodal reasoning and action. arXiv preprint arXiv:2303.11381, 2023. 3 +[77] Zhao Yang, Jiaxuan Liu, Yucheng Han, Xin Chen, Zebiao Huang, Bin Fu, and Gang Yu. Appagent: Multimodal agents as smartphone users. arXiv preprint arXiv:2312.13771, 2023. 2, 3 +[78] Shaokai Ye, Jessy Lauer, Mu Zhou, Alexander Mathis, and Mackenzie Mathis. Amadeusgpt: a natural language interface for interactive animal behavioral analysis. Advances in neural information processing systems, 36:6297-6329, 2023. 2 +[79] Shaokai Ye, Jessy Lauer, Mu Zhou, Alexander Mathis, and Mackenzie Mathis. Amadeusgpt: a natural language interface for interactive animal behavioral analysis. Advances in neural information processing systems, 36, 2024. 3 +[80] Siyu Yuan, Kaitao Song, Jiangjie Chen, Xu Tan, Yongliang Shen, Ren Kan, Dongsheng Li, and Deqing Yang. Easytool: Enhancing llm-based agents with concise tool instruction. arXiv preprint arXiv:2401.06201, 2024. 3 +[81] Hongwen Zhang, Yating Tian, Xinchi Zhou, Wanli Ouyang, Yebin Liu, Limin Wang, and Zhenan Sun. Pymaf: 3d human pose and shape regression with pyramidal mesh alignment feedback loop. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11446-11456, 2021. 3 +[82] Longwen Zhang, Qiwei Qiu, Hongyang Lin, Qixuan Zhang, Cheng Shi, Wei Yang, Ye Shi, Sibei Yang, Lan Xu, and Jingyi Yu. Dreamface: Progressive generation of animatable 3d faces under text guidance, 2023. 3 +[83] Ruochen Zhao, Hailin Chen, Weishi Wang, Fangkai Jiao, Xuan Long Do, Chengwei Qin, Bosheng Ding, Xiaobao Guo, Minzhi Li, Xingxuan Li, et al. Retrieving multimodal information for augmented generation: A survey. arXiv preprint arXiv:2303.10868, 2023. 3 +[84] Shenhao Zhu, Junming Leo Chen, Zuozhuo Dai, Yinghui Xu, Xun Cao, Yao Yao, Hao Zhu, and Siyu Zhu. Champ: + +Controllable and consistent human image animation with 3d parametric guidance, 2024. 1, 3, 5 \ No newline at end of file diff --git a/2025/ChatHuman_ Chatting about 3D Humans with Tools/images.zip b/2025/ChatHuman_ Chatting about 3D Humans with Tools/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..791bfd9c3f50c9a8654263f6d066f2c4a816d2c2 --- /dev/null +++ b/2025/ChatHuman_ Chatting about 3D Humans with Tools/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:93f74f3d06ea7b2642be307793010777beeba41d74c6907f90a46bbd48869096 +size 669407 diff --git a/2025/ChatHuman_ Chatting about 3D Humans with Tools/layout.json b/2025/ChatHuman_ Chatting about 3D Humans with Tools/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..1343f9e6476cfe964a605df3e17ba39f7cfca7f0 --- /dev/null +++ b/2025/ChatHuman_ Chatting about 3D Humans with Tools/layout.json @@ -0,0 +1,8812 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 143, + 103, + 468, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 103, + 468, + 121 + ], + "spans": [ + { + "bbox": [ + 143, + 103, + 468, + 121 + ], + "type": "text", + "content": "ChatHuman: Chatting about 3D Humans with Tools" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 143, + 141, + 466, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 141, + 466, + 157 + ], + "spans": [ + { + "bbox": [ + 143, + 141, + 466, + 157 + ], + "type": "text", + "content": "Jing Lin" + }, + { + "bbox": [ + 143, + 141, + 466, + 157 + ], + "type": "inline_equation", + "content": "^{3,4,\\star}" + }, + { + "bbox": [ + 143, + 141, + 466, + 157 + ], + "type": "text", + "content": " Yao Feng" + }, + { + "bbox": [ + 143, + 141, + 466, + 157 + ], + "type": "inline_equation", + "content": "^{2,3,\\star}" + }, + { + "bbox": [ + 143, + 141, + 466, + 157 + ], + "type": "text", + "content": " Weiyang Liu" + }, + { + "bbox": [ + 143, + 141, + 466, + 157 + ], + "type": "inline_equation", + "content": "^{1,5}" + }, + { + "bbox": [ + 143, + 141, + 466, + 157 + ], + "type": "text", + "content": " Michael J. Black" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 117, + 157, + 493, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 117, + 157, + 493, + 186 + ], + "spans": [ + { + "bbox": [ + 117, + 157, + 493, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 117, + 157, + 493, + 186 + ], + "type": "text", + "content": "Max Planck Institute for Intelligent Systems, Tübingen " + }, + { + "bbox": [ + 117, + 157, + 493, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 117, + 157, + 493, + 186 + ], + "type": "text", + "content": "Stanford University " + }, + { + "bbox": [ + 117, + 157, + 493, + 186 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 117, + 157, + 493, + 186 + ], + "type": "text", + "content": "Meshcapade " + }, + { + "bbox": [ + 117, + 157, + 493, + 186 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 117, + 157, + 493, + 186 + ], + "type": "text", + "content": "Tsinghua University " + }, + { + "bbox": [ + 117, + 157, + 493, + 186 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 117, + 157, + 493, + 186 + ], + "type": "text", + "content": "University of Cambridge" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 251, + 188, + 356, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 251, + 188, + 356, + 198 + ], + "spans": [ + { + "bbox": [ + 251, + 188, + 356, + 198 + ], + "type": "text", + "content": "chathuman.github.io" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 60, + 209, + 553, + 349 + ], + "blocks": [ + { + "bbox": [ + 60, + 209, + 553, + 349 + ], + "lines": [ + { + "bbox": [ + 60, + 209, + 553, + 349 + ], + "spans": [ + { + "bbox": [ + 60, + 209, + 553, + 349 + ], + "type": "image", + "image_path": "19ac7f316394b3d5dd28cec3f9e9bb30fde837e0cdd86cfd1d4936421a0a83b6.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 361, + 555, + 383 + ], + "lines": [ + { + "bbox": [ + 55, + 361, + 555, + 383 + ], + "spans": [ + { + "bbox": [ + 55, + 361, + 555, + 383 + ], + "type": "text", + "content": "Figure 1. ChatHuman is a LLM-based agent that uses a multimodal LLM to exploit and combine tools, discriminate their results, and integrate the results to solve tasks related to 3D Humans." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 151, + 394, + 200, + 406 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 394, + 200, + 406 + ], + "spans": [ + { + "bbox": [ + 151, + 394, + 200, + 406 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 54, + 420, + 297, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 420, + 297, + 684 + ], + "spans": [ + { + "bbox": [ + 54, + 420, + 297, + 684 + ], + "type": "text", + "content": "Numerous methods have been proposed to detect, estimate, and analyze properties of people in images, including 3D pose, shape, contact, human-object interaction, and emotion. While widely applicable in vision and other areas, such methods require expert knowledge to select, use, and interpret the results. To address this, we introduce ChatHuman, a language-driven system that integrates the capabilities of specialized methods into a unified framework. ChatHuman functions as an assistant proficient in utilizing, analyzing, and interacting with tools specific to 3D human tasks, adeptly discussing and resolving related challenges. Built on a Large Language Model (LLM) framework, ChatHuman is trained to autonomously select, apply, and interpret a diverse set of tools in response to user inputs. Our approach overcomes significant hurdles in adapting LLMs to 3D human tasks, including the need for domain-specific knowledge and the ability to interpret complex 3D outputs. The innovations of ChatHuman include leveraging academic publications to instruct the LLM on tool usage, employing a retrieval-augmented generation model to create in-context learning examples for managing new tools, and effectively discriminating between and integrating tool results by transforming" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 396, + 556, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 396, + 556, + 491 + ], + "spans": [ + { + "bbox": [ + 313, + 396, + 556, + 491 + ], + "type": "text", + "content": "specialized 3D outputs into comprehensible formats. Experiments demonstrate that ChatHuman surpasses existing models in both tool selection accuracy and overall performance across various 3D human tasks, and it supports interactive chatting with users. ChatHuman represents a significant step toward consolidating diverse analytical methods into a unified, robust system for 3D human tasks. Code and data are available at chathuman.github.io." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 502, + 395, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 502, + 395, + 514 + ], + "spans": [ + { + "bbox": [ + 314, + 502, + 395, + 514 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 312, + 521, + 557, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 521, + 557, + 715 + ], + "spans": [ + { + "bbox": [ + 312, + 521, + 557, + 715 + ], + "type": "text", + "content": "Research on 3D humans has progressed rapidly, resulting in the creation of many tools that can perform tasks like estimating a human's 3D pose from a single image [18, 25, 27, 36, 37], predicting face/body shapes [8, 15], capturing emotions [9, 15], and identifying regions of touch/contact [46, 68], generating human poses from text descriptions [10], and animating human images [84]. Each of these tools, however, focuses on a specific problem, functioning as isolated \"specialists\". Moreover, these separate tools cannot benefit from the expertise of others, and combining them to solve more complex tasks requires significant domain expertise. Ideally, we would have a single model that can adaptively leverage different tools to solve complex 3D human-related problems while offering intuitive user interaction through natural language input. Recent work such as ChatPose [16] has taken initial steps in this direction, unifying" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 693, + 297, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 693, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 693, + 297, + 714 + ], + "type": "text", + "content": "*Equal contribution. This work was done while YF and JL were at Meshcapade." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "8150" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "type": "text", + "content": "ing pose generation, estimation, and general understanding within an LLM framework. Unfortunately, ChatPose lacks the accuracy of the best specialist methods." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 110, + 295, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 110, + 295, + 266 + ], + "spans": [ + { + "bbox": [ + 54, + 110, + 295, + 266 + ], + "type": "text", + "content": "To address these issues, we build a multi-modal LLM, called ChatHuman, that specializes in using digital human modeling tools, enabling it to autonomously interpret instructions and complete diverse tasks related to 3D humans; see Fig. 1. Specifically, we teach an LLM to use a wide range of specialized human-related models for tasks like 3D pose estimation, emotion recognition, contact reasoning, and more, effectively extending the LLM's capabilities to the domain of 3D humans. This goes beyond providing a natural-language interface to these tools, as the LLM can use its broad understanding of humans to augment tool results or to analyze and integrate their outputs, providing better responses than any single tool alone." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 266, + 297, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 266, + 297, + 495 + ], + "spans": [ + { + "bbox": [ + 56, + 266, + 297, + 495 + ], + "type": "text", + "content": "With ChatHuman, we introduce a novel approach by fine-tuning an LLM to act as an agent that autonomously calls appropriate tools based on user inputs, completing tasks and enhancing responses with tool-generated results. Similar in spirit, recent works have employed off-the-shelf or fine-tuned LLMs for tasks like basic vision (e.g., Visual ChatGPT [53]), mobile applications (e.g., AppAgent [77]), biology (e.g., AmadeusGPT [78]) and system automation (e.g., GPT4Tools [74]). Our work, however, differs by focusing specifically on the unique challenges of 3D human understanding. This domain requires precise, specialized terminology and a nuanced understanding of 3D-specific tools, which conventional LLMs lack. To teach the network this specialized terminology, we do what we would do as humans – we have the LLM read the papers describing the methods. Even with that knowledge, the LLM needs to understand the task goals, select an appropriate tool or tools, interpret results, and resolve differing results. These skills are all beyond the abilities of general LLMs." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 496, + 295, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 496, + 295, + 616 + ], + "spans": [ + { + "bbox": [ + 55, + 496, + 295, + 616 + ], + "type": "text", + "content": "To address these challenges, we design the following training pipeline: 1) We utilize relevant literature about the tools to familiarize the LLM with domain knowledge, helping it know when and how to use these tools; 2) After using a tool, the LLM evaluates the reliability of the outcome using its \"judgment\" and compares different methods to identify the most reliable results; 3) It combines these results with its general knowledge to create response. This pipeline represents several key innovations, laying a foundation for LLMs to effectively handle complex, tool-driven 3D human tasks." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 618, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 295, + 714 + ], + "type": "text", + "content": "Retrieval-Augmented Tool Use: Details about tools are typically present in corresponding academic paper. We give the LLM access to these papers and demonstrate that \"reading the paper\" improves tool use performance. We further analyze which paper sections are most effective for instructing tool use. When encountering a new tool, users often turn to the user guide for assistance. We compile the complete documentation for these tools and utilize a paper-based" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 555, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 192 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 192 + ], + "type": "text", + "content": "Retrieval-Augmented Generation (RAG) mechanism to improve the LLM's understanding and management of new tools. This means that, although the LLM has not encountered such tools during fine-tuning, it can still effectively use the tools with the help of the paper-based RAG mechanism. In some cases, tasks require combining multiple tools. To address a broader range of tool usage scenarios, we employ a graph-based invocation system, which includes a node for single-tool use, a chain for sequential tool execution, and a DAG for multi-tool combinations as shown in Fig. 3." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 193, + 556, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 193, + 556, + 337 + ], + "spans": [ + { + "bbox": [ + 313, + 193, + 556, + 337 + ], + "type": "text", + "content": "3D Human-Related Tool Result Integration: Analyzing outputs from tools is crucial, as these outputs, such as body meshes, model parameters (e.g., SMPL pose), or motion sequences, are highly varied and complex. To make these results compatible with our LLM analysis system, we convert them into visual formats that the LLM can easily interpret. Guided by Cognitive Load Theory [63], we present these outputs as multiple-choice options, streamlining the selection process and enhancing the LLM's effectiveness in handling 3D human-related tasks. Combined with the LLM's extensive general knowledge, these integrated results enable it to generate sophisticated responses about 3D humans." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 337, + 556, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 337, + 556, + 480 + ], + "spans": [ + { + "bbox": [ + 313, + 337, + 556, + 480 + ], + "type": "text", + "content": "Specifically, ChatHuman consists of a multimodal LLM LLaVA [41], and 26 tools involving 3D Humans and general vision tasks. The LLM is finetuned to use these tools and incorporate their results. User requests can be in the form of text descriptions, images or other 3D information (if applicable), and the model produces text descriptions, images, or other 3D outputs after tool reasoning. Extensive evaluations demonstrate that ChatHuman not only surpasses previous models in tool-use accuracy but also improves performance on various human-related tasks. It achieves this by reasoning about multiple outputs, evaluating their veracity, and combining them with its own knowledge." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 482, + 556, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 482, + 556, + 650 + ], + "spans": [ + { + "bbox": [ + 313, + 482, + 556, + 650 + ], + "type": "text", + "content": "Summarizing, our key contributions include: (1) a framework that leverages LLMs to interact with users and address human-centric tasks using specialist tools; (2) a scientific-paper-based RAG mechanism that ensures precise tool use by comprehending tool descriptions from scholarly articles, enhancing tool applications and interactions; and (3) the integration of tool outcomes with LLMs, enabling the LLM to effectively explain tool results and interact with users. Additionally, the LLM is fine-tuned to distinguish between optimal and suboptimal tool results, improving overall accuracy. ChatHuman achieves superior performance in tool use and human-related tasks compared with other LLM-based methods or task-specific methods. The code, trained models, and datasets are available for research purposes." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 657, + 399, + 669 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 657, + 399, + 669 + ], + "spans": [ + { + "bbox": [ + 314, + 657, + 399, + 669 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "text", + "content": "3D Humans: Many 3D human analysis tasks leverage parametric models like SMPL [43], SMPL-X [48], or GHUM [73] for the body, BFM [49] or FLAME [35] for" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 313, + 757 + ], + "type": "text", + "content": "8151" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 300 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 300 + ], + "type": "text", + "content": "faces, and MANO [55] for hands. These models enable the representation of the human body, face, and hands as low-dimensional vectors, facilitating subsequent applications in estimation and generation. Estimation of human pose and shape either relies on optimization methods [4, 24] or regression methods [8, 14, 18, 25, 27, 32, 36, 37, 56, 81], which estimate body shape and pose parameters from a single image. Similarly, face reconstruction methods [12, 15, 66] estimate shape and expression parameters of the face model from single images. The analysis of contact, vital for understanding human-environment interaction and social touch, has seen recent attention [19, 46, 68]. Generative modeling techniques such as PoseScript [10] and PoseFix [11] provide methods for synthesizing and correcting 3D human poses based on textual descriptions, while language-to-3D generation methods [5, 20, 82] facilitate 3D avatar creation. Additionally, numerous methods generate human motions [1, 23, 33, 34, 45, 50, 64, 65], and recent language-to-video models are even able to generate humans moving [3, 84]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 303, + 296, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 303, + 296, + 495 + ], + "spans": [ + { + "bbox": [ + 55, + 303, + 296, + 495 + ], + "type": "text", + "content": "These basic methods excel in their respective scenarios, but are typically treated in isolation. When mature, such tools are often incorporated into software systems for animators that require significant domain knowledge. In contrast, recent generative models provide language interfaces to image, video, and 3D generation tools, making them accessible to novices. Until recently, such language-based control has not been possible for 3D humans. ChatPose [16] makes a step in this direction, unifying pose generation, estimation, and an LLM's general understanding into one model, but remains limited in its task capabilities. In contrast, our model integrates the performance of 26 3D human-related tasks into a single, LLM-based model. ChatHuman enables non-experts to solve real-world tasks by invoking appropriate tools and adding an extra layer of language-driven understanding that effectively leverages the tool outputs." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 498, + 298, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 298, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 298, + 713 + ], + "type": "text", + "content": "Large Language Models and Tool Use: To expand LLM capabilities without expensive retraining, recent work has focused on enabling them to use specialized tools. In this approach, a tool library is constructed and LLMs act as planners to coordinate tool usage. Various tools have been adopted, e.g., vision modules [53, 62, 76], mobile applications [77], community tools [58], special tools [79], and system tools [74]. However, general-purpose LLMs often lack a deep understanding of specific tools, especially those requiring domain knowledge. To address this, recent work [13, 28, 71] proposes to fintune general-purpose LLMs (e.g., LLaMA [67], LLaVA [40, 42]) with domain-specific tool-use data. Some methods use additional tool documentation to improve accuracy [21, 80], while others compose different tools to accomplish complex tasks [29, 57, 72]. Distinct from previous work, ChatHuman focuses on 3D human tasks through language interaction by leveraging off-the-shelf human-related tools." + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 319, + 73, + 553, + 159 + ], + "blocks": [ + { + "bbox": [ + 319, + 73, + 553, + 159 + ], + "lines": [ + { + "bbox": [ + 319, + 73, + 553, + 159 + ], + "spans": [ + { + "bbox": [ + 319, + 73, + 553, + 159 + ], + "type": "image", + "image_path": "c8a28cec09c4397d56651e279c93958dabb01518558dd93f02c2e68e2b1e75a8.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 164, + 555, + 220 + ], + "lines": [ + { + "bbox": [ + 313, + 164, + 555, + 220 + ], + "spans": [ + { + "bbox": [ + 313, + 164, + 555, + 220 + ], + "type": "text", + "content": "Figure 2. Paper-based Retrieval-Augmented Tool Use. We feed academic papers describing each tool to GPT-4 to build a document for each tool. During inference, given a user query, a relevant sample is retrieved from the documents and provided to the LLM-based agent as an in-context example to improve tool use accuracy." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 232, + 556, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 232, + 556, + 293 + ], + "spans": [ + { + "bbox": [ + 313, + 232, + 556, + 293 + ], + "type": "text", + "content": "Retrieval Augmented Generation: RAG [17, 31, 83] is a technique to enhance generative tasks by retrieving relevant information from external databases, allowing for continual knowledge update. Here, we design a RAG mechanism to facilitate the use of new unseen tools." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 303, + 370, + 316 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 303, + 370, + 316 + ], + "spans": [ + { + "bbox": [ + 314, + 303, + 370, + 316 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 322, + 412, + 335 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 322, + 412, + 335 + ], + "spans": [ + { + "bbox": [ + 314, + 322, + 412, + 335 + ], + "type": "text", + "content": "3.1. Overall Pipeline" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 341, + 556, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 341, + 556, + 449 + ], + "spans": [ + { + "bbox": [ + 313, + 341, + 556, + 449 + ], + "type": "text", + "content": "ChatHuman consists of a multimodal LLM " + }, + { + "bbox": [ + 313, + 341, + 556, + 449 + ], + "type": "inline_equation", + "content": "f_{\\phi}(\\cdot)" + }, + { + "bbox": [ + 313, + 341, + 556, + 449 + ], + "type": "text", + "content": ", along with a set of 3D human-related functions " + }, + { + "bbox": [ + 313, + 341, + 556, + 449 + ], + "type": "inline_equation", + "content": "f_{\\theta_1}(\\cdot), f_{\\theta_2}(\\cdot), \\ldots" + }, + { + "bbox": [ + 313, + 341, + 556, + 449 + ], + "type": "text", + "content": ". These functions serve as tools for various tasks, such as 3D human pose estimation, pose generation, and 3D face reconstruction. Our model takes input text queries " + }, + { + "bbox": [ + 313, + 341, + 556, + 449 + ], + "type": "inline_equation", + "content": "X_{q}" + }, + { + "bbox": [ + 313, + 341, + 556, + 449 + ], + "type": "text", + "content": ", images " + }, + { + "bbox": [ + 313, + 341, + 556, + 449 + ], + "type": "inline_equation", + "content": "X_{v}" + }, + { + "bbox": [ + 313, + 341, + 556, + 449 + ], + "type": "text", + "content": ", and optionally " + }, + { + "bbox": [ + 313, + 341, + 556, + 449 + ], + "type": "inline_equation", + "content": "X_{m}" + }, + { + "bbox": [ + 313, + 341, + 556, + 449 + ], + "type": "text", + "content": " representing other 3D human-related modalities (e.g., SMPL parameters for 3D human poses). Then it invokes tools and integrates their results to generate outputs as text " + }, + { + "bbox": [ + 313, + 341, + 556, + 449 + ], + "type": "inline_equation", + "content": "Y_{t}" + }, + { + "bbox": [ + 313, + 341, + 556, + 449 + ], + "type": "text", + "content": ", images " + }, + { + "bbox": [ + 313, + 341, + 556, + 449 + ], + "type": "inline_equation", + "content": "Y_{v}" + }, + { + "bbox": [ + 313, + 341, + 556, + 449 + ], + "type": "text", + "content": ", or 3D human-related modalities " + }, + { + "bbox": [ + 313, + 341, + 556, + 449 + ], + "type": "inline_equation", + "content": "Y_{m}" + }, + { + "bbox": [ + 313, + 341, + 556, + 449 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 456, + 492, + 468 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 456, + 492, + 468 + ], + "spans": [ + { + "bbox": [ + 313, + 456, + 492, + 468 + ], + "type": "text", + "content": "3.2. Retrieval-Augmented Tool Usage" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 474, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 474, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 474, + 556, + 715 + ], + "type": "text", + "content": "Teaching LLMs to decide when and how to use tools effectively is a significant challenge. A basic approach [53, 74] might involve including tool usage scenarios and input arguments within the LLM prompt, represented as " + }, + { + "bbox": [ + 313, + 474, + 556, + 715 + ], + "type": "inline_equation", + "content": "Y_{\\mathrm{tool}} = f_{\\phi}(X_q,X_t)" + }, + { + "bbox": [ + 313, + 474, + 556, + 715 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 474, + 556, + 715 + ], + "type": "inline_equation", + "content": "X_{t}" + }, + { + "bbox": [ + 313, + 474, + 556, + 715 + ], + "type": "text", + "content": " denotes tool definitions. However, this approach often falls short for specialized tools, especially given the variety of advanced tools for 3D human tasks. Many tools require background knowledge for correct use and have multiple application scenarios. For instance, the HMR tool [18] may be queried with requests like, \"Can you estimate this person's pose?\", \"What are the SMPL parameters?\", or \"Provide the 3D mesh of this person.\" Capturing all possible usage scenarios succinctly in a prompt is difficult, and as tools proliferate, prompt descriptions become unwieldy. To address these challenges, we introduce paper-based Retrieval-Augmented Generation (RAG) [30] and build a tool graph for tool combination. As shown in Fig. 2, we feed academic papers associated with each tool into GPT-4, prompting it to summarize the tool's functions and generate possible user queries for tool activation. These" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8152" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 72, + 294, + 131 + ], + "blocks": [ + { + "bbox": [ + 58, + 72, + 294, + 131 + ], + "lines": [ + { + "bbox": [ + 58, + 72, + 294, + 131 + ], + "spans": [ + { + "bbox": [ + 58, + 72, + 294, + 131 + ], + "type": "image", + "image_path": "a1e29eac086005c243e703ba19fe4add3ee45b7dfba3245a57fa16497403ddf5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 134, + 295, + 167 + ], + "lines": [ + { + "bbox": [ + 55, + 134, + 295, + 167 + ], + "spans": [ + { + "bbox": [ + 55, + 134, + 295, + 167 + ], + "type": "text", + "content": "Figure 3. Examples of tool use graphs. Tool use patterns include: a Node, which uses a single tool; a Chain, which requires sequential tool execution; and a DAG, which combines multiple tools." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 59, + 178, + 293, + 284 + ], + "blocks": [ + { + "bbox": [ + 59, + 178, + 293, + 284 + ], + "lines": [ + { + "bbox": [ + 59, + 178, + 293, + 284 + ], + "spans": [ + { + "bbox": [ + 59, + 178, + 293, + 284 + ], + "type": "image", + "image_path": "426956da5ff2035e46c7cfebbe5860d6a16b611375086f24a36f24a46f749f51.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 288, + 296, + 333 + ], + "lines": [ + { + "bbox": [ + 55, + 288, + 296, + 333 + ], + "spans": [ + { + "bbox": [ + 55, + 288, + 296, + 333 + ], + "type": "text", + "content": "Figure 4. Illustration of tool result integration. After tool invocation, results are transformed into VLM-compatible representations. The VLM then incorporates these transformed outputs to generate accurate, informative responses to user queries." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 353, + 296, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 353, + 296, + 498 + ], + "spans": [ + { + "bbox": [ + 55, + 353, + 296, + 498 + ], + "type": "text", + "content": "papers, with their rich background and detailed instructions, enable the generation of user queries that cover diverse use cases. By combining these queries with each tool's structured arguments, we compile a document of question-answer pairs for each tool's operation. Figure 2 provides an example from one of these documents. These documents serve as an auxiliary knowledge base during inference, from which we retrieve a relevant example " + }, + { + "bbox": [ + 55, + 353, + 296, + 498 + ], + "type": "inline_equation", + "content": "X_{e}" + }, + { + "bbox": [ + 55, + 353, + 296, + 498 + ], + "type": "text", + "content": " in response to a user query " + }, + { + "bbox": [ + 55, + 353, + 296, + 498 + ], + "type": "inline_equation", + "content": "X_{q}" + }, + { + "bbox": [ + 55, + 353, + 296, + 498 + ], + "type": "text", + "content": ". The retrieval process matches the text embedding of the query with embeddings in the tool documents using a text embedding model [61]. The retrieved sample is then presented to the agent " + }, + { + "bbox": [ + 55, + 353, + 296, + 498 + ], + "type": "inline_equation", + "content": "f_{\\phi}" + }, + { + "bbox": [ + 55, + 353, + 296, + 498 + ], + "type": "text", + "content": " as an in-context learning example:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 93, + 514, + 296, + 528 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 514, + 296, + 528 + ], + "spans": [ + { + "bbox": [ + 93, + 514, + 296, + 528 + ], + "type": "interline_equation", + "content": "X _ {e} = f _ {r} \\left(X _ {q}\\right), Y _ {\\text {t o o l}} = f _ {\\phi} \\left(X _ {q}, X _ {e}, X _ {t}\\right), \\tag {1}", + "image_path": "ebac8ac6c2b208ec376fc9a39fbd86a43d2c8a711a4f3817364b29ede104c0cc.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 542, + 297, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 542, + 297, + 578 + ], + "spans": [ + { + "bbox": [ + 55, + 542, + 297, + 578 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 542, + 297, + 578 + ], + "type": "inline_equation", + "content": "f_{r}" + }, + { + "bbox": [ + 55, + 542, + 297, + 578 + ], + "type": "text", + "content": " is the retrieval function, and " + }, + { + "bbox": [ + 55, + 542, + 297, + 578 + ], + "type": "inline_equation", + "content": "Y_{\\mathrm{tool}}" + }, + { + "bbox": [ + 55, + 542, + 297, + 578 + ], + "type": "text", + "content": " is a textual description of the tool invocation, specifying tool selection, names, and input arguments for tool calls." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 582, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 297, + 714 + ], + "type": "text", + "content": "Graph-based Tool Invocation. Note that the tool use description " + }, + { + "bbox": [ + 55, + 582, + 297, + 714 + ], + "type": "inline_equation", + "content": "Y_{tool}" + }, + { + "bbox": [ + 55, + 582, + 297, + 714 + ], + "type": "text", + "content": " varies depending on task settings, as shown for a single tool case in Fig. 2. However, some complex tasks require combining multiple tools. To handle this, we introduce a graph-based mechanism for tool invocation. We construct a tool graph with three structure types: nodes (single tool calls), chains (tool sequences for dependent tasks), and directed acyclic graphs (DAGs) [59] for complex multibranch operations. For each user query, the model predicts an appropriate tool graph and invokes the tools accordingly. Examples of tool graphs are shown in Fig. 3." + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 318, + 74, + 550, + 174 + ], + "blocks": [ + { + "bbox": [ + 318, + 74, + 550, + 174 + ], + "lines": [ + { + "bbox": [ + 318, + 74, + 550, + 174 + ], + "spans": [ + { + "bbox": [ + 318, + 74, + 550, + 174 + ], + "type": "image", + "image_path": "f00e8d6c697c375a36deb56f780b917385776a2f1bef2eeddaff0cff60e8a3af.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 180, + 555, + 213 + ], + "lines": [ + { + "bbox": [ + 313, + 180, + 555, + 213 + ], + "spans": [ + { + "bbox": [ + 313, + 180, + 555, + 213 + ], + "type": "text", + "content": "Figure 5. Illustration of tool results discrimination. When multiple plausible tools exist for a task, ChatHuman discriminates and chooses the best result as the final response." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 226, + 446, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 226, + 446, + 239 + ], + "spans": [ + { + "bbox": [ + 313, + 226, + 446, + 239 + ], + "type": "text", + "content": "3.3. Tool Result Integration" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 244, + 556, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 244, + 556, + 460 + ], + "spans": [ + { + "bbox": [ + 313, + 244, + 556, + 460 + ], + "type": "text", + "content": "After using tools, integrating their results is essential to effectively engage with users and solve problems. However, outputs from different tools vary widely, appearing as language, images, or vectors (like SMPL poses), which can challenge current multimodal LLMs, such as LLaVA [41], that process only text and images. To utilize these varied results and enhance the LLM's understanding of 3D humans, thereby improving its ability to apply world knowledge to problem-solving, we introduce a tool-conditioned transformation, " + }, + { + "bbox": [ + 313, + 244, + 556, + 460 + ], + "type": "inline_equation", + "content": "\\Psi(\\cdot)" + }, + { + "bbox": [ + 313, + 244, + 556, + 460 + ], + "type": "text", + "content": ". As shown in Fig. 4, this transformation converts tool outputs " + }, + { + "bbox": [ + 313, + 244, + 556, + 460 + ], + "type": "inline_equation", + "content": "Y_{m}" + }, + { + "bbox": [ + 313, + 244, + 556, + 460 + ], + "type": "text", + "content": " into textual or visual formats the LLM can process. For example, we transform the vertex-wise contact label from DECO [68] into body part-level descriptions using SMPL's [43] vertex-to-part mapping dictionary, and render the mesh generated by PoseScipt [10] into an RGB image using rendering techniques. See Sup. Mat. for more details. The transformed results are then merged with the user query as context for response generation:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 388, + 468, + 555, + 483 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 388, + 468, + 555, + 483 + ], + "spans": [ + { + "bbox": [ + 388, + 468, + 555, + 483 + ], + "type": "interline_equation", + "content": "Y _ {t} = f _ {\\phi} \\left(X _ {q}, \\Psi \\left(Y _ {m}\\right)\\right). \\tag {2}", + "image_path": "c0d523b6e85ec9dd9c342a560ef84f10ae2cc2dcdd52b05fe2423c47afb4f89a.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 490, + 555, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 490, + 555, + 526 + ], + "spans": [ + { + "bbox": [ + 313, + 490, + 555, + 526 + ], + "type": "text", + "content": "In scenarios where multiple tools can address a request (Fig. 5), we present outcomes as multiple-choice questions, prompting the model to select the most relevant answer:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 361, + 534, + 555, + 548 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 534, + 555, + 548 + ], + "spans": [ + { + "bbox": [ + 361, + 534, + 555, + 548 + ], + "type": "interline_equation", + "content": "Y _ {t} = f _ {\\phi} \\left(X _ {q}, \\Psi \\left(Y _ {m 1}\\right), \\Psi \\left(Y _ {m 2}\\right), \\dots\\right), \\tag {3}", + "image_path": "5e06e1b724875b098bdaa0477b867b31b99e5b4277006a53b287b675fa9d4d17.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 556, + 556, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 556, + 556, + 604 + ], + "spans": [ + { + "bbox": [ + 313, + 556, + 556, + 604 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 556, + 556, + 604 + ], + "type": "inline_equation", + "content": "Y_{mi}" + }, + { + "bbox": [ + 313, + 556, + 556, + 604 + ], + "type": "text", + "content": " denotes the " + }, + { + "bbox": [ + 313, + 556, + 556, + 604 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 556, + 556, + 604 + ], + "type": "text", + "content": "-th tool result. Since different tools have different failure modes, this process enables ChatHuman to identify the best method case by case, producing more accurate output than any individual method alone." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 623, + 468, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 623, + 468, + 635 + ], + "spans": [ + { + "bbox": [ + 313, + 623, + 468, + 635 + ], + "type": "text", + "content": "3.4. Training Data Construction" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "text", + "content": "Tool Usage Instruction-following Data. To teach the LLM-based agent to correctly use tools, we construct 90K instruction-response pairs about tool usage. Following GPT4Tools [74], we provide GPT-4 [47] with a textual description of COCO training images [39] and a tool-related prompt containing a tool description. To improve efficiency," + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "8153" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 64, + 550, + 277 + ], + "blocks": [ + { + "bbox": [ + 57, + 64, + 550, + 277 + ], + "lines": [ + { + "bbox": [ + 57, + 64, + 550, + 277 + ], + "spans": [ + { + "bbox": [ + 57, + 64, + 550, + 277 + ], + "type": "image", + "image_path": "04cef13e70446eb9d50cafc9e26de846c05d88e22127698bd1e3ec71a070a82e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 281, + 555, + 304 + ], + "lines": [ + { + "bbox": [ + 55, + 281, + 555, + 304 + ], + "spans": [ + { + "bbox": [ + 55, + 281, + 555, + 304 + ], + "type": "text", + "content": "Figure 6. Illustration of our instruction-following data construction pipeline. We construct tool usage and feedback data by providing GPT-4 with various tool-related information, image content, and ground truth labels. Gray text shows some example instructions." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 319, + 296, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 319, + 296, + 354 + ], + "spans": [ + { + "bbox": [ + 55, + 319, + 296, + 354 + ], + "type": "text", + "content": "we first prompt GPT-4 to summarize paper content, re-articulate tool functions, and enumerate 50 potential user queries for tool activation (see Fig. 6(a))." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 354, + 296, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 354, + 296, + 426 + ], + "spans": [ + { + "bbox": [ + 55, + 354, + 296, + 426 + ], + "type": "text", + "content": "Tool Feedback Instruction-following Data. To help the multimodal LLM model discriminate and integrate the tool results, we construct 88K pairs of instruction-following data based on existing datasets 3DPW [70], MOYO [69], PoseScript [10] and SHAPY [8] (see Fig. 6(b)(c)). Please see Sup. Mat. for more details about data construction." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 433, + 151, + 445 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 433, + 151, + 445 + ], + "spans": [ + { + "bbox": [ + 55, + 433, + 151, + 445 + ], + "type": "text", + "content": "3.5. Model Training" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 451, + 296, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 451, + 296, + 525 + ], + "spans": [ + { + "bbox": [ + 55, + 451, + 296, + 525 + ], + "type": "text", + "content": "Once we have data, we use LoRA [22] to finetune the LLM " + }, + { + "bbox": [ + 55, + 451, + 296, + 525 + ], + "type": "inline_equation", + "content": "f_{\\phi}(\\cdot)" + }, + { + "bbox": [ + 55, + 451, + 296, + 525 + ], + "type": "text", + "content": " with the cross entropy loss. More specifically, with the ground truth tool invocation labels " + }, + { + "bbox": [ + 55, + 451, + 296, + 525 + ], + "type": "inline_equation", + "content": "\\hat{Y}_{tool}" + }, + { + "bbox": [ + 55, + 451, + 296, + 525 + ], + "type": "text", + "content": " and response label " + }, + { + "bbox": [ + 55, + 451, + 296, + 525 + ], + "type": "inline_equation", + "content": "\\hat{Y}_t" + }, + { + "bbox": [ + 55, + 451, + 296, + 525 + ], + "type": "text", + "content": ", we optimize the model using the following objective function: " + }, + { + "bbox": [ + 55, + 451, + 296, + 525 + ], + "type": "inline_equation", + "content": "\\mathcal{L} = \\mathbf{CE}(\\hat{Y}_{tool}, Y_{tool}) + \\mathbf{CE}(\\hat{Y}_t, Y_t)" + }, + { + "bbox": [ + 55, + 451, + 296, + 525 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 451, + 296, + 525 + ], + "type": "inline_equation", + "content": "\\mathbf{CE}" + }, + { + "bbox": [ + 55, + 451, + 296, + 525 + ], + "type": "text", + "content": " denotes the cross-entropy loss. See Sup. Mat. for details." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 532, + 198, + 545 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 532, + 198, + 545 + ], + "spans": [ + { + "bbox": [ + 55, + 532, + 198, + 545 + ], + "type": "text", + "content": "4. Experiments and Results" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 552, + 189, + 564 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 552, + 189, + 564 + ], + "spans": [ + { + "bbox": [ + 55, + 552, + 189, + 564 + ], + "type": "text", + "content": "4.1. Implementation Details" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "content": "We use LLaVA-1.5 [41] as the VLM backbone, with CLIP [51] for vision encoding and Vicuna [7] for the LLM backbone. For retrieval, we adopt INSTRUCTOR [61] for text embedding and utilize Chroma's vector similarity searching algorithm to identify relevant examples. To preserve the generalization of the pretrained multi-modal LLM, we use LoRA [22] to perform efficient finetuning, with rank 128 and alpha 256. We implement tool utilization with LangChain [6], which enables automatic parsing of tool names and input parameters, followed by tool invocation Optimization uses AdamW [44], with a learning rate of 2e-4 and weight decay of 0. All models are finetuned over 2 epochs with a mix" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 315, + 311, + 553, + 384 + ], + "blocks": [ + { + "bbox": [ + 315, + 311, + 553, + 384 + ], + "lines": [ + { + "bbox": [ + 315, + 311, + 553, + 384 + ], + "spans": [ + { + "bbox": [ + 315, + 311, + 553, + 384 + ], + "type": "table", + "html": "
PerceptionReasoningGeneration
Body Pose Estimation [18]Selective Person Pose Detection [18, 41]Text-to-Pose Generation [10]
Body Shape Measurement [2]Specific Person Shape Measurement [2, 41]Speculative Pose Generation [10, 41]
Hand Pose Estimation [38]Targeted Hand Pose Estimation [38, 41]Text-to-Image Generation [54]
Face Reconstruction [15]Described Person Face Reconstruction [15, 41]Text-based Pose Editing [11]
Human Segmentation [26]Described Person Segmentation [26, 41]Remove Someone From The Photo [26, 41, 54]
HOI Detection [68]Selective Person Contact Estimation [41, 68]Replace Someone From The Photo [26, 41, 54]
Pose Description [10]Visual Question Answering [41]Instruct Image Using Text [54]
Image Caption [41]Text-to-Motion Generation [50]
Motion Capture [60]Text-to-Video Generation [50, 54, 84]
Image-to-Video Generation [50, 84]
", + "image_path": "fb2b54d2dc0029c0be1c5f3448e02ed55275852958ad8cb250d0d57671fc02c2.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 315, + 429, + 553, + 498 + ], + "blocks": [ + { + "bbox": [ + 313, + 388, + 555, + 421 + ], + "lines": [ + { + "bbox": [ + 313, + 388, + 555, + 421 + ], + "spans": [ + { + "bbox": [ + 313, + 388, + 555, + 421 + ], + "type": "text", + "content": "Table 1. ChatHuman supports 26 human-related tools, including 9 perception tools, 10 generation tools, and 7 reasoning tools. Tools in grey are unseen tools that are not included in the training data." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 429, + 553, + 498 + ], + "lines": [ + { + "bbox": [ + 315, + 429, + 553, + 498 + ], + "spans": [ + { + "bbox": [ + 315, + 429, + 553, + 498 + ], + "type": "table", + "html": "
MethodSeen ToolsUnseen Tools
SRtSRactSRargsSRIoUSRtSRactSRargsSRIoU
GPT4Tools [75]0.6090.5470.5250.5200.5660.6120.5460.5420.5250.573
GPT4Tools-FT [75]0.8250.7100.6870.6900.7410.9040.8070.6900.7470.800
Visual ChatGPT-3.5 [53]0.4980.3190.2370.2510.7910.5070.3140.2260.2930.803
Visual ChatGPT-4 [53]0.8920.8020.7150.7530.7970.9980.9130.8010.8720.907
ChatHuman1.0000.9740.9500.9700.9750.9990.9670.8930.9540.953
", + "image_path": "9f8f939c4bc3bc2810f5ced72bd5f80279d68dce1b696357d34fd5217f1b11ed.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 504, + 555, + 537 + ], + "lines": [ + { + "bbox": [ + 313, + 504, + 555, + 537 + ], + "spans": [ + { + "bbox": [ + 313, + 504, + 555, + 537 + ], + "type": "text", + "content": "Table 2. Tool use accuracy comparison. Successful rate of thought " + }, + { + "bbox": [ + 313, + 504, + 555, + 537 + ], + "type": "inline_equation", + "content": "(\\mathrm{SR}_{\\mathrm{t}})" + }, + { + "bbox": [ + 313, + 504, + 555, + 537 + ], + "type": "text", + "content": ", action " + }, + { + "bbox": [ + 313, + 504, + 555, + 537 + ], + "type": "inline_equation", + "content": "(\\mathrm{SR}_{\\mathrm{act}})" + }, + { + "bbox": [ + 313, + 504, + 555, + 537 + ], + "type": "text", + "content": ", arguments " + }, + { + "bbox": [ + 313, + 504, + 555, + 537 + ], + "type": "inline_equation", + "content": "(\\mathrm{SR}_{\\mathrm{args}})" + }, + { + "bbox": [ + 313, + 504, + 555, + 537 + ], + "type": "text", + "content": ", execution (SR), and IoU are reported." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 549, + 555, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 549, + 555, + 609 + ], + "spans": [ + { + "bbox": [ + 313, + 549, + 555, + 609 + ], + "type": "text", + "content": "ture of tool usage, tool feedback, and LLaVA multimodal instruction-tuning data, using 8 Nvidia A100-80G GPUs with the DeepSpeed [52] engine. Unless otherwise specified, we use LLaVA-1.5-7B as the base model for the ablation study. We support 26 tools, as listed in Tab. 1." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 613, + 457, + 625 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 613, + 457, + 625 + ], + "spans": [ + { + "bbox": [ + 313, + 613, + 457, + 625 + ], + "type": "text", + "content": "4.2. Evaluation on Tool Usage" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 630, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 555, + 713 + ], + "type": "text", + "content": "Tool Usage Benchmark. To evaluate tool usage accuracy, we construct a validation and test set. The validation set has 1000 samples with the same tools as the training set, while the test set includes 689 samples related to 3 tools unseen during training. Split of seen and unseen tools are detailed in Table 1. Similar to our training data construction, we feed a textual description of COCO validation set image, a" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8154" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 70, + 292, + 163 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 292, + 163 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 292, + 163 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 292, + 163 + ], + "type": "image", + "image_path": "48c2839a53c2d970e59df4baa3a9287a51c1c8cb7b72537477c6f0a9dcac77b4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 165, + 296, + 209 + ], + "lines": [ + { + "bbox": [ + 55, + 165, + 296, + 209 + ], + "spans": [ + { + "bbox": [ + 55, + 165, + 296, + 209 + ], + "type": "text", + "content": "Figure 7. Visualization of Animation Processing. Left: Conversation between the user and ChatHuman. Right: ChatHuman automatically calls tools to solve the task. ChatHuman can handle tasks beyond the capabilities of individual tools." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 214, + 295, + 286 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 214, + 295, + 286 + ], + "spans": [ + { + "bbox": [ + 54, + 214, + 295, + 286 + ], + "type": "text", + "content": "tool description, and some examples summarized from the tool paper into GPT-4 and prompt it to generate instruction-following data. We use the image description captioned by LLaVA [41] instead of the original image captions to ensure differences between training and test sets. Finally, we manually check all question-answering pairs for accuracy." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 287, + 296, + 393 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 287, + 296, + 393 + ], + "spans": [ + { + "bbox": [ + 54, + 287, + 296, + 393 + ], + "type": "text", + "content": "Baselines. We compare our method with Visual ChatGPT [53] and GPT4Tools [75] on the proposed evaluation set and report 5 metrics proposed in GPT4Tools [75]. See Sup. Mat. for details of the metrics. For Visual ChatGPT, we experiment with two versions of GPT: \"gpt-3.5-turbo-1106\" and \"gpt-4-turbo-preview\". For GPT4Tools, we adopt the official pretrained 13B model. For a fair comparison, we also finetune GPT4Tools with our training data using the official training code and obtain a variant, GPT4Tools-FT." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 396, + 296, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 396, + 296, + 491 + ], + "spans": [ + { + "bbox": [ + 54, + 396, + 296, + 491 + ], + "type": "text", + "content": "Table 2 shows that the original GPT4Tools struggles on our benchmark due to mismatches between its tools and our human-centric ones. Visual ChatGPT-4 exhibits impressive tool use accuracy, showing its powerful zero-shot ability to follow a standardized format and use tools accurately. ChatHuman consistently outperforms other methods, particularly with tools not seen in training, thanks to our paper-based RAG mechanism." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 496, + 265, + 508 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 496, + 265, + 508 + ], + "spans": [ + { + "bbox": [ + 55, + 496, + 265, + 508 + ], + "type": "text", + "content": "4.3. Evaluation on 3D Human Related Tasks" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 515, + 296, + 683 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 515, + 296, + 683 + ], + "spans": [ + { + "bbox": [ + 55, + 515, + 296, + 683 + ], + "type": "text", + "content": "Character Animation. ChatHuman employs tools for text-to-motion and image-to-video generation. We demonstrate how these tools are utilized to interact with users and reason about motions based on conversations in Fig. 7 and Fig. 1. ChatHuman can also tackle tasks that cannot be resolved with a single tool. For instance, text-to-human video generation poses significant challenges due to the complexity of motion. Therefore, another option is to first generate a motion sequence via text-to-motion generation, then apply a video generation model conditioned on this sequence. The internal processing within ChatHuman, detailing how it analyzes and solves tasks, is visualized in Fig. 7. We also compare our text-to-video generation results with those of Pika1. The qualitative comparisons are shown in Fig. 8." + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 316, + 71, + 553, + 170 + ], + "blocks": [ + { + "bbox": [ + 316, + 71, + 553, + 170 + ], + "lines": [ + { + "bbox": [ + 316, + 71, + 553, + 170 + ], + "spans": [ + { + "bbox": [ + 316, + 71, + 553, + 170 + ], + "type": "table", + "html": "
Method3DPW [70]RPE Benchmark [16]
MPJPE ↓PA-MPJPE ↓MPJRE ↓MPJPE ↓PA-MPJPE ↓MPJRE ↓
SPIN [27]102.962.910.1244.9107.312.4
HMR 2.0 [18]91.058.49.2225.2105.112.1
LLaVA-S [41]440.8205.421.8490.7207.421.1
LLaVA*-S [41]232.1101.112.8---
GPT4-S [47]322.0136.716.0---
LLaVA-P [41]335.2172.316.5391.5191.917.8
GPT4-P [47]396.5203.418.6---
ChatPose [16]163.681.910.4253.6103.811.7
ChatHuman91.358.79.2147.279.110.3
", + "image_path": "c17b93646c4df1699a3baba4f1aeb049e25b41917a4741d70d287b0983075cff.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 172, + 555, + 250 + ], + "lines": [ + { + "bbox": [ + 313, + 172, + 555, + 250 + ], + "spans": [ + { + "bbox": [ + 313, + 172, + 555, + 250 + ], + "type": "text", + "content": "Table 3. Comparison of vanilla human pose estimation and reasoning-based pose estimation on 3DPW [70] and RPE [16]. LLaVA* is fine-tuned with human keypoint data. \"S\" uses multimodal LLMs for keypoint detection and SMPLify [4] for pose optimization. \"P\" utilizes multimodal LLMs for textual pose descriptions, processed by PoseScript [10] to generate poses. MPJPE (in mm), MP-MPJPE (in mm), and MPJRE (" + }, + { + "bbox": [ + 313, + 172, + 555, + 250 + ], + "type": "inline_equation", + "content": "\\times 100" + }, + { + "bbox": [ + 313, + 172, + 555, + 250 + ], + "type": "text", + "content": ") are reported." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 317, + 258, + 553, + 350 + ], + "blocks": [ + { + "bbox": [ + 317, + 258, + 553, + 350 + ], + "lines": [ + { + "bbox": [ + 317, + 258, + 553, + 350 + ], + "spans": [ + { + "bbox": [ + 317, + 258, + 553, + 350 + ], + "type": "image", + "image_path": "03654e6214846602acdd54266ced783c595f6630aecbfad01c8e70afaf5d4632.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 353, + 555, + 387 + ], + "lines": [ + { + "bbox": [ + 313, + 353, + 555, + 387 + ], + "spans": [ + { + "bbox": [ + 313, + 353, + 555, + 387 + ], + "type": "text", + "content": "Figure 8. Left: Comparison to " + }, + { + "bbox": [ + 313, + 353, + 555, + 387 + ], + "type": "inline_equation", + "content": "\\mathrm{Pika}^1" + }, + { + "bbox": [ + 313, + 353, + 555, + 387 + ], + "type": "text", + "content": " on text to video generation. Right: Qualitative comparison with ChatPose [16] and HMR 2.0 [18] on reasoning-based human pose estimation task." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 399, + 556, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 399, + 556, + 650 + ], + "spans": [ + { + "bbox": [ + 313, + 399, + 556, + 650 + ], + "type": "text", + "content": "Pose Estimation. Following ChatPose [16], we evaluate the performance of our method on both classical and reasoning-based pose estimation (RPE) tasks. MPJPE, PA-MPJPE, and MPJRE on the 3DPW [70] and RPE [16] benchmarks are reported. For the reasoning-based pose estimation task, ChatHuman first grounds a human based on a textual description and feeds it into the pose estimation tool to get the result. As shown in Table 3, ChatHuman achieves comparable performance to the task-specific models on the classical pose estimation task. For reasoning-based human pose estimation, which involves both reasoning ability and advanced human pose estimation ability, ChatHuman outperforms both task-specific and multi-modal LLM methods by a large margin " + }, + { + "bbox": [ + 313, + 399, + 556, + 650 + ], + "type": "inline_equation", + "content": "(34.6\\% \\downarrow" + }, + { + "bbox": [ + 313, + 399, + 556, + 650 + ], + "type": "text", + "content": " in MPVPE). As shown in Fig. 8, only our method achieves a satisfactory result. The multimodal LLM competitor ChatPose finds the correct person but fails to obtain an accurate pose due to its limited perception ability, while the task-specific tool does not match the correct person due to the lack of reasoning ability. This demonstrates the advantage of ChatHuman, which combines task-specific tool use expertise with the general reasoning ability of an LLM." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 654, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 556, + 713 + ], + "type": "text", + "content": "Pose Generation. Here we evaluate the pose generation capability of ChatHuman on the classical text-to-pose generation task and the speculative pose generation task (SPG) [16]. Following previous work [10, 16], we report the text-to-pose recall rate " + }, + { + "bbox": [ + 313, + 654, + 556, + 713 + ], + "type": "inline_equation", + "content": "R^{T2P}" + }, + { + "bbox": [ + 313, + 654, + 556, + 713 + ], + "type": "text", + "content": " and pose-to-text recall rate " + }, + { + "bbox": [ + 313, + 654, + 556, + 713 + ], + "type": "inline_equation", + "content": "R^{P2T}" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 55, + 693, + 295, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 693, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 693, + 295, + 712 + ], + "type": "text", + "content": "1We use the demo available at https://pika.art/ (as of May 2025) to obtain the results." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8155" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 57, + 70, + 294, + 137 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 294, + 137 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 294, + 137 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 294, + 137 + ], + "type": "table", + "html": "
MethodPoseScript [10]SPG Benchmark [16]
\\( R^{P2T} \\)↑\\( R^{T2P} \\)↑\\( R^{P2T} \\)↑\\( R^{T2P} \\)↑
PoseScript [10]40.452.365.041.454.165.91.53.56.21.42.35.1
ChatPose [16]17.625.335.828.039.054.43.35.58.23.55.811.0
LLaVA-P [41]------2.14.07.12.13.36.1
GPT4-P [47]------2.74.79.22.75.38.2
ChatHuman41.852.665.142.152.366.53.25.09.93.56.510.6
", + "image_path": "d3cf2e7d69cf859eff8e4e40dc817c7432c35bef38b6a682ce6b22587e9db5ef.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 55, + 140, + 295, + 185 + ], + "lines": [ + { + "bbox": [ + 55, + 140, + 295, + 185 + ], + "spans": [ + { + "bbox": [ + 55, + 140, + 295, + 185 + ], + "type": "text", + "content": "Table 4. Comparison of classical and speculative pose generation on PoseScript [10] and SPG [16]. \"P\" denotes using LLMs to rephrase textual pose descriptions, which are then processed by PoseScript [10] to generate poses. Top 5, 10, 20 recall rates are reported." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 193, + 295, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 193, + 295, + 312 + ], + "spans": [ + { + "bbox": [ + 54, + 193, + 295, + 312 + ], + "type": "text", + "content": "of the retrieval models trained on real poses and evaluated on generated poses. For the SPG task, ChatHuman first rephrases the indirect pose descriptions into explicit ones and adopts PoseScript (journal version) [10] to generate a pose. As shown in Table 4, our method archives comparable performance to the SOTA methods on both benchmarks. In contrast, the previous LLM-based method, ChatPose, performs poorly on the classical pose generation benchmark, while the task-specific model, PoseScript, lags in the SPG benchmark due to limited reasoning ability." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 316, + 295, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 316, + 295, + 519 + ], + "spans": [ + { + "bbox": [ + 55, + 316, + 295, + 519 + ], + "type": "text", + "content": "Body Shape Measurement. We evaluate the body shape measurement accuracy of ChatHuman. We randomly sample 100 images from the HBW validation set [8] and compare our method with a multimodal LLM, LLaVA [41], and a SOTA body shape estimation method, CLIFF-BEDLAM [2]. For LLaVA and ChatHuman, we ask them the same question to inquire about the height, weight, chest, waist, and hip circumferences of a person and then prompt GPT-3.5 to extract the value from the model output. The details of the question and prompt are available in Sup. Mat. CLIFF-BEDLAM predicts the body shape parameters, which are then converted to measurements based on the shape-to-measurement function from SHAPY [8]. Anthropometric measurement errors are reported in Table 5(a). ChatHuman achieves superior performance in most measurements, outperforming the multimodal LLM competitor LLaVA by " + }, + { + "bbox": [ + 55, + 316, + 295, + 519 + ], + "type": "inline_equation", + "content": "42\\%" + }, + { + "bbox": [ + 55, + 316, + 295, + 519 + ], + "type": "text", + "content": " and CLIFF-BEDLAM by " + }, + { + "bbox": [ + 55, + 316, + 295, + 519 + ], + "type": "inline_equation", + "content": "15.7\\%" + }, + { + "bbox": [ + 55, + 316, + 295, + 519 + ], + "type": "text", + "content": " in average metric accuracy." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 522, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 522, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 522, + 295, + 713 + ], + "type": "text", + "content": "Human-Object Interaction (HoI). We evaluate the human-object interaction understanding ability of ChatHuman on the DECO [68] test set. The ground truth (GT) labels are obtained by converting the vertex-level contact labels into body part-level contact labels with SMPL's vertex-to-part mapping dictionary. Given a human image, we ask the multimodal LLM to detect the body parts contacting objects and prompt GPT-3.5 to extract the body part labels from the answer. Subsequently, we compare the predicted body parts with the GT label and compute the average detection precision, recall rate, and F1 Score. Table 5(b) shows that ChatHuman achieves SOTA precision and F1 score, demonstrating superior human-object interaction understanding ability. Notably, although LLaVA has a high recall rate, its precision and F1 score are rather poor, which means that it tends to predict all the body parts to be in contact with objects." + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 315, + 71, + 439, + 138 + ], + "blocks": [ + { + "bbox": [ + 315, + 71, + 439, + 138 + ], + "lines": [ + { + "bbox": [ + 315, + 71, + 439, + 138 + ], + "spans": [ + { + "bbox": [ + 315, + 71, + 439, + 138 + ], + "type": "table", + "html": "
MethodHeight ↓Chest ↓Waist ↓Hip ↓
LLaVA [41]6.716.522.917.6
CLIFF-BEDLAM [2]7.88.613.57.0
ChatHuman6.76.113.06.4
(a) Comparison of body shape measurement. Measurement errors (in cm) on HBW [8] are reported.
", + "image_path": "cd4187b31d0f42024c3a4a71201ada83c787e56b6a7800aeb6998383466528f7.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 332, + 142, + 534, + 152 + ], + "lines": [ + { + "bbox": [ + 332, + 142, + 534, + 152 + ], + "spans": [ + { + "bbox": [ + 332, + 142, + 534, + 152 + ], + "type": "text", + "content": "Table 5. Comparison of body shape and HOI estimation." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 443, + 71, + 551, + 138 + ], + "blocks": [ + { + "bbox": [ + 443, + 71, + 551, + 138 + ], + "lines": [ + { + "bbox": [ + 443, + 71, + 551, + 138 + ], + "spans": [ + { + "bbox": [ + 443, + 71, + 551, + 138 + ], + "type": "table", + "html": "
MethodPrecision↑ Recall↑ F1Score ↑
LLaVA [41]0.260.810.39
GPT-4 [47]0.610.480.49
ChatHuman0.670.670.63
(b) Comparison of HOI estimation. Precision, Recall, and F1 score on DECO [68] are reported.
", + "image_path": "4333de6ee1269e1a100485f6381d0119ac836b07550f4369fa2a56bcdc4931fa.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 162, + 408, + 175 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 162, + 408, + 175 + ], + "spans": [ + { + "bbox": [ + 313, + 162, + 408, + 175 + ], + "type": "text", + "content": "4.4. Ablation Study" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 182, + 555, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 182, + 555, + 396 + ], + "spans": [ + { + "bbox": [ + 313, + 182, + 555, + 396 + ], + "type": "text", + "content": "Paper-based RAG Mechanism. To improve tool use accuracy, we design a paper-based RAG mechanism. We perform a breakdown ablation to investigate the effect of each component and their interactions. The baseline model, created by removing the RAG operation and trained with instruction-following data without paper content, is compared in Table 6(a). The baseline model's success rate (SR) is 0.96 for seen tools and 0.82 for unseen tools. Adding RAG increases the SR for unseen tools to 0.89, demonstrating its effectiveness in zero-shot settings. Further incorporating articles into training data boosts the performance: the successful rate of arguments " + }, + { + "bbox": [ + 313, + 182, + 555, + 396 + ], + "type": "inline_equation", + "content": "(\\mathrm{SR}_{\\mathrm{args}})" + }, + { + "bbox": [ + 313, + 182, + 555, + 396 + ], + "type": "text", + "content": " rises from 0.93 to 0.95 for the seen tools and 0.84 to 0.94 for the unseen tools. This suggests that scholarly articles can help create high-quality instruction-following data and tool documents due to their detailed use instructions and diverse application scenarios. For a detailed analysis of each component's effect on instructing tool usage, please see Sup. Mat." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 398, + 556, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 398, + 556, + 518 + ], + "spans": [ + { + "bbox": [ + 313, + 398, + 556, + 518 + ], + "type": "text", + "content": "Multiple Tools Invocation. One of the advantages of using a VLM as an agent is its powerful generalization capacity. To test the robustness and generalization ability of ChatHuman, we conduct the following ablation study. During training, we only include the tool graphs with no more than three tools, while during evaluation, the user queries might need up to five tools to solve. Table 6(b) depicts the results. As shown, ChatHuman exhibits an excellent robustness in this out-of-domain setting (more than three tools combination) with an action accuracy higher than " + }, + { + "bbox": [ + 313, + 398, + 556, + 518 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 313, + 398, + 556, + 518 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 521, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 556, + 713 + ], + "type": "text", + "content": "Tool Result Integration. We first conduct an ablation to study how the tools can enhance the human understanding capacity of multimodal LLM. The model without tools is our multimodal LLM backbone, LLaVA-1.5-7B [41], and the model with tools is our ChatHuman. The quantitative results are listed in Table 7. When equipped with tools, the HOI contact detection F1 score increases from 0.39 to 0.63 and the average body shape measurement error declines by " + }, + { + "bbox": [ + 313, + 521, + 556, + 713 + ], + "type": "inline_equation", + "content": "38\\%" + }, + { + "bbox": [ + 313, + 521, + 556, + 713 + ], + "type": "text", + "content": ". These results demonstrate the effectiveness of tools in enriching the LLM's comprehension of human models and behaviors. Additionally, we study whether ChatHuman can utilize its world knowledge to discriminate and improve the tool performance. We design two discrimination schemes, i.e., selection and modification, and conduct an ablation study on two human-related tasks by comparing ChatHuman with the SOTA task-specific tools. For the selection scheme, we ex" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8156" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 83, + 70, + 329, + 127 + ], + "blocks": [ + { + "bbox": [ + 83, + 70, + 329, + 127 + ], + "lines": [ + { + "bbox": [ + 83, + 70, + 329, + 127 + ], + "spans": [ + { + "bbox": [ + 83, + 70, + 329, + 127 + ], + "type": "table", + "html": "
PaperRAGSeen ToolsUnseen Tools
SRtSRactSRargsSRIoUSRtSRactSRargsSRIoU
××0.9980.9670.9280.9600.9640.9460.8940.7750.8220.872
×✓1.0000.9670.9280.9610.9650.9960.9450.8420.8910.927
✓✓1.0000.9740.9500.9700.9750.9990.9670.8930.9540.953
", + "image_path": "1e6e12559e9ae42a8516043f9ab9450827d2e7d7ef54bb6535f7f36eb76f920f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 121, + 129, + 289, + 139 + ], + "lines": [ + { + "bbox": [ + 121, + 129, + 289, + 139 + ], + "spans": [ + { + "bbox": [ + 121, + 129, + 289, + 139 + ], + "type": "text", + "content": "(a) Ablation study of paper-based RAG mechanism." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 363, + 73, + 525, + 127 + ], + "blocks": [ + { + "bbox": [ + 363, + 73, + 525, + 127 + ], + "lines": [ + { + "bbox": [ + 363, + 73, + 525, + 127 + ], + "spans": [ + { + "bbox": [ + 363, + 73, + 525, + 127 + ], + "type": "table", + "html": "
Tool NumerSRtSRactSRargsSRIoU
20.9970.9600.9430.9280.973
30.9980.9590.9310.9320.974
40.9980.9430.9280.8750.968
50.9960.9290.8990.8470.950
", + "image_path": "59c9cd37d004dc0341c7ce6e8f0ae452ed8a572b06e45904f833380c5a8ca8a6.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 388, + 129, + 499, + 139 + ], + "lines": [ + { + "bbox": [ + 388, + 129, + 499, + 139 + ], + "spans": [ + { + "bbox": [ + 388, + 129, + 499, + 139 + ], + "type": "text", + "content": "(b) Ablation study of tool number." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_footnote" + }, + { + "bbox": [ + 91, + 140, + 517, + 152 + ], + "lines": [ + { + "bbox": [ + 91, + 140, + 517, + 152 + ], + "spans": [ + { + "bbox": [ + 91, + 140, + 517, + 152 + ], + "type": "text", + "content": "Table 6. Ablations related to tool usage. Successful rate of thought, action, arguments, execution, and IoU are reported." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 56, + 163, + 293, + 195 + ], + "blocks": [ + { + "bbox": [ + 56, + 163, + 293, + 195 + ], + "lines": [ + { + "bbox": [ + 56, + 163, + 293, + 195 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 293, + 195 + ], + "type": "table", + "html": "
MethodPrecision↑Recall↑F1Score↑MethodHeight↓Chest↓Waist↓Hip↓
w/o Tool0.260.810.39w/o Tool6.716.522.917.6
w/ Tool0.670.670.63w/ Tool6.76.113.06.4
", + "image_path": "f6024c8c892d5c29b7079fb047c6203a9a686c38257206cbe95f87aae6b3ef06.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 70, + 198, + 159, + 206 + ], + "lines": [ + { + "bbox": [ + 70, + 198, + 159, + 206 + ], + "spans": [ + { + "bbox": [ + 70, + 198, + 159, + 206 + ], + "type": "text", + "content": "(a) HOI Contact Detection." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_footnote" + }, + { + "bbox": [ + 183, + 198, + 282, + 207 + ], + "lines": [ + { + "bbox": [ + 183, + 198, + 282, + 207 + ], + "spans": [ + { + "bbox": [ + 183, + 198, + 282, + 207 + ], + "type": "text", + "content": "(b) Body Shape Measurement." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 56, + 233, + 176, + 271 + ], + "blocks": [ + { + "bbox": [ + 56, + 233, + 176, + 271 + ], + "lines": [ + { + "bbox": [ + 56, + 233, + 176, + 271 + ], + "spans": [ + { + "bbox": [ + 56, + 233, + 176, + 271 + ], + "type": "table", + "html": "
MethodMPJPE ↓PA-MPJPE ↓PA-MPVPE ↓
Tool A126.281.4101.9
Tool B124.084.6104.7
ChatHuman119.678.298.3
", + "image_path": "cbae583db8abd3c1c334636552c8c54224cd698ad609d1a3fba89e43cc6bbba4.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 56, + 273, + 175, + 282 + ], + "lines": [ + { + "bbox": [ + 56, + 273, + 175, + 282 + ], + "spans": [ + { + "bbox": [ + 56, + 273, + 175, + 282 + ], + "type": "text", + "content": "(a) Mesh Error (in mm) on MixPose." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 175, + 232, + 292, + 262 + ], + "blocks": [ + { + "bbox": [ + 56, + 209, + 295, + 231 + ], + "lines": [ + { + "bbox": [ + 56, + 209, + 295, + 231 + ], + "spans": [ + { + "bbox": [ + 56, + 209, + 295, + 231 + ], + "type": "text", + "content": "Table 7. Ablation study on the impact of tool usage for human-object contact detection and body shape estimation." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 175, + 232, + 292, + 262 + ], + "lines": [ + { + "bbox": [ + 175, + 232, + 292, + 262 + ], + "spans": [ + { + "bbox": [ + 175, + 232, + 292, + 262 + ], + "type": "table", + "html": "
MethodHeight ↓Chest ↓Waist ↓Hip ↓
Tool [2]7.88.613.57.0
ChatHuman6.76.113.06.4
", + "image_path": "8e3783a73a94862f6019a2977a0b3b84ccef7f178846c54ec67e7bc3ff80149e.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 179, + 264, + 292, + 282 + ], + "lines": [ + { + "bbox": [ + 179, + 264, + 292, + 282 + ], + "spans": [ + { + "bbox": [ + 179, + 264, + 292, + 282 + ], + "type": "text", + "content": "(b) Body Shape Measurement Error (in cm) on HBW [8]." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 284, + 296, + 306 + ], + "lines": [ + { + "bbox": [ + 56, + 284, + 296, + 306 + ], + "spans": [ + { + "bbox": [ + 56, + 284, + 296, + 306 + ], + "type": "text", + "content": "Table 8. Study revealing how tool use improves human understanding on pose estimation and body shape measurement tasks." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 58, + 308, + 171, + 365 + ], + "blocks": [ + { + "bbox": [ + 58, + 308, + 171, + 365 + ], + "lines": [ + { + "bbox": [ + 58, + 308, + 171, + 365 + ], + "spans": [ + { + "bbox": [ + 58, + 308, + 171, + 365 + ], + "type": "image", + "image_path": "5392627be968c7daa2879649bdbc37834a41ec515c7e336dfb41dd9a2384464d.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 78, + 366, + 162, + 373 + ], + "lines": [ + { + "bbox": [ + 78, + 366, + 162, + 373 + ], + "spans": [ + { + "bbox": [ + 78, + 366, + 162, + 373 + ], + "type": "text", + "content": "(a) Tool generates unreasonable output" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 376, + 296, + 409 + ], + "lines": [ + { + "bbox": [ + 55, + 376, + 296, + 409 + ], + "spans": [ + { + "bbox": [ + 55, + 376, + 296, + 409 + ], + "type": "text", + "content": "Figure 9. Illustration of how ChatHuman discriminates and integrates tool results. The Agent will fix the unreasonable tool result and integrate the reasonable tool result to generate a final response." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 171, + 308, + 288, + 365 + ], + "blocks": [ + { + "bbox": [ + 171, + 308, + 288, + 365 + ], + "lines": [ + { + "bbox": [ + 171, + 308, + 288, + 365 + ], + "spans": [ + { + "bbox": [ + 171, + 308, + 288, + 365 + ], + "type": "image", + "image_path": "728061aad1464158687002c1bc86383ed210f88f91d32c7b96b015e60ae2676f.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 191, + 366, + 270, + 373 + ], + "lines": [ + { + "bbox": [ + 191, + 366, + 270, + 373 + ], + "spans": [ + { + "bbox": [ + 191, + 366, + 270, + 373 + ], + "type": "text", + "content": "(b) Tool generates reasonable output" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "bbox": [ + 55, + 413, + 296, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 413, + 296, + 665 + ], + "spans": [ + { + "bbox": [ + 55, + 413, + 296, + 665 + ], + "type": "text", + "content": "experiment with the pose estimation task and select two SOTA methods, HMR 2.0 [18] and CLIFF-SMPLify [4, 36], as our tools to generate two poses of each person. We then prompt the LLM-based agent to discriminate the results and choose the better one as the final response. Different tools excel in different scenarios and, to cover more diverse human poses and camera views, we build a new benchmark MixPose by selecting 100 images with extreme camera views from the MoYo [69] test set, 100 full-body samples and 100 severely-truncated samples from 3DPW [70] test set. Details of the prompt and MixPose benchmark are in Sup. Mat. As shown in Table 8 (a), ChatHuman archives a lower reconstruction error on the MixPose benchmark, validating the agent's effectiveness as a discriminator to improve tool performance. For the modification scheme, we validate on the body shape measurement task. We use CLIFF-BEDLAM [2] as tool and prompt the agent to discriminate and modify the tool result. The result is reported in Table 8 (b), and Fig. 9. The LLM-based agent enhances tool performance by using its general world knowledge to identify and correct unreasonable tool results, such as height and weight in Fig. 9 (a)." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 55, + 672, + 258, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 672, + 258, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 672, + 258, + 685 + ], + "type": "text", + "content": "5. Discussion and Concluding Remarks" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 55, + 689, + 295, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 712 + ], + "type": "text", + "content": "We introduce ChatHuman, an LLM-based model designed to learn the use of tools related to 3D humans and assist users" + } + ] + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 317, + 163, + 554, + 232 + ], + "blocks": [ + { + "bbox": [ + 317, + 163, + 554, + 232 + ], + "lines": [ + { + "bbox": [ + 317, + 163, + 554, + 232 + ], + "spans": [ + { + "bbox": [ + 317, + 163, + 554, + 232 + ], + "type": "image", + "image_path": "26d1acca3eb15e0015bcb7dce7482816c37cf4da2ce94530c81c44d99a1ffac9.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 241, + 554, + 263 + ], + "lines": [ + { + "bbox": [ + 313, + 241, + 554, + 263 + ], + "spans": [ + { + "bbox": [ + 313, + 241, + 554, + 263 + ], + "type": "text", + "content": "Figure 10. Human interaction can improve the performance and tool usage accuracy of ChatHuman." + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 280, + 555, + 328 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 280, + 555, + 328 + ], + "spans": [ + { + "bbox": [ + 313, + 280, + 555, + 328 + ], + "type": "text", + "content": "in solving tasks associated with 3D humans. ChatHuman processes requests from users, analyzes the needs, and utilizes the necessary tools. It then evaluates the tools' outputs to respond to the user's queries." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 332, + 556, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 332, + 556, + 571 + ], + "spans": [ + { + "bbox": [ + 313, + 332, + 556, + 571 + ], + "type": "text", + "content": "Limitations. ChatHuman may fail in certain calling scenarios, particularly when the user request is vague, and subsequent LLM internal analysis cannot rectify an incorrect initial function call. However, further interaction with users can remedy this if they provide additional information. Fig. 10 illustrates an instance of using body estimation and face reconstruction tools for avatar creation. Even with the application and analysis of the tool, outcomes like height estimation may not be entirely precise. One contributing factor is the accuracy of the training data; for instance, most height labels in datasets use the official height of models or celebrities, which may not account for variations like shoe height, such as a 7-inch heel. Incorporating more cues from users, combined with the LLM's knowledge of the world and reasoning capabilities, can enhance result accuracy, as shown in Fig. 10. Our system is currently limited by the academic methods used. Incorporating better academic methods will enhance model performance. Notably, adding new tools requires no additional training, allowing our method to evolve and improve as new techniques are developed." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 576, + 555, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 576, + 555, + 624 + ], + "spans": [ + { + "bbox": [ + 313, + 576, + 555, + 624 + ], + "type": "text", + "content": "Future Work. ChatHuman offers several avenues for future development. In particular, user interaction/dialog offers opportunities to learn from user feedback. This could exploit reinforcement learning to refine the model's understanding." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 636, + 499, + 649 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 636, + 499, + 649 + ], + "spans": [ + { + "bbox": [ + 313, + 636, + 499, + 649 + ], + "type": "text", + "content": "Acknowledgements and Disclosures" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 313, + 654, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 555, + 713 + ], + "type": "text", + "content": "We thank Naureen Mahmood, Nicolas Keller and Nicolas Heron for the support of the data collection. While MJB is a co-founder and Chief Scientist at Meshcapade, his research in this project was performed solely at, and funded solely by, the Max Planck Society." + } + ] + } + ], + "index": 28 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8157" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 297, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 91, + 297, + 133 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 297, + 133 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 297, + 133 + ], + "type": "text", + "content": "[1] Kfir Aberman, Rundi Wu, Dani Lischinski, Baoquan Chen, and Daniel Cohen-Or. Learning character-agnostic motion for motion retargeting in 2d. arXiv preprint arXiv:1905.01680, 2019.3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 136, + 296, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 136, + 296, + 190 + ], + "spans": [ + { + "bbox": [ + 61, + 136, + 296, + 190 + ], + "type": "text", + "content": "[2] Michael J Black, Priyanka Patel, Joachim Tesch, and Jinlong Yang. Bedlam: A synthetic dataset of bodies exhibiting detailed lifelike animated motion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8726-8737, 2023. 5, 7, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 191, + 296, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 191, + 296, + 246 + ], + "spans": [ + { + "bbox": [ + 62, + 191, + 296, + 246 + ], + "type": "text", + "content": "[3] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 247, + 296, + 290 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 247, + 296, + 290 + ], + "spans": [ + { + "bbox": [ + 62, + 247, + 296, + 290 + ], + "type": "text", + "content": "[4] Federica Bogo, Angjoo Kanazawa, Christoph Lassner, Peter Gehler, Javier Romero, and Michael J. Black. Keep it SMPL: Automatic estimation of 3D human pose and shape from a single image. In ECCV, 2016. 3, 6, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 291, + 296, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 291, + 296, + 334 + ], + "spans": [ + { + "bbox": [ + 62, + 291, + 296, + 334 + ], + "type": "text", + "content": "[5] Yukang Cao, Yan-Pei Cao, Kai Han, Ying Shan, and KwanYee K Wong. Dreamavatar: Text-and-shape guided 3d human avatar generation via diffusion models. arXiv preprint arXiv:2304.00916, 2023. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 335, + 296, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 335, + 296, + 357 + ], + "spans": [ + { + "bbox": [ + 62, + 335, + 296, + 357 + ], + "type": "text", + "content": "[6] Harrison Chase and LangChain Contributors. Langchain, 2022. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 358, + 296, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 358, + 296, + 413 + ], + "spans": [ + { + "bbox": [ + 62, + 358, + 296, + 413 + ], + "type": "text", + "content": "[7] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhanghao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E. Gonzalez, Ion Stoica, and Eric P. Xing. Vicuna: An open-source chatbot impressing gpt-4 with " + }, + { + "bbox": [ + 62, + 358, + 296, + 413 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 62, + 358, + 296, + 413 + ], + "type": "text", + "content": " * chatgpt quality, 2023. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 414, + 296, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 414, + 296, + 479 + ], + "spans": [ + { + "bbox": [ + 62, + 414, + 296, + 479 + ], + "type": "text", + "content": "[8] Vasileios Choutas, Lea Müller, Chun-Hao P Huang, Siyu Tang, Dimitrios Tzionas, and Michael J Black. Accurate 3d body shape regression using metric and semantic attributes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2718-2728, 2022. 1, 3, 5, 7, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 480, + 296, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 480, + 296, + 524 + ], + "spans": [ + { + "bbox": [ + 62, + 480, + 296, + 524 + ], + "type": "text", + "content": "[9] Radek Daněček, Michael J Black, and Timo Bolkart. Emoca: Emotion driven monocular face capture and animation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20311-20322, 2022. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 525, + 296, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 525, + 296, + 567 + ], + "spans": [ + { + "bbox": [ + 57, + 525, + 296, + 567 + ], + "type": "text", + "content": "[10] Ginger Delmas, Philippe Weinzaepfel, Thomas Lucas, Francesc Moreno-Noguer, and Grégory Rogez. Posescript: 3d human poses from natural language. In ECCV, 2022. 1, 3, 4, 5, 6, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 569, + 296, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 569, + 296, + 624 + ], + "spans": [ + { + "bbox": [ + 57, + 569, + 296, + 624 + ], + "type": "text", + "content": "[11] Ginger Delmas, Philippe Weinzaepfel, Francesc Moreno-Noguer, and Grégory Rogez. Posefix: Correcting 3d human poses with natural language. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15018-15028, 2023. 3, 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 625, + 296, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 625, + 296, + 677 + ], + "spans": [ + { + "bbox": [ + 57, + 625, + 296, + 677 + ], + "type": "text", + "content": "[12] Yu Deng, Jiaolong Yang, Sicheng Xu, Dong Chen, Yunde Jia, and Xin Tong. Accurate 3d face reconstruction with weakly-supervised learning: From single image to image set. In IEEE Computer Vision and Pattern Recognition Workshops, 2019. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 680, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 680, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 680, + 296, + 713 + ], + "type": "text", + "content": "[13] Yu Du, Fangyun Wei, and Hongyang Zhang. Anytool: Self-reflective, hierarchical agents for large-scale api calls. arXiv preprint arXiv:2402.04253, 2024. 3" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 316, + 73, + 555, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 555, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 555, + 106 + ], + "type": "text", + "content": "[14] Yao Feng, Vasileios Choutas, Timo Bolkart, Dimitrios Tzionas, and Michael J. Black. Collaborative regression of expressive bodies using moderation. In 3DV, 2021. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 107, + 555, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 107, + 555, + 150 + ], + "spans": [ + { + "bbox": [ + 317, + 107, + 555, + 150 + ], + "type": "text", + "content": "[15] Yao Feng, Haiwen Feng, Michael J Black, and Timo Bolkart. Learning an animatable detailed 3d face model from in-the-wild images. ACM Transactions on Graphics, 40(4):1-13, 2021. 1, 3, 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 152, + 554, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 554, + 185 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 554, + 185 + ], + "type": "text", + "content": "[16] Yao Feng, Jing Lin, Sai Kumar Dwivedi, Yu Sun, Priyanka Patel, and Michael J. Black. ChatPose: Chatting about 3d human pose. In CVPR, 2024. 1, 3, 6, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 186, + 555, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 186, + 555, + 230 + ], + "spans": [ + { + "bbox": [ + 317, + 186, + 555, + 230 + ], + "type": "text", + "content": "[17] Yunfan Gao, Yun Xiong, Xinyu Gao, Kangxiang Jia, Jinliu Pan, Yuxi Bi, Yi Dai, Jiawei Sun, and Haofen Wang. Retrievalaugmented generation for large language models: A survey. arXiv preprint arXiv:2312.10997, 2023. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 232, + 555, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 232, + 555, + 274 + ], + "spans": [ + { + "bbox": [ + 317, + 232, + 555, + 274 + ], + "type": "text", + "content": "[18] Shubham Goel, Georgios Pavlakos, Jathushan Rajasegaran, Angjoo Kanazawa, and Jitendra Malik. Humans in 4D: Reconstructing and tracking humans with transformers. In ICCV, 2023. 1, 3, 5, 6, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 276, + 555, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 276, + 555, + 330 + ], + "spans": [ + { + "bbox": [ + 317, + 276, + 555, + 330 + ], + "type": "text", + "content": "[19] Sookwan Han and Hanbyul Joo. Chorus: Learning canonicalized 3d human-object spatial relations from unbounded synthesized images. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15835-15846, 2023. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 331, + 555, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 331, + 555, + 374 + ], + "spans": [ + { + "bbox": [ + 316, + 331, + 555, + 374 + ], + "type": "text", + "content": "[20] Fangzhou Hong, Mingyuan Zhang, Liang Pan, Zhongang Cai, Lei Yang, and Ziwei Liu. Avatarclip: Zero-shot text-driven generation and animation of 3d avatars. arXiv preprint arXiv:2205.08535, 2022. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 376, + 554, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 376, + 554, + 430 + ], + "spans": [ + { + "bbox": [ + 316, + 376, + 554, + 430 + ], + "type": "text", + "content": "[21] Cheng-Yu Hsieh, Si-An Chen, Chun-Liang Li, Yasuhisa Fujii, Alexander Ratner, Chen-Yu Lee, Ranjay Krishna, and Tomas Pfister. Tool documentation enables zero-shot tool-usage with large language models. arXiv preprint arXiv:2308.00675, 2023. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 432, + 555, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 432, + 555, + 475 + ], + "spans": [ + { + "bbox": [ + 316, + 432, + 555, + 475 + ], + "type": "text", + "content": "[22] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv:2106.09685, 2021. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 477, + 555, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 477, + 555, + 510 + ], + "spans": [ + { + "bbox": [ + 316, + 477, + 555, + 510 + ], + "type": "text", + "content": "[23] Yifeng Jiang, Jungdam Won, Yuting Ye, and C Karen Liu. Drop: Dynamics responses from human motion prior and projective dynamics. SIGGRAPH Asia, 2023. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 512, + 555, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 512, + 555, + 544 + ], + "spans": [ + { + "bbox": [ + 316, + 512, + 555, + 544 + ], + "type": "text", + "content": "[24] Hanbyul Joo, Natalia Neverova, and Andrea Vedaldi. Exemplar fine-tuning for 3D human pose fitting towards in-the-wild 3D human pose estimation. In 3DV, 2020. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 545, + 554, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 545, + 554, + 578 + ], + "spans": [ + { + "bbox": [ + 317, + 545, + 554, + 578 + ], + "type": "text", + "content": "[25] Angjoo Kanazawa, Michael J. Black, David W. Jacobs, and Jitendra Malik. End-to-end recovery of human shape and pose. In CVPR, 2018. 1, 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 579, + 555, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 579, + 555, + 635 + ], + "spans": [ + { + "bbox": [ + 317, + 579, + 555, + 635 + ], + "type": "text", + "content": "[26] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 5" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 635, + 554, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 635, + 554, + 678 + ], + "spans": [ + { + "bbox": [ + 317, + 635, + 554, + 678 + ], + "type": "text", + "content": "[27] Nikos Kolotouros, Georgios Pavlakos, Michael J. Black, and Kostas Daniilidis. Learning to reconstruct 3D human pose and shape via model-fitting in the loop. In ICCV, pages 2252-2261, 2019. 1, 3, 6" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 680, + 555, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 680, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 680, + 555, + 713 + ], + "type": "text", + "content": "[28] Yilun Kong, Jingqing Ruan, Yihong Chen, Bin Zhang, Tianpeng Bao, Shiwei Shi, Guoqing Du, Xiaoru Hu, Hangyu Mao, Ziyue Li, et al. Tptu-v2: Boosting task planning and tool" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8158" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 77, + 73, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 73, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 77, + 73, + 294, + 95 + ], + "type": "text", + "content": "usage of large language model-based agents in real-world systems. arXiv preprint arXiv:2311.11315, 2023. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 95, + 296, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 95, + 296, + 149 + ], + "spans": [ + { + "bbox": [ + 57, + 95, + 296, + 149 + ], + "type": "text", + "content": "[29] Yilun Kong, Jingqing Ruan, Yihong Chen, Bin Zhang, Tianpeng Bao, Shiwei Shi, Guoqing Du, Xiaoru Hu, Hangyu Mao, Ziyue Li, et al. Tptu-v2: Boosting task planning and tool usage of large language model-based agents in real-world systems. arXiv preprint arXiv:2311.11315, 2023. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 150, + 296, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 150, + 296, + 215 + ], + "spans": [ + { + "bbox": [ + 56, + 150, + 296, + 215 + ], + "type": "text", + "content": "[30] Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Kuttler, Mike Lewis, Wen-tau Yih, Tim Roektaschel, et al. Retrievalaugmented generation for knowledge-intensive nlp tasks. Advances in neural information processing systems, 33:9459-9474, 2020. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 217, + 296, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 217, + 296, + 281 + ], + "spans": [ + { + "bbox": [ + 56, + 217, + 296, + 281 + ], + "type": "text", + "content": "[31] Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Kuttler, Mike Lewis, Wen-tau Yih, Tim Roektaschel, et al. Retrievalaugmented generation for knowledge-intensive nlp tasks. Advances in Neural Information Processing Systems, 33:9459-9474, 2020. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 282, + 296, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 282, + 296, + 326 + ], + "spans": [ + { + "bbox": [ + 56, + 282, + 296, + 326 + ], + "type": "text", + "content": "[32] Jiefeng Li, Chao Xu, Zhicun Chen, Siyuan Bian, Lixin Yang, and Cewu Lu. HybrIK: A hybrid analytical-neural inverse kinematics solution for 3D human pose and shape estimation. In CVPR, 2021. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 327, + 294, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 327, + 294, + 358 + ], + "spans": [ + { + "bbox": [ + 56, + 327, + 294, + 358 + ], + "type": "text", + "content": "[33] Jiaman Li, Jiajun Wu, and C Karen Liu. Object motion guided human motion synthesis. ACM Transactions on Graphics (TOG), 42(6):1-11, 2023. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 360, + 294, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 360, + 294, + 403 + ], + "spans": [ + { + "bbox": [ + 56, + 360, + 294, + 403 + ], + "type": "text", + "content": "[34] Peizhuo Li, Kfir Aberman, Rana Hanocka, Libin Liu, Olga Sorkine-Hornung, and Baoquan Chen. Learning skeletal articulations with neural blend shapes. ACM Transactions on Graphics (TOG), 40(4):1, 2021. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 404, + 296, + 447 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 404, + 296, + 447 + ], + "spans": [ + { + "bbox": [ + 56, + 404, + 296, + 447 + ], + "type": "text", + "content": "[35] Tianye Li, Timo Bolkart, Michael. J. Black, Hao Li, and Javier Romero. Learning a model of facial shape and expression from 4D scans. ACM Transactions on Graphics, (Proc. SIGGRAPH Asia), 36(6):194:1-194:17, 2017. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 449, + 296, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 449, + 296, + 491 + ], + "spans": [ + { + "bbox": [ + 56, + 449, + 296, + 491 + ], + "type": "text", + "content": "[36] Zhihao Li, Jianzhuang Liu, Zhensong Zhang, Songcen Xu, and Youliang Yan. CLIFF: Carrying location information in full frames into human pose and shape estimation. In ECCV, 2022. 1, 3, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 492, + 294, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 492, + 294, + 524 + ], + "spans": [ + { + "bbox": [ + 56, + 492, + 294, + 524 + ], + "type": "text", + "content": "[37] Jing Lin, Ailing Zeng, Haoqian Wang, Lei Zhang, and Yu Li. One-stage 3d whole-body mesh recovery with component aware transformer. CVPR, 2023. 1, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 525, + 296, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 525, + 296, + 569 + ], + "spans": [ + { + "bbox": [ + 56, + 525, + 296, + 569 + ], + "type": "text", + "content": "[38] Kevin Lin, Lijuan Wang, and Zicheng Liu. End-to-end human pose and mesh reconstruction with transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1954-1963, 2021. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 570, + 294, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 570, + 294, + 612 + ], + "spans": [ + { + "bbox": [ + 56, + 570, + 294, + 612 + ], + "type": "text", + "content": "[39] Tsung-Yi Lin, Michael Maire, Serge J. Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C. Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 614, + 296, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 614, + 296, + 645 + ], + "spans": [ + { + "bbox": [ + 56, + 614, + 296, + 645 + ], + "type": "text", + "content": "[40] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. arXiv preprint arXiv:2304.08485, 2023. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 647, + 296, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 296, + 669 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 296, + 669 + ], + "type": "text", + "content": "[41] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, 2023. 2, 4, 5, 6, 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 670, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 296, + 713 + ], + "type": "text", + "content": "[42] Shilong Liu, Hao Cheng, Haotian Liu, Hao Zhang, Feng Li, Tianhe Ren, Xueyan Zou, Jianwei Yang, Hang Su, Jun Zhu, Lei Zhang, Jianfeng Gao, and Chunyuan Li. Llava-plus: Learning to use tools for creating multimodal agents, 2023. 3" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 316, + 73, + 555, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 555, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 555, + 106 + ], + "type": "text", + "content": "[43] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J. Black. SMPL: A skinned multiperson linear model. In ACM TOG, 2015. 2, 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 107, + 554, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 554, + 129 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 554, + 129 + ], + "type": "text", + "content": "[44] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 130, + 554, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 130, + 554, + 174 + ], + "spans": [ + { + "bbox": [ + 316, + 130, + 554, + 174 + ], + "type": "text", + "content": "[45] Thomas Lucas*, Fabien Baradel*, Philippe Weinzaepfel, and Grégory Rogez. Posegpt: Quantization-based 3d human motion generation and forecasting. In European Conference on Computer Vision (ECCV), 2022. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 175, + 555, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 555, + 219 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 555, + 219 + ], + "type": "text", + "content": "[46] Lea Muller, Ahmed AA Osman, Siyu Tang, Chun-Hao P Huang, and Michael J Black. On self-contact and human pose. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9990-9999, 2021. 1, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 220, + 504, + 231 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 220, + 504, + 231 + ], + "spans": [ + { + "bbox": [ + 316, + 220, + 504, + 231 + ], + "type": "text", + "content": "[47] OpenAI. GPT-4 technical report. 2023. 4, 6, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 232, + 555, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 232, + 555, + 275 + ], + "spans": [ + { + "bbox": [ + 316, + 232, + 555, + 275 + ], + "type": "text", + "content": "[48] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed A. A. Osman, Dimitrios Tzionas, and Michael J. Black. Expressive body capture: 3D hands, face, and body from a single image. In CVPR, 2019. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 277, + 555, + 309 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 277, + 555, + 309 + ], + "spans": [ + { + "bbox": [ + 316, + 277, + 555, + 309 + ], + "type": "text", + "content": "[49] P. Paysan, R. Knothe, B. Amberg, S. Romdhani, and T. Vetter. A 3D face model for pose and illumination invariant face recognition. Genova, Italy, 2009. IEEE. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 311, + 554, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 311, + 554, + 354 + ], + "spans": [ + { + "bbox": [ + 316, + 311, + 554, + 354 + ], + "type": "text", + "content": "[50] Mathis Petrovich, Michael J. Black, and Gül Varol. TMR: Text-to-motion retrieval using contrastive 3D human motion synthesis. In International Conference on Computer Vision (ICCV), 2023. 3, 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 356, + 555, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 356, + 555, + 421 + ], + "spans": [ + { + "bbox": [ + 316, + 356, + 555, + 421 + ], + "type": "text", + "content": "[51] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 422, + 555, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 422, + 555, + 466 + ], + "spans": [ + { + "bbox": [ + 316, + 422, + 555, + 466 + ], + "type": "text", + "content": "[52] Jeff Rasley, Samyam Rajbhandari, Olatunjri Ruwase, and Yuxiong He. Deepspeed: System optimizations enable training deep learning models with over 100 billion parameters. arXiv preprint arXiv:2002.11681, 2020. 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 468, + 555, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 468, + 555, + 510 + ], + "spans": [ + { + "bbox": [ + 316, + 468, + 555, + 510 + ], + "type": "text", + "content": "[53] Jeff Rasley, Samyam Rajbhandari, Olatunj Ruwase, and Yuxiong He. Visual chatgpt: Talking, drawing and editing with visual foundation models. arXiv preprint arXiv:2303.04671, 2023. 2, 3, 5, 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 513, + 555, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 513, + 555, + 567 + ], + "spans": [ + { + "bbox": [ + 316, + 513, + 555, + 567 + ], + "type": "text", + "content": "[54] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 5" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 568, + 555, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 568, + 555, + 612 + ], + "spans": [ + { + "bbox": [ + 316, + 568, + 555, + 612 + ], + "type": "text", + "content": "[55] Javier Romero, Dimitrios Tzionas, and Michael J. Black. Embodied hands: Modeling and capturing hands and bodies together. ACM Transactions on Graphics, (Proc. SIGGRAPH Asia), 36(6), 2017. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 613, + 555, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 613, + 555, + 657 + ], + "spans": [ + { + "bbox": [ + 316, + 613, + 555, + 657 + ], + "type": "text", + "content": "[56] Yu Rong, Takaaki Shiratori, and Hanbyul Joo. Frankmocap: A monocular 3d whole-body pose estimation system via regression and integration. In IEEE International Conference on Computer Vision Workshops, 2021. 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 658, + 555, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 555, + 713 + ], + "type": "text", + "content": "[57] Jingqing Ruan, Yihong Chen, Bin Zhang, Zhiwei Xu, Tianpeng Bao, Hangyu Mao, Ziyue Li, Xingyu Zeng, Rui Zhao, et al. Tptu: Task planning and tool usage of large language model-based ai agents. In NeurIPS 2023 Foundation Models for Decision Making Workshop, 2023. 3" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8159" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 73, + 296, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 296, + 115 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 296, + 115 + ], + "type": "text", + "content": "[58] Yongliang Shen, Kaitao Song, Xu Tan, Dongsheng Li, Weiming Lu, and Yueting Zhuang. Huggingppt: Solving ai tasks with chatgpt and its friends in huggingface. arXiv preprint arXiv:2303.17580, 2023. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 117, + 296, + 170 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 117, + 296, + 170 + ], + "spans": [ + { + "bbox": [ + 56, + 117, + 296, + 170 + ], + "type": "text", + "content": "[59] Yongliang Shen, Kaitao Song, Xu Tan, Wenqi Zhang, Kan Ren, Siyu Yuan, Weiming Lu, Dongsheng Li, and Yueting Zhuang. Taskbench: Benchmarking large language models for task automation. arXiv preprint arXiv:2311.18760, 2023.4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 172, + 296, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 172, + 296, + 205 + ], + "spans": [ + { + "bbox": [ + 56, + 172, + 296, + 205 + ], + "type": "text", + "content": "[60] Soyong Shin, Juyong Kim, Eni Halilaj, and Michael J. Black. Wham: Reconstructing world-grounded humans with accurate 3d motion. In CVPR, 2024. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 206, + 296, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 206, + 296, + 258 + ], + "spans": [ + { + "bbox": [ + 56, + 206, + 296, + 258 + ], + "type": "text", + "content": "[61] Hongjin Su, Weijia Shi, Jungo Kasai, Yizhong Wang, Yushi Hu, Mari Ostendorf, Wen-tau Yih, Noah A Smith, Luke Zettlemoyer, and Tao Yu. One embedder, any task: Instructionfinetuned text embeddings. arXiv preprint arXiv:2212.09741, 2022. 4.5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 261, + 296, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 261, + 296, + 304 + ], + "spans": [ + { + "bbox": [ + 56, + 261, + 296, + 304 + ], + "type": "text", + "content": "[62] Dídac Surís, Sachit Menon, and Carl Vondrick. Vipergpt: Visual inference via python execution for reasoning. Proceedings of IEEE International Conference on Computer Vision (ICCV), 2023. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 305, + 296, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 305, + 296, + 326 + ], + "spans": [ + { + "bbox": [ + 56, + 305, + 296, + 326 + ], + "type": "text", + "content": "[63] John Sweller, Paul Ayres, and Slava Kalyuga. Cognitive Load Theory. Springer, New York, NY, 2011. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 327, + 296, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 327, + 296, + 392 + ], + "spans": [ + { + "bbox": [ + 56, + 327, + 296, + 392 + ], + "type": "text", + "content": "[64] Guy Tevet, Brian Gordon, Amir Hertz, Amit H Bermano, and Daniel Cohen-Or. Motionclip: Exposing human motion generation to clip space. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23–27, 2022, Proceedings, Part XXII, pages 358–374. Springer, 2022. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 394, + 296, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 394, + 296, + 437 + ], + "spans": [ + { + "bbox": [ + 56, + 394, + 296, + 437 + ], + "type": "text", + "content": "[65] Guy Tevet, Sigal Raab, Brian Gordon, Yoni Shafir, Daniel Cohen-or, and Amit Haim Bermano. Human motion diffusion model. In The Eleventh International Conference on Learning Representations, 2023. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 438, + 296, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 438, + 296, + 503 + ], + "spans": [ + { + "bbox": [ + 56, + 438, + 296, + 503 + ], + "type": "text", + "content": "[66] Ayush Tewari, Michael Zollhofer, Hyeongwoo Kim, Pablo Garrido, Florian Bernard, Patrick Perez, and Christian Theobalt. Mofa: Model-based deep convolutional face autoencoder for unsupervised monocular reconstruction. In Proceedings of the IEEE international conference on computer vision workshops, pages 1274-1283, 2017. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 504, + 296, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 504, + 296, + 557 + ], + "spans": [ + { + "bbox": [ + 56, + 504, + 296, + 557 + ], + "type": "text", + "content": "[67] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 559, + 296, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 559, + 296, + 613 + ], + "spans": [ + { + "bbox": [ + 56, + 559, + 296, + 613 + ], + "type": "text", + "content": "[68] Shashank Tripathi, Agniv Chatterjee, Jean-Claude Passy, Hongwei Yi, Dimitrios Tzionas, and Michael J Black. Deco: Dense estimation of 3d human-scene contact in the wild. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8001-8013, 2023. 1, 3, 4, 5, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 614, + 296, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 614, + 296, + 668 + ], + "spans": [ + { + "bbox": [ + 56, + 614, + 296, + 668 + ], + "type": "text", + "content": "[69] Shashank Tripathi, Lea Müller, Chun-Hao P Huang, Omid Taheri, Michael J Black, and Dimitrios Tzionas. 3d human pose estimation via intuitive physics. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4713-4725, 2023. 5, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 670, + 296, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 296, + 712 + ], + "type": "text", + "content": "[70] Timo von Marcard, Roberto Henschel, Michael J. Black, Bodo Rosenhahn, and Gerard Pons-Moll. Recovering accurate 3D human pose in the wild using IMUs and a moving camera. In ECCV, 2018. 5, 6, 8" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 316, + 73, + 555, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 555, + 126 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 555, + 126 + ], + "type": "text", + "content": "[71] Chenyu Wang, Weixin Luo, Qianyu Chen, Haonan Mai, Jindi Guo, Sixun Dong, Xiaohua (Michael) Xuan, Zhengxin Li, Lin Ma, and Shenghua Gao. Mllm-tool: A multimodal large language model for tool agent learning. arXiv preprint arXiv:2401.10727, 2024. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 129, + 555, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 555, + 171 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 555, + 171 + ], + "type": "text", + "content": "[72] Qinzhuo Wu, Wei Liu, Jian Luan, and Bin Wang. Toolplanner: A tool augmented llm for multi granularity instructions with path planning and feedback. arXiv preprint arXiv:2409.14826, 2024.3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 173, + 555, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 173, + 555, + 215 + ], + "spans": [ + { + "bbox": [ + 316, + 173, + 555, + 215 + ], + "type": "text", + "content": "[73] Hongyi Xu, Eduard Gabriel Bazavan, Andrei Zanfir, William T. Freeman, Rahul Sukthankar, and Cristian Sminchisescu. GHUM & GHUML: Generative 3D human shape and articulated pose models. In CVPR, 2020. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 217, + 555, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 217, + 555, + 258 + ], + "spans": [ + { + "bbox": [ + 316, + 217, + 555, + 258 + ], + "type": "text", + "content": "[74] Rui Yang, Lin Song, Yanwei Li, Sijie Zhao, Yixiao Ge, Xiu Li, and Ying Shan. GPT4Tools: Teaching llm to use tools via self-instruction. arXiv preprint arXiv:2305.18752, 2023. 2, 3, 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 261, + 554, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 261, + 554, + 293 + ], + "spans": [ + { + "bbox": [ + 316, + 261, + 554, + 293 + ], + "type": "text", + "content": "[75] Rui Yang, Lin Song, Yanwei Li, Sijie Zhao, Yixiao Ge, Xiu Li, and Ying Shan. Gpt4tools: Teaching large language model to use tools via self-instruction, 2023. 5, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 294, + 555, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 294, + 555, + 347 + ], + "spans": [ + { + "bbox": [ + 316, + 294, + 555, + 347 + ], + "type": "text", + "content": "[76] Zhengyuan Yang, Linjie Li, Jianfeng Wang, Kevin Lin, Ehsan Azarnasab, Faisal Ahmed, Zicheng Liu, Ce Liu, Michael Zeng, and Lijuan Wang. Mm-react: Prompting chatgpt for multimodal reasoning and action. arXiv preprint arXiv:2303.11381, 2023. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 350, + 555, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 350, + 555, + 392 + ], + "spans": [ + { + "bbox": [ + 316, + 350, + 555, + 392 + ], + "type": "text", + "content": "[77] Zhao Yang, Jiaxuan Liu, Yucheng Han, Xin Chen, Zebiao Huang, Bin Fu, and Gang Yu. Appagent: Multimodal agents as smartphone users. arXiv preprint arXiv:2312.13771, 2023. 2, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 394, + 554, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 394, + 554, + 436 + ], + "spans": [ + { + "bbox": [ + 316, + 394, + 554, + 436 + ], + "type": "text", + "content": "[78] Shaokai Ye, Jessy Lauer, Mu Zhou, Alexander Mathis, and Mackenzie Mathis. Amadeusgpt: a natural language interface for interactive animal behavioral analysis. Advances in neural information processing systems, 36:6297-6329, 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 438, + 554, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 438, + 554, + 480 + ], + "spans": [ + { + "bbox": [ + 316, + 438, + 554, + 480 + ], + "type": "text", + "content": "[79] Shaokai Ye, Jessy Lauer, Mu Zhou, Alexander Mathis, and Mackenzie Mathis. Amadeusgpt: a natural language interface for interactive animal behavioral analysis. Advances in neural information processing systems, 36, 2024. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 482, + 555, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 482, + 555, + 525 + ], + "spans": [ + { + "bbox": [ + 316, + 482, + 555, + 525 + ], + "type": "text", + "content": "[80] Siyu Yuan, Kaitao Song, Jiangjie Chen, Xu Tan, Yongliang Shen, Ren Kan, Dongsheng Li, and Deqing Yang. Easytool: Enhancing llm-based agents with concise tool instruction. arXiv preprint arXiv:2401.06201, 2024. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 526, + 555, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 526, + 555, + 590 + ], + "spans": [ + { + "bbox": [ + 316, + 526, + 555, + 590 + ], + "type": "text", + "content": "[81] Hongwen Zhang, Yating Tian, Xinchi Zhou, Wanli Ouyang, Yebin Liu, Limin Wang, and Zhenan Sun. Pymaf: 3d human pose and shape regression with pyramidal mesh alignment feedback loop. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11446-11456, 2021. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 593, + 555, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 593, + 555, + 635 + ], + "spans": [ + { + "bbox": [ + 316, + 593, + 555, + 635 + ], + "type": "text", + "content": "[82] Longwen Zhang, Qiwei Qiu, Hongyang Lin, Qixuan Zhang, Cheng Shi, Wei Yang, Ye Shi, Sibei Yang, Lan Xu, and Jingyi Yu. Dreamface: Progressive generation of animatable 3d faces under text guidance, 2023. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 636, + 555, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 636, + 555, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 636, + 555, + 689 + ], + "type": "text", + "content": "[83] Ruochen Zhao, Hailin Chen, Weishi Wang, Fangkai Jiao, Xuan Long Do, Chengwei Qin, Bosheng Ding, Xiaobao Guo, Minzhi Li, Xingxuan Li, et al. Retrieving multimodal information for augmented generation: A survey. arXiv preprint arXiv:2303.10868, 2023. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 691, + 555, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 555, + 712 + ], + "type": "text", + "content": "[84] Shenhao Zhu, Junming Leo Chen, Zuozhuo Dai, Yinghui Xu, Xun Cao, Yao Yao, Hao Zhu, and Siyu Zhu. Champ:" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "8160" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 76, + 72, + 295, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 295, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 295, + 95 + ], + "type": "text", + "content": "Controllable and consistent human image animation with 3d parametric guidance, 2024. 1, 3, 5" + } + ] + } + ], + "index": 0 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "8161" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/CheXWorld_ Exploring Image World Modeling for Radiograph Representation Learning/f79a692c-48ff-4b64-8cb4-11cdf9303afb_content_list.json b/2025/CheXWorld_ Exploring Image World Modeling for Radiograph Representation Learning/f79a692c-48ff-4b64-8cb4-11cdf9303afb_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..b794a8cdef435783c8298d7bb1e3d80cdacc2bcf --- /dev/null +++ b/2025/CheXWorld_ Exploring Image World Modeling for Radiograph Representation Learning/f79a692c-48ff-4b64-8cb4-11cdf9303afb_content_list.json @@ -0,0 +1,2185 @@ +[ + { + "type": "text", + "text": "CheXWorld: Exploring Image World Modeling for Radiograph Representation Learning", + "text_level": 1, + "bbox": [ + 259, + 130, + 743, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yang Yue $^{1*}$ Yulin Wang $^{1*}$ Chenxin Tao $^{1}$ Pan Liu $^{2}$ Shiji Song $^{1}$ Gao Huang $^{1\\boxtimes}$", + "bbox": [ + 153, + 210, + 844, + 229 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Tsinghua University $^{2}$ PLA General Hospital", + "bbox": [ + 354, + 231, + 645, + 247 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "yueyang22@mails.tsinghua.edu.cn, gaohuang@tsinghua.edu.cn", + "bbox": [ + 241, + 247, + 750, + 262 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 296, + 326, + 313 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Humans can develop internal world models that encode common sense knowledge, telling them how the world works and predicting the consequences of their actions. This concept has emerged as a promising direction for establishing general-purpose machine-learning models in recent preliminary works, e.g., for visual representation learning. In this paper, we present CheXWorld, the first effort towards a self-supervised world model for radiographic images. Specifically, our work develops a unified framework that simultaneously models three aspects of medical knowledge essential for qualified radiologists, including 1) local anatomical structures describing the fine-grained characteristics of local tissues (e.g., architectures, shapes, and textures); 2) global anatomical layouts describing the global organization of the human body (e.g., layouts of organs and skeletons); and 3) domain variations that encourage CheXWorld to model the transitions across different appearance domains of radiographs (e.g., varying clarity, contrast, and exposure caused by collecting radiographs from different hospitals, devices, or patients). Empirically, we design tailored qualitative and quantitative analyses, revealing that CheX-World successfully captures these three dimensions of medical knowledge. Furthermore, transfer learning experiments across eight medical image classification and segmentation benchmarks showcase that CheXWorld significantly outperforms existing SSL methods and large-scale medical foundation models. Code & pre-trained models are available at https://github.com/LeapLabTHU/CheXWorld.", + "bbox": [ + 89, + 329, + 485, + 753 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 784, + 223, + 800 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Intelligent agents like humans learn extensive background knowledge about the world [39]. This common-sense information is embedded in the agents' internal models of the world, playing a pivotal role in their perception, learning,", + "bbox": [ + 89, + 810, + 483, + 872 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "and decision-making processes by simulating the world's dynamics, telling them what is plausible or impossible, and predicting the outcomes of their actions. Consequently, these world models enable agents to acquire new concepts and skills with minimal demonstrations and trials [55]. Recent works [2, 21] have preliminarily verified the effectiveness of establishing visual representation learning approaches through the lens of world modeling. Pre-trained world models can produce semantically rich embeddings and adapt to various downstream tasks with limited data.", + "bbox": [ + 511, + 297, + 906, + 450 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Like many other research areas, the field of medical imaging is experiencing a paradigm shift from task-specific models to general-purpose foundation models [20, 34, 45, 68], which are pre-trained on massive data and expected to encode meaningful medical knowledge. The idea of world modeling offers an approach to training a medical foundation model by capturing common sense information (e.g. human anatomy) from medical images, which is a promising yet under-explored research direction.", + "bbox": [ + 511, + 455, + 908, + 592 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we present CheXWorld, the first initiative toward world modeling for self-supervised representation learning on radiographic images. We propose and integrate three world modeling tasks that capture different dimensions of medical knowledge essential for qualified radiologists, depicted in Figure 1. In particular, a radiologist with a strong understanding of human anatomy can identify different anatomical structures and determine the relative geometry between anatomical regions. Given the hierarchical nature of human anatomy, we introduce world modeling tasks that focus on both local and global levels of anatomy: local anatomical structure modeling aims to predict specific anatomical characteristics and structures within a localized anatomical region, such as bones, airways, blood vessels, and lung segments; global anatomical layout modeling learns the overall arrangement and spatial relationships of various anatomical regions within the human body, understanding the relative positioning of organs and tissues such as the heart, lungs, diaphragm, and rib cage. Additionally, medical images are", + "bbox": [ + 511, + 598, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution.", + "bbox": [ + 107, + 885, + 220, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "Corresponding authors.", + "bbox": [ + 263, + 885, + 400, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "20778", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/41433b7f88a116e6c0b7e4bf4c94a8f533c04eb500b67bb7b701d3dc200ba63d.jpg", + "image_caption": [ + "(a) Local Anatomical Structure Modeling" + ], + "image_footnote": [], + "bbox": [ + 117, + 95, + 369, + 311 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/b6145e1e58a76ea4ed64cf4907960ac6830c978549b234b7ce10b94301f58e91.jpg", + "image_caption": [ + "(b) Global Anatomical Layout Modeling" + ], + "image_footnote": [], + "bbox": [ + 375, + 95, + 629, + 309 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/c40bc4c47d56a9ab85b6869a589f38b621864ac17511376b643e827846072086.jpg", + "image_caption": [ + "(c) Domain Variation Modeling" + ], + "image_footnote": [], + "bbox": [ + 635, + 90, + 880, + 314 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/2fa63e404893b79cd981aaf0a11bec89b40fd099fe7432fd45bc0da5f5bf06e6.jpg", + "image_caption": [ + "(d) Unified World Modeling", + "Figure 1. Overview of the CheXWorld framework1. The upper part of the figure depicts three dimensions of medical knowledge that are formulated in our framework, including (a) local anatomical structures describing the fine-grained characteristics of local tissues, (b) global anatomical layouts describing the global organization of the human body and (c) domain variations that encourage CheXWorld to model the transitions across different appearance domains of radiographs. The middle part of the figure illustrates the world modeling tasks corresponding to these aspects of medical knowledge. (d) shows our unified pipeline that combines the merits of all three tasks." + ], + "image_footnote": [], + "bbox": [ + 114, + 340, + 866, + 455 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "typically gathered from multiple domains with notable appearance differences due to varying equipment and acquisition techniques [52]. Experienced radiologists can recognize the physical reality behind domains with different appearances and facilitate the diagnosis process. Motivated by this, we hypothesize that radiologists quickly adapt to diverse domain changes by learning sophisticated internal world models that are capable of simulating transitions between domains. Driven by our hypothesis, we introduce the domain variation modeling task, which aims to learn an expressive feature space that encodes predictable transitions between domains. Finally, we design an integrated pipeline that performs the three world modeling tasks simultaneously, effectively incorporating medical anatomical knowledge and achieving robustness across domains.", + "bbox": [ + 88, + 546, + 482, + 772 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "It is noteworthy that the three aspects we consider, i.e., local anatomical structure modeling, global anatomical layout modeling, and domain variation modeling, represent core elements of radiographic expertise, each contributing to a comprehensive framework for capturing medical knowledge. While they may not encompass every dimen", + "bbox": [ + 88, + 773, + 482, + 864 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "sion of medical understanding, they are fundamental to radiology and critical for radiograph representation learning, as extensively demonstrated in [23, 49, 69, 70].", + "bbox": [ + 511, + 546, + 906, + 590 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Empirically, we validate that CheXWorld effectively captures the three dimensions of medical knowledge through a series of analytical experiments, including visualizations of predictor outputs with the help of a generative model and a domain variation sensitivity test. Moreover, we extensively evaluate CheXWorld on eight medical image classification and segmentation benchmarks. Our model consistently outperforms competitive self-supervised learning baselines with comparable backbone capacity, pretraining data, and pre-training computational cost.", + "bbox": [ + 509, + 592, + 908, + 743 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 511, + 753, + 656, + 768 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "World Modeling. In the machine learning community, world modeling originates from model-predictive control [9, 10] and is a common practice in reinforcement learning, where it predicts the future state of the environment based on the agent's action [24-27]. Recently, building general world models featuring profound comprehension of common sense knowledge is believed to be a crucial step towards general artificial intelligence [39, 53]. Video gener", + "bbox": [ + 511, + 779, + 908, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "1The x-rays are taken from kenhub.com, Radiopedia [43] and ChestXray14 [61].", + "bbox": [ + 89, + 875, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "20779", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ation models [37] trained on vast data (e.g. SoRA[47] and VideoPoet [38]) are world models that perform predictions in input space. Another line of work makes predictions in the latent space [2, 7, 28, 33, 39], which more closely resembles the world model in the human brain. World modeling is also an effective approach for visual representation learning, where the model predicts missing parts of the visual inputs [2, 4, 5, 7, 21]. Our framework builds on this foundation, where our model acquires knowledge—such as human anatomy and domain variations—through learning to predict unobserved outcomes from specific actions. Our key contribution lies in providing insights into modeling medical knowledge from radiographic images and designing a framework that unifies three world modeling tasks.", + "bbox": [ + 89, + 90, + 480, + 301 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Self-supervised learning in medical imaging. Self-supervised learning (SSL) [19, 31, 48, 62] provides a practical solution to mitigating annotation scarcity in medical imaging. A large body of works incorporate restorative SSL tasks, which typically involve reconstructing the original data from certain image corruptions, such as random masking [63, 67, 68], pixel shuffling [69], and patch order shuffling [50, 58, 70]. Another line of works [3, 12, 51, 56, 66] adapt global or dense contrastive learning to medical imaging. TransVW[29] and DiRA [30, 57] explore combining discriminative and restorative methods. Additionally, learning from anatomy is a common topic for medical SSL. SAM [65] enforces pixel-level anatomical correspondence between images, and Adam [32] learns part-whole correspondence that reflects the hierarchy of human anatomy. Recently, SSL-based foundation models trained on largescale datasets gained significant attention, such as RET-Found [68] for retinal imaging and UNI [13] for pathology.", + "bbox": [ + 89, + 301, + 482, + 574 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Compared with existing works, CheXWorld is the first to introduce the concept of world modeling into medical SSL, capturing task-relevant medical knowledge through three tailored world modeling tasks. Unlike contrastive-based approaches such as Adam [32], CheXWorld takes a technically distinct path by constructing equivariant [16, 17] representations—where input transformations result in predictable changes within the embedding space—rather than invariant representations. This approach allows CheXWorld to better model medical data without discarding valuable information. Moreover, our unified SSL framework delivers state-of-the-art performance across eight benchmarks, demonstrating the significant potential of designing SSL methods grounded in the philosophy of world modeling.", + "bbox": [ + 89, + 575, + 482, + 785 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Basic Framework of World Modeling", + "text_level": 1, + "bbox": [ + 89, + 799, + 426, + 816 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "World modeling. We begin by briefly introducing the concept of world modeling [25, 39], which forms the basis for our proposed method. The primary motivation behind world modeling is to predict the unobserved parts $y$ of a world (e.g., a visual environment, an image, or a video)", + "bbox": [ + 89, + 824, + 482, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "based on an observed context $x$ . This prediction can be formulated across various dimensions, such as spatial (predicting the unseen regions of the data) and temporal (foreseeing the consequence of an action). In fact, the paradigm of world modeling takes inspiration from human visual cognition. For example, when presented with the right upper lobe region of a chest radiograph as the context $x$ , the internal world of a radiologist could imagine that there are two lobes (right middle and right lower) below the visible region and two lobes on the left, forming the target $y$ .", + "bbox": [ + 511, + 90, + 903, + 242 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "A basic framework.", + "text_level": 1, + "bbox": [ + 535, + 248, + 674, + 261 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Here, we introduce our basic framework of instantiating world modeling, upon which we will further discuss how to establish a radiology world model by formulating various medical", + "bbox": [ + 511, + 262, + 676, + 382 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/ed454c2075b415059dbcd9c2a4cac7a970b903ad8eb4c96ac6716d65250a6627.jpg", + "image_caption": [ + "Figure 2. A basic framework of world modeling." + ], + "image_footnote": [], + "bbox": [ + 684, + 256, + 898, + 347 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "knowledge in Section 4. In specific, our basic framework mainly follows the joint-embedding predictive architecture (JEPA) [2, 39], as shown in Figure 2. We consider two encoders: the context encoder $f_{\\theta}$ and the target encoder $f_{\\theta'}'$ , which encode the context $x$ and target $y$ into representations $h_x$ and $h_y$ , respectively. A predictor $g_{\\phi}$ learns to predict the target embedding $h_y$ using the context embedding $h_x$ , conditioned on an additional latent variable $z$ that indicates how the target relates to the context. For example, $z$ can be the spatial locations of $y$ with respect to $x$ , or an action that leads to the transition from context to target $x \\rightarrow y$ (e.g., image transformations). The latent variable $z$ carries the information that makes $y$ predictable based on $x$ , effectively modeling the inherent variation and uncertainty of the real world. The primary objective of world modeling is:", + "bbox": [ + 511, + 383, + 903, + 609 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\text {m i n i m i z e} \\quad D \\left(h _ {y}, \\hat {h} _ {y}\\right) = D \\left(h _ {y}, g _ {\\phi} \\left(h _ {x}; z\\right)\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 562, + 614, + 903, + 633 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $D(\\cdot, \\cdot)$ denotes the prediction error. Notably, the predictions are performed in the abstract representation space, contrasting with generative world models that predict every detail of the target. This mirrors an important characteristic of intelligent agents: information filtering, i.e., eliminating the irrelevant details in the data during the perception process. For example, a radiologist can infer the anatomical layout of a partly missed image, but it's nearly impossible for him to recover each pixel value of the unseen regions.", + "bbox": [ + 511, + 638, + 905, + 773 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In implementation, we set the target encoder to be the exponential moving average of the context encoder following [2, 4, 7]. Moreover, we use the Vision Transformer (ViT) [18] architecture for the context/target encoder and the predictor. For the rest of this paper, the context and target feature $h_x, h_y$ are sequences of patch embeddings produced by the ViTs. To perform patch-level feature predictions, we attach mask tokens to the context features and feed the com", + "bbox": [ + 511, + 780, + 905, + 898 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "20780", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "binned sequence to the predictor. The output representations of the mask tokens are used as the final prediction.", + "bbox": [ + 89, + 90, + 482, + 122 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. CheXWorld: World Modeling for Radiograph Representation Learning", + "text_level": 1, + "bbox": [ + 89, + 132, + 482, + 170 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This section extends the idea of world modeling to radiology. We propose three world modeling tasks tailored for radiographs and establish a unified CheXWorld framework that seamlessly integrates them. Importantly, the three tasks are designed to model three critical dimensions of medical knowledge essential for qualified radiologists. To be specific, these tasks are organized hierarchically from low-level to high-level, namely 1) local anatomical structure modeling that aims to learn the fine-grained structures of local anatomical regions, 2) global anatomical layout modeling that seeks to learn the global geometry of the human body (e.g., the layout of organs and skeletons), and 3) domain variation modeling that learns to model the transition across different appearance domains of radiographs. We first elaborate on the details of each task. Then, we present a unified framework that comprehensively combines the characteristics and merits of all three tasks, yielding a powerful foundation model that produces semantically rich and transferable representations.", + "bbox": [ + 89, + 176, + 483, + 464 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Local Anatomical Structure Modeling", + "text_level": 1, + "bbox": [ + 89, + 470, + 419, + 488 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "At a relatively low, fundamental level, various types of tissues (e.g., bones, muscles, and epithelial tissues) form the structural organization of the human body. Understanding the fine-grained characteristics (e.g., shapes, sizes, appearance, and textures) of these micro-structures is essential knowledge that radiologists must possess. To build up a learning procedure that enables neural networks to acquire such local anatomical knowledge, we consider a mask-and-reconstruction task. The model predicts fine-grained details of a masked tissue region based on its peripheral, surrounding information, as shown in Figure 1(a). This encourages the networks to develop an internal model that captures the intricate structures and local continuities of human microanatomy. The large, continuous prediction target also prevents reliance on low-level features.", + "bbox": [ + 89, + 493, + 483, + 718 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Specifically, the image mask is a union of four randomly selected rectangular regions of the image following [2]. Let $M$ denote the set of 2D patch locations that are masked from the context input. Given an input image (crop) $I$ , the image patches with locations contained in $M$ are dropped from the image to create the context $x = \\operatorname{Mask}(I, M)$ , while the target is the entire image. The context and target are then fed to the corresponding encoders, producing representations $h_x = f_\\theta(x)$ and $h_y = f_{\\theta'}'(y)$ . Note that the context encoder exclusively processes visible patches, following [2, 31]. In the predictor, the mask tokens $m$ carrying the positional encoding of the masked locations are concatenated with the", + "bbox": [ + 89, + 719, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "context feature $h_x$ , which forms the input token sequence to the predictor. The predictor output $\\dot{h}_y$ is given by:", + "bbox": [ + 511, + 90, + 906, + 122 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\hat {h} _ {y} = g _ {\\phi} (h _ {x}; M) \\\\ = g _ {\\phi} \\left(\\left(h _ {x} + p _ {x}\\right) \\oplus \\{m + \\operatorname {P E} (u, v) \\} _ {(u, v) \\in M}\\right), \\tag {2} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 527, + 127, + 903, + 174 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\oplus$ is the concatenation operation along sequence dimension, $p_x$ is the positional embedding corresponding to the context, $u$ and $v$ are image patch location indices along height and width dimensions, $\\mathrm{PE}(\\cdot)$ is the sinusoidal positional encoding proposed by [59]. The prediction loss is computed only on the masked locations, which is given by:", + "bbox": [ + 511, + 178, + 906, + 268 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {l o c a l}} (x, y) = \\sum_ {c \\in M} \\| g _ {\\phi} \\left(f _ {\\theta} (x); M\\right) _ {c} - f _ {\\theta^ {\\prime}} ^ {\\prime} (y) _ {c} \\| _ {2} ^ {2}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 534, + 285, + 906, + 319 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $c$ is the 2D image patch location that belongs to $M$ .", + "bbox": [ + 511, + 321, + 898, + 338 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Global Anatomical Layout Modeling", + "text_level": 1, + "bbox": [ + 511, + 354, + 831, + 372 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In addition to understanding local microstructures like tissues, it is also important for radiologists to possess comprehensive medical knowledge regarding how groups of tissues or systems of organs are anatomically organized and assembled into the full bodies of humans. Inspired by this, we develop the global anatomical layout modeling task (see Figure 1(b)), seeking to predict the feature of an out-of-context area based on its relative position to a given context, thus facilitating understanding of the global structure of human bodies. Notably, this idea is orthogonal to the technique proposed in section 4.1 (see section 5.3 for ablation studies) as it focuses on formulating the long-range topological relationships of different tissues or organs instead of the fine-grained anatomical structures within local tissues.", + "bbox": [ + 511, + 380, + 906, + 592 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As illustrated in Figure 1(b), we randomly crop two anatomical areas from the original radiograph, serving as context $x$ and target $y$ , respectively. The world model learns to predict $y$ from $x$ given their relative position information $\\Delta_{x\\rightarrow y}$ . In particular, we assume the context and target image", + "bbox": [ + 511, + 594, + 678, + 792 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/e1a55e97415c78eaeeef8cf47ae44dc6091f8137a742da23b4598b120c451331.jpg", + "image_caption": [ + "Figure 3. Formulation of global anatomical layout modeling." + ], + "image_footnote": [], + "bbox": [ + 684, + 601, + 898, + 750 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "crop share the same spatial size of $h$ pixels in height and $w$ pixels in width, with the left-top corner located at $(i_x, j_x)$ and $(i_y, j_y)$ in the pixel space. The relative position information is defined as the proportional relative displacement between the two regions:", + "bbox": [ + 511, + 790, + 906, + 867 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\Delta_ {x \\rightarrow y} = (\\Delta h, \\Delta w) = \\left(\\frac {i _ {y} - i _ {x}}{h}, \\frac {j _ {y} - j _ {x}}{w}\\right). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 557, + 869, + 906, + 905 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "20781", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To properly represent the location of the image patches in the target $y$ , we establish a coordinate system over the context image patches, with the top-left corner patch located at the origin $(0,0)$ and the bottom-right corner patch located at $(N_h - 1, N_w - 1)$ , where $N_h, N_w$ are the number of tokens along the height and width dimensions. As demonstrated in Figure 3, the coordinates of the image patch at the $u$ -th row and $v$ -th column in the target image are given by:", + "bbox": [ + 89, + 90, + 483, + 212 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\phi_ {x \\rightarrow y} (u, v) = (\\Delta h \\cdot N _ {h} + u, \\Delta w \\cdot N _ {w} + v) \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 91, + 217, + 483, + 239 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $u\\in [0,N_h - 1],v\\in [0,N_w - 1]$", + "bbox": [ + 169, + 237, + 450, + 252 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Essentially, $\\phi_{x\\to y}(u,v)$ describes the target's location from the context's perspective, enabling the model to predict the target based on the context. Sinusoidal position encoding $\\mathrm{PE}(\\cdot)$ is then applied to the relative coordinates $\\phi_{x\\rightarrow y}(u,v)$ , and the mask tokens carrying the positional embeddings of all the target image patches are fed to the predictor. The predictor output is given by:", + "bbox": [ + 89, + 258, + 483, + 364 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\hat {h} _ {y} = g _ {\\phi} (h _ {x}; \\Delta_ {x \\to y}) \\\\ = g _ {\\phi} \\left(\\left(h _ {x} + p _ {x}\\right) \\oplus \\left\\{m + \\operatorname {P E} \\left(\\phi_ {x \\rightarrow y} (u, v)\\right)\\right\\} _ {u, v}\\right), \\tag {6} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 94, + 368, + 480, + 415 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $u\\in [0,N_h - 1],v\\in [0,N_w - 1]$", + "bbox": [ + 112, + 417, + 393, + 434 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "and the loss function is given by:", + "bbox": [ + 89, + 439, + 312, + 454 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {g l o b a l}} (x, y) = \\left\\| g _ {\\phi} \\left(f _ {\\theta} (x); \\Delta_ {x \\rightarrow y}\\right) - f _ {\\theta^ {\\prime}} ^ {\\prime} (y) \\right\\| _ {2} ^ {2}. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 117, + 459, + 480, + 479 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Domain Variation Modeling", + "text_level": 1, + "bbox": [ + 89, + 486, + 341, + 502 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Thus far, we have discussed two levels of anatomical knowledge of human bodies. Beyond this, radiologists are usually capable of flexibly adapting to the specific characteristics of each radiograph and understanding them despite variations [15, 23, 49]. As shown in Figure 1(c), medical images often come from different sources (e.g., hospitals, techniques, equipment) and show diverse appearances [52] (e.g., clarity, contrast, exposure). Nevertheless, experienced radiologists can imagine the physical reality behind the radiographs with varying appearances, which facilitates the diagnosis process. We hypothesize that radiologists' internal world models can simulate the variations across domains, thus obtaining an objective view of the scanned body.", + "bbox": [ + 89, + 508, + 483, + 704 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Inspired by this, we propose a domain variation modeling task that learns how image features change across domains, enabling cross-domain adaptability. We simulate domain shifts using data augmentation and construct context-target pairs to model these transitions. The model is tasked with predicting the feature outcomes of the inverse effect of an image transformation determined by certain augmentation parameters, as shown in Figure 1(c). The learned representation space is designed to be equivariant [16, 17], i.e., input transformation leads to predictable output change, which preserves sufficient information across domains, mimicking radiologists' internal models.", + "bbox": [ + 89, + 704, + 483, + 885 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Specifically, the target $y$ is an augmented version of", + "bbox": [ + 109, + 885, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "the original image. Another transform parameterized by $k$ scalars (denoted by $a \\in \\mathbb{R}^k$ ) is further applied to the target to produce the context $x = \\mathcal{T}_a(y)$ , where $\\mathcal{T}$ is an augmentation pipeline consisting of Gaussian blur, brightness, contrast, and gamma adjustments. The parameter $a$ contains the strength and other configurations of the augmentation. The world model learns to model the feature transformation from the context to the target domain (i.e., the inverse effect of $\\mathcal{T}_a$ ) conditioned on the parameter $a$ , which can be regarded as the \"action\" of domain transition. The prediction is also performed with mask tokens to be compatible with the other two world modeling tasks. The parameter $a$ is combined with the mask tokens via a lightweight policy network $\\pi$ . The forward pass and loss function of domain variation modeling are given by:", + "bbox": [ + 511, + 90, + 906, + 318 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} m _ {a, p} = \\pi (m + p, a), \\\\ \\hat {h} _ {y} = g _ {\\phi} \\left(h _ {x}; a\\right) \\\\ = g _ {\\phi} \\left(\\left(h _ {x} + p _ {x}\\right) \\oplus \\left\\{m _ {a, \\mathrm {P E} (u, v)} \\right\\} _ {u, v}\\right), \\tag {8} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 573, + 323, + 903, + 388 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {d o m a i n}} (x, y) = \\left\\| g _ {\\phi} \\left(f _ {\\theta} (x); a\\right) - f _ {\\theta^ {\\prime}} ^ {\\prime} (y) \\right\\| _ {2} ^ {2},\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 391, + 807, + 412 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $p$ is the positional embedding of an image patch, $m_{a,p}$ denotes the mask token carrying both spatial and domain transition information.", + "bbox": [ + 511, + 415, + 905, + 460 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.4. The Unified Framework", + "text_level": 1, + "bbox": [ + 513, + 489, + 733, + 503 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Having established the individual tasks of local anatomical structure modeling, global anatomical layout modeling, and domain variation modeling, we now integrate these tasks into a cohesive system. This integration is important for creating a comprehensive model that can leverage multiple dimensions of medical knowledge simultaneously and maximize the combined benefits of each task.", + "bbox": [ + 511, + 518, + 905, + 625 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Instead of simply adding three loss functions together, a more elegant solution is to design unified contexts, targets, and latent variables to perform the three world modeling tasks in a single forward pass. As illustrated in Figure 1(d), the training pipeline of CheXWorld starts with sampling two different image crops from the image $y_{1},y_{2}$ which serve as the targets. The targets are then augmented and masked with two different configurations $(a_{1},M_{1})$ and $(a_{2},M_{2})$ , producing the contexts $x_{1},x_{2}$ :", + "bbox": [ + 511, + 631, + 906, + 768 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nx _ {i} = \\operatorname {M a s k} \\left(\\mathcal {T} _ {a _ {i}} \\left(y _ {i}\\right), M _ {i}\\right), i = 1, 2. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 591, + 773, + 903, + 789 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The contexts $x_{1}, x_{2}$ and targets $y_{1}, y_{2}$ are then encoded into representations $h_{x_1}, h_{x_2}, h_{y_1}, h_{y_2}$ . Note that the targets $h_{y_1}, h_{y_2}$ are predictable from the contexts $h_{x_1}, h_{x_2}$ given the proper conditions. Specifically, since $x_{1}$ is obtained by applying augmentation $\\mathcal{T}_{a_1}$ and mask $M_{1}$ to $y_{1}$ , predicting $y_{1}$ from $x_{1}$ with conditions $(a_{1}, M_{1})$ simultaneously performs local anatomical structure modeling and domain vari", + "bbox": [ + 511, + 795, + 905, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "20782", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/326a376dec9e0bf39f24f7676a0b89436ef8bbddd6c495f2e7aa7c3eb3b20aaf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 178, + 93, + 232, + 143 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/3c393711d543ed0ddef7ca91d62b1b34bc1610251ce66e06778252f2efbd5519.jpg", + "image_caption": [ + "Context" + ], + "image_footnote": [], + "bbox": [ + 236, + 103, + 292, + 143 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e05d0f362b2847b042066c536b5391fcd1c9256bbd54baae17c3ed9a43cd38d2.jpg", + "image_caption": [ + "", + "World Model Prediction" + ], + "image_footnote": [], + "bbox": [ + 295, + 103, + 509, + 143 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4a6a87624a25db5f3084e7d5c3daa7c5a4d444d6634baf1452eb4dc387d0f251.jpg", + "image_caption": [ + "Target v.s. Prediction" + ], + "image_footnote": [], + "bbox": [ + 514, + 103, + 617, + 143 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1578f9d8ae5ea7e7ea8f89cefe781aef84d40ef352241f8ca925dccc0d83964d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 178, + 147, + 233, + 189 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/8b7daefee0531d89e9f371355a2141d01f1d9b99df9a6a8c0878974a68567031.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 236, + 147, + 292, + 189 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1fe7a232935c3719ac2603fb66cc49393bb50ff7bd23b1ce1339c09aaaf94123.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 295, + 147, + 349, + 189 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f7fdbea78ae50e67feec66900a240ecbca2281bfc61710f623a75b0d32120159.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 351, + 147, + 406, + 189 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/405a4f111d7f3da1fb17199a51746be3e0c2c06566e2c79c3981df678f649037.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 408, + 147, + 511, + 189 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1acfa791be86c8840b8164fcc64c9abc5b3d7523715900cc91ea4a9602201e2b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 514, + 147, + 565, + 189 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/cb54c4b6fff1665566c101437ccf6d012fc8d227e674c0adc8012c0ffe54a839.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 566, + 147, + 617, + 189 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4e9914b5b09de9100aeaebdf3f6f44d2d312e95efd9e31917159462bc45a8a9d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 178, + 193, + 232, + 234 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4f9d63a1bd528c4e6f4fe5319f73e5242562c69ea8a3a7c9d0848b2b3fe4b99c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 236, + 193, + 292, + 234 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5b68524b0361cd794535970be23d278a2f94a5d70ff7e59bb3b41a71be1449ce.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 295, + 193, + 349, + 234 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/316738ab436ce6819a06be4f6a4cc231c005acead4ef759e1e5598e3d51e1d9b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 351, + 193, + 406, + 234 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b74369d4637276f88f0aff867a281e3159aecd75bb4f5e445d87404f2703e9b8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 408, + 193, + 511, + 234 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/fac77acb0ce143ec46c2d7b6754d4d9d33b47a6f3568d33eb76902867fd6d7e8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 514, + 193, + 565, + 234 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/17de1a0bddc331484e9b41551160d397d63b2b5cb677ddb04dd1db832dfd2e8c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 566, + 193, + 617, + 234 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b2e0f4c24cf93edbbb8c995c78b0b4cf8fbea4b3d50dfcb2c7bf6af8502eec0c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 178, + 239, + 233, + 280 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/89217775023f4aabd9b313a4c9b7187fb54dd398bd8e7e158237f440c9c930be.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 236, + 239, + 292, + 280 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d6ec51fe22edc9e09e50bac10bf179e8dfa88cd0b201882bf50a5bb947ff1adc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 295, + 239, + 349, + 280 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/00de25365320a54a54e319665857f8a1ef07c8e7cc7855144c581c004420fba1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 351, + 239, + 406, + 280 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/132de9a0a72ba287f220e3e8bc707bfd6b349e1a8b95fd4acb32ee58eaf5c0a7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 408, + 239, + 511, + 280 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/8e40d3f2edd650a2a8054a19486f165ab07875cedd1dbd78afb881e8b7dadc51.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 514, + 239, + 565, + 280 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/6078cf6bd9568bf678dbdbfaa2209fb319f4dd52ad3a2ba018db042c099a8514.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 566, + 239, + 617, + 280 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/127370d9c02fe8a49b741b948cedab78c0e79676abc581127f75bf1ff0afb6f7.jpg", + "image_caption": [ + "Context&Target World Model Prediction" + ], + "image_footnote": [], + "bbox": [ + 640, + 103, + 815, + 143 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ba3fda43c6bfbf81898b88f642c048df3780af2c95d93120ecb7688dc3bbb931.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 640, + 147, + 756, + 189 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5b308fb86f8b761fae3d16bc3e92382890f3e0cb2b2c68c0d9c358aa9b1ca72a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 759, + 146, + 813, + 189 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a0dc1f2400ca8df47b7b74e261892668f7ca35427301e2ab51609fe1f8f9a16d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 640, + 193, + 694, + 233 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/2af7c6f860936e742feaa1ffcdba3f2b981fe9c6447fe8ec105f109f7c6f419e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 699, + 193, + 751, + 233 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c330d2f8b5b992503bc406ad510d8af4f56ee01781e1097742c8f2813a118f8a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 759, + 193, + 813, + 233 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/0d908be3ef7d7d9f628243f83e199333b3bcb4ff8c35103d86a12bc0f67e350c.jpg", + "image_caption": [ + "(b) Global Layout Modeling" + ], + "image_footnote": [], + "bbox": [ + 640, + 238, + 694, + 281 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/0059af19b42a4ffcbd012761f7583551f6e5aa1436a8ef33867a4fb9e6d4b377.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 699, + 238, + 756, + 281 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b2fdf30bec8b2577924aaf4f5faec368317a8ccd61d3e1258c606e3fcb32c8a1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 759, + 238, + 813, + 281 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4470cd4eef2e787a3003ce98a6b5b65d6cb55f865e7b25a5be4945edc7efd31a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 140, + 353, + 217, + 416 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e7428427db91350638451feb75d41c54ea8cc456647fd392ddbe29f4dff0623b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 233, + 354, + 305, + 416 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/674bdda378e763f892e1811dc2bf9b89245fc03c983577af3a8c2747c7d3f34e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 323, + 354, + 398, + 416 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/9e6aea537c52cc755f421ea085c45ff7fe1acb68129f0c219ecacc815c20bfc0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 142, + 419, + 217, + 448 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e624d15008386098ae6d1b94d2d91744218561430d95970cb57e11f3d2c449e4.jpg", + "image_caption": [ + "Figure 5. CheXWorld prioritizes relevant medical features over spurious signals (e.g., lateral markers) in the image. Regions in red boxes denote the predictor outputs." + ], + "image_footnote": [], + "bbox": [ + 233, + 420, + 307, + 449 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e0dc553cf827785bfe5397648b2d9b8438c935de2f226aa0219ee5ef6488fb93.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 323, + 420, + 397, + 449 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f1e9db7cb9dcd0bec2cec68238f447c7b69345564a860bca83b60abee864ab82.jpg", + "image_caption": [ + "Figure 4. Visualization of the CheXWorld predictor outputs (zooming in for details). The images presented in this figure were not included in the pre-training of CheXWorld or the training of the diffusion model. Regions in red bounding boxes denote the predictor outputs that are mapped to pixel space using the RCDM [8] framework. In (a), gray areas indicate masked regions excluded from the context. In (b), the two overlapping regions alternately serve as context and target.", + "Figure 6. CheXWorld learns anatomical correspondence. We compute pixel-level embeddings using RoI-pooling [22] and calculate the embedding similarity between four anatomical landmarks in a reference image and each pixel in the test image to create feature similarity heatmaps." + ], + "image_footnote": [], + "bbox": [ + 491, + 352, + 867, + 441 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ation modeling:", + "bbox": [ + 89, + 512, + 199, + 527 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {1 \\rightarrow 1} = \\mathcal {L} _ {\\text {l o c a l} + \\text {d o m a i n}} \\left(x _ {1}, y _ {1}\\right) \\\\ = \\sum_ {c \\in M _ {1}} \\| g _ {\\phi} \\left(h _ {x _ {1}}; M _ {1}, a _ {1}\\right) _ {c} - f _ {\\theta^ {\\prime}} ^ {\\prime} \\left(y _ {1}\\right) _ {c} \\| _ {2} ^ {2}. \\tag {10} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 112, + 532, + 480, + 584 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Moreover, $x_{1}$ can be derived from $y_{2}$ by moving the input window by $-\\Delta_{x_1\\to y_2}$ and applying augmentation $\\mathcal{T}_{a_1}$ , hence the predictor conditioning on $(\\Delta_{x_1\\to y_2},a_1)$ performs global anatomical structure modeling and domain variation modeling at the same time.", + "bbox": [ + 89, + 589, + 483, + 664 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {1 \\rightarrow 2} = \\mathcal {L} _ {\\text {g l o b a l} + \\text {d o m a i n}} \\left(x _ {1}, y _ {2}\\right) \\tag {11} \\\\ = \\left\\| g _ {\\phi} \\left(h _ {x _ {1}}; \\Delta_ {x _ {1} \\rightarrow y _ {2}}, a _ {1}\\right) - f _ {\\theta^ {\\prime}} ^ {\\prime} (y _ {2}) \\right\\| _ {2} ^ {2}. \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 122, + 670, + 480, + 709 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Similarly, $\\mathcal{L}_{2\\rightarrow 2},\\mathcal{L}_{2\\rightarrow 1}$ can be computed between $(x_{2},y_{2})$ and $(x_{2},y_{1})$ , respectively. Note that the context features and target features are reused twice when computing the loss (10) and (11). Finally, the model is trained with the combination of $2\\times 2 = 4$ supervisions computed from the two pairs of context and target.", + "bbox": [ + 89, + 714, + 483, + 805 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {1 \\rightarrow 1} + \\mathcal {L} _ {2 \\rightarrow 2} + \\mathcal {L} _ {1 \\rightarrow 2} + \\mathcal {L} _ {2 \\rightarrow 1}. \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 158, + 811, + 480, + 825 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "With this unified objective, CheXWorld showcases deep comprehension of human anatomy in different scales and pixel intensity distributions across various domains.", + "bbox": [ + 89, + 832, + 483, + 878 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 513, + 510, + 643, + 527 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we first demonstrate CheXWorld's world modeling capability with a series of analytical experiments. Then, we empirically compare CheXWorld with other self-supervised methods as baselines on eight downstream tasks. Finally, we provide ablation studies to highlight the effect of each building block of our method.", + "bbox": [ + 511, + 536, + 906, + 626 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation details. CheXWorld is pre-trained on $\\sim 0.5\\mathrm{M}$ frontal chest X-rays from several public datasets. The visual backbone is a ViT-Base [18]. The model is trained for 300 epochs, taking 16 hours on 8 RTX 4090 GPUs. Please refer to the appendix for more details.", + "bbox": [ + 511, + 627, + 906, + 702 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. CheXWorld is a Strong World Predictor", + "text_level": 1, + "bbox": [ + 511, + 710, + 857, + 728 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Visualization of anatomical modeling. To better understand the predictions made by CheXWorld, we train a generative model that maps the average-pooled predictor output back to pixel space following the RCDM framework [8]. Specifically, we construct context-target pairs in the same way as in the local and global anatomical modeling tasks. Then, we feed the context to a frozen pre-trained CheX-World model to obtain the model's feature prediction, on which a diffusion model conditions to produce the target image in pixel space. As shown in Figure 4(a), the CheX-World predictor perfectly recovers the appearance of bones", + "bbox": [ + 511, + 734, + 906, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "20783", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/034fb371fd2957b8b2d5edfa65f0da436d18be671962daf31ffbe7d7f3932778.jpg", + "table_caption": [ + "Table 1. Performance on five downstream classification benchmarks. Accuracy is reported for RSNA, and the area under the ROC curve (AUROC) is reported for the rest of the datasets. The best two results are bold-faced and underlined, respectively." + ], + "table_footnote": [ + "Results on ImageNet pre-trained models are adopted from [42]. ${}^{ \\dagger }$ Results reported by the original authors [31, 32]. ${}^{ \\ddagger }$ Rad-DINO [51] is included for reference but not for direct comparison, as it requires 2560 GPU-hours (20 times more than ours) and is trained on both public and private datasets." + ], + "table_body": "
MethodPre-train DataBackboneVinDr-CXRShenZhenChestX-ray14RSNA (CLS)CheXpert
Scratch-ViT-B70.22±1.9582.24±0.6071.69±0.3266.59±0.3980.78±0.03
MoCo-v3 [14]IN-1k (1.3M)ViT-B87.25±0.6392.85±1.0079.20±0.3072.79±0.5287.12±0.36
DINO [11]IN-1k (1.3M)ViT-B82.89±1.1090.39±4.2978.37±0.4771.27±0.4587.01±0.62
BEiT [6]IN-21k (12M)ViT-B85.93±1.9892.87±1.0879.91±0.2472.78±0.3787.77±0.38
LVM-Med [44]Medical (1.3M)ViT-B88.22±0.4494.81±1.3280.08±0.0972.75±0.4488.07±0.25
MAE [31, 63]X-rays (0.5M)ViT-B92.76±0.1897.63±0.2183.0†73.75±0.2489.3†
SimMIM [42, 64]IN-21k (12M) & X-rays (0.9M)Swin-B92.81±0.3198.09±0.1383.04±0.1574.09±0.3989.14±0.22
Adam-v2 [32]IN-21k (12M) & X-rays (0.9M)ConvNeXt-B91.46±0.3397.80†83.4†73.40±0.8888.90±0.36
CheXWorldX-rays (0.5M)ViT-B95.24±0.1398.88±0.0683.58±0.0575.03±0.3989.63±0.13
Rad-DINO‡[51]LVD (142M) & X-rays (public+private)ViT-B95.16±0.1698.20±0.1783.61±0.0874.51±0.4688.94±0.15
", + "bbox": [ + 93, + 122, + 867, + 333 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "and tissues within various masked areas, and Figure 4(b) shows that the model makes global-level predictions that are consistent with the context. Further visualizations also show that CheXWorld filters artifacts in the image (Figure 5) and learns anatomical correspondence (Figure 6).", + "bbox": [ + 88, + 372, + 480, + 448 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Awareness to domain variations. We design a domain sensitivity test to verify the model's ability to capture the uncertainty in a radiograph's appearance attributes. Specifically, we first generate a candidate set of targets $\\{y_i\\}_{i=1}^n$ by applying augmentations with different configurations to the same image. Then, we construct a set of context-target-latent triplets $\\{(x_i, y_i, a_i)\\}$ by applying another augmentation $x_i = \\mathcal{T}_{a_i}(y_i)$ . The model is asked to predict the target $y_i$ given the context $x_i$ and latent $a_i$ . For each predicted output $\\hat{y}_i$ , we calculate the top-k recall rate of the true target $y_i$ over the entire candidate set using L2 distance2. As shown in Table 4, our model achieves an average top-5 recall of 77.67 (10 times higher than random choice), demonstrating the model's strong discriminative ability across domains. We also observe a significant drop in the recall rate when removing the domain condition $a$ , indicating that the model indeed makes predictions conditioned on $a$ .", + "bbox": [ + 91, + 450, + 483, + 705 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. CheXWorld Produces Transferable Representations", + "text_level": 1, + "bbox": [ + 89, + 719, + 482, + 750 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Setup and baselines. We compare CheXWorld with several competitive baselines across eight classification and segmentation tasks using the following datasets: VinDr-CXR [46], ShenZhen-CXR [36], NIH ChestX-ray14 [61], CheXpert [35], MedFMC-ChestDR [60], SIIM-ACR Pneumothorax [1] and RSNA Pneumonia [54].", + "bbox": [ + 89, + 758, + 482, + 849 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/ce56c3d13f4282de30e178b7c70311f75a21241855cdf8fe0a450db76bb8732d.jpg", + "table_caption": [ + "Table 4. Effectiveness of domain variation modeling." + ], + "table_footnote": [], + "table_body": "
MethodRecall@1Recall@3Recall@5
Random Choice1.564.687.81
w/o domain condition15.8733.1244.37
CheXWorld38.4965.9077.67
", + "bbox": [ + 535, + 381, + 880, + 445 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We attach task-specific heads (linear heads and U-Net decoders) and perform full fine-tuning on the pre-trained CheXWorld encoder. The input image resolution is fixed to $224^2$ . We compare CheXWorld with several self-supervised learning methods designed for general domain images [6, 11, 14, 31, 64] and medical images [32, 44] using base-scale backbones [18, 40, 41]. We use the radiology version of these methods [42, 51, 63] if available.", + "bbox": [ + 511, + 446, + 906, + 568 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Classification results are shown in Table 1. Our method consistently outperforms the best alternatives on all the benchmarks. Although SimMIM and Adam-v2 are initialized from ImageNet weights and pre-trained on a large dataset (0.9M X-rays) and, yet still lag behind CheXWorld with considerable margins. For instance, CheXWorld surpasses Adam-v2 by 4.04 in AUROC on VinDr-CXR and SimMIM by 0.54 on ChestX-ray14. Our approach performs comparably to Rad-DINO across all benchmarks, despite Rad-DINO requiring 20 times more computational resources and utilizing private training data. Additionally, one can observe a significant contrast in performance between models trained on photometric images (ImageNet) and models trained on medical images, highlighting the importance of in-domain transfer.", + "bbox": [ + 511, + 569, + 908, + 794 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Segmentation results are presented in Table 2 and Figure 8. CheXWorld surpasses the leading baselines in terms of dice score and produces more accurate segmentation masks, demonstrating its superior capability in both instance-level and dense prediction settings.", + "bbox": [ + 511, + 795, + 906, + 869 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Few-shot learning results on the MedFMC dataset are presented in Table 2. CheXWorld demonstrates a clear ad", + "bbox": [ + 511, + 869, + 906, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_footnote", + "text": "2In our experiments the candidate set size is fixed at $n = 64$ . This process is repeated across multiple images, and the final result is obtained by averaging the outcomes.", + "bbox": [ + 89, + 862, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "20784", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/9180052ec78c08af70339ee4350518ed141c1c2537ade1dc812736284a503891.jpg", + "table_caption": [ + "Table 2. Results on segmentation (left) and few-shot learning (right) tasks. The dice score and the AUROC score are reported for the segmentation and few-shot learning benchmarks respectively." + ], + "table_footnote": [], + "table_body": "
MethodSegmentationFew-shot (MedFMC-ChestDR)
SIIM-ACRRSNA (SEG)1-shot5-shot10-shot
LVM-Med [44]82.19±0.3078.47±0.2457.55±0.8166.95±0.6067.44±0.71
MAE [31, 63]83.01±0.3977.39±0.2561.31±0.2774.03±0.3075.26±0.36
SimMIM [42, 64]82.89±0.0778.31±0.4160.64±1.9272.14±0.5474.42±0.38
Adam-v2 [32]83.60±0.3378.53±0.1959.16±1.0370.28±0.3970.67±0.71
CheXWorld84.58±0.3479.02±0.2564.60±1.0075.19±0.5176.40±0.25
", + "bbox": [ + 125, + 137, + 629, + 239 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/aa90d4e8eb98306fc27d669e2cd72040d0b5e7a2cfbff6ccd8b9336516d91b68.jpg", + "image_caption": [ + "Figure 7. Fine-tuning with $1\\%$ , $10\\%$ , and $100\\%$ training data on VinDr-CXR." + ], + "image_footnote": [], + "bbox": [ + 647, + 90, + 864, + 207 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/246c7835cd0224b1b6afc476f33c92a363d0667afee956e86dd0b7963ab13b18.jpg", + "image_caption": [ + "Figure 8. Visualization of segmentation masks on SIIM-ACR pneumothorax dataset. GT stands for the ground truth masks." + ], + "image_footnote": [], + "bbox": [ + 116, + 250, + 352, + 407 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/0e8a526b7479675368b2ddd95d5a3069833e70be9c5d409125530f07767b46c5.jpg", + "table_caption": [ + "Table 3. Ablation studies of the world modeling tasks and latent variables. We report the results on VinDr-CXR and RSNA with $1\\%$ proportion of training data. AUROC and accuracy are reported for the respective datasets. \"Local,\" \"Global,\" and \"Domain\" are abbreviations for local anatomical structure modeling, global anatomical layout modeling, and domain variation modeling, respectively. \"Unified\" indicates whether our unified formulation is applied, as opposed to a naive combination of losses." + ], + "table_footnote": [], + "table_body": "
World Modeling TaskLatent Variables Δx→y aUnifiedVinDr 1%RSNA 1%
Local.Global.Domain.
✓N.A.84.71±1.0065.89±0.54
✓✓✓✓88.31±0.8867.47±0.51
✓✓✓89.73±0.5068.27±1.00
✓✓✓✓✓90.53±1.0170.01±0.33
✓✓✓✓✓89.12±0.7668.37±0.50
✓✓✓✓✓89.91±0.6068.98±0.60
✓✓✓✓87.60±0.7266.56±0.39
", + "bbox": [ + 415, + 335, + 859, + 459 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "vantage on the 1-shot learning task and consistently outperforms the second-best methods in both 5-shot and 10-shot scenarios, showcasing quick adaptation to new concepts by leveraging knowledge acquired through world modeling.", + "bbox": [ + 89, + 469, + 482, + 530 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Sample efficiency results. The superiority of our method is further highlighted by reducing the training data proportion. As shown in Figure 7, we conduct experiments on VinDr-CXR with $1\\%$ , $10\\%$ , and $100\\%$ of the training data. Notably, our method with $10\\%$ of the data significantly outperforms all baselines trained on the entire dataset, reducing the annotation cost by over 90 percent.", + "bbox": [ + 89, + 530, + 483, + 638 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.3. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 643, + 243, + 660 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effectiveness of the world modeling tasks is validated in the upper part of Table 3. Combining local anatomical structure modeling with either of the other two world modeling tasks yields significant performance gains. The best performance is achieved when all three world modeling tasks are integrated within our unified framework, indicating that these tasks reflect different aspects of medical knowledge that are effectively captured by CheXWorld.", + "bbox": [ + 89, + 664, + 482, + 785 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Role of the latent variables. To verify that the latent variables (such as relative position $\\Delta_{x\\rightarrow y}$ for global anatomical layout modeling and transformation parameter $a$ for domain variation modeling) contribute to transfer learning performance, we test the models pre-trained without certain latent variables, shown in the bottom part of Table", + "bbox": [ + 89, + 785, + 483, + 876 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3. Domain variation modeling without the condition $a$ degrades into a blind image restoration task, which yields inferior results compared with the original method. Removing the condition $\\Delta_{x \\to y}$ from global anatomical layout modeling leads to a sharp decrease in performance because it confounds the model with erroneous target spatial locations.", + "bbox": [ + 511, + 469, + 906, + 561 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 575, + 633, + 590 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we proposed CheXWorld, a self-supervised world modeling framework for radiographic images that simultaneously encodes local and global anatomical knowledge and appearance domain variations. By integrating three world modeling tasks into a unified pipeline, our model effectively captures the spatial and domain uncertainties of the image world, as verified by qualitative and quantitative analytical experiments. Extensive transfer learning experiments on eight medical image analysis benchmarks demonstrate the state-of-the-art performance of our model. Our work offers new insights into representing and extracting knowledge from medical images, paving the way toward a general-purpose foundation model for medical vision.", + "bbox": [ + 511, + 601, + 906, + 799 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements.", + "text_level": 1, + "bbox": [ + 511, + 814, + 684, + 830 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The work is supported in part by the National Key R&D Program of China under Grant 2024YFB4708200 and National Natural Science Foundation of China under Grant 62321005.", + "bbox": [ + 511, + 839, + 906, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_footnote", + "text": "3Detailed numerical results can be found in the appendix.", + "bbox": [ + 107, + 886, + 413, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "20785", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 90, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Zawacki Anna, Wu Carol, Shih George, Elliott Julia, Fomitchev Mikhail, Hussain Mohannad, Lakhani Paras, Culliton Phil, and Bao Shunxing. Siim-acr pneumothorax segmentation. 2019. 7", + "[2] Mahmoud Assran, Quentin Duval, Ishan Misra, Piotr Bojanowski, Pascal Vincent, Michael Rabbat, Yann LeCun, and Nicolas Ballas. Self-supervised learning from images with a joint-embedding predictive architecture. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15619-15629, 2023. 1, 3, 4", + "[3] Shekoofeh Azizi, Basil Mustafa, Fiona Ryan, Zachary Beaver, Jan Freyberg, Jonathan Deaton, Aaron Loh, Alan Karthikesalingam, Simon Kornblith, Ting Chen, et al. Big self-supervised models advance medical image classification. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3478-3488, 2021. 3", + "[4] Alexei Baevski, Wei-Ning Hsu, Qiantong Xu, Arun Babu, Jiatao Gu, and Michael Auli. Data2vec: A general framework for self-supervised learning in speech, vision and language. In International Conference on Machine Learning, pages 1298-1312. PMLR, 2022. 3", + "[5] Alexei Baevski, Arun Babu, Wei-Ning Hsu, and Michael Auli. Efficient self-supervised learning with contextualized target representations for vision, speech and language. In International Conference on Machine Learning, pages 1416-1429. PMLR, 2023. 3", + "[6] Hangbo Bao, Li Dong, Songhao Piao, and Furu Wei. Beit: Bert pre-training of image transformers. arXiv preprint arXiv:2106.08254, 2021.7", + "[7] Adrien Bardes, Quentin Garrido, Jean Ponce, Xinlei Chen, Michael Rabbat, Yann LeCun, Mido Assran, and Nicolas Ballas. V-jepa: Latent video prediction for visual representation learning. 2023. 3", + "[8] Florian Bordes, Randall Balestriero, and Pascal Vincent. High fidelity visualization of what your self-supervised representation knows about. arXiv preprint arXiv:2112.09164, 2021. 6", + "[9] Arthur Earl Bryson. Applied optimal control: optimization, estimation and control. Routledge, 2018. 2", + "[10] Eduardo F Camacho, Carlos Bordons, Eduardo F Camacho, and Carlos Bordons. Constrained model predictive control. Springer, 2007. 2", + "[11] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 7", + "[12] Krishna Chaitanya, Ertunc Erdil, Neerav Karani, and Ender Konukoglu. Contrastive learning of global and local features for medical image segmentation with limited annotations. Advances in neural information processing systems, 33:12546-12558, 2020. 3", + "[13] Richard J Chen, Tong Ding, Ming Y Lu, Drew FK Williamson, Guillaume Jaume, Andrew H Song, Bowen Chen, Andrew Zhang, Daniel Shao, Muhammad Shaban, et al. Towards a general-purpose foundation model for" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "computational pathology. Nature Medicine, 30(3):850-862, 2024. 3", + "[14] Xinlei Chen, Saining Xie, and Kaiming He. An empirical study of training self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9640-9649, 2021. 7", + "[15] Peng Cui, Yang Yue, Zhijie Deng, and Jun Zhu. Confidence-based reliable learning under dual noises. Advances in Neural Information Processing Systems, 35:35116-35129, 2022. 5", + "[16] Rumen Dangovski, Li Jing, Charlotte Loh, Seungwook Han, Akash Srivastava, Brian Cheung, Pulkit Agrawal, and Marin Soljacic. Equivariant contrastive learning. arXiv preprint arXiv:2111.00899, 2021. 3, 5", + "[17] Alexandre Devillers and Mathieu Lefort. Equimod: An equivariance module to improve visual instance discrimination. In International Conference on Learning Representations, 2023. 3, 5", + "[18] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 3, 6, 7", + "[19] Chaoqun Du, Yulin Wang, Shiji Song, and Gao Huang. Probabilistic contrastive learning for long-tailed visual recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 3", + "[20] Cong Gao, Benjamin D Killeen, Yicheng Hu, Robert B Grupp, Russell H Taylor, Mehran Armand, and Mathias Unberath. Synthetic data accelerates the development of generalizable learning-based algorithms for x-ray image analysis. Nature Machine Intelligence, 5(3):294-308, 2023. 1", + "[21] Quentin Garrido, Mahmoud Assran, Nicolas Ballas, Adrien Bardes, Laurent Najman, and Yann LeCun. Learning and leveraging world models in visual representation learning. arXiv preprint arXiv:2403.00504, 2024. 1, 3", + "[22] Ross Girshick, Jeff Donahue, Trevor Darrell, and Jitendra Malik. Rich feature hierarchies for accurate object detection and semantic segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 580-587, 2014. 6", + "[23] Hao Guan and Mingxia Liu. Domain adaptation for medical image analysis: a survey. IEEE Transactions on Biomedical Engineering, 69(3):1173-1185, 2021. 2, 5", + "[24] David Ha and Jürgen Schmidhuber. Recurrent world models facilitate policy evolution. Advances in neural information processing systems, 31, 2018. 2", + "[25] David Ha and Jürgen Schmidhuber. World models. arXiv preprint arXiv:1803.10122, 2018. 3", + "[26] Danijar Hafner, Timothy Lillicrap, Jimmy Ba, and Mohammad Norouzi. Dream to control: Learning behaviors by latent imagination. arXiv preprint arXiv:1912.01603, 2019.", + "[27] Danijar Hafner, Timothy Lillicrap, Mohammad Norouzi, and Jimmy Ba. Mastering atari with discrete world models. arXiv preprint arXiv:2010.02193, 2020. 2" + ], + "bbox": [ + 517, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "20786", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[28] Danijar Hafner, Jurgis Pasukonis, Jimmy Ba, and Timothy Lillicrap. Mastering diverse domains through world models. arXiv preprint arXiv:2301.04104, 2023. 3", + "[29] Fatemeh Haghighi, Mohammad Reza Hosseinzadeh Taher, Zongwei Zhou, Michael B Gotway, and Jianming Liang. Transferable visual words: Exploiting the semantics of anatomical patterns for self-supervised learning. IEEE transactions on medical imaging, 40(10):2857-2868, 2021. 3", + "[30] Fatemeh Haghighi, Mohammad Reza Hosseinzadeh Taher, Michael B Gotway, and Jianming Liang. Dira: Discriminative, restorative, and adversarial learning for self-supervised medical image analysis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20824-20834, 2022. 3", + "[31] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 3, 4, 7, 8", + "[32] Mohammad Reza Hosseinzadeh Taher, Michael Gotway, and Jianming Liang. Representing part-whole hierarchies in foundation models by learning localizability, composability, and decomposability from anatomy via self-supervision. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2024. 3, 7, 8", + "[33] Anthony Hu, Lloyd Russell, Hudson Yeo, Zak Murez, George Fedoseev, Alex Kendall, Jamie Shotton, and Gianluca Corrado. Gaia-1: A generative world model for autonomous driving. arXiv preprint arXiv:2309.17080, 2023. 3", + "[34] Zhi Huang, Federico Bianchi, Mert Yuksekgonul, Thomas J Montine, and James Zou. A visual-language foundation model for pathology image analysis using medical twitter. Nature medicine, 29(9):2307-2316, 2023. 1", + "[35] Jeremy Irvin, Pranav Rajpurkar, Michael Ko, Yifan Yu, Silviana Ciurea-Ilicus, Chris Chute, Henrik Marklund, Behzad Haghgoo, Robyn Ball, Katie Shpanskaya, et al. Chexpert: A large chest radiograph dataset with uncertainty labels and expert comparison. In Proceedings of the AAAI conference on artificial intelligence, pages 590-597, 2019. 7", + "[36] Stefan Jaeger, Sema Candemir, Sameer Antani, Yi-Xiang J Wang, Pu-Xuan Lu, and George Thoma. Two public chest x-ray datasets for computer-aided screening of pulmonary diseases. Quantitative imaging in medicine and surgery, 4(6): 475, 2014. 7", + "[37] Bingyi Kang, Yang Yue, Rui Lu, Zhijie Lin, Yang Zhao, Kaixin Wang, Gao Huang, and Jiashi Feng. How far is video generation from world model: A physical law perspective. arXiv preprint arXiv:2411.02385, 2024. 3", + "[38] Dan Kondratyuk, Lijun Yu, Xiuye Gu, José Lezama, Jonathan Huang, Rachel Hornung, Hartwig Adam, Hassan Akbari, Yair Alon, Vighnesh Birodkar, et al. Videopoet: A large language model for zero-shot video generation. arXiv preprint arXiv:2312.14125, 2023. 3", + "[39] Yann LeCun. A path towards autonomous machine intelligence version 0.9.2, 2022-06-27. Open Review, 62(1), 2022. 1, 2, 3" + ], + "bbox": [ + 91, + 92, + 482, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[40] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 7", + "[41] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11976-11986, 2022. 7", + "[42] DongAo Ma, Mohammad Reza Hosseinzadeh Taher, Jiaxuan Pang, Nahid UI Islam, Fatemeh Haghighi, Michael B Gotway, and Jianming Liang. Benchmarking and boosting transformers for medical image classification. In MICCAI Workshop on Domain Adaptation and Representation Transfer, pages 12-22. Springer, 2022. 7, 8", + "[43] Phillip Marsh. Case courtesy of phillip marsh, radiopaedia.org. https://radiopaedia.org/cases/58938, 2023. Case ID: rID: 58938. 2", + "[44] Duy MH Nguyen, Hoang Nguyen, Nghiem Diep, Tan Ngoc Pham, Tri Cao, Binh Nguyen, Paul Swoboda, Nhat Ho, Shadi Albarqouni, Pengtao Xie, et al. Lvm-med: Learning large-scale self-supervised vision models for medical imaging via second-order graph matching. Advances in Neural Information Processing Systems, 36, 2024. 7, 8", + "[45] Michael Moor, Oishi Banerjee, Zahra Shakeri Hossein Abad, Harlan M Krumholz, Jure Leskovec, Eric J Topol, and Pranav Rajpurkar. Foundation models for generalist medical artificial intelligence. Nature, 616(7956):259-265, 2023. 1", + "[46] Ha Q Nguyen, Khanh Lam, Linh T Le, Hieu H Pham, Dat Q Tran, Dung B Nguyen, Dung D Le, Chi M Pham, Hang TT Tong, Diep H Dinh, et al. Vindr-cxr: An open dataset of chest x-rays with radiologist's annotations. Scientific Data, 9(1):429, 2022. 7", + "[47] OpenAI. Video generation models as world simulators. https://openai.com/index/video-generation-models-as-world-simulators/, 2024. Accessed: Feb, 2024. 3", + "[48] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. Transactions on Machine Learning Research Journal, pages 1-31, 2024. 3", + "[49] Cheng Ouyang, Chen Chen, Surui Li, Zeju Li, Chen Qin, Wenjia Bai, and Daniel Rueckert. Causality-inspired single-source domain generalization for medical image segmentation. IEEE Transactions on Medical Imaging, 42(4):1095-1106, 2022. 2, 5", + "[50] Jiaxuan Pang, Fatemeh Haghighi, DongAo Ma, Nahid Ul Islam, Mohammad Reza Hosseinzadeh Taher, Michael B Gotway, and Jianming Liang. Popar: Patch order prediction and appearance recovery for self-supervised medical image analysis. In MICCAI Workshop on Domain Adaptation and Representation Transfer, pages 77-87. Springer, 2022. 3", + "[51] Fernando Pérez-García, Harshita Sharma, Sam Bond-Taylor, Kenya Bouzid, Valentina Salvatelli, Maximilian Ilse, Shruthi" + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "20787", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Bannur, Daniel C Castro, Anton Schwaighofer, Matthew P Lungren, et al. Rad-dino: Exploring scalable medical image encoders beyond text supervision. arXiv preprint arXiv:2401.10815, 2024. 3, 7", + "[52] Eric Postal. Why don't radiology textbooks have imperfect images? https://www.diagnosticimaging.com/view/why-dont-radiology-textbooks-have-imperfect-images, 2019. Accessed: Feb, 2024. 2, 5", + "[53] RunwayML. Introducing general world models. https://research.runwayml.com/introducing-general-world-models, 2023. Accessed: May, 2023. 2", + "[54] George Shih, Carol C Wu, Safwan S Halabi, Marc D Kohli, Luciano M Prevedello, Tessa S Cook, Arjun Sharma, Judith K Amorosa, Veronica Arteaga, Maya Galperin-Aizenberg, et al. Augmenting the national institutes of health chest radiograph dataset with expert annotations of possible pneumonia. Radiology: Artificial Intelligence, 1(1): e180041, 2019. 7", + "[55] Ben Sorscher, Surya Ganguli, and Haim Sompolinsky. Neural representational geometry underlies few-shot concept learning. Proceedings of the National Academy of Sciences, 119(43):e2200800119, 2022. 1", + "[56] Hari Sowrirajan, Jingbo Yang, Andrew Y Ng, and Pranav Rajpurkar. Moco pretraining improves representation and transferability of chest x-ray models. In Medical Imaging with Deep Learning, pages 728-744. PMLR, 2021. 3", + "[57] Mohammad Reza Hosseinzadeh Taher, Fatemeh Haghighi, Michael B Gotway, and Jianming Liang. Caid: a self-supervised learning framework for empowering instance discrimination in medical imaging. Proceedings of Machine Learning Research, 3, 2022. 3", + "[58] Xing Tao, Yuexiang Li, Wenhui Zhou, Kai Ma, and Yefeng Zheng. Revisiting rubik's cube: self-supervised learning with volume-wise transformation for 3d medical image segmentation. In Medical Image Computing and Computer Assisted Intervention-MICCAI 2020: 23rd International Conference, Lima, Peru, October 4-8, 2020, Proceedings, Part IV 23, pages 238-248. Springer, 2020. 3", + "[59] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 4", + "[60] Dequan Wang, Xiaosong Wang, Lilong Wang, Mengzhang Li, Qian Da, Xiaoqiang Liu, Xiangyu Gao, Jun Shen, Junjun He, Tian Shen, et al. A real-world dataset and benchmark for foundation model adaptation in medical image classification. Scientific Data, 10(1):574, 2023. 7", + "[61] Xiaosong Wang, Yifan Peng, Le Lu, Zhiyong Lu, Mohammadhadi Bagheri, and Ronald M Summers. Chestxray8: Hospital-scale chest x-ray database and benchmarks on weakly-supervised classification and localization of common thorax diseases. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2097-2106, 2017. 2, 7", + "[62] Yulin Wang, Yang Yue, Rui Lu, Tianjiao Liu, Zhao Zhong, Shiji Song, and Gao Huang. Efficienttrain: Exploring generalized curriculum learning for training visual backbones. In" + ], + "bbox": [ + 91, + 90, + 483, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5852-5864, 2023. 3", + "[63] Junfei Xiao, Yutong Bai, Alan Yuille, and Zongwei Zhou. Delving into masked autoencoders for multi-label thorax disease classification. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 3588-3600, 2023. 3, 7, 8", + "[64] Zhenda Xie, Zheng Zhang, Yue Cao, Yutong Lin, Jianmin Bao, Zhuliang Yao, Qi Dai, and Han Hu. Simmim: A simple framework for masked image modeling. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9653-9663, 2022. 7, 8", + "[65] Ke Yan, Jinzheng Cai, Dakai Jin, Shun Miao, Dazhou Guo, Adam P Harrison, Youbao Tang, Jing Xiao, Jingjing Lu, and Le Lu. Sam: Self-supervised learning of pixel-wise anatomical embeddings in radiological images. IEEE Transactions on Medical Imaging, 41(10):2658-2669, 2022. 3", + "[66] Hong-Yu Zhou, Shuang Yu, Cheng Bian, Yifan Hu, Kai Ma, and Yefeng Zheng. Comparing to learn: Surpassing image-genet pretraining on radiographs by comparing image representations. In Medical Image Computing and Computer-Assisted Intervention-MICCAI 2020: 23rd International Conference, Lima, Peru, October 4-8, 2020, Proceedings, Part I 23, pages 398-407. Springer, 2020. 3", + "[67] Lei Zhou, Huidong Liu, Joseph Bae, Junjun He, Dimitris Samaras, and Prateek Prasanna. Self pre-training with masked autoencoders for medical image classification and segmentation. In 2023 IEEE 20th International Symposium on Biomedical Imaging (ISBI), pages 1-6. IEEE, 2023. 3", + "[68] Yukun Zhou, Mark A Chia, Siegfried K Wagner, Murat S Ayhan, Dominic J Williamson, Robbert R Struyven, Timing Liu, Moucheng Xu, Mateo G Lozano, Peter Woodward-Court, et al. A foundation model for generalizable disease detection from retinal images. Nature, 622(7981):156-163, 2023. 1, 3", + "[69] Zongwei Zhou, Vatsal Sodha, Jiaxuan Pang, Michael B Gotway, and Jianming Liang. Models genesis. Medical image analysis, 67:101840, 2021. 2, 3", + "[70] Ziyu Zhou, Haozhe Luo, Jiaxuan Pang, Xiaowei Ding, Michael Gotway, and Jianming Liang. Learning anatomically consistent embedding for chest radiography. arXiv preprint arXiv:2312.00335, 2023. 2, 3" + ], + "bbox": [ + 516, + 92, + 906, + 683 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "20788", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/CheXWorld_ Exploring Image World Modeling for Radiograph Representation Learning/f79a692c-48ff-4b64-8cb4-11cdf9303afb_model.json b/2025/CheXWorld_ Exploring Image World Modeling for Radiograph Representation Learning/f79a692c-48ff-4b64-8cb4-11cdf9303afb_model.json new file mode 100644 index 0000000000000000000000000000000000000000..f47827f9df6f0269486d8669794133c90a793172 --- /dev/null +++ b/2025/CheXWorld_ Exploring Image World Modeling for Radiograph Representation Learning/f79a692c-48ff-4b64-8cb4-11cdf9303afb_model.json @@ -0,0 +1,2994 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.26, + 0.131, + 0.744, + 0.178 + ], + "angle": 0, + "content": "CheXWorld: Exploring Image World Modeling for Radiograph Representation Learning" + }, + { + "type": "text", + "bbox": [ + 0.154, + 0.211, + 0.846, + 0.231 + ], + "angle": 0, + "content": "Yang Yue\\(^{1*}\\) Yulin Wang\\(^{1*}\\) Chenxin Tao\\(^{1}\\) Pan Liu\\(^{2}\\) Shiji Song\\(^{1}\\) Gao Huang\\(^{1\\boxtimes}\\)" + }, + { + "type": "text", + "bbox": [ + 0.355, + 0.232, + 0.646, + 0.248 + ], + "angle": 0, + "content": "\\(^{1}\\)Tsinghua University \\(^{2}\\)PLA General Hospital" + }, + { + "type": "text", + "bbox": [ + 0.243, + 0.248, + 0.75, + 0.263 + ], + "angle": 0, + "content": "yueyang22@mails.tsinghua.edu.cn, gaohuang@tsinghua.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.297, + 0.327, + 0.314 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.33, + 0.486, + 0.754 + ], + "angle": 0, + "content": "Humans can develop internal world models that encode common sense knowledge, telling them how the world works and predicting the consequences of their actions. This concept has emerged as a promising direction for establishing general-purpose machine-learning models in recent preliminary works, e.g., for visual representation learning. In this paper, we present CheXWorld, the first effort towards a self-supervised world model for radiographic images. Specifically, our work develops a unified framework that simultaneously models three aspects of medical knowledge essential for qualified radiologists, including 1) local anatomical structures describing the fine-grained characteristics of local tissues (e.g., architectures, shapes, and textures); 2) global anatomical layouts describing the global organization of the human body (e.g., layouts of organs and skeletons); and 3) domain variations that encourage CheXWorld to model the transitions across different appearance domains of radiographs (e.g., varying clarity, contrast, and exposure caused by collecting radiographs from different hospitals, devices, or patients). Empirically, we design tailored qualitative and quantitative analyses, revealing that CheX-World successfully captures these three dimensions of medical knowledge. Furthermore, transfer learning experiments across eight medical image classification and segmentation benchmarks showcase that CheXWorld significantly outperforms existing SSL methods and large-scale medical foundation models. Code & pre-trained models are available at https://github.com/LeapLabTHU/CheXWorld." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.785, + 0.224, + 0.801 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.485, + 0.873 + ], + "angle": 0, + "content": "Intelligent agents like humans learn extensive background knowledge about the world [39]. This common-sense information is embedded in the agents' internal models of the world, playing a pivotal role in their perception, learning," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.299, + 0.908, + 0.451 + ], + "angle": 0, + "content": "and decision-making processes by simulating the world's dynamics, telling them what is plausible or impossible, and predicting the outcomes of their actions. Consequently, these world models enable agents to acquire new concepts and skills with minimal demonstrations and trials [55]. Recent works [2, 21] have preliminarily verified the effectiveness of establishing visual representation learning approaches through the lens of world modeling. Pre-trained world models can produce semantically rich embeddings and adapt to various downstream tasks with limited data." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.457, + 0.909, + 0.593 + ], + "angle": 0, + "content": "Like many other research areas, the field of medical imaging is experiencing a paradigm shift from task-specific models to general-purpose foundation models [20, 34, 45, 68], which are pre-trained on massive data and expected to encode meaningful medical knowledge. The idea of world modeling offers an approach to training a medical foundation model by capturing common sense information (e.g. human anatomy) from medical images, which is a promising yet under-explored research direction." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.599, + 0.909, + 0.903 + ], + "angle": 0, + "content": "In this paper, we present CheXWorld, the first initiative toward world modeling for self-supervised representation learning on radiographic images. We propose and integrate three world modeling tasks that capture different dimensions of medical knowledge essential for qualified radiologists, depicted in Figure 1. In particular, a radiologist with a strong understanding of human anatomy can identify different anatomical structures and determine the relative geometry between anatomical regions. Given the hierarchical nature of human anatomy, we introduce world modeling tasks that focus on both local and global levels of anatomy: local anatomical structure modeling aims to predict specific anatomical characteristics and structures within a localized anatomical region, such as bones, airways, blood vessels, and lung segments; global anatomical layout modeling learns the overall arrangement and spatial relationships of various anatomical regions within the human body, understanding the relative positioning of organs and tissues such as the heart, lungs, diaphragm, and rib cage. Additionally, medical images are" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.886, + 0.221, + 0.901 + ], + "angle": 0, + "content": "*Equal contribution." + }, + { + "type": "page_footnote", + "bbox": [ + 0.264, + 0.886, + 0.401, + 0.901 + ], + "angle": 0, + "content": "Corresponding authors." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20778" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.118, + 0.096, + 0.37, + 0.312 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.131, + 0.317, + 0.361, + 0.33 + ], + "angle": 0, + "content": "(a) Local Anatomical Structure Modeling" + }, + { + "type": "image", + "bbox": [ + 0.376, + 0.097, + 0.63, + 0.31 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.39, + 0.317, + 0.614, + 0.33 + ], + "angle": 0, + "content": "(b) Global Anatomical Layout Modeling" + }, + { + "type": "image", + "bbox": [ + 0.636, + 0.092, + 0.882, + 0.315 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.667, + 0.317, + 0.841, + 0.33 + ], + "angle": 0, + "content": "(c) Domain Variation Modeling" + }, + { + "type": "image", + "bbox": [ + 0.116, + 0.342, + 0.867, + 0.457 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.34, + 0.46, + 0.496, + 0.472 + ], + "angle": 0, + "content": "(d) Unified World Modeling" + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.472, + 0.907, + 0.543 + ], + "angle": 0, + "content": "Figure 1. Overview of the CheXWorld framework1. The upper part of the figure depicts three dimensions of medical knowledge that are formulated in our framework, including (a) local anatomical structures describing the fine-grained characteristics of local tissues, (b) global anatomical layouts describing the global organization of the human body and (c) domain variations that encourage CheXWorld to model the transitions across different appearance domains of radiographs. The middle part of the figure illustrates the world modeling tasks corresponding to these aspects of medical knowledge. (d) shows our unified pipeline that combines the merits of all three tasks." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.547, + 0.483, + 0.773 + ], + "angle": 0, + "content": "typically gathered from multiple domains with notable appearance differences due to varying equipment and acquisition techniques [52]. Experienced radiologists can recognize the physical reality behind domains with different appearances and facilitate the diagnosis process. Motivated by this, we hypothesize that radiologists quickly adapt to diverse domain changes by learning sophisticated internal world models that are capable of simulating transitions between domains. Driven by our hypothesis, we introduce the domain variation modeling task, which aims to learn an expressive feature space that encodes predictable transitions between domains. Finally, we design an integrated pipeline that performs the three world modeling tasks simultaneously, effectively incorporating medical anatomical knowledge and achieving robustness across domains." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.774, + 0.483, + 0.866 + ], + "angle": 0, + "content": "It is noteworthy that the three aspects we consider, i.e., local anatomical structure modeling, global anatomical layout modeling, and domain variation modeling, represent core elements of radiographic expertise, each contributing to a comprehensive framework for capturing medical knowledge. While they may not encompass every dimen" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.547, + 0.907, + 0.592 + ], + "angle": 0, + "content": "sion of medical understanding, they are fundamental to radiology and critical for radiograph representation learning, as extensively demonstrated in [23, 49, 69, 70]." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.593, + 0.909, + 0.744 + ], + "angle": 0, + "content": "Empirically, we validate that CheXWorld effectively captures the three dimensions of medical knowledge through a series of analytical experiments, including visualizations of predictor outputs with the help of a generative model and a domain variation sensitivity test. Moreover, we extensively evaluate CheXWorld on eight medical image classification and segmentation benchmarks. Our model consistently outperforms competitive self-supervised learning baselines with comparable backbone capacity, pretraining data, and pre-training computational cost." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.755, + 0.657, + 0.77 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.909, + 0.903 + ], + "angle": 0, + "content": "World Modeling. In the machine learning community, world modeling originates from model-predictive control [9, 10] and is a common practice in reinforcement learning, where it predicts the future state of the environment based on the agent's action [24-27]. Recently, building general world models featuring profound comprehension of common sense knowledge is believed to be a crucial step towards general artificial intelligence [39, 53]. Video gener" + }, + { + "type": "page_footnote", + "bbox": [ + 0.09, + 0.875, + 0.483, + 0.902 + ], + "angle": 0, + "content": "1The x-rays are taken from kenhub.com, Radiopedia [43] and ChestXray14 [61]." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20779" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.302 + ], + "angle": 0, + "content": "ation models [37] trained on vast data (e.g. SoRA[47] and VideoPoet [38]) are world models that perform predictions in input space. Another line of work makes predictions in the latent space [2, 7, 28, 33, 39], which more closely resembles the world model in the human brain. World modeling is also an effective approach for visual representation learning, where the model predicts missing parts of the visual inputs [2, 4, 5, 7, 21]. Our framework builds on this foundation, where our model acquires knowledge—such as human anatomy and domain variations—through learning to predict unobserved outcomes from specific actions. Our key contribution lies in providing insights into modeling medical knowledge from radiographic images and designing a framework that unifies three world modeling tasks." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.303, + 0.483, + 0.575 + ], + "angle": 0, + "content": "Self-supervised learning in medical imaging. Self-supervised learning (SSL) [19, 31, 48, 62] provides a practical solution to mitigating annotation scarcity in medical imaging. A large body of works incorporate restorative SSL tasks, which typically involve reconstructing the original data from certain image corruptions, such as random masking [63, 67, 68], pixel shuffling [69], and patch order shuffling [50, 58, 70]. Another line of works [3, 12, 51, 56, 66] adapt global or dense contrastive learning to medical imaging. TransVW[29] and DiRA [30, 57] explore combining discriminative and restorative methods. Additionally, learning from anatomy is a common topic for medical SSL. SAM [65] enforces pixel-level anatomical correspondence between images, and Adam [32] learns part-whole correspondence that reflects the hierarchy of human anatomy. Recently, SSL-based foundation models trained on largescale datasets gained significant attention, such as RET-Found [68] for retinal imaging and UNI [13] for pathology." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.576, + 0.483, + 0.786 + ], + "angle": 0, + "content": "Compared with existing works, CheXWorld is the first to introduce the concept of world modeling into medical SSL, capturing task-relevant medical knowledge through three tailored world modeling tasks. Unlike contrastive-based approaches such as Adam [32], CheXWorld takes a technically distinct path by constructing equivariant [16, 17] representations—where input transformations result in predictable changes within the embedding space—rather than invariant representations. This approach allows CheXWorld to better model medical data without discarding valuable information. Moreover, our unified SSL framework delivers state-of-the-art performance across eight benchmarks, demonstrating the significant potential of designing SSL methods grounded in the philosophy of world modeling." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.8, + 0.427, + 0.817 + ], + "angle": 0, + "content": "3. Basic Framework of World Modeling" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.483, + 0.902 + ], + "angle": 0, + "content": "World modeling. We begin by briefly introducing the concept of world modeling [25, 39], which forms the basis for our proposed method. The primary motivation behind world modeling is to predict the unobserved parts \\( y \\) of a world (e.g., a visual environment, an image, or a video)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.243 + ], + "angle": 0, + "content": "based on an observed context \\( x \\). This prediction can be formulated across various dimensions, such as spatial (predicting the unseen regions of the data) and temporal (foreseeing the consequence of an action). In fact, the paradigm of world modeling takes inspiration from human visual cognition. For example, when presented with the right upper lobe region of a chest radiograph as the context \\( x \\), the internal world of a radiologist could imagine that there are two lobes (right middle and right lower) below the visible region and two lobes on the left, forming the target \\( y \\)." + }, + { + "type": "title", + "bbox": [ + 0.536, + 0.249, + 0.675, + 0.262 + ], + "angle": 0, + "content": "A basic framework." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.263, + 0.677, + 0.383 + ], + "angle": 0, + "content": "Here, we introduce our basic framework of instantiating world modeling, upon which we will further discuss how to establish a radiology world model by formulating various medical" + }, + { + "type": "image", + "bbox": [ + 0.686, + 0.257, + 0.9, + 0.348 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.684, + 0.351, + 0.903, + 0.38 + ], + "angle": 0, + "content": "Figure 2. A basic framework of world modeling." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.385, + 0.905, + 0.611 + ], + "angle": 0, + "content": "knowledge in Section 4. In specific, our basic framework mainly follows the joint-embedding predictive architecture (JEPA) [2, 39], as shown in Figure 2. We consider two encoders: the context encoder \\( f_{\\theta} \\) and the target encoder \\( f_{\\theta'}' \\), which encode the context \\( x \\) and target \\( y \\) into representations \\( h_x \\) and \\( h_y \\), respectively. A predictor \\( g_{\\phi} \\) learns to predict the target embedding \\( h_y \\) using the context embedding \\( h_x \\), conditioned on an additional latent variable \\( z \\) that indicates how the target relates to the context. For example, \\( z \\) can be the spatial locations of \\( y \\) with respect to \\( x \\), or an action that leads to the transition from context to target \\( x \\rightarrow y \\) (e.g., image transformations). The latent variable \\( z \\) carries the information that makes \\( y \\) predictable based on \\( x \\), effectively modeling the inherent variation and uncertainty of the real world. The primary objective of world modeling is:" + }, + { + "type": "equation", + "bbox": [ + 0.563, + 0.616, + 0.905, + 0.634 + ], + "angle": 0, + "content": "\\[\n\\text {m i n i m i z e} \\quad D \\left(h _ {y}, \\hat {h} _ {y}\\right) = D \\left(h _ {y}, g _ {\\phi} \\left(h _ {x}; z\\right)\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.639, + 0.906, + 0.774 + ], + "angle": 0, + "content": "where \\( D(\\cdot, \\cdot) \\) denotes the prediction error. Notably, the predictions are performed in the abstract representation space, contrasting with generative world models that predict every detail of the target. This mirrors an important characteristic of intelligent agents: information filtering, i.e., eliminating the irrelevant details in the data during the perception process. For example, a radiologist can infer the anatomical layout of a partly missed image, but it's nearly impossible for him to recover each pixel value of the unseen regions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.781, + 0.906, + 0.9 + ], + "angle": 0, + "content": "In implementation, we set the target encoder to be the exponential moving average of the context encoder following [2, 4, 7]. Moreover, we use the Vision Transformer (ViT) [18] architecture for the context/target encoder and the predictor. For the rest of this paper, the context and target feature \\( h_x, h_y \\) are sequences of patch embeddings produced by the ViTs. To perform patch-level feature predictions, we attach mask tokens to the context features and feed the com" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20780" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.123 + ], + "angle": 0, + "content": "binned sequence to the predictor. The output representations of the mask tokens are used as the final prediction." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.133, + 0.483, + 0.171 + ], + "angle": 0, + "content": "4. CheXWorld: World Modeling for Radiograph Representation Learning" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.177, + 0.484, + 0.465 + ], + "angle": 0, + "content": "This section extends the idea of world modeling to radiology. We propose three world modeling tasks tailored for radiographs and establish a unified CheXWorld framework that seamlessly integrates them. Importantly, the three tasks are designed to model three critical dimensions of medical knowledge essential for qualified radiologists. To be specific, these tasks are organized hierarchically from low-level to high-level, namely 1) local anatomical structure modeling that aims to learn the fine-grained structures of local anatomical regions, 2) global anatomical layout modeling that seeks to learn the global geometry of the human body (e.g., the layout of organs and skeletons), and 3) domain variation modeling that learns to model the transition across different appearance domains of radiographs. We first elaborate on the details of each task. Then, we present a unified framework that comprehensively combines the characteristics and merits of all three tasks, yielding a powerful foundation model that produces semantically rich and transferable representations." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.472, + 0.42, + 0.489 + ], + "angle": 0, + "content": "4.1. Local Anatomical Structure Modeling" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.494, + 0.484, + 0.719 + ], + "angle": 0, + "content": "At a relatively low, fundamental level, various types of tissues (e.g., bones, muscles, and epithelial tissues) form the structural organization of the human body. Understanding the fine-grained characteristics (e.g., shapes, sizes, appearance, and textures) of these micro-structures is essential knowledge that radiologists must possess. To build up a learning procedure that enables neural networks to acquire such local anatomical knowledge, we consider a mask-and-reconstruction task. The model predicts fine-grained details of a masked tissue region based on its peripheral, surrounding information, as shown in Figure 1(a). This encourages the networks to develop an internal model that captures the intricate structures and local continuities of human microanatomy. The large, continuous prediction target also prevents reliance on low-level features." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Specifically, the image mask is a union of four randomly selected rectangular regions of the image following [2]. Let \\(M\\) denote the set of 2D patch locations that are masked from the context input. Given an input image (crop) \\(I\\), the image patches with locations contained in \\(M\\) are dropped from the image to create the context \\(x = \\operatorname{Mask}(I, M)\\), while the target is the entire image. The context and target are then fed to the corresponding encoders, producing representations \\(h_x = f_\\theta(x)\\) and \\(h_y = f_{\\theta'}'(y)\\). Note that the context encoder exclusively processes visible patches, following [2, 31]. In the predictor, the mask tokens \\(m\\) carrying the positional encoding of the masked locations are concatenated with the" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.123 + ], + "angle": 0, + "content": "context feature \\( h_x \\), which forms the input token sequence to the predictor. The predictor output \\( \\dot{h}_y \\) is given by:" + }, + { + "type": "equation", + "bbox": [ + 0.528, + 0.128, + 0.905, + 0.175 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\hat {h} _ {y} = g _ {\\phi} (h _ {x}; M) \\\\ = g _ {\\phi} \\left(\\left(h _ {x} + p _ {x}\\right) \\oplus \\{m + \\operatorname {P E} (u, v) \\} _ {(u, v) \\in M}\\right), \\tag {2} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.179, + 0.907, + 0.27 + ], + "angle": 0, + "content": "where \\(\\oplus\\) is the concatenation operation along sequence dimension, \\(p_x\\) is the positional embedding corresponding to the context, \\(u\\) and \\(v\\) are image patch location indices along height and width dimensions, \\(\\mathrm{PE}(\\cdot)\\) is the sinusoidal positional encoding proposed by [59]. The prediction loss is computed only on the masked locations, which is given by:" + }, + { + "type": "equation", + "bbox": [ + 0.535, + 0.286, + 0.907, + 0.32 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {l o c a l}} (x, y) = \\sum_ {c \\in M} \\| g _ {\\phi} \\left(f _ {\\theta} (x); M\\right) _ {c} - f _ {\\theta^ {\\prime}} ^ {\\prime} (y) _ {c} \\| _ {2} ^ {2}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.323, + 0.899, + 0.339 + ], + "angle": 0, + "content": "where \\(c\\) is the 2D image patch location that belongs to \\(M\\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.356, + 0.833, + 0.373 + ], + "angle": 0, + "content": "4.2. Global Anatomical Layout Modeling" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.381, + 0.907, + 0.593 + ], + "angle": 0, + "content": "In addition to understanding local microstructures like tissues, it is also important for radiologists to possess comprehensive medical knowledge regarding how groups of tissues or systems of organs are anatomically organized and assembled into the full bodies of humans. Inspired by this, we develop the global anatomical layout modeling task (see Figure 1(b)), seeking to predict the feature of an out-of-context area based on its relative position to a given context, thus facilitating understanding of the global structure of human bodies. Notably, this idea is orthogonal to the technique proposed in section 4.1 (see section 5.3 for ablation studies) as it focuses on formulating the long-range topological relationships of different tissues or organs instead of the fine-grained anatomical structures within local tissues." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.595, + 0.679, + 0.793 + ], + "angle": 0, + "content": "As illustrated in Figure 1(b), we randomly crop two anatomical areas from the original radiograph, serving as context \\( x \\) and target \\( y \\), respectively. The world model learns to predict \\( y \\) from \\( x \\) given their relative position information \\( \\Delta_{x\\rightarrow y} \\). In particular, we assume the context and target image" + }, + { + "type": "image", + "bbox": [ + 0.686, + 0.602, + 0.9, + 0.751 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.683, + 0.754, + 0.902, + 0.782 + ], + "angle": 0, + "content": "Figure 3. Formulation of global anatomical layout modeling." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.791, + 0.907, + 0.868 + ], + "angle": 0, + "content": "crop share the same spatial size of \\( h \\) pixels in height and \\( w \\) pixels in width, with the left-top corner located at \\( (i_x, j_x) \\) and \\( (i_y, j_y) \\) in the pixel space. The relative position information is defined as the proportional relative displacement between the two regions:" + }, + { + "type": "equation", + "bbox": [ + 0.558, + 0.871, + 0.907, + 0.906 + ], + "angle": 0, + "content": "\\[\n\\Delta_ {x \\rightarrow y} = (\\Delta h, \\Delta w) = \\left(\\frac {i _ {y} - i _ {x}}{h}, \\frac {j _ {y} - j _ {x}}{w}\\right). \\tag {4}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "20781" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.484, + 0.213 + ], + "angle": 0, + "content": "To properly represent the location of the image patches in the target \\(y\\), we establish a coordinate system over the context image patches, with the top-left corner patch located at the origin \\((0,0)\\) and the bottom-right corner patch located at \\((N_h - 1, N_w - 1)\\), where \\(N_h, N_w\\) are the number of tokens along the height and width dimensions. As demonstrated in Figure 3, the coordinates of the image patch at the \\(u\\)-th row and \\(v\\)-th column in the target image are given by:" + }, + { + "type": "equation", + "bbox": [ + 0.093, + 0.218, + 0.484, + 0.24 + ], + "angle": 0, + "content": "\\[\n\\phi_ {x \\rightarrow y} (u, v) = (\\Delta h \\cdot N _ {h} + u, \\Delta w \\cdot N _ {w} + v) \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.171, + 0.238, + 0.452, + 0.253 + ], + "angle": 0, + "content": "where \\(u\\in [0,N_h - 1],v\\in [0,N_w - 1]\\)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.259, + 0.484, + 0.365 + ], + "angle": 0, + "content": "Essentially, \\(\\phi_{x\\to y}(u,v)\\) describes the target's location from the context's perspective, enabling the model to predict the target based on the context. Sinusoidal position encoding \\(\\mathrm{PE}(\\cdot)\\) is then applied to the relative coordinates \\(\\phi_{x\\rightarrow y}(u,v)\\), and the mask tokens carrying the positional embeddings of all the target image patches are fed to the predictor. The predictor output is given by:" + }, + { + "type": "equation", + "bbox": [ + 0.095, + 0.369, + 0.482, + 0.416 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\hat {h} _ {y} = g _ {\\phi} (h _ {x}; \\Delta_ {x \\to y}) \\\\ = g _ {\\phi} \\left(\\left(h _ {x} + p _ {x}\\right) \\oplus \\left\\{m + \\operatorname {P E} \\left(\\phi_ {x \\rightarrow y} (u, v)\\right)\\right\\} _ {u, v}\\right), \\tag {6} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.419, + 0.395, + 0.435 + ], + "angle": 0, + "content": "where \\(u\\in [0,N_h - 1],v\\in [0,N_w - 1]\\)" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.44, + 0.313, + 0.455 + ], + "angle": 0, + "content": "and the loss function is given by:" + }, + { + "type": "equation", + "bbox": [ + 0.119, + 0.46, + 0.482, + 0.481 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {g l o b a l}} (x, y) = \\left\\| g _ {\\phi} \\left(f _ {\\theta} (x); \\Delta_ {x \\rightarrow y}\\right) - f _ {\\theta^ {\\prime}} ^ {\\prime} (y) \\right\\| _ {2} ^ {2}. \\tag {7}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.487, + 0.343, + 0.503 + ], + "angle": 0, + "content": "4.3. Domain Variation Modeling" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.509, + 0.484, + 0.705 + ], + "angle": 0, + "content": "Thus far, we have discussed two levels of anatomical knowledge of human bodies. Beyond this, radiologists are usually capable of flexibly adapting to the specific characteristics of each radiograph and understanding them despite variations [15, 23, 49]. As shown in Figure 1(c), medical images often come from different sources (e.g., hospitals, techniques, equipment) and show diverse appearances [52] (e.g., clarity, contrast, exposure). Nevertheless, experienced radiologists can imagine the physical reality behind the radiographs with varying appearances, which facilitates the diagnosis process. We hypothesize that radiologists' internal world models can simulate the variations across domains, thus obtaining an objective view of the scanned body." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.484, + 0.886 + ], + "angle": 0, + "content": "Inspired by this, we propose a domain variation modeling task that learns how image features change across domains, enabling cross-domain adaptability. We simulate domain shifts using data augmentation and construct context-target pairs to model these transitions. The model is tasked with predicting the feature outcomes of the inverse effect of an image transformation determined by certain augmentation parameters, as shown in Figure 1(c). The learned representation space is designed to be equivariant [16, 17], i.e., input transformation leads to predictable output change, which preserves sufficient information across domains, mimicking radiologists' internal models." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.886, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Specifically, the target \\(y\\) is an augmented version of" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.319 + ], + "angle": 0, + "content": "the original image. Another transform parameterized by \\( k \\) scalars (denoted by \\( a \\in \\mathbb{R}^k \\)) is further applied to the target to produce the context \\( x = \\mathcal{T}_a(y) \\), where \\( \\mathcal{T} \\) is an augmentation pipeline consisting of Gaussian blur, brightness, contrast, and gamma adjustments. The parameter \\( a \\) contains the strength and other configurations of the augmentation. The world model learns to model the feature transformation from the context to the target domain (i.e., the inverse effect of \\( \\mathcal{T}_a \\)) conditioned on the parameter \\( a \\), which can be regarded as the \"action\" of domain transition. The prediction is also performed with mask tokens to be compatible with the other two world modeling tasks. The parameter \\( a \\) is combined with the mask tokens via a lightweight policy network \\( \\pi \\). The forward pass and loss function of domain variation modeling are given by:" + }, + { + "type": "equation", + "bbox": [ + 0.574, + 0.324, + 0.905, + 0.39 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} m _ {a, p} = \\pi (m + p, a), \\\\ \\hat {h} _ {y} = g _ {\\phi} \\left(h _ {x}; a\\right) \\\\ = g _ {\\phi} \\left(\\left(h _ {x} + p _ {x}\\right) \\oplus \\left\\{m _ {a, \\mathrm {P E} (u, v)} \\right\\} _ {u, v}\\right), \\tag {8} \\\\ \\end{array}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.516, + 0.392, + 0.808, + 0.413 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {d o m a i n}} (x, y) = \\left\\| g _ {\\phi} \\left(f _ {\\theta} (x); a\\right) - f _ {\\theta^ {\\prime}} ^ {\\prime} (y) \\right\\| _ {2} ^ {2},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.416, + 0.906, + 0.461 + ], + "angle": 0, + "content": "where \\( p \\) is the positional embedding of an image patch, \\( m_{a,p} \\) denotes the mask token carrying both spatial and domain transition information." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.491, + 0.735, + 0.505 + ], + "angle": 0, + "content": "4.4. The Unified Framework" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.52, + 0.906, + 0.625 + ], + "angle": 0, + "content": "Having established the individual tasks of local anatomical structure modeling, global anatomical layout modeling, and domain variation modeling, we now integrate these tasks into a cohesive system. This integration is important for creating a comprehensive model that can leverage multiple dimensions of medical knowledge simultaneously and maximize the combined benefits of each task." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.632, + 0.907, + 0.769 + ], + "angle": 0, + "content": "Instead of simply adding three loss functions together, a more elegant solution is to design unified contexts, targets, and latent variables to perform the three world modeling tasks in a single forward pass. As illustrated in Figure 1(d), the training pipeline of CheXWorld starts with sampling two different image crops from the image \\( y_{1},y_{2} \\) which serve as the targets. The targets are then augmented and masked with two different configurations \\( (a_{1},M_{1}) \\) and \\( (a_{2},M_{2}) \\), producing the contexts \\( x_{1},x_{2} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.592, + 0.774, + 0.905, + 0.79 + ], + "angle": 0, + "content": "\\[\nx _ {i} = \\operatorname {M a s k} \\left(\\mathcal {T} _ {a _ {i}} \\left(y _ {i}\\right), M _ {i}\\right), i = 1, 2. \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.906, + 0.901 + ], + "angle": 0, + "content": "The contexts \\(x_{1}, x_{2}\\) and targets \\(y_{1}, y_{2}\\) are then encoded into representations \\(h_{x_1}, h_{x_2}, h_{y_1}, h_{y_2}\\). Note that the targets \\(h_{y_1}, h_{y_2}\\) are predictable from the contexts \\(h_{x_1}, h_{x_2}\\) given the proper conditions. Specifically, since \\(x_{1}\\) is obtained by applying augmentation \\(\\mathcal{T}_{a_1}\\) and mask \\(M_{1}\\) to \\(y_{1}\\), predicting \\(y_{1}\\) from \\(x_{1}\\) with conditions \\((a_{1}, M_{1})\\) simultaneously performs local anatomical structure modeling and domain vari" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20782" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.179, + 0.094, + 0.233, + 0.145 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.246, + 0.094, + 0.285, + 0.103 + ], + "angle": 0, + "content": "Context" + }, + { + "type": "image", + "bbox": [ + 0.238, + 0.104, + 0.293, + 0.144 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.296, + 0.094, + 0.32, + 0.103 + ], + "angle": 0, + "content": "" + }, + { + "type": "image", + "bbox": [ + 0.296, + 0.104, + 0.511, + 0.145 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.352, + 0.094, + 0.456, + 0.102 + ], + "angle": 0, + "content": "World Model Prediction" + }, + { + "type": "image_caption", + "bbox": [ + 0.521, + 0.094, + 0.614, + 0.103 + ], + "angle": 0, + "content": "Target v.s. Prediction" + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.104, + 0.619, + 0.145 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.179, + 0.148, + 0.234, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.238, + 0.148, + 0.293, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.296, + 0.148, + 0.351, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.352, + 0.148, + 0.407, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.409, + 0.148, + 0.512, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.148, + 0.566, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.568, + 0.148, + 0.619, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.179, + 0.194, + 0.233, + 0.236 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.238, + 0.194, + 0.293, + 0.236 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.296, + 0.194, + 0.35, + 0.236 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.352, + 0.194, + 0.407, + 0.236 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.409, + 0.194, + 0.512, + 0.236 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.194, + 0.566, + 0.236 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.568, + 0.194, + 0.619, + 0.236 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.179, + 0.24, + 0.234, + 0.281 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.238, + 0.24, + 0.293, + 0.281 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.296, + 0.24, + 0.35, + 0.281 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.352, + 0.24, + 0.407, + 0.281 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.409, + 0.24, + 0.512, + 0.281 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.24, + 0.566, + 0.281 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.568, + 0.24, + 0.619, + 0.281 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.634, + 0.094, + 0.814, + 0.103 + ], + "angle": 0, + "content": "Context&Target World Model Prediction" + }, + { + "type": "image", + "bbox": [ + 0.641, + 0.104, + 0.816, + 0.145 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.642, + 0.148, + 0.757, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.76, + 0.147, + 0.815, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.641, + 0.194, + 0.696, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.7, + 0.194, + 0.753, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.76, + 0.194, + 0.815, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.641, + 0.239, + 0.696, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.7, + 0.239, + 0.758, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.76, + 0.239, + 0.815, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.664, + 0.284, + 0.795, + 0.294 + ], + "angle": 0, + "content": "(b) Global Layout Modeling" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.294, + 0.907, + 0.349 + ], + "angle": 0, + "content": "Figure 4. Visualization of the CheXWorld predictor outputs (zooming in for details). The images presented in this figure were not included in the pre-training of CheXWorld or the training of the diffusion model. Regions in red bounding boxes denote the predictor outputs that are mapped to pixel space using the RCDM [8] framework. In (a), gray areas indicate masked regions excluded from the context. In (b), the two overlapping regions alternately serve as context and target." + }, + { + "type": "image", + "bbox": [ + 0.142, + 0.354, + 0.218, + 0.417 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.234, + 0.355, + 0.307, + 0.417 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.324, + 0.355, + 0.399, + 0.417 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.143, + 0.42, + 0.218, + 0.449 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.234, + 0.421, + 0.308, + 0.45 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.325, + 0.421, + 0.398, + 0.45 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.092, + 0.457, + 0.446, + 0.499 + ], + "angle": 0, + "content": "Figure 5. CheXWorld prioritizes relevant medical features over spurious signals (e.g., lateral markers) in the image. Regions in red boxes denote the predictor outputs." + }, + { + "type": "image", + "bbox": [ + 0.493, + 0.353, + 0.868, + 0.442 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.454, + 0.445, + 0.905, + 0.501 + ], + "angle": 0, + "content": "Figure 6. CheXWorld learns anatomical correspondence. We compute pixel-level embeddings using RoI-pooling [22] and calculate the embedding similarity between four anatomical landmarks in a reference image and each pixel in the test image to create feature similarity heatmaps." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.513, + 0.2, + 0.528 + ], + "angle": 0, + "content": "ation modeling:" + }, + { + "type": "equation", + "bbox": [ + 0.114, + 0.533, + 0.482, + 0.585 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {1 \\rightarrow 1} = \\mathcal {L} _ {\\text {l o c a l} + \\text {d o m a i n}} \\left(x _ {1}, y _ {1}\\right) \\\\ = \\sum_ {c \\in M _ {1}} \\| g _ {\\phi} \\left(h _ {x _ {1}}; M _ {1}, a _ {1}\\right) _ {c} - f _ {\\theta^ {\\prime}} ^ {\\prime} \\left(y _ {1}\\right) _ {c} \\| _ {2} ^ {2}. \\tag {10} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.59, + 0.484, + 0.665 + ], + "angle": 0, + "content": "Moreover, \\( x_{1} \\) can be derived from \\( y_{2} \\) by moving the input window by \\( -\\Delta_{x_1\\to y_2} \\) and applying augmentation \\( \\mathcal{T}_{a_1} \\), hence the predictor conditioning on \\( (\\Delta_{x_1\\to y_2},a_1) \\) performs global anatomical structure modeling and domain variation modeling at the same time." + }, + { + "type": "equation", + "bbox": [ + 0.124, + 0.671, + 0.482, + 0.71 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {1 \\rightarrow 2} = \\mathcal {L} _ {\\text {g l o b a l} + \\text {d o m a i n}} \\left(x _ {1}, y _ {2}\\right) \\tag {11} \\\\ = \\left\\| g _ {\\phi} \\left(h _ {x _ {1}}; \\Delta_ {x _ {1} \\rightarrow y _ {2}}, a _ {1}\\right) - f _ {\\theta^ {\\prime}} ^ {\\prime} (y _ {2}) \\right\\| _ {2} ^ {2}. \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.715, + 0.484, + 0.806 + ], + "angle": 0, + "content": "Similarly, \\(\\mathcal{L}_{2\\rightarrow 2},\\mathcal{L}_{2\\rightarrow 1}\\) can be computed between \\((x_{2},y_{2})\\) and \\((x_{2},y_{1})\\), respectively. Note that the context features and target features are reused twice when computing the loss (10) and (11). Finally, the model is trained with the combination of \\(2\\times 2 = 4\\) supervisions computed from the two pairs of context and target." + }, + { + "type": "equation", + "bbox": [ + 0.16, + 0.812, + 0.482, + 0.827 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {1 \\rightarrow 1} + \\mathcal {L} _ {2 \\rightarrow 2} + \\mathcal {L} _ {1 \\rightarrow 2} + \\mathcal {L} _ {2 \\rightarrow 1}. \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.833, + 0.484, + 0.879 + ], + "angle": 0, + "content": "With this unified objective, CheXWorld showcases deep comprehension of human anatomy in different scales and pixel intensity distributions across various domains." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.511, + 0.645, + 0.528 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.537, + 0.907, + 0.627 + ], + "angle": 0, + "content": "In this section, we first demonstrate CheXWorld's world modeling capability with a series of analytical experiments. Then, we empirically compare CheXWorld with other self-supervised methods as baselines on eight downstream tasks. Finally, we provide ablation studies to highlight the effect of each building block of our method." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.628, + 0.907, + 0.703 + ], + "angle": 0, + "content": "Implementation details. CheXWorld is pre-trained on \\(\\sim 0.5\\mathrm{M}\\) frontal chest X-rays from several public datasets. The visual backbone is a ViT-Base [18]. The model is trained for 300 epochs, taking 16 hours on 8 RTX 4090 GPUs. Please refer to the appendix for more details." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.712, + 0.858, + 0.729 + ], + "angle": 0, + "content": "5.1. CheXWorld is a Strong World Predictor" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Visualization of anatomical modeling. To better understand the predictions made by CheXWorld, we train a generative model that maps the average-pooled predictor output back to pixel space following the RCDM framework [8]. Specifically, we construct context-target pairs in the same way as in the local and global anatomical modeling tasks. Then, we feed the context to a frozen pre-trained CheX-World model to obtain the model's feature prediction, on which a diffusion model conditions to produce the target image in pixel space. As shown in Figure 4(a), the CheX-World predictor perfectly recovers the appearance of bones" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "20783" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.908, + 0.119 + ], + "angle": 0, + "content": "Table 1. Performance on five downstream classification benchmarks. Accuracy is reported for RSNA, and the area under the ROC curve (AUROC) is reported for the rest of the datasets. The best two results are bold-faced and underlined, respectively." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.123, + 0.868, + 0.334 + ], + "angle": 0, + "content": "
MethodPre-train DataBackboneVinDr-CXRShenZhenChestX-ray14RSNA (CLS)CheXpert
Scratch-ViT-B70.22±1.9582.24±0.6071.69±0.3266.59±0.3980.78±0.03
MoCo-v3 [14]IN-1k (1.3M)ViT-B87.25±0.6392.85±1.0079.20±0.3072.79±0.5287.12±0.36
DINO [11]IN-1k (1.3M)ViT-B82.89±1.1090.39±4.2978.37±0.4771.27±0.4587.01±0.62
BEiT [6]IN-21k (12M)ViT-B85.93±1.9892.87±1.0879.91±0.2472.78±0.3787.77±0.38
LVM-Med [44]Medical (1.3M)ViT-B88.22±0.4494.81±1.3280.08±0.0972.75±0.4488.07±0.25
MAE [31, 63]X-rays (0.5M)ViT-B92.76±0.1897.63±0.2183.0†73.75±0.2489.3†
SimMIM [42, 64]IN-21k (12M) & X-rays (0.9M)Swin-B92.81±0.3198.09±0.1383.04±0.1574.09±0.3989.14±0.22
Adam-v2 [32]IN-21k (12M) & X-rays (0.9M)ConvNeXt-B91.46±0.3397.80†83.4†73.40±0.8888.90±0.36
CheXWorldX-rays (0.5M)ViT-B95.24±0.1398.88±0.0683.58±0.0575.03±0.3989.63±0.13
Rad-DINO‡[51]LVD (142M) & X-rays (public+private)ViT-B95.16±0.1698.20±0.1783.61±0.0874.51±0.4688.94±0.15
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.09, + 0.334, + 0.871, + 0.361 + ], + "angle": 0, + "content": "Results on ImageNet pre-trained models are adopted from [42]. \\( {}^{ \\dagger } \\) Results reported by the original authors [31, 32]. \\( {}^{ \\ddagger } \\) Rad-DINO [51] is included for reference but not for direct comparison, as it requires 2560 GPU-hours (20 times more than ours) and is trained on both public and private datasets." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.373, + 0.482, + 0.449 + ], + "angle": 0, + "content": "and tissues within various masked areas, and Figure 4(b) shows that the model makes global-level predictions that are consistent with the context. Further visualizations also show that CheXWorld filters artifacts in the image (Figure 5) and learns anatomical correspondence (Figure 6)." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.451, + 0.485, + 0.707 + ], + "angle": 0, + "content": "Awareness to domain variations. We design a domain sensitivity test to verify the model's ability to capture the uncertainty in a radiograph's appearance attributes. Specifically, we first generate a candidate set of targets \\(\\{y_i\\}_{i=1}^n\\) by applying augmentations with different configurations to the same image. Then, we construct a set of context-target-latent triplets \\(\\{(x_i, y_i, a_i)\\}\\) by applying another augmentation \\(x_i = \\mathcal{T}_{a_i}(y_i)\\). The model is asked to predict the target \\(y_i\\) given the context \\(x_i\\) and latent \\(a_i\\). For each predicted output \\(\\hat{y}_i\\), we calculate the top-k recall rate of the true target \\(y_i\\) over the entire candidate set using L2 distance2. As shown in Table 4, our model achieves an average top-5 recall of 77.67 (10 times higher than random choice), demonstrating the model's strong discriminative ability across domains. We also observe a significant drop in the recall rate when removing the domain condition \\(a\\), indicating that the model indeed makes predictions conditioned on \\(a\\)." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.72, + 0.483, + 0.751 + ], + "angle": 0, + "content": "5.2. CheXWorld Produces Transferable Representations" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.759, + 0.483, + 0.851 + ], + "angle": 0, + "content": "Setup and baselines. We compare CheXWorld with several competitive baselines across eight classification and segmentation tasks using the following datasets: VinDr-CXR [46], ShenZhen-CXR [36], NIH ChestX-ray14 [61], CheXpert [35], MedFMC-ChestDR [60], SIIM-ACR Pneumothorax [1] and RSNA Pneumonia [54]." + }, + { + "type": "table_caption", + "bbox": [ + 0.545, + 0.367, + 0.874, + 0.38 + ], + "angle": 0, + "content": "Table 4. Effectiveness of domain variation modeling." + }, + { + "type": "table", + "bbox": [ + 0.536, + 0.382, + 0.882, + 0.446 + ], + "angle": 0, + "content": "
MethodRecall@1Recall@3Recall@5
Random Choice1.564.687.81
w/o domain condition15.8733.1244.37
CheXWorld38.4965.9077.67
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.448, + 0.907, + 0.569 + ], + "angle": 0, + "content": "We attach task-specific heads (linear heads and U-Net decoders) and perform full fine-tuning on the pre-trained CheXWorld encoder. The input image resolution is fixed to \\(224^2\\). We compare CheXWorld with several self-supervised learning methods designed for general domain images [6, 11, 14, 31, 64] and medical images [32, 44] using base-scale backbones [18, 40, 41]. We use the radiology version of these methods [42, 51, 63] if available." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.57, + 0.909, + 0.795 + ], + "angle": 0, + "content": "Classification results are shown in Table 1. Our method consistently outperforms the best alternatives on all the benchmarks. Although SimMIM and Adam-v2 are initialized from ImageNet weights and pre-trained on a large dataset (0.9M X-rays) and, yet still lag behind CheXWorld with considerable margins. For instance, CheXWorld surpasses Adam-v2 by 4.04 in AUROC on VinDr-CXR and SimMIM by 0.54 on ChestX-ray14. Our approach performs comparably to Rad-DINO across all benchmarks, despite Rad-DINO requiring 20 times more computational resources and utilizing private training data. Additionally, one can observe a significant contrast in performance between models trained on photometric images (ImageNet) and models trained on medical images, highlighting the importance of in-domain transfer." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.907, + 0.871 + ], + "angle": 0, + "content": "Segmentation results are presented in Table 2 and Figure 8. CheXWorld surpasses the leading baselines in terms of dice score and produces more accurate segmentation masks, demonstrating its superior capability in both instance-level and dense prediction settings." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Few-shot learning results on the MedFMC dataset are presented in Table 2. CheXWorld demonstrates a clear ad" + }, + { + "type": "page_footnote", + "bbox": [ + 0.09, + 0.863, + 0.483, + 0.901 + ], + "angle": 0, + "content": "2In our experiments the candidate set size is fixed at \\( n = 64 \\). This process is repeated across multiple images, and the final result is obtained by averaging the outcomes." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20784" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.121, + 0.091, + 0.63, + 0.133 + ], + "angle": 0, + "content": "Table 2. Results on segmentation (left) and few-shot learning (right) tasks. The dice score and the AUROC score are reported for the segmentation and few-shot learning benchmarks respectively." + }, + { + "type": "table", + "bbox": [ + 0.126, + 0.138, + 0.63, + 0.24 + ], + "angle": 0, + "content": "
MethodSegmentationFew-shot (MedFMC-ChestDR)
SIIM-ACRRSNA (SEG)1-shot5-shot10-shot
LVM-Med [44]82.19±0.3078.47±0.2457.55±0.8166.95±0.6067.44±0.71
MAE [31, 63]83.01±0.3977.39±0.2561.31±0.2774.03±0.3075.26±0.36
SimMIM [42, 64]82.89±0.0778.31±0.4160.64±1.9272.14±0.5474.42±0.38
Adam-v2 [32]83.60±0.3378.53±0.1959.16±1.0370.28±0.3970.67±0.71
CheXWorld84.58±0.3479.02±0.2564.60±1.0075.19±0.5176.40±0.25
" + }, + { + "type": "image", + "bbox": [ + 0.648, + 0.091, + 0.865, + 0.208 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.638, + 0.211, + 0.875, + 0.239 + ], + "angle": 0, + "content": "Figure 7. Fine-tuning with \\(1\\%\\), \\(10\\%\\), and \\(100\\%\\) training data on VinDr-CXR." + }, + { + "type": "image", + "bbox": [ + 0.117, + 0.251, + 0.353, + 0.409 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.097, + 0.415, + 0.374, + 0.457 + ], + "angle": 0, + "content": "Figure 8. Visualization of segmentation masks on SIIM-ACR pneumothorax dataset. GT stands for the ground truth masks." + }, + { + "type": "table_caption", + "bbox": [ + 0.371, + 0.248, + 0.9, + 0.332 + ], + "angle": 0, + "content": "Table 3. Ablation studies of the world modeling tasks and latent variables. We report the results on VinDr-CXR and RSNA with \\(1\\%\\) proportion of training data. AUROC and accuracy are reported for the respective datasets. \"Local,\" \"Global,\" and \"Domain\" are abbreviations for local anatomical structure modeling, global anatomical layout modeling, and domain variation modeling, respectively. \"Unified\" indicates whether our unified formulation is applied, as opposed to a naive combination of losses." + }, + { + "type": "table", + "bbox": [ + 0.416, + 0.337, + 0.86, + 0.46 + ], + "angle": 0, + "content": "
World Modeling TaskLatent Variables Δx→y aUnifiedVinDr 1%RSNA 1%
Local.Global.Domain.
✓N.A.84.71±1.0065.89±0.54
✓✓✓✓88.31±0.8867.47±0.51
✓✓✓89.73±0.5068.27±1.00
✓✓✓✓✓90.53±1.0170.01±0.33
✓✓✓✓✓89.12±0.7668.37±0.50
✓✓✓✓✓89.91±0.6068.98±0.60
✓✓✓✓87.60±0.7266.56±0.39
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.47, + 0.483, + 0.531 + ], + "angle": 0, + "content": "vantage on the 1-shot learning task and consistently outperforms the second-best methods in both 5-shot and 10-shot scenarios, showcasing quick adaptation to new concepts by leveraging knowledge acquired through world modeling." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.531, + 0.484, + 0.639 + ], + "angle": 0, + "content": "Sample efficiency results. The superiority of our method is further highlighted by reducing the training data proportion. As shown in Figure 7, we conduct experiments on VinDr-CXR with \\(1\\%\\), \\(10\\%\\), and \\(100\\%\\) of the training data. Notably, our method with \\(10\\%\\) of the data significantly outperforms all baselines trained on the entire dataset, reducing the annotation cost by over 90 percent." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.645, + 0.245, + 0.661 + ], + "angle": 0, + "content": "5.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.665, + 0.483, + 0.786 + ], + "angle": 0, + "content": "Effectiveness of the world modeling tasks is validated in the upper part of Table 3. Combining local anatomical structure modeling with either of the other two world modeling tasks yields significant performance gains. The best performance is achieved when all three world modeling tasks are integrated within our unified framework, indicating that these tasks reflect different aspects of medical knowledge that are effectively captured by CheXWorld." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.786, + 0.484, + 0.877 + ], + "angle": 0, + "content": "Role of the latent variables. To verify that the latent variables (such as relative position \\(\\Delta_{x\\rightarrow y}\\) for global anatomical layout modeling and transformation parameter \\(a\\) for domain variation modeling) contribute to transfer learning performance, we test the models pre-trained without certain latent variables, shown in the bottom part of Table" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.47, + 0.907, + 0.562 + ], + "angle": 0, + "content": "3. Domain variation modeling without the condition \\( a \\) degrades into a blind image restoration task, which yields inferior results compared with the original method. Removing the condition \\( \\Delta_{x \\to y} \\) from global anatomical layout modeling leads to a sharp decrease in performance because it confounds the model with erroneous target spatial locations." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.577, + 0.634, + 0.592 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.602, + 0.907, + 0.8 + ], + "angle": 0, + "content": "In this paper, we proposed CheXWorld, a self-supervised world modeling framework for radiographic images that simultaneously encodes local and global anatomical knowledge and appearance domain variations. By integrating three world modeling tasks into a unified pipeline, our model effectively captures the spatial and domain uncertainties of the image world, as verified by qualitative and quantitative analytical experiments. Extensive transfer learning experiments on eight medical image analysis benchmarks demonstrate the state-of-the-art performance of our model. Our work offers new insights into representing and extracting knowledge from medical images, paving the way toward a general-purpose foundation model for medical vision." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.815, + 0.686, + 0.831 + ], + "angle": 0, + "content": "Acknowledgements." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.907, + 0.9 + ], + "angle": 0, + "content": "The work is supported in part by the National Key R&D Program of China under Grant 2024YFB4708200 and National Natural Science Foundation of China under Grant 62321005." + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.887, + 0.414, + 0.901 + ], + "angle": 0, + "content": "3Detailed numerical results can be found in the appendix." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "20785" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.091, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.169 + ], + "angle": 0, + "content": "[1] Zawacki Anna, Wu Carol, Shih George, Elliott Julia, Fomitchev Mikhail, Hussain Mohannad, Lakhani Paras, Culliton Phil, and Bao Shunxing. Siim-acr pneumothorax segmentation. 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.172, + 0.484, + 0.254 + ], + "angle": 0, + "content": "[2] Mahmoud Assran, Quentin Duval, Ishan Misra, Piotr Bojanowski, Pascal Vincent, Michael Rabbat, Yann LeCun, and Nicolas Ballas. Self-supervised learning from images with a joint-embedding predictive architecture. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15619-15629, 2023. 1, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.256, + 0.483, + 0.338 + ], + "angle": 0, + "content": "[3] Shekoofeh Azizi, Basil Mustafa, Fiona Ryan, Zachary Beaver, Jan Freyberg, Jonathan Deaton, Aaron Loh, Alan Karthikesalingam, Simon Kornblith, Ting Chen, et al. Big self-supervised models advance medical image classification. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3478-3488, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.34, + 0.482, + 0.408 + ], + "angle": 0, + "content": "[4] Alexei Baevski, Wei-Ning Hsu, Qiantong Xu, Arun Babu, Jiatao Gu, and Michael Auli. Data2vec: A general framework for self-supervised learning in speech, vision and language. In International Conference on Machine Learning, pages 1298-1312. PMLR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.41, + 0.483, + 0.478 + ], + "angle": 0, + "content": "[5] Alexei Baevski, Arun Babu, Wei-Ning Hsu, and Michael Auli. Efficient self-supervised learning with contextualized target representations for vision, speech and language. In International Conference on Machine Learning, pages 1416-1429. PMLR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.48, + 0.482, + 0.521 + ], + "angle": 0, + "content": "[6] Hangbo Bao, Li Dong, Songhao Piao, and Furu Wei. Beit: Bert pre-training of image transformers. arXiv preprint arXiv:2106.08254, 2021.7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.523, + 0.482, + 0.577 + ], + "angle": 0, + "content": "[7] Adrien Bardes, Quentin Garrido, Jean Ponce, Xinlei Chen, Michael Rabbat, Yann LeCun, Mido Assran, and Nicolas Ballas. V-jepa: Latent video prediction for visual representation learning. 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.579, + 0.482, + 0.632 + ], + "angle": 0, + "content": "[8] Florian Bordes, Randall Balestriero, and Pascal Vincent. High fidelity visualization of what your self-supervised representation knows about. arXiv preprint arXiv:2112.09164, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.635, + 0.481, + 0.662 + ], + "angle": 0, + "content": "[9] Arthur Earl Bryson. Applied optimal control: optimization, estimation and control. Routledge, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.664, + 0.482, + 0.704 + ], + "angle": 0, + "content": "[10] Eduardo F Camacho, Carlos Bordons, Eduardo F Camacho, and Carlos Bordons. Constrained model predictive control. Springer, 2007. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.706, + 0.482, + 0.774 + ], + "angle": 0, + "content": "[11] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.776, + 0.482, + 0.843 + ], + "angle": 0, + "content": "[12] Krishna Chaitanya, Ertunc Erdil, Neerav Karani, and Ender Konukoglu. Contrastive learning of global and local features for medical image segmentation with limited annotations. Advances in neural information processing systems, 33:12546-12558, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[13] Richard J Chen, Tong Ding, Ming Y Lu, Drew FK Williamson, Guillaume Jaume, Andrew H Song, Bowen Chen, Andrew Zhang, Daniel Shao, Muhammad Shaban, et al. Towards a general-purpose foundation model for" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.093, + 0.905, + 0.119 + ], + "angle": 0, + "content": "computational pathology. Nature Medicine, 30(3):850-862, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.123, + 0.905, + 0.178 + ], + "angle": 0, + "content": "[14] Xinlei Chen, Saining Xie, and Kaiming He. An empirical study of training self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9640-9649, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.181, + 0.905, + 0.234 + ], + "angle": 0, + "content": "[15] Peng Cui, Yang Yue, Zhijie Deng, and Jun Zhu. Confidence-based reliable learning under dual noises. Advances in Neural Information Processing Systems, 35:35116-35129, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.238, + 0.905, + 0.293 + ], + "angle": 0, + "content": "[16] Rumen Dangovski, Li Jing, Charlotte Loh, Seungwook Han, Akash Srivastava, Brian Cheung, Pulkit Agrawal, and Marin Soljacic. Equivariant contrastive learning. arXiv preprint arXiv:2111.00899, 2021. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.296, + 0.905, + 0.351 + ], + "angle": 0, + "content": "[17] Alexandre Devillers and Mathieu Lefort. Equimod: An equivariance module to improve visual instance discrimination. In International Conference on Learning Representations, 2023. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.354, + 0.905, + 0.436 + ], + "angle": 0, + "content": "[18] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.439, + 0.905, + 0.494 + ], + "angle": 0, + "content": "[19] Chaoqun Du, Yulin Wang, Shiji Song, and Gao Huang. Probabilistic contrastive learning for long-tailed visual recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.497, + 0.905, + 0.566 + ], + "angle": 0, + "content": "[20] Cong Gao, Benjamin D Killeen, Yicheng Hu, Robert B Grupp, Russell H Taylor, Mehran Armand, and Mathias Unberath. Synthetic data accelerates the development of generalizable learning-based algorithms for x-ray image analysis. Nature Machine Intelligence, 5(3):294-308, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.568, + 0.905, + 0.623 + ], + "angle": 0, + "content": "[21] Quentin Garrido, Mahmoud Assran, Nicolas Ballas, Adrien Bardes, Laurent Najman, and Yann LeCun. Learning and leveraging world models in visual representation learning. arXiv preprint arXiv:2403.00504, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.626, + 0.905, + 0.695 + ], + "angle": 0, + "content": "[22] Ross Girshick, Jeff Donahue, Trevor Darrell, and Jitendra Malik. Rich feature hierarchies for accurate object detection and semantic segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 580-587, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.698, + 0.905, + 0.739 + ], + "angle": 0, + "content": "[23] Hao Guan and Mingxia Liu. Domain adaptation for medical image analysis: a survey. IEEE Transactions on Biomedical Engineering, 69(3):1173-1185, 2021. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.742, + 0.905, + 0.783 + ], + "angle": 0, + "content": "[24] David Ha and Jürgen Schmidhuber. Recurrent world models facilitate policy evolution. Advances in neural information processing systems, 31, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.786, + 0.905, + 0.813 + ], + "angle": 0, + "content": "[25] David Ha and Jürgen Schmidhuber. World models. arXiv preprint arXiv:1803.10122, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.816, + 0.905, + 0.857 + ], + "angle": 0, + "content": "[26] Danijar Hafner, Timothy Lillicrap, Jimmy Ba, and Mohammad Norouzi. Dream to control: Learning behaviors by latent imagination. arXiv preprint arXiv:1912.01603, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.859, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[27] Danijar Hafner, Timothy Lillicrap, Mohammad Norouzi, and Jimmy Ba. Mastering atari with discrete world models. arXiv preprint arXiv:2010.02193, 2020. 2" + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "20786" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.134 + ], + "angle": 0, + "content": "[28] Danijar Hafner, Jurgis Pasukonis, Jimmy Ba, and Timothy Lillicrap. Mastering diverse domains through world models. arXiv preprint arXiv:2301.04104, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.483, + 0.207 + ], + "angle": 0, + "content": "[29] Fatemeh Haghighi, Mohammad Reza Hosseinzadeh Taher, Zongwei Zhou, Michael B Gotway, and Jianming Liang. Transferable visual words: Exploiting the semantics of anatomical patterns for self-supervised learning. IEEE transactions on medical imaging, 40(10):2857-2868, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.208, + 0.482, + 0.29 + ], + "angle": 0, + "content": "[30] Fatemeh Haghighi, Mohammad Reza Hosseinzadeh Taher, Michael B Gotway, and Jianming Liang. Dira: Discriminative, restorative, and adversarial learning for self-supervised medical image analysis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20824-20834, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.292, + 0.482, + 0.36 + ], + "angle": 0, + "content": "[31] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 3, 4, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.363, + 0.482, + 0.445 + ], + "angle": 0, + "content": "[32] Mohammad Reza Hosseinzadeh Taher, Michael Gotway, and Jianming Liang. Representing part-whole hierarchies in foundation models by learning localizability, composability, and decomposability from anatomy via self-supervision. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2024. 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.447, + 0.482, + 0.516 + ], + "angle": 0, + "content": "[33] Anthony Hu, Lloyd Russell, Hudson Yeo, Zak Murez, George Fedoseev, Alex Kendall, Jamie Shotton, and Gianluca Corrado. Gaia-1: A generative world model for autonomous driving. arXiv preprint arXiv:2309.17080, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.519, + 0.482, + 0.573 + ], + "angle": 0, + "content": "[34] Zhi Huang, Federico Bianchi, Mert Yuksekgonul, Thomas J Montine, and James Zou. A visual-language foundation model for pathology image analysis using medical twitter. Nature medicine, 29(9):2307-2316, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.576, + 0.482, + 0.658 + ], + "angle": 0, + "content": "[35] Jeremy Irvin, Pranav Rajpurkar, Michael Ko, Yifan Yu, Silviana Ciurea-Ilicus, Chris Chute, Henrik Marklund, Behzad Haghgoo, Robyn Ball, Katie Shpanskaya, et al. Chexpert: A large chest radiograph dataset with uncertainty labels and expert comparison. In Proceedings of the AAAI conference on artificial intelligence, pages 590-597, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.66, + 0.482, + 0.729 + ], + "angle": 0, + "content": "[36] Stefan Jaeger, Sema Candemir, Sameer Antani, Yi-Xiang J Wang, Pu-Xuan Lu, and George Thoma. Two public chest x-ray datasets for computer-aided screening of pulmonary diseases. Quantitative imaging in medicine and surgery, 4(6): 475, 2014. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.731, + 0.482, + 0.787 + ], + "angle": 0, + "content": "[37] Bingyi Kang, Yang Yue, Rui Lu, Zhijie Lin, Yang Zhao, Kaixin Wang, Gao Huang, and Jiashi Feng. How far is video generation from world model: A physical law perspective. arXiv preprint arXiv:2411.02385, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.789, + 0.482, + 0.857 + ], + "angle": 0, + "content": "[38] Dan Kondratyuk, Lijun Yu, Xiuye Gu, José Lezama, Jonathan Huang, Rachel Hornung, Hartwig Adam, Hassan Akbari, Yair Alon, Vighnesh Birodkar, et al. Videopoet: A large language model for zero-shot video generation. arXiv preprint arXiv:2312.14125, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.859, + 0.482, + 0.899 + ], + "angle": 0, + "content": "[39] Yann LeCun. A path towards autonomous machine intelligence version 0.9.2, 2022-06-27. Open Review, 62(1), 2022. 1, 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.162 + ], + "angle": 0, + "content": "[40] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.163, + 0.905, + 0.231 + ], + "angle": 0, + "content": "[41] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11976-11986, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.233, + 0.905, + 0.315 + ], + "angle": 0, + "content": "[42] DongAo Ma, Mohammad Reza Hosseinzadeh Taher, Jiaxuan Pang, Nahid UI Islam, Fatemeh Haghighi, Michael B Gotway, and Jianming Liang. Benchmarking and boosting transformers for medical image classification. In MICCAI Workshop on Domain Adaptation and Representation Transfer, pages 12-22. Springer, 2022. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.317, + 0.904, + 0.357 + ], + "angle": 0, + "content": "[43] Phillip Marsh. Case courtesy of phillip marsh, radiopaedia.org. https://radiopaedia.org/cases/58938, 2023. Case ID: rID: 58938. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.358, + 0.905, + 0.44 + ], + "angle": 0, + "content": "[44] Duy MH Nguyen, Hoang Nguyen, Nghiem Diep, Tan Ngoc Pham, Tri Cao, Binh Nguyen, Paul Swoboda, Nhat Ho, Shadi Albarqouni, Pengtao Xie, et al. Lvm-med: Learning large-scale self-supervised vision models for medical imaging via second-order graph matching. Advances in Neural Information Processing Systems, 36, 2024. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.442, + 0.905, + 0.508 + ], + "angle": 0, + "content": "[45] Michael Moor, Oishi Banerjee, Zahra Shakeri Hossein Abad, Harlan M Krumholz, Jure Leskovec, Eric J Topol, and Pranav Rajpurkar. Foundation models for generalist medical artificial intelligence. Nature, 616(7956):259-265, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.511, + 0.905, + 0.579 + ], + "angle": 0, + "content": "[46] Ha Q Nguyen, Khanh Lam, Linh T Le, Hieu H Pham, Dat Q Tran, Dung B Nguyen, Dung D Le, Chi M Pham, Hang TT Tong, Diep H Dinh, et al. Vindr-cxr: An open dataset of chest x-rays with radiologist's annotations. Scientific Data, 9(1):429, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.581, + 0.905, + 0.635 + ], + "angle": 0, + "content": "[47] OpenAI. Video generation models as world simulators. https://openai.com/index/video-generation-models-as-world-simulators/, 2024. Accessed: Feb, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.637, + 0.905, + 0.718 + ], + "angle": 0, + "content": "[48] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. Transactions on Machine Learning Research Journal, pages 1-31, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.72, + 0.905, + 0.787 + ], + "angle": 0, + "content": "[49] Cheng Ouyang, Chen Chen, Surui Li, Zeju Li, Chen Qin, Wenjia Bai, and Daniel Rueckert. Causality-inspired single-source domain generalization for medical image segmentation. IEEE Transactions on Medical Imaging, 42(4):1095-1106, 2022. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.79, + 0.905, + 0.872 + ], + "angle": 0, + "content": "[50] Jiaxuan Pang, Fatemeh Haghighi, DongAo Ma, Nahid Ul Islam, Mohammad Reza Hosseinzadeh Taher, Michael B Gotway, and Jianming Liang. Popar: Patch order prediction and appearance recovery for self-supervised medical image analysis. In MICCAI Workshop on Domain Adaptation and Representation Transfer, pages 77-87. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[51] Fernando Pérez-García, Harshita Sharma, Sam Bond-Taylor, Kenya Bouzid, Valentina Salvatelli, Maximilian Ilse, Shruthi" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20787" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.484, + 0.147 + ], + "angle": 0, + "content": "Bannur, Daniel C Castro, Anton Schwaighofer, Matthew P Lungren, et al. Rad-dino: Exploring scalable medical image encoders beyond text supervision. arXiv preprint arXiv:2401.10815, 2024. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.483, + 0.204 + ], + "angle": 0, + "content": "[52] Eric Postal. Why don't radiology textbooks have imperfect images? https://www.diagnosticimaging.com/view/why-dont-radiology-textbooks-have-imperfect-images, 2019. Accessed: Feb, 2024. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.205, + 0.482, + 0.258 + ], + "angle": 0, + "content": "[53] RunwayML. Introducing general world models. https://research.runwayml.com/introducing-general-world-models, 2023. Accessed: May, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.26, + 0.482, + 0.356 + ], + "angle": 0, + "content": "[54] George Shih, Carol C Wu, Safwan S Halabi, Marc D Kohli, Luciano M Prevedello, Tessa S Cook, Arjun Sharma, Judith K Amorosa, Veronica Arteaga, Maya Galperin-Aizenberg, et al. Augmenting the national institutes of health chest radiograph dataset with expert annotations of possible pneumonia. Radiology: Artificial Intelligence, 1(1): e180041, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.358, + 0.482, + 0.412 + ], + "angle": 0, + "content": "[55] Ben Sorscher, Surya Ganguli, and Haim Sompolinsky. Neural representational geometry underlies few-shot concept learning. Proceedings of the National Academy of Sciences, 119(43):e2200800119, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.414, + 0.482, + 0.469 + ], + "angle": 0, + "content": "[56] Hari Sowrirajan, Jingbo Yang, Andrew Y Ng, and Pranav Rajpurkar. Moco pretraining improves representation and transferability of chest x-ray models. In Medical Imaging with Deep Learning, pages 728-744. PMLR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.47, + 0.482, + 0.538 + ], + "angle": 0, + "content": "[57] Mohammad Reza Hosseinzadeh Taher, Fatemeh Haghighi, Michael B Gotway, and Jianming Liang. Caid: a self-supervised learning framework for empowering instance discrimination in medical imaging. Proceedings of Machine Learning Research, 3, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.54, + 0.482, + 0.636 + ], + "angle": 0, + "content": "[58] Xing Tao, Yuexiang Li, Wenhui Zhou, Kai Ma, and Yefeng Zheng. Revisiting rubik's cube: self-supervised learning with volume-wise transformation for 3d medical image segmentation. In Medical Image Computing and Computer Assisted Intervention-MICCAI 2020: 23rd International Conference, Lima, Peru, October 4-8, 2020, Proceedings, Part IV 23, pages 238-248. Springer, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.637, + 0.482, + 0.692 + ], + "angle": 0, + "content": "[59] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.693, + 0.482, + 0.761 + ], + "angle": 0, + "content": "[60] Dequan Wang, Xiaosong Wang, Lilong Wang, Mengzhang Li, Qian Da, Xiaoqiang Liu, Xiangyu Gao, Jun Shen, Junjun He, Tian Shen, et al. A real-world dataset and benchmark for foundation model adaptation in medical image classification. Scientific Data, 10(1):574, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.763, + 0.482, + 0.858 + ], + "angle": 0, + "content": "[61] Xiaosong Wang, Yifan Peng, Le Lu, Zhiyong Lu, Mohammadhadi Bagheri, and Ronald M Summers. Chestxray8: Hospital-scale chest x-ray database and benchmarks on weakly-supervised classification and localization of common thorax diseases. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2097-2106, 2017. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.86, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[62] Yulin Wang, Yang Yue, Rui Lu, Tianjiao Liu, Zhao Zhong, Shiji Song, and Gao Huang. Efficienttrain: Exploring generalized curriculum learning for training visual backbones. In" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5852-5864, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.189 + ], + "angle": 0, + "content": "[63] Junfei Xiao, Yutong Bai, Alan Yuille, and Zongwei Zhou. Delving into masked autoencoders for multi-label thorax disease classification. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 3588-3600, 2023. 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.192, + 0.907, + 0.261 + ], + "angle": 0, + "content": "[64] Zhenda Xie, Zheng Zhang, Yue Cao, Yutong Lin, Jianmin Bao, Zhuliang Yao, Qi Dai, and Han Hu. Simmim: A simple framework for masked image modeling. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9653-9663, 2022. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.263, + 0.905, + 0.331 + ], + "angle": 0, + "content": "[65] Ke Yan, Jinzheng Cai, Dakai Jin, Shun Miao, Dazhou Guo, Adam P Harrison, Youbao Tang, Jing Xiao, Jingjing Lu, and Le Lu. Sam: Self-supervised learning of pixel-wise anatomical embeddings in radiological images. IEEE Transactions on Medical Imaging, 41(10):2658-2669, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.333, + 0.905, + 0.43 + ], + "angle": 0, + "content": "[66] Hong-Yu Zhou, Shuang Yu, Cheng Bian, Yifan Hu, Kai Ma, and Yefeng Zheng. Comparing to learn: Surpassing image-genet pretraining on radiographs by comparing image representations. In Medical Image Computing and Computer-Assisted Intervention-MICCAI 2020: 23rd International Conference, Lima, Peru, October 4-8, 2020, Proceedings, Part I 23, pages 398-407. Springer, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.432, + 0.905, + 0.5 + ], + "angle": 0, + "content": "[67] Lei Zhou, Huidong Liu, Joseph Bae, Junjun He, Dimitris Samaras, and Prateek Prasanna. Self pre-training with masked autoencoders for medical image classification and segmentation. In 2023 IEEE 20th International Symposium on Biomedical Imaging (ISBI), pages 1-6. IEEE, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.502, + 0.905, + 0.583 + ], + "angle": 0, + "content": "[68] Yukun Zhou, Mark A Chia, Siegfried K Wagner, Murat S Ayhan, Dominic J Williamson, Robbert R Struyven, Timing Liu, Moucheng Xu, Mateo G Lozano, Peter Woodward-Court, et al. A foundation model for generalizable disease detection from retinal images. Nature, 622(7981):156-163, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.586, + 0.905, + 0.627 + ], + "angle": 0, + "content": "[69] Zongwei Zhou, Vatsal Sodha, Jiaxuan Pang, Michael B Gotway, and Jianming Liang. Models genesis. Medical image analysis, 67:101840, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.629, + 0.905, + 0.684 + ], + "angle": 0, + "content": "[70] Ziyu Zhou, Haozhe Luo, Jiaxuan Pang, Xiaowei Ding, Michael Gotway, and Jianming Liang. Learning anatomically consistent embedding for chest radiography. arXiv preprint arXiv:2312.00335, 2023. 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.684 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20788" + } + ] +] \ No newline at end of file diff --git a/2025/CheXWorld_ Exploring Image World Modeling for Radiograph Representation Learning/f79a692c-48ff-4b64-8cb4-11cdf9303afb_origin.pdf b/2025/CheXWorld_ Exploring Image World Modeling for Radiograph Representation Learning/f79a692c-48ff-4b64-8cb4-11cdf9303afb_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..cef00b3f258c93f139f221e7f98659ced6c3aafd --- /dev/null +++ b/2025/CheXWorld_ Exploring Image World Modeling for Radiograph Representation Learning/f79a692c-48ff-4b64-8cb4-11cdf9303afb_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e2e37baced807ad0d3a672f75c19e1d7e4e9a5c07207ace499cee06fba78a227 +size 2705617 diff --git a/2025/CheXWorld_ Exploring Image World Modeling for Radiograph Representation Learning/full.md b/2025/CheXWorld_ Exploring Image World Modeling for Radiograph Representation Learning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..b61c64e85b059dfd12c7cc250a67c490b9081292 --- /dev/null +++ b/2025/CheXWorld_ Exploring Image World Modeling for Radiograph Representation Learning/full.md @@ -0,0 +1,432 @@ +# CheXWorld: Exploring Image World Modeling for Radiograph Representation Learning + +Yang Yue $^{1*}$ Yulin Wang $^{1*}$ Chenxin Tao $^{1}$ Pan Liu $^{2}$ Shiji Song $^{1}$ Gao Huang $^{1\boxtimes}$ + +$^{1}$ Tsinghua University $^{2}$ PLA General Hospital + +yueyang22@mails.tsinghua.edu.cn, gaohuang@tsinghua.edu.cn + +# Abstract + +Humans can develop internal world models that encode common sense knowledge, telling them how the world works and predicting the consequences of their actions. This concept has emerged as a promising direction for establishing general-purpose machine-learning models in recent preliminary works, e.g., for visual representation learning. In this paper, we present CheXWorld, the first effort towards a self-supervised world model for radiographic images. Specifically, our work develops a unified framework that simultaneously models three aspects of medical knowledge essential for qualified radiologists, including 1) local anatomical structures describing the fine-grained characteristics of local tissues (e.g., architectures, shapes, and textures); 2) global anatomical layouts describing the global organization of the human body (e.g., layouts of organs and skeletons); and 3) domain variations that encourage CheXWorld to model the transitions across different appearance domains of radiographs (e.g., varying clarity, contrast, and exposure caused by collecting radiographs from different hospitals, devices, or patients). Empirically, we design tailored qualitative and quantitative analyses, revealing that CheX-World successfully captures these three dimensions of medical knowledge. Furthermore, transfer learning experiments across eight medical image classification and segmentation benchmarks showcase that CheXWorld significantly outperforms existing SSL methods and large-scale medical foundation models. Code & pre-trained models are available at https://github.com/LeapLabTHU/CheXWorld. + +# 1. Introduction + +Intelligent agents like humans learn extensive background knowledge about the world [39]. This common-sense information is embedded in the agents' internal models of the world, playing a pivotal role in their perception, learning, + +and decision-making processes by simulating the world's dynamics, telling them what is plausible or impossible, and predicting the outcomes of their actions. Consequently, these world models enable agents to acquire new concepts and skills with minimal demonstrations and trials [55]. Recent works [2, 21] have preliminarily verified the effectiveness of establishing visual representation learning approaches through the lens of world modeling. Pre-trained world models can produce semantically rich embeddings and adapt to various downstream tasks with limited data. + +Like many other research areas, the field of medical imaging is experiencing a paradigm shift from task-specific models to general-purpose foundation models [20, 34, 45, 68], which are pre-trained on massive data and expected to encode meaningful medical knowledge. The idea of world modeling offers an approach to training a medical foundation model by capturing common sense information (e.g. human anatomy) from medical images, which is a promising yet under-explored research direction. + +In this paper, we present CheXWorld, the first initiative toward world modeling for self-supervised representation learning on radiographic images. We propose and integrate three world modeling tasks that capture different dimensions of medical knowledge essential for qualified radiologists, depicted in Figure 1. In particular, a radiologist with a strong understanding of human anatomy can identify different anatomical structures and determine the relative geometry between anatomical regions. Given the hierarchical nature of human anatomy, we introduce world modeling tasks that focus on both local and global levels of anatomy: local anatomical structure modeling aims to predict specific anatomical characteristics and structures within a localized anatomical region, such as bones, airways, blood vessels, and lung segments; global anatomical layout modeling learns the overall arrangement and spatial relationships of various anatomical regions within the human body, understanding the relative positioning of organs and tissues such as the heart, lungs, diaphragm, and rib cage. Additionally, medical images are + +![](images/41433b7f88a116e6c0b7e4bf4c94a8f533c04eb500b67bb7b701d3dc200ba63d.jpg) +(a) Local Anatomical Structure Modeling + +![](images/b6145e1e58a76ea4ed64cf4907960ac6830c978549b234b7ce10b94301f58e91.jpg) +(b) Global Anatomical Layout Modeling + +![](images/c40bc4c47d56a9ab85b6869a589f38b621864ac17511376b643e827846072086.jpg) +(c) Domain Variation Modeling + +![](images/2fa63e404893b79cd981aaf0a11bec89b40fd099fe7432fd45bc0da5f5bf06e6.jpg) +(d) Unified World Modeling +Figure 1. Overview of the CheXWorld framework1. The upper part of the figure depicts three dimensions of medical knowledge that are formulated in our framework, including (a) local anatomical structures describing the fine-grained characteristics of local tissues, (b) global anatomical layouts describing the global organization of the human body and (c) domain variations that encourage CheXWorld to model the transitions across different appearance domains of radiographs. The middle part of the figure illustrates the world modeling tasks corresponding to these aspects of medical knowledge. (d) shows our unified pipeline that combines the merits of all three tasks. + +typically gathered from multiple domains with notable appearance differences due to varying equipment and acquisition techniques [52]. Experienced radiologists can recognize the physical reality behind domains with different appearances and facilitate the diagnosis process. Motivated by this, we hypothesize that radiologists quickly adapt to diverse domain changes by learning sophisticated internal world models that are capable of simulating transitions between domains. Driven by our hypothesis, we introduce the domain variation modeling task, which aims to learn an expressive feature space that encodes predictable transitions between domains. Finally, we design an integrated pipeline that performs the three world modeling tasks simultaneously, effectively incorporating medical anatomical knowledge and achieving robustness across domains. + +It is noteworthy that the three aspects we consider, i.e., local anatomical structure modeling, global anatomical layout modeling, and domain variation modeling, represent core elements of radiographic expertise, each contributing to a comprehensive framework for capturing medical knowledge. While they may not encompass every dimen + +sion of medical understanding, they are fundamental to radiology and critical for radiograph representation learning, as extensively demonstrated in [23, 49, 69, 70]. + +Empirically, we validate that CheXWorld effectively captures the three dimensions of medical knowledge through a series of analytical experiments, including visualizations of predictor outputs with the help of a generative model and a domain variation sensitivity test. Moreover, we extensively evaluate CheXWorld on eight medical image classification and segmentation benchmarks. Our model consistently outperforms competitive self-supervised learning baselines with comparable backbone capacity, pretraining data, and pre-training computational cost. + +# 2. Related Work + +World Modeling. In the machine learning community, world modeling originates from model-predictive control [9, 10] and is a common practice in reinforcement learning, where it predicts the future state of the environment based on the agent's action [24-27]. Recently, building general world models featuring profound comprehension of common sense knowledge is believed to be a crucial step towards general artificial intelligence [39, 53]. Video gener + +ation models [37] trained on vast data (e.g. SoRA[47] and VideoPoet [38]) are world models that perform predictions in input space. Another line of work makes predictions in the latent space [2, 7, 28, 33, 39], which more closely resembles the world model in the human brain. World modeling is also an effective approach for visual representation learning, where the model predicts missing parts of the visual inputs [2, 4, 5, 7, 21]. Our framework builds on this foundation, where our model acquires knowledge—such as human anatomy and domain variations—through learning to predict unobserved outcomes from specific actions. Our key contribution lies in providing insights into modeling medical knowledge from radiographic images and designing a framework that unifies three world modeling tasks. + +Self-supervised learning in medical imaging. Self-supervised learning (SSL) [19, 31, 48, 62] provides a practical solution to mitigating annotation scarcity in medical imaging. A large body of works incorporate restorative SSL tasks, which typically involve reconstructing the original data from certain image corruptions, such as random masking [63, 67, 68], pixel shuffling [69], and patch order shuffling [50, 58, 70]. Another line of works [3, 12, 51, 56, 66] adapt global or dense contrastive learning to medical imaging. TransVW[29] and DiRA [30, 57] explore combining discriminative and restorative methods. Additionally, learning from anatomy is a common topic for medical SSL. SAM [65] enforces pixel-level anatomical correspondence between images, and Adam [32] learns part-whole correspondence that reflects the hierarchy of human anatomy. Recently, SSL-based foundation models trained on largescale datasets gained significant attention, such as RET-Found [68] for retinal imaging and UNI [13] for pathology. + +Compared with existing works, CheXWorld is the first to introduce the concept of world modeling into medical SSL, capturing task-relevant medical knowledge through three tailored world modeling tasks. Unlike contrastive-based approaches such as Adam [32], CheXWorld takes a technically distinct path by constructing equivariant [16, 17] representations—where input transformations result in predictable changes within the embedding space—rather than invariant representations. This approach allows CheXWorld to better model medical data without discarding valuable information. Moreover, our unified SSL framework delivers state-of-the-art performance across eight benchmarks, demonstrating the significant potential of designing SSL methods grounded in the philosophy of world modeling. + +# 3. Basic Framework of World Modeling + +World modeling. We begin by briefly introducing the concept of world modeling [25, 39], which forms the basis for our proposed method. The primary motivation behind world modeling is to predict the unobserved parts $y$ of a world (e.g., a visual environment, an image, or a video) + +based on an observed context $x$ . This prediction can be formulated across various dimensions, such as spatial (predicting the unseen regions of the data) and temporal (foreseeing the consequence of an action). In fact, the paradigm of world modeling takes inspiration from human visual cognition. For example, when presented with the right upper lobe region of a chest radiograph as the context $x$ , the internal world of a radiologist could imagine that there are two lobes (right middle and right lower) below the visible region and two lobes on the left, forming the target $y$ . + +# A basic framework. + +Here, we introduce our basic framework of instantiating world modeling, upon which we will further discuss how to establish a radiology world model by formulating various medical + +![](images/ed454c2075b415059dbcd9c2a4cac7a970b903ad8eb4c96ac6716d65250a6627.jpg) +Figure 2. A basic framework of world modeling. + +knowledge in Section 4. In specific, our basic framework mainly follows the joint-embedding predictive architecture (JEPA) [2, 39], as shown in Figure 2. We consider two encoders: the context encoder $f_{\theta}$ and the target encoder $f_{\theta'}'$ , which encode the context $x$ and target $y$ into representations $h_x$ and $h_y$ , respectively. A predictor $g_{\phi}$ learns to predict the target embedding $h_y$ using the context embedding $h_x$ , conditioned on an additional latent variable $z$ that indicates how the target relates to the context. For example, $z$ can be the spatial locations of $y$ with respect to $x$ , or an action that leads to the transition from context to target $x \rightarrow y$ (e.g., image transformations). The latent variable $z$ carries the information that makes $y$ predictable based on $x$ , effectively modeling the inherent variation and uncertainty of the real world. The primary objective of world modeling is: + +$$ +\text {m i n i m i z e} \quad D \left(h _ {y}, \hat {h} _ {y}\right) = D \left(h _ {y}, g _ {\phi} \left(h _ {x}; z\right)\right), \tag {1} +$$ + +where $D(\cdot, \cdot)$ denotes the prediction error. Notably, the predictions are performed in the abstract representation space, contrasting with generative world models that predict every detail of the target. This mirrors an important characteristic of intelligent agents: information filtering, i.e., eliminating the irrelevant details in the data during the perception process. For example, a radiologist can infer the anatomical layout of a partly missed image, but it's nearly impossible for him to recover each pixel value of the unseen regions. + +In implementation, we set the target encoder to be the exponential moving average of the context encoder following [2, 4, 7]. Moreover, we use the Vision Transformer (ViT) [18] architecture for the context/target encoder and the predictor. For the rest of this paper, the context and target feature $h_x, h_y$ are sequences of patch embeddings produced by the ViTs. To perform patch-level feature predictions, we attach mask tokens to the context features and feed the com + +binned sequence to the predictor. The output representations of the mask tokens are used as the final prediction. + +# 4. CheXWorld: World Modeling for Radiograph Representation Learning + +This section extends the idea of world modeling to radiology. We propose three world modeling tasks tailored for radiographs and establish a unified CheXWorld framework that seamlessly integrates them. Importantly, the three tasks are designed to model three critical dimensions of medical knowledge essential for qualified radiologists. To be specific, these tasks are organized hierarchically from low-level to high-level, namely 1) local anatomical structure modeling that aims to learn the fine-grained structures of local anatomical regions, 2) global anatomical layout modeling that seeks to learn the global geometry of the human body (e.g., the layout of organs and skeletons), and 3) domain variation modeling that learns to model the transition across different appearance domains of radiographs. We first elaborate on the details of each task. Then, we present a unified framework that comprehensively combines the characteristics and merits of all three tasks, yielding a powerful foundation model that produces semantically rich and transferable representations. + +# 4.1. Local Anatomical Structure Modeling + +At a relatively low, fundamental level, various types of tissues (e.g., bones, muscles, and epithelial tissues) form the structural organization of the human body. Understanding the fine-grained characteristics (e.g., shapes, sizes, appearance, and textures) of these micro-structures is essential knowledge that radiologists must possess. To build up a learning procedure that enables neural networks to acquire such local anatomical knowledge, we consider a mask-and-reconstruction task. The model predicts fine-grained details of a masked tissue region based on its peripheral, surrounding information, as shown in Figure 1(a). This encourages the networks to develop an internal model that captures the intricate structures and local continuities of human microanatomy. The large, continuous prediction target also prevents reliance on low-level features. + +Specifically, the image mask is a union of four randomly selected rectangular regions of the image following [2]. Let $M$ denote the set of 2D patch locations that are masked from the context input. Given an input image (crop) $I$ , the image patches with locations contained in $M$ are dropped from the image to create the context $x = \operatorname{Mask}(I, M)$ , while the target is the entire image. The context and target are then fed to the corresponding encoders, producing representations $h_x = f_\theta(x)$ and $h_y = f_{\theta'}'(y)$ . Note that the context encoder exclusively processes visible patches, following [2, 31]. In the predictor, the mask tokens $m$ carrying the positional encoding of the masked locations are concatenated with the + +context feature $h_x$ , which forms the input token sequence to the predictor. The predictor output $\dot{h}_y$ is given by: + +$$ +\begin{array}{l} \hat {h} _ {y} = g _ {\phi} (h _ {x}; M) \\ = g _ {\phi} \left(\left(h _ {x} + p _ {x}\right) \oplus \{m + \operatorname {P E} (u, v) \} _ {(u, v) \in M}\right), \tag {2} \\ \end{array} +$$ + +where $\oplus$ is the concatenation operation along sequence dimension, $p_x$ is the positional embedding corresponding to the context, $u$ and $v$ are image patch location indices along height and width dimensions, $\mathrm{PE}(\cdot)$ is the sinusoidal positional encoding proposed by [59]. The prediction loss is computed only on the masked locations, which is given by: + +$$ +\mathcal {L} _ {\text {l o c a l}} (x, y) = \sum_ {c \in M} \| g _ {\phi} \left(f _ {\theta} (x); M\right) _ {c} - f _ {\theta^ {\prime}} ^ {\prime} (y) _ {c} \| _ {2} ^ {2}, \tag {3} +$$ + +where $c$ is the 2D image patch location that belongs to $M$ . + +# 4.2. Global Anatomical Layout Modeling + +In addition to understanding local microstructures like tissues, it is also important for radiologists to possess comprehensive medical knowledge regarding how groups of tissues or systems of organs are anatomically organized and assembled into the full bodies of humans. Inspired by this, we develop the global anatomical layout modeling task (see Figure 1(b)), seeking to predict the feature of an out-of-context area based on its relative position to a given context, thus facilitating understanding of the global structure of human bodies. Notably, this idea is orthogonal to the technique proposed in section 4.1 (see section 5.3 for ablation studies) as it focuses on formulating the long-range topological relationships of different tissues or organs instead of the fine-grained anatomical structures within local tissues. + +As illustrated in Figure 1(b), we randomly crop two anatomical areas from the original radiograph, serving as context $x$ and target $y$ , respectively. The world model learns to predict $y$ from $x$ given their relative position information $\Delta_{x\rightarrow y}$ . In particular, we assume the context and target image + +![](images/e1a55e97415c78eaeeef8cf47ae44dc6091f8137a742da23b4598b120c451331.jpg) +Figure 3. Formulation of global anatomical layout modeling. + +crop share the same spatial size of $h$ pixels in height and $w$ pixels in width, with the left-top corner located at $(i_x, j_x)$ and $(i_y, j_y)$ in the pixel space. The relative position information is defined as the proportional relative displacement between the two regions: + +$$ +\Delta_ {x \rightarrow y} = (\Delta h, \Delta w) = \left(\frac {i _ {y} - i _ {x}}{h}, \frac {j _ {y} - j _ {x}}{w}\right). \tag {4} +$$ + +To properly represent the location of the image patches in the target $y$ , we establish a coordinate system over the context image patches, with the top-left corner patch located at the origin $(0,0)$ and the bottom-right corner patch located at $(N_h - 1, N_w - 1)$ , where $N_h, N_w$ are the number of tokens along the height and width dimensions. As demonstrated in Figure 3, the coordinates of the image patch at the $u$ -th row and $v$ -th column in the target image are given by: + +$$ +\phi_ {x \rightarrow y} (u, v) = (\Delta h \cdot N _ {h} + u, \Delta w \cdot N _ {w} + v) \tag {5} +$$ + +where $u\in [0,N_h - 1],v\in [0,N_w - 1]$ + +Essentially, $\phi_{x\to y}(u,v)$ describes the target's location from the context's perspective, enabling the model to predict the target based on the context. Sinusoidal position encoding $\mathrm{PE}(\cdot)$ is then applied to the relative coordinates $\phi_{x\rightarrow y}(u,v)$ , and the mask tokens carrying the positional embeddings of all the target image patches are fed to the predictor. The predictor output is given by: + +$$ +\begin{array}{l} \hat {h} _ {y} = g _ {\phi} (h _ {x}; \Delta_ {x \to y}) \\ = g _ {\phi} \left(\left(h _ {x} + p _ {x}\right) \oplus \left\{m + \operatorname {P E} \left(\phi_ {x \rightarrow y} (u, v)\right)\right\} _ {u, v}\right), \tag {6} \\ \end{array} +$$ + +where $u\in [0,N_h - 1],v\in [0,N_w - 1]$ + +and the loss function is given by: + +$$ +\mathcal {L} _ {\text {g l o b a l}} (x, y) = \left\| g _ {\phi} \left(f _ {\theta} (x); \Delta_ {x \rightarrow y}\right) - f _ {\theta^ {\prime}} ^ {\prime} (y) \right\| _ {2} ^ {2}. \tag {7} +$$ + +# 4.3. Domain Variation Modeling + +Thus far, we have discussed two levels of anatomical knowledge of human bodies. Beyond this, radiologists are usually capable of flexibly adapting to the specific characteristics of each radiograph and understanding them despite variations [15, 23, 49]. As shown in Figure 1(c), medical images often come from different sources (e.g., hospitals, techniques, equipment) and show diverse appearances [52] (e.g., clarity, contrast, exposure). Nevertheless, experienced radiologists can imagine the physical reality behind the radiographs with varying appearances, which facilitates the diagnosis process. We hypothesize that radiologists' internal world models can simulate the variations across domains, thus obtaining an objective view of the scanned body. + +Inspired by this, we propose a domain variation modeling task that learns how image features change across domains, enabling cross-domain adaptability. We simulate domain shifts using data augmentation and construct context-target pairs to model these transitions. The model is tasked with predicting the feature outcomes of the inverse effect of an image transformation determined by certain augmentation parameters, as shown in Figure 1(c). The learned representation space is designed to be equivariant [16, 17], i.e., input transformation leads to predictable output change, which preserves sufficient information across domains, mimicking radiologists' internal models. + +Specifically, the target $y$ is an augmented version of + +the original image. Another transform parameterized by $k$ scalars (denoted by $a \in \mathbb{R}^k$ ) is further applied to the target to produce the context $x = \mathcal{T}_a(y)$ , where $\mathcal{T}$ is an augmentation pipeline consisting of Gaussian blur, brightness, contrast, and gamma adjustments. The parameter $a$ contains the strength and other configurations of the augmentation. The world model learns to model the feature transformation from the context to the target domain (i.e., the inverse effect of $\mathcal{T}_a$ ) conditioned on the parameter $a$ , which can be regarded as the "action" of domain transition. The prediction is also performed with mask tokens to be compatible with the other two world modeling tasks. The parameter $a$ is combined with the mask tokens via a lightweight policy network $\pi$ . The forward pass and loss function of domain variation modeling are given by: + +$$ +\begin{array}{l} m _ {a, p} = \pi (m + p, a), \\ \hat {h} _ {y} = g _ {\phi} \left(h _ {x}; a\right) \\ = g _ {\phi} \left(\left(h _ {x} + p _ {x}\right) \oplus \left\{m _ {a, \mathrm {P E} (u, v)} \right\} _ {u, v}\right), \tag {8} \\ \end{array} +$$ + +$$ +\mathcal {L} _ {\mathrm {d o m a i n}} (x, y) = \left\| g _ {\phi} \left(f _ {\theta} (x); a\right) - f _ {\theta^ {\prime}} ^ {\prime} (y) \right\| _ {2} ^ {2}, +$$ + +where $p$ is the positional embedding of an image patch, $m_{a,p}$ denotes the mask token carrying both spatial and domain transition information. + +# 4.4. The Unified Framework + +Having established the individual tasks of local anatomical structure modeling, global anatomical layout modeling, and domain variation modeling, we now integrate these tasks into a cohesive system. This integration is important for creating a comprehensive model that can leverage multiple dimensions of medical knowledge simultaneously and maximize the combined benefits of each task. + +Instead of simply adding three loss functions together, a more elegant solution is to design unified contexts, targets, and latent variables to perform the three world modeling tasks in a single forward pass. As illustrated in Figure 1(d), the training pipeline of CheXWorld starts with sampling two different image crops from the image $y_{1},y_{2}$ which serve as the targets. The targets are then augmented and masked with two different configurations $(a_{1},M_{1})$ and $(a_{2},M_{2})$ , producing the contexts $x_{1},x_{2}$ : + +$$ +x _ {i} = \operatorname {M a s k} \left(\mathcal {T} _ {a _ {i}} \left(y _ {i}\right), M _ {i}\right), i = 1, 2. \tag {9} +$$ + +The contexts $x_{1}, x_{2}$ and targets $y_{1}, y_{2}$ are then encoded into representations $h_{x_1}, h_{x_2}, h_{y_1}, h_{y_2}$ . Note that the targets $h_{y_1}, h_{y_2}$ are predictable from the contexts $h_{x_1}, h_{x_2}$ given the proper conditions. Specifically, since $x_{1}$ is obtained by applying augmentation $\mathcal{T}_{a_1}$ and mask $M_{1}$ to $y_{1}$ , predicting $y_{1}$ from $x_{1}$ with conditions $(a_{1}, M_{1})$ simultaneously performs local anatomical structure modeling and domain vari + +![](images/326a376dec9e0bf39f24f7676a0b89436ef8bbddd6c495f2e7aa7c3eb3b20aaf.jpg) + +![](images/3c393711d543ed0ddef7ca91d62b1b34bc1610251ce66e06778252f2efbd5519.jpg) +Context + +![](images/e05d0f362b2847b042066c536b5391fcd1c9256bbd54baae17c3ed9a43cd38d2.jpg) + +World Model Prediction + +![](images/4a6a87624a25db5f3084e7d5c3daa7c5a4d444d6634baf1452eb4dc387d0f251.jpg) +Target v.s. Prediction + +![](images/1578f9d8ae5ea7e7ea8f89cefe781aef84d40ef352241f8ca925dccc0d83964d.jpg) + +![](images/8b7daefee0531d89e9f371355a2141d01f1d9b99df9a6a8c0878974a68567031.jpg) + +![](images/1fe7a232935c3719ac2603fb66cc49393bb50ff7bd23b1ce1339c09aaaf94123.jpg) + +![](images/f7fdbea78ae50e67feec66900a240ecbca2281bfc61710f623a75b0d32120159.jpg) + +![](images/405a4f111d7f3da1fb17199a51746be3e0c2c06566e2c79c3981df678f649037.jpg) + +![](images/1acfa791be86c8840b8164fcc64c9abc5b3d7523715900cc91ea4a9602201e2b.jpg) + +![](images/cb54c4b6fff1665566c101437ccf6d012fc8d227e674c0adc8012c0ffe54a839.jpg) + +![](images/4e9914b5b09de9100aeaebdf3f6f44d2d312e95efd9e31917159462bc45a8a9d.jpg) + +![](images/4f9d63a1bd528c4e6f4fe5319f73e5242562c69ea8a3a7c9d0848b2b3fe4b99c.jpg) + +![](images/5b68524b0361cd794535970be23d278a2f94a5d70ff7e59bb3b41a71be1449ce.jpg) + +![](images/316738ab436ce6819a06be4f6a4cc231c005acead4ef759e1e5598e3d51e1d9b.jpg) + +![](images/b74369d4637276f88f0aff867a281e3159aecd75bb4f5e445d87404f2703e9b8.jpg) + +![](images/fac77acb0ce143ec46c2d7b6754d4d9d33b47a6f3568d33eb76902867fd6d7e8.jpg) + +![](images/17de1a0bddc331484e9b41551160d397d63b2b5cb677ddb04dd1db832dfd2e8c.jpg) + +![](images/b2e0f4c24cf93edbbb8c995c78b0b4cf8fbea4b3d50dfcb2c7bf6af8502eec0c.jpg) + +![](images/89217775023f4aabd9b313a4c9b7187fb54dd398bd8e7e158237f440c9c930be.jpg) + +![](images/d6ec51fe22edc9e09e50bac10bf179e8dfa88cd0b201882bf50a5bb947ff1adc.jpg) + +![](images/00de25365320a54a54e319665857f8a1ef07c8e7cc7855144c581c004420fba1.jpg) + +![](images/132de9a0a72ba287f220e3e8bc707bfd6b349e1a8b95fd4acb32ee58eaf5c0a7.jpg) + +![](images/8e40d3f2edd650a2a8054a19486f165ab07875cedd1dbd78afb881e8b7dadc51.jpg) + +![](images/6078cf6bd9568bf678dbdbfaa2209fb319f4dd52ad3a2ba018db042c099a8514.jpg) + +![](images/127370d9c02fe8a49b741b948cedab78c0e79676abc581127f75bf1ff0afb6f7.jpg) +Context&Target World Model Prediction + +![](images/ba3fda43c6bfbf81898b88f642c048df3780af2c95d93120ecb7688dc3bbb931.jpg) + +![](images/5b308fb86f8b761fae3d16bc3e92382890f3e0cb2b2c68c0d9c358aa9b1ca72a.jpg) + +![](images/a0dc1f2400ca8df47b7b74e261892668f7ca35427301e2ab51609fe1f8f9a16d.jpg) + +![](images/2af7c6f860936e742feaa1ffcdba3f2b981fe9c6447fe8ec105f109f7c6f419e.jpg) + +![](images/c330d2f8b5b992503bc406ad510d8af4f56ee01781e1097742c8f2813a118f8a.jpg) + +![](images/0d908be3ef7d7d9f628243f83e199333b3bcb4ff8c35103d86a12bc0f67e350c.jpg) +(b) Global Layout Modeling + +![](images/0059af19b42a4ffcbd012761f7583551f6e5aa1436a8ef33867a4fb9e6d4b377.jpg) + +![](images/b2fdf30bec8b2577924aaf4f5faec368317a8ccd61d3e1258c606e3fcb32c8a1.jpg) + +![](images/4470cd4eef2e787a3003ce98a6b5b65d6cb55f865e7b25a5be4945edc7efd31a.jpg) + +![](images/e7428427db91350638451feb75d41c54ea8cc456647fd392ddbe29f4dff0623b.jpg) + +![](images/674bdda378e763f892e1811dc2bf9b89245fc03c983577af3a8c2747c7d3f34e.jpg) + +![](images/9e6aea537c52cc755f421ea085c45ff7fe1acb68129f0c219ecacc815c20bfc0.jpg) + +![](images/e624d15008386098ae6d1b94d2d91744218561430d95970cb57e11f3d2c449e4.jpg) +Figure 5. CheXWorld prioritizes relevant medical features over spurious signals (e.g., lateral markers) in the image. Regions in red boxes denote the predictor outputs. + +![](images/e0dc553cf827785bfe5397648b2d9b8438c935de2f226aa0219ee5ef6488fb93.jpg) + +![](images/f1e9db7cb9dcd0bec2cec68238f447c7b69345564a860bca83b60abee864ab82.jpg) +Figure 4. Visualization of the CheXWorld predictor outputs (zooming in for details). The images presented in this figure were not included in the pre-training of CheXWorld or the training of the diffusion model. Regions in red bounding boxes denote the predictor outputs that are mapped to pixel space using the RCDM [8] framework. In (a), gray areas indicate masked regions excluded from the context. In (b), the two overlapping regions alternately serve as context and target. +Figure 6. CheXWorld learns anatomical correspondence. We compute pixel-level embeddings using RoI-pooling [22] and calculate the embedding similarity between four anatomical landmarks in a reference image and each pixel in the test image to create feature similarity heatmaps. + +ation modeling: + +$$ +\begin{array}{l} \mathcal {L} _ {1 \rightarrow 1} = \mathcal {L} _ {\text {l o c a l} + \text {d o m a i n}} \left(x _ {1}, y _ {1}\right) \\ = \sum_ {c \in M _ {1}} \| g _ {\phi} \left(h _ {x _ {1}}; M _ {1}, a _ {1}\right) _ {c} - f _ {\theta^ {\prime}} ^ {\prime} \left(y _ {1}\right) _ {c} \| _ {2} ^ {2}. \tag {10} \\ \end{array} +$$ + +Moreover, $x_{1}$ can be derived from $y_{2}$ by moving the input window by $-\Delta_{x_1\to y_2}$ and applying augmentation $\mathcal{T}_{a_1}$ , hence the predictor conditioning on $(\Delta_{x_1\to y_2},a_1)$ performs global anatomical structure modeling and domain variation modeling at the same time. + +$$ +\begin{array}{l} \mathcal {L} _ {1 \rightarrow 2} = \mathcal {L} _ {\text {g l o b a l} + \text {d o m a i n}} \left(x _ {1}, y _ {2}\right) \tag {11} \\ = \left\| g _ {\phi} \left(h _ {x _ {1}}; \Delta_ {x _ {1} \rightarrow y _ {2}}, a _ {1}\right) - f _ {\theta^ {\prime}} ^ {\prime} (y _ {2}) \right\| _ {2} ^ {2}. \\ \end{array} +$$ + +Similarly, $\mathcal{L}_{2\rightarrow 2},\mathcal{L}_{2\rightarrow 1}$ can be computed between $(x_{2},y_{2})$ and $(x_{2},y_{1})$ , respectively. Note that the context features and target features are reused twice when computing the loss (10) and (11). Finally, the model is trained with the combination of $2\times 2 = 4$ supervisions computed from the two pairs of context and target. + +$$ +\mathcal {L} = \mathcal {L} _ {1 \rightarrow 1} + \mathcal {L} _ {2 \rightarrow 2} + \mathcal {L} _ {1 \rightarrow 2} + \mathcal {L} _ {2 \rightarrow 1}. \tag {12} +$$ + +With this unified objective, CheXWorld showcases deep comprehension of human anatomy in different scales and pixel intensity distributions across various domains. + +# 5. Experiments + +In this section, we first demonstrate CheXWorld's world modeling capability with a series of analytical experiments. Then, we empirically compare CheXWorld with other self-supervised methods as baselines on eight downstream tasks. Finally, we provide ablation studies to highlight the effect of each building block of our method. + +Implementation details. CheXWorld is pre-trained on $\sim 0.5\mathrm{M}$ frontal chest X-rays from several public datasets. The visual backbone is a ViT-Base [18]. The model is trained for 300 epochs, taking 16 hours on 8 RTX 4090 GPUs. Please refer to the appendix for more details. + +# 5.1. CheXWorld is a Strong World Predictor + +Visualization of anatomical modeling. To better understand the predictions made by CheXWorld, we train a generative model that maps the average-pooled predictor output back to pixel space following the RCDM framework [8]. Specifically, we construct context-target pairs in the same way as in the local and global anatomical modeling tasks. Then, we feed the context to a frozen pre-trained CheX-World model to obtain the model's feature prediction, on which a diffusion model conditions to produce the target image in pixel space. As shown in Figure 4(a), the CheX-World predictor perfectly recovers the appearance of bones + +Table 1. Performance on five downstream classification benchmarks. Accuracy is reported for RSNA, and the area under the ROC curve (AUROC) is reported for the rest of the datasets. The best two results are bold-faced and underlined, respectively. + +
MethodPre-train DataBackboneVinDr-CXRShenZhenChestX-ray14RSNA (CLS)CheXpert
Scratch-ViT-B70.22±1.9582.24±0.6071.69±0.3266.59±0.3980.78±0.03
MoCo-v3 [14]IN-1k (1.3M)ViT-B87.25±0.6392.85±1.0079.20±0.3072.79±0.5287.12±0.36
DINO [11]IN-1k (1.3M)ViT-B82.89±1.1090.39±4.2978.37±0.4771.27±0.4587.01±0.62
BEiT [6]IN-21k (12M)ViT-B85.93±1.9892.87±1.0879.91±0.2472.78±0.3787.77±0.38
LVM-Med [44]Medical (1.3M)ViT-B88.22±0.4494.81±1.3280.08±0.0972.75±0.4488.07±0.25
MAE [31, 63]X-rays (0.5M)ViT-B92.76±0.1897.63±0.2183.0†73.75±0.2489.3†
SimMIM [42, 64]IN-21k (12M) & X-rays (0.9M)Swin-B92.81±0.3198.09±0.1383.04±0.1574.09±0.3989.14±0.22
Adam-v2 [32]IN-21k (12M) & X-rays (0.9M)ConvNeXt-B91.46±0.3397.80†83.4†73.40±0.8888.90±0.36
CheXWorldX-rays (0.5M)ViT-B95.24±0.1398.88±0.0683.58±0.0575.03±0.3989.63±0.13
Rad-DINO‡[51]LVD (142M) & X-rays (public+private)ViT-B95.16±0.1698.20±0.1783.61±0.0874.51±0.4688.94±0.15
+ +Results on ImageNet pre-trained models are adopted from [42]. ${}^{ \dagger }$ Results reported by the original authors [31, 32]. ${}^{ \ddagger }$ Rad-DINO [51] is included for reference but not for direct comparison, as it requires 2560 GPU-hours (20 times more than ours) and is trained on both public and private datasets. + +and tissues within various masked areas, and Figure 4(b) shows that the model makes global-level predictions that are consistent with the context. Further visualizations also show that CheXWorld filters artifacts in the image (Figure 5) and learns anatomical correspondence (Figure 6). + +Awareness to domain variations. We design a domain sensitivity test to verify the model's ability to capture the uncertainty in a radiograph's appearance attributes. Specifically, we first generate a candidate set of targets $\{y_i\}_{i=1}^n$ by applying augmentations with different configurations to the same image. Then, we construct a set of context-target-latent triplets $\{(x_i, y_i, a_i)\}$ by applying another augmentation $x_i = \mathcal{T}_{a_i}(y_i)$ . The model is asked to predict the target $y_i$ given the context $x_i$ and latent $a_i$ . For each predicted output $\hat{y}_i$ , we calculate the top-k recall rate of the true target $y_i$ over the entire candidate set using L2 distance2. As shown in Table 4, our model achieves an average top-5 recall of 77.67 (10 times higher than random choice), demonstrating the model's strong discriminative ability across domains. We also observe a significant drop in the recall rate when removing the domain condition $a$ , indicating that the model indeed makes predictions conditioned on $a$ . + +# 5.2. CheXWorld Produces Transferable Representations + +Setup and baselines. We compare CheXWorld with several competitive baselines across eight classification and segmentation tasks using the following datasets: VinDr-CXR [46], ShenZhen-CXR [36], NIH ChestX-ray14 [61], CheXpert [35], MedFMC-ChestDR [60], SIIM-ACR Pneumothorax [1] and RSNA Pneumonia [54]. + +Table 4. Effectiveness of domain variation modeling. + +
MethodRecall@1Recall@3Recall@5
Random Choice1.564.687.81
w/o domain condition15.8733.1244.37
CheXWorld38.4965.9077.67
+ +We attach task-specific heads (linear heads and U-Net decoders) and perform full fine-tuning on the pre-trained CheXWorld encoder. The input image resolution is fixed to $224^2$ . We compare CheXWorld with several self-supervised learning methods designed for general domain images [6, 11, 14, 31, 64] and medical images [32, 44] using base-scale backbones [18, 40, 41]. We use the radiology version of these methods [42, 51, 63] if available. + +Classification results are shown in Table 1. Our method consistently outperforms the best alternatives on all the benchmarks. Although SimMIM and Adam-v2 are initialized from ImageNet weights and pre-trained on a large dataset (0.9M X-rays) and, yet still lag behind CheXWorld with considerable margins. For instance, CheXWorld surpasses Adam-v2 by 4.04 in AUROC on VinDr-CXR and SimMIM by 0.54 on ChestX-ray14. Our approach performs comparably to Rad-DINO across all benchmarks, despite Rad-DINO requiring 20 times more computational resources and utilizing private training data. Additionally, one can observe a significant contrast in performance between models trained on photometric images (ImageNet) and models trained on medical images, highlighting the importance of in-domain transfer. + +Segmentation results are presented in Table 2 and Figure 8. CheXWorld surpasses the leading baselines in terms of dice score and produces more accurate segmentation masks, demonstrating its superior capability in both instance-level and dense prediction settings. + +Few-shot learning results on the MedFMC dataset are presented in Table 2. CheXWorld demonstrates a clear ad + +Table 2. Results on segmentation (left) and few-shot learning (right) tasks. The dice score and the AUROC score are reported for the segmentation and few-shot learning benchmarks respectively. + +
MethodSegmentationFew-shot (MedFMC-ChestDR)
SIIM-ACRRSNA (SEG)1-shot5-shot10-shot
LVM-Med [44]82.19±0.3078.47±0.2457.55±0.8166.95±0.6067.44±0.71
MAE [31, 63]83.01±0.3977.39±0.2561.31±0.2774.03±0.3075.26±0.36
SimMIM [42, 64]82.89±0.0778.31±0.4160.64±1.9272.14±0.5474.42±0.38
Adam-v2 [32]83.60±0.3378.53±0.1959.16±1.0370.28±0.3970.67±0.71
CheXWorld84.58±0.3479.02±0.2564.60±1.0075.19±0.5176.40±0.25
+ +![](images/aa90d4e8eb98306fc27d669e2cd72040d0b5e7a2cfbff6ccd8b9336516d91b68.jpg) +Figure 7. Fine-tuning with $1\%$ , $10\%$ , and $100\%$ training data on VinDr-CXR. + +![](images/246c7835cd0224b1b6afc476f33c92a363d0667afee956e86dd0b7963ab13b18.jpg) +Figure 8. Visualization of segmentation masks on SIIM-ACR pneumothorax dataset. GT stands for the ground truth masks. + +Table 3. Ablation studies of the world modeling tasks and latent variables. We report the results on VinDr-CXR and RSNA with $1\%$ proportion of training data. AUROC and accuracy are reported for the respective datasets. "Local," "Global," and "Domain" are abbreviations for local anatomical structure modeling, global anatomical layout modeling, and domain variation modeling, respectively. "Unified" indicates whether our unified formulation is applied, as opposed to a naive combination of losses. + +
World Modeling TaskLatent Variables Δx→y aUnifiedVinDr 1%RSNA 1%
Local.Global.Domain.
✓N.A.84.71±1.0065.89±0.54
✓✓✓✓88.31±0.8867.47±0.51
✓✓✓89.73±0.5068.27±1.00
✓✓✓✓✓90.53±1.0170.01±0.33
✓✓✓✓✓89.12±0.7668.37±0.50
✓✓✓✓✓89.91±0.6068.98±0.60
✓✓✓✓87.60±0.7266.56±0.39
+ +vantage on the 1-shot learning task and consistently outperforms the second-best methods in both 5-shot and 10-shot scenarios, showcasing quick adaptation to new concepts by leveraging knowledge acquired through world modeling. + +Sample efficiency results. The superiority of our method is further highlighted by reducing the training data proportion. As shown in Figure 7, we conduct experiments on VinDr-CXR with $1\%$ , $10\%$ , and $100\%$ of the training data. Notably, our method with $10\%$ of the data significantly outperforms all baselines trained on the entire dataset, reducing the annotation cost by over 90 percent. + +# 5.3. Ablation Study + +Effectiveness of the world modeling tasks is validated in the upper part of Table 3. Combining local anatomical structure modeling with either of the other two world modeling tasks yields significant performance gains. The best performance is achieved when all three world modeling tasks are integrated within our unified framework, indicating that these tasks reflect different aspects of medical knowledge that are effectively captured by CheXWorld. + +Role of the latent variables. To verify that the latent variables (such as relative position $\Delta_{x\rightarrow y}$ for global anatomical layout modeling and transformation parameter $a$ for domain variation modeling) contribute to transfer learning performance, we test the models pre-trained without certain latent variables, shown in the bottom part of Table + +3. Domain variation modeling without the condition $a$ degrades into a blind image restoration task, which yields inferior results compared with the original method. Removing the condition $\Delta_{x \to y}$ from global anatomical layout modeling leads to a sharp decrease in performance because it confounds the model with erroneous target spatial locations. + +# 6. Conclusion + +In this paper, we proposed CheXWorld, a self-supervised world modeling framework for radiographic images that simultaneously encodes local and global anatomical knowledge and appearance domain variations. By integrating three world modeling tasks into a unified pipeline, our model effectively captures the spatial and domain uncertainties of the image world, as verified by qualitative and quantitative analytical experiments. Extensive transfer learning experiments on eight medical image analysis benchmarks demonstrate the state-of-the-art performance of our model. Our work offers new insights into representing and extracting knowledge from medical images, paving the way toward a general-purpose foundation model for medical vision. + +# Acknowledgements. + +The work is supported in part by the National Key R&D Program of China under Grant 2024YFB4708200 and National Natural Science Foundation of China under Grant 62321005. + +# References + +[1] Zawacki Anna, Wu Carol, Shih George, Elliott Julia, Fomitchev Mikhail, Hussain Mohannad, Lakhani Paras, Culliton Phil, and Bao Shunxing. Siim-acr pneumothorax segmentation. 2019. 7 +[2] Mahmoud Assran, Quentin Duval, Ishan Misra, Piotr Bojanowski, Pascal Vincent, Michael Rabbat, Yann LeCun, and Nicolas Ballas. Self-supervised learning from images with a joint-embedding predictive architecture. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15619-15629, 2023. 1, 3, 4 +[3] Shekoofeh Azizi, Basil Mustafa, Fiona Ryan, Zachary Beaver, Jan Freyberg, Jonathan Deaton, Aaron Loh, Alan Karthikesalingam, Simon Kornblith, Ting Chen, et al. Big self-supervised models advance medical image classification. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3478-3488, 2021. 3 +[4] Alexei Baevski, Wei-Ning Hsu, Qiantong Xu, Arun Babu, Jiatao Gu, and Michael Auli. Data2vec: A general framework for self-supervised learning in speech, vision and language. In International Conference on Machine Learning, pages 1298-1312. PMLR, 2022. 3 +[5] Alexei Baevski, Arun Babu, Wei-Ning Hsu, and Michael Auli. Efficient self-supervised learning with contextualized target representations for vision, speech and language. In International Conference on Machine Learning, pages 1416-1429. PMLR, 2023. 3 +[6] Hangbo Bao, Li Dong, Songhao Piao, and Furu Wei. Beit: Bert pre-training of image transformers. arXiv preprint arXiv:2106.08254, 2021.7 +[7] Adrien Bardes, Quentin Garrido, Jean Ponce, Xinlei Chen, Michael Rabbat, Yann LeCun, Mido Assran, and Nicolas Ballas. V-jepa: Latent video prediction for visual representation learning. 2023. 3 +[8] Florian Bordes, Randall Balestriero, and Pascal Vincent. High fidelity visualization of what your self-supervised representation knows about. arXiv preprint arXiv:2112.09164, 2021. 6 +[9] Arthur Earl Bryson. Applied optimal control: optimization, estimation and control. Routledge, 2018. 2 +[10] Eduardo F Camacho, Carlos Bordons, Eduardo F Camacho, and Carlos Bordons. Constrained model predictive control. Springer, 2007. 2 +[11] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 7 +[12] Krishna Chaitanya, Ertunc Erdil, Neerav Karani, and Ender Konukoglu. Contrastive learning of global and local features for medical image segmentation with limited annotations. Advances in neural information processing systems, 33:12546-12558, 2020. 3 +[13] Richard J Chen, Tong Ding, Ming Y Lu, Drew FK Williamson, Guillaume Jaume, Andrew H Song, Bowen Chen, Andrew Zhang, Daniel Shao, Muhammad Shaban, et al. Towards a general-purpose foundation model for + +computational pathology. Nature Medicine, 30(3):850-862, 2024. 3 +[14] Xinlei Chen, Saining Xie, and Kaiming He. An empirical study of training self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9640-9649, 2021. 7 +[15] Peng Cui, Yang Yue, Zhijie Deng, and Jun Zhu. Confidence-based reliable learning under dual noises. Advances in Neural Information Processing Systems, 35:35116-35129, 2022. 5 +[16] Rumen Dangovski, Li Jing, Charlotte Loh, Seungwook Han, Akash Srivastava, Brian Cheung, Pulkit Agrawal, and Marin Soljacic. Equivariant contrastive learning. arXiv preprint arXiv:2111.00899, 2021. 3, 5 +[17] Alexandre Devillers and Mathieu Lefort. Equimod: An equivariance module to improve visual instance discrimination. In International Conference on Learning Representations, 2023. 3, 5 +[18] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 3, 6, 7 +[19] Chaoqun Du, Yulin Wang, Shiji Song, and Gao Huang. Probabilistic contrastive learning for long-tailed visual recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 3 +[20] Cong Gao, Benjamin D Killeen, Yicheng Hu, Robert B Grupp, Russell H Taylor, Mehran Armand, and Mathias Unberath. Synthetic data accelerates the development of generalizable learning-based algorithms for x-ray image analysis. Nature Machine Intelligence, 5(3):294-308, 2023. 1 +[21] Quentin Garrido, Mahmoud Assran, Nicolas Ballas, Adrien Bardes, Laurent Najman, and Yann LeCun. Learning and leveraging world models in visual representation learning. arXiv preprint arXiv:2403.00504, 2024. 1, 3 +[22] Ross Girshick, Jeff Donahue, Trevor Darrell, and Jitendra Malik. Rich feature hierarchies for accurate object detection and semantic segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 580-587, 2014. 6 +[23] Hao Guan and Mingxia Liu. Domain adaptation for medical image analysis: a survey. IEEE Transactions on Biomedical Engineering, 69(3):1173-1185, 2021. 2, 5 +[24] David Ha and Jürgen Schmidhuber. Recurrent world models facilitate policy evolution. Advances in neural information processing systems, 31, 2018. 2 +[25] David Ha and Jürgen Schmidhuber. World models. arXiv preprint arXiv:1803.10122, 2018. 3 +[26] Danijar Hafner, Timothy Lillicrap, Jimmy Ba, and Mohammad Norouzi. Dream to control: Learning behaviors by latent imagination. arXiv preprint arXiv:1912.01603, 2019. +[27] Danijar Hafner, Timothy Lillicrap, Mohammad Norouzi, and Jimmy Ba. Mastering atari with discrete world models. arXiv preprint arXiv:2010.02193, 2020. 2 + +[28] Danijar Hafner, Jurgis Pasukonis, Jimmy Ba, and Timothy Lillicrap. Mastering diverse domains through world models. arXiv preprint arXiv:2301.04104, 2023. 3 +[29] Fatemeh Haghighi, Mohammad Reza Hosseinzadeh Taher, Zongwei Zhou, Michael B Gotway, and Jianming Liang. Transferable visual words: Exploiting the semantics of anatomical patterns for self-supervised learning. IEEE transactions on medical imaging, 40(10):2857-2868, 2021. 3 +[30] Fatemeh Haghighi, Mohammad Reza Hosseinzadeh Taher, Michael B Gotway, and Jianming Liang. Dira: Discriminative, restorative, and adversarial learning for self-supervised medical image analysis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20824-20834, 2022. 3 +[31] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 3, 4, 7, 8 +[32] Mohammad Reza Hosseinzadeh Taher, Michael Gotway, and Jianming Liang. Representing part-whole hierarchies in foundation models by learning localizability, composability, and decomposability from anatomy via self-supervision. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2024. 3, 7, 8 +[33] Anthony Hu, Lloyd Russell, Hudson Yeo, Zak Murez, George Fedoseev, Alex Kendall, Jamie Shotton, and Gianluca Corrado. Gaia-1: A generative world model for autonomous driving. arXiv preprint arXiv:2309.17080, 2023. 3 +[34] Zhi Huang, Federico Bianchi, Mert Yuksekgonul, Thomas J Montine, and James Zou. A visual-language foundation model for pathology image analysis using medical twitter. Nature medicine, 29(9):2307-2316, 2023. 1 +[35] Jeremy Irvin, Pranav Rajpurkar, Michael Ko, Yifan Yu, Silviana Ciurea-Ilicus, Chris Chute, Henrik Marklund, Behzad Haghgoo, Robyn Ball, Katie Shpanskaya, et al. Chexpert: A large chest radiograph dataset with uncertainty labels and expert comparison. In Proceedings of the AAAI conference on artificial intelligence, pages 590-597, 2019. 7 +[36] Stefan Jaeger, Sema Candemir, Sameer Antani, Yi-Xiang J Wang, Pu-Xuan Lu, and George Thoma. Two public chest x-ray datasets for computer-aided screening of pulmonary diseases. Quantitative imaging in medicine and surgery, 4(6): 475, 2014. 7 +[37] Bingyi Kang, Yang Yue, Rui Lu, Zhijie Lin, Yang Zhao, Kaixin Wang, Gao Huang, and Jiashi Feng. How far is video generation from world model: A physical law perspective. arXiv preprint arXiv:2411.02385, 2024. 3 +[38] Dan Kondratyuk, Lijun Yu, Xiuye Gu, José Lezama, Jonathan Huang, Rachel Hornung, Hartwig Adam, Hassan Akbari, Yair Alon, Vighnesh Birodkar, et al. Videopoet: A large language model for zero-shot video generation. arXiv preprint arXiv:2312.14125, 2023. 3 +[39] Yann LeCun. A path towards autonomous machine intelligence version 0.9.2, 2022-06-27. Open Review, 62(1), 2022. 1, 2, 3 + +[40] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 7 +[41] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11976-11986, 2022. 7 +[42] DongAo Ma, Mohammad Reza Hosseinzadeh Taher, Jiaxuan Pang, Nahid UI Islam, Fatemeh Haghighi, Michael B Gotway, and Jianming Liang. Benchmarking and boosting transformers for medical image classification. In MICCAI Workshop on Domain Adaptation and Representation Transfer, pages 12-22. Springer, 2022. 7, 8 +[43] Phillip Marsh. Case courtesy of phillip marsh, radiopaedia.org. https://radiopaedia.org/cases/58938, 2023. Case ID: rID: 58938. 2 +[44] Duy MH Nguyen, Hoang Nguyen, Nghiem Diep, Tan Ngoc Pham, Tri Cao, Binh Nguyen, Paul Swoboda, Nhat Ho, Shadi Albarqouni, Pengtao Xie, et al. Lvm-med: Learning large-scale self-supervised vision models for medical imaging via second-order graph matching. Advances in Neural Information Processing Systems, 36, 2024. 7, 8 +[45] Michael Moor, Oishi Banerjee, Zahra Shakeri Hossein Abad, Harlan M Krumholz, Jure Leskovec, Eric J Topol, and Pranav Rajpurkar. Foundation models for generalist medical artificial intelligence. Nature, 616(7956):259-265, 2023. 1 +[46] Ha Q Nguyen, Khanh Lam, Linh T Le, Hieu H Pham, Dat Q Tran, Dung B Nguyen, Dung D Le, Chi M Pham, Hang TT Tong, Diep H Dinh, et al. Vindr-cxr: An open dataset of chest x-rays with radiologist's annotations. Scientific Data, 9(1):429, 2022. 7 +[47] OpenAI. Video generation models as world simulators. https://openai.com/index/video-generation-models-as-world-simulators/, 2024. Accessed: Feb, 2024. 3 +[48] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. Transactions on Machine Learning Research Journal, pages 1-31, 2024. 3 +[49] Cheng Ouyang, Chen Chen, Surui Li, Zeju Li, Chen Qin, Wenjia Bai, and Daniel Rueckert. Causality-inspired single-source domain generalization for medical image segmentation. IEEE Transactions on Medical Imaging, 42(4):1095-1106, 2022. 2, 5 +[50] Jiaxuan Pang, Fatemeh Haghighi, DongAo Ma, Nahid Ul Islam, Mohammad Reza Hosseinzadeh Taher, Michael B Gotway, and Jianming Liang. Popar: Patch order prediction and appearance recovery for self-supervised medical image analysis. In MICCAI Workshop on Domain Adaptation and Representation Transfer, pages 77-87. Springer, 2022. 3 +[51] Fernando Pérez-García, Harshita Sharma, Sam Bond-Taylor, Kenya Bouzid, Valentina Salvatelli, Maximilian Ilse, Shruthi + +Bannur, Daniel C Castro, Anton Schwaighofer, Matthew P Lungren, et al. Rad-dino: Exploring scalable medical image encoders beyond text supervision. arXiv preprint arXiv:2401.10815, 2024. 3, 7 +[52] Eric Postal. Why don't radiology textbooks have imperfect images? https://www.diagnosticimaging.com/view/why-dont-radiology-textbooks-have-imperfect-images, 2019. Accessed: Feb, 2024. 2, 5 +[53] RunwayML. Introducing general world models. https://research.runwayml.com/introducing-general-world-models, 2023. Accessed: May, 2023. 2 +[54] George Shih, Carol C Wu, Safwan S Halabi, Marc D Kohli, Luciano M Prevedello, Tessa S Cook, Arjun Sharma, Judith K Amorosa, Veronica Arteaga, Maya Galperin-Aizenberg, et al. Augmenting the national institutes of health chest radiograph dataset with expert annotations of possible pneumonia. Radiology: Artificial Intelligence, 1(1): e180041, 2019. 7 +[55] Ben Sorscher, Surya Ganguli, and Haim Sompolinsky. Neural representational geometry underlies few-shot concept learning. Proceedings of the National Academy of Sciences, 119(43):e2200800119, 2022. 1 +[56] Hari Sowrirajan, Jingbo Yang, Andrew Y Ng, and Pranav Rajpurkar. Moco pretraining improves representation and transferability of chest x-ray models. In Medical Imaging with Deep Learning, pages 728-744. PMLR, 2021. 3 +[57] Mohammad Reza Hosseinzadeh Taher, Fatemeh Haghighi, Michael B Gotway, and Jianming Liang. Caid: a self-supervised learning framework for empowering instance discrimination in medical imaging. Proceedings of Machine Learning Research, 3, 2022. 3 +[58] Xing Tao, Yuexiang Li, Wenhui Zhou, Kai Ma, and Yefeng Zheng. Revisiting rubik's cube: self-supervised learning with volume-wise transformation for 3d medical image segmentation. In Medical Image Computing and Computer Assisted Intervention-MICCAI 2020: 23rd International Conference, Lima, Peru, October 4-8, 2020, Proceedings, Part IV 23, pages 238-248. Springer, 2020. 3 +[59] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 4 +[60] Dequan Wang, Xiaosong Wang, Lilong Wang, Mengzhang Li, Qian Da, Xiaoqiang Liu, Xiangyu Gao, Jun Shen, Junjun He, Tian Shen, et al. A real-world dataset and benchmark for foundation model adaptation in medical image classification. Scientific Data, 10(1):574, 2023. 7 +[61] Xiaosong Wang, Yifan Peng, Le Lu, Zhiyong Lu, Mohammadhadi Bagheri, and Ronald M Summers. Chestxray8: Hospital-scale chest x-ray database and benchmarks on weakly-supervised classification and localization of common thorax diseases. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2097-2106, 2017. 2, 7 +[62] Yulin Wang, Yang Yue, Rui Lu, Tianjiao Liu, Zhao Zhong, Shiji Song, and Gao Huang. Efficienttrain: Exploring generalized curriculum learning for training visual backbones. In + +Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5852-5864, 2023. 3 +[63] Junfei Xiao, Yutong Bai, Alan Yuille, and Zongwei Zhou. Delving into masked autoencoders for multi-label thorax disease classification. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 3588-3600, 2023. 3, 7, 8 +[64] Zhenda Xie, Zheng Zhang, Yue Cao, Yutong Lin, Jianmin Bao, Zhuliang Yao, Qi Dai, and Han Hu. Simmim: A simple framework for masked image modeling. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9653-9663, 2022. 7, 8 +[65] Ke Yan, Jinzheng Cai, Dakai Jin, Shun Miao, Dazhou Guo, Adam P Harrison, Youbao Tang, Jing Xiao, Jingjing Lu, and Le Lu. Sam: Self-supervised learning of pixel-wise anatomical embeddings in radiological images. IEEE Transactions on Medical Imaging, 41(10):2658-2669, 2022. 3 +[66] Hong-Yu Zhou, Shuang Yu, Cheng Bian, Yifan Hu, Kai Ma, and Yefeng Zheng. Comparing to learn: Surpassing image-genet pretraining on radiographs by comparing image representations. In Medical Image Computing and Computer-Assisted Intervention-MICCAI 2020: 23rd International Conference, Lima, Peru, October 4-8, 2020, Proceedings, Part I 23, pages 398-407. Springer, 2020. 3 +[67] Lei Zhou, Huidong Liu, Joseph Bae, Junjun He, Dimitris Samaras, and Prateek Prasanna. Self pre-training with masked autoencoders for medical image classification and segmentation. In 2023 IEEE 20th International Symposium on Biomedical Imaging (ISBI), pages 1-6. IEEE, 2023. 3 +[68] Yukun Zhou, Mark A Chia, Siegfried K Wagner, Murat S Ayhan, Dominic J Williamson, Robbert R Struyven, Timing Liu, Moucheng Xu, Mateo G Lozano, Peter Woodward-Court, et al. A foundation model for generalizable disease detection from retinal images. Nature, 622(7981):156-163, 2023. 1, 3 +[69] Zongwei Zhou, Vatsal Sodha, Jiaxuan Pang, Michael B Gotway, and Jianming Liang. Models genesis. Medical image analysis, 67:101840, 2021. 2, 3 +[70] Ziyu Zhou, Haozhe Luo, Jiaxuan Pang, Xiaowei Ding, Michael Gotway, and Jianming Liang. Learning anatomically consistent embedding for chest radiography. arXiv preprint arXiv:2312.00335, 2023. 2, 3 \ No newline at end of file diff --git a/2025/CheXWorld_ Exploring Image World Modeling for Radiograph Representation Learning/images.zip b/2025/CheXWorld_ Exploring Image World Modeling for Radiograph Representation Learning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..208415da04e4256e77256ac8901de5a75a07c0b7 --- /dev/null +++ b/2025/CheXWorld_ Exploring Image World Modeling for Radiograph Representation Learning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:304161aa2b9a76b36e9e927575d94088cf1ef94ec8fa67733e3e4e33607d336a +size 582162 diff --git a/2025/CheXWorld_ Exploring Image World Modeling for Radiograph Representation Learning/layout.json b/2025/CheXWorld_ Exploring Image World Modeling for Radiograph Representation Learning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..13facd38629ef3007ef84eb2b1210c6497cf39c5 --- /dev/null +++ b/2025/CheXWorld_ Exploring Image World Modeling for Radiograph Representation Learning/layout.json @@ -0,0 +1,12156 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 159, + 103, + 455, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 103, + 455, + 140 + ], + "spans": [ + { + "bbox": [ + 159, + 103, + 455, + 140 + ], + "type": "text", + "content": "CheXWorld: Exploring Image World Modeling for Radiograph Representation Learning" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 94, + 167, + 517, + 182 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 167, + 517, + 182 + ], + "spans": [ + { + "bbox": [ + 94, + 167, + 517, + 182 + ], + "type": "text", + "content": "Yang Yue" + }, + { + "bbox": [ + 94, + 167, + 517, + 182 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 94, + 167, + 517, + 182 + ], + "type": "text", + "content": " Yulin Wang" + }, + { + "bbox": [ + 94, + 167, + 517, + 182 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 94, + 167, + 517, + 182 + ], + "type": "text", + "content": " Chenxin Tao" + }, + { + "bbox": [ + 94, + 167, + 517, + 182 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 94, + 167, + 517, + 182 + ], + "type": "text", + "content": " Pan Liu" + }, + { + "bbox": [ + 94, + 167, + 517, + 182 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 94, + 167, + 517, + 182 + ], + "type": "text", + "content": " Shiji Song" + }, + { + "bbox": [ + 94, + 167, + 517, + 182 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 94, + 167, + 517, + 182 + ], + "type": "text", + "content": " Gao Huang" + }, + { + "bbox": [ + 94, + 167, + 517, + 182 + ], + "type": "inline_equation", + "content": "^{1\\boxtimes}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 217, + 183, + 395, + 196 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 217, + 183, + 395, + 196 + ], + "spans": [ + { + "bbox": [ + 217, + 183, + 395, + 196 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 217, + 183, + 395, + 196 + ], + "type": "text", + "content": "Tsinghua University " + }, + { + "bbox": [ + 217, + 183, + 395, + 196 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 217, + 183, + 395, + 196 + ], + "type": "text", + "content": "PLA General Hospital" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 148, + 196, + 459, + 208 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 196, + 459, + 208 + ], + "spans": [ + { + "bbox": [ + 148, + 196, + 459, + 208 + ], + "type": "text", + "content": "yueyang22@mails.tsinghua.edu.cn, gaohuang@tsinghua.edu.cn" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 151, + 235, + 200, + 248 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 235, + 200, + 248 + ], + "spans": [ + { + "bbox": [ + 151, + 235, + 200, + 248 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 261, + 297, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 261, + 297, + 597 + ], + "spans": [ + { + "bbox": [ + 55, + 261, + 297, + 597 + ], + "type": "text", + "content": "Humans can develop internal world models that encode common sense knowledge, telling them how the world works and predicting the consequences of their actions. This concept has emerged as a promising direction for establishing general-purpose machine-learning models in recent preliminary works, e.g., for visual representation learning. In this paper, we present CheXWorld, the first effort towards a self-supervised world model for radiographic images. Specifically, our work develops a unified framework that simultaneously models three aspects of medical knowledge essential for qualified radiologists, including 1) local anatomical structures describing the fine-grained characteristics of local tissues (e.g., architectures, shapes, and textures); 2) global anatomical layouts describing the global organization of the human body (e.g., layouts of organs and skeletons); and 3) domain variations that encourage CheXWorld to model the transitions across different appearance domains of radiographs (e.g., varying clarity, contrast, and exposure caused by collecting radiographs from different hospitals, devices, or patients). Empirically, we design tailored qualitative and quantitative analyses, revealing that CheX-World successfully captures these three dimensions of medical knowledge. Furthermore, transfer learning experiments across eight medical image classification and segmentation benchmarks showcase that CheXWorld significantly outperforms existing SSL methods and large-scale medical foundation models. Code & pre-trained models are available at https://github.com/LeapLabTHU/CheXWorld." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 621, + 137, + 634 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 621, + 137, + 634 + ], + "spans": [ + { + "bbox": [ + 56, + 621, + 137, + 634 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 642, + 296, + 691 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 691 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 691 + ], + "type": "text", + "content": "Intelligent agents like humans learn extensive background knowledge about the world [39]. This common-sense information is embedded in the agents' internal models of the world, playing a pivotal role in their perception, learning," + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 236, + 555, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 236, + 555, + 357 + ], + "spans": [ + { + "bbox": [ + 313, + 236, + 555, + 357 + ], + "type": "text", + "content": "and decision-making processes by simulating the world's dynamics, telling them what is plausible or impossible, and predicting the outcomes of their actions. Consequently, these world models enable agents to acquire new concepts and skills with minimal demonstrations and trials [55]. Recent works [2, 21] have preliminarily verified the effectiveness of establishing visual representation learning approaches through the lens of world modeling. Pre-trained world models can produce semantically rich embeddings and adapt to various downstream tasks with limited data." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 361, + 556, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 361, + 556, + 469 + ], + "spans": [ + { + "bbox": [ + 313, + 361, + 556, + 469 + ], + "type": "text", + "content": "Like many other research areas, the field of medical imaging is experiencing a paradigm shift from task-specific models to general-purpose foundation models [20, 34, 45, 68], which are pre-trained on massive data and expected to encode meaningful medical knowledge. The idea of world modeling offers an approach to training a medical foundation model by capturing common sense information (e.g. human anatomy) from medical images, which is a promising yet under-explored research direction." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 474, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 474, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 474, + 556, + 715 + ], + "type": "text", + "content": "In this paper, we present CheXWorld, the first initiative toward world modeling for self-supervised representation learning on radiographic images. We propose and integrate three world modeling tasks that capture different dimensions of medical knowledge essential for qualified radiologists, depicted in Figure 1. In particular, a radiologist with a strong understanding of human anatomy can identify different anatomical structures and determine the relative geometry between anatomical regions. Given the hierarchical nature of human anatomy, we introduce world modeling tasks that focus on both local and global levels of anatomy: local anatomical structure modeling aims to predict specific anatomical characteristics and structures within a localized anatomical region, such as bones, airways, blood vessels, and lung segments; global anatomical layout modeling learns the overall arrangement and spatial relationships of various anatomical regions within the human body, understanding the relative positioning of organs and tissues such as the heart, lungs, diaphragm, and rib cage. Additionally, medical images are" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 701, + 135, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 701, + 135, + 713 + ], + "spans": [ + { + "bbox": [ + 66, + 701, + 135, + 713 + ], + "type": "text", + "content": "*Equal contribution." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 161, + 701, + 245, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 161, + 701, + 245, + 713 + ], + "spans": [ + { + "bbox": [ + 161, + 701, + 245, + 713 + ], + "type": "text", + "content": "Corresponding authors." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20778" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 72, + 76, + 226, + 247 + ], + "blocks": [ + { + "bbox": [ + 72, + 76, + 226, + 247 + ], + "lines": [ + { + "bbox": [ + 72, + 76, + 226, + 247 + ], + "spans": [ + { + "bbox": [ + 72, + 76, + 226, + 247 + ], + "type": "image", + "image_path": "41433b7f88a116e6c0b7e4bf4c94a8f533c04eb500b67bb7b701d3dc200ba63d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 80, + 251, + 220, + 261 + ], + "lines": [ + { + "bbox": [ + 80, + 251, + 220, + 261 + ], + "spans": [ + { + "bbox": [ + 80, + 251, + 220, + 261 + ], + "type": "text", + "content": "(a) Local Anatomical Structure Modeling" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 230, + 76, + 385, + 245 + ], + "blocks": [ + { + "bbox": [ + 230, + 76, + 385, + 245 + ], + "lines": [ + { + "bbox": [ + 230, + 76, + 385, + 245 + ], + "spans": [ + { + "bbox": [ + 230, + 76, + 385, + 245 + ], + "type": "image", + "image_path": "b6145e1e58a76ea4ed64cf4907960ac6830c978549b234b7ce10b94301f58e91.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 238, + 251, + 375, + 261 + ], + "lines": [ + { + "bbox": [ + 238, + 251, + 375, + 261 + ], + "spans": [ + { + "bbox": [ + 238, + 251, + 375, + 261 + ], + "type": "text", + "content": "(b) Global Anatomical Layout Modeling" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 389, + 72, + 539, + 249 + ], + "blocks": [ + { + "bbox": [ + 389, + 72, + 539, + 249 + ], + "lines": [ + { + "bbox": [ + 389, + 72, + 539, + 249 + ], + "spans": [ + { + "bbox": [ + 389, + 72, + 539, + 249 + ], + "type": "image", + "image_path": "c40bc4c47d56a9ab85b6869a589f38b621864ac17511376b643e827846072086.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 408, + 251, + 514, + 261 + ], + "lines": [ + { + "bbox": [ + 408, + 251, + 514, + 261 + ], + "spans": [ + { + "bbox": [ + 408, + 251, + 514, + 261 + ], + "type": "text", + "content": "(c) Domain Variation Modeling" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 70, + 270, + 530, + 361 + ], + "blocks": [ + { + "bbox": [ + 70, + 270, + 530, + 361 + ], + "lines": [ + { + "bbox": [ + 70, + 270, + 530, + 361 + ], + "spans": [ + { + "bbox": [ + 70, + 270, + 530, + 361 + ], + "type": "image", + "image_path": "2fa63e404893b79cd981aaf0a11bec89b40fd099fe7432fd45bc0da5f5bf06e6.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 208, + 364, + 303, + 373 + ], + "lines": [ + { + "bbox": [ + 208, + 364, + 303, + 373 + ], + "spans": [ + { + "bbox": [ + 208, + 364, + 303, + 373 + ], + "type": "text", + "content": "(d) Unified World Modeling" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 54, + 373, + 555, + 430 + ], + "lines": [ + { + "bbox": [ + 54, + 373, + 555, + 430 + ], + "spans": [ + { + "bbox": [ + 54, + 373, + 555, + 430 + ], + "type": "text", + "content": "Figure 1. Overview of the CheXWorld framework1. The upper part of the figure depicts three dimensions of medical knowledge that are formulated in our framework, including (a) local anatomical structures describing the fine-grained characteristics of local tissues, (b) global anatomical layouts describing the global organization of the human body and (c) domain variations that encourage CheXWorld to model the transitions across different appearance domains of radiographs. The middle part of the figure illustrates the world modeling tasks corresponding to these aspects of medical knowledge. (d) shows our unified pipeline that combines the merits of all three tasks." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 433, + 295, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 433, + 295, + 612 + ], + "spans": [ + { + "bbox": [ + 54, + 433, + 295, + 612 + ], + "type": "text", + "content": "typically gathered from multiple domains with notable appearance differences due to varying equipment and acquisition techniques [52]. Experienced radiologists can recognize the physical reality behind domains with different appearances and facilitate the diagnosis process. Motivated by this, we hypothesize that radiologists quickly adapt to diverse domain changes by learning sophisticated internal world models that are capable of simulating transitions between domains. Driven by our hypothesis, we introduce the domain variation modeling task, which aims to learn an expressive feature space that encodes predictable transitions between domains. Finally, we design an integrated pipeline that performs the three world modeling tasks simultaneously, effectively incorporating medical anatomical knowledge and achieving robustness across domains." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 54, + 613, + 295, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 613, + 295, + 685 + ], + "spans": [ + { + "bbox": [ + 54, + 613, + 295, + 685 + ], + "type": "text", + "content": "It is noteworthy that the three aspects we consider, i.e., local anatomical structure modeling, global anatomical layout modeling, and domain variation modeling, represent core elements of radiographic expertise, each contributing to a comprehensive framework for capturing medical knowledge. While they may not encompass every dimen" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 433, + 555, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 433, + 555, + 468 + ], + "spans": [ + { + "bbox": [ + 313, + 433, + 555, + 468 + ], + "type": "text", + "content": "sion of medical understanding, they are fundamental to radiology and critical for radiograph representation learning, as extensively demonstrated in [23, 49, 69, 70]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 312, + 469, + 556, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 469, + 556, + 589 + ], + "spans": [ + { + "bbox": [ + 312, + 469, + 556, + 589 + ], + "type": "text", + "content": "Empirically, we validate that CheXWorld effectively captures the three dimensions of medical knowledge through a series of analytical experiments, including visualizations of predictor outputs with the help of a generative model and a domain variation sensitivity test. Moreover, we extensively evaluate CheXWorld on eight medical image classification and segmentation benchmarks. Our model consistently outperforms competitive self-supervised learning baselines with comparable backbone capacity, pretraining data, and pre-training computational cost." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 597, + 402, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 597, + 402, + 609 + ], + "spans": [ + { + "bbox": [ + 313, + 597, + 402, + 609 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "text", + "content": "World Modeling. In the machine learning community, world modeling originates from model-predictive control [9, 10] and is a common practice in reinforcement learning, where it predicts the future state of the environment based on the agent's action [24-27]. Recently, building general world models featuring profound comprehension of common sense knowledge is believed to be a crucial step towards general artificial intelligence [39, 53]. Video gener" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 55, + 693, + 295, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 693, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 693, + 295, + 714 + ], + "type": "text", + "content": "1The x-rays are taken from kenhub.com, Radiopedia [43] and ChestXray14 [61]." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20779" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 239 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 239 + ], + "type": "text", + "content": "ation models [37] trained on vast data (e.g. SoRA[47] and VideoPoet [38]) are world models that perform predictions in input space. Another line of work makes predictions in the latent space [2, 7, 28, 33, 39], which more closely resembles the world model in the human brain. World modeling is also an effective approach for visual representation learning, where the model predicts missing parts of the visual inputs [2, 4, 5, 7, 21]. Our framework builds on this foundation, where our model acquires knowledge—such as human anatomy and domain variations—through learning to predict unobserved outcomes from specific actions. Our key contribution lies in providing insights into modeling medical knowledge from radiographic images and designing a framework that unifies three world modeling tasks." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 239, + 295, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 239, + 295, + 455 + ], + "spans": [ + { + "bbox": [ + 55, + 239, + 295, + 455 + ], + "type": "text", + "content": "Self-supervised learning in medical imaging. Self-supervised learning (SSL) [19, 31, 48, 62] provides a practical solution to mitigating annotation scarcity in medical imaging. A large body of works incorporate restorative SSL tasks, which typically involve reconstructing the original data from certain image corruptions, such as random masking [63, 67, 68], pixel shuffling [69], and patch order shuffling [50, 58, 70]. Another line of works [3, 12, 51, 56, 66] adapt global or dense contrastive learning to medical imaging. TransVW[29] and DiRA [30, 57] explore combining discriminative and restorative methods. Additionally, learning from anatomy is a common topic for medical SSL. SAM [65] enforces pixel-level anatomical correspondence between images, and Adam [32] learns part-whole correspondence that reflects the hierarchy of human anatomy. Recently, SSL-based foundation models trained on largescale datasets gained significant attention, such as RET-Found [68] for retinal imaging and UNI [13] for pathology." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 456, + 295, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 456, + 295, + 622 + ], + "spans": [ + { + "bbox": [ + 55, + 456, + 295, + 622 + ], + "type": "text", + "content": "Compared with existing works, CheXWorld is the first to introduce the concept of world modeling into medical SSL, capturing task-relevant medical knowledge through three tailored world modeling tasks. Unlike contrastive-based approaches such as Adam [32], CheXWorld takes a technically distinct path by constructing equivariant [16, 17] representations—where input transformations result in predictable changes within the embedding space—rather than invariant representations. This approach allows CheXWorld to better model medical data without discarding valuable information. Moreover, our unified SSL framework delivers state-of-the-art performance across eight benchmarks, demonstrating the significant potential of designing SSL methods grounded in the philosophy of world modeling." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 633, + 261, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 633, + 261, + 647 + ], + "spans": [ + { + "bbox": [ + 55, + 633, + 261, + 647 + ], + "type": "text", + "content": "3. Basic Framework of World Modeling" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 653, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 295, + 714 + ], + "type": "text", + "content": "World modeling. We begin by briefly introducing the concept of world modeling [25, 39], which forms the basis for our proposed method. The primary motivation behind world modeling is to predict the unobserved parts " + }, + { + "bbox": [ + 55, + 653, + 295, + 714 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 55, + 653, + 295, + 714 + ], + "type": "text", + "content": " of a world (e.g., a visual environment, an image, or a video)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 553, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 192 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 192 + ], + "type": "text", + "content": "based on an observed context " + }, + { + "bbox": [ + 313, + 72, + 553, + 192 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 72, + 553, + 192 + ], + "type": "text", + "content": ". This prediction can be formulated across various dimensions, such as spatial (predicting the unseen regions of the data) and temporal (foreseeing the consequence of an action). In fact, the paradigm of world modeling takes inspiration from human visual cognition. For example, when presented with the right upper lobe region of a chest radiograph as the context " + }, + { + "bbox": [ + 313, + 72, + 553, + 192 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 72, + 553, + 192 + ], + "type": "text", + "content": ", the internal world of a radiologist could imagine that there are two lobes (right middle and right lower) below the visible region and two lobes on the left, forming the target " + }, + { + "bbox": [ + 313, + 72, + 553, + 192 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 313, + 72, + 553, + 192 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 328, + 197, + 413, + 207 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 197, + 413, + 207 + ], + "spans": [ + { + "bbox": [ + 328, + 197, + 413, + 207 + ], + "type": "text", + "content": "A basic framework." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 208, + 414, + 303 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 208, + 414, + 303 + ], + "spans": [ + { + "bbox": [ + 313, + 208, + 414, + 303 + ], + "type": "text", + "content": "Here, we introduce our basic framework of instantiating world modeling, upon which we will further discuss how to establish a radiology world model by formulating various medical" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 419, + 203, + 550, + 275 + ], + "blocks": [ + { + "bbox": [ + 419, + 203, + 550, + 275 + ], + "lines": [ + { + "bbox": [ + 419, + 203, + 550, + 275 + ], + "spans": [ + { + "bbox": [ + 419, + 203, + 550, + 275 + ], + "type": "image", + "image_path": "ed454c2075b415059dbcd9c2a4cac7a970b903ad8eb4c96ac6716d65250a6627.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 418, + 277, + 552, + 300 + ], + "lines": [ + { + "bbox": [ + 418, + 277, + 552, + 300 + ], + "spans": [ + { + "bbox": [ + 418, + 277, + 552, + 300 + ], + "type": "text", + "content": "Figure 2. A basic framework of world modeling." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 304, + 553, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 304, + 553, + 483 + ], + "spans": [ + { + "bbox": [ + 313, + 304, + 553, + 483 + ], + "type": "text", + "content": "knowledge in Section 4. In specific, our basic framework mainly follows the joint-embedding predictive architecture (JEPA) [2, 39], as shown in Figure 2. We consider two encoders: the context encoder " + }, + { + "bbox": [ + 313, + 304, + 553, + 483 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 313, + 304, + 553, + 483 + ], + "type": "text", + "content": " and the target encoder " + }, + { + "bbox": [ + 313, + 304, + 553, + 483 + ], + "type": "inline_equation", + "content": "f_{\\theta'}'" + }, + { + "bbox": [ + 313, + 304, + 553, + 483 + ], + "type": "text", + "content": ", which encode the context " + }, + { + "bbox": [ + 313, + 304, + 553, + 483 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 304, + 553, + 483 + ], + "type": "text", + "content": " and target " + }, + { + "bbox": [ + 313, + 304, + 553, + 483 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 313, + 304, + 553, + 483 + ], + "type": "text", + "content": " into representations " + }, + { + "bbox": [ + 313, + 304, + 553, + 483 + ], + "type": "inline_equation", + "content": "h_x" + }, + { + "bbox": [ + 313, + 304, + 553, + 483 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 304, + 553, + 483 + ], + "type": "inline_equation", + "content": "h_y" + }, + { + "bbox": [ + 313, + 304, + 553, + 483 + ], + "type": "text", + "content": ", respectively. A predictor " + }, + { + "bbox": [ + 313, + 304, + 553, + 483 + ], + "type": "inline_equation", + "content": "g_{\\phi}" + }, + { + "bbox": [ + 313, + 304, + 553, + 483 + ], + "type": "text", + "content": " learns to predict the target embedding " + }, + { + "bbox": [ + 313, + 304, + 553, + 483 + ], + "type": "inline_equation", + "content": "h_y" + }, + { + "bbox": [ + 313, + 304, + 553, + 483 + ], + "type": "text", + "content": " using the context embedding " + }, + { + "bbox": [ + 313, + 304, + 553, + 483 + ], + "type": "inline_equation", + "content": "h_x" + }, + { + "bbox": [ + 313, + 304, + 553, + 483 + ], + "type": "text", + "content": ", conditioned on an additional latent variable " + }, + { + "bbox": [ + 313, + 304, + 553, + 483 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 313, + 304, + 553, + 483 + ], + "type": "text", + "content": " that indicates how the target relates to the context. For example, " + }, + { + "bbox": [ + 313, + 304, + 553, + 483 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 313, + 304, + 553, + 483 + ], + "type": "text", + "content": " can be the spatial locations of " + }, + { + "bbox": [ + 313, + 304, + 553, + 483 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 313, + 304, + 553, + 483 + ], + "type": "text", + "content": " with respect to " + }, + { + "bbox": [ + 313, + 304, + 553, + 483 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 304, + 553, + 483 + ], + "type": "text", + "content": ", or an action that leads to the transition from context to target " + }, + { + "bbox": [ + 313, + 304, + 553, + 483 + ], + "type": "inline_equation", + "content": "x \\rightarrow y" + }, + { + "bbox": [ + 313, + 304, + 553, + 483 + ], + "type": "text", + "content": " (e.g., image transformations). The latent variable " + }, + { + "bbox": [ + 313, + 304, + 553, + 483 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 313, + 304, + 553, + 483 + ], + "type": "text", + "content": " carries the information that makes " + }, + { + "bbox": [ + 313, + 304, + 553, + 483 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 313, + 304, + 553, + 483 + ], + "type": "text", + "content": " predictable based on " + }, + { + "bbox": [ + 313, + 304, + 553, + 483 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 304, + 553, + 483 + ], + "type": "text", + "content": ", effectively modeling the inherent variation and uncertainty of the real world. The primary objective of world modeling is:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 344, + 487, + 553, + 502 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 344, + 487, + 553, + 502 + ], + "spans": [ + { + "bbox": [ + 344, + 487, + 553, + 502 + ], + "type": "interline_equation", + "content": "\\text {m i n i m i z e} \\quad D \\left(h _ {y}, \\hat {h} _ {y}\\right) = D \\left(h _ {y}, g _ {\\phi} \\left(h _ {x}; z\\right)\\right), \\tag {1}", + "image_path": "5b6abd8bcaca2d6e455dc96ba56ede525f93e8a3562bd07de47e8e1372d3f0be.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 506, + 554, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 506, + 554, + 613 + ], + "spans": [ + { + "bbox": [ + 313, + 506, + 554, + 613 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 506, + 554, + 613 + ], + "type": "inline_equation", + "content": "D(\\cdot, \\cdot)" + }, + { + "bbox": [ + 313, + 506, + 554, + 613 + ], + "type": "text", + "content": " denotes the prediction error. Notably, the predictions are performed in the abstract representation space, contrasting with generative world models that predict every detail of the target. This mirrors an important characteristic of intelligent agents: information filtering, i.e., eliminating the irrelevant details in the data during the perception process. For example, a radiologist can infer the anatomical layout of a partly missed image, but it's nearly impossible for him to recover each pixel value of the unseen regions." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 618, + 554, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 618, + 554, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 618, + 554, + 712 + ], + "type": "text", + "content": "In implementation, we set the target encoder to be the exponential moving average of the context encoder following [2, 4, 7]. Moreover, we use the Vision Transformer (ViT) [18] architecture for the context/target encoder and the predictor. For the rest of this paper, the context and target feature " + }, + { + "bbox": [ + 313, + 618, + 554, + 712 + ], + "type": "inline_equation", + "content": "h_x, h_y" + }, + { + "bbox": [ + 313, + 618, + 554, + 712 + ], + "type": "text", + "content": " are sequences of patch embeddings produced by the ViTs. To perform patch-level feature predictions, we attach mask tokens to the context features and feed the com" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20780" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "text", + "content": "binned sequence to the predictor. The output representations of the mask tokens are used as the final prediction." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 105, + 295, + 135 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 105, + 295, + 135 + ], + "spans": [ + { + "bbox": [ + 55, + 105, + 295, + 135 + ], + "type": "text", + "content": "4. CheXWorld: World Modeling for Radiograph Representation Learning" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 140, + 296, + 368 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 140, + 296, + 368 + ], + "spans": [ + { + "bbox": [ + 55, + 140, + 296, + 368 + ], + "type": "text", + "content": "This section extends the idea of world modeling to radiology. We propose three world modeling tasks tailored for radiographs and establish a unified CheXWorld framework that seamlessly integrates them. Importantly, the three tasks are designed to model three critical dimensions of medical knowledge essential for qualified radiologists. To be specific, these tasks are organized hierarchically from low-level to high-level, namely 1) local anatomical structure modeling that aims to learn the fine-grained structures of local anatomical regions, 2) global anatomical layout modeling that seeks to learn the global geometry of the human body (e.g., the layout of organs and skeletons), and 3) domain variation modeling that learns to model the transition across different appearance domains of radiographs. We first elaborate on the details of each task. Then, we present a unified framework that comprehensively combines the characteristics and merits of all three tasks, yielding a powerful foundation model that produces semantically rich and transferable representations." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 373, + 257, + 387 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 373, + 257, + 387 + ], + "spans": [ + { + "bbox": [ + 55, + 373, + 257, + 387 + ], + "type": "text", + "content": "4.1. Local Anatomical Structure Modeling" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 391, + 296, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 391, + 296, + 569 + ], + "spans": [ + { + "bbox": [ + 55, + 391, + 296, + 569 + ], + "type": "text", + "content": "At a relatively low, fundamental level, various types of tissues (e.g., bones, muscles, and epithelial tissues) form the structural organization of the human body. Understanding the fine-grained characteristics (e.g., shapes, sizes, appearance, and textures) of these micro-structures is essential knowledge that radiologists must possess. To build up a learning procedure that enables neural networks to acquire such local anatomical knowledge, we consider a mask-and-reconstruction task. The model predicts fine-grained details of a masked tissue region based on its peripheral, surrounding information, as shown in Figure 1(a). This encourages the networks to develop an internal model that captures the intricate structures and local continuities of human microanatomy. The large, continuous prediction target also prevents reliance on low-level features." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "content": "Specifically, the image mask is a union of four randomly selected rectangular regions of the image following [2]. Let " + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "content": " denote the set of 2D patch locations that are masked from the context input. Given an input image (crop) " + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "content": ", the image patches with locations contained in " + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "content": " are dropped from the image to create the context " + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "inline_equation", + "content": "x = \\operatorname{Mask}(I, M)" + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "content": ", while the target is the entire image. The context and target are then fed to the corresponding encoders, producing representations " + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "inline_equation", + "content": "h_x = f_\\theta(x)" + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "inline_equation", + "content": "h_y = f_{\\theta'}'(y)" + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "content": ". Note that the context encoder exclusively processes visible patches, following [2, 31]. In the predictor, the mask tokens " + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "content": " carrying the positional encoding of the masked locations are concatenated with the" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 555, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 97 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 97 + ], + "type": "text", + "content": "context feature " + }, + { + "bbox": [ + 313, + 72, + 555, + 97 + ], + "type": "inline_equation", + "content": "h_x" + }, + { + "bbox": [ + 313, + 72, + 555, + 97 + ], + "type": "text", + "content": ", which forms the input token sequence to the predictor. The predictor output " + }, + { + "bbox": [ + 313, + 72, + 555, + 97 + ], + "type": "inline_equation", + "content": "\\dot{h}_y" + }, + { + "bbox": [ + 313, + 72, + 555, + 97 + ], + "type": "text", + "content": " is given by:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 323, + 101, + 553, + 138 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 101, + 553, + 138 + ], + "spans": [ + { + "bbox": [ + 323, + 101, + 553, + 138 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\hat {h} _ {y} = g _ {\\phi} (h _ {x}; M) \\\\ = g _ {\\phi} \\left(\\left(h _ {x} + p _ {x}\\right) \\oplus \\{m + \\operatorname {P E} (u, v) \\} _ {(u, v) \\in M}\\right), \\tag {2} \\\\ \\end{array}", + "image_path": "962154815edce2da1ef530a012c458663872d077efab990b8c686b0b6703463a.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 141, + 555, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 141, + 555, + 213 + ], + "spans": [ + { + "bbox": [ + 313, + 141, + 555, + 213 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 141, + 555, + 213 + ], + "type": "inline_equation", + "content": "\\oplus" + }, + { + "bbox": [ + 313, + 141, + 555, + 213 + ], + "type": "text", + "content": " is the concatenation operation along sequence dimension, " + }, + { + "bbox": [ + 313, + 141, + 555, + 213 + ], + "type": "inline_equation", + "content": "p_x" + }, + { + "bbox": [ + 313, + 141, + 555, + 213 + ], + "type": "text", + "content": " is the positional embedding corresponding to the context, " + }, + { + "bbox": [ + 313, + 141, + 555, + 213 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 313, + 141, + 555, + 213 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 141, + 555, + 213 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 313, + 141, + 555, + 213 + ], + "type": "text", + "content": " are image patch location indices along height and width dimensions, " + }, + { + "bbox": [ + 313, + 141, + 555, + 213 + ], + "type": "inline_equation", + "content": "\\mathrm{PE}(\\cdot)" + }, + { + "bbox": [ + 313, + 141, + 555, + 213 + ], + "type": "text", + "content": " is the sinusoidal positional encoding proposed by [59]. The prediction loss is computed only on the masked locations, which is given by:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 327, + 226, + 555, + 253 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 226, + 555, + 253 + ], + "spans": [ + { + "bbox": [ + 327, + 226, + 555, + 253 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {l o c a l}} (x, y) = \\sum_ {c \\in M} \\| g _ {\\phi} \\left(f _ {\\theta} (x); M\\right) _ {c} - f _ {\\theta^ {\\prime}} ^ {\\prime} (y) _ {c} \\| _ {2} ^ {2}, \\tag {3}", + "image_path": "52ee16bee947512115497bba4ce1a6e96bd8ea46a7ecd9c0707b5f0346cc15cb.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 255, + 550, + 268 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 255, + 550, + 268 + ], + "spans": [ + { + "bbox": [ + 313, + 255, + 550, + 268 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 255, + 550, + 268 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 255, + 550, + 268 + ], + "type": "text", + "content": " is the 2D image patch location that belongs to " + }, + { + "bbox": [ + 313, + 255, + 550, + 268 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 255, + 550, + 268 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 281, + 509, + 295 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 281, + 509, + 295 + ], + "spans": [ + { + "bbox": [ + 313, + 281, + 509, + 295 + ], + "type": "text", + "content": "4.2. Global Anatomical Layout Modeling" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 301, + 555, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 301, + 555, + 469 + ], + "spans": [ + { + "bbox": [ + 313, + 301, + 555, + 469 + ], + "type": "text", + "content": "In addition to understanding local microstructures like tissues, it is also important for radiologists to possess comprehensive medical knowledge regarding how groups of tissues or systems of organs are anatomically organized and assembled into the full bodies of humans. Inspired by this, we develop the global anatomical layout modeling task (see Figure 1(b)), seeking to predict the feature of an out-of-context area based on its relative position to a given context, thus facilitating understanding of the global structure of human bodies. Notably, this idea is orthogonal to the technique proposed in section 4.1 (see section 5.3 for ablation studies) as it focuses on formulating the long-range topological relationships of different tissues or organs instead of the fine-grained anatomical structures within local tissues." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 471, + 415, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 471, + 415, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 471, + 415, + 628 + ], + "type": "text", + "content": "As illustrated in Figure 1(b), we randomly crop two anatomical areas from the original radiograph, serving as context " + }, + { + "bbox": [ + 313, + 471, + 415, + 628 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 471, + 415, + 628 + ], + "type": "text", + "content": " and target " + }, + { + "bbox": [ + 313, + 471, + 415, + 628 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 313, + 471, + 415, + 628 + ], + "type": "text", + "content": ", respectively. The world model learns to predict " + }, + { + "bbox": [ + 313, + 471, + 415, + 628 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 313, + 471, + 415, + 628 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 313, + 471, + 415, + 628 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 471, + 415, + 628 + ], + "type": "text", + "content": " given their relative position information " + }, + { + "bbox": [ + 313, + 471, + 415, + 628 + ], + "type": "inline_equation", + "content": "\\Delta_{x\\rightarrow y}" + }, + { + "bbox": [ + 313, + 471, + 415, + 628 + ], + "type": "text", + "content": ". In particular, we assume the context and target image" + } + ] + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 419, + 476, + 550, + 594 + ], + "blocks": [ + { + "bbox": [ + 419, + 476, + 550, + 594 + ], + "lines": [ + { + "bbox": [ + 419, + 476, + 550, + 594 + ], + "spans": [ + { + "bbox": [ + 419, + 476, + 550, + 594 + ], + "type": "image", + "image_path": "e1a55e97415c78eaeeef8cf47ae44dc6091f8137a742da23b4598b120c451331.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 417, + 597, + 552, + 619 + ], + "lines": [ + { + "bbox": [ + 417, + 597, + 552, + 619 + ], + "spans": [ + { + "bbox": [ + 417, + 597, + 552, + 619 + ], + "type": "text", + "content": "Figure 3. Formulation of global anatomical layout modeling." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 626, + 555, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 626, + 555, + 687 + ], + "spans": [ + { + "bbox": [ + 313, + 626, + 555, + 687 + ], + "type": "text", + "content": "crop share the same spatial size of " + }, + { + "bbox": [ + 313, + 626, + 555, + 687 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 313, + 626, + 555, + 687 + ], + "type": "text", + "content": " pixels in height and " + }, + { + "bbox": [ + 313, + 626, + 555, + 687 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 313, + 626, + 555, + 687 + ], + "type": "text", + "content": " pixels in width, with the left-top corner located at " + }, + { + "bbox": [ + 313, + 626, + 555, + 687 + ], + "type": "inline_equation", + "content": "(i_x, j_x)" + }, + { + "bbox": [ + 313, + 626, + 555, + 687 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 626, + 555, + 687 + ], + "type": "inline_equation", + "content": "(i_y, j_y)" + }, + { + "bbox": [ + 313, + 626, + 555, + 687 + ], + "type": "text", + "content": " in the pixel space. The relative position information is defined as the proportional relative displacement between the two regions:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 341, + 689, + 555, + 717 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 341, + 689, + 555, + 717 + ], + "spans": [ + { + "bbox": [ + 341, + 689, + 555, + 717 + ], + "type": "interline_equation", + "content": "\\Delta_ {x \\rightarrow y} = (\\Delta h, \\Delta w) = \\left(\\frac {i _ {y} - i _ {x}}{h}, \\frac {j _ {y} - j _ {x}}{w}\\right). \\tag {4}", + "image_path": "e0109f7651974cf5f48af38123d777c913666cb27df42205383d76cd06381df4.jpg" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20781" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "text", + "content": "To properly represent the location of the image patches in the target " + }, + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "text", + "content": ", we establish a coordinate system over the context image patches, with the top-left corner patch located at the origin " + }, + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "inline_equation", + "content": "(0,0)" + }, + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "text", + "content": " and the bottom-right corner patch located at " + }, + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "inline_equation", + "content": "(N_h - 1, N_w - 1)" + }, + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "inline_equation", + "content": "N_h, N_w" + }, + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "text", + "content": " are the number of tokens along the height and width dimensions. As demonstrated in Figure 3, the coordinates of the image patch at the " + }, + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "text", + "content": "-th row and " + }, + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "text", + "content": "-th column in the target image are given by:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 172, + 296, + 190 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 172, + 296, + 190 + ], + "spans": [ + { + "bbox": [ + 56, + 172, + 296, + 190 + ], + "type": "interline_equation", + "content": "\\phi_ {x \\rightarrow y} (u, v) = (\\Delta h \\cdot N _ {h} + u, \\Delta w \\cdot N _ {w} + v) \\tag {5}", + "image_path": "83e32177c558d54d98a4ec9bc995c27ca194ba0f9fb22d6d0272b692e808804b.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 104, + 188, + 276, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 188, + 276, + 200 + ], + "spans": [ + { + "bbox": [ + 104, + 188, + 276, + 200 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 104, + 188, + 276, + 200 + ], + "type": "inline_equation", + "content": "u\\in [0,N_h - 1],v\\in [0,N_w - 1]" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 205, + 296, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 205, + 296, + 289 + ], + "spans": [ + { + "bbox": [ + 55, + 205, + 296, + 289 + ], + "type": "text", + "content": "Essentially, " + }, + { + "bbox": [ + 55, + 205, + 296, + 289 + ], + "type": "inline_equation", + "content": "\\phi_{x\\to y}(u,v)" + }, + { + "bbox": [ + 55, + 205, + 296, + 289 + ], + "type": "text", + "content": " describes the target's location from the context's perspective, enabling the model to predict the target based on the context. Sinusoidal position encoding " + }, + { + "bbox": [ + 55, + 205, + 296, + 289 + ], + "type": "inline_equation", + "content": "\\mathrm{PE}(\\cdot)" + }, + { + "bbox": [ + 55, + 205, + 296, + 289 + ], + "type": "text", + "content": " is then applied to the relative coordinates " + }, + { + "bbox": [ + 55, + 205, + 296, + 289 + ], + "type": "inline_equation", + "content": "\\phi_{x\\rightarrow y}(u,v)" + }, + { + "bbox": [ + 55, + 205, + 296, + 289 + ], + "type": "text", + "content": ", and the mask tokens carrying the positional embeddings of all the target image patches are fed to the predictor. The predictor output is given by:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 58, + 292, + 294, + 329 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 292, + 294, + 329 + ], + "spans": [ + { + "bbox": [ + 58, + 292, + 294, + 329 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\hat {h} _ {y} = g _ {\\phi} (h _ {x}; \\Delta_ {x \\to y}) \\\\ = g _ {\\phi} \\left(\\left(h _ {x} + p _ {x}\\right) \\oplus \\left\\{m + \\operatorname {P E} \\left(\\phi_ {x \\rightarrow y} (u, v)\\right)\\right\\} _ {u, v}\\right), \\tag {6} \\\\ \\end{array}", + "image_path": "b9108e0941ec6ad931767628bbad16189e6ba85ef2a354113d9e921e11c3dfaa.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 69, + 331, + 241, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 331, + 241, + 344 + ], + "spans": [ + { + "bbox": [ + 69, + 331, + 241, + 344 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 69, + 331, + 241, + 344 + ], + "type": "inline_equation", + "content": "u\\in [0,N_h - 1],v\\in [0,N_w - 1]" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 348, + 191, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 348, + 191, + 360 + ], + "spans": [ + { + "bbox": [ + 55, + 348, + 191, + 360 + ], + "type": "text", + "content": "and the loss function is given by:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 72, + 364, + 294, + 380 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 364, + 294, + 380 + ], + "spans": [ + { + "bbox": [ + 72, + 364, + 294, + 380 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {g l o b a l}} (x, y) = \\left\\| g _ {\\phi} \\left(f _ {\\theta} (x); \\Delta_ {x \\rightarrow y}\\right) - f _ {\\theta^ {\\prime}} ^ {\\prime} (y) \\right\\| _ {2} ^ {2}. \\tag {7}", + "image_path": "d5e72e7db5489718167134bf21dd7766070651f12a35fd03777c0c6128cc1c1f.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 385, + 209, + 398 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 385, + 209, + 398 + ], + "spans": [ + { + "bbox": [ + 55, + 385, + 209, + 398 + ], + "type": "text", + "content": "4.3. Domain Variation Modeling" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 403, + 296, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 403, + 296, + 558 + ], + "spans": [ + { + "bbox": [ + 55, + 403, + 296, + 558 + ], + "type": "text", + "content": "Thus far, we have discussed two levels of anatomical knowledge of human bodies. Beyond this, radiologists are usually capable of flexibly adapting to the specific characteristics of each radiograph and understanding them despite variations [15, 23, 49]. As shown in Figure 1(c), medical images often come from different sources (e.g., hospitals, techniques, equipment) and show diverse appearances [52] (e.g., clarity, contrast, exposure). Nevertheless, experienced radiologists can imagine the physical reality behind the radiographs with varying appearances, which facilitates the diagnosis process. We hypothesize that radiologists' internal world models can simulate the variations across domains, thus obtaining an objective view of the scanned body." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 558, + 296, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 296, + 701 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 296, + 701 + ], + "type": "text", + "content": "Inspired by this, we propose a domain variation modeling task that learns how image features change across domains, enabling cross-domain adaptability. We simulate domain shifts using data augmentation and construct context-target pairs to model these transitions. The model is tasked with predicting the feature outcomes of the inverse effect of an image transformation determined by certain augmentation parameters, as shown in Figure 1(c). The learned representation space is designed to be equivariant [16, 17], i.e., input transformation leads to predictable output change, which preserves sufficient information across domains, mimicking radiologists' internal models." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 67, + 701, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 701, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 701, + 296, + 713 + ], + "type": "text", + "content": "Specifically, the target " + }, + { + "bbox": [ + 67, + 701, + 296, + 713 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 67, + 701, + 296, + 713 + ], + "type": "text", + "content": " is an augmented version of" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 72, + 555, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 252 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 252 + ], + "type": "text", + "content": "the original image. Another transform parameterized by " + }, + { + "bbox": [ + 313, + 72, + 555, + 252 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 72, + 555, + 252 + ], + "type": "text", + "content": " scalars (denoted by " + }, + { + "bbox": [ + 313, + 72, + 555, + 252 + ], + "type": "inline_equation", + "content": "a \\in \\mathbb{R}^k" + }, + { + "bbox": [ + 313, + 72, + 555, + 252 + ], + "type": "text", + "content": ") is further applied to the target to produce the context " + }, + { + "bbox": [ + 313, + 72, + 555, + 252 + ], + "type": "inline_equation", + "content": "x = \\mathcal{T}_a(y)" + }, + { + "bbox": [ + 313, + 72, + 555, + 252 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 72, + 555, + 252 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 313, + 72, + 555, + 252 + ], + "type": "text", + "content": " is an augmentation pipeline consisting of Gaussian blur, brightness, contrast, and gamma adjustments. The parameter " + }, + { + "bbox": [ + 313, + 72, + 555, + 252 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 313, + 72, + 555, + 252 + ], + "type": "text", + "content": " contains the strength and other configurations of the augmentation. The world model learns to model the feature transformation from the context to the target domain (i.e., the inverse effect of " + }, + { + "bbox": [ + 313, + 72, + 555, + 252 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_a" + }, + { + "bbox": [ + 313, + 72, + 555, + 252 + ], + "type": "text", + "content": ") conditioned on the parameter " + }, + { + "bbox": [ + 313, + 72, + 555, + 252 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 313, + 72, + 555, + 252 + ], + "type": "text", + "content": ", which can be regarded as the \"action\" of domain transition. The prediction is also performed with mask tokens to be compatible with the other two world modeling tasks. The parameter " + }, + { + "bbox": [ + 313, + 72, + 555, + 252 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 313, + 72, + 555, + 252 + ], + "type": "text", + "content": " is combined with the mask tokens via a lightweight policy network " + }, + { + "bbox": [ + 313, + 72, + 555, + 252 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 313, + 72, + 555, + 252 + ], + "type": "text", + "content": ". The forward pass and loss function of domain variation modeling are given by:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 351, + 256, + 553, + 308 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 256, + 553, + 308 + ], + "spans": [ + { + "bbox": [ + 351, + 256, + 553, + 308 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} m _ {a, p} = \\pi (m + p, a), \\\\ \\hat {h} _ {y} = g _ {\\phi} \\left(h _ {x}; a\\right) \\\\ = g _ {\\phi} \\left(\\left(h _ {x} + p _ {x}\\right) \\oplus \\left\\{m _ {a, \\mathrm {P E} (u, v)} \\right\\} _ {u, v}\\right), \\tag {8} \\\\ \\end{array}", + "image_path": "8b0e04865b459c3bc3f976eb39b4dc3129457bf39c054bfd63527e7aeb27d5a6.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 315, + 310, + 494, + 327 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 310, + 494, + 327 + ], + "spans": [ + { + "bbox": [ + 315, + 310, + 494, + 327 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {d o m a i n}} (x, y) = \\left\\| g _ {\\phi} \\left(f _ {\\theta} (x); a\\right) - f _ {\\theta^ {\\prime}} ^ {\\prime} (y) \\right\\| _ {2} ^ {2},", + "image_path": "61a11630eb4e71dbe7940fa31b396b4350c0d1f87b4a4b8bd9be083427087ddd.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 329, + 554, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 329, + 554, + 365 + ], + "spans": [ + { + "bbox": [ + 313, + 329, + 554, + 365 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 329, + 554, + 365 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 313, + 329, + 554, + 365 + ], + "type": "text", + "content": " is the positional embedding of an image patch, " + }, + { + "bbox": [ + 313, + 329, + 554, + 365 + ], + "type": "inline_equation", + "content": "m_{a,p}" + }, + { + "bbox": [ + 313, + 329, + 554, + 365 + ], + "type": "text", + "content": " denotes the mask token carrying both spatial and domain transition information." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 314, + 388, + 449, + 399 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 388, + 449, + 399 + ], + "spans": [ + { + "bbox": [ + 314, + 388, + 449, + 399 + ], + "type": "text", + "content": "4.4. The Unified Framework" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 411, + 554, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 411, + 554, + 495 + ], + "spans": [ + { + "bbox": [ + 313, + 411, + 554, + 495 + ], + "type": "text", + "content": "Having established the individual tasks of local anatomical structure modeling, global anatomical layout modeling, and domain variation modeling, we now integrate these tasks into a cohesive system. This integration is important for creating a comprehensive model that can leverage multiple dimensions of medical knowledge simultaneously and maximize the combined benefits of each task." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 500, + 555, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 500, + 555, + 609 + ], + "spans": [ + { + "bbox": [ + 313, + 500, + 555, + 609 + ], + "type": "text", + "content": "Instead of simply adding three loss functions together, a more elegant solution is to design unified contexts, targets, and latent variables to perform the three world modeling tasks in a single forward pass. As illustrated in Figure 1(d), the training pipeline of CheXWorld starts with sampling two different image crops from the image " + }, + { + "bbox": [ + 313, + 500, + 555, + 609 + ], + "type": "inline_equation", + "content": "y_{1},y_{2}" + }, + { + "bbox": [ + 313, + 500, + 555, + 609 + ], + "type": "text", + "content": " which serve as the targets. The targets are then augmented and masked with two different configurations " + }, + { + "bbox": [ + 313, + 500, + 555, + 609 + ], + "type": "inline_equation", + "content": "(a_{1},M_{1})" + }, + { + "bbox": [ + 313, + 500, + 555, + 609 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 500, + 555, + 609 + ], + "type": "inline_equation", + "content": "(a_{2},M_{2})" + }, + { + "bbox": [ + 313, + 500, + 555, + 609 + ], + "type": "text", + "content": ", producing the contexts " + }, + { + "bbox": [ + 313, + 500, + 555, + 609 + ], + "type": "inline_equation", + "content": "x_{1},x_{2}" + }, + { + "bbox": [ + 313, + 500, + 555, + 609 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 362, + 613, + 553, + 625 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 362, + 613, + 553, + 625 + ], + "spans": [ + { + "bbox": [ + 362, + 613, + 553, + 625 + ], + "type": "interline_equation", + "content": "x _ {i} = \\operatorname {M a s k} \\left(\\mathcal {T} _ {a _ {i}} \\left(y _ {i}\\right), M _ {i}\\right), i = 1, 2. \\tag {9}", + "image_path": "fa3fd4fb788d656a9f70729e5d20c08367c55f0ab26340190d0bd42a2b950786.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "text", + "content": "The contexts " + }, + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "inline_equation", + "content": "x_{1}, x_{2}" + }, + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "text", + "content": " and targets " + }, + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "inline_equation", + "content": "y_{1}, y_{2}" + }, + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "text", + "content": " are then encoded into representations " + }, + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "inline_equation", + "content": "h_{x_1}, h_{x_2}, h_{y_1}, h_{y_2}" + }, + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "text", + "content": ". Note that the targets " + }, + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "inline_equation", + "content": "h_{y_1}, h_{y_2}" + }, + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "text", + "content": " are predictable from the contexts " + }, + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "inline_equation", + "content": "h_{x_1}, h_{x_2}" + }, + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "text", + "content": " given the proper conditions. Specifically, since " + }, + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "inline_equation", + "content": "x_{1}" + }, + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "text", + "content": " is obtained by applying augmentation " + }, + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{a_1}" + }, + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "text", + "content": " and mask " + }, + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "inline_equation", + "content": "M_{1}" + }, + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "inline_equation", + "content": "y_{1}" + }, + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "text", + "content": ", predicting " + }, + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "inline_equation", + "content": "y_{1}" + }, + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "inline_equation", + "content": "x_{1}" + }, + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "text", + "content": " with conditions " + }, + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "inline_equation", + "content": "(a_{1}, M_{1})" + }, + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "text", + "content": " simultaneously performs local anatomical structure modeling and domain vari" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20782" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 109, + 74, + 142, + 114 + ], + "blocks": [ + { + "bbox": [ + 109, + 74, + 142, + 114 + ], + "lines": [ + { + "bbox": [ + 109, + 74, + 142, + 114 + ], + "spans": [ + { + "bbox": [ + 109, + 74, + 142, + 114 + ], + "type": "image", + "image_path": "326a376dec9e0bf39f24f7676a0b89436ef8bbddd6c495f2e7aa7c3eb3b20aaf.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 145, + 82, + 179, + 114 + ], + "blocks": [ + { + "bbox": [ + 150, + 74, + 174, + 81 + ], + "lines": [ + { + "bbox": [ + 150, + 74, + 174, + 81 + ], + "spans": [ + { + "bbox": [ + 150, + 74, + 174, + 81 + ], + "type": "text", + "content": "Context" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 145, + 82, + 179, + 114 + ], + "lines": [ + { + "bbox": [ + 145, + 82, + 179, + 114 + ], + "spans": [ + { + "bbox": [ + 145, + 82, + 179, + 114 + ], + "type": "image", + "image_path": "3c393711d543ed0ddef7ca91d62b1b34bc1610251ce66e06778252f2efbd5519.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 181, + 82, + 312, + 114 + ], + "blocks": [ + { + "bbox": [ + 181, + 74, + 195, + 81 + ], + "lines": [ + { + "bbox": [ + 181, + 74, + 195, + 81 + ], + "spans": [ + { + "bbox": [ + 181, + 74, + 195, + 81 + ], + "type": "text", + "content": "" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 181, + 82, + 312, + 114 + ], + "lines": [ + { + "bbox": [ + 181, + 82, + 312, + 114 + ], + "spans": [ + { + "bbox": [ + 181, + 82, + 312, + 114 + ], + "type": "image", + "image_path": "e05d0f362b2847b042066c536b5391fcd1c9256bbd54baae17c3ed9a43cd38d2.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 215, + 74, + 279, + 80 + ], + "lines": [ + { + "bbox": [ + 215, + 74, + 279, + 80 + ], + "spans": [ + { + "bbox": [ + 215, + 74, + 279, + 80 + ], + "type": "text", + "content": "World Model Prediction" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 315, + 82, + 378, + 114 + ], + "blocks": [ + { + "bbox": [ + 318, + 74, + 375, + 81 + ], + "lines": [ + { + "bbox": [ + 318, + 74, + 375, + 81 + ], + "spans": [ + { + "bbox": [ + 318, + 74, + 375, + 81 + ], + "type": "text", + "content": "Target v.s. Prediction" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 315, + 82, + 378, + 114 + ], + "lines": [ + { + "bbox": [ + 315, + 82, + 378, + 114 + ], + "spans": [ + { + "bbox": [ + 315, + 82, + 378, + 114 + ], + "type": "image", + "image_path": "4a6a87624a25db5f3084e7d5c3daa7c5a4d444d6634baf1452eb4dc387d0f251.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 109, + 117, + 143, + 150 + ], + "blocks": [ + { + "bbox": [ + 109, + 117, + 143, + 150 + ], + "lines": [ + { + "bbox": [ + 109, + 117, + 143, + 150 + ], + "spans": [ + { + "bbox": [ + 109, + 117, + 143, + 150 + ], + "type": "image", + "image_path": "1578f9d8ae5ea7e7ea8f89cefe781aef84d40ef352241f8ca925dccc0d83964d.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 145, + 117, + 179, + 150 + ], + "blocks": [ + { + "bbox": [ + 145, + 117, + 179, + 150 + ], + "lines": [ + { + "bbox": [ + 145, + 117, + 179, + 150 + ], + "spans": [ + { + "bbox": [ + 145, + 117, + 179, + 150 + ], + "type": "image", + "image_path": "8b7daefee0531d89e9f371355a2141d01f1d9b99df9a6a8c0878974a68567031.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 181, + 117, + 214, + 150 + ], + "blocks": [ + { + "bbox": [ + 181, + 117, + 214, + 150 + ], + "lines": [ + { + "bbox": [ + 181, + 117, + 214, + 150 + ], + "spans": [ + { + "bbox": [ + 181, + 117, + 214, + 150 + ], + "type": "image", + "image_path": "1fe7a232935c3719ac2603fb66cc49393bb50ff7bd23b1ce1339c09aaaf94123.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 215, + 117, + 249, + 150 + ], + "blocks": [ + { + "bbox": [ + 215, + 117, + 249, + 150 + ], + "lines": [ + { + "bbox": [ + 215, + 117, + 249, + 150 + ], + "spans": [ + { + "bbox": [ + 215, + 117, + 249, + 150 + ], + "type": "image", + "image_path": "f7fdbea78ae50e67feec66900a240ecbca2281bfc61710f623a75b0d32120159.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 250, + 117, + 313, + 150 + ], + "blocks": [ + { + "bbox": [ + 250, + 117, + 313, + 150 + ], + "lines": [ + { + "bbox": [ + 250, + 117, + 313, + 150 + ], + "spans": [ + { + "bbox": [ + 250, + 117, + 313, + 150 + ], + "type": "image", + "image_path": "405a4f111d7f3da1fb17199a51746be3e0c2c06566e2c79c3981df678f649037.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 315, + 117, + 346, + 150 + ], + "blocks": [ + { + "bbox": [ + 315, + 117, + 346, + 150 + ], + "lines": [ + { + "bbox": [ + 315, + 117, + 346, + 150 + ], + "spans": [ + { + "bbox": [ + 315, + 117, + 346, + 150 + ], + "type": "image", + "image_path": "1acfa791be86c8840b8164fcc64c9abc5b3d7523715900cc91ea4a9602201e2b.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 347, + 117, + 378, + 150 + ], + "blocks": [ + { + "bbox": [ + 347, + 117, + 378, + 150 + ], + "lines": [ + { + "bbox": [ + 347, + 117, + 378, + 150 + ], + "spans": [ + { + "bbox": [ + 347, + 117, + 378, + 150 + ], + "type": "image", + "image_path": "cb54c4b6fff1665566c101437ccf6d012fc8d227e674c0adc8012c0ffe54a839.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 109, + 153, + 142, + 186 + ], + "blocks": [ + { + "bbox": [ + 109, + 153, + 142, + 186 + ], + "lines": [ + { + "bbox": [ + 109, + 153, + 142, + 186 + ], + "spans": [ + { + "bbox": [ + 109, + 153, + 142, + 186 + ], + "type": "image", + "image_path": "4e9914b5b09de9100aeaebdf3f6f44d2d312e95efd9e31917159462bc45a8a9d.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 145, + 153, + 179, + 186 + ], + "blocks": [ + { + "bbox": [ + 145, + 153, + 179, + 186 + ], + "lines": [ + { + "bbox": [ + 145, + 153, + 179, + 186 + ], + "spans": [ + { + "bbox": [ + 145, + 153, + 179, + 186 + ], + "type": "image", + "image_path": "4f9d63a1bd528c4e6f4fe5319f73e5242562c69ea8a3a7c9d0848b2b3fe4b99c.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 181, + 153, + 214, + 186 + ], + "blocks": [ + { + "bbox": [ + 181, + 153, + 214, + 186 + ], + "lines": [ + { + "bbox": [ + 181, + 153, + 214, + 186 + ], + "spans": [ + { + "bbox": [ + 181, + 153, + 214, + 186 + ], + "type": "image", + "image_path": "5b68524b0361cd794535970be23d278a2f94a5d70ff7e59bb3b41a71be1449ce.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 215, + 153, + 249, + 186 + ], + "blocks": [ + { + "bbox": [ + 215, + 153, + 249, + 186 + ], + "lines": [ + { + "bbox": [ + 215, + 153, + 249, + 186 + ], + "spans": [ + { + "bbox": [ + 215, + 153, + 249, + 186 + ], + "type": "image", + "image_path": "316738ab436ce6819a06be4f6a4cc231c005acead4ef759e1e5598e3d51e1d9b.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 250, + 153, + 313, + 186 + ], + "blocks": [ + { + "bbox": [ + 250, + 153, + 313, + 186 + ], + "lines": [ + { + "bbox": [ + 250, + 153, + 313, + 186 + ], + "spans": [ + { + "bbox": [ + 250, + 153, + 313, + 186 + ], + "type": "image", + "image_path": "b74369d4637276f88f0aff867a281e3159aecd75bb4f5e445d87404f2703e9b8.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 315, + 153, + 346, + 186 + ], + "blocks": [ + { + "bbox": [ + 315, + 153, + 346, + 186 + ], + "lines": [ + { + "bbox": [ + 315, + 153, + 346, + 186 + ], + "spans": [ + { + "bbox": [ + 315, + 153, + 346, + 186 + ], + "type": "image", + "image_path": "fac77acb0ce143ec46c2d7b6754d4d9d33b47a6f3568d33eb76902867fd6d7e8.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 347, + 153, + 378, + 186 + ], + "blocks": [ + { + "bbox": [ + 347, + 153, + 378, + 186 + ], + "lines": [ + { + "bbox": [ + 347, + 153, + 378, + 186 + ], + "spans": [ + { + "bbox": [ + 347, + 153, + 378, + 186 + ], + "type": "image", + "image_path": "17de1a0bddc331484e9b41551160d397d63b2b5cb677ddb04dd1db832dfd2e8c.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 109, + 190, + 143, + 222 + ], + "blocks": [ + { + "bbox": [ + 109, + 190, + 143, + 222 + ], + "lines": [ + { + "bbox": [ + 109, + 190, + 143, + 222 + ], + "spans": [ + { + "bbox": [ + 109, + 190, + 143, + 222 + ], + "type": "image", + "image_path": "b2e0f4c24cf93edbbb8c995c78b0b4cf8fbea4b3d50dfcb2c7bf6af8502eec0c.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 145, + 190, + 179, + 222 + ], + "blocks": [ + { + "bbox": [ + 145, + 190, + 179, + 222 + ], + "lines": [ + { + "bbox": [ + 145, + 190, + 179, + 222 + ], + "spans": [ + { + "bbox": [ + 145, + 190, + 179, + 222 + ], + "type": "image", + "image_path": "89217775023f4aabd9b313a4c9b7187fb54dd398bd8e7e158237f440c9c930be.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 181, + 190, + 214, + 222 + ], + "blocks": [ + { + "bbox": [ + 181, + 190, + 214, + 222 + ], + "lines": [ + { + "bbox": [ + 181, + 190, + 214, + 222 + ], + "spans": [ + { + "bbox": [ + 181, + 190, + 214, + 222 + ], + "type": "image", + "image_path": "d6ec51fe22edc9e09e50bac10bf179e8dfa88cd0b201882bf50a5bb947ff1adc.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 215, + 190, + 249, + 222 + ], + "blocks": [ + { + "bbox": [ + 215, + 190, + 249, + 222 + ], + "lines": [ + { + "bbox": [ + 215, + 190, + 249, + 222 + ], + "spans": [ + { + "bbox": [ + 215, + 190, + 249, + 222 + ], + "type": "image", + "image_path": "00de25365320a54a54e319665857f8a1ef07c8e7cc7855144c581c004420fba1.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 250, + 190, + 313, + 222 + ], + "blocks": [ + { + "bbox": [ + 250, + 190, + 313, + 222 + ], + "lines": [ + { + "bbox": [ + 250, + 190, + 313, + 222 + ], + "spans": [ + { + "bbox": [ + 250, + 190, + 313, + 222 + ], + "type": "image", + "image_path": "132de9a0a72ba287f220e3e8bc707bfd6b349e1a8b95fd4acb32ee58eaf5c0a7.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 315, + 190, + 346, + 222 + ], + "blocks": [ + { + "bbox": [ + 315, + 190, + 346, + 222 + ], + "lines": [ + { + "bbox": [ + 315, + 190, + 346, + 222 + ], + "spans": [ + { + "bbox": [ + 315, + 190, + 346, + 222 + ], + "type": "image", + "image_path": "8e40d3f2edd650a2a8054a19486f165ab07875cedd1dbd78afb881e8b7dadc51.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 347, + 190, + 378, + 222 + ], + "blocks": [ + { + "bbox": [ + 347, + 190, + 378, + 222 + ], + "lines": [ + { + "bbox": [ + 347, + 190, + 378, + 222 + ], + "spans": [ + { + "bbox": [ + 347, + 190, + 378, + 222 + ], + "type": "image", + "image_path": "6078cf6bd9568bf678dbdbfaa2209fb319f4dd52ad3a2ba018db042c099a8514.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 392, + 82, + 499, + 114 + ], + "blocks": [ + { + "bbox": [ + 388, + 74, + 498, + 81 + ], + "lines": [ + { + "bbox": [ + 388, + 74, + 498, + 81 + ], + "spans": [ + { + "bbox": [ + 388, + 74, + 498, + 81 + ], + "type": "text", + "content": "Context&Target World Model Prediction" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 392, + 82, + 499, + 114 + ], + "lines": [ + { + "bbox": [ + 392, + 82, + 499, + 114 + ], + "spans": [ + { + "bbox": [ + 392, + 82, + 499, + 114 + ], + "type": "image", + "image_path": "127370d9c02fe8a49b741b948cedab78c0e79676abc581127f75bf1ff0afb6f7.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 392, + 117, + 463, + 150 + ], + "blocks": [ + { + "bbox": [ + 392, + 117, + 463, + 150 + ], + "lines": [ + { + "bbox": [ + 392, + 117, + 463, + 150 + ], + "spans": [ + { + "bbox": [ + 392, + 117, + 463, + 150 + ], + "type": "image", + "image_path": "ba3fda43c6bfbf81898b88f642c048df3780af2c95d93120ecb7688dc3bbb931.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 465, + 116, + 498, + 150 + ], + "blocks": [ + { + "bbox": [ + 465, + 116, + 498, + 150 + ], + "lines": [ + { + "bbox": [ + 465, + 116, + 498, + 150 + ], + "spans": [ + { + "bbox": [ + 465, + 116, + 498, + 150 + ], + "type": "image", + "image_path": "5b308fb86f8b761fae3d16bc3e92382890f3e0cb2b2c68c0d9c358aa9b1ca72a.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 392, + 153, + 425, + 185 + ], + "blocks": [ + { + "bbox": [ + 392, + 153, + 425, + 185 + ], + "lines": [ + { + "bbox": [ + 392, + 153, + 425, + 185 + ], + "spans": [ + { + "bbox": [ + 392, + 153, + 425, + 185 + ], + "type": "image", + "image_path": "a0dc1f2400ca8df47b7b74e261892668f7ca35427301e2ab51609fe1f8f9a16d.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + } + ], + "index": 33 + }, + { + "type": "image", + "bbox": [ + 428, + 153, + 460, + 185 + ], + "blocks": [ + { + "bbox": [ + 428, + 153, + 460, + 185 + ], + "lines": [ + { + "bbox": [ + 428, + 153, + 460, + 185 + ], + "spans": [ + { + "bbox": [ + 428, + 153, + 460, + 185 + ], + "type": "image", + "image_path": "2af7c6f860936e742feaa1ffcdba3f2b981fe9c6447fe8ec105f109f7c6f419e.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 465, + 153, + 498, + 185 + ], + "blocks": [ + { + "bbox": [ + 465, + 153, + 498, + 185 + ], + "lines": [ + { + "bbox": [ + 465, + 153, + 498, + 185 + ], + "spans": [ + { + "bbox": [ + 465, + 153, + 498, + 185 + ], + "type": "image", + "image_path": "c330d2f8b5b992503bc406ad510d8af4f56ee01781e1097742c8f2813a118f8a.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + } + ], + "index": 35 + }, + { + "type": "image", + "bbox": [ + 392, + 189, + 425, + 223 + ], + "blocks": [ + { + "bbox": [ + 392, + 189, + 425, + 223 + ], + "lines": [ + { + "bbox": [ + 392, + 189, + 425, + 223 + ], + "spans": [ + { + "bbox": [ + 392, + 189, + 425, + 223 + ], + "type": "image", + "image_path": "0d908be3ef7d7d9f628243f83e199333b3bcb4ff8c35103d86a12bc0f67e350c.jpg" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 406, + 224, + 486, + 232 + ], + "lines": [ + { + "bbox": [ + 406, + 224, + 486, + 232 + ], + "spans": [ + { + "bbox": [ + 406, + 224, + 486, + 232 + ], + "type": "text", + "content": "(b) Global Layout Modeling" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_caption" + } + ], + "index": 36 + }, + { + "type": "image", + "bbox": [ + 428, + 189, + 463, + 223 + ], + "blocks": [ + { + "bbox": [ + 428, + 189, + 463, + 223 + ], + "lines": [ + { + "bbox": [ + 428, + 189, + 463, + 223 + ], + "spans": [ + { + "bbox": [ + 428, + 189, + 463, + 223 + ], + "type": "image", + "image_path": "0059af19b42a4ffcbd012761f7583551f6e5aa1436a8ef33867a4fb9e6d4b377.jpg" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_body" + } + ], + "index": 37 + }, + { + "type": "image", + "bbox": [ + 465, + 189, + 498, + 223 + ], + "blocks": [ + { + "bbox": [ + 465, + 189, + 498, + 223 + ], + "lines": [ + { + "bbox": [ + 465, + 189, + 498, + 223 + ], + "spans": [ + { + "bbox": [ + 465, + 189, + 498, + 223 + ], + "type": "image", + "image_path": "b2fdf30bec8b2577924aaf4f5faec368317a8ccd61d3e1258c606e3fcb32c8a1.jpg" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_body" + } + ], + "index": 38 + }, + { + "type": "image", + "bbox": [ + 86, + 280, + 133, + 330 + ], + "blocks": [ + { + "bbox": [ + 86, + 280, + 133, + 330 + ], + "lines": [ + { + "bbox": [ + 86, + 280, + 133, + 330 + ], + "spans": [ + { + "bbox": [ + 86, + 280, + 133, + 330 + ], + "type": "image", + "image_path": "4470cd4eef2e787a3003ce98a6b5b65d6cb55f865e7b25a5be4945edc7efd31a.jpg" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_body" + } + ], + "index": 41 + }, + { + "type": "image", + "bbox": [ + 143, + 281, + 187, + 330 + ], + "blocks": [ + { + "bbox": [ + 143, + 281, + 187, + 330 + ], + "lines": [ + { + "bbox": [ + 143, + 281, + 187, + 330 + ], + "spans": [ + { + "bbox": [ + 143, + 281, + 187, + 330 + ], + "type": "image", + "image_path": "e7428427db91350638451feb75d41c54ea8cc456647fd392ddbe29f4dff0623b.jpg" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_body" + } + ], + "index": 42 + }, + { + "type": "image", + "bbox": [ + 198, + 281, + 244, + 330 + ], + "blocks": [ + { + "bbox": [ + 198, + 281, + 244, + 330 + ], + "lines": [ + { + "bbox": [ + 198, + 281, + 244, + 330 + ], + "spans": [ + { + "bbox": [ + 198, + 281, + 244, + 330 + ], + "type": "image", + "image_path": "674bdda378e763f892e1811dc2bf9b89245fc03c983577af3a8c2747c7d3f34e.jpg" + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_body" + } + ], + "index": 43 + }, + { + "type": "image", + "bbox": [ + 87, + 332, + 133, + 355 + ], + "blocks": [ + { + "bbox": [ + 87, + 332, + 133, + 355 + ], + "lines": [ + { + "bbox": [ + 87, + 332, + 133, + 355 + ], + "spans": [ + { + "bbox": [ + 87, + 332, + 133, + 355 + ], + "type": "image", + "image_path": "9e6aea537c52cc755f421ea085c45ff7fe1acb68129f0c219ecacc815c20bfc0.jpg" + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "image_body" + } + ], + "index": 44 + }, + { + "type": "image", + "bbox": [ + 143, + 333, + 188, + 356 + ], + "blocks": [ + { + "bbox": [ + 143, + 333, + 188, + 356 + ], + "lines": [ + { + "bbox": [ + 143, + 333, + 188, + 356 + ], + "spans": [ + { + "bbox": [ + 143, + 333, + 188, + 356 + ], + "type": "image", + "image_path": "e624d15008386098ae6d1b94d2d91744218561430d95970cb57e11f3d2c449e4.jpg" + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 56, + 361, + 272, + 395 + ], + "lines": [ + { + "bbox": [ + 56, + 361, + 272, + 395 + ], + "spans": [ + { + "bbox": [ + 56, + 361, + 272, + 395 + ], + "type": "text", + "content": "Figure 5. CheXWorld prioritizes relevant medical features over spurious signals (e.g., lateral markers) in the image. Regions in red boxes denote the predictor outputs." + } + ] + } + ], + "index": 47, + "angle": 0, + "type": "image_caption" + } + ], + "index": 45 + }, + { + "type": "image", + "bbox": [ + 198, + 333, + 243, + 356 + ], + "blocks": [ + { + "bbox": [ + 198, + 333, + 243, + 356 + ], + "lines": [ + { + "bbox": [ + 198, + 333, + 243, + 356 + ], + "spans": [ + { + "bbox": [ + 198, + 333, + 243, + 356 + ], + "type": "image", + "image_path": "e0dc553cf827785bfe5397648b2d9b8438c935de2f226aa0219ee5ef6488fb93.jpg" + } + ] + } + ], + "index": 46, + "angle": 0, + "type": "image_body" + } + ], + "index": 46 + }, + { + "type": "image", + "bbox": [ + 301, + 279, + 531, + 350 + ], + "blocks": [ + { + "bbox": [ + 55, + 232, + 555, + 276 + ], + "lines": [ + { + "bbox": [ + 55, + 232, + 555, + 276 + ], + "spans": [ + { + "bbox": [ + 55, + 232, + 555, + 276 + ], + "type": "text", + "content": "Figure 4. Visualization of the CheXWorld predictor outputs (zooming in for details). The images presented in this figure were not included in the pre-training of CheXWorld or the training of the diffusion model. Regions in red bounding boxes denote the predictor outputs that are mapped to pixel space using the RCDM [8] framework. In (a), gray areas indicate masked regions excluded from the context. In (b), the two overlapping regions alternately serve as context and target." + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 301, + 279, + 531, + 350 + ], + "lines": [ + { + "bbox": [ + 301, + 279, + 531, + 350 + ], + "spans": [ + { + "bbox": [ + 301, + 279, + 531, + 350 + ], + "type": "image", + "image_path": "f1e9db7cb9dcd0bec2cec68238f447c7b69345564a860bca83b60abee864ab82.jpg" + } + ] + } + ], + "index": 48, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 277, + 352, + 553, + 396 + ], + "lines": [ + { + "bbox": [ + 277, + 352, + 553, + 396 + ], + "spans": [ + { + "bbox": [ + 277, + 352, + 553, + 396 + ], + "type": "text", + "content": "Figure 6. CheXWorld learns anatomical correspondence. We compute pixel-level embeddings using RoI-pooling [22] and calculate the embedding similarity between four anatomical landmarks in a reference image and each pixel in the test image to create feature similarity heatmaps." + } + ] + } + ], + "index": 49, + "angle": 0, + "type": "image_caption" + } + ], + "index": 48 + }, + { + "bbox": [ + 55, + 406, + 122, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 406, + 122, + 418 + ], + "spans": [ + { + "bbox": [ + 55, + 406, + 122, + 418 + ], + "type": "text", + "content": "ation modeling:" + } + ] + } + ], + "index": 50 + }, + { + "bbox": [ + 69, + 422, + 294, + 463 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 422, + 294, + 463 + ], + "spans": [ + { + "bbox": [ + 69, + 422, + 294, + 463 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {1 \\rightarrow 1} = \\mathcal {L} _ {\\text {l o c a l} + \\text {d o m a i n}} \\left(x _ {1}, y _ {1}\\right) \\\\ = \\sum_ {c \\in M _ {1}} \\| g _ {\\phi} \\left(h _ {x _ {1}}; M _ {1}, a _ {1}\\right) _ {c} - f _ {\\theta^ {\\prime}} ^ {\\prime} \\left(y _ {1}\\right) _ {c} \\| _ {2} ^ {2}. \\tag {10} \\\\ \\end{array}", + "image_path": "2de8b50c5f813b07dcd8eafb0b3374f88a56f47009a9e01d23498a28c6ca69f8.jpg" + } + ] + } + ], + "index": 51 + }, + { + "bbox": [ + 55, + 467, + 296, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 467, + 296, + 526 + ], + "spans": [ + { + "bbox": [ + 55, + 467, + 296, + 526 + ], + "type": "text", + "content": "Moreover, " + }, + { + "bbox": [ + 55, + 467, + 296, + 526 + ], + "type": "inline_equation", + "content": "x_{1}" + }, + { + "bbox": [ + 55, + 467, + 296, + 526 + ], + "type": "text", + "content": " can be derived from " + }, + { + "bbox": [ + 55, + 467, + 296, + 526 + ], + "type": "inline_equation", + "content": "y_{2}" + }, + { + "bbox": [ + 55, + 467, + 296, + 526 + ], + "type": "text", + "content": " by moving the input window by " + }, + { + "bbox": [ + 55, + 467, + 296, + 526 + ], + "type": "inline_equation", + "content": "-\\Delta_{x_1\\to y_2}" + }, + { + "bbox": [ + 55, + 467, + 296, + 526 + ], + "type": "text", + "content": " and applying augmentation " + }, + { + "bbox": [ + 55, + 467, + 296, + 526 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{a_1}" + }, + { + "bbox": [ + 55, + 467, + 296, + 526 + ], + "type": "text", + "content": ", hence the predictor conditioning on " + }, + { + "bbox": [ + 55, + 467, + 296, + 526 + ], + "type": "inline_equation", + "content": "(\\Delta_{x_1\\to y_2},a_1)" + }, + { + "bbox": [ + 55, + 467, + 296, + 526 + ], + "type": "text", + "content": " performs global anatomical structure modeling and domain variation modeling at the same time." + } + ] + } + ], + "index": 52 + }, + { + "bbox": [ + 75, + 531, + 294, + 562 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 531, + 294, + 562 + ], + "spans": [ + { + "bbox": [ + 75, + 531, + 294, + 562 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {1 \\rightarrow 2} = \\mathcal {L} _ {\\text {g l o b a l} + \\text {d o m a i n}} \\left(x _ {1}, y _ {2}\\right) \\tag {11} \\\\ = \\left\\| g _ {\\phi} \\left(h _ {x _ {1}}; \\Delta_ {x _ {1} \\rightarrow y _ {2}}, a _ {1}\\right) - f _ {\\theta^ {\\prime}} ^ {\\prime} (y _ {2}) \\right\\| _ {2} ^ {2}. \\\\ \\end{array}", + "image_path": "c724d7492f09fa504d30afd86a4ce3cb7d57d592583fe448ff986d44d798ba7c.jpg" + } + ] + } + ], + "index": 53 + }, + { + "bbox": [ + 55, + 566, + 296, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 566, + 296, + 638 + ], + "spans": [ + { + "bbox": [ + 55, + 566, + 296, + 638 + ], + "type": "text", + "content": "Similarly, " + }, + { + "bbox": [ + 55, + 566, + 296, + 638 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{2\\rightarrow 2},\\mathcal{L}_{2\\rightarrow 1}" + }, + { + "bbox": [ + 55, + 566, + 296, + 638 + ], + "type": "text", + "content": " can be computed between " + }, + { + "bbox": [ + 55, + 566, + 296, + 638 + ], + "type": "inline_equation", + "content": "(x_{2},y_{2})" + }, + { + "bbox": [ + 55, + 566, + 296, + 638 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 566, + 296, + 638 + ], + "type": "inline_equation", + "content": "(x_{2},y_{1})" + }, + { + "bbox": [ + 55, + 566, + 296, + 638 + ], + "type": "text", + "content": ", respectively. Note that the context features and target features are reused twice when computing the loss (10) and (11). Finally, the model is trained with the combination of " + }, + { + "bbox": [ + 55, + 566, + 296, + 638 + ], + "type": "inline_equation", + "content": "2\\times 2 = 4" + }, + { + "bbox": [ + 55, + 566, + 296, + 638 + ], + "type": "text", + "content": " supervisions computed from the two pairs of context and target." + } + ] + } + ], + "index": 54 + }, + { + "bbox": [ + 97, + 643, + 294, + 654 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 643, + 294, + 654 + ], + "spans": [ + { + "bbox": [ + 97, + 643, + 294, + 654 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {1 \\rightarrow 1} + \\mathcal {L} _ {2 \\rightarrow 2} + \\mathcal {L} _ {1 \\rightarrow 2} + \\mathcal {L} _ {2 \\rightarrow 1}. \\tag {12}", + "image_path": "a0b8422cec79ffc95ffcf92fec09d8deb2eccab655956ffc2b5ac13edc91d876.jpg" + } + ] + } + ], + "index": 55 + }, + { + "bbox": [ + 55, + 659, + 296, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 659, + 296, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 659, + 296, + 696 + ], + "type": "text", + "content": "With this unified objective, CheXWorld showcases deep comprehension of human anatomy in different scales and pixel intensity distributions across various domains." + } + ] + } + ], + "index": 56 + }, + { + "bbox": [ + 314, + 404, + 394, + 418 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 404, + 394, + 418 + ], + "spans": [ + { + "bbox": [ + 314, + 404, + 394, + 418 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 57 + }, + { + "bbox": [ + 313, + 425, + 555, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 425, + 555, + 496 + ], + "spans": [ + { + "bbox": [ + 313, + 425, + 555, + 496 + ], + "type": "text", + "content": "In this section, we first demonstrate CheXWorld's world modeling capability with a series of analytical experiments. Then, we empirically compare CheXWorld with other self-supervised methods as baselines on eight downstream tasks. Finally, we provide ablation studies to highlight the effect of each building block of our method." + } + ] + } + ], + "index": 58 + }, + { + "bbox": [ + 313, + 497, + 555, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 497, + 555, + 556 + ], + "spans": [ + { + "bbox": [ + 313, + 497, + 555, + 556 + ], + "type": "text", + "content": "Implementation details. CheXWorld is pre-trained on " + }, + { + "bbox": [ + 313, + 497, + 555, + 556 + ], + "type": "inline_equation", + "content": "\\sim 0.5\\mathrm{M}" + }, + { + "bbox": [ + 313, + 497, + 555, + 556 + ], + "type": "text", + "content": " frontal chest X-rays from several public datasets. The visual backbone is a ViT-Base [18]. The model is trained for 300 epochs, taking 16 hours on 8 RTX 4090 GPUs. Please refer to the appendix for more details." + } + ] + } + ], + "index": 59 + }, + { + "bbox": [ + 313, + 563, + 525, + 577 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 563, + 525, + 577 + ], + "spans": [ + { + "bbox": [ + 313, + 563, + 525, + 577 + ], + "type": "text", + "content": "5.1. CheXWorld is a Strong World Predictor" + } + ] + } + ], + "index": 60 + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "content": "Visualization of anatomical modeling. To better understand the predictions made by CheXWorld, we train a generative model that maps the average-pooled predictor output back to pixel space following the RCDM framework [8]. Specifically, we construct context-target pairs in the same way as in the local and global anatomical modeling tasks. Then, we feed the context to a frozen pre-trained CheX-World model to obtain the model's feature prediction, on which a diffusion model conditions to produce the target image in pixel space. As shown in Figure 4(a), the CheX-World predictor perfectly recovers the appearance of bones" + } + ] + } + ], + "index": 61 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20783" + } + ] + } + ], + "index": 62 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 57, + 97, + 531, + 264 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "text", + "content": "Table 1. Performance on five downstream classification benchmarks. Accuracy is reported for RSNA, and the area under the ROC curve (AUROC) is reported for the rest of the datasets. The best two results are bold-faced and underlined, respectively." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 97, + 531, + 264 + ], + "lines": [ + { + "bbox": [ + 57, + 97, + 531, + 264 + ], + "spans": [ + { + "bbox": [ + 57, + 97, + 531, + 264 + ], + "type": "table", + "html": "
MethodPre-train DataBackboneVinDr-CXRShenZhenChestX-ray14RSNA (CLS)CheXpert
Scratch-ViT-B70.22±1.9582.24±0.6071.69±0.3266.59±0.3980.78±0.03
MoCo-v3 [14]IN-1k (1.3M)ViT-B87.25±0.6392.85±1.0079.20±0.3072.79±0.5287.12±0.36
DINO [11]IN-1k (1.3M)ViT-B82.89±1.1090.39±4.2978.37±0.4771.27±0.4587.01±0.62
BEiT [6]IN-21k (12M)ViT-B85.93±1.9892.87±1.0879.91±0.2472.78±0.3787.77±0.38
LVM-Med [44]Medical (1.3M)ViT-B88.22±0.4494.81±1.3280.08±0.0972.75±0.4488.07±0.25
MAE [31, 63]X-rays (0.5M)ViT-B92.76±0.1897.63±0.2183.0†73.75±0.2489.3†
SimMIM [42, 64]IN-21k (12M) & X-rays (0.9M)Swin-B92.81±0.3198.09±0.1383.04±0.1574.09±0.3989.14±0.22
Adam-v2 [32]IN-21k (12M) & X-rays (0.9M)ConvNeXt-B91.46±0.3397.80†83.4†73.40±0.8888.90±0.36
CheXWorldX-rays (0.5M)ViT-B95.24±0.1398.88±0.0683.58±0.0575.03±0.3989.63±0.13
Rad-DINO‡[51]LVD (142M) & X-rays (public+private)ViT-B95.16±0.1698.20±0.1783.61±0.0874.51±0.4688.94±0.15
", + "image_path": "034fb371fd2957b8b2d5edfa65f0da436d18be671962daf31ffbe7d7f3932778.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 55, + 264, + 533, + 285 + ], + "lines": [ + { + "bbox": [ + 55, + 264, + 533, + 285 + ], + "spans": [ + { + "bbox": [ + 55, + 264, + 533, + 285 + ], + "type": "text", + "content": "Results on ImageNet pre-trained models are adopted from [42]. " + }, + { + "bbox": [ + 55, + 264, + 533, + 285 + ], + "type": "inline_equation", + "content": "{}^{ \\dagger }" + }, + { + "bbox": [ + 55, + 264, + 533, + 285 + ], + "type": "text", + "content": " Results reported by the original authors [31, 32]. " + }, + { + "bbox": [ + 55, + 264, + 533, + 285 + ], + "type": "inline_equation", + "content": "{}^{ \\ddagger }" + }, + { + "bbox": [ + 55, + 264, + 533, + 285 + ], + "type": "text", + "content": " Rad-DINO [51] is included for reference but not for direct comparison, as it requires 2560 GPU-hours (20 times more than ours) and is trained on both public and private datasets." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 295, + 294, + 355 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 295, + 294, + 355 + ], + "spans": [ + { + "bbox": [ + 54, + 295, + 294, + 355 + ], + "type": "text", + "content": "and tissues within various masked areas, and Figure 4(b) shows that the model makes global-level predictions that are consistent with the context. Further visualizations also show that CheXWorld filters artifacts in the image (Figure 5) and learns anatomical correspondence (Figure 6)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 357, + 296, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 357, + 296, + 559 + ], + "spans": [ + { + "bbox": [ + 56, + 357, + 296, + 559 + ], + "type": "text", + "content": "Awareness to domain variations. We design a domain sensitivity test to verify the model's ability to capture the uncertainty in a radiograph's appearance attributes. Specifically, we first generate a candidate set of targets " + }, + { + "bbox": [ + 56, + 357, + 296, + 559 + ], + "type": "inline_equation", + "content": "\\{y_i\\}_{i=1}^n" + }, + { + "bbox": [ + 56, + 357, + 296, + 559 + ], + "type": "text", + "content": " by applying augmentations with different configurations to the same image. Then, we construct a set of context-target-latent triplets " + }, + { + "bbox": [ + 56, + 357, + 296, + 559 + ], + "type": "inline_equation", + "content": "\\{(x_i, y_i, a_i)\\}" + }, + { + "bbox": [ + 56, + 357, + 296, + 559 + ], + "type": "text", + "content": " by applying another augmentation " + }, + { + "bbox": [ + 56, + 357, + 296, + 559 + ], + "type": "inline_equation", + "content": "x_i = \\mathcal{T}_{a_i}(y_i)" + }, + { + "bbox": [ + 56, + 357, + 296, + 559 + ], + "type": "text", + "content": ". The model is asked to predict the target " + }, + { + "bbox": [ + 56, + 357, + 296, + 559 + ], + "type": "inline_equation", + "content": "y_i" + }, + { + "bbox": [ + 56, + 357, + 296, + 559 + ], + "type": "text", + "content": " given the context " + }, + { + "bbox": [ + 56, + 357, + 296, + 559 + ], + "type": "inline_equation", + "content": "x_i" + }, + { + "bbox": [ + 56, + 357, + 296, + 559 + ], + "type": "text", + "content": " and latent " + }, + { + "bbox": [ + 56, + 357, + 296, + 559 + ], + "type": "inline_equation", + "content": "a_i" + }, + { + "bbox": [ + 56, + 357, + 296, + 559 + ], + "type": "text", + "content": ". For each predicted output " + }, + { + "bbox": [ + 56, + 357, + 296, + 559 + ], + "type": "inline_equation", + "content": "\\hat{y}_i" + }, + { + "bbox": [ + 56, + 357, + 296, + 559 + ], + "type": "text", + "content": ", we calculate the top-k recall rate of the true target " + }, + { + "bbox": [ + 56, + 357, + 296, + 559 + ], + "type": "inline_equation", + "content": "y_i" + }, + { + "bbox": [ + 56, + 357, + 296, + 559 + ], + "type": "text", + "content": " over the entire candidate set using L2 distance2. As shown in Table 4, our model achieves an average top-5 recall of 77.67 (10 times higher than random choice), demonstrating the model's strong discriminative ability across domains. We also observe a significant drop in the recall rate when removing the domain condition " + }, + { + "bbox": [ + 56, + 357, + 296, + 559 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 56, + 357, + 296, + 559 + ], + "type": "text", + "content": ", indicating that the model indeed makes predictions conditioned on " + }, + { + "bbox": [ + 56, + 357, + 296, + 559 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 56, + 357, + 296, + 559 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 570, + 295, + 594 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 295, + 594 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 295, + 594 + ], + "type": "text", + "content": "5.2. CheXWorld Produces Transferable Representations" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 601, + 295, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 601, + 295, + 673 + ], + "spans": [ + { + "bbox": [ + 55, + 601, + 295, + 673 + ], + "type": "text", + "content": "Setup and baselines. We compare CheXWorld with several competitive baselines across eight classification and segmentation tasks using the following datasets: VinDr-CXR [46], ShenZhen-CXR [36], NIH ChestX-ray14 [61], CheXpert [35], MedFMC-ChestDR [60], SIIM-ACR Pneumothorax [1] and RSNA Pneumonia [54]." + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 328, + 302, + 539, + 353 + ], + "blocks": [ + { + "bbox": [ + 333, + 290, + 534, + 300 + ], + "lines": [ + { + "bbox": [ + 333, + 290, + 534, + 300 + ], + "spans": [ + { + "bbox": [ + 333, + 290, + 534, + 300 + ], + "type": "text", + "content": "Table 4. Effectiveness of domain variation modeling." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 328, + 302, + 539, + 353 + ], + "lines": [ + { + "bbox": [ + 328, + 302, + 539, + 353 + ], + "spans": [ + { + "bbox": [ + 328, + 302, + 539, + 353 + ], + "type": "table", + "html": "
MethodRecall@1Recall@3Recall@5
Random Choice1.564.687.81
w/o domain condition15.8733.1244.37
CheXWorld38.4965.9077.67
", + "image_path": "ce56c3d13f4282de30e178b7c70311f75a21241855cdf8fe0a450db76bb8732d.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 354, + 555, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 354, + 555, + 450 + ], + "spans": [ + { + "bbox": [ + 313, + 354, + 555, + 450 + ], + "type": "text", + "content": "We attach task-specific heads (linear heads and U-Net decoders) and perform full fine-tuning on the pre-trained CheXWorld encoder. The input image resolution is fixed to " + }, + { + "bbox": [ + 313, + 354, + 555, + 450 + ], + "type": "inline_equation", + "content": "224^2" + }, + { + "bbox": [ + 313, + 354, + 555, + 450 + ], + "type": "text", + "content": ". We compare CheXWorld with several self-supervised learning methods designed for general domain images [6, 11, 14, 31, 64] and medical images [32, 44] using base-scale backbones [18, 40, 41]. We use the radiology version of these methods [42, 51, 63] if available." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 451, + 556, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 451, + 556, + 629 + ], + "spans": [ + { + "bbox": [ + 313, + 451, + 556, + 629 + ], + "type": "text", + "content": "Classification results are shown in Table 1. Our method consistently outperforms the best alternatives on all the benchmarks. Although SimMIM and Adam-v2 are initialized from ImageNet weights and pre-trained on a large dataset (0.9M X-rays) and, yet still lag behind CheXWorld with considerable margins. For instance, CheXWorld surpasses Adam-v2 by 4.04 in AUROC on VinDr-CXR and SimMIM by 0.54 on ChestX-ray14. Our approach performs comparably to Rad-DINO across all benchmarks, despite Rad-DINO requiring 20 times more computational resources and utilizing private training data. Additionally, one can observe a significant contrast in performance between models trained on photometric images (ImageNet) and models trained on medical images, highlighting the importance of in-domain transfer." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 630, + 555, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 555, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 555, + 689 + ], + "type": "text", + "content": "Segmentation results are presented in Table 2 and Figure 8. CheXWorld surpasses the leading baselines in terms of dice score and produces more accurate segmentation masks, demonstrating its superior capability in both instance-level and dense prediction settings." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": "Few-shot learning results on the MedFMC dataset are presented in Table 2. CheXWorld demonstrates a clear ad" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 55, + 683, + 295, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 683, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 683, + 295, + 713 + ], + "type": "text", + "content": "2In our experiments the candidate set size is fixed at " + }, + { + "bbox": [ + 55, + 683, + 295, + 713 + ], + "type": "inline_equation", + "content": "n = 64" + }, + { + "bbox": [ + 55, + 683, + 295, + 713 + ], + "type": "text", + "content": ". This process is repeated across multiple images, and the final result is obtained by averaging the outcomes." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20784" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 77, + 109, + 385, + 190 + ], + "blocks": [ + { + "bbox": [ + 74, + 72, + 385, + 105 + ], + "lines": [ + { + "bbox": [ + 74, + 72, + 385, + 105 + ], + "spans": [ + { + "bbox": [ + 74, + 72, + 385, + 105 + ], + "type": "text", + "content": "Table 2. Results on segmentation (left) and few-shot learning (right) tasks. The dice score and the AUROC score are reported for the segmentation and few-shot learning benchmarks respectively." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 77, + 109, + 385, + 190 + ], + "lines": [ + { + "bbox": [ + 77, + 109, + 385, + 190 + ], + "spans": [ + { + "bbox": [ + 77, + 109, + 385, + 190 + ], + "type": "table", + "html": "
MethodSegmentationFew-shot (MedFMC-ChestDR)
SIIM-ACRRSNA (SEG)1-shot5-shot10-shot
LVM-Med [44]82.19±0.3078.47±0.2457.55±0.8166.95±0.6067.44±0.71
MAE [31, 63]83.01±0.3977.39±0.2561.31±0.2774.03±0.3075.26±0.36
SimMIM [42, 64]82.89±0.0778.31±0.4160.64±1.9272.14±0.5474.42±0.38
Adam-v2 [32]83.60±0.3378.53±0.1959.16±1.0370.28±0.3970.67±0.71
CheXWorld84.58±0.3479.02±0.2564.60±1.0075.19±0.5176.40±0.25
", + "image_path": "9180052ec78c08af70339ee4350518ed141c1c2537ade1dc812736284a503891.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 396, + 72, + 529, + 164 + ], + "blocks": [ + { + "bbox": [ + 396, + 72, + 529, + 164 + ], + "lines": [ + { + "bbox": [ + 396, + 72, + 529, + 164 + ], + "spans": [ + { + "bbox": [ + 396, + 72, + 529, + 164 + ], + "type": "image", + "image_path": "aa90d4e8eb98306fc27d669e2cd72040d0b5e7a2cfbff6ccd8b9336516d91b68.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 390, + 167, + 535, + 189 + ], + "lines": [ + { + "bbox": [ + 390, + 167, + 535, + 189 + ], + "spans": [ + { + "bbox": [ + 390, + 167, + 535, + 189 + ], + "type": "text", + "content": "Figure 7. Fine-tuning with " + }, + { + "bbox": [ + 390, + 167, + 535, + 189 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 390, + 167, + 535, + 189 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 390, + 167, + 535, + 189 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 390, + 167, + 535, + 189 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 390, + 167, + 535, + 189 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 390, + 167, + 535, + 189 + ], + "type": "text", + "content": " training data on VinDr-CXR." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 71, + 198, + 216, + 323 + ], + "blocks": [ + { + "bbox": [ + 71, + 198, + 216, + 323 + ], + "lines": [ + { + "bbox": [ + 71, + 198, + 216, + 323 + ], + "spans": [ + { + "bbox": [ + 71, + 198, + 216, + 323 + ], + "type": "image", + "image_path": "246c7835cd0224b1b6afc476f33c92a363d0667afee956e86dd0b7963ab13b18.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 59, + 328, + 228, + 361 + ], + "lines": [ + { + "bbox": [ + 59, + 328, + 228, + 361 + ], + "spans": [ + { + "bbox": [ + 59, + 328, + 228, + 361 + ], + "type": "text", + "content": "Figure 8. Visualization of segmentation masks on SIIM-ACR pneumothorax dataset. GT stands for the ground truth masks." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 254, + 266, + 526, + 364 + ], + "blocks": [ + { + "bbox": [ + 227, + 196, + 550, + 262 + ], + "lines": [ + { + "bbox": [ + 227, + 196, + 550, + 262 + ], + "spans": [ + { + "bbox": [ + 227, + 196, + 550, + 262 + ], + "type": "text", + "content": "Table 3. Ablation studies of the world modeling tasks and latent variables. We report the results on VinDr-CXR and RSNA with " + }, + { + "bbox": [ + 227, + 196, + 550, + 262 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 227, + 196, + 550, + 262 + ], + "type": "text", + "content": " proportion of training data. AUROC and accuracy are reported for the respective datasets. \"Local,\" \"Global,\" and \"Domain\" are abbreviations for local anatomical structure modeling, global anatomical layout modeling, and domain variation modeling, respectively. \"Unified\" indicates whether our unified formulation is applied, as opposed to a naive combination of losses." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 254, + 266, + 526, + 364 + ], + "lines": [ + { + "bbox": [ + 254, + 266, + 526, + 364 + ], + "spans": [ + { + "bbox": [ + 254, + 266, + 526, + 364 + ], + "type": "table", + "html": "
World Modeling TaskLatent Variables Δx→y aUnifiedVinDr 1%RSNA 1%
Local.Global.Domain.
✓N.A.84.71±1.0065.89±0.54
✓✓✓✓88.31±0.8867.47±0.51
✓✓✓89.73±0.5068.27±1.00
✓✓✓✓✓90.53±1.0170.01±0.33
✓✓✓✓✓89.12±0.7668.37±0.50
✓✓✓✓✓89.91±0.6068.98±0.60
✓✓✓✓87.60±0.7266.56±0.39
", + "image_path": "0e8a526b7479675368b2ddd95d5a3069833e70be9c5d409125530f07767b46c5.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 372, + 295, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 372, + 295, + 420 + ], + "spans": [ + { + "bbox": [ + 55, + 372, + 295, + 420 + ], + "type": "text", + "content": "vantage on the 1-shot learning task and consistently outperforms the second-best methods in both 5-shot and 10-shot scenarios, showcasing quick adaptation to new concepts by leveraging knowledge acquired through world modeling." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 420, + 296, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 420, + 296, + 506 + ], + "spans": [ + { + "bbox": [ + 55, + 420, + 296, + 506 + ], + "type": "text", + "content": "Sample efficiency results. The superiority of our method is further highlighted by reducing the training data proportion. As shown in Figure 7, we conduct experiments on VinDr-CXR with " + }, + { + "bbox": [ + 55, + 420, + 296, + 506 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 55, + 420, + 296, + 506 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 420, + 296, + 506 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 55, + 420, + 296, + 506 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 420, + 296, + 506 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 55, + 420, + 296, + 506 + ], + "type": "text", + "content": " of the training data. Notably, our method with " + }, + { + "bbox": [ + 55, + 420, + 296, + 506 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 55, + 420, + 296, + 506 + ], + "type": "text", + "content": " of the data significantly outperforms all baselines trained on the entire dataset, reducing the annotation cost by over 90 percent." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 510, + 149, + 523 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 149, + 523 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 149, + 523 + ], + "type": "text", + "content": "5.3. Ablation Study" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 526, + 295, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 526, + 295, + 622 + ], + "spans": [ + { + "bbox": [ + 55, + 526, + 295, + 622 + ], + "type": "text", + "content": "Effectiveness of the world modeling tasks is validated in the upper part of Table 3. Combining local anatomical structure modeling with either of the other two world modeling tasks yields significant performance gains. The best performance is achieved when all three world modeling tasks are integrated within our unified framework, indicating that these tasks reflect different aspects of medical knowledge that are effectively captured by CheXWorld." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 622, + 296, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 622, + 296, + 694 + ], + "spans": [ + { + "bbox": [ + 55, + 622, + 296, + 694 + ], + "type": "text", + "content": "Role of the latent variables. To verify that the latent variables (such as relative position " + }, + { + "bbox": [ + 55, + 622, + 296, + 694 + ], + "type": "inline_equation", + "content": "\\Delta_{x\\rightarrow y}" + }, + { + "bbox": [ + 55, + 622, + 296, + 694 + ], + "type": "text", + "content": " for global anatomical layout modeling and transformation parameter " + }, + { + "bbox": [ + 55, + 622, + 296, + 694 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 55, + 622, + 296, + 694 + ], + "type": "text", + "content": " for domain variation modeling) contribute to transfer learning performance, we test the models pre-trained without certain latent variables, shown in the bottom part of Table" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 372, + 555, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 372, + 555, + 445 + ], + "spans": [ + { + "bbox": [ + 313, + 372, + 555, + 445 + ], + "type": "text", + "content": "3. Domain variation modeling without the condition " + }, + { + "bbox": [ + 313, + 372, + 555, + 445 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 313, + 372, + 555, + 445 + ], + "type": "text", + "content": " degrades into a blind image restoration task, which yields inferior results compared with the original method. Removing the condition " + }, + { + "bbox": [ + 313, + 372, + 555, + 445 + ], + "type": "inline_equation", + "content": "\\Delta_{x \\to y}" + }, + { + "bbox": [ + 313, + 372, + 555, + 445 + ], + "type": "text", + "content": " from global anatomical layout modeling leads to a sharp decrease in performance because it confounds the model with erroneous target spatial locations." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 456, + 388, + 468 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 456, + 388, + 468 + ], + "spans": [ + { + "bbox": [ + 313, + 456, + 388, + 468 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 476, + 555, + 633 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 476, + 555, + 633 + ], + "spans": [ + { + "bbox": [ + 313, + 476, + 555, + 633 + ], + "type": "text", + "content": "In this paper, we proposed CheXWorld, a self-supervised world modeling framework for radiographic images that simultaneously encodes local and global anatomical knowledge and appearance domain variations. By integrating three world modeling tasks into a unified pipeline, our model effectively captures the spatial and domain uncertainties of the image world, as verified by qualitative and quantitative analytical experiments. Extensive transfer learning experiments on eight medical image analysis benchmarks demonstrate the state-of-the-art performance of our model. Our work offers new insights into representing and extracting knowledge from medical images, paving the way toward a general-purpose foundation model for medical vision." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 645, + 419, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 645, + 419, + 658 + ], + "spans": [ + { + "bbox": [ + 313, + 645, + 419, + 658 + ], + "type": "text", + "content": "Acknowledgements." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 665, + 555, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 555, + 712 + ], + "type": "text", + "content": "The work is supported in part by the National Key R&D Program of China under Grant 2024YFB4708200 and National Natural Science Foundation of China under Grant 62321005." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 66, + 702, + 253, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 702, + 253, + 713 + ], + "spans": [ + { + "bbox": [ + 66, + 702, + 253, + 713 + ], + "type": "text", + "content": "3Detailed numerical results can be found in the appendix." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20785" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 133 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 133 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 133 + ], + "type": "text", + "content": "[1] Zawacki Anna, Wu Carol, Shih George, Elliott Julia, Fomitchev Mikhail, Hussain Mohannad, Lakhani Paras, Culliton Phil, and Bao Shunxing. Siim-acr pneumothorax segmentation. 2019. 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 136, + 296, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 136, + 296, + 201 + ], + "spans": [ + { + "bbox": [ + 61, + 136, + 296, + 201 + ], + "type": "text", + "content": "[2] Mahmoud Assran, Quentin Duval, Ishan Misra, Piotr Bojanowski, Pascal Vincent, Michael Rabbat, Yann LeCun, and Nicolas Ballas. Self-supervised learning from images with a joint-embedding predictive architecture. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15619-15629, 2023. 1, 3, 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 202, + 295, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 202, + 295, + 267 + ], + "spans": [ + { + "bbox": [ + 62, + 202, + 295, + 267 + ], + "type": "text", + "content": "[3] Shekoofeh Azizi, Basil Mustafa, Fiona Ryan, Zachary Beaver, Jan Freyberg, Jonathan Deaton, Aaron Loh, Alan Karthikesalingam, Simon Kornblith, Ting Chen, et al. Big self-supervised models advance medical image classification. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3478-3488, 2021. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 269, + 294, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 269, + 294, + 323 + ], + "spans": [ + { + "bbox": [ + 62, + 269, + 294, + 323 + ], + "type": "text", + "content": "[4] Alexei Baevski, Wei-Ning Hsu, Qiantong Xu, Arun Babu, Jiatao Gu, and Michael Auli. Data2vec: A general framework for self-supervised learning in speech, vision and language. In International Conference on Machine Learning, pages 1298-1312. PMLR, 2022. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 324, + 295, + 378 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 324, + 295, + 378 + ], + "spans": [ + { + "bbox": [ + 62, + 324, + 295, + 378 + ], + "type": "text", + "content": "[5] Alexei Baevski, Arun Babu, Wei-Ning Hsu, and Michael Auli. Efficient self-supervised learning with contextualized target representations for vision, speech and language. In International Conference on Machine Learning, pages 1416-1429. PMLR, 2023. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 380, + 294, + 412 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 380, + 294, + 412 + ], + "spans": [ + { + "bbox": [ + 62, + 380, + 294, + 412 + ], + "type": "text", + "content": "[6] Hangbo Bao, Li Dong, Songhao Piao, and Furu Wei. Beit: Bert pre-training of image transformers. arXiv preprint arXiv:2106.08254, 2021.7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 414, + 294, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 414, + 294, + 456 + ], + "spans": [ + { + "bbox": [ + 62, + 414, + 294, + 456 + ], + "type": "text", + "content": "[7] Adrien Bardes, Quentin Garrido, Jean Ponce, Xinlei Chen, Michael Rabbat, Yann LeCun, Mido Assran, and Nicolas Ballas. V-jepa: Latent video prediction for visual representation learning. 2023. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 458, + 294, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 458, + 294, + 500 + ], + "spans": [ + { + "bbox": [ + 62, + 458, + 294, + 500 + ], + "type": "text", + "content": "[8] Florian Bordes, Randall Balestriero, and Pascal Vincent. High fidelity visualization of what your self-supervised representation knows about. arXiv preprint arXiv:2112.09164, 2021. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 502, + 294, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 502, + 294, + 524 + ], + "spans": [ + { + "bbox": [ + 62, + 502, + 294, + 524 + ], + "type": "text", + "content": "[9] Arthur Earl Bryson. Applied optimal control: optimization, estimation and control. Routledge, 2018. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 525, + 294, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 525, + 294, + 557 + ], + "spans": [ + { + "bbox": [ + 57, + 525, + 294, + 557 + ], + "type": "text", + "content": "[10] Eduardo F Camacho, Carlos Bordons, Eduardo F Camacho, and Carlos Bordons. Constrained model predictive control. Springer, 2007. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 559, + 294, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 559, + 294, + 613 + ], + "spans": [ + { + "bbox": [ + 57, + 559, + 294, + 613 + ], + "type": "text", + "content": "[11] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 614, + 294, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 614, + 294, + 667 + ], + "spans": [ + { + "bbox": [ + 57, + 614, + 294, + 667 + ], + "type": "text", + "content": "[12] Krishna Chaitanya, Ertunc Erdil, Neerav Karani, and Ender Konukoglu. Contrastive learning of global and local features for medical image segmentation with limited annotations. Advances in neural information processing systems, 33:12546-12558, 2020. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 670, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 294, + 713 + ], + "type": "text", + "content": "[13] Richard J Chen, Tong Ding, Ming Y Lu, Drew FK Williamson, Guillaume Jaume, Andrew H Song, Bowen Chen, Andrew Zhang, Daniel Shao, Muhammad Shaban, et al. Towards a general-purpose foundation model for" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 317, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "type": "text", + "content": "computational pathology. Nature Medicine, 30(3):850-862, 2024. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 97, + 553, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 97, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 317, + 97, + 553, + 140 + ], + "type": "text", + "content": "[14] Xinlei Chen, Saining Xie, and Kaiming He. An empirical study of training self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9640-9649, 2021. 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 143, + 553, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 143, + 553, + 185 + ], + "spans": [ + { + "bbox": [ + 317, + 143, + 553, + 185 + ], + "type": "text", + "content": "[15] Peng Cui, Yang Yue, Zhijie Deng, and Jun Zhu. Confidence-based reliable learning under dual noises. Advances in Neural Information Processing Systems, 35:35116-35129, 2022. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 188, + 553, + 232 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 188, + 553, + 232 + ], + "spans": [ + { + "bbox": [ + 317, + 188, + 553, + 232 + ], + "type": "text", + "content": "[16] Rumen Dangovski, Li Jing, Charlotte Loh, Seungwook Han, Akash Srivastava, Brian Cheung, Pulkit Agrawal, and Marin Soljacic. Equivariant contrastive learning. arXiv preprint arXiv:2111.00899, 2021. 3, 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 234, + 553, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 234, + 553, + 277 + ], + "spans": [ + { + "bbox": [ + 317, + 234, + 553, + 277 + ], + "type": "text", + "content": "[17] Alexandre Devillers and Mathieu Lefort. Equimod: An equivariance module to improve visual instance discrimination. In International Conference on Learning Representations, 2023. 3, 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 280, + 553, + 345 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 280, + 553, + 345 + ], + "spans": [ + { + "bbox": [ + 317, + 280, + 553, + 345 + ], + "type": "text", + "content": "[18] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 3, 6, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 347, + 553, + 391 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 347, + 553, + 391 + ], + "spans": [ + { + "bbox": [ + 317, + 347, + 553, + 391 + ], + "type": "text", + "content": "[19] Chaoqun Du, Yulin Wang, Shiji Song, and Gao Huang. Probabilistic contrastive learning for long-tailed visual recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 393, + 553, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 393, + 553, + 448 + ], + "spans": [ + { + "bbox": [ + 317, + 393, + 553, + 448 + ], + "type": "text", + "content": "[20] Cong Gao, Benjamin D Killeen, Yicheng Hu, Robert B Grupp, Russell H Taylor, Mehran Armand, and Mathias Unberath. Synthetic data accelerates the development of generalizable learning-based algorithms for x-ray image analysis. Nature Machine Intelligence, 5(3):294-308, 2023. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 449, + 553, + 493 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 449, + 553, + 493 + ], + "spans": [ + { + "bbox": [ + 317, + 449, + 553, + 493 + ], + "type": "text", + "content": "[21] Quentin Garrido, Mahmoud Assran, Nicolas Ballas, Adrien Bardes, Laurent Najman, and Yann LeCun. Learning and leveraging world models in visual representation learning. arXiv preprint arXiv:2403.00504, 2024. 1, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 495, + 553, + 550 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 495, + 553, + 550 + ], + "spans": [ + { + "bbox": [ + 317, + 495, + 553, + 550 + ], + "type": "text", + "content": "[22] Ross Girshick, Jeff Donahue, Trevor Darrell, and Jitendra Malik. Rich feature hierarchies for accurate object detection and semantic segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 580-587, 2014. 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 552, + 553, + 585 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 552, + 553, + 585 + ], + "spans": [ + { + "bbox": [ + 317, + 552, + 553, + 585 + ], + "type": "text", + "content": "[23] Hao Guan and Mingxia Liu. Domain adaptation for medical image analysis: a survey. IEEE Transactions on Biomedical Engineering, 69(3):1173-1185, 2021. 2, 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 587, + 553, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 587, + 553, + 620 + ], + "spans": [ + { + "bbox": [ + 317, + 587, + 553, + 620 + ], + "type": "text", + "content": "[24] David Ha and Jürgen Schmidhuber. Recurrent world models facilitate policy evolution. Advances in neural information processing systems, 31, 2018. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 622, + 553, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 622, + 553, + 643 + ], + "spans": [ + { + "bbox": [ + 317, + 622, + 553, + 643 + ], + "type": "text", + "content": "[25] David Ha and Jürgen Schmidhuber. World models. arXiv preprint arXiv:1803.10122, 2018. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 646, + 553, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 646, + 553, + 678 + ], + "spans": [ + { + "bbox": [ + 317, + 646, + 553, + 678 + ], + "type": "text", + "content": "[26] Danijar Hafner, Timothy Lillicrap, Jimmy Ba, and Mohammad Norouzi. Dream to control: Learning behaviors by latent imagination. arXiv preprint arXiv:1912.01603, 2019." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 317, + 680, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 680, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 680, + 553, + 713 + ], + "type": "text", + "content": "[27] Danijar Hafner, Timothy Lillicrap, Mohammad Norouzi, and Jimmy Ba. Mastering atari with discrete world models. arXiv preprint arXiv:2010.02193, 2020. 2" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "20786" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "type": "text", + "content": "[28] Danijar Hafner, Jurgis Pasukonis, Jimmy Ba, and Timothy Lillicrap. Mastering diverse domains through world models. arXiv preprint arXiv:2301.04104, 2023. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 295, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 295, + 163 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 295, + 163 + ], + "type": "text", + "content": "[29] Fatemeh Haghighi, Mohammad Reza Hosseinzadeh Taher, Zongwei Zhou, Michael B Gotway, and Jianming Liang. Transferable visual words: Exploiting the semantics of anatomical patterns for self-supervised learning. IEEE transactions on medical imaging, 40(10):2857-2868, 2021. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 164, + 294, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 164, + 294, + 229 + ], + "spans": [ + { + "bbox": [ + 56, + 164, + 294, + 229 + ], + "type": "text", + "content": "[30] Fatemeh Haghighi, Mohammad Reza Hosseinzadeh Taher, Michael B Gotway, and Jianming Liang. Dira: Discriminative, restorative, and adversarial learning for self-supervised medical image analysis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20824-20834, 2022. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 231, + 294, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 231, + 294, + 285 + ], + "spans": [ + { + "bbox": [ + 56, + 231, + 294, + 285 + ], + "type": "text", + "content": "[31] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 3, 4, 7, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 287, + 294, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 287, + 294, + 352 + ], + "spans": [ + { + "bbox": [ + 56, + 287, + 294, + 352 + ], + "type": "text", + "content": "[32] Mohammad Reza Hosseinzadeh Taher, Michael Gotway, and Jianming Liang. Representing part-whole hierarchies in foundation models by learning localizability, composability, and decomposability from anatomy via self-supervision. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2024. 3, 7, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 354, + 294, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 354, + 294, + 408 + ], + "spans": [ + { + "bbox": [ + 56, + 354, + 294, + 408 + ], + "type": "text", + "content": "[33] Anthony Hu, Lloyd Russell, Hudson Yeo, Zak Murez, George Fedoseev, Alex Kendall, Jamie Shotton, and Gianluca Corrado. Gaia-1: A generative world model for autonomous driving. arXiv preprint arXiv:2309.17080, 2023. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 411, + 294, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 411, + 294, + 453 + ], + "spans": [ + { + "bbox": [ + 56, + 411, + 294, + 453 + ], + "type": "text", + "content": "[34] Zhi Huang, Federico Bianchi, Mert Yuksekgonul, Thomas J Montine, and James Zou. A visual-language foundation model for pathology image analysis using medical twitter. Nature medicine, 29(9):2307-2316, 2023. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 456, + 294, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 456, + 294, + 521 + ], + "spans": [ + { + "bbox": [ + 56, + 456, + 294, + 521 + ], + "type": "text", + "content": "[35] Jeremy Irvin, Pranav Rajpurkar, Michael Ko, Yifan Yu, Silviana Ciurea-Ilicus, Chris Chute, Henrik Marklund, Behzad Haghgoo, Robyn Ball, Katie Shpanskaya, et al. Chexpert: A large chest radiograph dataset with uncertainty labels and expert comparison. In Proceedings of the AAAI conference on artificial intelligence, pages 590-597, 2019. 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 522, + 294, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 522, + 294, + 577 + ], + "spans": [ + { + "bbox": [ + 56, + 522, + 294, + 577 + ], + "type": "text", + "content": "[36] Stefan Jaeger, Sema Candemir, Sameer Antani, Yi-Xiang J Wang, Pu-Xuan Lu, and George Thoma. Two public chest x-ray datasets for computer-aided screening of pulmonary diseases. Quantitative imaging in medicine and surgery, 4(6): 475, 2014. 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 578, + 294, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 578, + 294, + 623 + ], + "spans": [ + { + "bbox": [ + 56, + 578, + 294, + 623 + ], + "type": "text", + "content": "[37] Bingyi Kang, Yang Yue, Rui Lu, Zhijie Lin, Yang Zhao, Kaixin Wang, Gao Huang, and Jiashi Feng. How far is video generation from world model: A physical law perspective. arXiv preprint arXiv:2411.02385, 2024. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 624, + 294, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 294, + 678 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 294, + 678 + ], + "type": "text", + "content": "[38] Dan Kondratyuk, Lijun Yu, Xiuye Gu, José Lezama, Jonathan Huang, Rachel Hornung, Hartwig Adam, Hassan Akbari, Yair Alon, Vighnesh Birodkar, et al. Videopoet: A large language model for zero-shot video generation. arXiv preprint arXiv:2312.14125, 2023. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 680, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 680, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 680, + 294, + 712 + ], + "type": "text", + "content": "[39] Yann LeCun. A path towards autonomous machine intelligence version 0.9.2, 2022-06-27. Open Review, 62(1), 2022. 1, 2, 3" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 128 + ], + "type": "text", + "content": "[40] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 129, + 553, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 553, + 182 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 553, + 182 + ], + "type": "text", + "content": "[41] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11976-11986, 2022. 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 184, + 553, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 184, + 553, + 249 + ], + "spans": [ + { + "bbox": [ + 316, + 184, + 553, + 249 + ], + "type": "text", + "content": "[42] DongAo Ma, Mohammad Reza Hosseinzadeh Taher, Jiaxuan Pang, Nahid UI Islam, Fatemeh Haghighi, Michael B Gotway, and Jianming Liang. Benchmarking and boosting transformers for medical image classification. In MICCAI Workshop on Domain Adaptation and Representation Transfer, pages 12-22. Springer, 2022. 7, 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 251, + 553, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 251, + 553, + 282 + ], + "spans": [ + { + "bbox": [ + 316, + 251, + 553, + 282 + ], + "type": "text", + "content": "[43] Phillip Marsh. Case courtesy of phillip marsh, radiopaedia.org. https://radiopaedia.org/cases/58938, 2023. Case ID: rID: 58938. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 283, + 553, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 283, + 553, + 348 + ], + "spans": [ + { + "bbox": [ + 316, + 283, + 553, + 348 + ], + "type": "text", + "content": "[44] Duy MH Nguyen, Hoang Nguyen, Nghiem Diep, Tan Ngoc Pham, Tri Cao, Binh Nguyen, Paul Swoboda, Nhat Ho, Shadi Albarqouni, Pengtao Xie, et al. Lvm-med: Learning large-scale self-supervised vision models for medical imaging via second-order graph matching. Advances in Neural Information Processing Systems, 36, 2024. 7, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 350, + 553, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 350, + 553, + 402 + ], + "spans": [ + { + "bbox": [ + 316, + 350, + 553, + 402 + ], + "type": "text", + "content": "[45] Michael Moor, Oishi Banerjee, Zahra Shakeri Hossein Abad, Harlan M Krumholz, Jure Leskovec, Eric J Topol, and Pranav Rajpurkar. Foundation models for generalist medical artificial intelligence. Nature, 616(7956):259-265, 2023. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 404, + 553, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 404, + 553, + 458 + ], + "spans": [ + { + "bbox": [ + 316, + 404, + 553, + 458 + ], + "type": "text", + "content": "[46] Ha Q Nguyen, Khanh Lam, Linh T Le, Hieu H Pham, Dat Q Tran, Dung B Nguyen, Dung D Le, Chi M Pham, Hang TT Tong, Diep H Dinh, et al. Vindr-cxr: An open dataset of chest x-rays with radiologist's annotations. Scientific Data, 9(1):429, 2022. 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 460, + 553, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 460, + 553, + 502 + ], + "spans": [ + { + "bbox": [ + 316, + 460, + 553, + 502 + ], + "type": "text", + "content": "[47] OpenAI. Video generation models as world simulators. https://openai.com/index/video-generation-models-as-world-simulators/, 2024. Accessed: Feb, 2024. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 504, + 553, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 504, + 553, + 568 + ], + "spans": [ + { + "bbox": [ + 316, + 504, + 553, + 568 + ], + "type": "text", + "content": "[48] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. Transactions on Machine Learning Research Journal, pages 1-31, 2024. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 570, + 553, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 570, + 553, + 623 + ], + "spans": [ + { + "bbox": [ + 316, + 570, + 553, + 623 + ], + "type": "text", + "content": "[49] Cheng Ouyang, Chen Chen, Surui Li, Zeju Li, Chen Qin, Wenjia Bai, and Daniel Rueckert. Causality-inspired single-source domain generalization for medical image segmentation. IEEE Transactions on Medical Imaging, 42(4):1095-1106, 2022. 2, 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 625, + 553, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 625, + 553, + 690 + ], + "spans": [ + { + "bbox": [ + 316, + 625, + 553, + 690 + ], + "type": "text", + "content": "[50] Jiaxuan Pang, Fatemeh Haghighi, DongAo Ma, Nahid Ul Islam, Mohammad Reza Hosseinzadeh Taher, Michael B Gotway, and Jianming Liang. Popar: Patch order prediction and appearance recovery for self-supervised medical image analysis. In MICCAI Workshop on Domain Adaptation and Representation Transfer, pages 77-87. Springer, 2022. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 692, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 553, + 712 + ], + "type": "text", + "content": "[51] Fernando Pérez-García, Harshita Sharma, Sam Bond-Taylor, Kenya Bouzid, Valentina Salvatelli, Maximilian Ilse, Shruthi" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20787" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 76, + 72, + 296, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 296, + 116 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 296, + 116 + ], + "type": "text", + "content": "Bannur, Daniel C Castro, Anton Schwaighofer, Matthew P Lungren, et al. Rad-dino: Exploring scalable medical image encoders beyond text supervision. arXiv preprint arXiv:2401.10815, 2024. 3, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 295, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 295, + 161 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 295, + 161 + ], + "type": "text", + "content": "[52] Eric Postal. Why don't radiology textbooks have imperfect images? https://www.diagnosticimaging.com/view/why-dont-radiology-textbooks-have-imperfect-images, 2019. Accessed: Feb, 2024. 2, 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 162, + 294, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 162, + 294, + 204 + ], + "spans": [ + { + "bbox": [ + 56, + 162, + 294, + 204 + ], + "type": "text", + "content": "[53] RunwayML. Introducing general world models. https://research.runwayml.com/introducing-general-world-models, 2023. Accessed: May, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 205, + 294, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 205, + 294, + 281 + ], + "spans": [ + { + "bbox": [ + 56, + 205, + 294, + 281 + ], + "type": "text", + "content": "[54] George Shih, Carol C Wu, Safwan S Halabi, Marc D Kohli, Luciano M Prevedello, Tessa S Cook, Arjun Sharma, Judith K Amorosa, Veronica Arteaga, Maya Galperin-Aizenberg, et al. Augmenting the national institutes of health chest radiograph dataset with expert annotations of possible pneumonia. Radiology: Artificial Intelligence, 1(1): e180041, 2019. 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 283, + 294, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 283, + 294, + 326 + ], + "spans": [ + { + "bbox": [ + 56, + 283, + 294, + 326 + ], + "type": "text", + "content": "[55] Ben Sorscher, Surya Ganguli, and Haim Sompolinsky. Neural representational geometry underlies few-shot concept learning. Proceedings of the National Academy of Sciences, 119(43):e2200800119, 2022. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 327, + 294, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 327, + 294, + 371 + ], + "spans": [ + { + "bbox": [ + 56, + 327, + 294, + 371 + ], + "type": "text", + "content": "[56] Hari Sowrirajan, Jingbo Yang, Andrew Y Ng, and Pranav Rajpurkar. Moco pretraining improves representation and transferability of chest x-ray models. In Medical Imaging with Deep Learning, pages 728-744. PMLR, 2021. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 372, + 294, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 372, + 294, + 426 + ], + "spans": [ + { + "bbox": [ + 56, + 372, + 294, + 426 + ], + "type": "text", + "content": "[57] Mohammad Reza Hosseinzadeh Taher, Fatemeh Haghighi, Michael B Gotway, and Jianming Liang. Caid: a self-supervised learning framework for empowering instance discrimination in medical imaging. Proceedings of Machine Learning Research, 3, 2022. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 427, + 294, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 427, + 294, + 503 + ], + "spans": [ + { + "bbox": [ + 56, + 427, + 294, + 503 + ], + "type": "text", + "content": "[58] Xing Tao, Yuexiang Li, Wenhui Zhou, Kai Ma, and Yefeng Zheng. Revisiting rubik's cube: self-supervised learning with volume-wise transformation for 3d medical image segmentation. In Medical Image Computing and Computer Assisted Intervention-MICCAI 2020: 23rd International Conference, Lima, Peru, October 4-8, 2020, Proceedings, Part IV 23, pages 238-248. Springer, 2020. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 504, + 294, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 504, + 294, + 548 + ], + "spans": [ + { + "bbox": [ + 56, + 504, + 294, + 548 + ], + "type": "text", + "content": "[59] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 548, + 294, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 548, + 294, + 602 + ], + "spans": [ + { + "bbox": [ + 56, + 548, + 294, + 602 + ], + "type": "text", + "content": "[60] Dequan Wang, Xiaosong Wang, Lilong Wang, Mengzhang Li, Qian Da, Xiaoqiang Liu, Xiangyu Gao, Jun Shen, Junjun He, Tian Shen, et al. A real-world dataset and benchmark for foundation model adaptation in medical image classification. Scientific Data, 10(1):574, 2023. 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 604, + 294, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 604, + 294, + 679 + ], + "spans": [ + { + "bbox": [ + 56, + 604, + 294, + 679 + ], + "type": "text", + "content": "[61] Xiaosong Wang, Yifan Peng, Le Lu, Zhiyong Lu, Mohammadhadi Bagheri, and Ronald M Summers. Chestxray8: Hospital-scale chest x-ray database and benchmarks on weakly-supervised classification and localization of common thorax diseases. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2097-2106, 2017. 2, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 681, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 681, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 681, + 294, + 713 + ], + "type": "text", + "content": "[62] Yulin Wang, Yang Yue, Rui Lu, Tianjiao Liu, Zhao Zhong, Shiji Song, and Gao Huang. Efficienttrain: Exploring generalized curriculum learning for training visual backbones. In" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 541 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5852-5864, 2023. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 96, + 553, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 149 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 149 + ], + "type": "text", + "content": "[63] Junfei Xiao, Yutong Bai, Alan Yuille, and Zongwei Zhou. Delving into masked autoencoders for multi-label thorax disease classification. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 3588-3600, 2023. 3, 7, 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 152, + 555, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 555, + 206 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 555, + 206 + ], + "type": "text", + "content": "[64] Zhenda Xie, Zheng Zhang, Yue Cao, Yutong Lin, Jianmin Bao, Zhuliang Yao, Qi Dai, and Han Hu. Simmim: A simple framework for masked image modeling. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9653-9663, 2022. 7, 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 208, + 553, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 208, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 316, + 208, + 553, + 262 + ], + "type": "text", + "content": "[65] Ke Yan, Jinzheng Cai, Dakai Jin, Shun Miao, Dazhou Guo, Adam P Harrison, Youbao Tang, Jing Xiao, Jingjing Lu, and Le Lu. Sam: Self-supervised learning of pixel-wise anatomical embeddings in radiological images. IEEE Transactions on Medical Imaging, 41(10):2658-2669, 2022. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 263, + 553, + 340 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 263, + 553, + 340 + ], + "spans": [ + { + "bbox": [ + 316, + 263, + 553, + 340 + ], + "type": "text", + "content": "[66] Hong-Yu Zhou, Shuang Yu, Cheng Bian, Yifan Hu, Kai Ma, and Yefeng Zheng. Comparing to learn: Surpassing image-genet pretraining on radiographs by comparing image representations. In Medical Image Computing and Computer-Assisted Intervention-MICCAI 2020: 23rd International Conference, Lima, Peru, October 4-8, 2020, Proceedings, Part I 23, pages 398-407. Springer, 2020. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 342, + 553, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 342, + 553, + 396 + ], + "spans": [ + { + "bbox": [ + 316, + 342, + 553, + 396 + ], + "type": "text", + "content": "[67] Lei Zhou, Huidong Liu, Joseph Bae, Junjun He, Dimitris Samaras, and Prateek Prasanna. Self pre-training with masked autoencoders for medical image classification and segmentation. In 2023 IEEE 20th International Symposium on Biomedical Imaging (ISBI), pages 1-6. IEEE, 2023. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 397, + 553, + 461 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 397, + 553, + 461 + ], + "spans": [ + { + "bbox": [ + 316, + 397, + 553, + 461 + ], + "type": "text", + "content": "[68] Yukun Zhou, Mark A Chia, Siegfried K Wagner, Murat S Ayhan, Dominic J Williamson, Robbert R Struyven, Timing Liu, Moucheng Xu, Mateo G Lozano, Peter Woodward-Court, et al. A foundation model for generalizable disease detection from retinal images. Nature, 622(7981):156-163, 2023. 1, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 464, + 553, + 496 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 464, + 553, + 496 + ], + "spans": [ + { + "bbox": [ + 316, + 464, + 553, + 496 + ], + "type": "text", + "content": "[69] Zongwei Zhou, Vatsal Sodha, Jiaxuan Pang, Michael B Gotway, and Jianming Liang. Models genesis. Medical image analysis, 67:101840, 2021. 2, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 498, + 553, + 541 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 498, + 553, + 541 + ], + "spans": [ + { + "bbox": [ + 316, + 498, + 553, + 541 + ], + "type": "text", + "content": "[70] Ziyu Zhou, Haozhe Luo, Jiaxuan Pang, Xiaowei Ding, Michael Gotway, and Jianming Liang. Learning anatomically consistent embedding for chest radiography. arXiv preprint arXiv:2312.00335, 2023. 2, 3" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20788" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/CheXwhatsApp_ A Dataset for Exploring Challenges in the Diagnosis of Chest X-rays through Mobile Devices/8cf11147-8ad4-4f89-a3ee-6b5492c8beb0_content_list.json b/2025/CheXwhatsApp_ A Dataset for Exploring Challenges in the Diagnosis of Chest X-rays through Mobile Devices/8cf11147-8ad4-4f89-a3ee-6b5492c8beb0_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..a401a12cde36ca04ecb08ee9008b502aa4a81238 --- /dev/null +++ b/2025/CheXwhatsApp_ A Dataset for Exploring Challenges in the Diagnosis of Chest X-rays through Mobile Devices/8cf11147-8ad4-4f89-a3ee-6b5492c8beb0_content_list.json @@ -0,0 +1,1565 @@ +[ + { + "type": "text", + "text": "CheXwhatsapp: A Dataset for Exploring Challenges in the Diagnosis of Chest X-rays through Mobile Devices", + "text_level": 1, + "bbox": [ + 106, + 128, + 893, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Mariamma Antony1", + "bbox": [ + 148, + 202, + 313, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Rajiv Porana", + "bbox": [ + 354, + 202, + 467, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Sahil M Lathiya1", + "bbox": [ + 511, + 203, + 648, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Siva Teja Kakileti2", + "bbox": [ + 694, + 203, + 846, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Chiranjib Bhattacharyya", + "bbox": [ + 397, + 220, + 599, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Indian Institute of Science, Bangalore, India $^{2}$ Niramai Health Analytix, Bangalore, India", + "bbox": [ + 143, + 238, + 849, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 291, + 326, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Mobile health (mHealth) has emerged as a transformative solution to enhance healthcare accessibility and affordability, particularly in resource-constrained regions and low-to-middle-income countries. mHealth leverages mobile platforms to improve healthcare accessibility, addressing radiologist shortages in low-resource settings by enabling remote diagnosis and consultation through mobile devices. Mobile phones allow healthcare workers to transmit radiographic images, such as chest X-rays (CXR), to specialists or AI-driven models for interpretation. However, AI-based diagnosis using CXR images shared via apps like WhatsApp suffers from reduced predictability and explainability due to compression artifacts, and there is a lack of datasets to systematically study these challenges. To address this, we introduce CheXwhatsapp, a dataset of 141,804 paired original and WhatsApp-compressed CXR images. We present a benchmarking study which shows the dataset improves prediction stability and explainability of state-of-the-art models by up to $80\\%$ , while also enhancing localization performance. CheXwhatsapp is open-sourced to support advancements in mHealth applications for CXR analysis1.", + "bbox": [ + 88, + 321, + 485, + 642 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 669, + 220, + 686 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent studies highlight the wide opportunities of messaging apps such as WhatsApp as an efficient tool for remote image sharing in mHealth applications across digital healthcare, including digital radiology [1, 10], digital pathology [11-13], cardiac imaging [14], orthopedics [32], diabetes management [41], and oral health [8]. mHealth provides a platform for health practitioners to enable disease management and speed-up diagnosis by facilitating the transfer of medical images, providing advice, facilitating drug adherence, and monitoring outbreaks. The outbreak of COVID-19 pandemic has accelerated the adoption of mHealth for digital radiology, such as for analyzing Chest X-ray (CXR)", + "bbox": [ + 89, + 695, + 483, + 878 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "imaging.", + "bbox": [ + 511, + 292, + 576, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "mHealth is especially beneficial in resource constrained areas such as Low-to-Middle-Income countries (LMICs) where the ratio of radiologists to radiographers is significantly imbalanced. For example, in India, there is only one radiologist for every 100,000 people [2]. Compounding the issue is the fact that most radiologists are located in urban areas, often far from the majority of the X-ray imaging facilities located in rural areas. Due to the high demand for chest radiography and the global shortage of radiologists, numerous applications for AI-based diagnosis of CXR images have been developed.", + "bbox": [ + 511, + 308, + 906, + 474 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In recent years, smartphone penetration has significantly increased, and instant messaging applications such as WhatsApp have a large user-base with 535 million users in India, 118.5 million users in Brazil, and a penetration rate exceeding $97\\%$ in African nations like Kenya, South Africa, and Nigeria [15]. This presents new opportunities for the quick and easy sharing of radiographic images to aid clinical decision-making [3, 22, 23, 26, 32]. In rural hospitals and diagnostic centers without on-site radiologists, the ability to transmit anonymized images instantly to a mHealth app for interpretation can be extremely valuable to facilitate rapid triaging of infected individuals [22].", + "bbox": [ + 511, + 476, + 908, + 656 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Leveraging common messaging platforms like WhatsApp eliminates the need for additional training since it is a widely used communication medium. However, it is important to note that these messaging platforms often compress images by upto $90\\%$ to overcome bandwidth limitations. In medial applications, it is crucial to obtain explanation of predictions or localizations of findings in CXR images. Existing literature [1, 7, 9] identifies challenges related to explainability, such as predictions driven by features present outside the lung region in AI-based diagnostic systems for CXR images. Specifically, these systems suffer from the following challenges.", + "bbox": [ + 511, + 657, + 908, + 839 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Prediction Instability Problem(PIP): PIP refers to a situation where model's prediction differs between the original image and the image transmitted through WhatsApp. This undesirable effect can have critical implications in mHealth", + "bbox": [ + 511, + 839, + 908, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "1https://github.com/cheXwhatsApp/cheXwhatsApp", + "bbox": [ + 107, + 886, + 464, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "25887", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "applications.", + "bbox": [ + 89, + 90, + 178, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Out of Lung Saliency(OLS): In medical diagnosis, it is crucial to provide explainable predictions. Saliency maps, which identify localized regions in an image that significantly contribute to the prediction, are commonly used to provide explanations. Out of Lung Saliency refers to the scenario where the model uses regions outside the lungs as major contributors to its prediction. Figure 1 displays instances of saliency maps generated by state-of-the-art models on original and corresponding WhatsApp-compressed CXR image pairs. In each of these images, the saliency maps highlight regions outside the lungs as significant contributors to the prediction. This effect is even more pronounced in WhatsApp-compressed images as illustrated in Figure 1.", + "bbox": [ + 89, + 107, + 480, + 318 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Localization Instability Problem(LIP): LIP refers to a situation where an object detection model exhibits localizations which are different on the transmitted image when compared with the original image. Figure 2 displays several instances of localization instability between the original image and the corresponding WhatsApp compressed image. In each of these images, the localizations predicted by the model for the original image is either absent or incorrectly placed in the WhatsApp-compressed image.", + "bbox": [ + 89, + 321, + 480, + 455 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Current methods for understanding the variations due to image perturbations are primarily visual. In this paper, we discuss and propose PIP, OLS, and LIP to quantitatively study the changes in predictions, saliency, and localizations. While several databases of CXR images have been released, most include only disease diagnosis labels or bounding boxes indicating regions of interest. However, a dataset that illustrates the aforementioned challenges is not yet publicly available.", + "bbox": [ + 89, + 458, + 480, + 594 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Contributions: Our contributions are summarized as follows:", + "bbox": [ + 89, + 597, + 482, + 626 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce CheXwhatsApp, the first paired dataset of 141,804 pairs of original and WhatsApp-compressed CXR images, which can be used for understanding the effect of WhatsApp on localization of various thoracic diseases.", + "- We identify that prediction of state-of-the-art (SOTA) AI models on CXR images transmitted via low-bandwidth messaging applications such as WhatsApp are affected by prediction and explainability challenges. We find that several classification models exhibit the problem of PIP, where predictions differ between the original and transmitted images. While this is a concern, but even more worrisome is many of them also exhibit OLS where predictions are driven by regions outside the lung area, raising important questions on the explainability of the models. Similarly, models that localize lung abnormalities using bounding boxes exhibit the problem of LIP. We contribute several metrics to evaluate PIP, OLS and LIP." + ], + "bbox": [ + 91, + 628, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Benchmark results indicate that models trained on open-source CXR images can suffer from high PIP, OLS and LIP. However, models trained on CheXwhat-sApp, which includes both original and WhatsApp-compressed CXR images, reduces the effect of PIP, OLS and LIP by upto $80\\%$ , compared to models trained on open-source CXR images.", + "- We open-source the dataset and evaluation pipeline for evaluation of PIP, OLS and LIP in models on CheXwhat-sApp, to streamline the development and analysis of AI-based diagnostic systems for mHealth." + ], + "bbox": [ + 513, + 90, + 903, + 257 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related works", + "text_level": 1, + "bbox": [ + 513, + 271, + 658, + 286 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "There are several publicly available large-scale CXR datasets. The NIH dataset [38] is one of the first, containing 112,120 CXR images with labels from 30,805 unique patients. It is a multi-label dataset with 14 labels. Similar to NIH, the Chexpert dataset [18] has 224316 train and 234 validation images, both labeled with 14 categories. The Chexpert dataset was obtained from 65,240 unique patients at Stanford Hospital. The CovidNet [37] dataset contains 30,130 train and 400 testing images categorized as COVID-19, pneumonia, or normal. The MIMIC-CXR dataset [21] includes 227,835 CXR imaging studies for 65,379 patients, along with radiology reports from Beth Israel. ChestX-ray8 [39] is a CXR dataset that consists of 108,948 frontalview X-ray images of 32,717 unique patients with eight disease labels extracted through text mining. PadChest [4] contains 160,000 CXR images obtained from 67,000 patients in San Juan Hospital (Spain). BRAX [29] is a Brazilian labeled CXR dataset consisting of 24,959 CXR images annotated with fourteen labels derived from radiology reports using NLP.", + "bbox": [ + 511, + 296, + 903, + 595 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "RSNA [34] and VinDr-CXR [25] are two publicly available CXR datasets with bounding box annotations. The RSNA dataset [34] consists of CXR images of patients infected with pneumonia and normal cases. The pneumonia findings are localized by bounding boxes. The VinDr CXR dataset [25] contains 15,000 training images and 3,000 testing images with critical findings localized by bounding boxes. The CXR images are annotated for the presence of 22 different lung abnormalities. CheXphoto [27] comprises a training set of natural photos and synthetic transformations of 10,507 CXR's from 3,000 unique patients that were sampled at random from the CheXpert training set. It also includes a validation set of natural and synthetic transformations of 234 CXR's from 200 patients in the CheXpert validation set. In addition to these large datasets, there are also several smaller CXR datasets with a approximately 1,000 images, annotated either programmatically or manually by radiologists. However, the literature does not contain a large dataset with original and corresponding WhatsApp-compressed CXR images.", + "bbox": [ + 511, + 598, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "25888", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/6a4c48b9d37231cc43ae5de841e474fdca7932541b531822e45b129ee75c4165.jpg", + "image_caption": [ + "Original Image" + ], + "image_footnote": [], + "bbox": [ + 189, + 106, + 326, + 212 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/f2428e51a506c7f911d3ce7b9de3d35d0def57fbde27ff002db54d175883c1c0.jpg", + "image_caption": [ + "WhatsApp Image" + ], + "image_footnote": [], + "bbox": [ + 344, + 106, + 483, + 212 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/ed03c9d0381af79a0ae566d39b2892547647ecc18df6544542be4368a61a571c.jpg", + "image_caption": [ + "Original Image" + ], + "image_footnote": [], + "bbox": [ + 514, + 106, + 651, + 212 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/2f76e865bb7d4cea0bbd784a23623a5d8da016d64e55e13e13a474c5e412bc3b.jpg", + "image_caption": [ + "WhatsApp Image" + ], + "image_footnote": [], + "bbox": [ + 671, + 106, + 807, + 212 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/31286531a74904127f15d365c440e679ab6fa25698f1adb4141d012c7fbdcb39.jpg", + "image_caption": [ + "Figure 1. Exemplified localization generated with DenseNet architecture on (a) original images and (b) their corresponding WhatsApp compressed images from CheXWhatsApp. Though the images look visually similar after compression, the predictions and localizations generated by GradCAM can significantly differ." + ], + "image_footnote": [], + "bbox": [ + 189, + 226, + 326, + 332 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/4bdae44af782ead033d8d6c812f01130ad33a9023ba1c37b696db5e542003ef8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 344, + 226, + 483, + 332 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/190c5814fc5ad204a338e9da72ec87aa144172349810ab5b10062ca979e3beff.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 514, + 226, + 651, + 332 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/1e4d8d63f6dc8e9843b01db89ca6fc316b43977b39977c94988b7e4ea31a1693.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 669, + 226, + 808, + 332 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/6a87775742df2a4cccc8403e0d9c183fd69445550a1be96b463f8bcbefdc4b04.jpg", + "image_caption": [ + "Original Image" + ], + "image_footnote": [], + "bbox": [ + 187, + 426, + 328, + 534 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/81e452e66ac3b15fb16ac310bc0943b99a39b7d2f1a9027e11f0729b7e3b77b3.jpg", + "image_caption": [ + "WhatsApp Image" + ], + "image_footnote": [], + "bbox": [ + 344, + 426, + 483, + 532 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/bebe7dfe4c9af867ab09137cb3e64c13d8e2467a7ffc458d0bab9475ba665dcd.jpg", + "image_caption": [ + "Original Image" + ], + "image_footnote": [], + "bbox": [ + 514, + 426, + 651, + 532 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/045042c4a8e93f7522b8fa60e373d4ac5039cb3047ed38db3ec019b8bbf44c60.jpg", + "image_caption": [ + "WhatsApp Image" + ], + "image_footnote": [], + "bbox": [ + 669, + 426, + 808, + 534 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/0768f44fb39ba0149b40ad45460b28aa0d934d3b28e00d4f0210d671eea8a3f9.jpg", + "image_caption": [ + "Figure 2. Exemplified localization generated with Faster-RCNN architecture on (a) original images and (b) their corresponding WhatsApp compressed images from CheXwhatsApp. The red boxes indicate the ground truth, and the blue boxes indicate the model's prediction. Though the images look visually similar after compression, the localizations generated by Faster-RCNN are either incorrectly placed (bottom-left), partially missing (top-right) or fully missing (top-left, bottom-right) in WhatsApp-compressed images when compared with the original images." + ], + "image_footnote": [], + "bbox": [ + 191, + 546, + 328, + 651 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/09ce89bf123617e0141fdabf8d65917035348d17938c50e1a7f124b2b6f30eff.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 346, + 546, + 483, + 651 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/e57213d1a506f8c111233c3119b831e6669c00871fb7d51fe9813c9ca4f3aeab.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 514, + 546, + 651, + 652 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/ada2940d9ad213f858403adad94696bacfe8abe7a22d049deb044d6b4a31c6b1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 669, + 546, + 805, + 651 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Challenges in diagnosing Xray images received through Smart phones", + "text_level": 1, + "bbox": [ + 89, + 744, + 482, + 780 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In deep learning based models for image classification, the prediction outcome is provided by the model and explanation of predictions is obtained through class-activation maps. In mHealth settings, as envisaged here, such systems face two challenges: prediction instability and out-of-lung saliency. Object detection models employ a combined approach to prediction and explanation by lo", + "bbox": [ + 89, + 794, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "calizing the findings using bounding boxes. However, the output of object detection models can also suffer from localization instability when images are obtained through WhatsApp. This section delves deeper into these three challenges and defines the metrics used to evaluate them.", + "bbox": [ + 511, + 747, + 906, + 821 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. The Prediction Instability Problem", + "text_level": 1, + "bbox": [ + 511, + 830, + 816, + 847 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "PIP of a predictive model is defined as the probability of disagreement between the predictions on a randomly perturbed instance and the true instance.", + "bbox": [ + 511, + 854, + 906, + 898 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "25889", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1.1. Metrics to evaluate PIP", + "text_level": 1, + "bbox": [ + 89, + 90, + 302, + 104 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Prediction Instability Score (PI Score): Consider a deep learning model $\\mathsf{M}$ that classifies a CXR image into one or multiple categories. Consider a dataset $D = \\{X_{i},\\mathsf{R}(X_{i})\\}_{i = 1}^{N}$ consisting of $N$ pairs of original and WhatsApp-compressed CXR images denoted by $X_{i}$ , and $\\mathsf{R}(X_i)$ respectively. The estimate of PI Score due to image perturbation by WhatsApp can now be defined as the fraction of pairs where the predictions differed expressed as percentage.", + "bbox": [ + 89, + 109, + 483, + 247 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nP I S c o r e (\\mathbb {M}; D) = \\frac {1}{N} \\left(\\sum_ {\\{X _ {i}, \\mathbb {R} (X _ {i}) \\} \\in D} P I \\left(X _ {i}, \\mathbb {R} \\left(X _ {i}\\right), \\mathbb {M}\\right)\\right) \\times 1 0 0\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 252, + 516, + 311 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "and $PI(X_i, \\mathbf{R}(X_i), \\mathbb{M}) = \\mathbf{I}(y(X_i; \\mathbb{M})) \\neq \\mathbf{I}(y(\\mathbf{R}(X_i); \\mathbb{M}))$", + "bbox": [ + 109, + 311, + 478, + 329 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Out-of-Lung Saliency", + "text_level": 1, + "bbox": [ + 89, + 335, + 294, + 351 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For a Deep Learning Model diagnosing lung abnormalities in CXR images, out-of-lung saliency (OLS) is defined as the correlation between attributed pixels outside the lung region and model prediction.", + "bbox": [ + 89, + 357, + 483, + 417 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.1. Metrics to evaluate OLS", + "text_level": 1, + "bbox": [ + 89, + 422, + 307, + 436 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Out of Lung Saliency Score (OLS Score): The OLS Score quantifies the impact of OLS on the diagnosis made by state-of-the-art models. It is derived using the Intersection Over Lung-Region (IOL) metric, which measures the extent of OLS on each CXR image. Let heatmap-region $(y(x; \\mathbb{M}))$ indicate the set of pixels identified by a saliency based feature attribution technique such as GradCAM[33] applied to model $\\mathbb{M}$ on CXR image $x$ , resulting in prediction $y$ . Let $lung(x)$ represents the pixels within image $x$ that belong to the lung region. IOL is defined as the ratio of the number of pixels where the heatmap region and the lung region overlap in a CXR image, divided by the total number of pixels in the heatmap region. Mathematically, for a given CXR image $x$ (represented by its pixels and intensity values), define", + "bbox": [ + 89, + 441, + 483, + 667 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nI O L (x; \\mathbb {M}) = \\frac {\\operatorname {P i x} (\\text {h e a t m a p - r e g i o n} (y (x ; \\mathbb {M})) \\cap l u n g (x))}{\\operatorname {P i x} (\\text {h e a t m a p - r e g i o n} (y (x ; \\mathbb {M})))} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 674, + 483, + 719 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\operatorname{P}\\mathrm{i}\\mathrm{x}(R(x))$ counts the number of Pixels in region $R(x)$ , a subset of the pixels in the original image $x$ . The IOL value ranges from 0 to 1, where 0 indicates that prediction relies entirely on pixels outside the lung region and 1 indicates otherwise. The OLS Score of a model $\\mathbb{M}$ on a dataset $D$ consisting of $N$ images is defined as the percentage of images in $D$ with an IOL value less than a threshold $\\eta$ .", + "bbox": [ + 89, + 719, + 483, + 840 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nO L S \\operatorname {S c o r e} _ {\\eta} (D; \\mathbb {M}) = \\frac {1}{N} \\left(\\sum_ {x \\in D} I O L (x; \\mathbb {M}) < \\eta\\right) \\times 1 0 0 \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 96, + 845, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The value of $\\eta$ in this paper is set to 0.4. For brevity, we will exclude $\\eta$ from the subscript. A high OLS Score suggests that a majority of the data is affected by OLS, while a low score implies lesser impact.", + "bbox": [ + 511, + 90, + 905, + 151 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. The Localization Instability Problem", + "text_level": 1, + "bbox": [ + 511, + 160, + 833, + 176 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "LIP of an object detection model is defined as the probability of disagreement between the localization of findings identified by the model that agree with the ground truth, on a randomly perturbed instance versus the true instance. Object detection models localize findings using bounding boxes and thus simultaneously provide both the prediction and explanation of a finding.", + "bbox": [ + 511, + 181, + 906, + 287 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3.1. Metrics to evaluate LIP", + "text_level": 1, + "bbox": [ + 511, + 295, + 723, + 309 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Localization Instability Score (LI Score): Consider a model M that utilizes bounding boxes to localize findings (instances of lung abnormalities) on both the original image $x$ and its perturbed version $\\mathsf{R}(x)$ . Let $\\mathsf{Gt}_x$ denote the set of ground truth localizations of $x$ . Let $y_{\\mathsf{f}}(x, \\mathsf{M})$ be the localization of a finding $\\mathsf{f}$ returned by M. $y_{\\mathsf{f}}(x, \\mathsf{M})$ agrees with ground truth of the finding $\\mathsf{g}\\mathsf{t}_{x,\\mathsf{f}} \\in \\mathsf{G}\\mathsf{t}_x$ if intersection over union (IOU) between localization returned by model $y_{\\mathsf{f}}(x, \\mathsf{M})$ and ground truth $\\mathsf{g}\\mathsf{t}_{x,\\mathsf{f}}$ is greater than a threshold. ie, IOU $(y_{\\mathsf{f}}(x, \\mathsf{M}), \\mathsf{g}\\mathsf{t}_{x,\\mathsf{f}}) > t$ , where $t$ is a threshold.", + "bbox": [ + 511, + 314, + 905, + 465 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Let $\\mathsf{F}_{x,t}$ be the set of findings localized by M on image $x$ w.r.t the ground-truth. $\\mathsf{F}_{x,t} = \\{\\mathsf{f}|\\mathsf{IOU}(y_{\\mathsf{f}}(x,\\mathsf{M}),\\mathsf{gt}_{x,\\mathsf{f}}) > t,\\forall \\mathsf{f}\\in \\mathsf{Gt}_x\\}$ Let $\\mathsf{F}_{\\mathbb{R}(x),t}$ be the set of findings localized by M on perturbed image $\\mathbb{R}(x)$ w.r.t the ground-truth. $\\mathsf{F}_{\\mathbb{R}(x),t} = \\{\\mathsf{f}|\\mathsf{IOU}(y_{\\mathsf{f}}(\\mathbb{R}(x),\\mathsf{M}),\\mathsf{gt}_{x,\\mathsf{f}}) > t,\\forall \\mathsf{f}\\in \\mathsf{Gt}_x\\}$ . LI Score due to image perturbation by WhatsApp is defined as the fraction of image pairs, where the localization of findings on original and WhatsApp-compressed images that agree with ground truth differ, expressed as percentage.", + "bbox": [ + 511, + 465, + 906, + 602 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL I S c o r e _ {t} (D; \\mathbb {M}) = \\frac {1}{N} \\left(\\sum_ {x \\in D} \\mathrm {F} _ {x, t} \\neq \\mathrm {F} _ {\\mathrm {R} (x), t}\\right) \\times 1 0 0 \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 625, + 906, + 666 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Currently, there does not exist a comprehensive dataset to evaluate the problems of PIP, OLS and LIP associated with CXR images transmitted through WhatsApp for mHealth applications.", + "bbox": [ + 511, + 675, + 905, + 737 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. CheXwhatsapp : A new dataset of WhatsApp compressed CXR images paired with original images", + "text_level": 1, + "bbox": [ + 511, + 750, + 906, + 801 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section we review open source CXR datasets and discuss the steps involved in creation of a new CXR dataset called CheXwhatsapp that consists of both original and their WhatsApp images", + "bbox": [ + 511, + 809, + 905, + 869 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Discussion of Existing CXR Datasets: We study the most extensively used datasets - NIH [38] and RSNA[34], in", + "bbox": [ + 511, + 869, + 905, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "25890", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/0781c06f13d45771f93d73b191a0628a1629a27e80a628ad7881764d366962f9.jpg", + "image_caption": [ + "(a) Original image" + ], + "image_footnote": [], + "bbox": [ + 189, + 88, + 390, + 239 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/adc47ef8c69c707af9bd1f8028ba28c17cee17d7e49dffc3850bb0a3377e8470.jpg", + "image_caption": [ + "(b) Compressed image", + "Figure 3. Illustration of a sample (a) original image and its corresponding (b) compressed image from CheXwhatsApp. The original image is of dimension $1024 \\times 1024$ and has image size of 668KB and the WhatsApp image is of dimension $1024 \\times 1024$ and has an image size of 134KB, indicating an $80\\%$ compression." + ], + "image_footnote": [], + "bbox": [ + 611, + 88, + 810, + 241 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "the literature. These are publicly available anonymized datasets for research purposes [6, 28, 36, 37]. It has been duly verified that the NIH and RSNA datasets permits modification and redistribution of data with proper citation. It is important to note that the collection of these datasets adhered to established protocols and guidelines, following the ethical regulations stipulated by the respective institutions. Informed consent was obtained from the participating subjects. Further information regarding these datasets can be found in [34, 38]. An overview of these datasets, are as follows.", + "bbox": [ + 88, + 335, + 480, + 500 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- NIH: NIH [38] dataset contains 86524 CXR images in train set and 25596 CXR images in test set from 30,805 unique patients. It is a multi-label dataset where patients can have normal CXR or be associated with one or more of the 14 different pathologies: Atelectasis, Cardiomegaly, Effusion, Infiltration, Mass, Nodule, Pneumonia, Pneumothorax, Consolidation, Edema, Emphysema, Fibrosis, Pleural-Thickening, and Hernia.", + "- RSNA: The RSNA[34] consists of CXR images categorized as either pneumonia or normal cases, with manually annotated bounding boxes for pneumonia by expert radiologists. The dataset consists of 29,684 CXR images in training and validation set, as provided by the original authors." + ], + "bbox": [ + 89, + 505, + 482, + 715 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Additional details regarding the dataset can be found in the Supplementary Material.", + "bbox": [ + 104, + 717, + 482, + 747 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. CheXwhatsapp: A new dataset of original and WhatsApp compressed images", + "text_level": 1, + "bbox": [ + 89, + 757, + 482, + 789 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To train models that are robust to image perturbations due to compression caused by low-bandwidth messaging applications like WhatsApp, it is crucial to have a dataset that includes compressed images generated by these messaging applications. However, currently, no open-source datasets are available with image pairs consisting of original CXR images and their corresponding compressed versions. To", + "bbox": [ + 89, + 795, + 482, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "address this gap, we created a new in-house dataset using WhatsApp and the previously discussed datasets, called CheXwhatsapp.", + "bbox": [ + 511, + 335, + 903, + 380 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "CheXwhatsapp consists of 141,804 pairs of original and WhatsApp-compressed CXR images. CheXwhatsapp is an amalgamation of the two datasets: NIH [38] and RSNA [34] along with their corresponding WhatsApp-compressed images. The original and corresponding WhatsApp-compressed images in NIH [38] are known as NIH-WA. Similarly, the original and corresponding WhatsApp-compressed images in RSNA [34] are known as RSNA-WA. CheXwhatsapp is the union of NIH-WA and RSNA-WA.", + "bbox": [ + 511, + 382, + 905, + 532 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To create the image pairs, all the images in NIH [38] and RSNA [34] were sequentially sent through WhatsApp to a fixed mobile number. The creation of CheXwhatisApp is performed in a fully automated manner. The primary challenge in automating this process is to synchronize sending images to WhatsApp and receiving images from WhatsApp. This is primarily due to two reasons - 1) WhatsApp sends images out-of-order in certain scenarios and 2) the metadata of images is lost after sending, and images are renamed to some random name by WhatsApp at the recipient's end. Since we do not have control over WhatsApp API, the image name is also sent along with the image. The steps involved in creating the CheXwhatisApp are given below:", + "bbox": [ + 511, + 534, + 905, + 744 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. Preprocess the image by adding the image name to the image itself. This is done by adding a small white border of 15pixel $\\times$ image-width at the bottom of the image and placing the image name in the white border. The preprocessed images are uploaded to an Amazon S3 bucket to be sent to the WhatsApp Business API.", + "2. The pre-processed images are send by the WhatsApp business API one by one to a different WhatsApp account. This is established by using platforms for WhatsApp bulk-messaging. Such platforms are con" + ], + "bbox": [ + 511, + 750, + 903, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "25891", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "figured using a phone number enabled with WhatsApp and allow bulk-messaging of images and text.", + "bbox": [ + 109, + 90, + 480, + 119 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "3. The images received via WhatsApp are downloaded to an edge device, from where they are uploaded to a cloud service provider.", + "4. The white border at the bottom of the images is cropped to read the image name via an OCR technique called Tesseract. The image name read via OCR is used to rename the image." + ], + "bbox": [ + 89, + 121, + 482, + 227 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "At the conclusion of the WhatsApp-compression experimentation, it was observed that the 141,804 WhatsApp compressed CXR images in CheXwhatsapp occupied only 18.43 GB of memory, whereas the original 141,804 CXR images occupied 54.80 GB. This indicates a compression factor of $66.36\\%$ achieved by WhatsApp. However, this compression can alter the intricate intensity values of the images and potentially lead to inaccurate diagnoses if the models are trained solely on high-resolution images. Figure 3 presents an original and corresponding compressed images from NIH-WA in CheXwhWhatsApp. The existing state-of-the-art (SOTA) models predicted different diagnoses for the high-resolution (original) image compared to the low-resolution (compressed) image. Additionally, the activation maps used for localization showed significant differences, often highlighting non-lung regions. Clearly, this calls for more research in this area.", + "bbox": [ + 89, + 231, + 483, + 487 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Benchmarking using CheXwhat-sApp dataset", + "text_level": 1, + "bbox": [ + 89, + 502, + 483, + 537 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The study shows that introducing the proposed WhatsApp dataset mitigates the effect of PIP, OLS and LIP of SOTA models. The NIH and NIH-WA datasets are evaluated using classification models. The RSNA and RSNA-WA datasets are evaluated using object detection models. This section begins with a discussion of the benchmarking setup for classification and object detection models. As stated previously, the PI Score and OLS Score are employed to analyze the effect of PIP and OLS on the prediction outcome of classification models. LI Score is used to analyze LIP in object detection models.", + "bbox": [ + 89, + 546, + 482, + 710 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Benchmark setup. State-of-the-art models considered in this study are DenseNet [17], ResNet [16], ResNeXt [40] and Inception [35]. These models are implemented using Pytorch. The models are trained using the Adam optimizer with an initial learning rate of 1e-4. The input image dimension are $512 \\times 512$ . To fine-tune the learning process, ReduceLROOnPlateau scheduler is employed with patience 4 and a factor 0.1. The batch size for training data is 16 and the models are trained for 25 epochs.", + "bbox": [ + 89, + 734, + 482, + 869 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For object detection tasks, the Faster-RCNN [30] is trained using SGD optimizer with learning-rate 0.0055, mo", + "bbox": [ + 89, + 869, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "mentum 0.9 and weight-decay 0.0005. The training batch-size is 10. The learning-rate schedular is StepLR (step-size=3, gamma = 0.1) and the model is trained for 15 epochs. YOLOV5s [19] and YOLOV8s [20] models are trained with Auto optimizer with learning-rate $\\mathrm{lr}0 = 0.001$ , $\\mathrm{lrf} = 0.001$ , final-lr $= lr0 \\times lr f$ and momentum $= 0.937$ . The training batch-size is 16 and the models are trained for 50 epochs. The training is performed using 8 NVIDIA Tesla V100-SXM2-32GB GPUs. The training code is available2.", + "bbox": [ + 511, + 90, + 903, + 227 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Benchmarking results. In the experiments, SOTA models DenseNet [17], ResNet [16], ResNeXt [40] and Inception [35] are trained using the existing dataset- NIH and the proposed paired dataset - NIH-WA. The models trained using training images from existing and proposed datasets are evaluated for PIP and OLS using the original and WhatsApp-compressed images in the test dataset. The results are presented in Table 1. The results clearly indicate that models trained using NIH-WA reduces the effect of PIP, when compared with those trained using NIH. The PI Score of networks trained using NIH-WA is upto $80\\%$ lower than those trained using NIH.", + "bbox": [ + 511, + 239, + 903, + 421 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To evaluate OLS Score of SOTA models trained using the existing dataset - NIH and the proposed paired dataset - NIH-WA, heatmaps of CXR images were generated using GradCAM [33], GradCAM++ [5] and EigenCAM [24]. For each CXR image, the lung region was identified using a U-Net [31] to evaluate whether the model's activation map falls within the lung regions or extends into non-lung regions. We use the highest activation region in the heatmap to determine out-of-lung saliency, considering activation values greater than 0.5 (on a scale of 0 to 1). The lung region extracted by U-Net from the CXR image, along with the highest activation region in the heatmap is utilized for the calculation of the IOL Score, which is then used to evaluate OLS Score. The results (Table 2) clearly show that models trained using NIH-WA reduces the effect of OLS, when compared to those trained using NIH. Moreover, the problem of OLS is further exacerbated in CXR images compressed by WhatsApp.", + "bbox": [ + 511, + 422, + 905, + 694 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The SOTA object detection networks, Faster-RCNN [30],YOLOv5s [19] and YOLOv8s [20] are trained using the existing bounding box CXR dataset - RSNA and the proposed paired dataset - RSNA-WA. These models are evaluated for LIP using the original and WhatsApp-compressed images in the test dataset. From the LI Score results evaluated with threshold $t = 0.7$ (Table 3), it is evident that the networks trained using RSNA-WA reduces the problem of LIP by upto $81\\%$ when compared with those trained using RSNA. LI Score of networks trained using RSNA and RSNA-WA for different values of threshold $t$ is given in", + "bbox": [ + 511, + 696, + 905, + 861 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "https://github.com/cheXwhatsApp/cheXwhatsApp_benchmark", + "bbox": [ + 511, + 875, + 903, + 898 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "25892", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/ab33f10f8fb7db5eeae65b8d8a9f4d7d583a6522f814dd76ea9ac601cfde54d0.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelsPI Score
NIHNIH-WA
DenseNet15.743.31 (79%)
InceptionNet15.083.14 (79%)
ResNet-5014.603.30 (77%)
ResNeXt-5016.863.90 (77%)
", + "bbox": [ + 372, + 88, + 624, + 202 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/e2ec4968aefb990666e0afc55d26c865d3bb94d6addd9f8cc9a644812fa14319.jpg", + "table_caption": [ + "Table 1. PI Score for different networks trained on NIH and NIH-WA and evaluated on original and WhatsApp test images. From the table, it is evident that networks trained on NIH-WA has substantially lower prediction instability problem (PIP) compared to those trained on NIH." + ], + "table_footnote": [], + "table_body": "
ModelsDatsetOLS Score(GradCAM)OLS Score(GradCAM++)OLS Score(EigenCAM)
OriginalWhatsAppOriginalWhatsAppOriginalWhatsApp
DenseNetNIH18.7926.6930.7941.2546.0660.66
NIH-WA12.53 (33%)21.05 (21%)13.12 (57%)29.39 (29%)13.16 (71%)20.57 (66%)
InceptionNIH37.7547.4552.3362.2195.5896.70
NIH-WA22.10 (41%)32.64 (31%)18.04 (66%)34.26 (45%)94.55 (01%)95.31 (01%)
ResNetNIH24.6533.6542.1251.4097.5198.89
NIH-WA14.94 (39%)15.34 (54%)21.89 (48%)39.58 (23%)95.01 (02%)95.16 (03%)
ResNeXtNIH33.1242.5661.2469.3999.6799.79
NIH-WA11.81 (64%)26.26 (38%)10.51 (82%)26.58 (62%)96.28 (03%)96.48 (03%)
", + "bbox": [ + 133, + 267, + 864, + 464 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/3c61fa4a3c445e8869111474effc31b3ed1c5f2e25a9b64591126e729a218545.jpg", + "table_caption": [ + "Table 2. OLS Score score of attribution maps generated using different networks trained on NIH and NIH-WA training dataset and evaluated on original and WhatsApp images in test dataset. The results show that attribution maps generated using models trained on NIH-WA has lower out-of-lung saliency problem (OLS) compared to those trained on NIH." + ], + "table_footnote": [], + "table_body": "
ModelsLI Score
RSNARSNA-WA
YOLOV5s3.010.85 (72%)
YOLOV8s4.020.78 (81%)
Faster-RCNN12.4310.86 (13%)
", + "bbox": [ + 364, + 526, + 632, + 626 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. LI Score for different networks trained on the bounding-box dataset - RSNA and RSNA-WA and evaluated using original and WhatsApp test dataset. From the table, it is evident that the networks trained on RSNA-WA reduces loalization instability problem (LIP) by upto $81\\%$ compared to those trained on RSNA.", + "bbox": [ + 89, + 636, + 906, + 679 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/68b12e8b9cd4c7686d8157cd51eaac14baa995e57d0cb5d0cacfd6db42726603.jpg", + "image_caption": [ + "(a) LI Score of YOLOv5s [19]" + ], + "image_footnote": [], + "bbox": [ + 99, + 693, + 334, + 813 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/73ebc14075285b97c8af41f1d9f17234dd16d9ba1ff9c249e0e1991463cad3da.jpg", + "image_caption": [ + "(b) LI Score of YOLOv8s [20]", + "Figure 4. LI Score of Faster-RCNN [30], YOLOv5s [19] and YOLOv8s [20] on RSNA and RSNA-WA. From the results, it can be seen that models trained on RSNA-WA mitigate the effect of loalization instability problem (LIP) for different values of threshold $t$ ." + ], + "image_footnote": [], + "bbox": [ + 348, + 693, + 578, + 813 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/738d3d69929d5fb9ce25a82604cb6bceb501d751ef7934b527cbc59e70dc9b90.jpg", + "image_caption": [ + "(c) LI Score of Faster-RCNN [30]" + ], + "image_footnote": [], + "bbox": [ + 594, + 693, + 823, + 813 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "25893", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Figure 4. The figure also shows that networks trained using RSNA-WA have substantially lower LIP for all values of $t$ . In particular, YOLO trained using RSNA-WA reduces LI score by more than $70\\%$ .", + "bbox": [ + 89, + 90, + 480, + 150 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "For LI score, the choice of parameter $t$ represents the threshold of overlap between predicted and true boundingboxes. If the parameter $t$ is too small, then most predicted bounding boxes will overlap with the true bounding boxes in both the original and WhatsApp images and LIP will be low. However, if $t$ is too large, then only few predicted bounding boxes will overlap with true bounding boxes in both original and WhatsApp images and again LIP will be low. Therefore, we must select a value for $t$ that balances these two scenarios. This optimal value may vary depending on the dataset. In this paper, we selected $t$ that corresponds to the highest LI Score to study the maximum localization instability.", + "bbox": [ + 88, + 152, + 482, + 347 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To understand the generalizability of model trained on CheXwhatsapp to images compressed by other compression algorithms such as ISO JPEG compression, we created test datasets by compressing the original test images using various JPEG quality factors - 20, 40, 60, 80, and 90 with OpenCV library. Our observations indicated that instability increases as the quality factor decreases. Furthermore, models trained with CheXwhatsapp demonstrated an average of $26\\%$ lower prediction instability and $10\\%$ lower localization instability compared to models trained on original images alone. Additional results are available in supplementary material.", + "bbox": [ + 89, + 347, + 482, + 527 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Discussion and Conclusion", + "text_level": 1, + "bbox": [ + 89, + 542, + 338, + 558 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we present a large-scale CheXwhatsApp dataset to aid the development of mHealth applications for diagnosing CXR images. The existing datasets consist of high resolution images and the models trained on these datasets are found to face high PIP, OLS and LIP. To study this, we proposed quantitative metrics called PI, OLS, and LI scores and obtained these metrics using the WhatsApp compressed CXR images. The high scores indicate evidence of the problem, which is due to the compression techniques used by mHealth applications.", + "bbox": [ + 89, + 568, + 482, + 718 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The growing body of research on using smart-phones and WhatsApp [3, 8, 22, 23, 32, 41] in the medical domain calls for large datasets of WhatsApp-compressed images. Our proposed dataset comprise 141,804 pairs of high-resolution and their WhatsApp compressed CXR images. As there are no available WhatsApp CXR datasets, this proposed dataset is a valuable resource for the research community to address the challenges posed by image compression and transmission over low-bandwidth networks. Further, this proposed dataset was not easy to create. To create this dataset, we used a WhatsApp business api which incurred charges per each compression and we needed to", + "bbox": [ + 89, + 719, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "sequentially send the images from one phone and receive them on another phone, followed by proper mapping of the high resolution and their corresponding WhatsApp compressed images. The conducted experiments showed that the models trained with this dataset demonstrate much better visual and numerical results when tested on unseen images, thus highlighting its high robustness. As observed from Tables 1, 2, 3, training with proposed dataset helped reduce PI, OLS and LI scores significantly, demonstrating the importance of our proposed dataset.", + "bbox": [ + 511, + 90, + 903, + 241 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Furthermore, the OLS scores on the high resolution dataset improved by $70\\%$ in some cases when trained with proposed dataset. This shows that the proposed dataset can aid in better explainability of results with fewer predictions outside the lung region. In healthcare, explainability plays a vital role in assisting physicians in making informed decisions. While this paper is the first to offer a large paired dataset of original and WhatsApp-compressed CXR images, there are limitations. One limitation is that we did not collect a new dataset; instead it is derived from existing datasets. Another limitation is that we did not use the new dataset to create or improve existing mHealth applications.", + "bbox": [ + 511, + 242, + 903, + 422 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Additionally, we also experimented with available JPEG compression algorithms and used them to train the models. It was interesting to note that when trained with the JPEG compressed CXR images, the PIP, OLS and LIP further increased. More details of this experimentation are provided in the supplementary material. This highlights the importance of the proposed dataset in revolutionizing healthcare delivery with mHealth applications, especially in resource-constrained regions, by facilitating rapid patient triaging and assisting inexperienced rural doctors in interpretation. We encourage researchers to use this dataset to advance research in CXR diagnosis for mHealth applications.", + "bbox": [ + 511, + 424, + 903, + 604 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 514, + 619, + 609, + 635 + ], + "page_idx": 7 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Mariamma Antony, Siva Teja Kakileti, Rachit Shah, Sabyasachi Sahoo, Chiranjib Bhattacharyya, and Geetha Manjunath. Challenges of ai driven diagnosis of chest x-rays transmitted through smart phones: a case study in Covid-19. Scientific Reports, 13(1):18102, 2023. 1", + "[2] Richa Arora. The training and practice of radiology in india: current trends. Quantitative imaging in medicine and surgery, 4(6):449-450, 2014. 25525575[pmid]. 1", + "[3] Vera Helen Buss, Stuart Leesong, Margo Barr, Marlien Varnfield, and Mark Harris. Primary prevention of cardiovascular disease and type 2 diabetes mellitus using mobile health technology: systematic review of the literature. Journal of medical Internet research, 22(10):e21159, 2020. 1, 8", + "[4] Aurelia Bustos, Antonio Pertusa, Jose-Maria Salinas, and Maria De La Iglesia-Vaya. Padchest: A large chest x-ray image dataset with multi-label annotated reports. Medical image analysis, 66:101797, 2020. 2", + "[5] Aditya Chattopadhy, Anirban Sarkar, Prantik Howlader," + ], + "bbox": [ + 521, + 643, + 906, + 901 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "25894", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "and Vineeth N Balasubramanian. Grad-cam++: Generalized gradient-based visual explanations for deep convolutional networks. In 2018 IEEE Winter Conference on Applications of Computer Vision (WACV), pages 839-847, 2018. 6", + "[6] Muhammad EH Chowdhury, Tawsifur Rahman, Amith Khandakar, Rashid Mazhar, Muhammad Abdul Kadir, Zaid Bin Mahbub, Khandakar Reajul Islam, Muhammad Salman Khan, Atif Iqbal, Nasser Al Emadi, et al. Can ai help in screening viral and covid-19 pneumonia? IEEE Access, 8:132665-132676, 2020. 5", + "[7] Alex J. DeGrave, Joseph D. Janizek, and Su-In Lee. Ai for radiographic Covid-19 detection selects shortcuts over signal. Nature Machine Intelligence, 3(7):610–619, 2021. 1", + "[8] Olga Di Fede, Gaetano La Mantia, Mario GCA Cimino, and Giuseppina Campisi. Protection of patient data in digital oral and general health care: A scoping review with respect to the current regulations. Oral, 3(2):155-165, 2023. 1, 8", + "[9] Robert Geirhos, Jörn-Henrik Jacobsen, Claudio Michaelis, Richard Zemel, Wieland Brendel, Matthias Bethge, and Felix A. Wichmann. Shortcut learning in deep neural networks. Nature Machine Intelligence, 2(11):665-673, 2020. 1", + "[10] Daniele Giansanti. Teleradiology today: the quality concept and the Italian point of view. Telemedicine and e-Health, 23 (5):453–455, 2017. 1", + "[11] Daniele Giansanti, Livia Castrichella, and Maria Rosaria Giovagnoli. The design of a health technology assessment system in telepathology. Telemedicine and e-Health, 14(6): 570-575, 2008. 1", + "[12] Daniele Giansanti, Marco Pochini, and Maria Rosaria Giovagnoli. How tablet technology is going to change cooperative diagnosis in the cytology e-laboratory. Telemedicine and e-Health, 19(12):991-993, 2013.", + "[13] Daniele Giansanti, Marco Pochini, and Maria Rosaria Giovagnoli. Integration of tablet technologies in the e-laboratory of cytology: a health technology assessment. Telemedicine and e-Health, 20(10):909–915, 2014. 1", + "[14] Daniele and Giansanti. Whatsapp in mhealth: an overview on the potentialities and the opportunities in medical imaging. Mhealth, 6, 2020. 1", + "[15] GSMA. The state of mobile internet connectivity 2022, 2022. 1", + "[16] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 770-778, 2016. 6", + "[17] Gao Huang, Zhuang Liu, Laurens Van Der Maaten, and Kilian Q. Weinberger. Densely connected convolutional networks. In 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2261-2269, 2017. 6", + "[18] Jeremy Irvin, Pranav Rajpurkar, Michael Ko, Yifan Yu, Silviana Ciurea-Ilicus, Chris Chute, Henrik Marklund, Behzad Haghgoo, Robyn Ball, Katie Shpanskaya, et al. Chexpert: A large chest radiograph dataset with uncertainty labels and expert comparison. In Proceedings of the AAAI conference on artificial intelligence, pages 590-597, 2019. 2", + "[19] Glenn Jocher. Ultralytics yolov5, 2020. 6, 7" + ], + "bbox": [ + 93, + 90, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[20] Glenn Jocher, Ayush Chaurasia, and Jing Qiu. Ultralytics yolov8, 2023. 6, 7", + "[21] Alistair EW Johnson, Tom J Pollard, Seth J Berkowitz, Nathaniel R Greenbaum, Matthew P Lungren, Chih-ying Deng, Roger G Mark, and Steven Horng. Mimic-cxr, a de-identified publicly available database of chest radiographs with free-text reports. Scientific data, 6(1):317, 2019. 2", + "[22] Arjun Kalyanpur. WhatsApp and its role in teleradiology. Indian Journal of Radiology and Imaging, 2024. 1, 8", + "[23] J Austin Lee, Benjamin W Wachira, John Kennedy, Nicholas Asselin, and Nee-Kofi Mould-Millman. Utilisation of what-sapp for emergency medical services in garissa, kenya. African Journal of Emergency Medicine, 14(1):38-44, 2024. 1, 8", + "[24] Mohammed Bany Muhammad and Mohammed Yeasin. Eigen-cam: Class activation map using principal components. In 2020 international joint conference on neural networks (IJCNN), pages 1-7. IEEE, 2020. 6", + "[25] Ha Q Nguyen, Khanh Lam, Linh T Le, Hieu H Pham, Dat Q Tran, Dung B Nguyen, Dung D Le, Chi M Pham, Hang TT Tong, Diep H Dinh, et al. Vindr-cxr: An open dataset of chest x-rays with radiologist's annotations. Scientific Data, 9(1):429, 2022. 2", + "[26] Unathi Ntja, Jacques Janse van Rensburg, and Gina Joubert. Diagnostic accuracy and reliability of smartphone captured radiologic images communicated via Whatsapp®. African Journal of Emergency Medicine, 12(1):67-70, 2022. 1", + "[27] Nick A Phillips, Pranav Rajpurkar, Mark Sabini, Rayan Krishnan, Sharon Zhou, Anuj Pareek, Nguyen Minh Phu, Chris Wang, Mudit Jain, Nguyen Duong Du, et al. Chexphoto: $10,000+$ photos and transformations of chest x-rays for benchmarking deep learning robustness. In Machine Learning for Health, pages 318-327. PMLR, 2020. 2", + "[28] Sivaramakrishnan Rajaraman, Jenifer Siegelman, Philip O. Alderson, Lucas S. Folio, Les R. Folio, and Sameer K. Antani. Iteratively pruned deep learning ensembles for Covid-19 detection in chest x-rays. IEEE Access, 8:115041-115050, 2020. 5", + "[29] Eduardo P Reis, Joselisa PQ De Paiva, Maria CB Da Silva, Guilherme AS Ribeiro, Victor F Paiva, Lucas Bulgarelli, Henrique MH Lee, Paulo V Santos, Vanessa M Brito, Lucas TW Amaral, et al. Brax, brazilian labeled chest x-ray dataset. Scientific Data, 9(1):487, 2022. 2", + "[30] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. IEEE Transactions on Pattern Analysis and Machine Intelligence, 39(6):1137-1149, 2017. 6, 7", + "[31] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention - MICCAI 2015, pages 234-241, Cham, 2015. Springer International Publishing. 6", + "[32] Ravi Saini, Madhan Jeyaraman, Naveen Jeyaraman, Vijay Kumar Jain, Swaminathan Ramasubramanian, and Karthikeyan P Iyengar. Advancing orthopaedic trauma care through Whatsapp: An analysis of clinical and non-clinical applications, challenges, and future directions. World Journal of Orthopedics, 15(6):529, 2024. 1, 8" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "25895", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[33] Ramprasaath R. Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In 2017 IEEE International Conference on Computer Vision (ICCV), pages 618-626, 2017. 4, 6", + "[34] George Shih, Carol C Wu, Safwan S Halabi, Marc D Kohli, Luciano M Prevedello, Tessa S Cook, Arjun Sharma, Judith K Amorosa, Veronica Arteaga, Maya Galperin-Aizenberg, et al. Augmenting the national institutes of health chest radiograph dataset with expert annotations of possible pneumonia. Radiology. Artificial intelligence, 1(1), 2019. 2, 4, 5", + "[35] Christian Szegedy, Sergey Ioffe, Vincent Vanhoucke, and Alexander A Alemi. Inception-v4, inception-resnet and the impact of residual connections on learning. In Thirty-first AAAI conference on artificial intelligence, 2017. 6", + "[36] Yu-Xing Tang, You-Bao Tang, Yifan Peng, Ke Yan, Mohammadhadi Bagheri, Bernadette A Redd, Catherine J Brandon, Zhiyong Lu, Mei Han, Jing Xiao, et al. Automated abnormality classification of chest radiographs using deep convolutional neural networks. NPJ digital medicine, 3(1):1-8, 2020. 5", + "[37] Linda Wang, Zhong Qiu Lin, and Alexander Wong. Covidnet: a tailored deep convolutional neural network design for detection of Covid-19 cases from chest x-ray images. Scientific Reports, 10(1):19549, 2020. 2, 5", + "[38] Xiaosong Wang, Yifan Peng, Le Lu, Zhiyong Lu, Mohammadhadi Bagheri, and Ronald M. Summers. ChestXray8: Hospital-scale chest x-ray database and benchmarks on weakly-supervised classification and localization of common thorax diseases. In 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2017. 2, 4, 5", + "[39] Xiaosong Wang, Yifan Peng, Le Lu, Zhiyong Lu, Mohammadhadi Bagheri, and Ronald M Summers. Chestxray8: Hospital-scale chest x-ray database and benchmarks on weakly-supervised classification and localization of common thorax diseases. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2097-2106, 2017. 2", + "[40] Saining Xie, Ross Girshick, Piotr Dólár, Zhuowen Tu, and Kaiming He. Aggregated residual transformations for deep neural networks. In 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 5987-5995, 2017. 6", + "[41] Esmaeel Yaagoob, Regina Lee, Michelle Stubbs, Fatimah Shuaib, Raja Johar, and Sally Chan. WhatsApp-based intervention for people with type 2 diabetes: A randomized controlled trial. Nursing & Health Sciences, 26(2):e13117, 2024. 1, 8" + ], + "bbox": [ + 91, + 92, + 480, + 792 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "25896", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/CheXwhatsApp_ A Dataset for Exploring Challenges in the Diagnosis of Chest X-rays through Mobile Devices/8cf11147-8ad4-4f89-a3ee-6b5492c8beb0_model.json b/2025/CheXwhatsApp_ A Dataset for Exploring Challenges in the Diagnosis of Chest X-rays through Mobile Devices/8cf11147-8ad4-4f89-a3ee-6b5492c8beb0_model.json new file mode 100644 index 0000000000000000000000000000000000000000..3b5399cbcf3518fcb2d0bbe3403d8b3d95d33cbf --- /dev/null +++ b/2025/CheXwhatsApp_ A Dataset for Exploring Challenges in the Diagnosis of Chest X-rays through Mobile Devices/8cf11147-8ad4-4f89-a3ee-6b5492c8beb0_model.json @@ -0,0 +1,2189 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.107, + 0.13, + 0.895, + 0.177 + ], + "angle": 0, + "content": "CheXwhatsapp: A Dataset for Exploring Challenges in the Diagnosis of Chest X-rays through Mobile Devices" + }, + { + "type": "text", + "bbox": [ + 0.15, + 0.203, + 0.314, + 0.222 + ], + "angle": 0, + "content": "Mariamma Antony1" + }, + { + "type": "text", + "bbox": [ + 0.356, + 0.203, + 0.468, + 0.222 + ], + "angle": 0, + "content": "Rajiv Porana" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.204, + 0.65, + 0.222 + ], + "angle": 0, + "content": "Sahil M Lathiya1" + }, + { + "type": "text", + "bbox": [ + 0.695, + 0.204, + 0.847, + 0.222 + ], + "angle": 0, + "content": "Siva Teja Kakileti2" + }, + { + "type": "text", + "bbox": [ + 0.398, + 0.222, + 0.6, + 0.24 + ], + "angle": 0, + "content": "Chiranjib Bhattacharyya" + }, + { + "type": "text", + "bbox": [ + 0.144, + 0.239, + 0.85, + 0.258 + ], + "angle": 0, + "content": "\\(^{1}\\)Indian Institute of Science, Bangalore, India \\(^{2}\\)Niramai Health Analytix, Bangalore, India" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.292, + 0.327, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.323, + 0.486, + 0.643 + ], + "angle": 0, + "content": "Mobile health (mHealth) has emerged as a transformative solution to enhance healthcare accessibility and affordability, particularly in resource-constrained regions and low-to-middle-income countries. mHealth leverages mobile platforms to improve healthcare accessibility, addressing radiologist shortages in low-resource settings by enabling remote diagnosis and consultation through mobile devices. Mobile phones allow healthcare workers to transmit radiographic images, such as chest X-rays (CXR), to specialists or AI-driven models for interpretation. However, AI-based diagnosis using CXR images shared via apps like WhatsApp suffers from reduced predictability and explainability due to compression artifacts, and there is a lack of datasets to systematically study these challenges. To address this, we introduce CheXwhatsapp, a dataset of 141,804 paired original and WhatsApp-compressed CXR images. We present a benchmarking study which shows the dataset improves prediction stability and explainability of state-of-the-art models by up to \\(80\\%\\), while also enhancing localization performance. CheXwhatsapp is open-sourced to support advancements in mHealth applications for CXR analysis1." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.67, + 0.222, + 0.687 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.696, + 0.485, + 0.879 + ], + "angle": 0, + "content": "Recent studies highlight the wide opportunities of messaging apps such as WhatsApp as an efficient tool for remote image sharing in mHealth applications across digital healthcare, including digital radiology [1, 10], digital pathology [11-13], cardiac imaging [14], orthopedics [32], diabetes management [41], and oral health [8]. mHealth provides a platform for health practitioners to enable disease management and speed-up diagnosis by facilitating the transfer of medical images, providing advice, facilitating drug adherence, and monitoring outbreaks. The outbreak of COVID-19 pandemic has accelerated the adoption of mHealth for digital radiology, such as for analyzing Chest X-ray (CXR)" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.294, + 0.578, + 0.308 + ], + "angle": 0, + "content": "imaging." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.309, + 0.907, + 0.476 + ], + "angle": 0, + "content": "mHealth is especially beneficial in resource constrained areas such as Low-to-Middle-Income countries (LMICs) where the ratio of radiologists to radiographers is significantly imbalanced. For example, in India, there is only one radiologist for every 100,000 people [2]. Compounding the issue is the fact that most radiologists are located in urban areas, often far from the majority of the X-ray imaging facilities located in rural areas. Due to the high demand for chest radiography and the global shortage of radiologists, numerous applications for AI-based diagnosis of CXR images have been developed." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.477, + 0.909, + 0.657 + ], + "angle": 0, + "content": "In recent years, smartphone penetration has significantly increased, and instant messaging applications such as WhatsApp have a large user-base with 535 million users in India, 118.5 million users in Brazil, and a penetration rate exceeding \\(97\\%\\) in African nations like Kenya, South Africa, and Nigeria [15]. This presents new opportunities for the quick and easy sharing of radiographic images to aid clinical decision-making [3, 22, 23, 26, 32]. In rural hospitals and diagnostic centers without on-site radiologists, the ability to transmit anonymized images instantly to a mHealth app for interpretation can be extremely valuable to facilitate rapid triaging of infected individuals [22]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.659, + 0.909, + 0.84 + ], + "angle": 0, + "content": "Leveraging common messaging platforms like WhatsApp eliminates the need for additional training since it is a widely used communication medium. However, it is important to note that these messaging platforms often compress images by upto \\(90\\%\\) to overcome bandwidth limitations. In medial applications, it is crucial to obtain explanation of predictions or localizations of findings in CXR images. Existing literature [1, 7, 9] identifies challenges related to explainability, such as predictions driven by features present outside the lung region in AI-based diagnostic systems for CXR images. Specifically, these systems suffer from the following challenges." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Prediction Instability Problem(PIP): PIP refers to a situation where model's prediction differs between the original image and the image transmitted through WhatsApp. This undesirable effect can have critical implications in mHealth" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.887, + 0.465, + 0.901 + ], + "angle": 0, + "content": "1https://github.com/cheXwhatsApp/cheXwhatsApp" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "25887" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.179, + 0.106 + ], + "angle": 0, + "content": "applications." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.108, + 0.482, + 0.319 + ], + "angle": 0, + "content": "Out of Lung Saliency(OLS): In medical diagnosis, it is crucial to provide explainable predictions. Saliency maps, which identify localized regions in an image that significantly contribute to the prediction, are commonly used to provide explanations. Out of Lung Saliency refers to the scenario where the model uses regions outside the lungs as major contributors to its prediction. Figure 1 displays instances of saliency maps generated by state-of-the-art models on original and corresponding WhatsApp-compressed CXR image pairs. In each of these images, the saliency maps highlight regions outside the lungs as significant contributors to the prediction. This effect is even more pronounced in WhatsApp-compressed images as illustrated in Figure 1." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.322, + 0.482, + 0.457 + ], + "angle": 0, + "content": "Localization Instability Problem(LIP): LIP refers to a situation where an object detection model exhibits localizations which are different on the transmitted image when compared with the original image. Figure 2 displays several instances of localization instability between the original image and the corresponding WhatsApp compressed image. In each of these images, the localizations predicted by the model for the original image is either absent or incorrectly placed in the WhatsApp-compressed image." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.459, + 0.482, + 0.595 + ], + "angle": 0, + "content": "Current methods for understanding the variations due to image perturbations are primarily visual. In this paper, we discuss and propose PIP, OLS, and LIP to quantitatively study the changes in predictions, saliency, and localizations. While several databases of CXR images have been released, most include only disease diagnosis labels or bounding boxes indicating regions of interest. However, a dataset that illustrates the aforementioned challenges is not yet publicly available." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.598, + 0.483, + 0.627 + ], + "angle": 0, + "content": "Contributions: Our contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.629, + 0.483, + 0.703 + ], + "angle": 0, + "content": "- We introduce CheXwhatsApp, the first paired dataset of 141,804 pairs of original and WhatsApp-compressed CXR images, which can be used for understanding the effect of WhatsApp on localization of various thoracic diseases." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.705, + 0.483, + 0.901 + ], + "angle": 0, + "content": "- We identify that prediction of state-of-the-art (SOTA) AI models on CXR images transmitted via low-bandwidth messaging applications such as WhatsApp are affected by prediction and explainability challenges. We find that several classification models exhibit the problem of PIP, where predictions differ between the original and transmitted images. While this is a concern, but even more worrisome is many of them also exhibit OLS where predictions are driven by regions outside the lung area, raising important questions on the explainability of the models. Similarly, models that localize lung abnormalities using bounding boxes exhibit the problem of LIP. We contribute several metrics to evaluate PIP, OLS and LIP." + }, + { + "type": "list", + "bbox": [ + 0.092, + 0.629, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.092, + 0.905, + 0.198 + ], + "angle": 0, + "content": "- Benchmark results indicate that models trained on open-source CXR images can suffer from high PIP, OLS and LIP. However, models trained on CheXwhat-sApp, which includes both original and WhatsApp-compressed CXR images, reduces the effect of PIP, OLS and LIP by upto \\(80\\%\\), compared to models trained on open-source CXR images." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.198, + 0.905, + 0.258 + ], + "angle": 0, + "content": "- We open-source the dataset and evaluation pipeline for evaluation of PIP, OLS and LIP in models on CheXwhat-sApp, to streamline the development and analysis of AI-based diagnostic systems for mHealth." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.092, + 0.905, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.272, + 0.659, + 0.287 + ], + "angle": 0, + "content": "2. Related works" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.297, + 0.905, + 0.597 + ], + "angle": 0, + "content": "There are several publicly available large-scale CXR datasets. The NIH dataset [38] is one of the first, containing 112,120 CXR images with labels from 30,805 unique patients. It is a multi-label dataset with 14 labels. Similar to NIH, the Chexpert dataset [18] has 224316 train and 234 validation images, both labeled with 14 categories. The Chexpert dataset was obtained from 65,240 unique patients at Stanford Hospital. The CovidNet [37] dataset contains 30,130 train and 400 testing images categorized as COVID-19, pneumonia, or normal. The MIMIC-CXR dataset [21] includes 227,835 CXR imaging studies for 65,379 patients, along with radiology reports from Beth Israel. ChestX-ray8 [39] is a CXR dataset that consists of 108,948 frontalview X-ray images of 32,717 unique patients with eight disease labels extracted through text mining. PadChest [4] contains 160,000 CXR images obtained from 67,000 patients in San Juan Hospital (Spain). BRAX [29] is a Brazilian labeled CXR dataset consisting of 24,959 CXR images annotated with fourteen labels derived from radiology reports using NLP." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.599, + 0.905, + 0.901 + ], + "angle": 0, + "content": "RSNA [34] and VinDr-CXR [25] are two publicly available CXR datasets with bounding box annotations. The RSNA dataset [34] consists of CXR images of patients infected with pneumonia and normal cases. The pneumonia findings are localized by bounding boxes. The VinDr CXR dataset [25] contains 15,000 training images and 3,000 testing images with critical findings localized by bounding boxes. The CXR images are annotated for the presence of 22 different lung abnormalities. CheXphoto [27] comprises a training set of natural photos and synthetic transformations of 10,507 CXR's from 3,000 unique patients that were sampled at random from the CheXpert training set. It also includes a validation set of natural and synthetic transformations of 234 CXR's from 200 patients in the CheXpert validation set. In addition to these large datasets, there are also several smaller CXR datasets with a approximately 1,000 images, annotated either programmatically or manually by radiologists. However, the literature does not contain a large dataset with original and corresponding WhatsApp-compressed CXR images." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.956 + ], + "angle": 0, + "content": "25888" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.212, + 0.09, + 0.305, + 0.104 + ], + "angle": 0, + "content": "Original Image" + }, + { + "type": "image", + "bbox": [ + 0.19, + 0.107, + 0.327, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.362, + 0.09, + 0.469, + 0.104 + ], + "angle": 0, + "content": "WhatsApp Image" + }, + { + "type": "image", + "bbox": [ + 0.346, + 0.107, + 0.484, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.537, + 0.09, + 0.63, + 0.104 + ], + "angle": 0, + "content": "Original Image" + }, + { + "type": "image", + "bbox": [ + 0.515, + 0.107, + 0.653, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.688, + 0.091, + 0.793, + 0.105 + ], + "angle": 0, + "content": "WhatsApp Image" + }, + { + "type": "image", + "bbox": [ + 0.672, + 0.107, + 0.808, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.19, + 0.227, + 0.328, + 0.333 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.346, + 0.227, + 0.484, + 0.333 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.515, + 0.227, + 0.653, + 0.333 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.671, + 0.227, + 0.809, + 0.333 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.349, + 0.908, + 0.391 + ], + "angle": 0, + "content": "Figure 1. Exemplified localization generated with DenseNet architecture on (a) original images and (b) their corresponding WhatsApp compressed images from CheXWhatsApp. Though the images look visually similar after compression, the predictions and localizations generated by GradCAM can significantly differ." + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.411, + 0.306, + 0.424 + ], + "angle": 0, + "content": "Original Image" + }, + { + "type": "image", + "bbox": [ + 0.189, + 0.428, + 0.329, + 0.535 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.363, + 0.411, + 0.469, + 0.424 + ], + "angle": 0, + "content": "WhatsApp Image" + }, + { + "type": "image", + "bbox": [ + 0.346, + 0.427, + 0.485, + 0.534 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.537, + 0.411, + 0.629, + 0.424 + ], + "angle": 0, + "content": "Original Image" + }, + { + "type": "image", + "bbox": [ + 0.515, + 0.427, + 0.653, + 0.533 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.687, + 0.411, + 0.791, + 0.425 + ], + "angle": 0, + "content": "WhatsApp Image" + }, + { + "type": "image", + "bbox": [ + 0.671, + 0.428, + 0.809, + 0.535 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.192, + 0.547, + 0.33, + 0.652 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.347, + 0.547, + 0.484, + 0.652 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.515, + 0.547, + 0.653, + 0.654 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.671, + 0.547, + 0.807, + 0.652 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.669, + 0.908, + 0.738 + ], + "angle": 0, + "content": "Figure 2. Exemplified localization generated with Faster-RCNN architecture on (a) original images and (b) their corresponding WhatsApp compressed images from CheXwhatsApp. The red boxes indicate the ground truth, and the blue boxes indicate the model's prediction. Though the images look visually similar after compression, the localizations generated by Faster-RCNN are either incorrectly placed (bottom-left), partially missing (top-right) or fully missing (top-left, bottom-right) in WhatsApp-compressed images when compared with the original images." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.746, + 0.483, + 0.781 + ], + "angle": 0, + "content": "3. Challenges in diagnosing Xray images received through Smart phones" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.484, + 0.901 + ], + "angle": 0, + "content": "In deep learning based models for image classification, the prediction outcome is provided by the model and explanation of predictions is obtained through class-activation maps. In mHealth settings, as envisaged here, such systems face two challenges: prediction instability and out-of-lung saliency. Object detection models employ a combined approach to prediction and explanation by lo" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.748, + 0.907, + 0.823 + ], + "angle": 0, + "content": "calizing the findings using bounding boxes. However, the output of object detection models can also suffer from localization instability when images are obtained through WhatsApp. This section delves deeper into these three challenges and defines the metrics used to evaluate them." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.832, + 0.818, + 0.848 + ], + "angle": 0, + "content": "3.1. The Prediction Instability Problem" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.908, + 0.9 + ], + "angle": 0, + "content": "PIP of a predictive model is defined as the probability of disagreement between the predictions on a randomly perturbed instance and the true instance." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "25889" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.092, + 0.303, + 0.106 + ], + "angle": 0, + "content": "3.1.1. Metrics to evaluate PIP" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.11, + 0.485, + 0.248 + ], + "angle": 0, + "content": "Prediction Instability Score (PI Score): Consider a deep learning model \\(\\mathsf{M}\\) that classifies a CXR image into one or multiple categories. Consider a dataset \\(D = \\{X_{i},\\mathsf{R}(X_{i})\\}_{i = 1}^{N}\\) consisting of \\(N\\) pairs of original and WhatsApp-compressed CXR images denoted by \\(X_{i}\\), and \\(\\mathsf{R}(X_i)\\) respectively. The estimate of PI Score due to image perturbation by WhatsApp can now be defined as the fraction of pairs where the predictions differed expressed as percentage." + }, + { + "type": "equation", + "bbox": [ + 0.091, + 0.253, + 0.517, + 0.312 + ], + "angle": 0, + "content": "\\[\nP I S c o r e (\\mathbb {M}; D) = \\frac {1}{N} \\left(\\sum_ {\\{X _ {i}, \\mathbb {R} (X _ {i}) \\} \\in D} P I \\left(X _ {i}, \\mathbb {R} \\left(X _ {i}\\right), \\mathbb {M}\\right)\\right) \\times 1 0 0\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.313, + 0.48, + 0.33 + ], + "angle": 0, + "content": "and \\(PI(X_i, \\mathbf{R}(X_i), \\mathbb{M}) = \\mathbf{I}(y(X_i; \\mathbb{M})) \\neq \\mathbf{I}(y(\\mathbf{R}(X_i); \\mathbb{M}))\\)" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.336, + 0.295, + 0.352 + ], + "angle": 0, + "content": "3.2. Out-of-Lung Saliency" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.358, + 0.484, + 0.419 + ], + "angle": 0, + "content": "For a Deep Learning Model diagnosing lung abnormalities in CXR images, out-of-lung saliency (OLS) is defined as the correlation between attributed pixels outside the lung region and model prediction." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.424, + 0.308, + 0.437 + ], + "angle": 0, + "content": "3.2.1. Metrics to evaluate OLS" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.442, + 0.484, + 0.669 + ], + "angle": 0, + "content": "Out of Lung Saliency Score (OLS Score): The OLS Score quantifies the impact of OLS on the diagnosis made by state-of-the-art models. It is derived using the Intersection Over Lung-Region (IOL) metric, which measures the extent of OLS on each CXR image. Let heatmap-region \\((y(x; \\mathbb{M}))\\) indicate the set of pixels identified by a saliency based feature attribution technique such as GradCAM[33] applied to model \\(\\mathbb{M}\\) on CXR image \\(x\\), resulting in prediction \\(y\\). Let \\(lung(x)\\) represents the pixels within image \\(x\\) that belong to the lung region. IOL is defined as the ratio of the number of pixels where the heatmap region and the lung region overlap in a CXR image, divided by the total number of pixels in the heatmap region. Mathematically, for a given CXR image \\(x\\) (represented by its pixels and intensity values), define" + }, + { + "type": "equation", + "bbox": [ + 0.091, + 0.675, + 0.484, + 0.72 + ], + "angle": 0, + "content": "\\[\nI O L (x; \\mathbb {M}) = \\frac {\\operatorname {P i x} (\\text {h e a t m a p - r e g i o n} (y (x ; \\mathbb {M})) \\cap l u n g (x))}{\\operatorname {P i x} (\\text {h e a t m a p - r e g i o n} (y (x ; \\mathbb {M})))} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.484, + 0.841 + ], + "angle": 0, + "content": "where \\(\\operatorname{P}\\mathrm{i}\\mathrm{x}(R(x))\\) counts the number of Pixels in region \\(R(x)\\), a subset of the pixels in the original image \\(x\\). The IOL value ranges from 0 to 1, where 0 indicates that prediction relies entirely on pixels outside the lung region and 1 indicates otherwise. The OLS Score of a model \\(\\mathbb{M}\\) on a dataset \\(D\\) consisting of \\(N\\) images is defined as the percentage of images in \\(D\\) with an IOL value less than a threshold \\(\\eta\\)." + }, + { + "type": "equation", + "bbox": [ + 0.097, + 0.847, + 0.484, + 0.902 + ], + "angle": 0, + "content": "\\[\nO L S \\operatorname {S c o r e} _ {\\eta} (D; \\mathbb {M}) = \\frac {1}{N} \\left(\\sum_ {x \\in D} I O L (x; \\mathbb {M}) < \\eta\\right) \\times 1 0 0 \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.152 + ], + "angle": 0, + "content": "The value of \\(\\eta\\) in this paper is set to 0.4. For brevity, we will exclude \\(\\eta\\) from the subscript. A high OLS Score suggests that a majority of the data is affected by OLS, while a low score implies lesser impact." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.161, + 0.834, + 0.178 + ], + "angle": 0, + "content": "3.3. The Localization Instability Problem" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.183, + 0.907, + 0.289 + ], + "angle": 0, + "content": "LIP of an object detection model is defined as the probability of disagreement between the localization of findings identified by the model that agree with the ground truth, on a randomly perturbed instance versus the true instance. Object detection models localize findings using bounding boxes and thus simultaneously provide both the prediction and explanation of a finding." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.296, + 0.725, + 0.31 + ], + "angle": 0, + "content": "3.3.1. Metrics to evaluate LIP" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.315, + 0.906, + 0.466 + ], + "angle": 0, + "content": "Localization Instability Score (LI Score): Consider a model M that utilizes bounding boxes to localize findings (instances of lung abnormalities) on both the original image \\( x \\) and its perturbed version \\( \\mathsf{R}(x) \\). Let \\( \\mathsf{Gt}_x \\) denote the set of ground truth localizations of \\( x \\). Let \\( y_{\\mathsf{f}}(x, \\mathsf{M}) \\) be the localization of a finding \\( \\mathsf{f} \\) returned by M. \\( y_{\\mathsf{f}}(x, \\mathsf{M}) \\) agrees with ground truth of the finding \\( \\mathsf{g}\\mathsf{t}_{x,\\mathsf{f}} \\in \\mathsf{G}\\mathsf{t}_x \\) if intersection over union (IOU) between localization returned by model \\( y_{\\mathsf{f}}(x, \\mathsf{M}) \\) and ground truth \\( \\mathsf{g}\\mathsf{t}_{x,\\mathsf{f}} \\) is greater than a threshold. ie, IOU \\( (y_{\\mathsf{f}}(x, \\mathsf{M}), \\mathsf{g}\\mathsf{t}_{x,\\mathsf{f}}) > t \\), where \\( t \\) is a threshold." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.467, + 0.907, + 0.603 + ], + "angle": 0, + "content": "Let \\(\\mathsf{F}_{x,t}\\) be the set of findings localized by M on image \\(x\\) w.r.t the ground-truth. \\(\\mathsf{F}_{x,t} = \\{\\mathsf{f}|\\mathsf{IOU}(y_{\\mathsf{f}}(x,\\mathsf{M}),\\mathsf{gt}_{x,\\mathsf{f}}) > t,\\forall \\mathsf{f}\\in \\mathsf{Gt}_x\\}\\) Let \\(\\mathsf{F}_{\\mathbb{R}(x),t}\\) be the set of findings localized by M on perturbed image \\(\\mathbb{R}(x)\\) w.r.t the ground-truth. \\(\\mathsf{F}_{\\mathbb{R}(x),t} = \\{\\mathsf{f}|\\mathsf{IOU}(y_{\\mathsf{f}}(\\mathbb{R}(x),\\mathsf{M}),\\mathsf{gt}_{x,\\mathsf{f}}) > t,\\forall \\mathsf{f}\\in \\mathsf{Gt}_x\\}\\). LI Score due to image perturbation by WhatsApp is defined as the fraction of image pairs, where the localization of findings on original and WhatsApp-compressed images that agree with ground truth differ, expressed as percentage." + }, + { + "type": "equation", + "bbox": [ + 0.525, + 0.625, + 0.907, + 0.667 + ], + "angle": 0, + "content": "\\[\nL I S c o r e _ {t} (D; \\mathbb {M}) = \\frac {1}{N} \\left(\\sum_ {x \\in D} \\mathrm {F} _ {x, t} \\neq \\mathrm {F} _ {\\mathrm {R} (x), t}\\right) \\times 1 0 0 \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.676, + 0.906, + 0.738 + ], + "angle": 0, + "content": "Currently, there does not exist a comprehensive dataset to evaluate the problems of PIP, OLS and LIP associated with CXR images transmitted through WhatsApp for mHealth applications." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.75, + 0.907, + 0.803 + ], + "angle": 0, + "content": "4. CheXwhatsapp : A new dataset of WhatsApp compressed CXR images paired with original images" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.906, + 0.87 + ], + "angle": 0, + "content": "In this section we review open source CXR datasets and discuss the steps involved in creation of a new CXR dataset called CheXwhatsapp that consists of both original and their WhatsApp images" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.906, + 0.902 + ], + "angle": 0, + "content": "Discussion of Existing CXR Datasets: We study the most extensively used datasets - NIH [38] and RSNA[34], in" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "25890" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.191, + 0.089, + 0.391, + 0.241 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.239, + 0.244, + 0.34, + 0.256 + ], + "angle": 0, + "content": "(a) Original image" + }, + { + "type": "image", + "bbox": [ + 0.612, + 0.089, + 0.811, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.649, + 0.244, + 0.771, + 0.256 + ], + "angle": 0, + "content": "(b) Compressed image" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.268, + 0.908, + 0.311 + ], + "angle": 0, + "content": "Figure 3. Illustration of a sample (a) original image and its corresponding (b) compressed image from CheXwhatsApp. The original image is of dimension \\(1024 \\times 1024\\) and has image size of 668KB and the WhatsApp image is of dimension \\(1024 \\times 1024\\) and has an image size of 134KB, indicating an \\(80\\%\\) compression." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.337, + 0.482, + 0.501 + ], + "angle": 0, + "content": "the literature. These are publicly available anonymized datasets for research purposes [6, 28, 36, 37]. It has been duly verified that the NIH and RSNA datasets permits modification and redistribution of data with proper citation. It is important to note that the collection of these datasets adhered to established protocols and guidelines, following the ethical regulations stipulated by the respective institutions. Informed consent was obtained from the participating subjects. Further information regarding these datasets can be found in [34, 38]. An overview of these datasets, are as follows." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.506, + 0.483, + 0.626 + ], + "angle": 0, + "content": "- NIH: NIH [38] dataset contains 86524 CXR images in train set and 25596 CXR images in test set from 30,805 unique patients. It is a multi-label dataset where patients can have normal CXR or be associated with one or more of the 14 different pathologies: Atelectasis, Cardiomegaly, Effusion, Infiltration, Mass, Nodule, Pneumonia, Pneumothorax, Consolidation, Edema, Emphysema, Fibrosis, Pleural-Thickening, and Hernia." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.627, + 0.483, + 0.716 + ], + "angle": 0, + "content": "- RSNA: The RSNA[34] consists of CXR images categorized as either pneumonia or normal cases, with manually annotated bounding boxes for pneumonia by expert radiologists. The dataset consists of 29,684 CXR images in training and validation set, as provided by the original authors." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.506, + 0.483, + 0.716 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.106, + 0.718, + 0.483, + 0.748 + ], + "angle": 0, + "content": "Additional details regarding the dataset can be found in the Supplementary Material." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.758, + 0.483, + 0.79 + ], + "angle": 0, + "content": "4.1. CheXwhatsapp: A new dataset of original and WhatsApp compressed images" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.483, + 0.901 + ], + "angle": 0, + "content": "To train models that are robust to image perturbations due to compression caused by low-bandwidth messaging applications like WhatsApp, it is crucial to have a dataset that includes compressed images generated by these messaging applications. However, currently, no open-source datasets are available with image pairs consisting of original CXR images and their corresponding compressed versions. To" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.337, + 0.905, + 0.381 + ], + "angle": 0, + "content": "address this gap, we created a new in-house dataset using WhatsApp and the previously discussed datasets, called CheXwhatsapp." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.383, + 0.906, + 0.533 + ], + "angle": 0, + "content": "CheXwhatsapp consists of 141,804 pairs of original and WhatsApp-compressed CXR images. CheXwhatsapp is an amalgamation of the two datasets: NIH [38] and RSNA [34] along with their corresponding WhatsApp-compressed images. The original and corresponding WhatsApp-compressed images in NIH [38] are known as NIH-WA. Similarly, the original and corresponding WhatsApp-compressed images in RSNA [34] are known as RSNA-WA. CheXwhatsapp is the union of NIH-WA and RSNA-WA." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.535, + 0.906, + 0.746 + ], + "angle": 0, + "content": "To create the image pairs, all the images in NIH [38] and RSNA [34] were sequentially sent through WhatsApp to a fixed mobile number. The creation of CheXwhatisApp is performed in a fully automated manner. The primary challenge in automating this process is to synchronize sending images to WhatsApp and receiving images from WhatsApp. This is primarily due to two reasons - 1) WhatsApp sends images out-of-order in certain scenarios and 2) the metadata of images is lost after sending, and images are renamed to some random name by WhatsApp at the recipient's end. Since we do not have control over WhatsApp API, the image name is also sent along with the image. The steps involved in creating the CheXwhatisApp are given below:" + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.75, + 0.905, + 0.84 + ], + "angle": 0, + "content": "1. Preprocess the image by adding the image name to the image itself. This is done by adding a small white border of 15pixel \\(\\times\\) image-width at the bottom of the image and placing the image name in the white border. The preprocessed images are uploaded to an Amazon S3 bucket to be sent to the WhatsApp Business API." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.84, + 0.905, + 0.901 + ], + "angle": 0, + "content": "2. The pre-processed images are send by the WhatsApp business API one by one to a different WhatsApp account. This is established by using platforms for WhatsApp bulk-messaging. Such platforms are con" + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.75, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "25891" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.11, + 0.092, + 0.482, + 0.121 + ], + "angle": 0, + "content": "figured using a phone number enabled with WhatsApp and allow bulk-messaging of images and text." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.122, + 0.483, + 0.167 + ], + "angle": 0, + "content": "3. The images received via WhatsApp are downloaded to an edge device, from where they are uploaded to a cloud service provider." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.168, + 0.483, + 0.228 + ], + "angle": 0, + "content": "4. The white border at the bottom of the images is cropped to read the image name via an OCR technique called Tesseract. The image name read via OCR is used to rename the image." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.122, + 0.483, + 0.228 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.232, + 0.484, + 0.488 + ], + "angle": 0, + "content": "At the conclusion of the WhatsApp-compression experimentation, it was observed that the 141,804 WhatsApp compressed CXR images in CheXwhatsapp occupied only 18.43 GB of memory, whereas the original 141,804 CXR images occupied 54.80 GB. This indicates a compression factor of \\(66.36\\%\\) achieved by WhatsApp. However, this compression can alter the intricate intensity values of the images and potentially lead to inaccurate diagnoses if the models are trained solely on high-resolution images. Figure 3 presents an original and corresponding compressed images from NIH-WA in CheXwhWhatsApp. The existing state-of-the-art (SOTA) models predicted different diagnoses for the high-resolution (original) image compared to the low-resolution (compressed) image. Additionally, the activation maps used for localization showed significant differences, often highlighting non-lung regions. Clearly, this calls for more research in this area." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.503, + 0.484, + 0.539 + ], + "angle": 0, + "content": "5. Benchmarking using CheXwhat-sApp dataset" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.547, + 0.483, + 0.712 + ], + "angle": 0, + "content": "The study shows that introducing the proposed WhatsApp dataset mitigates the effect of PIP, OLS and LIP of SOTA models. The NIH and NIH-WA datasets are evaluated using classification models. The RSNA and RSNA-WA datasets are evaluated using object detection models. This section begins with a discussion of the benchmarking setup for classification and object detection models. As stated previously, the PI Score and OLS Score are employed to analyze the effect of PIP and OLS on the prediction outcome of classification models. LI Score is used to analyze LIP in object detection models." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.483, + 0.87 + ], + "angle": 0, + "content": "Benchmark setup. State-of-the-art models considered in this study are DenseNet [17], ResNet [16], ResNeXt [40] and Inception [35]. These models are implemented using Pytorch. The models are trained using the Adam optimizer with an initial learning rate of 1e-4. The input image dimension are \\(512 \\times 512\\). To fine-tune the learning process, ReduceLROOnPlateau scheduler is employed with patience 4 and a factor 0.1. The batch size for training data is 16 and the models are trained for 25 epochs." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.901 + ], + "angle": 0, + "content": "For object detection tasks, the Faster-RCNN [30] is trained using SGD optimizer with learning-rate 0.0055, mo" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.228 + ], + "angle": 0, + "content": "mentum 0.9 and weight-decay 0.0005. The training batch-size is 10. The learning-rate schedular is StepLR (step-size=3, gamma = 0.1) and the model is trained for 15 epochs. YOLOV5s [19] and YOLOV8s [20] models are trained with Auto optimizer with learning-rate \\(\\mathrm{lr}0 = 0.001\\), \\(\\mathrm{lrf} = 0.001\\), final-lr \\(= lr0 \\times lr f\\) and momentum \\(= 0.937\\). The training batch-size is 16 and the models are trained for 50 epochs. The training is performed using 8 NVIDIA Tesla V100-SXM2-32GB GPUs. The training code is available2." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.241, + 0.905, + 0.422 + ], + "angle": 0, + "content": "Benchmarking results. In the experiments, SOTA models DenseNet [17], ResNet [16], ResNeXt [40] and Inception [35] are trained using the existing dataset- NIH and the proposed paired dataset - NIH-WA. The models trained using training images from existing and proposed datasets are evaluated for PIP and OLS using the original and WhatsApp-compressed images in the test dataset. The results are presented in Table 1. The results clearly indicate that models trained using NIH-WA reduces the effect of PIP, when compared with those trained using NIH. The PI Score of networks trained using NIH-WA is upto \\(80\\%\\) lower than those trained using NIH." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.424, + 0.906, + 0.695 + ], + "angle": 0, + "content": "To evaluate OLS Score of SOTA models trained using the existing dataset - NIH and the proposed paired dataset - NIH-WA, heatmaps of CXR images were generated using GradCAM [33], GradCAM++ [5] and EigenCAM [24]. For each CXR image, the lung region was identified using a U-Net [31] to evaluate whether the model's activation map falls within the lung regions or extends into non-lung regions. We use the highest activation region in the heatmap to determine out-of-lung saliency, considering activation values greater than 0.5 (on a scale of 0 to 1). The lung region extracted by U-Net from the CXR image, along with the highest activation region in the heatmap is utilized for the calculation of the IOL Score, which is then used to evaluate OLS Score. The results (Table 2) clearly show that models trained using NIH-WA reduces the effect of OLS, when compared to those trained using NIH. Moreover, the problem of OLS is further exacerbated in CXR images compressed by WhatsApp." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.697, + 0.906, + 0.862 + ], + "angle": 0, + "content": "The SOTA object detection networks, Faster-RCNN [30],YOLOv5s [19] and YOLOv8s [20] are trained using the existing bounding box CXR dataset - RSNA and the proposed paired dataset - RSNA-WA. These models are evaluated for LIP using the original and WhatsApp-compressed images in the test dataset. From the LI Score results evaluated with threshold \\( t = 0.7 \\) (Table 3), it is evident that the networks trained using RSNA-WA reduces the problem of LIP by upto \\( 81\\% \\) when compared with those trained using RSNA. LI Score of networks trained using RSNA and RSNA-WA for different values of threshold \\( t \\) is given in" + }, + { + "type": "page_footnote", + "bbox": [ + 0.513, + 0.875, + 0.905, + 0.9 + ], + "angle": 0, + "content": "https://github.com/cheXwhatsApp/cheXwhatsApp_benchmark" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "25892" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.374, + 0.089, + 0.625, + 0.203 + ], + "angle": 0, + "content": "
ModelsPI Score
NIHNIH-WA
DenseNet15.743.31 (79%)
InceptionNet15.083.14 (79%)
ResNet-5014.603.30 (77%)
ResNeXt-5016.863.90 (77%)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.213, + 0.907, + 0.255 + ], + "angle": 0, + "content": "Table 1. PI Score for different networks trained on NIH and NIH-WA and evaluated on original and WhatsApp test images. From the table, it is evident that networks trained on NIH-WA has substantially lower prediction instability problem (PIP) compared to those trained on NIH." + }, + { + "type": "table", + "bbox": [ + 0.134, + 0.268, + 0.865, + 0.465 + ], + "angle": 0, + "content": "
ModelsDatsetOLS Score(GradCAM)OLS Score(GradCAM++)OLS Score(EigenCAM)
OriginalWhatsAppOriginalWhatsAppOriginalWhatsApp
DenseNetNIH18.7926.6930.7941.2546.0660.66
NIH-WA12.53 (33%)21.05 (21%)13.12 (57%)29.39 (29%)13.16 (71%)20.57 (66%)
InceptionNIH37.7547.4552.3362.2195.5896.70
NIH-WA22.10 (41%)32.64 (31%)18.04 (66%)34.26 (45%)94.55 (01%)95.31 (01%)
ResNetNIH24.6533.6542.1251.4097.5198.89
NIH-WA14.94 (39%)15.34 (54%)21.89 (48%)39.58 (23%)95.01 (02%)95.16 (03%)
ResNeXtNIH33.1242.5661.2469.3999.6799.79
NIH-WA11.81 (64%)26.26 (38%)10.51 (82%)26.58 (62%)96.28 (03%)96.48 (03%)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.472, + 0.907, + 0.515 + ], + "angle": 0, + "content": "Table 2. OLS Score score of attribution maps generated using different networks trained on NIH and NIH-WA training dataset and evaluated on original and WhatsApp images in test dataset. The results show that attribution maps generated using models trained on NIH-WA has lower out-of-lung saliency problem (OLS) compared to those trained on NIH." + }, + { + "type": "table", + "bbox": [ + 0.365, + 0.527, + 0.633, + 0.627 + ], + "angle": 0, + "content": "
ModelsLI Score
RSNARSNA-WA
YOLOV5s3.010.85 (72%)
YOLOV8s4.020.78 (81%)
Faster-RCNN12.4310.86 (13%)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.637, + 0.907, + 0.68 + ], + "angle": 0, + "content": "Table 3. LI Score for different networks trained on the bounding-box dataset - RSNA and RSNA-WA and evaluated using original and WhatsApp test dataset. From the table, it is evident that the networks trained on RSNA-WA reduces loalization instability problem (LIP) by upto \\(81\\%\\) compared to those trained on RSNA." + }, + { + "type": "image", + "bbox": [ + 0.1, + 0.694, + 0.335, + 0.814 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.134, + 0.818, + 0.299, + 0.832 + ], + "angle": 0, + "content": "(a) LI Score of YOLOv5s [19]" + }, + { + "type": "image", + "bbox": [ + 0.349, + 0.694, + 0.58, + 0.814 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.381, + 0.818, + 0.545, + 0.832 + ], + "angle": 0, + "content": "(b) LI Score of YOLOv8s [20]" + }, + { + "type": "image", + "bbox": [ + 0.596, + 0.694, + 0.825, + 0.814 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.62, + 0.818, + 0.8, + 0.832 + ], + "angle": 0, + "content": "(c) LI Score of Faster-RCNN [30]" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.844, + 0.907, + 0.873 + ], + "angle": 0, + "content": "Figure 4. LI Score of Faster-RCNN [30], YOLOv5s [19] and YOLOv8s [20] on RSNA and RSNA-WA. From the results, it can be seen that models trained on RSNA-WA mitigate the effect of loalization instability problem (LIP) for different values of threshold \\( t \\)." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "25893" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.151 + ], + "angle": 0, + "content": "Figure 4. The figure also shows that networks trained using RSNA-WA have substantially lower LIP for all values of \\( t \\). In particular, YOLO trained using RSNA-WA reduces LI score by more than \\( 70\\% \\)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.153, + 0.483, + 0.348 + ], + "angle": 0, + "content": "For LI score, the choice of parameter \\( t \\) represents the threshold of overlap between predicted and true boundingboxes. If the parameter \\( t \\) is too small, then most predicted bounding boxes will overlap with the true bounding boxes in both the original and WhatsApp images and LIP will be low. However, if \\( t \\) is too large, then only few predicted bounding boxes will overlap with true bounding boxes in both original and WhatsApp images and again LIP will be low. Therefore, we must select a value for \\( t \\) that balances these two scenarios. This optimal value may vary depending on the dataset. In this paper, we selected \\( t \\) that corresponds to the highest LI Score to study the maximum localization instability." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.348, + 0.483, + 0.529 + ], + "angle": 0, + "content": "To understand the generalizability of model trained on CheXwhatsapp to images compressed by other compression algorithms such as ISO JPEG compression, we created test datasets by compressing the original test images using various JPEG quality factors - 20, 40, 60, 80, and 90 with OpenCV library. Our observations indicated that instability increases as the quality factor decreases. Furthermore, models trained with CheXwhatsapp demonstrated an average of \\(26\\%\\) lower prediction instability and \\(10\\%\\) lower localization instability compared to models trained on original images alone. Additional results are available in supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.544, + 0.339, + 0.559 + ], + "angle": 0, + "content": "6. Discussion and Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.569, + 0.483, + 0.719 + ], + "angle": 0, + "content": "In this paper, we present a large-scale CheXwhatsApp dataset to aid the development of mHealth applications for diagnosing CXR images. The existing datasets consist of high resolution images and the models trained on these datasets are found to face high PIP, OLS and LIP. To study this, we proposed quantitative metrics called PI, OLS, and LI scores and obtained these metrics using the WhatsApp compressed CXR images. The high scores indicate evidence of the problem, which is due to the compression techniques used by mHealth applications." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.483, + 0.901 + ], + "angle": 0, + "content": "The growing body of research on using smart-phones and WhatsApp [3, 8, 22, 23, 32, 41] in the medical domain calls for large datasets of WhatsApp-compressed images. Our proposed dataset comprise 141,804 pairs of high-resolution and their WhatsApp compressed CXR images. As there are no available WhatsApp CXR datasets, this proposed dataset is a valuable resource for the research community to address the challenges posed by image compression and transmission over low-bandwidth networks. Further, this proposed dataset was not easy to create. To create this dataset, we used a WhatsApp business api which incurred charges per each compression and we needed to" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.242 + ], + "angle": 0, + "content": "sequentially send the images from one phone and receive them on another phone, followed by proper mapping of the high resolution and their corresponding WhatsApp compressed images. The conducted experiments showed that the models trained with this dataset demonstrate much better visual and numerical results when tested on unseen images, thus highlighting its high robustness. As observed from Tables 1, 2, 3, training with proposed dataset helped reduce PI, OLS and LI scores significantly, demonstrating the importance of our proposed dataset." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.243, + 0.905, + 0.424 + ], + "angle": 0, + "content": "Furthermore, the OLS scores on the high resolution dataset improved by \\(70\\%\\) in some cases when trained with proposed dataset. This shows that the proposed dataset can aid in better explainability of results with fewer predictions outside the lung region. In healthcare, explainability plays a vital role in assisting physicians in making informed decisions. While this paper is the first to offer a large paired dataset of original and WhatsApp-compressed CXR images, there are limitations. One limitation is that we did not collect a new dataset; instead it is derived from existing datasets. Another limitation is that we did not use the new dataset to create or improve existing mHealth applications." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.425, + 0.905, + 0.606 + ], + "angle": 0, + "content": "Additionally, we also experimented with available JPEG compression algorithms and used them to train the models. It was interesting to note that when trained with the JPEG compressed CXR images, the PIP, OLS and LIP further increased. More details of this experimentation are provided in the supplementary material. This highlights the importance of the proposed dataset in revolutionizing healthcare delivery with mHealth applications, especially in resource-constrained regions, by facilitating rapid patient triaging and assisting inexperienced rural doctors in interpretation. We encourage researchers to use this dataset to advance research in CXR diagnosis for mHealth applications." + }, + { + "type": "title", + "bbox": [ + 0.515, + 0.62, + 0.611, + 0.636 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.522, + 0.645, + 0.905, + 0.715 + ], + "angle": 0, + "content": "[1] Mariamma Antony, Siva Teja Kakileti, Rachit Shah, Sabyasachi Sahoo, Chiranjib Bhattacharyya, and Geetha Manjunath. Challenges of ai driven diagnosis of chest x-rays transmitted through smart phones: a case study in Covid-19. Scientific Reports, 13(1):18102, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.522, + 0.716, + 0.907, + 0.758 + ], + "angle": 0, + "content": "[2] Richa Arora. The training and practice of radiology in india: current trends. Quantitative imaging in medicine and surgery, 4(6):449-450, 2014. 25525575[pmid]. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.522, + 0.759, + 0.907, + 0.829 + ], + "angle": 0, + "content": "[3] Vera Helen Buss, Stuart Leesong, Margo Barr, Marlien Varnfield, and Mark Harris. Primary prevention of cardiovascular disease and type 2 diabetes mellitus using mobile health technology: systematic review of the literature. Journal of medical Internet research, 22(10):e21159, 2020. 1, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.522, + 0.83, + 0.905, + 0.884 + ], + "angle": 0, + "content": "[4] Aurelia Bustos, Antonio Pertusa, Jose-Maria Salinas, and Maria De La Iglesia-Vaya. Padchest: A large chest x-ray image dataset with multi-label annotated reports. Medical image analysis, 66:101797, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.522, + 0.887, + 0.905, + 0.902 + ], + "angle": 0, + "content": "[5] Aditya Chattopadhy, Anirban Sarkar, Prantik Howlader," + }, + { + "type": "list", + "bbox": [ + 0.522, + 0.645, + 0.907, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "25894" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.484, + 0.161 + ], + "angle": 0, + "content": "and Vineeth N Balasubramanian. Grad-cam++: Generalized gradient-based visual explanations for deep convolutional networks. In 2018 IEEE Winter Conference on Applications of Computer Vision (WACV), pages 839-847, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.164, + 0.484, + 0.245 + ], + "angle": 0, + "content": "[6] Muhammad EH Chowdhury, Tawsifur Rahman, Amith Khandakar, Rashid Mazhar, Muhammad Abdul Kadir, Zaid Bin Mahbub, Khandakar Reajul Islam, Muhammad Salman Khan, Atif Iqbal, Nasser Al Emadi, et al. Can ai help in screening viral and covid-19 pneumonia? IEEE Access, 8:132665-132676, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.248, + 0.483, + 0.289 + ], + "angle": 0, + "content": "[7] Alex J. DeGrave, Joseph D. Janizek, and Su-In Lee. Ai for radiographic Covid-19 detection selects shortcuts over signal. Nature Machine Intelligence, 3(7):610–619, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.291, + 0.483, + 0.346 + ], + "angle": 0, + "content": "[8] Olga Di Fede, Gaetano La Mantia, Mario GCA Cimino, and Giuseppina Campisi. Protection of patient data in digital oral and general health care: A scoping review with respect to the current regulations. Oral, 3(2):155-165, 2023. 1, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.348, + 0.483, + 0.403 + ], + "angle": 0, + "content": "[9] Robert Geirhos, Jörn-Henrik Jacobsen, Claudio Michaelis, Richard Zemel, Wieland Brendel, Matthias Bethge, and Felix A. Wichmann. Shortcut learning in deep neural networks. Nature Machine Intelligence, 2(11):665-673, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.405, + 0.482, + 0.445 + ], + "angle": 0, + "content": "[10] Daniele Giansanti. Teleradiology today: the quality concept and the Italian point of view. Telemedicine and e-Health, 23 (5):453–455, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.447, + 0.483, + 0.502 + ], + "angle": 0, + "content": "[11] Daniele Giansanti, Livia Castrichella, and Maria Rosaria Giovagnoli. The design of a health technology assessment system in telepathology. Telemedicine and e-Health, 14(6): 570-575, 2008. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.504, + 0.483, + 0.559 + ], + "angle": 0, + "content": "[12] Daniele Giansanti, Marco Pochini, and Maria Rosaria Giovagnoli. How tablet technology is going to change cooperative diagnosis in the cytology e-laboratory. Telemedicine and e-Health, 19(12):991-993, 2013." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.561, + 0.483, + 0.616 + ], + "angle": 0, + "content": "[13] Daniele Giansanti, Marco Pochini, and Maria Rosaria Giovagnoli. Integration of tablet technologies in the e-laboratory of cytology: a health technology assessment. Telemedicine and e-Health, 20(10):909–915, 2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.617, + 0.483, + 0.658 + ], + "angle": 0, + "content": "[14] Daniele and Giansanti. Whatsapp in mhealth: an overview on the potentialities and the opportunities in medical imaging. Mhealth, 6, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.66, + 0.483, + 0.687 + ], + "angle": 0, + "content": "[15] GSMA. The state of mobile internet connectivity 2022, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.689, + 0.483, + 0.745 + ], + "angle": 0, + "content": "[16] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 770-778, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.746, + 0.483, + 0.802 + ], + "angle": 0, + "content": "[17] Gao Huang, Zhuang Liu, Laurens Van Der Maaten, and Kilian Q. Weinberger. Densely connected convolutional networks. In 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2261-2269, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.803, + 0.483, + 0.886 + ], + "angle": 0, + "content": "[18] Jeremy Irvin, Pranav Rajpurkar, Michael Ko, Yifan Yu, Silviana Ciurea-Ilicus, Chris Chute, Henrik Marklund, Behzad Haghgoo, Robyn Ball, Katie Shpanskaya, et al. Chexpert: A large chest radiograph dataset with uncertainty labels and expert comparison. In Proceedings of the AAAI conference on artificial intelligence, pages 590-597, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.887, + 0.388, + 0.901 + ], + "angle": 0, + "content": "[19] Glenn Jocher. Ultralytics yolov5, 2020. 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.092, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "[20] Glenn Jocher, Ayush Chaurasia, and Jing Qiu. Ultralytics yolov8, 2023. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.122, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[21] Alistair EW Johnson, Tom J Pollard, Seth J Berkowitz, Nathaniel R Greenbaum, Matthew P Lungren, Chih-ying Deng, Roger G Mark, and Steven Horng. Mimic-cxr, a de-identified publicly available database of chest radiographs with free-text reports. Scientific data, 6(1):317, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.192, + 0.905, + 0.218 + ], + "angle": 0, + "content": "[22] Arjun Kalyanpur. WhatsApp and its role in teleradiology. Indian Journal of Radiology and Imaging, 2024. 1, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.22, + 0.905, + 0.286 + ], + "angle": 0, + "content": "[23] J Austin Lee, Benjamin W Wachira, John Kennedy, Nicholas Asselin, and Nee-Kofi Mould-Millman. Utilisation of what-sapp for emergency medical services in garissa, kenya. African Journal of Emergency Medicine, 14(1):38-44, 2024. 1, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.289, + 0.905, + 0.343 + ], + "angle": 0, + "content": "[24] Mohammed Bany Muhammad and Mohammed Yeasin. Eigen-cam: Class activation map using principal components. In 2020 international joint conference on neural networks (IJCNN), pages 1-7. IEEE, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.345, + 0.905, + 0.413 + ], + "angle": 0, + "content": "[25] Ha Q Nguyen, Khanh Lam, Linh T Le, Hieu H Pham, Dat Q Tran, Dung B Nguyen, Dung D Le, Chi M Pham, Hang TT Tong, Diep H Dinh, et al. Vindr-cxr: An open dataset of chest x-rays with radiologist's annotations. Scientific Data, 9(1):429, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.415, + 0.905, + 0.469 + ], + "angle": 0, + "content": "[26] Unathi Ntja, Jacques Janse van Rensburg, and Gina Joubert. Diagnostic accuracy and reliability of smartphone captured radiologic images communicated via Whatsapp®. African Journal of Emergency Medicine, 12(1):67-70, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.471, + 0.905, + 0.552 + ], + "angle": 0, + "content": "[27] Nick A Phillips, Pranav Rajpurkar, Mark Sabini, Rayan Krishnan, Sharon Zhou, Anuj Pareek, Nguyen Minh Phu, Chris Wang, Mudit Jain, Nguyen Duong Du, et al. Chexphoto: \\(10,000+\\) photos and transformations of chest x-rays for benchmarking deep learning robustness. In Machine Learning for Health, pages 318-327. PMLR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.554, + 0.905, + 0.621 + ], + "angle": 0, + "content": "[28] Sivaramakrishnan Rajaraman, Jenifer Siegelman, Philip O. Alderson, Lucas S. Folio, Les R. Folio, and Sameer K. Antani. Iteratively pruned deep learning ensembles for Covid-19 detection in chest x-rays. IEEE Access, 8:115041-115050, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.624, + 0.905, + 0.691 + ], + "angle": 0, + "content": "[29] Eduardo P Reis, Joselisa PQ De Paiva, Maria CB Da Silva, Guilherme AS Ribeiro, Victor F Paiva, Lucas Bulgarelli, Henrique MH Lee, Paulo V Santos, Vanessa M Brito, Lucas TW Amaral, et al. Brax, brazilian labeled chest x-ray dataset. Scientific Data, 9(1):487, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.693, + 0.905, + 0.748 + ], + "angle": 0, + "content": "[30] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. IEEE Transactions on Pattern Analysis and Machine Intelligence, 39(6):1137-1149, 2017. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.749, + 0.905, + 0.817 + ], + "angle": 0, + "content": "[31] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention - MICCAI 2015, pages 234-241, Cham, 2015. Springer International Publishing. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.818, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[32] Ravi Saini, Madhan Jeyaraman, Naveen Jeyaraman, Vijay Kumar Jain, Swaminathan Ramasubramanian, and Karthikeyan P Iyengar. Advancing orthopaedic trauma care through Whatsapp: An analysis of clinical and non-clinical applications, challenges, and future directions. World Journal of Orthopedics, 15(6):529, 2024. 1, 8" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "25895" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.174 + ], + "angle": 0, + "content": "[33] Ramprasaath R. Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In 2017 IEEE International Conference on Computer Vision (ICCV), pages 618-626, 2017. 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.178, + 0.482, + 0.272 + ], + "angle": 0, + "content": "[34] George Shih, Carol C Wu, Safwan S Halabi, Marc D Kohli, Luciano M Prevedello, Tessa S Cook, Arjun Sharma, Judith K Amorosa, Veronica Arteaga, Maya Galperin-Aizenberg, et al. Augmenting the national institutes of health chest radiograph dataset with expert annotations of possible pneumonia. Radiology. Artificial intelligence, 1(1), 2019. 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.276, + 0.482, + 0.33 + ], + "angle": 0, + "content": "[35] Christian Szegedy, Sergey Ioffe, Vincent Vanhoucke, and Alexander A Alemi. Inception-v4, inception-resnet and the impact of residual connections on learning. In Thirty-first AAAI conference on artificial intelligence, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.332, + 0.482, + 0.413 + ], + "angle": 0, + "content": "[36] Yu-Xing Tang, You-Bao Tang, Yifan Peng, Ke Yan, Mohammadhadi Bagheri, Bernadette A Redd, Catherine J Brandon, Zhiyong Lu, Mei Han, Jing Xiao, et al. Automated abnormality classification of chest radiographs using deep convolutional neural networks. NPJ digital medicine, 3(1):1-8, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.417, + 0.482, + 0.47 + ], + "angle": 0, + "content": "[37] Linda Wang, Zhong Qiu Lin, and Alexander Wong. Covidnet: a tailored deep convolutional neural network design for detection of Covid-19 cases from chest x-ray images. Scientific Reports, 10(1):19549, 2020. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.473, + 0.482, + 0.555 + ], + "angle": 0, + "content": "[38] Xiaosong Wang, Yifan Peng, Le Lu, Zhiyong Lu, Mohammadhadi Bagheri, and Ronald M. Summers. ChestXray8: Hospital-scale chest x-ray database and benchmarks on weakly-supervised classification and localization of common thorax diseases. In 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2017. 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.558, + 0.482, + 0.652 + ], + "angle": 0, + "content": "[39] Xiaosong Wang, Yifan Peng, Le Lu, Zhiyong Lu, Mohammadhadi Bagheri, and Ronald M Summers. Chestxray8: Hospital-scale chest x-ray database and benchmarks on weakly-supervised classification and localization of common thorax diseases. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2097-2106, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.656, + 0.482, + 0.722 + ], + "angle": 0, + "content": "[40] Saining Xie, Ross Girshick, Piotr Dólár, Zhuowen Tu, and Kaiming He. Aggregated residual transformations for deep neural networks. In 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 5987-5995, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.726, + 0.482, + 0.793 + ], + "angle": 0, + "content": "[41] Esmaeel Yaagoob, Regina Lee, Michelle Stubbs, Fatimah Shuaib, Raja Johar, and Sally Chan. WhatsApp-based intervention for people with type 2 diabetes: A randomized controlled trial. Nursing & Health Sciences, 26(2):e13117, 2024. 1, 8" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.793 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "25896" + } + ] +] \ No newline at end of file diff --git a/2025/CheXwhatsApp_ A Dataset for Exploring Challenges in the Diagnosis of Chest X-rays through Mobile Devices/8cf11147-8ad4-4f89-a3ee-6b5492c8beb0_origin.pdf b/2025/CheXwhatsApp_ A Dataset for Exploring Challenges in the Diagnosis of Chest X-rays through Mobile Devices/8cf11147-8ad4-4f89-a3ee-6b5492c8beb0_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..f349760be7bc4c70329e1ab6dd26b521b62eb025 --- /dev/null +++ b/2025/CheXwhatsApp_ A Dataset for Exploring Challenges in the Diagnosis of Chest X-rays through Mobile Devices/8cf11147-8ad4-4f89-a3ee-6b5492c8beb0_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4b50aa51e0d7c7d4bfaf5c06a149420bbba7177e7f579be0b92ffa0a9176445d +size 3410810 diff --git a/2025/CheXwhatsApp_ A Dataset for Exploring Challenges in the Diagnosis of Chest X-rays through Mobile Devices/full.md b/2025/CheXwhatsApp_ A Dataset for Exploring Challenges in the Diagnosis of Chest X-rays through Mobile Devices/full.md new file mode 100644 index 0000000000000000000000000000000000000000..fcb367a9b5039e50d145c0af42e4264dd9ddb9aa --- /dev/null +++ b/2025/CheXwhatsApp_ A Dataset for Exploring Challenges in the Diagnosis of Chest X-rays through Mobile Devices/full.md @@ -0,0 +1,295 @@ +# CheXwhatsapp: A Dataset for Exploring Challenges in the Diagnosis of Chest X-rays through Mobile Devices + +Mariamma Antony1 + +Rajiv Porana + +Sahil M Lathiya1 + +Siva Teja Kakileti2 + +Chiranjib Bhattacharyya + +$^{1}$ Indian Institute of Science, Bangalore, India $^{2}$ Niramai Health Analytix, Bangalore, India + +# Abstract + +Mobile health (mHealth) has emerged as a transformative solution to enhance healthcare accessibility and affordability, particularly in resource-constrained regions and low-to-middle-income countries. mHealth leverages mobile platforms to improve healthcare accessibility, addressing radiologist shortages in low-resource settings by enabling remote diagnosis and consultation through mobile devices. Mobile phones allow healthcare workers to transmit radiographic images, such as chest X-rays (CXR), to specialists or AI-driven models for interpretation. However, AI-based diagnosis using CXR images shared via apps like WhatsApp suffers from reduced predictability and explainability due to compression artifacts, and there is a lack of datasets to systematically study these challenges. To address this, we introduce CheXwhatsapp, a dataset of 141,804 paired original and WhatsApp-compressed CXR images. We present a benchmarking study which shows the dataset improves prediction stability and explainability of state-of-the-art models by up to $80\%$ , while also enhancing localization performance. CheXwhatsapp is open-sourced to support advancements in mHealth applications for CXR analysis1. + +# 1. Introduction + +Recent studies highlight the wide opportunities of messaging apps such as WhatsApp as an efficient tool for remote image sharing in mHealth applications across digital healthcare, including digital radiology [1, 10], digital pathology [11-13], cardiac imaging [14], orthopedics [32], diabetes management [41], and oral health [8]. mHealth provides a platform for health practitioners to enable disease management and speed-up diagnosis by facilitating the transfer of medical images, providing advice, facilitating drug adherence, and monitoring outbreaks. The outbreak of COVID-19 pandemic has accelerated the adoption of mHealth for digital radiology, such as for analyzing Chest X-ray (CXR) + +imaging. + +mHealth is especially beneficial in resource constrained areas such as Low-to-Middle-Income countries (LMICs) where the ratio of radiologists to radiographers is significantly imbalanced. For example, in India, there is only one radiologist for every 100,000 people [2]. Compounding the issue is the fact that most radiologists are located in urban areas, often far from the majority of the X-ray imaging facilities located in rural areas. Due to the high demand for chest radiography and the global shortage of radiologists, numerous applications for AI-based diagnosis of CXR images have been developed. + +In recent years, smartphone penetration has significantly increased, and instant messaging applications such as WhatsApp have a large user-base with 535 million users in India, 118.5 million users in Brazil, and a penetration rate exceeding $97\%$ in African nations like Kenya, South Africa, and Nigeria [15]. This presents new opportunities for the quick and easy sharing of radiographic images to aid clinical decision-making [3, 22, 23, 26, 32]. In rural hospitals and diagnostic centers without on-site radiologists, the ability to transmit anonymized images instantly to a mHealth app for interpretation can be extremely valuable to facilitate rapid triaging of infected individuals [22]. + +Leveraging common messaging platforms like WhatsApp eliminates the need for additional training since it is a widely used communication medium. However, it is important to note that these messaging platforms often compress images by upto $90\%$ to overcome bandwidth limitations. In medial applications, it is crucial to obtain explanation of predictions or localizations of findings in CXR images. Existing literature [1, 7, 9] identifies challenges related to explainability, such as predictions driven by features present outside the lung region in AI-based diagnostic systems for CXR images. Specifically, these systems suffer from the following challenges. + +Prediction Instability Problem(PIP): PIP refers to a situation where model's prediction differs between the original image and the image transmitted through WhatsApp. This undesirable effect can have critical implications in mHealth + +applications. + +Out of Lung Saliency(OLS): In medical diagnosis, it is crucial to provide explainable predictions. Saliency maps, which identify localized regions in an image that significantly contribute to the prediction, are commonly used to provide explanations. Out of Lung Saliency refers to the scenario where the model uses regions outside the lungs as major contributors to its prediction. Figure 1 displays instances of saliency maps generated by state-of-the-art models on original and corresponding WhatsApp-compressed CXR image pairs. In each of these images, the saliency maps highlight regions outside the lungs as significant contributors to the prediction. This effect is even more pronounced in WhatsApp-compressed images as illustrated in Figure 1. + +Localization Instability Problem(LIP): LIP refers to a situation where an object detection model exhibits localizations which are different on the transmitted image when compared with the original image. Figure 2 displays several instances of localization instability between the original image and the corresponding WhatsApp compressed image. In each of these images, the localizations predicted by the model for the original image is either absent or incorrectly placed in the WhatsApp-compressed image. + +Current methods for understanding the variations due to image perturbations are primarily visual. In this paper, we discuss and propose PIP, OLS, and LIP to quantitatively study the changes in predictions, saliency, and localizations. While several databases of CXR images have been released, most include only disease diagnosis labels or bounding boxes indicating regions of interest. However, a dataset that illustrates the aforementioned challenges is not yet publicly available. + +Contributions: Our contributions are summarized as follows: + +- We introduce CheXwhatsApp, the first paired dataset of 141,804 pairs of original and WhatsApp-compressed CXR images, which can be used for understanding the effect of WhatsApp on localization of various thoracic diseases. +- We identify that prediction of state-of-the-art (SOTA) AI models on CXR images transmitted via low-bandwidth messaging applications such as WhatsApp are affected by prediction and explainability challenges. We find that several classification models exhibit the problem of PIP, where predictions differ between the original and transmitted images. While this is a concern, but even more worrisome is many of them also exhibit OLS where predictions are driven by regions outside the lung area, raising important questions on the explainability of the models. Similarly, models that localize lung abnormalities using bounding boxes exhibit the problem of LIP. We contribute several metrics to evaluate PIP, OLS and LIP. + +- Benchmark results indicate that models trained on open-source CXR images can suffer from high PIP, OLS and LIP. However, models trained on CheXwhat-sApp, which includes both original and WhatsApp-compressed CXR images, reduces the effect of PIP, OLS and LIP by upto $80\%$ , compared to models trained on open-source CXR images. +- We open-source the dataset and evaluation pipeline for evaluation of PIP, OLS and LIP in models on CheXwhat-sApp, to streamline the development and analysis of AI-based diagnostic systems for mHealth. + +# 2. Related works + +There are several publicly available large-scale CXR datasets. The NIH dataset [38] is one of the first, containing 112,120 CXR images with labels from 30,805 unique patients. It is a multi-label dataset with 14 labels. Similar to NIH, the Chexpert dataset [18] has 224316 train and 234 validation images, both labeled with 14 categories. The Chexpert dataset was obtained from 65,240 unique patients at Stanford Hospital. The CovidNet [37] dataset contains 30,130 train and 400 testing images categorized as COVID-19, pneumonia, or normal. The MIMIC-CXR dataset [21] includes 227,835 CXR imaging studies for 65,379 patients, along with radiology reports from Beth Israel. ChestX-ray8 [39] is a CXR dataset that consists of 108,948 frontalview X-ray images of 32,717 unique patients with eight disease labels extracted through text mining. PadChest [4] contains 160,000 CXR images obtained from 67,000 patients in San Juan Hospital (Spain). BRAX [29] is a Brazilian labeled CXR dataset consisting of 24,959 CXR images annotated with fourteen labels derived from radiology reports using NLP. + +RSNA [34] and VinDr-CXR [25] are two publicly available CXR datasets with bounding box annotations. The RSNA dataset [34] consists of CXR images of patients infected with pneumonia and normal cases. The pneumonia findings are localized by bounding boxes. The VinDr CXR dataset [25] contains 15,000 training images and 3,000 testing images with critical findings localized by bounding boxes. The CXR images are annotated for the presence of 22 different lung abnormalities. CheXphoto [27] comprises a training set of natural photos and synthetic transformations of 10,507 CXR's from 3,000 unique patients that were sampled at random from the CheXpert training set. It also includes a validation set of natural and synthetic transformations of 234 CXR's from 200 patients in the CheXpert validation set. In addition to these large datasets, there are also several smaller CXR datasets with a approximately 1,000 images, annotated either programmatically or manually by radiologists. However, the literature does not contain a large dataset with original and corresponding WhatsApp-compressed CXR images. + +![](images/6a4c48b9d37231cc43ae5de841e474fdca7932541b531822e45b129ee75c4165.jpg) +Original Image + +![](images/f2428e51a506c7f911d3ce7b9de3d35d0def57fbde27ff002db54d175883c1c0.jpg) +WhatsApp Image + +![](images/ed03c9d0381af79a0ae566d39b2892547647ecc18df6544542be4368a61a571c.jpg) +Original Image + +![](images/2f76e865bb7d4cea0bbd784a23623a5d8da016d64e55e13e13a474c5e412bc3b.jpg) +WhatsApp Image + +![](images/31286531a74904127f15d365c440e679ab6fa25698f1adb4141d012c7fbdcb39.jpg) +Figure 1. Exemplified localization generated with DenseNet architecture on (a) original images and (b) their corresponding WhatsApp compressed images from CheXWhatsApp. Though the images look visually similar after compression, the predictions and localizations generated by GradCAM can significantly differ. + +![](images/4bdae44af782ead033d8d6c812f01130ad33a9023ba1c37b696db5e542003ef8.jpg) + +![](images/190c5814fc5ad204a338e9da72ec87aa144172349810ab5b10062ca979e3beff.jpg) + +![](images/1e4d8d63f6dc8e9843b01db89ca6fc316b43977b39977c94988b7e4ea31a1693.jpg) + +![](images/6a87775742df2a4cccc8403e0d9c183fd69445550a1be96b463f8bcbefdc4b04.jpg) +Original Image + +![](images/81e452e66ac3b15fb16ac310bc0943b99a39b7d2f1a9027e11f0729b7e3b77b3.jpg) +WhatsApp Image + +![](images/bebe7dfe4c9af867ab09137cb3e64c13d8e2467a7ffc458d0bab9475ba665dcd.jpg) +Original Image + +![](images/045042c4a8e93f7522b8fa60e373d4ac5039cb3047ed38db3ec019b8bbf44c60.jpg) +WhatsApp Image + +![](images/0768f44fb39ba0149b40ad45460b28aa0d934d3b28e00d4f0210d671eea8a3f9.jpg) +Figure 2. Exemplified localization generated with Faster-RCNN architecture on (a) original images and (b) their corresponding WhatsApp compressed images from CheXwhatsApp. The red boxes indicate the ground truth, and the blue boxes indicate the model's prediction. Though the images look visually similar after compression, the localizations generated by Faster-RCNN are either incorrectly placed (bottom-left), partially missing (top-right) or fully missing (top-left, bottom-right) in WhatsApp-compressed images when compared with the original images. + +![](images/09ce89bf123617e0141fdabf8d65917035348d17938c50e1a7f124b2b6f30eff.jpg) + +![](images/e57213d1a506f8c111233c3119b831e6669c00871fb7d51fe9813c9ca4f3aeab.jpg) + +![](images/ada2940d9ad213f858403adad94696bacfe8abe7a22d049deb044d6b4a31c6b1.jpg) + +# 3. Challenges in diagnosing Xray images received through Smart phones + +In deep learning based models for image classification, the prediction outcome is provided by the model and explanation of predictions is obtained through class-activation maps. In mHealth settings, as envisaged here, such systems face two challenges: prediction instability and out-of-lung saliency. Object detection models employ a combined approach to prediction and explanation by lo + +calizing the findings using bounding boxes. However, the output of object detection models can also suffer from localization instability when images are obtained through WhatsApp. This section delves deeper into these three challenges and defines the metrics used to evaluate them. + +# 3.1. The Prediction Instability Problem + +PIP of a predictive model is defined as the probability of disagreement between the predictions on a randomly perturbed instance and the true instance. + +# 3.1.1. Metrics to evaluate PIP + +Prediction Instability Score (PI Score): Consider a deep learning model $\mathsf{M}$ that classifies a CXR image into one or multiple categories. Consider a dataset $D = \{X_{i},\mathsf{R}(X_{i})\}_{i = 1}^{N}$ consisting of $N$ pairs of original and WhatsApp-compressed CXR images denoted by $X_{i}$ , and $\mathsf{R}(X_i)$ respectively. The estimate of PI Score due to image perturbation by WhatsApp can now be defined as the fraction of pairs where the predictions differed expressed as percentage. + +$$ +P I S c o r e (\mathbb {M}; D) = \frac {1}{N} \left(\sum_ {\{X _ {i}, \mathbb {R} (X _ {i}) \} \in D} P I \left(X _ {i}, \mathbb {R} \left(X _ {i}\right), \mathbb {M}\right)\right) \times 1 0 0 +$$ + +and $PI(X_i, \mathbf{R}(X_i), \mathbb{M}) = \mathbf{I}(y(X_i; \mathbb{M})) \neq \mathbf{I}(y(\mathbf{R}(X_i); \mathbb{M}))$ + +# 3.2. Out-of-Lung Saliency + +For a Deep Learning Model diagnosing lung abnormalities in CXR images, out-of-lung saliency (OLS) is defined as the correlation between attributed pixels outside the lung region and model prediction. + +# 3.2.1. Metrics to evaluate OLS + +Out of Lung Saliency Score (OLS Score): The OLS Score quantifies the impact of OLS on the diagnosis made by state-of-the-art models. It is derived using the Intersection Over Lung-Region (IOL) metric, which measures the extent of OLS on each CXR image. Let heatmap-region $(y(x; \mathbb{M}))$ indicate the set of pixels identified by a saliency based feature attribution technique such as GradCAM[33] applied to model $\mathbb{M}$ on CXR image $x$ , resulting in prediction $y$ . Let $lung(x)$ represents the pixels within image $x$ that belong to the lung region. IOL is defined as the ratio of the number of pixels where the heatmap region and the lung region overlap in a CXR image, divided by the total number of pixels in the heatmap region. Mathematically, for a given CXR image $x$ (represented by its pixels and intensity values), define + +$$ +I O L (x; \mathbb {M}) = \frac {\operatorname {P i x} (\text {h e a t m a p - r e g i o n} (y (x ; \mathbb {M})) \cap l u n g (x))}{\operatorname {P i x} (\text {h e a t m a p - r e g i o n} (y (x ; \mathbb {M})))} \tag {2} +$$ + +where $\operatorname{P}\mathrm{i}\mathrm{x}(R(x))$ counts the number of Pixels in region $R(x)$ , a subset of the pixels in the original image $x$ . The IOL value ranges from 0 to 1, where 0 indicates that prediction relies entirely on pixels outside the lung region and 1 indicates otherwise. The OLS Score of a model $\mathbb{M}$ on a dataset $D$ consisting of $N$ images is defined as the percentage of images in $D$ with an IOL value less than a threshold $\eta$ . + +$$ +O L S \operatorname {S c o r e} _ {\eta} (D; \mathbb {M}) = \frac {1}{N} \left(\sum_ {x \in D} I O L (x; \mathbb {M}) < \eta\right) \times 1 0 0 \tag {3} +$$ + +The value of $\eta$ in this paper is set to 0.4. For brevity, we will exclude $\eta$ from the subscript. A high OLS Score suggests that a majority of the data is affected by OLS, while a low score implies lesser impact. + +# 3.3. The Localization Instability Problem + +LIP of an object detection model is defined as the probability of disagreement between the localization of findings identified by the model that agree with the ground truth, on a randomly perturbed instance versus the true instance. Object detection models localize findings using bounding boxes and thus simultaneously provide both the prediction and explanation of a finding. + +# 3.3.1. Metrics to evaluate LIP + +Localization Instability Score (LI Score): Consider a model M that utilizes bounding boxes to localize findings (instances of lung abnormalities) on both the original image $x$ and its perturbed version $\mathsf{R}(x)$ . Let $\mathsf{Gt}_x$ denote the set of ground truth localizations of $x$ . Let $y_{\mathsf{f}}(x, \mathsf{M})$ be the localization of a finding $\mathsf{f}$ returned by M. $y_{\mathsf{f}}(x, \mathsf{M})$ agrees with ground truth of the finding $\mathsf{g}\mathsf{t}_{x,\mathsf{f}} \in \mathsf{G}\mathsf{t}_x$ if intersection over union (IOU) between localization returned by model $y_{\mathsf{f}}(x, \mathsf{M})$ and ground truth $\mathsf{g}\mathsf{t}_{x,\mathsf{f}}$ is greater than a threshold. ie, IOU $(y_{\mathsf{f}}(x, \mathsf{M}), \mathsf{g}\mathsf{t}_{x,\mathsf{f}}) > t$ , where $t$ is a threshold. + +Let $\mathsf{F}_{x,t}$ be the set of findings localized by M on image $x$ w.r.t the ground-truth. $\mathsf{F}_{x,t} = \{\mathsf{f}|\mathsf{IOU}(y_{\mathsf{f}}(x,\mathsf{M}),\mathsf{gt}_{x,\mathsf{f}}) > t,\forall \mathsf{f}\in \mathsf{Gt}_x\}$ Let $\mathsf{F}_{\mathbb{R}(x),t}$ be the set of findings localized by M on perturbed image $\mathbb{R}(x)$ w.r.t the ground-truth. $\mathsf{F}_{\mathbb{R}(x),t} = \{\mathsf{f}|\mathsf{IOU}(y_{\mathsf{f}}(\mathbb{R}(x),\mathsf{M}),\mathsf{gt}_{x,\mathsf{f}}) > t,\forall \mathsf{f}\in \mathsf{Gt}_x\}$ . LI Score due to image perturbation by WhatsApp is defined as the fraction of image pairs, where the localization of findings on original and WhatsApp-compressed images that agree with ground truth differ, expressed as percentage. + +$$ +L I S c o r e _ {t} (D; \mathbb {M}) = \frac {1}{N} \left(\sum_ {x \in D} \mathrm {F} _ {x, t} \neq \mathrm {F} _ {\mathrm {R} (x), t}\right) \times 1 0 0 \tag {4} +$$ + +Currently, there does not exist a comprehensive dataset to evaluate the problems of PIP, OLS and LIP associated with CXR images transmitted through WhatsApp for mHealth applications. + +# 4. CheXwhatsapp : A new dataset of WhatsApp compressed CXR images paired with original images + +In this section we review open source CXR datasets and discuss the steps involved in creation of a new CXR dataset called CheXwhatsapp that consists of both original and their WhatsApp images + +Discussion of Existing CXR Datasets: We study the most extensively used datasets - NIH [38] and RSNA[34], in + +![](images/0781c06f13d45771f93d73b191a0628a1629a27e80a628ad7881764d366962f9.jpg) +(a) Original image + +![](images/adc47ef8c69c707af9bd1f8028ba28c17cee17d7e49dffc3850bb0a3377e8470.jpg) +(b) Compressed image +Figure 3. Illustration of a sample (a) original image and its corresponding (b) compressed image from CheXwhatsApp. The original image is of dimension $1024 \times 1024$ and has image size of 668KB and the WhatsApp image is of dimension $1024 \times 1024$ and has an image size of 134KB, indicating an $80\%$ compression. + +the literature. These are publicly available anonymized datasets for research purposes [6, 28, 36, 37]. It has been duly verified that the NIH and RSNA datasets permits modification and redistribution of data with proper citation. It is important to note that the collection of these datasets adhered to established protocols and guidelines, following the ethical regulations stipulated by the respective institutions. Informed consent was obtained from the participating subjects. Further information regarding these datasets can be found in [34, 38]. An overview of these datasets, are as follows. + +- NIH: NIH [38] dataset contains 86524 CXR images in train set and 25596 CXR images in test set from 30,805 unique patients. It is a multi-label dataset where patients can have normal CXR or be associated with one or more of the 14 different pathologies: Atelectasis, Cardiomegaly, Effusion, Infiltration, Mass, Nodule, Pneumonia, Pneumothorax, Consolidation, Edema, Emphysema, Fibrosis, Pleural-Thickening, and Hernia. +- RSNA: The RSNA[34] consists of CXR images categorized as either pneumonia or normal cases, with manually annotated bounding boxes for pneumonia by expert radiologists. The dataset consists of 29,684 CXR images in training and validation set, as provided by the original authors. + +Additional details regarding the dataset can be found in the Supplementary Material. + +# 4.1. CheXwhatsapp: A new dataset of original and WhatsApp compressed images + +To train models that are robust to image perturbations due to compression caused by low-bandwidth messaging applications like WhatsApp, it is crucial to have a dataset that includes compressed images generated by these messaging applications. However, currently, no open-source datasets are available with image pairs consisting of original CXR images and their corresponding compressed versions. To + +address this gap, we created a new in-house dataset using WhatsApp and the previously discussed datasets, called CheXwhatsapp. + +CheXwhatsapp consists of 141,804 pairs of original and WhatsApp-compressed CXR images. CheXwhatsapp is an amalgamation of the two datasets: NIH [38] and RSNA [34] along with their corresponding WhatsApp-compressed images. The original and corresponding WhatsApp-compressed images in NIH [38] are known as NIH-WA. Similarly, the original and corresponding WhatsApp-compressed images in RSNA [34] are known as RSNA-WA. CheXwhatsapp is the union of NIH-WA and RSNA-WA. + +To create the image pairs, all the images in NIH [38] and RSNA [34] were sequentially sent through WhatsApp to a fixed mobile number. The creation of CheXwhatisApp is performed in a fully automated manner. The primary challenge in automating this process is to synchronize sending images to WhatsApp and receiving images from WhatsApp. This is primarily due to two reasons - 1) WhatsApp sends images out-of-order in certain scenarios and 2) the metadata of images is lost after sending, and images are renamed to some random name by WhatsApp at the recipient's end. Since we do not have control over WhatsApp API, the image name is also sent along with the image. The steps involved in creating the CheXwhatisApp are given below: + +1. Preprocess the image by adding the image name to the image itself. This is done by adding a small white border of 15pixel $\times$ image-width at the bottom of the image and placing the image name in the white border. The preprocessed images are uploaded to an Amazon S3 bucket to be sent to the WhatsApp Business API. +2. The pre-processed images are send by the WhatsApp business API one by one to a different WhatsApp account. This is established by using platforms for WhatsApp bulk-messaging. Such platforms are con + +figured using a phone number enabled with WhatsApp and allow bulk-messaging of images and text. + +3. The images received via WhatsApp are downloaded to an edge device, from where they are uploaded to a cloud service provider. +4. The white border at the bottom of the images is cropped to read the image name via an OCR technique called Tesseract. The image name read via OCR is used to rename the image. + +At the conclusion of the WhatsApp-compression experimentation, it was observed that the 141,804 WhatsApp compressed CXR images in CheXwhatsapp occupied only 18.43 GB of memory, whereas the original 141,804 CXR images occupied 54.80 GB. This indicates a compression factor of $66.36\%$ achieved by WhatsApp. However, this compression can alter the intricate intensity values of the images and potentially lead to inaccurate diagnoses if the models are trained solely on high-resolution images. Figure 3 presents an original and corresponding compressed images from NIH-WA in CheXwhWhatsApp. The existing state-of-the-art (SOTA) models predicted different diagnoses for the high-resolution (original) image compared to the low-resolution (compressed) image. Additionally, the activation maps used for localization showed significant differences, often highlighting non-lung regions. Clearly, this calls for more research in this area. + +# 5. Benchmarking using CheXwhat-sApp dataset + +The study shows that introducing the proposed WhatsApp dataset mitigates the effect of PIP, OLS and LIP of SOTA models. The NIH and NIH-WA datasets are evaluated using classification models. The RSNA and RSNA-WA datasets are evaluated using object detection models. This section begins with a discussion of the benchmarking setup for classification and object detection models. As stated previously, the PI Score and OLS Score are employed to analyze the effect of PIP and OLS on the prediction outcome of classification models. LI Score is used to analyze LIP in object detection models. + +Benchmark setup. State-of-the-art models considered in this study are DenseNet [17], ResNet [16], ResNeXt [40] and Inception [35]. These models are implemented using Pytorch. The models are trained using the Adam optimizer with an initial learning rate of 1e-4. The input image dimension are $512 \times 512$ . To fine-tune the learning process, ReduceLROOnPlateau scheduler is employed with patience 4 and a factor 0.1. The batch size for training data is 16 and the models are trained for 25 epochs. + +For object detection tasks, the Faster-RCNN [30] is trained using SGD optimizer with learning-rate 0.0055, mo + +mentum 0.9 and weight-decay 0.0005. The training batch-size is 10. The learning-rate schedular is StepLR (step-size=3, gamma = 0.1) and the model is trained for 15 epochs. YOLOV5s [19] and YOLOV8s [20] models are trained with Auto optimizer with learning-rate $\mathrm{lr}0 = 0.001$ , $\mathrm{lrf} = 0.001$ , final-lr $= lr0 \times lr f$ and momentum $= 0.937$ . The training batch-size is 16 and the models are trained for 50 epochs. The training is performed using 8 NVIDIA Tesla V100-SXM2-32GB GPUs. The training code is available2. + +Benchmarking results. In the experiments, SOTA models DenseNet [17], ResNet [16], ResNeXt [40] and Inception [35] are trained using the existing dataset- NIH and the proposed paired dataset - NIH-WA. The models trained using training images from existing and proposed datasets are evaluated for PIP and OLS using the original and WhatsApp-compressed images in the test dataset. The results are presented in Table 1. The results clearly indicate that models trained using NIH-WA reduces the effect of PIP, when compared with those trained using NIH. The PI Score of networks trained using NIH-WA is upto $80\%$ lower than those trained using NIH. + +To evaluate OLS Score of SOTA models trained using the existing dataset - NIH and the proposed paired dataset - NIH-WA, heatmaps of CXR images were generated using GradCAM [33], GradCAM++ [5] and EigenCAM [24]. For each CXR image, the lung region was identified using a U-Net [31] to evaluate whether the model's activation map falls within the lung regions or extends into non-lung regions. We use the highest activation region in the heatmap to determine out-of-lung saliency, considering activation values greater than 0.5 (on a scale of 0 to 1). The lung region extracted by U-Net from the CXR image, along with the highest activation region in the heatmap is utilized for the calculation of the IOL Score, which is then used to evaluate OLS Score. The results (Table 2) clearly show that models trained using NIH-WA reduces the effect of OLS, when compared to those trained using NIH. Moreover, the problem of OLS is further exacerbated in CXR images compressed by WhatsApp. + +The SOTA object detection networks, Faster-RCNN [30],YOLOv5s [19] and YOLOv8s [20] are trained using the existing bounding box CXR dataset - RSNA and the proposed paired dataset - RSNA-WA. These models are evaluated for LIP using the original and WhatsApp-compressed images in the test dataset. From the LI Score results evaluated with threshold $t = 0.7$ (Table 3), it is evident that the networks trained using RSNA-WA reduces the problem of LIP by upto $81\%$ when compared with those trained using RSNA. LI Score of networks trained using RSNA and RSNA-WA for different values of threshold $t$ is given in + +
ModelsPI Score
NIHNIH-WA
DenseNet15.743.31 (79%)
InceptionNet15.083.14 (79%)
ResNet-5014.603.30 (77%)
ResNeXt-5016.863.90 (77%)
+ +Table 1. PI Score for different networks trained on NIH and NIH-WA and evaluated on original and WhatsApp test images. From the table, it is evident that networks trained on NIH-WA has substantially lower prediction instability problem (PIP) compared to those trained on NIH. + +
ModelsDatsetOLS Score(GradCAM)OLS Score(GradCAM++)OLS Score(EigenCAM)
OriginalWhatsAppOriginalWhatsAppOriginalWhatsApp
DenseNetNIH18.7926.6930.7941.2546.0660.66
NIH-WA12.53 (33%)21.05 (21%)13.12 (57%)29.39 (29%)13.16 (71%)20.57 (66%)
InceptionNIH37.7547.4552.3362.2195.5896.70
NIH-WA22.10 (41%)32.64 (31%)18.04 (66%)34.26 (45%)94.55 (01%)95.31 (01%)
ResNetNIH24.6533.6542.1251.4097.5198.89
NIH-WA14.94 (39%)15.34 (54%)21.89 (48%)39.58 (23%)95.01 (02%)95.16 (03%)
ResNeXtNIH33.1242.5661.2469.3999.6799.79
NIH-WA11.81 (64%)26.26 (38%)10.51 (82%)26.58 (62%)96.28 (03%)96.48 (03%)
+ +Table 2. OLS Score score of attribution maps generated using different networks trained on NIH and NIH-WA training dataset and evaluated on original and WhatsApp images in test dataset. The results show that attribution maps generated using models trained on NIH-WA has lower out-of-lung saliency problem (OLS) compared to those trained on NIH. + +
ModelsLI Score
RSNARSNA-WA
YOLOV5s3.010.85 (72%)
YOLOV8s4.020.78 (81%)
Faster-RCNN12.4310.86 (13%)
+ +Table 3. LI Score for different networks trained on the bounding-box dataset - RSNA and RSNA-WA and evaluated using original and WhatsApp test dataset. From the table, it is evident that the networks trained on RSNA-WA reduces loalization instability problem (LIP) by upto $81\%$ compared to those trained on RSNA. + +![](images/68b12e8b9cd4c7686d8157cd51eaac14baa995e57d0cb5d0cacfd6db42726603.jpg) +(a) LI Score of YOLOv5s [19] + +![](images/73ebc14075285b97c8af41f1d9f17234dd16d9ba1ff9c249e0e1991463cad3da.jpg) +(b) LI Score of YOLOv8s [20] +Figure 4. LI Score of Faster-RCNN [30], YOLOv5s [19] and YOLOv8s [20] on RSNA and RSNA-WA. From the results, it can be seen that models trained on RSNA-WA mitigate the effect of loalization instability problem (LIP) for different values of threshold $t$ . + +![](images/738d3d69929d5fb9ce25a82604cb6bceb501d751ef7934b527cbc59e70dc9b90.jpg) +(c) LI Score of Faster-RCNN [30] + +Figure 4. The figure also shows that networks trained using RSNA-WA have substantially lower LIP for all values of $t$ . In particular, YOLO trained using RSNA-WA reduces LI score by more than $70\%$ . + +For LI score, the choice of parameter $t$ represents the threshold of overlap between predicted and true boundingboxes. If the parameter $t$ is too small, then most predicted bounding boxes will overlap with the true bounding boxes in both the original and WhatsApp images and LIP will be low. However, if $t$ is too large, then only few predicted bounding boxes will overlap with true bounding boxes in both original and WhatsApp images and again LIP will be low. Therefore, we must select a value for $t$ that balances these two scenarios. This optimal value may vary depending on the dataset. In this paper, we selected $t$ that corresponds to the highest LI Score to study the maximum localization instability. + +To understand the generalizability of model trained on CheXwhatsapp to images compressed by other compression algorithms such as ISO JPEG compression, we created test datasets by compressing the original test images using various JPEG quality factors - 20, 40, 60, 80, and 90 with OpenCV library. Our observations indicated that instability increases as the quality factor decreases. Furthermore, models trained with CheXwhatsapp demonstrated an average of $26\%$ lower prediction instability and $10\%$ lower localization instability compared to models trained on original images alone. Additional results are available in supplementary material. + +# 6. Discussion and Conclusion + +In this paper, we present a large-scale CheXwhatsApp dataset to aid the development of mHealth applications for diagnosing CXR images. The existing datasets consist of high resolution images and the models trained on these datasets are found to face high PIP, OLS and LIP. To study this, we proposed quantitative metrics called PI, OLS, and LI scores and obtained these metrics using the WhatsApp compressed CXR images. The high scores indicate evidence of the problem, which is due to the compression techniques used by mHealth applications. + +The growing body of research on using smart-phones and WhatsApp [3, 8, 22, 23, 32, 41] in the medical domain calls for large datasets of WhatsApp-compressed images. Our proposed dataset comprise 141,804 pairs of high-resolution and their WhatsApp compressed CXR images. As there are no available WhatsApp CXR datasets, this proposed dataset is a valuable resource for the research community to address the challenges posed by image compression and transmission over low-bandwidth networks. Further, this proposed dataset was not easy to create. To create this dataset, we used a WhatsApp business api which incurred charges per each compression and we needed to + +sequentially send the images from one phone and receive them on another phone, followed by proper mapping of the high resolution and their corresponding WhatsApp compressed images. The conducted experiments showed that the models trained with this dataset demonstrate much better visual and numerical results when tested on unseen images, thus highlighting its high robustness. As observed from Tables 1, 2, 3, training with proposed dataset helped reduce PI, OLS and LI scores significantly, demonstrating the importance of our proposed dataset. + +Furthermore, the OLS scores on the high resolution dataset improved by $70\%$ in some cases when trained with proposed dataset. This shows that the proposed dataset can aid in better explainability of results with fewer predictions outside the lung region. In healthcare, explainability plays a vital role in assisting physicians in making informed decisions. While this paper is the first to offer a large paired dataset of original and WhatsApp-compressed CXR images, there are limitations. One limitation is that we did not collect a new dataset; instead it is derived from existing datasets. Another limitation is that we did not use the new dataset to create or improve existing mHealth applications. + +Additionally, we also experimented with available JPEG compression algorithms and used them to train the models. It was interesting to note that when trained with the JPEG compressed CXR images, the PIP, OLS and LIP further increased. More details of this experimentation are provided in the supplementary material. This highlights the importance of the proposed dataset in revolutionizing healthcare delivery with mHealth applications, especially in resource-constrained regions, by facilitating rapid patient triaging and assisting inexperienced rural doctors in interpretation. We encourage researchers to use this dataset to advance research in CXR diagnosis for mHealth applications. + +# References + +[1] Mariamma Antony, Siva Teja Kakileti, Rachit Shah, Sabyasachi Sahoo, Chiranjib Bhattacharyya, and Geetha Manjunath. Challenges of ai driven diagnosis of chest x-rays transmitted through smart phones: a case study in Covid-19. Scientific Reports, 13(1):18102, 2023. 1 +[2] Richa Arora. The training and practice of radiology in india: current trends. Quantitative imaging in medicine and surgery, 4(6):449-450, 2014. 25525575[pmid]. 1 +[3] Vera Helen Buss, Stuart Leesong, Margo Barr, Marlien Varnfield, and Mark Harris. Primary prevention of cardiovascular disease and type 2 diabetes mellitus using mobile health technology: systematic review of the literature. Journal of medical Internet research, 22(10):e21159, 2020. 1, 8 +[4] Aurelia Bustos, Antonio Pertusa, Jose-Maria Salinas, and Maria De La Iglesia-Vaya. Padchest: A large chest x-ray image dataset with multi-label annotated reports. Medical image analysis, 66:101797, 2020. 2 +[5] Aditya Chattopadhy, Anirban Sarkar, Prantik Howlader, + +and Vineeth N Balasubramanian. Grad-cam++: Generalized gradient-based visual explanations for deep convolutional networks. In 2018 IEEE Winter Conference on Applications of Computer Vision (WACV), pages 839-847, 2018. 6 +[6] Muhammad EH Chowdhury, Tawsifur Rahman, Amith Khandakar, Rashid Mazhar, Muhammad Abdul Kadir, Zaid Bin Mahbub, Khandakar Reajul Islam, Muhammad Salman Khan, Atif Iqbal, Nasser Al Emadi, et al. Can ai help in screening viral and covid-19 pneumonia? IEEE Access, 8:132665-132676, 2020. 5 +[7] Alex J. DeGrave, Joseph D. Janizek, and Su-In Lee. Ai for radiographic Covid-19 detection selects shortcuts over signal. Nature Machine Intelligence, 3(7):610–619, 2021. 1 +[8] Olga Di Fede, Gaetano La Mantia, Mario GCA Cimino, and Giuseppina Campisi. Protection of patient data in digital oral and general health care: A scoping review with respect to the current regulations. Oral, 3(2):155-165, 2023. 1, 8 +[9] Robert Geirhos, Jörn-Henrik Jacobsen, Claudio Michaelis, Richard Zemel, Wieland Brendel, Matthias Bethge, and Felix A. Wichmann. Shortcut learning in deep neural networks. Nature Machine Intelligence, 2(11):665-673, 2020. 1 +[10] Daniele Giansanti. Teleradiology today: the quality concept and the Italian point of view. Telemedicine and e-Health, 23 (5):453–455, 2017. 1 +[11] Daniele Giansanti, Livia Castrichella, and Maria Rosaria Giovagnoli. The design of a health technology assessment system in telepathology. Telemedicine and e-Health, 14(6): 570-575, 2008. 1 +[12] Daniele Giansanti, Marco Pochini, and Maria Rosaria Giovagnoli. How tablet technology is going to change cooperative diagnosis in the cytology e-laboratory. Telemedicine and e-Health, 19(12):991-993, 2013. +[13] Daniele Giansanti, Marco Pochini, and Maria Rosaria Giovagnoli. Integration of tablet technologies in the e-laboratory of cytology: a health technology assessment. Telemedicine and e-Health, 20(10):909–915, 2014. 1 +[14] Daniele and Giansanti. Whatsapp in mhealth: an overview on the potentialities and the opportunities in medical imaging. Mhealth, 6, 2020. 1 +[15] GSMA. The state of mobile internet connectivity 2022, 2022. 1 +[16] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 770-778, 2016. 6 +[17] Gao Huang, Zhuang Liu, Laurens Van Der Maaten, and Kilian Q. Weinberger. Densely connected convolutional networks. In 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2261-2269, 2017. 6 +[18] Jeremy Irvin, Pranav Rajpurkar, Michael Ko, Yifan Yu, Silviana Ciurea-Ilicus, Chris Chute, Henrik Marklund, Behzad Haghgoo, Robyn Ball, Katie Shpanskaya, et al. Chexpert: A large chest radiograph dataset with uncertainty labels and expert comparison. In Proceedings of the AAAI conference on artificial intelligence, pages 590-597, 2019. 2 +[19] Glenn Jocher. Ultralytics yolov5, 2020. 6, 7 + +[20] Glenn Jocher, Ayush Chaurasia, and Jing Qiu. Ultralytics yolov8, 2023. 6, 7 +[21] Alistair EW Johnson, Tom J Pollard, Seth J Berkowitz, Nathaniel R Greenbaum, Matthew P Lungren, Chih-ying Deng, Roger G Mark, and Steven Horng. Mimic-cxr, a de-identified publicly available database of chest radiographs with free-text reports. Scientific data, 6(1):317, 2019. 2 +[22] Arjun Kalyanpur. WhatsApp and its role in teleradiology. Indian Journal of Radiology and Imaging, 2024. 1, 8 +[23] J Austin Lee, Benjamin W Wachira, John Kennedy, Nicholas Asselin, and Nee-Kofi Mould-Millman. Utilisation of what-sapp for emergency medical services in garissa, kenya. African Journal of Emergency Medicine, 14(1):38-44, 2024. 1, 8 +[24] Mohammed Bany Muhammad and Mohammed Yeasin. Eigen-cam: Class activation map using principal components. In 2020 international joint conference on neural networks (IJCNN), pages 1-7. IEEE, 2020. 6 +[25] Ha Q Nguyen, Khanh Lam, Linh T Le, Hieu H Pham, Dat Q Tran, Dung B Nguyen, Dung D Le, Chi M Pham, Hang TT Tong, Diep H Dinh, et al. Vindr-cxr: An open dataset of chest x-rays with radiologist's annotations. Scientific Data, 9(1):429, 2022. 2 +[26] Unathi Ntja, Jacques Janse van Rensburg, and Gina Joubert. Diagnostic accuracy and reliability of smartphone captured radiologic images communicated via Whatsapp®. African Journal of Emergency Medicine, 12(1):67-70, 2022. 1 +[27] Nick A Phillips, Pranav Rajpurkar, Mark Sabini, Rayan Krishnan, Sharon Zhou, Anuj Pareek, Nguyen Minh Phu, Chris Wang, Mudit Jain, Nguyen Duong Du, et al. Chexphoto: $10,000+$ photos and transformations of chest x-rays for benchmarking deep learning robustness. In Machine Learning for Health, pages 318-327. PMLR, 2020. 2 +[28] Sivaramakrishnan Rajaraman, Jenifer Siegelman, Philip O. Alderson, Lucas S. Folio, Les R. Folio, and Sameer K. Antani. Iteratively pruned deep learning ensembles for Covid-19 detection in chest x-rays. IEEE Access, 8:115041-115050, 2020. 5 +[29] Eduardo P Reis, Joselisa PQ De Paiva, Maria CB Da Silva, Guilherme AS Ribeiro, Victor F Paiva, Lucas Bulgarelli, Henrique MH Lee, Paulo V Santos, Vanessa M Brito, Lucas TW Amaral, et al. Brax, brazilian labeled chest x-ray dataset. Scientific Data, 9(1):487, 2022. 2 +[30] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. IEEE Transactions on Pattern Analysis and Machine Intelligence, 39(6):1137-1149, 2017. 6, 7 +[31] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention - MICCAI 2015, pages 234-241, Cham, 2015. Springer International Publishing. 6 +[32] Ravi Saini, Madhan Jeyaraman, Naveen Jeyaraman, Vijay Kumar Jain, Swaminathan Ramasubramanian, and Karthikeyan P Iyengar. Advancing orthopaedic trauma care through Whatsapp: An analysis of clinical and non-clinical applications, challenges, and future directions. World Journal of Orthopedics, 15(6):529, 2024. 1, 8 + +[33] Ramprasaath R. Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In 2017 IEEE International Conference on Computer Vision (ICCV), pages 618-626, 2017. 4, 6 +[34] George Shih, Carol C Wu, Safwan S Halabi, Marc D Kohli, Luciano M Prevedello, Tessa S Cook, Arjun Sharma, Judith K Amorosa, Veronica Arteaga, Maya Galperin-Aizenberg, et al. Augmenting the national institutes of health chest radiograph dataset with expert annotations of possible pneumonia. Radiology. Artificial intelligence, 1(1), 2019. 2, 4, 5 +[35] Christian Szegedy, Sergey Ioffe, Vincent Vanhoucke, and Alexander A Alemi. Inception-v4, inception-resnet and the impact of residual connections on learning. In Thirty-first AAAI conference on artificial intelligence, 2017. 6 +[36] Yu-Xing Tang, You-Bao Tang, Yifan Peng, Ke Yan, Mohammadhadi Bagheri, Bernadette A Redd, Catherine J Brandon, Zhiyong Lu, Mei Han, Jing Xiao, et al. Automated abnormality classification of chest radiographs using deep convolutional neural networks. NPJ digital medicine, 3(1):1-8, 2020. 5 +[37] Linda Wang, Zhong Qiu Lin, and Alexander Wong. Covidnet: a tailored deep convolutional neural network design for detection of Covid-19 cases from chest x-ray images. Scientific Reports, 10(1):19549, 2020. 2, 5 +[38] Xiaosong Wang, Yifan Peng, Le Lu, Zhiyong Lu, Mohammadhadi Bagheri, and Ronald M. Summers. ChestXray8: Hospital-scale chest x-ray database and benchmarks on weakly-supervised classification and localization of common thorax diseases. In 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2017. 2, 4, 5 +[39] Xiaosong Wang, Yifan Peng, Le Lu, Zhiyong Lu, Mohammadhadi Bagheri, and Ronald M Summers. Chestxray8: Hospital-scale chest x-ray database and benchmarks on weakly-supervised classification and localization of common thorax diseases. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2097-2106, 2017. 2 +[40] Saining Xie, Ross Girshick, Piotr Dólár, Zhuowen Tu, and Kaiming He. Aggregated residual transformations for deep neural networks. In 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 5987-5995, 2017. 6 +[41] Esmaeel Yaagoob, Regina Lee, Michelle Stubbs, Fatimah Shuaib, Raja Johar, and Sally Chan. WhatsApp-based intervention for people with type 2 diabetes: A randomized controlled trial. Nursing & Health Sciences, 26(2):e13117, 2024. 1, 8 \ No newline at end of file diff --git a/2025/CheXwhatsApp_ A Dataset for Exploring Challenges in the Diagnosis of Chest X-rays through Mobile Devices/images.zip b/2025/CheXwhatsApp_ A Dataset for Exploring Challenges in the Diagnosis of Chest X-rays through Mobile Devices/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..6f733e72a23d7326dfd1be9b68283af0c69087b5 --- /dev/null +++ b/2025/CheXwhatsApp_ A Dataset for Exploring Challenges in the Diagnosis of Chest X-rays through Mobile Devices/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:725ec22dc01251214400d3f0638edf5c1316a64c078442d05b924c8d5d18823f +size 311083 diff --git a/2025/CheXwhatsApp_ A Dataset for Exploring Challenges in the Diagnosis of Chest X-rays through Mobile Devices/layout.json b/2025/CheXwhatsApp_ A Dataset for Exploring Challenges in the Diagnosis of Chest X-rays through Mobile Devices/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..864820252d487e05f8fa686fed0b7c38e6f3ff33 --- /dev/null +++ b/2025/CheXwhatsApp_ A Dataset for Exploring Challenges in the Diagnosis of Chest X-rays through Mobile Devices/layout.json @@ -0,0 +1,8201 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 65, + 102, + 547, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 102, + 547, + 140 + ], + "spans": [ + { + "bbox": [ + 65, + 102, + 547, + 140 + ], + "type": "text", + "content": "CheXwhatsapp: A Dataset for Exploring Challenges in the Diagnosis of Chest X-rays through Mobile Devices" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 91, + 160, + 192, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 160, + 192, + 175 + ], + "spans": [ + { + "bbox": [ + 91, + 160, + 192, + 175 + ], + "type": "text", + "content": "Mariamma Antony1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 217, + 160, + 286, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 217, + 160, + 286, + 175 + ], + "spans": [ + { + "bbox": [ + 217, + 160, + 286, + 175 + ], + "type": "text", + "content": "Rajiv Porana" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 161, + 397, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 161, + 397, + 175 + ], + "spans": [ + { + "bbox": [ + 313, + 161, + 397, + 175 + ], + "type": "text", + "content": "Sahil M Lathiya1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 425, + 161, + 518, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 425, + 161, + 518, + 175 + ], + "spans": [ + { + "bbox": [ + 425, + 161, + 518, + 175 + ], + "type": "text", + "content": "Siva Teja Kakileti2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 243, + 175, + 367, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 243, + 175, + 367, + 190 + ], + "spans": [ + { + "bbox": [ + 243, + 175, + 367, + 190 + ], + "type": "text", + "content": "Chiranjib Bhattacharyya" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 88, + 189, + 520, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 189, + 520, + 204 + ], + "spans": [ + { + "bbox": [ + 88, + 189, + 520, + 204 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 88, + 189, + 520, + 204 + ], + "type": "text", + "content": "Indian Institute of Science, Bangalore, India " + }, + { + "bbox": [ + 88, + 189, + 520, + 204 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 88, + 189, + 520, + 204 + ], + "type": "text", + "content": "Niramai Health Analytix, Bangalore, India" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "spans": [ + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 54, + 255, + 297, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 255, + 297, + 509 + ], + "spans": [ + { + "bbox": [ + 54, + 255, + 297, + 509 + ], + "type": "text", + "content": "Mobile health (mHealth) has emerged as a transformative solution to enhance healthcare accessibility and affordability, particularly in resource-constrained regions and low-to-middle-income countries. mHealth leverages mobile platforms to improve healthcare accessibility, addressing radiologist shortages in low-resource settings by enabling remote diagnosis and consultation through mobile devices. Mobile phones allow healthcare workers to transmit radiographic images, such as chest X-rays (CXR), to specialists or AI-driven models for interpretation. However, AI-based diagnosis using CXR images shared via apps like WhatsApp suffers from reduced predictability and explainability due to compression artifacts, and there is a lack of datasets to systematically study these challenges. To address this, we introduce CheXwhatsapp, a dataset of 141,804 paired original and WhatsApp-compressed CXR images. We present a benchmarking study which shows the dataset improves prediction stability and explainability of state-of-the-art models by up to " + }, + { + "bbox": [ + 54, + 255, + 297, + 509 + ], + "type": "inline_equation", + "content": "80\\%" + }, + { + "bbox": [ + 54, + 255, + 297, + 509 + ], + "type": "text", + "content": ", while also enhancing localization performance. CheXwhatsapp is open-sourced to support advancements in mHealth applications for CXR analysis1." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 530, + 135, + 544 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 530, + 135, + 544 + ], + "spans": [ + { + "bbox": [ + 56, + 530, + 135, + 544 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 551, + 296, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 551, + 296, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 551, + 296, + 696 + ], + "type": "text", + "content": "Recent studies highlight the wide opportunities of messaging apps such as WhatsApp as an efficient tool for remote image sharing in mHealth applications across digital healthcare, including digital radiology [1, 10], digital pathology [11-13], cardiac imaging [14], orthopedics [32], diabetes management [41], and oral health [8]. mHealth provides a platform for health practitioners to enable disease management and speed-up diagnosis by facilitating the transfer of medical images, providing advice, facilitating drug adherence, and monitoring outbreaks. The outbreak of COVID-19 pandemic has accelerated the adoption of mHealth for digital radiology, such as for analyzing Chest X-ray (CXR)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 232, + 353, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 232, + 353, + 243 + ], + "spans": [ + { + "bbox": [ + 313, + 232, + 353, + 243 + ], + "type": "text", + "content": "imaging." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 244, + 555, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 244, + 555, + 376 + ], + "spans": [ + { + "bbox": [ + 313, + 244, + 555, + 376 + ], + "type": "text", + "content": "mHealth is especially beneficial in resource constrained areas such as Low-to-Middle-Income countries (LMICs) where the ratio of radiologists to radiographers is significantly imbalanced. For example, in India, there is only one radiologist for every 100,000 people [2]. Compounding the issue is the fact that most radiologists are located in urban areas, often far from the majority of the X-ray imaging facilities located in rural areas. Due to the high demand for chest radiography and the global shortage of radiologists, numerous applications for AI-based diagnosis of CXR images have been developed." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 377, + 556, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 377, + 556, + 520 + ], + "spans": [ + { + "bbox": [ + 313, + 377, + 556, + 520 + ], + "type": "text", + "content": "In recent years, smartphone penetration has significantly increased, and instant messaging applications such as WhatsApp have a large user-base with 535 million users in India, 118.5 million users in Brazil, and a penetration rate exceeding " + }, + { + "bbox": [ + 313, + 377, + 556, + 520 + ], + "type": "inline_equation", + "content": "97\\%" + }, + { + "bbox": [ + 313, + 377, + 556, + 520 + ], + "type": "text", + "content": " in African nations like Kenya, South Africa, and Nigeria [15]. This presents new opportunities for the quick and easy sharing of radiographic images to aid clinical decision-making [3, 22, 23, 26, 32]. In rural hospitals and diagnostic centers without on-site radiologists, the ability to transmit anonymized images instantly to a mHealth app for interpretation can be extremely valuable to facilitate rapid triaging of infected individuals [22]." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 521, + 556, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 556, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 556, + 665 + ], + "type": "text", + "content": "Leveraging common messaging platforms like WhatsApp eliminates the need for additional training since it is a widely used communication medium. However, it is important to note that these messaging platforms often compress images by upto " + }, + { + "bbox": [ + 313, + 521, + 556, + 665 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 313, + 521, + 556, + 665 + ], + "type": "text", + "content": " to overcome bandwidth limitations. In medial applications, it is crucial to obtain explanation of predictions or localizations of findings in CXR images. Existing literature [1, 7, 9] identifies challenges related to explainability, such as predictions driven by features present outside the lung region in AI-based diagnostic systems for CXR images. Specifically, these systems suffer from the following challenges." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 665, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 556, + 713 + ], + "type": "text", + "content": "Prediction Instability Problem(PIP): PIP refers to a situation where model's prediction differs between the original image and the image transmitted through WhatsApp. This undesirable effect can have critical implications in mHealth" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 702, + 284, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 702, + 284, + 713 + ], + "spans": [ + { + "bbox": [ + 66, + 702, + 284, + 713 + ], + "type": "text", + "content": "1https://github.com/cheXwhatsApp/cheXwhatsApp" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "25887" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 109, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 109, + 83 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 109, + 83 + ], + "type": "text", + "content": "applications." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 85, + 294, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 85, + 294, + 252 + ], + "spans": [ + { + "bbox": [ + 55, + 85, + 294, + 252 + ], + "type": "text", + "content": "Out of Lung Saliency(OLS): In medical diagnosis, it is crucial to provide explainable predictions. Saliency maps, which identify localized regions in an image that significantly contribute to the prediction, are commonly used to provide explanations. Out of Lung Saliency refers to the scenario where the model uses regions outside the lungs as major contributors to its prediction. Figure 1 displays instances of saliency maps generated by state-of-the-art models on original and corresponding WhatsApp-compressed CXR image pairs. In each of these images, the saliency maps highlight regions outside the lungs as significant contributors to the prediction. This effect is even more pronounced in WhatsApp-compressed images as illustrated in Figure 1." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 255, + 294, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 255, + 294, + 361 + ], + "spans": [ + { + "bbox": [ + 55, + 255, + 294, + 361 + ], + "type": "text", + "content": "Localization Instability Problem(LIP): LIP refers to a situation where an object detection model exhibits localizations which are different on the transmitted image when compared with the original image. Figure 2 displays several instances of localization instability between the original image and the corresponding WhatsApp compressed image. In each of these images, the localizations predicted by the model for the original image is either absent or incorrectly placed in the WhatsApp-compressed image." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 363, + 294, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 363, + 294, + 471 + ], + "spans": [ + { + "bbox": [ + 55, + 363, + 294, + 471 + ], + "type": "text", + "content": "Current methods for understanding the variations due to image perturbations are primarily visual. In this paper, we discuss and propose PIP, OLS, and LIP to quantitatively study the changes in predictions, saliency, and localizations. While several databases of CXR images have been released, most include only disease diagnosis labels or bounding boxes indicating regions of interest. However, a dataset that illustrates the aforementioned challenges is not yet publicly available." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 473, + 295, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 473, + 295, + 496 + ], + "spans": [ + { + "bbox": [ + 55, + 473, + 295, + 496 + ], + "type": "text", + "content": "Contributions: Our contributions are summarized as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 498, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 56, + 498, + 295, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 498, + 295, + 556 + ], + "spans": [ + { + "bbox": [ + 56, + 498, + 295, + 556 + ], + "type": "text", + "content": "- We introduce CheXwhatsApp, the first paired dataset of 141,804 pairs of original and WhatsApp-compressed CXR images, which can be used for understanding the effect of WhatsApp on localization of various thoracic diseases." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 558, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 558, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 558, + 295, + 713 + ], + "type": "text", + "content": "- We identify that prediction of state-of-the-art (SOTA) AI models on CXR images transmitted via low-bandwidth messaging applications such as WhatsApp are affected by prediction and explainability challenges. We find that several classification models exhibit the problem of PIP, where predictions differ between the original and transmitted images. While this is a concern, but even more worrisome is many of them also exhibit OLS where predictions are driven by regions outside the lung area, raising important questions on the explainability of the models. Similarly, models that localize lung abnormalities using bounding boxes exhibit the problem of LIP. We contribute several metrics to evaluate PIP, OLS and LIP." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 72, + 553, + 204 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 314, + 72, + 553, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 553, + 156 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 553, + 156 + ], + "type": "text", + "content": "- Benchmark results indicate that models trained on open-source CXR images can suffer from high PIP, OLS and LIP. However, models trained on CheXwhat-sApp, which includes both original and WhatsApp-compressed CXR images, reduces the effect of PIP, OLS and LIP by upto " + }, + { + "bbox": [ + 314, + 72, + 553, + 156 + ], + "type": "inline_equation", + "content": "80\\%" + }, + { + "bbox": [ + 314, + 72, + 553, + 156 + ], + "type": "text", + "content": ", compared to models trained on open-source CXR images." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 156, + 553, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 156, + 553, + 204 + ], + "spans": [ + { + "bbox": [ + 314, + 156, + 553, + 204 + ], + "type": "text", + "content": "- We open-source the dataset and evaluation pipeline for evaluation of PIP, OLS and LIP in models on CheXwhat-sApp, to streamline the development and analysis of AI-based diagnostic systems for mHealth." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 215, + 403, + 227 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 215, + 403, + 227 + ], + "spans": [ + { + "bbox": [ + 314, + 215, + 403, + 227 + ], + "type": "text", + "content": "2. Related works" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 235, + 553, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 235, + 553, + 472 + ], + "spans": [ + { + "bbox": [ + 313, + 235, + 553, + 472 + ], + "type": "text", + "content": "There are several publicly available large-scale CXR datasets. The NIH dataset [38] is one of the first, containing 112,120 CXR images with labels from 30,805 unique patients. It is a multi-label dataset with 14 labels. Similar to NIH, the Chexpert dataset [18] has 224316 train and 234 validation images, both labeled with 14 categories. The Chexpert dataset was obtained from 65,240 unique patients at Stanford Hospital. The CovidNet [37] dataset contains 30,130 train and 400 testing images categorized as COVID-19, pneumonia, or normal. The MIMIC-CXR dataset [21] includes 227,835 CXR imaging studies for 65,379 patients, along with radiology reports from Beth Israel. ChestX-ray8 [39] is a CXR dataset that consists of 108,948 frontalview X-ray images of 32,717 unique patients with eight disease labels extracted through text mining. PadChest [4] contains 160,000 CXR images obtained from 67,000 patients in San Juan Hospital (Spain). BRAX [29] is a Brazilian labeled CXR dataset consisting of 24,959 CXR images annotated with fourteen labels derived from radiology reports using NLP." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 474, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 474, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 474, + 553, + 713 + ], + "type": "text", + "content": "RSNA [34] and VinDr-CXR [25] are two publicly available CXR datasets with bounding box annotations. The RSNA dataset [34] consists of CXR images of patients infected with pneumonia and normal cases. The pneumonia findings are localized by bounding boxes. The VinDr CXR dataset [25] contains 15,000 training images and 3,000 testing images with critical findings localized by bounding boxes. The CXR images are annotated for the presence of 22 different lung abnormalities. CheXphoto [27] comprises a training set of natural photos and synthetic transformations of 10,507 CXR's from 3,000 unique patients that were sampled at random from the CheXpert training set. It also includes a validation set of natural and synthetic transformations of 234 CXR's from 200 patients in the CheXpert validation set. In addition to these large datasets, there are also several smaller CXR datasets with a approximately 1,000 images, annotated either programmatically or manually by radiologists. However, the literature does not contain a large dataset with original and corresponding WhatsApp-compressed CXR images." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25888" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 116, + 84, + 200, + 168 + ], + "blocks": [ + { + "bbox": [ + 129, + 71, + 186, + 82 + ], + "lines": [ + { + "bbox": [ + 129, + 71, + 186, + 82 + ], + "spans": [ + { + "bbox": [ + 129, + 71, + 186, + 82 + ], + "type": "text", + "content": "Original Image" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 116, + 84, + 200, + 168 + ], + "lines": [ + { + "bbox": [ + 116, + 84, + 200, + 168 + ], + "spans": [ + { + "bbox": [ + 116, + 84, + 200, + 168 + ], + "type": "image", + "image_path": "6a4c48b9d37231cc43ae5de841e474fdca7932541b531822e45b129ee75c4165.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 211, + 84, + 296, + 168 + ], + "blocks": [ + { + "bbox": [ + 221, + 71, + 287, + 82 + ], + "lines": [ + { + "bbox": [ + 221, + 71, + 287, + 82 + ], + "spans": [ + { + "bbox": [ + 221, + 71, + 287, + 82 + ], + "type": "text", + "content": "WhatsApp Image" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 211, + 84, + 296, + 168 + ], + "lines": [ + { + "bbox": [ + 211, + 84, + 296, + 168 + ], + "spans": [ + { + "bbox": [ + 211, + 84, + 296, + 168 + ], + "type": "image", + "image_path": "f2428e51a506c7f911d3ce7b9de3d35d0def57fbde27ff002db54d175883c1c0.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 315, + 84, + 399, + 168 + ], + "blocks": [ + { + "bbox": [ + 328, + 71, + 385, + 82 + ], + "lines": [ + { + "bbox": [ + 328, + 71, + 385, + 82 + ], + "spans": [ + { + "bbox": [ + 328, + 71, + 385, + 82 + ], + "type": "text", + "content": "Original Image" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 315, + 84, + 399, + 168 + ], + "lines": [ + { + "bbox": [ + 315, + 84, + 399, + 168 + ], + "spans": [ + { + "bbox": [ + 315, + 84, + 399, + 168 + ], + "type": "image", + "image_path": "ed03c9d0381af79a0ae566d39b2892547647ecc18df6544542be4368a61a571c.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 411, + 84, + 494, + 168 + ], + "blocks": [ + { + "bbox": [ + 421, + 72, + 485, + 83 + ], + "lines": [ + { + "bbox": [ + 421, + 72, + 485, + 83 + ], + "spans": [ + { + "bbox": [ + 421, + 72, + 485, + 83 + ], + "type": "text", + "content": "WhatsApp Image" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 411, + 84, + 494, + 168 + ], + "lines": [ + { + "bbox": [ + 411, + 84, + 494, + 168 + ], + "spans": [ + { + "bbox": [ + 411, + 84, + 494, + 168 + ], + "type": "image", + "image_path": "2f76e865bb7d4cea0bbd784a23623a5d8da016d64e55e13e13a474c5e412bc3b.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 116, + 179, + 200, + 263 + ], + "blocks": [ + { + "bbox": [ + 116, + 179, + 200, + 263 + ], + "lines": [ + { + "bbox": [ + 116, + 179, + 200, + 263 + ], + "spans": [ + { + "bbox": [ + 116, + 179, + 200, + 263 + ], + "type": "image", + "image_path": "31286531a74904127f15d365c440e679ab6fa25698f1adb4141d012c7fbdcb39.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 276, + 555, + 309 + ], + "lines": [ + { + "bbox": [ + 55, + 276, + 555, + 309 + ], + "spans": [ + { + "bbox": [ + 55, + 276, + 555, + 309 + ], + "type": "text", + "content": "Figure 1. Exemplified localization generated with DenseNet architecture on (a) original images and (b) their corresponding WhatsApp compressed images from CheXWhatsApp. Though the images look visually similar after compression, the predictions and localizations generated by GradCAM can significantly differ." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 211, + 179, + 296, + 263 + ], + "blocks": [ + { + "bbox": [ + 211, + 179, + 296, + 263 + ], + "lines": [ + { + "bbox": [ + 211, + 179, + 296, + 263 + ], + "spans": [ + { + "bbox": [ + 211, + 179, + 296, + 263 + ], + "type": "image", + "image_path": "4bdae44af782ead033d8d6c812f01130ad33a9023ba1c37b696db5e542003ef8.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 315, + 179, + 399, + 263 + ], + "blocks": [ + { + "bbox": [ + 315, + 179, + 399, + 263 + ], + "lines": [ + { + "bbox": [ + 315, + 179, + 399, + 263 + ], + "spans": [ + { + "bbox": [ + 315, + 179, + 399, + 263 + ], + "type": "image", + "image_path": "190c5814fc5ad204a338e9da72ec87aa144172349810ab5b10062ca979e3beff.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 410, + 179, + 495, + 263 + ], + "blocks": [ + { + "bbox": [ + 410, + 179, + 495, + 263 + ], + "lines": [ + { + "bbox": [ + 410, + 179, + 495, + 263 + ], + "spans": [ + { + "bbox": [ + 410, + 179, + 495, + 263 + ], + "type": "image", + "image_path": "1e4d8d63f6dc8e9843b01db89ca6fc316b43977b39977c94988b7e4ea31a1693.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 115, + 338, + 201, + 423 + ], + "blocks": [ + { + "bbox": [ + 130, + 325, + 187, + 335 + ], + "lines": [ + { + "bbox": [ + 130, + 325, + 187, + 335 + ], + "spans": [ + { + "bbox": [ + 130, + 325, + 187, + 335 + ], + "type": "text", + "content": "Original Image" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 115, + 338, + 201, + 423 + ], + "lines": [ + { + "bbox": [ + 115, + 338, + 201, + 423 + ], + "spans": [ + { + "bbox": [ + 115, + 338, + 201, + 423 + ], + "type": "image", + "image_path": "6a87775742df2a4cccc8403e0d9c183fd69445550a1be96b463f8bcbefdc4b04.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 211, + 338, + 296, + 422 + ], + "blocks": [ + { + "bbox": [ + 222, + 325, + 287, + 335 + ], + "lines": [ + { + "bbox": [ + 222, + 325, + 287, + 335 + ], + "spans": [ + { + "bbox": [ + 222, + 325, + 287, + 335 + ], + "type": "text", + "content": "WhatsApp Image" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 211, + 338, + 296, + 422 + ], + "lines": [ + { + "bbox": [ + 211, + 338, + 296, + 422 + ], + "spans": [ + { + "bbox": [ + 211, + 338, + 296, + 422 + ], + "type": "image", + "image_path": "81e452e66ac3b15fb16ac310bc0943b99a39b7d2f1a9027e11f0729b7e3b77b3.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 315, + 338, + 399, + 422 + ], + "blocks": [ + { + "bbox": [ + 328, + 325, + 384, + 335 + ], + "lines": [ + { + "bbox": [ + 328, + 325, + 384, + 335 + ], + "spans": [ + { + "bbox": [ + 328, + 325, + 384, + 335 + ], + "type": "text", + "content": "Original Image" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 315, + 338, + 399, + 422 + ], + "lines": [ + { + "bbox": [ + 315, + 338, + 399, + 422 + ], + "spans": [ + { + "bbox": [ + 315, + 338, + 399, + 422 + ], + "type": "image", + "image_path": "bebe7dfe4c9af867ab09137cb3e64c13d8e2467a7ffc458d0bab9475ba665dcd.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 410, + 338, + 495, + 423 + ], + "blocks": [ + { + "bbox": [ + 420, + 325, + 484, + 336 + ], + "lines": [ + { + "bbox": [ + 420, + 325, + 484, + 336 + ], + "spans": [ + { + "bbox": [ + 420, + 325, + 484, + 336 + ], + "type": "text", + "content": "WhatsApp Image" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 410, + 338, + 495, + 423 + ], + "lines": [ + { + "bbox": [ + 410, + 338, + 495, + 423 + ], + "spans": [ + { + "bbox": [ + 410, + 338, + 495, + 423 + ], + "type": "image", + "image_path": "045042c4a8e93f7522b8fa60e373d4ac5039cb3047ed38db3ec019b8bbf44c60.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 117, + 433, + 201, + 516 + ], + "blocks": [ + { + "bbox": [ + 117, + 433, + 201, + 516 + ], + "lines": [ + { + "bbox": [ + 117, + 433, + 201, + 516 + ], + "spans": [ + { + "bbox": [ + 117, + 433, + 201, + 516 + ], + "type": "image", + "image_path": "0768f44fb39ba0149b40ad45460b28aa0d934d3b28e00d4f0210d671eea8a3f9.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 529, + 555, + 584 + ], + "lines": [ + { + "bbox": [ + 55, + 529, + 555, + 584 + ], + "spans": [ + { + "bbox": [ + 55, + 529, + 555, + 584 + ], + "type": "text", + "content": "Figure 2. Exemplified localization generated with Faster-RCNN architecture on (a) original images and (b) their corresponding WhatsApp compressed images from CheXwhatsApp. The red boxes indicate the ground truth, and the blue boxes indicate the model's prediction. Though the images look visually similar after compression, the localizations generated by Faster-RCNN are either incorrectly placed (bottom-left), partially missing (top-right) or fully missing (top-left, bottom-right) in WhatsApp-compressed images when compared with the original images." + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 212, + 433, + 296, + 516 + ], + "blocks": [ + { + "bbox": [ + 212, + 433, + 296, + 516 + ], + "lines": [ + { + "bbox": [ + 212, + 433, + 296, + 516 + ], + "spans": [ + { + "bbox": [ + 212, + 433, + 296, + 516 + ], + "type": "image", + "image_path": "09ce89bf123617e0141fdabf8d65917035348d17938c50e1a7f124b2b6f30eff.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 315, + 433, + 399, + 517 + ], + "blocks": [ + { + "bbox": [ + 315, + 433, + 399, + 517 + ], + "lines": [ + { + "bbox": [ + 315, + 433, + 399, + 517 + ], + "spans": [ + { + "bbox": [ + 315, + 433, + 399, + 517 + ], + "type": "image", + "image_path": "e57213d1a506f8c111233c3119b831e6669c00871fb7d51fe9813c9ca4f3aeab.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 410, + 433, + 493, + 516 + ], + "blocks": [ + { + "bbox": [ + 410, + 433, + 493, + 516 + ], + "lines": [ + { + "bbox": [ + 410, + 433, + 493, + 516 + ], + "spans": [ + { + "bbox": [ + 410, + 433, + 493, + 516 + ], + "type": "image", + "image_path": "ada2940d9ad213f858403adad94696bacfe8abe7a22d049deb044d6b4a31c6b1.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "bbox": [ + 55, + 590, + 295, + 618 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 590, + 295, + 618 + ], + "spans": [ + { + "bbox": [ + 55, + 590, + 295, + 618 + ], + "type": "text", + "content": "3. Challenges in diagnosing Xray images received through Smart phones" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "content": "In deep learning based models for image classification, the prediction outcome is provided by the model and explanation of predictions is obtained through class-activation maps. In mHealth settings, as envisaged here, such systems face two challenges: prediction instability and out-of-lung saliency. Object detection models employ a combined approach to prediction and explanation by lo" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 313, + 592, + 555, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 592, + 555, + 651 + ], + "spans": [ + { + "bbox": [ + 313, + 592, + 555, + 651 + ], + "type": "text", + "content": "calizing the findings using bounding boxes. However, the output of object detection models can also suffer from localization instability when images are obtained through WhatsApp. This section delves deeper into these three challenges and defines the metrics used to evaluate them." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 313, + 658, + 500, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 658, + 500, + 671 + ], + "spans": [ + { + "bbox": [ + 313, + 658, + 500, + 671 + ], + "type": "text", + "content": "3.1. The Prediction Instability Problem" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 313, + 677, + 555, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 712 + ], + "type": "text", + "content": "PIP of a predictive model is defined as the probability of disagreement between the predictions on a randomly perturbed instance and the true instance." + } + ] + } + ], + "index": 30 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "25889" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 185, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 185, + 83 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 185, + 83 + ], + "type": "text", + "content": "3.1.1. Metrics to evaluate PIP" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 87, + 296, + 196 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 87, + 296, + 196 + ], + "spans": [ + { + "bbox": [ + 55, + 87, + 296, + 196 + ], + "type": "text", + "content": "Prediction Instability Score (PI Score): Consider a deep learning model " + }, + { + "bbox": [ + 55, + 87, + 296, + 196 + ], + "type": "inline_equation", + "content": "\\mathsf{M}" + }, + { + "bbox": [ + 55, + 87, + 296, + 196 + ], + "type": "text", + "content": " that classifies a CXR image into one or multiple categories. Consider a dataset " + }, + { + "bbox": [ + 55, + 87, + 296, + 196 + ], + "type": "inline_equation", + "content": "D = \\{X_{i},\\mathsf{R}(X_{i})\\}_{i = 1}^{N}" + }, + { + "bbox": [ + 55, + 87, + 296, + 196 + ], + "type": "text", + "content": " consisting of " + }, + { + "bbox": [ + 55, + 87, + 296, + 196 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 87, + 296, + 196 + ], + "type": "text", + "content": " pairs of original and WhatsApp-compressed CXR images denoted by " + }, + { + "bbox": [ + 55, + 87, + 296, + 196 + ], + "type": "inline_equation", + "content": "X_{i}" + }, + { + "bbox": [ + 55, + 87, + 296, + 196 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 87, + 296, + 196 + ], + "type": "inline_equation", + "content": "\\mathsf{R}(X_i)" + }, + { + "bbox": [ + 55, + 87, + 296, + 196 + ], + "type": "text", + "content": " respectively. The estimate of PI Score due to image perturbation by WhatsApp can now be defined as the fraction of pairs where the predictions differed expressed as percentage." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 200, + 316, + 247 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 200, + 316, + 247 + ], + "spans": [ + { + "bbox": [ + 55, + 200, + 316, + 247 + ], + "type": "interline_equation", + "content": "P I S c o r e (\\mathbb {M}; D) = \\frac {1}{N} \\left(\\sum_ {\\{X _ {i}, \\mathbb {R} (X _ {i}) \\} \\in D} P I \\left(X _ {i}, \\mathbb {R} \\left(X _ {i}\\right), \\mathbb {M}\\right)\\right) \\times 1 0 0", + "image_path": "1e5af04c4d2cdc5a2140419243b8b76f590867f62bcba5b640fc6557f5b82421.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 247, + 293, + 261 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 247, + 293, + 261 + ], + "spans": [ + { + "bbox": [ + 67, + 247, + 293, + 261 + ], + "type": "text", + "content": "and " + }, + { + "bbox": [ + 67, + 247, + 293, + 261 + ], + "type": "inline_equation", + "content": "PI(X_i, \\mathbf{R}(X_i), \\mathbb{M}) = \\mathbf{I}(y(X_i; \\mathbb{M})) \\neq \\mathbf{I}(y(\\mathbf{R}(X_i); \\mathbb{M}))" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 266, + 180, + 278 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 266, + 180, + 278 + ], + "spans": [ + { + "bbox": [ + 55, + 266, + 180, + 278 + ], + "type": "text", + "content": "3.2. Out-of-Lung Saliency" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 283, + 296, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 283, + 296, + 331 + ], + "spans": [ + { + "bbox": [ + 55, + 283, + 296, + 331 + ], + "type": "text", + "content": "For a Deep Learning Model diagnosing lung abnormalities in CXR images, out-of-lung saliency (OLS) is defined as the correlation between attributed pixels outside the lung region and model prediction." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 335, + 188, + 346 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 335, + 188, + 346 + ], + "spans": [ + { + "bbox": [ + 55, + 335, + 188, + 346 + ], + "type": "text", + "content": "3.2.1. Metrics to evaluate OLS" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 350, + 296, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 350, + 296, + 529 + ], + "spans": [ + { + "bbox": [ + 55, + 350, + 296, + 529 + ], + "type": "text", + "content": "Out of Lung Saliency Score (OLS Score): The OLS Score quantifies the impact of OLS on the diagnosis made by state-of-the-art models. It is derived using the Intersection Over Lung-Region (IOL) metric, which measures the extent of OLS on each CXR image. Let heatmap-region " + }, + { + "bbox": [ + 55, + 350, + 296, + 529 + ], + "type": "inline_equation", + "content": "(y(x; \\mathbb{M}))" + }, + { + "bbox": [ + 55, + 350, + 296, + 529 + ], + "type": "text", + "content": " indicate the set of pixels identified by a saliency based feature attribution technique such as GradCAM[33] applied to model " + }, + { + "bbox": [ + 55, + 350, + 296, + 529 + ], + "type": "inline_equation", + "content": "\\mathbb{M}" + }, + { + "bbox": [ + 55, + 350, + 296, + 529 + ], + "type": "text", + "content": " on CXR image " + }, + { + "bbox": [ + 55, + 350, + 296, + 529 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 350, + 296, + 529 + ], + "type": "text", + "content": ", resulting in prediction " + }, + { + "bbox": [ + 55, + 350, + 296, + 529 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 55, + 350, + 296, + 529 + ], + "type": "text", + "content": ". Let " + }, + { + "bbox": [ + 55, + 350, + 296, + 529 + ], + "type": "inline_equation", + "content": "lung(x)" + }, + { + "bbox": [ + 55, + 350, + 296, + 529 + ], + "type": "text", + "content": " represents the pixels within image " + }, + { + "bbox": [ + 55, + 350, + 296, + 529 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 350, + 296, + 529 + ], + "type": "text", + "content": " that belong to the lung region. IOL is defined as the ratio of the number of pixels where the heatmap region and the lung region overlap in a CXR image, divided by the total number of pixels in the heatmap region. Mathematically, for a given CXR image " + }, + { + "bbox": [ + 55, + 350, + 296, + 529 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 350, + 296, + 529 + ], + "type": "text", + "content": " (represented by its pixels and intensity values), define" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 534, + 296, + 570 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 296, + 570 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 296, + 570 + ], + "type": "interline_equation", + "content": "I O L (x; \\mathbb {M}) = \\frac {\\operatorname {P i x} (\\text {h e a t m a p - r e g i o n} (y (x ; \\mathbb {M})) \\cap l u n g (x))}{\\operatorname {P i x} (\\text {h e a t m a p - r e g i o n} (y (x ; \\mathbb {M})))} \\tag {2}", + "image_path": "ba61516ba8afc1c29b11f22c027fc14ad25c9cca181748db98e7785573cd46b1.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 570, + 296, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 296, + 666 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 296, + 666 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 570, + 296, + 666 + ], + "type": "inline_equation", + "content": "\\operatorname{P}\\mathrm{i}\\mathrm{x}(R(x))" + }, + { + "bbox": [ + 55, + 570, + 296, + 666 + ], + "type": "text", + "content": " counts the number of Pixels in region " + }, + { + "bbox": [ + 55, + 570, + 296, + 666 + ], + "type": "inline_equation", + "content": "R(x)" + }, + { + "bbox": [ + 55, + 570, + 296, + 666 + ], + "type": "text", + "content": ", a subset of the pixels in the original image " + }, + { + "bbox": [ + 55, + 570, + 296, + 666 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 570, + 296, + 666 + ], + "type": "text", + "content": ". The IOL value ranges from 0 to 1, where 0 indicates that prediction relies entirely on pixels outside the lung region and 1 indicates otherwise. The OLS Score of a model " + }, + { + "bbox": [ + 55, + 570, + 296, + 666 + ], + "type": "inline_equation", + "content": "\\mathbb{M}" + }, + { + "bbox": [ + 55, + 570, + 296, + 666 + ], + "type": "text", + "content": " on a dataset " + }, + { + "bbox": [ + 55, + 570, + 296, + 666 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 55, + 570, + 296, + 666 + ], + "type": "text", + "content": " consisting of " + }, + { + "bbox": [ + 55, + 570, + 296, + 666 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 570, + 296, + 666 + ], + "type": "text", + "content": " images is defined as the percentage of images in " + }, + { + "bbox": [ + 55, + 570, + 296, + 666 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 55, + 570, + 296, + 666 + ], + "type": "text", + "content": " with an IOL value less than a threshold " + }, + { + "bbox": [ + 55, + 570, + 296, + 666 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 55, + 570, + 296, + 666 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 59, + 670, + 296, + 714 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 670, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 59, + 670, + 296, + 714 + ], + "type": "interline_equation", + "content": "O L S \\operatorname {S c o r e} _ {\\eta} (D; \\mathbb {M}) = \\frac {1}{N} \\left(\\sum_ {x \\in D} I O L (x; \\mathbb {M}) < \\eta\\right) \\times 1 0 0 \\tag {3}", + "image_path": "a3c8fa79002e383a9625004757794284e1c0c5728a7840d8b2097550eb557888.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "text", + "content": "The value of " + }, + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "text", + "content": " in this paper is set to 0.4. For brevity, we will exclude " + }, + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "text", + "content": " from the subscript. A high OLS Score suggests that a majority of the data is affected by OLS, while a low score implies lesser impact." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 127, + 510, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 127, + 510, + 140 + ], + "spans": [ + { + "bbox": [ + 313, + 127, + 510, + 140 + ], + "type": "text", + "content": "3.3. The Localization Instability Problem" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 144, + 555, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 144, + 555, + 228 + ], + "spans": [ + { + "bbox": [ + 313, + 144, + 555, + 228 + ], + "type": "text", + "content": "LIP of an object detection model is defined as the probability of disagreement between the localization of findings identified by the model that agree with the ground truth, on a randomly perturbed instance versus the true instance. Object detection models localize findings using bounding boxes and thus simultaneously provide both the prediction and explanation of a finding." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 234, + 443, + 245 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 234, + 443, + 245 + ], + "spans": [ + { + "bbox": [ + 313, + 234, + 443, + 245 + ], + "type": "text", + "content": "3.3.1. Metrics to evaluate LIP" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 249, + 554, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 249, + 554, + 369 + ], + "spans": [ + { + "bbox": [ + 313, + 249, + 554, + 369 + ], + "type": "text", + "content": "Localization Instability Score (LI Score): Consider a model M that utilizes bounding boxes to localize findings (instances of lung abnormalities) on both the original image " + }, + { + "bbox": [ + 313, + 249, + 554, + 369 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 249, + 554, + 369 + ], + "type": "text", + "content": " and its perturbed version " + }, + { + "bbox": [ + 313, + 249, + 554, + 369 + ], + "type": "inline_equation", + "content": "\\mathsf{R}(x)" + }, + { + "bbox": [ + 313, + 249, + 554, + 369 + ], + "type": "text", + "content": ". Let " + }, + { + "bbox": [ + 313, + 249, + 554, + 369 + ], + "type": "inline_equation", + "content": "\\mathsf{Gt}_x" + }, + { + "bbox": [ + 313, + 249, + 554, + 369 + ], + "type": "text", + "content": " denote the set of ground truth localizations of " + }, + { + "bbox": [ + 313, + 249, + 554, + 369 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 249, + 554, + 369 + ], + "type": "text", + "content": ". Let " + }, + { + "bbox": [ + 313, + 249, + 554, + 369 + ], + "type": "inline_equation", + "content": "y_{\\mathsf{f}}(x, \\mathsf{M})" + }, + { + "bbox": [ + 313, + 249, + 554, + 369 + ], + "type": "text", + "content": " be the localization of a finding " + }, + { + "bbox": [ + 313, + 249, + 554, + 369 + ], + "type": "inline_equation", + "content": "\\mathsf{f}" + }, + { + "bbox": [ + 313, + 249, + 554, + 369 + ], + "type": "text", + "content": " returned by M. " + }, + { + "bbox": [ + 313, + 249, + 554, + 369 + ], + "type": "inline_equation", + "content": "y_{\\mathsf{f}}(x, \\mathsf{M})" + }, + { + "bbox": [ + 313, + 249, + 554, + 369 + ], + "type": "text", + "content": " agrees with ground truth of the finding " + }, + { + "bbox": [ + 313, + 249, + 554, + 369 + ], + "type": "inline_equation", + "content": "\\mathsf{g}\\mathsf{t}_{x,\\mathsf{f}} \\in \\mathsf{G}\\mathsf{t}_x" + }, + { + "bbox": [ + 313, + 249, + 554, + 369 + ], + "type": "text", + "content": " if intersection over union (IOU) between localization returned by model " + }, + { + "bbox": [ + 313, + 249, + 554, + 369 + ], + "type": "inline_equation", + "content": "y_{\\mathsf{f}}(x, \\mathsf{M})" + }, + { + "bbox": [ + 313, + 249, + 554, + 369 + ], + "type": "text", + "content": " and ground truth " + }, + { + "bbox": [ + 313, + 249, + 554, + 369 + ], + "type": "inline_equation", + "content": "\\mathsf{g}\\mathsf{t}_{x,\\mathsf{f}}" + }, + { + "bbox": [ + 313, + 249, + 554, + 369 + ], + "type": "text", + "content": " is greater than a threshold. ie, IOU " + }, + { + "bbox": [ + 313, + 249, + 554, + 369 + ], + "type": "inline_equation", + "content": "(y_{\\mathsf{f}}(x, \\mathsf{M}), \\mathsf{g}\\mathsf{t}_{x,\\mathsf{f}}) > t" + }, + { + "bbox": [ + 313, + 249, + 554, + 369 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 249, + 554, + 369 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 249, + 554, + 369 + ], + "type": "text", + "content": " is a threshold." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 369, + 555, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 369, + 555, + 477 + ], + "spans": [ + { + "bbox": [ + 313, + 369, + 555, + 477 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 313, + 369, + 555, + 477 + ], + "type": "inline_equation", + "content": "\\mathsf{F}_{x,t}" + }, + { + "bbox": [ + 313, + 369, + 555, + 477 + ], + "type": "text", + "content": " be the set of findings localized by M on image " + }, + { + "bbox": [ + 313, + 369, + 555, + 477 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 369, + 555, + 477 + ], + "type": "text", + "content": " w.r.t the ground-truth. " + }, + { + "bbox": [ + 313, + 369, + 555, + 477 + ], + "type": "inline_equation", + "content": "\\mathsf{F}_{x,t} = \\{\\mathsf{f}|\\mathsf{IOU}(y_{\\mathsf{f}}(x,\\mathsf{M}),\\mathsf{gt}_{x,\\mathsf{f}}) > t,\\forall \\mathsf{f}\\in \\mathsf{Gt}_x\\}" + }, + { + "bbox": [ + 313, + 369, + 555, + 477 + ], + "type": "text", + "content": " Let " + }, + { + "bbox": [ + 313, + 369, + 555, + 477 + ], + "type": "inline_equation", + "content": "\\mathsf{F}_{\\mathbb{R}(x),t}" + }, + { + "bbox": [ + 313, + 369, + 555, + 477 + ], + "type": "text", + "content": " be the set of findings localized by M on perturbed image " + }, + { + "bbox": [ + 313, + 369, + 555, + 477 + ], + "type": "inline_equation", + "content": "\\mathbb{R}(x)" + }, + { + "bbox": [ + 313, + 369, + 555, + 477 + ], + "type": "text", + "content": " w.r.t the ground-truth. " + }, + { + "bbox": [ + 313, + 369, + 555, + 477 + ], + "type": "inline_equation", + "content": "\\mathsf{F}_{\\mathbb{R}(x),t} = \\{\\mathsf{f}|\\mathsf{IOU}(y_{\\mathsf{f}}(\\mathbb{R}(x),\\mathsf{M}),\\mathsf{gt}_{x,\\mathsf{f}}) > t,\\forall \\mathsf{f}\\in \\mathsf{Gt}_x\\}" + }, + { + "bbox": [ + 313, + 369, + 555, + 477 + ], + "type": "text", + "content": ". LI Score due to image perturbation by WhatsApp is defined as the fraction of image pairs, where the localization of findings on original and WhatsApp-compressed images that agree with ground truth differ, expressed as percentage." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 321, + 495, + 555, + 528 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 495, + 555, + 528 + ], + "spans": [ + { + "bbox": [ + 321, + 495, + 555, + 528 + ], + "type": "interline_equation", + "content": "L I S c o r e _ {t} (D; \\mathbb {M}) = \\frac {1}{N} \\left(\\sum_ {x \\in D} \\mathrm {F} _ {x, t} \\neq \\mathrm {F} _ {\\mathrm {R} (x), t}\\right) \\times 1 0 0 \\tag {4}", + "image_path": "9a14aae4291a9146173fb9456dfb4cfd05afe7a20fd2b61605f961eb1e9b715f.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 535, + 554, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 535, + 554, + 584 + ], + "spans": [ + { + "bbox": [ + 313, + 535, + 554, + 584 + ], + "type": "text", + "content": "Currently, there does not exist a comprehensive dataset to evaluate the problems of PIP, OLS and LIP associated with CXR images transmitted through WhatsApp for mHealth applications." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 594, + 555, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 635 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 635 + ], + "type": "text", + "content": "4. CheXwhatsapp : A new dataset of WhatsApp compressed CXR images paired with original images" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 641, + 554, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 554, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 554, + 689 + ], + "type": "text", + "content": "In this section we review open source CXR datasets and discuss the steps involved in creation of a new CXR dataset called CheXwhatsapp that consists of both original and their WhatsApp images" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "content": "Discussion of Existing CXR Datasets: We study the most extensively used datasets - NIH [38] and RSNA[34], in" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "25890" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 116, + 70, + 239, + 190 + ], + "blocks": [ + { + "bbox": [ + 116, + 70, + 239, + 190 + ], + "lines": [ + { + "bbox": [ + 116, + 70, + 239, + 190 + ], + "spans": [ + { + "bbox": [ + 116, + 70, + 239, + 190 + ], + "type": "image", + "image_path": "0781c06f13d45771f93d73b191a0628a1629a27e80a628ad7881764d366962f9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 146, + 193, + 208, + 202 + ], + "lines": [ + { + "bbox": [ + 146, + 193, + 208, + 202 + ], + "spans": [ + { + "bbox": [ + 146, + 193, + 208, + 202 + ], + "type": "text", + "content": "(a) Original image" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 374, + 70, + 496, + 191 + ], + "blocks": [ + { + "bbox": [ + 374, + 70, + 496, + 191 + ], + "lines": [ + { + "bbox": [ + 374, + 70, + 496, + 191 + ], + "spans": [ + { + "bbox": [ + 374, + 70, + 496, + 191 + ], + "type": "image", + "image_path": "adc47ef8c69c707af9bd1f8028ba28c17cee17d7e49dffc3850bb0a3377e8470.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 397, + 193, + 471, + 202 + ], + "lines": [ + { + "bbox": [ + 397, + 193, + 471, + 202 + ], + "spans": [ + { + "bbox": [ + 397, + 193, + 471, + 202 + ], + "type": "text", + "content": "(b) Compressed image" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 212, + 555, + 246 + ], + "lines": [ + { + "bbox": [ + 55, + 212, + 555, + 246 + ], + "spans": [ + { + "bbox": [ + 55, + 212, + 555, + 246 + ], + "type": "text", + "content": "Figure 3. Illustration of a sample (a) original image and its corresponding (b) compressed image from CheXwhatsApp. The original image is of dimension " + }, + { + "bbox": [ + 55, + 212, + 555, + 246 + ], + "type": "inline_equation", + "content": "1024 \\times 1024" + }, + { + "bbox": [ + 55, + 212, + 555, + 246 + ], + "type": "text", + "content": " and has image size of 668KB and the WhatsApp image is of dimension " + }, + { + "bbox": [ + 55, + 212, + 555, + 246 + ], + "type": "inline_equation", + "content": "1024 \\times 1024" + }, + { + "bbox": [ + 55, + 212, + 555, + 246 + ], + "type": "text", + "content": " and has an image size of 134KB, indicating an " + }, + { + "bbox": [ + 55, + 212, + 555, + 246 + ], + "type": "inline_equation", + "content": "80\\%" + }, + { + "bbox": [ + 55, + 212, + 555, + 246 + ], + "type": "text", + "content": " compression." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 266, + 294, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 266, + 294, + 396 + ], + "spans": [ + { + "bbox": [ + 54, + 266, + 294, + 396 + ], + "type": "text", + "content": "the literature. These are publicly available anonymized datasets for research purposes [6, 28, 36, 37]. It has been duly verified that the NIH and RSNA datasets permits modification and redistribution of data with proper citation. It is important to note that the collection of these datasets adhered to established protocols and guidelines, following the ethical regulations stipulated by the respective institutions. Informed consent was obtained from the participating subjects. Further information regarding these datasets can be found in [34, 38]. An overview of these datasets, are as follows." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 400, + 295, + 567 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 55, + 400, + 295, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 400, + 295, + 495 + ], + "spans": [ + { + "bbox": [ + 55, + 400, + 295, + 495 + ], + "type": "text", + "content": "- NIH: NIH [38] dataset contains 86524 CXR images in train set and 25596 CXR images in test set from 30,805 unique patients. It is a multi-label dataset where patients can have normal CXR or be associated with one or more of the 14 different pathologies: Atelectasis, Cardiomegaly, Effusion, Infiltration, Mass, Nodule, Pneumonia, Pneumothorax, Consolidation, Edema, Emphysema, Fibrosis, Pleural-Thickening, and Hernia." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 496, + 295, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 496, + 295, + 567 + ], + "spans": [ + { + "bbox": [ + 55, + 496, + 295, + 567 + ], + "type": "text", + "content": "- RSNA: The RSNA[34] consists of CXR images categorized as either pneumonia or normal cases, with manually annotated bounding boxes for pneumonia by expert radiologists. The dataset consists of 29,684 CXR images in training and validation set, as provided by the original authors." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 64, + 568, + 295, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 568, + 295, + 592 + ], + "spans": [ + { + "bbox": [ + 64, + 568, + 295, + 592 + ], + "type": "text", + "content": "Additional details regarding the dataset can be found in the Supplementary Material." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 600, + 295, + 625 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 600, + 295, + 625 + ], + "spans": [ + { + "bbox": [ + 55, + 600, + 295, + 625 + ], + "type": "text", + "content": "4.1. CheXwhatsapp: A new dataset of original and WhatsApp compressed images" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "type": "text", + "content": "To train models that are robust to image perturbations due to compression caused by low-bandwidth messaging applications like WhatsApp, it is crucial to have a dataset that includes compressed images generated by these messaging applications. However, currently, no open-source datasets are available with image pairs consisting of original CXR images and their corresponding compressed versions. To" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 266, + 553, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 266, + 553, + 301 + ], + "spans": [ + { + "bbox": [ + 313, + 266, + 553, + 301 + ], + "type": "text", + "content": "address this gap, we created a new in-house dataset using WhatsApp and the previously discussed datasets, called CheXwhatsapp." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 303, + 554, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 303, + 554, + 422 + ], + "spans": [ + { + "bbox": [ + 313, + 303, + 554, + 422 + ], + "type": "text", + "content": "CheXwhatsapp consists of 141,804 pairs of original and WhatsApp-compressed CXR images. CheXwhatsapp is an amalgamation of the two datasets: NIH [38] and RSNA [34] along with their corresponding WhatsApp-compressed images. The original and corresponding WhatsApp-compressed images in NIH [38] are known as NIH-WA. Similarly, the original and corresponding WhatsApp-compressed images in RSNA [34] are known as RSNA-WA. CheXwhatsapp is the union of NIH-WA and RSNA-WA." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 423, + 554, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 423, + 554, + 590 + ], + "spans": [ + { + "bbox": [ + 313, + 423, + 554, + 590 + ], + "type": "text", + "content": "To create the image pairs, all the images in NIH [38] and RSNA [34] were sequentially sent through WhatsApp to a fixed mobile number. The creation of CheXwhatisApp is performed in a fully automated manner. The primary challenge in automating this process is to synchronize sending images to WhatsApp and receiving images from WhatsApp. This is primarily due to two reasons - 1) WhatsApp sends images out-of-order in certain scenarios and 2) the metadata of images is lost after sending, and images are renamed to some random name by WhatsApp at the recipient's end. Since we do not have control over WhatsApp API, the image name is also sent along with the image. The steps involved in creating the CheXwhatisApp are given below:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 594, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 315, + 594, + 553, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 594, + 553, + 665 + ], + "spans": [ + { + "bbox": [ + 315, + 594, + 553, + 665 + ], + "type": "text", + "content": "1. Preprocess the image by adding the image name to the image itself. This is done by adding a small white border of 15pixel " + }, + { + "bbox": [ + 315, + 594, + 553, + 665 + ], + "type": "inline_equation", + "content": "\\times" + }, + { + "bbox": [ + 315, + 594, + 553, + 665 + ], + "type": "text", + "content": " image-width at the bottom of the image and placing the image name in the white border. The preprocessed images are uploaded to an Amazon S3 bucket to be sent to the WhatsApp Business API." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "type": "text", + "content": "2. The pre-processed images are send by the WhatsApp business API one by one to a different WhatsApp account. This is established by using platforms for WhatsApp bulk-messaging. Such platforms are con" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25891" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 67, + 72, + 294, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 294, + 95 + ], + "type": "text", + "content": "figured using a phone number enabled with WhatsApp and allow bulk-messaging of images and text." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 96, + 295, + 180 + ], + "type": "list", + "angle": 0, + "index": 3, + "blocks": [ + { + "bbox": [ + 55, + 96, + 295, + 132 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 96, + 295, + 132 + ], + "spans": [ + { + "bbox": [ + 55, + 96, + 295, + 132 + ], + "type": "text", + "content": "3. The images received via WhatsApp are downloaded to an edge device, from where they are uploaded to a cloud service provider." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 133, + 295, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 133, + 295, + 180 + ], + "spans": [ + { + "bbox": [ + 55, + 133, + 295, + 180 + ], + "type": "text", + "content": "4. The white border at the bottom of the images is cropped to read the image name via an OCR technique called Tesseract. The image name read via OCR is used to rename the image." + } + ] + } + ], + "index": 2 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 183, + 296, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 183, + 296, + 386 + ], + "spans": [ + { + "bbox": [ + 55, + 183, + 296, + 386 + ], + "type": "text", + "content": "At the conclusion of the WhatsApp-compression experimentation, it was observed that the 141,804 WhatsApp compressed CXR images in CheXwhatsapp occupied only 18.43 GB of memory, whereas the original 141,804 CXR images occupied 54.80 GB. This indicates a compression factor of " + }, + { + "bbox": [ + 55, + 183, + 296, + 386 + ], + "type": "inline_equation", + "content": "66.36\\%" + }, + { + "bbox": [ + 55, + 183, + 296, + 386 + ], + "type": "text", + "content": " achieved by WhatsApp. However, this compression can alter the intricate intensity values of the images and potentially lead to inaccurate diagnoses if the models are trained solely on high-resolution images. Figure 3 presents an original and corresponding compressed images from NIH-WA in CheXwhWhatsApp. The existing state-of-the-art (SOTA) models predicted different diagnoses for the high-resolution (original) image compared to the low-resolution (compressed) image. Additionally, the activation maps used for localization showed significant differences, often highlighting non-lung regions. Clearly, this calls for more research in this area." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 398, + 296, + 426 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 398, + 296, + 426 + ], + "spans": [ + { + "bbox": [ + 55, + 398, + 296, + 426 + ], + "type": "text", + "content": "5. Benchmarking using CheXwhat-sApp dataset" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 433, + 295, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 433, + 295, + 563 + ], + "spans": [ + { + "bbox": [ + 55, + 433, + 295, + 563 + ], + "type": "text", + "content": "The study shows that introducing the proposed WhatsApp dataset mitigates the effect of PIP, OLS and LIP of SOTA models. The NIH and NIH-WA datasets are evaluated using classification models. The RSNA and RSNA-WA datasets are evaluated using object detection models. This section begins with a discussion of the benchmarking setup for classification and object detection models. As stated previously, the PI Score and OLS Score are employed to analyze the effect of PIP and OLS on the prediction outcome of classification models. LI Score is used to analyze LIP in object detection models." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 582, + 295, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 295, + 689 + ], + "type": "text", + "content": "Benchmark setup. State-of-the-art models considered in this study are DenseNet [17], ResNet [16], ResNeXt [40] and Inception [35]. These models are implemented using Pytorch. The models are trained using the Adam optimizer with an initial learning rate of 1e-4. The input image dimension are " + }, + { + "bbox": [ + 55, + 582, + 295, + 689 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 55, + 582, + 295, + 689 + ], + "type": "text", + "content": ". To fine-tune the learning process, ReduceLROOnPlateau scheduler is employed with patience 4 and a factor 0.1. The batch size for training data is 16 and the models are trained for 25 epochs." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "content": "For object detection tasks, the Faster-RCNN [30] is trained using SGD optimizer with learning-rate 0.0055, mo" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "type": "text", + "content": "mentum 0.9 and weight-decay 0.0005. The training batch-size is 10. The learning-rate schedular is StepLR (step-size=3, gamma = 0.1) and the model is trained for 15 epochs. YOLOV5s [19] and YOLOV8s [20] models are trained with Auto optimizer with learning-rate " + }, + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "type": "inline_equation", + "content": "\\mathrm{lr}0 = 0.001" + }, + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "type": "inline_equation", + "content": "\\mathrm{lrf} = 0.001" + }, + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "type": "text", + "content": ", final-lr " + }, + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "type": "inline_equation", + "content": "= lr0 \\times lr f" + }, + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "type": "text", + "content": " and momentum " + }, + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "type": "inline_equation", + "content": "= 0.937" + }, + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "type": "text", + "content": ". The training batch-size is 16 and the models are trained for 50 epochs. The training is performed using 8 NVIDIA Tesla V100-SXM2-32GB GPUs. The training code is available2." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 190, + 553, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 190, + 553, + 334 + ], + "spans": [ + { + "bbox": [ + 313, + 190, + 553, + 334 + ], + "type": "text", + "content": "Benchmarking results. In the experiments, SOTA models DenseNet [17], ResNet [16], ResNeXt [40] and Inception [35] are trained using the existing dataset- NIH and the proposed paired dataset - NIH-WA. The models trained using training images from existing and proposed datasets are evaluated for PIP and OLS using the original and WhatsApp-compressed images in the test dataset. The results are presented in Table 1. The results clearly indicate that models trained using NIH-WA reduces the effect of PIP, when compared with those trained using NIH. The PI Score of networks trained using NIH-WA is upto " + }, + { + "bbox": [ + 313, + 190, + 553, + 334 + ], + "type": "inline_equation", + "content": "80\\%" + }, + { + "bbox": [ + 313, + 190, + 553, + 334 + ], + "type": "text", + "content": " lower than those trained using NIH." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 335, + 554, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 335, + 554, + 550 + ], + "spans": [ + { + "bbox": [ + 313, + 335, + 554, + 550 + ], + "type": "text", + "content": "To evaluate OLS Score of SOTA models trained using the existing dataset - NIH and the proposed paired dataset - NIH-WA, heatmaps of CXR images were generated using GradCAM [33], GradCAM++ [5] and EigenCAM [24]. For each CXR image, the lung region was identified using a U-Net [31] to evaluate whether the model's activation map falls within the lung regions or extends into non-lung regions. We use the highest activation region in the heatmap to determine out-of-lung saliency, considering activation values greater than 0.5 (on a scale of 0 to 1). The lung region extracted by U-Net from the CXR image, along with the highest activation region in the heatmap is utilized for the calculation of the IOL Score, which is then used to evaluate OLS Score. The results (Table 2) clearly show that models trained using NIH-WA reduces the effect of OLS, when compared to those trained using NIH. Moreover, the problem of OLS is further exacerbated in CXR images compressed by WhatsApp." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 552, + 554, + 682 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 552, + 554, + 682 + ], + "spans": [ + { + "bbox": [ + 313, + 552, + 554, + 682 + ], + "type": "text", + "content": "The SOTA object detection networks, Faster-RCNN [30],YOLOv5s [19] and YOLOv8s [20] are trained using the existing bounding box CXR dataset - RSNA and the proposed paired dataset - RSNA-WA. These models are evaluated for LIP using the original and WhatsApp-compressed images in the test dataset. From the LI Score results evaluated with threshold " + }, + { + "bbox": [ + 313, + 552, + 554, + 682 + ], + "type": "inline_equation", + "content": "t = 0.7" + }, + { + "bbox": [ + 313, + 552, + 554, + 682 + ], + "type": "text", + "content": " (Table 3), it is evident that the networks trained using RSNA-WA reduces the problem of LIP by upto " + }, + { + "bbox": [ + 313, + 552, + 554, + 682 + ], + "type": "inline_equation", + "content": "81\\%" + }, + { + "bbox": [ + 313, + 552, + 554, + 682 + ], + "type": "text", + "content": " when compared with those trained using RSNA. LI Score of networks trained using RSNA and RSNA-WA for different values of threshold " + }, + { + "bbox": [ + 313, + 552, + 554, + 682 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 552, + 554, + 682 + ], + "type": "text", + "content": " is given in" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 313, + 693, + 553, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 693, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 693, + 553, + 712 + ], + "type": "text", + "content": "https://github.com/cheXwhatsApp/cheXwhatsApp_benchmark" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25892" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 228, + 70, + 382, + 160 + ], + "blocks": [ + { + "bbox": [ + 228, + 70, + 382, + 160 + ], + "lines": [ + { + "bbox": [ + 228, + 70, + 382, + 160 + ], + "spans": [ + { + "bbox": [ + 228, + 70, + 382, + 160 + ], + "type": "table", + "html": "
ModelsPI Score
NIHNIH-WA
DenseNet15.743.31 (79%)
InceptionNet15.083.14 (79%)
ResNet-5014.603.30 (77%)
ResNeXt-5016.863.90 (77%)
", + "image_path": "ab33f10f8fb7db5eeae65b8d8a9f4d7d583a6522f814dd76ea9ac601cfde54d0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 82, + 212, + 529, + 368 + ], + "blocks": [ + { + "bbox": [ + 55, + 168, + 555, + 201 + ], + "lines": [ + { + "bbox": [ + 55, + 168, + 555, + 201 + ], + "spans": [ + { + "bbox": [ + 55, + 168, + 555, + 201 + ], + "type": "text", + "content": "Table 1. PI Score for different networks trained on NIH and NIH-WA and evaluated on original and WhatsApp test images. From the table, it is evident that networks trained on NIH-WA has substantially lower prediction instability problem (PIP) compared to those trained on NIH." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 82, + 212, + 529, + 368 + ], + "lines": [ + { + "bbox": [ + 82, + 212, + 529, + 368 + ], + "spans": [ + { + "bbox": [ + 82, + 212, + 529, + 368 + ], + "type": "table", + "html": "
ModelsDatsetOLS Score(GradCAM)OLS Score(GradCAM++)OLS Score(EigenCAM)
OriginalWhatsAppOriginalWhatsAppOriginalWhatsApp
DenseNetNIH18.7926.6930.7941.2546.0660.66
NIH-WA12.53 (33%)21.05 (21%)13.12 (57%)29.39 (29%)13.16 (71%)20.57 (66%)
InceptionNIH37.7547.4552.3362.2195.5896.70
NIH-WA22.10 (41%)32.64 (31%)18.04 (66%)34.26 (45%)94.55 (01%)95.31 (01%)
ResNetNIH24.6533.6542.1251.4097.5198.89
NIH-WA14.94 (39%)15.34 (54%)21.89 (48%)39.58 (23%)95.01 (02%)95.16 (03%)
ResNeXtNIH33.1242.5661.2469.3999.6799.79
NIH-WA11.81 (64%)26.26 (38%)10.51 (82%)26.58 (62%)96.28 (03%)96.48 (03%)
", + "image_path": "e2ec4968aefb990666e0afc55d26c865d3bb94d6addd9f8cc9a644812fa14319.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 223, + 417, + 387, + 496 + ], + "blocks": [ + { + "bbox": [ + 55, + 373, + 555, + 407 + ], + "lines": [ + { + "bbox": [ + 55, + 373, + 555, + 407 + ], + "spans": [ + { + "bbox": [ + 55, + 373, + 555, + 407 + ], + "type": "text", + "content": "Table 2. OLS Score score of attribution maps generated using different networks trained on NIH and NIH-WA training dataset and evaluated on original and WhatsApp images in test dataset. The results show that attribution maps generated using models trained on NIH-WA has lower out-of-lung saliency problem (OLS) compared to those trained on NIH." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 223, + 417, + 387, + 496 + ], + "lines": [ + { + "bbox": [ + 223, + 417, + 387, + 496 + ], + "spans": [ + { + "bbox": [ + 223, + 417, + 387, + 496 + ], + "type": "table", + "html": "
ModelsLI Score
RSNARSNA-WA
YOLOV5s3.010.85 (72%)
YOLOV8s4.020.78 (81%)
Faster-RCNN12.4310.86 (13%)
", + "image_path": "3c61fa4a3c445e8869111474effc31b3ed1c5f2e25a9b64591126e729a218545.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 504, + 555, + 538 + ], + "lines": [ + { + "bbox": [ + 55, + 504, + 555, + 538 + ], + "spans": [ + { + "bbox": [ + 55, + 504, + 555, + 538 + ], + "type": "text", + "content": "Table 3. LI Score for different networks trained on the bounding-box dataset - RSNA and RSNA-WA and evaluated using original and WhatsApp test dataset. From the table, it is evident that the networks trained on RSNA-WA reduces loalization instability problem (LIP) by upto " + }, + { + "bbox": [ + 55, + 504, + 555, + 538 + ], + "type": "inline_equation", + "content": "81\\%" + }, + { + "bbox": [ + 55, + 504, + 555, + 538 + ], + "type": "text", + "content": " compared to those trained on RSNA." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 61, + 549, + 205, + 644 + ], + "blocks": [ + { + "bbox": [ + 61, + 549, + 205, + 644 + ], + "lines": [ + { + "bbox": [ + 61, + 549, + 205, + 644 + ], + "spans": [ + { + "bbox": [ + 61, + 549, + 205, + 644 + ], + "type": "image", + "image_path": "68b12e8b9cd4c7686d8157cd51eaac14baa995e57d0cb5d0cacfd6db42726603.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 82, + 647, + 182, + 658 + ], + "lines": [ + { + "bbox": [ + 82, + 647, + 182, + 658 + ], + "spans": [ + { + "bbox": [ + 82, + 647, + 182, + 658 + ], + "type": "text", + "content": "(a) LI Score of YOLOv5s [19]" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 213, + 549, + 354, + 644 + ], + "blocks": [ + { + "bbox": [ + 213, + 549, + 354, + 644 + ], + "lines": [ + { + "bbox": [ + 213, + 549, + 354, + 644 + ], + "spans": [ + { + "bbox": [ + 213, + 549, + 354, + 644 + ], + "type": "image", + "image_path": "73ebc14075285b97c8af41f1d9f17234dd16d9ba1ff9c249e0e1991463cad3da.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 233, + 647, + 333, + 658 + ], + "lines": [ + { + "bbox": [ + 233, + 647, + 333, + 658 + ], + "spans": [ + { + "bbox": [ + 233, + 647, + 333, + 658 + ], + "type": "text", + "content": "(b) LI Score of YOLOv8s [20]" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 668, + 555, + 691 + ], + "lines": [ + { + "bbox": [ + 55, + 668, + 555, + 691 + ], + "spans": [ + { + "bbox": [ + 55, + 668, + 555, + 691 + ], + "type": "text", + "content": "Figure 4. LI Score of Faster-RCNN [30], YOLOv5s [19] and YOLOv8s [20] on RSNA and RSNA-WA. From the results, it can be seen that models trained on RSNA-WA mitigate the effect of loalization instability problem (LIP) for different values of threshold " + }, + { + "bbox": [ + 55, + 668, + 555, + 691 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 668, + 555, + 691 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 364, + 549, + 504, + 644 + ], + "blocks": [ + { + "bbox": [ + 364, + 549, + 504, + 644 + ], + "lines": [ + { + "bbox": [ + 364, + 549, + 504, + 644 + ], + "spans": [ + { + "bbox": [ + 364, + 549, + 504, + 644 + ], + "type": "image", + "image_path": "738d3d69929d5fb9ce25a82604cb6bceb501d751ef7934b527cbc59e70dc9b90.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 379, + 647, + 489, + 658 + ], + "lines": [ + { + "bbox": [ + 379, + 647, + 489, + 658 + ], + "spans": [ + { + "bbox": [ + 379, + 647, + 489, + 658 + ], + "type": "text", + "content": "(c) LI Score of Faster-RCNN [30]" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25893" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 119 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 119 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 119 + ], + "type": "text", + "content": "Figure 4. The figure also shows that networks trained using RSNA-WA have substantially lower LIP for all values of " + }, + { + "bbox": [ + 55, + 72, + 294, + 119 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 72, + 294, + 119 + ], + "type": "text", + "content": ". In particular, YOLO trained using RSNA-WA reduces LI score by more than " + }, + { + "bbox": [ + 55, + 72, + 294, + 119 + ], + "type": "inline_equation", + "content": "70\\%" + }, + { + "bbox": [ + 55, + 72, + 294, + 119 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 121, + 295, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 121, + 295, + 275 + ], + "spans": [ + { + "bbox": [ + 54, + 121, + 295, + 275 + ], + "type": "text", + "content": "For LI score, the choice of parameter " + }, + { + "bbox": [ + 54, + 121, + 295, + 275 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 54, + 121, + 295, + 275 + ], + "type": "text", + "content": " represents the threshold of overlap between predicted and true boundingboxes. If the parameter " + }, + { + "bbox": [ + 54, + 121, + 295, + 275 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 54, + 121, + 295, + 275 + ], + "type": "text", + "content": " is too small, then most predicted bounding boxes will overlap with the true bounding boxes in both the original and WhatsApp images and LIP will be low. However, if " + }, + { + "bbox": [ + 54, + 121, + 295, + 275 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 54, + 121, + 295, + 275 + ], + "type": "text", + "content": " is too large, then only few predicted bounding boxes will overlap with true bounding boxes in both original and WhatsApp images and again LIP will be low. Therefore, we must select a value for " + }, + { + "bbox": [ + 54, + 121, + 295, + 275 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 54, + 121, + 295, + 275 + ], + "type": "text", + "content": " that balances these two scenarios. This optimal value may vary depending on the dataset. In this paper, we selected " + }, + { + "bbox": [ + 54, + 121, + 295, + 275 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 54, + 121, + 295, + 275 + ], + "type": "text", + "content": " that corresponds to the highest LI Score to study the maximum localization instability." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 275, + 295, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 275, + 295, + 418 + ], + "spans": [ + { + "bbox": [ + 55, + 275, + 295, + 418 + ], + "type": "text", + "content": "To understand the generalizability of model trained on CheXwhatsapp to images compressed by other compression algorithms such as ISO JPEG compression, we created test datasets by compressing the original test images using various JPEG quality factors - 20, 40, 60, 80, and 90 with OpenCV library. Our observations indicated that instability increases as the quality factor decreases. Furthermore, models trained with CheXwhatsapp demonstrated an average of " + }, + { + "bbox": [ + 55, + 275, + 295, + 418 + ], + "type": "inline_equation", + "content": "26\\%" + }, + { + "bbox": [ + 55, + 275, + 295, + 418 + ], + "type": "text", + "content": " lower prediction instability and " + }, + { + "bbox": [ + 55, + 275, + 295, + 418 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 55, + 275, + 295, + 418 + ], + "type": "text", + "content": " lower localization instability compared to models trained on original images alone. Additional results are available in supplementary material." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 430, + 207, + 442 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 430, + 207, + 442 + ], + "spans": [ + { + "bbox": [ + 55, + 430, + 207, + 442 + ], + "type": "text", + "content": "6. Discussion and Conclusion" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 450, + 295, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 450, + 295, + 569 + ], + "spans": [ + { + "bbox": [ + 55, + 450, + 295, + 569 + ], + "type": "text", + "content": "In this paper, we present a large-scale CheXwhatsApp dataset to aid the development of mHealth applications for diagnosing CXR images. The existing datasets consist of high resolution images and the models trained on these datasets are found to face high PIP, OLS and LIP. To study this, we proposed quantitative metrics called PI, OLS, and LI scores and obtained these metrics using the WhatsApp compressed CXR images. The high scores indicate evidence of the problem, which is due to the compression techniques used by mHealth applications." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 570, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 295, + 713 + ], + "type": "text", + "content": "The growing body of research on using smart-phones and WhatsApp [3, 8, 22, 23, 32, 41] in the medical domain calls for large datasets of WhatsApp-compressed images. Our proposed dataset comprise 141,804 pairs of high-resolution and their WhatsApp compressed CXR images. As there are no available WhatsApp CXR datasets, this proposed dataset is a valuable resource for the research community to address the challenges posed by image compression and transmission over low-bandwidth networks. Further, this proposed dataset was not easy to create. To create this dataset, we used a WhatsApp business api which incurred charges per each compression and we needed to" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 553, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 191 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 191 + ], + "type": "text", + "content": "sequentially send the images from one phone and receive them on another phone, followed by proper mapping of the high resolution and their corresponding WhatsApp compressed images. The conducted experiments showed that the models trained with this dataset demonstrate much better visual and numerical results when tested on unseen images, thus highlighting its high robustness. As observed from Tables 1, 2, 3, training with proposed dataset helped reduce PI, OLS and LI scores significantly, demonstrating the importance of our proposed dataset." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 192, + 553, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 192, + 553, + 335 + ], + "spans": [ + { + "bbox": [ + 313, + 192, + 553, + 335 + ], + "type": "text", + "content": "Furthermore, the OLS scores on the high resolution dataset improved by " + }, + { + "bbox": [ + 313, + 192, + 553, + 335 + ], + "type": "inline_equation", + "content": "70\\%" + }, + { + "bbox": [ + 313, + 192, + 553, + 335 + ], + "type": "text", + "content": " in some cases when trained with proposed dataset. This shows that the proposed dataset can aid in better explainability of results with fewer predictions outside the lung region. In healthcare, explainability plays a vital role in assisting physicians in making informed decisions. While this paper is the first to offer a large paired dataset of original and WhatsApp-compressed CXR images, there are limitations. One limitation is that we did not collect a new dataset; instead it is derived from existing datasets. Another limitation is that we did not use the new dataset to create or improve existing mHealth applications." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 336, + 553, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 336, + 553, + 479 + ], + "spans": [ + { + "bbox": [ + 313, + 336, + 553, + 479 + ], + "type": "text", + "content": "Additionally, we also experimented with available JPEG compression algorithms and used them to train the models. It was interesting to note that when trained with the JPEG compressed CXR images, the PIP, OLS and LIP further increased. More details of this experimentation are provided in the supplementary material. This highlights the importance of the proposed dataset in revolutionizing healthcare delivery with mHealth applications, especially in resource-constrained regions, by facilitating rapid patient triaging and assisting inexperienced rural doctors in interpretation. We encourage researchers to use this dataset to advance research in CXR diagnosis for mHealth applications." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 315, + 491, + 373, + 503 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 491, + 373, + 503 + ], + "spans": [ + { + "bbox": [ + 315, + 491, + 373, + 503 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 319, + 510, + 555, + 714 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 319, + 510, + 553, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 510, + 553, + 566 + ], + "spans": [ + { + "bbox": [ + 319, + 510, + 553, + 566 + ], + "type": "text", + "content": "[1] Mariamma Antony, Siva Teja Kakileti, Rachit Shah, Sabyasachi Sahoo, Chiranjib Bhattacharyya, and Geetha Manjunath. Challenges of ai driven diagnosis of chest x-rays transmitted through smart phones: a case study in Covid-19. Scientific Reports, 13(1):18102, 2023. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 319, + 567, + 555, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 567, + 555, + 600 + ], + "spans": [ + { + "bbox": [ + 319, + 567, + 555, + 600 + ], + "type": "text", + "content": "[2] Richa Arora. The training and practice of radiology in india: current trends. Quantitative imaging in medicine and surgery, 4(6):449-450, 2014. 25525575[pmid]. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 319, + 601, + 555, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 601, + 555, + 656 + ], + "spans": [ + { + "bbox": [ + 319, + 601, + 555, + 656 + ], + "type": "text", + "content": "[3] Vera Helen Buss, Stuart Leesong, Margo Barr, Marlien Varnfield, and Mark Harris. Primary prevention of cardiovascular disease and type 2 diabetes mellitus using mobile health technology: systematic review of the literature. Journal of medical Internet research, 22(10):e21159, 2020. 1, 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 319, + 657, + 553, + 700 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 657, + 553, + 700 + ], + "spans": [ + { + "bbox": [ + 319, + 657, + 553, + 700 + ], + "type": "text", + "content": "[4] Aurelia Bustos, Antonio Pertusa, Jose-Maria Salinas, and Maria De La Iglesia-Vaya. Padchest: A large chest x-ray image dataset with multi-label annotated reports. Medical image analysis, 66:101797, 2020. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 319, + 702, + 553, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 702, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 319, + 702, + 553, + 714 + ], + "type": "text", + "content": "[5] Aditya Chattopadhy, Anirban Sarkar, Prantik Howlader," + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25894" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 57, + 72, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 76, + 72, + 296, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 296, + 127 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 296, + 127 + ], + "type": "text", + "content": "and Vineeth N Balasubramanian. Grad-cam++: Generalized gradient-based visual explanations for deep convolutional networks. In 2018 IEEE Winter Conference on Applications of Computer Vision (WACV), pages 839-847, 2018. 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 129, + 296, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 129, + 296, + 194 + ], + "spans": [ + { + "bbox": [ + 61, + 129, + 296, + 194 + ], + "type": "text", + "content": "[6] Muhammad EH Chowdhury, Tawsifur Rahman, Amith Khandakar, Rashid Mazhar, Muhammad Abdul Kadir, Zaid Bin Mahbub, Khandakar Reajul Islam, Muhammad Salman Khan, Atif Iqbal, Nasser Al Emadi, et al. Can ai help in screening viral and covid-19 pneumonia? IEEE Access, 8:132665-132676, 2020. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 196, + 295, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 196, + 295, + 228 + ], + "spans": [ + { + "bbox": [ + 61, + 196, + 295, + 228 + ], + "type": "text", + "content": "[7] Alex J. DeGrave, Joseph D. Janizek, and Su-In Lee. Ai for radiographic Covid-19 detection selects shortcuts over signal. Nature Machine Intelligence, 3(7):610–619, 2021. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 230, + 295, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 230, + 295, + 274 + ], + "spans": [ + { + "bbox": [ + 62, + 230, + 295, + 274 + ], + "type": "text", + "content": "[8] Olga Di Fede, Gaetano La Mantia, Mario GCA Cimino, and Giuseppina Campisi. Protection of patient data in digital oral and general health care: A scoping review with respect to the current regulations. Oral, 3(2):155-165, 2023. 1, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 275, + 295, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 275, + 295, + 319 + ], + "spans": [ + { + "bbox": [ + 62, + 275, + 295, + 319 + ], + "type": "text", + "content": "[9] Robert Geirhos, Jörn-Henrik Jacobsen, Claudio Michaelis, Richard Zemel, Wieland Brendel, Matthias Bethge, and Felix A. Wichmann. Shortcut learning in deep neural networks. Nature Machine Intelligence, 2(11):665-673, 2020. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 57, + 320, + 294, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 320, + 294, + 352 + ], + "spans": [ + { + "bbox": [ + 57, + 320, + 294, + 352 + ], + "type": "text", + "content": "[10] Daniele Giansanti. Teleradiology today: the quality concept and the Italian point of view. Telemedicine and e-Health, 23 (5):453–455, 2017. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 57, + 354, + 295, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 354, + 295, + 397 + ], + "spans": [ + { + "bbox": [ + 57, + 354, + 295, + 397 + ], + "type": "text", + "content": "[11] Daniele Giansanti, Livia Castrichella, and Maria Rosaria Giovagnoli. The design of a health technology assessment system in telepathology. Telemedicine and e-Health, 14(6): 570-575, 2008. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 57, + 399, + 295, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 399, + 295, + 442 + ], + "spans": [ + { + "bbox": [ + 57, + 399, + 295, + 442 + ], + "type": "text", + "content": "[12] Daniele Giansanti, Marco Pochini, and Maria Rosaria Giovagnoli. How tablet technology is going to change cooperative diagnosis in the cytology e-laboratory. Telemedicine and e-Health, 19(12):991-993, 2013." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 57, + 444, + 295, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 444, + 295, + 487 + ], + "spans": [ + { + "bbox": [ + 57, + 444, + 295, + 487 + ], + "type": "text", + "content": "[13] Daniele Giansanti, Marco Pochini, and Maria Rosaria Giovagnoli. Integration of tablet technologies in the e-laboratory of cytology: a health technology assessment. Telemedicine and e-Health, 20(10):909–915, 2014. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 57, + 488, + 295, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 488, + 295, + 521 + ], + "spans": [ + { + "bbox": [ + 57, + 488, + 295, + 521 + ], + "type": "text", + "content": "[14] Daniele and Giansanti. Whatsapp in mhealth: an overview on the potentialities and the opportunities in medical imaging. Mhealth, 6, 2020. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 522, + 295, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 522, + 295, + 544 + ], + "spans": [ + { + "bbox": [ + 57, + 522, + 295, + 544 + ], + "type": "text", + "content": "[15] GSMA. The state of mobile internet connectivity 2022, 2022. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 545, + 295, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 545, + 295, + 590 + ], + "spans": [ + { + "bbox": [ + 57, + 545, + 295, + 590 + ], + "type": "text", + "content": "[16] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 770-778, 2016. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 590, + 295, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 590, + 295, + 635 + ], + "spans": [ + { + "bbox": [ + 57, + 590, + 295, + 635 + ], + "type": "text", + "content": "[17] Gao Huang, Zhuang Liu, Laurens Van Der Maaten, and Kilian Q. Weinberger. Densely connected convolutional networks. In 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2261-2269, 2017. 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 635, + 295, + 701 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 635, + 295, + 701 + ], + "spans": [ + { + "bbox": [ + 57, + 635, + 295, + 701 + ], + "type": "text", + "content": "[18] Jeremy Irvin, Pranav Rajpurkar, Michael Ko, Yifan Yu, Silviana Ciurea-Ilicus, Chris Chute, Henrik Marklund, Behzad Haghgoo, Robyn Ball, Katie Shpanskaya, et al. Chexpert: A large chest radiograph dataset with uncertainty labels and expert comparison. In Proceedings of the AAAI conference on artificial intelligence, pages 590-597, 2019. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 702, + 237, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 702, + 237, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 702, + 237, + 713 + ], + "type": "text", + "content": "[19] Glenn Jocher. Ultralytics yolov5, 2020. 6, 7" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 95 + ], + "type": "text", + "content": "[20] Glenn Jocher, Ayush Chaurasia, and Jing Qiu. Ultralytics yolov8, 2023. 6, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 96, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 96, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 317, + 96, + 553, + 150 + ], + "type": "text", + "content": "[21] Alistair EW Johnson, Tom J Pollard, Seth J Berkowitz, Nathaniel R Greenbaum, Matthew P Lungren, Chih-ying Deng, Roger G Mark, and Steven Horng. Mimic-cxr, a de-identified publicly available database of chest radiographs with free-text reports. Scientific data, 6(1):317, 2019. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 152, + 553, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 553, + 172 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 553, + 172 + ], + "type": "text", + "content": "[22] Arjun Kalyanpur. WhatsApp and its role in teleradiology. Indian Journal of Radiology and Imaging, 2024. 1, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 174, + 553, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 174, + 553, + 226 + ], + "spans": [ + { + "bbox": [ + 317, + 174, + 553, + 226 + ], + "type": "text", + "content": "[23] J Austin Lee, Benjamin W Wachira, John Kennedy, Nicholas Asselin, and Nee-Kofi Mould-Millman. Utilisation of what-sapp for emergency medical services in garissa, kenya. African Journal of Emergency Medicine, 14(1):38-44, 2024. 1, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 228, + 553, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 228, + 553, + 271 + ], + "spans": [ + { + "bbox": [ + 316, + 228, + 553, + 271 + ], + "type": "text", + "content": "[24] Mohammed Bany Muhammad and Mohammed Yeasin. Eigen-cam: Class activation map using principal components. In 2020 international joint conference on neural networks (IJCNN), pages 1-7. IEEE, 2020. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 273, + 553, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 273, + 553, + 327 + ], + "spans": [ + { + "bbox": [ + 317, + 273, + 553, + 327 + ], + "type": "text", + "content": "[25] Ha Q Nguyen, Khanh Lam, Linh T Le, Hieu H Pham, Dat Q Tran, Dung B Nguyen, Dung D Le, Chi M Pham, Hang TT Tong, Diep H Dinh, et al. Vindr-cxr: An open dataset of chest x-rays with radiologist's annotations. Scientific Data, 9(1):429, 2022. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 328, + 553, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 328, + 553, + 371 + ], + "spans": [ + { + "bbox": [ + 316, + 328, + 553, + 371 + ], + "type": "text", + "content": "[26] Unathi Ntja, Jacques Janse van Rensburg, and Gina Joubert. Diagnostic accuracy and reliability of smartphone captured radiologic images communicated via Whatsapp®. African Journal of Emergency Medicine, 12(1):67-70, 2022. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 373, + 553, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 373, + 553, + 437 + ], + "spans": [ + { + "bbox": [ + 316, + 373, + 553, + 437 + ], + "type": "text", + "content": "[27] Nick A Phillips, Pranav Rajpurkar, Mark Sabini, Rayan Krishnan, Sharon Zhou, Anuj Pareek, Nguyen Minh Phu, Chris Wang, Mudit Jain, Nguyen Duong Du, et al. Chexphoto: " + }, + { + "bbox": [ + 316, + 373, + 553, + 437 + ], + "type": "inline_equation", + "content": "10,000+" + }, + { + "bbox": [ + 316, + 373, + 553, + 437 + ], + "type": "text", + "content": " photos and transformations of chest x-rays for benchmarking deep learning robustness. In Machine Learning for Health, pages 318-327. PMLR, 2020. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 438, + 553, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 438, + 553, + 491 + ], + "spans": [ + { + "bbox": [ + 316, + 438, + 553, + 491 + ], + "type": "text", + "content": "[28] Sivaramakrishnan Rajaraman, Jenifer Siegelman, Philip O. Alderson, Lucas S. Folio, Les R. Folio, and Sameer K. Antani. Iteratively pruned deep learning ensembles for Covid-19 detection in chest x-rays. IEEE Access, 8:115041-115050, 2020. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 494, + 553, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 494, + 553, + 547 + ], + "spans": [ + { + "bbox": [ + 317, + 494, + 553, + 547 + ], + "type": "text", + "content": "[29] Eduardo P Reis, Joselisa PQ De Paiva, Maria CB Da Silva, Guilherme AS Ribeiro, Victor F Paiva, Lucas Bulgarelli, Henrique MH Lee, Paulo V Santos, Vanessa M Brito, Lucas TW Amaral, et al. Brax, brazilian labeled chest x-ray dataset. Scientific Data, 9(1):487, 2022. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 548, + 553, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 548, + 553, + 592 + ], + "spans": [ + { + "bbox": [ + 317, + 548, + 553, + 592 + ], + "type": "text", + "content": "[30] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. IEEE Transactions on Pattern Analysis and Machine Intelligence, 39(6):1137-1149, 2017. 6, 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 593, + 553, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 593, + 553, + 647 + ], + "spans": [ + { + "bbox": [ + 317, + 593, + 553, + 647 + ], + "type": "text", + "content": "[31] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention - MICCAI 2015, pages 234-241, Cham, 2015. Springer International Publishing. 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 647, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 647, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 647, + 553, + 713 + ], + "type": "text", + "content": "[32] Ravi Saini, Madhan Jeyaraman, Naveen Jeyaraman, Vijay Kumar Jain, Swaminathan Ramasubramanian, and Karthikeyan P Iyengar. Advancing orthopaedic trauma care through Whatsapp: An analysis of clinical and non-clinical applications, challenges, and future directions. World Journal of Orthopedics, 15(6):529, 2024. 1, 8" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25895" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 628 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 137 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 137 + ], + "type": "text", + "content": "[33] Ramprasaath R. Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In 2017 IEEE International Conference on Computer Vision (ICCV), pages 618-626, 2017. 4, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 140, + 294, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 140, + 294, + 215 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 294, + 215 + ], + "type": "text", + "content": "[34] George Shih, Carol C Wu, Safwan S Halabi, Marc D Kohli, Luciano M Prevedello, Tessa S Cook, Arjun Sharma, Judith K Amorosa, Veronica Arteaga, Maya Galperin-Aizenberg, et al. Augmenting the national institutes of health chest radiograph dataset with expert annotations of possible pneumonia. Radiology. Artificial intelligence, 1(1), 2019. 2, 4, 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 218, + 294, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 218, + 294, + 261 + ], + "spans": [ + { + "bbox": [ + 56, + 218, + 294, + 261 + ], + "type": "text", + "content": "[35] Christian Szegedy, Sergey Ioffe, Vincent Vanhoucke, and Alexander A Alemi. Inception-v4, inception-resnet and the impact of residual connections on learning. In Thirty-first AAAI conference on artificial intelligence, 2017. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 262, + 294, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 262, + 294, + 327 + ], + "spans": [ + { + "bbox": [ + 56, + 262, + 294, + 327 + ], + "type": "text", + "content": "[36] Yu-Xing Tang, You-Bao Tang, Yifan Peng, Ke Yan, Mohammadhadi Bagheri, Bernadette A Redd, Catherine J Brandon, Zhiyong Lu, Mei Han, Jing Xiao, et al. Automated abnormality classification of chest radiographs using deep convolutional neural networks. NPJ digital medicine, 3(1):1-8, 2020. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 330, + 294, + 372 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 330, + 294, + 372 + ], + "spans": [ + { + "bbox": [ + 56, + 330, + 294, + 372 + ], + "type": "text", + "content": "[37] Linda Wang, Zhong Qiu Lin, and Alexander Wong. Covidnet: a tailored deep convolutional neural network design for detection of Covid-19 cases from chest x-ray images. Scientific Reports, 10(1):19549, 2020. 2, 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 374, + 294, + 439 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 374, + 294, + 439 + ], + "spans": [ + { + "bbox": [ + 56, + 374, + 294, + 439 + ], + "type": "text", + "content": "[38] Xiaosong Wang, Yifan Peng, Le Lu, Zhiyong Lu, Mohammadhadi Bagheri, and Ronald M. Summers. ChestXray8: Hospital-scale chest x-ray database and benchmarks on weakly-supervised classification and localization of common thorax diseases. In 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2017. 2, 4, 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 441, + 294, + 516 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 441, + 294, + 516 + ], + "spans": [ + { + "bbox": [ + 56, + 441, + 294, + 516 + ], + "type": "text", + "content": "[39] Xiaosong Wang, Yifan Peng, Le Lu, Zhiyong Lu, Mohammadhadi Bagheri, and Ronald M Summers. Chestxray8: Hospital-scale chest x-ray database and benchmarks on weakly-supervised classification and localization of common thorax diseases. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2097-2106, 2017. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 519, + 294, + 571 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 519, + 294, + 571 + ], + "spans": [ + { + "bbox": [ + 56, + 519, + 294, + 571 + ], + "type": "text", + "content": "[40] Saining Xie, Ross Girshick, Piotr Dólár, Zhuowen Tu, and Kaiming He. Aggregated residual transformations for deep neural networks. In 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 5987-5995, 2017. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 574, + 294, + 628 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 574, + 294, + 628 + ], + "spans": [ + { + "bbox": [ + 56, + 574, + 294, + 628 + ], + "type": "text", + "content": "[41] Esmaeel Yaagoob, Regina Lee, Michelle Stubbs, Fatimah Shuaib, Raja Johar, and Sally Chan. WhatsApp-based intervention for people with type 2 diabetes: A randomized controlled trial. Nursing & Health Sciences, 26(2):e13117, 2024. 1, 8" + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "25896" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Cheb-GR_ Rethinking K-nearest Neighbor Search in Re-ranking for Person Re-identification/9e933435-e0fc-412e-b745-089375b60f23_content_list.json b/2025/Cheb-GR_ Rethinking K-nearest Neighbor Search in Re-ranking for Person Re-identification/9e933435-e0fc-412e-b745-089375b60f23_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..3b01a83baf274b3e93430cc2ba01008d25fc5c18 --- /dev/null +++ b/2025/Cheb-GR_ Rethinking K-nearest Neighbor Search in Re-ranking for Person Re-identification/9e933435-e0fc-412e-b745-089375b60f23_content_list.json @@ -0,0 +1,1524 @@ +[ + { + "type": "text", + "text": "Cheb-GR: Rethinking k-nearest neighbor search in Re-ranking for Person Re-identification", + "text_level": 1, + "bbox": [ + 125, + 128, + 872, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jinxi Yang * He Li * Bo Du Mang Ye† \nNational Engineering Research Center for Multimedia Software, School of Computer Science, Wuhan University, Wuhan, China \n{jinxiyang, lihe404, dubo, yemang}@whu.edu.cn \nhttps://github.com/Jinxi-Yang-WHU/Fast-GCR.git", + "bbox": [ + 220, + 202, + 769, + 292 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 325, + 326, + 343 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Person re-identification (ReID) is the task of matching individuals across different camera views. Existing approaches typically employ neural networks to extract discriminative features, ranking gallery images based on their similarities to probe images. While effective, these methods are often enhanced through re-ranking, a post-processing step that refines initial retrieval results without requiring additional model training. However, current re-ranking methods mostly rely on $k$ -nearest neighbor search to extract similar images that might have the same identity as the query, which is time-consuming with a high computation burden, limiting their applications in reality. We rethink the effect of the $k$ -nearest neighbor search and introduce the Chebyshev's Theorem-guided Graph Re-ranking (ChebGR) method, which adopts the adaptive neighbor search guided by Chebyshev's Theorem over the $k$ -nearest neighbor search for efficient neighbor selection. Our method leverages graph convolution operations to refine image features and achieve robust re-ranking, leading to enhanced retrieval performance. Furthermore, we provide a theoretical analysis based on Chebyshev's Inequality to elucidate the factors contributing to the strong performance of the proposed method. Our method significantly reduces the computation costs while maintaining relatively strong performance. Through extensive experiments in both general and cross-domain settings, we demonstrate the effectiveness of Cheb-GR and its potential for real-world applications.", + "bbox": [ + 88, + 357, + 485, + 767 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 794, + 223, + 810 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Person re-identification (ReID) is a computer vision task that aims to identify the same person across different camera views in a multi-camera surveillance system [5, 40]. It", + "bbox": [ + 89, + 820, + 483, + 866 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "involves matching individuals captured by different cameras based on their appearance, even when they are viewed from different angles or under varying lighting conditions [10]. Deep learning has become the predominant approach in ReID, enabling automatic feature extraction and significantly enhancing performance [17, 21, 29, 30, 32, 33, 36, 41]. However, these methods primarily rely on pairwise similarity comparisons between the query image and individual gallery images, neglecting the contextual relationships among images within the ranking list. Consequently, they are susceptible to noise in feature representations arising from variations in illumination, pose, or occlusion.", + "bbox": [ + 511, + 327, + 906, + 508 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To tackle this issue, many re-ranking methods have emerged as a prominent strategy for enhancing performance without the need for additional model training. Prevalent approaches involve utilizing k-nearest neighbor search to identify similar images, thereby extracting contextual information [25, 42]. However, this reliance on k-nearest neighbor search presents a significant bottleneck in real-world applications due to its computational complexity. Meanwhile, many works utilize Graph Convolutional Networks (GCN) to learn pairwise similarity information and relational information between images in person ReID tasks [1, 14, 27]. Notably, Zhang et al. [38] demonstrate the effectiveness of GCNs for re-ranking by refining embeddings through feature propagation. Despite their potential, these GCN-based methods still necessitate k-nearest neighbor search, which is inherently sequential and hinders parallel computation, leading to substantial time consumption.", + "bbox": [ + 511, + 516, + 908, + 773 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To improve re-rank efficiency, we propose the Chebyshev's Theorem-guided Graph Re-ranking (Cheb-GR) method that eliminates the need for the k-nearest neighbor search and can be executed in parallel. Specifically, we construct a graph where each node represents an image. Unlike GCR [38], which connects each node to its k-nearest neighbors, Cheb-GR uses feature distances as edge weights, discarding edges with extreme distances. The remaining dis", + "bbox": [ + 511, + 779, + 910, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution", + "bbox": [ + 107, + 875, + 217, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding Author: Mang Ye (yemang@whu.edu.cn)", + "bbox": [ + 109, + 887, + 418, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "19261", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tances are then converted into similarities. This approach directly constrains feature propagation to similar images, a key factor in achieving strong performance. A notable advantage of our method is that all operations are straightforward and highly parallelizable, resulting in significantly reduced computation costs and much faster execution compared to existing methods [25, 38, 42]. Additionally, ChebGR exhibits robust performance in cross-domain scenarios, further demonstrating the versatility of our method. We also provide a theoretical analysis to explain why the proposed approach performs effectively without relying on k-nearest neighbor search, which is a method commonly used by various re-ranking methods.", + "bbox": [ + 89, + 90, + 480, + 287 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To the best of our knowledge, Cheb-GR is the fastest re-ranking method currently available, delivering competitive performance. In summary, our contributions are threefold:", + "bbox": [ + 89, + 287, + 480, + 333 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose Chebyshev's Theorem-guided Graph Re-ranking (Cheb-GR) method with no need for finding k-nearest neighbors, which runs faster than all re-rank methods we know. Besides, Cheb-GR achieves stronger performance on cross-domain settings than most existing methods.", + "- A reasonable theoretical explanation of the proposed method is presented to reveal why Cheb-GR can obtain strong performance without k-nearest neighbor search.", + "- Extensive experimentation across a diverse range of datasets provides compelling evidence of our method's strong performance and low computation costs." + ], + "bbox": [ + 89, + 333, + 480, + 513 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 527, + 232, + 542 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Person Re-identification. Recent advancements in artificial intelligence have propelled person re-identification (Re-ID), with some methods even surpassing human performance. Current approaches can be broadly divided into two categories: the first focuses on designing metric learning loss functions to learn discriminative features, while the second modifies model architectures to capture contextual information, improving feature representation.", + "bbox": [ + 89, + 553, + 480, + 672 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "For model architecture modification, many works have been proposed [4, 8, 31, 33, 35]. Sun et al. [30] propose a part-based convolutional baseline (PCB), utilizing uniform partitioning to capture local image information and employing a refined part pooling (RPP) method to enhance within-part consistency. Luo et al. [20] propose BoT, a strong baseline that adds a BNNeck structure after ResNet50. BNNeck introduces a batch normalization layer to separate ID and triplet loss optimization, improving convergence and performance in person ReID. He et al. [13] introduced TransReID, the first transformer-based model for person re-identification to achieve performance comparable to convolutional neural networks. For metric learning loss design, Softmax loss or its modification [7, 9] have been widely used in person Re-ID for feature learning. Loss functions", + "bbox": [ + 89, + 674, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "utilized in metric learning, such as triplet loss [26] and contrastive loss [12], are also applied to the person Re-ID task. Some works also design novel loss functions to guide the model to learn distinguishable representations [28, 35]. Sun et al. [28] re-weight each similarity score to obtain flexible optimization and a definite convergence target. Ye et al. [35] propose Weighted Regularization Triplet loss to take full advantage of relative distance optimization between positive and negative pairs without introducing additional margin parameters.", + "bbox": [ + 511, + 90, + 903, + 242 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "All these works are committed to extracting distinguishable features of images and then calculating the similarity between the feature of a query image and the feature of each gallery image to generate a ranking list. However, these works ignore abundant contextual information within the ranking list.", + "bbox": [ + 511, + 243, + 903, + 335 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Re-ranking for Person Re-identification. Most existing methods for Re-ID concentrate on extracting distinctive image features and assessing the similarity between the features of the query image and those of the gallery images. Additionally, a significant number of researchers have directed their attention towards re-ranking methods within ReID tasks, aiming to harness gallery-to-gallery similarity [25, 34, 42] or to integrate various ReID approaches [2, 22]. While traditional re-ranking methods have primarily emphasized the design of innovative similarity metrics, a recent advancement involves employing the graph convolutional network (GCN) to update features of images directly, which is more efficient and scalable [38].", + "bbox": [ + 511, + 335, + 903, + 532 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Most re-ranking methods design novel distance metrics to consider global contextual information or directly refine the features of images to improve performance [25, 37, 38, 42]. Zhong et al. [42] propose a k-reciprocal encoding method that leverages contextual information between k-reciprocal nearest neighbors to improve performance. Zhang et al. [38] propose a feature propagation-based method, constructing a graph by connecting each image with its k-nearest neighbors and propagating features by graph convolution on the graph.", + "bbox": [ + 511, + 534, + 903, + 686 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, existing methods rely on computationally expensive k-nearest neighbor searches to identify similar images, limiting their practical applicability, particularly for large-scale datasets. So, we propose a method that circumvents the need to find k-nearest neighbors, significantly reducing computation costs.", + "bbox": [ + 511, + 686, + 903, + 777 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Graph Convolutional Network in ReID. Graph Convolutional Network (GCN) [11] utilizes a message-passing mechanism to allow nodes to exchange and aggregate information with neighbors. Through a convolutional operation, each node's representation is updated by a weighted sum of its features and those of its neighbors, computed via matrix multiplication with a learned weight matrix [15]. This iterative process allows GCNs to learn rich representations that", + "bbox": [ + 511, + 779, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "19262", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "capture both local and global information within the graph.", + "bbox": [ + 89, + 90, + 480, + 106 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Since graphs are a great data structure for describing relationships and GCNs can learn graph-structure information, many ReID methods utilize GCNs to learn pairwise similarity and relational information among image sets. Jiang et al. [14] propose a method based on a hierarchical graph to integrate local, global, and structural information. Zhang et al. [37] and GCR [38] show that GCN can be leveraged in the re-ranking area. This paper employs a simplified GCN that achieves feature propagation without a learned weight matrix. This parameter-free approach enhances efficiency by eliminating the need for training.", + "bbox": [ + 88, + 106, + 482, + 273 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Preliminary and Motivation", + "text_level": 1, + "bbox": [ + 89, + 284, + 349, + 301 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Efficiency of K-nearest Neighbor Search. Current person ReID re-ranking methods rely heavily on k-nearest neighbor search, a key operation used to select potential same-identity images. K-nearest neighbor search computes the distances between the query image and all other images in the gallery to identify the k-nearest neighbors, which is computationally expensive and presents significant challenges when parallelizing on GPUs. The time complexity of sorting N images is $\\Theta (N\\log N)$ . In order to select the k-nearest neighbors for each node, we need to perform the sorting operations $N$ times and select the top-k images each time. Thus, the total time complexity of the k-nearest neighbor search is $\\Theta (N^2\\log N)$ , which becomes the primary computational bottleneck in the re-ranking process.", + "bbox": [ + 91, + 309, + 483, + 521 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Graph based Re-ranking. Zhang et al. [38] propose Graph Convolutional-based Re-ranking (GCR). In GCR, given a set of images, a graph is constructed that encompasses both query and gallery images. Each query and gallery image corresponds to a node in the graph, with its associated feature vector serving as the initial node feature. Edges are established between nodes representing visually similar images, encoding potential identity relationships within the dataset. This constructed graph then serves as the input to a GCN, which performs iterative message-passing operations to refine the node features. The message-passing operation in a GCN layer can be written as:", + "bbox": [ + 89, + 521, + 483, + 703 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {X} ^ {(l)} = \\mathbf {D} ^ {- \\frac {1}{2}} \\mathbf {A D} ^ {- \\frac {1}{2}} \\mathbf {X} ^ {(l - 1)} \\mathbf {W}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 189, + 710, + 482, + 729 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbf{X}^{(l)}\\in \\mathbb{R}^{N\\times d}$ is the node features in the $l^{th}$ layer, $N$ denotes the number of nodes, $d$ denotes the feature dimensionality. $\\mathbf{A}\\in \\mathbb{R}^{N\\times N}$ is the adjacent matrix of the graph, $\\mathbf{D}\\in \\mathbb{R}^{N\\times N}$ is the degree matrix of the graph which is a diagonal matrix and $\\mathbf{D}_{i,i} = \\sum_{j}\\mathbf{A}_{i,j}$ . $\\mathbf{W}\\in \\mathbb{R}^{d\\times d'}$ denotes a learnable weighted matrix in the GCN layer, which can project the features vector into a certain space.", + "bbox": [ + 89, + 736, + 482, + 844 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "When the weighted matrix $\\mathbf{W}$ is an identity matrix, the graph convolution operation can be written as:", + "bbox": [ + 89, + 844, + 483, + 875 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {X} ^ {(l)} = \\mathbf {D} ^ {- \\frac {1}{2}} \\mathbf {A} \\mathbf {D} ^ {- \\frac {1}{2}} \\mathbf {X} ^ {(l - 1)}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 196, + 882, + 482, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As formalized in Eq. (2), this parameter-free operation iteratively updates each node's feature embedding by aggregating the features of its neighbors and itself. The weight assigned to each neighbor's contribution is determined by the corresponding edge weight in the graph. Ideally, Graph Convolution-based Re-ranking enhances feature representations by promoting propagation between images of the same identity. This process brings together features of matching identities while pushing apart features of distinct identities, leading to improved re-ranking performance. Therefore, the primary challenge lies in constructing a graph that maximizes connections between images of the same identity while minimizing connections between images of different identities. As illustrated in Fig. 1(b), GCR employs a k-nearest neighbor search to identify neighboring nodes. According to Eq. (2), the time complexity of a graph convolution operation in GCR [38] is equal to matrix multiplication, which is $\\Theta(N^3)$ . However, since the matrix multiplication is highly parallelized on GPUs, it runs much faster than the k-nearest neighbor search in practice. Thus the efficiency bottleneck of GCR is still mainly affected by k-nearest neighbor search.", + "bbox": [ + 511, + 90, + 906, + 424 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Proposed Method", + "text_level": 1, + "bbox": [ + 511, + 436, + 687, + 455 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we first revisit the neighbor search and introduce an efficient strategy to replace the k-nearest neighbor search in Sec. 4.1. After that, we introduce our method in detail in Sec. 4.2.", + "bbox": [ + 511, + 463, + 906, + 523 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1. Revisiting Neighbor Search", + "text_level": 1, + "bbox": [ + 511, + 534, + 761, + 550 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Since the role of the k-nearest neighbor search in re-ranking is to select potential positive samples, we need to find an alternative strategy to find potential positive samples with low time complexity to lower the computation costs. Since feature distance is inversely related to the probability of shared identity, a threshold can be established based on the distance between the features of an anchor image and other images. Images with a distance below this threshold are then considered positive samples, sharing the same identity as the anchor image.", + "bbox": [ + 511, + 556, + 905, + 707 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The determination of an appropriate threshold for feature distance warrants further consideration. The triplet loss can be written as:", + "bbox": [ + 511, + 708, + 905, + 752 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\max (d (a, p) - d (a, n) + \\alpha , 0), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 766, + 906, + 782 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\alpha$ is the margin hyperparameter. Eq. (3) implies that, ideally, the distance between images of different identities should exceed the distance between images of the same identity by at least a factor of $\\alpha$ . Given the significant disparity in the number of positive and negative samples within the datasets, with far fewer positive samples, we can simplify our analysis by focusing on the distribution of the", + "bbox": [ + 511, + 794, + 905, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "19263", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/00a74549813f566b92ceb391b299b4ff9ec46b8ee8738b2fc5efcfdff8395385.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 133, + 89, + 517, + 262 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/2ea39a22ba03492d6e32fb733ff801034e95d0b8b624e30c5030c923e1648423.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 522, + 90, + 862, + 262 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/4a031eff605da7e3a3d47c8586770dc32d8db3f0f5430c49f8642f7881e95c44.jpg", + "image_caption": [ + "Figure 1. Comparison between traditional re-rank methods. GCR [38] and the proposed Chebyshev's Guided Graph Re-ranking (Cheb-GR). (a) Most traditional re-rank methods, such as k-reciprocal [42] and ECN [25], operate in three stages: neighborhood set construction, dissimilarity measurement between these sets, and weighted distance aggregation combining direct and neighborhood-based distances. (b) GCR connects each node to its k-nearest neighbors to construct the graph. GCNs are then employed to propagate the node features through the graph, refining the image representations for improved re-ranking. (c) Cheb-GR leverages an adaptive neighbor search instead of a k-nearest neighbor search to select neighbors and construct the graph. Subsequently, feature propagation is achieved through the application of a GCN." + ], + "image_footnote": [], + "bbox": [ + 133, + 263, + 624, + 435 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/4833588ef2c759bfe2eae85ff9ce94f713f2a95190371833db9e1c497c45dc8b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 635, + 263, + 862, + 435 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "distance between negative samples in the embedding space. Thus, the problem is reframed as one of identifying the minimum distance separating the negative samples. According to Chebyshev's Theorem [3]:", + "bbox": [ + 89, + 571, + 482, + 632 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nP r [ | R | \\geq x ] \\leq \\frac {E x [ | R | ^ {\\alpha} ]}{x ^ {\\alpha}},\n$$\n", + "text_format": "latex", + "bbox": [ + 132, + 645, + 315, + 676 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nP r [ | R - E x [ R ] | \\geq x ] \\leq \\frac {E x [ | R - E x [ R ] | ^ {\\alpha} ]}{x ^ {\\alpha}}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 135, + 678, + 482, + 709 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nP r [ | R - E x [ R ] | \\geq x ] \\leq \\frac {V a r [ R ]}{x ^ {2}},\n$$\n", + "text_format": "latex", + "bbox": [ + 135, + 710, + 364, + 741 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $R$ denotes a random variable, $Ex[R]$ denotes the expected value of the random variable $R$ , $Var[R]$ denotes the variance of the random variable $R$ , $x$ denotes a real number. We can replace $x$ with $\\lambda \\sigma$ , where $\\lambda$ is a constant hyperparameter and $\\sigma$ is standard deviation, to get the inequality:", + "bbox": [ + 89, + 753, + 482, + 830 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nP r [ | R - E x [ R ] | \\geq \\lambda \\sigma ] \\leq \\frac {1}{\\lambda^ {2}}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 181, + 843, + 482, + 873 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Then, removing the absolute value operation and ignoring", + "bbox": [ + 89, + 885, + 482, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the scenario that $R \\geq Ex[R] + \\lambda \\sigma$ , we get the inequality:", + "bbox": [ + 511, + 571, + 895, + 587 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nP r [ R \\leq E x [ R ] - \\lambda \\sigma ] \\leq \\frac {1}{\\lambda^ {2}}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 609, + 598, + 906, + 628 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "which shows that only a small number of elements fall below the mean minus $\\lambda$ times the standard deviation when $\\lambda$ is sufficiently large. Thus, given an image $i$ , we can set threshold $\\mathcal{T}_i$ for image $i$ as follows:", + "bbox": [ + 511, + 636, + 906, + 696 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {T} _ {i} = \\mu_ {i} - \\lambda \\sigma_ {i},\n$$\n", + "text_format": "latex", + "bbox": [ + 606, + 708, + 712, + 724 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mu_ {i} = \\frac {1}{N} \\sum_ {j = 1} ^ {N} D i s _ {i, j}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 607, + 727, + 905, + 771 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\sigma_ {i} = \\sqrt {\\frac {\\sum_ {j = 1} ^ {N} (D i s _ {i , j} - \\mu_ {i}) ^ {2}}{N}},\n$$\n", + "text_format": "latex", + "bbox": [ + 607, + 773, + 812, + 815 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $Dis_{i,j}$ denotes the distance between features of image $i$ and image $j$ , $N$ denotes the number of images in the dataset, $\\lambda$ is a hyper-parameter, $\\mu_i$ is the mean of distances between features of an image in the dataset and image $i$ , $\\sigma_i$ is the standard deviations of distances between features of", + "bbox": [ + 511, + 825, + 906, + 898 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "19264", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "an image in the dataset and image $i$ . Given the presence of a margin $\\alpha$ between the distances of positive and negative samples, there exists a specific value of $\\lambda$ that ensures that the threshold $\\mathcal{T}_i$ surpasses the distances between positive samples while remaining below the distances between negative samples. Therefore, by carefully selecting an appropriate value for $\\lambda$ , we can effectively filter out negative samples while preserving positive samples within the graph construction process. The time complexity of calculating the mean and standard deviation of $N$ elements is $\\Theta(N)$ , so the time complexity of calculating the threshold for $N$ images is $\\Theta(N^2)$ , which is less than the time complexity of the k-nearest neighbor search. Moreover, our proposed method can be easily parallelized on GPUs and implemented on the PyTorch platform [23], so the computation cost of our method is much less than the computation costs of the k-nearest neighbor search in practice.", + "bbox": [ + 89, + 90, + 483, + 349 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Chebyshev's Guided Graph Re-ranking", + "text_level": 1, + "bbox": [ + 89, + 354, + 433, + 372 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The pipeline of our method is shown in Fig. 1(c). We first construct a directed complete graph, and the weight of each edge is the distance between the nodes the edge connects. The adjacent matrix of this graph can be computed by:", + "bbox": [ + 89, + 377, + 483, + 439 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nA _ {i, j} = \\left\\| \\mathbf {x} _ {i} - \\mathbf {x} _ {j} \\right\\| _ {2}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 218, + 445, + 482, + 463 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\| \\cdot \\|_2$ denotes Euclidean distance. Intuitively, the larger the distance between the features of two images, the less likely they have the same identity. To ensure that feature propagation is constrained to images of the same identity as far as possible, a threshold-based edge deletion strategy is employed. Edges with weights exceeding a predefined threshold, which is computed as Eq. (7), are removed from the graph's adjacency matrix.", + "bbox": [ + 89, + 468, + 483, + 589 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We delete the edges whose weights exceed the corresponding threshold and get the pruned graph whose adjacent matrix can be expressed by:", + "bbox": [ + 89, + 590, + 483, + 635 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {A} _ {i, j} = \\left\\{ \\begin{array}{l l} \\| \\mathbf {x} _ {i} - \\mathbf {x} _ {j} \\| _ {2} & \\| \\mathbf {x} _ {i} - \\mathbf {x} _ {j} \\| _ {2} \\leq \\mathcal {T} _ {i} \\\\ \\infty & \\text {o t h e r w i s e ,} \\end{array} \\right. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 145, + 642, + 482, + 681 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where threshold $\\mathcal{T}_i$ is computed by Eq. (7). Setting the weight of deleted edges to infinity ensures that, after converting distance to similarity, these edges effectively become non-existent. This approach is justified by the interpretation that if two nodes are not connected, their distance is considered infinite. The desired outcome is that edge weights should positively correlate with the similarity of features between connected nodes. However, the adjacent matrix computed by Eq. (9) is just the opposite. So, we convert the distance to similarity for edge weight by:", + "bbox": [ + 89, + 686, + 483, + 839 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {S i m} (i, j) = \\exp (- \\| \\mathbf {x} _ {i} - \\mathbf {x} _ {j} \\| _ {2} / \\gamma) \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 161, + 845, + 482, + 864 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $Sim(i,j)$ denotes the similarity between image $i$ and image $j$ and $\\gamma$ is the temperature hyper-parameter. The final", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Algorithm 1 Chebyshev's Guided Graph Re-ranking", + "text_level": 1, + "bbox": [ + 514, + 90, + 883, + 106 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Input: $\\mathbf{X}_0$ - Original query and gallery image features", + "bbox": [ + 514, + 107, + 872, + 122 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Output: $\\mathbf{X}_T$ - Final features after re-ranking", + "bbox": [ + 514, + 123, + 810, + 137 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1: $A_{i,j} = \\parallel \\mathbf{x}_i - \\mathbf{x}_j\\parallel_2$ . Calculate the original adjacent matrix $\\mathbf{A}$", + "2: $\\mathcal{T}_i = \\mu_i - \\lambda \\sigma_i$ . Calculate the threshold $\\mathcal{T}_i$ .", + "3: $\\tilde{A}_{i,j} = \\left\\{ \\begin{array}{ll}\\| \\mathbf{x}_i - \\mathbf{x}_j\\| _2 & \\| \\mathbf{x}_i - \\mathbf{x}_j\\| _2\\leq \\mathcal{T}_i\\\\ \\infty & \\mathrm{otherwise} \\end{array} \\right.$ . Re-build the graph based on the $\\mathcal{T}_i$", + "4: $\\dot{A}_{i,j} = \\exp (-\\dot{A}_{i,j})$ . Convert the distances to similarities to generate the final adjacent matrix.", + "5: $D_{i,i} = \\sum_{j}A_{i,j}$ . Calculate the degree matrix $\\mathbf{D}$", + "6: Propagate the features through $\\mathbf{A}$ ." + ], + "bbox": [ + 521, + 138, + 905, + 301 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {X} _ {t + 1} = \\mathbf {D} ^ {- \\frac {1}{2}} \\hat {\\mathbf {A}} \\mathbf {D} ^ {- \\frac {1}{2}} \\mathbf {X} _ {t}\n$$\n", + "text_format": "latex", + "bbox": [ + 643, + 311, + 800, + 330 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "adjacent matrix can be computed by:", + "bbox": [ + 513, + 371, + 759, + 386 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {A} _ {i, j} = \\left\\{ \\begin{array}{l l} \\exp (- \\| \\mathbf {x} _ {i} - \\mathbf {x} _ {j} \\| _ {2} / \\gamma) & \\| \\mathbf {x} _ {i} - \\mathbf {x} _ {j} \\| _ {2} \\leq \\mathcal {T} _ {i} \\\\ 0 & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 529, + 393, + 903, + 446 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "According from Eq. (7) to Eq. (11), the operation used to compute the final adjacent matrix can be easily parallelized by CUDA [18] and implemented on PyTorch [23]. The details of Cheb-GR are presented in Algorithm 1.", + "bbox": [ + 511, + 448, + 905, + 508 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 513, + 521, + 645, + 537 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Datasets and Evaluation Protocol", + "text_level": 1, + "bbox": [ + 511, + 546, + 805, + 561 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets We conduct extensive experiments on three popular Person ReID datasets Market-1501 (Market) [39], Duke-MTMC-re-ID (Duke) [24] and CUHK03 [16]. We utilize the Duke dataset in order to make reasonable comparisons with previous work. Market1501 is a large-scale dataset consisting of 1501 identities captured by 6 cameras, featuring challenging variations in viewpoint, illumination, and occlusion. Duke is another large-scale dataset with 1404 identities captured by 8 cameras, offering more challenging scenarios due to the inclusion of pedestrian trajectories and multiple camera views. CUHK03 is a smaller dataset with 1360 identities, but it focuses on difficult cases of cross-view matching, as images are taken from two non-overlapping camera views.", + "bbox": [ + 511, + 568, + 906, + 779 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation Metrics Performance is evaluated using the standard metrics of Cumulative Matching Characteristic (CMC), mean Average Precision (mAP), and the more recently proposed mean Inverse Negative Penalty (mINP)[35]. This combination of metrics provides a comprehensive assessment of retrieval effectiveness. To evaluate the computation costs, we measure the different methods five times and average the execution times.", + "bbox": [ + 511, + 779, + 908, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "19265", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/b58548780f3b56691e30cc85879fc42e6b07d84b8ac551c34a89229cba54dd4d.jpg", + "table_caption": [ + "Table 1. Person ReID performance. We compare our methods with several re-ranking methods on Market1501, DukeMTMC, and CUHK03 datasets. mAP, Rank-1, and mINP are calculated to evaluate performance. Bolding indicates the lowest computation time, and red subscript text denotes the performance improvement compared to the baseline." + ], + "table_footnote": [], + "table_body": "
BaseMethodVenueMarketDukeCUHK03
mAP Rank-1 mINP time(s)mAP Rank-1 mINP time(s)mAP Rank-1 mINP time(s)
AGW [35]-TPAMI2188.695.366.3-79.989.845.4-66.467.355.6-
QE [6]ICCV0791.895.874.71.384.991.954.31.072.168.466.90.6
LBR [19]ICCV1991.795.576.63.984.691.856.02.171.269.664.31.0
KR [42]CVPR1794.895.989.189.890.291.775.696.078.576.176.013.4
ECN [25]CVPR1894.796.190.3161.190.792.478.2148.678.075.676.414.6
GCR [38]TMM2394.995.890.343.390.091.276.248.879.677.676.86.2
Cheb-GR-93.096.080.30.986.591.460.40.973.972.967.90.2
+4.4+0.7+14.0+6.6+1.6+15.0+7.5+5.6+12.3
TransReID [13]-ICCV2189.095.169.3-82.290.750.3-75.477.665.4-
QE [6]ICCV0791.995.377.01.787.392.859.41.680.879.075.30.6
LBR [19]ICCV1990.894.277.451.886.691.959.646.878.477.472.814.9
KR [42]CVPR1794.095.387.673.391.392.876.587.086.484.784.118.3
ECN [25]CVPR1894.095.788.7147.991.493.078.7153.786.084.484.133.6
GCR [38]TMM2392.593.886.155.590.791.775.664.384.083.781.08.0
Cheb-GR-93.295.283.51.590.493.071.21.585.483.983.80.5
+4.2+0.1+14.2+8.2+2.3+20.9+10.0+6.3+18.4
", + "bbox": [ + 112, + 136, + 883, + 444 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f116b623d140f85fa49116a98a7731582b012b9a53960b4a79cf7f2d371dd735.jpg", + "image_caption": [ + "(a) Rank-1 versus computation cost" + ], + "image_footnote": [], + "bbox": [ + 106, + 458, + 359, + 604 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/fb1213bf4b65bd7188606b92bcf55ea8e18204e973e4d69ac51942a0b12c2cfa.jpg", + "image_caption": [ + "(b) mAP versus computation cost", + "Figure 2. Comparison of computation costs versus performance improvement of different methods for person re-identification task on CUHK03 dataset. We compare our method with k-reciprocal (KR) [41], GCR [38], ECN [25], QE [6], LBR [19]. Our method requires the least computation costs among these methods and outperforms QE [6], and LBR [19]. Although our method underperforms k-reciprocal (KR) [41], GCR [38], and ECN [25], our methods only cost about one-fiftieth of their computation time. The direction of the arrow indicates that the indicator is optimal." + ], + "image_footnote": [], + "bbox": [ + 367, + 458, + 624, + 604 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e6381dd76247570dace6899bf2577bf381b4e020b52288f4e9003d18a1ecfac3.jpg", + "image_caption": [ + "(c) mINP versus computation cost" + ], + "image_footnote": [], + "bbox": [ + 633, + 458, + 885, + 604 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details The original features are extracted using the pre-trained models of AGW [35] and TransReID [13]. The values of the temperature parameter $\\gamma$ and the numbers of iterations $T$ are 0.2 and 3, respectively, which is the same as the setting of GCR [38]. Given the fact that there is a substantial discrepancy in the feature spaces produced by different models across various datasets, we adjust the value of parameter $\\lambda$ in Eq. (7) based on the specific model and dataset used. $\\lambda$ is set to 6.0/5.3/6.0/4.8/4.0 for AGW on Market/Duke/CUHK03/Market2Duke/Market2CUHK03 respectively, and 5.3/4.3/4.2 for TransReID on Mar", + "bbox": [ + 89, + 710, + 485, + 893 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ket/Duke/CUHK03 respectively. All the re-ranking methods except ECN run on GPU since running ECN on GPU causes CUDA to run out of memory.", + "bbox": [ + 511, + 710, + 906, + 757 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Comparison with State-of-Art Methods", + "text_level": 1, + "bbox": [ + 511, + 771, + 852, + 787 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We compare our methods with k-reciprocal (KR) [41], GCR [38], ECN [25], QE [6], LBR [19] on three Person ReID datasets. We run each method five times on each dataset and calculate the average computation time to measure the computation costs. The comparison result is shown in Tab. 1. Fig. 2 uses a scatter plot to visualize the trade-offs between performance metrics and", + "bbox": [ + 511, + 794, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "19266", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c17d8c33c0bb17b1c5eddfe245d803ff9729e03ca4615e717e11ee556eb1e953.jpg", + "image_caption": [ + "Figure 3. The impact of parameter $\\lambda$ on Person ReID on the CUHK03 dataset." + ], + "image_footnote": [], + "bbox": [ + 119, + 93, + 369, + 251 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/41222d296995047b0c2ad2c81da7831a6f2f39b37fcda9ab3d23ac13befd1c53.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 372, + 93, + 624, + 251 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ad1ef4c14b00fe3beb1da3ca307c8bf61f504fcd59f4f5a419e168ba4bb9068e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 625, + 94, + 875, + 250 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/4d187df7a0a1ada9e23ffe34e928008e3ff9a308513ff18b0b6387e668dfee77.jpg", + "table_caption": [ + "Table 2. Results on cross-domain settings. Market2Duke denotes training on the Market1501 dataset and testing on the DukeMTMC dataset, and Market2CUHK03 denotes training on the Market1501 dataset and testing on the CUHK03 dataset. Bolding indicates the least computation time, and red subscript text denotes the performance improvement compared to the baseline." + ], + "table_footnote": [], + "table_body": "
BaseMethodMarket2DukeMarket2CUHK03
mAP Rank-1time(s)mAP Rank-1time(s)
AGW [35]-24.338.9-8.68.2-
QE [6]28.943.40.98.67.80.6
LBR [19]31.449.12.311.111.11.1
KR [42]35.545.295.411.69.513.2
ECN [25]35.746.1152.911.18.814.7
GCR [38]31.941.746.810.28.15.8
Cheb-GR33.045.30.910.48.60.2
+8.7+6.4+1.8+0.4
", + "bbox": [ + 94, + 383, + 480, + 540 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "computational costs on the CUHK03 dataset. As we can see, our proposed method, Cheb-GR, consistently demonstrates significantly lower computational costs compared to previous approaches, achieving a runtime that is less than the fastest existing method, regardless of the dataset or baseline model used. Notably, on the Market1501 dataset, Cheb-GR achieves a runtime of 0.9 seconds, significantly faster than the 43.3 seconds required by AGW + GCR [38] and runs on average about 50 times faster than widely used k-reciprocal re-ranking. This trend persists across the DukeMTMC and CUHK03 datasets, highlighting the method's efficiency in reducing computational overhead. Furthermore, our method achieves competitive performance, surpassing methods such as QE [6] and LBR [19] in most scenarios and even demonstrating performance comparable to or exceeding GCR [38] in some cases. For instance, on the Market1501 dataset, TransReID + Cheb-GR attains an mAP of $93.2\\%$ , surpassing TransReID $^+$ QE $(91.9\\%)$ ,TransReID $^+$ LBR $(90.8\\%)$ ,and TransReID $^+$ GCR $(92.5\\%)$ while also approaching the performance of TransReID $^+$ KR and TransReID $^+$ ECN $(94.0\\%)$ . Similar observations can be made on the DukeMTMC and", + "bbox": [ + 89, + 568, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "CUHK03 datasets, demonstrating the method's ability to maintain strong performance while achieving substantial computational savings. Therefore, our method offers a compelling solution for person ReID by balancing strong performance with significantly reduced computational cost, making it a practical and efficient choice for real-world applications where both accuracy and speed are paramount.", + "bbox": [ + 511, + 291, + 906, + 398 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Comparison on Cross-domain Settings", + "text_level": 1, + "bbox": [ + 511, + 407, + 844, + 422 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To investigate the effectiveness of various re-ranking methods in cross-domain scenarios, we fine-tune a pre-trained model of AGW [35] on the Market1501 dataset and evaluate its performance on the DukeMTMC and CUHK03 dataset. We apply different re-rank methods in these cross-domain settings, and we adjust the value of k in GCR [38] from its original setting to 7, as the original setting resulted in suboptimal performance in cross-domain settings; the result is shown in Tab. 2. We can observe that performance improvement achieved by all re-ranking methods in cross-domain settings is reduced compared to the general case, highlighting the inherent intractability of the cross-domain generalization. As shown in Tab. 2, Cheb-GR exhibits competitive performance with the least computation time, showcasing its versatility across different domains. Furthermore, Cheb-GR obtains superior performance than GCR [38], demonstrating a divergence from the general trend observed, which might suggest that the adaptive neighbor search might hold a potential advantage over the k-nearest neighbor search in certain contexts.", + "bbox": [ + 511, + 429, + 906, + 731 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.4. Parameter Analysis", + "text_level": 1, + "bbox": [ + 511, + 742, + 700, + 758 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To analyze the impact of the hyper-parameter $\\lambda$ , we conduct experiments on the CUHK03 dataset. According to the theoretical analysis in Sec. 4.1, a sufficiently large value for $\\lambda$ is necessary to suppress negative samples effectively. However, an excessively large $\\lambda$ risks eliminating positive samples, hindering feature propagation. In extreme circumstances, if $\\lambda$ is so excessively large that it leads to the elimination of all samples, the resulting graph structure collapses into isolated nodes, thereby nullifying the re-ranking effect.", + "bbox": [ + 511, + 763, + 905, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "19267", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3bdc3b32b454e7a2dc9787a4f6aafa3c1ba50b1d89375e3a694bb8f877959821.jpg", + "image_caption": [ + "Figure 4. The impact of parameter $\\gamma$ and $T$ on Person ReID on the CUHK03 dataset under different values of $\\lambda$ ." + ], + "image_footnote": [], + "bbox": [ + 127, + 95, + 374, + 242 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/97c4169288a7d88ea4cd62282ada2f26dfc94da412567f6cd80f20dff300ca39.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 375, + 95, + 620, + 242 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/690c536326703222dcb7aaf11690408c2d33205b85abb62dc9995e84a134767e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 625, + 95, + 870, + 242 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Therefore, an optimal $\\lambda$ value must strike a balance between reducing noise and facilitating feature propagation. As shown in Fig. 3, our method surpasses the baseline with various values of the parameter $\\lambda$ demonstrating its competitive performance, and as the value of $\\lambda$ increases, mAP, mINP, and Rank-1 accuracy generally exhibit an initial rise, culminating in a peak value, followed by a subsequent decline, matching the theoretical analysis we propose. Moreover, the two baseline models exhibit distinct sensitivities to variations in $\\lambda$ and differing optimal $\\lambda$ values for peak performance, highlighting the disparity between the feature spaces learned by the CNN-based and transformer-based models, the transformer-based model demonstrates a considerably greater sensitivity to variations in $\\lambda$ compared to the more robust CNN-based model. This suggests that the inherent characteristics of CNN-based feature spaces may contribute to enhanced robustness when employed within GCN-based re-ranking methods.", + "bbox": [ + 88, + 295, + 485, + 566 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We analyze the impact of parameters $\\gamma$ and $T$ under three circumstances: $\\lambda$ is optimal ( $\\lambda = 4.2$ ), $\\lambda$ is insufficiently large ( $\\lambda = 3.0$ ), and $\\lambda$ is excessively large ( $\\lambda = 6.0$ ). $T$ denotes the number of iterations and $T = 0$ denotes the baseline. In each iteration, the graph convolution operation smoothes the features of connected nodes, drawing them together in the embedding space. As shown in Fig. 4, in most cases, with the increase of $T$ from 0 to 5, mAP, Rank-1, and mINP first rise and then decline. This occurs because a small value of $T$ leads to insufficient smoothing, limiting the effectiveness of feature re-ranking. In contrast, a large value of $T$ causes excessive smoothing, which may result in falsely connected features appearing closer together and separating features of unconnected positive samples. $\\gamma$ is a temperature parameter controlling the sharpness of the distribution of edges' weight. From Fig. 4, we can observe that smaller $\\gamma$ performs better when $\\lambda$ is small, and larger $\\gamma$ performs better when $\\lambda$ is large. Thus, a smaller $\\gamma$ improves performance when neighbors are redundant and noisy by facilitating the propagation of features to more similar neighbors, and conversely, a larger $\\gamma$ proves advantageous when neighbors are limited, as it allows more features to spread", + "bbox": [ + 91, + 569, + 483, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "to neighbors more equally.", + "bbox": [ + 511, + 295, + 692, + 311 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion and Discussions", + "text_level": 1, + "bbox": [ + 511, + 319, + 769, + 335 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we introduce Cheb-GR, a novel and efficient GCN-based re-ranking method for person re-identification (ReID) that addresses the critical need for both accuracy and efficiency in this challenging task. We propose a novel adaptive neighbor search for neighbor selection, replacing the computationally expensive k-nearest neighbor search and enabling significant parallelization, which is particularly well-suited for GPU acceleration. The effectiveness of Cheb-GR is rigorously validated through extensive experiments on standard benchmarks and challenging cross-domain scenarios, consistently demonstrating its competitive performance. Furthermore, we provide a theoretical analysis of the underlying principles that govern the method's efficiency and accuracy, shedding light on its ability to distinguish between similar individuals and refine the ranking process effectively. Cheb-GR's high performance and efficiency pave the way for real-time ReID applications, particularly in scenarios where speed is paramount. Beyond its demonstrated efficacy in ReID, Cheb-GR's efficient graph construction and feature refinement capabilities hold promise for broader applicability to other similarity-based retrieval tasks, suggesting its potential.", + "bbox": [ + 511, + 340, + 906, + 672 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations Cheb-GR's adaptive neighbor search, while efficient, is sensitive to the trade-off parameter $\\lambda$ , which balances removing negative and preserving positive edges. Careful tuning is crucial: an overly small $\\lambda$ hinders re-ranking, while an excessively large $\\lambda$ removes beneficial edges. Moreover, quadratic memory scaling of the adjacency matrix limits scalability, and performance depends on using features from triplet loss (or similar) trained models.", + "bbox": [ + 511, + 674, + 908, + 794 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement This work is partially supported by the National Key Research and Development Program of China (2024YFC3308400), National Natural Science Foundation of China under Grant (62176188, 62225113). The numerical calculations in this paper were supported by the supercomputing system at the Supercomputing Center of Wuhan University.", + "bbox": [ + 511, + 795, + 908, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "19268", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 90, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Yang Bai, Yucheng Ji, Min Cao, Jinqiao Wang, and Mang Ye. Chat-based person retrieval via dialogue-refined cross-modal alignment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2025. 1", + "[2] Arko Barman and Shishir K Shah. Shape: A novel graph theoretic algorithm for making consensus-based decisions in person re-identification systems. In Proceedings of the IEEE International Conference on Computer Vision, pages 1115–1124, 2017. 2", + "[3] Pafnutii Lvovich Chebyshev. Des valeurs moyennes. J. Math. Pures Appl, 12(2):177-184, 1867. 4", + "[4] Cuiqun Chen, Mang Ye, and Ding Jiang. Towards modality-agnostic person re-identification with descriptive query. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15128-15137, 2023. 2", + "[5] Yanbei Chen, Xiatian Zhu, and Shaogang Gong. Person re-identification by deep learning multi-scale representations. In Proceedings of the IEEE international conference on computer vision workshops, pages 2590–2600, 2017. 1", + "[6] Ondrej Chum, James Philbin, Josef Sivic, Michael Isard, and Andrew Zisserman. Total recall: Automatic query expansion with a generative feature model for object retrieval. In 2007 IEEE 11th International Conference on Computer Vision, pages 1-8. IEEE, 2007. 6, 7", + "[7] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4690-4699, 2019. 2", + "[8] Bo Du, Fangyi Liu, and YE Mang. Person re-identification method based on perspective-guided multi-adversarial attention, Oct. 31 2023. US Patent 11,804,036. 2", + "[9] Xing Fan, Wei Jiang, Hao Luo, and Mengjuan Fei. Spherereid: Deep hypersphere manifold embedding for person re-identification. Journal of Visual Communication and Image Representation, 60:51-58, 2019. 2", + "[10] Niloofar Gheissari, Thomas B Sebastian, and Richard Hartley. Person reidentification using spatiotemporal appearance. In 2006 IEEE computer society conference on computer vision and pattern recognition (CVPR'06), volume 2, pages 1528-1535. IEEE, 2006. 1", + "[11] Justin Gilmer, Samuel S Schoenholz, Patrick F Riley, Oriol Vinyals, and George E Dahl. Neural message passing for quantum chemistry. In International conference on machine learning, pages 1263-1272. PMLR, 2017. 2", + "[12] Raia Hadsell, Sumit Chopra, and Yann LeCun. Dimensionality reduction by learning an invariant mapping. In 2006 IEEE computer society conference on computer vision and pattern recognition (CVPR'06), volume 2, pages 1735-1742. IEEE, 2006. 2", + "[13] Shuting He, Hao Luo, Pichao Wang, Fan Wang, Hao Li, and Wei Jiang. Transreid: Transformer-based object re-identification. In Proceedings of the IEEE/CVF international conference on computer vision, pages 15013-15022, 2021. 2, 6", + "[14] Bo Jiang, Xixi Wang, Aihua Zheng, Jin Tang, and Bin Luo. Ph-gen: Person retrieval with part-based hierarchical graph" + ], + "bbox": [ + 93, + 114, + 483, + 888 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "convolutional network. IEEE Transactions on Multimedia, 24:3218-3228, 2021. 1, 3", + "[15] Thomas N Kipf and Max Welling. Semi-supervised classification with graph convolutional networks. arXiv preprint arXiv:1609.02907, 2016. 2", + "[16] Wei Li, Rui Zhao, Tong Xiao, and Xiaogang Wang. Deepreid: Deep filter pairing neural network for person re-identification. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 152–159, 2014. 5", + "[17] Fangyi Liu, Mang Ye, and Bo Du. Learning a generalizable re-identification model from unlabelled data with domain-agnostic expert. Visual Intelligence, 2(1):28, 2024. 1", + "[18] David Luebke. Cuda: Scalable parallel programming for high-performance scientific computing. In 2008 5th IEEE international symposium on biomedical imaging: from nano to macro, pages 836-838. IEEE, 2008. 5", + "[19] Chuanchen Luo, Yuntao Chen, Naiyan Wang, and Zhaoxiang Zhang. Spectral feature transformation for person re-identification. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4976–4985, 2019. 6, 7", + "[20] Hao Luo, Youzhi Gu, Xingyu Liao, Shenqi Lai, and Wei Jiang. Bag of tricks and a strong baseline for deep person re-identification. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 0–0, 2019. 2", + "[21] Jiaxu Miao, Yu Wu, Ping Liu, Yuhang Ding, and Yi Yang. Pose-guided feature alignment for occluded person re-identification. In Proceedings of the IEEE/CVF international conference on computer vision, pages 542–551, 2019. 1", + "[22] Sakrapee Paisitkriangkrai, Chunhua Shen, and Anton Van Den Hengel. Learning to rank in person re-identification with metric ensembles. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1846–1855, 2015. 2", + "[23] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019. 5", + "[24] Ergys Ristani, Francesco Solera, Roger Zou, Rita Cucchiara, and Carlo Tomasi. Performance measures and a data set for multi-target, multi-camera tracking. In European conference on computer vision, pages 17-35. Springer, 2016. 5", + "[25] M Saquib Sarfraz, Arne Schumann, Andreas Eberle, and Rainer Stiefelhagen. A pose-sensitive embedding for person re-identification with expanded cross neighborhood reranking. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 420–429, 2018. 1, 2, 4, 6, 7", + "[26] Florian Schroff, Dmitry Kalenichenko, and James Philbin. Facenet: A unified embedding for face recognition and clustering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 815-823, 2015. 2", + "[27] Yantao Shen, Hongsheng Li, Shuai Yi, Dapeng Chen, and Xiaogang Wang. Person re-identification with deep" + ], + "bbox": [ + 516, + 93, + 903, + 892 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "19269", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "similarity-guided graph neural network. In Proceedings of the European conference on computer vision (ECCV), pages 486-504, 2018. 1", + "[28] Yifan Sun, Changmao Cheng, Yuhan Zhang, Chi Zhang, Liang Zheng, Zhongdao Wang, and Yichen Wei. Circle loss: A unified perspective of pair similarity optimization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6398-6407, 2020. 2", + "[29] Yifan Sun, Qin Xu, Yali Li, Chi Zhang, Yikang Li, Shengjin Wang, and Jian Sun. Perceive where to focus: Learning visibility-aware part-level features for partial person re-identification. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 393–402, 2019. 1", + "[30] Yifan Sun, Liang Zheng, Yi Yang, Qi Tian, and Shengjin Wang. Beyond part models: Person retrieval with refined part pooling (and a strong convolutional baseline). In Proceedings of the European conference on computer vision (ECCV), pages 480-496, 2018. 1, 2", + "[31] Haochen Wang, Jiayi Shen, Yongtuo Liu, Yan Gao, and Efstratos Gavves. Nformer: Robust person re-identification with neighbor transformer. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7297-7307, 2022. 2", + "[32] Yingjia Xu, Mengxia Wu, Zixin Guo, Min Cao, Mang Ye, and Jorma Laaksonen. Efficient text-to-video retrieval via multi-modal multi-tags derived pre-screening. Visual Intelligence, 3(1):1-13, 2025. 1", + "[33] Mang Ye, Shuoyi Chen, Chenyue Li, Wei-Shi Zheng, David Crandall, and Bo Du. Transformer for object re-identification: A survey. International Journal of Computer Vision, pages 1–31, 2024. 1, 2", + "[34] Mang Ye, Chao Liang, Zheng Wang, Qingming Leng, and Jun Chen. Ranking optimization for person re-identification via similarity and dissimilarity. In Proceedings of the 23rd ACM international conference on Multimedia, pages 1239–1242, 2015. 2", + "[35] Mang Ye, Jianbing Shen, Gaojie Lin, Tao Xiang, Ling Shao, and Steven CH Hoi. Deep learning for person re-identification: A survey and outlook. IEEE transactions on pattern analysis and machine intelligence, 44(6):2872-2893, 2021. 2, 5, 6, 7", + "[36] Mang Ye, Zesen Wu, and Bo Du. Dual-level matching with outlier filtering for unsupervised visible-infrared person re-identification. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025. 1", + "[37] Xuanmeng Zhang, Minyue Jiang, Zhedong Zheng, Xiao Tan, Errui Ding, and Yi Yang. Understanding image retrieval reranking: A graph neural network perspective. arXiv preprint arXiv:2012.07620, 2020. 2, 3", + "[38] Yuqi Zhang, Qi Qian, Hongsong Wang, Chong Liu, Weihua Chen, and Fan Wang. Graph convolution based efficient reranking for visual retrieval. IEEE Transactions on Multimedia, 26:1089-1101, 2023. 1, 2, 3, 4, 6, 7", + "[39] Liang Zheng, Liyue Shen, Lu Tian, Shengjin Wang, Jingdong Wang, and Qi Tian. Scalable person re-identification: A benchmark. In Proceedings of the IEEE international conference on computer vision, pages 1116–1124, 2015. 5", + "[40] Liang Zheng, Yi Yang, and Alexander G Hauptmann. Per" + ], + "bbox": [ + 91, + 90, + 483, + 892 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "son re-identification: Past, present and future. arXiv preprint arXiv:1610.02984, 2016. 1", + "[41] Meng Zheng, Srikrishna Karanam, Ziyan Wu, and Richard J Radke. Re-identification with consistent attentive siamese networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5735–5744, 2019. 1, 6", + "[42] Zhun Zhong, Liang Zheng, Donglin Cao, and Shaozi Li. Reranking person re-identification with k-reciprocal encoding. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1318–1327, 2017. 1, 2, 4, 6, 7" + ], + "bbox": [ + 516, + 92, + 903, + 253 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "19270", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Cheb-GR_ Rethinking K-nearest Neighbor Search in Re-ranking for Person Re-identification/9e933435-e0fc-412e-b745-089375b60f23_model.json b/2025/Cheb-GR_ Rethinking K-nearest Neighbor Search in Re-ranking for Person Re-identification/9e933435-e0fc-412e-b745-089375b60f23_model.json new file mode 100644 index 0000000000000000000000000000000000000000..5cdfd77691e74ab5efa63143afc8a668e4f3faca --- /dev/null +++ b/2025/Cheb-GR_ Rethinking K-nearest Neighbor Search in Re-ranking for Person Re-identification/9e933435-e0fc-412e-b745-089375b60f23_model.json @@ -0,0 +1,2101 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.126, + 0.13, + 0.873, + 0.175 + ], + "angle": 0, + "content": "Cheb-GR: Rethinking k-nearest neighbor search in Re-ranking for Person Re-identification" + }, + { + "type": "text", + "bbox": [ + 0.221, + 0.203, + 0.77, + 0.294 + ], + "angle": 0, + "content": "Jinxi Yang * He Li * Bo Du Mang Ye† \nNational Engineering Research Center for Multimedia Software, School of Computer Science, Wuhan University, Wuhan, China \n{jinxiyang, lihe404, dubo, yemang}@whu.edu.cn \nhttps://github.com/Jinxi-Yang-WHU/Fast-GCR.git" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.327, + 0.327, + 0.344 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.358, + 0.486, + 0.768 + ], + "angle": 0, + "content": "Person re-identification (ReID) is the task of matching individuals across different camera views. Existing approaches typically employ neural networks to extract discriminative features, ranking gallery images based on their similarities to probe images. While effective, these methods are often enhanced through re-ranking, a post-processing step that refines initial retrieval results without requiring additional model training. However, current re-ranking methods mostly rely on \\( k \\)-nearest neighbor search to extract similar images that might have the same identity as the query, which is time-consuming with a high computation burden, limiting their applications in reality. We rethink the effect of the \\( k \\)-nearest neighbor search and introduce the Chebyshev's Theorem-guided Graph Re-ranking (ChebGR) method, which adopts the adaptive neighbor search guided by Chebyshev's Theorem over the \\( k \\)-nearest neighbor search for efficient neighbor selection. Our method leverages graph convolution operations to refine image features and achieve robust re-ranking, leading to enhanced retrieval performance. Furthermore, we provide a theoretical analysis based on Chebyshev's Inequality to elucidate the factors contributing to the strong performance of the proposed method. Our method significantly reduces the computation costs while maintaining relatively strong performance. Through extensive experiments in both general and cross-domain settings, we demonstrate the effectiveness of Cheb-GR and its potential for real-world applications." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.795, + 0.224, + 0.811 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.821, + 0.485, + 0.867 + ], + "angle": 0, + "content": "Person re-identification (ReID) is a computer vision task that aims to identify the same person across different camera views in a multi-camera surveillance system [5, 40]. It" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.328, + 0.908, + 0.51 + ], + "angle": 0, + "content": "involves matching individuals captured by different cameras based on their appearance, even when they are viewed from different angles or under varying lighting conditions [10]. Deep learning has become the predominant approach in ReID, enabling automatic feature extraction and significantly enhancing performance [17, 21, 29, 30, 32, 33, 36, 41]. However, these methods primarily rely on pairwise similarity comparisons between the query image and individual gallery images, neglecting the contextual relationships among images within the ranking list. Consequently, they are susceptible to noise in feature representations arising from variations in illumination, pose, or occlusion." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.517, + 0.909, + 0.775 + ], + "angle": 0, + "content": "To tackle this issue, many re-ranking methods have emerged as a prominent strategy for enhancing performance without the need for additional model training. Prevalent approaches involve utilizing k-nearest neighbor search to identify similar images, thereby extracting contextual information [25, 42]. However, this reliance on k-nearest neighbor search presents a significant bottleneck in real-world applications due to its computational complexity. Meanwhile, many works utilize Graph Convolutional Networks (GCN) to learn pairwise similarity information and relational information between images in person ReID tasks [1, 14, 27]. Notably, Zhang et al. [38] demonstrate the effectiveness of GCNs for re-ranking by refining embeddings through feature propagation. Despite their potential, these GCN-based methods still necessitate k-nearest neighbor search, which is inherently sequential and hinders parallel computation, leading to substantial time consumption." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.911, + 0.903 + ], + "angle": 0, + "content": "To improve re-rank efficiency, we propose the Chebyshev's Theorem-guided Graph Re-ranking (Cheb-GR) method that eliminates the need for the k-nearest neighbor search and can be executed in parallel. Specifically, we construct a graph where each node represents an image. Unlike GCR [38], which connects each node to its k-nearest neighbors, Cheb-GR uses feature distances as edge weights, discarding edges with extreme distances. The remaining dis" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.875, + 0.218, + 0.888 + ], + "angle": 0, + "content": "*Equal contribution" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.419, + 0.901 + ], + "angle": 0, + "content": "† Corresponding Author: Mang Ye (yemang@whu.edu.cn)" + }, + { + "type": "list", + "bbox": [ + 0.109, + 0.875, + 0.419, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "19261" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.288 + ], + "angle": 0, + "content": "tances are then converted into similarities. This approach directly constrains feature propagation to similar images, a key factor in achieving strong performance. A notable advantage of our method is that all operations are straightforward and highly parallelizable, resulting in significantly reduced computation costs and much faster execution compared to existing methods [25, 38, 42]. Additionally, ChebGR exhibits robust performance in cross-domain scenarios, further demonstrating the versatility of our method. We also provide a theoretical analysis to explain why the proposed approach performs effectively without relying on k-nearest neighbor search, which is a method commonly used by various re-ranking methods." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.288, + 0.482, + 0.334 + ], + "angle": 0, + "content": "To the best of our knowledge, Cheb-GR is the fastest re-ranking method currently available, delivering competitive performance. In summary, our contributions are threefold:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.334, + 0.482, + 0.422 + ], + "angle": 0, + "content": "- We propose Chebyshev's Theorem-guided Graph Re-ranking (Cheb-GR) method with no need for finding k-nearest neighbors, which runs faster than all re-rank methods we know. Besides, Cheb-GR achieves stronger performance on cross-domain settings than most existing methods." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.424, + 0.482, + 0.469 + ], + "angle": 0, + "content": "- A reasonable theoretical explanation of the proposed method is presented to reveal why Cheb-GR can obtain strong performance without k-nearest neighbor search." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.469, + 0.482, + 0.515 + ], + "angle": 0, + "content": "- Extensive experimentation across a diverse range of datasets provides compelling evidence of our method's strong performance and low computation costs." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.334, + 0.482, + 0.515 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.529, + 0.233, + 0.544 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.554, + 0.482, + 0.674 + ], + "angle": 0, + "content": "Person Re-identification. Recent advancements in artificial intelligence have propelled person re-identification (Re-ID), with some methods even surpassing human performance. Current approaches can be broadly divided into two categories: the first focuses on designing metric learning loss functions to learn discriminative features, while the second modifies model architectures to capture contextual information, improving feature representation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.675, + 0.483, + 0.901 + ], + "angle": 0, + "content": "For model architecture modification, many works have been proposed [4, 8, 31, 33, 35]. Sun et al. [30] propose a part-based convolutional baseline (PCB), utilizing uniform partitioning to capture local image information and employing a refined part pooling (RPP) method to enhance within-part consistency. Luo et al. [20] propose BoT, a strong baseline that adds a BNNeck structure after ResNet50. BNNeck introduces a batch normalization layer to separate ID and triplet loss optimization, improving convergence and performance in person ReID. He et al. [13] introduced TransReID, the first transformer-based model for person re-identification to achieve performance comparable to convolutional neural networks. For metric learning loss design, Softmax loss or its modification [7, 9] have been widely used in person Re-ID for feature learning. Loss functions" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.243 + ], + "angle": 0, + "content": "utilized in metric learning, such as triplet loss [26] and contrastive loss [12], are also applied to the person Re-ID task. Some works also design novel loss functions to guide the model to learn distinguishable representations [28, 35]. Sun et al. [28] re-weight each similarity score to obtain flexible optimization and a definite convergence target. Ye et al. [35] propose Weighted Regularization Triplet loss to take full advantage of relative distance optimization between positive and negative pairs without introducing additional margin parameters." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.244, + 0.905, + 0.336 + ], + "angle": 0, + "content": "All these works are committed to extracting distinguishable features of images and then calculating the similarity between the feature of a query image and the feature of each gallery image to generate a ranking list. However, these works ignore abundant contextual information within the ranking list." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.337, + 0.905, + 0.534 + ], + "angle": 0, + "content": "Re-ranking for Person Re-identification. Most existing methods for Re-ID concentrate on extracting distinctive image features and assessing the similarity between the features of the query image and those of the gallery images. Additionally, a significant number of researchers have directed their attention towards re-ranking methods within ReID tasks, aiming to harness gallery-to-gallery similarity [25, 34, 42] or to integrate various ReID approaches [2, 22]. While traditional re-ranking methods have primarily emphasized the design of innovative similarity metrics, a recent advancement involves employing the graph convolutional network (GCN) to update features of images directly, which is more efficient and scalable [38]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.535, + 0.905, + 0.687 + ], + "angle": 0, + "content": "Most re-ranking methods design novel distance metrics to consider global contextual information or directly refine the features of images to improve performance [25, 37, 38, 42]. Zhong et al. [42] propose a k-reciprocal encoding method that leverages contextual information between k-reciprocal nearest neighbors to improve performance. Zhang et al. [38] propose a feature propagation-based method, constructing a graph by connecting each image with its k-nearest neighbors and propagating features by graph convolution on the graph." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.688, + 0.905, + 0.779 + ], + "angle": 0, + "content": "However, existing methods rely on computationally expensive k-nearest neighbor searches to identify similar images, limiting their practical applicability, particularly for large-scale datasets. So, we propose a method that circumvents the need to find k-nearest neighbors, significantly reducing computation costs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Graph Convolutional Network in ReID. Graph Convolutional Network (GCN) [11] utilizes a message-passing mechanism to allow nodes to exchange and aggregate information with neighbors. Through a convolutional operation, each node's representation is updated by a weighted sum of its features and those of its neighbors, computed via matrix multiplication with a learned weight matrix [15]. This iterative process allows GCNs to learn rich representations that" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19262" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.107 + ], + "angle": 0, + "content": "capture both local and global information within the graph." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.107, + 0.483, + 0.274 + ], + "angle": 0, + "content": "Since graphs are a great data structure for describing relationships and GCNs can learn graph-structure information, many ReID methods utilize GCNs to learn pairwise similarity and relational information among image sets. Jiang et al. [14] propose a method based on a hierarchical graph to integrate local, global, and structural information. Zhang et al. [37] and GCR [38] show that GCN can be leveraged in the re-ranking area. This paper employs a simplified GCN that achieves feature propagation without a learned weight matrix. This parameter-free approach enhances efficiency by eliminating the need for training." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.285, + 0.351, + 0.303 + ], + "angle": 0, + "content": "3. Preliminary and Motivation" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.31, + 0.484, + 0.522 + ], + "angle": 0, + "content": "Efficiency of K-nearest Neighbor Search. Current person ReID re-ranking methods rely heavily on k-nearest neighbor search, a key operation used to select potential same-identity images. K-nearest neighbor search computes the distances between the query image and all other images in the gallery to identify the k-nearest neighbors, which is computationally expensive and presents significant challenges when parallelizing on GPUs. The time complexity of sorting N images is \\( \\Theta (N\\log N) \\). In order to select the k-nearest neighbors for each node, we need to perform the sorting operations \\( N \\) times and select the top-k images each time. Thus, the total time complexity of the k-nearest neighbor search is \\( \\Theta (N^2\\log N) \\), which becomes the primary computational bottleneck in the re-ranking process." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.522, + 0.484, + 0.704 + ], + "angle": 0, + "content": "Graph based Re-ranking. Zhang et al. [38] propose Graph Convolutional-based Re-ranking (GCR). In GCR, given a set of images, a graph is constructed that encompasses both query and gallery images. Each query and gallery image corresponds to a node in the graph, with its associated feature vector serving as the initial node feature. Edges are established between nodes representing visually similar images, encoding potential identity relationships within the dataset. This constructed graph then serves as the input to a GCN, which performs iterative message-passing operations to refine the node features. The message-passing operation in a GCN layer can be written as:" + }, + { + "type": "equation", + "bbox": [ + 0.19, + 0.711, + 0.483, + 0.731 + ], + "angle": 0, + "content": "\\[\n\\mathbf {X} ^ {(l)} = \\mathbf {D} ^ {- \\frac {1}{2}} \\mathbf {A D} ^ {- \\frac {1}{2}} \\mathbf {X} ^ {(l - 1)} \\mathbf {W}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.737, + 0.483, + 0.845 + ], + "angle": 0, + "content": "where \\(\\mathbf{X}^{(l)}\\in \\mathbb{R}^{N\\times d}\\) is the node features in the \\(l^{th}\\) layer, \\(N\\) denotes the number of nodes, \\(d\\) denotes the feature dimensionality. \\(\\mathbf{A}\\in \\mathbb{R}^{N\\times N}\\) is the adjacent matrix of the graph, \\(\\mathbf{D}\\in \\mathbb{R}^{N\\times N}\\) is the degree matrix of the graph which is a diagonal matrix and \\(\\mathbf{D}_{i,i} = \\sum_{j}\\mathbf{A}_{i,j}\\). \\(\\mathbf{W}\\in \\mathbb{R}^{d\\times d'}\\) denotes a learnable weighted matrix in the GCN layer, which can project the features vector into a certain space." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.845, + 0.484, + 0.875 + ], + "angle": 0, + "content": "When the weighted matrix \\(\\mathbf{W}\\) is an identity matrix, the graph convolution operation can be written as:" + }, + { + "type": "equation", + "bbox": [ + 0.197, + 0.883, + 0.483, + 0.902 + ], + "angle": 0, + "content": "\\[\n\\mathbf {X} ^ {(l)} = \\mathbf {D} ^ {- \\frac {1}{2}} \\mathbf {A} \\mathbf {D} ^ {- \\frac {1}{2}} \\mathbf {X} ^ {(l - 1)}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.425 + ], + "angle": 0, + "content": "As formalized in Eq. (2), this parameter-free operation iteratively updates each node's feature embedding by aggregating the features of its neighbors and itself. The weight assigned to each neighbor's contribution is determined by the corresponding edge weight in the graph. Ideally, Graph Convolution-based Re-ranking enhances feature representations by promoting propagation between images of the same identity. This process brings together features of matching identities while pushing apart features of distinct identities, leading to improved re-ranking performance. Therefore, the primary challenge lies in constructing a graph that maximizes connections between images of the same identity while minimizing connections between images of different identities. As illustrated in Fig. 1(b), GCR employs a k-nearest neighbor search to identify neighboring nodes. According to Eq. (2), the time complexity of a graph convolution operation in GCR [38] is equal to matrix multiplication, which is \\(\\Theta(N^3)\\). However, since the matrix multiplication is highly parallelized on GPUs, it runs much faster than the k-nearest neighbor search in practice. Thus the efficiency bottleneck of GCR is still mainly affected by k-nearest neighbor search." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.438, + 0.688, + 0.456 + ], + "angle": 0, + "content": "4. Proposed Method" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.464, + 0.907, + 0.524 + ], + "angle": 0, + "content": "In this section, we first revisit the neighbor search and introduce an efficient strategy to replace the k-nearest neighbor search in Sec. 4.1. After that, we introduce our method in detail in Sec. 4.2." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.535, + 0.762, + 0.551 + ], + "angle": 0, + "content": "4.1. Revisiting Neighbor Search" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.557, + 0.906, + 0.708 + ], + "angle": 0, + "content": "Since the role of the k-nearest neighbor search in re-ranking is to select potential positive samples, we need to find an alternative strategy to find potential positive samples with low time complexity to lower the computation costs. Since feature distance is inversely related to the probability of shared identity, a threshold can be established based on the distance between the features of an anchor image and other images. Images with a distance below this threshold are then considered positive samples, sharing the same identity as the anchor image." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.709, + 0.906, + 0.753 + ], + "angle": 0, + "content": "The determination of an appropriate threshold for feature distance warrants further consideration. The triplet loss can be written as:" + }, + { + "type": "equation", + "bbox": [ + 0.587, + 0.767, + 0.907, + 0.784 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\max (d (a, p) - d (a, n) + \\alpha , 0), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.906, + 0.902 + ], + "angle": 0, + "content": "where \\(\\alpha\\) is the margin hyperparameter. Eq. (3) implies that, ideally, the distance between images of different identities should exceed the distance between images of the same identity by at least a factor of \\(\\alpha\\). Given the significant disparity in the number of positive and negative samples within the datasets, with far fewer positive samples, we can simplify our analysis by focusing on the distribution of the" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19263" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.135, + 0.09, + 0.518, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.524, + 0.091, + 0.864, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.134, + 0.264, + 0.625, + 0.436 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.636, + 0.264, + 0.864, + 0.436 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.448, + 0.907, + 0.546 + ], + "angle": 0, + "content": "Figure 1. Comparison between traditional re-rank methods. GCR [38] and the proposed Chebyshev's Guided Graph Re-ranking (Cheb-GR). (a) Most traditional re-rank methods, such as k-reciprocal [42] and ECN [25], operate in three stages: neighborhood set construction, dissimilarity measurement between these sets, and weighted distance aggregation combining direct and neighborhood-based distances. (b) GCR connects each node to its k-nearest neighbors to construct the graph. GCNs are then employed to propagate the node features through the graph, refining the image representations for improved re-ranking. (c) Cheb-GR leverages an adaptive neighbor search instead of a k-nearest neighbor search to select neighbors and construct the graph. Subsequently, feature propagation is achieved through the application of a GCN." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.572, + 0.483, + 0.633 + ], + "angle": 0, + "content": "distance between negative samples in the embedding space. Thus, the problem is reframed as one of identifying the minimum distance separating the negative samples. According to Chebyshev's Theorem [3]:" + }, + { + "type": "equation", + "bbox": [ + 0.133, + 0.646, + 0.316, + 0.678 + ], + "angle": 0, + "content": "\\[\nP r [ | R | \\geq x ] \\leq \\frac {E x [ | R | ^ {\\alpha} ]}{x ^ {\\alpha}},\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.136, + 0.679, + 0.483, + 0.71 + ], + "angle": 0, + "content": "\\[\nP r [ | R - E x [ R ] | \\geq x ] \\leq \\frac {E x [ | R - E x [ R ] | ^ {\\alpha} ]}{x ^ {\\alpha}}, \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.136, + 0.711, + 0.366, + 0.742 + ], + "angle": 0, + "content": "\\[\nP r [ | R - E x [ R ] | \\geq x ] \\leq \\frac {V a r [ R ]}{x ^ {2}},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.755, + 0.483, + 0.831 + ], + "angle": 0, + "content": "where \\(R\\) denotes a random variable, \\(Ex[R]\\) denotes the expected value of the random variable \\(R\\), \\(Var[R]\\) denotes the variance of the random variable \\(R\\), \\(x\\) denotes a real number. We can replace \\(x\\) with \\(\\lambda \\sigma\\), where \\(\\lambda\\) is a constant hyperparameter and \\(\\sigma\\) is standard deviation, to get the inequality:" + }, + { + "type": "equation", + "bbox": [ + 0.183, + 0.844, + 0.483, + 0.874 + ], + "angle": 0, + "content": "\\[\nP r [ | R - E x [ R ] | \\geq \\lambda \\sigma ] \\leq \\frac {1}{\\lambda^ {2}}. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.886, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Then, removing the absolute value operation and ignoring" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.572, + 0.896, + 0.588 + ], + "angle": 0, + "content": "the scenario that \\(R \\geq Ex[R] + \\lambda \\sigma\\), we get the inequality:" + }, + { + "type": "equation", + "bbox": [ + 0.61, + 0.599, + 0.907, + 0.629 + ], + "angle": 0, + "content": "\\[\nP r [ R \\leq E x [ R ] - \\lambda \\sigma ] \\leq \\frac {1}{\\lambda^ {2}}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.637, + 0.907, + 0.698 + ], + "angle": 0, + "content": "which shows that only a small number of elements fall below the mean minus \\(\\lambda\\) times the standard deviation when \\(\\lambda\\) is sufficiently large. Thus, given an image \\(i\\), we can set threshold \\(\\mathcal{T}_i\\) for image \\(i\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.607, + 0.709, + 0.713, + 0.726 + ], + "angle": 0, + "content": "\\[\n\\mathcal {T} _ {i} = \\mu_ {i} - \\lambda \\sigma_ {i},\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.608, + 0.728, + 0.906, + 0.772 + ], + "angle": 0, + "content": "\\[\n\\mu_ {i} = \\frac {1}{N} \\sum_ {j = 1} ^ {N} D i s _ {i, j}, \\tag {7}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.608, + 0.775, + 0.813, + 0.816 + ], + "angle": 0, + "content": "\\[\n\\sigma_ {i} = \\sqrt {\\frac {\\sum_ {j = 1} ^ {N} (D i s _ {i , j} - \\mu_ {i}) ^ {2}}{N}},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.907, + 0.9 + ], + "angle": 0, + "content": "where \\( Dis_{i,j} \\) denotes the distance between features of image \\( i \\) and image \\( j \\), \\( N \\) denotes the number of images in the dataset, \\( \\lambda \\) is a hyper-parameter, \\( \\mu_i \\) is the mean of distances between features of an image in the dataset and image \\( i \\), \\( \\sigma_i \\) is the standard deviations of distances between features of" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19264" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.35 + ], + "angle": 0, + "content": "an image in the dataset and image \\( i \\). Given the presence of a margin \\( \\alpha \\) between the distances of positive and negative samples, there exists a specific value of \\( \\lambda \\) that ensures that the threshold \\( \\mathcal{T}_i \\) surpasses the distances between positive samples while remaining below the distances between negative samples. Therefore, by carefully selecting an appropriate value for \\( \\lambda \\), we can effectively filter out negative samples while preserving positive samples within the graph construction process. The time complexity of calculating the mean and standard deviation of \\( N \\) elements is \\( \\Theta(N) \\), so the time complexity of calculating the threshold for \\( N \\) images is \\( \\Theta(N^2) \\), which is less than the time complexity of the k-nearest neighbor search. Moreover, our proposed method can be easily parallelized on GPUs and implemented on the PyTorch platform [23], so the computation cost of our method is much less than the computation costs of the k-nearest neighbor search in practice." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.356, + 0.434, + 0.373 + ], + "angle": 0, + "content": "4.2. Chebyshev's Guided Graph Re-ranking" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.378, + 0.484, + 0.44 + ], + "angle": 0, + "content": "The pipeline of our method is shown in Fig. 1(c). We first construct a directed complete graph, and the weight of each edge is the distance between the nodes the edge connects. The adjacent matrix of this graph can be computed by:" + }, + { + "type": "equation", + "bbox": [ + 0.22, + 0.446, + 0.483, + 0.464 + ], + "angle": 0, + "content": "\\[\nA _ {i, j} = \\left\\| \\mathbf {x} _ {i} - \\mathbf {x} _ {j} \\right\\| _ {2}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.469, + 0.484, + 0.59 + ], + "angle": 0, + "content": "where \\(\\| \\cdot \\|_2\\) denotes Euclidean distance. Intuitively, the larger the distance between the features of two images, the less likely they have the same identity. To ensure that feature propagation is constrained to images of the same identity as far as possible, a threshold-based edge deletion strategy is employed. Edges with weights exceeding a predefined threshold, which is computed as Eq. (7), are removed from the graph's adjacency matrix." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.591, + 0.484, + 0.636 + ], + "angle": 0, + "content": "We delete the edges whose weights exceed the corresponding threshold and get the pruned graph whose adjacent matrix can be expressed by:" + }, + { + "type": "equation", + "bbox": [ + 0.146, + 0.643, + 0.483, + 0.683 + ], + "angle": 0, + "content": "\\[\n\\tilde {A} _ {i, j} = \\left\\{ \\begin{array}{l l} \\| \\mathbf {x} _ {i} - \\mathbf {x} _ {j} \\| _ {2} & \\| \\mathbf {x} _ {i} - \\mathbf {x} _ {j} \\| _ {2} \\leq \\mathcal {T} _ {i} \\\\ \\infty & \\text {o t h e r w i s e ,} \\end{array} \\right. \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.688, + 0.484, + 0.84 + ], + "angle": 0, + "content": "where threshold \\(\\mathcal{T}_i\\) is computed by Eq. (7). Setting the weight of deleted edges to infinity ensures that, after converting distance to similarity, these edges effectively become non-existent. This approach is justified by the interpretation that if two nodes are not connected, their distance is considered infinite. The desired outcome is that edge weights should positively correlate with the similarity of features between connected nodes. However, the adjacent matrix computed by Eq. (9) is just the opposite. So, we convert the distance to similarity for edge weight by:" + }, + { + "type": "equation", + "bbox": [ + 0.162, + 0.847, + 0.483, + 0.865 + ], + "angle": 0, + "content": "\\[\n\\operatorname {S i m} (i, j) = \\exp (- \\| \\mathbf {x} _ {i} - \\mathbf {x} _ {j} \\| _ {2} / \\gamma) \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "where \\( Sim(i,j) \\) denotes the similarity between image \\( i \\) and image \\( j \\) and \\( \\gamma \\) is the temperature hyper-parameter. The final" + }, + { + "type": "title", + "bbox": [ + 0.515, + 0.091, + 0.885, + 0.107 + ], + "angle": 0, + "content": "Algorithm 1 Chebyshev's Guided Graph Re-ranking" + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.108, + 0.873, + 0.123 + ], + "angle": 0, + "content": "Input: \\(\\mathbf{X}_0\\) - Original query and gallery image features" + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.124, + 0.812, + 0.138 + ], + "angle": 0, + "content": "Output: \\(\\mathbf{X}_T\\) - Final features after re-ranking" + }, + { + "type": "text", + "bbox": [ + 0.522, + 0.14, + 0.906, + 0.17 + ], + "angle": 0, + "content": "1: \\( A_{i,j} = \\parallel \\mathbf{x}_i - \\mathbf{x}_j\\parallel_2 \\). Calculate the original adjacent matrix \\( \\mathbf{A} \\)" + }, + { + "type": "text", + "bbox": [ + 0.522, + 0.171, + 0.823, + 0.186 + ], + "angle": 0, + "content": "2: \\(\\mathcal{T}_i = \\mu_i - \\lambda \\sigma_i\\). Calculate the threshold \\(\\mathcal{T}_i\\)." + }, + { + "type": "text", + "bbox": [ + 0.522, + 0.186, + 0.906, + 0.237 + ], + "angle": 0, + "content": "3: \\(\\tilde{A}_{i,j} = \\left\\{ \\begin{array}{ll}\\| \\mathbf{x}_i - \\mathbf{x}_j\\| _2 & \\| \\mathbf{x}_i - \\mathbf{x}_j\\| _2\\leq \\mathcal{T}_i\\\\ \\infty & \\mathrm{otherwise} \\end{array} \\right.\\) . Re-build the graph based on the \\(\\mathcal{T}_i\\)" + }, + { + "type": "text", + "bbox": [ + 0.522, + 0.238, + 0.905, + 0.267 + ], + "angle": 0, + "content": "4: \\(\\dot{A}_{i,j} = \\exp (-\\dot{A}_{i,j})\\) . Convert the distances to similarities to generate the final adjacent matrix." + }, + { + "type": "text", + "bbox": [ + 0.522, + 0.268, + 0.857, + 0.285 + ], + "angle": 0, + "content": "5: \\(D_{i,i} = \\sum_{j}A_{i,j}\\) . Calculate the degree matrix \\(\\mathbf{D}\\)" + }, + { + "type": "text", + "bbox": [ + 0.522, + 0.287, + 0.766, + 0.302 + ], + "angle": 0, + "content": "6: Propagate the features through \\(\\mathbf{A}\\)." + }, + { + "type": "list", + "bbox": [ + 0.522, + 0.14, + 0.906, + 0.302 + ], + "angle": 0, + "content": null + }, + { + "type": "equation", + "bbox": [ + 0.645, + 0.312, + 0.801, + 0.331 + ], + "angle": 0, + "content": "\\[\n\\mathbf {X} _ {t + 1} = \\mathbf {D} ^ {- \\frac {1}{2}} \\hat {\\mathbf {A}} \\mathbf {D} ^ {- \\frac {1}{2}} \\mathbf {X} _ {t}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.372, + 0.76, + 0.387 + ], + "angle": 0, + "content": "adjacent matrix can be computed by:" + }, + { + "type": "equation", + "bbox": [ + 0.531, + 0.395, + 0.905, + 0.448 + ], + "angle": 0, + "content": "\\[\n\\hat {A} _ {i, j} = \\left\\{ \\begin{array}{l l} \\exp (- \\| \\mathbf {x} _ {i} - \\mathbf {x} _ {j} \\| _ {2} / \\gamma) & \\| \\mathbf {x} _ {i} - \\mathbf {x} _ {j} \\| _ {2} \\leq \\mathcal {T} _ {i} \\\\ 0 & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.449, + 0.906, + 0.51 + ], + "angle": 0, + "content": "According from Eq. (7) to Eq. (11), the operation used to compute the final adjacent matrix can be easily parallelized by CUDA [18] and implemented on PyTorch [23]. The details of Cheb-GR are presented in Algorithm 1." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.522, + 0.646, + 0.539 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.547, + 0.806, + 0.562 + ], + "angle": 0, + "content": "5.1. Datasets and Evaluation Protocol" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.569, + 0.907, + 0.78 + ], + "angle": 0, + "content": "Datasets We conduct extensive experiments on three popular Person ReID datasets Market-1501 (Market) [39], Duke-MTMC-re-ID (Duke) [24] and CUHK03 [16]. We utilize the Duke dataset in order to make reasonable comparisons with previous work. Market1501 is a large-scale dataset consisting of 1501 identities captured by 6 cameras, featuring challenging variations in viewpoint, illumination, and occlusion. Duke is another large-scale dataset with 1404 identities captured by 8 cameras, offering more challenging scenarios due to the inclusion of pedestrian trajectories and multiple camera views. CUHK03 is a smaller dataset with 1360 identities, but it focuses on difficult cases of cross-view matching, as images are taken from two non-overlapping camera views." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Evaluation Metrics Performance is evaluated using the standard metrics of Cumulative Matching Characteristic (CMC), mean Average Precision (mAP), and the more recently proposed mean Inverse Negative Penalty (mINP)[35]. This combination of metrics provides a comprehensive assessment of retrieval effectiveness. To evaluate the computation costs, we measure the different methods five times and average the execution times." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19265" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.908, + 0.131 + ], + "angle": 0, + "content": "Table 1. Person ReID performance. We compare our methods with several re-ranking methods on Market1501, DukeMTMC, and CUHK03 datasets. mAP, Rank-1, and mINP are calculated to evaluate performance. Bolding indicates the lowest computation time, and red subscript text denotes the performance improvement compared to the baseline." + }, + { + "type": "table", + "bbox": [ + 0.113, + 0.137, + 0.885, + 0.445 + ], + "angle": 0, + "content": "
BaseMethodVenueMarketDukeCUHK03
mAP Rank-1 mINP time(s)mAP Rank-1 mINP time(s)mAP Rank-1 mINP time(s)
AGW [35]-TPAMI2188.695.366.3-79.989.845.4-66.467.355.6-
QE [6]ICCV0791.895.874.71.384.991.954.31.072.168.466.90.6
LBR [19]ICCV1991.795.576.63.984.691.856.02.171.269.664.31.0
KR [42]CVPR1794.895.989.189.890.291.775.696.078.576.176.013.4
ECN [25]CVPR1894.796.190.3161.190.792.478.2148.678.075.676.414.6
GCR [38]TMM2394.995.890.343.390.091.276.248.879.677.676.86.2
Cheb-GR-93.096.080.30.986.591.460.40.973.972.967.90.2
+4.4+0.7+14.0+6.6+1.6+15.0+7.5+5.6+12.3
TransReID [13]-ICCV2189.095.169.3-82.290.750.3-75.477.665.4-
QE [6]ICCV0791.995.377.01.787.392.859.41.680.879.075.30.6
LBR [19]ICCV1990.894.277.451.886.691.959.646.878.477.472.814.9
KR [42]CVPR1794.095.387.673.391.392.876.587.086.484.784.118.3
ECN [25]CVPR1894.095.788.7147.991.493.078.7153.786.084.484.133.6
GCR [38]TMM2392.593.886.155.590.791.775.664.384.083.781.08.0
Cheb-GR-93.295.283.51.590.493.071.21.585.483.983.80.5
+4.2+0.1+14.2+8.2+2.3+20.9+10.0+6.3+18.4
" + }, + { + "type": "image", + "bbox": [ + 0.107, + 0.459, + 0.36, + 0.605 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.14, + 0.61, + 0.33, + 0.622 + ], + "angle": 0, + "content": "(a) Rank-1 versus computation cost" + }, + { + "type": "image", + "bbox": [ + 0.369, + 0.459, + 0.625, + 0.605 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.408, + 0.61, + 0.586, + 0.622 + ], + "angle": 0, + "content": "(b) mAP versus computation cost" + }, + { + "type": "image", + "bbox": [ + 0.635, + 0.459, + 0.886, + 0.605 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.67, + 0.61, + 0.851, + 0.622 + ], + "angle": 0, + "content": "(c) mINP versus computation cost" + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.628, + 0.908, + 0.699 + ], + "angle": 0, + "content": "Figure 2. Comparison of computation costs versus performance improvement of different methods for person re-identification task on CUHK03 dataset. We compare our method with k-reciprocal (KR) [41], GCR [38], ECN [25], QE [6], LBR [19]. Our method requires the least computation costs among these methods and outperforms QE [6], and LBR [19]. Although our method underperforms k-reciprocal (KR) [41], GCR [38], and ECN [25], our methods only cost about one-fiftieth of their computation time. The direction of the arrow indicates that the indicator is optimal." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.712, + 0.486, + 0.894 + ], + "angle": 0, + "content": "Implementation Details The original features are extracted using the pre-trained models of AGW [35] and TransReID [13]. The values of the temperature parameter \\(\\gamma\\) and the numbers of iterations \\(T\\) are 0.2 and 3, respectively, which is the same as the setting of GCR [38]. Given the fact that there is a substantial discrepancy in the feature spaces produced by different models across various datasets, we adjust the value of parameter \\(\\lambda\\) in Eq. (7) based on the specific model and dataset used. \\(\\lambda\\) is set to 6.0/5.3/6.0/4.8/4.0 for AGW on Market/Duke/CUHK03/Market2Duke/Market2CUHK03 respectively, and 5.3/4.3/4.2 for TransReID on Mar" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.712, + 0.907, + 0.758 + ], + "angle": 0, + "content": "ket/Duke/CUHK03 respectively. All the re-ranking methods except ECN run on GPU since running ECN on GPU causes CUDA to run out of memory." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.772, + 0.854, + 0.788 + ], + "angle": 0, + "content": "5.2. Comparison with State-of-Art Methods" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.909, + 0.903 + ], + "angle": 0, + "content": "We compare our methods with k-reciprocal (KR) [41], GCR [38], ECN [25], QE [6], LBR [19] on three Person ReID datasets. We run each method five times on each dataset and calculate the average computation time to measure the computation costs. The comparison result is shown in Tab. 1. Fig. 2 uses a scatter plot to visualize the trade-offs between performance metrics and" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19266" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.12, + 0.094, + 0.37, + 0.252 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.373, + 0.094, + 0.625, + 0.252 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.627, + 0.095, + 0.876, + 0.251 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.251, + 0.263, + 0.744, + 0.278 + ], + "angle": 0, + "content": "Figure 3. The impact of parameter \\(\\lambda\\) on Person ReID on the CUHK03 dataset." + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.29, + 0.483, + 0.373 + ], + "angle": 0, + "content": "Table 2. Results on cross-domain settings. Market2Duke denotes training on the Market1501 dataset and testing on the DukeMTMC dataset, and Market2CUHK03 denotes training on the Market1501 dataset and testing on the CUHK03 dataset. Bolding indicates the least computation time, and red subscript text denotes the performance improvement compared to the baseline." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.384, + 0.481, + 0.541 + ], + "angle": 0, + "content": "
BaseMethodMarket2DukeMarket2CUHK03
mAP Rank-1time(s)mAP Rank-1time(s)
AGW [35]-24.338.9-8.68.2-
QE [6]28.943.40.98.67.80.6
LBR [19]31.449.12.311.111.11.1
KR [42]35.545.295.411.69.513.2
ECN [25]35.746.1152.911.18.814.7
GCR [38]31.941.746.810.28.15.8
Cheb-GR33.045.30.910.48.60.2
+8.7+6.4+1.8+0.4
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.569, + 0.483, + 0.901 + ], + "angle": 0, + "content": "computational costs on the CUHK03 dataset. As we can see, our proposed method, Cheb-GR, consistently demonstrates significantly lower computational costs compared to previous approaches, achieving a runtime that is less than the fastest existing method, regardless of the dataset or baseline model used. Notably, on the Market1501 dataset, Cheb-GR achieves a runtime of 0.9 seconds, significantly faster than the 43.3 seconds required by AGW + GCR [38] and runs on average about 50 times faster than widely used k-reciprocal re-ranking. This trend persists across the DukeMTMC and CUHK03 datasets, highlighting the method's efficiency in reducing computational overhead. Furthermore, our method achieves competitive performance, surpassing methods such as QE [6] and LBR [19] in most scenarios and even demonstrating performance comparable to or exceeding GCR [38] in some cases. For instance, on the Market1501 dataset, TransReID + Cheb-GR attains an mAP of \\(93.2\\%\\) , surpassing TransReID \\(^+\\) QE \\((91.9\\%)\\) ,TransReID \\(^+\\) LBR \\((90.8\\%)\\) ,and TransReID \\(^+\\) GCR \\((92.5\\%)\\) while also approaching the performance of TransReID \\(^+\\) KR and TransReID \\(^+\\) ECN \\((94.0\\%)\\) . Similar observations can be made on the DukeMTMC and" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.292, + 0.907, + 0.399 + ], + "angle": 0, + "content": "CUHK03 datasets, demonstrating the method's ability to maintain strong performance while achieving substantial computational savings. Therefore, our method offers a compelling solution for person ReID by balancing strong performance with significantly reduced computational cost, making it a practical and efficient choice for real-world applications where both accuracy and speed are paramount." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.408, + 0.845, + 0.424 + ], + "angle": 0, + "content": "5.3. Comparison on Cross-domain Settings" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.43, + 0.907, + 0.732 + ], + "angle": 0, + "content": "To investigate the effectiveness of various re-ranking methods in cross-domain scenarios, we fine-tune a pre-trained model of AGW [35] on the Market1501 dataset and evaluate its performance on the DukeMTMC and CUHK03 dataset. We apply different re-rank methods in these cross-domain settings, and we adjust the value of k in GCR [38] from its original setting to 7, as the original setting resulted in suboptimal performance in cross-domain settings; the result is shown in Tab. 2. We can observe that performance improvement achieved by all re-ranking methods in cross-domain settings is reduced compared to the general case, highlighting the inherent intractability of the cross-domain generalization. As shown in Tab. 2, Cheb-GR exhibits competitive performance with the least computation time, showcasing its versatility across different domains. Furthermore, Cheb-GR obtains superior performance than GCR [38], demonstrating a divergence from the general trend observed, which might suggest that the adaptive neighbor search might hold a potential advantage over the k-nearest neighbor search in certain contexts." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.743, + 0.702, + 0.759 + ], + "angle": 0, + "content": "5.4. Parameter Analysis" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.906, + 0.901 + ], + "angle": 0, + "content": "To analyze the impact of the hyper-parameter \\(\\lambda\\), we conduct experiments on the CUHK03 dataset. According to the theoretical analysis in Sec. 4.1, a sufficiently large value for \\(\\lambda\\) is necessary to suppress negative samples effectively. However, an excessively large \\(\\lambda\\) risks eliminating positive samples, hindering feature propagation. In extreme circumstances, if \\(\\lambda\\) is so excessively large that it leads to the elimination of all samples, the resulting graph structure collapses into isolated nodes, thereby nullifying the re-ranking effect." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "19267" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.129, + 0.097, + 0.375, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.377, + 0.097, + 0.621, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.626, + 0.097, + 0.871, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.145, + 0.261, + 0.851, + 0.276 + ], + "angle": 0, + "content": "Figure 4. The impact of parameter \\(\\gamma\\) and \\(T\\) on Person ReID on the CUHK03 dataset under different values of \\(\\lambda\\)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.296, + 0.486, + 0.568 + ], + "angle": 0, + "content": "Therefore, an optimal \\(\\lambda\\) value must strike a balance between reducing noise and facilitating feature propagation. As shown in Fig. 3, our method surpasses the baseline with various values of the parameter \\(\\lambda\\) demonstrating its competitive performance, and as the value of \\(\\lambda\\) increases, mAP, mINP, and Rank-1 accuracy generally exhibit an initial rise, culminating in a peak value, followed by a subsequent decline, matching the theoretical analysis we propose. Moreover, the two baseline models exhibit distinct sensitivities to variations in \\(\\lambda\\) and differing optimal \\(\\lambda\\) values for peak performance, highlighting the disparity between the feature spaces learned by the CNN-based and transformer-based models, the transformer-based model demonstrates a considerably greater sensitivity to variations in \\(\\lambda\\) compared to the more robust CNN-based model. This suggests that the inherent characteristics of CNN-based feature spaces may contribute to enhanced robustness when employed within GCN-based re-ranking methods." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.57, + 0.485, + 0.902 + ], + "angle": 0, + "content": "We analyze the impact of parameters \\(\\gamma\\) and \\(T\\) under three circumstances: \\(\\lambda\\) is optimal (\\(\\lambda = 4.2\\)), \\(\\lambda\\) is insufficiently large (\\(\\lambda = 3.0\\)), and \\(\\lambda\\) is excessively large (\\(\\lambda = 6.0\\)). \\(T\\) denotes the number of iterations and \\(T = 0\\) denotes the baseline. In each iteration, the graph convolution operation smoothes the features of connected nodes, drawing them together in the embedding space. As shown in Fig. 4, in most cases, with the increase of \\(T\\) from 0 to 5, mAP, Rank-1, and mINP first rise and then decline. This occurs because a small value of \\(T\\) leads to insufficient smoothing, limiting the effectiveness of feature re-ranking. In contrast, a large value of \\(T\\) causes excessive smoothing, which may result in falsely connected features appearing closer together and separating features of unconnected positive samples. \\(\\gamma\\) is a temperature parameter controlling the sharpness of the distribution of edges' weight. From Fig. 4, we can observe that smaller \\(\\gamma\\) performs better when \\(\\lambda\\) is small, and larger \\(\\gamma\\) performs better when \\(\\lambda\\) is large. Thus, a smaller \\(\\gamma\\) improves performance when neighbors are redundant and noisy by facilitating the propagation of features to more similar neighbors, and conversely, a larger \\(\\gamma\\) proves advantageous when neighbors are limited, as it allows more features to spread" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.296, + 0.694, + 0.312 + ], + "angle": 0, + "content": "to neighbors more equally." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.32, + 0.771, + 0.336 + ], + "angle": 0, + "content": "6. Conclusion and Discussions" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.342, + 0.907, + 0.674 + ], + "angle": 0, + "content": "In this paper, we introduce Cheb-GR, a novel and efficient GCN-based re-ranking method for person re-identification (ReID) that addresses the critical need for both accuracy and efficiency in this challenging task. We propose a novel adaptive neighbor search for neighbor selection, replacing the computationally expensive k-nearest neighbor search and enabling significant parallelization, which is particularly well-suited for GPU acceleration. The effectiveness of Cheb-GR is rigorously validated through extensive experiments on standard benchmarks and challenging cross-domain scenarios, consistently demonstrating its competitive performance. Furthermore, we provide a theoretical analysis of the underlying principles that govern the method's efficiency and accuracy, shedding light on its ability to distinguish between similar individuals and refine the ranking process effectively. Cheb-GR's high performance and efficiency pave the way for real-time ReID applications, particularly in scenarios where speed is paramount. Beyond its demonstrated efficacy in ReID, Cheb-GR's efficient graph construction and feature refinement capabilities hold promise for broader applicability to other similarity-based retrieval tasks, suggesting its potential." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.909, + 0.795 + ], + "angle": 0, + "content": "Limitations Cheb-GR's adaptive neighbor search, while efficient, is sensitive to the trade-off parameter \\(\\lambda\\), which balances removing negative and preserving positive edges. Careful tuning is crucial: an overly small \\(\\lambda\\) hinders re-ranking, while an excessively large \\(\\lambda\\) removes beneficial edges. Moreover, quadratic memory scaling of the adjacency matrix limits scalability, and performance depends on using features from triplet loss (or similar) trained models." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Acknowledgement This work is partially supported by the National Key Research and Development Program of China (2024YFC3308400), National Natural Science Foundation of China under Grant (62176188, 62225113). The numerical calculations in this paper were supported by the supercomputing system at the Supercomputing Center of Wuhan University." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19268" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.091, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.116, + 0.484, + 0.17 + ], + "angle": 0, + "content": "[1] Yang Bai, Yucheng Ji, Min Cao, Jinqiao Wang, and Mang Ye. Chat-based person retrieval via dialogue-refined cross-modal alignment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.17, + 0.484, + 0.237 + ], + "angle": 0, + "content": "[2] Arko Barman and Shishir K Shah. Shape: A novel graph theoretic algorithm for making consensus-based decisions in person re-identification systems. In Proceedings of the IEEE International Conference on Computer Vision, pages 1115–1124, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.238, + 0.482, + 0.264 + ], + "angle": 0, + "content": "[3] Pafnutii Lvovich Chebyshev. Des valeurs moyennes. J. Math. Pures Appl, 12(2):177-184, 1867. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.265, + 0.483, + 0.319 + ], + "angle": 0, + "content": "[4] Cuiqun Chen, Mang Ye, and Ding Jiang. Towards modality-agnostic person re-identification with descriptive query. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15128-15137, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.319, + 0.482, + 0.373 + ], + "angle": 0, + "content": "[5] Yanbei Chen, Xiatian Zhu, and Shaogang Gong. Person re-identification by deep learning multi-scale representations. In Proceedings of the IEEE international conference on computer vision workshops, pages 2590–2600, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.373, + 0.482, + 0.44 + ], + "angle": 0, + "content": "[6] Ondrej Chum, James Philbin, Josef Sivic, Michael Isard, and Andrew Zisserman. Total recall: Automatic query expansion with a generative feature model for object retrieval. In 2007 IEEE 11th International Conference on Computer Vision, pages 1-8. IEEE, 2007. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.441, + 0.482, + 0.508 + ], + "angle": 0, + "content": "[7] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4690-4699, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.509, + 0.482, + 0.548 + ], + "angle": 0, + "content": "[8] Bo Du, Fangyi Liu, and YE Mang. Person re-identification method based on perspective-guided multi-adversarial attention, Oct. 31 2023. US Patent 11,804,036. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.549, + 0.482, + 0.603 + ], + "angle": 0, + "content": "[9] Xing Fan, Wei Jiang, Hao Luo, and Mengjuan Fei. Spherereid: Deep hypersphere manifold embedding for person re-identification. Journal of Visual Communication and Image Representation, 60:51-58, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.603, + 0.482, + 0.67 + ], + "angle": 0, + "content": "[10] Niloofar Gheissari, Thomas B Sebastian, and Richard Hartley. Person reidentification using spatiotemporal appearance. In 2006 IEEE computer society conference on computer vision and pattern recognition (CVPR'06), volume 2, pages 1528-1535. IEEE, 2006. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.671, + 0.482, + 0.725 + ], + "angle": 0, + "content": "[11] Justin Gilmer, Samuel S Schoenholz, Patrick F Riley, Oriol Vinyals, and George E Dahl. Neural message passing for quantum chemistry. In International conference on machine learning, pages 1263-1272. PMLR, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.726, + 0.482, + 0.792 + ], + "angle": 0, + "content": "[12] Raia Hadsell, Sumit Chopra, and Yann LeCun. Dimensionality reduction by learning an invariant mapping. In 2006 IEEE computer society conference on computer vision and pattern recognition (CVPR'06), volume 2, pages 1735-1742. IEEE, 2006. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.793, + 0.482, + 0.859 + ], + "angle": 0, + "content": "[13] Shuting He, Hao Luo, Pichao Wang, Fan Wang, Hao Li, and Wei Jiang. Transreid: Transformer-based object re-identification. In Proceedings of the IEEE/CVF international conference on computer vision, pages 15013-15022, 2021. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.86, + 0.482, + 0.889 + ], + "angle": 0, + "content": "[14] Bo Jiang, Xixi Wang, Aihua Zheng, Jin Tang, and Bin Luo. Ph-gen: Person retrieval with part-based hierarchical graph" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.116, + 0.484, + 0.889 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.548, + 0.094, + 0.905, + 0.119 + ], + "angle": 0, + "content": "convolutional network. IEEE Transactions on Multimedia, 24:3218-3228, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.12, + 0.905, + 0.159 + ], + "angle": 0, + "content": "[15] Thomas N Kipf and Max Welling. Semi-supervised classification with graph convolutional networks. arXiv preprint arXiv:1609.02907, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.16, + 0.905, + 0.227 + ], + "angle": 0, + "content": "[16] Wei Li, Rui Zhao, Tong Xiao, and Xiaogang Wang. Deepreid: Deep filter pairing neural network for person re-identification. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 152–159, 2014. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.228, + 0.905, + 0.269 + ], + "angle": 0, + "content": "[17] Fangyi Liu, Mang Ye, and Bo Du. Learning a generalizable re-identification model from unlabelled data with domain-agnostic expert. Visual Intelligence, 2(1):28, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.269, + 0.905, + 0.323 + ], + "angle": 0, + "content": "[18] David Luebke. Cuda: Scalable parallel programming for high-performance scientific computing. In 2008 5th IEEE international symposium on biomedical imaging: from nano to macro, pages 836-838. IEEE, 2008. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.323, + 0.905, + 0.389 + ], + "angle": 0, + "content": "[19] Chuanchen Luo, Yuntao Chen, Naiyan Wang, and Zhaoxiang Zhang. Spectral feature transformation for person re-identification. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4976–4985, 2019. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.39, + 0.905, + 0.459 + ], + "angle": 0, + "content": "[20] Hao Luo, Youzhi Gu, Xingyu Liao, Shenqi Lai, and Wei Jiang. Bag of tricks and a strong baseline for deep person re-identification. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 0–0, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.459, + 0.905, + 0.525 + ], + "angle": 0, + "content": "[21] Jiaxu Miao, Yu Wu, Ping Liu, Yuhang Ding, and Yi Yang. Pose-guided feature alignment for occluded person re-identification. In Proceedings of the IEEE/CVF international conference on computer vision, pages 542–551, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.526, + 0.905, + 0.593 + ], + "angle": 0, + "content": "[22] Sakrapee Paisitkriangkrai, Chunhua Shen, and Anton Van Den Hengel. Learning to rank in person re-identification with metric ensembles. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1846–1855, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.594, + 0.905, + 0.675 + ], + "angle": 0, + "content": "[23] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.676, + 0.905, + 0.731 + ], + "angle": 0, + "content": "[24] Ergys Ristani, Francesco Solera, Roger Zou, Rita Cucchiara, and Carlo Tomasi. Performance measures and a data set for multi-target, multi-camera tracking. In European conference on computer vision, pages 17-35. Springer, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.731, + 0.905, + 0.811 + ], + "angle": 0, + "content": "[25] M Saquib Sarfraz, Arne Schumann, Andreas Eberle, and Rainer Stiefelhagen. A pose-sensitive embedding for person re-identification with expanded cross neighborhood reranking. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 420–429, 2018. 1, 2, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.812, + 0.905, + 0.867 + ], + "angle": 0, + "content": "[26] Florian Schroff, Dmitry Kalenichenko, and James Philbin. Facenet: A unified embedding for face recognition and clustering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 815-823, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.867, + 0.905, + 0.893 + ], + "angle": 0, + "content": "[27] Yantao Shen, Hongsheng Li, Shuai Yi, Dapeng Chen, and Xiaogang Wang. Person re-identification with deep" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.094, + 0.905, + 0.893 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19269" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.484, + 0.133 + ], + "angle": 0, + "content": "similarity-guided graph neural network. In Proceedings of the European conference on computer vision (ECCV), pages 486-504, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.134, + 0.483, + 0.202 + ], + "angle": 0, + "content": "[28] Yifan Sun, Changmao Cheng, Yuhan Zhang, Chi Zhang, Liang Zheng, Zhongdao Wang, and Yichen Wei. Circle loss: A unified perspective of pair similarity optimization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6398-6407, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.203, + 0.482, + 0.282 + ], + "angle": 0, + "content": "[29] Yifan Sun, Qin Xu, Yali Li, Chi Zhang, Yikang Li, Shengjin Wang, and Jian Sun. Perceive where to focus: Learning visibility-aware part-level features for partial person re-identification. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 393–402, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.284, + 0.482, + 0.351 + ], + "angle": 0, + "content": "[30] Yifan Sun, Liang Zheng, Yi Yang, Qi Tian, and Shengjin Wang. Beyond part models: Person retrieval with refined part pooling (and a strong convolutional baseline). In Proceedings of the European conference on computer vision (ECCV), pages 480-496, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.352, + 0.482, + 0.419 + ], + "angle": 0, + "content": "[31] Haochen Wang, Jiayi Shen, Yongtuo Liu, Yan Gao, and Efstratos Gavves. Nformer: Robust person re-identification with neighbor transformer. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7297-7307, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.42, + 0.482, + 0.473 + ], + "angle": 0, + "content": "[32] Yingjia Xu, Mengxia Wu, Zixin Guo, Min Cao, Mang Ye, and Jorma Laaksonen. Efficient text-to-video retrieval via multi-modal multi-tags derived pre-screening. Visual Intelligence, 3(1):1-13, 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.474, + 0.482, + 0.528 + ], + "angle": 0, + "content": "[33] Mang Ye, Shuoyi Chen, Chenyue Li, Wei-Shi Zheng, David Crandall, and Bo Du. Transformer for object re-identification: A survey. International Journal of Computer Vision, pages 1–31, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.528, + 0.482, + 0.594 + ], + "angle": 0, + "content": "[34] Mang Ye, Chao Liang, Zheng Wang, Qingming Leng, and Jun Chen. Ranking optimization for person re-identification via similarity and dissimilarity. In Proceedings of the 23rd ACM international conference on Multimedia, pages 1239–1242, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.596, + 0.482, + 0.663 + ], + "angle": 0, + "content": "[35] Mang Ye, Jianbing Shen, Gaojie Lin, Tao Xiang, Ling Shao, and Steven CH Hoi. Deep learning for person re-identification: A survey and outlook. IEEE transactions on pattern analysis and machine intelligence, 44(6):2872-2893, 2021. 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.664, + 0.482, + 0.718 + ], + "angle": 0, + "content": "[36] Mang Ye, Zesen Wu, and Bo Du. Dual-level matching with outlier filtering for unsupervised visible-infrared person re-identification. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.718, + 0.482, + 0.771 + ], + "angle": 0, + "content": "[37] Xuanmeng Zhang, Minyue Jiang, Zhedong Zheng, Xiao Tan, Errui Ding, and Yi Yang. Understanding image retrieval reranking: A graph neural network perspective. arXiv preprint arXiv:2012.07620, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.772, + 0.482, + 0.825 + ], + "angle": 0, + "content": "[38] Yuqi Zhang, Qi Qian, Hongsong Wang, Chong Liu, Weihua Chen, and Fan Wang. Graph convolution based efficient reranking for visual retrieval. IEEE Transactions on Multimedia, 26:1089-1101, 2023. 1, 2, 3, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.826, + 0.482, + 0.88 + ], + "angle": 0, + "content": "[39] Liang Zheng, Liyue Shen, Lu Tian, Shengjin Wang, Jingdong Wang, and Qi Tian. Scalable person re-identification: A benchmark. In Proceedings of the IEEE international conference on computer vision, pages 1116–1124, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.88, + 0.482, + 0.893 + ], + "angle": 0, + "content": "[40] Liang Zheng, Yi Yang, and Alexander G Hauptmann. Per" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.893 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.093, + 0.905, + 0.119 + ], + "angle": 0, + "content": "son re-identification: Past, present and future. arXiv preprint arXiv:1610.02984, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.121, + 0.905, + 0.187 + ], + "angle": 0, + "content": "[41] Meng Zheng, Srikrishna Karanam, Ziyan Wu, and Richard J Radke. Re-identification with consistent attentive siamese networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5735–5744, 2019. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.188, + 0.905, + 0.254 + ], + "angle": 0, + "content": "[42] Zhun Zhong, Liang Zheng, Donglin Cao, and Shaozi Li. Reranking person re-identification with k-reciprocal encoding. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1318–1327, 2017. 1, 2, 4, 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.254 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19270" + } + ] +] \ No newline at end of file diff --git a/2025/Cheb-GR_ Rethinking K-nearest Neighbor Search in Re-ranking for Person Re-identification/9e933435-e0fc-412e-b745-089375b60f23_origin.pdf b/2025/Cheb-GR_ Rethinking K-nearest Neighbor Search in Re-ranking for Person Re-identification/9e933435-e0fc-412e-b745-089375b60f23_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..dbaba938537724a4618205d27a28677917d7cc1d --- /dev/null +++ b/2025/Cheb-GR_ Rethinking K-nearest Neighbor Search in Re-ranking for Person Re-identification/9e933435-e0fc-412e-b745-089375b60f23_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d7dc2c049c88cb1448065cc17dfe1ae16f51796c82e12a620b6f32000e25d57b +size 2064957 diff --git a/2025/Cheb-GR_ Rethinking K-nearest Neighbor Search in Re-ranking for Person Re-identification/full.md b/2025/Cheb-GR_ Rethinking K-nearest Neighbor Search in Re-ranking for Person Re-identification/full.md new file mode 100644 index 0000000000000000000000000000000000000000..35f17491518771e7c61a5c738a85bf08c496efb9 --- /dev/null +++ b/2025/Cheb-GR_ Rethinking K-nearest Neighbor Search in Re-ranking for Person Re-identification/full.md @@ -0,0 +1,315 @@ +# Cheb-GR: Rethinking k-nearest neighbor search in Re-ranking for Person Re-identification + +Jinxi Yang * He Li * Bo Du Mang Ye† +National Engineering Research Center for Multimedia Software, School of Computer Science, Wuhan University, Wuhan, China +{jinxiyang, lihe404, dubo, yemang}@whu.edu.cn +https://github.com/Jinxi-Yang-WHU/Fast-GCR.git + +# Abstract + +Person re-identification (ReID) is the task of matching individuals across different camera views. Existing approaches typically employ neural networks to extract discriminative features, ranking gallery images based on their similarities to probe images. While effective, these methods are often enhanced through re-ranking, a post-processing step that refines initial retrieval results without requiring additional model training. However, current re-ranking methods mostly rely on $k$ -nearest neighbor search to extract similar images that might have the same identity as the query, which is time-consuming with a high computation burden, limiting their applications in reality. We rethink the effect of the $k$ -nearest neighbor search and introduce the Chebyshev's Theorem-guided Graph Re-ranking (ChebGR) method, which adopts the adaptive neighbor search guided by Chebyshev's Theorem over the $k$ -nearest neighbor search for efficient neighbor selection. Our method leverages graph convolution operations to refine image features and achieve robust re-ranking, leading to enhanced retrieval performance. Furthermore, we provide a theoretical analysis based on Chebyshev's Inequality to elucidate the factors contributing to the strong performance of the proposed method. Our method significantly reduces the computation costs while maintaining relatively strong performance. Through extensive experiments in both general and cross-domain settings, we demonstrate the effectiveness of Cheb-GR and its potential for real-world applications. + +# 1. Introduction + +Person re-identification (ReID) is a computer vision task that aims to identify the same person across different camera views in a multi-camera surveillance system [5, 40]. It + +involves matching individuals captured by different cameras based on their appearance, even when they are viewed from different angles or under varying lighting conditions [10]. Deep learning has become the predominant approach in ReID, enabling automatic feature extraction and significantly enhancing performance [17, 21, 29, 30, 32, 33, 36, 41]. However, these methods primarily rely on pairwise similarity comparisons between the query image and individual gallery images, neglecting the contextual relationships among images within the ranking list. Consequently, they are susceptible to noise in feature representations arising from variations in illumination, pose, or occlusion. + +To tackle this issue, many re-ranking methods have emerged as a prominent strategy for enhancing performance without the need for additional model training. Prevalent approaches involve utilizing k-nearest neighbor search to identify similar images, thereby extracting contextual information [25, 42]. However, this reliance on k-nearest neighbor search presents a significant bottleneck in real-world applications due to its computational complexity. Meanwhile, many works utilize Graph Convolutional Networks (GCN) to learn pairwise similarity information and relational information between images in person ReID tasks [1, 14, 27]. Notably, Zhang et al. [38] demonstrate the effectiveness of GCNs for re-ranking by refining embeddings through feature propagation. Despite their potential, these GCN-based methods still necessitate k-nearest neighbor search, which is inherently sequential and hinders parallel computation, leading to substantial time consumption. + +To improve re-rank efficiency, we propose the Chebyshev's Theorem-guided Graph Re-ranking (Cheb-GR) method that eliminates the need for the k-nearest neighbor search and can be executed in parallel. Specifically, we construct a graph where each node represents an image. Unlike GCR [38], which connects each node to its k-nearest neighbors, Cheb-GR uses feature distances as edge weights, discarding edges with extreme distances. The remaining dis + +tances are then converted into similarities. This approach directly constrains feature propagation to similar images, a key factor in achieving strong performance. A notable advantage of our method is that all operations are straightforward and highly parallelizable, resulting in significantly reduced computation costs and much faster execution compared to existing methods [25, 38, 42]. Additionally, ChebGR exhibits robust performance in cross-domain scenarios, further demonstrating the versatility of our method. We also provide a theoretical analysis to explain why the proposed approach performs effectively without relying on k-nearest neighbor search, which is a method commonly used by various re-ranking methods. + +To the best of our knowledge, Cheb-GR is the fastest re-ranking method currently available, delivering competitive performance. In summary, our contributions are threefold: + +- We propose Chebyshev's Theorem-guided Graph Re-ranking (Cheb-GR) method with no need for finding k-nearest neighbors, which runs faster than all re-rank methods we know. Besides, Cheb-GR achieves stronger performance on cross-domain settings than most existing methods. +- A reasonable theoretical explanation of the proposed method is presented to reveal why Cheb-GR can obtain strong performance without k-nearest neighbor search. +- Extensive experimentation across a diverse range of datasets provides compelling evidence of our method's strong performance and low computation costs. + +# 2. Related Work + +Person Re-identification. Recent advancements in artificial intelligence have propelled person re-identification (Re-ID), with some methods even surpassing human performance. Current approaches can be broadly divided into two categories: the first focuses on designing metric learning loss functions to learn discriminative features, while the second modifies model architectures to capture contextual information, improving feature representation. + +For model architecture modification, many works have been proposed [4, 8, 31, 33, 35]. Sun et al. [30] propose a part-based convolutional baseline (PCB), utilizing uniform partitioning to capture local image information and employing a refined part pooling (RPP) method to enhance within-part consistency. Luo et al. [20] propose BoT, a strong baseline that adds a BNNeck structure after ResNet50. BNNeck introduces a batch normalization layer to separate ID and triplet loss optimization, improving convergence and performance in person ReID. He et al. [13] introduced TransReID, the first transformer-based model for person re-identification to achieve performance comparable to convolutional neural networks. For metric learning loss design, Softmax loss or its modification [7, 9] have been widely used in person Re-ID for feature learning. Loss functions + +utilized in metric learning, such as triplet loss [26] and contrastive loss [12], are also applied to the person Re-ID task. Some works also design novel loss functions to guide the model to learn distinguishable representations [28, 35]. Sun et al. [28] re-weight each similarity score to obtain flexible optimization and a definite convergence target. Ye et al. [35] propose Weighted Regularization Triplet loss to take full advantage of relative distance optimization between positive and negative pairs without introducing additional margin parameters. + +All these works are committed to extracting distinguishable features of images and then calculating the similarity between the feature of a query image and the feature of each gallery image to generate a ranking list. However, these works ignore abundant contextual information within the ranking list. + +Re-ranking for Person Re-identification. Most existing methods for Re-ID concentrate on extracting distinctive image features and assessing the similarity between the features of the query image and those of the gallery images. Additionally, a significant number of researchers have directed their attention towards re-ranking methods within ReID tasks, aiming to harness gallery-to-gallery similarity [25, 34, 42] or to integrate various ReID approaches [2, 22]. While traditional re-ranking methods have primarily emphasized the design of innovative similarity metrics, a recent advancement involves employing the graph convolutional network (GCN) to update features of images directly, which is more efficient and scalable [38]. + +Most re-ranking methods design novel distance metrics to consider global contextual information or directly refine the features of images to improve performance [25, 37, 38, 42]. Zhong et al. [42] propose a k-reciprocal encoding method that leverages contextual information between k-reciprocal nearest neighbors to improve performance. Zhang et al. [38] propose a feature propagation-based method, constructing a graph by connecting each image with its k-nearest neighbors and propagating features by graph convolution on the graph. + +However, existing methods rely on computationally expensive k-nearest neighbor searches to identify similar images, limiting their practical applicability, particularly for large-scale datasets. So, we propose a method that circumvents the need to find k-nearest neighbors, significantly reducing computation costs. + +Graph Convolutional Network in ReID. Graph Convolutional Network (GCN) [11] utilizes a message-passing mechanism to allow nodes to exchange and aggregate information with neighbors. Through a convolutional operation, each node's representation is updated by a weighted sum of its features and those of its neighbors, computed via matrix multiplication with a learned weight matrix [15]. This iterative process allows GCNs to learn rich representations that + +capture both local and global information within the graph. + +Since graphs are a great data structure for describing relationships and GCNs can learn graph-structure information, many ReID methods utilize GCNs to learn pairwise similarity and relational information among image sets. Jiang et al. [14] propose a method based on a hierarchical graph to integrate local, global, and structural information. Zhang et al. [37] and GCR [38] show that GCN can be leveraged in the re-ranking area. This paper employs a simplified GCN that achieves feature propagation without a learned weight matrix. This parameter-free approach enhances efficiency by eliminating the need for training. + +# 3. Preliminary and Motivation + +Efficiency of K-nearest Neighbor Search. Current person ReID re-ranking methods rely heavily on k-nearest neighbor search, a key operation used to select potential same-identity images. K-nearest neighbor search computes the distances between the query image and all other images in the gallery to identify the k-nearest neighbors, which is computationally expensive and presents significant challenges when parallelizing on GPUs. The time complexity of sorting N images is $\Theta (N\log N)$ . In order to select the k-nearest neighbors for each node, we need to perform the sorting operations $N$ times and select the top-k images each time. Thus, the total time complexity of the k-nearest neighbor search is $\Theta (N^2\log N)$ , which becomes the primary computational bottleneck in the re-ranking process. + +Graph based Re-ranking. Zhang et al. [38] propose Graph Convolutional-based Re-ranking (GCR). In GCR, given a set of images, a graph is constructed that encompasses both query and gallery images. Each query and gallery image corresponds to a node in the graph, with its associated feature vector serving as the initial node feature. Edges are established between nodes representing visually similar images, encoding potential identity relationships within the dataset. This constructed graph then serves as the input to a GCN, which performs iterative message-passing operations to refine the node features. The message-passing operation in a GCN layer can be written as: + +$$ +\mathbf {X} ^ {(l)} = \mathbf {D} ^ {- \frac {1}{2}} \mathbf {A D} ^ {- \frac {1}{2}} \mathbf {X} ^ {(l - 1)} \mathbf {W}, \tag {1} +$$ + +where $\mathbf{X}^{(l)}\in \mathbb{R}^{N\times d}$ is the node features in the $l^{th}$ layer, $N$ denotes the number of nodes, $d$ denotes the feature dimensionality. $\mathbf{A}\in \mathbb{R}^{N\times N}$ is the adjacent matrix of the graph, $\mathbf{D}\in \mathbb{R}^{N\times N}$ is the degree matrix of the graph which is a diagonal matrix and $\mathbf{D}_{i,i} = \sum_{j}\mathbf{A}_{i,j}$ . $\mathbf{W}\in \mathbb{R}^{d\times d'}$ denotes a learnable weighted matrix in the GCN layer, which can project the features vector into a certain space. + +When the weighted matrix $\mathbf{W}$ is an identity matrix, the graph convolution operation can be written as: + +$$ +\mathbf {X} ^ {(l)} = \mathbf {D} ^ {- \frac {1}{2}} \mathbf {A} \mathbf {D} ^ {- \frac {1}{2}} \mathbf {X} ^ {(l - 1)}. \tag {2} +$$ + +As formalized in Eq. (2), this parameter-free operation iteratively updates each node's feature embedding by aggregating the features of its neighbors and itself. The weight assigned to each neighbor's contribution is determined by the corresponding edge weight in the graph. Ideally, Graph Convolution-based Re-ranking enhances feature representations by promoting propagation between images of the same identity. This process brings together features of matching identities while pushing apart features of distinct identities, leading to improved re-ranking performance. Therefore, the primary challenge lies in constructing a graph that maximizes connections between images of the same identity while minimizing connections between images of different identities. As illustrated in Fig. 1(b), GCR employs a k-nearest neighbor search to identify neighboring nodes. According to Eq. (2), the time complexity of a graph convolution operation in GCR [38] is equal to matrix multiplication, which is $\Theta(N^3)$ . However, since the matrix multiplication is highly parallelized on GPUs, it runs much faster than the k-nearest neighbor search in practice. Thus the efficiency bottleneck of GCR is still mainly affected by k-nearest neighbor search. + +# 4. Proposed Method + +In this section, we first revisit the neighbor search and introduce an efficient strategy to replace the k-nearest neighbor search in Sec. 4.1. After that, we introduce our method in detail in Sec. 4.2. + +# 4.1. Revisiting Neighbor Search + +Since the role of the k-nearest neighbor search in re-ranking is to select potential positive samples, we need to find an alternative strategy to find potential positive samples with low time complexity to lower the computation costs. Since feature distance is inversely related to the probability of shared identity, a threshold can be established based on the distance between the features of an anchor image and other images. Images with a distance below this threshold are then considered positive samples, sharing the same identity as the anchor image. + +The determination of an appropriate threshold for feature distance warrants further consideration. The triplet loss can be written as: + +$$ +\mathcal {L} = \max (d (a, p) - d (a, n) + \alpha , 0), \tag {3} +$$ + +where $\alpha$ is the margin hyperparameter. Eq. (3) implies that, ideally, the distance between images of different identities should exceed the distance between images of the same identity by at least a factor of $\alpha$ . Given the significant disparity in the number of positive and negative samples within the datasets, with far fewer positive samples, we can simplify our analysis by focusing on the distribution of the + +![](images/00a74549813f566b92ceb391b299b4ff9ec46b8ee8738b2fc5efcfdff8395385.jpg) + +![](images/2ea39a22ba03492d6e32fb733ff801034e95d0b8b624e30c5030c923e1648423.jpg) + +![](images/4a031eff605da7e3a3d47c8586770dc32d8db3f0f5430c49f8642f7881e95c44.jpg) +Figure 1. Comparison between traditional re-rank methods. GCR [38] and the proposed Chebyshev's Guided Graph Re-ranking (Cheb-GR). (a) Most traditional re-rank methods, such as k-reciprocal [42] and ECN [25], operate in three stages: neighborhood set construction, dissimilarity measurement between these sets, and weighted distance aggregation combining direct and neighborhood-based distances. (b) GCR connects each node to its k-nearest neighbors to construct the graph. GCNs are then employed to propagate the node features through the graph, refining the image representations for improved re-ranking. (c) Cheb-GR leverages an adaptive neighbor search instead of a k-nearest neighbor search to select neighbors and construct the graph. Subsequently, feature propagation is achieved through the application of a GCN. + +![](images/4833588ef2c759bfe2eae85ff9ce94f713f2a95190371833db9e1c497c45dc8b.jpg) + +distance between negative samples in the embedding space. Thus, the problem is reframed as one of identifying the minimum distance separating the negative samples. According to Chebyshev's Theorem [3]: + +$$ +P r [ | R | \geq x ] \leq \frac {E x [ | R | ^ {\alpha} ]}{x ^ {\alpha}}, +$$ + +$$ +P r [ | R - E x [ R ] | \geq x ] \leq \frac {E x [ | R - E x [ R ] | ^ {\alpha} ]}{x ^ {\alpha}}, \tag {4} +$$ + +$$ +P r [ | R - E x [ R ] | \geq x ] \leq \frac {V a r [ R ]}{x ^ {2}}, +$$ + +where $R$ denotes a random variable, $Ex[R]$ denotes the expected value of the random variable $R$ , $Var[R]$ denotes the variance of the random variable $R$ , $x$ denotes a real number. We can replace $x$ with $\lambda \sigma$ , where $\lambda$ is a constant hyperparameter and $\sigma$ is standard deviation, to get the inequality: + +$$ +P r [ | R - E x [ R ] | \geq \lambda \sigma ] \leq \frac {1}{\lambda^ {2}}. \tag {5} +$$ + +Then, removing the absolute value operation and ignoring + +the scenario that $R \geq Ex[R] + \lambda \sigma$ , we get the inequality: + +$$ +P r [ R \leq E x [ R ] - \lambda \sigma ] \leq \frac {1}{\lambda^ {2}}, \tag {6} +$$ + +which shows that only a small number of elements fall below the mean minus $\lambda$ times the standard deviation when $\lambda$ is sufficiently large. Thus, given an image $i$ , we can set threshold $\mathcal{T}_i$ for image $i$ as follows: + +$$ +\mathcal {T} _ {i} = \mu_ {i} - \lambda \sigma_ {i}, +$$ + +$$ +\mu_ {i} = \frac {1}{N} \sum_ {j = 1} ^ {N} D i s _ {i, j}, \tag {7} +$$ + +$$ +\sigma_ {i} = \sqrt {\frac {\sum_ {j = 1} ^ {N} (D i s _ {i , j} - \mu_ {i}) ^ {2}}{N}}, +$$ + +where $Dis_{i,j}$ denotes the distance between features of image $i$ and image $j$ , $N$ denotes the number of images in the dataset, $\lambda$ is a hyper-parameter, $\mu_i$ is the mean of distances between features of an image in the dataset and image $i$ , $\sigma_i$ is the standard deviations of distances between features of + +an image in the dataset and image $i$ . Given the presence of a margin $\alpha$ between the distances of positive and negative samples, there exists a specific value of $\lambda$ that ensures that the threshold $\mathcal{T}_i$ surpasses the distances between positive samples while remaining below the distances between negative samples. Therefore, by carefully selecting an appropriate value for $\lambda$ , we can effectively filter out negative samples while preserving positive samples within the graph construction process. The time complexity of calculating the mean and standard deviation of $N$ elements is $\Theta(N)$ , so the time complexity of calculating the threshold for $N$ images is $\Theta(N^2)$ , which is less than the time complexity of the k-nearest neighbor search. Moreover, our proposed method can be easily parallelized on GPUs and implemented on the PyTorch platform [23], so the computation cost of our method is much less than the computation costs of the k-nearest neighbor search in practice. + +# 4.2. Chebyshev's Guided Graph Re-ranking + +The pipeline of our method is shown in Fig. 1(c). We first construct a directed complete graph, and the weight of each edge is the distance between the nodes the edge connects. The adjacent matrix of this graph can be computed by: + +$$ +A _ {i, j} = \left\| \mathbf {x} _ {i} - \mathbf {x} _ {j} \right\| _ {2}, \tag {8} +$$ + +where $\| \cdot \|_2$ denotes Euclidean distance. Intuitively, the larger the distance between the features of two images, the less likely they have the same identity. To ensure that feature propagation is constrained to images of the same identity as far as possible, a threshold-based edge deletion strategy is employed. Edges with weights exceeding a predefined threshold, which is computed as Eq. (7), are removed from the graph's adjacency matrix. + +We delete the edges whose weights exceed the corresponding threshold and get the pruned graph whose adjacent matrix can be expressed by: + +$$ +\tilde {A} _ {i, j} = \left\{ \begin{array}{l l} \| \mathbf {x} _ {i} - \mathbf {x} _ {j} \| _ {2} & \| \mathbf {x} _ {i} - \mathbf {x} _ {j} \| _ {2} \leq \mathcal {T} _ {i} \\ \infty & \text {o t h e r w i s e ,} \end{array} \right. \tag {9} +$$ + +where threshold $\mathcal{T}_i$ is computed by Eq. (7). Setting the weight of deleted edges to infinity ensures that, after converting distance to similarity, these edges effectively become non-existent. This approach is justified by the interpretation that if two nodes are not connected, their distance is considered infinite. The desired outcome is that edge weights should positively correlate with the similarity of features between connected nodes. However, the adjacent matrix computed by Eq. (9) is just the opposite. So, we convert the distance to similarity for edge weight by: + +$$ +\operatorname {S i m} (i, j) = \exp (- \| \mathbf {x} _ {i} - \mathbf {x} _ {j} \| _ {2} / \gamma) \tag {10} +$$ + +where $Sim(i,j)$ denotes the similarity between image $i$ and image $j$ and $\gamma$ is the temperature hyper-parameter. The final + +# Algorithm 1 Chebyshev's Guided Graph Re-ranking + +Input: $\mathbf{X}_0$ - Original query and gallery image features + +Output: $\mathbf{X}_T$ - Final features after re-ranking + +1: $A_{i,j} = \parallel \mathbf{x}_i - \mathbf{x}_j\parallel_2$ . Calculate the original adjacent matrix $\mathbf{A}$ +2: $\mathcal{T}_i = \mu_i - \lambda \sigma_i$ . Calculate the threshold $\mathcal{T}_i$ . +3: $\tilde{A}_{i,j} = \left\{ \begin{array}{ll}\| \mathbf{x}_i - \mathbf{x}_j\| _2 & \| \mathbf{x}_i - \mathbf{x}_j\| _2\leq \mathcal{T}_i\\ \infty & \mathrm{otherwise} \end{array} \right.$ . Re-build the graph based on the $\mathcal{T}_i$ +4: $\dot{A}_{i,j} = \exp (-\dot{A}_{i,j})$ . Convert the distances to similarities to generate the final adjacent matrix. +5: $D_{i,i} = \sum_{j}A_{i,j}$ . Calculate the degree matrix $\mathbf{D}$ +6: Propagate the features through $\mathbf{A}$ . + +$$ +\mathbf {X} _ {t + 1} = \mathbf {D} ^ {- \frac {1}{2}} \hat {\mathbf {A}} \mathbf {D} ^ {- \frac {1}{2}} \mathbf {X} _ {t} +$$ + +adjacent matrix can be computed by: + +$$ +\hat {A} _ {i, j} = \left\{ \begin{array}{l l} \exp (- \| \mathbf {x} _ {i} - \mathbf {x} _ {j} \| _ {2} / \gamma) & \| \mathbf {x} _ {i} - \mathbf {x} _ {j} \| _ {2} \leq \mathcal {T} _ {i} \\ 0 & \text {o t h e r w i s e .} \end{array} \right. \tag {11} +$$ + +According from Eq. (7) to Eq. (11), the operation used to compute the final adjacent matrix can be easily parallelized by CUDA [18] and implemented on PyTorch [23]. The details of Cheb-GR are presented in Algorithm 1. + +# 5. Experiments + +# 5.1. Datasets and Evaluation Protocol + +Datasets We conduct extensive experiments on three popular Person ReID datasets Market-1501 (Market) [39], Duke-MTMC-re-ID (Duke) [24] and CUHK03 [16]. We utilize the Duke dataset in order to make reasonable comparisons with previous work. Market1501 is a large-scale dataset consisting of 1501 identities captured by 6 cameras, featuring challenging variations in viewpoint, illumination, and occlusion. Duke is another large-scale dataset with 1404 identities captured by 8 cameras, offering more challenging scenarios due to the inclusion of pedestrian trajectories and multiple camera views. CUHK03 is a smaller dataset with 1360 identities, but it focuses on difficult cases of cross-view matching, as images are taken from two non-overlapping camera views. + +Evaluation Metrics Performance is evaluated using the standard metrics of Cumulative Matching Characteristic (CMC), mean Average Precision (mAP), and the more recently proposed mean Inverse Negative Penalty (mINP)[35]. This combination of metrics provides a comprehensive assessment of retrieval effectiveness. To evaluate the computation costs, we measure the different methods five times and average the execution times. + +Table 1. Person ReID performance. We compare our methods with several re-ranking methods on Market1501, DukeMTMC, and CUHK03 datasets. mAP, Rank-1, and mINP are calculated to evaluate performance. Bolding indicates the lowest computation time, and red subscript text denotes the performance improvement compared to the baseline. + +
BaseMethodVenueMarketDukeCUHK03
mAP Rank-1 mINP time(s)mAP Rank-1 mINP time(s)mAP Rank-1 mINP time(s)
AGW [35]-TPAMI2188.695.366.3-79.989.845.4-66.467.355.6-
QE [6]ICCV0791.895.874.71.384.991.954.31.072.168.466.90.6
LBR [19]ICCV1991.795.576.63.984.691.856.02.171.269.664.31.0
KR [42]CVPR1794.895.989.189.890.291.775.696.078.576.176.013.4
ECN [25]CVPR1894.796.190.3161.190.792.478.2148.678.075.676.414.6
GCR [38]TMM2394.995.890.343.390.091.276.248.879.677.676.86.2
Cheb-GR-93.096.080.30.986.591.460.40.973.972.967.90.2
+4.4+0.7+14.0+6.6+1.6+15.0+7.5+5.6+12.3
TransReID [13]-ICCV2189.095.169.3-82.290.750.3-75.477.665.4-
QE [6]ICCV0791.995.377.01.787.392.859.41.680.879.075.30.6
LBR [19]ICCV1990.894.277.451.886.691.959.646.878.477.472.814.9
KR [42]CVPR1794.095.387.673.391.392.876.587.086.484.784.118.3
ECN [25]CVPR1894.095.788.7147.991.493.078.7153.786.084.484.133.6
GCR [38]TMM2392.593.886.155.590.791.775.664.384.083.781.08.0
Cheb-GR-93.295.283.51.590.493.071.21.585.483.983.80.5
+4.2+0.1+14.2+8.2+2.3+20.9+10.0+6.3+18.4
+ +![](images/f116b623d140f85fa49116a98a7731582b012b9a53960b4a79cf7f2d371dd735.jpg) +(a) Rank-1 versus computation cost + +![](images/fb1213bf4b65bd7188606b92bcf55ea8e18204e973e4d69ac51942a0b12c2cfa.jpg) +(b) mAP versus computation cost +Figure 2. Comparison of computation costs versus performance improvement of different methods for person re-identification task on CUHK03 dataset. We compare our method with k-reciprocal (KR) [41], GCR [38], ECN [25], QE [6], LBR [19]. Our method requires the least computation costs among these methods and outperforms QE [6], and LBR [19]. Although our method underperforms k-reciprocal (KR) [41], GCR [38], and ECN [25], our methods only cost about one-fiftieth of their computation time. The direction of the arrow indicates that the indicator is optimal. + +![](images/e6381dd76247570dace6899bf2577bf381b4e020b52288f4e9003d18a1ecfac3.jpg) +(c) mINP versus computation cost + +Implementation Details The original features are extracted using the pre-trained models of AGW [35] and TransReID [13]. The values of the temperature parameter $\gamma$ and the numbers of iterations $T$ are 0.2 and 3, respectively, which is the same as the setting of GCR [38]. Given the fact that there is a substantial discrepancy in the feature spaces produced by different models across various datasets, we adjust the value of parameter $\lambda$ in Eq. (7) based on the specific model and dataset used. $\lambda$ is set to 6.0/5.3/6.0/4.8/4.0 for AGW on Market/Duke/CUHK03/Market2Duke/Market2CUHK03 respectively, and 5.3/4.3/4.2 for TransReID on Mar + +ket/Duke/CUHK03 respectively. All the re-ranking methods except ECN run on GPU since running ECN on GPU causes CUDA to run out of memory. + +# 5.2. Comparison with State-of-Art Methods + +We compare our methods with k-reciprocal (KR) [41], GCR [38], ECN [25], QE [6], LBR [19] on three Person ReID datasets. We run each method five times on each dataset and calculate the average computation time to measure the computation costs. The comparison result is shown in Tab. 1. Fig. 2 uses a scatter plot to visualize the trade-offs between performance metrics and + +![](images/c17d8c33c0bb17b1c5eddfe245d803ff9729e03ca4615e717e11ee556eb1e953.jpg) +Figure 3. The impact of parameter $\lambda$ on Person ReID on the CUHK03 dataset. + +![](images/41222d296995047b0c2ad2c81da7831a6f2f39b37fcda9ab3d23ac13befd1c53.jpg) + +![](images/ad1ef4c14b00fe3beb1da3ca307c8bf61f504fcd59f4f5a419e168ba4bb9068e.jpg) + +Table 2. Results on cross-domain settings. Market2Duke denotes training on the Market1501 dataset and testing on the DukeMTMC dataset, and Market2CUHK03 denotes training on the Market1501 dataset and testing on the CUHK03 dataset. Bolding indicates the least computation time, and red subscript text denotes the performance improvement compared to the baseline. + +
BaseMethodMarket2DukeMarket2CUHK03
mAP Rank-1time(s)mAP Rank-1time(s)
AGW [35]-24.338.9-8.68.2-
QE [6]28.943.40.98.67.80.6
LBR [19]31.449.12.311.111.11.1
KR [42]35.545.295.411.69.513.2
ECN [25]35.746.1152.911.18.814.7
GCR [38]31.941.746.810.28.15.8
Cheb-GR33.045.30.910.48.60.2
+8.7+6.4+1.8+0.4
+ +computational costs on the CUHK03 dataset. As we can see, our proposed method, Cheb-GR, consistently demonstrates significantly lower computational costs compared to previous approaches, achieving a runtime that is less than the fastest existing method, regardless of the dataset or baseline model used. Notably, on the Market1501 dataset, Cheb-GR achieves a runtime of 0.9 seconds, significantly faster than the 43.3 seconds required by AGW + GCR [38] and runs on average about 50 times faster than widely used k-reciprocal re-ranking. This trend persists across the DukeMTMC and CUHK03 datasets, highlighting the method's efficiency in reducing computational overhead. Furthermore, our method achieves competitive performance, surpassing methods such as QE [6] and LBR [19] in most scenarios and even demonstrating performance comparable to or exceeding GCR [38] in some cases. For instance, on the Market1501 dataset, TransReID + Cheb-GR attains an mAP of $93.2\%$ , surpassing TransReID $^+$ QE $(91.9\%)$ ,TransReID $^+$ LBR $(90.8\%)$ ,and TransReID $^+$ GCR $(92.5\%)$ while also approaching the performance of TransReID $^+$ KR and TransReID $^+$ ECN $(94.0\%)$ . Similar observations can be made on the DukeMTMC and + +CUHK03 datasets, demonstrating the method's ability to maintain strong performance while achieving substantial computational savings. Therefore, our method offers a compelling solution for person ReID by balancing strong performance with significantly reduced computational cost, making it a practical and efficient choice for real-world applications where both accuracy and speed are paramount. + +# 5.3. Comparison on Cross-domain Settings + +To investigate the effectiveness of various re-ranking methods in cross-domain scenarios, we fine-tune a pre-trained model of AGW [35] on the Market1501 dataset and evaluate its performance on the DukeMTMC and CUHK03 dataset. We apply different re-rank methods in these cross-domain settings, and we adjust the value of k in GCR [38] from its original setting to 7, as the original setting resulted in suboptimal performance in cross-domain settings; the result is shown in Tab. 2. We can observe that performance improvement achieved by all re-ranking methods in cross-domain settings is reduced compared to the general case, highlighting the inherent intractability of the cross-domain generalization. As shown in Tab. 2, Cheb-GR exhibits competitive performance with the least computation time, showcasing its versatility across different domains. Furthermore, Cheb-GR obtains superior performance than GCR [38], demonstrating a divergence from the general trend observed, which might suggest that the adaptive neighbor search might hold a potential advantage over the k-nearest neighbor search in certain contexts. + +# 5.4. Parameter Analysis + +To analyze the impact of the hyper-parameter $\lambda$ , we conduct experiments on the CUHK03 dataset. According to the theoretical analysis in Sec. 4.1, a sufficiently large value for $\lambda$ is necessary to suppress negative samples effectively. However, an excessively large $\lambda$ risks eliminating positive samples, hindering feature propagation. In extreme circumstances, if $\lambda$ is so excessively large that it leads to the elimination of all samples, the resulting graph structure collapses into isolated nodes, thereby nullifying the re-ranking effect. + +![](images/3bdc3b32b454e7a2dc9787a4f6aafa3c1ba50b1d89375e3a694bb8f877959821.jpg) +Figure 4. The impact of parameter $\gamma$ and $T$ on Person ReID on the CUHK03 dataset under different values of $\lambda$ . + +![](images/97c4169288a7d88ea4cd62282ada2f26dfc94da412567f6cd80f20dff300ca39.jpg) + +![](images/690c536326703222dcb7aaf11690408c2d33205b85abb62dc9995e84a134767e.jpg) + +Therefore, an optimal $\lambda$ value must strike a balance between reducing noise and facilitating feature propagation. As shown in Fig. 3, our method surpasses the baseline with various values of the parameter $\lambda$ demonstrating its competitive performance, and as the value of $\lambda$ increases, mAP, mINP, and Rank-1 accuracy generally exhibit an initial rise, culminating in a peak value, followed by a subsequent decline, matching the theoretical analysis we propose. Moreover, the two baseline models exhibit distinct sensitivities to variations in $\lambda$ and differing optimal $\lambda$ values for peak performance, highlighting the disparity between the feature spaces learned by the CNN-based and transformer-based models, the transformer-based model demonstrates a considerably greater sensitivity to variations in $\lambda$ compared to the more robust CNN-based model. This suggests that the inherent characteristics of CNN-based feature spaces may contribute to enhanced robustness when employed within GCN-based re-ranking methods. + +We analyze the impact of parameters $\gamma$ and $T$ under three circumstances: $\lambda$ is optimal ( $\lambda = 4.2$ ), $\lambda$ is insufficiently large ( $\lambda = 3.0$ ), and $\lambda$ is excessively large ( $\lambda = 6.0$ ). $T$ denotes the number of iterations and $T = 0$ denotes the baseline. In each iteration, the graph convolution operation smoothes the features of connected nodes, drawing them together in the embedding space. As shown in Fig. 4, in most cases, with the increase of $T$ from 0 to 5, mAP, Rank-1, and mINP first rise and then decline. This occurs because a small value of $T$ leads to insufficient smoothing, limiting the effectiveness of feature re-ranking. In contrast, a large value of $T$ causes excessive smoothing, which may result in falsely connected features appearing closer together and separating features of unconnected positive samples. $\gamma$ is a temperature parameter controlling the sharpness of the distribution of edges' weight. From Fig. 4, we can observe that smaller $\gamma$ performs better when $\lambda$ is small, and larger $\gamma$ performs better when $\lambda$ is large. Thus, a smaller $\gamma$ improves performance when neighbors are redundant and noisy by facilitating the propagation of features to more similar neighbors, and conversely, a larger $\gamma$ proves advantageous when neighbors are limited, as it allows more features to spread + +to neighbors more equally. + +# 6. Conclusion and Discussions + +In this paper, we introduce Cheb-GR, a novel and efficient GCN-based re-ranking method for person re-identification (ReID) that addresses the critical need for both accuracy and efficiency in this challenging task. We propose a novel adaptive neighbor search for neighbor selection, replacing the computationally expensive k-nearest neighbor search and enabling significant parallelization, which is particularly well-suited for GPU acceleration. The effectiveness of Cheb-GR is rigorously validated through extensive experiments on standard benchmarks and challenging cross-domain scenarios, consistently demonstrating its competitive performance. Furthermore, we provide a theoretical analysis of the underlying principles that govern the method's efficiency and accuracy, shedding light on its ability to distinguish between similar individuals and refine the ranking process effectively. Cheb-GR's high performance and efficiency pave the way for real-time ReID applications, particularly in scenarios where speed is paramount. Beyond its demonstrated efficacy in ReID, Cheb-GR's efficient graph construction and feature refinement capabilities hold promise for broader applicability to other similarity-based retrieval tasks, suggesting its potential. + +Limitations Cheb-GR's adaptive neighbor search, while efficient, is sensitive to the trade-off parameter $\lambda$ , which balances removing negative and preserving positive edges. Careful tuning is crucial: an overly small $\lambda$ hinders re-ranking, while an excessively large $\lambda$ removes beneficial edges. Moreover, quadratic memory scaling of the adjacency matrix limits scalability, and performance depends on using features from triplet loss (or similar) trained models. + +Acknowledgement This work is partially supported by the National Key Research and Development Program of China (2024YFC3308400), National Natural Science Foundation of China under Grant (62176188, 62225113). The numerical calculations in this paper were supported by the supercomputing system at the Supercomputing Center of Wuhan University. + +# References + +[1] Yang Bai, Yucheng Ji, Min Cao, Jinqiao Wang, and Mang Ye. Chat-based person retrieval via dialogue-refined cross-modal alignment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2025. 1 +[2] Arko Barman and Shishir K Shah. Shape: A novel graph theoretic algorithm for making consensus-based decisions in person re-identification systems. In Proceedings of the IEEE International Conference on Computer Vision, pages 1115–1124, 2017. 2 +[3] Pafnutii Lvovich Chebyshev. Des valeurs moyennes. J. Math. Pures Appl, 12(2):177-184, 1867. 4 +[4] Cuiqun Chen, Mang Ye, and Ding Jiang. Towards modality-agnostic person re-identification with descriptive query. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15128-15137, 2023. 2 +[5] Yanbei Chen, Xiatian Zhu, and Shaogang Gong. Person re-identification by deep learning multi-scale representations. In Proceedings of the IEEE international conference on computer vision workshops, pages 2590–2600, 2017. 1 +[6] Ondrej Chum, James Philbin, Josef Sivic, Michael Isard, and Andrew Zisserman. Total recall: Automatic query expansion with a generative feature model for object retrieval. In 2007 IEEE 11th International Conference on Computer Vision, pages 1-8. IEEE, 2007. 6, 7 +[7] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4690-4699, 2019. 2 +[8] Bo Du, Fangyi Liu, and YE Mang. Person re-identification method based on perspective-guided multi-adversarial attention, Oct. 31 2023. US Patent 11,804,036. 2 +[9] Xing Fan, Wei Jiang, Hao Luo, and Mengjuan Fei. Spherereid: Deep hypersphere manifold embedding for person re-identification. Journal of Visual Communication and Image Representation, 60:51-58, 2019. 2 +[10] Niloofar Gheissari, Thomas B Sebastian, and Richard Hartley. Person reidentification using spatiotemporal appearance. In 2006 IEEE computer society conference on computer vision and pattern recognition (CVPR'06), volume 2, pages 1528-1535. IEEE, 2006. 1 +[11] Justin Gilmer, Samuel S Schoenholz, Patrick F Riley, Oriol Vinyals, and George E Dahl. Neural message passing for quantum chemistry. In International conference on machine learning, pages 1263-1272. PMLR, 2017. 2 +[12] Raia Hadsell, Sumit Chopra, and Yann LeCun. Dimensionality reduction by learning an invariant mapping. In 2006 IEEE computer society conference on computer vision and pattern recognition (CVPR'06), volume 2, pages 1735-1742. IEEE, 2006. 2 +[13] Shuting He, Hao Luo, Pichao Wang, Fan Wang, Hao Li, and Wei Jiang. Transreid: Transformer-based object re-identification. In Proceedings of the IEEE/CVF international conference on computer vision, pages 15013-15022, 2021. 2, 6 +[14] Bo Jiang, Xixi Wang, Aihua Zheng, Jin Tang, and Bin Luo. Ph-gen: Person retrieval with part-based hierarchical graph + +convolutional network. IEEE Transactions on Multimedia, 24:3218-3228, 2021. 1, 3 +[15] Thomas N Kipf and Max Welling. Semi-supervised classification with graph convolutional networks. arXiv preprint arXiv:1609.02907, 2016. 2 +[16] Wei Li, Rui Zhao, Tong Xiao, and Xiaogang Wang. Deepreid: Deep filter pairing neural network for person re-identification. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 152–159, 2014. 5 +[17] Fangyi Liu, Mang Ye, and Bo Du. Learning a generalizable re-identification model from unlabelled data with domain-agnostic expert. Visual Intelligence, 2(1):28, 2024. 1 +[18] David Luebke. Cuda: Scalable parallel programming for high-performance scientific computing. In 2008 5th IEEE international symposium on biomedical imaging: from nano to macro, pages 836-838. IEEE, 2008. 5 +[19] Chuanchen Luo, Yuntao Chen, Naiyan Wang, and Zhaoxiang Zhang. Spectral feature transformation for person re-identification. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4976–4985, 2019. 6, 7 +[20] Hao Luo, Youzhi Gu, Xingyu Liao, Shenqi Lai, and Wei Jiang. Bag of tricks and a strong baseline for deep person re-identification. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 0–0, 2019. 2 +[21] Jiaxu Miao, Yu Wu, Ping Liu, Yuhang Ding, and Yi Yang. Pose-guided feature alignment for occluded person re-identification. In Proceedings of the IEEE/CVF international conference on computer vision, pages 542–551, 2019. 1 +[22] Sakrapee Paisitkriangkrai, Chunhua Shen, and Anton Van Den Hengel. Learning to rank in person re-identification with metric ensembles. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1846–1855, 2015. 2 +[23] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019. 5 +[24] Ergys Ristani, Francesco Solera, Roger Zou, Rita Cucchiara, and Carlo Tomasi. Performance measures and a data set for multi-target, multi-camera tracking. In European conference on computer vision, pages 17-35. Springer, 2016. 5 +[25] M Saquib Sarfraz, Arne Schumann, Andreas Eberle, and Rainer Stiefelhagen. A pose-sensitive embedding for person re-identification with expanded cross neighborhood reranking. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 420–429, 2018. 1, 2, 4, 6, 7 +[26] Florian Schroff, Dmitry Kalenichenko, and James Philbin. Facenet: A unified embedding for face recognition and clustering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 815-823, 2015. 2 +[27] Yantao Shen, Hongsheng Li, Shuai Yi, Dapeng Chen, and Xiaogang Wang. Person re-identification with deep + +similarity-guided graph neural network. In Proceedings of the European conference on computer vision (ECCV), pages 486-504, 2018. 1 +[28] Yifan Sun, Changmao Cheng, Yuhan Zhang, Chi Zhang, Liang Zheng, Zhongdao Wang, and Yichen Wei. Circle loss: A unified perspective of pair similarity optimization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6398-6407, 2020. 2 +[29] Yifan Sun, Qin Xu, Yali Li, Chi Zhang, Yikang Li, Shengjin Wang, and Jian Sun. Perceive where to focus: Learning visibility-aware part-level features for partial person re-identification. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 393–402, 2019. 1 +[30] Yifan Sun, Liang Zheng, Yi Yang, Qi Tian, and Shengjin Wang. Beyond part models: Person retrieval with refined part pooling (and a strong convolutional baseline). In Proceedings of the European conference on computer vision (ECCV), pages 480-496, 2018. 1, 2 +[31] Haochen Wang, Jiayi Shen, Yongtuo Liu, Yan Gao, and Efstratos Gavves. Nformer: Robust person re-identification with neighbor transformer. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7297-7307, 2022. 2 +[32] Yingjia Xu, Mengxia Wu, Zixin Guo, Min Cao, Mang Ye, and Jorma Laaksonen. Efficient text-to-video retrieval via multi-modal multi-tags derived pre-screening. Visual Intelligence, 3(1):1-13, 2025. 1 +[33] Mang Ye, Shuoyi Chen, Chenyue Li, Wei-Shi Zheng, David Crandall, and Bo Du. Transformer for object re-identification: A survey. International Journal of Computer Vision, pages 1–31, 2024. 1, 2 +[34] Mang Ye, Chao Liang, Zheng Wang, Qingming Leng, and Jun Chen. Ranking optimization for person re-identification via similarity and dissimilarity. In Proceedings of the 23rd ACM international conference on Multimedia, pages 1239–1242, 2015. 2 +[35] Mang Ye, Jianbing Shen, Gaojie Lin, Tao Xiang, Ling Shao, and Steven CH Hoi. Deep learning for person re-identification: A survey and outlook. IEEE transactions on pattern analysis and machine intelligence, 44(6):2872-2893, 2021. 2, 5, 6, 7 +[36] Mang Ye, Zesen Wu, and Bo Du. Dual-level matching with outlier filtering for unsupervised visible-infrared person re-identification. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025. 1 +[37] Xuanmeng Zhang, Minyue Jiang, Zhedong Zheng, Xiao Tan, Errui Ding, and Yi Yang. Understanding image retrieval reranking: A graph neural network perspective. arXiv preprint arXiv:2012.07620, 2020. 2, 3 +[38] Yuqi Zhang, Qi Qian, Hongsong Wang, Chong Liu, Weihua Chen, and Fan Wang. Graph convolution based efficient reranking for visual retrieval. IEEE Transactions on Multimedia, 26:1089-1101, 2023. 1, 2, 3, 4, 6, 7 +[39] Liang Zheng, Liyue Shen, Lu Tian, Shengjin Wang, Jingdong Wang, and Qi Tian. Scalable person re-identification: A benchmark. In Proceedings of the IEEE international conference on computer vision, pages 1116–1124, 2015. 5 +[40] Liang Zheng, Yi Yang, and Alexander G Hauptmann. Per + +son re-identification: Past, present and future. arXiv preprint arXiv:1610.02984, 2016. 1 +[41] Meng Zheng, Srikrishna Karanam, Ziyan Wu, and Richard J Radke. Re-identification with consistent attentive siamese networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5735–5744, 2019. 1, 6 +[42] Zhun Zhong, Liang Zheng, Donglin Cao, and Shaozi Li. Reranking person re-identification with k-reciprocal encoding. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1318–1327, 2017. 1, 2, 4, 6, 7 \ No newline at end of file diff --git a/2025/Cheb-GR_ Rethinking K-nearest Neighbor Search in Re-ranking for Person Re-identification/images.zip b/2025/Cheb-GR_ Rethinking K-nearest Neighbor Search in Re-ranking for Person Re-identification/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..911e950ca0f0813a3db97c3e51e0c46856c572c1 --- /dev/null +++ b/2025/Cheb-GR_ Rethinking K-nearest Neighbor Search in Re-ranking for Person Re-identification/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8e0031a673f33c0d57e89372e906c2776ec8355074f3ccc45bb0089e7a5d8bb6 +size 548697 diff --git a/2025/Cheb-GR_ Rethinking K-nearest Neighbor Search in Re-ranking for Person Re-identification/layout.json b/2025/Cheb-GR_ Rethinking K-nearest Neighbor Search in Re-ranking for Person Re-identification/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..b70d4a502d10bc5c7f8f719dddf3c53d5f70401e --- /dev/null +++ b/2025/Cheb-GR_ Rethinking K-nearest Neighbor Search in Re-ranking for Person Re-identification/layout.json @@ -0,0 +1,8774 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 77, + 102, + 534, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 102, + 534, + 138 + ], + "spans": [ + { + "bbox": [ + 77, + 102, + 534, + 138 + ], + "type": "text", + "content": "Cheb-GR: Rethinking k-nearest neighbor search in Re-ranking for Person Re-identification" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 135, + 160, + 471, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 160, + 471, + 232 + ], + "spans": [ + { + "bbox": [ + 135, + 160, + 471, + 232 + ], + "type": "text", + "content": "Jinxi Yang * He Li * Bo Du Mang Ye† \nNational Engineering Research Center for Multimedia Software, School of Computer Science, Wuhan University, Wuhan, China \n{jinxiyang, lihe404, dubo, yemang}@whu.edu.cn \nhttps://github.com/Jinxi-Yang-WHU/Fast-GCR.git" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 151, + 258, + 200, + 272 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 258, + 200, + 272 + ], + "spans": [ + { + "bbox": [ + 151, + 258, + 200, + 272 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 283, + 297, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 283, + 297, + 608 + ], + "spans": [ + { + "bbox": [ + 54, + 283, + 297, + 608 + ], + "type": "text", + "content": "Person re-identification (ReID) is the task of matching individuals across different camera views. Existing approaches typically employ neural networks to extract discriminative features, ranking gallery images based on their similarities to probe images. While effective, these methods are often enhanced through re-ranking, a post-processing step that refines initial retrieval results without requiring additional model training. However, current re-ranking methods mostly rely on " + }, + { + "bbox": [ + 54, + 283, + 297, + 608 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 54, + 283, + 297, + 608 + ], + "type": "text", + "content": "-nearest neighbor search to extract similar images that might have the same identity as the query, which is time-consuming with a high computation burden, limiting their applications in reality. We rethink the effect of the " + }, + { + "bbox": [ + 54, + 283, + 297, + 608 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 54, + 283, + 297, + 608 + ], + "type": "text", + "content": "-nearest neighbor search and introduce the Chebyshev's Theorem-guided Graph Re-ranking (ChebGR) method, which adopts the adaptive neighbor search guided by Chebyshev's Theorem over the " + }, + { + "bbox": [ + 54, + 283, + 297, + 608 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 54, + 283, + 297, + 608 + ], + "type": "text", + "content": "-nearest neighbor search for efficient neighbor selection. Our method leverages graph convolution operations to refine image features and achieve robust re-ranking, leading to enhanced retrieval performance. Furthermore, we provide a theoretical analysis based on Chebyshev's Inequality to elucidate the factors contributing to the strong performance of the proposed method. Our method significantly reduces the computation costs while maintaining relatively strong performance. Through extensive experiments in both general and cross-domain settings, we demonstrate the effectiveness of Cheb-GR and its potential for real-world applications." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 629, + 137, + 642 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 137, + 642 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 137, + 642 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 650, + 296, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 650, + 296, + 686 + ], + "spans": [ + { + "bbox": [ + 55, + 650, + 296, + 686 + ], + "type": "text", + "content": "Person re-identification (ReID) is a computer vision task that aims to identify the same person across different camera views in a multi-camera surveillance system [5, 40]. It" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 259, + 555, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 259, + 555, + 403 + ], + "spans": [ + { + "bbox": [ + 313, + 259, + 555, + 403 + ], + "type": "text", + "content": "involves matching individuals captured by different cameras based on their appearance, even when they are viewed from different angles or under varying lighting conditions [10]. Deep learning has become the predominant approach in ReID, enabling automatic feature extraction and significantly enhancing performance [17, 21, 29, 30, 32, 33, 36, 41]. However, these methods primarily rely on pairwise similarity comparisons between the query image and individual gallery images, neglecting the contextual relationships among images within the ranking list. Consequently, they are susceptible to noise in feature representations arising from variations in illumination, pose, or occlusion." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 409, + 556, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 409, + 556, + 613 + ], + "spans": [ + { + "bbox": [ + 313, + 409, + 556, + 613 + ], + "type": "text", + "content": "To tackle this issue, many re-ranking methods have emerged as a prominent strategy for enhancing performance without the need for additional model training. Prevalent approaches involve utilizing k-nearest neighbor search to identify similar images, thereby extracting contextual information [25, 42]. However, this reliance on k-nearest neighbor search presents a significant bottleneck in real-world applications due to its computational complexity. Meanwhile, many works utilize Graph Convolutional Networks (GCN) to learn pairwise similarity information and relational information between images in person ReID tasks [1, 14, 27]. Notably, Zhang et al. [38] demonstrate the effectiveness of GCNs for re-ranking by refining embeddings through feature propagation. Despite their potential, these GCN-based methods still necessitate k-nearest neighbor search, which is inherently sequential and hinders parallel computation, leading to substantial time consumption." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 617, + 557, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 557, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 557, + 715 + ], + "type": "text", + "content": "To improve re-rank efficiency, we propose the Chebyshev's Theorem-guided Graph Re-ranking (Cheb-GR) method that eliminates the need for the k-nearest neighbor search and can be executed in parallel. Specifically, we construct a graph where each node represents an image. Unlike GCR [38], which connects each node to its k-nearest neighbors, Cheb-GR uses feature distances as edge weights, discarding edges with extreme distances. The remaining dis" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 693, + 133, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 693, + 133, + 703 + ], + "spans": [ + { + "bbox": [ + 66, + 693, + 133, + 703 + ], + "type": "text", + "content": "*Equal contribution" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 67, + 703, + 256, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 256, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 256, + 713 + ], + "type": "text", + "content": "† Corresponding Author: Mang Ye (yemang@whu.edu.cn)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19261" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 228 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 228 + ], + "type": "text", + "content": "tances are then converted into similarities. This approach directly constrains feature propagation to similar images, a key factor in achieving strong performance. A notable advantage of our method is that all operations are straightforward and highly parallelizable, resulting in significantly reduced computation costs and much faster execution compared to existing methods [25, 38, 42]. Additionally, ChebGR exhibits robust performance in cross-domain scenarios, further demonstrating the versatility of our method. We also provide a theoretical analysis to explain why the proposed approach performs effectively without relying on k-nearest neighbor search, which is a method commonly used by various re-ranking methods." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 228, + 294, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 228, + 294, + 264 + ], + "spans": [ + { + "bbox": [ + 55, + 228, + 294, + 264 + ], + "type": "text", + "content": "To the best of our knowledge, Cheb-GR is the fastest re-ranking method currently available, delivering competitive performance. In summary, our contributions are threefold:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 264, + 294, + 407 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 55, + 264, + 294, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 264, + 294, + 334 + ], + "spans": [ + { + "bbox": [ + 55, + 264, + 294, + 334 + ], + "type": "text", + "content": "- We propose Chebyshev's Theorem-guided Graph Re-ranking (Cheb-GR) method with no need for finding k-nearest neighbors, which runs faster than all re-rank methods we know. Besides, Cheb-GR achieves stronger performance on cross-domain settings than most existing methods." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 335, + 294, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 335, + 294, + 371 + ], + "spans": [ + { + "bbox": [ + 55, + 335, + 294, + 371 + ], + "type": "text", + "content": "- A reasonable theoretical explanation of the proposed method is presented to reveal why Cheb-GR can obtain strong performance without k-nearest neighbor search." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 371, + 294, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 371, + 294, + 407 + ], + "spans": [ + { + "bbox": [ + 55, + 371, + 294, + 407 + ], + "type": "text", + "content": "- Extensive experimentation across a diverse range of datasets provides compelling evidence of our method's strong performance and low computation costs." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 418, + 142, + 430 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 418, + 142, + 430 + ], + "spans": [ + { + "bbox": [ + 55, + 418, + 142, + 430 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 438, + 294, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 438, + 294, + 533 + ], + "spans": [ + { + "bbox": [ + 55, + 438, + 294, + 533 + ], + "type": "text", + "content": "Person Re-identification. Recent advancements in artificial intelligence have propelled person re-identification (Re-ID), with some methods even surpassing human performance. Current approaches can be broadly divided into two categories: the first focuses on designing metric learning loss functions to learn discriminative features, while the second modifies model architectures to capture contextual information, improving feature representation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "text", + "content": "For model architecture modification, many works have been proposed [4, 8, 31, 33, 35]. Sun et al. [30] propose a part-based convolutional baseline (PCB), utilizing uniform partitioning to capture local image information and employing a refined part pooling (RPP) method to enhance within-part consistency. Luo et al. [20] propose BoT, a strong baseline that adds a BNNeck structure after ResNet50. BNNeck introduces a batch normalization layer to separate ID and triplet loss optimization, improving convergence and performance in person ReID. He et al. [13] introduced TransReID, the first transformer-based model for person re-identification to achieve performance comparable to convolutional neural networks. For metric learning loss design, Softmax loss or its modification [7, 9] have been widely used in person Re-ID for feature learning. Loss functions" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 553, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 192 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 192 + ], + "type": "text", + "content": "utilized in metric learning, such as triplet loss [26] and contrastive loss [12], are also applied to the person Re-ID task. Some works also design novel loss functions to guide the model to learn distinguishable representations [28, 35]. Sun et al. [28] re-weight each similarity score to obtain flexible optimization and a definite convergence target. Ye et al. [35] propose Weighted Regularization Triplet loss to take full advantage of relative distance optimization between positive and negative pairs without introducing additional margin parameters." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 193, + 553, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 193, + 553, + 266 + ], + "spans": [ + { + "bbox": [ + 313, + 193, + 553, + 266 + ], + "type": "text", + "content": "All these works are committed to extracting distinguishable features of images and then calculating the similarity between the feature of a query image and the feature of each gallery image to generate a ranking list. However, these works ignore abundant contextual information within the ranking list." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 266, + 553, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 266, + 553, + 422 + ], + "spans": [ + { + "bbox": [ + 313, + 266, + 553, + 422 + ], + "type": "text", + "content": "Re-ranking for Person Re-identification. Most existing methods for Re-ID concentrate on extracting distinctive image features and assessing the similarity between the features of the query image and those of the gallery images. Additionally, a significant number of researchers have directed their attention towards re-ranking methods within ReID tasks, aiming to harness gallery-to-gallery similarity [25, 34, 42] or to integrate various ReID approaches [2, 22]. While traditional re-ranking methods have primarily emphasized the design of innovative similarity metrics, a recent advancement involves employing the graph convolutional network (GCN) to update features of images directly, which is more efficient and scalable [38]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 423, + 553, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 423, + 553, + 544 + ], + "spans": [ + { + "bbox": [ + 313, + 423, + 553, + 544 + ], + "type": "text", + "content": "Most re-ranking methods design novel distance metrics to consider global contextual information or directly refine the features of images to improve performance [25, 37, 38, 42]. Zhong et al. [42] propose a k-reciprocal encoding method that leverages contextual information between k-reciprocal nearest neighbors to improve performance. Zhang et al. [38] propose a feature propagation-based method, constructing a graph by connecting each image with its k-nearest neighbors and propagating features by graph convolution on the graph." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 544, + 553, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 544, + 553, + 616 + ], + "spans": [ + { + "bbox": [ + 313, + 544, + 553, + 616 + ], + "type": "text", + "content": "However, existing methods rely on computationally expensive k-nearest neighbor searches to identify similar images, limiting their practical applicability, particularly for large-scale datasets. So, we propose a method that circumvents the need to find k-nearest neighbors, significantly reducing computation costs." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 617, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 553, + 713 + ], + "type": "text", + "content": "Graph Convolutional Network in ReID. Graph Convolutional Network (GCN) [11] utilizes a message-passing mechanism to allow nodes to exchange and aggregate information with neighbors. Through a convolutional operation, each node's representation is updated by a weighted sum of its features and those of its neighbors, computed via matrix multiplication with a learned weight matrix [15]. This iterative process allows GCNs to learn rich representations that" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19262" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 84 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 84 + ], + "type": "text", + "content": "capture both local and global information within the graph." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 84, + 295, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 84, + 295, + 217 + ], + "spans": [ + { + "bbox": [ + 54, + 84, + 295, + 217 + ], + "type": "text", + "content": "Since graphs are a great data structure for describing relationships and GCNs can learn graph-structure information, many ReID methods utilize GCNs to learn pairwise similarity and relational information among image sets. Jiang et al. [14] propose a method based on a hierarchical graph to integrate local, global, and structural information. Zhang et al. [37] and GCR [38] show that GCN can be leveraged in the re-ranking area. This paper employs a simplified GCN that achieves feature propagation without a learned weight matrix. This parameter-free approach enhances efficiency by eliminating the need for training." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 225, + 214, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 225, + 214, + 239 + ], + "spans": [ + { + "bbox": [ + 55, + 225, + 214, + 239 + ], + "type": "text", + "content": "3. Preliminary and Motivation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 245, + 296, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 245, + 296, + 413 + ], + "spans": [ + { + "bbox": [ + 56, + 245, + 296, + 413 + ], + "type": "text", + "content": "Efficiency of K-nearest Neighbor Search. Current person ReID re-ranking methods rely heavily on k-nearest neighbor search, a key operation used to select potential same-identity images. K-nearest neighbor search computes the distances between the query image and all other images in the gallery to identify the k-nearest neighbors, which is computationally expensive and presents significant challenges when parallelizing on GPUs. The time complexity of sorting N images is " + }, + { + "bbox": [ + 56, + 245, + 296, + 413 + ], + "type": "inline_equation", + "content": "\\Theta (N\\log N)" + }, + { + "bbox": [ + 56, + 245, + 296, + 413 + ], + "type": "text", + "content": ". In order to select the k-nearest neighbors for each node, we need to perform the sorting operations " + }, + { + "bbox": [ + 56, + 245, + 296, + 413 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 56, + 245, + 296, + 413 + ], + "type": "text", + "content": " times and select the top-k images each time. Thus, the total time complexity of the k-nearest neighbor search is " + }, + { + "bbox": [ + 56, + 245, + 296, + 413 + ], + "type": "inline_equation", + "content": "\\Theta (N^2\\log N)" + }, + { + "bbox": [ + 56, + 245, + 296, + 413 + ], + "type": "text", + "content": ", which becomes the primary computational bottleneck in the re-ranking process." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 413, + 296, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 413, + 296, + 557 + ], + "spans": [ + { + "bbox": [ + 55, + 413, + 296, + 557 + ], + "type": "text", + "content": "Graph based Re-ranking. Zhang et al. [38] propose Graph Convolutional-based Re-ranking (GCR). In GCR, given a set of images, a graph is constructed that encompasses both query and gallery images. Each query and gallery image corresponds to a node in the graph, with its associated feature vector serving as the initial node feature. Edges are established between nodes representing visually similar images, encoding potential identity relationships within the dataset. This constructed graph then serves as the input to a GCN, which performs iterative message-passing operations to refine the node features. The message-passing operation in a GCN layer can be written as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 116, + 563, + 295, + 578 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 563, + 295, + 578 + ], + "spans": [ + { + "bbox": [ + 116, + 563, + 295, + 578 + ], + "type": "interline_equation", + "content": "\\mathbf {X} ^ {(l)} = \\mathbf {D} ^ {- \\frac {1}{2}} \\mathbf {A D} ^ {- \\frac {1}{2}} \\mathbf {X} ^ {(l - 1)} \\mathbf {W}, \\tag {1}", + "image_path": "57463a3a177d14039fc4ba2ecffa54878017bd66de3733b8060b87946fd1b7b4.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 583, + 295, + 669 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 583, + 295, + 669 + ], + "spans": [ + { + "bbox": [ + 55, + 583, + 295, + 669 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 583, + 295, + 669 + ], + "type": "inline_equation", + "content": "\\mathbf{X}^{(l)}\\in \\mathbb{R}^{N\\times d}" + }, + { + "bbox": [ + 55, + 583, + 295, + 669 + ], + "type": "text", + "content": " is the node features in the " + }, + { + "bbox": [ + 55, + 583, + 295, + 669 + ], + "type": "inline_equation", + "content": "l^{th}" + }, + { + "bbox": [ + 55, + 583, + 295, + 669 + ], + "type": "text", + "content": " layer, " + }, + { + "bbox": [ + 55, + 583, + 295, + 669 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 583, + 295, + 669 + ], + "type": "text", + "content": " denotes the number of nodes, " + }, + { + "bbox": [ + 55, + 583, + 295, + 669 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 55, + 583, + 295, + 669 + ], + "type": "text", + "content": " denotes the feature dimensionality. " + }, + { + "bbox": [ + 55, + 583, + 295, + 669 + ], + "type": "inline_equation", + "content": "\\mathbf{A}\\in \\mathbb{R}^{N\\times N}" + }, + { + "bbox": [ + 55, + 583, + 295, + 669 + ], + "type": "text", + "content": " is the adjacent matrix of the graph, " + }, + { + "bbox": [ + 55, + 583, + 295, + 669 + ], + "type": "inline_equation", + "content": "\\mathbf{D}\\in \\mathbb{R}^{N\\times N}" + }, + { + "bbox": [ + 55, + 583, + 295, + 669 + ], + "type": "text", + "content": " is the degree matrix of the graph which is a diagonal matrix and " + }, + { + "bbox": [ + 55, + 583, + 295, + 669 + ], + "type": "inline_equation", + "content": "\\mathbf{D}_{i,i} = \\sum_{j}\\mathbf{A}_{i,j}" + }, + { + "bbox": [ + 55, + 583, + 295, + 669 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 55, + 583, + 295, + 669 + ], + "type": "inline_equation", + "content": "\\mathbf{W}\\in \\mathbb{R}^{d\\times d'}" + }, + { + "bbox": [ + 55, + 583, + 295, + 669 + ], + "type": "text", + "content": " denotes a learnable weighted matrix in the GCN layer, which can project the features vector into a certain space." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 669, + 296, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 669, + 296, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 669, + 296, + 693 + ], + "type": "text", + "content": "When the weighted matrix " + }, + { + "bbox": [ + 55, + 669, + 296, + 693 + ], + "type": "inline_equation", + "content": "\\mathbf{W}" + }, + { + "bbox": [ + 55, + 669, + 296, + 693 + ], + "type": "text", + "content": " is an identity matrix, the graph convolution operation can be written as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 120, + 699, + 295, + 714 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 120, + 699, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 120, + 699, + 295, + 714 + ], + "type": "interline_equation", + "content": "\\mathbf {X} ^ {(l)} = \\mathbf {D} ^ {- \\frac {1}{2}} \\mathbf {A} \\mathbf {D} ^ {- \\frac {1}{2}} \\mathbf {X} ^ {(l - 1)}. \\tag {2}", + "image_path": "e56e784f3b0b11cdf6eb6daabbea4d27634919ebf5f2855bf8cf7cf5791c3424.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 555, + 336 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 336 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 336 + ], + "type": "text", + "content": "As formalized in Eq. (2), this parameter-free operation iteratively updates each node's feature embedding by aggregating the features of its neighbors and itself. The weight assigned to each neighbor's contribution is determined by the corresponding edge weight in the graph. Ideally, Graph Convolution-based Re-ranking enhances feature representations by promoting propagation between images of the same identity. This process brings together features of matching identities while pushing apart features of distinct identities, leading to improved re-ranking performance. Therefore, the primary challenge lies in constructing a graph that maximizes connections between images of the same identity while minimizing connections between images of different identities. As illustrated in Fig. 1(b), GCR employs a k-nearest neighbor search to identify neighboring nodes. According to Eq. (2), the time complexity of a graph convolution operation in GCR [38] is equal to matrix multiplication, which is " + }, + { + "bbox": [ + 313, + 72, + 555, + 336 + ], + "type": "inline_equation", + "content": "\\Theta(N^3)" + }, + { + "bbox": [ + 313, + 72, + 555, + 336 + ], + "type": "text", + "content": ". However, since the matrix multiplication is highly parallelized on GPUs, it runs much faster than the k-nearest neighbor search in practice. Thus the efficiency bottleneck of GCR is still mainly affected by k-nearest neighbor search." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 346, + 421, + 361 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 346, + 421, + 361 + ], + "spans": [ + { + "bbox": [ + 313, + 346, + 421, + 361 + ], + "type": "text", + "content": "4. Proposed Method" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 367, + 555, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 367, + 555, + 415 + ], + "spans": [ + { + "bbox": [ + 313, + 367, + 555, + 415 + ], + "type": "text", + "content": "In this section, we first revisit the neighbor search and introduce an efficient strategy to replace the k-nearest neighbor search in Sec. 4.1. After that, we introduce our method in detail in Sec. 4.2." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 423, + 466, + 436 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 423, + 466, + 436 + ], + "spans": [ + { + "bbox": [ + 313, + 423, + 466, + 436 + ], + "type": "text", + "content": "4.1. Revisiting Neighbor Search" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 441, + 554, + 560 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 441, + 554, + 560 + ], + "spans": [ + { + "bbox": [ + 313, + 441, + 554, + 560 + ], + "type": "text", + "content": "Since the role of the k-nearest neighbor search in re-ranking is to select potential positive samples, we need to find an alternative strategy to find potential positive samples with low time complexity to lower the computation costs. Since feature distance is inversely related to the probability of shared identity, a threshold can be established based on the distance between the features of an anchor image and other images. Images with a distance below this threshold are then considered positive samples, sharing the same identity as the anchor image." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 561, + 554, + 596 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 561, + 554, + 596 + ], + "spans": [ + { + "bbox": [ + 313, + 561, + 554, + 596 + ], + "type": "text", + "content": "The determination of an appropriate threshold for feature distance warrants further consideration. The triplet loss can be written as:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 359, + 607, + 555, + 620 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 607, + 555, + 620 + ], + "spans": [ + { + "bbox": [ + 359, + 607, + 555, + 620 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\max (d (a, p) - d (a, n) + \\alpha , 0), \\tag {3}", + "image_path": "d690447b978eb599f8b056064b66f5de9ece9480ab6ff40bcbffbcd7934a1697.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 629, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 554, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 629, + 554, + 714 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 629, + 554, + 714 + ], + "type": "text", + "content": " is the margin hyperparameter. Eq. (3) implies that, ideally, the distance between images of different identities should exceed the distance between images of the same identity by at least a factor of " + }, + { + "bbox": [ + 313, + 629, + 554, + 714 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 629, + 554, + 714 + ], + "type": "text", + "content": ". Given the significant disparity in the number of positive and negative samples within the datasets, with far fewer positive samples, we can simplify our analysis by focusing on the distribution of the" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19263" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 82, + 71, + 317, + 208 + ], + "blocks": [ + { + "bbox": [ + 82, + 71, + 317, + 208 + ], + "lines": [ + { + "bbox": [ + 82, + 71, + 317, + 208 + ], + "spans": [ + { + "bbox": [ + 82, + 71, + 317, + 208 + ], + "type": "image", + "image_path": "00a74549813f566b92ceb391b299b4ff9ec46b8ee8738b2fc5efcfdff8395385.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 320, + 72, + 528, + 208 + ], + "blocks": [ + { + "bbox": [ + 320, + 72, + 528, + 208 + ], + "lines": [ + { + "bbox": [ + 320, + 72, + 528, + 208 + ], + "spans": [ + { + "bbox": [ + 320, + 72, + 528, + 208 + ], + "type": "image", + "image_path": "2ea39a22ba03492d6e32fb733ff801034e95d0b8b624e30c5030c923e1648423.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 82, + 209, + 382, + 345 + ], + "blocks": [ + { + "bbox": [ + 82, + 209, + 382, + 345 + ], + "lines": [ + { + "bbox": [ + 82, + 209, + 382, + 345 + ], + "spans": [ + { + "bbox": [ + 82, + 209, + 382, + 345 + ], + "type": "image", + "image_path": "4a031eff605da7e3a3d47c8586770dc32d8db3f0f5430c49f8642f7881e95c44.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 354, + 555, + 432 + ], + "lines": [ + { + "bbox": [ + 54, + 354, + 555, + 432 + ], + "spans": [ + { + "bbox": [ + 54, + 354, + 555, + 432 + ], + "type": "text", + "content": "Figure 1. Comparison between traditional re-rank methods. GCR [38] and the proposed Chebyshev's Guided Graph Re-ranking (Cheb-GR). (a) Most traditional re-rank methods, such as k-reciprocal [42] and ECN [25], operate in three stages: neighborhood set construction, dissimilarity measurement between these sets, and weighted distance aggregation combining direct and neighborhood-based distances. (b) GCR connects each node to its k-nearest neighbors to construct the graph. GCNs are then employed to propagate the node features through the graph, refining the image representations for improved re-ranking. (c) Cheb-GR leverages an adaptive neighbor search instead of a k-nearest neighbor search to select neighbors and construct the graph. Subsequently, feature propagation is achieved through the application of a GCN." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 389, + 209, + 528, + 345 + ], + "blocks": [ + { + "bbox": [ + 389, + 209, + 528, + 345 + ], + "lines": [ + { + "bbox": [ + 389, + 209, + 528, + 345 + ], + "spans": [ + { + "bbox": [ + 389, + 209, + 528, + 345 + ], + "type": "image", + "image_path": "4833588ef2c759bfe2eae85ff9ce94f713f2a95190371833db9e1c497c45dc8b.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 453, + 295, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 453, + 295, + 501 + ], + "spans": [ + { + "bbox": [ + 55, + 453, + 295, + 501 + ], + "type": "text", + "content": "distance between negative samples in the embedding space. Thus, the problem is reframed as one of identifying the minimum distance separating the negative samples. According to Chebyshev's Theorem [3]:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 81, + 511, + 193, + 536 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 511, + 193, + 536 + ], + "spans": [ + { + "bbox": [ + 81, + 511, + 193, + 536 + ], + "type": "interline_equation", + "content": "P r [ | R | \\geq x ] \\leq \\frac {E x [ | R | ^ {\\alpha} ]}{x ^ {\\alpha}},", + "image_path": "a1b34c2a0c15c16494dfc332050761adf843d25b9849fc04106fc6b5e0b8b004.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 83, + 537, + 295, + 562 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 537, + 295, + 562 + ], + "spans": [ + { + "bbox": [ + 83, + 537, + 295, + 562 + ], + "type": "interline_equation", + "content": "P r [ | R - E x [ R ] | \\geq x ] \\leq \\frac {E x [ | R - E x [ R ] | ^ {\\alpha} ]}{x ^ {\\alpha}}, \\tag {4}", + "image_path": "1e622a97791d6a4f083b899960ba43454344fa5a45f227da798802153f8c146c.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 83, + 563, + 223, + 587 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 563, + 223, + 587 + ], + "spans": [ + { + "bbox": [ + 83, + 563, + 223, + 587 + ], + "type": "interline_equation", + "content": "P r [ | R - E x [ R ] | \\geq x ] \\leq \\frac {V a r [ R ]}{x ^ {2}},", + "image_path": "f4b3ade6b055d1e9ad4c96cf0ca1f7f70e871d0e309d9bb2b5d42581404dadb5.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 597, + 295, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 597, + 295, + 658 + ], + "spans": [ + { + "bbox": [ + 55, + 597, + 295, + 658 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 597, + 295, + 658 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 55, + 597, + 295, + 658 + ], + "type": "text", + "content": " denotes a random variable, " + }, + { + "bbox": [ + 55, + 597, + 295, + 658 + ], + "type": "inline_equation", + "content": "Ex[R]" + }, + { + "bbox": [ + 55, + 597, + 295, + 658 + ], + "type": "text", + "content": " denotes the expected value of the random variable " + }, + { + "bbox": [ + 55, + 597, + 295, + 658 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 55, + 597, + 295, + 658 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 597, + 295, + 658 + ], + "type": "inline_equation", + "content": "Var[R]" + }, + { + "bbox": [ + 55, + 597, + 295, + 658 + ], + "type": "text", + "content": " denotes the variance of the random variable " + }, + { + "bbox": [ + 55, + 597, + 295, + 658 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 55, + 597, + 295, + 658 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 597, + 295, + 658 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 597, + 295, + 658 + ], + "type": "text", + "content": " denotes a real number. We can replace " + }, + { + "bbox": [ + 55, + 597, + 295, + 658 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 597, + 295, + 658 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 55, + 597, + 295, + 658 + ], + "type": "inline_equation", + "content": "\\lambda \\sigma" + }, + { + "bbox": [ + 55, + 597, + 295, + 658 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 597, + 295, + 658 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 597, + 295, + 658 + ], + "type": "text", + "content": " is a constant hyperparameter and " + }, + { + "bbox": [ + 55, + 597, + 295, + 658 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 55, + 597, + 295, + 658 + ], + "type": "text", + "content": " is standard deviation, to get the inequality:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 111, + 668, + 295, + 692 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 668, + 295, + 692 + ], + "spans": [ + { + "bbox": [ + 111, + 668, + 295, + 692 + ], + "type": "interline_equation", + "content": "P r [ | R - E x [ R ] | \\geq \\lambda \\sigma ] \\leq \\frac {1}{\\lambda^ {2}}. \\tag {5}", + "image_path": "0f5db02263c5442b0b85b2a954a91a728b33831c693b046c8848409dc05d85ca.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 701, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 701, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 701, + 295, + 714 + ], + "type": "text", + "content": "Then, removing the absolute value operation and ignoring" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 453, + 548, + 465 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 453, + 548, + 465 + ], + "spans": [ + { + "bbox": [ + 313, + 453, + 548, + 465 + ], + "type": "text", + "content": "the scenario that " + }, + { + "bbox": [ + 313, + 453, + 548, + 465 + ], + "type": "inline_equation", + "content": "R \\geq Ex[R] + \\lambda \\sigma" + }, + { + "bbox": [ + 313, + 453, + 548, + 465 + ], + "type": "text", + "content": ", we get the inequality:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 373, + 474, + 555, + 498 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 373, + 474, + 555, + 498 + ], + "spans": [ + { + "bbox": [ + 373, + 474, + 555, + 498 + ], + "type": "interline_equation", + "content": "P r [ R \\leq E x [ R ] - \\lambda \\sigma ] \\leq \\frac {1}{\\lambda^ {2}}, \\tag {6}", + "image_path": "64b9c11bfa00938ce8ee8233dde66803065ff13674dfcbdf54a59fe4b463d57c.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 504, + 555, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 504, + 555, + 552 + ], + "spans": [ + { + "bbox": [ + 313, + 504, + 555, + 552 + ], + "type": "text", + "content": "which shows that only a small number of elements fall below the mean minus " + }, + { + "bbox": [ + 313, + 504, + 555, + 552 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 504, + 555, + 552 + ], + "type": "text", + "content": " times the standard deviation when " + }, + { + "bbox": [ + 313, + 504, + 555, + 552 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 504, + 555, + 552 + ], + "type": "text", + "content": " is sufficiently large. Thus, given an image " + }, + { + "bbox": [ + 313, + 504, + 555, + 552 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 504, + 555, + 552 + ], + "type": "text", + "content": ", we can set threshold " + }, + { + "bbox": [ + 313, + 504, + 555, + 552 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_i" + }, + { + "bbox": [ + 313, + 504, + 555, + 552 + ], + "type": "text", + "content": " for image " + }, + { + "bbox": [ + 313, + 504, + 555, + 552 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 504, + 555, + 552 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 371, + 561, + 436, + 574 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 371, + 561, + 436, + 574 + ], + "spans": [ + { + "bbox": [ + 371, + 561, + 436, + 574 + ], + "type": "interline_equation", + "content": "\\mathcal {T} _ {i} = \\mu_ {i} - \\lambda \\sigma_ {i},", + "image_path": "6eb53164fbe4afca81dcb8346cb72b98de9d5dc76d02c5005d09128bdd1d3b41.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 372, + 576, + 554, + 611 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 576, + 554, + 611 + ], + "spans": [ + { + "bbox": [ + 372, + 576, + 554, + 611 + ], + "type": "interline_equation", + "content": "\\mu_ {i} = \\frac {1}{N} \\sum_ {j = 1} ^ {N} D i s _ {i, j}, \\tag {7}", + "image_path": "59c401f8a755d092e9431fc40d9ddb2bc0691f43ac4badddb43c4a62f7b8d885.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 372, + 613, + 497, + 646 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 613, + 497, + 646 + ], + "spans": [ + { + "bbox": [ + 372, + 613, + 497, + 646 + ], + "type": "interline_equation", + "content": "\\sigma_ {i} = \\sqrt {\\frac {\\sum_ {j = 1} ^ {N} (D i s _ {i , j} - \\mu_ {i}) ^ {2}}{N}},", + "image_path": "a5dc595ba769b1988da53864ef660aec6a0138f358b350a7ca31c7ff5564af4b.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 654, + 555, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 555, + 712 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 654, + 555, + 712 + ], + "type": "inline_equation", + "content": "Dis_{i,j}" + }, + { + "bbox": [ + 313, + 654, + 555, + 712 + ], + "type": "text", + "content": " denotes the distance between features of image " + }, + { + "bbox": [ + 313, + 654, + 555, + 712 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 654, + 555, + 712 + ], + "type": "text", + "content": " and image " + }, + { + "bbox": [ + 313, + 654, + 555, + 712 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 654, + 555, + 712 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 654, + 555, + 712 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 654, + 555, + 712 + ], + "type": "text", + "content": " denotes the number of images in the dataset, " + }, + { + "bbox": [ + 313, + 654, + 555, + 712 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 654, + 555, + 712 + ], + "type": "text", + "content": " is a hyper-parameter, " + }, + { + "bbox": [ + 313, + 654, + 555, + 712 + ], + "type": "inline_equation", + "content": "\\mu_i" + }, + { + "bbox": [ + 313, + 654, + 555, + 712 + ], + "type": "text", + "content": " is the mean of distances between features of an image in the dataset and image " + }, + { + "bbox": [ + 313, + 654, + 555, + 712 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 654, + 555, + 712 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 654, + 555, + 712 + ], + "type": "inline_equation", + "content": "\\sigma_i" + }, + { + "bbox": [ + 313, + 654, + 555, + 712 + ], + "type": "text", + "content": " is the standard deviations of distances between features of" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19264" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 277 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 277 + ], + "type": "text", + "content": "an image in the dataset and image " + }, + { + "bbox": [ + 55, + 72, + 296, + 277 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 72, + 296, + 277 + ], + "type": "text", + "content": ". Given the presence of a margin " + }, + { + "bbox": [ + 55, + 72, + 296, + 277 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 72, + 296, + 277 + ], + "type": "text", + "content": " between the distances of positive and negative samples, there exists a specific value of " + }, + { + "bbox": [ + 55, + 72, + 296, + 277 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 72, + 296, + 277 + ], + "type": "text", + "content": " that ensures that the threshold " + }, + { + "bbox": [ + 55, + 72, + 296, + 277 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_i" + }, + { + "bbox": [ + 55, + 72, + 296, + 277 + ], + "type": "text", + "content": " surpasses the distances between positive samples while remaining below the distances between negative samples. Therefore, by carefully selecting an appropriate value for " + }, + { + "bbox": [ + 55, + 72, + 296, + 277 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 72, + 296, + 277 + ], + "type": "text", + "content": ", we can effectively filter out negative samples while preserving positive samples within the graph construction process. The time complexity of calculating the mean and standard deviation of " + }, + { + "bbox": [ + 55, + 72, + 296, + 277 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 72, + 296, + 277 + ], + "type": "text", + "content": " elements is " + }, + { + "bbox": [ + 55, + 72, + 296, + 277 + ], + "type": "inline_equation", + "content": "\\Theta(N)" + }, + { + "bbox": [ + 55, + 72, + 296, + 277 + ], + "type": "text", + "content": ", so the time complexity of calculating the threshold for " + }, + { + "bbox": [ + 55, + 72, + 296, + 277 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 72, + 296, + 277 + ], + "type": "text", + "content": " images is " + }, + { + "bbox": [ + 55, + 72, + 296, + 277 + ], + "type": "inline_equation", + "content": "\\Theta(N^2)" + }, + { + "bbox": [ + 55, + 72, + 296, + 277 + ], + "type": "text", + "content": ", which is less than the time complexity of the k-nearest neighbor search. Moreover, our proposed method can be easily parallelized on GPUs and implemented on the PyTorch platform [23], so the computation cost of our method is much less than the computation costs of the k-nearest neighbor search in practice." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 281, + 265, + 295 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 281, + 265, + 295 + ], + "spans": [ + { + "bbox": [ + 55, + 281, + 265, + 295 + ], + "type": "text", + "content": "4.2. Chebyshev's Guided Graph Re-ranking" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 299, + 296, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 299, + 296, + 348 + ], + "spans": [ + { + "bbox": [ + 55, + 299, + 296, + 348 + ], + "type": "text", + "content": "The pipeline of our method is shown in Fig. 1(c). We first construct a directed complete graph, and the weight of each edge is the distance between the nodes the edge connects. The adjacent matrix of this graph can be computed by:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 134, + 353, + 295, + 367 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 353, + 295, + 367 + ], + "spans": [ + { + "bbox": [ + 134, + 353, + 295, + 367 + ], + "type": "interline_equation", + "content": "A _ {i, j} = \\left\\| \\mathbf {x} _ {i} - \\mathbf {x} _ {j} \\right\\| _ {2}, \\tag {8}", + "image_path": "f9999f76d4dbe4cc0e5e99dd47e76d8752cb471bcc1cb5b99fc8fafd090cdf03.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 371, + 296, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 371, + 296, + 467 + ], + "spans": [ + { + "bbox": [ + 55, + 371, + 296, + 467 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 371, + 296, + 467 + ], + "type": "inline_equation", + "content": "\\| \\cdot \\|_2" + }, + { + "bbox": [ + 55, + 371, + 296, + 467 + ], + "type": "text", + "content": " denotes Euclidean distance. Intuitively, the larger the distance between the features of two images, the less likely they have the same identity. To ensure that feature propagation is constrained to images of the same identity as far as possible, a threshold-based edge deletion strategy is employed. Edges with weights exceeding a predefined threshold, which is computed as Eq. (7), are removed from the graph's adjacency matrix." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 468, + 296, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 468, + 296, + 503 + ], + "spans": [ + { + "bbox": [ + 55, + 468, + 296, + 503 + ], + "type": "text", + "content": "We delete the edges whose weights exceed the corresponding threshold and get the pruned graph whose adjacent matrix can be expressed by:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 89, + 509, + 295, + 540 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 509, + 295, + 540 + ], + "spans": [ + { + "bbox": [ + 89, + 509, + 295, + 540 + ], + "type": "interline_equation", + "content": "\\tilde {A} _ {i, j} = \\left\\{ \\begin{array}{l l} \\| \\mathbf {x} _ {i} - \\mathbf {x} _ {j} \\| _ {2} & \\| \\mathbf {x} _ {i} - \\mathbf {x} _ {j} \\| _ {2} \\leq \\mathcal {T} _ {i} \\\\ \\infty & \\text {o t h e r w i s e ,} \\end{array} \\right. \\tag {9}", + "image_path": "21a4414d74a80ee64950c774019618fe11d55ad268e607d2f95110c95b012f29.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 544, + 296, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 544, + 296, + 665 + ], + "spans": [ + { + "bbox": [ + 55, + 544, + 296, + 665 + ], + "type": "text", + "content": "where threshold " + }, + { + "bbox": [ + 55, + 544, + 296, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_i" + }, + { + "bbox": [ + 55, + 544, + 296, + 665 + ], + "type": "text", + "content": " is computed by Eq. (7). Setting the weight of deleted edges to infinity ensures that, after converting distance to similarity, these edges effectively become non-existent. This approach is justified by the interpretation that if two nodes are not connected, their distance is considered infinite. The desired outcome is that edge weights should positively correlate with the similarity of features between connected nodes. However, the adjacent matrix computed by Eq. (9) is just the opposite. So, we convert the distance to similarity for edge weight by:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 99, + 670, + 295, + 685 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 670, + 295, + 685 + ], + "spans": [ + { + "bbox": [ + 99, + 670, + 295, + 685 + ], + "type": "interline_equation", + "content": "\\operatorname {S i m} (i, j) = \\exp (- \\| \\mathbf {x} _ {i} - \\mathbf {x} _ {j} \\| _ {2} / \\gamma) \\tag {10}", + "image_path": "2f2c70d86e4ea606caf07b31f49652078a92dd08e95efa10b8d5bf678c87736c.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "Sim(i,j)" + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": " denotes the similarity between image " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": " and image " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": " is the temperature hyper-parameter. The final" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 315, + 72, + 541, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 72, + 541, + 84 + ], + "spans": [ + { + "bbox": [ + 315, + 72, + 541, + 84 + ], + "type": "text", + "content": "Algorithm 1 Chebyshev's Guided Graph Re-ranking" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 315, + 85, + 534, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 85, + 534, + 97 + ], + "spans": [ + { + "bbox": [ + 315, + 85, + 534, + 97 + ], + "type": "text", + "content": "Input: " + }, + { + "bbox": [ + 315, + 85, + 534, + 97 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_0" + }, + { + "bbox": [ + 315, + 85, + 534, + 97 + ], + "type": "text", + "content": " - Original query and gallery image features" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 315, + 98, + 496, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 98, + 496, + 109 + ], + "spans": [ + { + "bbox": [ + 315, + 98, + 496, + 109 + ], + "type": "text", + "content": "Output: " + }, + { + "bbox": [ + 315, + 98, + 496, + 109 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_T" + }, + { + "bbox": [ + 315, + 98, + 496, + 109 + ], + "type": "text", + "content": " - Final features after re-ranking" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 319, + 110, + 554, + 239 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 319, + 110, + 554, + 134 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 110, + 554, + 134 + ], + "spans": [ + { + "bbox": [ + 319, + 110, + 554, + 134 + ], + "type": "text", + "content": "1: " + }, + { + "bbox": [ + 319, + 110, + 554, + 134 + ], + "type": "inline_equation", + "content": "A_{i,j} = \\parallel \\mathbf{x}_i - \\mathbf{x}_j\\parallel_2" + }, + { + "bbox": [ + 319, + 110, + 554, + 134 + ], + "type": "text", + "content": ". Calculate the original adjacent matrix " + }, + { + "bbox": [ + 319, + 110, + 554, + 134 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 319, + 135, + 503, + 147 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 135, + 503, + 147 + ], + "spans": [ + { + "bbox": [ + 319, + 135, + 503, + 147 + ], + "type": "text", + "content": "2: " + }, + { + "bbox": [ + 319, + 135, + 503, + 147 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_i = \\mu_i - \\lambda \\sigma_i" + }, + { + "bbox": [ + 319, + 135, + 503, + 147 + ], + "type": "text", + "content": ". Calculate the threshold " + }, + { + "bbox": [ + 319, + 135, + 503, + 147 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_i" + }, + { + "bbox": [ + 319, + 135, + 503, + 147 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 319, + 147, + 554, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 147, + 554, + 187 + ], + "spans": [ + { + "bbox": [ + 319, + 147, + 554, + 187 + ], + "type": "text", + "content": "3: " + }, + { + "bbox": [ + 319, + 147, + 554, + 187 + ], + "type": "inline_equation", + "content": "\\tilde{A}_{i,j} = \\left\\{ \\begin{array}{ll}\\| \\mathbf{x}_i - \\mathbf{x}_j\\| _2 & \\| \\mathbf{x}_i - \\mathbf{x}_j\\| _2\\leq \\mathcal{T}_i\\\\ \\infty & \\mathrm{otherwise} \\end{array} \\right." + }, + { + "bbox": [ + 319, + 147, + 554, + 187 + ], + "type": "text", + "content": " . Re-build the graph based on the " + }, + { + "bbox": [ + 319, + 147, + 554, + 187 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_i" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 319, + 188, + 553, + 211 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 188, + 553, + 211 + ], + "spans": [ + { + "bbox": [ + 319, + 188, + 553, + 211 + ], + "type": "text", + "content": "4: " + }, + { + "bbox": [ + 319, + 188, + 553, + 211 + ], + "type": "inline_equation", + "content": "\\dot{A}_{i,j} = \\exp (-\\dot{A}_{i,j})" + }, + { + "bbox": [ + 319, + 188, + 553, + 211 + ], + "type": "text", + "content": " . Convert the distances to similarities to generate the final adjacent matrix." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 319, + 212, + 524, + 225 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 212, + 524, + 225 + ], + "spans": [ + { + "bbox": [ + 319, + 212, + 524, + 225 + ], + "type": "text", + "content": "5: " + }, + { + "bbox": [ + 319, + 212, + 524, + 225 + ], + "type": "inline_equation", + "content": "D_{i,i} = \\sum_{j}A_{i,j}" + }, + { + "bbox": [ + 319, + 212, + 524, + 225 + ], + "type": "text", + "content": " . Calculate the degree matrix " + }, + { + "bbox": [ + 319, + 212, + 524, + 225 + ], + "type": "inline_equation", + "content": "\\mathbf{D}" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 319, + 227, + 468, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 227, + 468, + 239 + ], + "spans": [ + { + "bbox": [ + 319, + 227, + 468, + 239 + ], + "type": "text", + "content": "6: Propagate the features through " + }, + { + "bbox": [ + 319, + 227, + 468, + 239 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 319, + 227, + 468, + 239 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 394, + 247, + 490, + 262 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 394, + 247, + 490, + 262 + ], + "spans": [ + { + "bbox": [ + 394, + 247, + 490, + 262 + ], + "type": "interline_equation", + "content": "\\mathbf {X} _ {t + 1} = \\mathbf {D} ^ {- \\frac {1}{2}} \\hat {\\mathbf {A}} \\mathbf {D} ^ {- \\frac {1}{2}} \\mathbf {X} _ {t}", + "image_path": "939cbaf6255c66a4ea557e1459d32cd1e8e021f3be52cd51a8e2ffd369ecdd38.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 314, + 294, + 465, + 306 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 294, + 465, + 306 + ], + "spans": [ + { + "bbox": [ + 314, + 294, + 465, + 306 + ], + "type": "text", + "content": "adjacent matrix can be computed by:" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 324, + 312, + 553, + 354 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 312, + 553, + 354 + ], + "spans": [ + { + "bbox": [ + 324, + 312, + 553, + 354 + ], + "type": "interline_equation", + "content": "\\hat {A} _ {i, j} = \\left\\{ \\begin{array}{l l} \\exp (- \\| \\mathbf {x} _ {i} - \\mathbf {x} _ {j} \\| _ {2} / \\gamma) & \\| \\mathbf {x} _ {i} - \\mathbf {x} _ {j} \\| _ {2} \\leq \\mathcal {T} _ {i} \\\\ 0 & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {11}", + "image_path": "7ce2b22ac6116c52bc945017644fa4320e4c12d75c9791ca177e1df5a8000b84.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 355, + 554, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 355, + 554, + 403 + ], + "spans": [ + { + "bbox": [ + 313, + 355, + 554, + 403 + ], + "type": "text", + "content": "According from Eq. (7) to Eq. (11), the operation used to compute the final adjacent matrix can be easily parallelized by CUDA [18] and implemented on PyTorch [23]. The details of Cheb-GR are presented in Algorithm 1." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 314, + 413, + 395, + 426 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 413, + 395, + 426 + ], + "spans": [ + { + "bbox": [ + 314, + 413, + 395, + 426 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 433, + 493, + 445 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 433, + 493, + 445 + ], + "spans": [ + { + "bbox": [ + 313, + 433, + 493, + 445 + ], + "type": "text", + "content": "5.1. Datasets and Evaluation Protocol" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 450, + 555, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 450, + 555, + 617 + ], + "spans": [ + { + "bbox": [ + 313, + 450, + 555, + 617 + ], + "type": "text", + "content": "Datasets We conduct extensive experiments on three popular Person ReID datasets Market-1501 (Market) [39], Duke-MTMC-re-ID (Duke) [24] and CUHK03 [16]. We utilize the Duke dataset in order to make reasonable comparisons with previous work. Market1501 is a large-scale dataset consisting of 1501 identities captured by 6 cameras, featuring challenging variations in viewpoint, illumination, and occlusion. Duke is another large-scale dataset with 1404 identities captured by 8 cameras, offering more challenging scenarios due to the inclusion of pedestrian trajectories and multiple camera views. CUHK03 is a smaller dataset with 1360 identities, but it focuses on difficult cases of cross-view matching, as images are taken from two non-overlapping camera views." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "type": "text", + "content": "Evaluation Metrics Performance is evaluated using the standard metrics of Cumulative Matching Characteristic (CMC), mean Average Precision (mAP), and the more recently proposed mean Inverse Negative Penalty (mINP)[35]. This combination of metrics provides a comprehensive assessment of retrieval effectiveness. To evaluate the computation costs, we measure the different methods five times and average the execution times." + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19265" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 69, + 108, + 541, + 352 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 103 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 103 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 103 + ], + "type": "text", + "content": "Table 1. Person ReID performance. We compare our methods with several re-ranking methods on Market1501, DukeMTMC, and CUHK03 datasets. mAP, Rank-1, and mINP are calculated to evaluate performance. Bolding indicates the lowest computation time, and red subscript text denotes the performance improvement compared to the baseline." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 69, + 108, + 541, + 352 + ], + "lines": [ + { + "bbox": [ + 69, + 108, + 541, + 352 + ], + "spans": [ + { + "bbox": [ + 69, + 108, + 541, + 352 + ], + "type": "table", + "html": "
BaseMethodVenueMarketDukeCUHK03
mAP Rank-1 mINP time(s)mAP Rank-1 mINP time(s)mAP Rank-1 mINP time(s)
AGW [35]-TPAMI2188.695.366.3-79.989.845.4-66.467.355.6-
QE [6]ICCV0791.895.874.71.384.991.954.31.072.168.466.90.6
LBR [19]ICCV1991.795.576.63.984.691.856.02.171.269.664.31.0
KR [42]CVPR1794.895.989.189.890.291.775.696.078.576.176.013.4
ECN [25]CVPR1894.796.190.3161.190.792.478.2148.678.075.676.414.6
GCR [38]TMM2394.995.890.343.390.091.276.248.879.677.676.86.2
Cheb-GR-93.096.080.30.986.591.460.40.973.972.967.90.2
+4.4+0.7+14.0+6.6+1.6+15.0+7.5+5.6+12.3
TransReID [13]-ICCV2189.095.169.3-82.290.750.3-75.477.665.4-
QE [6]ICCV0791.995.377.01.787.392.859.41.680.879.075.30.6
LBR [19]ICCV1990.894.277.451.886.691.959.646.878.477.472.814.9
KR [42]CVPR1794.095.387.673.391.392.876.587.086.484.784.118.3
ECN [25]CVPR1894.095.788.7147.991.493.078.7153.786.084.484.133.6
GCR [38]TMM2392.593.886.155.590.791.775.664.384.083.781.08.0
Cheb-GR-93.295.283.51.590.493.071.21.585.483.983.80.5
+4.2+0.1+14.2+8.2+2.3+20.9+10.0+6.3+18.4
", + "image_path": "b58548780f3b56691e30cc85879fc42e6b07d84b8ac551c34a89229cba54dd4d.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 65, + 363, + 220, + 479 + ], + "blocks": [ + { + "bbox": [ + 65, + 363, + 220, + 479 + ], + "lines": [ + { + "bbox": [ + 65, + 363, + 220, + 479 + ], + "spans": [ + { + "bbox": [ + 65, + 363, + 220, + 479 + ], + "type": "image", + "image_path": "f116b623d140f85fa49116a98a7731582b012b9a53960b4a79cf7f2d371dd735.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 85, + 483, + 201, + 492 + ], + "lines": [ + { + "bbox": [ + 85, + 483, + 201, + 492 + ], + "spans": [ + { + "bbox": [ + 85, + 483, + 201, + 492 + ], + "type": "text", + "content": "(a) Rank-1 versus computation cost" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 225, + 363, + 382, + 479 + ], + "blocks": [ + { + "bbox": [ + 225, + 363, + 382, + 479 + ], + "lines": [ + { + "bbox": [ + 225, + 363, + 382, + 479 + ], + "spans": [ + { + "bbox": [ + 225, + 363, + 382, + 479 + ], + "type": "image", + "image_path": "fb1213bf4b65bd7188606b92bcf55ea8e18204e973e4d69ac51942a0b12c2cfa.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 249, + 483, + 358, + 492 + ], + "lines": [ + { + "bbox": [ + 249, + 483, + 358, + 492 + ], + "spans": [ + { + "bbox": [ + 249, + 483, + 358, + 492 + ], + "type": "text", + "content": "(b) mAP versus computation cost" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 54, + 497, + 555, + 553 + ], + "lines": [ + { + "bbox": [ + 54, + 497, + 555, + 553 + ], + "spans": [ + { + "bbox": [ + 54, + 497, + 555, + 553 + ], + "type": "text", + "content": "Figure 2. Comparison of computation costs versus performance improvement of different methods for person re-identification task on CUHK03 dataset. We compare our method with k-reciprocal (KR) [41], GCR [38], ECN [25], QE [6], LBR [19]. Our method requires the least computation costs among these methods and outperforms QE [6], and LBR [19]. Although our method underperforms k-reciprocal (KR) [41], GCR [38], and ECN [25], our methods only cost about one-fiftieth of their computation time. The direction of the arrow indicates that the indicator is optimal." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 388, + 363, + 542, + 479 + ], + "blocks": [ + { + "bbox": [ + 388, + 363, + 542, + 479 + ], + "lines": [ + { + "bbox": [ + 388, + 363, + 542, + 479 + ], + "spans": [ + { + "bbox": [ + 388, + 363, + 542, + 479 + ], + "type": "image", + "image_path": "e6381dd76247570dace6899bf2577bf381b4e020b52288f4e9003d18a1ecfac3.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 410, + 483, + 520, + 492 + ], + "lines": [ + { + "bbox": [ + 410, + 483, + 520, + 492 + ], + "spans": [ + { + "bbox": [ + 410, + 483, + 520, + 492 + ], + "type": "text", + "content": "(c) mINP versus computation cost" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 563, + 297, + 708 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 563, + 297, + 708 + ], + "spans": [ + { + "bbox": [ + 55, + 563, + 297, + 708 + ], + "type": "text", + "content": "Implementation Details The original features are extracted using the pre-trained models of AGW [35] and TransReID [13]. The values of the temperature parameter " + }, + { + "bbox": [ + 55, + 563, + 297, + 708 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 55, + 563, + 297, + 708 + ], + "type": "text", + "content": " and the numbers of iterations " + }, + { + "bbox": [ + 55, + 563, + 297, + 708 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 563, + 297, + 708 + ], + "type": "text", + "content": " are 0.2 and 3, respectively, which is the same as the setting of GCR [38]. Given the fact that there is a substantial discrepancy in the feature spaces produced by different models across various datasets, we adjust the value of parameter " + }, + { + "bbox": [ + 55, + 563, + 297, + 708 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 563, + 297, + 708 + ], + "type": "text", + "content": " in Eq. (7) based on the specific model and dataset used. " + }, + { + "bbox": [ + 55, + 563, + 297, + 708 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 563, + 297, + 708 + ], + "type": "text", + "content": " is set to 6.0/5.3/6.0/4.8/4.0 for AGW on Market/Duke/CUHK03/Market2Duke/Market2CUHK03 respectively, and 5.3/4.3/4.2 for TransReID on Mar" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 563, + 555, + 600 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 563, + 555, + 600 + ], + "spans": [ + { + "bbox": [ + 313, + 563, + 555, + 600 + ], + "type": "text", + "content": "ket/Duke/CUHK03 respectively. All the re-ranking methods except ECN run on GPU since running ECN on GPU causes CUDA to run out of memory." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 611, + 522, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 611, + 522, + 624 + ], + "spans": [ + { + "bbox": [ + 313, + 611, + 522, + 624 + ], + "type": "text", + "content": "5.2. Comparison with State-of-Art Methods" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "content": "We compare our methods with k-reciprocal (KR) [41], GCR [38], ECN [25], QE [6], LBR [19] on three Person ReID datasets. We run each method five times on each dataset and calculate the average computation time to measure the computation costs. The comparison result is shown in Tab. 1. Fig. 2 uses a scatter plot to visualize the trade-offs between performance metrics and" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19266" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 73, + 74, + 226, + 199 + ], + "blocks": [ + { + "bbox": [ + 73, + 74, + 226, + 199 + ], + "lines": [ + { + "bbox": [ + 73, + 74, + 226, + 199 + ], + "spans": [ + { + "bbox": [ + 73, + 74, + 226, + 199 + ], + "type": "image", + "image_path": "c17d8c33c0bb17b1c5eddfe245d803ff9729e03ca4615e717e11ee556eb1e953.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 153, + 208, + 455, + 220 + ], + "lines": [ + { + "bbox": [ + 153, + 208, + 455, + 220 + ], + "spans": [ + { + "bbox": [ + 153, + 208, + 455, + 220 + ], + "type": "text", + "content": "Figure 3. The impact of parameter " + }, + { + "bbox": [ + 153, + 208, + 455, + 220 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 153, + 208, + 455, + 220 + ], + "type": "text", + "content": " on Person ReID on the CUHK03 dataset." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 228, + 74, + 382, + 199 + ], + "blocks": [ + { + "bbox": [ + 228, + 74, + 382, + 199 + ], + "lines": [ + { + "bbox": [ + 228, + 74, + 382, + 199 + ], + "spans": [ + { + "bbox": [ + 228, + 74, + 382, + 199 + ], + "type": "image", + "image_path": "41222d296995047b0c2ad2c81da7831a6f2f39b37fcda9ab3d23ac13befd1c53.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 383, + 75, + 536, + 198 + ], + "blocks": [ + { + "bbox": [ + 383, + 75, + 536, + 198 + ], + "lines": [ + { + "bbox": [ + 383, + 75, + 536, + 198 + ], + "spans": [ + { + "bbox": [ + 383, + 75, + 536, + 198 + ], + "type": "image", + "image_path": "ad1ef4c14b00fe3beb1da3ca307c8bf61f504fcd59f4f5a419e168ba4bb9068e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 58, + 304, + 294, + 428 + ], + "blocks": [ + { + "bbox": [ + 55, + 229, + 295, + 295 + ], + "lines": [ + { + "bbox": [ + 55, + 229, + 295, + 295 + ], + "spans": [ + { + "bbox": [ + 55, + 229, + 295, + 295 + ], + "type": "text", + "content": "Table 2. Results on cross-domain settings. Market2Duke denotes training on the Market1501 dataset and testing on the DukeMTMC dataset, and Market2CUHK03 denotes training on the Market1501 dataset and testing on the CUHK03 dataset. Bolding indicates the least computation time, and red subscript text denotes the performance improvement compared to the baseline." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 304, + 294, + 428 + ], + "lines": [ + { + "bbox": [ + 58, + 304, + 294, + 428 + ], + "spans": [ + { + "bbox": [ + 58, + 304, + 294, + 428 + ], + "type": "table", + "html": "
BaseMethodMarket2DukeMarket2CUHK03
mAP Rank-1time(s)mAP Rank-1time(s)
AGW [35]-24.338.9-8.68.2-
QE [6]28.943.40.98.67.80.6
LBR [19]31.449.12.311.111.11.1
KR [42]35.545.295.411.69.513.2
ECN [25]35.746.1152.911.18.814.7
GCR [38]31.941.746.810.28.15.8
Cheb-GR33.045.30.910.48.60.2
+8.7+6.4+1.8+0.4
", + "image_path": "4d187df7a0a1ada9e23ffe34e928008e3ff9a308513ff18b0b6387e668dfee77.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 450, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 450, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 450, + 295, + 713 + ], + "type": "text", + "content": "computational costs on the CUHK03 dataset. As we can see, our proposed method, Cheb-GR, consistently demonstrates significantly lower computational costs compared to previous approaches, achieving a runtime that is less than the fastest existing method, regardless of the dataset or baseline model used. Notably, on the Market1501 dataset, Cheb-GR achieves a runtime of 0.9 seconds, significantly faster than the 43.3 seconds required by AGW + GCR [38] and runs on average about 50 times faster than widely used k-reciprocal re-ranking. This trend persists across the DukeMTMC and CUHK03 datasets, highlighting the method's efficiency in reducing computational overhead. Furthermore, our method achieves competitive performance, surpassing methods such as QE [6] and LBR [19] in most scenarios and even demonstrating performance comparable to or exceeding GCR [38] in some cases. For instance, on the Market1501 dataset, TransReID + Cheb-GR attains an mAP of " + }, + { + "bbox": [ + 55, + 450, + 295, + 713 + ], + "type": "inline_equation", + "content": "93.2\\%" + }, + { + "bbox": [ + 55, + 450, + 295, + 713 + ], + "type": "text", + "content": " , surpassing TransReID " + }, + { + "bbox": [ + 55, + 450, + 295, + 713 + ], + "type": "inline_equation", + "content": "^+" + }, + { + "bbox": [ + 55, + 450, + 295, + 713 + ], + "type": "text", + "content": " QE " + }, + { + "bbox": [ + 55, + 450, + 295, + 713 + ], + "type": "inline_equation", + "content": "(91.9\\%)" + }, + { + "bbox": [ + 55, + 450, + 295, + 713 + ], + "type": "text", + "content": " ,TransReID " + }, + { + "bbox": [ + 55, + 450, + 295, + 713 + ], + "type": "inline_equation", + "content": "^+" + }, + { + "bbox": [ + 55, + 450, + 295, + 713 + ], + "type": "text", + "content": " LBR " + }, + { + "bbox": [ + 55, + 450, + 295, + 713 + ], + "type": "inline_equation", + "content": "(90.8\\%)" + }, + { + "bbox": [ + 55, + 450, + 295, + 713 + ], + "type": "text", + "content": " ,and TransReID " + }, + { + "bbox": [ + 55, + 450, + 295, + 713 + ], + "type": "inline_equation", + "content": "^+" + }, + { + "bbox": [ + 55, + 450, + 295, + 713 + ], + "type": "text", + "content": " GCR " + }, + { + "bbox": [ + 55, + 450, + 295, + 713 + ], + "type": "inline_equation", + "content": "(92.5\\%)" + }, + { + "bbox": [ + 55, + 450, + 295, + 713 + ], + "type": "text", + "content": " while also approaching the performance of TransReID " + }, + { + "bbox": [ + 55, + 450, + 295, + 713 + ], + "type": "inline_equation", + "content": "^+" + }, + { + "bbox": [ + 55, + 450, + 295, + 713 + ], + "type": "text", + "content": " KR and TransReID " + }, + { + "bbox": [ + 55, + 450, + 295, + 713 + ], + "type": "inline_equation", + "content": "^+" + }, + { + "bbox": [ + 55, + 450, + 295, + 713 + ], + "type": "text", + "content": " ECN " + }, + { + "bbox": [ + 55, + 450, + 295, + 713 + ], + "type": "inline_equation", + "content": "(94.0\\%)" + }, + { + "bbox": [ + 55, + 450, + 295, + 713 + ], + "type": "text", + "content": " . Similar observations can be made on the DukeMTMC and" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 231, + 555, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 231, + 555, + 316 + ], + "spans": [ + { + "bbox": [ + 313, + 231, + 555, + 316 + ], + "type": "text", + "content": "CUHK03 datasets, demonstrating the method's ability to maintain strong performance while achieving substantial computational savings. Therefore, our method offers a compelling solution for person ReID by balancing strong performance with significantly reduced computational cost, making it a practical and efficient choice for real-world applications where both accuracy and speed are paramount." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 323, + 517, + 335 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 323, + 517, + 335 + ], + "spans": [ + { + "bbox": [ + 313, + 323, + 517, + 335 + ], + "type": "text", + "content": "5.3. Comparison on Cross-domain Settings" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 340, + 555, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 340, + 555, + 579 + ], + "spans": [ + { + "bbox": [ + 313, + 340, + 555, + 579 + ], + "type": "text", + "content": "To investigate the effectiveness of various re-ranking methods in cross-domain scenarios, we fine-tune a pre-trained model of AGW [35] on the Market1501 dataset and evaluate its performance on the DukeMTMC and CUHK03 dataset. We apply different re-rank methods in these cross-domain settings, and we adjust the value of k in GCR [38] from its original setting to 7, as the original setting resulted in suboptimal performance in cross-domain settings; the result is shown in Tab. 2. We can observe that performance improvement achieved by all re-ranking methods in cross-domain settings is reduced compared to the general case, highlighting the inherent intractability of the cross-domain generalization. As shown in Tab. 2, Cheb-GR exhibits competitive performance with the least computation time, showcasing its versatility across different domains. Furthermore, Cheb-GR obtains superior performance than GCR [38], demonstrating a divergence from the general trend observed, which might suggest that the adaptive neighbor search might hold a potential advantage over the k-nearest neighbor search in certain contexts." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 588, + 429, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 588, + 429, + 601 + ], + "spans": [ + { + "bbox": [ + 313, + 588, + 429, + 601 + ], + "type": "text", + "content": "5.4. Parameter Analysis" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "type": "text", + "content": "To analyze the impact of the hyper-parameter " + }, + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "type": "text", + "content": ", we conduct experiments on the CUHK03 dataset. According to the theoretical analysis in Sec. 4.1, a sufficiently large value for " + }, + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "type": "text", + "content": " is necessary to suppress negative samples effectively. However, an excessively large " + }, + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "type": "text", + "content": " risks eliminating positive samples, hindering feature propagation. In extreme circumstances, if " + }, + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "type": "text", + "content": " is so excessively large that it leads to the elimination of all samples, the resulting graph structure collapses into isolated nodes, thereby nullifying the re-ranking effect." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19267" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 78, + 76, + 229, + 192 + ], + "blocks": [ + { + "bbox": [ + 78, + 76, + 229, + 192 + ], + "lines": [ + { + "bbox": [ + 78, + 76, + 229, + 192 + ], + "spans": [ + { + "bbox": [ + 78, + 76, + 229, + 192 + ], + "type": "image", + "image_path": "3bdc3b32b454e7a2dc9787a4f6aafa3c1ba50b1d89375e3a694bb8f877959821.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 88, + 206, + 520, + 218 + ], + "lines": [ + { + "bbox": [ + 88, + 206, + 520, + 218 + ], + "spans": [ + { + "bbox": [ + 88, + 206, + 520, + 218 + ], + "type": "text", + "content": "Figure 4. The impact of parameter " + }, + { + "bbox": [ + 88, + 206, + 520, + 218 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 88, + 206, + 520, + 218 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 88, + 206, + 520, + 218 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 88, + 206, + 520, + 218 + ], + "type": "text", + "content": " on Person ReID on the CUHK03 dataset under different values of " + }, + { + "bbox": [ + 88, + 206, + 520, + 218 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 88, + 206, + 520, + 218 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 230, + 76, + 380, + 192 + ], + "blocks": [ + { + "bbox": [ + 230, + 76, + 380, + 192 + ], + "lines": [ + { + "bbox": [ + 230, + 76, + 380, + 192 + ], + "spans": [ + { + "bbox": [ + 230, + 76, + 380, + 192 + ], + "type": "image", + "image_path": "97c4169288a7d88ea4cd62282ada2f26dfc94da412567f6cd80f20dff300ca39.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 383, + 76, + 533, + 192 + ], + "blocks": [ + { + "bbox": [ + 383, + 76, + 533, + 192 + ], + "lines": [ + { + "bbox": [ + 383, + 76, + 533, + 192 + ], + "spans": [ + { + "bbox": [ + 383, + 76, + 533, + 192 + ], + "type": "image", + "image_path": "690c536326703222dcb7aaf11690408c2d33205b85abb62dc9995e84a134767e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 234, + 297, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 234, + 297, + 449 + ], + "spans": [ + { + "bbox": [ + 54, + 234, + 297, + 449 + ], + "type": "text", + "content": "Therefore, an optimal " + }, + { + "bbox": [ + 54, + 234, + 297, + 449 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 54, + 234, + 297, + 449 + ], + "type": "text", + "content": " value must strike a balance between reducing noise and facilitating feature propagation. As shown in Fig. 3, our method surpasses the baseline with various values of the parameter " + }, + { + "bbox": [ + 54, + 234, + 297, + 449 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 54, + 234, + 297, + 449 + ], + "type": "text", + "content": " demonstrating its competitive performance, and as the value of " + }, + { + "bbox": [ + 54, + 234, + 297, + 449 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 54, + 234, + 297, + 449 + ], + "type": "text", + "content": " increases, mAP, mINP, and Rank-1 accuracy generally exhibit an initial rise, culminating in a peak value, followed by a subsequent decline, matching the theoretical analysis we propose. Moreover, the two baseline models exhibit distinct sensitivities to variations in " + }, + { + "bbox": [ + 54, + 234, + 297, + 449 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 54, + 234, + 297, + 449 + ], + "type": "text", + "content": " and differing optimal " + }, + { + "bbox": [ + 54, + 234, + 297, + 449 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 54, + 234, + 297, + 449 + ], + "type": "text", + "content": " values for peak performance, highlighting the disparity between the feature spaces learned by the CNN-based and transformer-based models, the transformer-based model demonstrates a considerably greater sensitivity to variations in " + }, + { + "bbox": [ + 54, + 234, + 297, + 449 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 54, + 234, + 297, + 449 + ], + "type": "text", + "content": " compared to the more robust CNN-based model. This suggests that the inherent characteristics of CNN-based feature spaces may contribute to enhanced robustness when employed within GCN-based re-ranking methods." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 451, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 451, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 451, + 296, + 714 + ], + "type": "text", + "content": "We analyze the impact of parameters " + }, + { + "bbox": [ + 56, + 451, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 56, + 451, + 296, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 56, + 451, + 296, + 714 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 56, + 451, + 296, + 714 + ], + "type": "text", + "content": " under three circumstances: " + }, + { + "bbox": [ + 56, + 451, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 56, + 451, + 296, + 714 + ], + "type": "text", + "content": " is optimal (" + }, + { + "bbox": [ + 56, + 451, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\lambda = 4.2" + }, + { + "bbox": [ + 56, + 451, + 296, + 714 + ], + "type": "text", + "content": "), " + }, + { + "bbox": [ + 56, + 451, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 56, + 451, + 296, + 714 + ], + "type": "text", + "content": " is insufficiently large (" + }, + { + "bbox": [ + 56, + 451, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\lambda = 3.0" + }, + { + "bbox": [ + 56, + 451, + 296, + 714 + ], + "type": "text", + "content": "), and " + }, + { + "bbox": [ + 56, + 451, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 56, + 451, + 296, + 714 + ], + "type": "text", + "content": " is excessively large (" + }, + { + "bbox": [ + 56, + 451, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\lambda = 6.0" + }, + { + "bbox": [ + 56, + 451, + 296, + 714 + ], + "type": "text", + "content": "). " + }, + { + "bbox": [ + 56, + 451, + 296, + 714 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 56, + 451, + 296, + 714 + ], + "type": "text", + "content": " denotes the number of iterations and " + }, + { + "bbox": [ + 56, + 451, + 296, + 714 + ], + "type": "inline_equation", + "content": "T = 0" + }, + { + "bbox": [ + 56, + 451, + 296, + 714 + ], + "type": "text", + "content": " denotes the baseline. In each iteration, the graph convolution operation smoothes the features of connected nodes, drawing them together in the embedding space. As shown in Fig. 4, in most cases, with the increase of " + }, + { + "bbox": [ + 56, + 451, + 296, + 714 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 56, + 451, + 296, + 714 + ], + "type": "text", + "content": " from 0 to 5, mAP, Rank-1, and mINP first rise and then decline. This occurs because a small value of " + }, + { + "bbox": [ + 56, + 451, + 296, + 714 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 56, + 451, + 296, + 714 + ], + "type": "text", + "content": " leads to insufficient smoothing, limiting the effectiveness of feature re-ranking. In contrast, a large value of " + }, + { + "bbox": [ + 56, + 451, + 296, + 714 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 56, + 451, + 296, + 714 + ], + "type": "text", + "content": " causes excessive smoothing, which may result in falsely connected features appearing closer together and separating features of unconnected positive samples. " + }, + { + "bbox": [ + 56, + 451, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 56, + 451, + 296, + 714 + ], + "type": "text", + "content": " is a temperature parameter controlling the sharpness of the distribution of edges' weight. From Fig. 4, we can observe that smaller " + }, + { + "bbox": [ + 56, + 451, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 56, + 451, + 296, + 714 + ], + "type": "text", + "content": " performs better when " + }, + { + "bbox": [ + 56, + 451, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 56, + 451, + 296, + 714 + ], + "type": "text", + "content": " is small, and larger " + }, + { + "bbox": [ + 56, + 451, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 56, + 451, + 296, + 714 + ], + "type": "text", + "content": " performs better when " + }, + { + "bbox": [ + 56, + 451, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 56, + 451, + 296, + 714 + ], + "type": "text", + "content": " is large. Thus, a smaller " + }, + { + "bbox": [ + 56, + 451, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 56, + 451, + 296, + 714 + ], + "type": "text", + "content": " improves performance when neighbors are redundant and noisy by facilitating the propagation of features to more similar neighbors, and conversely, a larger " + }, + { + "bbox": [ + 56, + 451, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 56, + 451, + 296, + 714 + ], + "type": "text", + "content": " proves advantageous when neighbors are limited, as it allows more features to spread" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 234, + 424, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 234, + 424, + 247 + ], + "spans": [ + { + "bbox": [ + 313, + 234, + 424, + 247 + ], + "type": "text", + "content": "to neighbors more equally." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 253, + 471, + 266 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 253, + 471, + 266 + ], + "spans": [ + { + "bbox": [ + 313, + 253, + 471, + 266 + ], + "type": "text", + "content": "6. Conclusion and Discussions" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 270, + 555, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 270, + 555, + 533 + ], + "spans": [ + { + "bbox": [ + 313, + 270, + 555, + 533 + ], + "type": "text", + "content": "In this paper, we introduce Cheb-GR, a novel and efficient GCN-based re-ranking method for person re-identification (ReID) that addresses the critical need for both accuracy and efficiency in this challenging task. We propose a novel adaptive neighbor search for neighbor selection, replacing the computationally expensive k-nearest neighbor search and enabling significant parallelization, which is particularly well-suited for GPU acceleration. The effectiveness of Cheb-GR is rigorously validated through extensive experiments on standard benchmarks and challenging cross-domain scenarios, consistently demonstrating its competitive performance. Furthermore, we provide a theoretical analysis of the underlying principles that govern the method's efficiency and accuracy, shedding light on its ability to distinguish between similar individuals and refine the ranking process effectively. Cheb-GR's high performance and efficiency pave the way for real-time ReID applications, particularly in scenarios where speed is paramount. Beyond its demonstrated efficacy in ReID, Cheb-GR's efficient graph construction and feature refinement capabilities hold promise for broader applicability to other similarity-based retrieval tasks, suggesting its potential." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 534, + 556, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 556, + 629 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 556, + 629 + ], + "type": "text", + "content": "Limitations Cheb-GR's adaptive neighbor search, while efficient, is sensitive to the trade-off parameter " + }, + { + "bbox": [ + 313, + 534, + 556, + 629 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 534, + 556, + 629 + ], + "type": "text", + "content": ", which balances removing negative and preserving positive edges. Careful tuning is crucial: an overly small " + }, + { + "bbox": [ + 313, + 534, + 556, + 629 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 534, + 556, + 629 + ], + "type": "text", + "content": " hinders re-ranking, while an excessively large " + }, + { + "bbox": [ + 313, + 534, + 556, + 629 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 534, + 556, + 629 + ], + "type": "text", + "content": " removes beneficial edges. Moreover, quadratic memory scaling of the adjacency matrix limits scalability, and performance depends on using features from triplet loss (or similar) trained models." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "content": "Acknowledgement This work is partially supported by the National Key Research and Development Program of China (2024YFC3308400), National Natural Science Foundation of China under Grant (62176188, 62225113). The numerical calculations in this paper were supported by the supercomputing system at the Supercomputing Center of Wuhan University." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19268" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 704 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 134 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 134 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 134 + ], + "type": "text", + "content": "[1] Yang Bai, Yucheng Ji, Min Cao, Jinqiao Wang, and Mang Ye. Chat-based person retrieval via dialogue-refined cross-modal alignment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2025. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 134, + 296, + 187 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 134, + 296, + 187 + ], + "spans": [ + { + "bbox": [ + 61, + 134, + 296, + 187 + ], + "type": "text", + "content": "[2] Arko Barman and Shishir K Shah. Shape: A novel graph theoretic algorithm for making consensus-based decisions in person re-identification systems. In Proceedings of the IEEE International Conference on Computer Vision, pages 1115–1124, 2017. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 188, + 294, + 209 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 188, + 294, + 209 + ], + "spans": [ + { + "bbox": [ + 62, + 188, + 294, + 209 + ], + "type": "text", + "content": "[3] Pafnutii Lvovich Chebyshev. Des valeurs moyennes. J. Math. Pures Appl, 12(2):177-184, 1867. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 209, + 295, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 209, + 295, + 252 + ], + "spans": [ + { + "bbox": [ + 62, + 209, + 295, + 252 + ], + "type": "text", + "content": "[4] Cuiqun Chen, Mang Ye, and Ding Jiang. Towards modality-agnostic person re-identification with descriptive query. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15128-15137, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 252, + 294, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 252, + 294, + 295 + ], + "spans": [ + { + "bbox": [ + 62, + 252, + 294, + 295 + ], + "type": "text", + "content": "[5] Yanbei Chen, Xiatian Zhu, and Shaogang Gong. Person re-identification by deep learning multi-scale representations. In Proceedings of the IEEE international conference on computer vision workshops, pages 2590–2600, 2017. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 295, + 294, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 295, + 294, + 348 + ], + "spans": [ + { + "bbox": [ + 62, + 295, + 294, + 348 + ], + "type": "text", + "content": "[6] Ondrej Chum, James Philbin, Josef Sivic, Michael Isard, and Andrew Zisserman. Total recall: Automatic query expansion with a generative feature model for object retrieval. In 2007 IEEE 11th International Conference on Computer Vision, pages 1-8. IEEE, 2007. 6, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 349, + 294, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 349, + 294, + 402 + ], + "spans": [ + { + "bbox": [ + 62, + 349, + 294, + 402 + ], + "type": "text", + "content": "[7] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4690-4699, 2019. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 403, + 294, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 403, + 294, + 434 + ], + "spans": [ + { + "bbox": [ + 62, + 403, + 294, + 434 + ], + "type": "text", + "content": "[8] Bo Du, Fangyi Liu, and YE Mang. Person re-identification method based on perspective-guided multi-adversarial attention, Oct. 31 2023. US Patent 11,804,036. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 434, + 294, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 434, + 294, + 477 + ], + "spans": [ + { + "bbox": [ + 62, + 434, + 294, + 477 + ], + "type": "text", + "content": "[9] Xing Fan, Wei Jiang, Hao Luo, and Mengjuan Fei. Spherereid: Deep hypersphere manifold embedding for person re-identification. Journal of Visual Communication and Image Representation, 60:51-58, 2019. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 477, + 294, + 530 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 477, + 294, + 530 + ], + "spans": [ + { + "bbox": [ + 57, + 477, + 294, + 530 + ], + "type": "text", + "content": "[10] Niloofar Gheissari, Thomas B Sebastian, and Richard Hartley. Person reidentification using spatiotemporal appearance. In 2006 IEEE computer society conference on computer vision and pattern recognition (CVPR'06), volume 2, pages 1528-1535. IEEE, 2006. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 531, + 294, + 574 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 531, + 294, + 574 + ], + "spans": [ + { + "bbox": [ + 57, + 531, + 294, + 574 + ], + "type": "text", + "content": "[11] Justin Gilmer, Samuel S Schoenholz, Patrick F Riley, Oriol Vinyals, and George E Dahl. Neural message passing for quantum chemistry. In International conference on machine learning, pages 1263-1272. PMLR, 2017. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 574, + 294, + 627 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 574, + 294, + 627 + ], + "spans": [ + { + "bbox": [ + 57, + 574, + 294, + 627 + ], + "type": "text", + "content": "[12] Raia Hadsell, Sumit Chopra, and Yann LeCun. Dimensionality reduction by learning an invariant mapping. In 2006 IEEE computer society conference on computer vision and pattern recognition (CVPR'06), volume 2, pages 1735-1742. IEEE, 2006. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 628, + 294, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 628, + 294, + 680 + ], + "spans": [ + { + "bbox": [ + 57, + 628, + 294, + 680 + ], + "type": "text", + "content": "[13] Shuting He, Hao Luo, Pichao Wang, Fan Wang, Hao Li, and Wei Jiang. Transreid: Transformer-based object re-identification. In Proceedings of the IEEE/CVF international conference on computer vision, pages 15013-15022, 2021. 2, 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 681, + 294, + 704 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 681, + 294, + 704 + ], + "spans": [ + { + "bbox": [ + 57, + 681, + 294, + 704 + ], + "type": "text", + "content": "[14] Bo Jiang, Xixi Wang, Aihua Zheng, Jin Tang, and Bin Luo. Ph-gen: Person retrieval with part-based hierarchical graph" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 74, + 553, + 707 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 335, + 74, + 553, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 74, + 553, + 94 + ], + "spans": [ + { + "bbox": [ + 335, + 74, + 553, + 94 + ], + "type": "text", + "content": "convolutional network. IEEE Transactions on Multimedia, 24:3218-3228, 2021. 1, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 95, + 553, + 125 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 95, + 553, + 125 + ], + "spans": [ + { + "bbox": [ + 316, + 95, + 553, + 125 + ], + "type": "text", + "content": "[15] Thomas N Kipf and Max Welling. Semi-supervised classification with graph convolutional networks. arXiv preprint arXiv:1609.02907, 2016. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 126, + 553, + 179 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 126, + 553, + 179 + ], + "spans": [ + { + "bbox": [ + 316, + 126, + 553, + 179 + ], + "type": "text", + "content": "[16] Wei Li, Rui Zhao, Tong Xiao, and Xiaogang Wang. Deepreid: Deep filter pairing neural network for person re-identification. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 152–159, 2014. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 180, + 553, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 180, + 553, + 213 + ], + "spans": [ + { + "bbox": [ + 316, + 180, + 553, + 213 + ], + "type": "text", + "content": "[17] Fangyi Liu, Mang Ye, and Bo Du. Learning a generalizable re-identification model from unlabelled data with domain-agnostic expert. Visual Intelligence, 2(1):28, 2024. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 213, + 553, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 213, + 553, + 255 + ], + "spans": [ + { + "bbox": [ + 316, + 213, + 553, + 255 + ], + "type": "text", + "content": "[18] David Luebke. Cuda: Scalable parallel programming for high-performance scientific computing. In 2008 5th IEEE international symposium on biomedical imaging: from nano to macro, pages 836-838. IEEE, 2008. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 255, + 553, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 255, + 553, + 308 + ], + "spans": [ + { + "bbox": [ + 316, + 255, + 553, + 308 + ], + "type": "text", + "content": "[19] Chuanchen Luo, Yuntao Chen, Naiyan Wang, and Zhaoxiang Zhang. Spectral feature transformation for person re-identification. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4976–4985, 2019. 6, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 308, + 553, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 308, + 553, + 363 + ], + "spans": [ + { + "bbox": [ + 316, + 308, + 553, + 363 + ], + "type": "text", + "content": "[20] Hao Luo, Youzhi Gu, Xingyu Liao, Shenqi Lai, and Wei Jiang. Bag of tricks and a strong baseline for deep person re-identification. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 0–0, 2019. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 363, + 553, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 363, + 553, + 415 + ], + "spans": [ + { + "bbox": [ + 316, + 363, + 553, + 415 + ], + "type": "text", + "content": "[21] Jiaxu Miao, Yu Wu, Ping Liu, Yuhang Ding, and Yi Yang. Pose-guided feature alignment for occluded person re-identification. In Proceedings of the IEEE/CVF international conference on computer vision, pages 542–551, 2019. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 416, + 553, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 416, + 553, + 469 + ], + "spans": [ + { + "bbox": [ + 316, + 416, + 553, + 469 + ], + "type": "text", + "content": "[22] Sakrapee Paisitkriangkrai, Chunhua Shen, and Anton Van Den Hengel. Learning to rank in person re-identification with metric ensembles. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1846–1855, 2015. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 470, + 553, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 470, + 553, + 534 + ], + "spans": [ + { + "bbox": [ + 316, + 470, + 553, + 534 + ], + "type": "text", + "content": "[23] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019. 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 535, + 553, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 535, + 553, + 578 + ], + "spans": [ + { + "bbox": [ + 316, + 535, + 553, + 578 + ], + "type": "text", + "content": "[24] Ergys Ristani, Francesco Solera, Roger Zou, Rita Cucchiara, and Carlo Tomasi. Performance measures and a data set for multi-target, multi-camera tracking. In European conference on computer vision, pages 17-35. Springer, 2016. 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 578, + 553, + 642 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 578, + 553, + 642 + ], + "spans": [ + { + "bbox": [ + 316, + 578, + 553, + 642 + ], + "type": "text", + "content": "[25] M Saquib Sarfraz, Arne Schumann, Andreas Eberle, and Rainer Stiefelhagen. A pose-sensitive embedding for person re-identification with expanded cross neighborhood reranking. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 420–429, 2018. 1, 2, 4, 6, 7" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 643, + 553, + 686 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 643, + 553, + 686 + ], + "spans": [ + { + "bbox": [ + 316, + 643, + 553, + 686 + ], + "type": "text", + "content": "[26] Florian Schroff, Dmitry Kalenichenko, and James Philbin. Facenet: A unified embedding for face recognition and clustering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 815-823, 2015. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 686, + 553, + 707 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 686, + 553, + 707 + ], + "spans": [ + { + "bbox": [ + 316, + 686, + 553, + 707 + ], + "type": "text", + "content": "[27] Yantao Shen, Hongsheng Li, Shuai Yi, Dapeng Chen, and Xiaogang Wang. Person re-identification with deep" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19269" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 707 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 76, + 72, + 296, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 296, + 105 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 296, + 105 + ], + "type": "text", + "content": "similarity-guided graph neural network. In Proceedings of the European conference on computer vision (ECCV), pages 486-504, 2018. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 106, + 295, + 159 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 106, + 295, + 159 + ], + "spans": [ + { + "bbox": [ + 56, + 106, + 295, + 159 + ], + "type": "text", + "content": "[28] Yifan Sun, Changmao Cheng, Yuhan Zhang, Chi Zhang, Liang Zheng, Zhongdao Wang, and Yichen Wei. Circle loss: A unified perspective of pair similarity optimization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6398-6407, 2020. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 160, + 294, + 223 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 160, + 294, + 223 + ], + "spans": [ + { + "bbox": [ + 56, + 160, + 294, + 223 + ], + "type": "text", + "content": "[29] Yifan Sun, Qin Xu, Yali Li, Chi Zhang, Yikang Li, Shengjin Wang, and Jian Sun. Perceive where to focus: Learning visibility-aware part-level features for partial person re-identification. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 393–402, 2019. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 224, + 294, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 224, + 294, + 277 + ], + "spans": [ + { + "bbox": [ + 56, + 224, + 294, + 277 + ], + "type": "text", + "content": "[30] Yifan Sun, Liang Zheng, Yi Yang, Qi Tian, and Shengjin Wang. Beyond part models: Person retrieval with refined part pooling (and a strong convolutional baseline). In Proceedings of the European conference on computer vision (ECCV), pages 480-496, 2018. 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 278, + 294, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 278, + 294, + 331 + ], + "spans": [ + { + "bbox": [ + 56, + 278, + 294, + 331 + ], + "type": "text", + "content": "[31] Haochen Wang, Jiayi Shen, Yongtuo Liu, Yan Gao, and Efstratos Gavves. Nformer: Robust person re-identification with neighbor transformer. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7297-7307, 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 332, + 294, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 332, + 294, + 374 + ], + "spans": [ + { + "bbox": [ + 56, + 332, + 294, + 374 + ], + "type": "text", + "content": "[32] Yingjia Xu, Mengxia Wu, Zixin Guo, Min Cao, Mang Ye, and Jorma Laaksonen. Efficient text-to-video retrieval via multi-modal multi-tags derived pre-screening. Visual Intelligence, 3(1):1-13, 2025. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 375, + 294, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 375, + 294, + 418 + ], + "spans": [ + { + "bbox": [ + 56, + 375, + 294, + 418 + ], + "type": "text", + "content": "[33] Mang Ye, Shuoyi Chen, Chenyue Li, Wei-Shi Zheng, David Crandall, and Bo Du. Transformer for object re-identification: A survey. International Journal of Computer Vision, pages 1–31, 2024. 1, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 418, + 294, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 418, + 294, + 470 + ], + "spans": [ + { + "bbox": [ + 56, + 418, + 294, + 470 + ], + "type": "text", + "content": "[34] Mang Ye, Chao Liang, Zheng Wang, Qingming Leng, and Jun Chen. Ranking optimization for person re-identification via similarity and dissimilarity. In Proceedings of the 23rd ACM international conference on Multimedia, pages 1239–1242, 2015. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 472, + 294, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 472, + 294, + 525 + ], + "spans": [ + { + "bbox": [ + 56, + 472, + 294, + 525 + ], + "type": "text", + "content": "[35] Mang Ye, Jianbing Shen, Gaojie Lin, Tao Xiang, Ling Shao, and Steven CH Hoi. Deep learning for person re-identification: A survey and outlook. IEEE transactions on pattern analysis and machine intelligence, 44(6):2872-2893, 2021. 2, 5, 6, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 525, + 294, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 525, + 294, + 568 + ], + "spans": [ + { + "bbox": [ + 56, + 525, + 294, + 568 + ], + "type": "text", + "content": "[36] Mang Ye, Zesen Wu, and Bo Du. Dual-level matching with outlier filtering for unsupervised visible-infrared person re-identification. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 568, + 294, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 568, + 294, + 610 + ], + "spans": [ + { + "bbox": [ + 56, + 568, + 294, + 610 + ], + "type": "text", + "content": "[37] Xuanmeng Zhang, Minyue Jiang, Zhedong Zheng, Xiao Tan, Errui Ding, and Yi Yang. Understanding image retrieval reranking: A graph neural network perspective. arXiv preprint arXiv:2012.07620, 2020. 2, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 611, + 294, + 653 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 611, + 294, + 653 + ], + "spans": [ + { + "bbox": [ + 56, + 611, + 294, + 653 + ], + "type": "text", + "content": "[38] Yuqi Zhang, Qi Qian, Hongsong Wang, Chong Liu, Weihua Chen, and Fan Wang. Graph convolution based efficient reranking for visual retrieval. IEEE Transactions on Multimedia, 26:1089-1101, 2023. 1, 2, 3, 4, 6, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 654, + 294, + 696 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 654, + 294, + 696 + ], + "spans": [ + { + "bbox": [ + 56, + 654, + 294, + 696 + ], + "type": "text", + "content": "[39] Liang Zheng, Liyue Shen, Lu Tian, Shengjin Wang, Jingdong Wang, and Qi Tian. Scalable person re-identification: A benchmark. In Proceedings of the IEEE international conference on computer vision, pages 1116–1124, 2015. 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 696, + 294, + 707 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 696, + 294, + 707 + ], + "spans": [ + { + "bbox": [ + 56, + 696, + 294, + 707 + ], + "type": "text", + "content": "[40] Liang Zheng, Yi Yang, and Alexander G Hauptmann. Per" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 201 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "type": "text", + "content": "son re-identification: Past, present and future. arXiv preprint arXiv:1610.02984, 2016. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 95, + 553, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 95, + 553, + 148 + ], + "spans": [ + { + "bbox": [ + 316, + 95, + 553, + 148 + ], + "type": "text", + "content": "[41] Meng Zheng, Srikrishna Karanam, Ziyan Wu, and Richard J Radke. Re-identification with consistent attentive siamese networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5735–5744, 2019. 1, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 148, + 553, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 148, + 553, + 201 + ], + "spans": [ + { + "bbox": [ + 316, + 148, + 553, + 201 + ], + "type": "text", + "content": "[42] Zhun Zhong, Liang Zheng, Donglin Cao, and Shaozi Li. Reranking person re-identification with k-reciprocal encoding. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1318–1327, 2017. 1, 2, 4, 6, 7" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19270" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Chebyshev Attention Depth Permutation Texture Network with Latent Texture Attribute Loss/44aa142c-41b8-482e-a039-e4a4fc955e21_content_list.json b/2025/Chebyshev Attention Depth Permutation Texture Network with Latent Texture Attribute Loss/44aa142c-41b8-482e-a039-e4a4fc955e21_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..89bd375402bbc02c372fd8eddb518facb793d8c1 --- /dev/null +++ b/2025/Chebyshev Attention Depth Permutation Texture Network with Latent Texture Attribute Loss/44aa142c-41b8-482e-a039-e4a4fc955e21_content_list.json @@ -0,0 +1,1657 @@ +[ + { + "type": "text", + "text": "Chebyshev Attention Depth Permutation Texture Network with Latent Texture Attribute Loss", + "text_level": 1, + "bbox": [ + 204, + 130, + 794, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ravishankar Evani, Deepu Rajan, Shangbo Mao", + "bbox": [ + 285, + 204, + 709, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "College of Computing and Data Science, Nanyang Technological University, Singapore", + "bbox": [ + 238, + 223, + 756, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "S220007@e.ntu.edu.sg, ASDRajan@ntu.edu.sg, MAOS0003@e.ntu.edu.sg", + "bbox": [ + 274, + 243, + 718, + 255 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 291, + 326, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite recent advances in deep texture recognition, existing methods still lack representational diversity and struggle to capture and preserve discriminative cues across stages of representation hierarchies. Moreover, many rely on loss formulations that prioritize recognition accuracy while overlooking spatial coherence and statistical consistency in the feature space. To address these issues, we propose three key innovations: (1) Stochastic Local Texture Masking (SLTM), a regularization strategy that randomly occludes small texture patches to promote the learning of broader spatial and contextual dependencies; (2) the Chebyshev Attention Depth Permutation Texture Network (CAPTN), a novel architecture that learns expressive and persistent Latent Texture Attribute (LTA) representations. CAPTN integrates a Texture Frequency Attention (TFA) module that generates Latent Texture Attributes (LTAs) and enables frequency-aware interpretability, a Dual Depth Permutation $(D^{2}P)$ module to expose complementary channel adjacency patterns, and Learnable Chebyshev Polynomials (LCPs) to model high-order orderless LTA transformations via recursive Chebyshev basis expansion; and (3) a Latent Texture Attribute Loss that jointly optimizes classification accuracy, statistical alignment, and spatial fidelity. CAPTN supports end-to-end training without relying on fine-tuned CNN backbones and achieves state-of-the-art performance on several texture and material recognition benchmarks. (Code: https://github.com/RavishankarEvani/CAPTN)", + "bbox": [ + 88, + 325, + 485, + 750 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 782, + 220, + 797 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Visual texture perception plays a critical role in understanding an object's surface characteristics. Distinct from object recognition, which often relies on shape cues, texture recognition emphasizes the repetitive patterns or stochastic microstructures that define surface appearance. Over the years, texture recognition has evolved appreciably, be", + "bbox": [ + 89, + 809, + 483, + 902 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ginning with foundational work such as Julesz's theory of textons [14], Haralick et al.'s co-occurrence matrices [12], and Laws' texture energy measures based on convolution masks [16, 17]. These early methods established the basis for statistical and structural texture analysis and were widely adopted in classical image processing pipelines. However, while effective under controlled conditions, such traditional techniques often fail to capture the rich diversity and structural complexity of real-world textures.", + "bbox": [ + 511, + 292, + 906, + 429 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "With the rise of data-driven approaches, modern deep learning methods have largely supplanted handcrafted descriptors by offering more expressive and hierarchical feature representations. Despite this progress, deep models still encounter challenges when applied to texture recognition. In particular, successive convolution and pooling operations can suppress or discard fine-grained texture cues, limiting the network's ability to retain localized and discriminative information. Furthermore, textures often exhibit overlapping or perceptually similar traits [37], which contribute to inter-class confusion by blurring the boundaries between distinct texture categories. For instance, textures labeled as \"stratified\" might closely resemble those labeled \"cracked\" or \"zigzagged,\" making it difficult to consistently assign the correct class. In addition, intra-class variability arises from variations within the same category due to differences in lighting, scale, deformation, or inherent surface irregularities.", + "bbox": [ + 511, + 430, + 908, + 702 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The concept of texture attributes [10, 37, 39] has been shown to play a crucial role in representing perceptually similar textures. Latent texture attributes [10] refer to the underlying semantic components of visual texture attributes, where each latent texture attribute may correspond to one or more perceptual texture attributes. Effectively manipulating and transforming these latent attributes can enhance texture discrimination. Several studies employing Convolutional Neural Network (CNN)-based approaches have explored different aggregation methods, including global average pooling [1, 11, 37], Fisher vectors [7], feature encoding [35, 40], and cross-layer aggregation. However, certain aggregation and dimensionality", + "bbox": [ + 511, + 704, + 910, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "23423", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "reduction techniques lead to the loss of important texture cues, and little attention has been given to recovering discriminative information post-aggregation.", + "bbox": [ + 89, + 90, + 480, + 136 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address these limitations, we introduce a comprehensive texture recognition framework that advances both architectural design and training methodology. Central to our approach is the Chebyshev Attention Depth Permutation Texture Network (CAPTN), which extracts expressive and interpretable LTAs, depth permutation, and polynomial projection. In addition, we propose a stochastic local texture masking (SLTM) strategy and a tailored multi-objective loss function that together encourage spatially grounded and orderless texture representations. Our main contributions are as follows:", + "bbox": [ + 88, + 137, + 480, + 301 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose the CAPTN architecture, which integrates three key components:", + "- A Texture Frequency Attention (TFA) module that generates Latent Texture Attributes (LTAs) and enables visual interpretability by highlighting discriminative texture regions across frequency scales.", + "- A Dual Depth Permutation $(\\mathrm{D}^2\\mathrm{P})$ module that applies two distinct channel shuffling operations to expose complementary spatial relationships and enhance feature diversity.", + "- A Learnable Chebyshev Polynomial (LCP) expansion that enhances the representational power of each orderless LTA.", + "- We introduce a Stochastic Local Texture Masking (SLTM) strategy that randomly occludes small patches of the input image during training. This compels the model to leverage broader spatial context and enables a new supervision signal that aligns masked patches with regional latent features.", + "- We formulate a multi-objective LTA loss that jointly supervises classification accuracy, orderless latent attribute consistency, and spatial latent attribute alignment." + ], + "bbox": [ + 89, + 303, + 482, + 636 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 648, + 232, + 664 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Traditional Texture Recognition. Building upon the foundational work discussed in Sec. 1, influential methods such as the Bag of Textons (BoT) [9, 18] and Bag of Words (BoW) [8] introduced a paradigm for representing textures as orderless histograms over learned dictionaries of textural elements. These methods encode local patches, aggregate them into global descriptors, and classify the results using algorithms such as Support Vector Machines (SVM). Varma et al. [31, 32] advanced this line of work by learning texture descriptors directly from data. Subsequent developments led to residual encoding techniques—Jégou et al. [15] proposed VLAD, which captures first-order residuals relative to visual codewords, while Perronnin et al. [24] introduced Fisher Vectors (FV), a probabilistic extension that also models second-order statistics.", + "bbox": [ + 89, + 674, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Deep Texture Recognition. Deep learning-based texture recognition methods such as FV-CNN [7] and LFV [30] primarily focus on encoding features extracted from pretrained CNNs, but they do not support end-to-end training. To address this limitation, Zhang et al. [40] proposed DeepTEN, integrating residual encoding and dictionary learning into a single layer. The layer functions analogously to VLAD and FV while enabling end-to-end training with a CNN backbone. RPNet [22] simplifies Deep-TEN by replacing dictionary learning with residual pooling. DEP [35] enhances Deep-TEN further by combining orderless texture features with spatial cues to balance structure and texture.", + "bbox": [ + 511, + 90, + 903, + 271 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "MuLTER [13] improves DEP by integrating multi-layer CNN features, preserving fine texture and spatial details. MSBFEN [29] combines multiscale encoding and boosting to better capture discriminative texture features. Chen et al. [4] proposed using cross-layer statistical self-similarity to capture consistent patterns across CNN layers. Later, they introduced DTPNet [5], a local encoding scheme inspired by Local Binary Patterns (LBP), robust to rotation and grayscale changes. Yang et al. [36] proposed DFAEN, which uses frequency attention and first- and second-order statistics to enhance texture details. RADAM [27] fuses CNN activation maps via randomized aggregation, promoting texture diversity and reducing overfitting.", + "bbox": [ + 511, + 273, + 903, + 470 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "FENet [33] encodes spatial regularity via fractal analysis, while DSR-Net [38] models local spatial dependencies instead of orderless aggregation. Zhai et al. [37] earlier proposed MAP-Net, which learns texture-related visual attributes and uses spatially-adaptive pooling for feature fusion. Their recent work, MPAP [39], uses a multi-branch architecture to model bottom-up structures and top-down attribute relationships by learning local-to-global mappings. Reale et al. [25] proposed modeling triangular patches as graph nodes to encode texture patterns. More recently, GTN [10] introduced a graph-based deep approach designed to model directional relationships between Latent Texture Attributes (LTAs). GTN enhances LTAs through message passing and aggregation of semantically related information.", + "bbox": [ + 511, + 473, + 903, + 698 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Deep Polynomial Expansion. HistNet [23] features a trainable histogram layer that preserves spatial context using learnable bin centers and widths, combined with Radial Basis Functions (RBFs) for robustness to intra-class variability.", + "bbox": [ + 511, + 700, + 903, + 777 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "While graph-based methods explicitly learn attribute adjacencies for message passing to refine LTAs, our CAPTN framework adopts a more flexible and implicit approach. It enables rich and diverse feature interactions without the need for explicit graph construction. This not only simplifies the overall architecture but also implicitly regularizes the learning process by breaking fixed feature layouts, thereby enhancing texture robustness and generaliza", + "bbox": [ + 511, + 779, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "23424", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/30a2d67006fcfed2401ca74cc7386033f0c042feb7d6106c78018db81ed061f7.jpg", + "image_caption": [ + "Figure 1. Stochastic Local Texture Masking (SLTM) & components of Chebyshev Attention Depth Permutation Texture Network (CAPTN): SLTM facilitates contextual texture learning, Texture Frequency Attention (TFA) captures multi-frequency texture cues, Dual Depth Permutation $(\\mathrm{D}^2\\mathrm{P})$ promotes feature diversification, and Learnable Chebyshev Polynomials (LCP) map latent texture attributes to visual texture attributes. The training process optimizes Latent Texture Attribute (LTA) Loss, which balances classification accuracy with orderless and spatial texture consistency. || represents concatenation, $\\times$ represents matrix multiplied by a scalar." + ], + "image_footnote": [], + "bbox": [ + 94, + 111, + 911, + 354 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "tion. Unlike radial basis functions (RBFs), which often require numerous kernel centers and careful bandwidth tuning, we propose a Chebyshev polynomial-based method that provides a compact yet globally expressive basis. This approach eliminates the need for kernel-specific design choices while efficiently capturing complex and non-linear texture variations.", + "bbox": [ + 88, + 465, + 486, + 571 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Stochastic Local Texture Masking", + "text_level": 1, + "bbox": [ + 89, + 587, + 401, + 606 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To enhance the model's ability to capture local, global, and contextual texture representations, we introduce Stochastic Local Texture Masking (SLTM). This technique randomly masks a small patch of the input texture image prior to passing it through the pre-trained backbone, thereby preventing the Chebyshev Attention Depth Permutation Texture Network (CAPTN) from over-relying on fine-grained local details. Given an input image $\\mathbf{I} \\in \\mathbb{R}^{H \\times W \\times 3}$ , SLTM generates: (1) a masked image $\\mathbf{I}_r$ , where a randomly selected patch of size $P \\times P$ ( $P \\ll H, P \\ll W$ ) is zeroed out, and (2) the extracted patch $\\mathbf{G}_t \\in \\mathbb{R}^{P \\times P \\times 3}$ , which retains the original pixel values prior to masking. By selectively removing texture information, SLTM encourages CAPTN to exploit broader spatial dependencies, thereby enhancing global feature learning and reducing overfitting. Finally, the masked image $\\mathbf{I}_r$ is fed through the backbone and CAPTN for further processing. This module is activated during the computation of the Spatial Latent Texture Attribute Loss $(\\mathcal{L}_{SL})$ , as described in Section 5.", + "bbox": [ + 91, + 613, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Chebyshev Attention Depth Permutation Texture Network", + "text_level": 1, + "bbox": [ + 511, + 463, + 906, + 497 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The Chebyshev Attention Depth Permutation Texture Network (CAPTN) consists of three key modules: (1) Texture Frequency Attention (TFA) Network, (2) Dual Depth Permutation $(\\mathrm{D}^2\\mathrm{P})$ Network, and (3) Learnable Chebyshev Polynomials (LCP). Following [10], CAPTN extracts latent texture attributes (LTAs) by computing spatial texture frequency information from feature maps $\\mathcal{F} = \\{\\mathbf{X}^{o_m}\\mid 1\\leq m\\leq M - 1\\}$ obtained from an $M$ -layer pre-trained CNN backbone, excluding the patchify layer [19]. Depth permutation enhances LTA representation by diversifying feature space and reducing inter-class variability, while LCP improves expressiveness of orderless LTAs.", + "bbox": [ + 511, + 507, + 908, + 689 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1. Texture Frequency Attention", + "text_level": 1, + "bbox": [ + 511, + 696, + 772, + 713 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Textures display a wide range of spatial frequency characteristics, from high-frequency (fine-grained) patterns to low-frequency (smooth or homogeneous) regions. To effectively capture and interpret these variations, we propose a Texture Frequency Attention (TFA) mechanism. In addition to emphasizing discriminative texture regions, TFA also transforms backbone feature maps into Latent Texture Attributes (LTAs). Let $\\mathbf{X}^{o_m} \\in \\mathbb{R}^{H_m \\times W_m \\times C_m}$ be the $m^{\\text{th}}$ layer feature map. We compute transformed features as $\\mathbf{X}^{t_m} = \\mathbf{W}_l^m * \\mathbf{X}^{o_m} + \\mathbf{b}_l^m$ , where $\\mathbf{W}_l^m \\in \\mathbb{R}^{C_m \\times C_m \\times 1 \\times 1}$ is the convolution kernel, * denotes the convolution operation, and $\\mathbf{b}_l^m \\in \\mathbb{R}^{C_m}$ is a learnable bias vector. Next, we", + "bbox": [ + 511, + 719, + 908, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "23425", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/a67d04c227cb95c0ffaec0618099b40c232fc0bdd64550c26fd1f5280d1e9ebc.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 94, + 95, + 197, + 231 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/eeb164c67c23b70873281feae9075d050d251132db33e1e31961d72bbcb4857f.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 197, + 95, + 580, + 229 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/1ec616705a901841a4cbd66393e74b3dbad94522a53d96397e7e6030e4109e6a.jpg", + "image_caption": [ + "(c)" + ], + "image_footnote": [], + "bbox": [ + 581, + 95, + 898, + 229 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/85e080db0b7ffa8cfb51b76bdf45958f3a8d24a07db96c9a889df5e6f37366af.jpg", + "image_caption": [ + "(d)" + ], + "image_footnote": [], + "bbox": [ + 94, + 244, + 549, + 378 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/8065d32f9b6ac3509639a9700d48f47eee3e6c862290ffe33c4f252e3d8bafb9.jpg", + "image_caption": [ + "(e)", + "Figure 2. (a) SLTM: extract a small patch from an input texture image, (b) $\\mathrm{D}^2\\mathrm{P}$ : feature maps from backbone undergo depth permutation to diversify feature space, (c) LCP: enhanced Latent Texture Attributes undergo orderless aggregation and fit a learnable Chebyshev polynomial function, (d) TFA: use spatial texture frequencies to compute spatial attention, (e) SLAR: representation of spatial information for latent attributes and backbone features. SLTM and SLAR are used to compute Latent Texture Attribute Loss. || represents concatenation, $\\times$ represents matrix multiplied by a scalar, $\\cdot$ represents matrix multiplication." + ], + "image_footnote": [], + "bbox": [ + 550, + 244, + 901, + 378 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "compute per-location channel-wise maximum and average statistics", + "bbox": [ + 89, + 500, + 482, + 529 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nX _ {\\max } ^ {t _ {m}} (i, j) = \\max _ {0 \\leq c < C _ {m}} X ^ {t _ {m}} (i, j, c), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 161, + 563, + 482, + 587 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nX _ {\\text {a v g}} ^ {t _ {m}} (i, j) = \\frac {1}{C _ {m}} \\sum_ {c = 0} ^ {C _ {m} - 1} X ^ {t _ {m}} (i, j, c). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 169, + 590, + 480, + 632 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here, $\\mathbf{X}_{\\mathrm{max}}^{t_m}, \\mathbf{X}_{\\mathrm{avg}}^{t_m} \\in \\mathbb{R}^{H_m \\times W_m \\times 1}$ represent spatial maps of local maxima and average activations across channels. High-frequency textures tend to exhibit greater local variation, resulting in larger differences between maximum and average values, whereas low-frequency regions yield more consistent responses. To capture this contrast, we concatenate $\\mathbf{X}_{\\mathrm{max}}^{t_m}$ and $\\mathbf{X}_{\\mathrm{avg}}^{t_m}$ along the channel dimension and apply a convolution followed by a sigmoid activation", + "bbox": [ + 89, + 648, + 483, + 772 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {A} ^ {m} = \\frac {1}{1 + e ^ {- \\left[ \\mathbf {W} _ {a} ^ {m} * \\left(\\mathbf {X} _ {\\operatorname* {m a x}} ^ {t _ {m}} \\parallel \\mathbf {X} _ {\\operatorname* {a v g}} ^ {t _ {m}}\\right) \\right]}}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 171, + 791, + 482, + 825 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\parallel$ denotes channel-wise concatenation and $\\mathbf{W}_a^m\\in$ $\\mathbb{R}^{1\\times 2\\times 1\\times 1}$ is the convolution kernel. The resulting spatial attention map $\\mathbf{A}^m\\in \\mathbb{R}^{H_m\\times W_m\\times 1}$ is used to modulate $\\mathbf{X}^{t_m}$ via element-wise multiplication", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} X _ {\\text {a t t n}} ^ {t _ {m}} (i, j, c) = X ^ {t _ {m}} (i, j, c) \\times A ^ {m} (i, j), \\\\ \\forall (i, j, c) \\in [ 0, H _ {m} - 1 ] \\times [ 0, W _ {m} - 1 ] \\times [ 0, C _ {m} - 1 ], \\tag {4} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 539, + 521, + 903, + 570 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "A residual connection is then added, combining the attention-weighted features with the original $\\mathbf{X}^{t_m}$ . The result is passed through a transformation $g(\\cdot)$ comprising Batch Normalization, GELU activation, and dropout", + "bbox": [ + 511, + 571, + 905, + 631 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {X} _ {a} ^ {t _ {m}} = f _ {a} \\left(\\gamma_ {m}\\right) \\times g \\left(\\mathbf {X} _ {\\text {a t t n}} ^ {t _ {m}} + \\mathbf {X} ^ {t _ {m}}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 643, + 903, + 662 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\gamma_{m}$ is a learnable scalar, and $f_{a}(\\cdot)$ ensures the scaling factor remains within the range [0, 1]", + "bbox": [ + 511, + 667, + 905, + 698 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nf _ {a} (z) = \\min (\\max (0, z), 1), \\quad z \\in \\mathbb {R}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 581, + 712, + 903, + 729 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "While the convolution and $g(\\cdot)$ transformation generate the LTAs, the spatial attention map $\\mathbf{A}^m$ provides interpretable insights into the model's dynamic focus on spatially discriminative texture regions.", + "bbox": [ + 511, + 734, + 905, + 795 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Dual Depth Permutation", + "text_level": 1, + "bbox": [ + 511, + 801, + 740, + 819 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Dual Depth Permutation $(\\mathrm{D}^2\\mathrm{P})$ operates directly on backbone features to generate complementary representations through two distinct shuffling operations. Exposing varied adjacency patterns promotes rich cross-channel correlations. The outputs from both shuffling paths are later", + "bbox": [ + 511, + 824, + 905, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "23426", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "fused with LTAs from Sec. 4.1 to enhance representation diversity and improve texture robustness. Additionally, $\\mathrm{D}^2\\mathrm{P}$ serves as an implicit regularizer, reducing overfitting by disrupting fixed feature layouts. Drawing inspiration from constraining [2], $\\mathrm{D}^2\\mathrm{P}$ introduces multiple feature perspectives by permuting channel adjacency structures in a supervised setting. Let $\\mathbf{X}^{o_m} \\in \\mathbb{R}^{H_m \\times W_m \\times C_m}$ denote the feature map output from the $m^{\\text{th}}$ layer of the backbone. We apply two distinct channel shuffles using group sizes $s_0$ and $s_1$ , ensuring $\\frac{C_m}{s_0}, \\frac{C_m}{s_1} \\in \\mathbb{Z}$ . Let $p \\in \\{0,1\\}$ index the two shuffling paths, with corresponding group sizes $k_p^m \\in \\{s_0, s_1\\}$", + "bbox": [ + 89, + 90, + 483, + 258 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {X} ^ {o _ {m}} \\rightarrow \\mathbf {X} _ {\\mathrm {r}, p} ^ {o _ {m}} \\in \\mathbb {R} ^ {H _ {m} \\times W _ {m} \\times k _ {p} ^ {m} \\times g _ {k _ {p} ^ {m}}}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 161, + 273, + 482, + 296 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {X} _ {\\mathrm {r}, p} ^ {o _ {m}} \\rightarrow \\mathbf {X} _ {\\mathrm {r}, p} ^ {o _ {m} ^ {T}} \\in \\mathbb {R} ^ {H _ {m} \\times W _ {m} \\times g _ {k _ {p} ^ {m}} \\times k _ {p} ^ {m}}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 161, + 316, + 482, + 342 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $g_{k_p^m}$ is the number of groups formed using group size $k_{p}^{m}$ at layer $m$ . After flattening $\\mathbf{X}_{\\mathrm{r},p}^{o_m^T}$ , we obtain", + "bbox": [ + 89, + 352, + 483, + 391 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {X} _ {\\mathrm {s}, p} ^ {o _ {m}} \\in \\mathbb {R} ^ {H _ {m} \\times W _ {m} \\times C _ {m}}. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 205, + 406, + 482, + 426 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Each of the shuffled representations is processed by a convolution operation, where $\\mathbf{W}_p^m\\in \\mathbb{R}^{C_m\\times C_m\\times 1\\times 1}$ is the convolution kernel, $*$ denotes the convolution operation, and $\\mathbf{b}_p^m\\in \\mathbb{R}^{C_m}$ is a learnable bias vector. The result is then passed through a transformation $f_{p}(\\cdot)$ , consisting of Batch Normalization, GELU activation, and Dropout. The final output of the $\\mathrm{D}^2\\mathrm{P}$ module is obtained by summing the outputs from both shuffling paths", + "bbox": [ + 89, + 436, + 483, + 556 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {X} _ {\\mathrm {s}} ^ {d _ {m}} = f _ {a} (\\psi_ {m}) \\times \\left[ \\sum_ {p \\in \\{0, 1 \\}} f _ {p} \\left(\\mathbf {W} _ {p} ^ {m} * \\mathbf {X} _ {\\mathrm {s}, p} ^ {o _ {m}} + \\mathbf {b} _ {p} ^ {m}\\right) \\right],\n$$\n", + "text_format": "latex", + "bbox": [ + 104, + 583, + 464, + 633 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\psi_ {m} \\in (- \\infty , \\infty), \\quad f _ {a} (\\cdot) \\in [ 0, 1 ], \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 117, + 635, + 480, + 664 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\psi_{m}$ is a learnable scalar parameter, and the scaling function $f_{a}(\\cdot)$ is defined in Eq. (6) to constrain the modulation factor within the range [0, 1].", + "bbox": [ + 89, + 666, + 483, + 712 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Enhanced Orderless Latent Texture Attributes", + "text_level": 1, + "bbox": [ + 89, + 724, + 483, + 739 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "While GTN [10] enhances LTA representations via message passing and aggregation, CAPTN improves the LTAs generated by the TFA module, $\\mathbf{X}_a^{t_m}$ , by injecting dynamically adapted adjacency structures from the $\\mathrm{D}^2\\mathrm{P}$ module, $\\mathbf{X}_s^{d_m}$ , resulting in", + "bbox": [ + 89, + 748, + 483, + 823 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {X} _ {e} ^ {m} = \\mathbf {X} _ {a} ^ {t _ {m}} + \\mathbf {X} _ {\\mathrm {s}} ^ {d _ {m}}. \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 218, + 825, + 482, + 844 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We then compute the spatial mean over the height and width of the enhanced LTAs, $\\mathbf{X}_e^m\\in \\mathbb{R}^{H_m\\times W_m\\times C_m}$ , from layer $m$ to obtain an orderless representation, $\\mathbf{O}_{\\mathrm{e}}^{m}\\in \\mathbb{R}^{C_{m}}$ .", + "bbox": [ + 89, + 854, + 483, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.4. Learnable Chebyshev Polynomial", + "text_level": 1, + "bbox": [ + 513, + 90, + 808, + 106 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Chebyshev polynomials [26] offer a powerful global basis for modeling non-linear, oscillatory, and fine-grained variations in orderless LTAs. Their higher-order, non-monotonic structure enables expressive transformations that capture both smooth and oscillatory variations, allowing the model to distinguish subtle patterns while aligning semantically similar LTA values—even when numerically distant. To enhance the expressiveness of each orderless LTA dimension, we apply a Chebyshev polynomial $T_{n}(\\cdot)$ for $n \\in \\{0, \\dots, d\\}$ , each weighted by a sigmoid function parameterized by $\\theta_{n}$ . The resulting formulation, referred to as a Learnable Chebyshev Polynomial (LCP), takes the form", + "bbox": [ + 511, + 112, + 906, + 294 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nP \\left(\\mathbf {O} _ {\\mathrm {e}} ^ {m}\\right) = \\sum_ {n = 0} ^ {d} T _ {n} \\left(\\mathbf {O} _ {\\mathrm {e}} ^ {m}\\right) \\left[ \\frac {1}{1 + e ^ {- \\theta_ {n}}} \\right], \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 581, + 305, + 906, + 347 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The Chebyshev polynomials $T_{n}(\\mathbf{O}_{\\mathrm{e}}^{m})$ used in the expansion follow the recurrence relation", + "bbox": [ + 511, + 351, + 905, + 381 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nT _ {n} \\left(\\mathbf {O} _ {\\mathrm {e}} ^ {m}\\right) = \\left\\{ \\begin{array}{l l} \\mathbf {1}, & \\text {i f} n = 0, \\\\ \\mathbf {O} _ {\\mathrm {e}} ^ {m}, & \\text {i f} n = 1, \\\\ 2 \\mathbf {O} _ {\\mathrm {e}} ^ {m} T _ {n - 1} \\left(\\mathbf {O} _ {\\mathrm {e}} ^ {m}\\right) & \\text {f o r} n \\geq 2. \\\\ - T _ {n - 2} \\left(\\mathbf {O} _ {\\mathrm {e}} ^ {m}\\right), & \\end{array} \\right. \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 545, + 406, + 905, + 478 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "LCPs from all layers are then concatenated and passed through a fully connected layer, $f_{fc}$ , as", + "bbox": [ + 511, + 487, + 905, + 518 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {Y} = f _ {f c} \\left( \\right.\\left\\| \\right. _ {m = 1} ^ {N - 1} P \\left(\\mathbf {O} _ {\\mathrm {e}} ^ {m}\\right)\\left. \\right). \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 617, + 527, + 905, + 554 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The logits, $\\mathbf{Y}$ , are then evaluated using the Latent Texture Attribute Loss function, which will be discussed in Sec. 5.", + "bbox": [ + 511, + 558, + 905, + 588 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.5. Spatial Latent Attribute Representation", + "text_level": 1, + "bbox": [ + 511, + 597, + 856, + 614 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Finally, we aim to enforce some extent of structural similarity between a texture patch extracted from Sec. 3 and regional latent texture attribute information. To achieve this, we first concatenate the enhanced LTAs, $\\mathbf{X}_e^{M - 1}$ , with the original feature maps from the frozen backbone, $\\mathbf{X}^{O_M - 1}$ . Both representations are obtained from the final layer of the network. We integrate the frozen backbone features into the LTAs through a structured transformation consisting of three sequential operations: cropping $f_{c}(\\cdot)$ , a $3\\times 3$ convolution, and bilinear interpolation $f_{b}(\\cdot)$ . The transformation is formally defined as", + "bbox": [ + 511, + 619, + 906, + 785 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {G} _ {\\mathrm {p}} = f _ {b} \\left(\\mathbf {W} _ {s} * f _ {c} \\left(\\mathbf {X} _ {e} ^ {M - 1} \\| \\mathbf {X} ^ {O M - 1}\\right) + \\mathbf {b} _ {s}, P\\right), \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 539, + 811, + 906, + 830 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{W}_s\\in \\mathbb{R}^{3\\times C\\times 3\\times 3}$ and $\\mathbf{b}_s\\in \\mathbb{R}^3$ are the learnable weights and bias of the convolutional layer. The output of this transformation is the Spatial Latent Attribute Representation (SLAR), $\\mathbf{G}_{\\mathrm{p}}\\in \\mathbb{R}^{3\\times P\\times P}$ , where $P$ is the spatial size", + "bbox": [ + 511, + 838, + 906, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "23427", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "of the extracted patch. This ensures that the extracted regional texture attributes retain both the fine-grained information from LTAs and the structural guidance preserved in the frozen backbone features. Incorporating structure-aware features from the frozen backbone provides a grounded reference that constrains the adaptation of LTAs during training, preventing overcorrection while maintaining a balance between texture fidelity and spatial consistency. SLAR will be used to compute Spatial Latent Texture Attribute Loss (SL) in Sec. 5.", + "bbox": [ + 89, + 90, + 483, + 241 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Latent Texture Attribute Loss", + "text_level": 1, + "bbox": [ + 89, + 258, + 366, + 273 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We propose the Latent Texture Attribute (LTA) Loss to enhance classification accuracy while preserving texture consistency at both global and local levels. The LTA Loss integrates three complementary components: (1) Cross Entropy Loss $(\\mathcal{L}_{CE})$ to promote accurate class predictions, (2) Orderless LTA Loss $(\\mathcal{L}_{OL})$ to align global, spatially invariant texture distributions, and (3) Spatial LTA Loss $(\\mathcal{L}_{SL})$ to preserve fine-grained, spatial texture structure. By jointly optimizing these objectives, the model learns rich and coherent latent texture representations that support both accurate classification and texture understanding.", + "bbox": [ + 89, + 285, + 483, + 450 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Orderless Latent Texture Attribute Loss. This second texture loss component compares distribution of orderless LTAs with orderless features extracted directly from the frozen backbone. By aligning loss from learned texture representations with a stable baseline, we encourage consistency in the model's understanding of spatial invariant texture representations across images. We apply Batch Normalization, $f_{\\mathrm{bn}}$ , on feature maps from the $m^{\\mathrm{th}}$ layer of the backbone, $\\mathbf{X}^{o_m}$ , before calculating the spatial mean and concatenating multiple scales of orderless texture information across layers $m \\in [1, M - 1]$ of the backbone to obtain $\\mathbf{O}_{\\mathrm{b}}$ . We also concatenate orderless enhanced LTAs, $\\mathbf{O}_{\\mathrm{e}}^{m}$ , across the same range of layers to obtain $\\mathbf{O}_{\\mathrm{e}}$", + "bbox": [ + 89, + 452, + 483, + 648 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {O} _ {\\mathrm {b}} = \\left\\| \\right. _ {m = 1} ^ {M - 1} \\left(\\frac {\\sum_ {h = 0} ^ {H _ {m} - 1} \\sum_ {w = 0} ^ {W _ {m} - 1} f _ {\\mathrm {b n}} \\left(\\mathbf {X} ^ {o _ {m}}\\right) [ h , w , : ]}{H _ {m} \\times W _ {m}}\\right), \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 109, + 672, + 483, + 762 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {O} _ {\\mathrm {e}} = \\left\\| _ {m = 1} ^ {M - 1} \\mathbf {O} _ {\\mathrm {e}} ^ {m}. \\right. \\tag {17}\n$$\n", + "text_format": "latex", + "bbox": [ + 228, + 776, + 483, + 816 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We perform softmax normalization of $\\mathbf{O_b}$ and $\\mathbf{O_e}$ over the total number of channels in each sample $s$ from the minibatch, aggregated over layers $m\\in [1,M - 1]$ , where the total number of concatenated channels is $C = \\sum_{m = 1}^{M - 1}C_m$ . A small constant $\\epsilon$ is added for numerical stability", + "bbox": [ + 89, + 823, + 483, + 901 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\np _ {\\alpha , i} ^ {s} = \\frac {e ^ {\\mathbf {O} _ {\\alpha , i} ^ {s}}}{\\sum_ {j = 0} ^ {C - 1} e ^ {\\mathbf {O} _ {\\alpha , j} ^ {s}}} + \\epsilon , \\quad \\alpha \\in \\{\\mathrm {b}, \\mathrm {e} \\}. \\tag {18}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 102, + 903, + 142 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Inspired by Kullback-Leibler (KL) divergence, we calculate orderless LTA loss, $\\mathcal{L}_{OL}$ that measures the dissimilarity between two probability distributions: a fixed distribution from orderless feature maps extracted from the frozen backbone, $p_{\\mathrm{b},i}^{\\mathrm{s}}$ and that from orderless enhanced LTAs, $p_{\\mathrm{e},i}^{\\mathrm{s}}$", + "bbox": [ + 511, + 146, + 906, + 224 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {O L} = \\frac {1}{S} \\sum_ {s = 0} ^ {S - 1} \\sum_ {i = 0} ^ {C - 1} p _ {\\mathrm {b}, i} ^ {s} \\left[ \\log_ {e} \\left(p _ {\\mathrm {b}, i} ^ {s}\\right) - \\log_ {e} \\left(p _ {\\mathrm {e}, i} ^ {s}\\right) \\right], \\tag {19}\n$$\n", + "text_format": "latex", + "bbox": [ + 531, + 244, + 903, + 287 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $S$ represents the number of samples in the minibatch.", + "bbox": [ + 511, + 297, + 903, + 325 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Spatial Latent Texture Attribute Loss. The final LTA loss component measures the similarity between SLAR from Sec. 4.5 and a corresponding texture patch extracted from the original image. This component helps ensure that the latent representations preserve both regional and local spatial texture information by comparing the latent attributes $\\mathbf{G}_{\\mathrm{p}}$ with the extracted texture patch $\\mathbf{G}_{\\mathrm{t}}$ using the Mean Squared Error (MSE)", + "bbox": [ + 511, + 328, + 906, + 449 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {S L} = \\frac {1}{B C _ {M - 1} H _ {M - 1} W _ {M - 1}} \\| \\mathbf {G} _ {\\mathrm {p}} - \\mathbf {G} _ {\\mathrm {t}} \\| _ {F} ^ {2}, \\tag {20}\n$$\n", + "text_format": "latex", + "bbox": [ + 547, + 472, + 903, + 503 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $B, C_{M-1}, H_{M-1}, W_{M-1}$ represent the batch size, number of channels, height, and width of the feature maps, from the final $M-1$ layer, and $\\|\\cdot\\|_F$ denotes the Frobenius norm, which sums the squared differences over all elements. This loss encourages the model to preserve fine-grained spatial texture patterns within the latent space, ensuring that learned texture features remain spatially structured and contextually coherent. By aligning latent spatial representations with real texture patches, the model enhances its ability to reconstruct or classify textures with high fidelity.", + "bbox": [ + 511, + 513, + 905, + 665 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Latent Texture Attribute Loss. We formulate our loss during the training phase as a weighted sum of the three loss components", + "bbox": [ + 511, + 666, + 905, + 710 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {t r a i n}} = \\lambda \\mathcal {L} _ {C E} + \\mu \\mathcal {L} _ {O L} + \\nu \\mathcal {L} _ {S L}, \\tag {21}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 726, + 903, + 742 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\lambda, \\mu, \\nu \\in [0,1]$ are weights associated with each of the loss components.", + "bbox": [ + 511, + 750, + 906, + 779 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6. Experiments", + "text_level": 1, + "bbox": [ + 511, + 792, + 645, + 809 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6.1. Datasets", + "text_level": 1, + "bbox": [ + 511, + 818, + 616, + 832 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We evaluated our network using five benchmark datasets, summarized as follows: (a) Ground Terrain in Outdoor Scenes (GTOS) [34] contains 34,105 images of outdoor ground terrain across 40 categories, with 5 train/test splits.", + "bbox": [ + 511, + 839, + 905, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "23428", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Comparison of CAPTN with GTN [10] and RADAM [27] in terms of classification accuracy (\\%) on texture benchmarks. ConvNeXt-\\* variants are used as backbones. Pre-trained weights are frozen. The best result for each backbone-dataset combination is bolded.", + "bbox": [ + 89, + 88, + 549, + 143 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/22d5ac38b3ee23a8550022f2ee47a9506a543771b56ccb0012600b1763cbfb09.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Model-*DTDFMDKTHGTOSGTOS-M
RADAMN74.9 ±0.787.1 ±0.489.6 ±3.883.7 ±1.581.8
GTNN75.4 ±1.186.9 ±3.190.3 ±4.782.9 ±1.790.1
CAPTN (ours)N77.9 ±1.087.0 ±2.990.6 ±4.683.9 ±1.791.5
RADAMT77.0 ±0.788.7 ±0.490.7 ±4.084.2 ±1.785.3
GTNT77.5 ±1.087.2 ±3.492.0 ±4.984.7 ±1.691.9
CAPTN (ours)T78.9 ±0.887.2 ±3.692.2 ±4.485.1 ±1.794.2
RADAMB76.4 ±0.990.2 ±0.287.7 ±5.684.1 ±1.682.2
GTNB78.3 ±0.888.8 ±3.090.8 ±6.284.2 ±1.894.8
CAPTN (ours)B79.3±0.989.1 ±3.491.3±5.784.6 ±1.994.0
RADAML77.4 ±1.189.3±0.389.3 ±3.484.0 ±1.885.8
GTNL79.3 ±0.888.1 ±3.092.5 ±4.784.9 ±1.793.9
CAPTN (ours)L79.6±0.889.3±3.193.0 ±4.985.0 ±1.795.1
", + "bbox": [ + 91, + 155, + 547, + 316 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/59cd6d53e5c1d0274ce82752456819d0a915518f60fc5ad086a3ebb8c56de2b4.jpg", + "table_caption": [ + "Table 4. Classification accuracy (\\%) comparison between $1^{\\text{st}}$ -order mean pooling (w/ LCP) and a non-polynomial baseline: $2^{\\text{nd}}$ -order covariance pooling (w/o LCP), with ConvNeXt-B backbone." + ], + "table_footnote": [], + "table_body": "
PoolingDTDKTHGTOSGTOS-M
2nd-order76.4 ±1.091.1 ±6.383.5 ±1.793.0
1st-order79.3 ±0.991.3 ±5.784.6 ±1.994.0
", + "bbox": [ + 101, + 409, + 473, + 457 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/3a5f2940e19b1d8dbe3cd80ff01530337198ce75cd0daaedde2d71e42d0f4056.jpg", + "table_caption": [ + "Table 5. Classification accuracy (\\%) of CAPTN with different basis functions using ConvNeXt-L backbone." + ], + "table_footnote": [], + "table_body": "
Basis Fn.DTDKTHGTOSGTOS-M
RBF74.1 ±1.089.6 ±6.881.5 ±2.492.7
LCP79.6±0.893.0±4.985.0±1.795.1
", + "bbox": [ + 112, + 510, + 460, + 556 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "(b) GTOS-Mobile (GTOS-M) [35] includes 100,011 material samples from 31 categories, captured via mobile phone, with a single train/test split. (c) KTH-TIPS2-b (KTH) [3] provides 4,752 images in 11 material categories, with 4 train/test splits. (d) Flickr Material Dataset (FMD) [28] contains 1,000 images spanning 10 material categories, validated with 10-fold cross-validation and stratified sampling to maintain class proportions. (e) Describable Texture Dataset (DTD) [6] has 5,640 images across 47 categories, with 10 train/validation/test splits.", + "bbox": [ + 88, + 580, + 482, + 733 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.2. Implementation Details", + "text_level": 1, + "bbox": [ + 89, + 742, + 307, + 758 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our model was implemented in PyTorch and trained on a single NVIDIA RTX A6000 GPU. For all datasets, the learning rate was set to 0.0001 and followed a cosine annealing schedule [20]. We used the Adam optimizer with decoupled weight decay [21] and our novel Latent Texture Attribute Loss function. Our network was trained until the final epoch for each dataset, with test set performance reported. More design and implementation details can be found in the supplementary material.", + "bbox": [ + 88, + 763, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/92367dda5669648f81dfe51ef5152732d57e8731e15d3b4015d0856c9e29efd9.jpg", + "table_caption": [ + "Table 2. Evaluate effectiveness of TFA, ${\\mathrm{D}}^{2}\\mathrm{P}$ ,and LCP in terms of classification accuracy (%) . ConvNeXt-N is the backbone." + ], + "table_footnote": [], + "table_body": "
TFAD2PLCPDTDKTHGTOS
XXX75.4 ±1.087.2 ±5.381.7 ±2.0
✓XX77.1 ±1.190.4 ±4.883.5 ±1.8
✓✓X77.2 ±1.090.5 ±4.883.6 ±1.8
✓✓✓77.9 ±1.090.6 ±4.683.9 ±1.7
", + "bbox": [ + 560, + 142, + 903, + 210 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/2b7b51b2bfae0b142059a992b183e62f9887aac8c82698ef8ff4fabdf203f65e.jpg", + "table_caption": [ + "Table 3. Evaluate effectiveness of LTA Loss in terms of classification accuracy $(\\%)$ . ConvNeXt-T is the backbone." + ], + "table_footnote": [], + "table_body": "
Loss Fn.DTDKTHGTOSGTOS-M
CE Loss78.4 ±0.791.9 ±5.185.0 ±1.893.6
LTA Loss78.9 ±0.892.2 ±4.485.1 ±1.794.2
", + "bbox": [ + 560, + 265, + 903, + 311 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.3. Analysis", + "text_level": 1, + "bbox": [ + 513, + 357, + 614, + 373 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Recent studies [10, 27] have demonstrated the superiority of ConvNeXt backbones over ResNet backbones for texture-specific applications, particularly in multiscale texture representation. Hence, all our experiments will be based on ConvNeXt backbones.", + "bbox": [ + 511, + 386, + 903, + 460 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Quantitative. Tab. 1 depicts how CAPTN's performance compares with the current state-of-the-art RADAM [27] and GTN [10] across various ConvNeXt backbones pretrained on ImageNet. The average and standard deviation measures are calculated across the provided splits, where applicable. Following RADAM and GTN, we have frozen our ConvNeXt backbones. CAPTN out-performed both RADAM [27] and GTN [10] in most of the benchmark datasets across different ConvNeXt backbone sizes. Both RADAM and GTN do not have a mechanism to represent orderless texture features or attributes. On the other hand, CAPTN represents orderless latent texture attributes as a learnable polynomial improving its expressiveness and consequently optimizing the mapping from latent to visual texture attributes.", + "bbox": [ + 511, + 462, + 905, + 686 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Qualitative. The illustrations in Fig. 3 were generated using the ConvNeXt-L backbone with images from the DTD dataset. The Texture Frequency Attention (TFA) module from CAPTN enhances spatial attention in regions containing specific texture features. To some extent, the attention mechanism can identify the spatial regularity of grid textures. For spiraled textures, which exhibit greater irregularity, CAPTN demonstrates robustness to variations in illumination, shadows, occlusion, and background textures by selectively focusing on the spiraled region. Similarly, for polka-dotted textures, CAPTN disregards occlusions and concentrates on the textured areas. Furthermore, in all three cases, CAPTN recognizes masked regions within the images and assigns them lower importance.", + "bbox": [ + 511, + 688, + 905, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "23429", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/cf5dd0f9a5962ca2a49fe2d0416fd808563331fbbdd16d9a9e5e1e2d8d406111.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 91, + 87, + 197, + 169 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/78357eb568dbc8edd52514c0e6f6220560f14afc539096005be5c2bea3f375fb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 233, + 85, + 343, + 169 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/17385c32af57754918167cb618792d53f7fc1666d31273ef4740a722708fb5c6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 380, + 88, + 482, + 167 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a507883055b66b37a41b6c0d3be98b0f1b6ab9a7f7b009afeccd169e70229056.jpg", + "image_caption": [ + "Figure 3. Top row: Masked images from DTD dataset (grid, spiralled, polka dotted). Bottom row: Corresponding attention maps from Texture Frequency Attention (TFA). Red indicates high attention, blue indicates low attention." + ], + "image_footnote": [], + "bbox": [ + 91, + 181, + 194, + 261 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a67e76df2cf24ac9ffc717c5f99f7819613ecd321ff533703a89a8a2c8d59c45.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 233, + 181, + 339, + 261 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0c86009d5e3e61436ceaa5fc8704ad87976cae425ddc1000d0b48cadec304ffd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 380, + 183, + 482, + 261 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6.4. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 356, + 243, + 372 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We conducted ablation studies to evaluate the impact of TFA, $\\mathrm{D}^2\\mathrm{P}$ , and LCP in CAPTN. Overall, all components improve CAPTN's performance on both texture (DTD) and material (KTH, GTOS) datasets. The quantitative results are presented in Tab. 2. First, we assessed LCP by mapping orderless LTAs to visual texture attributes without using LCP. The deterioration of results demonstrated that LCP enhances expressiveness and injects non-linearities, improving the mapping of latent to visual texture attributes. Removing $\\mathrm{D}^2\\mathrm{P}$ led to a slight performance drop, as the LTA space became less robust and diverse. Finally, the removal of TFA further degraded CAPTN's performance, indicating that backbone features alone are less discriminative than LTAs. In addition, we performed an ablation study Tab. 3 to understand the effectiveness of LTA loss compared with Cross Entropy (CE) loss. The LTA loss function enhanced performance across texture and material datasets, underscoring its effectiveness for their recognition.", + "bbox": [ + 89, + 378, + 483, + 652 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6.5. Alternative Design Choices", + "text_level": 1, + "bbox": [ + 89, + 665, + 333, + 681 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Non-polynomial baselines. Table 4 demonstrates that first-order mean pooling combined with a LCP transformation outperforms the non-polynomial baseline of second-order covariance pooling. Covariance pooling captures second-order statistics by modeling inter-LTA correlations computed over spatial activations. While this can encode useful dependencies between LTA dimensions, it is also sensitive to noise, small artifacts, and spatial misalignments due to its reliance on precise variation patterns across spatial locations. In contrast, mean pooling followed by an LCP transformation offers a more robust encoding alternative. Mean pooling summarizes each LTA by computing the first-order global average across spatial dimensions, producing compact and stable representations that are less affected by lo", + "bbox": [ + 89, + 688, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "cal distortions. These pooled features serve as clean inputs to the LCP module, which applies expressive, non-linear transformations to each LTA independently using Chebyshev polynomials. This combination enhances the representational power of each LTA without the noise sensitivity inherent in second-order pooling. Together, mean pooling and LCP offer a robust and discriminative representation strategy that outperforms covariance pooling.", + "bbox": [ + 511, + 90, + 903, + 210 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Alternative Polynomial Bases. We replace the LCP module in CAPTN with a learnable RBF layer using Gaussian kernels. It computes scaled Euclidean distances to learnable centers and aggregates the activations via a learnable weight matrix. To avoid underfitting, we add a linear projection, LayerNorm, and GELU activation, along with kernel diversity for better generalization. As shown in Table 5, LCP outperforms the RBF variant across a range of texture and material datasets, while also requiring fewer learnable parameters. On the GTOS-M dataset, for example, CAPTN with a third-order LCP requires only 9.60M parameters, compared to 14.22M for its RBF-based counterpart. Although Gaussian RBFs are effective at modeling local variations, they often require many kernel centers and careful bandwidth tuning to capture diverse texture patterns, increasing computational cost. In contrast, LCPs form a global polynomial basis that efficiently captures long-range dependencies and oscillatory patterns with fewer parameters. In addition to RBFs, we also considered B-splines, which are widely used for smooth, piecewise function approximation. However, B-splines require careful knot placement and segmentwise assembly to represent multi-peaked or oscillatory signals, making them more sensitive to hyperparameter tuning and less scalable. LCPs, by comparison, provide a unified, smooth basis without discontinuities at segment boundaries, naturally capturing global texture variations while preserving sharp gradients and repetitive motifs. These properties make LCPs a more compact, scalable, and interpretable choice than both RBFs and B-splines for texture-specific applications.", + "bbox": [ + 511, + 212, + 906, + 664 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 678, + 633, + 694 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We propose a comprehensive framework for robust texture representation by integrating SLTM with CAPTN. SLTM enhances generalization by mitigating over-reliance on fine-grained details, while the TFA mechanism captures multifrequency texture cues to support visual interpretation of spatial texture information during the generation of LTAs. Furthermore, the $\\mathrm{D}^2\\mathrm{P}$ module enriches LTAs by diversifying feature representations, and LCP is specifically designed to enable a richer and more expressive encoding of orderless LTAs. To further strengthen the learning process, we introduce a unified LTA loss function that balances classification objectives with orderless alignment and spatial consistency.", + "bbox": [ + 511, + 703, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "23430", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Vincent Andrearczyk and Paul F Whelan. Using filter banks in convolutional neural networks for texture classification. Pattern Recognition Letters, 84:63-69, 2016. 1", + "[2] Avrim Blum and Tom Mitchell. Combining labeled and unlabeled data with co-training. In Proceedings of the eleventh annual conference on Computational learning theory, pages 92-100, 1998. 5", + "[3] Barbara Caputo, Eric Hayman, and P. Mallikarjuna. Class-specific material categorisation. In Tenth IEEE International Conference on Computer Vision (ICCV'05) Volume 1, pages 1597-1604 Vol. 2, 2005. 7", + "[4] Zhile Chen, Feng Li, Yuhui Quan, Yong Xu, and Hui Ji. Deep texture recognition via exploiting cross-layer statistical self-similarity. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5227-5236, 2021. 2", + "[5] Zhile Chen, Yuhui Quan, Ruotao Xu, Lianwen Jin, and Yong Xu. Enhancing texture representation with deep tracing pattern encoding. Pattern Recognition, 146:109959, 2024. 2", + "[6] Mircea Cimpoi, Subhransu Maji, Iasonas Kokkinos, Sammy Mohamed, and Andrea Vedaldi. Describing textures in the wild. In 2014 IEEE Conference on Computer Vision and Pattern Recognition, pages 3606-3613, 2014. 7", + "[7] Mircea Cimpoi, Subhransu Maji, and Andrea Vedaldi. Deep filter banks for texture recognition and segmentation. In 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 3828-3836, 2015. 1, 2", + "[8] Gabriella Csurka, Christopher R. Dance, Lixin Fan, Jutta Willamowski, and Cedric Bray. Visual categorization with bags of keypoints. In In Workshop on Statistical Learning in Computer Vision, ECCV, pages 1-22, 2004. 2", + "[9] O.G. Cula and K.J. Dana. Compact representation of bidirectional texture functions. In Proceedings of the 2001 IEEE Computer Society Conference on Computer Vision and Pattern Recognition. CVPR 2001, pages I-I, 2001. 2", + "[10] Ravishankar Evani, Deepu Rajan, and Shangbo Mao. Multiscale graph texture network. In Computer Vision - ECCV 2024, pages 218-235, Cham, 2025. Springer Nature Switzerland. 1, 2, 3, 5, 7", + "[11] Shin Fujieda, Kohei Takayama, and Toshiya Hachisuka. Wavelet convolutional neural networks for texture classification. arXiv preprint arXiv:1707.07394, 2017. 1", + "[12] Robert M. Haralick, K. Shanmugam, and Its'Hak Dinstein. Textural features for image classification. IEEE Transactions on Systems, Man, and Cybernetics, SMC-3(6):610-621, 1973. 1", + "[13] Yuting Hu, Zhiling Long, and Ghassan AlRegib. Multi-level texture encoding and representation (multer) based on deep neural networks. In 2019 IEEE International Conference on Image Processing (ICIP), pages 4410-4414, 2019. 2", + "[14] Béla Julesz. Visual pattern discrimination. IRE transactions on Information Theory, 8(2):84-92, 1962. 1", + "[15] Hervé Jégou, Matthijs Douze, Cordelia Schmid, and Patrick Pérez. Aggregating local descriptors into a compact image representation. In 2010 IEEE Computer Society Conference" + ], + "bbox": [ + 93, + 114, + 482, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "on Computer Vision and Pattern Recognition, pages 3304-3311, 2010. 2", + "[16] Kenneth I Laws. Texture energy measures. In Proceedings of Image Understanding Workshop, 1979. 1", + "[17] Kenneth I. Laws. Rapid Texture Identification. In Image Processing for Missile Guidance, pages 376 - 381. International Society for Optics and Photonics, SPIE, 1980. 1", + "[18] Thomas Leung and Jitendra Malik. Representing and recognizing the visual appearance of materials using three-dimensional textons. International journal of computer vision, 43:29-44, 2001. 2", + "[19] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11966-11976, 2022. 3", + "[20] Ilya Loshchilov and Frank Hutter. SGDR: stochastic gradient descent with warm restarts. In 5th International Conference on Learning Representations, ICLR 2017, Toulouse, France, 2017. 7", + "[21] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In 7th International Conference on Learning Representations, ICLR 2019, New Orleans, LA, USA, 2019. 7", + "[22] Shangbo Mao, Deepu Rajan, and Liang Tien Chia. Deep residual pooling network for texture recognition. Pattern Recognition, 112:107817, 2021. 2", + "[23] Joshua Peeples, Weihuang Xu, and Alina Zare. Histogram layers for texture analysis. IEEE Transactions on Artificial Intelligence, 3(4):541-552, 2022. 2", + "[24] Florent Perronnin and Christopher Dance. Fisher kernels on visual vocabularies for image categorization. In 2007 IEEE Conference on Computer Vision and Pattern Recognition, pages 1-8, 2007. 2", + "[25] Michael J. Reale, Matthew Ochrym, Micah Church, Nicholas Goutermout, John Rubado, and Maria Cornacchia. Shape and texture aware graph processing. In 2020 IEEE Applied Imagery Pattern Recognition Workshop (AIPR), pages 1-18, 2020. 2", + "[26] Theodore J Rivlin. Chebyshev polynomials. Courier Dover Publications, 2020. 5", + "[27] Leonardo Scabini, Kallil M. Zielinski, Lucas C. Ribas, Wesley N. Gonçalves, Bernard De Baets, and Odemir M. Bruno. Radam: Texture recognition through randomized aggregated encoding of deep activation maps. Pattern Recognition, 143: 109802, 2023. 2, 7", + "[28] Lavanya Sharan, Ruth Rosenholtz, and Edward Adelson. Material perception: What can you see in a brief glance? Journal of Vision, 9(8):784-784, 2009. 7", + "[29] Kaiyou Song, Hua Yang, and Zhouping Yin. Multi-scale boosting feature encoding network for texture recognition. IEEE Transactions on Circuits and Systems for Video Technology, 31(11):4269-4282, 2021. 2", + "[30] Yang Song, Fan Zhang, Qing Li, Heng Huang, Lauren J. O'Donnell, and Weidong Cai. Locally-transferred fisher vectors for texture classification. In 2017 IEEE International Conference on Computer Vision (ICCV), pages 4922-4930, 2017. 2" + ], + "bbox": [ + 516, + 93, + 903, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "23431", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[31] Manik Varma and Andrew Zisserman. Classifying images of materials: Achieving viewpoint and illumination independence. In Computer Vision — ECCV 2002, pages 255–271, Berlin, Heidelberg, 2002. Springer Berlin Heidelberg. 2", + "[32] Manik Varma and Andrew Zisserman. A statistical approach to texture classification from single images. International journal of computer vision, 62:61-81, 2005. 2", + "[33] Yong Xu, Feng Li, Zhile Chen, Jinxiu Liang, and Yuhui Quan. Encoding spatial distribution of convolutional features for texture representation. In Advances in Neural Information Processing Systems, pages 22732-22744. Curran Associates, Inc., 2021. 2", + "[34] Jia Xue, Hang Zhang, Kristin Dana, and Ko Nishino. Differential angular imaging for material recognition. In 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 6940-6949, 2017. 6", + "[35] Jia Xue, Hang Zhang, and Kristin Dana. Deep texture manifold for ground terrain recognition. In 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 558-567, 2018. 1, 2, 7", + "[36] Zhijing Yang, Shujian Lai, Xiaobin Hong, Yukai Shi, Yongqiang Cheng, and Chunmei Qing. Dfaen: Double-order knowledge fusion and attentional encoding network for texture recognition. Expert Systems with Applications, 209: 118223, 2022. 2", + "[37] Wei Zhai, Yang Cao, Jing Zhang, and Zheng-Jun Zha. Deep multiple-attribute-perceived network for real-world texture recognition. In 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 3612–3621, 2019. 1, 2", + "[38] Wei Zhai, Yang Cao, Zheng-Jun Zha, HaiYong Xie, and Feng Wu. Deep structure-revealed network for texture recognition. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11007-11016, 2020. 2", + "[39] Wei Zhai, Yang Cao, Jing Zhang, Haiyong Xie, Dacheng Tao, and Zheng-Jun Zha. On exploring multiplicity of primitives and attributes for texture recognition in the wild. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(1):403-420, 2024. 1, 2", + "[40] Hang Zhang, Jia Xue, and Kristin Dana. Deep ten: Texture encoding network. In 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2896-2905, 2017. 1, 2" + ], + "bbox": [ + 91, + 92, + 482, + 696 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "23432", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Chebyshev Attention Depth Permutation Texture Network with Latent Texture Attribute Loss/44aa142c-41b8-482e-a039-e4a4fc955e21_model.json b/2025/Chebyshev Attention Depth Permutation Texture Network with Latent Texture Attribute Loss/44aa142c-41b8-482e-a039-e4a4fc955e21_model.json new file mode 100644 index 0000000000000000000000000000000000000000..b04295a306c518fb0a4421debb32928ab2e6627f --- /dev/null +++ b/2025/Chebyshev Attention Depth Permutation Texture Network with Latent Texture Attribute Loss/44aa142c-41b8-482e-a039-e4a4fc955e21_model.json @@ -0,0 +1,2167 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.205, + 0.131, + 0.795, + 0.175 + ], + "angle": 0, + "content": "Chebyshev Attention Depth Permutation Texture Network with Latent Texture Attribute Loss" + }, + { + "type": "text", + "bbox": [ + 0.287, + 0.205, + 0.71, + 0.222 + ], + "angle": 0, + "content": "Ravishankar Evani, Deepu Rajan, Shangbo Mao" + }, + { + "type": "text", + "bbox": [ + 0.239, + 0.224, + 0.758, + 0.24 + ], + "angle": 0, + "content": "College of Computing and Data Science, Nanyang Technological University, Singapore" + }, + { + "type": "text", + "bbox": [ + 0.275, + 0.244, + 0.719, + 0.256 + ], + "angle": 0, + "content": "S220007@e.ntu.edu.sg, ASDRajan@ntu.edu.sg, MAOS0003@e.ntu.edu.sg" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.292, + 0.327, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.326, + 0.486, + 0.751 + ], + "angle": 0, + "content": "Despite recent advances in deep texture recognition, existing methods still lack representational diversity and struggle to capture and preserve discriminative cues across stages of representation hierarchies. Moreover, many rely on loss formulations that prioritize recognition accuracy while overlooking spatial coherence and statistical consistency in the feature space. To address these issues, we propose three key innovations: (1) Stochastic Local Texture Masking (SLTM), a regularization strategy that randomly occludes small texture patches to promote the learning of broader spatial and contextual dependencies; (2) the Chebyshev Attention Depth Permutation Texture Network (CAPTN), a novel architecture that learns expressive and persistent Latent Texture Attribute (LTA) representations. CAPTN integrates a Texture Frequency Attention (TFA) module that generates Latent Texture Attributes (LTAs) and enables frequency-aware interpretability, a Dual Depth Permutation \\((D^{2}P)\\) module to expose complementary channel adjacency patterns, and Learnable Chebyshev Polynomials (LCPs) to model high-order orderless LTA transformations via recursive Chebyshev basis expansion; and (3) a Latent Texture Attribute Loss that jointly optimizes classification accuracy, statistical alignment, and spatial fidelity. CAPTN supports end-to-end training without relying on fine-tuned CNN backbones and achieves state-of-the-art performance on several texture and material recognition benchmarks. (Code: https://github.com/RavishankarEvani/CAPTN)" + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.783, + 0.222, + 0.799 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.485, + 0.903 + ], + "angle": 0, + "content": "Visual texture perception plays a critical role in understanding an object's surface characteristics. Distinct from object recognition, which often relies on shape cues, texture recognition emphasizes the repetitive patterns or stochastic microstructures that define surface appearance. Over the years, texture recognition has evolved appreciably, be" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.293, + 0.908, + 0.43 + ], + "angle": 0, + "content": "ginning with foundational work such as Julesz's theory of textons [14], Haralick et al.'s co-occurrence matrices [12], and Laws' texture energy measures based on convolution masks [16, 17]. These early methods established the basis for statistical and structural texture analysis and were widely adopted in classical image processing pipelines. However, while effective under controlled conditions, such traditional techniques often fail to capture the rich diversity and structural complexity of real-world textures." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.431, + 0.909, + 0.703 + ], + "angle": 0, + "content": "With the rise of data-driven approaches, modern deep learning methods have largely supplanted handcrafted descriptors by offering more expressive and hierarchical feature representations. Despite this progress, deep models still encounter challenges when applied to texture recognition. In particular, successive convolution and pooling operations can suppress or discard fine-grained texture cues, limiting the network's ability to retain localized and discriminative information. Furthermore, textures often exhibit overlapping or perceptually similar traits [37], which contribute to inter-class confusion by blurring the boundaries between distinct texture categories. For instance, textures labeled as \"stratified\" might closely resemble those labeled \"cracked\" or \"zigzagged,\" making it difficult to consistently assign the correct class. In addition, intra-class variability arises from variations within the same category due to differences in lighting, scale, deformation, or inherent surface irregularities." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.911, + 0.903 + ], + "angle": 0, + "content": "The concept of texture attributes [10, 37, 39] has been shown to play a crucial role in representing perceptually similar textures. Latent texture attributes [10] refer to the underlying semantic components of visual texture attributes, where each latent texture attribute may correspond to one or more perceptual texture attributes. Effectively manipulating and transforming these latent attributes can enhance texture discrimination. Several studies employing Convolutional Neural Network (CNN)-based approaches have explored different aggregation methods, including global average pooling [1, 11, 37], Fisher vectors [7], feature encoding [35, 40], and cross-layer aggregation. However, certain aggregation and dimensionality" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23423" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.137 + ], + "angle": 0, + "content": "reduction techniques lead to the loss of important texture cues, and little attention has been given to recovering discriminative information post-aggregation." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.138, + 0.482, + 0.302 + ], + "angle": 0, + "content": "To address these limitations, we introduce a comprehensive texture recognition framework that advances both architectural design and training methodology. Central to our approach is the Chebyshev Attention Depth Permutation Texture Network (CAPTN), which extracts expressive and interpretable LTAs, depth permutation, and polynomial projection. In addition, we propose a stochastic local texture masking (SLTM) strategy and a tailored multi-objective loss function that together encourage spatially grounded and orderless texture representations. Our main contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.304, + 0.483, + 0.334 + ], + "angle": 0, + "content": "- We propose the CAPTN architecture, which integrates three key components:" + }, + { + "type": "text", + "bbox": [ + 0.104, + 0.335, + 0.483, + 0.395 + ], + "angle": 0, + "content": "- A Texture Frequency Attention (TFA) module that generates Latent Texture Attributes (LTAs) and enables visual interpretability by highlighting discriminative texture regions across frequency scales." + }, + { + "type": "text", + "bbox": [ + 0.104, + 0.396, + 0.483, + 0.455 + ], + "angle": 0, + "content": "- A Dual Depth Permutation \\((\\mathrm{D}^2\\mathrm{P})\\) module that applies two distinct channel shuffling operations to expose complementary spatial relationships and enhance feature diversity." + }, + { + "type": "text", + "bbox": [ + 0.104, + 0.456, + 0.483, + 0.499 + ], + "angle": 0, + "content": "- A Learnable Chebyshev Polynomial (LCP) expansion that enhances the representational power of each orderless LTA." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.5, + 0.483, + 0.59 + ], + "angle": 0, + "content": "- We introduce a Stochastic Local Texture Masking (SLTM) strategy that randomly occludes small patches of the input image during training. This compels the model to leverage broader spatial context and enables a new supervision signal that aligns masked patches with regional latent features." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.591, + 0.483, + 0.637 + ], + "angle": 0, + "content": "- We formulate a multi-objective LTA loss that jointly supervises classification accuracy, orderless latent attribute consistency, and spatial latent attribute alignment." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.304, + 0.483, + 0.637 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.649, + 0.233, + 0.665 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.675, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Traditional Texture Recognition. Building upon the foundational work discussed in Sec. 1, influential methods such as the Bag of Textons (BoT) [9, 18] and Bag of Words (BoW) [8] introduced a paradigm for representing textures as orderless histograms over learned dictionaries of textural elements. These methods encode local patches, aggregate them into global descriptors, and classify the results using algorithms such as Support Vector Machines (SVM). Varma et al. [31, 32] advanced this line of work by learning texture descriptors directly from data. Subsequent developments led to residual encoding techniques—Jégou et al. [15] proposed VLAD, which captures first-order residuals relative to visual codewords, while Perronnin et al. [24] introduced Fisher Vectors (FV), a probabilistic extension that also models second-order statistics." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.272 + ], + "angle": 0, + "content": "Deep Texture Recognition. Deep learning-based texture recognition methods such as FV-CNN [7] and LFV [30] primarily focus on encoding features extracted from pretrained CNNs, but they do not support end-to-end training. To address this limitation, Zhang et al. [40] proposed DeepTEN, integrating residual encoding and dictionary learning into a single layer. The layer functions analogously to VLAD and FV while enabling end-to-end training with a CNN backbone. RPNet [22] simplifies Deep-TEN by replacing dictionary learning with residual pooling. DEP [35] enhances Deep-TEN further by combining orderless texture features with spatial cues to balance structure and texture." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.275, + 0.905, + 0.472 + ], + "angle": 0, + "content": "MuLTER [13] improves DEP by integrating multi-layer CNN features, preserving fine texture and spatial details. MSBFEN [29] combines multiscale encoding and boosting to better capture discriminative texture features. Chen et al. [4] proposed using cross-layer statistical self-similarity to capture consistent patterns across CNN layers. Later, they introduced DTPNet [5], a local encoding scheme inspired by Local Binary Patterns (LBP), robust to rotation and grayscale changes. Yang et al. [36] proposed DFAEN, which uses frequency attention and first- and second-order statistics to enhance texture details. RADAM [27] fuses CNN activation maps via randomized aggregation, promoting texture diversity and reducing overfitting." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.474, + 0.905, + 0.699 + ], + "angle": 0, + "content": "FENet [33] encodes spatial regularity via fractal analysis, while DSR-Net [38] models local spatial dependencies instead of orderless aggregation. Zhai et al. [37] earlier proposed MAP-Net, which learns texture-related visual attributes and uses spatially-adaptive pooling for feature fusion. Their recent work, MPAP [39], uses a multi-branch architecture to model bottom-up structures and top-down attribute relationships by learning local-to-global mappings. Reale et al. [25] proposed modeling triangular patches as graph nodes to encode texture patterns. More recently, GTN [10] introduced a graph-based deep approach designed to model directional relationships between Latent Texture Attributes (LTAs). GTN enhances LTAs through message passing and aggregation of semantically related information." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.702, + 0.905, + 0.778 + ], + "angle": 0, + "content": "Deep Polynomial Expansion. HistNet [23] features a trainable histogram layer that preserves spatial context using learnable bin centers and widths, combined with Radial Basis Functions (RBFs) for robustness to intra-class variability." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.905, + 0.901 + ], + "angle": 0, + "content": "While graph-based methods explicitly learn attribute adjacencies for message passing to refine LTAs, our CAPTN framework adopts a more flexible and implicit approach. It enables rich and diverse feature interactions without the need for explicit graph construction. This not only simplifies the overall architecture but also implicitly regularizes the learning process by breaking fixed feature layouts, thereby enhancing texture robustness and generaliza" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23424" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.096, + 0.112, + 0.912, + 0.356 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.369, + 0.908, + 0.442 + ], + "angle": 0, + "content": "Figure 1. Stochastic Local Texture Masking (SLTM) & components of Chebyshev Attention Depth Permutation Texture Network (CAPTN): SLTM facilitates contextual texture learning, Texture Frequency Attention (TFA) captures multi-frequency texture cues, Dual Depth Permutation \\((\\mathrm{D}^2\\mathrm{P})\\) promotes feature diversification, and Learnable Chebyshev Polynomials (LCP) map latent texture attributes to visual texture attributes. The training process optimizes Latent Texture Attribute (LTA) Loss, which balances classification accuracy with orderless and spatial texture consistency. || represents concatenation, \\(\\times\\) represents matrix multiplied by a scalar." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.466, + 0.487, + 0.572 + ], + "angle": 0, + "content": "tion. Unlike radial basis functions (RBFs), which often require numerous kernel centers and careful bandwidth tuning, we propose a Chebyshev polynomial-based method that provides a compact yet globally expressive basis. This approach eliminates the need for kernel-specific design choices while efficiently capturing complex and non-linear texture variations." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.588, + 0.403, + 0.607 + ], + "angle": 0, + "content": "3. Stochastic Local Texture Masking" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.614, + 0.485, + 0.901 + ], + "angle": 0, + "content": "To enhance the model's ability to capture local, global, and contextual texture representations, we introduce Stochastic Local Texture Masking (SLTM). This technique randomly masks a small patch of the input texture image prior to passing it through the pre-trained backbone, thereby preventing the Chebyshev Attention Depth Permutation Texture Network (CAPTN) from over-relying on fine-grained local details. Given an input image \\(\\mathbf{I} \\in \\mathbb{R}^{H \\times W \\times 3}\\), SLTM generates: (1) a masked image \\(\\mathbf{I}_r\\), where a randomly selected patch of size \\(P \\times P\\) (\\(P \\ll H, P \\ll W\\)) is zeroed out, and (2) the extracted patch \\(\\mathbf{G}_t \\in \\mathbb{R}^{P \\times P \\times 3}\\), which retains the original pixel values prior to masking. By selectively removing texture information, SLTM encourages CAPTN to exploit broader spatial dependencies, thereby enhancing global feature learning and reducing overfitting. Finally, the masked image \\(\\mathbf{I}_r\\) is fed through the backbone and CAPTN for further processing. This module is activated during the computation of the Spatial Latent Texture Attribute Loss \\((\\mathcal{L}_{SL})\\), as described in Section 5." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.464, + 0.907, + 0.498 + ], + "angle": 0, + "content": "4. Chebyshev Attention Depth Permutation Texture Network" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.508, + 0.909, + 0.69 + ], + "angle": 0, + "content": "The Chebyshev Attention Depth Permutation Texture Network (CAPTN) consists of three key modules: (1) Texture Frequency Attention (TFA) Network, (2) Dual Depth Permutation \\((\\mathrm{D}^2\\mathrm{P})\\) Network, and (3) Learnable Chebyshev Polynomials (LCP). Following [10], CAPTN extracts latent texture attributes (LTAs) by computing spatial texture frequency information from feature maps \\(\\mathcal{F} = \\{\\mathbf{X}^{o_m}\\mid 1\\leq m\\leq M - 1\\}\\) obtained from an \\(M\\) -layer pre-trained CNN backbone, excluding the patchify layer [19]. Depth permutation enhances LTA representation by diversifying feature space and reducing inter-class variability, while LCP improves expressiveness of orderless LTAs." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.698, + 0.774, + 0.714 + ], + "angle": 0, + "content": "4.1. Texture Frequency Attention" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Textures display a wide range of spatial frequency characteristics, from high-frequency (fine-grained) patterns to low-frequency (smooth or homogeneous) regions. To effectively capture and interpret these variations, we propose a Texture Frequency Attention (TFA) mechanism. In addition to emphasizing discriminative texture regions, TFA also transforms backbone feature maps into Latent Texture Attributes (LTAs). Let \\(\\mathbf{X}^{o_m} \\in \\mathbb{R}^{H_m \\times W_m \\times C_m}\\) be the \\(m^{\\text{th}}\\) layer feature map. We compute transformed features as \\(\\mathbf{X}^{t_m} = \\mathbf{W}_l^m * \\mathbf{X}^{o_m} + \\mathbf{b}_l^m\\), where \\(\\mathbf{W}_l^m \\in \\mathbb{R}^{C_m \\times C_m \\times 1 \\times 1}\\) is the convolution kernel, * denotes the convolution operation, and \\(\\mathbf{b}_l^m \\in \\mathbb{R}^{C_m}\\) is a learnable bias vector. Next, we" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23425" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.097, + 0.198, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.135, + 0.234, + 0.15, + 0.243 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.199, + 0.097, + 0.581, + 0.231 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.383, + 0.233, + 0.398, + 0.244 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.583, + 0.097, + 0.899, + 0.231 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.738, + 0.233, + 0.753, + 0.243 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.245, + 0.55, + 0.38 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.312, + 0.382, + 0.328, + 0.393 + ], + "angle": 0, + "content": "(d)" + }, + { + "type": "image", + "bbox": [ + 0.552, + 0.245, + 0.902, + 0.38 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.72, + 0.383, + 0.735, + 0.393 + ], + "angle": 0, + "content": "(e)" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.404, + 0.907, + 0.475 + ], + "angle": 0, + "content": "Figure 2. (a) SLTM: extract a small patch from an input texture image, (b) \\(\\mathrm{D}^2\\mathrm{P}\\): feature maps from backbone undergo depth permutation to diversify feature space, (c) LCP: enhanced Latent Texture Attributes undergo orderless aggregation and fit a learnable Chebyshev polynomial function, (d) TFA: use spatial texture frequencies to compute spatial attention, (e) SLAR: representation of spatial information for latent attributes and backbone features. SLTM and SLAR are used to compute Latent Texture Attribute Loss. || represents concatenation, \\(\\times\\) represents matrix multiplied by a scalar, \\(\\cdot\\) represents matrix multiplication." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.501, + 0.483, + 0.53 + ], + "angle": 0, + "content": "compute per-location channel-wise maximum and average statistics" + }, + { + "type": "equation", + "bbox": [ + 0.162, + 0.564, + 0.483, + 0.588 + ], + "angle": 0, + "content": "\\[\nX _ {\\max } ^ {t _ {m}} (i, j) = \\max _ {0 \\leq c < C _ {m}} X ^ {t _ {m}} (i, j, c), \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.17, + 0.592, + 0.482, + 0.633 + ], + "angle": 0, + "content": "\\[\nX _ {\\text {a v g}} ^ {t _ {m}} (i, j) = \\frac {1}{C _ {m}} \\sum_ {c = 0} ^ {C _ {m} - 1} X ^ {t _ {m}} (i, j, c). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.65, + 0.484, + 0.773 + ], + "angle": 0, + "content": "Here, \\(\\mathbf{X}_{\\mathrm{max}}^{t_m}, \\mathbf{X}_{\\mathrm{avg}}^{t_m} \\in \\mathbb{R}^{H_m \\times W_m \\times 1}\\) represent spatial maps of local maxima and average activations across channels. High-frequency textures tend to exhibit greater local variation, resulting in larger differences between maximum and average values, whereas low-frequency regions yield more consistent responses. To capture this contrast, we concatenate \\(\\mathbf{X}_{\\mathrm{max}}^{t_m}\\) and \\(\\mathbf{X}_{\\mathrm{avg}}^{t_m}\\) along the channel dimension and apply a convolution followed by a sigmoid activation" + }, + { + "type": "equation", + "bbox": [ + 0.172, + 0.792, + 0.483, + 0.827 + ], + "angle": 0, + "content": "\\[\n\\mathbf {A} ^ {m} = \\frac {1}{1 + e ^ {- \\left[ \\mathbf {W} _ {a} ^ {m} * \\left(\\mathbf {X} _ {\\operatorname* {m a x}} ^ {t _ {m}} \\parallel \\mathbf {X} _ {\\operatorname* {a v g}} ^ {t _ {m}}\\right) \\right]}}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "where \\(\\parallel\\) denotes channel-wise concatenation and \\(\\mathbf{W}_a^m\\in\\) \\(\\mathbb{R}^{1\\times 2\\times 1\\times 1}\\) is the convolution kernel. The resulting spatial attention map \\(\\mathbf{A}^m\\in \\mathbb{R}^{H_m\\times W_m\\times 1}\\) is used to modulate \\(\\mathbf{X}^{t_m}\\) via element-wise multiplication" + }, + { + "type": "equation", + "bbox": [ + 0.54, + 0.522, + 0.905, + 0.571 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} X _ {\\text {a t t n}} ^ {t _ {m}} (i, j, c) = X ^ {t _ {m}} (i, j, c) \\times A ^ {m} (i, j), \\\\ \\forall (i, j, c) \\in [ 0, H _ {m} - 1 ] \\times [ 0, W _ {m} - 1 ] \\times [ 0, C _ {m} - 1 ], \\tag {4} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.572, + 0.906, + 0.632 + ], + "angle": 0, + "content": "A residual connection is then added, combining the attention-weighted features with the original \\(\\mathbf{X}^{t_m}\\). The result is passed through a transformation \\(g(\\cdot)\\) comprising Batch Normalization, GELU activation, and dropout" + }, + { + "type": "equation", + "bbox": [ + 0.587, + 0.645, + 0.905, + 0.663 + ], + "angle": 0, + "content": "\\[\n\\mathbf {X} _ {a} ^ {t _ {m}} = f _ {a} \\left(\\gamma_ {m}\\right) \\times g \\left(\\mathbf {X} _ {\\text {a t t n}} ^ {t _ {m}} + \\mathbf {X} ^ {t _ {m}}\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.668, + 0.906, + 0.699 + ], + "angle": 0, + "content": "where \\(\\gamma_{m}\\) is a learnable scalar, and \\(f_{a}(\\cdot)\\) ensures the scaling factor remains within the range [0, 1]" + }, + { + "type": "equation", + "bbox": [ + 0.583, + 0.713, + 0.905, + 0.73 + ], + "angle": 0, + "content": "\\[\nf _ {a} (z) = \\min (\\max (0, z), 1), \\quad z \\in \\mathbb {R}. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.906, + 0.796 + ], + "angle": 0, + "content": "While the convolution and \\( g(\\cdot) \\) transformation generate the LTAs, the spatial attention map \\( \\mathbf{A}^m \\) provides interpretable insights into the model's dynamic focus on spatially discriminative texture regions." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.803, + 0.741, + 0.82 + ], + "angle": 0, + "content": "4.2. Dual Depth Permutation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.906, + 0.902 + ], + "angle": 0, + "content": "Dual Depth Permutation \\((\\mathrm{D}^2\\mathrm{P})\\) operates directly on backbone features to generate complementary representations through two distinct shuffling operations. Exposing varied adjacency patterns promotes rich cross-channel correlations. The outputs from both shuffling paths are later" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "23426" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.26 + ], + "angle": 0, + "content": "fused with LTAs from Sec. 4.1 to enhance representation diversity and improve texture robustness. Additionally, \\(\\mathrm{D}^2\\mathrm{P}\\) serves as an implicit regularizer, reducing overfitting by disrupting fixed feature layouts. Drawing inspiration from constraining [2], \\(\\mathrm{D}^2\\mathrm{P}\\) introduces multiple feature perspectives by permuting channel adjacency structures in a supervised setting. Let \\(\\mathbf{X}^{o_m} \\in \\mathbb{R}^{H_m \\times W_m \\times C_m}\\) denote the feature map output from the \\(m^{\\text{th}}\\) layer of the backbone. We apply two distinct channel shuffles using group sizes \\(s_0\\) and \\(s_1\\), ensuring \\(\\frac{C_m}{s_0}, \\frac{C_m}{s_1} \\in \\mathbb{Z}\\). Let \\(p \\in \\{0,1\\}\\) index the two shuffling paths, with corresponding group sizes \\(k_p^m \\in \\{s_0, s_1\\}\\)" + }, + { + "type": "equation", + "bbox": [ + 0.162, + 0.274, + 0.483, + 0.297 + ], + "angle": 0, + "content": "\\[\n\\mathbf {X} ^ {o _ {m}} \\rightarrow \\mathbf {X} _ {\\mathrm {r}, p} ^ {o _ {m}} \\in \\mathbb {R} ^ {H _ {m} \\times W _ {m} \\times k _ {p} ^ {m} \\times g _ {k _ {p} ^ {m}}}, \\tag {7}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.163, + 0.318, + 0.483, + 0.343 + ], + "angle": 0, + "content": "\\[\n\\mathbf {X} _ {\\mathrm {r}, p} ^ {o _ {m}} \\rightarrow \\mathbf {X} _ {\\mathrm {r}, p} ^ {o _ {m} ^ {T}} \\in \\mathbb {R} ^ {H _ {m} \\times W _ {m} \\times g _ {k _ {p} ^ {m}} \\times k _ {p} ^ {m}}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.353, + 0.484, + 0.392 + ], + "angle": 0, + "content": "where \\(g_{k_p^m}\\) is the number of groups formed using group size \\(k_{p}^{m}\\) at layer \\(m\\). After flattening \\(\\mathbf{X}_{\\mathrm{r},p}^{o_m^T}\\), we obtain" + }, + { + "type": "equation", + "bbox": [ + 0.207, + 0.407, + 0.483, + 0.428 + ], + "angle": 0, + "content": "\\[\n\\mathbf {X} _ {\\mathrm {s}, p} ^ {o _ {m}} \\in \\mathbb {R} ^ {H _ {m} \\times W _ {m} \\times C _ {m}}. \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.437, + 0.484, + 0.558 + ], + "angle": 0, + "content": "Each of the shuffled representations is processed by a convolution operation, where \\(\\mathbf{W}_p^m\\in \\mathbb{R}^{C_m\\times C_m\\times 1\\times 1}\\) is the convolution kernel, \\(*\\) denotes the convolution operation, and \\(\\mathbf{b}_p^m\\in \\mathbb{R}^{C_m}\\) is a learnable bias vector. The result is then passed through a transformation \\(f_{p}(\\cdot)\\), consisting of Batch Normalization, GELU activation, and Dropout. The final output of the \\(\\mathrm{D}^2\\mathrm{P}\\) module is obtained by summing the outputs from both shuffling paths" + }, + { + "type": "equation", + "bbox": [ + 0.106, + 0.584, + 0.465, + 0.634 + ], + "angle": 0, + "content": "\\[\n\\mathbf {X} _ {\\mathrm {s}} ^ {d _ {m}} = f _ {a} (\\psi_ {m}) \\times \\left[ \\sum_ {p \\in \\{0, 1 \\}} f _ {p} \\left(\\mathbf {W} _ {p} ^ {m} * \\mathbf {X} _ {\\mathrm {s}, p} ^ {o _ {m}} + \\mathbf {b} _ {p} ^ {m}\\right) \\right],\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.118, + 0.636, + 0.482, + 0.665 + ], + "angle": 0, + "content": "\\[\n\\psi_ {m} \\in (- \\infty , \\infty), \\quad f _ {a} (\\cdot) \\in [ 0, 1 ], \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.667, + 0.484, + 0.713 + ], + "angle": 0, + "content": "where \\(\\psi_{m}\\) is a learnable scalar parameter, and the scaling function \\(f_{a}(\\cdot)\\) is defined in Eq. (6) to constrain the modulation factor within the range [0, 1]." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.725, + 0.484, + 0.74 + ], + "angle": 0, + "content": "4.3. Enhanced Orderless Latent Texture Attributes" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.749, + 0.484, + 0.824 + ], + "angle": 0, + "content": "While GTN [10] enhances LTA representations via message passing and aggregation, CAPTN improves the LTAs generated by the TFA module, \\(\\mathbf{X}_a^{t_m}\\), by injecting dynamically adapted adjacency structures from the \\(\\mathrm{D}^2\\mathrm{P}\\) module, \\(\\mathbf{X}_s^{d_m}\\), resulting in" + }, + { + "type": "equation", + "bbox": [ + 0.22, + 0.826, + 0.483, + 0.845 + ], + "angle": 0, + "content": "\\[\n\\mathbf {X} _ {e} ^ {m} = \\mathbf {X} _ {a} ^ {t _ {m}} + \\mathbf {X} _ {\\mathrm {s}} ^ {d _ {m}}. \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.903 + ], + "angle": 0, + "content": "We then compute the spatial mean over the height and width of the enhanced LTAs, \\(\\mathbf{X}_e^m\\in \\mathbb{R}^{H_m\\times W_m\\times C_m}\\), from layer \\(m\\) to obtain an orderless representation, \\(\\mathbf{O}_{\\mathrm{e}}^{m}\\in \\mathbb{R}^{C_{m}}\\)." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.092, + 0.809, + 0.107 + ], + "angle": 0, + "content": "4.4. Learnable Chebyshev Polynomial" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.113, + 0.907, + 0.295 + ], + "angle": 0, + "content": "Chebyshev polynomials [26] offer a powerful global basis for modeling non-linear, oscillatory, and fine-grained variations in orderless LTAs. Their higher-order, non-monotonic structure enables expressive transformations that capture both smooth and oscillatory variations, allowing the model to distinguish subtle patterns while aligning semantically similar LTA values—even when numerically distant. To enhance the expressiveness of each orderless LTA dimension, we apply a Chebyshev polynomial \\( T_{n}(\\cdot) \\) for \\( n \\in \\{0, \\dots, d\\} \\), each weighted by a sigmoid function parameterized by \\( \\theta_{n} \\). The resulting formulation, referred to as a Learnable Chebyshev Polynomial (LCP), takes the form" + }, + { + "type": "equation", + "bbox": [ + 0.583, + 0.306, + 0.907, + 0.348 + ], + "angle": 0, + "content": "\\[\nP \\left(\\mathbf {O} _ {\\mathrm {e}} ^ {m}\\right) = \\sum_ {n = 0} ^ {d} T _ {n} \\left(\\mathbf {O} _ {\\mathrm {e}} ^ {m}\\right) \\left[ \\frac {1}{1 + e ^ {- \\theta_ {n}}} \\right], \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.352, + 0.906, + 0.382 + ], + "angle": 0, + "content": "The Chebyshev polynomials \\( T_{n}(\\mathbf{O}_{\\mathrm{e}}^{m}) \\) used in the expansion follow the recurrence relation" + }, + { + "type": "equation", + "bbox": [ + 0.546, + 0.407, + 0.906, + 0.479 + ], + "angle": 0, + "content": "\\[\nT _ {n} \\left(\\mathbf {O} _ {\\mathrm {e}} ^ {m}\\right) = \\left\\{ \\begin{array}{l l} \\mathbf {1}, & \\text {i f} n = 0, \\\\ \\mathbf {O} _ {\\mathrm {e}} ^ {m}, & \\text {i f} n = 1, \\\\ 2 \\mathbf {O} _ {\\mathrm {e}} ^ {m} T _ {n - 1} \\left(\\mathbf {O} _ {\\mathrm {e}} ^ {m}\\right) & \\text {f o r} n \\geq 2. \\\\ - T _ {n - 2} \\left(\\mathbf {O} _ {\\mathrm {e}} ^ {m}\\right), & \\end{array} \\right. \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.488, + 0.906, + 0.52 + ], + "angle": 0, + "content": "LCPs from all layers are then concatenated and passed through a fully connected layer, \\( f_{fc} \\), as" + }, + { + "type": "equation", + "bbox": [ + 0.618, + 0.529, + 0.906, + 0.555 + ], + "angle": 0, + "content": "\\[\n\\mathbf {Y} = f _ {f c} \\left( \\right.\\left\\| \\right. _ {m = 1} ^ {N - 1} P \\left(\\mathbf {O} _ {\\mathrm {e}} ^ {m}\\right)\\left. \\right). \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.559, + 0.906, + 0.589 + ], + "angle": 0, + "content": "The logits, \\(\\mathbf{Y}\\), are then evaluated using the Latent Texture Attribute Loss function, which will be discussed in Sec. 5." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.598, + 0.857, + 0.615 + ], + "angle": 0, + "content": "4.5. Spatial Latent Attribute Representation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.62, + 0.907, + 0.786 + ], + "angle": 0, + "content": "Finally, we aim to enforce some extent of structural similarity between a texture patch extracted from Sec. 3 and regional latent texture attribute information. To achieve this, we first concatenate the enhanced LTAs, \\(\\mathbf{X}_e^{M - 1}\\), with the original feature maps from the frozen backbone, \\(\\mathbf{X}^{O_M - 1}\\). Both representations are obtained from the final layer of the network. We integrate the frozen backbone features into the LTAs through a structured transformation consisting of three sequential operations: cropping \\(f_{c}(\\cdot)\\), a \\(3\\times 3\\) convolution, and bilinear interpolation \\(f_{b}(\\cdot)\\). The transformation is formally defined as" + }, + { + "type": "equation", + "bbox": [ + 0.54, + 0.812, + 0.907, + 0.831 + ], + "angle": 0, + "content": "\\[\n\\mathbf {G} _ {\\mathrm {p}} = f _ {b} \\left(\\mathbf {W} _ {s} * f _ {c} \\left(\\mathbf {X} _ {e} ^ {M - 1} \\| \\mathbf {X} ^ {O M - 1}\\right) + \\mathbf {b} _ {s}, P\\right), \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.839, + 0.907, + 0.903 + ], + "angle": 0, + "content": "where \\(\\mathbf{W}_s\\in \\mathbb{R}^{3\\times C\\times 3\\times 3}\\) and \\(\\mathbf{b}_s\\in \\mathbb{R}^3\\) are the learnable weights and bias of the convolutional layer. The output of this transformation is the Spatial Latent Attribute Representation (SLAR), \\(\\mathbf{G}_{\\mathrm{p}}\\in \\mathbb{R}^{3\\times P\\times P}\\), where \\(P\\) is the spatial size" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23427" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.242 + ], + "angle": 0, + "content": "of the extracted patch. This ensures that the extracted regional texture attributes retain both the fine-grained information from LTAs and the structural guidance preserved in the frozen backbone features. Incorporating structure-aware features from the frozen backbone provides a grounded reference that constrains the adaptation of LTAs during training, preventing overcorrection while maintaining a balance between texture fidelity and spatial consistency. SLAR will be used to compute Spatial Latent Texture Attribute Loss (SL) in Sec. 5." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.259, + 0.367, + 0.275 + ], + "angle": 0, + "content": "5. Latent Texture Attribute Loss" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.286, + 0.484, + 0.451 + ], + "angle": 0, + "content": "We propose the Latent Texture Attribute (LTA) Loss to enhance classification accuracy while preserving texture consistency at both global and local levels. The LTA Loss integrates three complementary components: (1) Cross Entropy Loss \\((\\mathcal{L}_{CE})\\) to promote accurate class predictions, (2) Orderless LTA Loss \\((\\mathcal{L}_{OL})\\) to align global, spatially invariant texture distributions, and (3) Spatial LTA Loss \\((\\mathcal{L}_{SL})\\) to preserve fine-grained, spatial texture structure. By jointly optimizing these objectives, the model learns rich and coherent latent texture representations that support both accurate classification and texture understanding." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.453, + 0.484, + 0.65 + ], + "angle": 0, + "content": "Orderless Latent Texture Attribute Loss. This second texture loss component compares distribution of orderless LTAs with orderless features extracted directly from the frozen backbone. By aligning loss from learned texture representations with a stable baseline, we encourage consistency in the model's understanding of spatial invariant texture representations across images. We apply Batch Normalization, \\( f_{\\mathrm{bn}} \\), on feature maps from the \\( m^{\\mathrm{th}} \\) layer of the backbone, \\( \\mathbf{X}^{o_m} \\), before calculating the spatial mean and concatenating multiple scales of orderless texture information across layers \\( m \\in [1, M - 1] \\) of the backbone to obtain \\( \\mathbf{O}_{\\mathrm{b}} \\). We also concatenate orderless enhanced LTAs, \\( \\mathbf{O}_{\\mathrm{e}}^{m} \\), across the same range of layers to obtain \\( \\mathbf{O}_{\\mathrm{e}} \\)" + }, + { + "type": "equation", + "bbox": [ + 0.11, + 0.674, + 0.484, + 0.763 + ], + "angle": 0, + "content": "\\[\n\\mathbf {O} _ {\\mathrm {b}} = \\left\\| \\right. _ {m = 1} ^ {M - 1} \\left(\\frac {\\sum_ {h = 0} ^ {H _ {m} - 1} \\sum_ {w = 0} ^ {W _ {m} - 1} f _ {\\mathrm {b n}} \\left(\\mathbf {X} ^ {o _ {m}}\\right) [ h , w , : ]}{H _ {m} \\times W _ {m}}\\right), \\tag {16}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.229, + 0.777, + 0.484, + 0.817 + ], + "angle": 0, + "content": "\\[\n\\mathbf {O} _ {\\mathrm {e}} = \\left\\| _ {m = 1} ^ {M - 1} \\mathbf {O} _ {\\mathrm {e}} ^ {m}. \\right. \\tag {17}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.824, + 0.484, + 0.902 + ], + "angle": 0, + "content": "We perform softmax normalization of \\(\\mathbf{O_b}\\) and \\(\\mathbf{O_e}\\) over the total number of channels in each sample \\(s\\) from the minibatch, aggregated over layers \\(m\\in [1,M - 1]\\), where the total number of concatenated channels is \\(C = \\sum_{m = 1}^{M - 1}C_m\\). A small constant \\(\\epsilon\\) is added for numerical stability" + }, + { + "type": "equation", + "bbox": [ + 0.577, + 0.103, + 0.905, + 0.143 + ], + "angle": 0, + "content": "\\[\np _ {\\alpha , i} ^ {s} = \\frac {e ^ {\\mathbf {O} _ {\\alpha , i} ^ {s}}}{\\sum_ {j = 0} ^ {C - 1} e ^ {\\mathbf {O} _ {\\alpha , j} ^ {s}}} + \\epsilon , \\quad \\alpha \\in \\{\\mathrm {b}, \\mathrm {e} \\}. \\tag {18}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.147, + 0.907, + 0.226 + ], + "angle": 0, + "content": "Inspired by Kullback-Leibler (KL) divergence, we calculate orderless LTA loss, \\(\\mathcal{L}_{OL}\\) that measures the dissimilarity between two probability distributions: a fixed distribution from orderless feature maps extracted from the frozen backbone, \\(p_{\\mathrm{b},i}^{\\mathrm{s}}\\) and that from orderless enhanced LTAs, \\(p_{\\mathrm{e},i}^{\\mathrm{s}}\\)" + }, + { + "type": "equation", + "bbox": [ + 0.532, + 0.246, + 0.905, + 0.289 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {O L} = \\frac {1}{S} \\sum_ {s = 0} ^ {S - 1} \\sum_ {i = 0} ^ {C - 1} p _ {\\mathrm {b}, i} ^ {s} \\left[ \\log_ {e} \\left(p _ {\\mathrm {b}, i} ^ {s}\\right) - \\log_ {e} \\left(p _ {\\mathrm {e}, i} ^ {s}\\right) \\right], \\tag {19}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.299, + 0.905, + 0.327 + ], + "angle": 0, + "content": "where \\(S\\) represents the number of samples in the minibatch." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.329, + 0.907, + 0.45 + ], + "angle": 0, + "content": "Spatial Latent Texture Attribute Loss. The final LTA loss component measures the similarity between SLAR from Sec. 4.5 and a corresponding texture patch extracted from the original image. This component helps ensure that the latent representations preserve both regional and local spatial texture information by comparing the latent attributes \\(\\mathbf{G}_{\\mathrm{p}}\\) with the extracted texture patch \\(\\mathbf{G}_{\\mathrm{t}}\\) using the Mean Squared Error (MSE)" + }, + { + "type": "equation", + "bbox": [ + 0.548, + 0.473, + 0.905, + 0.505 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {S L} = \\frac {1}{B C _ {M - 1} H _ {M - 1} W _ {M - 1}} \\| \\mathbf {G} _ {\\mathrm {p}} - \\mathbf {G} _ {\\mathrm {t}} \\| _ {F} ^ {2}, \\tag {20}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.515, + 0.906, + 0.666 + ], + "angle": 0, + "content": "where \\(B, C_{M-1}, H_{M-1}, W_{M-1}\\) represent the batch size, number of channels, height, and width of the feature maps, from the final \\(M-1\\) layer, and \\(\\|\\cdot\\|_F\\) denotes the Frobenius norm, which sums the squared differences over all elements. This loss encourages the model to preserve fine-grained spatial texture patterns within the latent space, ensuring that learned texture features remain spatially structured and contextually coherent. By aligning latent spatial representations with real texture patches, the model enhances its ability to reconstruct or classify textures with high fidelity." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.667, + 0.906, + 0.711 + ], + "angle": 0, + "content": "Latent Texture Attribute Loss. We formulate our loss during the training phase as a weighted sum of the three loss components" + }, + { + "type": "equation", + "bbox": [ + 0.591, + 0.727, + 0.905, + 0.743 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {t r a i n}} = \\lambda \\mathcal {L} _ {C E} + \\mu \\mathcal {L} _ {O L} + \\nu \\mathcal {L} _ {S L}, \\tag {21}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.908, + 0.78 + ], + "angle": 0, + "content": "where \\(\\lambda, \\mu, \\nu \\in [0,1]\\) are weights associated with each of the loss components." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.794, + 0.646, + 0.81 + ], + "angle": 0, + "content": "6. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.819, + 0.617, + 0.833 + ], + "angle": 0, + "content": "6.1. Datasets" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.906, + 0.902 + ], + "angle": 0, + "content": "We evaluated our network using five benchmark datasets, summarized as follows: (a) Ground Terrain in Outdoor Scenes (GTOS) [34] contains 34,105 images of outdoor ground terrain across 40 categories, with 5 train/test splits." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23428" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.55, + 0.145 + ], + "angle": 0, + "content": "Table 1. Comparison of CAPTN with GTN [10] and RADAM [27] in terms of classification accuracy (\\%) on texture benchmarks. ConvNeXt-\\* variants are used as backbones. Pre-trained weights are frozen. The best result for each backbone-dataset combination is bolded." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.156, + 0.549, + 0.317 + ], + "angle": 0, + "content": "
Model-*DTDFMDKTHGTOSGTOS-M
RADAMN74.9 ±0.787.1 ±0.489.6 ±3.883.7 ±1.581.8
GTNN75.4 ±1.186.9 ±3.190.3 ±4.782.9 ±1.790.1
CAPTN (ours)N77.9 ±1.087.0 ±2.990.6 ±4.683.9 ±1.791.5
RADAMT77.0 ±0.788.7 ±0.490.7 ±4.084.2 ±1.785.3
GTNT77.5 ±1.087.2 ±3.492.0 ±4.984.7 ±1.691.9
CAPTN (ours)T78.9 ±0.887.2 ±3.692.2 ±4.485.1 ±1.794.2
RADAMB76.4 ±0.990.2 ±0.287.7 ±5.684.1 ±1.682.2
GTNB78.3 ±0.888.8 ±3.090.8 ±6.284.2 ±1.894.8
CAPTN (ours)B79.3±0.989.1 ±3.491.3±5.784.6 ±1.994.0
RADAML77.4 ±1.189.3±0.389.3 ±3.484.0 ±1.885.8
GTNL79.3 ±0.888.1 ±3.092.5 ±4.784.9 ±1.793.9
CAPTN (ours)L79.6±0.889.3±3.193.0 ±4.985.0 ±1.795.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.357, + 0.483, + 0.4 + ], + "angle": 0, + "content": "Table 4. Classification accuracy (\\%) comparison between \\(1^{\\text{st}}\\)-order mean pooling (w/ LCP) and a non-polynomial baseline: \\(2^{\\text{nd}}\\)-order covariance pooling (w/o LCP), with ConvNeXt-B backbone." + }, + { + "type": "table", + "bbox": [ + 0.102, + 0.41, + 0.475, + 0.458 + ], + "angle": 0, + "content": "
PoolingDTDKTHGTOSGTOS-M
2nd-order76.4 ±1.091.1 ±6.383.5 ±1.793.0
1st-order79.3 ±0.991.3 ±5.784.6 ±1.994.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.471, + 0.483, + 0.5 + ], + "angle": 0, + "content": "Table 5. Classification accuracy (\\%) of CAPTN with different basis functions using ConvNeXt-L backbone." + }, + { + "type": "table", + "bbox": [ + 0.114, + 0.511, + 0.462, + 0.558 + ], + "angle": 0, + "content": "
Basis Fn.DTDKTHGTOSGTOS-M
RBF74.1 ±1.089.6 ±6.881.5 ±2.492.7
LCP79.6±0.893.0±4.985.0±1.795.1
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.582, + 0.483, + 0.734 + ], + "angle": 0, + "content": "(b) GTOS-Mobile (GTOS-M) [35] includes 100,011 material samples from 31 categories, captured via mobile phone, with a single train/test split. (c) KTH-TIPS2-b (KTH) [3] provides 4,752 images in 11 material categories, with 4 train/test splits. (d) Flickr Material Dataset (FMD) [28] contains 1,000 images spanning 10 material categories, validated with 10-fold cross-validation and stratified sampling to maintain class proportions. (e) Describable Texture Dataset (DTD) [6] has 5,640 images across 47 categories, with 10 train/validation/test splits." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.743, + 0.308, + 0.759 + ], + "angle": 0, + "content": "6.2. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.765, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Our model was implemented in PyTorch and trained on a single NVIDIA RTX A6000 GPU. For all datasets, the learning rate was set to 0.0001 and followed a cosine annealing schedule [20]. We used the Adam optimizer with decoupled weight decay [21] and our novel Latent Texture Attribute Loss function. Our network was trained until the final epoch for each dataset, with test set performance reported. More design and implementation details can be found in the supplementary material." + }, + { + "type": "table_caption", + "bbox": [ + 0.559, + 0.089, + 0.905, + 0.131 + ], + "angle": 0, + "content": "Table 2. Evaluate effectiveness of TFA, \\( {\\mathrm{D}}^{2}\\mathrm{P} \\) ,and LCP in terms of classification accuracy (%) . ConvNeXt-N is the backbone." + }, + { + "type": "table", + "bbox": [ + 0.561, + 0.143, + 0.905, + 0.212 + ], + "angle": 0, + "content": "
TFAD2PLCPDTDKTHGTOS
XXX75.4 ±1.087.2 ±5.381.7 ±2.0
✓XX77.1 ±1.190.4 ±4.883.5 ±1.8
✓✓X77.2 ±1.090.5 ±4.883.6 ±1.8
✓✓✓77.9 ±1.090.6 ±4.683.9 ±1.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.559, + 0.227, + 0.905, + 0.255 + ], + "angle": 0, + "content": "Table 3. Evaluate effectiveness of LTA Loss in terms of classification accuracy \\((\\%)\\). ConvNeXt-T is the backbone." + }, + { + "type": "table", + "bbox": [ + 0.561, + 0.266, + 0.905, + 0.312 + ], + "angle": 0, + "content": "
Loss Fn.DTDKTHGTOSGTOS-M
CE Loss78.4 ±0.791.9 ±5.185.0 ±1.893.6
LTA Loss78.9 ±0.892.2 ±4.485.1 ±1.794.2
" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.358, + 0.616, + 0.374 + ], + "angle": 0, + "content": "6.3. Analysis" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.387, + 0.905, + 0.462 + ], + "angle": 0, + "content": "Recent studies [10, 27] have demonstrated the superiority of ConvNeXt backbones over ResNet backbones for texture-specific applications, particularly in multiscale texture representation. Hence, all our experiments will be based on ConvNeXt backbones." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.463, + 0.906, + 0.688 + ], + "angle": 0, + "content": "Quantitative. Tab. 1 depicts how CAPTN's performance compares with the current state-of-the-art RADAM [27] and GTN [10] across various ConvNeXt backbones pretrained on ImageNet. The average and standard deviation measures are calculated across the provided splits, where applicable. Following RADAM and GTN, we have frozen our ConvNeXt backbones. CAPTN out-performed both RADAM [27] and GTN [10] in most of the benchmark datasets across different ConvNeXt backbone sizes. Both RADAM and GTN do not have a mechanism to represent orderless texture features or attributes. On the other hand, CAPTN represents orderless latent texture attributes as a learnable polynomial improving its expressiveness and consequently optimizing the mapping from latent to visual texture attributes." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Qualitative. The illustrations in Fig. 3 were generated using the ConvNeXt-L backbone with images from the DTD dataset. The Texture Frequency Attention (TFA) module from CAPTN enhances spatial attention in regions containing specific texture features. To some extent, the attention mechanism can identify the spatial regularity of grid textures. For spiraled textures, which exhibit greater irregularity, CAPTN demonstrates robustness to variations in illumination, shadows, occlusion, and background textures by selectively focusing on the spiraled region. Similarly, for polka-dotted textures, CAPTN disregards occlusions and concentrates on the textured areas. Furthermore, in all three cases, CAPTN recognizes masked regions within the images and assigns them lower importance." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23429" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.088, + 0.198, + 0.17 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.234, + 0.087, + 0.344, + 0.17 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.382, + 0.089, + 0.483, + 0.168 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.092, + 0.182, + 0.195, + 0.262 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.235, + 0.182, + 0.34, + 0.262 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.382, + 0.184, + 0.483, + 0.262 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.271, + 0.483, + 0.327 + ], + "angle": 0, + "content": "Figure 3. Top row: Masked images from DTD dataset (grid, spiralled, polka dotted). Bottom row: Corresponding attention maps from Texture Frequency Attention (TFA). Red indicates high attention, blue indicates low attention." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.357, + 0.245, + 0.373 + ], + "angle": 0, + "content": "6.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.38, + 0.484, + 0.653 + ], + "angle": 0, + "content": "We conducted ablation studies to evaluate the impact of TFA, \\(\\mathrm{D}^2\\mathrm{P}\\), and LCP in CAPTN. Overall, all components improve CAPTN's performance on both texture (DTD) and material (KTH, GTOS) datasets. The quantitative results are presented in Tab. 2. First, we assessed LCP by mapping orderless LTAs to visual texture attributes without using LCP. The deterioration of results demonstrated that LCP enhances expressiveness and injects non-linearities, improving the mapping of latent to visual texture attributes. Removing \\(\\mathrm{D}^2\\mathrm{P}\\) led to a slight performance drop, as the LTA space became less robust and diverse. Finally, the removal of TFA further degraded CAPTN's performance, indicating that backbone features alone are less discriminative than LTAs. In addition, we performed an ablation study Tab. 3 to understand the effectiveness of LTA loss compared with Cross Entropy (CE) loss. The LTA loss function enhanced performance across texture and material datasets, underscoring its effectiveness for their recognition." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.666, + 0.334, + 0.683 + ], + "angle": 0, + "content": "6.5. Alternative Design Choices" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Non-polynomial baselines. Table 4 demonstrates that first-order mean pooling combined with a LCP transformation outperforms the non-polynomial baseline of second-order covariance pooling. Covariance pooling captures second-order statistics by modeling inter-LTA correlations computed over spatial activations. While this can encode useful dependencies between LTA dimensions, it is also sensitive to noise, small artifacts, and spatial misalignments due to its reliance on precise variation patterns across spatial locations. In contrast, mean pooling followed by an LCP transformation offers a more robust encoding alternative. Mean pooling summarizes each LTA by computing the first-order global average across spatial dimensions, producing compact and stable representations that are less affected by lo" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.212 + ], + "angle": 0, + "content": "cal distortions. These pooled features serve as clean inputs to the LCP module, which applies expressive, non-linear transformations to each LTA independently using Chebyshev polynomials. This combination enhances the representational power of each LTA without the noise sensitivity inherent in second-order pooling. Together, mean pooling and LCP offer a robust and discriminative representation strategy that outperforms covariance pooling." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.213, + 0.907, + 0.665 + ], + "angle": 0, + "content": "Alternative Polynomial Bases. We replace the LCP module in CAPTN with a learnable RBF layer using Gaussian kernels. It computes scaled Euclidean distances to learnable centers and aggregates the activations via a learnable weight matrix. To avoid underfitting, we add a linear projection, LayerNorm, and GELU activation, along with kernel diversity for better generalization. As shown in Table 5, LCP outperforms the RBF variant across a range of texture and material datasets, while also requiring fewer learnable parameters. On the GTOS-M dataset, for example, CAPTN with a third-order LCP requires only 9.60M parameters, compared to 14.22M for its RBF-based counterpart. Although Gaussian RBFs are effective at modeling local variations, they often require many kernel centers and careful bandwidth tuning to capture diverse texture patterns, increasing computational cost. In contrast, LCPs form a global polynomial basis that efficiently captures long-range dependencies and oscillatory patterns with fewer parameters. In addition to RBFs, we also considered B-splines, which are widely used for smooth, piecewise function approximation. However, B-splines require careful knot placement and segmentwise assembly to represent multi-peaked or oscillatory signals, making them more sensitive to hyperparameter tuning and less scalable. LCPs, by comparison, provide a unified, smooth basis without discontinuities at segment boundaries, naturally capturing global texture variations while preserving sharp gradients and repetitive motifs. These properties make LCPs a more compact, scalable, and interpretable choice than both RBFs and B-splines for texture-specific applications." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.679, + 0.634, + 0.695 + ], + "angle": 0, + "content": "7. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.704, + 0.907, + 0.901 + ], + "angle": 0, + "content": "We propose a comprehensive framework for robust texture representation by integrating SLTM with CAPTN. SLTM enhances generalization by mitigating over-reliance on fine-grained details, while the TFA mechanism captures multifrequency texture cues to support visual interpretation of spatial texture information during the generation of LTAs. Furthermore, the \\(\\mathrm{D}^2\\mathrm{P}\\) module enriches LTAs by diversifying feature representations, and LCP is specifically designed to enable a richer and more expressive encoding of orderless LTAs. To further strengthen the learning process, we introduce a unified LTA loss function that balances classification objectives with orderless alignment and spatial consistency." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "23430" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.116, + 0.482, + 0.156 + ], + "angle": 0, + "content": "[1] Vincent Andrearczyk and Paul F Whelan. Using filter banks in convolutional neural networks for texture classification. Pattern Recognition Letters, 84:63-69, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.159, + 0.483, + 0.213 + ], + "angle": 0, + "content": "[2] Avrim Blum and Tom Mitchell. Combining labeled and unlabeled data with co-training. In Proceedings of the eleventh annual conference on Computational learning theory, pages 92-100, 1998. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.216, + 0.483, + 0.271 + ], + "angle": 0, + "content": "[3] Barbara Caputo, Eric Hayman, and P. Mallikarjuna. Class-specific material categorisation. In Tenth IEEE International Conference on Computer Vision (ICCV'05) Volume 1, pages 1597-1604 Vol. 2, 2005. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.273, + 0.482, + 0.342 + ], + "angle": 0, + "content": "[4] Zhile Chen, Feng Li, Yuhui Quan, Yong Xu, and Hui Ji. Deep texture recognition via exploiting cross-layer statistical self-similarity. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5227-5236, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.344, + 0.482, + 0.385 + ], + "angle": 0, + "content": "[5] Zhile Chen, Yuhui Quan, Ruotao Xu, Lianwen Jin, and Yong Xu. Enhancing texture representation with deep tracing pattern encoding. Pattern Recognition, 146:109959, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.387, + 0.482, + 0.442 + ], + "angle": 0, + "content": "[6] Mircea Cimpoi, Subhransu Maji, Iasonas Kokkinos, Sammy Mohamed, and Andrea Vedaldi. Describing textures in the wild. In 2014 IEEE Conference on Computer Vision and Pattern Recognition, pages 3606-3613, 2014. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.445, + 0.482, + 0.499 + ], + "angle": 0, + "content": "[7] Mircea Cimpoi, Subhransu Maji, and Andrea Vedaldi. Deep filter banks for texture recognition and segmentation. In 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 3828-3836, 2015. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.501, + 0.482, + 0.557 + ], + "angle": 0, + "content": "[8] Gabriella Csurka, Christopher R. Dance, Lixin Fan, Jutta Willamowski, and Cedric Bray. Visual categorization with bags of keypoints. In In Workshop on Statistical Learning in Computer Vision, ECCV, pages 1-22, 2004. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.559, + 0.482, + 0.613 + ], + "angle": 0, + "content": "[9] O.G. Cula and K.J. Dana. Compact representation of bidirectional texture functions. In Proceedings of the 2001 IEEE Computer Society Conference on Computer Vision and Pattern Recognition. CVPR 2001, pages I-I, 2001. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.616, + 0.483, + 0.67 + ], + "angle": 0, + "content": "[10] Ravishankar Evani, Deepu Rajan, and Shangbo Mao. Multiscale graph texture network. In Computer Vision - ECCV 2024, pages 218-235, Cham, 2025. Springer Nature Switzerland. 1, 2, 3, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.673, + 0.482, + 0.714 + ], + "angle": 0, + "content": "[11] Shin Fujieda, Kohei Takayama, and Toshiya Hachisuka. Wavelet convolutional neural networks for texture classification. arXiv preprint arXiv:1707.07394, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.716, + 0.482, + 0.77 + ], + "angle": 0, + "content": "[12] Robert M. Haralick, K. Shanmugam, and Its'Hak Dinstein. Textural features for image classification. IEEE Transactions on Systems, Man, and Cybernetics, SMC-3(6):610-621, 1973. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.773, + 0.482, + 0.828 + ], + "angle": 0, + "content": "[13] Yuting Hu, Zhiling Long, and Ghassan AlRegib. Multi-level texture encoding and representation (multer) based on deep neural networks. In 2019 IEEE International Conference on Image Processing (ICIP), pages 4410-4414, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.83, + 0.482, + 0.857 + ], + "angle": 0, + "content": "[14] Béla Julesz. Visual pattern discrimination. IRE transactions on Information Theory, 8(2):84-92, 1962. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.859, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[15] Hervé Jégou, Matthijs Douze, Cordelia Schmid, and Patrick Pérez. Aggregating local descriptors into a compact image representation. In 2010 IEEE Computer Society Conference" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.116, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.094, + 0.905, + 0.12 + ], + "angle": 0, + "content": "on Computer Vision and Pattern Recognition, pages 3304-3311, 2010. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.122, + 0.905, + 0.148 + ], + "angle": 0, + "content": "[16] Kenneth I Laws. Texture energy measures. In Proceedings of Image Understanding Workshop, 1979. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.15, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[17] Kenneth I. Laws. Rapid Texture Identification. In Image Processing for Missile Guidance, pages 376 - 381. International Society for Optics and Photonics, SPIE, 1980. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.192, + 0.905, + 0.245 + ], + "angle": 0, + "content": "[18] Thomas Leung and Jitendra Malik. Representing and recognizing the visual appearance of materials using three-dimensional textons. International journal of computer vision, 43:29-44, 2001. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.247, + 0.905, + 0.313 + ], + "angle": 0, + "content": "[19] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11966-11976, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.316, + 0.905, + 0.37 + ], + "angle": 0, + "content": "[20] Ilya Loshchilov and Frank Hutter. SGDR: stochastic gradient descent with warm restarts. In 5th International Conference on Learning Representations, ICLR 2017, Toulouse, France, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.372, + 0.905, + 0.425 + ], + "angle": 0, + "content": "[21] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In 7th International Conference on Learning Representations, ICLR 2019, New Orleans, LA, USA, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.428, + 0.905, + 0.468 + ], + "angle": 0, + "content": "[22] Shangbo Mao, Deepu Rajan, and Liang Tien Chia. Deep residual pooling network for texture recognition. Pattern Recognition, 112:107817, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.47, + 0.905, + 0.51 + ], + "angle": 0, + "content": "[23] Joshua Peeples, Weihuang Xu, and Alina Zare. Histogram layers for texture analysis. IEEE Transactions on Artificial Intelligence, 3(4):541-552, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.512, + 0.905, + 0.567 + ], + "angle": 0, + "content": "[24] Florent Perronnin and Christopher Dance. Fisher kernels on visual vocabularies for image categorization. In 2007 IEEE Conference on Computer Vision and Pattern Recognition, pages 1-8, 2007. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.568, + 0.905, + 0.635 + ], + "angle": 0, + "content": "[25] Michael J. Reale, Matthew Ochrym, Micah Church, Nicholas Goutermout, John Rubado, and Maria Cornacchia. Shape and texture aware graph processing. In 2020 IEEE Applied Imagery Pattern Recognition Workshop (AIPR), pages 1-18, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.637, + 0.905, + 0.663 + ], + "angle": 0, + "content": "[26] Theodore J Rivlin. Chebyshev polynomials. Courier Dover Publications, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.665, + 0.905, + 0.733 + ], + "angle": 0, + "content": "[27] Leonardo Scabini, Kallil M. Zielinski, Lucas C. Ribas, Wesley N. Gonçalves, Bernard De Baets, and Odemir M. Bruno. Radam: Texture recognition through randomized aggregated encoding of deep activation maps. Pattern Recognition, 143: 109802, 2023. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.735, + 0.905, + 0.775 + ], + "angle": 0, + "content": "[28] Lavanya Sharan, Ruth Rosenholtz, and Edward Adelson. Material perception: What can you see in a brief glance? Journal of Vision, 9(8):784-784, 2009. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.777, + 0.905, + 0.831 + ], + "angle": 0, + "content": "[29] Kaiyou Song, Hua Yang, and Zhouping Yin. Multi-scale boosting feature encoding network for texture recognition. IEEE Transactions on Circuits and Systems for Video Technology, 31(11):4269-4282, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.833, + 0.905, + 0.899 + ], + "angle": 0, + "content": "[30] Yang Song, Fan Zhang, Qing Li, Heng Huang, Lauren J. O'Donnell, and Weidong Cai. Locally-transferred fisher vectors for texture classification. In 2017 IEEE International Conference on Computer Vision (ICCV), pages 4922-4930, 2017. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.094, + 0.905, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "23431" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[31] Manik Varma and Andrew Zisserman. Classifying images of materials: Achieving viewpoint and illumination independence. In Computer Vision — ECCV 2002, pages 255–271, Berlin, Heidelberg, 2002. Springer Berlin Heidelberg. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.483, + 0.19 + ], + "angle": 0, + "content": "[32] Manik Varma and Andrew Zisserman. A statistical approach to texture classification from single images. International journal of computer vision, 62:61-81, 2005. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.192, + 0.482, + 0.26 + ], + "angle": 0, + "content": "[33] Yong Xu, Feng Li, Zhile Chen, Jinxiu Liang, and Yuhui Quan. Encoding spatial distribution of convolutional features for texture representation. In Advances in Neural Information Processing Systems, pages 22732-22744. Curran Associates, Inc., 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.263, + 0.482, + 0.318 + ], + "angle": 0, + "content": "[34] Jia Xue, Hang Zhang, Kristin Dana, and Ko Nishino. Differential angular imaging for material recognition. In 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 6940-6949, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.32, + 0.482, + 0.374 + ], + "angle": 0, + "content": "[35] Jia Xue, Hang Zhang, and Kristin Dana. Deep texture manifold for ground terrain recognition. In 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 558-567, 2018. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.376, + 0.482, + 0.444 + ], + "angle": 0, + "content": "[36] Zhijing Yang, Shujian Lai, Xiaobin Hong, Yukai Shi, Yongqiang Cheng, and Chunmei Qing. Dfaen: Double-order knowledge fusion and attentional encoding network for texture recognition. Expert Systems with Applications, 209: 118223, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.446, + 0.482, + 0.502 + ], + "angle": 0, + "content": "[37] Wei Zhai, Yang Cao, Jing Zhang, and Zheng-Jun Zha. Deep multiple-attribute-perceived network for real-world texture recognition. In 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 3612–3621, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.504, + 0.482, + 0.571 + ], + "angle": 0, + "content": "[38] Wei Zhai, Yang Cao, Zheng-Jun Zha, HaiYong Xie, and Feng Wu. Deep structure-revealed network for texture recognition. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11007-11016, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.574, + 0.482, + 0.642 + ], + "angle": 0, + "content": "[39] Wei Zhai, Yang Cao, Jing Zhang, Haiyong Xie, Dacheng Tao, and Zheng-Jun Zha. On exploring multiplicity of primitives and attributes for texture recognition in the wild. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(1):403-420, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.644, + 0.482, + 0.698 + ], + "angle": 0, + "content": "[40] Hang Zhang, Jia Xue, and Kristin Dana. Deep ten: Texture encoding network. In 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2896-2905, 2017. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.698 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23432" + } + ] +] \ No newline at end of file diff --git a/2025/Chebyshev Attention Depth Permutation Texture Network with Latent Texture Attribute Loss/44aa142c-41b8-482e-a039-e4a4fc955e21_origin.pdf b/2025/Chebyshev Attention Depth Permutation Texture Network with Latent Texture Attribute Loss/44aa142c-41b8-482e-a039-e4a4fc955e21_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..686485e4a34f4bd2d8166d514cfd32d3421f0ddc --- /dev/null +++ b/2025/Chebyshev Attention Depth Permutation Texture Network with Latent Texture Attribute Loss/44aa142c-41b8-482e-a039-e4a4fc955e21_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:215b6d7455f9f8ef1857e346e7f9daece14b159eb1d44a56a08661ef78b26d90 +size 3863256 diff --git a/2025/Chebyshev Attention Depth Permutation Texture Network with Latent Texture Attribute Loss/full.md b/2025/Chebyshev Attention Depth Permutation Texture Network with Latent Texture Attribute Loss/full.md new file mode 100644 index 0000000000000000000000000000000000000000..4caa019d5fc5d000c583e2f40cbf4ec0dbb83811 --- /dev/null +++ b/2025/Chebyshev Attention Depth Permutation Texture Network with Latent Texture Attribute Loss/full.md @@ -0,0 +1,350 @@ +# Chebyshev Attention Depth Permutation Texture Network with Latent Texture Attribute Loss + +Ravishankar Evani, Deepu Rajan, Shangbo Mao + +College of Computing and Data Science, Nanyang Technological University, Singapore + +S220007@e.ntu.edu.sg, ASDRajan@ntu.edu.sg, MAOS0003@e.ntu.edu.sg + +# Abstract + +Despite recent advances in deep texture recognition, existing methods still lack representational diversity and struggle to capture and preserve discriminative cues across stages of representation hierarchies. Moreover, many rely on loss formulations that prioritize recognition accuracy while overlooking spatial coherence and statistical consistency in the feature space. To address these issues, we propose three key innovations: (1) Stochastic Local Texture Masking (SLTM), a regularization strategy that randomly occludes small texture patches to promote the learning of broader spatial and contextual dependencies; (2) the Chebyshev Attention Depth Permutation Texture Network (CAPTN), a novel architecture that learns expressive and persistent Latent Texture Attribute (LTA) representations. CAPTN integrates a Texture Frequency Attention (TFA) module that generates Latent Texture Attributes (LTAs) and enables frequency-aware interpretability, a Dual Depth Permutation $(D^{2}P)$ module to expose complementary channel adjacency patterns, and Learnable Chebyshev Polynomials (LCPs) to model high-order orderless LTA transformations via recursive Chebyshev basis expansion; and (3) a Latent Texture Attribute Loss that jointly optimizes classification accuracy, statistical alignment, and spatial fidelity. CAPTN supports end-to-end training without relying on fine-tuned CNN backbones and achieves state-of-the-art performance on several texture and material recognition benchmarks. (Code: https://github.com/RavishankarEvani/CAPTN) + +# 1. Introduction + +Visual texture perception plays a critical role in understanding an object's surface characteristics. Distinct from object recognition, which often relies on shape cues, texture recognition emphasizes the repetitive patterns or stochastic microstructures that define surface appearance. Over the years, texture recognition has evolved appreciably, be + +ginning with foundational work such as Julesz's theory of textons [14], Haralick et al.'s co-occurrence matrices [12], and Laws' texture energy measures based on convolution masks [16, 17]. These early methods established the basis for statistical and structural texture analysis and were widely adopted in classical image processing pipelines. However, while effective under controlled conditions, such traditional techniques often fail to capture the rich diversity and structural complexity of real-world textures. + +With the rise of data-driven approaches, modern deep learning methods have largely supplanted handcrafted descriptors by offering more expressive and hierarchical feature representations. Despite this progress, deep models still encounter challenges when applied to texture recognition. In particular, successive convolution and pooling operations can suppress or discard fine-grained texture cues, limiting the network's ability to retain localized and discriminative information. Furthermore, textures often exhibit overlapping or perceptually similar traits [37], which contribute to inter-class confusion by blurring the boundaries between distinct texture categories. For instance, textures labeled as "stratified" might closely resemble those labeled "cracked" or "zigzagged," making it difficult to consistently assign the correct class. In addition, intra-class variability arises from variations within the same category due to differences in lighting, scale, deformation, or inherent surface irregularities. + +The concept of texture attributes [10, 37, 39] has been shown to play a crucial role in representing perceptually similar textures. Latent texture attributes [10] refer to the underlying semantic components of visual texture attributes, where each latent texture attribute may correspond to one or more perceptual texture attributes. Effectively manipulating and transforming these latent attributes can enhance texture discrimination. Several studies employing Convolutional Neural Network (CNN)-based approaches have explored different aggregation methods, including global average pooling [1, 11, 37], Fisher vectors [7], feature encoding [35, 40], and cross-layer aggregation. However, certain aggregation and dimensionality + +reduction techniques lead to the loss of important texture cues, and little attention has been given to recovering discriminative information post-aggregation. + +To address these limitations, we introduce a comprehensive texture recognition framework that advances both architectural design and training methodology. Central to our approach is the Chebyshev Attention Depth Permutation Texture Network (CAPTN), which extracts expressive and interpretable LTAs, depth permutation, and polynomial projection. In addition, we propose a stochastic local texture masking (SLTM) strategy and a tailored multi-objective loss function that together encourage spatially grounded and orderless texture representations. Our main contributions are as follows: + +- We propose the CAPTN architecture, which integrates three key components: +- A Texture Frequency Attention (TFA) module that generates Latent Texture Attributes (LTAs) and enables visual interpretability by highlighting discriminative texture regions across frequency scales. +- A Dual Depth Permutation $(\mathrm{D}^2\mathrm{P})$ module that applies two distinct channel shuffling operations to expose complementary spatial relationships and enhance feature diversity. +- A Learnable Chebyshev Polynomial (LCP) expansion that enhances the representational power of each orderless LTA. +- We introduce a Stochastic Local Texture Masking (SLTM) strategy that randomly occludes small patches of the input image during training. This compels the model to leverage broader spatial context and enables a new supervision signal that aligns masked patches with regional latent features. +- We formulate a multi-objective LTA loss that jointly supervises classification accuracy, orderless latent attribute consistency, and spatial latent attribute alignment. + +# 2. Related Work + +Traditional Texture Recognition. Building upon the foundational work discussed in Sec. 1, influential methods such as the Bag of Textons (BoT) [9, 18] and Bag of Words (BoW) [8] introduced a paradigm for representing textures as orderless histograms over learned dictionaries of textural elements. These methods encode local patches, aggregate them into global descriptors, and classify the results using algorithms such as Support Vector Machines (SVM). Varma et al. [31, 32] advanced this line of work by learning texture descriptors directly from data. Subsequent developments led to residual encoding techniques—Jégou et al. [15] proposed VLAD, which captures first-order residuals relative to visual codewords, while Perronnin et al. [24] introduced Fisher Vectors (FV), a probabilistic extension that also models second-order statistics. + +Deep Texture Recognition. Deep learning-based texture recognition methods such as FV-CNN [7] and LFV [30] primarily focus on encoding features extracted from pretrained CNNs, but they do not support end-to-end training. To address this limitation, Zhang et al. [40] proposed DeepTEN, integrating residual encoding and dictionary learning into a single layer. The layer functions analogously to VLAD and FV while enabling end-to-end training with a CNN backbone. RPNet [22] simplifies Deep-TEN by replacing dictionary learning with residual pooling. DEP [35] enhances Deep-TEN further by combining orderless texture features with spatial cues to balance structure and texture. + +MuLTER [13] improves DEP by integrating multi-layer CNN features, preserving fine texture and spatial details. MSBFEN [29] combines multiscale encoding and boosting to better capture discriminative texture features. Chen et al. [4] proposed using cross-layer statistical self-similarity to capture consistent patterns across CNN layers. Later, they introduced DTPNet [5], a local encoding scheme inspired by Local Binary Patterns (LBP), robust to rotation and grayscale changes. Yang et al. [36] proposed DFAEN, which uses frequency attention and first- and second-order statistics to enhance texture details. RADAM [27] fuses CNN activation maps via randomized aggregation, promoting texture diversity and reducing overfitting. + +FENet [33] encodes spatial regularity via fractal analysis, while DSR-Net [38] models local spatial dependencies instead of orderless aggregation. Zhai et al. [37] earlier proposed MAP-Net, which learns texture-related visual attributes and uses spatially-adaptive pooling for feature fusion. Their recent work, MPAP [39], uses a multi-branch architecture to model bottom-up structures and top-down attribute relationships by learning local-to-global mappings. Reale et al. [25] proposed modeling triangular patches as graph nodes to encode texture patterns. More recently, GTN [10] introduced a graph-based deep approach designed to model directional relationships between Latent Texture Attributes (LTAs). GTN enhances LTAs through message passing and aggregation of semantically related information. + +Deep Polynomial Expansion. HistNet [23] features a trainable histogram layer that preserves spatial context using learnable bin centers and widths, combined with Radial Basis Functions (RBFs) for robustness to intra-class variability. + +While graph-based methods explicitly learn attribute adjacencies for message passing to refine LTAs, our CAPTN framework adopts a more flexible and implicit approach. It enables rich and diverse feature interactions without the need for explicit graph construction. This not only simplifies the overall architecture but also implicitly regularizes the learning process by breaking fixed feature layouts, thereby enhancing texture robustness and generaliza + +![](images/30a2d67006fcfed2401ca74cc7386033f0c042feb7d6106c78018db81ed061f7.jpg) +Figure 1. Stochastic Local Texture Masking (SLTM) & components of Chebyshev Attention Depth Permutation Texture Network (CAPTN): SLTM facilitates contextual texture learning, Texture Frequency Attention (TFA) captures multi-frequency texture cues, Dual Depth Permutation $(\mathrm{D}^2\mathrm{P})$ promotes feature diversification, and Learnable Chebyshev Polynomials (LCP) map latent texture attributes to visual texture attributes. The training process optimizes Latent Texture Attribute (LTA) Loss, which balances classification accuracy with orderless and spatial texture consistency. || represents concatenation, $\times$ represents matrix multiplied by a scalar. + +tion. Unlike radial basis functions (RBFs), which often require numerous kernel centers and careful bandwidth tuning, we propose a Chebyshev polynomial-based method that provides a compact yet globally expressive basis. This approach eliminates the need for kernel-specific design choices while efficiently capturing complex and non-linear texture variations. + +# 3. Stochastic Local Texture Masking + +To enhance the model's ability to capture local, global, and contextual texture representations, we introduce Stochastic Local Texture Masking (SLTM). This technique randomly masks a small patch of the input texture image prior to passing it through the pre-trained backbone, thereby preventing the Chebyshev Attention Depth Permutation Texture Network (CAPTN) from over-relying on fine-grained local details. Given an input image $\mathbf{I} \in \mathbb{R}^{H \times W \times 3}$ , SLTM generates: (1) a masked image $\mathbf{I}_r$ , where a randomly selected patch of size $P \times P$ ( $P \ll H, P \ll W$ ) is zeroed out, and (2) the extracted patch $\mathbf{G}_t \in \mathbb{R}^{P \times P \times 3}$ , which retains the original pixel values prior to masking. By selectively removing texture information, SLTM encourages CAPTN to exploit broader spatial dependencies, thereby enhancing global feature learning and reducing overfitting. Finally, the masked image $\mathbf{I}_r$ is fed through the backbone and CAPTN for further processing. This module is activated during the computation of the Spatial Latent Texture Attribute Loss $(\mathcal{L}_{SL})$ , as described in Section 5. + +# 4. Chebyshev Attention Depth Permutation Texture Network + +The Chebyshev Attention Depth Permutation Texture Network (CAPTN) consists of three key modules: (1) Texture Frequency Attention (TFA) Network, (2) Dual Depth Permutation $(\mathrm{D}^2\mathrm{P})$ Network, and (3) Learnable Chebyshev Polynomials (LCP). Following [10], CAPTN extracts latent texture attributes (LTAs) by computing spatial texture frequency information from feature maps $\mathcal{F} = \{\mathbf{X}^{o_m}\mid 1\leq m\leq M - 1\}$ obtained from an $M$ -layer pre-trained CNN backbone, excluding the patchify layer [19]. Depth permutation enhances LTA representation by diversifying feature space and reducing inter-class variability, while LCP improves expressiveness of orderless LTAs. + +# 4.1. Texture Frequency Attention + +Textures display a wide range of spatial frequency characteristics, from high-frequency (fine-grained) patterns to low-frequency (smooth or homogeneous) regions. To effectively capture and interpret these variations, we propose a Texture Frequency Attention (TFA) mechanism. In addition to emphasizing discriminative texture regions, TFA also transforms backbone feature maps into Latent Texture Attributes (LTAs). Let $\mathbf{X}^{o_m} \in \mathbb{R}^{H_m \times W_m \times C_m}$ be the $m^{\text{th}}$ layer feature map. We compute transformed features as $\mathbf{X}^{t_m} = \mathbf{W}_l^m * \mathbf{X}^{o_m} + \mathbf{b}_l^m$ , where $\mathbf{W}_l^m \in \mathbb{R}^{C_m \times C_m \times 1 \times 1}$ is the convolution kernel, * denotes the convolution operation, and $\mathbf{b}_l^m \in \mathbb{R}^{C_m}$ is a learnable bias vector. Next, we + +![](images/a67d04c227cb95c0ffaec0618099b40c232fc0bdd64550c26fd1f5280d1e9ebc.jpg) +(a) + +![](images/eeb164c67c23b70873281feae9075d050d251132db33e1e31961d72bbcb4857f.jpg) +(b) + +![](images/1ec616705a901841a4cbd66393e74b3dbad94522a53d96397e7e6030e4109e6a.jpg) +(c) + +![](images/85e080db0b7ffa8cfb51b76bdf45958f3a8d24a07db96c9a889df5e6f37366af.jpg) +(d) + +![](images/8065d32f9b6ac3509639a9700d48f47eee3e6c862290ffe33c4f252e3d8bafb9.jpg) +(e) +Figure 2. (a) SLTM: extract a small patch from an input texture image, (b) $\mathrm{D}^2\mathrm{P}$ : feature maps from backbone undergo depth permutation to diversify feature space, (c) LCP: enhanced Latent Texture Attributes undergo orderless aggregation and fit a learnable Chebyshev polynomial function, (d) TFA: use spatial texture frequencies to compute spatial attention, (e) SLAR: representation of spatial information for latent attributes and backbone features. SLTM and SLAR are used to compute Latent Texture Attribute Loss. || represents concatenation, $\times$ represents matrix multiplied by a scalar, $\cdot$ represents matrix multiplication. + +compute per-location channel-wise maximum and average statistics + +$$ +X _ {\max } ^ {t _ {m}} (i, j) = \max _ {0 \leq c < C _ {m}} X ^ {t _ {m}} (i, j, c), \tag {1} +$$ + +$$ +X _ {\text {a v g}} ^ {t _ {m}} (i, j) = \frac {1}{C _ {m}} \sum_ {c = 0} ^ {C _ {m} - 1} X ^ {t _ {m}} (i, j, c). \tag {2} +$$ + +Here, $\mathbf{X}_{\mathrm{max}}^{t_m}, \mathbf{X}_{\mathrm{avg}}^{t_m} \in \mathbb{R}^{H_m \times W_m \times 1}$ represent spatial maps of local maxima and average activations across channels. High-frequency textures tend to exhibit greater local variation, resulting in larger differences between maximum and average values, whereas low-frequency regions yield more consistent responses. To capture this contrast, we concatenate $\mathbf{X}_{\mathrm{max}}^{t_m}$ and $\mathbf{X}_{\mathrm{avg}}^{t_m}$ along the channel dimension and apply a convolution followed by a sigmoid activation + +$$ +\mathbf {A} ^ {m} = \frac {1}{1 + e ^ {- \left[ \mathbf {W} _ {a} ^ {m} * \left(\mathbf {X} _ {\operatorname* {m a x}} ^ {t _ {m}} \parallel \mathbf {X} _ {\operatorname* {a v g}} ^ {t _ {m}}\right) \right]}}, \tag {3} +$$ + +where $\parallel$ denotes channel-wise concatenation and $\mathbf{W}_a^m\in$ $\mathbb{R}^{1\times 2\times 1\times 1}$ is the convolution kernel. The resulting spatial attention map $\mathbf{A}^m\in \mathbb{R}^{H_m\times W_m\times 1}$ is used to modulate $\mathbf{X}^{t_m}$ via element-wise multiplication + +$$ +\begin{array}{l} X _ {\text {a t t n}} ^ {t _ {m}} (i, j, c) = X ^ {t _ {m}} (i, j, c) \times A ^ {m} (i, j), \\ \forall (i, j, c) \in [ 0, H _ {m} - 1 ] \times [ 0, W _ {m} - 1 ] \times [ 0, C _ {m} - 1 ], \tag {4} \\ \end{array} +$$ + +A residual connection is then added, combining the attention-weighted features with the original $\mathbf{X}^{t_m}$ . The result is passed through a transformation $g(\cdot)$ comprising Batch Normalization, GELU activation, and dropout + +$$ +\mathbf {X} _ {a} ^ {t _ {m}} = f _ {a} \left(\gamma_ {m}\right) \times g \left(\mathbf {X} _ {\text {a t t n}} ^ {t _ {m}} + \mathbf {X} ^ {t _ {m}}\right), \tag {5} +$$ + +where $\gamma_{m}$ is a learnable scalar, and $f_{a}(\cdot)$ ensures the scaling factor remains within the range [0, 1] + +$$ +f _ {a} (z) = \min (\max (0, z), 1), \quad z \in \mathbb {R}. \tag {6} +$$ + +While the convolution and $g(\cdot)$ transformation generate the LTAs, the spatial attention map $\mathbf{A}^m$ provides interpretable insights into the model's dynamic focus on spatially discriminative texture regions. + +# 4.2. Dual Depth Permutation + +Dual Depth Permutation $(\mathrm{D}^2\mathrm{P})$ operates directly on backbone features to generate complementary representations through two distinct shuffling operations. Exposing varied adjacency patterns promotes rich cross-channel correlations. The outputs from both shuffling paths are later + +fused with LTAs from Sec. 4.1 to enhance representation diversity and improve texture robustness. Additionally, $\mathrm{D}^2\mathrm{P}$ serves as an implicit regularizer, reducing overfitting by disrupting fixed feature layouts. Drawing inspiration from constraining [2], $\mathrm{D}^2\mathrm{P}$ introduces multiple feature perspectives by permuting channel adjacency structures in a supervised setting. Let $\mathbf{X}^{o_m} \in \mathbb{R}^{H_m \times W_m \times C_m}$ denote the feature map output from the $m^{\text{th}}$ layer of the backbone. We apply two distinct channel shuffles using group sizes $s_0$ and $s_1$ , ensuring $\frac{C_m}{s_0}, \frac{C_m}{s_1} \in \mathbb{Z}$ . Let $p \in \{0,1\}$ index the two shuffling paths, with corresponding group sizes $k_p^m \in \{s_0, s_1\}$ + +$$ +\mathbf {X} ^ {o _ {m}} \rightarrow \mathbf {X} _ {\mathrm {r}, p} ^ {o _ {m}} \in \mathbb {R} ^ {H _ {m} \times W _ {m} \times k _ {p} ^ {m} \times g _ {k _ {p} ^ {m}}}, \tag {7} +$$ + +$$ +\mathbf {X} _ {\mathrm {r}, p} ^ {o _ {m}} \rightarrow \mathbf {X} _ {\mathrm {r}, p} ^ {o _ {m} ^ {T}} \in \mathbb {R} ^ {H _ {m} \times W _ {m} \times g _ {k _ {p} ^ {m}} \times k _ {p} ^ {m}}, \tag {8} +$$ + +where $g_{k_p^m}$ is the number of groups formed using group size $k_{p}^{m}$ at layer $m$ . After flattening $\mathbf{X}_{\mathrm{r},p}^{o_m^T}$ , we obtain + +$$ +\mathbf {X} _ {\mathrm {s}, p} ^ {o _ {m}} \in \mathbb {R} ^ {H _ {m} \times W _ {m} \times C _ {m}}. \tag {9} +$$ + +Each of the shuffled representations is processed by a convolution operation, where $\mathbf{W}_p^m\in \mathbb{R}^{C_m\times C_m\times 1\times 1}$ is the convolution kernel, $*$ denotes the convolution operation, and $\mathbf{b}_p^m\in \mathbb{R}^{C_m}$ is a learnable bias vector. The result is then passed through a transformation $f_{p}(\cdot)$ , consisting of Batch Normalization, GELU activation, and Dropout. The final output of the $\mathrm{D}^2\mathrm{P}$ module is obtained by summing the outputs from both shuffling paths + +$$ +\mathbf {X} _ {\mathrm {s}} ^ {d _ {m}} = f _ {a} (\psi_ {m}) \times \left[ \sum_ {p \in \{0, 1 \}} f _ {p} \left(\mathbf {W} _ {p} ^ {m} * \mathbf {X} _ {\mathrm {s}, p} ^ {o _ {m}} + \mathbf {b} _ {p} ^ {m}\right) \right], +$$ + +$$ +\psi_ {m} \in (- \infty , \infty), \quad f _ {a} (\cdot) \in [ 0, 1 ], \tag {10} +$$ + +where $\psi_{m}$ is a learnable scalar parameter, and the scaling function $f_{a}(\cdot)$ is defined in Eq. (6) to constrain the modulation factor within the range [0, 1]. + +# 4.3. Enhanced Orderless Latent Texture Attributes + +While GTN [10] enhances LTA representations via message passing and aggregation, CAPTN improves the LTAs generated by the TFA module, $\mathbf{X}_a^{t_m}$ , by injecting dynamically adapted adjacency structures from the $\mathrm{D}^2\mathrm{P}$ module, $\mathbf{X}_s^{d_m}$ , resulting in + +$$ +\mathbf {X} _ {e} ^ {m} = \mathbf {X} _ {a} ^ {t _ {m}} + \mathbf {X} _ {\mathrm {s}} ^ {d _ {m}}. \tag {11} +$$ + +We then compute the spatial mean over the height and width of the enhanced LTAs, $\mathbf{X}_e^m\in \mathbb{R}^{H_m\times W_m\times C_m}$ , from layer $m$ to obtain an orderless representation, $\mathbf{O}_{\mathrm{e}}^{m}\in \mathbb{R}^{C_{m}}$ . + +# 4.4. Learnable Chebyshev Polynomial + +Chebyshev polynomials [26] offer a powerful global basis for modeling non-linear, oscillatory, and fine-grained variations in orderless LTAs. Their higher-order, non-monotonic structure enables expressive transformations that capture both smooth and oscillatory variations, allowing the model to distinguish subtle patterns while aligning semantically similar LTA values—even when numerically distant. To enhance the expressiveness of each orderless LTA dimension, we apply a Chebyshev polynomial $T_{n}(\cdot)$ for $n \in \{0, \dots, d\}$ , each weighted by a sigmoid function parameterized by $\theta_{n}$ . The resulting formulation, referred to as a Learnable Chebyshev Polynomial (LCP), takes the form + +$$ +P \left(\mathbf {O} _ {\mathrm {e}} ^ {m}\right) = \sum_ {n = 0} ^ {d} T _ {n} \left(\mathbf {O} _ {\mathrm {e}} ^ {m}\right) \left[ \frac {1}{1 + e ^ {- \theta_ {n}}} \right], \tag {12} +$$ + +The Chebyshev polynomials $T_{n}(\mathbf{O}_{\mathrm{e}}^{m})$ used in the expansion follow the recurrence relation + +$$ +T _ {n} \left(\mathbf {O} _ {\mathrm {e}} ^ {m}\right) = \left\{ \begin{array}{l l} \mathbf {1}, & \text {i f} n = 0, \\ \mathbf {O} _ {\mathrm {e}} ^ {m}, & \text {i f} n = 1, \\ 2 \mathbf {O} _ {\mathrm {e}} ^ {m} T _ {n - 1} \left(\mathbf {O} _ {\mathrm {e}} ^ {m}\right) & \text {f o r} n \geq 2. \\ - T _ {n - 2} \left(\mathbf {O} _ {\mathrm {e}} ^ {m}\right), & \end{array} \right. \tag {13} +$$ + +LCPs from all layers are then concatenated and passed through a fully connected layer, $f_{fc}$ , as + +$$ +\mathbf {Y} = f _ {f c} \left( \right.\left\| \right. _ {m = 1} ^ {N - 1} P \left(\mathbf {O} _ {\mathrm {e}} ^ {m}\right)\left. \right). \tag {14} +$$ + +The logits, $\mathbf{Y}$ , are then evaluated using the Latent Texture Attribute Loss function, which will be discussed in Sec. 5. + +# 4.5. Spatial Latent Attribute Representation + +Finally, we aim to enforce some extent of structural similarity between a texture patch extracted from Sec. 3 and regional latent texture attribute information. To achieve this, we first concatenate the enhanced LTAs, $\mathbf{X}_e^{M - 1}$ , with the original feature maps from the frozen backbone, $\mathbf{X}^{O_M - 1}$ . Both representations are obtained from the final layer of the network. We integrate the frozen backbone features into the LTAs through a structured transformation consisting of three sequential operations: cropping $f_{c}(\cdot)$ , a $3\times 3$ convolution, and bilinear interpolation $f_{b}(\cdot)$ . The transformation is formally defined as + +$$ +\mathbf {G} _ {\mathrm {p}} = f _ {b} \left(\mathbf {W} _ {s} * f _ {c} \left(\mathbf {X} _ {e} ^ {M - 1} \| \mathbf {X} ^ {O M - 1}\right) + \mathbf {b} _ {s}, P\right), \tag {15} +$$ + +where $\mathbf{W}_s\in \mathbb{R}^{3\times C\times 3\times 3}$ and $\mathbf{b}_s\in \mathbb{R}^3$ are the learnable weights and bias of the convolutional layer. The output of this transformation is the Spatial Latent Attribute Representation (SLAR), $\mathbf{G}_{\mathrm{p}}\in \mathbb{R}^{3\times P\times P}$ , where $P$ is the spatial size + +of the extracted patch. This ensures that the extracted regional texture attributes retain both the fine-grained information from LTAs and the structural guidance preserved in the frozen backbone features. Incorporating structure-aware features from the frozen backbone provides a grounded reference that constrains the adaptation of LTAs during training, preventing overcorrection while maintaining a balance between texture fidelity and spatial consistency. SLAR will be used to compute Spatial Latent Texture Attribute Loss (SL) in Sec. 5. + +# 5. Latent Texture Attribute Loss + +We propose the Latent Texture Attribute (LTA) Loss to enhance classification accuracy while preserving texture consistency at both global and local levels. The LTA Loss integrates three complementary components: (1) Cross Entropy Loss $(\mathcal{L}_{CE})$ to promote accurate class predictions, (2) Orderless LTA Loss $(\mathcal{L}_{OL})$ to align global, spatially invariant texture distributions, and (3) Spatial LTA Loss $(\mathcal{L}_{SL})$ to preserve fine-grained, spatial texture structure. By jointly optimizing these objectives, the model learns rich and coherent latent texture representations that support both accurate classification and texture understanding. + +Orderless Latent Texture Attribute Loss. This second texture loss component compares distribution of orderless LTAs with orderless features extracted directly from the frozen backbone. By aligning loss from learned texture representations with a stable baseline, we encourage consistency in the model's understanding of spatial invariant texture representations across images. We apply Batch Normalization, $f_{\mathrm{bn}}$ , on feature maps from the $m^{\mathrm{th}}$ layer of the backbone, $\mathbf{X}^{o_m}$ , before calculating the spatial mean and concatenating multiple scales of orderless texture information across layers $m \in [1, M - 1]$ of the backbone to obtain $\mathbf{O}_{\mathrm{b}}$ . We also concatenate orderless enhanced LTAs, $\mathbf{O}_{\mathrm{e}}^{m}$ , across the same range of layers to obtain $\mathbf{O}_{\mathrm{e}}$ + +$$ +\mathbf {O} _ {\mathrm {b}} = \left\| \right. _ {m = 1} ^ {M - 1} \left(\frac {\sum_ {h = 0} ^ {H _ {m} - 1} \sum_ {w = 0} ^ {W _ {m} - 1} f _ {\mathrm {b n}} \left(\mathbf {X} ^ {o _ {m}}\right) [ h , w , : ]}{H _ {m} \times W _ {m}}\right), \tag {16} +$$ + +$$ +\mathbf {O} _ {\mathrm {e}} = \left\| _ {m = 1} ^ {M - 1} \mathbf {O} _ {\mathrm {e}} ^ {m}. \right. \tag {17} +$$ + +We perform softmax normalization of $\mathbf{O_b}$ and $\mathbf{O_e}$ over the total number of channels in each sample $s$ from the minibatch, aggregated over layers $m\in [1,M - 1]$ , where the total number of concatenated channels is $C = \sum_{m = 1}^{M - 1}C_m$ . A small constant $\epsilon$ is added for numerical stability + +$$ +p _ {\alpha , i} ^ {s} = \frac {e ^ {\mathbf {O} _ {\alpha , i} ^ {s}}}{\sum_ {j = 0} ^ {C - 1} e ^ {\mathbf {O} _ {\alpha , j} ^ {s}}} + \epsilon , \quad \alpha \in \{\mathrm {b}, \mathrm {e} \}. \tag {18} +$$ + +Inspired by Kullback-Leibler (KL) divergence, we calculate orderless LTA loss, $\mathcal{L}_{OL}$ that measures the dissimilarity between two probability distributions: a fixed distribution from orderless feature maps extracted from the frozen backbone, $p_{\mathrm{b},i}^{\mathrm{s}}$ and that from orderless enhanced LTAs, $p_{\mathrm{e},i}^{\mathrm{s}}$ + +$$ +\mathcal {L} _ {O L} = \frac {1}{S} \sum_ {s = 0} ^ {S - 1} \sum_ {i = 0} ^ {C - 1} p _ {\mathrm {b}, i} ^ {s} \left[ \log_ {e} \left(p _ {\mathrm {b}, i} ^ {s}\right) - \log_ {e} \left(p _ {\mathrm {e}, i} ^ {s}\right) \right], \tag {19} +$$ + +where $S$ represents the number of samples in the minibatch. + +Spatial Latent Texture Attribute Loss. The final LTA loss component measures the similarity between SLAR from Sec. 4.5 and a corresponding texture patch extracted from the original image. This component helps ensure that the latent representations preserve both regional and local spatial texture information by comparing the latent attributes $\mathbf{G}_{\mathrm{p}}$ with the extracted texture patch $\mathbf{G}_{\mathrm{t}}$ using the Mean Squared Error (MSE) + +$$ +\mathcal {L} _ {S L} = \frac {1}{B C _ {M - 1} H _ {M - 1} W _ {M - 1}} \| \mathbf {G} _ {\mathrm {p}} - \mathbf {G} _ {\mathrm {t}} \| _ {F} ^ {2}, \tag {20} +$$ + +where $B, C_{M-1}, H_{M-1}, W_{M-1}$ represent the batch size, number of channels, height, and width of the feature maps, from the final $M-1$ layer, and $\|\cdot\|_F$ denotes the Frobenius norm, which sums the squared differences over all elements. This loss encourages the model to preserve fine-grained spatial texture patterns within the latent space, ensuring that learned texture features remain spatially structured and contextually coherent. By aligning latent spatial representations with real texture patches, the model enhances its ability to reconstruct or classify textures with high fidelity. + +Latent Texture Attribute Loss. We formulate our loss during the training phase as a weighted sum of the three loss components + +$$ +\mathcal {L} _ {\text {t r a i n}} = \lambda \mathcal {L} _ {C E} + \mu \mathcal {L} _ {O L} + \nu \mathcal {L} _ {S L}, \tag {21} +$$ + +where $\lambda, \mu, \nu \in [0,1]$ are weights associated with each of the loss components. + +# 6. Experiments + +# 6.1. Datasets + +We evaluated our network using five benchmark datasets, summarized as follows: (a) Ground Terrain in Outdoor Scenes (GTOS) [34] contains 34,105 images of outdoor ground terrain across 40 categories, with 5 train/test splits. + +Table 1. Comparison of CAPTN with GTN [10] and RADAM [27] in terms of classification accuracy (\%) on texture benchmarks. ConvNeXt-\* variants are used as backbones. Pre-trained weights are frozen. The best result for each backbone-dataset combination is bolded. + +
Model-*DTDFMDKTHGTOSGTOS-M
RADAMN74.9 ±0.787.1 ±0.489.6 ±3.883.7 ±1.581.8
GTNN75.4 ±1.186.9 ±3.190.3 ±4.782.9 ±1.790.1
CAPTN (ours)N77.9 ±1.087.0 ±2.990.6 ±4.683.9 ±1.791.5
RADAMT77.0 ±0.788.7 ±0.490.7 ±4.084.2 ±1.785.3
GTNT77.5 ±1.087.2 ±3.492.0 ±4.984.7 ±1.691.9
CAPTN (ours)T78.9 ±0.887.2 ±3.692.2 ±4.485.1 ±1.794.2
RADAMB76.4 ±0.990.2 ±0.287.7 ±5.684.1 ±1.682.2
GTNB78.3 ±0.888.8 ±3.090.8 ±6.284.2 ±1.894.8
CAPTN (ours)B79.3±0.989.1 ±3.491.3±5.784.6 ±1.994.0
RADAML77.4 ±1.189.3±0.389.3 ±3.484.0 ±1.885.8
GTNL79.3 ±0.888.1 ±3.092.5 ±4.784.9 ±1.793.9
CAPTN (ours)L79.6±0.889.3±3.193.0 ±4.985.0 ±1.795.1
+ +Table 4. Classification accuracy (\%) comparison between $1^{\text{st}}$ -order mean pooling (w/ LCP) and a non-polynomial baseline: $2^{\text{nd}}$ -order covariance pooling (w/o LCP), with ConvNeXt-B backbone. + +
PoolingDTDKTHGTOSGTOS-M
2nd-order76.4 ±1.091.1 ±6.383.5 ±1.793.0
1st-order79.3 ±0.991.3 ±5.784.6 ±1.994.0
+ +Table 5. Classification accuracy (\%) of CAPTN with different basis functions using ConvNeXt-L backbone. + +
Basis Fn.DTDKTHGTOSGTOS-M
RBF74.1 ±1.089.6 ±6.881.5 ±2.492.7
LCP79.6±0.893.0±4.985.0±1.795.1
+ +(b) GTOS-Mobile (GTOS-M) [35] includes 100,011 material samples from 31 categories, captured via mobile phone, with a single train/test split. (c) KTH-TIPS2-b (KTH) [3] provides 4,752 images in 11 material categories, with 4 train/test splits. (d) Flickr Material Dataset (FMD) [28] contains 1,000 images spanning 10 material categories, validated with 10-fold cross-validation and stratified sampling to maintain class proportions. (e) Describable Texture Dataset (DTD) [6] has 5,640 images across 47 categories, with 10 train/validation/test splits. + +# 6.2. Implementation Details + +Our model was implemented in PyTorch and trained on a single NVIDIA RTX A6000 GPU. For all datasets, the learning rate was set to 0.0001 and followed a cosine annealing schedule [20]. We used the Adam optimizer with decoupled weight decay [21] and our novel Latent Texture Attribute Loss function. Our network was trained until the final epoch for each dataset, with test set performance reported. More design and implementation details can be found in the supplementary material. + +Table 2. Evaluate effectiveness of TFA, ${\mathrm{D}}^{2}\mathrm{P}$ ,and LCP in terms of classification accuracy (%) . ConvNeXt-N is the backbone. + +
TFAD2PLCPDTDKTHGTOS
XXX75.4 ±1.087.2 ±5.381.7 ±2.0
✓XX77.1 ±1.190.4 ±4.883.5 ±1.8
✓✓X77.2 ±1.090.5 ±4.883.6 ±1.8
✓✓✓77.9 ±1.090.6 ±4.683.9 ±1.7
+ +Table 3. Evaluate effectiveness of LTA Loss in terms of classification accuracy $(\%)$ . ConvNeXt-T is the backbone. + +
Loss Fn.DTDKTHGTOSGTOS-M
CE Loss78.4 ±0.791.9 ±5.185.0 ±1.893.6
LTA Loss78.9 ±0.892.2 ±4.485.1 ±1.794.2
+ +# 6.3. Analysis + +Recent studies [10, 27] have demonstrated the superiority of ConvNeXt backbones over ResNet backbones for texture-specific applications, particularly in multiscale texture representation. Hence, all our experiments will be based on ConvNeXt backbones. + +Quantitative. Tab. 1 depicts how CAPTN's performance compares with the current state-of-the-art RADAM [27] and GTN [10] across various ConvNeXt backbones pretrained on ImageNet. The average and standard deviation measures are calculated across the provided splits, where applicable. Following RADAM and GTN, we have frozen our ConvNeXt backbones. CAPTN out-performed both RADAM [27] and GTN [10] in most of the benchmark datasets across different ConvNeXt backbone sizes. Both RADAM and GTN do not have a mechanism to represent orderless texture features or attributes. On the other hand, CAPTN represents orderless latent texture attributes as a learnable polynomial improving its expressiveness and consequently optimizing the mapping from latent to visual texture attributes. + +Qualitative. The illustrations in Fig. 3 were generated using the ConvNeXt-L backbone with images from the DTD dataset. The Texture Frequency Attention (TFA) module from CAPTN enhances spatial attention in regions containing specific texture features. To some extent, the attention mechanism can identify the spatial regularity of grid textures. For spiraled textures, which exhibit greater irregularity, CAPTN demonstrates robustness to variations in illumination, shadows, occlusion, and background textures by selectively focusing on the spiraled region. Similarly, for polka-dotted textures, CAPTN disregards occlusions and concentrates on the textured areas. Furthermore, in all three cases, CAPTN recognizes masked regions within the images and assigns them lower importance. + +![](images/cf5dd0f9a5962ca2a49fe2d0416fd808563331fbbdd16d9a9e5e1e2d8d406111.jpg) + +![](images/78357eb568dbc8edd52514c0e6f6220560f14afc539096005be5c2bea3f375fb.jpg) + +![](images/17385c32af57754918167cb618792d53f7fc1666d31273ef4740a722708fb5c6.jpg) + +![](images/a507883055b66b37a41b6c0d3be98b0f1b6ab9a7f7b009afeccd169e70229056.jpg) +Figure 3. Top row: Masked images from DTD dataset (grid, spiralled, polka dotted). Bottom row: Corresponding attention maps from Texture Frequency Attention (TFA). Red indicates high attention, blue indicates low attention. + +![](images/a67e76df2cf24ac9ffc717c5f99f7819613ecd321ff533703a89a8a2c8d59c45.jpg) + +![](images/0c86009d5e3e61436ceaa5fc8704ad87976cae425ddc1000d0b48cadec304ffd.jpg) + +# 6.4. Ablation Study + +We conducted ablation studies to evaluate the impact of TFA, $\mathrm{D}^2\mathrm{P}$ , and LCP in CAPTN. Overall, all components improve CAPTN's performance on both texture (DTD) and material (KTH, GTOS) datasets. The quantitative results are presented in Tab. 2. First, we assessed LCP by mapping orderless LTAs to visual texture attributes without using LCP. The deterioration of results demonstrated that LCP enhances expressiveness and injects non-linearities, improving the mapping of latent to visual texture attributes. Removing $\mathrm{D}^2\mathrm{P}$ led to a slight performance drop, as the LTA space became less robust and diverse. Finally, the removal of TFA further degraded CAPTN's performance, indicating that backbone features alone are less discriminative than LTAs. In addition, we performed an ablation study Tab. 3 to understand the effectiveness of LTA loss compared with Cross Entropy (CE) loss. The LTA loss function enhanced performance across texture and material datasets, underscoring its effectiveness for their recognition. + +# 6.5. Alternative Design Choices + +Non-polynomial baselines. Table 4 demonstrates that first-order mean pooling combined with a LCP transformation outperforms the non-polynomial baseline of second-order covariance pooling. Covariance pooling captures second-order statistics by modeling inter-LTA correlations computed over spatial activations. While this can encode useful dependencies between LTA dimensions, it is also sensitive to noise, small artifacts, and spatial misalignments due to its reliance on precise variation patterns across spatial locations. In contrast, mean pooling followed by an LCP transformation offers a more robust encoding alternative. Mean pooling summarizes each LTA by computing the first-order global average across spatial dimensions, producing compact and stable representations that are less affected by lo + +cal distortions. These pooled features serve as clean inputs to the LCP module, which applies expressive, non-linear transformations to each LTA independently using Chebyshev polynomials. This combination enhances the representational power of each LTA without the noise sensitivity inherent in second-order pooling. Together, mean pooling and LCP offer a robust and discriminative representation strategy that outperforms covariance pooling. + +Alternative Polynomial Bases. We replace the LCP module in CAPTN with a learnable RBF layer using Gaussian kernels. It computes scaled Euclidean distances to learnable centers and aggregates the activations via a learnable weight matrix. To avoid underfitting, we add a linear projection, LayerNorm, and GELU activation, along with kernel diversity for better generalization. As shown in Table 5, LCP outperforms the RBF variant across a range of texture and material datasets, while also requiring fewer learnable parameters. On the GTOS-M dataset, for example, CAPTN with a third-order LCP requires only 9.60M parameters, compared to 14.22M for its RBF-based counterpart. Although Gaussian RBFs are effective at modeling local variations, they often require many kernel centers and careful bandwidth tuning to capture diverse texture patterns, increasing computational cost. In contrast, LCPs form a global polynomial basis that efficiently captures long-range dependencies and oscillatory patterns with fewer parameters. In addition to RBFs, we also considered B-splines, which are widely used for smooth, piecewise function approximation. However, B-splines require careful knot placement and segmentwise assembly to represent multi-peaked or oscillatory signals, making them more sensitive to hyperparameter tuning and less scalable. LCPs, by comparison, provide a unified, smooth basis without discontinuities at segment boundaries, naturally capturing global texture variations while preserving sharp gradients and repetitive motifs. These properties make LCPs a more compact, scalable, and interpretable choice than both RBFs and B-splines for texture-specific applications. + +# 7. Conclusion + +We propose a comprehensive framework for robust texture representation by integrating SLTM with CAPTN. SLTM enhances generalization by mitigating over-reliance on fine-grained details, while the TFA mechanism captures multifrequency texture cues to support visual interpretation of spatial texture information during the generation of LTAs. Furthermore, the $\mathrm{D}^2\mathrm{P}$ module enriches LTAs by diversifying feature representations, and LCP is specifically designed to enable a richer and more expressive encoding of orderless LTAs. To further strengthen the learning process, we introduce a unified LTA loss function that balances classification objectives with orderless alignment and spatial consistency. + +# References + +[1] Vincent Andrearczyk and Paul F Whelan. Using filter banks in convolutional neural networks for texture classification. Pattern Recognition Letters, 84:63-69, 2016. 1 +[2] Avrim Blum and Tom Mitchell. Combining labeled and unlabeled data with co-training. In Proceedings of the eleventh annual conference on Computational learning theory, pages 92-100, 1998. 5 +[3] Barbara Caputo, Eric Hayman, and P. Mallikarjuna. Class-specific material categorisation. In Tenth IEEE International Conference on Computer Vision (ICCV'05) Volume 1, pages 1597-1604 Vol. 2, 2005. 7 +[4] Zhile Chen, Feng Li, Yuhui Quan, Yong Xu, and Hui Ji. Deep texture recognition via exploiting cross-layer statistical self-similarity. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5227-5236, 2021. 2 +[5] Zhile Chen, Yuhui Quan, Ruotao Xu, Lianwen Jin, and Yong Xu. Enhancing texture representation with deep tracing pattern encoding. Pattern Recognition, 146:109959, 2024. 2 +[6] Mircea Cimpoi, Subhransu Maji, Iasonas Kokkinos, Sammy Mohamed, and Andrea Vedaldi. Describing textures in the wild. In 2014 IEEE Conference on Computer Vision and Pattern Recognition, pages 3606-3613, 2014. 7 +[7] Mircea Cimpoi, Subhransu Maji, and Andrea Vedaldi. Deep filter banks for texture recognition and segmentation. In 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 3828-3836, 2015. 1, 2 +[8] Gabriella Csurka, Christopher R. Dance, Lixin Fan, Jutta Willamowski, and Cedric Bray. Visual categorization with bags of keypoints. In In Workshop on Statistical Learning in Computer Vision, ECCV, pages 1-22, 2004. 2 +[9] O.G. Cula and K.J. Dana. Compact representation of bidirectional texture functions. In Proceedings of the 2001 IEEE Computer Society Conference on Computer Vision and Pattern Recognition. CVPR 2001, pages I-I, 2001. 2 +[10] Ravishankar Evani, Deepu Rajan, and Shangbo Mao. Multiscale graph texture network. In Computer Vision - ECCV 2024, pages 218-235, Cham, 2025. Springer Nature Switzerland. 1, 2, 3, 5, 7 +[11] Shin Fujieda, Kohei Takayama, and Toshiya Hachisuka. Wavelet convolutional neural networks for texture classification. arXiv preprint arXiv:1707.07394, 2017. 1 +[12] Robert M. Haralick, K. Shanmugam, and Its'Hak Dinstein. Textural features for image classification. IEEE Transactions on Systems, Man, and Cybernetics, SMC-3(6):610-621, 1973. 1 +[13] Yuting Hu, Zhiling Long, and Ghassan AlRegib. Multi-level texture encoding and representation (multer) based on deep neural networks. In 2019 IEEE International Conference on Image Processing (ICIP), pages 4410-4414, 2019. 2 +[14] Béla Julesz. Visual pattern discrimination. IRE transactions on Information Theory, 8(2):84-92, 1962. 1 +[15] Hervé Jégou, Matthijs Douze, Cordelia Schmid, and Patrick Pérez. Aggregating local descriptors into a compact image representation. In 2010 IEEE Computer Society Conference + +on Computer Vision and Pattern Recognition, pages 3304-3311, 2010. 2 +[16] Kenneth I Laws. Texture energy measures. In Proceedings of Image Understanding Workshop, 1979. 1 +[17] Kenneth I. Laws. Rapid Texture Identification. In Image Processing for Missile Guidance, pages 376 - 381. International Society for Optics and Photonics, SPIE, 1980. 1 +[18] Thomas Leung and Jitendra Malik. Representing and recognizing the visual appearance of materials using three-dimensional textons. International journal of computer vision, 43:29-44, 2001. 2 +[19] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11966-11976, 2022. 3 +[20] Ilya Loshchilov and Frank Hutter. SGDR: stochastic gradient descent with warm restarts. In 5th International Conference on Learning Representations, ICLR 2017, Toulouse, France, 2017. 7 +[21] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In 7th International Conference on Learning Representations, ICLR 2019, New Orleans, LA, USA, 2019. 7 +[22] Shangbo Mao, Deepu Rajan, and Liang Tien Chia. Deep residual pooling network for texture recognition. Pattern Recognition, 112:107817, 2021. 2 +[23] Joshua Peeples, Weihuang Xu, and Alina Zare. Histogram layers for texture analysis. IEEE Transactions on Artificial Intelligence, 3(4):541-552, 2022. 2 +[24] Florent Perronnin and Christopher Dance. Fisher kernels on visual vocabularies for image categorization. In 2007 IEEE Conference on Computer Vision and Pattern Recognition, pages 1-8, 2007. 2 +[25] Michael J. Reale, Matthew Ochrym, Micah Church, Nicholas Goutermout, John Rubado, and Maria Cornacchia. Shape and texture aware graph processing. In 2020 IEEE Applied Imagery Pattern Recognition Workshop (AIPR), pages 1-18, 2020. 2 +[26] Theodore J Rivlin. Chebyshev polynomials. Courier Dover Publications, 2020. 5 +[27] Leonardo Scabini, Kallil M. Zielinski, Lucas C. Ribas, Wesley N. Gonçalves, Bernard De Baets, and Odemir M. Bruno. Radam: Texture recognition through randomized aggregated encoding of deep activation maps. Pattern Recognition, 143: 109802, 2023. 2, 7 +[28] Lavanya Sharan, Ruth Rosenholtz, and Edward Adelson. Material perception: What can you see in a brief glance? Journal of Vision, 9(8):784-784, 2009. 7 +[29] Kaiyou Song, Hua Yang, and Zhouping Yin. Multi-scale boosting feature encoding network for texture recognition. IEEE Transactions on Circuits and Systems for Video Technology, 31(11):4269-4282, 2021. 2 +[30] Yang Song, Fan Zhang, Qing Li, Heng Huang, Lauren J. O'Donnell, and Weidong Cai. Locally-transferred fisher vectors for texture classification. In 2017 IEEE International Conference on Computer Vision (ICCV), pages 4922-4930, 2017. 2 + +[31] Manik Varma and Andrew Zisserman. Classifying images of materials: Achieving viewpoint and illumination independence. In Computer Vision — ECCV 2002, pages 255–271, Berlin, Heidelberg, 2002. Springer Berlin Heidelberg. 2 +[32] Manik Varma and Andrew Zisserman. A statistical approach to texture classification from single images. International journal of computer vision, 62:61-81, 2005. 2 +[33] Yong Xu, Feng Li, Zhile Chen, Jinxiu Liang, and Yuhui Quan. Encoding spatial distribution of convolutional features for texture representation. In Advances in Neural Information Processing Systems, pages 22732-22744. Curran Associates, Inc., 2021. 2 +[34] Jia Xue, Hang Zhang, Kristin Dana, and Ko Nishino. Differential angular imaging for material recognition. In 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 6940-6949, 2017. 6 +[35] Jia Xue, Hang Zhang, and Kristin Dana. Deep texture manifold for ground terrain recognition. In 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 558-567, 2018. 1, 2, 7 +[36] Zhijing Yang, Shujian Lai, Xiaobin Hong, Yukai Shi, Yongqiang Cheng, and Chunmei Qing. Dfaen: Double-order knowledge fusion and attentional encoding network for texture recognition. Expert Systems with Applications, 209: 118223, 2022. 2 +[37] Wei Zhai, Yang Cao, Jing Zhang, and Zheng-Jun Zha. Deep multiple-attribute-perceived network for real-world texture recognition. In 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 3612–3621, 2019. 1, 2 +[38] Wei Zhai, Yang Cao, Zheng-Jun Zha, HaiYong Xie, and Feng Wu. Deep structure-revealed network for texture recognition. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11007-11016, 2020. 2 +[39] Wei Zhai, Yang Cao, Jing Zhang, Haiyong Xie, Dacheng Tao, and Zheng-Jun Zha. On exploring multiplicity of primitives and attributes for texture recognition in the wild. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(1):403-420, 2024. 1, 2 +[40] Hang Zhang, Jia Xue, and Kristin Dana. Deep ten: Texture encoding network. In 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2896-2905, 2017. 1, 2 \ No newline at end of file diff --git a/2025/Chebyshev Attention Depth Permutation Texture Network with Latent Texture Attribute Loss/images.zip b/2025/Chebyshev Attention Depth Permutation Texture Network with Latent Texture Attribute Loss/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..4bb5ed736ee6f631a2a1ac827eadb0e678627864 --- /dev/null +++ b/2025/Chebyshev Attention Depth Permutation Texture Network with Latent Texture Attribute Loss/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6d446a6f08b603a8fbfa61f863d2435bbc9eb13122e9277e818731fb2ca7147d +size 439722 diff --git a/2025/Chebyshev Attention Depth Permutation Texture Network with Latent Texture Attribute Loss/layout.json b/2025/Chebyshev Attention Depth Permutation Texture Network with Latent Texture Attribute Loss/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..be61bbbbc82a5026a9308efa5111c725e64a61e7 --- /dev/null +++ b/2025/Chebyshev Attention Depth Permutation Texture Network with Latent Texture Attribute Loss/layout.json @@ -0,0 +1,8886 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 125, + 103, + 486, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 125, + 103, + 486, + 138 + ], + "spans": [ + { + "bbox": [ + 125, + 103, + 486, + 138 + ], + "type": "text", + "content": "Chebyshev Attention Depth Permutation Texture Network with Latent Texture Attribute Loss" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 175, + 162, + 434, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 175, + 162, + 434, + 175 + ], + "spans": [ + { + "bbox": [ + 175, + 162, + 434, + 175 + ], + "type": "text", + "content": "Ravishankar Evani, Deepu Rajan, Shangbo Mao" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 146, + 177, + 463, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 177, + 463, + 190 + ], + "spans": [ + { + "bbox": [ + 146, + 177, + 463, + 190 + ], + "type": "text", + "content": "College of Computing and Data Science, Nanyang Technological University, Singapore" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 168, + 193, + 440, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 168, + 193, + 440, + 202 + ], + "spans": [ + { + "bbox": [ + 168, + 193, + 440, + 202 + ], + "type": "text", + "content": "S220007@e.ntu.edu.sg, ASDRajan@ntu.edu.sg, MAOS0003@e.ntu.edu.sg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "spans": [ + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 258, + 297, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 258, + 297, + 594 + ], + "spans": [ + { + "bbox": [ + 54, + 258, + 297, + 594 + ], + "type": "text", + "content": "Despite recent advances in deep texture recognition, existing methods still lack representational diversity and struggle to capture and preserve discriminative cues across stages of representation hierarchies. Moreover, many rely on loss formulations that prioritize recognition accuracy while overlooking spatial coherence and statistical consistency in the feature space. To address these issues, we propose three key innovations: (1) Stochastic Local Texture Masking (SLTM), a regularization strategy that randomly occludes small texture patches to promote the learning of broader spatial and contextual dependencies; (2) the Chebyshev Attention Depth Permutation Texture Network (CAPTN), a novel architecture that learns expressive and persistent Latent Texture Attribute (LTA) representations. CAPTN integrates a Texture Frequency Attention (TFA) module that generates Latent Texture Attributes (LTAs) and enables frequency-aware interpretability, a Dual Depth Permutation " + }, + { + "bbox": [ + 54, + 258, + 297, + 594 + ], + "type": "inline_equation", + "content": "(D^{2}P)" + }, + { + "bbox": [ + 54, + 258, + 297, + 594 + ], + "type": "text", + "content": " module to expose complementary channel adjacency patterns, and Learnable Chebyshev Polynomials (LCPs) to model high-order orderless LTA transformations via recursive Chebyshev basis expansion; and (3) a Latent Texture Attribute Loss that jointly optimizes classification accuracy, statistical alignment, and spatial fidelity. CAPTN supports end-to-end training without relying on fine-tuned CNN backbones and achieves state-of-the-art performance on several texture and material recognition benchmarks. (Code: https://github.com/RavishankarEvani/CAPTN)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 620, + 135, + 632 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 620, + 135, + 632 + ], + "spans": [ + { + "bbox": [ + 56, + 620, + 135, + 632 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 641, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 296, + 715 + ], + "type": "text", + "content": "Visual texture perception plays a critical role in understanding an object's surface characteristics. Distinct from object recognition, which often relies on shape cues, texture recognition emphasizes the repetitive patterns or stochastic microstructures that define surface appearance. Over the years, texture recognition has evolved appreciably, be" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 232, + 555, + 340 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 232, + 555, + 340 + ], + "spans": [ + { + "bbox": [ + 313, + 232, + 555, + 340 + ], + "type": "text", + "content": "ginning with foundational work such as Julesz's theory of textons [14], Haralick et al.'s co-occurrence matrices [12], and Laws' texture energy measures based on convolution masks [16, 17]. These early methods established the basis for statistical and structural texture analysis and were widely adopted in classical image processing pipelines. However, while effective under controlled conditions, such traditional techniques often fail to capture the rich diversity and structural complexity of real-world textures." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 341, + 556, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 341, + 556, + 556 + ], + "spans": [ + { + "bbox": [ + 313, + 341, + 556, + 556 + ], + "type": "text", + "content": "With the rise of data-driven approaches, modern deep learning methods have largely supplanted handcrafted descriptors by offering more expressive and hierarchical feature representations. Despite this progress, deep models still encounter challenges when applied to texture recognition. In particular, successive convolution and pooling operations can suppress or discard fine-grained texture cues, limiting the network's ability to retain localized and discriminative information. Furthermore, textures often exhibit overlapping or perceptually similar traits [37], which contribute to inter-class confusion by blurring the boundaries between distinct texture categories. For instance, textures labeled as \"stratified\" might closely resemble those labeled \"cracked\" or \"zigzagged,\" making it difficult to consistently assign the correct class. In addition, intra-class variability arises from variations within the same category due to differences in lighting, scale, deformation, or inherent surface irregularities." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 558, + 557, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 557, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 557, + 715 + ], + "type": "text", + "content": "The concept of texture attributes [10, 37, 39] has been shown to play a crucial role in representing perceptually similar textures. Latent texture attributes [10] refer to the underlying semantic components of visual texture attributes, where each latent texture attribute may correspond to one or more perceptual texture attributes. Effectively manipulating and transforming these latent attributes can enhance texture discrimination. Several studies employing Convolutional Neural Network (CNN)-based approaches have explored different aggregation methods, including global average pooling [1, 11, 37], Fisher vectors [7], feature encoding [35, 40], and cross-layer aggregation. However, certain aggregation and dimensionality" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23423" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "type": "text", + "content": "reduction techniques lead to the loss of important texture cues, and little attention has been given to recovering discriminative information post-aggregation." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 109, + 294, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 109, + 294, + 239 + ], + "spans": [ + { + "bbox": [ + 54, + 109, + 294, + 239 + ], + "type": "text", + "content": "To address these limitations, we introduce a comprehensive texture recognition framework that advances both architectural design and training methodology. Central to our approach is the Chebyshev Attention Depth Permutation Texture Network (CAPTN), which extracts expressive and interpretable LTAs, depth permutation, and polynomial projection. In addition, we propose a stochastic local texture masking (SLTM) strategy and a tailored multi-objective loss function that together encourage spatially grounded and orderless texture representations. Our main contributions are as follows:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 240, + 295, + 504 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 55, + 240, + 295, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 240, + 295, + 264 + ], + "spans": [ + { + "bbox": [ + 55, + 240, + 295, + 264 + ], + "type": "text", + "content": "- We propose the CAPTN architecture, which integrates three key components:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 63, + 265, + 295, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 265, + 295, + 312 + ], + "spans": [ + { + "bbox": [ + 63, + 265, + 295, + 312 + ], + "type": "text", + "content": "- A Texture Frequency Attention (TFA) module that generates Latent Texture Attributes (LTAs) and enables visual interpretability by highlighting discriminative texture regions across frequency scales." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 63, + 313, + 295, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 313, + 295, + 360 + ], + "spans": [ + { + "bbox": [ + 63, + 313, + 295, + 360 + ], + "type": "text", + "content": "- A Dual Depth Permutation " + }, + { + "bbox": [ + 63, + 313, + 295, + 360 + ], + "type": "inline_equation", + "content": "(\\mathrm{D}^2\\mathrm{P})" + }, + { + "bbox": [ + 63, + 313, + 295, + 360 + ], + "type": "text", + "content": " module that applies two distinct channel shuffling operations to expose complementary spatial relationships and enhance feature diversity." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 63, + 361, + 295, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 361, + 295, + 395 + ], + "spans": [ + { + "bbox": [ + 63, + 361, + 295, + 395 + ], + "type": "text", + "content": "- A Learnable Chebyshev Polynomial (LCP) expansion that enhances the representational power of each orderless LTA." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 396, + 295, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 396, + 295, + 467 + ], + "spans": [ + { + "bbox": [ + 55, + 396, + 295, + 467 + ], + "type": "text", + "content": "- We introduce a Stochastic Local Texture Masking (SLTM) strategy that randomly occludes small patches of the input image during training. This compels the model to leverage broader spatial context and enables a new supervision signal that aligns masked patches with regional latent features." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 468, + 295, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 468, + 295, + 504 + ], + "spans": [ + { + "bbox": [ + 55, + 468, + 295, + 504 + ], + "type": "text", + "content": "- We formulate a multi-objective LTA loss that jointly supervises classification accuracy, orderless latent attribute consistency, and spatial latent attribute alignment." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 514, + 142, + 526 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 514, + 142, + 526 + ], + "spans": [ + { + "bbox": [ + 55, + 514, + 142, + 526 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "text", + "content": "Traditional Texture Recognition. Building upon the foundational work discussed in Sec. 1, influential methods such as the Bag of Textons (BoT) [9, 18] and Bag of Words (BoW) [8] introduced a paradigm for representing textures as orderless histograms over learned dictionaries of textural elements. These methods encode local patches, aggregate them into global descriptors, and classify the results using algorithms such as Support Vector Machines (SVM). Varma et al. [31, 32] advanced this line of work by learning texture descriptors directly from data. Subsequent developments led to residual encoding techniques—Jégou et al. [15] proposed VLAD, which captures first-order residuals relative to visual codewords, while Perronnin et al. [24] introduced Fisher Vectors (FV), a probabilistic extension that also models second-order statistics." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 72, + 553, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 215 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 215 + ], + "type": "text", + "content": "Deep Texture Recognition. Deep learning-based texture recognition methods such as FV-CNN [7] and LFV [30] primarily focus on encoding features extracted from pretrained CNNs, but they do not support end-to-end training. To address this limitation, Zhang et al. [40] proposed DeepTEN, integrating residual encoding and dictionary learning into a single layer. The layer functions analogously to VLAD and FV while enabling end-to-end training with a CNN backbone. RPNet [22] simplifies Deep-TEN by replacing dictionary learning with residual pooling. DEP [35] enhances Deep-TEN further by combining orderless texture features with spatial cues to balance structure and texture." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 217, + 553, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 217, + 553, + 373 + ], + "spans": [ + { + "bbox": [ + 313, + 217, + 553, + 373 + ], + "type": "text", + "content": "MuLTER [13] improves DEP by integrating multi-layer CNN features, preserving fine texture and spatial details. MSBFEN [29] combines multiscale encoding and boosting to better capture discriminative texture features. Chen et al. [4] proposed using cross-layer statistical self-similarity to capture consistent patterns across CNN layers. Later, they introduced DTPNet [5], a local encoding scheme inspired by Local Binary Patterns (LBP), robust to rotation and grayscale changes. Yang et al. [36] proposed DFAEN, which uses frequency attention and first- and second-order statistics to enhance texture details. RADAM [27] fuses CNN activation maps via randomized aggregation, promoting texture diversity and reducing overfitting." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 375, + 553, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 375, + 553, + 553 + ], + "spans": [ + { + "bbox": [ + 313, + 375, + 553, + 553 + ], + "type": "text", + "content": "FENet [33] encodes spatial regularity via fractal analysis, while DSR-Net [38] models local spatial dependencies instead of orderless aggregation. Zhai et al. [37] earlier proposed MAP-Net, which learns texture-related visual attributes and uses spatially-adaptive pooling for feature fusion. Their recent work, MPAP [39], uses a multi-branch architecture to model bottom-up structures and top-down attribute relationships by learning local-to-global mappings. Reale et al. [25] proposed modeling triangular patches as graph nodes to encode texture patterns. More recently, GTN [10] introduced a graph-based deep approach designed to model directional relationships between Latent Texture Attributes (LTAs). GTN enhances LTAs through message passing and aggregation of semantically related information." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 555, + 553, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 555, + 553, + 616 + ], + "spans": [ + { + "bbox": [ + 313, + 555, + 553, + 616 + ], + "type": "text", + "content": "Deep Polynomial Expansion. HistNet [23] features a trainable histogram layer that preserves spatial context using learnable bin centers and widths, combined with Radial Basis Functions (RBFs) for robustness to intra-class variability." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 617, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 553, + 713 + ], + "type": "text", + "content": "While graph-based methods explicitly learn attribute adjacencies for message passing to refine LTAs, our CAPTN framework adopts a more flexible and implicit approach. It enables rich and diverse feature interactions without the need for explicit graph construction. This not only simplifies the overall architecture but also implicitly regularizes the learning process by breaking fixed feature layouts, thereby enhancing texture robustness and generaliza" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23424" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 88, + 558, + 281 + ], + "blocks": [ + { + "bbox": [ + 58, + 88, + 558, + 281 + ], + "lines": [ + { + "bbox": [ + 58, + 88, + 558, + 281 + ], + "spans": [ + { + "bbox": [ + 58, + 88, + 558, + 281 + ], + "type": "image", + "image_path": "30a2d67006fcfed2401ca74cc7386033f0c042feb7d6106c78018db81ed061f7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 292, + 555, + 350 + ], + "lines": [ + { + "bbox": [ + 54, + 292, + 555, + 350 + ], + "spans": [ + { + "bbox": [ + 54, + 292, + 555, + 350 + ], + "type": "text", + "content": "Figure 1. Stochastic Local Texture Masking (SLTM) & components of Chebyshev Attention Depth Permutation Texture Network (CAPTN): SLTM facilitates contextual texture learning, Texture Frequency Attention (TFA) captures multi-frequency texture cues, Dual Depth Permutation " + }, + { + "bbox": [ + 54, + 292, + 555, + 350 + ], + "type": "inline_equation", + "content": "(\\mathrm{D}^2\\mathrm{P})" + }, + { + "bbox": [ + 54, + 292, + 555, + 350 + ], + "type": "text", + "content": " promotes feature diversification, and Learnable Chebyshev Polynomials (LCP) map latent texture attributes to visual texture attributes. The training process optimizes Latent Texture Attribute (LTA) Loss, which balances classification accuracy with orderless and spatial texture consistency. || represents concatenation, " + }, + { + "bbox": [ + 54, + 292, + 555, + 350 + ], + "type": "inline_equation", + "content": "\\times" + }, + { + "bbox": [ + 54, + 292, + 555, + 350 + ], + "type": "text", + "content": " represents matrix multiplied by a scalar." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 369, + 298, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 369, + 298, + 453 + ], + "spans": [ + { + "bbox": [ + 54, + 369, + 298, + 453 + ], + "type": "text", + "content": "tion. Unlike radial basis functions (RBFs), which often require numerous kernel centers and careful bandwidth tuning, we propose a Chebyshev polynomial-based method that provides a compact yet globally expressive basis. This approach eliminates the need for kernel-specific design choices while efficiently capturing complex and non-linear texture variations." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 465, + 246, + 480 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 465, + 246, + 480 + ], + "spans": [ + { + "bbox": [ + 55, + 465, + 246, + 480 + ], + "type": "text", + "content": "3. Stochastic Local Texture Masking" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 486, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 486, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 486, + 296, + 713 + ], + "type": "text", + "content": "To enhance the model's ability to capture local, global, and contextual texture representations, we introduce Stochastic Local Texture Masking (SLTM). This technique randomly masks a small patch of the input texture image prior to passing it through the pre-trained backbone, thereby preventing the Chebyshev Attention Depth Permutation Texture Network (CAPTN) from over-relying on fine-grained local details. Given an input image " + }, + { + "bbox": [ + 56, + 486, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{I} \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 56, + 486, + 296, + 713 + ], + "type": "text", + "content": ", SLTM generates: (1) a masked image " + }, + { + "bbox": [ + 56, + 486, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_r" + }, + { + "bbox": [ + 56, + 486, + 296, + 713 + ], + "type": "text", + "content": ", where a randomly selected patch of size " + }, + { + "bbox": [ + 56, + 486, + 296, + 713 + ], + "type": "inline_equation", + "content": "P \\times P" + }, + { + "bbox": [ + 56, + 486, + 296, + 713 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 56, + 486, + 296, + 713 + ], + "type": "inline_equation", + "content": "P \\ll H, P \\ll W" + }, + { + "bbox": [ + 56, + 486, + 296, + 713 + ], + "type": "text", + "content": ") is zeroed out, and (2) the extracted patch " + }, + { + "bbox": [ + 56, + 486, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_t \\in \\mathbb{R}^{P \\times P \\times 3}" + }, + { + "bbox": [ + 56, + 486, + 296, + 713 + ], + "type": "text", + "content": ", which retains the original pixel values prior to masking. By selectively removing texture information, SLTM encourages CAPTN to exploit broader spatial dependencies, thereby enhancing global feature learning and reducing overfitting. Finally, the masked image " + }, + { + "bbox": [ + 56, + 486, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_r" + }, + { + "bbox": [ + 56, + 486, + 296, + 713 + ], + "type": "text", + "content": " is fed through the backbone and CAPTN for further processing. This module is activated during the computation of the Spatial Latent Texture Attribute Loss " + }, + { + "bbox": [ + 56, + 486, + 296, + 713 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{SL})" + }, + { + "bbox": [ + 56, + 486, + 296, + 713 + ], + "type": "text", + "content": ", as described in Section 5." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 367, + 555, + 394 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 367, + 555, + 394 + ], + "spans": [ + { + "bbox": [ + 313, + 367, + 555, + 394 + ], + "type": "text", + "content": "4. Chebyshev Attention Depth Permutation Texture Network" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 402, + 556, + 546 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 402, + 556, + 546 + ], + "spans": [ + { + "bbox": [ + 313, + 402, + 556, + 546 + ], + "type": "text", + "content": "The Chebyshev Attention Depth Permutation Texture Network (CAPTN) consists of three key modules: (1) Texture Frequency Attention (TFA) Network, (2) Dual Depth Permutation " + }, + { + "bbox": [ + 313, + 402, + 556, + 546 + ], + "type": "inline_equation", + "content": "(\\mathrm{D}^2\\mathrm{P})" + }, + { + "bbox": [ + 313, + 402, + 556, + 546 + ], + "type": "text", + "content": " Network, and (3) Learnable Chebyshev Polynomials (LCP). Following [10], CAPTN extracts latent texture attributes (LTAs) by computing spatial texture frequency information from feature maps " + }, + { + "bbox": [ + 313, + 402, + 556, + 546 + ], + "type": "inline_equation", + "content": "\\mathcal{F} = \\{\\mathbf{X}^{o_m}\\mid 1\\leq m\\leq M - 1\\}" + }, + { + "bbox": [ + 313, + 402, + 556, + 546 + ], + "type": "text", + "content": " obtained from an " + }, + { + "bbox": [ + 313, + 402, + 556, + 546 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 402, + 556, + 546 + ], + "type": "text", + "content": " -layer pre-trained CNN backbone, excluding the patchify layer [19]. Depth permutation enhances LTA representation by diversifying feature space and reducing inter-class variability, while LCP improves expressiveness of orderless LTAs." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 552, + 473, + 565 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 552, + 473, + 565 + ], + "spans": [ + { + "bbox": [ + 313, + 552, + 473, + 565 + ], + "type": "text", + "content": "4.1. Texture Frequency Attention" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "text", + "content": "Textures display a wide range of spatial frequency characteristics, from high-frequency (fine-grained) patterns to low-frequency (smooth or homogeneous) regions. To effectively capture and interpret these variations, we propose a Texture Frequency Attention (TFA) mechanism. In addition to emphasizing discriminative texture regions, TFA also transforms backbone feature maps into Latent Texture Attributes (LTAs). Let " + }, + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{X}^{o_m} \\in \\mathbb{R}^{H_m \\times W_m \\times C_m}" + }, + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "text", + "content": " be the " + }, + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "inline_equation", + "content": "m^{\\text{th}}" + }, + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "text", + "content": " layer feature map. We compute transformed features as " + }, + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{X}^{t_m} = \\mathbf{W}_l^m * \\mathbf{X}^{o_m} + \\mathbf{b}_l^m" + }, + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_l^m \\in \\mathbb{R}^{C_m \\times C_m \\times 1 \\times 1}" + }, + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "text", + "content": " is the convolution kernel, * denotes the convolution operation, and " + }, + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{b}_l^m \\in \\mathbb{R}^{C_m}" + }, + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "text", + "content": " is a learnable bias vector. Next, we" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23425" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 76, + 121, + 183 + ], + "blocks": [ + { + "bbox": [ + 58, + 76, + 121, + 183 + ], + "lines": [ + { + "bbox": [ + 58, + 76, + 121, + 183 + ], + "spans": [ + { + "bbox": [ + 58, + 76, + 121, + 183 + ], + "type": "image", + "image_path": "a67d04c227cb95c0ffaec0618099b40c232fc0bdd64550c26fd1f5280d1e9ebc.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 82, + 185, + 91, + 192 + ], + "lines": [ + { + "bbox": [ + 82, + 185, + 91, + 192 + ], + "spans": [ + { + "bbox": [ + 82, + 185, + 91, + 192 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 121, + 76, + 355, + 182 + ], + "blocks": [ + { + "bbox": [ + 121, + 76, + 355, + 182 + ], + "lines": [ + { + "bbox": [ + 121, + 76, + 355, + 182 + ], + "spans": [ + { + "bbox": [ + 121, + 76, + 355, + 182 + ], + "type": "image", + "image_path": "eeb164c67c23b70873281feae9075d050d251132db33e1e31961d72bbcb4857f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 234, + 184, + 243, + 193 + ], + "lines": [ + { + "bbox": [ + 234, + 184, + 243, + 193 + ], + "spans": [ + { + "bbox": [ + 234, + 184, + 243, + 193 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 356, + 76, + 550, + 182 + ], + "blocks": [ + { + "bbox": [ + 356, + 76, + 550, + 182 + ], + "lines": [ + { + "bbox": [ + 356, + 76, + 550, + 182 + ], + "spans": [ + { + "bbox": [ + 356, + 76, + 550, + 182 + ], + "type": "image", + "image_path": "1ec616705a901841a4cbd66393e74b3dbad94522a53d96397e7e6030e4109e6a.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 451, + 184, + 460, + 192 + ], + "lines": [ + { + "bbox": [ + 451, + 184, + 460, + 192 + ], + "spans": [ + { + "bbox": [ + 451, + 184, + 460, + 192 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 58, + 194, + 336, + 300 + ], + "blocks": [ + { + "bbox": [ + 58, + 194, + 336, + 300 + ], + "lines": [ + { + "bbox": [ + 58, + 194, + 336, + 300 + ], + "spans": [ + { + "bbox": [ + 58, + 194, + 336, + 300 + ], + "type": "image", + "image_path": "85e080db0b7ffa8cfb51b76bdf45958f3a8d24a07db96c9a889df5e6f37366af.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 190, + 302, + 200, + 311 + ], + "lines": [ + { + "bbox": [ + 190, + 302, + 200, + 311 + ], + "spans": [ + { + "bbox": [ + 190, + 302, + 200, + 311 + ], + "type": "text", + "content": "(d)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 337, + 194, + 552, + 300 + ], + "blocks": [ + { + "bbox": [ + 337, + 194, + 552, + 300 + ], + "lines": [ + { + "bbox": [ + 337, + 194, + 552, + 300 + ], + "spans": [ + { + "bbox": [ + 337, + 194, + 552, + 300 + ], + "type": "image", + "image_path": "8065d32f9b6ac3509639a9700d48f47eee3e6c862290ffe33c4f252e3d8bafb9.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 440, + 303, + 449, + 311 + ], + "lines": [ + { + "bbox": [ + 440, + 303, + 449, + 311 + ], + "spans": [ + { + "bbox": [ + 440, + 303, + 449, + 311 + ], + "type": "text", + "content": "(e)" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 319, + 555, + 376 + ], + "lines": [ + { + "bbox": [ + 55, + 319, + 555, + 376 + ], + "spans": [ + { + "bbox": [ + 55, + 319, + 555, + 376 + ], + "type": "text", + "content": "Figure 2. (a) SLTM: extract a small patch from an input texture image, (b) " + }, + { + "bbox": [ + 55, + 319, + 555, + 376 + ], + "type": "inline_equation", + "content": "\\mathrm{D}^2\\mathrm{P}" + }, + { + "bbox": [ + 55, + 319, + 555, + 376 + ], + "type": "text", + "content": ": feature maps from backbone undergo depth permutation to diversify feature space, (c) LCP: enhanced Latent Texture Attributes undergo orderless aggregation and fit a learnable Chebyshev polynomial function, (d) TFA: use spatial texture frequencies to compute spatial attention, (e) SLAR: representation of spatial information for latent attributes and backbone features. SLTM and SLAR are used to compute Latent Texture Attribute Loss. || represents concatenation, " + }, + { + "bbox": [ + 55, + 319, + 555, + 376 + ], + "type": "inline_equation", + "content": "\\times" + }, + { + "bbox": [ + 55, + 319, + 555, + 376 + ], + "type": "text", + "content": " represents matrix multiplied by a scalar, " + }, + { + "bbox": [ + 55, + 319, + 555, + 376 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 55, + 319, + 555, + 376 + ], + "type": "text", + "content": " represents matrix multiplication." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 396, + 295, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 396, + 295, + 419 + ], + "spans": [ + { + "bbox": [ + 55, + 396, + 295, + 419 + ], + "type": "text", + "content": "compute per-location channel-wise maximum and average statistics" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 99, + 446, + 295, + 465 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 446, + 295, + 465 + ], + "spans": [ + { + "bbox": [ + 99, + 446, + 295, + 465 + ], + "type": "interline_equation", + "content": "X _ {\\max } ^ {t _ {m}} (i, j) = \\max _ {0 \\leq c < C _ {m}} X ^ {t _ {m}} (i, j, c), \\tag {1}", + "image_path": "9fabfc56dd1865ca116ceade94b7d81a3b61db7e5443eb31e13f3fd0bddf4dcd.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 104, + 468, + 294, + 501 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 468, + 294, + 501 + ], + "spans": [ + { + "bbox": [ + 104, + 468, + 294, + 501 + ], + "type": "interline_equation", + "content": "X _ {\\text {a v g}} ^ {t _ {m}} (i, j) = \\frac {1}{C _ {m}} \\sum_ {c = 0} ^ {C _ {m} - 1} X ^ {t _ {m}} (i, j, c). \\tag {2}", + "image_path": "e50b2b66502e922c34e3e156534377a90e3da4929c972327f6d5601e6a8668e2.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 514, + 296, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 514, + 296, + 612 + ], + "spans": [ + { + "bbox": [ + 55, + 514, + 296, + 612 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 55, + 514, + 296, + 612 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_{\\mathrm{max}}^{t_m}, \\mathbf{X}_{\\mathrm{avg}}^{t_m} \\in \\mathbb{R}^{H_m \\times W_m \\times 1}" + }, + { + "bbox": [ + 55, + 514, + 296, + 612 + ], + "type": "text", + "content": " represent spatial maps of local maxima and average activations across channels. High-frequency textures tend to exhibit greater local variation, resulting in larger differences between maximum and average values, whereas low-frequency regions yield more consistent responses. To capture this contrast, we concatenate " + }, + { + "bbox": [ + 55, + 514, + 296, + 612 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_{\\mathrm{max}}^{t_m}" + }, + { + "bbox": [ + 55, + 514, + 296, + 612 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 514, + 296, + 612 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_{\\mathrm{avg}}^{t_m}" + }, + { + "bbox": [ + 55, + 514, + 296, + 612 + ], + "type": "text", + "content": " along the channel dimension and apply a convolution followed by a sigmoid activation" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 627, + 295, + 654 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 627, + 295, + 654 + ], + "spans": [ + { + "bbox": [ + 105, + 627, + 295, + 654 + ], + "type": "interline_equation", + "content": "\\mathbf {A} ^ {m} = \\frac {1}{1 + e ^ {- \\left[ \\mathbf {W} _ {a} ^ {m} * \\left(\\mathbf {X} _ {\\operatorname* {m a x}} ^ {t _ {m}} \\parallel \\mathbf {X} _ {\\operatorname* {a v g}} ^ {t _ {m}}\\right) \\right]}}, \\tag {3}", + "image_path": "bd538d32d9f2a67aabe6f6e88ca489001a8becc66681968bfb1d2f6dffd9cef3.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\parallel" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": " denotes channel-wise concatenation and " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_a^m\\in" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{1\\times 2\\times 1\\times 1}" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": " is the convolution kernel. The resulting spatial attention map " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{A}^m\\in \\mathbb{R}^{H_m\\times W_m\\times 1}" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": " is used to modulate " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{X}^{t_m}" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": " via element-wise multiplication" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 330, + 413, + 553, + 452 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 330, + 413, + 553, + 452 + ], + "spans": [ + { + "bbox": [ + 330, + 413, + 553, + 452 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} X _ {\\text {a t t n}} ^ {t _ {m}} (i, j, c) = X ^ {t _ {m}} (i, j, c) \\times A ^ {m} (i, j), \\\\ \\forall (i, j, c) \\in [ 0, H _ {m} - 1 ] \\times [ 0, W _ {m} - 1 ] \\times [ 0, C _ {m} - 1 ], \\tag {4} \\\\ \\end{array}", + "image_path": "fa32a7e210bf44ee00e2435c5312c81dee65baa30bb4610d5f2775bdd6b8dbc0.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 453, + 554, + 500 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 453, + 554, + 500 + ], + "spans": [ + { + "bbox": [ + 313, + 453, + 554, + 500 + ], + "type": "text", + "content": "A residual connection is then added, combining the attention-weighted features with the original " + }, + { + "bbox": [ + 313, + 453, + 554, + 500 + ], + "type": "inline_equation", + "content": "\\mathbf{X}^{t_m}" + }, + { + "bbox": [ + 313, + 453, + 554, + 500 + ], + "type": "text", + "content": ". The result is passed through a transformation " + }, + { + "bbox": [ + 313, + 453, + 554, + 500 + ], + "type": "inline_equation", + "content": "g(\\cdot)" + }, + { + "bbox": [ + 313, + 453, + 554, + 500 + ], + "type": "text", + "content": " comprising Batch Normalization, GELU activation, and dropout" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 359, + 510, + 553, + 525 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 510, + 553, + 525 + ], + "spans": [ + { + "bbox": [ + 359, + 510, + 553, + 525 + ], + "type": "interline_equation", + "content": "\\mathbf {X} _ {a} ^ {t _ {m}} = f _ {a} \\left(\\gamma_ {m}\\right) \\times g \\left(\\mathbf {X} _ {\\text {a t t n}} ^ {t _ {m}} + \\mathbf {X} ^ {t _ {m}}\\right), \\tag {5}", + "image_path": "792c7d213261e9bef76753d317ef9075f43965c06f57e0c824032ff3c30394f9.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 529, + 554, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 529, + 554, + 553 + ], + "spans": [ + { + "bbox": [ + 313, + 529, + 554, + 553 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 529, + 554, + 553 + ], + "type": "inline_equation", + "content": "\\gamma_{m}" + }, + { + "bbox": [ + 313, + 529, + 554, + 553 + ], + "type": "text", + "content": " is a learnable scalar, and " + }, + { + "bbox": [ + 313, + 529, + 554, + 553 + ], + "type": "inline_equation", + "content": "f_{a}(\\cdot)" + }, + { + "bbox": [ + 313, + 529, + 554, + 553 + ], + "type": "text", + "content": " ensures the scaling factor remains within the range [0, 1]" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 356, + 564, + 553, + 578 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 356, + 564, + 553, + 578 + ], + "spans": [ + { + "bbox": [ + 356, + 564, + 553, + 578 + ], + "type": "interline_equation", + "content": "f _ {a} (z) = \\min (\\max (0, z), 1), \\quad z \\in \\mathbb {R}. \\tag {6}", + "image_path": "47344fa05467995700d431d760bb6d88a73b27d43f57968284e604c064e47333.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 582, + 554, + 630 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 554, + 630 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 554, + 630 + ], + "type": "text", + "content": "While the convolution and " + }, + { + "bbox": [ + 313, + 582, + 554, + 630 + ], + "type": "inline_equation", + "content": "g(\\cdot)" + }, + { + "bbox": [ + 313, + 582, + 554, + 630 + ], + "type": "text", + "content": " transformation generate the LTAs, the spatial attention map " + }, + { + "bbox": [ + 313, + 582, + 554, + 630 + ], + "type": "inline_equation", + "content": "\\mathbf{A}^m" + }, + { + "bbox": [ + 313, + 582, + 554, + 630 + ], + "type": "text", + "content": " provides interpretable insights into the model's dynamic focus on spatially discriminative texture regions." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 635, + 453, + 649 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 635, + 453, + 649 + ], + "spans": [ + { + "bbox": [ + 313, + 635, + 453, + 649 + ], + "type": "text", + "content": "4.2. Dual Depth Permutation" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 653, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 554, + 714 + ], + "type": "text", + "content": "Dual Depth Permutation " + }, + { + "bbox": [ + 313, + 653, + 554, + 714 + ], + "type": "inline_equation", + "content": "(\\mathrm{D}^2\\mathrm{P})" + }, + { + "bbox": [ + 313, + 653, + 554, + 714 + ], + "type": "text", + "content": " operates directly on backbone features to generate complementary representations through two distinct shuffling operations. Exposing varied adjacency patterns promotes rich cross-channel correlations. The outputs from both shuffling paths are later" + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23426" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 205 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 205 + ], + "type": "text", + "content": "fused with LTAs from Sec. 4.1 to enhance representation diversity and improve texture robustness. Additionally, " + }, + { + "bbox": [ + 55, + 72, + 296, + 205 + ], + "type": "inline_equation", + "content": "\\mathrm{D}^2\\mathrm{P}" + }, + { + "bbox": [ + 55, + 72, + 296, + 205 + ], + "type": "text", + "content": " serves as an implicit regularizer, reducing overfitting by disrupting fixed feature layouts. Drawing inspiration from constraining [2], " + }, + { + "bbox": [ + 55, + 72, + 296, + 205 + ], + "type": "inline_equation", + "content": "\\mathrm{D}^2\\mathrm{P}" + }, + { + "bbox": [ + 55, + 72, + 296, + 205 + ], + "type": "text", + "content": " introduces multiple feature perspectives by permuting channel adjacency structures in a supervised setting. Let " + }, + { + "bbox": [ + 55, + 72, + 296, + 205 + ], + "type": "inline_equation", + "content": "\\mathbf{X}^{o_m} \\in \\mathbb{R}^{H_m \\times W_m \\times C_m}" + }, + { + "bbox": [ + 55, + 72, + 296, + 205 + ], + "type": "text", + "content": " denote the feature map output from the " + }, + { + "bbox": [ + 55, + 72, + 296, + 205 + ], + "type": "inline_equation", + "content": "m^{\\text{th}}" + }, + { + "bbox": [ + 55, + 72, + 296, + 205 + ], + "type": "text", + "content": " layer of the backbone. We apply two distinct channel shuffles using group sizes " + }, + { + "bbox": [ + 55, + 72, + 296, + 205 + ], + "type": "inline_equation", + "content": "s_0" + }, + { + "bbox": [ + 55, + 72, + 296, + 205 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 72, + 296, + 205 + ], + "type": "inline_equation", + "content": "s_1" + }, + { + "bbox": [ + 55, + 72, + 296, + 205 + ], + "type": "text", + "content": ", ensuring " + }, + { + "bbox": [ + 55, + 72, + 296, + 205 + ], + "type": "inline_equation", + "content": "\\frac{C_m}{s_0}, \\frac{C_m}{s_1} \\in \\mathbb{Z}" + }, + { + "bbox": [ + 55, + 72, + 296, + 205 + ], + "type": "text", + "content": ". Let " + }, + { + "bbox": [ + 55, + 72, + 296, + 205 + ], + "type": "inline_equation", + "content": "p \\in \\{0,1\\}" + }, + { + "bbox": [ + 55, + 72, + 296, + 205 + ], + "type": "text", + "content": " index the two shuffling paths, with corresponding group sizes " + }, + { + "bbox": [ + 55, + 72, + 296, + 205 + ], + "type": "inline_equation", + "content": "k_p^m \\in \\{s_0, s_1\\}" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 99, + 217, + 295, + 235 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 217, + 295, + 235 + ], + "spans": [ + { + "bbox": [ + 99, + 217, + 295, + 235 + ], + "type": "interline_equation", + "content": "\\mathbf {X} ^ {o _ {m}} \\rightarrow \\mathbf {X} _ {\\mathrm {r}, p} ^ {o _ {m}} \\in \\mathbb {R} ^ {H _ {m} \\times W _ {m} \\times k _ {p} ^ {m} \\times g _ {k _ {p} ^ {m}}}, \\tag {7}", + "image_path": "3176813a3382c3fbde48ddf4d60b1aa89d21c4b65cfc2d5ca4fdb75ec43d3abb.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 99, + 251, + 295, + 271 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 251, + 295, + 271 + ], + "spans": [ + { + "bbox": [ + 99, + 251, + 295, + 271 + ], + "type": "interline_equation", + "content": "\\mathbf {X} _ {\\mathrm {r}, p} ^ {o _ {m}} \\rightarrow \\mathbf {X} _ {\\mathrm {r}, p} ^ {o _ {m} ^ {T}} \\in \\mathbb {R} ^ {H _ {m} \\times W _ {m} \\times g _ {k _ {p} ^ {m}} \\times k _ {p} ^ {m}}, \\tag {8}", + "image_path": "1dc66e619f5fd8e3beb46920a8bb8309b9b837ca11ac3e0436ede249123ebf60.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 279, + 296, + 310 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 279, + 296, + 310 + ], + "spans": [ + { + "bbox": [ + 55, + 279, + 296, + 310 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 279, + 296, + 310 + ], + "type": "inline_equation", + "content": "g_{k_p^m}" + }, + { + "bbox": [ + 55, + 279, + 296, + 310 + ], + "type": "text", + "content": " is the number of groups formed using group size " + }, + { + "bbox": [ + 55, + 279, + 296, + 310 + ], + "type": "inline_equation", + "content": "k_{p}^{m}" + }, + { + "bbox": [ + 55, + 279, + 296, + 310 + ], + "type": "text", + "content": " at layer " + }, + { + "bbox": [ + 55, + 279, + 296, + 310 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 55, + 279, + 296, + 310 + ], + "type": "text", + "content": ". After flattening " + }, + { + "bbox": [ + 55, + 279, + 296, + 310 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_{\\mathrm{r},p}^{o_m^T}" + }, + { + "bbox": [ + 55, + 279, + 296, + 310 + ], + "type": "text", + "content": ", we obtain" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 126, + 322, + 295, + 338 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 126, + 322, + 295, + 338 + ], + "spans": [ + { + "bbox": [ + 126, + 322, + 295, + 338 + ], + "type": "interline_equation", + "content": "\\mathbf {X} _ {\\mathrm {s}, p} ^ {o _ {m}} \\in \\mathbb {R} ^ {H _ {m} \\times W _ {m} \\times C _ {m}}. \\tag {9}", + "image_path": "952c8242428ba50333f7f0f8ab762cfaacaf5c279cf82d436cd15f1e6d185321.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 346, + 296, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 346, + 296, + 441 + ], + "spans": [ + { + "bbox": [ + 55, + 346, + 296, + 441 + ], + "type": "text", + "content": "Each of the shuffled representations is processed by a convolution operation, where " + }, + { + "bbox": [ + 55, + 346, + 296, + 441 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_p^m\\in \\mathbb{R}^{C_m\\times C_m\\times 1\\times 1}" + }, + { + "bbox": [ + 55, + 346, + 296, + 441 + ], + "type": "text", + "content": " is the convolution kernel, " + }, + { + "bbox": [ + 55, + 346, + 296, + 441 + ], + "type": "inline_equation", + "content": "*" + }, + { + "bbox": [ + 55, + 346, + 296, + 441 + ], + "type": "text", + "content": " denotes the convolution operation, and " + }, + { + "bbox": [ + 55, + 346, + 296, + 441 + ], + "type": "inline_equation", + "content": "\\mathbf{b}_p^m\\in \\mathbb{R}^{C_m}" + }, + { + "bbox": [ + 55, + 346, + 296, + 441 + ], + "type": "text", + "content": " is a learnable bias vector. The result is then passed through a transformation " + }, + { + "bbox": [ + 55, + 346, + 296, + 441 + ], + "type": "inline_equation", + "content": "f_{p}(\\cdot)" + }, + { + "bbox": [ + 55, + 346, + 296, + 441 + ], + "type": "text", + "content": ", consisting of Batch Normalization, GELU activation, and Dropout. The final output of the " + }, + { + "bbox": [ + 55, + 346, + 296, + 441 + ], + "type": "inline_equation", + "content": "\\mathrm{D}^2\\mathrm{P}" + }, + { + "bbox": [ + 55, + 346, + 296, + 441 + ], + "type": "text", + "content": " module is obtained by summing the outputs from both shuffling paths" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 64, + 462, + 284, + 502 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 462, + 284, + 502 + ], + "spans": [ + { + "bbox": [ + 64, + 462, + 284, + 502 + ], + "type": "interline_equation", + "content": "\\mathbf {X} _ {\\mathrm {s}} ^ {d _ {m}} = f _ {a} (\\psi_ {m}) \\times \\left[ \\sum_ {p \\in \\{0, 1 \\}} f _ {p} \\left(\\mathbf {W} _ {p} ^ {m} * \\mathbf {X} _ {\\mathrm {s}, p} ^ {o _ {m}} + \\mathbf {b} _ {p} ^ {m}\\right) \\right],", + "image_path": "7d1822d47aafe0ee59e67bd41151d5e06e691b3107dbaf85ba81c4fa53a87718.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 72, + 503, + 294, + 526 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 503, + 294, + 526 + ], + "spans": [ + { + "bbox": [ + 72, + 503, + 294, + 526 + ], + "type": "interline_equation", + "content": "\\psi_ {m} \\in (- \\infty , \\infty), \\quad f _ {a} (\\cdot) \\in [ 0, 1 ], \\tag {10}", + "image_path": "2dbe24ec04d0c284fa9e49b1cc2f448bc255948dfa47782ee7d61bfbc4554286.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 528, + 296, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 528, + 296, + 564 + ], + "spans": [ + { + "bbox": [ + 55, + 528, + 296, + 564 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 528, + 296, + 564 + ], + "type": "inline_equation", + "content": "\\psi_{m}" + }, + { + "bbox": [ + 55, + 528, + 296, + 564 + ], + "type": "text", + "content": " is a learnable scalar parameter, and the scaling function " + }, + { + "bbox": [ + 55, + 528, + 296, + 564 + ], + "type": "inline_equation", + "content": "f_{a}(\\cdot)" + }, + { + "bbox": [ + 55, + 528, + 296, + 564 + ], + "type": "text", + "content": " is defined in Eq. (6) to constrain the modulation factor within the range [0, 1]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 574, + 296, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 574, + 296, + 586 + ], + "spans": [ + { + "bbox": [ + 55, + 574, + 296, + 586 + ], + "type": "text", + "content": "4.3. Enhanced Orderless Latent Texture Attributes" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 593, + 296, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 593, + 296, + 652 + ], + "spans": [ + { + "bbox": [ + 55, + 593, + 296, + 652 + ], + "type": "text", + "content": "While GTN [10] enhances LTA representations via message passing and aggregation, CAPTN improves the LTAs generated by the TFA module, " + }, + { + "bbox": [ + 55, + 593, + 296, + 652 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_a^{t_m}" + }, + { + "bbox": [ + 55, + 593, + 296, + 652 + ], + "type": "text", + "content": ", by injecting dynamically adapted adjacency structures from the " + }, + { + "bbox": [ + 55, + 593, + 296, + 652 + ], + "type": "inline_equation", + "content": "\\mathrm{D}^2\\mathrm{P}" + }, + { + "bbox": [ + 55, + 593, + 296, + 652 + ], + "type": "text", + "content": " module, " + }, + { + "bbox": [ + 55, + 593, + 296, + 652 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_s^{d_m}" + }, + { + "bbox": [ + 55, + 593, + 296, + 652 + ], + "type": "text", + "content": ", resulting in" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 134, + 654, + 295, + 669 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 654, + 295, + 669 + ], + "spans": [ + { + "bbox": [ + 134, + 654, + 295, + 669 + ], + "type": "interline_equation", + "content": "\\mathbf {X} _ {e} ^ {m} = \\mathbf {X} _ {a} ^ {t _ {m}} + \\mathbf {X} _ {\\mathrm {s}} ^ {d _ {m}}. \\tag {11}", + "image_path": "a59a2712cdc92dbccacd56492d507137ca64958aa537b87730ef3f3ab9e3c0a1.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": "We then compute the spatial mean over the height and width of the enhanced LTAs, " + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_e^m\\in \\mathbb{R}^{H_m\\times W_m\\times C_m}" + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": ", from layer " + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": " to obtain an orderless representation, " + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{O}_{\\mathrm{e}}^{m}\\in \\mathbb{R}^{C_{m}}" + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 72, + 495, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 495, + 84 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 495, + 84 + ], + "type": "text", + "content": "4.4. Learnable Chebyshev Polynomial" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 89, + 555, + 233 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 89, + 555, + 233 + ], + "spans": [ + { + "bbox": [ + 313, + 89, + 555, + 233 + ], + "type": "text", + "content": "Chebyshev polynomials [26] offer a powerful global basis for modeling non-linear, oscillatory, and fine-grained variations in orderless LTAs. Their higher-order, non-monotonic structure enables expressive transformations that capture both smooth and oscillatory variations, allowing the model to distinguish subtle patterns while aligning semantically similar LTA values—even when numerically distant. To enhance the expressiveness of each orderless LTA dimension, we apply a Chebyshev polynomial " + }, + { + "bbox": [ + 313, + 89, + 555, + 233 + ], + "type": "inline_equation", + "content": "T_{n}(\\cdot)" + }, + { + "bbox": [ + 313, + 89, + 555, + 233 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 313, + 89, + 555, + 233 + ], + "type": "inline_equation", + "content": "n \\in \\{0, \\dots, d\\}" + }, + { + "bbox": [ + 313, + 89, + 555, + 233 + ], + "type": "text", + "content": ", each weighted by a sigmoid function parameterized by " + }, + { + "bbox": [ + 313, + 89, + 555, + 233 + ], + "type": "inline_equation", + "content": "\\theta_{n}" + }, + { + "bbox": [ + 313, + 89, + 555, + 233 + ], + "type": "text", + "content": ". The resulting formulation, referred to as a Learnable Chebyshev Polynomial (LCP), takes the form" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 356, + 242, + 555, + 275 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 356, + 242, + 555, + 275 + ], + "spans": [ + { + "bbox": [ + 356, + 242, + 555, + 275 + ], + "type": "interline_equation", + "content": "P \\left(\\mathbf {O} _ {\\mathrm {e}} ^ {m}\\right) = \\sum_ {n = 0} ^ {d} T _ {n} \\left(\\mathbf {O} _ {\\mathrm {e}} ^ {m}\\right) \\left[ \\frac {1}{1 + e ^ {- \\theta_ {n}}} \\right], \\tag {12}", + "image_path": "0e246b1f648c327c193be2d28541340632c8562e7ac8c0471fc378f3773b71ba.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 278, + 554, + 302 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 278, + 554, + 302 + ], + "spans": [ + { + "bbox": [ + 313, + 278, + 554, + 302 + ], + "type": "text", + "content": "The Chebyshev polynomials " + }, + { + "bbox": [ + 313, + 278, + 554, + 302 + ], + "type": "inline_equation", + "content": "T_{n}(\\mathbf{O}_{\\mathrm{e}}^{m})" + }, + { + "bbox": [ + 313, + 278, + 554, + 302 + ], + "type": "text", + "content": " used in the expansion follow the recurrence relation" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 334, + 322, + 554, + 379 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 334, + 322, + 554, + 379 + ], + "spans": [ + { + "bbox": [ + 334, + 322, + 554, + 379 + ], + "type": "interline_equation", + "content": "T _ {n} \\left(\\mathbf {O} _ {\\mathrm {e}} ^ {m}\\right) = \\left\\{ \\begin{array}{l l} \\mathbf {1}, & \\text {i f} n = 0, \\\\ \\mathbf {O} _ {\\mathrm {e}} ^ {m}, & \\text {i f} n = 1, \\\\ 2 \\mathbf {O} _ {\\mathrm {e}} ^ {m} T _ {n - 1} \\left(\\mathbf {O} _ {\\mathrm {e}} ^ {m}\\right) & \\text {f o r} n \\geq 2. \\\\ - T _ {n - 2} \\left(\\mathbf {O} _ {\\mathrm {e}} ^ {m}\\right), & \\end{array} \\right. \\tag {13}", + "image_path": "75f36aa317801f18fe685d6a82874c5b1618baa5379568ed09b02629f84bd9d9.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 386, + 554, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 386, + 554, + 411 + ], + "spans": [ + { + "bbox": [ + 313, + 386, + 554, + 411 + ], + "type": "text", + "content": "LCPs from all layers are then concatenated and passed through a fully connected layer, " + }, + { + "bbox": [ + 313, + 386, + 554, + 411 + ], + "type": "inline_equation", + "content": "f_{fc}" + }, + { + "bbox": [ + 313, + 386, + 554, + 411 + ], + "type": "text", + "content": ", as" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 378, + 418, + 554, + 439 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 378, + 418, + 554, + 439 + ], + "spans": [ + { + "bbox": [ + 378, + 418, + 554, + 439 + ], + "type": "interline_equation", + "content": "\\mathbf {Y} = f _ {f c} \\left( \\right.\\left\\| \\right. _ {m = 1} ^ {N - 1} P \\left(\\mathbf {O} _ {\\mathrm {e}} ^ {m}\\right)\\left. \\right). \\tag {14}", + "image_path": "e6f72bd79cf1a90095aff73006c24a23bb0704e4ca94b9862b468febe3a9e870.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 442, + 554, + 466 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 442, + 554, + 466 + ], + "spans": [ + { + "bbox": [ + 313, + 442, + 554, + 466 + ], + "type": "text", + "content": "The logits, " + }, + { + "bbox": [ + 313, + 442, + 554, + 466 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}" + }, + { + "bbox": [ + 313, + 442, + 554, + 466 + ], + "type": "text", + "content": ", are then evaluated using the Latent Texture Attribute Loss function, which will be discussed in Sec. 5." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 473, + 524, + 487 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 473, + 524, + 487 + ], + "spans": [ + { + "bbox": [ + 313, + 473, + 524, + 487 + ], + "type": "text", + "content": "4.5. Spatial Latent Attribute Representation" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 491, + 555, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 491, + 555, + 622 + ], + "spans": [ + { + "bbox": [ + 313, + 491, + 555, + 622 + ], + "type": "text", + "content": "Finally, we aim to enforce some extent of structural similarity between a texture patch extracted from Sec. 3 and regional latent texture attribute information. To achieve this, we first concatenate the enhanced LTAs, " + }, + { + "bbox": [ + 313, + 491, + 555, + 622 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_e^{M - 1}" + }, + { + "bbox": [ + 313, + 491, + 555, + 622 + ], + "type": "text", + "content": ", with the original feature maps from the frozen backbone, " + }, + { + "bbox": [ + 313, + 491, + 555, + 622 + ], + "type": "inline_equation", + "content": "\\mathbf{X}^{O_M - 1}" + }, + { + "bbox": [ + 313, + 491, + 555, + 622 + ], + "type": "text", + "content": ". Both representations are obtained from the final layer of the network. We integrate the frozen backbone features into the LTAs through a structured transformation consisting of three sequential operations: cropping " + }, + { + "bbox": [ + 313, + 491, + 555, + 622 + ], + "type": "inline_equation", + "content": "f_{c}(\\cdot)" + }, + { + "bbox": [ + 313, + 491, + 555, + 622 + ], + "type": "text", + "content": ", a " + }, + { + "bbox": [ + 313, + 491, + 555, + 622 + ], + "type": "inline_equation", + "content": "3\\times 3" + }, + { + "bbox": [ + 313, + 491, + 555, + 622 + ], + "type": "text", + "content": " convolution, and bilinear interpolation " + }, + { + "bbox": [ + 313, + 491, + 555, + 622 + ], + "type": "inline_equation", + "content": "f_{b}(\\cdot)" + }, + { + "bbox": [ + 313, + 491, + 555, + 622 + ], + "type": "text", + "content": ". The transformation is formally defined as" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 330, + 643, + 555, + 658 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 330, + 643, + 555, + 658 + ], + "spans": [ + { + "bbox": [ + 330, + 643, + 555, + 658 + ], + "type": "interline_equation", + "content": "\\mathbf {G} _ {\\mathrm {p}} = f _ {b} \\left(\\mathbf {W} _ {s} * f _ {c} \\left(\\mathbf {X} _ {e} ^ {M - 1} \\| \\mathbf {X} ^ {O M - 1}\\right) + \\mathbf {b} _ {s}, P\\right), \\tag {15}", + "image_path": "839ec1c4b0fb4a13690907e6f19a39ad698b6f6fac0e3a27ac8abaedc589b0a6.jpg" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 664, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 664, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 664, + 555, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 664, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_s\\in \\mathbb{R}^{3\\times C\\times 3\\times 3}" + }, + { + "bbox": [ + 313, + 664, + 555, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 664, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{b}_s\\in \\mathbb{R}^3" + }, + { + "bbox": [ + 313, + 664, + 555, + 715 + ], + "type": "text", + "content": " are the learnable weights and bias of the convolutional layer. The output of this transformation is the Spatial Latent Attribute Representation (SLAR), " + }, + { + "bbox": [ + 313, + 664, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_{\\mathrm{p}}\\in \\mathbb{R}^{3\\times P\\times P}" + }, + { + "bbox": [ + 313, + 664, + 555, + 715 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 664, + 555, + 715 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 313, + 664, + 555, + 715 + ], + "type": "text", + "content": " is the spatial size" + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23427" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 191 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 191 + ], + "type": "text", + "content": "of the extracted patch. This ensures that the extracted regional texture attributes retain both the fine-grained information from LTAs and the structural guidance preserved in the frozen backbone features. Incorporating structure-aware features from the frozen backbone provides a grounded reference that constrains the adaptation of LTAs during training, preventing overcorrection while maintaining a balance between texture fidelity and spatial consistency. SLAR will be used to compute Spatial Latent Texture Attribute Loss (SL) in Sec. 5." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 205, + 224, + 217 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 205, + 224, + 217 + ], + "spans": [ + { + "bbox": [ + 55, + 205, + 224, + 217 + ], + "type": "text", + "content": "5. Latent Texture Attribute Loss" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 226, + 296, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 226, + 296, + 357 + ], + "spans": [ + { + "bbox": [ + 55, + 226, + 296, + 357 + ], + "type": "text", + "content": "We propose the Latent Texture Attribute (LTA) Loss to enhance classification accuracy while preserving texture consistency at both global and local levels. The LTA Loss integrates three complementary components: (1) Cross Entropy Loss " + }, + { + "bbox": [ + 55, + 226, + 296, + 357 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{CE})" + }, + { + "bbox": [ + 55, + 226, + 296, + 357 + ], + "type": "text", + "content": " to promote accurate class predictions, (2) Orderless LTA Loss " + }, + { + "bbox": [ + 55, + 226, + 296, + 357 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{OL})" + }, + { + "bbox": [ + 55, + 226, + 296, + 357 + ], + "type": "text", + "content": " to align global, spatially invariant texture distributions, and (3) Spatial LTA Loss " + }, + { + "bbox": [ + 55, + 226, + 296, + 357 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{SL})" + }, + { + "bbox": [ + 55, + 226, + 296, + 357 + ], + "type": "text", + "content": " to preserve fine-grained, spatial texture structure. By jointly optimizing these objectives, the model learns rich and coherent latent texture representations that support both accurate classification and texture understanding." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 358, + 296, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 358, + 296, + 514 + ], + "spans": [ + { + "bbox": [ + 55, + 358, + 296, + 514 + ], + "type": "text", + "content": "Orderless Latent Texture Attribute Loss. This second texture loss component compares distribution of orderless LTAs with orderless features extracted directly from the frozen backbone. By aligning loss from learned texture representations with a stable baseline, we encourage consistency in the model's understanding of spatial invariant texture representations across images. We apply Batch Normalization, " + }, + { + "bbox": [ + 55, + 358, + 296, + 514 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{bn}}" + }, + { + "bbox": [ + 55, + 358, + 296, + 514 + ], + "type": "text", + "content": ", on feature maps from the " + }, + { + "bbox": [ + 55, + 358, + 296, + 514 + ], + "type": "inline_equation", + "content": "m^{\\mathrm{th}}" + }, + { + "bbox": [ + 55, + 358, + 296, + 514 + ], + "type": "text", + "content": " layer of the backbone, " + }, + { + "bbox": [ + 55, + 358, + 296, + 514 + ], + "type": "inline_equation", + "content": "\\mathbf{X}^{o_m}" + }, + { + "bbox": [ + 55, + 358, + 296, + 514 + ], + "type": "text", + "content": ", before calculating the spatial mean and concatenating multiple scales of orderless texture information across layers " + }, + { + "bbox": [ + 55, + 358, + 296, + 514 + ], + "type": "inline_equation", + "content": "m \\in [1, M - 1]" + }, + { + "bbox": [ + 55, + 358, + 296, + 514 + ], + "type": "text", + "content": " of the backbone to obtain " + }, + { + "bbox": [ + 55, + 358, + 296, + 514 + ], + "type": "inline_equation", + "content": "\\mathbf{O}_{\\mathrm{b}}" + }, + { + "bbox": [ + 55, + 358, + 296, + 514 + ], + "type": "text", + "content": ". We also concatenate orderless enhanced LTAs, " + }, + { + "bbox": [ + 55, + 358, + 296, + 514 + ], + "type": "inline_equation", + "content": "\\mathbf{O}_{\\mathrm{e}}^{m}" + }, + { + "bbox": [ + 55, + 358, + 296, + 514 + ], + "type": "text", + "content": ", across the same range of layers to obtain " + }, + { + "bbox": [ + 55, + 358, + 296, + 514 + ], + "type": "inline_equation", + "content": "\\mathbf{O}_{\\mathrm{e}}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 533, + 296, + 604 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 533, + 296, + 604 + ], + "spans": [ + { + "bbox": [ + 67, + 533, + 296, + 604 + ], + "type": "interline_equation", + "content": "\\mathbf {O} _ {\\mathrm {b}} = \\left\\| \\right. _ {m = 1} ^ {M - 1} \\left(\\frac {\\sum_ {h = 0} ^ {H _ {m} - 1} \\sum_ {w = 0} ^ {W _ {m} - 1} f _ {\\mathrm {b n}} \\left(\\mathbf {X} ^ {o _ {m}}\\right) [ h , w , : ]}{H _ {m} \\times W _ {m}}\\right), \\tag {16}", + "image_path": "74dab86f7684d48920bd87cc2b1c59b4a33a04cc92d53befa3dc09615e749f75.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 140, + 615, + 296, + 647 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 615, + 296, + 647 + ], + "spans": [ + { + "bbox": [ + 140, + 615, + 296, + 647 + ], + "type": "interline_equation", + "content": "\\mathbf {O} _ {\\mathrm {e}} = \\left\\| _ {m = 1} ^ {M - 1} \\mathbf {O} _ {\\mathrm {e}} ^ {m}. \\right. \\tag {17}", + "image_path": "686a1900a9db2fb3fd0e9e181e07cbbce3dc377b30a94902497514a9ac7ce5ee.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 652, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 652, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 652, + 296, + 714 + ], + "type": "text", + "content": "We perform softmax normalization of " + }, + { + "bbox": [ + 55, + 652, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{O_b}" + }, + { + "bbox": [ + 55, + 652, + 296, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 652, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{O_e}" + }, + { + "bbox": [ + 55, + 652, + 296, + 714 + ], + "type": "text", + "content": " over the total number of channels in each sample " + }, + { + "bbox": [ + 55, + 652, + 296, + 714 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 55, + 652, + 296, + 714 + ], + "type": "text", + "content": " from the minibatch, aggregated over layers " + }, + { + "bbox": [ + 55, + 652, + 296, + 714 + ], + "type": "inline_equation", + "content": "m\\in [1,M - 1]" + }, + { + "bbox": [ + 55, + 652, + 296, + 714 + ], + "type": "text", + "content": ", where the total number of concatenated channels is " + }, + { + "bbox": [ + 55, + 652, + 296, + 714 + ], + "type": "inline_equation", + "content": "C = \\sum_{m = 1}^{M - 1}C_m" + }, + { + "bbox": [ + 55, + 652, + 296, + 714 + ], + "type": "text", + "content": ". A small constant " + }, + { + "bbox": [ + 55, + 652, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 55, + 652, + 296, + 714 + ], + "type": "text", + "content": " is added for numerical stability" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 353, + 81, + 553, + 113 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 81, + 553, + 113 + ], + "spans": [ + { + "bbox": [ + 353, + 81, + 553, + 113 + ], + "type": "interline_equation", + "content": "p _ {\\alpha , i} ^ {s} = \\frac {e ^ {\\mathbf {O} _ {\\alpha , i} ^ {s}}}{\\sum_ {j = 0} ^ {C - 1} e ^ {\\mathbf {O} _ {\\alpha , j} ^ {s}}} + \\epsilon , \\quad \\alpha \\in \\{\\mathrm {b}, \\mathrm {e} \\}. \\tag {18}", + "image_path": "0012e5c489e86e8e91127d169c38e15a16fbd73ad92ae1b48c8f565464772ef1.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 116, + 555, + 178 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 116, + 555, + 178 + ], + "spans": [ + { + "bbox": [ + 313, + 116, + 555, + 178 + ], + "type": "text", + "content": "Inspired by Kullback-Leibler (KL) divergence, we calculate orderless LTA loss, " + }, + { + "bbox": [ + 313, + 116, + 555, + 178 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{OL}" + }, + { + "bbox": [ + 313, + 116, + 555, + 178 + ], + "type": "text", + "content": " that measures the dissimilarity between two probability distributions: a fixed distribution from orderless feature maps extracted from the frozen backbone, " + }, + { + "bbox": [ + 313, + 116, + 555, + 178 + ], + "type": "inline_equation", + "content": "p_{\\mathrm{b},i}^{\\mathrm{s}}" + }, + { + "bbox": [ + 313, + 116, + 555, + 178 + ], + "type": "text", + "content": " and that from orderless enhanced LTAs, " + }, + { + "bbox": [ + 313, + 116, + 555, + 178 + ], + "type": "inline_equation", + "content": "p_{\\mathrm{e},i}^{\\mathrm{s}}" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 325, + 194, + 553, + 228 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 194, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 325, + 194, + 553, + 228 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {O L} = \\frac {1}{S} \\sum_ {s = 0} ^ {S - 1} \\sum_ {i = 0} ^ {C - 1} p _ {\\mathrm {b}, i} ^ {s} \\left[ \\log_ {e} \\left(p _ {\\mathrm {b}, i} ^ {s}\\right) - \\log_ {e} \\left(p _ {\\mathrm {e}, i} ^ {s}\\right) \\right], \\tag {19}", + "image_path": "12ea4d1cf9dfdd24e981588311bc6818e71f29083c3c10f0491dd44041429263.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 236, + 553, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 236, + 553, + 258 + ], + "spans": [ + { + "bbox": [ + 313, + 236, + 553, + 258 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 236, + 553, + 258 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 313, + 236, + 553, + 258 + ], + "type": "text", + "content": " represents the number of samples in the minibatch." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 260, + 555, + 356 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 260, + 555, + 356 + ], + "spans": [ + { + "bbox": [ + 313, + 260, + 555, + 356 + ], + "type": "text", + "content": "Spatial Latent Texture Attribute Loss. The final LTA loss component measures the similarity between SLAR from Sec. 4.5 and a corresponding texture patch extracted from the original image. This component helps ensure that the latent representations preserve both regional and local spatial texture information by comparing the latent attributes " + }, + { + "bbox": [ + 313, + 260, + 555, + 356 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_{\\mathrm{p}}" + }, + { + "bbox": [ + 313, + 260, + 555, + 356 + ], + "type": "text", + "content": " with the extracted texture patch " + }, + { + "bbox": [ + 313, + 260, + 555, + 356 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_{\\mathrm{t}}" + }, + { + "bbox": [ + 313, + 260, + 555, + 356 + ], + "type": "text", + "content": " using the Mean Squared Error (MSE)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 335, + 374, + 553, + 399 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 374, + 553, + 399 + ], + "spans": [ + { + "bbox": [ + 335, + 374, + 553, + 399 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {S L} = \\frac {1}{B C _ {M - 1} H _ {M - 1} W _ {M - 1}} \\| \\mathbf {G} _ {\\mathrm {p}} - \\mathbf {G} _ {\\mathrm {t}} \\| _ {F} ^ {2}, \\tag {20}", + "image_path": "8b73cd81e67d2786b33af86354cdf0058b9d446bbab41640d84e1ceb73cdbc97.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 407, + 554, + 527 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 407, + 554, + 527 + ], + "spans": [ + { + "bbox": [ + 313, + 407, + 554, + 527 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 407, + 554, + 527 + ], + "type": "inline_equation", + "content": "B, C_{M-1}, H_{M-1}, W_{M-1}" + }, + { + "bbox": [ + 313, + 407, + 554, + 527 + ], + "type": "text", + "content": " represent the batch size, number of channels, height, and width of the feature maps, from the final " + }, + { + "bbox": [ + 313, + 407, + 554, + 527 + ], + "type": "inline_equation", + "content": "M-1" + }, + { + "bbox": [ + 313, + 407, + 554, + 527 + ], + "type": "text", + "content": " layer, and " + }, + { + "bbox": [ + 313, + 407, + 554, + 527 + ], + "type": "inline_equation", + "content": "\\|\\cdot\\|_F" + }, + { + "bbox": [ + 313, + 407, + 554, + 527 + ], + "type": "text", + "content": " denotes the Frobenius norm, which sums the squared differences over all elements. This loss encourages the model to preserve fine-grained spatial texture patterns within the latent space, ensuring that learned texture features remain spatially structured and contextually coherent. By aligning latent spatial representations with real texture patches, the model enhances its ability to reconstruct or classify textures with high fidelity." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 528, + 554, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 528, + 554, + 563 + ], + "spans": [ + { + "bbox": [ + 313, + 528, + 554, + 563 + ], + "type": "text", + "content": "Latent Texture Attribute Loss. We formulate our loss during the training phase as a weighted sum of the three loss components" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 361, + 575, + 553, + 588 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 575, + 553, + 588 + ], + "spans": [ + { + "bbox": [ + 361, + 575, + 553, + 588 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {t r a i n}} = \\lambda \\mathcal {L} _ {C E} + \\mu \\mathcal {L} _ {O L} + \\nu \\mathcal {L} _ {S L}, \\tag {21}", + "image_path": "0d09e2371e434e2962c43125eb6670be24f344aae91872e8dbb1c163a803e303.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 594, + 555, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 617 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 617 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 594, + 555, + 617 + ], + "type": "inline_equation", + "content": "\\lambda, \\mu, \\nu \\in [0,1]" + }, + { + "bbox": [ + 313, + 594, + 555, + 617 + ], + "type": "text", + "content": " are weights associated with each of the loss components." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 628, + 395, + 641 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 628, + 395, + 641 + ], + "spans": [ + { + "bbox": [ + 313, + 628, + 395, + 641 + ], + "type": "text", + "content": "6. Experiments" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 648, + 377, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 648, + 377, + 659 + ], + "spans": [ + { + "bbox": [ + 313, + 648, + 377, + 659 + ], + "type": "text", + "content": "6.1. Datasets" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "type": "text", + "content": "We evaluated our network using five benchmark datasets, summarized as follows: (a) Ground Terrain in Outdoor Scenes (GTOS) [34] contains 34,105 images of outdoor ground terrain across 40 categories, with 5 train/test splits." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23428" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 70, + 336, + 114 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 336, + 114 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 336, + 114 + ], + "type": "text", + "content": "Table 1. Comparison of CAPTN with GTN [10] and RADAM [27] in terms of classification accuracy (\\%) on texture benchmarks. ConvNeXt-\\* variants are used as backbones. Pre-trained weights are frozen. The best result for each backbone-dataset combination is bolded." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "text" + }, + { + "type": "table", + "bbox": [ + 56, + 123, + 335, + 251 + ], + "blocks": [ + { + "bbox": [ + 56, + 123, + 335, + 251 + ], + "lines": [ + { + "bbox": [ + 56, + 123, + 335, + 251 + ], + "spans": [ + { + "bbox": [ + 56, + 123, + 335, + 251 + ], + "type": "table", + "html": "
Model-*DTDFMDKTHGTOSGTOS-M
RADAMN74.9 ±0.787.1 ±0.489.6 ±3.883.7 ±1.581.8
GTNN75.4 ±1.186.9 ±3.190.3 ±4.782.9 ±1.790.1
CAPTN (ours)N77.9 ±1.087.0 ±2.990.6 ±4.683.9 ±1.791.5
RADAMT77.0 ±0.788.7 ±0.490.7 ±4.084.2 ±1.785.3
GTNT77.5 ±1.087.2 ±3.492.0 ±4.984.7 ±1.691.9
CAPTN (ours)T78.9 ±0.887.2 ±3.692.2 ±4.485.1 ±1.794.2
RADAMB76.4 ±0.990.2 ±0.287.7 ±5.684.1 ±1.682.2
GTNB78.3 ±0.888.8 ±3.090.8 ±6.284.2 ±1.894.8
CAPTN (ours)B79.3±0.989.1 ±3.491.3±5.784.6 ±1.994.0
RADAML77.4 ±1.189.3±0.389.3 ±3.484.0 ±1.885.8
GTNL79.3 ±0.888.1 ±3.092.5 ±4.784.9 ±1.793.9
CAPTN (ours)L79.6±0.889.3±3.193.0 ±4.985.0 ±1.795.1
", + "image_path": "22d5ac38b3ee23a8550022f2ee47a9506a543771b56ccb0012600b1763cbfb09.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 62, + 324, + 290, + 362 + ], + "blocks": [ + { + "bbox": [ + 55, + 282, + 295, + 316 + ], + "lines": [ + { + "bbox": [ + 55, + 282, + 295, + 316 + ], + "spans": [ + { + "bbox": [ + 55, + 282, + 295, + 316 + ], + "type": "text", + "content": "Table 4. Classification accuracy (\\%) comparison between " + }, + { + "bbox": [ + 55, + 282, + 295, + 316 + ], + "type": "inline_equation", + "content": "1^{\\text{st}}" + }, + { + "bbox": [ + 55, + 282, + 295, + 316 + ], + "type": "text", + "content": "-order mean pooling (w/ LCP) and a non-polynomial baseline: " + }, + { + "bbox": [ + 55, + 282, + 295, + 316 + ], + "type": "inline_equation", + "content": "2^{\\text{nd}}" + }, + { + "bbox": [ + 55, + 282, + 295, + 316 + ], + "type": "text", + "content": "-order covariance pooling (w/o LCP), with ConvNeXt-B backbone." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 62, + 324, + 290, + 362 + ], + "lines": [ + { + "bbox": [ + 62, + 324, + 290, + 362 + ], + "spans": [ + { + "bbox": [ + 62, + 324, + 290, + 362 + ], + "type": "table", + "html": "
PoolingDTDKTHGTOSGTOS-M
2nd-order76.4 ±1.091.1 ±6.383.5 ±1.793.0
1st-order79.3 ±0.991.3 ±5.784.6 ±1.994.0
", + "image_path": "59cd6d53e5c1d0274ce82752456819d0a915518f60fc5ad086a3ebb8c56de2b4.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 69, + 404, + 282, + 441 + ], + "blocks": [ + { + "bbox": [ + 55, + 373, + 295, + 396 + ], + "lines": [ + { + "bbox": [ + 55, + 373, + 295, + 396 + ], + "spans": [ + { + "bbox": [ + 55, + 373, + 295, + 396 + ], + "type": "text", + "content": "Table 5. Classification accuracy (\\%) of CAPTN with different basis functions using ConvNeXt-L backbone." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 69, + 404, + 282, + 441 + ], + "lines": [ + { + "bbox": [ + 69, + 404, + 282, + 441 + ], + "spans": [ + { + "bbox": [ + 69, + 404, + 282, + 441 + ], + "type": "table", + "html": "
Basis Fn.DTDKTHGTOSGTOS-M
RBF74.1 ±1.089.6 ±6.881.5 ±2.492.7
LCP79.6±0.893.0±4.985.0±1.795.1
", + "image_path": "3a5f2940e19b1d8dbe3cd80ff01530337198ce75cd0daaedde2d71e42d0f4056.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 460, + 295, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 460, + 295, + 581 + ], + "spans": [ + { + "bbox": [ + 54, + 460, + 295, + 581 + ], + "type": "text", + "content": "(b) GTOS-Mobile (GTOS-M) [35] includes 100,011 material samples from 31 categories, captured via mobile phone, with a single train/test split. (c) KTH-TIPS2-b (KTH) [3] provides 4,752 images in 11 material categories, with 4 train/test splits. (d) Flickr Material Dataset (FMD) [28] contains 1,000 images spanning 10 material categories, validated with 10-fold cross-validation and stratified sampling to maintain class proportions. (e) Describable Texture Dataset (DTD) [6] has 5,640 images across 47 categories, with 10 train/validation/test splits." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 588, + 188, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 588, + 188, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 588, + 188, + 601 + ], + "type": "text", + "content": "6.2. Implementation Details" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 605, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 605, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 54, + 605, + 295, + 713 + ], + "type": "text", + "content": "Our model was implemented in PyTorch and trained on a single NVIDIA RTX A6000 GPU. For all datasets, the learning rate was set to 0.0001 and followed a cosine annealing schedule [20]. We used the Adam optimizer with decoupled weight decay [21] and our novel Latent Texture Attribute Loss function. Our network was trained until the final epoch for each dataset, with test set performance reported. More design and implementation details can be found in the supplementary material." + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 343, + 113, + 553, + 167 + ], + "blocks": [ + { + "bbox": [ + 342, + 70, + 553, + 103 + ], + "lines": [ + { + "bbox": [ + 342, + 70, + 553, + 103 + ], + "spans": [ + { + "bbox": [ + 342, + 70, + 553, + 103 + ], + "type": "text", + "content": "Table 2. Evaluate effectiveness of TFA, " + }, + { + "bbox": [ + 342, + 70, + 553, + 103 + ], + "type": "inline_equation", + "content": "{\\mathrm{D}}^{2}\\mathrm{P}" + }, + { + "bbox": [ + 342, + 70, + 553, + 103 + ], + "type": "text", + "content": " ,and LCP in terms of classification accuracy (%) . ConvNeXt-N is the backbone." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 343, + 113, + 553, + 167 + ], + "lines": [ + { + "bbox": [ + 343, + 113, + 553, + 167 + ], + "spans": [ + { + "bbox": [ + 343, + 113, + 553, + 167 + ], + "type": "table", + "html": "
TFAD2PLCPDTDKTHGTOS
XXX75.4 ±1.087.2 ±5.381.7 ±2.0
✓XX77.1 ±1.190.4 ±4.883.5 ±1.8
✓✓X77.2 ±1.090.5 ±4.883.6 ±1.8
✓✓✓77.9 ±1.090.6 ±4.683.9 ±1.7
", + "image_path": "92367dda5669648f81dfe51ef5152732d57e8731e15d3b4015d0856c9e29efd9.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 343, + 210, + 553, + 247 + ], + "blocks": [ + { + "bbox": [ + 342, + 179, + 553, + 201 + ], + "lines": [ + { + "bbox": [ + 342, + 179, + 553, + 201 + ], + "spans": [ + { + "bbox": [ + 342, + 179, + 553, + 201 + ], + "type": "text", + "content": "Table 3. Evaluate effectiveness of LTA Loss in terms of classification accuracy " + }, + { + "bbox": [ + 342, + 179, + 553, + 201 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 342, + 179, + 553, + 201 + ], + "type": "text", + "content": ". ConvNeXt-T is the backbone." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 343, + 210, + 553, + 247 + ], + "lines": [ + { + "bbox": [ + 343, + 210, + 553, + 247 + ], + "spans": [ + { + "bbox": [ + 343, + 210, + 553, + 247 + ], + "type": "table", + "html": "
Loss Fn.DTDKTHGTOSGTOS-M
CE Loss78.4 ±0.791.9 ±5.185.0 ±1.893.6
LTA Loss78.9 ±0.892.2 ±4.485.1 ±1.794.2
", + "image_path": "2b7b51b2bfae0b142059a992b183e62f9887aac8c82698ef8ff4fabdf203f65e.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 283, + 376, + 296 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 283, + 376, + 296 + ], + "spans": [ + { + "bbox": [ + 314, + 283, + 376, + 296 + ], + "type": "text", + "content": "6.3. Analysis" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 306, + 553, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 306, + 553, + 365 + ], + "spans": [ + { + "bbox": [ + 313, + 306, + 553, + 365 + ], + "type": "text", + "content": "Recent studies [10, 27] have demonstrated the superiority of ConvNeXt backbones over ResNet backbones for texture-specific applications, particularly in multiscale texture representation. Hence, all our experiments will be based on ConvNeXt backbones." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 366, + 554, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 366, + 554, + 544 + ], + "spans": [ + { + "bbox": [ + 313, + 366, + 554, + 544 + ], + "type": "text", + "content": "Quantitative. Tab. 1 depicts how CAPTN's performance compares with the current state-of-the-art RADAM [27] and GTN [10] across various ConvNeXt backbones pretrained on ImageNet. The average and standard deviation measures are calculated across the provided splits, where applicable. Following RADAM and GTN, we have frozen our ConvNeXt backbones. CAPTN out-performed both RADAM [27] and GTN [10] in most of the benchmark datasets across different ConvNeXt backbone sizes. Both RADAM and GTN do not have a mechanism to represent orderless texture features or attributes. On the other hand, CAPTN represents orderless latent texture attributes as a learnable polynomial improving its expressiveness and consequently optimizing the mapping from latent to visual texture attributes." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 545, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 554, + 713 + ], + "type": "text", + "content": "Qualitative. The illustrations in Fig. 3 were generated using the ConvNeXt-L backbone with images from the DTD dataset. The Texture Frequency Attention (TFA) module from CAPTN enhances spatial attention in regions containing specific texture features. To some extent, the attention mechanism can identify the spatial regularity of grid textures. For spiraled textures, which exhibit greater irregularity, CAPTN demonstrates robustness to variations in illumination, shadows, occlusion, and background textures by selectively focusing on the spiraled region. Similarly, for polka-dotted textures, CAPTN disregards occlusions and concentrates on the textured areas. Furthermore, in all three cases, CAPTN recognizes masked regions within the images and assigns them lower importance." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23429" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 69, + 121, + 134 + ], + "blocks": [ + { + "bbox": [ + 56, + 69, + 121, + 134 + ], + "lines": [ + { + "bbox": [ + 56, + 69, + 121, + 134 + ], + "spans": [ + { + "bbox": [ + 56, + 69, + 121, + 134 + ], + "type": "image", + "image_path": "cf5dd0f9a5962ca2a49fe2d0416fd808563331fbbdd16d9a9e5e1e2d8d406111.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 143, + 68, + 210, + 134 + ], + "blocks": [ + { + "bbox": [ + 143, + 68, + 210, + 134 + ], + "lines": [ + { + "bbox": [ + 143, + 68, + 210, + 134 + ], + "spans": [ + { + "bbox": [ + 143, + 68, + 210, + 134 + ], + "type": "image", + "image_path": "78357eb568dbc8edd52514c0e6f6220560f14afc539096005be5c2bea3f375fb.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 233, + 70, + 295, + 133 + ], + "blocks": [ + { + "bbox": [ + 233, + 70, + 295, + 133 + ], + "lines": [ + { + "bbox": [ + 233, + 70, + 295, + 133 + ], + "spans": [ + { + "bbox": [ + 233, + 70, + 295, + 133 + ], + "type": "image", + "image_path": "17385c32af57754918167cb618792d53f7fc1666d31273ef4740a722708fb5c6.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 56, + 144, + 119, + 207 + ], + "blocks": [ + { + "bbox": [ + 56, + 144, + 119, + 207 + ], + "lines": [ + { + "bbox": [ + 56, + 144, + 119, + 207 + ], + "spans": [ + { + "bbox": [ + 56, + 144, + 119, + 207 + ], + "type": "image", + "image_path": "a507883055b66b37a41b6c0d3be98b0f1b6ab9a7f7b009afeccd169e70229056.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 214, + 295, + 258 + ], + "lines": [ + { + "bbox": [ + 55, + 214, + 295, + 258 + ], + "spans": [ + { + "bbox": [ + 55, + 214, + 295, + 258 + ], + "type": "text", + "content": "Figure 3. Top row: Masked images from DTD dataset (grid, spiralled, polka dotted). Bottom row: Corresponding attention maps from Texture Frequency Attention (TFA). Red indicates high attention, blue indicates low attention." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 143, + 144, + 208, + 207 + ], + "blocks": [ + { + "bbox": [ + 143, + 144, + 208, + 207 + ], + "lines": [ + { + "bbox": [ + 143, + 144, + 208, + 207 + ], + "spans": [ + { + "bbox": [ + 143, + 144, + 208, + 207 + ], + "type": "image", + "image_path": "a67e76df2cf24ac9ffc717c5f99f7819613ecd321ff533703a89a8a2c8d59c45.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 233, + 145, + 295, + 207 + ], + "blocks": [ + { + "bbox": [ + 233, + 145, + 295, + 207 + ], + "lines": [ + { + "bbox": [ + 233, + 145, + 295, + 207 + ], + "spans": [ + { + "bbox": [ + 233, + 145, + 295, + 207 + ], + "type": "image", + "image_path": "0c86009d5e3e61436ceaa5fc8704ad87976cae425ddc1000d0b48cadec304ffd.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 282, + 149, + 295 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 282, + 149, + 295 + ], + "spans": [ + { + "bbox": [ + 55, + 282, + 149, + 295 + ], + "type": "text", + "content": "6.4. Ablation Study" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 300, + 296, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 300, + 296, + 517 + ], + "spans": [ + { + "bbox": [ + 55, + 300, + 296, + 517 + ], + "type": "text", + "content": "We conducted ablation studies to evaluate the impact of TFA, " + }, + { + "bbox": [ + 55, + 300, + 296, + 517 + ], + "type": "inline_equation", + "content": "\\mathrm{D}^2\\mathrm{P}" + }, + { + "bbox": [ + 55, + 300, + 296, + 517 + ], + "type": "text", + "content": ", and LCP in CAPTN. Overall, all components improve CAPTN's performance on both texture (DTD) and material (KTH, GTOS) datasets. The quantitative results are presented in Tab. 2. First, we assessed LCP by mapping orderless LTAs to visual texture attributes without using LCP. The deterioration of results demonstrated that LCP enhances expressiveness and injects non-linearities, improving the mapping of latent to visual texture attributes. Removing " + }, + { + "bbox": [ + 55, + 300, + 296, + 517 + ], + "type": "inline_equation", + "content": "\\mathrm{D}^2\\mathrm{P}" + }, + { + "bbox": [ + 55, + 300, + 296, + 517 + ], + "type": "text", + "content": " led to a slight performance drop, as the LTA space became less robust and diverse. Finally, the removal of TFA further degraded CAPTN's performance, indicating that backbone features alone are less discriminative than LTAs. In addition, we performed an ablation study Tab. 3 to understand the effectiveness of LTA loss compared with Cross Entropy (CE) loss. The LTA loss function enhanced performance across texture and material datasets, underscoring its effectiveness for their recognition." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 527, + 204, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 527, + 204, + 540 + ], + "spans": [ + { + "bbox": [ + 55, + 527, + 204, + 540 + ], + "type": "text", + "content": "6.5. Alternative Design Choices" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "text", + "content": "Non-polynomial baselines. Table 4 demonstrates that first-order mean pooling combined with a LCP transformation outperforms the non-polynomial baseline of second-order covariance pooling. Covariance pooling captures second-order statistics by modeling inter-LTA correlations computed over spatial activations. While this can encode useful dependencies between LTA dimensions, it is also sensitive to noise, small artifacts, and spatial misalignments due to its reliance on precise variation patterns across spatial locations. In contrast, mean pooling followed by an LCP transformation offers a more robust encoding alternative. Mean pooling summarizes each LTA by computing the first-order global average across spatial dimensions, producing compact and stable representations that are less affected by lo" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "text", + "content": "cal distortions. These pooled features serve as clean inputs to the LCP module, which applies expressive, non-linear transformations to each LTA independently using Chebyshev polynomials. This combination enhances the representational power of each LTA without the noise sensitivity inherent in second-order pooling. Together, mean pooling and LCP offer a robust and discriminative representation strategy that outperforms covariance pooling." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 168, + 555, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 168, + 555, + 526 + ], + "spans": [ + { + "bbox": [ + 313, + 168, + 555, + 526 + ], + "type": "text", + "content": "Alternative Polynomial Bases. We replace the LCP module in CAPTN with a learnable RBF layer using Gaussian kernels. It computes scaled Euclidean distances to learnable centers and aggregates the activations via a learnable weight matrix. To avoid underfitting, we add a linear projection, LayerNorm, and GELU activation, along with kernel diversity for better generalization. As shown in Table 5, LCP outperforms the RBF variant across a range of texture and material datasets, while also requiring fewer learnable parameters. On the GTOS-M dataset, for example, CAPTN with a third-order LCP requires only 9.60M parameters, compared to 14.22M for its RBF-based counterpart. Although Gaussian RBFs are effective at modeling local variations, they often require many kernel centers and careful bandwidth tuning to capture diverse texture patterns, increasing computational cost. In contrast, LCPs form a global polynomial basis that efficiently captures long-range dependencies and oscillatory patterns with fewer parameters. In addition to RBFs, we also considered B-splines, which are widely used for smooth, piecewise function approximation. However, B-splines require careful knot placement and segmentwise assembly to represent multi-peaked or oscillatory signals, making them more sensitive to hyperparameter tuning and less scalable. LCPs, by comparison, provide a unified, smooth basis without discontinuities at segment boundaries, naturally capturing global texture variations while preserving sharp gradients and repetitive motifs. These properties make LCPs a more compact, scalable, and interpretable choice than both RBFs and B-splines for texture-specific applications." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 537, + 388, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 537, + 388, + 550 + ], + "spans": [ + { + "bbox": [ + 314, + 537, + 388, + 550 + ], + "type": "text", + "content": "7. Conclusion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 557, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 557, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 557, + 555, + 713 + ], + "type": "text", + "content": "We propose a comprehensive framework for robust texture representation by integrating SLTM with CAPTN. SLTM enhances generalization by mitigating over-reliance on fine-grained details, while the TFA mechanism captures multifrequency texture cues to support visual interpretation of spatial texture information during the generation of LTAs. Furthermore, the " + }, + { + "bbox": [ + 313, + 557, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{D}^2\\mathrm{P}" + }, + { + "bbox": [ + 313, + 557, + 555, + 713 + ], + "type": "text", + "content": " module enriches LTAs by diversifying feature representations, and LCP is specifically designed to enable a richer and more expressive encoding of orderless LTAs. To further strengthen the learning process, we introduce a unified LTA loss function that balances classification objectives with orderless alignment and spatial consistency." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "23430" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 61, + 91, + 294, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 294, + 123 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 294, + 123 + ], + "type": "text", + "content": "[1] Vincent Andrearczyk and Paul F Whelan. Using filter banks in convolutional neural networks for texture classification. Pattern Recognition Letters, 84:63-69, 2016. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 125, + 295, + 168 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 125, + 295, + 168 + ], + "spans": [ + { + "bbox": [ + 61, + 125, + 295, + 168 + ], + "type": "text", + "content": "[2] Avrim Blum and Tom Mitchell. Combining labeled and unlabeled data with co-training. In Proceedings of the eleventh annual conference on Computational learning theory, pages 92-100, 1998. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 171, + 295, + 214 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 171, + 295, + 214 + ], + "spans": [ + { + "bbox": [ + 62, + 171, + 295, + 214 + ], + "type": "text", + "content": "[3] Barbara Caputo, Eric Hayman, and P. Mallikarjuna. Class-specific material categorisation. In Tenth IEEE International Conference on Computer Vision (ICCV'05) Volume 1, pages 1597-1604 Vol. 2, 2005. 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 216, + 294, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 216, + 294, + 270 + ], + "spans": [ + { + "bbox": [ + 62, + 216, + 294, + 270 + ], + "type": "text", + "content": "[4] Zhile Chen, Feng Li, Yuhui Quan, Yong Xu, and Hui Ji. Deep texture recognition via exploiting cross-layer statistical self-similarity. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5227-5236, 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 272, + 294, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 272, + 294, + 304 + ], + "spans": [ + { + "bbox": [ + 62, + 272, + 294, + 304 + ], + "type": "text", + "content": "[5] Zhile Chen, Yuhui Quan, Ruotao Xu, Lianwen Jin, and Yong Xu. Enhancing texture representation with deep tracing pattern encoding. Pattern Recognition, 146:109959, 2024. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 306, + 294, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 306, + 294, + 350 + ], + "spans": [ + { + "bbox": [ + 62, + 306, + 294, + 350 + ], + "type": "text", + "content": "[6] Mircea Cimpoi, Subhransu Maji, Iasonas Kokkinos, Sammy Mohamed, and Andrea Vedaldi. Describing textures in the wild. In 2014 IEEE Conference on Computer Vision and Pattern Recognition, pages 3606-3613, 2014. 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 352, + 294, + 395 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 352, + 294, + 395 + ], + "spans": [ + { + "bbox": [ + 62, + 352, + 294, + 395 + ], + "type": "text", + "content": "[7] Mircea Cimpoi, Subhransu Maji, and Andrea Vedaldi. Deep filter banks for texture recognition and segmentation. In 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 3828-3836, 2015. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 396, + 294, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 396, + 294, + 441 + ], + "spans": [ + { + "bbox": [ + 62, + 396, + 294, + 441 + ], + "type": "text", + "content": "[8] Gabriella Csurka, Christopher R. Dance, Lixin Fan, Jutta Willamowski, and Cedric Bray. Visual categorization with bags of keypoints. In In Workshop on Statistical Learning in Computer Vision, ECCV, pages 1-22, 2004. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 442, + 294, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 442, + 294, + 485 + ], + "spans": [ + { + "bbox": [ + 62, + 442, + 294, + 485 + ], + "type": "text", + "content": "[9] O.G. Cula and K.J. Dana. Compact representation of bidirectional texture functions. In Proceedings of the 2001 IEEE Computer Society Conference on Computer Vision and Pattern Recognition. CVPR 2001, pages I-I, 2001. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 487, + 295, + 530 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 487, + 295, + 530 + ], + "spans": [ + { + "bbox": [ + 57, + 487, + 295, + 530 + ], + "type": "text", + "content": "[10] Ravishankar Evani, Deepu Rajan, and Shangbo Mao. Multiscale graph texture network. In Computer Vision - ECCV 2024, pages 218-235, Cham, 2025. Springer Nature Switzerland. 1, 2, 3, 5, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 533, + 294, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 533, + 294, + 565 + ], + "spans": [ + { + "bbox": [ + 57, + 533, + 294, + 565 + ], + "type": "text", + "content": "[11] Shin Fujieda, Kohei Takayama, and Toshiya Hachisuka. Wavelet convolutional neural networks for texture classification. arXiv preprint arXiv:1707.07394, 2017. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 567, + 294, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 567, + 294, + 609 + ], + "spans": [ + { + "bbox": [ + 57, + 567, + 294, + 609 + ], + "type": "text", + "content": "[12] Robert M. Haralick, K. Shanmugam, and Its'Hak Dinstein. Textural features for image classification. IEEE Transactions on Systems, Man, and Cybernetics, SMC-3(6):610-621, 1973. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 612, + 294, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 612, + 294, + 655 + ], + "spans": [ + { + "bbox": [ + 57, + 612, + 294, + 655 + ], + "type": "text", + "content": "[13] Yuting Hu, Zhiling Long, and Ghassan AlRegib. Multi-level texture encoding and representation (multer) based on deep neural networks. In 2019 IEEE International Conference on Image Processing (ICIP), pages 4410-4414, 2019. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 657, + 294, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 657, + 294, + 678 + ], + "spans": [ + { + "bbox": [ + 57, + 657, + 294, + 678 + ], + "type": "text", + "content": "[14] Béla Julesz. Visual pattern discrimination. IRE transactions on Information Theory, 8(2):84-92, 1962. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 680, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 680, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 680, + 294, + 713 + ], + "type": "text", + "content": "[15] Hervé Jégou, Matthijs Douze, Cordelia Schmid, and Patrick Pérez. Aggregating local descriptors into a compact image representation. In 2010 IEEE Computer Society Conference" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 74, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 335, + 74, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 74, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 74, + 553, + 95 + ], + "type": "text", + "content": "on Computer Vision and Pattern Recognition, pages 3304-3311, 2010. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 96, + 553, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 96, + 553, + 117 + ], + "spans": [ + { + "bbox": [ + 317, + 96, + 553, + 117 + ], + "type": "text", + "content": "[16] Kenneth I Laws. Texture energy measures. In Proceedings of Image Understanding Workshop, 1979. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 118, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 118, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 317, + 118, + 553, + 150 + ], + "type": "text", + "content": "[17] Kenneth I. Laws. Rapid Texture Identification. In Image Processing for Missile Guidance, pages 376 - 381. International Society for Optics and Photonics, SPIE, 1980. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 152, + 553, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 152, + 553, + 194 + ], + "spans": [ + { + "bbox": [ + 317, + 152, + 553, + 194 + ], + "type": "text", + "content": "[18] Thomas Leung and Jitendra Malik. Representing and recognizing the visual appearance of materials using three-dimensional textons. International journal of computer vision, 43:29-44, 2001. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 195, + 553, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 195, + 553, + 247 + ], + "spans": [ + { + "bbox": [ + 317, + 195, + 553, + 247 + ], + "type": "text", + "content": "[19] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11966-11976, 2022. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 250, + 553, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 250, + 553, + 293 + ], + "spans": [ + { + "bbox": [ + 316, + 250, + 553, + 293 + ], + "type": "text", + "content": "[20] Ilya Loshchilov and Frank Hutter. SGDR: stochastic gradient descent with warm restarts. In 5th International Conference on Learning Representations, ICLR 2017, Toulouse, France, 2017. 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 294, + 553, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 294, + 553, + 336 + ], + "spans": [ + { + "bbox": [ + 316, + 294, + 553, + 336 + ], + "type": "text", + "content": "[21] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In 7th International Conference on Learning Representations, ICLR 2019, New Orleans, LA, USA, 2019. 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 338, + 553, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 338, + 553, + 370 + ], + "spans": [ + { + "bbox": [ + 316, + 338, + 553, + 370 + ], + "type": "text", + "content": "[22] Shangbo Mao, Deepu Rajan, and Liang Tien Chia. Deep residual pooling network for texture recognition. Pattern Recognition, 112:107817, 2021. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 372, + 553, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 372, + 553, + 403 + ], + "spans": [ + { + "bbox": [ + 316, + 372, + 553, + 403 + ], + "type": "text", + "content": "[23] Joshua Peeples, Weihuang Xu, and Alina Zare. Histogram layers for texture analysis. IEEE Transactions on Artificial Intelligence, 3(4):541-552, 2022. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 405, + 553, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 405, + 553, + 449 + ], + "spans": [ + { + "bbox": [ + 317, + 405, + 553, + 449 + ], + "type": "text", + "content": "[24] Florent Perronnin and Christopher Dance. Fisher kernels on visual vocabularies for image categorization. In 2007 IEEE Conference on Computer Vision and Pattern Recognition, pages 1-8, 2007. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 449, + 553, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 449, + 553, + 502 + ], + "spans": [ + { + "bbox": [ + 317, + 449, + 553, + 502 + ], + "type": "text", + "content": "[25] Michael J. Reale, Matthew Ochrym, Micah Church, Nicholas Goutermout, John Rubado, and Maria Cornacchia. Shape and texture aware graph processing. In 2020 IEEE Applied Imagery Pattern Recognition Workshop (AIPR), pages 1-18, 2020. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 504, + 553, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 504, + 553, + 525 + ], + "spans": [ + { + "bbox": [ + 317, + 504, + 553, + 525 + ], + "type": "text", + "content": "[26] Theodore J Rivlin. Chebyshev polynomials. Courier Dover Publications, 2020. 5" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 317, + 526, + 553, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 526, + 553, + 580 + ], + "spans": [ + { + "bbox": [ + 317, + 526, + 553, + 580 + ], + "type": "text", + "content": "[27] Leonardo Scabini, Kallil M. Zielinski, Lucas C. Ribas, Wesley N. Gonçalves, Bernard De Baets, and Odemir M. Bruno. Radam: Texture recognition through randomized aggregated encoding of deep activation maps. Pattern Recognition, 143: 109802, 2023. 2, 7" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 317, + 582, + 553, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 582, + 553, + 613 + ], + "spans": [ + { + "bbox": [ + 317, + 582, + 553, + 613 + ], + "type": "text", + "content": "[28] Lavanya Sharan, Ruth Rosenholtz, and Edward Adelson. Material perception: What can you see in a brief glance? Journal of Vision, 9(8):784-784, 2009. 7" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 317, + 615, + 553, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 615, + 553, + 658 + ], + "spans": [ + { + "bbox": [ + 317, + 615, + 553, + 658 + ], + "type": "text", + "content": "[29] Kaiyou Song, Hua Yang, and Zhouping Yin. Multi-scale boosting feature encoding network for texture recognition. IEEE Transactions on Circuits and Systems for Video Technology, 31(11):4269-4282, 2021. 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 317, + 659, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 659, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 317, + 659, + 553, + 712 + ], + "type": "text", + "content": "[30] Yang Song, Fan Zhang, Qing Li, Heng Huang, Lauren J. O'Donnell, and Weidong Cai. Locally-transferred fisher vectors for texture classification. In 2017 IEEE International Conference on Computer Vision (ICCV), pages 4922-4930, 2017. 2" + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "23431" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 552 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "text", + "content": "[31] Manik Varma and Andrew Zisserman. Classifying images of materials: Achieving viewpoint and illumination independence. In Computer Vision — ECCV 2002, pages 255–271, Berlin, Heidelberg, 2002. Springer Berlin Heidelberg. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 295, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 295, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 295, + 150 + ], + "type": "text", + "content": "[32] Manik Varma and Andrew Zisserman. A statistical approach to texture classification from single images. International journal of computer vision, 62:61-81, 2005. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 294, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 294, + 205 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 294, + 205 + ], + "type": "text", + "content": "[33] Yong Xu, Feng Li, Zhile Chen, Jinxiu Liang, and Yuhui Quan. Encoding spatial distribution of convolutional features for texture representation. In Advances in Neural Information Processing Systems, pages 22732-22744. Curran Associates, Inc., 2021. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 208, + 294, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 208, + 294, + 251 + ], + "spans": [ + { + "bbox": [ + 56, + 208, + 294, + 251 + ], + "type": "text", + "content": "[34] Jia Xue, Hang Zhang, Kristin Dana, and Ko Nishino. Differential angular imaging for material recognition. In 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 6940-6949, 2017. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 253, + 294, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 253, + 294, + 296 + ], + "spans": [ + { + "bbox": [ + 56, + 253, + 294, + 296 + ], + "type": "text", + "content": "[35] Jia Xue, Hang Zhang, and Kristin Dana. Deep texture manifold for ground terrain recognition. In 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 558-567, 2018. 1, 2, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 297, + 294, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 297, + 294, + 351 + ], + "spans": [ + { + "bbox": [ + 56, + 297, + 294, + 351 + ], + "type": "text", + "content": "[36] Zhijing Yang, Shujian Lai, Xiaobin Hong, Yukai Shi, Yongqiang Cheng, and Chunmei Qing. Dfaen: Double-order knowledge fusion and attentional encoding network for texture recognition. Expert Systems with Applications, 209: 118223, 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 353, + 294, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 353, + 294, + 397 + ], + "spans": [ + { + "bbox": [ + 56, + 353, + 294, + 397 + ], + "type": "text", + "content": "[37] Wei Zhai, Yang Cao, Jing Zhang, and Zheng-Jun Zha. Deep multiple-attribute-perceived network for real-world texture recognition. In 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 3612–3621, 2019. 1, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 399, + 294, + 452 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 399, + 294, + 452 + ], + "spans": [ + { + "bbox": [ + 56, + 399, + 294, + 452 + ], + "type": "text", + "content": "[38] Wei Zhai, Yang Cao, Zheng-Jun Zha, HaiYong Xie, and Feng Wu. Deep structure-revealed network for texture recognition. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11007-11016, 2020. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 454, + 294, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 454, + 294, + 508 + ], + "spans": [ + { + "bbox": [ + 56, + 454, + 294, + 508 + ], + "type": "text", + "content": "[39] Wei Zhai, Yang Cao, Jing Zhang, Haiyong Xie, Dacheng Tao, and Zheng-Jun Zha. On exploring multiplicity of primitives and attributes for texture recognition in the wild. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(1):403-420, 2024. 1, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 510, + 294, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 510, + 294, + 552 + ], + "spans": [ + { + "bbox": [ + 56, + 510, + 294, + 552 + ], + "type": "text", + "content": "[40] Hang Zhang, Jia Xue, and Kristin Dana. Deep ten: Texture encoding network. In 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2896-2905, 2017. 1, 2" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "23432" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/CheckManual_ A New Challenge and Benchmark for Manual-based Appliance Manipulation/c1aa26f1-f8e2-45ab-92dc-640acaf534e5_content_list.json b/2025/CheckManual_ A New Challenge and Benchmark for Manual-based Appliance Manipulation/c1aa26f1-f8e2-45ab-92dc-640acaf534e5_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..a2bb3f556343f07b74eea91786c2a07e7392c04f --- /dev/null +++ b/2025/CheckManual_ A New Challenge and Benchmark for Manual-based Appliance Manipulation/c1aa26f1-f8e2-45ab-92dc-640acaf534e5_content_list.json @@ -0,0 +1,1462 @@ +[ + { + "type": "text", + "text": "CheckManual: A New Challenge and Benchmark for Manual-based Appliance Manipulation", + "text_level": 1, + "bbox": [ + 230, + 130, + 769, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yuxing Long $^{1,2}$ , Jiyao Zhang $^{1,2}$ , Mingjie Pan $^{1,2}$ , Tianshu Wu $^{1}$ , Taewhan Kim $^{1,2}$ and Hao Dong $^{1,2\\dagger}$ $^{1}$ CFCS, School of Computer Science, Peking University $^{2}$ PKU-Agibot Lab", + "bbox": [ + 117, + 202, + 877, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 273, + 326, + 289 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Correct use of electrical appliances has significantly improved human life quality. Unlike simple tools that can be manipulated with common sense, different parts of electrical appliances have specific functions defined by manufacturers. If we want the robot to heat bread by microwave, we should enable them to review the microwave's manual first. From the manual, it can learn about component functions, interaction methods, and representative task steps about appliances. However, previous manual-related works remain limited to question-answering tasks while existing manipulation researchers ignore the manual's important role and fail to comprehend multi-page manuals. In this paper, we propose the first manual-based appliance manipulation benchmark CheckManual. Specifically, we design a large model-assisted human-revised data generation pipeline to create manuals based on CAD appliance models. With these manuals, we establish novel manual-based manipulation challenges, metrics, and simulator environments for model performance evaluation. Furthermore, we propose the first manual-based manipulation planning model ManualPlan to set up a group of baselines for the CheckManual benchmark. Our project page is available at https://sites.google.com/view/checkmanual.", + "bbox": [ + 89, + 306, + 483, + 654 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 676, + 220, + 691 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Since the twentieth century, the emergence of appliances has significantly changed human society. In daily life, we extensively use various electrical appliances to assist our living and working. In recent years, robots have demonstrated enormous potential and value in areas like component assembly [21], goods sorting [7], and indoor navigation [19]. Imagine if we could instruct robots to operate appliances to serve us, such as baking cakes with ovens, washing clothes with washing machines, and making juice with blenders, we would be further liberated from tedious household chores and have more free time.", + "bbox": [ + 89, + 703, + 482, + 867 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "However, appliances are different from simple tools like hammers or teapots that can be correctly used based on", + "bbox": [ + 89, + 869, + 483, + 900 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ab7c76859152a155849ee39e93795a72e549867bbcb954f049c26fb6c36d452e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 519, + 273, + 901, + 392 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e058bbca83b265920b4e7038e9411d4e7269fc9a798759b7a4b842c4acbe7641.jpg", + "image_caption": [ + "Figure 1. Comparison between appliance manipulation with and without manual. If the manual is absent, the concrete functions of the appliance part are unknown. As shown in the top image, the robot feels confused about tasks such as opening the microwave door. To correctly use the appliance, the robot must comprehend the manual content and follow its guidance to work." + ], + "image_footnote": [], + "bbox": [ + 521, + 398, + 901, + 522 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "common sense. One appliance may have multiple movable components and each component has specific functions even if they are the same type. Taking a microwave as an example, one button may control the temperature while the other button may set the time. Such function information is difficult to infer according to the appliance's visual appearance or human common sense. This time, we need to check the user manuals to learn about component functions and supporting tasks. Existing manipulation models like [4][3][12][10][11] already can adopt simple tools to do easy tasks but cannot read multi-page manuals not to say using the appliances to complete long-horizon complex tasks. Besides, previous manual-related research and academic datasets [23][32] are all collected by the NLP researchers1, which only focuses on page retrieval and question-answering tasks rather than manipulation. These", + "bbox": [ + 511, + 621, + 906, + 864 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "1Note that the copyright of real manual belongs to the company. Using content from the manual for academic research requires authorization.", + "bbox": [ + 511, + 875, + 906, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "22595", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "academic datasets manual datasets also do not provide appliance CAD models and cannot be used for quantitative manipulation evaluation in the simulator. Therefore, there are no manual-based manipulation evaluation datasets and benchmarks, nor manipulation models capable of following manuals to use appliances at present.", + "bbox": [ + 89, + 90, + 480, + 181 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address this problem, we propose the first manual-based manipulation benchmark CheckManual. In the benchmark, we collect new evaluation data, design novel manipulation challenges, and propose new baseline models. Specifically, we first design an LLM-assisted human-revised manual creation pipeline based on appliance CAD models. For every CAD model, this pipeline can follow our surveyed real manual analysis to generate diverse appliance manuals with different part functions, manipulation demonstrations, and page layouts. Then, we set up three manual-based appliance manipulation challenge tracks including manipulation planning, CAD-assisted manipulation, and CAD-free manipulation. They progressively cover the key steps in manual-based appliance manipulation, comprehensively evaluating the model's ability to operate appliances with manuals. Furthermore, we propose a manual-based manipulation planning method ManualPlan, and create a set of CAD-assisted primitive actions for appliance manipulation. The ManualPlan aligns manuals with appliances and plans detailed manipulation steps according to manual content, which can control primitive actions or open-vocabulary manipulation models to physically interact with appliances. These also act as the baseline methods for our CheckManual benchmark.", + "bbox": [ + 91, + 183, + 483, + 542 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, our main contributions are:", + "bbox": [ + 109, + 544, + 380, + 558 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We collect the first CAD model-aligned appliance manual dataset CheckManual through our devised large model-assisted pipeline. Every data collection stage undergoes careful human verification to guarantee correctness.", + "- We design novel challenges for manual-based appliance manipulation considering different difficulty levels and conditions. Furthermore, we establish the evaluation environment in the simulator for the benchmark.", + "- We propose the first manual-based manipulation planning model ManualPlan to set up a group of baselines for three-level challenge tracks in CheckManual." + ], + "bbox": [ + 89, + 559, + 482, + 724 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 731, + 232, + 746 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Manipulation Large Models", + "text_level": 1, + "bbox": [ + 89, + 756, + 341, + 772 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Large models pre-trained on Internet-scale text and image data emerge powerful capabilities for task planning and common sense reasoning. Therefore, leveraging large models has become a hotspot in embodied areas such as autonomous driving [29], navigation [20] [5] and manipulation. According to training and inference forms, existing manipulation methods can be divided into two main categories. The first category of work focuses on fine-tuning", + "bbox": [ + 89, + 780, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "large models with manipulation trajectories collected from the simulator environment and real-world. The representative models are Google RT-series from RT-1 [4], RT-2 [3], and RT-Trajectory [8] to STAR-RT [14]. These models take vision observation, robot states, and task instruction as input, and directly predict the low-level actions by multimodal large models. Similar open-sourced work includes ManipLLM [16], RoboFlamingo [15] and OpenVLA [12]. The other category of work makes efforts to effectively integrate large models' zero-shot inferences into the manipulation planning and decision process. SayCan [2] first utilizes a multimodal large model to score a group of kills to make progress toward the high-level task. Voxposer [10] drives a multimodal large model to create 3D value maps to guide the robot toward the target object part. CoPa [9] and ReKep [11] leverage multimodal large models to generate spatial-temporal constraints for complex tasks like \"insert the flower into vase\" or \"pour tea into water\". However, all these large model-based methods can only make plans based on their internal common sense. Considering heating the bread with the microwave, they even struggle to figure out the temperature knob and timer knob, not to say complete the whole task.", + "bbox": [ + 511, + 90, + 906, + 438 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Appliance Manual-based Tasks", + "text_level": 1, + "bbox": [ + 511, + 445, + 790, + 460 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "NLP researchers have recognized the manual's importance in addressing user questions. TechQA [24] collects 1400 user questions from online forums and annotates the corresponding answers from IBM technical documents to set up Document-based Question Answering task. S10 QA [23] manually creates 904 QA pairs based on the manual of the Samsung S10 phone. Nandy et al. [23] also collect 950 QA pairs based on the Samsung Smart TV manual to generate Smart TV/Remote QA dataset. These two datasets are utilized to establish a Manual-based Question Answering benchmark. The above datasets are all pure text corpus. PM209 [32] collects 209 product manuals from 27 well-known consumer electronic brands to set up Manual Page Retrieval task and Manual-based Multimodal Question Answering task. Although these datasets and tasks are related to the appliance manual, they are all designed for NLP evaluation and lack corresponding appliance CAD models. Therefore, manual-based manipulation benchmarks cannot be directly established on these existing related researches. We create the first manual & CAD model-aligned dataset and design novel benchmarks and baseline models for manual-based appliance manipulation.", + "bbox": [ + 511, + 468, + 908, + 801 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Manual-based Manipulation Benchmark CheckManual", + "text_level": 1, + "bbox": [ + 511, + 811, + 906, + 845 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this part, we will introduce the creation process of our CheckManual dataset in Section 3.1, analyze the numerical and visual statistics about CheckManual in Section 3.2, and", + "bbox": [ + 511, + 854, + 906, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "22596", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/5843d26d66f77e8664a28ca752bc3a37daa64329e2e35a744b0fbdebf840fd3b.jpg", + "image_caption": [ + "Figure 2. Generation workflow of CheckManual dataset. In the leftmost part (Section 3.1.1), we analyze real manuals to learn about their formats and collect different categories of appliance CAD models. The middle part (Section 3.1.2 - 3.1.4) demonstrates the creation of manual materials, including appliance creation, task generation, and figure design. Human verifies every step to guarantee correctness. Based on this information, the rightmost part (Section 3.1.5) generates appliance manuals with diverse formats through the LATEX codes." + ], + "image_footnote": [], + "bbox": [ + 94, + 88, + 225, + 369 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/7c4218757fdc77571c474a4c8935b8677945450d644d481960667ee100db0ae5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 236, + 88, + 436, + 369 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/869c06a7cf586df7fd7e134fb4bae8043a6dfde432fa60dcb7c390ce1cb3f97a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 446, + 88, + 694, + 369 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/c83d961fe54a2f97aa3cffdb04305b6350b706e86fcd95f59e983e896182dfc2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 694, + 88, + 903, + 369 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "define the evaluation for robot manipulation based on these manuals in Section 3.3.", + "bbox": [ + 89, + 450, + 482, + 479 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Appliance Manual Creation", + "text_level": 1, + "bbox": [ + 89, + 491, + 341, + 508 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To create an open-source available manual dataset, We design a large model-driven human-verified manual generation workflow based on appliance CAD models. Figure 2 demonstrates the generation workflow of the CheckManual dataset, which will be detailed in the following.", + "bbox": [ + 89, + 515, + 483, + 590 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1.1 Preparation", + "text_level": 1, + "bbox": [ + 89, + 609, + 228, + 626 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Before formally starting manual creation, we analyze a stack of real appliance manuals to learn about real manual formats and collect available articulated appliance CAD models for manual creation.", + "bbox": [ + 89, + 635, + 482, + 694 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Analysis of Real Appliance Manuals We collect 110 appliance manual PDFs from the Internet, which cover 11 categories of our dataset involved appliances. These manuals come from different countries and regions, including China, the European Union, Japan, and India. Then we manually analyze each manual from Functional Part Annotation Type, Guidance of Functional Part Operation, and Task Expression Format these three aspects. We find that the functional parts are usually explained in the point-line annotation formats. These lines connect the points on appliance parts and the function names or IDs of parts as shown in the middle of Figure 2. The guidance of functional", + "bbox": [ + 89, + 696, + 482, + 876 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "part operation is commonly represented in formats: 1 Pure text description like \"Rotate temperature knob $90^{\\circ}$ clockwise\". 2 Close-up figures of the manipulated part, such as the close and open states of the red microwave door in Figure 2. 3 Part motion decomposition diagram that shows the dynamic change of part during its motion process. Figure 2 shows one example of opening the white mixer lid. 4 Motion trajectory of the contact point on the part, like manipulating Figure 2's blue mixer slider from right to left. Besides, the manipulation tasks in the real manuals are often depicted in the text, bullet lists, ordered lists, and tables with one-column or multi-columns.", + "bbox": [ + 511, + 450, + 906, + 631 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Appliance CAD Models Our appliance CAD models come from the PartNet-Mobility [22] Dataset, which contains common categories of household appliances CAD models, including microwaves, dishwashers, and refrigerators. Appliance models in this dataset are annotated with manipulable parts, such as buttons, knobs, levers, sliders, and so on. In total, we collect 11 categories, 182 CAD models, which cover widely-used household appliances in our daily lives.", + "bbox": [ + 511, + 632, + 906, + 753 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1.2 Appliance Creation", + "text_level": 1, + "bbox": [ + 511, + 771, + 702, + 787 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "With the appliance CAD models, we can create concrete appliances by annotating each movable part with a specific function name (e.g., temperature knob) and corresponding function states (e.g., clockwise $60^{\\circ} - 120^{\\circ}C$ and clockwise $120^{\\circ} - 160^{\\circ}C$ ). This process is detailed in the following.", + "bbox": [ + 511, + 795, + 905, + 869 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Part Function Name Annotation Considering part function names are closely related to the appliance layout and", + "bbox": [ + 511, + 871, + 906, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "2We use examples from our ChcekManual to avoid any infringement", + "bbox": [ + 107, + 886, + 472, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "22597", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/d17051db036ef41e2b5673d328e01e347ba027b862a46bf0a159ede68e60c464.jpg", + "image_caption": [ + "Figure 3. One example of appliance manual from our CheckManual dataset. Please zoom in for details." + ], + "image_footnote": [], + "bbox": [ + 94, + 88, + 370, + 236 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/7afe3484708466aa39eea7c029df3af9dd4d5644a2736d57522c0e53dbef9dd6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 370, + 88, + 500, + 236 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/e07905b3275cf7d380266c9f671d0c865a5e19637e9cd9e9f761b8019bd3424f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 88, + 632, + 236 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/fc3fd8f20dbb3d1acc1dc20175535813ceb0a062a3253480a4733017957f5b36.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 632, + 89, + 903, + 236 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "parts' visual features, we drive the annotation process by the multimodal large language model (MLLM). Before annotation, manipulable part IDs and types are automatically labeled on the appliance overview figures through point-line annotation. Taking these figures as visual prompts, we instruct the MLLM to design the function name of each part. The multimodal large model should consider the appliance type and carefully analyze part visual information to create reasonable annotations. Re-generation mechanism would be triggered when the number of annotated parts is unequal to the total number of manipulable parts. By resolving the large model's prediction with regular expressions, we can obtain a dictionary recording the one-to-one relationships between part ID and corresponding function name. One such dictionary can be used to generate one appliance. We can create multiple appliances from one CAD model by repeating the above annotation process.", + "bbox": [ + 88, + 282, + 480, + 541 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Part Function State Annotation To annotate part function states, we first sample available states for each manipulable part according to their types. For buttons, we randomly sample available push times in a pre-defined range. For knobs and levers, we randomly sample several rotation degrees at intervals of $60^{\\circ}$ . Other manipulable parts like doors, lids, and containers, are assigned with \"open\" and \"close\" states. With part function names and available states, we instruct the large language model (LLM) to design concrete functions for each state. This way, each manipulable part on the appliance will be annotated with multiple function states.", + "bbox": [ + 89, + 547, + 482, + 727 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Human Verification about Part Functions Although large models are powerful annotation tools, they still make intolerable mistakes and thus require human verification. In the annotation about function names, MLLM sometimes fails to identify tiny visual elements on the CAD models like already-drawn button logos or scales, which may lead to mismatching between annotated function names and part visual appearance. In the annotation about function states, LLM may create an anti-commonsense design such as “ $500^{\\circ}\\mathrm{C}$ for oven temperature knob”. Therefore, we check large model-annotated function names and states", + "bbox": [ + 89, + 734, + 482, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "through our designed human verification website (See Appendix) to revise incorrect cases. In this stage, we manually verify more than 2200 part function annotations.", + "bbox": [ + 511, + 282, + 903, + 329 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1.3 Manipulation Task Creation", + "text_level": 1, + "bbox": [ + 511, + 348, + 764, + 363 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The manual often includes several representative tasks to introduce how to use the appliance. These tasks are also the core of our manual-based manipulation evaluation. We can create appliance-executable tasks based on their part function names and states.", + "bbox": [ + 511, + 372, + 903, + 446 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Task Proposal With function names and states of movable parts on the appliances, we instruct the LLM to propose short-horizon and long-horizon manipulation tasks for each appliance. Every task step should specify the part function name and the manipulated state in a resolvable format like \"1. Front Door, Open\", \"2. Power Button, Push 1 time\" ... To improve task quality and reduce repetition, the LLM only needs to design one task in one inference based on already created tasks.", + "bbox": [ + 511, + 448, + 905, + 583 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Human Verification about Manipulation Tasks In the task creation, LLM may also make anti-commonsense mistakes. To guarantee absolute correctness for evaluation, after the LLM completes task generation, we manually check every task step to adjust unreasonable step orders and revise incorrect steps. If the task is completely unreasonable, we will manually create a new one from scratch. Totally, we checked over 1400 manipulation tasks created by LLM.", + "bbox": [ + 511, + 584, + 905, + 705 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1.4 Manual Figure Design", + "text_level": 1, + "bbox": [ + 511, + 724, + 723, + 739 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "According to our analysis of real manuals, cover, control panels, and part manipulation guidance are the main figures in the manuals. Therefore, we devise methods to create these figures.", + "bbox": [ + 511, + 750, + 903, + 809 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Manual Cover The manual cover usually displays appliance appearance and typical usage scenes. we predefined a set of candidate camera poses to capture the appliance from different views. The views with the largest total areas of manipulable parts are selected to render RGB appliance overview figures. To synthesize custom usage scenes,", + "bbox": [ + 511, + 810, + 905, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "22598", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/b31e291063126d2190157b947b17f00848e2f16b36744699b0aa9ef97037bd52.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 94, + 90, + 289, + 172 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/6fbc62139ead53f34aeace96a753265895645e6888c749ab00a53b4c6113443c.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 290, + 88, + 459, + 180 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/b0341611a065ee7f7785289612ccb86d6a3827862164f77bd8d73424fb62af5c.jpg", + "image_caption": [ + "(c)", + "Figure 4. CheckManual dataset statistics. The above tables and figures from left to right are (a) Numerical statistics. (b) Proportion of different appliances. (c) Distribution of manipulable components on the appliances. (d) Distribution of manipulation task step lengths." + ], + "image_footnote": [], + "bbox": [ + 468, + 87, + 696, + 190 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/7412f40b543c4f6474d9a28f2d6cadbb643f557324d738371c190d292359b948.jpg", + "image_caption": [ + "(d)" + ], + "image_footnote": [], + "bbox": [ + 707, + 90, + 901, + 190 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "we further add the indoor background to the appliance overview through text-to-image generation model SDXL-Lightning [17]. The background room type is chosen based on the appliance category. For example, the oven is usually put into the kitchen while the washing machine is often arranged in the laundry room.", + "bbox": [ + 88, + 258, + 482, + 349 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Overview Diagram & Control Panel The overview diagram and control panel demonstrate the function name of every movable part on the appliance. Before annotation, we first apply the DBSCAN [6] algorithm to try clustering buttons and knobs based on their geometric centers. If clustered buttons and knobs are found, the areas with them are selected as control panels and cropped from the appliance overview figures as close-ups of control panels. Then, we adopt point-line annotation methods (Figure 2) from the real manuals to draw overview diagram and control panel based on part function names created in the previous stage. Furthermore, considering real manuals often deliver overview diagrams and control panels in line drawings, we create sketch-style figures by conducting edge detection on the RGB normal maps. This way, we can generate these two kinds of figures both in RGB style and sketch style.", + "bbox": [ + 91, + 351, + 483, + 593 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Part Manipulation Guidance This kind of figure usually introduces the physical interaction approach for appliance manipulable parts. To generate such figures, we first render manipulable parts according to their types and visible surface areas. For buttons, knobs, levers, and drawers, we render them with the camera axis parallel to their joint axes. Other manipulable parts, including doors, sliders, lids, and containers, are rendered with the camera axis perpendicular to their joint axes while ensuring maximum visible areas. Then, we draw manipulation guidance on these figures to display how to manipulate them. Following common representation in the real manuals from Section 3.1.1, we adopted these four different strategies to draw manipulation guidance as shown in Figure 2. To realize this, we manually annotate the 3D position of one contact point for every manipulable part on all appliance CAD models and calculate its motion trajectory according to the part-linked joint axis.", + "bbox": [ + 91, + 595, + 483, + 853 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Human Verification about Manual Figures When collecting appliance overviews based on the largest visible part areas, the view may not be suitable to observe every part.", + "bbox": [ + 89, + 854, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Besides, sketch-style figures may contain missing lines or abnormal shadows. Therefore, we conduct human verification on more than 1800 figures to trigger regeneration when these errors occur.", + "bbox": [ + 511, + 258, + 906, + 318 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/69ff07721b735d6418dc23756290ba87fcb640c7e5e9194274e5bfe80c1deacd.jpg", + "table_caption": [ + "Table 1. Comparison of different manual-based manipulation challenge tracks in our CheckManual benchmark." + ], + "table_footnote": [], + "table_body": "
AVAILABLE INPUTPREDICTION TARGET
Task InstructionRGB-DManual PDFAppliance CAD ModelHigh-level PlanningLow-level Manipulation
Track 1: Manual-CAD-Appliance Aligned Planing
✓✓✓✓✓
Track 2: Manual & CAD based Manipulation
✓✓✓✓✓
Track 3: Pure Manual-based Manipulation
✓✓✓✓
", + "bbox": [ + 516, + 369, + 906, + 453 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1.5 LATEX-based Manual Generation", + "text_level": 1, + "bbox": [ + 511, + 468, + 794, + 484 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As we know, $\\mathrm{LATEX}$ is a document programming language that offers both the advantages of standardization and flexibility. We can not only keep the same document style from page to page but also easily insert different types of figures and tables into the document by $\\mathrm{LATEX}$ coding. Recently, large language models emerged with powerful capabilities for code generation. Therefore, we decided to leverage LLM to generate the $\\mathrm{LATEX}$ codes for the manual and then compile them into a PDF format appliance manual as shown in the rightmost part of Figure 2.", + "bbox": [ + 509, + 491, + 906, + 643 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Before starting this process, we pre-defined a set of widely used $\\mathrm{LATEX}$ packages (e.g., multirow, xcolor, caption) and prepared a collection of commonly used icons (e.g., fragility reminder, temperature warning, environment protection). To increase diversity, we randomly sample different formats of covers, diagrams, and guidance demonstrations from the Section 3.1.4 as figure resources and randomly select a text style from real manuals (e.g., text, bullet list, table) for task description. In the generation, we synthesize the prompt with textual part function annotation, manipulation tasks, and above image paths to require the large model to continue one section of $\\mathrm{LATEX}$ codes based on the previous. To prevent uncompilable codes, we design a section-granularity compilation mechanism. The newly generated section of $\\mathrm{LATEX}$ codes is appended to the end of the previous correct code and compiled together. Regeneration will be triggered when errors occur in the code compi", + "bbox": [ + 511, + 643, + 908, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "22599", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "lation. Figure 3 displays one manual from our CheckManual dataset.", + "bbox": [ + 89, + 90, + 480, + 119 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2. Dataset Statistic Analysis", + "text_level": 1, + "bbox": [ + 89, + 132, + 321, + 148 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We totally created 1107 manuals for 369 appliances, which cover 11 common appliance categories in daily life, including washing machines, coffee makers, dishwashers, microwaves, and so on. These appliances contain 12 types, with 2211 different manipulable parts. On average, there are 6.00 manipulable parts on each appliance. In terms of quantity and diversity, parts contained in our dataset can support most appliance operations. Besides, our dataset includes 1464 manipulation tasks in total. Every task is accompanied by step-by-step manipulation parts and target states. These tasks include both single-step and long-horizon tasks with a maximum of 18 steps. The table and charts in Figure 4 display comprehensive statistics about our CheckManual dataset.", + "bbox": [ + 89, + 155, + 482, + 367 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.3. CheckManual Challenge Tracks", + "text_level": 1, + "bbox": [ + 89, + 377, + 375, + 393 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "With the collected manual data, we design challenge tracks for manual-based manipulation. The basic setting is given a task instruction $T$ and a corresponding appliance PDF manual $M$ , the robot should follow the manual to complete the tabletop appliance manipulation task with RGB-D observation $I_{c}$ and $I_{d}$ . Based on this basic setting, we design the challenge tracks as shown in Table 1 and the following.", + "bbox": [ + 89, + 401, + 482, + 507 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Track 1: CAD-Appliance Aligned Planing We first consider an ideal setting where the appliance CAD models are available and CAD-Appliance is aligned. The \"CAD-Appliance Aligned\" means part area and ID are accessible on the camera observation. In this case, the evaluation focuses on testing the robot's capability of analyzing appliance manuals and making manipulation plans based on the given task instruction and manual content. The evaluation is composed of two aspects. Firstly the robot should align all function names in the manuals with appliance parts. Secondly, the task planning should include correct manipulation target part and primitive actions in every task step. This challenge track is free from manipulation execution.", + "bbox": [ + 89, + 508, + 482, + 704 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Track 2: Manual & CAD based Manipulation In this task, the appliance CAD models are available but the mapping relationships among parts on manual, CAD model, and observation are not known. This setting is reasonable because it is easy for manufacturers to provide plain CAD models for their appliance products. For evaluation, the robot should follow the given task instruction to execute low-level manipulation with a two-finger gripper based on manual content and CAD model. To complete this challenge, the robot needs to first estimate the 6-DoF pose of observed objects before using the CAD model. After that, the robot should also align parts on the manual, CAD model, and current observation. After completing the part align", + "bbox": [ + 89, + 704, + 482, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ment, the robot can obtain part-linked joint position and rotation axis from the CAD model, which will help the robot to make manipulation planning and execution.", + "bbox": [ + 511, + 90, + 903, + 136 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Track 3: Pure Manual-based Manipulation We further consider the setting that aligns with real-world applications where only appliance manuals are available to the robot. The challenges of this setting come from two aspects. On the high level, the robot should analyze the appliance manual to learn about the function information (i.e., name and possible states) and manipulation methods of every movable part on the appliances. Then, the robot should plan the detailed manipulation steps for the given task based on the manual content and common sense. On the low level, the robot must execute the high-level plan step-by-step with its end-effector to complete the instruction-specified task. During the whole process, the robot can only rely on its RGB-D observation and manuals instead of any CAD-model assistance. For both Track 2 and Track 3, any procedural error during the operation will cause the entire task to fail.", + "bbox": [ + 511, + 137, + 906, + 377 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metrics To evaluate model performances on the above challenge tracks, we define the task-level and step-level metrics. On the step level, we evaluate the alignment success rate and planning success rate for Track 1 and task execution success rate and average task completion rate for Track 2 and 3. The alignment success is recognized when all CAD model parts are linked to the correct function names mentioned in the manual. The task planning success is confirmed when part IDs and manipulation actions in every planned step are all correct. The task execution is regarded as a success when all task steps are completed in the correct order. The task completion rate is calculated by \"Success Steps / # Task Steps\". Note that only steps before the failure step can be counted as successful steps. We take the average values of the above metrics as final results. On the step level, we check the state of the target part after each manipulation step to judge the success. Specifically, we define the whole movement limitation range of the target joint as $L$ . For revolute parts like knobs, levers, and doors, the acceptable error range is defined as $L \\pm 30^{\\circ}$ . For prismatic parts like the button, container, and slider, the successful manipulation is confirmed when the button moves $>25\\%$ $L$ or $MD_{part} / MD_{appliance}$ $>50\\%$ if target part and appliances move in the same direction, where $MD$ represented motion distance.", + "bbox": [ + 511, + 378, + 906, + 755 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Manual-based Manipulation Baselines", + "text_level": 1, + "bbox": [ + 511, + 768, + 857, + 787 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we design baseline models for the three tracks of the benchmark. For track 1, we propose the ManualPlan model, which focuses on manual-based task planning, as detailed in Section 4.1. For tasks involving robotic execution, corresponding to track 2 and track 3, we extend the ManualPlan model by integrating it as a high-level planning module with existing manipulation models, as dis", + "bbox": [ + 511, + 795, + 903, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "22600", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c405025e32b47094d4e7ed3ee1040f42db8a1d1b3ed5bb66a8de5012992e7047.jpg", + "image_caption": [ + "Figure 5. The framework of our ManualPlan model. It is composed of Manual Resolution, Manipulation Planing and Part Alignment modules. The ManualPlan can make high level planning to control the CAD-assisted primitive actions or open-vocabulary manipulation large model (e.g., VoxPoser [10]) to use the appliance, which serves as the baseline models for CheckManual benchmark." + ], + "image_footnote": [], + "bbox": [ + 96, + 88, + 903, + 218 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "cussed in Section 4.2.", + "bbox": [ + 89, + 291, + 236, + 305 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.1. ManualPlan for High-Level Task Planning", + "text_level": 1, + "bbox": [ + 89, + 316, + 452, + 333 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ManualPlan introduces a novel framework for high-level task planning of robotic manipulation with appliance manuals. It processes manual content, generates detailed plans, and aligns parts in the real appliance with those referenced in the manual.", + "bbox": [ + 89, + 340, + 483, + 415 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Manual Resolution Module. The appliance manual in our dataset is in PDF format, which is consistent with the real manuals. The PDF file cannot be directly processed by large language models. Therefore, we divide the manual PDF file into page-level images first. Then, we extract text content from these page images using the Paddle OCR model. We also utilize the multimodal large language model (MLLM) to detect visual elements like figures and tables on each page image. In this way, we can obtain the page-granularity text and vision information about the whole manual.", + "bbox": [ + 89, + 416, + 483, + 566 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Manipulation Planning Module. With extracted manual information, we call LLM to plan a step-by-step appliance manipulation process for the given task. The large language model must specify the part function name and manipulated states in the dictionary format like “{'Brew Strength Lever': 'Rotate 60 degrees'}” for every planning step. Only the manual-mentioned parts can be used in the plan. Considering just a part of the tasks are demonstrated in the manual, the LLM should make plans based on both common sense about appliance usage and the manual content. Any unresolvable prediction will trigger the regeneration request.", + "bbox": [ + 89, + 568, + 482, + 734 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Part Alignment Module. To achieve this goal, we incorporate a visual prompting mechanism known as Set-of-Mark (SoM) [28] to align parts in the field of view and parts depicted in the manual. SoM leverages segmentation models to partition an image into distinct regions and assign a numeric marker to each, which can significantly improve the visual grounding capabilities of multimodal large language models. Here, we utilize the open-vocabulary object detection model Grounding-DINO [18] to predict bounding boxes of movable parts and prompt Segment Anything [13] with these bounding boxes", + "bbox": [ + 89, + 734, + 483, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "to generate movable part masks. These masks are assigned different colors with transparency and overlayed on the observation image. The ID is marked at the geometric center of each part mask. Then, we provide the observation image and appliance diagrams from manual to MLLM requiring it to identify the mapping relationships from mask IDs to part function names. This way, we realize the part alignment between the observation and the manual. For the Manual & CAD Model-based Manipulation task, although the CAD model is available, the part function names still need to be aligned with parts on the CAD model. Therefore, we use a similar SoM-based approach to achieve such alignment with MLLM. For the other two tasks, the CAD model is unavailable for the Pure Manual-based Manipulation task while part alignment is already achieved in the Manual-CAD-Appliance Aligned Manipulation task.", + "bbox": [ + 511, + 291, + 906, + 547 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. ManualPlan-Driven Low-Level Manipulation", + "text_level": 1, + "bbox": [ + 511, + 574, + 897, + 590 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "When executing manipulation on the appliance, we employ ManualPlan as a high-level planner and design different manipulation policies for Track 2 (CAD Model Available) and Track 3 (CAD Model Unavailable).", + "bbox": [ + 511, + 598, + 905, + 657 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "CAD-assisted Manipulation Policy. When the CAD model is available, we can estimate the 6-DoF pose of the observed appliance in the camera frame by the pose estimation models like [25][31][30]. With such an estimated object pose, the observed appliance can be aligned with its CAD model and all joints' positions and axis directions on the observed appliance become available. Therefore, we design primitive actions for different appliance parts to synthesize CAD-assisted manipulation policy. For door and parts with slider (e.g., screen, slider, lid), we provide the 3D bounding box of the target part to constrain the AnyGrasp [7] model to predict part-level grasp poses. Then, the gripper poses in the motion sequences can be calculated based on the information about the joint and manipulation method. For the button part, we take its surface geometrical center as the contact point and control the closed gripper", + "bbox": [ + 511, + 657, + 906, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "22601", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/d16d2c8378ad67acbc542077cd9ac705f56619a634aec404e228ab613f06867b.jpg", + "table_caption": [ + "Table 2. Baseline model's performances on three challenge tracks of CheckManual benchmark. The results of Track 1 are reported in the \"Part Alignment Success Rate / Task Planning Success Rate\" format. For Track 2 and 3, their results are reported in \"Task Completion Rate / Task Success Rate\" format. The \"w/o manual\" means ablating appliance manuals input for the above methods." + ], + "table_footnote": [], + "table_body": "
BASELINE MODEL图图图图图图图图图图图Total
Track 1: Manual-CAD-Appliance Aligned Planing
ManualPlan w/o manual22.96 / 22.22 - / 4.2139.47 / 22.82 - / 2.6818.52 / 14.29 - / 20.0072.84 / 60.19 - / 48.154.17 / 3.91 - / 0.788.13 / 8.64 - / 2.4794.44 / 68.00 - / 36.0047.22 / 30.56 - / 6.9441.67 / 40.62 - / 15.6221.15 / 14.39 - / 4.3215.46 / 9.96 - / 2.8527.32 / 20.70 - / 7.99
Track 2: Manual & CAD based Manipulation
ManualPlan + Primitive Actions w/o manual2.39 / 1.15 1.05 / 0.387.12 / 4.70 1.76 / 1.3413.63 / 11.43 21.39 / 14.290.46 / 0.00 0.46 / 0.000.78 / 0.78 0.78 / 0.787.91 / 3.70 3.50 / 1.2310.00 / 4.00 2.00 / 0.0011.11 / 4.86 5.09 / 1.393.12 / 0.00 1.04 / 0.002.32 / 1.44 2.28 / 2.162.56 / 1.42 1.69 / 1.424.50 / 2.39 2.47 / 1.37
Track 3: Pure Manual-based Manipulation
ManualPlan + VoxPoser [10] w/o manual6.11 / 1.92 0.99 / 0.387.18 / 4.70 1.39 / 0.677.10 / 5.71 0.57 / 0.003.01 / 1.85 0.46 / 0.003.12 / 3.12 1.56 / 1.567.99 / 3.70 1.54 / 0.620.00 / 0.00 0.00 / 0.0010.66 / 6.25 2.50 / 0.696.75 / 3.12 0.00 / 0.000.48 / 0.00 0.00 / 0.002.96 / 2.14 0.98 / 0.715.18 / 2.87 1.11 / 0.55
", + "bbox": [ + 94, + 141, + 903, + 275 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "to push the button from the normal direction of the contact point. For the knob part, we control the open gripper to get close to the knob from the normal direction of its surface geometrical center and close the gripper to grasp. After the gripper successfully grasps the knob, it will rotate planned degrees around the normal of its surface geometrical center. CAD-free Manipulation Policy. We integrate a recently proposed open-vocabulary manipulation model VoxPoser [10] into our ManualPlan to complete the CAD-free manipulation task. In this framework, ManualPlan and VoxPoser cooperate like the cerebrum and cerebellum in the human brain. ManualPlan reads manuals, makes high-level planning, and aligns observed components with manual-depicted parts. Then, it grounds the target component and demonstrates the corresponding manipulation method in language format, which is transferred to VoxPoser with camera RGB-D observation to execute low-level manipulation action on the appliance.", + "bbox": [ + 88, + 291, + 485, + 564 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 89, + 577, + 223, + 593 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.1. Implementation Details", + "text_level": 1, + "bbox": [ + 89, + 601, + 307, + 617 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "When creating the CheckManual dataset, we utilize GPT-4o [1] as the large language model and multimodal large language model. Following previous work [26], we use the Franka Panda flying gripper in the simulation. The groundingDINO [18] model is utilized to ground the bounding boxes of appliance components. The predicted bounding boxes prompt SAM [13] model to generate component masks. FoundationPose [25] helps us estimate appliance 6-DoF pose in the camera observation. In the evaluation, manual comprehension and manipulation planning processes are assisted by GPT-4o [1] model.", + "bbox": [ + 89, + 625, + 483, + 791 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.2. Simulation Evaluation", + "text_level": 1, + "bbox": [ + 89, + 801, + 299, + 816 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We evaluate our proposed ManualPlan model and ManualPlan-driven framework using the CheckManual benchmark, which employs SAPIEN [27] as the simulation. This environment provides a realistic and flexible simulation for testing robotic manipulation tasks across a", + "bbox": [ + 89, + 824, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "variety of appliances. The experimental results on different tasks and appliance categories are summarized in Table 2. From the results of Track 1, we observe that the ManualPlan model is capable of generating a part of correct task plans based on multi-page manuals. However, even the most advanced Multi-Modal Large Language Models (MLLMs) still face significant challenges in aligning manual part annotations with the corresponding CAD models of appliances, which greatly limits the overall performance of ManualPlan in task planning. For Track 2 and Track 3, the performance of the ManualPlan-driven framework highlights the difficulties posed by long-horizon, manual-based manipulation tasks. After analyzing failure cases, we identified that the accumulation of errors in part identification, task planning, and multi-step execution are the primary causes of failure. Additionally, we conducted an ablation study by removing the manual (i.e., w/o manual) to investigate its impact on task performance. As shown in Table 2, the manual plays a crucial role in guiding the robot to accurately plan and execute manipulation tasks, underscoring its importance in completing appliance manipulation tasks.", + "bbox": [ + 511, + 291, + 906, + 609 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.3. Real Robot Deployment", + "text_level": 1, + "bbox": [ + 511, + 621, + 733, + 636 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To demonstrate the potential of our approach in real-world applications, we conducted real-world robot experiments following the task definition of Track 3 from the Check-Manual benchmark. The hardware setup involved a Franka Panda robot arm and a RealSense D415 RGB-D camera mounted on the end-effector. Considering error accumulation in the long-horizon execution, we follow [11] to cache correct manipulation actions in every step.", + "bbox": [ + 511, + 645, + 906, + 767 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 782, + 633, + 799 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we introduce CheckManual, the first benchmark for manual-based appliance manipulation, and define new challenges, metrics, and simulator environments for model evaluation. Additionally, we present the first manual-based manipulation planning model ManualPlan to set up a group of baselines for the CheckManual benchmark.", + "bbox": [ + 511, + 809, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "22602", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgment", + "text_level": 1, + "bbox": [ + 91, + 90, + 240, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported by the National Youth Talent Support Program (8200800081) and National Natural Science Foundation of China (No. 62376006). We also thank Yang Liu and Zhuoqun Xu from Polar Mirage for their technical support on assets.", + "bbox": [ + 89, + 113, + 483, + 184 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 195, + 187, + 210 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 8", + "[2] Michael Ahn, Anthony Brohan, Noah Brown, Yevgen Chebotar, Omar Cortes, Byron David, Chelsea Finn, Chuyuan Fu, Keerthana Gopalakrishnan, Karol Hausman, et al. Do as i can, not as i say: Grounding language in robotic affordances. arXiv preprint arXiv:2204.01691, 2022. 2", + "[3] Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Xi Chen, Krzysztof Choromanski, Tianli Ding, Danny Driess, Avinava Dubey, Chelsea Finn, Pete Florence, Chuyuan Fu, Montse Gonzalez Arenas, Keerthana Gopalakrishnan, Kehang Han, Karol Hausman, Alexander Herzog, Jasmine Hsu, Brian Ichter, Alex Irpan, Nikhil Joshi, Ryan Julian, Dmitry Kalashnikov, Yuheng Kuang, Isabel Leal, Lisa Lee, Tsang-Wei Edward Lee, Sergey Levine, Yao Lu, Henryk Michalewski, Igor Mordatch, Karl Pertsch, Kanishka Rao, Krista Reymann, Michael Ryoo, Grecia Salazar, Pannag Sanketi, Pierre Sermanet, Jaspiar Singh, Anikait Singh, Radu Soricut, Huong Tran, Vincent Vanhoucke, Quan Vuong, Ayzaan Wahid, Stefan Welker, Paul Wohlhart, Jialin Wu, Fei Xia, Ted Xiao, Peng Xu, Sichun Xu, Tianhe Yu, and Brianna Zitkovich. Rt-2: Vision-language-action models transfer web knowledge to robotic control, 2023. 1, 2", + "[4] Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Joseph Dabis, Chelsea Finn, Keerthana Gopalakrishnan, Karol Hausman, Alex Herzog, Jasmine Hsu, Julian Ibarz, Brian Ichter, Alex Irpan, Tomas Jackson, Sally Jesmonth, Nikhil J Joshi, Ryan Julian, Dmitry Kalashnikov, Yuheng Kuang, Isabel Leal, Kuang-Huei Lee, Sergey Levine, Yao Lu, Utsav Malla, Deeksha Manjunath, Igor Mordatch, Ofir Nachum, Carolina Parada, Jodilyn Peralta, Emily Perez, Karl Pertsch, Jornell Quiambao, Kanishka Rao, Michael Ryoo, Grecia Salazar, Pannag Sanketi, Kevin Sayed, Jaspiar Singh, Sumedh Sontakke, Austin Stone, Clayton Tan, Huong Tran, Vincent Vanhoucke, Steve Vega, Quan Vuong, Fei Xia, Ted Xiao, Peng Xu, Sichun Xu, Tianhe Yu, and Brianna Zitkovich. Rt-1: Robotics transformer for real-world control at scale, 2023. 1, 2", + "[5] Wenzhe Cai, Siyuan Huang, Guangran Cheng, Yuxing Long, Peng Gao, Changyin Sun, and Hao Dong. Bridging zero-shot object navigation and foundation models through pixel-guided navigation skill. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 5228-5234. IEEE, 2024. 2", + "[6] Martin Ester, Hans-Peter Kriegel, Jörg Sander, Xiaowei Xu, et al. A density-based algorithm for discovering clusters in" + ], + "bbox": [ + 99, + 220, + 483, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "large spatial databases with noise. In kdd, pages 226-231, 1996. 5", + "[7] Hao-Shu Fang, Chenxi Wang, Hongjie Fang, Minghao Gou, Jirong Liu, Hengxu Yan, Wenhai Liu, Yichen Xie, and Cewu Lu. Anygrasp: Robust and efficient grasp perception in spatial and temporal domains. IEEE Transactions on Robotics (T-RO), 2023. 1, 7", + "[8] Jiayuan Gu, Sean Kirmani, Paul Wohlhart, Yao Lu, Montserrat Gonzalez Arenas, Kanishka Rao, Wenhao Yu, Chuyuan Fu, Keerthana Gopalakrishnan, Zhuo Xu, Priya Sundaresan, Peng Xu, Hao Su, Karol Hausman, Chelsea Finn, Quan Vuong, and Ted Xiao. Rt-trajectory: Robotic task generalization via hindsight trajectory sketches, 2023. 2", + "[9] Haoxu Huang, Fanqi Lin, Yingdong Hu, Shengjie Wang, and Yang Gao. Copa: General robotic manipulation through spatial constraints of parts with foundation models. arXiv preprint arXiv:2403.08248, 2024. 2", + "[10] Wenlong Huang, Chen Wang, Ruohan Zhang, Yunzhu Li, Jiajun Wu, and Li Fei-Fei. Voxposer: Composable 3d value maps for robotic manipulation with language models. arXiv preprint arXiv:2307.05973, 2023. 1, 2, 7, 8", + "[11] Wenlong Huang, Chen Wang, Yunzhu Li, Ruohan Zhang, and Li Fei-Fei. Rekep: Spatio-temporal reasoning of relational keypoint constraints for robotic manipulation. arXiv preprint arXiv:2409.01652, 2024. 1, 2, 8", + "[12] Moo Jin Kim, Karl Pertsch, Siddharth Karamcheti, Ted Xiao, Ashwin Balakrishna, Suraj Nair, Rafael Rafailov, Ethan Foster, Grace Lam, Pannag Sanketi, Quan Vuong, Thomas Kollar, Benjamin Burchfiel, Russ Tedrake, Dorsa Sadigh, Sergey Levine, Percy Liang, and Chelsea Finn. Openvla: An open-source vision-language-action model. arXiv preprint arXiv:2406.09246, 2024. 1, 2", + "[13] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dollar, and Ross Girshick. Segment anything. arXiv:2304.02643, 2023. 7, 8", + "[14] Isabel Leal, Krzysztof Choromanski, Deepali Jain, Avinava Dubey, Jake Varley, Michael Ryoo, Yao Lu, Frederick Liu, Vikas Sindhwani, Quan Vuong, et al. Sara-rt: Scaling up robotics transformers with self-adaptive robust attention. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 6920–6927. IEEE, 2024. 2", + "[15] Xinghang Li, Minghuan Liu, Hanbo Zhang, Cunjun Yu, Jie Xu, Hongtao Wu, Chilam Cheang, Ya Jing, Weinan Zhang, Huaping Liu, et al. Vision-language foundation models as effective robot imitators. arXiv preprint arXiv:2311.01378, 2023. 2", + "[16] Xiaqi Li, Mingxu Zhang, Yiran Geng, Haoran Geng, Yuxing Long, Yan Shen, Renrui Zhang, Jiaming Liu, and Hao Dong. Maniplm: Embodied multimodal large language model for object-centric robotic manipulation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18061-18070, 2024. 2", + "[17] Shanchuan Lin, Anran Wang, and Xiao Yang. Sdxlightning: Progressive adversarial diffusion distillation, 2024. 5" + ], + "bbox": [ + 514, + 92, + 903, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "22603", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[18] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Qing Jiang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, and Lei Zhang. Grounding dino: Marrying dino with grounded pre-training for open-set object detection, 2024. 7, 8", + "[19] Yuxing Long, Wenzhe Cai, Hongcheng Wang, Guanqi Zhan, and Hao Dong. Instructnav: Zero-shot system for generic instruction navigation in unexplored environment. arXiv preprint arXiv:2406.04882, 2024. 1", + "[20] Yuxing Long, Xiaqi Li, Wenzhe Cai, and Hao Dong. Discuss before moving: Visual language navigation via multi-expert discussions. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 17380-17387. IEEE, 2024. 2", + "[21] Jianlan Luo, Charles Xu, Jeffrey Wu, and Sergey Levine. Hil-serl: Precise and dexterous robotic manipulation via human-in-the-loop reinforcement learning, 2024. 1", + "[22] Kaichun Mo, Shilin Zhu, Angel X. Chang, Li Yi, Subarna Tripathi, Leonidas J. Guibas, and Hao Su. PartNet: A large-scale benchmark for fine-grained and hierarchical part-level 3D object understanding. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 3", + "[23] Abhilash Nandy, Soumya Sharma, Shubham Maddhashiya, Kapil Sachdeva, Pawan Goyal, and NIloy Ganguly. Question answering over electronic devices: A new benchmark dataset and a multi-task learning based QA framework. In Findings of the Association for Computational Linguistics: EMNLP 2021, pages 4600-4609, Punta Cana, Dominican Republic, 2021. Association for Computational Linguistics. 1, 2", + "[24] Abhilash Nandy, Soumya Sharma, Shubham Maddhashiya, Kapil Sachdeva, Pawan Goyal, and NIloy Ganguly. Question answering over electronic devices: A new benchmark dataset and a multi-task learning based QA framework. In Findings of the Association for Computational Linguistics: EMNLP 2021, pages 4600-4609, Punta Cana, Dominican Republic, 2021. Association for Computational Linguistics. 2", + "[25] Bowen Wen, Wei Yang, Jan Kautz, and Stan Birchfield. FoundationPose: Unified 6d pose estimation and tracking of novel objects. In CVPR, 2024. 7, 8", + "[26] Ruihai Wu, Yan Zhao, Kaichun Mo, Zizheng Guo, Yian Wang, Tianhao Wu, Qingnan Fan, Xuelin Chen, Leonidas Guibas, and Hao Dong. VAT-mart: Learning visual action trajectory proposals for manipulating 3d ARTiculated objects. In International Conference on Learning Representations, 2022. 8", + "[27] Fanbo Xiang, Yuzhe Qin, Kaichun Mo, Yikuan Xia, Hao Zhu, Fangchen Liu, Minghua Liu, Hanxiao Jiang, Yifu Yuan, He Wang, Li Yi, Angel X. Chang, Leonidas J. Guibas, and Hao Su. SAPIEN: A simulated part-based interactive environment. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 8", + "[28] Jianwei Yang, Hao Zhang, Feng Li, Xueyan Zou, Chunyuan Li, and Jianfeng Gao. Set-of-mark prompting unleashes extraordinary visual grounding in gpt-4v. arXiv preprint arXiv:2310.11441, 2023. 7", + "[29] Senqiao Yang, Jiaming Liu, Ray Zhang, Mingjie Pan, Zoey Guo, Xiaqi Li, Zehui Chen, Peng Gao, Yandong Guo, and" + ], + "bbox": [ + 91, + 92, + 480, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Shanghang Zhang. Lidar-llm: Exploring the potential of large language models for 3d lidar understanding. arXiv preprint arXiv:2312.14074, 2023. 2", + "[30] Jiyao Zhang, Mingdong Wu, and Hao Dong. Generative category-level object pose estimation via diffusion models. Advances in Neural Information Processing Systems, 36: 54627-54644, 2023. 7", + "[31] Jiyao Zhang, Weiyao Huang, Bo Peng, Mingdong Wu, Fei Hu, Zijian Chen, Bo Zhao, and Hao Dong. Omni6dpose: A benchmark and model for universal 6d object pose estimation and tracking. In European Conference on Computer Vision, pages 199-216. Springer, 2024. 7", + "[32] Liang Zhang, Anwen Hu, Jing Zhang, Shuo Hu, and Qin Jin. Mpmqa: multimodal question answering on product manuals. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 13958-13966, 2023. 1, 2" + ], + "bbox": [ + 516, + 92, + 903, + 316 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "22604", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/CheckManual_ A New Challenge and Benchmark for Manual-based Appliance Manipulation/c1aa26f1-f8e2-45ab-92dc-640acaf534e5_model.json b/2025/CheckManual_ A New Challenge and Benchmark for Manual-based Appliance Manipulation/c1aa26f1-f8e2-45ab-92dc-640acaf534e5_model.json new file mode 100644 index 0000000000000000000000000000000000000000..a204eeb2a872957d6805e4754bdb094a77c73adb --- /dev/null +++ b/2025/CheckManual_ A New Challenge and Benchmark for Manual-based Appliance Manipulation/c1aa26f1-f8e2-45ab-92dc-640acaf534e5_model.json @@ -0,0 +1,1881 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.232, + 0.131, + 0.77, + 0.177 + ], + "angle": 0, + "content": "CheckManual: A New Challenge and Benchmark for Manual-based Appliance Manipulation" + }, + { + "type": "text", + "bbox": [ + 0.119, + 0.203, + 0.878, + 0.24 + ], + "angle": 0, + "content": "Yuxing Long\\(^{1,2}\\), Jiyao Zhang\\(^{1,2}\\), Mingjie Pan\\(^{1,2}\\), Tianshu Wu\\(^{1}\\), Taewhan Kim\\(^{1,2}\\) and Hao Dong\\(^{1,2\\dagger}\\) \n\\(^{1}\\)CFCS, School of Computer Science, Peking University \\(^{2}\\)PKU-Agibot Lab" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.274, + 0.327, + 0.29 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.307, + 0.485, + 0.655 + ], + "angle": 0, + "content": "Correct use of electrical appliances has significantly improved human life quality. Unlike simple tools that can be manipulated with common sense, different parts of electrical appliances have specific functions defined by manufacturers. If we want the robot to heat bread by microwave, we should enable them to review the microwave's manual first. From the manual, it can learn about component functions, interaction methods, and representative task steps about appliances. However, previous manual-related works remain limited to question-answering tasks while existing manipulation researchers ignore the manual's important role and fail to comprehend multi-page manuals. In this paper, we propose the first manual-based appliance manipulation benchmark CheckManual. Specifically, we design a large model-assisted human-revised data generation pipeline to create manuals based on CAD appliance models. With these manuals, we establish novel manual-based manipulation challenges, metrics, and simulator environments for model performance evaluation. Furthermore, we propose the first manual-based manipulation planning model ManualPlan to set up a group of baselines for the CheckManual benchmark. Our project page is available at https://sites.google.com/view/checkmanual." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.678, + 0.222, + 0.693 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.704, + 0.483, + 0.868 + ], + "angle": 0, + "content": "Since the twentieth century, the emergence of appliances has significantly changed human society. In daily life, we extensively use various electrical appliances to assist our living and working. In recent years, robots have demonstrated enormous potential and value in areas like component assembly [21], goods sorting [7], and indoor navigation [19]. Imagine if we could instruct robots to operate appliances to serve us, such as baking cakes with ovens, washing clothes with washing machines, and making juice with blenders, we would be further liberated from tedious household chores and have more free time." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.484, + 0.901 + ], + "angle": 0, + "content": "However, appliances are different from simple tools like hammers or teapots that can be correctly used based on" + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.275, + 0.903, + 0.393 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.399, + 0.903, + 0.523 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.535, + 0.907, + 0.619 + ], + "angle": 0, + "content": "Figure 1. Comparison between appliance manipulation with and without manual. If the manual is absent, the concrete functions of the appliance part are unknown. As shown in the top image, the robot feels confused about tasks such as opening the microwave door. To correctly use the appliance, the robot must comprehend the manual content and follow its guidance to work." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.622, + 0.907, + 0.865 + ], + "angle": 0, + "content": "common sense. One appliance may have multiple movable components and each component has specific functions even if they are the same type. Taking a microwave as an example, one button may control the temperature while the other button may set the time. Such function information is difficult to infer according to the appliance's visual appearance or human common sense. This time, we need to check the user manuals to learn about component functions and supporting tasks. Existing manipulation models like [4][3][12][10][11] already can adopt simple tools to do easy tasks but cannot read multi-page manuals not to say using the appliances to complete long-horizon complex tasks. Besides, previous manual-related research and academic datasets [23][32] are all collected by the NLP researchers1, which only focuses on page retrieval and question-answering tasks rather than manipulation. These" + }, + { + "type": "page_footnote", + "bbox": [ + 0.513, + 0.875, + 0.907, + 0.901 + ], + "angle": 0, + "content": "1Note that the copyright of real manual belongs to the company. Using content from the manual for academic research requires authorization." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "22595" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.182 + ], + "angle": 0, + "content": "academic datasets manual datasets also do not provide appliance CAD models and cannot be used for quantitative manipulation evaluation in the simulator. Therefore, there are no manual-based manipulation evaluation datasets and benchmarks, nor manipulation models capable of following manuals to use appliances at present." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.184, + 0.485, + 0.543 + ], + "angle": 0, + "content": "To address this problem, we propose the first manual-based manipulation benchmark CheckManual. In the benchmark, we collect new evaluation data, design novel manipulation challenges, and propose new baseline models. Specifically, we first design an LLM-assisted human-revised manual creation pipeline based on appliance CAD models. For every CAD model, this pipeline can follow our surveyed real manual analysis to generate diverse appliance manuals with different part functions, manipulation demonstrations, and page layouts. Then, we set up three manual-based appliance manipulation challenge tracks including manipulation planning, CAD-assisted manipulation, and CAD-free manipulation. They progressively cover the key steps in manual-based appliance manipulation, comprehensively evaluating the model's ability to operate appliances with manuals. Furthermore, we propose a manual-based manipulation planning method ManualPlan, and create a set of CAD-assisted primitive actions for appliance manipulation. The ManualPlan aligns manuals with appliances and plans detailed manipulation steps according to manual content, which can control primitive actions or open-vocabulary manipulation models to physically interact with appliances. These also act as the baseline methods for our CheckManual benchmark." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.545, + 0.382, + 0.559 + ], + "angle": 0, + "content": "In this paper, our main contributions are:" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.56, + 0.483, + 0.619 + ], + "angle": 0, + "content": "- We collect the first CAD model-aligned appliance manual dataset CheckManual through our devised large model-assisted pipeline. Every data collection stage undergoes careful human verification to guarantee correctness." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.62, + 0.483, + 0.679 + ], + "angle": 0, + "content": "- We design novel challenges for manual-based appliance manipulation considering different difficulty levels and conditions. Furthermore, we establish the evaluation environment in the simulator for the benchmark." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.681, + 0.483, + 0.725 + ], + "angle": 0, + "content": "- We propose the first manual-based manipulation planning model ManualPlan to set up a group of baselines for three-level challenge tracks in CheckManual." + }, + { + "type": "list", + "bbox": [ + 0.09, + 0.56, + 0.483, + 0.725 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.732, + 0.233, + 0.747 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.757, + 0.343, + 0.773 + ], + "angle": 0, + "content": "2.1. Manipulation Large Models" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Large models pre-trained on Internet-scale text and image data emerge powerful capabilities for task planning and common sense reasoning. Therefore, leveraging large models has become a hotspot in embodied areas such as autonomous driving [29], navigation [20] [5] and manipulation. According to training and inference forms, existing manipulation methods can be divided into two main categories. The first category of work focuses on fine-tuning" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.908, + 0.439 + ], + "angle": 0, + "content": "large models with manipulation trajectories collected from the simulator environment and real-world. The representative models are Google RT-series from RT-1 [4], RT-2 [3], and RT-Trajectory [8] to STAR-RT [14]. These models take vision observation, robot states, and task instruction as input, and directly predict the low-level actions by multimodal large models. Similar open-sourced work includes ManipLLM [16], RoboFlamingo [15] and OpenVLA [12]. The other category of work makes efforts to effectively integrate large models' zero-shot inferences into the manipulation planning and decision process. SayCan [2] first utilizes a multimodal large model to score a group of kills to make progress toward the high-level task. Voxposer [10] drives a multimodal large model to create 3D value maps to guide the robot toward the target object part. CoPa [9] and ReKep [11] leverage multimodal large models to generate spatial-temporal constraints for complex tasks like \"insert the flower into vase\" or \"pour tea into water\". However, all these large model-based methods can only make plans based on their internal common sense. Considering heating the bread with the microwave, they even struggle to figure out the temperature knob and timer knob, not to say complete the whole task." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.446, + 0.792, + 0.462 + ], + "angle": 0, + "content": "2.2. Appliance Manual-based Tasks" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.469, + 0.909, + 0.803 + ], + "angle": 0, + "content": "NLP researchers have recognized the manual's importance in addressing user questions. TechQA [24] collects 1400 user questions from online forums and annotates the corresponding answers from IBM technical documents to set up Document-based Question Answering task. S10 QA [23] manually creates 904 QA pairs based on the manual of the Samsung S10 phone. Nandy et al. [23] also collect 950 QA pairs based on the Samsung Smart TV manual to generate Smart TV/Remote QA dataset. These two datasets are utilized to establish a Manual-based Question Answering benchmark. The above datasets are all pure text corpus. PM209 [32] collects 209 product manuals from 27 well-known consumer electronic brands to set up Manual Page Retrieval task and Manual-based Multimodal Question Answering task. Although these datasets and tasks are related to the appliance manual, they are all designed for NLP evaluation and lack corresponding appliance CAD models. Therefore, manual-based manipulation benchmarks cannot be directly established on these existing related researches. We create the first manual & CAD model-aligned dataset and design novel benchmarks and baseline models for manual-based appliance manipulation." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.812, + 0.907, + 0.846 + ], + "angle": 0, + "content": "3. Manual-based Manipulation Benchmark CheckManual" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.908, + 0.901 + ], + "angle": 0, + "content": "In this part, we will introduce the creation process of our CheckManual dataset in Section 3.1, analyze the numerical and visual statistics about CheckManual in Section 3.2, and" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22596" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.089, + 0.226, + 0.371 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.237, + 0.089, + 0.437, + 0.371 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.447, + 0.089, + 0.695, + 0.371 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.695, + 0.089, + 0.905, + 0.371 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.381, + 0.907, + 0.439 + ], + "angle": 0, + "content": "Figure 2. Generation workflow of CheckManual dataset. In the leftmost part (Section 3.1.1), we analyze real manuals to learn about their formats and collect different categories of appliance CAD models. The middle part (Section 3.1.2 - 3.1.4) demonstrates the creation of manual materials, including appliance creation, task generation, and figure design. Human verifies every step to guarantee correctness. Based on this information, the rightmost part (Section 3.1.5) generates appliance manuals with diverse formats through the LATEX codes." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.452, + 0.483, + 0.481 + ], + "angle": 0, + "content": "define the evaluation for robot manipulation based on these manuals in Section 3.3." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.492, + 0.342, + 0.509 + ], + "angle": 0, + "content": "3.1. Appliance Manual Creation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.516, + 0.484, + 0.592 + ], + "angle": 0, + "content": "To create an open-source available manual dataset, We design a large model-driven human-verified manual generation workflow based on appliance CAD models. Figure 2 demonstrates the generation workflow of the CheckManual dataset, which will be detailed in the following." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.611, + 0.23, + 0.627 + ], + "angle": 0, + "content": "3.1.1 Preparation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.636, + 0.483, + 0.695 + ], + "angle": 0, + "content": "Before formally starting manual creation, we analyze a stack of real appliance manuals to learn about real manual formats and collect available articulated appliance CAD models for manual creation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.697, + 0.483, + 0.877 + ], + "angle": 0, + "content": "Analysis of Real Appliance Manuals We collect 110 appliance manual PDFs from the Internet, which cover 11 categories of our dataset involved appliances. These manuals come from different countries and regions, including China, the European Union, Japan, and India. Then we manually analyze each manual from Functional Part Annotation Type, Guidance of Functional Part Operation, and Task Expression Format these three aspects. We find that the functional parts are usually explained in the point-line annotation formats. These lines connect the points on appliance parts and the function names or IDs of parts as shown in the middle of Figure 2. The guidance of functional" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.452, + 0.907, + 0.632 + ], + "angle": 0, + "content": "part operation is commonly represented in formats: 1 Pure text description like \"Rotate temperature knob \\(90^{\\circ}\\) clockwise\". 2 Close-up figures of the manipulated part, such as the close and open states of the red microwave door in Figure 2. 3 Part motion decomposition diagram that shows the dynamic change of part during its motion process. Figure 2 shows one example of opening the white mixer lid. 4 Motion trajectory of the contact point on the part, like manipulating Figure 2's blue mixer slider from right to left. Besides, the manipulation tasks in the real manuals are often depicted in the text, bullet lists, ordered lists, and tables with one-column or multi-columns." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.633, + 0.907, + 0.754 + ], + "angle": 0, + "content": "Appliance CAD Models Our appliance CAD models come from the PartNet-Mobility [22] Dataset, which contains common categories of household appliances CAD models, including microwaves, dishwashers, and refrigerators. Appliance models in this dataset are annotated with manipulable parts, such as buttons, knobs, levers, sliders, and so on. In total, we collect 11 categories, 182 CAD models, which cover widely-used household appliances in our daily lives." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.772, + 0.704, + 0.788 + ], + "angle": 0, + "content": "3.1.2 Appliance Creation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.906, + 0.871 + ], + "angle": 0, + "content": "With the appliance CAD models, we can create concrete appliances by annotating each movable part with a specific function name (e.g., temperature knob) and corresponding function states (e.g., clockwise \\(60^{\\circ} - 120^{\\circ}C\\) and clockwise \\(120^{\\circ} - 160^{\\circ}C\\)). This process is detailed in the following." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.872, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Part Function Name Annotation Considering part function names are closely related to the appliance layout and" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.887, + 0.473, + 0.901 + ], + "angle": 0, + "content": "2We use examples from our ChcekManual to avoid any infringement" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22597" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.089, + 0.371, + 0.237 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.371, + 0.089, + 0.5, + 0.237 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.5, + 0.089, + 0.633, + 0.237 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.633, + 0.09, + 0.905, + 0.237 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.178, + 0.254, + 0.816, + 0.267 + ], + "angle": 0, + "content": "Figure 3. One example of appliance manual from our CheckManual dataset. Please zoom in for details." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.284, + 0.482, + 0.542 + ], + "angle": 0, + "content": "parts' visual features, we drive the annotation process by the multimodal large language model (MLLM). Before annotation, manipulable part IDs and types are automatically labeled on the appliance overview figures through point-line annotation. Taking these figures as visual prompts, we instruct the MLLM to design the function name of each part. The multimodal large model should consider the appliance type and carefully analyze part visual information to create reasonable annotations. Re-generation mechanism would be triggered when the number of annotated parts is unequal to the total number of manipulable parts. By resolving the large model's prediction with regular expressions, we can obtain a dictionary recording the one-to-one relationships between part ID and corresponding function name. One such dictionary can be used to generate one appliance. We can create multiple appliances from one CAD model by repeating the above annotation process." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.548, + 0.483, + 0.728 + ], + "angle": 0, + "content": "Part Function State Annotation To annotate part function states, we first sample available states for each manipulable part according to their types. For buttons, we randomly sample available push times in a pre-defined range. For knobs and levers, we randomly sample several rotation degrees at intervals of \\(60^{\\circ}\\). Other manipulable parts like doors, lids, and containers, are assigned with \"open\" and \"close\" states. With part function names and available states, we instruct the large language model (LLM) to design concrete functions for each state. This way, each manipulable part on the appliance will be annotated with multiple function states." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Human Verification about Part Functions Although large models are powerful annotation tools, they still make intolerable mistakes and thus require human verification. In the annotation about function names, MLLM sometimes fails to identify tiny visual elements on the CAD models like already-drawn button logos or scales, which may lead to mismatching between annotated function names and part visual appearance. In the annotation about function states, LLM may create an anti-commonsense design such as “\\(500^{\\circ}\\mathrm{C}\\) for oven temperature knob”. Therefore, we check large model-annotated function names and states" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.284, + 0.905, + 0.33 + ], + "angle": 0, + "content": "through our designed human verification website (See Appendix) to revise incorrect cases. In this stage, we manually verify more than 2200 part function annotations." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.349, + 0.766, + 0.364 + ], + "angle": 0, + "content": "3.1.3 Manipulation Task Creation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.373, + 0.905, + 0.447 + ], + "angle": 0, + "content": "The manual often includes several representative tasks to introduce how to use the appliance. These tasks are also the core of our manual-based manipulation evaluation. We can create appliance-executable tasks based on their part function names and states." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.449, + 0.906, + 0.584 + ], + "angle": 0, + "content": "Task Proposal With function names and states of movable parts on the appliances, we instruct the LLM to propose short-horizon and long-horizon manipulation tasks for each appliance. Every task step should specify the part function name and the manipulated state in a resolvable format like \"1. Front Door, Open\", \"2. Power Button, Push 1 time\" ... To improve task quality and reduce repetition, the LLM only needs to design one task in one inference based on already created tasks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.585, + 0.906, + 0.706 + ], + "angle": 0, + "content": "Human Verification about Manipulation Tasks In the task creation, LLM may also make anti-commonsense mistakes. To guarantee absolute correctness for evaluation, after the LLM completes task generation, we manually check every task step to adjust unreasonable step orders and revise incorrect steps. If the task is completely unreasonable, we will manually create a new one from scratch. Totally, we checked over 1400 manipulation tasks created by LLM." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.726, + 0.724, + 0.74 + ], + "angle": 0, + "content": "3.1.4 Manual Figure Design" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.905, + 0.81 + ], + "angle": 0, + "content": "According to our analysis of real manuals, cover, control panels, and part manipulation guidance are the main figures in the manuals. Therefore, we devise methods to create these figures." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Manual Cover The manual cover usually displays appliance appearance and typical usage scenes. we predefined a set of candidate camera poses to capture the appliance from different views. The views with the largest total areas of manipulable parts are selected to render RGB appliance overview figures. To synthesize custom usage scenes," + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "22598" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.092, + 0.29, + 0.174 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.169, + 0.193, + 0.186, + 0.205 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.291, + 0.089, + 0.46, + 0.181 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.371, + 0.193, + 0.388, + 0.205 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.47, + 0.088, + 0.697, + 0.191 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.581, + 0.193, + 0.596, + 0.205 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image", + "bbox": [ + 0.708, + 0.091, + 0.902, + 0.191 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.801, + 0.193, + 0.819, + 0.205 + ], + "angle": 0, + "content": "(d)" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.215, + 0.908, + 0.245 + ], + "angle": 0, + "content": "Figure 4. CheckManual dataset statistics. The above tables and figures from left to right are (a) Numerical statistics. (b) Proportion of different appliances. (c) Distribution of manipulable components on the appliances. (d) Distribution of manipulation task step lengths." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.259, + 0.483, + 0.35 + ], + "angle": 0, + "content": "we further add the indoor background to the appliance overview through text-to-image generation model SDXL-Lightning [17]. The background room type is chosen based on the appliance category. For example, the oven is usually put into the kitchen while the washing machine is often arranged in the laundry room." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.352, + 0.485, + 0.594 + ], + "angle": 0, + "content": "Overview Diagram & Control Panel The overview diagram and control panel demonstrate the function name of every movable part on the appliance. Before annotation, we first apply the DBSCAN [6] algorithm to try clustering buttons and knobs based on their geometric centers. If clustered buttons and knobs are found, the areas with them are selected as control panels and cropped from the appliance overview figures as close-ups of control panels. Then, we adopt point-line annotation methods (Figure 2) from the real manuals to draw overview diagram and control panel based on part function names created in the previous stage. Furthermore, considering real manuals often deliver overview diagrams and control panels in line drawings, we create sketch-style figures by conducting edge detection on the RGB normal maps. This way, we can generate these two kinds of figures both in RGB style and sketch style." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.597, + 0.484, + 0.854 + ], + "angle": 0, + "content": "Part Manipulation Guidance This kind of figure usually introduces the physical interaction approach for appliance manipulable parts. To generate such figures, we first render manipulable parts according to their types and visible surface areas. For buttons, knobs, levers, and drawers, we render them with the camera axis parallel to their joint axes. Other manipulable parts, including doors, sliders, lids, and containers, are rendered with the camera axis perpendicular to their joint axes while ensuring maximum visible areas. Then, we draw manipulation guidance on these figures to display how to manipulate them. Following common representation in the real manuals from Section 3.1.1, we adopted these four different strategies to draw manipulation guidance as shown in Figure 2. To realize this, we manually annotate the 3D position of one contact point for every manipulable part on all appliance CAD models and calculate its motion trajectory according to the part-linked joint axis." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Human Verification about Manual Figures When collecting appliance overviews based on the largest visible part areas, the view may not be suitable to observe every part." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.259, + 0.907, + 0.319 + ], + "angle": 0, + "content": "Besides, sketch-style figures may contain missing lines or abnormal shadows. Therefore, we conduct human verification on more than 1800 figures to trigger regeneration when these errors occur." + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.331, + 0.907, + 0.36 + ], + "angle": 0, + "content": "Table 1. Comparison of different manual-based manipulation challenge tracks in our CheckManual benchmark." + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.37, + 0.907, + 0.454 + ], + "angle": 0, + "content": "
AVAILABLE INPUTPREDICTION TARGET
Task InstructionRGB-DManual PDFAppliance CAD ModelHigh-level PlanningLow-level Manipulation
Track 1: Manual-CAD-Appliance Aligned Planing
✓✓✓✓✓
Track 2: Manual & CAD based Manipulation
✓✓✓✓✓
Track 3: Pure Manual-based Manipulation
✓✓✓✓
" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.469, + 0.795, + 0.485 + ], + "angle": 0, + "content": "3.1.5 LATEX-based Manual Generation" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.492, + 0.907, + 0.644 + ], + "angle": 0, + "content": "As we know, \\(\\mathrm{LATEX}\\) is a document programming language that offers both the advantages of standardization and flexibility. We can not only keep the same document style from page to page but also easily insert different types of figures and tables into the document by \\(\\mathrm{LATEX}\\) coding. Recently, large language models emerged with powerful capabilities for code generation. Therefore, we decided to leverage LLM to generate the \\(\\mathrm{LATEX}\\) codes for the manual and then compile them into a PDF format appliance manual as shown in the rightmost part of Figure 2." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.645, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Before starting this process, we pre-defined a set of widely used \\(\\mathrm{LATEX}\\) packages (e.g., multirow, xcolor, caption) and prepared a collection of commonly used icons (e.g., fragility reminder, temperature warning, environment protection). To increase diversity, we randomly sample different formats of covers, diagrams, and guidance demonstrations from the Section 3.1.4 as figure resources and randomly select a text style from real manuals (e.g., text, bullet list, table) for task description. In the generation, we synthesize the prompt with textual part function annotation, manipulation tasks, and above image paths to require the large model to continue one section of \\(\\mathrm{LATEX}\\) codes based on the previous. To prevent uncompilable codes, we design a section-granularity compilation mechanism. The newly generated section of \\(\\mathrm{LATEX}\\) codes is appended to the end of the previous correct code and compiled together. Regeneration will be triggered when errors occur in the code compi" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22599" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.121 + ], + "angle": 0, + "content": "lation. Figure 3 displays one manual from our CheckManual dataset." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.133, + 0.322, + 0.149 + ], + "angle": 0, + "content": "3.2. Dataset Statistic Analysis" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.156, + 0.483, + 0.368 + ], + "angle": 0, + "content": "We totally created 1107 manuals for 369 appliances, which cover 11 common appliance categories in daily life, including washing machines, coffee makers, dishwashers, microwaves, and so on. These appliances contain 12 types, with 2211 different manipulable parts. On average, there are 6.00 manipulable parts on each appliance. In terms of quantity and diversity, parts contained in our dataset can support most appliance operations. Besides, our dataset includes 1464 manipulation tasks in total. Every task is accompanied by step-by-step manipulation parts and target states. These tasks include both single-step and long-horizon tasks with a maximum of 18 steps. The table and charts in Figure 4 display comprehensive statistics about our CheckManual dataset." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.378, + 0.376, + 0.394 + ], + "angle": 0, + "content": "3.3. CheckManual Challenge Tracks" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.402, + 0.483, + 0.508 + ], + "angle": 0, + "content": "With the collected manual data, we design challenge tracks for manual-based manipulation. The basic setting is given a task instruction \\( T \\) and a corresponding appliance PDF manual \\( M \\), the robot should follow the manual to complete the tabletop appliance manipulation task with RGB-D observation \\( I_{c} \\) and \\( I_{d} \\). Based on this basic setting, we design the challenge tracks as shown in Table 1 and the following." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.509, + 0.483, + 0.705 + ], + "angle": 0, + "content": "Track 1: CAD-Appliance Aligned Planing We first consider an ideal setting where the appliance CAD models are available and CAD-Appliance is aligned. The \"CAD-Appliance Aligned\" means part area and ID are accessible on the camera observation. In this case, the evaluation focuses on testing the robot's capability of analyzing appliance manuals and making manipulation plans based on the given task instruction and manual content. The evaluation is composed of two aspects. Firstly the robot should align all function names in the manuals with appliance parts. Secondly, the task planning should include correct manipulation target part and primitive actions in every task step. This challenge track is free from manipulation execution." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Track 2: Manual & CAD based Manipulation In this task, the appliance CAD models are available but the mapping relationships among parts on manual, CAD model, and observation are not known. This setting is reasonable because it is easy for manufacturers to provide plain CAD models for their appliance products. For evaluation, the robot should follow the given task instruction to execute low-level manipulation with a two-finger gripper based on manual content and CAD model. To complete this challenge, the robot needs to first estimate the 6-DoF pose of observed objects before using the CAD model. After that, the robot should also align parts on the manual, CAD model, and current observation. After completing the part align" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.137 + ], + "angle": 0, + "content": "ment, the robot can obtain part-linked joint position and rotation axis from the CAD model, which will help the robot to make manipulation planning and execution." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.138, + 0.907, + 0.378 + ], + "angle": 0, + "content": "Track 3: Pure Manual-based Manipulation We further consider the setting that aligns with real-world applications where only appliance manuals are available to the robot. The challenges of this setting come from two aspects. On the high level, the robot should analyze the appliance manual to learn about the function information (i.e., name and possible states) and manipulation methods of every movable part on the appliances. Then, the robot should plan the detailed manipulation steps for the given task based on the manual content and common sense. On the low level, the robot must execute the high-level plan step-by-step with its end-effector to complete the instruction-specified task. During the whole process, the robot can only rely on its RGB-D observation and manuals instead of any CAD-model assistance. For both Track 2 and Track 3, any procedural error during the operation will cause the entire task to fail." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.379, + 0.908, + 0.756 + ], + "angle": 0, + "content": "Evaluation Metrics To evaluate model performances on the above challenge tracks, we define the task-level and step-level metrics. On the step level, we evaluate the alignment success rate and planning success rate for Track 1 and task execution success rate and average task completion rate for Track 2 and 3. The alignment success is recognized when all CAD model parts are linked to the correct function names mentioned in the manual. The task planning success is confirmed when part IDs and manipulation actions in every planned step are all correct. The task execution is regarded as a success when all task steps are completed in the correct order. The task completion rate is calculated by \"Success Steps / # Task Steps\". Note that only steps before the failure step can be counted as successful steps. We take the average values of the above metrics as final results. On the step level, we check the state of the target part after each manipulation step to judge the success. Specifically, we define the whole movement limitation range of the target joint as \\( L \\). For revolute parts like knobs, levers, and doors, the acceptable error range is defined as \\( L \\pm 30^{\\circ} \\). For prismatic parts like the button, container, and slider, the successful manipulation is confirmed when the button moves \\( >25\\% \\) \\( L \\) or \\( MD_{part} / MD_{appliance} \\) \\( >50\\% \\) if target part and appliances move in the same direction, where \\( MD \\) represented motion distance." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.77, + 0.859, + 0.788 + ], + "angle": 0, + "content": "4. Manual-based Manipulation Baselines" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.905, + 0.903 + ], + "angle": 0, + "content": "In this section, we design baseline models for the three tracks of the benchmark. For track 1, we propose the ManualPlan model, which focuses on manual-based task planning, as detailed in Section 4.1. For tasks involving robotic execution, corresponding to track 2 and track 3, we extend the ManualPlan model by integrating it as a high-level planning module with existing manipulation models, as dis" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.521, + 0.957 + ], + "angle": 0, + "content": "22600" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.098, + 0.089, + 0.904, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.233, + 0.908, + 0.278 + ], + "angle": 0, + "content": "Figure 5. The framework of our ManualPlan model. It is composed of Manual Resolution, Manipulation Planing and Part Alignment modules. The ManualPlan can make high level planning to control the CAD-assisted primitive actions or open-vocabulary manipulation large model (e.g., VoxPoser [10]) to use the appliance, which serves as the baseline models for CheckManual benchmark." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.292, + 0.238, + 0.306 + ], + "angle": 0, + "content": "cussed in Section 4.2." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.318, + 0.453, + 0.334 + ], + "angle": 0, + "content": "4.1. ManualPlan for High-Level Task Planning" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.341, + 0.484, + 0.416 + ], + "angle": 0, + "content": "ManualPlan introduces a novel framework for high-level task planning of robotic manipulation with appliance manuals. It processes manual content, generates detailed plans, and aligns parts in the real appliance with those referenced in the manual." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.417, + 0.484, + 0.568 + ], + "angle": 0, + "content": "Manual Resolution Module. The appliance manual in our dataset is in PDF format, which is consistent with the real manuals. The PDF file cannot be directly processed by large language models. Therefore, we divide the manual PDF file into page-level images first. Then, we extract text content from these page images using the Paddle OCR model. We also utilize the multimodal large language model (MLLM) to detect visual elements like figures and tables on each page image. In this way, we can obtain the page-granularity text and vision information about the whole manual." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.569, + 0.483, + 0.735 + ], + "angle": 0, + "content": "Manipulation Planning Module. With extracted manual information, we call LLM to plan a step-by-step appliance manipulation process for the given task. The large language model must specify the part function name and manipulated states in the dictionary format like “{'Brew Strength Lever': 'Rotate 60 degrees'}” for every planning step. Only the manual-mentioned parts can be used in the plan. Considering just a part of the tasks are demonstrated in the manual, the LLM should make plans based on both common sense about appliance usage and the manual content. Any unresolvable prediction will trigger the regeneration request." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Part Alignment Module. To achieve this goal, we incorporate a visual prompting mechanism known as Set-of-Mark (SoM) [28] to align parts in the field of view and parts depicted in the manual. SoM leverages segmentation models to partition an image into distinct regions and assign a numeric marker to each, which can significantly improve the visual grounding capabilities of multimodal large language models. Here, we utilize the open-vocabulary object detection model Grounding-DINO [18] to predict bounding boxes of movable parts and prompt Segment Anything [13] with these bounding boxes" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.292, + 0.907, + 0.549 + ], + "angle": 0, + "content": "to generate movable part masks. These masks are assigned different colors with transparency and overlayed on the observation image. The ID is marked at the geometric center of each part mask. Then, we provide the observation image and appliance diagrams from manual to MLLM requiring it to identify the mapping relationships from mask IDs to part function names. This way, we realize the part alignment between the observation and the manual. For the Manual & CAD Model-based Manipulation task, although the CAD model is available, the part function names still need to be aligned with parts on the CAD model. Therefore, we use a similar SoM-based approach to achieve such alignment with MLLM. For the other two tasks, the CAD model is unavailable for the Pure Manual-based Manipulation task while part alignment is already achieved in the Manual-CAD-Appliance Aligned Manipulation task." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.575, + 0.898, + 0.591 + ], + "angle": 0, + "content": "4.2. ManualPlan-Driven Low-Level Manipulation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.599, + 0.906, + 0.658 + ], + "angle": 0, + "content": "When executing manipulation on the appliance, we employ ManualPlan as a high-level planner and design different manipulation policies for Track 2 (CAD Model Available) and Track 3 (CAD Model Unavailable)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.659, + 0.907, + 0.902 + ], + "angle": 0, + "content": "CAD-assisted Manipulation Policy. When the CAD model is available, we can estimate the 6-DoF pose of the observed appliance in the camera frame by the pose estimation models like [25][31][30]. With such an estimated object pose, the observed appliance can be aligned with its CAD model and all joints' positions and axis directions on the observed appliance become available. Therefore, we design primitive actions for different appliance parts to synthesize CAD-assisted manipulation policy. For door and parts with slider (e.g., screen, slider, lid), we provide the 3D bounding box of the target part to constrain the AnyGrasp [7] model to predict part-level grasp poses. Then, the gripper poses in the motion sequences can be calculated based on the information about the joint and manipulation method. For the button part, we take its surface geometrical center as the contact point and control the closed gripper" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "22601" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.907, + 0.132 + ], + "angle": 0, + "content": "Table 2. Baseline model's performances on three challenge tracks of CheckManual benchmark. The results of Track 1 are reported in the \"Part Alignment Success Rate / Task Planning Success Rate\" format. For Track 2 and 3, their results are reported in \"Task Completion Rate / Task Success Rate\" format. The \"w/o manual\" means ablating appliance manuals input for the above methods." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.142, + 0.905, + 0.276 + ], + "angle": 0, + "content": "
BASELINE MODEL图图图图图图图图图图图Total
Track 1: Manual-CAD-Appliance Aligned Planing
ManualPlan w/o manual22.96 / 22.22 - / 4.2139.47 / 22.82 - / 2.6818.52 / 14.29 - / 20.0072.84 / 60.19 - / 48.154.17 / 3.91 - / 0.788.13 / 8.64 - / 2.4794.44 / 68.00 - / 36.0047.22 / 30.56 - / 6.9441.67 / 40.62 - / 15.6221.15 / 14.39 - / 4.3215.46 / 9.96 - / 2.8527.32 / 20.70 - / 7.99
Track 2: Manual & CAD based Manipulation
ManualPlan + Primitive Actions w/o manual2.39 / 1.15 1.05 / 0.387.12 / 4.70 1.76 / 1.3413.63 / 11.43 21.39 / 14.290.46 / 0.00 0.46 / 0.000.78 / 0.78 0.78 / 0.787.91 / 3.70 3.50 / 1.2310.00 / 4.00 2.00 / 0.0011.11 / 4.86 5.09 / 1.393.12 / 0.00 1.04 / 0.002.32 / 1.44 2.28 / 2.162.56 / 1.42 1.69 / 1.424.50 / 2.39 2.47 / 1.37
Track 3: Pure Manual-based Manipulation
ManualPlan + VoxPoser [10] w/o manual6.11 / 1.92 0.99 / 0.387.18 / 4.70 1.39 / 0.677.10 / 5.71 0.57 / 0.003.01 / 1.85 0.46 / 0.003.12 / 3.12 1.56 / 1.567.99 / 3.70 1.54 / 0.620.00 / 0.00 0.00 / 0.0010.66 / 6.25 2.50 / 0.696.75 / 3.12 0.00 / 0.000.48 / 0.00 0.00 / 0.002.96 / 2.14 0.98 / 0.715.18 / 2.87 1.11 / 0.55
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.292, + 0.486, + 0.565 + ], + "angle": 0, + "content": "to push the button from the normal direction of the contact point. For the knob part, we control the open gripper to get close to the knob from the normal direction of its surface geometrical center and close the gripper to grasp. After the gripper successfully grasps the knob, it will rotate planned degrees around the normal of its surface geometrical center. CAD-free Manipulation Policy. We integrate a recently proposed open-vocabulary manipulation model VoxPoser [10] into our ManualPlan to complete the CAD-free manipulation task. In this framework, ManualPlan and VoxPoser cooperate like the cerebrum and cerebellum in the human brain. ManualPlan reads manuals, makes high-level planning, and aligns observed components with manual-depicted parts. Then, it grounds the target component and demonstrates the corresponding manipulation method in language format, which is transferred to VoxPoser with camera RGB-D observation to execute low-level manipulation action on the appliance." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.578, + 0.224, + 0.594 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.602, + 0.308, + 0.618 + ], + "angle": 0, + "content": "5.1. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.626, + 0.484, + 0.792 + ], + "angle": 0, + "content": "When creating the CheckManual dataset, we utilize GPT-4o [1] as the large language model and multimodal large language model. Following previous work [26], we use the Franka Panda flying gripper in the simulation. The groundingDINO [18] model is utilized to ground the bounding boxes of appliance components. The predicted bounding boxes prompt SAM [13] model to generate component masks. FoundationPose [25] helps us estimate appliance 6-DoF pose in the camera observation. In the evaluation, manual comprehension and manipulation planning processes are assisted by GPT-4o [1] model." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.802, + 0.3, + 0.817 + ], + "angle": 0, + "content": "5.2. Simulation Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.484, + 0.901 + ], + "angle": 0, + "content": "We evaluate our proposed ManualPlan model and ManualPlan-driven framework using the CheckManual benchmark, which employs SAPIEN [27] as the simulation. This environment provides a realistic and flexible simulation for testing robotic manipulation tasks across a" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.292, + 0.908, + 0.61 + ], + "angle": 0, + "content": "variety of appliances. The experimental results on different tasks and appliance categories are summarized in Table 2. From the results of Track 1, we observe that the ManualPlan model is capable of generating a part of correct task plans based on multi-page manuals. However, even the most advanced Multi-Modal Large Language Models (MLLMs) still face significant challenges in aligning manual part annotations with the corresponding CAD models of appliances, which greatly limits the overall performance of ManualPlan in task planning. For Track 2 and Track 3, the performance of the ManualPlan-driven framework highlights the difficulties posed by long-horizon, manual-based manipulation tasks. After analyzing failure cases, we identified that the accumulation of errors in part identification, task planning, and multi-step execution are the primary causes of failure. Additionally, we conducted an ablation study by removing the manual (i.e., w/o manual) to investigate its impact on task performance. As shown in Table 2, the manual plays a crucial role in guiding the robot to accurately plan and execute manipulation tasks, underscoring its importance in completing appliance manipulation tasks." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.622, + 0.734, + 0.637 + ], + "angle": 0, + "content": "5.3. Real Robot Deployment" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.646, + 0.907, + 0.768 + ], + "angle": 0, + "content": "To demonstrate the potential of our approach in real-world applications, we conducted real-world robot experiments following the task definition of Track 3 from the Check-Manual benchmark. The hardware setup involved a Franka Panda robot arm and a RealSense D415 RGB-D camera mounted on the end-effector. Considering error accumulation in the long-horizon execution, we follow [11] to cache correct manipulation actions in every step." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.784, + 0.634, + 0.8 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In this work, we introduce CheckManual, the first benchmark for manual-based appliance manipulation, and define new challenges, metrics, and simulator environments for model evaluation. Additionally, we present the first manual-based manipulation planning model ManualPlan to set up a group of baselines for the CheckManual benchmark." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22602" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.241, + 0.108 + ], + "angle": 0, + "content": "Acknowledgment" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.114, + 0.484, + 0.185 + ], + "angle": 0, + "content": "This work was supported by the National Youth Talent Support Program (8200800081) and National Natural Science Foundation of China (No. 62376006). We also thank Yang Liu and Zhuoqun Xu from Polar Mirage for their technical support on assets." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.196, + 0.188, + 0.212 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.221, + 0.484, + 0.289 + ], + "angle": 0, + "content": "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.291, + 0.484, + 0.36 + ], + "angle": 0, + "content": "[2] Michael Ahn, Anthony Brohan, Noah Brown, Yevgen Chebotar, Omar Cortes, Byron David, Chelsea Finn, Chuyuan Fu, Keerthana Gopalakrishnan, Karol Hausman, et al. Do as i can, not as i say: Grounding language in robotic affordances. arXiv preprint arXiv:2204.01691, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.361, + 0.483, + 0.581 + ], + "angle": 0, + "content": "[3] Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Xi Chen, Krzysztof Choromanski, Tianli Ding, Danny Driess, Avinava Dubey, Chelsea Finn, Pete Florence, Chuyuan Fu, Montse Gonzalez Arenas, Keerthana Gopalakrishnan, Kehang Han, Karol Hausman, Alexander Herzog, Jasmine Hsu, Brian Ichter, Alex Irpan, Nikhil Joshi, Ryan Julian, Dmitry Kalashnikov, Yuheng Kuang, Isabel Leal, Lisa Lee, Tsang-Wei Edward Lee, Sergey Levine, Yao Lu, Henryk Michalewski, Igor Mordatch, Karl Pertsch, Kanishka Rao, Krista Reymann, Michael Ryoo, Grecia Salazar, Pannag Sanketi, Pierre Sermanet, Jaspiar Singh, Anikait Singh, Radu Soricut, Huong Tran, Vincent Vanhoucke, Quan Vuong, Ayzaan Wahid, Stefan Welker, Paul Wohlhart, Jialin Wu, Fei Xia, Ted Xiao, Peng Xu, Sichun Xu, Tianhe Yu, and Brianna Zitkovich. Rt-2: Vision-language-action models transfer web knowledge to robotic control, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.582, + 0.485, + 0.789 + ], + "angle": 0, + "content": "[4] Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Joseph Dabis, Chelsea Finn, Keerthana Gopalakrishnan, Karol Hausman, Alex Herzog, Jasmine Hsu, Julian Ibarz, Brian Ichter, Alex Irpan, Tomas Jackson, Sally Jesmonth, Nikhil J Joshi, Ryan Julian, Dmitry Kalashnikov, Yuheng Kuang, Isabel Leal, Kuang-Huei Lee, Sergey Levine, Yao Lu, Utsav Malla, Deeksha Manjunath, Igor Mordatch, Ofir Nachum, Carolina Parada, Jodilyn Peralta, Emily Perez, Karl Pertsch, Jornell Quiambao, Kanishka Rao, Michael Ryoo, Grecia Salazar, Pannag Sanketi, Kevin Sayed, Jaspiar Singh, Sumedh Sontakke, Austin Stone, Clayton Tan, Huong Tran, Vincent Vanhoucke, Steve Vega, Quan Vuong, Fei Xia, Ted Xiao, Peng Xu, Sichun Xu, Tianhe Yu, and Brianna Zitkovich. Rt-1: Robotics transformer for real-world control at scale, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.79, + 0.484, + 0.872 + ], + "angle": 0, + "content": "[5] Wenzhe Cai, Siyuan Huang, Guangran Cheng, Yuxing Long, Peng Gao, Changyin Sun, and Hao Dong. Bridging zero-shot object navigation and foundation models through pixel-guided navigation skill. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 5228-5234. IEEE, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.873, + 0.484, + 0.902 + ], + "angle": 0, + "content": "[6] Martin Ester, Hans-Peter Kriegel, Jörg Sander, Xiaowei Xu, et al. A density-based algorithm for discovering clusters in" + }, + { + "type": "list", + "bbox": [ + 0.1, + 0.221, + 0.485, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.119 + ], + "angle": 0, + "content": "large spatial databases with noise. In kdd, pages 226-231, 1996. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.122, + 0.905, + 0.191 + ], + "angle": 0, + "content": "[7] Hao-Shu Fang, Chenxi Wang, Hongjie Fang, Minghao Gou, Jirong Liu, Hengxu Yan, Wenhai Liu, Yichen Xie, and Cewu Lu. Anygrasp: Robust and efficient grasp perception in spatial and temporal domains. IEEE Transactions on Robotics (T-RO), 2023. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.192, + 0.905, + 0.277 + ], + "angle": 0, + "content": "[8] Jiayuan Gu, Sean Kirmani, Paul Wohlhart, Yao Lu, Montserrat Gonzalez Arenas, Kanishka Rao, Wenhao Yu, Chuyuan Fu, Keerthana Gopalakrishnan, Zhuo Xu, Priya Sundaresan, Peng Xu, Hao Su, Karol Hausman, Chelsea Finn, Quan Vuong, and Ted Xiao. Rt-trajectory: Robotic task generalization via hindsight trajectory sketches, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.278, + 0.905, + 0.333 + ], + "angle": 0, + "content": "[9] Haoxu Huang, Fanqi Lin, Yingdong Hu, Shengjie Wang, and Yang Gao. Copa: General robotic manipulation through spatial constraints of parts with foundation models. arXiv preprint arXiv:2403.08248, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.515, + 0.334, + 0.905, + 0.39 + ], + "angle": 0, + "content": "[10] Wenlong Huang, Chen Wang, Ruohan Zhang, Yunzhu Li, Jiajun Wu, and Li Fei-Fei. Voxposer: Composable 3d value maps for robotic manipulation with language models. arXiv preprint arXiv:2307.05973, 2023. 1, 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.515, + 0.392, + 0.905, + 0.448 + ], + "angle": 0, + "content": "[11] Wenlong Huang, Chen Wang, Yunzhu Li, Ruohan Zhang, and Li Fei-Fei. Rekep: Spatio-temporal reasoning of relational keypoint constraints for robotic manipulation. arXiv preprint arXiv:2409.01652, 2024. 1, 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.449, + 0.905, + 0.546 + ], + "angle": 0, + "content": "[12] Moo Jin Kim, Karl Pertsch, Siddharth Karamcheti, Ted Xiao, Ashwin Balakrishna, Suraj Nair, Rafael Rafailov, Ethan Foster, Grace Lam, Pannag Sanketi, Quan Vuong, Thomas Kollar, Benjamin Burchfiel, Russ Tedrake, Dorsa Sadigh, Sergey Levine, Percy Liang, and Chelsea Finn. Openvla: An open-source vision-language-action model. arXiv preprint arXiv:2406.09246, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.548, + 0.905, + 0.616 + ], + "angle": 0, + "content": "[13] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dollar, and Ross Girshick. Segment anything. arXiv:2304.02643, 2023. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.618, + 0.905, + 0.702 + ], + "angle": 0, + "content": "[14] Isabel Leal, Krzysztof Choromanski, Deepali Jain, Avinava Dubey, Jake Varley, Michael Ryoo, Yao Lu, Frederick Liu, Vikas Sindhwani, Quan Vuong, et al. Sara-rt: Scaling up robotics transformers with self-adaptive robust attention. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 6920–6927. IEEE, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.704, + 0.905, + 0.771 + ], + "angle": 0, + "content": "[15] Xinghang Li, Minghuan Liu, Hanbo Zhang, Cunjun Yu, Jie Xu, Hongtao Wu, Chilam Cheang, Ya Jing, Weinan Zhang, Huaping Liu, et al. Vision-language foundation models as effective robot imitators. arXiv preprint arXiv:2311.01378, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.774, + 0.905, + 0.858 + ], + "angle": 0, + "content": "[16] Xiaqi Li, Mingxu Zhang, Yiran Geng, Haoran Geng, Yuxing Long, Yan Shen, Renrui Zhang, Jiaming Liu, and Hao Dong. Maniplm: Embodied multimodal large language model for object-centric robotic manipulation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18061-18070, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.859, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[17] Shanchuan Lin, Anran Wang, and Xiao Yang. Sdxlightning: Progressive adversarial diffusion distillation, 2024. 5" + }, + { + "type": "list", + "bbox": [ + 0.515, + 0.093, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22603" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.161 + ], + "angle": 0, + "content": "[18] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Qing Jiang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, and Lei Zhang. Grounding dino: Marrying dino with grounded pre-training for open-set object detection, 2024. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.164, + 0.482, + 0.219 + ], + "angle": 0, + "content": "[19] Yuxing Long, Wenzhe Cai, Hongcheng Wang, Guanqi Zhan, and Hao Dong. Instructnav: Zero-shot system for generic instruction navigation in unexplored environment. arXiv preprint arXiv:2406.04882, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.221, + 0.482, + 0.289 + ], + "angle": 0, + "content": "[20] Yuxing Long, Xiaqi Li, Wenzhe Cai, and Hao Dong. Discuss before moving: Visual language navigation via multi-expert discussions. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 17380-17387. IEEE, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.291, + 0.482, + 0.333 + ], + "angle": 0, + "content": "[21] Jianlan Luo, Charles Xu, Jeffrey Wu, and Sergey Levine. Hil-serl: Precise and dexterous robotic manipulation via human-in-the-loop reinforcement learning, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.335, + 0.482, + 0.404 + ], + "angle": 0, + "content": "[22] Kaichun Mo, Shilin Zhu, Angel X. Chang, Li Yi, Subarna Tripathi, Leonidas J. Guibas, and Hao Su. PartNet: A large-scale benchmark for fine-grained and hierarchical part-level 3D object understanding. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.406, + 0.482, + 0.503 + ], + "angle": 0, + "content": "[23] Abhilash Nandy, Soumya Sharma, Shubham Maddhashiya, Kapil Sachdeva, Pawan Goyal, and NIloy Ganguly. Question answering over electronic devices: A new benchmark dataset and a multi-task learning based QA framework. In Findings of the Association for Computational Linguistics: EMNLP 2021, pages 4600-4609, Punta Cana, Dominican Republic, 2021. Association for Computational Linguistics. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.505, + 0.482, + 0.602 + ], + "angle": 0, + "content": "[24] Abhilash Nandy, Soumya Sharma, Shubham Maddhashiya, Kapil Sachdeva, Pawan Goyal, and NIloy Ganguly. Question answering over electronic devices: A new benchmark dataset and a multi-task learning based QA framework. In Findings of the Association for Computational Linguistics: EMNLP 2021, pages 4600-4609, Punta Cana, Dominican Republic, 2021. Association for Computational Linguistics. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.604, + 0.482, + 0.645 + ], + "angle": 0, + "content": "[25] Bowen Wen, Wei Yang, Jan Kautz, and Stan Birchfield. FoundationPose: Unified 6d pose estimation and tracking of novel objects. In CVPR, 2024. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.647, + 0.482, + 0.729 + ], + "angle": 0, + "content": "[26] Ruihai Wu, Yan Zhao, Kaichun Mo, Zizheng Guo, Yian Wang, Tianhao Wu, Qingnan Fan, Xuelin Chen, Leonidas Guibas, and Hao Dong. VAT-mart: Learning visual action trajectory proposals for manipulating 3d ARTiculated objects. In International Conference on Learning Representations, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.731, + 0.482, + 0.814 + ], + "angle": 0, + "content": "[27] Fanbo Xiang, Yuzhe Qin, Kaichun Mo, Yikuan Xia, Hao Zhu, Fangchen Liu, Minghua Liu, Hanxiao Jiang, Yifu Yuan, He Wang, Li Yi, Angel X. Chang, Leonidas J. Guibas, and Hao Su. SAPIEN: A simulated part-based interactive environment. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.817, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[28] Jianwei Yang, Hao Zhang, Feng Li, Xueyan Zou, Chunyuan Li, and Jianfeng Gao. Set-of-mark prompting unleashes extraordinary visual grounding in gpt-4v. arXiv preprint arXiv:2310.11441, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[29] Senqiao Yang, Jiaming Liu, Ray Zhang, Mingjie Pan, Zoey Guo, Xiaqi Li, Zehui Chen, Peng Gao, Yandong Guo, and" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.135 + ], + "angle": 0, + "content": "Shanghang Zhang. Lidar-llm: Exploring the potential of large language models for 3d lidar understanding. arXiv preprint arXiv:2312.14074, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[30] Jiyao Zhang, Mingdong Wu, and Hao Dong. Generative category-level object pose estimation via diffusion models. Advances in Neural Information Processing Systems, 36: 54627-54644, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.192, + 0.905, + 0.261 + ], + "angle": 0, + "content": "[31] Jiyao Zhang, Weiyao Huang, Bo Peng, Mingdong Wu, Fei Hu, Zijian Chen, Bo Zhao, and Hao Dong. Omni6dpose: A benchmark and model for universal 6d object pose estimation and tracking. In European Conference on Computer Vision, pages 199-216. Springer, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.263, + 0.905, + 0.318 + ], + "angle": 0, + "content": "[32] Liang Zhang, Anwen Hu, Jing Zhang, Shuo Hu, and Qin Jin. Mpmqa: multimodal question answering on product manuals. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 13958-13966, 2023. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.318 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22604" + } + ] +] \ No newline at end of file diff --git a/2025/CheckManual_ A New Challenge and Benchmark for Manual-based Appliance Manipulation/c1aa26f1-f8e2-45ab-92dc-640acaf534e5_origin.pdf b/2025/CheckManual_ A New Challenge and Benchmark for Manual-based Appliance Manipulation/c1aa26f1-f8e2-45ab-92dc-640acaf534e5_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..88d9dc63eaed08cea22b23daa093d9ae95b8bafd --- /dev/null +++ b/2025/CheckManual_ A New Challenge and Benchmark for Manual-based Appliance Manipulation/c1aa26f1-f8e2-45ab-92dc-640acaf534e5_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b2428660a414fa2750f623365025fa1f65ea3a3a3d6bc52e09005b74199cb86f +size 2697358 diff --git a/2025/CheckManual_ A New Challenge and Benchmark for Manual-based Appliance Manipulation/full.md b/2025/CheckManual_ A New Challenge and Benchmark for Manual-based Appliance Manipulation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..44cf21e66404a4807fd81952eb339aaa5b4557e2 --- /dev/null +++ b/2025/CheckManual_ A New Challenge and Benchmark for Manual-based Appliance Manipulation/full.md @@ -0,0 +1,258 @@ +# CheckManual: A New Challenge and Benchmark for Manual-based Appliance Manipulation + +Yuxing Long $^{1,2}$ , Jiyao Zhang $^{1,2}$ , Mingjie Pan $^{1,2}$ , Tianshu Wu $^{1}$ , Taewhan Kim $^{1,2}$ and Hao Dong $^{1,2\dagger}$ $^{1}$ CFCS, School of Computer Science, Peking University $^{2}$ PKU-Agibot Lab + +# Abstract + +Correct use of electrical appliances has significantly improved human life quality. Unlike simple tools that can be manipulated with common sense, different parts of electrical appliances have specific functions defined by manufacturers. If we want the robot to heat bread by microwave, we should enable them to review the microwave's manual first. From the manual, it can learn about component functions, interaction methods, and representative task steps about appliances. However, previous manual-related works remain limited to question-answering tasks while existing manipulation researchers ignore the manual's important role and fail to comprehend multi-page manuals. In this paper, we propose the first manual-based appliance manipulation benchmark CheckManual. Specifically, we design a large model-assisted human-revised data generation pipeline to create manuals based on CAD appliance models. With these manuals, we establish novel manual-based manipulation challenges, metrics, and simulator environments for model performance evaluation. Furthermore, we propose the first manual-based manipulation planning model ManualPlan to set up a group of baselines for the CheckManual benchmark. Our project page is available at https://sites.google.com/view/checkmanual. + +# 1. Introduction + +Since the twentieth century, the emergence of appliances has significantly changed human society. In daily life, we extensively use various electrical appliances to assist our living and working. In recent years, robots have demonstrated enormous potential and value in areas like component assembly [21], goods sorting [7], and indoor navigation [19]. Imagine if we could instruct robots to operate appliances to serve us, such as baking cakes with ovens, washing clothes with washing machines, and making juice with blenders, we would be further liberated from tedious household chores and have more free time. + +However, appliances are different from simple tools like hammers or teapots that can be correctly used based on + +![](images/ab7c76859152a155849ee39e93795a72e549867bbcb954f049c26fb6c36d452e.jpg) + +![](images/e058bbca83b265920b4e7038e9411d4e7269fc9a798759b7a4b842c4acbe7641.jpg) +Figure 1. Comparison between appliance manipulation with and without manual. If the manual is absent, the concrete functions of the appliance part are unknown. As shown in the top image, the robot feels confused about tasks such as opening the microwave door. To correctly use the appliance, the robot must comprehend the manual content and follow its guidance to work. + +common sense. One appliance may have multiple movable components and each component has specific functions even if they are the same type. Taking a microwave as an example, one button may control the temperature while the other button may set the time. Such function information is difficult to infer according to the appliance's visual appearance or human common sense. This time, we need to check the user manuals to learn about component functions and supporting tasks. Existing manipulation models like [4][3][12][10][11] already can adopt simple tools to do easy tasks but cannot read multi-page manuals not to say using the appliances to complete long-horizon complex tasks. Besides, previous manual-related research and academic datasets [23][32] are all collected by the NLP researchers1, which only focuses on page retrieval and question-answering tasks rather than manipulation. These + +academic datasets manual datasets also do not provide appliance CAD models and cannot be used for quantitative manipulation evaluation in the simulator. Therefore, there are no manual-based manipulation evaluation datasets and benchmarks, nor manipulation models capable of following manuals to use appliances at present. + +To address this problem, we propose the first manual-based manipulation benchmark CheckManual. In the benchmark, we collect new evaluation data, design novel manipulation challenges, and propose new baseline models. Specifically, we first design an LLM-assisted human-revised manual creation pipeline based on appliance CAD models. For every CAD model, this pipeline can follow our surveyed real manual analysis to generate diverse appliance manuals with different part functions, manipulation demonstrations, and page layouts. Then, we set up three manual-based appliance manipulation challenge tracks including manipulation planning, CAD-assisted manipulation, and CAD-free manipulation. They progressively cover the key steps in manual-based appliance manipulation, comprehensively evaluating the model's ability to operate appliances with manuals. Furthermore, we propose a manual-based manipulation planning method ManualPlan, and create a set of CAD-assisted primitive actions for appliance manipulation. The ManualPlan aligns manuals with appliances and plans detailed manipulation steps according to manual content, which can control primitive actions or open-vocabulary manipulation models to physically interact with appliances. These also act as the baseline methods for our CheckManual benchmark. + +In this paper, our main contributions are: + +- We collect the first CAD model-aligned appliance manual dataset CheckManual through our devised large model-assisted pipeline. Every data collection stage undergoes careful human verification to guarantee correctness. +- We design novel challenges for manual-based appliance manipulation considering different difficulty levels and conditions. Furthermore, we establish the evaluation environment in the simulator for the benchmark. +- We propose the first manual-based manipulation planning model ManualPlan to set up a group of baselines for three-level challenge tracks in CheckManual. + +# 2. Related Work + +# 2.1. Manipulation Large Models + +Large models pre-trained on Internet-scale text and image data emerge powerful capabilities for task planning and common sense reasoning. Therefore, leveraging large models has become a hotspot in embodied areas such as autonomous driving [29], navigation [20] [5] and manipulation. According to training and inference forms, existing manipulation methods can be divided into two main categories. The first category of work focuses on fine-tuning + +large models with manipulation trajectories collected from the simulator environment and real-world. The representative models are Google RT-series from RT-1 [4], RT-2 [3], and RT-Trajectory [8] to STAR-RT [14]. These models take vision observation, robot states, and task instruction as input, and directly predict the low-level actions by multimodal large models. Similar open-sourced work includes ManipLLM [16], RoboFlamingo [15] and OpenVLA [12]. The other category of work makes efforts to effectively integrate large models' zero-shot inferences into the manipulation planning and decision process. SayCan [2] first utilizes a multimodal large model to score a group of kills to make progress toward the high-level task. Voxposer [10] drives a multimodal large model to create 3D value maps to guide the robot toward the target object part. CoPa [9] and ReKep [11] leverage multimodal large models to generate spatial-temporal constraints for complex tasks like "insert the flower into vase" or "pour tea into water". However, all these large model-based methods can only make plans based on their internal common sense. Considering heating the bread with the microwave, they even struggle to figure out the temperature knob and timer knob, not to say complete the whole task. + +# 2.2. Appliance Manual-based Tasks + +NLP researchers have recognized the manual's importance in addressing user questions. TechQA [24] collects 1400 user questions from online forums and annotates the corresponding answers from IBM technical documents to set up Document-based Question Answering task. S10 QA [23] manually creates 904 QA pairs based on the manual of the Samsung S10 phone. Nandy et al. [23] also collect 950 QA pairs based on the Samsung Smart TV manual to generate Smart TV/Remote QA dataset. These two datasets are utilized to establish a Manual-based Question Answering benchmark. The above datasets are all pure text corpus. PM209 [32] collects 209 product manuals from 27 well-known consumer electronic brands to set up Manual Page Retrieval task and Manual-based Multimodal Question Answering task. Although these datasets and tasks are related to the appliance manual, they are all designed for NLP evaluation and lack corresponding appliance CAD models. Therefore, manual-based manipulation benchmarks cannot be directly established on these existing related researches. We create the first manual & CAD model-aligned dataset and design novel benchmarks and baseline models for manual-based appliance manipulation. + +# 3. Manual-based Manipulation Benchmark CheckManual + +In this part, we will introduce the creation process of our CheckManual dataset in Section 3.1, analyze the numerical and visual statistics about CheckManual in Section 3.2, and + +![](images/5843d26d66f77e8664a28ca752bc3a37daa64329e2e35a744b0fbdebf840fd3b.jpg) +Figure 2. Generation workflow of CheckManual dataset. In the leftmost part (Section 3.1.1), we analyze real manuals to learn about their formats and collect different categories of appliance CAD models. The middle part (Section 3.1.2 - 3.1.4) demonstrates the creation of manual materials, including appliance creation, task generation, and figure design. Human verifies every step to guarantee correctness. Based on this information, the rightmost part (Section 3.1.5) generates appliance manuals with diverse formats through the LATEX codes. + +![](images/7c4218757fdc77571c474a4c8935b8677945450d644d481960667ee100db0ae5.jpg) + +![](images/869c06a7cf586df7fd7e134fb4bae8043a6dfde432fa60dcb7c390ce1cb3f97a.jpg) + +![](images/c83d961fe54a2f97aa3cffdb04305b6350b706e86fcd95f59e983e896182dfc2.jpg) + +define the evaluation for robot manipulation based on these manuals in Section 3.3. + +# 3.1. Appliance Manual Creation + +To create an open-source available manual dataset, We design a large model-driven human-verified manual generation workflow based on appliance CAD models. Figure 2 demonstrates the generation workflow of the CheckManual dataset, which will be detailed in the following. + +# 3.1.1 Preparation + +Before formally starting manual creation, we analyze a stack of real appliance manuals to learn about real manual formats and collect available articulated appliance CAD models for manual creation. + +Analysis of Real Appliance Manuals We collect 110 appliance manual PDFs from the Internet, which cover 11 categories of our dataset involved appliances. These manuals come from different countries and regions, including China, the European Union, Japan, and India. Then we manually analyze each manual from Functional Part Annotation Type, Guidance of Functional Part Operation, and Task Expression Format these three aspects. We find that the functional parts are usually explained in the point-line annotation formats. These lines connect the points on appliance parts and the function names or IDs of parts as shown in the middle of Figure 2. The guidance of functional + +part operation is commonly represented in formats: 1 Pure text description like "Rotate temperature knob $90^{\circ}$ clockwise". 2 Close-up figures of the manipulated part, such as the close and open states of the red microwave door in Figure 2. 3 Part motion decomposition diagram that shows the dynamic change of part during its motion process. Figure 2 shows one example of opening the white mixer lid. 4 Motion trajectory of the contact point on the part, like manipulating Figure 2's blue mixer slider from right to left. Besides, the manipulation tasks in the real manuals are often depicted in the text, bullet lists, ordered lists, and tables with one-column or multi-columns. + +Appliance CAD Models Our appliance CAD models come from the PartNet-Mobility [22] Dataset, which contains common categories of household appliances CAD models, including microwaves, dishwashers, and refrigerators. Appliance models in this dataset are annotated with manipulable parts, such as buttons, knobs, levers, sliders, and so on. In total, we collect 11 categories, 182 CAD models, which cover widely-used household appliances in our daily lives. + +# 3.1.2 Appliance Creation + +With the appliance CAD models, we can create concrete appliances by annotating each movable part with a specific function name (e.g., temperature knob) and corresponding function states (e.g., clockwise $60^{\circ} - 120^{\circ}C$ and clockwise $120^{\circ} - 160^{\circ}C$ ). This process is detailed in the following. + +Part Function Name Annotation Considering part function names are closely related to the appliance layout and + +![](images/d17051db036ef41e2b5673d328e01e347ba027b862a46bf0a159ede68e60c464.jpg) +Figure 3. One example of appliance manual from our CheckManual dataset. Please zoom in for details. + +![](images/7afe3484708466aa39eea7c029df3af9dd4d5644a2736d57522c0e53dbef9dd6.jpg) + +![](images/e07905b3275cf7d380266c9f671d0c865a5e19637e9cd9e9f761b8019bd3424f.jpg) + +![](images/fc3fd8f20dbb3d1acc1dc20175535813ceb0a062a3253480a4733017957f5b36.jpg) + +parts' visual features, we drive the annotation process by the multimodal large language model (MLLM). Before annotation, manipulable part IDs and types are automatically labeled on the appliance overview figures through point-line annotation. Taking these figures as visual prompts, we instruct the MLLM to design the function name of each part. The multimodal large model should consider the appliance type and carefully analyze part visual information to create reasonable annotations. Re-generation mechanism would be triggered when the number of annotated parts is unequal to the total number of manipulable parts. By resolving the large model's prediction with regular expressions, we can obtain a dictionary recording the one-to-one relationships between part ID and corresponding function name. One such dictionary can be used to generate one appliance. We can create multiple appliances from one CAD model by repeating the above annotation process. + +Part Function State Annotation To annotate part function states, we first sample available states for each manipulable part according to their types. For buttons, we randomly sample available push times in a pre-defined range. For knobs and levers, we randomly sample several rotation degrees at intervals of $60^{\circ}$ . Other manipulable parts like doors, lids, and containers, are assigned with "open" and "close" states. With part function names and available states, we instruct the large language model (LLM) to design concrete functions for each state. This way, each manipulable part on the appliance will be annotated with multiple function states. + +Human Verification about Part Functions Although large models are powerful annotation tools, they still make intolerable mistakes and thus require human verification. In the annotation about function names, MLLM sometimes fails to identify tiny visual elements on the CAD models like already-drawn button logos or scales, which may lead to mismatching between annotated function names and part visual appearance. In the annotation about function states, LLM may create an anti-commonsense design such as “ $500^{\circ}\mathrm{C}$ for oven temperature knob”. Therefore, we check large model-annotated function names and states + +through our designed human verification website (See Appendix) to revise incorrect cases. In this stage, we manually verify more than 2200 part function annotations. + +# 3.1.3 Manipulation Task Creation + +The manual often includes several representative tasks to introduce how to use the appliance. These tasks are also the core of our manual-based manipulation evaluation. We can create appliance-executable tasks based on their part function names and states. + +Task Proposal With function names and states of movable parts on the appliances, we instruct the LLM to propose short-horizon and long-horizon manipulation tasks for each appliance. Every task step should specify the part function name and the manipulated state in a resolvable format like "1. Front Door, Open", "2. Power Button, Push 1 time" ... To improve task quality and reduce repetition, the LLM only needs to design one task in one inference based on already created tasks. + +Human Verification about Manipulation Tasks In the task creation, LLM may also make anti-commonsense mistakes. To guarantee absolute correctness for evaluation, after the LLM completes task generation, we manually check every task step to adjust unreasonable step orders and revise incorrect steps. If the task is completely unreasonable, we will manually create a new one from scratch. Totally, we checked over 1400 manipulation tasks created by LLM. + +# 3.1.4 Manual Figure Design + +According to our analysis of real manuals, cover, control panels, and part manipulation guidance are the main figures in the manuals. Therefore, we devise methods to create these figures. + +Manual Cover The manual cover usually displays appliance appearance and typical usage scenes. we predefined a set of candidate camera poses to capture the appliance from different views. The views with the largest total areas of manipulable parts are selected to render RGB appliance overview figures. To synthesize custom usage scenes, + +![](images/b31e291063126d2190157b947b17f00848e2f16b36744699b0aa9ef97037bd52.jpg) +(a) + +![](images/6fbc62139ead53f34aeace96a753265895645e6888c749ab00a53b4c6113443c.jpg) +(b) + +![](images/b0341611a065ee7f7785289612ccb86d6a3827862164f77bd8d73424fb62af5c.jpg) +(c) +Figure 4. CheckManual dataset statistics. The above tables and figures from left to right are (a) Numerical statistics. (b) Proportion of different appliances. (c) Distribution of manipulable components on the appliances. (d) Distribution of manipulation task step lengths. + +![](images/7412f40b543c4f6474d9a28f2d6cadbb643f557324d738371c190d292359b948.jpg) +(d) + +we further add the indoor background to the appliance overview through text-to-image generation model SDXL-Lightning [17]. The background room type is chosen based on the appliance category. For example, the oven is usually put into the kitchen while the washing machine is often arranged in the laundry room. + +Overview Diagram & Control Panel The overview diagram and control panel demonstrate the function name of every movable part on the appliance. Before annotation, we first apply the DBSCAN [6] algorithm to try clustering buttons and knobs based on their geometric centers. If clustered buttons and knobs are found, the areas with them are selected as control panels and cropped from the appliance overview figures as close-ups of control panels. Then, we adopt point-line annotation methods (Figure 2) from the real manuals to draw overview diagram and control panel based on part function names created in the previous stage. Furthermore, considering real manuals often deliver overview diagrams and control panels in line drawings, we create sketch-style figures by conducting edge detection on the RGB normal maps. This way, we can generate these two kinds of figures both in RGB style and sketch style. + +Part Manipulation Guidance This kind of figure usually introduces the physical interaction approach for appliance manipulable parts. To generate such figures, we first render manipulable parts according to their types and visible surface areas. For buttons, knobs, levers, and drawers, we render them with the camera axis parallel to their joint axes. Other manipulable parts, including doors, sliders, lids, and containers, are rendered with the camera axis perpendicular to their joint axes while ensuring maximum visible areas. Then, we draw manipulation guidance on these figures to display how to manipulate them. Following common representation in the real manuals from Section 3.1.1, we adopted these four different strategies to draw manipulation guidance as shown in Figure 2. To realize this, we manually annotate the 3D position of one contact point for every manipulable part on all appliance CAD models and calculate its motion trajectory according to the part-linked joint axis. + +Human Verification about Manual Figures When collecting appliance overviews based on the largest visible part areas, the view may not be suitable to observe every part. + +Besides, sketch-style figures may contain missing lines or abnormal shadows. Therefore, we conduct human verification on more than 1800 figures to trigger regeneration when these errors occur. + +Table 1. Comparison of different manual-based manipulation challenge tracks in our CheckManual benchmark. + +
AVAILABLE INPUTPREDICTION TARGET
Task InstructionRGB-DManual PDFAppliance CAD ModelHigh-level PlanningLow-level Manipulation
Track 1: Manual-CAD-Appliance Aligned Planing
✓✓✓✓✓
Track 2: Manual & CAD based Manipulation
✓✓✓✓✓
Track 3: Pure Manual-based Manipulation
✓✓✓✓
+ +# 3.1.5 LATEX-based Manual Generation + +As we know, $\mathrm{LATEX}$ is a document programming language that offers both the advantages of standardization and flexibility. We can not only keep the same document style from page to page but also easily insert different types of figures and tables into the document by $\mathrm{LATEX}$ coding. Recently, large language models emerged with powerful capabilities for code generation. Therefore, we decided to leverage LLM to generate the $\mathrm{LATEX}$ codes for the manual and then compile them into a PDF format appliance manual as shown in the rightmost part of Figure 2. + +Before starting this process, we pre-defined a set of widely used $\mathrm{LATEX}$ packages (e.g., multirow, xcolor, caption) and prepared a collection of commonly used icons (e.g., fragility reminder, temperature warning, environment protection). To increase diversity, we randomly sample different formats of covers, diagrams, and guidance demonstrations from the Section 3.1.4 as figure resources and randomly select a text style from real manuals (e.g., text, bullet list, table) for task description. In the generation, we synthesize the prompt with textual part function annotation, manipulation tasks, and above image paths to require the large model to continue one section of $\mathrm{LATEX}$ codes based on the previous. To prevent uncompilable codes, we design a section-granularity compilation mechanism. The newly generated section of $\mathrm{LATEX}$ codes is appended to the end of the previous correct code and compiled together. Regeneration will be triggered when errors occur in the code compi + +lation. Figure 3 displays one manual from our CheckManual dataset. + +# 3.2. Dataset Statistic Analysis + +We totally created 1107 manuals for 369 appliances, which cover 11 common appliance categories in daily life, including washing machines, coffee makers, dishwashers, microwaves, and so on. These appliances contain 12 types, with 2211 different manipulable parts. On average, there are 6.00 manipulable parts on each appliance. In terms of quantity and diversity, parts contained in our dataset can support most appliance operations. Besides, our dataset includes 1464 manipulation tasks in total. Every task is accompanied by step-by-step manipulation parts and target states. These tasks include both single-step and long-horizon tasks with a maximum of 18 steps. The table and charts in Figure 4 display comprehensive statistics about our CheckManual dataset. + +# 3.3. CheckManual Challenge Tracks + +With the collected manual data, we design challenge tracks for manual-based manipulation. The basic setting is given a task instruction $T$ and a corresponding appliance PDF manual $M$ , the robot should follow the manual to complete the tabletop appliance manipulation task with RGB-D observation $I_{c}$ and $I_{d}$ . Based on this basic setting, we design the challenge tracks as shown in Table 1 and the following. + +Track 1: CAD-Appliance Aligned Planing We first consider an ideal setting where the appliance CAD models are available and CAD-Appliance is aligned. The "CAD-Appliance Aligned" means part area and ID are accessible on the camera observation. In this case, the evaluation focuses on testing the robot's capability of analyzing appliance manuals and making manipulation plans based on the given task instruction and manual content. The evaluation is composed of two aspects. Firstly the robot should align all function names in the manuals with appliance parts. Secondly, the task planning should include correct manipulation target part and primitive actions in every task step. This challenge track is free from manipulation execution. + +Track 2: Manual & CAD based Manipulation In this task, the appliance CAD models are available but the mapping relationships among parts on manual, CAD model, and observation are not known. This setting is reasonable because it is easy for manufacturers to provide plain CAD models for their appliance products. For evaluation, the robot should follow the given task instruction to execute low-level manipulation with a two-finger gripper based on manual content and CAD model. To complete this challenge, the robot needs to first estimate the 6-DoF pose of observed objects before using the CAD model. After that, the robot should also align parts on the manual, CAD model, and current observation. After completing the part align + +ment, the robot can obtain part-linked joint position and rotation axis from the CAD model, which will help the robot to make manipulation planning and execution. + +Track 3: Pure Manual-based Manipulation We further consider the setting that aligns with real-world applications where only appliance manuals are available to the robot. The challenges of this setting come from two aspects. On the high level, the robot should analyze the appliance manual to learn about the function information (i.e., name and possible states) and manipulation methods of every movable part on the appliances. Then, the robot should plan the detailed manipulation steps for the given task based on the manual content and common sense. On the low level, the robot must execute the high-level plan step-by-step with its end-effector to complete the instruction-specified task. During the whole process, the robot can only rely on its RGB-D observation and manuals instead of any CAD-model assistance. For both Track 2 and Track 3, any procedural error during the operation will cause the entire task to fail. + +Evaluation Metrics To evaluate model performances on the above challenge tracks, we define the task-level and step-level metrics. On the step level, we evaluate the alignment success rate and planning success rate for Track 1 and task execution success rate and average task completion rate for Track 2 and 3. The alignment success is recognized when all CAD model parts are linked to the correct function names mentioned in the manual. The task planning success is confirmed when part IDs and manipulation actions in every planned step are all correct. The task execution is regarded as a success when all task steps are completed in the correct order. The task completion rate is calculated by "Success Steps / # Task Steps". Note that only steps before the failure step can be counted as successful steps. We take the average values of the above metrics as final results. On the step level, we check the state of the target part after each manipulation step to judge the success. Specifically, we define the whole movement limitation range of the target joint as $L$ . For revolute parts like knobs, levers, and doors, the acceptable error range is defined as $L \pm 30^{\circ}$ . For prismatic parts like the button, container, and slider, the successful manipulation is confirmed when the button moves $>25\%$ $L$ or $MD_{part} / MD_{appliance}$ $>50\%$ if target part and appliances move in the same direction, where $MD$ represented motion distance. + +# 4. Manual-based Manipulation Baselines + +In this section, we design baseline models for the three tracks of the benchmark. For track 1, we propose the ManualPlan model, which focuses on manual-based task planning, as detailed in Section 4.1. For tasks involving robotic execution, corresponding to track 2 and track 3, we extend the ManualPlan model by integrating it as a high-level planning module with existing manipulation models, as dis + +![](images/c405025e32b47094d4e7ed3ee1040f42db8a1d1b3ed5bb66a8de5012992e7047.jpg) +Figure 5. The framework of our ManualPlan model. It is composed of Manual Resolution, Manipulation Planing and Part Alignment modules. The ManualPlan can make high level planning to control the CAD-assisted primitive actions or open-vocabulary manipulation large model (e.g., VoxPoser [10]) to use the appliance, which serves as the baseline models for CheckManual benchmark. + +cussed in Section 4.2. + +# 4.1. ManualPlan for High-Level Task Planning + +ManualPlan introduces a novel framework for high-level task planning of robotic manipulation with appliance manuals. It processes manual content, generates detailed plans, and aligns parts in the real appliance with those referenced in the manual. + +Manual Resolution Module. The appliance manual in our dataset is in PDF format, which is consistent with the real manuals. The PDF file cannot be directly processed by large language models. Therefore, we divide the manual PDF file into page-level images first. Then, we extract text content from these page images using the Paddle OCR model. We also utilize the multimodal large language model (MLLM) to detect visual elements like figures and tables on each page image. In this way, we can obtain the page-granularity text and vision information about the whole manual. + +Manipulation Planning Module. With extracted manual information, we call LLM to plan a step-by-step appliance manipulation process for the given task. The large language model must specify the part function name and manipulated states in the dictionary format like “{'Brew Strength Lever': 'Rotate 60 degrees'}” for every planning step. Only the manual-mentioned parts can be used in the plan. Considering just a part of the tasks are demonstrated in the manual, the LLM should make plans based on both common sense about appliance usage and the manual content. Any unresolvable prediction will trigger the regeneration request. + +Part Alignment Module. To achieve this goal, we incorporate a visual prompting mechanism known as Set-of-Mark (SoM) [28] to align parts in the field of view and parts depicted in the manual. SoM leverages segmentation models to partition an image into distinct regions and assign a numeric marker to each, which can significantly improve the visual grounding capabilities of multimodal large language models. Here, we utilize the open-vocabulary object detection model Grounding-DINO [18] to predict bounding boxes of movable parts and prompt Segment Anything [13] with these bounding boxes + +to generate movable part masks. These masks are assigned different colors with transparency and overlayed on the observation image. The ID is marked at the geometric center of each part mask. Then, we provide the observation image and appliance diagrams from manual to MLLM requiring it to identify the mapping relationships from mask IDs to part function names. This way, we realize the part alignment between the observation and the manual. For the Manual & CAD Model-based Manipulation task, although the CAD model is available, the part function names still need to be aligned with parts on the CAD model. Therefore, we use a similar SoM-based approach to achieve such alignment with MLLM. For the other two tasks, the CAD model is unavailable for the Pure Manual-based Manipulation task while part alignment is already achieved in the Manual-CAD-Appliance Aligned Manipulation task. + +# 4.2. ManualPlan-Driven Low-Level Manipulation + +When executing manipulation on the appliance, we employ ManualPlan as a high-level planner and design different manipulation policies for Track 2 (CAD Model Available) and Track 3 (CAD Model Unavailable). + +CAD-assisted Manipulation Policy. When the CAD model is available, we can estimate the 6-DoF pose of the observed appliance in the camera frame by the pose estimation models like [25][31][30]. With such an estimated object pose, the observed appliance can be aligned with its CAD model and all joints' positions and axis directions on the observed appliance become available. Therefore, we design primitive actions for different appliance parts to synthesize CAD-assisted manipulation policy. For door and parts with slider (e.g., screen, slider, lid), we provide the 3D bounding box of the target part to constrain the AnyGrasp [7] model to predict part-level grasp poses. Then, the gripper poses in the motion sequences can be calculated based on the information about the joint and manipulation method. For the button part, we take its surface geometrical center as the contact point and control the closed gripper + +Table 2. Baseline model's performances on three challenge tracks of CheckManual benchmark. The results of Track 1 are reported in the "Part Alignment Success Rate / Task Planning Success Rate" format. For Track 2 and 3, their results are reported in "Task Completion Rate / Task Success Rate" format. The "w/o manual" means ablating appliance manuals input for the above methods. + +
BASELINE MODEL图图图图图图图图图图图Total
Track 1: Manual-CAD-Appliance Aligned Planing
ManualPlan w/o manual22.96 / 22.22 - / 4.2139.47 / 22.82 - / 2.6818.52 / 14.29 - / 20.0072.84 / 60.19 - / 48.154.17 / 3.91 - / 0.788.13 / 8.64 - / 2.4794.44 / 68.00 - / 36.0047.22 / 30.56 - / 6.9441.67 / 40.62 - / 15.6221.15 / 14.39 - / 4.3215.46 / 9.96 - / 2.8527.32 / 20.70 - / 7.99
Track 2: Manual & CAD based Manipulation
ManualPlan + Primitive Actions w/o manual2.39 / 1.15 1.05 / 0.387.12 / 4.70 1.76 / 1.3413.63 / 11.43 21.39 / 14.290.46 / 0.00 0.46 / 0.000.78 / 0.78 0.78 / 0.787.91 / 3.70 3.50 / 1.2310.00 / 4.00 2.00 / 0.0011.11 / 4.86 5.09 / 1.393.12 / 0.00 1.04 / 0.002.32 / 1.44 2.28 / 2.162.56 / 1.42 1.69 / 1.424.50 / 2.39 2.47 / 1.37
Track 3: Pure Manual-based Manipulation
ManualPlan + VoxPoser [10] w/o manual6.11 / 1.92 0.99 / 0.387.18 / 4.70 1.39 / 0.677.10 / 5.71 0.57 / 0.003.01 / 1.85 0.46 / 0.003.12 / 3.12 1.56 / 1.567.99 / 3.70 1.54 / 0.620.00 / 0.00 0.00 / 0.0010.66 / 6.25 2.50 / 0.696.75 / 3.12 0.00 / 0.000.48 / 0.00 0.00 / 0.002.96 / 2.14 0.98 / 0.715.18 / 2.87 1.11 / 0.55
+ +to push the button from the normal direction of the contact point. For the knob part, we control the open gripper to get close to the knob from the normal direction of its surface geometrical center and close the gripper to grasp. After the gripper successfully grasps the knob, it will rotate planned degrees around the normal of its surface geometrical center. CAD-free Manipulation Policy. We integrate a recently proposed open-vocabulary manipulation model VoxPoser [10] into our ManualPlan to complete the CAD-free manipulation task. In this framework, ManualPlan and VoxPoser cooperate like the cerebrum and cerebellum in the human brain. ManualPlan reads manuals, makes high-level planning, and aligns observed components with manual-depicted parts. Then, it grounds the target component and demonstrates the corresponding manipulation method in language format, which is transferred to VoxPoser with camera RGB-D observation to execute low-level manipulation action on the appliance. + +# 5. Experiments + +# 5.1. Implementation Details + +When creating the CheckManual dataset, we utilize GPT-4o [1] as the large language model and multimodal large language model. Following previous work [26], we use the Franka Panda flying gripper in the simulation. The groundingDINO [18] model is utilized to ground the bounding boxes of appliance components. The predicted bounding boxes prompt SAM [13] model to generate component masks. FoundationPose [25] helps us estimate appliance 6-DoF pose in the camera observation. In the evaluation, manual comprehension and manipulation planning processes are assisted by GPT-4o [1] model. + +# 5.2. Simulation Evaluation + +We evaluate our proposed ManualPlan model and ManualPlan-driven framework using the CheckManual benchmark, which employs SAPIEN [27] as the simulation. This environment provides a realistic and flexible simulation for testing robotic manipulation tasks across a + +variety of appliances. The experimental results on different tasks and appliance categories are summarized in Table 2. From the results of Track 1, we observe that the ManualPlan model is capable of generating a part of correct task plans based on multi-page manuals. However, even the most advanced Multi-Modal Large Language Models (MLLMs) still face significant challenges in aligning manual part annotations with the corresponding CAD models of appliances, which greatly limits the overall performance of ManualPlan in task planning. For Track 2 and Track 3, the performance of the ManualPlan-driven framework highlights the difficulties posed by long-horizon, manual-based manipulation tasks. After analyzing failure cases, we identified that the accumulation of errors in part identification, task planning, and multi-step execution are the primary causes of failure. Additionally, we conducted an ablation study by removing the manual (i.e., w/o manual) to investigate its impact on task performance. As shown in Table 2, the manual plays a crucial role in guiding the robot to accurately plan and execute manipulation tasks, underscoring its importance in completing appliance manipulation tasks. + +# 5.3. Real Robot Deployment + +To demonstrate the potential of our approach in real-world applications, we conducted real-world robot experiments following the task definition of Track 3 from the Check-Manual benchmark. The hardware setup involved a Franka Panda robot arm and a RealSense D415 RGB-D camera mounted on the end-effector. Considering error accumulation in the long-horizon execution, we follow [11] to cache correct manipulation actions in every step. + +# 6. Conclusion + +In this work, we introduce CheckManual, the first benchmark for manual-based appliance manipulation, and define new challenges, metrics, and simulator environments for model evaluation. Additionally, we present the first manual-based manipulation planning model ManualPlan to set up a group of baselines for the CheckManual benchmark. + +# Acknowledgment + +This work was supported by the National Youth Talent Support Program (8200800081) and National Natural Science Foundation of China (No. 62376006). We also thank Yang Liu and Zhuoqun Xu from Polar Mirage for their technical support on assets. + +# References + +[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 8 +[2] Michael Ahn, Anthony Brohan, Noah Brown, Yevgen Chebotar, Omar Cortes, Byron David, Chelsea Finn, Chuyuan Fu, Keerthana Gopalakrishnan, Karol Hausman, et al. Do as i can, not as i say: Grounding language in robotic affordances. arXiv preprint arXiv:2204.01691, 2022. 2 +[3] Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Xi Chen, Krzysztof Choromanski, Tianli Ding, Danny Driess, Avinava Dubey, Chelsea Finn, Pete Florence, Chuyuan Fu, Montse Gonzalez Arenas, Keerthana Gopalakrishnan, Kehang Han, Karol Hausman, Alexander Herzog, Jasmine Hsu, Brian Ichter, Alex Irpan, Nikhil Joshi, Ryan Julian, Dmitry Kalashnikov, Yuheng Kuang, Isabel Leal, Lisa Lee, Tsang-Wei Edward Lee, Sergey Levine, Yao Lu, Henryk Michalewski, Igor Mordatch, Karl Pertsch, Kanishka Rao, Krista Reymann, Michael Ryoo, Grecia Salazar, Pannag Sanketi, Pierre Sermanet, Jaspiar Singh, Anikait Singh, Radu Soricut, Huong Tran, Vincent Vanhoucke, Quan Vuong, Ayzaan Wahid, Stefan Welker, Paul Wohlhart, Jialin Wu, Fei Xia, Ted Xiao, Peng Xu, Sichun Xu, Tianhe Yu, and Brianna Zitkovich. Rt-2: Vision-language-action models transfer web knowledge to robotic control, 2023. 1, 2 +[4] Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Joseph Dabis, Chelsea Finn, Keerthana Gopalakrishnan, Karol Hausman, Alex Herzog, Jasmine Hsu, Julian Ibarz, Brian Ichter, Alex Irpan, Tomas Jackson, Sally Jesmonth, Nikhil J Joshi, Ryan Julian, Dmitry Kalashnikov, Yuheng Kuang, Isabel Leal, Kuang-Huei Lee, Sergey Levine, Yao Lu, Utsav Malla, Deeksha Manjunath, Igor Mordatch, Ofir Nachum, Carolina Parada, Jodilyn Peralta, Emily Perez, Karl Pertsch, Jornell Quiambao, Kanishka Rao, Michael Ryoo, Grecia Salazar, Pannag Sanketi, Kevin Sayed, Jaspiar Singh, Sumedh Sontakke, Austin Stone, Clayton Tan, Huong Tran, Vincent Vanhoucke, Steve Vega, Quan Vuong, Fei Xia, Ted Xiao, Peng Xu, Sichun Xu, Tianhe Yu, and Brianna Zitkovich. Rt-1: Robotics transformer for real-world control at scale, 2023. 1, 2 +[5] Wenzhe Cai, Siyuan Huang, Guangran Cheng, Yuxing Long, Peng Gao, Changyin Sun, and Hao Dong. Bridging zero-shot object navigation and foundation models through pixel-guided navigation skill. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 5228-5234. IEEE, 2024. 2 +[6] Martin Ester, Hans-Peter Kriegel, Jörg Sander, Xiaowei Xu, et al. A density-based algorithm for discovering clusters in + +large spatial databases with noise. In kdd, pages 226-231, 1996. 5 +[7] Hao-Shu Fang, Chenxi Wang, Hongjie Fang, Minghao Gou, Jirong Liu, Hengxu Yan, Wenhai Liu, Yichen Xie, and Cewu Lu. Anygrasp: Robust and efficient grasp perception in spatial and temporal domains. IEEE Transactions on Robotics (T-RO), 2023. 1, 7 +[8] Jiayuan Gu, Sean Kirmani, Paul Wohlhart, Yao Lu, Montserrat Gonzalez Arenas, Kanishka Rao, Wenhao Yu, Chuyuan Fu, Keerthana Gopalakrishnan, Zhuo Xu, Priya Sundaresan, Peng Xu, Hao Su, Karol Hausman, Chelsea Finn, Quan Vuong, and Ted Xiao. Rt-trajectory: Robotic task generalization via hindsight trajectory sketches, 2023. 2 +[9] Haoxu Huang, Fanqi Lin, Yingdong Hu, Shengjie Wang, and Yang Gao. Copa: General robotic manipulation through spatial constraints of parts with foundation models. arXiv preprint arXiv:2403.08248, 2024. 2 +[10] Wenlong Huang, Chen Wang, Ruohan Zhang, Yunzhu Li, Jiajun Wu, and Li Fei-Fei. Voxposer: Composable 3d value maps for robotic manipulation with language models. arXiv preprint arXiv:2307.05973, 2023. 1, 2, 7, 8 +[11] Wenlong Huang, Chen Wang, Yunzhu Li, Ruohan Zhang, and Li Fei-Fei. Rekep: Spatio-temporal reasoning of relational keypoint constraints for robotic manipulation. arXiv preprint arXiv:2409.01652, 2024. 1, 2, 8 +[12] Moo Jin Kim, Karl Pertsch, Siddharth Karamcheti, Ted Xiao, Ashwin Balakrishna, Suraj Nair, Rafael Rafailov, Ethan Foster, Grace Lam, Pannag Sanketi, Quan Vuong, Thomas Kollar, Benjamin Burchfiel, Russ Tedrake, Dorsa Sadigh, Sergey Levine, Percy Liang, and Chelsea Finn. Openvla: An open-source vision-language-action model. arXiv preprint arXiv:2406.09246, 2024. 1, 2 +[13] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dollar, and Ross Girshick. Segment anything. arXiv:2304.02643, 2023. 7, 8 +[14] Isabel Leal, Krzysztof Choromanski, Deepali Jain, Avinava Dubey, Jake Varley, Michael Ryoo, Yao Lu, Frederick Liu, Vikas Sindhwani, Quan Vuong, et al. Sara-rt: Scaling up robotics transformers with self-adaptive robust attention. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 6920–6927. IEEE, 2024. 2 +[15] Xinghang Li, Minghuan Liu, Hanbo Zhang, Cunjun Yu, Jie Xu, Hongtao Wu, Chilam Cheang, Ya Jing, Weinan Zhang, Huaping Liu, et al. Vision-language foundation models as effective robot imitators. arXiv preprint arXiv:2311.01378, 2023. 2 +[16] Xiaqi Li, Mingxu Zhang, Yiran Geng, Haoran Geng, Yuxing Long, Yan Shen, Renrui Zhang, Jiaming Liu, and Hao Dong. Maniplm: Embodied multimodal large language model for object-centric robotic manipulation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18061-18070, 2024. 2 +[17] Shanchuan Lin, Anran Wang, and Xiao Yang. Sdxlightning: Progressive adversarial diffusion distillation, 2024. 5 + +[18] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Qing Jiang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, and Lei Zhang. Grounding dino: Marrying dino with grounded pre-training for open-set object detection, 2024. 7, 8 +[19] Yuxing Long, Wenzhe Cai, Hongcheng Wang, Guanqi Zhan, and Hao Dong. Instructnav: Zero-shot system for generic instruction navigation in unexplored environment. arXiv preprint arXiv:2406.04882, 2024. 1 +[20] Yuxing Long, Xiaqi Li, Wenzhe Cai, and Hao Dong. Discuss before moving: Visual language navigation via multi-expert discussions. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 17380-17387. IEEE, 2024. 2 +[21] Jianlan Luo, Charles Xu, Jeffrey Wu, and Sergey Levine. Hil-serl: Precise and dexterous robotic manipulation via human-in-the-loop reinforcement learning, 2024. 1 +[22] Kaichun Mo, Shilin Zhu, Angel X. Chang, Li Yi, Subarna Tripathi, Leonidas J. Guibas, and Hao Su. PartNet: A large-scale benchmark for fine-grained and hierarchical part-level 3D object understanding. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 3 +[23] Abhilash Nandy, Soumya Sharma, Shubham Maddhashiya, Kapil Sachdeva, Pawan Goyal, and NIloy Ganguly. Question answering over electronic devices: A new benchmark dataset and a multi-task learning based QA framework. In Findings of the Association for Computational Linguistics: EMNLP 2021, pages 4600-4609, Punta Cana, Dominican Republic, 2021. Association for Computational Linguistics. 1, 2 +[24] Abhilash Nandy, Soumya Sharma, Shubham Maddhashiya, Kapil Sachdeva, Pawan Goyal, and NIloy Ganguly. Question answering over electronic devices: A new benchmark dataset and a multi-task learning based QA framework. In Findings of the Association for Computational Linguistics: EMNLP 2021, pages 4600-4609, Punta Cana, Dominican Republic, 2021. Association for Computational Linguistics. 2 +[25] Bowen Wen, Wei Yang, Jan Kautz, and Stan Birchfield. FoundationPose: Unified 6d pose estimation and tracking of novel objects. In CVPR, 2024. 7, 8 +[26] Ruihai Wu, Yan Zhao, Kaichun Mo, Zizheng Guo, Yian Wang, Tianhao Wu, Qingnan Fan, Xuelin Chen, Leonidas Guibas, and Hao Dong. VAT-mart: Learning visual action trajectory proposals for manipulating 3d ARTiculated objects. In International Conference on Learning Representations, 2022. 8 +[27] Fanbo Xiang, Yuzhe Qin, Kaichun Mo, Yikuan Xia, Hao Zhu, Fangchen Liu, Minghua Liu, Hanxiao Jiang, Yifu Yuan, He Wang, Li Yi, Angel X. Chang, Leonidas J. Guibas, and Hao Su. SAPIEN: A simulated part-based interactive environment. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 8 +[28] Jianwei Yang, Hao Zhang, Feng Li, Xueyan Zou, Chunyuan Li, and Jianfeng Gao. Set-of-mark prompting unleashes extraordinary visual grounding in gpt-4v. arXiv preprint arXiv:2310.11441, 2023. 7 +[29] Senqiao Yang, Jiaming Liu, Ray Zhang, Mingjie Pan, Zoey Guo, Xiaqi Li, Zehui Chen, Peng Gao, Yandong Guo, and + +Shanghang Zhang. Lidar-llm: Exploring the potential of large language models for 3d lidar understanding. arXiv preprint arXiv:2312.14074, 2023. 2 +[30] Jiyao Zhang, Mingdong Wu, and Hao Dong. Generative category-level object pose estimation via diffusion models. Advances in Neural Information Processing Systems, 36: 54627-54644, 2023. 7 +[31] Jiyao Zhang, Weiyao Huang, Bo Peng, Mingdong Wu, Fei Hu, Zijian Chen, Bo Zhao, and Hao Dong. Omni6dpose: A benchmark and model for universal 6d object pose estimation and tracking. In European Conference on Computer Vision, pages 199-216. Springer, 2024. 7 +[32] Liang Zhang, Anwen Hu, Jing Zhang, Shuo Hu, and Qin Jin. Mpmqa: multimodal question answering on product manuals. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 13958-13966, 2023. 1, 2 \ No newline at end of file diff --git a/2025/CheckManual_ A New Challenge and Benchmark for Manual-based Appliance Manipulation/images.zip b/2025/CheckManual_ A New Challenge and Benchmark for Manual-based Appliance Manipulation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..2d1760f08c74abf7429cd69eb111645b31e0d473 --- /dev/null +++ b/2025/CheckManual_ A New Challenge and Benchmark for Manual-based Appliance Manipulation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:105c2f0f3e75a98d22876e7fcb3695df36a456af25d655771efa5f39fe7afaff +size 446872 diff --git a/2025/CheckManual_ A New Challenge and Benchmark for Manual-based Appliance Manipulation/layout.json b/2025/CheckManual_ A New Challenge and Benchmark for Manual-based Appliance Manipulation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..02663cd6bf1401c5c14daee26cfb9753730260b0 --- /dev/null +++ b/2025/CheckManual_ A New Challenge and Benchmark for Manual-based Appliance Manipulation/layout.json @@ -0,0 +1,6374 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 141, + 103, + 471, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 103, + 471, + 140 + ], + "spans": [ + { + "bbox": [ + 141, + 103, + 471, + 140 + ], + "type": "text", + "content": "CheckManual: A New Challenge and Benchmark for Manual-based Appliance Manipulation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 72, + 160, + 537, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 160, + 537, + 190 + ], + "spans": [ + { + "bbox": [ + 72, + 160, + 537, + 190 + ], + "type": "text", + "content": "Yuxing Long" + }, + { + "bbox": [ + 72, + 160, + 537, + 190 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 72, + 160, + 537, + 190 + ], + "type": "text", + "content": ", Jiyao Zhang" + }, + { + "bbox": [ + 72, + 160, + 537, + 190 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 72, + 160, + 537, + 190 + ], + "type": "text", + "content": ", Mingjie Pan" + }, + { + "bbox": [ + 72, + 160, + 537, + 190 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 72, + 160, + 537, + 190 + ], + "type": "text", + "content": ", Tianshu Wu" + }, + { + "bbox": [ + 72, + 160, + 537, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 72, + 160, + 537, + 190 + ], + "type": "text", + "content": ", Taewhan Kim" + }, + { + "bbox": [ + 72, + 160, + 537, + 190 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 72, + 160, + 537, + 190 + ], + "type": "text", + "content": " and Hao Dong" + }, + { + "bbox": [ + 72, + 160, + 537, + 190 + ], + "type": "inline_equation", + "content": "^{1,2\\dagger}" + }, + { + "bbox": [ + 72, + 160, + 537, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 72, + 160, + 537, + 190 + ], + "type": "text", + "content": "CFCS, School of Computer Science, Peking University " + }, + { + "bbox": [ + 72, + 160, + 537, + 190 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 72, + 160, + 537, + 190 + ], + "type": "text", + "content": "PKU-Agibot Lab" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 151, + 217, + 200, + 229 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 217, + 200, + 229 + ], + "spans": [ + { + "bbox": [ + 151, + 217, + 200, + 229 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 243, + 296, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 243, + 296, + 518 + ], + "spans": [ + { + "bbox": [ + 55, + 243, + 296, + 518 + ], + "type": "text", + "content": "Correct use of electrical appliances has significantly improved human life quality. Unlike simple tools that can be manipulated with common sense, different parts of electrical appliances have specific functions defined by manufacturers. If we want the robot to heat bread by microwave, we should enable them to review the microwave's manual first. From the manual, it can learn about component functions, interaction methods, and representative task steps about appliances. However, previous manual-related works remain limited to question-answering tasks while existing manipulation researchers ignore the manual's important role and fail to comprehend multi-page manuals. In this paper, we propose the first manual-based appliance manipulation benchmark CheckManual. Specifically, we design a large model-assisted human-revised data generation pipeline to create manuals based on CAD appliance models. With these manuals, we establish novel manual-based manipulation challenges, metrics, and simulator environments for model performance evaluation. Furthermore, we propose the first manual-based manipulation planning model ManualPlan to set up a group of baselines for the CheckManual benchmark. Our project page is available at https://sites.google.com/view/checkmanual." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 536, + 135, + 548 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 536, + 135, + 548 + ], + "spans": [ + { + "bbox": [ + 56, + 536, + 135, + 548 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 557, + 295, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 557, + 295, + 687 + ], + "spans": [ + { + "bbox": [ + 55, + 557, + 295, + 687 + ], + "type": "text", + "content": "Since the twentieth century, the emergence of appliances has significantly changed human society. In daily life, we extensively use various electrical appliances to assist our living and working. In recent years, robots have demonstrated enormous potential and value in areas like component assembly [21], goods sorting [7], and indoor navigation [19]. Imagine if we could instruct robots to operate appliances to serve us, such as baking cakes with ovens, washing clothes with washing machines, and making juice with blenders, we would be further liberated from tedious household chores and have more free time." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": "However, appliances are different from simple tools like hammers or teapots that can be correctly used based on" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 318, + 217, + 552, + 311 + ], + "blocks": [ + { + "bbox": [ + 318, + 217, + 552, + 311 + ], + "lines": [ + { + "bbox": [ + 318, + 217, + 552, + 311 + ], + "spans": [ + { + "bbox": [ + 318, + 217, + 552, + 311 + ], + "type": "image", + "image_path": "ab7c76859152a155849ee39e93795a72e549867bbcb954f049c26fb6c36d452e.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 319, + 316, + 552, + 414 + ], + "blocks": [ + { + "bbox": [ + 319, + 316, + 552, + 414 + ], + "lines": [ + { + "bbox": [ + 319, + 316, + 552, + 414 + ], + "spans": [ + { + "bbox": [ + 319, + 316, + 552, + 414 + ], + "type": "image", + "image_path": "e058bbca83b265920b4e7038e9411d4e7269fc9a798759b7a4b842c4acbe7641.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 423, + 555, + 490 + ], + "lines": [ + { + "bbox": [ + 313, + 423, + 555, + 490 + ], + "spans": [ + { + "bbox": [ + 313, + 423, + 555, + 490 + ], + "type": "text", + "content": "Figure 1. Comparison between appliance manipulation with and without manual. If the manual is absent, the concrete functions of the appliance part are unknown. As shown in the top image, the robot feels confused about tasks such as opening the microwave door. To correctly use the appliance, the robot must comprehend the manual content and follow its guidance to work." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 492, + 555, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 492, + 555, + 685 + ], + "spans": [ + { + "bbox": [ + 313, + 492, + 555, + 685 + ], + "type": "text", + "content": "common sense. One appliance may have multiple movable components and each component has specific functions even if they are the same type. Taking a microwave as an example, one button may control the temperature while the other button may set the time. Such function information is difficult to infer according to the appliance's visual appearance or human common sense. This time, we need to check the user manuals to learn about component functions and supporting tasks. Existing manipulation models like [4][3][12][10][11] already can adopt simple tools to do easy tasks but cannot read multi-page manuals not to say using the appliances to complete long-horizon complex tasks. Besides, previous manual-related research and academic datasets [23][32] are all collected by the NLP researchers1, which only focuses on page retrieval and question-answering tasks rather than manipulation. These" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 313, + 693, + 555, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 693, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 693, + 555, + 713 + ], + "type": "text", + "content": "1Note that the copyright of real manual belongs to the company. Using content from the manual for academic research requires authorization." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22595" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "content": "academic datasets manual datasets also do not provide appliance CAD models and cannot be used for quantitative manipulation evaluation in the simulator. Therefore, there are no manual-based manipulation evaluation datasets and benchmarks, nor manipulation models capable of following manuals to use appliances at present." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 145, + 296, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 145, + 296, + 430 + ], + "spans": [ + { + "bbox": [ + 56, + 145, + 296, + 430 + ], + "type": "text", + "content": "To address this problem, we propose the first manual-based manipulation benchmark CheckManual. In the benchmark, we collect new evaluation data, design novel manipulation challenges, and propose new baseline models. Specifically, we first design an LLM-assisted human-revised manual creation pipeline based on appliance CAD models. For every CAD model, this pipeline can follow our surveyed real manual analysis to generate diverse appliance manuals with different part functions, manipulation demonstrations, and page layouts. Then, we set up three manual-based appliance manipulation challenge tracks including manipulation planning, CAD-assisted manipulation, and CAD-free manipulation. They progressively cover the key steps in manual-based appliance manipulation, comprehensively evaluating the model's ability to operate appliances with manuals. Furthermore, we propose a manual-based manipulation planning method ManualPlan, and create a set of CAD-assisted primitive actions for appliance manipulation. The ManualPlan aligns manuals with appliances and plans detailed manipulation steps according to manual content, which can control primitive actions or open-vocabulary manipulation models to physically interact with appliances. These also act as the baseline methods for our CheckManual benchmark." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 431, + 233, + 442 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 431, + 233, + 442 + ], + "spans": [ + { + "bbox": [ + 67, + 431, + 233, + 442 + ], + "type": "text", + "content": "In this paper, our main contributions are:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 443, + 295, + 574 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 55, + 443, + 295, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 443, + 295, + 490 + ], + "spans": [ + { + "bbox": [ + 55, + 443, + 295, + 490 + ], + "type": "text", + "content": "- We collect the first CAD model-aligned appliance manual dataset CheckManual through our devised large model-assisted pipeline. Every data collection stage undergoes careful human verification to guarantee correctness." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 491, + 295, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 491, + 295, + 537 + ], + "spans": [ + { + "bbox": [ + 55, + 491, + 295, + 537 + ], + "type": "text", + "content": "- We design novel challenges for manual-based appliance manipulation considering different difficulty levels and conditions. Furthermore, we establish the evaluation environment in the simulator for the benchmark." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 539, + 295, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 539, + 295, + 574 + ], + "spans": [ + { + "bbox": [ + 55, + 539, + 295, + 574 + ], + "type": "text", + "content": "- We propose the first manual-based manipulation planning model ManualPlan to set up a group of baselines for three-level challenge tracks in CheckManual." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 579, + 142, + 591 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 579, + 142, + 591 + ], + "spans": [ + { + "bbox": [ + 55, + 579, + 142, + 591 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 599, + 209, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 599, + 209, + 612 + ], + "spans": [ + { + "bbox": [ + 55, + 599, + 209, + 612 + ], + "type": "text", + "content": "2.1. Manipulation Large Models" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 618, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 295, + 713 + ], + "type": "text", + "content": "Large models pre-trained on Internet-scale text and image data emerge powerful capabilities for task planning and common sense reasoning. Therefore, leveraging large models has become a hotspot in embodied areas such as autonomous driving [29], navigation [20] [5] and manipulation. According to training and inference forms, existing manipulation methods can be divided into two main categories. The first category of work focuses on fine-tuning" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 555, + 347 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 347 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 347 + ], + "type": "text", + "content": "large models with manipulation trajectories collected from the simulator environment and real-world. The representative models are Google RT-series from RT-1 [4], RT-2 [3], and RT-Trajectory [8] to STAR-RT [14]. These models take vision observation, robot states, and task instruction as input, and directly predict the low-level actions by multimodal large models. Similar open-sourced work includes ManipLLM [16], RoboFlamingo [15] and OpenVLA [12]. The other category of work makes efforts to effectively integrate large models' zero-shot inferences into the manipulation planning and decision process. SayCan [2] first utilizes a multimodal large model to score a group of kills to make progress toward the high-level task. Voxposer [10] drives a multimodal large model to create 3D value maps to guide the robot toward the target object part. CoPa [9] and ReKep [11] leverage multimodal large models to generate spatial-temporal constraints for complex tasks like \"insert the flower into vase\" or \"pour tea into water\". However, all these large model-based methods can only make plans based on their internal common sense. Considering heating the bread with the microwave, they even struggle to figure out the temperature knob and timer knob, not to say complete the whole task." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 353, + 484, + 365 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 353, + 484, + 365 + ], + "spans": [ + { + "bbox": [ + 313, + 353, + 484, + 365 + ], + "type": "text", + "content": "2.2. Appliance Manual-based Tasks" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 371, + 556, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 371, + 556, + 635 + ], + "spans": [ + { + "bbox": [ + 313, + 371, + 556, + 635 + ], + "type": "text", + "content": "NLP researchers have recognized the manual's importance in addressing user questions. TechQA [24] collects 1400 user questions from online forums and annotates the corresponding answers from IBM technical documents to set up Document-based Question Answering task. S10 QA [23] manually creates 904 QA pairs based on the manual of the Samsung S10 phone. Nandy et al. [23] also collect 950 QA pairs based on the Samsung Smart TV manual to generate Smart TV/Remote QA dataset. These two datasets are utilized to establish a Manual-based Question Answering benchmark. The above datasets are all pure text corpus. PM209 [32] collects 209 product manuals from 27 well-known consumer electronic brands to set up Manual Page Retrieval task and Manual-based Multimodal Question Answering task. Although these datasets and tasks are related to the appliance manual, they are all designed for NLP evaluation and lack corresponding appliance CAD models. Therefore, manual-based manipulation benchmarks cannot be directly established on these existing related researches. We create the first manual & CAD model-aligned dataset and design novel benchmarks and baseline models for manual-based appliance manipulation." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 643, + 555, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 643, + 555, + 670 + ], + "spans": [ + { + "bbox": [ + 313, + 643, + 555, + 670 + ], + "type": "text", + "content": "3. Manual-based Manipulation Benchmark CheckManual" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "content": "In this part, we will introduce the creation process of our CheckManual dataset in Section 3.1, analyze the numerical and visual statistics about CheckManual in Section 3.2, and" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22596" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 70, + 138, + 293 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 138, + 293 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 138, + 293 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 138, + 293 + ], + "type": "image", + "image_path": "5843d26d66f77e8664a28ca752bc3a37daa64329e2e35a744b0fbdebf840fd3b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 301, + 555, + 347 + ], + "lines": [ + { + "bbox": [ + 55, + 301, + 555, + 347 + ], + "spans": [ + { + "bbox": [ + 55, + 301, + 555, + 347 + ], + "type": "text", + "content": "Figure 2. Generation workflow of CheckManual dataset. In the leftmost part (Section 3.1.1), we analyze real manuals to learn about their formats and collect different categories of appliance CAD models. The middle part (Section 3.1.2 - 3.1.4) demonstrates the creation of manual materials, including appliance creation, task generation, and figure design. Human verifies every step to guarantee correctness. Based on this information, the rightmost part (Section 3.1.5) generates appliance manuals with diverse formats through the LATEX codes." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 145, + 70, + 267, + 293 + ], + "blocks": [ + { + "bbox": [ + 145, + 70, + 267, + 293 + ], + "lines": [ + { + "bbox": [ + 145, + 70, + 267, + 293 + ], + "spans": [ + { + "bbox": [ + 145, + 70, + 267, + 293 + ], + "type": "image", + "image_path": "7c4218757fdc77571c474a4c8935b8677945450d644d481960667ee100db0ae5.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 273, + 70, + 425, + 293 + ], + "blocks": [ + { + "bbox": [ + 273, + 70, + 425, + 293 + ], + "lines": [ + { + "bbox": [ + 273, + 70, + 425, + 293 + ], + "spans": [ + { + "bbox": [ + 273, + 70, + 425, + 293 + ], + "type": "image", + "image_path": "869c06a7cf586df7fd7e134fb4bae8043a6dfde432fa60dcb7c390ce1cb3f97a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 425, + 70, + 553, + 293 + ], + "blocks": [ + { + "bbox": [ + 425, + 70, + 553, + 293 + ], + "lines": [ + { + "bbox": [ + 425, + 70, + 553, + 293 + ], + "spans": [ + { + "bbox": [ + 425, + 70, + 553, + 293 + ], + "type": "image", + "image_path": "c83d961fe54a2f97aa3cffdb04305b6350b706e86fcd95f59e983e896182dfc2.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 357, + 295, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 357, + 295, + 380 + ], + "spans": [ + { + "bbox": [ + 55, + 357, + 295, + 380 + ], + "type": "text", + "content": "define the evaluation for robot manipulation based on these manuals in Section 3.3." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 389, + 209, + 403 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 389, + 209, + 403 + ], + "spans": [ + { + "bbox": [ + 55, + 389, + 209, + 403 + ], + "type": "text", + "content": "3.1. Appliance Manual Creation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 408, + 296, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 408, + 296, + 468 + ], + "spans": [ + { + "bbox": [ + 55, + 408, + 296, + 468 + ], + "type": "text", + "content": "To create an open-source available manual dataset, We design a large model-driven human-verified manual generation workflow based on appliance CAD models. Figure 2 demonstrates the generation workflow of the CheckManual dataset, which will be detailed in the following." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 483, + 140, + 496 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 483, + 140, + 496 + ], + "spans": [ + { + "bbox": [ + 55, + 483, + 140, + 496 + ], + "type": "text", + "content": "3.1.1 Preparation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 503, + 295, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 503, + 295, + 550 + ], + "spans": [ + { + "bbox": [ + 55, + 503, + 295, + 550 + ], + "type": "text", + "content": "Before formally starting manual creation, we analyze a stack of real appliance manuals to learn about real manual formats and collect available articulated appliance CAD models for manual creation." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 552, + 295, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 552, + 295, + 694 + ], + "spans": [ + { + "bbox": [ + 55, + 552, + 295, + 694 + ], + "type": "text", + "content": "Analysis of Real Appliance Manuals We collect 110 appliance manual PDFs from the Internet, which cover 11 categories of our dataset involved appliances. These manuals come from different countries and regions, including China, the European Union, Japan, and India. Then we manually analyze each manual from Functional Part Annotation Type, Guidance of Functional Part Operation, and Task Expression Format these three aspects. We find that the functional parts are usually explained in the point-line annotation formats. These lines connect the points on appliance parts and the function names or IDs of parts as shown in the middle of Figure 2. The guidance of functional" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 357, + 555, + 500 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 357, + 555, + 500 + ], + "spans": [ + { + "bbox": [ + 313, + 357, + 555, + 500 + ], + "type": "text", + "content": "part operation is commonly represented in formats: 1 Pure text description like \"Rotate temperature knob " + }, + { + "bbox": [ + 313, + 357, + 555, + 500 + ], + "type": "inline_equation", + "content": "90^{\\circ}" + }, + { + "bbox": [ + 313, + 357, + 555, + 500 + ], + "type": "text", + "content": " clockwise\". 2 Close-up figures of the manipulated part, such as the close and open states of the red microwave door in Figure 2. 3 Part motion decomposition diagram that shows the dynamic change of part during its motion process. Figure 2 shows one example of opening the white mixer lid. 4 Motion trajectory of the contact point on the part, like manipulating Figure 2's blue mixer slider from right to left. Besides, the manipulation tasks in the real manuals are often depicted in the text, bullet lists, ordered lists, and tables with one-column or multi-columns." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 501, + 555, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 501, + 555, + 597 + ], + "spans": [ + { + "bbox": [ + 313, + 501, + 555, + 597 + ], + "type": "text", + "content": "Appliance CAD Models Our appliance CAD models come from the PartNet-Mobility [22] Dataset, which contains common categories of household appliances CAD models, including microwaves, dishwashers, and refrigerators. Appliance models in this dataset are annotated with manipulable parts, such as buttons, knobs, levers, sliders, and so on. In total, we collect 11 categories, 182 CAD models, which cover widely-used household appliances in our daily lives." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 611, + 430, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 611, + 430, + 624 + ], + "spans": [ + { + "bbox": [ + 313, + 611, + 430, + 624 + ], + "type": "text", + "content": "3.1.2 Appliance Creation" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 630, + 554, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 554, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 554, + 689 + ], + "type": "text", + "content": "With the appliance CAD models, we can create concrete appliances by annotating each movable part with a specific function name (e.g., temperature knob) and corresponding function states (e.g., clockwise " + }, + { + "bbox": [ + 313, + 630, + 554, + 689 + ], + "type": "inline_equation", + "content": "60^{\\circ} - 120^{\\circ}C" + }, + { + "bbox": [ + 313, + 630, + 554, + 689 + ], + "type": "text", + "content": " and clockwise " + }, + { + "bbox": [ + 313, + 630, + 554, + 689 + ], + "type": "inline_equation", + "content": "120^{\\circ} - 160^{\\circ}C" + }, + { + "bbox": [ + 313, + 630, + 554, + 689 + ], + "type": "text", + "content": "). This process is detailed in the following." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 690, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 690, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 690, + 555, + 713 + ], + "type": "text", + "content": "Part Function Name Annotation Considering part function names are closely related to the appliance layout and" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 66, + 702, + 289, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 702, + 289, + 713 + ], + "spans": [ + { + "bbox": [ + 66, + 702, + 289, + 713 + ], + "type": "text", + "content": "2We use examples from our ChcekManual to avoid any infringement" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22597" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 70, + 227, + 187 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 227, + 187 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 227, + 187 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 227, + 187 + ], + "type": "image", + "image_path": "d17051db036ef41e2b5673d328e01e347ba027b862a46bf0a159ede68e60c464.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 108, + 201, + 499, + 211 + ], + "lines": [ + { + "bbox": [ + 108, + 201, + 499, + 211 + ], + "spans": [ + { + "bbox": [ + 108, + 201, + 499, + 211 + ], + "type": "text", + "content": "Figure 3. One example of appliance manual from our CheckManual dataset. Please zoom in for details." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 227, + 70, + 306, + 187 + ], + "blocks": [ + { + "bbox": [ + 227, + 70, + 306, + 187 + ], + "lines": [ + { + "bbox": [ + 227, + 70, + 306, + 187 + ], + "spans": [ + { + "bbox": [ + 227, + 70, + 306, + 187 + ], + "type": "image", + "image_path": "7afe3484708466aa39eea7c029df3af9dd4d5644a2736d57522c0e53dbef9dd6.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 306, + 70, + 387, + 187 + ], + "blocks": [ + { + "bbox": [ + 306, + 70, + 387, + 187 + ], + "lines": [ + { + "bbox": [ + 306, + 70, + 387, + 187 + ], + "spans": [ + { + "bbox": [ + 306, + 70, + 387, + 187 + ], + "type": "image", + "image_path": "e07905b3275cf7d380266c9f671d0c865a5e19637e9cd9e9f761b8019bd3424f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 387, + 71, + 553, + 187 + ], + "blocks": [ + { + "bbox": [ + 387, + 71, + 553, + 187 + ], + "lines": [ + { + "bbox": [ + 387, + 71, + 553, + 187 + ], + "spans": [ + { + "bbox": [ + 387, + 71, + 553, + 187 + ], + "type": "image", + "image_path": "fc3fd8f20dbb3d1acc1dc20175535813ceb0a062a3253480a4733017957f5b36.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 224, + 294, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 224, + 294, + 429 + ], + "spans": [ + { + "bbox": [ + 54, + 224, + 294, + 429 + ], + "type": "text", + "content": "parts' visual features, we drive the annotation process by the multimodal large language model (MLLM). Before annotation, manipulable part IDs and types are automatically labeled on the appliance overview figures through point-line annotation. Taking these figures as visual prompts, we instruct the MLLM to design the function name of each part. The multimodal large model should consider the appliance type and carefully analyze part visual information to create reasonable annotations. Re-generation mechanism would be triggered when the number of annotated parts is unequal to the total number of manipulable parts. By resolving the large model's prediction with regular expressions, we can obtain a dictionary recording the one-to-one relationships between part ID and corresponding function name. One such dictionary can be used to generate one appliance. We can create multiple appliances from one CAD model by repeating the above annotation process." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 434, + 295, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 434, + 295, + 576 + ], + "spans": [ + { + "bbox": [ + 55, + 434, + 295, + 576 + ], + "type": "text", + "content": "Part Function State Annotation To annotate part function states, we first sample available states for each manipulable part according to their types. For buttons, we randomly sample available push times in a pre-defined range. For knobs and levers, we randomly sample several rotation degrees at intervals of " + }, + { + "bbox": [ + 55, + 434, + 295, + 576 + ], + "type": "inline_equation", + "content": "60^{\\circ}" + }, + { + "bbox": [ + 55, + 434, + 295, + 576 + ], + "type": "text", + "content": ". Other manipulable parts like doors, lids, and containers, are assigned with \"open\" and \"close\" states. With part function names and available states, we instruct the large language model (LLM) to design concrete functions for each state. This way, each manipulable part on the appliance will be annotated with multiple function states." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "text", + "content": "Human Verification about Part Functions Although large models are powerful annotation tools, they still make intolerable mistakes and thus require human verification. In the annotation about function names, MLLM sometimes fails to identify tiny visual elements on the CAD models like already-drawn button logos or scales, which may lead to mismatching between annotated function names and part visual appearance. In the annotation about function states, LLM may create an anti-commonsense design such as “" + }, + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "inline_equation", + "content": "500^{\\circ}\\mathrm{C}" + }, + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "text", + "content": " for oven temperature knob”. Therefore, we check large model-annotated function names and states" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 224, + 553, + 261 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 224, + 553, + 261 + ], + "spans": [ + { + "bbox": [ + 313, + 224, + 553, + 261 + ], + "type": "text", + "content": "through our designed human verification website (See Appendix) to revise incorrect cases. In this stage, we manually verify more than 2200 part function annotations." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 276, + 468, + 288 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 276, + 468, + 288 + ], + "spans": [ + { + "bbox": [ + 313, + 276, + 468, + 288 + ], + "type": "text", + "content": "3.1.3 Manipulation Task Creation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 295, + 553, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 295, + 553, + 354 + ], + "spans": [ + { + "bbox": [ + 313, + 295, + 553, + 354 + ], + "type": "text", + "content": "The manual often includes several representative tasks to introduce how to use the appliance. These tasks are also the core of our manual-based manipulation evaluation. We can create appliance-executable tasks based on their part function names and states." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 355, + 554, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 355, + 554, + 462 + ], + "spans": [ + { + "bbox": [ + 313, + 355, + 554, + 462 + ], + "type": "text", + "content": "Task Proposal With function names and states of movable parts on the appliances, we instruct the LLM to propose short-horizon and long-horizon manipulation tasks for each appliance. Every task step should specify the part function name and the manipulated state in a resolvable format like \"1. Front Door, Open\", \"2. Power Button, Push 1 time\" ... To improve task quality and reduce repetition, the LLM only needs to design one task in one inference based on already created tasks." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 463, + 554, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 463, + 554, + 559 + ], + "spans": [ + { + "bbox": [ + 313, + 463, + 554, + 559 + ], + "type": "text", + "content": "Human Verification about Manipulation Tasks In the task creation, LLM may also make anti-commonsense mistakes. To guarantee absolute correctness for evaluation, after the LLM completes task generation, we manually check every task step to adjust unreasonable step orders and revise incorrect steps. If the task is completely unreasonable, we will manually create a new one from scratch. Totally, we checked over 1400 manipulation tasks created by LLM." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 574, + 443, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 574, + 443, + 586 + ], + "spans": [ + { + "bbox": [ + 313, + 574, + 443, + 586 + ], + "type": "text", + "content": "3.1.4 Manual Figure Design" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 594, + 553, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 553, + 641 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 553, + 641 + ], + "type": "text", + "content": "According to our analysis of real manuals, cover, control panels, and part manipulation guidance are the main figures in the manuals. Therefore, we devise methods to create these figures." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 642, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 554, + 713 + ], + "type": "text", + "content": "Manual Cover The manual cover usually displays appliance appearance and typical usage scenes. we predefined a set of candidate camera poses to capture the appliance from different views. The views with the largest total areas of manipulable parts are selected to render RGB appliance overview figures. To synthesize custom usage scenes," + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "22598" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 72, + 177, + 137 + ], + "blocks": [ + { + "bbox": [ + 58, + 72, + 177, + 137 + ], + "lines": [ + { + "bbox": [ + 58, + 72, + 177, + 137 + ], + "spans": [ + { + "bbox": [ + 58, + 72, + 177, + 137 + ], + "type": "image", + "image_path": "b31e291063126d2190157b947b17f00848e2f16b36744699b0aa9ef97037bd52.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 103, + 152, + 113, + 162 + ], + "lines": [ + { + "bbox": [ + 103, + 152, + 113, + 162 + ], + "spans": [ + { + "bbox": [ + 103, + 152, + 113, + 162 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 178, + 70, + 281, + 143 + ], + "blocks": [ + { + "bbox": [ + 178, + 70, + 281, + 143 + ], + "lines": [ + { + "bbox": [ + 178, + 70, + 281, + 143 + ], + "spans": [ + { + "bbox": [ + 178, + 70, + 281, + 143 + ], + "type": "image", + "image_path": "6fbc62139ead53f34aeace96a753265895645e6888c749ab00a53b4c6113443c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 227, + 152, + 237, + 162 + ], + "lines": [ + { + "bbox": [ + 227, + 152, + 237, + 162 + ], + "spans": [ + { + "bbox": [ + 227, + 152, + 237, + 162 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 287, + 69, + 426, + 151 + ], + "blocks": [ + { + "bbox": [ + 287, + 69, + 426, + 151 + ], + "lines": [ + { + "bbox": [ + 287, + 69, + 426, + 151 + ], + "spans": [ + { + "bbox": [ + 287, + 69, + 426, + 151 + ], + "type": "image", + "image_path": "b0341611a065ee7f7785289612ccb86d6a3827862164f77bd8d73424fb62af5c.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 355, + 152, + 364, + 162 + ], + "lines": [ + { + "bbox": [ + 355, + 152, + 364, + 162 + ], + "spans": [ + { + "bbox": [ + 355, + 152, + 364, + 162 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 170, + 555, + 194 + ], + "lines": [ + { + "bbox": [ + 55, + 170, + 555, + 194 + ], + "spans": [ + { + "bbox": [ + 55, + 170, + 555, + 194 + ], + "type": "text", + "content": "Figure 4. CheckManual dataset statistics. The above tables and figures from left to right are (a) Numerical statistics. (b) Proportion of different appliances. (c) Distribution of manipulable components on the appliances. (d) Distribution of manipulation task step lengths." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 433, + 72, + 552, + 151 + ], + "blocks": [ + { + "bbox": [ + 433, + 72, + 552, + 151 + ], + "lines": [ + { + "bbox": [ + 433, + 72, + 552, + 151 + ], + "spans": [ + { + "bbox": [ + 433, + 72, + 552, + 151 + ], + "type": "image", + "image_path": "7412f40b543c4f6474d9a28f2d6cadbb643f557324d738371c190d292359b948.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 490, + 152, + 501, + 162 + ], + "lines": [ + { + "bbox": [ + 490, + 152, + 501, + 162 + ], + "spans": [ + { + "bbox": [ + 490, + 152, + 501, + 162 + ], + "type": "text", + "content": "(d)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 205, + 295, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 205, + 295, + 277 + ], + "spans": [ + { + "bbox": [ + 54, + 205, + 295, + 277 + ], + "type": "text", + "content": "we further add the indoor background to the appliance overview through text-to-image generation model SDXL-Lightning [17]. The background room type is chosen based on the appliance category. For example, the oven is usually put into the kitchen while the washing machine is often arranged in the laundry room." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 278, + 296, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 278, + 296, + 470 + ], + "spans": [ + { + "bbox": [ + 56, + 278, + 296, + 470 + ], + "type": "text", + "content": "Overview Diagram & Control Panel The overview diagram and control panel demonstrate the function name of every movable part on the appliance. Before annotation, we first apply the DBSCAN [6] algorithm to try clustering buttons and knobs based on their geometric centers. If clustered buttons and knobs are found, the areas with them are selected as control panels and cropped from the appliance overview figures as close-ups of control panels. Then, we adopt point-line annotation methods (Figure 2) from the real manuals to draw overview diagram and control panel based on part function names created in the previous stage. Furthermore, considering real manuals often deliver overview diagrams and control panels in line drawings, we create sketch-style figures by conducting edge detection on the RGB normal maps. This way, we can generate these two kinds of figures both in RGB style and sketch style." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 472, + 296, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 472, + 296, + 676 + ], + "spans": [ + { + "bbox": [ + 56, + 472, + 296, + 676 + ], + "type": "text", + "content": "Part Manipulation Guidance This kind of figure usually introduces the physical interaction approach for appliance manipulable parts. To generate such figures, we first render manipulable parts according to their types and visible surface areas. For buttons, knobs, levers, and drawers, we render them with the camera axis parallel to their joint axes. Other manipulable parts, including doors, sliders, lids, and containers, are rendered with the camera axis perpendicular to their joint axes while ensuring maximum visible areas. Then, we draw manipulation guidance on these figures to display how to manipulate them. Following common representation in the real manuals from Section 3.1.1, we adopted these four different strategies to draw manipulation guidance as shown in Figure 2. To realize this, we manually annotate the 3D position of one contact point for every manipulable part on all appliance CAD models and calculate its motion trajectory according to the part-linked joint axis." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": "Human Verification about Manual Figures When collecting appliance overviews based on the largest visible part areas, the view may not be suitable to observe every part." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 205, + 555, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 205, + 555, + 252 + ], + "spans": [ + { + "bbox": [ + 313, + 205, + 555, + 252 + ], + "type": "text", + "content": "Besides, sketch-style figures may contain missing lines or abnormal shadows. Therefore, we conduct human verification on more than 1800 figures to trigger regeneration when these errors occur." + } + ] + } + ], + "index": 13 + }, + { + "type": "table", + "bbox": [ + 316, + 293, + 555, + 359 + ], + "blocks": [ + { + "bbox": [ + 313, + 262, + 555, + 285 + ], + "lines": [ + { + "bbox": [ + 313, + 262, + 555, + 285 + ], + "spans": [ + { + "bbox": [ + 313, + 262, + 555, + 285 + ], + "type": "text", + "content": "Table 1. Comparison of different manual-based manipulation challenge tracks in our CheckManual benchmark." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 316, + 293, + 555, + 359 + ], + "lines": [ + { + "bbox": [ + 316, + 293, + 555, + 359 + ], + "spans": [ + { + "bbox": [ + 316, + 293, + 555, + 359 + ], + "type": "table", + "html": "
AVAILABLE INPUTPREDICTION TARGET
Task InstructionRGB-DManual PDFAppliance CAD ModelHigh-level PlanningLow-level Manipulation
Track 1: Manual-CAD-Appliance Aligned Planing
✓✓✓✓✓
Track 2: Manual & CAD based Manipulation
✓✓✓✓✓
Track 3: Pure Manual-based Manipulation
✓✓✓✓
", + "image_path": "69ff07721b735d6418dc23756290ba87fcb640c7e5e9194274e5bfe80c1deacd.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_body" + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 371, + 486, + 384 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 371, + 486, + 384 + ], + "spans": [ + { + "bbox": [ + 313, + 371, + 486, + 384 + ], + "type": "text", + "content": "3.1.5 LATEX-based Manual Generation" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 312, + 389, + 555, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 389, + 555, + 510 + ], + "spans": [ + { + "bbox": [ + 312, + 389, + 555, + 510 + ], + "type": "text", + "content": "As we know, " + }, + { + "bbox": [ + 312, + 389, + 555, + 510 + ], + "type": "inline_equation", + "content": "\\mathrm{LATEX}" + }, + { + "bbox": [ + 312, + 389, + 555, + 510 + ], + "type": "text", + "content": " is a document programming language that offers both the advantages of standardization and flexibility. We can not only keep the same document style from page to page but also easily insert different types of figures and tables into the document by " + }, + { + "bbox": [ + 312, + 389, + 555, + 510 + ], + "type": "inline_equation", + "content": "\\mathrm{LATEX}" + }, + { + "bbox": [ + 312, + 389, + 555, + 510 + ], + "type": "text", + "content": " coding. Recently, large language models emerged with powerful capabilities for code generation. Therefore, we decided to leverage LLM to generate the " + }, + { + "bbox": [ + 312, + 389, + 555, + 510 + ], + "type": "inline_equation", + "content": "\\mathrm{LATEX}" + }, + { + "bbox": [ + 312, + 389, + 555, + 510 + ], + "type": "text", + "content": " codes for the manual and then compile them into a PDF format appliance manual as shown in the rightmost part of Figure 2." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 510, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 556, + 714 + ], + "type": "text", + "content": "Before starting this process, we pre-defined a set of widely used " + }, + { + "bbox": [ + 313, + 510, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{LATEX}" + }, + { + "bbox": [ + 313, + 510, + 556, + 714 + ], + "type": "text", + "content": " packages (e.g., multirow, xcolor, caption) and prepared a collection of commonly used icons (e.g., fragility reminder, temperature warning, environment protection). To increase diversity, we randomly sample different formats of covers, diagrams, and guidance demonstrations from the Section 3.1.4 as figure resources and randomly select a text style from real manuals (e.g., text, bullet list, table) for task description. In the generation, we synthesize the prompt with textual part function annotation, manipulation tasks, and above image paths to require the large model to continue one section of " + }, + { + "bbox": [ + 313, + 510, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{LATEX}" + }, + { + "bbox": [ + 313, + 510, + 556, + 714 + ], + "type": "text", + "content": " codes based on the previous. To prevent uncompilable codes, we design a section-granularity compilation mechanism. The newly generated section of " + }, + { + "bbox": [ + 313, + 510, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{LATEX}" + }, + { + "bbox": [ + 313, + 510, + 556, + 714 + ], + "type": "text", + "content": " codes is appended to the end of the previous correct code and compiled together. Regeneration will be triggered when errors occur in the code compi" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22599" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 95 + ], + "type": "text", + "content": "lation. Figure 3 displays one manual from our CheckManual dataset." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 105, + 197, + 118 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 105, + 197, + 118 + ], + "spans": [ + { + "bbox": [ + 55, + 105, + 197, + 118 + ], + "type": "text", + "content": "3.2. Dataset Statistic Analysis" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 123, + 295, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 123, + 295, + 291 + ], + "spans": [ + { + "bbox": [ + 55, + 123, + 295, + 291 + ], + "type": "text", + "content": "We totally created 1107 manuals for 369 appliances, which cover 11 common appliance categories in daily life, including washing machines, coffee makers, dishwashers, microwaves, and so on. These appliances contain 12 types, with 2211 different manipulable parts. On average, there are 6.00 manipulable parts on each appliance. In terms of quantity and diversity, parts contained in our dataset can support most appliance operations. Besides, our dataset includes 1464 manipulation tasks in total. Every task is accompanied by step-by-step manipulation parts and target states. These tasks include both single-step and long-horizon tasks with a maximum of 18 steps. The table and charts in Figure 4 display comprehensive statistics about our CheckManual dataset." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 299, + 230, + 312 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 299, + 230, + 312 + ], + "spans": [ + { + "bbox": [ + 55, + 299, + 230, + 312 + ], + "type": "text", + "content": "3.3. CheckManual Challenge Tracks" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 318, + 295, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 318, + 295, + 402 + ], + "spans": [ + { + "bbox": [ + 55, + 318, + 295, + 402 + ], + "type": "text", + "content": "With the collected manual data, we design challenge tracks for manual-based manipulation. The basic setting is given a task instruction " + }, + { + "bbox": [ + 55, + 318, + 295, + 402 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 318, + 295, + 402 + ], + "type": "text", + "content": " and a corresponding appliance PDF manual " + }, + { + "bbox": [ + 55, + 318, + 295, + 402 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 318, + 295, + 402 + ], + "type": "text", + "content": ", the robot should follow the manual to complete the tabletop appliance manipulation task with RGB-D observation " + }, + { + "bbox": [ + 55, + 318, + 295, + 402 + ], + "type": "inline_equation", + "content": "I_{c}" + }, + { + "bbox": [ + 55, + 318, + 295, + 402 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 318, + 295, + 402 + ], + "type": "inline_equation", + "content": "I_{d}" + }, + { + "bbox": [ + 55, + 318, + 295, + 402 + ], + "type": "text", + "content": ". Based on this basic setting, we design the challenge tracks as shown in Table 1 and the following." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 403, + 295, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 403, + 295, + 558 + ], + "spans": [ + { + "bbox": [ + 55, + 403, + 295, + 558 + ], + "type": "text", + "content": "Track 1: CAD-Appliance Aligned Planing We first consider an ideal setting where the appliance CAD models are available and CAD-Appliance is aligned. The \"CAD-Appliance Aligned\" means part area and ID are accessible on the camera observation. In this case, the evaluation focuses on testing the robot's capability of analyzing appliance manuals and making manipulation plans based on the given task instruction and manual content. The evaluation is composed of two aspects. Firstly the robot should align all function names in the manuals with appliance parts. Secondly, the task planning should include correct manipulation target part and primitive actions in every task step. This challenge track is free from manipulation execution." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 558, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 295, + 714 + ], + "type": "text", + "content": "Track 2: Manual & CAD based Manipulation In this task, the appliance CAD models are available but the mapping relationships among parts on manual, CAD model, and observation are not known. This setting is reasonable because it is easy for manufacturers to provide plain CAD models for their appliance products. For evaluation, the robot should follow the given task instruction to execute low-level manipulation with a two-finger gripper based on manual content and CAD model. To complete this challenge, the robot needs to first estimate the 6-DoF pose of observed objects before using the CAD model. After that, the robot should also align parts on the manual, CAD model, and current observation. After completing the part align" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "text", + "content": "ment, the robot can obtain part-linked joint position and rotation axis from the CAD model, which will help the robot to make manipulation planning and execution." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 109, + 555, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 109, + 555, + 299 + ], + "spans": [ + { + "bbox": [ + 313, + 109, + 555, + 299 + ], + "type": "text", + "content": "Track 3: Pure Manual-based Manipulation We further consider the setting that aligns with real-world applications where only appliance manuals are available to the robot. The challenges of this setting come from two aspects. On the high level, the robot should analyze the appliance manual to learn about the function information (i.e., name and possible states) and manipulation methods of every movable part on the appliances. Then, the robot should plan the detailed manipulation steps for the given task based on the manual content and common sense. On the low level, the robot must execute the high-level plan step-by-step with its end-effector to complete the instruction-specified task. During the whole process, the robot can only rely on its RGB-D observation and manuals instead of any CAD-model assistance. For both Track 2 and Track 3, any procedural error during the operation will cause the entire task to fail." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 300, + 555, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 300, + 555, + 598 + ], + "spans": [ + { + "bbox": [ + 313, + 300, + 555, + 598 + ], + "type": "text", + "content": "Evaluation Metrics To evaluate model performances on the above challenge tracks, we define the task-level and step-level metrics. On the step level, we evaluate the alignment success rate and planning success rate for Track 1 and task execution success rate and average task completion rate for Track 2 and 3. The alignment success is recognized when all CAD model parts are linked to the correct function names mentioned in the manual. The task planning success is confirmed when part IDs and manipulation actions in every planned step are all correct. The task execution is regarded as a success when all task steps are completed in the correct order. The task completion rate is calculated by \"Success Steps / # Task Steps\". Note that only steps before the failure step can be counted as successful steps. We take the average values of the above metrics as final results. On the step level, we check the state of the target part after each manipulation step to judge the success. Specifically, we define the whole movement limitation range of the target joint as " + }, + { + "bbox": [ + 313, + 300, + 555, + 598 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 313, + 300, + 555, + 598 + ], + "type": "text", + "content": ". For revolute parts like knobs, levers, and doors, the acceptable error range is defined as " + }, + { + "bbox": [ + 313, + 300, + 555, + 598 + ], + "type": "inline_equation", + "content": "L \\pm 30^{\\circ}" + }, + { + "bbox": [ + 313, + 300, + 555, + 598 + ], + "type": "text", + "content": ". For prismatic parts like the button, container, and slider, the successful manipulation is confirmed when the button moves " + }, + { + "bbox": [ + 313, + 300, + 555, + 598 + ], + "type": "inline_equation", + "content": ">25\\%" + }, + { + "bbox": [ + 313, + 300, + 555, + 598 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 313, + 300, + 555, + 598 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 313, + 300, + 555, + 598 + ], + "type": "inline_equation", + "content": "MD_{part} / MD_{appliance}" + }, + { + "bbox": [ + 313, + 300, + 555, + 598 + ], + "type": "inline_equation", + "content": ">50\\%" + }, + { + "bbox": [ + 313, + 300, + 555, + 598 + ], + "type": "text", + "content": " if target part and appliances move in the same direction, where " + }, + { + "bbox": [ + 313, + 300, + 555, + 598 + ], + "type": "inline_equation", + "content": "MD" + }, + { + "bbox": [ + 313, + 300, + 555, + 598 + ], + "type": "text", + "content": " represented motion distance." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 609, + 525, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 609, + 525, + 624 + ], + "spans": [ + { + "bbox": [ + 313, + 609, + 525, + 624 + ], + "type": "text", + "content": "4. Manual-based Manipulation Baselines" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 630, + 553, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 553, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 553, + 715 + ], + "type": "text", + "content": "In this section, we design baseline models for the three tracks of the benchmark. For track 1, we propose the ManualPlan model, which focuses on manual-based task planning, as detailed in Section 4.1. For tasks involving robotic execution, corresponding to track 2 and track 3, we extend the ManualPlan model by integrating it as a high-level planning module with existing manipulation models, as dis" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "22600" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 70, + 553, + 173 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 553, + 173 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 553, + 173 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 553, + 173 + ], + "type": "image", + "image_path": "c405025e32b47094d4e7ed3ee1040f42db8a1d1b3ed5bb66a8de5012992e7047.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 184, + 555, + 220 + ], + "lines": [ + { + "bbox": [ + 55, + 184, + 555, + 220 + ], + "spans": [ + { + "bbox": [ + 55, + 184, + 555, + 220 + ], + "type": "text", + "content": "Figure 5. The framework of our ManualPlan model. It is composed of Manual Resolution, Manipulation Planing and Part Alignment modules. The ManualPlan can make high level planning to control the CAD-assisted primitive actions or open-vocabulary manipulation large model (e.g., VoxPoser [10]) to use the appliance, which serves as the baseline models for CheckManual benchmark." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 231, + 145, + 242 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 231, + 145, + 242 + ], + "spans": [ + { + "bbox": [ + 55, + 231, + 145, + 242 + ], + "type": "text", + "content": "cussed in Section 4.2." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 251, + 277, + 264 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 251, + 277, + 264 + ], + "spans": [ + { + "bbox": [ + 55, + 251, + 277, + 264 + ], + "type": "text", + "content": "4.1. ManualPlan for High-Level Task Planning" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 270, + 296, + 329 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 270, + 296, + 329 + ], + "spans": [ + { + "bbox": [ + 55, + 270, + 296, + 329 + ], + "type": "text", + "content": "ManualPlan introduces a novel framework for high-level task planning of robotic manipulation with appliance manuals. It processes manual content, generates detailed plans, and aligns parts in the real appliance with those referenced in the manual." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 330, + 296, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 330, + 296, + 449 + ], + "spans": [ + { + "bbox": [ + 55, + 330, + 296, + 449 + ], + "type": "text", + "content": "Manual Resolution Module. The appliance manual in our dataset is in PDF format, which is consistent with the real manuals. The PDF file cannot be directly processed by large language models. Therefore, we divide the manual PDF file into page-level images first. Then, we extract text content from these page images using the Paddle OCR model. We also utilize the multimodal large language model (MLLM) to detect visual elements like figures and tables on each page image. In this way, we can obtain the page-granularity text and vision information about the whole manual." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 450, + 295, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 450, + 295, + 582 + ], + "spans": [ + { + "bbox": [ + 55, + 450, + 295, + 582 + ], + "type": "text", + "content": "Manipulation Planning Module. With extracted manual information, we call LLM to plan a step-by-step appliance manipulation process for the given task. The large language model must specify the part function name and manipulated states in the dictionary format like “{'Brew Strength Lever': 'Rotate 60 degrees'}” for every planning step. Only the manual-mentioned parts can be used in the plan. Considering just a part of the tasks are demonstrated in the manual, the LLM should make plans based on both common sense about appliance usage and the manual content. Any unresolvable prediction will trigger the regeneration request." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "content": "Part Alignment Module. To achieve this goal, we incorporate a visual prompting mechanism known as Set-of-Mark (SoM) [28] to align parts in the field of view and parts depicted in the manual. SoM leverages segmentation models to partition an image into distinct regions and assign a numeric marker to each, which can significantly improve the visual grounding capabilities of multimodal large language models. Here, we utilize the open-vocabulary object detection model Grounding-DINO [18] to predict bounding boxes of movable parts and prompt Segment Anything [13] with these bounding boxes" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 231, + 555, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 231, + 555, + 434 + ], + "spans": [ + { + "bbox": [ + 313, + 231, + 555, + 434 + ], + "type": "text", + "content": "to generate movable part masks. These masks are assigned different colors with transparency and overlayed on the observation image. The ID is marked at the geometric center of each part mask. Then, we provide the observation image and appliance diagrams from manual to MLLM requiring it to identify the mapping relationships from mask IDs to part function names. This way, we realize the part alignment between the observation and the manual. For the Manual & CAD Model-based Manipulation task, although the CAD model is available, the part function names still need to be aligned with parts on the CAD model. Therefore, we use a similar SoM-based approach to achieve such alignment with MLLM. For the other two tasks, the CAD model is unavailable for the Pure Manual-based Manipulation task while part alignment is already achieved in the Manual-CAD-Appliance Aligned Manipulation task." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 455, + 549, + 468 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 455, + 549, + 468 + ], + "spans": [ + { + "bbox": [ + 313, + 455, + 549, + 468 + ], + "type": "text", + "content": "4.2. ManualPlan-Driven Low-Level Manipulation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 474, + 554, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 474, + 554, + 521 + ], + "spans": [ + { + "bbox": [ + 313, + 474, + 554, + 521 + ], + "type": "text", + "content": "When executing manipulation on the appliance, we employ ManualPlan as a high-level planner and design different manipulation policies for Track 2 (CAD Model Available) and Track 3 (CAD Model Unavailable)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 521, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 555, + 714 + ], + "type": "text", + "content": "CAD-assisted Manipulation Policy. When the CAD model is available, we can estimate the 6-DoF pose of the observed appliance in the camera frame by the pose estimation models like [25][31][30]. With such an estimated object pose, the observed appliance can be aligned with its CAD model and all joints' positions and axis directions on the observed appliance become available. Therefore, we design primitive actions for different appliance parts to synthesize CAD-assisted manipulation policy. For door and parts with slider (e.g., screen, slider, lid), we provide the 3D bounding box of the target part to constrain the AnyGrasp [7] model to predict part-level grasp poses. Then, the gripper poses in the motion sequences can be calculated based on the information about the joint and manipulation method. For the button part, we take its surface geometrical center as the contact point and control the closed gripper" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22601" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 112, + 553, + 218 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 104 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 104 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 104 + ], + "type": "text", + "content": "Table 2. Baseline model's performances on three challenge tracks of CheckManual benchmark. The results of Track 1 are reported in the \"Part Alignment Success Rate / Task Planning Success Rate\" format. For Track 2 and 3, their results are reported in \"Task Completion Rate / Task Success Rate\" format. The \"w/o manual\" means ablating appliance manuals input for the above methods." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 112, + 553, + 218 + ], + "lines": [ + { + "bbox": [ + 58, + 112, + 553, + 218 + ], + "spans": [ + { + "bbox": [ + 58, + 112, + 553, + 218 + ], + "type": "table", + "html": "
BASELINE MODEL图图图图图图图图图图图Total
Track 1: Manual-CAD-Appliance Aligned Planing
ManualPlan w/o manual22.96 / 22.22 - / 4.2139.47 / 22.82 - / 2.6818.52 / 14.29 - / 20.0072.84 / 60.19 - / 48.154.17 / 3.91 - / 0.788.13 / 8.64 - / 2.4794.44 / 68.00 - / 36.0047.22 / 30.56 - / 6.9441.67 / 40.62 - / 15.6221.15 / 14.39 - / 4.3215.46 / 9.96 - / 2.8527.32 / 20.70 - / 7.99
Track 2: Manual & CAD based Manipulation
ManualPlan + Primitive Actions w/o manual2.39 / 1.15 1.05 / 0.387.12 / 4.70 1.76 / 1.3413.63 / 11.43 21.39 / 14.290.46 / 0.00 0.46 / 0.000.78 / 0.78 0.78 / 0.787.91 / 3.70 3.50 / 1.2310.00 / 4.00 2.00 / 0.0011.11 / 4.86 5.09 / 1.393.12 / 0.00 1.04 / 0.002.32 / 1.44 2.28 / 2.162.56 / 1.42 1.69 / 1.424.50 / 2.39 2.47 / 1.37
Track 3: Pure Manual-based Manipulation
ManualPlan + VoxPoser [10] w/o manual6.11 / 1.92 0.99 / 0.387.18 / 4.70 1.39 / 0.677.10 / 5.71 0.57 / 0.003.01 / 1.85 0.46 / 0.003.12 / 3.12 1.56 / 1.567.99 / 3.70 1.54 / 0.620.00 / 0.00 0.00 / 0.0010.66 / 6.25 2.50 / 0.696.75 / 3.12 0.00 / 0.000.48 / 0.00 0.00 / 0.002.96 / 2.14 0.98 / 0.715.18 / 2.87 1.11 / 0.55
", + "image_path": "d16d2c8378ad67acbc542077cd9ac705f56619a634aec404e228ab613f06867b.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 231, + 297, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 231, + 297, + 447 + ], + "spans": [ + { + "bbox": [ + 54, + 231, + 297, + 447 + ], + "type": "text", + "content": "to push the button from the normal direction of the contact point. For the knob part, we control the open gripper to get close to the knob from the normal direction of its surface geometrical center and close the gripper to grasp. After the gripper successfully grasps the knob, it will rotate planned degrees around the normal of its surface geometrical center. CAD-free Manipulation Policy. We integrate a recently proposed open-vocabulary manipulation model VoxPoser [10] into our ManualPlan to complete the CAD-free manipulation task. In this framework, ManualPlan and VoxPoser cooperate like the cerebrum and cerebellum in the human brain. ManualPlan reads manuals, makes high-level planning, and aligns observed components with manual-depicted parts. Then, it grounds the target component and demonstrates the corresponding manipulation method in language format, which is transferred to VoxPoser with camera RGB-D observation to execute low-level manipulation action on the appliance." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 457, + 137, + 470 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 457, + 137, + 470 + ], + "spans": [ + { + "bbox": [ + 55, + 457, + 137, + 470 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 476, + 188, + 489 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 476, + 188, + 489 + ], + "spans": [ + { + "bbox": [ + 55, + 476, + 188, + 489 + ], + "type": "text", + "content": "5.1. Implementation Details" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 495, + 296, + 627 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 495, + 296, + 627 + ], + "spans": [ + { + "bbox": [ + 55, + 495, + 296, + 627 + ], + "type": "text", + "content": "When creating the CheckManual dataset, we utilize GPT-4o [1] as the large language model and multimodal large language model. Following previous work [26], we use the Franka Panda flying gripper in the simulation. The groundingDINO [18] model is utilized to ground the bounding boxes of appliance components. The predicted bounding boxes prompt SAM [13] model to generate component masks. FoundationPose [25] helps us estimate appliance 6-DoF pose in the camera observation. In the evaluation, manual comprehension and manipulation planning processes are assisted by GPT-4o [1] model." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 635, + 183, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 635, + 183, + 647 + ], + "spans": [ + { + "bbox": [ + 55, + 635, + 183, + 647 + ], + "type": "text", + "content": "5.2. Simulation Evaluation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "type": "text", + "content": "We evaluate our proposed ManualPlan model and ManualPlan-driven framework using the CheckManual benchmark, which employs SAPIEN [27] as the simulation. This environment provides a realistic and flexible simulation for testing robotic manipulation tasks across a" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 231, + 555, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 231, + 555, + 483 + ], + "spans": [ + { + "bbox": [ + 313, + 231, + 555, + 483 + ], + "type": "text", + "content": "variety of appliances. The experimental results on different tasks and appliance categories are summarized in Table 2. From the results of Track 1, we observe that the ManualPlan model is capable of generating a part of correct task plans based on multi-page manuals. However, even the most advanced Multi-Modal Large Language Models (MLLMs) still face significant challenges in aligning manual part annotations with the corresponding CAD models of appliances, which greatly limits the overall performance of ManualPlan in task planning. For Track 2 and Track 3, the performance of the ManualPlan-driven framework highlights the difficulties posed by long-horizon, manual-based manipulation tasks. After analyzing failure cases, we identified that the accumulation of errors in part identification, task planning, and multi-step execution are the primary causes of failure. Additionally, we conducted an ablation study by removing the manual (i.e., w/o manual) to investigate its impact on task performance. As shown in Table 2, the manual plays a crucial role in guiding the robot to accurately plan and execute manipulation tasks, underscoring its importance in completing appliance manipulation tasks." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 492, + 449, + 504 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 492, + 449, + 504 + ], + "spans": [ + { + "bbox": [ + 313, + 492, + 449, + 504 + ], + "type": "text", + "content": "5.3. Real Robot Deployment" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 511, + 555, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 511, + 555, + 608 + ], + "spans": [ + { + "bbox": [ + 313, + 511, + 555, + 608 + ], + "type": "text", + "content": "To demonstrate the potential of our approach in real-world applications, we conducted real-world robot experiments following the task definition of Track 3 from the Check-Manual benchmark. The hardware setup involved a Franka Panda robot arm and a RealSense D415 RGB-D camera mounted on the end-effector. Considering error accumulation in the long-horizon execution, we follow [11] to cache correct manipulation actions in every step." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 620, + 388, + 633 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 620, + 388, + 633 + ], + "spans": [ + { + "bbox": [ + 313, + 620, + 388, + 633 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 641, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 555, + 713 + ], + "type": "text", + "content": "In this work, we introduce CheckManual, the first benchmark for manual-based appliance manipulation, and define new challenges, metrics, and simulator environments for model evaluation. Additionally, we present the first manual-based manipulation planning model ManualPlan to set up a group of baselines for the CheckManual benchmark." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22602" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "type": "text", + "content": "Acknowledgment" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 90, + 296, + 146 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 90, + 296, + 146 + ], + "spans": [ + { + "bbox": [ + 55, + 90, + 296, + 146 + ], + "type": "text", + "content": "This work was supported by the National Youth Talent Support Program (8200800081) and National Natural Science Foundation of China (No. 62376006). We also thank Yang Liu and Zhuoqun Xu from Polar Mirage for their technical support on assets." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 155, + 115, + 167 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 155, + 115, + 167 + ], + "spans": [ + { + "bbox": [ + 56, + 155, + 115, + 167 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 175, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 61, + 175, + 296, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 175, + 296, + 228 + ], + "spans": [ + { + "bbox": [ + 61, + 175, + 296, + 228 + ], + "type": "text", + "content": "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 230, + 296, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 230, + 296, + 285 + ], + "spans": [ + { + "bbox": [ + 61, + 230, + 296, + 285 + ], + "type": "text", + "content": "[2] Michael Ahn, Anthony Brohan, Noah Brown, Yevgen Chebotar, Omar Cortes, Byron David, Chelsea Finn, Chuyuan Fu, Keerthana Gopalakrishnan, Karol Hausman, et al. Do as i can, not as i say: Grounding language in robotic affordances. arXiv preprint arXiv:2204.01691, 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 285, + 295, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 285, + 295, + 460 + ], + "spans": [ + { + "bbox": [ + 61, + 285, + 295, + 460 + ], + "type": "text", + "content": "[3] Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Xi Chen, Krzysztof Choromanski, Tianli Ding, Danny Driess, Avinava Dubey, Chelsea Finn, Pete Florence, Chuyuan Fu, Montse Gonzalez Arenas, Keerthana Gopalakrishnan, Kehang Han, Karol Hausman, Alexander Herzog, Jasmine Hsu, Brian Ichter, Alex Irpan, Nikhil Joshi, Ryan Julian, Dmitry Kalashnikov, Yuheng Kuang, Isabel Leal, Lisa Lee, Tsang-Wei Edward Lee, Sergey Levine, Yao Lu, Henryk Michalewski, Igor Mordatch, Karl Pertsch, Kanishka Rao, Krista Reymann, Michael Ryoo, Grecia Salazar, Pannag Sanketi, Pierre Sermanet, Jaspiar Singh, Anikait Singh, Radu Soricut, Huong Tran, Vincent Vanhoucke, Quan Vuong, Ayzaan Wahid, Stefan Welker, Paul Wohlhart, Jialin Wu, Fei Xia, Ted Xiao, Peng Xu, Sichun Xu, Tianhe Yu, and Brianna Zitkovich. Rt-2: Vision-language-action models transfer web knowledge to robotic control, 2023. 1, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 61, + 460, + 296, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 460, + 296, + 624 + ], + "spans": [ + { + "bbox": [ + 61, + 460, + 296, + 624 + ], + "type": "text", + "content": "[4] Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Joseph Dabis, Chelsea Finn, Keerthana Gopalakrishnan, Karol Hausman, Alex Herzog, Jasmine Hsu, Julian Ibarz, Brian Ichter, Alex Irpan, Tomas Jackson, Sally Jesmonth, Nikhil J Joshi, Ryan Julian, Dmitry Kalashnikov, Yuheng Kuang, Isabel Leal, Kuang-Huei Lee, Sergey Levine, Yao Lu, Utsav Malla, Deeksha Manjunath, Igor Mordatch, Ofir Nachum, Carolina Parada, Jodilyn Peralta, Emily Perez, Karl Pertsch, Jornell Quiambao, Kanishka Rao, Michael Ryoo, Grecia Salazar, Pannag Sanketi, Kevin Sayed, Jaspiar Singh, Sumedh Sontakke, Austin Stone, Clayton Tan, Huong Tran, Vincent Vanhoucke, Steve Vega, Quan Vuong, Fei Xia, Ted Xiao, Peng Xu, Sichun Xu, Tianhe Yu, and Brianna Zitkovich. Rt-1: Robotics transformer for real-world control at scale, 2023. 1, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 61, + 625, + 296, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 625, + 296, + 690 + ], + "spans": [ + { + "bbox": [ + 61, + 625, + 296, + 690 + ], + "type": "text", + "content": "[5] Wenzhe Cai, Siyuan Huang, Guangran Cheng, Yuxing Long, Peng Gao, Changyin Sun, and Hao Dong. Bridging zero-shot object navigation and foundation models through pixel-guided navigation skill. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 5228-5234. IEEE, 2024. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 61, + 691, + 296, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 691, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 61, + 691, + 296, + 714 + ], + "type": "text", + "content": "[6] Martin Ester, Hans-Peter Kriegel, Jörg Sander, Xiaowei Xu, et al. A density-based algorithm for discovering clusters in" + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 94 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 94 + ], + "type": "text", + "content": "large spatial databases with noise. In kdd, pages 226-231, 1996. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 320, + 96, + 553, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 96, + 553, + 151 + ], + "spans": [ + { + "bbox": [ + 320, + 96, + 553, + 151 + ], + "type": "text", + "content": "[7] Hao-Shu Fang, Chenxi Wang, Hongjie Fang, Minghao Gou, Jirong Liu, Hengxu Yan, Wenhai Liu, Yichen Xie, and Cewu Lu. Anygrasp: Robust and efficient grasp perception in spatial and temporal domains. IEEE Transactions on Robotics (T-RO), 2023. 1, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 320, + 152, + 553, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 152, + 553, + 219 + ], + "spans": [ + { + "bbox": [ + 320, + 152, + 553, + 219 + ], + "type": "text", + "content": "[8] Jiayuan Gu, Sean Kirmani, Paul Wohlhart, Yao Lu, Montserrat Gonzalez Arenas, Kanishka Rao, Wenhao Yu, Chuyuan Fu, Keerthana Gopalakrishnan, Zhuo Xu, Priya Sundaresan, Peng Xu, Hao Su, Karol Hausman, Chelsea Finn, Quan Vuong, and Ted Xiao. Rt-trajectory: Robotic task generalization via hindsight trajectory sketches, 2023. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 320, + 220, + 553, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 220, + 553, + 263 + ], + "spans": [ + { + "bbox": [ + 320, + 220, + 553, + 263 + ], + "type": "text", + "content": "[9] Haoxu Huang, Fanqi Lin, Yingdong Hu, Shengjie Wang, and Yang Gao. Copa: General robotic manipulation through spatial constraints of parts with foundation models. arXiv preprint arXiv:2403.08248, 2024. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 315, + 264, + 553, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 264, + 553, + 308 + ], + "spans": [ + { + "bbox": [ + 315, + 264, + 553, + 308 + ], + "type": "text", + "content": "[10] Wenlong Huang, Chen Wang, Ruohan Zhang, Yunzhu Li, Jiajun Wu, and Li Fei-Fei. Voxposer: Composable 3d value maps for robotic manipulation with language models. arXiv preprint arXiv:2307.05973, 2023. 1, 2, 7, 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 315, + 310, + 553, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 310, + 553, + 354 + ], + "spans": [ + { + "bbox": [ + 315, + 310, + 553, + 354 + ], + "type": "text", + "content": "[11] Wenlong Huang, Chen Wang, Yunzhu Li, Ruohan Zhang, and Li Fei-Fei. Rekep: Spatio-temporal reasoning of relational keypoint constraints for robotic manipulation. arXiv preprint arXiv:2409.01652, 2024. 1, 2, 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 315, + 355, + 553, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 355, + 553, + 432 + ], + "spans": [ + { + "bbox": [ + 315, + 355, + 553, + 432 + ], + "type": "text", + "content": "[12] Moo Jin Kim, Karl Pertsch, Siddharth Karamcheti, Ted Xiao, Ashwin Balakrishna, Suraj Nair, Rafael Rafailov, Ethan Foster, Grace Lam, Pannag Sanketi, Quan Vuong, Thomas Kollar, Benjamin Burchfiel, Russ Tedrake, Dorsa Sadigh, Sergey Levine, Percy Liang, and Chelsea Finn. Openvla: An open-source vision-language-action model. arXiv preprint arXiv:2406.09246, 2024. 1, 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 315, + 434, + 553, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 434, + 553, + 487 + ], + "spans": [ + { + "bbox": [ + 315, + 434, + 553, + 487 + ], + "type": "text", + "content": "[13] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dollar, and Ross Girshick. Segment anything. arXiv:2304.02643, 2023. 7, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 315, + 489, + 553, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 489, + 553, + 555 + ], + "spans": [ + { + "bbox": [ + 315, + 489, + 553, + 555 + ], + "type": "text", + "content": "[14] Isabel Leal, Krzysztof Choromanski, Deepali Jain, Avinava Dubey, Jake Varley, Michael Ryoo, Yao Lu, Frederick Liu, Vikas Sindhwani, Quan Vuong, et al. Sara-rt: Scaling up robotics transformers with self-adaptive robust attention. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 6920–6927. IEEE, 2024. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 557, + 553, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 557, + 553, + 610 + ], + "spans": [ + { + "bbox": [ + 316, + 557, + 553, + 610 + ], + "type": "text", + "content": "[15] Xinghang Li, Minghuan Liu, Hanbo Zhang, Cunjun Yu, Jie Xu, Hongtao Wu, Chilam Cheang, Ya Jing, Weinan Zhang, Huaping Liu, et al. Vision-language foundation models as effective robot imitators. arXiv preprint arXiv:2311.01378, 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 613, + 553, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 613, + 553, + 679 + ], + "spans": [ + { + "bbox": [ + 316, + 613, + 553, + 679 + ], + "type": "text", + "content": "[16] Xiaqi Li, Mingxu Zhang, Yiran Geng, Haoran Geng, Yuxing Long, Yan Shen, Renrui Zhang, Jiaming Liu, and Hao Dong. Maniplm: Embodied multimodal large language model for object-centric robotic manipulation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18061-18070, 2024. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 680, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 680, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 680, + 553, + 712 + ], + "type": "text", + "content": "[17] Shanchuan Lin, Anran Wang, and Xiao Yang. Sdxlightning: Progressive adversarial diffusion distillation, 2024. 5" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22603" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 127 + ], + "type": "text", + "content": "[18] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Qing Jiang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, and Lei Zhang. Grounding dino: Marrying dino with grounded pre-training for open-set object detection, 2024. 7, 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 294, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 294, + 173 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 294, + 173 + ], + "type": "text", + "content": "[19] Yuxing Long, Wenzhe Cai, Hongcheng Wang, Guanqi Zhan, and Hao Dong. Instructnav: Zero-shot system for generic instruction navigation in unexplored environment. arXiv preprint arXiv:2406.04882, 2024. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 175, + 294, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 294, + 228 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 294, + 228 + ], + "type": "text", + "content": "[20] Yuxing Long, Xiaqi Li, Wenzhe Cai, and Hao Dong. Discuss before moving: Visual language navigation via multi-expert discussions. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 17380-17387. IEEE, 2024. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 230, + 294, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 230, + 294, + 263 + ], + "spans": [ + { + "bbox": [ + 56, + 230, + 294, + 263 + ], + "type": "text", + "content": "[21] Jianlan Luo, Charles Xu, Jeffrey Wu, and Sergey Levine. Hil-serl: Precise and dexterous robotic manipulation via human-in-the-loop reinforcement learning, 2024. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 265, + 294, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 265, + 294, + 319 + ], + "spans": [ + { + "bbox": [ + 56, + 265, + 294, + 319 + ], + "type": "text", + "content": "[22] Kaichun Mo, Shilin Zhu, Angel X. Chang, Li Yi, Subarna Tripathi, Leonidas J. Guibas, and Hao Su. PartNet: A large-scale benchmark for fine-grained and hierarchical part-level 3D object understanding. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 321, + 294, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 321, + 294, + 398 + ], + "spans": [ + { + "bbox": [ + 56, + 321, + 294, + 398 + ], + "type": "text", + "content": "[23] Abhilash Nandy, Soumya Sharma, Shubham Maddhashiya, Kapil Sachdeva, Pawan Goyal, and NIloy Ganguly. Question answering over electronic devices: A new benchmark dataset and a multi-task learning based QA framework. In Findings of the Association for Computational Linguistics: EMNLP 2021, pages 4600-4609, Punta Cana, Dominican Republic, 2021. Association for Computational Linguistics. 1, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 399, + 294, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 399, + 294, + 476 + ], + "spans": [ + { + "bbox": [ + 56, + 399, + 294, + 476 + ], + "type": "text", + "content": "[24] Abhilash Nandy, Soumya Sharma, Shubham Maddhashiya, Kapil Sachdeva, Pawan Goyal, and NIloy Ganguly. Question answering over electronic devices: A new benchmark dataset and a multi-task learning based QA framework. In Findings of the Association for Computational Linguistics: EMNLP 2021, pages 4600-4609, Punta Cana, Dominican Republic, 2021. Association for Computational Linguistics. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 478, + 294, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 478, + 294, + 510 + ], + "spans": [ + { + "bbox": [ + 56, + 478, + 294, + 510 + ], + "type": "text", + "content": "[25] Bowen Wen, Wei Yang, Jan Kautz, and Stan Birchfield. FoundationPose: Unified 6d pose estimation and tracking of novel objects. In CVPR, 2024. 7, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 512, + 294, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 512, + 294, + 577 + ], + "spans": [ + { + "bbox": [ + 56, + 512, + 294, + 577 + ], + "type": "text", + "content": "[26] Ruihai Wu, Yan Zhao, Kaichun Mo, Zizheng Guo, Yian Wang, Tianhao Wu, Qingnan Fan, Xuelin Chen, Leonidas Guibas, and Hao Dong. VAT-mart: Learning visual action trajectory proposals for manipulating 3d ARTiculated objects. In International Conference on Learning Representations, 2022. 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 578, + 294, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 578, + 294, + 644 + ], + "spans": [ + { + "bbox": [ + 56, + 578, + 294, + 644 + ], + "type": "text", + "content": "[27] Fanbo Xiang, Yuzhe Qin, Kaichun Mo, Yikuan Xia, Hao Zhu, Fangchen Liu, Minghua Liu, Hanxiao Jiang, Yifu Yuan, He Wang, Li Yi, Angel X. Chang, Leonidas J. Guibas, and Hao Su. SAPIEN: A simulated part-based interactive environment. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 647, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 294, + 689 + ], + "type": "text", + "content": "[28] Jianwei Yang, Hao Zhang, Feng Li, Xueyan Zou, Chunyuan Li, and Jianfeng Gao. Set-of-mark prompting unleashes extraordinary visual grounding in gpt-4v. arXiv preprint arXiv:2310.11441, 2023. 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "text", + "content": "[29] Senqiao Yang, Jiaming Liu, Ray Zhang, Mingjie Pan, Zoey Guo, Xiaqi Li, Zehui Chen, Peng Gao, Yandong Guo, and" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 251 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "text", + "content": "Shanghang Zhang. Lidar-llm: Exploring the potential of large language models for 3d lidar understanding. arXiv preprint arXiv:2312.14074, 2023. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 107, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 150 + ], + "type": "text", + "content": "[30] Jiyao Zhang, Mingdong Wu, and Hao Dong. Generative category-level object pose estimation via diffusion models. Advances in Neural Information Processing Systems, 36: 54627-54644, 2023. 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 152, + 553, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 553, + 206 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 553, + 206 + ], + "type": "text", + "content": "[31] Jiyao Zhang, Weiyao Huang, Bo Peng, Mingdong Wu, Fei Hu, Zijian Chen, Bo Zhao, and Hao Dong. Omni6dpose: A benchmark and model for universal 6d object pose estimation and tracking. In European Conference on Computer Vision, pages 199-216. Springer, 2024. 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 208, + 553, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 208, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 316, + 208, + 553, + 251 + ], + "type": "text", + "content": "[32] Liang Zhang, Anwen Hu, Jing Zhang, Shuo Hu, and Qin Jin. Mpmqa: multimodal question answering on product manuals. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 13958-13966, 2023. 1, 2" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "22604" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/CholecTrack20_ A Multi-Perspective Tracking Dataset for Surgical Tools/3671c43b-afb0-4272-8573-cad641013bba_content_list.json b/2025/CholecTrack20_ A Multi-Perspective Tracking Dataset for Surgical Tools/3671c43b-afb0-4272-8573-cad641013bba_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..8a0c9681a04b750201b72cdff5398eb7e7a2c233 --- /dev/null +++ b/2025/CholecTrack20_ A Multi-Perspective Tracking Dataset for Surgical Tools/3671c43b-afb0-4272-8573-cad641013bba_content_list.json @@ -0,0 +1,1997 @@ +[ + { + "type": "text", + "text": "CholecTrack20: A Multi-Perspective Tracking Dataset for Surgical Tools", + "text_level": 1, + "bbox": [ + 132, + 130, + 867, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Chinedu Innocent Nwoye $^{1,3}$ Kareem Elgohary $^{1}$ Anvita Srinivas $^{1}$ Fauzan Zaid $^{1}$ Joel L. Lavanchy $^{2}$ Nicolas Padoy $^{1,3}$", + "bbox": [ + 223, + 179, + 772, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1University of Strasbourg, CNRS, INSERM, ICube, UMR7357, Strasbourg, France", + "bbox": [ + 220, + 220, + 776, + 238 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2University of Basel, University Digestive Health Care Center, Clarunis, Switzerland", + "bbox": [ + 220, + 239, + 777, + 255 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{3}$ IHU Strasbourg, France", + "bbox": [ + 416, + 257, + 581, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "https://github.com/camma-public/cholectrack20", + "bbox": [ + 294, + 277, + 696, + 290 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 325, + 326, + 342 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Tool tracking in surgical videos is essential for advancing computer-assisted interventions, such as skill assessment, safety zone estimation, and human-machine collaboration. However, the lack of context-rich datasets limits AI applications in this field. Existing datasets rely on overly generic tracking formalizations that fail to capture surgical-specific dynamics, such as tools moving out of the camera's view or exiting the body. This results in less clinically relevant trajectories and a lack of flexibility for real-world surgical applications. Methods trained on these datasets often struggle with visual challenges such as smoke, reflection, and bleeding, further exposing the limitations of current approaches. We introduce CholecTrack20, a specialized dataset for multi-class, multi-tool tracking in surgical procedures. It redefines tracking formalization with three perspectives: (1) intraoperative, (2) intracorporeal, and (3) visibility, enabling adaptable and clinically meaningful tool trajectories. The dataset comprises 20 full-length surgical videos, annotated at 1 fps, yielding over 35K frames and 65K labeled tool instances. Annotations include spatial location, category, identity, operator, phase, and visual challenges. Benchmarking state-of-the-art methods on CholecTrack20 reveals significant performance gaps, with current approaches (< 45% HOTA) failing to meet the accuracy required for clinical translation. These findings motivate the need for advanced and intuitive tracking algorithms and establish CholecTrack20 as a foundation for developing robust AI-driven surgical assistance systems.", + "bbox": [ + 89, + 358, + 485, + 782 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 813, + 220, + 829 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The true impact of computer vision research lies in its practical applications, especially in critical fields like healthcare. Among these, surgery represents one of the most demanding domains, providing a definitive test for the capa", + "bbox": [ + 89, + 839, + 483, + 902 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "bilities of vision technologies. Surgical data science, on its own, has significantly advanced interventional healthcare by leveraging data-driven techniques to provide critical decision support to medical professionals [34]. A key area of this advancement is the analysis of endoscopic video data, which offers real-time insights into surgical procedures, aids in skill assessment, and helps predict complications [9, 56, 65]. Accurate tracking of surgical tools is central to these analyses, as it guides temporal progression of procedural phases and correlates with surgical actions and management of adverse events [38, 59]. Despite progress in computer vision, research [42] shows that deep learning models pretrained on general datasets often struggle with surgical contexts due to complex scene dynamics, diverse tool types, and challenging visual conditions such as bleeding, smoke, and variable lighting [12, 20, 51]. This highlights the need for domain-specific datasets tailored to the unique requirements of surgical tool tracking. Obtaining medical and surgical data for research is challenging due to ethical and practical constraints, and annotating it requires expert knowledge. Current methods for tool tracking largely focus on Single Object Tracking (SOT) [68], Multi-Class Tracking (MCT) with one tool per class [36, 37], or Multi-Object Tracking (MOT) treating all tools as a single class [16, 46]. However, these approaches often miss the complexities of Multi-Class Multi-Object Tracking (MC-MOT) specific to surgical contexts, where tools interact dynamically and may move out of the camera's field of view or within the body cavity.", + "bbox": [ + 511, + 327, + 906, + 767 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Multi-perspective tracking addresses these challenges by defining tool trajectories across different viewpoints during surgical procedures. It includes three critical perspectives (illustrated in Fig. 1): (1) intraoperative - covering the entire procedure duration to monitor tool usage and assess surgical proficiency; (2) intracorporeal - focusing on tool tracks within the body to evaluate specific tasks and predict risks; and (3) visibility - tracking tools within the camera's", + "bbox": [ + 511, + 779, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "8942", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/fb7f4a50edb064bbba40a2063ab477c772e2c10e47cee16d12dc6ec670f06693.jpg", + "image_caption": [ + "Surgeon operators positioning and tool placement" + ], + "image_footnote": [], + "bbox": [ + 98, + 102, + 336, + 279 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/d55b6a8fed8b89cd2f3ab3d8e1468aab64f78aa9e614eb87498504c462869ee2.jpg", + "image_caption": [ + "Multi-perspective of surgical tool trajectory" + ], + "image_footnote": [], + "bbox": [ + 343, + 102, + 669, + 255 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/2a8b05911b3f6726b8b0b485f2fa6562255f4383d24838cdab50923c7503b2af.jpg", + "image_caption": [ + "Sample data and label overlaid" + ], + "image_footnote": [], + "bbox": [ + 643, + 104, + 897, + 232 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/a1e3fce7f8d791780bb9dae68750a87f66ed65884246ea358f566051640babe3.jpg", + "image_caption": [ + "Considered surgical tool categories" + ], + "image_footnote": [], + "bbox": [ + 102, + 301, + 148, + 324 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/2221ca3216cc5592037105c619f4d105fe230a2f696aed73c433fdab8bc762dc.jpg", + "image_caption": [ + "bipolar" + ], + "image_footnote": [], + "bbox": [ + 153, + 301, + 200, + 324 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/e6df6ba0977b91ab3652d9ec1c2bd85b594a667e9997ae40d4178698f54b260d.jpg", + "image_caption": [ + "hook" + ], + "image_footnote": [], + "bbox": [ + 202, + 301, + 253, + 324 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/2b3f98256588cb824ae067c4ca0be875d5ec9afe922f11a28fe12797af82e020.jpg", + "image_caption": [ + "scissors" + ], + "image_footnote": [], + "bbox": [ + 254, + 301, + 300, + 324 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/97bb9c4cd069f1d9e0b835a0f4d61815dc77c0feb12eeae99df7faacf88c87ee.jpg", + "image_caption": [ + "grasper" + ], + "image_footnote": [], + "bbox": [ + 120, + 335, + 169, + 364 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/9a6618930d7de982a75cd5bcd79885d5012884ae6665818bc848556faf5522aa.jpg", + "image_caption": [ + "Figure 1. Illustration of multi-perspective tracking in surgical domain and CholecTrack20 dataset labels for surgical tool tracking." + ], + "image_footnote": [], + "bbox": [ + 179, + 335, + 223, + 364 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/66360be32b62168d2ffdb758a6d43526e9aa634684e13853376e1b7ed1ee49eb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 230, + 335, + 277, + 364 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/f4e498405819c4c83448ab7bfcaa82de25ba95094f588a7ee9b3cf452174901f.jpg", + "image_caption": [ + "Visibility tracking FoV", + "1", + "Operating room", + "$\\left( 4\\right)$", + "$\\therefore m - 1 \\neq 0$ ;", + "$\\therefore m - 1 \\neq 0$ ;", + "" + ], + "image_footnote": [], + "bbox": [ + 318, + 290, + 398, + 352 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "room", + "bbox": [ + 447, + 311, + 465, + 319 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "$\\frac{1 + u}{1} - \\frac{u}{1} = \\frac{\\left( {1 + u}\\right) u}{1} < \\frac{u}{1} = u$", + "bbox": [ + 447, + 319, + 465, + 325 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/7b2f28c59d7f50c8b0b572afea6afa97a72cb33fe11d4e82a08413594a194c71.jpg", + "image_caption": [ + "Intracorporeal tracking Fov + OOCV" + ], + "image_footnote": [], + "bbox": [ + 553, + 290, + 666, + 352 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/bc78089af8921d384819590d80ba041c3235476bff83da1acc8f5c9aae9ca30f.jpg", + "image_caption": [ + "Sample surgical visual challenges" + ], + "image_footnote": [], + "bbox": [ + 686, + 253, + 754, + 284 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/16729a3eafacbdcea5c9786f5a15b60f9716df083e1561d869a55f7c2649cffc.jpg", + "image_caption": [ + "Blurring", + "Trocar view" + ], + "image_footnote": [], + "bbox": [ + 686, + 291, + 754, + 321 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/7b710929e27e75ae90105524806d27f6e76d4df1cb9dd6c16d065663c3c27177.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 756, + 253, + 823, + 284 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/c761f4abba5cfaad4418a4f9328434cd71bd448fb237fc9b9e805277c8174407.jpg", + "image_caption": [ + "Bleeding", + "Reflection" + ], + "image_footnote": [], + "bbox": [ + 756, + 291, + 823, + 321 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/1ab595c418f3f88042bde5c82b69227b8634f8f679f2ff8c963bde084fb8438a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 826, + 253, + 897, + 284 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/c20fd387dde5b8625bc4945b9a75bf5f4e0c913ad1c84f7daa96fc04ca2cba7b.jpg", + "image_caption": [ + "Crowded", + "Smoke" + ], + "image_footnote": [], + "bbox": [ + 826, + 291, + 895, + 321 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/2d9c6b533680090b55241b6e7e461bfee4690e482a7c90c6f473c827d74b4723.jpg", + "image_caption": [ + "Occlusion" + ], + "image_footnote": [], + "bbox": [ + 686, + 330, + 754, + 361 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/87d9566a6f14a86753e0a91700e3ded82fe6eafb79da1a37960515ed70e4b06e.jpg", + "image_caption": [ + "Lens fouling" + ], + "image_footnote": [], + "bbox": [ + 756, + 330, + 823, + 361 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/4a07ecb3239a181d166aaa80984ab3a38fab92751fd597b1205296c22d919770.jpg", + "image_caption": [ + "Blood stain" + ], + "image_footnote": [], + "bbox": [ + 826, + 330, + 897, + 361 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "field of view to provide real-time feedback to surgeons. Existing surgical tracking datasets [18, 53] often lack these levels of granularity and adaptability needed for comprehensive tool modeling. They generally follow generic tracking formalizations and struggle to capture the intricacies of surgical tool interactions, particularly when tools are replaced or move beyond camera visibility [6, 28, 53].", + "bbox": [ + 88, + 422, + 482, + 527 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address this gap, we introduce CholecTrack20, a novel dataset, based on laparoscopic cholecystectomy surgery, designed for multi-class multi-tool tracking from intraoperative, intracorporeal, and visibility perspectives. Derived from raw laparoscopic videos [39, 56], it includes detailed annotations such as spatial coordinates, tool categories, track identities (IDs), visual challenges, phase labels, and other scene attributes. This dataset enhances benchmarking resources for computer vision research and supports the development of AI models tailored to surgical tool tracking, phase recognition, and surgeon performance assessment. This paper provides a comprehensive overview of the data acquisition and annotation methodology, detailed data analysis, and technical validation. In addition, we conduct extensive benchmark experiments using state-of-the-art deep learning methods for object detection and tracking, evaluating performance across various surgical phases and visual challenges, discussing insightful findings.", + "bbox": [ + 89, + 529, + 482, + 816 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, the main contributions are:", + "bbox": [ + 109, + 819, + 375, + 833 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. Introduction of CholecTrack20, a pioneering dataset for multi-perspective tracking with extensive annotations.", + "2. Extensive experimental analysis validating the dataset's effectiveness for surgical tool detection and tracking." + ], + "bbox": [ + 89, + 839, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Insights from model performance analysis under diverse visual challenges, highlighting the utility of each tracking perspective for AI-driven surgical solutions.", + "bbox": [ + 511, + 422, + 903, + 467 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 513, + 502, + 661, + 517 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Object detection and tracking. Advances in object detection and tracking have been driven by datasets like COCO [29], KITTI [20], MOTChallenge [12], VisDrone [70], DanceTrack [51], and TAO [11], enabling progress in Single Object Tracking (SOT) [10], Multi-Object Tracking (MOT) [64, 66], and Multi-Class Multi-Object Tracking (MCMOT) [13, 26]. Applying these techniques to surgical tool tracking poses challenges especially in the phase of bleeding, smoke, rapid movements, and variable lighting. Traditional tracking, centered on visibility, struggles when tools leave the camera's view or are replaced during surgery.", + "bbox": [ + 511, + 534, + 906, + 702 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Surgical tool tracking and configurations. While electromagnetic and optical tracking methods [8, 17] have been explored in surgical domain, image-based approaches [53] better align with surgeons' view but face issues like identity fragmentation, identity switch, and low tracking accuracy [6, 16, 28, 36, 37, 46] especially with mid-procedure tool replacements and tools exiting/re-entering the field of view or body cavity. To fully capture tool usage complexity, it is crucial to consider multiple perspectives on tool trajectories. Existing datasets [16, 41] provide insights but lack comprehensive coverage of these scenarios, highlighting the need for detailed datasets addressing these complexities.", + "bbox": [ + 511, + 719, + 906, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "8943", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 89, + 89, + 225, + 107 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "CholecTrack20 is a detailed dataset for surgical tool tracking in laparoscopic cholecystectomy, offering binary and spatial annotations for tools, including identity, category, bounding box location, motion, operator, phase, activity, usage conditions, and visual challenges, essential for training and benchmarking surgical AI tool tracking algorithms.", + "bbox": [ + 89, + 114, + 483, + 205 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Data Acquisition and Collection", + "text_level": 1, + "bbox": [ + 89, + 218, + 372, + 234 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Data source. The raw videos are sourced from the publicly available Cholec80 [56] and CholecT50 [39] datasets, with appropriate permissions and adherence to license terms. Recorded at the University Hospital of Strasbourg with the aid of laparoscopic cameras, these videos document laparoscopic cholecystectomy surgeries.", + "bbox": [ + 89, + 242, + 482, + 333 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Video selection. Long videos were systematically chosen to represent surgical complexities and tool variability, capturing all key phases of laparoscopic cholecystectomy. High-quality videos with clear visuals were selected, subsampled from 25 FPS to 1 FPS for annotation, maintaining temporal consistency and clarity.", + "bbox": [ + 89, + 339, + 483, + 430 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Sensitive data handling. To protect patient and staff identities, out-of-body frames potentially revealing sensitive information, such as the identity of patients, clinical staff, or the operating room, were checked and anonymized following established techniques [25]. This ensures compliance with ethical standards and privacy regulations.", + "bbox": [ + 89, + 435, + 483, + 526 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Track Formalization", + "text_level": 1, + "bbox": [ + 89, + 537, + 285, + 551 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given a video dataset $D = \\{S_{1}, S_{2}, \\ldots, S_{n}\\}$ of laparoscopic surgeries, each sequence $S_{i}$ includes frames annotated with bounding boxes $B = [B_{1}, B_{2}, \\ldots, B_{M}]$ for tool locations and classes $C = \\{C_{1}, C_{2}, \\ldots, C_{N}\\}$ . Tools, manipulated by operators linked to trocar ports $P = [P_{1}, P_{2}, \\ldots, P_{M}]$ , are assigned unique track identities (ID) through time. The ID reassignment is guided by visual cues and clinical knowledge. Visually, it is based on class $c \\in C$ and location $b \\in B$ . Contextually, it considers the role and hand position of the surgeon operator linked to a unique trocar port $p \\in P$ . Tool tracking solves the association matrix $A(t)$ , where $A_{i,j} = 1$ indicates the $i$ -th tool in frame $t$ is associated with the $j$ -th tool in frame $t + 1$ , and $A_{i,j} = 0$ otherwise. The aim is to obtain tool trajectories $T = \\{T_{1}, T_{2}, \\ldots, T_{K}\\}$ , each uniquely identified by ID, taking into consideration the visibility, intracorporeal, and intraoperative use.", + "bbox": [ + 89, + 561, + 482, + 819 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.3. Multi-Perspective (MP) Trajectory", + "text_level": 1, + "bbox": [ + 89, + 830, + 393, + 848 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Defining tool trajectories in surgical procedures necessitates a unique approach, given the variability in tracking across different perspectives, formalized as follows:", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/cdd3ad09fa90314ec31990633fc7bc5d027b5ccaf34e715ad1c0f1c6a811af47.jpg", + "image_caption": [ + "Figure 2. Multi-perspective trajectories of surgical tool." + ], + "image_footnote": [], + "bbox": [ + 526, + 88, + 893, + 199 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "1. Intraoperative trajectory: This lifelong tracking starts with a tool's first appearance and ends at its last in a patient's body during a procedure. It requires re-identification post-occlusion, out-of-camera view, or reinsertion. This approach is vital for applications such as tool usage monitoring [2], inventory management [35], surgeon training [27], skill assessment [16], and tool usage pattern analysis [52].", + "bbox": [ + 511, + 253, + 903, + 358 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. Intracorporeal trajectory: Unique tracks begin when a tool enters the body and end when it exits through a trocar port, even if off-camera. If a tool exits outside the camera's view, the exit is inferred if another tool enters through the same trocar or the initial tool releases its grasp. This is essential for understanding surgical workflow, as some actions occur outside the camera's focus, like graspers holding tissue out of view to facilitate other tools' actions [27]. Intracorporeal tracking supports a range of AI tasks including action evaluation [55], skill assessment [16, 27], tool usage optimization [44], and surgical risk estimation [52].", + "bbox": [ + 511, + 359, + 905, + 525 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Visibility trajectory: Tracking starts with a tool's first appearance within the camera view and ends when it leaves the view. It requires re-identification (re-ID) after occlusions or brief periods of disappearance within a two-second tolerance. This method is useful for assessing surgeon actions [27] and skill training [14, 24], providing go-no-go decision support, and measuring economy of motion [48].", + "bbox": [ + 511, + 525, + 905, + 631 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Some existing studies [16, 37] follow the intraoperative tracking format, others [24, 46] employs visibility trajectory. The intracorporeal trajectory, being the most complex to annotate, are not well-represented in the literature. Our dataset is the first to provide fine-grained labels for all the three (Fig. 2), enabling statistical analyses such as tool usage counts, events counts, abdominal entries/ exits, tool idleness, out-of-camera view occurrences, and mean tracklets per perspective.", + "bbox": [ + 511, + 635, + 905, + 772 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.4. Data Annotation", + "text_level": 1, + "bbox": [ + 511, + 785, + 679, + 800 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Annotators and tools. Bounding boxes were annotated by four researchers skilled in surgical workflow analysis, supplemented by prior study data [37, 40, 57]. Annotation tools used include Anonymized System, and a custom Python tool for visualization, merging, and validation. A pre-designed guide, refined by surgical experts, described", + "bbox": [ + 511, + 810, + 905, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "8944", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/2ef49139e401ce7df0b6679663e28c05e35ea2784c443e6a2180d25bc0fe2a37.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 102, + 88, + 287, + 170 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/42a6ae560d1cd6bb2c532e6e510ca4a9867a782033eba65927042769b612b143.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 287, + 88, + 470, + 170 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/991eb9e1aca2d3fe836d87c367a7b876a7681aa7b15335098f5daa4ea7569b30.jpg", + "image_caption": [ + "Figure 3. Examples of images from CholecTrack20 tracking dataset with the labels overlaid on the raw images." + ], + "image_footnote": [], + "bbox": [ + 101, + 171, + 285, + 268 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/9697a46f490c604d565d8cb985f205b53829a40ad466e3b5f57f9f1f793cd190.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 285, + 171, + 470, + 268 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the labels and provided image guidance as needed.", + "bbox": [ + 89, + 345, + 426, + 361 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Label types and categories. Tool spatial positions were annotated using bounding box coordinates, while other labels were represented by class identities. Seven predominant tool categories were defined: (1) cold grasper, (2) bipolar grasper, (3) monopolar hook, (4) monopolar scissors, (5) clipper or clip applicer, (6) irrigator or suction device, and (7) specimen bag. Four operator categories were defined: (1) main surgeon left hand MSLH, (2) main surgeon right hand MSRH, (3) assistant surgeon right hand ASRH, and (4) null operator NULL. The assistant surgeon left hand ASLH holding the endoscopic camera is unreported. A total of eight visual challenges were noted: (1) blurring, (2) bleeding, (3) camera lens fouling, (4) crowded scene, (5) occlusion, (6) smoke, (7) specular light reflection, and (8) trocar view or under-coverage. The seven commonest surgical phases were annotated: (1) preparation, (2) calot triangle dissection, (3) gallbladder dissection, (4) clipping & cutting, (5) gallbladder packaging, (6) cleaning and coagulation, and (7) gallbladder extraction.", + "bbox": [ + 88, + 373, + 483, + 660 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Annotation process. Annotations involved drawing bounding boxes $[x, y, w, h]$ over tool tips tagged with tool class $c \\in C$ and operator class $p \\in P$ , following trocar ports for accurate surgeon identification. Surgical details such as phase, out-of-view statuses (camera/abdomen), tool entry/exit, and visual challenge attributes were annotated to aid accurate track assignment. Tool-tissue interaction labels from CholecT50 [39] provides additional help in perpetuating track identities. Annotations were reviewed at 25 FPS in uncertain cases, and underwent rigorous quality control, ensuring high-quality labels over two years. Fig. 3 presents samples of images from the CholecTrack20 dataset alongside their respective annotations illustrating the meticulous labeling system employed to ensure a rich dataset for detailed surgical tool tracking analysis.", + "bbox": [ + 89, + 674, + 483, + 902 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/727d093663a8add9db71df69c8ad56964edc3a0cf9236aecb28ebb6730f0c256.jpg", + "image_caption": [ + "Figure 4. Dataset statistics on the distributions of (a) surgical scene visual challenges across data splits (b) track labels across perspectives, averaged across videos. Track length in seconds." + ], + "image_footnote": [], + "bbox": [ + 517, + 87, + 903, + 244 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.5. Quality Assurance", + "text_level": 1, + "bbox": [ + 513, + 320, + 692, + 337 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Label agreement. Two label agreement metrics validate dataset quality: Jaccard Index for spatial overlap of bounding boxes and Cohen's Kappa Statistic for category labels. The findings of our three validation forms include:", + "bbox": [ + 511, + 345, + 906, + 405 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. Intra-rater agreement involves self-correction. We observe a Jaccard Index of $99.4\\%$ and Cohen's Kappa Scores of $94.6\\%$ for tools and $94.0\\%$ for operators.", + "2. Inter-rater agreement is evaluated on 20 random samples across raters. The Jaccard Index is $91.8\\%$ , while tool class labels achieve $95.2\\%$ and operator labels $92.7\\%$ Cohen's Kappa. Minor differences reflect high-quality annotations.", + "3. Label mediation uses a board-certified surgeon for ambiguities, particularly in operator labels. Out of 758 uncertain samples, 133 needed correction post-mediation." + ], + "bbox": [ + 511, + 409, + 906, + 575 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.6. Data Statistics", + "text_level": 1, + "bbox": [ + 513, + 589, + 661, + 604 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Quantitative overview. The dataset includes 20 surgical videos totaling over 14 hours, averaging 42 minutes per surgery. As detailed in Tab. 1, annotations cover 35,000 frames at 1 FPS ( $\\sim$ 875,000 at 25 FPS), with 65,000 bounding box labels, averaging two tools per frame. Fig. 4(a) summarizes track configurations across perspectives, showing tool usage by type and trajectory. For example, graspers average 27 minutes 39 seconds inside the body and are re-inserted 8.4 times, while irrigators are used for about 1 minute 30 seconds per surgery. The visibility perspective contributes approximately 2,000 trajectories. Fig. 4(b) outlines tool attributes and visual challenges in the dataset, with occlusion as most prevalent occurring up to 23,000 seconds, bleeding totals around 18,700 seconds. This analysis provides insights into the varied visual challenges encountered in laparoscopic cholecystectomy. Evaluating tracking methods across these challenges will reveal their strengths and weaknesses. Statistical details across dataset splits are presented in Fig. 4(b).", + "bbox": [ + 511, + 613, + 906, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "8945", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/68430eaeaa23c28a3064150a877355ee8fac76a6e759a10acf156d2416158043.jpg", + "table_caption": [ + "Table 1. Dataset comparison showing the scope, statistics, and attributes. Dataset marked $\\ddagger$ are not full-length videos." + ], + "table_footnote": [], + "table_body": "
DatasetTaskTrack PerspectivesStatisticsAttributes
VisibilityIntraoperativeIntracorporealNo. of VideosTotal Duration (s)Frame CountTool BoxesNo. of TrajectoriesSurgoen OperatorVisual ChallengeSurgical Phase
ATLAS Dione [47]‡Detection✓XX99-22467--XXX
Cholec80-locations [49]Detection✓XX-401140116471-XXX
Bouget et.al [5]Detection✓XX14-24763819-XXX
m2cai 16-tool-locations [21]Detection✓XX--25323038-XXX
EndoVis'15 [15]‡Tracking✓XX1654013500--XXX
Fathollahi et el [16]‡TrackingX✓X1527002700--XXX
RMIT [53]Detection & Tracking✓XX415001171XXX
CholecTrack20 (Ours)Detection & Tracking✓✓✓205058135000652002,624✓✓✓
", + "bbox": [ + 112, + 114, + 883, + 237 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Dataset comparison. Existing publicly available datasets primarily focus on single perspective trajectory [16, 46]. In contrast, our dataset introduces a novel approach by annotation different perspectives, as presented in Tab. 1, including visual challenges, phase details, activity labels, etc.", + "bbox": [ + 89, + 261, + 480, + 337 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Dataset split. The dataset is split at the procedure level into non-overlapping training, validation, and testing sets in a 5:1:4 ratio, preventing data leakage. Video distribution is balanced by procedure duration to ensure similar complexity and difficulty levels across all splits.", + "bbox": [ + 89, + 340, + 483, + 417 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.7. Data Analysis", + "text_level": 1, + "bbox": [ + 89, + 426, + 235, + 441 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We conduct a comprehensive analysis of the dataset to gain insights into label alignments and feature similarities, revealing correlations across labels. This would guide data preprocessing and feature selection when using the data. Our analysis as illustrated in Fig. 5, Fig. 6, and Fig. 7 encompasses four distinct dimensions and discussed further.", + "bbox": [ + 89, + 450, + 483, + 541 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Tracking vs. surgical tool type correlation. This analysis explores the relationship between different tool types and their tracks, providing insights into unique patterns associated with each tool during surgery. Fig. 5(a) illustrates center point locations of tools, color-coded by category over time in three videos. This shows that while some trajectories appear separable, they are mostly densely interwoven, suggesting the need for advanced modeling in tool tracking.", + "bbox": [ + 89, + 545, + 483, + 667 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Tracking vs. surgeon tool operator correlation. We analyzed the alignment between tool operator identities and tool locations, revealing dynamic interactions between surgeons and instruments. Fig. 5(b) shows a strong correlation between operators and their tools, attributed to the distinct positioning of trocars. This underscores the value of operator information in accurate track label generation.", + "bbox": [ + 89, + 670, + 483, + 776 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Tracking vs. surgical phase segmentation. This analysis delves into tracking tools across surgical phases, uncovering how tool utilization varies with procedural stages. Fig. 6 shows that complex phases, such as calot triangle and gallbladder dissection, exhibit densely packed trajectories due to prolonged duration and frequent tool manipulation. Simpler phases, like preparation and clipping, feature fewer trajectories, facilitating modeling. These insights inform deep", + "bbox": [ + 89, + 780, + 483, + 902 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/2704ea2eedc13fe4f82ba0dfcccc32e912588f04c97168c2a9529f1b83b00a1f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 526, + 260, + 643, + 334 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/e30acdeb9ffaf6458871f74f2c78a5a43eff1b0046104beac080fc5453784d20.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 645, + 261, + 763, + 333 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/df30a378ce2b2f618497745f07be265f0d2b8362c6a0be190fabad607ad066bb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 764, + 261, + 893, + 333 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/fd20cb4ebc8019ab066faa0e1295b19af1b36f0fb629cf34a62b3ed5c48672ea.jpg", + "image_caption": [ + "Figure 5. 3D visualization of label alignments showing the tool position over track time. The coloring is for grouping features according to: (a) tool classes and (b) tool operators." + ], + "image_footnote": [], + "bbox": [ + 524, + 340, + 643, + 412 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/eec379849419c74f4da22e6f8c89cdb6018b0fefe8e20085b144472b9aebb9e6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 645, + 342, + 764, + 412 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/42c63b944bf9a009018462133333d698939af8d13cad7e78bee608c7b0e93aa3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 767, + 342, + 893, + 412 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/3d86bcf08c4bd17bf3d61267a52a4078b20e098d3b11afc1020733986c438788.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 526, + 478, + 645, + 555 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/39700c595b61332c0e958a66956dfafc5292bc11bb732eb967609c2eeabe7b51.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 651, + 478, + 769, + 555 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/1fa3ae87e774469ab1b5874eff703d188649e6145e2a2c66d6d5efdb4f5a2fe9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 774, + 478, + 893, + 555 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/165faeef204e8994e051ad323cff3aa62bc454ac87e905eda49aa79ca01436ed.jpg", + "image_caption": [ + "Figure 6. 3D visualization of tracking across different surgical phases for some randomly selected videos." + ], + "image_footnote": [], + "bbox": [ + 527, + 559, + 648, + 637 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/fc8efa217cf617101faa9213f4fe7f70bdf65fba0f9c4c0d439170fed131677e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 651, + 559, + 767, + 637 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/82b7c4f8969ef70591d9fedc36a2984035c139e7c53df56f99af599f5b7306ed.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 774, + 559, + 893, + 637 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/d05cce21a7f9088eb75d4494392e2ac5866343b6c06e829dd50182d8a005d3b4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 524, + 685, + 643, + 755 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/4c67c568558578949091197344d85ce42020c8a7418f8c5edeadb1cbf6a61aa2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 651, + 685, + 769, + 756 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/a135a11defd19d4258a115911fe7da6820e18c2864f13af8a5ef4d0c931e2ae0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 774, + 685, + 893, + 756 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/27506cc6dc91c6f54b713a43bc59c9764373f1ea8d3ba19cdc98a2d18fa2b2ce.jpg", + "image_caption": [ + "Figure 7. Trajectories of selected tools with EMA over time. Plotted positions are computed as weighted combinations of the center coordinate of the bounding boxes, scaled by image size." + ], + "image_footnote": [], + "bbox": [ + 524, + 760, + 643, + 828 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/ae6516bec1d140b08e62fc15fed07bc071490342df98d577929bce0c3d209000.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 651, + 760, + 769, + 828 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/6ee3105fd554f60069040c5b19ed3184e560172ae9323787f64439938ae422b1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 774, + 760, + 893, + 828 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/00b52bd8d89ca5fca38d70292519e9e3ef5ced20bcbb9d81a2838e3419fbf648.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 633, + 829, + 772, + 844 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "8946", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/6d30ec45b110a782dae3335563cabadbb3e9d83d3691a539d4e8b3e80a88012c.jpg", + "table_caption": [ + "Table 2. Benchmark Results of SOTA Object Detectors on Surgical Tool Detection Dataset." + ], + "table_footnote": [], + "table_body": "
Detector \nModelDetection AP accross 3 thresholdsDetection AP per category. (% AP @ Θ = 0.5)Detection AP across surgical visual challengesSpeed
AP0.5↑AP0.75↑AP0.5:0.95↑GrasperBipolarHookScissorsClipperIrrigatorBagBleedingBlurSmokeCrowdedOccludedReflectionFoul LensTrocarFPS↑
Faster-RCNN [43]56.038.134.653.565.080.160.970.126.831.857.941.054.543.555.046.941.235.77.6
Cascade-RCNN [7]51.739.034.752.058.979.745.744.923.717.953.939.048.139.546.429.133.733.77.0
CenterNet [69]53.039.535.060.261.486.456.368.025.810.258.042.150.236.751.746.035.830.833.8
FCOS [54]43.531.528.151.244.374.749.254.221.97.247.840.651.515.140.842.729.717.67.7
SSD [30]61.937.836.175.262.291.663.472.922.540.864.549.358.357.562.453.947.742.630.9
PAA [23]64.544.941.169.679.089.268.774.237.628.967.155.665.055.064.656.051.247.58.5
Def-DETR [71]58.442.038.360.666.583.861.972.039.923.862.442.758.637.157.443.941.547.410.2
Swin-T [31]62.344.340.263.364.883.080.277.238.026.863.553.862.835.361.166.255.245.79.8
YOLOX [19]64.748.944.269.672.289.475.479.137.327.168.255.666.045.964.252.558.143.123.6
YOLOv7 [62]80.662.056.190.586.496.082.389.349.166.280.261.280.179.582.165.671.266.720.6
YOLOv8 [58]79.162.455.687.984.596.280.087.248.465.077.158.374.476.280.470.357.462.929.0
YOLOv9 [61]80.262.656.588.587.696.079.387.150.167.778.154.078.278.681.165.363.463.123.7
YOLOv10 [60]80.162.155.887.686.696.081.989.053.861.377.861.978.777.581.266.759.365.428.6
", + "bbox": [ + 91, + 114, + 903, + 263 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "learning model assessment and AI model development for surgical tool tracking.", + "bbox": [ + 89, + 287, + 480, + 318 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Tracking variance. Using Exponential Moving Average (EMA), we analyze tool tracking data, as shown in Fig. 7. By overlaying EMAs on tool trajectories, we visualize variance between actual trajectories and modeled tracks. This approach highlights high-variance frames, serving as challenging cases for benchmarking model robustness and accuracy. Through this analysis, we provide a valuable resource for researchers, emphasizing the need to focus on complex scenarios for a rigorous assessment of model capabilities.", + "bbox": [ + 89, + 321, + 482, + 455 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Benchmark and Experiments", + "text_level": 1, + "bbox": [ + 89, + 469, + 361, + 486 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Tool Detection", + "text_level": 1, + "bbox": [ + 89, + 493, + 238, + 508 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Models. Owing that tool detection is a fundamental part of tool tracking, we showcase the usability of the CholecTrack20 dataset for this task, by benchmarking several object detectors representing diverse methodologies. Faster-RCNN [43] and Cascade-RCNN [7] are anchor-based models, with Cascade-RCNN employing a multi-stage approach to refine detection accuracy. CenterNet [69] and FCOS [54] are anchor-free models, utilizing center points and direct bounding box regression for efficient detection. SSD [30] provides real-time performance with its multi-scale approach. Deformable-DETR [71] applies a transformer-based method for flexible feature processing, while Swin-T [31] uses hierarchical transformers with shifted windows. The YOLO models [19, 58, 60–62] feature advanced multiscale strategies for high accuracy and speed.", + "bbox": [ + 89, + 517, + 482, + 743 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation metrics. We evaluate tool detection using COCO standard average precision (AP) metrics (pycocotools, not ultralytics) across several thresholds, categories, and visual challenges. We also report model inference speed in frames per seconds (FPS) on a single NVIDIA GTX 1080 Ti (10 GB) GPU.", + "bbox": [ + 89, + 746, + 482, + 837 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Overall tool detection results. The detection performance of the models is summarized in Tab. 2. YOLOv7 [62] achieves the highest Average Precision (AP) of $56.1\\%$ , surpassing other models. YOLOX [19] follows with an AP", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "of $44.2\\%$ . Deformable-DETR [71] and Swin-T [31] show competitive results with APs of $38.3\\%$ and $40.2\\%$ , respectively. The anchor-free models such as CenterNet [69] and FCOS [54] demonstrate robust performance, with CenterNet achieving an AP of $35.0\\%$ and FCOS $28.1\\%$ . In comparison, Faster-RCNN [43] and Cascade-RCNN [7] deliver APs of $34.6\\%$ and $34.7\\%$ , respectively, showcasing their efficacy with anchor-based approaches. At IoU thresholds of 0.5 and 0.75, YOLOv7 leads with scores of $80.6\\%$ and $62.0\\%$ , respectively. In terms of inference speed, YOLO networks excel with real-time capacities exceeding 20 FPS, with CenterNet achieving the highest speed at 33.8 FPS.", + "bbox": [ + 511, + 287, + 903, + 469 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Class-wise detection results. Analyzing tool detection results per category (Tab. 2), YOLOv7 emerges as the top performer, dominating in all the 7 categories, achieving above $90\\%$ accuracy in 2 tool categories and above $80\\%$ in 5. Notably, the hook exhibits the highest tendency among tools, with AP scores ranging from $74.7\\%$ to $96.0\\%$ across all models. Conversely, irrigator and specimen bag pose challenges, likely due to unclear tool tip boundaries and the bag's deformable nature, respectively. Grasper, bipolar, scissors, and clipper show high detection rates.", + "bbox": [ + 511, + 472, + 903, + 625 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Detection under visual challenges. Tab. 2 presents the performance of the benchmark object detection models across different surgical visual challenges. Notably, YOLOv7 achieves the highest detection accuracy across most challenges, particularly excelling in scenarios involving bleeding, smoke, and crowded scenes. Conversely, detecting tools in blurred scenes, near trocars, and specular light reflection pose significant challenges for all models, with lower detection rates observed across the board.", + "bbox": [ + 511, + 626, + 903, + 762 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Tool Tracking", + "text_level": 1, + "bbox": [ + 511, + 771, + 658, + 787 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Models. We train and evaluate several state-of-the-art multi-object tracking (MOT) methods on the CholecTrack20 dataset, focusing on their ability to track surgical tools. OCSORT [33] and TransTrack [50] employ sophisticated tracking-by-detection frameworks, with TransTrack utilizing transformers to improve feature association. ByteTrack [67], Bot-SORT [1], and SMILETrack [63] use ad", + "bbox": [ + 511, + 794, + 903, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "8947", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/d34c8182af67dcccc990ee8c6782272582603e2bd9e5ac8d3e5ba2f19026685b.jpg", + "table_caption": [ + "Table 3. Benchmark Multi-Perspective Multi-Tool Tracking Results @ 25 FPS." + ], + "table_footnote": [], + "table_body": "
ModelHOTA MetricsCLEAR MetricsIdentity MetricsCount MetricsSpeed
TrackerHOTA↑DetA↑LocA↑AssA↑MOTA↑MOTP↑MT↑PT↓ML↓IDF1↑IDSW↓Frag↓#Dets#IDsFPS↑
Intraoperative Trajectory (Groundtruth counts: #Dets = 29994, #IDs = 70)
OCSORT [33]14.652.786.74.149.285.02432149.52921273121936333610.2
FairMOT [66]5.825.875.91.35.073.9324434.34227192415252445614.2
TransTrack [50]7.431.584.41.74.282.9936254.2475718992164040796.7
ByteTrack [67]15.870.685.73.667.084.0541229.54648242928440538316.4
Bot-SORT [1]17.470.785.44.469.683.75811110.2390723762930245018.7
SMILETrack [63]15.971.085.53.766.483.8551329.24968236928821576111.2
Intracorporeal Trajectory (Groundtruth counts: #Dets = 29994, #IDs = 247)
OCSORT [33]23.751.486.511.047.184.8115874518.12953279621797352610.2
FairMOT [66]7.519.776.12.95.474.019601686.02890149611287396214.2
TransTrack [50]13.131.584.45.54.682.98079888.7464817912164040796.7
ByteTrack [67]24.770.685.78.767.484.0176482316.94515229028440538316.4
Bot-SORT [1]27.070.785.410.470.083.7188382118.9377122382930045018.7
SMILETrack [63]24.966.785.58.966.783.8186392216.94868223228820577911.2
Visibility Trajectory (Groundtruth counts: #Dets = 29994, #IDs = 916)
SORT [4]17.439.585.27.821.483.313939937813.46619213816595884419.5
OCSORT [33]37.052.686.526.250.284.830037124535.92317226022197358710.2
FairMOT [66]15.325.075.89.57.173.75821864014.43140157415338487514.2
TransTrack [50]19.231.684.411.85.882.922428041216.1427314032164040796.7
ByteTrack [67]41.570.785.724.869.384.059121710836.83930170428440538316.4
Bot-SORT [1]44.770.885.528.772.083.76381849441.4318316382930045058.7
SMILETrack [63]41.371.085.624.468.983.861919210536.54227164128821575211.2
", + "bbox": [ + 109, + 114, + 883, + 429 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "vanced tracking algorithms, with Bot-SORT and SMILE-Track incorporating extra modules for enhanced robustness.", + "bbox": [ + 89, + 455, + 480, + 484 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Evaluation metrics. We assess benchmark models on variety of tracking metrics: higher-order tracking accuracy (HOTA) [32], CLEAR MOT metrics [3], identity metrics [45], counting metrics, and tracking speed. A pull request is made to the standard TrackEval [22] library integrating CholecTrack20 benchmark with all its exhaustive performance evaluation protocols recommended by this study.", + "bbox": [ + 89, + 484, + 480, + 592 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Multi-object tracking results. The performance of the evaluated tracking methods is summarized in Tab. 3. Models such as FairMOT and TransTrack show the lowest HOTA score of $5.8\\%$ and $7.4\\%$ respectively, highlighting challenges in maintaining tool identities over time. ByteTrack, Bot-SORT, and SMILETrack achieve higher HOTA scores, ranging from $15.7\\%$ to $17.4\\%$ , but still face difficulties in tool re-ID due to similarities among tools.", + "bbox": [ + 89, + 594, + 480, + 715 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Despite these advancements, there remains room for enhancement in identity association and tracking precision. The results also include metrics on the detection counts, unique identities assigned, and tracking speed, providing a comprehensive view of each method's performance.", + "bbox": [ + 89, + 715, + 480, + 792 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Multi-perspective tracking results. Looking at the different trajectory perspectives, Tab. 3 shows that visibility tracking is the easiest with most of the existing models showcasing their strengths. This is expected because deep learning models mostly rely on visual cues, which are captured by camera in the visibility track scenario. Here, Bot-SORT record a landslide top performance scores of $44.7\\%$", + "bbox": [ + 89, + 794, + 480, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "HOTA, $72.0\\%$ MOTA, and $41.4\\%$ IDF1. The intracorporeal tracking is the most challenging since the major factors marking the entry and exit of the tools from the body are not readily visible. A maximum of $27.0\\%$ HOTA suggest a decline on the leading Bot-SORT. New methods could leverage rich fine-grained history to estimate the out-of-view and out-of-body status of the tools for improve re-ID. The intraoperative trajectory comes in the middle in terms of difficulty. While it may be challenging to ascertain the persistence of a trajectory after re-insertion, the class features are also helpful especially for tools of different categories. Again the Bot-SORT, leverage camera compensation details, shows a better tendency of estimating the persistent identity of different tools of the same class with a $+1.5\\%$ HOTA higher than similarity and appearance features.", + "bbox": [ + 511, + 455, + 903, + 681 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Multi-class tracking results. In Fig. 8, we analyze tracking performance by tool class and observe that the grasper, despite having the most instances and being the most frequently used tool in the dataset, achieves the highest tracking accuracy across perspectives. Class-agnostic results reveal medium tracking accuracy for other commonly used tools like the bipolar, hook, and clipper, while rarely used tools (e.g., scissors, irrigator) have lower scores. Specimen bag tracking is affected by shape deformation, contents, states (open/closed, empty/filled), tool interactions, and fluid stains. Tracking surgical tools beyond the visual perspective remains challenging for all models tested.", + "bbox": [ + 511, + 685, + 903, + 867 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Tracking results under visual challenges. We evaluate tool tracking performance across various surgical condi", + "bbox": [ + 511, + 869, + 903, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "8948", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6c0ebe6b0fbb1a8b8a189ff7f167a91470dbae44ff502f32292b612acdc1b78f.jpg", + "image_caption": [ + "Figure 8. Results of multi-perspective tracking across seven surgical tool categories in the CholecTrack20 dataset." + ], + "image_footnote": [], + "bbox": [ + 93, + 87, + 480, + 257 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0bd0e055ac20de3fea600d3d4020d0c43922dd460bff00bd65b5230a9fa8769a.jpg", + "image_caption": [ + "Figure 9. Results of multi-perspective tracking across eight surgical visual challenges in the CholecTrack20 dataset." + ], + "image_footnote": [], + "bbox": [ + 91, + 308, + 483, + 479 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9c4a38df3ba2bf363b6978e0a35c9a44d5ee5beb1a89866fe74e773c32829407.jpg", + "image_caption": [ + "Figure 10. Results of multi-perspective tracking across seven surgical phases in the CholecTrack20 dataset." + ], + "image_footnote": [], + "bbox": [ + 91, + 529, + 483, + 709 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "tions using HOTA metrics in Fig. 9, providing insights into model interactions with complex surgical environments. The model performs well under blurring, reflections, and limited camera coverage, likely due to effective data augmentation. However, lens fouling, smoke, and occlusion present significant challenges, reducing accuracy.", + "bbox": [ + 89, + 773, + 482, + 866 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Tracking results by surgical phase. Fig. 10 shows performance across seven surgical phases, with the clipping and", + "bbox": [ + 89, + 869, + 483, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "cutting phases proving easiest to track due to limited activities and a linear progression. Preparation phase shows similar performance. Phases like Calot triangle dissection and gallbladder dissection exhibit comparable tracking results, while gallbladder packaging shows the most consistent tracking across perspectives. Overall, OC-SORT struggles the most, while Bot-SORT achieves the best result.", + "bbox": [ + 511, + 90, + 903, + 196 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Limitations and Gaps to Address", + "text_level": 1, + "bbox": [ + 511, + 204, + 803, + 220 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The SOTA tracking methods trained on CholecTrack20 reveal substantial limitations, with performance under $45\\%$ HOTA, which is insufficient for clinical translation. These models struggle with various visual challenges, such as smoke, bleeding, and specular light reflection, affecting detection and re-ID. Since location and appearance features alone are inadequate, especially for tools with similar appearances, this highlights the need to move beyond current cues and innovate more intuitive, context-aware methods for re-ID. CholecTrack20 serves as a critical foundation for exploring this direction, offering a dataset rich in diverse tracking perspectives and challenges, essential for developing more robust and clinically viable tracking solutions.", + "bbox": [ + 511, + 227, + 906, + 422 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 435, + 633, + 450 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we presented the CholecTrack20 dataset, a novel resource designed to advance the state-of-the-art in surgical tool tracking within computer vision. CholecTrack20 addresses a critical gap by providing comprehensive annotations and diverse tracking scenarios and tasks across various surgical phases and visual challenges. Key innovations of CholecTrack20 include multi-perspective tracking, which defines the start and termination of a tool track differently based on visibility, intracorporeal, or intraoperative contexts. It also features detailed annotations of surgical visual challenges and precise surgical phase segmentation. Our extensive benchmark experiments demonstrate the dataset's effectiveness in developing models for tool detection and multi-object tracking across these three distinct trajectory perspectives. Through evaluating several deep learning methodologies on the CholecTrack20 dataset, we gain insights into their strengths and weaknesses in handling multiple viewpoints or tracking perspectives, across surgical phases and diverse surgical scene visual challenges.", + "bbox": [ + 511, + 459, + 906, + 747 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "By introducing CholecTrack20 to the computer vision research community, we aim to stimulate new research directions and foster collaborations between the computer vision and surgical communities. This dataset serves as a benchmark for evaluating state-of-the-art algorithms and promotes the development of robust and reliable surgical assistance systems. We anticipate that CholecTrack20 will inspire innovative approaches and contribute significantly to advancements in surgical tool tracking and related fields. The dataset is released under the CC BY-NC license.", + "bbox": [ + 511, + 747, + 908, + 897 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8949", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 91, + 90, + 236, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported by French state funds managed within the Plan Investissements d'Avenir by the ANR under references: National AI Chair AI4ORSafety [ANR-20-CHIA-0029-01], DeepSurg [ANR-16-CE33-0009], IHU Strasbourg [ANR-10-IAHU-02] and by BPI France under references: project CONDOR, project 5G-OR. Joel L. Lavanchy received funding by the Swiss National Science Foundation (P500PM_206724, P5R5PM_217663). This work was granted access to the servers/HPC resources managed by CAMMA, IHU Strasbourg, Unistra Mesocentre, and GENCI-IDRIS [Grant 2021-AD011011638R3, 2021-AD011011638R4].", + "bbox": [ + 89, + 112, + 485, + 282 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 305, + 187, + 321 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Nir Aharon, Roy Orfaig, and Ben-Zion Bobrovsky. Botsort: Robust associations multi-pedestrian tracking. arXiv preprint arXiv:2206.14651, 2022. 6, 7", + "[2] Hassan Al Hajj, Mathieu Lamard, Pierre-Henri Conze, Béatrice Cochener, and Gwenolé Quellec. Monitoring tool usage in surgery videos using boosted convolutional and recurrent neural networks. Medical image analysis, 47:203-218, 2018. 3", + "[3] Keni Bernstein and Rainer Stiefelhagen. Evaluating multiple object tracking performance: the clear mot metrics. EURASIP Journal on Image and Video Processing, 2008:1-10, 2008. 7", + "[4] Alex Bewley, Zongyuan Ge, Lionel Ott, Fabio Ramos, and Ben Upcroft. Simple online and realtime tracking. In 2016 IEEE international conference on image processing (ICIP), pages 3464-3468. IEEE, 2016. 7", + "[5] David Bouget, Rodrigo Benenson, Mohamed Omran, Laurent Riffaud, Bernt Schiele, and Pierre Jannin. Detecting surgical tools by modelling local appearance and global shape. IEEE transactions on medical imaging, 34(12):2603-2617, 2015. 5", + "[6] David Bouget, Max Allan, Danail Stoyanov, and Pierre Jannin. Vision-based and marker-less surgical tool detection and tracking: a review of the literature. Medical image analysis, 35:633-654, 2017. 2", + "[7] Zhaowei Cai and Nuno Vasconcelos. Cascade r-cnn: High quality object detection and instance segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence, page 1-1, 2019. 6", + "[8] Magdalena K Chmarra, CA Grimbergen, and J Dankelman. Systems for tracking minimally invasive surgical instruments. Minimally Invasive Therapy & Allied Technologies, 16(6):328-340, 2007. 2", + "[9] Tobias Czempiel, Magdalini Paschali, Matthias Keicher, Walter Simson, Hubertus Feussner, Seong Tae Kim, and Nassir Navab. Tecn: Surgical phase recognition with multistage temporal convolutional networks. In Medical Image Computing and Computer Assisted Intervention MICCAI, pages 343-352. Springer, 2020. 1", + "[10] Martin Danelljan, Goutam Bhat, Fahad Shahbaz Khan, and Michael Felsberg. Eco: Efficient convolution operators" + ], + "bbox": [ + 93, + 330, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "for tracking. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6638-6646, 2017. 2", + "[11] Achal Dave, Tarasha Khurana, Pavel Tokmakov, Cordelia Schmid, and Deva Ramanan. Tao: A large-scale benchmark for tracking any object. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part V 16, pages 436–454. Springer, 2020. 2", + "[12] Patrick Dendorfer, Hamid Rezatofighi, Anton Milan, Javen Shi, Daniel Cremers, Ian Reid, Stefan Roth, Konrad Schindler, and Laura Leal-Taixe. Mot20: A benchmark for multi object tracking in crowded scenes. arXiv preprint arXiv:2003.09003, 2020. 1, 2", + "[13] Yunhao Du, Junfeng Wan, Yanyun Zhao, Binyu Zhang, Zhihang Tong, and Junhao Dong. Giaotracker: A comprehensive framework for mcmot with global information and optimizing strategies in visdrone 2021. In Proceedings of the IEEE/CVF International conference on computer vision, pages 2809-2819, 2021. 2", + "[14] Ariel Kate Dubin, Danielle Julian, Alyssa Tanaka, Patricia Mattingly, and Roger Smith. A model for predicting the gears score from virtual reality surgical simulator metrics. Surgical endoscopy, 32:3576-3581, 2018. 3", + "[15] EndoVis Sub-Instrument Challenge. Endovis sub-instrument grand challenge, 2024. Accessed: 2021-11-12. 5", + "[16] Mona Fathollahi, Mohammad Hasan Sarhan, Ramon Pena, Lela DiMonte, Anshu Gupta, Aishani Ataliwala, and Jocelyn Barker. Video-based surgical skills assessment using long term tool tracking. In Medical Image Computing and Computer Assisted Intervention-MICCAI 2022: 25th International Conference, Singapore, September 18-22, 2022, Proceedings, Part VII, pages 541-550. Springer, 2022. 1, 2, 3, 5", + "[17] Marvin P Fried, Jonathan Kleefield, Harsha Gopal, Edward Reardon, Bryan T Ho, and Frederick A Kuhn. Image-guided endoscopic surgery: results of accuracy and performance in a multicenter clinical study using an electromagnetic tracking system. The Laryngoscope, 107(5):594-601, 1997. 2", + "[18] Yixin Gao, S Swaroop Vedula, Carol E Reiley, Narges Ah-midi, Balakrishnan Varadarajan, Henry C Lin, Lingling Tao, Luca Zappella, Benjamin Bejar, David D Yuh, et al. Jhu-isi gesture and skill assessment working set (jigsaws): A surgical activity dataset for human motion modeling. In MICCAI workshop: M2cai, page 3, 2014. 2", + "[19] Zheng Ge, Songtao Liu, Feng Wang, Zeming Li, and Jian Sun. YOLOX: Exceeding yolo series in 2021. arXiv preprint arXiv:2107.08430, 2021. 6", + "[20] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In 2012 IEEE conference on computer vision and pattern recognition, pages 3354-3361. IEEE, 2012. 1, 2", + "[21] Amy Jin, Serena Yeung, Jeffrey Jopling, Jonathan Krause, Dan Azagury, Arnold Milstein, and Li Fei-Fei. Tool detection and operative skill assessment in surgical videos using region-based convolutional neural networks. In 2018 IEEE winter conference on applications of computer vision (WACV), pages 691-699. IEEE, 2018. 5" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "8950", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[22] Arne Hoffhues Jonathon Luiten. Trackeval. https://github.com/JonathonLuiten/TrackEval, 2020. Accessed: 16 October 2023. 7", + "[23] Kang Kim and Hee Seok Lee. Probabilistic anchor assignment with iou prediction for object detection. In ECCV, 2020. 6", + "[24] Joel L Lavanchy, Joel Zindel, Kadir Kirtac, Isabell Twick, Enes Hosgor, Daniel Candinas, and Guido Beldi. Automation of surgical skill assessment using a three-stage machine learning algorithm. Scientific reports, 11(1):5197, 2021. 3", + "[25] Joel L Lavanchy, Armine Vardazaryan, Pietro Mascagni, Didier Mutter, and Nicolas Padoy. Preserving privacy in surgical video analysis using a deep learning classifier to identify out-of-body scenes in endoscopic videos. Scientific reports, 13(1):9235, 2023. 3", + "[26] Byungjae Lee, Enkhbayar Erdenee, Songguo Jin, Mi Young Nam, Young Giu Jung, and Phill Kyu Rhee. Multi-class multi-object tracking using changing point detection. In Computer Vision-ECCV 2016 Workshops: Amsterdam, The Netherlands, October 8-10 and 15-16, 2016, Proceedings, Part II 14, pages 68-83. Springer, 2016. 2", + "[27] Dongheon Lee, Hyeong Won Yu, Hyungju Kwon, HyounJoong Kong, Kyu Eun Lee, and Hee Chan Kim. Evaluation of surgical skills during robotic surgery by deep learning-based multiple surgical instrument tracking in training and actual operations. Journal of clinical medicine, 9(6):1964, 2020. 3", + "[28] Eung-Joo Lee, William Plishker, Xinyang Liu, Shuvra S Bhattacharyya, and Raj Shekhar. Weakly supervised segmentation for real-time surgical tool tracking. Healthcare technology letters, 6(6):231-236, 2019. 2", + "[29] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 2", + "[30] Wei Liu, Dragomir Anguelov, Dumitru Erhan, Christian Szegedy, Scott Reed, Cheng-Yang Fu, and Alexander C. Berg. Ssd: Single shot multibox detector. ECCV, 2016. 6", + "[31] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. arXiv preprint arXiv:2103.14030, 2021. 6", + "[32] Jonathon Luiten, Aljosa Osep, Patrick Dendorfer, Philip Torr, Andreas Geiger, Laura Leal-Taixe, and Bastian Leibe. Hota: A higher order metric for evaluating multi-object tracking. International journal of computer vision, 129:548-578, 2021. 7", + "[33] Gerard Maggiolino, Adnan Ahmad, Jinkun Cao, and Kris Kitani. Deep oc-sort: Multi-pedestrian tracking by adaptive re-identification. arXiv preprint arXiv:2302.11813, 2023. 6, 7", + "[34] Lena Maier-Hein, Swaroop S Vedula, Stefanie Speidel, Nassir Navab, Ron Kikinis, Adrian Park, Matthias Eisenmann, Hubertus Feussner, Germain Forestier, Stamatia Giannarou, et al. Surgical data science for next-generation interventions. Nature Biomedical Engineering, 1(9):691-696, 2017. 1" + ], + "bbox": [ + 91, + 92, + 480, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[35] Afrooz Moatari-Kazerouni and Ygal Bendavid. Improving logistics processes of surgical instruments: case of RFID technology. Business Process Management Journal, 23(2):448-466, 2017. 3", + "[36] Chinedu Innocent Nwoye. Deep learning methods for the detection and recognition of surgical tools and activities in laparoscopic videos. PhD thesis, Université de Strasbourg, 2021. 1, 2", + "[37] Chinedu Innocent Nwoye, Didier Mutter, Jacques Marescaux, and Nicolas Padoy. Weakly supervised convolutional LSTM approach for tool tracking in laparoscopic videos. International journal of computer assisted radiology and surgery, 14:1059-1067, 2019. 1, 2, 3", + "[38] Chinedu Innocent Nwoye, Cristians Gonzalez, Tong Yu, Pietro Mascagni, Didier Mutter, Jacques Marescaux, and Nicolas Padoy. Recognition of instrument-tissue interactions in endoscopic videos via action triplets. In International Conference on Medical Image Computing and Computer-Assisted Intervention MICCAI, pages 364-374. Springer, 2020. 1", + "[39] Chinedu Innocent Nwoye, Tong Yu, Cristians Gonzalez, Barbara Seeliger, Pietro Mascagni, Didier Mutter, Jacques Marescaux, and Nicolas Padoy. Rendezvous: Attention mechanisms for the recognition of surgical action triplets in endoscopic videos. 78:102433, 2022. 2, 3, 4", + "[40] Chinedu Innocent Nwoye, Tong Yu, Saurav Sharma, Aditya Murali, Deepak Alapatt, Armine Vardazaryan, Kun Yuan, Jonas Hajek, Wolfgang Reiter, Amine Yamlahi, et al. Cholectriplet2022: Show me a tool and tell me the triplet—an endoscopic vision challenge for surgical action triplet detection. arXiv preprint arXiv:2302.06294, 2023. 3", + "[41] Liang Qiu, Changsheng Li, and Hongliang Ren. Real-time surgical instrument tracking in robot-assisted surgery using multi-domain convolutional neural network. Healthcare technology letters, 6(6):159-164, 2019. 2", + "[42] Sanat Ramesh, Vinkle Srivastav, Deepak Alapatt, Tong Yu, Aditya Murali, Luca Sestini, Chinedu Innocent Nwoye, Idris Hamoud, Saurav Sharma, Antoine Fleurentin, et al. Dissecting self-supervised learning methods for surgical computer vision. Medical Image Analysis, 88:102844, 2023. 1", + "[43] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2017. 6", + "[44] Rogério Richa, Marcin Balicki, Eric Meisner, Raphael Sznitman, Russell Taylor, and Gregory Hager. Visual tracking of surgical tools for proximity detection in retinal surgery. In Information Processing in Computer-Assisted Interventions: Second International Conference, IPCAI 2011, Berlin, Germany, June 22, 2011. Proceedings 2, pages 55-66. Springer, 2011. 3", + "[45] Ergys Ristani, Francesco Solera, Roger Zou, Rita Cucchiara, and Carlo Tomasi. Performance measures and a data set for multi-target, multi-camera tracking. In European conference on computer vision, pages 17-35. Springer, 2016. 7", + "[46] Maria Robu, Abdelrahim Kadkhodamohammadi, Imanol Luengo, and Danail Stoyanov. Towards real-time multiple" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "8951", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "surgical tool tracking. Computer Methods in Biomechanics and Biomedical Engineering: Imaging & Visualization, 9(3): 279-285, 2021. 1, 2, 3, 5", + "[47] Duygu Sarikaya, Jason J Corso, and Khurshid A Guru. Detection and localization of robotic tools in robot-assisted surgery videos using deep neural networks for region proposal and detection. IEEE transactions on medical imaging, 36(7):1542-1549, 2017. 5", + "[48] Mohammad A Shbool, Ammar Al-Bazi, Alma Kokash, AlAlaween Wafa'H, Nibal T Albashabsheh, and Raed Al-Taher. The economy of motion for laparoscopic ball clamping surgery: A feedback educational tool. MethodsX, 10: 102168, 2023. 3", + "[49] Pan Shi, Zijian Zhao, Sanyuan Hu, and Faliang Chang. Realtime surgical tool detection in minimally invasive surgery based on attention-guided convolutional neural network. IEEE Access, 8:228853-228862, 2020. 5", + "[50] Peize Sun, Jinkun Cao, Yi Jiang, Rufeng Zhang, Enze Xie, Zehuan Yuan, Changhu Wang, and Ping Luo. Transtrack: Multiple object tracking with transformer. arXiv preprint arXiv:2012.15460, 2020. 6, 7", + "[51] Peize Sun, Jinkun Cao, Yi Jiang, Zehuan Yuan, Song Bai, Kris Kitani, and Ping Luo. Dancetrack: Multi-object tracking in uniform appearance and diverse motion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20993-21002, 2022. 1, 2", + "[52] Xiaochuan Sun and Shahram Payandeh. Estimation of incision patterns based on visual tracking of surgical tools in minimally invasive surgery. In ASME International Mechanical Engineering Congress and Exposition, pages 75-83, 2010. 3", + "[53] Raphael Sznitman, Karim Ali, Rogerio Richa, Russell H Taylor, Gregory D Hager, and Pascal Fua. Data-driven visual tracking in retinal microsurgery. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 568-575. Springer, 2012. 2, 5", + "[54] Zhi Tian, Chunhua Shen, Hao Chen, and Tong He. Fcos: Fully convolutional one-stage object detection. arXiv preprint arXiv:1904.01355, 2019. 6", + "[55] Jay Toor, Avneesh Bhangu, Jesse Wolfstadt, Garry Bassi, Stanley Chung, Raja Rampersaud, William Mitchell, Joseph Milner, and Martin Koyle. Optimizing the surgical instrument tray to immediately increase efficiency and lower costs in the operating room. Canadian Journal of Surgery, 65(2): E275, 2022. 3", + "[56] Andru P Twinanda, Sherif Shehata, Didier Mutter, Jacques Marescaux, Michel De Mathelin, and Nicolas Padoy. Endonet: a deep architecture for recognition tasks on laparoscopic videos. IEEE transactions on medical imaging, 36 (1):86-97, 2016. 1, 2, 3", + "[57] Armine Vardazaryan, Didier Mutter, Jacques Marescaux, and Nicolas Padoy. Weakly-supervised learning for tool localization in laparoscopic videos. In MICCAI-LABEL, pages 169–179. Springer, 2018. 3", + "[58] Rejin Varghese and M Sambath. Yolov8: A novel object detection algorithm with enhanced performance and robustness. In 2024 International Conference on Advances in Data" + ], + "bbox": [ + 91, + 90, + 483, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Engineering and Intelligent Computing Systems (ADICS), pages 1-6. IEEE, 2024. 6", + "[59] Martin Wagner, Beat-Peter Müller-Stich, Anna Kisilenko, Duc Tran, Patrick Heger, Lars Mündermann, David M Lubotsky, Benjamin Müller, Tornike Davitashvili, Manuela Capek, et al. Comparative validation of machine learning algorithms for surgical workflow and skill analysis with the heichole benchmark. Medical Image Analysis, 86:102770, 2023. 1", + "[60] Ao Wang, Hui Chen, Lihao Liu, Kai Chen, Zijia Lin, Jungong Han, et al. Yolov10: Real-time end-to-end object detection. Advances in Neural Information Processing Systems, 37:107984-108011, 2024. 6", + "[61] Chien-Yao Wang and Hong-Yuan Mark Liao. Yolov9: Learning what you want to learn using programmable gradient information. 2024. 6", + "[62] Chien-Yao Wang, Alexey Bochkovskyi, and Hong-Yuan Mark Liao. Yolov7: Trainable bag-of-freebies sets new state-of-the-art for real-time object detectors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7464-7475, 2023. 6", + "[63] Yu-Hsiang Wang, Jun-Wei Hsieh, Ping-Yang Chen, Ming-Ching Chang, Hung Hin So, and Xin Li. Smiletrack: Similarity learning for occlusion-aware multiple object tracking, 2023. 6, 7", + "[64] Nicolai Wojke, Alex Bewley, and Dietrich Paulus. Simple online and realtime tracking with a deep association metric. In 2017 IEEE international conference on image processing (ICIP), pages 3645-3649. IEEE, 2017. 2", + "[65] Lisheng Xu, Haoran Zhang, Jiaole Wang, Ang Li, Shuang Song, Hongliang Ren, Lin Qi, Jason J Gu, and Max Q-H Meng. Information loss challenges in surgical navigation systems: From information fusion to ai-based approaches. Information Fusion, 2022. 1", + "[66] Yifu Zhang, Chunyu Wang, Xinggang Wang, Wenjun Zeng, and Wenyu Liu. Fairmot: On the fairness of detection and re-identification in multiple object tracking. International Journal of Computer Vision, 129:3069–3087, 2021. 2, 7", + "[67] Yifu Zhang, Peize Sun, Yi Jiang, Dongdong Yu, Fucheng Weng, Zehuan Yuan, Ping Luo, Wenyu Liu, and Xinggang Wang. Bytetrack: Multi-object tracking by associating every detection box. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXII, pages 1-21. Springer, 2022. 6, 7", + "[68] Zijian Zhao, Sandrine Voros, Zhaorui Chen, and Xiaolin Cheng. Surgical tool tracking based on two cnns: from coarse to fine. The Journal of Engineering, 2019(14):467-472, 2019. 1", + "[69] Xingyi Zhou, Dequan Wang, and Philipp Krahenbuhl. Objects as points. arXiv preprint arXiv:1904.07850, 2019. 6", + "[70] Pengfei Zhu, Longyin Wen, Dawei Du, Xiao Bian, Heng Fan, Qinghua Hu, and Haibin Ling. Detection and tracking meet drones challenge. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(11):7380-7399, 2021. 2", + "[71] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. In International Conference on Learning Representations, 2021. 6" + ], + "bbox": [ + 516, + 90, + 903, + 900 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "8952", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/CholecTrack20_ A Multi-Perspective Tracking Dataset for Surgical Tools/3671c43b-afb0-4272-8573-cad641013bba_model.json b/2025/CholecTrack20_ A Multi-Perspective Tracking Dataset for Surgical Tools/3671c43b-afb0-4272-8573-cad641013bba_model.json new file mode 100644 index 0000000000000000000000000000000000000000..5e6e496505f6e3c53295166ea4a4ce9d4db3883e --- /dev/null +++ b/2025/CholecTrack20_ A Multi-Perspective Tracking Dataset for Surgical Tools/3671c43b-afb0-4272-8573-cad641013bba_model.json @@ -0,0 +1,3027 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.133, + 0.131, + 0.868, + 0.154 + ], + "angle": 0, + "content": "CholecTrack20: A Multi-Perspective Tracking Dataset for Surgical Tools" + }, + { + "type": "text", + "bbox": [ + 0.224, + 0.18, + 0.773, + 0.218 + ], + "angle": 0, + "content": "Chinedu Innocent Nwoye\\(^{1,3}\\) Kareem Elgohary\\(^{1}\\) Anvita Srinivas\\(^{1}\\) Fauzan Zaid\\(^{1}\\) Joel L. Lavanchy\\(^{2}\\) Nicolas Padoy\\(^{1,3}\\)" + }, + { + "type": "text", + "bbox": [ + 0.221, + 0.222, + 0.777, + 0.239 + ], + "angle": 0, + "content": "1University of Strasbourg, CNRS, INSERM, ICube, UMR7357, Strasbourg, France" + }, + { + "type": "text", + "bbox": [ + 0.221, + 0.24, + 0.778, + 0.256 + ], + "angle": 0, + "content": "2University of Basel, University Digestive Health Care Center, Clarunis, Switzerland" + }, + { + "type": "text", + "bbox": [ + 0.417, + 0.258, + 0.583, + 0.274 + ], + "angle": 0, + "content": "\\(^{3}\\)IHU Strasbourg, France" + }, + { + "type": "text", + "bbox": [ + 0.295, + 0.278, + 0.697, + 0.291 + ], + "angle": 0, + "content": "https://github.com/camma-public/cholectrack20" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.327, + 0.327, + 0.343 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.359, + 0.486, + 0.784 + ], + "angle": 0, + "content": "Tool tracking in surgical videos is essential for advancing computer-assisted interventions, such as skill assessment, safety zone estimation, and human-machine collaboration. However, the lack of context-rich datasets limits AI applications in this field. Existing datasets rely on overly generic tracking formalizations that fail to capture surgical-specific dynamics, such as tools moving out of the camera's view or exiting the body. This results in less clinically relevant trajectories and a lack of flexibility for real-world surgical applications. Methods trained on these datasets often struggle with visual challenges such as smoke, reflection, and bleeding, further exposing the limitations of current approaches. We introduce CholecTrack20, a specialized dataset for multi-class, multi-tool tracking in surgical procedures. It redefines tracking formalization with three perspectives: (1) intraoperative, (2) intracorporeal, and (3) visibility, enabling adaptable and clinically meaningful tool trajectories. The dataset comprises 20 full-length surgical videos, annotated at 1 fps, yielding over 35K frames and 65K labeled tool instances. Annotations include spatial location, category, identity, operator, phase, and visual challenges. Benchmarking state-of-the-art methods on CholecTrack20 reveals significant performance gaps, with current approaches (< 45% HOTA) failing to meet the accuracy required for clinical translation. These findings motivate the need for advanced and intuitive tracking algorithms and establish CholecTrack20 as a foundation for developing robust AI-driven surgical assistance systems." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.814, + 0.222, + 0.83 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.485, + 0.903 + ], + "angle": 0, + "content": "The true impact of computer vision research lies in its practical applications, especially in critical fields like healthcare. Among these, surgery represents one of the most demanding domains, providing a definitive test for the capa" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.328, + 0.908, + 0.768 + ], + "angle": 0, + "content": "bilities of vision technologies. Surgical data science, on its own, has significantly advanced interventional healthcare by leveraging data-driven techniques to provide critical decision support to medical professionals [34]. A key area of this advancement is the analysis of endoscopic video data, which offers real-time insights into surgical procedures, aids in skill assessment, and helps predict complications [9, 56, 65]. Accurate tracking of surgical tools is central to these analyses, as it guides temporal progression of procedural phases and correlates with surgical actions and management of adverse events [38, 59]. Despite progress in computer vision, research [42] shows that deep learning models pretrained on general datasets often struggle with surgical contexts due to complex scene dynamics, diverse tool types, and challenging visual conditions such as bleeding, smoke, and variable lighting [12, 20, 51]. This highlights the need for domain-specific datasets tailored to the unique requirements of surgical tool tracking. Obtaining medical and surgical data for research is challenging due to ethical and practical constraints, and annotating it requires expert knowledge. Current methods for tool tracking largely focus on Single Object Tracking (SOT) [68], Multi-Class Tracking (MCT) with one tool per class [36, 37], or Multi-Object Tracking (MOT) treating all tools as a single class [16, 46]. However, these approaches often miss the complexities of Multi-Class Multi-Object Tracking (MC-MOT) specific to surgical contexts, where tools interact dynamically and may move out of the camera's field of view or within the body cavity." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Multi-perspective tracking addresses these challenges by defining tool trajectories across different viewpoints during surgical procedures. It includes three critical perspectives (illustrated in Fig. 1): (1) intraoperative - covering the entire procedure duration to monitor tool usage and assess surgical proficiency; (2) intracorporeal - focusing on tool tracks within the body to evaluate specific tasks and predict risks; and (3) visibility - tracking tools within the camera's" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8942" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.099, + 0.09, + 0.332, + 0.101 + ], + "angle": 0, + "content": "Surgeon operators positioning and tool placement" + }, + { + "type": "image_caption", + "bbox": [ + 0.383, + 0.091, + 0.583, + 0.101 + ], + "angle": 0, + "content": "Multi-perspective of surgical tool trajectory" + }, + { + "type": "image_caption", + "bbox": [ + 0.699, + 0.091, + 0.842, + 0.101 + ], + "angle": 0, + "content": "Sample data and label overlaid" + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.103, + 0.337, + 0.28 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.344, + 0.103, + 0.671, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.645, + 0.105, + 0.898, + 0.233 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.123, + 0.29, + 0.285, + 0.3 + ], + "angle": 0, + "content": "Considered surgical tool categories" + }, + { + "type": "image", + "bbox": [ + 0.104, + 0.302, + 0.15, + 0.325 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.114, + 0.327, + 0.142, + 0.334 + ], + "angle": 0, + "content": "grasper" + }, + { + "type": "image", + "bbox": [ + 0.154, + 0.302, + 0.201, + 0.325 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.165, + 0.327, + 0.192, + 0.334 + ], + "angle": 0, + "content": "bipolar" + }, + { + "type": "image", + "bbox": [ + 0.204, + 0.302, + 0.254, + 0.325 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.218, + 0.327, + 0.24, + 0.334 + ], + "angle": 0, + "content": "hook" + }, + { + "type": "image", + "bbox": [ + 0.255, + 0.302, + 0.302, + 0.325 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.261, + 0.327, + 0.293, + 0.334 + ], + "angle": 0, + "content": "scissors" + }, + { + "type": "image", + "bbox": [ + 0.121, + 0.336, + 0.17, + 0.366 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.181, + 0.336, + 0.225, + 0.366 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.231, + 0.336, + 0.278, + 0.366 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.319, + 0.291, + 0.399, + 0.353 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.338, + 0.353, + 0.394, + 0.366 + ], + "angle": 0, + "content": "Visibility tracking FoV" + }, + { + "type": "image_caption", + "bbox": [ + 0.4, + 0.291, + 0.411, + 0.298 + ], + "angle": 0, + "content": "1" + }, + { + "type": "image_caption", + "bbox": [ + 0.412, + 0.313, + 0.468, + 0.32 + ], + "angle": 0, + "content": "Operating room" + }, + { + "type": "image_caption", + "bbox": [ + 0.412, + 0.32, + 0.468, + 0.326 + ], + "angle": 0, + "content": "\\( \\left( 4\\right) \\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.412, + 0.326, + 0.468, + 0.333 + ], + "angle": 0, + "content": "\\( \\therefore m - 1 \\neq 0 \\) ;" + }, + { + "type": "image_caption", + "bbox": [ + 0.412, + 0.333, + 0.468, + 0.339 + ], + "angle": 0, + "content": "\\( \\therefore m - 1 \\neq 0 \\) ;" + }, + { + "type": "image_caption", + "bbox": [ + 0.412, + 0.339, + 0.468, + 0.346 + ], + "angle": 0, + "content": "" + }, + { + "type": "image_caption", + "bbox": [ + 0.448, + 0.313, + 0.467, + 0.32 + ], + "angle": 0, + "content": "room" + }, + { + "type": "image_caption", + "bbox": [ + 0.448, + 0.32, + 0.467, + 0.326 + ], + "angle": 0, + "content": "\\( \\frac{1 + u}{1} - \\frac{u}{1} = \\frac{\\left( {1 + u}\\right) u}{1} < \\frac{u}{1} = u \\)" + }, + { + "type": "image", + "bbox": [ + 0.555, + 0.291, + 0.668, + 0.353 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.581, + 0.353, + 0.66, + 0.366 + ], + "angle": 0, + "content": "Intracorporeal tracking Fov + OOCV" + }, + { + "type": "image_caption", + "bbox": [ + 0.715, + 0.243, + 0.869, + 0.252 + ], + "angle": 0, + "content": "Sample surgical visual challenges" + }, + { + "type": "image", + "bbox": [ + 0.687, + 0.254, + 0.755, + 0.285 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.706, + 0.286, + 0.735, + 0.292 + ], + "angle": 0, + "content": "Blurring" + }, + { + "type": "image", + "bbox": [ + 0.687, + 0.292, + 0.755, + 0.323 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.701, + 0.324, + 0.742, + 0.329 + ], + "angle": 0, + "content": "Trocar view" + }, + { + "type": "image", + "bbox": [ + 0.757, + 0.254, + 0.825, + 0.285 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.776, + 0.286, + 0.805, + 0.292 + ], + "angle": 0, + "content": "Bleeding" + }, + { + "type": "image", + "bbox": [ + 0.757, + 0.292, + 0.825, + 0.323 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.773, + 0.324, + 0.808, + 0.329 + ], + "angle": 0, + "content": "Reflection" + }, + { + "type": "image", + "bbox": [ + 0.828, + 0.254, + 0.898, + 0.285 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.847, + 0.286, + 0.878, + 0.292 + ], + "angle": 0, + "content": "Crowded" + }, + { + "type": "image", + "bbox": [ + 0.828, + 0.292, + 0.897, + 0.323 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.851, + 0.324, + 0.873, + 0.329 + ], + "angle": 0, + "content": "Smoke" + }, + { + "type": "image", + "bbox": [ + 0.687, + 0.332, + 0.755, + 0.362 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.704, + 0.363, + 0.737, + 0.368 + ], + "angle": 0, + "content": "Occlusion" + }, + { + "type": "image", + "bbox": [ + 0.757, + 0.332, + 0.825, + 0.362 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.771, + 0.363, + 0.812, + 0.369 + ], + "angle": 0, + "content": "Lens fouling" + }, + { + "type": "image", + "bbox": [ + 0.828, + 0.332, + 0.898, + 0.362 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.845, + 0.363, + 0.88, + 0.368 + ], + "angle": 0, + "content": "Blood stain" + }, + { + "type": "image_caption", + "bbox": [ + 0.115, + 0.382, + 0.88, + 0.396 + ], + "angle": 0, + "content": "Figure 1. Illustration of multi-perspective tracking in surgical domain and CholecTrack20 dataset labels for surgical tool tracking." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.423, + 0.483, + 0.528 + ], + "angle": 0, + "content": "field of view to provide real-time feedback to surgeons. Existing surgical tracking datasets [18, 53] often lack these levels of granularity and adaptability needed for comprehensive tool modeling. They generally follow generic tracking formalizations and struggle to capture the intricacies of surgical tool interactions, particularly when tools are replaced or move beyond camera visibility [6, 28, 53]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.53, + 0.483, + 0.817 + ], + "angle": 0, + "content": "To address this gap, we introduce CholecTrack20, a novel dataset, based on laparoscopic cholecystectomy surgery, designed for multi-class multi-tool tracking from intraoperative, intracorporeal, and visibility perspectives. Derived from raw laparoscopic videos [39, 56], it includes detailed annotations such as spatial coordinates, tool categories, track identities (IDs), visual challenges, phase labels, and other scene attributes. This dataset enhances benchmarking resources for computer vision research and supports the development of AI models tailored to surgical tool tracking, phase recognition, and surgeon performance assessment. This paper provides a comprehensive overview of the data acquisition and annotation methodology, detailed data analysis, and technical validation. In addition, we conduct extensive benchmark experiments using state-of-the-art deep learning methods for object detection and tracking, evaluating performance across various surgical phases and visual challenges, discussing insightful findings." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.82, + 0.376, + 0.834 + ], + "angle": 0, + "content": "In summary, the main contributions are:" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.84, + 0.483, + 0.871 + ], + "angle": 0, + "content": "1. Introduction of CholecTrack20, a pioneering dataset for multi-perspective tracking with extensive annotations." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.482, + 0.901 + ], + "angle": 0, + "content": "2. Extensive experimental analysis validating the dataset's effectiveness for surgical tool detection and tracking." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.84, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.423, + 0.905, + 0.468 + ], + "angle": 0, + "content": "3. Insights from model performance analysis under diverse visual challenges, highlighting the utility of each tracking perspective for AI-driven surgical solutions." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.503, + 0.663, + 0.518 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.535, + 0.907, + 0.703 + ], + "angle": 0, + "content": "Object detection and tracking. Advances in object detection and tracking have been driven by datasets like COCO [29], KITTI [20], MOTChallenge [12], VisDrone [70], DanceTrack [51], and TAO [11], enabling progress in Single Object Tracking (SOT) [10], Multi-Object Tracking (MOT) [64, 66], and Multi-Class Multi-Object Tracking (MCMOT) [13, 26]. Applying these techniques to surgical tool tracking poses challenges especially in the phase of bleeding, smoke, rapid movements, and variable lighting. Traditional tracking, centered on visibility, struggles when tools leave the camera's view or are replaced during surgery." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Surgical tool tracking and configurations. While electromagnetic and optical tracking methods [8, 17] have been explored in surgical domain, image-based approaches [53] better align with surgeons' view but face issues like identity fragmentation, identity switch, and low tracking accuracy [6, 16, 28, 36, 37, 46] especially with mid-procedure tool replacements and tools exiting/re-entering the field of view or body cavity. To fully capture tool usage complexity, it is crucial to consider multiple perspectives on tool trajectories. Existing datasets [16, 41] provide insights but lack comprehensive coverage of these scenarios, highlighting the need for detailed datasets addressing these complexities." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8943" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.09, + 0.226, + 0.108 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.116, + 0.485, + 0.207 + ], + "angle": 0, + "content": "CholecTrack20 is a detailed dataset for surgical tool tracking in laparoscopic cholecystectomy, offering binary and spatial annotations for tools, including identity, category, bounding box location, motion, operator, phase, activity, usage conditions, and visual challenges, essential for training and benchmarking surgical AI tool tracking algorithms." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.219, + 0.373, + 0.236 + ], + "angle": 0, + "content": "3.1. Data Acquisition and Collection" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.243, + 0.483, + 0.334 + ], + "angle": 0, + "content": "Data source. The raw videos are sourced from the publicly available Cholec80 [56] and CholecT50 [39] datasets, with appropriate permissions and adherence to license terms. Recorded at the University Hospital of Strasbourg with the aid of laparoscopic cameras, these videos document laparoscopic cholecystectomy surgeries." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.34, + 0.484, + 0.431 + ], + "angle": 0, + "content": "Video selection. Long videos were systematically chosen to represent surgical complexities and tool variability, capturing all key phases of laparoscopic cholecystectomy. High-quality videos with clear visuals were selected, subsampled from 25 FPS to 1 FPS for annotation, maintaining temporal consistency and clarity." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.436, + 0.484, + 0.527 + ], + "angle": 0, + "content": "Sensitive data handling. To protect patient and staff identities, out-of-body frames potentially revealing sensitive information, such as the identity of patients, clinical staff, or the operating room, were checked and anonymized following established techniques [25]. This ensures compliance with ethical standards and privacy regulations." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.538, + 0.286, + 0.552 + ], + "angle": 0, + "content": "3.2. Track Formalization" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.562, + 0.483, + 0.82 + ], + "angle": 0, + "content": "Given a video dataset \\( D = \\{S_{1}, S_{2}, \\ldots, S_{n}\\} \\) of laparoscopic surgeries, each sequence \\( S_{i} \\) includes frames annotated with bounding boxes \\( B = [B_{1}, B_{2}, \\ldots, B_{M}] \\) for tool locations and classes \\( C = \\{C_{1}, C_{2}, \\ldots, C_{N}\\} \\). Tools, manipulated by operators linked to trocar ports \\( P = [P_{1}, P_{2}, \\ldots, P_{M}] \\), are assigned unique track identities (ID) through time. The ID reassignment is guided by visual cues and clinical knowledge. Visually, it is based on class \\( c \\in C \\) and location \\( b \\in B \\). Contextually, it considers the role and hand position of the surgeon operator linked to a unique trocar port \\( p \\in P \\). Tool tracking solves the association matrix \\( A(t) \\), where \\( A_{i,j} = 1 \\) indicates the \\( i \\)-th tool in frame \\( t \\) is associated with the \\( j \\)-th tool in frame \\( t + 1 \\), and \\( A_{i,j} = 0 \\) otherwise. The aim is to obtain tool trajectories \\( T = \\{T_{1}, T_{2}, \\ldots, T_{K}\\} \\), each uniquely identified by ID, taking into consideration the visibility, intracorporeal, and intraoperative use." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.832, + 0.394, + 0.849 + ], + "angle": 0, + "content": "3.3. Multi-Perspective (MP) Trajectory" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Defining tool trajectories in surgical procedures necessitates a unique approach, given the variability in tracking across different perspectives, formalized as follows:" + }, + { + "type": "image", + "bbox": [ + 0.527, + 0.089, + 0.895, + 0.2 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.544, + 0.21, + 0.875, + 0.224 + ], + "angle": 0, + "content": "Figure 2. Multi-perspective trajectories of surgical tool." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.254, + 0.905, + 0.359 + ], + "angle": 0, + "content": "1. Intraoperative trajectory: This lifelong tracking starts with a tool's first appearance and ends at its last in a patient's body during a procedure. It requires re-identification post-occlusion, out-of-camera view, or reinsertion. This approach is vital for applications such as tool usage monitoring [2], inventory management [35], surgeon training [27], skill assessment [16], and tool usage pattern analysis [52]." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.36, + 0.906, + 0.526 + ], + "angle": 0, + "content": "2. Intracorporeal trajectory: Unique tracks begin when a tool enters the body and end when it exits through a trocar port, even if off-camera. If a tool exits outside the camera's view, the exit is inferred if another tool enters through the same trocar or the initial tool releases its grasp. This is essential for understanding surgical workflow, as some actions occur outside the camera's focus, like graspers holding tissue out of view to facilitate other tools' actions [27]. Intracorporeal tracking supports a range of AI tasks including action evaluation [55], skill assessment [16, 27], tool usage optimization [44], and surgical risk estimation [52]." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.526, + 0.906, + 0.632 + ], + "angle": 0, + "content": "3. Visibility trajectory: Tracking starts with a tool's first appearance within the camera view and ends when it leaves the view. It requires re-identification (re-ID) after occlusions or brief periods of disappearance within a two-second tolerance. This method is useful for assessing surgeon actions [27] and skill training [14, 24], providing go-no-go decision support, and measuring economy of motion [48]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.636, + 0.906, + 0.773 + ], + "angle": 0, + "content": "Some existing studies [16, 37] follow the intraoperative tracking format, others [24, 46] employs visibility trajectory. The intracorporeal trajectory, being the most complex to annotate, are not well-represented in the literature. Our dataset is the first to provide fine-grained labels for all the three (Fig. 2), enabling statistical analyses such as tool usage counts, events counts, abdominal entries/ exits, tool idleness, out-of-camera view occurrences, and mean tracklets per perspective." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.786, + 0.681, + 0.801 + ], + "angle": 0, + "content": "3.4. Data Annotation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.906, + 0.903 + ], + "angle": 0, + "content": "Annotators and tools. Bounding boxes were annotated by four researchers skilled in surgical workflow analysis, supplemented by prior study data [37, 40, 57]. Annotation tools used include Anonymized System, and a custom Python tool for visualization, merging, and validation. A pre-designed guide, refined by surgical experts, described" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8944" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.104, + 0.089, + 0.288, + 0.171 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.288, + 0.089, + 0.471, + 0.171 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.172, + 0.286, + 0.269 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.286, + 0.172, + 0.471, + 0.269 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.282, + 0.483, + 0.311 + ], + "angle": 0, + "content": "Figure 3. Examples of images from CholecTrack20 tracking dataset with the labels overlaid on the raw images." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.346, + 0.427, + 0.362 + ], + "angle": 0, + "content": "the labels and provided image guidance as needed." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.374, + 0.485, + 0.661 + ], + "angle": 0, + "content": "Label types and categories. Tool spatial positions were annotated using bounding box coordinates, while other labels were represented by class identities. Seven predominant tool categories were defined: (1) cold grasper, (2) bipolar grasper, (3) monopolar hook, (4) monopolar scissors, (5) clipper or clip applicer, (6) irrigator or suction device, and (7) specimen bag. Four operator categories were defined: (1) main surgeon left hand MSLH, (2) main surgeon right hand MSRH, (3) assistant surgeon right hand ASRH, and (4) null operator NULL. The assistant surgeon left hand ASLH holding the endoscopic camera is unreported. A total of eight visual challenges were noted: (1) blurring, (2) bleeding, (3) camera lens fouling, (4) crowded scene, (5) occlusion, (6) smoke, (7) specular light reflection, and (8) trocar view or under-coverage. The seven commonest surgical phases were annotated: (1) preparation, (2) calot triangle dissection, (3) gallbladder dissection, (4) clipping & cutting, (5) gallbladder packaging, (6) cleaning and coagulation, and (7) gallbladder extraction." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.675, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Annotation process. Annotations involved drawing bounding boxes \\([x, y, w, h]\\) over tool tips tagged with tool class \\(c \\in C\\) and operator class \\(p \\in P\\), following trocar ports for accurate surgeon identification. Surgical details such as phase, out-of-view statuses (camera/abdomen), tool entry/exit, and visual challenge attributes were annotated to aid accurate track assignment. Tool-tissue interaction labels from CholecT50 [39] provides additional help in perpetuating track identities. Annotations were reviewed at 25 FPS in uncertain cases, and underwent rigorous quality control, ensuring high-quality labels over two years. Fig. 3 presents samples of images from the CholecTrack20 dataset alongside their respective annotations illustrating the meticulous labeling system employed to ensure a rich dataset for detailed surgical tool tracking analysis." + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.088, + 0.905, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.253, + 0.908, + 0.296 + ], + "angle": 0, + "content": "Figure 4. Dataset statistics on the distributions of (a) surgical scene visual challenges across data splits (b) track labels across perspectives, averaged across videos. Track length in seconds." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.321, + 0.694, + 0.338 + ], + "angle": 0, + "content": "3.5. Quality Assurance" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.346, + 0.907, + 0.406 + ], + "angle": 0, + "content": "Label agreement. Two label agreement metrics validate dataset quality: Jaccard Index for spatial overlap of bounding boxes and Cohen's Kappa Statistic for category labels. The findings of our three validation forms include:" + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.41, + 0.907, + 0.455 + ], + "angle": 0, + "content": "1. Intra-rater agreement involves self-correction. We observe a Jaccard Index of \\(99.4\\%\\) and Cohen's Kappa Scores of \\(94.6\\%\\) for tools and \\(94.0\\%\\) for operators." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.455, + 0.906, + 0.53 + ], + "angle": 0, + "content": "2. Inter-rater agreement is evaluated on 20 random samples across raters. The Jaccard Index is \\(91.8\\%\\), while tool class labels achieve \\(95.2\\%\\) and operator labels \\(92.7\\%\\) Cohen's Kappa. Minor differences reflect high-quality annotations." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.53, + 0.905, + 0.577 + ], + "angle": 0, + "content": "3. Label mediation uses a board-certified surgeon for ambiguities, particularly in operator labels. Out of 758 uncertain samples, 133 needed correction post-mediation." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.41, + 0.907, + 0.577 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.59, + 0.662, + 0.605 + ], + "angle": 0, + "content": "3.6. Data Statistics" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.614, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Quantitative overview. The dataset includes 20 surgical videos totaling over 14 hours, averaging 42 minutes per surgery. As detailed in Tab. 1, annotations cover 35,000 frames at 1 FPS (\\(\\sim\\)875,000 at 25 FPS), with 65,000 bounding box labels, averaging two tools per frame. Fig. 4(a) summarizes track configurations across perspectives, showing tool usage by type and trajectory. For example, graspers average 27 minutes 39 seconds inside the body and are re-inserted 8.4 times, while irrigators are used for about 1 minute 30 seconds per surgery. The visibility perspective contributes approximately 2,000 trajectories. Fig. 4(b) outlines tool attributes and visual challenges in the dataset, with occlusion as most prevalent occurring up to 23,000 seconds, bleeding totals around 18,700 seconds. This analysis provides insights into the varied visual challenges encountered in laparoscopic cholecystectomy. Evaluating tracking methods across these challenges will reveal their strengths and weaknesses. Statistical details across dataset splits are presented in Fig. 4(b)." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8945" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.147, + 0.09, + 0.848, + 0.104 + ], + "angle": 0, + "content": "Table 1. Dataset comparison showing the scope, statistics, and attributes. Dataset marked \\( \\ddagger \\) are not full-length videos." + }, + { + "type": "table", + "bbox": [ + 0.113, + 0.115, + 0.885, + 0.238 + ], + "angle": 0, + "content": "
DatasetTaskTrack PerspectivesStatisticsAttributes
VisibilityIntraoperativeIntracorporealNo. of VideosTotal Duration (s)Frame CountTool BoxesNo. of TrajectoriesSurgoen OperatorVisual ChallengeSurgical Phase
ATLAS Dione [47]‡Detection✓XX99-22467--XXX
Cholec80-locations [49]Detection✓XX-401140116471-XXX
Bouget et.al [5]Detection✓XX14-24763819-XXX
m2cai 16-tool-locations [21]Detection✓XX--25323038-XXX
EndoVis'15 [15]‡Tracking✓XX1654013500--XXX
Fathollahi et el [16]‡TrackingX✓X1527002700--XXX
RMIT [53]Detection & Tracking✓XX415001171XXX
CholecTrack20 (Ours)Detection & Tracking✓✓✓205058135000652002,624✓✓✓
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.262, + 0.482, + 0.338 + ], + "angle": 0, + "content": "Dataset comparison. Existing publicly available datasets primarily focus on single perspective trajectory [16, 46]. In contrast, our dataset introduces a novel approach by annotation different perspectives, as presented in Tab. 1, including visual challenges, phase details, activity labels, etc." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.342, + 0.484, + 0.418 + ], + "angle": 0, + "content": "Dataset split. The dataset is split at the procedure level into non-overlapping training, validation, and testing sets in a 5:1:4 ratio, preventing data leakage. Video distribution is balanced by procedure duration to ensure similar complexity and difficulty levels across all splits." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.427, + 0.236, + 0.443 + ], + "angle": 0, + "content": "3.7. Data Analysis" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.451, + 0.484, + 0.542 + ], + "angle": 0, + "content": "We conduct a comprehensive analysis of the dataset to gain insights into label alignments and feature similarities, revealing correlations across labels. This would guide data preprocessing and feature selection when using the data. Our analysis as illustrated in Fig. 5, Fig. 6, and Fig. 7 encompasses four distinct dimensions and discussed further." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.546, + 0.484, + 0.669 + ], + "angle": 0, + "content": "Tracking vs. surgical tool type correlation. This analysis explores the relationship between different tool types and their tracks, providing insights into unique patterns associated with each tool during surgery. Fig. 5(a) illustrates center point locations of tools, color-coded by category over time in three videos. This shows that while some trajectories appear separable, they are mostly densely interwoven, suggesting the need for advanced modeling in tool tracking." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.671, + 0.484, + 0.777 + ], + "angle": 0, + "content": "Tracking vs. surgeon tool operator correlation. We analyzed the alignment between tool operator identities and tool locations, revealing dynamic interactions between surgeons and instruments. Fig. 5(b) shows a strong correlation between operators and their tools, attributed to the distinct positioning of trocars. This underscores the value of operator information in accurate track label generation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Tracking vs. surgical phase segmentation. This analysis delves into tracking tools across surgical phases, uncovering how tool utilization varies with procedural stages. Fig. 6 shows that complex phases, such as calot triangle and gallbladder dissection, exhibit densely packed trajectories due to prolonged duration and frequent tool manipulation. Simpler phases, like preparation and clipping, feature fewer trajectories, facilitating modeling. These insights inform deep" + }, + { + "type": "image", + "bbox": [ + 0.527, + 0.261, + 0.645, + 0.335 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.647, + 0.262, + 0.764, + 0.334 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.766, + 0.262, + 0.895, + 0.334 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.526, + 0.342, + 0.645, + 0.414 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.647, + 0.343, + 0.765, + 0.413 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.768, + 0.343, + 0.895, + 0.414 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.427, + 0.907, + 0.469 + ], + "angle": 0, + "content": "Figure 5. 3D visualization of label alignments showing the tool position over track time. The coloring is for grouping features according to: (a) tool classes and (b) tool operators." + }, + { + "type": "image", + "bbox": [ + 0.527, + 0.479, + 0.646, + 0.556 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.652, + 0.479, + 0.77, + 0.556 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.776, + 0.479, + 0.895, + 0.556 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.528, + 0.56, + 0.649, + 0.638 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.653, + 0.56, + 0.769, + 0.638 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.776, + 0.56, + 0.895, + 0.638 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.648, + 0.907, + 0.677 + ], + "angle": 0, + "content": "Figure 6. 3D visualization of tracking across different surgical phases for some randomly selected videos." + }, + { + "type": "image", + "bbox": [ + 0.526, + 0.686, + 0.645, + 0.756 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.652, + 0.686, + 0.77, + 0.757 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.776, + 0.686, + 0.895, + 0.757 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.526, + 0.761, + 0.644, + 0.829 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.652, + 0.761, + 0.77, + 0.829 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.776, + 0.761, + 0.895, + 0.829 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.634, + 0.83, + 0.774, + 0.845 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.855, + 0.907, + 0.898 + ], + "angle": 0, + "content": "Figure 7. Trajectories of selected tools with EMA over time. Plotted positions are computed as weighted combinations of the center coordinate of the bounding boxes, scaled by image size." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8946" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.227, + 0.09, + 0.768, + 0.104 + ], + "angle": 0, + "content": "Table 2. Benchmark Results of SOTA Object Detectors on Surgical Tool Detection Dataset." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.115, + 0.905, + 0.264 + ], + "angle": 0, + "content": "
Detector \nModelDetection AP accross 3 thresholdsDetection AP per category. (% AP @ Θ = 0.5)Detection AP across surgical visual challengesSpeed
AP0.5↑AP0.75↑AP0.5:0.95↑GrasperBipolarHookScissorsClipperIrrigatorBagBleedingBlurSmokeCrowdedOccludedReflectionFoul LensTrocarFPS↑
Faster-RCNN [43]56.038.134.653.565.080.160.970.126.831.857.941.054.543.555.046.941.235.77.6
Cascade-RCNN [7]51.739.034.752.058.979.745.744.923.717.953.939.048.139.546.429.133.733.77.0
CenterNet [69]53.039.535.060.261.486.456.368.025.810.258.042.150.236.751.746.035.830.833.8
FCOS [54]43.531.528.151.244.374.749.254.221.97.247.840.651.515.140.842.729.717.67.7
SSD [30]61.937.836.175.262.291.663.472.922.540.864.549.358.357.562.453.947.742.630.9
PAA [23]64.544.941.169.679.089.268.774.237.628.967.155.665.055.064.656.051.247.58.5
Def-DETR [71]58.442.038.360.666.583.861.972.039.923.862.442.758.637.157.443.941.547.410.2
Swin-T [31]62.344.340.263.364.883.080.277.238.026.863.553.862.835.361.166.255.245.79.8
YOLOX [19]64.748.944.269.672.289.475.479.137.327.168.255.666.045.964.252.558.143.123.6
YOLOv7 [62]80.662.056.190.586.496.082.389.349.166.280.261.280.179.582.165.671.266.720.6
YOLOv8 [58]79.162.455.687.984.596.280.087.248.465.077.158.374.476.280.470.357.462.929.0
YOLOv9 [61]80.262.656.588.587.696.079.387.150.167.778.154.078.278.681.165.363.463.123.7
YOLOv10 [60]80.162.155.887.686.696.081.989.053.861.377.861.978.777.581.266.759.365.428.6
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.289, + 0.482, + 0.319 + ], + "angle": 0, + "content": "learning model assessment and AI model development for surgical tool tracking." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.322, + 0.483, + 0.457 + ], + "angle": 0, + "content": "Tracking variance. Using Exponential Moving Average (EMA), we analyze tool tracking data, as shown in Fig. 7. By overlaying EMAs on tool trajectories, we visualize variance between actual trajectories and modeled tracks. This approach highlights high-variance frames, serving as challenging cases for benchmarking model robustness and accuracy. Through this analysis, we provide a valuable resource for researchers, emphasizing the need to focus on complex scenarios for a rigorous assessment of model capabilities." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.47, + 0.362, + 0.487 + ], + "angle": 0, + "content": "4. Benchmark and Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.494, + 0.239, + 0.509 + ], + "angle": 0, + "content": "4.1. Tool Detection" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.518, + 0.483, + 0.744 + ], + "angle": 0, + "content": "Models. Owing that tool detection is a fundamental part of tool tracking, we showcase the usability of the CholecTrack20 dataset for this task, by benchmarking several object detectors representing diverse methodologies. Faster-RCNN [43] and Cascade-RCNN [7] are anchor-based models, with Cascade-RCNN employing a multi-stage approach to refine detection accuracy. CenterNet [69] and FCOS [54] are anchor-free models, utilizing center points and direct bounding box regression for efficient detection. SSD [30] provides real-time performance with its multi-scale approach. Deformable-DETR [71] applies a transformer-based method for flexible feature processing, while Swin-T [31] uses hierarchical transformers with shifted windows. The YOLO models [19, 58, 60–62] feature advanced multiscale strategies for high accuracy and speed." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.747, + 0.483, + 0.838 + ], + "angle": 0, + "content": "Evaluation metrics. We evaluate tool detection using COCO standard average precision (AP) metrics (pycocotools, not ultralytics) across several thresholds, categories, and visual challenges. We also report model inference speed in frames per seconds (FPS) on a single NVIDIA GTX 1080 Ti (10 GB) GPU." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Overall tool detection results. The detection performance of the models is summarized in Tab. 2. YOLOv7 [62] achieves the highest Average Precision (AP) of \\(56.1\\%\\), surpassing other models. YOLOX [19] follows with an AP" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.288, + 0.905, + 0.47 + ], + "angle": 0, + "content": "of \\(44.2\\%\\). Deformable-DETR [71] and Swin-T [31] show competitive results with APs of \\(38.3\\%\\) and \\(40.2\\%\\), respectively. The anchor-free models such as CenterNet [69] and FCOS [54] demonstrate robust performance, with CenterNet achieving an AP of \\(35.0\\%\\) and FCOS \\(28.1\\%\\). In comparison, Faster-RCNN [43] and Cascade-RCNN [7] deliver APs of \\(34.6\\%\\) and \\(34.7\\%\\), respectively, showcasing their efficacy with anchor-based approaches. At IoU thresholds of 0.5 and 0.75, YOLOv7 leads with scores of \\(80.6\\%\\) and \\(62.0\\%\\), respectively. In terms of inference speed, YOLO networks excel with real-time capacities exceeding 20 FPS, with CenterNet achieving the highest speed at 33.8 FPS." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.473, + 0.905, + 0.625 + ], + "angle": 0, + "content": "Class-wise detection results. Analyzing tool detection results per category (Tab. 2), YOLOv7 emerges as the top performer, dominating in all the 7 categories, achieving above \\(90\\%\\) accuracy in 2 tool categories and above \\(80\\%\\) in 5. Notably, the hook exhibits the highest tendency among tools, with AP scores ranging from \\(74.7\\%\\) to \\(96.0\\%\\) across all models. Conversely, irrigator and specimen bag pose challenges, likely due to unclear tool tip boundaries and the bag's deformable nature, respectively. Grasper, bipolar, scissors, and clipper show high detection rates." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.627, + 0.905, + 0.763 + ], + "angle": 0, + "content": "Detection under visual challenges. Tab. 2 presents the performance of the benchmark object detection models across different surgical visual challenges. Notably, YOLOv7 achieves the highest detection accuracy across most challenges, particularly excelling in scenarios involving bleeding, smoke, and crowded scenes. Conversely, detecting tools in blurred scenes, near trocars, and specular light reflection pose significant challenges for all models, with lower detection rates observed across the board." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.772, + 0.659, + 0.789 + ], + "angle": 0, + "content": "4.2. Tool Tracking" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Models. We train and evaluate several state-of-the-art multi-object tracking (MOT) methods on the CholecTrack20 dataset, focusing on their ability to track surgical tools. OCSORT [33] and TransTrack [50] employ sophisticated tracking-by-detection frameworks, with TransTrack utilizing transformers to improve feature association. ByteTrack [67], Bot-SORT [1], and SMILETrack [63] use ad" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8947" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.264, + 0.089, + 0.733, + 0.104 + ], + "angle": 0, + "content": "Table 3. Benchmark Multi-Perspective Multi-Tool Tracking Results @ 25 FPS." + }, + { + "type": "table", + "bbox": [ + 0.111, + 0.115, + 0.885, + 0.43 + ], + "angle": 0, + "content": "
ModelHOTA MetricsCLEAR MetricsIdentity MetricsCount MetricsSpeed
TrackerHOTA↑DetA↑LocA↑AssA↑MOTA↑MOTP↑MT↑PT↓ML↓IDF1↑IDSW↓Frag↓#Dets#IDsFPS↑
Intraoperative Trajectory (Groundtruth counts: #Dets = 29994, #IDs = 70)
OCSORT [33]14.652.786.74.149.285.02432149.52921273121936333610.2
FairMOT [66]5.825.875.91.35.073.9324434.34227192415252445614.2
TransTrack [50]7.431.584.41.74.282.9936254.2475718992164040796.7
ByteTrack [67]15.870.685.73.667.084.0541229.54648242928440538316.4
Bot-SORT [1]17.470.785.44.469.683.75811110.2390723762930245018.7
SMILETrack [63]15.971.085.53.766.483.8551329.24968236928821576111.2
Intracorporeal Trajectory (Groundtruth counts: #Dets = 29994, #IDs = 247)
OCSORT [33]23.751.486.511.047.184.8115874518.12953279621797352610.2
FairMOT [66]7.519.776.12.95.474.019601686.02890149611287396214.2
TransTrack [50]13.131.584.45.54.682.98079888.7464817912164040796.7
ByteTrack [67]24.770.685.78.767.484.0176482316.94515229028440538316.4
Bot-SORT [1]27.070.785.410.470.083.7188382118.9377122382930045018.7
SMILETrack [63]24.966.785.58.966.783.8186392216.94868223228820577911.2
Visibility Trajectory (Groundtruth counts: #Dets = 29994, #IDs = 916)
SORT [4]17.439.585.27.821.483.313939937813.46619213816595884419.5
OCSORT [33]37.052.686.526.250.284.830037124535.92317226022197358710.2
FairMOT [66]15.325.075.89.57.173.75821864014.43140157415338487514.2
TransTrack [50]19.231.684.411.85.882.922428041216.1427314032164040796.7
ByteTrack [67]41.570.785.724.869.384.059121710836.83930170428440538316.4
Bot-SORT [1]44.770.885.528.772.083.76381849441.4318316382930045058.7
SMILETrack [63]41.371.085.624.468.983.861919210536.54227164128821575211.2
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.456, + 0.482, + 0.486 + ], + "angle": 0, + "content": "vanced tracking algorithms, with Bot-SORT and SMILE-Track incorporating extra modules for enhanced robustness." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.486, + 0.482, + 0.593 + ], + "angle": 0, + "content": "Evaluation metrics. We assess benchmark models on variety of tracking metrics: higher-order tracking accuracy (HOTA) [32], CLEAR MOT metrics [3], identity metrics [45], counting metrics, and tracking speed. A pull request is made to the standard TrackEval [22] library integrating CholecTrack20 benchmark with all its exhaustive performance evaluation protocols recommended by this study." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.595, + 0.482, + 0.716 + ], + "angle": 0, + "content": "Multi-object tracking results. The performance of the evaluated tracking methods is summarized in Tab. 3. Models such as FairMOT and TransTrack show the lowest HOTA score of \\(5.8\\%\\) and \\(7.4\\%\\) respectively, highlighting challenges in maintaining tool identities over time. ByteTrack, Bot-SORT, and SMILETrack achieve higher HOTA scores, ranging from \\(15.7\\%\\) to \\(17.4\\%\\), but still face difficulties in tool re-ID due to similarities among tools." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.717, + 0.482, + 0.793 + ], + "angle": 0, + "content": "Despite these advancements, there remains room for enhancement in identity association and tracking precision. The results also include metrics on the detection counts, unique identities assigned, and tracking speed, providing a comprehensive view of each method's performance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.482, + 0.901 + ], + "angle": 0, + "content": "Multi-perspective tracking results. Looking at the different trajectory perspectives, Tab. 3 shows that visibility tracking is the easiest with most of the existing models showcasing their strengths. This is expected because deep learning models mostly rely on visual cues, which are captured by camera in the visibility track scenario. Here, Bot-SORT record a landslide top performance scores of \\(44.7\\%\\)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.456, + 0.905, + 0.683 + ], + "angle": 0, + "content": "HOTA, \\(72.0\\%\\) MOTA, and \\(41.4\\%\\) IDF1. The intracorporeal tracking is the most challenging since the major factors marking the entry and exit of the tools from the body are not readily visible. A maximum of \\(27.0\\%\\) HOTA suggest a decline on the leading Bot-SORT. New methods could leverage rich fine-grained history to estimate the out-of-view and out-of-body status of the tools for improve re-ID. The intraoperative trajectory comes in the middle in terms of difficulty. While it may be challenging to ascertain the persistence of a trajectory after re-insertion, the class features are also helpful especially for tools of different categories. Again the Bot-SORT, leverage camera compensation details, shows a better tendency of estimating the persistent identity of different tools of the same class with a \\(+1.5\\%\\) HOTA higher than similarity and appearance features." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.686, + 0.905, + 0.868 + ], + "angle": 0, + "content": "Multi-class tracking results. In Fig. 8, we analyze tracking performance by tool class and observe that the grasper, despite having the most instances and being the most frequently used tool in the dataset, achieves the highest tracking accuracy across perspectives. Class-agnostic results reveal medium tracking accuracy for other commonly used tools like the bipolar, hook, and clipper, while rarely used tools (e.g., scissors, irrigator) have lower scores. Specimen bag tracking is affected by shape deformation, contents, states (open/closed, empty/filled), tool interactions, and fluid stains. Tracking surgical tools beyond the visual perspective remains challenging for all models tested." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Tracking results under visual challenges. We evaluate tool tracking performance across various surgical condi" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8948" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.088, + 0.482, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.267, + 0.483, + 0.295 + ], + "angle": 0, + "content": "Figure 8. Results of multi-perspective tracking across seven surgical tool categories in the CholecTrack20 dataset." + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.309, + 0.485, + 0.48 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.489, + 0.483, + 0.517 + ], + "angle": 0, + "content": "Figure 9. Results of multi-perspective tracking across eight surgical visual challenges in the CholecTrack20 dataset." + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.53, + 0.485, + 0.71 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.719, + 0.483, + 0.748 + ], + "angle": 0, + "content": "Figure 10. Results of multi-perspective tracking across seven surgical phases in the CholecTrack20 dataset." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.775, + 0.483, + 0.867 + ], + "angle": 0, + "content": "tions using HOTA metrics in Fig. 9, providing insights into model interactions with complex surgical environments. The model performs well under blurring, reflections, and limited camera coverage, likely due to effective data augmentation. However, lens fouling, smoke, and occlusion present significant challenges, reducing accuracy." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Tracking results by surgical phase. Fig. 10 shows performance across seven surgical phases, with the clipping and" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.198 + ], + "angle": 0, + "content": "cutting phases proving easiest to track due to limited activities and a linear progression. Preparation phase shows similar performance. Phases like Calot triangle dissection and gallbladder dissection exhibit comparable tracking results, while gallbladder packaging shows the most consistent tracking across perspectives. Overall, OC-SORT struggles the most, while Bot-SORT achieves the best result." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.205, + 0.805, + 0.221 + ], + "angle": 0, + "content": "4.3. Limitations and Gaps to Address" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.228, + 0.907, + 0.424 + ], + "angle": 0, + "content": "The SOTA tracking methods trained on CholecTrack20 reveal substantial limitations, with performance under \\(45\\%\\) HOTA, which is insufficient for clinical translation. These models struggle with various visual challenges, such as smoke, bleeding, and specular light reflection, affecting detection and re-ID. Since location and appearance features alone are inadequate, especially for tools with similar appearances, this highlights the need to move beyond current cues and innovate more intuitive, context-aware methods for re-ID. CholecTrack20 serves as a critical foundation for exploring this direction, offering a dataset rich in diverse tracking perspectives and challenges, essential for developing more robust and clinically viable tracking solutions." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.436, + 0.634, + 0.451 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.46, + 0.907, + 0.748 + ], + "angle": 0, + "content": "In this work, we presented the CholecTrack20 dataset, a novel resource designed to advance the state-of-the-art in surgical tool tracking within computer vision. CholecTrack20 addresses a critical gap by providing comprehensive annotations and diverse tracking scenarios and tasks across various surgical phases and visual challenges. Key innovations of CholecTrack20 include multi-perspective tracking, which defines the start and termination of a tool track differently based on visibility, intracorporeal, or intraoperative contexts. It also features detailed annotations of surgical visual challenges and precise surgical phase segmentation. Our extensive benchmark experiments demonstrate the dataset's effectiveness in developing models for tool detection and multi-object tracking across these three distinct trajectory perspectives. Through evaluating several deep learning methodologies on the CholecTrack20 dataset, we gain insights into their strengths and weaknesses in handling multiple viewpoints or tracking perspectives, across surgical phases and diverse surgical scene visual challenges." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.748, + 0.909, + 0.898 + ], + "angle": 0, + "content": "By introducing CholecTrack20 to the computer vision research community, we aim to stimulate new research directions and foster collaborations between the computer vision and surgical communities. This dataset serves as a benchmark for evaluating state-of-the-art algorithms and promotes the development of robust and reliable surgical assistance systems. We anticipate that CholecTrack20 will inspire innovative approaches and contribute significantly to advancements in surgical tool tracking and related fields. The dataset is released under the CC BY-NC license." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8949" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.092, + 0.238, + 0.108 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.113, + 0.486, + 0.283 + ], + "angle": 0, + "content": "This work was supported by French state funds managed within the Plan Investissements d'Avenir by the ANR under references: National AI Chair AI4ORSafety [ANR-20-CHIA-0029-01], DeepSurg [ANR-16-CE33-0009], IHU Strasbourg [ANR-10-IAHU-02] and by BPI France under references: project CONDOR, project 5G-OR. Joel L. Lavanchy received funding by the Swiss National Science Foundation (P500PM_206724, P5R5PM_217663). This work was granted access to the servers/HPC resources managed by CAMMA, IHU Strasbourg, Unistra Mesocentre, and GENCI-IDRIS [Grant 2021-AD011011638R3, 2021-AD011011638R4]." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.306, + 0.188, + 0.322 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.331, + 0.484, + 0.373 + ], + "angle": 0, + "content": "[1] Nir Aharon, Roy Orfaig, and Ben-Zion Bobrovsky. Botsort: Robust associations multi-pedestrian tracking. arXiv preprint arXiv:2206.14651, 2022. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.373, + 0.484, + 0.442 + ], + "angle": 0, + "content": "[2] Hassan Al Hajj, Mathieu Lamard, Pierre-Henri Conze, Béatrice Cochener, and Gwenolé Quellec. Monitoring tool usage in surgery videos using boosted convolutional and recurrent neural networks. Medical image analysis, 47:203-218, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.443, + 0.484, + 0.497 + ], + "angle": 0, + "content": "[3] Keni Bernstein and Rainer Stiefelhagen. Evaluating multiple object tracking performance: the clear mot metrics. EURASIP Journal on Image and Video Processing, 2008:1-10, 2008. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.499, + 0.484, + 0.553 + ], + "angle": 0, + "content": "[4] Alex Bewley, Zongyuan Ge, Lionel Ott, Fabio Ramos, and Ben Upcroft. Simple online and realtime tracking. In 2016 IEEE international conference on image processing (ICIP), pages 3464-3468. IEEE, 2016. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.554, + 0.484, + 0.621 + ], + "angle": 0, + "content": "[5] David Bouget, Rodrigo Benenson, Mohamed Omran, Laurent Riffaud, Bernt Schiele, and Pierre Jannin. Detecting surgical tools by modelling local appearance and global shape. IEEE transactions on medical imaging, 34(12):2603-2617, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.623, + 0.484, + 0.678 + ], + "angle": 0, + "content": "[6] David Bouget, Max Allan, Danail Stoyanov, and Pierre Jannin. Vision-based and marker-less surgical tool detection and tracking: a review of the literature. Medical image analysis, 35:633-654, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.679, + 0.484, + 0.734 + ], + "angle": 0, + "content": "[7] Zhaowei Cai and Nuno Vasconcelos. Cascade r-cnn: High quality object detection and instance segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence, page 1-1, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.735, + 0.484, + 0.789 + ], + "angle": 0, + "content": "[8] Magdalena K Chmarra, CA Grimbergen, and J Dankelman. Systems for tracking minimally invasive surgical instruments. Minimally Invasive Therapy & Allied Technologies, 16(6):328-340, 2007. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.79, + 0.484, + 0.874 + ], + "angle": 0, + "content": "[9] Tobias Czempiel, Magdalini Paschali, Matthias Keicher, Walter Simson, Hubertus Feussner, Seong Tae Kim, and Nassir Navab. Tecn: Surgical phase recognition with multistage temporal convolutional networks. In Medical Image Computing and Computer Assisted Intervention MICCAI, pages 343-352. Springer, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.874, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[10] Martin Danelljan, Goutam Bhat, Fahad Shahbaz Khan, and Michael Felsberg. Eco: Efficient convolution operators" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.331, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.907, + 0.133 + ], + "angle": 0, + "content": "for tracking. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6638-6646, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.135, + 0.907, + 0.217 + ], + "angle": 0, + "content": "[11] Achal Dave, Tarasha Khurana, Pavel Tokmakov, Cordelia Schmid, and Deva Ramanan. Tao: A large-scale benchmark for tracking any object. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part V 16, pages 436–454. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.218, + 0.906, + 0.288 + ], + "angle": 0, + "content": "[12] Patrick Dendorfer, Hamid Rezatofighi, Anton Milan, Javen Shi, Daniel Cremers, Ian Reid, Stefan Roth, Konrad Schindler, and Laura Leal-Taixe. Mot20: A benchmark for multi object tracking in crowded scenes. arXiv preprint arXiv:2003.09003, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.289, + 0.907, + 0.37 + ], + "angle": 0, + "content": "[13] Yunhao Du, Junfeng Wan, Yanyun Zhao, Binyu Zhang, Zhihang Tong, and Junhao Dong. Giaotracker: A comprehensive framework for mcmot with global information and optimizing strategies in visdrone 2021. In Proceedings of the IEEE/CVF International conference on computer vision, pages 2809-2819, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.371, + 0.906, + 0.427 + ], + "angle": 0, + "content": "[14] Ariel Kate Dubin, Danielle Julian, Alyssa Tanaka, Patricia Mattingly, and Roger Smith. A model for predicting the gears score from virtual reality surgical simulator metrics. Surgical endoscopy, 32:3576-3581, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.428, + 0.906, + 0.455 + ], + "angle": 0, + "content": "[15] EndoVis Sub-Instrument Challenge. Endovis sub-instrument grand challenge, 2024. Accessed: 2021-11-12. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.457, + 0.906, + 0.565 + ], + "angle": 0, + "content": "[16] Mona Fathollahi, Mohammad Hasan Sarhan, Ramon Pena, Lela DiMonte, Anshu Gupta, Aishani Ataliwala, and Jocelyn Barker. Video-based surgical skills assessment using long term tool tracking. In Medical Image Computing and Computer Assisted Intervention-MICCAI 2022: 25th International Conference, Singapore, September 18-22, 2022, Proceedings, Part VII, pages 541-550. Springer, 2022. 1, 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.567, + 0.906, + 0.636 + ], + "angle": 0, + "content": "[17] Marvin P Fried, Jonathan Kleefield, Harsha Gopal, Edward Reardon, Bryan T Ho, and Frederick A Kuhn. Image-guided endoscopic surgery: results of accuracy and performance in a multicenter clinical study using an electromagnetic tracking system. The Laryngoscope, 107(5):594-601, 1997. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.637, + 0.906, + 0.72 + ], + "angle": 0, + "content": "[18] Yixin Gao, S Swaroop Vedula, Carol E Reiley, Narges Ah-midi, Balakrishnan Varadarajan, Henry C Lin, Lingling Tao, Luca Zappella, Benjamin Bejar, David D Yuh, et al. Jhu-isi gesture and skill assessment working set (jigsaws): A surgical activity dataset for human motion modeling. In MICCAI workshop: M2cai, page 3, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.721, + 0.906, + 0.761 + ], + "angle": 0, + "content": "[19] Zheng Ge, Songtao Liu, Feng Wang, Zeming Li, and Jian Sun. YOLOX: Exceeding yolo series in 2021. arXiv preprint arXiv:2107.08430, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.762, + 0.906, + 0.817 + ], + "angle": 0, + "content": "[20] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In 2012 IEEE conference on computer vision and pattern recognition, pages 3354-3361. IEEE, 2012. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.818, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[21] Amy Jin, Serena Yeung, Jeffrey Jopling, Jonathan Krause, Dan Azagury, Arnold Milstein, and Li Fei-Fei. Tool detection and operative skill assessment in surgical videos using region-based convolutional neural networks. In 2018 IEEE winter conference on applications of computer vision (WACV), pages 691-699. IEEE, 2018. 5" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "8950" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.133 + ], + "angle": 0, + "content": "[22] Arne Hoffhues Jonathon Luiten. Trackeval. https://github.com/JonathonLuiten/TrackEval, 2020. Accessed: 16 October 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.134, + 0.482, + 0.174 + ], + "angle": 0, + "content": "[23] Kang Kim and Hee Seok Lee. Probabilistic anchor assignment with iou prediction for object detection. In ECCV, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.176, + 0.482, + 0.232 + ], + "angle": 0, + "content": "[24] Joel L Lavanchy, Joel Zindel, Kadir Kirtac, Isabell Twick, Enes Hosgor, Daniel Candinas, and Guido Beldi. Automation of surgical skill assessment using a three-stage machine learning algorithm. Scientific reports, 11(1):5197, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.233, + 0.482, + 0.3 + ], + "angle": 0, + "content": "[25] Joel L Lavanchy, Armine Vardazaryan, Pietro Mascagni, Didier Mutter, and Nicolas Padoy. Preserving privacy in surgical video analysis using a deep learning classifier to identify out-of-body scenes in endoscopic videos. Scientific reports, 13(1):9235, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.302, + 0.482, + 0.385 + ], + "angle": 0, + "content": "[26] Byungjae Lee, Enkhbayar Erdenee, Songguo Jin, Mi Young Nam, Young Giu Jung, and Phill Kyu Rhee. Multi-class multi-object tracking using changing point detection. In Computer Vision-ECCV 2016 Workshops: Amsterdam, The Netherlands, October 8-10 and 15-16, 2016, Proceedings, Part II 14, pages 68-83. Springer, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.386, + 0.482, + 0.467 + ], + "angle": 0, + "content": "[27] Dongheon Lee, Hyeong Won Yu, Hyungju Kwon, HyounJoong Kong, Kyu Eun Lee, and Hee Chan Kim. Evaluation of surgical skills during robotic surgery by deep learning-based multiple surgical instrument tracking in training and actual operations. Journal of clinical medicine, 9(6):1964, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.469, + 0.482, + 0.524 + ], + "angle": 0, + "content": "[28] Eung-Joo Lee, William Plishker, Xinyang Liu, Shuvra S Bhattacharyya, and Raj Shekhar. Weakly supervised segmentation for real-time surgical tool tracking. Healthcare technology letters, 6(6):231-236, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.525, + 0.482, + 0.608 + ], + "angle": 0, + "content": "[29] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.609, + 0.482, + 0.65 + ], + "angle": 0, + "content": "[30] Wei Liu, Dragomir Anguelov, Dumitru Erhan, Christian Szegedy, Scott Reed, Cheng-Yang Fu, and Alexander C. Berg. Ssd: Single shot multibox detector. ECCV, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.651, + 0.482, + 0.705 + ], + "angle": 0, + "content": "[31] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. arXiv preprint arXiv:2103.14030, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.707, + 0.482, + 0.773 + ], + "angle": 0, + "content": "[32] Jonathon Luiten, Aljosa Osep, Patrick Dendorfer, Philip Torr, Andreas Geiger, Laura Leal-Taixe, and Bastian Leibe. Hota: A higher order metric for evaluating multi-object tracking. International journal of computer vision, 129:548-578, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.776, + 0.482, + 0.829 + ], + "angle": 0, + "content": "[33] Gerard Maggiolino, Adnan Ahmad, Jinkun Cao, and Kris Kitani. Deep oc-sort: Multi-pedestrian tracking by adaptive re-identification. arXiv preprint arXiv:2302.11813, 2023. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[34] Lena Maier-Hein, Swaroop S Vedula, Stefanie Speidel, Nassir Navab, Ron Kikinis, Adrian Park, Matthias Eisenmann, Hubertus Feussner, Germain Forestier, Stamatia Giannarou, et al. Surgical data science for next-generation interventions. Nature Biomedical Engineering, 1(9):691-696, 2017. 1" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[35] Afrooz Moatari-Kazerouni and Ygal Bendavid. Improving logistics processes of surgical instruments: case of RFID technology. Business Process Management Journal, 23(2):448-466, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.15, + 0.905, + 0.205 + ], + "angle": 0, + "content": "[36] Chinedu Innocent Nwoye. Deep learning methods for the detection and recognition of surgical tools and activities in laparoscopic videos. PhD thesis, Université de Strasbourg, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.207, + 0.905, + 0.276 + ], + "angle": 0, + "content": "[37] Chinedu Innocent Nwoye, Didier Mutter, Jacques Marescaux, and Nicolas Padoy. Weakly supervised convolutional LSTM approach for tool tracking in laparoscopic videos. International journal of computer assisted radiology and surgery, 14:1059-1067, 2019. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.278, + 0.905, + 0.373 + ], + "angle": 0, + "content": "[38] Chinedu Innocent Nwoye, Cristians Gonzalez, Tong Yu, Pietro Mascagni, Didier Mutter, Jacques Marescaux, and Nicolas Padoy. Recognition of instrument-tissue interactions in endoscopic videos via action triplets. In International Conference on Medical Image Computing and Computer-Assisted Intervention MICCAI, pages 364-374. Springer, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.376, + 0.905, + 0.445 + ], + "angle": 0, + "content": "[39] Chinedu Innocent Nwoye, Tong Yu, Cristians Gonzalez, Barbara Seeliger, Pietro Mascagni, Didier Mutter, Jacques Marescaux, and Nicolas Padoy. Rendezvous: Attention mechanisms for the recognition of surgical action triplets in endoscopic videos. 78:102433, 2022. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.447, + 0.905, + 0.531 + ], + "angle": 0, + "content": "[40] Chinedu Innocent Nwoye, Tong Yu, Saurav Sharma, Aditya Murali, Deepak Alapatt, Armine Vardazaryan, Kun Yuan, Jonas Hajek, Wolfgang Reiter, Amine Yamlahi, et al. Cholectriplet2022: Show me a tool and tell me the triplet—an endoscopic vision challenge for surgical action triplet detection. arXiv preprint arXiv:2302.06294, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.532, + 0.905, + 0.587 + ], + "angle": 0, + "content": "[41] Liang Qiu, Changsheng Li, and Hongliang Ren. Real-time surgical instrument tracking in robot-assisted surgery using multi-domain convolutional neural network. Healthcare technology letters, 6(6):159-164, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.589, + 0.905, + 0.658 + ], + "angle": 0, + "content": "[42] Sanat Ramesh, Vinkle Srivastav, Deepak Alapatt, Tong Yu, Aditya Murali, Luca Sestini, Chinedu Innocent Nwoye, Idris Hamoud, Saurav Sharma, Antoine Fleurentin, et al. Dissecting self-supervised learning methods for surgical computer vision. Medical Image Analysis, 88:102844, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.66, + 0.905, + 0.716 + ], + "angle": 0, + "content": "[43] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.718, + 0.905, + 0.813 + ], + "angle": 0, + "content": "[44] Rogério Richa, Marcin Balicki, Eric Meisner, Raphael Sznitman, Russell Taylor, and Gregory Hager. Visual tracking of surgical tools for proximity detection in retinal surgery. In Information Processing in Computer-Assisted Interventions: Second International Conference, IPCAI 2011, Berlin, Germany, June 22, 2011. Proceedings 2, pages 55-66. Springer, 2011. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.816, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[45] Ergys Ristani, Francesco Solera, Roger Zou, Rita Cucchiara, and Carlo Tomasi. Performance measures and a data set for multi-target, multi-camera tracking. In European conference on computer vision, pages 17-35. Springer, 2016. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[46] Maria Robu, Abdelrahim Kadkhodamohammadi, Imanol Luengo, and Danail Stoyanov. Towards real-time multiple" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.956 + ], + "angle": 0, + "content": "8951" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.484, + 0.134 + ], + "angle": 0, + "content": "surgical tool tracking. Computer Methods in Biomechanics and Biomedical Engineering: Imaging & Visualization, 9(3): 279-285, 2021. 1, 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.484, + 0.205 + ], + "angle": 0, + "content": "[47] Duygu Sarikaya, Jason J Corso, and Khurshid A Guru. Detection and localization of robotic tools in robot-assisted surgery videos using deep neural networks for region proposal and detection. IEEE transactions on medical imaging, 36(7):1542-1549, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.207, + 0.484, + 0.275 + ], + "angle": 0, + "content": "[48] Mohammad A Shbool, Ammar Al-Bazi, Alma Kokash, AlAlaween Wafa'H, Nibal T Albashabsheh, and Raed Al-Taher. The economy of motion for laparoscopic ball clamping surgery: A feedback educational tool. MethodsX, 10: 102168, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.278, + 0.484, + 0.332 + ], + "angle": 0, + "content": "[49] Pan Shi, Zijian Zhao, Sanyuan Hu, and Faliang Chang. Realtime surgical tool detection in minimally invasive surgery based on attention-guided convolutional neural network. IEEE Access, 8:228853-228862, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.334, + 0.484, + 0.389 + ], + "angle": 0, + "content": "[50] Peize Sun, Jinkun Cao, Yi Jiang, Rufeng Zhang, Enze Xie, Zehuan Yuan, Changhu Wang, and Ping Luo. Transtrack: Multiple object tracking with transformer. arXiv preprint arXiv:2012.15460, 2020. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.391, + 0.484, + 0.461 + ], + "angle": 0, + "content": "[51] Peize Sun, Jinkun Cao, Yi Jiang, Zehuan Yuan, Song Bai, Kris Kitani, and Ping Luo. Dancetrack: Multi-object tracking in uniform appearance and diverse motion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20993-21002, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.462, + 0.484, + 0.531 + ], + "angle": 0, + "content": "[52] Xiaochuan Sun and Shahram Payandeh. Estimation of incision patterns based on visual tracking of surgical tools in minimally invasive surgery. In ASME International Mechanical Engineering Congress and Exposition, pages 75-83, 2010. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.533, + 0.484, + 0.603 + ], + "angle": 0, + "content": "[53] Raphael Sznitman, Karim Ali, Rogerio Richa, Russell H Taylor, Gregory D Hager, and Pascal Fua. Data-driven visual tracking in retinal microsurgery. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 568-575. Springer, 2012. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.604, + 0.484, + 0.646 + ], + "angle": 0, + "content": "[54] Zhi Tian, Chunhua Shen, Hao Chen, and Tong He. Fcos: Fully convolutional one-stage object detection. arXiv preprint arXiv:1904.01355, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.647, + 0.484, + 0.73 + ], + "angle": 0, + "content": "[55] Jay Toor, Avneesh Bhangu, Jesse Wolfstadt, Garry Bassi, Stanley Chung, Raja Rampersaud, William Mitchell, Joseph Milner, and Martin Koyle. Optimizing the surgical instrument tray to immediately increase efficiency and lower costs in the operating room. Canadian Journal of Surgery, 65(2): E275, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.732, + 0.484, + 0.8 + ], + "angle": 0, + "content": "[56] Andru P Twinanda, Sherif Shehata, Didier Mutter, Jacques Marescaux, Michel De Mathelin, and Nicolas Padoy. Endonet: a deep architecture for recognition tasks on laparoscopic videos. IEEE transactions on medical imaging, 36 (1):86-97, 2016. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.802, + 0.484, + 0.858 + ], + "angle": 0, + "content": "[57] Armine Vardazaryan, Didier Mutter, Jacques Marescaux, and Nicolas Padoy. Weakly-supervised learning for tool localization in laparoscopic videos. In MICCAI-LABEL, pages 169–179. Springer, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.86, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[58] Rejin Varghese and M Sambath. Yolov8: A novel object detection algorithm with enhanced performance and robustness. In 2024 International Conference on Advances in Data" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.092, + 0.905, + 0.12 + ], + "angle": 0, + "content": "Engineering and Intelligent Computing Systems (ADICS), pages 1-6. IEEE, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.121, + 0.905, + 0.217 + ], + "angle": 0, + "content": "[59] Martin Wagner, Beat-Peter Müller-Stich, Anna Kisilenko, Duc Tran, Patrick Heger, Lars Mündermann, David M Lubotsky, Benjamin Müller, Tornike Davitashvili, Manuela Capek, et al. Comparative validation of machine learning algorithms for surgical workflow and skill analysis with the heichole benchmark. Medical Image Analysis, 86:102770, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.218, + 0.905, + 0.272 + ], + "angle": 0, + "content": "[60] Ao Wang, Hui Chen, Lihao Liu, Kai Chen, Zijia Lin, Jungong Han, et al. Yolov10: Real-time end-to-end object detection. Advances in Neural Information Processing Systems, 37:107984-108011, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.274, + 0.905, + 0.315 + ], + "angle": 0, + "content": "[61] Chien-Yao Wang and Hong-Yuan Mark Liao. Yolov9: Learning what you want to learn using programmable gradient information. 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.316, + 0.905, + 0.385 + ], + "angle": 0, + "content": "[62] Chien-Yao Wang, Alexey Bochkovskyi, and Hong-Yuan Mark Liao. Yolov7: Trainable bag-of-freebies sets new state-of-the-art for real-time object detectors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7464-7475, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.386, + 0.905, + 0.44 + ], + "angle": 0, + "content": "[63] Yu-Hsiang Wang, Jun-Wei Hsieh, Ping-Yang Chen, Ming-Ching Chang, Hung Hin So, and Xin Li. Smiletrack: Similarity learning for occlusion-aware multiple object tracking, 2023. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.441, + 0.905, + 0.496 + ], + "angle": 0, + "content": "[64] Nicolai Wojke, Alex Bewley, and Dietrich Paulus. Simple online and realtime tracking with a deep association metric. In 2017 IEEE international conference on image processing (ICIP), pages 3645-3649. IEEE, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.497, + 0.905, + 0.566 + ], + "angle": 0, + "content": "[65] Lisheng Xu, Haoran Zhang, Jiaole Wang, Ang Li, Shuang Song, Hongliang Ren, Lin Qi, Jason J Gu, and Max Q-H Meng. Information loss challenges in surgical navigation systems: From information fusion to ai-based approaches. Information Fusion, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.567, + 0.905, + 0.622 + ], + "angle": 0, + "content": "[66] Yifu Zhang, Chunyu Wang, Xinggang Wang, Wenjun Zeng, and Wenyu Liu. Fairmot: On the fairness of detection and re-identification in multiple object tracking. International Journal of Computer Vision, 129:3069–3087, 2021. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.622, + 0.905, + 0.705 + ], + "angle": 0, + "content": "[67] Yifu Zhang, Peize Sun, Yi Jiang, Dongdong Yu, Fucheng Weng, Zehuan Yuan, Ping Luo, Wenyu Liu, and Xinggang Wang. Bytetrack: Multi-object tracking by associating every detection box. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXII, pages 1-21. Springer, 2022. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.706, + 0.905, + 0.761 + ], + "angle": 0, + "content": "[68] Zijian Zhao, Sandrine Voros, Zhaorui Chen, and Xiaolin Cheng. Surgical tool tracking based on two cnns: from coarse to fine. The Journal of Engineering, 2019(14):467-472, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.761, + 0.905, + 0.789 + ], + "angle": 0, + "content": "[69] Xingyi Zhou, Dequan Wang, and Philipp Krahenbuhl. Objects as points. arXiv preprint arXiv:1904.07850, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.79, + 0.905, + 0.845 + ], + "angle": 0, + "content": "[70] Pengfei Zhu, Longyin Wen, Dawei Du, Xiao Bian, Heng Fan, Qinghua Hu, and Haibin Ling. Detection and tracking meet drones challenge. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(11):7380-7399, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[71] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. In International Conference on Learning Representations, 2021. 6" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8952" + } + ] +] \ No newline at end of file diff --git a/2025/CholecTrack20_ A Multi-Perspective Tracking Dataset for Surgical Tools/3671c43b-afb0-4272-8573-cad641013bba_origin.pdf b/2025/CholecTrack20_ A Multi-Perspective Tracking Dataset for Surgical Tools/3671c43b-afb0-4272-8573-cad641013bba_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..9c905575d2ae3d0f3eb5f1ea0a26c42c3a59eb3b --- /dev/null +++ b/2025/CholecTrack20_ A Multi-Perspective Tracking Dataset for Surgical Tools/3671c43b-afb0-4272-8573-cad641013bba_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ef6b4e9d776ef3770108607ced23121987a270a61b38f926c0c997c4b24b4447 +size 6457668 diff --git a/2025/CholecTrack20_ A Multi-Perspective Tracking Dataset for Surgical Tools/full.md b/2025/CholecTrack20_ A Multi-Perspective Tracking Dataset for Surgical Tools/full.md new file mode 100644 index 0000000000000000000000000000000000000000..de37a60b2d335b812c174878155b52bd23c2fdec --- /dev/null +++ b/2025/CholecTrack20_ A Multi-Perspective Tracking Dataset for Surgical Tools/full.md @@ -0,0 +1,396 @@ +# CholecTrack20: A Multi-Perspective Tracking Dataset for Surgical Tools + +Chinedu Innocent Nwoye $^{1,3}$ Kareem Elgohary $^{1}$ Anvita Srinivas $^{1}$ Fauzan Zaid $^{1}$ Joel L. Lavanchy $^{2}$ Nicolas Padoy $^{1,3}$ + +1University of Strasbourg, CNRS, INSERM, ICube, UMR7357, Strasbourg, France + +2University of Basel, University Digestive Health Care Center, Clarunis, Switzerland + +$^{3}$ IHU Strasbourg, France + +https://github.com/camma-public/cholectrack20 + +# Abstract + +Tool tracking in surgical videos is essential for advancing computer-assisted interventions, such as skill assessment, safety zone estimation, and human-machine collaboration. However, the lack of context-rich datasets limits AI applications in this field. Existing datasets rely on overly generic tracking formalizations that fail to capture surgical-specific dynamics, such as tools moving out of the camera's view or exiting the body. This results in less clinically relevant trajectories and a lack of flexibility for real-world surgical applications. Methods trained on these datasets often struggle with visual challenges such as smoke, reflection, and bleeding, further exposing the limitations of current approaches. We introduce CholecTrack20, a specialized dataset for multi-class, multi-tool tracking in surgical procedures. It redefines tracking formalization with three perspectives: (1) intraoperative, (2) intracorporeal, and (3) visibility, enabling adaptable and clinically meaningful tool trajectories. The dataset comprises 20 full-length surgical videos, annotated at 1 fps, yielding over 35K frames and 65K labeled tool instances. Annotations include spatial location, category, identity, operator, phase, and visual challenges. Benchmarking state-of-the-art methods on CholecTrack20 reveals significant performance gaps, with current approaches (< 45% HOTA) failing to meet the accuracy required for clinical translation. These findings motivate the need for advanced and intuitive tracking algorithms and establish CholecTrack20 as a foundation for developing robust AI-driven surgical assistance systems. + +# 1. Introduction + +The true impact of computer vision research lies in its practical applications, especially in critical fields like healthcare. Among these, surgery represents one of the most demanding domains, providing a definitive test for the capa + +bilities of vision technologies. Surgical data science, on its own, has significantly advanced interventional healthcare by leveraging data-driven techniques to provide critical decision support to medical professionals [34]. A key area of this advancement is the analysis of endoscopic video data, which offers real-time insights into surgical procedures, aids in skill assessment, and helps predict complications [9, 56, 65]. Accurate tracking of surgical tools is central to these analyses, as it guides temporal progression of procedural phases and correlates with surgical actions and management of adverse events [38, 59]. Despite progress in computer vision, research [42] shows that deep learning models pretrained on general datasets often struggle with surgical contexts due to complex scene dynamics, diverse tool types, and challenging visual conditions such as bleeding, smoke, and variable lighting [12, 20, 51]. This highlights the need for domain-specific datasets tailored to the unique requirements of surgical tool tracking. Obtaining medical and surgical data for research is challenging due to ethical and practical constraints, and annotating it requires expert knowledge. Current methods for tool tracking largely focus on Single Object Tracking (SOT) [68], Multi-Class Tracking (MCT) with one tool per class [36, 37], or Multi-Object Tracking (MOT) treating all tools as a single class [16, 46]. However, these approaches often miss the complexities of Multi-Class Multi-Object Tracking (MC-MOT) specific to surgical contexts, where tools interact dynamically and may move out of the camera's field of view or within the body cavity. + +Multi-perspective tracking addresses these challenges by defining tool trajectories across different viewpoints during surgical procedures. It includes three critical perspectives (illustrated in Fig. 1): (1) intraoperative - covering the entire procedure duration to monitor tool usage and assess surgical proficiency; (2) intracorporeal - focusing on tool tracks within the body to evaluate specific tasks and predict risks; and (3) visibility - tracking tools within the camera's + +![](images/fb7f4a50edb064bbba40a2063ab477c772e2c10e47cee16d12dc6ec670f06693.jpg) +Surgeon operators positioning and tool placement + +![](images/d55b6a8fed8b89cd2f3ab3d8e1468aab64f78aa9e614eb87498504c462869ee2.jpg) +Multi-perspective of surgical tool trajectory + +![](images/2a8b05911b3f6726b8b0b485f2fa6562255f4383d24838cdab50923c7503b2af.jpg) +Sample data and label overlaid + +![](images/a1e3fce7f8d791780bb9dae68750a87f66ed65884246ea358f566051640babe3.jpg) +Considered surgical tool categories + +![](images/2221ca3216cc5592037105c619f4d105fe230a2f696aed73c433fdab8bc762dc.jpg) +bipolar + +![](images/e6df6ba0977b91ab3652d9ec1c2bd85b594a667e9997ae40d4178698f54b260d.jpg) +hook + +![](images/2b3f98256588cb824ae067c4ca0be875d5ec9afe922f11a28fe12797af82e020.jpg) +scissors + +![](images/97bb9c4cd069f1d9e0b835a0f4d61815dc77c0feb12eeae99df7faacf88c87ee.jpg) +grasper + +![](images/9a6618930d7de982a75cd5bcd79885d5012884ae6665818bc848556faf5522aa.jpg) +Figure 1. Illustration of multi-perspective tracking in surgical domain and CholecTrack20 dataset labels for surgical tool tracking. + +![](images/66360be32b62168d2ffdb758a6d43526e9aa634684e13853376e1b7ed1ee49eb.jpg) + +![](images/f4e498405819c4c83448ab7bfcaa82de25ba95094f588a7ee9b3cf452174901f.jpg) +Visibility tracking FoV +1 +Operating room +$\left( 4\right)$ +$\therefore m - 1 \neq 0$ ; +$\therefore m - 1 \neq 0$ ; + +room + +$\frac{1 + u}{1} - \frac{u}{1} = \frac{\left( {1 + u}\right) u}{1} < \frac{u}{1} = u$ + +![](images/7b2f28c59d7f50c8b0b572afea6afa97a72cb33fe11d4e82a08413594a194c71.jpg) +Intracorporeal tracking Fov + OOCV + +![](images/bc78089af8921d384819590d80ba041c3235476bff83da1acc8f5c9aae9ca30f.jpg) +Sample surgical visual challenges + +![](images/16729a3eafacbdcea5c9786f5a15b60f9716df083e1561d869a55f7c2649cffc.jpg) +Blurring +Trocar view + +![](images/7b710929e27e75ae90105524806d27f6e76d4df1cb9dd6c16d065663c3c27177.jpg) + +![](images/c761f4abba5cfaad4418a4f9328434cd71bd448fb237fc9b9e805277c8174407.jpg) +Bleeding +Reflection + +![](images/1ab595c418f3f88042bde5c82b69227b8634f8f679f2ff8c963bde084fb8438a.jpg) + +![](images/c20fd387dde5b8625bc4945b9a75bf5f4e0c913ad1c84f7daa96fc04ca2cba7b.jpg) +Crowded +Smoke + +![](images/2d9c6b533680090b55241b6e7e461bfee4690e482a7c90c6f473c827d74b4723.jpg) +Occlusion + +![](images/87d9566a6f14a86753e0a91700e3ded82fe6eafb79da1a37960515ed70e4b06e.jpg) +Lens fouling + +![](images/4a07ecb3239a181d166aaa80984ab3a38fab92751fd597b1205296c22d919770.jpg) +Blood stain + +field of view to provide real-time feedback to surgeons. Existing surgical tracking datasets [18, 53] often lack these levels of granularity and adaptability needed for comprehensive tool modeling. They generally follow generic tracking formalizations and struggle to capture the intricacies of surgical tool interactions, particularly when tools are replaced or move beyond camera visibility [6, 28, 53]. + +To address this gap, we introduce CholecTrack20, a novel dataset, based on laparoscopic cholecystectomy surgery, designed for multi-class multi-tool tracking from intraoperative, intracorporeal, and visibility perspectives. Derived from raw laparoscopic videos [39, 56], it includes detailed annotations such as spatial coordinates, tool categories, track identities (IDs), visual challenges, phase labels, and other scene attributes. This dataset enhances benchmarking resources for computer vision research and supports the development of AI models tailored to surgical tool tracking, phase recognition, and surgeon performance assessment. This paper provides a comprehensive overview of the data acquisition and annotation methodology, detailed data analysis, and technical validation. In addition, we conduct extensive benchmark experiments using state-of-the-art deep learning methods for object detection and tracking, evaluating performance across various surgical phases and visual challenges, discussing insightful findings. + +In summary, the main contributions are: + +1. Introduction of CholecTrack20, a pioneering dataset for multi-perspective tracking with extensive annotations. +2. Extensive experimental analysis validating the dataset's effectiveness for surgical tool detection and tracking. + +3. Insights from model performance analysis under diverse visual challenges, highlighting the utility of each tracking perspective for AI-driven surgical solutions. + +# 2. Related Works + +Object detection and tracking. Advances in object detection and tracking have been driven by datasets like COCO [29], KITTI [20], MOTChallenge [12], VisDrone [70], DanceTrack [51], and TAO [11], enabling progress in Single Object Tracking (SOT) [10], Multi-Object Tracking (MOT) [64, 66], and Multi-Class Multi-Object Tracking (MCMOT) [13, 26]. Applying these techniques to surgical tool tracking poses challenges especially in the phase of bleeding, smoke, rapid movements, and variable lighting. Traditional tracking, centered on visibility, struggles when tools leave the camera's view or are replaced during surgery. + +Surgical tool tracking and configurations. While electromagnetic and optical tracking methods [8, 17] have been explored in surgical domain, image-based approaches [53] better align with surgeons' view but face issues like identity fragmentation, identity switch, and low tracking accuracy [6, 16, 28, 36, 37, 46] especially with mid-procedure tool replacements and tools exiting/re-entering the field of view or body cavity. To fully capture tool usage complexity, it is crucial to consider multiple perspectives on tool trajectories. Existing datasets [16, 41] provide insights but lack comprehensive coverage of these scenarios, highlighting the need for detailed datasets addressing these complexities. + +# 3. Methodology + +CholecTrack20 is a detailed dataset for surgical tool tracking in laparoscopic cholecystectomy, offering binary and spatial annotations for tools, including identity, category, bounding box location, motion, operator, phase, activity, usage conditions, and visual challenges, essential for training and benchmarking surgical AI tool tracking algorithms. + +# 3.1. Data Acquisition and Collection + +Data source. The raw videos are sourced from the publicly available Cholec80 [56] and CholecT50 [39] datasets, with appropriate permissions and adherence to license terms. Recorded at the University Hospital of Strasbourg with the aid of laparoscopic cameras, these videos document laparoscopic cholecystectomy surgeries. + +Video selection. Long videos were systematically chosen to represent surgical complexities and tool variability, capturing all key phases of laparoscopic cholecystectomy. High-quality videos with clear visuals were selected, subsampled from 25 FPS to 1 FPS for annotation, maintaining temporal consistency and clarity. + +Sensitive data handling. To protect patient and staff identities, out-of-body frames potentially revealing sensitive information, such as the identity of patients, clinical staff, or the operating room, were checked and anonymized following established techniques [25]. This ensures compliance with ethical standards and privacy regulations. + +# 3.2. Track Formalization + +Given a video dataset $D = \{S_{1}, S_{2}, \ldots, S_{n}\}$ of laparoscopic surgeries, each sequence $S_{i}$ includes frames annotated with bounding boxes $B = [B_{1}, B_{2}, \ldots, B_{M}]$ for tool locations and classes $C = \{C_{1}, C_{2}, \ldots, C_{N}\}$ . Tools, manipulated by operators linked to trocar ports $P = [P_{1}, P_{2}, \ldots, P_{M}]$ , are assigned unique track identities (ID) through time. The ID reassignment is guided by visual cues and clinical knowledge. Visually, it is based on class $c \in C$ and location $b \in B$ . Contextually, it considers the role and hand position of the surgeon operator linked to a unique trocar port $p \in P$ . Tool tracking solves the association matrix $A(t)$ , where $A_{i,j} = 1$ indicates the $i$ -th tool in frame $t$ is associated with the $j$ -th tool in frame $t + 1$ , and $A_{i,j} = 0$ otherwise. The aim is to obtain tool trajectories $T = \{T_{1}, T_{2}, \ldots, T_{K}\}$ , each uniquely identified by ID, taking into consideration the visibility, intracorporeal, and intraoperative use. + +# 3.3. Multi-Perspective (MP) Trajectory + +Defining tool trajectories in surgical procedures necessitates a unique approach, given the variability in tracking across different perspectives, formalized as follows: + +![](images/cdd3ad09fa90314ec31990633fc7bc5d027b5ccaf34e715ad1c0f1c6a811af47.jpg) +Figure 2. Multi-perspective trajectories of surgical tool. + +1. Intraoperative trajectory: This lifelong tracking starts with a tool's first appearance and ends at its last in a patient's body during a procedure. It requires re-identification post-occlusion, out-of-camera view, or reinsertion. This approach is vital for applications such as tool usage monitoring [2], inventory management [35], surgeon training [27], skill assessment [16], and tool usage pattern analysis [52]. + +2. Intracorporeal trajectory: Unique tracks begin when a tool enters the body and end when it exits through a trocar port, even if off-camera. If a tool exits outside the camera's view, the exit is inferred if another tool enters through the same trocar or the initial tool releases its grasp. This is essential for understanding surgical workflow, as some actions occur outside the camera's focus, like graspers holding tissue out of view to facilitate other tools' actions [27]. Intracorporeal tracking supports a range of AI tasks including action evaluation [55], skill assessment [16, 27], tool usage optimization [44], and surgical risk estimation [52]. + +3. Visibility trajectory: Tracking starts with a tool's first appearance within the camera view and ends when it leaves the view. It requires re-identification (re-ID) after occlusions or brief periods of disappearance within a two-second tolerance. This method is useful for assessing surgeon actions [27] and skill training [14, 24], providing go-no-go decision support, and measuring economy of motion [48]. + +Some existing studies [16, 37] follow the intraoperative tracking format, others [24, 46] employs visibility trajectory. The intracorporeal trajectory, being the most complex to annotate, are not well-represented in the literature. Our dataset is the first to provide fine-grained labels for all the three (Fig. 2), enabling statistical analyses such as tool usage counts, events counts, abdominal entries/ exits, tool idleness, out-of-camera view occurrences, and mean tracklets per perspective. + +# 3.4. Data Annotation + +Annotators and tools. Bounding boxes were annotated by four researchers skilled in surgical workflow analysis, supplemented by prior study data [37, 40, 57]. Annotation tools used include Anonymized System, and a custom Python tool for visualization, merging, and validation. A pre-designed guide, refined by surgical experts, described + +![](images/2ef49139e401ce7df0b6679663e28c05e35ea2784c443e6a2180d25bc0fe2a37.jpg) + +![](images/42a6ae560d1cd6bb2c532e6e510ca4a9867a782033eba65927042769b612b143.jpg) + +![](images/991eb9e1aca2d3fe836d87c367a7b876a7681aa7b15335098f5daa4ea7569b30.jpg) +Figure 3. Examples of images from CholecTrack20 tracking dataset with the labels overlaid on the raw images. + +![](images/9697a46f490c604d565d8cb985f205b53829a40ad466e3b5f57f9f1f793cd190.jpg) + +the labels and provided image guidance as needed. + +Label types and categories. Tool spatial positions were annotated using bounding box coordinates, while other labels were represented by class identities. Seven predominant tool categories were defined: (1) cold grasper, (2) bipolar grasper, (3) monopolar hook, (4) monopolar scissors, (5) clipper or clip applicer, (6) irrigator or suction device, and (7) specimen bag. Four operator categories were defined: (1) main surgeon left hand MSLH, (2) main surgeon right hand MSRH, (3) assistant surgeon right hand ASRH, and (4) null operator NULL. The assistant surgeon left hand ASLH holding the endoscopic camera is unreported. A total of eight visual challenges were noted: (1) blurring, (2) bleeding, (3) camera lens fouling, (4) crowded scene, (5) occlusion, (6) smoke, (7) specular light reflection, and (8) trocar view or under-coverage. The seven commonest surgical phases were annotated: (1) preparation, (2) calot triangle dissection, (3) gallbladder dissection, (4) clipping & cutting, (5) gallbladder packaging, (6) cleaning and coagulation, and (7) gallbladder extraction. + +Annotation process. Annotations involved drawing bounding boxes $[x, y, w, h]$ over tool tips tagged with tool class $c \in C$ and operator class $p \in P$ , following trocar ports for accurate surgeon identification. Surgical details such as phase, out-of-view statuses (camera/abdomen), tool entry/exit, and visual challenge attributes were annotated to aid accurate track assignment. Tool-tissue interaction labels from CholecT50 [39] provides additional help in perpetuating track identities. Annotations were reviewed at 25 FPS in uncertain cases, and underwent rigorous quality control, ensuring high-quality labels over two years. Fig. 3 presents samples of images from the CholecTrack20 dataset alongside their respective annotations illustrating the meticulous labeling system employed to ensure a rich dataset for detailed surgical tool tracking analysis. + +![](images/727d093663a8add9db71df69c8ad56964edc3a0cf9236aecb28ebb6730f0c256.jpg) +Figure 4. Dataset statistics on the distributions of (a) surgical scene visual challenges across data splits (b) track labels across perspectives, averaged across videos. Track length in seconds. + +# 3.5. Quality Assurance + +Label agreement. Two label agreement metrics validate dataset quality: Jaccard Index for spatial overlap of bounding boxes and Cohen's Kappa Statistic for category labels. The findings of our three validation forms include: + +1. Intra-rater agreement involves self-correction. We observe a Jaccard Index of $99.4\%$ and Cohen's Kappa Scores of $94.6\%$ for tools and $94.0\%$ for operators. +2. Inter-rater agreement is evaluated on 20 random samples across raters. The Jaccard Index is $91.8\%$ , while tool class labels achieve $95.2\%$ and operator labels $92.7\%$ Cohen's Kappa. Minor differences reflect high-quality annotations. +3. Label mediation uses a board-certified surgeon for ambiguities, particularly in operator labels. Out of 758 uncertain samples, 133 needed correction post-mediation. + +# 3.6. Data Statistics + +Quantitative overview. The dataset includes 20 surgical videos totaling over 14 hours, averaging 42 minutes per surgery. As detailed in Tab. 1, annotations cover 35,000 frames at 1 FPS ( $\sim$ 875,000 at 25 FPS), with 65,000 bounding box labels, averaging two tools per frame. Fig. 4(a) summarizes track configurations across perspectives, showing tool usage by type and trajectory. For example, graspers average 27 minutes 39 seconds inside the body and are re-inserted 8.4 times, while irrigators are used for about 1 minute 30 seconds per surgery. The visibility perspective contributes approximately 2,000 trajectories. Fig. 4(b) outlines tool attributes and visual challenges in the dataset, with occlusion as most prevalent occurring up to 23,000 seconds, bleeding totals around 18,700 seconds. This analysis provides insights into the varied visual challenges encountered in laparoscopic cholecystectomy. Evaluating tracking methods across these challenges will reveal their strengths and weaknesses. Statistical details across dataset splits are presented in Fig. 4(b). + +Table 1. Dataset comparison showing the scope, statistics, and attributes. Dataset marked $\ddagger$ are not full-length videos. + +
DatasetTaskTrack PerspectivesStatisticsAttributes
VisibilityIntraoperativeIntracorporealNo. of VideosTotal Duration (s)Frame CountTool BoxesNo. of TrajectoriesSurgoen OperatorVisual ChallengeSurgical Phase
ATLAS Dione [47]‡Detection✓XX99-22467--XXX
Cholec80-locations [49]Detection✓XX-401140116471-XXX
Bouget et.al [5]Detection✓XX14-24763819-XXX
m2cai 16-tool-locations [21]Detection✓XX--25323038-XXX
EndoVis'15 [15]‡Tracking✓XX1654013500--XXX
Fathollahi et el [16]‡TrackingX✓X1527002700--XXX
RMIT [53]Detection & Tracking✓XX415001171XXX
CholecTrack20 (Ours)Detection & Tracking✓✓✓205058135000652002,624✓✓✓
+ +Dataset comparison. Existing publicly available datasets primarily focus on single perspective trajectory [16, 46]. In contrast, our dataset introduces a novel approach by annotation different perspectives, as presented in Tab. 1, including visual challenges, phase details, activity labels, etc. + +Dataset split. The dataset is split at the procedure level into non-overlapping training, validation, and testing sets in a 5:1:4 ratio, preventing data leakage. Video distribution is balanced by procedure duration to ensure similar complexity and difficulty levels across all splits. + +# 3.7. Data Analysis + +We conduct a comprehensive analysis of the dataset to gain insights into label alignments and feature similarities, revealing correlations across labels. This would guide data preprocessing and feature selection when using the data. Our analysis as illustrated in Fig. 5, Fig. 6, and Fig. 7 encompasses four distinct dimensions and discussed further. + +Tracking vs. surgical tool type correlation. This analysis explores the relationship between different tool types and their tracks, providing insights into unique patterns associated with each tool during surgery. Fig. 5(a) illustrates center point locations of tools, color-coded by category over time in three videos. This shows that while some trajectories appear separable, they are mostly densely interwoven, suggesting the need for advanced modeling in tool tracking. + +Tracking vs. surgeon tool operator correlation. We analyzed the alignment between tool operator identities and tool locations, revealing dynamic interactions between surgeons and instruments. Fig. 5(b) shows a strong correlation between operators and their tools, attributed to the distinct positioning of trocars. This underscores the value of operator information in accurate track label generation. + +Tracking vs. surgical phase segmentation. This analysis delves into tracking tools across surgical phases, uncovering how tool utilization varies with procedural stages. Fig. 6 shows that complex phases, such as calot triangle and gallbladder dissection, exhibit densely packed trajectories due to prolonged duration and frequent tool manipulation. Simpler phases, like preparation and clipping, feature fewer trajectories, facilitating modeling. These insights inform deep + +![](images/2704ea2eedc13fe4f82ba0dfcccc32e912588f04c97168c2a9529f1b83b00a1f.jpg) + +![](images/e30acdeb9ffaf6458871f74f2c78a5a43eff1b0046104beac080fc5453784d20.jpg) + +![](images/df30a378ce2b2f618497745f07be265f0d2b8362c6a0be190fabad607ad066bb.jpg) + +![](images/fd20cb4ebc8019ab066faa0e1295b19af1b36f0fb629cf34a62b3ed5c48672ea.jpg) +Figure 5. 3D visualization of label alignments showing the tool position over track time. The coloring is for grouping features according to: (a) tool classes and (b) tool operators. + +![](images/eec379849419c74f4da22e6f8c89cdb6018b0fefe8e20085b144472b9aebb9e6.jpg) + +![](images/42c63b944bf9a009018462133333d698939af8d13cad7e78bee608c7b0e93aa3.jpg) + +![](images/3d86bcf08c4bd17bf3d61267a52a4078b20e098d3b11afc1020733986c438788.jpg) + +![](images/39700c595b61332c0e958a66956dfafc5292bc11bb732eb967609c2eeabe7b51.jpg) + +![](images/1fa3ae87e774469ab1b5874eff703d188649e6145e2a2c66d6d5efdb4f5a2fe9.jpg) + +![](images/165faeef204e8994e051ad323cff3aa62bc454ac87e905eda49aa79ca01436ed.jpg) +Figure 6. 3D visualization of tracking across different surgical phases for some randomly selected videos. + +![](images/fc8efa217cf617101faa9213f4fe7f70bdf65fba0f9c4c0d439170fed131677e.jpg) + +![](images/82b7c4f8969ef70591d9fedc36a2984035c139e7c53df56f99af599f5b7306ed.jpg) + +![](images/d05cce21a7f9088eb75d4494392e2ac5866343b6c06e829dd50182d8a005d3b4.jpg) + +![](images/4c67c568558578949091197344d85ce42020c8a7418f8c5edeadb1cbf6a61aa2.jpg) + +![](images/a135a11defd19d4258a115911fe7da6820e18c2864f13af8a5ef4d0c931e2ae0.jpg) + +![](images/27506cc6dc91c6f54b713a43bc59c9764373f1ea8d3ba19cdc98a2d18fa2b2ce.jpg) +Figure 7. Trajectories of selected tools with EMA over time. Plotted positions are computed as weighted combinations of the center coordinate of the bounding boxes, scaled by image size. + +![](images/ae6516bec1d140b08e62fc15fed07bc071490342df98d577929bce0c3d209000.jpg) + +![](images/6ee3105fd554f60069040c5b19ed3184e560172ae9323787f64439938ae422b1.jpg) + +![](images/00b52bd8d89ca5fca38d70292519e9e3ef5ced20bcbb9d81a2838e3419fbf648.jpg) + +Table 2. Benchmark Results of SOTA Object Detectors on Surgical Tool Detection Dataset. + +
Detector +ModelDetection AP accross 3 thresholdsDetection AP per category. (% AP @ Θ = 0.5)Detection AP across surgical visual challengesSpeed
AP0.5↑AP0.75↑AP0.5:0.95↑GrasperBipolarHookScissorsClipperIrrigatorBagBleedingBlurSmokeCrowdedOccludedReflectionFoul LensTrocarFPS↑
Faster-RCNN [43]56.038.134.653.565.080.160.970.126.831.857.941.054.543.555.046.941.235.77.6
Cascade-RCNN [7]51.739.034.752.058.979.745.744.923.717.953.939.048.139.546.429.133.733.77.0
CenterNet [69]53.039.535.060.261.486.456.368.025.810.258.042.150.236.751.746.035.830.833.8
FCOS [54]43.531.528.151.244.374.749.254.221.97.247.840.651.515.140.842.729.717.67.7
SSD [30]61.937.836.175.262.291.663.472.922.540.864.549.358.357.562.453.947.742.630.9
PAA [23]64.544.941.169.679.089.268.774.237.628.967.155.665.055.064.656.051.247.58.5
Def-DETR [71]58.442.038.360.666.583.861.972.039.923.862.442.758.637.157.443.941.547.410.2
Swin-T [31]62.344.340.263.364.883.080.277.238.026.863.553.862.835.361.166.255.245.79.8
YOLOX [19]64.748.944.269.672.289.475.479.137.327.168.255.666.045.964.252.558.143.123.6
YOLOv7 [62]80.662.056.190.586.496.082.389.349.166.280.261.280.179.582.165.671.266.720.6
YOLOv8 [58]79.162.455.687.984.596.280.087.248.465.077.158.374.476.280.470.357.462.929.0
YOLOv9 [61]80.262.656.588.587.696.079.387.150.167.778.154.078.278.681.165.363.463.123.7
YOLOv10 [60]80.162.155.887.686.696.081.989.053.861.377.861.978.777.581.266.759.365.428.6
+ +learning model assessment and AI model development for surgical tool tracking. + +Tracking variance. Using Exponential Moving Average (EMA), we analyze tool tracking data, as shown in Fig. 7. By overlaying EMAs on tool trajectories, we visualize variance between actual trajectories and modeled tracks. This approach highlights high-variance frames, serving as challenging cases for benchmarking model robustness and accuracy. Through this analysis, we provide a valuable resource for researchers, emphasizing the need to focus on complex scenarios for a rigorous assessment of model capabilities. + +# 4. Benchmark and Experiments + +# 4.1. Tool Detection + +Models. Owing that tool detection is a fundamental part of tool tracking, we showcase the usability of the CholecTrack20 dataset for this task, by benchmarking several object detectors representing diverse methodologies. Faster-RCNN [43] and Cascade-RCNN [7] are anchor-based models, with Cascade-RCNN employing a multi-stage approach to refine detection accuracy. CenterNet [69] and FCOS [54] are anchor-free models, utilizing center points and direct bounding box regression for efficient detection. SSD [30] provides real-time performance with its multi-scale approach. Deformable-DETR [71] applies a transformer-based method for flexible feature processing, while Swin-T [31] uses hierarchical transformers with shifted windows. The YOLO models [19, 58, 60–62] feature advanced multiscale strategies for high accuracy and speed. + +Evaluation metrics. We evaluate tool detection using COCO standard average precision (AP) metrics (pycocotools, not ultralytics) across several thresholds, categories, and visual challenges. We also report model inference speed in frames per seconds (FPS) on a single NVIDIA GTX 1080 Ti (10 GB) GPU. + +Overall tool detection results. The detection performance of the models is summarized in Tab. 2. YOLOv7 [62] achieves the highest Average Precision (AP) of $56.1\%$ , surpassing other models. YOLOX [19] follows with an AP + +of $44.2\%$ . Deformable-DETR [71] and Swin-T [31] show competitive results with APs of $38.3\%$ and $40.2\%$ , respectively. The anchor-free models such as CenterNet [69] and FCOS [54] demonstrate robust performance, with CenterNet achieving an AP of $35.0\%$ and FCOS $28.1\%$ . In comparison, Faster-RCNN [43] and Cascade-RCNN [7] deliver APs of $34.6\%$ and $34.7\%$ , respectively, showcasing their efficacy with anchor-based approaches. At IoU thresholds of 0.5 and 0.75, YOLOv7 leads with scores of $80.6\%$ and $62.0\%$ , respectively. In terms of inference speed, YOLO networks excel with real-time capacities exceeding 20 FPS, with CenterNet achieving the highest speed at 33.8 FPS. + +Class-wise detection results. Analyzing tool detection results per category (Tab. 2), YOLOv7 emerges as the top performer, dominating in all the 7 categories, achieving above $90\%$ accuracy in 2 tool categories and above $80\%$ in 5. Notably, the hook exhibits the highest tendency among tools, with AP scores ranging from $74.7\%$ to $96.0\%$ across all models. Conversely, irrigator and specimen bag pose challenges, likely due to unclear tool tip boundaries and the bag's deformable nature, respectively. Grasper, bipolar, scissors, and clipper show high detection rates. + +Detection under visual challenges. Tab. 2 presents the performance of the benchmark object detection models across different surgical visual challenges. Notably, YOLOv7 achieves the highest detection accuracy across most challenges, particularly excelling in scenarios involving bleeding, smoke, and crowded scenes. Conversely, detecting tools in blurred scenes, near trocars, and specular light reflection pose significant challenges for all models, with lower detection rates observed across the board. + +# 4.2. Tool Tracking + +Models. We train and evaluate several state-of-the-art multi-object tracking (MOT) methods on the CholecTrack20 dataset, focusing on their ability to track surgical tools. OCSORT [33] and TransTrack [50] employ sophisticated tracking-by-detection frameworks, with TransTrack utilizing transformers to improve feature association. ByteTrack [67], Bot-SORT [1], and SMILETrack [63] use ad + +Table 3. Benchmark Multi-Perspective Multi-Tool Tracking Results @ 25 FPS. + +
ModelHOTA MetricsCLEAR MetricsIdentity MetricsCount MetricsSpeed
TrackerHOTA↑DetA↑LocA↑AssA↑MOTA↑MOTP↑MT↑PT↓ML↓IDF1↑IDSW↓Frag↓#Dets#IDsFPS↑
Intraoperative Trajectory (Groundtruth counts: #Dets = 29994, #IDs = 70)
OCSORT [33]14.652.786.74.149.285.02432149.52921273121936333610.2
FairMOT [66]5.825.875.91.35.073.9324434.34227192415252445614.2
TransTrack [50]7.431.584.41.74.282.9936254.2475718992164040796.7
ByteTrack [67]15.870.685.73.667.084.0541229.54648242928440538316.4
Bot-SORT [1]17.470.785.44.469.683.75811110.2390723762930245018.7
SMILETrack [63]15.971.085.53.766.483.8551329.24968236928821576111.2
Intracorporeal Trajectory (Groundtruth counts: #Dets = 29994, #IDs = 247)
OCSORT [33]23.751.486.511.047.184.8115874518.12953279621797352610.2
FairMOT [66]7.519.776.12.95.474.019601686.02890149611287396214.2
TransTrack [50]13.131.584.45.54.682.98079888.7464817912164040796.7
ByteTrack [67]24.770.685.78.767.484.0176482316.94515229028440538316.4
Bot-SORT [1]27.070.785.410.470.083.7188382118.9377122382930045018.7
SMILETrack [63]24.966.785.58.966.783.8186392216.94868223228820577911.2
Visibility Trajectory (Groundtruth counts: #Dets = 29994, #IDs = 916)
SORT [4]17.439.585.27.821.483.313939937813.46619213816595884419.5
OCSORT [33]37.052.686.526.250.284.830037124535.92317226022197358710.2
FairMOT [66]15.325.075.89.57.173.75821864014.43140157415338487514.2
TransTrack [50]19.231.684.411.85.882.922428041216.1427314032164040796.7
ByteTrack [67]41.570.785.724.869.384.059121710836.83930170428440538316.4
Bot-SORT [1]44.770.885.528.772.083.76381849441.4318316382930045058.7
SMILETrack [63]41.371.085.624.468.983.861919210536.54227164128821575211.2
+ +vanced tracking algorithms, with Bot-SORT and SMILE-Track incorporating extra modules for enhanced robustness. + +Evaluation metrics. We assess benchmark models on variety of tracking metrics: higher-order tracking accuracy (HOTA) [32], CLEAR MOT metrics [3], identity metrics [45], counting metrics, and tracking speed. A pull request is made to the standard TrackEval [22] library integrating CholecTrack20 benchmark with all its exhaustive performance evaluation protocols recommended by this study. + +Multi-object tracking results. The performance of the evaluated tracking methods is summarized in Tab. 3. Models such as FairMOT and TransTrack show the lowest HOTA score of $5.8\%$ and $7.4\%$ respectively, highlighting challenges in maintaining tool identities over time. ByteTrack, Bot-SORT, and SMILETrack achieve higher HOTA scores, ranging from $15.7\%$ to $17.4\%$ , but still face difficulties in tool re-ID due to similarities among tools. + +Despite these advancements, there remains room for enhancement in identity association and tracking precision. The results also include metrics on the detection counts, unique identities assigned, and tracking speed, providing a comprehensive view of each method's performance. + +Multi-perspective tracking results. Looking at the different trajectory perspectives, Tab. 3 shows that visibility tracking is the easiest with most of the existing models showcasing their strengths. This is expected because deep learning models mostly rely on visual cues, which are captured by camera in the visibility track scenario. Here, Bot-SORT record a landslide top performance scores of $44.7\%$ + +HOTA, $72.0\%$ MOTA, and $41.4\%$ IDF1. The intracorporeal tracking is the most challenging since the major factors marking the entry and exit of the tools from the body are not readily visible. A maximum of $27.0\%$ HOTA suggest a decline on the leading Bot-SORT. New methods could leverage rich fine-grained history to estimate the out-of-view and out-of-body status of the tools for improve re-ID. The intraoperative trajectory comes in the middle in terms of difficulty. While it may be challenging to ascertain the persistence of a trajectory after re-insertion, the class features are also helpful especially for tools of different categories. Again the Bot-SORT, leverage camera compensation details, shows a better tendency of estimating the persistent identity of different tools of the same class with a $+1.5\%$ HOTA higher than similarity and appearance features. + +Multi-class tracking results. In Fig. 8, we analyze tracking performance by tool class and observe that the grasper, despite having the most instances and being the most frequently used tool in the dataset, achieves the highest tracking accuracy across perspectives. Class-agnostic results reveal medium tracking accuracy for other commonly used tools like the bipolar, hook, and clipper, while rarely used tools (e.g., scissors, irrigator) have lower scores. Specimen bag tracking is affected by shape deformation, contents, states (open/closed, empty/filled), tool interactions, and fluid stains. Tracking surgical tools beyond the visual perspective remains challenging for all models tested. + +Tracking results under visual challenges. We evaluate tool tracking performance across various surgical condi + +![](images/6c0ebe6b0fbb1a8b8a189ff7f167a91470dbae44ff502f32292b612acdc1b78f.jpg) +Figure 8. Results of multi-perspective tracking across seven surgical tool categories in the CholecTrack20 dataset. + +![](images/0bd0e055ac20de3fea600d3d4020d0c43922dd460bff00bd65b5230a9fa8769a.jpg) +Figure 9. Results of multi-perspective tracking across eight surgical visual challenges in the CholecTrack20 dataset. + +![](images/9c4a38df3ba2bf363b6978e0a35c9a44d5ee5beb1a89866fe74e773c32829407.jpg) +Figure 10. Results of multi-perspective tracking across seven surgical phases in the CholecTrack20 dataset. + +tions using HOTA metrics in Fig. 9, providing insights into model interactions with complex surgical environments. The model performs well under blurring, reflections, and limited camera coverage, likely due to effective data augmentation. However, lens fouling, smoke, and occlusion present significant challenges, reducing accuracy. + +Tracking results by surgical phase. Fig. 10 shows performance across seven surgical phases, with the clipping and + +cutting phases proving easiest to track due to limited activities and a linear progression. Preparation phase shows similar performance. Phases like Calot triangle dissection and gallbladder dissection exhibit comparable tracking results, while gallbladder packaging shows the most consistent tracking across perspectives. Overall, OC-SORT struggles the most, while Bot-SORT achieves the best result. + +# 4.3. Limitations and Gaps to Address + +The SOTA tracking methods trained on CholecTrack20 reveal substantial limitations, with performance under $45\%$ HOTA, which is insufficient for clinical translation. These models struggle with various visual challenges, such as smoke, bleeding, and specular light reflection, affecting detection and re-ID. Since location and appearance features alone are inadequate, especially for tools with similar appearances, this highlights the need to move beyond current cues and innovate more intuitive, context-aware methods for re-ID. CholecTrack20 serves as a critical foundation for exploring this direction, offering a dataset rich in diverse tracking perspectives and challenges, essential for developing more robust and clinically viable tracking solutions. + +# 5. Conclusion + +In this work, we presented the CholecTrack20 dataset, a novel resource designed to advance the state-of-the-art in surgical tool tracking within computer vision. CholecTrack20 addresses a critical gap by providing comprehensive annotations and diverse tracking scenarios and tasks across various surgical phases and visual challenges. Key innovations of CholecTrack20 include multi-perspective tracking, which defines the start and termination of a tool track differently based on visibility, intracorporeal, or intraoperative contexts. It also features detailed annotations of surgical visual challenges and precise surgical phase segmentation. Our extensive benchmark experiments demonstrate the dataset's effectiveness in developing models for tool detection and multi-object tracking across these three distinct trajectory perspectives. Through evaluating several deep learning methodologies on the CholecTrack20 dataset, we gain insights into their strengths and weaknesses in handling multiple viewpoints or tracking perspectives, across surgical phases and diverse surgical scene visual challenges. + +By introducing CholecTrack20 to the computer vision research community, we aim to stimulate new research directions and foster collaborations between the computer vision and surgical communities. This dataset serves as a benchmark for evaluating state-of-the-art algorithms and promotes the development of robust and reliable surgical assistance systems. We anticipate that CholecTrack20 will inspire innovative approaches and contribute significantly to advancements in surgical tool tracking and related fields. The dataset is released under the CC BY-NC license. + +# Acknowledgments + +This work was supported by French state funds managed within the Plan Investissements d'Avenir by the ANR under references: National AI Chair AI4ORSafety [ANR-20-CHIA-0029-01], DeepSurg [ANR-16-CE33-0009], IHU Strasbourg [ANR-10-IAHU-02] and by BPI France under references: project CONDOR, project 5G-OR. Joel L. Lavanchy received funding by the Swiss National Science Foundation (P500PM_206724, P5R5PM_217663). This work was granted access to the servers/HPC resources managed by CAMMA, IHU Strasbourg, Unistra Mesocentre, and GENCI-IDRIS [Grant 2021-AD011011638R3, 2021-AD011011638R4]. + +# References + +[1] Nir Aharon, Roy Orfaig, and Ben-Zion Bobrovsky. Botsort: Robust associations multi-pedestrian tracking. arXiv preprint arXiv:2206.14651, 2022. 6, 7 +[2] Hassan Al Hajj, Mathieu Lamard, Pierre-Henri Conze, Béatrice Cochener, and Gwenolé Quellec. Monitoring tool usage in surgery videos using boosted convolutional and recurrent neural networks. Medical image analysis, 47:203-218, 2018. 3 +[3] Keni Bernstein and Rainer Stiefelhagen. Evaluating multiple object tracking performance: the clear mot metrics. EURASIP Journal on Image and Video Processing, 2008:1-10, 2008. 7 +[4] Alex Bewley, Zongyuan Ge, Lionel Ott, Fabio Ramos, and Ben Upcroft. Simple online and realtime tracking. In 2016 IEEE international conference on image processing (ICIP), pages 3464-3468. IEEE, 2016. 7 +[5] David Bouget, Rodrigo Benenson, Mohamed Omran, Laurent Riffaud, Bernt Schiele, and Pierre Jannin. Detecting surgical tools by modelling local appearance and global shape. IEEE transactions on medical imaging, 34(12):2603-2617, 2015. 5 +[6] David Bouget, Max Allan, Danail Stoyanov, and Pierre Jannin. Vision-based and marker-less surgical tool detection and tracking: a review of the literature. Medical image analysis, 35:633-654, 2017. 2 +[7] Zhaowei Cai and Nuno Vasconcelos. Cascade r-cnn: High quality object detection and instance segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence, page 1-1, 2019. 6 +[8] Magdalena K Chmarra, CA Grimbergen, and J Dankelman. Systems for tracking minimally invasive surgical instruments. Minimally Invasive Therapy & Allied Technologies, 16(6):328-340, 2007. 2 +[9] Tobias Czempiel, Magdalini Paschali, Matthias Keicher, Walter Simson, Hubertus Feussner, Seong Tae Kim, and Nassir Navab. Tecn: Surgical phase recognition with multistage temporal convolutional networks. In Medical Image Computing and Computer Assisted Intervention MICCAI, pages 343-352. Springer, 2020. 1 +[10] Martin Danelljan, Goutam Bhat, Fahad Shahbaz Khan, and Michael Felsberg. Eco: Efficient convolution operators + +for tracking. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6638-6646, 2017. 2 +[11] Achal Dave, Tarasha Khurana, Pavel Tokmakov, Cordelia Schmid, and Deva Ramanan. Tao: A large-scale benchmark for tracking any object. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part V 16, pages 436–454. Springer, 2020. 2 +[12] Patrick Dendorfer, Hamid Rezatofighi, Anton Milan, Javen Shi, Daniel Cremers, Ian Reid, Stefan Roth, Konrad Schindler, and Laura Leal-Taixe. Mot20: A benchmark for multi object tracking in crowded scenes. arXiv preprint arXiv:2003.09003, 2020. 1, 2 +[13] Yunhao Du, Junfeng Wan, Yanyun Zhao, Binyu Zhang, Zhihang Tong, and Junhao Dong. Giaotracker: A comprehensive framework for mcmot with global information and optimizing strategies in visdrone 2021. In Proceedings of the IEEE/CVF International conference on computer vision, pages 2809-2819, 2021. 2 +[14] Ariel Kate Dubin, Danielle Julian, Alyssa Tanaka, Patricia Mattingly, and Roger Smith. A model for predicting the gears score from virtual reality surgical simulator metrics. Surgical endoscopy, 32:3576-3581, 2018. 3 +[15] EndoVis Sub-Instrument Challenge. Endovis sub-instrument grand challenge, 2024. Accessed: 2021-11-12. 5 +[16] Mona Fathollahi, Mohammad Hasan Sarhan, Ramon Pena, Lela DiMonte, Anshu Gupta, Aishani Ataliwala, and Jocelyn Barker. Video-based surgical skills assessment using long term tool tracking. In Medical Image Computing and Computer Assisted Intervention-MICCAI 2022: 25th International Conference, Singapore, September 18-22, 2022, Proceedings, Part VII, pages 541-550. Springer, 2022. 1, 2, 3, 5 +[17] Marvin P Fried, Jonathan Kleefield, Harsha Gopal, Edward Reardon, Bryan T Ho, and Frederick A Kuhn. Image-guided endoscopic surgery: results of accuracy and performance in a multicenter clinical study using an electromagnetic tracking system. The Laryngoscope, 107(5):594-601, 1997. 2 +[18] Yixin Gao, S Swaroop Vedula, Carol E Reiley, Narges Ah-midi, Balakrishnan Varadarajan, Henry C Lin, Lingling Tao, Luca Zappella, Benjamin Bejar, David D Yuh, et al. Jhu-isi gesture and skill assessment working set (jigsaws): A surgical activity dataset for human motion modeling. In MICCAI workshop: M2cai, page 3, 2014. 2 +[19] Zheng Ge, Songtao Liu, Feng Wang, Zeming Li, and Jian Sun. YOLOX: Exceeding yolo series in 2021. arXiv preprint arXiv:2107.08430, 2021. 6 +[20] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In 2012 IEEE conference on computer vision and pattern recognition, pages 3354-3361. IEEE, 2012. 1, 2 +[21] Amy Jin, Serena Yeung, Jeffrey Jopling, Jonathan Krause, Dan Azagury, Arnold Milstein, and Li Fei-Fei. Tool detection and operative skill assessment in surgical videos using region-based convolutional neural networks. In 2018 IEEE winter conference on applications of computer vision (WACV), pages 691-699. IEEE, 2018. 5 + +[22] Arne Hoffhues Jonathon Luiten. Trackeval. https://github.com/JonathonLuiten/TrackEval, 2020. Accessed: 16 October 2023. 7 +[23] Kang Kim and Hee Seok Lee. Probabilistic anchor assignment with iou prediction for object detection. In ECCV, 2020. 6 +[24] Joel L Lavanchy, Joel Zindel, Kadir Kirtac, Isabell Twick, Enes Hosgor, Daniel Candinas, and Guido Beldi. Automation of surgical skill assessment using a three-stage machine learning algorithm. Scientific reports, 11(1):5197, 2021. 3 +[25] Joel L Lavanchy, Armine Vardazaryan, Pietro Mascagni, Didier Mutter, and Nicolas Padoy. Preserving privacy in surgical video analysis using a deep learning classifier to identify out-of-body scenes in endoscopic videos. Scientific reports, 13(1):9235, 2023. 3 +[26] Byungjae Lee, Enkhbayar Erdenee, Songguo Jin, Mi Young Nam, Young Giu Jung, and Phill Kyu Rhee. Multi-class multi-object tracking using changing point detection. In Computer Vision-ECCV 2016 Workshops: Amsterdam, The Netherlands, October 8-10 and 15-16, 2016, Proceedings, Part II 14, pages 68-83. Springer, 2016. 2 +[27] Dongheon Lee, Hyeong Won Yu, Hyungju Kwon, HyounJoong Kong, Kyu Eun Lee, and Hee Chan Kim. Evaluation of surgical skills during robotic surgery by deep learning-based multiple surgical instrument tracking in training and actual operations. Journal of clinical medicine, 9(6):1964, 2020. 3 +[28] Eung-Joo Lee, William Plishker, Xinyang Liu, Shuvra S Bhattacharyya, and Raj Shekhar. Weakly supervised segmentation for real-time surgical tool tracking. Healthcare technology letters, 6(6):231-236, 2019. 2 +[29] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 2 +[30] Wei Liu, Dragomir Anguelov, Dumitru Erhan, Christian Szegedy, Scott Reed, Cheng-Yang Fu, and Alexander C. Berg. Ssd: Single shot multibox detector. ECCV, 2016. 6 +[31] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. arXiv preprint arXiv:2103.14030, 2021. 6 +[32] Jonathon Luiten, Aljosa Osep, Patrick Dendorfer, Philip Torr, Andreas Geiger, Laura Leal-Taixe, and Bastian Leibe. Hota: A higher order metric for evaluating multi-object tracking. International journal of computer vision, 129:548-578, 2021. 7 +[33] Gerard Maggiolino, Adnan Ahmad, Jinkun Cao, and Kris Kitani. Deep oc-sort: Multi-pedestrian tracking by adaptive re-identification. arXiv preprint arXiv:2302.11813, 2023. 6, 7 +[34] Lena Maier-Hein, Swaroop S Vedula, Stefanie Speidel, Nassir Navab, Ron Kikinis, Adrian Park, Matthias Eisenmann, Hubertus Feussner, Germain Forestier, Stamatia Giannarou, et al. Surgical data science for next-generation interventions. Nature Biomedical Engineering, 1(9):691-696, 2017. 1 + +[35] Afrooz Moatari-Kazerouni and Ygal Bendavid. Improving logistics processes of surgical instruments: case of RFID technology. Business Process Management Journal, 23(2):448-466, 2017. 3 +[36] Chinedu Innocent Nwoye. Deep learning methods for the detection and recognition of surgical tools and activities in laparoscopic videos. PhD thesis, Université de Strasbourg, 2021. 1, 2 +[37] Chinedu Innocent Nwoye, Didier Mutter, Jacques Marescaux, and Nicolas Padoy. Weakly supervised convolutional LSTM approach for tool tracking in laparoscopic videos. International journal of computer assisted radiology and surgery, 14:1059-1067, 2019. 1, 2, 3 +[38] Chinedu Innocent Nwoye, Cristians Gonzalez, Tong Yu, Pietro Mascagni, Didier Mutter, Jacques Marescaux, and Nicolas Padoy. Recognition of instrument-tissue interactions in endoscopic videos via action triplets. In International Conference on Medical Image Computing and Computer-Assisted Intervention MICCAI, pages 364-374. Springer, 2020. 1 +[39] Chinedu Innocent Nwoye, Tong Yu, Cristians Gonzalez, Barbara Seeliger, Pietro Mascagni, Didier Mutter, Jacques Marescaux, and Nicolas Padoy. Rendezvous: Attention mechanisms for the recognition of surgical action triplets in endoscopic videos. 78:102433, 2022. 2, 3, 4 +[40] Chinedu Innocent Nwoye, Tong Yu, Saurav Sharma, Aditya Murali, Deepak Alapatt, Armine Vardazaryan, Kun Yuan, Jonas Hajek, Wolfgang Reiter, Amine Yamlahi, et al. Cholectriplet2022: Show me a tool and tell me the triplet—an endoscopic vision challenge for surgical action triplet detection. arXiv preprint arXiv:2302.06294, 2023. 3 +[41] Liang Qiu, Changsheng Li, and Hongliang Ren. Real-time surgical instrument tracking in robot-assisted surgery using multi-domain convolutional neural network. Healthcare technology letters, 6(6):159-164, 2019. 2 +[42] Sanat Ramesh, Vinkle Srivastav, Deepak Alapatt, Tong Yu, Aditya Murali, Luca Sestini, Chinedu Innocent Nwoye, Idris Hamoud, Saurav Sharma, Antoine Fleurentin, et al. Dissecting self-supervised learning methods for surgical computer vision. Medical Image Analysis, 88:102844, 2023. 1 +[43] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2017. 6 +[44] Rogério Richa, Marcin Balicki, Eric Meisner, Raphael Sznitman, Russell Taylor, and Gregory Hager. Visual tracking of surgical tools for proximity detection in retinal surgery. In Information Processing in Computer-Assisted Interventions: Second International Conference, IPCAI 2011, Berlin, Germany, June 22, 2011. Proceedings 2, pages 55-66. Springer, 2011. 3 +[45] Ergys Ristani, Francesco Solera, Roger Zou, Rita Cucchiara, and Carlo Tomasi. Performance measures and a data set for multi-target, multi-camera tracking. In European conference on computer vision, pages 17-35. Springer, 2016. 7 +[46] Maria Robu, Abdelrahim Kadkhodamohammadi, Imanol Luengo, and Danail Stoyanov. Towards real-time multiple + +surgical tool tracking. Computer Methods in Biomechanics and Biomedical Engineering: Imaging & Visualization, 9(3): 279-285, 2021. 1, 2, 3, 5 +[47] Duygu Sarikaya, Jason J Corso, and Khurshid A Guru. Detection and localization of robotic tools in robot-assisted surgery videos using deep neural networks for region proposal and detection. IEEE transactions on medical imaging, 36(7):1542-1549, 2017. 5 +[48] Mohammad A Shbool, Ammar Al-Bazi, Alma Kokash, AlAlaween Wafa'H, Nibal T Albashabsheh, and Raed Al-Taher. The economy of motion for laparoscopic ball clamping surgery: A feedback educational tool. MethodsX, 10: 102168, 2023. 3 +[49] Pan Shi, Zijian Zhao, Sanyuan Hu, and Faliang Chang. Realtime surgical tool detection in minimally invasive surgery based on attention-guided convolutional neural network. IEEE Access, 8:228853-228862, 2020. 5 +[50] Peize Sun, Jinkun Cao, Yi Jiang, Rufeng Zhang, Enze Xie, Zehuan Yuan, Changhu Wang, and Ping Luo. Transtrack: Multiple object tracking with transformer. arXiv preprint arXiv:2012.15460, 2020. 6, 7 +[51] Peize Sun, Jinkun Cao, Yi Jiang, Zehuan Yuan, Song Bai, Kris Kitani, and Ping Luo. Dancetrack: Multi-object tracking in uniform appearance and diverse motion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20993-21002, 2022. 1, 2 +[52] Xiaochuan Sun and Shahram Payandeh. Estimation of incision patterns based on visual tracking of surgical tools in minimally invasive surgery. In ASME International Mechanical Engineering Congress and Exposition, pages 75-83, 2010. 3 +[53] Raphael Sznitman, Karim Ali, Rogerio Richa, Russell H Taylor, Gregory D Hager, and Pascal Fua. Data-driven visual tracking in retinal microsurgery. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 568-575. Springer, 2012. 2, 5 +[54] Zhi Tian, Chunhua Shen, Hao Chen, and Tong He. Fcos: Fully convolutional one-stage object detection. arXiv preprint arXiv:1904.01355, 2019. 6 +[55] Jay Toor, Avneesh Bhangu, Jesse Wolfstadt, Garry Bassi, Stanley Chung, Raja Rampersaud, William Mitchell, Joseph Milner, and Martin Koyle. Optimizing the surgical instrument tray to immediately increase efficiency and lower costs in the operating room. Canadian Journal of Surgery, 65(2): E275, 2022. 3 +[56] Andru P Twinanda, Sherif Shehata, Didier Mutter, Jacques Marescaux, Michel De Mathelin, and Nicolas Padoy. Endonet: a deep architecture for recognition tasks on laparoscopic videos. IEEE transactions on medical imaging, 36 (1):86-97, 2016. 1, 2, 3 +[57] Armine Vardazaryan, Didier Mutter, Jacques Marescaux, and Nicolas Padoy. Weakly-supervised learning for tool localization in laparoscopic videos. In MICCAI-LABEL, pages 169–179. Springer, 2018. 3 +[58] Rejin Varghese and M Sambath. Yolov8: A novel object detection algorithm with enhanced performance and robustness. In 2024 International Conference on Advances in Data + +Engineering and Intelligent Computing Systems (ADICS), pages 1-6. IEEE, 2024. 6 +[59] Martin Wagner, Beat-Peter Müller-Stich, Anna Kisilenko, Duc Tran, Patrick Heger, Lars Mündermann, David M Lubotsky, Benjamin Müller, Tornike Davitashvili, Manuela Capek, et al. Comparative validation of machine learning algorithms for surgical workflow and skill analysis with the heichole benchmark. Medical Image Analysis, 86:102770, 2023. 1 +[60] Ao Wang, Hui Chen, Lihao Liu, Kai Chen, Zijia Lin, Jungong Han, et al. Yolov10: Real-time end-to-end object detection. Advances in Neural Information Processing Systems, 37:107984-108011, 2024. 6 +[61] Chien-Yao Wang and Hong-Yuan Mark Liao. Yolov9: Learning what you want to learn using programmable gradient information. 2024. 6 +[62] Chien-Yao Wang, Alexey Bochkovskyi, and Hong-Yuan Mark Liao. Yolov7: Trainable bag-of-freebies sets new state-of-the-art for real-time object detectors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7464-7475, 2023. 6 +[63] Yu-Hsiang Wang, Jun-Wei Hsieh, Ping-Yang Chen, Ming-Ching Chang, Hung Hin So, and Xin Li. Smiletrack: Similarity learning for occlusion-aware multiple object tracking, 2023. 6, 7 +[64] Nicolai Wojke, Alex Bewley, and Dietrich Paulus. Simple online and realtime tracking with a deep association metric. In 2017 IEEE international conference on image processing (ICIP), pages 3645-3649. IEEE, 2017. 2 +[65] Lisheng Xu, Haoran Zhang, Jiaole Wang, Ang Li, Shuang Song, Hongliang Ren, Lin Qi, Jason J Gu, and Max Q-H Meng. Information loss challenges in surgical navigation systems: From information fusion to ai-based approaches. Information Fusion, 2022. 1 +[66] Yifu Zhang, Chunyu Wang, Xinggang Wang, Wenjun Zeng, and Wenyu Liu. Fairmot: On the fairness of detection and re-identification in multiple object tracking. International Journal of Computer Vision, 129:3069–3087, 2021. 2, 7 +[67] Yifu Zhang, Peize Sun, Yi Jiang, Dongdong Yu, Fucheng Weng, Zehuan Yuan, Ping Luo, Wenyu Liu, and Xinggang Wang. Bytetrack: Multi-object tracking by associating every detection box. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXII, pages 1-21. Springer, 2022. 6, 7 +[68] Zijian Zhao, Sandrine Voros, Zhaorui Chen, and Xiaolin Cheng. Surgical tool tracking based on two cnns: from coarse to fine. The Journal of Engineering, 2019(14):467-472, 2019. 1 +[69] Xingyi Zhou, Dequan Wang, and Philipp Krahenbuhl. Objects as points. arXiv preprint arXiv:1904.07850, 2019. 6 +[70] Pengfei Zhu, Longyin Wen, Dawei Du, Xiao Bian, Heng Fan, Qinghua Hu, and Haibin Ling. Detection and tracking meet drones challenge. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(11):7380-7399, 2021. 2 +[71] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. In International Conference on Learning Representations, 2021. 6 \ No newline at end of file diff --git a/2025/CholecTrack20_ A Multi-Perspective Tracking Dataset for Surgical Tools/images.zip b/2025/CholecTrack20_ A Multi-Perspective Tracking Dataset for Surgical Tools/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..0bcee357054b26d7cb1a373b26da67f63004b190 --- /dev/null +++ b/2025/CholecTrack20_ A Multi-Perspective Tracking Dataset for Surgical Tools/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4c106df61b1df6136b6f9d526a822b37e477417adaae079a9372505679fb6f81 +size 764675 diff --git a/2025/CholecTrack20_ A Multi-Perspective Tracking Dataset for Surgical Tools/layout.json b/2025/CholecTrack20_ A Multi-Perspective Tracking Dataset for Surgical Tools/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..28430c645cd678eca1df3bab34ccc006779c8a95 --- /dev/null +++ b/2025/CholecTrack20_ A Multi-Perspective Tracking Dataset for Surgical Tools/layout.json @@ -0,0 +1,10761 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 81, + 103, + 531, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 103, + 531, + 121 + ], + "spans": [ + { + "bbox": [ + 81, + 103, + 531, + 121 + ], + "type": "text", + "content": "CholecTrack20: A Multi-Perspective Tracking Dataset for Surgical Tools" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 137, + 142, + 473, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 142, + 473, + 172 + ], + "spans": [ + { + "bbox": [ + 137, + 142, + 473, + 172 + ], + "type": "text", + "content": "Chinedu Innocent Nwoye" + }, + { + "bbox": [ + 137, + 142, + 473, + 172 + ], + "type": "inline_equation", + "content": "^{1,3}" + }, + { + "bbox": [ + 137, + 142, + 473, + 172 + ], + "type": "text", + "content": " Kareem Elgohary" + }, + { + "bbox": [ + 137, + 142, + 473, + 172 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 137, + 142, + 473, + 172 + ], + "type": "text", + "content": " Anvita Srinivas" + }, + { + "bbox": [ + 137, + 142, + 473, + 172 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 137, + 142, + 473, + 172 + ], + "type": "text", + "content": " Fauzan Zaid" + }, + { + "bbox": [ + 137, + 142, + 473, + 172 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 137, + 142, + 473, + 172 + ], + "type": "text", + "content": " Joel L. Lavanchy" + }, + { + "bbox": [ + 137, + 142, + 473, + 172 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 137, + 142, + 473, + 172 + ], + "type": "text", + "content": " Nicolas Padoy" + }, + { + "bbox": [ + 137, + 142, + 473, + 172 + ], + "type": "inline_equation", + "content": "^{1,3}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 135, + 175, + 475, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 175, + 475, + 189 + ], + "spans": [ + { + "bbox": [ + 135, + 175, + 475, + 189 + ], + "type": "text", + "content": "1University of Strasbourg, CNRS, INSERM, ICube, UMR7357, Strasbourg, France" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 135, + 190, + 476, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 190, + 476, + 202 + ], + "spans": [ + { + "bbox": [ + 135, + 190, + 476, + 202 + ], + "type": "text", + "content": "2University of Basel, University Digestive Health Care Center, Clarunis, Switzerland" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 255, + 204, + 356, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 255, + 204, + 356, + 217 + ], + "spans": [ + { + "bbox": [ + 255, + 204, + 356, + 217 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 255, + 204, + 356, + 217 + ], + "type": "text", + "content": "IHU Strasbourg, France" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 180, + 220, + 426, + 230 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 180, + 220, + 426, + 230 + ], + "spans": [ + { + "bbox": [ + 180, + 220, + 426, + 230 + ], + "type": "text", + "content": "https://github.com/camma-public/cholectrack20" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 151, + 258, + 200, + 271 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 258, + 200, + 271 + ], + "spans": [ + { + "bbox": [ + 151, + 258, + 200, + 271 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 284, + 297, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 284, + 297, + 620 + ], + "spans": [ + { + "bbox": [ + 55, + 284, + 297, + 620 + ], + "type": "text", + "content": "Tool tracking in surgical videos is essential for advancing computer-assisted interventions, such as skill assessment, safety zone estimation, and human-machine collaboration. However, the lack of context-rich datasets limits AI applications in this field. Existing datasets rely on overly generic tracking formalizations that fail to capture surgical-specific dynamics, such as tools moving out of the camera's view or exiting the body. This results in less clinically relevant trajectories and a lack of flexibility for real-world surgical applications. Methods trained on these datasets often struggle with visual challenges such as smoke, reflection, and bleeding, further exposing the limitations of current approaches. We introduce CholecTrack20, a specialized dataset for multi-class, multi-tool tracking in surgical procedures. It redefines tracking formalization with three perspectives: (1) intraoperative, (2) intracorporeal, and (3) visibility, enabling adaptable and clinically meaningful tool trajectories. The dataset comprises 20 full-length surgical videos, annotated at 1 fps, yielding over 35K frames and 65K labeled tool instances. Annotations include spatial location, category, identity, operator, phase, and visual challenges. Benchmarking state-of-the-art methods on CholecTrack20 reveals significant performance gaps, with current approaches (< 45% HOTA) failing to meet the accuracy required for clinical translation. These findings motivate the need for advanced and intuitive tracking algorithms and establish CholecTrack20 as a foundation for developing robust AI-driven surgical assistance systems." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 644, + 135, + 657 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 644, + 135, + 657 + ], + "spans": [ + { + "bbox": [ + 56, + 644, + 135, + 657 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "content": "The true impact of computer vision research lies in its practical applications, especially in critical fields like healthcare. Among these, surgery represents one of the most demanding domains, providing a definitive test for the capa" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 259, + 555, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 259, + 555, + 608 + ], + "spans": [ + { + "bbox": [ + 313, + 259, + 555, + 608 + ], + "type": "text", + "content": "bilities of vision technologies. Surgical data science, on its own, has significantly advanced interventional healthcare by leveraging data-driven techniques to provide critical decision support to medical professionals [34]. A key area of this advancement is the analysis of endoscopic video data, which offers real-time insights into surgical procedures, aids in skill assessment, and helps predict complications [9, 56, 65]. Accurate tracking of surgical tools is central to these analyses, as it guides temporal progression of procedural phases and correlates with surgical actions and management of adverse events [38, 59]. Despite progress in computer vision, research [42] shows that deep learning models pretrained on general datasets often struggle with surgical contexts due to complex scene dynamics, diverse tool types, and challenging visual conditions such as bleeding, smoke, and variable lighting [12, 20, 51]. This highlights the need for domain-specific datasets tailored to the unique requirements of surgical tool tracking. Obtaining medical and surgical data for research is challenging due to ethical and practical constraints, and annotating it requires expert knowledge. Current methods for tool tracking largely focus on Single Object Tracking (SOT) [68], Multi-Class Tracking (MCT) with one tool per class [36, 37], or Multi-Object Tracking (MOT) treating all tools as a single class [16, 46]. However, these approaches often miss the complexities of Multi-Class Multi-Object Tracking (MC-MOT) specific to surgical contexts, where tools interact dynamically and may move out of the camera's field of view or within the body cavity." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "text", + "content": "Multi-perspective tracking addresses these challenges by defining tool trajectories across different viewpoints during surgical procedures. It includes three critical perspectives (illustrated in Fig. 1): (1) intraoperative - covering the entire procedure duration to monitor tool usage and assess surgical proficiency; (2) intracorporeal - focusing on tool tracks within the body to evaluate specific tasks and predict risks; and (3) visibility - tracking tools within the camera's" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8942" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 81, + 206, + 221 + ], + "blocks": [ + { + "bbox": [ + 60, + 71, + 203, + 79 + ], + "lines": [ + { + "bbox": [ + 60, + 71, + 203, + 79 + ], + "spans": [ + { + "bbox": [ + 60, + 71, + 203, + 79 + ], + "type": "text", + "content": "Surgeon operators positioning and tool placement" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 60, + 81, + 206, + 221 + ], + "lines": [ + { + "bbox": [ + 60, + 81, + 206, + 221 + ], + "spans": [ + { + "bbox": [ + 60, + 81, + 206, + 221 + ], + "type": "image", + "image_path": "fb7f4a50edb064bbba40a2063ab477c772e2c10e47cee16d12dc6ec670f06693.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 210, + 81, + 410, + 202 + ], + "blocks": [ + { + "bbox": [ + 234, + 72, + 356, + 79 + ], + "lines": [ + { + "bbox": [ + 234, + 72, + 356, + 79 + ], + "spans": [ + { + "bbox": [ + 234, + 72, + 356, + 79 + ], + "type": "text", + "content": "Multi-perspective of surgical tool trajectory" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 210, + 81, + 410, + 202 + ], + "lines": [ + { + "bbox": [ + 210, + 81, + 410, + 202 + ], + "spans": [ + { + "bbox": [ + 210, + 81, + 410, + 202 + ], + "type": "image", + "image_path": "d55b6a8fed8b89cd2f3ab3d8e1468aab64f78aa9e614eb87498504c462869ee2.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 394, + 83, + 549, + 184 + ], + "blocks": [ + { + "bbox": [ + 427, + 72, + 515, + 79 + ], + "lines": [ + { + "bbox": [ + 427, + 72, + 515, + 79 + ], + "spans": [ + { + "bbox": [ + 427, + 72, + 515, + 79 + ], + "type": "text", + "content": "Sample data and label overlaid" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 394, + 83, + 549, + 184 + ], + "lines": [ + { + "bbox": [ + 394, + 83, + 549, + 184 + ], + "spans": [ + { + "bbox": [ + 394, + 83, + 549, + 184 + ], + "type": "image", + "image_path": "2a8b05911b3f6726b8b0b485f2fa6562255f4383d24838cdab50923c7503b2af.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 63, + 239, + 91, + 257 + ], + "blocks": [ + { + "bbox": [ + 75, + 229, + 174, + 237 + ], + "lines": [ + { + "bbox": [ + 75, + 229, + 174, + 237 + ], + "spans": [ + { + "bbox": [ + 75, + 229, + 174, + 237 + ], + "type": "text", + "content": "Considered surgical tool categories" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 63, + 239, + 91, + 257 + ], + "lines": [ + { + "bbox": [ + 63, + 239, + 91, + 257 + ], + "spans": [ + { + "bbox": [ + 63, + 239, + 91, + 257 + ], + "type": "image", + "image_path": "a1e3fce7f8d791780bb9dae68750a87f66ed65884246ea358f566051640babe3.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 94, + 239, + 123, + 257 + ], + "blocks": [ + { + "bbox": [ + 94, + 239, + 123, + 257 + ], + "lines": [ + { + "bbox": [ + 94, + 239, + 123, + 257 + ], + "spans": [ + { + "bbox": [ + 94, + 239, + 123, + 257 + ], + "type": "image", + "image_path": "2221ca3216cc5592037105c619f4d105fe230a2f696aed73c433fdab8bc762dc.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 100, + 258, + 117, + 264 + ], + "lines": [ + { + "bbox": [ + 100, + 258, + 117, + 264 + ], + "spans": [ + { + "bbox": [ + 100, + 258, + 117, + 264 + ], + "type": "text", + "content": "bipolar" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 124, + 239, + 155, + 257 + ], + "blocks": [ + { + "bbox": [ + 124, + 239, + 155, + 257 + ], + "lines": [ + { + "bbox": [ + 124, + 239, + 155, + 257 + ], + "spans": [ + { + "bbox": [ + 124, + 239, + 155, + 257 + ], + "type": "image", + "image_path": "e6df6ba0977b91ab3652d9ec1c2bd85b594a667e9997ae40d4178698f54b260d.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 133, + 258, + 146, + 264 + ], + "lines": [ + { + "bbox": [ + 133, + 258, + 146, + 264 + ], + "spans": [ + { + "bbox": [ + 133, + 258, + 146, + 264 + ], + "type": "text", + "content": "hook" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 156, + 239, + 184, + 257 + ], + "blocks": [ + { + "bbox": [ + 156, + 239, + 184, + 257 + ], + "lines": [ + { + "bbox": [ + 156, + 239, + 184, + 257 + ], + "spans": [ + { + "bbox": [ + 156, + 239, + 184, + 257 + ], + "type": "image", + "image_path": "2b3f98256588cb824ae067c4ca0be875d5ec9afe922f11a28fe12797af82e020.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 159, + 258, + 179, + 264 + ], + "lines": [ + { + "bbox": [ + 159, + 258, + 179, + 264 + ], + "spans": [ + { + "bbox": [ + 159, + 258, + 179, + 264 + ], + "type": "text", + "content": "scissors" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 74, + 266, + 104, + 289 + ], + "blocks": [ + { + "bbox": [ + 69, + 258, + 86, + 264 + ], + "lines": [ + { + "bbox": [ + 69, + 258, + 86, + 264 + ], + "spans": [ + { + "bbox": [ + 69, + 258, + 86, + 264 + ], + "type": "text", + "content": "grasper" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 74, + 266, + 104, + 289 + ], + "lines": [ + { + "bbox": [ + 74, + 266, + 104, + 289 + ], + "spans": [ + { + "bbox": [ + 74, + 266, + 104, + 289 + ], + "type": "image", + "image_path": "97bb9c4cd069f1d9e0b835a0f4d61815dc77c0feb12eeae99df7faacf88c87ee.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 110, + 266, + 137, + 289 + ], + "blocks": [ + { + "bbox": [ + 110, + 266, + 137, + 289 + ], + "lines": [ + { + "bbox": [ + 110, + 266, + 137, + 289 + ], + "spans": [ + { + "bbox": [ + 110, + 266, + 137, + 289 + ], + "type": "image", + "image_path": "9a6618930d7de982a75cd5bcd79885d5012884ae6665818bc848556faf5522aa.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 70, + 302, + 538, + 313 + ], + "lines": [ + { + "bbox": [ + 70, + 302, + 538, + 313 + ], + "spans": [ + { + "bbox": [ + 70, + 302, + 538, + 313 + ], + "type": "text", + "content": "Figure 1. Illustration of multi-perspective tracking in surgical domain and CholecTrack20 dataset labels for surgical tool tracking." + } + ] + } + ], + "index": 49, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 141, + 266, + 170, + 289 + ], + "blocks": [ + { + "bbox": [ + 141, + 266, + 170, + 289 + ], + "lines": [ + { + "bbox": [ + 141, + 266, + 170, + 289 + ], + "spans": [ + { + "bbox": [ + 141, + 266, + 170, + 289 + ], + "type": "image", + "image_path": "66360be32b62168d2ffdb758a6d43526e9aa634684e13853376e1b7ed1ee49eb.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 195, + 230, + 244, + 279 + ], + "blocks": [ + { + "bbox": [ + 195, + 230, + 244, + 279 + ], + "lines": [ + { + "bbox": [ + 195, + 230, + 244, + 279 + ], + "spans": [ + { + "bbox": [ + 195, + 230, + 244, + 279 + ], + "type": "image", + "image_path": "f4e498405819c4c83448ab7bfcaa82de25ba95094f588a7ee9b3cf452174901f.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 206, + 279, + 241, + 289 + ], + "lines": [ + { + "bbox": [ + 206, + 279, + 241, + 289 + ], + "spans": [ + { + "bbox": [ + 206, + 279, + 241, + 289 + ], + "type": "text", + "content": "Visibility tracking FoV" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 244, + 230, + 251, + 236 + ], + "lines": [ + { + "bbox": [ + 244, + 230, + 251, + 236 + ], + "spans": [ + { + "bbox": [ + 244, + 230, + 251, + 236 + ], + "type": "text", + "content": "1" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 252, + 247, + 286, + 253 + ], + "lines": [ + { + "bbox": [ + 252, + 247, + 286, + 253 + ], + "spans": [ + { + "bbox": [ + 252, + 247, + 286, + 253 + ], + "type": "text", + "content": "Operating room" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 252, + 253, + 286, + 258 + ], + "lines": [ + { + "bbox": [ + 252, + 253, + 286, + 258 + ], + "spans": [ + { + "bbox": [ + 252, + 253, + 286, + 258 + ], + "type": "inline_equation", + "content": "\\left( 4\\right)" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 252, + 258, + 286, + 263 + ], + "lines": [ + { + "bbox": [ + 252, + 258, + 286, + 263 + ], + "spans": [ + { + "bbox": [ + 252, + 258, + 286, + 263 + ], + "type": "inline_equation", + "content": "\\therefore m - 1 \\neq 0" + }, + { + "bbox": [ + 252, + 258, + 286, + 263 + ], + "type": "text", + "content": " ;" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 252, + 263, + 286, + 268 + ], + "lines": [ + { + "bbox": [ + 252, + 263, + 286, + 268 + ], + "spans": [ + { + "bbox": [ + 252, + 263, + 286, + 268 + ], + "type": "inline_equation", + "content": "\\therefore m - 1 \\neq 0" + }, + { + "bbox": [ + 252, + 263, + 286, + 268 + ], + "type": "text", + "content": " ;" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 252, + 268, + 286, + 274 + ], + "lines": [ + { + "bbox": [ + 252, + 268, + 286, + 274 + ], + "spans": [ + { + "bbox": [ + 252, + 268, + 286, + 274 + ], + "type": "text", + "content": "" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "bbox": [ + 274, + 247, + 285, + 253 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 274, + 247, + 285, + 253 + ], + "spans": [ + { + "bbox": [ + 274, + 247, + 285, + 253 + ], + "type": "text", + "content": "room" + } + ] + } + ], + "index": 26, + "type": "text" + }, + { + "bbox": [ + 274, + 253, + 285, + 258 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 274, + 253, + 285, + 258 + ], + "spans": [ + { + "bbox": [ + 274, + 253, + 285, + 258 + ], + "type": "inline_equation", + "content": "\\frac{1 + u}{1} - \\frac{u}{1} = \\frac{\\left( {1 + u}\\right) u}{1} < \\frac{u}{1} = u" + } + ] + } + ], + "index": 27, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 339, + 230, + 408, + 279 + ], + "blocks": [ + { + "bbox": [ + 339, + 230, + 408, + 279 + ], + "lines": [ + { + "bbox": [ + 339, + 230, + 408, + 279 + ], + "spans": [ + { + "bbox": [ + 339, + 230, + 408, + 279 + ], + "type": "image", + "image_path": "7b2f28c59d7f50c8b0b572afea6afa97a72cb33fe11d4e82a08413594a194c71.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 355, + 279, + 403, + 289 + ], + "lines": [ + { + "bbox": [ + 355, + 279, + 403, + 289 + ], + "spans": [ + { + "bbox": [ + 355, + 279, + 403, + 289 + ], + "type": "text", + "content": "Intracorporeal tracking Fov + OOCV" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 420, + 201, + 462, + 225 + ], + "blocks": [ + { + "bbox": [ + 437, + 192, + 531, + 199 + ], + "lines": [ + { + "bbox": [ + 437, + 192, + 531, + 199 + ], + "spans": [ + { + "bbox": [ + 437, + 192, + 531, + 199 + ], + "type": "text", + "content": "Sample surgical visual challenges" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 420, + 201, + 462, + 225 + ], + "lines": [ + { + "bbox": [ + 420, + 201, + 462, + 225 + ], + "spans": [ + { + "bbox": [ + 420, + 201, + 462, + 225 + ], + "type": "image", + "image_path": "bc78089af8921d384819590d80ba041c3235476bff83da1acc8f5c9aae9ca30f.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 420, + 231, + 462, + 255 + ], + "blocks": [ + { + "bbox": [ + 432, + 226, + 449, + 231 + ], + "lines": [ + { + "bbox": [ + 432, + 226, + 449, + 231 + ], + "spans": [ + { + "bbox": [ + 432, + 226, + 449, + 231 + ], + "type": "text", + "content": "Blurring" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 420, + 231, + 462, + 255 + ], + "lines": [ + { + "bbox": [ + 420, + 231, + 462, + 255 + ], + "spans": [ + { + "bbox": [ + 420, + 231, + 462, + 255 + ], + "type": "image", + "image_path": "16729a3eafacbdcea5c9786f5a15b60f9716df083e1561d869a55f7c2649cffc.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 429, + 256, + 454, + 260 + ], + "lines": [ + { + "bbox": [ + 429, + 256, + 454, + 260 + ], + "spans": [ + { + "bbox": [ + 429, + 256, + 454, + 260 + ], + "type": "text", + "content": "Trocar view" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_caption" + } + ], + "index": 33 + }, + { + "type": "image", + "bbox": [ + 463, + 201, + 504, + 225 + ], + "blocks": [ + { + "bbox": [ + 463, + 201, + 504, + 225 + ], + "lines": [ + { + "bbox": [ + 463, + 201, + 504, + 225 + ], + "spans": [ + { + "bbox": [ + 463, + 201, + 504, + 225 + ], + "type": "image", + "image_path": "7b710929e27e75ae90105524806d27f6e76d4df1cb9dd6c16d065663c3c27177.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + } + ], + "index": 35 + }, + { + "type": "image", + "bbox": [ + 463, + 231, + 504, + 255 + ], + "blocks": [ + { + "bbox": [ + 474, + 226, + 492, + 231 + ], + "lines": [ + { + "bbox": [ + 474, + 226, + 492, + 231 + ], + "spans": [ + { + "bbox": [ + 474, + 226, + 492, + 231 + ], + "type": "text", + "content": "Bleeding" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 463, + 231, + 504, + 255 + ], + "lines": [ + { + "bbox": [ + 463, + 231, + 504, + 255 + ], + "spans": [ + { + "bbox": [ + 463, + 231, + 504, + 255 + ], + "type": "image", + "image_path": "c761f4abba5cfaad4418a4f9328434cd71bd448fb237fc9b9e805277c8174407.jpg" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 473, + 256, + 494, + 260 + ], + "lines": [ + { + "bbox": [ + 473, + 256, + 494, + 260 + ], + "spans": [ + { + "bbox": [ + 473, + 256, + 494, + 260 + ], + "type": "text", + "content": "Reflection" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_caption" + } + ], + "index": 37 + }, + { + "type": "image", + "bbox": [ + 506, + 201, + 549, + 225 + ], + "blocks": [ + { + "bbox": [ + 506, + 201, + 549, + 225 + ], + "lines": [ + { + "bbox": [ + 506, + 201, + 549, + 225 + ], + "spans": [ + { + "bbox": [ + 506, + 201, + 549, + 225 + ], + "type": "image", + "image_path": "1ab595c418f3f88042bde5c82b69227b8634f8f679f2ff8c963bde084fb8438a.jpg" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_body" + } + ], + "index": 39 + }, + { + "type": "image", + "bbox": [ + 506, + 231, + 548, + 255 + ], + "blocks": [ + { + "bbox": [ + 518, + 226, + 537, + 231 + ], + "lines": [ + { + "bbox": [ + 518, + 226, + 537, + 231 + ], + "spans": [ + { + "bbox": [ + 518, + 226, + 537, + 231 + ], + "type": "text", + "content": "Crowded" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 506, + 231, + 548, + 255 + ], + "lines": [ + { + "bbox": [ + 506, + 231, + 548, + 255 + ], + "spans": [ + { + "bbox": [ + 506, + 231, + 548, + 255 + ], + "type": "image", + "image_path": "c20fd387dde5b8625bc4945b9a75bf5f4e0c913ad1c84f7daa96fc04ca2cba7b.jpg" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 520, + 256, + 534, + 260 + ], + "lines": [ + { + "bbox": [ + 520, + 256, + 534, + 260 + ], + "spans": [ + { + "bbox": [ + 520, + 256, + 534, + 260 + ], + "type": "text", + "content": "Smoke" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_caption" + } + ], + "index": 41 + }, + { + "type": "image", + "bbox": [ + 420, + 262, + 462, + 286 + ], + "blocks": [ + { + "bbox": [ + 420, + 262, + 462, + 286 + ], + "lines": [ + { + "bbox": [ + 420, + 262, + 462, + 286 + ], + "spans": [ + { + "bbox": [ + 420, + 262, + 462, + 286 + ], + "type": "image", + "image_path": "2d9c6b533680090b55241b6e7e461bfee4690e482a7c90c6f473c827d74b4723.jpg" + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 430, + 287, + 451, + 291 + ], + "lines": [ + { + "bbox": [ + 430, + 287, + 451, + 291 + ], + "spans": [ + { + "bbox": [ + 430, + 287, + 451, + 291 + ], + "type": "text", + "content": "Occlusion" + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "image_caption" + } + ], + "index": 43 + }, + { + "type": "image", + "bbox": [ + 463, + 262, + 504, + 286 + ], + "blocks": [ + { + "bbox": [ + 463, + 262, + 504, + 286 + ], + "lines": [ + { + "bbox": [ + 463, + 262, + 504, + 286 + ], + "spans": [ + { + "bbox": [ + 463, + 262, + 504, + 286 + ], + "type": "image", + "image_path": "87d9566a6f14a86753e0a91700e3ded82fe6eafb79da1a37960515ed70e4b06e.jpg" + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 471, + 287, + 496, + 292 + ], + "lines": [ + { + "bbox": [ + 471, + 287, + 496, + 292 + ], + "spans": [ + { + "bbox": [ + 471, + 287, + 496, + 292 + ], + "type": "text", + "content": "Lens fouling" + } + ] + } + ], + "index": 46, + "angle": 0, + "type": "image_caption" + } + ], + "index": 45 + }, + { + "type": "image", + "bbox": [ + 506, + 262, + 549, + 286 + ], + "blocks": [ + { + "bbox": [ + 506, + 262, + 549, + 286 + ], + "lines": [ + { + "bbox": [ + 506, + 262, + 549, + 286 + ], + "spans": [ + { + "bbox": [ + 506, + 262, + 549, + 286 + ], + "type": "image", + "image_path": "4a07ecb3239a181d166aaa80984ab3a38fab92751fd597b1205296c22d919770.jpg" + } + ] + } + ], + "index": 47, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 517, + 287, + 538, + 291 + ], + "lines": [ + { + "bbox": [ + 517, + 287, + 538, + 291 + ], + "spans": [ + { + "bbox": [ + 517, + 287, + 538, + 291 + ], + "type": "text", + "content": "Blood stain" + } + ] + } + ], + "index": 48, + "angle": 0, + "type": "image_caption" + } + ], + "index": 47 + }, + { + "bbox": [ + 54, + 335, + 295, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 335, + 295, + 418 + ], + "spans": [ + { + "bbox": [ + 54, + 335, + 295, + 418 + ], + "type": "text", + "content": "field of view to provide real-time feedback to surgeons. Existing surgical tracking datasets [18, 53] often lack these levels of granularity and adaptability needed for comprehensive tool modeling. They generally follow generic tracking formalizations and struggle to capture the intricacies of surgical tool interactions, particularly when tools are replaced or move beyond camera visibility [6, 28, 53]." + } + ] + } + ], + "index": 50 + }, + { + "bbox": [ + 55, + 419, + 295, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 419, + 295, + 647 + ], + "spans": [ + { + "bbox": [ + 55, + 419, + 295, + 647 + ], + "type": "text", + "content": "To address this gap, we introduce CholecTrack20, a novel dataset, based on laparoscopic cholecystectomy surgery, designed for multi-class multi-tool tracking from intraoperative, intracorporeal, and visibility perspectives. Derived from raw laparoscopic videos [39, 56], it includes detailed annotations such as spatial coordinates, tool categories, track identities (IDs), visual challenges, phase labels, and other scene attributes. This dataset enhances benchmarking resources for computer vision research and supports the development of AI models tailored to surgical tool tracking, phase recognition, and surgeon performance assessment. This paper provides a comprehensive overview of the data acquisition and annotation methodology, detailed data analysis, and technical validation. In addition, we conduct extensive benchmark experiments using state-of-the-art deep learning methods for object detection and tracking, evaluating performance across various surgical phases and visual challenges, discussing insightful findings." + } + ] + } + ], + "index": 51 + }, + { + "bbox": [ + 67, + 649, + 230, + 660 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 649, + 230, + 660 + ], + "spans": [ + { + "bbox": [ + 67, + 649, + 230, + 660 + ], + "type": "text", + "content": "In summary, the main contributions are:" + } + ] + } + ], + "index": 52 + }, + { + "bbox": [ + 55, + 665, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 55, + "blocks": [ + { + "bbox": [ + 56, + 665, + 295, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 665, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 665, + 295, + 689 + ], + "type": "text", + "content": "1. Introduction of CholecTrack20, a pioneering dataset for multi-perspective tracking with extensive annotations." + } + ] + } + ], + "index": 53 + }, + { + "bbox": [ + 55, + 689, + 294, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 294, + 713 + ], + "type": "text", + "content": "2. Extensive experimental analysis validating the dataset's effectiveness for surgical tool detection and tracking." + } + ] + } + ], + "index": 54 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 335, + 553, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 335, + 553, + 370 + ], + "spans": [ + { + "bbox": [ + 313, + 335, + 553, + 370 + ], + "type": "text", + "content": "3. Insights from model performance analysis under diverse visual challenges, highlighting the utility of each tracking perspective for AI-driven surgical solutions." + } + ] + } + ], + "index": 56 + }, + { + "bbox": [ + 314, + 398, + 405, + 410 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 398, + 405, + 410 + ], + "spans": [ + { + "bbox": [ + 314, + 398, + 405, + 410 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 57 + }, + { + "bbox": [ + 313, + 423, + 555, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 423, + 555, + 556 + ], + "spans": [ + { + "bbox": [ + 313, + 423, + 555, + 556 + ], + "type": "text", + "content": "Object detection and tracking. Advances in object detection and tracking have been driven by datasets like COCO [29], KITTI [20], MOTChallenge [12], VisDrone [70], DanceTrack [51], and TAO [11], enabling progress in Single Object Tracking (SOT) [10], Multi-Object Tracking (MOT) [64, 66], and Multi-Class Multi-Object Tracking (MCMOT) [13, 26]. Applying these techniques to surgical tool tracking poses challenges especially in the phase of bleeding, smoke, rapid movements, and variable lighting. Traditional tracking, centered on visibility, struggles when tools leave the camera's view or are replaced during surgery." + } + ] + } + ], + "index": 58 + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "content": "Surgical tool tracking and configurations. While electromagnetic and optical tracking methods [8, 17] have been explored in surgical domain, image-based approaches [53] better align with surgeons' view but face issues like identity fragmentation, identity switch, and low tracking accuracy [6, 16, 28, 36, 37, 46] especially with mid-procedure tool replacements and tools exiting/re-entering the field of view or body cavity. To fully capture tool usage complexity, it is crucial to consider multiple perspectives on tool trajectories. Existing datasets [16, 41] provide insights but lack comprehensive coverage of these scenarios, highlighting the need for detailed datasets addressing these complexities." + } + ] + } + ], + "index": 59 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8943" + } + ] + } + ], + "index": 60 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 71, + 138, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 71, + 138, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 138, + 85 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 163 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 163 + ], + "type": "text", + "content": "CholecTrack20 is a detailed dataset for surgical tool tracking in laparoscopic cholecystectomy, offering binary and spatial annotations for tools, including identity, category, bounding box location, motion, operator, phase, activity, usage conditions, and visual challenges, essential for training and benchmarking surgical AI tool tracking algorithms." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 173, + 228, + 186 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 173, + 228, + 186 + ], + "spans": [ + { + "bbox": [ + 55, + 173, + 228, + 186 + ], + "type": "text", + "content": "3.1. Data Acquisition and Collection" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 192, + 295, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 192, + 295, + 264 + ], + "spans": [ + { + "bbox": [ + 55, + 192, + 295, + 264 + ], + "type": "text", + "content": "Data source. The raw videos are sourced from the publicly available Cholec80 [56] and CholecT50 [39] datasets, with appropriate permissions and adherence to license terms. Recorded at the University Hospital of Strasbourg with the aid of laparoscopic cameras, these videos document laparoscopic cholecystectomy surgeries." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 269, + 296, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 269, + 296, + 341 + ], + "spans": [ + { + "bbox": [ + 55, + 269, + 296, + 341 + ], + "type": "text", + "content": "Video selection. Long videos were systematically chosen to represent surgical complexities and tool variability, capturing all key phases of laparoscopic cholecystectomy. High-quality videos with clear visuals were selected, subsampled from 25 FPS to 1 FPS for annotation, maintaining temporal consistency and clarity." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 345, + 296, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 345, + 296, + 417 + ], + "spans": [ + { + "bbox": [ + 55, + 345, + 296, + 417 + ], + "type": "text", + "content": "Sensitive data handling. To protect patient and staff identities, out-of-body frames potentially revealing sensitive information, such as the identity of patients, clinical staff, or the operating room, were checked and anonymized following established techniques [25]. This ensures compliance with ethical standards and privacy regulations." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 426, + 175, + 437 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 426, + 175, + 437 + ], + "spans": [ + { + "bbox": [ + 55, + 426, + 175, + 437 + ], + "type": "text", + "content": "3.2. Track Formalization" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 445, + 295, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 445, + 295, + 649 + ], + "spans": [ + { + "bbox": [ + 55, + 445, + 295, + 649 + ], + "type": "text", + "content": "Given a video dataset " + }, + { + "bbox": [ + 55, + 445, + 295, + 649 + ], + "type": "inline_equation", + "content": "D = \\{S_{1}, S_{2}, \\ldots, S_{n}\\}" + }, + { + "bbox": [ + 55, + 445, + 295, + 649 + ], + "type": "text", + "content": " of laparoscopic surgeries, each sequence " + }, + { + "bbox": [ + 55, + 445, + 295, + 649 + ], + "type": "inline_equation", + "content": "S_{i}" + }, + { + "bbox": [ + 55, + 445, + 295, + 649 + ], + "type": "text", + "content": " includes frames annotated with bounding boxes " + }, + { + "bbox": [ + 55, + 445, + 295, + 649 + ], + "type": "inline_equation", + "content": "B = [B_{1}, B_{2}, \\ldots, B_{M}]" + }, + { + "bbox": [ + 55, + 445, + 295, + 649 + ], + "type": "text", + "content": " for tool locations and classes " + }, + { + "bbox": [ + 55, + 445, + 295, + 649 + ], + "type": "inline_equation", + "content": "C = \\{C_{1}, C_{2}, \\ldots, C_{N}\\}" + }, + { + "bbox": [ + 55, + 445, + 295, + 649 + ], + "type": "text", + "content": ". Tools, manipulated by operators linked to trocar ports " + }, + { + "bbox": [ + 55, + 445, + 295, + 649 + ], + "type": "inline_equation", + "content": "P = [P_{1}, P_{2}, \\ldots, P_{M}]" + }, + { + "bbox": [ + 55, + 445, + 295, + 649 + ], + "type": "text", + "content": ", are assigned unique track identities (ID) through time. The ID reassignment is guided by visual cues and clinical knowledge. Visually, it is based on class " + }, + { + "bbox": [ + 55, + 445, + 295, + 649 + ], + "type": "inline_equation", + "content": "c \\in C" + }, + { + "bbox": [ + 55, + 445, + 295, + 649 + ], + "type": "text", + "content": " and location " + }, + { + "bbox": [ + 55, + 445, + 295, + 649 + ], + "type": "inline_equation", + "content": "b \\in B" + }, + { + "bbox": [ + 55, + 445, + 295, + 649 + ], + "type": "text", + "content": ". Contextually, it considers the role and hand position of the surgeon operator linked to a unique trocar port " + }, + { + "bbox": [ + 55, + 445, + 295, + 649 + ], + "type": "inline_equation", + "content": "p \\in P" + }, + { + "bbox": [ + 55, + 445, + 295, + 649 + ], + "type": "text", + "content": ". Tool tracking solves the association matrix " + }, + { + "bbox": [ + 55, + 445, + 295, + 649 + ], + "type": "inline_equation", + "content": "A(t)" + }, + { + "bbox": [ + 55, + 445, + 295, + 649 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 445, + 295, + 649 + ], + "type": "inline_equation", + "content": "A_{i,j} = 1" + }, + { + "bbox": [ + 55, + 445, + 295, + 649 + ], + "type": "text", + "content": " indicates the " + }, + { + "bbox": [ + 55, + 445, + 295, + 649 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 445, + 295, + 649 + ], + "type": "text", + "content": "-th tool in frame " + }, + { + "bbox": [ + 55, + 445, + 295, + 649 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 445, + 295, + 649 + ], + "type": "text", + "content": " is associated with the " + }, + { + "bbox": [ + 55, + 445, + 295, + 649 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 55, + 445, + 295, + 649 + ], + "type": "text", + "content": "-th tool in frame " + }, + { + "bbox": [ + 55, + 445, + 295, + 649 + ], + "type": "inline_equation", + "content": "t + 1" + }, + { + "bbox": [ + 55, + 445, + 295, + 649 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 445, + 295, + 649 + ], + "type": "inline_equation", + "content": "A_{i,j} = 0" + }, + { + "bbox": [ + 55, + 445, + 295, + 649 + ], + "type": "text", + "content": " otherwise. The aim is to obtain tool trajectories " + }, + { + "bbox": [ + 55, + 445, + 295, + 649 + ], + "type": "inline_equation", + "content": "T = \\{T_{1}, T_{2}, \\ldots, T_{K}\\}" + }, + { + "bbox": [ + 55, + 445, + 295, + 649 + ], + "type": "text", + "content": ", each uniquely identified by ID, taking into consideration the visibility, intracorporeal, and intraoperative use." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 658, + 241, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 658, + 241, + 672 + ], + "spans": [ + { + "bbox": [ + 55, + 658, + 241, + 672 + ], + "type": "text", + "content": "3.3. Multi-Perspective (MP) Trajectory" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "Defining tool trajectories in surgical procedures necessitates a unique approach, given the variability in tracking across different perspectives, formalized as follows:" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 322, + 70, + 547, + 158 + ], + "blocks": [ + { + "bbox": [ + 322, + 70, + 547, + 158 + ], + "lines": [ + { + "bbox": [ + 322, + 70, + 547, + 158 + ], + "spans": [ + { + "bbox": [ + 322, + 70, + 547, + 158 + ], + "type": "image", + "image_path": "cdd3ad09fa90314ec31990633fc7bc5d027b5ccaf34e715ad1c0f1c6a811af47.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 332, + 166, + 535, + 177 + ], + "lines": [ + { + "bbox": [ + 332, + 166, + 535, + 177 + ], + "spans": [ + { + "bbox": [ + 332, + 166, + 535, + 177 + ], + "type": "text", + "content": "Figure 2. Multi-perspective trajectories of surgical tool." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 201, + 553, + 284 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 201, + 553, + 284 + ], + "spans": [ + { + "bbox": [ + 313, + 201, + 553, + 284 + ], + "type": "text", + "content": "1. Intraoperative trajectory: This lifelong tracking starts with a tool's first appearance and ends at its last in a patient's body during a procedure. It requires re-identification post-occlusion, out-of-camera view, or reinsertion. This approach is vital for applications such as tool usage monitoring [2], inventory management [35], surgeon training [27], skill assessment [16], and tool usage pattern analysis [52]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 285, + 554, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 285, + 554, + 416 + ], + "spans": [ + { + "bbox": [ + 313, + 285, + 554, + 416 + ], + "type": "text", + "content": "2. Intracorporeal trajectory: Unique tracks begin when a tool enters the body and end when it exits through a trocar port, even if off-camera. If a tool exits outside the camera's view, the exit is inferred if another tool enters through the same trocar or the initial tool releases its grasp. This is essential for understanding surgical workflow, as some actions occur outside the camera's focus, like graspers holding tissue out of view to facilitate other tools' actions [27]. Intracorporeal tracking supports a range of AI tasks including action evaluation [55], skill assessment [16, 27], tool usage optimization [44], and surgical risk estimation [52]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 416, + 554, + 500 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 416, + 554, + 500 + ], + "spans": [ + { + "bbox": [ + 313, + 416, + 554, + 500 + ], + "type": "text", + "content": "3. Visibility trajectory: Tracking starts with a tool's first appearance within the camera view and ends when it leaves the view. It requires re-identification (re-ID) after occlusions or brief periods of disappearance within a two-second tolerance. This method is useful for assessing surgeon actions [27] and skill training [14, 24], providing go-no-go decision support, and measuring economy of motion [48]." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 503, + 554, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 503, + 554, + 612 + ], + "spans": [ + { + "bbox": [ + 313, + 503, + 554, + 612 + ], + "type": "text", + "content": "Some existing studies [16, 37] follow the intraoperative tracking format, others [24, 46] employs visibility trajectory. The intracorporeal trajectory, being the most complex to annotate, are not well-represented in the literature. Our dataset is the first to provide fine-grained labels for all the three (Fig. 2), enabling statistical analyses such as tool usage counts, events counts, abdominal entries/ exits, tool idleness, out-of-camera view occurrences, and mean tracklets per perspective." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 622, + 416, + 634 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 622, + 416, + 634 + ], + "spans": [ + { + "bbox": [ + 313, + 622, + 416, + 634 + ], + "type": "text", + "content": "3.4. Data Annotation" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 642, + 554, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 554, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 554, + 715 + ], + "type": "text", + "content": "Annotators and tools. Bounding boxes were annotated by four researchers skilled in surgical workflow analysis, supplemented by prior study data [37, 40, 57]. Annotation tools used include Anonymized System, and a custom Python tool for visualization, merging, and validation. A pre-designed guide, refined by surgical experts, described" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8944" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 63, + 70, + 176, + 135 + ], + "blocks": [ + { + "bbox": [ + 63, + 70, + 176, + 135 + ], + "lines": [ + { + "bbox": [ + 63, + 70, + 176, + 135 + ], + "spans": [ + { + "bbox": [ + 63, + 70, + 176, + 135 + ], + "type": "image", + "image_path": "2ef49139e401ce7df0b6679663e28c05e35ea2784c443e6a2180d25bc0fe2a37.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 176, + 70, + 288, + 135 + ], + "blocks": [ + { + "bbox": [ + 176, + 70, + 288, + 135 + ], + "lines": [ + { + "bbox": [ + 176, + 70, + 288, + 135 + ], + "spans": [ + { + "bbox": [ + 176, + 70, + 288, + 135 + ], + "type": "image", + "image_path": "42a6ae560d1cd6bb2c532e6e510ca4a9867a782033eba65927042769b612b143.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 62, + 136, + 175, + 213 + ], + "blocks": [ + { + "bbox": [ + 62, + 136, + 175, + 213 + ], + "lines": [ + { + "bbox": [ + 62, + 136, + 175, + 213 + ], + "spans": [ + { + "bbox": [ + 62, + 136, + 175, + 213 + ], + "type": "image", + "image_path": "991eb9e1aca2d3fe836d87c367a7b876a7681aa7b15335098f5daa4ea7569b30.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 223, + 295, + 246 + ], + "lines": [ + { + "bbox": [ + 55, + 223, + 295, + 246 + ], + "spans": [ + { + "bbox": [ + 55, + 223, + 295, + 246 + ], + "type": "text", + "content": "Figure 3. Examples of images from CholecTrack20 tracking dataset with the labels overlaid on the raw images." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 175, + 136, + 288, + 213 + ], + "blocks": [ + { + "bbox": [ + 175, + 136, + 288, + 213 + ], + "lines": [ + { + "bbox": [ + 175, + 136, + 288, + 213 + ], + "spans": [ + { + "bbox": [ + 175, + 136, + 288, + 213 + ], + "type": "image", + "image_path": "9697a46f490c604d565d8cb985f205b53829a40ad466e3b5f57f9f1f793cd190.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 274, + 261, + 286 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 274, + 261, + 286 + ], + "spans": [ + { + "bbox": [ + 55, + 274, + 261, + 286 + ], + "type": "text", + "content": "the labels and provided image guidance as needed." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 296, + 296, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 296, + 296, + 523 + ], + "spans": [ + { + "bbox": [ + 54, + 296, + 296, + 523 + ], + "type": "text", + "content": "Label types and categories. Tool spatial positions were annotated using bounding box coordinates, while other labels were represented by class identities. Seven predominant tool categories were defined: (1) cold grasper, (2) bipolar grasper, (3) monopolar hook, (4) monopolar scissors, (5) clipper or clip applicer, (6) irrigator or suction device, and (7) specimen bag. Four operator categories were defined: (1) main surgeon left hand MSLH, (2) main surgeon right hand MSRH, (3) assistant surgeon right hand ASRH, and (4) null operator NULL. The assistant surgeon left hand ASLH holding the endoscopic camera is unreported. A total of eight visual challenges were noted: (1) blurring, (2) bleeding, (3) camera lens fouling, (4) crowded scene, (5) occlusion, (6) smoke, (7) specular light reflection, and (8) trocar view or under-coverage. The seven commonest surgical phases were annotated: (1) preparation, (2) calot triangle dissection, (3) gallbladder dissection, (4) clipping & cutting, (5) gallbladder packaging, (6) cleaning and coagulation, and (7) gallbladder extraction." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 534, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 296, + 715 + ], + "type": "text", + "content": "Annotation process. Annotations involved drawing bounding boxes " + }, + { + "bbox": [ + 55, + 534, + 296, + 715 + ], + "type": "inline_equation", + "content": "[x, y, w, h]" + }, + { + "bbox": [ + 55, + 534, + 296, + 715 + ], + "type": "text", + "content": " over tool tips tagged with tool class " + }, + { + "bbox": [ + 55, + 534, + 296, + 715 + ], + "type": "inline_equation", + "content": "c \\in C" + }, + { + "bbox": [ + 55, + 534, + 296, + 715 + ], + "type": "text", + "content": " and operator class " + }, + { + "bbox": [ + 55, + 534, + 296, + 715 + ], + "type": "inline_equation", + "content": "p \\in P" + }, + { + "bbox": [ + 55, + 534, + 296, + 715 + ], + "type": "text", + "content": ", following trocar ports for accurate surgeon identification. Surgical details such as phase, out-of-view statuses (camera/abdomen), tool entry/exit, and visual challenge attributes were annotated to aid accurate track assignment. Tool-tissue interaction labels from CholecT50 [39] provides additional help in perpetuating track identities. Annotations were reviewed at 25 FPS in uncertain cases, and underwent rigorous quality control, ensuring high-quality labels over two years. Fig. 3 presents samples of images from the CholecTrack20 dataset alongside their respective annotations illustrating the meticulous labeling system employed to ensure a rich dataset for detailed surgical tool tracking analysis." + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 317, + 69, + 553, + 194 + ], + "blocks": [ + { + "bbox": [ + 317, + 69, + 553, + 194 + ], + "lines": [ + { + "bbox": [ + 317, + 69, + 553, + 194 + ], + "spans": [ + { + "bbox": [ + 317, + 69, + 553, + 194 + ], + "type": "image", + "image_path": "727d093663a8add9db71df69c8ad56964edc3a0cf9236aecb28ebb6730f0c256.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 200, + 555, + 234 + ], + "lines": [ + { + "bbox": [ + 313, + 200, + 555, + 234 + ], + "spans": [ + { + "bbox": [ + 313, + 200, + 555, + 234 + ], + "type": "text", + "content": "Figure 4. Dataset statistics on the distributions of (a) surgical scene visual challenges across data splits (b) track labels across perspectives, averaged across videos. Track length in seconds." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 254, + 424, + 267 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 254, + 424, + 267 + ], + "spans": [ + { + "bbox": [ + 314, + 254, + 424, + 267 + ], + "type": "text", + "content": "3.5. Quality Assurance" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 274, + 555, + 321 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 274, + 555, + 321 + ], + "spans": [ + { + "bbox": [ + 313, + 274, + 555, + 321 + ], + "type": "text", + "content": "Label agreement. Two label agreement metrics validate dataset quality: Jaccard Index for spatial overlap of bounding boxes and Cohen's Kappa Statistic for category labels. The findings of our three validation forms include:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 324, + 555, + 456 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 315, + 324, + 555, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 324, + 555, + 360 + ], + "spans": [ + { + "bbox": [ + 315, + 324, + 555, + 360 + ], + "type": "text", + "content": "1. Intra-rater agreement involves self-correction. We observe a Jaccard Index of " + }, + { + "bbox": [ + 315, + 324, + 555, + 360 + ], + "type": "inline_equation", + "content": "99.4\\%" + }, + { + "bbox": [ + 315, + 324, + 555, + 360 + ], + "type": "text", + "content": " and Cohen's Kappa Scores of " + }, + { + "bbox": [ + 315, + 324, + 555, + 360 + ], + "type": "inline_equation", + "content": "94.6\\%" + }, + { + "bbox": [ + 315, + 324, + 555, + 360 + ], + "type": "text", + "content": " for tools and " + }, + { + "bbox": [ + 315, + 324, + 555, + 360 + ], + "type": "inline_equation", + "content": "94.0\\%" + }, + { + "bbox": [ + 315, + 324, + 555, + 360 + ], + "type": "text", + "content": " for operators." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 360, + 554, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 360, + 554, + 419 + ], + "spans": [ + { + "bbox": [ + 313, + 360, + 554, + 419 + ], + "type": "text", + "content": "2. Inter-rater agreement is evaluated on 20 random samples across raters. The Jaccard Index is " + }, + { + "bbox": [ + 313, + 360, + 554, + 419 + ], + "type": "inline_equation", + "content": "91.8\\%" + }, + { + "bbox": [ + 313, + 360, + 554, + 419 + ], + "type": "text", + "content": ", while tool class labels achieve " + }, + { + "bbox": [ + 313, + 360, + 554, + 419 + ], + "type": "inline_equation", + "content": "95.2\\%" + }, + { + "bbox": [ + 313, + 360, + 554, + 419 + ], + "type": "text", + "content": " and operator labels " + }, + { + "bbox": [ + 313, + 360, + 554, + 419 + ], + "type": "inline_equation", + "content": "92.7\\%" + }, + { + "bbox": [ + 313, + 360, + 554, + 419 + ], + "type": "text", + "content": " Cohen's Kappa. Minor differences reflect high-quality annotations." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 419, + 553, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 419, + 553, + 456 + ], + "spans": [ + { + "bbox": [ + 313, + 419, + 553, + 456 + ], + "type": "text", + "content": "3. Label mediation uses a board-certified surgeon for ambiguities, particularly in operator labels. Out of 758 uncertain samples, 133 needed correction post-mediation." + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 467, + 405, + 479 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 467, + 405, + 479 + ], + "spans": [ + { + "bbox": [ + 314, + 467, + 405, + 479 + ], + "type": "text", + "content": "3.6. Data Statistics" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 486, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 486, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 486, + 555, + 714 + ], + "type": "text", + "content": "Quantitative overview. The dataset includes 20 surgical videos totaling over 14 hours, averaging 42 minutes per surgery. As detailed in Tab. 1, annotations cover 35,000 frames at 1 FPS (" + }, + { + "bbox": [ + 313, + 486, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\sim" + }, + { + "bbox": [ + 313, + 486, + 555, + 714 + ], + "type": "text", + "content": "875,000 at 25 FPS), with 65,000 bounding box labels, averaging two tools per frame. Fig. 4(a) summarizes track configurations across perspectives, showing tool usage by type and trajectory. For example, graspers average 27 minutes 39 seconds inside the body and are re-inserted 8.4 times, while irrigators are used for about 1 minute 30 seconds per surgery. The visibility perspective contributes approximately 2,000 trajectories. Fig. 4(b) outlines tool attributes and visual challenges in the dataset, with occlusion as most prevalent occurring up to 23,000 seconds, bleeding totals around 18,700 seconds. This analysis provides insights into the varied visual challenges encountered in laparoscopic cholecystectomy. Evaluating tracking methods across these challenges will reveal their strengths and weaknesses. Statistical details across dataset splits are presented in Fig. 4(b)." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8945" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 69, + 91, + 541, + 188 + ], + "blocks": [ + { + "bbox": [ + 89, + 71, + 518, + 82 + ], + "lines": [ + { + "bbox": [ + 89, + 71, + 518, + 82 + ], + "spans": [ + { + "bbox": [ + 89, + 71, + 518, + 82 + ], + "type": "text", + "content": "Table 1. Dataset comparison showing the scope, statistics, and attributes. Dataset marked " + }, + { + "bbox": [ + 89, + 71, + 518, + 82 + ], + "type": "inline_equation", + "content": "\\ddagger" + }, + { + "bbox": [ + 89, + 71, + 518, + 82 + ], + "type": "text", + "content": " are not full-length videos." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 69, + 91, + 541, + 188 + ], + "lines": [ + { + "bbox": [ + 69, + 91, + 541, + 188 + ], + "spans": [ + { + "bbox": [ + 69, + 91, + 541, + 188 + ], + "type": "table", + "html": "
DatasetTaskTrack PerspectivesStatisticsAttributes
VisibilityIntraoperativeIntracorporealNo. of VideosTotal Duration (s)Frame CountTool BoxesNo. of TrajectoriesSurgoen OperatorVisual ChallengeSurgical Phase
ATLAS Dione [47]‡Detection✓XX99-22467--XXX
Cholec80-locations [49]Detection✓XX-401140116471-XXX
Bouget et.al [5]Detection✓XX14-24763819-XXX
m2cai 16-tool-locations [21]Detection✓XX--25323038-XXX
EndoVis'15 [15]‡Tracking✓XX1654013500--XXX
Fathollahi et el [16]‡TrackingX✓X1527002700--XXX
RMIT [53]Detection & Tracking✓XX415001171XXX
CholecTrack20 (Ours)Detection & Tracking✓✓✓205058135000652002,624✓✓✓
", + "image_path": "68430eaeaa23c28a3064150a877355ee8fac76a6e759a10acf156d2416158043.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 207, + 294, + 267 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 207, + 294, + 267 + ], + "spans": [ + { + "bbox": [ + 55, + 207, + 294, + 267 + ], + "type": "text", + "content": "Dataset comparison. Existing publicly available datasets primarily focus on single perspective trajectory [16, 46]. In contrast, our dataset introduces a novel approach by annotation different perspectives, as presented in Tab. 1, including visual challenges, phase details, activity labels, etc." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 270, + 296, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 270, + 296, + 331 + ], + "spans": [ + { + "bbox": [ + 55, + 270, + 296, + 331 + ], + "type": "text", + "content": "Dataset split. The dataset is split at the procedure level into non-overlapping training, validation, and testing sets in a 5:1:4 ratio, preventing data leakage. Video distribution is balanced by procedure duration to ensure similar complexity and difficulty levels across all splits." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 338, + 144, + 350 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 338, + 144, + 350 + ], + "spans": [ + { + "bbox": [ + 55, + 338, + 144, + 350 + ], + "type": "text", + "content": "3.7. Data Analysis" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 357, + 296, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 357, + 296, + 429 + ], + "spans": [ + { + "bbox": [ + 55, + 357, + 296, + 429 + ], + "type": "text", + "content": "We conduct a comprehensive analysis of the dataset to gain insights into label alignments and feature similarities, revealing correlations across labels. This would guide data preprocessing and feature selection when using the data. Our analysis as illustrated in Fig. 5, Fig. 6, and Fig. 7 encompasses four distinct dimensions and discussed further." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 432, + 296, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 432, + 296, + 529 + ], + "spans": [ + { + "bbox": [ + 55, + 432, + 296, + 529 + ], + "type": "text", + "content": "Tracking vs. surgical tool type correlation. This analysis explores the relationship between different tool types and their tracks, providing insights into unique patterns associated with each tool during surgery. Fig. 5(a) illustrates center point locations of tools, color-coded by category over time in three videos. This shows that while some trajectories appear separable, they are mostly densely interwoven, suggesting the need for advanced modeling in tool tracking." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 531, + 296, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 531, + 296, + 615 + ], + "spans": [ + { + "bbox": [ + 55, + 531, + 296, + 615 + ], + "type": "text", + "content": "Tracking vs. surgeon tool operator correlation. We analyzed the alignment between tool operator identities and tool locations, revealing dynamic interactions between surgeons and instruments. Fig. 5(b) shows a strong correlation between operators and their tools, attributed to the distinct positioning of trocars. This underscores the value of operator information in accurate track label generation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 618, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 296, + 715 + ], + "type": "text", + "content": "Tracking vs. surgical phase segmentation. This analysis delves into tracking tools across surgical phases, uncovering how tool utilization varies with procedural stages. Fig. 6 shows that complex phases, such as calot triangle and gallbladder dissection, exhibit densely packed trajectories due to prolonged duration and frequent tool manipulation. Simpler phases, like preparation and clipping, feature fewer trajectories, facilitating modeling. These insights inform deep" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 322, + 206, + 394, + 265 + ], + "blocks": [ + { + "bbox": [ + 322, + 206, + 394, + 265 + ], + "lines": [ + { + "bbox": [ + 322, + 206, + 394, + 265 + ], + "spans": [ + { + "bbox": [ + 322, + 206, + 394, + 265 + ], + "type": "image", + "image_path": "2704ea2eedc13fe4f82ba0dfcccc32e912588f04c97168c2a9529f1b83b00a1f.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 395, + 207, + 467, + 264 + ], + "blocks": [ + { + "bbox": [ + 395, + 207, + 467, + 264 + ], + "lines": [ + { + "bbox": [ + 395, + 207, + 467, + 264 + ], + "spans": [ + { + "bbox": [ + 395, + 207, + 467, + 264 + ], + "type": "image", + "image_path": "e30acdeb9ffaf6458871f74f2c78a5a43eff1b0046104beac080fc5453784d20.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 468, + 207, + 547, + 264 + ], + "blocks": [ + { + "bbox": [ + 468, + 207, + 547, + 264 + ], + "lines": [ + { + "bbox": [ + 468, + 207, + 547, + 264 + ], + "spans": [ + { + "bbox": [ + 468, + 207, + 547, + 264 + ], + "type": "image", + "image_path": "df30a378ce2b2f618497745f07be265f0d2b8362c6a0be190fabad607ad066bb.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 321, + 270, + 394, + 327 + ], + "blocks": [ + { + "bbox": [ + 321, + 270, + 394, + 327 + ], + "lines": [ + { + "bbox": [ + 321, + 270, + 394, + 327 + ], + "spans": [ + { + "bbox": [ + 321, + 270, + 394, + 327 + ], + "type": "image", + "image_path": "fd20cb4ebc8019ab066faa0e1295b19af1b36f0fb629cf34a62b3ed5c48672ea.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 338, + 555, + 371 + ], + "lines": [ + { + "bbox": [ + 313, + 338, + 555, + 371 + ], + "spans": [ + { + "bbox": [ + 313, + 338, + 555, + 371 + ], + "type": "text", + "content": "Figure 5. 3D visualization of label alignments showing the tool position over track time. The coloring is for grouping features according to: (a) tool classes and (b) tool operators." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 395, + 271, + 468, + 327 + ], + "blocks": [ + { + "bbox": [ + 395, + 271, + 468, + 327 + ], + "lines": [ + { + "bbox": [ + 395, + 271, + 468, + 327 + ], + "spans": [ + { + "bbox": [ + 395, + 271, + 468, + 327 + ], + "type": "image", + "image_path": "eec379849419c74f4da22e6f8c89cdb6018b0fefe8e20085b144472b9aebb9e6.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 470, + 271, + 547, + 327 + ], + "blocks": [ + { + "bbox": [ + 470, + 271, + 547, + 327 + ], + "lines": [ + { + "bbox": [ + 470, + 271, + 547, + 327 + ], + "spans": [ + { + "bbox": [ + 470, + 271, + 547, + 327 + ], + "type": "image", + "image_path": "42c63b944bf9a009018462133333d698939af8d13cad7e78bee608c7b0e93aa3.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 322, + 379, + 395, + 440 + ], + "blocks": [ + { + "bbox": [ + 322, + 379, + 395, + 440 + ], + "lines": [ + { + "bbox": [ + 322, + 379, + 395, + 440 + ], + "spans": [ + { + "bbox": [ + 322, + 379, + 395, + 440 + ], + "type": "image", + "image_path": "3d86bcf08c4bd17bf3d61267a52a4078b20e098d3b11afc1020733986c438788.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 399, + 379, + 471, + 440 + ], + "blocks": [ + { + "bbox": [ + 399, + 379, + 471, + 440 + ], + "lines": [ + { + "bbox": [ + 399, + 379, + 471, + 440 + ], + "spans": [ + { + "bbox": [ + 399, + 379, + 471, + 440 + ], + "type": "image", + "image_path": "39700c595b61332c0e958a66956dfafc5292bc11bb732eb967609c2eeabe7b51.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 474, + 379, + 547, + 440 + ], + "blocks": [ + { + "bbox": [ + 474, + 379, + 547, + 440 + ], + "lines": [ + { + "bbox": [ + 474, + 379, + 547, + 440 + ], + "spans": [ + { + "bbox": [ + 474, + 379, + 547, + 440 + ], + "type": "image", + "image_path": "1fa3ae87e774469ab1b5874eff703d188649e6145e2a2c66d6d5efdb4f5a2fe9.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 323, + 443, + 397, + 505 + ], + "blocks": [ + { + "bbox": [ + 323, + 443, + 397, + 505 + ], + "lines": [ + { + "bbox": [ + 323, + 443, + 397, + 505 + ], + "spans": [ + { + "bbox": [ + 323, + 443, + 397, + 505 + ], + "type": "image", + "image_path": "165faeef204e8994e051ad323cff3aa62bc454ac87e905eda49aa79ca01436ed.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 513, + 555, + 536 + ], + "lines": [ + { + "bbox": [ + 313, + 513, + 555, + 536 + ], + "spans": [ + { + "bbox": [ + 313, + 513, + 555, + 536 + ], + "type": "text", + "content": "Figure 6. 3D visualization of tracking across different surgical phases for some randomly selected videos." + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 399, + 443, + 470, + 505 + ], + "blocks": [ + { + "bbox": [ + 399, + 443, + 470, + 505 + ], + "lines": [ + { + "bbox": [ + 399, + 443, + 470, + 505 + ], + "spans": [ + { + "bbox": [ + 399, + 443, + 470, + 505 + ], + "type": "image", + "image_path": "fc8efa217cf617101faa9213f4fe7f70bdf65fba0f9c4c0d439170fed131677e.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 474, + 443, + 547, + 505 + ], + "blocks": [ + { + "bbox": [ + 474, + 443, + 547, + 505 + ], + "lines": [ + { + "bbox": [ + 474, + 443, + 547, + 505 + ], + "spans": [ + { + "bbox": [ + 474, + 443, + 547, + 505 + ], + "type": "image", + "image_path": "82b7c4f8969ef70591d9fedc36a2984035c139e7c53df56f99af599f5b7306ed.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 321, + 543, + 394, + 598 + ], + "blocks": [ + { + "bbox": [ + 321, + 543, + 394, + 598 + ], + "lines": [ + { + "bbox": [ + 321, + 543, + 394, + 598 + ], + "spans": [ + { + "bbox": [ + 321, + 543, + 394, + 598 + ], + "type": "image", + "image_path": "d05cce21a7f9088eb75d4494392e2ac5866343b6c06e829dd50182d8a005d3b4.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 399, + 543, + 471, + 599 + ], + "blocks": [ + { + "bbox": [ + 399, + 543, + 471, + 599 + ], + "lines": [ + { + "bbox": [ + 399, + 543, + 471, + 599 + ], + "spans": [ + { + "bbox": [ + 399, + 543, + 471, + 599 + ], + "type": "image", + "image_path": "4c67c568558578949091197344d85ce42020c8a7418f8c5edeadb1cbf6a61aa2.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 474, + 543, + 547, + 599 + ], + "blocks": [ + { + "bbox": [ + 474, + 543, + 547, + 599 + ], + "lines": [ + { + "bbox": [ + 474, + 543, + 547, + 599 + ], + "spans": [ + { + "bbox": [ + 474, + 543, + 547, + 599 + ], + "type": "image", + "image_path": "a135a11defd19d4258a115911fe7da6820e18c2864f13af8a5ef4d0c931e2ae0.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 321, + 602, + 394, + 656 + ], + "blocks": [ + { + "bbox": [ + 321, + 602, + 394, + 656 + ], + "lines": [ + { + "bbox": [ + 321, + 602, + 394, + 656 + ], + "spans": [ + { + "bbox": [ + 321, + 602, + 394, + 656 + ], + "type": "image", + "image_path": "27506cc6dc91c6f54b713a43bc59c9764373f1ea8d3ba19cdc98a2d18fa2b2ce.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 677, + 555, + 711 + ], + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 711 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 711 + ], + "type": "text", + "content": "Figure 7. Trajectories of selected tools with EMA over time. Plotted positions are computed as weighted combinations of the center coordinate of the bounding boxes, scaled by image size." + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 399, + 602, + 471, + 656 + ], + "blocks": [ + { + "bbox": [ + 399, + 602, + 471, + 656 + ], + "lines": [ + { + "bbox": [ + 399, + 602, + 471, + 656 + ], + "spans": [ + { + "bbox": [ + 399, + 602, + 471, + 656 + ], + "type": "image", + "image_path": "ae6516bec1d140b08e62fc15fed07bc071490342df98d577929bce0c3d209000.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 474, + 602, + 547, + 656 + ], + "blocks": [ + { + "bbox": [ + 474, + 602, + 547, + 656 + ], + "lines": [ + { + "bbox": [ + 474, + 602, + 547, + 656 + ], + "spans": [ + { + "bbox": [ + 474, + 602, + 547, + 656 + ], + "type": "image", + "image_path": "6ee3105fd554f60069040c5b19ed3184e560172ae9323787f64439938ae422b1.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 388, + 657, + 473, + 669 + ], + "blocks": [ + { + "bbox": [ + 388, + 657, + 473, + 669 + ], + "lines": [ + { + "bbox": [ + 388, + 657, + 473, + 669 + ], + "spans": [ + { + "bbox": [ + 388, + 657, + 473, + 669 + ], + "type": "image", + "image_path": "00b52bd8d89ca5fca38d70292519e9e3ef5ced20bcbb9d81a2838e3419fbf648.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + } + ], + "index": 29 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "8946" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 91, + 553, + 209 + ], + "blocks": [ + { + "bbox": [ + 138, + 71, + 470, + 82 + ], + "lines": [ + { + "bbox": [ + 138, + 71, + 470, + 82 + ], + "spans": [ + { + "bbox": [ + 138, + 71, + 470, + 82 + ], + "type": "text", + "content": "Table 2. Benchmark Results of SOTA Object Detectors on Surgical Tool Detection Dataset." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 91, + 553, + 209 + ], + "lines": [ + { + "bbox": [ + 56, + 91, + 553, + 209 + ], + "spans": [ + { + "bbox": [ + 56, + 91, + 553, + 209 + ], + "type": "table", + "html": "
Detector \nModelDetection AP accross 3 thresholdsDetection AP per category. (% AP @ Θ = 0.5)Detection AP across surgical visual challengesSpeed
AP0.5↑AP0.75↑AP0.5:0.95↑GrasperBipolarHookScissorsClipperIrrigatorBagBleedingBlurSmokeCrowdedOccludedReflectionFoul LensTrocarFPS↑
Faster-RCNN [43]56.038.134.653.565.080.160.970.126.831.857.941.054.543.555.046.941.235.77.6
Cascade-RCNN [7]51.739.034.752.058.979.745.744.923.717.953.939.048.139.546.429.133.733.77.0
CenterNet [69]53.039.535.060.261.486.456.368.025.810.258.042.150.236.751.746.035.830.833.8
FCOS [54]43.531.528.151.244.374.749.254.221.97.247.840.651.515.140.842.729.717.67.7
SSD [30]61.937.836.175.262.291.663.472.922.540.864.549.358.357.562.453.947.742.630.9
PAA [23]64.544.941.169.679.089.268.774.237.628.967.155.665.055.064.656.051.247.58.5
Def-DETR [71]58.442.038.360.666.583.861.972.039.923.862.442.758.637.157.443.941.547.410.2
Swin-T [31]62.344.340.263.364.883.080.277.238.026.863.553.862.835.361.166.255.245.79.8
YOLOX [19]64.748.944.269.672.289.475.479.137.327.168.255.666.045.964.252.558.143.123.6
YOLOv7 [62]80.662.056.190.586.496.082.389.349.166.280.261.280.179.582.165.671.266.720.6
YOLOv8 [58]79.162.455.687.984.596.280.087.248.465.077.158.374.476.280.470.357.462.929.0
YOLOv9 [61]80.262.656.588.587.696.079.387.150.167.778.154.078.278.681.165.363.463.123.7
YOLOv10 [60]80.162.155.887.686.696.081.989.053.861.377.861.978.777.581.266.759.365.428.6
", + "image_path": "6d30ec45b110a782dae3335563cabadbb3e9d83d3691a539d4e8b3e80a88012c.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 228, + 294, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 228, + 294, + 252 + ], + "spans": [ + { + "bbox": [ + 55, + 228, + 294, + 252 + ], + "type": "text", + "content": "learning model assessment and AI model development for surgical tool tracking." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 255, + 295, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 255, + 295, + 361 + ], + "spans": [ + { + "bbox": [ + 55, + 255, + 295, + 361 + ], + "type": "text", + "content": "Tracking variance. Using Exponential Moving Average (EMA), we analyze tool tracking data, as shown in Fig. 7. By overlaying EMAs on tool trajectories, we visualize variance between actual trajectories and modeled tracks. This approach highlights high-variance frames, serving as challenging cases for benchmarking model robustness and accuracy. Through this analysis, we provide a valuable resource for researchers, emphasizing the need to focus on complex scenarios for a rigorous assessment of model capabilities." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 372, + 221, + 385 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 372, + 221, + 385 + ], + "spans": [ + { + "bbox": [ + 55, + 372, + 221, + 385 + ], + "type": "text", + "content": "4. Benchmark and Experiments" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 391, + 146, + 403 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 391, + 146, + 403 + ], + "spans": [ + { + "bbox": [ + 55, + 391, + 146, + 403 + ], + "type": "text", + "content": "4.1. Tool Detection" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 410, + 295, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 410, + 295, + 589 + ], + "spans": [ + { + "bbox": [ + 55, + 410, + 295, + 589 + ], + "type": "text", + "content": "Models. Owing that tool detection is a fundamental part of tool tracking, we showcase the usability of the CholecTrack20 dataset for this task, by benchmarking several object detectors representing diverse methodologies. Faster-RCNN [43] and Cascade-RCNN [7] are anchor-based models, with Cascade-RCNN employing a multi-stage approach to refine detection accuracy. CenterNet [69] and FCOS [54] are anchor-free models, utilizing center points and direct bounding box regression for efficient detection. SSD [30] provides real-time performance with its multi-scale approach. Deformable-DETR [71] applies a transformer-based method for flexible feature processing, while Swin-T [31] uses hierarchical transformers with shifted windows. The YOLO models [19, 58, 60–62] feature advanced multiscale strategies for high accuracy and speed." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 591, + 295, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 591, + 295, + 663 + ], + "spans": [ + { + "bbox": [ + 55, + 591, + 295, + 663 + ], + "type": "text", + "content": "Evaluation metrics. We evaluate tool detection using COCO standard average precision (AP) metrics (pycocotools, not ultralytics) across several thresholds, categories, and visual challenges. We also report model inference speed in frames per seconds (FPS) on a single NVIDIA GTX 1080 Ti (10 GB) GPU." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "Overall tool detection results. The detection performance of the models is summarized in Tab. 2. YOLOv7 [62] achieves the highest Average Precision (AP) of " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "56.1\\%" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": ", surpassing other models. YOLOX [19] follows with an AP" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 228, + 553, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 228, + 553, + 372 + ], + "spans": [ + { + "bbox": [ + 313, + 228, + 553, + 372 + ], + "type": "text", + "content": "of " + }, + { + "bbox": [ + 313, + 228, + 553, + 372 + ], + "type": "inline_equation", + "content": "44.2\\%" + }, + { + "bbox": [ + 313, + 228, + 553, + 372 + ], + "type": "text", + "content": ". Deformable-DETR [71] and Swin-T [31] show competitive results with APs of " + }, + { + "bbox": [ + 313, + 228, + 553, + 372 + ], + "type": "inline_equation", + "content": "38.3\\%" + }, + { + "bbox": [ + 313, + 228, + 553, + 372 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 228, + 553, + 372 + ], + "type": "inline_equation", + "content": "40.2\\%" + }, + { + "bbox": [ + 313, + 228, + 553, + 372 + ], + "type": "text", + "content": ", respectively. The anchor-free models such as CenterNet [69] and FCOS [54] demonstrate robust performance, with CenterNet achieving an AP of " + }, + { + "bbox": [ + 313, + 228, + 553, + 372 + ], + "type": "inline_equation", + "content": "35.0\\%" + }, + { + "bbox": [ + 313, + 228, + 553, + 372 + ], + "type": "text", + "content": " and FCOS " + }, + { + "bbox": [ + 313, + 228, + 553, + 372 + ], + "type": "inline_equation", + "content": "28.1\\%" + }, + { + "bbox": [ + 313, + 228, + 553, + 372 + ], + "type": "text", + "content": ". In comparison, Faster-RCNN [43] and Cascade-RCNN [7] deliver APs of " + }, + { + "bbox": [ + 313, + 228, + 553, + 372 + ], + "type": "inline_equation", + "content": "34.6\\%" + }, + { + "bbox": [ + 313, + 228, + 553, + 372 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 228, + 553, + 372 + ], + "type": "inline_equation", + "content": "34.7\\%" + }, + { + "bbox": [ + 313, + 228, + 553, + 372 + ], + "type": "text", + "content": ", respectively, showcasing their efficacy with anchor-based approaches. At IoU thresholds of 0.5 and 0.75, YOLOv7 leads with scores of " + }, + { + "bbox": [ + 313, + 228, + 553, + 372 + ], + "type": "inline_equation", + "content": "80.6\\%" + }, + { + "bbox": [ + 313, + 228, + 553, + 372 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 228, + 553, + 372 + ], + "type": "inline_equation", + "content": "62.0\\%" + }, + { + "bbox": [ + 313, + 228, + 553, + 372 + ], + "type": "text", + "content": ", respectively. In terms of inference speed, YOLO networks excel with real-time capacities exceeding 20 FPS, with CenterNet achieving the highest speed at 33.8 FPS." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 374, + 553, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 374, + 553, + 495 + ], + "spans": [ + { + "bbox": [ + 313, + 374, + 553, + 495 + ], + "type": "text", + "content": "Class-wise detection results. Analyzing tool detection results per category (Tab. 2), YOLOv7 emerges as the top performer, dominating in all the 7 categories, achieving above " + }, + { + "bbox": [ + 313, + 374, + 553, + 495 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 313, + 374, + 553, + 495 + ], + "type": "text", + "content": " accuracy in 2 tool categories and above " + }, + { + "bbox": [ + 313, + 374, + 553, + 495 + ], + "type": "inline_equation", + "content": "80\\%" + }, + { + "bbox": [ + 313, + 374, + 553, + 495 + ], + "type": "text", + "content": " in 5. Notably, the hook exhibits the highest tendency among tools, with AP scores ranging from " + }, + { + "bbox": [ + 313, + 374, + 553, + 495 + ], + "type": "inline_equation", + "content": "74.7\\%" + }, + { + "bbox": [ + 313, + 374, + 553, + 495 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 374, + 553, + 495 + ], + "type": "inline_equation", + "content": "96.0\\%" + }, + { + "bbox": [ + 313, + 374, + 553, + 495 + ], + "type": "text", + "content": " across all models. Conversely, irrigator and specimen bag pose challenges, likely due to unclear tool tip boundaries and the bag's deformable nature, respectively. Grasper, bipolar, scissors, and clipper show high detection rates." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 496, + 553, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 496, + 553, + 604 + ], + "spans": [ + { + "bbox": [ + 313, + 496, + 553, + 604 + ], + "type": "text", + "content": "Detection under visual challenges. Tab. 2 presents the performance of the benchmark object detection models across different surgical visual challenges. Notably, YOLOv7 achieves the highest detection accuracy across most challenges, particularly excelling in scenarios involving bleeding, smoke, and crowded scenes. Conversely, detecting tools in blurred scenes, near trocars, and specular light reflection pose significant challenges for all models, with lower detection rates observed across the board." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 611, + 403, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 611, + 403, + 624 + ], + "spans": [ + { + "bbox": [ + 313, + 611, + 403, + 624 + ], + "type": "text", + "content": "4.2. Tool Tracking" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 629, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 553, + 713 + ], + "type": "text", + "content": "Models. We train and evaluate several state-of-the-art multi-object tracking (MOT) methods on the CholecTrack20 dataset, focusing on their ability to track surgical tools. OCSORT [33] and TransTrack [50] employ sophisticated tracking-by-detection frameworks, with TransTrack utilizing transformers to improve feature association. ByteTrack [67], Bot-SORT [1], and SMILETrack [63] use ad" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8947" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 67, + 91, + 541, + 340 + ], + "blocks": [ + { + "bbox": [ + 161, + 70, + 448, + 82 + ], + "lines": [ + { + "bbox": [ + 161, + 70, + 448, + 82 + ], + "spans": [ + { + "bbox": [ + 161, + 70, + 448, + 82 + ], + "type": "text", + "content": "Table 3. Benchmark Multi-Perspective Multi-Tool Tracking Results @ 25 FPS." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 67, + 91, + 541, + 340 + ], + "lines": [ + { + "bbox": [ + 67, + 91, + 541, + 340 + ], + "spans": [ + { + "bbox": [ + 67, + 91, + 541, + 340 + ], + "type": "table", + "html": "
ModelHOTA MetricsCLEAR MetricsIdentity MetricsCount MetricsSpeed
TrackerHOTA↑DetA↑LocA↑AssA↑MOTA↑MOTP↑MT↑PT↓ML↓IDF1↑IDSW↓Frag↓#Dets#IDsFPS↑
Intraoperative Trajectory (Groundtruth counts: #Dets = 29994, #IDs = 70)
OCSORT [33]14.652.786.74.149.285.02432149.52921273121936333610.2
FairMOT [66]5.825.875.91.35.073.9324434.34227192415252445614.2
TransTrack [50]7.431.584.41.74.282.9936254.2475718992164040796.7
ByteTrack [67]15.870.685.73.667.084.0541229.54648242928440538316.4
Bot-SORT [1]17.470.785.44.469.683.75811110.2390723762930245018.7
SMILETrack [63]15.971.085.53.766.483.8551329.24968236928821576111.2
Intracorporeal Trajectory (Groundtruth counts: #Dets = 29994, #IDs = 247)
OCSORT [33]23.751.486.511.047.184.8115874518.12953279621797352610.2
FairMOT [66]7.519.776.12.95.474.019601686.02890149611287396214.2
TransTrack [50]13.131.584.45.54.682.98079888.7464817912164040796.7
ByteTrack [67]24.770.685.78.767.484.0176482316.94515229028440538316.4
Bot-SORT [1]27.070.785.410.470.083.7188382118.9377122382930045018.7
SMILETrack [63]24.966.785.58.966.783.8186392216.94868223228820577911.2
Visibility Trajectory (Groundtruth counts: #Dets = 29994, #IDs = 916)
SORT [4]17.439.585.27.821.483.313939937813.46619213816595884419.5
OCSORT [33]37.052.686.526.250.284.830037124535.92317226022197358710.2
FairMOT [66]15.325.075.89.57.173.75821864014.43140157415338487514.2
TransTrack [50]19.231.684.411.85.882.922428041216.1427314032164040796.7
ByteTrack [67]41.570.785.724.869.384.059121710836.83930170428440538316.4
Bot-SORT [1]44.770.885.528.772.083.76381849441.4318316382930045058.7
SMILETrack [63]41.371.085.624.468.983.861919210536.54227164128821575211.2
", + "image_path": "d34c8182af67dcccc990ee8c6782272582603e2bd9e5ac8d3e5ba2f19026685b.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 361, + 294, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 361, + 294, + 384 + ], + "spans": [ + { + "bbox": [ + 55, + 361, + 294, + 384 + ], + "type": "text", + "content": "vanced tracking algorithms, with Bot-SORT and SMILE-Track incorporating extra modules for enhanced robustness." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 384, + 294, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 384, + 294, + 469 + ], + "spans": [ + { + "bbox": [ + 55, + 384, + 294, + 469 + ], + "type": "text", + "content": "Evaluation metrics. We assess benchmark models on variety of tracking metrics: higher-order tracking accuracy (HOTA) [32], CLEAR MOT metrics [3], identity metrics [45], counting metrics, and tracking speed. A pull request is made to the standard TrackEval [22] library integrating CholecTrack20 benchmark with all its exhaustive performance evaluation protocols recommended by this study." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 471, + 294, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 471, + 294, + 567 + ], + "spans": [ + { + "bbox": [ + 55, + 471, + 294, + 567 + ], + "type": "text", + "content": "Multi-object tracking results. The performance of the evaluated tracking methods is summarized in Tab. 3. Models such as FairMOT and TransTrack show the lowest HOTA score of " + }, + { + "bbox": [ + 55, + 471, + 294, + 567 + ], + "type": "inline_equation", + "content": "5.8\\%" + }, + { + "bbox": [ + 55, + 471, + 294, + 567 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 471, + 294, + 567 + ], + "type": "inline_equation", + "content": "7.4\\%" + }, + { + "bbox": [ + 55, + 471, + 294, + 567 + ], + "type": "text", + "content": " respectively, highlighting challenges in maintaining tool identities over time. ByteTrack, Bot-SORT, and SMILETrack achieve higher HOTA scores, ranging from " + }, + { + "bbox": [ + 55, + 471, + 294, + 567 + ], + "type": "inline_equation", + "content": "15.7\\%" + }, + { + "bbox": [ + 55, + 471, + 294, + 567 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 471, + 294, + 567 + ], + "type": "inline_equation", + "content": "17.4\\%" + }, + { + "bbox": [ + 55, + 471, + 294, + 567 + ], + "type": "text", + "content": ", but still face difficulties in tool re-ID due to similarities among tools." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 567, + 294, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 567, + 294, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 567, + 294, + 628 + ], + "type": "text", + "content": "Despite these advancements, there remains room for enhancement in identity association and tracking precision. The results also include metrics on the detection counts, unique identities assigned, and tracking speed, providing a comprehensive view of each method's performance." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 629, + 294, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 294, + 713 + ], + "type": "text", + "content": "Multi-perspective tracking results. Looking at the different trajectory perspectives, Tab. 3 shows that visibility tracking is the easiest with most of the existing models showcasing their strengths. This is expected because deep learning models mostly rely on visual cues, which are captured by camera in the visibility track scenario. Here, Bot-SORT record a landslide top performance scores of " + }, + { + "bbox": [ + 55, + 629, + 294, + 713 + ], + "type": "inline_equation", + "content": "44.7\\%" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 361, + 553, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 361, + 553, + 540 + ], + "spans": [ + { + "bbox": [ + 313, + 361, + 553, + 540 + ], + "type": "text", + "content": "HOTA, " + }, + { + "bbox": [ + 313, + 361, + 553, + 540 + ], + "type": "inline_equation", + "content": "72.0\\%" + }, + { + "bbox": [ + 313, + 361, + 553, + 540 + ], + "type": "text", + "content": " MOTA, and " + }, + { + "bbox": [ + 313, + 361, + 553, + 540 + ], + "type": "inline_equation", + "content": "41.4\\%" + }, + { + "bbox": [ + 313, + 361, + 553, + 540 + ], + "type": "text", + "content": " IDF1. The intracorporeal tracking is the most challenging since the major factors marking the entry and exit of the tools from the body are not readily visible. A maximum of " + }, + { + "bbox": [ + 313, + 361, + 553, + 540 + ], + "type": "inline_equation", + "content": "27.0\\%" + }, + { + "bbox": [ + 313, + 361, + 553, + 540 + ], + "type": "text", + "content": " HOTA suggest a decline on the leading Bot-SORT. New methods could leverage rich fine-grained history to estimate the out-of-view and out-of-body status of the tools for improve re-ID. The intraoperative trajectory comes in the middle in terms of difficulty. While it may be challenging to ascertain the persistence of a trajectory after re-insertion, the class features are also helpful especially for tools of different categories. Again the Bot-SORT, leverage camera compensation details, shows a better tendency of estimating the persistent identity of different tools of the same class with a " + }, + { + "bbox": [ + 313, + 361, + 553, + 540 + ], + "type": "inline_equation", + "content": "+1.5\\%" + }, + { + "bbox": [ + 313, + 361, + 553, + 540 + ], + "type": "text", + "content": " HOTA higher than similarity and appearance features." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 543, + 553, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 543, + 553, + 687 + ], + "spans": [ + { + "bbox": [ + 313, + 543, + 553, + 687 + ], + "type": "text", + "content": "Multi-class tracking results. In Fig. 8, we analyze tracking performance by tool class and observe that the grasper, despite having the most instances and being the most frequently used tool in the dataset, achieves the highest tracking accuracy across perspectives. Class-agnostic results reveal medium tracking accuracy for other commonly used tools like the bipolar, hook, and clipper, while rarely used tools (e.g., scissors, irrigator) have lower scores. Specimen bag tracking is affected by shape deformation, contents, states (open/closed, empty/filled), tool interactions, and fluid stains. Tracking surgical tools beyond the visual perspective remains challenging for all models tested." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "content": "Tracking results under visual challenges. We evaluate tool tracking performance across various surgical condi" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8948" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 69, + 294, + 204 + ], + "blocks": [ + { + "bbox": [ + 57, + 69, + 294, + 204 + ], + "lines": [ + { + "bbox": [ + 57, + 69, + 294, + 204 + ], + "spans": [ + { + "bbox": [ + 57, + 69, + 294, + 204 + ], + "type": "image", + "image_path": "6c0ebe6b0fbb1a8b8a189ff7f167a91470dbae44ff502f32292b612acdc1b78f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 211, + 295, + 233 + ], + "lines": [ + { + "bbox": [ + 55, + 211, + 295, + 233 + ], + "spans": [ + { + "bbox": [ + 55, + 211, + 295, + 233 + ], + "type": "text", + "content": "Figure 8. Results of multi-perspective tracking across seven surgical tool categories in the CholecTrack20 dataset." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 56, + 244, + 296, + 380 + ], + "blocks": [ + { + "bbox": [ + 56, + 244, + 296, + 380 + ], + "lines": [ + { + "bbox": [ + 56, + 244, + 296, + 380 + ], + "spans": [ + { + "bbox": [ + 56, + 244, + 296, + 380 + ], + "type": "image", + "image_path": "0bd0e055ac20de3fea600d3d4020d0c43922dd460bff00bd65b5230a9fa8769a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 387, + 295, + 409 + ], + "lines": [ + { + "bbox": [ + 55, + 387, + 295, + 409 + ], + "spans": [ + { + "bbox": [ + 55, + 387, + 295, + 409 + ], + "type": "text", + "content": "Figure 9. Results of multi-perspective tracking across eight surgical visual challenges in the CholecTrack20 dataset." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 56, + 419, + 296, + 562 + ], + "blocks": [ + { + "bbox": [ + 56, + 419, + 296, + 562 + ], + "lines": [ + { + "bbox": [ + 56, + 419, + 296, + 562 + ], + "spans": [ + { + "bbox": [ + 56, + 419, + 296, + 562 + ], + "type": "image", + "image_path": "9c4a38df3ba2bf363b6978e0a35c9a44d5ee5beb1a89866fe74e773c32829407.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 569, + 295, + 592 + ], + "lines": [ + { + "bbox": [ + 55, + 569, + 295, + 592 + ], + "spans": [ + { + "bbox": [ + 55, + 569, + 295, + 592 + ], + "type": "text", + "content": "Figure 10. Results of multi-perspective tracking across seven surgical phases in the CholecTrack20 dataset." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 613, + 295, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 613, + 295, + 686 + ], + "spans": [ + { + "bbox": [ + 55, + 613, + 295, + 686 + ], + "type": "text", + "content": "tions using HOTA metrics in Fig. 9, providing insights into model interactions with complex surgical environments. The model performs well under blurring, reflections, and limited camera coverage, likely due to effective data augmentation. However, lens fouling, smoke, and occlusion present significant challenges, reducing accuracy." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": "Tracking results by surgical phase. Fig. 10 shows performance across seven surgical phases, with the clipping and" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "content": "cutting phases proving easiest to track due to limited activities and a linear progression. Preparation phase shows similar performance. Phases like Calot triangle dissection and gallbladder dissection exhibit comparable tracking results, while gallbladder packaging shows the most consistent tracking across perspectives. Overall, OC-SORT struggles the most, while Bot-SORT achieves the best result." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 162, + 492, + 175 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 162, + 492, + 175 + ], + "spans": [ + { + "bbox": [ + 313, + 162, + 492, + 175 + ], + "type": "text", + "content": "4.3. Limitations and Gaps to Address" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 180, + 555, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 180, + 555, + 335 + ], + "spans": [ + { + "bbox": [ + 313, + 180, + 555, + 335 + ], + "type": "text", + "content": "The SOTA tracking methods trained on CholecTrack20 reveal substantial limitations, with performance under " + }, + { + "bbox": [ + 313, + 180, + 555, + 335 + ], + "type": "inline_equation", + "content": "45\\%" + }, + { + "bbox": [ + 313, + 180, + 555, + 335 + ], + "type": "text", + "content": " HOTA, which is insufficient for clinical translation. These models struggle with various visual challenges, such as smoke, bleeding, and specular light reflection, affecting detection and re-ID. Since location and appearance features alone are inadequate, especially for tools with similar appearances, this highlights the need to move beyond current cues and innovate more intuitive, context-aware methods for re-ID. CholecTrack20 serves as a critical foundation for exploring this direction, offering a dataset rich in diverse tracking perspectives and challenges, essential for developing more robust and clinically viable tracking solutions." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 345, + 388, + 357 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 345, + 388, + 357 + ], + "spans": [ + { + "bbox": [ + 313, + 345, + 388, + 357 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 364, + 555, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 364, + 555, + 592 + ], + "spans": [ + { + "bbox": [ + 313, + 364, + 555, + 592 + ], + "type": "text", + "content": "In this work, we presented the CholecTrack20 dataset, a novel resource designed to advance the state-of-the-art in surgical tool tracking within computer vision. CholecTrack20 addresses a critical gap by providing comprehensive annotations and diverse tracking scenarios and tasks across various surgical phases and visual challenges. Key innovations of CholecTrack20 include multi-perspective tracking, which defines the start and termination of a tool track differently based on visibility, intracorporeal, or intraoperative contexts. It also features detailed annotations of surgical visual challenges and precise surgical phase segmentation. Our extensive benchmark experiments demonstrate the dataset's effectiveness in developing models for tool detection and multi-object tracking across these three distinct trajectory perspectives. Through evaluating several deep learning methodologies on the CholecTrack20 dataset, we gain insights into their strengths and weaknesses in handling multiple viewpoints or tracking perspectives, across surgical phases and diverse surgical scene visual challenges." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 592, + 556, + 711 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 592, + 556, + 711 + ], + "spans": [ + { + "bbox": [ + 313, + 592, + 556, + 711 + ], + "type": "text", + "content": "By introducing CholecTrack20 to the computer vision research community, we aim to stimulate new research directions and foster collaborations between the computer vision and surgical communities. This dataset serves as a benchmark for evaluating state-of-the-art algorithms and promotes the development of robust and reliable surgical assistance systems. We anticipate that CholecTrack20 will inspire innovative approaches and contribute significantly to advancements in surgical tool tracking and related fields. The dataset is released under the CC BY-NC license." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8949" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 145, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 145, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 145, + 85 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 89, + 297, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 89, + 297, + 224 + ], + "spans": [ + { + "bbox": [ + 55, + 89, + 297, + 224 + ], + "type": "text", + "content": "This work was supported by French state funds managed within the Plan Investissements d'Avenir by the ANR under references: National AI Chair AI4ORSafety [ANR-20-CHIA-0029-01], DeepSurg [ANR-16-CE33-0009], IHU Strasbourg [ANR-10-IAHU-02] and by BPI France under references: project CONDOR, project 5G-OR. Joel L. Lavanchy received funding by the Swiss National Science Foundation (P500PM_206724, P5R5PM_217663). This work was granted access to the servers/HPC resources managed by CAMMA, IHU Strasbourg, Unistra Mesocentre, and GENCI-IDRIS [Grant 2021-AD011011638R3, 2021-AD011011638R4]." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 242, + 115, + 255 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 242, + 115, + 255 + ], + "spans": [ + { + "bbox": [ + 56, + 242, + 115, + 255 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 262, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 262, + 296, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 262, + 296, + 295 + ], + "spans": [ + { + "bbox": [ + 61, + 262, + 296, + 295 + ], + "type": "text", + "content": "[1] Nir Aharon, Roy Orfaig, and Ben-Zion Bobrovsky. Botsort: Robust associations multi-pedestrian tracking. arXiv preprint arXiv:2206.14651, 2022. 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 295, + 296, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 295, + 296, + 350 + ], + "spans": [ + { + "bbox": [ + 61, + 295, + 296, + 350 + ], + "type": "text", + "content": "[2] Hassan Al Hajj, Mathieu Lamard, Pierre-Henri Conze, Béatrice Cochener, and Gwenolé Quellec. Monitoring tool usage in surgery videos using boosted convolutional and recurrent neural networks. Medical image analysis, 47:203-218, 2018. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 350, + 296, + 393 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 350, + 296, + 393 + ], + "spans": [ + { + "bbox": [ + 62, + 350, + 296, + 393 + ], + "type": "text", + "content": "[3] Keni Bernstein and Rainer Stiefelhagen. Evaluating multiple object tracking performance: the clear mot metrics. EURASIP Journal on Image and Video Processing, 2008:1-10, 2008. 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 395, + 296, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 395, + 296, + 437 + ], + "spans": [ + { + "bbox": [ + 62, + 395, + 296, + 437 + ], + "type": "text", + "content": "[4] Alex Bewley, Zongyuan Ge, Lionel Ott, Fabio Ramos, and Ben Upcroft. Simple online and realtime tracking. In 2016 IEEE international conference on image processing (ICIP), pages 3464-3468. IEEE, 2016. 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 438, + 296, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 438, + 296, + 491 + ], + "spans": [ + { + "bbox": [ + 62, + 438, + 296, + 491 + ], + "type": "text", + "content": "[5] David Bouget, Rodrigo Benenson, Mohamed Omran, Laurent Riffaud, Bernt Schiele, and Pierre Jannin. Detecting surgical tools by modelling local appearance and global shape. IEEE transactions on medical imaging, 34(12):2603-2617, 2015. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 493, + 296, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 493, + 296, + 536 + ], + "spans": [ + { + "bbox": [ + 62, + 493, + 296, + 536 + ], + "type": "text", + "content": "[6] David Bouget, Max Allan, Danail Stoyanov, and Pierre Jannin. Vision-based and marker-less surgical tool detection and tracking: a review of the literature. Medical image analysis, 35:633-654, 2017. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 537, + 296, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 537, + 296, + 581 + ], + "spans": [ + { + "bbox": [ + 62, + 537, + 296, + 581 + ], + "type": "text", + "content": "[7] Zhaowei Cai and Nuno Vasconcelos. Cascade r-cnn: High quality object detection and instance segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence, page 1-1, 2019. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 582, + 296, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 582, + 296, + 624 + ], + "spans": [ + { + "bbox": [ + 62, + 582, + 296, + 624 + ], + "type": "text", + "content": "[8] Magdalena K Chmarra, CA Grimbergen, and J Dankelman. Systems for tracking minimally invasive surgical instruments. Minimally Invasive Therapy & Allied Technologies, 16(6):328-340, 2007. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 625, + 296, + 692 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 625, + 296, + 692 + ], + "spans": [ + { + "bbox": [ + 62, + 625, + 296, + 692 + ], + "type": "text", + "content": "[9] Tobias Czempiel, Magdalini Paschali, Matthias Keicher, Walter Simson, Hubertus Feussner, Seong Tae Kim, and Nassir Navab. Tecn: Surgical phase recognition with multistage temporal convolutional networks. In Medical Image Computing and Computer Assisted Intervention MICCAI, pages 343-352. Springer, 2020. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 692, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 692, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 692, + 296, + 713 + ], + "type": "text", + "content": "[10] Martin Danelljan, Goutam Bhat, Fahad Shahbaz Khan, and Michael Felsberg. Eco: Efficient convolution operators" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 333, + 73, + 555, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 555, + 105 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 555, + 105 + ], + "type": "text", + "content": "for tracking. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6638-6646, 2017. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 106, + 555, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 106, + 555, + 171 + ], + "spans": [ + { + "bbox": [ + 316, + 106, + 555, + 171 + ], + "type": "text", + "content": "[11] Achal Dave, Tarasha Khurana, Pavel Tokmakov, Cordelia Schmid, and Deva Ramanan. Tao: A large-scale benchmark for tracking any object. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part V 16, pages 436–454. Springer, 2020. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 172, + 554, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 172, + 554, + 228 + ], + "spans": [ + { + "bbox": [ + 316, + 172, + 554, + 228 + ], + "type": "text", + "content": "[12] Patrick Dendorfer, Hamid Rezatofighi, Anton Milan, Javen Shi, Daniel Cremers, Ian Reid, Stefan Roth, Konrad Schindler, and Laura Leal-Taixe. Mot20: A benchmark for multi object tracking in crowded scenes. arXiv preprint arXiv:2003.09003, 2020. 1, 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 228, + 555, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 228, + 555, + 293 + ], + "spans": [ + { + "bbox": [ + 316, + 228, + 555, + 293 + ], + "type": "text", + "content": "[13] Yunhao Du, Junfeng Wan, Yanyun Zhao, Binyu Zhang, Zhihang Tong, and Junhao Dong. Giaotracker: A comprehensive framework for mcmot with global information and optimizing strategies in visdrone 2021. In Proceedings of the IEEE/CVF International conference on computer vision, pages 2809-2819, 2021. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 293, + 554, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 293, + 554, + 338 + ], + "spans": [ + { + "bbox": [ + 316, + 293, + 554, + 338 + ], + "type": "text", + "content": "[14] Ariel Kate Dubin, Danielle Julian, Alyssa Tanaka, Patricia Mattingly, and Roger Smith. A model for predicting the gears score from virtual reality surgical simulator metrics. Surgical endoscopy, 32:3576-3581, 2018. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 338, + 554, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 338, + 554, + 360 + ], + "spans": [ + { + "bbox": [ + 316, + 338, + 554, + 360 + ], + "type": "text", + "content": "[15] EndoVis Sub-Instrument Challenge. Endovis sub-instrument grand challenge, 2024. Accessed: 2021-11-12. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 361, + 554, + 447 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 361, + 554, + 447 + ], + "spans": [ + { + "bbox": [ + 317, + 361, + 554, + 447 + ], + "type": "text", + "content": "[16] Mona Fathollahi, Mohammad Hasan Sarhan, Ramon Pena, Lela DiMonte, Anshu Gupta, Aishani Ataliwala, and Jocelyn Barker. Video-based surgical skills assessment using long term tool tracking. In Medical Image Computing and Computer Assisted Intervention-MICCAI 2022: 25th International Conference, Singapore, September 18-22, 2022, Proceedings, Part VII, pages 541-550. Springer, 2022. 1, 2, 3, 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 449, + 554, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 449, + 554, + 503 + ], + "spans": [ + { + "bbox": [ + 317, + 449, + 554, + 503 + ], + "type": "text", + "content": "[17] Marvin P Fried, Jonathan Kleefield, Harsha Gopal, Edward Reardon, Bryan T Ho, and Frederick A Kuhn. Image-guided endoscopic surgery: results of accuracy and performance in a multicenter clinical study using an electromagnetic tracking system. The Laryngoscope, 107(5):594-601, 1997. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 504, + 554, + 570 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 504, + 554, + 570 + ], + "spans": [ + { + "bbox": [ + 317, + 504, + 554, + 570 + ], + "type": "text", + "content": "[18] Yixin Gao, S Swaroop Vedula, Carol E Reiley, Narges Ah-midi, Balakrishnan Varadarajan, Henry C Lin, Lingling Tao, Luca Zappella, Benjamin Bejar, David D Yuh, et al. Jhu-isi gesture and skill assessment working set (jigsaws): A surgical activity dataset for human motion modeling. In MICCAI workshop: M2cai, page 3, 2014. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 571, + 554, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 571, + 554, + 602 + ], + "spans": [ + { + "bbox": [ + 317, + 571, + 554, + 602 + ], + "type": "text", + "content": "[19] Zheng Ge, Songtao Liu, Feng Wang, Zeming Li, and Jian Sun. YOLOX: Exceeding yolo series in 2021. arXiv preprint arXiv:2107.08430, 2021. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 603, + 554, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 603, + 554, + 647 + ], + "spans": [ + { + "bbox": [ + 316, + 603, + 554, + 647 + ], + "type": "text", + "content": "[20] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In 2012 IEEE conference on computer vision and pattern recognition, pages 3354-3361. IEEE, 2012. 1, 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 647, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 554, + 713 + ], + "type": "text", + "content": "[21] Amy Jin, Serena Yeung, Jeffrey Jopling, Jonathan Krause, Dan Azagury, Arnold Milstein, and Li Fei-Fei. Tool detection and operative skill assessment in surgical videos using region-based convolutional neural networks. In 2018 IEEE winter conference on applications of computer vision (WACV), pages 691-699. IEEE, 2018. 5" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "8950" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 105 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 105 + ], + "type": "text", + "content": "[22] Arne Hoffhues Jonathon Luiten. Trackeval. https://github.com/JonathonLuiten/TrackEval, 2020. Accessed: 16 October 2023. 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 106, + 294, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 106, + 294, + 137 + ], + "spans": [ + { + "bbox": [ + 56, + 106, + 294, + 137 + ], + "type": "text", + "content": "[23] Kang Kim and Hee Seok Lee. Probabilistic anchor assignment with iou prediction for object detection. In ECCV, 2020. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 139, + 294, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 139, + 294, + 183 + ], + "spans": [ + { + "bbox": [ + 56, + 139, + 294, + 183 + ], + "type": "text", + "content": "[24] Joel L Lavanchy, Joel Zindel, Kadir Kirtac, Isabell Twick, Enes Hosgor, Daniel Candinas, and Guido Beldi. Automation of surgical skill assessment using a three-stage machine learning algorithm. Scientific reports, 11(1):5197, 2021. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 184, + 294, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 184, + 294, + 237 + ], + "spans": [ + { + "bbox": [ + 56, + 184, + 294, + 237 + ], + "type": "text", + "content": "[25] Joel L Lavanchy, Armine Vardazaryan, Pietro Mascagni, Didier Mutter, and Nicolas Padoy. Preserving privacy in surgical video analysis using a deep learning classifier to identify out-of-body scenes in endoscopic videos. Scientific reports, 13(1):9235, 2023. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 239, + 294, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 239, + 294, + 304 + ], + "spans": [ + { + "bbox": [ + 56, + 239, + 294, + 304 + ], + "type": "text", + "content": "[26] Byungjae Lee, Enkhbayar Erdenee, Songguo Jin, Mi Young Nam, Young Giu Jung, and Phill Kyu Rhee. Multi-class multi-object tracking using changing point detection. In Computer Vision-ECCV 2016 Workshops: Amsterdam, The Netherlands, October 8-10 and 15-16, 2016, Proceedings, Part II 14, pages 68-83. Springer, 2016. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 305, + 294, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 305, + 294, + 369 + ], + "spans": [ + { + "bbox": [ + 56, + 305, + 294, + 369 + ], + "type": "text", + "content": "[27] Dongheon Lee, Hyeong Won Yu, Hyungju Kwon, HyounJoong Kong, Kyu Eun Lee, and Hee Chan Kim. Evaluation of surgical skills during robotic surgery by deep learning-based multiple surgical instrument tracking in training and actual operations. Journal of clinical medicine, 9(6):1964, 2020. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 371, + 294, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 371, + 294, + 415 + ], + "spans": [ + { + "bbox": [ + 56, + 371, + 294, + 415 + ], + "type": "text", + "content": "[28] Eung-Joo Lee, William Plishker, Xinyang Liu, Shuvra S Bhattacharyya, and Raj Shekhar. Weakly supervised segmentation for real-time surgical tool tracking. Healthcare technology letters, 6(6):231-236, 2019. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 415, + 294, + 481 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 415, + 294, + 481 + ], + "spans": [ + { + "bbox": [ + 56, + 415, + 294, + 481 + ], + "type": "text", + "content": "[29] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 482, + 294, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 482, + 294, + 514 + ], + "spans": [ + { + "bbox": [ + 56, + 482, + 294, + 514 + ], + "type": "text", + "content": "[30] Wei Liu, Dragomir Anguelov, Dumitru Erhan, Christian Szegedy, Scott Reed, Cheng-Yang Fu, and Alexander C. Berg. Ssd: Single shot multibox detector. ECCV, 2016. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 515, + 294, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 515, + 294, + 558 + ], + "spans": [ + { + "bbox": [ + 56, + 515, + 294, + 558 + ], + "type": "text", + "content": "[31] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. arXiv preprint arXiv:2103.14030, 2021. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 559, + 294, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 559, + 294, + 612 + ], + "spans": [ + { + "bbox": [ + 56, + 559, + 294, + 612 + ], + "type": "text", + "content": "[32] Jonathon Luiten, Aljosa Osep, Patrick Dendorfer, Philip Torr, Andreas Geiger, Laura Leal-Taixe, and Bastian Leibe. Hota: A higher order metric for evaluating multi-object tracking. International journal of computer vision, 129:548-578, 2021. 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 614, + 294, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 614, + 294, + 656 + ], + "spans": [ + { + "bbox": [ + 56, + 614, + 294, + 656 + ], + "type": "text", + "content": "[33] Gerard Maggiolino, Adnan Ahmad, Jinkun Cao, and Kris Kitani. Deep oc-sort: Multi-pedestrian tracking by adaptive re-identification. arXiv preprint arXiv:2302.11813, 2023. 6, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 658, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 294, + 713 + ], + "type": "text", + "content": "[34] Lena Maier-Hein, Swaroop S Vedula, Stefanie Speidel, Nassir Navab, Ron Kikinis, Adrian Park, Matthias Eisenmann, Hubertus Feussner, Germain Forestier, Stamatia Giannarou, et al. Surgical data science for next-generation interventions. Nature Biomedical Engineering, 1(9):691-696, 2017. 1" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "text", + "content": "[35] Afrooz Moatari-Kazerouni and Ygal Bendavid. Improving logistics processes of surgical instruments: case of RFID technology. Business Process Management Journal, 23(2):448-466, 2017. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 118, + 553, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 553, + 162 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 553, + 162 + ], + "type": "text", + "content": "[36] Chinedu Innocent Nwoye. Deep learning methods for the detection and recognition of surgical tools and activities in laparoscopic videos. PhD thesis, Université de Strasbourg, 2021. 1, 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 163, + 553, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 553, + 218 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 553, + 218 + ], + "type": "text", + "content": "[37] Chinedu Innocent Nwoye, Didier Mutter, Jacques Marescaux, and Nicolas Padoy. Weakly supervised convolutional LSTM approach for tool tracking in laparoscopic videos. International journal of computer assisted radiology and surgery, 14:1059-1067, 2019. 1, 2, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 220, + 553, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 220, + 553, + 295 + ], + "spans": [ + { + "bbox": [ + 316, + 220, + 553, + 295 + ], + "type": "text", + "content": "[38] Chinedu Innocent Nwoye, Cristians Gonzalez, Tong Yu, Pietro Mascagni, Didier Mutter, Jacques Marescaux, and Nicolas Padoy. Recognition of instrument-tissue interactions in endoscopic videos via action triplets. In International Conference on Medical Image Computing and Computer-Assisted Intervention MICCAI, pages 364-374. Springer, 2020. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 297, + 553, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 297, + 553, + 352 + ], + "spans": [ + { + "bbox": [ + 316, + 297, + 553, + 352 + ], + "type": "text", + "content": "[39] Chinedu Innocent Nwoye, Tong Yu, Cristians Gonzalez, Barbara Seeliger, Pietro Mascagni, Didier Mutter, Jacques Marescaux, and Nicolas Padoy. Rendezvous: Attention mechanisms for the recognition of surgical action triplets in endoscopic videos. 78:102433, 2022. 2, 3, 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 354, + 553, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 354, + 553, + 420 + ], + "spans": [ + { + "bbox": [ + 316, + 354, + 553, + 420 + ], + "type": "text", + "content": "[40] Chinedu Innocent Nwoye, Tong Yu, Saurav Sharma, Aditya Murali, Deepak Alapatt, Armine Vardazaryan, Kun Yuan, Jonas Hajek, Wolfgang Reiter, Amine Yamlahi, et al. Cholectriplet2022: Show me a tool and tell me the triplet—an endoscopic vision challenge for surgical action triplet detection. arXiv preprint arXiv:2302.06294, 2023. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 421, + 553, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 421, + 553, + 464 + ], + "spans": [ + { + "bbox": [ + 316, + 421, + 553, + 464 + ], + "type": "text", + "content": "[41] Liang Qiu, Changsheng Li, and Hongliang Ren. Real-time surgical instrument tracking in robot-assisted surgery using multi-domain convolutional neural network. Healthcare technology letters, 6(6):159-164, 2019. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 466, + 553, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 466, + 553, + 521 + ], + "spans": [ + { + "bbox": [ + 316, + 466, + 553, + 521 + ], + "type": "text", + "content": "[42] Sanat Ramesh, Vinkle Srivastav, Deepak Alapatt, Tong Yu, Aditya Murali, Luca Sestini, Chinedu Innocent Nwoye, Idris Hamoud, Saurav Sharma, Antoine Fleurentin, et al. Dissecting self-supervised learning methods for surgical computer vision. Medical Image Analysis, 88:102844, 2023. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 522, + 553, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 522, + 553, + 567 + ], + "spans": [ + { + "bbox": [ + 316, + 522, + 553, + 567 + ], + "type": "text", + "content": "[43] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2017. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 568, + 553, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 568, + 553, + 643 + ], + "spans": [ + { + "bbox": [ + 316, + 568, + 553, + 643 + ], + "type": "text", + "content": "[44] Rogério Richa, Marcin Balicki, Eric Meisner, Raphael Sznitman, Russell Taylor, and Gregory Hager. Visual tracking of surgical tools for proximity detection in retinal surgery. In Information Processing in Computer-Assisted Interventions: Second International Conference, IPCAI 2011, Berlin, Germany, June 22, 2011. Proceedings 2, pages 55-66. Springer, 2011. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 646, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 646, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 646, + 553, + 689 + ], + "type": "text", + "content": "[45] Ergys Ristani, Francesco Solera, Roger Zou, Rita Cucchiara, and Carlo Tomasi. Performance measures and a data set for multi-target, multi-camera tracking. In European conference on computer vision, pages 17-35. Springer, 2016. 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "text", + "content": "[46] Maria Robu, Abdelrahim Kadkhodamohammadi, Imanol Luengo, and Danail Stoyanov. Towards real-time multiple" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "8951" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 76, + 72, + 296, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 296, + 106 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 296, + 106 + ], + "type": "text", + "content": "surgical tool tracking. Computer Methods in Biomechanics and Biomedical Engineering: Imaging & Visualization, 9(3): 279-285, 2021. 1, 2, 3, 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 296, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 296, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 296, + 162 + ], + "type": "text", + "content": "[47] Duygu Sarikaya, Jason J Corso, and Khurshid A Guru. Detection and localization of robotic tools in robot-assisted surgery videos using deep neural networks for region proposal and detection. IEEE transactions on medical imaging, 36(7):1542-1549, 2017. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 296, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 296, + 217 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 296, + 217 + ], + "type": "text", + "content": "[48] Mohammad A Shbool, Ammar Al-Bazi, Alma Kokash, AlAlaween Wafa'H, Nibal T Albashabsheh, and Raed Al-Taher. The economy of motion for laparoscopic ball clamping surgery: A feedback educational tool. MethodsX, 10: 102168, 2023. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 220, + 296, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 220, + 296, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 220, + 296, + 262 + ], + "type": "text", + "content": "[49] Pan Shi, Zijian Zhao, Sanyuan Hu, and Faliang Chang. Realtime surgical tool detection in minimally invasive surgery based on attention-guided convolutional neural network. IEEE Access, 8:228853-228862, 2020. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 264, + 296, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 264, + 296, + 308 + ], + "spans": [ + { + "bbox": [ + 56, + 264, + 296, + 308 + ], + "type": "text", + "content": "[50] Peize Sun, Jinkun Cao, Yi Jiang, Rufeng Zhang, Enze Xie, Zehuan Yuan, Changhu Wang, and Ping Luo. Transtrack: Multiple object tracking with transformer. arXiv preprint arXiv:2012.15460, 2020. 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 309, + 296, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 309, + 296, + 365 + ], + "spans": [ + { + "bbox": [ + 56, + 309, + 296, + 365 + ], + "type": "text", + "content": "[51] Peize Sun, Jinkun Cao, Yi Jiang, Zehuan Yuan, Song Bai, Kris Kitani, and Ping Luo. Dancetrack: Multi-object tracking in uniform appearance and diverse motion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20993-21002, 2022. 1, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 365, + 296, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 365, + 296, + 420 + ], + "spans": [ + { + "bbox": [ + 56, + 365, + 296, + 420 + ], + "type": "text", + "content": "[52] Xiaochuan Sun and Shahram Payandeh. Estimation of incision patterns based on visual tracking of surgical tools in minimally invasive surgery. In ASME International Mechanical Engineering Congress and Exposition, pages 75-83, 2010. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 422, + 296, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 422, + 296, + 477 + ], + "spans": [ + { + "bbox": [ + 56, + 422, + 296, + 477 + ], + "type": "text", + "content": "[53] Raphael Sznitman, Karim Ali, Rogerio Richa, Russell H Taylor, Gregory D Hager, and Pascal Fua. Data-driven visual tracking in retinal microsurgery. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 568-575. Springer, 2012. 2, 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 478, + 296, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 478, + 296, + 511 + ], + "spans": [ + { + "bbox": [ + 56, + 478, + 296, + 511 + ], + "type": "text", + "content": "[54] Zhi Tian, Chunhua Shen, Hao Chen, and Tong He. Fcos: Fully convolutional one-stage object detection. arXiv preprint arXiv:1904.01355, 2019. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 512, + 296, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 512, + 296, + 578 + ], + "spans": [ + { + "bbox": [ + 56, + 512, + 296, + 578 + ], + "type": "text", + "content": "[55] Jay Toor, Avneesh Bhangu, Jesse Wolfstadt, Garry Bassi, Stanley Chung, Raja Rampersaud, William Mitchell, Joseph Milner, and Martin Koyle. Optimizing the surgical instrument tray to immediately increase efficiency and lower costs in the operating room. Canadian Journal of Surgery, 65(2): E275, 2022. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 579, + 296, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 579, + 296, + 633 + ], + "spans": [ + { + "bbox": [ + 56, + 579, + 296, + 633 + ], + "type": "text", + "content": "[56] Andru P Twinanda, Sherif Shehata, Didier Mutter, Jacques Marescaux, Michel De Mathelin, and Nicolas Padoy. Endonet: a deep architecture for recognition tasks on laparoscopic videos. IEEE transactions on medical imaging, 36 (1):86-97, 2016. 1, 2, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 635, + 296, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 296, + 679 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 296, + 679 + ], + "type": "text", + "content": "[57] Armine Vardazaryan, Didier Mutter, Jacques Marescaux, and Nicolas Padoy. Weakly-supervised learning for tool localization in laparoscopic videos. In MICCAI-LABEL, pages 169–179. Springer, 2018. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 681, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 681, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 681, + 296, + 713 + ], + "type": "text", + "content": "[58] Rejin Varghese and M Sambath. Yolov8: A novel object detection algorithm with enhanced performance and robustness. In 2024 International Conference on Advances in Data" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 333, + 72, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 72, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 72, + 553, + 95 + ], + "type": "text", + "content": "Engineering and Intelligent Computing Systems (ADICS), pages 1-6. IEEE, 2024. 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 95, + 553, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 95, + 553, + 171 + ], + "spans": [ + { + "bbox": [ + 316, + 95, + 553, + 171 + ], + "type": "text", + "content": "[59] Martin Wagner, Beat-Peter Müller-Stich, Anna Kisilenko, Duc Tran, Patrick Heger, Lars Mündermann, David M Lubotsky, Benjamin Müller, Tornike Davitashvili, Manuela Capek, et al. Comparative validation of machine learning algorithms for surgical workflow and skill analysis with the heichole benchmark. Medical Image Analysis, 86:102770, 2023. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 172, + 553, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 172, + 553, + 215 + ], + "spans": [ + { + "bbox": [ + 316, + 172, + 553, + 215 + ], + "type": "text", + "content": "[60] Ao Wang, Hui Chen, Lihao Liu, Kai Chen, Zijia Lin, Jungong Han, et al. Yolov10: Real-time end-to-end object detection. Advances in Neural Information Processing Systems, 37:107984-108011, 2024. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 217, + 553, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 217, + 553, + 249 + ], + "spans": [ + { + "bbox": [ + 316, + 217, + 553, + 249 + ], + "type": "text", + "content": "[61] Chien-Yao Wang and Hong-Yuan Mark Liao. Yolov9: Learning what you want to learn using programmable gradient information. 2024. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 250, + 553, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 250, + 553, + 304 + ], + "spans": [ + { + "bbox": [ + 316, + 250, + 553, + 304 + ], + "type": "text", + "content": "[62] Chien-Yao Wang, Alexey Bochkovskyi, and Hong-Yuan Mark Liao. Yolov7: Trainable bag-of-freebies sets new state-of-the-art for real-time object detectors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7464-7475, 2023. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 305, + 553, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 305, + 553, + 348 + ], + "spans": [ + { + "bbox": [ + 316, + 305, + 553, + 348 + ], + "type": "text", + "content": "[63] Yu-Hsiang Wang, Jun-Wei Hsieh, Ping-Yang Chen, Ming-Ching Chang, Hung Hin So, and Xin Li. Smiletrack: Similarity learning for occlusion-aware multiple object tracking, 2023. 6, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 349, + 553, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 349, + 553, + 392 + ], + "spans": [ + { + "bbox": [ + 316, + 349, + 553, + 392 + ], + "type": "text", + "content": "[64] Nicolai Wojke, Alex Bewley, and Dietrich Paulus. Simple online and realtime tracking with a deep association metric. In 2017 IEEE international conference on image processing (ICIP), pages 3645-3649. IEEE, 2017. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 393, + 553, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 393, + 553, + 448 + ], + "spans": [ + { + "bbox": [ + 316, + 393, + 553, + 448 + ], + "type": "text", + "content": "[65] Lisheng Xu, Haoran Zhang, Jiaole Wang, Ang Li, Shuang Song, Hongliang Ren, Lin Qi, Jason J Gu, and Max Q-H Meng. Information loss challenges in surgical navigation systems: From information fusion to ai-based approaches. Information Fusion, 2022. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 449, + 553, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 449, + 553, + 492 + ], + "spans": [ + { + "bbox": [ + 316, + 449, + 553, + 492 + ], + "type": "text", + "content": "[66] Yifu Zhang, Chunyu Wang, Xinggang Wang, Wenjun Zeng, and Wenyu Liu. Fairmot: On the fairness of detection and re-identification in multiple object tracking. International Journal of Computer Vision, 129:3069–3087, 2021. 2, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 492, + 553, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 492, + 553, + 558 + ], + "spans": [ + { + "bbox": [ + 316, + 492, + 553, + 558 + ], + "type": "text", + "content": "[67] Yifu Zhang, Peize Sun, Yi Jiang, Dongdong Yu, Fucheng Weng, Zehuan Yuan, Ping Luo, Wenyu Liu, and Xinggang Wang. Bytetrack: Multi-object tracking by associating every detection box. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXII, pages 1-21. Springer, 2022. 6, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 559, + 553, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 559, + 553, + 602 + ], + "spans": [ + { + "bbox": [ + 316, + 559, + 553, + 602 + ], + "type": "text", + "content": "[68] Zijian Zhao, Sandrine Voros, Zhaorui Chen, and Xiaolin Cheng. Surgical tool tracking based on two cnns: from coarse to fine. The Journal of Engineering, 2019(14):467-472, 2019. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 602, + 553, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 602, + 553, + 624 + ], + "spans": [ + { + "bbox": [ + 316, + 602, + 553, + 624 + ], + "type": "text", + "content": "[69] Xingyi Zhou, Dequan Wang, and Philipp Krahenbuhl. Objects as points. arXiv preprint arXiv:1904.07850, 2019. 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 625, + 553, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 625, + 553, + 669 + ], + "spans": [ + { + "bbox": [ + 316, + 625, + 553, + 669 + ], + "type": "text", + "content": "[70] Pengfei Zhu, Longyin Wen, Dawei Du, Xiao Bian, Heng Fan, Qinghua Hu, and Haibin Ling. Detection and tracking meet drones challenge. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(11):7380-7399, 2021. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "text", + "content": "[71] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. In International Conference on Learning Representations, 2021. 6" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8952" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Circumventing Shortcuts in Audio-visual Deepfake Detection Datasets with Unsupervised Learning/371069ba-8973-4b9c-bb98-ecda46029053_content_list.json b/2025/Circumventing Shortcuts in Audio-visual Deepfake Detection Datasets with Unsupervised Learning/371069ba-8973-4b9c-bb98-ecda46029053_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..82218f8ad9ddeb49ca13fe0774168457b5b78966 --- /dev/null +++ b/2025/Circumventing Shortcuts in Audio-visual Deepfake Detection Datasets with Unsupervised Learning/371069ba-8973-4b9c-bb98-ecda46029053_content_list.json @@ -0,0 +1,1629 @@ +[ + { + "type": "text", + "text": "Circumventing shortcuts in audio-visual deepfake detection datasets with unsupervised learning", + "text_level": 1, + "bbox": [ + 130, + 130, + 867, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Stefan Smeu\\*1 Dragos-Alexandru Boldisor\\*1 Dan Oneata2 Elisabeta Oneata1 1Bitdefender 2POLITEHNICA Bucharest", + "bbox": [ + 166, + 202, + 826, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{ssmeu, dboldisor, eoneata}@bitdefender.com dan. oneata@gmail.com", + "bbox": [ + 207, + 241, + 784, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 291, + 326, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Good datasets are essential for developing and benchmarking any machine learning system. Their importance is even more extreme for safety critical applications such as deepfake detection—the focus of this paper. Here we reveal that two of the most widely used audio-video deepfake datasets suffer from a previously unidentified spurious feature: the leading silence. Fake videos start with a very brief moment of silence and, on the basis of this feature alone, we can separate the real and fake samples almost perfectly. As such, previous audio-only and audio-video models exploit the presence of silence in the fake videos and consequently perform worse when the leading silence is removed. To circumvent latching on such an unwanted artifact and possibly other unrevealed ones, we propose a shift from supervised to unsupervised learning by training models exclusively on real data. We show that by aligning self-supervised audio-video representations we remove the risk of relying on dataset-specific biases and improve robustness in deepfake detection.", + "bbox": [ + 86, + 321, + 483, + 609 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 638, + 220, + 656 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Manipulated videos represent a threat to society as they have the potential of misleading people into believing actors with malicious intents. By spreading misinformation on social media platforms, people may be exposed to scams (e.g., identity theft operations), conspiracy theories and political misinformation. Therefore, deepfake detection methods are essential tools on a global scale.", + "bbox": [ + 89, + 665, + 482, + 770 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The progress in automated deepfake detection is fueled by the datasets developed by the research community. Good quality data is essential for both training and benchmarking the progress of these methods. In recent years, numerous datasets have been proposed. Among the audio-video deepfake datasets, there have been released datasets that alter both streams [6, 23] or only one [36, 46, 56]; datasets with", + "bbox": [ + 89, + 771, + 482, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/88ea4fd6679afa749cb4379af534874a4687fbf669bfa7767bf2cbb9d0414013.jpg", + "image_caption": [ + "Figure 1. Audio-visual deepfake detection datasets have a silence bias: fake samples start with a brief moment of silence, which is not the case for real samples. Here we show the first $62.5\\mathrm{ms}$ of the audio waveform for a real and the corresponding fake sample from the AV-Deepfake1M dataset [6]." + ], + "image_footnote": [], + "bbox": [ + 545, + 292, + 874, + 426 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "full-video [23] or only local [4, 6] manipulations.", + "bbox": [ + 511, + 536, + 839, + 551 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "However, care has to be taken to ensure a good deepfake detection dataset. Any asymmetry in the preparation of fakes and reals can result in biases that correlate spuriously with the groundtruth label. For example, in the image domain, different preprocessing pipelines [8] or types of resizing [43] between real and fake samples paint an overly optimistic picture. In the audio domain, the very popular ASVspoof19 dataset [56] has been shown to leak information about labels in the form of silence duration [35] or bitrate information [3]. Since many of the deepfake detection models are high capacity, they can easily use such artifacts as shortcuts for learning, consequently greatly impacting their ability to generalize in a real case scenario.", + "bbox": [ + 509, + 551, + 906, + 748 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this work, we first expose a bias present in two widely-adopted audio-video deepfake detection datasets, FakeAVCeleb [23] and AV-Deepfake1M [6]: a short moment of silence present at the beginning of manipulated videos (see Figure 1). Based on this information, a very simple silence classifier can reach a near-perfect performance of over $98\\%$ on both FakeAVCeleb and AV-Deepfake1M. Moreover, we observe that the silence distribution is similar across datasets, implying that prior methods may have potentially over estimated the generalization", + "bbox": [ + 511, + 750, + 908, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution.", + "bbox": [ + 109, + 887, + 218, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "18815", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "performance based on this shortcut.", + "bbox": [ + 89, + 90, + 328, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Introducing such biases when creating the datasets is unavoidable and there might be others that are more subtle and harder to reveal. To circumvent this problem, we show that one solution is a shift in learning paradigm, from the more common supervised setup to the unsupervised one [17, 44]. Removing fake samples from training eliminates the focus on the asymmetries induced by spurious artifacts. Moreover, this also has the potential to improve generalization among different manipulation techniques, since supervised detection methods tend to over-rely on generator-specific fingerprints [34, 59]. By limiting to real data, we can also naturally leverage self-supervised representations, which have been shown to improve generalization [38].", + "bbox": [ + 89, + 106, + 480, + 301 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To this end, we propose an approach, named AVH-Align, that learns a frame-level audio-video alignment score on top of AV-HuBERT features [49]. Since tampered videos are expected to have greater audio-video desynchronizations, this alignment score can effectively differentiate fake and real videos. We show that this approach is robust with respect to the identified shortcut of leading silence, and also outperforms other methods that do not use the silence bias, while not seeing any fake samples at training time.", + "bbox": [ + 89, + 303, + 480, + 438 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To summarize, our work makes the following contributions: 1. We expose a previously unknown spurious feature in two of the most widely adopted deepfake detection datasets. 2. We analyze the impact of this shortcut on various state-of-the-art models. 3. We show that a way to mitigate such shortcuts is by training on real data only and we introduce a new method in this direction. Our code is available at: https://github.com/bit-ml/AVH-Align.", + "bbox": [ + 89, + 439, + 480, + 559 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 89, + 573, + 227, + 589 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Audio-video deepfake detection. Many approaches for deepfake detection on videos have focused on the visual stream of information only [1, 5, 18, 19, 21, 50, 60]. But with the recent introduction of audio-visual datasets, (FakeAVCeleb [23], Deepfake TIMIT [25, 47], KODF [28], LAV-DF [4], AV-Deepfake-1M [6]), more research has shifted towards models that exploit both audio and video cues [17, 20, 27, 39, 48, 58]. An emerging trend in this direction is the use of pretrained representations in a self-supervised way [17, 39]. But, different from our approach, these methods train the representations from scratch and use them as a first step in a more elaborated pipeline: supervised classification [39] or anomaly detection [17]. There are also works that similarly to us exploit the pretrained audio-video AV-HuBERT model [49] to extract representations [20, 48], but all of those methods are trained in the fully supervised learning paradigm.", + "bbox": [ + 89, + 606, + 482, + 863 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Unsupervised deepfake detection. To make detection more generalizable across generators, a new direction is", + "bbox": [ + 89, + 869, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "to depart from the supervised paradigm and resort only to real data. There are two main classes of such unsupervised approaches: methods that rely on consistency checks and methods that treat the problem as an anomaly detection one. Among the first class, prior work proposed to verify that audio and visual streams align at a semantic level, for example, from the point of view of spoken content [2, 29], or at a representation level, based on the alignment of audio and video features [17, 44]. Consistency checks have also been used with respect to the identity of a speaker (comparing a query sample against real audio [41] or real images [44] of the target speaker) or between image and text modalities [44]. For the second class, deepfake detection as anomaly detection, recent work in the image domain has used the reconstruction loss of the query image to tell whether it is anomalous [14, 45]; for example, Ricker et al. [45] make the observation that images generated by a latent diffusion model (LDM) are easier to reconstruct by the LDM than real images. For the audio-visual domain, Feng et al. [17] use both classes of approaches: they use consistency checks to estimate synchronization between the two streams and then flag anomalies using density estimation.", + "bbox": [ + 511, + 90, + 903, + 422 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Silence bias in audio-video datasets", + "text_level": 1, + "bbox": [ + 511, + 439, + 831, + 455 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this section, we show that two popular datasets (Sec. 3.1) have a silence bias. We analyze its behavior and show that a simple classifier based on the leading silence alone can obtain almost perfect separation between fake and real samples (Sec. 3.2). This implies that the performance of prior work is susceptible to have been overestimated. For this reason we analyze its impact on various audio and audio-visual methods (Sec. 3.3).", + "bbox": [ + 511, + 465, + 903, + 585 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3.1. Datasets", + "text_level": 1, + "bbox": [ + 511, + 597, + 614, + 611 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We consider two audio-visual datasets in our analysis: FakeAVCeleb [23] and AV-Deepfake1M [6]. They distinguish mainly in the fact that the first contains fully-generated video sequences, while the second contains partially-manipulated sequences. Both are based on the VoxCeleb2 dataset [11], which consists of YouTube audio-video of celebrities. Apart from the real samples—real video real audio (RVRA)—both datasets include three types of fake videos: real video fake audio (RVFA), fake video real audio (FVRA), and fake video fake audio (FVFA).", + "bbox": [ + 511, + 621, + 903, + 772 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "FakeAVCeleb contains 500 real videos from VoxCeleb2 and 19.5k fake videos (10k FVFA, 9k FVRA, 500 RVFA). The fake visual content was generated with face swapping methods (Faceswap [26] and FSGAN [37]) or the Wav2Lip lipSyncing approach [42]. The fake audio content was generate with the voice cloning tool SV2TTS [22]. The dataset is diverse across age groups, genders, races, as well as with respect to the number of subjects in a single video, their", + "bbox": [ + 511, + 779, + 906, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "18816", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/ac2fa88c7ea9a187a45c81f5d8eb3c9c641ae51ddf7a2a38b3ad0df6f89f25ee.jpg", + "image_caption": [ + "Figure 2. Normalized distribution plots of the leading silence duration for real and fake videos in the FakeAVCeleb (left) and AV-Deepfake1M (right) datasets. The fake samples start with 25–30 ms of silence." + ], + "image_footnote": [], + "bbox": [ + 94, + 88, + 480, + 229 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "placement, the visual and audio quality. In our experiments, we split the dataset in $70\\%$ (train and validation) and $30\\%$ (test). This split is kept consistent across all experiments.", + "bbox": [ + 89, + 325, + 482, + 372 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "AV-Deepfake1M is a large scale dataset, which consists of over one million videos and 2k subjects. As opposed to the FakeAVCeleb dataset, the manipulations here are local and consists of word-level replacements, insertions and deletions. The text manipulations are generated with the ChatGPT large language model. The fake video content is generated with the lipSyncing method TalkLip [55], while the fake audio content is generated with the VITS [24] or YourTTS [7] methods. The authors ensure that the synthesized words share the same background noise with the full audio, by first extracting the audio noise with the Denoiser method [15] and then adding it to the synthesized words. The dataset is originally split into train, validation and test, with the test split having a different set of speakers than those encountered in the train and validation splits. For our experiments, we select training and validation samples from the original training split, and evaluate on 10k samples from the original validation split or on the full official test set.", + "bbox": [ + 89, + 378, + 483, + 652 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Analysis of leading silence", + "text_level": 1, + "bbox": [ + 89, + 662, + 328, + 680 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We start by analyzing the silence distribution of the real and fake samples in the two considered datasets. We define the duration of the leading silence as the moment when the magnitude of the audio exceeds a certain threshold $\\tau$ . For this experiment, we select this threshold to be $5 \\cdot 10^{-4}$ , but as we will shortly see, the results are robust to its choice. We carry this analysis only on the real (RVRA) and fully fake (FVFA) videos from each dataset's test set.", + "bbox": [ + 89, + 686, + 482, + 808 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The results are shown in Figure 2. We observe that the real videos start with noise, while the fake samples have a leading silence of around $25 - 30\\mathrm{ms}$ . The silence duration of fake samples is similar for both datasets, although the distribution is much sharper for the AV-Deepfake1M dataset. If we were to rank the samples based on this feature we", + "bbox": [ + 89, + 810, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/e6c33b744cbca2d22ca98c8b5ac0f177e281b915cfcd47b16348cdd8f2267e97.jpg", + "image_caption": [ + "Figure 3. Left: The impact of the silence threshold on the leading silence classifier. Right: The impact of the leading duration on the maximum amplitude classifier." + ], + "image_footnote": [], + "bbox": [ + 514, + 87, + 906, + 229 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "would obtain an area under curve of the receiver operating characteristic curve (AUC) of over $98\\%$ for either datasets.", + "bbox": [ + 511, + 310, + 906, + 339 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "What counts as silence? For the previous experiment we have considered that silence is the signal that has an amplitude lower than $\\tau = 5 \\cdot 10^{-4}$ . We investigate how sensitive the performance is to this threshold. We vary $\\tau$ across a grid of values and show the results in Figure 3 (left). We observe that the results are strong as long as this threshold is small enough.", + "bbox": [ + 511, + 347, + 908, + 455 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "A different perspective: Maximum amplitude. Instead of looking at the silence duration, we can alternatively measure the maximum amplitude in the first $\\delta$ seconds of the audio. Figure 3 shows the AUC obtained by ranking the samples based on this feature for various leading durations $\\delta$ . We observe that this alternative measure also yields strong results (over $98\\%$ AUC), and that the optimum is obtained for a $\\delta$ of around $30~\\mathrm{ms}$ , in line with the previous experiment. As we extend the window, the performance gets closer to random chance, AUC of $50\\%$ .", + "bbox": [ + 511, + 460, + 908, + 612 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Other biases: Volume and trailing silence. If we extend the duration $\\delta$ to cover the entire audio, we get an estimate of the maximum volume of the audio. This feature yields an AUC of $67.6\\%$ on FakeAVCeleb, which is much less than $98.4\\%$ , but still over random chance (50% AUC). Similarly, we have also investigated the trailing silence (the silence at the end of the audio) and have obtained AUC of over $99\\%$ for FakeAVCeleb. AV-Deepfake1M is less sensitive to these other biases, with values closer to $50\\%$ . While these biases are not as consistent as the leading silence, they are still problematic.", + "bbox": [ + 511, + 619, + 908, + 787 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Why do fake samples have a leading silence? Given that we do not have access to generation process of the two datasets, it is challenging to pinpoint the exact reason for the occurrence of the leading silence bias. However, we speculate that this happens when the audio may be slightly shorter than the video counterpart. Note that this is a different reason from the silence observed in audio only datasets", + "bbox": [ + 511, + 794, + 908, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "18817", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "[35]. There the real audios had a leading silence, while the synthesized speech was silence-free. In the case of audio-video datasets it might be challenging to completely avoid this problem, but an easy solution (for this particular bias) is to trim the leading silence. This is what we do in the next experiments.", + "bbox": [ + 89, + 90, + 480, + 181 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Biases in other datasets. We consider three more datasets: LAV-DF [4], AVLips [31], DFDC [16]. We find that LAV-DF has a slight trailing silence bias (65% AUC), but it does not exhibit the other two biases (leading silence and volume). In the case of AVLips and DFDC we did not identify any of the three biases. This may happen because the samples in these datasets have real audio and the unmodified audio matches the visual sequence more precisely. This is confirmed by a fine-grained analysis on FakeAVCeleb, where we observe that on the FVRA split the biases are diminished (62.4% AUC for lead silence), while they persist on the RVFA split (100% AUC for lead silence).", + "bbox": [ + 89, + 188, + 482, + 369 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Impact on prior work", + "text_level": 1, + "bbox": [ + 89, + 378, + 294, + 395 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We have shown that real and fake samples differ in terms of the leading silence. This is a simple feature which could be learned by the high-capacity neural networks. Here we investigate whether that is indeed the case for three existing methods:", + "bbox": [ + 89, + 402, + 482, + 474 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- RawNet2 [53], which is an audio-only method that operates on the raw waveform. Its architecture sequences sinc layers, convolutions, gated recurrent units and fully connected layers.", + "- MDS [10] (modality dissonance score), which is an audio-visual method that estimates the mismatch between audio and video segments. The score is computed as the distance between audio and visual features.", + "- AVAD [17] (audio-visual anomaly detection), which is an audio-visual method trained on real data only. The approach has two steps: first, it estimates the desynchronization between audio and video; then it estimates whether these patterns are typical of real data or anomalous." + ], + "bbox": [ + 91, + 477, + 482, + 672 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We train the RawNet2 and MDS methods on both datasets using the code provided by the authors. For the AVAD we use the provided checkpoint (trained on LRS [52]) and do not retrain it since training code is not available. As a baseline we include our leading silence duration classifier described in the previous section.", + "bbox": [ + 89, + 674, + 482, + 763 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We evaluate all methods in two settings: on the corresponding evaluation set and on a trimmed version of the same evaluation set. For the trimmed version, we discard the beginning of videos such that each of them starts with the first non-silent segment. We find the duration to discard by first computing the leading silence (using the same $5 \\cdot 10^{-4}$ threshold on the magnitude) and then rounding it up to the nearest multiple of $1 / \\text{FPS}$ (the reciprocal of the video's frame rate); this rounding ensures that the audio and", + "bbox": [ + 89, + 765, + 482, + 900 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/a4684493e663c7f8ab10a1945e8dd7e3b2783046f6764ab5a4c9f240e4548984.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodMod.FakeAVCelebAV-Deepfake1M
Trim: XTrim: √Trim: XTrim: √
Silence classifierA98.454.8 ↓ 43.698.250.6 ↓ 47.6
RawNet2 [53]A99.997.3 ↓ 2.699.988.1 ↓ 11.8
MDS [10]AV90.473.8 ↓ 16.699.254.9 ↓ 44.3
AVAD [17]AV95.295.2 ≈ 0.052.952.9 ≈ 0.0
", + "bbox": [ + 514, + 88, + 905, + 183 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Table 1. The impact of leading silence on the performance (area under the receiver operator characteristic curve; AUC) for three existing deepfake detection methods and the silence classifier. Results are shown for RVRA-FVFA subsets.", + "bbox": [ + 511, + 193, + 903, + 247 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "video channels remain synchronized. For training, we do not perform trimming, but use the original dataset.", + "bbox": [ + 511, + 279, + 903, + 309 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The results are shown in Table 1. First, we observe that the audio-based RawNet2 is the top performing method. Moreover, we see that the leading silence classifier is generally better than the two other approaches: MDS and AVAD. The relatively stable performance of AVAD indicates that this method dose not latch on the silence information.", + "bbox": [ + 511, + 311, + 903, + 400 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In terms of the impact of the leading silence, we observe that removing it through trimming affects methods differently. As expected, the silence classifier is affected the most and its performance drops down to random chance on the trimmed data. RawNet2 is not as affected on the FakeAVCeleb, presumably because there is still enough information throughout the signal, but it suffers a larger hit on the more challenging AV-DeepFake1M, which has only partially-manipulated samples. MDS is affected by the leading silence bias on both FakeAVCeleb and AV-Deepfake1M, though significantly stronger on AV-Deepfake1M. AVAD is the most robust method, since this approach does not explicitly model the silence information. Finally, we notice that the RawNet2 audio-only method is best even on the trimmed settings. This is noteworthy since it is the first time an audio-only method has been applied to these datasets. This suggests that the audio stream is an important source of information that is often overlooked by prior work.", + "bbox": [ + 511, + 402, + 905, + 689 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Modeling real data for deepfake detection", + "text_level": 1, + "bbox": [ + 511, + 707, + 885, + 724 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The previous section indicated that audio-visual datasets exhibit a silence bias which can easily be exploited. We want to develop a method that is robust to this bias (and possibly other uncovered ones) and still performs well. We have seen that models trained only on real data [17] are promising in being robust to the silence shortcut, but the performance was modest. To further improve them we propose to build on top of audio-focused self-supervised features. Self-supervised features have shown strong generalization for both visual [13, 38] and audio [40, 41] deepfake detection. We choose audio-focused self-supervised features be", + "bbox": [ + 511, + 734, + 905, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "18818", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "cause the audio models showed strong performance in the previous section. Note that we cannot rely on audio-only models because there are cases where manipulations appear only in the visual domain (the fake video, real audio case).", + "bbox": [ + 89, + 90, + 483, + 151 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Method", + "text_level": 1, + "bbox": [ + 89, + 162, + 187, + 176 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We propose a method that aligns AV-HuBERT [49] features on real data. First, we extract audio and visual frame-level features with a pretrained AV-HuBERT model. Then, on top of these features we learn a network to better align them. The alignment network is learnt on real samples by matching each video frame to its corresponding audio frame [17]. We call our method AVH-Align (AV-HuBERT Aligned) and show its depiction in Figure 4.", + "bbox": [ + 89, + 186, + 483, + 308 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Self-supervised features. We use AV-HuBERT to represent both the audio and visual content of a video. AV-HuBERT is a Transformer network trained in a self-supervised way to predict iteratively refined centroids from masked inputs. The features extracted by AV-HuBERT encode audio information as proved by its strong performance on tasks such as lip reading or noisy audio-visual speech recognition. We extract the audio and visual representations independently: to extract audio features we mask the visual input, to extract visual features we mask the audio input. For a video, we obtain audio features $\\mathbf{a}_i$ and visual features $\\mathbf{v}_i$ features for each time step $i$ . Both representations are 1024 dimensional and have a temporal resolution of 25 frames per second.", + "bbox": [ + 89, + 315, + 483, + 526 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Alignment network. To tell how well the audio and visual features match each other, we first independently L2 normalize the feature vectors and then feed them into a network $\\Phi$ . This is implemented as a multi-layer perceptron (MLP) over the concatenated normalized audio and visual features:", + "bbox": [ + 89, + 534, + 483, + 625 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\Phi_ {i j} = \\operatorname {M L P} \\left(\\left[ \\mathbf {a} _ {i}; \\mathbf {v} _ {j} \\right]\\right). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 209, + 637, + 482, + 656 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The MLP has four layers, which progressively reduce the feature dimensionality, with layers mapping from the AV-HuBERT feature size of 1024 to 512, 256, 128, and finally to a single output. Each hidden layer includes Layer Normalization and ReLU activations.", + "bbox": [ + 89, + 667, + 483, + 743 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Loss function. To learn the alignment network $\\Phi$ we maximize the probability of an audio frame $\\mathbf{a}_i$ to match the corresponding video frame $\\mathbf{v}_i$ ; this probability is defined as:", + "bbox": [ + 89, + 751, + 483, + 797 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\np \\left(\\mathbf {v} _ {i} \\mid \\mathbf {a} _ {i}\\right) = \\frac {\\exp \\Phi_ {i i}}{\\sum_ {k \\in \\mathcal {N} (i)} \\exp \\Phi_ {i k}}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 181, + 808, + 482, + 844 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{N}(i)$ represents the temporal neighborhood around the frame $i$ . In our case $\\mathcal{N}(i)$ contains the 30 neighboring frames around $i$ . We define the final loss as the negative", + "bbox": [ + 89, + 854, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "probabilities averaged across the entire video:", + "bbox": [ + 511, + 90, + 816, + 106 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = - \\frac {1}{T} \\sum_ {i = 1} ^ {T} \\log p (\\mathbf {v} _ {i} | \\mathbf {a} _ {j}). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 617, + 116, + 905, + 157 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This loss is similar to the contrastive loss InfoNCE [54], which was also used for deepfake detection [17, 39].", + "bbox": [ + 511, + 166, + 905, + 198 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Inference. Once $\\Phi$ is learned, we can estimate the fakeness score as the negative of $\\Phi_{ii}$ for each audio-video frame pair in a video; aligned audio-video frame pairs should yield lower scores. Then we compute an overall alignment score for the entire video by pooling the per-frame scores using the log-sum-exp function (a smooth version of the max function).", + "bbox": [ + 511, + 204, + 905, + 309 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Supervised variant. To understand the impact of the silence bias on the standard supervised learning paradigm, we design a supervised variant, AVH-Align/sup, that uses the same features and alignment network as AVH-Align, but a classification loss. In this setup we assume that apart from the real videos, we also have access to fake videos in the training set, with corresponding labels $y$ . To obtain a per-video fakeness probability, we first pool the negated per-frame scores $\\Phi_{ii}$ with the log-sum-exp function and then we apply the sigmoid function $\\sigma$ . Finally, we optimize the binary cross-entropy (BCE) loss:", + "bbox": [ + 511, + 316, + 906, + 484 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\sup } = \\operatorname {B C E} \\left(\\sigma \\left(\\log \\operatorname {s u m} \\exp (- \\Phi_ {i i})\\right), y\\right). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 571, + 494, + 906, + 512 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Experimental setup", + "text_level": 1, + "bbox": [ + 511, + 530, + 700, + 547 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets and metrics. We conducted our experiments on the two datasets introduced in Sec. 3.1: FakeAVCeleb and AV-Deepfake1M. For FakeAVCeleb, we use $70\\%$ of the dataset for training and validation, and the rest of $30\\%$ for testing. For AV-Deepfake1M, we evaluate on 10k samples from the original validation set, but for the best performing models we also report results on the official withheld test set. To train AVH-Align, we use 50k real samples from the training set of AV-Deepfake1M (45k samples for training and 5k for validation). Since AV-Deepfake1M is based on VoxCeleb2, these samples originate from VoxCeleb2. To train the supervised variant, AVH-Align/sup, on AV-Deepfake1M, we select the same number of samples (45k for training and 5k for validation), but this time coming from both the real and fake classes. For both datasets, we evaluate video-level detection. Following prior work, we report results in terms of the area under the receiver operator characteristic curve (AUC) and average precision (AP), with the fake class being the positive class.", + "bbox": [ + 511, + 561, + 906, + 849 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation details. For AVH-Align we use a learning rate scheduler with a patience of 5 epochs and a factor of 0.1, with a starting learning rate of $10^{-5}$ . The training is", + "bbox": [ + 511, + 854, + 908, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "18819", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/fef535300c508cd85c42f608f28343135960f9d4e6659e6e4c7197083836fe04.jpg", + "image_caption": [ + "Figure 4. Overview of the AVH-Align method. A: We use the pretrained AV-HuBERT model to extract self-supervised features which we further align with a learnable network $\\Phi$ . Note that we use a single AV-HuBERT model, but make two forward passes to obtain audio-only and video-only features (instead of a single set of multimodal features). B: At training we maximize the alignment score $\\Phi_{ii}$ , between the audio features $\\mathbf{a}_i$ at time step $i$ and the corresponding video features $\\mathbf{v}_i$ , while minimizing the alignment $\\Phi_{ik}$ to the other features $\\mathbf{v}_k$ in a neighboring window $\\mathcal{N}(i)$ ." + ], + "image_footnote": [], + "bbox": [ + 210, + 88, + 584, + 232 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/543f0c7bf486300dc621cb5661b715db57f07236c69304ea86d63f3f43615941.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 591, + 88, + 759, + 232 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/b1a045c6c622e1664f95a6fe757766a84ba5afc349f8745f266e9bbc29e2440c.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodModalityTrain typeTrain dataMetric: AUCMetric: AP
FakeAVCelebAV-Deepfake1MFakeAVCelebAV-Deepfake1M
Trim: XTrim: ✓Trim: XTrim: ✓Trim: XTrim: ✓Trim: XTrim: ✓
AVH-Align/supAVsup.FAVC99.299.2 ≈69.063.6 ↓100.0100.0 ≈84.182.2 ↓
AVH-Align/supAVsup.AV1M77.570.8 ↓100.083.1 ↓99.499.1 ≈100.094.4 ↓
AVAD [17]AVunsup.LRS84.584.7 ≈54.354.3 ≈99.599.5 ≈76.376.3 ≈
SpeechForensics [30]AVunsup.VoxCeleb298.898.8 ≈68.868.2 ≈100.0100.0 ≈83.783.5 ≈
AVH-AlignAVunsup.VoxCeleb294.694.6 ≈85.983.5 ↓99.899.8 ≈94.393.5 ≈
", + "bbox": [ + 91, + 327, + 906, + 460 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Comparison of AVH-Align and AVH-Align/sup on FakeAVCeleb (FAVC) and AV-Deepfake1M (AV1M) original and trimmed datasets. AVH-Align is not impacted by the presence of leading silence in the original datasets, showing similar performance on the trimmed versions. In contrast, the AUC performance of AVH-Align/sup degrades by $16.9\\%$ when tested on the trimmed variant of AV1M.", + "bbox": [ + 89, + 469, + 906, + 513 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "stopped if there is no loss improvement on the validation set for 10 consecutive epochs. AVH-Align/sup is learned using an Adam optimizer with a learning rate of $10^{-3}$ . For the AVHuBERT feature extractor we use the checkpoint pretrained on LRS3 and VoxCeleb2 and finetuned on LRS3 for visual speech recognition.", + "bbox": [ + 89, + 539, + 482, + 628 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Experimental results", + "text_level": 1, + "bbox": [ + 89, + 641, + 287, + 657 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We train our method, AVH-Align, as well as its supervised variant, AVH-Align/sup, on the two datasets, FakeAVCeleb and AV-Deepfake1M, and evaluate on either of them (that is, both in-domain and out-of-domain). For each evaluation, in addition to the original set, we also consider the trimmed version of the validation dataset (trim: $\\checkmark$ ). The trimming is performed as described in Sec. 3.3. We always use the untrimmed, original data for training. Results are shown in Table 2.", + "bbox": [ + 89, + 665, + 482, + 800 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Impact of leading silence bias. We observe that AVH-Align is robust to the leading silence bias: trimming the silence has no effect on FakeAVCeleb and only a slight effect on AV-Deepfake1M (2.4% AUC). On the other hand, the performance of the supervised variant, AVHAlign/sup, decreases considerably when testing on AV", + "bbox": [ + 89, + 809, + 483, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Deepfake1M (by $16.9\\%$ or $5.4\\%$ AUC, depending on the train set). On FakeAVCeleb removing the leading silence has no impact for AVH-Align/sup, when training also on FakeAVCeleb. This may happen because FakeAVCeleb has full manipulations and there is other useful information in the video. Instead, when training on the locally-manipulated AV-Deepfake1M, the difference becomes significant $(6.7\\%)$ AUC.", + "bbox": [ + 511, + 537, + 906, + 657 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Comparison with other unsupervised methods. We compare our results with those of AVAD [17] and SpeechForensics [30], which are also unsupervised methods, trained on real data only. Similar to AVH-Align they are not impacted by the spurious leading silence, showing nearly identical results for trimmed and untrimmed datasets. However, their overall performance is considerably worse than ours on the AV-Deepfake1M dataset.", + "bbox": [ + 511, + 666, + 908, + 787 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Score visualization. Figure 5 shows the per frame scores obtained by the two methods, AVH-Align and AVH-Align/sup, together with the groundtruth manipulated interval. For the AVH-Align method, the scores represent the misalignment probability between the audio and visual streams at each time frame; for the AVH-Align/sup method, the scores represent the probability of an audio-video frame", + "bbox": [ + 511, + 794, + 908, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "18820", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/3a5e7050545c78dd216795fba5e17c516d365fcdc4fd74a3bf82eac464359703.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 101, + 89, + 346, + 162 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9fc5dc4342f77a1f76df917b818f00ee06880fe6068c5daef0ace80fae4a4c84.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 364, + 88, + 609, + 161 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3eb58c549dfade0519d30bdb06fc7a18cd3fabf2bf600c74cd88a91167ecaeec.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 625, + 88, + 872, + 161 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d684a2c335c1d0c7b8ca09f075030673d4aa06ea8848ed5f7118661714106f12.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 101, + 162, + 346, + 234 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/355e011d05424346819a2982958bbec153a7fba9682f6dc4343b90e39e78cc91.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 364, + 164, + 609, + 234 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d95be9185d9f5913eec140e6ae8a6ffd894d101050f2b8df55693bbc180e41bd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 625, + 164, + 872, + 234 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d279d263aa7afd4d98915b491760dec98e3721e50dfb8903fee1002589843c35.jpg", + "image_caption": [ + "Figure 5. Per frame fakeness probabilities for AVH-Align and AVH-Align/sup on AV-Deepfake1M. AVH-Align/sup always marks the first frame—corresponding usually to the leading silence—as fake, thus confirming that it uses the bias to distinguish between real and fake videos. AVH-Align is not affected by the presence of the leading silence. The fakeness probabilities for AVH-Align can be interpreted as misalignment probabilities, which is why they are higher during or after the manipulated region." + ], + "image_footnote": [], + "bbox": [ + 101, + 238, + 348, + 311 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/16d9b9bb8ba13e5ea84c2ddcee3d06d930ebcf47abc8260fcedf9544826f4db5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 364, + 238, + 609, + 310 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/cdfe86de160fa666ba4a3447feb789efc16039559628f24d475d68ff11186488.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 625, + 238, + 872, + 310 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "of being fake. We see that AVH-Align/sup always predicts the first frame as being fake, confirming once again that it has learned to associate the spurious leading silence with a video being fake. When detected, the actual fake region is given a considerable lower fakeness probability than that assigned to the leading silence. On the other hand, AVH-Align is not affected by the presence of leading silence. In this case the manipulated areas have higher misalignment scores, a frame-level evaluation yielding $77.7\\%$ AUC. We do notice that there are other regions in the video that get a high misalignment score. This happens because the separation and re-composition of the audio and video streams can introduce desynchronizations between the two even in areas that were not intentionally manipulated.", + "bbox": [ + 88, + 407, + 485, + 621 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Comparison on the official AV-Deepfake1M test", + "text_level": 1, + "bbox": [ + 89, + 633, + 483, + 648 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To further verify our conclusions, we evaluate our approach on the official test set of the AV-Deepfake1M dataset for which the labels are withheld. The main difference between the validation set, used in the previous section, and the test set are the subject identities: the test subjects are unseen at train time. We obtain the results by submitting our prediction to the official competition server. Our results are compared against those of other methods based on the values reported in the AV-Deepfake1M paper [6].", + "bbox": [ + 89, + 657, + 483, + 794 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The results shown in Table 3 are in line to those obtained on the validation split (Table 2). Compared to the rest of the methods, AVH-Align has the highest performance (85.24% AUC) even if has seen only real data at training. The silence classifier described in Sec. 3 obtains again a performance of over 98% AUC, indicating that the test set suffers from the same spurious feature as the training and validation splits.", + "bbox": [ + 89, + 795, + 483, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/3d115afac7bbfe0f7741c3eabbd35c0486407845981daa2ff5871a1113a46b2d.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsModalityAUC
Segment-level methods
Meso4 [1]V54.53
MesoInception4 [1]V57.16
MDS [10]AV56.57
MARLIN [5]V58.03
Frame-level methods
Meso4 [1]V63.05
MesoInception4 [1]V64.04
Xception [9]V68.68
EfficientViT [12]V65.51
Ours
AVH-AlignAV85.24
AVH-Align/supAV99.90
Silence classifierA98.44
", + "bbox": [ + 514, + 405, + 908, + 597 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Results on the official AV-Deepfake1M test set. Except the methods that have access to the leading silence bias (shown in red), AVH-Align performs best with $85.24\\%$ AUC. Having access to the leading silence bias, the silence classifier and AVH-Align/sup show almost perfect performance of over $98\\%$ AUC.", + "bbox": [ + 511, + 607, + 906, + 678 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Other methods that are trained at video level do not have access to the silence bias, and hence show only modest performance. Instead, we expect audio-based methods to use the bias and return over optimistic results. This is indeed the case for the supervised variant, AVH-Align/sup, which returns an AUC of $99.90\\%$ on the official test set.", + "bbox": [ + 511, + 705, + 906, + 797 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Further analysis", + "text_level": 1, + "bbox": [ + 511, + 808, + 678, + 825 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Ablation of AVH-Align components. Here we analyze the impact of various design choices on the performance of the AVH-Align method. Specifically, we investigate the following: 1) removing feature normalization; 2) reducing the", + "bbox": [ + 511, + 839, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "18821", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/3ba86d5b55e8be0b3a03c57ca69a06f5ef149460ad7bc55b841334d9b97bc606.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ConfigurationFAVCAV1MAVLips
AVH-Align94.685.986.3
1... Feature norm: ×92.484.886.1
2... Train size: 8.7k89.182.583.0
3... Score pooling: mean97.476.888.8
4... AV-HuBERT: LRS384.860.380.1
5... Alignment net: Linear31.551.943.4
", + "bbox": [ + 91, + 88, + 482, + 188 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "training set size from 45,000 to 8,782 samples (the real ones from the AVH-Align/sup train set); 3) using mean score pooling instead of log-sum-exp; 4) extracting features with the AV-HuBERT checkpoint pretrained on LRS3 only instead of LRS3 and VoxCeleb2; 5) using a linear layer instead of MLP. Results are shown in Table 4. All ablations lead to a decrease in performance with the exception of mean score pooling for FakeAVCeleb and AVLips. This is expected, however, since fake videos in FakeAVCeleb and AVLips are manipulated at every frame, while AVDeepfake1M has only locally-manipulated fake videos.", + "bbox": [ + 88, + 266, + 482, + 434 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Impact of VoxCeleb2 pretraining. Both FakeAVCeleb and AV1M datasets use VoxCeleb2 as their source dataset. Can there be a data leakage from AV-HuBERT VoxCeleb2 features to these datasets? We evaluate on a dataset that is not using VoxCeleb2, AVLips [31]. We observe that VoxCeleb2 features perform better even on this dataset, suggesting that these features are simply stronger than LRS3 features. This is may be explained by the size of VoxCeleb data (about four times larger than LRS3) and is in line with the results on other downstream tasks (e.g., lip reading [49]).", + "bbox": [ + 88, + 439, + 482, + 592 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Analysis of AVH-Align/sup architecture. We investigate how a simpler alignment network impacts the supervised model. To this end, we train a linear layer on top of the AV-HuBERT representations instead of the MLP network. In Table 5, we observe that unlike the unsupervised case where a linear layer was too weak to learn, here a linear layer performs as well or even better than the MLP. The largest difference between the MLP and then linear layer is observed when training on FakeAVCeleb and testing on AV-Deepfake1M. With a linear layer, the performance drop from untrimmed to trimmed AV-Deepafake1M is $19.3\\%$ (from $85.9\\%$ to $66.6\\%$ AUC), while with an MLP, the drop is $5.4\\%$ (from $69.0\\%$ to $63.6\\%$ AUC).", + "bbox": [ + 89, + 598, + 482, + 794 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Discussion", + "text_level": 1, + "bbox": [ + 89, + 806, + 204, + 821 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We discuss how our conclusions fit into the broader scope of deepfake detection.", + "bbox": [ + 89, + 833, + 482, + 863 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "A different evaluation paradigm. We observed near-perfect performance on datasets such as FakeAVCeleb and", + "bbox": [ + 89, + 869, + 482, + 901 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/6cffe6b10e144d030fbba8cd5bef62604f942caaeb58878bc015733d39d40837.jpg", + "table_caption": [ + "Table 4. Ablation of the main components of AVH-Align in terms of AUC. Mean score pooling helps when videos are fully generated (FAVC, AVLips). Other ablations degrade the performance." + ], + "table_footnote": [], + "table_body": "
ArchitectureTrain dataFakeAVCelebAV-Deepfake1M
Trim: XTrim: √Trim: XTrim: √
LinearFakeAVCeleb99.699.585.966.6
MLPFakeAVCeleb99.299.269.063.6
LinearAV-Deepfake1M76.269.799.983.2
MLPAV-Deepfake1M77.570.8100.083.1
", + "bbox": [ + 517, + 88, + 903, + 183 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. Architecture analysis of AVH-Align/sup. In most cases, the linear and MLP architectures perform similarly.", + "bbox": [ + 511, + 194, + 903, + 222 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "AV-Deepfake1M. The reasons are that the same generative models were used in both train and test, as well as the unintentional leakage of spurious features. The latter also affects cross-dataset performance, e.g., training on FakeAVCeleb and testing on AV-Deepfake1M still yields strong performance. In a realistic scenario, however, we are not given access to the generators, nor to the pre-processing or post-processing steps, which introduce unintentional alterations. Each deepfake released in the wild may be created through a different set of tools, unknown apriori. As such, we believe that another way of gauging the progress of deepfake detection is by refraining the training to real data only.", + "bbox": [ + 511, + 253, + 906, + 435 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Circumventing shortcuts through alignment. Modeling real data is a way to avoid shortcuts in the data. A different direction is taken in the very recent work of Rajan et al. [43]. The idea is to generate fake samples by reconstructing real samples through a generator of choice. This approach ensures that fakes and reals are aligned, avoiding spurious features. On the other hand, this forces the model to focus on the fingerprint of the generator, hindering generalizability. Moreover, by equating fake to a fingerprint, we become susceptible to laundering [32], lose the ability to localise [33, 51] and distinguish benign fakes (e.g., superresolution images [57]). We believe that both directions (alignment and real data) are complementary perspectives that should be considered to tackle the multi-faceted problem of deepfake detection.", + "bbox": [ + 511, + 444, + 908, + 671 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusions", + "text_level": 1, + "bbox": [ + 511, + 691, + 638, + 707 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper we exposed a previously unknown bias in two widely adopted audio-video deepfake detection datasets—a leading silence in fake videos. We showed that models exposed to this bias during training are prone to rely on it when deciding the authenticity of a video, thus displaying overly optimistic results. As an alternative, we propose to shift the learning paradigm towards unsupervised learning on real data only. Specifically, we find that self-supervised audio-video representations coupled with an alignment network trained on real videos produce more robust and consistent results. Our work raises awareness regarding dataset design and evaluation of deepfake detection.", + "bbox": [ + 511, + 719, + 908, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "18822", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 91, + 90, + 250, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported in part by the EU Horizon projects AI4TRUST (No. 101070190) and ELIAS (No. 101120237), and by CNCS-UEFISCDI (PN-IV-P7-7.1-PTE-2024-0600).", + "bbox": [ + 89, + 116, + 483, + 176 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 194, + 187, + 210 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Darius Afchar, Vincent Nozick, Junichi Yamagishi, and Isao Echizen. MesoNet: A compact facial video forgery detection network. In WIFS, page 1-7, 2018. 2, 7", + "[2] Matyas Bohacek and Hany Farid. Lost in translation: Lipsync deepfake detection from audio-video mismatch. In CVPR, pages 4315-4323, 2024. 2", + "[3] Stefano Borzi, Oliver Giudice, Filippo Stanco, and Dario Allegra. Is synthetic voice detection research going into the right direction? In CVPRW, pages 71-80, 2022. 1", + "[4] Zhixi Cai, Shreya Ghosh, Abhinav Dhall, Tom Gedeon, Kalin Stefanov, and Munawar Hayat. Glitch in the matrix: A large scale benchmark for content driven audio-visual forgery detection and localization. Comput. Vis. Image Underst., 236:103818, 2023. 1, 2, 4", + "[5] Zhixi Cai, Shreya Ghosh, Kalin Stefanov, Abhinav Dhall, Jianfei Cai, Hamid Rezatofighi, Reza Haffari, and Munawar Hayat. MARLIN: Masked autoencoder for facial video representation learning, 2023. 2, 7", + "[6] Zhixi Cai, Shreya Ghosh, Aman Pankaj Adatia, Munawar Hayat, Abhinav Dhall, Tom Gedeon, and Kalin Stefanov. AV-Deepfake1M: A large-scale LLM-driven audio-visual deepfake dataset, 2024. 1, 2, 7", + "[7] Edresson Casanova, Julian Weber, Christopher Dane Shulby, Arnaldo Candido Junior, Eren Gölgge, and Moacir A. Ponti. YourTTS: Towards zero-shot multi-speaker TTS and zero-shot voice conversion for everyone. In ICML, pages 2709-2720, 2022. 3", + "[8] Lucy Chai, David Bau, Ser-Nam Lim, and Phillip Isola. What makes fake images detectable? understanding properties that generalize. In ECCV, pages 103-120, 2020. 1", + "[9] François Chollet. Xception: Deep learning with depthwise separable convolutions. In CVPR, pages 1800-1807, 2017. 7", + "[10] Komal Chugh, Parul Gupta, Abhinav Dhall, and Ramanathan Subramanian. Not made for each other-audiovisual dissonance-based deepfake detection and localization. In ACMM, pages 439-447, 2020. 4, 7", + "[11] J. S. Chung, A. Nagrani, and A. Zisserman. VoxCeleb2: Deep speaker recognition. In Interspeech, 2018. 2", + "[12] Davide Alessandro Coccomini, Nicola Messina, Claudio Gennaro, and Fabrizio Falchi. Combining EfficientNet and vision transformers for video deepfake detection. In ICIAP, pages 219-229, 2022. 7", + "[13] Davide Cozzolino, Giovanni Poggi, Riccardo Corvi, Matthias Nießner, and Luisa Verdoliva. Raising the bar of AI-generated image detection with CLIP. In CVPR, pages 4356-4366, 2024. 4" + ], + "bbox": [ + 93, + 219, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Davide Cozzolino, Giovanni Poggi, Matthias Nießner, and Luisa Verdoliva. Zero-shot detection of AI-generated images. In ECCV, pages 54-72, 2024. 2", + "[15] Alexandre Défossez, Gabriel Synnaeve, and Yossi Adi. Real time speech enhancement in the waveform domain. In *Interspeech*, pages 3291–3295, 2020. 3", + "[16] Brian Dolhansky, Joanna Bitton, Ben Pflaum, Jikuo Lu, Russ Howes, Menglin Wang, and Cristian Canton Ferrer. The deepfake detection challenge (dfdc) dataset. arXiv preprint arXiv:2006.07397, 2020. 4", + "[17] Chao Feng, Ziyang Chen, and Andrew Owens. Self-supervised video forensics by audio-visual anomaly detection. In CVPR, pages 10491–10503, 2023. 2, 4, 5, 6", + "[18] Alexandros Haliassos, Konstantinos Vougioukas, Stavros Petridis, and Maja Pantic. Lips don't lie: A generalisable and robust approach to face forgery detection. In CVPR, pages 5039-5049, 2021. 2", + "[19] Alexandros Haliassos, Rodrigo Mira, Stavros Petridis, and Maja Pantic. Leveraging real talking faces via self-supervision for robust forgery detection. In CVPR, pages 14950-14962, 2022. 2", + "[20] Ammarah Hashmi, Sahibzada Adil Shahzad, Chia-Wen Lin, Yu Tsao, and Hsin-Min Wang. AVTENet: Audio-visual transformer-based ensemble network exploiting multiple experts for video deepfake detection. CoRR, abs/2310.13103, 2023. 2", + "[21] Baojin Huang, Zhongyuan Wang, Jifan Yang, Jiaxin Ai, Qin Zou, Qian Wang, and Dengpan Ye. Implicit identity driven deepfake face swapping detection. In CVPR, pages 4490-4499, 2023. 2", + "[22] Ye Jia, Yu Zhang, Ron J. Weiss, Quan Wang, Jonathan Shen, Fei Ren, Zhifeng Chen, Patrick Nguyen, Ruoming Pang, Ignacio López-Moreno, and Yonghui Wu. Transfer learning from speaker verification to multispeaker text-to-speech synthesis. In NeurIPS, pages 4485-4495, 2018. 2", + "[23] Hasam Khalid, Shahroz Tariq, Minha Kim, and Simon S. Woo. FakeAVCeleb: A novel audio-video multimodal deepfake dataset. In NeurIPS Datasets and Benchmarks, 2021. 1, 2", + "[24] Jaehyeon Kim, Jungil Kong, and Juhee Son. Conditional variational autoencoder with adversarial learning for end-to-end text-to-speech. In ICML, pages 5530-5540, 2021. 3", + "[25] Pavel Korshunov and Sébastien Marcel. DeepFakes: A new threat to face recognition? Assessment and detection. CoRR, abs/1812.08685, 2018. 2", + "[26] Iryna Korshunova, Wenzhe Shi, Joni Dambre, and Lucas Theis. Fast face-swap using convolutional neural networks. In ICCV, pages 3697-3705, 2017. 2", + "[27] Christos Koutlis and Symeon Papadopoulos. DiMoDif: Discourse modality-information differentiation for audio-visual deepfake detection and localization. CoRR, abs/2411.10193, 2024. 2", + "[28] Patrick Kwon, Jaeseong You, Gyuhyeon Nam, Sungwoo Park, and Gyeongsu Chae. KoDF: A large-scale Korean deepfake detection dataset. In ICCV, pages 10744-10753, 2021. 2" + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "18823", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[29] Xiaolou Li, Zehua Liu, Chen Chen, Lantian Li, Li Guo, and Dong Wang. Zero-shot fake video detection by audio-visual consistency. CoRR, abs/2406.07854, 2024. 2", + "[30] Yachao Liang, Min Yu, Gang Li, Jianguo Jiang, Boquan Li, Feng Yu, Ning Zhang, Xiang Meng, and Weiqing Huang. SpeechForensics: Audio-visual speech representation learning for face forgery detection. In NeurIPS, 2024. 6", + "[31] Weifeng Liu, Tianyi She, Jiawei Liu, Boheng Li, Dongyu Yao, Ziyou Liang, and Run Wang. Lips are lying: Spotting the temporal inconsistency between audio and visual in lip-syncing deepfakes. In NeurIPS, pages 91131-91155, 2024. 4, 8", + "[32] Sara Mandelli, Paolo Bestagini, and Stefano Tubaro. When synthetic traces hide real content: Analysis of stable diffusion image laundering. CoRR, abs/2407.10736, 2024. 8", + "[33] Hannes Mareen, Dimitrios Karageorgiou, Glenn Van Wallendael, Peter Lambert, and Symeon Papadopoulos. TGIF: Text-guided inpainting forgery dataset. CoRR, abs/2407.11566, 2024. 8", + "[34] Francesco Marra, Diego Gragnaniello, Luisa Verdoliva, and Giovanni Poggi. Do GANs leave artificial fingerprints? In MIPR, pages 506-511, 2019. 2", + "[35] Nicolas M. Müller, Franziska Dieckmann, Pavel Czempin, Roman Canals, and Konstantin Bötttinger. Speech is silver, silence is golden: What do ASVspoof-trained models really learn? CoRR, abs/2106.12914, 2021. 1, 4", + "[36] Nicolas M. Müller, Piotr Kawa, Wei Herng Choong, Edresson Casanova, Eren Gölgge, Thorsten Müller, Piotr Syga, Philip Sperl, and Konstantin Bötttinger. MLAAD: The multilanguage audio anti-spoofing dataset. In IJCNN, pages 1-7, 2024. 1", + "[37] Yuval Nirkin, Yosi Keller, and Tal Hassner. FSGAN: Subject agnostic face swapping and reenactment. In ICCV, pages 7183-7192, 2019. 2", + "[38] Utkarsh Ojha, Yuheng Li, and Yong Jae Lee. Towards universal fake image detectors that generalize across generative models. In CVPR, pages 24480-24489, 2023. 2, 4", + "[39] Trevine Oorloff, Surya Koppisetti, Nicolò Bonettini, Dipyaraj Solanki, Ben Colman, Yaser Yacoob, Ali Shahriyari, and Gaurav Bharaj. AVFF: Audio-visual feature fusion for video deepfake detection. In CVPR, pages 27102-27112, 2024. 2, 5", + "[40] Octavian Pascu, Adriana Stan, Dan Oneata, Elisabeta Oneata, and Horia Cucu. Towards generalisable and calibrated audio deepfake detection with self-supervised representations. In Interspeech, pages 4828-4832, 2024. 4", + "[41] Alessandro Pianese, Davide Cozzolino, Giovanni Poggi, and Luisa Verdoliva. Training-free deepfake voice recognition by leveraging large-scale pre-trained models. In ACM Workshop on Information Hiding and Multimedia Security, page 289–294, 2024. 2, 4", + "[42] K R Prajwal, Rudrabha Mukhopadhyay, Vinay P. Namboodiri, and C.V. Jawahar. A lip sync expert is all you need for speech to lip generation in the wild. In ACMM, 2020. 2", + "[43] Anirudh Sundara Rajan, Utkarsh Ojha, Jedidiah Schloesser, and Yong Jae Lee. On the effectiveness of dataset alignment for fake image detection. CoRR, abs/2410.11835, 2024. 1, 8" + ], + "bbox": [ + 91, + 90, + 480, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[44] Tal Reiss, Bar Cavia, and Yedid Hoshen. Detecting deep-fakes without seeing any. CoRR, abs/2311.01458, 2023. 2", + "[45] Jonas Ricker, Denis Lukovnikov, and Asja Fischer. AEROBLADE: Training-free detection of latent diffusion images using autoencoder reconstruction error. In CVPR, pages 9130-9140, 2024. 2", + "[46] Andreas Rossler, Davide Cozzolino, Luisa Verdoliva, Christian Riess, Justus Thies, and Matthias Niessner. FaceForensics: Learning to detect manipulated facial images. In ICCV, 2019. 1", + "[47] Davide Salvi, Brian C. Hosler, Paolo Bestagini, Matthew C. Stamm, and Stefano Tubaro. TIMIT-TTS: A text-to-speech dataset for multimodal synthetic media detection. IEEE Access, 11:50851-50866, 2023. 2", + "[48] Sahibzada Adil Shahzad, Ammarah Hashmi, Yan-Tsung Peng, Yu Tsao, and Hsin-Min Wang. AV-Lip-Sync+: Leveraging AV-HuBERT to exploit multimodal inconsistency for video deepfake detection. CoRR, abs/2311.02733, 2023. 2", + "[49] Bowen Shi, Wei-Ning Hsu, Kushal Lakhotia, and Abdelrahman Mohamed. Learning audio-visual speech representation by masked multimodal cluster prediction. In ICLR, 2022. 2, 5, 8", + "[50] Kaede Shiohara and Toshihiko Yamasaki. Detecting deepfakes with self-blended images. In CVPR, pages 18699-18708, 2022. 2", + "[51] Stefan Smeu, Elisabeta Oneata, and Dan Oneata. DeCLIP: Decoding CLIP representations for deepfake localization. CoRR, abs/2409.08849, 2024. 8", + "[52] Joon Son Chung, Andrew Senior, Oriol Vinyals, and Andrew Zisserman. Lip reading sentences in the wild. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6447-6456, 2017. 4", + "[53] Hemlata Tak, Jose Patino, Massimiliano Todisco, Andreas Nautsch, Nicholas Evans, and Anthony Larcher. End-to-end anti-spoofing with RawNet2. In ICASSP, pages 6369-6373, 2021. 4", + "[54] Aäron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. CoRR, abs/1807.03748, 2018. 5", + "[55] Jiadong Wang, Xinyuan Qian, Malu Zhang, Robby T. Tan, and Haizhou Li. Seeing what you said: Talking face generation guided by a lip reading expert. In CVPR, pages 14653-14662, 2023. 3", + "[56] Xin Wang, Junichi Yamagishi, Massimiliano Todisco, Héctor Delgado, Andreas Nautsch, Nicholas W. D. Evans, Md. Sahidullah, Ville Vestman, Tomi Kinnunen, Kong Aik Lee, Lauri Juvela, Paavo Alku, Yu-Huai Peng, Hsin-Te Hwang, Yu Tsao, Hsin-Min Wang, Sébastien Le Maguer, Markus Becker, and Zhen-Hua Ling. ASVspoof 2019: A large-scale public database of synthesized, converted and replayed speech. Comput. Speech Lang., 64:101114, 2020. 1", + "[57] Zhiyuan Yan, Taiping Yao, Shen Chen, Yandan Zhao, Xinghe Fu, Junwei Zhu, Donghao Luo, Li Yuan, Chengjie Wang, Shouhong Ding, and Yunsheng Wu. DF40: Toward next-generation deepfake detection. CoRR, abs/2406.13495, 2024.8", + "[58] Wenyuan Yang, Xiaoyu Zhou, Zhikai Chen, Bofei Guo, Zhongjie Ba, Zhihua Xia, Xiaochun Cao, and Kui Ren." + ], + "bbox": [ + 516, + 90, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "18824", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "AVoiD-DF: Audio-visual joint learning for detecting deepfake. IEEE Trans. Inf. Forensics Secur., 18:2015-2029, 2023. 2", + "[59] Ning Yu, Larry S Davis, and Mario Fritz. Attributing fake images to GANs: Learning and analyzing GAN fingerprints. In CVPR, pages 7556-7566, 2019. 2", + "[60] Yinglin Zheng, Jianmin Bao, Dong Chen, Ming Zeng, and Fang Wen. Exploring temporal coherence for more general video face forgery detection. In ICCV, pages 15024-15034, 2021. 2" + ], + "bbox": [ + 91, + 90, + 482, + 231 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "18825", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Circumventing Shortcuts in Audio-visual Deepfake Detection Datasets with Unsupervised Learning/371069ba-8973-4b9c-bb98-ecda46029053_model.json b/2025/Circumventing Shortcuts in Audio-visual Deepfake Detection Datasets with Unsupervised Learning/371069ba-8973-4b9c-bb98-ecda46029053_model.json new file mode 100644 index 0000000000000000000000000000000000000000..0dd5312c5e02967f9ed7d30c358964e3b51860df --- /dev/null +++ b/2025/Circumventing Shortcuts in Audio-visual Deepfake Detection Datasets with Unsupervised Learning/371069ba-8973-4b9c-bb98-ecda46029053_model.json @@ -0,0 +1,2268 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.132, + 0.131, + 0.868, + 0.177 + ], + "angle": 0, + "content": "Circumventing shortcuts in audio-visual deepfake detection datasets with unsupervised learning" + }, + { + "type": "text", + "bbox": [ + 0.168, + 0.203, + 0.827, + 0.24 + ], + "angle": 0, + "content": "Stefan Smeu\\*1 Dragos-Alexandru Boldisor\\*1 Dan Oneata2 Elisabeta Oneata1 1Bitdefender 2POLITEHNICA Bucharest" + }, + { + "type": "text", + "bbox": [ + 0.208, + 0.242, + 0.785, + 0.258 + ], + "angle": 0, + "content": "{ssmeu, dboldisor, eoneata}@bitdefender.com dan. oneata@gmail.com" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.292, + 0.327, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.323, + 0.485, + 0.611 + ], + "angle": 0, + "content": "Good datasets are essential for developing and benchmarking any machine learning system. Their importance is even more extreme for safety critical applications such as deepfake detection—the focus of this paper. Here we reveal that two of the most widely used audio-video deepfake datasets suffer from a previously unidentified spurious feature: the leading silence. Fake videos start with a very brief moment of silence and, on the basis of this feature alone, we can separate the real and fake samples almost perfectly. As such, previous audio-only and audio-video models exploit the presence of silence in the fake videos and consequently perform worse when the leading silence is removed. To circumvent latching on such an unwanted artifact and possibly other unrevealed ones, we propose a shift from supervised to unsupervised learning by training models exclusively on real data. We show that by aligning self-supervised audio-video representations we remove the risk of relying on dataset-specific biases and improve robustness in deepfake detection." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.64, + 0.222, + 0.657 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.666, + 0.483, + 0.771 + ], + "angle": 0, + "content": "Manipulated videos represent a threat to society as they have the potential of misleading people into believing actors with malicious intents. By spreading misinformation on social media platforms, people may be exposed to scams (e.g., identity theft operations), conspiracy theories and political misinformation. Therefore, deepfake detection methods are essential tools on a global scale." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.772, + 0.483, + 0.879 + ], + "angle": 0, + "content": "The progress in automated deepfake detection is fueled by the datasets developed by the research community. Good quality data is essential for both training and benchmarking the progress of these methods. In recent years, numerous datasets have been proposed. Among the audio-video deepfake datasets, there have been released datasets that alter both streams [6, 23] or only one [36, 46, 56]; datasets with" + }, + { + "type": "image", + "bbox": [ + 0.546, + 0.293, + 0.875, + 0.427 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.439, + 0.908, + 0.509 + ], + "angle": 0, + "content": "Figure 1. Audio-visual deepfake detection datasets have a silence bias: fake samples start with a brief moment of silence, which is not the case for real samples. Here we show the first \\(62.5\\mathrm{ms}\\) of the audio waveform for a real and the corresponding fake sample from the AV-Deepfake1M dataset [6]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.537, + 0.84, + 0.553 + ], + "angle": 0, + "content": "full-video [23] or only local [4, 6] manipulations." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.553, + 0.907, + 0.749 + ], + "angle": 0, + "content": "However, care has to be taken to ensure a good deepfake detection dataset. Any asymmetry in the preparation of fakes and reals can result in biases that correlate spuriously with the groundtruth label. For example, in the image domain, different preprocessing pipelines [8] or types of resizing [43] between real and fake samples paint an overly optimistic picture. In the audio domain, the very popular ASVspoof19 dataset [56] has been shown to leak information about labels in the form of silence duration [35] or bitrate information [3]. Since many of the deepfake detection models are high capacity, they can easily use such artifacts as shortcuts for learning, consequently greatly impacting their ability to generalize in a real case scenario." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.909, + 0.902 + ], + "angle": 0, + "content": "In this work, we first expose a bias present in two widely-adopted audio-video deepfake detection datasets, FakeAVCeleb [23] and AV-Deepfake1M [6]: a short moment of silence present at the beginning of manipulated videos (see Figure 1). Based on this information, a very simple silence classifier can reach a near-perfect performance of over \\(98\\%\\) on both FakeAVCeleb and AV-Deepfake1M. Moreover, we observe that the silence distribution is similar across datasets, implying that prior methods may have potentially over estimated the generalization" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.22, + 0.901 + ], + "angle": 0, + "content": "*Equal contribution." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18815" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.329, + 0.106 + ], + "angle": 0, + "content": "performance based on this shortcut." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.107, + 0.482, + 0.303 + ], + "angle": 0, + "content": "Introducing such biases when creating the datasets is unavoidable and there might be others that are more subtle and harder to reveal. To circumvent this problem, we show that one solution is a shift in learning paradigm, from the more common supervised setup to the unsupervised one [17, 44]. Removing fake samples from training eliminates the focus on the asymmetries induced by spurious artifacts. Moreover, this also has the potential to improve generalization among different manipulation techniques, since supervised detection methods tend to over-rely on generator-specific fingerprints [34, 59]. By limiting to real data, we can also naturally leverage self-supervised representations, which have been shown to improve generalization [38]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.304, + 0.482, + 0.439 + ], + "angle": 0, + "content": "To this end, we propose an approach, named AVH-Align, that learns a frame-level audio-video alignment score on top of AV-HuBERT features [49]. Since tampered videos are expected to have greater audio-video desynchronizations, this alignment score can effectively differentiate fake and real videos. We show that this approach is robust with respect to the identified shortcut of leading silence, and also outperforms other methods that do not use the silence bias, while not seeing any fake samples at training time." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.44, + 0.482, + 0.56 + ], + "angle": 0, + "content": "To summarize, our work makes the following contributions: 1. We expose a previously unknown spurious feature in two of the most widely adopted deepfake detection datasets. 2. We analyze the impact of this shortcut on various state-of-the-art models. 3. We show that a way to mitigate such shortcuts is by training on real data only and we introduce a new method in this direction. Our code is available at: https://github.com/bit-ml/AVH-Align." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.574, + 0.228, + 0.59 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.607, + 0.483, + 0.864 + ], + "angle": 0, + "content": "Audio-video deepfake detection. Many approaches for deepfake detection on videos have focused on the visual stream of information only [1, 5, 18, 19, 21, 50, 60]. But with the recent introduction of audio-visual datasets, (FakeAVCeleb [23], Deepfake TIMIT [25, 47], KODF [28], LAV-DF [4], AV-Deepfake-1M [6]), more research has shifted towards models that exploit both audio and video cues [17, 20, 27, 39, 48, 58]. An emerging trend in this direction is the use of pretrained representations in a self-supervised way [17, 39]. But, different from our approach, these methods train the representations from scratch and use them as a first step in a more elaborated pipeline: supervised classification [39] or anomaly detection [17]. There are also works that similarly to us exploit the pretrained audio-video AV-HuBERT model [49] to extract representations [20, 48], but all of those methods are trained in the fully supervised learning paradigm." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Unsupervised deepfake detection. To make detection more generalizable across generators, a new direction is" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.424 + ], + "angle": 0, + "content": "to depart from the supervised paradigm and resort only to real data. There are two main classes of such unsupervised approaches: methods that rely on consistency checks and methods that treat the problem as an anomaly detection one. Among the first class, prior work proposed to verify that audio and visual streams align at a semantic level, for example, from the point of view of spoken content [2, 29], or at a representation level, based on the alignment of audio and video features [17, 44]. Consistency checks have also been used with respect to the identity of a speaker (comparing a query sample against real audio [41] or real images [44] of the target speaker) or between image and text modalities [44]. For the second class, deepfake detection as anomaly detection, recent work in the image domain has used the reconstruction loss of the query image to tell whether it is anomalous [14, 45]; for example, Ricker et al. [45] make the observation that images generated by a latent diffusion model (LDM) are easier to reconstruct by the LDM than real images. For the audio-visual domain, Feng et al. [17] use both classes of approaches: they use consistency checks to estimate synchronization between the two streams and then flag anomalies using density estimation." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.44, + 0.833, + 0.456 + ], + "angle": 0, + "content": "3. Silence bias in audio-video datasets" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.466, + 0.905, + 0.587 + ], + "angle": 0, + "content": "In this section, we show that two popular datasets (Sec. 3.1) have a silence bias. We analyze its behavior and show that a simple classifier based on the leading silence alone can obtain almost perfect separation between fake and real samples (Sec. 3.2). This implies that the performance of prior work is susceptible to have been overestimated. For this reason we analyze its impact on various audio and audio-visual methods (Sec. 3.3)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.598, + 0.615, + 0.612 + ], + "angle": 0, + "content": "3.1. Datasets" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.622, + 0.905, + 0.773 + ], + "angle": 0, + "content": "We consider two audio-visual datasets in our analysis: FakeAVCeleb [23] and AV-Deepfake1M [6]. They distinguish mainly in the fact that the first contains fully-generated video sequences, while the second contains partially-manipulated sequences. Both are based on the VoxCeleb2 dataset [11], which consists of YouTube audio-video of celebrities. Apart from the real samples—real video real audio (RVRA)—both datasets include three types of fake videos: real video fake audio (RVFA), fake video real audio (FVRA), and fake video fake audio (FVFA)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.907, + 0.903 + ], + "angle": 0, + "content": "FakeAVCeleb contains 500 real videos from VoxCeleb2 and 19.5k fake videos (10k FVFA, 9k FVRA, 500 RVFA). The fake visual content was generated with face swapping methods (Faceswap [26] and FSGAN [37]) or the Wav2Lip lipSyncing approach [42]. The fake audio content was generate with the voice cloning tool SV2TTS [22]. The dataset is diverse across age groups, genders, races, as well as with respect to the number of subjects in a single video, their" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18816" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.089, + 0.482, + 0.231 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.243, + 0.485, + 0.3 + ], + "angle": 0, + "content": "Figure 2. Normalized distribution plots of the leading silence duration for real and fake videos in the FakeAVCeleb (left) and AV-Deepfake1M (right) datasets. The fake samples start with 25–30 ms of silence." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.327, + 0.483, + 0.373 + ], + "angle": 0, + "content": "placement, the visual and audio quality. In our experiments, we split the dataset in \\(70\\%\\) (train and validation) and \\(30\\%\\) (test). This split is kept consistent across all experiments." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.38, + 0.484, + 0.653 + ], + "angle": 0, + "content": "AV-Deepfake1M is a large scale dataset, which consists of over one million videos and 2k subjects. As opposed to the FakeAVCeleb dataset, the manipulations here are local and consists of word-level replacements, insertions and deletions. The text manipulations are generated with the ChatGPT large language model. The fake video content is generated with the lipSyncing method TalkLip [55], while the fake audio content is generated with the VITS [24] or YourTTS [7] methods. The authors ensure that the synthesized words share the same background noise with the full audio, by first extracting the audio noise with the Denoiser method [15] and then adding it to the synthesized words. The dataset is originally split into train, validation and test, with the test split having a different set of speakers than those encountered in the train and validation splits. For our experiments, we select training and validation samples from the original training split, and evaluate on 10k samples from the original validation split or on the full official test set." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.664, + 0.33, + 0.681 + ], + "angle": 0, + "content": "3.2. Analysis of leading silence" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.688, + 0.483, + 0.809 + ], + "angle": 0, + "content": "We start by analyzing the silence distribution of the real and fake samples in the two considered datasets. We define the duration of the leading silence as the moment when the magnitude of the audio exceeds a certain threshold \\(\\tau\\). For this experiment, we select this threshold to be \\(5 \\cdot 10^{-4}\\), but as we will shortly see, the results are robust to its choice. We carry this analysis only on the real (RVRA) and fully fake (FVFA) videos from each dataset's test set." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.484, + 0.901 + ], + "angle": 0, + "content": "The results are shown in Figure 2. We observe that the real videos start with noise, while the fake samples have a leading silence of around \\(25 - 30\\mathrm{ms}\\). The silence duration of fake samples is similar for both datasets, although the distribution is much sharper for the AV-Deepfake1M dataset. If we were to rank the samples based on this feature we" + }, + { + "type": "image", + "bbox": [ + 0.515, + 0.088, + 0.907, + 0.231 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.24, + 0.908, + 0.283 + ], + "angle": 0, + "content": "Figure 3. Left: The impact of the silence threshold on the leading silence classifier. Right: The impact of the leading duration on the maximum amplitude classifier." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.311, + 0.907, + 0.34 + ], + "angle": 0, + "content": "would obtain an area under curve of the receiver operating characteristic curve (AUC) of over \\(98\\%\\) for either datasets." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.348, + 0.909, + 0.456 + ], + "angle": 0, + "content": "What counts as silence? For the previous experiment we have considered that silence is the signal that has an amplitude lower than \\(\\tau = 5 \\cdot 10^{-4}\\). We investigate how sensitive the performance is to this threshold. We vary \\(\\tau\\) across a grid of values and show the results in Figure 3 (left). We observe that the results are strong as long as this threshold is small enough." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.462, + 0.909, + 0.613 + ], + "angle": 0, + "content": "A different perspective: Maximum amplitude. Instead of looking at the silence duration, we can alternatively measure the maximum amplitude in the first \\(\\delta\\) seconds of the audio. Figure 3 shows the AUC obtained by ranking the samples based on this feature for various leading durations \\(\\delta\\). We observe that this alternative measure also yields strong results (over \\(98\\%\\) AUC), and that the optimum is obtained for a \\(\\delta\\) of around \\(30~\\mathrm{ms}\\), in line with the previous experiment. As we extend the window, the performance gets closer to random chance, AUC of \\(50\\%\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.621, + 0.909, + 0.788 + ], + "angle": 0, + "content": "Other biases: Volume and trailing silence. If we extend the duration \\(\\delta\\) to cover the entire audio, we get an estimate of the maximum volume of the audio. This feature yields an AUC of \\(67.6\\%\\) on FakeAVCeleb, which is much less than \\(98.4\\%\\), but still over random chance (50% AUC). Similarly, we have also investigated the trailing silence (the silence at the end of the audio) and have obtained AUC of over \\(99\\%\\) for FakeAVCeleb. AV-Deepfake1M is less sensitive to these other biases, with values closer to \\(50\\%\\). While these biases are not as consistent as the leading silence, they are still problematic." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Why do fake samples have a leading silence? Given that we do not have access to generation process of the two datasets, it is challenging to pinpoint the exact reason for the occurrence of the leading silence bias. However, we speculate that this happens when the audio may be slightly shorter than the video counterpart. Note that this is a different reason from the silence observed in audio only datasets" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18817" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.091, + 0.482, + 0.182 + ], + "angle": 0, + "content": "[35]. There the real audios had a leading silence, while the synthesized speech was silence-free. In the case of audio-video datasets it might be challenging to completely avoid this problem, but an easy solution (for this particular bias) is to trim the leading silence. This is what we do in the next experiments." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.189, + 0.483, + 0.371 + ], + "angle": 0, + "content": "Biases in other datasets. We consider three more datasets: LAV-DF [4], AVLips [31], DFDC [16]. We find that LAV-DF has a slight trailing silence bias (65% AUC), but it does not exhibit the other two biases (leading silence and volume). In the case of AVLips and DFDC we did not identify any of the three biases. This may happen because the samples in these datasets have real audio and the unmodified audio matches the visual sequence more precisely. This is confirmed by a fine-grained analysis on FakeAVCeleb, where we observe that on the FVRA split the biases are diminished (62.4% AUC for lead silence), while they persist on the RVFA split (100% AUC for lead silence)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.38, + 0.295, + 0.396 + ], + "angle": 0, + "content": "3.3. Impact on prior work" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.403, + 0.483, + 0.476 + ], + "angle": 0, + "content": "We have shown that real and fake samples differ in terms of the leading silence. This is a simple feature which could be learned by the high-capacity neural networks. Here we investigate whether that is indeed the case for three existing methods:" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.478, + 0.483, + 0.539 + ], + "angle": 0, + "content": "- RawNet2 [53], which is an audio-only method that operates on the raw waveform. Its architecture sequences sinc layers, convolutions, gated recurrent units and fully connected layers." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.539, + 0.483, + 0.599 + ], + "angle": 0, + "content": "- MDS [10] (modality dissonance score), which is an audio-visual method that estimates the mismatch between audio and video segments. The score is computed as the distance between audio and visual features." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.6, + 0.483, + 0.674 + ], + "angle": 0, + "content": "- AVAD [17] (audio-visual anomaly detection), which is an audio-visual method trained on real data only. The approach has two steps: first, it estimates the desynchronization between audio and video; then it estimates whether these patterns are typical of real data or anomalous." + }, + { + "type": "list", + "bbox": [ + 0.092, + 0.478, + 0.483, + 0.674 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.675, + 0.483, + 0.765 + ], + "angle": 0, + "content": "We train the RawNet2 and MDS methods on both datasets using the code provided by the authors. For the AVAD we use the provided checkpoint (trained on LRS [52]) and do not retrain it since training code is not available. As a baseline we include our leading silence duration classifier described in the previous section." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.766, + 0.483, + 0.901 + ], + "angle": 0, + "content": "We evaluate all methods in two settings: on the corresponding evaluation set and on a trimmed version of the same evaluation set. For the trimmed version, we discard the beginning of videos such that each of them starts with the first non-silent segment. We find the duration to discard by first computing the leading silence (using the same \\(5 \\cdot 10^{-4}\\) threshold on the magnitude) and then rounding it up to the nearest multiple of \\(1 / \\text{FPS}\\) (the reciprocal of the video's frame rate); this rounding ensures that the audio and" + }, + { + "type": "table", + "bbox": [ + 0.515, + 0.089, + 0.906, + 0.184 + ], + "angle": 0, + "content": "
MethodMod.FakeAVCelebAV-Deepfake1M
Trim: XTrim: √Trim: XTrim: √
Silence classifierA98.454.8 ↓ 43.698.250.6 ↓ 47.6
RawNet2 [53]A99.997.3 ↓ 2.699.988.1 ↓ 11.8
MDS [10]AV90.473.8 ↓ 16.699.254.9 ↓ 44.3
AVAD [17]AV95.295.2 ≈ 0.052.952.9 ≈ 0.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.512, + 0.194, + 0.905, + 0.248 + ], + "angle": 0, + "content": "Table 1. The impact of leading silence on the performance (area under the receiver operator characteristic curve; AUC) for three existing deepfake detection methods and the silence classifier. Results are shown for RVRA-FVFA subsets." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.28, + 0.905, + 0.31 + ], + "angle": 0, + "content": "video channels remain synchronized. For training, we do not perform trimming, but use the original dataset." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.312, + 0.905, + 0.401 + ], + "angle": 0, + "content": "The results are shown in Table 1. First, we observe that the audio-based RawNet2 is the top performing method. Moreover, we see that the leading silence classifier is generally better than the two other approaches: MDS and AVAD. The relatively stable performance of AVAD indicates that this method dose not latch on the silence information." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.403, + 0.906, + 0.69 + ], + "angle": 0, + "content": "In terms of the impact of the leading silence, we observe that removing it through trimming affects methods differently. As expected, the silence classifier is affected the most and its performance drops down to random chance on the trimmed data. RawNet2 is not as affected on the FakeAVCeleb, presumably because there is still enough information throughout the signal, but it suffers a larger hit on the more challenging AV-DeepFake1M, which has only partially-manipulated samples. MDS is affected by the leading silence bias on both FakeAVCeleb and AV-Deepfake1M, though significantly stronger on AV-Deepfake1M. AVAD is the most robust method, since this approach does not explicitly model the silence information. Finally, we notice that the RawNet2 audio-only method is best even on the trimmed settings. This is noteworthy since it is the first time an audio-only method has been applied to these datasets. This suggests that the audio stream is an important source of information that is often overlooked by prior work." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.708, + 0.887, + 0.725 + ], + "angle": 0, + "content": "4. Modeling real data for deepfake detection" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.906, + 0.901 + ], + "angle": 0, + "content": "The previous section indicated that audio-visual datasets exhibit a silence bias which can easily be exploited. We want to develop a method that is robust to this bias (and possibly other uncovered ones) and still performs well. We have seen that models trained only on real data [17] are promising in being robust to the silence shortcut, but the performance was modest. To further improve them we propose to build on top of audio-focused self-supervised features. Self-supervised features have shown strong generalization for both visual [13, 38] and audio [40, 41] deepfake detection. We choose audio-focused self-supervised features be" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "18818" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.152 + ], + "angle": 0, + "content": "cause the audio models showed strong performance in the previous section. Note that we cannot rely on audio-only models because there are cases where manipulations appear only in the visual domain (the fake video, real audio case)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.163, + 0.188, + 0.178 + ], + "angle": 0, + "content": "4.1. Method" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.187, + 0.485, + 0.309 + ], + "angle": 0, + "content": "We propose a method that aligns AV-HuBERT [49] features on real data. First, we extract audio and visual frame-level features with a pretrained AV-HuBERT model. Then, on top of these features we learn a network to better align them. The alignment network is learnt on real samples by matching each video frame to its corresponding audio frame [17]. We call our method AVH-Align (AV-HuBERT Aligned) and show its depiction in Figure 4." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.316, + 0.484, + 0.527 + ], + "angle": 0, + "content": "Self-supervised features. We use AV-HuBERT to represent both the audio and visual content of a video. AV-HuBERT is a Transformer network trained in a self-supervised way to predict iteratively refined centroids from masked inputs. The features extracted by AV-HuBERT encode audio information as proved by its strong performance on tasks such as lip reading or noisy audio-visual speech recognition. We extract the audio and visual representations independently: to extract audio features we mask the visual input, to extract visual features we mask the audio input. For a video, we obtain audio features \\(\\mathbf{a}_i\\) and visual features \\(\\mathbf{v}_i\\) features for each time step \\(i\\). Both representations are 1024 dimensional and have a temporal resolution of 25 frames per second." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.535, + 0.484, + 0.625 + ], + "angle": 0, + "content": "Alignment network. To tell how well the audio and visual features match each other, we first independently L2 normalize the feature vectors and then feed them into a network \\(\\Phi\\). This is implemented as a multi-layer perceptron (MLP) over the concatenated normalized audio and visual features:" + }, + { + "type": "equation", + "bbox": [ + 0.21, + 0.638, + 0.483, + 0.657 + ], + "angle": 0, + "content": "\\[\n\\Phi_ {i j} = \\operatorname {M L P} \\left(\\left[ \\mathbf {a} _ {i}; \\mathbf {v} _ {j} \\right]\\right). \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.668, + 0.484, + 0.744 + ], + "angle": 0, + "content": "The MLP has four layers, which progressively reduce the feature dimensionality, with layers mapping from the AV-HuBERT feature size of 1024 to 512, 256, 128, and finally to a single output. Each hidden layer includes Layer Normalization and ReLU activations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.752, + 0.484, + 0.798 + ], + "angle": 0, + "content": "Loss function. To learn the alignment network \\(\\Phi\\) we maximize the probability of an audio frame \\(\\mathbf{a}_i\\) to match the corresponding video frame \\(\\mathbf{v}_i\\); this probability is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.182, + 0.809, + 0.483, + 0.845 + ], + "angle": 0, + "content": "\\[\np \\left(\\mathbf {v} _ {i} \\mid \\mathbf {a} _ {i}\\right) = \\frac {\\exp \\Phi_ {i i}}{\\sum_ {k \\in \\mathcal {N} (i)} \\exp \\Phi_ {i k}}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.902 + ], + "angle": 0, + "content": "where \\(\\mathcal{N}(i)\\) represents the temporal neighborhood around the frame \\(i\\). In our case \\(\\mathcal{N}(i)\\) contains the 30 neighboring frames around \\(i\\). We define the final loss as the negative" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.818, + 0.107 + ], + "angle": 0, + "content": "probabilities averaged across the entire video:" + }, + { + "type": "equation", + "bbox": [ + 0.619, + 0.117, + 0.906, + 0.158 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = - \\frac {1}{T} \\sum_ {i = 1} ^ {T} \\log p (\\mathbf {v} _ {i} | \\mathbf {a} _ {j}). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.167, + 0.906, + 0.199 + ], + "angle": 0, + "content": "This loss is similar to the contrastive loss InfoNCE [54], which was also used for deepfake detection [17, 39]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.205, + 0.906, + 0.31 + ], + "angle": 0, + "content": "Inference. Once \\(\\Phi\\) is learned, we can estimate the fakeness score as the negative of \\(\\Phi_{ii}\\) for each audio-video frame pair in a video; aligned audio-video frame pairs should yield lower scores. Then we compute an overall alignment score for the entire video by pooling the per-frame scores using the log-sum-exp function (a smooth version of the max function)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.318, + 0.907, + 0.485 + ], + "angle": 0, + "content": "Supervised variant. To understand the impact of the silence bias on the standard supervised learning paradigm, we design a supervised variant, AVH-Align/sup, that uses the same features and alignment network as AVH-Align, but a classification loss. In this setup we assume that apart from the real videos, we also have access to fake videos in the training set, with corresponding labels \\( y \\). To obtain a per-video fakeness probability, we first pool the negated per-frame scores \\( \\Phi_{ii} \\) with the log-sum-exp function and then we apply the sigmoid function \\( \\sigma \\). Finally, we optimize the binary cross-entropy (BCE) loss:" + }, + { + "type": "equation", + "bbox": [ + 0.573, + 0.495, + 0.907, + 0.513 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\sup } = \\operatorname {B C E} \\left(\\sigma \\left(\\log \\operatorname {s u m} \\exp (- \\Phi_ {i i})\\right), y\\right). \\tag {4}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.531, + 0.702, + 0.548 + ], + "angle": 0, + "content": "4.2. Experimental setup" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.562, + 0.907, + 0.85 + ], + "angle": 0, + "content": "Datasets and metrics. We conducted our experiments on the two datasets introduced in Sec. 3.1: FakeAVCeleb and AV-Deepfake1M. For FakeAVCeleb, we use \\(70\\%\\) of the dataset for training and validation, and the rest of \\(30\\%\\) for testing. For AV-Deepfake1M, we evaluate on 10k samples from the original validation set, but for the best performing models we also report results on the official withheld test set. To train AVH-Align, we use 50k real samples from the training set of AV-Deepfake1M (45k samples for training and 5k for validation). Since AV-Deepfake1M is based on VoxCeleb2, these samples originate from VoxCeleb2. To train the supervised variant, AVH-Align/sup, on AV-Deepfake1M, we select the same number of samples (45k for training and 5k for validation), but this time coming from both the real and fake classes. For both datasets, we evaluate video-level detection. Following prior work, we report results in terms of the area under the receiver operator characteristic curve (AUC) and average precision (AP), with the fake class being the positive class." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Implementation details. For AVH-Align we use a learning rate scheduler with a patience of 5 epochs and a factor of 0.1, with a starting learning rate of \\(10^{-5}\\). The training is" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18819" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.212, + 0.089, + 0.585, + 0.233 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.593, + 0.089, + 0.761, + 0.233 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.245, + 0.908, + 0.318 + ], + "angle": 0, + "content": "Figure 4. Overview of the AVH-Align method. A: We use the pretrained AV-HuBERT model to extract self-supervised features which we further align with a learnable network \\(\\Phi\\). Note that we use a single AV-HuBERT model, but make two forward passes to obtain audio-only and video-only features (instead of a single set of multimodal features). B: At training we maximize the alignment score \\(\\Phi_{ii}\\), between the audio features \\(\\mathbf{a}_i\\) at time step \\(i\\) and the corresponding video features \\(\\mathbf{v}_i\\), while minimizing the alignment \\(\\Phi_{ik}\\) to the other features \\(\\mathbf{v}_k\\) in a neighboring window \\(\\mathcal{N}(i)\\)." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.328, + 0.907, + 0.461 + ], + "angle": 0, + "content": "
MethodModalityTrain typeTrain dataMetric: AUCMetric: AP
FakeAVCelebAV-Deepfake1MFakeAVCelebAV-Deepfake1M
Trim: XTrim: ✓Trim: XTrim: ✓Trim: XTrim: ✓Trim: XTrim: ✓
AVH-Align/supAVsup.FAVC99.299.2 ≈69.063.6 ↓100.0100.0 ≈84.182.2 ↓
AVH-Align/supAVsup.AV1M77.570.8 ↓100.083.1 ↓99.499.1 ≈100.094.4 ↓
AVAD [17]AVunsup.LRS84.584.7 ≈54.354.3 ≈99.599.5 ≈76.376.3 ≈
SpeechForensics [30]AVunsup.VoxCeleb298.898.8 ≈68.868.2 ≈100.0100.0 ≈83.783.5 ≈
AVH-AlignAVunsup.VoxCeleb294.694.6 ≈85.983.5 ↓99.899.8 ≈94.393.5 ≈
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.47, + 0.908, + 0.514 + ], + "angle": 0, + "content": "Table 2. Comparison of AVH-Align and AVH-Align/sup on FakeAVCeleb (FAVC) and AV-Deepfake1M (AV1M) original and trimmed datasets. AVH-Align is not impacted by the presence of leading silence in the original datasets, showing similar performance on the trimmed versions. In contrast, the AUC performance of AVH-Align/sup degrades by \\(16.9\\%\\) when tested on the trimmed variant of AV1M." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.54, + 0.483, + 0.63 + ], + "angle": 0, + "content": "stopped if there is no loss improvement on the validation set for 10 consecutive epochs. AVH-Align/sup is learned using an Adam optimizer with a learning rate of \\(10^{-3}\\). For the AVHuBERT feature extractor we use the checkpoint pretrained on LRS3 and VoxCeleb2 and finetuned on LRS3 for visual speech recognition." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.642, + 0.288, + 0.658 + ], + "angle": 0, + "content": "4.3. Experimental results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.666, + 0.483, + 0.801 + ], + "angle": 0, + "content": "We train our method, AVH-Align, as well as its supervised variant, AVH-Align/sup, on the two datasets, FakeAVCeleb and AV-Deepfake1M, and evaluate on either of them (that is, both in-domain and out-of-domain). For each evaluation, in addition to the original set, we also consider the trimmed version of the validation dataset (trim: \\(\\checkmark\\)). The trimming is performed as described in Sec. 3.3. We always use the untrimmed, original data for training. Results are shown in Table 2." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Impact of leading silence bias. We observe that AVH-Align is robust to the leading silence bias: trimming the silence has no effect on FakeAVCeleb and only a slight effect on AV-Deepfake1M (2.4% AUC). On the other hand, the performance of the supervised variant, AVHAlign/sup, decreases considerably when testing on AV" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.539, + 0.907, + 0.659 + ], + "angle": 0, + "content": "Deepfake1M (by \\(16.9\\%\\) or \\(5.4\\%\\) AUC, depending on the train set). On FakeAVCeleb removing the leading silence has no impact for AVH-Align/sup, when training also on FakeAVCeleb. This may happen because FakeAVCeleb has full manipulations and there is other useful information in the video. Instead, when training on the locally-manipulated AV-Deepfake1M, the difference becomes significant \\((6.7\\%)\\) AUC." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.667, + 0.909, + 0.789 + ], + "angle": 0, + "content": "Comparison with other unsupervised methods. We compare our results with those of AVAD [17] and SpeechForensics [30], which are also unsupervised methods, trained on real data only. Similar to AVH-Align they are not impacted by the spurious leading silence, showing nearly identical results for trimmed and untrimmed datasets. However, their overall performance is considerably worse than ours on the AV-Deepfake1M dataset." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Score visualization. Figure 5 shows the per frame scores obtained by the two methods, AVH-Align and AVH-Align/sup, together with the groundtruth manipulated interval. For the AVH-Align method, the scores represent the misalignment probability between the audio and visual streams at each time frame; for the AVH-Align/sup method, the scores represent the probability of an audio-video frame" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18820" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.102, + 0.09, + 0.348, + 0.163 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.365, + 0.089, + 0.611, + 0.162 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.626, + 0.089, + 0.874, + 0.162 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.164, + 0.348, + 0.236 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.365, + 0.165, + 0.61, + 0.236 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.627, + 0.165, + 0.874, + 0.236 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.239, + 0.349, + 0.312 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.365, + 0.239, + 0.61, + 0.311 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.627, + 0.239, + 0.874, + 0.311 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.326, + 0.908, + 0.385 + ], + "angle": 0, + "content": "Figure 5. Per frame fakeness probabilities for AVH-Align and AVH-Align/sup on AV-Deepfake1M. AVH-Align/sup always marks the first frame—corresponding usually to the leading silence—as fake, thus confirming that it uses the bias to distinguish between real and fake videos. AVH-Align is not affected by the presence of the leading silence. The fakeness probabilities for AVH-Align can be interpreted as misalignment probabilities, which is why they are higher during or after the manipulated region." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.409, + 0.486, + 0.622 + ], + "angle": 0, + "content": "of being fake. We see that AVH-Align/sup always predicts the first frame as being fake, confirming once again that it has learned to associate the spurious leading silence with a video being fake. When detected, the actual fake region is given a considerable lower fakeness probability than that assigned to the leading silence. On the other hand, AVH-Align is not affected by the presence of leading silence. In this case the manipulated areas have higher misalignment scores, a frame-level evaluation yielding \\(77.7\\%\\) AUC. We do notice that there are other regions in the video that get a high misalignment score. This happens because the separation and re-composition of the audio and video streams can introduce desynchronizations between the two even in areas that were not intentionally manipulated." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.634, + 0.484, + 0.65 + ], + "angle": 0, + "content": "4.4. Comparison on the official AV-Deepfake1M test" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.658, + 0.484, + 0.795 + ], + "angle": 0, + "content": "To further verify our conclusions, we evaluate our approach on the official test set of the AV-Deepfake1M dataset for which the labels are withheld. The main difference between the validation set, used in the previous section, and the test set are the subject identities: the test subjects are unseen at train time. We obtain the results by submitting our prediction to the official competition server. Our results are compared against those of other methods based on the values reported in the AV-Deepfake1M paper [6]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.484, + 0.903 + ], + "angle": 0, + "content": "The results shown in Table 3 are in line to those obtained on the validation split (Table 2). Compared to the rest of the methods, AVH-Align has the highest performance (85.24% AUC) even if has seen only real data at training. The silence classifier described in Sec. 3 obtains again a performance of over 98% AUC, indicating that the test set suffers from the same spurious feature as the training and validation splits." + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.406, + 0.909, + 0.598 + ], + "angle": 0, + "content": "
MethodsModalityAUC
Segment-level methods
Meso4 [1]V54.53
MesoInception4 [1]V57.16
MDS [10]AV56.57
MARLIN [5]V58.03
Frame-level methods
Meso4 [1]V63.05
MesoInception4 [1]V64.04
Xception [9]V68.68
EfficientViT [12]V65.51
Ours
AVH-AlignAV85.24
AVH-Align/supAV99.90
Silence classifierA98.44
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.608, + 0.907, + 0.679 + ], + "angle": 0, + "content": "Table 3. Results on the official AV-Deepfake1M test set. Except the methods that have access to the leading silence bias (shown in red), AVH-Align performs best with \\(85.24\\%\\) AUC. Having access to the leading silence bias, the silence classifier and AVH-Align/sup show almost perfect performance of over \\(98\\%\\) AUC." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.706, + 0.907, + 0.798 + ], + "angle": 0, + "content": "Other methods that are trained at video level do not have access to the silence bias, and hence show only modest performance. Instead, we expect audio-based methods to use the bias and return over optimistic results. This is indeed the case for the supervised variant, AVH-Align/sup, which returns an AUC of \\(99.90\\%\\) on the official test set." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.809, + 0.679, + 0.827 + ], + "angle": 0, + "content": "4.5. Further analysis" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Ablation of AVH-Align components. Here we analyze the impact of various design choices on the performance of the AVH-Align method. Specifically, we investigate the following: 1) removing feature normalization; 2) reducing the" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "18821" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.093, + 0.089, + 0.483, + 0.189 + ], + "angle": 0, + "content": "
ConfigurationFAVCAV1MAVLips
AVH-Align94.685.986.3
1... Feature norm: ×92.484.886.1
2... Train size: 8.7k89.182.583.0
3... Score pooling: mean97.476.888.8
4... AV-HuBERT: LRS384.860.380.1
5... Alignment net: Linear31.551.943.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.2, + 0.483, + 0.243 + ], + "angle": 0, + "content": "Table 4. Ablation of the main components of AVH-Align in terms of AUC. Mean score pooling helps when videos are fully generated (FAVC, AVLips). Other ablations degrade the performance." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.267, + 0.483, + 0.435 + ], + "angle": 0, + "content": "training set size from 45,000 to 8,782 samples (the real ones from the AVH-Align/sup train set); 3) using mean score pooling instead of log-sum-exp; 4) extracting features with the AV-HuBERT checkpoint pretrained on LRS3 only instead of LRS3 and VoxCeleb2; 5) using a linear layer instead of MLP. Results are shown in Table 4. All ablations lead to a decrease in performance with the exception of mean score pooling for FakeAVCeleb and AVLips. This is expected, however, since fake videos in FakeAVCeleb and AVLips are manipulated at every frame, while AVDeepfake1M has only locally-manipulated fake videos." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.44, + 0.483, + 0.593 + ], + "angle": 0, + "content": "Impact of VoxCeleb2 pretraining. Both FakeAVCeleb and AV1M datasets use VoxCeleb2 as their source dataset. Can there be a data leakage from AV-HuBERT VoxCeleb2 features to these datasets? We evaluate on a dataset that is not using VoxCeleb2, AVLips [31]. We observe that VoxCeleb2 features perform better even on this dataset, suggesting that these features are simply stronger than LRS3 features. This is may be explained by the size of VoxCeleb data (about four times larger than LRS3) and is in line with the results on other downstream tasks (e.g., lip reading [49])." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.599, + 0.483, + 0.795 + ], + "angle": 0, + "content": "Analysis of AVH-Align/sup architecture. We investigate how a simpler alignment network impacts the supervised model. To this end, we train a linear layer on top of the AV-HuBERT representations instead of the MLP network. In Table 5, we observe that unlike the unsupervised case where a linear layer was too weak to learn, here a linear layer performs as well or even better than the MLP. The largest difference between the MLP and then linear layer is observed when training on FakeAVCeleb and testing on AV-Deepfake1M. With a linear layer, the performance drop from untrimmed to trimmed AV-Deepafake1M is \\(19.3\\%\\) (from \\(85.9\\%\\) to \\(66.6\\%\\) AUC), while with an MLP, the drop is \\(5.4\\%\\) (from \\(69.0\\%\\) to \\(63.6\\%\\) AUC)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.808, + 0.205, + 0.823 + ], + "angle": 0, + "content": "5. Discussion" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.834, + 0.483, + 0.864 + ], + "angle": 0, + "content": "We discuss how our conclusions fit into the broader scope of deepfake detection." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.902 + ], + "angle": 0, + "content": "A different evaluation paradigm. We observed near-perfect performance on datasets such as FakeAVCeleb and" + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.089, + 0.905, + 0.184 + ], + "angle": 0, + "content": "
ArchitectureTrain dataFakeAVCelebAV-Deepfake1M
Trim: XTrim: √Trim: XTrim: √
LinearFakeAVCeleb99.699.585.966.6
MLPFakeAVCeleb99.299.269.063.6
LinearAV-Deepfake1M76.269.799.983.2
MLPAV-Deepfake1M77.570.8100.083.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.195, + 0.905, + 0.223 + ], + "angle": 0, + "content": "Table 5. Architecture analysis of AVH-Align/sup. In most cases, the linear and MLP architectures perform similarly." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.254, + 0.907, + 0.436 + ], + "angle": 0, + "content": "AV-Deepfake1M. The reasons are that the same generative models were used in both train and test, as well as the unintentional leakage of spurious features. The latter also affects cross-dataset performance, e.g., training on FakeAVCeleb and testing on AV-Deepfake1M still yields strong performance. In a realistic scenario, however, we are not given access to the generators, nor to the pre-processing or post-processing steps, which introduce unintentional alterations. Each deepfake released in the wild may be created through a different set of tools, unknown apriori. As such, we believe that another way of gauging the progress of deepfake detection is by refraining the training to real data only." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.445, + 0.909, + 0.672 + ], + "angle": 0, + "content": "Circumventing shortcuts through alignment. Modeling real data is a way to avoid shortcuts in the data. A different direction is taken in the very recent work of Rajan et al. [43]. The idea is to generate fake samples by reconstructing real samples through a generator of choice. This approach ensures that fakes and reals are aligned, avoiding spurious features. On the other hand, this forces the model to focus on the fingerprint of the generator, hindering generalizability. Moreover, by equating fake to a fingerprint, we become susceptible to laundering [32], lose the ability to localise [33, 51] and distinguish benign fakes (e.g., superresolution images [57]). We believe that both directions (alignment and real data) are complementary perspectives that should be considered to tackle the multi-faceted problem of deepfake detection." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.692, + 0.64, + 0.708 + ], + "angle": 0, + "content": "6. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.909, + 0.901 + ], + "angle": 0, + "content": "In this paper we exposed a previously unknown bias in two widely adopted audio-video deepfake detection datasets—a leading silence in fake videos. We showed that models exposed to this bias during training are prone to rely on it when deciding the authenticity of a video, thus displaying overly optimistic results. As an alternative, we propose to shift the learning paradigm towards unsupervised learning on real data only. Specifically, we find that self-supervised audio-video representations coupled with an alignment network trained on real videos produce more robust and consistent results. Our work raises awareness regarding dataset design and evaluation of deepfake detection." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18822" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.091, + 0.251, + 0.108 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.117, + 0.484, + 0.178 + ], + "angle": 0, + "content": "This work was supported in part by the EU Horizon projects AI4TRUST (No. 101070190) and ELIAS (No. 101120237), and by CNCS-UEFISCDI (PN-IV-P7-7.1-PTE-2024-0600)." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.195, + 0.188, + 0.211 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.22, + 0.484, + 0.262 + ], + "angle": 0, + "content": "[1] Darius Afchar, Vincent Nozick, Junichi Yamagishi, and Isao Echizen. MesoNet: A compact facial video forgery detection network. In WIFS, page 1-7, 2018. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.264, + 0.484, + 0.305 + ], + "angle": 0, + "content": "[2] Matyas Bohacek and Hany Farid. Lost in translation: Lipsync deepfake detection from audio-video mismatch. In CVPR, pages 4315-4323, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.308, + 0.483, + 0.35 + ], + "angle": 0, + "content": "[3] Stefano Borzi, Oliver Giudice, Filippo Stanco, and Dario Allegra. Is synthetic voice detection research going into the right direction? In CVPRW, pages 71-80, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.352, + 0.483, + 0.42 + ], + "angle": 0, + "content": "[4] Zhixi Cai, Shreya Ghosh, Abhinav Dhall, Tom Gedeon, Kalin Stefanov, and Munawar Hayat. Glitch in the matrix: A large scale benchmark for content driven audio-visual forgery detection and localization. Comput. Vis. Image Underst., 236:103818, 2023. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.424, + 0.483, + 0.478 + ], + "angle": 0, + "content": "[5] Zhixi Cai, Shreya Ghosh, Kalin Stefanov, Abhinav Dhall, Jianfei Cai, Hamid Rezatofighi, Reza Haffari, and Munawar Hayat. MARLIN: Masked autoencoder for facial video representation learning, 2023. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.482, + 0.483, + 0.537 + ], + "angle": 0, + "content": "[6] Zhixi Cai, Shreya Ghosh, Aman Pankaj Adatia, Munawar Hayat, Abhinav Dhall, Tom Gedeon, and Kalin Stefanov. AV-Deepfake1M: A large-scale LLM-driven audio-visual deepfake dataset, 2024. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.54, + 0.483, + 0.608 + ], + "angle": 0, + "content": "[7] Edresson Casanova, Julian Weber, Christopher Dane Shulby, Arnaldo Candido Junior, Eren Gölgge, and Moacir A. Ponti. YourTTS: Towards zero-shot multi-speaker TTS and zero-shot voice conversion for everyone. In ICML, pages 2709-2720, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.612, + 0.483, + 0.653 + ], + "angle": 0, + "content": "[8] Lucy Chai, David Bau, Ser-Nam Lim, and Phillip Isola. What makes fake images detectable? understanding properties that generalize. In ECCV, pages 103-120, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.656, + 0.483, + 0.695 + ], + "angle": 0, + "content": "[9] François Chollet. Xception: Deep learning with depthwise separable convolutions. In CVPR, pages 1800-1807, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.699, + 0.483, + 0.755 + ], + "angle": 0, + "content": "[10] Komal Chugh, Parul Gupta, Abhinav Dhall, and Ramanathan Subramanian. Not made for each other-audiovisual dissonance-based deepfake detection and localization. In ACMM, pages 439-447, 2020. 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.758, + 0.482, + 0.785 + ], + "angle": 0, + "content": "[11] J. S. Chung, A. Nagrani, and A. Zisserman. VoxCeleb2: Deep speaker recognition. In Interspeech, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.788, + 0.483, + 0.843 + ], + "angle": 0, + "content": "[12] Davide Alessandro Coccomini, Nicola Messina, Claudio Gennaro, and Fabrizio Falchi. Combining EfficientNet and vision transformers for video deepfake detection. In ICIAP, pages 219-229, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.483, + 0.9 + ], + "angle": 0, + "content": "[13] Davide Cozzolino, Giovanni Poggi, Riccardo Corvi, Matthias Nießner, and Luisa Verdoliva. Raising the bar of AI-generated image detection with CLIP. In CVPR, pages 4356-4366, 2024. 4" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.22, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.135 + ], + "angle": 0, + "content": "[14] Davide Cozzolino, Giovanni Poggi, Matthias Nießner, and Luisa Verdoliva. Zero-shot detection of AI-generated images. In ECCV, pages 54-72, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.137, + 0.905, + 0.178 + ], + "angle": 0, + "content": "[15] Alexandre Défossez, Gabriel Synnaeve, and Yossi Adi. Real time speech enhancement in the waveform domain. In *Interspeech*, pages 3291–3295, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.181, + 0.905, + 0.235 + ], + "angle": 0, + "content": "[16] Brian Dolhansky, Joanna Bitton, Ben Pflaum, Jikuo Lu, Russ Howes, Menglin Wang, and Cristian Canton Ferrer. The deepfake detection challenge (dfdc) dataset. arXiv preprint arXiv:2006.07397, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.239, + 0.905, + 0.28 + ], + "angle": 0, + "content": "[17] Chao Feng, Ziyang Chen, and Andrew Owens. Self-supervised video forensics by audio-visual anomaly detection. In CVPR, pages 10491–10503, 2023. 2, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.282, + 0.905, + 0.337 + ], + "angle": 0, + "content": "[18] Alexandros Haliassos, Konstantinos Vougioukas, Stavros Petridis, and Maja Pantic. Lips don't lie: A generalisable and robust approach to face forgery detection. In CVPR, pages 5039-5049, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.34, + 0.905, + 0.395 + ], + "angle": 0, + "content": "[19] Alexandros Haliassos, Rodrigo Mira, Stavros Petridis, and Maja Pantic. Leveraging real talking faces via self-supervision for robust forgery detection. In CVPR, pages 14950-14962, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.398, + 0.905, + 0.466 + ], + "angle": 0, + "content": "[20] Ammarah Hashmi, Sahibzada Adil Shahzad, Chia-Wen Lin, Yu Tsao, and Hsin-Min Wang. AVTENet: Audio-visual transformer-based ensemble network exploiting multiple experts for video deepfake detection. CoRR, abs/2310.13103, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.469, + 0.905, + 0.523 + ], + "angle": 0, + "content": "[21] Baojin Huang, Zhongyuan Wang, Jifan Yang, Jiaxin Ai, Qin Zou, Qian Wang, and Dengpan Ye. Implicit identity driven deepfake face swapping detection. In CVPR, pages 4490-4499, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.527, + 0.905, + 0.596 + ], + "angle": 0, + "content": "[22] Ye Jia, Yu Zhang, Ron J. Weiss, Quan Wang, Jonathan Shen, Fei Ren, Zhifeng Chen, Patrick Nguyen, Ruoming Pang, Ignacio López-Moreno, and Yonghui Wu. Transfer learning from speaker verification to multispeaker text-to-speech synthesis. In NeurIPS, pages 4485-4495, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.599, + 0.905, + 0.652 + ], + "angle": 0, + "content": "[23] Hasam Khalid, Shahroz Tariq, Minha Kim, and Simon S. Woo. FakeAVCeleb: A novel audio-video multimodal deepfake dataset. In NeurIPS Datasets and Benchmarks, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.656, + 0.905, + 0.698 + ], + "angle": 0, + "content": "[24] Jaehyeon Kim, Jungil Kong, and Juhee Son. Conditional variational autoencoder with adversarial learning for end-to-end text-to-speech. In ICML, pages 5530-5540, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.701, + 0.905, + 0.74 + ], + "angle": 0, + "content": "[25] Pavel Korshunov and Sébastien Marcel. DeepFakes: A new threat to face recognition? Assessment and detection. CoRR, abs/1812.08685, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.744, + 0.905, + 0.785 + ], + "angle": 0, + "content": "[26] Iryna Korshunova, Wenzhe Shi, Joni Dambre, and Lucas Theis. Fast face-swap using convolutional neural networks. In ICCV, pages 3697-3705, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.788, + 0.905, + 0.842 + ], + "angle": 0, + "content": "[27] Christos Koutlis and Symeon Papadopoulos. DiMoDif: Discourse modality-information differentiation for audio-visual deepfake detection and localization. CoRR, abs/2411.10193, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[28] Patrick Kwon, Jaeseong You, Gyuhyeon Nam, Sungwoo Park, and Gyeongsu Chae. KoDF: A large-scale Korean deepfake detection dataset. In ICCV, pages 10744-10753, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18823" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.134 + ], + "angle": 0, + "content": "[29] Xiaolou Li, Zehua Liu, Chen Chen, Lantian Li, Li Guo, and Dong Wang. Zero-shot fake video detection by audio-visual consistency. CoRR, abs/2406.07854, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.482, + 0.19 + ], + "angle": 0, + "content": "[30] Yachao Liang, Min Yu, Gang Li, Jianguo Jiang, Boquan Li, Feng Yu, Ning Zhang, Xiang Meng, and Weiqing Huang. SpeechForensics: Audio-visual speech representation learning for face forgery detection. In NeurIPS, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.193, + 0.482, + 0.261 + ], + "angle": 0, + "content": "[31] Weifeng Liu, Tianyi She, Jiawei Liu, Boheng Li, Dongyu Yao, Ziyou Liang, and Run Wang. Lips are lying: Spotting the temporal inconsistency between audio and visual in lip-syncing deepfakes. In NeurIPS, pages 91131-91155, 2024. 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.263, + 0.482, + 0.304 + ], + "angle": 0, + "content": "[32] Sara Mandelli, Paolo Bestagini, and Stefano Tubaro. When synthetic traces hide real content: Analysis of stable diffusion image laundering. CoRR, abs/2407.10736, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.307, + 0.482, + 0.36 + ], + "angle": 0, + "content": "[33] Hannes Mareen, Dimitrios Karageorgiou, Glenn Van Wallendael, Peter Lambert, and Symeon Papadopoulos. TGIF: Text-guided inpainting forgery dataset. CoRR, abs/2407.11566, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.362, + 0.482, + 0.404 + ], + "angle": 0, + "content": "[34] Francesco Marra, Diego Gragnaniello, Luisa Verdoliva, and Giovanni Poggi. Do GANs leave artificial fingerprints? In MIPR, pages 506-511, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.406, + 0.482, + 0.46 + ], + "angle": 0, + "content": "[35] Nicolas M. Müller, Franziska Dieckmann, Pavel Czempin, Roman Canals, and Konstantin Bötttinger. Speech is silver, silence is golden: What do ASVspoof-trained models really learn? CoRR, abs/2106.12914, 2021. 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.462, + 0.482, + 0.531 + ], + "angle": 0, + "content": "[36] Nicolas M. Müller, Piotr Kawa, Wei Herng Choong, Edresson Casanova, Eren Gölgge, Thorsten Müller, Piotr Syga, Philip Sperl, and Konstantin Bötttinger. MLAAD: The multilanguage audio anti-spoofing dataset. In IJCNN, pages 1-7, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.533, + 0.482, + 0.574 + ], + "angle": 0, + "content": "[37] Yuval Nirkin, Yosi Keller, and Tal Hassner. FSGAN: Subject agnostic face swapping and reenactment. In ICCV, pages 7183-7192, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.576, + 0.482, + 0.617 + ], + "angle": 0, + "content": "[38] Utkarsh Ojha, Yuheng Li, and Yong Jae Lee. Towards universal fake image detectors that generalize across generative models. In CVPR, pages 24480-24489, 2023. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.619, + 0.482, + 0.687 + ], + "angle": 0, + "content": "[39] Trevine Oorloff, Surya Koppisetti, Nicolò Bonettini, Dipyaraj Solanki, Ben Colman, Yaser Yacoob, Ali Shahriyari, and Gaurav Bharaj. AVFF: Audio-visual feature fusion for video deepfake detection. In CVPR, pages 27102-27112, 2024. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.689, + 0.482, + 0.744 + ], + "angle": 0, + "content": "[40] Octavian Pascu, Adriana Stan, Dan Oneata, Elisabeta Oneata, and Horia Cucu. Towards generalisable and calibrated audio deepfake detection with self-supervised representations. In Interspeech, pages 4828-4832, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.746, + 0.482, + 0.814 + ], + "angle": 0, + "content": "[41] Alessandro Pianese, Davide Cozzolino, Giovanni Poggi, and Luisa Verdoliva. Training-free deepfake voice recognition by leveraging large-scale pre-trained models. In ACM Workshop on Information Hiding and Multimedia Security, page 289–294, 2024. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.817, + 0.482, + 0.858 + ], + "angle": 0, + "content": "[42] K R Prajwal, Rudrabha Mukhopadhyay, Vinay P. Namboodiri, and C.V. Jawahar. A lip sync expert is all you need for speech to lip generation in the wild. In ACMM, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.86, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[43] Anirudh Sundara Rajan, Utkarsh Ojha, Jedidiah Schloesser, and Yong Jae Lee. On the effectiveness of dataset alignment for fake image detection. CoRR, abs/2410.11835, 2024. 1, 8" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.12 + ], + "angle": 0, + "content": "[44] Tal Reiss, Bar Cavia, and Yedid Hoshen. Detecting deep-fakes without seeing any. CoRR, abs/2311.01458, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.175 + ], + "angle": 0, + "content": "[45] Jonas Ricker, Denis Lukovnikov, and Asja Fischer. AEROBLADE: Training-free detection of latent diffusion images using autoencoder reconstruction error. In CVPR, pages 9130-9140, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.178, + 0.905, + 0.231 + ], + "angle": 0, + "content": "[46] Andreas Rossler, Davide Cozzolino, Luisa Verdoliva, Christian Riess, Justus Thies, and Matthias Niessner. FaceForensics: Learning to detect manipulated facial images. In ICCV, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.233, + 0.905, + 0.287 + ], + "angle": 0, + "content": "[47] Davide Salvi, Brian C. Hosler, Paolo Bestagini, Matthew C. Stamm, and Stefano Tubaro. TIMIT-TTS: A text-to-speech dataset for multimodal synthetic media detection. IEEE Access, 11:50851-50866, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.289, + 0.905, + 0.343 + ], + "angle": 0, + "content": "[48] Sahibzada Adil Shahzad, Ammarah Hashmi, Yan-Tsung Peng, Yu Tsao, and Hsin-Min Wang. AV-Lip-Sync+: Leveraging AV-HuBERT to exploit multimodal inconsistency for video deepfake detection. CoRR, abs/2311.02733, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.345, + 0.905, + 0.397 + ], + "angle": 0, + "content": "[49] Bowen Shi, Wei-Ning Hsu, Kushal Lakhotia, and Abdelrahman Mohamed. Learning audio-visual speech representation by masked multimodal cluster prediction. In ICLR, 2022. 2, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.4, + 0.905, + 0.439 + ], + "angle": 0, + "content": "[50] Kaede Shiohara and Toshihiko Yamasaki. Detecting deepfakes with self-blended images. In CVPR, pages 18699-18708, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.442, + 0.904, + 0.482 + ], + "angle": 0, + "content": "[51] Stefan Smeu, Elisabeta Oneata, and Dan Oneata. DeCLIP: Decoding CLIP representations for deepfake localization. CoRR, abs/2409.08849, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.484, + 0.905, + 0.538 + ], + "angle": 0, + "content": "[52] Joon Son Chung, Andrew Senior, Oriol Vinyals, and Andrew Zisserman. Lip reading sentences in the wild. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6447-6456, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.54, + 0.905, + 0.593 + ], + "angle": 0, + "content": "[53] Hemlata Tak, Jose Patino, Massimiliano Todisco, Andreas Nautsch, Nicholas Evans, and Anthony Larcher. End-to-end anti-spoofing with RawNet2. In ICASSP, pages 6369-6373, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.596, + 0.905, + 0.636 + ], + "angle": 0, + "content": "[54] Aäron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. CoRR, abs/1807.03748, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.638, + 0.905, + 0.691 + ], + "angle": 0, + "content": "[55] Jiadong Wang, Xinyuan Qian, Malu Zhang, Robby T. Tan, and Haizhou Li. Seeing what you said: Talking face generation guided by a lip reading expert. In CVPR, pages 14653-14662, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.693, + 0.905, + 0.803 + ], + "angle": 0, + "content": "[56] Xin Wang, Junichi Yamagishi, Massimiliano Todisco, Héctor Delgado, Andreas Nautsch, Nicholas W. D. Evans, Md. Sahidullah, Ville Vestman, Tomi Kinnunen, Kong Aik Lee, Lauri Juvela, Paavo Alku, Yu-Huai Peng, Hsin-Te Hwang, Yu Tsao, Hsin-Min Wang, Sébastien Le Maguer, Markus Becker, and Zhen-Hua Ling. ASVspoof 2019: A large-scale public database of synthesized, converted and replayed speech. Comput. Speech Lang., 64:101114, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.805, + 0.905, + 0.87 + ], + "angle": 0, + "content": "[57] Zhiyuan Yan, Taiping Yao, Shen Chen, Yandan Zhao, Xinghe Fu, Junwei Zhu, Donghao Luo, Li Yuan, Chengjie Wang, Shouhong Ding, and Yunsheng Wu. DF40: Toward next-generation deepfake detection. CoRR, abs/2406.13495, 2024.8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[58] Wenyuan Yang, Xiaoyu Zhou, Zhikai Chen, Bofei Guo, Zhongjie Ba, Zhihua Xia, Xiaochun Cao, and Kui Ren." + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18824" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.124, + 0.092, + 0.482, + 0.133 + ], + "angle": 0, + "content": "AVoiD-DF: Audio-visual joint learning for detecting deepfake. IEEE Trans. Inf. Forensics Secur., 18:2015-2029, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.135, + 0.483, + 0.178 + ], + "angle": 0, + "content": "[59] Ning Yu, Larry S Davis, and Mario Fritz. Attributing fake images to GANs: Learning and analyzing GAN fingerprints. In CVPR, pages 7556-7566, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.179, + 0.482, + 0.232 + ], + "angle": 0, + "content": "[60] Yinglin Zheng, Jianmin Bao, Dong Chen, Ming Zeng, and Fang Wen. Exploring temporal coherence for more general video face forgery detection. In ICCV, pages 15024-15034, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.518, + 0.956 + ], + "angle": 0, + "content": "18825" + } + ] +] \ No newline at end of file diff --git a/2025/Circumventing Shortcuts in Audio-visual Deepfake Detection Datasets with Unsupervised Learning/371069ba-8973-4b9c-bb98-ecda46029053_origin.pdf b/2025/Circumventing Shortcuts in Audio-visual Deepfake Detection Datasets with Unsupervised Learning/371069ba-8973-4b9c-bb98-ecda46029053_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..1162ade3720a943c9c4071580a822053d367663f --- /dev/null +++ b/2025/Circumventing Shortcuts in Audio-visual Deepfake Detection Datasets with Unsupervised Learning/371069ba-8973-4b9c-bb98-ecda46029053_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c745ae198aba129139f911e52a66d1ca6ef9c3da0c59ad780cdcafe41cc2b591 +size 2180085 diff --git a/2025/Circumventing Shortcuts in Audio-visual Deepfake Detection Datasets with Unsupervised Learning/full.md b/2025/Circumventing Shortcuts in Audio-visual Deepfake Detection Datasets with Unsupervised Learning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..119f6c27f53a89cb845fc849171764b4cb4ccb45 --- /dev/null +++ b/2025/Circumventing Shortcuts in Audio-visual Deepfake Detection Datasets with Unsupervised Learning/full.md @@ -0,0 +1,312 @@ +# Circumventing shortcuts in audio-visual deepfake detection datasets with unsupervised learning + +Stefan Smeu\*1 Dragos-Alexandru Boldisor\*1 Dan Oneata2 Elisabeta Oneata1 1Bitdefender 2POLITEHNICA Bucharest + +{ssmeu, dboldisor, eoneata}@bitdefender.com dan. oneata@gmail.com + +# Abstract + +Good datasets are essential for developing and benchmarking any machine learning system. Their importance is even more extreme for safety critical applications such as deepfake detection—the focus of this paper. Here we reveal that two of the most widely used audio-video deepfake datasets suffer from a previously unidentified spurious feature: the leading silence. Fake videos start with a very brief moment of silence and, on the basis of this feature alone, we can separate the real and fake samples almost perfectly. As such, previous audio-only and audio-video models exploit the presence of silence in the fake videos and consequently perform worse when the leading silence is removed. To circumvent latching on such an unwanted artifact and possibly other unrevealed ones, we propose a shift from supervised to unsupervised learning by training models exclusively on real data. We show that by aligning self-supervised audio-video representations we remove the risk of relying on dataset-specific biases and improve robustness in deepfake detection. + +# 1. Introduction + +Manipulated videos represent a threat to society as they have the potential of misleading people into believing actors with malicious intents. By spreading misinformation on social media platforms, people may be exposed to scams (e.g., identity theft operations), conspiracy theories and political misinformation. Therefore, deepfake detection methods are essential tools on a global scale. + +The progress in automated deepfake detection is fueled by the datasets developed by the research community. Good quality data is essential for both training and benchmarking the progress of these methods. In recent years, numerous datasets have been proposed. Among the audio-video deepfake datasets, there have been released datasets that alter both streams [6, 23] or only one [36, 46, 56]; datasets with + +![](images/88ea4fd6679afa749cb4379af534874a4687fbf669bfa7767bf2cbb9d0414013.jpg) +Figure 1. Audio-visual deepfake detection datasets have a silence bias: fake samples start with a brief moment of silence, which is not the case for real samples. Here we show the first $62.5\mathrm{ms}$ of the audio waveform for a real and the corresponding fake sample from the AV-Deepfake1M dataset [6]. + +full-video [23] or only local [4, 6] manipulations. + +However, care has to be taken to ensure a good deepfake detection dataset. Any asymmetry in the preparation of fakes and reals can result in biases that correlate spuriously with the groundtruth label. For example, in the image domain, different preprocessing pipelines [8] or types of resizing [43] between real and fake samples paint an overly optimistic picture. In the audio domain, the very popular ASVspoof19 dataset [56] has been shown to leak information about labels in the form of silence duration [35] or bitrate information [3]. Since many of the deepfake detection models are high capacity, they can easily use such artifacts as shortcuts for learning, consequently greatly impacting their ability to generalize in a real case scenario. + +In this work, we first expose a bias present in two widely-adopted audio-video deepfake detection datasets, FakeAVCeleb [23] and AV-Deepfake1M [6]: a short moment of silence present at the beginning of manipulated videos (see Figure 1). Based on this information, a very simple silence classifier can reach a near-perfect performance of over $98\%$ on both FakeAVCeleb and AV-Deepfake1M. Moreover, we observe that the silence distribution is similar across datasets, implying that prior methods may have potentially over estimated the generalization + +performance based on this shortcut. + +Introducing such biases when creating the datasets is unavoidable and there might be others that are more subtle and harder to reveal. To circumvent this problem, we show that one solution is a shift in learning paradigm, from the more common supervised setup to the unsupervised one [17, 44]. Removing fake samples from training eliminates the focus on the asymmetries induced by spurious artifacts. Moreover, this also has the potential to improve generalization among different manipulation techniques, since supervised detection methods tend to over-rely on generator-specific fingerprints [34, 59]. By limiting to real data, we can also naturally leverage self-supervised representations, which have been shown to improve generalization [38]. + +To this end, we propose an approach, named AVH-Align, that learns a frame-level audio-video alignment score on top of AV-HuBERT features [49]. Since tampered videos are expected to have greater audio-video desynchronizations, this alignment score can effectively differentiate fake and real videos. We show that this approach is robust with respect to the identified shortcut of leading silence, and also outperforms other methods that do not use the silence bias, while not seeing any fake samples at training time. + +To summarize, our work makes the following contributions: 1. We expose a previously unknown spurious feature in two of the most widely adopted deepfake detection datasets. 2. We analyze the impact of this shortcut on various state-of-the-art models. 3. We show that a way to mitigate such shortcuts is by training on real data only and we introduce a new method in this direction. Our code is available at: https://github.com/bit-ml/AVH-Align. + +# 2. Related work + +Audio-video deepfake detection. Many approaches for deepfake detection on videos have focused on the visual stream of information only [1, 5, 18, 19, 21, 50, 60]. But with the recent introduction of audio-visual datasets, (FakeAVCeleb [23], Deepfake TIMIT [25, 47], KODF [28], LAV-DF [4], AV-Deepfake-1M [6]), more research has shifted towards models that exploit both audio and video cues [17, 20, 27, 39, 48, 58]. An emerging trend in this direction is the use of pretrained representations in a self-supervised way [17, 39]. But, different from our approach, these methods train the representations from scratch and use them as a first step in a more elaborated pipeline: supervised classification [39] or anomaly detection [17]. There are also works that similarly to us exploit the pretrained audio-video AV-HuBERT model [49] to extract representations [20, 48], but all of those methods are trained in the fully supervised learning paradigm. + +Unsupervised deepfake detection. To make detection more generalizable across generators, a new direction is + +to depart from the supervised paradigm and resort only to real data. There are two main classes of such unsupervised approaches: methods that rely on consistency checks and methods that treat the problem as an anomaly detection one. Among the first class, prior work proposed to verify that audio and visual streams align at a semantic level, for example, from the point of view of spoken content [2, 29], or at a representation level, based on the alignment of audio and video features [17, 44]. Consistency checks have also been used with respect to the identity of a speaker (comparing a query sample against real audio [41] or real images [44] of the target speaker) or between image and text modalities [44]. For the second class, deepfake detection as anomaly detection, recent work in the image domain has used the reconstruction loss of the query image to tell whether it is anomalous [14, 45]; for example, Ricker et al. [45] make the observation that images generated by a latent diffusion model (LDM) are easier to reconstruct by the LDM than real images. For the audio-visual domain, Feng et al. [17] use both classes of approaches: they use consistency checks to estimate synchronization between the two streams and then flag anomalies using density estimation. + +# 3. Silence bias in audio-video datasets + +In this section, we show that two popular datasets (Sec. 3.1) have a silence bias. We analyze its behavior and show that a simple classifier based on the leading silence alone can obtain almost perfect separation between fake and real samples (Sec. 3.2). This implies that the performance of prior work is susceptible to have been overestimated. For this reason we analyze its impact on various audio and audio-visual methods (Sec. 3.3). + +# 3.1. Datasets + +We consider two audio-visual datasets in our analysis: FakeAVCeleb [23] and AV-Deepfake1M [6]. They distinguish mainly in the fact that the first contains fully-generated video sequences, while the second contains partially-manipulated sequences. Both are based on the VoxCeleb2 dataset [11], which consists of YouTube audio-video of celebrities. Apart from the real samples—real video real audio (RVRA)—both datasets include three types of fake videos: real video fake audio (RVFA), fake video real audio (FVRA), and fake video fake audio (FVFA). + +FakeAVCeleb contains 500 real videos from VoxCeleb2 and 19.5k fake videos (10k FVFA, 9k FVRA, 500 RVFA). The fake visual content was generated with face swapping methods (Faceswap [26] and FSGAN [37]) or the Wav2Lip lipSyncing approach [42]. The fake audio content was generate with the voice cloning tool SV2TTS [22]. The dataset is diverse across age groups, genders, races, as well as with respect to the number of subjects in a single video, their + +![](images/ac2fa88c7ea9a187a45c81f5d8eb3c9c641ae51ddf7a2a38b3ad0df6f89f25ee.jpg) +Figure 2. Normalized distribution plots of the leading silence duration for real and fake videos in the FakeAVCeleb (left) and AV-Deepfake1M (right) datasets. The fake samples start with 25–30 ms of silence. + +placement, the visual and audio quality. In our experiments, we split the dataset in $70\%$ (train and validation) and $30\%$ (test). This split is kept consistent across all experiments. + +AV-Deepfake1M is a large scale dataset, which consists of over one million videos and 2k subjects. As opposed to the FakeAVCeleb dataset, the manipulations here are local and consists of word-level replacements, insertions and deletions. The text manipulations are generated with the ChatGPT large language model. The fake video content is generated with the lipSyncing method TalkLip [55], while the fake audio content is generated with the VITS [24] or YourTTS [7] methods. The authors ensure that the synthesized words share the same background noise with the full audio, by first extracting the audio noise with the Denoiser method [15] and then adding it to the synthesized words. The dataset is originally split into train, validation and test, with the test split having a different set of speakers than those encountered in the train and validation splits. For our experiments, we select training and validation samples from the original training split, and evaluate on 10k samples from the original validation split or on the full official test set. + +# 3.2. Analysis of leading silence + +We start by analyzing the silence distribution of the real and fake samples in the two considered datasets. We define the duration of the leading silence as the moment when the magnitude of the audio exceeds a certain threshold $\tau$ . For this experiment, we select this threshold to be $5 \cdot 10^{-4}$ , but as we will shortly see, the results are robust to its choice. We carry this analysis only on the real (RVRA) and fully fake (FVFA) videos from each dataset's test set. + +The results are shown in Figure 2. We observe that the real videos start with noise, while the fake samples have a leading silence of around $25 - 30\mathrm{ms}$ . The silence duration of fake samples is similar for both datasets, although the distribution is much sharper for the AV-Deepfake1M dataset. If we were to rank the samples based on this feature we + +![](images/e6c33b744cbca2d22ca98c8b5ac0f177e281b915cfcd47b16348cdd8f2267e97.jpg) +Figure 3. Left: The impact of the silence threshold on the leading silence classifier. Right: The impact of the leading duration on the maximum amplitude classifier. + +would obtain an area under curve of the receiver operating characteristic curve (AUC) of over $98\%$ for either datasets. + +What counts as silence? For the previous experiment we have considered that silence is the signal that has an amplitude lower than $\tau = 5 \cdot 10^{-4}$ . We investigate how sensitive the performance is to this threshold. We vary $\tau$ across a grid of values and show the results in Figure 3 (left). We observe that the results are strong as long as this threshold is small enough. + +A different perspective: Maximum amplitude. Instead of looking at the silence duration, we can alternatively measure the maximum amplitude in the first $\delta$ seconds of the audio. Figure 3 shows the AUC obtained by ranking the samples based on this feature for various leading durations $\delta$ . We observe that this alternative measure also yields strong results (over $98\%$ AUC), and that the optimum is obtained for a $\delta$ of around $30~\mathrm{ms}$ , in line with the previous experiment. As we extend the window, the performance gets closer to random chance, AUC of $50\%$ . + +Other biases: Volume and trailing silence. If we extend the duration $\delta$ to cover the entire audio, we get an estimate of the maximum volume of the audio. This feature yields an AUC of $67.6\%$ on FakeAVCeleb, which is much less than $98.4\%$ , but still over random chance (50% AUC). Similarly, we have also investigated the trailing silence (the silence at the end of the audio) and have obtained AUC of over $99\%$ for FakeAVCeleb. AV-Deepfake1M is less sensitive to these other biases, with values closer to $50\%$ . While these biases are not as consistent as the leading silence, they are still problematic. + +Why do fake samples have a leading silence? Given that we do not have access to generation process of the two datasets, it is challenging to pinpoint the exact reason for the occurrence of the leading silence bias. However, we speculate that this happens when the audio may be slightly shorter than the video counterpart. Note that this is a different reason from the silence observed in audio only datasets + +[35]. There the real audios had a leading silence, while the synthesized speech was silence-free. In the case of audio-video datasets it might be challenging to completely avoid this problem, but an easy solution (for this particular bias) is to trim the leading silence. This is what we do in the next experiments. + +Biases in other datasets. We consider three more datasets: LAV-DF [4], AVLips [31], DFDC [16]. We find that LAV-DF has a slight trailing silence bias (65% AUC), but it does not exhibit the other two biases (leading silence and volume). In the case of AVLips and DFDC we did not identify any of the three biases. This may happen because the samples in these datasets have real audio and the unmodified audio matches the visual sequence more precisely. This is confirmed by a fine-grained analysis on FakeAVCeleb, where we observe that on the FVRA split the biases are diminished (62.4% AUC for lead silence), while they persist on the RVFA split (100% AUC for lead silence). + +# 3.3. Impact on prior work + +We have shown that real and fake samples differ in terms of the leading silence. This is a simple feature which could be learned by the high-capacity neural networks. Here we investigate whether that is indeed the case for three existing methods: + +- RawNet2 [53], which is an audio-only method that operates on the raw waveform. Its architecture sequences sinc layers, convolutions, gated recurrent units and fully connected layers. +- MDS [10] (modality dissonance score), which is an audio-visual method that estimates the mismatch between audio and video segments. The score is computed as the distance between audio and visual features. +- AVAD [17] (audio-visual anomaly detection), which is an audio-visual method trained on real data only. The approach has two steps: first, it estimates the desynchronization between audio and video; then it estimates whether these patterns are typical of real data or anomalous. + +We train the RawNet2 and MDS methods on both datasets using the code provided by the authors. For the AVAD we use the provided checkpoint (trained on LRS [52]) and do not retrain it since training code is not available. As a baseline we include our leading silence duration classifier described in the previous section. + +We evaluate all methods in two settings: on the corresponding evaluation set and on a trimmed version of the same evaluation set. For the trimmed version, we discard the beginning of videos such that each of them starts with the first non-silent segment. We find the duration to discard by first computing the leading silence (using the same $5 \cdot 10^{-4}$ threshold on the magnitude) and then rounding it up to the nearest multiple of $1 / \text{FPS}$ (the reciprocal of the video's frame rate); this rounding ensures that the audio and + +
MethodMod.FakeAVCelebAV-Deepfake1M
Trim: XTrim: √Trim: XTrim: √
Silence classifierA98.454.8 ↓ 43.698.250.6 ↓ 47.6
RawNet2 [53]A99.997.3 ↓ 2.699.988.1 ↓ 11.8
MDS [10]AV90.473.8 ↓ 16.699.254.9 ↓ 44.3
AVAD [17]AV95.295.2 ≈ 0.052.952.9 ≈ 0.0
+ +Table 1. The impact of leading silence on the performance (area under the receiver operator characteristic curve; AUC) for three existing deepfake detection methods and the silence classifier. Results are shown for RVRA-FVFA subsets. + +video channels remain synchronized. For training, we do not perform trimming, but use the original dataset. + +The results are shown in Table 1. First, we observe that the audio-based RawNet2 is the top performing method. Moreover, we see that the leading silence classifier is generally better than the two other approaches: MDS and AVAD. The relatively stable performance of AVAD indicates that this method dose not latch on the silence information. + +In terms of the impact of the leading silence, we observe that removing it through trimming affects methods differently. As expected, the silence classifier is affected the most and its performance drops down to random chance on the trimmed data. RawNet2 is not as affected on the FakeAVCeleb, presumably because there is still enough information throughout the signal, but it suffers a larger hit on the more challenging AV-DeepFake1M, which has only partially-manipulated samples. MDS is affected by the leading silence bias on both FakeAVCeleb and AV-Deepfake1M, though significantly stronger on AV-Deepfake1M. AVAD is the most robust method, since this approach does not explicitly model the silence information. Finally, we notice that the RawNet2 audio-only method is best even on the trimmed settings. This is noteworthy since it is the first time an audio-only method has been applied to these datasets. This suggests that the audio stream is an important source of information that is often overlooked by prior work. + +# 4. Modeling real data for deepfake detection + +The previous section indicated that audio-visual datasets exhibit a silence bias which can easily be exploited. We want to develop a method that is robust to this bias (and possibly other uncovered ones) and still performs well. We have seen that models trained only on real data [17] are promising in being robust to the silence shortcut, but the performance was modest. To further improve them we propose to build on top of audio-focused self-supervised features. Self-supervised features have shown strong generalization for both visual [13, 38] and audio [40, 41] deepfake detection. We choose audio-focused self-supervised features be + +cause the audio models showed strong performance in the previous section. Note that we cannot rely on audio-only models because there are cases where manipulations appear only in the visual domain (the fake video, real audio case). + +# 4.1. Method + +We propose a method that aligns AV-HuBERT [49] features on real data. First, we extract audio and visual frame-level features with a pretrained AV-HuBERT model. Then, on top of these features we learn a network to better align them. The alignment network is learnt on real samples by matching each video frame to its corresponding audio frame [17]. We call our method AVH-Align (AV-HuBERT Aligned) and show its depiction in Figure 4. + +Self-supervised features. We use AV-HuBERT to represent both the audio and visual content of a video. AV-HuBERT is a Transformer network trained in a self-supervised way to predict iteratively refined centroids from masked inputs. The features extracted by AV-HuBERT encode audio information as proved by its strong performance on tasks such as lip reading or noisy audio-visual speech recognition. We extract the audio and visual representations independently: to extract audio features we mask the visual input, to extract visual features we mask the audio input. For a video, we obtain audio features $\mathbf{a}_i$ and visual features $\mathbf{v}_i$ features for each time step $i$ . Both representations are 1024 dimensional and have a temporal resolution of 25 frames per second. + +Alignment network. To tell how well the audio and visual features match each other, we first independently L2 normalize the feature vectors and then feed them into a network $\Phi$ . This is implemented as a multi-layer perceptron (MLP) over the concatenated normalized audio and visual features: + +$$ +\Phi_ {i j} = \operatorname {M L P} \left(\left[ \mathbf {a} _ {i}; \mathbf {v} _ {j} \right]\right). \tag {1} +$$ + +The MLP has four layers, which progressively reduce the feature dimensionality, with layers mapping from the AV-HuBERT feature size of 1024 to 512, 256, 128, and finally to a single output. Each hidden layer includes Layer Normalization and ReLU activations. + +Loss function. To learn the alignment network $\Phi$ we maximize the probability of an audio frame $\mathbf{a}_i$ to match the corresponding video frame $\mathbf{v}_i$ ; this probability is defined as: + +$$ +p \left(\mathbf {v} _ {i} \mid \mathbf {a} _ {i}\right) = \frac {\exp \Phi_ {i i}}{\sum_ {k \in \mathcal {N} (i)} \exp \Phi_ {i k}}, \tag {2} +$$ + +where $\mathcal{N}(i)$ represents the temporal neighborhood around the frame $i$ . In our case $\mathcal{N}(i)$ contains the 30 neighboring frames around $i$ . We define the final loss as the negative + +probabilities averaged across the entire video: + +$$ +\mathcal {L} = - \frac {1}{T} \sum_ {i = 1} ^ {T} \log p (\mathbf {v} _ {i} | \mathbf {a} _ {j}). \tag {3} +$$ + +This loss is similar to the contrastive loss InfoNCE [54], which was also used for deepfake detection [17, 39]. + +Inference. Once $\Phi$ is learned, we can estimate the fakeness score as the negative of $\Phi_{ii}$ for each audio-video frame pair in a video; aligned audio-video frame pairs should yield lower scores. Then we compute an overall alignment score for the entire video by pooling the per-frame scores using the log-sum-exp function (a smooth version of the max function). + +Supervised variant. To understand the impact of the silence bias on the standard supervised learning paradigm, we design a supervised variant, AVH-Align/sup, that uses the same features and alignment network as AVH-Align, but a classification loss. In this setup we assume that apart from the real videos, we also have access to fake videos in the training set, with corresponding labels $y$ . To obtain a per-video fakeness probability, we first pool the negated per-frame scores $\Phi_{ii}$ with the log-sum-exp function and then we apply the sigmoid function $\sigma$ . Finally, we optimize the binary cross-entropy (BCE) loss: + +$$ +\mathcal {L} _ {\sup } = \operatorname {B C E} \left(\sigma \left(\log \operatorname {s u m} \exp (- \Phi_ {i i})\right), y\right). \tag {4} +$$ + +# 4.2. Experimental setup + +Datasets and metrics. We conducted our experiments on the two datasets introduced in Sec. 3.1: FakeAVCeleb and AV-Deepfake1M. For FakeAVCeleb, we use $70\%$ of the dataset for training and validation, and the rest of $30\%$ for testing. For AV-Deepfake1M, we evaluate on 10k samples from the original validation set, but for the best performing models we also report results on the official withheld test set. To train AVH-Align, we use 50k real samples from the training set of AV-Deepfake1M (45k samples for training and 5k for validation). Since AV-Deepfake1M is based on VoxCeleb2, these samples originate from VoxCeleb2. To train the supervised variant, AVH-Align/sup, on AV-Deepfake1M, we select the same number of samples (45k for training and 5k for validation), but this time coming from both the real and fake classes. For both datasets, we evaluate video-level detection. Following prior work, we report results in terms of the area under the receiver operator characteristic curve (AUC) and average precision (AP), with the fake class being the positive class. + +Implementation details. For AVH-Align we use a learning rate scheduler with a patience of 5 epochs and a factor of 0.1, with a starting learning rate of $10^{-5}$ . The training is + +![](images/fef535300c508cd85c42f608f28343135960f9d4e6659e6e4c7197083836fe04.jpg) +Figure 4. Overview of the AVH-Align method. A: We use the pretrained AV-HuBERT model to extract self-supervised features which we further align with a learnable network $\Phi$ . Note that we use a single AV-HuBERT model, but make two forward passes to obtain audio-only and video-only features (instead of a single set of multimodal features). B: At training we maximize the alignment score $\Phi_{ii}$ , between the audio features $\mathbf{a}_i$ at time step $i$ and the corresponding video features $\mathbf{v}_i$ , while minimizing the alignment $\Phi_{ik}$ to the other features $\mathbf{v}_k$ in a neighboring window $\mathcal{N}(i)$ . + +![](images/543f0c7bf486300dc621cb5661b715db57f07236c69304ea86d63f3f43615941.jpg) + +
MethodModalityTrain typeTrain dataMetric: AUCMetric: AP
FakeAVCelebAV-Deepfake1MFakeAVCelebAV-Deepfake1M
Trim: XTrim: ✓Trim: XTrim: ✓Trim: XTrim: ✓Trim: XTrim: ✓
AVH-Align/supAVsup.FAVC99.299.2 ≈69.063.6 ↓100.0100.0 ≈84.182.2 ↓
AVH-Align/supAVsup.AV1M77.570.8 ↓100.083.1 ↓99.499.1 ≈100.094.4 ↓
AVAD [17]AVunsup.LRS84.584.7 ≈54.354.3 ≈99.599.5 ≈76.376.3 ≈
SpeechForensics [30]AVunsup.VoxCeleb298.898.8 ≈68.868.2 ≈100.0100.0 ≈83.783.5 ≈
AVH-AlignAVunsup.VoxCeleb294.694.6 ≈85.983.5 ↓99.899.8 ≈94.393.5 ≈
+ +Table 2. Comparison of AVH-Align and AVH-Align/sup on FakeAVCeleb (FAVC) and AV-Deepfake1M (AV1M) original and trimmed datasets. AVH-Align is not impacted by the presence of leading silence in the original datasets, showing similar performance on the trimmed versions. In contrast, the AUC performance of AVH-Align/sup degrades by $16.9\%$ when tested on the trimmed variant of AV1M. + +stopped if there is no loss improvement on the validation set for 10 consecutive epochs. AVH-Align/sup is learned using an Adam optimizer with a learning rate of $10^{-3}$ . For the AVHuBERT feature extractor we use the checkpoint pretrained on LRS3 and VoxCeleb2 and finetuned on LRS3 for visual speech recognition. + +# 4.3. Experimental results + +We train our method, AVH-Align, as well as its supervised variant, AVH-Align/sup, on the two datasets, FakeAVCeleb and AV-Deepfake1M, and evaluate on either of them (that is, both in-domain and out-of-domain). For each evaluation, in addition to the original set, we also consider the trimmed version of the validation dataset (trim: $\checkmark$ ). The trimming is performed as described in Sec. 3.3. We always use the untrimmed, original data for training. Results are shown in Table 2. + +Impact of leading silence bias. We observe that AVH-Align is robust to the leading silence bias: trimming the silence has no effect on FakeAVCeleb and only a slight effect on AV-Deepfake1M (2.4% AUC). On the other hand, the performance of the supervised variant, AVHAlign/sup, decreases considerably when testing on AV + +Deepfake1M (by $16.9\%$ or $5.4\%$ AUC, depending on the train set). On FakeAVCeleb removing the leading silence has no impact for AVH-Align/sup, when training also on FakeAVCeleb. This may happen because FakeAVCeleb has full manipulations and there is other useful information in the video. Instead, when training on the locally-manipulated AV-Deepfake1M, the difference becomes significant $(6.7\%)$ AUC. + +Comparison with other unsupervised methods. We compare our results with those of AVAD [17] and SpeechForensics [30], which are also unsupervised methods, trained on real data only. Similar to AVH-Align they are not impacted by the spurious leading silence, showing nearly identical results for trimmed and untrimmed datasets. However, their overall performance is considerably worse than ours on the AV-Deepfake1M dataset. + +Score visualization. Figure 5 shows the per frame scores obtained by the two methods, AVH-Align and AVH-Align/sup, together with the groundtruth manipulated interval. For the AVH-Align method, the scores represent the misalignment probability between the audio and visual streams at each time frame; for the AVH-Align/sup method, the scores represent the probability of an audio-video frame + +![](images/3a5e7050545c78dd216795fba5e17c516d365fcdc4fd74a3bf82eac464359703.jpg) + +![](images/9fc5dc4342f77a1f76df917b818f00ee06880fe6068c5daef0ace80fae4a4c84.jpg) + +![](images/3eb58c549dfade0519d30bdb06fc7a18cd3fabf2bf600c74cd88a91167ecaeec.jpg) + +![](images/d684a2c335c1d0c7b8ca09f075030673d4aa06ea8848ed5f7118661714106f12.jpg) + +![](images/355e011d05424346819a2982958bbec153a7fba9682f6dc4343b90e39e78cc91.jpg) + +![](images/d95be9185d9f5913eec140e6ae8a6ffd894d101050f2b8df55693bbc180e41bd.jpg) + +![](images/d279d263aa7afd4d98915b491760dec98e3721e50dfb8903fee1002589843c35.jpg) +Figure 5. Per frame fakeness probabilities for AVH-Align and AVH-Align/sup on AV-Deepfake1M. AVH-Align/sup always marks the first frame—corresponding usually to the leading silence—as fake, thus confirming that it uses the bias to distinguish between real and fake videos. AVH-Align is not affected by the presence of the leading silence. The fakeness probabilities for AVH-Align can be interpreted as misalignment probabilities, which is why they are higher during or after the manipulated region. + +![](images/16d9b9bb8ba13e5ea84c2ddcee3d06d930ebcf47abc8260fcedf9544826f4db5.jpg) + +![](images/cdfe86de160fa666ba4a3447feb789efc16039559628f24d475d68ff11186488.jpg) + +of being fake. We see that AVH-Align/sup always predicts the first frame as being fake, confirming once again that it has learned to associate the spurious leading silence with a video being fake. When detected, the actual fake region is given a considerable lower fakeness probability than that assigned to the leading silence. On the other hand, AVH-Align is not affected by the presence of leading silence. In this case the manipulated areas have higher misalignment scores, a frame-level evaluation yielding $77.7\%$ AUC. We do notice that there are other regions in the video that get a high misalignment score. This happens because the separation and re-composition of the audio and video streams can introduce desynchronizations between the two even in areas that were not intentionally manipulated. + +# 4.4. Comparison on the official AV-Deepfake1M test + +To further verify our conclusions, we evaluate our approach on the official test set of the AV-Deepfake1M dataset for which the labels are withheld. The main difference between the validation set, used in the previous section, and the test set are the subject identities: the test subjects are unseen at train time. We obtain the results by submitting our prediction to the official competition server. Our results are compared against those of other methods based on the values reported in the AV-Deepfake1M paper [6]. + +The results shown in Table 3 are in line to those obtained on the validation split (Table 2). Compared to the rest of the methods, AVH-Align has the highest performance (85.24% AUC) even if has seen only real data at training. The silence classifier described in Sec. 3 obtains again a performance of over 98% AUC, indicating that the test set suffers from the same spurious feature as the training and validation splits. + +
MethodsModalityAUC
Segment-level methods
Meso4 [1]V54.53
MesoInception4 [1]V57.16
MDS [10]AV56.57
MARLIN [5]V58.03
Frame-level methods
Meso4 [1]V63.05
MesoInception4 [1]V64.04
Xception [9]V68.68
EfficientViT [12]V65.51
Ours
AVH-AlignAV85.24
AVH-Align/supAV99.90
Silence classifierA98.44
+ +Table 3. Results on the official AV-Deepfake1M test set. Except the methods that have access to the leading silence bias (shown in red), AVH-Align performs best with $85.24\%$ AUC. Having access to the leading silence bias, the silence classifier and AVH-Align/sup show almost perfect performance of over $98\%$ AUC. + +Other methods that are trained at video level do not have access to the silence bias, and hence show only modest performance. Instead, we expect audio-based methods to use the bias and return over optimistic results. This is indeed the case for the supervised variant, AVH-Align/sup, which returns an AUC of $99.90\%$ on the official test set. + +# 4.5. Further analysis + +Ablation of AVH-Align components. Here we analyze the impact of various design choices on the performance of the AVH-Align method. Specifically, we investigate the following: 1) removing feature normalization; 2) reducing the + +
ConfigurationFAVCAV1MAVLips
AVH-Align94.685.986.3
1... Feature norm: ×92.484.886.1
2... Train size: 8.7k89.182.583.0
3... Score pooling: mean97.476.888.8
4... AV-HuBERT: LRS384.860.380.1
5... Alignment net: Linear31.551.943.4
+ +training set size from 45,000 to 8,782 samples (the real ones from the AVH-Align/sup train set); 3) using mean score pooling instead of log-sum-exp; 4) extracting features with the AV-HuBERT checkpoint pretrained on LRS3 only instead of LRS3 and VoxCeleb2; 5) using a linear layer instead of MLP. Results are shown in Table 4. All ablations lead to a decrease in performance with the exception of mean score pooling for FakeAVCeleb and AVLips. This is expected, however, since fake videos in FakeAVCeleb and AVLips are manipulated at every frame, while AVDeepfake1M has only locally-manipulated fake videos. + +Impact of VoxCeleb2 pretraining. Both FakeAVCeleb and AV1M datasets use VoxCeleb2 as their source dataset. Can there be a data leakage from AV-HuBERT VoxCeleb2 features to these datasets? We evaluate on a dataset that is not using VoxCeleb2, AVLips [31]. We observe that VoxCeleb2 features perform better even on this dataset, suggesting that these features are simply stronger than LRS3 features. This is may be explained by the size of VoxCeleb data (about four times larger than LRS3) and is in line with the results on other downstream tasks (e.g., lip reading [49]). + +Analysis of AVH-Align/sup architecture. We investigate how a simpler alignment network impacts the supervised model. To this end, we train a linear layer on top of the AV-HuBERT representations instead of the MLP network. In Table 5, we observe that unlike the unsupervised case where a linear layer was too weak to learn, here a linear layer performs as well or even better than the MLP. The largest difference between the MLP and then linear layer is observed when training on FakeAVCeleb and testing on AV-Deepfake1M. With a linear layer, the performance drop from untrimmed to trimmed AV-Deepafake1M is $19.3\%$ (from $85.9\%$ to $66.6\%$ AUC), while with an MLP, the drop is $5.4\%$ (from $69.0\%$ to $63.6\%$ AUC). + +# 5. Discussion + +We discuss how our conclusions fit into the broader scope of deepfake detection. + +A different evaluation paradigm. We observed near-perfect performance on datasets such as FakeAVCeleb and + +Table 4. Ablation of the main components of AVH-Align in terms of AUC. Mean score pooling helps when videos are fully generated (FAVC, AVLips). Other ablations degrade the performance. + +
ArchitectureTrain dataFakeAVCelebAV-Deepfake1M
Trim: XTrim: √Trim: XTrim: √
LinearFakeAVCeleb99.699.585.966.6
MLPFakeAVCeleb99.299.269.063.6
LinearAV-Deepfake1M76.269.799.983.2
MLPAV-Deepfake1M77.570.8100.083.1
+ +Table 5. Architecture analysis of AVH-Align/sup. In most cases, the linear and MLP architectures perform similarly. + +AV-Deepfake1M. The reasons are that the same generative models were used in both train and test, as well as the unintentional leakage of spurious features. The latter also affects cross-dataset performance, e.g., training on FakeAVCeleb and testing on AV-Deepfake1M still yields strong performance. In a realistic scenario, however, we are not given access to the generators, nor to the pre-processing or post-processing steps, which introduce unintentional alterations. Each deepfake released in the wild may be created through a different set of tools, unknown apriori. As such, we believe that another way of gauging the progress of deepfake detection is by refraining the training to real data only. + +Circumventing shortcuts through alignment. Modeling real data is a way to avoid shortcuts in the data. A different direction is taken in the very recent work of Rajan et al. [43]. The idea is to generate fake samples by reconstructing real samples through a generator of choice. This approach ensures that fakes and reals are aligned, avoiding spurious features. On the other hand, this forces the model to focus on the fingerprint of the generator, hindering generalizability. Moreover, by equating fake to a fingerprint, we become susceptible to laundering [32], lose the ability to localise [33, 51] and distinguish benign fakes (e.g., superresolution images [57]). We believe that both directions (alignment and real data) are complementary perspectives that should be considered to tackle the multi-faceted problem of deepfake detection. + +# 6. Conclusions + +In this paper we exposed a previously unknown bias in two widely adopted audio-video deepfake detection datasets—a leading silence in fake videos. We showed that models exposed to this bias during training are prone to rely on it when deciding the authenticity of a video, thus displaying overly optimistic results. As an alternative, we propose to shift the learning paradigm towards unsupervised learning on real data only. Specifically, we find that self-supervised audio-video representations coupled with an alignment network trained on real videos produce more robust and consistent results. Our work raises awareness regarding dataset design and evaluation of deepfake detection. + +# Acknowledgments + +This work was supported in part by the EU Horizon projects AI4TRUST (No. 101070190) and ELIAS (No. 101120237), and by CNCS-UEFISCDI (PN-IV-P7-7.1-PTE-2024-0600). + +# References + +[1] Darius Afchar, Vincent Nozick, Junichi Yamagishi, and Isao Echizen. MesoNet: A compact facial video forgery detection network. In WIFS, page 1-7, 2018. 2, 7 +[2] Matyas Bohacek and Hany Farid. Lost in translation: Lipsync deepfake detection from audio-video mismatch. In CVPR, pages 4315-4323, 2024. 2 +[3] Stefano Borzi, Oliver Giudice, Filippo Stanco, and Dario Allegra. Is synthetic voice detection research going into the right direction? In CVPRW, pages 71-80, 2022. 1 +[4] Zhixi Cai, Shreya Ghosh, Abhinav Dhall, Tom Gedeon, Kalin Stefanov, and Munawar Hayat. Glitch in the matrix: A large scale benchmark for content driven audio-visual forgery detection and localization. Comput. Vis. Image Underst., 236:103818, 2023. 1, 2, 4 +[5] Zhixi Cai, Shreya Ghosh, Kalin Stefanov, Abhinav Dhall, Jianfei Cai, Hamid Rezatofighi, Reza Haffari, and Munawar Hayat. MARLIN: Masked autoencoder for facial video representation learning, 2023. 2, 7 +[6] Zhixi Cai, Shreya Ghosh, Aman Pankaj Adatia, Munawar Hayat, Abhinav Dhall, Tom Gedeon, and Kalin Stefanov. AV-Deepfake1M: A large-scale LLM-driven audio-visual deepfake dataset, 2024. 1, 2, 7 +[7] Edresson Casanova, Julian Weber, Christopher Dane Shulby, Arnaldo Candido Junior, Eren Gölgge, and Moacir A. Ponti. YourTTS: Towards zero-shot multi-speaker TTS and zero-shot voice conversion for everyone. In ICML, pages 2709-2720, 2022. 3 +[8] Lucy Chai, David Bau, Ser-Nam Lim, and Phillip Isola. What makes fake images detectable? understanding properties that generalize. In ECCV, pages 103-120, 2020. 1 +[9] François Chollet. Xception: Deep learning with depthwise separable convolutions. In CVPR, pages 1800-1807, 2017. 7 +[10] Komal Chugh, Parul Gupta, Abhinav Dhall, and Ramanathan Subramanian. Not made for each other-audiovisual dissonance-based deepfake detection and localization. In ACMM, pages 439-447, 2020. 4, 7 +[11] J. S. Chung, A. Nagrani, and A. Zisserman. VoxCeleb2: Deep speaker recognition. In Interspeech, 2018. 2 +[12] Davide Alessandro Coccomini, Nicola Messina, Claudio Gennaro, and Fabrizio Falchi. Combining EfficientNet and vision transformers for video deepfake detection. In ICIAP, pages 219-229, 2022. 7 +[13] Davide Cozzolino, Giovanni Poggi, Riccardo Corvi, Matthias Nießner, and Luisa Verdoliva. Raising the bar of AI-generated image detection with CLIP. In CVPR, pages 4356-4366, 2024. 4 + +[14] Davide Cozzolino, Giovanni Poggi, Matthias Nießner, and Luisa Verdoliva. Zero-shot detection of AI-generated images. In ECCV, pages 54-72, 2024. 2 +[15] Alexandre Défossez, Gabriel Synnaeve, and Yossi Adi. Real time speech enhancement in the waveform domain. In *Interspeech*, pages 3291–3295, 2020. 3 +[16] Brian Dolhansky, Joanna Bitton, Ben Pflaum, Jikuo Lu, Russ Howes, Menglin Wang, and Cristian Canton Ferrer. The deepfake detection challenge (dfdc) dataset. arXiv preprint arXiv:2006.07397, 2020. 4 +[17] Chao Feng, Ziyang Chen, and Andrew Owens. Self-supervised video forensics by audio-visual anomaly detection. In CVPR, pages 10491–10503, 2023. 2, 4, 5, 6 +[18] Alexandros Haliassos, Konstantinos Vougioukas, Stavros Petridis, and Maja Pantic. Lips don't lie: A generalisable and robust approach to face forgery detection. In CVPR, pages 5039-5049, 2021. 2 +[19] Alexandros Haliassos, Rodrigo Mira, Stavros Petridis, and Maja Pantic. Leveraging real talking faces via self-supervision for robust forgery detection. In CVPR, pages 14950-14962, 2022. 2 +[20] Ammarah Hashmi, Sahibzada Adil Shahzad, Chia-Wen Lin, Yu Tsao, and Hsin-Min Wang. AVTENet: Audio-visual transformer-based ensemble network exploiting multiple experts for video deepfake detection. CoRR, abs/2310.13103, 2023. 2 +[21] Baojin Huang, Zhongyuan Wang, Jifan Yang, Jiaxin Ai, Qin Zou, Qian Wang, and Dengpan Ye. Implicit identity driven deepfake face swapping detection. In CVPR, pages 4490-4499, 2023. 2 +[22] Ye Jia, Yu Zhang, Ron J. Weiss, Quan Wang, Jonathan Shen, Fei Ren, Zhifeng Chen, Patrick Nguyen, Ruoming Pang, Ignacio López-Moreno, and Yonghui Wu. Transfer learning from speaker verification to multispeaker text-to-speech synthesis. In NeurIPS, pages 4485-4495, 2018. 2 +[23] Hasam Khalid, Shahroz Tariq, Minha Kim, and Simon S. Woo. FakeAVCeleb: A novel audio-video multimodal deepfake dataset. In NeurIPS Datasets and Benchmarks, 2021. 1, 2 +[24] Jaehyeon Kim, Jungil Kong, and Juhee Son. Conditional variational autoencoder with adversarial learning for end-to-end text-to-speech. In ICML, pages 5530-5540, 2021. 3 +[25] Pavel Korshunov and Sébastien Marcel. DeepFakes: A new threat to face recognition? Assessment and detection. CoRR, abs/1812.08685, 2018. 2 +[26] Iryna Korshunova, Wenzhe Shi, Joni Dambre, and Lucas Theis. Fast face-swap using convolutional neural networks. In ICCV, pages 3697-3705, 2017. 2 +[27] Christos Koutlis and Symeon Papadopoulos. DiMoDif: Discourse modality-information differentiation for audio-visual deepfake detection and localization. CoRR, abs/2411.10193, 2024. 2 +[28] Patrick Kwon, Jaeseong You, Gyuhyeon Nam, Sungwoo Park, and Gyeongsu Chae. KoDF: A large-scale Korean deepfake detection dataset. In ICCV, pages 10744-10753, 2021. 2 + +[29] Xiaolou Li, Zehua Liu, Chen Chen, Lantian Li, Li Guo, and Dong Wang. Zero-shot fake video detection by audio-visual consistency. CoRR, abs/2406.07854, 2024. 2 +[30] Yachao Liang, Min Yu, Gang Li, Jianguo Jiang, Boquan Li, Feng Yu, Ning Zhang, Xiang Meng, and Weiqing Huang. SpeechForensics: Audio-visual speech representation learning for face forgery detection. In NeurIPS, 2024. 6 +[31] Weifeng Liu, Tianyi She, Jiawei Liu, Boheng Li, Dongyu Yao, Ziyou Liang, and Run Wang. Lips are lying: Spotting the temporal inconsistency between audio and visual in lip-syncing deepfakes. In NeurIPS, pages 91131-91155, 2024. 4, 8 +[32] Sara Mandelli, Paolo Bestagini, and Stefano Tubaro. When synthetic traces hide real content: Analysis of stable diffusion image laundering. CoRR, abs/2407.10736, 2024. 8 +[33] Hannes Mareen, Dimitrios Karageorgiou, Glenn Van Wallendael, Peter Lambert, and Symeon Papadopoulos. TGIF: Text-guided inpainting forgery dataset. CoRR, abs/2407.11566, 2024. 8 +[34] Francesco Marra, Diego Gragnaniello, Luisa Verdoliva, and Giovanni Poggi. Do GANs leave artificial fingerprints? In MIPR, pages 506-511, 2019. 2 +[35] Nicolas M. Müller, Franziska Dieckmann, Pavel Czempin, Roman Canals, and Konstantin Bötttinger. Speech is silver, silence is golden: What do ASVspoof-trained models really learn? CoRR, abs/2106.12914, 2021. 1, 4 +[36] Nicolas M. Müller, Piotr Kawa, Wei Herng Choong, Edresson Casanova, Eren Gölgge, Thorsten Müller, Piotr Syga, Philip Sperl, and Konstantin Bötttinger. MLAAD: The multilanguage audio anti-spoofing dataset. In IJCNN, pages 1-7, 2024. 1 +[37] Yuval Nirkin, Yosi Keller, and Tal Hassner. FSGAN: Subject agnostic face swapping and reenactment. In ICCV, pages 7183-7192, 2019. 2 +[38] Utkarsh Ojha, Yuheng Li, and Yong Jae Lee. Towards universal fake image detectors that generalize across generative models. In CVPR, pages 24480-24489, 2023. 2, 4 +[39] Trevine Oorloff, Surya Koppisetti, Nicolò Bonettini, Dipyaraj Solanki, Ben Colman, Yaser Yacoob, Ali Shahriyari, and Gaurav Bharaj. AVFF: Audio-visual feature fusion for video deepfake detection. In CVPR, pages 27102-27112, 2024. 2, 5 +[40] Octavian Pascu, Adriana Stan, Dan Oneata, Elisabeta Oneata, and Horia Cucu. Towards generalisable and calibrated audio deepfake detection with self-supervised representations. In Interspeech, pages 4828-4832, 2024. 4 +[41] Alessandro Pianese, Davide Cozzolino, Giovanni Poggi, and Luisa Verdoliva. Training-free deepfake voice recognition by leveraging large-scale pre-trained models. In ACM Workshop on Information Hiding and Multimedia Security, page 289–294, 2024. 2, 4 +[42] K R Prajwal, Rudrabha Mukhopadhyay, Vinay P. Namboodiri, and C.V. Jawahar. A lip sync expert is all you need for speech to lip generation in the wild. In ACMM, 2020. 2 +[43] Anirudh Sundara Rajan, Utkarsh Ojha, Jedidiah Schloesser, and Yong Jae Lee. On the effectiveness of dataset alignment for fake image detection. CoRR, abs/2410.11835, 2024. 1, 8 + +[44] Tal Reiss, Bar Cavia, and Yedid Hoshen. Detecting deep-fakes without seeing any. CoRR, abs/2311.01458, 2023. 2 +[45] Jonas Ricker, Denis Lukovnikov, and Asja Fischer. AEROBLADE: Training-free detection of latent diffusion images using autoencoder reconstruction error. In CVPR, pages 9130-9140, 2024. 2 +[46] Andreas Rossler, Davide Cozzolino, Luisa Verdoliva, Christian Riess, Justus Thies, and Matthias Niessner. FaceForensics: Learning to detect manipulated facial images. In ICCV, 2019. 1 +[47] Davide Salvi, Brian C. Hosler, Paolo Bestagini, Matthew C. Stamm, and Stefano Tubaro. TIMIT-TTS: A text-to-speech dataset for multimodal synthetic media detection. IEEE Access, 11:50851-50866, 2023. 2 +[48] Sahibzada Adil Shahzad, Ammarah Hashmi, Yan-Tsung Peng, Yu Tsao, and Hsin-Min Wang. AV-Lip-Sync+: Leveraging AV-HuBERT to exploit multimodal inconsistency for video deepfake detection. CoRR, abs/2311.02733, 2023. 2 +[49] Bowen Shi, Wei-Ning Hsu, Kushal Lakhotia, and Abdelrahman Mohamed. Learning audio-visual speech representation by masked multimodal cluster prediction. In ICLR, 2022. 2, 5, 8 +[50] Kaede Shiohara and Toshihiko Yamasaki. Detecting deepfakes with self-blended images. In CVPR, pages 18699-18708, 2022. 2 +[51] Stefan Smeu, Elisabeta Oneata, and Dan Oneata. DeCLIP: Decoding CLIP representations for deepfake localization. CoRR, abs/2409.08849, 2024. 8 +[52] Joon Son Chung, Andrew Senior, Oriol Vinyals, and Andrew Zisserman. Lip reading sentences in the wild. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6447-6456, 2017. 4 +[53] Hemlata Tak, Jose Patino, Massimiliano Todisco, Andreas Nautsch, Nicholas Evans, and Anthony Larcher. End-to-end anti-spoofing with RawNet2. In ICASSP, pages 6369-6373, 2021. 4 +[54] Aäron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. CoRR, abs/1807.03748, 2018. 5 +[55] Jiadong Wang, Xinyuan Qian, Malu Zhang, Robby T. Tan, and Haizhou Li. Seeing what you said: Talking face generation guided by a lip reading expert. In CVPR, pages 14653-14662, 2023. 3 +[56] Xin Wang, Junichi Yamagishi, Massimiliano Todisco, Héctor Delgado, Andreas Nautsch, Nicholas W. D. Evans, Md. Sahidullah, Ville Vestman, Tomi Kinnunen, Kong Aik Lee, Lauri Juvela, Paavo Alku, Yu-Huai Peng, Hsin-Te Hwang, Yu Tsao, Hsin-Min Wang, Sébastien Le Maguer, Markus Becker, and Zhen-Hua Ling. ASVspoof 2019: A large-scale public database of synthesized, converted and replayed speech. Comput. Speech Lang., 64:101114, 2020. 1 +[57] Zhiyuan Yan, Taiping Yao, Shen Chen, Yandan Zhao, Xinghe Fu, Junwei Zhu, Donghao Luo, Li Yuan, Chengjie Wang, Shouhong Ding, and Yunsheng Wu. DF40: Toward next-generation deepfake detection. CoRR, abs/2406.13495, 2024.8 +[58] Wenyuan Yang, Xiaoyu Zhou, Zhikai Chen, Bofei Guo, Zhongjie Ba, Zhihua Xia, Xiaochun Cao, and Kui Ren. + +AVoiD-DF: Audio-visual joint learning for detecting deepfake. IEEE Trans. Inf. Forensics Secur., 18:2015-2029, 2023. 2 +[59] Ning Yu, Larry S Davis, and Mario Fritz. Attributing fake images to GANs: Learning and analyzing GAN fingerprints. In CVPR, pages 7556-7566, 2019. 2 +[60] Yinglin Zheng, Jianmin Bao, Dong Chen, Ming Zeng, and Fang Wen. Exploring temporal coherence for more general video face forgery detection. In ICCV, pages 15024-15034, 2021. 2 \ No newline at end of file diff --git a/2025/Circumventing Shortcuts in Audio-visual Deepfake Detection Datasets with Unsupervised Learning/images.zip b/2025/Circumventing Shortcuts in Audio-visual Deepfake Detection Datasets with Unsupervised Learning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..ca35eb2129bad3eb8c6bd5d0dd80697a625187f7 --- /dev/null +++ b/2025/Circumventing Shortcuts in Audio-visual Deepfake Detection Datasets with Unsupervised Learning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9a68e845aa93d69efe9ab6849ce0ba6073aac4defbe2ebe9724768824c8231e0 +size 359735 diff --git a/2025/Circumventing Shortcuts in Audio-visual Deepfake Detection Datasets with Unsupervised Learning/layout.json b/2025/Circumventing Shortcuts in Audio-visual Deepfake Detection Datasets with Unsupervised Learning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..cfb99579f8a33831696a80ad7753f7a2923b3d23 --- /dev/null +++ b/2025/Circumventing Shortcuts in Audio-visual Deepfake Detection Datasets with Unsupervised Learning/layout.json @@ -0,0 +1,8302 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 80, + 103, + 531, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 103, + 531, + 140 + ], + "spans": [ + { + "bbox": [ + 80, + 103, + 531, + 140 + ], + "type": "text", + "content": "Circumventing shortcuts in audio-visual deepfake detection datasets with unsupervised learning" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 102, + 160, + 506, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 160, + 506, + 190 + ], + "spans": [ + { + "bbox": [ + 102, + 160, + 506, + 190 + ], + "type": "text", + "content": "Stefan Smeu\\*1 Dragos-Alexandru Boldisor\\*1 Dan Oneata2 Elisabeta Oneata1 1Bitdefender 2POLITEHNICA Bucharest" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 127, + 191, + 480, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 191, + 480, + 204 + ], + "spans": [ + { + "bbox": [ + 127, + 191, + 480, + 204 + ], + "type": "text", + "content": "{ssmeu, dboldisor, eoneata}@bitdefender.com dan. oneata@gmail.com" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "spans": [ + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 255, + 296, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 255, + 296, + 483 + ], + "spans": [ + { + "bbox": [ + 53, + 255, + 296, + 483 + ], + "type": "text", + "content": "Good datasets are essential for developing and benchmarking any machine learning system. Their importance is even more extreme for safety critical applications such as deepfake detection—the focus of this paper. Here we reveal that two of the most widely used audio-video deepfake datasets suffer from a previously unidentified spurious feature: the leading silence. Fake videos start with a very brief moment of silence and, on the basis of this feature alone, we can separate the real and fake samples almost perfectly. As such, previous audio-only and audio-video models exploit the presence of silence in the fake videos and consequently perform worse when the leading silence is removed. To circumvent latching on such an unwanted artifact and possibly other unrevealed ones, we propose a shift from supervised to unsupervised learning by training models exclusively on real data. We show that by aligning self-supervised audio-video representations we remove the risk of relying on dataset-specific biases and improve robustness in deepfake detection." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 506, + 135, + 520 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 506, + 135, + 520 + ], + "spans": [ + { + "bbox": [ + 55, + 506, + 135, + 520 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 527, + 295, + 610 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 527, + 295, + 610 + ], + "spans": [ + { + "bbox": [ + 55, + 527, + 295, + 610 + ], + "type": "text", + "content": "Manipulated videos represent a threat to society as they have the potential of misleading people into believing actors with malicious intents. By spreading misinformation on social media platforms, people may be exposed to scams (e.g., identity theft operations), conspiracy theories and political misinformation. Therefore, deepfake detection methods are essential tools on a global scale." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 611, + 295, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 611, + 295, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 611, + 295, + 696 + ], + "type": "text", + "content": "The progress in automated deepfake detection is fueled by the datasets developed by the research community. Good quality data is essential for both training and benchmarking the progress of these methods. In recent years, numerous datasets have been proposed. Among the audio-video deepfake datasets, there have been released datasets that alter both streams [6, 23] or only one [36, 46, 56]; datasets with" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 334, + 232, + 535, + 338 + ], + "blocks": [ + { + "bbox": [ + 334, + 232, + 535, + 338 + ], + "lines": [ + { + "bbox": [ + 334, + 232, + 535, + 338 + ], + "spans": [ + { + "bbox": [ + 334, + 232, + 535, + 338 + ], + "type": "image", + "image_path": "88ea4fd6679afa749cb4379af534874a4687fbf669bfa7767bf2cbb9d0414013.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 347, + 555, + 403 + ], + "lines": [ + { + "bbox": [ + 313, + 347, + 555, + 403 + ], + "spans": [ + { + "bbox": [ + 313, + 347, + 555, + 403 + ], + "type": "text", + "content": "Figure 1. Audio-visual deepfake detection datasets have a silence bias: fake samples start with a brief moment of silence, which is not the case for real samples. Here we show the first " + }, + { + "bbox": [ + 313, + 347, + 555, + 403 + ], + "type": "inline_equation", + "content": "62.5\\mathrm{ms}" + }, + { + "bbox": [ + 313, + 347, + 555, + 403 + ], + "type": "text", + "content": " of the audio waveform for a real and the corresponding fake sample from the AV-Deepfake1M dataset [6]." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 425, + 514, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 425, + 514, + 437 + ], + "spans": [ + { + "bbox": [ + 313, + 425, + 514, + 437 + ], + "type": "text", + "content": "full-video [23] or only local [4, 6] manipulations." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 312, + 437, + 555, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 437, + 555, + 593 + ], + "spans": [ + { + "bbox": [ + 312, + 437, + 555, + 593 + ], + "type": "text", + "content": "However, care has to be taken to ensure a good deepfake detection dataset. Any asymmetry in the preparation of fakes and reals can result in biases that correlate spuriously with the groundtruth label. For example, in the image domain, different preprocessing pipelines [8] or types of resizing [43] between real and fake samples paint an overly optimistic picture. In the audio domain, the very popular ASVspoof19 dataset [56] has been shown to leak information about labels in the form of silence duration [35] or bitrate information [3]. Since many of the deepfake detection models are high capacity, they can easily use such artifacts as shortcuts for learning, consequently greatly impacting their ability to generalize in a real case scenario." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 594, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 556, + 714 + ], + "type": "text", + "content": "In this work, we first expose a bias present in two widely-adopted audio-video deepfake detection datasets, FakeAVCeleb [23] and AV-Deepfake1M [6]: a short moment of silence present at the beginning of manipulated videos (see Figure 1). Based on this information, a very simple silence classifier can reach a near-perfect performance of over " + }, + { + "bbox": [ + 313, + 594, + 556, + 714 + ], + "type": "inline_equation", + "content": "98\\%" + }, + { + "bbox": [ + 313, + 594, + 556, + 714 + ], + "type": "text", + "content": " on both FakeAVCeleb and AV-Deepfake1M. Moreover, we observe that the silence distribution is similar across datasets, implying that prior methods may have potentially over estimated the generalization" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 703, + 134, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 134, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 134, + 713 + ], + "type": "text", + "content": "*Equal contribution." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18815" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 201, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 201, + 83 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 201, + 83 + ], + "type": "text", + "content": "performance based on this shortcut." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 84, + 294, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 84, + 294, + 239 + ], + "spans": [ + { + "bbox": [ + 55, + 84, + 294, + 239 + ], + "type": "text", + "content": "Introducing such biases when creating the datasets is unavoidable and there might be others that are more subtle and harder to reveal. To circumvent this problem, we show that one solution is a shift in learning paradigm, from the more common supervised setup to the unsupervised one [17, 44]. Removing fake samples from training eliminates the focus on the asymmetries induced by spurious artifacts. Moreover, this also has the potential to improve generalization among different manipulation techniques, since supervised detection methods tend to over-rely on generator-specific fingerprints [34, 59]. By limiting to real data, we can also naturally leverage self-supervised representations, which have been shown to improve generalization [38]." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 240, + 294, + 347 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 240, + 294, + 347 + ], + "spans": [ + { + "bbox": [ + 55, + 240, + 294, + 347 + ], + "type": "text", + "content": "To this end, we propose an approach, named AVH-Align, that learns a frame-level audio-video alignment score on top of AV-HuBERT features [49]. Since tampered videos are expected to have greater audio-video desynchronizations, this alignment score can effectively differentiate fake and real videos. We show that this approach is robust with respect to the identified shortcut of leading silence, and also outperforms other methods that do not use the silence bias, while not seeing any fake samples at training time." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 348, + 294, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 348, + 294, + 443 + ], + "spans": [ + { + "bbox": [ + 55, + 348, + 294, + 443 + ], + "type": "text", + "content": "To summarize, our work makes the following contributions: 1. We expose a previously unknown spurious feature in two of the most widely adopted deepfake detection datasets. 2. We analyze the impact of this shortcut on various state-of-the-art models. 3. We show that a way to mitigate such shortcuts is by training on real data only and we introduce a new method in this direction. Our code is available at: https://github.com/bit-ml/AVH-Align." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 454, + 139, + 467 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 454, + 139, + 467 + ], + "spans": [ + { + "bbox": [ + 55, + 454, + 139, + 467 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 480, + 295, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 480, + 295, + 684 + ], + "spans": [ + { + "bbox": [ + 55, + 480, + 295, + 684 + ], + "type": "text", + "content": "Audio-video deepfake detection. Many approaches for deepfake detection on videos have focused on the visual stream of information only [1, 5, 18, 19, 21, 50, 60]. But with the recent introduction of audio-visual datasets, (FakeAVCeleb [23], Deepfake TIMIT [25, 47], KODF [28], LAV-DF [4], AV-Deepfake-1M [6]), more research has shifted towards models that exploit both audio and video cues [17, 20, 27, 39, 48, 58]. An emerging trend in this direction is the use of pretrained representations in a self-supervised way [17, 39]. But, different from our approach, these methods train the representations from scratch and use them as a first step in a more elaborated pipeline: supervised classification [39] or anomaly detection [17]. There are also works that similarly to us exploit the pretrained audio-video AV-HuBERT model [49] to extract representations [20, 48], but all of those methods are trained in the fully supervised learning paradigm." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": "Unsupervised deepfake detection. To make detection more generalizable across generators, a new direction is" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 553, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 335 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 335 + ], + "type": "text", + "content": "to depart from the supervised paradigm and resort only to real data. There are two main classes of such unsupervised approaches: methods that rely on consistency checks and methods that treat the problem as an anomaly detection one. Among the first class, prior work proposed to verify that audio and visual streams align at a semantic level, for example, from the point of view of spoken content [2, 29], or at a representation level, based on the alignment of audio and video features [17, 44]. Consistency checks have also been used with respect to the identity of a speaker (comparing a query sample against real audio [41] or real images [44] of the target speaker) or between image and text modalities [44]. For the second class, deepfake detection as anomaly detection, recent work in the image domain has used the reconstruction loss of the query image to tell whether it is anomalous [14, 45]; for example, Ricker et al. [45] make the observation that images generated by a latent diffusion model (LDM) are easier to reconstruct by the LDM than real images. For the audio-visual domain, Feng et al. [17] use both classes of approaches: they use consistency checks to estimate synchronization between the two streams and then flag anomalies using density estimation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 348, + 509, + 361 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 348, + 509, + 361 + ], + "spans": [ + { + "bbox": [ + 313, + 348, + 509, + 361 + ], + "type": "text", + "content": "3. Silence bias in audio-video datasets" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 369, + 553, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 369, + 553, + 464 + ], + "spans": [ + { + "bbox": [ + 313, + 369, + 553, + 464 + ], + "type": "text", + "content": "In this section, we show that two popular datasets (Sec. 3.1) have a silence bias. We analyze its behavior and show that a simple classifier based on the leading silence alone can obtain almost perfect separation between fake and real samples (Sec. 3.2). This implies that the performance of prior work is susceptible to have been overestimated. For this reason we analyze its impact on various audio and audio-visual methods (Sec. 3.3)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 473, + 376, + 484 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 473, + 376, + 484 + ], + "spans": [ + { + "bbox": [ + 313, + 473, + 376, + 484 + ], + "type": "text", + "content": "3.1. Datasets" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 492, + 553, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 492, + 553, + 612 + ], + "spans": [ + { + "bbox": [ + 313, + 492, + 553, + 612 + ], + "type": "text", + "content": "We consider two audio-visual datasets in our analysis: FakeAVCeleb [23] and AV-Deepfake1M [6]. They distinguish mainly in the fact that the first contains fully-generated video sequences, while the second contains partially-manipulated sequences. Both are based on the VoxCeleb2 dataset [11], which consists of YouTube audio-video of celebrities. Apart from the real samples—real video real audio (RVRA)—both datasets include three types of fake videos: real video fake audio (RVFA), fake video real audio (FVRA), and fake video fake audio (FVFA)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 617, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 555, + 715 + ], + "type": "text", + "content": "FakeAVCeleb contains 500 real videos from VoxCeleb2 and 19.5k fake videos (10k FVFA, 9k FVRA, 500 RVFA). The fake visual content was generated with face swapping methods (Faceswap [26] and FSGAN [37]) or the Wav2Lip lipSyncing approach [42]. The fake audio content was generate with the voice cloning tool SV2TTS [22]. The dataset is diverse across age groups, genders, races, as well as with respect to the number of subjects in a single video, their" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "18816" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 70, + 294, + 182 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 294, + 182 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 294, + 182 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 294, + 182 + ], + "type": "image", + "image_path": "ac2fa88c7ea9a187a45c81f5d8eb3c9c641ae51ddf7a2a38b3ad0df6f89f25ee.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 192, + 296, + 237 + ], + "lines": [ + { + "bbox": [ + 55, + 192, + 296, + 237 + ], + "spans": [ + { + "bbox": [ + 55, + 192, + 296, + 237 + ], + "type": "text", + "content": "Figure 2. Normalized distribution plots of the leading silence duration for real and fake videos in the FakeAVCeleb (left) and AV-Deepfake1M (right) datasets. The fake samples start with 25–30 ms of silence." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 258, + 295, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 258, + 295, + 295 + ], + "spans": [ + { + "bbox": [ + 55, + 258, + 295, + 295 + ], + "type": "text", + "content": "placement, the visual and audio quality. In our experiments, we split the dataset in " + }, + { + "bbox": [ + 55, + 258, + 295, + 295 + ], + "type": "inline_equation", + "content": "70\\%" + }, + { + "bbox": [ + 55, + 258, + 295, + 295 + ], + "type": "text", + "content": " (train and validation) and " + }, + { + "bbox": [ + 55, + 258, + 295, + 295 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 55, + 258, + 295, + 295 + ], + "type": "text", + "content": " (test). This split is kept consistent across all experiments." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 300, + 296, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 300, + 296, + 517 + ], + "spans": [ + { + "bbox": [ + 55, + 300, + 296, + 517 + ], + "type": "text", + "content": "AV-Deepfake1M is a large scale dataset, which consists of over one million videos and 2k subjects. As opposed to the FakeAVCeleb dataset, the manipulations here are local and consists of word-level replacements, insertions and deletions. The text manipulations are generated with the ChatGPT large language model. The fake video content is generated with the lipSyncing method TalkLip [55], while the fake audio content is generated with the VITS [24] or YourTTS [7] methods. The authors ensure that the synthesized words share the same background noise with the full audio, by first extracting the audio noise with the Denoiser method [15] and then adding it to the synthesized words. The dataset is originally split into train, validation and test, with the test split having a different set of speakers than those encountered in the train and validation splits. For our experiments, we select training and validation samples from the original training split, and evaluate on 10k samples from the original validation split or on the full official test set." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 525, + 201, + 539 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 525, + 201, + 539 + ], + "spans": [ + { + "bbox": [ + 55, + 525, + 201, + 539 + ], + "type": "text", + "content": "3.2. Analysis of leading silence" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 544, + 295, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 544, + 295, + 640 + ], + "spans": [ + { + "bbox": [ + 55, + 544, + 295, + 640 + ], + "type": "text", + "content": "We start by analyzing the silence distribution of the real and fake samples in the two considered datasets. We define the duration of the leading silence as the moment when the magnitude of the audio exceeds a certain threshold " + }, + { + "bbox": [ + 55, + 544, + 295, + 640 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 55, + 544, + 295, + 640 + ], + "type": "text", + "content": ". For this experiment, we select this threshold to be " + }, + { + "bbox": [ + 55, + 544, + 295, + 640 + ], + "type": "inline_equation", + "content": "5 \\cdot 10^{-4}" + }, + { + "bbox": [ + 55, + 544, + 295, + 640 + ], + "type": "text", + "content": ", but as we will shortly see, the results are robust to its choice. We carry this analysis only on the real (RVRA) and fully fake (FVFA) videos from each dataset's test set." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "type": "text", + "content": "The results are shown in Figure 2. We observe that the real videos start with noise, while the fake samples have a leading silence of around " + }, + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "type": "inline_equation", + "content": "25 - 30\\mathrm{ms}" + }, + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "type": "text", + "content": ". The silence duration of fake samples is similar for both datasets, although the distribution is much sharper for the AV-Deepfake1M dataset. If we were to rank the samples based on this feature we" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 315, + 69, + 555, + 182 + ], + "blocks": [ + { + "bbox": [ + 315, + 69, + 555, + 182 + ], + "lines": [ + { + "bbox": [ + 315, + 69, + 555, + 182 + ], + "spans": [ + { + "bbox": [ + 315, + 69, + 555, + 182 + ], + "type": "image", + "image_path": "e6c33b744cbca2d22ca98c8b5ac0f177e281b915cfcd47b16348cdd8f2267e97.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 190, + 555, + 224 + ], + "lines": [ + { + "bbox": [ + 313, + 190, + 555, + 224 + ], + "spans": [ + { + "bbox": [ + 313, + 190, + 555, + 224 + ], + "type": "text", + "content": "Figure 3. Left: The impact of the silence threshold on the leading silence classifier. Right: The impact of the leading duration on the maximum amplitude classifier." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 246, + 555, + 269 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 246, + 555, + 269 + ], + "spans": [ + { + "bbox": [ + 313, + 246, + 555, + 269 + ], + "type": "text", + "content": "would obtain an area under curve of the receiver operating characteristic curve (AUC) of over " + }, + { + "bbox": [ + 313, + 246, + 555, + 269 + ], + "type": "inline_equation", + "content": "98\\%" + }, + { + "bbox": [ + 313, + 246, + 555, + 269 + ], + "type": "text", + "content": " for either datasets." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 275, + 556, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 275, + 556, + 361 + ], + "spans": [ + { + "bbox": [ + 313, + 275, + 556, + 361 + ], + "type": "text", + "content": "What counts as silence? For the previous experiment we have considered that silence is the signal that has an amplitude lower than " + }, + { + "bbox": [ + 313, + 275, + 556, + 361 + ], + "type": "inline_equation", + "content": "\\tau = 5 \\cdot 10^{-4}" + }, + { + "bbox": [ + 313, + 275, + 556, + 361 + ], + "type": "text", + "content": ". We investigate how sensitive the performance is to this threshold. We vary " + }, + { + "bbox": [ + 313, + 275, + 556, + 361 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 275, + 556, + 361 + ], + "type": "text", + "content": " across a grid of values and show the results in Figure 3 (left). We observe that the results are strong as long as this threshold is small enough." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 365, + 556, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 365, + 556, + 485 + ], + "spans": [ + { + "bbox": [ + 313, + 365, + 556, + 485 + ], + "type": "text", + "content": "A different perspective: Maximum amplitude. Instead of looking at the silence duration, we can alternatively measure the maximum amplitude in the first " + }, + { + "bbox": [ + 313, + 365, + 556, + 485 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 313, + 365, + 556, + 485 + ], + "type": "text", + "content": " seconds of the audio. Figure 3 shows the AUC obtained by ranking the samples based on this feature for various leading durations " + }, + { + "bbox": [ + 313, + 365, + 556, + 485 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 313, + 365, + 556, + 485 + ], + "type": "text", + "content": ". We observe that this alternative measure also yields strong results (over " + }, + { + "bbox": [ + 313, + 365, + 556, + 485 + ], + "type": "inline_equation", + "content": "98\\%" + }, + { + "bbox": [ + 313, + 365, + 556, + 485 + ], + "type": "text", + "content": " AUC), and that the optimum is obtained for a " + }, + { + "bbox": [ + 313, + 365, + 556, + 485 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 313, + 365, + 556, + 485 + ], + "type": "text", + "content": " of around " + }, + { + "bbox": [ + 313, + 365, + 556, + 485 + ], + "type": "inline_equation", + "content": "30~\\mathrm{ms}" + }, + { + "bbox": [ + 313, + 365, + 556, + 485 + ], + "type": "text", + "content": ", in line with the previous experiment. As we extend the window, the performance gets closer to random chance, AUC of " + }, + { + "bbox": [ + 313, + 365, + 556, + 485 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 313, + 365, + 556, + 485 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 491, + 556, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 491, + 556, + 624 + ], + "spans": [ + { + "bbox": [ + 313, + 491, + 556, + 624 + ], + "type": "text", + "content": "Other biases: Volume and trailing silence. If we extend the duration " + }, + { + "bbox": [ + 313, + 491, + 556, + 624 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 313, + 491, + 556, + 624 + ], + "type": "text", + "content": " to cover the entire audio, we get an estimate of the maximum volume of the audio. This feature yields an AUC of " + }, + { + "bbox": [ + 313, + 491, + 556, + 624 + ], + "type": "inline_equation", + "content": "67.6\\%" + }, + { + "bbox": [ + 313, + 491, + 556, + 624 + ], + "type": "text", + "content": " on FakeAVCeleb, which is much less than " + }, + { + "bbox": [ + 313, + 491, + 556, + 624 + ], + "type": "inline_equation", + "content": "98.4\\%" + }, + { + "bbox": [ + 313, + 491, + 556, + 624 + ], + "type": "text", + "content": ", but still over random chance (50% AUC). Similarly, we have also investigated the trailing silence (the silence at the end of the audio) and have obtained AUC of over " + }, + { + "bbox": [ + 313, + 491, + 556, + 624 + ], + "type": "inline_equation", + "content": "99\\%" + }, + { + "bbox": [ + 313, + 491, + 556, + 624 + ], + "type": "text", + "content": " for FakeAVCeleb. AV-Deepfake1M is less sensitive to these other biases, with values closer to " + }, + { + "bbox": [ + 313, + 491, + 556, + 624 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 313, + 491, + 556, + 624 + ], + "type": "text", + "content": ". While these biases are not as consistent as the leading silence, they are still problematic." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "content": "Why do fake samples have a leading silence? Given that we do not have access to generation process of the two datasets, it is challenging to pinpoint the exact reason for the occurrence of the leading silence bias. However, we speculate that this happens when the audio may be slightly shorter than the video counterpart. Note that this is a different reason from the silence observed in audio only datasets" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18817" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "content": "[35]. There the real audios had a leading silence, while the synthesized speech was silence-free. In the case of audio-video datasets it might be challenging to completely avoid this problem, but an easy solution (for this particular bias) is to trim the leading silence. This is what we do in the next experiments." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 149, + 295, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 149, + 295, + 293 + ], + "spans": [ + { + "bbox": [ + 55, + 149, + 295, + 293 + ], + "type": "text", + "content": "Biases in other datasets. We consider three more datasets: LAV-DF [4], AVLips [31], DFDC [16]. We find that LAV-DF has a slight trailing silence bias (65% AUC), but it does not exhibit the other two biases (leading silence and volume). In the case of AVLips and DFDC we did not identify any of the three biases. This may happen because the samples in these datasets have real audio and the unmodified audio matches the visual sequence more precisely. This is confirmed by a fine-grained analysis on FakeAVCeleb, where we observe that on the FVRA split the biases are diminished (62.4% AUC for lead silence), while they persist on the RVFA split (100% AUC for lead silence)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 300, + 180, + 313 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 300, + 180, + 313 + ], + "spans": [ + { + "bbox": [ + 55, + 300, + 180, + 313 + ], + "type": "text", + "content": "3.3. Impact on prior work" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 319, + 295, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 319, + 295, + 376 + ], + "spans": [ + { + "bbox": [ + 55, + 319, + 295, + 376 + ], + "type": "text", + "content": "We have shown that real and fake samples differ in terms of the leading silence. This is a simple feature which could be learned by the high-capacity neural networks. Here we investigate whether that is indeed the case for three existing methods:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 378, + 295, + 533 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 56, + 378, + 295, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 378, + 295, + 426 + ], + "spans": [ + { + "bbox": [ + 56, + 378, + 295, + 426 + ], + "type": "text", + "content": "- RawNet2 [53], which is an audio-only method that operates on the raw waveform. Its architecture sequences sinc layers, convolutions, gated recurrent units and fully connected layers." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 426, + 295, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 426, + 295, + 474 + ], + "spans": [ + { + "bbox": [ + 56, + 426, + 295, + 474 + ], + "type": "text", + "content": "- MDS [10] (modality dissonance score), which is an audio-visual method that estimates the mismatch between audio and video segments. The score is computed as the distance between audio and visual features." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 475, + 295, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 475, + 295, + 533 + ], + "spans": [ + { + "bbox": [ + 56, + 475, + 295, + 533 + ], + "type": "text", + "content": "- AVAD [17] (audio-visual anomaly detection), which is an audio-visual method trained on real data only. The approach has two steps: first, it estimates the desynchronization between audio and video; then it estimates whether these patterns are typical of real data or anomalous." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 534, + 295, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 295, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 295, + 605 + ], + "type": "text", + "content": "We train the RawNet2 and MDS methods on both datasets using the code provided by the authors. For the AVAD we use the provided checkpoint (trained on LRS [52]) and do not retrain it since training code is not available. As a baseline we include our leading silence duration classifier described in the previous section." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 606, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 606, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 606, + 295, + 713 + ], + "type": "text", + "content": "We evaluate all methods in two settings: on the corresponding evaluation set and on a trimmed version of the same evaluation set. For the trimmed version, we discard the beginning of videos such that each of them starts with the first non-silent segment. We find the duration to discard by first computing the leading silence (using the same " + }, + { + "bbox": [ + 55, + 606, + 295, + 713 + ], + "type": "inline_equation", + "content": "5 \\cdot 10^{-4}" + }, + { + "bbox": [ + 55, + 606, + 295, + 713 + ], + "type": "text", + "content": " threshold on the magnitude) and then rounding it up to the nearest multiple of " + }, + { + "bbox": [ + 55, + 606, + 295, + 713 + ], + "type": "inline_equation", + "content": "1 / \\text{FPS}" + }, + { + "bbox": [ + 55, + 606, + 295, + 713 + ], + "type": "text", + "content": " (the reciprocal of the video's frame rate); this rounding ensures that the audio and" + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 315, + 70, + 554, + 145 + ], + "blocks": [ + { + "bbox": [ + 315, + 70, + 554, + 145 + ], + "lines": [ + { + "bbox": [ + 315, + 70, + 554, + 145 + ], + "spans": [ + { + "bbox": [ + 315, + 70, + 554, + 145 + ], + "type": "table", + "html": "
MethodMod.FakeAVCelebAV-Deepfake1M
Trim: XTrim: √Trim: XTrim: √
Silence classifierA98.454.8 ↓ 43.698.250.6 ↓ 47.6
RawNet2 [53]A99.997.3 ↓ 2.699.988.1 ↓ 11.8
MDS [10]AV90.473.8 ↓ 16.699.254.9 ↓ 44.3
AVAD [17]AV95.295.2 ≈ 0.052.952.9 ≈ 0.0
", + "image_path": "a4684493e663c7f8ab10a1945e8dd7e3b2783046f6764ab5a4c9f240e4548984.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 153, + 553, + 196 + ], + "lines": [ + { + "bbox": [ + 313, + 153, + 553, + 196 + ], + "spans": [ + { + "bbox": [ + 313, + 153, + 553, + 196 + ], + "type": "text", + "content": "Table 1. The impact of leading silence on the performance (area under the receiver operator characteristic curve; AUC) for three existing deepfake detection methods and the silence classifier. Results are shown for RVRA-FVFA subsets." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 221, + 553, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 221, + 553, + 245 + ], + "spans": [ + { + "bbox": [ + 313, + 221, + 553, + 245 + ], + "type": "text", + "content": "video channels remain synchronized. For training, we do not perform trimming, but use the original dataset." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 247, + 553, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 247, + 553, + 317 + ], + "spans": [ + { + "bbox": [ + 313, + 247, + 553, + 317 + ], + "type": "text", + "content": "The results are shown in Table 1. First, we observe that the audio-based RawNet2 is the top performing method. Moreover, we see that the leading silence classifier is generally better than the two other approaches: MDS and AVAD. The relatively stable performance of AVAD indicates that this method dose not latch on the silence information." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 319, + 554, + 546 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 319, + 554, + 546 + ], + "spans": [ + { + "bbox": [ + 313, + 319, + 554, + 546 + ], + "type": "text", + "content": "In terms of the impact of the leading silence, we observe that removing it through trimming affects methods differently. As expected, the silence classifier is affected the most and its performance drops down to random chance on the trimmed data. RawNet2 is not as affected on the FakeAVCeleb, presumably because there is still enough information throughout the signal, but it suffers a larger hit on the more challenging AV-DeepFake1M, which has only partially-manipulated samples. MDS is affected by the leading silence bias on both FakeAVCeleb and AV-Deepfake1M, though significantly stronger on AV-Deepfake1M. AVAD is the most robust method, since this approach does not explicitly model the silence information. Finally, we notice that the RawNet2 audio-only method is best even on the trimmed settings. This is noteworthy since it is the first time an audio-only method has been applied to these datasets. This suggests that the audio stream is an important source of information that is often overlooked by prior work." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 560, + 542, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 560, + 542, + 574 + ], + "spans": [ + { + "bbox": [ + 313, + 560, + 542, + 574 + ], + "type": "text", + "content": "4. Modeling real data for deepfake detection" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 582, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 554, + 713 + ], + "type": "text", + "content": "The previous section indicated that audio-visual datasets exhibit a silence bias which can easily be exploited. We want to develop a method that is robust to this bias (and possibly other uncovered ones) and still performs well. We have seen that models trained only on real data [17] are promising in being robust to the silence shortcut, but the performance was modest. To further improve them we propose to build on top of audio-focused self-supervised features. Self-supervised features have shown strong generalization for both visual [13, 38] and audio [40, 41] deepfake detection. We choose audio-focused self-supervised features be" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18818" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "text", + "content": "cause the audio models showed strong performance in the previous section. Note that we cannot rely on audio-only models because there are cases where manipulations appear only in the visual domain (the fake video, real audio case)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 129, + 115, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 129, + 115, + 140 + ], + "spans": [ + { + "bbox": [ + 55, + 129, + 115, + 140 + ], + "type": "text", + "content": "4.1. Method" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 148, + 296, + 244 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 148, + 296, + 244 + ], + "spans": [ + { + "bbox": [ + 55, + 148, + 296, + 244 + ], + "type": "text", + "content": "We propose a method that aligns AV-HuBERT [49] features on real data. First, we extract audio and visual frame-level features with a pretrained AV-HuBERT model. Then, on top of these features we learn a network to better align them. The alignment network is learnt on real samples by matching each video frame to its corresponding audio frame [17]. We call our method AVH-Align (AV-HuBERT Aligned) and show its depiction in Figure 4." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 250, + 296, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 250, + 296, + 417 + ], + "spans": [ + { + "bbox": [ + 55, + 250, + 296, + 417 + ], + "type": "text", + "content": "Self-supervised features. We use AV-HuBERT to represent both the audio and visual content of a video. AV-HuBERT is a Transformer network trained in a self-supervised way to predict iteratively refined centroids from masked inputs. The features extracted by AV-HuBERT encode audio information as proved by its strong performance on tasks such as lip reading or noisy audio-visual speech recognition. We extract the audio and visual representations independently: to extract audio features we mask the visual input, to extract visual features we mask the audio input. For a video, we obtain audio features " + }, + { + "bbox": [ + 55, + 250, + 296, + 417 + ], + "type": "inline_equation", + "content": "\\mathbf{a}_i" + }, + { + "bbox": [ + 55, + 250, + 296, + 417 + ], + "type": "text", + "content": " and visual features " + }, + { + "bbox": [ + 55, + 250, + 296, + 417 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_i" + }, + { + "bbox": [ + 55, + 250, + 296, + 417 + ], + "type": "text", + "content": " features for each time step " + }, + { + "bbox": [ + 55, + 250, + 296, + 417 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 250, + 296, + 417 + ], + "type": "text", + "content": ". Both representations are 1024 dimensional and have a temporal resolution of 25 frames per second." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 423, + 296, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 423, + 296, + 495 + ], + "spans": [ + { + "bbox": [ + 55, + 423, + 296, + 495 + ], + "type": "text", + "content": "Alignment network. To tell how well the audio and visual features match each other, we first independently L2 normalize the feature vectors and then feed them into a network " + }, + { + "bbox": [ + 55, + 423, + 296, + 495 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 55, + 423, + 296, + 495 + ], + "type": "text", + "content": ". This is implemented as a multi-layer perceptron (MLP) over the concatenated normalized audio and visual features:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 128, + 505, + 295, + 520 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 128, + 505, + 295, + 520 + ], + "spans": [ + { + "bbox": [ + 128, + 505, + 295, + 520 + ], + "type": "interline_equation", + "content": "\\Phi_ {i j} = \\operatorname {M L P} \\left(\\left[ \\mathbf {a} _ {i}; \\mathbf {v} _ {j} \\right]\\right). \\tag {1}", + "image_path": "0b276796187a77ed68e8d7054a35d969a62efc4dc890e5236a371d82e110c928.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 529, + 296, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 529, + 296, + 589 + ], + "spans": [ + { + "bbox": [ + 55, + 529, + 296, + 589 + ], + "type": "text", + "content": "The MLP has four layers, which progressively reduce the feature dimensionality, with layers mapping from the AV-HuBERT feature size of 1024 to 512, 256, 128, and finally to a single output. Each hidden layer includes Layer Normalization and ReLU activations." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 595, + 296, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 595, + 296, + 632 + ], + "spans": [ + { + "bbox": [ + 55, + 595, + 296, + 632 + ], + "type": "text", + "content": "Loss function. To learn the alignment network " + }, + { + "bbox": [ + 55, + 595, + 296, + 632 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 55, + 595, + 296, + 632 + ], + "type": "text", + "content": " we maximize the probability of an audio frame " + }, + { + "bbox": [ + 55, + 595, + 296, + 632 + ], + "type": "inline_equation", + "content": "\\mathbf{a}_i" + }, + { + "bbox": [ + 55, + 595, + 296, + 632 + ], + "type": "text", + "content": " to match the corresponding video frame " + }, + { + "bbox": [ + 55, + 595, + 296, + 632 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_i" + }, + { + "bbox": [ + 55, + 595, + 296, + 632 + ], + "type": "text", + "content": "; this probability is defined as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 111, + 640, + 295, + 669 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 640, + 295, + 669 + ], + "spans": [ + { + "bbox": [ + 111, + 640, + 295, + 669 + ], + "type": "interline_equation", + "content": "p \\left(\\mathbf {v} _ {i} \\mid \\mathbf {a} _ {i}\\right) = \\frac {\\exp \\Phi_ {i i}}{\\sum_ {k \\in \\mathcal {N} (i)} \\exp \\Phi_ {i k}}, \\tag {2}", + "image_path": "0c964aec050c465cdab4c29a49cde88778a14c42d9c28f7e85e8013ab7c8b006.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(i)" + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": " represents the temporal neighborhood around the frame " + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": ". In our case " + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(i)" + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": " contains the 30 neighboring frames around " + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": ". We define the final loss as the negative" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 500, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 500, + 84 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 500, + 84 + ], + "type": "text", + "content": "probabilities averaged across the entire video:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 378, + 92, + 554, + 125 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 378, + 92, + 554, + 125 + ], + "spans": [ + { + "bbox": [ + 378, + 92, + 554, + 125 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = - \\frac {1}{T} \\sum_ {i = 1} ^ {T} \\log p (\\mathbf {v} _ {i} | \\mathbf {a} _ {j}). \\tag {3}", + "image_path": "410b51b659bd6d41935dc76547022628360ecb9ff1500feb59d514c6400192ce.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 132, + 554, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 132, + 554, + 157 + ], + "spans": [ + { + "bbox": [ + 313, + 132, + 554, + 157 + ], + "type": "text", + "content": "This loss is similar to the contrastive loss InfoNCE [54], which was also used for deepfake detection [17, 39]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 162, + 554, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 162, + 554, + 245 + ], + "spans": [ + { + "bbox": [ + 313, + 162, + 554, + 245 + ], + "type": "text", + "content": "Inference. Once " + }, + { + "bbox": [ + 313, + 162, + 554, + 245 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 313, + 162, + 554, + 245 + ], + "type": "text", + "content": " is learned, we can estimate the fakeness score as the negative of " + }, + { + "bbox": [ + 313, + 162, + 554, + 245 + ], + "type": "inline_equation", + "content": "\\Phi_{ii}" + }, + { + "bbox": [ + 313, + 162, + 554, + 245 + ], + "type": "text", + "content": " for each audio-video frame pair in a video; aligned audio-video frame pairs should yield lower scores. Then we compute an overall alignment score for the entire video by pooling the per-frame scores using the log-sum-exp function (a smooth version of the max function)." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 251, + 555, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 251, + 555, + 384 + ], + "spans": [ + { + "bbox": [ + 313, + 251, + 555, + 384 + ], + "type": "text", + "content": "Supervised variant. To understand the impact of the silence bias on the standard supervised learning paradigm, we design a supervised variant, AVH-Align/sup, that uses the same features and alignment network as AVH-Align, but a classification loss. In this setup we assume that apart from the real videos, we also have access to fake videos in the training set, with corresponding labels " + }, + { + "bbox": [ + 313, + 251, + 555, + 384 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 313, + 251, + 555, + 384 + ], + "type": "text", + "content": ". To obtain a per-video fakeness probability, we first pool the negated per-frame scores " + }, + { + "bbox": [ + 313, + 251, + 555, + 384 + ], + "type": "inline_equation", + "content": "\\Phi_{ii}" + }, + { + "bbox": [ + 313, + 251, + 555, + 384 + ], + "type": "text", + "content": " with the log-sum-exp function and then we apply the sigmoid function " + }, + { + "bbox": [ + 313, + 251, + 555, + 384 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 313, + 251, + 555, + 384 + ], + "type": "text", + "content": ". Finally, we optimize the binary cross-entropy (BCE) loss:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 350, + 392, + 555, + 406 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 392, + 555, + 406 + ], + "spans": [ + { + "bbox": [ + 350, + 392, + 555, + 406 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\sup } = \\operatorname {B C E} \\left(\\sigma \\left(\\log \\operatorname {s u m} \\exp (- \\Phi_ {i i})\\right), y\\right). \\tag {4}", + "image_path": "b3adebc42448b166966c6effd88bebc6612c4c4c8c4fb9d92e1a24b8643ec74a.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 420, + 429, + 434 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 420, + 429, + 434 + ], + "spans": [ + { + "bbox": [ + 313, + 420, + 429, + 434 + ], + "type": "text", + "content": "4.2. Experimental setup" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 445, + 555, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 445, + 555, + 673 + ], + "spans": [ + { + "bbox": [ + 313, + 445, + 555, + 673 + ], + "type": "text", + "content": "Datasets and metrics. We conducted our experiments on the two datasets introduced in Sec. 3.1: FakeAVCeleb and AV-Deepfake1M. For FakeAVCeleb, we use " + }, + { + "bbox": [ + 313, + 445, + 555, + 673 + ], + "type": "inline_equation", + "content": "70\\%" + }, + { + "bbox": [ + 313, + 445, + 555, + 673 + ], + "type": "text", + "content": " of the dataset for training and validation, and the rest of " + }, + { + "bbox": [ + 313, + 445, + 555, + 673 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 313, + 445, + 555, + 673 + ], + "type": "text", + "content": " for testing. For AV-Deepfake1M, we evaluate on 10k samples from the original validation set, but for the best performing models we also report results on the official withheld test set. To train AVH-Align, we use 50k real samples from the training set of AV-Deepfake1M (45k samples for training and 5k for validation). Since AV-Deepfake1M is based on VoxCeleb2, these samples originate from VoxCeleb2. To train the supervised variant, AVH-Align/sup, on AV-Deepfake1M, we select the same number of samples (45k for training and 5k for validation), but this time coming from both the real and fake classes. For both datasets, we evaluate video-level detection. Following prior work, we report results in terms of the area under the receiver operator characteristic curve (AUC) and average precision (AP), with the fake class being the positive class." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "text", + "content": "Implementation details. For AVH-Align we use a learning rate scheduler with a patience of 5 epochs and a factor of 0.1, with a starting learning rate of " + }, + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "inline_equation", + "content": "10^{-5}" + }, + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "text", + "content": ". The training is" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18819" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 129, + 70, + 358, + 184 + ], + "blocks": [ + { + "bbox": [ + 129, + 70, + 358, + 184 + ], + "lines": [ + { + "bbox": [ + 129, + 70, + 358, + 184 + ], + "spans": [ + { + "bbox": [ + 129, + 70, + 358, + 184 + ], + "type": "image", + "image_path": "fef535300c508cd85c42f608f28343135960f9d4e6659e6e4c7197083836fe04.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 194, + 555, + 251 + ], + "lines": [ + { + "bbox": [ + 55, + 194, + 555, + 251 + ], + "spans": [ + { + "bbox": [ + 55, + 194, + 555, + 251 + ], + "type": "text", + "content": "Figure 4. Overview of the AVH-Align method. A: We use the pretrained AV-HuBERT model to extract self-supervised features which we further align with a learnable network " + }, + { + "bbox": [ + 55, + 194, + 555, + 251 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 55, + 194, + 555, + 251 + ], + "type": "text", + "content": ". Note that we use a single AV-HuBERT model, but make two forward passes to obtain audio-only and video-only features (instead of a single set of multimodal features). B: At training we maximize the alignment score " + }, + { + "bbox": [ + 55, + 194, + 555, + 251 + ], + "type": "inline_equation", + "content": "\\Phi_{ii}" + }, + { + "bbox": [ + 55, + 194, + 555, + 251 + ], + "type": "text", + "content": ", between the audio features " + }, + { + "bbox": [ + 55, + 194, + 555, + 251 + ], + "type": "inline_equation", + "content": "\\mathbf{a}_i" + }, + { + "bbox": [ + 55, + 194, + 555, + 251 + ], + "type": "text", + "content": " at time step " + }, + { + "bbox": [ + 55, + 194, + 555, + 251 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 194, + 555, + 251 + ], + "type": "text", + "content": " and the corresponding video features " + }, + { + "bbox": [ + 55, + 194, + 555, + 251 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_i" + }, + { + "bbox": [ + 55, + 194, + 555, + 251 + ], + "type": "text", + "content": ", while minimizing the alignment " + }, + { + "bbox": [ + 55, + 194, + 555, + 251 + ], + "type": "inline_equation", + "content": "\\Phi_{ik}" + }, + { + "bbox": [ + 55, + 194, + 555, + 251 + ], + "type": "text", + "content": " to the other features " + }, + { + "bbox": [ + 55, + 194, + 555, + 251 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_k" + }, + { + "bbox": [ + 55, + 194, + 555, + 251 + ], + "type": "text", + "content": " in a neighboring window " + }, + { + "bbox": [ + 55, + 194, + 555, + 251 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(i)" + }, + { + "bbox": [ + 55, + 194, + 555, + 251 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 362, + 70, + 465, + 184 + ], + "blocks": [ + { + "bbox": [ + 362, + 70, + 465, + 184 + ], + "lines": [ + { + "bbox": [ + 362, + 70, + 465, + 184 + ], + "spans": [ + { + "bbox": [ + 362, + 70, + 465, + 184 + ], + "type": "image", + "image_path": "543f0c7bf486300dc621cb5661b715db57f07236c69304ea86d63f3f43615941.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 56, + 259, + 555, + 365 + ], + "blocks": [ + { + "bbox": [ + 56, + 259, + 555, + 365 + ], + "lines": [ + { + "bbox": [ + 56, + 259, + 555, + 365 + ], + "spans": [ + { + "bbox": [ + 56, + 259, + 555, + 365 + ], + "type": "table", + "html": "
MethodModalityTrain typeTrain dataMetric: AUCMetric: AP
FakeAVCelebAV-Deepfake1MFakeAVCelebAV-Deepfake1M
Trim: XTrim: ✓Trim: XTrim: ✓Trim: XTrim: ✓Trim: XTrim: ✓
AVH-Align/supAVsup.FAVC99.299.2 ≈69.063.6 ↓100.0100.0 ≈84.182.2 ↓
AVH-Align/supAVsup.AV1M77.570.8 ↓100.083.1 ↓99.499.1 ≈100.094.4 ↓
AVAD [17]AVunsup.LRS84.584.7 ≈54.354.3 ≈99.599.5 ≈76.376.3 ≈
SpeechForensics [30]AVunsup.VoxCeleb298.898.8 ≈68.868.2 ≈100.0100.0 ≈83.783.5 ≈
AVH-AlignAVunsup.VoxCeleb294.694.6 ≈85.983.5 ↓99.899.8 ≈94.393.5 ≈
", + "image_path": "b1a045c6c622e1664f95a6fe757766a84ba5afc349f8745f266e9bbc29e2440c.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 372, + 555, + 407 + ], + "lines": [ + { + "bbox": [ + 55, + 372, + 555, + 407 + ], + "spans": [ + { + "bbox": [ + 55, + 372, + 555, + 407 + ], + "type": "text", + "content": "Table 2. Comparison of AVH-Align and AVH-Align/sup on FakeAVCeleb (FAVC) and AV-Deepfake1M (AV1M) original and trimmed datasets. AVH-Align is not impacted by the presence of leading silence in the original datasets, showing similar performance on the trimmed versions. In contrast, the AUC performance of AVH-Align/sup degrades by " + }, + { + "bbox": [ + 55, + 372, + 555, + 407 + ], + "type": "inline_equation", + "content": "16.9\\%" + }, + { + "bbox": [ + 55, + 372, + 555, + 407 + ], + "type": "text", + "content": " when tested on the trimmed variant of AV1M." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 427, + 295, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 427, + 295, + 498 + ], + "spans": [ + { + "bbox": [ + 55, + 427, + 295, + 498 + ], + "type": "text", + "content": "stopped if there is no loss improvement on the validation set for 10 consecutive epochs. AVH-Align/sup is learned using an Adam optimizer with a learning rate of " + }, + { + "bbox": [ + 55, + 427, + 295, + 498 + ], + "type": "inline_equation", + "content": "10^{-3}" + }, + { + "bbox": [ + 55, + 427, + 295, + 498 + ], + "type": "text", + "content": ". For the AVHuBERT feature extractor we use the checkpoint pretrained on LRS3 and VoxCeleb2 and finetuned on LRS3 for visual speech recognition." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 508, + 176, + 521 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 508, + 176, + 521 + ], + "spans": [ + { + "bbox": [ + 55, + 508, + 176, + 521 + ], + "type": "text", + "content": "4.3. Experimental results" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 527, + 295, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 527, + 295, + 634 + ], + "spans": [ + { + "bbox": [ + 55, + 527, + 295, + 634 + ], + "type": "text", + "content": "We train our method, AVH-Align, as well as its supervised variant, AVH-Align/sup, on the two datasets, FakeAVCeleb and AV-Deepfake1M, and evaluate on either of them (that is, both in-domain and out-of-domain). For each evaluation, in addition to the original set, we also consider the trimmed version of the validation dataset (trim: " + }, + { + "bbox": [ + 55, + 527, + 295, + 634 + ], + "type": "inline_equation", + "content": "\\checkmark" + }, + { + "bbox": [ + 55, + 527, + 295, + 634 + ], + "type": "text", + "content": "). The trimming is performed as described in Sec. 3.3. We always use the untrimmed, original data for training. Results are shown in Table 2." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": "Impact of leading silence bias. We observe that AVH-Align is robust to the leading silence bias: trimming the silence has no effect on FakeAVCeleb and only a slight effect on AV-Deepfake1M (2.4% AUC). On the other hand, the performance of the supervised variant, AVHAlign/sup, decreases considerably when testing on AV" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 426, + 555, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 555, + 521 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 555, + 521 + ], + "type": "text", + "content": "Deepfake1M (by " + }, + { + "bbox": [ + 313, + 426, + 555, + 521 + ], + "type": "inline_equation", + "content": "16.9\\%" + }, + { + "bbox": [ + 313, + 426, + 555, + 521 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 313, + 426, + 555, + 521 + ], + "type": "inline_equation", + "content": "5.4\\%" + }, + { + "bbox": [ + 313, + 426, + 555, + 521 + ], + "type": "text", + "content": " AUC, depending on the train set). On FakeAVCeleb removing the leading silence has no impact for AVH-Align/sup, when training also on FakeAVCeleb. This may happen because FakeAVCeleb has full manipulations and there is other useful information in the video. Instead, when training on the locally-manipulated AV-Deepfake1M, the difference becomes significant " + }, + { + "bbox": [ + 313, + 426, + 555, + 521 + ], + "type": "inline_equation", + "content": "(6.7\\%)" + }, + { + "bbox": [ + 313, + 426, + 555, + 521 + ], + "type": "text", + "content": " AUC." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 528, + 556, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 528, + 556, + 624 + ], + "spans": [ + { + "bbox": [ + 313, + 528, + 556, + 624 + ], + "type": "text", + "content": "Comparison with other unsupervised methods. We compare our results with those of AVAD [17] and SpeechForensics [30], which are also unsupervised methods, trained on real data only. Similar to AVH-Align they are not impacted by the spurious leading silence, showing nearly identical results for trimmed and untrimmed datasets. However, their overall performance is considerably worse than ours on the AV-Deepfake1M dataset." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 629, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 556, + 714 + ], + "type": "text", + "content": "Score visualization. Figure 5 shows the per frame scores obtained by the two methods, AVH-Align and AVH-Align/sup, together with the groundtruth manipulated interval. For the AVH-Align method, the scores represent the misalignment probability between the audio and visual streams at each time frame; for the AVH-Align/sup method, the scores represent the probability of an audio-video frame" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18820" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 71, + 212, + 129 + ], + "blocks": [ + { + "bbox": [ + 62, + 71, + 212, + 129 + ], + "lines": [ + { + "bbox": [ + 62, + 71, + 212, + 129 + ], + "spans": [ + { + "bbox": [ + 62, + 71, + 212, + 129 + ], + "type": "image", + "image_path": "3a5e7050545c78dd216795fba5e17c516d365fcdc4fd74a3bf82eac464359703.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 223, + 70, + 373, + 128 + ], + "blocks": [ + { + "bbox": [ + 223, + 70, + 373, + 128 + ], + "lines": [ + { + "bbox": [ + 223, + 70, + 373, + 128 + ], + "spans": [ + { + "bbox": [ + 223, + 70, + 373, + 128 + ], + "type": "image", + "image_path": "9fc5dc4342f77a1f76df917b818f00ee06880fe6068c5daef0ace80fae4a4c84.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 383, + 70, + 534, + 128 + ], + "blocks": [ + { + "bbox": [ + 383, + 70, + 534, + 128 + ], + "lines": [ + { + "bbox": [ + 383, + 70, + 534, + 128 + ], + "spans": [ + { + "bbox": [ + 383, + 70, + 534, + 128 + ], + "type": "image", + "image_path": "3eb58c549dfade0519d30bdb06fc7a18cd3fabf2bf600c74cd88a91167ecaeec.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 62, + 129, + 212, + 186 + ], + "blocks": [ + { + "bbox": [ + 62, + 129, + 212, + 186 + ], + "lines": [ + { + "bbox": [ + 62, + 129, + 212, + 186 + ], + "spans": [ + { + "bbox": [ + 62, + 129, + 212, + 186 + ], + "type": "image", + "image_path": "d684a2c335c1d0c7b8ca09f075030673d4aa06ea8848ed5f7118661714106f12.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 223, + 130, + 373, + 186 + ], + "blocks": [ + { + "bbox": [ + 223, + 130, + 373, + 186 + ], + "lines": [ + { + "bbox": [ + 223, + 130, + 373, + 186 + ], + "spans": [ + { + "bbox": [ + 223, + 130, + 373, + 186 + ], + "type": "image", + "image_path": "355e011d05424346819a2982958bbec153a7fba9682f6dc4343b90e39e78cc91.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 383, + 130, + 534, + 186 + ], + "blocks": [ + { + "bbox": [ + 383, + 130, + 534, + 186 + ], + "lines": [ + { + "bbox": [ + 383, + 130, + 534, + 186 + ], + "spans": [ + { + "bbox": [ + 383, + 130, + 534, + 186 + ], + "type": "image", + "image_path": "d95be9185d9f5913eec140e6ae8a6ffd894d101050f2b8df55693bbc180e41bd.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 62, + 189, + 213, + 247 + ], + "blocks": [ + { + "bbox": [ + 62, + 189, + 213, + 247 + ], + "lines": [ + { + "bbox": [ + 62, + 189, + 213, + 247 + ], + "spans": [ + { + "bbox": [ + 62, + 189, + 213, + 247 + ], + "type": "image", + "image_path": "d279d263aa7afd4d98915b491760dec98e3721e50dfb8903fee1002589843c35.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 258, + 555, + 304 + ], + "lines": [ + { + "bbox": [ + 54, + 258, + 555, + 304 + ], + "spans": [ + { + "bbox": [ + 54, + 258, + 555, + 304 + ], + "type": "text", + "content": "Figure 5. Per frame fakeness probabilities for AVH-Align and AVH-Align/sup on AV-Deepfake1M. AVH-Align/sup always marks the first frame—corresponding usually to the leading silence—as fake, thus confirming that it uses the bias to distinguish between real and fake videos. AVH-Align is not affected by the presence of the leading silence. The fakeness probabilities for AVH-Align can be interpreted as misalignment probabilities, which is why they are higher during or after the manipulated region." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 223, + 189, + 373, + 246 + ], + "blocks": [ + { + "bbox": [ + 223, + 189, + 373, + 246 + ], + "lines": [ + { + "bbox": [ + 223, + 189, + 373, + 246 + ], + "spans": [ + { + "bbox": [ + 223, + 189, + 373, + 246 + ], + "type": "image", + "image_path": "16d9b9bb8ba13e5ea84c2ddcee3d06d930ebcf47abc8260fcedf9544826f4db5.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 383, + 189, + 534, + 246 + ], + "blocks": [ + { + "bbox": [ + 383, + 189, + 534, + 246 + ], + "lines": [ + { + "bbox": [ + 383, + 189, + 534, + 246 + ], + "spans": [ + { + "bbox": [ + 383, + 189, + 534, + 246 + ], + "type": "image", + "image_path": "cdfe86de160fa666ba4a3447feb789efc16039559628f24d475d68ff11186488.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 54, + 323, + 297, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 323, + 297, + 492 + ], + "spans": [ + { + "bbox": [ + 54, + 323, + 297, + 492 + ], + "type": "text", + "content": "of being fake. We see that AVH-Align/sup always predicts the first frame as being fake, confirming once again that it has learned to associate the spurious leading silence with a video being fake. When detected, the actual fake region is given a considerable lower fakeness probability than that assigned to the leading silence. On the other hand, AVH-Align is not affected by the presence of leading silence. In this case the manipulated areas have higher misalignment scores, a frame-level evaluation yielding " + }, + { + "bbox": [ + 54, + 323, + 297, + 492 + ], + "type": "inline_equation", + "content": "77.7\\%" + }, + { + "bbox": [ + 54, + 323, + 297, + 492 + ], + "type": "text", + "content": " AUC. We do notice that there are other regions in the video that get a high misalignment score. This happens because the separation and re-composition of the audio and video streams can introduce desynchronizations between the two even in areas that were not intentionally manipulated." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 502, + 296, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 502, + 296, + 514 + ], + "spans": [ + { + "bbox": [ + 55, + 502, + 296, + 514 + ], + "type": "text", + "content": "4.4. Comparison on the official AV-Deepfake1M test" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 521, + 296, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 296, + 629 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 296, + 629 + ], + "type": "text", + "content": "To further verify our conclusions, we evaluate our approach on the official test set of the AV-Deepfake1M dataset for which the labels are withheld. The main difference between the validation set, used in the previous section, and the test set are the subject identities: the test subjects are unseen at train time. We obtain the results by submitting our prediction to the official competition server. Our results are compared against those of other methods based on the values reported in the AV-Deepfake1M paper [6]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 630, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 296, + 715 + ], + "type": "text", + "content": "The results shown in Table 3 are in line to those obtained on the validation split (Table 2). Compared to the rest of the methods, AVH-Align has the highest performance (85.24% AUC) even if has seen only real data at training. The silence classifier described in Sec. 3 obtains again a performance of over 98% AUC, indicating that the test set suffers from the same spurious feature as the training and validation splits." + } + ] + } + ], + "index": 13 + }, + { + "type": "table", + "bbox": [ + 315, + 321, + 556, + 473 + ], + "blocks": [ + { + "bbox": [ + 315, + 321, + 556, + 473 + ], + "lines": [ + { + "bbox": [ + 315, + 321, + 556, + 473 + ], + "spans": [ + { + "bbox": [ + 315, + 321, + 556, + 473 + ], + "type": "table", + "html": "
MethodsModalityAUC
Segment-level methods
Meso4 [1]V54.53
MesoInception4 [1]V57.16
MDS [10]AV56.57
MARLIN [5]V58.03
Frame-level methods
Meso4 [1]V63.05
MesoInception4 [1]V64.04
Xception [9]V68.68
EfficientViT [12]V65.51
Ours
AVH-AlignAV85.24
AVH-Align/supAV99.90
Silence classifierA98.44
", + "image_path": "3d115afac7bbfe0f7741c3eabbd35c0486407845981daa2ff5871a1113a46b2d.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 481, + 555, + 537 + ], + "lines": [ + { + "bbox": [ + 313, + 481, + 555, + 537 + ], + "spans": [ + { + "bbox": [ + 313, + 481, + 555, + 537 + ], + "type": "text", + "content": "Table 3. Results on the official AV-Deepfake1M test set. Except the methods that have access to the leading silence bias (shown in red), AVH-Align performs best with " + }, + { + "bbox": [ + 313, + 481, + 555, + 537 + ], + "type": "inline_equation", + "content": "85.24\\%" + }, + { + "bbox": [ + 313, + 481, + 555, + 537 + ], + "type": "text", + "content": " AUC. Having access to the leading silence bias, the silence classifier and AVH-Align/sup show almost perfect performance of over " + }, + { + "bbox": [ + 313, + 481, + 555, + 537 + ], + "type": "inline_equation", + "content": "98\\%" + }, + { + "bbox": [ + 313, + 481, + 555, + 537 + ], + "type": "text", + "content": " AUC." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 559, + 555, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 559, + 555, + 632 + ], + "spans": [ + { + "bbox": [ + 313, + 559, + 555, + 632 + ], + "type": "text", + "content": "Other methods that are trained at video level do not have access to the silence bias, and hence show only modest performance. Instead, we expect audio-based methods to use the bias and return over optimistic results. This is indeed the case for the supervised variant, AVH-Align/sup, which returns an AUC of " + }, + { + "bbox": [ + 313, + 559, + 555, + 632 + ], + "type": "inline_equation", + "content": "99.90\\%" + }, + { + "bbox": [ + 313, + 559, + 555, + 632 + ], + "type": "text", + "content": " on the official test set." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 640, + 415, + 654 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 640, + 415, + 654 + ], + "spans": [ + { + "bbox": [ + 313, + 640, + 415, + 654 + ], + "type": "text", + "content": "4.5. Further analysis" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "content": "Ablation of AVH-Align components. Here we analyze the impact of various design choices on the performance of the AVH-Align method. Specifically, we investigate the following: 1) removing feature normalization; 2) reducing the" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18821" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 70, + 295, + 149 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 295, + 149 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 295, + 149 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 295, + 149 + ], + "type": "table", + "html": "
ConfigurationFAVCAV1MAVLips
AVH-Align94.685.986.3
1... Feature norm: ×92.484.886.1
2... Train size: 8.7k89.182.583.0
3... Score pooling: mean97.476.888.8
4... AV-HuBERT: LRS384.860.380.1
5... Alignment net: Linear31.551.943.4
", + "image_path": "3ba86d5b55e8be0b3a03c57ca69a06f5ef149460ad7bc55b841334d9b97bc606.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 211, + 295, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 211, + 295, + 344 + ], + "spans": [ + { + "bbox": [ + 54, + 211, + 295, + 344 + ], + "type": "text", + "content": "training set size from 45,000 to 8,782 samples (the real ones from the AVH-Align/sup train set); 3) using mean score pooling instead of log-sum-exp; 4) extracting features with the AV-HuBERT checkpoint pretrained on LRS3 only instead of LRS3 and VoxCeleb2; 5) using a linear layer instead of MLP. Results are shown in Table 4. All ablations lead to a decrease in performance with the exception of mean score pooling for FakeAVCeleb and AVLips. This is expected, however, since fake videos in FakeAVCeleb and AVLips are manipulated at every frame, while AVDeepfake1M has only locally-manipulated fake videos." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 348, + 295, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 348, + 295, + 469 + ], + "spans": [ + { + "bbox": [ + 54, + 348, + 295, + 469 + ], + "type": "text", + "content": "Impact of VoxCeleb2 pretraining. Both FakeAVCeleb and AV1M datasets use VoxCeleb2 as their source dataset. Can there be a data leakage from AV-HuBERT VoxCeleb2 features to these datasets? We evaluate on a dataset that is not using VoxCeleb2, AVLips [31]. We observe that VoxCeleb2 features perform better even on this dataset, suggesting that these features are simply stronger than LRS3 features. This is may be explained by the size of VoxCeleb data (about four times larger than LRS3) and is in line with the results on other downstream tasks (e.g., lip reading [49])." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 474, + 295, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 474, + 295, + 629 + ], + "spans": [ + { + "bbox": [ + 55, + 474, + 295, + 629 + ], + "type": "text", + "content": "Analysis of AVH-Align/sup architecture. We investigate how a simpler alignment network impacts the supervised model. To this end, we train a linear layer on top of the AV-HuBERT representations instead of the MLP network. In Table 5, we observe that unlike the unsupervised case where a linear layer was too weak to learn, here a linear layer performs as well or even better than the MLP. The largest difference between the MLP and then linear layer is observed when training on FakeAVCeleb and testing on AV-Deepfake1M. With a linear layer, the performance drop from untrimmed to trimmed AV-Deepafake1M is " + }, + { + "bbox": [ + 55, + 474, + 295, + 629 + ], + "type": "inline_equation", + "content": "19.3\\%" + }, + { + "bbox": [ + 55, + 474, + 295, + 629 + ], + "type": "text", + "content": " (from " + }, + { + "bbox": [ + 55, + 474, + 295, + 629 + ], + "type": "inline_equation", + "content": "85.9\\%" + }, + { + "bbox": [ + 55, + 474, + 295, + 629 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 474, + 295, + 629 + ], + "type": "inline_equation", + "content": "66.6\\%" + }, + { + "bbox": [ + 55, + 474, + 295, + 629 + ], + "type": "text", + "content": " AUC), while with an MLP, the drop is " + }, + { + "bbox": [ + 55, + 474, + 295, + 629 + ], + "type": "inline_equation", + "content": "5.4\\%" + }, + { + "bbox": [ + 55, + 474, + 295, + 629 + ], + "type": "text", + "content": " (from " + }, + { + "bbox": [ + 55, + 474, + 295, + 629 + ], + "type": "inline_equation", + "content": "69.0\\%" + }, + { + "bbox": [ + 55, + 474, + 295, + 629 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 474, + 295, + 629 + ], + "type": "inline_equation", + "content": "63.6\\%" + }, + { + "bbox": [ + 55, + 474, + 295, + 629 + ], + "type": "text", + "content": " AUC)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 639, + 125, + 651 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 639, + 125, + 651 + ], + "spans": [ + { + "bbox": [ + 55, + 639, + 125, + 651 + ], + "type": "text", + "content": "5. Discussion" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 660, + 295, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 660, + 295, + 684 + ], + "spans": [ + { + "bbox": [ + 55, + 660, + 295, + 684 + ], + "type": "text", + "content": "We discuss how our conclusions fit into the broader scope of deepfake detection." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": "A different evaluation paradigm. We observed near-perfect performance on datasets such as FakeAVCeleb and" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 317, + 70, + 553, + 145 + ], + "blocks": [ + { + "bbox": [ + 55, + 158, + 295, + 192 + ], + "lines": [ + { + "bbox": [ + 55, + 158, + 295, + 192 + ], + "spans": [ + { + "bbox": [ + 55, + 158, + 295, + 192 + ], + "type": "text", + "content": "Table 4. Ablation of the main components of AVH-Align in terms of AUC. Mean score pooling helps when videos are fully generated (FAVC, AVLips). Other ablations degrade the performance." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 70, + 553, + 145 + ], + "lines": [ + { + "bbox": [ + 317, + 70, + 553, + 145 + ], + "spans": [ + { + "bbox": [ + 317, + 70, + 553, + 145 + ], + "type": "table", + "html": "
ArchitectureTrain dataFakeAVCelebAV-Deepfake1M
Trim: XTrim: √Trim: XTrim: √
LinearFakeAVCeleb99.699.585.966.6
MLPFakeAVCeleb99.299.269.063.6
LinearAV-Deepfake1M76.269.799.983.2
MLPAV-Deepfake1M77.570.8100.083.1
", + "image_path": "6cffe6b10e144d030fbba8cd5bef62604f942caaeb58878bc015733d39d40837.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 154, + 553, + 176 + ], + "lines": [ + { + "bbox": [ + 313, + 154, + 553, + 176 + ], + "spans": [ + { + "bbox": [ + 313, + 154, + 553, + 176 + ], + "type": "text", + "content": "Table 5. Architecture analysis of AVH-Align/sup. In most cases, the linear and MLP architectures perform similarly." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 201, + 555, + 345 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 201, + 555, + 345 + ], + "spans": [ + { + "bbox": [ + 313, + 201, + 555, + 345 + ], + "type": "text", + "content": "AV-Deepfake1M. The reasons are that the same generative models were used in both train and test, as well as the unintentional leakage of spurious features. The latter also affects cross-dataset performance, e.g., training on FakeAVCeleb and testing on AV-Deepfake1M still yields strong performance. In a realistic scenario, however, we are not given access to the generators, nor to the pre-processing or post-processing steps, which introduce unintentional alterations. Each deepfake released in the wild may be created through a different set of tools, unknown apriori. As such, we believe that another way of gauging the progress of deepfake detection is by refraining the training to real data only." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 352, + 556, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 352, + 556, + 532 + ], + "spans": [ + { + "bbox": [ + 313, + 352, + 556, + 532 + ], + "type": "text", + "content": "Circumventing shortcuts through alignment. Modeling real data is a way to avoid shortcuts in the data. A different direction is taken in the very recent work of Rajan et al. [43]. The idea is to generate fake samples by reconstructing real samples through a generator of choice. This approach ensures that fakes and reals are aligned, avoiding spurious features. On the other hand, this forces the model to focus on the fingerprint of the generator, hindering generalizability. Moreover, by equating fake to a fingerprint, we become susceptible to laundering [32], lose the ability to localise [33, 51] and distinguish benign fakes (e.g., superresolution images [57]). We believe that both directions (alignment and real data) are complementary perspectives that should be considered to tackle the multi-faceted problem of deepfake detection." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 548, + 391, + 560 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 548, + 391, + 560 + ], + "spans": [ + { + "bbox": [ + 313, + 548, + 391, + 560 + ], + "type": "text", + "content": "6. Conclusions" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "text", + "content": "In this paper we exposed a previously unknown bias in two widely adopted audio-video deepfake detection datasets—a leading silence in fake videos. We showed that models exposed to this bias during training are prone to rely on it when deciding the authenticity of a video, thus displaying overly optimistic results. As an alternative, we propose to shift the learning paradigm towards unsupervised learning on real data only. Specifically, we find that self-supervised audio-video representations coupled with an alignment network trained on real videos produce more robust and consistent results. Our work raises awareness regarding dataset design and evaluation of deepfake detection." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18822" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 92, + 296, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 92, + 296, + 140 + ], + "spans": [ + { + "bbox": [ + 55, + 92, + 296, + 140 + ], + "type": "text", + "content": "This work was supported in part by the EU Horizon projects AI4TRUST (No. 101070190) and ELIAS (No. 101120237), and by CNCS-UEFISCDI (PN-IV-P7-7.1-PTE-2024-0600)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 154, + 115, + 167 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 154, + 115, + 167 + ], + "spans": [ + { + "bbox": [ + 56, + 154, + 115, + 167 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 174, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 61, + 174, + 296, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 174, + 296, + 207 + ], + "spans": [ + { + "bbox": [ + 61, + 174, + 296, + 207 + ], + "type": "text", + "content": "[1] Darius Afchar, Vincent Nozick, Junichi Yamagishi, and Isao Echizen. MesoNet: A compact facial video forgery detection network. In WIFS, page 1-7, 2018. 2, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 209, + 296, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 209, + 296, + 241 + ], + "spans": [ + { + "bbox": [ + 61, + 209, + 296, + 241 + ], + "type": "text", + "content": "[2] Matyas Bohacek and Hany Farid. Lost in translation: Lipsync deepfake detection from audio-video mismatch. In CVPR, pages 4315-4323, 2024. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 243, + 295, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 243, + 295, + 277 + ], + "spans": [ + { + "bbox": [ + 62, + 243, + 295, + 277 + ], + "type": "text", + "content": "[3] Stefano Borzi, Oliver Giudice, Filippo Stanco, and Dario Allegra. Is synthetic voice detection research going into the right direction? In CVPRW, pages 71-80, 2022. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 278, + 295, + 332 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 278, + 295, + 332 + ], + "spans": [ + { + "bbox": [ + 62, + 278, + 295, + 332 + ], + "type": "text", + "content": "[4] Zhixi Cai, Shreya Ghosh, Abhinav Dhall, Tom Gedeon, Kalin Stefanov, and Munawar Hayat. Glitch in the matrix: A large scale benchmark for content driven audio-visual forgery detection and localization. Comput. Vis. Image Underst., 236:103818, 2023. 1, 2, 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 335, + 295, + 378 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 335, + 295, + 378 + ], + "spans": [ + { + "bbox": [ + 62, + 335, + 295, + 378 + ], + "type": "text", + "content": "[5] Zhixi Cai, Shreya Ghosh, Kalin Stefanov, Abhinav Dhall, Jianfei Cai, Hamid Rezatofighi, Reza Haffari, and Munawar Hayat. MARLIN: Masked autoencoder for facial video representation learning, 2023. 2, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 381, + 295, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 381, + 295, + 425 + ], + "spans": [ + { + "bbox": [ + 62, + 381, + 295, + 425 + ], + "type": "text", + "content": "[6] Zhixi Cai, Shreya Ghosh, Aman Pankaj Adatia, Munawar Hayat, Abhinav Dhall, Tom Gedeon, and Kalin Stefanov. AV-Deepfake1M: A large-scale LLM-driven audio-visual deepfake dataset, 2024. 1, 2, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 427, + 295, + 481 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 427, + 295, + 481 + ], + "spans": [ + { + "bbox": [ + 62, + 427, + 295, + 481 + ], + "type": "text", + "content": "[7] Edresson Casanova, Julian Weber, Christopher Dane Shulby, Arnaldo Candido Junior, Eren Gölgge, and Moacir A. Ponti. YourTTS: Towards zero-shot multi-speaker TTS and zero-shot voice conversion for everyone. In ICML, pages 2709-2720, 2022. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 484, + 295, + 517 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 484, + 295, + 517 + ], + "spans": [ + { + "bbox": [ + 62, + 484, + 295, + 517 + ], + "type": "text", + "content": "[8] Lucy Chai, David Bau, Ser-Nam Lim, and Phillip Isola. What makes fake images detectable? understanding properties that generalize. In ECCV, pages 103-120, 2020. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 519, + 295, + 550 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 519, + 295, + 550 + ], + "spans": [ + { + "bbox": [ + 62, + 519, + 295, + 550 + ], + "type": "text", + "content": "[9] François Chollet. Xception: Deep learning with depthwise separable convolutions. In CVPR, pages 1800-1807, 2017. 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 553, + 295, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 553, + 295, + 597 + ], + "spans": [ + { + "bbox": [ + 57, + 553, + 295, + 597 + ], + "type": "text", + "content": "[10] Komal Chugh, Parul Gupta, Abhinav Dhall, and Ramanathan Subramanian. Not made for each other-audiovisual dissonance-based deepfake detection and localization. In ACMM, pages 439-447, 2020. 4, 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 600, + 294, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 600, + 294, + 621 + ], + "spans": [ + { + "bbox": [ + 57, + 600, + 294, + 621 + ], + "type": "text", + "content": "[11] J. S. Chung, A. Nagrani, and A. Zisserman. VoxCeleb2: Deep speaker recognition. In Interspeech, 2018. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 624, + 295, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 624, + 295, + 667 + ], + "spans": [ + { + "bbox": [ + 57, + 624, + 295, + 667 + ], + "type": "text", + "content": "[12] Davide Alessandro Coccomini, Nicola Messina, Claudio Gennaro, and Fabrizio Falchi. Combining EfficientNet and vision transformers for video deepfake detection. In ICIAP, pages 219-229, 2022. 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 670, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 295, + 712 + ], + "type": "text", + "content": "[13] Davide Cozzolino, Giovanni Poggi, Riccardo Corvi, Matthias Nießner, and Luisa Verdoliva. Raising the bar of AI-generated image detection with CLIP. In CVPR, pages 4356-4366, 2024. 4" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "text", + "content": "[14] Davide Cozzolino, Giovanni Poggi, Matthias Nießner, and Luisa Verdoliva. Zero-shot detection of AI-generated images. In ECCV, pages 54-72, 2024. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 108, + 553, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 108, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 317, + 108, + 553, + 140 + ], + "type": "text", + "content": "[15] Alexandre Défossez, Gabriel Synnaeve, and Yossi Adi. Real time speech enhancement in the waveform domain. In *Interspeech*, pages 3291–3295, 2020. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 143, + 553, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 143, + 553, + 186 + ], + "spans": [ + { + "bbox": [ + 317, + 143, + 553, + 186 + ], + "type": "text", + "content": "[16] Brian Dolhansky, Joanna Bitton, Ben Pflaum, Jikuo Lu, Russ Howes, Menglin Wang, and Cristian Canton Ferrer. The deepfake detection challenge (dfdc) dataset. arXiv preprint arXiv:2006.07397, 2020. 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 189, + 553, + 221 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 189, + 553, + 221 + ], + "spans": [ + { + "bbox": [ + 317, + 189, + 553, + 221 + ], + "type": "text", + "content": "[17] Chao Feng, Ziyang Chen, and Andrew Owens. Self-supervised video forensics by audio-visual anomaly detection. In CVPR, pages 10491–10503, 2023. 2, 4, 5, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 223, + 553, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 223, + 553, + 266 + ], + "spans": [ + { + "bbox": [ + 316, + 223, + 553, + 266 + ], + "type": "text", + "content": "[18] Alexandros Haliassos, Konstantinos Vougioukas, Stavros Petridis, and Maja Pantic. Lips don't lie: A generalisable and robust approach to face forgery detection. In CVPR, pages 5039-5049, 2021. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 269, + 553, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 269, + 553, + 312 + ], + "spans": [ + { + "bbox": [ + 316, + 269, + 553, + 312 + ], + "type": "text", + "content": "[19] Alexandros Haliassos, Rodrigo Mira, Stavros Petridis, and Maja Pantic. Leveraging real talking faces via self-supervision for robust forgery detection. In CVPR, pages 14950-14962, 2022. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 315, + 553, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 315, + 553, + 369 + ], + "spans": [ + { + "bbox": [ + 316, + 315, + 553, + 369 + ], + "type": "text", + "content": "[20] Ammarah Hashmi, Sahibzada Adil Shahzad, Chia-Wen Lin, Yu Tsao, and Hsin-Min Wang. AVTENet: Audio-visual transformer-based ensemble network exploiting multiple experts for video deepfake detection. CoRR, abs/2310.13103, 2023. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 371, + 553, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 371, + 553, + 414 + ], + "spans": [ + { + "bbox": [ + 316, + 371, + 553, + 414 + ], + "type": "text", + "content": "[21] Baojin Huang, Zhongyuan Wang, Jifan Yang, Jiaxin Ai, Qin Zou, Qian Wang, and Dengpan Ye. Implicit identity driven deepfake face swapping detection. In CVPR, pages 4490-4499, 2023. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 417, + 553, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 417, + 553, + 472 + ], + "spans": [ + { + "bbox": [ + 316, + 417, + 553, + 472 + ], + "type": "text", + "content": "[22] Ye Jia, Yu Zhang, Ron J. Weiss, Quan Wang, Jonathan Shen, Fei Ren, Zhifeng Chen, Patrick Nguyen, Ruoming Pang, Ignacio López-Moreno, and Yonghui Wu. Transfer learning from speaker verification to multispeaker text-to-speech synthesis. In NeurIPS, pages 4485-4495, 2018. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 474, + 553, + 516 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 474, + 553, + 516 + ], + "spans": [ + { + "bbox": [ + 316, + 474, + 553, + 516 + ], + "type": "text", + "content": "[23] Hasam Khalid, Shahroz Tariq, Minha Kim, and Simon S. Woo. FakeAVCeleb: A novel audio-video multimodal deepfake dataset. In NeurIPS Datasets and Benchmarks, 2021. 1, 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 519, + 553, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 519, + 553, + 552 + ], + "spans": [ + { + "bbox": [ + 316, + 519, + 553, + 552 + ], + "type": "text", + "content": "[24] Jaehyeon Kim, Jungil Kong, and Juhee Son. Conditional variational autoencoder with adversarial learning for end-to-end text-to-speech. In ICML, pages 5530-5540, 2021. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 555, + 553, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 555, + 553, + 586 + ], + "spans": [ + { + "bbox": [ + 316, + 555, + 553, + 586 + ], + "type": "text", + "content": "[25] Pavel Korshunov and Sébastien Marcel. DeepFakes: A new threat to face recognition? Assessment and detection. CoRR, abs/1812.08685, 2018. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 589, + 553, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 589, + 553, + 621 + ], + "spans": [ + { + "bbox": [ + 316, + 589, + 553, + 621 + ], + "type": "text", + "content": "[26] Iryna Korshunova, Wenzhe Shi, Joni Dambre, and Lucas Theis. Fast face-swap using convolutional neural networks. In ICCV, pages 3697-3705, 2017. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 624, + 553, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 553, + 666 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 553, + 666 + ], + "type": "text", + "content": "[27] Christos Koutlis and Symeon Papadopoulos. DiMoDif: Discourse modality-information differentiation for audio-visual deepfake detection and localization. CoRR, abs/2411.10193, 2024. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "text", + "content": "[28] Patrick Kwon, Jaeseong You, Gyuhyeon Nam, Sungwoo Park, and Gyeongsu Chae. KoDF: A large-scale Korean deepfake detection dataset. In ICCV, pages 10744-10753, 2021. 2" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18823" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "type": "text", + "content": "[29] Xiaolou Li, Zehua Liu, Chen Chen, Lantian Li, Li Guo, and Dong Wang. Zero-shot fake video detection by audio-visual consistency. CoRR, abs/2406.07854, 2024. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 294, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 294, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 294, + 150 + ], + "type": "text", + "content": "[30] Yachao Liang, Min Yu, Gang Li, Jianguo Jiang, Boquan Li, Feng Yu, Ning Zhang, Xiang Meng, and Weiqing Huang. SpeechForensics: Audio-visual speech representation learning for face forgery detection. In NeurIPS, 2024. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 294, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 294, + 206 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 294, + 206 + ], + "type": "text", + "content": "[31] Weifeng Liu, Tianyi She, Jiawei Liu, Boheng Li, Dongyu Yao, Ziyou Liang, and Run Wang. Lips are lying: Spotting the temporal inconsistency between audio and visual in lip-syncing deepfakes. In NeurIPS, pages 91131-91155, 2024. 4, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 208, + 294, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 208, + 294, + 240 + ], + "spans": [ + { + "bbox": [ + 56, + 208, + 294, + 240 + ], + "type": "text", + "content": "[32] Sara Mandelli, Paolo Bestagini, and Stefano Tubaro. When synthetic traces hide real content: Analysis of stable diffusion image laundering. CoRR, abs/2407.10736, 2024. 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 243, + 294, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 243, + 294, + 285 + ], + "spans": [ + { + "bbox": [ + 56, + 243, + 294, + 285 + ], + "type": "text", + "content": "[33] Hannes Mareen, Dimitrios Karageorgiou, Glenn Van Wallendael, Peter Lambert, and Symeon Papadopoulos. TGIF: Text-guided inpainting forgery dataset. CoRR, abs/2407.11566, 2024. 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 286, + 294, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 286, + 294, + 319 + ], + "spans": [ + { + "bbox": [ + 56, + 286, + 294, + 319 + ], + "type": "text", + "content": "[34] Francesco Marra, Diego Gragnaniello, Luisa Verdoliva, and Giovanni Poggi. Do GANs leave artificial fingerprints? In MIPR, pages 506-511, 2019. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 321, + 294, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 321, + 294, + 364 + ], + "spans": [ + { + "bbox": [ + 56, + 321, + 294, + 364 + ], + "type": "text", + "content": "[35] Nicolas M. Müller, Franziska Dieckmann, Pavel Czempin, Roman Canals, and Konstantin Bötttinger. Speech is silver, silence is golden: What do ASVspoof-trained models really learn? CoRR, abs/2106.12914, 2021. 1, 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 365, + 294, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 365, + 294, + 420 + ], + "spans": [ + { + "bbox": [ + 56, + 365, + 294, + 420 + ], + "type": "text", + "content": "[36] Nicolas M. Müller, Piotr Kawa, Wei Herng Choong, Edresson Casanova, Eren Gölgge, Thorsten Müller, Piotr Syga, Philip Sperl, and Konstantin Bötttinger. MLAAD: The multilanguage audio anti-spoofing dataset. In IJCNN, pages 1-7, 2024. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 422, + 294, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 422, + 294, + 454 + ], + "spans": [ + { + "bbox": [ + 56, + 422, + 294, + 454 + ], + "type": "text", + "content": "[37] Yuval Nirkin, Yosi Keller, and Tal Hassner. FSGAN: Subject agnostic face swapping and reenactment. In ICCV, pages 7183-7192, 2019. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 456, + 294, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 456, + 294, + 488 + ], + "spans": [ + { + "bbox": [ + 56, + 456, + 294, + 488 + ], + "type": "text", + "content": "[38] Utkarsh Ojha, Yuheng Li, and Yong Jae Lee. Towards universal fake image detectors that generalize across generative models. In CVPR, pages 24480-24489, 2023. 2, 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 490, + 294, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 490, + 294, + 544 + ], + "spans": [ + { + "bbox": [ + 56, + 490, + 294, + 544 + ], + "type": "text", + "content": "[39] Trevine Oorloff, Surya Koppisetti, Nicolò Bonettini, Dipyaraj Solanki, Ben Colman, Yaser Yacoob, Ali Shahriyari, and Gaurav Bharaj. AVFF: Audio-visual feature fusion for video deepfake detection. In CVPR, pages 27102-27112, 2024. 2, 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 545, + 294, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 545, + 294, + 589 + ], + "spans": [ + { + "bbox": [ + 56, + 545, + 294, + 589 + ], + "type": "text", + "content": "[40] Octavian Pascu, Adriana Stan, Dan Oneata, Elisabeta Oneata, and Horia Cucu. Towards generalisable and calibrated audio deepfake detection with self-supervised representations. In Interspeech, pages 4828-4832, 2024. 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 590, + 294, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 590, + 294, + 644 + ], + "spans": [ + { + "bbox": [ + 56, + 590, + 294, + 644 + ], + "type": "text", + "content": "[41] Alessandro Pianese, Davide Cozzolino, Giovanni Poggi, and Luisa Verdoliva. Training-free deepfake voice recognition by leveraging large-scale pre-trained models. In ACM Workshop on Information Hiding and Multimedia Security, page 289–294, 2024. 2, 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 647, + 294, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 294, + 679 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 294, + 679 + ], + "type": "text", + "content": "[42] K R Prajwal, Rudrabha Mukhopadhyay, Vinay P. Namboodiri, and C.V. Jawahar. A lip sync expert is all you need for speech to lip generation in the wild. In ACMM, 2020. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 681, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 681, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 681, + 294, + 713 + ], + "type": "text", + "content": "[43] Anirudh Sundara Rajan, Utkarsh Ojha, Jedidiah Schloesser, and Yong Jae Lee. On the effectiveness of dataset alignment for fake image detection. CoRR, abs/2410.11835, 2024. 1, 8" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 95 + ], + "type": "text", + "content": "[44] Tal Reiss, Bar Cavia, and Yedid Hoshen. Detecting deep-fakes without seeing any. CoRR, abs/2311.01458, 2023. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 96, + 553, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 138 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 138 + ], + "type": "text", + "content": "[45] Jonas Ricker, Denis Lukovnikov, and Asja Fischer. AEROBLADE: Training-free detection of latent diffusion images using autoencoder reconstruction error. In CVPR, pages 9130-9140, 2024. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 140, + 553, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 140, + 553, + 182 + ], + "spans": [ + { + "bbox": [ + 316, + 140, + 553, + 182 + ], + "type": "text", + "content": "[46] Andreas Rossler, Davide Cozzolino, Luisa Verdoliva, Christian Riess, Justus Thies, and Matthias Niessner. FaceForensics: Learning to detect manipulated facial images. In ICCV, 2019. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 184, + 553, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 184, + 553, + 227 + ], + "spans": [ + { + "bbox": [ + 316, + 184, + 553, + 227 + ], + "type": "text", + "content": "[47] Davide Salvi, Brian C. Hosler, Paolo Bestagini, Matthew C. Stamm, and Stefano Tubaro. TIMIT-TTS: A text-to-speech dataset for multimodal synthetic media detection. IEEE Access, 11:50851-50866, 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 228, + 553, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 228, + 553, + 271 + ], + "spans": [ + { + "bbox": [ + 316, + 228, + 553, + 271 + ], + "type": "text", + "content": "[48] Sahibzada Adil Shahzad, Ammarah Hashmi, Yan-Tsung Peng, Yu Tsao, and Hsin-Min Wang. AV-Lip-Sync+: Leveraging AV-HuBERT to exploit multimodal inconsistency for video deepfake detection. CoRR, abs/2311.02733, 2023. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 273, + 553, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 273, + 553, + 314 + ], + "spans": [ + { + "bbox": [ + 316, + 273, + 553, + 314 + ], + "type": "text", + "content": "[49] Bowen Shi, Wei-Ning Hsu, Kushal Lakhotia, and Abdelrahman Mohamed. Learning audio-visual speech representation by masked multimodal cluster prediction. In ICLR, 2022. 2, 5, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 316, + 553, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 316, + 553, + 347 + ], + "spans": [ + { + "bbox": [ + 316, + 316, + 553, + 347 + ], + "type": "text", + "content": "[50] Kaede Shiohara and Toshihiko Yamasaki. Detecting deepfakes with self-blended images. In CVPR, pages 18699-18708, 2022. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 350, + 553, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 350, + 553, + 381 + ], + "spans": [ + { + "bbox": [ + 316, + 350, + 553, + 381 + ], + "type": "text", + "content": "[51] Stefan Smeu, Elisabeta Oneata, and Dan Oneata. DeCLIP: Decoding CLIP representations for deepfake localization. CoRR, abs/2409.08849, 2024. 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 383, + 553, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 383, + 553, + 426 + ], + "spans": [ + { + "bbox": [ + 316, + 383, + 553, + 426 + ], + "type": "text", + "content": "[52] Joon Son Chung, Andrew Senior, Oriol Vinyals, and Andrew Zisserman. Lip reading sentences in the wild. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6447-6456, 2017. 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 427, + 553, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 427, + 553, + 469 + ], + "spans": [ + { + "bbox": [ + 316, + 427, + 553, + 469 + ], + "type": "text", + "content": "[53] Hemlata Tak, Jose Patino, Massimiliano Todisco, Andreas Nautsch, Nicholas Evans, and Anthony Larcher. End-to-end anti-spoofing with RawNet2. In ICASSP, pages 6369-6373, 2021. 4" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 472, + 553, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 472, + 553, + 503 + ], + "spans": [ + { + "bbox": [ + 316, + 472, + 553, + 503 + ], + "type": "text", + "content": "[54] Aäron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. CoRR, abs/1807.03748, 2018. 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 505, + 553, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 505, + 553, + 547 + ], + "spans": [ + { + "bbox": [ + 316, + 505, + 553, + 547 + ], + "type": "text", + "content": "[55] Jiadong Wang, Xinyuan Qian, Malu Zhang, Robby T. Tan, and Haizhou Li. Seeing what you said: Talking face generation guided by a lip reading expert. In CVPR, pages 14653-14662, 2023. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 548, + 553, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 548, + 553, + 635 + ], + "spans": [ + { + "bbox": [ + 316, + 548, + 553, + 635 + ], + "type": "text", + "content": "[56] Xin Wang, Junichi Yamagishi, Massimiliano Todisco, Héctor Delgado, Andreas Nautsch, Nicholas W. D. Evans, Md. Sahidullah, Ville Vestman, Tomi Kinnunen, Kong Aik Lee, Lauri Juvela, Paavo Alku, Yu-Huai Peng, Hsin-Te Hwang, Yu Tsao, Hsin-Min Wang, Sébastien Le Maguer, Markus Becker, and Zhen-Hua Ling. ASVspoof 2019: A large-scale public database of synthesized, converted and replayed speech. Comput. Speech Lang., 64:101114, 2020. 1" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 637, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 637, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 637, + 553, + 689 + ], + "type": "text", + "content": "[57] Zhiyuan Yan, Taiping Yao, Shen Chen, Yandan Zhao, Xinghe Fu, Junwei Zhu, Donghao Luo, Li Yuan, Chengjie Wang, Shouhong Ding, and Yunsheng Wu. DF40: Toward next-generation deepfake detection. CoRR, abs/2406.13495, 2024.8" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "text", + "content": "[58] Wenyuan Yang, Xiaoyu Zhou, Zhikai Chen, Bofei Guo, Zhongjie Ba, Zhihua Xia, Xiaochun Cao, and Kui Ren." + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18824" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 183 + ], + "type": "list", + "angle": 0, + "index": 3, + "blocks": [ + { + "bbox": [ + 75, + 72, + 294, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 72, + 294, + 105 + ], + "spans": [ + { + "bbox": [ + 75, + 72, + 294, + 105 + ], + "type": "text", + "content": "AVoiD-DF: Audio-visual joint learning for detecting deepfake. IEEE Trans. Inf. Forensics Secur., 18:2015-2029, 2023. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 106, + 295, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 106, + 295, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 106, + 295, + 140 + ], + "type": "text", + "content": "[59] Ning Yu, Larry S Davis, and Mario Fritz. Attributing fake images to GANs: Learning and analyzing GAN fingerprints. In CVPR, pages 7556-7566, 2019. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 141, + 294, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 141, + 294, + 183 + ], + "spans": [ + { + "bbox": [ + 56, + 141, + 294, + 183 + ], + "type": "text", + "content": "[60] Yinglin Zheng, Jianmin Bao, Dong Chen, Ming Zeng, and Fang Wen. Exploring temporal coherence for more general video face forgery detection. In ICCV, pages 15024-15034, 2021. 2" + } + ] + } + ], + "index": 2 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "18825" + } + ] + } + ], + "index": 4 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/CityWalker_ Learning Embodied Urban Navigation from Web-Scale Videos/a9edaa72-0d38-4c4e-a25a-9480eb60d753_content_list.json b/2025/CityWalker_ Learning Embodied Urban Navigation from Web-Scale Videos/a9edaa72-0d38-4c4e-a25a-9480eb60d753_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..209d912f092cc3296d6ebed9f9072c60872688fd --- /dev/null +++ b/2025/CityWalker_ Learning Embodied Urban Navigation from Web-Scale Videos/a9edaa72-0d38-4c4e-a25a-9480eb60d753_content_list.json @@ -0,0 +1,1602 @@ +[ + { + "type": "text", + "text": "CityWalker: Learning Embodied Urban Navigation from Web-Scale Videos", + "text_level": 1, + "bbox": [ + 117, + 130, + 880, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xinhao Liu* Jintong Li* Yicheng Jiang Niranjan Sujay Zhicheng Yang \nJuexiao Zhang John Abanes Jing Zhang Chen Feng† \nNew York University", + "bbox": [ + 184, + 180, + 805, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "https://ai4ce.github.io/CityWalker/", + "bbox": [ + 338, + 236, + 653, + 250 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a1fec92d74c553c785a46a93318511fffcceb86a4406b4c578d6f80625d6d74e.jpg", + "image_caption": [ + "Figure 1. Embodied Urban Navigation. Navigating urban spaces is challenging for (especially off-street) mobile agents. The differently colored pins $(\\bullet)$ along the route highlight various critical scenarios unique to complex and dynamic urban landscapes. Thumbsails on the right with corresponding colored pins demonstrate the real-world observation of these challenging cases. Our CityWalker model is trained with over 2000 hours of city walking videos and fine-tuned with a small amount of expert data to address these challenges effectively." + ], + "image_footnote": [], + "bbox": [ + 94, + 263, + 297, + 459 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/39c6b186b02948833978a4d38b8df20737fd408be235e3a3581711d0a87f12a0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 302, + 262, + 411, + 459 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c59de79307c0829cbafaa166c9aca5ce5fd4239a7ecbcf07eff72dabb96b9abc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 413, + 266, + 534, + 459 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3933f2cb71831601565ec90e622646d0e5d0c66910bdda4c578ce902f85546aa.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 534, + 266, + 653, + 459 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/816003afc62fae5ff8d1c4875a468feab2afd20e5f308e411bd63cfc55a9b542.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 655, + 267, + 777, + 459 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5d8098eb857d237ff52acff78373d423fa0be62db101fda42e13f16c506b3ce3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 777, + 267, + 897, + 459 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 530, + 326, + 545 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Navigating dynamic urban environments presents significant challenges for embodied agents, requiring advanced spatial reasoning and adherence to common-sense norms. Despite progress, existing visual navigation methods struggle in map-free or off-street settings, limiting the deployment of autonomous agents like last-mile delivery robots. To overcome these obstacles, we propose a scalable, data-driven approach for human-like urban navigation by training agents on thousands of hours of in-the-wild city walking and driving videos sourced from the web. We introduce a simple and scalable data processing pipeline that extracts action supervision from these videos, enabling large-scale imitation learning without costly annotations. Our model learns sophisticated navigation policies to handle diverse challenges and critical scenarios. Experimental results show that training on large-scale, diverse datasets significantly enhances navigation performance, surpassing current methods. This work shows the potential of using abundant online video data to develop robust navigation policies for embodied agents in dynamic urban settings.", + "bbox": [ + 88, + 564, + 485, + 867 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 514, + 530, + 643, + 545 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Visual navigation is a crucial capability for mobile agents. In previously unseen environments, humans typically rely on navigation tools (such as Google Maps) to reach a goal location by following a series of waypoints. These tools provide high-level guidance, but the actual navigation between waypoints requires sophisticated spatial awareness and decision-making. In urban settings, this involves understanding and adhering to complex rules, such as common sense regulations and norms, while dynamically responding to obstacles and environmental changes. The diversity and complexity encountered in urban navigation scenarios pose significant challenges for navigation policies but are essential for deploying mobile robots in real-world settings. Applications like delivery robots and robotaxis, especially without high-definition maps, highlight the growing need for embodied agents to navigate efficiently and safely in dynamic urban landscapes. How can we teach embodied agents to navigate dynamic urban landscapes as effectively as humans do?", + "bbox": [ + 511, + 561, + 906, + 848 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The question has been explored in early works [1-3] before the deep learning era. These approaches combined SLAM-based mapping with modular systems for search", + "bbox": [ + 511, + 854, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution ( {xinhao.liu, jintong.li}@nyu.edu)", + "bbox": [ + 107, + 875, + 477, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "†Corresponding author (cfeng@nyu.edu)", + "bbox": [ + 109, + 887, + 346, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "6875", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ing, detection, analysis, and planning. The development of high-performance photorealistic simulators [4-7] significantly advanced research in visual navigation by enabling reinforcement learning (RL) in static indoor environments, achieving near-perfect point-goal navigation [8]. This simulation-RL paradigm was also extended to more complex settings, including real-world panoramic street views [9-12]. Recent works begin to explore imitation learning from expert demonstrations for long-horizon or cross-embodiment navigation tasks [13, 14].", + "bbox": [ + 89, + 90, + 480, + 242 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Despite their success in simulators or moderately complex environments, existing methods fall short in addressing the complexity of embodied urban navigation. Public urban spaces are inherently dynamic and unpredictable, characterized by \"multifarious terrains, diverse obstacles, and dense pedestrians\" that require real-time adaptability [15]. Moreover, effective navigation in such settings demands adherence to common sense navigation rules and social norms, such as using sidewalks, obeying traffic signals, and maintaining appropriate personal space to avoid conflicts [16-18]. Current reinforcement learning and imitation learning approaches typically excel in static or controlled environments but struggle with real-world urban navigation. This is because these nuanced behaviors and delicate constraints are challenging to incorporate into simulation environments while ensuring reliable and generalizable sim-to-real transfer. Moreover, coverage of these complexities is still lacking in existing teleoperation demonstration datasets. As a result, embodied urban navigation remains an unsolved problem.", + "bbox": [ + 91, + 244, + 483, + 547 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "While a straightforward approach to the problem is to collect expert trajectories through teleoperation, this is limited by data size and lack of diversity, which weakens the agent's ability to generalize across various urban scenarios. Is there a data-driven approach that requires minimum annotation but maximizes generalizability? Inspired by the success of scaling laws in language, vision, and robotics tasks [19-22], we propose a scalable framework, CityWalker, which leverages web-scale city walking and driving videos to train models for embodied urban navigation. Our model is trained on over two thousand hours of internet-sourced videos across different geolocations, weather conditions, and times of day.", + "bbox": [ + 89, + 550, + 482, + 746 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The subsequent question is how to effectively obtain action supervision from in-the-wild videos for imitation learning. Current solutions rely heavily on prompting proprietary VLMs to generate action labels from videos [23, 24], which is costly and challenging to control, making it difficult to scale for large datasets. We argue that noisy pseudo labels from off-the-shelf visual odometry (VO) models [25] are sufficiently effective for imitation learning in embodied urban navigation. Moreover, this data processing method generalizes beyond walking data. We demonstrate that", + "bbox": [ + 89, + 750, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "a model trained with driving videos also performs well on quadruped agents, with further improvements observed when training in a cross-domain setting that combines both city walking and driving data.", + "bbox": [ + 511, + 90, + 903, + 151 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We summarize our contribution as the following:", + "bbox": [ + 532, + 152, + 856, + 166 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We identify embodied urban navigation as an intriguing yet unsolved problem. To address this problem, we propose a scalable, data-driven solution by leveraging web-scale city walking and driving videos.", + "- We introduce a simple yet scalable data processing paradigm, making large-scale imitation learning feasible for urban navigation without extensive manual labeling.", + "- We demonstrate training on web-scale data significantly improves navigation performance in real-world experiments, allowing embodied agents to handle the complexities of urban scenarios. We also show model performance scales well with training data size." + ], + "bbox": [ + 511, + 167, + 903, + 349 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 362, + 653, + 378 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Navigation in Simulation. The introduction of high-performance photorealistic simulators [4-7] has significantly boosted navigation research in indoor household environments. Based on the form of goal specification, navigation tasks are usually categorized into point-goal [4, 8, 26, 27], image-goal [28-31], object-goal [32-34], and vision-language navigation [35-37]. While tasks like image-goal and object-goal navigation focus on efficient goal-finding or instruction understanding, point-goal navigation has a more straightforward goal specification and emphasizes safe and efficient trajectory planning to the goal. Our problem formulation resembles point-goal navigation in that accurate waypoints (represented as GPS coordinates) are readily available from navigation tools. Recently, point-goal navigation has been considered a \"solved\" problem due to near-perfect performance achieved in simulators [8]. However, our investigation shows that waypoint-goal navigation in public urban spaces remains an unsolved challenge.", + "bbox": [ + 511, + 386, + 903, + 657 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Real-World Navigation. Real-world robot navigation has a long history, predating the deep learning era. Early methods relied on ranging sensors and modular SLAM systems for navigation [1-3]. Witnessing the success in simulators, various methods have been proposed to bridge the sim-to-real gap [38-41]. Another thread of work learns navigation policies from expert trajectories [42, 43]. By incorporating data from different embodiments [14] and carefully designed heuristics [13], these methods have been extended to long-horizon navigation on different robots [44, 45]. While deployed on real-world embodied agents, these methods still operate in relatively simple and static environments such as suburban areas, parking garages, or controlled obstacle courses. They are not guaranteed to generalize to significantly different, dynamic, and complex urban landscapes. Our work follows this line of research by training naviga", + "bbox": [ + 511, + 659, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "6876", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/6a7facbf87e4bdf11ed8d4bd2974e20170545956d39babae1641fd92bed4f3e7.jpg", + "image_caption": [ + "Figure 2. Overall Illustration of CityWalker. Our training pipeline starts with internet-sourced videos, using visual odometry to obtain relative poses between frames. At each time step, the model receives past observations, past trajectory, and target location as input. They are encoded via a frozen image encoder and a trainable coordinate encoder. A transformer processes these inputs to generate future tokens. An action head and an arrival head decode these tokens into action and arrival status predictions. During training, future frame tokens from future frames guide the transformer to hallucinate future tokens." + ], + "image_footnote": [], + "bbox": [ + 138, + 90, + 857, + 381 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "tion models with imitation learning. But we focus on learning the rules and norms essential for urban environments, which is crucial for enabling applications such as delivery robots and robot-taxis.", + "bbox": [ + 89, + 473, + 482, + 532 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In-the-Wild Video Learning. Learning from web-scale data has proven successful in language and vision tasks [19-21, 46]. While some works utilize in-the-wild videos to extract intermediate representations, priors, or reward functions [47, 48], a key challenge remains: the lack of action labels. Patel et al. [49] use a reconstruction pipeline to obtain hand-object trajectories as action labels for learning robot manipulation. SelfD [50] argues BEV model is better than VO in pseudo-labeling, while we show VO works well with large-scale data. Recent works also rely on prompting closed VLMs to get action labels [23, 51]. LeLaN [24], a concurrent work most relevant to ours, also relies on VLM prompting and uses pretrained navigation models to generate action labels. In contrast, we process in-the-wild video data differently by using only the noisy pseudolabels from off-the-shelf visual SLAM tools [25]. This approach enables scalable and cost-effective action label generation for imitation learning in urban navigation tasks.", + "bbox": [ + 89, + 535, + 482, + 806 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Embodied Urban Navigation", + "text_level": 1, + "bbox": [ + 89, + 821, + 356, + 839 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Problem Formulation", + "text_level": 1, + "bbox": [ + 89, + 847, + 294, + 862 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this work, we address the problem of embodied urban navigation for agents in dynamic urban environments. The", + "bbox": [ + 89, + 869, + 482, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "agent's task is to navigate from its current location to a specified target waypoint location, following a series of waypoints provided by a navigation tool. Formally, we define this as a point-goal navigation problem in real-world urban settings. At each time step $t$ , the agent receives an RGB observation $o_t$ , its current GPS location $p_t$ , and a sub-goal waypoint $w_t$ . The agent aims to learn a policy $\\pi(a_t \\mid o_{(t-k):t}, p_{(t-k):t}, w_t)$ that maps the past observations and positional information to an action $a_t$ from the action space $A$ , represented as a series of action waypoints in Euclidean space. Typically, we take $k = 5$ and also predict actions for the next 5 time steps.", + "bbox": [ + 511, + 473, + 906, + 654 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As accurate waypoints are easily available from navigation tools (e.g. Google Maps API), we focus our problem on the navigation between two consecutive waypoints. The agent treats the current sub-goal $w_{t}$ as the immediate target and is responsible for determining when it has reached this sub-goal based on its observations and positional data. Upon reaching $w_{t}$ , the agent proceeds to the next waypoint in the sequence.", + "bbox": [ + 511, + 655, + 908, + 777 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Evaluation Metrics", + "text_level": 1, + "bbox": [ + 511, + 787, + 697, + 801 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To enable analysis and assessment of different methods on the problem, our evaluation metrics are twofold. We use the ground truth trajectory in teleoperation-collected data to provide a thorough analysis of each method's performance across various scenarios. Below, we discuss the metrics we used in offline data evaluation. For real-world deployment,", + "bbox": [ + 511, + 810, + 906, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "6877", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/dadde6cde915baf037f3e33ff4541f0b884cbd63c0971954b01df46fa4829efc.jpg", + "image_caption": [ + "Figure 3. Evaluation Metrics. Left: The orientation error is defined to be the angle between each predicted and ground truth action pair, labeled by $\\theta_{n}$ in the figure. The action angle $\\varphi_{\\mathrm{action}}$ and target angle $\\varphi_{\\mathrm{target}}$ are defined with respect to the positive y-axis. Right: Both green and red trajectories are predicted actions. The green trajectory is the preferred one, having a large L2 distance but a small AOE. Vice versa for the red trajectory." + ], + "image_footnote": [], + "bbox": [ + 93, + 80, + 285, + 239 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/d957e4dc9fa6bdda9adae73a099f6720b20a085f50a68eca2fb063f5f5f94cb7.jpg", + "image_caption": [ + "Figure 4. Data Sample and Visual Odometry (VO) Result. Our internet-source training data includes both walking and driving videos. These videos cover various scenarios in the urban environment. The VO tool gives noisy trajectories globally, but trustworthy local relative pose within a short time period." + ], + "image_footnote": [], + "bbox": [ + 285, + 80, + 478, + 239 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "we primarily use navigation success rate as the key metric for performance comparison.", + "bbox": [ + 89, + 353, + 482, + 383 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Average Orientation Error. We use the error between each predicted action and the ground truth action as the main evaluation metric on the offline teleoperation data. While L2-distance in Euclidean space is a straightforward metric, we find it does not adequately capture the problem. As shown on the right side of Fig. 3, the red trajectory has a smaller L2 error compared to the green trajectory but moves in a direction that deviates from the target. This observation motivates us to come up with a new evaluation metric that better reflects the quality of the predicted actions. We define average orientation error (AOE) as the angle difference between the predicted and ground truth actions, averaged across all samples at each step (left side of Fig. 3):", + "bbox": [ + 89, + 383, + 483, + 580 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {A O E} (k) = \\frac {1}{n} \\sum_ {i} ^ {n} \\theta_ {i _ {k}} = \\frac {1}{n} \\sum_ {i} ^ {n} \\arccos \\frac {\\left\\langle a _ {\\hat {i} _ {k}} ^ {\\prime} , a _ {i _ {k}} \\right\\rangle}{\\| a _ {\\hat {i} _ {k}} \\| \\| a _ {i _ {k}} \\|}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 107, + 590, + 482, + 631 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $k$ is the index of predicted action and $n$ is the number of data. Additionally, we observe that taking the mean across $k$ predicted future time steps can lead to an underestimation due to steps with very small errors. Hence, we propose using the maximum average orientation error (MAOE) over all predicted actions as an overall estimation:", + "bbox": [ + 89, + 642, + 483, + 733 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {M A O E} = \\frac {1}{n} \\sum_ {i} ^ {n} \\max _ {k} \\theta_ {i _ {k}}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 197, + 743, + 482, + 782 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As orientation error might be noisy when action distance is small, we also use L2-distance as a complimentary evaluation metric.", + "bbox": [ + 89, + 794, + 482, + 838 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Critical Scenarios. Not all time steps are equally critical in a trajectory. For example, errors at road crossings or turning points are more likely to result in navigation failure compared to mistakes during forward movement in open areas.", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/8cd6e988b5616c6a39f97a64c82c9d0084ca7375cb3a91e95dfe052cb4bc82e0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 517, + 73, + 908, + 186 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We identify several key scenarios that are most critical during navigation. AOE and MAOE are calculated separately for each scenario, and we use the scenario mean rather than the sample mean for a more comprehensive evaluation. Detailed definitions of each scenario are provided in Sec. 4.", + "bbox": [ + 511, + 279, + 905, + 354 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Learning from In-the-Wild Videos", + "text_level": 1, + "bbox": [ + 511, + 364, + 815, + 380 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "City Walking Videos. To facilitate large-scale imitation learning for human-like urban navigation, we utilize city walking videos sourced from the internet. Walking in urban environments is intuitive for humans, resulting in an abundance of first-person view videos available online. These videos naturally capture the complexities of navigating dynamic urban landscapes, including interactions with pedestrians, adherence to traffic signals, and maneuvering around various obstacles. Motivated by the extensive availability of such data and its inherent alignment with our navigation objectives, we curate a dataset comprising over 2000 hours of city walking videos, encompassing diverse weather and lighting conditions. These videos exhibit consistent motion patterns characteristic of walking, providing continuous motion cues that can directly serve as action labels.", + "bbox": [ + 511, + 386, + 906, + 612 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Action Labels from Videos. To generate actionable labels for imitation learning, we employ VO [25, 52-54] tools to extract trajectory poses from the collected city walking videos. Specifically, we utilize the state-of-the-art DPVO [25] while any other VO method should also work. The resulting VO trajectories are illustrated in Fig. 4. However, using VO poses presents two primary challenges. First, global trajectory accuracy is compromised by accumulation errors inherent in VO systems. Nevertheless, since our model focuses on predicting a limited sequence of actions (5 past and 5 future), we mitigate this issue by relying solely on relative poses within short temporal windows. Second, VO methods suffer from scale ambiguity, leading to inconsistencies in trajectory lengths across different data. For instance, as shown in Fig. 4, a driving video may exhibit a longer actual trajectory compared to a walking video, yet the DPVO output incorrectly reflects a shorter trajectory. To address this, we normalize each action by the average step length within its trajectory, following the approach in", + "bbox": [ + 511, + 613, + 908, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "6878", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "previous work [14]. This normalization resolves both scale ambiguity and inconsistencies in step lengths between different types of embodiments all at once. This enables the learning of policies in a consistent, abstracted action space. During deployment, the predicted actions are denormalized using the specific step length of the target robot, ensuring appropriate movement execution in real-world settings.", + "bbox": [ + 89, + 90, + 480, + 196 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Generalizability and Scalability. Our data processing pipeline is both generalizable and scalable. It can be applied to any video sources beyond city walking, as long as the videos are captured with egomotion, such as driving videos, which are also abundantly available online. This flexibility enables the training of more generalizable navigation policies that can be zero-shot applied or fine-tuned for cross-domain and cross-embediment tasks, such as transitioning from pedestrian to vehicular navigation. Furthermore, our method offers significant scalability advantages. Unlike pipelines that rely on prompting large language models for action label generation [23, 24], our visual odometry-based approach can be easily parallelized. Processing 2000 hours of video data with our pipeline requires negligible wall-clock time, making it both feasible and cost-efficient to scale up the training dataset.", + "bbox": [ + 89, + 198, + 482, + 438 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Pipeline and Training", + "text_level": 1, + "bbox": [ + 89, + 446, + 294, + 462 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Pipeline Architecture. Figure 2 shows our pipeline. It builds upon previous works [44, 45]. At the core of our model is a transformer that processes a temporal sequence of input tokens, including $k$ image features and one coordinate embeddings. Specifically, the coordinate input is stacked by $k$ past positions and one single target position, and outputs a sequence of the same length. We use an action head and an arrival prediction head to decode the output of the transformer. We utilize DINOv2 [21] representation and freeze the backbone during training.", + "bbox": [ + 89, + 468, + 482, + 619 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Feature Hallucination. Inspired by previous works on feature learning [55-57], we use feature hallucination as an auxiliary loss during training. Specifically, we compute the MSE loss between the output image tokens and that directly extracted from future frames. We hope this will guide the transformer to predict more informative future tokens that best mimic future frames, which in turn helps the two MLP heads predict the final action and arrival status.", + "bbox": [ + 89, + 619, + 482, + 739 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Loss Functions. As mentioned in Sec. 3.2, L1 or L2 loss can be less effective than orientation error in evaluating the quality of the predicted actions. We also employ orientation error to supervise the training. We define the orientation loss to be the negative cosine similarity between the predicted actions and the ground truth actions:", + "bbox": [ + 89, + 739, + 482, + 830 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {o r i}} = - \\frac {1}{k} \\sum_ {i = 1} ^ {k} \\frac {\\left\\langle \\hat {a} _ {i} , a _ {i} \\right\\rangle}{\\left\\| \\hat {a} _ {i} \\right\\| \\left\\| a _ {i} \\right\\|}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 197, + 838, + 482, + 878 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Together with the L1 loss for predicted action and the BCE", + "bbox": [ + 89, + 885, + 482, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "loss for predicted arriving status, the final loss function is a weight sum of four individual losses:", + "bbox": [ + 511, + 90, + 906, + 121 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\omega_ {l 1} \\mathcal {L} _ {l 1} + \\omega_ {\\text {o r i}} \\mathcal {L} _ {\\text {o r i}} + \\omega_ {\\text {a r r}} \\mathcal {L} _ {\\text {a r r}} + \\omega_ {\\text {f e a t}} \\mathcal {L} _ {\\text {f e a t}}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 539, + 133, + 906, + 150 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We choose each weight so that each loss items are within the same magnitude.", + "bbox": [ + 511, + 161, + 906, + 191 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiment", + "text_level": 1, + "bbox": [ + 511, + 205, + 638, + 222 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Setup", + "text_level": 1, + "bbox": [ + 511, + 229, + 596, + 244 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Baselines. We compare our model with outdoor navigation models closely related to our setup, including GNM [14], ViNT [44], and NoMaD [45]. Although originally developed for image-goal navigation, we tested them using goal-images from our collected data. Additionally, ViNT as a foundation model for visual navigation supports finetuning, which we also evaluated. We acknowledge CoNVOI [58] as a recent work with a similar setup, but could not test it due to the lack of open source code and prompt. We also experimented with VLM-based approaches similar to CoNVOI and defer the results to the appendix.", + "bbox": [ + 511, + 251, + 906, + 417 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Data. We collected expert data via teleoperation for fine-tuning and offline testing. We collect data via a Unitree Go1 quadruped equipped with a Livox Mid-360 LiDAR and a webcam for RGB observations. We employed the LiDAR-SLAM method [59] to obtain the robot's poses as ground truth actions. Additionally, we retrieved location data from a smartphone using a webpage. We reduce noise from the quadruped's vibrations by holding the phone [38]. In total, we gathered 15 hours of teleoperation data across various areas in New York City, allocating 6 hours for fine-tuning and 9 hours for testing.", + "bbox": [ + 511, + 417, + 905, + 584 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Critical Scenarios As discussed in Sec. 3.2, we identify several critical scenarios in our dataset that we want to emphasize in our evaluation. We analyze the ground truth trajectory and run object detection [60] to aid us in standardizing the definition. We define these scenarios as follows:", + "bbox": [ + 511, + 584, + 905, + 657 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Turn: When the ground truth action changes direction significantly. Defined when $\\varphi_{\\mathrm{action}} > 20^{\\circ}$ (see Fig. 3).", + "- Crossing: When the agent is at a road crossing. Defined if any traffic light is detected with score $>0.5$ .", + "- Detour: When the action angle deviates from the target angle. Defined when $|\\varphi_{\\text{action}} - \\varphi_{\\text{target}}| > 45^{\\circ}$ (see Fig. 3).", + "- Proximity: When any person is close to the agent. Defined when the largest bounding box of any detected person is larger than $25\\%$ of the image area.", + "- **Crowd:** When the agent is surrounded by a crowd of people. Defined when $\\geq 5$ people are detected." + ], + "bbox": [ + 511, + 659, + 903, + 824 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We define these scenarios in a way that they are not mutually exclusive. A single data sample can belong to multiple scenarios. Although these scenarios account for less than half of the data, we argue they are the most significant factors contributing to successful urban navigation.", + "bbox": [ + 511, + 825, + 905, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "6879", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/c07603de6a50fc5a23577aaa4a9c99bce841123d79655c139f79e93e600cb4d5.jpg", + "table_caption": [ + "Table 1. Benchmark on Offline Data. We evaluate three metrics in each critical scenario for all methods. Percentages under scenarios indicate their data proportions. The \"Mean\" column shows scenario means averaged over six scenarios; \"All\" shows sample means over all data samples. We highlight the best performance in bold and the second-best performance with underline." + ], + "table_footnote": [], + "table_body": "
MethodMetricMeanTurn 8%Crossing 12%Detour 12%Proximity 6%Crowd 7%Other 55%All 100%
GNM [14] (fine-tuned)↓L2 (m)1.222.361.361.420.880.760.550.74
↓MAOE (°)16.231.114.812.514.712.811.012.1
↑Arrival (%)68.666.469.766.469.069.270.770.0
ViNT [44] (fine-tuned)↓L2 (m)1.301.911.131.140.770.660.570.70
↓MAOE (°)16.531.115.412.914.813.311.612.6
↑Arrival (%)70.571.268.770.773.068.671.070.7
NoMaD [45] (fine-tuned)↓L2 (m)1.392.491.561.551.060.950.760.74
↓MAOE (°)19.135.118.515.618.114.312.812.1
↑Arrival (%)68.666.469.766.469.069.370.770.0
Ours (zero-shot)↓L2 (m)1.341.301.091.331.441.481.391.38
↓MAOE (°)16.526.815.516.316.313.111.312.7
↑Arrival (%)79.169.371.478.884.184.586.484.1
Ours (fine-tuned)↓L2 (m)1.111.271.001.151.061.121.061.07
↓MAOE (°)15.226.614.113.914.312.010.411.5
↑Arrival (%)81.868.975.378.590.687.590.287.8
", + "bbox": [ + 174, + 142, + 823, + 400 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Performance Benchmarking", + "text_level": 1, + "bbox": [ + 89, + 422, + 344, + 439 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this subsection, we aim to answer the question Q1: Can our CityWalker model navigate successfully in complex urban environments? To answer this, we benchmark our model and the baselines both with offline data and in real-world navigation deployment.", + "bbox": [ + 89, + 446, + 482, + 521 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Overall Benchmark. Table 1 shows the benchmark in different critical scenarios. Overall, our fine-tuned model has superior performance on all metrics in all scenarios except \"detour\". We think the less competent performance in the detour scenario is due to the small portion of such data in our video training data. It can be verified by the fact that the fine-tuning performance gain is most significant in this scenario. It is worth noting that our zero-shot model has similar or even better performance compared to the fine-tuned baselines. This demonstrates the power of large-scale pretraining, especially considering the domain and modality gap from human walking videos.", + "bbox": [ + 89, + 522, + 482, + 704 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Real-world Depolyment. We deploy our model and the baseline methods on the same Unitree Go1 quadruped for real-world navigation, in previously unseen urban environments. We adapt the PD controller provided in [44] to provide velocity command to the quadruped. We choose area with relatively constant density of dynamics such as pedestrian and traffic to ensure the consistency among different trials. In each trial, the target location is generally 50-100m far away from the starting location. We categorize the real-world test into forward, left turn, and right turn cases to provide case-by-case analysis. We conduct 8-14 trails for each cases. Each trial is marked success when the model predicts arrival within a 5m distance to the target location.", + "bbox": [ + 89, + 704, + 482, + 901 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/46e348324a1c7eaeb2891907e89c6679d48390f533d5a586ab0acbbc3d75e7d2.jpg", + "table_caption": [ + "Table 2. Real-World Navigation. The table shows the success rate of real-world experiments in different scenarios. * and † indicate zero-shot inference and fine-tuned models respectively." + ], + "table_footnote": [], + "table_body": "
MethodAllForwardLeft turnRight turn
ViNT* [44]37.762.50.050.0
ViNT† [44]57.1100.025.025.0
NoMad* [45]42.975.016.728.6
Ours†77.3100.062.566.7
", + "bbox": [ + 535, + 468, + 880, + 551 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "All human interruptions due to potential collision or timeout are treated as failure cases.", + "bbox": [ + 511, + 563, + 903, + 593 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2 shows the success rate on all real-world test cases. Our model achieves the highest success rate among all cases with a significant gap to the second-best performing finetuned ViNT. As the baselines are trained with navigation data mostly in suburban or off-road environments [14, 61], they are limited in handling complex maneuvers, making them unsuitable for urban navigation tasks, as shown in Fig. 5. Our model's robust performance in both forward and turns highlights its ability to manage dynamic and varied urban scenarios effectively, allowing it to be more reliably used in real-world environments where frequent and precise directional changes are essential. The substantial improvement not only increases the overall navigation success but also extends the applicability of our model to more demanding and realistic urban settings, positioning it as a superior solution for embodied agents in dynamic environments.", + "bbox": [ + 511, + 594, + 906, + 835 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Power of Data Scaling", + "text_level": 1, + "bbox": [ + 511, + 847, + 720, + 864 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our model is significantly boosted by the large-scale training data. We ask Q2: To what extent does increasing the", + "bbox": [ + 511, + 869, + 905, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6880", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e2e21c1fff13ed9f1ecb15548d16055943dbfa72f0c842560a92675de7264991.jpg", + "image_caption": [ + "Current Observation" + ], + "image_footnote": [], + "bbox": [ + 116, + 112, + 267, + 229 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7a65bb3bfa3d90537fdabf04d4bf3837becdff4952ca5f0d3c3f4873e749e697.jpg", + "image_caption": [ + "X (m) Coordinates" + ], + "image_footnote": [], + "bbox": [ + 274, + 108, + 424, + 229 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0fe13c8cae1a774f18e9382a452b96400165389aca4549183e61382c09bba991.jpg", + "image_caption": [ + "ViNT Zero-shot", + "ViNT Fine-tune", + "X (m) Coordinates" + ], + "image_footnote": [], + "bbox": [ + 424, + 106, + 571, + 229 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5b2f72ea50dc2532803c1996d73c3040eb534f21fe04ebadf5490c50cfe3fa3c.jpg", + "image_caption": [ + "NoMad Zero-shot", + "X (m) Coordinates" + ], + "image_footnote": [], + "bbox": [ + 573, + 107, + 725, + 229 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/33380879e63d2d78c8e92dea8df834cd596485aa2832d9e2d3a684ed65b228fc.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 730, + 107, + 875, + 234 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7341e9001dee630a1d8f34b6f08cc51d7117895d198d9f5e56519b34cd167117.jpg", + "image_caption": [ + "Input past traj." + ], + "image_footnote": [], + "bbox": [ + 116, + 243, + 269, + 359 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/deafe897752aa4918eaeeb125d39a1b2198df2be710e84b0fa980e6319617fd4.jpg", + "image_caption": [ + "Actual past traj." + ], + "image_footnote": [], + "bbox": [ + 277, + 243, + 423, + 366 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2076933fbec325f325f7415aa8b0a289875b223f6c67030a84a69dd66368d491.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 428, + 243, + 573, + 366 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7503e3b96efdae1061cffe23de953bd09a4c58e828f94417e03853306a75f805.jpg", + "image_caption": [ + "Predicted action", + "GTaction", + "$\\star$ Target" + ], + "image_footnote": [], + "bbox": [ + 578, + 243, + 725, + 367 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ab2d565d131d2703bf9beed86839e64c15a34f1f6a091815665c16d402ed9be5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 728, + 243, + 875, + 366 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b77d40fa67f01460f9f1fb4a415fc7fc4e69cd6c17e41b360ff23bfdc32c5063.jpg", + "image_caption": [ + "Figure 5. Qualitative Results. Left image shows current observations of two samples. Right plots displays input trajectory, ground truth actions, and predicted actions in the current coordinate system with the agent at the origin.", + "Figure 6. Performance and Data Size. We show the model performance evaluated by MAOE with respect to the size of the training data measured by video length in hours. We also show the zero-shot performance of our model trained with only driving videos and mixed driving and walking videos" + ], + "image_footnote": [], + "bbox": [ + 107, + 433, + 454, + 595 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "size of the training dataset improve the performance of our urban navigation model?", + "bbox": [ + 89, + 699, + 483, + 729 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Performance Gain and Data Scale. Figure 6 shows the model performance when trained with different data sizes. The plot demonstrates that our model's zero-shot performance improves significantly as we increase the amount of training data. Notably, when trained with more than 1000 hours of videos, our zero-shot model (red line) achieves a better performance than the fine-tuned ViNT model $(\\square)$ . This indicates that large-scale training on diverse video data can even surpass the performance of models that are fine-tuned with limited expert data, highlighting the potential of leveraging web-scale datasets for embodied navigation.", + "bbox": [ + 89, + 734, + 483, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Domain and Embodiment Gap. An unexpected observation is that our model without the feature hallucination loss outperforms the one with it in zero-shot inference. We hypothesize this phenomenon mainly due the domain and embodiment gap between human walking videos and quadruped navigation policies. The feature hallucination loss aims to guide the transformer model to produce output tokens that mimic future observations, essentially encouraging the model to anticipate future visual inputs. This enforcement might inadvertently hinder performance by focusing on predicting future observations based on human motion. The model may learn representations that are less applicable or even misleading for quadruped navigation. Consequently, the feature hallucination loss could introduce a bias that detracts from the model's ability to make accurate navigation decisions in the robot's operational domain.", + "bbox": [ + 511, + 431, + 906, + 672 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Feature Hallucination Loss. Fortunately, as shown in Tab. 1, this issue no longer exists after fine-tuning. We also show the benefit of feature hallucination loss from another perspective. Figure 7 shows two training loss curve for different data size and training setups with and without feature hallucination loss. We observe that although feature hallucination loss should increase the absolute value of the total loss, the training with feature hallucination loss lead to a lower total loss for 2000-hour training and a lower orientation loss in both cases. The training with feature hallucination loss also shows a steeper decreasing trend, showing a potential of having even lower error with more training steps. Therefore, this is another proof of the effectiveness of our proposed feature hallucination loss.", + "bbox": [ + 511, + 672, + 908, + 883 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Driving Videos. As discussed in Sec. 3.3, our data process", + "bbox": [ + 511, + 885, + 903, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "6881", + "bbox": [ + 482, + 945, + 513, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1be800e77262e5e99074bb8830b35bd432270f3e860ac7f776841ec62b4a9925.jpg", + "image_caption": [ + "Figure 7. Improvement from Feature Hallucination. The plot shows the orientation loss for different amounts of training data and loss setups. For all four setups, we trian with $\\omega_{\\mathrm{II}} = 1.0, \\omega_{\\mathrm{arr}} = 1.0, \\text{and}, \\omega_{\\mathrm{ori}} = 5.0$ . The unproportional training steps are due to different batch sizes used for different data sizes. All models are trained for 10 epochs." + ], + "image_footnote": [], + "bbox": [ + 117, + 93, + 444, + 247 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ing pipeline also enables the use of abundant online driving videos. We evaluate models trained exclusively on driving videos and on a combination of walking and driving videos, results shown in Fig. 6. We can observe that the model trained only with driving data $(+)$ has a similar performance to the zero-shot baselines. As a result, we can think of driving as a form of navigation from a different domain and embodiment. However, combining driving and walking data yields significant performance gains (x). With just 250 hours of mixed data, the model nearly matches the performance of models trained on 1000 hours of walking data alone. This signifies the generalizability and robustness provided by cross-domain cross-embodiment data.", + "bbox": [ + 88, + 364, + 482, + 561 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Pipeline component analysis", + "text_level": 1, + "bbox": [ + 89, + 574, + 344, + 590 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To provide a more in-depth analysis of our model and pipeline, we ask Q3: How do the individual components of our pipeline enhance the performance and reliability of our model for real-world urban navigation?", + "bbox": [ + 89, + 597, + 482, + 657 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "AOE at Different Time Steps. In Tab. 4, we study the AOE at different time steps, i.e. $\\mathrm{AOE}(k)$ for $k$ from 1 to 5. ViNT exhibits increasing orientation errors as the prediction steps progress, indicating a degradation in accuracy over time. In contrast, our model maintains relatively consistent AOE from step 2 to step 5, which likely contributes to its superior performance in real-world navigation deployments with stable action prediction across multiple future steps. All methods experience higher errors at the first step. We attribute it to the proximity of the initial action to the origin, as minor linear deviations result in substantial angular error. It is worth noting that the MAOE is significantly larger than any individual AOE from steps 1 to 5, suggesting that the peak errors occur at varying time steps across different samples. This variability suggests the potential benefit of taking softmax for orientation errors instead of simple averaging, as it", + "bbox": [ + 89, + 659, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/d745a21044632dbe22c9dca2fbeb9e843d5c233f3f29ec5fc0f60955aaed931e.jpg", + "table_caption": [ + "Table 3. Ablation Study. We ablate on the contribution of orientation loss, feature hallucination, and fine-tuning with expert data. We show the MAOE(°) in scenario-mean. All results are trained with 1000-hour walking data." + ], + "table_footnote": [], + "table_body": "
Training ComponentsMAOE
Ori. LossFeature Hall.Fine-tuningMean
17.03
✓17.00
✓✓17.02
✓15.23
✓✓15.21
✓✓✓15.16
", + "bbox": [ + 557, + 156, + 864, + 276 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/47c1064d91cb27e634db81bf66f87fa716dd384b1a98bbde93e157051efcafc5.jpg", + "table_caption": [ + "Table 4. AOE at Different Time Steps. The table shows MAOE and different AOE of scenario-mean, all in degrees $(^{\\circ})$ . * and † indicates zero-shot inference and fine-tuned models respectively." + ], + "table_footnote": [], + "table_body": "
MethodAOE(1)AOE(2)AOE(3)AOE(4)AOE(5)MAOE
ViNT*[44]10.577.939.0910.2111.4517.52
ViNT†[44]11.247.968.408.939.8716.53
Ours*12.918.928.648.468.4316.54
Ours†11.447.777.687.757.9715.23
", + "bbox": [ + 517, + 351, + 903, + 425 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "could better capture the distribution of peak errors.", + "bbox": [ + 511, + 453, + 849, + 468 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ablation Study. Table 3 examines the individual contributions of different pipeline components to our model's performance. A clear improvement in MAOE when fine-tuning with expert data, demonstrating the effectiveness of using in-domain expert information. The improvement from orientation loss and feature hallucination loss is marginal, within error range of our limited testing data. We defer more detailed ablation study to the appendix.", + "bbox": [ + 511, + 473, + 906, + 595 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 623, + 633, + 638 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Summary. In this work, we addressed embodied urban navigation by training agents with thousands of hours of in-the-wild city walking videos, introducing a scalable pipeline for imitation learning from diverse walking videos. Our experiments show that large-scale, diverse training data significantly enhances navigation performance, surpassing existing methods and demonstrating the benefits of data scaling. Our findings highlight the potential of leveraging abundant online videos to develop robust navigation policies for embodied agents in dynamic urban environments.", + "bbox": [ + 511, + 654, + 906, + 805 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitation and Future Work. During real-world deployment, our current system implementation is sensitive to potential large location noises read naively from iPhone location services, which could be improved with better GPS hardware. Future work could also enhance the robustness of the model to such location noises.", + "bbox": [ + 511, + 809, + 908, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "6882", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgment", + "text_level": 1, + "bbox": [ + 91, + 90, + 240, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The work was supported by NSF grants 2238968, 2121391, 2322242 and 2345139; and in part through the NYU IT High Performance Computing resources, services, and staff expertise. We also thank Xingyu Liu and Zixuan Hu for their help in data collection.", + "bbox": [ + 89, + 114, + 483, + 191 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 203, + 187, + 218 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Quirin Muhlbauer, Stefan Sosnowski, Tingting Xu, Tianguang Zhang, Kolja Kuhnlenz, and Martin Buss. Navigation through urban environments by visual perception and interaction. In ICRA, pages 3558-3564. IEEE, 2009. 1, 2", + "[2] Rainer Kümmerle, Michael Ruhnke, Bastian Steder, Cyril Stachniss, and Wolfram Burgard. A navigation system for robots operating in crowded urban environments. In ICRA, pages 3225-3232. IEEE, 2013.", + "[3] Yoichi Morales, Alexander Carballo, Eijiro Takeuchi, Atsushi Aburadani, and Takashi Tsubouchi. Autonomous robot navigation in outdoor cluttered pedestrian walkways. Journal of Field Robotics, 26(8):609-635, 2009. 1, 2", + "[4] Manolis Savva, Abhishek Kadian, Oleksandr Maksymets, Yili Zhao, Erik Wijmans, Bhavana Jain, Julian Straub, Jia Liu, Vladlen Koltun, Jitendra Malik, et al. Habitat: A platform for embodied ai research. In CVPR, pages 9339-9347, 2019. 2", + "[5] Fei Xia, Amir R Zamir, Zhiyang He, Alexander Sax, Jitendra Malik, and Silvio Savarese. Gibson env: Real-world perception for embodied agents. In CVPR, pages 9068-9079, 2018.", + "[6] Eric Kolve, Roozbeh Mottaghi, Winson Han, Eli VanderBilt, Luca Weihs, Alvaro Herrasti, Matt Deitke, Kiana Ehsani, Daniel Gordon, Yuke Zhu, et al. Ai2-thor: An interactive 3d environment for visual ai. arXiv preprint arXiv:1712.05474, 2017.", + "[7] Angel Chang, Angela Dai, Thomas Funkhouser, Maciej Halber, Matthias Niessner, Manolis Savva, Shuran Song, Andy Zeng, and Yinda Zhang. Matterport3d: Learning from RGB-D data in indoor environments. 3DV, 2017. 2", + "[8] Erik Wijmans, Abhishek Kadian, Ari S. Morcos, Stefan Lee, Irfan Essa, Devi Parikh, Manolis Savva, and Dhruv Batra. Dd-ppo: Learning near-perfect pointgoal navigators from 2.5 billion frames. In ICLR, 2019. 2", + "[9] Piotr Mirowski, Matt Grimes, Mateusz Malinowski, Karl Moritz Hermann, Keith Anderson, Denis Teplyashin, Karen Simonyan, Andrew Zisserman, Raia Hadsell, et al. Learning to navigate in cities without a map. NeurIPS, 31, 2018. 2", + "[10] Ang Li, Huiyi Hu, Piotr Mirowski, and Mehrdad Farajtabar. Cross-view policy learning for street navigation. In ICCV, pages 8100-8109, 2019.", + "[11] Karl Moritz Hermann, Mateusz Malinowski, Piotr Mirowski, Andras Banki-Horvath, Keith Anderson, and Raia Hadsell. Learning to follow directions in street view. In AAAI, volume 34, pages 11773-11781, 2020.", + "[12] Piotr Mirowski, Andras Banki-Horvath, Keith Anderson, Denis Teptyashin, Karl Moritz Hermann, Mateusz Mali" + ], + "bbox": [ + 93, + 227, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "nowski, Matthew Koichi Grimes, Karen Simonyan, Koray Kavukcuoglu, Andrew Zisserman, et al. The streetlearn environment and dataset. arXiv preprint arXiv:1903.01292, 2019. 2", + "[13] Dhruv Shah and Sergey Levine. ViKiNG: Vision-Based Kilometer-Scale Navigation with Geographic Hints. In RSS, 2022. 2", + "[14] Dhruv Shah, Ajay Sridhar, Arjun Bhorkar, Noriaki Hirose, and Sergey Levine. Gnm: A general navigation model to drive any robot. In ICRA, pages 7226-7233. IEEE, 2023. 2, 5, 6", + "[15] Wayne Wu, Honglin He, Yiran Wang, Chenda Duan, Jack He, Zhizheng Liu, Quanyi Li, and Bolei Zhou. Metaurban: A simulation platform for embodied ai in urban spaces. arXiv preprint arXiv:2407.08725, 2024. 2", + "[16] Adarsh Jagan Sathyamoorthy, Utsav Patel, Moumita Paul, Nithish K Sanjeev Kumar, Yash Savle, and Dinesh Manocha. Comet: Modeling group cohesion for socially compliant robot navigation in crowded scenes. IEEE Robotics and Automation Letters, 7(2):1008-1015, 2021. 2", + "[17] Muchen Sun, Francesca Baldini, Peter Trautman, and Todd Murphey. Move beyond trajectories: Distribution space coupling for crowd navigation. In RSS, 2021.", + "[18] Yuya Onozuka, Ryosuke Matsumi, and Motoki Shino. Autonomous mobile robot navigation independent of road boundary using driving recommendation map. In IROS, pages 4501-4508. IEEE, 2021. 2", + "[19] Tom B Brown. Language models are few-shot learners. arXiv preprint arXiv:2005.14165, 2020. 2, 3", + "[20] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In ICCV, pages 4015-4026, 2023.", + "[21] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel HAZIZA, Francisco Massa, Alaaeldin El-Nouby, Mido Assran, Nicolas Ballas, Wojciech Galuba, Russell Howes, Po-Yao Huang, Shang-Wen Li, Ishan Misra, Michael Rabbat, Vasu Sharma, Gabriel Synnaeve, Hu Xu, Herve Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. DINOv2: Learning robust visual features without supervision. Transactions on Machine Learning Research, 2024. 3, 5", + "[22] Danny Driess, Fei Xia, Mehdi SM Sajjadi, Corey Lynch, Aakanksha Chowdhery, Brian Ichter, Ayzaan Wahid, Jonathan Tompson, Quan Vuong, Tianhe Yu, et al. Palme: An embodied multimodal language model. arXiv preprint arXiv:2303.03378, 2023. 2", + "[23] Beichen Wang, Juexiao Zhang, Shuwen Dong, Irving Fang, and Chen Feng. Vlm see, robot do: Human demo video to robot action plan via vision language model. arXiv preprint arXiv:2410.08792, 2024. 2, 3, 5", + "[24] Noriaki Hirose, Catherine Glossop, Ajay Sridhar, Dhruv Shah, Oier Mees, and Sergey Levine. Lelan: Learning a language-conditioned navigation policy from in-the-wild videos. arXiv preprint arXiv:2410.03603, 2024. 2, 3, 5", + "[25] Zachary Teed, Lahav Lipson, and Jia Deng. Deep patch visual odometry. NeurIPS, 36, 2024. 2, 3, 4" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "6883", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[26] Max Jaderberg, Volodymyr Mnih, Wojciech Marian Czarnecki, Tom Schaul, Joel Z Leibo, David Silver, and Koray Kavukcuoglu. Reinforcement learning with unsupervised auxiliary tasks. In ICLR, 2017. 2", + "[27] Saurabh Gupta, James Davidson, Sergey Levine, Rahul Sukthankar, and Jitendra Malik. Cognitive mapping and planning for visual navigation. In CVPR, pages 2616-2625, 2017. 2", + "[28] Nikolay Savinov, Alexey Dosovitskiy, and Vladlen Koltun. Semi-parametric topological memory for navigation. In ICLR, 2018. 2", + "[29] Meera Hahn, Devendra Singh Chaplot, Shubham Tulsiani, Mustafa Mukadam, James M Rehg, and Abhinav Gupta. No rl, no simulation: Learning to navigate without navigating. NeurIPS, 34:26661-26673, 2021.", + "[30] Obin Kwon, Nuri Kim, Yunho Choi, Hwiyeon Yoo, Jeongho Park, and Songhwai Oh. Visual graph memory with unsupervised representation for visual navigation. In CVPR, pages 15890-15899, 2021.", + "[31] Jacob Krantz, Theophile Gervet, Karmesh Yadav, Austin Wang, Chris Paxton, Roozbeh Mottaghi, Dhruv Batra, Jitendra Malik, Stefan Lee, and Devendra Singh Chaplot. Navigating to objects specified by images. In CVPR, pages 10916-10925, 2023. 2", + "[32] Theophile Gervet, Soumith Chintala, Dhruv Batra, Jitendra Malik, and Devendra Singh Chaplot. Navigating to objects in the real world. Science Robotics, 8(79):eadf6991, 2023. 2", + "[33] Devendra Singh Chaplot, Dhiraj Prakashchand Gandhi, Abhinav Gupta, and Russ R Salakhutdinov. Object goal navigation using goal-oriented semantic exploration. NeurIPS, 33:4247-4258, 2020.", + "[34] Arjun Majumdar, Gunjan Aggarwal, Bhavika Devnani, Judy Hoffman, and Dhruv Batra. Zson: Zero-shot object-goal navigation using multimodal goal embeddings. NeurIPS, 35:32340-32352, 2022. 2", + "[35] Abhishek Das, Samyak Datta, Georgia Gkioxari, Stefan Lee, Devi Parikh, and Dhruv Batra. Embodied question answering. In CVPR, pages 1-10, 2018. 2", + "[36] Hao Tan, Licheng Yu, and Mohit Bansal. Learning to navigate unseen environments: Back translation with environmental dropout. arXiv preprint arXiv:1904.04195, 2019.", + "[37] Jacob Krantz, Erik Wijmans, Arjun Majumdar, Dhruv Batra, and Stefan Lee. Beyond the nav-graph: Vision-and-language navigation in continuous environments. In ECCV, pages 104-120. Springer, 2020. 2", + "[38] Maks Sorokin, Jie Tan, C Karen Liu, and Sehoon Ha. Learning to navigate sidewalks in outdoor environments. IEEE Robotics and Automation Letters, 7(2):3906-3913, 2022. 2, 5", + "[39] Joanne Truong, Max Rudolph, Naoki Harrison Yokoyama, Sonia Chernova, Dhruv Batra, and Akshara Rai. Rethinking sim2real: Lower fidelity simulation leads to higher sim2real transfer in navigation. In CoRL, pages 859-870. PMLR, 2023.", + "[40] Joanne Truong, Denis Yarats, Tianyu Li, Franziska Meier, Sonia Chernova, Dhruv Batra, and Akshara Rai. Learning navigation skills for legged robots with learned robot embeddings. In IROS, pages 484-491. IEEE, 2021." + ], + "bbox": [ + 91, + 92, + 480, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[41] Wei Liu, Huihua Zhao, Chenran Li, Joydeep Biswas, Billy Okal, Pulkit Goyal, Yan Chang, and Soha Pouya. X-mobility: End-to-end generalizable navigation via world modeling. arXiv preprint arXiv:2410.17491, 2024. 2", + "[42] Gregory Kahn, Pieter Abbeel, and Sergey Levine. Badgr: An autonomous self-supervised learning-based navigation system. IEEE Robotics and Automation Letters, 6(2):1312-1319, 2021. 2", + "[43] Gregory Kahn, Pieter Abbeel, and Sergey Levine. Land: Learning to navigate from disengagements. IEEE Robotics and Automation Letters, 6(2):1872-1879, 2021. 2", + "[44] Dhruv Shah, Ajay Sridhar, Nitish Dashora, Kyle Stachowicz, Kevin Black, Noriaki Hirose, and Sergey Levine. ViNT: A foundation model for visual navigation. In CoRL, 2023. 2, 5, 6, 8", + "[45] Ajay Sridhar, Dhruv Shah, Catherine Glossop, and Sergey Levine. Nomad: Goal masked diffusion policies for navigation and exploration. In ICRA, pages 63-70, 2024. 2, 5, 6", + "[46] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763. PMLR, 2021. 3", + "[47] Annie S Chen, Suraj Nair, and Chelsea Finn. Learning generalizable robotic reward functions from \"in-the-wild\" human videos. arXiv preprint arXiv:2103.16817, 2021. 3", + "[48] Shikhar Bahl, Abhinav Gupta, and Deepak Pathak. Human-to-robot imitation in the wild. In RSS, 2022. 3", + "[49] Austin Patel, Andrew Wang, Ilija Radosavovic, and Jitendra Malik. Learning to imitate object interactions from internet videos. arXiv preprint arXiv:2211.13225, 2022. 3", + "[50] Jimuyang Zhang, Ruizhao Zhu, and Eshed Ohn-Bar. Selfd: Self-learning large-scale driving policies from the web. In CVPR, pages 17316-17326, 2022. 3", + "[51] Naoki Wake, Atsushi Kanehira, Kazuhiro Sasabuchi, Jun Takamatsu, and Katsushi Ikeuchi. Gpt-4v (ision) for robotics: Multimodal task planning from human demonstration. IEEE Robotics and Automation Letters, 2024. 3", + "[52] Jakob Engel, Vladlen Koltun, and Daniel Cremers. Direct sparse odometry. TPAMI, 40(3):611-625, 2017. 4", + "[53] Raul Mur-Artal, Jose Maria Martinez Montiel, and Juan D Tardos. Orb-slam: a versatile and accurate monocular slam system. IEEE transactions on robotics, 31(5):1147-1163, 2015.", + "[54] Zachary Teed and Jia Deng. Droid-slam: Deep visual slam for monocular, stereo, and rgb-d cameras. NeurIPS, 34:16558-16569, 2021. 4", + "[55] Yuhang He, Irving Fang, Yiming Li, Rushi Bhavesh Shah, and Chen Feng. Metric-Free Exploration for Topological Mapping by Task and Motion Imitation in Feature Space. In RSS, July 2023. 5", + "[56] Mahmoud Assran, Quentin Duval, Ishan Misra, Piotr Bojanowski, Pascal Vincent, Michael Rabbat, Yann LeCun, and Nicolas Ballas. Self-supervised learning from images with a joint-embedding predictive architecture. In CVPR, pages 15619–15629, 2023." + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "6884", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[57] Adrien Bardes, Quentin Garrido, Jean Ponce, Xinlei Chen, Michael Rabbat, Yann LeCun, Mahmoud Assran, and Nicolas Ballas. Revisiting feature prediction for learning visual representations from video. arXiv preprint arXiv:2404.08471, 2024. 5", + "[58] Adarsh Jagan Sathyamoorthy, Kasun Weerakoon, Mohamed Elnoor, Anuj Zore, Brian Ichter, Fei Xia, Jie Tan, Wenhao Yu, and Dinesh Manocha. Convoi: Context-aware navigation using vision language models in outdoor and indoor environments. arXiv preprint arXiv:2403.15637, 2024. 5", + "[59] Wei Xu and Fu Zhang. Fast-lio: A fast, robust lidar-inertial odometry package by tightly-coupled iterated kalman filter. IEEE Robotics and Automation Letters, 6(2):3317-3324, 2021. 5", + "[60] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 5", + "[61] Noriaki Hirose, Dhruv Shah, Ajay Sridhar, and Sergey Levine. SACson: Scalable autonomous control for social navigation. IEEE Robotics and Automation Letters, 2023. 6" + ], + "bbox": [ + 91, + 92, + 480, + 375 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "6885", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/CityWalker_ Learning Embodied Urban Navigation from Web-Scale Videos/a9edaa72-0d38-4c4e-a25a-9480eb60d753_model.json b/2025/CityWalker_ Learning Embodied Urban Navigation from Web-Scale Videos/a9edaa72-0d38-4c4e-a25a-9480eb60d753_model.json new file mode 100644 index 0000000000000000000000000000000000000000..28f348b51dd9d44bc75cae4cb019123e0ee4ef51 --- /dev/null +++ b/2025/CityWalker_ Learning Embodied Urban Navigation from Web-Scale Videos/a9edaa72-0d38-4c4e-a25a-9480eb60d753_model.json @@ -0,0 +1,2466 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.118, + 0.131, + 0.882, + 0.153 + ], + "angle": 0, + "content": "CityWalker: Learning Embodied Urban Navigation from Web-Scale Videos" + }, + { + "type": "text", + "bbox": [ + 0.186, + 0.181, + 0.807, + 0.234 + ], + "angle": 0, + "content": "Xinhao Liu* Jintong Li* Yicheng Jiang Niranjan Sujay Zhicheng Yang \nJuexiao Zhang John Abanes Jing Zhang Chen Feng† \nNew York University" + }, + { + "type": "text", + "bbox": [ + 0.339, + 0.237, + 0.654, + 0.251 + ], + "angle": 0, + "content": "https://ai4ce.github.io/CityWalker/" + }, + { + "type": "image", + "bbox": [ + 0.096, + 0.264, + 0.299, + 0.46 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.303, + 0.263, + 0.412, + 0.46 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.414, + 0.267, + 0.535, + 0.46 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.535, + 0.267, + 0.655, + 0.46 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.656, + 0.268, + 0.778, + 0.46 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.779, + 0.268, + 0.898, + 0.46 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.467, + 0.908, + 0.524 + ], + "angle": 0, + "content": "Figure 1. Embodied Urban Navigation. Navigating urban spaces is challenging for (especially off-street) mobile agents. The differently colored pins \\((\\bullet)\\) along the route highlight various critical scenarios unique to complex and dynamic urban landscapes. Thumbsails on the right with corresponding colored pins demonstrate the real-world observation of these challenging cases. Our CityWalker model is trained with over 2000 hours of city walking videos and fine-tuned with a small amount of expert data to address these challenges effectively." + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.531, + 0.327, + 0.546 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.565, + 0.486, + 0.868 + ], + "angle": 0, + "content": "Navigating dynamic urban environments presents significant challenges for embodied agents, requiring advanced spatial reasoning and adherence to common-sense norms. Despite progress, existing visual navigation methods struggle in map-free or off-street settings, limiting the deployment of autonomous agents like last-mile delivery robots. To overcome these obstacles, we propose a scalable, data-driven approach for human-like urban navigation by training agents on thousands of hours of in-the-wild city walking and driving videos sourced from the web. We introduce a simple and scalable data processing pipeline that extracts action supervision from these videos, enabling large-scale imitation learning without costly annotations. Our model learns sophisticated navigation policies to handle diverse challenges and critical scenarios. Experimental results show that training on large-scale, diverse datasets significantly enhances navigation performance, surpassing current methods. This work shows the potential of using abundant online video data to develop robust navigation policies for embodied agents in dynamic urban settings." + }, + { + "type": "title", + "bbox": [ + 0.515, + 0.531, + 0.645, + 0.546 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.562, + 0.907, + 0.849 + ], + "angle": 0, + "content": "Visual navigation is a crucial capability for mobile agents. In previously unseen environments, humans typically rely on navigation tools (such as Google Maps) to reach a goal location by following a series of waypoints. These tools provide high-level guidance, but the actual navigation between waypoints requires sophisticated spatial awareness and decision-making. In urban settings, this involves understanding and adhering to complex rules, such as common sense regulations and norms, while dynamically responding to obstacles and environmental changes. The diversity and complexity encountered in urban navigation scenarios pose significant challenges for navigation policies but are essential for deploying mobile robots in real-world settings. Applications like delivery robots and robotaxis, especially without high-definition maps, highlight the growing need for embodied agents to navigate efficiently and safely in dynamic urban landscapes. How can we teach embodied agents to navigate dynamic urban landscapes as effectively as humans do?" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.856, + 0.909, + 0.903 + ], + "angle": 0, + "content": "The question has been explored in early works [1-3] before the deep learning era. These approaches combined SLAM-based mapping with modular systems for search" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.875, + 0.478, + 0.888 + ], + "angle": 0, + "content": "*Equal contribution ( {xinhao.liu, jintong.li}@nyu.edu)" + }, + { + "type": "page_footnote", + "bbox": [ + 0.111, + 0.888, + 0.348, + 0.9 + ], + "angle": 0, + "content": "†Corresponding author (cfeng@nyu.edu)" + }, + { + "type": "list", + "bbox": [ + 0.109, + 0.875, + 0.478, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6875" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.243 + ], + "angle": 0, + "content": "ing, detection, analysis, and planning. The development of high-performance photorealistic simulators [4-7] significantly advanced research in visual navigation by enabling reinforcement learning (RL) in static indoor environments, achieving near-perfect point-goal navigation [8]. This simulation-RL paradigm was also extended to more complex settings, including real-world panoramic street views [9-12]. Recent works begin to explore imitation learning from expert demonstrations for long-horizon or cross-embodiment navigation tasks [13, 14]." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.246, + 0.485, + 0.548 + ], + "angle": 0, + "content": "Despite their success in simulators or moderately complex environments, existing methods fall short in addressing the complexity of embodied urban navigation. Public urban spaces are inherently dynamic and unpredictable, characterized by \"multifarious terrains, diverse obstacles, and dense pedestrians\" that require real-time adaptability [15]. Moreover, effective navigation in such settings demands adherence to common sense navigation rules and social norms, such as using sidewalks, obeying traffic signals, and maintaining appropriate personal space to avoid conflicts [16-18]. Current reinforcement learning and imitation learning approaches typically excel in static or controlled environments but struggle with real-world urban navigation. This is because these nuanced behaviors and delicate constraints are challenging to incorporate into simulation environments while ensuring reliable and generalizable sim-to-real transfer. Moreover, coverage of these complexities is still lacking in existing teleoperation demonstration datasets. As a result, embodied urban navigation remains an unsolved problem." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.551, + 0.483, + 0.747 + ], + "angle": 0, + "content": "While a straightforward approach to the problem is to collect expert trajectories through teleoperation, this is limited by data size and lack of diversity, which weakens the agent's ability to generalize across various urban scenarios. Is there a data-driven approach that requires minimum annotation but maximizes generalizability? Inspired by the success of scaling laws in language, vision, and robotics tasks [19-22], we propose a scalable framework, CityWalker, which leverages web-scale city walking and driving videos to train models for embodied urban navigation. Our model is trained on over two thousand hours of internet-sourced videos across different geolocations, weather conditions, and times of day." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.902 + ], + "angle": 0, + "content": "The subsequent question is how to effectively obtain action supervision from in-the-wild videos for imitation learning. Current solutions rely heavily on prompting proprietary VLMs to generate action labels from videos [23, 24], which is costly and challenging to control, making it difficult to scale for large datasets. We argue that noisy pseudo labels from off-the-shelf visual odometry (VO) models [25] are sufficiently effective for imitation learning in embodied urban navigation. Moreover, this data processing method generalizes beyond walking data. We demonstrate that" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.152 + ], + "angle": 0, + "content": "a model trained with driving videos also performs well on quadruped agents, with further improvements observed when training in a cross-domain setting that combines both city walking and driving data." + }, + { + "type": "text", + "bbox": [ + 0.534, + 0.153, + 0.857, + 0.167 + ], + "angle": 0, + "content": "We summarize our contribution as the following:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.169, + 0.905, + 0.229 + ], + "angle": 0, + "content": "- We identify embodied urban navigation as an intriguing yet unsolved problem. To address this problem, we propose a scalable, data-driven solution by leveraging web-scale city walking and driving videos." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.23, + 0.905, + 0.274 + ], + "angle": 0, + "content": "- We introduce a simple yet scalable data processing paradigm, making large-scale imitation learning feasible for urban navigation without extensive manual labeling." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.275, + 0.905, + 0.35 + ], + "angle": 0, + "content": "- We demonstrate training on web-scale data significantly improves navigation performance in real-world experiments, allowing embodied agents to handle the complexities of urban scenarios. We also show model performance scales well with training data size." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.169, + 0.905, + 0.35 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.363, + 0.655, + 0.379 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.387, + 0.905, + 0.659 + ], + "angle": 0, + "content": "Navigation in Simulation. The introduction of high-performance photorealistic simulators [4-7] has significantly boosted navigation research in indoor household environments. Based on the form of goal specification, navigation tasks are usually categorized into point-goal [4, 8, 26, 27], image-goal [28-31], object-goal [32-34], and vision-language navigation [35-37]. While tasks like image-goal and object-goal navigation focus on efficient goal-finding or instruction understanding, point-goal navigation has a more straightforward goal specification and emphasizes safe and efficient trajectory planning to the goal. Our problem formulation resembles point-goal navigation in that accurate waypoints (represented as GPS coordinates) are readily available from navigation tools. Recently, point-goal navigation has been considered a \"solved\" problem due to near-perfect performance achieved in simulators [8]. However, our investigation shows that waypoint-goal navigation in public urban spaces remains an unsolved challenge." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.66, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Real-World Navigation. Real-world robot navigation has a long history, predating the deep learning era. Early methods relied on ranging sensors and modular SLAM systems for navigation [1-3]. Witnessing the success in simulators, various methods have been proposed to bridge the sim-to-real gap [38-41]. Another thread of work learns navigation policies from expert trajectories [42, 43]. By incorporating data from different embodiments [14] and carefully designed heuristics [13], these methods have been extended to long-horizon navigation on different robots [44, 45]. While deployed on real-world embodied agents, these methods still operate in relatively simple and static environments such as suburban areas, parking garages, or controlled obstacle courses. They are not guaranteed to generalize to significantly different, dynamic, and complex urban landscapes. Our work follows this line of research by training naviga" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6876" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.139, + 0.092, + 0.859, + 0.382 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.392, + 0.907, + 0.462 + ], + "angle": 0, + "content": "Figure 2. Overall Illustration of CityWalker. Our training pipeline starts with internet-sourced videos, using visual odometry to obtain relative poses between frames. At each time step, the model receives past observations, past trajectory, and target location as input. They are encoded via a frozen image encoder and a trainable coordinate encoder. A transformer processes these inputs to generate future tokens. An action head and an arrival head decode these tokens into action and arrival status predictions. During training, future frame tokens from future frames guide the transformer to hallucinate future tokens." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.474, + 0.483, + 0.534 + ], + "angle": 0, + "content": "tion models with imitation learning. But we focus on learning the rules and norms essential for urban environments, which is crucial for enabling applications such as delivery robots and robot-taxis." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.536, + 0.483, + 0.807 + ], + "angle": 0, + "content": "In-the-Wild Video Learning. Learning from web-scale data has proven successful in language and vision tasks [19-21, 46]. While some works utilize in-the-wild videos to extract intermediate representations, priors, or reward functions [47, 48], a key challenge remains: the lack of action labels. Patel et al. [49] use a reconstruction pipeline to obtain hand-object trajectories as action labels for learning robot manipulation. SelfD [50] argues BEV model is better than VO in pseudo-labeling, while we show VO works well with large-scale data. Recent works also rely on prompting closed VLMs to get action labels [23, 51]. LeLaN [24], a concurrent work most relevant to ours, also relies on VLM prompting and uses pretrained navigation models to generate action labels. In contrast, we process in-the-wild video data differently by using only the noisy pseudolabels from off-the-shelf visual SLAM tools [25]. This approach enables scalable and cost-effective action label generation for imitation learning in urban navigation tasks." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.823, + 0.357, + 0.84 + ], + "angle": 0, + "content": "3. Embodied Urban Navigation" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.848, + 0.295, + 0.863 + ], + "angle": 0, + "content": "3.1. Problem Formulation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.901 + ], + "angle": 0, + "content": "In this work, we address the problem of embodied urban navigation for agents in dynamic urban environments. The" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.474, + 0.907, + 0.655 + ], + "angle": 0, + "content": "agent's task is to navigate from its current location to a specified target waypoint location, following a series of waypoints provided by a navigation tool. Formally, we define this as a point-goal navigation problem in real-world urban settings. At each time step \\( t \\), the agent receives an RGB observation \\( o_t \\), its current GPS location \\( p_t \\), and a sub-goal waypoint \\( w_t \\). The agent aims to learn a policy \\( \\pi(a_t \\mid o_{(t-k):t}, p_{(t-k):t}, w_t) \\) that maps the past observations and positional information to an action \\( a_t \\) from the action space \\( A \\), represented as a series of action waypoints in Euclidean space. Typically, we take \\( k = 5 \\) and also predict actions for the next 5 time steps." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.656, + 0.909, + 0.779 + ], + "angle": 0, + "content": "As accurate waypoints are easily available from navigation tools (e.g. Google Maps API), we focus our problem on the navigation between two consecutive waypoints. The agent treats the current sub-goal \\( w_{t} \\) as the immediate target and is responsible for determining when it has reached this sub-goal based on its observations and positional data. Upon reaching \\( w_{t} \\), the agent proceeds to the next waypoint in the sequence." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.788, + 0.698, + 0.803 + ], + "angle": 0, + "content": "3.2. Evaluation Metrics" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.907, + 0.903 + ], + "angle": 0, + "content": "To enable analysis and assessment of different methods on the problem, our evaluation metrics are twofold. We use the ground truth trajectory in teleoperation-collected data to provide a thorough analysis of each method's performance across various scenarios. Below, we discuss the metrics we used in offline data evaluation. For real-world deployment," + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6877" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.082, + 0.287, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.287, + 0.082, + 0.48, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.244, + 0.485, + 0.342 + ], + "angle": 0, + "content": "Figure 3. Evaluation Metrics. Left: The orientation error is defined to be the angle between each predicted and ground truth action pair, labeled by \\(\\theta_{n}\\) in the figure. The action angle \\(\\varphi_{\\mathrm{action}}\\) and target angle \\(\\varphi_{\\mathrm{target}}\\) are defined with respect to the positive y-axis. Right: Both green and red trajectories are predicted actions. The green trajectory is the preferred one, having a large L2 distance but a small AOE. Vice versa for the red trajectory." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.354, + 0.483, + 0.384 + ], + "angle": 0, + "content": "we primarily use navigation success rate as the key metric for performance comparison." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.384, + 0.484, + 0.581 + ], + "angle": 0, + "content": "Average Orientation Error. We use the error between each predicted action and the ground truth action as the main evaluation metric on the offline teleoperation data. While L2-distance in Euclidean space is a straightforward metric, we find it does not adequately capture the problem. As shown on the right side of Fig. 3, the red trajectory has a smaller L2 error compared to the green trajectory but moves in a direction that deviates from the target. This observation motivates us to come up with a new evaluation metric that better reflects the quality of the predicted actions. We define average orientation error (AOE) as the angle difference between the predicted and ground truth actions, averaged across all samples at each step (left side of Fig. 3):" + }, + { + "type": "equation", + "bbox": [ + 0.108, + 0.592, + 0.483, + 0.632 + ], + "angle": 0, + "content": "\\[\n\\operatorname {A O E} (k) = \\frac {1}{n} \\sum_ {i} ^ {n} \\theta_ {i _ {k}} = \\frac {1}{n} \\sum_ {i} ^ {n} \\arccos \\frac {\\left\\langle a _ {\\hat {i} _ {k}} ^ {\\prime} , a _ {i _ {k}} \\right\\rangle}{\\| a _ {\\hat {i} _ {k}} \\| \\| a _ {i _ {k}} \\|}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.643, + 0.484, + 0.734 + ], + "angle": 0, + "content": "where \\( k \\) is the index of predicted action and \\( n \\) is the number of data. Additionally, we observe that taking the mean across \\( k \\) predicted future time steps can lead to an underestimation due to steps with very small errors. Hence, we propose using the maximum average orientation error (MAOE) over all predicted actions as an overall estimation:" + }, + { + "type": "equation", + "bbox": [ + 0.199, + 0.744, + 0.483, + 0.784 + ], + "angle": 0, + "content": "\\[\n\\mathrm {M A O E} = \\frac {1}{n} \\sum_ {i} ^ {n} \\max _ {k} \\theta_ {i _ {k}}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.483, + 0.839 + ], + "angle": 0, + "content": "As orientation error might be noisy when action distance is small, we also use L2-distance as a complimentary evaluation metric." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Critical Scenarios. Not all time steps are equally critical in a trajectory. For example, errors at road crossings or turning points are more likely to result in navigation failure compared to mistakes during forward movement in open areas." + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.074, + 0.909, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.198, + 0.907, + 0.268 + ], + "angle": 0, + "content": "Figure 4. Data Sample and Visual Odometry (VO) Result. Our internet-source training data includes both walking and driving videos. These videos cover various scenarios in the urban environment. The VO tool gives noisy trajectories globally, but trustworthy local relative pose within a short time period." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.28, + 0.906, + 0.356 + ], + "angle": 0, + "content": "We identify several key scenarios that are most critical during navigation. AOE and MAOE are calculated separately for each scenario, and we use the scenario mean rather than the sample mean for a more comprehensive evaluation. Detailed definitions of each scenario are provided in Sec. 4." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.365, + 0.816, + 0.381 + ], + "angle": 0, + "content": "3.3. Learning from In-the-Wild Videos" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.387, + 0.907, + 0.613 + ], + "angle": 0, + "content": "City Walking Videos. To facilitate large-scale imitation learning for human-like urban navigation, we utilize city walking videos sourced from the internet. Walking in urban environments is intuitive for humans, resulting in an abundance of first-person view videos available online. These videos naturally capture the complexities of navigating dynamic urban landscapes, including interactions with pedestrians, adherence to traffic signals, and maneuvering around various obstacles. Motivated by the extensive availability of such data and its inherent alignment with our navigation objectives, we curate a dataset comprising over 2000 hours of city walking videos, encompassing diverse weather and lighting conditions. These videos exhibit consistent motion patterns characteristic of walking, providing continuous motion cues that can directly serve as action labels." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.614, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Action Labels from Videos. To generate actionable labels for imitation learning, we employ VO [25, 52-54] tools to extract trajectory poses from the collected city walking videos. Specifically, we utilize the state-of-the-art DPVO [25] while any other VO method should also work. The resulting VO trajectories are illustrated in Fig. 4. However, using VO poses presents two primary challenges. First, global trajectory accuracy is compromised by accumulation errors inherent in VO systems. Nevertheless, since our model focuses on predicting a limited sequence of actions (5 past and 5 future), we mitigate this issue by relying solely on relative poses within short temporal windows. Second, VO methods suffer from scale ambiguity, leading to inconsistencies in trajectory lengths across different data. For instance, as shown in Fig. 4, a driving video may exhibit a longer actual trajectory compared to a walking video, yet the DPVO output incorrectly reflects a shorter trajectory. To address this, we normalize each action by the average step length within its trajectory, following the approach in" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "6878" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.198 + ], + "angle": 0, + "content": "previous work [14]. This normalization resolves both scale ambiguity and inconsistencies in step lengths between different types of embodiments all at once. This enables the learning of policies in a consistent, abstracted action space. During deployment, the predicted actions are denormalized using the specific step length of the target robot, ensuring appropriate movement execution in real-world settings." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.199, + 0.483, + 0.439 + ], + "angle": 0, + "content": "Generalizability and Scalability. Our data processing pipeline is both generalizable and scalable. It can be applied to any video sources beyond city walking, as long as the videos are captured with egomotion, such as driving videos, which are also abundantly available online. This flexibility enables the training of more generalizable navigation policies that can be zero-shot applied or fine-tuned for cross-domain and cross-embediment tasks, such as transitioning from pedestrian to vehicular navigation. Furthermore, our method offers significant scalability advantages. Unlike pipelines that rely on prompting large language models for action label generation [23, 24], our visual odometry-based approach can be easily parallelized. Processing 2000 hours of video data with our pipeline requires negligible wall-clock time, making it both feasible and cost-efficient to scale up the training dataset." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.447, + 0.295, + 0.463 + ], + "angle": 0, + "content": "3.4. Pipeline and Training" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.469, + 0.483, + 0.62 + ], + "angle": 0, + "content": "Pipeline Architecture. Figure 2 shows our pipeline. It builds upon previous works [44, 45]. At the core of our model is a transformer that processes a temporal sequence of input tokens, including \\( k \\) image features and one coordinate embeddings. Specifically, the coordinate input is stacked by \\( k \\) past positions and one single target position, and outputs a sequence of the same length. We use an action head and an arrival prediction head to decode the output of the transformer. We utilize DINOv2 [21] representation and freeze the backbone during training." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.62, + 0.483, + 0.74 + ], + "angle": 0, + "content": "Feature Hallucination. Inspired by previous works on feature learning [55-57], we use feature hallucination as an auxiliary loss during training. Specifically, we compute the MSE loss between the output image tokens and that directly extracted from future frames. We hope this will guide the transformer to predict more informative future tokens that best mimic future frames, which in turn helps the two MLP heads predict the final action and arrival status." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.74, + 0.483, + 0.831 + ], + "angle": 0, + "content": "Loss Functions. As mentioned in Sec. 3.2, L1 or L2 loss can be less effective than orientation error in evaluating the quality of the predicted actions. We also employ orientation error to supervise the training. We define the orientation loss to be the negative cosine similarity between the predicted actions and the ground truth actions:" + }, + { + "type": "equation", + "bbox": [ + 0.199, + 0.839, + 0.483, + 0.88 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {o r i}} = - \\frac {1}{k} \\sum_ {i = 1} ^ {k} \\frac {\\left\\langle \\hat {a} _ {i} , a _ {i} \\right\\rangle}{\\left\\| \\hat {a} _ {i} \\right\\| \\left\\| a _ {i} \\right\\|}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.886, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Together with the L1 loss for predicted action and the BCE" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.907, + 0.122 + ], + "angle": 0, + "content": "loss for predicted arriving status, the final loss function is a weight sum of four individual losses:" + }, + { + "type": "equation", + "bbox": [ + 0.54, + 0.134, + 0.907, + 0.151 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\omega_ {l 1} \\mathcal {L} _ {l 1} + \\omega_ {\\text {o r i}} \\mathcal {L} _ {\\text {o r i}} + \\omega_ {\\text {a r r}} \\mathcal {L} _ {\\text {a r r}} + \\omega_ {\\text {f e a t}} \\mathcal {L} _ {\\text {f e a t}}. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.162, + 0.907, + 0.192 + ], + "angle": 0, + "content": "We choose each weight so that each loss items are within the same magnitude." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.206, + 0.64, + 0.223 + ], + "angle": 0, + "content": "4. Experiment" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.23, + 0.597, + 0.246 + ], + "angle": 0, + "content": "4.1. Setup" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.252, + 0.907, + 0.418 + ], + "angle": 0, + "content": "Baselines. We compare our model with outdoor navigation models closely related to our setup, including GNM [14], ViNT [44], and NoMaD [45]. Although originally developed for image-goal navigation, we tested them using goal-images from our collected data. Additionally, ViNT as a foundation model for visual navigation supports finetuning, which we also evaluated. We acknowledge CoNVOI [58] as a recent work with a similar setup, but could not test it due to the lack of open source code and prompt. We also experimented with VLM-based approaches similar to CoNVOI and defer the results to the appendix." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.419, + 0.906, + 0.585 + ], + "angle": 0, + "content": "Data. We collected expert data via teleoperation for fine-tuning and offline testing. We collect data via a Unitree Go1 quadruped equipped with a Livox Mid-360 LiDAR and a webcam for RGB observations. We employed the LiDAR-SLAM method [59] to obtain the robot's poses as ground truth actions. Additionally, we retrieved location data from a smartphone using a webpage. We reduce noise from the quadruped's vibrations by holding the phone [38]. In total, we gathered 15 hours of teleoperation data across various areas in New York City, allocating 6 hours for fine-tuning and 9 hours for testing." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.585, + 0.906, + 0.659 + ], + "angle": 0, + "content": "Critical Scenarios As discussed in Sec. 3.2, we identify several critical scenarios in our dataset that we want to emphasize in our evaluation. We analyze the ground truth trajectory and run object detection [60] to aid us in standardizing the definition. We define these scenarios as follows:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.66, + 0.905, + 0.689 + ], + "angle": 0, + "content": "- Turn: When the ground truth action changes direction significantly. Defined when \\(\\varphi_{\\mathrm{action}} > 20^{\\circ}\\) (see Fig. 3)." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.69, + 0.905, + 0.719 + ], + "angle": 0, + "content": "- Crossing: When the agent is at a road crossing. Defined if any traffic light is detected with score \\(>0.5\\)." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.72, + 0.905, + 0.751 + ], + "angle": 0, + "content": "- Detour: When the action angle deviates from the target angle. Defined when \\( |\\varphi_{\\text{action}} - \\varphi_{\\text{target}}| > 45^{\\circ} \\) (see Fig. 3)." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.751, + 0.905, + 0.795 + ], + "angle": 0, + "content": "- Proximity: When any person is close to the agent. Defined when the largest bounding box of any detected person is larger than \\(25\\%\\) of the image area." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.796, + 0.905, + 0.825 + ], + "angle": 0, + "content": "- **Crowd:** When the agent is surrounded by a crowd of people. Defined when \\(\\geq 5\\) people are detected." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.66, + 0.905, + 0.825 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.906, + 0.901 + ], + "angle": 0, + "content": "We define these scenarios in a way that they are not mutually exclusive. A single data sample can belong to multiple scenarios. Although these scenarios account for less than half of the data, we argue they are the most significant factors contributing to successful urban navigation." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6879" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.907, + 0.133 + ], + "angle": 0, + "content": "Table 1. Benchmark on Offline Data. We evaluate three metrics in each critical scenario for all methods. Percentages under scenarios indicate their data proportions. The \"Mean\" column shows scenario means averaged over six scenarios; \"All\" shows sample means over all data samples. We highlight the best performance in bold and the second-best performance with underline." + }, + { + "type": "table", + "bbox": [ + 0.176, + 0.143, + 0.825, + 0.401 + ], + "angle": 0, + "content": "
MethodMetricMeanTurn 8%Crossing 12%Detour 12%Proximity 6%Crowd 7%Other 55%All 100%
GNM [14] (fine-tuned)↓L2 (m)1.222.361.361.420.880.760.550.74
↓MAOE (°)16.231.114.812.514.712.811.012.1
↑Arrival (%)68.666.469.766.469.069.270.770.0
ViNT [44] (fine-tuned)↓L2 (m)1.301.911.131.140.770.660.570.70
↓MAOE (°)16.531.115.412.914.813.311.612.6
↑Arrival (%)70.571.268.770.773.068.671.070.7
NoMaD [45] (fine-tuned)↓L2 (m)1.392.491.561.551.060.950.760.74
↓MAOE (°)19.135.118.515.618.114.312.812.1
↑Arrival (%)68.666.469.766.469.069.370.770.0
Ours (zero-shot)↓L2 (m)1.341.301.091.331.441.481.391.38
↓MAOE (°)16.526.815.516.316.313.111.312.7
↑Arrival (%)79.169.371.478.884.184.586.484.1
Ours (fine-tuned)↓L2 (m)1.111.271.001.151.061.121.061.07
↓MAOE (°)15.226.614.113.914.312.010.411.5
↑Arrival (%)81.868.975.378.590.687.590.287.8
" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.424, + 0.345, + 0.44 + ], + "angle": 0, + "content": "4.2. Performance Benchmarking" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.447, + 0.483, + 0.522 + ], + "angle": 0, + "content": "In this subsection, we aim to answer the question Q1: Can our CityWalker model navigate successfully in complex urban environments? To answer this, we benchmark our model and the baselines both with offline data and in real-world navigation deployment." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.523, + 0.483, + 0.705 + ], + "angle": 0, + "content": "Overall Benchmark. Table 1 shows the benchmark in different critical scenarios. Overall, our fine-tuned model has superior performance on all metrics in all scenarios except \"detour\". We think the less competent performance in the detour scenario is due to the small portion of such data in our video training data. It can be verified by the fact that the fine-tuning performance gain is most significant in this scenario. It is worth noting that our zero-shot model has similar or even better performance compared to the fine-tuned baselines. This demonstrates the power of large-scale pretraining, especially considering the domain and modality gap from human walking videos." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Real-world Depolyment. We deploy our model and the baseline methods on the same Unitree Go1 quadruped for real-world navigation, in previously unseen urban environments. We adapt the PD controller provided in [44] to provide velocity command to the quadruped. We choose area with relatively constant density of dynamics such as pedestrian and traffic to ensure the consistency among different trials. In each trial, the target location is generally 50-100m far away from the starting location. We categorize the real-world test into forward, left turn, and right turn cases to provide case-by-case analysis. We conduct 8-14 trails for each cases. Each trial is marked success when the model predicts arrival within a 5m distance to the target location." + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.423, + 0.907, + 0.465 + ], + "angle": 0, + "content": "Table 2. Real-World Navigation. The table shows the success rate of real-world experiments in different scenarios. * and † indicate zero-shot inference and fine-tuned models respectively." + }, + { + "type": "table", + "bbox": [ + 0.536, + 0.469, + 0.882, + 0.553 + ], + "angle": 0, + "content": "
MethodAllForwardLeft turnRight turn
ViNT* [44]37.762.50.050.0
ViNT† [44]57.1100.025.025.0
NoMad* [45]42.975.016.728.6
Ours†77.3100.062.566.7
" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.564, + 0.905, + 0.594 + ], + "angle": 0, + "content": "All human interruptions due to potential collision or timeout are treated as failure cases." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.595, + 0.907, + 0.837 + ], + "angle": 0, + "content": "Table 2 shows the success rate on all real-world test cases. Our model achieves the highest success rate among all cases with a significant gap to the second-best performing finetuned ViNT. As the baselines are trained with navigation data mostly in suburban or off-road environments [14, 61], they are limited in handling complex maneuvers, making them unsuitable for urban navigation tasks, as shown in Fig. 5. Our model's robust performance in both forward and turns highlights its ability to manage dynamic and varied urban scenarios effectively, allowing it to be more reliably used in real-world environments where frequent and precise directional changes are essential. The substantial improvement not only increases the overall navigation success but also extends the applicability of our model to more demanding and realistic urban settings, positioning it as a superior solution for embodied agents in dynamic environments." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.848, + 0.722, + 0.865 + ], + "angle": 0, + "content": "4.3. Power of Data Scaling" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.906, + 0.903 + ], + "angle": 0, + "content": "Our model is significantly boosted by the large-scale training data. We ask Q2: To what extent does increasing the" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "6880" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.123, + 0.092, + 0.258, + 0.104 + ], + "angle": 0, + "content": "Current Observation" + }, + { + "type": "image", + "bbox": [ + 0.117, + 0.113, + 0.268, + 0.231 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.316, + 0.092, + 0.414, + 0.103 + ], + "angle": 0, + "content": "ViNT Zero-shot" + }, + { + "type": "image", + "bbox": [ + 0.275, + 0.109, + 0.425, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.339, + 0.231, + 0.385, + 0.244 + ], + "angle": 0, + "content": "X (m) Coordinates" + }, + { + "type": "image_caption", + "bbox": [ + 0.462, + 0.092, + 0.56, + 0.103 + ], + "angle": 0, + "content": "ViNT Fine-tune" + }, + { + "type": "image", + "bbox": [ + 0.426, + 0.107, + 0.573, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.491, + 0.231, + 0.535, + 0.244 + ], + "angle": 0, + "content": "X (m) Coordinates" + }, + { + "type": "image_caption", + "bbox": [ + 0.608, + 0.092, + 0.723, + 0.103 + ], + "angle": 0, + "content": "NoMad Zero-shot" + }, + { + "type": "image", + "bbox": [ + 0.575, + 0.108, + 0.727, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.64, + 0.231, + 0.687, + 0.244 + ], + "angle": 0, + "content": "X (m) Coordinates" + }, + { + "type": "image_caption", + "bbox": [ + 0.797, + 0.092, + 0.831, + 0.104 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.731, + 0.108, + 0.877, + 0.236 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.117, + 0.244, + 0.27, + 0.36 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.233, + 0.371, + 0.321, + 0.381 + ], + "angle": 0, + "content": "Input past traj." + }, + { + "type": "image", + "bbox": [ + 0.278, + 0.244, + 0.424, + 0.367 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.359, + 0.371, + 0.452, + 0.38 + ], + "angle": 0, + "content": "Actual past traj." + }, + { + "type": "image", + "bbox": [ + 0.429, + 0.244, + 0.575, + 0.367 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.489, + 0.371, + 0.586, + 0.38 + ], + "angle": 0, + "content": "Predicted action" + }, + { + "type": "image", + "bbox": [ + 0.579, + 0.244, + 0.726, + 0.368 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.624, + 0.371, + 0.688, + 0.38 + ], + "angle": 0, + "content": "GTaction" + }, + { + "type": "image", + "bbox": [ + 0.73, + 0.244, + 0.877, + 0.367 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.726, + 0.371, + 0.771, + 0.381 + ], + "angle": 0, + "content": "\\(\\star\\) Target" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.388, + 0.907, + 0.416 + ], + "angle": 0, + "content": "Figure 5. Qualitative Results. Left image shows current observations of two samples. Right plots displays input trajectory, ground truth actions, and predicted actions in the current coordinate system with the agent at the origin." + }, + { + "type": "image", + "bbox": [ + 0.109, + 0.434, + 0.455, + 0.596 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.602, + 0.484, + 0.673 + ], + "angle": 0, + "content": "Figure 6. Performance and Data Size. We show the model performance evaluated by MAOE with respect to the size of the training data measured by video length in hours. We also show the zero-shot performance of our model trained with only driving videos and mixed driving and walking videos" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.7, + 0.484, + 0.731 + ], + "angle": 0, + "content": "size of the training dataset improve the performance of our urban navigation model?" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Performance Gain and Data Scale. Figure 6 shows the model performance when trained with different data sizes. The plot demonstrates that our model's zero-shot performance improves significantly as we increase the amount of training data. Notably, when trained with more than 1000 hours of videos, our zero-shot model (red line) achieves a better performance than the fine-tuned ViNT model \\((\\square)\\). This indicates that large-scale training on diverse video data can even surpass the performance of models that are fine-tuned with limited expert data, highlighting the potential of leveraging web-scale datasets for embodied navigation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.432, + 0.907, + 0.673 + ], + "angle": 0, + "content": "Domain and Embodiment Gap. An unexpected observation is that our model without the feature hallucination loss outperforms the one with it in zero-shot inference. We hypothesize this phenomenon mainly due the domain and embodiment gap between human walking videos and quadruped navigation policies. The feature hallucination loss aims to guide the transformer model to produce output tokens that mimic future observations, essentially encouraging the model to anticipate future visual inputs. This enforcement might inadvertently hinder performance by focusing on predicting future observations based on human motion. The model may learn representations that are less applicable or even misleading for quadruped navigation. Consequently, the feature hallucination loss could introduce a bias that detracts from the model's ability to make accurate navigation decisions in the robot's operational domain." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.674, + 0.909, + 0.885 + ], + "angle": 0, + "content": "Feature Hallucination Loss. Fortunately, as shown in Tab. 1, this issue no longer exists after fine-tuning. We also show the benefit of feature hallucination loss from another perspective. Figure 7 shows two training loss curve for different data size and training setups with and without feature hallucination loss. We observe that although feature hallucination loss should increase the absolute value of the total loss, the training with feature hallucination loss lead to a lower total loss for 2000-hour training and a lower orientation loss in both cases. The training with feature hallucination loss also shows a steeper decreasing trend, showing a potential of having even lower error with more training steps. Therefore, this is another proof of the effectiveness of our proposed feature hallucination loss." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.886, + 0.905, + 0.902 + ], + "angle": 0, + "content": "Driving Videos. As discussed in Sec. 3.3, our data process" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.514, + 0.957 + ], + "angle": 0, + "content": "6881" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.119, + 0.094, + 0.445, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.261, + 0.485, + 0.346 + ], + "angle": 0, + "content": "Figure 7. Improvement from Feature Hallucination. The plot shows the orientation loss for different amounts of training data and loss setups. For all four setups, we trian with \\(\\omega_{\\mathrm{II}} = 1.0, \\omega_{\\mathrm{arr}} = 1.0, \\text{and}, \\omega_{\\mathrm{ori}} = 5.0\\). The unproportional training steps are due to different batch sizes used for different data sizes. All models are trained for 10 epochs." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.366, + 0.483, + 0.562 + ], + "angle": 0, + "content": "ing pipeline also enables the use of abundant online driving videos. We evaluate models trained exclusively on driving videos and on a combination of walking and driving videos, results shown in Fig. 6. We can observe that the model trained only with driving data \\((+)\\) has a similar performance to the zero-shot baselines. As a result, we can think of driving as a form of navigation from a different domain and embodiment. However, combining driving and walking data yields significant performance gains (x). With just 250 hours of mixed data, the model nearly matches the performance of models trained on 1000 hours of walking data alone. This signifies the generalizability and robustness provided by cross-domain cross-embodiment data." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.575, + 0.345, + 0.592 + ], + "angle": 0, + "content": "4.4. Pipeline component analysis" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.598, + 0.483, + 0.659 + ], + "angle": 0, + "content": "To provide a more in-depth analysis of our model and pipeline, we ask Q3: How do the individual components of our pipeline enhance the performance and reliability of our model for real-world urban navigation?" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.66, + 0.483, + 0.901 + ], + "angle": 0, + "content": "AOE at Different Time Steps. In Tab. 4, we study the AOE at different time steps, i.e. \\(\\mathrm{AOE}(k)\\) for \\(k\\) from 1 to 5. ViNT exhibits increasing orientation errors as the prediction steps progress, indicating a degradation in accuracy over time. In contrast, our model maintains relatively consistent AOE from step 2 to step 5, which likely contributes to its superior performance in real-world navigation deployments with stable action prediction across multiple future steps. All methods experience higher errors at the first step. We attribute it to the proximity of the initial action to the origin, as minor linear deviations result in substantial angular error. It is worth noting that the MAOE is significantly larger than any individual AOE from steps 1 to 5, suggesting that the peak errors occur at varying time steps across different samples. This variability suggests the potential benefit of taking softmax for orientation errors instead of simple averaging, as it" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.09, + 0.905, + 0.146 + ], + "angle": 0, + "content": "Table 3. Ablation Study. We ablate on the contribution of orientation loss, feature hallucination, and fine-tuning with expert data. We show the MAOE(°) in scenario-mean. All results are trained with 1000-hour walking data." + }, + { + "type": "table", + "bbox": [ + 0.558, + 0.157, + 0.865, + 0.277 + ], + "angle": 0, + "content": "
Training ComponentsMAOE
Ori. LossFeature Hall.Fine-tuningMean
17.03
✓17.00
✓✓17.02
✓15.23
✓✓15.21
✓✓✓15.16
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.298, + 0.907, + 0.342 + ], + "angle": 0, + "content": "Table 4. AOE at Different Time Steps. The table shows MAOE and different AOE of scenario-mean, all in degrees \\((^{\\circ})\\). * and † indicates zero-shot inference and fine-tuned models respectively." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.352, + 0.905, + 0.426 + ], + "angle": 0, + "content": "
MethodAOE(1)AOE(2)AOE(3)AOE(4)AOE(5)MAOE
ViNT*[44]10.577.939.0910.2111.4517.52
ViNT†[44]11.247.968.408.939.8716.53
Ours*12.918.928.648.468.4316.54
Ours†11.447.777.687.757.9715.23
" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.454, + 0.851, + 0.469 + ], + "angle": 0, + "content": "could better capture the distribution of peak errors." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.474, + 0.907, + 0.596 + ], + "angle": 0, + "content": "Ablation Study. Table 3 examines the individual contributions of different pipeline components to our model's performance. A clear improvement in MAOE when fine-tuning with expert data, demonstrating the effectiveness of using in-domain expert information. The improvement from orientation loss and feature hallucination loss is marginal, within error range of our limited testing data. We defer more detailed ablation study to the appendix." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.624, + 0.634, + 0.64 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.655, + 0.907, + 0.806 + ], + "angle": 0, + "content": "Summary. In this work, we addressed embodied urban navigation by training agents with thousands of hours of in-the-wild city walking videos, introducing a scalable pipeline for imitation learning from diverse walking videos. Our experiments show that large-scale, diverse training data significantly enhances navigation performance, surpassing existing methods and demonstrating the benefits of data scaling. Our findings highlight the potential of leveraging abundant online videos to develop robust navigation policies for embodied agents in dynamic urban environments." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.909, + 0.9 + ], + "angle": 0, + "content": "Limitation and Future Work. During real-world deployment, our current system implementation is sensitive to potential large location noises read naively from iPhone location services, which could be improved with better GPS hardware. Future work could also enhance the robustness of the model to such location noises." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "6882" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.241, + 0.108 + ], + "angle": 0, + "content": "Acknowledgment" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.115, + 0.484, + 0.192 + ], + "angle": 0, + "content": "The work was supported by NSF grants 2238968, 2121391, 2322242 and 2345139; and in part through the NYU IT High Performance Computing resources, services, and staff expertise. We also thank Xingyu Liu and Zixuan Hu for their help in data collection." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.204, + 0.188, + 0.219 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.228, + 0.484, + 0.284 + ], + "angle": 0, + "content": "[1] Quirin Muhlbauer, Stefan Sosnowski, Tingting Xu, Tianguang Zhang, Kolja Kuhnlenz, and Martin Buss. Navigation through urban environments by visual perception and interaction. In ICRA, pages 3558-3564. IEEE, 2009. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.285, + 0.484, + 0.34 + ], + "angle": 0, + "content": "[2] Rainer Kümmerle, Michael Ruhnke, Bastian Steder, Cyril Stachniss, and Wolfram Burgard. A navigation system for robots operating in crowded urban environments. In ICRA, pages 3225-3232. IEEE, 2013." + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.342, + 0.483, + 0.396 + ], + "angle": 0, + "content": "[3] Yoichi Morales, Alexander Carballo, Eijiro Takeuchi, Atsushi Aburadani, and Takashi Tsubouchi. Autonomous robot navigation in outdoor cluttered pedestrian walkways. Journal of Field Robotics, 26(8):609-635, 2009. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.398, + 0.484, + 0.464 + ], + "angle": 0, + "content": "[4] Manolis Savva, Abhishek Kadian, Oleksandr Maksymets, Yili Zhao, Erik Wijmans, Bhavana Jain, Julian Straub, Jia Liu, Vladlen Koltun, Jitendra Malik, et al. Habitat: A platform for embodied ai research. In CVPR, pages 9339-9347, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.467, + 0.482, + 0.52 + ], + "angle": 0, + "content": "[5] Fei Xia, Amir R Zamir, Zhiyang He, Alexander Sax, Jitendra Malik, and Silvio Savarese. Gibson env: Real-world perception for embodied agents. In CVPR, pages 9068-9079, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.523, + 0.483, + 0.59 + ], + "angle": 0, + "content": "[6] Eric Kolve, Roozbeh Mottaghi, Winson Han, Eli VanderBilt, Luca Weihs, Alvaro Herrasti, Matt Deitke, Kiana Ehsani, Daniel Gordon, Yuke Zhu, et al. Ai2-thor: An interactive 3d environment for visual ai. arXiv preprint arXiv:1712.05474, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.593, + 0.482, + 0.647 + ], + "angle": 0, + "content": "[7] Angel Chang, Angela Dai, Thomas Funkhouser, Maciej Halber, Matthias Niessner, Manolis Savva, Shuran Song, Andy Zeng, and Yinda Zhang. Matterport3d: Learning from RGB-D data in indoor environments. 3DV, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.649, + 0.482, + 0.703 + ], + "angle": 0, + "content": "[8] Erik Wijmans, Abhishek Kadian, Ari S. Morcos, Stefan Lee, Irfan Essa, Devi Parikh, Manolis Savva, and Dhruv Batra. Dd-ppo: Learning near-perfect pointgoal navigators from 2.5 billion frames. In ICLR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.705, + 0.482, + 0.771 + ], + "angle": 0, + "content": "[9] Piotr Mirowski, Matt Grimes, Mateusz Malinowski, Karl Moritz Hermann, Keith Anderson, Denis Teplyashin, Karen Simonyan, Andrew Zisserman, Raia Hadsell, et al. Learning to navigate in cities without a map. NeurIPS, 31, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.775, + 0.482, + 0.816 + ], + "angle": 0, + "content": "[10] Ang Li, Huiyi Hu, Piotr Mirowski, and Mehrdad Farajtabar. Cross-view policy learning for street navigation. In ICCV, pages 8100-8109, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.818, + 0.482, + 0.872 + ], + "angle": 0, + "content": "[11] Karl Moritz Hermann, Mateusz Malinowski, Piotr Mirowski, Andras Banki-Horvath, Keith Anderson, and Raia Hadsell. Learning to follow directions in street view. In AAAI, volume 34, pages 11773-11781, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.874, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[12] Piotr Mirowski, Andras Banki-Horvath, Keith Anderson, Denis Teptyashin, Karl Moritz Hermann, Mateusz Mali" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.228, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "nowski, Matthew Koichi Grimes, Karen Simonyan, Koray Kavukcuoglu, Andrew Zisserman, et al. The streetlearn environment and dataset. arXiv preprint arXiv:1903.01292, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.149, + 0.905, + 0.188 + ], + "angle": 0, + "content": "[13] Dhruv Shah and Sergey Levine. ViKiNG: Vision-Based Kilometer-Scale Navigation with Geographic Hints. In RSS, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.19, + 0.905, + 0.245 + ], + "angle": 0, + "content": "[14] Dhruv Shah, Ajay Sridhar, Arjun Bhorkar, Noriaki Hirose, and Sergey Levine. Gnm: A general navigation model to drive any robot. In ICRA, pages 7226-7233. IEEE, 2023. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.247, + 0.905, + 0.301 + ], + "angle": 0, + "content": "[15] Wayne Wu, Honglin He, Yiran Wang, Chenda Duan, Jack He, Zhizheng Liu, Quanyi Li, and Bolei Zhou. Metaurban: A simulation platform for embodied ai in urban spaces. arXiv preprint arXiv:2407.08725, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.303, + 0.905, + 0.37 + ], + "angle": 0, + "content": "[16] Adarsh Jagan Sathyamoorthy, Utsav Patel, Moumita Paul, Nithish K Sanjeev Kumar, Yash Savle, and Dinesh Manocha. Comet: Modeling group cohesion for socially compliant robot navigation in crowded scenes. IEEE Robotics and Automation Letters, 7(2):1008-1015, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.372, + 0.905, + 0.413 + ], + "angle": 0, + "content": "[17] Muchen Sun, Francesca Baldini, Peter Trautman, and Todd Murphey. Move beyond trajectories: Distribution space coupling for crowd navigation. In RSS, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.415, + 0.905, + 0.469 + ], + "angle": 0, + "content": "[18] Yuya Onozuka, Ryosuke Matsumi, and Motoki Shino. Autonomous mobile robot navigation independent of road boundary using driving recommendation map. In IROS, pages 4501-4508. IEEE, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.47, + 0.904, + 0.496 + ], + "angle": 0, + "content": "[19] Tom B Brown. Language models are few-shot learners. arXiv preprint arXiv:2005.14165, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.498, + 0.905, + 0.552 + ], + "angle": 0, + "content": "[20] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In ICCV, pages 4015-4026, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.554, + 0.905, + 0.691 + ], + "angle": 0, + "content": "[21] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel HAZIZA, Francisco Massa, Alaaeldin El-Nouby, Mido Assran, Nicolas Ballas, Wojciech Galuba, Russell Howes, Po-Yao Huang, Shang-Wen Li, Ishan Misra, Michael Rabbat, Vasu Sharma, Gabriel Synnaeve, Hu Xu, Herve Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. DINOv2: Learning robust visual features without supervision. Transactions on Machine Learning Research, 2024. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.692, + 0.905, + 0.76 + ], + "angle": 0, + "content": "[22] Danny Driess, Fei Xia, Mehdi SM Sajjadi, Corey Lynch, Aakanksha Chowdhery, Brian Ichter, Ayzaan Wahid, Jonathan Tompson, Quan Vuong, Tianhe Yu, et al. Palme: An embodied multimodal language model. arXiv preprint arXiv:2303.03378, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.762, + 0.905, + 0.816 + ], + "angle": 0, + "content": "[23] Beichen Wang, Juexiao Zhang, Shuwen Dong, Irving Fang, and Chen Feng. Vlm see, robot do: Human demo video to robot action plan via vision language model. arXiv preprint arXiv:2410.08792, 2024. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.818, + 0.905, + 0.872 + ], + "angle": 0, + "content": "[24] Noriaki Hirose, Catherine Glossop, Ajay Sridhar, Dhruv Shah, Oier Mees, and Sergey Levine. Lelan: Learning a language-conditioned navigation policy from in-the-wild videos. arXiv preprint arXiv:2410.03603, 2024. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.904, + 0.901 + ], + "angle": 0, + "content": "[25] Zachary Teed, Lahav Lipson, and Jia Deng. Deep patch visual odometry. NeurIPS, 36, 2024. 2, 3, 4" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "6883" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[26] Max Jaderberg, Volodymyr Mnih, Wojciech Marian Czarnecki, Tom Schaul, Joel Z Leibo, David Silver, and Koray Kavukcuoglu. Reinforcement learning with unsupervised auxiliary tasks. In ICLR, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.203 + ], + "angle": 0, + "content": "[27] Saurabh Gupta, James Davidson, Sergey Levine, Rahul Sukthankar, and Jitendra Malik. Cognitive mapping and planning for visual navigation. In CVPR, pages 2616-2625, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.205, + 0.482, + 0.245 + ], + "angle": 0, + "content": "[28] Nikolay Savinov, Alexey Dosovitskiy, and Vladlen Koltun. Semi-parametric topological memory for navigation. In ICLR, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.247, + 0.482, + 0.3 + ], + "angle": 0, + "content": "[29] Meera Hahn, Devendra Singh Chaplot, Shubham Tulsiani, Mustafa Mukadam, James M Rehg, and Abhinav Gupta. No rl, no simulation: Learning to navigate without navigating. NeurIPS, 34:26661-26673, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.302, + 0.482, + 0.357 + ], + "angle": 0, + "content": "[30] Obin Kwon, Nuri Kim, Yunho Choi, Hwiyeon Yoo, Jeongho Park, and Songhwai Oh. Visual graph memory with unsupervised representation for visual navigation. In CVPR, pages 15890-15899, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.358, + 0.482, + 0.425 + ], + "angle": 0, + "content": "[31] Jacob Krantz, Theophile Gervet, Karmesh Yadav, Austin Wang, Chris Paxton, Roozbeh Mottaghi, Dhruv Batra, Jitendra Malik, Stefan Lee, and Devendra Singh Chaplot. Navigating to objects specified by images. In CVPR, pages 10916-10925, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.427, + 0.482, + 0.468 + ], + "angle": 0, + "content": "[32] Theophile Gervet, Soumith Chintala, Dhruv Batra, Jitendra Malik, and Devendra Singh Chaplot. Navigating to objects in the real world. Science Robotics, 8(79):eadf6991, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.47, + 0.482, + 0.523 + ], + "angle": 0, + "content": "[33] Devendra Singh Chaplot, Dhiraj Prakashchand Gandhi, Abhinav Gupta, and Russ R Salakhutdinov. Object goal navigation using goal-oriented semantic exploration. NeurIPS, 33:4247-4258, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.525, + 0.482, + 0.579 + ], + "angle": 0, + "content": "[34] Arjun Majumdar, Gunjan Aggarwal, Bhavika Devnani, Judy Hoffman, and Dhruv Batra. Zson: Zero-shot object-goal navigation using multimodal goal embeddings. NeurIPS, 35:32340-32352, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.581, + 0.482, + 0.622 + ], + "angle": 0, + "content": "[35] Abhishek Das, Samyak Datta, Georgia Gkioxari, Stefan Lee, Devi Parikh, and Dhruv Batra. Embodied question answering. In CVPR, pages 1-10, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.623, + 0.482, + 0.664 + ], + "angle": 0, + "content": "[36] Hao Tan, Licheng Yu, and Mohit Bansal. Learning to navigate unseen environments: Back translation with environmental dropout. arXiv preprint arXiv:1904.04195, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.665, + 0.482, + 0.72 + ], + "angle": 0, + "content": "[37] Jacob Krantz, Erik Wijmans, Arjun Majumdar, Dhruv Batra, and Stefan Lee. Beyond the nav-graph: Vision-and-language navigation in continuous environments. In ECCV, pages 104-120. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.721, + 0.482, + 0.774 + ], + "angle": 0, + "content": "[38] Maks Sorokin, Jie Tan, C Karen Liu, and Sehoon Ha. Learning to navigate sidewalks in outdoor environments. IEEE Robotics and Automation Letters, 7(2):3906-3913, 2022. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.776, + 0.482, + 0.843 + ], + "angle": 0, + "content": "[39] Joanne Truong, Max Rudolph, Naoki Harrison Yokoyama, Sonia Chernova, Dhruv Batra, and Akshara Rai. Rethinking sim2real: Lower fidelity simulation leads to higher sim2real transfer in navigation. In CoRL, pages 859-870. PMLR, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.845, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[40] Joanne Truong, Denis Yarats, Tianyu Li, Franziska Meier, Sonia Chernova, Dhruv Batra, and Akshara Rai. Learning navigation skills for legged robots with learned robot embeddings. In IROS, pages 484-491. IEEE, 2021." + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[41] Wei Liu, Huihua Zhao, Chenran Li, Joydeep Biswas, Billy Okal, Pulkit Goyal, Yan Chang, and Soha Pouya. X-mobility: End-to-end generalizable navigation via world modeling. arXiv preprint arXiv:2410.17491, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.149, + 0.905, + 0.203 + ], + "angle": 0, + "content": "[42] Gregory Kahn, Pieter Abbeel, and Sergey Levine. Badgr: An autonomous self-supervised learning-based navigation system. IEEE Robotics and Automation Letters, 6(2):1312-1319, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.206, + 0.905, + 0.247 + ], + "angle": 0, + "content": "[43] Gregory Kahn, Pieter Abbeel, and Sergey Levine. Land: Learning to navigate from disengagements. IEEE Robotics and Automation Letters, 6(2):1872-1879, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.249, + 0.905, + 0.302 + ], + "angle": 0, + "content": "[44] Dhruv Shah, Ajay Sridhar, Nitish Dashora, Kyle Stachowicz, Kevin Black, Noriaki Hirose, and Sergey Levine. ViNT: A foundation model for visual navigation. In CoRL, 2023. 2, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.305, + 0.905, + 0.358 + ], + "angle": 0, + "content": "[45] Ajay Sridhar, Dhruv Shah, Catherine Glossop, and Sergey Levine. Nomad: Goal masked diffusion policies for navigation and exploration. In ICRA, pages 63-70, 2024. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.362, + 0.905, + 0.431 + ], + "angle": 0, + "content": "[46] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763. PMLR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.433, + 0.905, + 0.473 + ], + "angle": 0, + "content": "[47] Annie S Chen, Suraj Nair, and Chelsea Finn. Learning generalizable robotic reward functions from \"in-the-wild\" human videos. arXiv preprint arXiv:2103.16817, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.475, + 0.905, + 0.503 + ], + "angle": 0, + "content": "[48] Shikhar Bahl, Abhinav Gupta, and Deepak Pathak. Human-to-robot imitation in the wild. In RSS, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.504, + 0.905, + 0.545 + ], + "angle": 0, + "content": "[49] Austin Patel, Andrew Wang, Ilija Radosavovic, and Jitendra Malik. Learning to imitate object interactions from internet videos. arXiv preprint arXiv:2211.13225, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.547, + 0.905, + 0.588 + ], + "angle": 0, + "content": "[50] Jimuyang Zhang, Ruizhao Zhu, and Eshed Ohn-Bar. Selfd: Self-learning large-scale driving policies from the web. In CVPR, pages 17316-17326, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.59, + 0.905, + 0.644 + ], + "angle": 0, + "content": "[51] Naoki Wake, Atsushi Kanehira, Kazuhiro Sasabuchi, Jun Takamatsu, and Katsushi Ikeuchi. Gpt-4v (ision) for robotics: Multimodal task planning from human demonstration. IEEE Robotics and Automation Letters, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.646, + 0.905, + 0.674 + ], + "angle": 0, + "content": "[52] Jakob Engel, Vladlen Koltun, and Daniel Cremers. Direct sparse odometry. TPAMI, 40(3):611-625, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.676, + 0.905, + 0.73 + ], + "angle": 0, + "content": "[53] Raul Mur-Artal, Jose Maria Martinez Montiel, and Juan D Tardos. Orb-slam: a versatile and accurate monocular slam system. IEEE transactions on robotics, 31(5):1147-1163, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.732, + 0.905, + 0.772 + ], + "angle": 0, + "content": "[54] Zachary Teed and Jia Deng. Droid-slam: Deep visual slam for monocular, stereo, and rgb-d cameras. NeurIPS, 34:16558-16569, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.775, + 0.905, + 0.83 + ], + "angle": 0, + "content": "[55] Yuhang He, Irving Fang, Yiming Li, Rushi Bhavesh Shah, and Chen Feng. Metric-Free Exploration for Topological Mapping by Task and Motion Imitation in Feature Space. In RSS, July 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.832, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[56] Mahmoud Assran, Quentin Duval, Ishan Misra, Piotr Bojanowski, Pascal Vincent, Michael Rabbat, Yann LeCun, and Nicolas Ballas. Self-supervised learning from images with a joint-embedding predictive architecture. In CVPR, pages 15619–15629, 2023." + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "6884" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.161 + ], + "angle": 0, + "content": "[57] Adrien Bardes, Quentin Garrido, Jean Ponce, Xinlei Chen, Michael Rabbat, Yann LeCun, Mahmoud Assran, and Nicolas Ballas. Revisiting feature prediction for learning visual representations from video. arXiv preprint arXiv:2404.08471, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.164, + 0.482, + 0.233 + ], + "angle": 0, + "content": "[58] Adarsh Jagan Sathyamoorthy, Kasun Weerakoon, Mohamed Elnoor, Anuj Zore, Brian Ichter, Fei Xia, Jie Tan, Wenhao Yu, and Dinesh Manocha. Convoi: Context-aware navigation using vision language models in outdoor and indoor environments. arXiv preprint arXiv:2403.15637, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.234, + 0.482, + 0.288 + ], + "angle": 0, + "content": "[59] Wei Xu and Fu Zhang. Fast-lio: A fast, robust lidar-inertial odometry package by tightly-coupled iterated kalman filter. IEEE Robotics and Automation Letters, 6(2):3317-3324, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.291, + 0.482, + 0.332 + ], + "angle": 0, + "content": "[60] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.334, + 0.482, + 0.375 + ], + "angle": 0, + "content": "[61] Noriaki Hirose, Dhruv Shah, Ajay Sridhar, and Sergey Levine. SACson: Scalable autonomous control for social navigation. IEEE Robotics and Automation Letters, 2023. 6" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.375 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.515, + 0.956 + ], + "angle": 0, + "content": "6885" + } + ] +] \ No newline at end of file diff --git a/2025/CityWalker_ Learning Embodied Urban Navigation from Web-Scale Videos/a9edaa72-0d38-4c4e-a25a-9480eb60d753_origin.pdf b/2025/CityWalker_ Learning Embodied Urban Navigation from Web-Scale Videos/a9edaa72-0d38-4c4e-a25a-9480eb60d753_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..bfff0391d389f2ac2bbfc4955b89dfd1a38ecc30 --- /dev/null +++ b/2025/CityWalker_ Learning Embodied Urban Navigation from Web-Scale Videos/a9edaa72-0d38-4c4e-a25a-9480eb60d753_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:612174410b5dcb5144bd8bd24741915de136fc067516fbe4c8eea7c685dd0b85 +size 1869731 diff --git a/2025/CityWalker_ Learning Embodied Urban Navigation from Web-Scale Videos/full.md b/2025/CityWalker_ Learning Embodied Urban Navigation from Web-Scale Videos/full.md new file mode 100644 index 0000000000000000000000000000000000000000..5cee990bac57af7d1b74b7e04668f49932906230 --- /dev/null +++ b/2025/CityWalker_ Learning Embodied Urban Navigation from Web-Scale Videos/full.md @@ -0,0 +1,324 @@ +# CityWalker: Learning Embodied Urban Navigation from Web-Scale Videos + +Xinhao Liu* Jintong Li* Yicheng Jiang Niranjan Sujay Zhicheng Yang +Juexiao Zhang John Abanes Jing Zhang Chen Feng† +New York University + +https://ai4ce.github.io/CityWalker/ + +![](images/a1fec92d74c553c785a46a93318511fffcceb86a4406b4c578d6f80625d6d74e.jpg) +Figure 1. Embodied Urban Navigation. Navigating urban spaces is challenging for (especially off-street) mobile agents. The differently colored pins $(\bullet)$ along the route highlight various critical scenarios unique to complex and dynamic urban landscapes. Thumbsails on the right with corresponding colored pins demonstrate the real-world observation of these challenging cases. Our CityWalker model is trained with over 2000 hours of city walking videos and fine-tuned with a small amount of expert data to address these challenges effectively. + +![](images/39c6b186b02948833978a4d38b8df20737fd408be235e3a3581711d0a87f12a0.jpg) + +![](images/c59de79307c0829cbafaa166c9aca5ce5fd4239a7ecbcf07eff72dabb96b9abc.jpg) + +![](images/3933f2cb71831601565ec90e622646d0e5d0c66910bdda4c578ce902f85546aa.jpg) + +![](images/816003afc62fae5ff8d1c4875a468feab2afd20e5f308e411bd63cfc55a9b542.jpg) + +![](images/5d8098eb857d237ff52acff78373d423fa0be62db101fda42e13f16c506b3ce3.jpg) + +# Abstract + +Navigating dynamic urban environments presents significant challenges for embodied agents, requiring advanced spatial reasoning and adherence to common-sense norms. Despite progress, existing visual navigation methods struggle in map-free or off-street settings, limiting the deployment of autonomous agents like last-mile delivery robots. To overcome these obstacles, we propose a scalable, data-driven approach for human-like urban navigation by training agents on thousands of hours of in-the-wild city walking and driving videos sourced from the web. We introduce a simple and scalable data processing pipeline that extracts action supervision from these videos, enabling large-scale imitation learning without costly annotations. Our model learns sophisticated navigation policies to handle diverse challenges and critical scenarios. Experimental results show that training on large-scale, diverse datasets significantly enhances navigation performance, surpassing current methods. This work shows the potential of using abundant online video data to develop robust navigation policies for embodied agents in dynamic urban settings. + +# 1. Introduction + +Visual navigation is a crucial capability for mobile agents. In previously unseen environments, humans typically rely on navigation tools (such as Google Maps) to reach a goal location by following a series of waypoints. These tools provide high-level guidance, but the actual navigation between waypoints requires sophisticated spatial awareness and decision-making. In urban settings, this involves understanding and adhering to complex rules, such as common sense regulations and norms, while dynamically responding to obstacles and environmental changes. The diversity and complexity encountered in urban navigation scenarios pose significant challenges for navigation policies but are essential for deploying mobile robots in real-world settings. Applications like delivery robots and robotaxis, especially without high-definition maps, highlight the growing need for embodied agents to navigate efficiently and safely in dynamic urban landscapes. How can we teach embodied agents to navigate dynamic urban landscapes as effectively as humans do? + +The question has been explored in early works [1-3] before the deep learning era. These approaches combined SLAM-based mapping with modular systems for search + +ing, detection, analysis, and planning. The development of high-performance photorealistic simulators [4-7] significantly advanced research in visual navigation by enabling reinforcement learning (RL) in static indoor environments, achieving near-perfect point-goal navigation [8]. This simulation-RL paradigm was also extended to more complex settings, including real-world panoramic street views [9-12]. Recent works begin to explore imitation learning from expert demonstrations for long-horizon or cross-embodiment navigation tasks [13, 14]. + +Despite their success in simulators or moderately complex environments, existing methods fall short in addressing the complexity of embodied urban navigation. Public urban spaces are inherently dynamic and unpredictable, characterized by "multifarious terrains, diverse obstacles, and dense pedestrians" that require real-time adaptability [15]. Moreover, effective navigation in such settings demands adherence to common sense navigation rules and social norms, such as using sidewalks, obeying traffic signals, and maintaining appropriate personal space to avoid conflicts [16-18]. Current reinforcement learning and imitation learning approaches typically excel in static or controlled environments but struggle with real-world urban navigation. This is because these nuanced behaviors and delicate constraints are challenging to incorporate into simulation environments while ensuring reliable and generalizable sim-to-real transfer. Moreover, coverage of these complexities is still lacking in existing teleoperation demonstration datasets. As a result, embodied urban navigation remains an unsolved problem. + +While a straightforward approach to the problem is to collect expert trajectories through teleoperation, this is limited by data size and lack of diversity, which weakens the agent's ability to generalize across various urban scenarios. Is there a data-driven approach that requires minimum annotation but maximizes generalizability? Inspired by the success of scaling laws in language, vision, and robotics tasks [19-22], we propose a scalable framework, CityWalker, which leverages web-scale city walking and driving videos to train models for embodied urban navigation. Our model is trained on over two thousand hours of internet-sourced videos across different geolocations, weather conditions, and times of day. + +The subsequent question is how to effectively obtain action supervision from in-the-wild videos for imitation learning. Current solutions rely heavily on prompting proprietary VLMs to generate action labels from videos [23, 24], which is costly and challenging to control, making it difficult to scale for large datasets. We argue that noisy pseudo labels from off-the-shelf visual odometry (VO) models [25] are sufficiently effective for imitation learning in embodied urban navigation. Moreover, this data processing method generalizes beyond walking data. We demonstrate that + +a model trained with driving videos also performs well on quadruped agents, with further improvements observed when training in a cross-domain setting that combines both city walking and driving data. + +We summarize our contribution as the following: + +- We identify embodied urban navigation as an intriguing yet unsolved problem. To address this problem, we propose a scalable, data-driven solution by leveraging web-scale city walking and driving videos. +- We introduce a simple yet scalable data processing paradigm, making large-scale imitation learning feasible for urban navigation without extensive manual labeling. +- We demonstrate training on web-scale data significantly improves navigation performance in real-world experiments, allowing embodied agents to handle the complexities of urban scenarios. We also show model performance scales well with training data size. + +# 2. Related Work + +Navigation in Simulation. The introduction of high-performance photorealistic simulators [4-7] has significantly boosted navigation research in indoor household environments. Based on the form of goal specification, navigation tasks are usually categorized into point-goal [4, 8, 26, 27], image-goal [28-31], object-goal [32-34], and vision-language navigation [35-37]. While tasks like image-goal and object-goal navigation focus on efficient goal-finding or instruction understanding, point-goal navigation has a more straightforward goal specification and emphasizes safe and efficient trajectory planning to the goal. Our problem formulation resembles point-goal navigation in that accurate waypoints (represented as GPS coordinates) are readily available from navigation tools. Recently, point-goal navigation has been considered a "solved" problem due to near-perfect performance achieved in simulators [8]. However, our investigation shows that waypoint-goal navigation in public urban spaces remains an unsolved challenge. + +Real-World Navigation. Real-world robot navigation has a long history, predating the deep learning era. Early methods relied on ranging sensors and modular SLAM systems for navigation [1-3]. Witnessing the success in simulators, various methods have been proposed to bridge the sim-to-real gap [38-41]. Another thread of work learns navigation policies from expert trajectories [42, 43]. By incorporating data from different embodiments [14] and carefully designed heuristics [13], these methods have been extended to long-horizon navigation on different robots [44, 45]. While deployed on real-world embodied agents, these methods still operate in relatively simple and static environments such as suburban areas, parking garages, or controlled obstacle courses. They are not guaranteed to generalize to significantly different, dynamic, and complex urban landscapes. Our work follows this line of research by training naviga + +![](images/6a7facbf87e4bdf11ed8d4bd2974e20170545956d39babae1641fd92bed4f3e7.jpg) +Figure 2. Overall Illustration of CityWalker. Our training pipeline starts with internet-sourced videos, using visual odometry to obtain relative poses between frames. At each time step, the model receives past observations, past trajectory, and target location as input. They are encoded via a frozen image encoder and a trainable coordinate encoder. A transformer processes these inputs to generate future tokens. An action head and an arrival head decode these tokens into action and arrival status predictions. During training, future frame tokens from future frames guide the transformer to hallucinate future tokens. + +tion models with imitation learning. But we focus on learning the rules and norms essential for urban environments, which is crucial for enabling applications such as delivery robots and robot-taxis. + +In-the-Wild Video Learning. Learning from web-scale data has proven successful in language and vision tasks [19-21, 46]. While some works utilize in-the-wild videos to extract intermediate representations, priors, or reward functions [47, 48], a key challenge remains: the lack of action labels. Patel et al. [49] use a reconstruction pipeline to obtain hand-object trajectories as action labels for learning robot manipulation. SelfD [50] argues BEV model is better than VO in pseudo-labeling, while we show VO works well with large-scale data. Recent works also rely on prompting closed VLMs to get action labels [23, 51]. LeLaN [24], a concurrent work most relevant to ours, also relies on VLM prompting and uses pretrained navigation models to generate action labels. In contrast, we process in-the-wild video data differently by using only the noisy pseudolabels from off-the-shelf visual SLAM tools [25]. This approach enables scalable and cost-effective action label generation for imitation learning in urban navigation tasks. + +# 3. Embodied Urban Navigation + +# 3.1. Problem Formulation + +In this work, we address the problem of embodied urban navigation for agents in dynamic urban environments. The + +agent's task is to navigate from its current location to a specified target waypoint location, following a series of waypoints provided by a navigation tool. Formally, we define this as a point-goal navigation problem in real-world urban settings. At each time step $t$ , the agent receives an RGB observation $o_t$ , its current GPS location $p_t$ , and a sub-goal waypoint $w_t$ . The agent aims to learn a policy $\pi(a_t \mid o_{(t-k):t}, p_{(t-k):t}, w_t)$ that maps the past observations and positional information to an action $a_t$ from the action space $A$ , represented as a series of action waypoints in Euclidean space. Typically, we take $k = 5$ and also predict actions for the next 5 time steps. + +As accurate waypoints are easily available from navigation tools (e.g. Google Maps API), we focus our problem on the navigation between two consecutive waypoints. The agent treats the current sub-goal $w_{t}$ as the immediate target and is responsible for determining when it has reached this sub-goal based on its observations and positional data. Upon reaching $w_{t}$ , the agent proceeds to the next waypoint in the sequence. + +# 3.2. Evaluation Metrics + +To enable analysis and assessment of different methods on the problem, our evaluation metrics are twofold. We use the ground truth trajectory in teleoperation-collected data to provide a thorough analysis of each method's performance across various scenarios. Below, we discuss the metrics we used in offline data evaluation. For real-world deployment, + +![](images/dadde6cde915baf037f3e33ff4541f0b884cbd63c0971954b01df46fa4829efc.jpg) +Figure 3. Evaluation Metrics. Left: The orientation error is defined to be the angle between each predicted and ground truth action pair, labeled by $\theta_{n}$ in the figure. The action angle $\varphi_{\mathrm{action}}$ and target angle $\varphi_{\mathrm{target}}$ are defined with respect to the positive y-axis. Right: Both green and red trajectories are predicted actions. The green trajectory is the preferred one, having a large L2 distance but a small AOE. Vice versa for the red trajectory. + +![](images/d957e4dc9fa6bdda9adae73a099f6720b20a085f50a68eca2fb063f5f5f94cb7.jpg) +Figure 4. Data Sample and Visual Odometry (VO) Result. Our internet-source training data includes both walking and driving videos. These videos cover various scenarios in the urban environment. The VO tool gives noisy trajectories globally, but trustworthy local relative pose within a short time period. + +we primarily use navigation success rate as the key metric for performance comparison. + +Average Orientation Error. We use the error between each predicted action and the ground truth action as the main evaluation metric on the offline teleoperation data. While L2-distance in Euclidean space is a straightforward metric, we find it does not adequately capture the problem. As shown on the right side of Fig. 3, the red trajectory has a smaller L2 error compared to the green trajectory but moves in a direction that deviates from the target. This observation motivates us to come up with a new evaluation metric that better reflects the quality of the predicted actions. We define average orientation error (AOE) as the angle difference between the predicted and ground truth actions, averaged across all samples at each step (left side of Fig. 3): + +$$ +\operatorname {A O E} (k) = \frac {1}{n} \sum_ {i} ^ {n} \theta_ {i _ {k}} = \frac {1}{n} \sum_ {i} ^ {n} \arccos \frac {\left\langle a _ {\hat {i} _ {k}} ^ {\prime} , a _ {i _ {k}} \right\rangle}{\| a _ {\hat {i} _ {k}} \| \| a _ {i _ {k}} \|}, \tag {1} +$$ + +where $k$ is the index of predicted action and $n$ is the number of data. Additionally, we observe that taking the mean across $k$ predicted future time steps can lead to an underestimation due to steps with very small errors. Hence, we propose using the maximum average orientation error (MAOE) over all predicted actions as an overall estimation: + +$$ +\mathrm {M A O E} = \frac {1}{n} \sum_ {i} ^ {n} \max _ {k} \theta_ {i _ {k}}. \tag {2} +$$ + +As orientation error might be noisy when action distance is small, we also use L2-distance as a complimentary evaluation metric. + +Critical Scenarios. Not all time steps are equally critical in a trajectory. For example, errors at road crossings or turning points are more likely to result in navigation failure compared to mistakes during forward movement in open areas. + +![](images/8cd6e988b5616c6a39f97a64c82c9d0084ca7375cb3a91e95dfe052cb4bc82e0.jpg) + +We identify several key scenarios that are most critical during navigation. AOE and MAOE are calculated separately for each scenario, and we use the scenario mean rather than the sample mean for a more comprehensive evaluation. Detailed definitions of each scenario are provided in Sec. 4. + +# 3.3. Learning from In-the-Wild Videos + +City Walking Videos. To facilitate large-scale imitation learning for human-like urban navigation, we utilize city walking videos sourced from the internet. Walking in urban environments is intuitive for humans, resulting in an abundance of first-person view videos available online. These videos naturally capture the complexities of navigating dynamic urban landscapes, including interactions with pedestrians, adherence to traffic signals, and maneuvering around various obstacles. Motivated by the extensive availability of such data and its inherent alignment with our navigation objectives, we curate a dataset comprising over 2000 hours of city walking videos, encompassing diverse weather and lighting conditions. These videos exhibit consistent motion patterns characteristic of walking, providing continuous motion cues that can directly serve as action labels. + +Action Labels from Videos. To generate actionable labels for imitation learning, we employ VO [25, 52-54] tools to extract trajectory poses from the collected city walking videos. Specifically, we utilize the state-of-the-art DPVO [25] while any other VO method should also work. The resulting VO trajectories are illustrated in Fig. 4. However, using VO poses presents two primary challenges. First, global trajectory accuracy is compromised by accumulation errors inherent in VO systems. Nevertheless, since our model focuses on predicting a limited sequence of actions (5 past and 5 future), we mitigate this issue by relying solely on relative poses within short temporal windows. Second, VO methods suffer from scale ambiguity, leading to inconsistencies in trajectory lengths across different data. For instance, as shown in Fig. 4, a driving video may exhibit a longer actual trajectory compared to a walking video, yet the DPVO output incorrectly reflects a shorter trajectory. To address this, we normalize each action by the average step length within its trajectory, following the approach in + +previous work [14]. This normalization resolves both scale ambiguity and inconsistencies in step lengths between different types of embodiments all at once. This enables the learning of policies in a consistent, abstracted action space. During deployment, the predicted actions are denormalized using the specific step length of the target robot, ensuring appropriate movement execution in real-world settings. + +Generalizability and Scalability. Our data processing pipeline is both generalizable and scalable. It can be applied to any video sources beyond city walking, as long as the videos are captured with egomotion, such as driving videos, which are also abundantly available online. This flexibility enables the training of more generalizable navigation policies that can be zero-shot applied or fine-tuned for cross-domain and cross-embediment tasks, such as transitioning from pedestrian to vehicular navigation. Furthermore, our method offers significant scalability advantages. Unlike pipelines that rely on prompting large language models for action label generation [23, 24], our visual odometry-based approach can be easily parallelized. Processing 2000 hours of video data with our pipeline requires negligible wall-clock time, making it both feasible and cost-efficient to scale up the training dataset. + +# 3.4. Pipeline and Training + +Pipeline Architecture. Figure 2 shows our pipeline. It builds upon previous works [44, 45]. At the core of our model is a transformer that processes a temporal sequence of input tokens, including $k$ image features and one coordinate embeddings. Specifically, the coordinate input is stacked by $k$ past positions and one single target position, and outputs a sequence of the same length. We use an action head and an arrival prediction head to decode the output of the transformer. We utilize DINOv2 [21] representation and freeze the backbone during training. + +Feature Hallucination. Inspired by previous works on feature learning [55-57], we use feature hallucination as an auxiliary loss during training. Specifically, we compute the MSE loss between the output image tokens and that directly extracted from future frames. We hope this will guide the transformer to predict more informative future tokens that best mimic future frames, which in turn helps the two MLP heads predict the final action and arrival status. + +Loss Functions. As mentioned in Sec. 3.2, L1 or L2 loss can be less effective than orientation error in evaluating the quality of the predicted actions. We also employ orientation error to supervise the training. We define the orientation loss to be the negative cosine similarity between the predicted actions and the ground truth actions: + +$$ +\mathcal {L} _ {\mathrm {o r i}} = - \frac {1}{k} \sum_ {i = 1} ^ {k} \frac {\left\langle \hat {a} _ {i} , a _ {i} \right\rangle}{\left\| \hat {a} _ {i} \right\| \left\| a _ {i} \right\|}. \tag {3} +$$ + +Together with the L1 loss for predicted action and the BCE + +loss for predicted arriving status, the final loss function is a weight sum of four individual losses: + +$$ +\mathcal {L} = \omega_ {l 1} \mathcal {L} _ {l 1} + \omega_ {\text {o r i}} \mathcal {L} _ {\text {o r i}} + \omega_ {\text {a r r}} \mathcal {L} _ {\text {a r r}} + \omega_ {\text {f e a t}} \mathcal {L} _ {\text {f e a t}}. \tag {4} +$$ + +We choose each weight so that each loss items are within the same magnitude. + +# 4. Experiment + +# 4.1. Setup + +Baselines. We compare our model with outdoor navigation models closely related to our setup, including GNM [14], ViNT [44], and NoMaD [45]. Although originally developed for image-goal navigation, we tested them using goal-images from our collected data. Additionally, ViNT as a foundation model for visual navigation supports finetuning, which we also evaluated. We acknowledge CoNVOI [58] as a recent work with a similar setup, but could not test it due to the lack of open source code and prompt. We also experimented with VLM-based approaches similar to CoNVOI and defer the results to the appendix. + +Data. We collected expert data via teleoperation for fine-tuning and offline testing. We collect data via a Unitree Go1 quadruped equipped with a Livox Mid-360 LiDAR and a webcam for RGB observations. We employed the LiDAR-SLAM method [59] to obtain the robot's poses as ground truth actions. Additionally, we retrieved location data from a smartphone using a webpage. We reduce noise from the quadruped's vibrations by holding the phone [38]. In total, we gathered 15 hours of teleoperation data across various areas in New York City, allocating 6 hours for fine-tuning and 9 hours for testing. + +Critical Scenarios As discussed in Sec. 3.2, we identify several critical scenarios in our dataset that we want to emphasize in our evaluation. We analyze the ground truth trajectory and run object detection [60] to aid us in standardizing the definition. We define these scenarios as follows: + +- Turn: When the ground truth action changes direction significantly. Defined when $\varphi_{\mathrm{action}} > 20^{\circ}$ (see Fig. 3). +- Crossing: When the agent is at a road crossing. Defined if any traffic light is detected with score $>0.5$ . +- Detour: When the action angle deviates from the target angle. Defined when $|\varphi_{\text{action}} - \varphi_{\text{target}}| > 45^{\circ}$ (see Fig. 3). +- Proximity: When any person is close to the agent. Defined when the largest bounding box of any detected person is larger than $25\%$ of the image area. +- **Crowd:** When the agent is surrounded by a crowd of people. Defined when $\geq 5$ people are detected. + +We define these scenarios in a way that they are not mutually exclusive. A single data sample can belong to multiple scenarios. Although these scenarios account for less than half of the data, we argue they are the most significant factors contributing to successful urban navigation. + +Table 1. Benchmark on Offline Data. We evaluate three metrics in each critical scenario for all methods. Percentages under scenarios indicate their data proportions. The "Mean" column shows scenario means averaged over six scenarios; "All" shows sample means over all data samples. We highlight the best performance in bold and the second-best performance with underline. + +
MethodMetricMeanTurn 8%Crossing 12%Detour 12%Proximity 6%Crowd 7%Other 55%All 100%
GNM [14] (fine-tuned)↓L2 (m)1.222.361.361.420.880.760.550.74
↓MAOE (°)16.231.114.812.514.712.811.012.1
↑Arrival (%)68.666.469.766.469.069.270.770.0
ViNT [44] (fine-tuned)↓L2 (m)1.301.911.131.140.770.660.570.70
↓MAOE (°)16.531.115.412.914.813.311.612.6
↑Arrival (%)70.571.268.770.773.068.671.070.7
NoMaD [45] (fine-tuned)↓L2 (m)1.392.491.561.551.060.950.760.74
↓MAOE (°)19.135.118.515.618.114.312.812.1
↑Arrival (%)68.666.469.766.469.069.370.770.0
Ours (zero-shot)↓L2 (m)1.341.301.091.331.441.481.391.38
↓MAOE (°)16.526.815.516.316.313.111.312.7
↑Arrival (%)79.169.371.478.884.184.586.484.1
Ours (fine-tuned)↓L2 (m)1.111.271.001.151.061.121.061.07
↓MAOE (°)15.226.614.113.914.312.010.411.5
↑Arrival (%)81.868.975.378.590.687.590.287.8
+ +# 4.2. Performance Benchmarking + +In this subsection, we aim to answer the question Q1: Can our CityWalker model navigate successfully in complex urban environments? To answer this, we benchmark our model and the baselines both with offline data and in real-world navigation deployment. + +Overall Benchmark. Table 1 shows the benchmark in different critical scenarios. Overall, our fine-tuned model has superior performance on all metrics in all scenarios except "detour". We think the less competent performance in the detour scenario is due to the small portion of such data in our video training data. It can be verified by the fact that the fine-tuning performance gain is most significant in this scenario. It is worth noting that our zero-shot model has similar or even better performance compared to the fine-tuned baselines. This demonstrates the power of large-scale pretraining, especially considering the domain and modality gap from human walking videos. + +Real-world Depolyment. We deploy our model and the baseline methods on the same Unitree Go1 quadruped for real-world navigation, in previously unseen urban environments. We adapt the PD controller provided in [44] to provide velocity command to the quadruped. We choose area with relatively constant density of dynamics such as pedestrian and traffic to ensure the consistency among different trials. In each trial, the target location is generally 50-100m far away from the starting location. We categorize the real-world test into forward, left turn, and right turn cases to provide case-by-case analysis. We conduct 8-14 trails for each cases. Each trial is marked success when the model predicts arrival within a 5m distance to the target location. + +Table 2. Real-World Navigation. The table shows the success rate of real-world experiments in different scenarios. * and † indicate zero-shot inference and fine-tuned models respectively. + +
MethodAllForwardLeft turnRight turn
ViNT* [44]37.762.50.050.0
ViNT† [44]57.1100.025.025.0
NoMad* [45]42.975.016.728.6
Ours†77.3100.062.566.7
+ +All human interruptions due to potential collision or timeout are treated as failure cases. + +Table 2 shows the success rate on all real-world test cases. Our model achieves the highest success rate among all cases with a significant gap to the second-best performing finetuned ViNT. As the baselines are trained with navigation data mostly in suburban or off-road environments [14, 61], they are limited in handling complex maneuvers, making them unsuitable for urban navigation tasks, as shown in Fig. 5. Our model's robust performance in both forward and turns highlights its ability to manage dynamic and varied urban scenarios effectively, allowing it to be more reliably used in real-world environments where frequent and precise directional changes are essential. The substantial improvement not only increases the overall navigation success but also extends the applicability of our model to more demanding and realistic urban settings, positioning it as a superior solution for embodied agents in dynamic environments. + +# 4.3. Power of Data Scaling + +Our model is significantly boosted by the large-scale training data. We ask Q2: To what extent does increasing the + +![](images/e2e21c1fff13ed9f1ecb15548d16055943dbfa72f0c842560a92675de7264991.jpg) +Current Observation + +![](images/7a65bb3bfa3d90537fdabf04d4bf3837becdff4952ca5f0d3c3f4873e749e697.jpg) +X (m) Coordinates + +![](images/0fe13c8cae1a774f18e9382a452b96400165389aca4549183e61382c09bba991.jpg) +ViNT Zero-shot +ViNT Fine-tune +X (m) Coordinates + +![](images/5b2f72ea50dc2532803c1996d73c3040eb534f21fe04ebadf5490c50cfe3fa3c.jpg) +NoMad Zero-shot +X (m) Coordinates + +![](images/33380879e63d2d78c8e92dea8df834cd596485aa2832d9e2d3a684ed65b228fc.jpg) +Ours + +![](images/7341e9001dee630a1d8f34b6f08cc51d7117895d198d9f5e56519b34cd167117.jpg) +Input past traj. + +![](images/deafe897752aa4918eaeeb125d39a1b2198df2be710e84b0fa980e6319617fd4.jpg) +Actual past traj. + +![](images/2076933fbec325f325f7415aa8b0a289875b223f6c67030a84a69dd66368d491.jpg) + +![](images/7503e3b96efdae1061cffe23de953bd09a4c58e828f94417e03853306a75f805.jpg) +Predicted action +GTaction +$\star$ Target + +![](images/ab2d565d131d2703bf9beed86839e64c15a34f1f6a091815665c16d402ed9be5.jpg) + +![](images/b77d40fa67f01460f9f1fb4a415fc7fc4e69cd6c17e41b360ff23bfdc32c5063.jpg) +Figure 5. Qualitative Results. Left image shows current observations of two samples. Right plots displays input trajectory, ground truth actions, and predicted actions in the current coordinate system with the agent at the origin. +Figure 6. Performance and Data Size. We show the model performance evaluated by MAOE with respect to the size of the training data measured by video length in hours. We also show the zero-shot performance of our model trained with only driving videos and mixed driving and walking videos + +size of the training dataset improve the performance of our urban navigation model? + +Performance Gain and Data Scale. Figure 6 shows the model performance when trained with different data sizes. The plot demonstrates that our model's zero-shot performance improves significantly as we increase the amount of training data. Notably, when trained with more than 1000 hours of videos, our zero-shot model (red line) achieves a better performance than the fine-tuned ViNT model $(\square)$ . This indicates that large-scale training on diverse video data can even surpass the performance of models that are fine-tuned with limited expert data, highlighting the potential of leveraging web-scale datasets for embodied navigation. + +Domain and Embodiment Gap. An unexpected observation is that our model without the feature hallucination loss outperforms the one with it in zero-shot inference. We hypothesize this phenomenon mainly due the domain and embodiment gap between human walking videos and quadruped navigation policies. The feature hallucination loss aims to guide the transformer model to produce output tokens that mimic future observations, essentially encouraging the model to anticipate future visual inputs. This enforcement might inadvertently hinder performance by focusing on predicting future observations based on human motion. The model may learn representations that are less applicable or even misleading for quadruped navigation. Consequently, the feature hallucination loss could introduce a bias that detracts from the model's ability to make accurate navigation decisions in the robot's operational domain. + +Feature Hallucination Loss. Fortunately, as shown in Tab. 1, this issue no longer exists after fine-tuning. We also show the benefit of feature hallucination loss from another perspective. Figure 7 shows two training loss curve for different data size and training setups with and without feature hallucination loss. We observe that although feature hallucination loss should increase the absolute value of the total loss, the training with feature hallucination loss lead to a lower total loss for 2000-hour training and a lower orientation loss in both cases. The training with feature hallucination loss also shows a steeper decreasing trend, showing a potential of having even lower error with more training steps. Therefore, this is another proof of the effectiveness of our proposed feature hallucination loss. + +Driving Videos. As discussed in Sec. 3.3, our data process + +![](images/1be800e77262e5e99074bb8830b35bd432270f3e860ac7f776841ec62b4a9925.jpg) +Figure 7. Improvement from Feature Hallucination. The plot shows the orientation loss for different amounts of training data and loss setups. For all four setups, we trian with $\omega_{\mathrm{II}} = 1.0, \omega_{\mathrm{arr}} = 1.0, \text{and}, \omega_{\mathrm{ori}} = 5.0$ . The unproportional training steps are due to different batch sizes used for different data sizes. All models are trained for 10 epochs. + +ing pipeline also enables the use of abundant online driving videos. We evaluate models trained exclusively on driving videos and on a combination of walking and driving videos, results shown in Fig. 6. We can observe that the model trained only with driving data $(+)$ has a similar performance to the zero-shot baselines. As a result, we can think of driving as a form of navigation from a different domain and embodiment. However, combining driving and walking data yields significant performance gains (x). With just 250 hours of mixed data, the model nearly matches the performance of models trained on 1000 hours of walking data alone. This signifies the generalizability and robustness provided by cross-domain cross-embodiment data. + +# 4.4. Pipeline component analysis + +To provide a more in-depth analysis of our model and pipeline, we ask Q3: How do the individual components of our pipeline enhance the performance and reliability of our model for real-world urban navigation? + +AOE at Different Time Steps. In Tab. 4, we study the AOE at different time steps, i.e. $\mathrm{AOE}(k)$ for $k$ from 1 to 5. ViNT exhibits increasing orientation errors as the prediction steps progress, indicating a degradation in accuracy over time. In contrast, our model maintains relatively consistent AOE from step 2 to step 5, which likely contributes to its superior performance in real-world navigation deployments with stable action prediction across multiple future steps. All methods experience higher errors at the first step. We attribute it to the proximity of the initial action to the origin, as minor linear deviations result in substantial angular error. It is worth noting that the MAOE is significantly larger than any individual AOE from steps 1 to 5, suggesting that the peak errors occur at varying time steps across different samples. This variability suggests the potential benefit of taking softmax for orientation errors instead of simple averaging, as it + +Table 3. Ablation Study. We ablate on the contribution of orientation loss, feature hallucination, and fine-tuning with expert data. We show the MAOE(°) in scenario-mean. All results are trained with 1000-hour walking data. + +
Training ComponentsMAOE
Ori. LossFeature Hall.Fine-tuningMean
17.03
✓17.00
✓✓17.02
✓15.23
✓✓15.21
✓✓✓15.16
+ +Table 4. AOE at Different Time Steps. The table shows MAOE and different AOE of scenario-mean, all in degrees $(^{\circ})$ . * and † indicates zero-shot inference and fine-tuned models respectively. + +
MethodAOE(1)AOE(2)AOE(3)AOE(4)AOE(5)MAOE
ViNT*[44]10.577.939.0910.2111.4517.52
ViNT†[44]11.247.968.408.939.8716.53
Ours*12.918.928.648.468.4316.54
Ours†11.447.777.687.757.9715.23
+ +could better capture the distribution of peak errors. + +Ablation Study. Table 3 examines the individual contributions of different pipeline components to our model's performance. A clear improvement in MAOE when fine-tuning with expert data, demonstrating the effectiveness of using in-domain expert information. The improvement from orientation loss and feature hallucination loss is marginal, within error range of our limited testing data. We defer more detailed ablation study to the appendix. + +# 5. Conclusion + +Summary. In this work, we addressed embodied urban navigation by training agents with thousands of hours of in-the-wild city walking videos, introducing a scalable pipeline for imitation learning from diverse walking videos. Our experiments show that large-scale, diverse training data significantly enhances navigation performance, surpassing existing methods and demonstrating the benefits of data scaling. Our findings highlight the potential of leveraging abundant online videos to develop robust navigation policies for embodied agents in dynamic urban environments. + +Limitation and Future Work. During real-world deployment, our current system implementation is sensitive to potential large location noises read naively from iPhone location services, which could be improved with better GPS hardware. Future work could also enhance the robustness of the model to such location noises. + +# Acknowledgment + +The work was supported by NSF grants 2238968, 2121391, 2322242 and 2345139; and in part through the NYU IT High Performance Computing resources, services, and staff expertise. We also thank Xingyu Liu and Zixuan Hu for their help in data collection. + +# References + +[1] Quirin Muhlbauer, Stefan Sosnowski, Tingting Xu, Tianguang Zhang, Kolja Kuhnlenz, and Martin Buss. Navigation through urban environments by visual perception and interaction. In ICRA, pages 3558-3564. IEEE, 2009. 1, 2 +[2] Rainer Kümmerle, Michael Ruhnke, Bastian Steder, Cyril Stachniss, and Wolfram Burgard. A navigation system for robots operating in crowded urban environments. In ICRA, pages 3225-3232. IEEE, 2013. +[3] Yoichi Morales, Alexander Carballo, Eijiro Takeuchi, Atsushi Aburadani, and Takashi Tsubouchi. Autonomous robot navigation in outdoor cluttered pedestrian walkways. Journal of Field Robotics, 26(8):609-635, 2009. 1, 2 +[4] Manolis Savva, Abhishek Kadian, Oleksandr Maksymets, Yili Zhao, Erik Wijmans, Bhavana Jain, Julian Straub, Jia Liu, Vladlen Koltun, Jitendra Malik, et al. Habitat: A platform for embodied ai research. In CVPR, pages 9339-9347, 2019. 2 +[5] Fei Xia, Amir R Zamir, Zhiyang He, Alexander Sax, Jitendra Malik, and Silvio Savarese. Gibson env: Real-world perception for embodied agents. In CVPR, pages 9068-9079, 2018. +[6] Eric Kolve, Roozbeh Mottaghi, Winson Han, Eli VanderBilt, Luca Weihs, Alvaro Herrasti, Matt Deitke, Kiana Ehsani, Daniel Gordon, Yuke Zhu, et al. Ai2-thor: An interactive 3d environment for visual ai. arXiv preprint arXiv:1712.05474, 2017. +[7] Angel Chang, Angela Dai, Thomas Funkhouser, Maciej Halber, Matthias Niessner, Manolis Savva, Shuran Song, Andy Zeng, and Yinda Zhang. Matterport3d: Learning from RGB-D data in indoor environments. 3DV, 2017. 2 +[8] Erik Wijmans, Abhishek Kadian, Ari S. Morcos, Stefan Lee, Irfan Essa, Devi Parikh, Manolis Savva, and Dhruv Batra. Dd-ppo: Learning near-perfect pointgoal navigators from 2.5 billion frames. In ICLR, 2019. 2 +[9] Piotr Mirowski, Matt Grimes, Mateusz Malinowski, Karl Moritz Hermann, Keith Anderson, Denis Teplyashin, Karen Simonyan, Andrew Zisserman, Raia Hadsell, et al. Learning to navigate in cities without a map. NeurIPS, 31, 2018. 2 +[10] Ang Li, Huiyi Hu, Piotr Mirowski, and Mehrdad Farajtabar. Cross-view policy learning for street navigation. In ICCV, pages 8100-8109, 2019. +[11] Karl Moritz Hermann, Mateusz Malinowski, Piotr Mirowski, Andras Banki-Horvath, Keith Anderson, and Raia Hadsell. Learning to follow directions in street view. In AAAI, volume 34, pages 11773-11781, 2020. +[12] Piotr Mirowski, Andras Banki-Horvath, Keith Anderson, Denis Teptyashin, Karl Moritz Hermann, Mateusz Mali + +nowski, Matthew Koichi Grimes, Karen Simonyan, Koray Kavukcuoglu, Andrew Zisserman, et al. The streetlearn environment and dataset. arXiv preprint arXiv:1903.01292, 2019. 2 +[13] Dhruv Shah and Sergey Levine. ViKiNG: Vision-Based Kilometer-Scale Navigation with Geographic Hints. In RSS, 2022. 2 +[14] Dhruv Shah, Ajay Sridhar, Arjun Bhorkar, Noriaki Hirose, and Sergey Levine. Gnm: A general navigation model to drive any robot. In ICRA, pages 7226-7233. IEEE, 2023. 2, 5, 6 +[15] Wayne Wu, Honglin He, Yiran Wang, Chenda Duan, Jack He, Zhizheng Liu, Quanyi Li, and Bolei Zhou. Metaurban: A simulation platform for embodied ai in urban spaces. arXiv preprint arXiv:2407.08725, 2024. 2 +[16] Adarsh Jagan Sathyamoorthy, Utsav Patel, Moumita Paul, Nithish K Sanjeev Kumar, Yash Savle, and Dinesh Manocha. Comet: Modeling group cohesion for socially compliant robot navigation in crowded scenes. IEEE Robotics and Automation Letters, 7(2):1008-1015, 2021. 2 +[17] Muchen Sun, Francesca Baldini, Peter Trautman, and Todd Murphey. Move beyond trajectories: Distribution space coupling for crowd navigation. In RSS, 2021. +[18] Yuya Onozuka, Ryosuke Matsumi, and Motoki Shino. Autonomous mobile robot navigation independent of road boundary using driving recommendation map. In IROS, pages 4501-4508. IEEE, 2021. 2 +[19] Tom B Brown. Language models are few-shot learners. arXiv preprint arXiv:2005.14165, 2020. 2, 3 +[20] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In ICCV, pages 4015-4026, 2023. +[21] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel HAZIZA, Francisco Massa, Alaaeldin El-Nouby, Mido Assran, Nicolas Ballas, Wojciech Galuba, Russell Howes, Po-Yao Huang, Shang-Wen Li, Ishan Misra, Michael Rabbat, Vasu Sharma, Gabriel Synnaeve, Hu Xu, Herve Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. DINOv2: Learning robust visual features without supervision. Transactions on Machine Learning Research, 2024. 3, 5 +[22] Danny Driess, Fei Xia, Mehdi SM Sajjadi, Corey Lynch, Aakanksha Chowdhery, Brian Ichter, Ayzaan Wahid, Jonathan Tompson, Quan Vuong, Tianhe Yu, et al. Palme: An embodied multimodal language model. arXiv preprint arXiv:2303.03378, 2023. 2 +[23] Beichen Wang, Juexiao Zhang, Shuwen Dong, Irving Fang, and Chen Feng. Vlm see, robot do: Human demo video to robot action plan via vision language model. arXiv preprint arXiv:2410.08792, 2024. 2, 3, 5 +[24] Noriaki Hirose, Catherine Glossop, Ajay Sridhar, Dhruv Shah, Oier Mees, and Sergey Levine. Lelan: Learning a language-conditioned navigation policy from in-the-wild videos. arXiv preprint arXiv:2410.03603, 2024. 2, 3, 5 +[25] Zachary Teed, Lahav Lipson, and Jia Deng. Deep patch visual odometry. NeurIPS, 36, 2024. 2, 3, 4 + +[26] Max Jaderberg, Volodymyr Mnih, Wojciech Marian Czarnecki, Tom Schaul, Joel Z Leibo, David Silver, and Koray Kavukcuoglu. Reinforcement learning with unsupervised auxiliary tasks. In ICLR, 2017. 2 +[27] Saurabh Gupta, James Davidson, Sergey Levine, Rahul Sukthankar, and Jitendra Malik. Cognitive mapping and planning for visual navigation. In CVPR, pages 2616-2625, 2017. 2 +[28] Nikolay Savinov, Alexey Dosovitskiy, and Vladlen Koltun. Semi-parametric topological memory for navigation. In ICLR, 2018. 2 +[29] Meera Hahn, Devendra Singh Chaplot, Shubham Tulsiani, Mustafa Mukadam, James M Rehg, and Abhinav Gupta. No rl, no simulation: Learning to navigate without navigating. NeurIPS, 34:26661-26673, 2021. +[30] Obin Kwon, Nuri Kim, Yunho Choi, Hwiyeon Yoo, Jeongho Park, and Songhwai Oh. Visual graph memory with unsupervised representation for visual navigation. In CVPR, pages 15890-15899, 2021. +[31] Jacob Krantz, Theophile Gervet, Karmesh Yadav, Austin Wang, Chris Paxton, Roozbeh Mottaghi, Dhruv Batra, Jitendra Malik, Stefan Lee, and Devendra Singh Chaplot. Navigating to objects specified by images. In CVPR, pages 10916-10925, 2023. 2 +[32] Theophile Gervet, Soumith Chintala, Dhruv Batra, Jitendra Malik, and Devendra Singh Chaplot. Navigating to objects in the real world. Science Robotics, 8(79):eadf6991, 2023. 2 +[33] Devendra Singh Chaplot, Dhiraj Prakashchand Gandhi, Abhinav Gupta, and Russ R Salakhutdinov. Object goal navigation using goal-oriented semantic exploration. NeurIPS, 33:4247-4258, 2020. +[34] Arjun Majumdar, Gunjan Aggarwal, Bhavika Devnani, Judy Hoffman, and Dhruv Batra. Zson: Zero-shot object-goal navigation using multimodal goal embeddings. NeurIPS, 35:32340-32352, 2022. 2 +[35] Abhishek Das, Samyak Datta, Georgia Gkioxari, Stefan Lee, Devi Parikh, and Dhruv Batra. Embodied question answering. In CVPR, pages 1-10, 2018. 2 +[36] Hao Tan, Licheng Yu, and Mohit Bansal. Learning to navigate unseen environments: Back translation with environmental dropout. arXiv preprint arXiv:1904.04195, 2019. +[37] Jacob Krantz, Erik Wijmans, Arjun Majumdar, Dhruv Batra, and Stefan Lee. Beyond the nav-graph: Vision-and-language navigation in continuous environments. In ECCV, pages 104-120. Springer, 2020. 2 +[38] Maks Sorokin, Jie Tan, C Karen Liu, and Sehoon Ha. Learning to navigate sidewalks in outdoor environments. IEEE Robotics and Automation Letters, 7(2):3906-3913, 2022. 2, 5 +[39] Joanne Truong, Max Rudolph, Naoki Harrison Yokoyama, Sonia Chernova, Dhruv Batra, and Akshara Rai. Rethinking sim2real: Lower fidelity simulation leads to higher sim2real transfer in navigation. In CoRL, pages 859-870. PMLR, 2023. +[40] Joanne Truong, Denis Yarats, Tianyu Li, Franziska Meier, Sonia Chernova, Dhruv Batra, and Akshara Rai. Learning navigation skills for legged robots with learned robot embeddings. In IROS, pages 484-491. IEEE, 2021. + +[41] Wei Liu, Huihua Zhao, Chenran Li, Joydeep Biswas, Billy Okal, Pulkit Goyal, Yan Chang, and Soha Pouya. X-mobility: End-to-end generalizable navigation via world modeling. arXiv preprint arXiv:2410.17491, 2024. 2 +[42] Gregory Kahn, Pieter Abbeel, and Sergey Levine. Badgr: An autonomous self-supervised learning-based navigation system. IEEE Robotics and Automation Letters, 6(2):1312-1319, 2021. 2 +[43] Gregory Kahn, Pieter Abbeel, and Sergey Levine. Land: Learning to navigate from disengagements. IEEE Robotics and Automation Letters, 6(2):1872-1879, 2021. 2 +[44] Dhruv Shah, Ajay Sridhar, Nitish Dashora, Kyle Stachowicz, Kevin Black, Noriaki Hirose, and Sergey Levine. ViNT: A foundation model for visual navigation. In CoRL, 2023. 2, 5, 6, 8 +[45] Ajay Sridhar, Dhruv Shah, Catherine Glossop, and Sergey Levine. Nomad: Goal masked diffusion policies for navigation and exploration. In ICRA, pages 63-70, 2024. 2, 5, 6 +[46] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763. PMLR, 2021. 3 +[47] Annie S Chen, Suraj Nair, and Chelsea Finn. Learning generalizable robotic reward functions from "in-the-wild" human videos. arXiv preprint arXiv:2103.16817, 2021. 3 +[48] Shikhar Bahl, Abhinav Gupta, and Deepak Pathak. Human-to-robot imitation in the wild. In RSS, 2022. 3 +[49] Austin Patel, Andrew Wang, Ilija Radosavovic, and Jitendra Malik. Learning to imitate object interactions from internet videos. arXiv preprint arXiv:2211.13225, 2022. 3 +[50] Jimuyang Zhang, Ruizhao Zhu, and Eshed Ohn-Bar. Selfd: Self-learning large-scale driving policies from the web. In CVPR, pages 17316-17326, 2022. 3 +[51] Naoki Wake, Atsushi Kanehira, Kazuhiro Sasabuchi, Jun Takamatsu, and Katsushi Ikeuchi. Gpt-4v (ision) for robotics: Multimodal task planning from human demonstration. IEEE Robotics and Automation Letters, 2024. 3 +[52] Jakob Engel, Vladlen Koltun, and Daniel Cremers. Direct sparse odometry. TPAMI, 40(3):611-625, 2017. 4 +[53] Raul Mur-Artal, Jose Maria Martinez Montiel, and Juan D Tardos. Orb-slam: a versatile and accurate monocular slam system. IEEE transactions on robotics, 31(5):1147-1163, 2015. +[54] Zachary Teed and Jia Deng. Droid-slam: Deep visual slam for monocular, stereo, and rgb-d cameras. NeurIPS, 34:16558-16569, 2021. 4 +[55] Yuhang He, Irving Fang, Yiming Li, Rushi Bhavesh Shah, and Chen Feng. Metric-Free Exploration for Topological Mapping by Task and Motion Imitation in Feature Space. In RSS, July 2023. 5 +[56] Mahmoud Assran, Quentin Duval, Ishan Misra, Piotr Bojanowski, Pascal Vincent, Michael Rabbat, Yann LeCun, and Nicolas Ballas. Self-supervised learning from images with a joint-embedding predictive architecture. In CVPR, pages 15619–15629, 2023. + +[57] Adrien Bardes, Quentin Garrido, Jean Ponce, Xinlei Chen, Michael Rabbat, Yann LeCun, Mahmoud Assran, and Nicolas Ballas. Revisiting feature prediction for learning visual representations from video. arXiv preprint arXiv:2404.08471, 2024. 5 +[58] Adarsh Jagan Sathyamoorthy, Kasun Weerakoon, Mohamed Elnoor, Anuj Zore, Brian Ichter, Fei Xia, Jie Tan, Wenhao Yu, and Dinesh Manocha. Convoi: Context-aware navigation using vision language models in outdoor and indoor environments. arXiv preprint arXiv:2403.15637, 2024. 5 +[59] Wei Xu and Fu Zhang. Fast-lio: A fast, robust lidar-inertial odometry package by tightly-coupled iterated kalman filter. IEEE Robotics and Automation Letters, 6(2):3317-3324, 2021. 5 +[60] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 5 +[61] Noriaki Hirose, Dhruv Shah, Ajay Sridhar, and Sergey Levine. SACson: Scalable autonomous control for social navigation. IEEE Robotics and Automation Letters, 2023. 6 \ No newline at end of file diff --git a/2025/CityWalker_ Learning Embodied Urban Navigation from Web-Scale Videos/images.zip b/2025/CityWalker_ Learning Embodied Urban Navigation from Web-Scale Videos/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..5d0e58f3f1df1d51b65104ac8cb1b83143918fe2 --- /dev/null +++ b/2025/CityWalker_ Learning Embodied Urban Navigation from Web-Scale Videos/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:eea8860298a3aaca3ad53d6c5e26531ebdd01d8539b0ead0732f5e380d4b3196 +size 554800 diff --git a/2025/CityWalker_ Learning Embodied Urban Navigation from Web-Scale Videos/layout.json b/2025/CityWalker_ Learning Embodied Urban Navigation from Web-Scale Videos/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..c70a89a24a3801f0ed8b34898e33e7fd96cfd63e --- /dev/null +++ b/2025/CityWalker_ Learning Embodied Urban Navigation from Web-Scale Videos/layout.json @@ -0,0 +1,8217 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 72, + 103, + 539, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 103, + 539, + 121 + ], + "spans": [ + { + "bbox": [ + 72, + 103, + 539, + 121 + ], + "type": "text", + "content": "CityWalker: Learning Embodied Urban Navigation from Web-Scale Videos" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 113, + 143, + 493, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 143, + 493, + 185 + ], + "spans": [ + { + "bbox": [ + 113, + 143, + 493, + 185 + ], + "type": "text", + "content": "Xinhao Liu* Jintong Li* Yicheng Jiang Niranjan Sujay Zhicheng Yang \nJuexiao Zhang John Abanes Jing Zhang Chen Feng† \nNew York University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 207, + 187, + 400, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 207, + 187, + 400, + 198 + ], + "spans": [ + { + "bbox": [ + 207, + 187, + 400, + 198 + ], + "type": "text", + "content": "https://ai4ce.github.io/CityWalker/" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 58, + 209, + 182, + 364 + ], + "blocks": [ + { + "bbox": [ + 58, + 209, + 182, + 364 + ], + "lines": [ + { + "bbox": [ + 58, + 209, + 182, + 364 + ], + "spans": [ + { + "bbox": [ + 58, + 209, + 182, + 364 + ], + "type": "image", + "image_path": "a1fec92d74c553c785a46a93318511fffcceb86a4406b4c578d6f80625d6d74e.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 369, + 555, + 415 + ], + "lines": [ + { + "bbox": [ + 55, + 369, + 555, + 415 + ], + "spans": [ + { + "bbox": [ + 55, + 369, + 555, + 415 + ], + "type": "text", + "content": "Figure 1. Embodied Urban Navigation. Navigating urban spaces is challenging for (especially off-street) mobile agents. The differently colored pins " + }, + { + "bbox": [ + 55, + 369, + 555, + 415 + ], + "type": "inline_equation", + "content": "(\\bullet)" + }, + { + "bbox": [ + 55, + 369, + 555, + 415 + ], + "type": "text", + "content": " along the route highlight various critical scenarios unique to complex and dynamic urban landscapes. Thumbsails on the right with corresponding colored pins demonstrate the real-world observation of these challenging cases. Our CityWalker model is trained with over 2000 hours of city walking videos and fine-tuned with a small amount of expert data to address these challenges effectively." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 185, + 208, + 252, + 364 + ], + "blocks": [ + { + "bbox": [ + 185, + 208, + 252, + 364 + ], + "lines": [ + { + "bbox": [ + 185, + 208, + 252, + 364 + ], + "spans": [ + { + "bbox": [ + 185, + 208, + 252, + 364 + ], + "type": "image", + "image_path": "39c6b186b02948833978a4d38b8df20737fd408be235e3a3581711d0a87f12a0.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 253, + 211, + 327, + 364 + ], + "blocks": [ + { + "bbox": [ + 253, + 211, + 327, + 364 + ], + "lines": [ + { + "bbox": [ + 253, + 211, + 327, + 364 + ], + "spans": [ + { + "bbox": [ + 253, + 211, + 327, + 364 + ], + "type": "image", + "image_path": "c59de79307c0829cbafaa166c9aca5ce5fd4239a7ecbcf07eff72dabb96b9abc.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 327, + 211, + 400, + 364 + ], + "blocks": [ + { + "bbox": [ + 327, + 211, + 400, + 364 + ], + "lines": [ + { + "bbox": [ + 327, + 211, + 400, + 364 + ], + "spans": [ + { + "bbox": [ + 327, + 211, + 400, + 364 + ], + "type": "image", + "image_path": "3933f2cb71831601565ec90e622646d0e5d0c66910bdda4c578ce902f85546aa.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 401, + 212, + 476, + 364 + ], + "blocks": [ + { + "bbox": [ + 401, + 212, + 476, + 364 + ], + "lines": [ + { + "bbox": [ + 401, + 212, + 476, + 364 + ], + "spans": [ + { + "bbox": [ + 401, + 212, + 476, + 364 + ], + "type": "image", + "image_path": "816003afc62fae5ff8d1c4875a468feab2afd20e5f308e411bd63cfc55a9b542.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 476, + 212, + 549, + 364 + ], + "blocks": [ + { + "bbox": [ + 476, + 212, + 549, + 364 + ], + "lines": [ + { + "bbox": [ + 476, + 212, + 549, + 364 + ], + "spans": [ + { + "bbox": [ + 476, + 212, + 549, + 364 + ], + "type": "image", + "image_path": "5d8098eb857d237ff52acff78373d423fa0be62db101fda42e13f16c506b3ce3.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 152, + 420, + 200, + 432 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 420, + 200, + 432 + ], + "spans": [ + { + "bbox": [ + 152, + 420, + 200, + 432 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 54, + 447, + 297, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 447, + 297, + 687 + ], + "spans": [ + { + "bbox": [ + 54, + 447, + 297, + 687 + ], + "type": "text", + "content": "Navigating dynamic urban environments presents significant challenges for embodied agents, requiring advanced spatial reasoning and adherence to common-sense norms. Despite progress, existing visual navigation methods struggle in map-free or off-street settings, limiting the deployment of autonomous agents like last-mile delivery robots. To overcome these obstacles, we propose a scalable, data-driven approach for human-like urban navigation by training agents on thousands of hours of in-the-wild city walking and driving videos sourced from the web. We introduce a simple and scalable data processing pipeline that extracts action supervision from these videos, enabling large-scale imitation learning without costly annotations. Our model learns sophisticated navigation policies to handle diverse challenges and critical scenarios. Experimental results show that training on large-scale, diverse datasets significantly enhances navigation performance, surpassing current methods. This work shows the potential of using abundant online video data to develop robust navigation policies for embodied agents in dynamic urban settings." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 315, + 420, + 394, + 432 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 420, + 394, + 432 + ], + "spans": [ + { + "bbox": [ + 315, + 420, + 394, + 432 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 445, + 555, + 672 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 445, + 555, + 672 + ], + "spans": [ + { + "bbox": [ + 313, + 445, + 555, + 672 + ], + "type": "text", + "content": "Visual navigation is a crucial capability for mobile agents. In previously unseen environments, humans typically rely on navigation tools (such as Google Maps) to reach a goal location by following a series of waypoints. These tools provide high-level guidance, but the actual navigation between waypoints requires sophisticated spatial awareness and decision-making. In urban settings, this involves understanding and adhering to complex rules, such as common sense regulations and norms, while dynamically responding to obstacles and environmental changes. The diversity and complexity encountered in urban navigation scenarios pose significant challenges for navigation policies but are essential for deploying mobile robots in real-world settings. Applications like delivery robots and robotaxis, especially without high-definition maps, highlight the growing need for embodied agents to navigate efficiently and safely in dynamic urban landscapes. How can we teach embodied agents to navigate dynamic urban landscapes as effectively as humans do?" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "text", + "content": "The question has been explored in early works [1-3] before the deep learning era. These approaches combined SLAM-based mapping with modular systems for search" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 693, + 292, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 693, + 292, + 703 + ], + "spans": [ + { + "bbox": [ + 66, + 693, + 292, + 703 + ], + "type": "text", + "content": "*Equal contribution ( {xinhao.liu, jintong.li}@nyu.edu)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 67, + 703, + 212, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 212, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 212, + 712 + ], + "type": "text", + "content": "†Corresponding author (cfeng@nyu.edu)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6875" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 192 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 192 + ], + "type": "text", + "content": "ing, detection, analysis, and planning. The development of high-performance photorealistic simulators [4-7] significantly advanced research in visual navigation by enabling reinforcement learning (RL) in static indoor environments, achieving near-perfect point-goal navigation [8]. This simulation-RL paradigm was also extended to more complex settings, including real-world panoramic street views [9-12]. Recent works begin to explore imitation learning from expert demonstrations for long-horizon or cross-embodiment navigation tasks [13, 14]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 194, + 296, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 194, + 296, + 434 + ], + "spans": [ + { + "bbox": [ + 56, + 194, + 296, + 434 + ], + "type": "text", + "content": "Despite their success in simulators or moderately complex environments, existing methods fall short in addressing the complexity of embodied urban navigation. Public urban spaces are inherently dynamic and unpredictable, characterized by \"multifarious terrains, diverse obstacles, and dense pedestrians\" that require real-time adaptability [15]. Moreover, effective navigation in such settings demands adherence to common sense navigation rules and social norms, such as using sidewalks, obeying traffic signals, and maintaining appropriate personal space to avoid conflicts [16-18]. Current reinforcement learning and imitation learning approaches typically excel in static or controlled environments but struggle with real-world urban navigation. This is because these nuanced behaviors and delicate constraints are challenging to incorporate into simulation environments while ensuring reliable and generalizable sim-to-real transfer. Moreover, coverage of these complexities is still lacking in existing teleoperation demonstration datasets. As a result, embodied urban navigation remains an unsolved problem." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 436, + 295, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 436, + 295, + 591 + ], + "spans": [ + { + "bbox": [ + 55, + 436, + 295, + 591 + ], + "type": "text", + "content": "While a straightforward approach to the problem is to collect expert trajectories through teleoperation, this is limited by data size and lack of diversity, which weakens the agent's ability to generalize across various urban scenarios. Is there a data-driven approach that requires minimum annotation but maximizes generalizability? Inspired by the success of scaling laws in language, vision, and robotics tasks [19-22], we propose a scalable framework, CityWalker, which leverages web-scale city walking and driving videos to train models for embodied urban navigation. Our model is trained on over two thousand hours of internet-sourced videos across different geolocations, weather conditions, and times of day." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "content": "The subsequent question is how to effectively obtain action supervision from in-the-wild videos for imitation learning. Current solutions rely heavily on prompting proprietary VLMs to generate action labels from videos [23, 24], which is costly and challenging to control, making it difficult to scale for large datasets. We argue that noisy pseudo labels from off-the-shelf visual odometry (VO) models [25] are sufficiently effective for imitation learning in embodied urban navigation. Moreover, this data processing method generalizes beyond walking data. We demonstrate that" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "content": "a model trained with driving videos also performs well on quadruped agents, with further improvements observed when training in a cross-domain setting that combines both city walking and driving data." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 326, + 121, + 524, + 132 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 121, + 524, + 132 + ], + "spans": [ + { + "bbox": [ + 326, + 121, + 524, + 132 + ], + "type": "text", + "content": "We summarize our contribution as the following:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 133, + 553, + 277 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 314, + 133, + 553, + 181 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 133, + 553, + 181 + ], + "spans": [ + { + "bbox": [ + 314, + 133, + 553, + 181 + ], + "type": "text", + "content": "- We identify embodied urban navigation as an intriguing yet unsolved problem. To address this problem, we propose a scalable, data-driven solution by leveraging web-scale city walking and driving videos." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 182, + 553, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 182, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 313, + 182, + 553, + 217 + ], + "type": "text", + "content": "- We introduce a simple yet scalable data processing paradigm, making large-scale imitation learning feasible for urban navigation without extensive manual labeling." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 217, + 553, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 217, + 553, + 277 + ], + "spans": [ + { + "bbox": [ + 314, + 217, + 553, + 277 + ], + "type": "text", + "content": "- We demonstrate training on web-scale data significantly improves navigation performance in real-world experiments, allowing embodied agents to handle the complexities of urban scenarios. We also show model performance scales well with training data size." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 287, + 400, + 300 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 287, + 400, + 300 + ], + "spans": [ + { + "bbox": [ + 314, + 287, + 400, + 300 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 306, + 553, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 306, + 553, + 521 + ], + "spans": [ + { + "bbox": [ + 313, + 306, + 553, + 521 + ], + "type": "text", + "content": "Navigation in Simulation. The introduction of high-performance photorealistic simulators [4-7] has significantly boosted navigation research in indoor household environments. Based on the form of goal specification, navigation tasks are usually categorized into point-goal [4, 8, 26, 27], image-goal [28-31], object-goal [32-34], and vision-language navigation [35-37]. While tasks like image-goal and object-goal navigation focus on efficient goal-finding or instruction understanding, point-goal navigation has a more straightforward goal specification and emphasizes safe and efficient trajectory planning to the goal. Our problem formulation resembles point-goal navigation in that accurate waypoints (represented as GPS coordinates) are readily available from navigation tools. Recently, point-goal navigation has been considered a \"solved\" problem due to near-perfect performance achieved in simulators [8]. However, our investigation shows that waypoint-goal navigation in public urban spaces remains an unsolved challenge." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 522, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 522, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 522, + 553, + 713 + ], + "type": "text", + "content": "Real-World Navigation. Real-world robot navigation has a long history, predating the deep learning era. Early methods relied on ranging sensors and modular SLAM systems for navigation [1-3]. Witnessing the success in simulators, various methods have been proposed to bridge the sim-to-real gap [38-41]. Another thread of work learns navigation policies from expert trajectories [42, 43]. By incorporating data from different embodiments [14] and carefully designed heuristics [13], these methods have been extended to long-horizon navigation on different robots [44, 45]. While deployed on real-world embodied agents, these methods still operate in relatively simple and static environments such as suburban areas, parking garages, or controlled obstacle courses. They are not guaranteed to generalize to significantly different, dynamic, and complex urban landscapes. Our work follows this line of research by training naviga" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "6876" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 85, + 72, + 525, + 302 + ], + "blocks": [ + { + "bbox": [ + 85, + 72, + 525, + 302 + ], + "lines": [ + { + "bbox": [ + 85, + 72, + 525, + 302 + ], + "spans": [ + { + "bbox": [ + 85, + 72, + 525, + 302 + ], + "type": "image", + "image_path": "6a7facbf87e4bdf11ed8d4bd2974e20170545956d39babae1641fd92bed4f3e7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 310, + 555, + 365 + ], + "lines": [ + { + "bbox": [ + 55, + 310, + 555, + 365 + ], + "spans": [ + { + "bbox": [ + 55, + 310, + 555, + 365 + ], + "type": "text", + "content": "Figure 2. Overall Illustration of CityWalker. Our training pipeline starts with internet-sourced videos, using visual odometry to obtain relative poses between frames. At each time step, the model receives past observations, past trajectory, and target location as input. They are encoded via a frozen image encoder and a trainable coordinate encoder. A transformer processes these inputs to generate future tokens. An action head and an arrival head decode these tokens into action and arrival status predictions. During training, future frame tokens from future frames guide the transformer to hallucinate future tokens." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 375, + 295, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 375, + 295, + 422 + ], + "spans": [ + { + "bbox": [ + 55, + 375, + 295, + 422 + ], + "type": "text", + "content": "tion models with imitation learning. But we focus on learning the rules and norms essential for urban environments, which is crucial for enabling applications such as delivery robots and robot-taxis." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 424, + 295, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 424, + 295, + 639 + ], + "spans": [ + { + "bbox": [ + 55, + 424, + 295, + 639 + ], + "type": "text", + "content": "In-the-Wild Video Learning. Learning from web-scale data has proven successful in language and vision tasks [19-21, 46]. While some works utilize in-the-wild videos to extract intermediate representations, priors, or reward functions [47, 48], a key challenge remains: the lack of action labels. Patel et al. [49] use a reconstruction pipeline to obtain hand-object trajectories as action labels for learning robot manipulation. SelfD [50] argues BEV model is better than VO in pseudo-labeling, while we show VO works well with large-scale data. Recent works also rely on prompting closed VLMs to get action labels [23, 51]. LeLaN [24], a concurrent work most relevant to ours, also relies on VLM prompting and uses pretrained navigation models to generate action labels. In contrast, we process in-the-wild video data differently by using only the noisy pseudolabels from off-the-shelf visual SLAM tools [25]. This approach enables scalable and cost-effective action label generation for imitation learning in urban navigation tasks." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 651, + 218, + 665 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 651, + 218, + 665 + ], + "spans": [ + { + "bbox": [ + 55, + 651, + 218, + 665 + ], + "type": "text", + "content": "3. Embodied Urban Navigation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 671, + 180, + 683 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 671, + 180, + 683 + ], + "spans": [ + { + "bbox": [ + 55, + 671, + 180, + 683 + ], + "type": "text", + "content": "3.1. Problem Formulation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "content": "In this work, we address the problem of embodied urban navigation for agents in dynamic urban environments. The" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 375, + 555, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 375, + 555, + 518 + ], + "spans": [ + { + "bbox": [ + 313, + 375, + 555, + 518 + ], + "type": "text", + "content": "agent's task is to navigate from its current location to a specified target waypoint location, following a series of waypoints provided by a navigation tool. Formally, we define this as a point-goal navigation problem in real-world urban settings. At each time step " + }, + { + "bbox": [ + 313, + 375, + 555, + 518 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 375, + 555, + 518 + ], + "type": "text", + "content": ", the agent receives an RGB observation " + }, + { + "bbox": [ + 313, + 375, + 555, + 518 + ], + "type": "inline_equation", + "content": "o_t" + }, + { + "bbox": [ + 313, + 375, + 555, + 518 + ], + "type": "text", + "content": ", its current GPS location " + }, + { + "bbox": [ + 313, + 375, + 555, + 518 + ], + "type": "inline_equation", + "content": "p_t" + }, + { + "bbox": [ + 313, + 375, + 555, + 518 + ], + "type": "text", + "content": ", and a sub-goal waypoint " + }, + { + "bbox": [ + 313, + 375, + 555, + 518 + ], + "type": "inline_equation", + "content": "w_t" + }, + { + "bbox": [ + 313, + 375, + 555, + 518 + ], + "type": "text", + "content": ". The agent aims to learn a policy " + }, + { + "bbox": [ + 313, + 375, + 555, + 518 + ], + "type": "inline_equation", + "content": "\\pi(a_t \\mid o_{(t-k):t}, p_{(t-k):t}, w_t)" + }, + { + "bbox": [ + 313, + 375, + 555, + 518 + ], + "type": "text", + "content": " that maps the past observations and positional information to an action " + }, + { + "bbox": [ + 313, + 375, + 555, + 518 + ], + "type": "inline_equation", + "content": "a_t" + }, + { + "bbox": [ + 313, + 375, + 555, + 518 + ], + "type": "text", + "content": " from the action space " + }, + { + "bbox": [ + 313, + 375, + 555, + 518 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 313, + 375, + 555, + 518 + ], + "type": "text", + "content": ", represented as a series of action waypoints in Euclidean space. Typically, we take " + }, + { + "bbox": [ + 313, + 375, + 555, + 518 + ], + "type": "inline_equation", + "content": "k = 5" + }, + { + "bbox": [ + 313, + 375, + 555, + 518 + ], + "type": "text", + "content": " and also predict actions for the next 5 time steps." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 519, + 556, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 519, + 556, + 616 + ], + "spans": [ + { + "bbox": [ + 313, + 519, + 556, + 616 + ], + "type": "text", + "content": "As accurate waypoints are easily available from navigation tools (e.g. Google Maps API), we focus our problem on the navigation between two consecutive waypoints. The agent treats the current sub-goal " + }, + { + "bbox": [ + 313, + 519, + 556, + 616 + ], + "type": "inline_equation", + "content": "w_{t}" + }, + { + "bbox": [ + 313, + 519, + 556, + 616 + ], + "type": "text", + "content": " as the immediate target and is responsible for determining when it has reached this sub-goal based on its observations and positional data. Upon reaching " + }, + { + "bbox": [ + 313, + 519, + 556, + 616 + ], + "type": "inline_equation", + "content": "w_{t}" + }, + { + "bbox": [ + 313, + 519, + 556, + 616 + ], + "type": "text", + "content": ", the agent proceeds to the next waypoint in the sequence." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 624, + 427, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 624, + 427, + 635 + ], + "spans": [ + { + "bbox": [ + 313, + 624, + 427, + 635 + ], + "type": "text", + "content": "3.2. Evaluation Metrics" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 642, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 555, + 715 + ], + "type": "text", + "content": "To enable analysis and assessment of different methods on the problem, our evaluation metrics are twofold. We use the ground truth trajectory in teleoperation-collected data to provide a thorough analysis of each method's performance across various scenarios. Below, we discuss the metrics we used in offline data evaluation. For real-world deployment," + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6877" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 64, + 175, + 190 + ], + "blocks": [ + { + "bbox": [ + 57, + 64, + 175, + 190 + ], + "lines": [ + { + "bbox": [ + 57, + 64, + 175, + 190 + ], + "spans": [ + { + "bbox": [ + 57, + 64, + 175, + 190 + ], + "type": "image", + "image_path": "dadde6cde915baf037f3e33ff4541f0b884cbd63c0971954b01df46fa4829efc.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 193, + 296, + 270 + ], + "lines": [ + { + "bbox": [ + 55, + 193, + 296, + 270 + ], + "spans": [ + { + "bbox": [ + 55, + 193, + 296, + 270 + ], + "type": "text", + "content": "Figure 3. Evaluation Metrics. Left: The orientation error is defined to be the angle between each predicted and ground truth action pair, labeled by " + }, + { + "bbox": [ + 55, + 193, + 296, + 270 + ], + "type": "inline_equation", + "content": "\\theta_{n}" + }, + { + "bbox": [ + 55, + 193, + 296, + 270 + ], + "type": "text", + "content": " in the figure. The action angle " + }, + { + "bbox": [ + 55, + 193, + 296, + 270 + ], + "type": "inline_equation", + "content": "\\varphi_{\\mathrm{action}}" + }, + { + "bbox": [ + 55, + 193, + 296, + 270 + ], + "type": "text", + "content": " and target angle " + }, + { + "bbox": [ + 55, + 193, + 296, + 270 + ], + "type": "inline_equation", + "content": "\\varphi_{\\mathrm{target}}" + }, + { + "bbox": [ + 55, + 193, + 296, + 270 + ], + "type": "text", + "content": " are defined with respect to the positive y-axis. Right: Both green and red trajectories are predicted actions. The green trajectory is the preferred one, having a large L2 distance but a small AOE. Vice versa for the red trajectory." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 175, + 64, + 293, + 190 + ], + "blocks": [ + { + "bbox": [ + 175, + 64, + 293, + 190 + ], + "lines": [ + { + "bbox": [ + 175, + 64, + 293, + 190 + ], + "spans": [ + { + "bbox": [ + 175, + 64, + 293, + 190 + ], + "type": "image", + "image_path": "d957e4dc9fa6bdda9adae73a099f6720b20a085f50a68eca2fb063f5f5f94cb7.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 156, + 555, + 212 + ], + "lines": [ + { + "bbox": [ + 313, + 156, + 555, + 212 + ], + "spans": [ + { + "bbox": [ + 313, + 156, + 555, + 212 + ], + "type": "text", + "content": "Figure 4. Data Sample and Visual Odometry (VO) Result. Our internet-source training data includes both walking and driving videos. These videos cover various scenarios in the urban environment. The VO tool gives noisy trajectories globally, but trustworthy local relative pose within a short time period." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 280, + 295, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 280, + 295, + 304 + ], + "spans": [ + { + "bbox": [ + 55, + 280, + 295, + 304 + ], + "type": "text", + "content": "we primarily use navigation success rate as the key metric for performance comparison." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 304, + 296, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 304, + 296, + 460 + ], + "spans": [ + { + "bbox": [ + 55, + 304, + 296, + 460 + ], + "type": "text", + "content": "Average Orientation Error. We use the error between each predicted action and the ground truth action as the main evaluation metric on the offline teleoperation data. While L2-distance in Euclidean space is a straightforward metric, we find it does not adequately capture the problem. As shown on the right side of Fig. 3, the red trajectory has a smaller L2 error compared to the green trajectory but moves in a direction that deviates from the target. This observation motivates us to come up with a new evaluation metric that better reflects the quality of the predicted actions. We define average orientation error (AOE) as the angle difference between the predicted and ground truth actions, averaged across all samples at each step (left side of Fig. 3):" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 66, + 468, + 295, + 500 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 468, + 295, + 500 + ], + "spans": [ + { + "bbox": [ + 66, + 468, + 295, + 500 + ], + "type": "interline_equation", + "content": "\\operatorname {A O E} (k) = \\frac {1}{n} \\sum_ {i} ^ {n} \\theta_ {i _ {k}} = \\frac {1}{n} \\sum_ {i} ^ {n} \\arccos \\frac {\\left\\langle a _ {\\hat {i} _ {k}} ^ {\\prime} , a _ {i _ {k}} \\right\\rangle}{\\| a _ {\\hat {i} _ {k}} \\| \\| a _ {i _ {k}} \\|}, \\tag {1}", + "image_path": "4442f7a7786785be3a49e8b5ce77f2b86afbe39eb73ed019b5e545bab7cb84aa.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 509, + 296, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 509, + 296, + 581 + ], + "spans": [ + { + "bbox": [ + 55, + 509, + 296, + 581 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 509, + 296, + 581 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 509, + 296, + 581 + ], + "type": "text", + "content": " is the index of predicted action and " + }, + { + "bbox": [ + 55, + 509, + 296, + 581 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 509, + 296, + 581 + ], + "type": "text", + "content": " is the number of data. Additionally, we observe that taking the mean across " + }, + { + "bbox": [ + 55, + 509, + 296, + 581 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 509, + 296, + 581 + ], + "type": "text", + "content": " predicted future time steps can lead to an underestimation due to steps with very small errors. Hence, we propose using the maximum average orientation error (MAOE) over all predicted actions as an overall estimation:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 121, + 589, + 295, + 620 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 589, + 295, + 620 + ], + "spans": [ + { + "bbox": [ + 121, + 589, + 295, + 620 + ], + "type": "interline_equation", + "content": "\\mathrm {M A O E} = \\frac {1}{n} \\sum_ {i} ^ {n} \\max _ {k} \\theta_ {i _ {k}}. \\tag {2}", + "image_path": "3fa8408457a70a72eb5b51d6a08799603b379c7b79718838003b7efed72b133e.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 629, + 295, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 295, + 664 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 295, + 664 + ], + "type": "text", + "content": "As orientation error might be noisy when action distance is small, we also use L2-distance as a complimentary evaluation metric." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "Critical Scenarios. Not all time steps are equally critical in a trajectory. For example, errors at road crossings or turning points are more likely to result in navigation failure compared to mistakes during forward movement in open areas." + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 317, + 58, + 556, + 148 + ], + "blocks": [ + { + "bbox": [ + 317, + 58, + 556, + 148 + ], + "lines": [ + { + "bbox": [ + 317, + 58, + 556, + 148 + ], + "spans": [ + { + "bbox": [ + 317, + 58, + 556, + 148 + ], + "type": "image", + "image_path": "8cd6e988b5616c6a39f97a64c82c9d0084ca7375cb3a91e95dfe052cb4bc82e0.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 221, + 554, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 221, + 554, + 281 + ], + "spans": [ + { + "bbox": [ + 313, + 221, + 554, + 281 + ], + "type": "text", + "content": "We identify several key scenarios that are most critical during navigation. AOE and MAOE are calculated separately for each scenario, and we use the scenario mean rather than the sample mean for a more comprehensive evaluation. Detailed definitions of each scenario are provided in Sec. 4." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 289, + 499, + 301 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 289, + 499, + 301 + ], + "spans": [ + { + "bbox": [ + 313, + 289, + 499, + 301 + ], + "type": "text", + "content": "3.3. Learning from In-the-Wild Videos" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 306, + 555, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 306, + 555, + 485 + ], + "spans": [ + { + "bbox": [ + 313, + 306, + 555, + 485 + ], + "type": "text", + "content": "City Walking Videos. To facilitate large-scale imitation learning for human-like urban navigation, we utilize city walking videos sourced from the internet. Walking in urban environments is intuitive for humans, resulting in an abundance of first-person view videos available online. These videos naturally capture the complexities of navigating dynamic urban landscapes, including interactions with pedestrians, adherence to traffic signals, and maneuvering around various obstacles. Motivated by the extensive availability of such data and its inherent alignment with our navigation objectives, we curate a dataset comprising over 2000 hours of city walking videos, encompassing diverse weather and lighting conditions. These videos exhibit consistent motion patterns characteristic of walking, providing continuous motion cues that can directly serve as action labels." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 486, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 486, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 486, + 556, + 714 + ], + "type": "text", + "content": "Action Labels from Videos. To generate actionable labels for imitation learning, we employ VO [25, 52-54] tools to extract trajectory poses from the collected city walking videos. Specifically, we utilize the state-of-the-art DPVO [25] while any other VO method should also work. The resulting VO trajectories are illustrated in Fig. 4. However, using VO poses presents two primary challenges. First, global trajectory accuracy is compromised by accumulation errors inherent in VO systems. Nevertheless, since our model focuses on predicting a limited sequence of actions (5 past and 5 future), we mitigate this issue by relying solely on relative poses within short temporal windows. Second, VO methods suffer from scale ambiguity, leading to inconsistencies in trajectory lengths across different data. For instance, as shown in Fig. 4, a driving video may exhibit a longer actual trajectory compared to a walking video, yet the DPVO output incorrectly reflects a shorter trajectory. To address this, we normalize each action by the average step length within its trajectory, following the approach in" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6878" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": "previous work [14]. This normalization resolves both scale ambiguity and inconsistencies in step lengths between different types of embodiments all at once. This enables the learning of policies in a consistent, abstracted action space. During deployment, the predicted actions are denormalized using the specific step length of the target robot, ensuring appropriate movement execution in real-world settings." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 157, + 295, + 347 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 157, + 295, + 347 + ], + "spans": [ + { + "bbox": [ + 55, + 157, + 295, + 347 + ], + "type": "text", + "content": "Generalizability and Scalability. Our data processing pipeline is both generalizable and scalable. It can be applied to any video sources beyond city walking, as long as the videos are captured with egomotion, such as driving videos, which are also abundantly available online. This flexibility enables the training of more generalizable navigation policies that can be zero-shot applied or fine-tuned for cross-domain and cross-embediment tasks, such as transitioning from pedestrian to vehicular navigation. Furthermore, our method offers significant scalability advantages. Unlike pipelines that rely on prompting large language models for action label generation [23, 24], our visual odometry-based approach can be easily parallelized. Processing 2000 hours of video data with our pipeline requires negligible wall-clock time, making it both feasible and cost-efficient to scale up the training dataset." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 354, + 180, + 366 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 354, + 180, + 366 + ], + "spans": [ + { + "bbox": [ + 55, + 354, + 180, + 366 + ], + "type": "text", + "content": "3.4. Pipeline and Training" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 371, + 295, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 371, + 295, + 491 + ], + "spans": [ + { + "bbox": [ + 55, + 371, + 295, + 491 + ], + "type": "text", + "content": "Pipeline Architecture. Figure 2 shows our pipeline. It builds upon previous works [44, 45]. At the core of our model is a transformer that processes a temporal sequence of input tokens, including " + }, + { + "bbox": [ + 55, + 371, + 295, + 491 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 371, + 295, + 491 + ], + "type": "text", + "content": " image features and one coordinate embeddings. Specifically, the coordinate input is stacked by " + }, + { + "bbox": [ + 55, + 371, + 295, + 491 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 371, + 295, + 491 + ], + "type": "text", + "content": " past positions and one single target position, and outputs a sequence of the same length. We use an action head and an arrival prediction head to decode the output of the transformer. We utilize DINOv2 [21] representation and freeze the backbone during training." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 491, + 295, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 491, + 295, + 586 + ], + "spans": [ + { + "bbox": [ + 55, + 491, + 295, + 586 + ], + "type": "text", + "content": "Feature Hallucination. Inspired by previous works on feature learning [55-57], we use feature hallucination as an auxiliary loss during training. Specifically, we compute the MSE loss between the output image tokens and that directly extracted from future frames. We hope this will guide the transformer to predict more informative future tokens that best mimic future frames, which in turn helps the two MLP heads predict the final action and arrival status." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 586, + 295, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 586, + 295, + 658 + ], + "spans": [ + { + "bbox": [ + 55, + 586, + 295, + 658 + ], + "type": "text", + "content": "Loss Functions. As mentioned in Sec. 3.2, L1 or L2 loss can be less effective than orientation error in evaluating the quality of the predicted actions. We also employ orientation error to supervise the training. We define the orientation loss to be the negative cosine similarity between the predicted actions and the ground truth actions:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 121, + 664, + 295, + 696 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 664, + 295, + 696 + ], + "spans": [ + { + "bbox": [ + 121, + 664, + 295, + 696 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {o r i}} = - \\frac {1}{k} \\sum_ {i = 1} ^ {k} \\frac {\\left\\langle \\hat {a} _ {i} , a _ {i} \\right\\rangle}{\\left\\| \\hat {a} _ {i} \\right\\| \\left\\| a _ {i} \\right\\|}. \\tag {3}", + "image_path": "dc8dc50e38b27d1672bbf4cb4eb9b56d37b963ed1a5c23c96ba120a7c8e45ea1.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 701, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 701, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 701, + 295, + 713 + ], + "type": "text", + "content": "Together with the L1 loss for predicted action and the BCE" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "type": "text", + "content": "loss for predicted arriving status, the final loss function is a weight sum of four individual losses:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 330, + 106, + 555, + 119 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 330, + 106, + 555, + 119 + ], + "spans": [ + { + "bbox": [ + 330, + 106, + 555, + 119 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\omega_ {l 1} \\mathcal {L} _ {l 1} + \\omega_ {\\text {o r i}} \\mathcal {L} _ {\\text {o r i}} + \\omega_ {\\text {a r r}} \\mathcal {L} _ {\\text {a r r}} + \\omega_ {\\text {f e a t}} \\mathcal {L} _ {\\text {f e a t}}. \\tag {4}", + "image_path": "17641e0000ae03bc0ddb794fd6d34b237f265ba2c57d4c6d2862bf60a5705aa0.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 128, + 555, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 128, + 555, + 152 + ], + "spans": [ + { + "bbox": [ + 313, + 128, + 555, + 152 + ], + "type": "text", + "content": "We choose each weight so that each loss items are within the same magnitude." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 163, + 391, + 176 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 163, + 391, + 176 + ], + "spans": [ + { + "bbox": [ + 313, + 163, + 391, + 176 + ], + "type": "text", + "content": "4. Experiment" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 182, + 365, + 194 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 182, + 365, + 194 + ], + "spans": [ + { + "bbox": [ + 313, + 182, + 365, + 194 + ], + "type": "text", + "content": "4.1. Setup" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 199, + 555, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 199, + 555, + 331 + ], + "spans": [ + { + "bbox": [ + 313, + 199, + 555, + 331 + ], + "type": "text", + "content": "Baselines. We compare our model with outdoor navigation models closely related to our setup, including GNM [14], ViNT [44], and NoMaD [45]. Although originally developed for image-goal navigation, we tested them using goal-images from our collected data. Additionally, ViNT as a foundation model for visual navigation supports finetuning, which we also evaluated. We acknowledge CoNVOI [58] as a recent work with a similar setup, but could not test it due to the lack of open source code and prompt. We also experimented with VLM-based approaches similar to CoNVOI and defer the results to the appendix." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 331, + 554, + 463 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 331, + 554, + 463 + ], + "spans": [ + { + "bbox": [ + 313, + 331, + 554, + 463 + ], + "type": "text", + "content": "Data. We collected expert data via teleoperation for fine-tuning and offline testing. We collect data via a Unitree Go1 quadruped equipped with a Livox Mid-360 LiDAR and a webcam for RGB observations. We employed the LiDAR-SLAM method [59] to obtain the robot's poses as ground truth actions. Additionally, we retrieved location data from a smartphone using a webpage. We reduce noise from the quadruped's vibrations by holding the phone [38]. In total, we gathered 15 hours of teleoperation data across various areas in New York City, allocating 6 hours for fine-tuning and 9 hours for testing." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 463, + 554, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 463, + 554, + 521 + ], + "spans": [ + { + "bbox": [ + 313, + 463, + 554, + 521 + ], + "type": "text", + "content": "Critical Scenarios As discussed in Sec. 3.2, we identify several critical scenarios in our dataset that we want to emphasize in our evaluation. We analyze the ground truth trajectory and run object detection [60] to aid us in standardizing the definition. We define these scenarios as follows:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 522, + 553, + 653 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 314, + 522, + 553, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 522, + 553, + 545 + ], + "spans": [ + { + "bbox": [ + 314, + 522, + 553, + 545 + ], + "type": "text", + "content": "- Turn: When the ground truth action changes direction significantly. Defined when " + }, + { + "bbox": [ + 314, + 522, + 553, + 545 + ], + "type": "inline_equation", + "content": "\\varphi_{\\mathrm{action}} > 20^{\\circ}" + }, + { + "bbox": [ + 314, + 522, + 553, + 545 + ], + "type": "text", + "content": " (see Fig. 3)." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 314, + 546, + 553, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 546, + 553, + 569 + ], + "spans": [ + { + "bbox": [ + 314, + 546, + 553, + 569 + ], + "type": "text", + "content": "- Crossing: When the agent is at a road crossing. Defined if any traffic light is detected with score " + }, + { + "bbox": [ + 314, + 546, + 553, + 569 + ], + "type": "inline_equation", + "content": ">0.5" + }, + { + "bbox": [ + 314, + 546, + 553, + 569 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 314, + 570, + 553, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 570, + 553, + 594 + ], + "spans": [ + { + "bbox": [ + 314, + 570, + 553, + 594 + ], + "type": "text", + "content": "- Detour: When the action angle deviates from the target angle. Defined when " + }, + { + "bbox": [ + 314, + 570, + 553, + 594 + ], + "type": "inline_equation", + "content": "|\\varphi_{\\text{action}} - \\varphi_{\\text{target}}| > 45^{\\circ}" + }, + { + "bbox": [ + 314, + 570, + 553, + 594 + ], + "type": "text", + "content": " (see Fig. 3)." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 314, + 594, + 553, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 594, + 553, + 629 + ], + "spans": [ + { + "bbox": [ + 314, + 594, + 553, + 629 + ], + "type": "text", + "content": "- Proximity: When any person is close to the agent. Defined when the largest bounding box of any detected person is larger than " + }, + { + "bbox": [ + 314, + 594, + 553, + 629 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 314, + 594, + 553, + 629 + ], + "type": "text", + "content": " of the image area." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 630, + 553, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 553, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 553, + 653 + ], + "type": "text", + "content": "- **Crowd:** When the agent is surrounded by a crowd of people. Defined when " + }, + { + "bbox": [ + 313, + 630, + 553, + 653 + ], + "type": "inline_equation", + "content": "\\geq 5" + }, + { + "bbox": [ + 313, + 630, + 553, + 653 + ], + "type": "text", + "content": " people are detected." + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "text", + "content": "We define these scenarios in a way that they are not mutually exclusive. A single data sample can belong to multiple scenarios. Although these scenarios account for less than half of the data, we argue they are the most significant factors contributing to successful urban navigation." + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6879" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 107, + 113, + 504, + 317 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 105 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 105 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 105 + ], + "type": "text", + "content": "Table 1. Benchmark on Offline Data. We evaluate three metrics in each critical scenario for all methods. Percentages under scenarios indicate their data proportions. The \"Mean\" column shows scenario means averaged over six scenarios; \"All\" shows sample means over all data samples. We highlight the best performance in bold and the second-best performance with underline." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 107, + 113, + 504, + 317 + ], + "lines": [ + { + "bbox": [ + 107, + 113, + 504, + 317 + ], + "spans": [ + { + "bbox": [ + 107, + 113, + 504, + 317 + ], + "type": "table", + "html": "
MethodMetricMeanTurn 8%Crossing 12%Detour 12%Proximity 6%Crowd 7%Other 55%All 100%
GNM [14] (fine-tuned)↓L2 (m)1.222.361.361.420.880.760.550.74
↓MAOE (°)16.231.114.812.514.712.811.012.1
↑Arrival (%)68.666.469.766.469.069.270.770.0
ViNT [44] (fine-tuned)↓L2 (m)1.301.911.131.140.770.660.570.70
↓MAOE (°)16.531.115.412.914.813.311.612.6
↑Arrival (%)70.571.268.770.773.068.671.070.7
NoMaD [45] (fine-tuned)↓L2 (m)1.392.491.561.551.060.950.760.74
↓MAOE (°)19.135.118.515.618.114.312.812.1
↑Arrival (%)68.666.469.766.469.069.370.770.0
Ours (zero-shot)↓L2 (m)1.341.301.091.331.441.481.391.38
↓MAOE (°)16.526.815.516.316.313.111.312.7
↑Arrival (%)79.169.371.478.884.184.586.484.1
Ours (fine-tuned)↓L2 (m)1.111.271.001.151.061.121.061.07
↓MAOE (°)15.226.614.113.914.312.010.411.5
↑Arrival (%)81.868.975.378.590.687.590.287.8
", + "image_path": "c07603de6a50fc5a23577aaa4a9c99bce841123d79655c139f79e93e600cb4d5.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 335, + 211, + 348 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 335, + 211, + 348 + ], + "spans": [ + { + "bbox": [ + 55, + 335, + 211, + 348 + ], + "type": "text", + "content": "4.2. Performance Benchmarking" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 354, + 295, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 354, + 295, + 413 + ], + "spans": [ + { + "bbox": [ + 55, + 354, + 295, + 413 + ], + "type": "text", + "content": "In this subsection, we aim to answer the question Q1: Can our CityWalker model navigate successfully in complex urban environments? To answer this, we benchmark our model and the baselines both with offline data and in real-world navigation deployment." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 414, + 295, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 414, + 295, + 558 + ], + "spans": [ + { + "bbox": [ + 55, + 414, + 295, + 558 + ], + "type": "text", + "content": "Overall Benchmark. Table 1 shows the benchmark in different critical scenarios. Overall, our fine-tuned model has superior performance on all metrics in all scenarios except \"detour\". We think the less competent performance in the detour scenario is due to the small portion of such data in our video training data. It can be verified by the fact that the fine-tuning performance gain is most significant in this scenario. It is worth noting that our zero-shot model has similar or even better performance compared to the fine-tuned baselines. This demonstrates the power of large-scale pretraining, especially considering the domain and modality gap from human walking videos." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 558, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 295, + 714 + ], + "type": "text", + "content": "Real-world Depolyment. We deploy our model and the baseline methods on the same Unitree Go1 quadruped for real-world navigation, in previously unseen urban environments. We adapt the PD controller provided in [44] to provide velocity command to the quadruped. We choose area with relatively constant density of dynamics such as pedestrian and traffic to ensure the consistency among different trials. In each trial, the target location is generally 50-100m far away from the starting location. We categorize the real-world test into forward, left turn, and right turn cases to provide case-by-case analysis. We conduct 8-14 trails for each cases. Each trial is marked success when the model predicts arrival within a 5m distance to the target location." + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 328, + 371, + 539, + 437 + ], + "blocks": [ + { + "bbox": [ + 313, + 335, + 555, + 368 + ], + "lines": [ + { + "bbox": [ + 313, + 335, + 555, + 368 + ], + "spans": [ + { + "bbox": [ + 313, + 335, + 555, + 368 + ], + "type": "text", + "content": "Table 2. Real-World Navigation. The table shows the success rate of real-world experiments in different scenarios. * and † indicate zero-shot inference and fine-tuned models respectively." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 328, + 371, + 539, + 437 + ], + "lines": [ + { + "bbox": [ + 328, + 371, + 539, + 437 + ], + "spans": [ + { + "bbox": [ + 328, + 371, + 539, + 437 + ], + "type": "table", + "html": "
MethodAllForwardLeft turnRight turn
ViNT* [44]37.762.50.050.0
ViNT† [44]57.1100.025.025.0
NoMad* [45]42.975.016.728.6
Ours†77.3100.062.566.7
", + "image_path": "46e348324a1c7eaeb2891907e89c6679d48390f533d5a586ab0acbbc3d75e7d2.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 446, + 553, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 446, + 553, + 470 + ], + "spans": [ + { + "bbox": [ + 313, + 446, + 553, + 470 + ], + "type": "text", + "content": "All human interruptions due to potential collision or timeout are treated as failure cases." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 471, + 555, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 471, + 555, + 662 + ], + "spans": [ + { + "bbox": [ + 313, + 471, + 555, + 662 + ], + "type": "text", + "content": "Table 2 shows the success rate on all real-world test cases. Our model achieves the highest success rate among all cases with a significant gap to the second-best performing finetuned ViNT. As the baselines are trained with navigation data mostly in suburban or off-road environments [14, 61], they are limited in handling complex maneuvers, making them unsuitable for urban navigation tasks, as shown in Fig. 5. Our model's robust performance in both forward and turns highlights its ability to manage dynamic and varied urban scenarios effectively, allowing it to be more reliably used in real-world environments where frequent and precise directional changes are essential. The substantial improvement not only increases the overall navigation success but also extends the applicability of our model to more demanding and realistic urban settings, positioning it as a superior solution for embodied agents in dynamic environments." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 671, + 441, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 671, + 441, + 685 + ], + "spans": [ + { + "bbox": [ + 313, + 671, + 441, + 685 + ], + "type": "text", + "content": "4.3. Power of Data Scaling" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 689, + 554, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 715 + ], + "type": "text", + "content": "Our model is significantly boosted by the large-scale training data. We ask Q2: To what extent does increasing the" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6880" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 71, + 89, + 164, + 182 + ], + "blocks": [ + { + "bbox": [ + 75, + 72, + 157, + 82 + ], + "lines": [ + { + "bbox": [ + 75, + 72, + 157, + 82 + ], + "spans": [ + { + "bbox": [ + 75, + 72, + 157, + 82 + ], + "type": "text", + "content": "Current Observation" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 71, + 89, + 164, + 182 + ], + "lines": [ + { + "bbox": [ + 71, + 89, + 164, + 182 + ], + "spans": [ + { + "bbox": [ + 71, + 89, + 164, + 182 + ], + "type": "image", + "image_path": "e2e21c1fff13ed9f1ecb15548d16055943dbfa72f0c842560a92675de7264991.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 168, + 86, + 260, + 182 + ], + "blocks": [ + { + "bbox": [ + 168, + 86, + 260, + 182 + ], + "lines": [ + { + "bbox": [ + 168, + 86, + 260, + 182 + ], + "spans": [ + { + "bbox": [ + 168, + 86, + 260, + 182 + ], + "type": "image", + "image_path": "7a65bb3bfa3d90537fdabf04d4bf3837becdff4952ca5f0d3c3f4873e749e697.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 207, + 182, + 235, + 193 + ], + "lines": [ + { + "bbox": [ + 207, + 182, + 235, + 193 + ], + "spans": [ + { + "bbox": [ + 207, + 182, + 235, + 193 + ], + "type": "text", + "content": "X (m) Coordinates" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 260, + 84, + 350, + 182 + ], + "blocks": [ + { + "bbox": [ + 193, + 72, + 253, + 81 + ], + "lines": [ + { + "bbox": [ + 193, + 72, + 253, + 81 + ], + "spans": [ + { + "bbox": [ + 193, + 72, + 253, + 81 + ], + "type": "text", + "content": "ViNT Zero-shot" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 282, + 72, + 342, + 81 + ], + "lines": [ + { + "bbox": [ + 282, + 72, + 342, + 81 + ], + "spans": [ + { + "bbox": [ + 282, + 72, + 342, + 81 + ], + "type": "text", + "content": "ViNT Fine-tune" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 260, + 84, + 350, + 182 + ], + "lines": [ + { + "bbox": [ + 260, + 84, + 350, + 182 + ], + "spans": [ + { + "bbox": [ + 260, + 84, + 350, + 182 + ], + "type": "image", + "image_path": "0fe13c8cae1a774f18e9382a452b96400165389aca4549183e61382c09bba991.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 300, + 182, + 327, + 193 + ], + "lines": [ + { + "bbox": [ + 300, + 182, + 327, + 193 + ], + "spans": [ + { + "bbox": [ + 300, + 182, + 327, + 193 + ], + "type": "text", + "content": "X (m) Coordinates" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 351, + 85, + 444, + 182 + ], + "blocks": [ + { + "bbox": [ + 372, + 72, + 442, + 81 + ], + "lines": [ + { + "bbox": [ + 372, + 72, + 442, + 81 + ], + "spans": [ + { + "bbox": [ + 372, + 72, + 442, + 81 + ], + "type": "text", + "content": "NoMad Zero-shot" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 351, + 85, + 444, + 182 + ], + "lines": [ + { + "bbox": [ + 351, + 85, + 444, + 182 + ], + "spans": [ + { + "bbox": [ + 351, + 85, + 444, + 182 + ], + "type": "image", + "image_path": "5b2f72ea50dc2532803c1996d73c3040eb534f21fe04ebadf5490c50cfe3fa3c.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 391, + 182, + 420, + 193 + ], + "lines": [ + { + "bbox": [ + 391, + 182, + 420, + 193 + ], + "spans": [ + { + "bbox": [ + 391, + 182, + 420, + 193 + ], + "type": "text", + "content": "X (m) Coordinates" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 447, + 85, + 536, + 186 + ], + "blocks": [ + { + "bbox": [ + 487, + 72, + 508, + 82 + ], + "lines": [ + { + "bbox": [ + 487, + 72, + 508, + 82 + ], + "spans": [ + { + "bbox": [ + 487, + 72, + 508, + 82 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 447, + 85, + 536, + 186 + ], + "lines": [ + { + "bbox": [ + 447, + 85, + 536, + 186 + ], + "spans": [ + { + "bbox": [ + 447, + 85, + 536, + 186 + ], + "type": "image", + "image_path": "33380879e63d2d78c8e92dea8df834cd596485aa2832d9e2d3a684ed65b228fc.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 71, + 193, + 165, + 285 + ], + "blocks": [ + { + "bbox": [ + 71, + 193, + 165, + 285 + ], + "lines": [ + { + "bbox": [ + 71, + 193, + 165, + 285 + ], + "spans": [ + { + "bbox": [ + 71, + 193, + 165, + 285 + ], + "type": "image", + "image_path": "7341e9001dee630a1d8f34b6f08cc51d7117895d198d9f5e56519b34cd167117.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 142, + 293, + 196, + 301 + ], + "lines": [ + { + "bbox": [ + 142, + 293, + 196, + 301 + ], + "spans": [ + { + "bbox": [ + 142, + 293, + 196, + 301 + ], + "type": "text", + "content": "Input past traj." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 170, + 193, + 259, + 290 + ], + "blocks": [ + { + "bbox": [ + 170, + 193, + 259, + 290 + ], + "lines": [ + { + "bbox": [ + 170, + 193, + 259, + 290 + ], + "spans": [ + { + "bbox": [ + 170, + 193, + 259, + 290 + ], + "type": "image", + "image_path": "deafe897752aa4918eaeeb125d39a1b2198df2be710e84b0fa980e6319617fd4.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 219, + 293, + 276, + 300 + ], + "lines": [ + { + "bbox": [ + 219, + 293, + 276, + 300 + ], + "spans": [ + { + "bbox": [ + 219, + 293, + 276, + 300 + ], + "type": "text", + "content": "Actual past traj." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 262, + 193, + 351, + 290 + ], + "blocks": [ + { + "bbox": [ + 262, + 193, + 351, + 290 + ], + "lines": [ + { + "bbox": [ + 262, + 193, + 351, + 290 + ], + "spans": [ + { + "bbox": [ + 262, + 193, + 351, + 290 + ], + "type": "image", + "image_path": "2076933fbec325f325f7415aa8b0a289875b223f6c67030a84a69dd66368d491.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 354, + 193, + 444, + 291 + ], + "blocks": [ + { + "bbox": [ + 299, + 293, + 358, + 300 + ], + "lines": [ + { + "bbox": [ + 299, + 293, + 358, + 300 + ], + "spans": [ + { + "bbox": [ + 299, + 293, + 358, + 300 + ], + "type": "text", + "content": "Predicted action" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 354, + 193, + 444, + 291 + ], + "lines": [ + { + "bbox": [ + 354, + 193, + 444, + 291 + ], + "spans": [ + { + "bbox": [ + 354, + 193, + 444, + 291 + ], + "type": "image", + "image_path": "7503e3b96efdae1061cffe23de953bd09a4c58e828f94417e03853306a75f805.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 381, + 293, + 421, + 300 + ], + "lines": [ + { + "bbox": [ + 381, + 293, + 421, + 300 + ], + "spans": [ + { + "bbox": [ + 381, + 293, + 421, + 300 + ], + "type": "text", + "content": "GTaction" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 444, + 293, + 471, + 301 + ], + "lines": [ + { + "bbox": [ + 444, + 293, + 471, + 301 + ], + "spans": [ + { + "bbox": [ + 444, + 293, + 471, + 301 + ], + "type": "inline_equation", + "content": "\\star" + }, + { + "bbox": [ + 444, + 293, + 471, + 301 + ], + "type": "text", + "content": " Target" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 446, + 193, + 536, + 290 + ], + "blocks": [ + { + "bbox": [ + 446, + 193, + 536, + 290 + ], + "lines": [ + { + "bbox": [ + 446, + 193, + 536, + 290 + ], + "spans": [ + { + "bbox": [ + 446, + 193, + 536, + 290 + ], + "type": "image", + "image_path": "ab2d565d131d2703bf9beed86839e64c15a34f1f6a091815665c16d402ed9be5.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 66, + 343, + 278, + 472 + ], + "blocks": [ + { + "bbox": [ + 55, + 307, + 555, + 329 + ], + "lines": [ + { + "bbox": [ + 55, + 307, + 555, + 329 + ], + "spans": [ + { + "bbox": [ + 55, + 307, + 555, + 329 + ], + "type": "text", + "content": "Figure 5. Qualitative Results. Left image shows current observations of two samples. Right plots displays input trajectory, ground truth actions, and predicted actions in the current coordinate system with the agent at the origin." + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 66, + 343, + 278, + 472 + ], + "lines": [ + { + "bbox": [ + 66, + 343, + 278, + 472 + ], + "spans": [ + { + "bbox": [ + 66, + 343, + 278, + 472 + ], + "type": "image", + "image_path": "b77d40fa67f01460f9f1fb4a415fc7fc4e69cd6c17e41b360ff23bfdc32c5063.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 476, + 296, + 533 + ], + "lines": [ + { + "bbox": [ + 55, + 476, + 296, + 533 + ], + "spans": [ + { + "bbox": [ + 55, + 476, + 296, + 533 + ], + "type": "text", + "content": "Figure 6. Performance and Data Size. We show the model performance evaluated by MAOE with respect to the size of the training data measured by video length in hours. We also show the zero-shot performance of our model trained with only driving videos and mixed driving and walking videos" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "bbox": [ + 55, + 554, + 296, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 554, + 296, + 578 + ], + "spans": [ + { + "bbox": [ + 55, + 554, + 296, + 578 + ], + "type": "text", + "content": "size of the training dataset improve the performance of our urban navigation model?" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "content": "Performance Gain and Data Scale. Figure 6 shows the model performance when trained with different data sizes. The plot demonstrates that our model's zero-shot performance improves significantly as we increase the amount of training data. Notably, when trained with more than 1000 hours of videos, our zero-shot model (red line) achieves a better performance than the fine-tuned ViNT model " + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "inline_equation", + "content": "(\\square)" + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "content": ". This indicates that large-scale training on diverse video data can even surpass the performance of models that are fine-tuned with limited expert data, highlighting the potential of leveraging web-scale datasets for embodied navigation." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 313, + 342, + 555, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 342, + 555, + 533 + ], + "spans": [ + { + "bbox": [ + 313, + 342, + 555, + 533 + ], + "type": "text", + "content": "Domain and Embodiment Gap. An unexpected observation is that our model without the feature hallucination loss outperforms the one with it in zero-shot inference. We hypothesize this phenomenon mainly due the domain and embodiment gap between human walking videos and quadruped navigation policies. The feature hallucination loss aims to guide the transformer model to produce output tokens that mimic future observations, essentially encouraging the model to anticipate future visual inputs. This enforcement might inadvertently hinder performance by focusing on predicting future observations based on human motion. The model may learn representations that are less applicable or even misleading for quadruped navigation. Consequently, the feature hallucination loss could introduce a bias that detracts from the model's ability to make accurate navigation decisions in the robot's operational domain." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 313, + 533, + 556, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 556, + 700 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 556, + 700 + ], + "type": "text", + "content": "Feature Hallucination Loss. Fortunately, as shown in Tab. 1, this issue no longer exists after fine-tuning. We also show the benefit of feature hallucination loss from another perspective. Figure 7 shows two training loss curve for different data size and training setups with and without feature hallucination loss. We observe that although feature hallucination loss should increase the absolute value of the total loss, the training with feature hallucination loss lead to a lower total loss for 2000-hour training and a lower orientation loss in both cases. The training with feature hallucination loss also shows a steeper decreasing trend, showing a potential of having even lower error with more training steps. Therefore, this is another proof of the effectiveness of our proposed feature hallucination loss." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 313, + 701, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 701, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 701, + 553, + 714 + ], + "type": "text", + "content": "Driving Videos. As discussed in Sec. 3.3, our data process" + } + ] + } + ], + "index": 30 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "text", + "content": "6881" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 72, + 74, + 272, + 196 + ], + "blocks": [ + { + "bbox": [ + 72, + 74, + 272, + 196 + ], + "lines": [ + { + "bbox": [ + 72, + 74, + 272, + 196 + ], + "spans": [ + { + "bbox": [ + 72, + 74, + 272, + 196 + ], + "type": "image", + "image_path": "1be800e77262e5e99074bb8830b35bd432270f3e860ac7f776841ec62b4a9925.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 206, + 296, + 274 + ], + "lines": [ + { + "bbox": [ + 55, + 206, + 296, + 274 + ], + "spans": [ + { + "bbox": [ + 55, + 206, + 296, + 274 + ], + "type": "text", + "content": "Figure 7. Improvement from Feature Hallucination. The plot shows the orientation loss for different amounts of training data and loss setups. For all four setups, we trian with " + }, + { + "bbox": [ + 55, + 206, + 296, + 274 + ], + "type": "inline_equation", + "content": "\\omega_{\\mathrm{II}} = 1.0, \\omega_{\\mathrm{arr}} = 1.0, \\text{and}, \\omega_{\\mathrm{ori}} = 5.0" + }, + { + "bbox": [ + 55, + 206, + 296, + 274 + ], + "type": "text", + "content": ". The unproportional training steps are due to different batch sizes used for different data sizes. All models are trained for 10 epochs." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 289, + 295, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 289, + 295, + 445 + ], + "spans": [ + { + "bbox": [ + 54, + 289, + 295, + 445 + ], + "type": "text", + "content": "ing pipeline also enables the use of abundant online driving videos. We evaluate models trained exclusively on driving videos and on a combination of walking and driving videos, results shown in Fig. 6. We can observe that the model trained only with driving data " + }, + { + "bbox": [ + 54, + 289, + 295, + 445 + ], + "type": "inline_equation", + "content": "(+)" + }, + { + "bbox": [ + 54, + 289, + 295, + 445 + ], + "type": "text", + "content": " has a similar performance to the zero-shot baselines. As a result, we can think of driving as a form of navigation from a different domain and embodiment. However, combining driving and walking data yields significant performance gains (x). With just 250 hours of mixed data, the model nearly matches the performance of models trained on 1000 hours of walking data alone. This signifies the generalizability and robustness provided by cross-domain cross-embodiment data." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 455, + 211, + 468 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 455, + 211, + 468 + ], + "spans": [ + { + "bbox": [ + 55, + 455, + 211, + 468 + ], + "type": "text", + "content": "4.4. Pipeline component analysis" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 473, + 295, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 473, + 295, + 521 + ], + "spans": [ + { + "bbox": [ + 55, + 473, + 295, + 521 + ], + "type": "text", + "content": "To provide a more in-depth analysis of our model and pipeline, we ask Q3: How do the individual components of our pipeline enhance the performance and reliability of our model for real-world urban navigation?" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 522, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 522, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 522, + 295, + 713 + ], + "type": "text", + "content": "AOE at Different Time Steps. In Tab. 4, we study the AOE at different time steps, i.e. " + }, + { + "bbox": [ + 55, + 522, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{AOE}(k)" + }, + { + "bbox": [ + 55, + 522, + 295, + 713 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 55, + 522, + 295, + 713 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 522, + 295, + 713 + ], + "type": "text", + "content": " from 1 to 5. ViNT exhibits increasing orientation errors as the prediction steps progress, indicating a degradation in accuracy over time. In contrast, our model maintains relatively consistent AOE from step 2 to step 5, which likely contributes to its superior performance in real-world navigation deployments with stable action prediction across multiple future steps. All methods experience higher errors at the first step. We attribute it to the proximity of the initial action to the origin, as minor linear deviations result in substantial angular error. It is worth noting that the MAOE is significantly larger than any individual AOE from steps 1 to 5, suggesting that the peak errors occur at varying time steps across different samples. This variability suggests the potential benefit of taking softmax for orientation errors instead of simple averaging, as it" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 341, + 124, + 529, + 219 + ], + "blocks": [ + { + "bbox": [ + 313, + 71, + 553, + 115 + ], + "lines": [ + { + "bbox": [ + 313, + 71, + 553, + 115 + ], + "spans": [ + { + "bbox": [ + 313, + 71, + 553, + 115 + ], + "type": "text", + "content": "Table 3. Ablation Study. We ablate on the contribution of orientation loss, feature hallucination, and fine-tuning with expert data. We show the MAOE(°) in scenario-mean. All results are trained with 1000-hour walking data." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 341, + 124, + 529, + 219 + ], + "lines": [ + { + "bbox": [ + 341, + 124, + 529, + 219 + ], + "spans": [ + { + "bbox": [ + 341, + 124, + 529, + 219 + ], + "type": "table", + "html": "
Training ComponentsMAOE
Ori. LossFeature Hall.Fine-tuningMean
17.03
✓17.00
✓✓17.02
✓15.23
✓✓15.21
✓✓✓15.16
", + "image_path": "d745a21044632dbe22c9dca2fbeb9e843d5c233f3f29ec5fc0f60955aaed931e.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 317, + 278, + 553, + 337 + ], + "blocks": [ + { + "bbox": [ + 313, + 236, + 555, + 270 + ], + "lines": [ + { + "bbox": [ + 313, + 236, + 555, + 270 + ], + "spans": [ + { + "bbox": [ + 313, + 236, + 555, + 270 + ], + "type": "text", + "content": "Table 4. AOE at Different Time Steps. The table shows MAOE and different AOE of scenario-mean, all in degrees " + }, + { + "bbox": [ + 313, + 236, + 555, + 270 + ], + "type": "inline_equation", + "content": "(^{\\circ})" + }, + { + "bbox": [ + 313, + 236, + 555, + 270 + ], + "type": "text", + "content": ". * and † indicates zero-shot inference and fine-tuned models respectively." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 278, + 553, + 337 + ], + "lines": [ + { + "bbox": [ + 317, + 278, + 553, + 337 + ], + "spans": [ + { + "bbox": [ + 317, + 278, + 553, + 337 + ], + "type": "table", + "html": "
MethodAOE(1)AOE(2)AOE(3)AOE(4)AOE(5)MAOE
ViNT*[44]10.577.939.0910.2111.4517.52
ViNT†[44]11.247.968.408.939.8716.53
Ours*12.918.928.648.468.4316.54
Ours†11.447.777.687.757.9715.23
", + "image_path": "47c1064d91cb27e634db81bf66f87fa716dd384b1a98bbde93e157051efcafc5.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 359, + 520, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 359, + 520, + 371 + ], + "spans": [ + { + "bbox": [ + 313, + 359, + 520, + 371 + ], + "type": "text", + "content": "could better capture the distribution of peak errors." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 375, + 555, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 375, + 555, + 472 + ], + "spans": [ + { + "bbox": [ + 313, + 375, + 555, + 472 + ], + "type": "text", + "content": "Ablation Study. Table 3 examines the individual contributions of different pipeline components to our model's performance. A clear improvement in MAOE when fine-tuning with expert data, demonstrating the effectiveness of using in-domain expert information. The improvement from orientation loss and feature hallucination loss is marginal, within error range of our limited testing data. We defer more detailed ablation study to the appendix." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 494, + 388, + 506 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 494, + 388, + 506 + ], + "spans": [ + { + "bbox": [ + 313, + 494, + 388, + 506 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 518, + 555, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 518, + 555, + 638 + ], + "spans": [ + { + "bbox": [ + 313, + 518, + 555, + 638 + ], + "type": "text", + "content": "Summary. In this work, we addressed embodied urban navigation by training agents with thousands of hours of in-the-wild city walking videos, introducing a scalable pipeline for imitation learning from diverse walking videos. Our experiments show that large-scale, diverse training data significantly enhances navigation performance, surpassing existing methods and demonstrating the benefits of data scaling. Our findings highlight the potential of leveraging abundant online videos to develop robust navigation policies for embodied agents in dynamic urban environments." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 641, + 556, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 556, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 556, + 712 + ], + "type": "text", + "content": "Limitation and Future Work. During real-world deployment, our current system implementation is sensitive to potential large location noises read naively from iPhone location services, which could be improved with better GPS hardware. Future work could also enhance the robustness of the model to such location noises." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6882" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "type": "text", + "content": "Acknowledgment" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 152 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 152 + ], + "type": "text", + "content": "The work was supported by NSF grants 2238968, 2121391, 2322242 and 2345139; and in part through the NYU IT High Performance Computing resources, services, and staff expertise. We also thank Xingyu Liu and Zixuan Hu for their help in data collection." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 161, + 115, + 173 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 161, + 115, + 173 + ], + "spans": [ + { + "bbox": [ + 56, + 161, + 115, + 173 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 180, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 180, + 296, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 180, + 296, + 224 + ], + "spans": [ + { + "bbox": [ + 61, + 180, + 296, + 224 + ], + "type": "text", + "content": "[1] Quirin Muhlbauer, Stefan Sosnowski, Tingting Xu, Tianguang Zhang, Kolja Kuhnlenz, and Martin Buss. Navigation through urban environments by visual perception and interaction. In ICRA, pages 3558-3564. IEEE, 2009. 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 225, + 296, + 269 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 225, + 296, + 269 + ], + "spans": [ + { + "bbox": [ + 61, + 225, + 296, + 269 + ], + "type": "text", + "content": "[2] Rainer Kümmerle, Michael Ruhnke, Bastian Steder, Cyril Stachniss, and Wolfram Burgard. A navigation system for robots operating in crowded urban environments. In ICRA, pages 3225-3232. IEEE, 2013." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 270, + 295, + 313 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 270, + 295, + 313 + ], + "spans": [ + { + "bbox": [ + 61, + 270, + 295, + 313 + ], + "type": "text", + "content": "[3] Yoichi Morales, Alexander Carballo, Eijiro Takeuchi, Atsushi Aburadani, and Takashi Tsubouchi. Autonomous robot navigation in outdoor cluttered pedestrian walkways. Journal of Field Robotics, 26(8):609-635, 2009. 1, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 315, + 296, + 367 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 315, + 296, + 367 + ], + "spans": [ + { + "bbox": [ + 62, + 315, + 296, + 367 + ], + "type": "text", + "content": "[4] Manolis Savva, Abhishek Kadian, Oleksandr Maksymets, Yili Zhao, Erik Wijmans, Bhavana Jain, Julian Straub, Jia Liu, Vladlen Koltun, Jitendra Malik, et al. Habitat: A platform for embodied ai research. In CVPR, pages 9339-9347, 2019. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 369, + 294, + 411 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 369, + 294, + 411 + ], + "spans": [ + { + "bbox": [ + 62, + 369, + 294, + 411 + ], + "type": "text", + "content": "[5] Fei Xia, Amir R Zamir, Zhiyang He, Alexander Sax, Jitendra Malik, and Silvio Savarese. Gibson env: Real-world perception for embodied agents. In CVPR, pages 9068-9079, 2018." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 414, + 295, + 467 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 414, + 295, + 467 + ], + "spans": [ + { + "bbox": [ + 62, + 414, + 295, + 467 + ], + "type": "text", + "content": "[6] Eric Kolve, Roozbeh Mottaghi, Winson Han, Eli VanderBilt, Luca Weihs, Alvaro Herrasti, Matt Deitke, Kiana Ehsani, Daniel Gordon, Yuke Zhu, et al. Ai2-thor: An interactive 3d environment for visual ai. arXiv preprint arXiv:1712.05474, 2017." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 469, + 294, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 469, + 294, + 512 + ], + "spans": [ + { + "bbox": [ + 62, + 469, + 294, + 512 + ], + "type": "text", + "content": "[7] Angel Chang, Angela Dai, Thomas Funkhouser, Maciej Halber, Matthias Niessner, Manolis Savva, Shuran Song, Andy Zeng, and Yinda Zhang. Matterport3d: Learning from RGB-D data in indoor environments. 3DV, 2017. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 514, + 294, + 556 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 514, + 294, + 556 + ], + "spans": [ + { + "bbox": [ + 62, + 514, + 294, + 556 + ], + "type": "text", + "content": "[8] Erik Wijmans, Abhishek Kadian, Ari S. Morcos, Stefan Lee, Irfan Essa, Devi Parikh, Manolis Savva, and Dhruv Batra. Dd-ppo: Learning near-perfect pointgoal navigators from 2.5 billion frames. In ICLR, 2019. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 558, + 294, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 558, + 294, + 610 + ], + "spans": [ + { + "bbox": [ + 62, + 558, + 294, + 610 + ], + "type": "text", + "content": "[9] Piotr Mirowski, Matt Grimes, Mateusz Malinowski, Karl Moritz Hermann, Keith Anderson, Denis Teplyashin, Karen Simonyan, Andrew Zisserman, Raia Hadsell, et al. Learning to navigate in cities without a map. NeurIPS, 31, 2018. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 613, + 294, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 613, + 294, + 646 + ], + "spans": [ + { + "bbox": [ + 57, + 613, + 294, + 646 + ], + "type": "text", + "content": "[10] Ang Li, Huiyi Hu, Piotr Mirowski, and Mehrdad Farajtabar. Cross-view policy learning for street navigation. In ICCV, pages 8100-8109, 2019." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 647, + 294, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 294, + 690 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 294, + 690 + ], + "type": "text", + "content": "[11] Karl Moritz Hermann, Mateusz Malinowski, Piotr Mirowski, Andras Banki-Horvath, Keith Anderson, and Raia Hadsell. Learning to follow directions in street view. In AAAI, volume 34, pages 11773-11781, 2020." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 692, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 692, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 692, + 294, + 713 + ], + "type": "text", + "content": "[12] Piotr Mirowski, Andras Banki-Horvath, Keith Anderson, Denis Teptyashin, Karl Moritz Hermann, Mateusz Mali" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 116 + ], + "type": "text", + "content": "nowski, Matthew Koichi Grimes, Karen Simonyan, Koray Kavukcuoglu, Andrew Zisserman, et al. The streetlearn environment and dataset. arXiv preprint arXiv:1903.01292, 2019. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 118, + 553, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 553, + 148 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 553, + 148 + ], + "type": "text", + "content": "[13] Dhruv Shah and Sergey Levine. ViKiNG: Vision-Based Kilometer-Scale Navigation with Geographic Hints. In RSS, 2022. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 150, + 553, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 150, + 553, + 194 + ], + "spans": [ + { + "bbox": [ + 316, + 150, + 553, + 194 + ], + "type": "text", + "content": "[14] Dhruv Shah, Ajay Sridhar, Arjun Bhorkar, Noriaki Hirose, and Sergey Levine. Gnm: A general navigation model to drive any robot. In ICRA, pages 7226-7233. IEEE, 2023. 2, 5, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 195, + 553, + 238 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 195, + 553, + 238 + ], + "spans": [ + { + "bbox": [ + 316, + 195, + 553, + 238 + ], + "type": "text", + "content": "[15] Wayne Wu, Honglin He, Yiran Wang, Chenda Duan, Jack He, Zhizheng Liu, Quanyi Li, and Bolei Zhou. Metaurban: A simulation platform for embodied ai in urban spaces. arXiv preprint arXiv:2407.08725, 2024. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 239, + 553, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 239, + 553, + 293 + ], + "spans": [ + { + "bbox": [ + 316, + 239, + 553, + 293 + ], + "type": "text", + "content": "[16] Adarsh Jagan Sathyamoorthy, Utsav Patel, Moumita Paul, Nithish K Sanjeev Kumar, Yash Savle, and Dinesh Manocha. Comet: Modeling group cohesion for socially compliant robot navigation in crowded scenes. IEEE Robotics and Automation Letters, 7(2):1008-1015, 2021. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 294, + 553, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 294, + 553, + 327 + ], + "spans": [ + { + "bbox": [ + 316, + 294, + 553, + 327 + ], + "type": "text", + "content": "[17] Muchen Sun, Francesca Baldini, Peter Trautman, and Todd Murphey. Move beyond trajectories: Distribution space coupling for crowd navigation. In RSS, 2021." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 328, + 553, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 328, + 553, + 371 + ], + "spans": [ + { + "bbox": [ + 316, + 328, + 553, + 371 + ], + "type": "text", + "content": "[18] Yuya Onozuka, Ryosuke Matsumi, and Motoki Shino. Autonomous mobile robot navigation independent of road boundary using driving recommendation map. In IROS, pages 4501-4508. IEEE, 2021. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 372, + 553, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 372, + 553, + 392 + ], + "spans": [ + { + "bbox": [ + 316, + 372, + 553, + 392 + ], + "type": "text", + "content": "[19] Tom B Brown. Language models are few-shot learners. arXiv preprint arXiv:2005.14165, 2020. 2, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 394, + 553, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 394, + 553, + 437 + ], + "spans": [ + { + "bbox": [ + 316, + 394, + 553, + 437 + ], + "type": "text", + "content": "[20] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In ICCV, pages 4015-4026, 2023." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 438, + 553, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 438, + 553, + 547 + ], + "spans": [ + { + "bbox": [ + 317, + 438, + 553, + 547 + ], + "type": "text", + "content": "[21] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel HAZIZA, Francisco Massa, Alaaeldin El-Nouby, Mido Assran, Nicolas Ballas, Wojciech Galuba, Russell Howes, Po-Yao Huang, Shang-Wen Li, Ishan Misra, Michael Rabbat, Vasu Sharma, Gabriel Synnaeve, Hu Xu, Herve Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. DINOv2: Learning robust visual features without supervision. Transactions on Machine Learning Research, 2024. 3, 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 548, + 553, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 548, + 553, + 601 + ], + "spans": [ + { + "bbox": [ + 317, + 548, + 553, + 601 + ], + "type": "text", + "content": "[22] Danny Driess, Fei Xia, Mehdi SM Sajjadi, Corey Lynch, Aakanksha Chowdhery, Brian Ichter, Ayzaan Wahid, Jonathan Tompson, Quan Vuong, Tianhe Yu, et al. Palme: An embodied multimodal language model. arXiv preprint arXiv:2303.03378, 2023. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 603, + 553, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 603, + 553, + 646 + ], + "spans": [ + { + "bbox": [ + 316, + 603, + 553, + 646 + ], + "type": "text", + "content": "[23] Beichen Wang, Juexiao Zhang, Shuwen Dong, Irving Fang, and Chen Feng. Vlm see, robot do: Human demo video to robot action plan via vision language model. arXiv preprint arXiv:2410.08792, 2024. 2, 3, 5" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 647, + 553, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 690 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 690 + ], + "type": "text", + "content": "[24] Noriaki Hirose, Catherine Glossop, Ajay Sridhar, Dhruv Shah, Oier Mees, and Sergey Levine. Lelan: Learning a language-conditioned navigation policy from in-the-wild videos. arXiv preprint arXiv:2410.03603, 2024. 2, 3, 5" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "text", + "content": "[25] Zachary Teed, Lahav Lipson, and Jia Deng. Deep patch visual odometry. NeurIPS, 36, 2024. 2, 3, 4" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6883" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "text", + "content": "[26] Max Jaderberg, Volodymyr Mnih, Wojciech Marian Czarnecki, Tom Schaul, Joel Z Leibo, David Silver, and Koray Kavukcuoglu. Reinforcement learning with unsupervised auxiliary tasks. In ICLR, 2017. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 160 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 160 + ], + "type": "text", + "content": "[27] Saurabh Gupta, James Davidson, Sergey Levine, Rahul Sukthankar, and Jitendra Malik. Cognitive mapping and planning for visual navigation. In CVPR, pages 2616-2625, 2017. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 162, + 294, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 162, + 294, + 194 + ], + "spans": [ + { + "bbox": [ + 56, + 162, + 294, + 194 + ], + "type": "text", + "content": "[28] Nikolay Savinov, Alexey Dosovitskiy, and Vladlen Koltun. Semi-parametric topological memory for navigation. In ICLR, 2018. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 195, + 294, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 195, + 294, + 237 + ], + "spans": [ + { + "bbox": [ + 56, + 195, + 294, + 237 + ], + "type": "text", + "content": "[29] Meera Hahn, Devendra Singh Chaplot, Shubham Tulsiani, Mustafa Mukadam, James M Rehg, and Abhinav Gupta. No rl, no simulation: Learning to navigate without navigating. NeurIPS, 34:26661-26673, 2021." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 239, + 294, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 239, + 294, + 282 + ], + "spans": [ + { + "bbox": [ + 56, + 239, + 294, + 282 + ], + "type": "text", + "content": "[30] Obin Kwon, Nuri Kim, Yunho Choi, Hwiyeon Yoo, Jeongho Park, and Songhwai Oh. Visual graph memory with unsupervised representation for visual navigation. In CVPR, pages 15890-15899, 2021." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 283, + 294, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 283, + 294, + 336 + ], + "spans": [ + { + "bbox": [ + 56, + 283, + 294, + 336 + ], + "type": "text", + "content": "[31] Jacob Krantz, Theophile Gervet, Karmesh Yadav, Austin Wang, Chris Paxton, Roozbeh Mottaghi, Dhruv Batra, Jitendra Malik, Stefan Lee, and Devendra Singh Chaplot. Navigating to objects specified by images. In CVPR, pages 10916-10925, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 338, + 294, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 338, + 294, + 370 + ], + "spans": [ + { + "bbox": [ + 56, + 338, + 294, + 370 + ], + "type": "text", + "content": "[32] Theophile Gervet, Soumith Chintala, Dhruv Batra, Jitendra Malik, and Devendra Singh Chaplot. Navigating to objects in the real world. Science Robotics, 8(79):eadf6991, 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 372, + 294, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 372, + 294, + 414 + ], + "spans": [ + { + "bbox": [ + 56, + 372, + 294, + 414 + ], + "type": "text", + "content": "[33] Devendra Singh Chaplot, Dhiraj Prakashchand Gandhi, Abhinav Gupta, and Russ R Salakhutdinov. Object goal navigation using goal-oriented semantic exploration. NeurIPS, 33:4247-4258, 2020." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 415, + 294, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 415, + 294, + 458 + ], + "spans": [ + { + "bbox": [ + 56, + 415, + 294, + 458 + ], + "type": "text", + "content": "[34] Arjun Majumdar, Gunjan Aggarwal, Bhavika Devnani, Judy Hoffman, and Dhruv Batra. Zson: Zero-shot object-goal navigation using multimodal goal embeddings. NeurIPS, 35:32340-32352, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 460, + 294, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 460, + 294, + 492 + ], + "spans": [ + { + "bbox": [ + 56, + 460, + 294, + 492 + ], + "type": "text", + "content": "[35] Abhishek Das, Samyak Datta, Georgia Gkioxari, Stefan Lee, Devi Parikh, and Dhruv Batra. Embodied question answering. In CVPR, pages 1-10, 2018. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 493, + 294, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 493, + 294, + 525 + ], + "spans": [ + { + "bbox": [ + 56, + 493, + 294, + 525 + ], + "type": "text", + "content": "[36] Hao Tan, Licheng Yu, and Mohit Bansal. Learning to navigate unseen environments: Back translation with environmental dropout. arXiv preprint arXiv:1904.04195, 2019." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 526, + 294, + 570 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 526, + 294, + 570 + ], + "spans": [ + { + "bbox": [ + 56, + 526, + 294, + 570 + ], + "type": "text", + "content": "[37] Jacob Krantz, Erik Wijmans, Arjun Majumdar, Dhruv Batra, and Stefan Lee. Beyond the nav-graph: Vision-and-language navigation in continuous environments. In ECCV, pages 104-120. Springer, 2020. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 571, + 294, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 571, + 294, + 613 + ], + "spans": [ + { + "bbox": [ + 56, + 571, + 294, + 613 + ], + "type": "text", + "content": "[38] Maks Sorokin, Jie Tan, C Karen Liu, and Sehoon Ha. Learning to navigate sidewalks in outdoor environments. IEEE Robotics and Automation Letters, 7(2):3906-3913, 2022. 2, 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 614, + 294, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 614, + 294, + 667 + ], + "spans": [ + { + "bbox": [ + 56, + 614, + 294, + 667 + ], + "type": "text", + "content": "[39] Joanne Truong, Max Rudolph, Naoki Harrison Yokoyama, Sonia Chernova, Dhruv Batra, and Akshara Rai. Rethinking sim2real: Lower fidelity simulation leads to higher sim2real transfer in navigation. In CoRL, pages 859-870. PMLR, 2023." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 669, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 669, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 669, + 294, + 713 + ], + "type": "text", + "content": "[40] Joanne Truong, Denis Yarats, Tianyu Li, Franziska Meier, Sonia Chernova, Dhruv Batra, and Akshara Rai. Learning navigation skills for legged robots with learned robot embeddings. In IROS, pages 484-491. IEEE, 2021." + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "text", + "content": "[41] Wei Liu, Huihua Zhao, Chenran Li, Joydeep Biswas, Billy Okal, Pulkit Goyal, Yan Chang, and Soha Pouya. X-mobility: End-to-end generalizable navigation via world modeling. arXiv preprint arXiv:2410.17491, 2024. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 118, + 553, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 553, + 160 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 553, + 160 + ], + "type": "text", + "content": "[42] Gregory Kahn, Pieter Abbeel, and Sergey Levine. Badgr: An autonomous self-supervised learning-based navigation system. IEEE Robotics and Automation Letters, 6(2):1312-1319, 2021. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 163, + 553, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 553, + 195 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 553, + 195 + ], + "type": "text", + "content": "[43] Gregory Kahn, Pieter Abbeel, and Sergey Levine. Land: Learning to navigate from disengagements. IEEE Robotics and Automation Letters, 6(2):1872-1879, 2021. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 197, + 553, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 197, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 316, + 197, + 553, + 239 + ], + "type": "text", + "content": "[44] Dhruv Shah, Ajay Sridhar, Nitish Dashora, Kyle Stachowicz, Kevin Black, Noriaki Hirose, and Sergey Levine. ViNT: A foundation model for visual navigation. In CoRL, 2023. 2, 5, 6, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 241, + 553, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 241, + 553, + 283 + ], + "spans": [ + { + "bbox": [ + 316, + 241, + 553, + 283 + ], + "type": "text", + "content": "[45] Ajay Sridhar, Dhruv Shah, Catherine Glossop, and Sergey Levine. Nomad: Goal masked diffusion policies for navigation and exploration. In ICRA, pages 63-70, 2024. 2, 5, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 286, + 553, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 286, + 553, + 341 + ], + "spans": [ + { + "bbox": [ + 316, + 286, + 553, + 341 + ], + "type": "text", + "content": "[46] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763. PMLR, 2021. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 342, + 553, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 342, + 553, + 374 + ], + "spans": [ + { + "bbox": [ + 316, + 342, + 553, + 374 + ], + "type": "text", + "content": "[47] Annie S Chen, Suraj Nair, and Chelsea Finn. Learning generalizable robotic reward functions from \"in-the-wild\" human videos. arXiv preprint arXiv:2103.16817, 2021. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 376, + 553, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 376, + 553, + 398 + ], + "spans": [ + { + "bbox": [ + 316, + 376, + 553, + 398 + ], + "type": "text", + "content": "[48] Shikhar Bahl, Abhinav Gupta, and Deepak Pathak. Human-to-robot imitation in the wild. In RSS, 2022. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 399, + 553, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 399, + 553, + 431 + ], + "spans": [ + { + "bbox": [ + 316, + 399, + 553, + 431 + ], + "type": "text", + "content": "[49] Austin Patel, Andrew Wang, Ilija Radosavovic, and Jitendra Malik. Learning to imitate object interactions from internet videos. arXiv preprint arXiv:2211.13225, 2022. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 433, + 553, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 433, + 553, + 465 + ], + "spans": [ + { + "bbox": [ + 316, + 433, + 553, + 465 + ], + "type": "text", + "content": "[50] Jimuyang Zhang, Ruizhao Zhu, and Eshed Ohn-Bar. Selfd: Self-learning large-scale driving policies from the web. In CVPR, pages 17316-17326, 2022. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 467, + 553, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 467, + 553, + 510 + ], + "spans": [ + { + "bbox": [ + 316, + 467, + 553, + 510 + ], + "type": "text", + "content": "[51] Naoki Wake, Atsushi Kanehira, Kazuhiro Sasabuchi, Jun Takamatsu, and Katsushi Ikeuchi. Gpt-4v (ision) for robotics: Multimodal task planning from human demonstration. IEEE Robotics and Automation Letters, 2024. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 511, + 553, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 511, + 553, + 533 + ], + "spans": [ + { + "bbox": [ + 316, + 511, + 553, + 533 + ], + "type": "text", + "content": "[52] Jakob Engel, Vladlen Koltun, and Daniel Cremers. Direct sparse odometry. TPAMI, 40(3):611-625, 2017. 4" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 535, + 553, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 535, + 553, + 578 + ], + "spans": [ + { + "bbox": [ + 316, + 535, + 553, + 578 + ], + "type": "text", + "content": "[53] Raul Mur-Artal, Jose Maria Martinez Montiel, and Juan D Tardos. Orb-slam: a versatile and accurate monocular slam system. IEEE transactions on robotics, 31(5):1147-1163, 2015." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 579, + 553, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 579, + 553, + 611 + ], + "spans": [ + { + "bbox": [ + 316, + 579, + 553, + 611 + ], + "type": "text", + "content": "[54] Zachary Teed and Jia Deng. Droid-slam: Deep visual slam for monocular, stereo, and rgb-d cameras. NeurIPS, 34:16558-16569, 2021. 4" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 613, + 553, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 613, + 553, + 657 + ], + "spans": [ + { + "bbox": [ + 316, + 613, + 553, + 657 + ], + "type": "text", + "content": "[55] Yuhang He, Irving Fang, Yiming Li, Rushi Bhavesh Shah, and Chen Feng. Metric-Free Exploration for Topological Mapping by Task and Motion Imitation in Feature Space. In RSS, July 2023. 5" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 658, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 553, + 712 + ], + "type": "text", + "content": "[56] Mahmoud Assran, Quentin Duval, Ishan Misra, Piotr Bojanowski, Pascal Vincent, Michael Rabbat, Yann LeCun, and Nicolas Ballas. Self-supervised learning from images with a joint-embedding predictive architecture. In CVPR, pages 15619–15629, 2023." + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6884" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 297 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 127 + ], + "type": "text", + "content": "[57] Adrien Bardes, Quentin Garrido, Jean Ponce, Xinlei Chen, Michael Rabbat, Yann LeCun, Mahmoud Assran, and Nicolas Ballas. Revisiting feature prediction for learning visual representations from video. arXiv preprint arXiv:2404.08471, 2024. 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 294, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 294, + 184 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 294, + 184 + ], + "type": "text", + "content": "[58] Adarsh Jagan Sathyamoorthy, Kasun Weerakoon, Mohamed Elnoor, Anuj Zore, Brian Ichter, Fei Xia, Jie Tan, Wenhao Yu, and Dinesh Manocha. Convoi: Context-aware navigation using vision language models in outdoor and indoor environments. arXiv preprint arXiv:2403.15637, 2024. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 185, + 294, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 185, + 294, + 228 + ], + "spans": [ + { + "bbox": [ + 56, + 185, + 294, + 228 + ], + "type": "text", + "content": "[59] Wei Xu and Fu Zhang. Fast-lio: A fast, robust lidar-inertial odometry package by tightly-coupled iterated kalman filter. IEEE Robotics and Automation Letters, 6(2):3317-3324, 2021. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 230, + 294, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 230, + 294, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 230, + 294, + 262 + ], + "type": "text", + "content": "[60] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 264, + 294, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 264, + 294, + 297 + ], + "spans": [ + { + "bbox": [ + 56, + 264, + 294, + 297 + ], + "type": "text", + "content": "[61] Noriaki Hirose, Dhruv Shah, Ajay Sridhar, and Sergey Levine. SACson: Scalable autonomous control for social navigation. IEEE Robotics and Automation Letters, 2023. 6" + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "6885" + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Classic Video Denoising in a Machine Learning World_ Robust, Fast, and Controllable/079f9f38-09f8-4442-836f-ae05472182ae_content_list.json b/2025/Classic Video Denoising in a Machine Learning World_ Robust, Fast, and Controllable/079f9f38-09f8-4442-836f-ae05472182ae_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..364a96c4f2cdbfb5f91e9c1d83a2aa11edce85ae --- /dev/null +++ b/2025/Classic Video Denoising in a Machine Learning World_ Robust, Fast, and Controllable/079f9f38-09f8-4442-836f-ae05472182ae_content_list.json @@ -0,0 +1,1674 @@ +[ + { + "type": "text", + "text": "Classic Video Denoising in a Machine Learning World: Robust, Fast, and Controllable", + "text_level": 1, + "bbox": [ + 220, + 130, + 776, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xin Jin $^{1*}$ Simon Niklaus $^{2\\dagger}$ Zhoutong Zhang $^{3}$ Zhihao Xia $^{3}$", + "bbox": [ + 251, + 202, + 743, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Chunle Guo $^{1,4‡}$ Yuting Yang $^{3}$ Jiawen Chen $^{3}$ Chongyi Li $^{1,4}$", + "bbox": [ + 251, + 220, + 743, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ VCIP, CS, Nankai University $^{2}$ Adobe Research $^{3}$ Adobe $^{4}$ NKIARI, Shenzhen Futian", + "bbox": [ + 194, + 239, + 803, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "xjin@mail.nankai.edu.cn, sniklaus@adobe.com,", + "bbox": [ + 303, + 258, + 689, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{guochunle,lichongyi}@nankai.edu.cn", + "bbox": [ + 339, + 277, + 655, + 292 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "https://srameo.github.io/projects/levd", + "bbox": [ + 326, + 295, + 666, + 309 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4efc2cd0f1ea64e14b5777b5312b57420d30070c0a38ba489869a982e3741489.jpg", + "image_caption": [ + "Input" + ], + "image_footnote": [], + "bbox": [ + 93, + 321, + 225, + 465 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/451109cddf2cdc2c1ce4f5be28fb772a8728cb2f5b25c3a3954fd18a5d576ebb.jpg", + "image_caption": [ + "$\\mathrm{VRT}^{\\dagger}$ [26]", + "(sophisticated model)" + ], + "image_footnote": [], + "bbox": [ + 230, + 321, + 364, + 465 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ea7e17ec687dedc64ec0ba245e20bf2b17c88cb9c50091cf261f9dfdb1789655.jpg", + "image_caption": [ + "Real-ESRGAN [46]", + "(sophisticated dataset)", + "Figure 1. In real-world videos, noise can manifest in vastly different ways, causing supervised methods to fail when the input is too far outside the training distribution. Footage provided by Robert Kjettrup, face pixelated after inference. We denote retrained models with a $\\dagger$ ." + ], + "image_footnote": [], + "bbox": [ + 367, + 321, + 500, + 465 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0498becf27d39aa6541777cd5cd727475a3c9511cb9ab028960376cefe66e847.jpg", + "image_caption": [ + "MF2F [14]", + "(sophisticated supervision)" + ], + "image_footnote": [], + "bbox": [ + 503, + 321, + 637, + 465 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/33790d91080c2092151131a3f11e8d9e74a5f794b46f9a11c48eb3a67c5faed7.jpg", + "image_caption": [ + "UDVD [40]", + "(sophisticated supervision)" + ], + "image_footnote": [], + "bbox": [ + 638, + 321, + 772, + 465 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/fbc1097a89e12e7276e5d2b7f0909f438ed2fb90f4ffcb476e97cfd293bc984c.jpg", + "image_caption": [ + "Ours - RFCVD" + ], + "image_footnote": [], + "bbox": [ + 776, + 321, + 911, + 465 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 546, + 326, + 561 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Denoising is a crucial step in many video processing pipelines such as in interactive editing, where high quality, speed, and user control are essential. While recent approaches achieve significant improvements in denoising quality by leveraging deep learning, they are prone to unexpected failures due to discrepancies between training data distributions and the wide variety of noise patterns found in real-world videos. These methods also tend to be slow and lack user control. In contrast, traditional denoising methods perform reliably on in-the-wild videos and run relatively quickly on modern hardware. However, they require manually tuning parameters for each input video, which is not only tedious but also requires skill. We bridge the gap between these two paradigms by proposing a differentiable denoising pipeline based on traditional methods. A neural network is then trained to predict the optimal denoising parameters for each specific input, resulting in a robust and efficient approach that also supports user control.", + "bbox": [ + 88, + 578, + 485, + 849 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 514, + 546, + 643, + 561 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Video denoising is a fundamental part of any video editing pipeline, and it is typically applied prior to color grading. After interviewing professional video editors, we have found that they not only want clean results but they also want the denoising to be quick as not to interrupt their workflow, and they want controllability to assert their artistic expression. Specifically, editors often have to decide to leave some noise in the footage in favor of over-smoothing or vice versa, and the answer can differ for each case. For this reason, professional cameras like the Arri Alexa even allow the user to choose between different noise profiles to better support different post-production workflows.", + "bbox": [ + 511, + 577, + 906, + 758 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Unfortunately, recent work on video denoising tends to focus only on the first aspect, the denoising quality, which makes these solutions impractical for the typical video editing workflow. At the same time, we have found that such video denoising approaches are subject to failure cases like the one in Fig. 1 even though they focus on quality. This is not surprising though, considering that noise in real videos can manifest itself in wildly different ways. That is, in addition to the typical degradation that images are subject to,", + "bbox": [ + 511, + 763, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*This project was done during Xin Jin's internship at Adobe Research.", + "bbox": [ + 107, + 875, + 480, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "†Project lead. ‡ Corresponding author.", + "bbox": [ + 109, + 887, + 323, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "2084", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c613ff942e19a76b73e678ff027bca564ac3f5d4528da137eae75312d5b858d9.jpg", + "image_caption": [ + "Figure 2. High-level comparison of how related work approaches video denoising (left) and our proposed approach (right)." + ], + "image_footnote": [], + "bbox": [ + 91, + 85, + 477, + 198 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "videos also leverage temporal compression. In the H.264 codec, for example, there are P-frames and B-frames that copy information from I-frames which means that the noise is often temporally correlated. Furthermore, there are many different video codecs that affect the noise profile in different ways, and even for the same codec, the noise can vary significantly when using different encoders or settings.", + "bbox": [ + 88, + 271, + 482, + 378 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To tackle the complexity of video noise in a deep learning context and as summarized in Fig. 2 (a), we have found that popular approaches either propose a sophisticated way to supervise the model [14, 40], utilize an intricate pipeline to simulate noise [46], leverage an image formation that by design is less prone to overfitting [13], or a combination of these. Yet, the aforementioned example in Fig. 1 demonstrates that this is not always sufficient.", + "bbox": [ + 88, + 385, + 482, + 507 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In contrast, classic denoisers not only work reasonably well but they are also relatively fast, especially on modern hardware. However, they require manually tuning parameters which is not only tedious but also requires a certain amount of skill. We see this as an opportunity. Specifically and as shown in Fig. 2 (b), by implementing a traditional denoising pipeline in a differentiable manner, we can have a neural network learn to predict what the optimal denoising parameters for a given input should be. As shown in Fig. 3, not only does this end up being robust but it is also comparatively fast while providing support for user control. To be clear on our objectives, if the noise profile is known, such as when developing a denoiser for a specific camera model subject to a constrained video encoder, any reasonable deep learning approach would provide better denoising results. However, such an approach would in turn be limited in terms of robustness and controllability.", + "bbox": [ + 89, + 513, + 482, + 772 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In short, our contributions are (1) an approach to leverage a traditional video denoising pipeline in a deep learning setting where we decouple the analysis of the noise from the denoising itself to improve efficiency by avoiding redundant compute, (2) a not only robust and fast but also controllable video denosier (RFCVD), and (3) a training augmentation pipeline based on additive white Gaussian noise together with H.264 transcoding that works surprisingly well.", + "bbox": [ + 88, + 780, + 482, + 902 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/792d63a9c11df280a50ebf42916cc13ae56dabea86898464513706235a06f898.jpg", + "image_caption": [ + "Figure 3. Video denoising on the CRVD (sRGB) benchmark [50] using the PSNR across all ISO values with respect to the computational efficiency on an RTX 3090 GPU in FPS (frames per second). We improved some models through retraining, denoted with a $\\dagger$ ." + ], + "image_footnote": [], + "bbox": [ + 517, + 87, + 903, + 284 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 358, + 653, + 375 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Traditional Video Denoising. Most traditional video denoising methods can be viewed as a series of linear and non-linear filtering processes, where the filter strength is a function of the estimated noise level. Many such methods are based on block matching [10-12, 31], where similar patches are first aggregated within and across frames before being filtered and merged to produce the denoised result. There are many variations of this process of course, for example by improving the matching [28] or by leveraging bilateral filtering [8, 18, 36, 43] for spatial and temporal denoising [1, 37]. In doing so, image pyramids [3, 16] can be used to improve the computational efficiency and recent iterations of classic multi-frame denoising methods are even quick to run on commodity smartphones [20, 27, 48].", + "bbox": [ + 511, + 388, + 906, + 602 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Machine Learning in Video Denoising. In recent years, it has become increasingly popular to tackle video denoising with the help of machine learning. This includes methods that, like traditional video denoisers, perform explicit matching [13, 44, 49] as well as those where correspondences are only implicit [24, 32, 42]. And while most of these methods denoise each frame independently, there are also IIR-like approaches where information is passed along such that previously denoised frames can guide future ones [5, 6]. On the model architecture side, the rise of attention in the vision and language community has also led to various attention mechanisms for video restoration [4, 26, 29, 41, 45]. But with machine learning there is a catch, if an input has a noise pattern that has not been seen during training then the denoising is unpredictable.", + "bbox": [ + 511, + 607, + 908, + 835 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Noise Simulation. A common approach to facilitate better generalizability on in-the-wild footage is to employ a more sophisticated degradation pipeline when adding synthetic noise to the ground truth. This includes not only do", + "bbox": [ + 511, + 839, + 908, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2085", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/eea26d388a9b6825f76e7393350f7cac0dd3d0bba7410ef509e6081aa9480a39.jpg", + "image_caption": [ + "Noise Profiling on an Anchor Frame Through a Hypernetwork", + "Figure 4. Video denoising fundamentally first needs to analyze the noise and then remove it. We mimic this in our pipeline by first estimating a noise profile on a random anchor frame (top left in green) before using this profile to denoise the video (bottom in yellow). Specifically, we leverage a hypernetwork configuration where the noise profile $\\theta$ is essentially the parameters of the subsequent denoiser. That is, our denoiser is a traditional pipeline consisting of (1) a Wiener filter that performs temporal denoising of neighboring frames that were aligned via optical flow and (2) a bilateral Laplacian pyramid filter for spatial denoising of the temporally merged frames, where a small neural network $\\mathcal{P}(\\cdot ;\\theta)$ predicts spatially-varying parameters for the Wiener merger and the bilateral filters. This separation of concerns improves the overall efficiency since it avoids having to redundantly analyze the noise over and over again." + ], + "image_footnote": [], + "bbox": [ + 99, + 107, + 903, + 401 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ing one but two rounds of degradations [46], shuffling the degradations [51], or integrating more advanced operators such as tone mapping [52]. Although these approaches improve the ability to denoise real-world videos, they can only reduce but not entirely resolve the generalizability gap.", + "bbox": [ + 89, + 513, + 482, + 590 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Self-Supervised Learning. Another strategy to improve the generalizability in video denoising is to adopt self-supervision on real-world videos without requiring access to a ground truth. For example, this can be achieved by leveraging the nature of regression losses [25] or by building on the blind spot idea [14, 17, 23, 40]. While these approaches make it possible to train on noisy in-the-wild videos, it is not necessarily feasible to collect a dataset that contains all possible types of noise that occur in the real world. Therefore, self-supervision often also relies on test-time adaptation which comes with its own challenges.", + "bbox": [ + 89, + 595, + 482, + 762 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 782, + 181, + 797 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Video denoising fundamentally first needs to analyze the noise and then remove it. In contrast, recent approaches perform both tasks at once by utilizing a single neural network that is given the noisy input frames and is tasked with producing the clean output. The noise profile in a video does typically not change over time though, yet these ap", + "bbox": [ + 89, + 809, + 482, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "proaches independently denoise one frame after another so they have to analyze the noise over and over again.", + "bbox": [ + 511, + 515, + 903, + 544 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To avoid this redundancy, we follow the natural separation of analyzing the noise and then removing it. As shown in Fig. 4, we first estimate the noise profile on an anchor frame and then denoise the video using the estimated profile. Specifically, we leverage a hypernetwork setup where the noise profile $\\theta$ provides the parameters of the subsequent denoiser. That is, our denoiser is a traditional pipeline consisting of (1) a Wiener filter that performs temporal denoising of neighboring frames that were aligned via optical flow and (2) a bilateral Laplacian pyramid filter for spatial denoising of the temporally merged frames, where a small neural network $\\mathcal{P}(\\cdot ;\\theta)$ predicts spatially-varying parameters for the Wiener merger and the bilateral filters.", + "bbox": [ + 511, + 546, + 905, + 741 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We will subsequently discuss these parts in turn before providing more details on user control and the training.", + "bbox": [ + 511, + 743, + 903, + 773 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Noise Profiling", + "text_level": 1, + "bbox": [ + 511, + 785, + 663, + 801 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Since noise is not necessarily spatially uniform but often dependent on the signal, we argue that a noise profile should not be just a fixed set of values for the entire input but rather be a descriptor that makes it possible to derive spatially-varying denoising parameters. We see this as an ideal use case for hypernetworks [19], where a noise profiler esti", + "bbox": [ + 511, + 809, + 905, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "2086", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/03eb94de23ae0d8a136bd6e105428eeb4057be9a3566073988ccd7b21a83c1f1.jpg", + "image_caption": [ + "Input" + ], + "image_footnote": [], + "bbox": [ + 91, + 87, + 225, + 219 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/ad9e2eb9e981b29dfd671fe69868eeb2f3ad0d935aec408f3ae3e1069df1f2f1.jpg", + "image_caption": [ + "reduced temporal only $(0.2\\cdot \\sigma^2,0\\cdot \\sigma_d,0\\cdot \\sigma_r)$" + ], + "image_footnote": [], + "bbox": [ + 228, + 88, + 364, + 219 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/ee0a4435d21cba2ce07442037b033b55e7a52fff9663f7627660cf0d649de330.jpg", + "image_caption": [ + "temporal only $(1\\cdot \\sigma^2,0\\cdot \\sigma_d,0\\cdot \\sigma_r)$", + "Figure 5. By scaling the denoising parameters estimated by $\\mathcal{P}(\\cdot ;\\theta)$ , our framework makes it easy to control the amount of temporal denoising (through $\\sigma^2$ ) as well as the amount of spatial denoising (through $\\sigma_d$ and $\\sigma_r$ ). We scale the denoising parameters equivalently for the luma and chroma channels in this sample, but they can be scaled independently for more control. Please see the supplementary for a video demo that demonstrates how users can control the denoising in an interactive manner. Footage provided by Amit Zinman." + ], + "image_footnote": [], + "bbox": [ + 367, + 88, + 500, + 219 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/712c0b43eaf5055d26349b9a6f8881bb85100a2ef1a47be6b8ed107beb14b91a.jpg", + "image_caption": [ + "reduced spatial only $(0\\cdot \\sigma^{2},0.2\\cdot \\sigma_{d},0.2\\cdot \\sigma_{r})$" + ], + "image_footnote": [], + "bbox": [ + 501, + 88, + 637, + 219 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/e5d5f37b5cf91cd079f40ccd6ae4b57a05565d5d5897ccf4de24ae648121e492.jpg", + "image_caption": [ + "spatial only $(0\\cdot \\sigma^{2},1\\cdot \\sigma_{d},1\\cdot \\sigma_{r})$" + ], + "image_footnote": [], + "bbox": [ + 638, + 88, + 774, + 219 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/182ed8d1157ea58a335c7c3636b019f63f2f9e999ff20ffe761ce505120f6f8a.jpg", + "image_caption": [ + "default settings $(1\\cdot \\sigma^{2},1\\cdot \\sigma_{d},1\\cdot \\sigma_{r})$" + ], + "image_footnote": [], + "bbox": [ + 776, + 88, + 911, + 219 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "mates the weights $\\theta$ of a small neural network $\\mathcal{P}(\\cdot ;\\theta)$ that predicts spatially-varying denoising parameters. However, instead of directly predicting $\\theta$ , we stabilize the training through NPA [35] where $\\theta = \\theta^0 +\\Delta \\theta$ with $\\theta^0$ being learnable parameters and $\\Delta \\theta$ being the hypernet prediction.", + "bbox": [ + 89, + 309, + 483, + 383 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Intuitively, analyzing the noise profile of a given image is a mixture of low-level image processing and high-level semantics. Specifically, one first needs to understand what an image depicts before being able to examine the details to discern between unintended noise and actual texture. For this reason, we leverage a pre-trained backbone $\\mathcal{B}$ in the form of a ConvNext [30] with a random MLP head to serve as the hypernet. We have found this backbone to be crucial for the quality of the predicted noise profile.", + "bbox": [ + 89, + 385, + 483, + 521 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Lastly, which anchor frame should we choose to analyze the noise of a given video? We would like this choice not to matter, such that the denoising result is independent of having to choose a \"good\" anchor. To achieve this, we always pick the first frame of a video as the anchor frame because it is as good as any, and then utilize a consistency loss", + "bbox": [ + 89, + 521, + 483, + 612 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {c s t s y} = \\left\\| \\Delta \\theta - \\Delta \\theta_ {i} \\right\\| _ {2}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 200, + 625, + 482, + 643 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where we encourage the hypernet prediction of our anchor frame $\\Delta \\theta$ and any random neighbor $\\Delta \\theta_{i}$ to be equal.", + "bbox": [ + 89, + 654, + 483, + 685 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Denoising the Video", + "text_level": 1, + "bbox": [ + 89, + 695, + 281, + 710 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Once we have the noise profile $\\theta$ , we process the input video one output frame at a tim where we obtain spatially-varying denoising parameters through a small neural network $\\mathcal{P}(\\cdot ;\\theta)$ with three convolutional layers. We use these parameters to do temporal and then spatial denoising.", + "bbox": [ + 89, + 719, + 482, + 795 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For the temporal denoising, to denoise a given target frame, we first align the two preceding and the two subsequent frames. Specifically, we estimate the optical flow between the target frame and the neighbors using an off-the-shelf SpyNet [38] at a quarter of the target resolution for improved computational efficiency as well as robustness to noise, and warp the neighbors to the target frame using the", + "bbox": [ + 89, + 795, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "estimated optical flow. We then merge the aligned frames using a Wiener filter as in Hasinoff et al. [20] but using a tiles size of $8 \\times 8$ and the noise maps from $\\mathcal{P}(\\cdot; \\theta)$ instead of approximating the noise as the RMS in each tile. As is typical and to facilitate more control, we perform Wiener filtering on the luma and the chroma channels independently.", + "bbox": [ + 511, + 309, + 905, + 400 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We then apply spatial denoising to the temporally denoised target frame through a bilateral Laplacian pyramid filter with three levels. This borrows many ideas from multiresolution bilateral filtering [53] but using a Laplacian instead of a Gaussian pyramid and using bilateral filtering throughout instead of wavelet thresholding for the first pyramid levels. More specifically, we have $\\mathcal{P}(\\cdot ;\\theta)$ estimate noise map pyramids and then use these as $\\sigma_{s}$ and $\\sigma_r$ for the bilateral filtering. Furthermore, just like with the Wiener filter, we perform the denoising separately on the luma and the chroma channels to facilitate user control.", + "bbox": [ + 511, + 400, + 905, + 565 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. User Control", + "text_level": 1, + "bbox": [ + 513, + 575, + 650, + 589 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Through interviews with creative professionals, we have found that they want to assert their artistic expression when denoising a clip. That is, they often have to decide to leave some noise in the footage in favor of over-smoothing and vice versa. The key to facilitating user control in our method is the spatially-varying parameters from $\\mathcal{P}(\\cdot ;\\theta)$ that guide the temporal and spatial denoising. Specifically, we have two of these maps for $\\sigma^2$ in the Wiener filter (one map for chroma and one for luma) and in the bilateral Laplacian pyramid filter we have two map pyramids for $\\sigma_{d}$ as well as two map pyramids for $\\sigma_r$ (one map pyramid for chroma and one for luma). This makes for a total of six sets of parameters where we can facilitate user control.", + "bbox": [ + 511, + 599, + 905, + 794 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "How do we attenuate these parameters though? As demonstrated in Fig. 5, we have found that simply scaling the spatially-varying denoising parameters uniformly works reasonably well, so we can just provide six tunable knobs to end users. And since our denoising framework runs in real time, changing the knobs provides instantaneous feedback which makes it easy to attain the desired result.", + "bbox": [ + 511, + 795, + 905, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "2087", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Augmentation Pipeline", + "text_level": 1, + "bbox": [ + 91, + 90, + 302, + 107 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To train our model, we need pairs of clean and noisy frame sequences just like any other denoiser that leverages machine learning without self-supervision. Specifically, we adopt the typical paradigm of taking a set of \"noise-free\" videos that serve as the ground truth, and augmenting them with various degradations to obtain the corresponding noisy inputs. For this to work well, we have found it to be crucial to include temporal compression in this degradation pipeline since it makes the noise temporally correlated.", + "bbox": [ + 88, + 113, + 480, + 250 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "But let's start with noise-free videos, which we obtain from the REDS dataset [33]. It provides 240 videos and was shot at a relatively high 120 frames per second, allowing us to augment the frame rate via sub-sampling. We extract random frame sequences from these videos and first degrade them through additive white Gaussian noise with a variance sampled from $\\mathcal{U}(1,50)$ . We then transcode them using an H.264 codec through libx264 with a CRF sampled from $\\mathcal{U}(18,30)$ to make the noise temporally correlated.", + "bbox": [ + 88, + 250, + 480, + 385 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We have found this data pipeline to work surprisingly well. So well in fact that we retrained the models that we compare to in the evaluation since we were able to get much better results on the CRVD benchmark [50] with our retrained versions than the original checkpoints. However, we did not retrain Real-ESRGAN [46] since it brings its own much more complex augmentation pipeline, and we also did not retrain MF2F [14] and UDVD [40] since these leverage sophisticated ways of supervising the models which are difficult for us to replicate and properly do justice.", + "bbox": [ + 88, + 385, + 480, + 537 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Implementation Details", + "text_level": 1, + "bbox": [ + 89, + 544, + 307, + 559 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In terms of architecture details, we utilize the base size of a ConvNext [30] for the backbone $\\mathcal{B}$ in the hypernetwork. To translate the output from this backbone to the residual noise profile $\\Delta \\theta$ , we utilize an MLP with five layers and PReLU activations [21] in between each layer. And for the neural network $\\mathcal{P}(\\cdot ;\\theta)$ that predicts the spatially-varying denoising parameters, we leverage three convolution layers with a stride of two and PReLU activations in between. We additionally constrain the output of this network with a softmax [15] to facilitate non-negative denoising parameters. Furthermore, we have found it beneficial to not only provide the (aligned) frames to $\\mathcal{P}(\\cdot ;\\theta)$ but also their gradients, approximated by a Sobel filter, as well as masks that indicate whether or not a pixel in an aligned frame is valid (it may have been warped from outside the frame) [2].", + "bbox": [ + 88, + 566, + 482, + 792 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In our reconstruction loss, we minimize the difference between the temporally denoised image $I^{\\mathrm{T}}$ and the ground truth $I^{\\mathrm{GT}}$ as well as the difference between all levels of the spatial denoising pyramid $I^{\\mathrm{ST}}$ and $I^{\\mathrm{GT}}$ as", + "bbox": [ + 88, + 794, + 482, + 854 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {r e c} = \\left\\| I ^ {\\mathrm {T}} - I ^ {\\mathrm {G T}} \\right\\| _ {2} + \\sum_ {l = 1} ^ {3} \\left\\| I _ {l} ^ {\\mathrm {S T}} - I _ {l} ^ {\\mathrm {G T}} \\right\\| _ {2}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 140, + 862, + 482, + 902 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/829fbf10a8350c0805f77332216be38cc2949cf5cdc8672b68c6c42497a2cd50.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
PSNRdelta (higher PSNR is better)SSIMdelta (higher SSIM is better)LPIPSdelta (lower LPIPS is better)
NAFNet [9]30.25-0.747-0.358-
NAFNet†[9]35.32+ 5.070.937+ 0.1900.089- 0.269
FastDVDNet [42]27.89-0.565-0.502-
FastDVDNet†[42]34.63+ 6.750.914+ 0.3490.124- 0.378
TOFlow [49]25.96-0.673-0.251-
TOFlow†[49]34.08+ 8.110.903+ 0.2310.147- 0.104
BasicVSR++ [6]31.98-0.769-0.326-
BasicVSR++†[6]34.34+ 2.350.873+ 0.1040.167- 0.158
VRT [26]31.99-0.784-0.296-
VRT†[26]33.86+ 1.860.848+ 0.0640.192- 0.104
", + "bbox": [ + 514, + 89, + 905, + 272 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1. Denoising results on the CRVD (sRGB) benchmark for various methods that we compare to, both for their original version as well as our retrained one which we denote with a $\\dagger$ .", + "bbox": [ + 511, + 277, + 903, + 319 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $l$ is the $l$ -th level in spatial denoising pyramid. Recall that we also have a consistency loss that we outlined in Eq. (1) in the section about noise profiling.", + "bbox": [ + 511, + 333, + 903, + 378 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We train our model using Adam [22] with an initial learning rate of $2 \\times 10^{-4}$ that decays to $1 \\times 10^{-7}$ using a cosine annealing schedule. In total, we train the model for 400 thousand iterations with a batch size of 24 consisting of patches with $512 \\times 512$ pixels. Since our pipeline is computationally lightweight, this takes less than two days.", + "bbox": [ + 511, + 378, + 905, + 470 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 482, + 643, + 500 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We quantitatively evaluate our approach on two types of videos, real and synthetic. For the former, we leverage the CRVD benchmark [50] consisting of RAW videos which we convert to sRGB using the provided deep ISP [39]. For the latter, we use the validation samples from the REDS dataset [33] which we augmented in two ways that are notably different from the AWGN subject to H.264 transcoding during training. Specifically, one way is using film grain noise [34] with AV1 transcoding, and the other way is using spatially correlated Gaussian noise with H.265 transcoding. As for evaluation metrics, we follow the typical paradigm of reporting PSNR, SSIM [47], and LPIPS [54]. Since we argue that our method is computationally efficient, we also report the frames per second (FPS) which we measured on an RTX 3090 GPU with proper CUDA synchronization.", + "bbox": [ + 511, + 508, + 905, + 734 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In terms of data for our qualitative evaluation, we reached out to industry professionals who sent us footage where they had trouble removing the noise. Since these professionals did not have a release form for all the actors, we sometimes had to anonymize faces post inference.", + "bbox": [ + 511, + 734, + 903, + 809 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Lastly, we compare our approach to various kinds of denoisers. This includes image denoisers [7, 9], video denoisers [42, 49], recurrent methods [6], transformers [26], models with a sophisticated augmentation pipeline [46], as well as methods that leverage self-supervision [14, 40]. Since we have found that our augmentation pipeline works sur", + "bbox": [ + 511, + 810, + 903, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "2088", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/320b47a166857f110f0538d369558864636bcdddf64de8327b2351dc396a3732.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ISO 1600ISO 3200ISO 6400ISO 12800ISO 25600OverallSpeed
PSNRrank (higher PSNR is better)PSNRrank (higher PSNR is better)PSNRrank (higher PSNR is better)PSNRrank (higher PSNR is better)PSNRrank (higher PSNR is better)PSNRrank (higher PSNR is better)FPSrank (higher FPS is better)
\\( SID^† \\)[7]38.85\\( 7^{th} \\) of 1037.68\\( 7^{th} \\) of 1035.82\\( 4^{th} \\) of 1033.51\\( 4^{th} \\) of 1029.18\\( 3^{rd} \\) of 1035.01\\( 4^{th} \\) of 106.95\\( 3^{rd} \\) of 10
\\( NAFNet^† \\)[9]39.48\\( 3^{rd} \\) of 1038.12\\( 5^{th} \\) of 1035.94\\( 3^{rd} \\) of 1033.53\\( 3^{rd} \\) of 1029.55\\( 2^{nd} \\) of 1035.32\\( 2^{nd} \\) of 101.69\\( 7^{th} \\) of 10
\\( FastDVDNet^† \\)[42]39.16\\( 5^{th} \\) of 1037.92\\( 6^{th} \\) of 1035.60\\( 5^{th} \\) of 1032.56\\( 5^{th} \\) of 1027.93\\( 6^{th} \\) of 1034.63\\( 5^{th} \\) of 105.72\\( 4^{th} \\) of 10
\\( TOFlow^† \\)[49]38.25\\( 8^{th} \\) of 1036.97\\( 8^{th} \\) of 1034.90\\( 7^{th} \\) of 1032.21\\( 6^{th} \\) of 1028.07\\( 5^{th} \\) of 1034.08\\( 7^{th} \\) of 102.84\\( 6^{th} \\) of 10
\\( BasicVSR++^† \\)[6]39.40\\( 4^{th} \\) of 1038.24\\( 2^{nd} \\) of 1035.55\\( 6^{th} \\) of 1031.72\\( 7^{th} \\) of 1026.78\\( 9^{th} \\) of 1034.34\\( 6^{th} \\) of 107.41\\( 2^{nd} \\) of 10
\\( VRT^† \\)[26]39.55\\( 2^{nd} \\) of 1038.12\\( 4^{th} \\) of 1034.82\\( 8^{th} \\) of 1030.77\\( 8^{th} \\) of 1026.01\\( 10^{th} \\) of 1033.86\\( 8^{th} \\) of 100.05\\( 10^{th} \\) of 10
Real-ESRGAN [46]29.98\\( 10^{th} \\) of 1028.27\\( 10^{th} \\) of 1028.04\\( 10^{th} \\) of 1027.52\\( 10^{th} \\) of 1027.63\\( 8^{th} \\) of 1028.29\\( 10^{th} \\) of 100.24\\( 8^{th} \\) of 10
UDVD [40]31.15\\( 9^{th} \\) of 1030.72\\( 9^{th} \\) of 1030.23\\( 9^{th} \\) of 1029.10\\( 9^{th} \\) of 1027.63\\( 7^{th} \\) of 1029.77\\( 9^{th} \\) of 100.16\\( 9^{th} \\) of 10
MF2F [14]39.09\\( 6^{th} \\) of 1038.20\\( 3^{rd} \\) of 10\\( \\underline{36.36} \\)\\( 1^{st} \\) of 1033.57\\( 2^{nd} \\) of 1029.04\\( 4^{th} \\) of 1035.25\\( 3^{rd} \\) of 104.62\\( 5^{th} \\) of 10
Ours - RFCVD40.35\\( 1^{st} \\) of 10\\( \\underline{38.60} \\)\\( 1^{st} \\) of 1036.28\\( 2^{nd} \\) of 1033.86\\( 1^{st} \\) of 1031.12\\( 1^{st} \\) of 1036.04\\( 1^{st} \\) of 1031.66\\( 1^{st} \\) of 10
", + "bbox": [ + 91, + 88, + 906, + 295 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/5f0b875e59c15221ca2dddd7f2ff9faad5de8a4c651f36750f682954679e99b6.jpg", + "table_caption": [ + "Table 2. Video denoising results on the CRVD (sRGB) benchmark. Not only does our approach perform best overall, it is also four times faster than the second-fastest place. Please kindly see the supplementary for SSIM and LPIPS where our approach ranks first as well." + ], + "table_footnote": [], + "table_body": "
film grain noise w/ AV1 compressionspatially correlated noise w/ H.265 compressionSpeed
PSNR (higher PSNR is better)rank SSIM (higher SSIM is better)rank (lower LPIPS is better)LPIPS rank (lower LPIPS is better)rank (higher PSNR is better)rank (higher PSNR is better)SSIM (higher SSIM is better)rank (lower LPIPS is better)LPIPS rank (lower LPIPS is better)rank (higher FPS is better)FPS rank (higher FPS is better)
\\( SID^† \\)[7]27.05 5thof 70.664 5thof 70.293 4thof 728.44 3rdof 70.771 4thof 70.282 5thof 715.00 3rdof 7
\\( NAFNet^† \\)[9]27.16 4thof 70.672 4thof 70.294 5thof 728.40 4thof 70.764 6thof 70.257 3rdof 73.812 6thof 7
\\( FastDVDNet^† \\)[42]27.39 3rdof 70.686 3rdof 70.272 3rdof 727.92 6thof 70.769 5thof 70.296 6thof 712.09 4thof 7
\\( TOFlow^† \\)[49]28.15 2ndof 70.750 2ndof 70.220 1stof 728.39 5thof 70.788 3rdof 70.258 4thof 75.665 5thof 7
\\( BasicVSR++^† \\)[6]26.90 6thof 70.651 6thof 70.313 6thof 727.48 7thof 70.728 7thof 70.358 7thof 715.32 2ndof 7
\\( VRT^† \\)[26]26.55 7thof 70.629 7thof 70.331 7thof 728.78 2ndof 70.803 2ndof 70.206 1stof 70.105 7thof 7
Ours - RFCVD28.59 1stof 70.774 1stof 70.247 2ndof 728.93 1stof 70.808 1stof 70.239 2ndof 769.73 1stof 7
", + "bbox": [ + 91, + 327, + 906, + 492 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 3. Testing the generalizability when training on AWGN with H.264 transcoding but testing with two different degradation schemes. Since our approach is based on classic methods, the denoising itself only has a few parameters which naturally reduces domain gaps.", + "bbox": [ + 89, + 494, + 906, + 523 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "prisingly well, we have retrained all methods that neither leverage a sophisticated augmentation pipeline already nor utilize self-supervision. This makes for a more fair comparison since, as shown in Tab. 1, we have improved their performance on the CRVD benchmark with our retrained version. Throughout this paper, we denote all retrained models with a $\\dagger$ to prevent any potential confusions.", + "bbox": [ + 88, + 549, + 482, + 657 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Quantitative Comparison", + "text_level": 1, + "bbox": [ + 89, + 665, + 323, + 683 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Please see Tab. 2 for a summary of the quantitative evaluation on the CRVD benchmark. In short, not only does our approach perform best overall, it is also four times faster than the second-fastest method. We attribute these favorable results to our utilization of classic denoising techniques which greatly helps with the ability to generalize to unseen noise patterns. Specifically, the denoising itself is driven by only a few parameters which naturally reduces overfitting and helps with bridging domain gaps. To test this hypothesis, we conducted an experiment with synthetic data where all models are trained on AWGN with H.264 transcoding but tested on two different degradation pipelines. As shown in Tab. 3, our approach performs favorably in this evaluation as well which supports our initial hypothesis.", + "bbox": [ + 88, + 688, + 482, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Qualitative Comparison", + "text_level": 1, + "bbox": [ + 511, + 549, + 735, + 566 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To test video denoising in the real world, we reached out to various creative professionals who kindly shared footage where they had trouble removing the noise. We show a representative excerpt of the denoising results on this footage in Fig. 6 but kindly refer to the supplementary which includes video results of all methods. Like with the quantitative evaluation, we have found that our approach performs favorably, and once again we hypothesize that this is due to the underlying utilization of classic denoising techniques.", + "bbox": [ + 511, + 573, + 906, + 710 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Ablative Experiments", + "text_level": 1, + "bbox": [ + 511, + 723, + 718, + 739 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "One of the main things we wondered ourselves was the performance with respect to the choice of anchor frame. After all, the noise profile of I-frames can be quite different from P- and B-frames. Nevertheless as shown in Tab. 4 (top), we have always used the first frame as the anchor throughout the experiments section and would have drawn the same conclusions with the middle or the last frame. And as shown in Tab. 4 (bottom), we have also found that $\\mathcal{L}_{cstsy}$ helps with providing a slightly more consistent experience.", + "bbox": [ + 511, + 748, + 906, + 883 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "A key part of our integration of classic denoising tech", + "bbox": [ + 531, + 885, + 903, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "2089", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1df6be4b5a87c7f1cc7e60530297622ad88009c6ee6832c78a28aef66251d7fe.jpg", + "image_caption": [ + "Figure 6. Video denoising results from exemplary still frames. Due to space constraints, we only share the results from a representative subset of the methods that we compare to. Please kindly refer to the supplementary to find video results of all methods. We are grateful that Amit Zinman (top row), Robert Kjettrup (middle row), and an anonymous artist (last row) were willing to provide test footage." + ], + "image_footnote": [], + "bbox": [ + 91, + 87, + 915, + 579 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "niques into a machine learning pipeline is that we account for spatially-varying noise. To assess the importance of this design decision, we trained an ablation that only predicts a set of scalar denoising parameters. As expected and as shown in Tab. 6, this \"w/o spat. varying\" experiment exhibits a significant drop in quality. As shown in Fig. 7, this ablation produces much more blurry denoising results since it cannot easily account for signal-dependent noise.", + "bbox": [ + 89, + 654, + 482, + 775 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To best support $\\mathcal{P}(\\cdot ;\\theta)$ in predicting the denoising parameters, we not only provide it access to the (merged) input frames but also their gradients as well as a mask indicating whether or not a pixel in an aligned frame is valid. To test the effectiveness of this design, we trained two ablations where we refrained from providing one (\"w/o img. gradients\") or the other (\"w/o align. mask\") and found that, as shown in Tab. 6, this context is indeed beneficial.", + "bbox": [ + 89, + 779, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We argue that analyzing the noise profile of a given image is a mixture of both low-level image processing and high-level semantics, because one first needs an understanding of what an image depicts before being able to go into the low-level details to discern between unintended noise and actual texture. For this reason, we leverage a classification backbone $\\mathcal{B}$ in the hypernetwork that is predicting the noise profile $\\theta$ , and we trained an ablation without this backbone to evaluate this hypothesis. As intuitively expected and as shown in Tab. 6, this \"w/o backbone $\\mathcal{B}$ \" experiment indeed exhibits a significant drop in quality. This drop is also visually exemplified in Fig. 7, where the lack of this backbone leads to the noise not fully being removed.", + "bbox": [ + 511, + 654, + 906, + 851 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Lastly, we not only fine-tune the aforementioned backbone $\\mathcal{B}$ but also stabilize the training of the hypernetwork through NPA [35]. To analyze the effect of these measures,", + "bbox": [ + 511, + 854, + 906, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "2090", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/b4e0511ae9b8b9641516ea398d424619a5f80ad61d73e75cf0927089e7e8b9ca.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
PSNRdelta (higher PSNR is better)SSIMdelta (higher SSIM is better)LPIPSdelta (lower LPIPS is better)
Avg. w/ Lcstsy35.95-0.9477-0.0757-
first frame ← Ours36.04+ 0.090.9472- 0.00050.0763+ 0.0006
middle frame35.92- 0.030.9480+ 0.00030.0753- 0.0003
last frame35.90- 0.060.9480+ 0.00030.0754- 0.0003
Avg. w/o Lcstsy35.86-0.9460-0.0754-
first frame35.80- 0.050.9454- 0.00050.0750- 0.0004
middle frame36.02+ 0.160.9473+ 0.00140.0753- 0.0001
last frame35.75- 0.110.9451- 0.00080.0759+ 0.0005
", + "bbox": [ + 91, + 88, + 483, + 250 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/b20f714afce042ef99277613972615ff4b4c6ceba10de0d20f79561ffe5cec5d.jpg", + "table_caption": [ + "Table 4. Denoising results on the CRVD benchmark with respect to the anchor frame choice. We find that our approach is relatively robust to this choice (top), partly thanks to $\\mathcal{L}_{cstsy}$ (bottom)." + ], + "table_footnote": [], + "table_body": "
PSNRdelta (higher PSNR is better)SSIMdelta (higher SSIM is better)LPIPSdelta (lower LPIPS is better)
Ours36.04-0.9472-0.0763-
fine-tuned on GT36.75+ 0.700.9494+ 0.00220.0591- 0.0172
", + "bbox": [ + 91, + 316, + 483, + 387 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "we trained an ablation for the former (\"w/o fine-tuning $\\mathcal{B}$ \" as well as the latter (\"w/o NPA\") and found that, as shown in Tab. 6, both contribute to the denoising quality.", + "bbox": [ + 89, + 457, + 482, + 503 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Limitations", + "text_level": 1, + "bbox": [ + 89, + 521, + 217, + 535 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "While our noise profile $\\theta$ with the subsequent denoising parameter prediction from $\\mathcal{P}(\\cdot ;\\theta)$ works reasonably well as demonstrated throughout the experiments section, it is far from perfect. Specifically and as shown in Tab. 5, we fine-tuned the predicted denoising parameters on the ground truth and found that this hypothetical oracle is able to improve our results by a rather significant margin. Nevertheless, we leave bridging this gap to future research.", + "bbox": [ + 88, + 547, + 482, + 667 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In the case where the noise profile is entirely known a priori, such as when developing a denoiser for a specific camera model subject to a constrained encoder, any reasonable deep learning method would provide better results than our approach. And with a baked-in noise profile it would perhaps even be faster. However, such a method would in turn be limited in terms of robustness and control.", + "bbox": [ + 88, + 671, + 482, + 776 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Lastly, and on a more practical note, deploying hypernetworks in end-user applications is not well supported by the common inference ecosystems yet. Specifically, popular libraries like CoreML, WinML, or OpenVINO expect the model weights to be fixed which is obviously not the case for hypernetworks. As such, while our approach is easy to implement within frameworks like PyTorch, it is not straightforward to deploy in production environments.", + "bbox": [ + 88, + 779, + 482, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d617da301db3f1bdba758af444dbdfcab7b4df741dc9fd6d3e24380b2057bc9f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 516, + 88, + 612, + 164 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a6147f8108f4b894c00419764e36d1f746cdb23aecdc4fc353bba3b5aaaf48ff.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 516, + 165, + 611, + 210 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/180577504c3d21230b9c91032cf119b2c6a956c0fec859267a84e0865cb2eac1.jpg", + "image_caption": [ + "Input" + ], + "image_footnote": [], + "bbox": [ + 516, + 212, + 611, + 286 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/eaa422a565e2b480918ec025a8b466b88925dc1d38c01191b6cec38c30e3cf6e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 612, + 88, + 709, + 164 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5c34871bb5c19595625021fa1f101a9261a4c10068a1a26b6b041902c7c096c2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 612, + 165, + 709, + 210 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/22a767d79f8d800880a9161affb4f0ab1bdd976f843f4c3315c2daed2f8b1ef4.jpg", + "image_caption": [ + "w/o spat. varying" + ], + "image_footnote": [], + "bbox": [ + 612, + 212, + 709, + 286 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/96bc151ab373809c214ffbbff0e65f4ac71871d1fe2f61295b4e34343477b2a1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 710, + 88, + 805, + 164 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/59f2eaae540e41a729af029582b29ef35a405c799151bd9f7ea48c7ef3299880.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 710, + 165, + 805, + 210 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/95a7282e092419b8effe5e0f9ff2b0152145fd6b4fddeccbbc5ce563b8d9761b.jpg", + "image_caption": [ + "w/o backbone $\\mathcal{B}$" + ], + "image_footnote": [], + "bbox": [ + 710, + 212, + 805, + 286 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b101c90257dcf2a870cfee103439369810981e8bdc94895ed837dc4763e388cb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 808, + 88, + 903, + 164 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/fec91e0d23dbf73ac3a7bc97da582be1d81334a9498716733cceb39f243ac12f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 808, + 165, + 903, + 210 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/17345923f8d9819f678594c7a1db336f4bfd0602988913cf9edab0f8723f18c3.jpg", + "image_caption": [ + "Ours", + "Figure 7. Visual comparison of the two most significant ablations on a representative sample from the CRVD dataset." + ], + "image_footnote": [], + "bbox": [ + 808, + 212, + 903, + 286 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/7e1bf30bde2f724afe6bef1e87ea8ff40953b1f2b7e7d72cfa3f146dc0b9a33f.jpg", + "table_caption": [ + "Table 5. Fine-tuning our estimated denoising parameters on the ground truth of the CRVD benchmark indicates that our estimates are good but not perfect if we had access to an oracle." + ], + "table_footnote": [], + "table_body": "
PSNRdelta (higher PSNR is better)SSIMdelta (higher SSIM is better)LPIPSdelta (lower LPIPS is better)
Ours36.04-0.9472-0.0763-
w/o spat. varying33.62-2.420.9359-0.01120.1184+0.0421
w/o img. gradients35.61-0.440.9469-0.00030.0756-0.0007
w/o align. mask35.26-0.780.9468-0.00030.0740-0.0023
w/o backbone B34.38-1.670.9059-0.04120.1444+0.0680
w/o fine-tuning B35.71-0.340.9464-0.00080.0732-0.0032
w/o NPA [35]35.83-0.210.9453-0.00180.0814+0.0051
", + "bbox": [ + 514, + 347, + 905, + 489 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 6. Various ablative experiments, please see Section 4.3 for more context and details of each individual ablation.", + "bbox": [ + 511, + 492, + 905, + 520 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 541, + 633, + 556 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We show that traditional denoising techniques are still very much relevant in the modern machine learning world. Not only have we found them to be robust and fast, but also controllable which is an important property in many video denoising applications where creative professionals want to assert their artistic expression. To bridge the gap between classic denoising and modern approaches, we train a model to estimate the various parameters of a traditional denoising approach for a given input video, which is otherwise not only tedious but also requires a certain amount of skill.", + "bbox": [ + 509, + 566, + 906, + 717 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Furthermore, in an effort to make best use of the computational efficiency of classic denoising techniques, our method separates the analysis of the noise from the denoising in order to avoid having to estimate the noise profile over and over again as is typical for machine learning approaches to video denoising. Lastly, we have found a simple yet effective degradation pipeline for training video denoiseers based on AWGN subject to H.264 video transcoding. This pipeline ensures that the noise is temporally correlated just like it often is the case in the real world, and we have found it not only to work well for our approach but we have also been able to retrain and improve existing models.", + "bbox": [ + 509, + 719, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "2091", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement. This work was supported in part by the National Natural Science Foundation of China (62306153, 62225604), the Fundamental Research Funds for the Central Universities (Nankai University, 070-63243143), the Natural Science Foundation of Tianjin, China (24JCJQJC00020), and Shenzhen Science and Technology Program (JCYJ20240813114237048). The computational devices is partly supported by the Supercomputing Center of Nankai University (NKSC).", + "bbox": [ + 89, + 90, + 485, + 227 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 242, + 187, + 258 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Eric P Bennett and Leonard McMillan. Video enhancement using per-pixel virtual exposures. In ACM SIGGRAPH 2005 Papers, 2005. 2", + "[2] Goutam Bhat, Michael Gharbi, Jiawen Chen, Luc Van Gool, and Zhihao Xia. Self-supervised burst super-resolution. In ICCV, pages 10571-10580. IEEE, 2023. 5", + "[3] Peter J Burt and Edward H Adelson. The laplacian pyramid as a compact image code. In Readings in computer vision, 1987. 2", + "[4] Jiezhang Cao, Yawei Li, Kai Zhang, and Luc Van Gool. Video super-resolution transformer. arXiv:2106.06847, 2021. 2", + "[5] Kelvin CK Chan, Xintao Wang, Ke Yu, Chao Dong, and Chen Change Loy. Basicvsr: The search for essential components in video super-resolution and beyond. In CVPR, 2021. 2", + "[6] Kelvin CK Chan, Shangchen Zhou, Xiangyu Xu, and Chen Change Loy. Basicvsr++: Improving video superresolution with enhanced propagation and alignment. In CVPR, 2022. 2, 5, 6", + "[7] Chen Chen, Qifeng Chen, Jia Xu, and Vladlen Koltun. Learning to see in the dark. In CVPR, 2018. 2, 5, 6", + "[8] Jiawen Chen, Sylvain Paris, and Frédo Durand. Real-time edge-aware image processing with the bilateral grid. ACM TOG, 2007. 2", + "[9] Liangyu Chen, Xiaojie Chu, Xiangyu Zhang, and Jian Sun. Simple baselines for image restoration. In ECCV, 2022. 2, 5, 6", + "[10] Kostadin Dabov, Alessandro Foi, Vladimir Katkovnik, and Karen Egiazarian. Image denoising with block-matching and 3d filtering. In Image Processing, 2006. 2", + "[11] Kostadin Dabov, Alessandro Foi, and Karen Egiazarian. Video denoising by sparse 3d transform domain collaborative filtering. EURASIP, 2007.", + "[12] Kostadin Dabov, Alessandro Foi, Vladimir Katkovnik, and Karen Egiazarian. Image denoising by sparse 3-d transform-domain collaborative filtering. IEEE TIP, 2007. 2", + "[13] Axel Davy, Thibaud Ehret, Jean-Michel Morel, Pablo Arias, and Gabriele Facciolo. A non-local cnn for video denoising. In ICIP, 2019. 2", + "[14] Valery Dewil, Jérémy Anger, Axel Davy, Thibaud Ehret, Gabriele Facciolo, and Pablo Arias. Self-supervised training for blind multi-frame video denoising. In WACV, 2021. 1, 2, 3, 5, 6, 7" + ], + "bbox": [ + 91, + 266, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Charles Dugas, Yoshua Bengio, François Bélisle, Claude Nadeau, and René Garcia. Incorporating second-order functional knowledge for better option pricing. In NIPS, 2000. 5", + "[16] Jana Ehmann, Lun-Cheng Chu, Sung-Fang Tsai, and Chia-Kai Liang. Real-time video denoising on mobile phones. In ICIP, 2018. 2", + "[17] Thibaud Ehret, Axel Davy, Jean-Michel Morel, Gabriele Facciolo, and Pablo Arias. Model-blind video denoising via frame-to-frame training. In CVPR, 2019. 3", + "[18] Ruturaj G Gavaskar and Kunal N Chaudhury. Fast adaptive bilateral filtering. IEEE TIP, 2018. 2", + "[19] David Ha, Andrew M. Dai, and Quoc V. Le. Hypernetworks. arXiv:1609.09106, 2016. 3", + "[20] Samuel W Hasinoff, Dillon Sharlet, Ryan Geiss, Andrew Adams, Jonathan T Barron, Florian Kainz, Jiawen Chen, and Marc Levoy. Burst photography for high dynamic range and low-light imaging on mobile cameras. ACM TOG, 2016. 2, 4", + "[21] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Delving deep into rectifiers: Surpassing human-level performance on imagenet classification. In ICCV, 2015. 5", + "[22] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR, 2015. 5", + "[23] Samuli Laine, Tero Karras, Jaakko Lehtinen, and Timo Aila. High-quality self-supervised deep image denoising. In NeurIPS, 2019. 3", + "[24] Seunghwan Lee, Donghyeon Cho, Jiwon Kim, and Tae Hyun Kim. Restore from restored: Video restoration with pseudo clean video. In CVPR, 2021. 2", + "[25] Jaakko Lehtinen, Jacob Munkberg, Jon Hasselgren, Samuli Laine, Tero Karras, Miika Aittala, and Timo Aila. Noise2noise: Learning image restoration without clean data. In ICML, 2018. 3", + "[26] Jingyun Liang, Jiezhang Cao, Yuchen Fan, Kai Zhang, Rakesh Ranjan, Yawei Li, Radu Timofte, and Luc Van Gool. Vrt: A video restoration transformer. IEEE TIP, 2024. 1, 2, 5, 6, 7", + "[27] Orly Liba, Kiran Murthy, Yun-Ta Tsai, Tim Brooks, Tianfan Xue, Nikhil Karnad, Qiurui He, Jonathan T Barron, Dillon Sharlet, Ryan Geiss, et al. Handheld mobile photography in very low light. ACM TOG, 2019. 2", + "[28] Ce Liu and William T Freeman. A high-quality video denoising algorithm based on reliable motion estimation. In ECCV, 2010. 2", + "[29] Ding Liu, Zhaowen Wang, Yuchen Fan, Xianming Liu, Zhangyang Wang, Shiyu Chang, and Thomas Huang. Robust video super-resolution with learned temporal dynamics. In ICCV, 2017. 2", + "[30] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In CVPR, 2022. 4, 5", + "[31] Matteo Maggioni, Vladimir Katkovnik, Karen Egiazarian, and Alessandro Foi. Nonlocal transform-domain filter for volumetric data denoising and reconstruction. IEEE TIP, 2012. 2" + ], + "bbox": [ + 516, + 90, + 905, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "2092", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[32] Matteo Maggioni, Yibin Huang, Cheng Li, Shuai Xiao, Zhongqian Fu, and Fenglong Song. Efficient multi-stage video denoising with recurrent spatio-temporal fusion. In CVPR, 2021. 2", + "[33] Seungjun Nah, Sungyong Baik, Seokil Hong, Gyeongsik Moon, Sanghyun Son, Radu Timofte, and Kyoung Mu Lee. NTIRE 2019 challenge on video deblurring and superresolution: Dataset and study. In CVPRW, 2019. 5", + "[34] Alasdair Newson, Noura Faraj, Bruno Galerne, and Julie Delon. Realistic film grain rendering. IPOL, 2017. 5", + "[35] Jose Javier Gonzalez Ortiz, John V. Guttag, and Adrian V. Dalca. Non-proportional parametrizations for stable hypernetwork learning. arXiv:2304.07645, 2023. 3, 4, 7, 8", + "[36] Sylvain Paris, Pierre Kornprobst, Jack Tumblin, Frédo Durand, et al. Bilateral filtering: Theory and applications. Foundations and Trends® in Computer Graphics and Vision, 2009. 2", + "[37] Andrea Petreto, Thomas Romera, Florian Lemaitre, Ian Masliah, Boris Gaillard, Manuel Bouyer, Quentin L Meunier, and Lionel Lacassagne. A new real-time embedded video denoising algorithm. In DASIP, 2019. 2", + "[38] Anurag Ranjan and Michael J. Black. Optical flow estimation using a spatial pyramid network. In CVPR, 2017. 4", + "[39] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In MICCAI, 2015. 5", + "[40] Dev Yashpal Sheth, Sreyas Mohan, Joshua L Vincent, Ramon Manzorro, Peter A Crozier, Mitesh M Khapra, Eero P Simoncelli, and Carlos Fernandez-Granda. Unsupervised deep video denoising. In ICCV, 2021. 1, 2, 3, 5, 6, 7", + "[41] Maitreya Suin and AN Rajagopalan. Gated spatio-temporal attention-guided video deblurring. In CVPR, 2021. 2", + "[42] Matias Tassano, Julie Delon, and Thomas Veit. Fastdvnet: Towards real-time deep video denoising without flow estimation. In CVPR, 2020. 2, 5, 6", + "[43] Carlo Tomasi and Roberto Manduchi. Bilateral filtering for gray and color images. In ICCV, 1998. 2", + "[44] Gregory Vaksman, Michael Elad, and Peyman Milanfar. Patch craft: Video denoising by deep modeling and patch matching. In ICCV, 2021. 2", + "[45] Xintao Wang, Kelvin CK Chan, Ke Yu, Chao Dong, and Chen Change Loy. Edvr: Video restoration with enhanced deformable convolutional networks. In CVPRW, 2019. 2", + "[46] Xintao Wang, Liangbin Xie, Chao Dong, and Ying Shan. Real-esrgan: Training real-world blind super-resolution with pure synthetic data. In ICCV Workshop, 2021. 1, 2, 3, 5, 6, 7", + "[47] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE TIP, 2004. 5", + "[48] Bartlomiej Wronski, Ignacio Garcia-Dorado, Manfred Ernst, Damien Kelly, Michael Krainin, Chia-Kai Liang, Marc Levoy, and Peyman Milanfar. Handheld multi-frame superresolution. ACM TOG, 2019. 2", + "[49] Tianfan Xue, Baian Chen, Jiajun Wu, Donglai Wei, and William T Freeman. Video enhancement with task-oriented flow. IJCV, 2019. 2, 5, 6" + ], + "bbox": [ + 91, + 90, + 482, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[50] Huanjing Yue, Cong Cao, Lei Liao, Ronghe Chu, and Jingyu Yang. Supervised raw video denoising with a benchmark dataset on dynamic scenes. In CVPR, 2020. 2, 5", + "[51] Kai Zhang, Jingyun Liang, Luc Van Gool, and Radu Timofte. Designing a practical degradation model for deep blind image super-resolution. In CVPR, 2021. 3", + "[52] Kai Zhang, Yawei Li, Jingyun Liang, Jiezhang Cao, Yu-lun Zhang, Hao Tang, Deng-Ping Fan, Radu Timofte, and Luc Van Gool. Practical blind image denoising via swim-conv-unet and data synthesis. MIR, 2023. 3", + "[53] Ming Zhang and Bahadir K. Gunturk. Multiresolution bilateral filtering for image denoising. IEEE TIP, 2008. 4", + "[54] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 5" + ], + "bbox": [ + 516, + 92, + 903, + 304 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "2093", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Classic Video Denoising in a Machine Learning World_ Robust, Fast, and Controllable/079f9f38-09f8-4442-836f-ae05472182ae_model.json b/2025/Classic Video Denoising in a Machine Learning World_ Robust, Fast, and Controllable/079f9f38-09f8-4442-836f-ae05472182ae_model.json new file mode 100644 index 0000000000000000000000000000000000000000..65363ca2c7d71e7930283b87d3633bff00259d50 --- /dev/null +++ b/2025/Classic Video Denoising in a Machine Learning World_ Robust, Fast, and Controllable/079f9f38-09f8-4442-836f-ae05472182ae_model.json @@ -0,0 +1,2431 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.222, + 0.131, + 0.777, + 0.175 + ], + "angle": 0, + "content": "Classic Video Denoising in a Machine Learning World: Robust, Fast, and Controllable" + }, + { + "type": "text", + "bbox": [ + 0.253, + 0.203, + 0.744, + 0.221 + ], + "angle": 0, + "content": "Xin Jin\\(^{1*}\\) Simon Niklaus\\(^{2\\dagger}\\) Zhoutong Zhang\\(^{3}\\) Zhihao Xia\\(^{3}\\)" + }, + { + "type": "text", + "bbox": [ + 0.253, + 0.221, + 0.744, + 0.24 + ], + "angle": 0, + "content": "Chunle Guo\\(^{1,4‡}\\) Yuting Yang\\(^{3}\\) Jiawen Chen\\(^{3}\\) Chongyi Li\\(^{1,4}\\)" + }, + { + "type": "text", + "bbox": [ + 0.195, + 0.24, + 0.804, + 0.257 + ], + "angle": 0, + "content": "\\(^{1}\\)VCIP, CS, Nankai University \\(^{2}\\)Adobe Research \\(^{3}\\)Adobe \\(^{4}\\)NKIARI, Shenzhen Futian" + }, + { + "type": "text", + "bbox": [ + 0.304, + 0.26, + 0.691, + 0.275 + ], + "angle": 0, + "content": "xjin@mail.nankai.edu.cn, sniklaus@adobe.com," + }, + { + "type": "text", + "bbox": [ + 0.34, + 0.278, + 0.656, + 0.293 + ], + "angle": 0, + "content": "{guochunle,lichongyi}@nankai.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.327, + 0.296, + 0.668, + 0.31 + ], + "angle": 0, + "content": "https://srameo.github.io/projects/levd" + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.322, + 0.227, + 0.466 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.145, + 0.471, + 0.177, + 0.484 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image", + "bbox": [ + 0.231, + 0.322, + 0.365, + 0.466 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.267, + 0.47, + 0.328, + 0.483 + ], + "angle": 0, + "content": "\\(\\mathrm{VRT}^{\\dagger}\\) [26]" + }, + { + "type": "image_caption", + "bbox": [ + 0.246, + 0.484, + 0.348, + 0.494 + ], + "angle": 0, + "content": "(sophisticated model)" + }, + { + "type": "image", + "bbox": [ + 0.368, + 0.322, + 0.501, + 0.466 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.379, + 0.471, + 0.488, + 0.483 + ], + "angle": 0, + "content": "Real-ESRGAN [46]" + }, + { + "type": "image_caption", + "bbox": [ + 0.382, + 0.484, + 0.486, + 0.494 + ], + "angle": 0, + "content": "(sophisticated dataset)" + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.322, + 0.638, + 0.466 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.541, + 0.471, + 0.602, + 0.483 + ], + "angle": 0, + "content": "MF2F [14]" + }, + { + "type": "image_caption", + "bbox": [ + 0.508, + 0.484, + 0.634, + 0.495 + ], + "angle": 0, + "content": "(sophisticated supervision)" + }, + { + "type": "image", + "bbox": [ + 0.64, + 0.322, + 0.774, + 0.466 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.674, + 0.471, + 0.739, + 0.483 + ], + "angle": 0, + "content": "UDVD [40]" + }, + { + "type": "image_caption", + "bbox": [ + 0.645, + 0.484, + 0.771, + 0.495 + ], + "angle": 0, + "content": "(sophisticated supervision)" + }, + { + "type": "image", + "bbox": [ + 0.777, + 0.322, + 0.912, + 0.466 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.801, + 0.471, + 0.884, + 0.483 + ], + "angle": 0, + "content": "Ours - RFCVD" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.501, + 0.908, + 0.53 + ], + "angle": 0, + "content": "Figure 1. In real-world videos, noise can manifest in vastly different ways, causing supervised methods to fail when the input is too far outside the training distribution. Footage provided by Robert Kjettrup, face pixelated after inference. We denote retrained models with a \\(\\dagger\\)." + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.547, + 0.327, + 0.563 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.579, + 0.486, + 0.851 + ], + "angle": 0, + "content": "Denoising is a crucial step in many video processing pipelines such as in interactive editing, where high quality, speed, and user control are essential. While recent approaches achieve significant improvements in denoising quality by leveraging deep learning, they are prone to unexpected failures due to discrepancies between training data distributions and the wide variety of noise patterns found in real-world videos. These methods also tend to be slow and lack user control. In contrast, traditional denoising methods perform reliably on in-the-wild videos and run relatively quickly on modern hardware. However, they require manually tuning parameters for each input video, which is not only tedious but also requires skill. We bridge the gap between these two paradigms by proposing a differentiable denoising pipeline based on traditional methods. A neural network is then trained to predict the optimal denoising parameters for each specific input, resulting in a robust and efficient approach that also supports user control." + }, + { + "type": "title", + "bbox": [ + 0.515, + 0.547, + 0.645, + 0.563 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.578, + 0.907, + 0.76 + ], + "angle": 0, + "content": "Video denoising is a fundamental part of any video editing pipeline, and it is typically applied prior to color grading. After interviewing professional video editors, we have found that they not only want clean results but they also want the denoising to be quick as not to interrupt their workflow, and they want controllability to assert their artistic expression. Specifically, editors often have to decide to leave some noise in the footage in favor of over-smoothing or vice versa, and the answer can differ for each case. For this reason, professional cameras like the Arri Alexa even allow the user to choose between different noise profiles to better support different post-production workflows." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Unfortunately, recent work on video denoising tends to focus only on the first aspect, the denoising quality, which makes these solutions impractical for the typical video editing workflow. At the same time, we have found that such video denoising approaches are subject to failure cases like the one in Fig. 1 even though they focus on quality. This is not surprising though, considering that noise in real videos can manifest itself in wildly different ways. That is, in addition to the typical degradation that images are subject to," + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.875, + 0.482, + 0.888 + ], + "angle": 0, + "content": "*This project was done during Xin Jin's internship at Adobe Research." + }, + { + "type": "page_footnote", + "bbox": [ + 0.111, + 0.888, + 0.325, + 0.901 + ], + "angle": 0, + "content": "†Project lead. ‡ Corresponding author." + }, + { + "type": "list", + "bbox": [ + 0.109, + 0.875, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2084" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.093, + 0.087, + 0.478, + 0.199 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.207, + 0.483, + 0.237 + ], + "angle": 0, + "content": "Figure 2. High-level comparison of how related work approaches video denoising (left) and our proposed approach (right)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.272, + 0.483, + 0.379 + ], + "angle": 0, + "content": "videos also leverage temporal compression. In the H.264 codec, for example, there are P-frames and B-frames that copy information from I-frames which means that the noise is often temporally correlated. Furthermore, there are many different video codecs that affect the noise profile in different ways, and even for the same codec, the noise can vary significantly when using different encoders or settings." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.386, + 0.483, + 0.508 + ], + "angle": 0, + "content": "To tackle the complexity of video noise in a deep learning context and as summarized in Fig. 2 (a), we have found that popular approaches either propose a sophisticated way to supervise the model [14, 40], utilize an intricate pipeline to simulate noise [46], leverage an image formation that by design is less prone to overfitting [13], or a combination of these. Yet, the aforementioned example in Fig. 1 demonstrates that this is not always sufficient." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.515, + 0.483, + 0.773 + ], + "angle": 0, + "content": "In contrast, classic denoisers not only work reasonably well but they are also relatively fast, especially on modern hardware. However, they require manually tuning parameters which is not only tedious but also requires a certain amount of skill. We see this as an opportunity. Specifically and as shown in Fig. 2 (b), by implementing a traditional denoising pipeline in a differentiable manner, we can have a neural network learn to predict what the optimal denoising parameters for a given input should be. As shown in Fig. 3, not only does this end up being robust but it is also comparatively fast while providing support for user control. To be clear on our objectives, if the noise profile is known, such as when developing a denoiser for a specific camera model subject to a constrained video encoder, any reasonable deep learning approach would provide better denoising results. However, such an approach would in turn be limited in terms of robustness and controllability." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.781, + 0.483, + 0.903 + ], + "angle": 0, + "content": "In short, our contributions are (1) an approach to leverage a traditional video denoising pipeline in a deep learning setting where we decouple the analysis of the noise from the denoising itself to improve efficiency by avoiding redundant compute, (2) a not only robust and fast but also controllable video denosier (RFCVD), and (3) a training augmentation pipeline based on additive white Gaussian noise together with H.264 transcoding that works surprisingly well." + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.088, + 0.905, + 0.285 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.286, + 0.907, + 0.343 + ], + "angle": 0, + "content": "Figure 3. Video denoising on the CRVD (sRGB) benchmark [50] using the PSNR across all ISO values with respect to the computational efficiency on an RTX 3090 GPU in FPS (frames per second). We improved some models through retraining, denoted with a \\(\\dagger\\)." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.359, + 0.655, + 0.375 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.39, + 0.907, + 0.603 + ], + "angle": 0, + "content": "Traditional Video Denoising. Most traditional video denoising methods can be viewed as a series of linear and non-linear filtering processes, where the filter strength is a function of the estimated noise level. Many such methods are based on block matching [10-12, 31], where similar patches are first aggregated within and across frames before being filtered and merged to produce the denoised result. There are many variations of this process of course, for example by improving the matching [28] or by leveraging bilateral filtering [8, 18, 36, 43] for spatial and temporal denoising [1, 37]. In doing so, image pyramids [3, 16] can be used to improve the computational efficiency and recent iterations of classic multi-frame denoising methods are even quick to run on commodity smartphones [20, 27, 48]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.608, + 0.909, + 0.836 + ], + "angle": 0, + "content": "Machine Learning in Video Denoising. In recent years, it has become increasingly popular to tackle video denoising with the help of machine learning. This includes methods that, like traditional video denoisers, perform explicit matching [13, 44, 49] as well as those where correspondences are only implicit [24, 32, 42]. And while most of these methods denoise each frame independently, there are also IIR-like approaches where information is passed along such that previously denoised frames can guide future ones [5, 6]. On the model architecture side, the rise of attention in the vision and language community has also led to various attention mechanisms for video restoration [4, 26, 29, 41, 45]. But with machine learning there is a catch, if an input has a noise pattern that has not been seen during training then the denoising is unpredictable." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Noise Simulation. A common approach to facilitate better generalizability on in-the-wild footage is to employ a more sophisticated degradation pipeline when adding synthetic noise to the ground truth. This includes not only do" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2085" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.104, + 0.091, + 0.506, + 0.106 + ], + "angle": 0, + "content": "Noise Profiling on an Anchor Frame Through a Hypernetwork" + }, + { + "type": "image", + "bbox": [ + 0.1, + 0.108, + 0.905, + 0.402 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.406, + 0.907, + 0.504 + ], + "angle": 0, + "content": "Figure 4. Video denoising fundamentally first needs to analyze the noise and then remove it. We mimic this in our pipeline by first estimating a noise profile on a random anchor frame (top left in green) before using this profile to denoise the video (bottom in yellow). Specifically, we leverage a hypernetwork configuration where the noise profile \\(\\theta\\) is essentially the parameters of the subsequent denoiser. That is, our denoiser is a traditional pipeline consisting of (1) a Wiener filter that performs temporal denoising of neighboring frames that were aligned via optical flow and (2) a bilateral Laplacian pyramid filter for spatial denoising of the temporally merged frames, where a small neural network \\(\\mathcal{P}(\\cdot ;\\theta)\\) predicts spatially-varying parameters for the Wiener merger and the bilateral filters. This separation of concerns improves the overall efficiency since it avoids having to redundantly analyze the noise over and over again." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.515, + 0.483, + 0.592 + ], + "angle": 0, + "content": "ing one but two rounds of degradations [46], shuffling the degradations [51], or integrating more advanced operators such as tone mapping [52]. Although these approaches improve the ability to denoise real-world videos, they can only reduce but not entirely resolve the generalizability gap." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.597, + 0.483, + 0.763 + ], + "angle": 0, + "content": "Self-Supervised Learning. Another strategy to improve the generalizability in video denoising is to adopt self-supervision on real-world videos without requiring access to a ground truth. For example, this can be achieved by leveraging the nature of regression losses [25] or by building on the blind spot idea [14, 17, 23, 40]. While these approaches make it possible to train on noisy in-the-wild videos, it is not necessarily feasible to collect a dataset that contains all possible types of noise that occur in the real world. Therefore, self-supervision often also relies on test-time adaptation which comes with its own challenges." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.783, + 0.182, + 0.799 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Video denoising fundamentally first needs to analyze the noise and then remove it. In contrast, recent approaches perform both tasks at once by utilizing a single neural network that is given the noisy input frames and is tasked with producing the clean output. The noise profile in a video does typically not change over time though, yet these ap" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.516, + 0.905, + 0.545 + ], + "angle": 0, + "content": "proaches independently denoise one frame after another so they have to analyze the noise over and over again." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.547, + 0.906, + 0.742 + ], + "angle": 0, + "content": "To avoid this redundancy, we follow the natural separation of analyzing the noise and then removing it. As shown in Fig. 4, we first estimate the noise profile on an anchor frame and then denoise the video using the estimated profile. Specifically, we leverage a hypernetwork setup where the noise profile \\(\\theta\\) provides the parameters of the subsequent denoiser. That is, our denoiser is a traditional pipeline consisting of (1) a Wiener filter that performs temporal denoising of neighboring frames that were aligned via optical flow and (2) a bilateral Laplacian pyramid filter for spatial denoising of the temporally merged frames, where a small neural network \\(\\mathcal{P}(\\cdot ;\\theta)\\) predicts spatially-varying parameters for the Wiener merger and the bilateral filters." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.744, + 0.905, + 0.774 + ], + "angle": 0, + "content": "We will subsequently discuss these parts in turn before providing more details on user control and the training." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.786, + 0.665, + 0.803 + ], + "angle": 0, + "content": "3.1. Noise Profiling" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Since noise is not necessarily spatially uniform but often dependent on the signal, we argue that a noise profile should not be just a fixed set of values for the entire input but rather be a descriptor that makes it possible to derive spatially-varying denoising parameters. We see this as an ideal use case for hypernetworks [19], where a noise profiler esti" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "2086" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.093, + 0.088, + 0.227, + 0.22 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.145, + 0.223, + 0.177, + 0.235 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image", + "bbox": [ + 0.229, + 0.089, + 0.365, + 0.22 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.238, + 0.223, + 0.357, + 0.245 + ], + "angle": 0, + "content": "reduced temporal only \\((0.2\\cdot \\sigma^2,0\\cdot \\sigma_d,0\\cdot \\sigma_r)\\)" + }, + { + "type": "image", + "bbox": [ + 0.368, + 0.089, + 0.501, + 0.22 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.383, + 0.223, + 0.484, + 0.245 + ], + "angle": 0, + "content": "temporal only \\((1\\cdot \\sigma^2,0\\cdot \\sigma_d,0\\cdot \\sigma_r)\\)" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.089, + 0.638, + 0.22 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.223, + 0.63, + 0.245 + ], + "angle": 0, + "content": "reduced spatial only \\((0\\cdot \\sigma^{2},0.2\\cdot \\sigma_{d},0.2\\cdot \\sigma_{r})\\)" + }, + { + "type": "image", + "bbox": [ + 0.64, + 0.089, + 0.775, + 0.22 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.658, + 0.223, + 0.757, + 0.245 + ], + "angle": 0, + "content": "spatial only \\((0\\cdot \\sigma^{2},1\\cdot \\sigma_{d},1\\cdot \\sigma_{r})\\)" + }, + { + "type": "image", + "bbox": [ + 0.777, + 0.089, + 0.912, + 0.22 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.795, + 0.223, + 0.895, + 0.245 + ], + "angle": 0, + "content": "default settings \\((1\\cdot \\sigma^{2},1\\cdot \\sigma_{d},1\\cdot \\sigma_{r})\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.245, + 0.907, + 0.3 + ], + "angle": 0, + "content": "Figure 5. By scaling the denoising parameters estimated by \\(\\mathcal{P}(\\cdot ;\\theta)\\), our framework makes it easy to control the amount of temporal denoising (through \\(\\sigma^2\\)) as well as the amount of spatial denoising (through \\(\\sigma_d\\) and \\(\\sigma_r\\)). We scale the denoising parameters equivalently for the luma and chroma channels in this sample, but they can be scaled independently for more control. Please see the supplementary for a video demo that demonstrates how users can control the denoising in an interactive manner. Footage provided by Amit Zinman." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.31, + 0.484, + 0.385 + ], + "angle": 0, + "content": "mates the weights \\(\\theta\\) of a small neural network \\(\\mathcal{P}(\\cdot ;\\theta)\\) that predicts spatially-varying denoising parameters. However, instead of directly predicting \\(\\theta\\), we stabilize the training through NPA [35] where \\(\\theta = \\theta^0 +\\Delta \\theta\\) with \\(\\theta^0\\) being learnable parameters and \\(\\Delta \\theta\\) being the hypernet prediction." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.386, + 0.484, + 0.522 + ], + "angle": 0, + "content": "Intuitively, analyzing the noise profile of a given image is a mixture of low-level image processing and high-level semantics. Specifically, one first needs to understand what an image depicts before being able to examine the details to discern between unintended noise and actual texture. For this reason, we leverage a pre-trained backbone \\(\\mathcal{B}\\) in the form of a ConvNext [30] with a random MLP head to serve as the hypernet. We have found this backbone to be crucial for the quality of the predicted noise profile." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.522, + 0.484, + 0.613 + ], + "angle": 0, + "content": "Lastly, which anchor frame should we choose to analyze the noise of a given video? We would like this choice not to matter, such that the denoising result is independent of having to choose a \"good\" anchor. To achieve this, we always pick the first frame of a video as the anchor frame because it is as good as any, and then utilize a consistency loss" + }, + { + "type": "equation", + "bbox": [ + 0.202, + 0.625, + 0.483, + 0.645 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {c s t s y} = \\left\\| \\Delta \\theta - \\Delta \\theta_ {i} \\right\\| _ {2}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.655, + 0.484, + 0.686 + ], + "angle": 0, + "content": "where we encourage the hypernet prediction of our anchor frame \\(\\Delta \\theta\\) and any random neighbor \\(\\Delta \\theta_{i}\\) to be equal." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.696, + 0.282, + 0.712 + ], + "angle": 0, + "content": "3.2. Denoising the Video" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.483, + 0.796 + ], + "angle": 0, + "content": "Once we have the noise profile \\(\\theta\\), we process the input video one output frame at a tim where we obtain spatially-varying denoising parameters through a small neural network \\(\\mathcal{P}(\\cdot ;\\theta)\\) with three convolutional layers. We use these parameters to do temporal and then spatial denoising." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.484, + 0.902 + ], + "angle": 0, + "content": "For the temporal denoising, to denoise a given target frame, we first align the two preceding and the two subsequent frames. Specifically, we estimate the optical flow between the target frame and the neighbors using an off-the-shelf SpyNet [38] at a quarter of the target resolution for improved computational efficiency as well as robustness to noise, and warp the neighbors to the target frame using the" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.31, + 0.906, + 0.401 + ], + "angle": 0, + "content": "estimated optical flow. We then merge the aligned frames using a Wiener filter as in Hasinoff et al. [20] but using a tiles size of \\(8 \\times 8\\) and the noise maps from \\(\\mathcal{P}(\\cdot; \\theta)\\) instead of approximating the noise as the RMS in each tile. As is typical and to facilitate more control, we perform Wiener filtering on the luma and the chroma channels independently." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.401, + 0.906, + 0.566 + ], + "angle": 0, + "content": "We then apply spatial denoising to the temporally denoised target frame through a bilateral Laplacian pyramid filter with three levels. This borrows many ideas from multiresolution bilateral filtering [53] but using a Laplacian instead of a Gaussian pyramid and using bilateral filtering throughout instead of wavelet thresholding for the first pyramid levels. More specifically, we have \\(\\mathcal{P}(\\cdot ;\\theta)\\) estimate noise map pyramids and then use these as \\(\\sigma_{s}\\) and \\(\\sigma_r\\) for the bilateral filtering. Furthermore, just like with the Wiener filter, we perform the denoising separately on the luma and the chroma channels to facilitate user control." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.576, + 0.651, + 0.59 + ], + "angle": 0, + "content": "3.3. User Control" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.6, + 0.906, + 0.795 + ], + "angle": 0, + "content": "Through interviews with creative professionals, we have found that they want to assert their artistic expression when denoising a clip. That is, they often have to decide to leave some noise in the footage in favor of over-smoothing and vice versa. The key to facilitating user control in our method is the spatially-varying parameters from \\(\\mathcal{P}(\\cdot ;\\theta)\\) that guide the temporal and spatial denoising. Specifically, we have two of these maps for \\(\\sigma^2\\) in the Wiener filter (one map for chroma and one for luma) and in the bilateral Laplacian pyramid filter we have two map pyramids for \\(\\sigma_{d}\\) as well as two map pyramids for \\(\\sigma_r\\) (one map pyramid for chroma and one for luma). This makes for a total of six sets of parameters where we can facilitate user control." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.906, + 0.901 + ], + "angle": 0, + "content": "How do we attenuate these parameters though? As demonstrated in Fig. 5, we have found that simply scaling the spatially-varying denoising parameters uniformly works reasonably well, so we can just provide six tunable knobs to end users. And since our denoising framework runs in real time, changing the knobs provides instantaneous feedback which makes it easy to attain the desired result." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2087" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.303, + 0.108 + ], + "angle": 0, + "content": "3.4. Augmentation Pipeline" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.114, + 0.482, + 0.25 + ], + "angle": 0, + "content": "To train our model, we need pairs of clean and noisy frame sequences just like any other denoiser that leverages machine learning without self-supervision. Specifically, we adopt the typical paradigm of taking a set of \"noise-free\" videos that serve as the ground truth, and augmenting them with various degradations to obtain the corresponding noisy inputs. For this to work well, we have found it to be crucial to include temporal compression in this degradation pipeline since it makes the noise temporally correlated." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.25, + 0.482, + 0.386 + ], + "angle": 0, + "content": "But let's start with noise-free videos, which we obtain from the REDS dataset [33]. It provides 240 videos and was shot at a relatively high 120 frames per second, allowing us to augment the frame rate via sub-sampling. We extract random frame sequences from these videos and first degrade them through additive white Gaussian noise with a variance sampled from \\(\\mathcal{U}(1,50)\\). We then transcode them using an H.264 codec through libx264 with a CRF sampled from \\(\\mathcal{U}(18,30)\\) to make the noise temporally correlated." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.386, + 0.482, + 0.538 + ], + "angle": 0, + "content": "We have found this data pipeline to work surprisingly well. So well in fact that we retrained the models that we compare to in the evaluation since we were able to get much better results on the CRVD benchmark [50] with our retrained versions than the original checkpoints. However, we did not retrain Real-ESRGAN [46] since it brings its own much more complex augmentation pipeline, and we also did not retrain MF2F [14] and UDVD [40] since these leverage sophisticated ways of supervising the models which are difficult for us to replicate and properly do justice." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.545, + 0.308, + 0.56 + ], + "angle": 0, + "content": "3.5. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.568, + 0.483, + 0.794 + ], + "angle": 0, + "content": "In terms of architecture details, we utilize the base size of a ConvNext [30] for the backbone \\(\\mathcal{B}\\) in the hypernetwork. To translate the output from this backbone to the residual noise profile \\(\\Delta \\theta\\), we utilize an MLP with five layers and PReLU activations [21] in between each layer. And for the neural network \\(\\mathcal{P}(\\cdot ;\\theta)\\) that predicts the spatially-varying denoising parameters, we leverage three convolution layers with a stride of two and PReLU activations in between. We additionally constrain the output of this network with a softmax [15] to facilitate non-negative denoising parameters. Furthermore, we have found it beneficial to not only provide the (aligned) frames to \\(\\mathcal{P}(\\cdot ;\\theta)\\) but also their gradients, approximated by a Sobel filter, as well as masks that indicate whether or not a pixel in an aligned frame is valid (it may have been warped from outside the frame) [2]." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.795, + 0.483, + 0.856 + ], + "angle": 0, + "content": "In our reconstruction loss, we minimize the difference between the temporally denoised image \\( I^{\\mathrm{T}} \\) and the ground truth \\( I^{\\mathrm{GT}} \\) as well as the difference between all levels of the spatial denoising pyramid \\( I^{\\mathrm{ST}} \\) and \\( I^{\\mathrm{GT}} \\) as" + }, + { + "type": "equation", + "bbox": [ + 0.141, + 0.863, + 0.483, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {r e c} = \\left\\| I ^ {\\mathrm {T}} - I ^ {\\mathrm {G T}} \\right\\| _ {2} + \\sum_ {l = 1} ^ {3} \\left\\| I _ {l} ^ {\\mathrm {S T}} - I _ {l} ^ {\\mathrm {G T}} \\right\\| _ {2}, \\tag {2}\n\\]" + }, + { + "type": "table", + "bbox": [ + 0.515, + 0.09, + 0.906, + 0.273 + ], + "angle": 0, + "content": "
PSNRdelta (higher PSNR is better)SSIMdelta (higher SSIM is better)LPIPSdelta (lower LPIPS is better)
NAFNet [9]30.25-0.747-0.358-
NAFNet†[9]35.32+ 5.070.937+ 0.1900.089- 0.269
FastDVDNet [42]27.89-0.565-0.502-
FastDVDNet†[42]34.63+ 6.750.914+ 0.3490.124- 0.378
TOFlow [49]25.96-0.673-0.251-
TOFlow†[49]34.08+ 8.110.903+ 0.2310.147- 0.104
BasicVSR++ [6]31.98-0.769-0.326-
BasicVSR++†[6]34.34+ 2.350.873+ 0.1040.167- 0.158
VRT [26]31.99-0.784-0.296-
VRT†[26]33.86+ 1.860.848+ 0.0640.192- 0.104
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.278, + 0.905, + 0.32 + ], + "angle": 0, + "content": "Table 1. Denoising results on the CRVD (sRGB) benchmark for various methods that we compare to, both for their original version as well as our retrained one which we denote with a \\(\\dagger\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.334, + 0.905, + 0.379 + ], + "angle": 0, + "content": "where \\( l \\) is the \\( l \\)-th level in spatial denoising pyramid. Recall that we also have a consistency loss that we outlined in Eq. (1) in the section about noise profiling." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.379, + 0.906, + 0.471 + ], + "angle": 0, + "content": "We train our model using Adam [22] with an initial learning rate of \\( 2 \\times 10^{-4} \\) that decays to \\( 1 \\times 10^{-7} \\) using a cosine annealing schedule. In total, we train the model for 400 thousand iterations with a batch size of 24 consisting of patches with \\( 512 \\times 512 \\) pixels. Since our pipeline is computationally lightweight, this takes less than two days." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.483, + 0.645, + 0.5 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.509, + 0.906, + 0.735 + ], + "angle": 0, + "content": "We quantitatively evaluate our approach on two types of videos, real and synthetic. For the former, we leverage the CRVD benchmark [50] consisting of RAW videos which we convert to sRGB using the provided deep ISP [39]. For the latter, we use the validation samples from the REDS dataset [33] which we augmented in two ways that are notably different from the AWGN subject to H.264 transcoding during training. Specifically, one way is using film grain noise [34] with AV1 transcoding, and the other way is using spatially correlated Gaussian noise with H.265 transcoding. As for evaluation metrics, we follow the typical paradigm of reporting PSNR, SSIM [47], and LPIPS [54]. Since we argue that our method is computationally efficient, we also report the frames per second (FPS) which we measured on an RTX 3090 GPU with proper CUDA synchronization." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.905, + 0.81 + ], + "angle": 0, + "content": "In terms of data for our qualitative evaluation, we reached out to industry professionals who sent us footage where they had trouble removing the noise. Since these professionals did not have a release form for all the actors, we sometimes had to anonymize faces post inference." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Lastly, we compare our approach to various kinds of denoisers. This includes image denoisers [7, 9], video denoisers [42, 49], recurrent methods [6], transformers [26], models with a sophisticated augmentation pipeline [46], as well as methods that leverage self-supervision [14, 40]. Since we have found that our augmentation pipeline works sur" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "2088" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.093, + 0.089, + 0.907, + 0.296 + ], + "angle": 0, + "content": "
ISO 1600ISO 3200ISO 6400ISO 12800ISO 25600OverallSpeed
PSNRrank (higher PSNR is better)PSNRrank (higher PSNR is better)PSNRrank (higher PSNR is better)PSNRrank (higher PSNR is better)PSNRrank (higher PSNR is better)PSNRrank (higher PSNR is better)FPSrank (higher FPS is better)
\\( SID^† \\)[7]38.85\\( 7^{th} \\) of 1037.68\\( 7^{th} \\) of 1035.82\\( 4^{th} \\) of 1033.51\\( 4^{th} \\) of 1029.18\\( 3^{rd} \\) of 1035.01\\( 4^{th} \\) of 106.95\\( 3^{rd} \\) of 10
\\( NAFNet^† \\)[9]39.48\\( 3^{rd} \\) of 1038.12\\( 5^{th} \\) of 1035.94\\( 3^{rd} \\) of 1033.53\\( 3^{rd} \\) of 1029.55\\( 2^{nd} \\) of 1035.32\\( 2^{nd} \\) of 101.69\\( 7^{th} \\) of 10
\\( FastDVDNet^† \\)[42]39.16\\( 5^{th} \\) of 1037.92\\( 6^{th} \\) of 1035.60\\( 5^{th} \\) of 1032.56\\( 5^{th} \\) of 1027.93\\( 6^{th} \\) of 1034.63\\( 5^{th} \\) of 105.72\\( 4^{th} \\) of 10
\\( TOFlow^† \\)[49]38.25\\( 8^{th} \\) of 1036.97\\( 8^{th} \\) of 1034.90\\( 7^{th} \\) of 1032.21\\( 6^{th} \\) of 1028.07\\( 5^{th} \\) of 1034.08\\( 7^{th} \\) of 102.84\\( 6^{th} \\) of 10
\\( BasicVSR++^† \\)[6]39.40\\( 4^{th} \\) of 1038.24\\( 2^{nd} \\) of 1035.55\\( 6^{th} \\) of 1031.72\\( 7^{th} \\) of 1026.78\\( 9^{th} \\) of 1034.34\\( 6^{th} \\) of 107.41\\( 2^{nd} \\) of 10
\\( VRT^† \\)[26]39.55\\( 2^{nd} \\) of 1038.12\\( 4^{th} \\) of 1034.82\\( 8^{th} \\) of 1030.77\\( 8^{th} \\) of 1026.01\\( 10^{th} \\) of 1033.86\\( 8^{th} \\) of 100.05\\( 10^{th} \\) of 10
Real-ESRGAN [46]29.98\\( 10^{th} \\) of 1028.27\\( 10^{th} \\) of 1028.04\\( 10^{th} \\) of 1027.52\\( 10^{th} \\) of 1027.63\\( 8^{th} \\) of 1028.29\\( 10^{th} \\) of 100.24\\( 8^{th} \\) of 10
UDVD [40]31.15\\( 9^{th} \\) of 1030.72\\( 9^{th} \\) of 1030.23\\( 9^{th} \\) of 1029.10\\( 9^{th} \\) of 1027.63\\( 7^{th} \\) of 1029.77\\( 9^{th} \\) of 100.16\\( 9^{th} \\) of 10
MF2F [14]39.09\\( 6^{th} \\) of 1038.20\\( 3^{rd} \\) of 10\\( \\underline{36.36} \\)\\( 1^{st} \\) of 1033.57\\( 2^{nd} \\) of 1029.04\\( 4^{th} \\) of 1035.25\\( 3^{rd} \\) of 104.62\\( 5^{th} \\) of 10
Ours - RFCVD40.35\\( 1^{st} \\) of 10\\( \\underline{38.60} \\)\\( 1^{st} \\) of 1036.28\\( 2^{nd} \\) of 1033.86\\( 1^{st} \\) of 1031.12\\( 1^{st} \\) of 1036.04\\( 1^{st} \\) of 1031.66\\( 1^{st} \\) of 10
" + }, + { + "type": "table_caption", + "bbox": [ + 0.093, + 0.298, + 0.907, + 0.327 + ], + "angle": 0, + "content": "Table 2. Video denoising results on the CRVD (sRGB) benchmark. Not only does our approach perform best overall, it is also four times faster than the second-fastest place. Please kindly see the supplementary for SSIM and LPIPS where our approach ranks first as well." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.328, + 0.907, + 0.493 + ], + "angle": 0, + "content": "
film grain noise w/ AV1 compressionspatially correlated noise w/ H.265 compressionSpeed
PSNR (higher PSNR is better)rank SSIM (higher SSIM is better)rank (lower LPIPS is better)LPIPS rank (lower LPIPS is better)rank (higher PSNR is better)rank (higher PSNR is better)SSIM (higher SSIM is better)rank (lower LPIPS is better)LPIPS rank (lower LPIPS is better)rank (higher FPS is better)FPS rank (higher FPS is better)
\\( SID^† \\)[7]27.05 5thof 70.664 5thof 70.293 4thof 728.44 3rdof 70.771 4thof 70.282 5thof 715.00 3rdof 7
\\( NAFNet^† \\)[9]27.16 4thof 70.672 4thof 70.294 5thof 728.40 4thof 70.764 6thof 70.257 3rdof 73.812 6thof 7
\\( FastDVDNet^† \\)[42]27.39 3rdof 70.686 3rdof 70.272 3rdof 727.92 6thof 70.769 5thof 70.296 6thof 712.09 4thof 7
\\( TOFlow^† \\)[49]28.15 2ndof 70.750 2ndof 70.220 1stof 728.39 5thof 70.788 3rdof 70.258 4thof 75.665 5thof 7
\\( BasicVSR++^† \\)[6]26.90 6thof 70.651 6thof 70.313 6thof 727.48 7thof 70.728 7thof 70.358 7thof 715.32 2ndof 7
\\( VRT^† \\)[26]26.55 7thof 70.629 7thof 70.331 7thof 728.78 2ndof 70.803 2ndof 70.206 1stof 70.105 7thof 7
Ours - RFCVD28.59 1stof 70.774 1stof 70.247 2ndof 728.93 1stof 70.808 1stof 70.239 2ndof 769.73 1stof 7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.495, + 0.907, + 0.525 + ], + "angle": 0, + "content": "Table 3. Testing the generalizability when training on AWGN with H.264 transcoding but testing with two different degradation schemes. Since our approach is based on classic methods, the denoising itself only has a few parameters which naturally reduces domain gaps." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.55, + 0.483, + 0.658 + ], + "angle": 0, + "content": "prisingly well, we have retrained all methods that neither leverage a sophisticated augmentation pipeline already nor utilize self-supervision. This makes for a more fair comparison since, as shown in Tab. 1, we have improved their performance on the CRVD benchmark with our retrained version. Throughout this paper, we denote all retrained models with a \\(\\dagger\\) to prevent any potential confusions." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.666, + 0.325, + 0.684 + ], + "angle": 0, + "content": "4.1. Quantitative Comparison" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.689, + 0.483, + 0.903 + ], + "angle": 0, + "content": "Please see Tab. 2 for a summary of the quantitative evaluation on the CRVD benchmark. In short, not only does our approach perform best overall, it is also four times faster than the second-fastest method. We attribute these favorable results to our utilization of classic denoising techniques which greatly helps with the ability to generalize to unseen noise patterns. Specifically, the denoising itself is driven by only a few parameters which naturally reduces overfitting and helps with bridging domain gaps. To test this hypothesis, we conducted an experiment with synthetic data where all models are trained on AWGN with H.264 transcoding but tested on two different degradation pipelines. As shown in Tab. 3, our approach performs favorably in this evaluation as well which supports our initial hypothesis." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.55, + 0.736, + 0.567 + ], + "angle": 0, + "content": "4.2. Qualitative Comparison" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.574, + 0.907, + 0.712 + ], + "angle": 0, + "content": "To test video denoising in the real world, we reached out to various creative professionals who kindly shared footage where they had trouble removing the noise. We show a representative excerpt of the denoising results on this footage in Fig. 6 but kindly refer to the supplementary which includes video results of all methods. Like with the quantitative evaluation, we have found that our approach performs favorably, and once again we hypothesize that this is due to the underlying utilization of classic denoising techniques." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.724, + 0.719, + 0.74 + ], + "angle": 0, + "content": "4.3. Ablative Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.749, + 0.907, + 0.885 + ], + "angle": 0, + "content": "One of the main things we wondered ourselves was the performance with respect to the choice of anchor frame. After all, the noise profile of I-frames can be quite different from P- and B-frames. Nevertheless as shown in Tab. 4 (top), we have always used the first frame as the anchor throughout the experiments section and would have drawn the same conclusions with the middle or the last frame. And as shown in Tab. 4 (bottom), we have also found that \\(\\mathcal{L}_{cstsy}\\) helps with providing a slightly more consistent experience." + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.886, + 0.905, + 0.902 + ], + "angle": 0, + "content": "A key part of our integration of classic denoising tech" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2089" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.093, + 0.088, + 0.916, + 0.58 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.587, + 0.908, + 0.63 + ], + "angle": 0, + "content": "Figure 6. Video denoising results from exemplary still frames. Due to space constraints, we only share the results from a representative subset of the methods that we compare to. Please kindly refer to the supplementary to find video results of all methods. We are grateful that Amit Zinman (top row), Robert Kjettrup (middle row), and an anonymous artist (last row) were willing to provide test footage." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.655, + 0.483, + 0.776 + ], + "angle": 0, + "content": "niques into a machine learning pipeline is that we account for spatially-varying noise. To assess the importance of this design decision, we trained an ablation that only predicts a set of scalar denoising parameters. As expected and as shown in Tab. 6, this \"w/o spat. varying\" experiment exhibits a significant drop in quality. As shown in Fig. 7, this ablation produces much more blurry denoising results since it cannot easily account for signal-dependent noise." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.483, + 0.901 + ], + "angle": 0, + "content": "To best support \\(\\mathcal{P}(\\cdot ;\\theta)\\) in predicting the denoising parameters, we not only provide it access to the (merged) input frames but also their gradients as well as a mask indicating whether or not a pixel in an aligned frame is valid. To test the effectiveness of this design, we trained two ablations where we refrained from providing one (\"w/o img. gradients\") or the other (\"w/o align. mask\") and found that, as shown in Tab. 6, this context is indeed beneficial." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.655, + 0.907, + 0.852 + ], + "angle": 0, + "content": "We argue that analyzing the noise profile of a given image is a mixture of both low-level image processing and high-level semantics, because one first needs an understanding of what an image depicts before being able to go into the low-level details to discern between unintended noise and actual texture. For this reason, we leverage a classification backbone \\(\\mathcal{B}\\) in the hypernetwork that is predicting the noise profile \\(\\theta\\), and we trained an ablation without this backbone to evaluate this hypothesis. As intuitively expected and as shown in Tab. 6, this \"w/o backbone \\(\\mathcal{B}\\)\" experiment indeed exhibits a significant drop in quality. This drop is also visually exemplified in Fig. 7, where the lack of this backbone leads to the noise not fully being removed." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Lastly, we not only fine-tune the aforementioned backbone \\(\\mathcal{B}\\) but also stabilize the training of the hypernetwork through NPA [35]. To analyze the effect of these measures," + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2090" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.093, + 0.089, + 0.485, + 0.251 + ], + "angle": 0, + "content": "
PSNRdelta (higher PSNR is better)SSIMdelta (higher SSIM is better)LPIPSdelta (lower LPIPS is better)
Avg. w/ Lcstsy35.95-0.9477-0.0757-
first frame ← Ours36.04+ 0.090.9472- 0.00050.0763+ 0.0006
middle frame35.92- 0.030.9480+ 0.00030.0753- 0.0003
last frame35.90- 0.060.9480+ 0.00030.0754- 0.0003
Avg. w/o Lcstsy35.86-0.9460-0.0754-
first frame35.80- 0.050.9454- 0.00050.0750- 0.0004
middle frame36.02+ 0.160.9473+ 0.00140.0753- 0.0001
last frame35.75- 0.110.9451- 0.00080.0759+ 0.0005
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.255, + 0.483, + 0.298 + ], + "angle": 0, + "content": "Table 4. Denoising results on the CRVD benchmark with respect to the anchor frame choice. We find that our approach is relatively robust to this choice (top), partly thanks to \\( \\mathcal{L}_{cstsy} \\) (bottom)." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.317, + 0.485, + 0.388 + ], + "angle": 0, + "content": "
PSNRdelta (higher PSNR is better)SSIMdelta (higher SSIM is better)LPIPSdelta (lower LPIPS is better)
Ours36.04-0.9472-0.0763-
fine-tuned on GT36.75+ 0.700.9494+ 0.00220.0591- 0.0172
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.391, + 0.483, + 0.434 + ], + "angle": 0, + "content": "Table 5. Fine-tuning our estimated denoising parameters on the ground truth of the CRVD benchmark indicates that our estimates are good but not perfect if we had access to an oracle." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.458, + 0.483, + 0.505 + ], + "angle": 0, + "content": "we trained an ablation for the former (\"w/o fine-tuning \\(\\mathcal{B}\\)\" as well as the latter (\"w/o NPA\") and found that, as shown in Tab. 6, both contribute to the denoising quality." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.522, + 0.218, + 0.536 + ], + "angle": 0, + "content": "4.4. Limitations" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.548, + 0.483, + 0.669 + ], + "angle": 0, + "content": "While our noise profile \\(\\theta\\) with the subsequent denoising parameter prediction from \\(\\mathcal{P}(\\cdot ;\\theta)\\) works reasonably well as demonstrated throughout the experiments section, it is far from perfect. Specifically and as shown in Tab. 5, we fine-tuned the predicted denoising parameters on the ground truth and found that this hypothetical oracle is able to improve our results by a rather significant margin. Nevertheless, we leave bridging this gap to future research." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.672, + 0.483, + 0.777 + ], + "angle": 0, + "content": "In the case where the noise profile is entirely known a priori, such as when developing a denoiser for a specific camera model subject to a constrained encoder, any reasonable deep learning method would provide better results than our approach. And with a baked-in noise profile it would perhaps even be faster. However, such a method would in turn be limited in terms of robustness and control." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.78, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Lastly, and on a more practical note, deploying hypernetworks in end-user applications is not well supported by the common inference ecosystems yet. Specifically, popular libraries like CoreML, WinML, or OpenVINO expect the model weights to be fixed which is obviously not the case for hypernetworks. As such, while our approach is easy to implement within frameworks like PyTorch, it is not straightforward to deploy in production environments." + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.089, + 0.613, + 0.165 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.166, + 0.612, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.213, + 0.612, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.549, + 0.293, + 0.58, + 0.305 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image", + "bbox": [ + 0.613, + 0.089, + 0.71, + 0.165 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.613, + 0.166, + 0.71, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.613, + 0.213, + 0.71, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.616, + 0.293, + 0.709, + 0.305 + ], + "angle": 0, + "content": "w/o spat. varying" + }, + { + "type": "image", + "bbox": [ + 0.712, + 0.089, + 0.807, + 0.165 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.712, + 0.166, + 0.807, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.712, + 0.213, + 0.807, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.715, + 0.293, + 0.804, + 0.305 + ], + "angle": 0, + "content": "w/o backbone \\(\\mathcal{B}\\)" + }, + { + "type": "image", + "bbox": [ + 0.81, + 0.089, + 0.905, + 0.165 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.81, + 0.166, + 0.905, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.81, + 0.213, + 0.905, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.843, + 0.293, + 0.871, + 0.304 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.311, + 0.905, + 0.34 + ], + "angle": 0, + "content": "Figure 7. Visual comparison of the two most significant ablations on a representative sample from the CRVD dataset." + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.348, + 0.906, + 0.49 + ], + "angle": 0, + "content": "
PSNRdelta (higher PSNR is better)SSIMdelta (higher SSIM is better)LPIPSdelta (lower LPIPS is better)
Ours36.04-0.9472-0.0763-
w/o spat. varying33.62-2.420.9359-0.01120.1184+0.0421
w/o img. gradients35.61-0.440.9469-0.00030.0756-0.0007
w/o align. mask35.26-0.780.9468-0.00030.0740-0.0023
w/o backbone B34.38-1.670.9059-0.04120.1444+0.0680
w/o fine-tuning B35.71-0.340.9464-0.00080.0732-0.0032
w/o NPA [35]35.83-0.210.9453-0.00180.0814+0.0051
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.493, + 0.906, + 0.521 + ], + "angle": 0, + "content": "Table 6. Various ablative experiments, please see Section 4.3 for more context and details of each individual ablation." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.542, + 0.634, + 0.558 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.568, + 0.907, + 0.718 + ], + "angle": 0, + "content": "We show that traditional denoising techniques are still very much relevant in the modern machine learning world. Not only have we found them to be robust and fast, but also controllable which is an important property in many video denoising applications where creative professionals want to assert their artistic expression. To bridge the gap between classic denoising and modern approaches, we train a model to estimate the various parameters of a traditional denoising approach for a given input video, which is otherwise not only tedious but also requires a certain amount of skill." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.72, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Furthermore, in an effort to make best use of the computational efficiency of classic denoising techniques, our method separates the analysis of the noise from the denoising in order to avoid having to estimate the noise profile over and over again as is typical for machine learning approaches to video denoising. Lastly, we have found a simple yet effective degradation pipeline for training video denoiseers based on AWGN subject to H.264 video transcoding. This pipeline ensures that the noise is temporally correlated just like it often is the case in the real world, and we have found it not only to work well for our approach but we have also been able to retrain and improve existing models." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "2091" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.228 + ], + "angle": 0, + "content": "Acknowledgement. This work was supported in part by the National Natural Science Foundation of China (62306153, 62225604), the Fundamental Research Funds for the Central Universities (Nankai University, 070-63243143), the Natural Science Foundation of Tianjin, China (24JCJQJC00020), and Shenzhen Science and Technology Program (JCYJ20240813114237048). The computational devices is partly supported by the Supercomputing Center of Nankai University (NKSC)." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.243, + 0.188, + 0.259 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.267, + 0.484, + 0.31 + ], + "angle": 0, + "content": "[1] Eric P Bennett and Leonard McMillan. Video enhancement using per-pixel virtual exposures. In ACM SIGGRAPH 2005 Papers, 2005. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.311, + 0.484, + 0.354 + ], + "angle": 0, + "content": "[2] Goutam Bhat, Michael Gharbi, Jiawen Chen, Luc Van Gool, and Zhihao Xia. Self-supervised burst super-resolution. In ICCV, pages 10571-10580. IEEE, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.354, + 0.484, + 0.396 + ], + "angle": 0, + "content": "[3] Peter J Burt and Edward H Adelson. The laplacian pyramid as a compact image code. In Readings in computer vision, 1987. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.398, + 0.484, + 0.438 + ], + "angle": 0, + "content": "[4] Jiezhang Cao, Yawei Li, Kai Zhang, and Luc Van Gool. Video super-resolution transformer. arXiv:2106.06847, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.441, + 0.484, + 0.495 + ], + "angle": 0, + "content": "[5] Kelvin CK Chan, Xintao Wang, Ke Yu, Chao Dong, and Chen Change Loy. Basicvsr: The search for essential components in video super-resolution and beyond. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.497, + 0.484, + 0.554 + ], + "angle": 0, + "content": "[6] Kelvin CK Chan, Shangchen Zhou, Xiangyu Xu, and Chen Change Loy. Basicvsr++: Improving video superresolution with enhanced propagation and alignment. In CVPR, 2022. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.556, + 0.484, + 0.584 + ], + "angle": 0, + "content": "[7] Chen Chen, Qifeng Chen, Jia Xu, and Vladlen Koltun. Learning to see in the dark. In CVPR, 2018. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.585, + 0.484, + 0.626 + ], + "angle": 0, + "content": "[8] Jiawen Chen, Sylvain Paris, and Frédo Durand. Real-time edge-aware image processing with the bilateral grid. ACM TOG, 2007. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.628, + 0.484, + 0.669 + ], + "angle": 0, + "content": "[9] Liangyu Chen, Xiaojie Chu, Xiangyu Zhang, and Jian Sun. Simple baselines for image restoration. In ECCV, 2022. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.671, + 0.484, + 0.714 + ], + "angle": 0, + "content": "[10] Kostadin Dabov, Alessandro Foi, Vladimir Katkovnik, and Karen Egiazarian. Image denoising with block-matching and 3d filtering. In Image Processing, 2006. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.715, + 0.484, + 0.757 + ], + "angle": 0, + "content": "[11] Kostadin Dabov, Alessandro Foi, and Karen Egiazarian. Video denoising by sparse 3d transform domain collaborative filtering. EURASIP, 2007." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.758, + 0.484, + 0.801 + ], + "angle": 0, + "content": "[12] Kostadin Dabov, Alessandro Foi, Vladimir Katkovnik, and Karen Egiazarian. Image denoising by sparse 3-d transform-domain collaborative filtering. IEEE TIP, 2007. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.802, + 0.484, + 0.843 + ], + "angle": 0, + "content": "[13] Axel Davy, Thibaud Ehret, Jean-Michel Morel, Pablo Arias, and Gabriele Facciolo. A non-local cnn for video denoising. In ICIP, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.845, + 0.484, + 0.9 + ], + "angle": 0, + "content": "[14] Valery Dewil, Jérémy Anger, Axel Davy, Thibaud Ehret, Gabriele Facciolo, and Pablo Arias. Self-supervised training for blind multi-frame video denoising. In WACV, 2021. 1, 2, 3, 5, 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.267, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.906, + 0.146 + ], + "angle": 0, + "content": "[15] Charles Dugas, Yoshua Bengio, François Bélisle, Claude Nadeau, and René Garcia. Incorporating second-order functional knowledge for better option pricing. In NIPS, 2000. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.15, + 0.906, + 0.19 + ], + "angle": 0, + "content": "[16] Jana Ehmann, Lun-Cheng Chu, Sung-Fang Tsai, and Chia-Kai Liang. Real-time video denoising on mobile phones. In ICIP, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.193, + 0.906, + 0.235 + ], + "angle": 0, + "content": "[17] Thibaud Ehret, Axel Davy, Jean-Michel Morel, Gabriele Facciolo, and Pablo Arias. Model-blind video denoising via frame-to-frame training. In CVPR, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.236, + 0.906, + 0.265 + ], + "angle": 0, + "content": "[18] Ruturaj G Gavaskar and Kunal N Chaudhury. Fast adaptive bilateral filtering. IEEE TIP, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.267, + 0.906, + 0.294 + ], + "angle": 0, + "content": "[19] David Ha, Andrew M. Dai, and Quoc V. Le. Hypernetworks. arXiv:1609.09106, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.297, + 0.906, + 0.365 + ], + "angle": 0, + "content": "[20] Samuel W Hasinoff, Dillon Sharlet, Ryan Geiss, Andrew Adams, Jonathan T Barron, Florian Kainz, Jiawen Chen, and Marc Levoy. Burst photography for high dynamic range and low-light imaging on mobile cameras. ACM TOG, 2016. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.368, + 0.906, + 0.41 + ], + "angle": 0, + "content": "[21] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Delving deep into rectifiers: Surpassing human-level performance on imagenet classification. In ICCV, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.411, + 0.906, + 0.439 + ], + "angle": 0, + "content": "[22] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.441, + 0.906, + 0.482 + ], + "angle": 0, + "content": "[23] Samuli Laine, Tero Karras, Jaakko Lehtinen, and Timo Aila. High-quality self-supervised deep image denoising. In NeurIPS, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.485, + 0.906, + 0.527 + ], + "angle": 0, + "content": "[24] Seunghwan Lee, Donghyeon Cho, Jiwon Kim, and Tae Hyun Kim. Restore from restored: Video restoration with pseudo clean video. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.529, + 0.906, + 0.583 + ], + "angle": 0, + "content": "[25] Jaakko Lehtinen, Jacob Munkberg, Jon Hasselgren, Samuli Laine, Tero Karras, Miika Aittala, and Timo Aila. Noise2noise: Learning image restoration without clean data. In ICML, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.586, + 0.906, + 0.641 + ], + "angle": 0, + "content": "[26] Jingyun Liang, Jiezhang Cao, Yuchen Fan, Kai Zhang, Rakesh Ranjan, Yawei Li, Radu Timofte, and Luc Van Gool. Vrt: A video restoration transformer. IEEE TIP, 2024. 1, 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.643, + 0.906, + 0.699 + ], + "angle": 0, + "content": "[27] Orly Liba, Kiran Murthy, Yun-Ta Tsai, Tim Brooks, Tianfan Xue, Nikhil Karnad, Qiurui He, Jonathan T Barron, Dillon Sharlet, Ryan Geiss, et al. Handheld mobile photography in very low light. ACM TOG, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.701, + 0.906, + 0.742 + ], + "angle": 0, + "content": "[28] Ce Liu and William T Freeman. A high-quality video denoising algorithm based on reliable motion estimation. In ECCV, 2010. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.744, + 0.906, + 0.8 + ], + "angle": 0, + "content": "[29] Ding Liu, Zhaowen Wang, Yuchen Fan, Xianming Liu, Zhangyang Wang, Shiyu Chang, and Thomas Huang. Robust video super-resolution with learned temporal dynamics. In ICCV, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.802, + 0.906, + 0.843 + ], + "angle": 0, + "content": "[30] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In CVPR, 2022. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.845, + 0.906, + 0.9 + ], + "angle": 0, + "content": "[31] Matteo Maggioni, Vladimir Katkovnik, Karen Egiazarian, and Alessandro Foi. Nonlocal transform-domain filter for volumetric data denoising and reconstruction. IEEE TIP, 2012. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.906, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "2092" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.146 + ], + "angle": 0, + "content": "[32] Matteo Maggioni, Yibin Huang, Cheng Li, Shuai Xiao, Zhongqian Fu, and Fenglong Song. Efficient multi-stage video denoising with recurrent spatio-temporal fusion. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.15, + 0.482, + 0.205 + ], + "angle": 0, + "content": "[33] Seungjun Nah, Sungyong Baik, Seokil Hong, Gyeongsik Moon, Sanghyun Son, Radu Timofte, and Kyoung Mu Lee. NTIRE 2019 challenge on video deblurring and superresolution: Dataset and study. In CVPRW, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.208, + 0.482, + 0.235 + ], + "angle": 0, + "content": "[34] Alasdair Newson, Noura Faraj, Bruno Galerne, and Julie Delon. Realistic film grain rendering. IPOL, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.238, + 0.483, + 0.278 + ], + "angle": 0, + "content": "[35] Jose Javier Gonzalez Ortiz, John V. Guttag, and Adrian V. Dalca. Non-proportional parametrizations for stable hypernetwork learning. arXiv:2304.07645, 2023. 3, 4, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.281, + 0.482, + 0.333 + ], + "angle": 0, + "content": "[36] Sylvain Paris, Pierre Kornprobst, Jack Tumblin, Frédo Durand, et al. Bilateral filtering: Theory and applications. Foundations and Trends® in Computer Graphics and Vision, 2009. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.338, + 0.482, + 0.393 + ], + "angle": 0, + "content": "[37] Andrea Petreto, Thomas Romera, Florian Lemaitre, Ian Masliah, Boris Gaillard, Manuel Bouyer, Quentin L Meunier, and Lionel Lacassagne. A new real-time embedded video denoising algorithm. In DASIP, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.396, + 0.482, + 0.422 + ], + "angle": 0, + "content": "[38] Anurag Ranjan and Michael J. Black. Optical flow estimation using a spatial pyramid network. In CVPR, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.425, + 0.482, + 0.465 + ], + "angle": 0, + "content": "[39] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In MICCAI, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.469, + 0.482, + 0.523 + ], + "angle": 0, + "content": "[40] Dev Yashpal Sheth, Sreyas Mohan, Joshua L Vincent, Ramon Manzorro, Peter A Crozier, Mitesh M Khapra, Eero P Simoncelli, and Carlos Fernandez-Granda. Unsupervised deep video denoising. In ICCV, 2021. 1, 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.526, + 0.482, + 0.552 + ], + "angle": 0, + "content": "[41] Maitreya Suin and AN Rajagopalan. Gated spatio-temporal attention-guided video deblurring. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.556, + 0.482, + 0.596 + ], + "angle": 0, + "content": "[42] Matias Tassano, Julie Delon, and Thomas Veit. Fastdvnet: Towards real-time deep video denoising without flow estimation. In CVPR, 2020. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.599, + 0.482, + 0.626 + ], + "angle": 0, + "content": "[43] Carlo Tomasi and Roberto Manduchi. Bilateral filtering for gray and color images. In ICCV, 1998. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.629, + 0.482, + 0.669 + ], + "angle": 0, + "content": "[44] Gregory Vaksman, Michael Elad, and Peyman Milanfar. Patch craft: Video denoising by deep modeling and patch matching. In ICCV, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.672, + 0.482, + 0.712 + ], + "angle": 0, + "content": "[45] Xintao Wang, Kelvin CK Chan, Ke Yu, Chao Dong, and Chen Change Loy. Edvr: Video restoration with enhanced deformable convolutional networks. In CVPRW, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.716, + 0.482, + 0.757 + ], + "angle": 0, + "content": "[46] Xintao Wang, Liangbin Xie, Chao Dong, and Ying Shan. Real-esrgan: Training real-world blind super-resolution with pure synthetic data. In ICCV Workshop, 2021. 1, 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.759, + 0.482, + 0.8 + ], + "angle": 0, + "content": "[47] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE TIP, 2004. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.803, + 0.482, + 0.856 + ], + "angle": 0, + "content": "[48] Bartlomiej Wronski, Ignacio Garcia-Dorado, Manfred Ernst, Damien Kelly, Michael Krainin, Chia-Kai Liang, Marc Levoy, and Peyman Milanfar. Handheld multi-frame superresolution. ACM TOG, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.86, + 0.482, + 0.899 + ], + "angle": 0, + "content": "[49] Tianfan Xue, Baian Chen, Jiajun Wu, Donglai Wei, and William T Freeman. Video enhancement with task-oriented flow. IJCV, 2019. 2, 5, 6" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "[50] Huanjing Yue, Cong Cao, Lei Liao, Ronghe Chu, and Jingyu Yang. Supervised raw video denoising with a benchmark dataset on dynamic scenes. In CVPR, 2020. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.136, + 0.905, + 0.177 + ], + "angle": 0, + "content": "[51] Kai Zhang, Jingyun Liang, Luc Van Gool, and Radu Timofte. Designing a practical degradation model for deep blind image super-resolution. In CVPR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.179, + 0.905, + 0.233 + ], + "angle": 0, + "content": "[52] Kai Zhang, Yawei Li, Jingyun Liang, Jiezhang Cao, Yu-lun Zhang, Hao Tang, Deng-Ping Fan, Radu Timofte, and Luc Van Gool. Practical blind image denoising via swim-conv-unet and data synthesis. MIR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.235, + 0.904, + 0.262 + ], + "angle": 0, + "content": "[53] Ming Zhang and Bahadir K. Gunturk. Multiresolution bilateral filtering for image denoising. IEEE TIP, 2008. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.264, + 0.905, + 0.305 + ], + "angle": 0, + "content": "[54] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 5" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.305 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2093" + } + ] +] \ No newline at end of file diff --git a/2025/Classic Video Denoising in a Machine Learning World_ Robust, Fast, and Controllable/079f9f38-09f8-4442-836f-ae05472182ae_origin.pdf b/2025/Classic Video Denoising in a Machine Learning World_ Robust, Fast, and Controllable/079f9f38-09f8-4442-836f-ae05472182ae_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..8a14027b50a0cfe9f2e8766f71c0c06ace42268a --- /dev/null +++ b/2025/Classic Video Denoising in a Machine Learning World_ Robust, Fast, and Controllable/079f9f38-09f8-4442-836f-ae05472182ae_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7c06436e4e71414f900064d431823a1b942700a7aca797e5161ee7250e5d3cb3 +size 9671312 diff --git a/2025/Classic Video Denoising in a Machine Learning World_ Robust, Fast, and Controllable/full.md b/2025/Classic Video Denoising in a Machine Learning World_ Robust, Fast, and Controllable/full.md new file mode 100644 index 0000000000000000000000000000000000000000..7fe3a87d9eaa90266640ff24420c567908246f6b --- /dev/null +++ b/2025/Classic Video Denoising in a Machine Learning World_ Robust, Fast, and Controllable/full.md @@ -0,0 +1,325 @@ +# Classic Video Denoising in a Machine Learning World: Robust, Fast, and Controllable + +Xin Jin $^{1*}$ Simon Niklaus $^{2\dagger}$ Zhoutong Zhang $^{3}$ Zhihao Xia $^{3}$ + +Chunle Guo $^{1,4‡}$ Yuting Yang $^{3}$ Jiawen Chen $^{3}$ Chongyi Li $^{1,4}$ + +$^{1}$ VCIP, CS, Nankai University $^{2}$ Adobe Research $^{3}$ Adobe $^{4}$ NKIARI, Shenzhen Futian + +xjin@mail.nankai.edu.cn, sniklaus@adobe.com, + +{guochunle,lichongyi}@nankai.edu.cn + +https://srameo.github.io/projects/levd + +![](images/4efc2cd0f1ea64e14b5777b5312b57420d30070c0a38ba489869a982e3741489.jpg) +Input + +![](images/451109cddf2cdc2c1ce4f5be28fb772a8728cb2f5b25c3a3954fd18a5d576ebb.jpg) +$\mathrm{VRT}^{\dagger}$ [26] +(sophisticated model) + +![](images/ea7e17ec687dedc64ec0ba245e20bf2b17c88cb9c50091cf261f9dfdb1789655.jpg) +Real-ESRGAN [46] +(sophisticated dataset) +Figure 1. In real-world videos, noise can manifest in vastly different ways, causing supervised methods to fail when the input is too far outside the training distribution. Footage provided by Robert Kjettrup, face pixelated after inference. We denote retrained models with a $\dagger$ . + +![](images/0498becf27d39aa6541777cd5cd727475a3c9511cb9ab028960376cefe66e847.jpg) +MF2F [14] +(sophisticated supervision) + +![](images/33790d91080c2092151131a3f11e8d9e74a5f794b46f9a11c48eb3a67c5faed7.jpg) +UDVD [40] +(sophisticated supervision) + +![](images/fbc1097a89e12e7276e5d2b7f0909f438ed2fb90f4ffcb476e97cfd293bc984c.jpg) +Ours - RFCVD + +# Abstract + +Denoising is a crucial step in many video processing pipelines such as in interactive editing, where high quality, speed, and user control are essential. While recent approaches achieve significant improvements in denoising quality by leveraging deep learning, they are prone to unexpected failures due to discrepancies between training data distributions and the wide variety of noise patterns found in real-world videos. These methods also tend to be slow and lack user control. In contrast, traditional denoising methods perform reliably on in-the-wild videos and run relatively quickly on modern hardware. However, they require manually tuning parameters for each input video, which is not only tedious but also requires skill. We bridge the gap between these two paradigms by proposing a differentiable denoising pipeline based on traditional methods. A neural network is then trained to predict the optimal denoising parameters for each specific input, resulting in a robust and efficient approach that also supports user control. + +# 1. Introduction + +Video denoising is a fundamental part of any video editing pipeline, and it is typically applied prior to color grading. After interviewing professional video editors, we have found that they not only want clean results but they also want the denoising to be quick as not to interrupt their workflow, and they want controllability to assert their artistic expression. Specifically, editors often have to decide to leave some noise in the footage in favor of over-smoothing or vice versa, and the answer can differ for each case. For this reason, professional cameras like the Arri Alexa even allow the user to choose between different noise profiles to better support different post-production workflows. + +Unfortunately, recent work on video denoising tends to focus only on the first aspect, the denoising quality, which makes these solutions impractical for the typical video editing workflow. At the same time, we have found that such video denoising approaches are subject to failure cases like the one in Fig. 1 even though they focus on quality. This is not surprising though, considering that noise in real videos can manifest itself in wildly different ways. That is, in addition to the typical degradation that images are subject to, + +![](images/c613ff942e19a76b73e678ff027bca564ac3f5d4528da137eae75312d5b858d9.jpg) +Figure 2. High-level comparison of how related work approaches video denoising (left) and our proposed approach (right). + +videos also leverage temporal compression. In the H.264 codec, for example, there are P-frames and B-frames that copy information from I-frames which means that the noise is often temporally correlated. Furthermore, there are many different video codecs that affect the noise profile in different ways, and even for the same codec, the noise can vary significantly when using different encoders or settings. + +To tackle the complexity of video noise in a deep learning context and as summarized in Fig. 2 (a), we have found that popular approaches either propose a sophisticated way to supervise the model [14, 40], utilize an intricate pipeline to simulate noise [46], leverage an image formation that by design is less prone to overfitting [13], or a combination of these. Yet, the aforementioned example in Fig. 1 demonstrates that this is not always sufficient. + +In contrast, classic denoisers not only work reasonably well but they are also relatively fast, especially on modern hardware. However, they require manually tuning parameters which is not only tedious but also requires a certain amount of skill. We see this as an opportunity. Specifically and as shown in Fig. 2 (b), by implementing a traditional denoising pipeline in a differentiable manner, we can have a neural network learn to predict what the optimal denoising parameters for a given input should be. As shown in Fig. 3, not only does this end up being robust but it is also comparatively fast while providing support for user control. To be clear on our objectives, if the noise profile is known, such as when developing a denoiser for a specific camera model subject to a constrained video encoder, any reasonable deep learning approach would provide better denoising results. However, such an approach would in turn be limited in terms of robustness and controllability. + +In short, our contributions are (1) an approach to leverage a traditional video denoising pipeline in a deep learning setting where we decouple the analysis of the noise from the denoising itself to improve efficiency by avoiding redundant compute, (2) a not only robust and fast but also controllable video denosier (RFCVD), and (3) a training augmentation pipeline based on additive white Gaussian noise together with H.264 transcoding that works surprisingly well. + +![](images/792d63a9c11df280a50ebf42916cc13ae56dabea86898464513706235a06f898.jpg) +Figure 3. Video denoising on the CRVD (sRGB) benchmark [50] using the PSNR across all ISO values with respect to the computational efficiency on an RTX 3090 GPU in FPS (frames per second). We improved some models through retraining, denoted with a $\dagger$ . + +# 2. Related Work + +Traditional Video Denoising. Most traditional video denoising methods can be viewed as a series of linear and non-linear filtering processes, where the filter strength is a function of the estimated noise level. Many such methods are based on block matching [10-12, 31], where similar patches are first aggregated within and across frames before being filtered and merged to produce the denoised result. There are many variations of this process of course, for example by improving the matching [28] or by leveraging bilateral filtering [8, 18, 36, 43] for spatial and temporal denoising [1, 37]. In doing so, image pyramids [3, 16] can be used to improve the computational efficiency and recent iterations of classic multi-frame denoising methods are even quick to run on commodity smartphones [20, 27, 48]. + +Machine Learning in Video Denoising. In recent years, it has become increasingly popular to tackle video denoising with the help of machine learning. This includes methods that, like traditional video denoisers, perform explicit matching [13, 44, 49] as well as those where correspondences are only implicit [24, 32, 42]. And while most of these methods denoise each frame independently, there are also IIR-like approaches where information is passed along such that previously denoised frames can guide future ones [5, 6]. On the model architecture side, the rise of attention in the vision and language community has also led to various attention mechanisms for video restoration [4, 26, 29, 41, 45]. But with machine learning there is a catch, if an input has a noise pattern that has not been seen during training then the denoising is unpredictable. + +Noise Simulation. A common approach to facilitate better generalizability on in-the-wild footage is to employ a more sophisticated degradation pipeline when adding synthetic noise to the ground truth. This includes not only do + +![](images/eea26d388a9b6825f76e7393350f7cac0dd3d0bba7410ef509e6081aa9480a39.jpg) +Noise Profiling on an Anchor Frame Through a Hypernetwork +Figure 4. Video denoising fundamentally first needs to analyze the noise and then remove it. We mimic this in our pipeline by first estimating a noise profile on a random anchor frame (top left in green) before using this profile to denoise the video (bottom in yellow). Specifically, we leverage a hypernetwork configuration where the noise profile $\theta$ is essentially the parameters of the subsequent denoiser. That is, our denoiser is a traditional pipeline consisting of (1) a Wiener filter that performs temporal denoising of neighboring frames that were aligned via optical flow and (2) a bilateral Laplacian pyramid filter for spatial denoising of the temporally merged frames, where a small neural network $\mathcal{P}(\cdot ;\theta)$ predicts spatially-varying parameters for the Wiener merger and the bilateral filters. This separation of concerns improves the overall efficiency since it avoids having to redundantly analyze the noise over and over again. + +ing one but two rounds of degradations [46], shuffling the degradations [51], or integrating more advanced operators such as tone mapping [52]. Although these approaches improve the ability to denoise real-world videos, they can only reduce but not entirely resolve the generalizability gap. + +Self-Supervised Learning. Another strategy to improve the generalizability in video denoising is to adopt self-supervision on real-world videos without requiring access to a ground truth. For example, this can be achieved by leveraging the nature of regression losses [25] or by building on the blind spot idea [14, 17, 23, 40]. While these approaches make it possible to train on noisy in-the-wild videos, it is not necessarily feasible to collect a dataset that contains all possible types of noise that occur in the real world. Therefore, self-supervision often also relies on test-time adaptation which comes with its own challenges. + +# 3. Method + +Video denoising fundamentally first needs to analyze the noise and then remove it. In contrast, recent approaches perform both tasks at once by utilizing a single neural network that is given the noisy input frames and is tasked with producing the clean output. The noise profile in a video does typically not change over time though, yet these ap + +proaches independently denoise one frame after another so they have to analyze the noise over and over again. + +To avoid this redundancy, we follow the natural separation of analyzing the noise and then removing it. As shown in Fig. 4, we first estimate the noise profile on an anchor frame and then denoise the video using the estimated profile. Specifically, we leverage a hypernetwork setup where the noise profile $\theta$ provides the parameters of the subsequent denoiser. That is, our denoiser is a traditional pipeline consisting of (1) a Wiener filter that performs temporal denoising of neighboring frames that were aligned via optical flow and (2) a bilateral Laplacian pyramid filter for spatial denoising of the temporally merged frames, where a small neural network $\mathcal{P}(\cdot ;\theta)$ predicts spatially-varying parameters for the Wiener merger and the bilateral filters. + +We will subsequently discuss these parts in turn before providing more details on user control and the training. + +# 3.1. Noise Profiling + +Since noise is not necessarily spatially uniform but often dependent on the signal, we argue that a noise profile should not be just a fixed set of values for the entire input but rather be a descriptor that makes it possible to derive spatially-varying denoising parameters. We see this as an ideal use case for hypernetworks [19], where a noise profiler esti + +![](images/03eb94de23ae0d8a136bd6e105428eeb4057be9a3566073988ccd7b21a83c1f1.jpg) +Input + +![](images/ad9e2eb9e981b29dfd671fe69868eeb2f3ad0d935aec408f3ae3e1069df1f2f1.jpg) +reduced temporal only $(0.2\cdot \sigma^2,0\cdot \sigma_d,0\cdot \sigma_r)$ + +![](images/ee0a4435d21cba2ce07442037b033b55e7a52fff9663f7627660cf0d649de330.jpg) +temporal only $(1\cdot \sigma^2,0\cdot \sigma_d,0\cdot \sigma_r)$ +Figure 5. By scaling the denoising parameters estimated by $\mathcal{P}(\cdot ;\theta)$ , our framework makes it easy to control the amount of temporal denoising (through $\sigma^2$ ) as well as the amount of spatial denoising (through $\sigma_d$ and $\sigma_r$ ). We scale the denoising parameters equivalently for the luma and chroma channels in this sample, but they can be scaled independently for more control. Please see the supplementary for a video demo that demonstrates how users can control the denoising in an interactive manner. Footage provided by Amit Zinman. + +![](images/712c0b43eaf5055d26349b9a6f8881bb85100a2ef1a47be6b8ed107beb14b91a.jpg) +reduced spatial only $(0\cdot \sigma^{2},0.2\cdot \sigma_{d},0.2\cdot \sigma_{r})$ + +![](images/e5d5f37b5cf91cd079f40ccd6ae4b57a05565d5d5897ccf4de24ae648121e492.jpg) +spatial only $(0\cdot \sigma^{2},1\cdot \sigma_{d},1\cdot \sigma_{r})$ + +![](images/182ed8d1157ea58a335c7c3636b019f63f2f9e999ff20ffe761ce505120f6f8a.jpg) +default settings $(1\cdot \sigma^{2},1\cdot \sigma_{d},1\cdot \sigma_{r})$ + +mates the weights $\theta$ of a small neural network $\mathcal{P}(\cdot ;\theta)$ that predicts spatially-varying denoising parameters. However, instead of directly predicting $\theta$ , we stabilize the training through NPA [35] where $\theta = \theta^0 +\Delta \theta$ with $\theta^0$ being learnable parameters and $\Delta \theta$ being the hypernet prediction. + +Intuitively, analyzing the noise profile of a given image is a mixture of low-level image processing and high-level semantics. Specifically, one first needs to understand what an image depicts before being able to examine the details to discern between unintended noise and actual texture. For this reason, we leverage a pre-trained backbone $\mathcal{B}$ in the form of a ConvNext [30] with a random MLP head to serve as the hypernet. We have found this backbone to be crucial for the quality of the predicted noise profile. + +Lastly, which anchor frame should we choose to analyze the noise of a given video? We would like this choice not to matter, such that the denoising result is independent of having to choose a "good" anchor. To achieve this, we always pick the first frame of a video as the anchor frame because it is as good as any, and then utilize a consistency loss + +$$ +\mathcal {L} _ {c s t s y} = \left\| \Delta \theta - \Delta \theta_ {i} \right\| _ {2}, \tag {1} +$$ + +where we encourage the hypernet prediction of our anchor frame $\Delta \theta$ and any random neighbor $\Delta \theta_{i}$ to be equal. + +# 3.2. Denoising the Video + +Once we have the noise profile $\theta$ , we process the input video one output frame at a tim where we obtain spatially-varying denoising parameters through a small neural network $\mathcal{P}(\cdot ;\theta)$ with three convolutional layers. We use these parameters to do temporal and then spatial denoising. + +For the temporal denoising, to denoise a given target frame, we first align the two preceding and the two subsequent frames. Specifically, we estimate the optical flow between the target frame and the neighbors using an off-the-shelf SpyNet [38] at a quarter of the target resolution for improved computational efficiency as well as robustness to noise, and warp the neighbors to the target frame using the + +estimated optical flow. We then merge the aligned frames using a Wiener filter as in Hasinoff et al. [20] but using a tiles size of $8 \times 8$ and the noise maps from $\mathcal{P}(\cdot; \theta)$ instead of approximating the noise as the RMS in each tile. As is typical and to facilitate more control, we perform Wiener filtering on the luma and the chroma channels independently. + +We then apply spatial denoising to the temporally denoised target frame through a bilateral Laplacian pyramid filter with three levels. This borrows many ideas from multiresolution bilateral filtering [53] but using a Laplacian instead of a Gaussian pyramid and using bilateral filtering throughout instead of wavelet thresholding for the first pyramid levels. More specifically, we have $\mathcal{P}(\cdot ;\theta)$ estimate noise map pyramids and then use these as $\sigma_{s}$ and $\sigma_r$ for the bilateral filtering. Furthermore, just like with the Wiener filter, we perform the denoising separately on the luma and the chroma channels to facilitate user control. + +# 3.3. User Control + +Through interviews with creative professionals, we have found that they want to assert their artistic expression when denoising a clip. That is, they often have to decide to leave some noise in the footage in favor of over-smoothing and vice versa. The key to facilitating user control in our method is the spatially-varying parameters from $\mathcal{P}(\cdot ;\theta)$ that guide the temporal and spatial denoising. Specifically, we have two of these maps for $\sigma^2$ in the Wiener filter (one map for chroma and one for luma) and in the bilateral Laplacian pyramid filter we have two map pyramids for $\sigma_{d}$ as well as two map pyramids for $\sigma_r$ (one map pyramid for chroma and one for luma). This makes for a total of six sets of parameters where we can facilitate user control. + +How do we attenuate these parameters though? As demonstrated in Fig. 5, we have found that simply scaling the spatially-varying denoising parameters uniformly works reasonably well, so we can just provide six tunable knobs to end users. And since our denoising framework runs in real time, changing the knobs provides instantaneous feedback which makes it easy to attain the desired result. + +# 3.4. Augmentation Pipeline + +To train our model, we need pairs of clean and noisy frame sequences just like any other denoiser that leverages machine learning without self-supervision. Specifically, we adopt the typical paradigm of taking a set of "noise-free" videos that serve as the ground truth, and augmenting them with various degradations to obtain the corresponding noisy inputs. For this to work well, we have found it to be crucial to include temporal compression in this degradation pipeline since it makes the noise temporally correlated. + +But let's start with noise-free videos, which we obtain from the REDS dataset [33]. It provides 240 videos and was shot at a relatively high 120 frames per second, allowing us to augment the frame rate via sub-sampling. We extract random frame sequences from these videos and first degrade them through additive white Gaussian noise with a variance sampled from $\mathcal{U}(1,50)$ . We then transcode them using an H.264 codec through libx264 with a CRF sampled from $\mathcal{U}(18,30)$ to make the noise temporally correlated. + +We have found this data pipeline to work surprisingly well. So well in fact that we retrained the models that we compare to in the evaluation since we were able to get much better results on the CRVD benchmark [50] with our retrained versions than the original checkpoints. However, we did not retrain Real-ESRGAN [46] since it brings its own much more complex augmentation pipeline, and we also did not retrain MF2F [14] and UDVD [40] since these leverage sophisticated ways of supervising the models which are difficult for us to replicate and properly do justice. + +# 3.5. Implementation Details + +In terms of architecture details, we utilize the base size of a ConvNext [30] for the backbone $\mathcal{B}$ in the hypernetwork. To translate the output from this backbone to the residual noise profile $\Delta \theta$ , we utilize an MLP with five layers and PReLU activations [21] in between each layer. And for the neural network $\mathcal{P}(\cdot ;\theta)$ that predicts the spatially-varying denoising parameters, we leverage three convolution layers with a stride of two and PReLU activations in between. We additionally constrain the output of this network with a softmax [15] to facilitate non-negative denoising parameters. Furthermore, we have found it beneficial to not only provide the (aligned) frames to $\mathcal{P}(\cdot ;\theta)$ but also their gradients, approximated by a Sobel filter, as well as masks that indicate whether or not a pixel in an aligned frame is valid (it may have been warped from outside the frame) [2]. + +In our reconstruction loss, we minimize the difference between the temporally denoised image $I^{\mathrm{T}}$ and the ground truth $I^{\mathrm{GT}}$ as well as the difference between all levels of the spatial denoising pyramid $I^{\mathrm{ST}}$ and $I^{\mathrm{GT}}$ as + +$$ +\mathcal {L} _ {r e c} = \left\| I ^ {\mathrm {T}} - I ^ {\mathrm {G T}} \right\| _ {2} + \sum_ {l = 1} ^ {3} \left\| I _ {l} ^ {\mathrm {S T}} - I _ {l} ^ {\mathrm {G T}} \right\| _ {2}, \tag {2} +$$ + +
PSNRdelta (higher PSNR is better)SSIMdelta (higher SSIM is better)LPIPSdelta (lower LPIPS is better)
NAFNet [9]30.25-0.747-0.358-
NAFNet†[9]35.32+ 5.070.937+ 0.1900.089- 0.269
FastDVDNet [42]27.89-0.565-0.502-
FastDVDNet†[42]34.63+ 6.750.914+ 0.3490.124- 0.378
TOFlow [49]25.96-0.673-0.251-
TOFlow†[49]34.08+ 8.110.903+ 0.2310.147- 0.104
BasicVSR++ [6]31.98-0.769-0.326-
BasicVSR++†[6]34.34+ 2.350.873+ 0.1040.167- 0.158
VRT [26]31.99-0.784-0.296-
VRT†[26]33.86+ 1.860.848+ 0.0640.192- 0.104
+ +Table 1. Denoising results on the CRVD (sRGB) benchmark for various methods that we compare to, both for their original version as well as our retrained one which we denote with a $\dagger$ . + +where $l$ is the $l$ -th level in spatial denoising pyramid. Recall that we also have a consistency loss that we outlined in Eq. (1) in the section about noise profiling. + +We train our model using Adam [22] with an initial learning rate of $2 \times 10^{-4}$ that decays to $1 \times 10^{-7}$ using a cosine annealing schedule. In total, we train the model for 400 thousand iterations with a batch size of 24 consisting of patches with $512 \times 512$ pixels. Since our pipeline is computationally lightweight, this takes less than two days. + +# 4. Experiments + +We quantitatively evaluate our approach on two types of videos, real and synthetic. For the former, we leverage the CRVD benchmark [50] consisting of RAW videos which we convert to sRGB using the provided deep ISP [39]. For the latter, we use the validation samples from the REDS dataset [33] which we augmented in two ways that are notably different from the AWGN subject to H.264 transcoding during training. Specifically, one way is using film grain noise [34] with AV1 transcoding, and the other way is using spatially correlated Gaussian noise with H.265 transcoding. As for evaluation metrics, we follow the typical paradigm of reporting PSNR, SSIM [47], and LPIPS [54]. Since we argue that our method is computationally efficient, we also report the frames per second (FPS) which we measured on an RTX 3090 GPU with proper CUDA synchronization. + +In terms of data for our qualitative evaluation, we reached out to industry professionals who sent us footage where they had trouble removing the noise. Since these professionals did not have a release form for all the actors, we sometimes had to anonymize faces post inference. + +Lastly, we compare our approach to various kinds of denoisers. This includes image denoisers [7, 9], video denoisers [42, 49], recurrent methods [6], transformers [26], models with a sophisticated augmentation pipeline [46], as well as methods that leverage self-supervision [14, 40]. Since we have found that our augmentation pipeline works sur + +
ISO 1600ISO 3200ISO 6400ISO 12800ISO 25600OverallSpeed
PSNRrank (higher PSNR is better)PSNRrank (higher PSNR is better)PSNRrank (higher PSNR is better)PSNRrank (higher PSNR is better)PSNRrank (higher PSNR is better)PSNRrank (higher PSNR is better)FPSrank (higher FPS is better)
\( SID^† \)[7]38.85\( 7^{th} \) of 1037.68\( 7^{th} \) of 1035.82\( 4^{th} \) of 1033.51\( 4^{th} \) of 1029.18\( 3^{rd} \) of 1035.01\( 4^{th} \) of 106.95\( 3^{rd} \) of 10
\( NAFNet^† \)[9]39.48\( 3^{rd} \) of 1038.12\( 5^{th} \) of 1035.94\( 3^{rd} \) of 1033.53\( 3^{rd} \) of 1029.55\( 2^{nd} \) of 1035.32\( 2^{nd} \) of 101.69\( 7^{th} \) of 10
\( FastDVDNet^† \)[42]39.16\( 5^{th} \) of 1037.92\( 6^{th} \) of 1035.60\( 5^{th} \) of 1032.56\( 5^{th} \) of 1027.93\( 6^{th} \) of 1034.63\( 5^{th} \) of 105.72\( 4^{th} \) of 10
\( TOFlow^† \)[49]38.25\( 8^{th} \) of 1036.97\( 8^{th} \) of 1034.90\( 7^{th} \) of 1032.21\( 6^{th} \) of 1028.07\( 5^{th} \) of 1034.08\( 7^{th} \) of 102.84\( 6^{th} \) of 10
\( BasicVSR++^† \)[6]39.40\( 4^{th} \) of 1038.24\( 2^{nd} \) of 1035.55\( 6^{th} \) of 1031.72\( 7^{th} \) of 1026.78\( 9^{th} \) of 1034.34\( 6^{th} \) of 107.41\( 2^{nd} \) of 10
\( VRT^† \)[26]39.55\( 2^{nd} \) of 1038.12\( 4^{th} \) of 1034.82\( 8^{th} \) of 1030.77\( 8^{th} \) of 1026.01\( 10^{th} \) of 1033.86\( 8^{th} \) of 100.05\( 10^{th} \) of 10
Real-ESRGAN [46]29.98\( 10^{th} \) of 1028.27\( 10^{th} \) of 1028.04\( 10^{th} \) of 1027.52\( 10^{th} \) of 1027.63\( 8^{th} \) of 1028.29\( 10^{th} \) of 100.24\( 8^{th} \) of 10
UDVD [40]31.15\( 9^{th} \) of 1030.72\( 9^{th} \) of 1030.23\( 9^{th} \) of 1029.10\( 9^{th} \) of 1027.63\( 7^{th} \) of 1029.77\( 9^{th} \) of 100.16\( 9^{th} \) of 10
MF2F [14]39.09\( 6^{th} \) of 1038.20\( 3^{rd} \) of 10\( \underline{36.36} \)\( 1^{st} \) of 1033.57\( 2^{nd} \) of 1029.04\( 4^{th} \) of 1035.25\( 3^{rd} \) of 104.62\( 5^{th} \) of 10
Ours - RFCVD40.35\( 1^{st} \) of 10\( \underline{38.60} \)\( 1^{st} \) of 1036.28\( 2^{nd} \) of 1033.86\( 1^{st} \) of 1031.12\( 1^{st} \) of 1036.04\( 1^{st} \) of 1031.66\( 1^{st} \) of 10
+ +Table 2. Video denoising results on the CRVD (sRGB) benchmark. Not only does our approach perform best overall, it is also four times faster than the second-fastest place. Please kindly see the supplementary for SSIM and LPIPS where our approach ranks first as well. + +
film grain noise w/ AV1 compressionspatially correlated noise w/ H.265 compressionSpeed
PSNR (higher PSNR is better)rank SSIM (higher SSIM is better)rank (lower LPIPS is better)LPIPS rank (lower LPIPS is better)rank (higher PSNR is better)rank (higher PSNR is better)SSIM (higher SSIM is better)rank (lower LPIPS is better)LPIPS rank (lower LPIPS is better)rank (higher FPS is better)FPS rank (higher FPS is better)
\( SID^† \)[7]27.05 5thof 70.664 5thof 70.293 4thof 728.44 3rdof 70.771 4thof 70.282 5thof 715.00 3rdof 7
\( NAFNet^† \)[9]27.16 4thof 70.672 4thof 70.294 5thof 728.40 4thof 70.764 6thof 70.257 3rdof 73.812 6thof 7
\( FastDVDNet^† \)[42]27.39 3rdof 70.686 3rdof 70.272 3rdof 727.92 6thof 70.769 5thof 70.296 6thof 712.09 4thof 7
\( TOFlow^† \)[49]28.15 2ndof 70.750 2ndof 70.220 1stof 728.39 5thof 70.788 3rdof 70.258 4thof 75.665 5thof 7
\( BasicVSR++^† \)[6]26.90 6thof 70.651 6thof 70.313 6thof 727.48 7thof 70.728 7thof 70.358 7thof 715.32 2ndof 7
\( VRT^† \)[26]26.55 7thof 70.629 7thof 70.331 7thof 728.78 2ndof 70.803 2ndof 70.206 1stof 70.105 7thof 7
Ours - RFCVD28.59 1stof 70.774 1stof 70.247 2ndof 728.93 1stof 70.808 1stof 70.239 2ndof 769.73 1stof 7
+ +Table 3. Testing the generalizability when training on AWGN with H.264 transcoding but testing with two different degradation schemes. Since our approach is based on classic methods, the denoising itself only has a few parameters which naturally reduces domain gaps. + +prisingly well, we have retrained all methods that neither leverage a sophisticated augmentation pipeline already nor utilize self-supervision. This makes for a more fair comparison since, as shown in Tab. 1, we have improved their performance on the CRVD benchmark with our retrained version. Throughout this paper, we denote all retrained models with a $\dagger$ to prevent any potential confusions. + +# 4.1. Quantitative Comparison + +Please see Tab. 2 for a summary of the quantitative evaluation on the CRVD benchmark. In short, not only does our approach perform best overall, it is also four times faster than the second-fastest method. We attribute these favorable results to our utilization of classic denoising techniques which greatly helps with the ability to generalize to unseen noise patterns. Specifically, the denoising itself is driven by only a few parameters which naturally reduces overfitting and helps with bridging domain gaps. To test this hypothesis, we conducted an experiment with synthetic data where all models are trained on AWGN with H.264 transcoding but tested on two different degradation pipelines. As shown in Tab. 3, our approach performs favorably in this evaluation as well which supports our initial hypothesis. + +# 4.2. Qualitative Comparison + +To test video denoising in the real world, we reached out to various creative professionals who kindly shared footage where they had trouble removing the noise. We show a representative excerpt of the denoising results on this footage in Fig. 6 but kindly refer to the supplementary which includes video results of all methods. Like with the quantitative evaluation, we have found that our approach performs favorably, and once again we hypothesize that this is due to the underlying utilization of classic denoising techniques. + +# 4.3. Ablative Experiments + +One of the main things we wondered ourselves was the performance with respect to the choice of anchor frame. After all, the noise profile of I-frames can be quite different from P- and B-frames. Nevertheless as shown in Tab. 4 (top), we have always used the first frame as the anchor throughout the experiments section and would have drawn the same conclusions with the middle or the last frame. And as shown in Tab. 4 (bottom), we have also found that $\mathcal{L}_{cstsy}$ helps with providing a slightly more consistent experience. + +A key part of our integration of classic denoising tech + +![](images/1df6be4b5a87c7f1cc7e60530297622ad88009c6ee6832c78a28aef66251d7fe.jpg) +Figure 6. Video denoising results from exemplary still frames. Due to space constraints, we only share the results from a representative subset of the methods that we compare to. Please kindly refer to the supplementary to find video results of all methods. We are grateful that Amit Zinman (top row), Robert Kjettrup (middle row), and an anonymous artist (last row) were willing to provide test footage. + +niques into a machine learning pipeline is that we account for spatially-varying noise. To assess the importance of this design decision, we trained an ablation that only predicts a set of scalar denoising parameters. As expected and as shown in Tab. 6, this "w/o spat. varying" experiment exhibits a significant drop in quality. As shown in Fig. 7, this ablation produces much more blurry denoising results since it cannot easily account for signal-dependent noise. + +To best support $\mathcal{P}(\cdot ;\theta)$ in predicting the denoising parameters, we not only provide it access to the (merged) input frames but also their gradients as well as a mask indicating whether or not a pixel in an aligned frame is valid. To test the effectiveness of this design, we trained two ablations where we refrained from providing one ("w/o img. gradients") or the other ("w/o align. mask") and found that, as shown in Tab. 6, this context is indeed beneficial. + +We argue that analyzing the noise profile of a given image is a mixture of both low-level image processing and high-level semantics, because one first needs an understanding of what an image depicts before being able to go into the low-level details to discern between unintended noise and actual texture. For this reason, we leverage a classification backbone $\mathcal{B}$ in the hypernetwork that is predicting the noise profile $\theta$ , and we trained an ablation without this backbone to evaluate this hypothesis. As intuitively expected and as shown in Tab. 6, this "w/o backbone $\mathcal{B}$ " experiment indeed exhibits a significant drop in quality. This drop is also visually exemplified in Fig. 7, where the lack of this backbone leads to the noise not fully being removed. + +Lastly, we not only fine-tune the aforementioned backbone $\mathcal{B}$ but also stabilize the training of the hypernetwork through NPA [35]. To analyze the effect of these measures, + +
PSNRdelta (higher PSNR is better)SSIMdelta (higher SSIM is better)LPIPSdelta (lower LPIPS is better)
Avg. w/ Lcstsy35.95-0.9477-0.0757-
first frame ← Ours36.04+ 0.090.9472- 0.00050.0763+ 0.0006
middle frame35.92- 0.030.9480+ 0.00030.0753- 0.0003
last frame35.90- 0.060.9480+ 0.00030.0754- 0.0003
Avg. w/o Lcstsy35.86-0.9460-0.0754-
first frame35.80- 0.050.9454- 0.00050.0750- 0.0004
middle frame36.02+ 0.160.9473+ 0.00140.0753- 0.0001
last frame35.75- 0.110.9451- 0.00080.0759+ 0.0005
+ +Table 4. Denoising results on the CRVD benchmark with respect to the anchor frame choice. We find that our approach is relatively robust to this choice (top), partly thanks to $\mathcal{L}_{cstsy}$ (bottom). + +
PSNRdelta (higher PSNR is better)SSIMdelta (higher SSIM is better)LPIPSdelta (lower LPIPS is better)
Ours36.04-0.9472-0.0763-
fine-tuned on GT36.75+ 0.700.9494+ 0.00220.0591- 0.0172
+ +we trained an ablation for the former ("w/o fine-tuning $\mathcal{B}$ " as well as the latter ("w/o NPA") and found that, as shown in Tab. 6, both contribute to the denoising quality. + +# 4.4. Limitations + +While our noise profile $\theta$ with the subsequent denoising parameter prediction from $\mathcal{P}(\cdot ;\theta)$ works reasonably well as demonstrated throughout the experiments section, it is far from perfect. Specifically and as shown in Tab. 5, we fine-tuned the predicted denoising parameters on the ground truth and found that this hypothetical oracle is able to improve our results by a rather significant margin. Nevertheless, we leave bridging this gap to future research. + +In the case where the noise profile is entirely known a priori, such as when developing a denoiser for a specific camera model subject to a constrained encoder, any reasonable deep learning method would provide better results than our approach. And with a baked-in noise profile it would perhaps even be faster. However, such a method would in turn be limited in terms of robustness and control. + +Lastly, and on a more practical note, deploying hypernetworks in end-user applications is not well supported by the common inference ecosystems yet. Specifically, popular libraries like CoreML, WinML, or OpenVINO expect the model weights to be fixed which is obviously not the case for hypernetworks. As such, while our approach is easy to implement within frameworks like PyTorch, it is not straightforward to deploy in production environments. + +![](images/d617da301db3f1bdba758af444dbdfcab7b4df741dc9fd6d3e24380b2057bc9f.jpg) + +![](images/a6147f8108f4b894c00419764e36d1f746cdb23aecdc4fc353bba3b5aaaf48ff.jpg) + +![](images/180577504c3d21230b9c91032cf119b2c6a956c0fec859267a84e0865cb2eac1.jpg) +Input + +![](images/eaa422a565e2b480918ec025a8b466b88925dc1d38c01191b6cec38c30e3cf6e.jpg) + +![](images/5c34871bb5c19595625021fa1f101a9261a4c10068a1a26b6b041902c7c096c2.jpg) + +![](images/22a767d79f8d800880a9161affb4f0ab1bdd976f843f4c3315c2daed2f8b1ef4.jpg) +w/o spat. varying + +![](images/96bc151ab373809c214ffbbff0e65f4ac71871d1fe2f61295b4e34343477b2a1.jpg) + +![](images/59f2eaae540e41a729af029582b29ef35a405c799151bd9f7ea48c7ef3299880.jpg) + +![](images/95a7282e092419b8effe5e0f9ff2b0152145fd6b4fddeccbbc5ce563b8d9761b.jpg) +w/o backbone $\mathcal{B}$ + +![](images/b101c90257dcf2a870cfee103439369810981e8bdc94895ed837dc4763e388cb.jpg) + +![](images/fec91e0d23dbf73ac3a7bc97da582be1d81334a9498716733cceb39f243ac12f.jpg) + +![](images/17345923f8d9819f678594c7a1db336f4bfd0602988913cf9edab0f8723f18c3.jpg) +Ours +Figure 7. Visual comparison of the two most significant ablations on a representative sample from the CRVD dataset. + +Table 5. Fine-tuning our estimated denoising parameters on the ground truth of the CRVD benchmark indicates that our estimates are good but not perfect if we had access to an oracle. + +
PSNRdelta (higher PSNR is better)SSIMdelta (higher SSIM is better)LPIPSdelta (lower LPIPS is better)
Ours36.04-0.9472-0.0763-
w/o spat. varying33.62-2.420.9359-0.01120.1184+0.0421
w/o img. gradients35.61-0.440.9469-0.00030.0756-0.0007
w/o align. mask35.26-0.780.9468-0.00030.0740-0.0023
w/o backbone B34.38-1.670.9059-0.04120.1444+0.0680
w/o fine-tuning B35.71-0.340.9464-0.00080.0732-0.0032
w/o NPA [35]35.83-0.210.9453-0.00180.0814+0.0051
+ +Table 6. Various ablative experiments, please see Section 4.3 for more context and details of each individual ablation. + +# 5. Conclusion + +We show that traditional denoising techniques are still very much relevant in the modern machine learning world. Not only have we found them to be robust and fast, but also controllable which is an important property in many video denoising applications where creative professionals want to assert their artistic expression. To bridge the gap between classic denoising and modern approaches, we train a model to estimate the various parameters of a traditional denoising approach for a given input video, which is otherwise not only tedious but also requires a certain amount of skill. + +Furthermore, in an effort to make best use of the computational efficiency of classic denoising techniques, our method separates the analysis of the noise from the denoising in order to avoid having to estimate the noise profile over and over again as is typical for machine learning approaches to video denoising. Lastly, we have found a simple yet effective degradation pipeline for training video denoiseers based on AWGN subject to H.264 video transcoding. This pipeline ensures that the noise is temporally correlated just like it often is the case in the real world, and we have found it not only to work well for our approach but we have also been able to retrain and improve existing models. + +Acknowledgement. This work was supported in part by the National Natural Science Foundation of China (62306153, 62225604), the Fundamental Research Funds for the Central Universities (Nankai University, 070-63243143), the Natural Science Foundation of Tianjin, China (24JCJQJC00020), and Shenzhen Science and Technology Program (JCYJ20240813114237048). The computational devices is partly supported by the Supercomputing Center of Nankai University (NKSC). + +# References + +[1] Eric P Bennett and Leonard McMillan. Video enhancement using per-pixel virtual exposures. In ACM SIGGRAPH 2005 Papers, 2005. 2 +[2] Goutam Bhat, Michael Gharbi, Jiawen Chen, Luc Van Gool, and Zhihao Xia. Self-supervised burst super-resolution. In ICCV, pages 10571-10580. IEEE, 2023. 5 +[3] Peter J Burt and Edward H Adelson. The laplacian pyramid as a compact image code. In Readings in computer vision, 1987. 2 +[4] Jiezhang Cao, Yawei Li, Kai Zhang, and Luc Van Gool. Video super-resolution transformer. arXiv:2106.06847, 2021. 2 +[5] Kelvin CK Chan, Xintao Wang, Ke Yu, Chao Dong, and Chen Change Loy. Basicvsr: The search for essential components in video super-resolution and beyond. In CVPR, 2021. 2 +[6] Kelvin CK Chan, Shangchen Zhou, Xiangyu Xu, and Chen Change Loy. Basicvsr++: Improving video superresolution with enhanced propagation and alignment. In CVPR, 2022. 2, 5, 6 +[7] Chen Chen, Qifeng Chen, Jia Xu, and Vladlen Koltun. Learning to see in the dark. In CVPR, 2018. 2, 5, 6 +[8] Jiawen Chen, Sylvain Paris, and Frédo Durand. Real-time edge-aware image processing with the bilateral grid. ACM TOG, 2007. 2 +[9] Liangyu Chen, Xiaojie Chu, Xiangyu Zhang, and Jian Sun. Simple baselines for image restoration. In ECCV, 2022. 2, 5, 6 +[10] Kostadin Dabov, Alessandro Foi, Vladimir Katkovnik, and Karen Egiazarian. Image denoising with block-matching and 3d filtering. In Image Processing, 2006. 2 +[11] Kostadin Dabov, Alessandro Foi, and Karen Egiazarian. Video denoising by sparse 3d transform domain collaborative filtering. EURASIP, 2007. +[12] Kostadin Dabov, Alessandro Foi, Vladimir Katkovnik, and Karen Egiazarian. Image denoising by sparse 3-d transform-domain collaborative filtering. IEEE TIP, 2007. 2 +[13] Axel Davy, Thibaud Ehret, Jean-Michel Morel, Pablo Arias, and Gabriele Facciolo. A non-local cnn for video denoising. In ICIP, 2019. 2 +[14] Valery Dewil, Jérémy Anger, Axel Davy, Thibaud Ehret, Gabriele Facciolo, and Pablo Arias. Self-supervised training for blind multi-frame video denoising. In WACV, 2021. 1, 2, 3, 5, 6, 7 + +[15] Charles Dugas, Yoshua Bengio, François Bélisle, Claude Nadeau, and René Garcia. Incorporating second-order functional knowledge for better option pricing. In NIPS, 2000. 5 +[16] Jana Ehmann, Lun-Cheng Chu, Sung-Fang Tsai, and Chia-Kai Liang. Real-time video denoising on mobile phones. In ICIP, 2018. 2 +[17] Thibaud Ehret, Axel Davy, Jean-Michel Morel, Gabriele Facciolo, and Pablo Arias. Model-blind video denoising via frame-to-frame training. In CVPR, 2019. 3 +[18] Ruturaj G Gavaskar and Kunal N Chaudhury. Fast adaptive bilateral filtering. IEEE TIP, 2018. 2 +[19] David Ha, Andrew M. Dai, and Quoc V. Le. Hypernetworks. arXiv:1609.09106, 2016. 3 +[20] Samuel W Hasinoff, Dillon Sharlet, Ryan Geiss, Andrew Adams, Jonathan T Barron, Florian Kainz, Jiawen Chen, and Marc Levoy. Burst photography for high dynamic range and low-light imaging on mobile cameras. ACM TOG, 2016. 2, 4 +[21] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Delving deep into rectifiers: Surpassing human-level performance on imagenet classification. In ICCV, 2015. 5 +[22] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR, 2015. 5 +[23] Samuli Laine, Tero Karras, Jaakko Lehtinen, and Timo Aila. High-quality self-supervised deep image denoising. In NeurIPS, 2019. 3 +[24] Seunghwan Lee, Donghyeon Cho, Jiwon Kim, and Tae Hyun Kim. Restore from restored: Video restoration with pseudo clean video. In CVPR, 2021. 2 +[25] Jaakko Lehtinen, Jacob Munkberg, Jon Hasselgren, Samuli Laine, Tero Karras, Miika Aittala, and Timo Aila. Noise2noise: Learning image restoration without clean data. In ICML, 2018. 3 +[26] Jingyun Liang, Jiezhang Cao, Yuchen Fan, Kai Zhang, Rakesh Ranjan, Yawei Li, Radu Timofte, and Luc Van Gool. Vrt: A video restoration transformer. IEEE TIP, 2024. 1, 2, 5, 6, 7 +[27] Orly Liba, Kiran Murthy, Yun-Ta Tsai, Tim Brooks, Tianfan Xue, Nikhil Karnad, Qiurui He, Jonathan T Barron, Dillon Sharlet, Ryan Geiss, et al. Handheld mobile photography in very low light. ACM TOG, 2019. 2 +[28] Ce Liu and William T Freeman. A high-quality video denoising algorithm based on reliable motion estimation. In ECCV, 2010. 2 +[29] Ding Liu, Zhaowen Wang, Yuchen Fan, Xianming Liu, Zhangyang Wang, Shiyu Chang, and Thomas Huang. Robust video super-resolution with learned temporal dynamics. In ICCV, 2017. 2 +[30] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In CVPR, 2022. 4, 5 +[31] Matteo Maggioni, Vladimir Katkovnik, Karen Egiazarian, and Alessandro Foi. Nonlocal transform-domain filter for volumetric data denoising and reconstruction. IEEE TIP, 2012. 2 + +[32] Matteo Maggioni, Yibin Huang, Cheng Li, Shuai Xiao, Zhongqian Fu, and Fenglong Song. Efficient multi-stage video denoising with recurrent spatio-temporal fusion. In CVPR, 2021. 2 +[33] Seungjun Nah, Sungyong Baik, Seokil Hong, Gyeongsik Moon, Sanghyun Son, Radu Timofte, and Kyoung Mu Lee. NTIRE 2019 challenge on video deblurring and superresolution: Dataset and study. In CVPRW, 2019. 5 +[34] Alasdair Newson, Noura Faraj, Bruno Galerne, and Julie Delon. Realistic film grain rendering. IPOL, 2017. 5 +[35] Jose Javier Gonzalez Ortiz, John V. Guttag, and Adrian V. Dalca. Non-proportional parametrizations for stable hypernetwork learning. arXiv:2304.07645, 2023. 3, 4, 7, 8 +[36] Sylvain Paris, Pierre Kornprobst, Jack Tumblin, Frédo Durand, et al. Bilateral filtering: Theory and applications. Foundations and Trends® in Computer Graphics and Vision, 2009. 2 +[37] Andrea Petreto, Thomas Romera, Florian Lemaitre, Ian Masliah, Boris Gaillard, Manuel Bouyer, Quentin L Meunier, and Lionel Lacassagne. A new real-time embedded video denoising algorithm. In DASIP, 2019. 2 +[38] Anurag Ranjan and Michael J. Black. Optical flow estimation using a spatial pyramid network. In CVPR, 2017. 4 +[39] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In MICCAI, 2015. 5 +[40] Dev Yashpal Sheth, Sreyas Mohan, Joshua L Vincent, Ramon Manzorro, Peter A Crozier, Mitesh M Khapra, Eero P Simoncelli, and Carlos Fernandez-Granda. Unsupervised deep video denoising. In ICCV, 2021. 1, 2, 3, 5, 6, 7 +[41] Maitreya Suin and AN Rajagopalan. Gated spatio-temporal attention-guided video deblurring. In CVPR, 2021. 2 +[42] Matias Tassano, Julie Delon, and Thomas Veit. Fastdvnet: Towards real-time deep video denoising without flow estimation. In CVPR, 2020. 2, 5, 6 +[43] Carlo Tomasi and Roberto Manduchi. Bilateral filtering for gray and color images. In ICCV, 1998. 2 +[44] Gregory Vaksman, Michael Elad, and Peyman Milanfar. Patch craft: Video denoising by deep modeling and patch matching. In ICCV, 2021. 2 +[45] Xintao Wang, Kelvin CK Chan, Ke Yu, Chao Dong, and Chen Change Loy. Edvr: Video restoration with enhanced deformable convolutional networks. In CVPRW, 2019. 2 +[46] Xintao Wang, Liangbin Xie, Chao Dong, and Ying Shan. Real-esrgan: Training real-world blind super-resolution with pure synthetic data. In ICCV Workshop, 2021. 1, 2, 3, 5, 6, 7 +[47] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE TIP, 2004. 5 +[48] Bartlomiej Wronski, Ignacio Garcia-Dorado, Manfred Ernst, Damien Kelly, Michael Krainin, Chia-Kai Liang, Marc Levoy, and Peyman Milanfar. Handheld multi-frame superresolution. ACM TOG, 2019. 2 +[49] Tianfan Xue, Baian Chen, Jiajun Wu, Donglai Wei, and William T Freeman. Video enhancement with task-oriented flow. IJCV, 2019. 2, 5, 6 + +[50] Huanjing Yue, Cong Cao, Lei Liao, Ronghe Chu, and Jingyu Yang. Supervised raw video denoising with a benchmark dataset on dynamic scenes. In CVPR, 2020. 2, 5 +[51] Kai Zhang, Jingyun Liang, Luc Van Gool, and Radu Timofte. Designing a practical degradation model for deep blind image super-resolution. In CVPR, 2021. 3 +[52] Kai Zhang, Yawei Li, Jingyun Liang, Jiezhang Cao, Yu-lun Zhang, Hao Tang, Deng-Ping Fan, Radu Timofte, and Luc Van Gool. Practical blind image denoising via swim-conv-unet and data synthesis. MIR, 2023. 3 +[53] Ming Zhang and Bahadir K. Gunturk. Multiresolution bilateral filtering for image denoising. IEEE TIP, 2008. 4 +[54] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 5 \ No newline at end of file diff --git a/2025/Classic Video Denoising in a Machine Learning World_ Robust, Fast, and Controllable/images.zip b/2025/Classic Video Denoising in a Machine Learning World_ Robust, Fast, and Controllable/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..e8544cec63820f911dc928017f576e1c04b01059 --- /dev/null +++ b/2025/Classic Video Denoising in a Machine Learning World_ Robust, Fast, and Controllable/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:01151e892fe88b4521ffb356288a33b5b9808894900157e033f46251c77dbfbd +size 890309 diff --git a/2025/Classic Video Denoising in a Machine Learning World_ Robust, Fast, and Controllable/layout.json b/2025/Classic Video Denoising in a Machine Learning World_ Robust, Fast, and Controllable/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..bdab5253e5c1b50b1c01a525e9c613b30175b8ae --- /dev/null +++ b/2025/Classic Video Denoising in a Machine Learning World_ Robust, Fast, and Controllable/layout.json @@ -0,0 +1,9070 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 135, + 103, + 475, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 103, + 475, + 138 + ], + "spans": [ + { + "bbox": [ + 135, + 103, + 475, + 138 + ], + "type": "text", + "content": "Classic Video Denoising in a Machine Learning World: Robust, Fast, and Controllable" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 154, + 160, + 455, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 154, + 160, + 455, + 175 + ], + "spans": [ + { + "bbox": [ + 154, + 160, + 455, + 175 + ], + "type": "text", + "content": "Xin Jin" + }, + { + "bbox": [ + 154, + 160, + 455, + 175 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 154, + 160, + 455, + 175 + ], + "type": "text", + "content": " Simon Niklaus" + }, + { + "bbox": [ + 154, + 160, + 455, + 175 + ], + "type": "inline_equation", + "content": "^{2\\dagger}" + }, + { + "bbox": [ + 154, + 160, + 455, + 175 + ], + "type": "text", + "content": " Zhoutong Zhang" + }, + { + "bbox": [ + 154, + 160, + 455, + 175 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 154, + 160, + 455, + 175 + ], + "type": "text", + "content": " Zhihao Xia" + }, + { + "bbox": [ + 154, + 160, + 455, + 175 + ], + "type": "inline_equation", + "content": "^{3}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 154, + 175, + 455, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 154, + 175, + 455, + 190 + ], + "spans": [ + { + "bbox": [ + 154, + 175, + 455, + 190 + ], + "type": "text", + "content": "Chunle Guo" + }, + { + "bbox": [ + 154, + 175, + 455, + 190 + ], + "type": "inline_equation", + "content": "^{1,4‡}" + }, + { + "bbox": [ + 154, + 175, + 455, + 190 + ], + "type": "text", + "content": " Yuting Yang" + }, + { + "bbox": [ + 154, + 175, + 455, + 190 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 154, + 175, + 455, + 190 + ], + "type": "text", + "content": " Jiawen Chen" + }, + { + "bbox": [ + 154, + 175, + 455, + 190 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 154, + 175, + 455, + 190 + ], + "type": "text", + "content": " Chongyi Li" + }, + { + "bbox": [ + 154, + 175, + 455, + 190 + ], + "type": "inline_equation", + "content": "^{1,4}" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 119, + 190, + 492, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 119, + 190, + 492, + 203 + ], + "spans": [ + { + "bbox": [ + 119, + 190, + 492, + 203 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 119, + 190, + 492, + 203 + ], + "type": "text", + "content": "VCIP, CS, Nankai University " + }, + { + "bbox": [ + 119, + 190, + 492, + 203 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 119, + 190, + 492, + 203 + ], + "type": "text", + "content": "Adobe Research " + }, + { + "bbox": [ + 119, + 190, + 492, + 203 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 119, + 190, + 492, + 203 + ], + "type": "text", + "content": "Adobe " + }, + { + "bbox": [ + 119, + 190, + 492, + 203 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 119, + 190, + 492, + 203 + ], + "type": "text", + "content": "NKIARI, Shenzhen Futian" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 186, + 205, + 422, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 186, + 205, + 422, + 217 + ], + "spans": [ + { + "bbox": [ + 186, + 205, + 422, + 217 + ], + "type": "text", + "content": "xjin@mail.nankai.edu.cn, sniklaus@adobe.com," + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 208, + 220, + 401, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 208, + 220, + 401, + 232 + ], + "spans": [ + { + "bbox": [ + 208, + 220, + 401, + 232 + ], + "type": "text", + "content": "{guochunle,lichongyi}@nankai.edu.cn" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 200, + 234, + 408, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 200, + 234, + 408, + 245 + ], + "spans": [ + { + "bbox": [ + 200, + 234, + 408, + 245 + ], + "type": "text", + "content": "https://srameo.github.io/projects/levd" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 57, + 255, + 138, + 369 + ], + "blocks": [ + { + "bbox": [ + 57, + 255, + 138, + 369 + ], + "lines": [ + { + "bbox": [ + 57, + 255, + 138, + 369 + ], + "spans": [ + { + "bbox": [ + 57, + 255, + 138, + 369 + ], + "type": "image", + "image_path": "4efc2cd0f1ea64e14b5777b5312b57420d30070c0a38ba489869a982e3741489.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 88, + 373, + 108, + 383 + ], + "lines": [ + { + "bbox": [ + 88, + 373, + 108, + 383 + ], + "spans": [ + { + "bbox": [ + 88, + 373, + 108, + 383 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 141, + 255, + 223, + 369 + ], + "blocks": [ + { + "bbox": [ + 141, + 255, + 223, + 369 + ], + "lines": [ + { + "bbox": [ + 141, + 255, + 223, + 369 + ], + "spans": [ + { + "bbox": [ + 141, + 255, + 223, + 369 + ], + "type": "image", + "image_path": "451109cddf2cdc2c1ce4f5be28fb772a8728cb2f5b25c3a3954fd18a5d576ebb.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 163, + 372, + 200, + 382 + ], + "lines": [ + { + "bbox": [ + 163, + 372, + 200, + 382 + ], + "spans": [ + { + "bbox": [ + 163, + 372, + 200, + 382 + ], + "type": "inline_equation", + "content": "\\mathrm{VRT}^{\\dagger}" + }, + { + "bbox": [ + 163, + 372, + 200, + 382 + ], + "type": "text", + "content": " [26]" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 150, + 383, + 212, + 391 + ], + "lines": [ + { + "bbox": [ + 150, + 383, + 212, + 391 + ], + "spans": [ + { + "bbox": [ + 150, + 383, + 212, + 391 + ], + "type": "text", + "content": "(sophisticated model)" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 225, + 255, + 306, + 369 + ], + "blocks": [ + { + "bbox": [ + 225, + 255, + 306, + 369 + ], + "lines": [ + { + "bbox": [ + 225, + 255, + 306, + 369 + ], + "spans": [ + { + "bbox": [ + 225, + 255, + 306, + 369 + ], + "type": "image", + "image_path": "ea7e17ec687dedc64ec0ba245e20bf2b17c88cb9c50091cf261f9dfdb1789655.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 231, + 373, + 298, + 382 + ], + "lines": [ + { + "bbox": [ + 231, + 373, + 298, + 382 + ], + "spans": [ + { + "bbox": [ + 231, + 373, + 298, + 382 + ], + "type": "text", + "content": "Real-ESRGAN [46]" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 233, + 383, + 297, + 391 + ], + "lines": [ + { + "bbox": [ + 233, + 383, + 297, + 391 + ], + "spans": [ + { + "bbox": [ + 233, + 383, + 297, + 391 + ], + "type": "text", + "content": "(sophisticated dataset)" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 396, + 555, + 419 + ], + "lines": [ + { + "bbox": [ + 55, + 396, + 555, + 419 + ], + "spans": [ + { + "bbox": [ + 55, + 396, + 555, + 419 + ], + "type": "text", + "content": "Figure 1. In real-world videos, noise can manifest in vastly different ways, causing supervised methods to fail when the input is too far outside the training distribution. Footage provided by Robert Kjettrup, face pixelated after inference. We denote retrained models with a " + }, + { + "bbox": [ + 55, + 396, + 555, + 419 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 55, + 396, + 555, + 419 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 308, + 255, + 390, + 369 + ], + "blocks": [ + { + "bbox": [ + 308, + 255, + 390, + 369 + ], + "lines": [ + { + "bbox": [ + 308, + 255, + 390, + 369 + ], + "spans": [ + { + "bbox": [ + 308, + 255, + 390, + 369 + ], + "type": "image", + "image_path": "0498becf27d39aa6541777cd5cd727475a3c9511cb9ab028960376cefe66e847.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 331, + 373, + 368, + 382 + ], + "lines": [ + { + "bbox": [ + 331, + 373, + 368, + 382 + ], + "spans": [ + { + "bbox": [ + 331, + 373, + 368, + 382 + ], + "type": "text", + "content": "MF2F [14]" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 310, + 383, + 388, + 392 + ], + "lines": [ + { + "bbox": [ + 310, + 383, + 388, + 392 + ], + "spans": [ + { + "bbox": [ + 310, + 383, + 388, + 392 + ], + "type": "text", + "content": "(sophisticated supervision)" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 391, + 255, + 473, + 369 + ], + "blocks": [ + { + "bbox": [ + 391, + 255, + 473, + 369 + ], + "lines": [ + { + "bbox": [ + 391, + 255, + 473, + 369 + ], + "spans": [ + { + "bbox": [ + 391, + 255, + 473, + 369 + ], + "type": "image", + "image_path": "33790d91080c2092151131a3f11e8d9e74a5f794b46f9a11c48eb3a67c5faed7.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 412, + 373, + 452, + 382 + ], + "lines": [ + { + "bbox": [ + 412, + 373, + 452, + 382 + ], + "spans": [ + { + "bbox": [ + 412, + 373, + 452, + 382 + ], + "type": "text", + "content": "UDVD [40]" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 394, + 383, + 471, + 392 + ], + "lines": [ + { + "bbox": [ + 394, + 383, + 471, + 392 + ], + "spans": [ + { + "bbox": [ + 394, + 383, + 471, + 392 + ], + "type": "text", + "content": "(sophisticated supervision)" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 475, + 255, + 558, + 369 + ], + "blocks": [ + { + "bbox": [ + 475, + 255, + 558, + 369 + ], + "lines": [ + { + "bbox": [ + 475, + 255, + 558, + 369 + ], + "spans": [ + { + "bbox": [ + 475, + 255, + 558, + 369 + ], + "type": "image", + "image_path": "fbc1097a89e12e7276e5d2b7f0909f438ed2fb90f4ffcb476e97cfd293bc984c.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 490, + 373, + 541, + 382 + ], + "lines": [ + { + "bbox": [ + 490, + 373, + 541, + 382 + ], + "spans": [ + { + "bbox": [ + 490, + 373, + 541, + 382 + ], + "type": "text", + "content": "Ours - RFCVD" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "bbox": [ + 152, + 433, + 200, + 445 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 433, + 200, + 445 + ], + "spans": [ + { + "bbox": [ + 152, + 433, + 200, + 445 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 54, + 458, + 297, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 458, + 297, + 673 + ], + "spans": [ + { + "bbox": [ + 54, + 458, + 297, + 673 + ], + "type": "text", + "content": "Denoising is a crucial step in many video processing pipelines such as in interactive editing, where high quality, speed, and user control are essential. While recent approaches achieve significant improvements in denoising quality by leveraging deep learning, they are prone to unexpected failures due to discrepancies between training data distributions and the wide variety of noise patterns found in real-world videos. These methods also tend to be slow and lack user control. In contrast, traditional denoising methods perform reliably on in-the-wild videos and run relatively quickly on modern hardware. However, they require manually tuning parameters for each input video, which is not only tedious but also requires skill. We bridge the gap between these two paradigms by proposing a differentiable denoising pipeline based on traditional methods. A neural network is then trained to predict the optimal denoising parameters for each specific input, resulting in a robust and efficient approach that also supports user control." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 315, + 433, + 394, + 445 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 433, + 394, + 445 + ], + "spans": [ + { + "bbox": [ + 315, + 433, + 394, + 445 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 313, + 457, + 555, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 457, + 555, + 601 + ], + "spans": [ + { + "bbox": [ + 313, + 457, + 555, + 601 + ], + "type": "text", + "content": "Video denoising is a fundamental part of any video editing pipeline, and it is typically applied prior to color grading. After interviewing professional video editors, we have found that they not only want clean results but they also want the denoising to be quick as not to interrupt their workflow, and they want controllability to assert their artistic expression. Specifically, editors often have to decide to leave some noise in the footage in favor of over-smoothing or vice versa, and the answer can differ for each case. For this reason, professional cameras like the Arri Alexa even allow the user to choose between different noise profiles to better support different post-production workflows." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "content": "Unfortunately, recent work on video denoising tends to focus only on the first aspect, the denoising quality, which makes these solutions impractical for the typical video editing workflow. At the same time, we have found that such video denoising approaches are subject to failure cases like the one in Fig. 1 even though they focus on quality. This is not surprising though, considering that noise in real videos can manifest itself in wildly different ways. That is, in addition to the typical degradation that images are subject to," + } + ] + } + ], + "index": 30 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 693, + 294, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 693, + 294, + 703 + ], + "spans": [ + { + "bbox": [ + 66, + 693, + 294, + 703 + ], + "type": "text", + "content": "*This project was done during Xin Jin's internship at Adobe Research." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 67, + 703, + 198, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 198, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 198, + 713 + ], + "type": "text", + "content": "†Project lead. ‡ Corresponding author." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2084" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 68, + 292, + 157 + ], + "blocks": [ + { + "bbox": [ + 56, + 68, + 292, + 157 + ], + "lines": [ + { + "bbox": [ + 56, + 68, + 292, + 157 + ], + "spans": [ + { + "bbox": [ + 56, + 68, + 292, + 157 + ], + "type": "image", + "image_path": "c613ff942e19a76b73e678ff027bca564ac3f5d4528da137eae75312d5b858d9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 163, + 295, + 187 + ], + "lines": [ + { + "bbox": [ + 55, + 163, + 295, + 187 + ], + "spans": [ + { + "bbox": [ + 55, + 163, + 295, + 187 + ], + "type": "text", + "content": "Figure 2. High-level comparison of how related work approaches video denoising (left) and our proposed approach (right)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 215, + 295, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 215, + 295, + 300 + ], + "spans": [ + { + "bbox": [ + 54, + 215, + 295, + 300 + ], + "type": "text", + "content": "videos also leverage temporal compression. In the H.264 codec, for example, there are P-frames and B-frames that copy information from I-frames which means that the noise is often temporally correlated. Furthermore, there are many different video codecs that affect the noise profile in different ways, and even for the same codec, the noise can vary significantly when using different encoders or settings." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 305, + 295, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 305, + 295, + 402 + ], + "spans": [ + { + "bbox": [ + 54, + 305, + 295, + 402 + ], + "type": "text", + "content": "To tackle the complexity of video noise in a deep learning context and as summarized in Fig. 2 (a), we have found that popular approaches either propose a sophisticated way to supervise the model [14, 40], utilize an intricate pipeline to simulate noise [46], leverage an image formation that by design is less prone to overfitting [13], or a combination of these. Yet, the aforementioned example in Fig. 1 demonstrates that this is not always sufficient." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 407, + 295, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 407, + 295, + 612 + ], + "spans": [ + { + "bbox": [ + 55, + 407, + 295, + 612 + ], + "type": "text", + "content": "In contrast, classic denoisers not only work reasonably well but they are also relatively fast, especially on modern hardware. However, they require manually tuning parameters which is not only tedious but also requires a certain amount of skill. We see this as an opportunity. Specifically and as shown in Fig. 2 (b), by implementing a traditional denoising pipeline in a differentiable manner, we can have a neural network learn to predict what the optimal denoising parameters for a given input should be. As shown in Fig. 3, not only does this end up being robust but it is also comparatively fast while providing support for user control. To be clear on our objectives, if the noise profile is known, such as when developing a denoiser for a specific camera model subject to a constrained video encoder, any reasonable deep learning approach would provide better denoising results. However, such an approach would in turn be limited in terms of robustness and controllability." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 618, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 618, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 54, + 618, + 295, + 715 + ], + "type": "text", + "content": "In short, our contributions are (1) an approach to leverage a traditional video denoising pipeline in a deep learning setting where we decouple the analysis of the noise from the denoising itself to improve efficiency by avoiding redundant compute, (2) a not only robust and fast but also controllable video denosier (RFCVD), and (3) a training augmentation pipeline based on additive white Gaussian noise together with H.264 transcoding that works surprisingly well." + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 317, + 69, + 553, + 225 + ], + "blocks": [ + { + "bbox": [ + 317, + 69, + 553, + 225 + ], + "lines": [ + { + "bbox": [ + 317, + 69, + 553, + 225 + ], + "spans": [ + { + "bbox": [ + 317, + 69, + 553, + 225 + ], + "type": "image", + "image_path": "792d63a9c11df280a50ebf42916cc13ae56dabea86898464513706235a06f898.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 226, + 555, + 271 + ], + "lines": [ + { + "bbox": [ + 313, + 226, + 555, + 271 + ], + "spans": [ + { + "bbox": [ + 313, + 226, + 555, + 271 + ], + "type": "text", + "content": "Figure 3. Video denoising on the CRVD (sRGB) benchmark [50] using the PSNR across all ISO values with respect to the computational efficiency on an RTX 3090 GPU in FPS (frames per second). We improved some models through retraining, denoted with a " + }, + { + "bbox": [ + 313, + 226, + 555, + 271 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 313, + 226, + 555, + 271 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 284, + 400, + 297 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 284, + 400, + 297 + ], + "spans": [ + { + "bbox": [ + 314, + 284, + 400, + 297 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 308, + 555, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 308, + 555, + 477 + ], + "spans": [ + { + "bbox": [ + 313, + 308, + 555, + 477 + ], + "type": "text", + "content": "Traditional Video Denoising. Most traditional video denoising methods can be viewed as a series of linear and non-linear filtering processes, where the filter strength is a function of the estimated noise level. Many such methods are based on block matching [10-12, 31], where similar patches are first aggregated within and across frames before being filtered and merged to produce the denoised result. There are many variations of this process of course, for example by improving the matching [28] or by leveraging bilateral filtering [8, 18, 36, 43] for spatial and temporal denoising [1, 37]. In doing so, image pyramids [3, 16] can be used to improve the computational efficiency and recent iterations of classic multi-frame denoising methods are even quick to run on commodity smartphones [20, 27, 48]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 481, + 556, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 481, + 556, + 662 + ], + "spans": [ + { + "bbox": [ + 313, + 481, + 556, + 662 + ], + "type": "text", + "content": "Machine Learning in Video Denoising. In recent years, it has become increasingly popular to tackle video denoising with the help of machine learning. This includes methods that, like traditional video denoisers, perform explicit matching [13, 44, 49] as well as those where correspondences are only implicit [24, 32, 42]. And while most of these methods denoise each frame independently, there are also IIR-like approaches where information is passed along such that previously denoised frames can guide future ones [5, 6]. On the model architecture side, the rise of attention in the vision and language community has also led to various attention mechanisms for video restoration [4, 26, 29, 41, 45]. But with machine learning there is a catch, if an input has a noise pattern that has not been seen during training then the denoising is unpredictable." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "content": "Noise Simulation. A common approach to facilitate better generalizability on in-the-wild footage is to employ a more sophisticated degradation pipeline when adding synthetic noise to the ground truth. This includes not only do" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2085" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 85, + 553, + 318 + ], + "blocks": [ + { + "bbox": [ + 63, + 72, + 309, + 83 + ], + "lines": [ + { + "bbox": [ + 63, + 72, + 309, + 83 + ], + "spans": [ + { + "bbox": [ + 63, + 72, + 309, + 83 + ], + "type": "text", + "content": "Noise Profiling on an Anchor Frame Through a Hypernetwork" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 61, + 85, + 553, + 318 + ], + "lines": [ + { + "bbox": [ + 61, + 85, + 553, + 318 + ], + "spans": [ + { + "bbox": [ + 61, + 85, + 553, + 318 + ], + "type": "image", + "image_path": "eea26d388a9b6825f76e7393350f7cac0dd3d0bba7410ef509e6081aa9480a39.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 321, + 555, + 399 + ], + "lines": [ + { + "bbox": [ + 55, + 321, + 555, + 399 + ], + "spans": [ + { + "bbox": [ + 55, + 321, + 555, + 399 + ], + "type": "text", + "content": "Figure 4. Video denoising fundamentally first needs to analyze the noise and then remove it. We mimic this in our pipeline by first estimating a noise profile on a random anchor frame (top left in green) before using this profile to denoise the video (bottom in yellow). Specifically, we leverage a hypernetwork configuration where the noise profile " + }, + { + "bbox": [ + 55, + 321, + 555, + 399 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 55, + 321, + 555, + 399 + ], + "type": "text", + "content": " is essentially the parameters of the subsequent denoiser. That is, our denoiser is a traditional pipeline consisting of (1) a Wiener filter that performs temporal denoising of neighboring frames that were aligned via optical flow and (2) a bilateral Laplacian pyramid filter for spatial denoising of the temporally merged frames, where a small neural network " + }, + { + "bbox": [ + 55, + 321, + 555, + 399 + ], + "type": "inline_equation", + "content": "\\mathcal{P}(\\cdot ;\\theta)" + }, + { + "bbox": [ + 55, + 321, + 555, + 399 + ], + "type": "text", + "content": " predicts spatially-varying parameters for the Wiener merger and the bilateral filters. This separation of concerns improves the overall efficiency since it avoids having to redundantly analyze the noise over and over again." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 407, + 295, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 407, + 295, + 468 + ], + "spans": [ + { + "bbox": [ + 55, + 407, + 295, + 468 + ], + "type": "text", + "content": "ing one but two rounds of degradations [46], shuffling the degradations [51], or integrating more advanced operators such as tone mapping [52]. Although these approaches improve the ability to denoise real-world videos, they can only reduce but not entirely resolve the generalizability gap." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 472, + 295, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 472, + 295, + 604 + ], + "spans": [ + { + "bbox": [ + 55, + 472, + 295, + 604 + ], + "type": "text", + "content": "Self-Supervised Learning. Another strategy to improve the generalizability in video denoising is to adopt self-supervision on real-world videos without requiring access to a ground truth. For example, this can be achieved by leveraging the nature of regression losses [25] or by building on the blind spot idea [14, 17, 23, 40]. While these approaches make it possible to train on noisy in-the-wild videos, it is not necessarily feasible to collect a dataset that contains all possible types of noise that occur in the real world. Therefore, self-supervision often also relies on test-time adaptation which comes with its own challenges." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 620, + 111, + 632 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 620, + 111, + 632 + ], + "spans": [ + { + "bbox": [ + 55, + 620, + 111, + 632 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 641, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 295, + 714 + ], + "type": "text", + "content": "Video denoising fundamentally first needs to analyze the noise and then remove it. In contrast, recent approaches perform both tasks at once by utilizing a single neural network that is given the noisy input frames and is tasked with producing the clean output. The noise profile in a video does typically not change over time though, yet these ap" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 408, + 553, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 408, + 553, + 431 + ], + "spans": [ + { + "bbox": [ + 313, + 408, + 553, + 431 + ], + "type": "text", + "content": "proaches independently denoise one frame after another so they have to analyze the noise over and over again." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 433, + 554, + 587 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 433, + 554, + 587 + ], + "spans": [ + { + "bbox": [ + 313, + 433, + 554, + 587 + ], + "type": "text", + "content": "To avoid this redundancy, we follow the natural separation of analyzing the noise and then removing it. As shown in Fig. 4, we first estimate the noise profile on an anchor frame and then denoise the video using the estimated profile. Specifically, we leverage a hypernetwork setup where the noise profile " + }, + { + "bbox": [ + 313, + 433, + 554, + 587 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 433, + 554, + 587 + ], + "type": "text", + "content": " provides the parameters of the subsequent denoiser. That is, our denoiser is a traditional pipeline consisting of (1) a Wiener filter that performs temporal denoising of neighboring frames that were aligned via optical flow and (2) a bilateral Laplacian pyramid filter for spatial denoising of the temporally merged frames, where a small neural network " + }, + { + "bbox": [ + 313, + 433, + 554, + 587 + ], + "type": "inline_equation", + "content": "\\mathcal{P}(\\cdot ;\\theta)" + }, + { + "bbox": [ + 313, + 433, + 554, + 587 + ], + "type": "text", + "content": " predicts spatially-varying parameters for the Wiener merger and the bilateral filters." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 589, + 553, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 589, + 553, + 613 + ], + "spans": [ + { + "bbox": [ + 313, + 589, + 553, + 613 + ], + "type": "text", + "content": "We will subsequently discuss these parts in turn before providing more details on user control and the training." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 622, + 406, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 622, + 406, + 635 + ], + "spans": [ + { + "bbox": [ + 313, + 622, + 406, + 635 + ], + "type": "text", + "content": "3.1. Noise Profiling" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 641, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 554, + 713 + ], + "type": "text", + "content": "Since noise is not necessarily spatially uniform but often dependent on the signal, we argue that a noise profile should not be just a fixed set of values for the entire input but rather be a descriptor that makes it possible to derive spatially-varying denoising parameters. We see this as an ideal use case for hypernetworks [19], where a noise profiler esti" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2086" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 69, + 138, + 174 + ], + "blocks": [ + { + "bbox": [ + 56, + 69, + 138, + 174 + ], + "lines": [ + { + "bbox": [ + 56, + 69, + 138, + 174 + ], + "spans": [ + { + "bbox": [ + 56, + 69, + 138, + 174 + ], + "type": "image", + "image_path": "03eb94de23ae0d8a136bd6e105428eeb4057be9a3566073988ccd7b21a83c1f1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 88, + 176, + 108, + 186 + ], + "lines": [ + { + "bbox": [ + 88, + 176, + 108, + 186 + ], + "spans": [ + { + "bbox": [ + 88, + 176, + 108, + 186 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 140, + 70, + 223, + 174 + ], + "blocks": [ + { + "bbox": [ + 140, + 70, + 223, + 174 + ], + "lines": [ + { + "bbox": [ + 140, + 70, + 223, + 174 + ], + "spans": [ + { + "bbox": [ + 140, + 70, + 223, + 174 + ], + "type": "image", + "image_path": "ad9e2eb9e981b29dfd671fe69868eeb2f3ad0d935aec408f3ae3e1069df1f2f1.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 145, + 176, + 218, + 194 + ], + "lines": [ + { + "bbox": [ + 145, + 176, + 218, + 194 + ], + "spans": [ + { + "bbox": [ + 145, + 176, + 218, + 194 + ], + "type": "text", + "content": "reduced temporal only " + }, + { + "bbox": [ + 145, + 176, + 218, + 194 + ], + "type": "inline_equation", + "content": "(0.2\\cdot \\sigma^2,0\\cdot \\sigma_d,0\\cdot \\sigma_r)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 225, + 70, + 306, + 174 + ], + "blocks": [ + { + "bbox": [ + 225, + 70, + 306, + 174 + ], + "lines": [ + { + "bbox": [ + 225, + 70, + 306, + 174 + ], + "spans": [ + { + "bbox": [ + 225, + 70, + 306, + 174 + ], + "type": "image", + "image_path": "ee0a4435d21cba2ce07442037b033b55e7a52fff9663f7627660cf0d649de330.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 234, + 176, + 296, + 194 + ], + "lines": [ + { + "bbox": [ + 234, + 176, + 296, + 194 + ], + "spans": [ + { + "bbox": [ + 234, + 176, + 296, + 194 + ], + "type": "text", + "content": "temporal only " + }, + { + "bbox": [ + 234, + 176, + 296, + 194 + ], + "type": "inline_equation", + "content": "(1\\cdot \\sigma^2,0\\cdot \\sigma_d,0\\cdot \\sigma_r)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 194, + 555, + 237 + ], + "lines": [ + { + "bbox": [ + 55, + 194, + 555, + 237 + ], + "spans": [ + { + "bbox": [ + 55, + 194, + 555, + 237 + ], + "type": "text", + "content": "Figure 5. By scaling the denoising parameters estimated by " + }, + { + "bbox": [ + 55, + 194, + 555, + 237 + ], + "type": "inline_equation", + "content": "\\mathcal{P}(\\cdot ;\\theta)" + }, + { + "bbox": [ + 55, + 194, + 555, + 237 + ], + "type": "text", + "content": ", our framework makes it easy to control the amount of temporal denoising (through " + }, + { + "bbox": [ + 55, + 194, + 555, + 237 + ], + "type": "inline_equation", + "content": "\\sigma^2" + }, + { + "bbox": [ + 55, + 194, + 555, + 237 + ], + "type": "text", + "content": ") as well as the amount of spatial denoising (through " + }, + { + "bbox": [ + 55, + 194, + 555, + 237 + ], + "type": "inline_equation", + "content": "\\sigma_d" + }, + { + "bbox": [ + 55, + 194, + 555, + 237 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 194, + 555, + 237 + ], + "type": "inline_equation", + "content": "\\sigma_r" + }, + { + "bbox": [ + 55, + 194, + 555, + 237 + ], + "type": "text", + "content": "). We scale the denoising parameters equivalently for the luma and chroma channels in this sample, but they can be scaled independently for more control. Please see the supplementary for a video demo that demonstrates how users can control the denoising in an interactive manner. Footage provided by Amit Zinman." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 307, + 70, + 390, + 174 + ], + "blocks": [ + { + "bbox": [ + 307, + 70, + 390, + 174 + ], + "lines": [ + { + "bbox": [ + 307, + 70, + 390, + 174 + ], + "spans": [ + { + "bbox": [ + 307, + 70, + 390, + 174 + ], + "type": "image", + "image_path": "712c0b43eaf5055d26349b9a6f8881bb85100a2ef1a47be6b8ed107beb14b91a.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 176, + 385, + 194 + ], + "lines": [ + { + "bbox": [ + 313, + 176, + 385, + 194 + ], + "spans": [ + { + "bbox": [ + 313, + 176, + 385, + 194 + ], + "type": "text", + "content": "reduced spatial only " + }, + { + "bbox": [ + 313, + 176, + 385, + 194 + ], + "type": "inline_equation", + "content": "(0\\cdot \\sigma^{2},0.2\\cdot \\sigma_{d},0.2\\cdot \\sigma_{r})" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 391, + 70, + 474, + 174 + ], + "blocks": [ + { + "bbox": [ + 391, + 70, + 474, + 174 + ], + "lines": [ + { + "bbox": [ + 391, + 70, + 474, + 174 + ], + "spans": [ + { + "bbox": [ + 391, + 70, + 474, + 174 + ], + "type": "image", + "image_path": "e5d5f37b5cf91cd079f40ccd6ae4b57a05565d5d5897ccf4de24ae648121e492.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 402, + 176, + 463, + 194 + ], + "lines": [ + { + "bbox": [ + 402, + 176, + 463, + 194 + ], + "spans": [ + { + "bbox": [ + 402, + 176, + 463, + 194 + ], + "type": "text", + "content": "spatial only " + }, + { + "bbox": [ + 402, + 176, + 463, + 194 + ], + "type": "inline_equation", + "content": "(0\\cdot \\sigma^{2},1\\cdot \\sigma_{d},1\\cdot \\sigma_{r})" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 475, + 70, + 558, + 174 + ], + "blocks": [ + { + "bbox": [ + 475, + 70, + 558, + 174 + ], + "lines": [ + { + "bbox": [ + 475, + 70, + 558, + 174 + ], + "spans": [ + { + "bbox": [ + 475, + 70, + 558, + 174 + ], + "type": "image", + "image_path": "182ed8d1157ea58a335c7c3636b019f63f2f9e999ff20ffe761ce505120f6f8a.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 486, + 176, + 547, + 194 + ], + "lines": [ + { + "bbox": [ + 486, + 176, + 547, + 194 + ], + "spans": [ + { + "bbox": [ + 486, + 176, + 547, + 194 + ], + "type": "text", + "content": "default settings " + }, + { + "bbox": [ + 486, + 176, + 547, + 194 + ], + "type": "inline_equation", + "content": "(1\\cdot \\sigma^{2},1\\cdot \\sigma_{d},1\\cdot \\sigma_{r})" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 245, + 296, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 245, + 296, + 304 + ], + "spans": [ + { + "bbox": [ + 55, + 245, + 296, + 304 + ], + "type": "text", + "content": "mates the weights " + }, + { + "bbox": [ + 55, + 245, + 296, + 304 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 55, + 245, + 296, + 304 + ], + "type": "text", + "content": " of a small neural network " + }, + { + "bbox": [ + 55, + 245, + 296, + 304 + ], + "type": "inline_equation", + "content": "\\mathcal{P}(\\cdot ;\\theta)" + }, + { + "bbox": [ + 55, + 245, + 296, + 304 + ], + "type": "text", + "content": " that predicts spatially-varying denoising parameters. However, instead of directly predicting " + }, + { + "bbox": [ + 55, + 245, + 296, + 304 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 55, + 245, + 296, + 304 + ], + "type": "text", + "content": ", we stabilize the training through NPA [35] where " + }, + { + "bbox": [ + 55, + 245, + 296, + 304 + ], + "type": "inline_equation", + "content": "\\theta = \\theta^0 +\\Delta \\theta" + }, + { + "bbox": [ + 55, + 245, + 296, + 304 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 55, + 245, + 296, + 304 + ], + "type": "inline_equation", + "content": "\\theta^0" + }, + { + "bbox": [ + 55, + 245, + 296, + 304 + ], + "type": "text", + "content": " being learnable parameters and " + }, + { + "bbox": [ + 55, + 245, + 296, + 304 + ], + "type": "inline_equation", + "content": "\\Delta \\theta" + }, + { + "bbox": [ + 55, + 245, + 296, + 304 + ], + "type": "text", + "content": " being the hypernet prediction." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 305, + 296, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 305, + 296, + 413 + ], + "spans": [ + { + "bbox": [ + 55, + 305, + 296, + 413 + ], + "type": "text", + "content": "Intuitively, analyzing the noise profile of a given image is a mixture of low-level image processing and high-level semantics. Specifically, one first needs to understand what an image depicts before being able to examine the details to discern between unintended noise and actual texture. For this reason, we leverage a pre-trained backbone " + }, + { + "bbox": [ + 55, + 305, + 296, + 413 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 55, + 305, + 296, + 413 + ], + "type": "text", + "content": " in the form of a ConvNext [30] with a random MLP head to serve as the hypernet. We have found this backbone to be crucial for the quality of the predicted noise profile." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 413, + 296, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 413, + 296, + 485 + ], + "spans": [ + { + "bbox": [ + 55, + 413, + 296, + 485 + ], + "type": "text", + "content": "Lastly, which anchor frame should we choose to analyze the noise of a given video? We would like this choice not to matter, such that the denoising result is independent of having to choose a \"good\" anchor. To achieve this, we always pick the first frame of a video as the anchor frame because it is as good as any, and then utilize a consistency loss" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 123, + 495, + 295, + 510 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 123, + 495, + 295, + 510 + ], + "spans": [ + { + "bbox": [ + 123, + 495, + 295, + 510 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {c s t s y} = \\left\\| \\Delta \\theta - \\Delta \\theta_ {i} \\right\\| _ {2}, \\tag {1}", + "image_path": "a7019b7a09ac5dd051a441b71796794abdee472193108cc15896f26be0c7f8b9.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 55, + 518, + 296, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 518, + 296, + 543 + ], + "spans": [ + { + "bbox": [ + 55, + 518, + 296, + 543 + ], + "type": "text", + "content": "where we encourage the hypernet prediction of our anchor frame " + }, + { + "bbox": [ + 55, + 518, + 296, + 543 + ], + "type": "inline_equation", + "content": "\\Delta \\theta" + }, + { + "bbox": [ + 55, + 518, + 296, + 543 + ], + "type": "text", + "content": " and any random neighbor " + }, + { + "bbox": [ + 55, + 518, + 296, + 543 + ], + "type": "inline_equation", + "content": "\\Delta \\theta_{i}" + }, + { + "bbox": [ + 55, + 518, + 296, + 543 + ], + "type": "text", + "content": " to be equal." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 55, + 551, + 172, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 551, + 172, + 563 + ], + "spans": [ + { + "bbox": [ + 55, + 551, + 172, + 563 + ], + "type": "text", + "content": "3.2. Denoising the Video" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 55, + 570, + 295, + 630 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 295, + 630 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 295, + 630 + ], + "type": "text", + "content": "Once we have the noise profile " + }, + { + "bbox": [ + 55, + 570, + 295, + 630 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 55, + 570, + 295, + 630 + ], + "type": "text", + "content": ", we process the input video one output frame at a tim where we obtain spatially-varying denoising parameters through a small neural network " + }, + { + "bbox": [ + 55, + 570, + 295, + 630 + ], + "type": "inline_equation", + "content": "\\mathcal{P}(\\cdot ;\\theta)" + }, + { + "bbox": [ + 55, + 570, + 295, + 630 + ], + "type": "text", + "content": " with three convolutional layers. We use these parameters to do temporal and then spatial denoising." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "text", + "content": "For the temporal denoising, to denoise a given target frame, we first align the two preceding and the two subsequent frames. Specifically, we estimate the optical flow between the target frame and the neighbors using an off-the-shelf SpyNet [38] at a quarter of the target resolution for improved computational efficiency as well as robustness to noise, and warp the neighbors to the target frame using the" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 245, + 554, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 245, + 554, + 317 + ], + "spans": [ + { + "bbox": [ + 313, + 245, + 554, + 317 + ], + "type": "text", + "content": "estimated optical flow. We then merge the aligned frames using a Wiener filter as in Hasinoff et al. [20] but using a tiles size of " + }, + { + "bbox": [ + 313, + 245, + 554, + 317 + ], + "type": "inline_equation", + "content": "8 \\times 8" + }, + { + "bbox": [ + 313, + 245, + 554, + 317 + ], + "type": "text", + "content": " and the noise maps from " + }, + { + "bbox": [ + 313, + 245, + 554, + 317 + ], + "type": "inline_equation", + "content": "\\mathcal{P}(\\cdot; \\theta)" + }, + { + "bbox": [ + 313, + 245, + 554, + 317 + ], + "type": "text", + "content": " instead of approximating the noise as the RMS in each tile. As is typical and to facilitate more control, we perform Wiener filtering on the luma and the chroma channels independently." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 317, + 554, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 317, + 554, + 448 + ], + "spans": [ + { + "bbox": [ + 313, + 317, + 554, + 448 + ], + "type": "text", + "content": "We then apply spatial denoising to the temporally denoised target frame through a bilateral Laplacian pyramid filter with three levels. This borrows many ideas from multiresolution bilateral filtering [53] but using a Laplacian instead of a Gaussian pyramid and using bilateral filtering throughout instead of wavelet thresholding for the first pyramid levels. More specifically, we have " + }, + { + "bbox": [ + 313, + 317, + 554, + 448 + ], + "type": "inline_equation", + "content": "\\mathcal{P}(\\cdot ;\\theta)" + }, + { + "bbox": [ + 313, + 317, + 554, + 448 + ], + "type": "text", + "content": " estimate noise map pyramids and then use these as " + }, + { + "bbox": [ + 313, + 317, + 554, + 448 + ], + "type": "inline_equation", + "content": "\\sigma_{s}" + }, + { + "bbox": [ + 313, + 317, + 554, + 448 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 317, + 554, + 448 + ], + "type": "inline_equation", + "content": "\\sigma_r" + }, + { + "bbox": [ + 313, + 317, + 554, + 448 + ], + "type": "text", + "content": " for the bilateral filtering. Furthermore, just like with the Wiener filter, we perform the denoising separately on the luma and the chroma channels to facilitate user control." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 314, + 456, + 398, + 467 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 456, + 398, + 467 + ], + "spans": [ + { + "bbox": [ + 314, + 456, + 398, + 467 + ], + "type": "text", + "content": "3.3. User Control" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 475, + 554, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 475, + 554, + 629 + ], + "spans": [ + { + "bbox": [ + 313, + 475, + 554, + 629 + ], + "type": "text", + "content": "Through interviews with creative professionals, we have found that they want to assert their artistic expression when denoising a clip. That is, they often have to decide to leave some noise in the footage in favor of over-smoothing and vice versa. The key to facilitating user control in our method is the spatially-varying parameters from " + }, + { + "bbox": [ + 313, + 475, + 554, + 629 + ], + "type": "inline_equation", + "content": "\\mathcal{P}(\\cdot ;\\theta)" + }, + { + "bbox": [ + 313, + 475, + 554, + 629 + ], + "type": "text", + "content": " that guide the temporal and spatial denoising. Specifically, we have two of these maps for " + }, + { + "bbox": [ + 313, + 475, + 554, + 629 + ], + "type": "inline_equation", + "content": "\\sigma^2" + }, + { + "bbox": [ + 313, + 475, + 554, + 629 + ], + "type": "text", + "content": " in the Wiener filter (one map for chroma and one for luma) and in the bilateral Laplacian pyramid filter we have two map pyramids for " + }, + { + "bbox": [ + 313, + 475, + 554, + 629 + ], + "type": "inline_equation", + "content": "\\sigma_{d}" + }, + { + "bbox": [ + 313, + 475, + 554, + 629 + ], + "type": "text", + "content": " as well as two map pyramids for " + }, + { + "bbox": [ + 313, + 475, + 554, + 629 + ], + "type": "inline_equation", + "content": "\\sigma_r" + }, + { + "bbox": [ + 313, + 475, + 554, + 629 + ], + "type": "text", + "content": " (one map pyramid for chroma and one for luma). This makes for a total of six sets of parameters where we can facilitate user control." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "text", + "content": "How do we attenuate these parameters though? As demonstrated in Fig. 5, we have found that simply scaling the spatially-varying denoising parameters uniformly works reasonably well, so we can just provide six tunable knobs to end users. And since our denoising framework runs in real time, changing the knobs provides instantaneous feedback which makes it easy to attain the desired result." + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2087" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 185, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 185, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 185, + 85 + ], + "type": "text", + "content": "3.4. Augmentation Pipeline" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 90, + 294, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 90, + 294, + 198 + ], + "spans": [ + { + "bbox": [ + 54, + 90, + 294, + 198 + ], + "type": "text", + "content": "To train our model, we need pairs of clean and noisy frame sequences just like any other denoiser that leverages machine learning without self-supervision. Specifically, we adopt the typical paradigm of taking a set of \"noise-free\" videos that serve as the ground truth, and augmenting them with various degradations to obtain the corresponding noisy inputs. For this to work well, we have found it to be crucial to include temporal compression in this degradation pipeline since it makes the noise temporally correlated." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 198, + 294, + 305 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 198, + 294, + 305 + ], + "spans": [ + { + "bbox": [ + 54, + 198, + 294, + 305 + ], + "type": "text", + "content": "But let's start with noise-free videos, which we obtain from the REDS dataset [33]. It provides 240 videos and was shot at a relatively high 120 frames per second, allowing us to augment the frame rate via sub-sampling. We extract random frame sequences from these videos and first degrade them through additive white Gaussian noise with a variance sampled from " + }, + { + "bbox": [ + 54, + 198, + 294, + 305 + ], + "type": "inline_equation", + "content": "\\mathcal{U}(1,50)" + }, + { + "bbox": [ + 54, + 198, + 294, + 305 + ], + "type": "text", + "content": ". We then transcode them using an H.264 codec through libx264 with a CRF sampled from " + }, + { + "bbox": [ + 54, + 198, + 294, + 305 + ], + "type": "inline_equation", + "content": "\\mathcal{U}(18,30)" + }, + { + "bbox": [ + 54, + 198, + 294, + 305 + ], + "type": "text", + "content": " to make the noise temporally correlated." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 305, + 294, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 305, + 294, + 426 + ], + "spans": [ + { + "bbox": [ + 54, + 305, + 294, + 426 + ], + "type": "text", + "content": "We have found this data pipeline to work surprisingly well. So well in fact that we retrained the models that we compare to in the evaluation since we were able to get much better results on the CRVD benchmark [50] with our retrained versions than the original checkpoints. However, we did not retrain Real-ESRGAN [46] since it brings its own much more complex augmentation pipeline, and we also did not retrain MF2F [14] and UDVD [40] since these leverage sophisticated ways of supervising the models which are difficult for us to replicate and properly do justice." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 431, + 188, + 443 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 431, + 188, + 443 + ], + "spans": [ + { + "bbox": [ + 55, + 431, + 188, + 443 + ], + "type": "text", + "content": "3.5. Implementation Details" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 449, + 295, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 449, + 295, + 628 + ], + "spans": [ + { + "bbox": [ + 54, + 449, + 295, + 628 + ], + "type": "text", + "content": "In terms of architecture details, we utilize the base size of a ConvNext [30] for the backbone " + }, + { + "bbox": [ + 54, + 449, + 295, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 54, + 449, + 295, + 628 + ], + "type": "text", + "content": " in the hypernetwork. To translate the output from this backbone to the residual noise profile " + }, + { + "bbox": [ + 54, + 449, + 295, + 628 + ], + "type": "inline_equation", + "content": "\\Delta \\theta" + }, + { + "bbox": [ + 54, + 449, + 295, + 628 + ], + "type": "text", + "content": ", we utilize an MLP with five layers and PReLU activations [21] in between each layer. And for the neural network " + }, + { + "bbox": [ + 54, + 449, + 295, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{P}(\\cdot ;\\theta)" + }, + { + "bbox": [ + 54, + 449, + 295, + 628 + ], + "type": "text", + "content": " that predicts the spatially-varying denoising parameters, we leverage three convolution layers with a stride of two and PReLU activations in between. We additionally constrain the output of this network with a softmax [15] to facilitate non-negative denoising parameters. Furthermore, we have found it beneficial to not only provide the (aligned) frames to " + }, + { + "bbox": [ + 54, + 449, + 295, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{P}(\\cdot ;\\theta)" + }, + { + "bbox": [ + 54, + 449, + 295, + 628 + ], + "type": "text", + "content": " but also their gradients, approximated by a Sobel filter, as well as masks that indicate whether or not a pixel in an aligned frame is valid (it may have been warped from outside the frame) [2]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 629, + 295, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 629, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 54, + 629, + 295, + 677 + ], + "type": "text", + "content": "In our reconstruction loss, we minimize the difference between the temporally denoised image " + }, + { + "bbox": [ + 54, + 629, + 295, + 677 + ], + "type": "inline_equation", + "content": "I^{\\mathrm{T}}" + }, + { + "bbox": [ + 54, + 629, + 295, + 677 + ], + "type": "text", + "content": " and the ground truth " + }, + { + "bbox": [ + 54, + 629, + 295, + 677 + ], + "type": "inline_equation", + "content": "I^{\\mathrm{GT}}" + }, + { + "bbox": [ + 54, + 629, + 295, + 677 + ], + "type": "text", + "content": " as well as the difference between all levels of the spatial denoising pyramid " + }, + { + "bbox": [ + 54, + 629, + 295, + 677 + ], + "type": "inline_equation", + "content": "I^{\\mathrm{ST}}" + }, + { + "bbox": [ + 54, + 629, + 295, + 677 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 629, + 295, + 677 + ], + "type": "inline_equation", + "content": "I^{\\mathrm{GT}}" + }, + { + "bbox": [ + 54, + 629, + 295, + 677 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 86, + 683, + 295, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 683, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 86, + 683, + 295, + 715 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {r e c} = \\left\\| I ^ {\\mathrm {T}} - I ^ {\\mathrm {G T}} \\right\\| _ {2} + \\sum_ {l = 1} ^ {3} \\left\\| I _ {l} ^ {\\mathrm {S T}} - I _ {l} ^ {\\mathrm {G T}} \\right\\| _ {2}, \\tag {2}", + "image_path": "8ef666158620adafc5acca7ae914998e43bcdaa5185b9687f2c745ef776c8ff3.jpg" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 315, + 71, + 554, + 216 + ], + "blocks": [ + { + "bbox": [ + 315, + 71, + 554, + 216 + ], + "lines": [ + { + "bbox": [ + 315, + 71, + 554, + 216 + ], + "spans": [ + { + "bbox": [ + 315, + 71, + 554, + 216 + ], + "type": "table", + "html": "
PSNRdelta (higher PSNR is better)SSIMdelta (higher SSIM is better)LPIPSdelta (lower LPIPS is better)
NAFNet [9]30.25-0.747-0.358-
NAFNet†[9]35.32+ 5.070.937+ 0.1900.089- 0.269
FastDVDNet [42]27.89-0.565-0.502-
FastDVDNet†[42]34.63+ 6.750.914+ 0.3490.124- 0.378
TOFlow [49]25.96-0.673-0.251-
TOFlow†[49]34.08+ 8.110.903+ 0.2310.147- 0.104
BasicVSR++ [6]31.98-0.769-0.326-
BasicVSR++†[6]34.34+ 2.350.873+ 0.1040.167- 0.158
VRT [26]31.99-0.784-0.296-
VRT†[26]33.86+ 1.860.848+ 0.0640.192- 0.104
", + "image_path": "829fbf10a8350c0805f77332216be38cc2949cf5cdc8672b68c6c42497a2cd50.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 220, + 553, + 253 + ], + "lines": [ + { + "bbox": [ + 313, + 220, + 553, + 253 + ], + "spans": [ + { + "bbox": [ + 313, + 220, + 553, + 253 + ], + "type": "text", + "content": "Table 1. Denoising results on the CRVD (sRGB) benchmark for various methods that we compare to, both for their original version as well as our retrained one which we denote with a " + }, + { + "bbox": [ + 313, + 220, + 553, + 253 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 313, + 220, + 553, + 253 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 264, + 553, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 264, + 553, + 300 + ], + "spans": [ + { + "bbox": [ + 313, + 264, + 553, + 300 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 264, + 553, + 300 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 313, + 264, + 553, + 300 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 313, + 264, + 553, + 300 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 313, + 264, + 553, + 300 + ], + "type": "text", + "content": "-th level in spatial denoising pyramid. Recall that we also have a consistency loss that we outlined in Eq. (1) in the section about noise profiling." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 300, + 554, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 300, + 554, + 373 + ], + "spans": [ + { + "bbox": [ + 313, + 300, + 554, + 373 + ], + "type": "text", + "content": "We train our model using Adam [22] with an initial learning rate of " + }, + { + "bbox": [ + 313, + 300, + 554, + 373 + ], + "type": "inline_equation", + "content": "2 \\times 10^{-4}" + }, + { + "bbox": [ + 313, + 300, + 554, + 373 + ], + "type": "text", + "content": " that decays to " + }, + { + "bbox": [ + 313, + 300, + 554, + 373 + ], + "type": "inline_equation", + "content": "1 \\times 10^{-7}" + }, + { + "bbox": [ + 313, + 300, + 554, + 373 + ], + "type": "text", + "content": " using a cosine annealing schedule. In total, we train the model for 400 thousand iterations with a batch size of 24 consisting of patches with " + }, + { + "bbox": [ + 313, + 300, + 554, + 373 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 313, + 300, + 554, + 373 + ], + "type": "text", + "content": " pixels. Since our pipeline is computationally lightweight, this takes less than two days." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 382, + 394, + 396 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 382, + 394, + 396 + ], + "spans": [ + { + "bbox": [ + 313, + 382, + 394, + 396 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 403, + 554, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 403, + 554, + 582 + ], + "spans": [ + { + "bbox": [ + 313, + 403, + 554, + 582 + ], + "type": "text", + "content": "We quantitatively evaluate our approach on two types of videos, real and synthetic. For the former, we leverage the CRVD benchmark [50] consisting of RAW videos which we convert to sRGB using the provided deep ISP [39]. For the latter, we use the validation samples from the REDS dataset [33] which we augmented in two ways that are notably different from the AWGN subject to H.264 transcoding during training. Specifically, one way is using film grain noise [34] with AV1 transcoding, and the other way is using spatially correlated Gaussian noise with H.265 transcoding. As for evaluation metrics, we follow the typical paradigm of reporting PSNR, SSIM [47], and LPIPS [54]. Since we argue that our method is computationally efficient, we also report the frames per second (FPS) which we measured on an RTX 3090 GPU with proper CUDA synchronization." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 582, + 553, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 553, + 641 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 553, + 641 + ], + "type": "text", + "content": "In terms of data for our qualitative evaluation, we reached out to industry professionals who sent us footage where they had trouble removing the noise. Since these professionals did not have a release form for all the actors, we sometimes had to anonymize faces post inference." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 642, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 553, + 713 + ], + "type": "text", + "content": "Lastly, we compare our approach to various kinds of denoisers. This includes image denoisers [7, 9], video denoisers [42, 49], recurrent methods [6], transformers [26], models with a sophisticated augmentation pipeline [46], as well as methods that leverage self-supervision [14, 40]. Since we have found that our augmentation pipeline works sur" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2088" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 70, + 555, + 234 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 555, + 234 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 555, + 234 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 555, + 234 + ], + "type": "table", + "html": "
ISO 1600ISO 3200ISO 6400ISO 12800ISO 25600OverallSpeed
PSNRrank (higher PSNR is better)PSNRrank (higher PSNR is better)PSNRrank (higher PSNR is better)PSNRrank (higher PSNR is better)PSNRrank (higher PSNR is better)PSNRrank (higher PSNR is better)FPSrank (higher FPS is better)
\\( SID^† \\)[7]38.85\\( 7^{th} \\) of 1037.68\\( 7^{th} \\) of 1035.82\\( 4^{th} \\) of 1033.51\\( 4^{th} \\) of 1029.18\\( 3^{rd} \\) of 1035.01\\( 4^{th} \\) of 106.95\\( 3^{rd} \\) of 10
\\( NAFNet^† \\)[9]39.48\\( 3^{rd} \\) of 1038.12\\( 5^{th} \\) of 1035.94\\( 3^{rd} \\) of 1033.53\\( 3^{rd} \\) of 1029.55\\( 2^{nd} \\) of 1035.32\\( 2^{nd} \\) of 101.69\\( 7^{th} \\) of 10
\\( FastDVDNet^† \\)[42]39.16\\( 5^{th} \\) of 1037.92\\( 6^{th} \\) of 1035.60\\( 5^{th} \\) of 1032.56\\( 5^{th} \\) of 1027.93\\( 6^{th} \\) of 1034.63\\( 5^{th} \\) of 105.72\\( 4^{th} \\) of 10
\\( TOFlow^† \\)[49]38.25\\( 8^{th} \\) of 1036.97\\( 8^{th} \\) of 1034.90\\( 7^{th} \\) of 1032.21\\( 6^{th} \\) of 1028.07\\( 5^{th} \\) of 1034.08\\( 7^{th} \\) of 102.84\\( 6^{th} \\) of 10
\\( BasicVSR++^† \\)[6]39.40\\( 4^{th} \\) of 1038.24\\( 2^{nd} \\) of 1035.55\\( 6^{th} \\) of 1031.72\\( 7^{th} \\) of 1026.78\\( 9^{th} \\) of 1034.34\\( 6^{th} \\) of 107.41\\( 2^{nd} \\) of 10
\\( VRT^† \\)[26]39.55\\( 2^{nd} \\) of 1038.12\\( 4^{th} \\) of 1034.82\\( 8^{th} \\) of 1030.77\\( 8^{th} \\) of 1026.01\\( 10^{th} \\) of 1033.86\\( 8^{th} \\) of 100.05\\( 10^{th} \\) of 10
Real-ESRGAN [46]29.98\\( 10^{th} \\) of 1028.27\\( 10^{th} \\) of 1028.04\\( 10^{th} \\) of 1027.52\\( 10^{th} \\) of 1027.63\\( 8^{th} \\) of 1028.29\\( 10^{th} \\) of 100.24\\( 8^{th} \\) of 10
UDVD [40]31.15\\( 9^{th} \\) of 1030.72\\( 9^{th} \\) of 1030.23\\( 9^{th} \\) of 1029.10\\( 9^{th} \\) of 1027.63\\( 7^{th} \\) of 1029.77\\( 9^{th} \\) of 100.16\\( 9^{th} \\) of 10
MF2F [14]39.09\\( 6^{th} \\) of 1038.20\\( 3^{rd} \\) of 10\\( \\underline{36.36} \\)\\( 1^{st} \\) of 1033.57\\( 2^{nd} \\) of 1029.04\\( 4^{th} \\) of 1035.25\\( 3^{rd} \\) of 104.62\\( 5^{th} \\) of 10
Ours - RFCVD40.35\\( 1^{st} \\) of 10\\( \\underline{38.60} \\)\\( 1^{st} \\) of 1036.28\\( 2^{nd} \\) of 1033.86\\( 1^{st} \\) of 1031.12\\( 1^{st} \\) of 1036.04\\( 1^{st} \\) of 1031.66\\( 1^{st} \\) of 10
", + "image_path": "320b47a166857f110f0538d369558864636bcdddf64de8327b2351dc396a3732.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 56, + 259, + 555, + 390 + ], + "blocks": [ + { + "bbox": [ + 56, + 236, + 555, + 258 + ], + "lines": [ + { + "bbox": [ + 56, + 236, + 555, + 258 + ], + "spans": [ + { + "bbox": [ + 56, + 236, + 555, + 258 + ], + "type": "text", + "content": "Table 2. Video denoising results on the CRVD (sRGB) benchmark. Not only does our approach perform best overall, it is also four times faster than the second-fastest place. Please kindly see the supplementary for SSIM and LPIPS where our approach ranks first as well." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 259, + 555, + 390 + ], + "lines": [ + { + "bbox": [ + 56, + 259, + 555, + 390 + ], + "spans": [ + { + "bbox": [ + 56, + 259, + 555, + 390 + ], + "type": "table", + "html": "
film grain noise w/ AV1 compressionspatially correlated noise w/ H.265 compressionSpeed
PSNR (higher PSNR is better)rank SSIM (higher SSIM is better)rank (lower LPIPS is better)LPIPS rank (lower LPIPS is better)rank (higher PSNR is better)rank (higher PSNR is better)SSIM (higher SSIM is better)rank (lower LPIPS is better)LPIPS rank (lower LPIPS is better)rank (higher FPS is better)FPS rank (higher FPS is better)
\\( SID^† \\)[7]27.05 5thof 70.664 5thof 70.293 4thof 728.44 3rdof 70.771 4thof 70.282 5thof 715.00 3rdof 7
\\( NAFNet^† \\)[9]27.16 4thof 70.672 4thof 70.294 5thof 728.40 4thof 70.764 6thof 70.257 3rdof 73.812 6thof 7
\\( FastDVDNet^† \\)[42]27.39 3rdof 70.686 3rdof 70.272 3rdof 727.92 6thof 70.769 5thof 70.296 6thof 712.09 4thof 7
\\( TOFlow^† \\)[49]28.15 2ndof 70.750 2ndof 70.220 1stof 728.39 5thof 70.788 3rdof 70.258 4thof 75.665 5thof 7
\\( BasicVSR++^† \\)[6]26.90 6thof 70.651 6thof 70.313 6thof 727.48 7thof 70.728 7thof 70.358 7thof 715.32 2ndof 7
\\( VRT^† \\)[26]26.55 7thof 70.629 7thof 70.331 7thof 728.78 2ndof 70.803 2ndof 70.206 1stof 70.105 7thof 7
Ours - RFCVD28.59 1stof 70.774 1stof 70.247 2ndof 728.93 1stof 70.808 1stof 70.239 2ndof 769.73 1stof 7
", + "image_path": "5f0b875e59c15221ca2dddd7f2ff9faad5de8a4c651f36750f682954679e99b6.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 392, + 555, + 415 + ], + "lines": [ + { + "bbox": [ + 55, + 392, + 555, + 415 + ], + "spans": [ + { + "bbox": [ + 55, + 392, + 555, + 415 + ], + "type": "text", + "content": "Table 3. Testing the generalizability when training on AWGN with H.264 transcoding but testing with two different degradation schemes. Since our approach is based on classic methods, the denoising itself only has a few parameters which naturally reduces domain gaps." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 435, + 295, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 435, + 295, + 521 + ], + "spans": [ + { + "bbox": [ + 54, + 435, + 295, + 521 + ], + "type": "text", + "content": "prisingly well, we have retrained all methods that neither leverage a sophisticated augmentation pipeline already nor utilize self-supervision. This makes for a more fair comparison since, as shown in Tab. 1, we have improved their performance on the CRVD benchmark with our retrained version. Throughout this paper, we denote all retrained models with a " + }, + { + "bbox": [ + 54, + 435, + 295, + 521 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 54, + 435, + 295, + 521 + ], + "type": "text", + "content": " to prevent any potential confusions." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 527, + 198, + 541 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 527, + 198, + 541 + ], + "spans": [ + { + "bbox": [ + 55, + 527, + 198, + 541 + ], + "type": "text", + "content": "4.1. Quantitative Comparison" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 545, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 545, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 54, + 545, + 295, + 715 + ], + "type": "text", + "content": "Please see Tab. 2 for a summary of the quantitative evaluation on the CRVD benchmark. In short, not only does our approach perform best overall, it is also four times faster than the second-fastest method. We attribute these favorable results to our utilization of classic denoising techniques which greatly helps with the ability to generalize to unseen noise patterns. Specifically, the denoising itself is driven by only a few parameters which naturally reduces overfitting and helps with bridging domain gaps. To test this hypothesis, we conducted an experiment with synthetic data where all models are trained on AWGN with H.264 transcoding but tested on two different degradation pipelines. As shown in Tab. 3, our approach performs favorably in this evaluation as well which supports our initial hypothesis." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 435, + 450, + 449 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 435, + 450, + 449 + ], + "spans": [ + { + "bbox": [ + 313, + 435, + 450, + 449 + ], + "type": "text", + "content": "4.2. Qualitative Comparison" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 454, + 555, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 454, + 555, + 563 + ], + "spans": [ + { + "bbox": [ + 313, + 454, + 555, + 563 + ], + "type": "text", + "content": "To test video denoising in the real world, we reached out to various creative professionals who kindly shared footage where they had trouble removing the noise. We show a representative excerpt of the denoising results on this footage in Fig. 6 but kindly refer to the supplementary which includes video results of all methods. Like with the quantitative evaluation, we have found that our approach performs favorably, and once again we hypothesize that this is due to the underlying utilization of classic denoising techniques." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 573, + 440, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 573, + 440, + 586 + ], + "spans": [ + { + "bbox": [ + 313, + 573, + 440, + 586 + ], + "type": "text", + "content": "4.3. Ablative Experiments" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 593, + 555, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 593, + 555, + 700 + ], + "spans": [ + { + "bbox": [ + 313, + 593, + 555, + 700 + ], + "type": "text", + "content": "One of the main things we wondered ourselves was the performance with respect to the choice of anchor frame. After all, the noise profile of I-frames can be quite different from P- and B-frames. Nevertheless as shown in Tab. 4 (top), we have always used the first frame as the anchor throughout the experiments section and would have drawn the same conclusions with the middle or the last frame. And as shown in Tab. 4 (bottom), we have also found that " + }, + { + "bbox": [ + 313, + 593, + 555, + 700 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{cstsy}" + }, + { + "bbox": [ + 313, + 593, + 555, + 700 + ], + "type": "text", + "content": " helps with providing a slightly more consistent experience." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 325, + 701, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 701, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 325, + 701, + 553, + 714 + ], + "type": "text", + "content": "A key part of our integration of classic denoising tech" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2089" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 69, + 560, + 459 + ], + "blocks": [ + { + "bbox": [ + 56, + 69, + 560, + 459 + ], + "lines": [ + { + "bbox": [ + 56, + 69, + 560, + 459 + ], + "spans": [ + { + "bbox": [ + 56, + 69, + 560, + 459 + ], + "type": "image", + "image_path": "1df6be4b5a87c7f1cc7e60530297622ad88009c6ee6832c78a28aef66251d7fe.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 464, + 555, + 498 + ], + "lines": [ + { + "bbox": [ + 55, + 464, + 555, + 498 + ], + "spans": [ + { + "bbox": [ + 55, + 464, + 555, + 498 + ], + "type": "text", + "content": "Figure 6. Video denoising results from exemplary still frames. Due to space constraints, we only share the results from a representative subset of the methods that we compare to. Please kindly refer to the supplementary to find video results of all methods. We are grateful that Amit Zinman (top row), Robert Kjettrup (middle row), and an anonymous artist (last row) were willing to provide test footage." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 518, + 295, + 614 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 518, + 295, + 614 + ], + "spans": [ + { + "bbox": [ + 55, + 518, + 295, + 614 + ], + "type": "text", + "content": "niques into a machine learning pipeline is that we account for spatially-varying noise. To assess the importance of this design decision, we trained an ablation that only predicts a set of scalar denoising parameters. As expected and as shown in Tab. 6, this \"w/o spat. varying\" experiment exhibits a significant drop in quality. As shown in Fig. 7, this ablation produces much more blurry denoising results since it cannot easily account for signal-dependent noise." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "text", + "content": "To best support " + }, + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{P}(\\cdot ;\\theta)" + }, + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "text", + "content": " in predicting the denoising parameters, we not only provide it access to the (merged) input frames but also their gradients as well as a mask indicating whether or not a pixel in an aligned frame is valid. To test the effectiveness of this design, we trained two ablations where we refrained from providing one (\"w/o img. gradients\") or the other (\"w/o align. mask\") and found that, as shown in Tab. 6, this context is indeed beneficial." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 518, + 555, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 518, + 555, + 674 + ], + "spans": [ + { + "bbox": [ + 313, + 518, + 555, + 674 + ], + "type": "text", + "content": "We argue that analyzing the noise profile of a given image is a mixture of both low-level image processing and high-level semantics, because one first needs an understanding of what an image depicts before being able to go into the low-level details to discern between unintended noise and actual texture. For this reason, we leverage a classification backbone " + }, + { + "bbox": [ + 313, + 518, + 555, + 674 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 313, + 518, + 555, + 674 + ], + "type": "text", + "content": " in the hypernetwork that is predicting the noise profile " + }, + { + "bbox": [ + 313, + 518, + 555, + 674 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 518, + 555, + 674 + ], + "type": "text", + "content": ", and we trained an ablation without this backbone to evaluate this hypothesis. As intuitively expected and as shown in Tab. 6, this \"w/o backbone " + }, + { + "bbox": [ + 313, + 518, + 555, + 674 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 313, + 518, + 555, + 674 + ], + "type": "text", + "content": "\" experiment indeed exhibits a significant drop in quality. This drop is also visually exemplified in Fig. 7, where the lack of this backbone leads to the noise not fully being removed." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "text", + "content": "Lastly, we not only fine-tune the aforementioned backbone " + }, + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "text", + "content": " but also stabilize the training of the hypernetwork through NPA [35]. To analyze the effect of these measures," + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2090" + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 70, + 296, + 198 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 296, + 198 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 296, + 198 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 296, + 198 + ], + "type": "table", + "html": "
PSNRdelta (higher PSNR is better)SSIMdelta (higher SSIM is better)LPIPSdelta (lower LPIPS is better)
Avg. w/ Lcstsy35.95-0.9477-0.0757-
first frame ← Ours36.04+ 0.090.9472- 0.00050.0763+ 0.0006
middle frame35.92- 0.030.9480+ 0.00030.0753- 0.0003
last frame35.90- 0.060.9480+ 0.00030.0754- 0.0003
Avg. w/o Lcstsy35.86-0.9460-0.0754-
first frame35.80- 0.050.9454- 0.00050.0750- 0.0004
middle frame36.02+ 0.160.9473+ 0.00140.0753- 0.0001
last frame35.75- 0.110.9451- 0.00080.0759+ 0.0005
", + "image_path": "b4e0511ae9b8b9641516ea398d424619a5f80ad61d73e75cf0927089e7e8b9ca.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 56, + 251, + 296, + 307 + ], + "blocks": [ + { + "bbox": [ + 55, + 201, + 295, + 236 + ], + "lines": [ + { + "bbox": [ + 55, + 201, + 295, + 236 + ], + "spans": [ + { + "bbox": [ + 55, + 201, + 295, + 236 + ], + "type": "text", + "content": "Table 4. Denoising results on the CRVD benchmark with respect to the anchor frame choice. We find that our approach is relatively robust to this choice (top), partly thanks to " + }, + { + "bbox": [ + 55, + 201, + 295, + 236 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{cstsy}" + }, + { + "bbox": [ + 55, + 201, + 295, + 236 + ], + "type": "text", + "content": " (bottom)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 251, + 296, + 307 + ], + "lines": [ + { + "bbox": [ + 56, + 251, + 296, + 307 + ], + "spans": [ + { + "bbox": [ + 56, + 251, + 296, + 307 + ], + "type": "table", + "html": "
PSNRdelta (higher PSNR is better)SSIMdelta (higher SSIM is better)LPIPSdelta (lower LPIPS is better)
Ours36.04-0.9472-0.0763-
fine-tuned on GT36.75+ 0.700.9494+ 0.00220.0591- 0.0172
", + "image_path": "b20f714afce042ef99277613972615ff4b4c6ceba10de0d20f79561ffe5cec5d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 362, + 295, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 362, + 295, + 399 + ], + "spans": [ + { + "bbox": [ + 55, + 362, + 295, + 399 + ], + "type": "text", + "content": "we trained an ablation for the former (\"w/o fine-tuning " + }, + { + "bbox": [ + 55, + 362, + 295, + 399 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 55, + 362, + 295, + 399 + ], + "type": "text", + "content": "\" as well as the latter (\"w/o NPA\") and found that, as shown in Tab. 6, both contribute to the denoising quality." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 413, + 133, + 424 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 413, + 133, + 424 + ], + "spans": [ + { + "bbox": [ + 55, + 413, + 133, + 424 + ], + "type": "text", + "content": "4.4. Limitations" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 434, + 295, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 434, + 295, + 529 + ], + "spans": [ + { + "bbox": [ + 54, + 434, + 295, + 529 + ], + "type": "text", + "content": "While our noise profile " + }, + { + "bbox": [ + 54, + 434, + 295, + 529 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 54, + 434, + 295, + 529 + ], + "type": "text", + "content": " with the subsequent denoising parameter prediction from " + }, + { + "bbox": [ + 54, + 434, + 295, + 529 + ], + "type": "inline_equation", + "content": "\\mathcal{P}(\\cdot ;\\theta)" + }, + { + "bbox": [ + 54, + 434, + 295, + 529 + ], + "type": "text", + "content": " works reasonably well as demonstrated throughout the experiments section, it is far from perfect. Specifically and as shown in Tab. 5, we fine-tuned the predicted denoising parameters on the ground truth and found that this hypothetical oracle is able to improve our results by a rather significant margin. Nevertheless, we leave bridging this gap to future research." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 532, + 295, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 532, + 295, + 615 + ], + "spans": [ + { + "bbox": [ + 54, + 532, + 295, + 615 + ], + "type": "text", + "content": "In the case where the noise profile is entirely known a priori, such as when developing a denoiser for a specific camera model subject to a constrained encoder, any reasonable deep learning method would provide better results than our approach. And with a baked-in noise profile it would perhaps even be faster. However, such a method would in turn be limited in terms of robustness and control." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 617, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 617, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 617, + 295, + 714 + ], + "type": "text", + "content": "Lastly, and on a more practical note, deploying hypernetworks in end-user applications is not well supported by the common inference ecosystems yet. Specifically, popular libraries like CoreML, WinML, or OpenVINO expect the model weights to be fixed which is obviously not the case for hypernetworks. As such, while our approach is easy to implement within frameworks like PyTorch, it is not straightforward to deploy in production environments." + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 316, + 70, + 375, + 130 + ], + "blocks": [ + { + "bbox": [ + 316, + 70, + 375, + 130 + ], + "lines": [ + { + "bbox": [ + 316, + 70, + 375, + 130 + ], + "spans": [ + { + "bbox": [ + 316, + 70, + 375, + 130 + ], + "type": "image", + "image_path": "d617da301db3f1bdba758af444dbdfcab7b4df741dc9fd6d3e24380b2057bc9f.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 316, + 131, + 374, + 167 + ], + "blocks": [ + { + "bbox": [ + 316, + 131, + 374, + 167 + ], + "lines": [ + { + "bbox": [ + 316, + 131, + 374, + 167 + ], + "spans": [ + { + "bbox": [ + 316, + 131, + 374, + 167 + ], + "type": "image", + "image_path": "a6147f8108f4b894c00419764e36d1f746cdb23aecdc4fc353bba3b5aaaf48ff.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 316, + 168, + 374, + 227 + ], + "blocks": [ + { + "bbox": [ + 316, + 168, + 374, + 227 + ], + "lines": [ + { + "bbox": [ + 316, + 168, + 374, + 227 + ], + "spans": [ + { + "bbox": [ + 316, + 168, + 374, + 227 + ], + "type": "image", + "image_path": "180577504c3d21230b9c91032cf119b2c6a956c0fec859267a84e0865cb2eac1.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 335, + 232, + 354, + 241 + ], + "lines": [ + { + "bbox": [ + 335, + 232, + 354, + 241 + ], + "spans": [ + { + "bbox": [ + 335, + 232, + 354, + 241 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 375, + 70, + 434, + 130 + ], + "blocks": [ + { + "bbox": [ + 375, + 70, + 434, + 130 + ], + "lines": [ + { + "bbox": [ + 375, + 70, + 434, + 130 + ], + "spans": [ + { + "bbox": [ + 375, + 70, + 434, + 130 + ], + "type": "image", + "image_path": "eaa422a565e2b480918ec025a8b466b88925dc1d38c01191b6cec38c30e3cf6e.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 375, + 131, + 434, + 167 + ], + "blocks": [ + { + "bbox": [ + 375, + 131, + 434, + 167 + ], + "lines": [ + { + "bbox": [ + 375, + 131, + 434, + 167 + ], + "spans": [ + { + "bbox": [ + 375, + 131, + 434, + 167 + ], + "type": "image", + "image_path": "5c34871bb5c19595625021fa1f101a9261a4c10068a1a26b6b041902c7c096c2.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 375, + 168, + 434, + 227 + ], + "blocks": [ + { + "bbox": [ + 375, + 168, + 434, + 227 + ], + "lines": [ + { + "bbox": [ + 375, + 168, + 434, + 227 + ], + "spans": [ + { + "bbox": [ + 375, + 168, + 434, + 227 + ], + "type": "image", + "image_path": "22a767d79f8d800880a9161affb4f0ab1bdd976f843f4c3315c2daed2f8b1ef4.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 376, + 232, + 433, + 241 + ], + "lines": [ + { + "bbox": [ + 376, + 232, + 433, + 241 + ], + "spans": [ + { + "bbox": [ + 376, + 232, + 433, + 241 + ], + "type": "text", + "content": "w/o spat. varying" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 435, + 70, + 493, + 130 + ], + "blocks": [ + { + "bbox": [ + 435, + 70, + 493, + 130 + ], + "lines": [ + { + "bbox": [ + 435, + 70, + 493, + 130 + ], + "spans": [ + { + "bbox": [ + 435, + 70, + 493, + 130 + ], + "type": "image", + "image_path": "96bc151ab373809c214ffbbff0e65f4ac71871d1fe2f61295b4e34343477b2a1.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 435, + 131, + 493, + 167 + ], + "blocks": [ + { + "bbox": [ + 435, + 131, + 493, + 167 + ], + "lines": [ + { + "bbox": [ + 435, + 131, + 493, + 167 + ], + "spans": [ + { + "bbox": [ + 435, + 131, + 493, + 167 + ], + "type": "image", + "image_path": "59f2eaae540e41a729af029582b29ef35a405c799151bd9f7ea48c7ef3299880.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 435, + 168, + 493, + 227 + ], + "blocks": [ + { + "bbox": [ + 435, + 168, + 493, + 227 + ], + "lines": [ + { + "bbox": [ + 435, + 168, + 493, + 227 + ], + "spans": [ + { + "bbox": [ + 435, + 168, + 493, + 227 + ], + "type": "image", + "image_path": "95a7282e092419b8effe5e0f9ff2b0152145fd6b4fddeccbbc5ce563b8d9761b.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 437, + 232, + 492, + 241 + ], + "lines": [ + { + "bbox": [ + 437, + 232, + 492, + 241 + ], + "spans": [ + { + "bbox": [ + 437, + 232, + 492, + 241 + ], + "type": "text", + "content": "w/o backbone " + }, + { + "bbox": [ + 437, + 232, + 492, + 241 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 495, + 70, + 553, + 130 + ], + "blocks": [ + { + "bbox": [ + 495, + 70, + 553, + 130 + ], + "lines": [ + { + "bbox": [ + 495, + 70, + 553, + 130 + ], + "spans": [ + { + "bbox": [ + 495, + 70, + 553, + 130 + ], + "type": "image", + "image_path": "b101c90257dcf2a870cfee103439369810981e8bdc94895ed837dc4763e388cb.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 495, + 131, + 553, + 167 + ], + "blocks": [ + { + "bbox": [ + 495, + 131, + 553, + 167 + ], + "lines": [ + { + "bbox": [ + 495, + 131, + 553, + 167 + ], + "spans": [ + { + "bbox": [ + 495, + 131, + 553, + 167 + ], + "type": "image", + "image_path": "fec91e0d23dbf73ac3a7bc97da582be1d81334a9498716733cceb39f243ac12f.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 495, + 168, + 553, + 227 + ], + "blocks": [ + { + "bbox": [ + 495, + 168, + 553, + 227 + ], + "lines": [ + { + "bbox": [ + 495, + 168, + 553, + 227 + ], + "spans": [ + { + "bbox": [ + 495, + 168, + 553, + 227 + ], + "type": "image", + "image_path": "17345923f8d9819f678594c7a1db336f4bfd0602988913cf9edab0f8723f18c3.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 515, + 232, + 533, + 240 + ], + "lines": [ + { + "bbox": [ + 515, + 232, + 533, + 240 + ], + "spans": [ + { + "bbox": [ + 515, + 232, + 533, + 240 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 246, + 553, + 269 + ], + "lines": [ + { + "bbox": [ + 313, + 246, + 553, + 269 + ], + "spans": [ + { + "bbox": [ + 313, + 246, + 553, + 269 + ], + "type": "text", + "content": "Figure 7. Visual comparison of the two most significant ablations on a representative sample from the CRVD dataset." + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "type": "table", + "bbox": [ + 315, + 275, + 554, + 388 + ], + "blocks": [ + { + "bbox": [ + 55, + 309, + 295, + 343 + ], + "lines": [ + { + "bbox": [ + 55, + 309, + 295, + 343 + ], + "spans": [ + { + "bbox": [ + 55, + 309, + 295, + 343 + ], + "type": "text", + "content": "Table 5. Fine-tuning our estimated denoising parameters on the ground truth of the CRVD benchmark indicates that our estimates are good but not perfect if we had access to an oracle." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 275, + 554, + 388 + ], + "lines": [ + { + "bbox": [ + 315, + 275, + 554, + 388 + ], + "spans": [ + { + "bbox": [ + 315, + 275, + 554, + 388 + ], + "type": "table", + "html": "
PSNRdelta (higher PSNR is better)SSIMdelta (higher SSIM is better)LPIPSdelta (lower LPIPS is better)
Ours36.04-0.9472-0.0763-
w/o spat. varying33.62-2.420.9359-0.01120.1184+0.0421
w/o img. gradients35.61-0.440.9469-0.00030.0756-0.0007
w/o align. mask35.26-0.780.9468-0.00030.0740-0.0023
w/o backbone B34.38-1.670.9059-0.04120.1444+0.0680
w/o fine-tuning B35.71-0.340.9464-0.00080.0732-0.0032
w/o NPA [35]35.83-0.210.9453-0.00180.0814+0.0051
", + "image_path": "7e1bf30bde2f724afe6bef1e87ea8ff40953b1f2b7e7d72cfa3f146dc0b9a33f.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "table_body" + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 390, + 554, + 412 + ], + "lines": [ + { + "bbox": [ + 313, + 390, + 554, + 412 + ], + "spans": [ + { + "bbox": [ + 313, + 390, + 554, + 412 + ], + "type": "text", + "content": "Table 6. Various ablative experiments, please see Section 4.3 for more context and details of each individual ablation." + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 429, + 388, + 441 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 429, + 388, + 441 + ], + "spans": [ + { + "bbox": [ + 313, + 429, + 388, + 441 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 312, + 449, + 555, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 449, + 555, + 568 + ], + "spans": [ + { + "bbox": [ + 312, + 449, + 555, + 568 + ], + "type": "text", + "content": "We show that traditional denoising techniques are still very much relevant in the modern machine learning world. Not only have we found them to be robust and fast, but also controllable which is an important property in many video denoising applications where creative professionals want to assert their artistic expression. To bridge the gap between classic denoising and modern approaches, we train a model to estimate the various parameters of a traditional denoising approach for a given input video, which is otherwise not only tedious but also requires a certain amount of skill." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 312, + 570, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 570, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 312, + 570, + 555, + 713 + ], + "type": "text", + "content": "Furthermore, in an effort to make best use of the computational efficiency of classic denoising techniques, our method separates the analysis of the noise from the denoising in order to avoid having to estimate the noise profile over and over again as is typical for machine learning approaches to video denoising. Lastly, we have found a simple yet effective degradation pipeline for training video denoiseers based on AWGN subject to H.264 video transcoding. This pipeline ensures that the noise is temporally correlated just like it often is the case in the real world, and we have found it not only to work well for our approach but we have also been able to retrain and improve existing models." + } + ] + } + ], + "index": 30 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "2091" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 180 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 180 + ], + "type": "text", + "content": "Acknowledgement. This work was supported in part by the National Natural Science Foundation of China (62306153, 62225604), the Fundamental Research Funds for the Central Universities (Nankai University, 070-63243143), the Natural Science Foundation of Tianjin, China (24JCJQJC00020), and Shenzhen Science and Technology Program (JCYJ20240813114237048). The computational devices is partly supported by the Supercomputing Center of Nankai University (NKSC)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 192, + 115, + 205 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 192, + 115, + 205 + ], + "spans": [ + { + "bbox": [ + 56, + 192, + 115, + 205 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 211, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 61, + 211, + 296, + 245 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 211, + 296, + 245 + ], + "spans": [ + { + "bbox": [ + 61, + 211, + 296, + 245 + ], + "type": "text", + "content": "[1] Eric P Bennett and Leonard McMillan. Video enhancement using per-pixel virtual exposures. In ACM SIGGRAPH 2005 Papers, 2005. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 246, + 296, + 280 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 246, + 296, + 280 + ], + "spans": [ + { + "bbox": [ + 61, + 246, + 296, + 280 + ], + "type": "text", + "content": "[2] Goutam Bhat, Michael Gharbi, Jiawen Chen, Luc Van Gool, and Zhihao Xia. Self-supervised burst super-resolution. In ICCV, pages 10571-10580. IEEE, 2023. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 280, + 296, + 313 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 280, + 296, + 313 + ], + "spans": [ + { + "bbox": [ + 61, + 280, + 296, + 313 + ], + "type": "text", + "content": "[3] Peter J Burt and Edward H Adelson. The laplacian pyramid as a compact image code. In Readings in computer vision, 1987. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 315, + 296, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 315, + 296, + 346 + ], + "spans": [ + { + "bbox": [ + 62, + 315, + 296, + 346 + ], + "type": "text", + "content": "[4] Jiezhang Cao, Yawei Li, Kai Zhang, and Luc Van Gool. Video super-resolution transformer. arXiv:2106.06847, 2021. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 349, + 296, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 349, + 296, + 392 + ], + "spans": [ + { + "bbox": [ + 62, + 349, + 296, + 392 + ], + "type": "text", + "content": "[5] Kelvin CK Chan, Xintao Wang, Ke Yu, Chao Dong, and Chen Change Loy. Basicvsr: The search for essential components in video super-resolution and beyond. In CVPR, 2021. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 393, + 296, + 438 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 393, + 296, + 438 + ], + "spans": [ + { + "bbox": [ + 62, + 393, + 296, + 438 + ], + "type": "text", + "content": "[6] Kelvin CK Chan, Shangchen Zhou, Xiangyu Xu, and Chen Change Loy. Basicvsr++: Improving video superresolution with enhanced propagation and alignment. In CVPR, 2022. 2, 5, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 440, + 296, + 462 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 440, + 296, + 462 + ], + "spans": [ + { + "bbox": [ + 62, + 440, + 296, + 462 + ], + "type": "text", + "content": "[7] Chen Chen, Qifeng Chen, Jia Xu, and Vladlen Koltun. Learning to see in the dark. In CVPR, 2018. 2, 5, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 463, + 296, + 495 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 463, + 296, + 495 + ], + "spans": [ + { + "bbox": [ + 62, + 463, + 296, + 495 + ], + "type": "text", + "content": "[8] Jiawen Chen, Sylvain Paris, and Frédo Durand. Real-time edge-aware image processing with the bilateral grid. ACM TOG, 2007. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 497, + 296, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 497, + 296, + 529 + ], + "spans": [ + { + "bbox": [ + 62, + 497, + 296, + 529 + ], + "type": "text", + "content": "[9] Liangyu Chen, Xiaojie Chu, Xiangyu Zhang, and Jian Sun. Simple baselines for image restoration. In ECCV, 2022. 2, 5, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 531, + 296, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 531, + 296, + 565 + ], + "spans": [ + { + "bbox": [ + 57, + 531, + 296, + 565 + ], + "type": "text", + "content": "[10] Kostadin Dabov, Alessandro Foi, Vladimir Katkovnik, and Karen Egiazarian. Image denoising with block-matching and 3d filtering. In Image Processing, 2006. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 566, + 296, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 566, + 296, + 599 + ], + "spans": [ + { + "bbox": [ + 57, + 566, + 296, + 599 + ], + "type": "text", + "content": "[11] Kostadin Dabov, Alessandro Foi, and Karen Egiazarian. Video denoising by sparse 3d transform domain collaborative filtering. EURASIP, 2007." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 600, + 296, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 600, + 296, + 634 + ], + "spans": [ + { + "bbox": [ + 56, + 600, + 296, + 634 + ], + "type": "text", + "content": "[12] Kostadin Dabov, Alessandro Foi, Vladimir Katkovnik, and Karen Egiazarian. Image denoising by sparse 3-d transform-domain collaborative filtering. IEEE TIP, 2007. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 635, + 296, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 296, + 667 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 296, + 667 + ], + "type": "text", + "content": "[13] Axel Davy, Thibaud Ehret, Jean-Michel Morel, Pablo Arias, and Gabriele Facciolo. A non-local cnn for video denoising. In ICIP, 2019. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 669, + 296, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 669, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 669, + 296, + 712 + ], + "type": "text", + "content": "[14] Valery Dewil, Jérémy Anger, Axel Davy, Thibaud Ehret, Gabriele Facciolo, and Pablo Arias. Self-supervised training for blind multi-frame video denoising. In WACV, 2021. 1, 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 554, + 712 + ], + "type": "list", + "angle": 0, + "index": 34, + "blocks": [ + { + "bbox": [ + 316, + 72, + 554, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 554, + 115 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 554, + 115 + ], + "type": "text", + "content": "[15] Charles Dugas, Yoshua Bengio, François Bélisle, Claude Nadeau, and René Garcia. Incorporating second-order functional knowledge for better option pricing. In NIPS, 2000. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 118, + 554, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 554, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 554, + 150 + ], + "type": "text", + "content": "[16] Jana Ehmann, Lun-Cheng Chu, Sung-Fang Tsai, and Chia-Kai Liang. Real-time video denoising on mobile phones. In ICIP, 2018. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 152, + 554, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 554, + 186 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 554, + 186 + ], + "type": "text", + "content": "[17] Thibaud Ehret, Axel Davy, Jean-Michel Morel, Gabriele Facciolo, and Pablo Arias. Model-blind video denoising via frame-to-frame training. In CVPR, 2019. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 186, + 554, + 209 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 186, + 554, + 209 + ], + "spans": [ + { + "bbox": [ + 316, + 186, + 554, + 209 + ], + "type": "text", + "content": "[18] Ruturaj G Gavaskar and Kunal N Chaudhury. Fast adaptive bilateral filtering. IEEE TIP, 2018. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 211, + 554, + 232 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 211, + 554, + 232 + ], + "spans": [ + { + "bbox": [ + 316, + 211, + 554, + 232 + ], + "type": "text", + "content": "[19] David Ha, Andrew M. Dai, and Quoc V. Le. Hypernetworks. arXiv:1609.09106, 2016. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 235, + 554, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 235, + 554, + 289 + ], + "spans": [ + { + "bbox": [ + 316, + 235, + 554, + 289 + ], + "type": "text", + "content": "[20] Samuel W Hasinoff, Dillon Sharlet, Ryan Geiss, Andrew Adams, Jonathan T Barron, Florian Kainz, Jiawen Chen, and Marc Levoy. Burst photography for high dynamic range and low-light imaging on mobile cameras. ACM TOG, 2016. 2, 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 291, + 554, + 324 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 291, + 554, + 324 + ], + "spans": [ + { + "bbox": [ + 316, + 291, + 554, + 324 + ], + "type": "text", + "content": "[21] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Delving deep into rectifiers: Surpassing human-level performance on imagenet classification. In ICCV, 2015. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 325, + 554, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 325, + 554, + 347 + ], + "spans": [ + { + "bbox": [ + 316, + 325, + 554, + 347 + ], + "type": "text", + "content": "[22] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR, 2015. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 349, + 554, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 349, + 554, + 381 + ], + "spans": [ + { + "bbox": [ + 316, + 349, + 554, + 381 + ], + "type": "text", + "content": "[23] Samuli Laine, Tero Karras, Jaakko Lehtinen, and Timo Aila. High-quality self-supervised deep image denoising. In NeurIPS, 2019. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 384, + 554, + 417 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 384, + 554, + 417 + ], + "spans": [ + { + "bbox": [ + 316, + 384, + 554, + 417 + ], + "type": "text", + "content": "[24] Seunghwan Lee, Donghyeon Cho, Jiwon Kim, and Tae Hyun Kim. Restore from restored: Video restoration with pseudo clean video. In CVPR, 2021. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 418, + 554, + 461 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 418, + 554, + 461 + ], + "spans": [ + { + "bbox": [ + 316, + 418, + 554, + 461 + ], + "type": "text", + "content": "[25] Jaakko Lehtinen, Jacob Munkberg, Jon Hasselgren, Samuli Laine, Tero Karras, Miika Aittala, and Timo Aila. Noise2noise: Learning image restoration without clean data. In ICML, 2018. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 464, + 554, + 507 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 464, + 554, + 507 + ], + "spans": [ + { + "bbox": [ + 316, + 464, + 554, + 507 + ], + "type": "text", + "content": "[26] Jingyun Liang, Jiezhang Cao, Yuchen Fan, Kai Zhang, Rakesh Ranjan, Yawei Li, Radu Timofte, and Luc Van Gool. Vrt: A video restoration transformer. IEEE TIP, 2024. 1, 2, 5, 6, 7" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 509, + 554, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 509, + 554, + 553 + ], + "spans": [ + { + "bbox": [ + 316, + 509, + 554, + 553 + ], + "type": "text", + "content": "[27] Orly Liba, Kiran Murthy, Yun-Ta Tsai, Tim Brooks, Tianfan Xue, Nikhil Karnad, Qiurui He, Jonathan T Barron, Dillon Sharlet, Ryan Geiss, et al. Handheld mobile photography in very low light. ACM TOG, 2019. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 555, + 554, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 555, + 554, + 587 + ], + "spans": [ + { + "bbox": [ + 316, + 555, + 554, + 587 + ], + "type": "text", + "content": "[28] Ce Liu and William T Freeman. A high-quality video denoising algorithm based on reliable motion estimation. In ECCV, 2010. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 589, + 554, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 589, + 554, + 633 + ], + "spans": [ + { + "bbox": [ + 316, + 589, + 554, + 633 + ], + "type": "text", + "content": "[29] Ding Liu, Zhaowen Wang, Yuchen Fan, Xianming Liu, Zhangyang Wang, Shiyu Chang, and Thomas Huang. Robust video super-resolution with learned temporal dynamics. In ICCV, 2017. 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 316, + 635, + 554, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 554, + 667 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 554, + 667 + ], + "type": "text", + "content": "[30] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In CVPR, 2022. 4, 5" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 316, + 669, + 554, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 669, + 554, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 669, + 554, + 712 + ], + "type": "text", + "content": "[31] Matteo Maggioni, Vladimir Katkovnik, Karen Egiazarian, and Alessandro Foi. Nonlocal transform-domain filter for volumetric data denoising and reconstruction. IEEE TIP, 2012. 2" + } + ] + } + ], + "index": 33 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "2092" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 115 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 115 + ], + "type": "text", + "content": "[32] Matteo Maggioni, Yibin Huang, Cheng Li, Shuai Xiao, Zhongqian Fu, and Fenglong Song. Efficient multi-stage video denoising with recurrent spatio-temporal fusion. In CVPR, 2021. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 162 + ], + "type": "text", + "content": "[33] Seungjun Nah, Sungyong Baik, Seokil Hong, Gyeongsik Moon, Sanghyun Son, Radu Timofte, and Kyoung Mu Lee. NTIRE 2019 challenge on video deblurring and superresolution: Dataset and study. In CVPRW, 2019. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 164, + 294, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 164, + 294, + 186 + ], + "spans": [ + { + "bbox": [ + 57, + 164, + 294, + 186 + ], + "type": "text", + "content": "[34] Alasdair Newson, Noura Faraj, Bruno Galerne, and Julie Delon. Realistic film grain rendering. IPOL, 2017. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 188, + 295, + 220 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 188, + 295, + 220 + ], + "spans": [ + { + "bbox": [ + 57, + 188, + 295, + 220 + ], + "type": "text", + "content": "[35] Jose Javier Gonzalez Ortiz, John V. Guttag, and Adrian V. Dalca. Non-proportional parametrizations for stable hypernetwork learning. arXiv:2304.07645, 2023. 3, 4, 7, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 57, + 222, + 294, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 222, + 294, + 263 + ], + "spans": [ + { + "bbox": [ + 57, + 222, + 294, + 263 + ], + "type": "text", + "content": "[36] Sylvain Paris, Pierre Kornprobst, Jack Tumblin, Frédo Durand, et al. Bilateral filtering: Theory and applications. Foundations and Trends® in Computer Graphics and Vision, 2009. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 57, + 267, + 294, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 267, + 294, + 311 + ], + "spans": [ + { + "bbox": [ + 57, + 267, + 294, + 311 + ], + "type": "text", + "content": "[37] Andrea Petreto, Thomas Romera, Florian Lemaitre, Ian Masliah, Boris Gaillard, Manuel Bouyer, Quentin L Meunier, and Lionel Lacassagne. A new real-time embedded video denoising algorithm. In DASIP, 2019. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 57, + 313, + 294, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 313, + 294, + 334 + ], + "spans": [ + { + "bbox": [ + 57, + 313, + 294, + 334 + ], + "type": "text", + "content": "[38] Anurag Ranjan and Michael J. Black. Optical flow estimation using a spatial pyramid network. In CVPR, 2017. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 57, + 336, + 294, + 368 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 336, + 294, + 368 + ], + "spans": [ + { + "bbox": [ + 57, + 336, + 294, + 368 + ], + "type": "text", + "content": "[39] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In MICCAI, 2015. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 57, + 371, + 294, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 371, + 294, + 414 + ], + "spans": [ + { + "bbox": [ + 57, + 371, + 294, + 414 + ], + "type": "text", + "content": "[40] Dev Yashpal Sheth, Sreyas Mohan, Joshua L Vincent, Ramon Manzorro, Peter A Crozier, Mitesh M Khapra, Eero P Simoncelli, and Carlos Fernandez-Granda. Unsupervised deep video denoising. In ICCV, 2021. 1, 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 57, + 416, + 294, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 416, + 294, + 437 + ], + "spans": [ + { + "bbox": [ + 57, + 416, + 294, + 437 + ], + "type": "text", + "content": "[41] Maitreya Suin and AN Rajagopalan. Gated spatio-temporal attention-guided video deblurring. In CVPR, 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 440, + 294, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 440, + 294, + 472 + ], + "spans": [ + { + "bbox": [ + 57, + 440, + 294, + 472 + ], + "type": "text", + "content": "[42] Matias Tassano, Julie Delon, and Thomas Veit. Fastdvnet: Towards real-time deep video denoising without flow estimation. In CVPR, 2020. 2, 5, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 474, + 294, + 495 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 474, + 294, + 495 + ], + "spans": [ + { + "bbox": [ + 57, + 474, + 294, + 495 + ], + "type": "text", + "content": "[43] Carlo Tomasi and Roberto Manduchi. Bilateral filtering for gray and color images. In ICCV, 1998. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 498, + 294, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 498, + 294, + 529 + ], + "spans": [ + { + "bbox": [ + 57, + 498, + 294, + 529 + ], + "type": "text", + "content": "[44] Gregory Vaksman, Michael Elad, and Peyman Milanfar. Patch craft: Video denoising by deep modeling and patch matching. In ICCV, 2021. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 532, + 294, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 532, + 294, + 563 + ], + "spans": [ + { + "bbox": [ + 57, + 532, + 294, + 563 + ], + "type": "text", + "content": "[45] Xintao Wang, Kelvin CK Chan, Ke Yu, Chao Dong, and Chen Change Loy. Edvr: Video restoration with enhanced deformable convolutional networks. In CVPRW, 2019. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 567, + 294, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 567, + 294, + 599 + ], + "spans": [ + { + "bbox": [ + 57, + 567, + 294, + 599 + ], + "type": "text", + "content": "[46] Xintao Wang, Liangbin Xie, Chao Dong, and Ying Shan. Real-esrgan: Training real-world blind super-resolution with pure synthetic data. In ICCV Workshop, 2021. 1, 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 601, + 294, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 601, + 294, + 633 + ], + "spans": [ + { + "bbox": [ + 57, + 601, + 294, + 633 + ], + "type": "text", + "content": "[47] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE TIP, 2004. 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 57, + 635, + 294, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 635, + 294, + 677 + ], + "spans": [ + { + "bbox": [ + 57, + 635, + 294, + 677 + ], + "type": "text", + "content": "[48] Bartlomiej Wronski, Ignacio Garcia-Dorado, Manfred Ernst, Damien Kelly, Michael Krainin, Chia-Kai Liang, Marc Levoy, and Peyman Milanfar. Handheld multi-frame superresolution. ACM TOG, 2019. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 57, + 681, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 681, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 681, + 294, + 712 + ], + "type": "text", + "content": "[49] Tianfan Xue, Baian Chen, Jiajun Wu, Donglai Wei, and William T Freeman. Video enhancement with task-oriented flow. IJCV, 2019. 2, 5, 6" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 241 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "text", + "content": "[50] Huanjing Yue, Cong Cao, Lei Liao, Ronghe Chu, and Jingyu Yang. Supervised raw video denoising with a benchmark dataset on dynamic scenes. In CVPR, 2020. 2, 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 107, + 553, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 107, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 317, + 107, + 553, + 140 + ], + "type": "text", + "content": "[51] Kai Zhang, Jingyun Liang, Luc Van Gool, and Radu Timofte. Designing a practical degradation model for deep blind image super-resolution. In CVPR, 2021. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 141, + 553, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 141, + 553, + 184 + ], + "spans": [ + { + "bbox": [ + 317, + 141, + 553, + 184 + ], + "type": "text", + "content": "[52] Kai Zhang, Yawei Li, Jingyun Liang, Jiezhang Cao, Yu-lun Zhang, Hao Tang, Deng-Ping Fan, Radu Timofte, and Luc Van Gool. Practical blind image denoising via swim-conv-unet and data synthesis. MIR, 2023. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 186, + 553, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 186, + 553, + 207 + ], + "spans": [ + { + "bbox": [ + 317, + 186, + 553, + 207 + ], + "type": "text", + "content": "[53] Ming Zhang and Bahadir K. Gunturk. Multiresolution bilateral filtering for image denoising. IEEE TIP, 2008. 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 209, + 553, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 209, + 553, + 241 + ], + "spans": [ + { + "bbox": [ + 317, + 209, + 553, + 241 + ], + "type": "text", + "content": "[54] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 5" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2093" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Classifier-Free Guidance Inside the Attraction Basin May Cause Memorization/a1a30dde-51b7-469a-a404-0c4a54031538_content_list.json b/2025/Classifier-Free Guidance Inside the Attraction Basin May Cause Memorization/a1a30dde-51b7-469a-a404-0c4a54031538_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..68b3945cd8d74b062d186352c5c0ec7d3acc801b --- /dev/null +++ b/2025/Classifier-Free Guidance Inside the Attraction Basin May Cause Memorization/a1a30dde-51b7-469a-a404-0c4a54031538_content_list.json @@ -0,0 +1,1468 @@ +[ + { + "type": "text", + "text": "Classifier-Free Guidance inside the Attraction Basin May Cause Memorization", + "text_level": 1, + "bbox": [ + 101, + 130, + 893, + 151 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Anubhav Jain1*, Yuya Kobayashi2, Takashi Shibuya2, Yuhta Takida2, Nasir Memon1, Julian Togelius1, Yuki Mitsufuji2,3", + "bbox": [ + 228, + 179, + 767, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ New York University, $^{2}$ Sony AI, $^{3}$ Sony Group Corporation", + "bbox": [ + 263, + 215, + 733, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{aj3281,memon,julian.togelius}@nyu.edu", + "bbox": [ + 310, + 234, + 684, + 251 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{u.kobayashi,takashi.tak.shibuya,yuta.takida,yuhki.mitsufuji}@sony.com", + "bbox": [ + 150, + 252, + 839, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 304, + 326, + 319 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Diffusion models are prone to exactly reproduce images from the training data. This exact reproduction of the training data is concerning as it can lead to copyright infringement and/or leakage of privacy-sensitive information. In this paper, we present a novel perspective on the memorization phenomenon and propose a simple yet effective approach to mitigate it. We argue that memorization occurs because of an attraction basin in the denoising process which steers the diffusion trajectory towards a memorized image. However, this can be mitigated by guiding the diffusion trajectory away from the attraction basin by not applying classifier-free guidance until an ideal transition point occurs from which classifier-free guidance is applied. This leads to the generation of non-memorized images that are high in image quality and well-aligned with the conditioning mechanism. To further improve on this, we present a new guidance technique, opposite guidance, that escapes the attraction basin sooner in the denoising process. We demonstrate the existence of attraction basins in various scenarios in which memorization occurs, and we show that our proposed approach successfully mitigates memorization. Our codebase is publicly available at https://github.com/SonyResearch/mitigating_memorization.", + "bbox": [ + 88, + 337, + 485, + 684 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 714, + 220, + 729 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent advancements in image-generation models have enabled the generation of hyper-realistic images. Diffusion models [21, 25] have been at the forefront of this advancement with their ability to align image generation with different conditioning mechanisms such as text giving users more control over the output. However, a key challenge has emerged: these models can memorize and reproduce training examples verbatim (verbatim memorization) [4, 23, 24], leading to concerns over copyright infringement and leak", + "bbox": [ + 89, + 739, + 485, + 876 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f959491e85289c4f63b35fd8f3ab1bad326e94efd43f273c778e7621afd9b104.jpg", + "image_caption": [ + "Figure 1. The diffusion trajectory contains an attraction basin (red region) which steers conditioned samples towards their memorized images. It can be avoided by applying zero classifier-free guidance when the trajectory is inside the attraction basin, such that there is an ideal transition point $\\tau^{*}$ after which applying CFG leads to non-memorized output. Applying CFG at an earlier point such as $\\tau_{1}$ inside the attraction basin leads to the same memorized sample." + ], + "image_footnote": [], + "bbox": [ + 522, + 308, + 901, + 502 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "age of privacy sensitive information. Alternatively, the model can generate images copying the composition and structure present in one or more training images (template memorization)[20]. Researchers have identified factors contributing to memorization and reproduction, such as the duplication of training images and repeated captions [24]. However, not all causes are understood, as memorization can still occur even when these issues are avoided. Therefore, there is a growing need for effective mitigation strategies, in particular methods that can be applied during inference without requiring computationally intensive retraining of the core diffusion models.", + "bbox": [ + 511, + 638, + 908, + 820 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In text-to-image (T2I) diffusion models, strong associations between a text prompt and an image in the training dataset lead to memorization. This can occur due to the presence of \"trigger tokens\" in overly specific text-prompts [20, 28], the model seeing multiple occurrences of these dur", + "bbox": [ + 511, + 825, + 908, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Work done during an internship at Sony AI.", + "bbox": [ + 107, + 886, + 346, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "12871", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d81ca9c631ca872b93e29fdcc7c5c867306e0a9a4ee22644eea158e8cadf89fa.jpg", + "image_caption": [ + "(a) Prompt: \"Here's What You Need to Know About St. Vincent's Apple Music Radio Show\"" + ], + "image_footnote": [], + "bbox": [ + 96, + 89, + 488, + 268 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/3c89e43d47393742467f6e21b22d3314396f0c775e181c7f42959f7e3c64e3c2.jpg", + "image_caption": [ + "(b) Prompt: \"Here's What You Need to Know About St. Vincent's Apple Music Radio Show\"", + "Figure 2. Plots showing magnitude of $\\epsilon_{\\theta}(x_t, e_p) - \\epsilon_{\\theta}(x_t, e_{\\emptyset})$ when denoising without classifier-free guidance (CFG) at each time step. The figures show the generated image if you start applying CFG at that time step. We get non-memorized output if we apply CFG after the ideal transition point $\\tau^*$ which coincides with the fall in the conditional noise prediction. This value is dependent on both the prompt and the initialization ((a) and (b) contain the same prompt with different initializations). More examples in the Appendix." + ], + "image_footnote": [], + "bbox": [ + 500, + 88, + 903, + 271 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ing training (data duplication) [24, 28], or other causes such as fine-tuning on smaller datasets [24]. Previous approaches have focused on perturbing either the text prompt/embeddings [24, 28] or the cross-attention between the text and image embeddings [20]. While these approaches help weaken the associations between prompts and corresponding images, they are limited to specific conditions in which they have been tested and fail to generalize effectively. Additionally, since these techniques work at the prompt level, they are only applicable to text-conditional generation, and also assume the ability to update the prompt.", + "bbox": [ + 88, + 388, + 485, + 555 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we introduce a new concept to explain the behavior of memorization in diffusion models based on our empirical observations. We first observe that applying classifier-free guidance (CFG) before a certain time step tends to produce memorized samples; we refer to this time step as the transition point. Interestingly, applying CFG after the transition point is unlikely to yield a memorized image. Based on these observations, we propose the concept of an attraction basin, in which applying CFG induces a strong steering force toward a memorized output, even with different initializations. Furthermore, following the approach of Wen et al. [28], we analyze the magnitude of the conditional noise predictions and discover uncharacteristically large values within the attraction basin, providing empirical evidence of a strong steering force in this region. Once the transition point is passed, the strong steering force is no longer pervasive. This magnitude-based analysis motivates us to detect the transition point by tracking the magnitude. Specifically, we initially apply no guidance until a transition point occurs—marked by a sharp drop in the discrepancy between conditional and unconditional guidance—after which applying CFG leads to non-memorized images.", + "bbox": [ + 88, + 569, + 486, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We summarize our contributions in this paper as follows:", + "bbox": [ + 529, + 388, + 906, + 404 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We present a new way to understand memorization in conditional diffusion models by understanding the diffusion trajectory from a dynamical systems theory perspective. We show that an attraction basin forms in the denoising trajectory.", + "- We present approaches to avoiding the attraction basin in various scenarios that incur no additional computational costs and require neither access to the prompt nor modification of weights.", + "- We present a new guidance technique that we call Opposite Guidance (not to be confused with negative prompts [3]) to push the trajectory away from the attraction basin.", + "- We comprehensively show that previous approaches are not generalizable in mitigating memorization and only work in certain scenarios studied in those papers. On the contrary, our approach is able to mitigate memorization in all the explored scenarios." + ], + "bbox": [ + 513, + 404, + 906, + 660 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 672, + 656, + 688 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Understanding Memorization", + "text_level": 1, + "bbox": [ + 511, + 696, + 779, + 713 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In deep generative models, when memorization occurs, a generated sample can match verbatim with a single or set of training images (verbatim memorization) [4, 23, 24] or copy the same template as the training image (template memorization) [20] leading to privacy and/or copyright infringement issues. Researchers [4, 23, 24, 27] have identified that when images or prompts are duplicated in the training dataset, it can lead to memorization. Additionally, training on smaller datasets has also been shown to be a major factor in causing memorization [24]. However, as Somepalli et al. [24] pointed out, even though data has been de-duplicated in the newer models [16] such as Stable Diffusion v2.1 (SDv2.1),", + "bbox": [ + 509, + 719, + 908, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "12872", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "the issue of memorization is yet to be resolved, and it does not explain much of the observed replication behavior. Factors such as conditioning and dataset complexity play a role even when the training datasets are large enough and are de-duplicated where \"simpler\" images and overly specific prompts have a higher chance of getting memorized [24]. Thus, there is a strong need for mitigation techniques beyond dataset manipulations.", + "bbox": [ + 89, + 90, + 483, + 210 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Detecting Memorization", + "text_level": 1, + "bbox": [ + 89, + 220, + 313, + 237 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Similarity scores based on the SSCD (self-supervised copy detection) model [18] can be used for getting an estimate on the extent of memorization wherein scores $>0.5$ can suggest that it is memorized [23]. However, this metric is prone to false positives when the textual similarity is high between a generated and real image [24]. Wen et al. [28] showed that for memorized prompts, the difference between the text-conditioned and the unconditioned score predictions is uncharacteristically high and can be informative in detecting memorization. They demonstrated high performance even in the first step of noise prediction. Ren et al. [20], on the other hand, showed that for memorized prompts, the cross-attention scores pertaining to certain \"trigger tokens\" in the memorized prompts are unusually high. Suggesting that they are given more importance, and thus lead to memorization. Other authors [4, 27] used a black box membership inference attack for finding memorized images. Daras et al. [8] showed that when an image is memorized, the model is able to fully reconstruct it from a noised version of it. This observation could be used for detection. Another line of research has tried to understand memorization by comparing diffusion models to associative memory networks [1, 11, 17].", + "bbox": [ + 89, + 244, + 483, + 578 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Mitigating Memorization", + "text_level": 1, + "bbox": [ + 89, + 587, + 321, + 603 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3.1. Training Time Mitigation", + "text_level": 1, + "bbox": [ + 89, + 609, + 313, + 625 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Wen et al. [28] proposed monitoring the text-conditioned noise prediction scores for each sample and excluding it from the current mini-batch if it surpasses a certain predetermined threshold. On similar lines, Ren et al. [20] proposed removing samples from the mini-batch when their cross-attention entropy is above a particular predetermined threshold. Somepalli et al. [24] proposed generating multiple BLIP captions per image during training. They also proposed perturbing the prompts by adding Gaussian noise or adding/replacing random words/numbers. Daras et al. [8] proposed training/finetuning on corrupted images can reduce instances of data replication by the model. Chavhan et al. [6] proposed pruning the UNet model weights assuming prior knowledge of memorized prompts.", + "bbox": [ + 89, + 628, + 483, + 839 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Most training time mitigations require re-training the base diffusion model from scratch. Not only is this computationally expensive, it also requires control over the entire training process including access to the training datasets.", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3.2. Inference Time Mitigation", + "text_level": 1, + "bbox": [ + 513, + 90, + 741, + 106 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Wen et al. [28] proposed perturbing the prompt embedding by minimizing the difference between the text-conditioned and unconditioned noise prediction scores such that it falls below a certain target loss. Ren et al. [20] proposed applying a mask and rescaling the attention scores to give less attention to \"trigger tokens\". However, this approach does not mitigate memorization when \"trigger tokens\" are not present. Somepalli et al. [24] proposed adding random tokens to the text prompt or repeating certain tokens during inference so as to add noise to the text embedding indirectly. Chen et al. [7] proposed different CFG weight schedulers to mitigate different causes of memorization. The approach presumes knowledge of not only whether memorization is present but also the type of memorization.", + "bbox": [ + 511, + 109, + 906, + 320 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Most of these approaches work by updating text prompt/embeddings leads to a trade-off between the text alignment and memorization. Another drawback of these approaches is that they work better when a trigger token is present such that we can reduce the importance/weightage of that token and they do not generalize well to other types of memorization as we show in this paper. Lastly, since they work directly on the text prompt/embeddings, they cannot be applied to other conditioning mechanisms such as class labels.", + "bbox": [ + 511, + 321, + 908, + 457 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Preliminaries", + "text_level": 1, + "bbox": [ + 513, + 468, + 651, + 484 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Diffusion models [25] such as Stable Diffusion (SD) [21] and Imagen [22] are trained with the objective of learning a model $\\epsilon_{\\theta}$ to denoise a noisy input vector at different levels of noise characterized by a time-dependent noise scheduler. We can define the diffusion process as a stochastic differential equation (SDE) in the form,", + "bbox": [ + 511, + 494, + 905, + 584 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nd x _ {t} = f (x _ {t}, t) d t + g (t) d w _ {t}.\n$$\n", + "text_format": "latex", + "bbox": [ + 612, + 592, + 803, + 609 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $x_{t} \\in X$ is a sample at time $t$ , $f(x_{t}, t)$ is the drift term, $g(t)$ is the diffusion coefficient, and $w_{t}$ is a standard Wiener process.", + "bbox": [ + 511, + 616, + 906, + 660 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "During training, the forward diffusion process comprises a Markov chain with fixed time steps $T$ . Given a data point $x_0 \\sim q(x)$ , we iteratively add Gaussian noise with variance $\\beta_t$ at each time step $t$ to $x_{t-1}$ to get $x_t$ such that $x_T \\sim \\mathcal{N}(0, I)$ . This process can be expressed as,", + "bbox": [ + 511, + 662, + 905, + 737 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nq (x _ {t} | x _ {t - 1}) = \\mathcal {N} (x _ {t}; \\sqrt {1 - \\beta_ {t}} x _ {t - 1}, \\beta_ {t} I), \\quad \\forall t \\in \\{1, \\dots , T \\}.\n$$\n", + "text_format": "latex", + "bbox": [ + 513, + 744, + 910, + 763 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We can get a closed-form expression of $x_{t}$ ,", + "bbox": [ + 532, + 770, + 818, + 785 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nx _ {t} = \\sqrt {\\bar {\\alpha} _ {t}} x _ {0} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon_ {t}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 617, + 792, + 906, + 809 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\bar{\\alpha}_t = \\prod_{i=1}^t (1 - \\beta_i)$ and $\\alpha_t = 1 - \\beta_t$ .", + "bbox": [ + 511, + 815, + 808, + 832 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "During training, we learn the reverse process through a network $\\epsilon_{\\theta}$ to iteratively denoise $x_{t}$ by estimating the noise $\\epsilon_{t}$ at each time step $t$ . The loss function is expressed as,", + "bbox": [ + 511, + 832, + 905, + 876 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathbb {E} _ {t \\in [ 1, T ], \\epsilon \\sim \\mathcal {N} (0, I)} \\left[ \\left\\| \\epsilon_ {t} - \\epsilon_ {\\theta} \\left(x _ {t}\\right) \\right\\| _ {2} ^ {2} \\right]. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 578, + 883, + 906, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "12873", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Using the learned noise estimator network $\\epsilon_{\\theta}$ , we can compute the previous sample $x_{t - 1}$ from $x_{t}$ as follows,", + "bbox": [ + 89, + 90, + 482, + 122 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nx _ {t - 1} = \\frac {1}{\\sqrt {\\alpha} _ {t}} \\left(x _ {t} - \\frac {1 - \\alpha_ {t}}{\\sqrt {1 - \\bar {\\alpha} _ {t}}} \\epsilon_ {\\theta} \\left(x _ {t}\\right)\\right). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 161, + 132, + 483, + 165 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The learned noise estimator network $\\epsilon_{\\theta}$ can be conditioned using a conditioning input such as text or class label, expressed as an embedding $e_p$ . Ho et al. [10] proposed classifier-free guidance (CFG) as a mechanism to guide the diffusion trajectory towards generating outputs that align with the conditioning. The trajectory is directed towards the conditional score predictions and away from the unconditional score predictions, where $s$ controls the degree of adjustment and $e_{\\emptyset}$ are empty prompt embeddings used for unconditional guidance.", + "bbox": [ + 89, + 175, + 483, + 325 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\epsilon} \\leftarrow \\epsilon_ {\\theta} \\left(x _ {t}, e _ {\\emptyset}\\right) + s \\underbrace {\\left(\\epsilon_ {\\theta} \\left(x _ {t} , e _ {p}\\right) - \\epsilon_ {\\theta} \\left(x _ {t} , e _ {\\emptyset}\\right)\\right)} _ {\\text {c o n d i t i o n a l g u i d a n c e}}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 138, + 338, + 483, + 375 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Understanding Diffusion Trajectory during Memorization", + "text_level": 1, + "bbox": [ + 89, + 386, + 415, + 422 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In text-to-image diffusion models, different initializations generally lead to different outputs for the same text prompt. However, when the diffusion model has memorized a particular sample, the outputs are similar and closely resemble one or more training data samples regardless of initialized noise [28].", + "bbox": [ + 89, + 429, + 483, + 518 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As shown in Figure 2, we observe that applying CFG after a certain time step avoids producing memorized outputs, while applying it from earlier time steps, including at the beginning $(t = T)$ , easily leads to memorized outputs. Inspired by these empirical findings and the perspective of dynamical systems [2, 15], we suspect that there is an attraction basin in the sample space of the diffusion trajectory, in which applying CFG results in memorized outputs. Intuitively, one can imagine this attraction basin as having a funnel-like shape. It is broader at $t = T$ and narrows as the reverse denoising process progresses. To define the attraction basin, we introduce the concept of a denoiser.", + "bbox": [ + 89, + 521, + 483, + 702 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Definition 1 (Denoiser). Let $X$ denote the entire sample space of the diffusion trajectory, and let $E$ denote the embedding space. A denoiser is a function $\\varphi : X \\times (0, T] \\times E \\to X$ that outputs the result of the reverse diffusion process based on Eq. (3) with CFG. Specifically, $\\varphi(x, t, e)$ represents the resulting sample at time $t = 0$ , starting from a state $(x, t)$ with embeddings $e \\in E$ , where $x \\in X$ and $t \\in (0, T]$ .", + "bbox": [ + 89, + 712, + 483, + 819 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For an attractor $x^{\\mathrm{a}}$ from the training dataset, we then define the attraction basin as follows:", + "bbox": [ + 89, + 829, + 482, + 859 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Definition 2 (Attractor and attraction basin). Suppose a prompt is given with embeddings denoted as $e_p$ . A sample", + "bbox": [ + 89, + 869, + 483, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$x^{\\mathrm{a}} \\in X$ is considered an $(\\epsilon)$ -attractor if the diffusion trajectories denoised with CFG tend to converge to samples within $B_{\\mathcal{D}}(x^{\\mathrm{a}}, \\epsilon)$ at $t = 0$ , regardless of the initial noise, i.e., $\\varphi(x, T, e_p) \\in B_{\\mathcal{D}}(x^{\\mathrm{a}}, \\epsilon)$ for any $x \\in X$ , where $B_{\\mathcal{D}}(\\cdot, \\cdot)$ represents a perceptual ball given by a certain perceptual distance $\\mathcal{D}: X \\times X \\to \\mathbb{R}_{\\geq 0}$ as", + "bbox": [ + 511, + 90, + 906, + 181 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nB _ {\\mathcal {D}} \\left(x ^ {\\mathrm {a}}, \\epsilon\\right) = \\left\\{x ^ {\\prime} \\in X \\mid \\mathcal {D} \\left(x ^ {\\mathrm {a}}, x ^ {\\prime}\\right) \\leq \\epsilon \\right\\}.\n$$\n", + "text_format": "latex", + "bbox": [ + 573, + 196, + 844, + 213 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "An attraction basin is a region in the state space within which samples will converge to around $x^{\\mathrm{a}}$ at $t = 0$ under CFG inference. More formally, the $((\\epsilon, \\delta))$ -attraction basin defined using $x^{\\mathrm{a}}$ is expressed as", + "bbox": [ + 511, + 227, + 906, + 287 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nX ^ {\\mathrm {b}} \\left(x ^ {\\mathrm {a}}, \\epsilon\\right) = \\left\\{\\left(x, t\\right) \\mid \\mathbb {P} \\left(\\varphi (x, t, e) \\in B _ {\\mathcal {D}} \\left(x ^ {\\mathrm {a}}, \\epsilon\\right)\\right) > 1 - \\delta \\right\\},\n$$\n", + "text_format": "latex", + "bbox": [ + 513, + 301, + 908, + 327 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "representing all points in the state space from which generated samples using CFG will reach samples perceptually similar to the attractor at $t = 0$ .", + "bbox": [ + 511, + 340, + 906, + 383 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Inspired from Wen et al. [28], we analyze the attraction basin by examining the magnitudes of the conditional guidance, $\\epsilon_{\\theta}(x_t,e_p) - \\epsilon_{\\theta}(x_t,e_{\\emptyset})$ . Interestingly, as shown in Figure 2, the conditional noise prediction $\\epsilon_{\\theta}(x_t,e_p) - \\epsilon_{\\theta}(x_t,e_{\\emptyset})$ remains high during the initial time steps and then steeply falls at the time step immediately before the denoising trajectory exits the attraction basin. We refer to this time step as the transition point, marking the change from producing memorized outputs to non-memorized outputs. In Figure 2(a), if you were to start applying CFG at any point on or before $t = 779$ , you get the same memorized output. However, surprisingly, applying CFG at exactly one denoising step after this $t = 759$ leads to a non-memorized output (step size is 20 when denoising with 50 inference steps). We define the transition point as follows.", + "bbox": [ + 511, + 398, + 908, + 626 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Definition 3 (Transition point). Given prompt embeddings, let $x^{\\mathrm{a}} \\in X$ be an attractor. We define a transition point as $(x_{\\tau}, \\tau)$ for $x_{\\tau} \\in X$ and $\\tau > 0$ such that $(x_{\\tau}, \\tau + 1)$ lies within its attraction basin, while $(x_{\\tau}, \\tau)$ does not. More formally, an $((\\epsilon, \\delta))$ -transition point satisfies", + "bbox": [ + 511, + 640, + 906, + 715 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {P} \\left(\\varphi \\left(x _ {\\tau}, \\tau + 1\\right) \\in B _ {\\mathcal {D}} \\left(x ^ {\\mathrm {a}}, \\epsilon\\right)\\right) > 1 - \\delta \\quad a n d\n$$\n", + "text_format": "latex", + "bbox": [ + 560, + 729, + 859, + 747 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {P} (\\varphi (x _ {\\tau}, \\tau) \\notin B _ {\\mathcal {D}} (x ^ {\\mathrm {a}}, \\epsilon)) < 1 - \\delta ,\n$$\n", + "text_format": "latex", + "bbox": [ + 560, + 750, + 792, + 766 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $B_{\\mathcal{D}}$ is a perceptual ball defined in Definition 2.", + "bbox": [ + 511, + 780, + 867, + 795 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This phenomenon occurs only for memorized samples as illustrated in Figure 6, the conditional guidance $\\epsilon_{\\theta}(x_t,e_p) - \\epsilon_{\\theta}(x_t,e_{\\emptyset})$ remains low and almost flat across different time steps for non-memorized samples. We further analyze the diffusion trajectory during memorization, demonstrating the presence of an attraction basin in Appendix 8.", + "bbox": [ + 511, + 810, + 905, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "12874", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "5. Simple Mitigation Strategy", + "text_level": 1, + "bbox": [ + 89, + 89, + 341, + 108 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Based on the intuition from the previous section, we can simply mitigate memorization by finding an ideal transition point to switch from zero CFG to applying CFG. Based on our preliminary experiments, we found that transition points can either be static or dynamic.", + "bbox": [ + 89, + 114, + 483, + 191 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/be1c77a4df291f24f0d967c3297837186a10fc276eeab0f141b6a03ec6de2d81.jpg", + "image_caption": [ + "Figure 3. Average magnitude of the text-conditioned $(\\epsilon_{\\theta}(x_t, e_p))$ and unconditional noise predictions $(\\epsilon_{\\theta}(x_t, e_{\\emptyset}))$ and their difference $(\\epsilon_{\\theta}(x_t, e_p) - \\epsilon_{\\theta}(x_t, e_{\\emptyset}))$ when applying zero CFG. We see a static transition point ( $t = 500$ ) appear when SDv2.1 is finetuned on the LAION-10k dataset [24]." + ], + "image_footnote": [], + "bbox": [ + 151, + 205, + 424, + 340 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/e87873fd6aeb16f7c3cf38ff8bc82c95c27358ed8b83276fe6b97703b148c80f.jpg", + "image_caption": [ + "Figure 4. Applying CFG before the static transition point $(\\mathrm{T} = 500)$ leads to memorized outputs while applying CFG after the fixed transition point leads to non-memorized outputs. Applying CFG too late results in poor-quality images that resemble the unconditional generations." + ], + "image_footnote": [], + "bbox": [ + 94, + 448, + 483, + 636 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Static Transition Point. In some trained diffusion models, we observed a universal trend in the mean conditional $\\epsilon_{\\theta}(x_t,e_p)$ and unconditional noise prediction $\\epsilon_{\\theta}(x_t,e_{\\emptyset})$ across different samples. As shown in Figure 3, the mean conditional noise prediction $\\epsilon_{\\theta}(x_t,e_p)$ is higher than unconditional noise prediction $\\epsilon_{\\theta}(x_t,e_{\\emptyset})$ in the initial time steps $(t\\geq 500)$ . At a static time step $(t = 500)$ , the two scores align in magnitude until the end of the denoising process, and at the same time, the $L_{2}$ distance between them drastically drops. This suggests that the zero CFG trajectory leaves the attraction basin at that fixed time step $(t = 500)$ .", + "bbox": [ + 89, + 734, + 485, + 902 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/0f61266e6523b9078ce2df0b2203aa2ad178e6a4798481e1b2869589d865117e.jpg", + "image_caption": [ + "Figure 5. In some models, transition points can occur at a different time step for each sample, as seen for pre-trained SDv1.4. In row 1 the transition point is approximately $t = 800$ while for row 2 it is $t = 650$ ." + ], + "image_footnote": [], + "bbox": [ + 517, + 90, + 903, + 286 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This can be further validated by applying CFG from any time step before the fixed transition point $(t = 500)$ , as illustrated in Figure 4. Applying CFG from an early time step leads to the same memorized image, but the output immediately switches at the transition point $(t = 500)$ such that we generate text-aligned non-memorized images. On the other hand, if we start applying CFG very late in the denoising process, the outputs are not well aligned with the text prompts and resemble their unconditional generations.", + "bbox": [ + 511, + 375, + 908, + 511 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Dynamic Transition Point. In other trained diffusion models, we saw that every prompt and initialization pair leads to a different transition point. As shown in Figures 6(a) and 5, different trajectories leave the attraction basin at different time steps, characterized by their fall in the conditional guidance $\\epsilon_{\\theta}(x_t,e_p) - \\epsilon_{\\theta}(x_t,e_{\\emptyset})$ . We found that it is dependent on both the conditional embedding $e_p$ and the initialization $x_{T}$ . We show an example in Figures 2(a) and 2(b), where the transition point occurs at different time steps even for the same prompt. Thus, we propose a method to find this dynamic transition point (DTP). As shown in Figure 2, it is the point after the first local minima in the graph. Thus, to mitigate memorization, this difference is tracked until the first local minimum occurs, and CFG is applied henceforth. This process requires no additional computations as the text-conditioned and unconditional noise predictions are computed by default when denoising using CFG. We present the pseudo-code for this approach in Algorithm 1.", + "bbox": [ + 511, + 518, + 908, + 792 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Opposite Guidance", + "text_level": 1, + "bbox": [ + 511, + 801, + 697, + 818 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "An issue with this method is that when the transition point occurs very late ( $t \\leq 500$ ) in the denoising process, CFG is applied for fewer time steps, and this can impact the image quality.", + "bbox": [ + 511, + 824, + 906, + 885 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To ensure that the transition point occurs earlier, we in", + "bbox": [ + 532, + 885, + 908, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "12875", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/3f50b349a1bd99b69a39980048240d1d08f250ab20741fe040d0787ff82c2cab.jpg", + "image_caption": [ + "(a) Applying zero CFG" + ], + "image_footnote": [], + "bbox": [ + 151, + 89, + 421, + 223 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1f3f1140f2a1236b8917829a23e348de47fd494d97987dca69da77361cea0132.jpg", + "image_caption": [ + "(b) Applying opposite guidance", + "Figure 6. Magnitude of $\\epsilon_{\\theta}(x_t, e_p) - \\epsilon_{\\theta}(x_t, e_{\\emptyset})$ when denoising (a) without CFG and (b) with opposite guidance. Transition point occurs sooner with opposite guidance." + ], + "image_footnote": [], + "bbox": [ + 151, + 243, + 421, + 376 + ], + "page_idx": 5 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1 Reverse Diffusion with Dynamic Transition Point Method and Opposite Guidance" + ], + "code_body": "Require: $x_{T}\\sim \\mathcal{N}(0,\\mathrm{I}_{d}),\\lambda >0,v_{\\mathrm{OG}}\\in \\{0,1\\}$ \n1: $d_{T + 2} = -\\infty$ . $d_{T + 1} = -\\infty$ \n2: $s = -\\lambda *v_{\\mathrm{OG}}$ \n3: for $t = T$ to 1 do \n4: $d_{t} = \\| \\epsilon_{\\theta}(x_{t},e_{p}) - \\epsilon_{\\theta}(x_{t},e_{\\emptyset})\\|_{2}^{2}$ \n5: if $d_{t + 2} > d_{t + 1}$ and $d_{t + 1} < d_t$ then \n6: $s = \\lambda$ \n7: end if \n8: $\\hat{\\epsilon} = \\epsilon_{\\theta}(x_t,e_\\emptyset) + s(\\epsilon_\\theta (x_t,e_p) - \\epsilon_\\theta (x_t,e_\\emptyset))$ \n9: $x_{t - 1} = \\frac{1}{\\sqrt{\\alpha}_t} (x_t - \\frac{1 - \\alpha_t}{\\sqrt{1 - \\alpha_t}}\\hat{\\epsilon})$ \n10: end for \n11: return $x_0$", + "bbox": [ + 91, + 491, + 483, + 662 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "troduce a new concept of opposite guidance (OG). Unlike negative prompting, we apply negative or opposite CFG, to push the diffusion trajectory in the opposite direction to that of traditional CFG and thus, push the trajectory away from the attraction basin of the traditional CFG trajectory sooner in the denoising process, as also shown in Figure 6(b). Opposite guidance can be expressed as follows for $s > 0$ ,", + "bbox": [ + 89, + 689, + 485, + 796 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\epsilon} \\leftarrow \\epsilon_ {\\theta} \\left(x _ {t}, e _ {\\emptyset}\\right) - s \\underbrace {\\left(\\epsilon_ {\\theta} \\left(x _ {t} , e _ {p}\\right) - \\epsilon_ {\\theta} \\left(x _ {t} , e _ {\\emptyset}\\right)\\right)} _ {\\text {c o n d i t i o n a l g u i d a n c e}}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 138, + 821, + 483, + 859 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Similar to unconditional denoising, when denoising using opposite guidance, there is an ideal time step where", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "the magnitude of conditional guidance falls. At this time step, switching from opposite guidance to traditional positive guidance leads to non-memorized samples with high image quality and text alignment. Thus, it can easily be integrated along with the previous method to find a static or dynamic transition point. We present the pseudo-code for this approach in conjunction with the DTP method in Algorithm 1 with the parameter $v_{\\mathrm{OG}} = 1$ .", + "bbox": [ + 511, + 90, + 906, + 212 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6. Memorization under Different Scenarios", + "text_level": 1, + "bbox": [ + 511, + 226, + 877, + 242 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We study memorization under different scenarios and show how this simple approach of finding an ideal transition point to avoid the attraction basin can be applied to effectively mitigate memorization.", + "bbox": [ + 511, + 251, + 906, + 311 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Scenario 1: Fine-tuned SDv2.1 on 10,000 LAION datapoints (followed by [24]). Memorization occurs due to overfitting on the small dataset.", + "- Scenario 2: Fine-tuned SDv2.1 on the Imagenette dataset [12] containing 10 ImageNet classes (followed by [24]). Exact memorization is not always observed but similarity with the training dataset increases.", + "- Scenario 3: Fine-tuned SDv1.4 on 200 prompts duplicated 200 times with an additional 120,000 prompts that were not duplicated (followed by [28]). Memorization occurs due to data duplication.", + "- Scenario 4: 500 memorized prompts for pre-trained SDv1.4 found using membership inference attack [4, 27] (followed by [20]). The exact cause of memorization for each prompt is unknown, but the strong presence of \"trigger tokens\" is noticeable. These are in the form of nouns such as celebrity and movie names with little to no additional descriptions. Results are in Appendix 9." + ], + "bbox": [ + 511, + 314, + 906, + 587 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Notably, different researchers have independently studied these different scenarios, but, as we show later, their approaches do not generalize to other scenarios as memorization occurs due to different causes in each scenario. Refer to Appendix 11 for other implementation details.", + "bbox": [ + 511, + 589, + 908, + 666 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metrics We utilize the similarity metric based on the SSCD [18] embeddings to judge the level of memorization [24]; the CLIP score [19] to assess how well the generated images align with the text prompts; and the Fréchet Inception Distance (FID) [9] for image quality.", + "bbox": [ + 511, + 671, + 908, + 748 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6.1. Scenario 1", + "text_level": 1, + "bbox": [ + 511, + 757, + 630, + 772 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We observed that in this scenario, a static transition point at $t = 500$ occurs as shown in Figure 3. We present qualitative and quantitative results in Figure 7 and Table 1, respectively. Our approach outperformed all previous methods, and we observed the previously proposed techniques that focused solely on the presence of \"trigger tokens\" [20, 28] did not work well. This shows that when memorization occurs due to factors other than data duplication or the presence of", + "bbox": [ + 511, + 780, + 908, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "12876", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b1c0a9055535ea261072cbbe07be2c967b5f1a18aeea5917f71058e67642405d.jpg", + "image_caption": [ + "Figure 7. Qualitative results comparing the proposed approach with the baselines in Scenario 1. Ren et al. [20] is unable to mitigate memorization in most cases while Wen et al. [28] has poor image quality and text-alignment. More enlarged examples (Fig. 14) and prompts for this figure are in Appendix 12." + ], + "image_footnote": [], + "bbox": [ + 235, + 87, + 767, + 441 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/2d72096b01a19ad8f646714819e8a9002b2d2ec4dc32ac167e171b2eb4feffa7.jpg", + "table_caption": [ + "Table 1. Comparison of results on Scenario 1. None of the other approaches perform well. The only comparable method [28] qualitatively results in poorer quality images (see visual examples in Fig. 7)." + ], + "table_footnote": [], + "table_body": "
Similarity (95pc)CLIP ScoreFID
No Mitigation0.65040.302716.8373
Add rand word [24]0.52540.294117.4142
Add rand numb [24]0.54160.299317.0245
Wen et al. [28]0.38530.289516.7176
Ren et al. [20]0.60280.295920.2931
Ours (STP)0.28570.297619.8494
Ours (OG + STP)0.38110.302015.6679
", + "bbox": [ + 94, + 570, + 483, + 678 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "trigger tokens, these methods may fail. We present further results when finetuning on the LAION-100k dataset [24] in Appendix 10.", + "bbox": [ + 89, + 702, + 483, + 748 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.2. Scenario 2", + "text_level": 1, + "bbox": [ + 89, + 757, + 209, + 772 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Even in this scenario, as illustrated in Figure 10, a static transition point occurs. We applied CFG from $t = 700$ and $t = 600$ to show that we can reduce the similarity with the training dataset at no cost to the image quality, as shown in Table 2. We would like to point out that previous approaches are specific to text-to-image models and thus cannot be extrapolated to other conditioning mechanisms. This shows a further advantage of our CFG-based mitigation", + "bbox": [ + 89, + 779, + 483, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "technique.", + "bbox": [ + 513, + 513, + 586, + 527 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/93ad06585487a5c36d2920625bd95a88b58e551bf062f965babb115d1f7d1cde.jpg", + "table_caption": [ + "Table 2. Results on Scenario 2. The value in the brackets in the time steps between which we apply CFG. Applying CFG later from $t = 700$ (row 2) and $t = 600$ (row 3) reduces similarity with comparable or improved FID score." + ], + "table_footnote": [], + "table_body": "
Similarity (95pc)FID
No mitigation0.370243.35
CFG=7.5 [700,0]0.301838.86
CFG=7.5 [600,0]0.275645.73
", + "bbox": [ + 544, + 601, + 875, + 667 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.3. Scenario 3", + "text_level": 1, + "bbox": [ + 513, + 691, + 632, + 705 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this scenario, we observed that a dynamic transition point occurs. We experimentally validate our approach of DTP method and opposite guidance qualitatively and quantitatively in Figure 8 and Table 3. We saw that our method performs at par with [28] that studied this specific scenario and even outperforms it in terms of memorization in the top 95 percentile. Approaches by Ren et al. [20] and Somepalli et al.[24] that were proposed for other scenarios do not work well, showcasing their lack of generalizability.", + "bbox": [ + 511, + 713, + 906, + 849 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Inference Time. One of the major advantages of our approach is its simplicity and lack of computational time overhead. Image generation using our approach as well as stan-", + "bbox": [ + 511, + 854, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "12877", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/56f91106059d11df40a99e806c65b6ebbf7aba033ab488b7dc661fdfb66c135f.jpg", + "image_caption": [ + "Figure 8. Qualitative results comparing the proposed approach with the baselines in Scenario 3. Prompts used to generate these figures are given in the Appendix. Rows 1, 2, 5, and 6 show examples, where for most baselines the output images remain closely related to their memorized versions. More enlarged examples (Fig. 15) and prompts for this figure are in Appendix 13." + ], + "image_footnote": [], + "bbox": [ + 236, + 88, + 767, + 445 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/9668c0ba64d63a8eceb6330fb0d8712c7888df80d598b993e26090acf724b63a.jpg", + "table_caption": [ + "Table 3. Table showcasing results in Scenario 3. Wen et al. [28] studied this scenario and is the only other method which performs well but does not generalize to other scenarios (see Table 1)." + ], + "table_footnote": [], + "table_body": "
SimilarityCLIP ScoreFID
95pcMean ± Std
No Mitigation0.79770.5513 ± 0.160.3105106.49
Add rand word [24]0.74800.4312 ± 0.210.3071116.92
Add rand numb [24]0.73660.3850 ± 0.210.3027126.54
Wen et al. [28] (l_target=3)0.77470.5147 ± 0.170.3100109.28
Wen et al. [28] (l_target=1)0.60380.2808 ± 0.160.3050136.34
Ren et al. [20]0.68810.4036 ± 0.160.3066124.38
Ours (DTP)0.58850.2866 ± 0.160.3020138.92
Ours (OG + DTP)0.69150.2844 ± 0.200.2910140.05
", + "bbox": [ + 93, + 563, + 483, + 679 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "dard CFG-based image generation using SDv2.1 on one A100 GPU takes 1.26 seconds, while in comparison the mitigation technique proposed by Wen et al. [28] took 2.86 seconds and that of Ren et al. [20] took 2.08 seconds. These are the average values across 10,000 generations.", + "bbox": [ + 89, + 705, + 483, + 782 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations. As transition points do not exist for nonmemorized samples, we require detecting memorization before we can apply our approach. This can either be done by denoising twice, once with traditional CFG to look at the trends in the magnitude of conditional guidance and the second time to apply our mitigation technique if the prompt is memorized. Alternatively, as [20, 28] showed, we can", + "bbox": [ + 89, + 794, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "do detection at $t = 0$ by simply looking at the magnitude of conditional guidance $\\| \\epsilon_{\\theta}(x_t, e_p) - \\epsilon_{\\theta}(x_t, e_{\\emptyset}) \\|_2^2$ . If this magnitude is above a certain threshold, it indicates that the sample is memorized with high accuracy and we can apply our guidance accordingly.", + "bbox": [ + 511, + 518, + 906, + 595 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 623, + 633, + 638 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We present a novel understanding of the dynamics behind memorization in conditional diffusion models. We showcase the presence of an attraction basin that steers randomly initialized latent vectors towards a memorized output. We propose an approach to steer away from the attraction basin and detect the point at which the trajectory leaves the attraction basin, referred to as the transition point. We show that applying CFG after the transition point leads to non-memorized outputs. We presented results in various scenarios such as the presence of data replication, fine-tuning on smaller datasets, and memorized examples found in existing pre-trained models to showcase its efficiency.", + "bbox": [ + 511, + 654, + 908, + 835 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To summarize, our method is simple, works well across all tested scenarios in terms of both image quality metrics and visual inspection, requires less computational time, and does not require altering the prompt.", + "bbox": [ + 511, + 839, + 908, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "12878", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Luca Ambrogioni. In search of dispersed memories: Generative diffusion models are associative memory networks. Entropy, 26(5):381, 2024. 3", + "[2] Joseph Auslander, Nam P Bhatia, and Peter Seibert. Attractors in dynamical systems. Technical report, 1964. 4", + "[3] Yuanhao Ban, Ruochen Wang, Tianyi Zhou, Minhao Cheng, Boqing Gong, and Cho-Jui Hsieh. Understanding the impact of negative prompts: When and how do they take effect? arXiv preprint arXiv:2406.02965, 2024. 2", + "[4] Nicolas Carlini, Jamie Hayes, Milad Nasr, Matthew Jagielski, Vikash Sehwag, Florian Tramer, Borja Balle, Daphne Ippolito, and Eric Wallace. Extracting training data from diffusion models. In 32nd USENIX Security Symposium (USENIX Security 23), pages 5253-5270, 2023. 1, 2, 3, 6", + "[5] Huiwen Chang, Han Zhang, Jarred Barber, AJ Maschinot, Jose Lezama, Lu Jiang, Ming-Hsuan Yang, Kevin Murphy, William T Freeman, Michael Rubinstein, et al. Muse: Text-to-image generation via masked generative transformers. arXiv preprint arXiv:2301.00704, 2023. 1", + "[6] Ruchika Chavhan, Ondrej Bohdal, Yongshuo Zong, Da Li, and Timothy Hospedales. Memorized images in diffusion models share a subspace that can be located and deleted. arXiv preprint arXiv:2406.18566, 2024. 3", + "[7] Chen Chen, Daochang Liu, and Chang Xu. Towards memorization-free diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8425-8434, 2024. 3", + "[8] Giannis Daras, Alexandros G Dimakis, and Constantinos Daskalakis. Consistent diffusion meets tweedie: Training exact ambient diffusion models with noisy data. arXiv preprint arXiv:2404.10177, 2024. 3", + "[9] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 6", + "[10] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 4", + "[11] Benjamin Hoover, Hendrik Strobelt, Dmitry Krotov, Judy Hoffman, Zsolt Kira, and Duen Horng Chau. Memory in plain sight: Surveying the uncanny resemblances of associative memories and diffusion models. arXiv preprint arXiv:2309.16750, 2023. 3", + "[12] Jeremy Howard. Imagenette: A smaller subset of 10 easily classified classes from imagenet, 2019. 6, 3", + "[13] Tuomas Kynkänniemi, Miika Aittala, Tero Karras, Samuli Laine, Timo Aila, and Jaakko Lehtinen. Applying guidance in a limited interval improves sample and distribution quality in diffusion models. arXiv preprint arXiv:2404.07724, 2024. 1", + "[14] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps. Advances in Neural Information Processing Systems, 35:5775-5787, 2022. 3" + ], + "bbox": [ + 93, + 114, + 485, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] John W. Milnor. Attractor. *Scholarpedia*, 1(11):1815, 2006. 4", + "[16] Alex Nichol. Dall-e 2 pre-training mitigations. https://openai.com/index/dall-e-2-pre-training-mitigations/, 2022. [Accessed 10-10-2024]. 2", + "[17] Bao Pham, Gabriel Raya, Matteo Negri, Mohammed J Zaki, Luca Ambrogioni, and Dmitry Krotov. Memorization to generalization: The emergence of diffusion models from associative memory. 2024. 3", + "[18] Ed Pizzi, Sreya Dutta Roy, Sugosh Nagavara Ravindra, Priya Goyal, and Matthijs Douze. A self-supervised descriptor for image copy detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14532-14542, 2022. 3, 6", + "[19] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 6", + "[20] Jie Ren, Yaxin Li, Shenglai Zen, Han Xu, Lingjuan Lyu, Yue Xing, and Jiliang Tang. Unveiling and mitigating memorization in text-to-image diffusion models through cross attention. arXiv preprint arXiv:2403.11052, 2024. 1, 2, 3, 6, 7, 8, 5", + "[21] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 3", + "[22] Chitwan Saharia, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural information processing systems, 35:36479-36494, 2022. 3", + "[23] Gowthami Somepalli, Vasu Singla, Micah Goldblum, Jonas Geiping, and Tom Goldstein. Diffusion art or digital forgery? investigating data replication in diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6048-6058, 2023. 1, 2, 3", + "[24] Gowthami Somepalli, Vasu Singla, Micah Goldblum, Jonas Geiping, and Tom Goldstein. Understanding and mitigating copying in diffusion models. Advances in Neural Information Processing Systems, 36:47783-47803, 2023. 1, 2, 3, 5, 6, 7, 8", + "[25] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 1, 3", + "[26] Xi Wang, Nicolas Dufour, Nefeli Andreou, Marie-Paule Cani, Victoria Fernández Abrevaya, David Picard, and Vicky Kalogeiton. Analysis of classifier-free guidance weight schedulers. arXiv preprint arXiv:2404.13040, 2024. 1", + "[27] Ryan Webster. A reproducible extraction of training images from diffusion models. arXiv preprint arXiv:2305.08694, 2023. 2, 3, 6", + "[28] Yuxin Wen, Yuchen Liu, Chen Chen, and Lingjuan Lyu. Detecting, explaining, and mitigating memorization in diffusion models. In The Twelfth International Conference on Learning Representations, 2024. 1, 2, 3, 4, 6, 7, 8, 5" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "12879", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + } +] \ No newline at end of file diff --git a/2025/Classifier-Free Guidance Inside the Attraction Basin May Cause Memorization/a1a30dde-51b7-469a-a404-0c4a54031538_model.json b/2025/Classifier-Free Guidance Inside the Attraction Basin May Cause Memorization/a1a30dde-51b7-469a-a404-0c4a54031538_model.json new file mode 100644 index 0000000000000000000000000000000000000000..7ec3029c8dbb2d1fc9461d14817f159579a18294 --- /dev/null +++ b/2025/Classifier-Free Guidance Inside the Attraction Basin May Cause Memorization/a1a30dde-51b7-469a-a404-0c4a54031538_model.json @@ -0,0 +1,1923 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.102, + 0.131, + 0.895, + 0.152 + ], + "angle": 0, + "content": "Classifier-Free Guidance inside the Attraction Basin May Cause Memorization" + }, + { + "type": "text", + "bbox": [ + 0.23, + 0.18, + 0.768, + 0.217 + ], + "angle": 0, + "content": "Anubhav Jain1*, Yuya Kobayashi2, Takashi Shibuya2, Yuhta Takida2, Nasir Memon1, Julian Togelius1, Yuki Mitsufuji2,3" + }, + { + "type": "text", + "bbox": [ + 0.264, + 0.217, + 0.734, + 0.235 + ], + "angle": 0, + "content": "\\(^{1}\\)New York University, \\(^{2}\\)Sony AI, \\(^{3}\\)Sony Group Corporation" + }, + { + "type": "text", + "bbox": [ + 0.312, + 0.235, + 0.686, + 0.252 + ], + "angle": 0, + "content": "{aj3281,memon,julian.togelius}@nyu.edu" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.253, + 0.841, + 0.27 + ], + "angle": 0, + "content": "{u.kobayashi,takashi.tak.shibuya,yuta.takida,yuhki.mitsufuji}@sony.com" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.305, + 0.327, + 0.32 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.338, + 0.486, + 0.685 + ], + "angle": 0, + "content": "Diffusion models are prone to exactly reproduce images from the training data. This exact reproduction of the training data is concerning as it can lead to copyright infringement and/or leakage of privacy-sensitive information. In this paper, we present a novel perspective on the memorization phenomenon and propose a simple yet effective approach to mitigate it. We argue that memorization occurs because of an attraction basin in the denoising process which steers the diffusion trajectory towards a memorized image. However, this can be mitigated by guiding the diffusion trajectory away from the attraction basin by not applying classifier-free guidance until an ideal transition point occurs from which classifier-free guidance is applied. This leads to the generation of non-memorized images that are high in image quality and well-aligned with the conditioning mechanism. To further improve on this, we present a new guidance technique, opposite guidance, that escapes the attraction basin sooner in the denoising process. We demonstrate the existence of attraction basins in various scenarios in which memorization occurs, and we show that our proposed approach successfully mitigates memorization. Our codebase is publicly available at https://github.com/SonyResearch/mitigating_memorization." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.715, + 0.222, + 0.73 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.74, + 0.486, + 0.877 + ], + "angle": 0, + "content": "Recent advancements in image-generation models have enabled the generation of hyper-realistic images. Diffusion models [21, 25] have been at the forefront of this advancement with their ability to align image generation with different conditioning mechanisms such as text giving users more control over the output. However, a key challenge has emerged: these models can memorize and reproduce training examples verbatim (verbatim memorization) [4, 23, 24], leading to concerns over copyright infringement and leak" + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.309, + 0.903, + 0.503 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.509, + 0.907, + 0.607 + ], + "angle": 0, + "content": "Figure 1. The diffusion trajectory contains an attraction basin (red region) which steers conditioned samples towards their memorized images. It can be avoided by applying zero classifier-free guidance when the trajectory is inside the attraction basin, such that there is an ideal transition point \\(\\tau^{*}\\) after which applying CFG leads to non-memorized output. Applying CFG at an earlier point such as \\(\\tau_{1}\\) inside the attraction basin leads to the same memorized sample." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.64, + 0.909, + 0.821 + ], + "angle": 0, + "content": "age of privacy sensitive information. Alternatively, the model can generate images copying the composition and structure present in one or more training images (template memorization)[20]. Researchers have identified factors contributing to memorization and reproduction, such as the duplication of training images and repeated captions [24]. However, not all causes are understood, as memorization can still occur even when these issues are avoided. Therefore, there is a growing need for effective mitigation strategies, in particular methods that can be applied during inference without requiring computationally intensive retraining of the core diffusion models." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.91, + 0.901 + ], + "angle": 0, + "content": "In text-to-image (T2I) diffusion models, strong associations between a text prompt and an image in the training dataset lead to memorization. This can occur due to the presence of \"trigger tokens\" in overly specific text-prompts [20, 28], the model seeing multiple occurrences of these dur" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.887, + 0.348, + 0.901 + ], + "angle": 0, + "content": "*Work done during an internship at Sony AI." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "12871" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.098, + 0.09, + 0.49, + 0.27 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.092, + 0.276, + 0.495, + 0.3 + ], + "angle": 0, + "content": "(a) Prompt: \"Here's What You Need to Know About St. Vincent's Apple Music Radio Show\"" + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.089, + 0.905, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.501, + 0.276, + 0.905, + 0.3 + ], + "angle": 0, + "content": "(b) Prompt: \"Here's What You Need to Know About St. Vincent's Apple Music Radio Show\"" + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.306, + 0.908, + 0.364 + ], + "angle": 0, + "content": "Figure 2. Plots showing magnitude of \\(\\epsilon_{\\theta}(x_t, e_p) - \\epsilon_{\\theta}(x_t, e_{\\emptyset})\\) when denoising without classifier-free guidance (CFG) at each time step. The figures show the generated image if you start applying CFG at that time step. We get non-memorized output if we apply CFG after the ideal transition point \\(\\tau^*\\) which coincides with the fall in the conditional noise prediction. This value is dependent on both the prompt and the initialization ((a) and (b) contain the same prompt with different initializations). More examples in the Appendix." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.389, + 0.486, + 0.556 + ], + "angle": 0, + "content": "ing training (data duplication) [24, 28], or other causes such as fine-tuning on smaller datasets [24]. Previous approaches have focused on perturbing either the text prompt/embeddings [24, 28] or the cross-attention between the text and image embeddings [20]. While these approaches help weaken the associations between prompts and corresponding images, they are limited to specific conditions in which they have been tested and fail to generalize effectively. Additionally, since these techniques work at the prompt level, they are only applicable to text-conditional generation, and also assume the ability to update the prompt." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.57, + 0.487, + 0.902 + ], + "angle": 0, + "content": "In this paper, we introduce a new concept to explain the behavior of memorization in diffusion models based on our empirical observations. We first observe that applying classifier-free guidance (CFG) before a certain time step tends to produce memorized samples; we refer to this time step as the transition point. Interestingly, applying CFG after the transition point is unlikely to yield a memorized image. Based on these observations, we propose the concept of an attraction basin, in which applying CFG induces a strong steering force toward a memorized output, even with different initializations. Furthermore, following the approach of Wen et al. [28], we analyze the magnitude of the conditional noise predictions and discover uncharacteristically large values within the attraction basin, providing empirical evidence of a strong steering force in this region. Once the transition point is passed, the strong steering force is no longer pervasive. This magnitude-based analysis motivates us to detect the transition point by tracking the magnitude. Specifically, we initially apply no guidance until a transition point occurs—marked by a sharp drop in the discrepancy between conditional and unconditional guidance—after which applying CFG leads to non-memorized images." + }, + { + "type": "text", + "bbox": [ + 0.531, + 0.389, + 0.908, + 0.405 + ], + "angle": 0, + "content": "We summarize our contributions in this paper as follows:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.405, + 0.908, + 0.48 + ], + "angle": 0, + "content": "- We present a new way to understand memorization in conditional diffusion models by understanding the diffusion trajectory from a dynamical systems theory perspective. We show that an attraction basin forms in the denoising trajectory." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.48, + 0.908, + 0.54 + ], + "angle": 0, + "content": "- We present approaches to avoiding the attraction basin in various scenarios that incur no additional computational costs and require neither access to the prompt nor modification of weights." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.541, + 0.907, + 0.585 + ], + "angle": 0, + "content": "- We present a new guidance technique that we call Opposite Guidance (not to be confused with negative prompts [3]) to push the trajectory away from the attraction basin." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.585, + 0.907, + 0.661 + ], + "angle": 0, + "content": "- We comprehensively show that previous approaches are not generalizable in mitigating memorization and only work in certain scenarios studied in those papers. On the contrary, our approach is able to mitigate memorization in all the explored scenarios." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.405, + 0.908, + 0.661 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.673, + 0.657, + 0.689 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.698, + 0.78, + 0.714 + ], + "angle": 0, + "content": "2.1. Understanding Memorization" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.72, + 0.909, + 0.902 + ], + "angle": 0, + "content": "In deep generative models, when memorization occurs, a generated sample can match verbatim with a single or set of training images (verbatim memorization) [4, 23, 24] or copy the same template as the training image (template memorization) [20] leading to privacy and/or copyright infringement issues. Researchers [4, 23, 24, 27] have identified that when images or prompts are duplicated in the training dataset, it can lead to memorization. Additionally, training on smaller datasets has also been shown to be a major factor in causing memorization [24]. However, as Somepalli et al. [24] pointed out, even though data has been de-duplicated in the newer models [16] such as Stable Diffusion v2.1 (SDv2.1)," + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "12872" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.212 + ], + "angle": 0, + "content": "the issue of memorization is yet to be resolved, and it does not explain much of the observed replication behavior. Factors such as conditioning and dataset complexity play a role even when the training datasets are large enough and are de-duplicated where \"simpler\" images and overly specific prompts have a higher chance of getting memorized [24]. Thus, there is a strong need for mitigation techniques beyond dataset manipulations." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.222, + 0.314, + 0.238 + ], + "angle": 0, + "content": "2.2. Detecting Memorization" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.245, + 0.485, + 0.579 + ], + "angle": 0, + "content": "Similarity scores based on the SSCD (self-supervised copy detection) model [18] can be used for getting an estimate on the extent of memorization wherein scores \\(>0.5\\) can suggest that it is memorized [23]. However, this metric is prone to false positives when the textual similarity is high between a generated and real image [24]. Wen et al. [28] showed that for memorized prompts, the difference between the text-conditioned and the unconditioned score predictions is uncharacteristically high and can be informative in detecting memorization. They demonstrated high performance even in the first step of noise prediction. Ren et al. [20], on the other hand, showed that for memorized prompts, the cross-attention scores pertaining to certain \"trigger tokens\" in the memorized prompts are unusually high. Suggesting that they are given more importance, and thus lead to memorization. Other authors [4, 27] used a black box membership inference attack for finding memorized images. Daras et al. [8] showed that when an image is memorized, the model is able to fully reconstruct it from a noised version of it. This observation could be used for detection. Another line of research has tried to understand memorization by comparing diffusion models to associative memory networks [1, 11, 17]." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.588, + 0.322, + 0.604 + ], + "angle": 0, + "content": "2.3. Mitigating Memorization" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.61, + 0.315, + 0.625 + ], + "angle": 0, + "content": "2.3.1. Training Time Mitigation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.629, + 0.484, + 0.84 + ], + "angle": 0, + "content": "Wen et al. [28] proposed monitoring the text-conditioned noise prediction scores for each sample and excluding it from the current mini-batch if it surpasses a certain predetermined threshold. On similar lines, Ren et al. [20] proposed removing samples from the mini-batch when their cross-attention entropy is above a particular predetermined threshold. Somepalli et al. [24] proposed generating multiple BLIP captions per image during training. They also proposed perturbing the prompts by adding Gaussian noise or adding/replacing random words/numbers. Daras et al. [8] proposed training/finetuning on corrupted images can reduce instances of data replication by the model. Chavhan et al. [6] proposed pruning the UNet model weights assuming prior knowledge of memorized prompts." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Most training time mitigations require re-training the base diffusion model from scratch. Not only is this computationally expensive, it also requires control over the entire training process including access to the training datasets." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.092, + 0.743, + 0.107 + ], + "angle": 0, + "content": "2.3.2. Inference Time Mitigation" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.11, + 0.907, + 0.321 + ], + "angle": 0, + "content": "Wen et al. [28] proposed perturbing the prompt embedding by minimizing the difference between the text-conditioned and unconditioned noise prediction scores such that it falls below a certain target loss. Ren et al. [20] proposed applying a mask and rescaling the attention scores to give less attention to \"trigger tokens\". However, this approach does not mitigate memorization when \"trigger tokens\" are not present. Somepalli et al. [24] proposed adding random tokens to the text prompt or repeating certain tokens during inference so as to add noise to the text embedding indirectly. Chen et al. [7] proposed different CFG weight schedulers to mitigate different causes of memorization. The approach presumes knowledge of not only whether memorization is present but also the type of memorization." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.322, + 0.909, + 0.458 + ], + "angle": 0, + "content": "Most of these approaches work by updating text prompt/embeddings leads to a trade-off between the text alignment and memorization. Another drawback of these approaches is that they work better when a trigger token is present such that we can reduce the importance/weightage of that token and they do not generalize well to other types of memorization as we show in this paper. Lastly, since they work directly on the text prompt/embeddings, they cannot be applied to other conditioning mechanisms such as class labels." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.469, + 0.652, + 0.485 + ], + "angle": 0, + "content": "3. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.495, + 0.906, + 0.585 + ], + "angle": 0, + "content": "Diffusion models [25] such as Stable Diffusion (SD) [21] and Imagen [22] are trained with the objective of learning a model \\(\\epsilon_{\\theta}\\) to denoise a noisy input vector at different levels of noise characterized by a time-dependent noise scheduler. We can define the diffusion process as a stochastic differential equation (SDE) in the form," + }, + { + "type": "equation", + "bbox": [ + 0.613, + 0.593, + 0.805, + 0.61 + ], + "angle": 0, + "content": "\\[\nd x _ {t} = f (x _ {t}, t) d t + g (t) d w _ {t}.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.617, + 0.907, + 0.661 + ], + "angle": 0, + "content": "where \\( x_{t} \\in X \\) is a sample at time \\( t \\), \\( f(x_{t}, t) \\) is the drift term, \\( g(t) \\) is the diffusion coefficient, and \\( w_{t} \\) is a standard Wiener process." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.663, + 0.906, + 0.738 + ], + "angle": 0, + "content": "During training, the forward diffusion process comprises a Markov chain with fixed time steps \\(T\\). Given a data point \\(x_0 \\sim q(x)\\), we iteratively add Gaussian noise with variance \\(\\beta_t\\) at each time step \\(t\\) to \\(x_{t-1}\\) to get \\(x_t\\) such that \\(x_T \\sim \\mathcal{N}(0, I)\\). This process can be expressed as," + }, + { + "type": "equation", + "bbox": [ + 0.514, + 0.746, + 0.911, + 0.765 + ], + "angle": 0, + "content": "\\[\nq (x _ {t} | x _ {t - 1}) = \\mathcal {N} (x _ {t}; \\sqrt {1 - \\beta_ {t}} x _ {t - 1}, \\beta_ {t} I), \\quad \\forall t \\in \\{1, \\dots , T \\}.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.534, + 0.771, + 0.82, + 0.786 + ], + "angle": 0, + "content": "We can get a closed-form expression of \\( x_{t} \\)," + }, + { + "type": "equation", + "bbox": [ + 0.618, + 0.793, + 0.907, + 0.81 + ], + "angle": 0, + "content": "\\[\nx _ {t} = \\sqrt {\\bar {\\alpha} _ {t}} x _ {0} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon_ {t}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.816, + 0.809, + 0.833 + ], + "angle": 0, + "content": "where \\(\\bar{\\alpha}_t = \\prod_{i=1}^t (1 - \\beta_i)\\) and \\(\\alpha_t = 1 - \\beta_t\\)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.833, + 0.906, + 0.877 + ], + "angle": 0, + "content": "During training, we learn the reverse process through a network \\(\\epsilon_{\\theta}\\) to iteratively denoise \\(x_{t}\\) by estimating the noise \\(\\epsilon_{t}\\) at each time step \\(t\\). The loss function is expressed as," + }, + { + "type": "equation", + "bbox": [ + 0.579, + 0.884, + 0.907, + 0.903 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathbb {E} _ {t \\in [ 1, T ], \\epsilon \\sim \\mathcal {N} (0, I)} \\left[ \\left\\| \\epsilon_ {t} - \\epsilon_ {\\theta} \\left(x _ {t}\\right) \\right\\| _ {2} ^ {2} \\right]. \\tag {2}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12873" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.483, + 0.123 + ], + "angle": 0, + "content": "Using the learned noise estimator network \\(\\epsilon_{\\theta}\\), we can compute the previous sample \\(x_{t - 1}\\) from \\(x_{t}\\) as follows," + }, + { + "type": "equation", + "bbox": [ + 0.162, + 0.133, + 0.484, + 0.166 + ], + "angle": 0, + "content": "\\[\nx _ {t - 1} = \\frac {1}{\\sqrt {\\alpha} _ {t}} \\left(x _ {t} - \\frac {1 - \\alpha_ {t}}{\\sqrt {1 - \\bar {\\alpha} _ {t}}} \\epsilon_ {\\theta} \\left(x _ {t}\\right)\\right). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.176, + 0.485, + 0.327 + ], + "angle": 0, + "content": "The learned noise estimator network \\(\\epsilon_{\\theta}\\) can be conditioned using a conditioning input such as text or class label, expressed as an embedding \\(e_p\\). Ho et al. [10] proposed classifier-free guidance (CFG) as a mechanism to guide the diffusion trajectory towards generating outputs that align with the conditioning. The trajectory is directed towards the conditional score predictions and away from the unconditional score predictions, where \\(s\\) controls the degree of adjustment and \\(e_{\\emptyset}\\) are empty prompt embeddings used for unconditional guidance." + }, + { + "type": "equation", + "bbox": [ + 0.139, + 0.339, + 0.484, + 0.376 + ], + "angle": 0, + "content": "\\[\n\\hat {\\epsilon} \\leftarrow \\epsilon_ {\\theta} \\left(x _ {t}, e _ {\\emptyset}\\right) + s \\underbrace {\\left(\\epsilon_ {\\theta} \\left(x _ {t} , e _ {p}\\right) - \\epsilon_ {\\theta} \\left(x _ {t} , e _ {\\emptyset}\\right)\\right)} _ {\\text {c o n d i t i o n a l g u i d a n c e}}. \\tag {4}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.387, + 0.416, + 0.423 + ], + "angle": 0, + "content": "4. Understanding Diffusion Trajectory during Memorization" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.43, + 0.484, + 0.52 + ], + "angle": 0, + "content": "In text-to-image diffusion models, different initializations generally lead to different outputs for the same text prompt. However, when the diffusion model has memorized a particular sample, the outputs are similar and closely resemble one or more training data samples regardless of initialized noise [28]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.522, + 0.484, + 0.703 + ], + "angle": 0, + "content": "As shown in Figure 2, we observe that applying CFG after a certain time step avoids producing memorized outputs, while applying it from earlier time steps, including at the beginning \\((t = T)\\), easily leads to memorized outputs. Inspired by these empirical findings and the perspective of dynamical systems [2, 15], we suspect that there is an attraction basin in the sample space of the diffusion trajectory, in which applying CFG results in memorized outputs. Intuitively, one can imagine this attraction basin as having a funnel-like shape. It is broader at \\( t = T \\) and narrows as the reverse denoising process progresses. To define the attraction basin, we introduce the concept of a denoiser." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.713, + 0.484, + 0.82 + ], + "angle": 0, + "content": "Definition 1 (Denoiser). Let \\( X \\) denote the entire sample space of the diffusion trajectory, and let \\( E \\) denote the embedding space. A denoiser is a function \\( \\varphi : X \\times (0, T] \\times E \\to X \\) that outputs the result of the reverse diffusion process based on Eq. (3) with CFG. Specifically, \\( \\varphi(x, t, e) \\) represents the resulting sample at time \\( t = 0 \\), starting from a state \\( (x, t) \\) with embeddings \\( e \\in E \\), where \\( x \\in X \\) and \\( t \\in (0, T] \\)." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.83, + 0.483, + 0.86 + ], + "angle": 0, + "content": "For an attractor \\( x^{\\mathrm{a}} \\) from the training dataset, we then define the attraction basin as follows:" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Definition 2 (Attractor and attraction basin). Suppose a prompt is given with embeddings denoted as \\( e_p \\). A sample" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.183 + ], + "angle": 0, + "content": "\\(x^{\\mathrm{a}} \\in X\\) is considered an \\((\\epsilon)\\)-attractor if the diffusion trajectories denoised with CFG tend to converge to samples within \\(B_{\\mathcal{D}}(x^{\\mathrm{a}}, \\epsilon)\\) at \\(t = 0\\), regardless of the initial noise, i.e., \\(\\varphi(x, T, e_p) \\in B_{\\mathcal{D}}(x^{\\mathrm{a}}, \\epsilon)\\) for any \\(x \\in X\\), where \\(B_{\\mathcal{D}}(\\cdot, \\cdot)\\) represents a perceptual ball given by a certain perceptual distance \\(\\mathcal{D}: X \\times X \\to \\mathbb{R}_{\\geq 0}\\) as" + }, + { + "type": "equation", + "bbox": [ + 0.574, + 0.197, + 0.845, + 0.214 + ], + "angle": 0, + "content": "\\[\nB _ {\\mathcal {D}} \\left(x ^ {\\mathrm {a}}, \\epsilon\\right) = \\left\\{x ^ {\\prime} \\in X \\mid \\mathcal {D} \\left(x ^ {\\mathrm {a}}, x ^ {\\prime}\\right) \\leq \\epsilon \\right\\}.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.228, + 0.907, + 0.289 + ], + "angle": 0, + "content": "An attraction basin is a region in the state space within which samples will converge to around \\( x^{\\mathrm{a}} \\) at \\( t = 0 \\) under CFG inference. More formally, the \\( ((\\epsilon, \\delta)) \\)-attraction basin defined using \\( x^{\\mathrm{a}} \\) is expressed as" + }, + { + "type": "equation", + "bbox": [ + 0.514, + 0.302, + 0.909, + 0.328 + ], + "angle": 0, + "content": "\\[\nX ^ {\\mathrm {b}} \\left(x ^ {\\mathrm {a}}, \\epsilon\\right) = \\left\\{\\left(x, t\\right) \\mid \\mathbb {P} \\left(\\varphi (x, t, e) \\in B _ {\\mathcal {D}} \\left(x ^ {\\mathrm {a}}, \\epsilon\\right)\\right) > 1 - \\delta \\right\\},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.341, + 0.908, + 0.385 + ], + "angle": 0, + "content": "representing all points in the state space from which generated samples using CFG will reach samples perceptually similar to the attractor at \\( t = 0 \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.4, + 0.909, + 0.627 + ], + "angle": 0, + "content": "Inspired from Wen et al. [28], we analyze the attraction basin by examining the magnitudes of the conditional guidance, \\(\\epsilon_{\\theta}(x_t,e_p) - \\epsilon_{\\theta}(x_t,e_{\\emptyset})\\). Interestingly, as shown in Figure 2, the conditional noise prediction \\(\\epsilon_{\\theta}(x_t,e_p) - \\epsilon_{\\theta}(x_t,e_{\\emptyset})\\) remains high during the initial time steps and then steeply falls at the time step immediately before the denoising trajectory exits the attraction basin. We refer to this time step as the transition point, marking the change from producing memorized outputs to non-memorized outputs. In Figure 2(a), if you were to start applying CFG at any point on or before \\(t = 779\\), you get the same memorized output. However, surprisingly, applying CFG at exactly one denoising step after this \\(t = 759\\) leads to a non-memorized output (step size is 20 when denoising with 50 inference steps). We define the transition point as follows." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.641, + 0.907, + 0.717 + ], + "angle": 0, + "content": "Definition 3 (Transition point). Given prompt embeddings, let \\( x^{\\mathrm{a}} \\in X \\) be an attractor. We define a transition point as \\( (x_{\\tau}, \\tau) \\) for \\( x_{\\tau} \\in X \\) and \\( \\tau > 0 \\) such that \\( (x_{\\tau}, \\tau + 1) \\) lies within its attraction basin, while \\( (x_{\\tau}, \\tau) \\) does not. More formally, an \\( ((\\epsilon, \\delta)) \\)-transition point satisfies" + }, + { + "type": "equation", + "bbox": [ + 0.561, + 0.731, + 0.86, + 0.748 + ], + "angle": 0, + "content": "\\[\n\\mathbb {P} \\left(\\varphi \\left(x _ {\\tau}, \\tau + 1\\right) \\in B _ {\\mathcal {D}} \\left(x ^ {\\mathrm {a}}, \\epsilon\\right)\\right) > 1 - \\delta \\quad a n d\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.562, + 0.75, + 0.794, + 0.767 + ], + "angle": 0, + "content": "\\[\n\\mathbb {P} (\\varphi (x _ {\\tau}, \\tau) \\notin B _ {\\mathcal {D}} (x ^ {\\mathrm {a}}, \\epsilon)) < 1 - \\delta ,\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.781, + 0.868, + 0.796 + ], + "angle": 0, + "content": "where \\(B_{\\mathcal{D}}\\) is a perceptual ball defined in Definition 2." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.906, + 0.902 + ], + "angle": 0, + "content": "This phenomenon occurs only for memorized samples as illustrated in Figure 6, the conditional guidance \\(\\epsilon_{\\theta}(x_t,e_p) - \\epsilon_{\\theta}(x_t,e_{\\emptyset})\\) remains low and almost flat across different time steps for non-memorized samples. We further analyze the diffusion trajectory during memorization, demonstrating the presence of an attraction basin in Appendix 8." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12874" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.09, + 0.342, + 0.109 + ], + "angle": 0, + "content": "5. Simple Mitigation Strategy" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.116, + 0.484, + 0.192 + ], + "angle": 0, + "content": "Based on the intuition from the previous section, we can simply mitigate memorization by finding an ideal transition point to switch from zero CFG to applying CFG. Based on our preliminary experiments, we found that transition points can either be static or dynamic." + }, + { + "type": "image", + "bbox": [ + 0.152, + 0.207, + 0.426, + 0.342 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.348, + 0.485, + 0.418 + ], + "angle": 0, + "content": "Figure 3. Average magnitude of the text-conditioned \\((\\epsilon_{\\theta}(x_t, e_p))\\) and unconditional noise predictions \\((\\epsilon_{\\theta}(x_t, e_{\\emptyset}))\\) and their difference \\((\\epsilon_{\\theta}(x_t, e_p) - \\epsilon_{\\theta}(x_t, e_{\\emptyset}))\\) when applying zero CFG. We see a static transition point (\\(t = 500\\)) appear when SDv2.1 is finetuned on the LAION-10k dataset [24]." + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.449, + 0.485, + 0.637 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.64, + 0.484, + 0.71 + ], + "angle": 0, + "content": "Figure 4. Applying CFG before the static transition point \\((\\mathrm{T} = 500)\\) leads to memorized outputs while applying CFG after the fixed transition point leads to non-memorized outputs. Applying CFG too late results in poor-quality images that resemble the unconditional generations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.486, + 0.903 + ], + "angle": 0, + "content": "Static Transition Point. In some trained diffusion models, we observed a universal trend in the mean conditional \\(\\epsilon_{\\theta}(x_t,e_p)\\) and unconditional noise prediction \\(\\epsilon_{\\theta}(x_t,e_{\\emptyset})\\) across different samples. As shown in Figure 3, the mean conditional noise prediction \\(\\epsilon_{\\theta}(x_t,e_p)\\) is higher than unconditional noise prediction \\(\\epsilon_{\\theta}(x_t,e_{\\emptyset})\\) in the initial time steps \\((t\\geq 500)\\). At a static time step \\((t = 500)\\), the two scores align in magnitude until the end of the denoising process, and at the same time, the \\(L_{2}\\) distance between them drastically drops. This suggests that the zero CFG trajectory leaves the attraction basin at that fixed time step \\((t = 500)\\)." + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.091, + 0.905, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.292, + 0.908, + 0.348 + ], + "angle": 0, + "content": "Figure 5. In some models, transition points can occur at a different time step for each sample, as seen for pre-trained SDv1.4. In row 1 the transition point is approximately \\( t = 800 \\) while for row 2 it is \\( t = 650 \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.375, + 0.909, + 0.512 + ], + "angle": 0, + "content": "This can be further validated by applying CFG from any time step before the fixed transition point \\((t = 500)\\), as illustrated in Figure 4. Applying CFG from an early time step leads to the same memorized image, but the output immediately switches at the transition point \\((t = 500)\\) such that we generate text-aligned non-memorized images. On the other hand, if we start applying CFG very late in the denoising process, the outputs are not well aligned with the text prompts and resemble their unconditional generations." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.52, + 0.909, + 0.793 + ], + "angle": 0, + "content": "Dynamic Transition Point. In other trained diffusion models, we saw that every prompt and initialization pair leads to a different transition point. As shown in Figures 6(a) and 5, different trajectories leave the attraction basin at different time steps, characterized by their fall in the conditional guidance \\(\\epsilon_{\\theta}(x_t,e_p) - \\epsilon_{\\theta}(x_t,e_{\\emptyset})\\). We found that it is dependent on both the conditional embedding \\(e_p\\) and the initialization \\(x_{T}\\). We show an example in Figures 2(a) and 2(b), where the transition point occurs at different time steps even for the same prompt. Thus, we propose a method to find this dynamic transition point (DTP). As shown in Figure 2, it is the point after the first local minima in the graph. Thus, to mitigate memorization, this difference is tracked until the first local minimum occurs, and CFG is applied henceforth. This process requires no additional computations as the text-conditioned and unconditional noise predictions are computed by default when denoising using CFG. We present the pseudo-code for this approach in Algorithm 1." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.802, + 0.698, + 0.819 + ], + "angle": 0, + "content": "5.1. Opposite Guidance" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.907, + 0.886 + ], + "angle": 0, + "content": "An issue with this method is that when the transition point occurs very late (\\(t \\leq 500\\)) in the denoising process, CFG is applied for fewer time steps, and this can impact the image quality." + }, + { + "type": "text", + "bbox": [ + 0.534, + 0.886, + 0.909, + 0.902 + ], + "angle": 0, + "content": "To ensure that the transition point occurs earlier, we in" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12875" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.152, + 0.09, + 0.422, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.225, + 0.228, + 0.351, + 0.242 + ], + "angle": 0, + "content": "(a) Applying zero CFG" + }, + { + "type": "image", + "bbox": [ + 0.153, + 0.244, + 0.422, + 0.377 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.202, + 0.382, + 0.372, + 0.395 + ], + "angle": 0, + "content": "(b) Applying opposite guidance" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.4, + 0.484, + 0.442 + ], + "angle": 0, + "content": "Figure 6. Magnitude of \\(\\epsilon_{\\theta}(x_t, e_p) - \\epsilon_{\\theta}(x_t, e_{\\emptyset})\\) when denoising (a) without CFG and (b) with opposite guidance. Transition point occurs sooner with opposite guidance." + }, + { + "type": "code_caption", + "bbox": [ + 0.09, + 0.458, + 0.484, + 0.488 + ], + "angle": 0, + "content": "Algorithm 1 Reverse Diffusion with Dynamic Transition Point Method and Opposite Guidance" + }, + { + "type": "algorithm", + "bbox": [ + 0.092, + 0.492, + 0.484, + 0.663 + ], + "angle": 0, + "content": "Require: \\(x_{T}\\sim \\mathcal{N}(0,\\mathrm{I}_{d}),\\lambda >0,v_{\\mathrm{OG}}\\in \\{0,1\\}\\) \n1: \\(d_{T + 2} = -\\infty\\) . \\(d_{T + 1} = -\\infty\\) \n2: \\(s = -\\lambda *v_{\\mathrm{OG}}\\) \n3: for \\(t = T\\) to 1 do \n4: \\(d_{t} = \\| \\epsilon_{\\theta}(x_{t},e_{p}) - \\epsilon_{\\theta}(x_{t},e_{\\emptyset})\\|_{2}^{2}\\) \n5: if \\(d_{t + 2} > d_{t + 1}\\) and \\(d_{t + 1} < d_t\\) then \n6: \\(s = \\lambda\\) \n7: end if \n8: \\(\\hat{\\epsilon} = \\epsilon_{\\theta}(x_t,e_\\emptyset) + s(\\epsilon_\\theta (x_t,e_p) - \\epsilon_\\theta (x_t,e_\\emptyset))\\) \n9: \\(x_{t - 1} = \\frac{1}{\\sqrt{\\alpha}_t} (x_t - \\frac{1 - \\alpha_t}{\\sqrt{1 - \\alpha_t}}\\hat{\\epsilon})\\) \n10: end for \n11: return \\(x_0\\)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.69, + 0.486, + 0.797 + ], + "angle": 0, + "content": "troduce a new concept of opposite guidance (OG). Unlike negative prompting, we apply negative or opposite CFG, to push the diffusion trajectory in the opposite direction to that of traditional CFG and thus, push the trajectory away from the attraction basin of the traditional CFG trajectory sooner in the denoising process, as also shown in Figure 6(b). Opposite guidance can be expressed as follows for \\( s > 0 \\)," + }, + { + "type": "equation", + "bbox": [ + 0.139, + 0.823, + 0.484, + 0.861 + ], + "angle": 0, + "content": "\\[\n\\hat {\\epsilon} \\leftarrow \\epsilon_ {\\theta} \\left(x _ {t}, e _ {\\emptyset}\\right) - s \\underbrace {\\left(\\epsilon_ {\\theta} \\left(x _ {t} , e _ {p}\\right) - \\epsilon_ {\\theta} \\left(x _ {t} , e _ {\\emptyset}\\right)\\right)} _ {\\text {c o n d i t i o n a l g u i d a n c e}}. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.485, + 0.902 + ], + "angle": 0, + "content": "Similar to unconditional denoising, when denoising using opposite guidance, there is an ideal time step where" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.213 + ], + "angle": 0, + "content": "the magnitude of conditional guidance falls. At this time step, switching from opposite guidance to traditional positive guidance leads to non-memorized samples with high image quality and text alignment. Thus, it can easily be integrated along with the previous method to find a static or dynamic transition point. We present the pseudo-code for this approach in conjunction with the DTP method in Algorithm 1 with the parameter \\( v_{\\mathrm{OG}} = 1 \\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.227, + 0.879, + 0.243 + ], + "angle": 0, + "content": "6. Memorization under Different Scenarios" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.252, + 0.907, + 0.313 + ], + "angle": 0, + "content": "We study memorization under different scenarios and show how this simple approach of finding an ideal transition point to avoid the attraction basin can be applied to effectively mitigate memorization." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.315, + 0.907, + 0.36 + ], + "angle": 0, + "content": "- Scenario 1: Fine-tuned SDv2.1 on 10,000 LAION datapoints (followed by [24]). Memorization occurs due to overfitting on the small dataset." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.361, + 0.907, + 0.421 + ], + "angle": 0, + "content": "- Scenario 2: Fine-tuned SDv2.1 on the Imagenette dataset [12] containing 10 ImageNet classes (followed by [24]). Exact memorization is not always observed but similarity with the training dataset increases." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.422, + 0.906, + 0.481 + ], + "angle": 0, + "content": "- Scenario 3: Fine-tuned SDv1.4 on 200 prompts duplicated 200 times with an additional 120,000 prompts that were not duplicated (followed by [28]). Memorization occurs due to data duplication." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.482, + 0.907, + 0.588 + ], + "angle": 0, + "content": "- Scenario 4: 500 memorized prompts for pre-trained SDv1.4 found using membership inference attack [4, 27] (followed by [20]). The exact cause of memorization for each prompt is unknown, but the strong presence of \"trigger tokens\" is noticeable. These are in the form of nouns such as celebrity and movie names with little to no additional descriptions. Results are in Appendix 9." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.315, + 0.907, + 0.588 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.59, + 0.909, + 0.667 + ], + "angle": 0, + "content": "Notably, different researchers have independently studied these different scenarios, but, as we show later, their approaches do not generalize to other scenarios as memorization occurs due to different causes in each scenario. Refer to Appendix 11 for other implementation details." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.672, + 0.909, + 0.749 + ], + "angle": 0, + "content": "Evaluation Metrics We utilize the similarity metric based on the SSCD [18] embeddings to judge the level of memorization [24]; the CLIP score [19] to assess how well the generated images align with the text prompts; and the Fréchet Inception Distance (FID) [9] for image quality." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.758, + 0.631, + 0.773 + ], + "angle": 0, + "content": "6.1. Scenario 1" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.781, + 0.909, + 0.901 + ], + "angle": 0, + "content": "We observed that in this scenario, a static transition point at \\( t = 500 \\) occurs as shown in Figure 3. We present qualitative and quantitative results in Figure 7 and Table 1, respectively. Our approach outperformed all previous methods, and we observed the previously proposed techniques that focused solely on the presence of \"trigger tokens\" [20, 28] did not work well. This shows that when memorization occurs due to factors other than data duplication or the presence of" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "12876" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.236, + 0.088, + 0.768, + 0.442 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.444, + 0.908, + 0.488 + ], + "angle": 0, + "content": "Figure 7. Qualitative results comparing the proposed approach with the baselines in Scenario 1. Ren et al. [20] is unable to mitigate memorization in most cases while Wen et al. [28] has poor image quality and text-alignment. More enlarged examples (Fig. 14) and prompts for this figure are in Appendix 12." + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.511, + 0.486, + 0.568 + ], + "angle": 0, + "content": "Table 1. Comparison of results on Scenario 1. None of the other approaches perform well. The only comparable method [28] qualitatively results in poorer quality images (see visual examples in Fig. 7)." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.571, + 0.484, + 0.679 + ], + "angle": 0, + "content": "
Similarity (95pc)CLIP ScoreFID
No Mitigation0.65040.302716.8373
Add rand word [24]0.52540.294117.4142
Add rand numb [24]0.54160.299317.0245
Wen et al. [28]0.38530.289516.7176
Ren et al. [20]0.60280.295920.2931
Ours (STP)0.28570.297619.8494
Ours (OG + STP)0.38110.302015.6679
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.703, + 0.484, + 0.749 + ], + "angle": 0, + "content": "trigger tokens, these methods may fail. We present further results when finetuning on the LAION-100k dataset [24] in Appendix 10." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.758, + 0.21, + 0.773 + ], + "angle": 0, + "content": "6.2. Scenario 2" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Even in this scenario, as illustrated in Figure 10, a static transition point occurs. We applied CFG from \\( t = 700 \\) and \\( t = 600 \\) to show that we can reduce the similarity with the training dataset at no cost to the image quality, as shown in Table 2. We would like to point out that previous approaches are specific to text-to-image models and thus cannot be extrapolated to other conditioning mechanisms. This shows a further advantage of our CFG-based mitigation" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.514, + 0.587, + 0.529 + ], + "angle": 0, + "content": "technique." + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.542, + 0.907, + 0.598 + ], + "angle": 0, + "content": "Table 2. Results on Scenario 2. The value in the brackets in the time steps between which we apply CFG. Applying CFG later from \\( t = 700 \\) (row 2) and \\( t = 600 \\) (row 3) reduces similarity with comparable or improved FID score." + }, + { + "type": "table", + "bbox": [ + 0.545, + 0.602, + 0.877, + 0.668 + ], + "angle": 0, + "content": "
Similarity (95pc)FID
No mitigation0.370243.35
CFG=7.5 [700,0]0.301838.86
CFG=7.5 [600,0]0.275645.73
" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.692, + 0.633, + 0.707 + ], + "angle": 0, + "content": "6.3. Scenario 3" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.714, + 0.907, + 0.851 + ], + "angle": 0, + "content": "In this scenario, we observed that a dynamic transition point occurs. We experimentally validate our approach of DTP method and opposite guidance qualitatively and quantitatively in Figure 8 and Table 3. We saw that our method performs at par with [28] that studied this specific scenario and even outperforms it in terms of memorization in the top 95 percentile. Approaches by Ren et al. [20] and Somepalli et al.[24] that were proposed for other scenarios do not work well, showcasing their lack of generalizability." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Inference Time. One of the major advantages of our approach is its simplicity and lack of computational time overhead. Image generation using our approach as well as stan-" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12877" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.237, + 0.089, + 0.768, + 0.446 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.451, + 0.908, + 0.494 + ], + "angle": 0, + "content": "Figure 8. Qualitative results comparing the proposed approach with the baselines in Scenario 3. Prompts used to generate these figures are given in the Appendix. Rows 1, 2, 5, and 6 show examples, where for most baselines the output images remain closely related to their memorized versions. More enlarged examples (Fig. 15) and prompts for this figure are in Appendix 13." + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.517, + 0.483, + 0.56 + ], + "angle": 0, + "content": "Table 3. Table showcasing results in Scenario 3. Wen et al. [28] studied this scenario and is the only other method which performs well but does not generalize to other scenarios (see Table 1)." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.564, + 0.485, + 0.68 + ], + "angle": 0, + "content": "
SimilarityCLIP ScoreFID
95pcMean ± Std
No Mitigation0.79770.5513 ± 0.160.3105106.49
Add rand word [24]0.74800.4312 ± 0.210.3071116.92
Add rand numb [24]0.73660.3850 ± 0.210.3027126.54
Wen et al. [28] (l_target=3)0.77470.5147 ± 0.170.3100109.28
Wen et al. [28] (l_target=1)0.60380.2808 ± 0.160.3050136.34
Ren et al. [20]0.68810.4036 ± 0.160.3066124.38
Ours (DTP)0.58850.2866 ± 0.160.3020138.92
Ours (OG + DTP)0.69150.2844 ± 0.200.2910140.05
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.706, + 0.484, + 0.783 + ], + "angle": 0, + "content": "dard CFG-based image generation using SDv2.1 on one A100 GPU takes 1.26 seconds, while in comparison the mitigation technique proposed by Wen et al. [28] took 2.86 seconds and that of Ren et al. [20] took 2.08 seconds. These are the average values across 10,000 generations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Limitations. As transition points do not exist for nonmemorized samples, we require detecting memorization before we can apply our approach. This can either be done by denoising twice, once with traditional CFG to look at the trends in the magnitude of conditional guidance and the second time to apply our mitigation technique if the prompt is memorized. Alternatively, as [20, 28] showed, we can" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.519, + 0.907, + 0.596 + ], + "angle": 0, + "content": "do detection at \\( t = 0 \\) by simply looking at the magnitude of conditional guidance \\( \\| \\epsilon_{\\theta}(x_t, e_p) - \\epsilon_{\\theta}(x_t, e_{\\emptyset}) \\|_2^2 \\). If this magnitude is above a certain threshold, it indicates that the sample is memorized with high accuracy and we can apply our guidance accordingly." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.624, + 0.634, + 0.64 + ], + "angle": 0, + "content": "7. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.655, + 0.909, + 0.837 + ], + "angle": 0, + "content": "We present a novel understanding of the dynamics behind memorization in conditional diffusion models. We showcase the presence of an attraction basin that steers randomly initialized latent vectors towards a memorized output. We propose an approach to steer away from the attraction basin and detect the point at which the trajectory leaves the attraction basin, referred to as the transition point. We show that applying CFG after the transition point leads to non-memorized outputs. We presented results in various scenarios such as the presence of data replication, fine-tuning on smaller datasets, and memorized examples found in existing pre-trained models to showcase its efficiency." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.909, + 0.903 + ], + "angle": 0, + "content": "To summarize, our method is simple, works well across all tested scenarios in terms of both image quality metrics and visual inspection, requires less computational time, and does not require altering the prompt." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12878" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.116, + 0.486, + 0.157 + ], + "angle": 0, + "content": "[1] Luca Ambrogioni. In search of dispersed memories: Generative diffusion models are associative memory networks. Entropy, 26(5):381, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.159, + 0.483, + 0.186 + ], + "angle": 0, + "content": "[2] Joseph Auslander, Nam P Bhatia, and Peter Seibert. Attractors in dynamical systems. Technical report, 1964. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.189, + 0.483, + 0.243 + ], + "angle": 0, + "content": "[3] Yuanhao Ban, Ruochen Wang, Tianyi Zhou, Minhao Cheng, Boqing Gong, and Cho-Jui Hsieh. Understanding the impact of negative prompts: When and how do they take effect? arXiv preprint arXiv:2406.02965, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.246, + 0.483, + 0.315 + ], + "angle": 0, + "content": "[4] Nicolas Carlini, Jamie Hayes, Milad Nasr, Matthew Jagielski, Vikash Sehwag, Florian Tramer, Borja Balle, Daphne Ippolito, and Eric Wallace. Extracting training data from diffusion models. In 32nd USENIX Security Symposium (USENIX Security 23), pages 5253-5270, 2023. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.317, + 0.484, + 0.386 + ], + "angle": 0, + "content": "[5] Huiwen Chang, Han Zhang, Jarred Barber, AJ Maschinot, Jose Lezama, Lu Jiang, Ming-Hsuan Yang, Kevin Murphy, William T Freeman, Michael Rubinstein, et al. Muse: Text-to-image generation via masked generative transformers. arXiv preprint arXiv:2301.00704, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.388, + 0.483, + 0.443 + ], + "angle": 0, + "content": "[6] Ruchika Chavhan, Ondrej Bohdal, Yongshuo Zong, Da Li, and Timothy Hospedales. Memorized images in diffusion models share a subspace that can be located and deleted. arXiv preprint arXiv:2406.18566, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.445, + 0.484, + 0.5 + ], + "angle": 0, + "content": "[7] Chen Chen, Daochang Liu, and Chang Xu. Towards memorization-free diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8425-8434, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.502, + 0.483, + 0.557 + ], + "angle": 0, + "content": "[8] Giannis Daras, Alexandros G Dimakis, and Constantinos Daskalakis. Consistent diffusion meets tweedie: Training exact ambient diffusion models with noisy data. arXiv preprint arXiv:2404.10177, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.56, + 0.484, + 0.627 + ], + "angle": 0, + "content": "[9] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.631, + 0.482, + 0.658 + ], + "angle": 0, + "content": "[10] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.661, + 0.483, + 0.728 + ], + "angle": 0, + "content": "[11] Benjamin Hoover, Hendrik Strobelt, Dmitry Krotov, Judy Hoffman, Zsolt Kira, and Duen Horng Chau. Memory in plain sight: Surveying the uncanny resemblances of associative memories and diffusion models. arXiv preprint arXiv:2309.16750, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.732, + 0.482, + 0.759 + ], + "angle": 0, + "content": "[12] Jeremy Howard. Imagenette: A smaller subset of 10 easily classified classes from imagenet, 2019. 6, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.761, + 0.483, + 0.829 + ], + "angle": 0, + "content": "[13] Tuomas Kynkänniemi, Miika Aittala, Tero Karras, Samuli Laine, Timo Aila, and Jaakko Lehtinen. Applying guidance in a limited interval improves sample and distribution quality in diffusion models. arXiv preprint arXiv:2404.07724, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.832, + 0.484, + 0.899 + ], + "angle": 0, + "content": "[14] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps. Advances in Neural Information Processing Systems, 35:5775-5787, 2022. 3" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.116, + 0.486, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.119 + ], + "angle": 0, + "content": "[15] John W. Milnor. Attractor. *Scholarpedia*, 1(11):1815, 2006. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.122, + 0.908, + 0.162 + ], + "angle": 0, + "content": "[16] Alex Nichol. Dall-e 2 pre-training mitigations. https://openai.com/index/dall-e-2-pre-training-mitigations/, 2022. [Accessed 10-10-2024]. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.164, + 0.907, + 0.218 + ], + "angle": 0, + "content": "[17] Bao Pham, Gabriel Raya, Matteo Negri, Mohammed J Zaki, Luca Ambrogioni, and Dmitry Krotov. Memorization to generalization: The emergence of diffusion models from associative memory. 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.22, + 0.907, + 0.287 + ], + "angle": 0, + "content": "[18] Ed Pizzi, Sreya Dutta Roy, Sugosh Nagavara Ravindra, Priya Goyal, and Matthijs Douze. A self-supervised descriptor for image copy detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14532-14542, 2022. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.289, + 0.907, + 0.37 + ], + "angle": 0, + "content": "[19] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.372, + 0.907, + 0.427 + ], + "angle": 0, + "content": "[20] Jie Ren, Yaxin Li, Shenglai Zen, Han Xu, Lingjuan Lyu, Yue Xing, and Jiliang Tang. Unveiling and mitigating memorization in text-to-image diffusion models through cross attention. arXiv preprint arXiv:2403.11052, 2024. 1, 2, 3, 6, 7, 8, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.429, + 0.907, + 0.496 + ], + "angle": 0, + "content": "[21] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.498, + 0.907, + 0.579 + ], + "angle": 0, + "content": "[22] Chitwan Saharia, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural information processing systems, 35:36479-36494, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.582, + 0.907, + 0.65 + ], + "angle": 0, + "content": "[23] Gowthami Somepalli, Vasu Singla, Micah Goldblum, Jonas Geiping, and Tom Goldstein. Diffusion art or digital forgery? investigating data replication in diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6048-6058, 2023. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.651, + 0.906, + 0.705 + ], + "angle": 0, + "content": "[24] Gowthami Somepalli, Vasu Singla, Micah Goldblum, Jonas Geiping, and Tom Goldstein. Understanding and mitigating copying in diffusion models. Advances in Neural Information Processing Systems, 36:47783-47803, 2023. 1, 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.707, + 0.907, + 0.747 + ], + "angle": 0, + "content": "[25] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.749, + 0.907, + 0.802 + ], + "angle": 0, + "content": "[26] Xi Wang, Nicolas Dufour, Nefeli Andreou, Marie-Paule Cani, Victoria Fernández Abrevaya, David Picard, and Vicky Kalogeiton. Analysis of classifier-free guidance weight schedulers. arXiv preprint arXiv:2404.13040, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.804, + 0.907, + 0.844 + ], + "angle": 0, + "content": "[27] Ryan Webster. A reproducible extraction of training images from diffusion models. arXiv preprint arXiv:2305.08694, 2023. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.907, + 0.901 + ], + "angle": 0, + "content": "[28] Yuxin Wen, Yuchen Liu, Chen Chen, and Lingjuan Lyu. Detecting, explaining, and mitigating memorization in diffusion models. In The Twelfth International Conference on Learning Representations, 2024. 1, 2, 3, 4, 6, 7, 8, 5" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "12879" + } + ] +] \ No newline at end of file diff --git a/2025/Classifier-Free Guidance Inside the Attraction Basin May Cause Memorization/a1a30dde-51b7-469a-a404-0c4a54031538_origin.pdf b/2025/Classifier-Free Guidance Inside the Attraction Basin May Cause Memorization/a1a30dde-51b7-469a-a404-0c4a54031538_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..3d1f66d8b2992cb04709669e28faa34eab01f24c --- /dev/null +++ b/2025/Classifier-Free Guidance Inside the Attraction Basin May Cause Memorization/a1a30dde-51b7-469a-a404-0c4a54031538_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8d8405e8152f5b40c897e2d53cac31b4bb3423eb5191c0dd3516108e07363311 +size 8597555 diff --git a/2025/Classifier-Free Guidance Inside the Attraction Basin May Cause Memorization/full.md b/2025/Classifier-Free Guidance Inside the Attraction Basin May Cause Memorization/full.md new file mode 100644 index 0000000000000000000000000000000000000000..e6e9d704a3fc79469e3ea10c37dad5ba8b864860 --- /dev/null +++ b/2025/Classifier-Free Guidance Inside the Attraction Basin May Cause Memorization/full.md @@ -0,0 +1,301 @@ +# Classifier-Free Guidance inside the Attraction Basin May Cause Memorization + +Anubhav Jain1*, Yuya Kobayashi2, Takashi Shibuya2, Yuhta Takida2, Nasir Memon1, Julian Togelius1, Yuki Mitsufuji2,3 + +$^{1}$ New York University, $^{2}$ Sony AI, $^{3}$ Sony Group Corporation + +{aj3281,memon,julian.togelius}@nyu.edu + +{u.kobayashi,takashi.tak.shibuya,yuta.takida,yuhki.mitsufuji}@sony.com + +# Abstract + +Diffusion models are prone to exactly reproduce images from the training data. This exact reproduction of the training data is concerning as it can lead to copyright infringement and/or leakage of privacy-sensitive information. In this paper, we present a novel perspective on the memorization phenomenon and propose a simple yet effective approach to mitigate it. We argue that memorization occurs because of an attraction basin in the denoising process which steers the diffusion trajectory towards a memorized image. However, this can be mitigated by guiding the diffusion trajectory away from the attraction basin by not applying classifier-free guidance until an ideal transition point occurs from which classifier-free guidance is applied. This leads to the generation of non-memorized images that are high in image quality and well-aligned with the conditioning mechanism. To further improve on this, we present a new guidance technique, opposite guidance, that escapes the attraction basin sooner in the denoising process. We demonstrate the existence of attraction basins in various scenarios in which memorization occurs, and we show that our proposed approach successfully mitigates memorization. Our codebase is publicly available at https://github.com/SonyResearch/mitigating_memorization. + +# 1. Introduction + +Recent advancements in image-generation models have enabled the generation of hyper-realistic images. Diffusion models [21, 25] have been at the forefront of this advancement with their ability to align image generation with different conditioning mechanisms such as text giving users more control over the output. However, a key challenge has emerged: these models can memorize and reproduce training examples verbatim (verbatim memorization) [4, 23, 24], leading to concerns over copyright infringement and leak + +![](images/f959491e85289c4f63b35fd8f3ab1bad326e94efd43f273c778e7621afd9b104.jpg) +Figure 1. The diffusion trajectory contains an attraction basin (red region) which steers conditioned samples towards their memorized images. It can be avoided by applying zero classifier-free guidance when the trajectory is inside the attraction basin, such that there is an ideal transition point $\tau^{*}$ after which applying CFG leads to non-memorized output. Applying CFG at an earlier point such as $\tau_{1}$ inside the attraction basin leads to the same memorized sample. + +age of privacy sensitive information. Alternatively, the model can generate images copying the composition and structure present in one or more training images (template memorization)[20]. Researchers have identified factors contributing to memorization and reproduction, such as the duplication of training images and repeated captions [24]. However, not all causes are understood, as memorization can still occur even when these issues are avoided. Therefore, there is a growing need for effective mitigation strategies, in particular methods that can be applied during inference without requiring computationally intensive retraining of the core diffusion models. + +In text-to-image (T2I) diffusion models, strong associations between a text prompt and an image in the training dataset lead to memorization. This can occur due to the presence of "trigger tokens" in overly specific text-prompts [20, 28], the model seeing multiple occurrences of these dur + +![](images/d81ca9c631ca872b93e29fdcc7c5c867306e0a9a4ee22644eea158e8cadf89fa.jpg) +(a) Prompt: "Here's What You Need to Know About St. Vincent's Apple Music Radio Show" + +![](images/3c89e43d47393742467f6e21b22d3314396f0c775e181c7f42959f7e3c64e3c2.jpg) +(b) Prompt: "Here's What You Need to Know About St. Vincent's Apple Music Radio Show" +Figure 2. Plots showing magnitude of $\epsilon_{\theta}(x_t, e_p) - \epsilon_{\theta}(x_t, e_{\emptyset})$ when denoising without classifier-free guidance (CFG) at each time step. The figures show the generated image if you start applying CFG at that time step. We get non-memorized output if we apply CFG after the ideal transition point $\tau^*$ which coincides with the fall in the conditional noise prediction. This value is dependent on both the prompt and the initialization ((a) and (b) contain the same prompt with different initializations). More examples in the Appendix. + +ing training (data duplication) [24, 28], or other causes such as fine-tuning on smaller datasets [24]. Previous approaches have focused on perturbing either the text prompt/embeddings [24, 28] or the cross-attention between the text and image embeddings [20]. While these approaches help weaken the associations between prompts and corresponding images, they are limited to specific conditions in which they have been tested and fail to generalize effectively. Additionally, since these techniques work at the prompt level, they are only applicable to text-conditional generation, and also assume the ability to update the prompt. + +In this paper, we introduce a new concept to explain the behavior of memorization in diffusion models based on our empirical observations. We first observe that applying classifier-free guidance (CFG) before a certain time step tends to produce memorized samples; we refer to this time step as the transition point. Interestingly, applying CFG after the transition point is unlikely to yield a memorized image. Based on these observations, we propose the concept of an attraction basin, in which applying CFG induces a strong steering force toward a memorized output, even with different initializations. Furthermore, following the approach of Wen et al. [28], we analyze the magnitude of the conditional noise predictions and discover uncharacteristically large values within the attraction basin, providing empirical evidence of a strong steering force in this region. Once the transition point is passed, the strong steering force is no longer pervasive. This magnitude-based analysis motivates us to detect the transition point by tracking the magnitude. Specifically, we initially apply no guidance until a transition point occurs—marked by a sharp drop in the discrepancy between conditional and unconditional guidance—after which applying CFG leads to non-memorized images. + +We summarize our contributions in this paper as follows: + +- We present a new way to understand memorization in conditional diffusion models by understanding the diffusion trajectory from a dynamical systems theory perspective. We show that an attraction basin forms in the denoising trajectory. +- We present approaches to avoiding the attraction basin in various scenarios that incur no additional computational costs and require neither access to the prompt nor modification of weights. +- We present a new guidance technique that we call Opposite Guidance (not to be confused with negative prompts [3]) to push the trajectory away from the attraction basin. +- We comprehensively show that previous approaches are not generalizable in mitigating memorization and only work in certain scenarios studied in those papers. On the contrary, our approach is able to mitigate memorization in all the explored scenarios. + +# 2. Related Work + +# 2.1. Understanding Memorization + +In deep generative models, when memorization occurs, a generated sample can match verbatim with a single or set of training images (verbatim memorization) [4, 23, 24] or copy the same template as the training image (template memorization) [20] leading to privacy and/or copyright infringement issues. Researchers [4, 23, 24, 27] have identified that when images or prompts are duplicated in the training dataset, it can lead to memorization. Additionally, training on smaller datasets has also been shown to be a major factor in causing memorization [24]. However, as Somepalli et al. [24] pointed out, even though data has been de-duplicated in the newer models [16] such as Stable Diffusion v2.1 (SDv2.1), + +the issue of memorization is yet to be resolved, and it does not explain much of the observed replication behavior. Factors such as conditioning and dataset complexity play a role even when the training datasets are large enough and are de-duplicated where "simpler" images and overly specific prompts have a higher chance of getting memorized [24]. Thus, there is a strong need for mitigation techniques beyond dataset manipulations. + +# 2.2. Detecting Memorization + +Similarity scores based on the SSCD (self-supervised copy detection) model [18] can be used for getting an estimate on the extent of memorization wherein scores $>0.5$ can suggest that it is memorized [23]. However, this metric is prone to false positives when the textual similarity is high between a generated and real image [24]. Wen et al. [28] showed that for memorized prompts, the difference between the text-conditioned and the unconditioned score predictions is uncharacteristically high and can be informative in detecting memorization. They demonstrated high performance even in the first step of noise prediction. Ren et al. [20], on the other hand, showed that for memorized prompts, the cross-attention scores pertaining to certain "trigger tokens" in the memorized prompts are unusually high. Suggesting that they are given more importance, and thus lead to memorization. Other authors [4, 27] used a black box membership inference attack for finding memorized images. Daras et al. [8] showed that when an image is memorized, the model is able to fully reconstruct it from a noised version of it. This observation could be used for detection. Another line of research has tried to understand memorization by comparing diffusion models to associative memory networks [1, 11, 17]. + +# 2.3. Mitigating Memorization + +# 2.3.1. Training Time Mitigation + +Wen et al. [28] proposed monitoring the text-conditioned noise prediction scores for each sample and excluding it from the current mini-batch if it surpasses a certain predetermined threshold. On similar lines, Ren et al. [20] proposed removing samples from the mini-batch when their cross-attention entropy is above a particular predetermined threshold. Somepalli et al. [24] proposed generating multiple BLIP captions per image during training. They also proposed perturbing the prompts by adding Gaussian noise or adding/replacing random words/numbers. Daras et al. [8] proposed training/finetuning on corrupted images can reduce instances of data replication by the model. Chavhan et al. [6] proposed pruning the UNet model weights assuming prior knowledge of memorized prompts. + +Most training time mitigations require re-training the base diffusion model from scratch. Not only is this computationally expensive, it also requires control over the entire training process including access to the training datasets. + +# 2.3.2. Inference Time Mitigation + +Wen et al. [28] proposed perturbing the prompt embedding by minimizing the difference between the text-conditioned and unconditioned noise prediction scores such that it falls below a certain target loss. Ren et al. [20] proposed applying a mask and rescaling the attention scores to give less attention to "trigger tokens". However, this approach does not mitigate memorization when "trigger tokens" are not present. Somepalli et al. [24] proposed adding random tokens to the text prompt or repeating certain tokens during inference so as to add noise to the text embedding indirectly. Chen et al. [7] proposed different CFG weight schedulers to mitigate different causes of memorization. The approach presumes knowledge of not only whether memorization is present but also the type of memorization. + +Most of these approaches work by updating text prompt/embeddings leads to a trade-off between the text alignment and memorization. Another drawback of these approaches is that they work better when a trigger token is present such that we can reduce the importance/weightage of that token and they do not generalize well to other types of memorization as we show in this paper. Lastly, since they work directly on the text prompt/embeddings, they cannot be applied to other conditioning mechanisms such as class labels. + +# 3. Preliminaries + +Diffusion models [25] such as Stable Diffusion (SD) [21] and Imagen [22] are trained with the objective of learning a model $\epsilon_{\theta}$ to denoise a noisy input vector at different levels of noise characterized by a time-dependent noise scheduler. We can define the diffusion process as a stochastic differential equation (SDE) in the form, + +$$ +d x _ {t} = f (x _ {t}, t) d t + g (t) d w _ {t}. +$$ + +where $x_{t} \in X$ is a sample at time $t$ , $f(x_{t}, t)$ is the drift term, $g(t)$ is the diffusion coefficient, and $w_{t}$ is a standard Wiener process. + +During training, the forward diffusion process comprises a Markov chain with fixed time steps $T$ . Given a data point $x_0 \sim q(x)$ , we iteratively add Gaussian noise with variance $\beta_t$ at each time step $t$ to $x_{t-1}$ to get $x_t$ such that $x_T \sim \mathcal{N}(0, I)$ . This process can be expressed as, + +$$ +q (x _ {t} | x _ {t - 1}) = \mathcal {N} (x _ {t}; \sqrt {1 - \beta_ {t}} x _ {t - 1}, \beta_ {t} I), \quad \forall t \in \{1, \dots , T \}. +$$ + +We can get a closed-form expression of $x_{t}$ , + +$$ +x _ {t} = \sqrt {\bar {\alpha} _ {t}} x _ {0} + \sqrt {1 - \bar {\alpha} _ {t}} \epsilon_ {t}, \tag {1} +$$ + +where $\bar{\alpha}_t = \prod_{i=1}^t (1 - \beta_i)$ and $\alpha_t = 1 - \beta_t$ . + +During training, we learn the reverse process through a network $\epsilon_{\theta}$ to iteratively denoise $x_{t}$ by estimating the noise $\epsilon_{t}$ at each time step $t$ . The loss function is expressed as, + +$$ +\mathcal {L} = \mathbb {E} _ {t \in [ 1, T ], \epsilon \sim \mathcal {N} (0, I)} \left[ \left\| \epsilon_ {t} - \epsilon_ {\theta} \left(x _ {t}\right) \right\| _ {2} ^ {2} \right]. \tag {2} +$$ + +Using the learned noise estimator network $\epsilon_{\theta}$ , we can compute the previous sample $x_{t - 1}$ from $x_{t}$ as follows, + +$$ +x _ {t - 1} = \frac {1}{\sqrt {\alpha} _ {t}} \left(x _ {t} - \frac {1 - \alpha_ {t}}{\sqrt {1 - \bar {\alpha} _ {t}}} \epsilon_ {\theta} \left(x _ {t}\right)\right). \tag {3} +$$ + +The learned noise estimator network $\epsilon_{\theta}$ can be conditioned using a conditioning input such as text or class label, expressed as an embedding $e_p$ . Ho et al. [10] proposed classifier-free guidance (CFG) as a mechanism to guide the diffusion trajectory towards generating outputs that align with the conditioning. The trajectory is directed towards the conditional score predictions and away from the unconditional score predictions, where $s$ controls the degree of adjustment and $e_{\emptyset}$ are empty prompt embeddings used for unconditional guidance. + +$$ +\hat {\epsilon} \leftarrow \epsilon_ {\theta} \left(x _ {t}, e _ {\emptyset}\right) + s \underbrace {\left(\epsilon_ {\theta} \left(x _ {t} , e _ {p}\right) - \epsilon_ {\theta} \left(x _ {t} , e _ {\emptyset}\right)\right)} _ {\text {c o n d i t i o n a l g u i d a n c e}}. \tag {4} +$$ + +# 4. Understanding Diffusion Trajectory during Memorization + +In text-to-image diffusion models, different initializations generally lead to different outputs for the same text prompt. However, when the diffusion model has memorized a particular sample, the outputs are similar and closely resemble one or more training data samples regardless of initialized noise [28]. + +As shown in Figure 2, we observe that applying CFG after a certain time step avoids producing memorized outputs, while applying it from earlier time steps, including at the beginning $(t = T)$ , easily leads to memorized outputs. Inspired by these empirical findings and the perspective of dynamical systems [2, 15], we suspect that there is an attraction basin in the sample space of the diffusion trajectory, in which applying CFG results in memorized outputs. Intuitively, one can imagine this attraction basin as having a funnel-like shape. It is broader at $t = T$ and narrows as the reverse denoising process progresses. To define the attraction basin, we introduce the concept of a denoiser. + +Definition 1 (Denoiser). Let $X$ denote the entire sample space of the diffusion trajectory, and let $E$ denote the embedding space. A denoiser is a function $\varphi : X \times (0, T] \times E \to X$ that outputs the result of the reverse diffusion process based on Eq. (3) with CFG. Specifically, $\varphi(x, t, e)$ represents the resulting sample at time $t = 0$ , starting from a state $(x, t)$ with embeddings $e \in E$ , where $x \in X$ and $t \in (0, T]$ . + +For an attractor $x^{\mathrm{a}}$ from the training dataset, we then define the attraction basin as follows: + +Definition 2 (Attractor and attraction basin). Suppose a prompt is given with embeddings denoted as $e_p$ . A sample + +$x^{\mathrm{a}} \in X$ is considered an $(\epsilon)$ -attractor if the diffusion trajectories denoised with CFG tend to converge to samples within $B_{\mathcal{D}}(x^{\mathrm{a}}, \epsilon)$ at $t = 0$ , regardless of the initial noise, i.e., $\varphi(x, T, e_p) \in B_{\mathcal{D}}(x^{\mathrm{a}}, \epsilon)$ for any $x \in X$ , where $B_{\mathcal{D}}(\cdot, \cdot)$ represents a perceptual ball given by a certain perceptual distance $\mathcal{D}: X \times X \to \mathbb{R}_{\geq 0}$ as + +$$ +B _ {\mathcal {D}} \left(x ^ {\mathrm {a}}, \epsilon\right) = \left\{x ^ {\prime} \in X \mid \mathcal {D} \left(x ^ {\mathrm {a}}, x ^ {\prime}\right) \leq \epsilon \right\}. +$$ + +An attraction basin is a region in the state space within which samples will converge to around $x^{\mathrm{a}}$ at $t = 0$ under CFG inference. More formally, the $((\epsilon, \delta))$ -attraction basin defined using $x^{\mathrm{a}}$ is expressed as + +$$ +X ^ {\mathrm {b}} \left(x ^ {\mathrm {a}}, \epsilon\right) = \left\{\left(x, t\right) \mid \mathbb {P} \left(\varphi (x, t, e) \in B _ {\mathcal {D}} \left(x ^ {\mathrm {a}}, \epsilon\right)\right) > 1 - \delta \right\}, +$$ + +representing all points in the state space from which generated samples using CFG will reach samples perceptually similar to the attractor at $t = 0$ . + +Inspired from Wen et al. [28], we analyze the attraction basin by examining the magnitudes of the conditional guidance, $\epsilon_{\theta}(x_t,e_p) - \epsilon_{\theta}(x_t,e_{\emptyset})$ . Interestingly, as shown in Figure 2, the conditional noise prediction $\epsilon_{\theta}(x_t,e_p) - \epsilon_{\theta}(x_t,e_{\emptyset})$ remains high during the initial time steps and then steeply falls at the time step immediately before the denoising trajectory exits the attraction basin. We refer to this time step as the transition point, marking the change from producing memorized outputs to non-memorized outputs. In Figure 2(a), if you were to start applying CFG at any point on or before $t = 779$ , you get the same memorized output. However, surprisingly, applying CFG at exactly one denoising step after this $t = 759$ leads to a non-memorized output (step size is 20 when denoising with 50 inference steps). We define the transition point as follows. + +Definition 3 (Transition point). Given prompt embeddings, let $x^{\mathrm{a}} \in X$ be an attractor. We define a transition point as $(x_{\tau}, \tau)$ for $x_{\tau} \in X$ and $\tau > 0$ such that $(x_{\tau}, \tau + 1)$ lies within its attraction basin, while $(x_{\tau}, \tau)$ does not. More formally, an $((\epsilon, \delta))$ -transition point satisfies + +$$ +\mathbb {P} \left(\varphi \left(x _ {\tau}, \tau + 1\right) \in B _ {\mathcal {D}} \left(x ^ {\mathrm {a}}, \epsilon\right)\right) > 1 - \delta \quad a n d +$$ + +$$ +\mathbb {P} (\varphi (x _ {\tau}, \tau) \notin B _ {\mathcal {D}} (x ^ {\mathrm {a}}, \epsilon)) < 1 - \delta , +$$ + +where $B_{\mathcal{D}}$ is a perceptual ball defined in Definition 2. + +This phenomenon occurs only for memorized samples as illustrated in Figure 6, the conditional guidance $\epsilon_{\theta}(x_t,e_p) - \epsilon_{\theta}(x_t,e_{\emptyset})$ remains low and almost flat across different time steps for non-memorized samples. We further analyze the diffusion trajectory during memorization, demonstrating the presence of an attraction basin in Appendix 8. + +# 5. Simple Mitigation Strategy + +Based on the intuition from the previous section, we can simply mitigate memorization by finding an ideal transition point to switch from zero CFG to applying CFG. Based on our preliminary experiments, we found that transition points can either be static or dynamic. + +![](images/be1c77a4df291f24f0d967c3297837186a10fc276eeab0f141b6a03ec6de2d81.jpg) +Figure 3. Average magnitude of the text-conditioned $(\epsilon_{\theta}(x_t, e_p))$ and unconditional noise predictions $(\epsilon_{\theta}(x_t, e_{\emptyset}))$ and their difference $(\epsilon_{\theta}(x_t, e_p) - \epsilon_{\theta}(x_t, e_{\emptyset}))$ when applying zero CFG. We see a static transition point ( $t = 500$ ) appear when SDv2.1 is finetuned on the LAION-10k dataset [24]. + +![](images/e87873fd6aeb16f7c3cf38ff8bc82c95c27358ed8b83276fe6b97703b148c80f.jpg) +Figure 4. Applying CFG before the static transition point $(\mathrm{T} = 500)$ leads to memorized outputs while applying CFG after the fixed transition point leads to non-memorized outputs. Applying CFG too late results in poor-quality images that resemble the unconditional generations. + +Static Transition Point. In some trained diffusion models, we observed a universal trend in the mean conditional $\epsilon_{\theta}(x_t,e_p)$ and unconditional noise prediction $\epsilon_{\theta}(x_t,e_{\emptyset})$ across different samples. As shown in Figure 3, the mean conditional noise prediction $\epsilon_{\theta}(x_t,e_p)$ is higher than unconditional noise prediction $\epsilon_{\theta}(x_t,e_{\emptyset})$ in the initial time steps $(t\geq 500)$ . At a static time step $(t = 500)$ , the two scores align in magnitude until the end of the denoising process, and at the same time, the $L_{2}$ distance between them drastically drops. This suggests that the zero CFG trajectory leaves the attraction basin at that fixed time step $(t = 500)$ . + +![](images/0f61266e6523b9078ce2df0b2203aa2ad178e6a4798481e1b2869589d865117e.jpg) +Figure 5. In some models, transition points can occur at a different time step for each sample, as seen for pre-trained SDv1.4. In row 1 the transition point is approximately $t = 800$ while for row 2 it is $t = 650$ . + +This can be further validated by applying CFG from any time step before the fixed transition point $(t = 500)$ , as illustrated in Figure 4. Applying CFG from an early time step leads to the same memorized image, but the output immediately switches at the transition point $(t = 500)$ such that we generate text-aligned non-memorized images. On the other hand, if we start applying CFG very late in the denoising process, the outputs are not well aligned with the text prompts and resemble their unconditional generations. + +Dynamic Transition Point. In other trained diffusion models, we saw that every prompt and initialization pair leads to a different transition point. As shown in Figures 6(a) and 5, different trajectories leave the attraction basin at different time steps, characterized by their fall in the conditional guidance $\epsilon_{\theta}(x_t,e_p) - \epsilon_{\theta}(x_t,e_{\emptyset})$ . We found that it is dependent on both the conditional embedding $e_p$ and the initialization $x_{T}$ . We show an example in Figures 2(a) and 2(b), where the transition point occurs at different time steps even for the same prompt. Thus, we propose a method to find this dynamic transition point (DTP). As shown in Figure 2, it is the point after the first local minima in the graph. Thus, to mitigate memorization, this difference is tracked until the first local minimum occurs, and CFG is applied henceforth. This process requires no additional computations as the text-conditioned and unconditional noise predictions are computed by default when denoising using CFG. We present the pseudo-code for this approach in Algorithm 1. + +# 5.1. Opposite Guidance + +An issue with this method is that when the transition point occurs very late ( $t \leq 500$ ) in the denoising process, CFG is applied for fewer time steps, and this can impact the image quality. + +To ensure that the transition point occurs earlier, we in + +![](images/3f50b349a1bd99b69a39980048240d1d08f250ab20741fe040d0787ff82c2cab.jpg) +(a) Applying zero CFG + +![](images/1f3f1140f2a1236b8917829a23e348de47fd494d97987dca69da77361cea0132.jpg) +(b) Applying opposite guidance +Figure 6. Magnitude of $\epsilon_{\theta}(x_t, e_p) - \epsilon_{\theta}(x_t, e_{\emptyset})$ when denoising (a) without CFG and (b) with opposite guidance. Transition point occurs sooner with opposite guidance. + +Algorithm 1 Reverse Diffusion with Dynamic Transition Point Method and Opposite Guidance +Require: $x_{T}\sim \mathcal{N}(0,\mathrm{I}_{d}),\lambda >0,v_{\mathrm{OG}}\in \{0,1\}$ +1: $d_{T + 2} = -\infty$ . $d_{T + 1} = -\infty$ +2: $s = -\lambda *v_{\mathrm{OG}}$ +3: for $t = T$ to 1 do +4: $d_{t} = \| \epsilon_{\theta}(x_{t},e_{p}) - \epsilon_{\theta}(x_{t},e_{\emptyset})\|_{2}^{2}$ +5: if $d_{t + 2} > d_{t + 1}$ and $d_{t + 1} < d_t$ then +6: $s = \lambda$ +7: end if +8: $\hat{\epsilon} = \epsilon_{\theta}(x_t,e_\emptyset) + s(\epsilon_\theta (x_t,e_p) - \epsilon_\theta (x_t,e_\emptyset))$ +9: $x_{t - 1} = \frac{1}{\sqrt{\alpha}_t} (x_t - \frac{1 - \alpha_t}{\sqrt{1 - \alpha_t}}\hat{\epsilon})$ +10: end for +11: return $x_0$ + +troduce a new concept of opposite guidance (OG). Unlike negative prompting, we apply negative or opposite CFG, to push the diffusion trajectory in the opposite direction to that of traditional CFG and thus, push the trajectory away from the attraction basin of the traditional CFG trajectory sooner in the denoising process, as also shown in Figure 6(b). Opposite guidance can be expressed as follows for $s > 0$ , + +$$ +\hat {\epsilon} \leftarrow \epsilon_ {\theta} \left(x _ {t}, e _ {\emptyset}\right) - s \underbrace {\left(\epsilon_ {\theta} \left(x _ {t} , e _ {p}\right) - \epsilon_ {\theta} \left(x _ {t} , e _ {\emptyset}\right)\right)} _ {\text {c o n d i t i o n a l g u i d a n c e}}. \tag {5} +$$ + +Similar to unconditional denoising, when denoising using opposite guidance, there is an ideal time step where + +the magnitude of conditional guidance falls. At this time step, switching from opposite guidance to traditional positive guidance leads to non-memorized samples with high image quality and text alignment. Thus, it can easily be integrated along with the previous method to find a static or dynamic transition point. We present the pseudo-code for this approach in conjunction with the DTP method in Algorithm 1 with the parameter $v_{\mathrm{OG}} = 1$ . + +# 6. Memorization under Different Scenarios + +We study memorization under different scenarios and show how this simple approach of finding an ideal transition point to avoid the attraction basin can be applied to effectively mitigate memorization. + +- Scenario 1: Fine-tuned SDv2.1 on 10,000 LAION datapoints (followed by [24]). Memorization occurs due to overfitting on the small dataset. +- Scenario 2: Fine-tuned SDv2.1 on the Imagenette dataset [12] containing 10 ImageNet classes (followed by [24]). Exact memorization is not always observed but similarity with the training dataset increases. +- Scenario 3: Fine-tuned SDv1.4 on 200 prompts duplicated 200 times with an additional 120,000 prompts that were not duplicated (followed by [28]). Memorization occurs due to data duplication. +- Scenario 4: 500 memorized prompts for pre-trained SDv1.4 found using membership inference attack [4, 27] (followed by [20]). The exact cause of memorization for each prompt is unknown, but the strong presence of "trigger tokens" is noticeable. These are in the form of nouns such as celebrity and movie names with little to no additional descriptions. Results are in Appendix 9. + +Notably, different researchers have independently studied these different scenarios, but, as we show later, their approaches do not generalize to other scenarios as memorization occurs due to different causes in each scenario. Refer to Appendix 11 for other implementation details. + +Evaluation Metrics We utilize the similarity metric based on the SSCD [18] embeddings to judge the level of memorization [24]; the CLIP score [19] to assess how well the generated images align with the text prompts; and the Fréchet Inception Distance (FID) [9] for image quality. + +# 6.1. Scenario 1 + +We observed that in this scenario, a static transition point at $t = 500$ occurs as shown in Figure 3. We present qualitative and quantitative results in Figure 7 and Table 1, respectively. Our approach outperformed all previous methods, and we observed the previously proposed techniques that focused solely on the presence of "trigger tokens" [20, 28] did not work well. This shows that when memorization occurs due to factors other than data duplication or the presence of + +![](images/b1c0a9055535ea261072cbbe07be2c967b5f1a18aeea5917f71058e67642405d.jpg) +Figure 7. Qualitative results comparing the proposed approach with the baselines in Scenario 1. Ren et al. [20] is unable to mitigate memorization in most cases while Wen et al. [28] has poor image quality and text-alignment. More enlarged examples (Fig. 14) and prompts for this figure are in Appendix 12. + +Table 1. Comparison of results on Scenario 1. None of the other approaches perform well. The only comparable method [28] qualitatively results in poorer quality images (see visual examples in Fig. 7). + +
Similarity (95pc)CLIP ScoreFID
No Mitigation0.65040.302716.8373
Add rand word [24]0.52540.294117.4142
Add rand numb [24]0.54160.299317.0245
Wen et al. [28]0.38530.289516.7176
Ren et al. [20]0.60280.295920.2931
Ours (STP)0.28570.297619.8494
Ours (OG + STP)0.38110.302015.6679
+ +trigger tokens, these methods may fail. We present further results when finetuning on the LAION-100k dataset [24] in Appendix 10. + +# 6.2. Scenario 2 + +Even in this scenario, as illustrated in Figure 10, a static transition point occurs. We applied CFG from $t = 700$ and $t = 600$ to show that we can reduce the similarity with the training dataset at no cost to the image quality, as shown in Table 2. We would like to point out that previous approaches are specific to text-to-image models and thus cannot be extrapolated to other conditioning mechanisms. This shows a further advantage of our CFG-based mitigation + +technique. + +Table 2. Results on Scenario 2. The value in the brackets in the time steps between which we apply CFG. Applying CFG later from $t = 700$ (row 2) and $t = 600$ (row 3) reduces similarity with comparable or improved FID score. + +
Similarity (95pc)FID
No mitigation0.370243.35
CFG=7.5 [700,0]0.301838.86
CFG=7.5 [600,0]0.275645.73
+ +# 6.3. Scenario 3 + +In this scenario, we observed that a dynamic transition point occurs. We experimentally validate our approach of DTP method and opposite guidance qualitatively and quantitatively in Figure 8 and Table 3. We saw that our method performs at par with [28] that studied this specific scenario and even outperforms it in terms of memorization in the top 95 percentile. Approaches by Ren et al. [20] and Somepalli et al.[24] that were proposed for other scenarios do not work well, showcasing their lack of generalizability. + +Inference Time. One of the major advantages of our approach is its simplicity and lack of computational time overhead. Image generation using our approach as well as stan- + +![](images/56f91106059d11df40a99e806c65b6ebbf7aba033ab488b7dc661fdfb66c135f.jpg) +Figure 8. Qualitative results comparing the proposed approach with the baselines in Scenario 3. Prompts used to generate these figures are given in the Appendix. Rows 1, 2, 5, and 6 show examples, where for most baselines the output images remain closely related to their memorized versions. More enlarged examples (Fig. 15) and prompts for this figure are in Appendix 13. + +Table 3. Table showcasing results in Scenario 3. Wen et al. [28] studied this scenario and is the only other method which performs well but does not generalize to other scenarios (see Table 1). + +
SimilarityCLIP ScoreFID
95pcMean ± Std
No Mitigation0.79770.5513 ± 0.160.3105106.49
Add rand word [24]0.74800.4312 ± 0.210.3071116.92
Add rand numb [24]0.73660.3850 ± 0.210.3027126.54
Wen et al. [28] (l_target=3)0.77470.5147 ± 0.170.3100109.28
Wen et al. [28] (l_target=1)0.60380.2808 ± 0.160.3050136.34
Ren et al. [20]0.68810.4036 ± 0.160.3066124.38
Ours (DTP)0.58850.2866 ± 0.160.3020138.92
Ours (OG + DTP)0.69150.2844 ± 0.200.2910140.05
+ +dard CFG-based image generation using SDv2.1 on one A100 GPU takes 1.26 seconds, while in comparison the mitigation technique proposed by Wen et al. [28] took 2.86 seconds and that of Ren et al. [20] took 2.08 seconds. These are the average values across 10,000 generations. + +Limitations. As transition points do not exist for nonmemorized samples, we require detecting memorization before we can apply our approach. This can either be done by denoising twice, once with traditional CFG to look at the trends in the magnitude of conditional guidance and the second time to apply our mitigation technique if the prompt is memorized. Alternatively, as [20, 28] showed, we can + +do detection at $t = 0$ by simply looking at the magnitude of conditional guidance $\| \epsilon_{\theta}(x_t, e_p) - \epsilon_{\theta}(x_t, e_{\emptyset}) \|_2^2$ . If this magnitude is above a certain threshold, it indicates that the sample is memorized with high accuracy and we can apply our guidance accordingly. + +# 7. Conclusion + +We present a novel understanding of the dynamics behind memorization in conditional diffusion models. We showcase the presence of an attraction basin that steers randomly initialized latent vectors towards a memorized output. We propose an approach to steer away from the attraction basin and detect the point at which the trajectory leaves the attraction basin, referred to as the transition point. We show that applying CFG after the transition point leads to non-memorized outputs. We presented results in various scenarios such as the presence of data replication, fine-tuning on smaller datasets, and memorized examples found in existing pre-trained models to showcase its efficiency. + +To summarize, our method is simple, works well across all tested scenarios in terms of both image quality metrics and visual inspection, requires less computational time, and does not require altering the prompt. + +# References + +[1] Luca Ambrogioni. In search of dispersed memories: Generative diffusion models are associative memory networks. Entropy, 26(5):381, 2024. 3 +[2] Joseph Auslander, Nam P Bhatia, and Peter Seibert. Attractors in dynamical systems. Technical report, 1964. 4 +[3] Yuanhao Ban, Ruochen Wang, Tianyi Zhou, Minhao Cheng, Boqing Gong, and Cho-Jui Hsieh. Understanding the impact of negative prompts: When and how do they take effect? arXiv preprint arXiv:2406.02965, 2024. 2 +[4] Nicolas Carlini, Jamie Hayes, Milad Nasr, Matthew Jagielski, Vikash Sehwag, Florian Tramer, Borja Balle, Daphne Ippolito, and Eric Wallace. Extracting training data from diffusion models. In 32nd USENIX Security Symposium (USENIX Security 23), pages 5253-5270, 2023. 1, 2, 3, 6 +[5] Huiwen Chang, Han Zhang, Jarred Barber, AJ Maschinot, Jose Lezama, Lu Jiang, Ming-Hsuan Yang, Kevin Murphy, William T Freeman, Michael Rubinstein, et al. Muse: Text-to-image generation via masked generative transformers. arXiv preprint arXiv:2301.00704, 2023. 1 +[6] Ruchika Chavhan, Ondrej Bohdal, Yongshuo Zong, Da Li, and Timothy Hospedales. Memorized images in diffusion models share a subspace that can be located and deleted. arXiv preprint arXiv:2406.18566, 2024. 3 +[7] Chen Chen, Daochang Liu, and Chang Xu. Towards memorization-free diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8425-8434, 2024. 3 +[8] Giannis Daras, Alexandros G Dimakis, and Constantinos Daskalakis. Consistent diffusion meets tweedie: Training exact ambient diffusion models with noisy data. arXiv preprint arXiv:2404.10177, 2024. 3 +[9] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 6 +[10] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 4 +[11] Benjamin Hoover, Hendrik Strobelt, Dmitry Krotov, Judy Hoffman, Zsolt Kira, and Duen Horng Chau. Memory in plain sight: Surveying the uncanny resemblances of associative memories and diffusion models. arXiv preprint arXiv:2309.16750, 2023. 3 +[12] Jeremy Howard. Imagenette: A smaller subset of 10 easily classified classes from imagenet, 2019. 6, 3 +[13] Tuomas Kynkänniemi, Miika Aittala, Tero Karras, Samuli Laine, Timo Aila, and Jaakko Lehtinen. Applying guidance in a limited interval improves sample and distribution quality in diffusion models. arXiv preprint arXiv:2404.07724, 2024. 1 +[14] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps. Advances in Neural Information Processing Systems, 35:5775-5787, 2022. 3 + +[15] John W. Milnor. Attractor. *Scholarpedia*, 1(11):1815, 2006. 4 +[16] Alex Nichol. Dall-e 2 pre-training mitigations. https://openai.com/index/dall-e-2-pre-training-mitigations/, 2022. [Accessed 10-10-2024]. 2 +[17] Bao Pham, Gabriel Raya, Matteo Negri, Mohammed J Zaki, Luca Ambrogioni, and Dmitry Krotov. Memorization to generalization: The emergence of diffusion models from associative memory. 2024. 3 +[18] Ed Pizzi, Sreya Dutta Roy, Sugosh Nagavara Ravindra, Priya Goyal, and Matthijs Douze. A self-supervised descriptor for image copy detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14532-14542, 2022. 3, 6 +[19] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 6 +[20] Jie Ren, Yaxin Li, Shenglai Zen, Han Xu, Lingjuan Lyu, Yue Xing, and Jiliang Tang. Unveiling and mitigating memorization in text-to-image diffusion models through cross attention. arXiv preprint arXiv:2403.11052, 2024. 1, 2, 3, 6, 7, 8, 5 +[21] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 3 +[22] Chitwan Saharia, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural information processing systems, 35:36479-36494, 2022. 3 +[23] Gowthami Somepalli, Vasu Singla, Micah Goldblum, Jonas Geiping, and Tom Goldstein. Diffusion art or digital forgery? investigating data replication in diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6048-6058, 2023. 1, 2, 3 +[24] Gowthami Somepalli, Vasu Singla, Micah Goldblum, Jonas Geiping, and Tom Goldstein. Understanding and mitigating copying in diffusion models. Advances in Neural Information Processing Systems, 36:47783-47803, 2023. 1, 2, 3, 5, 6, 7, 8 +[25] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 1, 3 +[26] Xi Wang, Nicolas Dufour, Nefeli Andreou, Marie-Paule Cani, Victoria Fernández Abrevaya, David Picard, and Vicky Kalogeiton. Analysis of classifier-free guidance weight schedulers. arXiv preprint arXiv:2404.13040, 2024. 1 +[27] Ryan Webster. A reproducible extraction of training images from diffusion models. arXiv preprint arXiv:2305.08694, 2023. 2, 3, 6 +[28] Yuxin Wen, Yuchen Liu, Chen Chen, and Lingjuan Lyu. Detecting, explaining, and mitigating memorization in diffusion models. In The Twelfth International Conference on Learning Representations, 2024. 1, 2, 3, 4, 6, 7, 8, 5 \ No newline at end of file diff --git a/2025/Classifier-Free Guidance Inside the Attraction Basin May Cause Memorization/images.zip b/2025/Classifier-Free Guidance Inside the Attraction Basin May Cause Memorization/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..c8f014167103a3356984bff5ee9e81dbf5f299d8 --- /dev/null +++ b/2025/Classifier-Free Guidance Inside the Attraction Basin May Cause Memorization/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:35eb833330afa13427a7c2e89ae4692653af44300b2f63203137bb666aaac576 +size 859130 diff --git a/2025/Classifier-Free Guidance Inside the Attraction Basin May Cause Memorization/layout.json b/2025/Classifier-Free Guidance Inside the Attraction Basin May Cause Memorization/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..88c3f8115c9fe4f2869dd1d21a010b876686a214 --- /dev/null +++ b/2025/Classifier-Free Guidance Inside the Attraction Basin May Cause Memorization/layout.json @@ -0,0 +1,8231 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 62, + 103, + 547, + 120 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 103, + 547, + 120 + ], + "spans": [ + { + "bbox": [ + 62, + 103, + 547, + 120 + ], + "type": "text", + "content": "Classifier-Free Guidance inside the Attraction Basin May Cause Memorization" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 140, + 142, + 470, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 142, + 470, + 171 + ], + "spans": [ + { + "bbox": [ + 140, + 142, + 470, + 171 + ], + "type": "text", + "content": "Anubhav Jain1*, Yuya Kobayashi2, Takashi Shibuya2, Yuhta Takida2, Nasir Memon1, Julian Togelius1, Yuki Mitsufuji2,3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 161, + 171, + 449, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 161, + 171, + 449, + 186 + ], + "spans": [ + { + "bbox": [ + 161, + 171, + 449, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 161, + 171, + 449, + 186 + ], + "type": "text", + "content": "New York University, " + }, + { + "bbox": [ + 161, + 171, + 449, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 161, + 171, + 449, + 186 + ], + "type": "text", + "content": "Sony AI, " + }, + { + "bbox": [ + 161, + 171, + 449, + 186 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 161, + 171, + 449, + 186 + ], + "type": "text", + "content": "Sony Group Corporation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 190, + 186, + 419, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 190, + 186, + 419, + 199 + ], + "spans": [ + { + "bbox": [ + 190, + 186, + 419, + 199 + ], + "type": "text", + "content": "{aj3281,memon,julian.togelius}@nyu.edu" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 92, + 200, + 514, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 92, + 200, + 514, + 213 + ], + "spans": [ + { + "bbox": [ + 92, + 200, + 514, + 213 + ], + "type": "text", + "content": "{u.kobayashi,takashi.tak.shibuya,yuta.takida,yuhki.mitsufuji}@sony.com" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 151, + 241, + 200, + 253 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 241, + 200, + 253 + ], + "spans": [ + { + "bbox": [ + 151, + 241, + 200, + 253 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 267, + 297, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 267, + 297, + 542 + ], + "spans": [ + { + "bbox": [ + 54, + 267, + 297, + 542 + ], + "type": "text", + "content": "Diffusion models are prone to exactly reproduce images from the training data. This exact reproduction of the training data is concerning as it can lead to copyright infringement and/or leakage of privacy-sensitive information. In this paper, we present a novel perspective on the memorization phenomenon and propose a simple yet effective approach to mitigate it. We argue that memorization occurs because of an attraction basin in the denoising process which steers the diffusion trajectory towards a memorized image. However, this can be mitigated by guiding the diffusion trajectory away from the attraction basin by not applying classifier-free guidance until an ideal transition point occurs from which classifier-free guidance is applied. This leads to the generation of non-memorized images that are high in image quality and well-aligned with the conditioning mechanism. To further improve on this, we present a new guidance technique, opposite guidance, that escapes the attraction basin sooner in the denoising process. We demonstrate the existence of attraction basins in various scenarios in which memorization occurs, and we show that our proposed approach successfully mitigates memorization. Our codebase is publicly available at https://github.com/SonyResearch/mitigating_memorization." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 566, + 135, + 578 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 566, + 135, + 578 + ], + "spans": [ + { + "bbox": [ + 55, + 566, + 135, + 578 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 586, + 297, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 586, + 297, + 694 + ], + "spans": [ + { + "bbox": [ + 55, + 586, + 297, + 694 + ], + "type": "text", + "content": "Recent advancements in image-generation models have enabled the generation of hyper-realistic images. Diffusion models [21, 25] have been at the forefront of this advancement with their ability to align image generation with different conditioning mechanisms such as text giving users more control over the output. However, a key challenge has emerged: these models can memorize and reproduce training examples verbatim (verbatim memorization) [4, 23, 24], leading to concerns over copyright infringement and leak" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 320, + 244, + 552, + 398 + ], + "blocks": [ + { + "bbox": [ + 320, + 244, + 552, + 398 + ], + "lines": [ + { + "bbox": [ + 320, + 244, + 552, + 398 + ], + "spans": [ + { + "bbox": [ + 320, + 244, + 552, + 398 + ], + "type": "image", + "image_path": "f959491e85289c4f63b35fd8f3ab1bad326e94efd43f273c778e7621afd9b104.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 403, + 555, + 480 + ], + "lines": [ + { + "bbox": [ + 313, + 403, + 555, + 480 + ], + "spans": [ + { + "bbox": [ + 313, + 403, + 555, + 480 + ], + "type": "text", + "content": "Figure 1. The diffusion trajectory contains an attraction basin (red region) which steers conditioned samples towards their memorized images. It can be avoided by applying zero classifier-free guidance when the trajectory is inside the attraction basin, such that there is an ideal transition point " + }, + { + "bbox": [ + 313, + 403, + 555, + 480 + ], + "type": "inline_equation", + "content": "\\tau^{*}" + }, + { + "bbox": [ + 313, + 403, + 555, + 480 + ], + "type": "text", + "content": " after which applying CFG leads to non-memorized output. Applying CFG at an earlier point such as " + }, + { + "bbox": [ + 313, + 403, + 555, + 480 + ], + "type": "inline_equation", + "content": "\\tau_{1}" + }, + { + "bbox": [ + 313, + 403, + 555, + 480 + ], + "type": "text", + "content": " inside the attraction basin leads to the same memorized sample." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 506, + 556, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 506, + 556, + 650 + ], + "spans": [ + { + "bbox": [ + 313, + 506, + 556, + 650 + ], + "type": "text", + "content": "age of privacy sensitive information. Alternatively, the model can generate images copying the composition and structure present in one or more training images (template memorization)[20]. Researchers have identified factors contributing to memorization and reproduction, such as the duplication of training images and repeated captions [24]. However, not all causes are understood, as memorization can still occur even when these issues are avoided. Therefore, there is a growing need for effective mitigation strategies, in particular methods that can be applied during inference without requiring computationally intensive retraining of the core diffusion models." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 654, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 556, + 713 + ], + "type": "text", + "content": "In text-to-image (T2I) diffusion models, strong associations between a text prompt and an image in the training dataset lead to memorization. This can occur due to the presence of \"trigger tokens\" in overly specific text-prompts [20, 28], the model seeing multiple occurrences of these dur" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 702, + 212, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 702, + 212, + 713 + ], + "spans": [ + { + "bbox": [ + 66, + 702, + 212, + 713 + ], + "type": "text", + "content": "*Work done during an internship at Sony AI." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12871" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 71, + 299, + 213 + ], + "blocks": [ + { + "bbox": [ + 59, + 71, + 299, + 213 + ], + "lines": [ + { + "bbox": [ + 59, + 71, + 299, + 213 + ], + "spans": [ + { + "bbox": [ + 59, + 71, + 299, + 213 + ], + "type": "image", + "image_path": "d81ca9c631ca872b93e29fdcc7c5c867306e0a9a4ee22644eea158e8cadf89fa.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 56, + 218, + 302, + 237 + ], + "lines": [ + { + "bbox": [ + 56, + 218, + 302, + 237 + ], + "spans": [ + { + "bbox": [ + 56, + 218, + 302, + 237 + ], + "type": "text", + "content": "(a) Prompt: \"Here's What You Need to Know About St. Vincent's Apple Music Radio Show\"" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 306, + 70, + 553, + 215 + ], + "blocks": [ + { + "bbox": [ + 306, + 70, + 553, + 215 + ], + "lines": [ + { + "bbox": [ + 306, + 70, + 553, + 215 + ], + "spans": [ + { + "bbox": [ + 306, + 70, + 553, + 215 + ], + "type": "image", + "image_path": "3c89e43d47393742467f6e21b22d3314396f0c775e181c7f42959f7e3c64e3c2.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 218, + 553, + 237 + ], + "lines": [ + { + "bbox": [ + 306, + 218, + 553, + 237 + ], + "spans": [ + { + "bbox": [ + 306, + 218, + 553, + 237 + ], + "type": "text", + "content": "(b) Prompt: \"Here's What You Need to Know About St. Vincent's Apple Music Radio Show\"" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 54, + 242, + 555, + 288 + ], + "lines": [ + { + "bbox": [ + 54, + 242, + 555, + 288 + ], + "spans": [ + { + "bbox": [ + 54, + 242, + 555, + 288 + ], + "type": "text", + "content": "Figure 2. Plots showing magnitude of " + }, + { + "bbox": [ + 54, + 242, + 555, + 288 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(x_t, e_p) - \\epsilon_{\\theta}(x_t, e_{\\emptyset})" + }, + { + "bbox": [ + 54, + 242, + 555, + 288 + ], + "type": "text", + "content": " when denoising without classifier-free guidance (CFG) at each time step. The figures show the generated image if you start applying CFG at that time step. We get non-memorized output if we apply CFG after the ideal transition point " + }, + { + "bbox": [ + 54, + 242, + 555, + 288 + ], + "type": "inline_equation", + "content": "\\tau^*" + }, + { + "bbox": [ + 54, + 242, + 555, + 288 + ], + "type": "text", + "content": " which coincides with the fall in the conditional noise prediction. This value is dependent on both the prompt and the initialization ((a) and (b) contain the same prompt with different initializations). More examples in the Appendix." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 308, + 297, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 308, + 297, + 440 + ], + "spans": [ + { + "bbox": [ + 54, + 308, + 297, + 440 + ], + "type": "text", + "content": "ing training (data duplication) [24, 28], or other causes such as fine-tuning on smaller datasets [24]. Previous approaches have focused on perturbing either the text prompt/embeddings [24, 28] or the cross-attention between the text and image embeddings [20]. While these approaches help weaken the associations between prompts and corresponding images, they are limited to specific conditions in which they have been tested and fail to generalize effectively. Additionally, since these techniques work at the prompt level, they are only applicable to text-conditional generation, and also assume the ability to update the prompt." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 451, + 298, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 451, + 298, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 451, + 298, + 714 + ], + "type": "text", + "content": "In this paper, we introduce a new concept to explain the behavior of memorization in diffusion models based on our empirical observations. We first observe that applying classifier-free guidance (CFG) before a certain time step tends to produce memorized samples; we refer to this time step as the transition point. Interestingly, applying CFG after the transition point is unlikely to yield a memorized image. Based on these observations, we propose the concept of an attraction basin, in which applying CFG induces a strong steering force toward a memorized output, even with different initializations. Furthermore, following the approach of Wen et al. [28], we analyze the magnitude of the conditional noise predictions and discover uncharacteristically large values within the attraction basin, providing empirical evidence of a strong steering force in this region. Once the transition point is passed, the strong steering force is no longer pervasive. This magnitude-based analysis motivates us to detect the transition point by tracking the magnitude. Specifically, we initially apply no guidance until a transition point occurs—marked by a sharp drop in the discrepancy between conditional and unconditional guidance—after which applying CFG leads to non-memorized images." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 324, + 308, + 555, + 320 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 308, + 555, + 320 + ], + "spans": [ + { + "bbox": [ + 324, + 308, + 555, + 320 + ], + "type": "text", + "content": "We summarize our contributions in this paper as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 320, + 555, + 523 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 314, + 320, + 555, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 320, + 555, + 380 + ], + "spans": [ + { + "bbox": [ + 314, + 320, + 555, + 380 + ], + "type": "text", + "content": "- We present a new way to understand memorization in conditional diffusion models by understanding the diffusion trajectory from a dynamical systems theory perspective. We show that an attraction basin forms in the denoising trajectory." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 380, + 555, + 427 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 380, + 555, + 427 + ], + "spans": [ + { + "bbox": [ + 314, + 380, + 555, + 427 + ], + "type": "text", + "content": "- We present approaches to avoiding the attraction basin in various scenarios that incur no additional computational costs and require neither access to the prompt nor modification of weights." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 428, + 555, + 463 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 428, + 555, + 463 + ], + "spans": [ + { + "bbox": [ + 314, + 428, + 555, + 463 + ], + "type": "text", + "content": "- We present a new guidance technique that we call Opposite Guidance (not to be confused with negative prompts [3]) to push the trajectory away from the attraction basin." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 463, + 555, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 463, + 555, + 523 + ], + "spans": [ + { + "bbox": [ + 314, + 463, + 555, + 523 + ], + "type": "text", + "content": "- We comprehensively show that previous approaches are not generalizable in mitigating memorization and only work in certain scenarios studied in those papers. On the contrary, our approach is able to mitigate memorization in all the explored scenarios." + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 533, + 402, + 545 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 533, + 402, + 545 + ], + "spans": [ + { + "bbox": [ + 314, + 533, + 402, + 545 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 552, + 477, + 565 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 552, + 477, + 565 + ], + "spans": [ + { + "bbox": [ + 313, + 552, + 477, + 565 + ], + "type": "text", + "content": "2.1. Understanding Memorization" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 312, + 570, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 570, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 312, + 570, + 556, + 714 + ], + "type": "text", + "content": "In deep generative models, when memorization occurs, a generated sample can match verbatim with a single or set of training images (verbatim memorization) [4, 23, 24] or copy the same template as the training image (template memorization) [20] leading to privacy and/or copyright infringement issues. Researchers [4, 23, 24, 27] have identified that when images or prompts are duplicated in the training dataset, it can lead to memorization. Additionally, training on smaller datasets has also been shown to be a major factor in causing memorization [24]. However, as Somepalli et al. [24] pointed out, even though data has been de-duplicated in the newer models [16] such as Stable Diffusion v2.1 (SDv2.1)," + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "12872" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 167 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 167 + ], + "type": "text", + "content": "the issue of memorization is yet to be resolved, and it does not explain much of the observed replication behavior. Factors such as conditioning and dataset complexity play a role even when the training datasets are large enough and are de-duplicated where \"simpler\" images and overly specific prompts have a higher chance of getting memorized [24]. Thus, there is a strong need for mitigation techniques beyond dataset manipulations." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 175, + 192, + 188 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 175, + 192, + 188 + ], + "spans": [ + { + "bbox": [ + 55, + 175, + 192, + 188 + ], + "type": "text", + "content": "2.2. Detecting Memorization" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 194, + 296, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 194, + 296, + 458 + ], + "spans": [ + { + "bbox": [ + 55, + 194, + 296, + 458 + ], + "type": "text", + "content": "Similarity scores based on the SSCD (self-supervised copy detection) model [18] can be used for getting an estimate on the extent of memorization wherein scores " + }, + { + "bbox": [ + 55, + 194, + 296, + 458 + ], + "type": "inline_equation", + "content": ">0.5" + }, + { + "bbox": [ + 55, + 194, + 296, + 458 + ], + "type": "text", + "content": " can suggest that it is memorized [23]. However, this metric is prone to false positives when the textual similarity is high between a generated and real image [24]. Wen et al. [28] showed that for memorized prompts, the difference between the text-conditioned and the unconditioned score predictions is uncharacteristically high and can be informative in detecting memorization. They demonstrated high performance even in the first step of noise prediction. Ren et al. [20], on the other hand, showed that for memorized prompts, the cross-attention scores pertaining to certain \"trigger tokens\" in the memorized prompts are unusually high. Suggesting that they are given more importance, and thus lead to memorization. Other authors [4, 27] used a black box membership inference attack for finding memorized images. Daras et al. [8] showed that when an image is memorized, the model is able to fully reconstruct it from a noised version of it. This observation could be used for detection. Another line of research has tried to understand memorization by comparing diffusion models to associative memory networks [1, 11, 17]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 465, + 197, + 478 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 465, + 197, + 478 + ], + "spans": [ + { + "bbox": [ + 55, + 465, + 197, + 478 + ], + "type": "text", + "content": "2.3. Mitigating Memorization" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 483, + 192, + 495 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 483, + 192, + 495 + ], + "spans": [ + { + "bbox": [ + 55, + 483, + 192, + 495 + ], + "type": "text", + "content": "2.3.1. Training Time Mitigation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 498, + 296, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 296, + 665 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 296, + 665 + ], + "type": "text", + "content": "Wen et al. [28] proposed monitoring the text-conditioned noise prediction scores for each sample and excluding it from the current mini-batch if it surpasses a certain predetermined threshold. On similar lines, Ren et al. [20] proposed removing samples from the mini-batch when their cross-attention entropy is above a particular predetermined threshold. Somepalli et al. [24] proposed generating multiple BLIP captions per image during training. They also proposed perturbing the prompts by adding Gaussian noise or adding/replacing random words/numbers. Daras et al. [8] proposed training/finetuning on corrupted images can reduce instances of data replication by the model. Chavhan et al. [6] proposed pruning the UNet model weights assuming prior knowledge of memorized prompts." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "Most training time mitigations require re-training the base diffusion model from scratch. Not only is this computationally expensive, it also requires control over the entire training process including access to the training datasets." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 72, + 454, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 454, + 84 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 454, + 84 + ], + "type": "text", + "content": "2.3.2. Inference Time Mitigation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 87, + 555, + 254 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 87, + 555, + 254 + ], + "spans": [ + { + "bbox": [ + 313, + 87, + 555, + 254 + ], + "type": "text", + "content": "Wen et al. [28] proposed perturbing the prompt embedding by minimizing the difference between the text-conditioned and unconditioned noise prediction scores such that it falls below a certain target loss. Ren et al. [20] proposed applying a mask and rescaling the attention scores to give less attention to \"trigger tokens\". However, this approach does not mitigate memorization when \"trigger tokens\" are not present. Somepalli et al. [24] proposed adding random tokens to the text prompt or repeating certain tokens during inference so as to add noise to the text embedding indirectly. Chen et al. [7] proposed different CFG weight schedulers to mitigate different causes of memorization. The approach presumes knowledge of not only whether memorization is present but also the type of memorization." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 255, + 556, + 362 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 255, + 556, + 362 + ], + "spans": [ + { + "bbox": [ + 313, + 255, + 556, + 362 + ], + "type": "text", + "content": "Most of these approaches work by updating text prompt/embeddings leads to a trade-off between the text alignment and memorization. Another drawback of these approaches is that they work better when a trigger token is present such that we can reduce the importance/weightage of that token and they do not generalize well to other types of memorization as we show in this paper. Lastly, since they work directly on the text prompt/embeddings, they cannot be applied to other conditioning mechanisms such as class labels." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 371, + 399, + 384 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 371, + 399, + 384 + ], + "spans": [ + { + "bbox": [ + 314, + 371, + 399, + 384 + ], + "type": "text", + "content": "3. Preliminaries" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 392, + 554, + 463 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 392, + 554, + 463 + ], + "spans": [ + { + "bbox": [ + 313, + 392, + 554, + 463 + ], + "type": "text", + "content": "Diffusion models [25] such as Stable Diffusion (SD) [21] and Imagen [22] are trained with the objective of learning a model " + }, + { + "bbox": [ + 313, + 392, + 554, + 463 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 313, + 392, + 554, + 463 + ], + "type": "text", + "content": " to denoise a noisy input vector at different levels of noise characterized by a time-dependent noise scheduler. We can define the diffusion process as a stochastic differential equation (SDE) in the form," + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 375, + 469, + 492, + 483 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 375, + 469, + 492, + 483 + ], + "spans": [ + { + "bbox": [ + 375, + 469, + 492, + 483 + ], + "type": "interline_equation", + "content": "d x _ {t} = f (x _ {t}, t) d t + g (t) d w _ {t}.", + "image_path": "bf754d5847132c580603b7e51146fb9817c1e3edec9159bcaa8bf8f581a63b56.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 488, + 555, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 488, + 555, + 523 + ], + "spans": [ + { + "bbox": [ + 313, + 488, + 555, + 523 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 488, + 555, + 523 + ], + "type": "inline_equation", + "content": "x_{t} \\in X" + }, + { + "bbox": [ + 313, + 488, + 555, + 523 + ], + "type": "text", + "content": " is a sample at time " + }, + { + "bbox": [ + 313, + 488, + 555, + 523 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 488, + 555, + 523 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 488, + 555, + 523 + ], + "type": "inline_equation", + "content": "f(x_{t}, t)" + }, + { + "bbox": [ + 313, + 488, + 555, + 523 + ], + "type": "text", + "content": " is the drift term, " + }, + { + "bbox": [ + 313, + 488, + 555, + 523 + ], + "type": "inline_equation", + "content": "g(t)" + }, + { + "bbox": [ + 313, + 488, + 555, + 523 + ], + "type": "text", + "content": " is the diffusion coefficient, and " + }, + { + "bbox": [ + 313, + 488, + 555, + 523 + ], + "type": "inline_equation", + "content": "w_{t}" + }, + { + "bbox": [ + 313, + 488, + 555, + 523 + ], + "type": "text", + "content": " is a standard Wiener process." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 525, + 554, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 525, + 554, + 584 + ], + "spans": [ + { + "bbox": [ + 313, + 525, + 554, + 584 + ], + "type": "text", + "content": "During training, the forward diffusion process comprises a Markov chain with fixed time steps " + }, + { + "bbox": [ + 313, + 525, + 554, + 584 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 525, + 554, + 584 + ], + "type": "text", + "content": ". Given a data point " + }, + { + "bbox": [ + 313, + 525, + 554, + 584 + ], + "type": "inline_equation", + "content": "x_0 \\sim q(x)" + }, + { + "bbox": [ + 313, + 525, + 554, + 584 + ], + "type": "text", + "content": ", we iteratively add Gaussian noise with variance " + }, + { + "bbox": [ + 313, + 525, + 554, + 584 + ], + "type": "inline_equation", + "content": "\\beta_t" + }, + { + "bbox": [ + 313, + 525, + 554, + 584 + ], + "type": "text", + "content": " at each time step " + }, + { + "bbox": [ + 313, + 525, + 554, + 584 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 525, + 554, + 584 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 525, + 554, + 584 + ], + "type": "inline_equation", + "content": "x_{t-1}" + }, + { + "bbox": [ + 313, + 525, + 554, + 584 + ], + "type": "text", + "content": " to get " + }, + { + "bbox": [ + 313, + 525, + 554, + 584 + ], + "type": "inline_equation", + "content": "x_t" + }, + { + "bbox": [ + 313, + 525, + 554, + 584 + ], + "type": "text", + "content": " such that " + }, + { + "bbox": [ + 313, + 525, + 554, + 584 + ], + "type": "inline_equation", + "content": "x_T \\sim \\mathcal{N}(0, I)" + }, + { + "bbox": [ + 313, + 525, + 554, + 584 + ], + "type": "text", + "content": ". This process can be expressed as," + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 314, + 590, + 557, + 605 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 590, + 557, + 605 + ], + "spans": [ + { + "bbox": [ + 314, + 590, + 557, + 605 + ], + "type": "interline_equation", + "content": "q (x _ {t} | x _ {t - 1}) = \\mathcal {N} (x _ {t}; \\sqrt {1 - \\beta_ {t}} x _ {t - 1}, \\beta_ {t} I), \\quad \\forall t \\in \\{1, \\dots , T \\}.", + "image_path": "c645de94eacbec725e3afc0dbc3d92010afd01cea99a910f42a9e6a0dedbcaf8.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 326, + 610, + 501, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 610, + 501, + 622 + ], + "spans": [ + { + "bbox": [ + 326, + 610, + 501, + 622 + ], + "type": "text", + "content": "We can get a closed-form expression of " + }, + { + "bbox": [ + 326, + 610, + 501, + 622 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 326, + 610, + 501, + 622 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 378, + 628, + 555, + 641 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 378, + 628, + 555, + 641 + ], + "spans": [ + { + "bbox": [ + 378, + 628, + 555, + 641 + ], + "type": "interline_equation", + "content": "x _ {t} = \\sqrt {\\bar {\\alpha} _ {t}} x _ {0} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon_ {t}, \\tag {1}", + "image_path": "d6b2bd3f5bb970d71cc6ebef704e8db5f4a64af0699f7945d57a13baa2da9877.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 646, + 495, + 659 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 646, + 495, + 659 + ], + "spans": [ + { + "bbox": [ + 313, + 646, + 495, + 659 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 646, + 495, + 659 + ], + "type": "inline_equation", + "content": "\\bar{\\alpha}_t = \\prod_{i=1}^t (1 - \\beta_i)" + }, + { + "bbox": [ + 313, + 646, + 495, + 659 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 646, + 495, + 659 + ], + "type": "inline_equation", + "content": "\\alpha_t = 1 - \\beta_t" + }, + { + "bbox": [ + 313, + 646, + 495, + 659 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 659, + 554, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 659, + 554, + 694 + ], + "spans": [ + { + "bbox": [ + 313, + 659, + 554, + 694 + ], + "type": "text", + "content": "During training, we learn the reverse process through a network " + }, + { + "bbox": [ + 313, + 659, + 554, + 694 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 313, + 659, + 554, + 694 + ], + "type": "text", + "content": " to iteratively denoise " + }, + { + "bbox": [ + 313, + 659, + 554, + 694 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 313, + 659, + 554, + 694 + ], + "type": "text", + "content": " by estimating the noise " + }, + { + "bbox": [ + 313, + 659, + 554, + 694 + ], + "type": "inline_equation", + "content": "\\epsilon_{t}" + }, + { + "bbox": [ + 313, + 659, + 554, + 694 + ], + "type": "text", + "content": " at each time step " + }, + { + "bbox": [ + 313, + 659, + 554, + 694 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 659, + 554, + 694 + ], + "type": "text", + "content": ". The loss function is expressed as," + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 354, + 700, + 555, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 700, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 354, + 700, + 555, + 715 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathbb {E} _ {t \\in [ 1, T ], \\epsilon \\sim \\mathcal {N} (0, I)} \\left[ \\left\\| \\epsilon_ {t} - \\epsilon_ {\\theta} \\left(x _ {t}\\right) \\right\\| _ {2} ^ {2} \\right]. \\tag {2}", + "image_path": "bab5b1c28c737b733357c0aa40df74285abf4b924f0ee4de7382d94a694eaaaf.jpg" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12873" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "text", + "content": "Using the learned noise estimator network " + }, + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "text", + "content": ", we can compute the previous sample " + }, + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "inline_equation", + "content": "x_{t - 1}" + }, + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "text", + "content": " as follows," + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 99, + 105, + 296, + 131 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 105, + 296, + 131 + ], + "spans": [ + { + "bbox": [ + 99, + 105, + 296, + 131 + ], + "type": "interline_equation", + "content": "x _ {t - 1} = \\frac {1}{\\sqrt {\\alpha} _ {t}} \\left(x _ {t} - \\frac {1 - \\alpha_ {t}}{\\sqrt {1 - \\bar {\\alpha} _ {t}}} \\epsilon_ {\\theta} \\left(x _ {t}\\right)\\right). \\tag {3}", + "image_path": "1105492267e93d4384dc6e5e347db2ac6367a01eaa718a5c9f874383b4c98fdc.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 139, + 296, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 139, + 296, + 258 + ], + "spans": [ + { + "bbox": [ + 55, + 139, + 296, + 258 + ], + "type": "text", + "content": "The learned noise estimator network " + }, + { + "bbox": [ + 55, + 139, + 296, + 258 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 55, + 139, + 296, + 258 + ], + "type": "text", + "content": " can be conditioned using a conditioning input such as text or class label, expressed as an embedding " + }, + { + "bbox": [ + 55, + 139, + 296, + 258 + ], + "type": "inline_equation", + "content": "e_p" + }, + { + "bbox": [ + 55, + 139, + 296, + 258 + ], + "type": "text", + "content": ". Ho et al. [10] proposed classifier-free guidance (CFG) as a mechanism to guide the diffusion trajectory towards generating outputs that align with the conditioning. The trajectory is directed towards the conditional score predictions and away from the unconditional score predictions, where " + }, + { + "bbox": [ + 55, + 139, + 296, + 258 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 55, + 139, + 296, + 258 + ], + "type": "text", + "content": " controls the degree of adjustment and " + }, + { + "bbox": [ + 55, + 139, + 296, + 258 + ], + "type": "inline_equation", + "content": "e_{\\emptyset}" + }, + { + "bbox": [ + 55, + 139, + 296, + 258 + ], + "type": "text", + "content": " are empty prompt embeddings used for unconditional guidance." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 85, + 268, + 296, + 297 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 268, + 296, + 297 + ], + "spans": [ + { + "bbox": [ + 85, + 268, + 296, + 297 + ], + "type": "interline_equation", + "content": "\\hat {\\epsilon} \\leftarrow \\epsilon_ {\\theta} \\left(x _ {t}, e _ {\\emptyset}\\right) + s \\underbrace {\\left(\\epsilon_ {\\theta} \\left(x _ {t} , e _ {p}\\right) - \\epsilon_ {\\theta} \\left(x _ {t} , e _ {\\emptyset}\\right)\\right)} _ {\\text {c o n d i t i o n a l g u i d a n c e}}. \\tag {4}", + "image_path": "51ea41ea5dee039ad922c558927ce2a6950b0f63d7254ecf5f564f1c07071b61.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 306, + 254, + 335 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 306, + 254, + 335 + ], + "spans": [ + { + "bbox": [ + 55, + 306, + 254, + 335 + ], + "type": "text", + "content": "4. Understanding Diffusion Trajectory during Memorization" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 340, + 296, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 340, + 296, + 411 + ], + "spans": [ + { + "bbox": [ + 55, + 340, + 296, + 411 + ], + "type": "text", + "content": "In text-to-image diffusion models, different initializations generally lead to different outputs for the same text prompt. However, when the diffusion model has memorized a particular sample, the outputs are similar and closely resemble one or more training data samples regardless of initialized noise [28]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 413, + 296, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 413, + 296, + 556 + ], + "spans": [ + { + "bbox": [ + 55, + 413, + 296, + 556 + ], + "type": "text", + "content": "As shown in Figure 2, we observe that applying CFG after a certain time step avoids producing memorized outputs, while applying it from earlier time steps, including at the beginning " + }, + { + "bbox": [ + 55, + 413, + 296, + 556 + ], + "type": "inline_equation", + "content": "(t = T)" + }, + { + "bbox": [ + 55, + 413, + 296, + 556 + ], + "type": "text", + "content": ", easily leads to memorized outputs. Inspired by these empirical findings and the perspective of dynamical systems [2, 15], we suspect that there is an attraction basin in the sample space of the diffusion trajectory, in which applying CFG results in memorized outputs. Intuitively, one can imagine this attraction basin as having a funnel-like shape. It is broader at " + }, + { + "bbox": [ + 55, + 413, + 296, + 556 + ], + "type": "inline_equation", + "content": "t = T" + }, + { + "bbox": [ + 55, + 413, + 296, + 556 + ], + "type": "text", + "content": " and narrows as the reverse denoising process progresses. To define the attraction basin, we introduce the concept of a denoiser." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 564, + 296, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 564, + 296, + 649 + ], + "spans": [ + { + "bbox": [ + 55, + 564, + 296, + 649 + ], + "type": "text", + "content": "Definition 1 (Denoiser). Let " + }, + { + "bbox": [ + 55, + 564, + 296, + 649 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 55, + 564, + 296, + 649 + ], + "type": "text", + "content": " denote the entire sample space of the diffusion trajectory, and let " + }, + { + "bbox": [ + 55, + 564, + 296, + 649 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 55, + 564, + 296, + 649 + ], + "type": "text", + "content": " denote the embedding space. A denoiser is a function " + }, + { + "bbox": [ + 55, + 564, + 296, + 649 + ], + "type": "inline_equation", + "content": "\\varphi : X \\times (0, T] \\times E \\to X" + }, + { + "bbox": [ + 55, + 564, + 296, + 649 + ], + "type": "text", + "content": " that outputs the result of the reverse diffusion process based on Eq. (3) with CFG. Specifically, " + }, + { + "bbox": [ + 55, + 564, + 296, + 649 + ], + "type": "inline_equation", + "content": "\\varphi(x, t, e)" + }, + { + "bbox": [ + 55, + 564, + 296, + 649 + ], + "type": "text", + "content": " represents the resulting sample at time " + }, + { + "bbox": [ + 55, + 564, + 296, + 649 + ], + "type": "inline_equation", + "content": "t = 0" + }, + { + "bbox": [ + 55, + 564, + 296, + 649 + ], + "type": "text", + "content": ", starting from a state " + }, + { + "bbox": [ + 55, + 564, + 296, + 649 + ], + "type": "inline_equation", + "content": "(x, t)" + }, + { + "bbox": [ + 55, + 564, + 296, + 649 + ], + "type": "text", + "content": " with embeddings " + }, + { + "bbox": [ + 55, + 564, + 296, + 649 + ], + "type": "inline_equation", + "content": "e \\in E" + }, + { + "bbox": [ + 55, + 564, + 296, + 649 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 564, + 296, + 649 + ], + "type": "inline_equation", + "content": "x \\in X" + }, + { + "bbox": [ + 55, + 564, + 296, + 649 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 564, + 296, + 649 + ], + "type": "inline_equation", + "content": "t \\in (0, T]" + }, + { + "bbox": [ + 55, + 564, + 296, + 649 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 657, + 295, + 681 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 657, + 295, + 681 + ], + "spans": [ + { + "bbox": [ + 55, + 657, + 295, + 681 + ], + "type": "text", + "content": "For an attractor " + }, + { + "bbox": [ + 55, + 657, + 295, + 681 + ], + "type": "inline_equation", + "content": "x^{\\mathrm{a}}" + }, + { + "bbox": [ + 55, + 657, + 295, + 681 + ], + "type": "text", + "content": " from the training dataset, we then define the attraction basin as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": "Definition 2 (Attractor and attraction basin). Suppose a prompt is given with embeddings denoted as " + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "inline_equation", + "content": "e_p" + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": ". A sample" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 555, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 144 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 144 + ], + "type": "inline_equation", + "content": "x^{\\mathrm{a}} \\in X" + }, + { + "bbox": [ + 313, + 72, + 555, + 144 + ], + "type": "text", + "content": " is considered an " + }, + { + "bbox": [ + 313, + 72, + 555, + 144 + ], + "type": "inline_equation", + "content": "(\\epsilon)" + }, + { + "bbox": [ + 313, + 72, + 555, + 144 + ], + "type": "text", + "content": "-attractor if the diffusion trajectories denoised with CFG tend to converge to samples within " + }, + { + "bbox": [ + 313, + 72, + 555, + 144 + ], + "type": "inline_equation", + "content": "B_{\\mathcal{D}}(x^{\\mathrm{a}}, \\epsilon)" + }, + { + "bbox": [ + 313, + 72, + 555, + 144 + ], + "type": "text", + "content": " at " + }, + { + "bbox": [ + 313, + 72, + 555, + 144 + ], + "type": "inline_equation", + "content": "t = 0" + }, + { + "bbox": [ + 313, + 72, + 555, + 144 + ], + "type": "text", + "content": ", regardless of the initial noise, i.e., " + }, + { + "bbox": [ + 313, + 72, + 555, + 144 + ], + "type": "inline_equation", + "content": "\\varphi(x, T, e_p) \\in B_{\\mathcal{D}}(x^{\\mathrm{a}}, \\epsilon)" + }, + { + "bbox": [ + 313, + 72, + 555, + 144 + ], + "type": "text", + "content": " for any " + }, + { + "bbox": [ + 313, + 72, + 555, + 144 + ], + "type": "inline_equation", + "content": "x \\in X" + }, + { + "bbox": [ + 313, + 72, + 555, + 144 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 72, + 555, + 144 + ], + "type": "inline_equation", + "content": "B_{\\mathcal{D}}(\\cdot, \\cdot)" + }, + { + "bbox": [ + 313, + 72, + 555, + 144 + ], + "type": "text", + "content": " represents a perceptual ball given by a certain perceptual distance " + }, + { + "bbox": [ + 313, + 72, + 555, + 144 + ], + "type": "inline_equation", + "content": "\\mathcal{D}: X \\times X \\to \\mathbb{R}_{\\geq 0}" + }, + { + "bbox": [ + 313, + 72, + 555, + 144 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 351, + 156, + 517, + 169 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 156, + 517, + 169 + ], + "spans": [ + { + "bbox": [ + 351, + 156, + 517, + 169 + ], + "type": "interline_equation", + "content": "B _ {\\mathcal {D}} \\left(x ^ {\\mathrm {a}}, \\epsilon\\right) = \\left\\{x ^ {\\prime} \\in X \\mid \\mathcal {D} \\left(x ^ {\\mathrm {a}}, x ^ {\\prime}\\right) \\leq \\epsilon \\right\\}.", + "image_path": "cb882c1592c57435591eeb44a8c4f16abaf7ceeafb9f7b4e4a9f8ef4bf3c64c0.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 180, + 555, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 180, + 555, + 228 + ], + "spans": [ + { + "bbox": [ + 313, + 180, + 555, + 228 + ], + "type": "text", + "content": "An attraction basin is a region in the state space within which samples will converge to around " + }, + { + "bbox": [ + 313, + 180, + 555, + 228 + ], + "type": "inline_equation", + "content": "x^{\\mathrm{a}}" + }, + { + "bbox": [ + 313, + 180, + 555, + 228 + ], + "type": "text", + "content": " at " + }, + { + "bbox": [ + 313, + 180, + 555, + 228 + ], + "type": "inline_equation", + "content": "t = 0" + }, + { + "bbox": [ + 313, + 180, + 555, + 228 + ], + "type": "text", + "content": " under CFG inference. More formally, the " + }, + { + "bbox": [ + 313, + 180, + 555, + 228 + ], + "type": "inline_equation", + "content": "((\\epsilon, \\delta))" + }, + { + "bbox": [ + 313, + 180, + 555, + 228 + ], + "type": "text", + "content": "-attraction basin defined using " + }, + { + "bbox": [ + 313, + 180, + 555, + 228 + ], + "type": "inline_equation", + "content": "x^{\\mathrm{a}}" + }, + { + "bbox": [ + 313, + 180, + 555, + 228 + ], + "type": "text", + "content": " is expressed as" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 239, + 556, + 259 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 239, + 556, + 259 + ], + "spans": [ + { + "bbox": [ + 314, + 239, + 556, + 259 + ], + "type": "interline_equation", + "content": "X ^ {\\mathrm {b}} \\left(x ^ {\\mathrm {a}}, \\epsilon\\right) = \\left\\{\\left(x, t\\right) \\mid \\mathbb {P} \\left(\\varphi (x, t, e) \\in B _ {\\mathcal {D}} \\left(x ^ {\\mathrm {a}}, \\epsilon\\right)\\right) > 1 - \\delta \\right\\},", + "image_path": "ab43c0fff4a080f1d842ff92bfda085cf738da9ac56132f5a6097c2e44fc00ce.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 270, + 555, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 270, + 555, + 304 + ], + "spans": [ + { + "bbox": [ + 313, + 270, + 555, + 304 + ], + "type": "text", + "content": "representing all points in the state space from which generated samples using CFG will reach samples perceptually similar to the attractor at " + }, + { + "bbox": [ + 313, + 270, + 555, + 304 + ], + "type": "inline_equation", + "content": "t = 0" + }, + { + "bbox": [ + 313, + 270, + 555, + 304 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 316, + 556, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 316, + 556, + 496 + ], + "spans": [ + { + "bbox": [ + 313, + 316, + 556, + 496 + ], + "type": "text", + "content": "Inspired from Wen et al. [28], we analyze the attraction basin by examining the magnitudes of the conditional guidance, " + }, + { + "bbox": [ + 313, + 316, + 556, + 496 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(x_t,e_p) - \\epsilon_{\\theta}(x_t,e_{\\emptyset})" + }, + { + "bbox": [ + 313, + 316, + 556, + 496 + ], + "type": "text", + "content": ". Interestingly, as shown in Figure 2, the conditional noise prediction " + }, + { + "bbox": [ + 313, + 316, + 556, + 496 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(x_t,e_p) - \\epsilon_{\\theta}(x_t,e_{\\emptyset})" + }, + { + "bbox": [ + 313, + 316, + 556, + 496 + ], + "type": "text", + "content": " remains high during the initial time steps and then steeply falls at the time step immediately before the denoising trajectory exits the attraction basin. We refer to this time step as the transition point, marking the change from producing memorized outputs to non-memorized outputs. In Figure 2(a), if you were to start applying CFG at any point on or before " + }, + { + "bbox": [ + 313, + 316, + 556, + 496 + ], + "type": "inline_equation", + "content": "t = 779" + }, + { + "bbox": [ + 313, + 316, + 556, + 496 + ], + "type": "text", + "content": ", you get the same memorized output. However, surprisingly, applying CFG at exactly one denoising step after this " + }, + { + "bbox": [ + 313, + 316, + 556, + 496 + ], + "type": "inline_equation", + "content": "t = 759" + }, + { + "bbox": [ + 313, + 316, + 556, + 496 + ], + "type": "text", + "content": " leads to a non-memorized output (step size is 20 when denoising with 50 inference steps). We define the transition point as follows." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 507, + 555, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 507, + 555, + 567 + ], + "spans": [ + { + "bbox": [ + 313, + 507, + 555, + 567 + ], + "type": "text", + "content": "Definition 3 (Transition point). Given prompt embeddings, let " + }, + { + "bbox": [ + 313, + 507, + 555, + 567 + ], + "type": "inline_equation", + "content": "x^{\\mathrm{a}} \\in X" + }, + { + "bbox": [ + 313, + 507, + 555, + 567 + ], + "type": "text", + "content": " be an attractor. We define a transition point as " + }, + { + "bbox": [ + 313, + 507, + 555, + 567 + ], + "type": "inline_equation", + "content": "(x_{\\tau}, \\tau)" + }, + { + "bbox": [ + 313, + 507, + 555, + 567 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 313, + 507, + 555, + 567 + ], + "type": "inline_equation", + "content": "x_{\\tau} \\in X" + }, + { + "bbox": [ + 313, + 507, + 555, + 567 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 507, + 555, + 567 + ], + "type": "inline_equation", + "content": "\\tau > 0" + }, + { + "bbox": [ + 313, + 507, + 555, + 567 + ], + "type": "text", + "content": " such that " + }, + { + "bbox": [ + 313, + 507, + 555, + 567 + ], + "type": "inline_equation", + "content": "(x_{\\tau}, \\tau + 1)" + }, + { + "bbox": [ + 313, + 507, + 555, + 567 + ], + "type": "text", + "content": " lies within its attraction basin, while " + }, + { + "bbox": [ + 313, + 507, + 555, + 567 + ], + "type": "inline_equation", + "content": "(x_{\\tau}, \\tau)" + }, + { + "bbox": [ + 313, + 507, + 555, + 567 + ], + "type": "text", + "content": " does not. More formally, an " + }, + { + "bbox": [ + 313, + 507, + 555, + 567 + ], + "type": "inline_equation", + "content": "((\\epsilon, \\delta))" + }, + { + "bbox": [ + 313, + 507, + 555, + 567 + ], + "type": "text", + "content": "-transition point satisfies" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 343, + 578, + 526, + 592 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 343, + 578, + 526, + 592 + ], + "spans": [ + { + "bbox": [ + 343, + 578, + 526, + 592 + ], + "type": "interline_equation", + "content": "\\mathbb {P} \\left(\\varphi \\left(x _ {\\tau}, \\tau + 1\\right) \\in B _ {\\mathcal {D}} \\left(x ^ {\\mathrm {a}}, \\epsilon\\right)\\right) > 1 - \\delta \\quad a n d", + "image_path": "0c006220a22a383613cd277d23042e1fec5995fbfbfd171684111de12293daeb.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 343, + 594, + 485, + 607 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 343, + 594, + 485, + 607 + ], + "spans": [ + { + "bbox": [ + 343, + 594, + 485, + 607 + ], + "type": "interline_equation", + "content": "\\mathbb {P} (\\varphi (x _ {\\tau}, \\tau) \\notin B _ {\\mathcal {D}} (x ^ {\\mathrm {a}}, \\epsilon)) < 1 - \\delta ,", + "image_path": "c49112705c874cdf8b2a4726b5f2fd99e7f1f0d9e6142b8d3aeb34a10d287260.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 618, + 531, + 630 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 618, + 531, + 630 + ], + "spans": [ + { + "bbox": [ + 313, + 618, + 531, + 630 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 618, + 531, + 630 + ], + "type": "inline_equation", + "content": "B_{\\mathcal{D}}" + }, + { + "bbox": [ + 313, + 618, + 531, + 630 + ], + "type": "text", + "content": " is a perceptual ball defined in Definition 2." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 642, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 554, + 714 + ], + "type": "text", + "content": "This phenomenon occurs only for memorized samples as illustrated in Figure 6, the conditional guidance " + }, + { + "bbox": [ + 313, + 642, + 554, + 714 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(x_t,e_p) - \\epsilon_{\\theta}(x_t,e_{\\emptyset})" + }, + { + "bbox": [ + 313, + 642, + 554, + 714 + ], + "type": "text", + "content": " remains low and almost flat across different time steps for non-memorized samples. We further analyze the diffusion trajectory during memorization, demonstrating the presence of an attraction basin in Appendix 8." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12874" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 71, + 209, + 86 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 71, + 209, + 86 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 209, + 86 + ], + "type": "text", + "content": "5. Simple Mitigation Strategy" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 152 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 152 + ], + "type": "text", + "content": "Based on the intuition from the previous section, we can simply mitigate memorization by finding an ideal transition point to switch from zero CFG to applying CFG. Based on our preliminary experiments, we found that transition points can either be static or dynamic." + } + ] + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 93, + 163, + 260, + 270 + ], + "blocks": [ + { + "bbox": [ + 93, + 163, + 260, + 270 + ], + "lines": [ + { + "bbox": [ + 93, + 163, + 260, + 270 + ], + "spans": [ + { + "bbox": [ + 93, + 163, + 260, + 270 + ], + "type": "image", + "image_path": "be1c77a4df291f24f0d967c3297837186a10fc276eeab0f141b6a03ec6de2d81.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 275, + 296, + 331 + ], + "lines": [ + { + "bbox": [ + 55, + 275, + 296, + 331 + ], + "spans": [ + { + "bbox": [ + 55, + 275, + 296, + 331 + ], + "type": "text", + "content": "Figure 3. Average magnitude of the text-conditioned " + }, + { + "bbox": [ + 55, + 275, + 296, + 331 + ], + "type": "inline_equation", + "content": "(\\epsilon_{\\theta}(x_t, e_p))" + }, + { + "bbox": [ + 55, + 275, + 296, + 331 + ], + "type": "text", + "content": " and unconditional noise predictions " + }, + { + "bbox": [ + 55, + 275, + 296, + 331 + ], + "type": "inline_equation", + "content": "(\\epsilon_{\\theta}(x_t, e_{\\emptyset}))" + }, + { + "bbox": [ + 55, + 275, + 296, + 331 + ], + "type": "text", + "content": " and their difference " + }, + { + "bbox": [ + 55, + 275, + 296, + 331 + ], + "type": "inline_equation", + "content": "(\\epsilon_{\\theta}(x_t, e_p) - \\epsilon_{\\theta}(x_t, e_{\\emptyset}))" + }, + { + "bbox": [ + 55, + 275, + 296, + 331 + ], + "type": "text", + "content": " when applying zero CFG. We see a static transition point (" + }, + { + "bbox": [ + 55, + 275, + 296, + 331 + ], + "type": "inline_equation", + "content": "t = 500" + }, + { + "bbox": [ + 55, + 275, + 296, + 331 + ], + "type": "text", + "content": ") appear when SDv2.1 is finetuned on the LAION-10k dataset [24]." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 58, + 355, + 296, + 504 + ], + "blocks": [ + { + "bbox": [ + 58, + 355, + 296, + 504 + ], + "lines": [ + { + "bbox": [ + 58, + 355, + 296, + 504 + ], + "spans": [ + { + "bbox": [ + 58, + 355, + 296, + 504 + ], + "type": "image", + "image_path": "e87873fd6aeb16f7c3cf38ff8bc82c95c27358ed8b83276fe6b97703b148c80f.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 506, + 296, + 562 + ], + "lines": [ + { + "bbox": [ + 55, + 506, + 296, + 562 + ], + "spans": [ + { + "bbox": [ + 55, + 506, + 296, + 562 + ], + "type": "text", + "content": "Figure 4. Applying CFG before the static transition point " + }, + { + "bbox": [ + 55, + 506, + 296, + 562 + ], + "type": "inline_equation", + "content": "(\\mathrm{T} = 500)" + }, + { + "bbox": [ + 55, + 506, + 296, + 562 + ], + "type": "text", + "content": " leads to memorized outputs while applying CFG after the fixed transition point leads to non-memorized outputs. Applying CFG too late results in poor-quality images that resemble the unconditional generations." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 582, + 297, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 297, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 297, + 715 + ], + "type": "text", + "content": "Static Transition Point. In some trained diffusion models, we observed a universal trend in the mean conditional " + }, + { + "bbox": [ + 55, + 582, + 297, + 715 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(x_t,e_p)" + }, + { + "bbox": [ + 55, + 582, + 297, + 715 + ], + "type": "text", + "content": " and unconditional noise prediction " + }, + { + "bbox": [ + 55, + 582, + 297, + 715 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(x_t,e_{\\emptyset})" + }, + { + "bbox": [ + 55, + 582, + 297, + 715 + ], + "type": "text", + "content": " across different samples. As shown in Figure 3, the mean conditional noise prediction " + }, + { + "bbox": [ + 55, + 582, + 297, + 715 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(x_t,e_p)" + }, + { + "bbox": [ + 55, + 582, + 297, + 715 + ], + "type": "text", + "content": " is higher than unconditional noise prediction " + }, + { + "bbox": [ + 55, + 582, + 297, + 715 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(x_t,e_{\\emptyset})" + }, + { + "bbox": [ + 55, + 582, + 297, + 715 + ], + "type": "text", + "content": " in the initial time steps " + }, + { + "bbox": [ + 55, + 582, + 297, + 715 + ], + "type": "inline_equation", + "content": "(t\\geq 500)" + }, + { + "bbox": [ + 55, + 582, + 297, + 715 + ], + "type": "text", + "content": ". At a static time step " + }, + { + "bbox": [ + 55, + 582, + 297, + 715 + ], + "type": "inline_equation", + "content": "(t = 500)" + }, + { + "bbox": [ + 55, + 582, + 297, + 715 + ], + "type": "text", + "content": ", the two scores align in magnitude until the end of the denoising process, and at the same time, the " + }, + { + "bbox": [ + 55, + 582, + 297, + 715 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 55, + 582, + 297, + 715 + ], + "type": "text", + "content": " distance between them drastically drops. This suggests that the zero CFG trajectory leaves the attraction basin at that fixed time step " + }, + { + "bbox": [ + 55, + 582, + 297, + 715 + ], + "type": "inline_equation", + "content": "(t = 500)" + }, + { + "bbox": [ + 55, + 582, + 297, + 715 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 317, + 72, + 553, + 227 + ], + "blocks": [ + { + "bbox": [ + 317, + 72, + 553, + 227 + ], + "lines": [ + { + "bbox": [ + 317, + 72, + 553, + 227 + ], + "spans": [ + { + "bbox": [ + 317, + 72, + 553, + 227 + ], + "type": "image", + "image_path": "0f61266e6523b9078ce2df0b2203aa2ad178e6a4798481e1b2869589d865117e.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 231, + 555, + 275 + ], + "lines": [ + { + "bbox": [ + 313, + 231, + 555, + 275 + ], + "spans": [ + { + "bbox": [ + 313, + 231, + 555, + 275 + ], + "type": "text", + "content": "Figure 5. In some models, transition points can occur at a different time step for each sample, as seen for pre-trained SDv1.4. In row 1 the transition point is approximately " + }, + { + "bbox": [ + 313, + 231, + 555, + 275 + ], + "type": "inline_equation", + "content": "t = 800" + }, + { + "bbox": [ + 313, + 231, + 555, + 275 + ], + "type": "text", + "content": " while for row 2 it is " + }, + { + "bbox": [ + 313, + 231, + 555, + 275 + ], + "type": "inline_equation", + "content": "t = 650" + }, + { + "bbox": [ + 313, + 231, + 555, + 275 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 297, + 556, + 405 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 297, + 556, + 405 + ], + "spans": [ + { + "bbox": [ + 313, + 297, + 556, + 405 + ], + "type": "text", + "content": "This can be further validated by applying CFG from any time step before the fixed transition point " + }, + { + "bbox": [ + 313, + 297, + 556, + 405 + ], + "type": "inline_equation", + "content": "(t = 500)" + }, + { + "bbox": [ + 313, + 297, + 556, + 405 + ], + "type": "text", + "content": ", as illustrated in Figure 4. Applying CFG from an early time step leads to the same memorized image, but the output immediately switches at the transition point " + }, + { + "bbox": [ + 313, + 297, + 556, + 405 + ], + "type": "inline_equation", + "content": "(t = 500)" + }, + { + "bbox": [ + 313, + 297, + 556, + 405 + ], + "type": "text", + "content": " such that we generate text-aligned non-memorized images. On the other hand, if we start applying CFG very late in the denoising process, the outputs are not well aligned with the text prompts and resemble their unconditional generations." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 411, + 556, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 411, + 556, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 411, + 556, + 628 + ], + "type": "text", + "content": "Dynamic Transition Point. In other trained diffusion models, we saw that every prompt and initialization pair leads to a different transition point. As shown in Figures 6(a) and 5, different trajectories leave the attraction basin at different time steps, characterized by their fall in the conditional guidance " + }, + { + "bbox": [ + 313, + 411, + 556, + 628 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(x_t,e_p) - \\epsilon_{\\theta}(x_t,e_{\\emptyset})" + }, + { + "bbox": [ + 313, + 411, + 556, + 628 + ], + "type": "text", + "content": ". We found that it is dependent on both the conditional embedding " + }, + { + "bbox": [ + 313, + 411, + 556, + 628 + ], + "type": "inline_equation", + "content": "e_p" + }, + { + "bbox": [ + 313, + 411, + 556, + 628 + ], + "type": "text", + "content": " and the initialization " + }, + { + "bbox": [ + 313, + 411, + 556, + 628 + ], + "type": "inline_equation", + "content": "x_{T}" + }, + { + "bbox": [ + 313, + 411, + 556, + 628 + ], + "type": "text", + "content": ". We show an example in Figures 2(a) and 2(b), where the transition point occurs at different time steps even for the same prompt. Thus, we propose a method to find this dynamic transition point (DTP). As shown in Figure 2, it is the point after the first local minima in the graph. Thus, to mitigate memorization, this difference is tracked until the first local minimum occurs, and CFG is applied henceforth. This process requires no additional computations as the text-conditioned and unconditional noise predictions are computed by default when denoising using CFG. We present the pseudo-code for this approach in Algorithm 1." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 635, + 427, + 648 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 635, + 427, + 648 + ], + "spans": [ + { + "bbox": [ + 313, + 635, + 427, + 648 + ], + "type": "text", + "content": "5.1. Opposite Guidance" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 653, + 555, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 555, + 701 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 555, + 701 + ], + "type": "text", + "content": "An issue with this method is that when the transition point occurs very late (" + }, + { + "bbox": [ + 313, + 653, + 555, + 701 + ], + "type": "inline_equation", + "content": "t \\leq 500" + }, + { + "bbox": [ + 313, + 653, + 555, + 701 + ], + "type": "text", + "content": ") in the denoising process, CFG is applied for fewer time steps, and this can impact the image quality." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 326, + 701, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 701, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 326, + 701, + 556, + 714 + ], + "type": "text", + "content": "To ensure that the transition point occurs earlier, we in" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12875" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 93, + 71, + 258, + 177 + ], + "blocks": [ + { + "bbox": [ + 93, + 71, + 258, + 177 + ], + "lines": [ + { + "bbox": [ + 93, + 71, + 258, + 177 + ], + "spans": [ + { + "bbox": [ + 93, + 71, + 258, + 177 + ], + "type": "image", + "image_path": "3f50b349a1bd99b69a39980048240d1d08f250ab20741fe040d0787ff82c2cab.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 137, + 180, + 214, + 191 + ], + "lines": [ + { + "bbox": [ + 137, + 180, + 214, + 191 + ], + "spans": [ + { + "bbox": [ + 137, + 180, + 214, + 191 + ], + "type": "text", + "content": "(a) Applying zero CFG" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 93, + 193, + 258, + 298 + ], + "blocks": [ + { + "bbox": [ + 93, + 193, + 258, + 298 + ], + "lines": [ + { + "bbox": [ + 93, + 193, + 258, + 298 + ], + "spans": [ + { + "bbox": [ + 93, + 193, + 258, + 298 + ], + "type": "image", + "image_path": "1f3f1140f2a1236b8917829a23e348de47fd494d97987dca69da77361cea0132.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 123, + 302, + 227, + 312 + ], + "lines": [ + { + "bbox": [ + 123, + 302, + 227, + 312 + ], + "spans": [ + { + "bbox": [ + 123, + 302, + 227, + 312 + ], + "type": "text", + "content": "(b) Applying opposite guidance" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 316, + 296, + 350 + ], + "lines": [ + { + "bbox": [ + 55, + 316, + 296, + 350 + ], + "spans": [ + { + "bbox": [ + 55, + 316, + 296, + 350 + ], + "type": "text", + "content": "Figure 6. Magnitude of " + }, + { + "bbox": [ + 55, + 316, + 296, + 350 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(x_t, e_p) - \\epsilon_{\\theta}(x_t, e_{\\emptyset})" + }, + { + "bbox": [ + 55, + 316, + 296, + 350 + ], + "type": "text", + "content": " when denoising (a) without CFG and (b) with opposite guidance. Transition point occurs sooner with opposite guidance." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "code", + "bbox": [ + 56, + 389, + 296, + 525 + ], + "blocks": [ + { + "bbox": [ + 55, + 362, + 296, + 386 + ], + "lines": [ + { + "bbox": [ + 55, + 362, + 296, + 386 + ], + "spans": [ + { + "bbox": [ + 55, + 362, + 296, + 386 + ], + "type": "text", + "content": "Algorithm 1 Reverse Diffusion with Dynamic Transition Point Method and Opposite Guidance" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 56, + 389, + 296, + 525 + ], + "lines": [ + { + "bbox": [ + 56, + 389, + 296, + 525 + ], + "spans": [ + { + "bbox": [ + 56, + 389, + 296, + 525 + ], + "type": "text", + "content": "Require: " + }, + { + "bbox": [ + 56, + 389, + 296, + 525 + ], + "type": "inline_equation", + "content": "x_{T}\\sim \\mathcal{N}(0,\\mathrm{I}_{d}),\\lambda >0,v_{\\mathrm{OG}}\\in \\{0,1\\}" + }, + { + "bbox": [ + 56, + 389, + 296, + 525 + ], + "type": "text", + "content": " \n1: " + }, + { + "bbox": [ + 56, + 389, + 296, + 525 + ], + "type": "inline_equation", + "content": "d_{T + 2} = -\\infty" + }, + { + "bbox": [ + 56, + 389, + 296, + 525 + ], + "type": "text", + "content": " . " + }, + { + "bbox": [ + 56, + 389, + 296, + 525 + ], + "type": "inline_equation", + "content": "d_{T + 1} = -\\infty" + }, + { + "bbox": [ + 56, + 389, + 296, + 525 + ], + "type": "text", + "content": " \n2: " + }, + { + "bbox": [ + 56, + 389, + 296, + 525 + ], + "type": "inline_equation", + "content": "s = -\\lambda *v_{\\mathrm{OG}}" + }, + { + "bbox": [ + 56, + 389, + 296, + 525 + ], + "type": "text", + "content": " \n3: for " + }, + { + "bbox": [ + 56, + 389, + 296, + 525 + ], + "type": "inline_equation", + "content": "t = T" + }, + { + "bbox": [ + 56, + 389, + 296, + 525 + ], + "type": "text", + "content": " to 1 do \n4: " + }, + { + "bbox": [ + 56, + 389, + 296, + 525 + ], + "type": "inline_equation", + "content": "d_{t} = \\| \\epsilon_{\\theta}(x_{t},e_{p}) - \\epsilon_{\\theta}(x_{t},e_{\\emptyset})\\|_{2}^{2}" + }, + { + "bbox": [ + 56, + 389, + 296, + 525 + ], + "type": "text", + "content": " \n5: if " + }, + { + "bbox": [ + 56, + 389, + 296, + 525 + ], + "type": "inline_equation", + "content": "d_{t + 2} > d_{t + 1}" + }, + { + "bbox": [ + 56, + 389, + 296, + 525 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 56, + 389, + 296, + 525 + ], + "type": "inline_equation", + "content": "d_{t + 1} < d_t" + }, + { + "bbox": [ + 56, + 389, + 296, + 525 + ], + "type": "text", + "content": " then \n6: " + }, + { + "bbox": [ + 56, + 389, + 296, + 525 + ], + "type": "inline_equation", + "content": "s = \\lambda" + }, + { + "bbox": [ + 56, + 389, + 296, + 525 + ], + "type": "text", + "content": " \n7: end if \n8: " + }, + { + "bbox": [ + 56, + 389, + 296, + 525 + ], + "type": "inline_equation", + "content": "\\hat{\\epsilon} = \\epsilon_{\\theta}(x_t,e_\\emptyset) + s(\\epsilon_\\theta (x_t,e_p) - \\epsilon_\\theta (x_t,e_\\emptyset))" + }, + { + "bbox": [ + 56, + 389, + 296, + 525 + ], + "type": "text", + "content": " \n9: " + }, + { + "bbox": [ + 56, + 389, + 296, + 525 + ], + "type": "inline_equation", + "content": "x_{t - 1} = \\frac{1}{\\sqrt{\\alpha}_t} (x_t - \\frac{1 - \\alpha_t}{\\sqrt{1 - \\alpha_t}}\\hat{\\epsilon})" + }, + { + "bbox": [ + 56, + 389, + 296, + 525 + ], + "type": "text", + "content": " \n10: end for \n11: return " + }, + { + "bbox": [ + 56, + 389, + 296, + 525 + ], + "type": "inline_equation", + "content": "x_0" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "code_body" + } + ], + "index": 6, + "sub_type": "algorithm" + }, + { + "bbox": [ + 55, + 546, + 297, + 631 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 546, + 297, + 631 + ], + "spans": [ + { + "bbox": [ + 55, + 546, + 297, + 631 + ], + "type": "text", + "content": "troduce a new concept of opposite guidance (OG). Unlike negative prompting, we apply negative or opposite CFG, to push the diffusion trajectory in the opposite direction to that of traditional CFG and thus, push the trajectory away from the attraction basin of the traditional CFG trajectory sooner in the denoising process, as also shown in Figure 6(b). Opposite guidance can be expressed as follows for " + }, + { + "bbox": [ + 55, + 546, + 297, + 631 + ], + "type": "inline_equation", + "content": "s > 0" + }, + { + "bbox": [ + 55, + 546, + 297, + 631 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 85, + 651, + 296, + 681 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 651, + 296, + 681 + ], + "spans": [ + { + "bbox": [ + 85, + 651, + 296, + 681 + ], + "type": "interline_equation", + "content": "\\hat {\\epsilon} \\leftarrow \\epsilon_ {\\theta} \\left(x _ {t}, e _ {\\emptyset}\\right) - s \\underbrace {\\left(\\epsilon_ {\\theta} \\left(x _ {t} , e _ {p}\\right) - \\epsilon_ {\\theta} \\left(x _ {t} , e _ {\\emptyset}\\right)\\right)} _ {\\text {c o n d i t i o n a l g u i d a n c e}}. \\tag {5}", + "image_path": "8297fc37374ab8bd3d521960939a0ec686f22a75146c5410fdbcc69936bd5fcb.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "Similar to unconditional denoising, when denoising using opposite guidance, there is an ideal time step where" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "type": "text", + "content": "the magnitude of conditional guidance falls. At this time step, switching from opposite guidance to traditional positive guidance leads to non-memorized samples with high image quality and text alignment. Thus, it can easily be integrated along with the previous method to find a static or dynamic transition point. We present the pseudo-code for this approach in conjunction with the DTP method in Algorithm 1 with the parameter " + }, + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "type": "inline_equation", + "content": "v_{\\mathrm{OG}} = 1" + }, + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 179, + 537, + 192 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 179, + 537, + 192 + ], + "spans": [ + { + "bbox": [ + 313, + 179, + 537, + 192 + ], + "type": "text", + "content": "6. Memorization under Different Scenarios" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 199, + 555, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 199, + 555, + 247 + ], + "spans": [ + { + "bbox": [ + 313, + 199, + 555, + 247 + ], + "type": "text", + "content": "We study memorization under different scenarios and show how this simple approach of finding an ideal transition point to avoid the attraction basin can be applied to effectively mitigate memorization." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 249, + 555, + 465 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 314, + 249, + 555, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 249, + 555, + 285 + ], + "spans": [ + { + "bbox": [ + 314, + 249, + 555, + 285 + ], + "type": "text", + "content": "- Scenario 1: Fine-tuned SDv2.1 on 10,000 LAION datapoints (followed by [24]). Memorization occurs due to overfitting on the small dataset." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 314, + 285, + 555, + 333 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 285, + 555, + 333 + ], + "spans": [ + { + "bbox": [ + 314, + 285, + 555, + 333 + ], + "type": "text", + "content": "- Scenario 2: Fine-tuned SDv2.1 on the Imagenette dataset [12] containing 10 ImageNet classes (followed by [24]). Exact memorization is not always observed but similarity with the training dataset increases." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 314, + 334, + 554, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 334, + 554, + 380 + ], + "spans": [ + { + "bbox": [ + 314, + 334, + 554, + 380 + ], + "type": "text", + "content": "- Scenario 3: Fine-tuned SDv1.4 on 200 prompts duplicated 200 times with an additional 120,000 prompts that were not duplicated (followed by [28]). Memorization occurs due to data duplication." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 381, + 555, + 465 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 381, + 555, + 465 + ], + "spans": [ + { + "bbox": [ + 313, + 381, + 555, + 465 + ], + "type": "text", + "content": "- Scenario 4: 500 memorized prompts for pre-trained SDv1.4 found using membership inference attack [4, 27] (followed by [20]). The exact cause of memorization for each prompt is unknown, but the strong presence of \"trigger tokens\" is noticeable. These are in the form of nouns such as celebrity and movie names with little to no additional descriptions. Results are in Appendix 9." + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 467, + 556, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 467, + 556, + 528 + ], + "spans": [ + { + "bbox": [ + 313, + 467, + 556, + 528 + ], + "type": "text", + "content": "Notably, different researchers have independently studied these different scenarios, but, as we show later, their approaches do not generalize to other scenarios as memorization occurs due to different causes in each scenario. Refer to Appendix 11 for other implementation details." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 532, + 556, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 532, + 556, + 593 + ], + "spans": [ + { + "bbox": [ + 313, + 532, + 556, + 593 + ], + "type": "text", + "content": "Evaluation Metrics We utilize the similarity metric based on the SSCD [18] embeddings to judge the level of memorization [24]; the CLIP score [19] to assess how well the generated images align with the text prompts; and the Fréchet Inception Distance (FID) [9] for image quality." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 600, + 386, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 600, + 386, + 612 + ], + "spans": [ + { + "bbox": [ + 313, + 600, + 386, + 612 + ], + "type": "text", + "content": "6.1. Scenario 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 618, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 618, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 618, + 556, + 713 + ], + "type": "text", + "content": "We observed that in this scenario, a static transition point at " + }, + { + "bbox": [ + 313, + 618, + 556, + 713 + ], + "type": "inline_equation", + "content": "t = 500" + }, + { + "bbox": [ + 313, + 618, + 556, + 713 + ], + "type": "text", + "content": " occurs as shown in Figure 3. We present qualitative and quantitative results in Figure 7 and Table 1, respectively. Our approach outperformed all previous methods, and we observed the previously proposed techniques that focused solely on the presence of \"trigger tokens\" [20, 28] did not work well. This shows that when memorization occurs due to factors other than data duplication or the presence of" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "12876" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 144, + 69, + 470, + 350 + ], + "blocks": [ + { + "bbox": [ + 144, + 69, + 470, + 350 + ], + "lines": [ + { + "bbox": [ + 144, + 69, + 470, + 350 + ], + "spans": [ + { + "bbox": [ + 144, + 69, + 470, + 350 + ], + "type": "image", + "image_path": "b1c0a9055535ea261072cbbe07be2c967b5f1a18aeea5917f71058e67642405d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 351, + 555, + 386 + ], + "lines": [ + { + "bbox": [ + 54, + 351, + 555, + 386 + ], + "spans": [ + { + "bbox": [ + 54, + 351, + 555, + 386 + ], + "type": "text", + "content": "Figure 7. Qualitative results comparing the proposed approach with the baselines in Scenario 1. Ren et al. [20] is unable to mitigate memorization in most cases while Wen et al. [28] has poor image quality and text-alignment. More enlarged examples (Fig. 14) and prompts for this figure are in Appendix 12." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 58, + 452, + 296, + 537 + ], + "blocks": [ + { + "bbox": [ + 55, + 404, + 297, + 449 + ], + "lines": [ + { + "bbox": [ + 55, + 404, + 297, + 449 + ], + "spans": [ + { + "bbox": [ + 55, + 404, + 297, + 449 + ], + "type": "text", + "content": "Table 1. Comparison of results on Scenario 1. None of the other approaches perform well. The only comparable method [28] qualitatively results in poorer quality images (see visual examples in Fig. 7)." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 452, + 296, + 537 + ], + "lines": [ + { + "bbox": [ + 58, + 452, + 296, + 537 + ], + "spans": [ + { + "bbox": [ + 58, + 452, + 296, + 537 + ], + "type": "table", + "html": "
Similarity (95pc)CLIP ScoreFID
No Mitigation0.65040.302716.8373
Add rand word [24]0.52540.294117.4142
Add rand numb [24]0.54160.299317.0245
Wen et al. [28]0.38530.289516.7176
Ren et al. [20]0.60280.295920.2931
Ours (STP)0.28570.297619.8494
Ours (OG + STP)0.38110.302015.6679
", + "image_path": "2d72096b01a19ad8f646714819e8a9002b2d2ec4dc32ac167e171b2eb4feffa7.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 556, + 296, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 556, + 296, + 593 + ], + "spans": [ + { + "bbox": [ + 55, + 556, + 296, + 593 + ], + "type": "text", + "content": "trigger tokens, these methods may fail. We present further results when finetuning on the LAION-100k dataset [24] in Appendix 10." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 600, + 128, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 600, + 128, + 612 + ], + "spans": [ + { + "bbox": [ + 55, + 600, + 128, + 612 + ], + "type": "text", + "content": "6.2. Scenario 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "text", + "content": "Even in this scenario, as illustrated in Figure 10, a static transition point occurs. We applied CFG from " + }, + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "inline_equation", + "content": "t = 700" + }, + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "inline_equation", + "content": "t = 600" + }, + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "text", + "content": " to show that we can reduce the similarity with the training dataset at no cost to the image quality, as shown in Table 2. We would like to point out that previous approaches are specific to text-to-image models and thus cannot be extrapolated to other conditioning mechanisms. This shows a further advantage of our CFG-based mitigation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 407, + 359, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 407, + 359, + 418 + ], + "spans": [ + { + "bbox": [ + 314, + 407, + 359, + 418 + ], + "type": "text", + "content": "technique." + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 333, + 476, + 536, + 529 + ], + "blocks": [ + { + "bbox": [ + 313, + 429, + 555, + 473 + ], + "lines": [ + { + "bbox": [ + 313, + 429, + 555, + 473 + ], + "spans": [ + { + "bbox": [ + 313, + 429, + 555, + 473 + ], + "type": "text", + "content": "Table 2. Results on Scenario 2. The value in the brackets in the time steps between which we apply CFG. Applying CFG later from " + }, + { + "bbox": [ + 313, + 429, + 555, + 473 + ], + "type": "inline_equation", + "content": "t = 700" + }, + { + "bbox": [ + 313, + 429, + 555, + 473 + ], + "type": "text", + "content": " (row 2) and " + }, + { + "bbox": [ + 313, + 429, + 555, + 473 + ], + "type": "inline_equation", + "content": "t = 600" + }, + { + "bbox": [ + 313, + 429, + 555, + 473 + ], + "type": "text", + "content": " (row 3) reduces similarity with comparable or improved FID score." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 333, + 476, + 536, + 529 + ], + "lines": [ + { + "bbox": [ + 333, + 476, + 536, + 529 + ], + "spans": [ + { + "bbox": [ + 333, + 476, + 536, + 529 + ], + "type": "table", + "html": "
Similarity (95pc)FID
No mitigation0.370243.35
CFG=7.5 [700,0]0.301838.86
CFG=7.5 [600,0]0.275645.73
", + "image_path": "93ad06585487a5c36d2920625bd95a88b58e551bf062f965babb115d1f7d1cde.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 548, + 387, + 559 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 548, + 387, + 559 + ], + "spans": [ + { + "bbox": [ + 314, + 548, + 387, + 559 + ], + "type": "text", + "content": "6.3. Scenario 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 565, + 555, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 565, + 555, + 673 + ], + "spans": [ + { + "bbox": [ + 313, + 565, + 555, + 673 + ], + "type": "text", + "content": "In this scenario, we observed that a dynamic transition point occurs. We experimentally validate our approach of DTP method and opposite guidance qualitatively and quantitatively in Figure 8 and Table 3. We saw that our method performs at par with [28] that studied this specific scenario and even outperforms it in terms of memorization in the top 95 percentile. Approaches by Ren et al. [20] and Somepalli et al.[24] that were proposed for other scenarios do not work well, showcasing their lack of generalizability." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "text", + "content": "Inference Time. One of the major advantages of our approach is its simplicity and lack of computational time overhead. Image generation using our approach as well as stan-" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12877" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 145, + 70, + 470, + 353 + ], + "blocks": [ + { + "bbox": [ + 145, + 70, + 470, + 353 + ], + "lines": [ + { + "bbox": [ + 145, + 70, + 470, + 353 + ], + "spans": [ + { + "bbox": [ + 145, + 70, + 470, + 353 + ], + "type": "image", + "image_path": "56f91106059d11df40a99e806c65b6ebbf7aba033ab488b7dc661fdfb66c135f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 357, + 555, + 391 + ], + "lines": [ + { + "bbox": [ + 55, + 357, + 555, + 391 + ], + "spans": [ + { + "bbox": [ + 55, + 357, + 555, + 391 + ], + "type": "text", + "content": "Figure 8. Qualitative results comparing the proposed approach with the baselines in Scenario 3. Prompts used to generate these figures are given in the Appendix. Rows 1, 2, 5, and 6 show examples, where for most baselines the output images remain closely related to their memorized versions. More enlarged examples (Fig. 15) and prompts for this figure are in Appendix 13." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 57, + 446, + 296, + 538 + ], + "blocks": [ + { + "bbox": [ + 55, + 409, + 295, + 443 + ], + "lines": [ + { + "bbox": [ + 55, + 409, + 295, + 443 + ], + "spans": [ + { + "bbox": [ + 55, + 409, + 295, + 443 + ], + "type": "text", + "content": "Table 3. Table showcasing results in Scenario 3. Wen et al. [28] studied this scenario and is the only other method which performs well but does not generalize to other scenarios (see Table 1)." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 446, + 296, + 538 + ], + "lines": [ + { + "bbox": [ + 57, + 446, + 296, + 538 + ], + "spans": [ + { + "bbox": [ + 57, + 446, + 296, + 538 + ], + "type": "table", + "html": "
SimilarityCLIP ScoreFID
95pcMean ± Std
No Mitigation0.79770.5513 ± 0.160.3105106.49
Add rand word [24]0.74800.4312 ± 0.210.3071116.92
Add rand numb [24]0.73660.3850 ± 0.210.3027126.54
Wen et al. [28] (l_target=3)0.77470.5147 ± 0.170.3100109.28
Wen et al. [28] (l_target=1)0.60380.2808 ± 0.160.3050136.34
Ren et al. [20]0.68810.4036 ± 0.160.3066124.38
Ours (DTP)0.58850.2866 ± 0.160.3020138.92
Ours (OG + DTP)0.69150.2844 ± 0.200.2910140.05
", + "image_path": "9668c0ba64d63a8eceb6330fb0d8712c7888df80d598b993e26090acf724b63a.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 559, + 296, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 559, + 296, + 620 + ], + "spans": [ + { + "bbox": [ + 55, + 559, + 296, + 620 + ], + "type": "text", + "content": "dard CFG-based image generation using SDv2.1 on one A100 GPU takes 1.26 seconds, while in comparison the mitigation technique proposed by Wen et al. [28] took 2.86 seconds and that of Ren et al. [20] took 2.08 seconds. These are the average values across 10,000 generations." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "content": "Limitations. As transition points do not exist for nonmemorized samples, we require detecting memorization before we can apply our approach. This can either be done by denoising twice, once with traditional CFG to look at the trends in the magnitude of conditional guidance and the second time to apply our mitigation technique if the prompt is memorized. Alternatively, as [20, 28] showed, we can" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 411, + 555, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 411, + 555, + 472 + ], + "spans": [ + { + "bbox": [ + 313, + 411, + 555, + 472 + ], + "type": "text", + "content": "do detection at " + }, + { + "bbox": [ + 313, + 411, + 555, + 472 + ], + "type": "inline_equation", + "content": "t = 0" + }, + { + "bbox": [ + 313, + 411, + 555, + 472 + ], + "type": "text", + "content": " by simply looking at the magnitude of conditional guidance " + }, + { + "bbox": [ + 313, + 411, + 555, + 472 + ], + "type": "inline_equation", + "content": "\\| \\epsilon_{\\theta}(x_t, e_p) - \\epsilon_{\\theta}(x_t, e_{\\emptyset}) \\|_2^2" + }, + { + "bbox": [ + 313, + 411, + 555, + 472 + ], + "type": "text", + "content": ". If this magnitude is above a certain threshold, it indicates that the sample is memorized with high accuracy and we can apply our guidance accordingly." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 494, + 388, + 506 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 494, + 388, + 506 + ], + "spans": [ + { + "bbox": [ + 314, + 494, + 388, + 506 + ], + "type": "text", + "content": "7. Conclusion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 518, + 556, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 518, + 556, + 662 + ], + "spans": [ + { + "bbox": [ + 313, + 518, + 556, + 662 + ], + "type": "text", + "content": "We present a novel understanding of the dynamics behind memorization in conditional diffusion models. We showcase the presence of an attraction basin that steers randomly initialized latent vectors towards a memorized output. We propose an approach to steer away from the attraction basin and detect the point at which the trajectory leaves the attraction basin, referred to as the transition point. We show that applying CFG after the transition point leads to non-memorized outputs. We presented results in various scenarios such as the presence of data replication, fine-tuning on smaller datasets, and memorized examples found in existing pre-trained models to showcase its efficiency." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "content": "To summarize, our method is simple, works well across all tested scenarios in terms of both image quality metrics and visual inspection, requires less computational time, and does not require altering the prompt." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12878" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 297, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 91, + 297, + 124 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 297, + 124 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 297, + 124 + ], + "type": "text", + "content": "[1] Luca Ambrogioni. In search of dispersed memories: Generative diffusion models are associative memory networks. Entropy, 26(5):381, 2024. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 125, + 295, + 147 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 125, + 295, + 147 + ], + "spans": [ + { + "bbox": [ + 61, + 125, + 295, + 147 + ], + "type": "text", + "content": "[2] Joseph Auslander, Nam P Bhatia, and Peter Seibert. Attractors in dynamical systems. Technical report, 1964. 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 149, + 295, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 149, + 295, + 192 + ], + "spans": [ + { + "bbox": [ + 62, + 149, + 295, + 192 + ], + "type": "text", + "content": "[3] Yuanhao Ban, Ruochen Wang, Tianyi Zhou, Minhao Cheng, Boqing Gong, and Cho-Jui Hsieh. Understanding the impact of negative prompts: When and how do they take effect? arXiv preprint arXiv:2406.02965, 2024. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 194, + 295, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 194, + 295, + 249 + ], + "spans": [ + { + "bbox": [ + 62, + 194, + 295, + 249 + ], + "type": "text", + "content": "[4] Nicolas Carlini, Jamie Hayes, Milad Nasr, Matthew Jagielski, Vikash Sehwag, Florian Tramer, Borja Balle, Daphne Ippolito, and Eric Wallace. Extracting training data from diffusion models. In 32nd USENIX Security Symposium (USENIX Security 23), pages 5253-5270, 2023. 1, 2, 3, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 251, + 296, + 305 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 251, + 296, + 305 + ], + "spans": [ + { + "bbox": [ + 62, + 251, + 296, + 305 + ], + "type": "text", + "content": "[5] Huiwen Chang, Han Zhang, Jarred Barber, AJ Maschinot, Jose Lezama, Lu Jiang, Ming-Hsuan Yang, Kevin Murphy, William T Freeman, Michael Rubinstein, et al. Muse: Text-to-image generation via masked generative transformers. arXiv preprint arXiv:2301.00704, 2023. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 307, + 295, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 307, + 295, + 350 + ], + "spans": [ + { + "bbox": [ + 62, + 307, + 295, + 350 + ], + "type": "text", + "content": "[6] Ruchika Chavhan, Ondrej Bohdal, Yongshuo Zong, Da Li, and Timothy Hospedales. Memorized images in diffusion models share a subspace that can be located and deleted. arXiv preprint arXiv:2406.18566, 2024. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 352, + 296, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 352, + 296, + 396 + ], + "spans": [ + { + "bbox": [ + 62, + 352, + 296, + 396 + ], + "type": "text", + "content": "[7] Chen Chen, Daochang Liu, and Chang Xu. Towards memorization-free diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8425-8434, 2024. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 397, + 295, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 397, + 295, + 441 + ], + "spans": [ + { + "bbox": [ + 62, + 397, + 295, + 441 + ], + "type": "text", + "content": "[8] Giannis Daras, Alexandros G Dimakis, and Constantinos Daskalakis. Consistent diffusion meets tweedie: Training exact ambient diffusion models with noisy data. arXiv preprint arXiv:2404.10177, 2024. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 443, + 296, + 496 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 443, + 296, + 496 + ], + "spans": [ + { + "bbox": [ + 62, + 443, + 296, + 496 + ], + "type": "text", + "content": "[9] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 499, + 294, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 499, + 294, + 521 + ], + "spans": [ + { + "bbox": [ + 57, + 499, + 294, + 521 + ], + "type": "text", + "content": "[10] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 523, + 295, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 523, + 295, + 576 + ], + "spans": [ + { + "bbox": [ + 57, + 523, + 295, + 576 + ], + "type": "text", + "content": "[11] Benjamin Hoover, Hendrik Strobelt, Dmitry Krotov, Judy Hoffman, Zsolt Kira, and Duen Horng Chau. Memory in plain sight: Surveying the uncanny resemblances of associative memories and diffusion models. arXiv preprint arXiv:2309.16750, 2023. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 579, + 294, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 579, + 294, + 601 + ], + "spans": [ + { + "bbox": [ + 57, + 579, + 294, + 601 + ], + "type": "text", + "content": "[12] Jeremy Howard. Imagenette: A smaller subset of 10 easily classified classes from imagenet, 2019. 6, 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 602, + 295, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 602, + 295, + 656 + ], + "spans": [ + { + "bbox": [ + 57, + 602, + 295, + 656 + ], + "type": "text", + "content": "[13] Tuomas Kynkänniemi, Miika Aittala, Tero Karras, Samuli Laine, Timo Aila, and Jaakko Lehtinen. Applying guidance in a limited interval improves sample and distribution quality in diffusion models. arXiv preprint arXiv:2404.07724, 2024. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 658, + 296, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 658, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 658, + 296, + 712 + ], + "type": "text", + "content": "[14] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps. Advances in Neural Information Processing Systems, 35:5775-5787, 2022. 3" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 316, + 73, + 555, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 555, + 94 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 555, + 94 + ], + "type": "text", + "content": "[15] John W. Milnor. Attractor. *Scholarpedia*, 1(11):1815, 2006. 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 96, + 555, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 96, + 555, + 128 + ], + "spans": [ + { + "bbox": [ + 317, + 96, + 555, + 128 + ], + "type": "text", + "content": "[16] Alex Nichol. Dall-e 2 pre-training mitigations. https://openai.com/index/dall-e-2-pre-training-mitigations/, 2022. [Accessed 10-10-2024]. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 129, + 555, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 129, + 555, + 172 + ], + "spans": [ + { + "bbox": [ + 317, + 129, + 555, + 172 + ], + "type": "text", + "content": "[17] Bao Pham, Gabriel Raya, Matteo Negri, Mohammed J Zaki, Luca Ambrogioni, and Dmitry Krotov. Memorization to generalization: The emergence of diffusion models from associative memory. 2024. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 174, + 555, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 174, + 555, + 227 + ], + "spans": [ + { + "bbox": [ + 317, + 174, + 555, + 227 + ], + "type": "text", + "content": "[18] Ed Pizzi, Sreya Dutta Roy, Sugosh Nagavara Ravindra, Priya Goyal, and Matthijs Douze. A self-supervised descriptor for image copy detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14532-14542, 2022. 3, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 228, + 555, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 228, + 555, + 293 + ], + "spans": [ + { + "bbox": [ + 317, + 228, + 555, + 293 + ], + "type": "text", + "content": "[19] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 294, + 555, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 294, + 555, + 338 + ], + "spans": [ + { + "bbox": [ + 317, + 294, + 555, + 338 + ], + "type": "text", + "content": "[20] Jie Ren, Yaxin Li, Shenglai Zen, Han Xu, Lingjuan Lyu, Yue Xing, and Jiliang Tang. Unveiling and mitigating memorization in text-to-image diffusion models through cross attention. arXiv preprint arXiv:2403.11052, 2024. 1, 2, 3, 6, 7, 8, 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 339, + 555, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 339, + 555, + 392 + ], + "spans": [ + { + "bbox": [ + 317, + 339, + 555, + 392 + ], + "type": "text", + "content": "[21] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 394, + 555, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 394, + 555, + 458 + ], + "spans": [ + { + "bbox": [ + 317, + 394, + 555, + 458 + ], + "type": "text", + "content": "[22] Chitwan Saharia, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural information processing systems, 35:36479-36494, 2022. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 460, + 555, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 460, + 555, + 514 + ], + "spans": [ + { + "bbox": [ + 317, + 460, + 555, + 514 + ], + "type": "text", + "content": "[23] Gowthami Somepalli, Vasu Singla, Micah Goldblum, Jonas Geiping, and Tom Goldstein. Diffusion art or digital forgery? investigating data replication in diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6048-6058, 2023. 1, 2, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 515, + 554, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 515, + 554, + 558 + ], + "spans": [ + { + "bbox": [ + 317, + 515, + 554, + 558 + ], + "type": "text", + "content": "[24] Gowthami Somepalli, Vasu Singla, Micah Goldblum, Jonas Geiping, and Tom Goldstein. Understanding and mitigating copying in diffusion models. Advances in Neural Information Processing Systems, 36:47783-47803, 2023. 1, 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 559, + 555, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 559, + 555, + 591 + ], + "spans": [ + { + "bbox": [ + 317, + 559, + 555, + 591 + ], + "type": "text", + "content": "[25] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 1, 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 593, + 555, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 593, + 555, + 635 + ], + "spans": [ + { + "bbox": [ + 317, + 593, + 555, + 635 + ], + "type": "text", + "content": "[26] Xi Wang, Nicolas Dufour, Nefeli Andreou, Marie-Paule Cani, Victoria Fernández Abrevaya, David Picard, and Vicky Kalogeiton. Analysis of classifier-free guidance weight schedulers. arXiv preprint arXiv:2404.13040, 2024. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 636, + 555, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 636, + 555, + 668 + ], + "spans": [ + { + "bbox": [ + 317, + 636, + 555, + 668 + ], + "type": "text", + "content": "[27] Ryan Webster. A reproducible extraction of training images from diffusion models. arXiv preprint arXiv:2305.08694, 2023. 2, 3, 6" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 670, + 555, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 555, + 713 + ], + "type": "text", + "content": "[28] Yuxin Wen, Yuchen Liu, Chen Chen, and Lingjuan Lyu. Detecting, explaining, and mitigating memorization in diffusion models. In The Twelfth International Conference on Learning Representations, 2024. 1, 2, 3, 4, 6, 7, 8, 5" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "12879" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Classifier-guided CLIP Distillation for Unsupervised Multi-label Classification/325c3ac3-6589-4c70-b2e1-2f7ac7877553_content_list.json b/2025/Classifier-guided CLIP Distillation for Unsupervised Multi-label Classification/325c3ac3-6589-4c70-b2e1-2f7ac7877553_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..9442b9ec2c558086990a905c76737990ce26ae82 --- /dev/null +++ b/2025/Classifier-guided CLIP Distillation for Unsupervised Multi-label Classification/325c3ac3-6589-4c70-b2e1-2f7ac7877553_content_list.json @@ -0,0 +1,1622 @@ +[ + { + "type": "text", + "text": "Classifier-guided CLIP Distillation for Unsupervised Multi-label Classification", + "text_level": 1, + "bbox": [ + 104, + 130, + 893, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Dongseob Kim \nSamsung Electronics, Republic of Korea", + "bbox": [ + 204, + 181, + 527, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "kou.kim@samsung.com", + "bbox": [ + 279, + 219, + 452, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hyunjung Shim KAIST, Republic of Korea", + "bbox": [ + 578, + 181, + 792, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "kateshim@kaist.ac.kr", + "bbox": [ + 591, + 219, + 772, + 232 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 268, + 326, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Multi-label classification is crucial for comprehensive image understanding, yet acquiring accurate annotations is challenging and costly. To address this, a recent study suggests exploiting unsupervised multi-label classification leveraging CLIP, a powerful vision-language model. Despite CLIP's proficiency, it suffers from view-dependent predictions and inherent bias, limiting its effectiveness. We propose a novel method that addresses these issues by leveraging multiple views near target objects, guided by Class Activation Mapping (CAM) of the classifier, and debiasing pseudo-labels derived from CLIP predictions. Our Classifier-guided CLIP Distillation (CCD) enables selecting multiple local views without extra labels and debiasing predictions to enhance classification performance. Experimental results validate our method's superiority over existing techniques across diverse datasets. The code is available at https://github.com/k0u-id/CCD.", + "bbox": [ + 89, + 300, + 483, + 556 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 587, + 220, + 603 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Multi-label classification plays a pivotal role in identifying multiple classes within an image, serving as a foundational component in image understanding [25, 27, 29, 36, 40]. Its applications are widespread, including recommendation systems, surveillance, and image retrieval [43]. However, acquiring multi-label annotations poses challenges due to the difficulty and significant costs involved in meticulously annotating all relevant objects within an image.", + "bbox": [ + 89, + 612, + 483, + 733 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To solve this problem, various label-efficient techniques have been proposed. A partial label setting [5, 10, 14, 20, 24, 30] annotates only a portion of the classes appearing in an image and considers the rest as unknown labels. In the extreme, single positive label settings [7, 19, 35, 42, 51] provide only one positive label per image, with all other labels assumed unknown. However, these methods still incur labeling costs. Besides, current studies simulate partial and single positive settings by randomly omitting labels from a fully labeled dataset, which may not accurately reflect the actual labeling practices. For example, humans tend to pri", + "bbox": [ + 89, + 734, + 483, + 901 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/81948ba10a0ea8b6dcd29b979461ede4a68b1358709dd6f4d8993e4a0a941df9.jpg", + "image_caption": [ + "H. 0.99 / P. 0.00 (a) original image", + "Figure 1. CLIP prediction probability of horse (H) and person (P) corresponding to different input patches from single image: (a) The original image and its CLIP prediction. (b) The $1/4$ sized patch and its CLIP prediction. (c) The $1/9$ sized patch and its CLIP prediction. (d) The $1/16$ sized patch and its CLIP prediction. The same colored box indicates each patch cropped from the corresponding box." + ], + "image_footnote": [], + "bbox": [ + 516, + 267, + 609, + 339 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0745cb1f1b35809050fcfae8df1076a832f20e325b1129bda770d7aa8bcaa314.jpg", + "image_caption": [ + "H:0.96 /P:0.00 (b)1/4 patch" + ], + "image_footnote": [], + "bbox": [ + 614, + 267, + 707, + 339 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/32655c6fe797e810dd6eefd3686ad278d2297679375e08c863db5d8139fc0794.jpg", + "image_caption": [ + "H:0.14/P:0.12 (c) 1/9" + ], + "image_footnote": [], + "bbox": [ + 710, + 267, + 805, + 339 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1c5bd6b89078ae350b37dd22545cf6fbae0052265c6e88ef062ccce7b84a042e.jpg", + "image_caption": [ + ":0.05/P:0.65 (d) 1/16 patch" + ], + "image_footnote": [], + "bbox": [ + 810, + 268, + 903, + 339 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "oritize labeling trivial objects first.", + "bbox": [ + 511, + 506, + 743, + 521 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To overcome these limitations, we focus on an unsupervised multi-label classification problem leveraging a vision-language model. One such model is CLIP [31], trained on an extensive dataset of 400M {image, text} pairs from the web. As a result, this model excels at recognizing diverse visual attributes and object categories in a zero-shot manner. Most recently, CDUL [2] applied CLIP for unsupervised multi-label classification by using it as a pseudolabeling function. CDUL generates pseudo-labels for both global and local views to help identify less prominent objects, training a multi-label classifier with each view-label pair. However, CDUL overlooks the limitations of CLIP, which hinders its prediction performance.", + "bbox": [ + 511, + 522, + 906, + 718 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we identified two reliability issues in CLIP, addressing them to fully harness CLIP's potential. Firstly, we demonstrate that CLIP's predictions are highly view-dependent, considerably impacting classification performance. In a toy experiment, we observe that changing the crop position and size of an image leads to notable changes in CLIP's predictions. For example, in Fig. 1, when the cropped area is slightly shifted or increased, the probabilities for the horse and person classes vary significantly. Based on these key observations, we use multiple local views near the target object to mitigate the uncertainty of CLIP's pseudo-labels caused by view changes. Specifically,", + "bbox": [ + 511, + 719, + 908, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation.", + "bbox": [ + 236, + 0, + 807, + 18 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "Except for this watermark, it is identical to the accepted version;", + "bbox": [ + 323, + 16, + 722, + 30 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 292, + 31, + 754, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "4661", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/60088c0668175707771d18910ba604eaea5adee25790c5032fd813a2987f6cce.jpg", + "image_caption": [ + "T:0.35/C:0.29" + ], + "image_footnote": [], + "bbox": [ + 91, + 87, + 222, + 155 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/37478862c79ef123fbaf49e2455cd93a54a16254b1cb9d166c4bfb19872fdb83.jpg", + "image_caption": [ + "P:0.50/C:0.21" + ], + "image_footnote": [], + "bbox": [ + 222, + 88, + 352, + 155 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/1d11a6b5723afdc8abad5b59af6e729df59af67e13f64a139337163f67269144.jpg", + "image_caption": [ + "H:1.00" + ], + "image_footnote": [], + "bbox": [ + 354, + 88, + 482, + 155 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/9a42edc92de4dd5a8a7bfbc895adec65846fe2bb2607fa46e6d8eca6a855ffb7.jpg", + "image_caption": [ + "(a) tv monitor (T)", + "(b) person (P)", + "(c) horse (H)", + "(d) CLIP probability distribution at VOC 2012", + "Figure 2. Sample images showcasing CLIP bias. C indicates the probability of chair. (a) Top 1 probability of \"tv monitor\" image is $35\\%$ . (b) Top 1 probability of a \"person\" image is $50\\%$ . (c) Top 1 probability of a \"horse\" image is $100\\%$ . (d) The mean class-wise probability of PASCAL VOC 2012. We can observe the class-wise prediction bias of CLIP from these results." + ], + "image_footnote": [], + "bbox": [ + 94, + 176, + 473, + 238 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "we acquire various local views extracted by the bounding box using the Class Activation Mapping (CAM) [49] method, where the CAM is computed from the trained classifier itself. Additionally, our method can draw more local views for hard samples whose prediction exceeds a certain threshold across multiple classes.", + "bbox": [ + 89, + 383, + 482, + 474 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Another notable issue with CLIP is its inherent prediction bias, which tends to reduce the predicted probability of specific classes. Due to polysemy [31], the words representing each class can be ambiguous in text embedding space. As illustrated in Fig. 2 (a) and (b), despite the presence of a single class (tv monitor and person, respectively), CLIP assigns considerably low top-1 probabilities of around $35\\%$ and $50\\%$ . Conversely, objects like horses exhibit notably higher probabilities, as depicted in Fig. 2 (c). Analyzing the average CLIP top-1 probabilities across classes, as shown in Fig. 2 (d), further highlights this bias, indicating a lower prediction for certain classes compared to others. The biased prediction is evident across various datasets such as MS COCO and NUSWIDE1. Since the proposed method relies on pseudo-labels generated by CLIP, its bias naturally transfers to the classifier. Consequently, the classifier shows a similar prediction bias and performance degradation. To mitigate this issue, we adjust all pseudo-labels by applying the inverse transform of the bias. While this adjustment can amplify unwanted noise in pseudo-labels, we apply consistency loss to handle this noise after debiasing. Our adjustment helps debiasing the classifier, thereby enhancing its generalization performance. As a result, our proposed Classifier-guided CLIP Distillation (CCD) achieves state-of-the-art performances on the PASCAL VOC 2012, PASCAL VOC 2007, COCO, and NUSWIDE benchmarks.", + "bbox": [ + 91, + 474, + 482, + 864 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 89, + 653, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Label efficient multi-label classification.", + "text_level": 1, + "bbox": [ + 511, + 114, + 854, + 130 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Unlike multi-class classification, where a single class is identified in an image, multi-label classification requires accurately identifying all classes present in an image. Labeling every class in an image is significantly more challenging and resource-intensive than identifying just one class. To address this issue, a partial label setting has been proposed [5, 10, 14, 20, 24, 30], which utilizes only a subset of labels associated with each image.", + "bbox": [ + 511, + 137, + 905, + 258 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Early approaches assumed any unannotated labels (i.e. \"unknown labels\") as negative labels [4, 37, 45]. These methods involved correcting these labels during the learning process. However, approaches that relied on probabilistic models [9, 16, 41, 46] for label correction were dependent on optimization techniques, resulting in limited scalability. Subsequently, alternative methods have been proposed, such as leveraging label-to-label similarity [14], cross-image semantic correlation [5], or category-specific representation [30] to refine labels. Some approaches suggest using soft-pseudo-labels instead of unknown labels [10, 26, 38].", + "bbox": [ + 511, + 258, + 905, + 440 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Research has extended to a more extreme scenario: single positive label setting [7, 19, 42, 51], where only a single positive label is provided per image. These methods build upon those proposed in the partial label setting and adapt them to function effectively in more challenging contexts.", + "bbox": [ + 511, + 441, + 905, + 516 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "More recently, CDUL [2] has addressed the unsupervised multi-label classification problem by utilizing CLIP. In a similar vein, our method tackles the unsupervised multi-label classification problem using CLIP while additionally providing local area information with CAM and addressing bias in CLIP predictions.", + "bbox": [ + 511, + 517, + 905, + 608 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Vision-language pretraining.", + "text_level": 1, + "bbox": [ + 511, + 619, + 769, + 636 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Vision-language pretraining (VLP) has emerged as a pivotal approach for bridging the gap between visual and textual modalities [15, 28, 34]. Among the pioneering VLPs, CLIP [31] stands out as the most commonly utilized framework. CLIP was trained on large {image, text} pairs sourced from the Internet. Recently, several studies have adopted VLP to provide rich visual-semantic knowledge for downstream image tasks, such as detection [12, 48], segmentation [18, 32, 50], human-object interation [17, 22], and generation [8, 39, 44].", + "bbox": [ + 511, + 642, + 905, + 794 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In multi-label classification, various methods have emerged to derive pseudo-labels from CLIP predictions [2, 13, 47]. Among these, CDUL [2] integrates local inferences with global inference, successfully expanding the utilization of the VLP to multi-label classification. However, CDUL fails to tackle two limitations inherent in CLIP: sensitivity to input view selection and inherent prediction bias.", + "bbox": [ + 511, + 795, + 905, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "1CLIP probability distribution of other datasets are in the supplementary material.", + "bbox": [ + 89, + 875, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "4662", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/c3877900e645a0a43a464b410debe40e899b33c285104aee0fbbedf6ab0a7b39.jpg", + "image_caption": [ + "87.6%", + "(a) Around GT box" + ], + "image_footnote": [], + "bbox": [ + 93, + 88, + 189, + 160 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/06361d85478502a965ba0baff7c63cc13ee8f4c4ae428e933ddc4c89de38f378.jpg", + "image_caption": [ + "87.1%", + "(b) GT box" + ], + "image_footnote": [], + "bbox": [ + 191, + 88, + 287, + 160 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/1ea59459653e8aad6e275be5e67fdc3ad76b7dd141c77be1c08da820d070d3bd.jpg", + "image_caption": [ + "87.1%", + "(c) Random box", + "Figure 3. The proof-of-concept study of local view selection method. We train the classifier with pseudo-label generated from four different local views: (a) Around GT boxes, (b) GT boxes, (c) Random boxes, (d) Uniform grid boxes. The numbers below each sample are the performance (mAP) of the classifier trained with the corresponding pseudo-label for the entire training set. The classifier trained with local view around GT boxes achieved $1.3\\%$ higher performance compared to classifier trained with uniform-grid local view." + ], + "image_footnote": [], + "bbox": [ + 289, + 88, + 383, + 161 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/0071ec44c4ecff1e6e0a54d1b0d10480d757f919bba0a8102ced51a68d7d2769.jpg", + "image_caption": [ + "86.4%", + "(d) Uniform grid box" + ], + "image_footnote": [], + "bbox": [ + 383, + 88, + 482, + 161 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Consequently, the pseudo-labels generated by CDUL exhibit noise and bias. In contrast, our approach leverages a fine-tuned classifier to suggest local views and addresses CLIP bias through calibration.", + "bbox": [ + 89, + 359, + 483, + 421 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 436, + 181, + 452 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Generating initial pseudo-labels", + "text_level": 1, + "bbox": [ + 89, + 462, + 370, + 479 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "CLIP, a powerful vision-language model, has been trained to align images with corresponding textual descriptions. Leveraging the powerful zero-shot capability of CLIP, we initially generate pseudo-labels similar to CDUL. For each image in the dataset, we extract its image embedding using the CLIP image encoder. Simultaneously, we obtain the text embedding by inputting the class name associated with each dataset and a fixed prompt (e.g. \"a photo of the [class name]\") into the text encoder. Next, we compute the similarity between the image embedding and the class-specific text embedding. This similarity score is then translated to compute softmax probabilities. These probabilities serve as the initial pseudo-labels. Mathematically, the similarity score for each class can be expressed as:", + "bbox": [ + 89, + 484, + 483, + 696 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\ns _ {c} = \\frac {f ^ {\\top} w _ {c}}{\\| f \\| \\cdot \\| w _ {c} \\|}, 1 \\leq c \\leq C \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 183, + 710, + 482, + 744 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $f$ is a CLIP image embedding, $w_{c}$ is a class-specific CLIP text embedding and $c$ indicates a class index. With class-wise similarity, we can calculate a class-wise softmax probability $p_{c}$ of each image as follows:", + "bbox": [ + 89, + 752, + 483, + 814 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\np _ {c} = \\frac {\\exp \\left(s _ {c} / \\tau\\right)}{\\sum_ {c = 1} ^ {C} \\exp \\left(s _ {c} / \\tau\\right)}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 202, + 825, + 482, + 863 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\tau$ is a CLIP temperature parameter. We concatenate each class-wise softmax probability to generate initial", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "pseudo-labels. Finally, the initial pseudo-label for each image is defined as:", + "bbox": [ + 511, + 90, + 903, + 121 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nl ^ {\\text {i n i t i a l}} = \\left\\{p _ {1} ^ {\\text {g l o b a l}}, p _ {2} ^ {\\text {g l o b a l}}, \\dots , p _ {c} ^ {\\text {g l o b a l}} \\right\\}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 570, + 131, + 903, + 157 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Following the acquisition of softmax probabilities, the initial pseudo-labels are used for the warm-up training of the classifier. Since the softmax calculation inherently normalizes the probabilities to ensure a sum of 1, it does not fit the multi-label classification problem where multiple classes have the same high probability (e.g. 1). Nevertheless, after the classifier undergoes warm-up training, it can predict each class independently, leading to recognizing multiple classes present in an image, albeit with limited accuracy. To address this performance limitation, in the subsequent section, we introduce a method to construct {training data, pseudo-label} pairs tailored for the multi-label classification problem, leveraging a warm-up trained classifier.", + "bbox": [ + 511, + 159, + 906, + 354 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Updating classifier-guided pseudo-label", + "text_level": 1, + "bbox": [ + 511, + 363, + 852, + 378 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Prior research [2] has shown that CLIP's inference results on the entire images may be suboptimal for multi-label classification. CLIP tends to assign high probabilities primarily to prominent objects in the input image. As a result, it does not effectively distribute probabilities across all existing objects. CDUL addresses this limitation by dividing the image into uniform patches in a grid fashion. Despite its simplicity, employing a uniform patch-based approach for local view acquisition may lead to the inference of background batch, where the patch does not contain any class object. More importantly, treating both simple images (i.e. those containing a single object) and complex images (i.e. those containing multiple objects) equally with the same number of local views is clearly inefficient in optimizing the resource (i.e. CLIP inference).", + "bbox": [ + 511, + 383, + 906, + 609 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Proof-of-concept study. We perform a proof-of-concept study to evaluate the impact of using object-specific locations in local view proposals. Our hypothesis is that incorporating class-related object locations enhances performance. We extract local views in four different scenarios, using ground truth bounding boxes (i.e., the bounding box that tightly covers the target object): (1) random cropping around the bounding box, (2) using only the bounding box itself $^{2}$ , (3) random cropping without guidance from the bounding box, and (4) uniform grid-based cropping. Each scenario produces nine patches per image. Subsequently, we infer pseudo-labels for these patches and train with each $\\{\\text{patch}, \\text{pseudo-label}\\}$ pair. As shown in Fig. 3, our results indicate that the first scenario-(a) random cropping around the bounding box-achieves the best performance.", + "bbox": [ + 511, + 611, + 905, + 837 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our proof-of-concept study reveals two critical insights: (1) the local view generation policy significantly influences", + "bbox": [ + 511, + 838, + 903, + 868 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "2For images with fewer than nine boxes, we additionally infer their resized versions.", + "bbox": [ + 511, + 875, + 906, + 898 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "4663", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/bc0ad15f1147ebf3b38b7c660904aba36e8ff3ea63bff11ae9ad98cdaae849bf.jpg", + "image_caption": [ + "Figure 4. The overview of label preparation. (a) We calculate the cosine similarity between text embeddings and image embeddings. The softmax probability of these similarities is the CLIP prediction of each image. The top 1 probabilities are highlighted in green boxes. By class-wise averaging the top-1 probabilities, the CLIP bias is derived (yellow box). Pseudo-labels are then generated by debiasing CLIP predictions. (b) For label updating, classes above a threshold are selected from the classifier output, and local views corresponding to these classes are extracted. The process for acquiring local labels for each patch mirrors the initial label acquisition. The final pseudo-label is obtained through a weighted sum of initial pseudo-labels and local labels." + ], + "image_footnote": [], + "bbox": [ + 93, + 88, + 903, + 364 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "CLIP's predictions, and (2) framing the view around the target object greatly improves classification accuracy.", + "bbox": [ + 88, + 498, + 480, + 527 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "CAM-based view-selection. Motivated by the above two findings, we decided to approximate the first scenario in our approach. We apply Class Activation Mapping (CAM) with the trained classifier to identify vital regions within the image. This is called a CAM-based view selection method. This method generates {training image, pseudo-label} pairs by focusing on the object periphery, without relying on any ground-truth annotations.", + "bbox": [ + 88, + 529, + 482, + 648 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "By employing the trained classifier, CAM visualizes activations that most significantly influence class prediction. CAM calculation is specifically based on the weights assigned to each channel of the feature map, which influence the classification score. This can be represented by the following formula:", + "bbox": [ + 89, + 651, + 482, + 742 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nG _ {c} (x) = \\sum_ {i = 1} ^ {Q} \\mathbf {w} _ {i} ^ {c \\top} \\cdot g _ {i} (x), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 197, + 757, + 482, + 800 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $x$ is input image, $g$ is CNN network, $g(x)$ is the feature map of the CNN, $\\mathbf{w}^c$ is class-wise weight of classifier, and $Q$ indicates the number of channels in the feature map.", + "bbox": [ + 89, + 809, + 482, + 854 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We first normalize the activation map to a range between 0 and 1, then apply a threshold to identify highly activated regions. Next, we generate candidate bounding boxes with", + "bbox": [ + 89, + 854, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "slight perturbation offsets. These bounding boxes define local views of the image. The regions highlighted by CAM act as reasonable approximations of the object's location, allowing us to extract local views without additional labeling costs. These localized view images are fed into CLIP to obtain softmax probabilities for the patches. The probabilities of local views are computed using the same method as for the global view, and then generating local labels that are focused on each local region.", + "bbox": [ + 511, + 498, + 906, + 636 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The number of local views. Additionally, we adjust the number of local views based on image complexity. For simpler images, the classifier typically provides confident, high-probability predictions. This translates to high logits for well-trained classes and near-zero for others. Conversely, complex images often produce uncertain predictions, characterized by higher logits distributed broadly across multiple classes. To account for this, we conduct additional local view sampling when the classifier output of a certain class surpasses a predefined threshold. Specifically, we crop the bounding box of any class exceeding this threshold value and obtain its local probability. Through our experiments, we empirically determined a fixed threshold that ensures appropriate average number of classes exceeding this threshold. Details on this ablation study can be found in the experiments section. This approach allocates fewer local views to simple images and more local views to", + "bbox": [ + 511, + 643, + 908, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4664", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "complex ones, optimizing the local view assignment.", + "bbox": [ + 89, + 90, + 439, + 104 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To obtain a single local pseudo-label for each input image, we select the maximum value of class-wise probability across all local patches from the corresponding image. Denoting the probability of $i$ -th patch for class $c$ as $p_{i,c}^{patch}$ , we conduct patch-wise aggregation using the maximum value as follows:", + "bbox": [ + 89, + 106, + 483, + 196 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\np _ {c} ^ {\\text {l o c a l}} = \\max _ {i = 1, \\dots , N} p _ {i, c} ^ {\\text {p a t c h}}, 1 \\leq c \\leq C \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 160, + 209, + 480, + 234 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $N$ represents the number of local patches corresponding to each image, dependent on the input image itself. We then aggregate the local class-wise probability for generating a local pseudo-label as follows:", + "bbox": [ + 89, + 239, + 483, + 299 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nl ^ {\\text {l o c a l}} = \\left\\{p _ {1} ^ {\\text {l o c a l}}, p _ {2} ^ {\\text {l o c a l}}, \\dots , p _ {c} ^ {\\text {l o c a l}} \\right\\}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 163, + 311, + 482, + 330 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We update the pseudo-label by the weighted sum of initial and local pseudo-labels. The final pseudo-label is calculated as follows:", + "bbox": [ + 89, + 335, + 482, + 378 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nl ^ {\\text {f i n a l}} = \\alpha l ^ {\\text {i n i t i a l}} + (1 - \\alpha) l ^ {\\text {l o c a l}}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 171, + 393, + 482, + 411 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\alpha$ is a hyperparameter and we used 0.4. Through this label update, the pseudo-label efficiently incorporates information about less salient objects in multi-label classification. The updated pseudo-label is utilized throughout the remainder of the training process.", + "bbox": [ + 89, + 417, + 483, + 494 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Debiasing CLIP pseudo-label", + "text_level": 1, + "bbox": [ + 89, + 500, + 352, + 517 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Earlier, we aimed to identify highly activated areas within images through Class Activation Mapping and to infer additional label information focused on these areas. However, a challenge arises when using pseudo-labels generated by CLIP due to its inherent prediction bias. Consequently, classifiers trained using these pseudo-labels may inadvertently inherit this bias. Previous work [3] reveals that while classifiers are robust at overcoming random noise, they cannot rectify consistent biases associated with specific classes. Therefore, we need to overcome the inherent prediction bias of CLIP to enhance the performance.", + "bbox": [ + 89, + 523, + 482, + 688 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To understand this bias, given CLIP's focus on salient objects, we treat global CLIP prediction as a single-label classification task. In a single-label classification, the predicted probability distribution for each class aligns with the class-wise average of the highest probabilities from each prediction. Following this principle, we compute the average prediction probability for each class, defining it as CLIP's prediction bias. Specifically, we compute CLIP bias as follows: (1) Obtain initial pseudo-labels by running each image in the \"target training set\" through CLIP; (2) Record top-1 probabilities for samples below the entropy threshold to constrain the task to single-label classification; (3) Count each class occurrences only if it achieves the top 1 probability; (4)Normalize each class's softmax probability by its", + "bbox": [ + 89, + 689, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/16611be1bb536704fa59c68cc0e2ac6a7a41dbdcba9a6e10c325bb5a06f6ac4d.jpg", + "image_caption": [ + "Figure 5. The training process of our method. We train the classifier using cross-entropy targeting initial pseudo-label during the warm-up phase which is illustrated as the green line. After the classifier-guided label update, we train the classifier using the cross-entropy loss targeting updated pseudo-label and the cross-entropy loss between the logits of differently augmented inputs which are illustrated as the green line and the blue dashed line." + ], + "image_footnote": [], + "bbox": [ + 516, + 95, + 903, + 258 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "occurrence count, defining this normalized probability as CLIP bias. This process does not require any ground-truth annotation, solely relying on the result of CLIP prediction.", + "bbox": [ + 511, + 393, + 903, + 439 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The bias obtained in this manner serves as a basis for calibrating CLIP-derived probabilities. Specifically, we calibrate by multiplying each class's softmax probability from CLIP by the inverse of the CLIP bias. This calibration process enables the classifier to learn a rectified probability distribution. Our debiasing technique leads to a noticeable improvement in the performance of classes previously challenged due to the inherent CLIP bias (see Tab. 3). Furthermore, we adjust excessively small probabilities to prevent the minimum probability value from approaching zero, aiding in learning the corrected distribution. In doing so, we guide the classifier as a whole to learn from this inverse distribution. This calibration is applied not only to the global probability but also to local inference, minimizing the influence of the CLIP bias on the classifier as much as possible.", + "bbox": [ + 511, + 441, + 906, + 667 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Overall training", + "text_level": 1, + "bbox": [ + 511, + 680, + 674, + 696 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Consistency loss. The proposed method tackles noisy data by utilizing pseudo-labels as target labels for all {training data, pseudo-label} pairs. Since pseudo-labels inherently carry noise and our debiasing process can potentially amplify this noise, we employ consistency loss between the data and its augmented data as additional critics. This approach is inspired by the common practice in a semi-supervised regime, improving feature extractors for handling noisy data. In our framework, we introduce consistency loss alongside the conventional binary cross-entropy loss during training. Consistency loss involves applying different augmentation intensities to the same image. Also, it passes each augmented version through the classifier to en", + "bbox": [ + 511, + 704, + 906, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "4665", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "sure consistent predictions. In Fig. 5, we employ weak augmentations, such as resizing and flipping, and strong augmentations, which comprise weak augmentation and additional color jittering.", + "bbox": [ + 89, + 90, + 480, + 151 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Training process. Initially, the entire training dataset undergoes inference through CLIP, followed by calculating the CLIP bias. Subsequently, we derive a global label by debiasing the global probability, utilizing it as an initial pseudolabel in Fig. 4 (a). The classifier is then warmed up with this pseudo-label as the target (the green line in Fig. 5). Following warm-up training, we update the pseudo-label with classifier-guided local inference. During this phase, the obtained local probabilities undergoes debiasing. Then they are combined to derive a local label, which then updates the existing pseudo-label as depicted in Fig. 4 (b). The classifier is subsequently trained with the updated pseudo-label for the remaining epochs (the green line and the blue dashed line in Fig. 5). For whole trainig, our loss function combines two terms:", + "bbox": [ + 89, + 152, + 482, + 380 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {t o t a l}} = L _ {c e} + \\beta L _ {\\text {c o n s i s t}}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 197, + 398, + 482, + 416 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_{ce}$ is cross-entropy loss and $\\mathcal{L}_{\\text{consist}}$ is consistency loss. $\\beta$ is a hyperparameter and we fix it as 0 during warm-up phase and 1 after.", + "bbox": [ + 89, + 425, + 482, + 470 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 488, + 223, + 505 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experimental setup", + "text_level": 1, + "bbox": [ + 89, + 513, + 277, + 530 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Dataset & evaluation Metric. We conducted evaluations on four distinct datasets. The PASCAL VOC 2012 [11] dataset comprises 20 classes, with 5,717 images in the training set and 5,823 in the validation set. Similarly, the PASCAL VOC 2007 dataset also encompasses 20 classes, with a combined total of 5,011 training and validation images and 4,952 test images. MS COCO [23] consists of 80 classes, encompassing 82,081 training images and 40,137 validation images. Additionally, the NUSWIDE [6] dataset comprises 81 classes, with a training set of 150,000 images and a test set of 60,200 images. In cases where a separate test set was not available, the validation set was utilized for testing purposes. Mean Average Precision (mAP) was employed as the performance metric, with ground truth labels utilized solely for performance evaluation and not incorporated during the model training phase.", + "bbox": [ + 89, + 537, + 482, + 777 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation detail. For a fair comparison, we endeavored to set the hyperparameters for CLIP inference as closely as possible to those of the primary competitive method, CDUL. We acquired labels using the frozen CLIP ResNet50×64, and employed the ImageNet pre-trained ResNet101 as the backbone network for our classifier. The entropy threshold for calculating CLIP bias and the classifier threshold for local view inference is both 0.5. It Our", + "bbox": [ + 89, + 780, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "warm-up classifier is trained for two epochs. While the entire training process can proceed up to 10 epochs, we obtain our final model via early stopping to prevent overfitting. Specifically, we terminate training based on the gradient of mean Average Precision (mAP) of the training labels throughout the learning process. Training was ceased when the gradient of mAP reached the first local minimum. We utilized a learning rate of $10^{-5}$ across all datasets and employed a batch size of 16. Undisclosed hyperparameters of CDUL were manually reproduced. For generating the initial pseudo-labels of global views, we maintained a fixed image resolution of $640 \\times 640$ . The CLIP prompts consistently used the phrase \"a photo of the [class name]\" except for the NUSWIDE dataset. (For a fair comparison with the existing method, CDUL, we employ the phrase \"a photo of a [class name]\" to match the pseudo-label performance.) For acquiring the initial pseudo-labels of local views, we introduced an offset value of 80 pixels for each vertexes when generating bounding boxes (i.e. a thresholded CAM-bounding box $\\pm 80$ as our local views). Additionally, cropped images were resized to have a longer side of 640 pixels while preserving their aspect ratios to obtain CLIP probabilities.", + "bbox": [ + 511, + 90, + 906, + 439 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Comparison with state-of-the-art methods", + "text_level": 1, + "bbox": [ + 511, + 452, + 872, + 468 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Competitive methods. Multi-label classification can be partitioned depending on the type of supervision: fully supervised, weakly supervised, and unsupervised approaches. Firstly, in a fully supervised setting, we choose (1) the Naïve method using only binary cross-entropy (BCE) loss and (2) the Naïve method with additional label smoothing (LS) [7]. Secondly, we compared our method with recent weakly supervised methods, including partial labeled methods utilizing only $10\\%$ of total labels such as (3) SARB [30], (4) ASL [33], (5) Chen et al. [5]. For single positive label methods, we choose (6) LL-R [19], (7) $\\mathrm{G}^2\\mathrm{NetPL}$ [1]. Finally, in an unsupervised scenario, we obtained initial labels using CLIP and evaluated the performance of (8) Naïve AN [21], (9) Szegedy et al. [38], (10) Aodha et al. [26], (11) Durand et al. [10], (12) ROLE [7], which leveraging CLIP generated initial labels. We conducted a comparative analysis with (13) CDUL [2], the current state-of-the-art in an unsupervised multi-label classification task.", + "bbox": [ + 511, + 476, + 906, + 762 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Comparison with fully & weakly methods. Compared to fully supervised methods, our Classifier Guided CLIP Distillation (CCD) demonstrates comparable performance on the PASCAL VOC 2012 and PASCAL VOC 2007 datasets. In particular, achieving similar performance without any human annotation is impressive. CCD outperforms various weakly supervised methods on the PASCAL VOC 2012 and PASCAL VOC 2007 datasets, although our performances on MS COCO and NUSWIDE are relatively lower. This", + "bbox": [ + 511, + 763, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "4666", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/413bed70b3ee40997e3dd510419a5517f6f81738d8f16621d3dad4e475d888c1.jpg", + "table_caption": [ + "Table 1. Mean average precision mAP in (%) for different multi-label classification methods under different supervision levels. Bold represents the best results among unsupervised methods. The proposed achieved state-of-the-art performance." + ], + "table_footnote": [], + "table_body": "
Supervision levelAnnotationMethodVOC12VOC07COCONUS
Fully supervisedFully labeledBCE90.191.378.550.7
BCE-LS [7]91.692.679.451.7
Weakly supervisedPartial labeled (10%)SARB [30]-85.772.5-
ASL [33]-82.969.7-
Chen et al. [5]-81.568.1-
Single positive labeledLL-R [19]89.790.672.647.4
G2NetPL [1]89.589.972.548.5
UnsupervisedAnnotation freeNaive AN [21]85.586.565.140.8
Szegedy et al. [38]86.887.965.541.3
Aodha et al. [26]84.286.263.940.1
Durand et al. [10]81.383.163.239.4
ROLE [7]82.684.667.143.2
CDUL [2]88.689.069.244.0
CCD (Ours)90.191.070.344.5
", + "bbox": [ + 91, + 127, + 906, + 343 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/90fbede7364eff033f0803ced1ced44738e77ed38fc5c8afc02a59e7cd851bdd.jpg", + "table_caption": [ + "Table 2. Ablation study of the proposed moduels. Mean average precision mAP in $(\\%)$ for different configuration uon the PASCAL VOC 2012 dataset. The best score is in bold." + ], + "table_footnote": [], + "table_body": "
Label updateCLIP debiasConsistencymAP (VOC12)
×××86.4
✓××88.7
✓✓×89.4
✓×✓88.8
✓✓✓90.1
", + "bbox": [ + 91, + 419, + 486, + 536 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "suggests that CCD is more effective on datasets with a smaller number of classes, such as PASCAL VOC 2012 and 2007, which consist of 20 classes. On the other hand, the performance gain is relatively small for complex datasets with many classes. It can be attributed to the simplistic prompt, \"a photo of the [class name],\" which might not accurately represent the target classes and co-occurrence issue, which bring out noisy local labels. The performance of initial labels of the training sets for PASCAL VOC 2012, MS COCO, and NUSWIDE is $85.3\\%$ , $65.4\\%$ , and $41.2\\%$ , respectively. CCD consistently shows about a $3.4\\%$ performance gain over all datasets.", + "bbox": [ + 88, + 565, + 482, + 747 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparison with unsupervised methods. Compared to unsupervised methods, CCD exhibits superior performance across all datasets. Particularly noteworthy are the significant improvements of $1.5\\% \\mathrm{p}$ , $2\\% \\mathrm{p}$ , and $1.1\\% \\mathrm{p}$ on PASCAL VOC 2012, PASCAL VOC 2007, and MS COCO, respectively, compared to the existing state-of-the-art methods. These results highlight the effectiveness of our method. However, in the case of the NUSWIDE dataset, the performance improvement is not as substantial. This can be attributed to the unique characteristics of the NUSWIDE", + "bbox": [ + 88, + 750, + 482, + 901 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c2a15663f866cca08d7ad4bb59d99283bb7cc3c225cca6d2f9d525a20cedc33b.jpg", + "image_caption": [ + "(a) The number of mean inferences", + "Figure 6. Classification results (mAP) on PASCAL VOC 2012 depending on (a) the change of the number of mean inferences (b) the change of ratio of pseudo-label updating." + ], + "image_footnote": [], + "bbox": [ + 517, + 366, + 712, + 446 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9a2493902e34083513eee6b395d270e7380f1bec5d0e1c419236a3f6ac82d349.jpg", + "image_caption": [ + "(b) Initial & Local weight $(\\alpha)$" + ], + "image_footnote": [], + "bbox": [ + 715, + 368, + 900, + 446 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "dataset, where $20\\%$ of the train/test images do not contain any class-related objects (i.e. \"none images\"). As a result, the initial pseudo-label acquisition process fails to form a high probability for specific classes in these images. Consequently, during local inference, a significant portion of inferences is assigned to \"none images,\" thereby reducing the quality of pseudo-labels of CCD.", + "bbox": [ + 511, + 551, + 906, + 657 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation", + "text_level": 1, + "bbox": [ + 511, + 667, + 617, + 681 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effects of each module. We evaluate the effectiveness of each component of our method in Tab. 2. The performance of the baseline model, the classifier trained with CLIP-generated global pseudo-labels, is $86.4\\%$ . When we update the pseudo-label with classifier-guided local inference, it shows a significant improvement of $2.3\\%$ p. This shows that our CAM-based local view selection method and adaptively allocating the number of CLIP inferences are effective. It already exceeds the performance of previous state-of-the-art methods. Debiasing CLIP also contributed $0.7\\%$ p, indicating that our proposed debiasing method effectively relieves the CLIP bias. The consistency loss alone adds only $0.1\\%$ p performance gain. However, when CLIP debiasing and consistency loss are combined, a substantial improvement of", + "bbox": [ + 511, + 688, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "4667", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/cbef149517b9033918d7e806b5a7d3c907015388817322b03726a709c1fd5b9a.jpg", + "table_caption": [ + "Table 3. Per-class AP in $(\\%)$ for without and with debiasing. We focus on the seven classes from the lowest mean probability. The result showcases that our debiasing is effective for enhancing the performance of biased classes." + ], + "table_footnote": [], + "table_body": "
Methodbottlechairtablepersonplantsofatv
w/o debias77.077.172.787.368.873.089.5
W debias77.679.074.889.272.176.692.0
", + "bbox": [ + 91, + 156, + 480, + 195 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "$1.4\\%$ p is achieved. This result implies that pseudo-labels generated by compensating for CLIP bias may amplify both accurate information and noise. Since the consistency loss is effective for dealing with noisy data, it creates a synergy with CLIP debiasing.", + "bbox": [ + 94, + 227, + 480, + 301 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The number of inference. We examine how model performance varies with different numbers of CLIP local inferences. We incrementally increase the average number of inferences to approximately 4, 6, 10, and 14. In Fig. 6 (a), the model achieves its best performance when the average number of inferences is around 10. Notably, increasing the number of inferences does not consistently enhance performance. Fewer inferences limit the allocation of local views to challenging samples, which hinders the ability to extract sufficient information from complex images, thereby negatively impacting negatively. In contrast, an excessive number of inferences results in the overallocation of local views, even for easy samples. It potentially introduces redundant inference for easy samples, making noisy pseudo-labels. In particular, when the average number of inferences is set to 10 or 14, the maximum number of local inferences remains capped at 20 (i.e. the total number of classes). It suggests that the inferences allocated to complex images have already reached saturation.", + "bbox": [ + 93, + 305, + 480, + 592 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Initial & local weight. We investigate the impact of adjusting the ratio $\\alpha$ between initial and local labels when updating the pseudo-labels (Eq. 7). In Fig. 6 (b), the results indicate that utilizing both global and local labels together leads to performance enhancement. This trend arises because two different label sources have complementary properties. The local label contains more information through multiple CLIP inferences of different local patches. However, since the reliability of CAM is not $100\\%$ , it may draw an invalid patch. Conversely, the initial label contains less information, but it is always generated from valid images. By blending these two sources, we reinforce consistent predictions between local and initial labels, thereby strengthening the signal of reliable labels. Additionally, we highlight that the use of our local labels alone (i.e. at weight 0) can outperforms CDUL. This confirms that our choice of local views and their labels extracts more informative learning signals than the mixture of the global view and uniformly sampled local views.", + "bbox": [ + 93, + 595, + 480, + 880 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Per-class mAP of biased classes. We assess the effective", + "bbox": [ + 94, + 886, + 480, + 898 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ness of our proposed method by examining the improvement in performance for specific target classes before and after CLIP debiasing. In Tab. 3, we focus on seven classes from the biased class, including bottle, chair, table, person, potted plant, sofa, and tv monitor as confirmed in the Fig. 2. Our results demonstrate consistent improvements in the performance of all these classes, ranging from $0.6\\% \\mathrm{p}$ (bottle) to $3.6\\% \\mathrm{p}$ (sofa). Moreover, we observe comparable or slightly enhanced performance for the remaining classes not shown in the table. The result affirms the effectiveness of our method across a broader spectrum. Detailed class-wise performance results for the entire dataset are provided in the supplementary material.", + "bbox": [ + 516, + 90, + 903, + 287 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Limitation", + "text_level": 1, + "bbox": [ + 516, + 297, + 630, + 310 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "A limitation of proposed method is that the generated local views often struggle to distinguish between co-occurring objects, an issue inherited from CAM. This challenge is especially evident in complex datasets like COCO, where similar classes—such as skies, snowboards, and skateboards—frequently appear in scene with human activity. These classes often co-occur with human feet, leading classifiers to associate high activations near the feet with these classes. This phenomenon makes them potential candidates for local views. This redundant information results in noisy pseudo-labels, reducing overall performance. Examples of such failure cases are provided in the supplementary materials. Developing an improved local view proposal method to better handle noisy inferences and focusing on the target objects will be an important direction for future work.", + "bbox": [ + 516, + 319, + 903, + 544 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 516, + 559, + 630, + 571 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we propose a novel unsupervised multi-label classification method, Classifier-guided CLIP Distillation (CCD). We address two critical issues in CLIP: inconsistent predictions due to subtle input changes and inherent prediction bias. To tackle these issues, we devised a classifier-guided label update and CLIP debiasing. Firstly, we identified the effectiveness of acquiring local views near class-related objects and proposed a classifier-guided label update. It selects local views utilizing the Class Activation Mapping (CAM) and acquires corresponding labels from CLIP. Our approach results in allocating more local views for complex images. Secondly, we address the CLIP's bias. We investigated the biased prediction distribution of CLIP and rectified the pseudo-labels. This step notably boosts performance in certain classes and reduces the impact of bias on CLIP. Through experiments on four benchmarks, CCD newly achieved state-of-the-art performances in unsupervised multi-label classification. Remarkably, CCD achieved comparable performance to fully supervised methods on PASCAL VOC 2012 and PASCAL VOC 2007 datasets without relying on annotations.", + "bbox": [ + 516, + 584, + 903, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "4668", + "bbox": [ + 483, + 946, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements. This research was supported by the Basic Science Research Program through the National Research Foundation of Korea (NRF) funded by the MSIP (RS-2025-00520207, RS-2023-00219019), KEIT grant funded by the Korean government (MOTIE) (No. 2022-0-00680, No. 2022-0-01045), the IITP grant funded by the Korean government (MSIT) (No. RS-2024-00457882, National AI Research Lab Project, No. 2021-0-02068 Artificial Intelligence Innovation Hub, RS-2019-II190075 Artificial Intelligence Graduate School Program (KAIST)), Samsung Electronics Co., Ltd (IO230508-06190-01) and SAMSUNG Research, Samsung Electronics Co., Ltd.", + "bbox": [ + 89, + 90, + 485, + 258 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 287, + 187, + 304 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Rabab Abdelfattah, Xin Zhang, Mostafa M Fouda, Xiaofeng Wang, and Song Wang. G2netpl: Generic game-theoretic network for partial-label image classification. arXiv preprint arXiv:2210.11469, 2022. 6, 7", + "[2] Rabab Abdelfattah, Qing Guo, Xiaoguang Li, Xiaofeng Wang, and Song Wang. Cdul: Clip-driven unsupervised learning for multi-label image classification. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1348-1357, 2023. 1, 2, 3, 6, 7", + "[3] Görkem Algan and Ülkay Ulusoy. Label noise types and their effects on deep learning, 2020. 5", + "[4] Serhat Selcuk Bucak, Rong Jin, and Anil K. Jain. Multi-label learning with incomplete class assignments. In CVPR 2011, pages 2801-2808, 2011. 2", + "[5] Tianshui Chen, Tao Pu, Hefeng Wu, Yuan Xie, and Liang Lin. Structured semantic transfer for multi-label recognition with partial labels. In Proceedings of the AAAI conference on artificial intelligence, pages 339-346, 2022. 1, 2, 6, 7", + "[6] Tat-Seng Chua, Jinhui Tang, Richang Hong, Haojie Li, Zhiping Luo, and Yantao Zheng. Nus-wide: a real-world web image database from national university ofSingapore. In Proceedings of the ACM international conference on image and video retrieval, pages 1-9, 2009. 6", + "[7] Elijah Cole, Oisin Mac Aodha, Titouan Lorieul, Pietro Perona, Dan Morris, and Nebojsa Jojic. Multi-label learning from single positive labels. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 933-942, 2021. 1, 2, 6, 7", + "[8] Katherine Crowson, Stella Biderman, Daniel Kornis, Dashiell Stander, Eric Hallahan, Louis Castricato, and Edward Raff. Vqgan-clip: Open domain image generation and editing with natural language guidance. In European Conference on Computer Vision, pages 88–105. Springer, 2022. 2", + "[9] Jia Deng, Olga Russakovsky, Jonathan Krause, Michael S Bernstein, Alex Berg, and Li Fei-Fei. Scalable multi-label annotation. In Proceedings of the SIGCHI Conference on Human Factors in Computing Systems, pages 3099–3102, 2014. 2", + "[10] Thibaut Durand, Nazanin Mehrasa, and Greg Mori. Learning a deep convnet for multi-label classification with partial" + ], + "bbox": [ + 93, + 313, + 482, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "labels. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 647-657, 2019. 1, 2, 6, 7", + "[11] M. Everingham, L. Van Gool, C. K. I. Williams, J. Winn, and A. Zisserman. The Pascal visual object classes (voc) challenge. International Journal of Computer Vision, 88(2): 303-338, 2010. 6", + "[12] Xiuye Gu, Tsung-Yi Lin, Weicheng Kuo, and Yin Cui. Open-vocabulary object detection via vision and language knowledge distillation. arXiv preprint arXiv:2104.13921, 2021. 2", + "[13] Ping Hu, Ximeng Sun, Stan Sclaroff, and Kate Saenko. Dualcoop++: Fast and effective adaptation to multi-label recognition with limited annotations. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 2", + "[14] Dat Huynh and Ehsan Elhamifar. Interactive multi-label cnn learning with partial labels. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9423-9432, 2020. 1, 2", + "[15] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International conference on machine learning, pages 4904-4916. PMLR, 2021. 2", + "[16] Ashish Kapoor, Raajay Viswanathan, and Prateek Jain. Multilabel classification using bayesian compressed sensing. Advances in neural information processing systems, 25, 2012. 2", + "[17] Apoorv Khandelwal, Luca Weihs, Roozbeh Mottaghi, and Aniruddha Kembhavi. Simple but effective: Clip embeddings for embodied ai. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14829-14838, 2022. 2", + "[18] Dongseob Kim, Seungho Lee, Junsuk Choe, and Hyunjung Shim. Weakly supervised semantic segmentation for driving scenes. arXiv preprint arXiv:2312.13646, 2023. 2", + "[19] Youngwook Kim, Jae Myung Kim, Zeynep Akata, and Jungwoo Lee. Large loss matters in weakly supervised multi-label classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14156-14165, 2022. 1, 2, 6, 7", + "[20] Youngwook Kim, Jae Myung Kim, Jieun Jeong, Cordelia Schmid, Zeynep Akata, and Jungwoo Lee. Bridging the gap between model explanations in partially annotated multi-label classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3408-3417, 2023. 1, 2", + "[21] Kaustav Kundu and Joseph Tighe. Exploiting weakly supervised visual patterns to learn from partial annotations. Advances in Neural Information Processing Systems, 33:561-572, 2020. 6, 7", + "[22] Yue Liao, Aixi Zhang, Miao Lu, Yongliang Wang, Xiaobo Li, and Si Liu. Gen-vlkt: Simplify association and enhance interaction understanding for hoi detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20123-20132, 2022. 2" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "4669", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[23] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 6", + "[24] Bing-Qing Liu, Bin-Bin Jia, and Min-Ling Zhang. Towards enabling binary decomposition for partial multi-label learning. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(11):13203-13217, 2023. 1, 2", + "[25] Weiwei Liu and Ivor Tsang. On the optimality of classifier chain for multi-label classification. Advances in Neural Information Processing Systems, 28, 2015. 1", + "[26] Oisin Mac Aodha, Elijah Cole, and Pietro Perona. Presence-only geographical priors for fine-grained image classification. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9596-9606, 2019. 2, 6, 7", + "[27] Dhruv Mahajan, Ross Girshick, Vignesh Ramanathan, Kaiming He, Manohar Paluri, Yixuan Li, Ashwin Bharambe, and Laurens Van Der Maaten. Exploring the limits of weakly supervised pretraining. In Proceedings of the European conference on computer vision (ECCV), pages 181-196, 2018. 1", + "[28] Antoine Miech, Jean-Baptiste Alayrac, Lucas Smaira, Ivan Laptev, Josef Sivic, and Andrew Zisserman. End-to-end learning of visual representations from uncurated instructional videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9879-9889, 2020. 2", + "[29] Ishan Misra, C Lawrence Zitnick, Margaret Mitchell, and Ross Girshick. Seeing through the human reporting bias: Visual classifiers from noisy human-centric labels. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2930–2939, 2016. 1", + "[30] Tao Pu, Tianshui Chen, Hefeng Wu, and Liang Lin. Semantic-aware representation blending for multi-label image recognition with partial labels. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2091-2098, 2022. 1, 2, 6, 7", + "[31] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 2", + "[32] Yongming Rao, Wenliang Zhao, Guangyi Chen, Yansong Tang, Zheng Zhu, Guan Huang, Jie Zhou, and Jiwen Lu. Denseclip: Language-guided dense prediction with context-aware prompting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18082-18091, 2022. 2", + "[33] Tal Ridnik, Emanuel Ben-Baruch, Nadav Zamir, Asaf Noy, Itamar Friedman, Matan Protter, and Lihi Zelnik-Manor. Asymmetric loss for multi-label classification. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 82-91, 2021. 6, 7" + ], + "bbox": [ + 91, + 90, + 480, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[34] Amanpreet Singh, Ronghang Hu, Vedanuj Goswami, Guillaume Couairon, Wojciech Galuba, Marcus Rohrbach, and Douwe Kiela. Flava: A foundational language and vision alignment model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15638-15650, 2022. 2", + "[35] GyeongRyeol Song, Noo-ri Kim, Jin-Seop Lee, and JeeHyong Lee. Ignore: Information gap-based false negative loss rejection for single positive multi-label learning. In European Conference on Computer Vision, pages 472–488. Springer, 2025. 1", + "[36] Chen Sun, Abhinav Shrivastava, Saurabh Singh, and Abhinav Gupta. Revisiting unreasonable effectiveness of data in deep learning era. In Proceedings of the IEEE international conference on computer vision, pages 843-852, 2017. 1", + "[37] Yu-Yin Sun, Yin Zhang, and Zhi-Hua Zhou. Multi-label learning with weak label. Proceedings of the AAAI Conference on Artificial Intelligence, 24(1):593-598, 2010. 2", + "[38] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2818-2826, 2016. 2, 6, 7", + "[39] Guy Tevet, Brian Gordon, Amir Hertz, Amit H Bermano, and Daniel Cohen-Or. Motionclip: Exposing human motion generation to clip space. In European Conference on Computer Vision, pages 358–374. Springer, 2022. 2", + "[40] Grigorios Tsoumakas and Ioannis Katakis. Multi-label classification: An overview. International Journal of Data Warehousing and Mining (IJDWM), 3(3):1-13, 2007. 1", + "[41] Deepak Vasisht, Andreas Damianou, Manik Varma, and Ashish Kapoor. Active learning for sparse bayesian multilabel classification. In Proceedings of the 20th ACM SIGKDD international conference on Knowledge discovery and data mining, pages 472-481, 2014. 2", + "[42] Thomas Verelst, Paul K. Rubenstein, Marcin Eichner, Tinne Tuytelaars, and Maxim Berman. Spatial consistency loss for training multi-label classifiers from single-label annotations. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 3879-3889, 2023. 1, 2", + "[43] Changhu Wang, Shuicheng Yan, Lei Zhang, and Hong-Jiang Zhang. Multi-label sparse coding for automatic image annotation. In 2009 IEEE conference on computer vision and pattern recognition, pages 1643-1650. IEEE, 2009. 1", + "[44] Can Wang, Menglei Chai, Mingming He, Dongdong Chen, and Jing Liao. Clip-nerf: Text-and-image driven manipulation of neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3835-3844, 2022. 2", + "[45] Qifan Wang, Bin Shen, Shumiao Wang, Liang Li, and Luo Si. Binary codes embedding for fast image tagging with incomplete labels. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part II 13, pages 425-439. Springer, 2014. 2", + "[46] Baoyuan Wu, Siwei Lyu, and Bernard Ghanem. Ml-mg: Multi-label learning with missing labels using a mixed" + ], + "bbox": [ + 516, + 90, + 903, + 901 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "4670", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "graph. In Proceedings of the IEEE international conference on computer vision, pages 4157-4165, 2015. 2", + "[47] Xin Xing, Zhexiao Xiong, Abby Stylianou, Srikumar Sastry, Liyu Gong, and Nathan Jacobs. Vision-language pseudolabels for single-positive multi-label learning. arXiv preprint arXiv:2310.15985, 2023. 2", + "[48] Alireza Zareian, Kevin Dela Rosa, Derek Hao Hu, and Shih-Fu Chang. Open-vocabulary object detection using captions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14393-14402, 2021. 2", + "[49] Bolei Zhou, Aditya Khosla, Agata Lapedriza, Aude Oliva, and Antonio Torralba. Learning deep features for discriminative localization. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2921-2929, 2016. 2", + "[50] Chong Zhou, Chen Change Loy, and Bo Dai. Extract free dense labels from clip. In European Conference on Computer Vision, pages 696-712. Springer, 2022. 2", + "[51] Donghao Zhou, Pengfei Chen, Qiong Wang, Guangyong Chen, and Pheng-Ann Heng. Acknowledging the unknown for multi-label learning with single positive labels. In European Conference on Computer Vision, pages 423-440. Springer, 2022. 1, 2" + ], + "bbox": [ + 91, + 90, + 483, + 430 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "4671", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Classifier-guided CLIP Distillation for Unsupervised Multi-label Classification/325c3ac3-6589-4c70-b2e1-2f7ac7877553_model.json b/2025/Classifier-guided CLIP Distillation for Unsupervised Multi-label Classification/325c3ac3-6589-4c70-b2e1-2f7ac7877553_model.json new file mode 100644 index 0000000000000000000000000000000000000000..419d0cbd429f0b8317d393f837c19dd308038a4a --- /dev/null +++ b/2025/Classifier-guided CLIP Distillation for Unsupervised Multi-label Classification/325c3ac3-6589-4c70-b2e1-2f7ac7877553_model.json @@ -0,0 +1,2378 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.02 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation." + }, + { + "type": "header", + "bbox": [ + 0.325, + 0.017, + 0.723, + 0.031 + ], + "angle": 0, + "content": "Except for this watermark, it is identical to the accepted version;" + }, + { + "type": "header", + "bbox": [ + 0.294, + 0.032, + 0.755, + 0.047 + ], + "angle": 0, + "content": "the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.105, + 0.131, + 0.895, + 0.154 + ], + "angle": 0, + "content": "Classifier-guided CLIP Distillation for Unsupervised Multi-label Classification" + }, + { + "type": "text", + "bbox": [ + 0.205, + 0.182, + 0.528, + 0.217 + ], + "angle": 0, + "content": "Dongseob Kim \nSamsung Electronics, Republic of Korea" + }, + { + "type": "text", + "bbox": [ + 0.28, + 0.22, + 0.453, + 0.234 + ], + "angle": 0, + "content": "kou.kim@samsung.com" + }, + { + "type": "text", + "bbox": [ + 0.579, + 0.182, + 0.793, + 0.217 + ], + "angle": 0, + "content": "Hyunjung Shim KAIST, Republic of Korea" + }, + { + "type": "text", + "bbox": [ + 0.593, + 0.22, + 0.774, + 0.233 + ], + "angle": 0, + "content": "kateshim@kaist.ac.kr" + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.269, + 0.327, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.301, + 0.485, + 0.558 + ], + "angle": 0, + "content": "Multi-label classification is crucial for comprehensive image understanding, yet acquiring accurate annotations is challenging and costly. To address this, a recent study suggests exploiting unsupervised multi-label classification leveraging CLIP, a powerful vision-language model. Despite CLIP's proficiency, it suffers from view-dependent predictions and inherent bias, limiting its effectiveness. We propose a novel method that addresses these issues by leveraging multiple views near target objects, guided by Class Activation Mapping (CAM) of the classifier, and debiasing pseudo-labels derived from CLIP predictions. Our Classifier-guided CLIP Distillation (CCD) enables selecting multiple local views without extra labels and debiasing predictions to enhance classification performance. Experimental results validate our method's superiority over existing techniques across diverse datasets. The code is available at https://github.com/k0u-id/CCD." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.588, + 0.222, + 0.604 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.613, + 0.484, + 0.734 + ], + "angle": 0, + "content": "Multi-label classification plays a pivotal role in identifying multiple classes within an image, serving as a foundational component in image understanding [25, 27, 29, 36, 40]. Its applications are widespread, including recommendation systems, surveillance, and image retrieval [43]. However, acquiring multi-label annotations poses challenges due to the difficulty and significant costs involved in meticulously annotating all relevant objects within an image." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.484, + 0.902 + ], + "angle": 0, + "content": "To solve this problem, various label-efficient techniques have been proposed. A partial label setting [5, 10, 14, 20, 24, 30] annotates only a portion of the classes appearing in an image and considers the rest as unknown labels. In the extreme, single positive label settings [7, 19, 35, 42, 51] provide only one positive label per image, with all other labels assumed unknown. However, these methods still incur labeling costs. Besides, current studies simulate partial and single positive settings by randomly omitting labels from a fully labeled dataset, which may not accurately reflect the actual labeling practices. For example, humans tend to pri" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.268, + 0.61, + 0.34 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.527, + 0.34, + 0.6, + 0.358 + ], + "angle": 0, + "content": "H. 0.99 / P. 0.00 (a) original image" + }, + { + "type": "image", + "bbox": [ + 0.615, + 0.268, + 0.708, + 0.34 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.624, + 0.341, + 0.696, + 0.358 + ], + "angle": 0, + "content": "H:0.96 /P:0.00 (b)1/4 patch" + }, + { + "type": "image", + "bbox": [ + 0.712, + 0.268, + 0.806, + 0.34 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.724, + 0.34, + 0.796, + 0.358 + ], + "angle": 0, + "content": "H:0.14/P:0.12 (c) 1/9" + }, + { + "type": "image", + "bbox": [ + 0.812, + 0.269, + 0.905, + 0.34 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.827, + 0.341, + 0.893, + 0.358 + ], + "angle": 0, + "content": ":0.05/P:0.65 (d) 1/16 patch" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.381, + 0.907, + 0.48 + ], + "angle": 0, + "content": "Figure 1. CLIP prediction probability of horse (H) and person (P) corresponding to different input patches from single image: (a) The original image and its CLIP prediction. (b) The \\(1/4\\) sized patch and its CLIP prediction. (c) The \\(1/9\\) sized patch and its CLIP prediction. (d) The \\(1/16\\) sized patch and its CLIP prediction. The same colored box indicates each patch cropped from the corresponding box." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.507, + 0.744, + 0.522 + ], + "angle": 0, + "content": "oritize labeling trivial objects first." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.523, + 0.907, + 0.719 + ], + "angle": 0, + "content": "To overcome these limitations, we focus on an unsupervised multi-label classification problem leveraging a vision-language model. One such model is CLIP [31], trained on an extensive dataset of 400M {image, text} pairs from the web. As a result, this model excels at recognizing diverse visual attributes and object categories in a zero-shot manner. Most recently, CDUL [2] applied CLIP for unsupervised multi-label classification by using it as a pseudolabeling function. CDUL generates pseudo-labels for both global and local views to help identify less prominent objects, training a multi-label classifier with each view-label pair. However, CDUL overlooks the limitations of CLIP, which hinders its prediction performance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.909, + 0.902 + ], + "angle": 0, + "content": "In this paper, we identified two reliability issues in CLIP, addressing them to fully harness CLIP's potential. Firstly, we demonstrate that CLIP's predictions are highly view-dependent, considerably impacting classification performance. In a toy experiment, we observe that changing the crop position and size of an image leads to notable changes in CLIP's predictions. For example, in Fig. 1, when the cropped area is slightly shifted or increased, the probabilities for the horse and person classes vary significantly. Based on these key observations, we use multiple local views near the target object to mitigate the uncertainty of CLIP's pseudo-labels caused by view changes. Specifically," + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "4661" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.088, + 0.223, + 0.156 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.123, + 0.156, + 0.193, + 0.163 + ], + "angle": 0, + "content": "T:0.35/C:0.29" + }, + { + "type": "image_caption", + "bbox": [ + 0.115, + 0.165, + 0.192, + 0.174 + ], + "angle": 0, + "content": "(a) tv monitor (T)" + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.089, + 0.353, + 0.156 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.253, + 0.156, + 0.323, + 0.163 + ], + "angle": 0, + "content": "P:0.50/C:0.21" + }, + { + "type": "image_caption", + "bbox": [ + 0.255, + 0.165, + 0.317, + 0.175 + ], + "angle": 0, + "content": "(b) person (P)" + }, + { + "type": "image", + "bbox": [ + 0.355, + 0.089, + 0.483, + 0.156 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.398, + 0.156, + 0.436, + 0.163 + ], + "angle": 0, + "content": "H:1.00" + }, + { + "type": "image_caption", + "bbox": [ + 0.391, + 0.165, + 0.448, + 0.175 + ], + "angle": 0, + "content": "(c) horse (H)" + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.177, + 0.475, + 0.239 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.187, + 0.239, + 0.388, + 0.249 + ], + "angle": 0, + "content": "(d) CLIP probability distribution at VOC 2012" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.274, + 0.483, + 0.358 + ], + "angle": 0, + "content": "Figure 2. Sample images showcasing CLIP bias. C indicates the probability of chair. (a) Top 1 probability of \"tv monitor\" image is \\(35\\%\\). (b) Top 1 probability of a \"person\" image is \\(50\\%\\). (c) Top 1 probability of a \"horse\" image is \\(100\\%\\). (d) The mean class-wise probability of PASCAL VOC 2012. We can observe the class-wise prediction bias of CLIP from these results." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.384, + 0.483, + 0.476 + ], + "angle": 0, + "content": "we acquire various local views extracted by the bounding box using the Class Activation Mapping (CAM) [49] method, where the CAM is computed from the trained classifier itself. Additionally, our method can draw more local views for hard samples whose prediction exceeds a certain threshold across multiple classes." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.476, + 0.483, + 0.866 + ], + "angle": 0, + "content": "Another notable issue with CLIP is its inherent prediction bias, which tends to reduce the predicted probability of specific classes. Due to polysemy [31], the words representing each class can be ambiguous in text embedding space. As illustrated in Fig. 2 (a) and (b), despite the presence of a single class (tv monitor and person, respectively), CLIP assigns considerably low top-1 probabilities of around \\(35\\%\\) and \\(50\\%\\). Conversely, objects like horses exhibit notably higher probabilities, as depicted in Fig. 2 (c). Analyzing the average CLIP top-1 probabilities across classes, as shown in Fig. 2 (d), further highlights this bias, indicating a lower prediction for certain classes compared to others. The biased prediction is evident across various datasets such as MS COCO and NUSWIDE1. Since the proposed method relies on pseudo-labels generated by CLIP, its bias naturally transfers to the classifier. Consequently, the classifier shows a similar prediction bias and performance degradation. To mitigate this issue, we adjust all pseudo-labels by applying the inverse transform of the bias. While this adjustment can amplify unwanted noise in pseudo-labels, we apply consistency loss to handle this noise after debiasing. Our adjustment helps debiasing the classifier, thereby enhancing its generalization performance. As a result, our proposed Classifier-guided CLIP Distillation (CCD) achieves state-of-the-art performances on the PASCAL VOC 2012, PASCAL VOC 2007, COCO, and NUSWIDE benchmarks." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.09, + 0.655, + 0.105 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.115, + 0.855, + 0.131 + ], + "angle": 0, + "content": "2.1. Label efficient multi-label classification." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.138, + 0.906, + 0.259 + ], + "angle": 0, + "content": "Unlike multi-class classification, where a single class is identified in an image, multi-label classification requires accurately identifying all classes present in an image. Labeling every class in an image is significantly more challenging and resource-intensive than identifying just one class. To address this issue, a partial label setting has been proposed [5, 10, 14, 20, 24, 30], which utilizes only a subset of labels associated with each image." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.26, + 0.906, + 0.441 + ], + "angle": 0, + "content": "Early approaches assumed any unannotated labels (i.e. \"unknown labels\") as negative labels [4, 37, 45]. These methods involved correcting these labels during the learning process. However, approaches that relied on probabilistic models [9, 16, 41, 46] for label correction were dependent on optimization techniques, resulting in limited scalability. Subsequently, alternative methods have been proposed, such as leveraging label-to-label similarity [14], cross-image semantic correlation [5], or category-specific representation [30] to refine labels. Some approaches suggest using soft-pseudo-labels instead of unknown labels [10, 26, 38]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.442, + 0.906, + 0.517 + ], + "angle": 0, + "content": "Research has extended to a more extreme scenario: single positive label setting [7, 19, 42, 51], where only a single positive label is provided per image. These methods build upon those proposed in the partial label setting and adapt them to function effectively in more challenging contexts." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.518, + 0.906, + 0.609 + ], + "angle": 0, + "content": "More recently, CDUL [2] has addressed the unsupervised multi-label classification problem by utilizing CLIP. In a similar vein, our method tackles the unsupervised multi-label classification problem using CLIP while additionally providing local area information with CAM and addressing bias in CLIP predictions." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.621, + 0.771, + 0.637 + ], + "angle": 0, + "content": "2.2. Vision-language pretraining." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.643, + 0.906, + 0.795 + ], + "angle": 0, + "content": "Vision-language pretraining (VLP) has emerged as a pivotal approach for bridging the gap between visual and textual modalities [15, 28, 34]. Among the pioneering VLPs, CLIP [31] stands out as the most commonly utilized framework. CLIP was trained on large {image, text} pairs sourced from the Internet. Recently, several studies have adopted VLP to provide rich visual-semantic knowledge for downstream image tasks, such as detection [12, 48], segmentation [18, 32, 50], human-object interation [17, 22], and generation [8, 39, 44]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.906, + 0.901 + ], + "angle": 0, + "content": "In multi-label classification, various methods have emerged to derive pseudo-labels from CLIP predictions [2, 13, 47]. Among these, CDUL [2] integrates local inferences with global inference, successfully expanding the utilization of the VLP to multi-label classification. However, CDUL fails to tackle two limitations inherent in CLIP: sensitivity to input view selection and inherent prediction bias." + }, + { + "type": "page_footnote", + "bbox": [ + 0.091, + 0.875, + 0.483, + 0.901 + ], + "angle": 0, + "content": "1CLIP probability distribution of other datasets are in the supplementary material." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4662" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.089, + 0.191, + 0.161 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.128, + 0.163, + 0.156, + 0.171 + ], + "angle": 0, + "content": "87.6%" + }, + { + "type": "image_caption", + "bbox": [ + 0.102, + 0.174, + 0.18, + 0.182 + ], + "angle": 0, + "content": "(a) Around GT box" + }, + { + "type": "image", + "bbox": [ + 0.192, + 0.089, + 0.288, + 0.161 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.225, + 0.163, + 0.254, + 0.171 + ], + "angle": 0, + "content": "87.1%" + }, + { + "type": "image_caption", + "bbox": [ + 0.217, + 0.174, + 0.262, + 0.182 + ], + "angle": 0, + "content": "(b) GT box" + }, + { + "type": "image", + "bbox": [ + 0.29, + 0.089, + 0.384, + 0.162 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.324, + 0.164, + 0.351, + 0.171 + ], + "angle": 0, + "content": "87.1%" + }, + { + "type": "image_caption", + "bbox": [ + 0.303, + 0.174, + 0.37, + 0.182 + ], + "angle": 0, + "content": "(c) Random box" + }, + { + "type": "image", + "bbox": [ + 0.385, + 0.089, + 0.483, + 0.162 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.42, + 0.164, + 0.448, + 0.171 + ], + "angle": 0, + "content": "86.4%" + }, + { + "type": "image_caption", + "bbox": [ + 0.392, + 0.174, + 0.477, + 0.183 + ], + "angle": 0, + "content": "(d) Uniform grid box" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.208, + 0.485, + 0.333 + ], + "angle": 0, + "content": "Figure 3. The proof-of-concept study of local view selection method. We train the classifier with pseudo-label generated from four different local views: (a) Around GT boxes, (b) GT boxes, (c) Random boxes, (d) Uniform grid boxes. The numbers below each sample are the performance (mAP) of the classifier trained with the corresponding pseudo-label for the entire training set. The classifier trained with local view around GT boxes achieved \\(1.3\\%\\) higher performance compared to classifier trained with uniform-grid local view." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.361, + 0.484, + 0.422 + ], + "angle": 0, + "content": "Consequently, the pseudo-labels generated by CDUL exhibit noise and bias. In contrast, our approach leverages a fine-tuned classifier to suggest local views and addresses CLIP bias through calibration." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.438, + 0.182, + 0.453 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.463, + 0.372, + 0.48 + ], + "angle": 0, + "content": "3.1. Generating initial pseudo-labels" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.486, + 0.484, + 0.698 + ], + "angle": 0, + "content": "CLIP, a powerful vision-language model, has been trained to align images with corresponding textual descriptions. Leveraging the powerful zero-shot capability of CLIP, we initially generate pseudo-labels similar to CDUL. For each image in the dataset, we extract its image embedding using the CLIP image encoder. Simultaneously, we obtain the text embedding by inputting the class name associated with each dataset and a fixed prompt (e.g. \"a photo of the [class name]\") into the text encoder. Next, we compute the similarity between the image embedding and the class-specific text embedding. This similarity score is then translated to compute softmax probabilities. These probabilities serve as the initial pseudo-labels. Mathematically, the similarity score for each class can be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.184, + 0.711, + 0.483, + 0.746 + ], + "angle": 0, + "content": "\\[\ns _ {c} = \\frac {f ^ {\\top} w _ {c}}{\\| f \\| \\cdot \\| w _ {c} \\|}, 1 \\leq c \\leq C \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.753, + 0.484, + 0.815 + ], + "angle": 0, + "content": "where \\( f \\) is a CLIP image embedding, \\( w_{c} \\) is a class-specific CLIP text embedding and \\( c \\) indicates a class index. With class-wise similarity, we can calculate a class-wise softmax probability \\( p_{c} \\) of each image as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.204, + 0.827, + 0.483, + 0.864 + ], + "angle": 0, + "content": "\\[\np _ {c} = \\frac {\\exp \\left(s _ {c} / \\tau\\right)}{\\sum_ {c = 1} ^ {C} \\exp \\left(s _ {c} / \\tau\\right)}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "where \\(\\tau\\) is a CLIP temperature parameter. We concatenate each class-wise softmax probability to generate initial" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.122 + ], + "angle": 0, + "content": "pseudo-labels. Finally, the initial pseudo-label for each image is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.571, + 0.132, + 0.905, + 0.158 + ], + "angle": 0, + "content": "\\[\nl ^ {\\text {i n i t i a l}} = \\left\\{p _ {1} ^ {\\text {g l o b a l}}, p _ {2} ^ {\\text {g l o b a l}}, \\dots , p _ {c} ^ {\\text {g l o b a l}} \\right\\}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.16, + 0.907, + 0.356 + ], + "angle": 0, + "content": "Following the acquisition of softmax probabilities, the initial pseudo-labels are used for the warm-up training of the classifier. Since the softmax calculation inherently normalizes the probabilities to ensure a sum of 1, it does not fit the multi-label classification problem where multiple classes have the same high probability (e.g. 1). Nevertheless, after the classifier undergoes warm-up training, it can predict each class independently, leading to recognizing multiple classes present in an image, albeit with limited accuracy. To address this performance limitation, in the subsequent section, we introduce a method to construct {training data, pseudo-label} pairs tailored for the multi-label classification problem, leveraging a warm-up trained classifier." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.364, + 0.854, + 0.38 + ], + "angle": 0, + "content": "3.2. Updating classifier-guided pseudo-label" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.385, + 0.907, + 0.61 + ], + "angle": 0, + "content": "Prior research [2] has shown that CLIP's inference results on the entire images may be suboptimal for multi-label classification. CLIP tends to assign high probabilities primarily to prominent objects in the input image. As a result, it does not effectively distribute probabilities across all existing objects. CDUL addresses this limitation by dividing the image into uniform patches in a grid fashion. Despite its simplicity, employing a uniform patch-based approach for local view acquisition may lead to the inference of background batch, where the patch does not contain any class object. More importantly, treating both simple images (i.e. those containing a single object) and complex images (i.e. those containing multiple objects) equally with the same number of local views is clearly inefficient in optimizing the resource (i.e. CLIP inference)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.612, + 0.906, + 0.838 + ], + "angle": 0, + "content": "Proof-of-concept study. We perform a proof-of-concept study to evaluate the impact of using object-specific locations in local view proposals. Our hypothesis is that incorporating class-related object locations enhances performance. We extract local views in four different scenarios, using ground truth bounding boxes (i.e., the bounding box that tightly covers the target object): (1) random cropping around the bounding box, (2) using only the bounding box itself\\(^{2}\\), (3) random cropping without guidance from the bounding box, and (4) uniform grid-based cropping. Each scenario produces nine patches per image. Subsequently, we infer pseudo-labels for these patches and train with each \\(\\{\\text{patch}, \\text{pseudo-label}\\}\\) pair. As shown in Fig. 3, our results indicate that the first scenario-(a) random cropping around the bounding box-achieves the best performance." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.839, + 0.905, + 0.869 + ], + "angle": 0, + "content": "Our proof-of-concept study reveals two critical insights: (1) the local view generation policy significantly influences" + }, + { + "type": "page_footnote", + "bbox": [ + 0.513, + 0.875, + 0.907, + 0.9 + ], + "angle": 0, + "content": "2For images with fewer than nine boxes, we additionally infer their resized versions." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4663" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.089, + 0.904, + 0.366 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.388, + 0.908, + 0.473 + ], + "angle": 0, + "content": "Figure 4. The overview of label preparation. (a) We calculate the cosine similarity between text embeddings and image embeddings. The softmax probability of these similarities is the CLIP prediction of each image. The top 1 probabilities are highlighted in green boxes. By class-wise averaging the top-1 probabilities, the CLIP bias is derived (yellow box). Pseudo-labels are then generated by debiasing CLIP predictions. (b) For label updating, classes above a threshold are selected from the classifier output, and local views corresponding to these classes are extracted. The process for acquiring local labels for each patch mirrors the initial label acquisition. The final pseudo-label is obtained through a weighted sum of initial pseudo-labels and local labels." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.499, + 0.482, + 0.529 + ], + "angle": 0, + "content": "CLIP's predictions, and (2) framing the view around the target object greatly improves classification accuracy." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.53, + 0.483, + 0.65 + ], + "angle": 0, + "content": "CAM-based view-selection. Motivated by the above two findings, we decided to approximate the first scenario in our approach. We apply Class Activation Mapping (CAM) with the trained classifier to identify vital regions within the image. This is called a CAM-based view selection method. This method generates {training image, pseudo-label} pairs by focusing on the object periphery, without relying on any ground-truth annotations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.652, + 0.483, + 0.743 + ], + "angle": 0, + "content": "By employing the trained classifier, CAM visualizes activations that most significantly influence class prediction. CAM calculation is specifically based on the weights assigned to each channel of the feature map, which influence the classification score. This can be represented by the following formula:" + }, + { + "type": "equation", + "bbox": [ + 0.198, + 0.758, + 0.483, + 0.801 + ], + "angle": 0, + "content": "\\[\nG _ {c} (x) = \\sum_ {i = 1} ^ {Q} \\mathbf {w} _ {i} ^ {c \\top} \\cdot g _ {i} (x), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.483, + 0.856 + ], + "angle": 0, + "content": "where \\( x \\) is input image, \\( g \\) is CNN network, \\( g(x) \\) is the feature map of the CNN, \\( \\mathbf{w}^c \\) is class-wise weight of classifier, and \\( Q \\) indicates the number of channels in the feature map." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.902 + ], + "angle": 0, + "content": "We first normalize the activation map to a range between 0 and 1, then apply a threshold to identify highly activated regions. Next, we generate candidate bounding boxes with" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.499, + 0.907, + 0.637 + ], + "angle": 0, + "content": "slight perturbation offsets. These bounding boxes define local views of the image. The regions highlighted by CAM act as reasonable approximations of the object's location, allowing us to extract local views without additional labeling costs. These localized view images are fed into CLIP to obtain softmax probabilities for the patches. The probabilities of local views are computed using the same method as for the global view, and then generating local labels that are focused on each local region." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.645, + 0.909, + 0.903 + ], + "angle": 0, + "content": "The number of local views. Additionally, we adjust the number of local views based on image complexity. For simpler images, the classifier typically provides confident, high-probability predictions. This translates to high logits for well-trained classes and near-zero for others. Conversely, complex images often produce uncertain predictions, characterized by higher logits distributed broadly across multiple classes. To account for this, we conduct additional local view sampling when the classifier output of a certain class surpasses a predefined threshold. Specifically, we crop the bounding box of any class exceeding this threshold value and obtain its local probability. Through our experiments, we empirically determined a fixed threshold that ensures appropriate average number of classes exceeding this threshold. Details on this ablation study can be found in the experiments section. This approach allocates fewer local views to simple images and more local views to" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4664" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.441, + 0.106 + ], + "angle": 0, + "content": "complex ones, optimizing the local view assignment." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.107, + 0.484, + 0.197 + ], + "angle": 0, + "content": "To obtain a single local pseudo-label for each input image, we select the maximum value of class-wise probability across all local patches from the corresponding image. Denoting the probability of \\(i\\)-th patch for class \\(c\\) as \\(p_{i,c}^{patch}\\), we conduct patch-wise aggregation using the maximum value as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.161, + 0.21, + 0.482, + 0.235 + ], + "angle": 0, + "content": "\\[\np _ {c} ^ {\\text {l o c a l}} = \\max _ {i = 1, \\dots , N} p _ {i, c} ^ {\\text {p a t c h}}, 1 \\leq c \\leq C \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.24, + 0.484, + 0.3 + ], + "angle": 0, + "content": "where \\(N\\) represents the number of local patches corresponding to each image, dependent on the input image itself. We then aggregate the local class-wise probability for generating a local pseudo-label as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.165, + 0.312, + 0.483, + 0.332 + ], + "angle": 0, + "content": "\\[\nl ^ {\\text {l o c a l}} = \\left\\{p _ {1} ^ {\\text {l o c a l}}, p _ {2} ^ {\\text {l o c a l}}, \\dots , p _ {c} ^ {\\text {l o c a l}} \\right\\}. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.336, + 0.483, + 0.38 + ], + "angle": 0, + "content": "We update the pseudo-label by the weighted sum of initial and local pseudo-labels. The final pseudo-label is calculated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.172, + 0.394, + 0.483, + 0.412 + ], + "angle": 0, + "content": "\\[\nl ^ {\\text {f i n a l}} = \\alpha l ^ {\\text {i n i t i a l}} + (1 - \\alpha) l ^ {\\text {l o c a l}}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.418, + 0.484, + 0.495 + ], + "angle": 0, + "content": "where \\(\\alpha\\) is a hyperparameter and we used 0.4. Through this label update, the pseudo-label efficiently incorporates information about less salient objects in multi-label classification. The updated pseudo-label is utilized throughout the remainder of the training process." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.501, + 0.353, + 0.518 + ], + "angle": 0, + "content": "3.3. Debiasing CLIP pseudo-label" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.524, + 0.483, + 0.689 + ], + "angle": 0, + "content": "Earlier, we aimed to identify highly activated areas within images through Class Activation Mapping and to infer additional label information focused on these areas. However, a challenge arises when using pseudo-labels generated by CLIP due to its inherent prediction bias. Consequently, classifiers trained using these pseudo-labels may inadvertently inherit this bias. Previous work [3] reveals that while classifiers are robust at overcoming random noise, they cannot rectify consistent biases associated with specific classes. Therefore, we need to overcome the inherent prediction bias of CLIP to enhance the performance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.69, + 0.484, + 0.902 + ], + "angle": 0, + "content": "To understand this bias, given CLIP's focus on salient objects, we treat global CLIP prediction as a single-label classification task. In a single-label classification, the predicted probability distribution for each class aligns with the class-wise average of the highest probabilities from each prediction. Following this principle, we compute the average prediction probability for each class, defining it as CLIP's prediction bias. Specifically, we compute CLIP bias as follows: (1) Obtain initial pseudo-labels by running each image in the \"target training set\" through CLIP; (2) Record top-1 probabilities for samples below the entropy threshold to constrain the task to single-label classification; (3) Count each class occurrences only if it achieves the top 1 probability; (4)Normalize each class's softmax probability by its" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.096, + 0.905, + 0.259 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.268, + 0.907, + 0.366 + ], + "angle": 0, + "content": "Figure 5. The training process of our method. We train the classifier using cross-entropy targeting initial pseudo-label during the warm-up phase which is illustrated as the green line. After the classifier-guided label update, we train the classifier using the cross-entropy loss targeting updated pseudo-label and the cross-entropy loss between the logits of differently augmented inputs which are illustrated as the green line and the blue dashed line." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.395, + 0.905, + 0.44 + ], + "angle": 0, + "content": "occurrence count, defining this normalized probability as CLIP bias. This process does not require any ground-truth annotation, solely relying on the result of CLIP prediction." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.442, + 0.907, + 0.669 + ], + "angle": 0, + "content": "The bias obtained in this manner serves as a basis for calibrating CLIP-derived probabilities. Specifically, we calibrate by multiplying each class's softmax probability from CLIP by the inverse of the CLIP bias. This calibration process enables the classifier to learn a rectified probability distribution. Our debiasing technique leads to a noticeable improvement in the performance of classes previously challenged due to the inherent CLIP bias (see Tab. 3). Furthermore, we adjust excessively small probabilities to prevent the minimum probability value from approaching zero, aiding in learning the corrected distribution. In doing so, we guide the classifier as a whole to learn from this inverse distribution. This calibration is applied not only to the global probability but also to local inference, minimizing the influence of the CLIP bias on the classifier as much as possible." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.681, + 0.676, + 0.698 + ], + "angle": 0, + "content": "3.4. Overall training" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Consistency loss. The proposed method tackles noisy data by utilizing pseudo-labels as target labels for all {training data, pseudo-label} pairs. Since pseudo-labels inherently carry noise and our debiasing process can potentially amplify this noise, we employ consistency loss between the data and its augmented data as additional critics. This approach is inspired by the common practice in a semi-supervised regime, improving feature extractors for handling noisy data. In our framework, we introduce consistency loss alongside the conventional binary cross-entropy loss during training. Consistency loss involves applying different augmentation intensities to the same image. Also, it passes each augmented version through the classifier to en" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4665" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.152 + ], + "angle": 0, + "content": "sure consistent predictions. In Fig. 5, we employ weak augmentations, such as resizing and flipping, and strong augmentations, which comprise weak augmentation and additional color jittering." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.153, + 0.483, + 0.381 + ], + "angle": 0, + "content": "Training process. Initially, the entire training dataset undergoes inference through CLIP, followed by calculating the CLIP bias. Subsequently, we derive a global label by debiasing the global probability, utilizing it as an initial pseudolabel in Fig. 4 (a). The classifier is then warmed up with this pseudo-label as the target (the green line in Fig. 5). Following warm-up training, we update the pseudo-label with classifier-guided local inference. During this phase, the obtained local probabilities undergoes debiasing. Then they are combined to derive a local label, which then updates the existing pseudo-label as depicted in Fig. 4 (b). The classifier is subsequently trained with the updated pseudo-label for the remaining epochs (the green line and the blue dashed line in Fig. 5). For whole trainig, our loss function combines two terms:" + }, + { + "type": "equation", + "bbox": [ + 0.198, + 0.399, + 0.483, + 0.417 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {t o t a l}} = L _ {c e} + \\beta L _ {\\text {c o n s i s t}}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.426, + 0.483, + 0.472 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_{ce}\\) is cross-entropy loss and \\(\\mathcal{L}_{\\text{consist}}\\) is consistency loss. \\(\\beta\\) is a hyperparameter and we fix it as 0 during warm-up phase and 1 after." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.489, + 0.224, + 0.506 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.515, + 0.279, + 0.531 + ], + "angle": 0, + "content": "4.1. Experimental setup" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.538, + 0.483, + 0.779 + ], + "angle": 0, + "content": "Dataset & evaluation Metric. We conducted evaluations on four distinct datasets. The PASCAL VOC 2012 [11] dataset comprises 20 classes, with 5,717 images in the training set and 5,823 in the validation set. Similarly, the PASCAL VOC 2007 dataset also encompasses 20 classes, with a combined total of 5,011 training and validation images and 4,952 test images. MS COCO [23] consists of 80 classes, encompassing 82,081 training images and 40,137 validation images. Additionally, the NUSWIDE [6] dataset comprises 81 classes, with a training set of 150,000 images and a test set of 60,200 images. In cases where a separate test set was not available, the validation set was utilized for testing purposes. Mean Average Precision (mAP) was employed as the performance metric, with ground truth labels utilized solely for performance evaluation and not incorporated during the model training phase." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Implementation detail. For a fair comparison, we endeavored to set the hyperparameters for CLIP inference as closely as possible to those of the primary competitive method, CDUL. We acquired labels using the frozen CLIP ResNet50×64, and employed the ImageNet pre-trained ResNet101 as the backbone network for our classifier. The entropy threshold for calculating CLIP bias and the classifier threshold for local view inference is both 0.5. It Our" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.44 + ], + "angle": 0, + "content": "warm-up classifier is trained for two epochs. While the entire training process can proceed up to 10 epochs, we obtain our final model via early stopping to prevent overfitting. Specifically, we terminate training based on the gradient of mean Average Precision (mAP) of the training labels throughout the learning process. Training was ceased when the gradient of mAP reached the first local minimum. We utilized a learning rate of \\(10^{-5}\\) across all datasets and employed a batch size of 16. Undisclosed hyperparameters of CDUL were manually reproduced. For generating the initial pseudo-labels of global views, we maintained a fixed image resolution of \\(640 \\times 640\\). The CLIP prompts consistently used the phrase \"a photo of the [class name]\" except for the NUSWIDE dataset. (For a fair comparison with the existing method, CDUL, we employ the phrase \"a photo of a [class name]\" to match the pseudo-label performance.) For acquiring the initial pseudo-labels of local views, we introduced an offset value of 80 pixels for each vertexes when generating bounding boxes (i.e. a thresholded CAM-bounding box \\(\\pm 80\\) as our local views). Additionally, cropped images were resized to have a longer side of 640 pixels while preserving their aspect ratios to obtain CLIP probabilities." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.453, + 0.874, + 0.469 + ], + "angle": 0, + "content": "4.2. Comparison with state-of-the-art methods" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.477, + 0.907, + 0.763 + ], + "angle": 0, + "content": "Competitive methods. Multi-label classification can be partitioned depending on the type of supervision: fully supervised, weakly supervised, and unsupervised approaches. Firstly, in a fully supervised setting, we choose (1) the Naïve method using only binary cross-entropy (BCE) loss and (2) the Naïve method with additional label smoothing (LS) [7]. Secondly, we compared our method with recent weakly supervised methods, including partial labeled methods utilizing only \\(10\\%\\) of total labels such as (3) SARB [30], (4) ASL [33], (5) Chen et al. [5]. For single positive label methods, we choose (6) LL-R [19], (7) \\(\\mathrm{G}^2\\mathrm{NetPL}\\) [1]. Finally, in an unsupervised scenario, we obtained initial labels using CLIP and evaluated the performance of (8) Naïve AN [21], (9) Szegedy et al. [38], (10) Aodha et al. [26], (11) Durand et al. [10], (12) ROLE [7], which leveraging CLIP generated initial labels. We conducted a comparative analysis with (13) CDUL [2], the current state-of-the-art in an unsupervised multi-label classification task." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Comparison with fully & weakly methods. Compared to fully supervised methods, our Classifier Guided CLIP Distillation (CCD) demonstrates comparable performance on the PASCAL VOC 2012 and PASCAL VOC 2007 datasets. In particular, achieving similar performance without any human annotation is impressive. CCD outperforms various weakly supervised methods on the PASCAL VOC 2012 and PASCAL VOC 2007 datasets, although our performances on MS COCO and NUSWIDE are relatively lower. This" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4666" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.907, + 0.119 + ], + "angle": 0, + "content": "Table 1. Mean average precision mAP in (%) for different multi-label classification methods under different supervision levels. Bold represents the best results among unsupervised methods. The proposed achieved state-of-the-art performance." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.128, + 0.907, + 0.344 + ], + "angle": 0, + "content": "
Supervision levelAnnotationMethodVOC12VOC07COCONUS
Fully supervisedFully labeledBCE90.191.378.550.7
BCE-LS [7]91.692.679.451.7
Weakly supervisedPartial labeled (10%)SARB [30]-85.772.5-
ASL [33]-82.969.7-
Chen et al. [5]-81.568.1-
Single positive labeledLL-R [19]89.790.672.647.4
G2NetPL [1]89.589.972.548.5
UnsupervisedAnnotation freeNaive AN [21]85.586.565.140.8
Szegedy et al. [38]86.887.965.541.3
Aodha et al. [26]84.286.263.940.1
Durand et al. [10]81.383.163.239.4
ROLE [7]82.684.667.143.2
CDUL [2]88.689.069.244.0
CCD (Ours)90.191.070.344.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.368, + 0.483, + 0.41 + ], + "angle": 0, + "content": "Table 2. Ablation study of the proposed moduels. Mean average precision mAP in \\((\\%)\\) for different configuration uon the PASCAL VOC 2012 dataset. The best score is in bold." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.42, + 0.487, + 0.537 + ], + "angle": 0, + "content": "
Label updateCLIP debiasConsistencymAP (VOC12)
×××86.4
✓××88.7
✓✓×89.4
✓×✓88.8
✓✓✓90.1
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.566, + 0.483, + 0.748 + ], + "angle": 0, + "content": "suggests that CCD is more effective on datasets with a smaller number of classes, such as PASCAL VOC 2012 and 2007, which consist of 20 classes. On the other hand, the performance gain is relatively small for complex datasets with many classes. It can be attributed to the simplistic prompt, \"a photo of the [class name],\" which might not accurately represent the target classes and co-occurrence issue, which bring out noisy local labels. The performance of initial labels of the training sets for PASCAL VOC 2012, MS COCO, and NUSWIDE is \\(85.3\\%\\), \\(65.4\\%\\), and \\(41.2\\%\\), respectively. CCD consistently shows about a \\(3.4\\%\\) performance gain over all datasets." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.75, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Comparison with unsupervised methods. Compared to unsupervised methods, CCD exhibits superior performance across all datasets. Particularly noteworthy are the significant improvements of \\(1.5\\% \\mathrm{p}\\), \\(2\\% \\mathrm{p}\\), and \\(1.1\\% \\mathrm{p}\\) on PASCAL VOC 2012, PASCAL VOC 2007, and MS COCO, respectively, compared to the existing state-of-the-art methods. These results highlight the effectiveness of our method. However, in the case of the NUSWIDE dataset, the performance improvement is not as substantial. This can be attributed to the unique characteristics of the NUSWIDE" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.367, + 0.714, + 0.447 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.539, + 0.448, + 0.702, + 0.459 + ], + "angle": 0, + "content": "(a) The number of mean inferences" + }, + { + "type": "image", + "bbox": [ + 0.716, + 0.369, + 0.901, + 0.447 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.749, + 0.448, + 0.886, + 0.459 + ], + "angle": 0, + "content": "(b) Initial & Local weight \\((\\alpha)\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.485, + 0.907, + 0.528 + ], + "angle": 0, + "content": "Figure 6. Classification results (mAP) on PASCAL VOC 2012 depending on (a) the change of the number of mean inferences (b) the change of ratio of pseudo-label updating." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.553, + 0.907, + 0.659 + ], + "angle": 0, + "content": "dataset, where \\(20\\%\\) of the train/test images do not contain any class-related objects (i.e. \"none images\"). As a result, the initial pseudo-label acquisition process fails to form a high probability for specific classes in these images. Consequently, during local inference, a significant portion of inferences is assigned to \"none images,\" thereby reducing the quality of pseudo-labels of CCD." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.668, + 0.618, + 0.683 + ], + "angle": 0, + "content": "4.3. Ablation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Effects of each module. We evaluate the effectiveness of each component of our method in Tab. 2. The performance of the baseline model, the classifier trained with CLIP-generated global pseudo-labels, is \\(86.4\\%\\). When we update the pseudo-label with classifier-guided local inference, it shows a significant improvement of \\(2.3\\%\\) p. This shows that our CAM-based local view selection method and adaptively allocating the number of CLIP inferences are effective. It already exceeds the performance of previous state-of-the-art methods. Debiasing CLIP also contributed \\(0.7\\%\\) p, indicating that our proposed debiasing method effectively relieves the CLIP bias. The consistency loss alone adds only \\(0.1\\%\\) p performance gain. However, when CLIP debiasing and consistency loss are combined, a substantial improvement of" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4667" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.093, + 0.09, + 0.482, + 0.145 + ], + "angle": 0, + "content": "Table 3. Per-class AP in \\((\\%)\\) for without and with debiasing. We focus on the seven classes from the lowest mean probability. The result showcases that our debiasing is effective for enhancing the performance of biased classes." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.157, + 0.482, + 0.196 + ], + "angle": 0, + "content": "
Methodbottlechairtablepersonplantsofatv
w/o debias77.077.172.787.368.873.089.5
W debias77.679.074.889.272.176.692.0
" + }, + { + "type": "text", + "bbox": [ + 0.095, + 0.228, + 0.481, + 0.303 + ], + "angle": 0, + "content": "\\(1.4\\%\\) p is achieved. This result implies that pseudo-labels generated by compensating for CLIP bias may amplify both accurate information and noise. Since the consistency loss is effective for dealing with noisy data, it creates a synergy with CLIP debiasing." + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.306, + 0.482, + 0.593 + ], + "angle": 0, + "content": "The number of inference. We examine how model performance varies with different numbers of CLIP local inferences. We incrementally increase the average number of inferences to approximately 4, 6, 10, and 14. In Fig. 6 (a), the model achieves its best performance when the average number of inferences is around 10. Notably, increasing the number of inferences does not consistently enhance performance. Fewer inferences limit the allocation of local views to challenging samples, which hinders the ability to extract sufficient information from complex images, thereby negatively impacting negatively. In contrast, an excessive number of inferences results in the overallocation of local views, even for easy samples. It potentially introduces redundant inference for easy samples, making noisy pseudo-labels. In particular, when the average number of inferences is set to 10 or 14, the maximum number of local inferences remains capped at 20 (i.e. the total number of classes). It suggests that the inferences allocated to complex images have already reached saturation." + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.597, + 0.482, + 0.881 + ], + "angle": 0, + "content": "Initial & local weight. We investigate the impact of adjusting the ratio \\(\\alpha\\) between initial and local labels when updating the pseudo-labels (Eq. 7). In Fig. 6 (b), the results indicate that utilizing both global and local labels together leads to performance enhancement. This trend arises because two different label sources have complementary properties. The local label contains more information through multiple CLIP inferences of different local patches. However, since the reliability of CAM is not \\(100\\%\\), it may draw an invalid patch. Conversely, the initial label contains less information, but it is always generated from valid images. By blending these two sources, we reinforce consistent predictions between local and initial labels, thereby strengthening the signal of reliable labels. Additionally, we highlight that the use of our local labels alone (i.e. at weight 0) can outperforms CDUL. This confirms that our choice of local views and their labels extracts more informative learning signals than the mixture of the global view and uniformly sampled local views." + }, + { + "type": "text", + "bbox": [ + 0.095, + 0.887, + 0.481, + 0.899 + ], + "angle": 0, + "content": "Per-class mAP of biased classes. We assess the effective" + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.288 + ], + "angle": 0, + "content": "ness of our proposed method by examining the improvement in performance for specific target classes before and after CLIP debiasing. In Tab. 3, we focus on seven classes from the biased class, including bottle, chair, table, person, potted plant, sofa, and tv monitor as confirmed in the Fig. 2. Our results demonstrate consistent improvements in the performance of all these classes, ranging from \\(0.6\\% \\mathrm{p}\\) (bottle) to \\(3.6\\% \\mathrm{p}\\) (sofa). Moreover, we observe comparable or slightly enhanced performance for the remaining classes not shown in the table. The result affirms the effectiveness of our method across a broader spectrum. Detailed class-wise performance results for the entire dataset are provided in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.517, + 0.298, + 0.631, + 0.311 + ], + "angle": 0, + "content": "4.4. Limitation" + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.32, + 0.905, + 0.545 + ], + "angle": 0, + "content": "A limitation of proposed method is that the generated local views often struggle to distinguish between co-occurring objects, an issue inherited from CAM. This challenge is especially evident in complex datasets like COCO, where similar classes—such as skies, snowboards, and skateboards—frequently appear in scene with human activity. These classes often co-occur with human feet, leading classifiers to associate high activations near the feet with these classes. This phenomenon makes them potential candidates for local views. This redundant information results in noisy pseudo-labels, reducing overall performance. Examples of such failure cases are provided in the supplementary materials. Developing an improved local view proposal method to better handle noisy inferences and focusing on the target objects will be an important direction for future work." + }, + { + "type": "title", + "bbox": [ + 0.517, + 0.56, + 0.631, + 0.573 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.585, + 0.905, + 0.9 + ], + "angle": 0, + "content": "In this paper, we propose a novel unsupervised multi-label classification method, Classifier-guided CLIP Distillation (CCD). We address two critical issues in CLIP: inconsistent predictions due to subtle input changes and inherent prediction bias. To tackle these issues, we devised a classifier-guided label update and CLIP debiasing. Firstly, we identified the effectiveness of acquiring local views near class-related objects and proposed a classifier-guided label update. It selects local views utilizing the Class Activation Mapping (CAM) and acquires corresponding labels from CLIP. Our approach results in allocating more local views for complex images. Secondly, we address the CLIP's bias. We investigated the biased prediction distribution of CLIP and rectified the pseudo-labels. This step notably boosts performance in certain classes and reduces the impact of bias on CLIP. Through experiments on four benchmarks, CCD newly achieved state-of-the-art performances in unsupervised multi-label classification. Remarkably, CCD achieved comparable performance to fully supervised methods on PASCAL VOC 2012 and PASCAL VOC 2007 datasets without relying on annotations." + }, + { + "type": "page_number", + "bbox": [ + 0.485, + 0.947, + 0.515, + 0.956 + ], + "angle": 0, + "content": "4668" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.26 + ], + "angle": 0, + "content": "Acknowledgements. This research was supported by the Basic Science Research Program through the National Research Foundation of Korea (NRF) funded by the MSIP (RS-2025-00520207, RS-2023-00219019), KEIT grant funded by the Korean government (MOTIE) (No. 2022-0-00680, No. 2022-0-01045), the IITP grant funded by the Korean government (MSIT) (No. RS-2024-00457882, National AI Research Lab Project, No. 2021-0-02068 Artificial Intelligence Innovation Hub, RS-2019-II190075 Artificial Intelligence Graduate School Program (KAIST)), Samsung Electronics Co., Ltd (IO230508-06190-01) and SAMSUNG Research, Samsung Electronics Co., Ltd." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.289, + 0.188, + 0.305 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.314, + 0.482, + 0.369 + ], + "angle": 0, + "content": "[1] Rabab Abdelfattah, Xin Zhang, Mostafa M Fouda, Xiaofeng Wang, and Song Wang. G2netpl: Generic game-theoretic network for partial-label image classification. arXiv preprint arXiv:2210.11469, 2022. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.372, + 0.482, + 0.441 + ], + "angle": 0, + "content": "[2] Rabab Abdelfattah, Qing Guo, Xiaoguang Li, Xiaofeng Wang, and Song Wang. Cdul: Clip-driven unsupervised learning for multi-label image classification. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1348-1357, 2023. 1, 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.443, + 0.483, + 0.471 + ], + "angle": 0, + "content": "[3] Görkem Algan and Ülkay Ulusoy. Label noise types and their effects on deep learning, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.473, + 0.482, + 0.514 + ], + "angle": 0, + "content": "[4] Serhat Selcuk Bucak, Rong Jin, and Anil K. Jain. Multi-label learning with incomplete class assignments. In CVPR 2011, pages 2801-2808, 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.517, + 0.482, + 0.572 + ], + "angle": 0, + "content": "[5] Tianshui Chen, Tao Pu, Hefeng Wu, Yuan Xie, and Liang Lin. Structured semantic transfer for multi-label recognition with partial labels. In Proceedings of the AAAI conference on artificial intelligence, pages 339-346, 2022. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.574, + 0.482, + 0.643 + ], + "angle": 0, + "content": "[6] Tat-Seng Chua, Jinhui Tang, Richang Hong, Haojie Li, Zhiping Luo, and Yantao Zheng. Nus-wide: a real-world web image database from national university ofSingapore. In Proceedings of the ACM international conference on image and video retrieval, pages 1-9, 2009. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.646, + 0.482, + 0.714 + ], + "angle": 0, + "content": "[7] Elijah Cole, Oisin Mac Aodha, Titouan Lorieul, Pietro Perona, Dan Morris, and Nebojsa Jojic. Multi-label learning from single positive labels. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 933-942, 2021. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.717, + 0.482, + 0.799 + ], + "angle": 0, + "content": "[8] Katherine Crowson, Stella Biderman, Daniel Kornis, Dashiell Stander, Eric Hallahan, Louis Castricato, and Edward Raff. Vqgan-clip: Open domain image generation and editing with natural language guidance. In European Conference on Computer Vision, pages 88–105. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.803, + 0.482, + 0.87 + ], + "angle": 0, + "content": "[9] Jia Deng, Olga Russakovsky, Jonathan Krause, Michael S Bernstein, Alex Berg, and Li Fei-Fei. Scalable multi-label annotation. In Proceedings of the SIGCHI Conference on Human Factors in Computing Systems, pages 3099–3102, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[10] Thibaut Durand, Nazanin Mehrasa, and Greg Mori. Learning a deep convnet for multi-label classification with partial" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.314, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "labels. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 647-657, 2019. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[11] M. Everingham, L. Van Gool, C. K. I. Williams, J. Winn, and A. Zisserman. The Pascal visual object classes (voc) challenge. International Journal of Computer Vision, 88(2): 303-338, 2010. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.193, + 0.905, + 0.247 + ], + "angle": 0, + "content": "[12] Xiuye Gu, Tsung-Yi Lin, Weicheng Kuo, and Yin Cui. Open-vocabulary object detection via vision and language knowledge distillation. arXiv preprint arXiv:2104.13921, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.25, + 0.905, + 0.305 + ], + "angle": 0, + "content": "[13] Ping Hu, Ximeng Sun, Stan Sclaroff, and Kate Saenko. Dualcoop++: Fast and effective adaptation to multi-label recognition with limited annotations. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.307, + 0.905, + 0.362 + ], + "angle": 0, + "content": "[14] Dat Huynh and Ehsan Elhamifar. Interactive multi-label cnn learning with partial labels. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9423-9432, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.364, + 0.905, + 0.445 + ], + "angle": 0, + "content": "[15] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International conference on machine learning, pages 4904-4916. PMLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.448, + 0.905, + 0.502 + ], + "angle": 0, + "content": "[16] Ashish Kapoor, Raajay Viswanathan, and Prateek Jain. Multilabel classification using bayesian compressed sensing. Advances in neural information processing systems, 25, 2012. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.505, + 0.905, + 0.574 + ], + "angle": 0, + "content": "[17] Apoorv Khandelwal, Luca Weihs, Roozbeh Mottaghi, and Aniruddha Kembhavi. Simple but effective: Clip embeddings for embodied ai. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14829-14838, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.577, + 0.905, + 0.617 + ], + "angle": 0, + "content": "[18] Dongseob Kim, Seungho Lee, Junsuk Choe, and Hyunjung Shim. Weakly supervised semantic segmentation for driving scenes. arXiv preprint arXiv:2312.13646, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.62, + 0.905, + 0.687 + ], + "angle": 0, + "content": "[19] Youngwook Kim, Jae Myung Kim, Zeynep Akata, and Jungwoo Lee. Large loss matters in weakly supervised multi-label classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14156-14165, 2022. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.69, + 0.905, + 0.772 + ], + "angle": 0, + "content": "[20] Youngwook Kim, Jae Myung Kim, Jieun Jeong, Cordelia Schmid, Zeynep Akata, and Jungwoo Lee. Bridging the gap between model explanations in partially annotated multi-label classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3408-3417, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.775, + 0.905, + 0.829 + ], + "angle": 0, + "content": "[21] Kaustav Kundu and Joseph Tighe. Exploiting weakly supervised visual patterns to learn from partial annotations. Advances in Neural Information Processing Systems, 33:561-572, 2020. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.832, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[22] Yue Liao, Aixi Zhang, Miao Lu, Yongliang Wang, Xiaobo Li, and Si Liu. Gen-vlkt: Simplify association and enhance interaction understanding for hoi detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20123-20132, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4669" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.176 + ], + "angle": 0, + "content": "[23] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.178, + 0.482, + 0.233 + ], + "angle": 0, + "content": "[24] Bing-Qing Liu, Bin-Bin Jia, and Min-Ling Zhang. Towards enabling binary decomposition for partial multi-label learning. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(11):13203-13217, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.235, + 0.482, + 0.276 + ], + "angle": 0, + "content": "[25] Weiwei Liu and Ivor Tsang. On the optimality of classifier chain for multi-label classification. Advances in Neural Information Processing Systems, 28, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.278, + 0.482, + 0.346 + ], + "angle": 0, + "content": "[26] Oisin Mac Aodha, Elijah Cole, and Pietro Perona. Presence-only geographical priors for fine-grained image classification. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9596-9606, 2019. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.349, + 0.482, + 0.432 + ], + "angle": 0, + "content": "[27] Dhruv Mahajan, Ross Girshick, Vignesh Ramanathan, Kaiming He, Manohar Paluri, Yixuan Li, Ashwin Bharambe, and Laurens Van Der Maaten. Exploring the limits of weakly supervised pretraining. In Proceedings of the European conference on computer vision (ECCV), pages 181-196, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.434, + 0.482, + 0.517 + ], + "angle": 0, + "content": "[28] Antoine Miech, Jean-Baptiste Alayrac, Lucas Smaira, Ivan Laptev, Josef Sivic, and Andrew Zisserman. End-to-end learning of visual representations from uncurated instructional videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9879-9889, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.519, + 0.482, + 0.589 + ], + "angle": 0, + "content": "[29] Ishan Misra, C Lawrence Zitnick, Margaret Mitchell, and Ross Girshick. Seeing through the human reporting bias: Visual classifiers from noisy human-centric labels. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2930–2939, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.591, + 0.482, + 0.659 + ], + "angle": 0, + "content": "[30] Tao Pu, Tianshui Chen, Hefeng Wu, and Liang Lin. Semantic-aware representation blending for multi-label image recognition with partial labels. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2091-2098, 2022. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.661, + 0.482, + 0.744 + ], + "angle": 0, + "content": "[31] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.746, + 0.482, + 0.829 + ], + "angle": 0, + "content": "[32] Yongming Rao, Wenliang Zhao, Guangyi Chen, Yansong Tang, Zheng Zhu, Guan Huang, Jie Zhou, and Jiwen Lu. Denseclip: Language-guided dense prediction with context-aware prompting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18082-18091, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.831, + 0.482, + 0.902 + ], + "angle": 0, + "content": "[33] Tal Ridnik, Emanuel Ben-Baruch, Nadav Zamir, Asaf Noy, Itamar Friedman, Matan Protter, and Lihi Zelnik-Manor. Asymmetric loss for multi-label classification. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 82-91, 2021. 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.175 + ], + "angle": 0, + "content": "[34] Amanpreet Singh, Ronghang Hu, Vedanuj Goswami, Guillaume Couairon, Wojciech Galuba, Marcus Rohrbach, and Douwe Kiela. Flava: A foundational language and vision alignment model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15638-15650, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.177, + 0.905, + 0.245 + ], + "angle": 0, + "content": "[35] GyeongRyeol Song, Noo-ri Kim, Jin-Seop Lee, and JeeHyong Lee. Ignore: Information gap-based false negative loss rejection for single positive multi-label learning. In European Conference on Computer Vision, pages 472–488. Springer, 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.247, + 0.905, + 0.302 + ], + "angle": 0, + "content": "[36] Chen Sun, Abhinav Shrivastava, Saurabh Singh, and Abhinav Gupta. Revisiting unreasonable effectiveness of data in deep learning era. In Proceedings of the IEEE international conference on computer vision, pages 843-852, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.302, + 0.905, + 0.344 + ], + "angle": 0, + "content": "[37] Yu-Yin Sun, Yin Zhang, and Zhi-Hua Zhou. Multi-label learning with weak label. Proceedings of the AAAI Conference on Artificial Intelligence, 24(1):593-598, 2010. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.344, + 0.905, + 0.412 + ], + "angle": 0, + "content": "[38] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2818-2826, 2016. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.413, + 0.905, + 0.469 + ], + "angle": 0, + "content": "[39] Guy Tevet, Brian Gordon, Amir Hertz, Amit H Bermano, and Daniel Cohen-Or. Motionclip: Exposing human motion generation to clip space. In European Conference on Computer Vision, pages 358–374. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.47, + 0.905, + 0.511 + ], + "angle": 0, + "content": "[40] Grigorios Tsoumakas and Ioannis Katakis. Multi-label classification: An overview. International Journal of Data Warehousing and Mining (IJDWM), 3(3):1-13, 2007. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.511, + 0.905, + 0.58 + ], + "angle": 0, + "content": "[41] Deepak Vasisht, Andreas Damianou, Manik Varma, and Ashish Kapoor. Active learning for sparse bayesian multilabel classification. In Proceedings of the 20th ACM SIGKDD international conference on Knowledge discovery and data mining, pages 472-481, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.58, + 0.905, + 0.663 + ], + "angle": 0, + "content": "[42] Thomas Verelst, Paul K. Rubenstein, Marcin Eichner, Tinne Tuytelaars, and Maxim Berman. Spatial consistency loss for training multi-label classifiers from single-label annotations. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 3879-3889, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.664, + 0.905, + 0.72 + ], + "angle": 0, + "content": "[43] Changhu Wang, Shuicheng Yan, Lei Zhang, and Hong-Jiang Zhang. Multi-label sparse coding for automatic image annotation. In 2009 IEEE conference on computer vision and pattern recognition, pages 1643-1650. IEEE, 2009. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.72, + 0.905, + 0.789 + ], + "angle": 0, + "content": "[44] Can Wang, Menglei Chai, Mingming He, Dongdong Chen, and Jing Liao. Clip-nerf: Text-and-image driven manipulation of neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3835-3844, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.79, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[45] Qifan Wang, Bin Shen, Shumiao Wang, Liang Li, and Luo Si. Binary codes embedding for fast image tagging with incomplete labels. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part II 13, pages 425-439. Springer, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.872, + 0.905, + 0.902 + ], + "angle": 0, + "content": "[46] Baoyuan Wu, Siwei Lyu, and Bernard Ghanem. Ml-mg: Multi-label learning with missing labels using a mixed" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.956 + ], + "angle": 0, + "content": "4670" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.484, + 0.12 + ], + "angle": 0, + "content": "graph. In Proceedings of the IEEE international conference on computer vision, pages 4157-4165, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.484, + 0.177 + ], + "angle": 0, + "content": "[47] Xin Xing, Zhexiao Xiong, Abby Stylianou, Srikumar Sastry, Liyu Gong, and Nathan Jacobs. Vision-language pseudolabels for single-positive multi-label learning. arXiv preprint arXiv:2310.15985, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.179, + 0.484, + 0.246 + ], + "angle": 0, + "content": "[48] Alireza Zareian, Kevin Dela Rosa, Derek Hao Hu, and Shih-Fu Chang. Open-vocabulary object detection using captions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14393-14402, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.249, + 0.484, + 0.317 + ], + "angle": 0, + "content": "[49] Bolei Zhou, Aditya Khosla, Agata Lapedriza, Aude Oliva, and Antonio Torralba. Learning deep features for discriminative localization. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2921-2929, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.32, + 0.484, + 0.361 + ], + "angle": 0, + "content": "[50] Chong Zhou, Chen Change Loy, and Bo Dai. Extract free dense labels from clip. In European Conference on Computer Vision, pages 696-712. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.362, + 0.484, + 0.431 + ], + "angle": 0, + "content": "[51] Donghao Zhou, Pengfei Chen, Qiong Wang, Guangyong Chen, and Pheng-Ann Heng. Acknowledging the unknown for multi-label learning with single positive labels. In European Conference on Computer Vision, pages 423-440. Springer, 2022. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.431 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.515, + 0.956 + ], + "angle": 0, + "content": "4671" + } + ] +] \ No newline at end of file diff --git a/2025/Classifier-guided CLIP Distillation for Unsupervised Multi-label Classification/325c3ac3-6589-4c70-b2e1-2f7ac7877553_origin.pdf b/2025/Classifier-guided CLIP Distillation for Unsupervised Multi-label Classification/325c3ac3-6589-4c70-b2e1-2f7ac7877553_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..ac9d57c778b0dd4967e3e4ba917952f55b04a397 --- /dev/null +++ b/2025/Classifier-guided CLIP Distillation for Unsupervised Multi-label Classification/325c3ac3-6589-4c70-b2e1-2f7ac7877553_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:240341b0be4aeefe4f786f3dddad95e1998e47380b36d46a7a2262914197334a +size 1190437 diff --git a/2025/Classifier-guided CLIP Distillation for Unsupervised Multi-label Classification/full.md b/2025/Classifier-guided CLIP Distillation for Unsupervised Multi-label Classification/full.md new file mode 100644 index 0000000000000000000000000000000000000000..65195a11386da1d74aa5a986772bae4b12234dd6 --- /dev/null +++ b/2025/Classifier-guided CLIP Distillation for Unsupervised Multi-label Classification/full.md @@ -0,0 +1,325 @@ +# Classifier-guided CLIP Distillation for Unsupervised Multi-label Classification + +Dongseob Kim +Samsung Electronics, Republic of Korea + +kou.kim@samsung.com + +Hyunjung Shim KAIST, Republic of Korea + +kateshim@kaist.ac.kr + +# Abstract + +Multi-label classification is crucial for comprehensive image understanding, yet acquiring accurate annotations is challenging and costly. To address this, a recent study suggests exploiting unsupervised multi-label classification leveraging CLIP, a powerful vision-language model. Despite CLIP's proficiency, it suffers from view-dependent predictions and inherent bias, limiting its effectiveness. We propose a novel method that addresses these issues by leveraging multiple views near target objects, guided by Class Activation Mapping (CAM) of the classifier, and debiasing pseudo-labels derived from CLIP predictions. Our Classifier-guided CLIP Distillation (CCD) enables selecting multiple local views without extra labels and debiasing predictions to enhance classification performance. Experimental results validate our method's superiority over existing techniques across diverse datasets. The code is available at https://github.com/k0u-id/CCD. + +# 1. Introduction + +Multi-label classification plays a pivotal role in identifying multiple classes within an image, serving as a foundational component in image understanding [25, 27, 29, 36, 40]. Its applications are widespread, including recommendation systems, surveillance, and image retrieval [43]. However, acquiring multi-label annotations poses challenges due to the difficulty and significant costs involved in meticulously annotating all relevant objects within an image. + +To solve this problem, various label-efficient techniques have been proposed. A partial label setting [5, 10, 14, 20, 24, 30] annotates only a portion of the classes appearing in an image and considers the rest as unknown labels. In the extreme, single positive label settings [7, 19, 35, 42, 51] provide only one positive label per image, with all other labels assumed unknown. However, these methods still incur labeling costs. Besides, current studies simulate partial and single positive settings by randomly omitting labels from a fully labeled dataset, which may not accurately reflect the actual labeling practices. For example, humans tend to pri + +![](images/81948ba10a0ea8b6dcd29b979461ede4a68b1358709dd6f4d8993e4a0a941df9.jpg) +H. 0.99 / P. 0.00 (a) original image +Figure 1. CLIP prediction probability of horse (H) and person (P) corresponding to different input patches from single image: (a) The original image and its CLIP prediction. (b) The $1/4$ sized patch and its CLIP prediction. (c) The $1/9$ sized patch and its CLIP prediction. (d) The $1/16$ sized patch and its CLIP prediction. The same colored box indicates each patch cropped from the corresponding box. + +![](images/0745cb1f1b35809050fcfae8df1076a832f20e325b1129bda770d7aa8bcaa314.jpg) +H:0.96 /P:0.00 (b)1/4 patch + +![](images/32655c6fe797e810dd6eefd3686ad278d2297679375e08c863db5d8139fc0794.jpg) +H:0.14/P:0.12 (c) 1/9 + +![](images/1c5bd6b89078ae350b37dd22545cf6fbae0052265c6e88ef062ccce7b84a042e.jpg) +:0.05/P:0.65 (d) 1/16 patch + +oritize labeling trivial objects first. + +To overcome these limitations, we focus on an unsupervised multi-label classification problem leveraging a vision-language model. One such model is CLIP [31], trained on an extensive dataset of 400M {image, text} pairs from the web. As a result, this model excels at recognizing diverse visual attributes and object categories in a zero-shot manner. Most recently, CDUL [2] applied CLIP for unsupervised multi-label classification by using it as a pseudolabeling function. CDUL generates pseudo-labels for both global and local views to help identify less prominent objects, training a multi-label classifier with each view-label pair. However, CDUL overlooks the limitations of CLIP, which hinders its prediction performance. + +In this paper, we identified two reliability issues in CLIP, addressing them to fully harness CLIP's potential. Firstly, we demonstrate that CLIP's predictions are highly view-dependent, considerably impacting classification performance. In a toy experiment, we observe that changing the crop position and size of an image leads to notable changes in CLIP's predictions. For example, in Fig. 1, when the cropped area is slightly shifted or increased, the probabilities for the horse and person classes vary significantly. Based on these key observations, we use multiple local views near the target object to mitigate the uncertainty of CLIP's pseudo-labels caused by view changes. Specifically, + +![](images/60088c0668175707771d18910ba604eaea5adee25790c5032fd813a2987f6cce.jpg) +T:0.35/C:0.29 + +![](images/37478862c79ef123fbaf49e2455cd93a54a16254b1cb9d166c4bfb19872fdb83.jpg) +P:0.50/C:0.21 + +![](images/1d11a6b5723afdc8abad5b59af6e729df59af67e13f64a139337163f67269144.jpg) +H:1.00 + +![](images/9a42edc92de4dd5a8a7bfbc895adec65846fe2bb2607fa46e6d8eca6a855ffb7.jpg) +(a) tv monitor (T) +(b) person (P) +(c) horse (H) +(d) CLIP probability distribution at VOC 2012 +Figure 2. Sample images showcasing CLIP bias. C indicates the probability of chair. (a) Top 1 probability of "tv monitor" image is $35\%$ . (b) Top 1 probability of a "person" image is $50\%$ . (c) Top 1 probability of a "horse" image is $100\%$ . (d) The mean class-wise probability of PASCAL VOC 2012. We can observe the class-wise prediction bias of CLIP from these results. + +we acquire various local views extracted by the bounding box using the Class Activation Mapping (CAM) [49] method, where the CAM is computed from the trained classifier itself. Additionally, our method can draw more local views for hard samples whose prediction exceeds a certain threshold across multiple classes. + +Another notable issue with CLIP is its inherent prediction bias, which tends to reduce the predicted probability of specific classes. Due to polysemy [31], the words representing each class can be ambiguous in text embedding space. As illustrated in Fig. 2 (a) and (b), despite the presence of a single class (tv monitor and person, respectively), CLIP assigns considerably low top-1 probabilities of around $35\%$ and $50\%$ . Conversely, objects like horses exhibit notably higher probabilities, as depicted in Fig. 2 (c). Analyzing the average CLIP top-1 probabilities across classes, as shown in Fig. 2 (d), further highlights this bias, indicating a lower prediction for certain classes compared to others. The biased prediction is evident across various datasets such as MS COCO and NUSWIDE1. Since the proposed method relies on pseudo-labels generated by CLIP, its bias naturally transfers to the classifier. Consequently, the classifier shows a similar prediction bias and performance degradation. To mitigate this issue, we adjust all pseudo-labels by applying the inverse transform of the bias. While this adjustment can amplify unwanted noise in pseudo-labels, we apply consistency loss to handle this noise after debiasing. Our adjustment helps debiasing the classifier, thereby enhancing its generalization performance. As a result, our proposed Classifier-guided CLIP Distillation (CCD) achieves state-of-the-art performances on the PASCAL VOC 2012, PASCAL VOC 2007, COCO, and NUSWIDE benchmarks. + +# 2. Related Work + +# 2.1. Label efficient multi-label classification. + +Unlike multi-class classification, where a single class is identified in an image, multi-label classification requires accurately identifying all classes present in an image. Labeling every class in an image is significantly more challenging and resource-intensive than identifying just one class. To address this issue, a partial label setting has been proposed [5, 10, 14, 20, 24, 30], which utilizes only a subset of labels associated with each image. + +Early approaches assumed any unannotated labels (i.e. "unknown labels") as negative labels [4, 37, 45]. These methods involved correcting these labels during the learning process. However, approaches that relied on probabilistic models [9, 16, 41, 46] for label correction were dependent on optimization techniques, resulting in limited scalability. Subsequently, alternative methods have been proposed, such as leveraging label-to-label similarity [14], cross-image semantic correlation [5], or category-specific representation [30] to refine labels. Some approaches suggest using soft-pseudo-labels instead of unknown labels [10, 26, 38]. + +Research has extended to a more extreme scenario: single positive label setting [7, 19, 42, 51], where only a single positive label is provided per image. These methods build upon those proposed in the partial label setting and adapt them to function effectively in more challenging contexts. + +More recently, CDUL [2] has addressed the unsupervised multi-label classification problem by utilizing CLIP. In a similar vein, our method tackles the unsupervised multi-label classification problem using CLIP while additionally providing local area information with CAM and addressing bias in CLIP predictions. + +# 2.2. Vision-language pretraining. + +Vision-language pretraining (VLP) has emerged as a pivotal approach for bridging the gap between visual and textual modalities [15, 28, 34]. Among the pioneering VLPs, CLIP [31] stands out as the most commonly utilized framework. CLIP was trained on large {image, text} pairs sourced from the Internet. Recently, several studies have adopted VLP to provide rich visual-semantic knowledge for downstream image tasks, such as detection [12, 48], segmentation [18, 32, 50], human-object interation [17, 22], and generation [8, 39, 44]. + +In multi-label classification, various methods have emerged to derive pseudo-labels from CLIP predictions [2, 13, 47]. Among these, CDUL [2] integrates local inferences with global inference, successfully expanding the utilization of the VLP to multi-label classification. However, CDUL fails to tackle two limitations inherent in CLIP: sensitivity to input view selection and inherent prediction bias. + +![](images/c3877900e645a0a43a464b410debe40e899b33c285104aee0fbbedf6ab0a7b39.jpg) +87.6% +(a) Around GT box + +![](images/06361d85478502a965ba0baff7c63cc13ee8f4c4ae428e933ddc4c89de38f378.jpg) +87.1% +(b) GT box + +![](images/1ea59459653e8aad6e275be5e67fdc3ad76b7dd141c77be1c08da820d070d3bd.jpg) +87.1% +(c) Random box +Figure 3. The proof-of-concept study of local view selection method. We train the classifier with pseudo-label generated from four different local views: (a) Around GT boxes, (b) GT boxes, (c) Random boxes, (d) Uniform grid boxes. The numbers below each sample are the performance (mAP) of the classifier trained with the corresponding pseudo-label for the entire training set. The classifier trained with local view around GT boxes achieved $1.3\%$ higher performance compared to classifier trained with uniform-grid local view. + +![](images/0071ec44c4ecff1e6e0a54d1b0d10480d757f919bba0a8102ced51a68d7d2769.jpg) +86.4% +(d) Uniform grid box + +Consequently, the pseudo-labels generated by CDUL exhibit noise and bias. In contrast, our approach leverages a fine-tuned classifier to suggest local views and addresses CLIP bias through calibration. + +# 3. Method + +# 3.1. Generating initial pseudo-labels + +CLIP, a powerful vision-language model, has been trained to align images with corresponding textual descriptions. Leveraging the powerful zero-shot capability of CLIP, we initially generate pseudo-labels similar to CDUL. For each image in the dataset, we extract its image embedding using the CLIP image encoder. Simultaneously, we obtain the text embedding by inputting the class name associated with each dataset and a fixed prompt (e.g. "a photo of the [class name]") into the text encoder. Next, we compute the similarity between the image embedding and the class-specific text embedding. This similarity score is then translated to compute softmax probabilities. These probabilities serve as the initial pseudo-labels. Mathematically, the similarity score for each class can be expressed as: + +$$ +s _ {c} = \frac {f ^ {\top} w _ {c}}{\| f \| \cdot \| w _ {c} \|}, 1 \leq c \leq C \tag {1} +$$ + +where $f$ is a CLIP image embedding, $w_{c}$ is a class-specific CLIP text embedding and $c$ indicates a class index. With class-wise similarity, we can calculate a class-wise softmax probability $p_{c}$ of each image as follows: + +$$ +p _ {c} = \frac {\exp \left(s _ {c} / \tau\right)}{\sum_ {c = 1} ^ {C} \exp \left(s _ {c} / \tau\right)}, \tag {2} +$$ + +where $\tau$ is a CLIP temperature parameter. We concatenate each class-wise softmax probability to generate initial + +pseudo-labels. Finally, the initial pseudo-label for each image is defined as: + +$$ +l ^ {\text {i n i t i a l}} = \left\{p _ {1} ^ {\text {g l o b a l}}, p _ {2} ^ {\text {g l o b a l}}, \dots , p _ {c} ^ {\text {g l o b a l}} \right\}. \tag {3} +$$ + +Following the acquisition of softmax probabilities, the initial pseudo-labels are used for the warm-up training of the classifier. Since the softmax calculation inherently normalizes the probabilities to ensure a sum of 1, it does not fit the multi-label classification problem where multiple classes have the same high probability (e.g. 1). Nevertheless, after the classifier undergoes warm-up training, it can predict each class independently, leading to recognizing multiple classes present in an image, albeit with limited accuracy. To address this performance limitation, in the subsequent section, we introduce a method to construct {training data, pseudo-label} pairs tailored for the multi-label classification problem, leveraging a warm-up trained classifier. + +# 3.2. Updating classifier-guided pseudo-label + +Prior research [2] has shown that CLIP's inference results on the entire images may be suboptimal for multi-label classification. CLIP tends to assign high probabilities primarily to prominent objects in the input image. As a result, it does not effectively distribute probabilities across all existing objects. CDUL addresses this limitation by dividing the image into uniform patches in a grid fashion. Despite its simplicity, employing a uniform patch-based approach for local view acquisition may lead to the inference of background batch, where the patch does not contain any class object. More importantly, treating both simple images (i.e. those containing a single object) and complex images (i.e. those containing multiple objects) equally with the same number of local views is clearly inefficient in optimizing the resource (i.e. CLIP inference). + +Proof-of-concept study. We perform a proof-of-concept study to evaluate the impact of using object-specific locations in local view proposals. Our hypothesis is that incorporating class-related object locations enhances performance. We extract local views in four different scenarios, using ground truth bounding boxes (i.e., the bounding box that tightly covers the target object): (1) random cropping around the bounding box, (2) using only the bounding box itself $^{2}$ , (3) random cropping without guidance from the bounding box, and (4) uniform grid-based cropping. Each scenario produces nine patches per image. Subsequently, we infer pseudo-labels for these patches and train with each $\{\text{patch}, \text{pseudo-label}\}$ pair. As shown in Fig. 3, our results indicate that the first scenario-(a) random cropping around the bounding box-achieves the best performance. + +Our proof-of-concept study reveals two critical insights: (1) the local view generation policy significantly influences + +![](images/bc0ad15f1147ebf3b38b7c660904aba36e8ff3ea63bff11ae9ad98cdaae849bf.jpg) +Figure 4. The overview of label preparation. (a) We calculate the cosine similarity between text embeddings and image embeddings. The softmax probability of these similarities is the CLIP prediction of each image. The top 1 probabilities are highlighted in green boxes. By class-wise averaging the top-1 probabilities, the CLIP bias is derived (yellow box). Pseudo-labels are then generated by debiasing CLIP predictions. (b) For label updating, classes above a threshold are selected from the classifier output, and local views corresponding to these classes are extracted. The process for acquiring local labels for each patch mirrors the initial label acquisition. The final pseudo-label is obtained through a weighted sum of initial pseudo-labels and local labels. + +CLIP's predictions, and (2) framing the view around the target object greatly improves classification accuracy. + +CAM-based view-selection. Motivated by the above two findings, we decided to approximate the first scenario in our approach. We apply Class Activation Mapping (CAM) with the trained classifier to identify vital regions within the image. This is called a CAM-based view selection method. This method generates {training image, pseudo-label} pairs by focusing on the object periphery, without relying on any ground-truth annotations. + +By employing the trained classifier, CAM visualizes activations that most significantly influence class prediction. CAM calculation is specifically based on the weights assigned to each channel of the feature map, which influence the classification score. This can be represented by the following formula: + +$$ +G _ {c} (x) = \sum_ {i = 1} ^ {Q} \mathbf {w} _ {i} ^ {c \top} \cdot g _ {i} (x), \tag {4} +$$ + +where $x$ is input image, $g$ is CNN network, $g(x)$ is the feature map of the CNN, $\mathbf{w}^c$ is class-wise weight of classifier, and $Q$ indicates the number of channels in the feature map. + +We first normalize the activation map to a range between 0 and 1, then apply a threshold to identify highly activated regions. Next, we generate candidate bounding boxes with + +slight perturbation offsets. These bounding boxes define local views of the image. The regions highlighted by CAM act as reasonable approximations of the object's location, allowing us to extract local views without additional labeling costs. These localized view images are fed into CLIP to obtain softmax probabilities for the patches. The probabilities of local views are computed using the same method as for the global view, and then generating local labels that are focused on each local region. + +The number of local views. Additionally, we adjust the number of local views based on image complexity. For simpler images, the classifier typically provides confident, high-probability predictions. This translates to high logits for well-trained classes and near-zero for others. Conversely, complex images often produce uncertain predictions, characterized by higher logits distributed broadly across multiple classes. To account for this, we conduct additional local view sampling when the classifier output of a certain class surpasses a predefined threshold. Specifically, we crop the bounding box of any class exceeding this threshold value and obtain its local probability. Through our experiments, we empirically determined a fixed threshold that ensures appropriate average number of classes exceeding this threshold. Details on this ablation study can be found in the experiments section. This approach allocates fewer local views to simple images and more local views to + +complex ones, optimizing the local view assignment. + +To obtain a single local pseudo-label for each input image, we select the maximum value of class-wise probability across all local patches from the corresponding image. Denoting the probability of $i$ -th patch for class $c$ as $p_{i,c}^{patch}$ , we conduct patch-wise aggregation using the maximum value as follows: + +$$ +p _ {c} ^ {\text {l o c a l}} = \max _ {i = 1, \dots , N} p _ {i, c} ^ {\text {p a t c h}}, 1 \leq c \leq C \tag {5} +$$ + +where $N$ represents the number of local patches corresponding to each image, dependent on the input image itself. We then aggregate the local class-wise probability for generating a local pseudo-label as follows: + +$$ +l ^ {\text {l o c a l}} = \left\{p _ {1} ^ {\text {l o c a l}}, p _ {2} ^ {\text {l o c a l}}, \dots , p _ {c} ^ {\text {l o c a l}} \right\}. \tag {6} +$$ + +We update the pseudo-label by the weighted sum of initial and local pseudo-labels. The final pseudo-label is calculated as follows: + +$$ +l ^ {\text {f i n a l}} = \alpha l ^ {\text {i n i t i a l}} + (1 - \alpha) l ^ {\text {l o c a l}}, \tag {7} +$$ + +where $\alpha$ is a hyperparameter and we used 0.4. Through this label update, the pseudo-label efficiently incorporates information about less salient objects in multi-label classification. The updated pseudo-label is utilized throughout the remainder of the training process. + +# 3.3. Debiasing CLIP pseudo-label + +Earlier, we aimed to identify highly activated areas within images through Class Activation Mapping and to infer additional label information focused on these areas. However, a challenge arises when using pseudo-labels generated by CLIP due to its inherent prediction bias. Consequently, classifiers trained using these pseudo-labels may inadvertently inherit this bias. Previous work [3] reveals that while classifiers are robust at overcoming random noise, they cannot rectify consistent biases associated with specific classes. Therefore, we need to overcome the inherent prediction bias of CLIP to enhance the performance. + +To understand this bias, given CLIP's focus on salient objects, we treat global CLIP prediction as a single-label classification task. In a single-label classification, the predicted probability distribution for each class aligns with the class-wise average of the highest probabilities from each prediction. Following this principle, we compute the average prediction probability for each class, defining it as CLIP's prediction bias. Specifically, we compute CLIP bias as follows: (1) Obtain initial pseudo-labels by running each image in the "target training set" through CLIP; (2) Record top-1 probabilities for samples below the entropy threshold to constrain the task to single-label classification; (3) Count each class occurrences only if it achieves the top 1 probability; (4)Normalize each class's softmax probability by its + +![](images/16611be1bb536704fa59c68cc0e2ac6a7a41dbdcba9a6e10c325bb5a06f6ac4d.jpg) +Figure 5. The training process of our method. We train the classifier using cross-entropy targeting initial pseudo-label during the warm-up phase which is illustrated as the green line. After the classifier-guided label update, we train the classifier using the cross-entropy loss targeting updated pseudo-label and the cross-entropy loss between the logits of differently augmented inputs which are illustrated as the green line and the blue dashed line. + +occurrence count, defining this normalized probability as CLIP bias. This process does not require any ground-truth annotation, solely relying on the result of CLIP prediction. + +The bias obtained in this manner serves as a basis for calibrating CLIP-derived probabilities. Specifically, we calibrate by multiplying each class's softmax probability from CLIP by the inverse of the CLIP bias. This calibration process enables the classifier to learn a rectified probability distribution. Our debiasing technique leads to a noticeable improvement in the performance of classes previously challenged due to the inherent CLIP bias (see Tab. 3). Furthermore, we adjust excessively small probabilities to prevent the minimum probability value from approaching zero, aiding in learning the corrected distribution. In doing so, we guide the classifier as a whole to learn from this inverse distribution. This calibration is applied not only to the global probability but also to local inference, minimizing the influence of the CLIP bias on the classifier as much as possible. + +# 3.4. Overall training + +Consistency loss. The proposed method tackles noisy data by utilizing pseudo-labels as target labels for all {training data, pseudo-label} pairs. Since pseudo-labels inherently carry noise and our debiasing process can potentially amplify this noise, we employ consistency loss between the data and its augmented data as additional critics. This approach is inspired by the common practice in a semi-supervised regime, improving feature extractors for handling noisy data. In our framework, we introduce consistency loss alongside the conventional binary cross-entropy loss during training. Consistency loss involves applying different augmentation intensities to the same image. Also, it passes each augmented version through the classifier to en + +sure consistent predictions. In Fig. 5, we employ weak augmentations, such as resizing and flipping, and strong augmentations, which comprise weak augmentation and additional color jittering. + +Training process. Initially, the entire training dataset undergoes inference through CLIP, followed by calculating the CLIP bias. Subsequently, we derive a global label by debiasing the global probability, utilizing it as an initial pseudolabel in Fig. 4 (a). The classifier is then warmed up with this pseudo-label as the target (the green line in Fig. 5). Following warm-up training, we update the pseudo-label with classifier-guided local inference. During this phase, the obtained local probabilities undergoes debiasing. Then they are combined to derive a local label, which then updates the existing pseudo-label as depicted in Fig. 4 (b). The classifier is subsequently trained with the updated pseudo-label for the remaining epochs (the green line and the blue dashed line in Fig. 5). For whole trainig, our loss function combines two terms: + +$$ +\mathcal {L} _ {\text {t o t a l}} = L _ {c e} + \beta L _ {\text {c o n s i s t}}, \tag {8} +$$ + +where $\mathcal{L}_{ce}$ is cross-entropy loss and $\mathcal{L}_{\text{consist}}$ is consistency loss. $\beta$ is a hyperparameter and we fix it as 0 during warm-up phase and 1 after. + +# 4. Experiments + +# 4.1. Experimental setup + +Dataset & evaluation Metric. We conducted evaluations on four distinct datasets. The PASCAL VOC 2012 [11] dataset comprises 20 classes, with 5,717 images in the training set and 5,823 in the validation set. Similarly, the PASCAL VOC 2007 dataset also encompasses 20 classes, with a combined total of 5,011 training and validation images and 4,952 test images. MS COCO [23] consists of 80 classes, encompassing 82,081 training images and 40,137 validation images. Additionally, the NUSWIDE [6] dataset comprises 81 classes, with a training set of 150,000 images and a test set of 60,200 images. In cases where a separate test set was not available, the validation set was utilized for testing purposes. Mean Average Precision (mAP) was employed as the performance metric, with ground truth labels utilized solely for performance evaluation and not incorporated during the model training phase. + +Implementation detail. For a fair comparison, we endeavored to set the hyperparameters for CLIP inference as closely as possible to those of the primary competitive method, CDUL. We acquired labels using the frozen CLIP ResNet50×64, and employed the ImageNet pre-trained ResNet101 as the backbone network for our classifier. The entropy threshold for calculating CLIP bias and the classifier threshold for local view inference is both 0.5. It Our + +warm-up classifier is trained for two epochs. While the entire training process can proceed up to 10 epochs, we obtain our final model via early stopping to prevent overfitting. Specifically, we terminate training based on the gradient of mean Average Precision (mAP) of the training labels throughout the learning process. Training was ceased when the gradient of mAP reached the first local minimum. We utilized a learning rate of $10^{-5}$ across all datasets and employed a batch size of 16. Undisclosed hyperparameters of CDUL were manually reproduced. For generating the initial pseudo-labels of global views, we maintained a fixed image resolution of $640 \times 640$ . The CLIP prompts consistently used the phrase "a photo of the [class name]" except for the NUSWIDE dataset. (For a fair comparison with the existing method, CDUL, we employ the phrase "a photo of a [class name]" to match the pseudo-label performance.) For acquiring the initial pseudo-labels of local views, we introduced an offset value of 80 pixels for each vertexes when generating bounding boxes (i.e. a thresholded CAM-bounding box $\pm 80$ as our local views). Additionally, cropped images were resized to have a longer side of 640 pixels while preserving their aspect ratios to obtain CLIP probabilities. + +# 4.2. Comparison with state-of-the-art methods + +Competitive methods. Multi-label classification can be partitioned depending on the type of supervision: fully supervised, weakly supervised, and unsupervised approaches. Firstly, in a fully supervised setting, we choose (1) the Naïve method using only binary cross-entropy (BCE) loss and (2) the Naïve method with additional label smoothing (LS) [7]. Secondly, we compared our method with recent weakly supervised methods, including partial labeled methods utilizing only $10\%$ of total labels such as (3) SARB [30], (4) ASL [33], (5) Chen et al. [5]. For single positive label methods, we choose (6) LL-R [19], (7) $\mathrm{G}^2\mathrm{NetPL}$ [1]. Finally, in an unsupervised scenario, we obtained initial labels using CLIP and evaluated the performance of (8) Naïve AN [21], (9) Szegedy et al. [38], (10) Aodha et al. [26], (11) Durand et al. [10], (12) ROLE [7], which leveraging CLIP generated initial labels. We conducted a comparative analysis with (13) CDUL [2], the current state-of-the-art in an unsupervised multi-label classification task. + +Comparison with fully & weakly methods. Compared to fully supervised methods, our Classifier Guided CLIP Distillation (CCD) demonstrates comparable performance on the PASCAL VOC 2012 and PASCAL VOC 2007 datasets. In particular, achieving similar performance without any human annotation is impressive. CCD outperforms various weakly supervised methods on the PASCAL VOC 2012 and PASCAL VOC 2007 datasets, although our performances on MS COCO and NUSWIDE are relatively lower. This + +Table 1. Mean average precision mAP in (%) for different multi-label classification methods under different supervision levels. Bold represents the best results among unsupervised methods. The proposed achieved state-of-the-art performance. + +
Supervision levelAnnotationMethodVOC12VOC07COCONUS
Fully supervisedFully labeledBCE90.191.378.550.7
BCE-LS [7]91.692.679.451.7
Weakly supervisedPartial labeled (10%)SARB [30]-85.772.5-
ASL [33]-82.969.7-
Chen et al. [5]-81.568.1-
Single positive labeledLL-R [19]89.790.672.647.4
G2NetPL [1]89.589.972.548.5
UnsupervisedAnnotation freeNaive AN [21]85.586.565.140.8
Szegedy et al. [38]86.887.965.541.3
Aodha et al. [26]84.286.263.940.1
Durand et al. [10]81.383.163.239.4
ROLE [7]82.684.667.143.2
CDUL [2]88.689.069.244.0
CCD (Ours)90.191.070.344.5
+ +Table 2. Ablation study of the proposed moduels. Mean average precision mAP in $(\%)$ for different configuration uon the PASCAL VOC 2012 dataset. The best score is in bold. + +
Label updateCLIP debiasConsistencymAP (VOC12)
×××86.4
✓××88.7
✓✓×89.4
✓×✓88.8
✓✓✓90.1
+ +suggests that CCD is more effective on datasets with a smaller number of classes, such as PASCAL VOC 2012 and 2007, which consist of 20 classes. On the other hand, the performance gain is relatively small for complex datasets with many classes. It can be attributed to the simplistic prompt, "a photo of the [class name]," which might not accurately represent the target classes and co-occurrence issue, which bring out noisy local labels. The performance of initial labels of the training sets for PASCAL VOC 2012, MS COCO, and NUSWIDE is $85.3\%$ , $65.4\%$ , and $41.2\%$ , respectively. CCD consistently shows about a $3.4\%$ performance gain over all datasets. + +Comparison with unsupervised methods. Compared to unsupervised methods, CCD exhibits superior performance across all datasets. Particularly noteworthy are the significant improvements of $1.5\% \mathrm{p}$ , $2\% \mathrm{p}$ , and $1.1\% \mathrm{p}$ on PASCAL VOC 2012, PASCAL VOC 2007, and MS COCO, respectively, compared to the existing state-of-the-art methods. These results highlight the effectiveness of our method. However, in the case of the NUSWIDE dataset, the performance improvement is not as substantial. This can be attributed to the unique characteristics of the NUSWIDE + +![](images/c2a15663f866cca08d7ad4bb59d99283bb7cc3c225cca6d2f9d525a20cedc33b.jpg) +(a) The number of mean inferences +Figure 6. Classification results (mAP) on PASCAL VOC 2012 depending on (a) the change of the number of mean inferences (b) the change of ratio of pseudo-label updating. + +![](images/9a2493902e34083513eee6b395d270e7380f1bec5d0e1c419236a3f6ac82d349.jpg) +(b) Initial & Local weight $(\alpha)$ + +dataset, where $20\%$ of the train/test images do not contain any class-related objects (i.e. "none images"). As a result, the initial pseudo-label acquisition process fails to form a high probability for specific classes in these images. Consequently, during local inference, a significant portion of inferences is assigned to "none images," thereby reducing the quality of pseudo-labels of CCD. + +# 4.3. Ablation + +Effects of each module. We evaluate the effectiveness of each component of our method in Tab. 2. The performance of the baseline model, the classifier trained with CLIP-generated global pseudo-labels, is $86.4\%$ . When we update the pseudo-label with classifier-guided local inference, it shows a significant improvement of $2.3\%$ p. This shows that our CAM-based local view selection method and adaptively allocating the number of CLIP inferences are effective. It already exceeds the performance of previous state-of-the-art methods. Debiasing CLIP also contributed $0.7\%$ p, indicating that our proposed debiasing method effectively relieves the CLIP bias. The consistency loss alone adds only $0.1\%$ p performance gain. However, when CLIP debiasing and consistency loss are combined, a substantial improvement of + +Table 3. Per-class AP in $(\%)$ for without and with debiasing. We focus on the seven classes from the lowest mean probability. The result showcases that our debiasing is effective for enhancing the performance of biased classes. + +
Methodbottlechairtablepersonplantsofatv
w/o debias77.077.172.787.368.873.089.5
W debias77.679.074.889.272.176.692.0
+ +$1.4\%$ p is achieved. This result implies that pseudo-labels generated by compensating for CLIP bias may amplify both accurate information and noise. Since the consistency loss is effective for dealing with noisy data, it creates a synergy with CLIP debiasing. + +The number of inference. We examine how model performance varies with different numbers of CLIP local inferences. We incrementally increase the average number of inferences to approximately 4, 6, 10, and 14. In Fig. 6 (a), the model achieves its best performance when the average number of inferences is around 10. Notably, increasing the number of inferences does not consistently enhance performance. Fewer inferences limit the allocation of local views to challenging samples, which hinders the ability to extract sufficient information from complex images, thereby negatively impacting negatively. In contrast, an excessive number of inferences results in the overallocation of local views, even for easy samples. It potentially introduces redundant inference for easy samples, making noisy pseudo-labels. In particular, when the average number of inferences is set to 10 or 14, the maximum number of local inferences remains capped at 20 (i.e. the total number of classes). It suggests that the inferences allocated to complex images have already reached saturation. + +Initial & local weight. We investigate the impact of adjusting the ratio $\alpha$ between initial and local labels when updating the pseudo-labels (Eq. 7). In Fig. 6 (b), the results indicate that utilizing both global and local labels together leads to performance enhancement. This trend arises because two different label sources have complementary properties. The local label contains more information through multiple CLIP inferences of different local patches. However, since the reliability of CAM is not $100\%$ , it may draw an invalid patch. Conversely, the initial label contains less information, but it is always generated from valid images. By blending these two sources, we reinforce consistent predictions between local and initial labels, thereby strengthening the signal of reliable labels. Additionally, we highlight that the use of our local labels alone (i.e. at weight 0) can outperforms CDUL. This confirms that our choice of local views and their labels extracts more informative learning signals than the mixture of the global view and uniformly sampled local views. + +Per-class mAP of biased classes. We assess the effective + +ness of our proposed method by examining the improvement in performance for specific target classes before and after CLIP debiasing. In Tab. 3, we focus on seven classes from the biased class, including bottle, chair, table, person, potted plant, sofa, and tv monitor as confirmed in the Fig. 2. Our results demonstrate consistent improvements in the performance of all these classes, ranging from $0.6\% \mathrm{p}$ (bottle) to $3.6\% \mathrm{p}$ (sofa). Moreover, we observe comparable or slightly enhanced performance for the remaining classes not shown in the table. The result affirms the effectiveness of our method across a broader spectrum. Detailed class-wise performance results for the entire dataset are provided in the supplementary material. + +# 4.4. Limitation + +A limitation of proposed method is that the generated local views often struggle to distinguish between co-occurring objects, an issue inherited from CAM. This challenge is especially evident in complex datasets like COCO, where similar classes—such as skies, snowboards, and skateboards—frequently appear in scene with human activity. These classes often co-occur with human feet, leading classifiers to associate high activations near the feet with these classes. This phenomenon makes them potential candidates for local views. This redundant information results in noisy pseudo-labels, reducing overall performance. Examples of such failure cases are provided in the supplementary materials. Developing an improved local view proposal method to better handle noisy inferences and focusing on the target objects will be an important direction for future work. + +# 5. Conclusion + +In this paper, we propose a novel unsupervised multi-label classification method, Classifier-guided CLIP Distillation (CCD). We address two critical issues in CLIP: inconsistent predictions due to subtle input changes and inherent prediction bias. To tackle these issues, we devised a classifier-guided label update and CLIP debiasing. Firstly, we identified the effectiveness of acquiring local views near class-related objects and proposed a classifier-guided label update. It selects local views utilizing the Class Activation Mapping (CAM) and acquires corresponding labels from CLIP. Our approach results in allocating more local views for complex images. Secondly, we address the CLIP's bias. We investigated the biased prediction distribution of CLIP and rectified the pseudo-labels. This step notably boosts performance in certain classes and reduces the impact of bias on CLIP. Through experiments on four benchmarks, CCD newly achieved state-of-the-art performances in unsupervised multi-label classification. Remarkably, CCD achieved comparable performance to fully supervised methods on PASCAL VOC 2012 and PASCAL VOC 2007 datasets without relying on annotations. + +Acknowledgements. This research was supported by the Basic Science Research Program through the National Research Foundation of Korea (NRF) funded by the MSIP (RS-2025-00520207, RS-2023-00219019), KEIT grant funded by the Korean government (MOTIE) (No. 2022-0-00680, No. 2022-0-01045), the IITP grant funded by the Korean government (MSIT) (No. RS-2024-00457882, National AI Research Lab Project, No. 2021-0-02068 Artificial Intelligence Innovation Hub, RS-2019-II190075 Artificial Intelligence Graduate School Program (KAIST)), Samsung Electronics Co., Ltd (IO230508-06190-01) and SAMSUNG Research, Samsung Electronics Co., Ltd. + +# References + +[1] Rabab Abdelfattah, Xin Zhang, Mostafa M Fouda, Xiaofeng Wang, and Song Wang. G2netpl: Generic game-theoretic network for partial-label image classification. arXiv preprint arXiv:2210.11469, 2022. 6, 7 +[2] Rabab Abdelfattah, Qing Guo, Xiaoguang Li, Xiaofeng Wang, and Song Wang. Cdul: Clip-driven unsupervised learning for multi-label image classification. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1348-1357, 2023. 1, 2, 3, 6, 7 +[3] Görkem Algan and Ülkay Ulusoy. Label noise types and their effects on deep learning, 2020. 5 +[4] Serhat Selcuk Bucak, Rong Jin, and Anil K. Jain. Multi-label learning with incomplete class assignments. In CVPR 2011, pages 2801-2808, 2011. 2 +[5] Tianshui Chen, Tao Pu, Hefeng Wu, Yuan Xie, and Liang Lin. Structured semantic transfer for multi-label recognition with partial labels. In Proceedings of the AAAI conference on artificial intelligence, pages 339-346, 2022. 1, 2, 6, 7 +[6] Tat-Seng Chua, Jinhui Tang, Richang Hong, Haojie Li, Zhiping Luo, and Yantao Zheng. Nus-wide: a real-world web image database from national university ofSingapore. In Proceedings of the ACM international conference on image and video retrieval, pages 1-9, 2009. 6 +[7] Elijah Cole, Oisin Mac Aodha, Titouan Lorieul, Pietro Perona, Dan Morris, and Nebojsa Jojic. Multi-label learning from single positive labels. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 933-942, 2021. 1, 2, 6, 7 +[8] Katherine Crowson, Stella Biderman, Daniel Kornis, Dashiell Stander, Eric Hallahan, Louis Castricato, and Edward Raff. Vqgan-clip: Open domain image generation and editing with natural language guidance. In European Conference on Computer Vision, pages 88–105. Springer, 2022. 2 +[9] Jia Deng, Olga Russakovsky, Jonathan Krause, Michael S Bernstein, Alex Berg, and Li Fei-Fei. Scalable multi-label annotation. In Proceedings of the SIGCHI Conference on Human Factors in Computing Systems, pages 3099–3102, 2014. 2 +[10] Thibaut Durand, Nazanin Mehrasa, and Greg Mori. Learning a deep convnet for multi-label classification with partial + +labels. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 647-657, 2019. 1, 2, 6, 7 +[11] M. Everingham, L. Van Gool, C. K. I. Williams, J. Winn, and A. Zisserman. The Pascal visual object classes (voc) challenge. International Journal of Computer Vision, 88(2): 303-338, 2010. 6 +[12] Xiuye Gu, Tsung-Yi Lin, Weicheng Kuo, and Yin Cui. Open-vocabulary object detection via vision and language knowledge distillation. arXiv preprint arXiv:2104.13921, 2021. 2 +[13] Ping Hu, Ximeng Sun, Stan Sclaroff, and Kate Saenko. Dualcoop++: Fast and effective adaptation to multi-label recognition with limited annotations. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 2 +[14] Dat Huynh and Ehsan Elhamifar. Interactive multi-label cnn learning with partial labels. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9423-9432, 2020. 1, 2 +[15] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International conference on machine learning, pages 4904-4916. PMLR, 2021. 2 +[16] Ashish Kapoor, Raajay Viswanathan, and Prateek Jain. Multilabel classification using bayesian compressed sensing. Advances in neural information processing systems, 25, 2012. 2 +[17] Apoorv Khandelwal, Luca Weihs, Roozbeh Mottaghi, and Aniruddha Kembhavi. Simple but effective: Clip embeddings for embodied ai. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14829-14838, 2022. 2 +[18] Dongseob Kim, Seungho Lee, Junsuk Choe, and Hyunjung Shim. Weakly supervised semantic segmentation for driving scenes. arXiv preprint arXiv:2312.13646, 2023. 2 +[19] Youngwook Kim, Jae Myung Kim, Zeynep Akata, and Jungwoo Lee. Large loss matters in weakly supervised multi-label classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14156-14165, 2022. 1, 2, 6, 7 +[20] Youngwook Kim, Jae Myung Kim, Jieun Jeong, Cordelia Schmid, Zeynep Akata, and Jungwoo Lee. Bridging the gap between model explanations in partially annotated multi-label classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3408-3417, 2023. 1, 2 +[21] Kaustav Kundu and Joseph Tighe. Exploiting weakly supervised visual patterns to learn from partial annotations. Advances in Neural Information Processing Systems, 33:561-572, 2020. 6, 7 +[22] Yue Liao, Aixi Zhang, Miao Lu, Yongliang Wang, Xiaobo Li, and Si Liu. Gen-vlkt: Simplify association and enhance interaction understanding for hoi detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20123-20132, 2022. 2 + +[23] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 6 +[24] Bing-Qing Liu, Bin-Bin Jia, and Min-Ling Zhang. Towards enabling binary decomposition for partial multi-label learning. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(11):13203-13217, 2023. 1, 2 +[25] Weiwei Liu and Ivor Tsang. On the optimality of classifier chain for multi-label classification. Advances in Neural Information Processing Systems, 28, 2015. 1 +[26] Oisin Mac Aodha, Elijah Cole, and Pietro Perona. Presence-only geographical priors for fine-grained image classification. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9596-9606, 2019. 2, 6, 7 +[27] Dhruv Mahajan, Ross Girshick, Vignesh Ramanathan, Kaiming He, Manohar Paluri, Yixuan Li, Ashwin Bharambe, and Laurens Van Der Maaten. Exploring the limits of weakly supervised pretraining. In Proceedings of the European conference on computer vision (ECCV), pages 181-196, 2018. 1 +[28] Antoine Miech, Jean-Baptiste Alayrac, Lucas Smaira, Ivan Laptev, Josef Sivic, and Andrew Zisserman. End-to-end learning of visual representations from uncurated instructional videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9879-9889, 2020. 2 +[29] Ishan Misra, C Lawrence Zitnick, Margaret Mitchell, and Ross Girshick. Seeing through the human reporting bias: Visual classifiers from noisy human-centric labels. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2930–2939, 2016. 1 +[30] Tao Pu, Tianshui Chen, Hefeng Wu, and Liang Lin. Semantic-aware representation blending for multi-label image recognition with partial labels. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2091-2098, 2022. 1, 2, 6, 7 +[31] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 2 +[32] Yongming Rao, Wenliang Zhao, Guangyi Chen, Yansong Tang, Zheng Zhu, Guan Huang, Jie Zhou, and Jiwen Lu. Denseclip: Language-guided dense prediction with context-aware prompting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18082-18091, 2022. 2 +[33] Tal Ridnik, Emanuel Ben-Baruch, Nadav Zamir, Asaf Noy, Itamar Friedman, Matan Protter, and Lihi Zelnik-Manor. Asymmetric loss for multi-label classification. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 82-91, 2021. 6, 7 + +[34] Amanpreet Singh, Ronghang Hu, Vedanuj Goswami, Guillaume Couairon, Wojciech Galuba, Marcus Rohrbach, and Douwe Kiela. Flava: A foundational language and vision alignment model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15638-15650, 2022. 2 +[35] GyeongRyeol Song, Noo-ri Kim, Jin-Seop Lee, and JeeHyong Lee. Ignore: Information gap-based false negative loss rejection for single positive multi-label learning. In European Conference on Computer Vision, pages 472–488. Springer, 2025. 1 +[36] Chen Sun, Abhinav Shrivastava, Saurabh Singh, and Abhinav Gupta. Revisiting unreasonable effectiveness of data in deep learning era. In Proceedings of the IEEE international conference on computer vision, pages 843-852, 2017. 1 +[37] Yu-Yin Sun, Yin Zhang, and Zhi-Hua Zhou. Multi-label learning with weak label. Proceedings of the AAAI Conference on Artificial Intelligence, 24(1):593-598, 2010. 2 +[38] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2818-2826, 2016. 2, 6, 7 +[39] Guy Tevet, Brian Gordon, Amir Hertz, Amit H Bermano, and Daniel Cohen-Or. Motionclip: Exposing human motion generation to clip space. In European Conference on Computer Vision, pages 358–374. Springer, 2022. 2 +[40] Grigorios Tsoumakas and Ioannis Katakis. Multi-label classification: An overview. International Journal of Data Warehousing and Mining (IJDWM), 3(3):1-13, 2007. 1 +[41] Deepak Vasisht, Andreas Damianou, Manik Varma, and Ashish Kapoor. Active learning for sparse bayesian multilabel classification. In Proceedings of the 20th ACM SIGKDD international conference on Knowledge discovery and data mining, pages 472-481, 2014. 2 +[42] Thomas Verelst, Paul K. Rubenstein, Marcin Eichner, Tinne Tuytelaars, and Maxim Berman. Spatial consistency loss for training multi-label classifiers from single-label annotations. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 3879-3889, 2023. 1, 2 +[43] Changhu Wang, Shuicheng Yan, Lei Zhang, and Hong-Jiang Zhang. Multi-label sparse coding for automatic image annotation. In 2009 IEEE conference on computer vision and pattern recognition, pages 1643-1650. IEEE, 2009. 1 +[44] Can Wang, Menglei Chai, Mingming He, Dongdong Chen, and Jing Liao. Clip-nerf: Text-and-image driven manipulation of neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3835-3844, 2022. 2 +[45] Qifan Wang, Bin Shen, Shumiao Wang, Liang Li, and Luo Si. Binary codes embedding for fast image tagging with incomplete labels. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part II 13, pages 425-439. Springer, 2014. 2 +[46] Baoyuan Wu, Siwei Lyu, and Bernard Ghanem. Ml-mg: Multi-label learning with missing labels using a mixed + +graph. In Proceedings of the IEEE international conference on computer vision, pages 4157-4165, 2015. 2 +[47] Xin Xing, Zhexiao Xiong, Abby Stylianou, Srikumar Sastry, Liyu Gong, and Nathan Jacobs. Vision-language pseudolabels for single-positive multi-label learning. arXiv preprint arXiv:2310.15985, 2023. 2 +[48] Alireza Zareian, Kevin Dela Rosa, Derek Hao Hu, and Shih-Fu Chang. Open-vocabulary object detection using captions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14393-14402, 2021. 2 +[49] Bolei Zhou, Aditya Khosla, Agata Lapedriza, Aude Oliva, and Antonio Torralba. Learning deep features for discriminative localization. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2921-2929, 2016. 2 +[50] Chong Zhou, Chen Change Loy, and Bo Dai. Extract free dense labels from clip. In European Conference on Computer Vision, pages 696-712. Springer, 2022. 2 +[51] Donghao Zhou, Pengfei Chen, Qiong Wang, Guangyong Chen, and Pheng-Ann Heng. Acknowledging the unknown for multi-label learning with single positive labels. In European Conference on Computer Vision, pages 423-440. Springer, 2022. 1, 2 \ No newline at end of file diff --git a/2025/Classifier-guided CLIP Distillation for Unsupervised Multi-label Classification/images.zip b/2025/Classifier-guided CLIP Distillation for Unsupervised Multi-label Classification/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..c8d69b7c90304b0c36c82f50c53fc4fb6fdcbf2e --- /dev/null +++ b/2025/Classifier-guided CLIP Distillation for Unsupervised Multi-label Classification/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:651cfd0ff8244331c998d8df673a271ae3315d3457cc85b253ad36edea68df45 +size 380414 diff --git a/2025/Classifier-guided CLIP Distillation for Unsupervised Multi-label Classification/layout.json b/2025/Classifier-guided CLIP Distillation for Unsupervised Multi-label Classification/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..bcce0eea57301b47de86d1acb70ea7617f95db71 --- /dev/null +++ b/2025/Classifier-guided CLIP Distillation for Unsupervised Multi-label Classification/layout.json @@ -0,0 +1,8393 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 64, + 103, + 547, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 103, + 547, + 121 + ], + "spans": [ + { + "bbox": [ + 64, + 103, + 547, + 121 + ], + "type": "text", + "content": "Classifier-guided CLIP Distillation for Unsupervised Multi-label Classification" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 125, + 144, + 323, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 125, + 144, + 323, + 171 + ], + "spans": [ + { + "bbox": [ + 125, + 144, + 323, + 171 + ], + "type": "text", + "content": "Dongseob Kim \nSamsung Electronics, Republic of Korea" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 171, + 174, + 277, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 171, + 174, + 277, + 185 + ], + "spans": [ + { + "bbox": [ + 171, + 174, + 277, + 185 + ], + "type": "text", + "content": "kou.kim@samsung.com" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 354, + 144, + 485, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 144, + 485, + 171 + ], + "spans": [ + { + "bbox": [ + 354, + 144, + 485, + 171 + ], + "type": "text", + "content": "Hyunjung Shim KAIST, Republic of Korea" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 362, + 174, + 473, + 184 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 362, + 174, + 473, + 184 + ], + "spans": [ + { + "bbox": [ + 362, + 174, + 473, + 184 + ], + "type": "text", + "content": "kateshim@kaist.ac.kr" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 152, + 213, + 200, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 213, + 200, + 225 + ], + "spans": [ + { + "bbox": [ + 152, + 213, + 200, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 238, + 296, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 238, + 296, + 441 + ], + "spans": [ + { + "bbox": [ + 55, + 238, + 296, + 441 + ], + "type": "text", + "content": "Multi-label classification is crucial for comprehensive image understanding, yet acquiring accurate annotations is challenging and costly. To address this, a recent study suggests exploiting unsupervised multi-label classification leveraging CLIP, a powerful vision-language model. Despite CLIP's proficiency, it suffers from view-dependent predictions and inherent bias, limiting its effectiveness. We propose a novel method that addresses these issues by leveraging multiple views near target objects, guided by Class Activation Mapping (CAM) of the classifier, and debiasing pseudo-labels derived from CLIP predictions. Our Classifier-guided CLIP Distillation (CCD) enables selecting multiple local views without extra labels and debiasing predictions to enhance classification performance. Experimental results validate our method's superiority over existing techniques across diverse datasets. The code is available at https://github.com/k0u-id/CCD." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 465, + 135, + 478 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 465, + 135, + 478 + ], + "spans": [ + { + "bbox": [ + 55, + 465, + 135, + 478 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 485, + 296, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 485, + 296, + 581 + ], + "spans": [ + { + "bbox": [ + 55, + 485, + 296, + 581 + ], + "type": "text", + "content": "Multi-label classification plays a pivotal role in identifying multiple classes within an image, serving as a foundational component in image understanding [25, 27, 29, 36, 40]. Its applications are widespread, including recommendation systems, surveillance, and image retrieval [43]. However, acquiring multi-label annotations poses challenges due to the difficulty and significant costs involved in meticulously annotating all relevant objects within an image." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "content": "To solve this problem, various label-efficient techniques have been proposed. A partial label setting [5, 10, 14, 20, 24, 30] annotates only a portion of the classes appearing in an image and considers the rest as unknown labels. In the extreme, single positive label settings [7, 19, 35, 42, 51] provide only one positive label per image, with all other labels assumed unknown. However, these methods still incur labeling costs. Besides, current studies simulate partial and single positive settings by randomly omitting labels from a fully labeled dataset, which may not accurately reflect the actual labeling practices. For example, humans tend to pri" + } + ] + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 316, + 212, + 373, + 269 + ], + "blocks": [ + { + "bbox": [ + 316, + 212, + 373, + 269 + ], + "lines": [ + { + "bbox": [ + 316, + 212, + 373, + 269 + ], + "spans": [ + { + "bbox": [ + 316, + 212, + 373, + 269 + ], + "type": "image", + "image_path": "81948ba10a0ea8b6dcd29b979461ede4a68b1358709dd6f4d8993e4a0a941df9.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 322, + 269, + 367, + 283 + ], + "lines": [ + { + "bbox": [ + 322, + 269, + 367, + 283 + ], + "spans": [ + { + "bbox": [ + 322, + 269, + 367, + 283 + ], + "type": "text", + "content": "H. 0.99 / P. 0.00 (a) original image" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 301, + 555, + 380 + ], + "lines": [ + { + "bbox": [ + 313, + 301, + 555, + 380 + ], + "spans": [ + { + "bbox": [ + 313, + 301, + 555, + 380 + ], + "type": "text", + "content": "Figure 1. CLIP prediction probability of horse (H) and person (P) corresponding to different input patches from single image: (a) The original image and its CLIP prediction. (b) The " + }, + { + "bbox": [ + 313, + 301, + 555, + 380 + ], + "type": "inline_equation", + "content": "1/4" + }, + { + "bbox": [ + 313, + 301, + 555, + 380 + ], + "type": "text", + "content": " sized patch and its CLIP prediction. (c) The " + }, + { + "bbox": [ + 313, + 301, + 555, + 380 + ], + "type": "inline_equation", + "content": "1/9" + }, + { + "bbox": [ + 313, + 301, + 555, + 380 + ], + "type": "text", + "content": " sized patch and its CLIP prediction. (d) The " + }, + { + "bbox": [ + 313, + 301, + 555, + 380 + ], + "type": "inline_equation", + "content": "1/16" + }, + { + "bbox": [ + 313, + 301, + 555, + 380 + ], + "type": "text", + "content": " sized patch and its CLIP prediction. The same colored box indicates each patch cropped from the corresponding box." + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 376, + 212, + 433, + 269 + ], + "blocks": [ + { + "bbox": [ + 376, + 212, + 433, + 269 + ], + "lines": [ + { + "bbox": [ + 376, + 212, + 433, + 269 + ], + "spans": [ + { + "bbox": [ + 376, + 212, + 433, + 269 + ], + "type": "image", + "image_path": "0745cb1f1b35809050fcfae8df1076a832f20e325b1129bda770d7aa8bcaa314.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 381, + 270, + 425, + 283 + ], + "lines": [ + { + "bbox": [ + 381, + 270, + 425, + 283 + ], + "spans": [ + { + "bbox": [ + 381, + 270, + 425, + 283 + ], + "type": "text", + "content": "H:0.96 /P:0.00 (b)1/4 patch" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 435, + 212, + 493, + 269 + ], + "blocks": [ + { + "bbox": [ + 435, + 212, + 493, + 269 + ], + "lines": [ + { + "bbox": [ + 435, + 212, + 493, + 269 + ], + "spans": [ + { + "bbox": [ + 435, + 212, + 493, + 269 + ], + "type": "image", + "image_path": "32655c6fe797e810dd6eefd3686ad278d2297679375e08c863db5d8139fc0794.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 443, + 269, + 487, + 283 + ], + "lines": [ + { + "bbox": [ + 443, + 269, + 487, + 283 + ], + "spans": [ + { + "bbox": [ + 443, + 269, + 487, + 283 + ], + "type": "text", + "content": "H:0.14/P:0.12 (c) 1/9" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 496, + 213, + 553, + 269 + ], + "blocks": [ + { + "bbox": [ + 496, + 213, + 553, + 269 + ], + "lines": [ + { + "bbox": [ + 496, + 213, + 553, + 269 + ], + "spans": [ + { + "bbox": [ + 496, + 213, + 553, + 269 + ], + "type": "image", + "image_path": "1c5bd6b89078ae350b37dd22545cf6fbae0052265c6e88ef062ccce7b84a042e.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 506, + 270, + 546, + 283 + ], + "lines": [ + { + "bbox": [ + 506, + 270, + 546, + 283 + ], + "spans": [ + { + "bbox": [ + 506, + 270, + 546, + 283 + ], + "type": "text", + "content": ":0.05/P:0.65 (d) 1/16 patch" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 401, + 455, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 401, + 455, + 413 + ], + "spans": [ + { + "bbox": [ + 313, + 401, + 455, + 413 + ], + "type": "text", + "content": "oritize labeling trivial objects first." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 414, + 555, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 414, + 555, + 569 + ], + "spans": [ + { + "bbox": [ + 313, + 414, + 555, + 569 + ], + "type": "text", + "content": "To overcome these limitations, we focus on an unsupervised multi-label classification problem leveraging a vision-language model. One such model is CLIP [31], trained on an extensive dataset of 400M {image, text} pairs from the web. As a result, this model excels at recognizing diverse visual attributes and object categories in a zero-shot manner. Most recently, CDUL [2] applied CLIP for unsupervised multi-label classification by using it as a pseudolabeling function. CDUL generates pseudo-labels for both global and local views to help identify less prominent objects, training a multi-label classifier with each view-label pair. However, CDUL overlooks the limitations of CLIP, which hinders its prediction performance." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "type": "text", + "content": "In this paper, we identified two reliability issues in CLIP, addressing them to fully harness CLIP's potential. Firstly, we demonstrate that CLIP's predictions are highly view-dependent, considerably impacting classification performance. In a toy experiment, we observe that changing the crop position and size of an image leads to notable changes in CLIP's predictions. For example, in Fig. 1, when the cropped area is slightly shifted or increased, the probabilities for the horse and person classes vary significantly. Based on these key observations, we use multiple local views near the target object to mitigate the uncertainty of CLIP's pseudo-labels caused by view changes. Specifically," + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 15 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 15 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 15 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 198, + 13, + 442, + 24 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 13, + 442, + 24 + ], + "spans": [ + { + "bbox": [ + 198, + 13, + 442, + 24 + ], + "type": "text", + "content": "Except for this watermark, it is identical to the accepted version;" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 179, + 25, + 462, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 25, + 462, + 37 + ], + "spans": [ + { + "bbox": [ + 179, + 25, + 462, + 37 + ], + "type": "text", + "content": "the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "4661" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 69, + 136, + 123 + ], + "blocks": [ + { + "bbox": [ + 56, + 69, + 136, + 123 + ], + "lines": [ + { + "bbox": [ + 56, + 69, + 136, + 123 + ], + "spans": [ + { + "bbox": [ + 56, + 69, + 136, + 123 + ], + "type": "image", + "image_path": "60088c0668175707771d18910ba604eaea5adee25790c5032fd813a2987f6cce.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 75, + 123, + 118, + 129 + ], + "lines": [ + { + "bbox": [ + 75, + 123, + 118, + 129 + ], + "spans": [ + { + "bbox": [ + 75, + 123, + 118, + 129 + ], + "type": "text", + "content": "T:0.35/C:0.29" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 136, + 70, + 216, + 123 + ], + "blocks": [ + { + "bbox": [ + 136, + 70, + 216, + 123 + ], + "lines": [ + { + "bbox": [ + 136, + 70, + 216, + 123 + ], + "spans": [ + { + "bbox": [ + 136, + 70, + 216, + 123 + ], + "type": "image", + "image_path": "37478862c79ef123fbaf49e2455cd93a54a16254b1cb9d166c4bfb19872fdb83.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 154, + 123, + 197, + 129 + ], + "lines": [ + { + "bbox": [ + 154, + 123, + 197, + 129 + ], + "spans": [ + { + "bbox": [ + 154, + 123, + 197, + 129 + ], + "type": "text", + "content": "P:0.50/C:0.21" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 217, + 70, + 295, + 123 + ], + "blocks": [ + { + "bbox": [ + 217, + 70, + 295, + 123 + ], + "lines": [ + { + "bbox": [ + 217, + 70, + 295, + 123 + ], + "spans": [ + { + "bbox": [ + 217, + 70, + 295, + 123 + ], + "type": "image", + "image_path": "1d11a6b5723afdc8abad5b59af6e729df59af67e13f64a139337163f67269144.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 243, + 123, + 266, + 129 + ], + "lines": [ + { + "bbox": [ + 243, + 123, + 266, + 129 + ], + "spans": [ + { + "bbox": [ + 243, + 123, + 266, + 129 + ], + "type": "text", + "content": "H:1.00" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 58, + 140, + 290, + 189 + ], + "blocks": [ + { + "bbox": [ + 70, + 130, + 117, + 137 + ], + "lines": [ + { + "bbox": [ + 70, + 130, + 117, + 137 + ], + "spans": [ + { + "bbox": [ + 70, + 130, + 117, + 137 + ], + "type": "text", + "content": "(a) tv monitor (T)" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 156, + 130, + 194, + 138 + ], + "lines": [ + { + "bbox": [ + 156, + 130, + 194, + 138 + ], + "spans": [ + { + "bbox": [ + 156, + 130, + 194, + 138 + ], + "type": "text", + "content": "(b) person (P)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 239, + 130, + 274, + 138 + ], + "lines": [ + { + "bbox": [ + 239, + 130, + 274, + 138 + ], + "spans": [ + { + "bbox": [ + 239, + 130, + 274, + 138 + ], + "type": "text", + "content": "(c) horse (H)" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 58, + 140, + 290, + 189 + ], + "lines": [ + { + "bbox": [ + 58, + 140, + 290, + 189 + ], + "spans": [ + { + "bbox": [ + 58, + 140, + 290, + 189 + ], + "type": "image", + "image_path": "9a42edc92de4dd5a8a7bfbc895adec65846fe2bb2607fa46e6d8eca6a855ffb7.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 114, + 189, + 237, + 197 + ], + "lines": [ + { + "bbox": [ + 114, + 189, + 237, + 197 + ], + "spans": [ + { + "bbox": [ + 114, + 189, + 237, + 197 + ], + "type": "text", + "content": "(d) CLIP probability distribution at VOC 2012" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 217, + 295, + 283 + ], + "lines": [ + { + "bbox": [ + 55, + 217, + 295, + 283 + ], + "spans": [ + { + "bbox": [ + 55, + 217, + 295, + 283 + ], + "type": "text", + "content": "Figure 2. Sample images showcasing CLIP bias. C indicates the probability of chair. (a) Top 1 probability of \"tv monitor\" image is " + }, + { + "bbox": [ + 55, + 217, + 295, + 283 + ], + "type": "inline_equation", + "content": "35\\%" + }, + { + "bbox": [ + 55, + 217, + 295, + 283 + ], + "type": "text", + "content": ". (b) Top 1 probability of a \"person\" image is " + }, + { + "bbox": [ + 55, + 217, + 295, + 283 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 55, + 217, + 295, + 283 + ], + "type": "text", + "content": ". (c) Top 1 probability of a \"horse\" image is " + }, + { + "bbox": [ + 55, + 217, + 295, + 283 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 55, + 217, + 295, + 283 + ], + "type": "text", + "content": ". (d) The mean class-wise probability of PASCAL VOC 2012. We can observe the class-wise prediction bias of CLIP from these results." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 304, + 295, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 304, + 295, + 376 + ], + "spans": [ + { + "bbox": [ + 55, + 304, + 295, + 376 + ], + "type": "text", + "content": "we acquire various local views extracted by the bounding box using the Class Activation Mapping (CAM) [49] method, where the CAM is computed from the trained classifier itself. Additionally, our method can draw more local views for hard samples whose prediction exceeds a certain threshold across multiple classes." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 376, + 295, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 376, + 295, + 685 + ], + "spans": [ + { + "bbox": [ + 56, + 376, + 295, + 685 + ], + "type": "text", + "content": "Another notable issue with CLIP is its inherent prediction bias, which tends to reduce the predicted probability of specific classes. Due to polysemy [31], the words representing each class can be ambiguous in text embedding space. As illustrated in Fig. 2 (a) and (b), despite the presence of a single class (tv monitor and person, respectively), CLIP assigns considerably low top-1 probabilities of around " + }, + { + "bbox": [ + 56, + 376, + 295, + 685 + ], + "type": "inline_equation", + "content": "35\\%" + }, + { + "bbox": [ + 56, + 376, + 295, + 685 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 56, + 376, + 295, + 685 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 56, + 376, + 295, + 685 + ], + "type": "text", + "content": ". Conversely, objects like horses exhibit notably higher probabilities, as depicted in Fig. 2 (c). Analyzing the average CLIP top-1 probabilities across classes, as shown in Fig. 2 (d), further highlights this bias, indicating a lower prediction for certain classes compared to others. The biased prediction is evident across various datasets such as MS COCO and NUSWIDE1. Since the proposed method relies on pseudo-labels generated by CLIP, its bias naturally transfers to the classifier. Consequently, the classifier shows a similar prediction bias and performance degradation. To mitigate this issue, we adjust all pseudo-labels by applying the inverse transform of the bias. While this adjustment can amplify unwanted noise in pseudo-labels, we apply consistency loss to handle this noise after debiasing. Our adjustment helps debiasing the classifier, thereby enhancing its generalization performance. As a result, our proposed Classifier-guided CLIP Distillation (CCD) achieves state-of-the-art performances on the PASCAL VOC 2012, PASCAL VOC 2007, COCO, and NUSWIDE benchmarks." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 314, + 71, + 400, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 71, + 400, + 83 + ], + "spans": [ + { + "bbox": [ + 314, + 71, + 400, + 83 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 91, + 523, + 103 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 91, + 523, + 103 + ], + "spans": [ + { + "bbox": [ + 313, + 91, + 523, + 103 + ], + "type": "text", + "content": "2.1. Label efficient multi-label classification." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 109, + 554, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 109, + 554, + 205 + ], + "spans": [ + { + "bbox": [ + 313, + 109, + 554, + 205 + ], + "type": "text", + "content": "Unlike multi-class classification, where a single class is identified in an image, multi-label classification requires accurately identifying all classes present in an image. Labeling every class in an image is significantly more challenging and resource-intensive than identifying just one class. To address this issue, a partial label setting has been proposed [5, 10, 14, 20, 24, 30], which utilizes only a subset of labels associated with each image." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 205, + 554, + 349 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 205, + 554, + 349 + ], + "spans": [ + { + "bbox": [ + 313, + 205, + 554, + 349 + ], + "type": "text", + "content": "Early approaches assumed any unannotated labels (i.e. \"unknown labels\") as negative labels [4, 37, 45]. These methods involved correcting these labels during the learning process. However, approaches that relied on probabilistic models [9, 16, 41, 46] for label correction were dependent on optimization techniques, resulting in limited scalability. Subsequently, alternative methods have been proposed, such as leveraging label-to-label similarity [14], cross-image semantic correlation [5], or category-specific representation [30] to refine labels. Some approaches suggest using soft-pseudo-labels instead of unknown labels [10, 26, 38]." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 350, + 554, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 350, + 554, + 409 + ], + "spans": [ + { + "bbox": [ + 313, + 350, + 554, + 409 + ], + "type": "text", + "content": "Research has extended to a more extreme scenario: single positive label setting [7, 19, 42, 51], where only a single positive label is provided per image. These methods build upon those proposed in the partial label setting and adapt them to function effectively in more challenging contexts." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 410, + 554, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 410, + 554, + 482 + ], + "spans": [ + { + "bbox": [ + 313, + 410, + 554, + 482 + ], + "type": "text", + "content": "More recently, CDUL [2] has addressed the unsupervised multi-label classification problem by utilizing CLIP. In a similar vein, our method tackles the unsupervised multi-label classification problem using CLIP while additionally providing local area information with CAM and addressing bias in CLIP predictions." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 491, + 471, + 504 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 491, + 471, + 504 + ], + "spans": [ + { + "bbox": [ + 313, + 491, + 471, + 504 + ], + "type": "text", + "content": "2.2. Vision-language pretraining." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 509, + 554, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 509, + 554, + 629 + ], + "spans": [ + { + "bbox": [ + 313, + 509, + 554, + 629 + ], + "type": "text", + "content": "Vision-language pretraining (VLP) has emerged as a pivotal approach for bridging the gap between visual and textual modalities [15, 28, 34]. Among the pioneering VLPs, CLIP [31] stands out as the most commonly utilized framework. CLIP was trained on large {image, text} pairs sourced from the Internet. Recently, several studies have adopted VLP to provide rich visual-semantic knowledge for downstream image tasks, such as detection [12, 48], segmentation [18, 32, 50], human-object interation [17, 22], and generation [8, 39, 44]." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "text", + "content": "In multi-label classification, various methods have emerged to derive pseudo-labels from CLIP predictions [2, 13, 47]. Among these, CDUL [2] integrates local inferences with global inference, successfully expanding the utilization of the VLP to multi-label classification. However, CDUL fails to tackle two limitations inherent in CLIP: sensitivity to input view selection and inherent prediction bias." + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "type": "text", + "content": "1CLIP probability distribution of other datasets are in the supplementary material." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4662" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 70, + 116, + 127 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 116, + 127 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 116, + 127 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 116, + 127 + ], + "type": "image", + "image_path": "c3877900e645a0a43a464b410debe40e899b33c285104aee0fbbedf6ab0a7b39.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 78, + 129, + 95, + 135 + ], + "lines": [ + { + "bbox": [ + 78, + 129, + 95, + 135 + ], + "spans": [ + { + "bbox": [ + 78, + 129, + 95, + 135 + ], + "type": "text", + "content": "87.6%" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 62, + 137, + 110, + 144 + ], + "lines": [ + { + "bbox": [ + 62, + 137, + 110, + 144 + ], + "spans": [ + { + "bbox": [ + 62, + 137, + 110, + 144 + ], + "type": "text", + "content": "(a) Around GT box" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 117, + 70, + 176, + 127 + ], + "blocks": [ + { + "bbox": [ + 117, + 70, + 176, + 127 + ], + "lines": [ + { + "bbox": [ + 117, + 70, + 176, + 127 + ], + "spans": [ + { + "bbox": [ + 117, + 70, + 176, + 127 + ], + "type": "image", + "image_path": "06361d85478502a965ba0baff7c63cc13ee8f4c4ae428e933ddc4c89de38f378.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 137, + 129, + 155, + 135 + ], + "lines": [ + { + "bbox": [ + 137, + 129, + 155, + 135 + ], + "spans": [ + { + "bbox": [ + 137, + 129, + 155, + 135 + ], + "type": "text", + "content": "87.1%" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 132, + 137, + 160, + 144 + ], + "lines": [ + { + "bbox": [ + 132, + 137, + 160, + 144 + ], + "spans": [ + { + "bbox": [ + 132, + 137, + 160, + 144 + ], + "type": "text", + "content": "(b) GT box" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 177, + 70, + 235, + 128 + ], + "blocks": [ + { + "bbox": [ + 177, + 70, + 235, + 128 + ], + "lines": [ + { + "bbox": [ + 177, + 70, + 235, + 128 + ], + "spans": [ + { + "bbox": [ + 177, + 70, + 235, + 128 + ], + "type": "image", + "image_path": "1ea59459653e8aad6e275be5e67fdc3ad76b7dd141c77be1c08da820d070d3bd.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 198, + 129, + 214, + 135 + ], + "lines": [ + { + "bbox": [ + 198, + 129, + 214, + 135 + ], + "spans": [ + { + "bbox": [ + 198, + 129, + 214, + 135 + ], + "type": "text", + "content": "87.1%" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 185, + 137, + 226, + 144 + ], + "lines": [ + { + "bbox": [ + 185, + 137, + 226, + 144 + ], + "spans": [ + { + "bbox": [ + 185, + 137, + 226, + 144 + ], + "type": "text", + "content": "(c) Random box" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 164, + 296, + 263 + ], + "lines": [ + { + "bbox": [ + 55, + 164, + 296, + 263 + ], + "spans": [ + { + "bbox": [ + 55, + 164, + 296, + 263 + ], + "type": "text", + "content": "Figure 3. The proof-of-concept study of local view selection method. We train the classifier with pseudo-label generated from four different local views: (a) Around GT boxes, (b) GT boxes, (c) Random boxes, (d) Uniform grid boxes. The numbers below each sample are the performance (mAP) of the classifier trained with the corresponding pseudo-label for the entire training set. The classifier trained with local view around GT boxes achieved " + }, + { + "bbox": [ + 55, + 164, + 296, + 263 + ], + "type": "inline_equation", + "content": "1.3\\%" + }, + { + "bbox": [ + 55, + 164, + 296, + 263 + ], + "type": "text", + "content": " higher performance compared to classifier trained with uniform-grid local view." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 235, + 70, + 295, + 128 + ], + "blocks": [ + { + "bbox": [ + 235, + 70, + 295, + 128 + ], + "lines": [ + { + "bbox": [ + 235, + 70, + 295, + 128 + ], + "spans": [ + { + "bbox": [ + 235, + 70, + 295, + 128 + ], + "type": "image", + "image_path": "0071ec44c4ecff1e6e0a54d1b0d10480d757f919bba0a8102ced51a68d7d2769.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 257, + 129, + 274, + 135 + ], + "lines": [ + { + "bbox": [ + 257, + 129, + 274, + 135 + ], + "spans": [ + { + "bbox": [ + 257, + 129, + 274, + 135 + ], + "type": "text", + "content": "86.4%" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 239, + 137, + 291, + 144 + ], + "lines": [ + { + "bbox": [ + 239, + 137, + 291, + 144 + ], + "spans": [ + { + "bbox": [ + 239, + 137, + 291, + 144 + ], + "type": "text", + "content": "(d) Uniform grid box" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 285, + 296, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 285, + 296, + 334 + ], + "spans": [ + { + "bbox": [ + 55, + 285, + 296, + 334 + ], + "type": "text", + "content": "Consequently, the pseudo-labels generated by CDUL exhibit noise and bias. In contrast, our approach leverages a fine-tuned classifier to suggest local views and addresses CLIP bias through calibration." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 346, + 111, + 358 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 346, + 111, + 358 + ], + "spans": [ + { + "bbox": [ + 55, + 346, + 111, + 358 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 366, + 227, + 380 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 366, + 227, + 380 + ], + "spans": [ + { + "bbox": [ + 55, + 366, + 227, + 380 + ], + "type": "text", + "content": "3.1. Generating initial pseudo-labels" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 55, + 384, + 296, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 384, + 296, + 552 + ], + "spans": [ + { + "bbox": [ + 55, + 384, + 296, + 552 + ], + "type": "text", + "content": "CLIP, a powerful vision-language model, has been trained to align images with corresponding textual descriptions. Leveraging the powerful zero-shot capability of CLIP, we initially generate pseudo-labels similar to CDUL. For each image in the dataset, we extract its image embedding using the CLIP image encoder. Simultaneously, we obtain the text embedding by inputting the class name associated with each dataset and a fixed prompt (e.g. \"a photo of the [class name]\") into the text encoder. Next, we compute the similarity between the image embedding and the class-specific text embedding. This similarity score is then translated to compute softmax probabilities. These probabilities serve as the initial pseudo-labels. Mathematically, the similarity score for each class can be expressed as:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 112, + 563, + 295, + 590 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 112, + 563, + 295, + 590 + ], + "spans": [ + { + "bbox": [ + 112, + 563, + 295, + 590 + ], + "type": "interline_equation", + "content": "s _ {c} = \\frac {f ^ {\\top} w _ {c}}{\\| f \\| \\cdot \\| w _ {c} \\|}, 1 \\leq c \\leq C \\tag {1}", + "image_path": "075f11fdb144e84fadf48f430e0f76165cbaca54b1637e9547e6526fbb4e8b69.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 55, + 596, + 296, + 645 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 596, + 296, + 645 + ], + "spans": [ + { + "bbox": [ + 55, + 596, + 296, + 645 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 596, + 296, + 645 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 55, + 596, + 296, + 645 + ], + "type": "text", + "content": " is a CLIP image embedding, " + }, + { + "bbox": [ + 55, + 596, + 296, + 645 + ], + "type": "inline_equation", + "content": "w_{c}" + }, + { + "bbox": [ + 55, + 596, + 296, + 645 + ], + "type": "text", + "content": " is a class-specific CLIP text embedding and " + }, + { + "bbox": [ + 55, + 596, + 296, + 645 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 596, + 296, + 645 + ], + "type": "text", + "content": " indicates a class index. With class-wise similarity, we can calculate a class-wise softmax probability " + }, + { + "bbox": [ + 55, + 596, + 296, + 645 + ], + "type": "inline_equation", + "content": "p_{c}" + }, + { + "bbox": [ + 55, + 596, + 296, + 645 + ], + "type": "text", + "content": " of each image as follows:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 124, + 654, + 295, + 684 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 124, + 654, + 295, + 684 + ], + "spans": [ + { + "bbox": [ + 124, + 654, + 295, + 684 + ], + "type": "interline_equation", + "content": "p _ {c} = \\frac {\\exp \\left(s _ {c} / \\tau\\right)}{\\sum_ {c = 1} ^ {C} \\exp \\left(s _ {c} / \\tau\\right)}, \\tag {2}", + "image_path": "2816ce073c9de54b8269b16a7fc1cf4f80cf2aaa1416d961139dea8e65d2bba1.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": " is a CLIP temperature parameter. We concatenate each class-wise softmax probability to generate initial" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "content": "pseudo-labels. Finally, the initial pseudo-label for each image is defined as:" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 349, + 104, + 553, + 125 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 349, + 104, + 553, + 125 + ], + "spans": [ + { + "bbox": [ + 349, + 104, + 553, + 125 + ], + "type": "interline_equation", + "content": "l ^ {\\text {i n i t i a l}} = \\left\\{p _ {1} ^ {\\text {g l o b a l}}, p _ {2} ^ {\\text {g l o b a l}}, \\dots , p _ {c} ^ {\\text {g l o b a l}} \\right\\}. \\tag {3}", + "image_path": "4092d502e5e67ab38f026459156ec8b8fd79494476a329d468777fabe36dd709.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 126, + 555, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 126, + 555, + 281 + ], + "spans": [ + { + "bbox": [ + 313, + 126, + 555, + 281 + ], + "type": "text", + "content": "Following the acquisition of softmax probabilities, the initial pseudo-labels are used for the warm-up training of the classifier. Since the softmax calculation inherently normalizes the probabilities to ensure a sum of 1, it does not fit the multi-label classification problem where multiple classes have the same high probability (e.g. 1). Nevertheless, after the classifier undergoes warm-up training, it can predict each class independently, leading to recognizing multiple classes present in an image, albeit with limited accuracy. To address this performance limitation, in the subsequent section, we introduce a method to construct {training data, pseudo-label} pairs tailored for the multi-label classification problem, leveraging a warm-up trained classifier." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 288, + 522, + 300 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 288, + 522, + 300 + ], + "spans": [ + { + "bbox": [ + 313, + 288, + 522, + 300 + ], + "type": "text", + "content": "3.2. Updating classifier-guided pseudo-label" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 304, + 555, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 304, + 555, + 483 + ], + "spans": [ + { + "bbox": [ + 313, + 304, + 555, + 483 + ], + "type": "text", + "content": "Prior research [2] has shown that CLIP's inference results on the entire images may be suboptimal for multi-label classification. CLIP tends to assign high probabilities primarily to prominent objects in the input image. As a result, it does not effectively distribute probabilities across all existing objects. CDUL addresses this limitation by dividing the image into uniform patches in a grid fashion. Despite its simplicity, employing a uniform patch-based approach for local view acquisition may lead to the inference of background batch, where the patch does not contain any class object. More importantly, treating both simple images (i.e. those containing a single object) and complex images (i.e. those containing multiple objects) equally with the same number of local views is clearly inefficient in optimizing the resource (i.e. CLIP inference)." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 484, + 554, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 484, + 554, + 663 + ], + "spans": [ + { + "bbox": [ + 313, + 484, + 554, + 663 + ], + "type": "text", + "content": "Proof-of-concept study. We perform a proof-of-concept study to evaluate the impact of using object-specific locations in local view proposals. Our hypothesis is that incorporating class-related object locations enhances performance. We extract local views in four different scenarios, using ground truth bounding boxes (i.e., the bounding box that tightly covers the target object): (1) random cropping around the bounding box, (2) using only the bounding box itself" + }, + { + "bbox": [ + 313, + 484, + 554, + 663 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 313, + 484, + 554, + 663 + ], + "type": "text", + "content": ", (3) random cropping without guidance from the bounding box, and (4) uniform grid-based cropping. Each scenario produces nine patches per image. Subsequently, we infer pseudo-labels for these patches and train with each " + }, + { + "bbox": [ + 313, + 484, + 554, + 663 + ], + "type": "inline_equation", + "content": "\\{\\text{patch}, \\text{pseudo-label}\\}" + }, + { + "bbox": [ + 313, + 484, + 554, + 663 + ], + "type": "text", + "content": " pair. As shown in Fig. 3, our results indicate that the first scenario-(a) random cropping around the bounding box-achieves the best performance." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 664, + 553, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 664, + 553, + 688 + ], + "spans": [ + { + "bbox": [ + 313, + 664, + 553, + 688 + ], + "type": "text", + "content": "Our proof-of-concept study reveals two critical insights: (1) the local view generation policy significantly influences" + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 313, + 693, + 555, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 693, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 693, + 555, + 712 + ], + "type": "text", + "content": "2For images with fewer than nine boxes, we additionally infer their resized versions." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4663" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 70, + 553, + 289 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 553, + 289 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 553, + 289 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 553, + 289 + ], + "type": "image", + "image_path": "bc0ad15f1147ebf3b38b7c660904aba36e8ff3ea63bff11ae9ad98cdaae849bf.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 307, + 555, + 374 + ], + "lines": [ + { + "bbox": [ + 54, + 307, + 555, + 374 + ], + "spans": [ + { + "bbox": [ + 54, + 307, + 555, + 374 + ], + "type": "text", + "content": "Figure 4. The overview of label preparation. (a) We calculate the cosine similarity between text embeddings and image embeddings. The softmax probability of these similarities is the CLIP prediction of each image. The top 1 probabilities are highlighted in green boxes. By class-wise averaging the top-1 probabilities, the CLIP bias is derived (yellow box). Pseudo-labels are then generated by debiasing CLIP predictions. (b) For label updating, classes above a threshold are selected from the classifier output, and local views corresponding to these classes are extracted. The process for acquiring local labels for each patch mirrors the initial label acquisition. The final pseudo-label is obtained through a weighted sum of initial pseudo-labels and local labels." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 395, + 294, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 395, + 294, + 418 + ], + "spans": [ + { + "bbox": [ + 54, + 395, + 294, + 418 + ], + "type": "text", + "content": "CLIP's predictions, and (2) framing the view around the target object greatly improves classification accuracy." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 419, + 295, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 419, + 295, + 514 + ], + "spans": [ + { + "bbox": [ + 54, + 419, + 295, + 514 + ], + "type": "text", + "content": "CAM-based view-selection. Motivated by the above two findings, we decided to approximate the first scenario in our approach. We apply Class Activation Mapping (CAM) with the trained classifier to identify vital regions within the image. This is called a CAM-based view selection method. This method generates {training image, pseudo-label} pairs by focusing on the object periphery, without relying on any ground-truth annotations." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 516, + 295, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 516, + 295, + 588 + ], + "spans": [ + { + "bbox": [ + 55, + 516, + 295, + 588 + ], + "type": "text", + "content": "By employing the trained classifier, CAM visualizes activations that most significantly influence class prediction. CAM calculation is specifically based on the weights assigned to each channel of the feature map, which influence the classification score. This can be represented by the following formula:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 121, + 600, + 295, + 634 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 600, + 295, + 634 + ], + "spans": [ + { + "bbox": [ + 121, + 600, + 295, + 634 + ], + "type": "interline_equation", + "content": "G _ {c} (x) = \\sum_ {i = 1} ^ {Q} \\mathbf {w} _ {i} ^ {c \\top} \\cdot g _ {i} (x), \\tag {4}", + "image_path": "414f5bb0936418f2a23e4dd589b332b271d00569a1a6a47c6cf0c7656db1cdc4.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 641, + 295, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 295, + 677 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 641, + 295, + 677 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 641, + 295, + 677 + ], + "type": "text", + "content": " is input image, " + }, + { + "bbox": [ + 55, + 641, + 295, + 677 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 55, + 641, + 295, + 677 + ], + "type": "text", + "content": " is CNN network, " + }, + { + "bbox": [ + 55, + 641, + 295, + 677 + ], + "type": "inline_equation", + "content": "g(x)" + }, + { + "bbox": [ + 55, + 641, + 295, + 677 + ], + "type": "text", + "content": " is the feature map of the CNN, " + }, + { + "bbox": [ + 55, + 641, + 295, + 677 + ], + "type": "inline_equation", + "content": "\\mathbf{w}^c" + }, + { + "bbox": [ + 55, + 641, + 295, + 677 + ], + "type": "text", + "content": " is class-wise weight of classifier, and " + }, + { + "bbox": [ + 55, + 641, + 295, + 677 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 55, + 641, + 295, + 677 + ], + "type": "text", + "content": " indicates the number of channels in the feature map." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": "We first normalize the activation map to a range between 0 and 1, then apply a threshold to identify highly activated regions. Next, we generate candidate bounding boxes with" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 395, + 555, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 395, + 555, + 504 + ], + "spans": [ + { + "bbox": [ + 313, + 395, + 555, + 504 + ], + "type": "text", + "content": "slight perturbation offsets. These bounding boxes define local views of the image. The regions highlighted by CAM act as reasonable approximations of the object's location, allowing us to extract local views without additional labeling costs. These localized view images are fed into CLIP to obtain softmax probabilities for the patches. The probabilities of local views are computed using the same method as for the global view, and then generating local labels that are focused on each local region." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "text", + "content": "The number of local views. Additionally, we adjust the number of local views based on image complexity. For simpler images, the classifier typically provides confident, high-probability predictions. This translates to high logits for well-trained classes and near-zero for others. Conversely, complex images often produce uncertain predictions, characterized by higher logits distributed broadly across multiple classes. To account for this, we conduct additional local view sampling when the classifier output of a certain class surpasses a predefined threshold. Specifically, we crop the bounding box of any class exceeding this threshold value and obtain its local probability. Through our experiments, we empirically determined a fixed threshold that ensures appropriate average number of classes exceeding this threshold. Details on this ablation study can be found in the experiments section. This approach allocates fewer local views to simple images and more local views to" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4664" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 269, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 269, + 83 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 269, + 83 + ], + "type": "text", + "content": "complex ones, optimizing the local view assignment." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 84, + 296, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 84, + 296, + 156 + ], + "spans": [ + { + "bbox": [ + 55, + 84, + 296, + 156 + ], + "type": "text", + "content": "To obtain a single local pseudo-label for each input image, we select the maximum value of class-wise probability across all local patches from the corresponding image. Denoting the probability of " + }, + { + "bbox": [ + 55, + 84, + 296, + 156 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 84, + 296, + 156 + ], + "type": "text", + "content": "-th patch for class " + }, + { + "bbox": [ + 55, + 84, + 296, + 156 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 84, + 296, + 156 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 55, + 84, + 296, + 156 + ], + "type": "inline_equation", + "content": "p_{i,c}^{patch}" + }, + { + "bbox": [ + 55, + 84, + 296, + 156 + ], + "type": "text", + "content": ", we conduct patch-wise aggregation using the maximum value as follows:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 98, + 166, + 294, + 186 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 98, + 166, + 294, + 186 + ], + "spans": [ + { + "bbox": [ + 98, + 166, + 294, + 186 + ], + "type": "interline_equation", + "content": "p _ {c} ^ {\\text {l o c a l}} = \\max _ {i = 1, \\dots , N} p _ {i, c} ^ {\\text {p a t c h}}, 1 \\leq c \\leq C \\tag {5}", + "image_path": "4cb9eeea6a8697cbde9541a11aafd808a8147606697a866145724ae0f69e8d42.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 190, + 296, + 237 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 190, + 296, + 237 + ], + "spans": [ + { + "bbox": [ + 55, + 190, + 296, + 237 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 190, + 296, + 237 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 190, + 296, + 237 + ], + "type": "text", + "content": " represents the number of local patches corresponding to each image, dependent on the input image itself. We then aggregate the local class-wise probability for generating a local pseudo-label as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 100, + 247, + 295, + 262 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 247, + 295, + 262 + ], + "spans": [ + { + "bbox": [ + 100, + 247, + 295, + 262 + ], + "type": "interline_equation", + "content": "l ^ {\\text {l o c a l}} = \\left\\{p _ {1} ^ {\\text {l o c a l}}, p _ {2} ^ {\\text {l o c a l}}, \\dots , p _ {c} ^ {\\text {l o c a l}} \\right\\}. \\tag {6}", + "image_path": "a3962db1e2a26404d8dc8158841fd80a24b83f165cb58a166422f13165c3db73.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 266, + 295, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 266, + 295, + 300 + ], + "spans": [ + { + "bbox": [ + 55, + 266, + 295, + 300 + ], + "type": "text", + "content": "We update the pseudo-label by the weighted sum of initial and local pseudo-labels. The final pseudo-label is calculated as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 312, + 295, + 326 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 312, + 295, + 326 + ], + "spans": [ + { + "bbox": [ + 105, + 312, + 295, + 326 + ], + "type": "interline_equation", + "content": "l ^ {\\text {f i n a l}} = \\alpha l ^ {\\text {i n i t i a l}} + (1 - \\alpha) l ^ {\\text {l o c a l}}, \\tag {7}", + "image_path": "d77a763d017319ad2f834093f5f9d1e7d92871a4a45739877f28f9dd14d375b5.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 331, + 296, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 331, + 296, + 392 + ], + "spans": [ + { + "bbox": [ + 55, + 331, + 296, + 392 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 331, + 296, + 392 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 331, + 296, + 392 + ], + "type": "text", + "content": " is a hyperparameter and we used 0.4. Through this label update, the pseudo-label efficiently incorporates information about less salient objects in multi-label classification. The updated pseudo-label is utilized throughout the remainder of the training process." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 396, + 216, + 410 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 396, + 216, + 410 + ], + "spans": [ + { + "bbox": [ + 55, + 396, + 216, + 410 + ], + "type": "text", + "content": "3.3. Debiasing CLIP pseudo-label" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 415, + 295, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 415, + 295, + 545 + ], + "spans": [ + { + "bbox": [ + 55, + 415, + 295, + 545 + ], + "type": "text", + "content": "Earlier, we aimed to identify highly activated areas within images through Class Activation Mapping and to infer additional label information focused on these areas. However, a challenge arises when using pseudo-labels generated by CLIP due to its inherent prediction bias. Consequently, classifiers trained using these pseudo-labels may inadvertently inherit this bias. Previous work [3] reveals that while classifiers are robust at overcoming random noise, they cannot rectify consistent biases associated with specific classes. Therefore, we need to overcome the inherent prediction bias of CLIP to enhance the performance." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 546, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 546, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 546, + 296, + 714 + ], + "type": "text", + "content": "To understand this bias, given CLIP's focus on salient objects, we treat global CLIP prediction as a single-label classification task. In a single-label classification, the predicted probability distribution for each class aligns with the class-wise average of the highest probabilities from each prediction. Following this principle, we compute the average prediction probability for each class, defining it as CLIP's prediction bias. Specifically, we compute CLIP bias as follows: (1) Obtain initial pseudo-labels by running each image in the \"target training set\" through CLIP; (2) Record top-1 probabilities for samples below the entropy threshold to constrain the task to single-label classification; (3) Count each class occurrences only if it achieves the top 1 probability; (4)Normalize each class's softmax probability by its" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 316, + 76, + 553, + 205 + ], + "blocks": [ + { + "bbox": [ + 316, + 76, + 553, + 205 + ], + "lines": [ + { + "bbox": [ + 316, + 76, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 316, + 76, + 553, + 205 + ], + "type": "image", + "image_path": "16611be1bb536704fa59c68cc0e2ac6a7a41dbdcba9a6e10c325bb5a06f6ac4d.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 212, + 555, + 289 + ], + "lines": [ + { + "bbox": [ + 313, + 212, + 555, + 289 + ], + "spans": [ + { + "bbox": [ + 313, + 212, + 555, + 289 + ], + "type": "text", + "content": "Figure 5. The training process of our method. We train the classifier using cross-entropy targeting initial pseudo-label during the warm-up phase which is illustrated as the green line. After the classifier-guided label update, we train the classifier using the cross-entropy loss targeting updated pseudo-label and the cross-entropy loss between the logits of differently augmented inputs which are illustrated as the green line and the blue dashed line." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 312, + 553, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 312, + 553, + 348 + ], + "spans": [ + { + "bbox": [ + 313, + 312, + 553, + 348 + ], + "type": "text", + "content": "occurrence count, defining this normalized probability as CLIP bias. This process does not require any ground-truth annotation, solely relying on the result of CLIP prediction." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 350, + 555, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 350, + 555, + 529 + ], + "spans": [ + { + "bbox": [ + 313, + 350, + 555, + 529 + ], + "type": "text", + "content": "The bias obtained in this manner serves as a basis for calibrating CLIP-derived probabilities. Specifically, we calibrate by multiplying each class's softmax probability from CLIP by the inverse of the CLIP bias. This calibration process enables the classifier to learn a rectified probability distribution. Our debiasing technique leads to a noticeable improvement in the performance of classes previously challenged due to the inherent CLIP bias (see Tab. 3). Furthermore, we adjust excessively small probabilities to prevent the minimum probability value from approaching zero, aiding in learning the corrected distribution. In doing so, we guide the classifier as a whole to learn from this inverse distribution. This calibration is applied not only to the global probability but also to local inference, minimizing the influence of the CLIP bias on the classifier as much as possible." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 539, + 413, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 539, + 413, + 552 + ], + "spans": [ + { + "bbox": [ + 313, + 539, + 413, + 552 + ], + "type": "text", + "content": "3.4. Overall training" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 558, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 555, + 714 + ], + "type": "text", + "content": "Consistency loss. The proposed method tackles noisy data by utilizing pseudo-labels as target labels for all {training data, pseudo-label} pairs. Since pseudo-labels inherently carry noise and our debiasing process can potentially amplify this noise, we employ consistency loss between the data and its augmented data as additional critics. This approach is inspired by the common practice in a semi-supervised regime, improving feature extractors for handling noisy data. In our framework, we introduce consistency loss alongside the conventional binary cross-entropy loss during training. Consistency loss involves applying different augmentation intensities to the same image. Also, it passes each augmented version through the classifier to en" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4665" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "content": "sure consistent predictions. In Fig. 5, we employ weak augmentations, such as resizing and flipping, and strong augmentations, which comprise weak augmentation and additional color jittering." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 121, + 295, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 121, + 295, + 301 + ], + "spans": [ + { + "bbox": [ + 55, + 121, + 295, + 301 + ], + "type": "text", + "content": "Training process. Initially, the entire training dataset undergoes inference through CLIP, followed by calculating the CLIP bias. Subsequently, we derive a global label by debiasing the global probability, utilizing it as an initial pseudolabel in Fig. 4 (a). The classifier is then warmed up with this pseudo-label as the target (the green line in Fig. 5). Following warm-up training, we update the pseudo-label with classifier-guided local inference. During this phase, the obtained local probabilities undergoes debiasing. Then they are combined to derive a local label, which then updates the existing pseudo-label as depicted in Fig. 4 (b). The classifier is subsequently trained with the updated pseudo-label for the remaining epochs (the green line and the blue dashed line in Fig. 5). For whole trainig, our loss function combines two terms:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 121, + 316, + 295, + 330 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 316, + 295, + 330 + ], + "spans": [ + { + "bbox": [ + 121, + 316, + 295, + 330 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {t o t a l}} = L _ {c e} + \\beta L _ {\\text {c o n s i s t}}, \\tag {8}", + "image_path": "a811adeac0c3313504bf0a820478e1f4217f26c8dd573ba51bf20418801bfbd3.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 337, + 295, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 337, + 295, + 373 + ], + "spans": [ + { + "bbox": [ + 55, + 337, + 295, + 373 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 337, + 295, + 373 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{ce}" + }, + { + "bbox": [ + 55, + 337, + 295, + 373 + ], + "type": "text", + "content": " is cross-entropy loss and " + }, + { + "bbox": [ + 55, + 337, + 295, + 373 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\text{consist}}" + }, + { + "bbox": [ + 55, + 337, + 295, + 373 + ], + "type": "text", + "content": " is consistency loss. " + }, + { + "bbox": [ + 55, + 337, + 295, + 373 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 55, + 337, + 295, + 373 + ], + "type": "text", + "content": " is a hyperparameter and we fix it as 0 during warm-up phase and 1 after." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 387, + 137, + 400 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 387, + 137, + 400 + ], + "spans": [ + { + "bbox": [ + 55, + 387, + 137, + 400 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 407, + 170, + 420 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 407, + 170, + 420 + ], + "spans": [ + { + "bbox": [ + 55, + 407, + 170, + 420 + ], + "type": "text", + "content": "4.1. Experimental setup" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 426, + 295, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 426, + 295, + 616 + ], + "spans": [ + { + "bbox": [ + 55, + 426, + 295, + 616 + ], + "type": "text", + "content": "Dataset & evaluation Metric. We conducted evaluations on four distinct datasets. The PASCAL VOC 2012 [11] dataset comprises 20 classes, with 5,717 images in the training set and 5,823 in the validation set. Similarly, the PASCAL VOC 2007 dataset also encompasses 20 classes, with a combined total of 5,011 training and validation images and 4,952 test images. MS COCO [23] consists of 80 classes, encompassing 82,081 training images and 40,137 validation images. Additionally, the NUSWIDE [6] dataset comprises 81 classes, with a training set of 150,000 images and a test set of 60,200 images. In cases where a separate test set was not available, the validation set was utilized for testing purposes. Mean Average Precision (mAP) was employed as the performance metric, with ground truth labels utilized solely for performance evaluation and not incorporated during the model training phase." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 618, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 295, + 713 + ], + "type": "text", + "content": "Implementation detail. For a fair comparison, we endeavored to set the hyperparameters for CLIP inference as closely as possible to those of the primary competitive method, CDUL. We acquired labels using the frozen CLIP ResNet50×64, and employed the ImageNet pre-trained ResNet101 as the backbone network for our classifier. The entropy threshold for calculating CLIP bias and the classifier threshold for local view inference is both 0.5. It Our" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 555, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 348 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 348 + ], + "type": "text", + "content": "warm-up classifier is trained for two epochs. While the entire training process can proceed up to 10 epochs, we obtain our final model via early stopping to prevent overfitting. Specifically, we terminate training based on the gradient of mean Average Precision (mAP) of the training labels throughout the learning process. Training was ceased when the gradient of mAP reached the first local minimum. We utilized a learning rate of " + }, + { + "bbox": [ + 313, + 72, + 555, + 348 + ], + "type": "inline_equation", + "content": "10^{-5}" + }, + { + "bbox": [ + 313, + 72, + 555, + 348 + ], + "type": "text", + "content": " across all datasets and employed a batch size of 16. Undisclosed hyperparameters of CDUL were manually reproduced. For generating the initial pseudo-labels of global views, we maintained a fixed image resolution of " + }, + { + "bbox": [ + 313, + 72, + 555, + 348 + ], + "type": "inline_equation", + "content": "640 \\times 640" + }, + { + "bbox": [ + 313, + 72, + 555, + 348 + ], + "type": "text", + "content": ". The CLIP prompts consistently used the phrase \"a photo of the [class name]\" except for the NUSWIDE dataset. (For a fair comparison with the existing method, CDUL, we employ the phrase \"a photo of a [class name]\" to match the pseudo-label performance.) For acquiring the initial pseudo-labels of local views, we introduced an offset value of 80 pixels for each vertexes when generating bounding boxes (i.e. a thresholded CAM-bounding box " + }, + { + "bbox": [ + 313, + 72, + 555, + 348 + ], + "type": "inline_equation", + "content": "\\pm 80" + }, + { + "bbox": [ + 313, + 72, + 555, + 348 + ], + "type": "text", + "content": " as our local views). Additionally, cropped images were resized to have a longer side of 640 pixels while preserving their aspect ratios to obtain CLIP probabilities." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 358, + 534, + 371 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 358, + 534, + 371 + ], + "spans": [ + { + "bbox": [ + 313, + 358, + 534, + 371 + ], + "type": "text", + "content": "4.2. Comparison with state-of-the-art methods" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 377, + 555, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 377, + 555, + 604 + ], + "spans": [ + { + "bbox": [ + 313, + 377, + 555, + 604 + ], + "type": "text", + "content": "Competitive methods. Multi-label classification can be partitioned depending on the type of supervision: fully supervised, weakly supervised, and unsupervised approaches. Firstly, in a fully supervised setting, we choose (1) the Naïve method using only binary cross-entropy (BCE) loss and (2) the Naïve method with additional label smoothing (LS) [7]. Secondly, we compared our method with recent weakly supervised methods, including partial labeled methods utilizing only " + }, + { + "bbox": [ + 313, + 377, + 555, + 604 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 313, + 377, + 555, + 604 + ], + "type": "text", + "content": " of total labels such as (3) SARB [30], (4) ASL [33], (5) Chen et al. [5]. For single positive label methods, we choose (6) LL-R [19], (7) " + }, + { + "bbox": [ + 313, + 377, + 555, + 604 + ], + "type": "inline_equation", + "content": "\\mathrm{G}^2\\mathrm{NetPL}" + }, + { + "bbox": [ + 313, + 377, + 555, + 604 + ], + "type": "text", + "content": " [1]. Finally, in an unsupervised scenario, we obtained initial labels using CLIP and evaluated the performance of (8) Naïve AN [21], (9) Szegedy et al. [38], (10) Aodha et al. [26], (11) Durand et al. [10], (12) ROLE [7], which leveraging CLIP generated initial labels. We conducted a comparative analysis with (13) CDUL [2], the current state-of-the-art in an unsupervised multi-label classification task." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "content": "Comparison with fully & weakly methods. Compared to fully supervised methods, our Classifier Guided CLIP Distillation (CCD) demonstrates comparable performance on the PASCAL VOC 2012 and PASCAL VOC 2007 datasets. In particular, achieving similar performance without any human annotation is impressive. CCD outperforms various weakly supervised methods on the PASCAL VOC 2012 and PASCAL VOC 2007 datasets, although our performances on MS COCO and NUSWIDE are relatively lower. This" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "4666" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 101, + 555, + 272 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "text", + "content": "Table 1. Mean average precision mAP in (%) for different multi-label classification methods under different supervision levels. Bold represents the best results among unsupervised methods. The proposed achieved state-of-the-art performance." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 101, + 555, + 272 + ], + "lines": [ + { + "bbox": [ + 56, + 101, + 555, + 272 + ], + "spans": [ + { + "bbox": [ + 56, + 101, + 555, + 272 + ], + "type": "table", + "html": "
Supervision levelAnnotationMethodVOC12VOC07COCONUS
Fully supervisedFully labeledBCE90.191.378.550.7
BCE-LS [7]91.692.679.451.7
Weakly supervisedPartial labeled (10%)SARB [30]-85.772.5-
ASL [33]-82.969.7-
Chen et al. [5]-81.568.1-
Single positive labeledLL-R [19]89.790.672.647.4
G2NetPL [1]89.589.972.548.5
UnsupervisedAnnotation freeNaive AN [21]85.586.565.140.8
Szegedy et al. [38]86.887.965.541.3
Aodha et al. [26]84.286.263.940.1
Durand et al. [10]81.383.163.239.4
ROLE [7]82.684.667.143.2
CDUL [2]88.689.069.244.0
CCD (Ours)90.191.070.344.5
", + "image_path": "413bed70b3ee40997e3dd510419a5517f6f81738d8f16621d3dad4e475d888c1.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 56, + 332, + 298, + 425 + ], + "blocks": [ + { + "bbox": [ + 55, + 291, + 295, + 324 + ], + "lines": [ + { + "bbox": [ + 55, + 291, + 295, + 324 + ], + "spans": [ + { + "bbox": [ + 55, + 291, + 295, + 324 + ], + "type": "text", + "content": "Table 2. Ablation study of the proposed moduels. Mean average precision mAP in " + }, + { + "bbox": [ + 55, + 291, + 295, + 324 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 55, + 291, + 295, + 324 + ], + "type": "text", + "content": " for different configuration uon the PASCAL VOC 2012 dataset. The best score is in bold." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 332, + 298, + 425 + ], + "lines": [ + { + "bbox": [ + 56, + 332, + 298, + 425 + ], + "spans": [ + { + "bbox": [ + 56, + 332, + 298, + 425 + ], + "type": "table", + "html": "
Label updateCLIP debiasConsistencymAP (VOC12)
×××86.4
✓××88.7
✓✓×89.4
✓×✓88.8
✓✓✓90.1
", + "image_path": "90fbede7364eff033f0803ced1ced44738e77ed38fc5c8afc02a59e7cd851bdd.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 448, + 295, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 448, + 295, + 592 + ], + "spans": [ + { + "bbox": [ + 54, + 448, + 295, + 592 + ], + "type": "text", + "content": "suggests that CCD is more effective on datasets with a smaller number of classes, such as PASCAL VOC 2012 and 2007, which consist of 20 classes. On the other hand, the performance gain is relatively small for complex datasets with many classes. It can be attributed to the simplistic prompt, \"a photo of the [class name],\" which might not accurately represent the target classes and co-occurrence issue, which bring out noisy local labels. The performance of initial labels of the training sets for PASCAL VOC 2012, MS COCO, and NUSWIDE is " + }, + { + "bbox": [ + 54, + 448, + 295, + 592 + ], + "type": "inline_equation", + "content": "85.3\\%" + }, + { + "bbox": [ + 54, + 448, + 295, + 592 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 54, + 448, + 295, + 592 + ], + "type": "inline_equation", + "content": "65.4\\%" + }, + { + "bbox": [ + 54, + 448, + 295, + 592 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 54, + 448, + 295, + 592 + ], + "type": "inline_equation", + "content": "41.2\\%" + }, + { + "bbox": [ + 54, + 448, + 295, + 592 + ], + "type": "text", + "content": ", respectively. CCD consistently shows about a " + }, + { + "bbox": [ + 54, + 448, + 295, + 592 + ], + "type": "inline_equation", + "content": "3.4\\%" + }, + { + "bbox": [ + 54, + 448, + 295, + 592 + ], + "type": "text", + "content": " performance gain over all datasets." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 594, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 594, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 594, + 295, + 714 + ], + "type": "text", + "content": "Comparison with unsupervised methods. Compared to unsupervised methods, CCD exhibits superior performance across all datasets. Particularly noteworthy are the significant improvements of " + }, + { + "bbox": [ + 54, + 594, + 295, + 714 + ], + "type": "inline_equation", + "content": "1.5\\% \\mathrm{p}" + }, + { + "bbox": [ + 54, + 594, + 295, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 54, + 594, + 295, + 714 + ], + "type": "inline_equation", + "content": "2\\% \\mathrm{p}" + }, + { + "bbox": [ + 54, + 594, + 295, + 714 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 54, + 594, + 295, + 714 + ], + "type": "inline_equation", + "content": "1.1\\% \\mathrm{p}" + }, + { + "bbox": [ + 54, + 594, + 295, + 714 + ], + "type": "text", + "content": " on PASCAL VOC 2012, PASCAL VOC 2007, and MS COCO, respectively, compared to the existing state-of-the-art methods. These results highlight the effectiveness of our method. However, in the case of the NUSWIDE dataset, the performance improvement is not as substantial. This can be attributed to the unique characteristics of the NUSWIDE" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 317, + 290, + 436, + 354 + ], + "blocks": [ + { + "bbox": [ + 317, + 290, + 436, + 354 + ], + "lines": [ + { + "bbox": [ + 317, + 290, + 436, + 354 + ], + "spans": [ + { + "bbox": [ + 317, + 290, + 436, + 354 + ], + "type": "image", + "image_path": "c2a15663f866cca08d7ad4bb59d99283bb7cc3c225cca6d2f9d525a20cedc33b.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 329, + 354, + 429, + 363 + ], + "lines": [ + { + "bbox": [ + 329, + 354, + 429, + 363 + ], + "spans": [ + { + "bbox": [ + 329, + 354, + 429, + 363 + ], + "type": "text", + "content": "(a) The number of mean inferences" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 384, + 555, + 418 + ], + "lines": [ + { + "bbox": [ + 313, + 384, + 555, + 418 + ], + "spans": [ + { + "bbox": [ + 313, + 384, + 555, + 418 + ], + "type": "text", + "content": "Figure 6. Classification results (mAP) on PASCAL VOC 2012 depending on (a) the change of the number of mean inferences (b) the change of ratio of pseudo-label updating." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 438, + 292, + 551, + 354 + ], + "blocks": [ + { + "bbox": [ + 438, + 292, + 551, + 354 + ], + "lines": [ + { + "bbox": [ + 438, + 292, + 551, + 354 + ], + "spans": [ + { + "bbox": [ + 438, + 292, + 551, + 354 + ], + "type": "image", + "image_path": "9a2493902e34083513eee6b395d270e7380f1bec5d0e1c419236a3f6ac82d349.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 458, + 354, + 542, + 363 + ], + "lines": [ + { + "bbox": [ + 458, + 354, + 542, + 363 + ], + "spans": [ + { + "bbox": [ + 458, + 354, + 542, + 363 + ], + "type": "text", + "content": "(b) Initial & Local weight " + }, + { + "bbox": [ + 458, + 354, + 542, + 363 + ], + "type": "inline_equation", + "content": "(\\alpha)" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 437, + 555, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 555, + 521 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 555, + 521 + ], + "type": "text", + "content": "dataset, where " + }, + { + "bbox": [ + 313, + 437, + 555, + 521 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 313, + 437, + 555, + 521 + ], + "type": "text", + "content": " of the train/test images do not contain any class-related objects (i.e. \"none images\"). As a result, the initial pseudo-label acquisition process fails to form a high probability for specific classes in these images. Consequently, during local inference, a significant portion of inferences is assigned to \"none images,\" thereby reducing the quality of pseudo-labels of CCD." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 529, + 378, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 529, + 378, + 540 + ], + "spans": [ + { + "bbox": [ + 313, + 529, + 378, + 540 + ], + "type": "text", + "content": "4.3. Ablation" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 545, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 556, + 715 + ], + "type": "text", + "content": "Effects of each module. We evaluate the effectiveness of each component of our method in Tab. 2. The performance of the baseline model, the classifier trained with CLIP-generated global pseudo-labels, is " + }, + { + "bbox": [ + 313, + 545, + 556, + 715 + ], + "type": "inline_equation", + "content": "86.4\\%" + }, + { + "bbox": [ + 313, + 545, + 556, + 715 + ], + "type": "text", + "content": ". When we update the pseudo-label with classifier-guided local inference, it shows a significant improvement of " + }, + { + "bbox": [ + 313, + 545, + 556, + 715 + ], + "type": "inline_equation", + "content": "2.3\\%" + }, + { + "bbox": [ + 313, + 545, + 556, + 715 + ], + "type": "text", + "content": " p. This shows that our CAM-based local view selection method and adaptively allocating the number of CLIP inferences are effective. It already exceeds the performance of previous state-of-the-art methods. Debiasing CLIP also contributed " + }, + { + "bbox": [ + 313, + 545, + 556, + 715 + ], + "type": "inline_equation", + "content": "0.7\\%" + }, + { + "bbox": [ + 313, + 545, + 556, + 715 + ], + "type": "text", + "content": " p, indicating that our proposed debiasing method effectively relieves the CLIP bias. The consistency loss alone adds only " + }, + { + "bbox": [ + 313, + 545, + 556, + 715 + ], + "type": "inline_equation", + "content": "0.1\\%" + }, + { + "bbox": [ + 313, + 545, + 556, + 715 + ], + "type": "text", + "content": " p performance gain. However, when CLIP debiasing and consistency loss are combined, a substantial improvement of" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "4667" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 124, + 294, + 155 + ], + "blocks": [ + { + "bbox": [ + 56, + 71, + 294, + 114 + ], + "lines": [ + { + "bbox": [ + 56, + 71, + 294, + 114 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 294, + 114 + ], + "type": "text", + "content": "Table 3. Per-class AP in " + }, + { + "bbox": [ + 56, + 71, + 294, + 114 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 56, + 71, + 294, + 114 + ], + "type": "text", + "content": " for without and with debiasing. We focus on the seven classes from the lowest mean probability. The result showcases that our debiasing is effective for enhancing the performance of biased classes." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 124, + 294, + 155 + ], + "lines": [ + { + "bbox": [ + 56, + 124, + 294, + 155 + ], + "spans": [ + { + "bbox": [ + 56, + 124, + 294, + 155 + ], + "type": "table", + "html": "
Methodbottlechairtablepersonplantsofatv
w/o debias77.077.172.787.368.873.089.5
W debias77.679.074.889.272.176.692.0
", + "image_path": "cbef149517b9033918d7e806b5a7d3c907015388817322b03726a709c1fd5b9a.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 180, + 294, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 180, + 294, + 239 + ], + "spans": [ + { + "bbox": [ + 58, + 180, + 294, + 239 + ], + "type": "inline_equation", + "content": "1.4\\%" + }, + { + "bbox": [ + 58, + 180, + 294, + 239 + ], + "type": "text", + "content": " p is achieved. This result implies that pseudo-labels generated by compensating for CLIP bias may amplify both accurate information and noise. Since the consistency loss is effective for dealing with noisy data, it creates a synergy with CLIP debiasing." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 242, + 294, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 242, + 294, + 469 + ], + "spans": [ + { + "bbox": [ + 57, + 242, + 294, + 469 + ], + "type": "text", + "content": "The number of inference. We examine how model performance varies with different numbers of CLIP local inferences. We incrementally increase the average number of inferences to approximately 4, 6, 10, and 14. In Fig. 6 (a), the model achieves its best performance when the average number of inferences is around 10. Notably, increasing the number of inferences does not consistently enhance performance. Fewer inferences limit the allocation of local views to challenging samples, which hinders the ability to extract sufficient information from complex images, thereby negatively impacting negatively. In contrast, an excessive number of inferences results in the overallocation of local views, even for easy samples. It potentially introduces redundant inference for easy samples, making noisy pseudo-labels. In particular, when the average number of inferences is set to 10 or 14, the maximum number of local inferences remains capped at 20 (i.e. the total number of classes). It suggests that the inferences allocated to complex images have already reached saturation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 57, + 472, + 294, + 697 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 472, + 294, + 697 + ], + "spans": [ + { + "bbox": [ + 57, + 472, + 294, + 697 + ], + "type": "text", + "content": "Initial & local weight. We investigate the impact of adjusting the ratio " + }, + { + "bbox": [ + 57, + 472, + 294, + 697 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 57, + 472, + 294, + 697 + ], + "type": "text", + "content": " between initial and local labels when updating the pseudo-labels (Eq. 7). In Fig. 6 (b), the results indicate that utilizing both global and local labels together leads to performance enhancement. This trend arises because two different label sources have complementary properties. The local label contains more information through multiple CLIP inferences of different local patches. However, since the reliability of CAM is not " + }, + { + "bbox": [ + 57, + 472, + 294, + 697 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 57, + 472, + 294, + 697 + ], + "type": "text", + "content": ", it may draw an invalid patch. Conversely, the initial label contains less information, but it is always generated from valid images. By blending these two sources, we reinforce consistent predictions between local and initial labels, thereby strengthening the signal of reliable labels. Additionally, we highlight that the use of our local labels alone (i.e. at weight 0) can outperforms CDUL. This confirms that our choice of local views and their labels extracts more informative learning signals than the mixture of the global view and uniformly sampled local views." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 58, + 702, + 294, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 294, + 712 + ], + "type": "text", + "content": "Per-class mAP of biased classes. We assess the effective" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 316, + 72, + 553, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 228 + ], + "type": "text", + "content": "ness of our proposed method by examining the improvement in performance for specific target classes before and after CLIP debiasing. In Tab. 3, we focus on seven classes from the biased class, including bottle, chair, table, person, potted plant, sofa, and tv monitor as confirmed in the Fig. 2. Our results demonstrate consistent improvements in the performance of all these classes, ranging from " + }, + { + "bbox": [ + 316, + 72, + 553, + 228 + ], + "type": "inline_equation", + "content": "0.6\\% \\mathrm{p}" + }, + { + "bbox": [ + 316, + 72, + 553, + 228 + ], + "type": "text", + "content": " (bottle) to " + }, + { + "bbox": [ + 316, + 72, + 553, + 228 + ], + "type": "inline_equation", + "content": "3.6\\% \\mathrm{p}" + }, + { + "bbox": [ + 316, + 72, + 553, + 228 + ], + "type": "text", + "content": " (sofa). Moreover, we observe comparable or slightly enhanced performance for the remaining classes not shown in the table. The result affirms the effectiveness of our method across a broader spectrum. Detailed class-wise performance results for the entire dataset are provided in the supplementary material." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 316, + 236, + 386, + 246 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 236, + 386, + 246 + ], + "spans": [ + { + "bbox": [ + 316, + 236, + 386, + 246 + ], + "type": "text", + "content": "4.4. Limitation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 316, + 253, + 553, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 253, + 553, + 431 + ], + "spans": [ + { + "bbox": [ + 316, + 253, + 553, + 431 + ], + "type": "text", + "content": "A limitation of proposed method is that the generated local views often struggle to distinguish between co-occurring objects, an issue inherited from CAM. This challenge is especially evident in complex datasets like COCO, where similar classes—such as skies, snowboards, and skateboards—frequently appear in scene with human activity. These classes often co-occur with human feet, leading classifiers to associate high activations near the feet with these classes. This phenomenon makes them potential candidates for local views. This redundant information results in noisy pseudo-labels, reducing overall performance. Examples of such failure cases are provided in the supplementary materials. Developing an improved local view proposal method to better handle noisy inferences and focusing on the target objects will be an important direction for future work." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 316, + 443, + 386, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 443, + 386, + 453 + ], + "spans": [ + { + "bbox": [ + 316, + 443, + 386, + 453 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 316, + 463, + 553, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 463, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 463, + 553, + 712 + ], + "type": "text", + "content": "In this paper, we propose a novel unsupervised multi-label classification method, Classifier-guided CLIP Distillation (CCD). We address two critical issues in CLIP: inconsistent predictions due to subtle input changes and inherent prediction bias. To tackle these issues, we devised a classifier-guided label update and CLIP debiasing. Firstly, we identified the effectiveness of acquiring local views near class-related objects and proposed a classifier-guided label update. It selects local views utilizing the Class Activation Mapping (CAM) and acquires corresponding labels from CLIP. Our approach results in allocating more local views for complex images. Secondly, we address the CLIP's bias. We investigated the biased prediction distribution of CLIP and rectified the pseudo-labels. This step notably boosts performance in certain classes and reduces the impact of bias on CLIP. Through experiments on four benchmarks, CCD newly achieved state-of-the-art performances in unsupervised multi-label classification. Remarkably, CCD achieved comparable performance to fully supervised methods on PASCAL VOC 2012 and PASCAL VOC 2007 datasets without relying on annotations." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 750, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 750, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 750, + 315, + 757 + ], + "type": "text", + "content": "4668" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 205 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 205 + ], + "type": "text", + "content": "Acknowledgements. This research was supported by the Basic Science Research Program through the National Research Foundation of Korea (NRF) funded by the MSIP (RS-2025-00520207, RS-2023-00219019), KEIT grant funded by the Korean government (MOTIE) (No. 2022-0-00680, No. 2022-0-01045), the IITP grant funded by the Korean government (MSIT) (No. RS-2024-00457882, National AI Research Lab Project, No. 2021-0-02068 Artificial Intelligence Innovation Hub, RS-2019-II190075 Artificial Intelligence Graduate School Program (KAIST)), Samsung Electronics Co., Ltd (IO230508-06190-01) and SAMSUNG Research, Samsung Electronics Co., Ltd." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 228, + 115, + 241 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 228, + 115, + 241 + ], + "spans": [ + { + "bbox": [ + 56, + 228, + 115, + 241 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 248, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 61, + 248, + 294, + 292 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 248, + 294, + 292 + ], + "spans": [ + { + "bbox": [ + 61, + 248, + 294, + 292 + ], + "type": "text", + "content": "[1] Rabab Abdelfattah, Xin Zhang, Mostafa M Fouda, Xiaofeng Wang, and Song Wang. G2netpl: Generic game-theoretic network for partial-label image classification. arXiv preprint arXiv:2210.11469, 2022. 6, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 294, + 294, + 349 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 294, + 294, + 349 + ], + "spans": [ + { + "bbox": [ + 61, + 294, + 294, + 349 + ], + "type": "text", + "content": "[2] Rabab Abdelfattah, Qing Guo, Xiaoguang Li, Xiaofeng Wang, and Song Wang. Cdul: Clip-driven unsupervised learning for multi-label image classification. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1348-1357, 2023. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 350, + 295, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 350, + 295, + 373 + ], + "spans": [ + { + "bbox": [ + 62, + 350, + 295, + 373 + ], + "type": "text", + "content": "[3] Görkem Algan and Ülkay Ulusoy. Label noise types and their effects on deep learning, 2020. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 374, + 294, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 374, + 294, + 407 + ], + "spans": [ + { + "bbox": [ + 62, + 374, + 294, + 407 + ], + "type": "text", + "content": "[4] Serhat Selcuk Bucak, Rong Jin, and Anil K. Jain. Multi-label learning with incomplete class assignments. In CVPR 2011, pages 2801-2808, 2011. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 409, + 294, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 409, + 294, + 453 + ], + "spans": [ + { + "bbox": [ + 62, + 409, + 294, + 453 + ], + "type": "text", + "content": "[5] Tianshui Chen, Tao Pu, Hefeng Wu, Yuan Xie, and Liang Lin. Structured semantic transfer for multi-label recognition with partial labels. In Proceedings of the AAAI conference on artificial intelligence, pages 339-346, 2022. 1, 2, 6, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 454, + 294, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 454, + 294, + 509 + ], + "spans": [ + { + "bbox": [ + 62, + 454, + 294, + 509 + ], + "type": "text", + "content": "[6] Tat-Seng Chua, Jinhui Tang, Richang Hong, Haojie Li, Zhiping Luo, and Yantao Zheng. Nus-wide: a real-world web image database from national university ofSingapore. In Proceedings of the ACM international conference on image and video retrieval, pages 1-9, 2009. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 511, + 294, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 511, + 294, + 565 + ], + "spans": [ + { + "bbox": [ + 62, + 511, + 294, + 565 + ], + "type": "text", + "content": "[7] Elijah Cole, Oisin Mac Aodha, Titouan Lorieul, Pietro Perona, Dan Morris, and Nebojsa Jojic. Multi-label learning from single positive labels. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 933-942, 2021. 1, 2, 6, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 567, + 294, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 567, + 294, + 632 + ], + "spans": [ + { + "bbox": [ + 62, + 567, + 294, + 632 + ], + "type": "text", + "content": "[8] Katherine Crowson, Stella Biderman, Daniel Kornis, Dashiell Stander, Eric Hallahan, Louis Castricato, and Edward Raff. Vqgan-clip: Open domain image generation and editing with natural language guidance. In European Conference on Computer Vision, pages 88–105. Springer, 2022. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 635, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 635, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 62, + 635, + 294, + 689 + ], + "type": "text", + "content": "[9] Jia Deng, Olga Russakovsky, Jonathan Krause, Michael S Bernstein, Alex Berg, and Li Fei-Fei. Scalable multi-label annotation. In Proceedings of the SIGCHI Conference on Human Factors in Computing Systems, pages 3099–3102, 2014. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 294, + 713 + ], + "type": "text", + "content": "[10] Thibaut Durand, Nazanin Mehrasa, and Greg Mori. Learning a deep convnet for multi-label classification with partial" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "text", + "content": "labels. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 647-657, 2019. 1, 2, 6, 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 107, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 150 + ], + "type": "text", + "content": "[11] M. Everingham, L. Van Gool, C. K. I. Williams, J. Winn, and A. Zisserman. The Pascal visual object classes (voc) challenge. International Journal of Computer Vision, 88(2): 303-338, 2010. 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 152, + 553, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 553, + 195 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 553, + 195 + ], + "type": "text", + "content": "[12] Xiuye Gu, Tsung-Yi Lin, Weicheng Kuo, and Yin Cui. Open-vocabulary object detection via vision and language knowledge distillation. arXiv preprint arXiv:2104.13921, 2021. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 198, + 553, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 198, + 553, + 241 + ], + "spans": [ + { + "bbox": [ + 316, + 198, + 553, + 241 + ], + "type": "text", + "content": "[13] Ping Hu, Ximeng Sun, Stan Sclaroff, and Kate Saenko. Dualcoop++: Fast and effective adaptation to multi-label recognition with limited annotations. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 243, + 553, + 286 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 243, + 553, + 286 + ], + "spans": [ + { + "bbox": [ + 316, + 243, + 553, + 286 + ], + "type": "text", + "content": "[14] Dat Huynh and Ehsan Elhamifar. Interactive multi-label cnn learning with partial labels. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9423-9432, 2020. 1, 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 288, + 553, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 288, + 553, + 352 + ], + "spans": [ + { + "bbox": [ + 317, + 288, + 553, + 352 + ], + "type": "text", + "content": "[15] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International conference on machine learning, pages 4904-4916. PMLR, 2021. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 354, + 553, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 354, + 553, + 397 + ], + "spans": [ + { + "bbox": [ + 317, + 354, + 553, + 397 + ], + "type": "text", + "content": "[16] Ashish Kapoor, Raajay Viswanathan, and Prateek Jain. Multilabel classification using bayesian compressed sensing. Advances in neural information processing systems, 25, 2012. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 399, + 553, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 399, + 553, + 454 + ], + "spans": [ + { + "bbox": [ + 317, + 399, + 553, + 454 + ], + "type": "text", + "content": "[17] Apoorv Khandelwal, Luca Weihs, Roozbeh Mottaghi, and Aniruddha Kembhavi. Simple but effective: Clip embeddings for embodied ai. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14829-14838, 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 456, + 553, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 456, + 553, + 488 + ], + "spans": [ + { + "bbox": [ + 317, + 456, + 553, + 488 + ], + "type": "text", + "content": "[18] Dongseob Kim, Seungho Lee, Junsuk Choe, and Hyunjung Shim. Weakly supervised semantic segmentation for driving scenes. arXiv preprint arXiv:2312.13646, 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 491, + 553, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 491, + 553, + 544 + ], + "spans": [ + { + "bbox": [ + 317, + 491, + 553, + 544 + ], + "type": "text", + "content": "[19] Youngwook Kim, Jae Myung Kim, Zeynep Akata, and Jungwoo Lee. Large loss matters in weakly supervised multi-label classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14156-14165, 2022. 1, 2, 6, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 546, + 553, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 546, + 553, + 611 + ], + "spans": [ + { + "bbox": [ + 317, + 546, + 553, + 611 + ], + "type": "text", + "content": "[20] Youngwook Kim, Jae Myung Kim, Jieun Jeong, Cordelia Schmid, Zeynep Akata, and Jungwoo Lee. Bridging the gap between model explanations in partially annotated multi-label classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3408-3417, 2023. 1, 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 613, + 553, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 613, + 553, + 656 + ], + "spans": [ + { + "bbox": [ + 317, + 613, + 553, + 656 + ], + "type": "text", + "content": "[21] Kaustav Kundu and Joseph Tighe. Exploiting weakly supervised visual patterns to learn from partial annotations. Advances in Neural Information Processing Systems, 33:561-572, 2020. 6, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 658, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 658, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 658, + 553, + 713 + ], + "type": "text", + "content": "[22] Yue Liao, Aixi Zhang, Miao Lu, Yongliang Wang, Xiaobo Li, and Si Liu. Gen-vlkt: Simplify association and enhance interaction understanding for hoi detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20123-20132, 2022. 2" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4669" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 714 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 139 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 139 + ], + "type": "text", + "content": "[23] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 140, + 294, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 140, + 294, + 184 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 294, + 184 + ], + "type": "text", + "content": "[24] Bing-Qing Liu, Bin-Bin Jia, and Min-Ling Zhang. Towards enabling binary decomposition for partial multi-label learning. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(11):13203-13217, 2023. 1, 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 186, + 294, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 186, + 294, + 218 + ], + "spans": [ + { + "bbox": [ + 56, + 186, + 294, + 218 + ], + "type": "text", + "content": "[25] Weiwei Liu and Ivor Tsang. On the optimality of classifier chain for multi-label classification. Advances in Neural Information Processing Systems, 28, 2015. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 220, + 294, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 220, + 294, + 274 + ], + "spans": [ + { + "bbox": [ + 56, + 220, + 294, + 274 + ], + "type": "text", + "content": "[26] Oisin Mac Aodha, Elijah Cole, and Pietro Perona. Presence-only geographical priors for fine-grained image classification. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9596-9606, 2019. 2, 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 276, + 294, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 276, + 294, + 342 + ], + "spans": [ + { + "bbox": [ + 56, + 276, + 294, + 342 + ], + "type": "text", + "content": "[27] Dhruv Mahajan, Ross Girshick, Vignesh Ramanathan, Kaiming He, Manohar Paluri, Yixuan Li, Ashwin Bharambe, and Laurens Van Der Maaten. Exploring the limits of weakly supervised pretraining. In Proceedings of the European conference on computer vision (ECCV), pages 181-196, 2018. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 343, + 294, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 343, + 294, + 409 + ], + "spans": [ + { + "bbox": [ + 56, + 343, + 294, + 409 + ], + "type": "text", + "content": "[28] Antoine Miech, Jean-Baptiste Alayrac, Lucas Smaira, Ivan Laptev, Josef Sivic, and Andrew Zisserman. End-to-end learning of visual representations from uncurated instructional videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9879-9889, 2020. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 411, + 294, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 411, + 294, + 466 + ], + "spans": [ + { + "bbox": [ + 56, + 411, + 294, + 466 + ], + "type": "text", + "content": "[29] Ishan Misra, C Lawrence Zitnick, Margaret Mitchell, and Ross Girshick. Seeing through the human reporting bias: Visual classifiers from noisy human-centric labels. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2930–2939, 2016. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 468, + 294, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 468, + 294, + 521 + ], + "spans": [ + { + "bbox": [ + 56, + 468, + 294, + 521 + ], + "type": "text", + "content": "[30] Tao Pu, Tianshui Chen, Hefeng Wu, and Liang Lin. Semantic-aware representation blending for multi-label image recognition with partial labels. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2091-2098, 2022. 1, 2, 6, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 523, + 294, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 523, + 294, + 589 + ], + "spans": [ + { + "bbox": [ + 56, + 523, + 294, + 589 + ], + "type": "text", + "content": "[31] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 590, + 294, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 590, + 294, + 656 + ], + "spans": [ + { + "bbox": [ + 56, + 590, + 294, + 656 + ], + "type": "text", + "content": "[32] Yongming Rao, Wenliang Zhao, Guangyi Chen, Yansong Tang, Zheng Zhu, Guan Huang, Jie Zhou, and Jiwen Lu. Denseclip: Language-guided dense prediction with context-aware prompting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18082-18091, 2022. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 658, + 294, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 294, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 294, + 714 + ], + "type": "text", + "content": "[33] Tal Ridnik, Emanuel Ben-Baruch, Nadav Zamir, Asaf Noy, Itamar Friedman, Matan Protter, and Lihi Zelnik-Manor. Asymmetric loss for multi-label classification. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 82-91, 2021. 6, 7" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 714 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 138 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 138 + ], + "type": "text", + "content": "[34] Amanpreet Singh, Ronghang Hu, Vedanuj Goswami, Guillaume Couairon, Wojciech Galuba, Marcus Rohrbach, and Douwe Kiela. Flava: A foundational language and vision alignment model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15638-15650, 2022. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 316, + 140, + 553, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 140, + 553, + 194 + ], + "spans": [ + { + "bbox": [ + 316, + 140, + 553, + 194 + ], + "type": "text", + "content": "[35] GyeongRyeol Song, Noo-ri Kim, Jin-Seop Lee, and JeeHyong Lee. Ignore: Information gap-based false negative loss rejection for single positive multi-label learning. In European Conference on Computer Vision, pages 472–488. Springer, 2025. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 195, + 553, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 195, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 316, + 195, + 553, + 239 + ], + "type": "text", + "content": "[36] Chen Sun, Abhinav Shrivastava, Saurabh Singh, and Abhinav Gupta. Revisiting unreasonable effectiveness of data in deep learning era. In Proceedings of the IEEE international conference on computer vision, pages 843-852, 2017. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 239, + 553, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 239, + 553, + 272 + ], + "spans": [ + { + "bbox": [ + 316, + 239, + 553, + 272 + ], + "type": "text", + "content": "[37] Yu-Yin Sun, Yin Zhang, and Zhi-Hua Zhou. Multi-label learning with weak label. Proceedings of the AAAI Conference on Artificial Intelligence, 24(1):593-598, 2010. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 272, + 553, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 272, + 553, + 326 + ], + "spans": [ + { + "bbox": [ + 316, + 272, + 553, + 326 + ], + "type": "text", + "content": "[38] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2818-2826, 2016. 2, 6, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 327, + 553, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 327, + 553, + 371 + ], + "spans": [ + { + "bbox": [ + 316, + 327, + 553, + 371 + ], + "type": "text", + "content": "[39] Guy Tevet, Brian Gordon, Amir Hertz, Amit H Bermano, and Daniel Cohen-Or. Motionclip: Exposing human motion generation to clip space. In European Conference on Computer Vision, pages 358–374. Springer, 2022. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 372, + 553, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 372, + 553, + 404 + ], + "spans": [ + { + "bbox": [ + 316, + 372, + 553, + 404 + ], + "type": "text", + "content": "[40] Grigorios Tsoumakas and Ioannis Katakis. Multi-label classification: An overview. International Journal of Data Warehousing and Mining (IJDWM), 3(3):1-13, 2007. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 404, + 553, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 404, + 553, + 459 + ], + "spans": [ + { + "bbox": [ + 316, + 404, + 553, + 459 + ], + "type": "text", + "content": "[41] Deepak Vasisht, Andreas Damianou, Manik Varma, and Ashish Kapoor. Active learning for sparse bayesian multilabel classification. In Proceedings of the 20th ACM SIGKDD international conference on Knowledge discovery and data mining, pages 472-481, 2014. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 459, + 553, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 459, + 553, + 525 + ], + "spans": [ + { + "bbox": [ + 316, + 459, + 553, + 525 + ], + "type": "text", + "content": "[42] Thomas Verelst, Paul K. Rubenstein, Marcin Eichner, Tinne Tuytelaars, and Maxim Berman. Spatial consistency loss for training multi-label classifiers from single-label annotations. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 3879-3889, 2023. 1, 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 525, + 553, + 570 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 525, + 553, + 570 + ], + "spans": [ + { + "bbox": [ + 316, + 525, + 553, + 570 + ], + "type": "text", + "content": "[43] Changhu Wang, Shuicheng Yan, Lei Zhang, and Hong-Jiang Zhang. Multi-label sparse coding for automatic image annotation. In 2009 IEEE conference on computer vision and pattern recognition, pages 1643-1650. IEEE, 2009. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 570, + 553, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 570, + 553, + 624 + ], + "spans": [ + { + "bbox": [ + 316, + 570, + 553, + 624 + ], + "type": "text", + "content": "[44] Can Wang, Menglei Chai, Mingming He, Dongdong Chen, and Jing Liao. Clip-nerf: Text-and-image driven manipulation of neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3835-3844, 2022. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 625, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 625, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 625, + 553, + 689 + ], + "type": "text", + "content": "[45] Qifan Wang, Bin Shen, Shumiao Wang, Liang Li, and Luo Si. Binary codes embedding for fast image tagging with incomplete labels. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part II 13, pages 425-439. Springer, 2014. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 690, + 553, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 690, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 690, + 553, + 714 + ], + "type": "text", + "content": "[46] Baoyuan Wu, Siwei Lyu, and Bernard Ghanem. Ml-mg: Multi-label learning with missing labels using a mixed" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "4670" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 341 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 76, + 72, + 296, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 296, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 296, + 95 + ], + "type": "text", + "content": "graph. In Proceedings of the IEEE international conference on computer vision, pages 4157-4165, 2015. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 296, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 296, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 296, + 140 + ], + "type": "text", + "content": "[47] Xin Xing, Zhexiao Xiong, Abby Stylianou, Srikumar Sastry, Liyu Gong, and Nathan Jacobs. Vision-language pseudolabels for single-positive multi-label learning. arXiv preprint arXiv:2310.15985, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 141, + 296, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 141, + 296, + 194 + ], + "spans": [ + { + "bbox": [ + 56, + 141, + 296, + 194 + ], + "type": "text", + "content": "[48] Alireza Zareian, Kevin Dela Rosa, Derek Hao Hu, and Shih-Fu Chang. Open-vocabulary object detection using captions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14393-14402, 2021. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 197, + 296, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 197, + 296, + 251 + ], + "spans": [ + { + "bbox": [ + 56, + 197, + 296, + 251 + ], + "type": "text", + "content": "[49] Bolei Zhou, Aditya Khosla, Agata Lapedriza, Aude Oliva, and Antonio Torralba. Learning deep features for discriminative localization. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2921-2929, 2016. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 253, + 296, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 253, + 296, + 285 + ], + "spans": [ + { + "bbox": [ + 56, + 253, + 296, + 285 + ], + "type": "text", + "content": "[50] Chong Zhou, Chen Change Loy, and Bo Dai. Extract free dense labels from clip. In European Conference on Computer Vision, pages 696-712. Springer, 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 286, + 296, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 286, + 296, + 341 + ], + "spans": [ + { + "bbox": [ + 56, + 286, + 296, + 341 + ], + "type": "text", + "content": "[51] Donghao Zhou, Pengfei Chen, Qiong Wang, Guangyong Chen, and Pheng-Ann Heng. Acknowledging the unknown for multi-label learning with single positive labels. In European Conference on Computer Vision, pages 423-440. Springer, 2022. 1, 2" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4671" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Classifier-to-Bias_ Toward Unsupervised Automatic Bias Detection for Visual Classifiers/303d2f58-e46e-4b7a-806e-0d4e64d53df2_content_list.json b/2025/Classifier-to-Bias_ Toward Unsupervised Automatic Bias Detection for Visual Classifiers/303d2f58-e46e-4b7a-806e-0d4e64d53df2_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..90cd0afa7acefcd63feb9387af833f05c11908af --- /dev/null +++ b/2025/Classifier-to-Bias_ Toward Unsupervised Automatic Bias Detection for Visual Classifiers/303d2f58-e46e-4b7a-806e-0d4e64d53df2_content_list.json @@ -0,0 +1,1588 @@ +[ + { + "type": "text", + "text": "CLASSIFIER-TO-BIAS: Toward Unsupervised Automatic Bias Detection for Visual Classifiers", + "text_level": 1, + "bbox": [ + 210, + 130, + 785, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Quentin Guimard $^{1}$ Moreno D'Incà $^{1}$ Massimiliano Mancini $^{1}$ Elisa Ricci $^{1,2}$ $^{1}$ University of Trento $^{2}$ Fondazione Bruno Kessler", + "bbox": [ + 178, + 202, + 816, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "https://github.com/mardgui/C2B", + "bbox": [ + 361, + 241, + 630, + 255 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 291, + 326, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "A person downloading a pre-trained model from the web should be aware of its biases. Existing approaches for bias identification rely on datasets containing labels for the task of interest, something that a non-expert may not have access to, or may not have the necessary resources to collect: this greatly limits the number of tasks where model biases can be identified. In this work, we present CLASSIFIER-TO-BIAS (C2B), the first bias discovery framework that works without access to any labeled data: it only relies on a textual description of the classification task to identify biases in the target classification model. This description is fed to a large language model to generate bias proposals and corresponding captions depicting biases together with task-specific target labels. A retrieval model collects images for those captions, which are then used to assess the accuracy of the model w.r.t. the given biases. C2B is training-free, does not require any annotations, has no constraints on the list of biases, and can be applied to any pre-trained model on any classification task. Experiments on two publicly available datasets show that C2B discovers biases beyond those of the original datasets and outperforms a recent state-of-the-art bias detection baseline that relies on task-specific annotations, being a promising first step toward addressing task-agnostic unsupervised bias detection.", + "bbox": [ + 89, + 325, + 483, + 688 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 722, + 220, + 738 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The wide availability of pre-trained machine learning models allows anyone to access them. A user may search in online libraries (e.g., HuggingFace [27]) for models addressing their task of interest, download the most suitable one, and use it for their specific use case. As the user did not develop the model, they may not be aware of potential failure modes [28, 60, 63] or harmful biases [9, 10, 24, 49, 69, 70]. These are usually reported on model cards [47], describing how the model has been developed, in which cases it has been tested, its recommended use, and its potential biases.", + "bbox": [ + 89, + 750, + 483, + 900 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1ad9d05a0618ab645e66313f37c59c01ce03526b801871aae1a393c66b8ba488.jpg", + "image_caption": [ + "Figure 1. We explore the novel task of bias discovery when we are only given a specific classification task and a pre-trained model. We propose CLASSIFIER-TO-BIAS (C2B), which automatically detects potential biases in the model, categorizes these biases, and assigns scores to each category." + ], + "image_footnote": [], + "bbox": [ + 516, + 287, + 903, + 500 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "While model cards are a useful tool to describe a model and its shortcomings, (i) a lot of models still come without them; (ii) even if they are present, they may not describe specific cases or biases of interest to the users. This means that a user would need to assess the weaknesses of the downloaded model independently.", + "bbox": [ + 511, + 609, + 906, + 700 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Various tools have already been proposed to measure and mitigate biases in datasets [8, 64] and models [2, 8, 11]. While these tools rely on bias labels, recent work has made progress toward open-set failure mode extraction [16, 19, 22, 30, 53] and bias detection [6, 17, 33, 34, 68] without bias labels, even removing the need for a predefined list of biases. However, in the context of discriminative models such as visual classifiers, all existing methods rely on task-specific labeled data to be able to identify failure modes and measure the biases, limiting the number of tasks where model biases can be identified, and preventing the automation of bias detection.", + "bbox": [ + 511, + 702, + 908, + 882 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To address this problem, we present CLASSIFIER-TO", + "bbox": [ + 532, + 885, + 903, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "15151", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "BIAS (C2B), the first framework that, given a pre-trained visual classifier and its task specification, automatically identifies its biases (Fig. 1) in an unsupervised manner, removing the need for task-specific labeled data. We build C2B on two components: a large language model (LLM) and a text-to-image retrieval engine. First, the task specifications are fed to the LLM to produce a series of candidate bias attributes for the task (e.g., gender, ethnicity), and their corresponding bias classes (e.g., male, female). The list of candidate bias classes is fed again to the LLM to generate a set of captions describing both the biases and the task-specific target classes they are associated to. These captions are used by the retrieval module to get images from a large database: the images will constitute our dataset, where, for each image, both the target and bias labels are given by the constructed caption itself. By testing the pre-trained model on this dataset, we can compute the standard bias metrics, verifying which bias class the model is most biased against.", + "bbox": [ + 89, + 90, + 480, + 362 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Importantly, C2B is fully unsupervised and does not suffer from the need of costly annotations, thus it can be applied to automatically detect biases for virtually any classification task. Moreover, the list of biases is class-specific, as different classification tasks and categories (e.g., birds, deers) may exhibit different biases (e.g., tail color, horns size). Finally, the model is modular and can readily incorporate more powerful LLMs, retrieval approaches, or databases as they become available. We test our framework on two datasets, i.e., CelebA [41], and ImageNet-X [28]. Results indicate that C2B can uncover biases beyond those annotated in the original datasets, outperforming a recent supervised open-set bias detector [33].", + "bbox": [ + 89, + 363, + 480, + 559 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Contributions. To summarize, our contribution is threefold: (i) We propose the new task of unsupervised bias detection in a pre-trained model given only a textual description of the classification task without requiring any annotated dataset or manual intervention, and define a protocol to evaluate methods on this setting; (ii) We present C2B, the first framework for this task that uses an LLM to propose a list of biases and generate captions, and multimodal retrieval to produce a dataset from which biases can be estimated; (iii) We show that C2B can discover biases beyond those annotated in the original datasets, being the first step toward addressing this challenging problem and democratizing bias discovery.", + "bbox": [ + 89, + 560, + 482, + 755 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 768, + 232, + 784 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Bias Detection and Mitigation. Bias detection and mitigation are essential to ensure fairness and robustness in machine learning models. Several approaches tackle hidden subgroup imbalances without relying on explicit group annotations [35, 37, 59, 61], e.g., by analyzing the mistakes on annotated task data and re-weighting misclassified samples [37]. Other methods focus on open-set or annotation-", + "bbox": [ + 89, + 794, + 480, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "free bias discovery [34, 36, 68]. UDIS [34] uses hierarchical clustering to surface systematic biases without predefined subgroup definitions. DebiAN [36] alternates training between a bias discoverer and a classifier to identify and mitigate multiple biases. DIM [68] decomposes feature spaces using Partial Least Squares to uncover multiple unknown biased subgroups and describes them using vision-language models. More recently, language-driven methods have also gained prominence, using captions and keyword extraction to detect and name potential biases learned by models [14, 33, 71]. B2T [33] and Zhao et al. [71] relies on captions derived from annotated datasets or misclassified samples with known labels, validating these biases using, e.g. vision-language similarity. SaMyNa [14] emphasizes interpretability by extracting semantic bias descriptors from large-scale captioning, supporting bias analysis and debiasing. However, all these methods require access to annotated, task-specific datasets.", + "bbox": [ + 511, + 90, + 903, + 362 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In contrast, our approach, C2B, operates without any kind of annotated data. It relies solely on a natural language description of the classification task and a pre-trained classifier. Using LLM-generated, class-specific bias proposals combined with retrieval-based evaluation, C2B enables fully unsupervised, task-agnostic bias discovery.", + "bbox": [ + 511, + 363, + 903, + 453 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Other related methods focus on bias detection or fairness in generative models [17, 32, 57]. For instance, Open-Bias [17] uses LLMs and VQA models to detect and quantify open-set biases in text-to-image generative models, while GELDA [32] reveals visual biases in image generators by generating domain-specific attributes with LLMs and annotating them using selected VLMs.", + "bbox": [ + 511, + 454, + 903, + 559 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Inspired by these works, we choose to leverage an LLM as well as a text-to-image retrieval engine to tackle the new task of unsupervised bias detection, where only a pretrained model and a textual description of the classification task are available, thus eliminating the need for annotated datasets or manual intervention.", + "bbox": [ + 511, + 561, + 903, + 650 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Explainability and Failure Mode Detection. A parallel research direction focuses on systematic failure mode discovery and interpretability [3, 4, 16, 25, 29]. VoG [3] highlights difficult examples by analyzing gradient variance, and Spotlight [16] identifies contiguous underperforming regions in representation space. Saliency-based methods [29] provide local visual explanations, while counterfactual explanations are generated in either natural language [25] or visual form using diffusion models [4].", + "bbox": [ + 511, + 657, + 903, + 792 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recent work has scaled these efforts using embeddings and generative models. Wiles et al. [65] leverages large-scale generative models to automatically find and describe failure clusters. Domino [19] uses cross-modal embeddings and mixture models to identify and describe coherent data slices where models underperform. Jain et al. [30] extracts latent directions corresponding to consistent model errors", + "bbox": [ + 511, + 795, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "15152", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "and generates challenging counterfactual examples for data augmentation. AdaVision [22] enables interactive discovery of failure modes by combining CLIP-based retrieval with GPT-generated prompts. PRIME [53] prioritizes interpretability by starting from human-defined tags and identifying minimal descriptive combinations that explain failures. MAIA [56] automates interpretability workflows using vision-language models for feature analysis and failure mode discovery. Csurka et al. [15] formalizes language-based error explainability and proposes evaluation metrics to assess the quality of generated explanations.", + "bbox": [ + 89, + 90, + 480, + 256 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "While these methods analyze model behavior and failure patterns through model outputs or datasets, our work introduces a complementary approach: starting from a high-level task description, C2B uses LLM-based bias proposals and retrieval-based validation to automatically identify, validate, and quantify biases in pre-trained classifiers, enabling large-scale, unsupervised bias discovery across arbitrary classification tasks.", + "bbox": [ + 89, + 257, + 480, + 378 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 393, + 181, + 407 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we first define bias attributes and bias classes (Sec. 3.1), and formulate the problem of identifying biases of pre-trained models from their task specifications (Sec. 3.2). We then describe C2B, that proposes a list of biases (Sec. 3.3) and collects data for testing them (Sec. 3.4). Finally, we describe how we measure model's biases from the collected data (Sec. 3.5).", + "bbox": [ + 89, + 417, + 483, + 523 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Bias definition", + "text_level": 1, + "bbox": [ + 89, + 534, + 236, + 549 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Formally, we denote the target classification model as $f: \\mathcal{X} \\to \\mathcal{Y}$ , mapping an image in the space $\\mathcal{X}$ to labels in the space $\\mathcal{Y}$ . We define as bias attributes $\\mathcal{B} = \\{\\mathbf{B}_1, \\dots, \\mathbf{B}_N\\}$ the different categories of bias, i.e., factors that can impact its performance [44]. For each bias attribute, we define a list of bias classes $\\mathbf{B}_i = \\{b_{i,1}, \\dots, b_{i,n_i}\\}$ , the different values that each bias attribute can take. For a given target class, we define the model to be biased or unfair w.r.t specific bias classes depending on the per-class accuracy difference of the model. Let us denote the per-class accuracy of a model $f$ on a target class $y$ on data that matches a bias class $b_{i,j}$ as $A_y(f, b_{i,j})$ . We define the bias score $\\phi_{y,i,j}$ for each bias class $b_{i,j}$ as", + "bbox": [ + 89, + 558, + 482, + 755 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\phi_ {y, i, j} = A _ {y} \\left(f, b _ {i, j}\\right) - \\frac {1}{n _ {i} - 1} \\sum_ {b _ {i, k} \\in \\mathbf {B} _ {i}, k \\neq j} A _ {y} \\left(f, b _ {i, k}\\right). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 766, + 482, + 804 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Problem formulation", + "text_level": 1, + "bbox": [ + 89, + 816, + 287, + 830 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given a pre-trained model $f$ , our goal is to identify its biases. We assume $f$ to be a generic classification model that we did not develop and that we aim to use (e.g., downloaded from the web). Thus, we also assume to have no access to", + "bbox": [ + 89, + 839, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "additional data $f$ had access to (e.g., the dataset set used for training). As we want to avoid costly annotation procedures, we also assume to have no access to any dataset containing labeled images of the classes in $\\mathcal{V}$ .", + "bbox": [ + 511, + 90, + 903, + 151 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To address this problem, we can rely only on what we know about the model: i.e., the task it has been designed for, the input domain it expects, and the output it predicts. This information can all be expressed using simple, natural language. Specifically, we assume to have a general description of the model's task $T$ , the input the model expects to process, and its output (i.e., classes in $\\mathcal{V}$ ).", + "bbox": [ + 511, + 152, + 905, + 258 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This task is challenging as (i) detecting biases requires measuring the performance of the model on a labeled dataset, potentially containing annotations for the factors of interest and (ii) we have no pre-defined list of biases to start with. This implies that we have to both generate a list of potential biases and to collect a dataset to test them. In the following, we will describe our framework, C2B, that exploits LLMs for proposing biases, and a combination of LLMs and multimodal retrieval for scoring them. An overview of C2B is illustrated in Fig. 2.", + "bbox": [ + 511, + 261, + 906, + 412 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.3. Generating bias proposals", + "text_level": 1, + "bbox": [ + 511, + 428, + 750, + 445 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The first step of C2B is to generate a list of bias attributes and classes that we can explicitly test for. To generate the list, we cannot rely on attribute annotations, as we have no access to labeled data. Moreover, we cannot use lists available online, as biases depend on the task of interest.", + "bbox": [ + 511, + 453, + 905, + 529 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To address this problem, we exploit the reasoning capabilities of LLMs. Specifically, let us denote as $\\Phi_{\\mathrm{LLM}}$ a generic LLM, mapping textual inputs to textual outputs, i.e., $\\Phi_{\\mathrm{LLM}}: \\mathcal{T} \\to \\mathcal{T}$ , with $\\mathcal{T}$ being the space of natural language. To generate the candidate set of biases, we prompt the LLM to produce a list of bias attributes and classes for the given model and task description. Because different classes may have different biases, we query the LLM to give us a list of potential biases $\\mathcal{B}^y$ for each class $y \\in \\mathcal{Y}$ . Formally, we have:", + "bbox": [ + 511, + 531, + 905, + 681 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {B} ^ {y} = \\Phi_ {\\mathrm {L L M}} \\left(\\pi_ {\\text {b i a s}} \\circ T \\circ y\\right) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 620, + 698, + 903, + 715 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\circ$ is the string concatenation operation, $\\pi_{\\mathrm{bias}}$ is the system prompt that instructs the LLM about the task it should perform (Supp. Mat., Appx. I), and $T$ is the model's task description, including input/output specification.", + "bbox": [ + 511, + 731, + 905, + 792 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The output of Eq. (2), $\\mathcal{B}^y$ , is a list1 of bias attributes $\\mathcal{B}^y = \\{\\mathbf{B}_1^y, \\dots, \\mathbf{B}_{N_y}^y\\}$ , of size $N_y$ . We prompt the LLM so that each bias attribute is associated to a list of $n_{y,i}$ bias classes $\\mathbf{B}_i^y = \\{b_{i,1}^y, \\dots, b_{i,n_{y,i}}^y\\}$ .", + "bbox": [ + 511, + 792, + 903, + 857 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "1For simplicity, we assume that the natural language output of the LLM is already formatted as a list.", + "bbox": [ + 511, + 875, + 903, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "15153", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/5f6973cc196021fad07be63facd83e17b2783b016e6560a09857c8222cbcf2c5.jpg", + "image_caption": [ + "Figure 2. Overview of C2B. Given a specific task with an associated description and a pre-trained classification model, our approach leverages an LLM to identify potential candidate bias attributes and corresponding classes. These candidate biases are then used to prompt the LLM to generate captions, which are subsequently utilized by a retrieval module to collect a dataset of images for bias testing. This dataset enables evaluation of the target classification model, producing scores for each identified bias category." + ], + "image_footnote": [], + "bbox": [ + 89, + 85, + 911, + 345 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Collecting data", + "text_level": 1, + "bbox": [ + 89, + 435, + 243, + 450 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given the list of candidate bias attributes and classes, we still need a dataset to score them. Ideally, for each target class $y \\in \\mathcal{V}$ and bias class $b^{y} \\in \\mathbf{B}^{y}$ , with $\\mathbf{B}^{y} \\in \\mathcal{B}^{y}$ , we would like to have a dataset $D(b^{y}, y)$ depicting images of class $y$ exhibiting the bias class $b^{y}$ , i.e., $D(b^{y}, y) = \\{(x_{i}, y_{i}, b_{i}^{y})\\}_{i=1}^{M}$ of size $M$ . Unfortunately, we cannot rely on labeled data for this task as (i) we do not have access to it; (ii) biases change w.r.t. the particular application, thus it is very hard to collect beforehand a dataset containing all possible target and bias classes of interest.", + "bbox": [ + 89, + 462, + 483, + 613 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To address this problem, we resort to external sources. In particular, we assume to have access to a large-scale database $\\mathcal{D}$ containing images. Note that $\\mathcal{D}$ has no annotations: we just assume it to be large-scale (i.e., potentially web-scale). From this database, we can collect images for each target-bias combination of interest using a text-to-image retrieval model. This is achieved in two steps: captions generation and cross-modal retrieval.", + "bbox": [ + 89, + 618, + 483, + 739 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Caption generation. A straightforward solution to create captions is to use a fixed template (e.g., an image of [target class] with [bias class]). However, this would not consider the specificity of the considered task and thus may lead to retrieving erroneous images. For instance, the biases and captions associated to a facial attribute classification task should be different from those in a natural scene classification task. To avoid this issue, we resort to using the LLM, querying it in two steps: first to create a template for the specific task, and then to gener", + "bbox": [ + 89, + 750, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ate captions specific for target and bias class. Formally, we obtain the initial template as:", + "bbox": [ + 511, + 436, + 906, + 465 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {t e m p} = \\Phi_ {\\mathrm {L L M}} \\left(\\pi_ {\\text {t e m p l a t e}} \\circ T\\right) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 612, + 478, + 906, + 494 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\pi_{\\mathrm{template}}$ is the system prompt that instructs the LLM about the task it should perform (Supp. Mat., Appx. 1). Then, we ask the LLM to refine this template for each specific target class $y$ and bias class $b$ , as follows:", + "bbox": [ + 511, + 506, + 905, + 566 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nc _ {b} ^ {y} = \\Phi_ {\\mathrm {L L M}} \\left(\\pi_ {\\text {c a p t i o n}} \\circ \\operatorname {t e m p} \\circ b \\circ y\\right). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 577, + 903, + 595 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $c_b^y$ is the resulting caption. Note that, while with this strategy we obtain one caption per target-bias class pair, in principle we can also generate multiple captions per pair to have more variability, beneficial for more complex tasks.", + "bbox": [ + 511, + 604, + 905, + 666 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Image retrieval. The obtained captions can be used to retrieve images from a large scale image database, and we implement the retrieval by using a contrastive vision-language model (VLM) that can score the similarity between a text and image. Formally, we can decompose the VLM in three elements: a visual encoder $\\varphi_{\\mathrm{vis}}:\\mathcal{X}\\to \\mathbb{R}^d$ a text encoder $\\varphi_{\\mathrm{txt}}:\\mathcal{T}\\rightarrow \\mathbb{R}^d$ , and a similarity function $\\sin :\\mathbb{R}^d\\times \\mathbb{R}^d\\to \\mathbb{R}$ (e.g., cosine similarity for CLIP [51]). Given a caption, we can retrieve a set of images $\\mathbf{X}_b^y$ for each class label $y$ and bias $b$ as follows:", + "bbox": [ + 511, + 672, + 906, + 823 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {X} _ {b} ^ {y} = \\underset {x \\in \\mathcal {D}} {\\operatorname {t o p - k}} \\operatorname {s i m} \\left(\\varphi_ {\\text {v i s}} (x), \\varphi_ {\\text {t x t}} \\left(c _ {b} ^ {y}\\right)\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 580, + 834, + 903, + 859 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where we retrieve the $k$ images with the highest matching score. While these images do not have a ground-truth,", + "bbox": [ + 511, + 869, + 905, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "15154", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "they can still help to estimate the biases of the model, with the fidelity of the latter depending on the retrieval accuracy (Sec. 4.4). We also experiment with a different type of retrieval system, a Web search engine (details in Sec. 4.1).", + "bbox": [ + 89, + 90, + 480, + 151 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Measuring biases", + "text_level": 1, + "bbox": [ + 89, + 162, + 261, + 176 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the previous section, we have collected a dataset with pseudo-annotations for all bias values and classes of interest. We can now test the target model on this dataset to compute its performance, using the variation across bias classes to score each bias. We use the formula defined in Eq. (1) to score each target-bias pair relative to other target-bias pairs. For each bias class, we obtain a bias score between -1 and 1. If the per-class accuracy is higher for a given bias class, the classifier is biased toward the bias class (positive bias score). Reciprocally, if the per-class accuracy is lower for a given bias class, the classifier is biased against the bias class (negative bias score). Intuitively, larger absolute bias score values also correspond to stronger biases.", + "bbox": [ + 89, + 185, + 482, + 380 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To quantitatively evaluate the quality of the biases detected by C2B, we propose two novel evaluation settings that complement each other. First, in Sec. 4.2.1, we propose to match the detected biases with ground-truth annotations using SBERT embeddings for tasks where datasets with additional bias annotations exist. Second, in Sec. 4.2.2, we propose to assign bias labels (based on the initial proposed biases) to ground-truth task-specific data with a VQA model, to check the agreement between the bias detection method and an expert VQA model.", + "bbox": [ + 89, + 382, + 482, + 532 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Collecting scores for all bias attributes and classes enables various types of analyses to explore the biases. Biases can be ranked by score, to assess which attributes are more critical than others. Bias class scores can also be aggregated across bias attributes, to measure which bias attributes exhibit a larger bias. We provide examples of qualitative analyses in Sec. 4.3, and more in the Supp. Mat., Appx. D.", + "bbox": [ + 89, + 534, + 482, + 638 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 654, + 220, + 671 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we present the results of our experiments, where we apply C2B to detect biases for different models on two different image classification tasks. We first describe our experimental protocol (Sec. 4.1), we then quantitatively compare the biases detected by C2B with a state-of-the-art competitor (Sec. 4.2), and we provide qualitative examples of detected biases (Sec. 4.3). Finally, we analyze the core component of our approach (Sec. 4.4).", + "bbox": [ + 89, + 679, + 482, + 800 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental protocol", + "text_level": 1, + "bbox": [ + 89, + 809, + 299, + 825 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Tasks and models. We test C2B on two different tasks: facial attribute classification and image classification.", + "bbox": [ + 89, + 839, + 482, + 869 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Facial attribute classification. We consider the CelebA dataset [41], which is composed of 200k images of celebrity", + "bbox": [ + 89, + 869, + 482, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "faces annotated with 40 facial attributes, such as male, blond hair, young, or high cheekbones. To evaluate C2B, we choose to detect the biases of a recent state-of-the-art model, FaceXFormer [48]. This model relies on a pretrained Swin/B [42] backbone, coupled with an MLP-based fusion module and an attention-based decoder, which are co-trained on several face datasets.", + "bbox": [ + 511, + 90, + 903, + 196 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Image classification. We focus on the widely used ImageNet-1K [54] dataset, where each image is associated to one of the 1000 ImageNet classes. We choose to detect the biases of four widely used models, easily accessible through TorchVision [43], comprising both convolutional networks and transformers: ResNet50_V2 [23], ResNet101_V2 [23], ResNet152_V2 [23], ViT_B_16_SWAG_E2E_V1 [58].", + "bbox": [ + 511, + 198, + 903, + 319 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Competitors. As we introduce the new problem of unsupervised bias detection relying solely on a textual description of the classification task, there are no existing baseline methods designed for this specific setting. Nevertheless, we still compare C2B to a recent supervised open-set bias detection method, Bias-to-Text (B2T) [33]. However, unlike C2B which does not require task-specific labeled data to identify biases in an image classifier, B2T depends on ground-truth task labels to identify model failures. It then employs a pre-trained captioning model to generate descriptions of misclassified images, extracts common keywords, and computes a bias score based on the difference between (i) the similarity (CLIP score) between the discovered bias keyword and misclassified images, and (ii) the similarity between the discovered bias keyword and well-classified images. Thus, we test B2T using the labeled validation sets of the two datasets (i.e., CelebA and ImageNet).", + "bbox": [ + 511, + 325, + 906, + 583 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation details. C2B comprises two components: the LLM that proposes biases and captions, and the retrieval module. For the former, we use Llama 3.1 8B [18] (see Supp. Mat., Appx. J for qualitative LLM comparisons). For the retrieval, we explore two strategies: (i) CC12M: we consider the large Conceptual Captions 12M (CC12M) dataset [12] and retrieve images based on CLIP embeddings [7], and (ii) BING: instead of using a VLM to retrieve images from an existing database where some image domains may not be well-represented, we propose to retrieve images from the Web using a publicly available search engine, Bing. For both datasets, we retrieve 20 images per caption, and we measure biases as in Eq. (1), comparing the difference in the model's per-class accuracy between images that match different candidate bias classes.", + "bbox": [ + 511, + 590, + 905, + 816 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Quantitative evaluation", + "text_level": 1, + "bbox": [ + 511, + 830, + 730, + 847 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we provide a quantitative evaluation to assess the quality of biases identified by C2B. First, we present results based on ground-truth bias annotations, i.e.,", + "bbox": [ + 511, + 854, + 903, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "15155", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "exploiting attributes annotated in the validation sets of the datasets. For biases not in the ground-truth annotations, we introduce an evaluation based on Visual Question Answering (VQA) to measure the relevance of the proposed biases.", + "bbox": [ + 89, + 90, + 483, + 151 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2.1 Evaluation with ground-truth annotations", + "text_level": 1, + "bbox": [ + 89, + 174, + 437, + 189 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Setting. In this setting, we exploit the available attribute annotations, checking if models can detect and quantify those biases correctly. In the case of CelebA, when we focus on a specific attribute classification task (e.g., blonde hair) we consider all other 39 attributes (e.g., young) as biases measurable with the ground-truth. For ImageNet-1K, we use ImageNet-X [28], i.e., the validation set of ImageNet-1K with 16 labeled \"factors\" distinguishing images from typical prototypes (e.g., orientation), as ground-truth biases.", + "bbox": [ + 88, + 205, + 482, + 340 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To obtain ground-truth biases, we test the model on each validation set and check the difference in per-class accuracy among images where the bias attribute is present and images where it is absent. We consider a ground-truth bias to be present if this difference is higher than a threshold (i.e., $\\tau = 0.05$ ) using the same threshold to assess bias identification on both C2B and B2T scores. We refer to the sign of the bias score as the direction of the bias. We provide more results with a varying threshold in the Supp. Mat., Appx. C.", + "bbox": [ + 89, + 342, + 483, + 478 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Due to the open-set nature of the detected biases for both C2B and B2T, there are almost no exact matches between them and ground-truth biases. To compare the two sets, we compute the cosine similarity between SBERT embeddings [52] of the ground-truth and detected biases. We empirically set the similarity threshold to 0.9 to avoid having too many false positives. We provide more details and qualitative examples of matches in the Supp. Mat., Appx. K.", + "bbox": [ + 89, + 479, + 483, + 599 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Metrics. We report results from two complementary perspectives: GT $\\rightarrow$ Detected and Detected $\\rightarrow$ GT. In GT $\\rightarrow$ Detected, we start from each ground-truth bias and check whether it was correctly detected with the same bias direction (HIT), detected with the opposite direction (false hit, FH), or not detected (M1ss). This perspective focuses on recall. However, methods that propose many biases may artificially reduce misses. Therefore, we also report Detected $\\rightarrow$ GT, where we start from each detected bias and verify whether it matches a ground-truth bias with the correct direction. This second view measures precision and helps assess the quality of the proposed biases.", + "bbox": [ + 89, + 606, + 483, + 787 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Results. In Tabs. 1 and 2, we report the results of our analysis. On CelebA (left side of Tab. 1, GT $\\rightarrow$ Detected), C2B outperforms B2T in detecting ground-truth biases, regardless of the retrieval source. We obtain the best results with Bing, detecting between $10.8\\%$ and $12.3\\%$ of the ground-truth biases, compared to only $4.5\\%$ for B2T. However, B2T exhibits a lower proportion of false hits $(3.1\\%)$ , whereas", + "bbox": [ + 89, + 794, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/9fa8d96b5edb8bcc4fe12537b00ffb3b2ccff3a4141fb73ed243cd60ce7d1d8c.jpg", + "table_caption": [ + "Table 1. Proportion $(\\%)$ of ground-truth biases detected on CelebA $\\mathrm{(GT\\to Detected)}$ and of detected biases corresponding to groundtruth ones (Detected $\\rightarrow$ GT). FH=False Hit." + ], + "table_footnote": [], + "table_body": "
METHODGT → DETECTEDDETECTED → GT
HIT(↑)FH(↓)MISS (↓)HIT(↑)FH(↓)MISS(↓)
FaceXFormer
B2T [33]4.533.1492.326.594.1089.29
C2B (BING)12.296.8880.8314.187.3478.48
C2B (CC12M)10.767.7581.4912.758.1579.09
", + "bbox": [ + 519, + 142, + 901, + 219 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/7b5e0cb2fb6bbc9550aa18decfdcaa8c789b486c7c57652a33ba0dee9cbc1858.jpg", + "table_caption": [ + "Table 2. Proportion (\\%) of ground-truth biases detected on ImageNet-X (GT $\\rightarrow$ Detected) and of detected biases corresponding to ground-truth ones (Detected $\\rightarrow$ GT). FH=False Hit." + ], + "table_footnote": [], + "table_body": "
METHODGT → DETECTEDDETECTED → GT
HIT(↑)FH(↓)Miss (↓)HIT(↑)FH (↓)Miss(↓)
ResNet50_V2
B2T [33]2.402.1995.410.800.7198.49
C2B (BING)7.808.5383.662.602.9894.42
C2B (CC12M)11.1811.3077.522.992.8794.14
ResNet101_V2
B2T [33]2.552.1195.330.840.6998.47
C2B (BING)7.918.4183.682.562.7794.67
C2B (CC12M)11.2111.6877.112.762.9694.28
ResNet152_V2
B2T [33]2.531.7295.750.830.5898.59
C2B (BING)7.537.9384.542.752.7994.46
C2B (CC12M)11.1012.0176.902.812.9594.24
ViT_B_16_SWAG
B2T [33]2.111.9895.900.720.6798.61
C2B (BING)7.727.1885.102.482.4895.05
C2B (CC12M)10.8511.0778.082.642.7094.66
", + "bbox": [ + 519, + 286, + 903, + 489 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "C2B shows slightly higher false hit rates (6.9–7.8%). This is largely due to retrieval noise and the broader search space introduced by LLM-generated proposals, which can sometimes lead to less precise bias candidates. Nevertheless, C2B misses fewer ground-truth biases (80.8–81.5% vs. B2T's 92.3%). These trends remain consistent when considering the proportion of detected biases that correspond to ground-truth biases (Detected → GT), where C2B achieves 12.8–14.2%, compared to 6.6% for B2T.", + "bbox": [ + 511, + 515, + 906, + 651 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "On ImageNet-X (Tab. 1), across four classifiers, C2B again detects more ground-truth biases than B2T (7.5–11.2% vs. 2.1–2.6%), at the cost of higher false hit rates (7.2–12.0% vs. 1.7–2.2%). As before, these false hits are attributable to noisy retrieval and the open-ended nature of bias proposals. Still, C2B misses significantly fewer ground-truth biases (76.9–85.1% vs. 95.3–95.9%). Interestingly, on ImageNet-X, retrieving from CC12M provides better results than Bing. The Detected → GT evaluation confirms these findings, with a larger fraction of C2B-detected biases aligning with ground-truth ones (2.5–3.0% vs. 0.6–0.7% for B2T).", + "bbox": [ + 511, + 652, + 908, + 833 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Discussion. Due to the limited human annotations available in CelebA and ImageNet-X, the proportion of matching biases is low across all methods, particularly on ImageNet-X. Furthermore, since we are using text embeddings to", + "bbox": [ + 511, + 839, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "15156", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/940d460151245fe75fad876c287a1a9bf795a2e1a396f8848581e8fe281fc8a3.jpg", + "table_caption": [ + "Table 3. Detected biases and VQA agreement on CelebA." + ], + "table_footnote": [], + "table_body": "
MODELB2T [33]C2B (BING)C2B (CC12M)
FaceXFormer0.100.220.25
", + "bbox": [ + 129, + 114, + 444, + 148 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/662c77f6608c654a85818b79a540e8045e9e2fee69393b800ae973222e84a073.jpg", + "table_caption": [ + "Table 4. Detected biases and VQA agreement on ImageNet-X." + ], + "table_footnote": [], + "table_body": "
MODELB2T [33]C2B (BING)C2B (CC12M)
ResNet50_V20.200.260.20
ResNet101_V20.190.280.20
ResNet152_V20.200.280.20
ViT_B_16_SWAG0.200.320.21
", + "bbox": [ + 127, + 188, + 446, + 253 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "match the biases, the numbers are also influenced by the semantic proximity between LLM-proposed biases, caption-extracted keywords, and ground-truth annotations. Moreover, a high number of Miss in the Detected $\\rightarrow$ GT case is not necessarily a bad thing, as C2B and B2T may be detecting accurate biases that are not present in the annotation. For these reasons, in the next section we introduce a different evaluation based on visual question answering.", + "bbox": [ + 89, + 281, + 483, + 402 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2.2 VQA-based evaluation", + "text_level": 1, + "bbox": [ + 89, + 422, + 302, + 436 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As ground-truth bias annotations may not always be available, we also propose an alternative evaluation scheme based on VQA, measuring the quality of the proposed biases that are not present in ground-truth annotations.", + "bbox": [ + 89, + 448, + 483, + 508 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Metric. We use a VQA model (LLaVA-1.5-13B [38, 39]) to pseudo-label each image of the validation sets with each of the proposed bias by C2B or B2T. Additional details can be found in the Supp. Mat., Appx. G. From these pseudolabels, we can use the ground-truth task labels (i.e., class annotations) to check how the per-class accuracy varies over different bias classes, using the same threshold $\\tau = 0.05$ of Sec. 4.2.1. For each bias class $b$ proposed by the model, there are three cases: a positive bias is detected $(\\phi_b \\geq \\tau)$ , a negative bias is detected $(\\phi_b \\leq -\\tau)$ , or the bias is not detected $(|\\phi_b| < \\tau)$ . For each bias class, we define the agreement score to be 1 if both scores fall in the same case, -1 if they detect opposite biases, and 0 otherwise.", + "bbox": [ + 89, + 515, + 483, + 712 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results. Tabs. 3 and 4 present the results of our VQA-based evaluation on CelebA and ImageNet-X. On CelebA, C2B achieves higher agreement with VQA pseudo-labels (0.22 for Bing and 0.25 for CC12M) than B2T (0.10), indicating that the biases discovered by C2B are more aligned with visually detectable and semantically coherent bias attributes. On ImageNet-X, while C2B often achieves higher agreement scores than B2T with Bing (i.e., +0.09 on average), results with CC12M are either the same or comparable to those of B2T (i.e., +0.005 on average). We attribute this to the challenge of relying on retrieval-based data collection, without access to annotated failure cases or", + "bbox": [ + 89, + 719, + 483, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "task-specific captions. In contrast, B2T benefits from captions derived from known misclassifications. Despite this more challenging and fully unsupervised setting, C2B still achieves agreements that are competitive with or superior to B2T. This confirms that C2B's proposals are not only meaningful and visually grounded but also relevant for bias detection in settings where no annotations are available.", + "bbox": [ + 511, + 90, + 905, + 196 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Discussion. The results presented in this section are influenced by both the types of questions posed to the VQA model and the model's own inherent biases. Given these limitations, evaluations based on ground-truth annotations and those derived from VQA outputs serve as complementary approaches. Overall, biases proposed by the LLM tend to offer more relevant bias indicators than keywords extracted from captions as in B2T [33], allowing C2B to detect a greater proportion of ground-truth biases and achieve higher agreement with the VQA model than B2T.", + "bbox": [ + 511, + 204, + 905, + 354 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Qualitative evaluation", + "text_level": 1, + "bbox": [ + 511, + 368, + 718, + 383 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we show examples of biases discovered by C2B for different models on different tasks.", + "bbox": [ + 511, + 392, + 903, + 421 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In Fig. 3, we show biases that C2B detects for the high cheekbones attribute on face attribute classification, with bias scores from different retrieval methods. We can see that a strong bias was detected on the expression attribute: the classifier appears to perform better when people are smiling and worse when they appear angry. Other smaller biases were detected, such as young, and male.", + "bbox": [ + 511, + 422, + 905, + 529 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Fig. 4 shows a similar analysis for biases detected by C2B for the minibus class on image classification for the ResNet50_V2 model. Strong biases were detected on the lighting, orientation, and presence of reflection bias attribute: the classifier appears stronger on daytime images, portrait-oriented images, and images without reflections, while performing worse on nighttime images, landscape-oriented images, and images with reflections. Finally, Fig. 5 depicts biases for the birdhouse class again for image classification, but for the ViT_B_16_SWAG model. A strong bias is found on the shape bias attribute: the classifier seems to be performing better on images of rectangular birdhouses, while it underperforms on images of spherical birdhouses.", + "bbox": [ + 511, + 531, + 905, + 727 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Evaluation of VLM-based retrieval", + "text_level": 1, + "bbox": [ + 511, + 739, + 818, + 753 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Retrieving images matching the target classes and biases present in the captions is essential for the detected biases to be trustworthy. Therefore, in this section, we analyze the CLIP-based retrieval approach used by C2B, by trying to retrieve images for which we have annotations. For a VQA-based analysis of both the Bing-based retrieval and CLIP-based retrieval on CC12M, please refer to Supp. Mat., Appx. F. To perform our analysis, we use CLIP to retrieve images from the validation sets of CelebA and ImageNet-", + "bbox": [ + 511, + 763, + 905, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "15157", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/517b6ddb58ba735e614704926cceb229630dc2224ccf7d35b1a7646f0150073c.jpg", + "image_caption": [ + "Figure 3. Examples of biases found by C2B for the high cheekbones target attribute on CelebA." + ], + "image_footnote": [], + "bbox": [ + 98, + 90, + 903, + 255 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1005d94e80770f02c61266a1f5476eed61590f5c79e0f5812e1d8c7849ac584b.jpg", + "image_caption": [ + "Figure 4. Examples of biases found by C2B for the minibus class on ImageNet-X (ResNet50_V2)." + ], + "image_footnote": [], + "bbox": [ + 96, + 306, + 480, + 450 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/40b395fe3a813bd492853383acccf65cabe9055b7085a593f2b01416b297e14a.jpg", + "image_caption": [ + "Figure 5. Examples of biases found by C2B for the birdhouse class on ImageNet-X (ViT_B_16_SWAG)." + ], + "image_footnote": [], + "bbox": [ + 94, + 513, + 480, + 661 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "X, using handcrafted captions as input: combinations of attributes on CelebA, and classes and factors on ImageNet-X.", + "bbox": [ + 89, + 733, + 482, + 763 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Metric. We use recall@ $K$ , where $K$ varies for each caption to match the number of corresponding ground-truth samples in the dataset. We show results using recall @ $A \\cdot K$ , where $A$ is a coefficient that varies between 0.01 and 1.0.", + "bbox": [ + 89, + 771, + 482, + 830 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Results. We present the results in Fig. 6. For CelebA (left), the recall is the same for both target and bias attributes, as expected, since both are drawn from the same set of attributes. On the same dataset, CLIP struggles to", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/69ace24a8a203f0e7f6e92391a298ddf676ffe6c72f6e9d9589910b3f69c440c.jpg", + "image_caption": [ + "Figure 6. Accuracy of VLM-based retrieval." + ], + "image_footnote": [], + "bbox": [ + 516, + 304, + 715, + 444 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5e0928000fd6dc910aaa55266bbb7847b560712dce1ead30e13f9b66bad03d5f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 717, + 304, + 901, + 443 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "retrieve images with combinations of attributes, with a recall @ $0.01 \\cdot K$ just above $40\\%$ . For ImageNet-X (right), CLIP is fairly accurate, especially for low values of $A$ , with more than $80\\%$ and $70\\%$ recall at $0.05 \\cdot K$ to retrieve images with specific biases and classes, respectively. Generally, the recall is higher when the number of retrieved images decreases relative to the number of potential matching images present in the dataset, which suggests that, the larger the dataset, the more accurate the retrieval is.", + "bbox": [ + 511, + 502, + 906, + 638 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 660, + 633, + 676 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We introduced CLASSIFIER-TO-BIAS (C2B), a novel method for automatically detecting biases in visual classifiers entirely in an unsupervised manner, without the need for task-specific labeled data. Starting from a simple textual description of the classification task and a pre-trained model, C2B generates bias hypotheses, retrieves relevant images to test them, and produces a list of biases with their scores. Although C2B is the first approach to operate in such challenging setting, our results show that it surpasses a recent supervised open-set bias detector across two evaluation settings: one based on ground-truth annotations and another on VQA-based pseudo-labels. We believe our work marks an important step toward automated, open-set, and task-agnostic bias detection.", + "bbox": [ + 511, + 688, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "15158", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 91, + 90, + 250, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was sponsored by the MUR PNRR project FAIR - Future AI Research (PE00000013), funded by NextGeneration, and the Italian Ministerial grants PRIN 2022: \"B-FAIR: Bias-Free Artificial Intelligence methods for automated visual Recognition\" (CUP E53D23008010006), and the EU project ELIAS (No.01120237).", + "bbox": [ + 89, + 114, + 483, + 205 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 217, + 187, + 233 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Marah Abdin, Sam Ade Jacobs, Ammar Ahmad Awan, Jyoti Aneja, Ahmed Awadallah, Hany Awadalla, Nguyen Bach, Amit Bahree, Arash Bakhtiari, Harkirat Behl, et al. Phi-3 technical report: A highly capable language model locally on your phone. arXiv preprint arXiv:2404.14219, 2024. 15", + "[2] Julius A Adebayo et al. *FairML: ToolBox for diagnosing bias in predictive modeling*. PhD thesis, Massachusetts Institute of Technology, 2016. 1", + "[3] Chirag Agarwal, Daniel D'souza, and Sara Hooker. Estimating example difficulty using variance of gradients. In CVPR, 2022. 2", + "[4] Maximilian Augustin, Valentyn Boreiko, Francesco Croce, and Matthias Hein. Diffusion visual counterfactual explanations. NeurIPS, 2022. 2", + "[5] Razvan Azamfirei, Sapna R. Kudchadkar, and James Fackler. Large language models and the perils of their hallucinations. Critical Care, 27(1):120, 2023. 2", + "[6] Yujia Bao and Regina Barzilay. Learning to split for automatic bias detection. arXiv preprint arXiv:2204.13749, 2022. 1", + "[7] Romain Beaumont. Clip retrieval: Easily compute clip embeddings and build a clip retrieval system with them. https://github.com/rom1504/clip-retrieval, 2022.5", + "[8] R. K. E. Bellamy, K. Dey, M. Hind, S. C. Hoffman, S. Houde, K. Kannan, P. Lohia, J. Martino, S. Mehta, A. Mojsilovic, S. Nagar, K. Natesan Ramamurthy, J. Richards, D. Saha, P. Sattigeri, M. Singh, K. R. Varshney, and Y. Zhang. Ai fairness 360: An extensible toolkit for detecting and mitigating algorithmic bias. IBM Journal of Research and Development, 63(4/5):4:1-4:15, 2019. 1", + "[9] Tolga Bolukbasi, Kai-Wei Chang, James Zou, Venkatesh Saligrama, and Adam Kalai. Man is to computer programmer as woman is to homemaker? debiasing word embeddings. In NeurIPS, 2016. 1", + "[10] Joy Buolamwini and Timnit Gebru. Gender shades: Intersectional accuracy disparities in commercial gender classification. In Proceedings of the 1st Conference on Fairness, Accountability and Transparency, 2018. 1", + "[11] Angel Alexander Cabrera, Will Epperson, Fred Hohman, Minsuk Kahng, Jamie Morgenstern, and Duen Horng Chau. Fairvis: Visual analytics for discovering intersectional bias in machine learning. In 2019 IEEE Conference on Visual Analytics Science and Technology (VAST), pages 46-56, 2019.", + "[12] Soravit Changpinyo, Piyush Sharma, Nan Ding, and Radu Soricut. Conceptual 12M: Pushing web-scale image-text" + ], + "bbox": [ + 93, + 243, + 482, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "pre-training to recognize long-tail visual concepts. In CVPR, 2021.5", + "[13] Davide Chicco, Niklas Totsch, and Giuseppe Jurman. The matthews correlation coefficient (mcc) is more reliable than balanced accuracy, bookmaker informedness, and markedness in two-class confusion matrix evaluation. BioData Mining, 14(1):13, 2021. 11", + "[14] Massimiliano Ciranni, Luca Molinaro, Carlo Alberto Barbano, Attilio Fiandrotti, Vittorio Murino, Vito Paolo Pastore, and Enzo Tartaglione. Say my name: a model's bias discovery framework. arXiv preprint arXiv:2408.09570, 2024. 2", + "[15] Gabriela Csurka, Tyler L Hayes, Diane Larlus, and Riccardo Volpi. What could go wrong? discovering and describing failure modes in computer vision. In ECCV-WS, 2024. 3", + "[16] Greg d'Eon, Jason d'Eon, James R Wright, and Kevin Leyton-Brown. The spotlight: A general method for discovering systematic errors in deep learning models. In Proceedings of the 2022 ACM Conference on Fairness, Accountability, and Transparency, pages 1962-1981, 2022. 1, 2", + "[17] Moreno D'Inca, Elia Peruzzo, Massimiliano Mancini, Dejia Xu, Vedit Goel, Xingqian Xu, Zhangyang Wang, Humphrey Shi, and Nicu Sebe. Openbias: Open-set bias detection in text-to-image generative models. In CVPR, 2024. 1, 2", + "[18] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 5, 15", + "[19] Sabri Eyuboglu, Maya Varma, Khaled Kamal Saab, Jean-Benoit Delbrouck, Christopher Lee-Messer, Jared Dunnmon, James Zou, and Christopher Re. Domino: Discovering systematic errors with cross-modal embeddings. In ICLR, 2022. 1, 2", + "[20] Samir Yitzhak Gadre, Gabriel Ilharco, Alex Fang, Jonathan Hayase, Georgios Smyrnis, Thao Nguyen, Ryan Marten, Mitchell Wortsman, Dhruba Ghosh, Jieyu Zhang, Eyal Orgad, Rahim Entezari, Giannis Daras, Sarah Pratt, Vivek Ramanujan, Yonatan Bitton, Kalyani Marathe, Stephen Mussmann, Richard Vencu, Mehdi Cherti, Ranjay Krishna, Pang Wei W Koh, Olga Saukh, Alexander J Ratner, Shuran Song, Hannaneh Hajishirzi, Ali Farhadi, Romain Beaumont, Sewoong Oh, Alex Dimakis, Jenia Jitsev, Yair Carmon, Vaishaal Shankar, and Ludwig Schmidt. Datacomp: In search of the next generation of multimodal datasets. In Advances in Neural Information Processing Systems, pages 27092-27112, 2023. 2", + "[21] Isabel O. Gallegos, Ryan A. Rossi, Joe Barrow, Md Mehrab Tanjim, Sungchul Kim, Franck Dernoncourt, Tong Yu, Ruiyi Zhang, and Nesreen K. Ahmed. Bias and fairness in large language models: A survey. Computational Linguistics, 50 (3):1097-1179, 2024. 2", + "[22] Irena Gao, Gabriel Ilharco, Scott Lundberg, and Marco Tulio Ribeiro. Adaptive testing of computer vision models. In ICCV, 2023. 1, 3", + "[23] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 5" + ], + "bbox": [ + 516, + 93, + 903, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "15159", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[24] Lisa Anne Hendricks, Kaylee Burns, Kate Saenko, Trevor Darrell, and Anna Rohrbach. Women also snowboard: Overcoming bias in captioning models. In ECCV, 2018. 1", + "[25] Lisa Anne Hendricks, Ronghang Hu, Trevor Darrell, and Zeynep Akata. Generating counterfactual explanations with natural language. In ICML-WS, 2018. 2", + "[26] Cheng-Yu Hsieh, Jieyu Zhang, Zixian Ma, Aniruddha Kembhavi, and Ranjay Krishna. Sugarcrepe: Fixing hackable benchmarks for vision-language compositionality. NeurIPS, 2024. 2", + "[27] https://huggingface.co/. 1", + "[28] Badr Youbi Idrissi, Diane Bouchacourt, Randall Balestriero, Ivan Evtimov, Caner Hazirbas, Nicolas Ballas, Pascal Vincent, Michal Drozdjal, David Lopez-Paz, and Mark Ibrahim. Imagenet-x: Understanding model mistakes with factor of variation annotations. In ICLR, 2023. 1, 2, 6", + "[29] Saachi Jain, Hadi Salman, Eric Wong, Pengchuan Zhang, Vibhav Vineet, Sai Vermprala, and Aleksander Madry. Missingness bias in model debugging. In ICLR, 2022. 2", + "[30] Saachi Jain, Hannah Lawrence, Ankur Moitra, and Aleksander Madry. Distilling model failures as directions in latent space. In ICLR, 2023. 1, 2", + "[31] Ziwei Ji, Nayeon Lee, Rita Frieske, Tiezheng Yu, Dan Su, Yan Xu, Etsuko Ishii, Ye Jin Bang, Andrea Madotto, and Pascale Fung. Survey of hallucination in natural language generation. ACM Comput. Surv., 55(12), 2023. 2", + "[32] Krish Kabra, Kathleen M. Lewis, and Guha Balakrishnan. Gelda: A generative language annotation framework to reveal visual biases in image generators. In CVPRW, 2024. 2", + "[33] Younghyun Kim, Sangwoo Mo, Minkyu Kim, Kyungmin Lee, Jaeho Lee, and Jinwoo Shin. Discovering and mitigating visual biases through keyword explanation. In CVPR, 2024. 1, 2, 5, 6, 7, 4", + "[34] Arvindkumar Krishnakumar, Viraj Prabhu, Sruthi Sudhakar, and Judy Hoffman. Udis: Unsupervised discovery of bias in deep visual recognition models. In BMVC, 2021. 1, 2", + "[35] Zhiheng Li and Chenliang Xu. Discover the unknown biased attribute of an image classifier. In ICCV, 2021. 2", + "[36] Zhiheng Li, Anthony Hoogs, and Chenliang Xu. Discover and mitigate unknown biases with debiasing alternate networks. In ECCV, 2022. 2", + "[37] Evan Z Liu, Behzad Haghgoo, Annie S Chen, Aditi Raghunathan, Pang Wei Koh, Shiori Sagawa, Percy Liang, and Chelsea Finn. Just train twice: Improving group robustness without training group information. In ICML, 2021. 2", + "[38] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, 2023. 7, 11", + "[39] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In CVPR, 2024. 7, 11", + "[40] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava next: Improved reasoning,OCR, and world knowledge, 2024. 11", + "[41] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaou Tang. Deep learning face attributes in the wild. In ICCV, 2015. 2, 5" + ], + "bbox": [ + 91, + 92, + 480, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[42] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, 2021. 5", + "[43] TorchVision maintainers and contributors. Torchvision: Pytorch's computer vision library. https://github.com/pytorch/vision, 2016.5", + "[44] Ninareh Mehrabi, Fred Morstatter, Nripsuta Saxena, Kristina Lerman, and Aram Galstyan. A survey on bias and fairness in machine learning. ACM computing surveys (CSUR), 54 (6):1-35, 2021. 3", + "[45] Thomas Mesnard, Cassidy Hardin, Robert Dadashi, Surya Bhupatiraju, Shreya Pathak, Laurent Sifre, Morgane Rivière, Mihir Sanjay Kale, Juliette Love, et al. Gemma: Open models based on gemini research and technology. arXiv preprint arXiv:2403.08295, 2024. 15", + "[46] Microsoft. Chat markup language chatml (preview), 2024. https://learn.microsoft.com/en-us/azure/ai-services/openai/how-to/chat-markup-language, Last accessed on 2024-11-21. 13", + "[47] Margaret Mitchell, Simone Wu, Andrew Zaldivar, Parker Barnes, Lucy Vasserman, Ben Hutchinson, Elena Spitzer, Inioluwa Deborah Raji, and Timnit Gebru. Model cards for model reporting. In Proceedings of the conference on fairness, accountability, and transparency, pages 220-229, 2019. 1", + "[48] Kartik Narayan, Vibashan VS, Rama Chellappa, and Vishal M Patel. Facexformer: A unified transformer for facial analysis. arXiv preprint arXiv:2403.12960, 2024. 5", + "[49] Ziad Obermeyer, Brian Powers, Christine Vogeli, and Sendhil Mullainathan. Dissecting racial bias in an algorithm used to manage the health of populations. Science, 366(6464): 447-453, 2019. 1", + "[50] Jesutofunmi A. Omiye, Jenna C. Lester, Simon Spichak, Veronica Rotemberg, and Roxana Daneshjou. Large language models propagate race-based medicine. npj Digital Medicine, 6(1):195, 2023. 2", + "[51] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 4", + "[52] Nils Reimers and Iryna Gurevych. Sentence-bert: Sentence embeddings using siamese bert-networks. In ENMLP, 2019. 6", + "[53] Keivan Rezaei, Mehrdad Saberi, Mazda Moayeri, and Soheil Feizi. PRIME: Prioritizing interpretability in failure mode extraction. In ICLR, 2024. 1, 3", + "[54] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, Alexander C. Berg, and Li Fei-Fei. ImageNet Large Scale Visual Recognition Challenge. IJCV, 115(3):211-252, 2015. 5", + "[55] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortman, Patrick Schramowski, Srivatsa Kundurthy, Katherine" + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "15160", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Crowson, Ludwig Schmidt, Robert Kaczmarczyk, and Jenia Jitsev. Laion-5b: An open large-scale dataset for training next generation image-text models. In Advances in Neural Information Processing Systems, pages 25278-25294, 2022. 2", + "[56] Tamar Rott Shaham, Sarah Schwettmann, Franklin Wang, Achyuta Rajaram, Evan Hernandez, Jacob Andreas, and Antonio Torralba. A multimodal automated interpretability agent. In ICML, 2024. 3", + "[57] Robik Shrestha, Yang Zou, Qiuyu Chen, Zhiheng Li, Yusheng Xie, and Siqi Deng. Fairrag: Fair human generation via fair retrieval augmentation. In CVPR, 2024. 2", + "[58] Mannat Singh, Laura Gustafson, Aaron Adcock, Vinicius de Freitas Reis, Bugra Gedik, Raj Prateek Kosaraju, Dhruv Mahajan, Ross Girshick, Piotr Dólár, and Laurens van der Maaten. Revisiting weakly supervised pre-training of visual perception models. In CVPR, 2022. 5", + "[59] Sahil Singla and Soheil Feizi. Salient imagenet: How to discover spurious features in deep learning? In ICLR, 2022. 2", + "[60] Sahil Singla, Besmira Nushi, Shital Shah, Ece Kamar, and Eric Horvitz. Understanding failures of deep networks via robust feature extraction. In CVPR, 2021. 1", + "[61] Nimit Sohoni, Jared Dunnmon, Geoffrey Angus, Albert Gu, and Christopher Ré. No subclass left behind: Fine-grained robustness in coarse-grained classification problems. NeurIPS, 2020. 2", + "[62] Tristan Thrush, Ryan Jiang, Max Bartolo, Amanpreet Singh, Adina Williams, Douwe Kiela, and Candace Ross. Winoground: Probing vision and language models for visio-linguistic compositionality. In CVPR, 2022. 2", + "[63] Vijay Vasudevan, Benjamin Caine, Raphael Gontijo Lopes, Sara Fridovich-Keil, and Rebecca Roelofs. When does dough become a bagel? analyzing the remaining mistakes on imagenet. In NeurIPS, 2022. 1", + "[64] Angelina Wang, Alexander Liu, Ryan Zhang, Anat Kleiman, Leslie Kim, Dora Zhao, Iroha Shirai, Arvind Narayanan, and Olga Russakovsky. Revise: A tool for measuring and mitigating bias in visual datasets. *IJCV*, 130(7):1790–1810, 2022. 1", + "[65] Olivia Wiles, Isabela Albuquerque, and Sven Gowal. Discovering bugs in vision models using off-the-shelf image generation and captioning. In NeurIPS-WS, 2022. 2", + "[66] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 11975-11986, 2023. 2", + "[67] Yue Zhang, Yafu Li, Leyang Cui, Deng Cai, Lemao Liu, Tingchen Fu, Xinting Huang, Enbo Zhao, Yu Zhang, Yu-long Chen, et al. Siren's song in the ai ocean: a survey on hallucination in large language models. arXiv preprint arXiv:2309.01219, 2023. 2", + "[68] Zeliang Zhang, Mingqian Feng, Zhiheng Li, and Chenliang Xu. Discover and mitigate multiple biased subgroups in image classifiers. In CVPR, 2024. 1, 2", + "[69] Dora Zhao, Angelina Wang, and Olga Russakovsky. Understanding and evaluating racial biases in image captioning. In ICCV, 2021. 1" + ], + "bbox": [ + 91, + 90, + 483, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[70] Jieyu Zhao, Tianlu Wang, Mark Yatskar, Vicente Ordonez, and Kai-Wei Chang. Men also like shopping: Reducing gender bias amplification using corpus-level constraints. In EMNLP, 2017. 1", + "[71] Zaiying Zhao, Soichiro Kumano, and Toshihiko Yamasaki. Language-guided detection and mitigation of unknown dataset bias. arXiv preprint arXiv:2406.02889, 2024. 2" + ], + "bbox": [ + 514, + 90, + 903, + 190 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "15161", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Classifier-to-Bias_ Toward Unsupervised Automatic Bias Detection for Visual Classifiers/303d2f58-e46e-4b7a-806e-0d4e64d53df2_model.json b/2025/Classifier-to-Bias_ Toward Unsupervised Automatic Bias Detection for Visual Classifiers/303d2f58-e46e-4b7a-806e-0d4e64d53df2_model.json new file mode 100644 index 0000000000000000000000000000000000000000..141852da6d21a6026370bc1153b0f60c57583e45 --- /dev/null +++ b/2025/Classifier-to-Bias_ Toward Unsupervised Automatic Bias Detection for Visual Classifiers/303d2f58-e46e-4b7a-806e-0d4e64d53df2_model.json @@ -0,0 +1,2367 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.212, + 0.131, + 0.787, + 0.175 + ], + "angle": 0, + "content": "CLASSIFIER-TO-BIAS: Toward Unsupervised Automatic Bias Detection for Visual Classifiers" + }, + { + "type": "text", + "bbox": [ + 0.179, + 0.203, + 0.818, + 0.24 + ], + "angle": 0, + "content": "Quentin Guimard\\(^{1}\\) Moreno D'Incà\\(^{1}\\) Massimiliano Mancini\\(^{1}\\) Elisa Ricci\\(^{1,2}\\) \n\\(^{1}\\)University of Trento\\(^{2}\\) Fondazione Bruno Kessler" + }, + { + "type": "text", + "bbox": [ + 0.362, + 0.242, + 0.631, + 0.256 + ], + "angle": 0, + "content": "https://github.com/mardgui/C2B" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.292, + 0.327, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.326, + 0.485, + 0.689 + ], + "angle": 0, + "content": "A person downloading a pre-trained model from the web should be aware of its biases. Existing approaches for bias identification rely on datasets containing labels for the task of interest, something that a non-expert may not have access to, or may not have the necessary resources to collect: this greatly limits the number of tasks where model biases can be identified. In this work, we present CLASSIFIER-TO-BIAS (C2B), the first bias discovery framework that works without access to any labeled data: it only relies on a textual description of the classification task to identify biases in the target classification model. This description is fed to a large language model to generate bias proposals and corresponding captions depicting biases together with task-specific target labels. A retrieval model collects images for those captions, which are then used to assess the accuracy of the model w.r.t. the given biases. C2B is training-free, does not require any annotations, has no constraints on the list of biases, and can be applied to any pre-trained model on any classification task. Experiments on two publicly available datasets show that C2B discovers biases beyond those of the original datasets and outperforms a recent state-of-the-art bias detection baseline that relies on task-specific annotations, being a promising first step toward addressing task-agnostic unsupervised bias detection." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.723, + 0.222, + 0.739 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.484, + 0.901 + ], + "angle": 0, + "content": "The wide availability of pre-trained machine learning models allows anyone to access them. A user may search in online libraries (e.g., HuggingFace [27]) for models addressing their task of interest, download the most suitable one, and use it for their specific use case. As the user did not develop the model, they may not be aware of potential failure modes [28, 60, 63] or harmful biases [9, 10, 24, 49, 69, 70]. These are usually reported on model cards [47], describing how the model has been developed, in which cases it has been tested, its recommended use, and its potential biases." + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.289, + 0.905, + 0.501 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.51, + 0.907, + 0.581 + ], + "angle": 0, + "content": "Figure 1. We explore the novel task of bias discovery when we are only given a specific classification task and a pre-trained model. We propose CLASSIFIER-TO-BIAS (C2B), which automatically detects potential biases in the model, categorizes these biases, and assigns scores to each category." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.61, + 0.907, + 0.701 + ], + "angle": 0, + "content": "While model cards are a useful tool to describe a model and its shortcomings, (i) a lot of models still come without them; (ii) even if they are present, they may not describe specific cases or biases of interest to the users. This means that a user would need to assess the weaknesses of the downloaded model independently." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.703, + 0.909, + 0.883 + ], + "angle": 0, + "content": "Various tools have already been proposed to measure and mitigate biases in datasets [8, 64] and models [2, 8, 11]. While these tools rely on bias labels, recent work has made progress toward open-set failure mode extraction [16, 19, 22, 30, 53] and bias detection [6, 17, 33, 34, 68] without bias labels, even removing the need for a predefined list of biases. However, in the context of discriminative models such as visual classifiers, all existing methods rely on task-specific labeled data to be able to identify failure modes and measure the biases, limiting the number of tasks where model biases can be identified, and preventing the automation of bias detection." + }, + { + "type": "text", + "bbox": [ + 0.534, + 0.886, + 0.905, + 0.901 + ], + "angle": 0, + "content": "To address this problem, we present CLASSIFIER-TO" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "15151" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.363 + ], + "angle": 0, + "content": "BIAS (C2B), the first framework that, given a pre-trained visual classifier and its task specification, automatically identifies its biases (Fig. 1) in an unsupervised manner, removing the need for task-specific labeled data. We build C2B on two components: a large language model (LLM) and a text-to-image retrieval engine. First, the task specifications are fed to the LLM to produce a series of candidate bias attributes for the task (e.g., gender, ethnicity), and their corresponding bias classes (e.g., male, female). The list of candidate bias classes is fed again to the LLM to generate a set of captions describing both the biases and the task-specific target classes they are associated to. These captions are used by the retrieval module to get images from a large database: the images will constitute our dataset, where, for each image, both the target and bias labels are given by the constructed caption itself. By testing the pre-trained model on this dataset, we can compute the standard bias metrics, verifying which bias class the model is most biased against." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.364, + 0.482, + 0.56 + ], + "angle": 0, + "content": "Importantly, C2B is fully unsupervised and does not suffer from the need of costly annotations, thus it can be applied to automatically detect biases for virtually any classification task. Moreover, the list of biases is class-specific, as different classification tasks and categories (e.g., birds, deers) may exhibit different biases (e.g., tail color, horns size). Finally, the model is modular and can readily incorporate more powerful LLMs, retrieval approaches, or databases as they become available. We test our framework on two datasets, i.e., CelebA [41], and ImageNet-X [28]. Results indicate that C2B can uncover biases beyond those annotated in the original datasets, outperforming a recent supervised open-set bias detector [33]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.561, + 0.483, + 0.756 + ], + "angle": 0, + "content": "Contributions. To summarize, our contribution is threefold: (i) We propose the new task of unsupervised bias detection in a pre-trained model given only a textual description of the classification task without requiring any annotated dataset or manual intervention, and define a protocol to evaluate methods on this setting; (ii) We present C2B, the first framework for this task that uses an LLM to propose a list of biases and generate captions, and multimodal retrieval to produce a dataset from which biases can be estimated; (iii) We show that C2B can discover biases beyond those annotated in the original datasets, being the first step toward addressing this challenging problem and democratizing bias discovery." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.77, + 0.233, + 0.785 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.482, + 0.901 + ], + "angle": 0, + "content": "Bias Detection and Mitigation. Bias detection and mitigation are essential to ensure fairness and robustness in machine learning models. Several approaches tackle hidden subgroup imbalances without relying on explicit group annotations [35, 37, 59, 61], e.g., by analyzing the mistakes on annotated task data and re-weighting misclassified samples [37]. Other methods focus on open-set or annotation-" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.363 + ], + "angle": 0, + "content": "free bias discovery [34, 36, 68]. UDIS [34] uses hierarchical clustering to surface systematic biases without predefined subgroup definitions. DebiAN [36] alternates training between a bias discoverer and a classifier to identify and mitigate multiple biases. DIM [68] decomposes feature spaces using Partial Least Squares to uncover multiple unknown biased subgroups and describes them using vision-language models. More recently, language-driven methods have also gained prominence, using captions and keyword extraction to detect and name potential biases learned by models [14, 33, 71]. B2T [33] and Zhao et al. [71] relies on captions derived from annotated datasets or misclassified samples with known labels, validating these biases using, e.g. vision-language similarity. SaMyNa [14] emphasizes interpretability by extracting semantic bias descriptors from large-scale captioning, supporting bias analysis and debiasing. However, all these methods require access to annotated, task-specific datasets." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.364, + 0.905, + 0.454 + ], + "angle": 0, + "content": "In contrast, our approach, C2B, operates without any kind of annotated data. It relies solely on a natural language description of the classification task and a pre-trained classifier. Using LLM-generated, class-specific bias proposals combined with retrieval-based evaluation, C2B enables fully unsupervised, task-agnostic bias discovery." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.455, + 0.905, + 0.56 + ], + "angle": 0, + "content": "Other related methods focus on bias detection or fairness in generative models [17, 32, 57]. For instance, Open-Bias [17] uses LLMs and VQA models to detect and quantify open-set biases in text-to-image generative models, while GELDA [32] reveals visual biases in image generators by generating domain-specific attributes with LLMs and annotating them using selected VLMs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.562, + 0.905, + 0.651 + ], + "angle": 0, + "content": "Inspired by these works, we choose to leverage an LLM as well as a text-to-image retrieval engine to tackle the new task of unsupervised bias detection, where only a pretrained model and a textual description of the classification task are available, thus eliminating the need for annotated datasets or manual intervention." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.659, + 0.905, + 0.794 + ], + "angle": 0, + "content": "Explainability and Failure Mode Detection. A parallel research direction focuses on systematic failure mode discovery and interpretability [3, 4, 16, 25, 29]. VoG [3] highlights difficult examples by analyzing gradient variance, and Spotlight [16] identifies contiguous underperforming regions in representation space. Saliency-based methods [29] provide local visual explanations, while counterfactual explanations are generated in either natural language [25] or visual form using diffusion models [4]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Recent work has scaled these efforts using embeddings and generative models. Wiles et al. [65] leverages large-scale generative models to automatically find and describe failure clusters. Domino [19] uses cross-modal embeddings and mixture models to identify and describe coherent data slices where models underperform. Jain et al. [30] extracts latent directions corresponding to consistent model errors" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "15152" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.257 + ], + "angle": 0, + "content": "and generates challenging counterfactual examples for data augmentation. AdaVision [22] enables interactive discovery of failure modes by combining CLIP-based retrieval with GPT-generated prompts. PRIME [53] prioritizes interpretability by starting from human-defined tags and identifying minimal descriptive combinations that explain failures. MAIA [56] automates interpretability workflows using vision-language models for feature analysis and failure mode discovery. Csurka et al. [15] formalizes language-based error explainability and proposes evaluation metrics to assess the quality of generated explanations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.258, + 0.482, + 0.379 + ], + "angle": 0, + "content": "While these methods analyze model behavior and failure patterns through model outputs or datasets, our work introduces a complementary approach: starting from a high-level task description, C2B uses LLM-based bias proposals and retrieval-based validation to automatically identify, validate, and quantify biases in pre-trained classifiers, enabling large-scale, unsupervised bias discovery across arbitrary classification tasks." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.394, + 0.182, + 0.409 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.419, + 0.484, + 0.525 + ], + "angle": 0, + "content": "In this section, we first define bias attributes and bias classes (Sec. 3.1), and formulate the problem of identifying biases of pre-trained models from their task specifications (Sec. 3.2). We then describe C2B, that proposes a list of biases (Sec. 3.3) and collects data for testing them (Sec. 3.4). Finally, we describe how we measure model's biases from the collected data (Sec. 3.5)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.535, + 0.238, + 0.55 + ], + "angle": 0, + "content": "3.1. Bias definition" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.559, + 0.483, + 0.756 + ], + "angle": 0, + "content": "Formally, we denote the target classification model as \\( f: \\mathcal{X} \\to \\mathcal{Y} \\), mapping an image in the space \\( \\mathcal{X} \\) to labels in the space \\( \\mathcal{Y} \\). We define as bias attributes \\( \\mathcal{B} = \\{\\mathbf{B}_1, \\dots, \\mathbf{B}_N\\} \\) the different categories of bias, i.e., factors that can impact its performance [44]. For each bias attribute, we define a list of bias classes \\( \\mathbf{B}_i = \\{b_{i,1}, \\dots, b_{i,n_i}\\} \\), the different values that each bias attribute can take. For a given target class, we define the model to be biased or unfair w.r.t specific bias classes depending on the per-class accuracy difference of the model. Let us denote the per-class accuracy of a model \\( f \\) on a target class \\( y \\) on data that matches a bias class \\( b_{i,j} \\) as \\( A_y(f, b_{i,j}) \\). We define the bias score \\( \\phi_{y,i,j} \\) for each bias class \\( b_{i,j} \\) as" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.767, + 0.483, + 0.805 + ], + "angle": 0, + "content": "\\[\n\\phi_ {y, i, j} = A _ {y} \\left(f, b _ {i, j}\\right) - \\frac {1}{n _ {i} - 1} \\sum_ {b _ {i, k} \\in \\mathbf {B} _ {i}, k \\neq j} A _ {y} \\left(f, b _ {i, k}\\right). \\tag {1}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.817, + 0.289, + 0.832 + ], + "angle": 0, + "content": "3.2. Problem formulation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Given a pre-trained model \\( f \\), our goal is to identify its biases. We assume \\( f \\) to be a generic classification model that we did not develop and that we aim to use (e.g., downloaded from the web). Thus, we also assume to have no access to" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.152 + ], + "angle": 0, + "content": "additional data \\( f \\) had access to (e.g., the dataset set used for training). As we want to avoid costly annotation procedures, we also assume to have no access to any dataset containing labeled images of the classes in \\( \\mathcal{V} \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.154, + 0.906, + 0.259 + ], + "angle": 0, + "content": "To address this problem, we can rely only on what we know about the model: i.e., the task it has been designed for, the input domain it expects, and the output it predicts. This information can all be expressed using simple, natural language. Specifically, we assume to have a general description of the model's task \\( T \\), the input the model expects to process, and its output (i.e., classes in \\( \\mathcal{V} \\))." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.262, + 0.907, + 0.413 + ], + "angle": 0, + "content": "This task is challenging as (i) detecting biases requires measuring the performance of the model on a labeled dataset, potentially containing annotations for the factors of interest and (ii) we have no pre-defined list of biases to start with. This implies that we have to both generate a list of potential biases and to collect a dataset to test them. In the following, we will describe our framework, C2B, that exploits LLMs for proposing biases, and a combination of LLMs and multimodal retrieval for scoring them. An overview of C2B is illustrated in Fig. 2." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.429, + 0.75, + 0.446 + ], + "angle": 0, + "content": "3.3. Generating bias proposals" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.454, + 0.906, + 0.53 + ], + "angle": 0, + "content": "The first step of C2B is to generate a list of bias attributes and classes that we can explicitly test for. To generate the list, we cannot rely on attribute annotations, as we have no access to labeled data. Moreover, we cannot use lists available online, as biases depend on the task of interest." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.532, + 0.906, + 0.683 + ], + "angle": 0, + "content": "To address this problem, we exploit the reasoning capabilities of LLMs. Specifically, let us denote as \\(\\Phi_{\\mathrm{LLM}}\\) a generic LLM, mapping textual inputs to textual outputs, i.e., \\(\\Phi_{\\mathrm{LLM}}: \\mathcal{T} \\to \\mathcal{T}\\), with \\(\\mathcal{T}\\) being the space of natural language. To generate the candidate set of biases, we prompt the LLM to produce a list of bias attributes and classes for the given model and task description. Because different classes may have different biases, we query the LLM to give us a list of potential biases \\(\\mathcal{B}^y\\) for each class \\(y \\in \\mathcal{Y}\\). Formally, we have:" + }, + { + "type": "equation", + "bbox": [ + 0.621, + 0.699, + 0.905, + 0.716 + ], + "angle": 0, + "content": "\\[\n\\mathcal {B} ^ {y} = \\Phi_ {\\mathrm {L L M}} \\left(\\pi_ {\\text {b i a s}} \\circ T \\circ y\\right) \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.732, + 0.906, + 0.793 + ], + "angle": 0, + "content": "where \\( \\circ \\) is the string concatenation operation, \\( \\pi_{\\mathrm{bias}} \\) is the system prompt that instructs the LLM about the task it should perform (Supp. Mat., Appx. I), and \\( T \\) is the model's task description, including input/output specification." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.794, + 0.905, + 0.858 + ], + "angle": 0, + "content": "The output of Eq. (2), \\(\\mathcal{B}^y\\), is a list1 of bias attributes \\(\\mathcal{B}^y = \\{\\mathbf{B}_1^y, \\dots, \\mathbf{B}_{N_y}^y\\}\\), of size \\(N_y\\). We prompt the LLM so that each bias attribute is associated to a list of \\(n_{y,i}\\) bias classes \\(\\mathbf{B}_i^y = \\{b_{i,1}^y, \\dots, b_{i,n_{y,i}}^y\\}\\)." + }, + { + "type": "page_footnote", + "bbox": [ + 0.513, + 0.875, + 0.905, + 0.901 + ], + "angle": 0, + "content": "1For simplicity, we assume that the natural language output of the LLM is already formatted as a list." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15153" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.091, + 0.086, + 0.912, + 0.346 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.354, + 0.908, + 0.412 + ], + "angle": 0, + "content": "Figure 2. Overview of C2B. Given a specific task with an associated description and a pre-trained classification model, our approach leverages an LLM to identify potential candidate bias attributes and corresponding classes. These candidate biases are then used to prompt the LLM to generate captions, which are subsequently utilized by a retrieval module to collect a dataset of images for bias testing. This dataset enables evaluation of the target classification model, producing scores for each identified bias category." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.436, + 0.245, + 0.452 + ], + "angle": 0, + "content": "3.4. Collecting data" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.463, + 0.484, + 0.614 + ], + "angle": 0, + "content": "Given the list of candidate bias attributes and classes, we still need a dataset to score them. Ideally, for each target class \\( y \\in \\mathcal{V} \\) and bias class \\( b^{y} \\in \\mathbf{B}^{y} \\), with \\( \\mathbf{B}^{y} \\in \\mathcal{B}^{y} \\), we would like to have a dataset \\( D(b^{y}, y) \\) depicting images of class \\( y \\) exhibiting the bias class \\( b^{y} \\), i.e., \\( D(b^{y}, y) = \\{(x_{i}, y_{i}, b_{i}^{y})\\}_{i=1}^{M} \\) of size \\( M \\). Unfortunately, we cannot rely on labeled data for this task as (i) we do not have access to it; (ii) biases change w.r.t. the particular application, thus it is very hard to collect beforehand a dataset containing all possible target and bias classes of interest." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.619, + 0.484, + 0.741 + ], + "angle": 0, + "content": "To address this problem, we resort to external sources. In particular, we assume to have access to a large-scale database \\(\\mathcal{D}\\) containing images. Note that \\(\\mathcal{D}\\) has no annotations: we just assume it to be large-scale (i.e., potentially web-scale). From this database, we can collect images for each target-bias combination of interest using a text-to-image retrieval model. This is achieved in two steps: captions generation and cross-modal retrieval." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Caption generation. A straightforward solution to create captions is to use a fixed template (e.g., an image of [target class] with [bias class]). However, this would not consider the specificity of the considered task and thus may lead to retrieving erroneous images. For instance, the biases and captions associated to a facial attribute classification task should be different from those in a natural scene classification task. To avoid this issue, we resort to using the LLM, querying it in two steps: first to create a template for the specific task, and then to gener" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.437, + 0.907, + 0.467 + ], + "angle": 0, + "content": "ate captions specific for target and bias class. Formally, we obtain the initial template as:" + }, + { + "type": "equation", + "bbox": [ + 0.614, + 0.479, + 0.907, + 0.496 + ], + "angle": 0, + "content": "\\[\n\\operatorname {t e m p} = \\Phi_ {\\mathrm {L L M}} \\left(\\pi_ {\\text {t e m p l a t e}} \\circ T\\right) \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.507, + 0.906, + 0.568 + ], + "angle": 0, + "content": "where \\(\\pi_{\\mathrm{template}}\\) is the system prompt that instructs the LLM about the task it should perform (Supp. Mat., Appx. 1). Then, we ask the LLM to refine this template for each specific target class \\(y\\) and bias class \\(b\\), as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.588, + 0.578, + 0.905, + 0.596 + ], + "angle": 0, + "content": "\\[\nc _ {b} ^ {y} = \\Phi_ {\\mathrm {L L M}} \\left(\\pi_ {\\text {c a p t i o n}} \\circ \\operatorname {t e m p} \\circ b \\circ y\\right). \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.606, + 0.906, + 0.667 + ], + "angle": 0, + "content": "where \\( c_b^y \\) is the resulting caption. Note that, while with this strategy we obtain one caption per target-bias class pair, in principle we can also generate multiple captions per pair to have more variability, beneficial for more complex tasks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.673, + 0.907, + 0.824 + ], + "angle": 0, + "content": "Image retrieval. The obtained captions can be used to retrieve images from a large scale image database, and we implement the retrieval by using a contrastive vision-language model (VLM) that can score the similarity between a text and image. Formally, we can decompose the VLM in three elements: a visual encoder \\(\\varphi_{\\mathrm{vis}}:\\mathcal{X}\\to \\mathbb{R}^d\\) a text encoder \\(\\varphi_{\\mathrm{txt}}:\\mathcal{T}\\rightarrow \\mathbb{R}^d\\) , and a similarity function \\(\\sin :\\mathbb{R}^d\\times \\mathbb{R}^d\\to \\mathbb{R}\\) (e.g., cosine similarity for CLIP [51]). Given a caption, we can retrieve a set of images \\(\\mathbf{X}_b^y\\) for each class label \\(y\\) and bias \\(b\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.581, + 0.835, + 0.905, + 0.861 + ], + "angle": 0, + "content": "\\[\n\\mathbf {X} _ {b} ^ {y} = \\underset {x \\in \\mathcal {D}} {\\operatorname {t o p - k}} \\operatorname {s i m} \\left(\\varphi_ {\\text {v i s}} (x), \\varphi_ {\\text {t x t}} \\left(c _ {b} ^ {y}\\right)\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.906, + 0.902 + ], + "angle": 0, + "content": "where we retrieve the \\(k\\) images with the highest matching score. While these images do not have a ground-truth," + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15154" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.152 + ], + "angle": 0, + "content": "they can still help to estimate the biases of the model, with the fidelity of the latter depending on the retrieval accuracy (Sec. 4.4). We also experiment with a different type of retrieval system, a Web search engine (details in Sec. 4.1)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.163, + 0.262, + 0.178 + ], + "angle": 0, + "content": "3.5. Measuring biases" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.186, + 0.483, + 0.381 + ], + "angle": 0, + "content": "In the previous section, we have collected a dataset with pseudo-annotations for all bias values and classes of interest. We can now test the target model on this dataset to compute its performance, using the variation across bias classes to score each bias. We use the formula defined in Eq. (1) to score each target-bias pair relative to other target-bias pairs. For each bias class, we obtain a bias score between -1 and 1. If the per-class accuracy is higher for a given bias class, the classifier is biased toward the bias class (positive bias score). Reciprocally, if the per-class accuracy is lower for a given bias class, the classifier is biased against the bias class (negative bias score). Intuitively, larger absolute bias score values also correspond to stronger biases." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.383, + 0.483, + 0.534 + ], + "angle": 0, + "content": "To quantitatively evaluate the quality of the biases detected by C2B, we propose two novel evaluation settings that complement each other. First, in Sec. 4.2.1, we propose to match the detected biases with ground-truth annotations using SBERT embeddings for tasks where datasets with additional bias annotations exist. Second, in Sec. 4.2.2, we propose to assign bias labels (based on the initial proposed biases) to ground-truth task-specific data with a VQA model, to check the agreement between the bias detection method and an expert VQA model." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.535, + 0.483, + 0.64 + ], + "angle": 0, + "content": "Collecting scores for all bias attributes and classes enables various types of analyses to explore the biases. Biases can be ranked by score, to assess which attributes are more critical than others. Bias class scores can also be aggregated across bias attributes, to measure which bias attributes exhibit a larger bias. We provide examples of qualitative analyses in Sec. 4.3, and more in the Supp. Mat., Appx. D." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.655, + 0.222, + 0.672 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.68, + 0.483, + 0.801 + ], + "angle": 0, + "content": "In this section, we present the results of our experiments, where we apply C2B to detect biases for different models on two different image classification tasks. We first describe our experimental protocol (Sec. 4.1), we then quantitatively compare the biases detected by C2B with a state-of-the-art competitor (Sec. 4.2), and we provide qualitative examples of detected biases (Sec. 4.3). Finally, we analyze the core component of our approach (Sec. 4.4)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.81, + 0.3, + 0.827 + ], + "angle": 0, + "content": "4.1. Experimental protocol" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.483, + 0.871 + ], + "angle": 0, + "content": "Tasks and models. We test C2B on two different tasks: facial attribute classification and image classification." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Facial attribute classification. We consider the CelebA dataset [41], which is composed of 200k images of celebrity" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.198 + ], + "angle": 0, + "content": "faces annotated with 40 facial attributes, such as male, blond hair, young, or high cheekbones. To evaluate C2B, we choose to detect the biases of a recent state-of-the-art model, FaceXFormer [48]. This model relies on a pretrained Swin/B [42] backbone, coupled with an MLP-based fusion module and an attention-based decoder, which are co-trained on several face datasets." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.199, + 0.905, + 0.32 + ], + "angle": 0, + "content": "Image classification. We focus on the widely used ImageNet-1K [54] dataset, where each image is associated to one of the 1000 ImageNet classes. We choose to detect the biases of four widely used models, easily accessible through TorchVision [43], comprising both convolutional networks and transformers: ResNet50_V2 [23], ResNet101_V2 [23], ResNet152_V2 [23], ViT_B_16_SWAG_E2E_V1 [58]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.327, + 0.907, + 0.584 + ], + "angle": 0, + "content": "Competitors. As we introduce the new problem of unsupervised bias detection relying solely on a textual description of the classification task, there are no existing baseline methods designed for this specific setting. Nevertheless, we still compare C2B to a recent supervised open-set bias detection method, Bias-to-Text (B2T) [33]. However, unlike C2B which does not require task-specific labeled data to identify biases in an image classifier, B2T depends on ground-truth task labels to identify model failures. It then employs a pre-trained captioning model to generate descriptions of misclassified images, extracts common keywords, and computes a bias score based on the difference between (i) the similarity (CLIP score) between the discovered bias keyword and misclassified images, and (ii) the similarity between the discovered bias keyword and well-classified images. Thus, we test B2T using the labeled validation sets of the two datasets (i.e., CelebA and ImageNet)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.592, + 0.906, + 0.818 + ], + "angle": 0, + "content": "Implementation details. C2B comprises two components: the LLM that proposes biases and captions, and the retrieval module. For the former, we use Llama 3.1 8B [18] (see Supp. Mat., Appx. J for qualitative LLM comparisons). For the retrieval, we explore two strategies: (i) CC12M: we consider the large Conceptual Captions 12M (CC12M) dataset [12] and retrieve images based on CLIP embeddings [7], and (ii) BING: instead of using a VLM to retrieve images from an existing database where some image domains may not be well-represented, we propose to retrieve images from the Web using a publicly available search engine, Bing. For both datasets, we retrieve 20 images per caption, and we measure biases as in Eq. (1), comparing the difference in the model's per-class accuracy between images that match different candidate bias classes." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.831, + 0.731, + 0.848 + ], + "angle": 0, + "content": "4.2. Quantitative evaluation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.905, + 0.902 + ], + "angle": 0, + "content": "In this section, we provide a quantitative evaluation to assess the quality of biases identified by C2B. First, we present results based on ground-truth bias annotations, i.e.," + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "15155" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.152 + ], + "angle": 0, + "content": "exploiting attributes annotated in the validation sets of the datasets. For biases not in the ground-truth annotations, we introduce an evaluation based on Visual Question Answering (VQA) to measure the relevance of the proposed biases." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.175, + 0.439, + 0.19 + ], + "angle": 0, + "content": "4.2.1 Evaluation with ground-truth annotations" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.206, + 0.483, + 0.341 + ], + "angle": 0, + "content": "Setting. In this setting, we exploit the available attribute annotations, checking if models can detect and quantify those biases correctly. In the case of CelebA, when we focus on a specific attribute classification task (e.g., blonde hair) we consider all other 39 attributes (e.g., young) as biases measurable with the ground-truth. For ImageNet-1K, we use ImageNet-X [28], i.e., the validation set of ImageNet-1K with 16 labeled \"factors\" distinguishing images from typical prototypes (e.g., orientation), as ground-truth biases." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.343, + 0.484, + 0.479 + ], + "angle": 0, + "content": "To obtain ground-truth biases, we test the model on each validation set and check the difference in per-class accuracy among images where the bias attribute is present and images where it is absent. We consider a ground-truth bias to be present if this difference is higher than a threshold (i.e., \\(\\tau = 0.05\\)) using the same threshold to assess bias identification on both C2B and B2T scores. We refer to the sign of the bias score as the direction of the bias. We provide more results with a varying threshold in the Supp. Mat., Appx. C." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.48, + 0.484, + 0.601 + ], + "angle": 0, + "content": "Due to the open-set nature of the detected biases for both C2B and B2T, there are almost no exact matches between them and ground-truth biases. To compare the two sets, we compute the cosine similarity between SBERT embeddings [52] of the ground-truth and detected biases. We empirically set the similarity threshold to 0.9 to avoid having too many false positives. We provide more details and qualitative examples of matches in the Supp. Mat., Appx. K." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.607, + 0.484, + 0.789 + ], + "angle": 0, + "content": "Metrics. We report results from two complementary perspectives: GT \\(\\rightarrow\\) Detected and Detected \\(\\rightarrow\\) GT. In GT \\(\\rightarrow\\) Detected, we start from each ground-truth bias and check whether it was correctly detected with the same bias direction (HIT), detected with the opposite direction (false hit, FH), or not detected (M1ss). This perspective focuses on recall. However, methods that propose many biases may artificially reduce misses. Therefore, we also report Detected \\(\\rightarrow\\) GT, where we start from each detected bias and verify whether it matches a ground-truth bias with the correct direction. This second view measures precision and helps assess the quality of the proposed biases." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Results. In Tabs. 1 and 2, we report the results of our analysis. On CelebA (left side of Tab. 1, GT \\(\\rightarrow\\) Detected), C2B outperforms B2T in detecting ground-truth biases, regardless of the retrieval source. We obtain the best results with Bing, detecting between \\(10.8\\%\\) and \\(12.3\\%\\) of the ground-truth biases, compared to only \\(4.5\\%\\) for B2T. However, B2T exhibits a lower proportion of false hits \\((3.1\\%)\\), whereas" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.09, + 0.907, + 0.132 + ], + "angle": 0, + "content": "Table 1. Proportion \\((\\%)\\) of ground-truth biases detected on CelebA \\(\\mathrm{(GT\\to Detected)}\\) and of detected biases corresponding to groundtruth ones (Detected \\(\\rightarrow\\) GT). FH=False Hit." + }, + { + "type": "table", + "bbox": [ + 0.52, + 0.143, + 0.903, + 0.22 + ], + "angle": 0, + "content": "
METHODGT → DETECTEDDETECTED → GT
HIT(↑)FH(↓)MISS (↓)HIT(↑)FH(↓)MISS(↓)
FaceXFormer
B2T [33]4.533.1492.326.594.1089.29
C2B (BING)12.296.8880.8314.187.3478.48
C2B (CC12M)10.767.7581.4912.758.1579.09
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.233, + 0.907, + 0.275 + ], + "angle": 0, + "content": "Table 2. Proportion (\\%) of ground-truth biases detected on ImageNet-X (GT \\(\\rightarrow\\) Detected) and of detected biases corresponding to ground-truth ones (Detected \\(\\rightarrow\\) GT). FH=False Hit." + }, + { + "type": "table", + "bbox": [ + 0.52, + 0.287, + 0.905, + 0.491 + ], + "angle": 0, + "content": "
METHODGT → DETECTEDDETECTED → GT
HIT(↑)FH(↓)Miss (↓)HIT(↑)FH (↓)Miss(↓)
ResNet50_V2
B2T [33]2.402.1995.410.800.7198.49
C2B (BING)7.808.5383.662.602.9894.42
C2B (CC12M)11.1811.3077.522.992.8794.14
ResNet101_V2
B2T [33]2.552.1195.330.840.6998.47
C2B (BING)7.918.4183.682.562.7794.67
C2B (CC12M)11.2111.6877.112.762.9694.28
ResNet152_V2
B2T [33]2.531.7295.750.830.5898.59
C2B (BING)7.537.9384.542.752.7994.46
C2B (CC12M)11.1012.0176.902.812.9594.24
ViT_B_16_SWAG
B2T [33]2.111.9895.900.720.6798.61
C2B (BING)7.727.1885.102.482.4895.05
C2B (CC12M)10.8511.0778.082.642.7094.66
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.516, + 0.907, + 0.652 + ], + "angle": 0, + "content": "C2B shows slightly higher false hit rates (6.9–7.8%). This is largely due to retrieval noise and the broader search space introduced by LLM-generated proposals, which can sometimes lead to less precise bias candidates. Nevertheless, C2B misses fewer ground-truth biases (80.8–81.5% vs. B2T's 92.3%). These trends remain consistent when considering the proportion of detected biases that correspond to ground-truth biases (Detected → GT), where C2B achieves 12.8–14.2%, compared to 6.6% for B2T." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.653, + 0.909, + 0.834 + ], + "angle": 0, + "content": "On ImageNet-X (Tab. 1), across four classifiers, C2B again detects more ground-truth biases than B2T (7.5–11.2% vs. 2.1–2.6%), at the cost of higher false hit rates (7.2–12.0% vs. 1.7–2.2%). As before, these false hits are attributable to noisy retrieval and the open-ended nature of bias proposals. Still, C2B misses significantly fewer ground-truth biases (76.9–85.1% vs. 95.3–95.9%). Interestingly, on ImageNet-X, retrieving from CC12M provides better results than Bing. The Detected → GT evaluation confirms these findings, with a larger fraction of C2B-detected biases aligning with ground-truth ones (2.5–3.0% vs. 0.6–0.7% for B2T)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Discussion. Due to the limited human annotations available in CelebA and ImageNet-X, the proportion of matching biases is low across all methods, particularly on ImageNet-X. Furthermore, since we are using text embeddings to" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "15156" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.116, + 0.09, + 0.46, + 0.104 + ], + "angle": 0, + "content": "Table 3. Detected biases and VQA agreement on CelebA." + }, + { + "type": "table", + "bbox": [ + 0.13, + 0.115, + 0.445, + 0.149 + ], + "angle": 0, + "content": "
MODELB2T [33]C2B (BING)C2B (CC12M)
FaceXFormer0.100.220.25
" + }, + { + "type": "table_caption", + "bbox": [ + 0.101, + 0.164, + 0.472, + 0.178 + ], + "angle": 0, + "content": "Table 4. Detected biases and VQA agreement on ImageNet-X." + }, + { + "type": "table", + "bbox": [ + 0.128, + 0.189, + 0.447, + 0.255 + ], + "angle": 0, + "content": "
MODELB2T [33]C2B (BING)C2B (CC12M)
ResNet50_V20.200.260.20
ResNet101_V20.190.280.20
ResNet152_V20.200.280.20
ViT_B_16_SWAG0.200.320.21
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.282, + 0.484, + 0.403 + ], + "angle": 0, + "content": "match the biases, the numbers are also influenced by the semantic proximity between LLM-proposed biases, caption-extracted keywords, and ground-truth annotations. Moreover, a high number of Miss in the Detected \\(\\rightarrow\\) GT case is not necessarily a bad thing, as C2B and B2T may be detecting accurate biases that are not present in the annotation. For these reasons, in the next section we introduce a different evaluation based on visual question answering." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.424, + 0.303, + 0.438 + ], + "angle": 0, + "content": "4.2.2 VQA-based evaluation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.449, + 0.484, + 0.51 + ], + "angle": 0, + "content": "As ground-truth bias annotations may not always be available, we also propose an alternative evaluation scheme based on VQA, measuring the quality of the proposed biases that are not present in ground-truth annotations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.516, + 0.484, + 0.713 + ], + "angle": 0, + "content": "Metric. We use a VQA model (LLaVA-1.5-13B [38, 39]) to pseudo-label each image of the validation sets with each of the proposed bias by C2B or B2T. Additional details can be found in the Supp. Mat., Appx. G. From these pseudolabels, we can use the ground-truth task labels (i.e., class annotations) to check how the per-class accuracy varies over different bias classes, using the same threshold \\(\\tau = 0.05\\) of Sec. 4.2.1. For each bias class \\(b\\) proposed by the model, there are three cases: a positive bias is detected \\((\\phi_b \\geq \\tau)\\), a negative bias is detected \\((\\phi_b \\leq -\\tau)\\), or the bias is not detected \\((|\\phi_b| < \\tau)\\). For each bias class, we define the agreement score to be 1 if both scores fall in the same case, -1 if they detect opposite biases, and 0 otherwise." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Results. Tabs. 3 and 4 present the results of our VQA-based evaluation on CelebA and ImageNet-X. On CelebA, C2B achieves higher agreement with VQA pseudo-labels (0.22 for Bing and 0.25 for CC12M) than B2T (0.10), indicating that the biases discovered by C2B are more aligned with visually detectable and semantically coherent bias attributes. On ImageNet-X, while C2B often achieves higher agreement scores than B2T with Bing (i.e., +0.09 on average), results with CC12M are either the same or comparable to those of B2T (i.e., +0.005 on average). We attribute this to the challenge of relying on retrieval-based data collection, without access to annotated failure cases or" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.198 + ], + "angle": 0, + "content": "task-specific captions. In contrast, B2T benefits from captions derived from known misclassifications. Despite this more challenging and fully unsupervised setting, C2B still achieves agreements that are competitive with or superior to B2T. This confirms that C2B's proposals are not only meaningful and visually grounded but also relevant for bias detection in settings where no annotations are available." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.205, + 0.906, + 0.356 + ], + "angle": 0, + "content": "Discussion. The results presented in this section are influenced by both the types of questions posed to the VQA model and the model's own inherent biases. Given these limitations, evaluations based on ground-truth annotations and those derived from VQA outputs serve as complementary approaches. Overall, biases proposed by the LLM tend to offer more relevant bias indicators than keywords extracted from captions as in B2T [33], allowing C2B to detect a greater proportion of ground-truth biases and achieve higher agreement with the VQA model than B2T." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.369, + 0.72, + 0.385 + ], + "angle": 0, + "content": "4.3. Qualitative evaluation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.393, + 0.905, + 0.422 + ], + "angle": 0, + "content": "In this section, we show examples of biases discovered by C2B for different models on different tasks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.424, + 0.906, + 0.53 + ], + "angle": 0, + "content": "In Fig. 3, we show biases that C2B detects for the high cheekbones attribute on face attribute classification, with bias scores from different retrieval methods. We can see that a strong bias was detected on the expression attribute: the classifier appears to perform better when people are smiling and worse when they appear angry. Other smaller biases were detected, such as young, and male." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.532, + 0.906, + 0.728 + ], + "angle": 0, + "content": "Fig. 4 shows a similar analysis for biases detected by C2B for the minibus class on image classification for the ResNet50_V2 model. Strong biases were detected on the lighting, orientation, and presence of reflection bias attribute: the classifier appears stronger on daytime images, portrait-oriented images, and images without reflections, while performing worse on nighttime images, landscape-oriented images, and images with reflections. Finally, Fig. 5 depicts biases for the birdhouse class again for image classification, but for the ViT_B_16_SWAG model. A strong bias is found on the shape bias attribute: the classifier seems to be performing better on images of rectangular birdhouses, while it underperforms on images of spherical birdhouses." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.741, + 0.819, + 0.755 + ], + "angle": 0, + "content": "4.4. Evaluation of VLM-based retrieval" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Retrieving images matching the target classes and biases present in the captions is essential for the detected biases to be trustworthy. Therefore, in this section, we analyze the CLIP-based retrieval approach used by C2B, by trying to retrieve images for which we have annotations. For a VQA-based analysis of both the Bing-based retrieval and CLIP-based retrieval on CC12M, please refer to Supp. Mat., Appx. F. To perform our analysis, we use CLIP to retrieve images from the validation sets of CelebA and ImageNet-" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "15157" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.092, + 0.905, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.213, + 0.265, + 0.784, + 0.28 + ], + "angle": 0, + "content": "Figure 3. Examples of biases found by C2B for the high cheekbones target attribute on CelebA." + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.307, + 0.482, + 0.452 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.465, + 0.483, + 0.494 + ], + "angle": 0, + "content": "Figure 4. Examples of biases found by C2B for the minibus class on ImageNet-X (ResNet50_V2)." + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.514, + 0.482, + 0.662 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.675, + 0.483, + 0.703 + ], + "angle": 0, + "content": "Figure 5. Examples of biases found by C2B for the birdhouse class on ImageNet-X (ViT_B_16_SWAG)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.734, + 0.483, + 0.765 + ], + "angle": 0, + "content": "X, using handcrafted captions as input: combinations of attributes on CelebA, and classes and factors on ImageNet-X." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.772, + 0.483, + 0.832 + ], + "angle": 0, + "content": "Metric. We use recall@ \\(K\\), where \\(K\\) varies for each caption to match the number of corresponding ground-truth samples in the dataset. We show results using recall @ \\(A \\cdot K\\), where \\(A\\) is a coefficient that varies between 0.01 and 1.0." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Results. We present the results in Fig. 6. For CelebA (left), the recall is the same for both target and bias attributes, as expected, since both are drawn from the same set of attributes. On the same dataset, CLIP struggles to" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.305, + 0.716, + 0.445 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.718, + 0.305, + 0.903, + 0.444 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.577, + 0.456, + 0.842, + 0.47 + ], + "angle": 0, + "content": "Figure 6. Accuracy of VLM-based retrieval." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.503, + 0.907, + 0.64 + ], + "angle": 0, + "content": "retrieve images with combinations of attributes, with a recall @ \\(0.01 \\cdot K\\) just above \\(40\\%\\). For ImageNet-X (right), CLIP is fairly accurate, especially for low values of \\(A\\), with more than \\(80\\%\\) and \\(70\\%\\) recall at \\(0.05 \\cdot K\\) to retrieve images with specific biases and classes, respectively. Generally, the recall is higher when the number of retrieved images decreases relative to the number of potential matching images present in the dataset, which suggests that, the larger the dataset, the more accurate the retrieval is." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.661, + 0.634, + 0.677 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.907, + 0.901 + ], + "angle": 0, + "content": "We introduced CLASSIFIER-TO-BIAS (C2B), a novel method for automatically detecting biases in visual classifiers entirely in an unsupervised manner, without the need for task-specific labeled data. Starting from a simple textual description of the classification task and a pre-trained model, C2B generates bias hypotheses, retrieves relevant images to test them, and produces a list of biases with their scores. Although C2B is the first approach to operate in such challenging setting, our results show that it surpasses a recent supervised open-set bias detector across two evaluation settings: one based on ground-truth annotations and another on VQA-based pseudo-labels. We believe our work marks an important step toward automated, open-set, and task-agnostic bias detection." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15158" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.251, + 0.108 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.484, + 0.207 + ], + "angle": 0, + "content": "This work was sponsored by the MUR PNRR project FAIR - Future AI Research (PE00000013), funded by NextGeneration, and the Italian Ministerial grants PRIN 2022: \"B-FAIR: Bias-Free Artificial Intelligence methods for automated visual Recognition\" (CUP E53D23008010006), and the EU project ELIAS (No.01120237)." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.218, + 0.188, + 0.234 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.244, + 0.483, + 0.313 + ], + "angle": 0, + "content": "[1] Marah Abdin, Sam Ade Jacobs, Ammar Ahmad Awan, Jyoti Aneja, Ahmed Awadallah, Hany Awadalla, Nguyen Bach, Amit Bahree, Arash Bakhtiari, Harkirat Behl, et al. Phi-3 technical report: A highly capable language model locally on your phone. arXiv preprint arXiv:2404.14219, 2024. 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.314, + 0.483, + 0.355 + ], + "angle": 0, + "content": "[2] Julius A Adebayo et al. *FairML: ToolBox for diagnosing bias in predictive modeling*. PhD thesis, Massachusetts Institute of Technology, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.357, + 0.483, + 0.396 + ], + "angle": 0, + "content": "[3] Chirag Agarwal, Daniel D'souza, and Sara Hooker. Estimating example difficulty using variance of gradients. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.399, + 0.482, + 0.438 + ], + "angle": 0, + "content": "[4] Maximilian Augustin, Valentyn Boreiko, Francesco Croce, and Matthias Hein. Diffusion visual counterfactual explanations. NeurIPS, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.441, + 0.482, + 0.481 + ], + "angle": 0, + "content": "[5] Razvan Azamfirei, Sapna R. Kudchadkar, and James Fackler. Large language models and the perils of their hallucinations. Critical Care, 27(1):120, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.483, + 0.482, + 0.522 + ], + "angle": 0, + "content": "[6] Yujia Bao and Regina Barzilay. Learning to split for automatic bias detection. arXiv preprint arXiv:2204.13749, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.525, + 0.482, + 0.579 + ], + "angle": 0, + "content": "[7] Romain Beaumont. Clip retrieval: Easily compute clip embeddings and build a clip retrieval system with them. https://github.com/rom1504/clip-retrieval, 2022.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.581, + 0.482, + 0.676 + ], + "angle": 0, + "content": "[8] R. K. E. Bellamy, K. Dey, M. Hind, S. C. Hoffman, S. Houde, K. Kannan, P. Lohia, J. Martino, S. Mehta, A. Mojsilovic, S. Nagar, K. Natesan Ramamurthy, J. Richards, D. Saha, P. Sattigeri, M. Singh, K. R. Varshney, and Y. Zhang. Ai fairness 360: An extensible toolkit for detecting and mitigating algorithmic bias. IBM Journal of Research and Development, 63(4/5):4:1-4:15, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.678, + 0.482, + 0.734 + ], + "angle": 0, + "content": "[9] Tolga Bolukbasi, Kai-Wei Chang, James Zou, Venkatesh Saligrama, and Adam Kalai. Man is to computer programmer as woman is to homemaker? debiasing word embeddings. In NeurIPS, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.735, + 0.482, + 0.789 + ], + "angle": 0, + "content": "[10] Joy Buolamwini and Timnit Gebru. Gender shades: Intersectional accuracy disparities in commercial gender classification. In Proceedings of the 1st Conference on Fairness, Accountability and Transparency, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.79, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[11] Angel Alexander Cabrera, Will Epperson, Fred Hohman, Minsuk Kahng, Jamie Morgenstern, and Duen Horng Chau. Fairvis: Visual analytics for discovering intersectional bias in machine learning. In 2019 IEEE Conference on Visual Analytics Science and Technology (VAST), pages 46-56, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.874, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[12] Soravit Changpinyo, Piyush Sharma, Nan Ding, and Radu Soricut. Conceptual 12M: Pushing web-scale image-text" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.244, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.094, + 0.905, + 0.119 + ], + "angle": 0, + "content": "pre-training to recognize long-tail visual concepts. In CVPR, 2021.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.122, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[13] Davide Chicco, Niklas Totsch, and Giuseppe Jurman. The matthews correlation coefficient (mcc) is more reliable than balanced accuracy, bookmaker informedness, and markedness in two-class confusion matrix evaluation. BioData Mining, 14(1):13, 2021. 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.192, + 0.905, + 0.258 + ], + "angle": 0, + "content": "[14] Massimiliano Ciranni, Luca Molinaro, Carlo Alberto Barbano, Attilio Fiandrotti, Vittorio Murino, Vito Paolo Pastore, and Enzo Tartaglione. Say my name: a model's bias discovery framework. arXiv preprint arXiv:2408.09570, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.261, + 0.905, + 0.302 + ], + "angle": 0, + "content": "[15] Gabriela Csurka, Tyler L Hayes, Diane Larlus, and Riccardo Volpi. What could go wrong? discovering and describing failure modes in computer vision. In ECCV-WS, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.304, + 0.905, + 0.372 + ], + "angle": 0, + "content": "[16] Greg d'Eon, Jason d'Eon, James R Wright, and Kevin Leyton-Brown. The spotlight: A general method for discovering systematic errors in deep learning models. In Proceedings of the 2022 ACM Conference on Fairness, Accountability, and Transparency, pages 1962-1981, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.373, + 0.905, + 0.428 + ], + "angle": 0, + "content": "[17] Moreno D'Inca, Elia Peruzzo, Massimiliano Mancini, Dejia Xu, Vedit Goel, Xingqian Xu, Zhangyang Wang, Humphrey Shi, and Nicu Sebe. Openbias: Open-set bias detection in text-to-image generative models. In CVPR, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.429, + 0.905, + 0.496 + ], + "angle": 0, + "content": "[18] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 5, 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.498, + 0.905, + 0.566 + ], + "angle": 0, + "content": "[19] Sabri Eyuboglu, Maya Varma, Khaled Kamal Saab, Jean-Benoit Delbrouck, Christopher Lee-Messer, Jared Dunnmon, James Zou, and Christopher Re. Domino: Discovering systematic errors with cross-modal embeddings. In ICLR, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.568, + 0.905, + 0.746 + ], + "angle": 0, + "content": "[20] Samir Yitzhak Gadre, Gabriel Ilharco, Alex Fang, Jonathan Hayase, Georgios Smyrnis, Thao Nguyen, Ryan Marten, Mitchell Wortsman, Dhruba Ghosh, Jieyu Zhang, Eyal Orgad, Rahim Entezari, Giannis Daras, Sarah Pratt, Vivek Ramanujan, Yonatan Bitton, Kalyani Marathe, Stephen Mussmann, Richard Vencu, Mehdi Cherti, Ranjay Krishna, Pang Wei W Koh, Olga Saukh, Alexander J Ratner, Shuran Song, Hannaneh Hajishirzi, Ali Farhadi, Romain Beaumont, Sewoong Oh, Alex Dimakis, Jenia Jitsev, Yair Carmon, Vaishaal Shankar, and Ludwig Schmidt. Datacomp: In search of the next generation of multimodal datasets. In Advances in Neural Information Processing Systems, pages 27092-27112, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.748, + 0.905, + 0.816 + ], + "angle": 0, + "content": "[21] Isabel O. Gallegos, Ryan A. Rossi, Joe Barrow, Md Mehrab Tanjim, Sungchul Kim, Franck Dernoncourt, Tong Yu, Ruiyi Zhang, and Nesreen K. Ahmed. Bias and fairness in large language models: A survey. Computational Linguistics, 50 (3):1097-1179, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.818, + 0.905, + 0.859 + ], + "angle": 0, + "content": "[22] Irena Gao, Gabriel Ilharco, Scott Lundberg, and Marco Tulio Ribeiro. Adaptive testing of computer vision models. In ICCV, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[23] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 5" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.094, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15159" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.134 + ], + "angle": 0, + "content": "[24] Lisa Anne Hendricks, Kaylee Burns, Kate Saenko, Trevor Darrell, and Anna Rohrbach. Women also snowboard: Overcoming bias in captioning models. In ECCV, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.482, + 0.177 + ], + "angle": 0, + "content": "[25] Lisa Anne Hendricks, Ronghang Hu, Trevor Darrell, and Zeynep Akata. Generating counterfactual explanations with natural language. In ICML-WS, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.179, + 0.482, + 0.232 + ], + "angle": 0, + "content": "[26] Cheng-Yu Hsieh, Jieyu Zhang, Zixian Ma, Aniruddha Kembhavi, and Ranjay Krishna. Sugarcrepe: Fixing hackable benchmarks for vision-language compositionality. NeurIPS, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.235, + 0.279, + 0.248 + ], + "angle": 0, + "content": "[27] https://huggingface.co/. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.25, + 0.482, + 0.318 + ], + "angle": 0, + "content": "[28] Badr Youbi Idrissi, Diane Bouchacourt, Randall Balestriero, Ivan Evtimov, Caner Hazirbas, Nicolas Ballas, Pascal Vincent, Michal Drozdjal, David Lopez-Paz, and Mark Ibrahim. Imagenet-x: Understanding model mistakes with factor of variation annotations. In ICLR, 2023. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.321, + 0.482, + 0.361 + ], + "angle": 0, + "content": "[29] Saachi Jain, Hadi Salman, Eric Wong, Pengchuan Zhang, Vibhav Vineet, Sai Vermprala, and Aleksander Madry. Missingness bias in model debugging. In ICLR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.363, + 0.482, + 0.403 + ], + "angle": 0, + "content": "[30] Saachi Jain, Hannah Lawrence, Ankur Moitra, and Aleksander Madry. Distilling model failures as directions in latent space. In ICLR, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.406, + 0.482, + 0.46 + ], + "angle": 0, + "content": "[31] Ziwei Ji, Nayeon Lee, Rita Frieske, Tiezheng Yu, Dan Su, Yan Xu, Etsuko Ishii, Ye Jin Bang, Andrea Madotto, and Pascale Fung. Survey of hallucination in natural language generation. ACM Comput. Surv., 55(12), 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.462, + 0.482, + 0.516 + ], + "angle": 0, + "content": "[32] Krish Kabra, Kathleen M. Lewis, and Guha Balakrishnan. Gelda: A generative language annotation framework to reveal visual biases in image generators. In CVPRW, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.519, + 0.482, + 0.572 + ], + "angle": 0, + "content": "[33] Younghyun Kim, Sangwoo Mo, Minkyu Kim, Kyungmin Lee, Jaeho Lee, and Jinwoo Shin. Discovering and mitigating visual biases through keyword explanation. In CVPR, 2024. 1, 2, 5, 6, 7, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.575, + 0.482, + 0.616 + ], + "angle": 0, + "content": "[34] Arvindkumar Krishnakumar, Viraj Prabhu, Sruthi Sudhakar, and Judy Hoffman. Udis: Unsupervised discovery of bias in deep visual recognition models. In BMVC, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.618, + 0.482, + 0.644 + ], + "angle": 0, + "content": "[35] Zhiheng Li and Chenliang Xu. Discover the unknown biased attribute of an image classifier. In ICCV, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.647, + 0.482, + 0.686 + ], + "angle": 0, + "content": "[36] Zhiheng Li, Anthony Hoogs, and Chenliang Xu. Discover and mitigate unknown biases with debiasing alternate networks. In ECCV, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.689, + 0.482, + 0.743 + ], + "angle": 0, + "content": "[37] Evan Z Liu, Behzad Haghgoo, Annie S Chen, Aditi Raghunathan, Pang Wei Koh, Shiori Sagawa, Percy Liang, and Chelsea Finn. Just train twice: Improving group robustness without training group information. In ICML, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.746, + 0.482, + 0.772 + ], + "angle": 0, + "content": "[38] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, 2023. 7, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.775, + 0.482, + 0.814 + ], + "angle": 0, + "content": "[39] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In CVPR, 2024. 7, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.817, + 0.482, + 0.858 + ], + "angle": 0, + "content": "[40] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava next: Improved reasoning,OCR, and world knowledge, 2024. 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.86, + 0.482, + 0.899 + ], + "angle": 0, + "content": "[41] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaou Tang. Deep learning face attributes in the wild. In ICCV, 2015. 2, 5" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[42] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.15, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[43] TorchVision maintainers and contributors. Torchvision: Pytorch's computer vision library. https://github.com/pytorch/vision, 2016.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.193, + 0.905, + 0.247 + ], + "angle": 0, + "content": "[44] Ninareh Mehrabi, Fred Morstatter, Nripsuta Saxena, Kristina Lerman, and Aram Galstyan. A survey on bias and fairness in machine learning. ACM computing surveys (CSUR), 54 (6):1-35, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.25, + 0.905, + 0.317 + ], + "angle": 0, + "content": "[45] Thomas Mesnard, Cassidy Hardin, Robert Dadashi, Surya Bhupatiraju, Shreya Pathak, Laurent Sifre, Morgane Rivière, Mihir Sanjay Kale, Juliette Love, et al. Gemma: Open models based on gemini research and technology. arXiv preprint arXiv:2403.08295, 2024. 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.321, + 0.905, + 0.375 + ], + "angle": 0, + "content": "[46] Microsoft. Chat markup language chatml (preview), 2024. https://learn.microsoft.com/en-us/azure/ai-services/openai/how-to/chat-markup-language, Last accessed on 2024-11-21. 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.378, + 0.905, + 0.459 + ], + "angle": 0, + "content": "[47] Margaret Mitchell, Simone Wu, Andrew Zaldivar, Parker Barnes, Lucy Vasserman, Ben Hutchinson, Elena Spitzer, Inioluwa Deborah Raji, and Timnit Gebru. Model cards for model reporting. In Proceedings of the conference on fairness, accountability, and transparency, pages 220-229, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.462, + 0.905, + 0.503 + ], + "angle": 0, + "content": "[48] Kartik Narayan, Vibashan VS, Rama Chellappa, and Vishal M Patel. Facexformer: A unified transformer for facial analysis. arXiv preprint arXiv:2403.12960, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.505, + 0.905, + 0.559 + ], + "angle": 0, + "content": "[49] Ziad Obermeyer, Brian Powers, Christine Vogeli, and Sendhil Mullainathan. Dissecting racial bias in an algorithm used to manage the health of populations. Science, 366(6464): 447-453, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.562, + 0.905, + 0.616 + ], + "angle": 0, + "content": "[50] Jesutofunmi A. Omiye, Jenna C. Lester, Simon Spichak, Veronica Rotemberg, and Roxana Daneshjou. Large language models propagate race-based medicine. npj Digital Medicine, 6(1):195, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.619, + 0.905, + 0.686 + ], + "angle": 0, + "content": "[51] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.689, + 0.905, + 0.728 + ], + "angle": 0, + "content": "[52] Nils Reimers and Iryna Gurevych. Sentence-bert: Sentence embeddings using siamese bert-networks. In ENMLP, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.733, + 0.905, + 0.772 + ], + "angle": 0, + "content": "[53] Keivan Rezaei, Mehrdad Saberi, Mazda Moayeri, and Soheil Feizi. PRIME: Prioritizing interpretability in failure mode extraction. In ICLR, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.775, + 0.905, + 0.843 + ], + "angle": 0, + "content": "[54] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, Alexander C. Berg, and Li Fei-Fei. ImageNet Large Scale Visual Recognition Challenge. IJCV, 115(3):211-252, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[55] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortman, Patrick Schramowski, Srivatsa Kundurthy, Katherine" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15160" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.484, + 0.161 + ], + "angle": 0, + "content": "Crowson, Ludwig Schmidt, Robert Kaczmarczyk, and Jenia Jitsev. Laion-5b: An open large-scale dataset for training next generation image-text models. In Advances in Neural Information Processing Systems, pages 25278-25294, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.163, + 0.482, + 0.218 + ], + "angle": 0, + "content": "[56] Tamar Rott Shaham, Sarah Schwettmann, Franklin Wang, Achyuta Rajaram, Evan Hernandez, Jacob Andreas, and Antonio Torralba. A multimodal automated interpretability agent. In ICML, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.219, + 0.482, + 0.26 + ], + "angle": 0, + "content": "[57] Robik Shrestha, Yang Zou, Qiuyu Chen, Zhiheng Li, Yusheng Xie, and Siqi Deng. Fairrag: Fair human generation via fair retrieval augmentation. In CVPR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.261, + 0.482, + 0.328 + ], + "angle": 0, + "content": "[58] Mannat Singh, Laura Gustafson, Aaron Adcock, Vinicius de Freitas Reis, Bugra Gedik, Raj Prateek Kosaraju, Dhruv Mahajan, Ross Girshick, Piotr Dólár, and Laurens van der Maaten. Revisiting weakly supervised pre-training of visual perception models. In CVPR, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.33, + 0.482, + 0.37 + ], + "angle": 0, + "content": "[59] Sahil Singla and Soheil Feizi. Salient imagenet: How to discover spurious features in deep learning? In ICLR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.372, + 0.482, + 0.412 + ], + "angle": 0, + "content": "[60] Sahil Singla, Besmira Nushi, Shital Shah, Ece Kamar, and Eric Horvitz. Understanding failures of deep networks via robust feature extraction. In CVPR, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.414, + 0.482, + 0.467 + ], + "angle": 0, + "content": "[61] Nimit Sohoni, Jared Dunnmon, Geoffrey Angus, Albert Gu, and Christopher Ré. No subclass left behind: Fine-grained robustness in coarse-grained classification problems. NeurIPS, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.47, + 0.482, + 0.524 + ], + "angle": 0, + "content": "[62] Tristan Thrush, Ryan Jiang, Max Bartolo, Amanpreet Singh, Adina Williams, Douwe Kiela, and Candace Ross. Winoground: Probing vision and language models for visio-linguistic compositionality. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.526, + 0.482, + 0.58 + ], + "angle": 0, + "content": "[63] Vijay Vasudevan, Benjamin Caine, Raphael Gontijo Lopes, Sara Fridovich-Keil, and Rebecca Roelofs. When does dough become a bagel? analyzing the remaining mistakes on imagenet. In NeurIPS, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.582, + 0.482, + 0.648 + ], + "angle": 0, + "content": "[64] Angelina Wang, Alexander Liu, Ryan Zhang, Anat Kleiman, Leslie Kim, Dora Zhao, Iroha Shirai, Arvind Narayanan, and Olga Russakovsky. Revise: A tool for measuring and mitigating bias in visual datasets. *IJCV*, 130(7):1790–1810, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.651, + 0.482, + 0.692 + ], + "angle": 0, + "content": "[65] Olivia Wiles, Isabela Albuquerque, and Sven Gowal. Discovering bugs in vision models using off-the-shelf image generation and captioning. In NeurIPS-WS, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.693, + 0.482, + 0.748 + ], + "angle": 0, + "content": "[66] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 11975-11986, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.749, + 0.482, + 0.816 + ], + "angle": 0, + "content": "[67] Yue Zhang, Yafu Li, Leyang Cui, Deng Cai, Lemao Liu, Tingchen Fu, Xinting Huang, Enbo Zhao, Yu Zhang, Yu-long Chen, et al. Siren's song in the ai ocean: a survey on hallucination in large language models. arXiv preprint arXiv:2309.01219, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.818, + 0.482, + 0.859 + ], + "angle": 0, + "content": "[68] Zeliang Zhang, Mingqian Feng, Zhiheng Li, and Chenliang Xu. Discover and mitigate multiple biased subgroups in image classifiers. In CVPR, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.86, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[69] Dora Zhao, Angelina Wang, and Olga Russakovsky. Understanding and evaluating racial biases in image captioning. In ICCV, 2021. 1" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.092, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[70] Jieyu Zhao, Tianlu Wang, Mark Yatskar, Vicente Ordonez, and Kai-Wei Chang. Men also like shopping: Reducing gender bias amplification using corpus-level constraints. In EMNLP, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.149, + 0.905, + 0.191 + ], + "angle": 0, + "content": "[71] Zaiying Zhao, Soichiro Kumano, and Toshihiko Yamasaki. Language-guided detection and mitigation of unknown dataset bias. arXiv preprint arXiv:2406.02889, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.092, + 0.905, + 0.191 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "15161" + } + ] +] \ No newline at end of file diff --git a/2025/Classifier-to-Bias_ Toward Unsupervised Automatic Bias Detection for Visual Classifiers/303d2f58-e46e-4b7a-806e-0d4e64d53df2_origin.pdf b/2025/Classifier-to-Bias_ Toward Unsupervised Automatic Bias Detection for Visual Classifiers/303d2f58-e46e-4b7a-806e-0d4e64d53df2_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..f7ae2e74e3e65e42ec0d32ed739d8a41ef05bf3b --- /dev/null +++ b/2025/Classifier-to-Bias_ Toward Unsupervised Automatic Bias Detection for Visual Classifiers/303d2f58-e46e-4b7a-806e-0d4e64d53df2_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6b16a79825f3fd7fe37e62f3e1f1404de31ab8a84923530741766c3a625a0a8b +size 2542508 diff --git a/2025/Classifier-to-Bias_ Toward Unsupervised Automatic Bias Detection for Visual Classifiers/full.md b/2025/Classifier-to-Bias_ Toward Unsupervised Automatic Bias Detection for Visual Classifiers/full.md new file mode 100644 index 0000000000000000000000000000000000000000..52156bfd73a9360659517e69328c0024cffd143e --- /dev/null +++ b/2025/Classifier-to-Bias_ Toward Unsupervised Automatic Bias Detection for Visual Classifiers/full.md @@ -0,0 +1,322 @@ +# CLASSIFIER-TO-BIAS: Toward Unsupervised Automatic Bias Detection for Visual Classifiers + +Quentin Guimard $^{1}$ Moreno D'Incà $^{1}$ Massimiliano Mancini $^{1}$ Elisa Ricci $^{1,2}$ $^{1}$ University of Trento $^{2}$ Fondazione Bruno Kessler + +https://github.com/mardgui/C2B + +# Abstract + +A person downloading a pre-trained model from the web should be aware of its biases. Existing approaches for bias identification rely on datasets containing labels for the task of interest, something that a non-expert may not have access to, or may not have the necessary resources to collect: this greatly limits the number of tasks where model biases can be identified. In this work, we present CLASSIFIER-TO-BIAS (C2B), the first bias discovery framework that works without access to any labeled data: it only relies on a textual description of the classification task to identify biases in the target classification model. This description is fed to a large language model to generate bias proposals and corresponding captions depicting biases together with task-specific target labels. A retrieval model collects images for those captions, which are then used to assess the accuracy of the model w.r.t. the given biases. C2B is training-free, does not require any annotations, has no constraints on the list of biases, and can be applied to any pre-trained model on any classification task. Experiments on two publicly available datasets show that C2B discovers biases beyond those of the original datasets and outperforms a recent state-of-the-art bias detection baseline that relies on task-specific annotations, being a promising first step toward addressing task-agnostic unsupervised bias detection. + +# 1. Introduction + +The wide availability of pre-trained machine learning models allows anyone to access them. A user may search in online libraries (e.g., HuggingFace [27]) for models addressing their task of interest, download the most suitable one, and use it for their specific use case. As the user did not develop the model, they may not be aware of potential failure modes [28, 60, 63] or harmful biases [9, 10, 24, 49, 69, 70]. These are usually reported on model cards [47], describing how the model has been developed, in which cases it has been tested, its recommended use, and its potential biases. + +![](images/1ad9d05a0618ab645e66313f37c59c01ce03526b801871aae1a393c66b8ba488.jpg) +Figure 1. We explore the novel task of bias discovery when we are only given a specific classification task and a pre-trained model. We propose CLASSIFIER-TO-BIAS (C2B), which automatically detects potential biases in the model, categorizes these biases, and assigns scores to each category. + +While model cards are a useful tool to describe a model and its shortcomings, (i) a lot of models still come without them; (ii) even if they are present, they may not describe specific cases or biases of interest to the users. This means that a user would need to assess the weaknesses of the downloaded model independently. + +Various tools have already been proposed to measure and mitigate biases in datasets [8, 64] and models [2, 8, 11]. While these tools rely on bias labels, recent work has made progress toward open-set failure mode extraction [16, 19, 22, 30, 53] and bias detection [6, 17, 33, 34, 68] without bias labels, even removing the need for a predefined list of biases. However, in the context of discriminative models such as visual classifiers, all existing methods rely on task-specific labeled data to be able to identify failure modes and measure the biases, limiting the number of tasks where model biases can be identified, and preventing the automation of bias detection. + +To address this problem, we present CLASSIFIER-TO + +BIAS (C2B), the first framework that, given a pre-trained visual classifier and its task specification, automatically identifies its biases (Fig. 1) in an unsupervised manner, removing the need for task-specific labeled data. We build C2B on two components: a large language model (LLM) and a text-to-image retrieval engine. First, the task specifications are fed to the LLM to produce a series of candidate bias attributes for the task (e.g., gender, ethnicity), and their corresponding bias classes (e.g., male, female). The list of candidate bias classes is fed again to the LLM to generate a set of captions describing both the biases and the task-specific target classes they are associated to. These captions are used by the retrieval module to get images from a large database: the images will constitute our dataset, where, for each image, both the target and bias labels are given by the constructed caption itself. By testing the pre-trained model on this dataset, we can compute the standard bias metrics, verifying which bias class the model is most biased against. + +Importantly, C2B is fully unsupervised and does not suffer from the need of costly annotations, thus it can be applied to automatically detect biases for virtually any classification task. Moreover, the list of biases is class-specific, as different classification tasks and categories (e.g., birds, deers) may exhibit different biases (e.g., tail color, horns size). Finally, the model is modular and can readily incorporate more powerful LLMs, retrieval approaches, or databases as they become available. We test our framework on two datasets, i.e., CelebA [41], and ImageNet-X [28]. Results indicate that C2B can uncover biases beyond those annotated in the original datasets, outperforming a recent supervised open-set bias detector [33]. + +Contributions. To summarize, our contribution is threefold: (i) We propose the new task of unsupervised bias detection in a pre-trained model given only a textual description of the classification task without requiring any annotated dataset or manual intervention, and define a protocol to evaluate methods on this setting; (ii) We present C2B, the first framework for this task that uses an LLM to propose a list of biases and generate captions, and multimodal retrieval to produce a dataset from which biases can be estimated; (iii) We show that C2B can discover biases beyond those annotated in the original datasets, being the first step toward addressing this challenging problem and democratizing bias discovery. + +# 2. Related Work + +Bias Detection and Mitigation. Bias detection and mitigation are essential to ensure fairness and robustness in machine learning models. Several approaches tackle hidden subgroup imbalances without relying on explicit group annotations [35, 37, 59, 61], e.g., by analyzing the mistakes on annotated task data and re-weighting misclassified samples [37]. Other methods focus on open-set or annotation- + +free bias discovery [34, 36, 68]. UDIS [34] uses hierarchical clustering to surface systematic biases without predefined subgroup definitions. DebiAN [36] alternates training between a bias discoverer and a classifier to identify and mitigate multiple biases. DIM [68] decomposes feature spaces using Partial Least Squares to uncover multiple unknown biased subgroups and describes them using vision-language models. More recently, language-driven methods have also gained prominence, using captions and keyword extraction to detect and name potential biases learned by models [14, 33, 71]. B2T [33] and Zhao et al. [71] relies on captions derived from annotated datasets or misclassified samples with known labels, validating these biases using, e.g. vision-language similarity. SaMyNa [14] emphasizes interpretability by extracting semantic bias descriptors from large-scale captioning, supporting bias analysis and debiasing. However, all these methods require access to annotated, task-specific datasets. + +In contrast, our approach, C2B, operates without any kind of annotated data. It relies solely on a natural language description of the classification task and a pre-trained classifier. Using LLM-generated, class-specific bias proposals combined with retrieval-based evaluation, C2B enables fully unsupervised, task-agnostic bias discovery. + +Other related methods focus on bias detection or fairness in generative models [17, 32, 57]. For instance, Open-Bias [17] uses LLMs and VQA models to detect and quantify open-set biases in text-to-image generative models, while GELDA [32] reveals visual biases in image generators by generating domain-specific attributes with LLMs and annotating them using selected VLMs. + +Inspired by these works, we choose to leverage an LLM as well as a text-to-image retrieval engine to tackle the new task of unsupervised bias detection, where only a pretrained model and a textual description of the classification task are available, thus eliminating the need for annotated datasets or manual intervention. + +Explainability and Failure Mode Detection. A parallel research direction focuses on systematic failure mode discovery and interpretability [3, 4, 16, 25, 29]. VoG [3] highlights difficult examples by analyzing gradient variance, and Spotlight [16] identifies contiguous underperforming regions in representation space. Saliency-based methods [29] provide local visual explanations, while counterfactual explanations are generated in either natural language [25] or visual form using diffusion models [4]. + +Recent work has scaled these efforts using embeddings and generative models. Wiles et al. [65] leverages large-scale generative models to automatically find and describe failure clusters. Domino [19] uses cross-modal embeddings and mixture models to identify and describe coherent data slices where models underperform. Jain et al. [30] extracts latent directions corresponding to consistent model errors + +and generates challenging counterfactual examples for data augmentation. AdaVision [22] enables interactive discovery of failure modes by combining CLIP-based retrieval with GPT-generated prompts. PRIME [53] prioritizes interpretability by starting from human-defined tags and identifying minimal descriptive combinations that explain failures. MAIA [56] automates interpretability workflows using vision-language models for feature analysis and failure mode discovery. Csurka et al. [15] formalizes language-based error explainability and proposes evaluation metrics to assess the quality of generated explanations. + +While these methods analyze model behavior and failure patterns through model outputs or datasets, our work introduces a complementary approach: starting from a high-level task description, C2B uses LLM-based bias proposals and retrieval-based validation to automatically identify, validate, and quantify biases in pre-trained classifiers, enabling large-scale, unsupervised bias discovery across arbitrary classification tasks. + +# 3. Method + +In this section, we first define bias attributes and bias classes (Sec. 3.1), and formulate the problem of identifying biases of pre-trained models from their task specifications (Sec. 3.2). We then describe C2B, that proposes a list of biases (Sec. 3.3) and collects data for testing them (Sec. 3.4). Finally, we describe how we measure model's biases from the collected data (Sec. 3.5). + +# 3.1. Bias definition + +Formally, we denote the target classification model as $f: \mathcal{X} \to \mathcal{Y}$ , mapping an image in the space $\mathcal{X}$ to labels in the space $\mathcal{Y}$ . We define as bias attributes $\mathcal{B} = \{\mathbf{B}_1, \dots, \mathbf{B}_N\}$ the different categories of bias, i.e., factors that can impact its performance [44]. For each bias attribute, we define a list of bias classes $\mathbf{B}_i = \{b_{i,1}, \dots, b_{i,n_i}\}$ , the different values that each bias attribute can take. For a given target class, we define the model to be biased or unfair w.r.t specific bias classes depending on the per-class accuracy difference of the model. Let us denote the per-class accuracy of a model $f$ on a target class $y$ on data that matches a bias class $b_{i,j}$ as $A_y(f, b_{i,j})$ . We define the bias score $\phi_{y,i,j}$ for each bias class $b_{i,j}$ as + +$$ +\phi_ {y, i, j} = A _ {y} \left(f, b _ {i, j}\right) - \frac {1}{n _ {i} - 1} \sum_ {b _ {i, k} \in \mathbf {B} _ {i}, k \neq j} A _ {y} \left(f, b _ {i, k}\right). \tag {1} +$$ + +# 3.2. Problem formulation + +Given a pre-trained model $f$ , our goal is to identify its biases. We assume $f$ to be a generic classification model that we did not develop and that we aim to use (e.g., downloaded from the web). Thus, we also assume to have no access to + +additional data $f$ had access to (e.g., the dataset set used for training). As we want to avoid costly annotation procedures, we also assume to have no access to any dataset containing labeled images of the classes in $\mathcal{V}$ . + +To address this problem, we can rely only on what we know about the model: i.e., the task it has been designed for, the input domain it expects, and the output it predicts. This information can all be expressed using simple, natural language. Specifically, we assume to have a general description of the model's task $T$ , the input the model expects to process, and its output (i.e., classes in $\mathcal{V}$ ). + +This task is challenging as (i) detecting biases requires measuring the performance of the model on a labeled dataset, potentially containing annotations for the factors of interest and (ii) we have no pre-defined list of biases to start with. This implies that we have to both generate a list of potential biases and to collect a dataset to test them. In the following, we will describe our framework, C2B, that exploits LLMs for proposing biases, and a combination of LLMs and multimodal retrieval for scoring them. An overview of C2B is illustrated in Fig. 2. + +# 3.3. Generating bias proposals + +The first step of C2B is to generate a list of bias attributes and classes that we can explicitly test for. To generate the list, we cannot rely on attribute annotations, as we have no access to labeled data. Moreover, we cannot use lists available online, as biases depend on the task of interest. + +To address this problem, we exploit the reasoning capabilities of LLMs. Specifically, let us denote as $\Phi_{\mathrm{LLM}}$ a generic LLM, mapping textual inputs to textual outputs, i.e., $\Phi_{\mathrm{LLM}}: \mathcal{T} \to \mathcal{T}$ , with $\mathcal{T}$ being the space of natural language. To generate the candidate set of biases, we prompt the LLM to produce a list of bias attributes and classes for the given model and task description. Because different classes may have different biases, we query the LLM to give us a list of potential biases $\mathcal{B}^y$ for each class $y \in \mathcal{Y}$ . Formally, we have: + +$$ +\mathcal {B} ^ {y} = \Phi_ {\mathrm {L L M}} \left(\pi_ {\text {b i a s}} \circ T \circ y\right) \tag {2} +$$ + +where $\circ$ is the string concatenation operation, $\pi_{\mathrm{bias}}$ is the system prompt that instructs the LLM about the task it should perform (Supp. Mat., Appx. I), and $T$ is the model's task description, including input/output specification. + +The output of Eq. (2), $\mathcal{B}^y$ , is a list1 of bias attributes $\mathcal{B}^y = \{\mathbf{B}_1^y, \dots, \mathbf{B}_{N_y}^y\}$ , of size $N_y$ . We prompt the LLM so that each bias attribute is associated to a list of $n_{y,i}$ bias classes $\mathbf{B}_i^y = \{b_{i,1}^y, \dots, b_{i,n_{y,i}}^y\}$ . + +![](images/5f6973cc196021fad07be63facd83e17b2783b016e6560a09857c8222cbcf2c5.jpg) +Figure 2. Overview of C2B. Given a specific task with an associated description and a pre-trained classification model, our approach leverages an LLM to identify potential candidate bias attributes and corresponding classes. These candidate biases are then used to prompt the LLM to generate captions, which are subsequently utilized by a retrieval module to collect a dataset of images for bias testing. This dataset enables evaluation of the target classification model, producing scores for each identified bias category. + +# 3.4. Collecting data + +Given the list of candidate bias attributes and classes, we still need a dataset to score them. Ideally, for each target class $y \in \mathcal{V}$ and bias class $b^{y} \in \mathbf{B}^{y}$ , with $\mathbf{B}^{y} \in \mathcal{B}^{y}$ , we would like to have a dataset $D(b^{y}, y)$ depicting images of class $y$ exhibiting the bias class $b^{y}$ , i.e., $D(b^{y}, y) = \{(x_{i}, y_{i}, b_{i}^{y})\}_{i=1}^{M}$ of size $M$ . Unfortunately, we cannot rely on labeled data for this task as (i) we do not have access to it; (ii) biases change w.r.t. the particular application, thus it is very hard to collect beforehand a dataset containing all possible target and bias classes of interest. + +To address this problem, we resort to external sources. In particular, we assume to have access to a large-scale database $\mathcal{D}$ containing images. Note that $\mathcal{D}$ has no annotations: we just assume it to be large-scale (i.e., potentially web-scale). From this database, we can collect images for each target-bias combination of interest using a text-to-image retrieval model. This is achieved in two steps: captions generation and cross-modal retrieval. + +Caption generation. A straightforward solution to create captions is to use a fixed template (e.g., an image of [target class] with [bias class]). However, this would not consider the specificity of the considered task and thus may lead to retrieving erroneous images. For instance, the biases and captions associated to a facial attribute classification task should be different from those in a natural scene classification task. To avoid this issue, we resort to using the LLM, querying it in two steps: first to create a template for the specific task, and then to gener + +ate captions specific for target and bias class. Formally, we obtain the initial template as: + +$$ +\operatorname {t e m p} = \Phi_ {\mathrm {L L M}} \left(\pi_ {\text {t e m p l a t e}} \circ T\right) \tag {3} +$$ + +where $\pi_{\mathrm{template}}$ is the system prompt that instructs the LLM about the task it should perform (Supp. Mat., Appx. 1). Then, we ask the LLM to refine this template for each specific target class $y$ and bias class $b$ , as follows: + +$$ +c _ {b} ^ {y} = \Phi_ {\mathrm {L L M}} \left(\pi_ {\text {c a p t i o n}} \circ \operatorname {t e m p} \circ b \circ y\right). \tag {4} +$$ + +where $c_b^y$ is the resulting caption. Note that, while with this strategy we obtain one caption per target-bias class pair, in principle we can also generate multiple captions per pair to have more variability, beneficial for more complex tasks. + +Image retrieval. The obtained captions can be used to retrieve images from a large scale image database, and we implement the retrieval by using a contrastive vision-language model (VLM) that can score the similarity between a text and image. Formally, we can decompose the VLM in three elements: a visual encoder $\varphi_{\mathrm{vis}}:\mathcal{X}\to \mathbb{R}^d$ a text encoder $\varphi_{\mathrm{txt}}:\mathcal{T}\rightarrow \mathbb{R}^d$ , and a similarity function $\sin :\mathbb{R}^d\times \mathbb{R}^d\to \mathbb{R}$ (e.g., cosine similarity for CLIP [51]). Given a caption, we can retrieve a set of images $\mathbf{X}_b^y$ for each class label $y$ and bias $b$ as follows: + +$$ +\mathbf {X} _ {b} ^ {y} = \underset {x \in \mathcal {D}} {\operatorname {t o p - k}} \operatorname {s i m} \left(\varphi_ {\text {v i s}} (x), \varphi_ {\text {t x t}} \left(c _ {b} ^ {y}\right)\right), \tag {5} +$$ + +where we retrieve the $k$ images with the highest matching score. While these images do not have a ground-truth, + +they can still help to estimate the biases of the model, with the fidelity of the latter depending on the retrieval accuracy (Sec. 4.4). We also experiment with a different type of retrieval system, a Web search engine (details in Sec. 4.1). + +# 3.5. Measuring biases + +In the previous section, we have collected a dataset with pseudo-annotations for all bias values and classes of interest. We can now test the target model on this dataset to compute its performance, using the variation across bias classes to score each bias. We use the formula defined in Eq. (1) to score each target-bias pair relative to other target-bias pairs. For each bias class, we obtain a bias score between -1 and 1. If the per-class accuracy is higher for a given bias class, the classifier is biased toward the bias class (positive bias score). Reciprocally, if the per-class accuracy is lower for a given bias class, the classifier is biased against the bias class (negative bias score). Intuitively, larger absolute bias score values also correspond to stronger biases. + +To quantitatively evaluate the quality of the biases detected by C2B, we propose two novel evaluation settings that complement each other. First, in Sec. 4.2.1, we propose to match the detected biases with ground-truth annotations using SBERT embeddings for tasks where datasets with additional bias annotations exist. Second, in Sec. 4.2.2, we propose to assign bias labels (based on the initial proposed biases) to ground-truth task-specific data with a VQA model, to check the agreement between the bias detection method and an expert VQA model. + +Collecting scores for all bias attributes and classes enables various types of analyses to explore the biases. Biases can be ranked by score, to assess which attributes are more critical than others. Bias class scores can also be aggregated across bias attributes, to measure which bias attributes exhibit a larger bias. We provide examples of qualitative analyses in Sec. 4.3, and more in the Supp. Mat., Appx. D. + +# 4. Experiments + +In this section, we present the results of our experiments, where we apply C2B to detect biases for different models on two different image classification tasks. We first describe our experimental protocol (Sec. 4.1), we then quantitatively compare the biases detected by C2B with a state-of-the-art competitor (Sec. 4.2), and we provide qualitative examples of detected biases (Sec. 4.3). Finally, we analyze the core component of our approach (Sec. 4.4). + +# 4.1. Experimental protocol + +Tasks and models. We test C2B on two different tasks: facial attribute classification and image classification. + +Facial attribute classification. We consider the CelebA dataset [41], which is composed of 200k images of celebrity + +faces annotated with 40 facial attributes, such as male, blond hair, young, or high cheekbones. To evaluate C2B, we choose to detect the biases of a recent state-of-the-art model, FaceXFormer [48]. This model relies on a pretrained Swin/B [42] backbone, coupled with an MLP-based fusion module and an attention-based decoder, which are co-trained on several face datasets. + +Image classification. We focus on the widely used ImageNet-1K [54] dataset, where each image is associated to one of the 1000 ImageNet classes. We choose to detect the biases of four widely used models, easily accessible through TorchVision [43], comprising both convolutional networks and transformers: ResNet50_V2 [23], ResNet101_V2 [23], ResNet152_V2 [23], ViT_B_16_SWAG_E2E_V1 [58]. + +Competitors. As we introduce the new problem of unsupervised bias detection relying solely on a textual description of the classification task, there are no existing baseline methods designed for this specific setting. Nevertheless, we still compare C2B to a recent supervised open-set bias detection method, Bias-to-Text (B2T) [33]. However, unlike C2B which does not require task-specific labeled data to identify biases in an image classifier, B2T depends on ground-truth task labels to identify model failures. It then employs a pre-trained captioning model to generate descriptions of misclassified images, extracts common keywords, and computes a bias score based on the difference between (i) the similarity (CLIP score) between the discovered bias keyword and misclassified images, and (ii) the similarity between the discovered bias keyword and well-classified images. Thus, we test B2T using the labeled validation sets of the two datasets (i.e., CelebA and ImageNet). + +Implementation details. C2B comprises two components: the LLM that proposes biases and captions, and the retrieval module. For the former, we use Llama 3.1 8B [18] (see Supp. Mat., Appx. J for qualitative LLM comparisons). For the retrieval, we explore two strategies: (i) CC12M: we consider the large Conceptual Captions 12M (CC12M) dataset [12] and retrieve images based on CLIP embeddings [7], and (ii) BING: instead of using a VLM to retrieve images from an existing database where some image domains may not be well-represented, we propose to retrieve images from the Web using a publicly available search engine, Bing. For both datasets, we retrieve 20 images per caption, and we measure biases as in Eq. (1), comparing the difference in the model's per-class accuracy between images that match different candidate bias classes. + +# 4.2. Quantitative evaluation + +In this section, we provide a quantitative evaluation to assess the quality of biases identified by C2B. First, we present results based on ground-truth bias annotations, i.e., + +exploiting attributes annotated in the validation sets of the datasets. For biases not in the ground-truth annotations, we introduce an evaluation based on Visual Question Answering (VQA) to measure the relevance of the proposed biases. + +# 4.2.1 Evaluation with ground-truth annotations + +Setting. In this setting, we exploit the available attribute annotations, checking if models can detect and quantify those biases correctly. In the case of CelebA, when we focus on a specific attribute classification task (e.g., blonde hair) we consider all other 39 attributes (e.g., young) as biases measurable with the ground-truth. For ImageNet-1K, we use ImageNet-X [28], i.e., the validation set of ImageNet-1K with 16 labeled "factors" distinguishing images from typical prototypes (e.g., orientation), as ground-truth biases. + +To obtain ground-truth biases, we test the model on each validation set and check the difference in per-class accuracy among images where the bias attribute is present and images where it is absent. We consider a ground-truth bias to be present if this difference is higher than a threshold (i.e., $\tau = 0.05$ ) using the same threshold to assess bias identification on both C2B and B2T scores. We refer to the sign of the bias score as the direction of the bias. We provide more results with a varying threshold in the Supp. Mat., Appx. C. + +Due to the open-set nature of the detected biases for both C2B and B2T, there are almost no exact matches between them and ground-truth biases. To compare the two sets, we compute the cosine similarity between SBERT embeddings [52] of the ground-truth and detected biases. We empirically set the similarity threshold to 0.9 to avoid having too many false positives. We provide more details and qualitative examples of matches in the Supp. Mat., Appx. K. + +Metrics. We report results from two complementary perspectives: GT $\rightarrow$ Detected and Detected $\rightarrow$ GT. In GT $\rightarrow$ Detected, we start from each ground-truth bias and check whether it was correctly detected with the same bias direction (HIT), detected with the opposite direction (false hit, FH), or not detected (M1ss). This perspective focuses on recall. However, methods that propose many biases may artificially reduce misses. Therefore, we also report Detected $\rightarrow$ GT, where we start from each detected bias and verify whether it matches a ground-truth bias with the correct direction. This second view measures precision and helps assess the quality of the proposed biases. + +Results. In Tabs. 1 and 2, we report the results of our analysis. On CelebA (left side of Tab. 1, GT $\rightarrow$ Detected), C2B outperforms B2T in detecting ground-truth biases, regardless of the retrieval source. We obtain the best results with Bing, detecting between $10.8\%$ and $12.3\%$ of the ground-truth biases, compared to only $4.5\%$ for B2T. However, B2T exhibits a lower proportion of false hits $(3.1\%)$ , whereas + +Table 1. Proportion $(\%)$ of ground-truth biases detected on CelebA $\mathrm{(GT\to Detected)}$ and of detected biases corresponding to groundtruth ones (Detected $\rightarrow$ GT). FH=False Hit. + +
METHODGT → DETECTEDDETECTED → GT
HIT(↑)FH(↓)MISS (↓)HIT(↑)FH(↓)MISS(↓)
FaceXFormer
B2T [33]4.533.1492.326.594.1089.29
C2B (BING)12.296.8880.8314.187.3478.48
C2B (CC12M)10.767.7581.4912.758.1579.09
+ +Table 2. Proportion (\%) of ground-truth biases detected on ImageNet-X (GT $\rightarrow$ Detected) and of detected biases corresponding to ground-truth ones (Detected $\rightarrow$ GT). FH=False Hit. + +
METHODGT → DETECTEDDETECTED → GT
HIT(↑)FH(↓)Miss (↓)HIT(↑)FH (↓)Miss(↓)
ResNet50_V2
B2T [33]2.402.1995.410.800.7198.49
C2B (BING)7.808.5383.662.602.9894.42
C2B (CC12M)11.1811.3077.522.992.8794.14
ResNet101_V2
B2T [33]2.552.1195.330.840.6998.47
C2B (BING)7.918.4183.682.562.7794.67
C2B (CC12M)11.2111.6877.112.762.9694.28
ResNet152_V2
B2T [33]2.531.7295.750.830.5898.59
C2B (BING)7.537.9384.542.752.7994.46
C2B (CC12M)11.1012.0176.902.812.9594.24
ViT_B_16_SWAG
B2T [33]2.111.9895.900.720.6798.61
C2B (BING)7.727.1885.102.482.4895.05
C2B (CC12M)10.8511.0778.082.642.7094.66
+ +C2B shows slightly higher false hit rates (6.9–7.8%). This is largely due to retrieval noise and the broader search space introduced by LLM-generated proposals, which can sometimes lead to less precise bias candidates. Nevertheless, C2B misses fewer ground-truth biases (80.8–81.5% vs. B2T's 92.3%). These trends remain consistent when considering the proportion of detected biases that correspond to ground-truth biases (Detected → GT), where C2B achieves 12.8–14.2%, compared to 6.6% for B2T. + +On ImageNet-X (Tab. 1), across four classifiers, C2B again detects more ground-truth biases than B2T (7.5–11.2% vs. 2.1–2.6%), at the cost of higher false hit rates (7.2–12.0% vs. 1.7–2.2%). As before, these false hits are attributable to noisy retrieval and the open-ended nature of bias proposals. Still, C2B misses significantly fewer ground-truth biases (76.9–85.1% vs. 95.3–95.9%). Interestingly, on ImageNet-X, retrieving from CC12M provides better results than Bing. The Detected → GT evaluation confirms these findings, with a larger fraction of C2B-detected biases aligning with ground-truth ones (2.5–3.0% vs. 0.6–0.7% for B2T). + +Discussion. Due to the limited human annotations available in CelebA and ImageNet-X, the proportion of matching biases is low across all methods, particularly on ImageNet-X. Furthermore, since we are using text embeddings to + +Table 3. Detected biases and VQA agreement on CelebA. + +
MODELB2T [33]C2B (BING)C2B (CC12M)
FaceXFormer0.100.220.25
+ +Table 4. Detected biases and VQA agreement on ImageNet-X. + +
MODELB2T [33]C2B (BING)C2B (CC12M)
ResNet50_V20.200.260.20
ResNet101_V20.190.280.20
ResNet152_V20.200.280.20
ViT_B_16_SWAG0.200.320.21
+ +match the biases, the numbers are also influenced by the semantic proximity between LLM-proposed biases, caption-extracted keywords, and ground-truth annotations. Moreover, a high number of Miss in the Detected $\rightarrow$ GT case is not necessarily a bad thing, as C2B and B2T may be detecting accurate biases that are not present in the annotation. For these reasons, in the next section we introduce a different evaluation based on visual question answering. + +# 4.2.2 VQA-based evaluation + +As ground-truth bias annotations may not always be available, we also propose an alternative evaluation scheme based on VQA, measuring the quality of the proposed biases that are not present in ground-truth annotations. + +Metric. We use a VQA model (LLaVA-1.5-13B [38, 39]) to pseudo-label each image of the validation sets with each of the proposed bias by C2B or B2T. Additional details can be found in the Supp. Mat., Appx. G. From these pseudolabels, we can use the ground-truth task labels (i.e., class annotations) to check how the per-class accuracy varies over different bias classes, using the same threshold $\tau = 0.05$ of Sec. 4.2.1. For each bias class $b$ proposed by the model, there are three cases: a positive bias is detected $(\phi_b \geq \tau)$ , a negative bias is detected $(\phi_b \leq -\tau)$ , or the bias is not detected $(|\phi_b| < \tau)$ . For each bias class, we define the agreement score to be 1 if both scores fall in the same case, -1 if they detect opposite biases, and 0 otherwise. + +Results. Tabs. 3 and 4 present the results of our VQA-based evaluation on CelebA and ImageNet-X. On CelebA, C2B achieves higher agreement with VQA pseudo-labels (0.22 for Bing and 0.25 for CC12M) than B2T (0.10), indicating that the biases discovered by C2B are more aligned with visually detectable and semantically coherent bias attributes. On ImageNet-X, while C2B often achieves higher agreement scores than B2T with Bing (i.e., +0.09 on average), results with CC12M are either the same or comparable to those of B2T (i.e., +0.005 on average). We attribute this to the challenge of relying on retrieval-based data collection, without access to annotated failure cases or + +task-specific captions. In contrast, B2T benefits from captions derived from known misclassifications. Despite this more challenging and fully unsupervised setting, C2B still achieves agreements that are competitive with or superior to B2T. This confirms that C2B's proposals are not only meaningful and visually grounded but also relevant for bias detection in settings where no annotations are available. + +Discussion. The results presented in this section are influenced by both the types of questions posed to the VQA model and the model's own inherent biases. Given these limitations, evaluations based on ground-truth annotations and those derived from VQA outputs serve as complementary approaches. Overall, biases proposed by the LLM tend to offer more relevant bias indicators than keywords extracted from captions as in B2T [33], allowing C2B to detect a greater proportion of ground-truth biases and achieve higher agreement with the VQA model than B2T. + +# 4.3. Qualitative evaluation + +In this section, we show examples of biases discovered by C2B for different models on different tasks. + +In Fig. 3, we show biases that C2B detects for the high cheekbones attribute on face attribute classification, with bias scores from different retrieval methods. We can see that a strong bias was detected on the expression attribute: the classifier appears to perform better when people are smiling and worse when they appear angry. Other smaller biases were detected, such as young, and male. + +Fig. 4 shows a similar analysis for biases detected by C2B for the minibus class on image classification for the ResNet50_V2 model. Strong biases were detected on the lighting, orientation, and presence of reflection bias attribute: the classifier appears stronger on daytime images, portrait-oriented images, and images without reflections, while performing worse on nighttime images, landscape-oriented images, and images with reflections. Finally, Fig. 5 depicts biases for the birdhouse class again for image classification, but for the ViT_B_16_SWAG model. A strong bias is found on the shape bias attribute: the classifier seems to be performing better on images of rectangular birdhouses, while it underperforms on images of spherical birdhouses. + +# 4.4. Evaluation of VLM-based retrieval + +Retrieving images matching the target classes and biases present in the captions is essential for the detected biases to be trustworthy. Therefore, in this section, we analyze the CLIP-based retrieval approach used by C2B, by trying to retrieve images for which we have annotations. For a VQA-based analysis of both the Bing-based retrieval and CLIP-based retrieval on CC12M, please refer to Supp. Mat., Appx. F. To perform our analysis, we use CLIP to retrieve images from the validation sets of CelebA and ImageNet- + +![](images/517b6ddb58ba735e614704926cceb229630dc2224ccf7d35b1a7646f0150073c.jpg) +Figure 3. Examples of biases found by C2B for the high cheekbones target attribute on CelebA. + +![](images/1005d94e80770f02c61266a1f5476eed61590f5c79e0f5812e1d8c7849ac584b.jpg) +Figure 4. Examples of biases found by C2B for the minibus class on ImageNet-X (ResNet50_V2). + +![](images/40b395fe3a813bd492853383acccf65cabe9055b7085a593f2b01416b297e14a.jpg) +Figure 5. Examples of biases found by C2B for the birdhouse class on ImageNet-X (ViT_B_16_SWAG). + +X, using handcrafted captions as input: combinations of attributes on CelebA, and classes and factors on ImageNet-X. + +Metric. We use recall@ $K$ , where $K$ varies for each caption to match the number of corresponding ground-truth samples in the dataset. We show results using recall @ $A \cdot K$ , where $A$ is a coefficient that varies between 0.01 and 1.0. + +Results. We present the results in Fig. 6. For CelebA (left), the recall is the same for both target and bias attributes, as expected, since both are drawn from the same set of attributes. On the same dataset, CLIP struggles to + +![](images/69ace24a8a203f0e7f6e92391a298ddf676ffe6c72f6e9d9589910b3f69c440c.jpg) +Figure 6. Accuracy of VLM-based retrieval. + +![](images/5e0928000fd6dc910aaa55266bbb7847b560712dce1ead30e13f9b66bad03d5f.jpg) + +retrieve images with combinations of attributes, with a recall @ $0.01 \cdot K$ just above $40\%$ . For ImageNet-X (right), CLIP is fairly accurate, especially for low values of $A$ , with more than $80\%$ and $70\%$ recall at $0.05 \cdot K$ to retrieve images with specific biases and classes, respectively. Generally, the recall is higher when the number of retrieved images decreases relative to the number of potential matching images present in the dataset, which suggests that, the larger the dataset, the more accurate the retrieval is. + +# 5. Conclusion + +We introduced CLASSIFIER-TO-BIAS (C2B), a novel method for automatically detecting biases in visual classifiers entirely in an unsupervised manner, without the need for task-specific labeled data. Starting from a simple textual description of the classification task and a pre-trained model, C2B generates bias hypotheses, retrieves relevant images to test them, and produces a list of biases with their scores. Although C2B is the first approach to operate in such challenging setting, our results show that it surpasses a recent supervised open-set bias detector across two evaluation settings: one based on ground-truth annotations and another on VQA-based pseudo-labels. We believe our work marks an important step toward automated, open-set, and task-agnostic bias detection. + +# Acknowledgments + +This work was sponsored by the MUR PNRR project FAIR - Future AI Research (PE00000013), funded by NextGeneration, and the Italian Ministerial grants PRIN 2022: "B-FAIR: Bias-Free Artificial Intelligence methods for automated visual Recognition" (CUP E53D23008010006), and the EU project ELIAS (No.01120237). + +# References + +[1] Marah Abdin, Sam Ade Jacobs, Ammar Ahmad Awan, Jyoti Aneja, Ahmed Awadallah, Hany Awadalla, Nguyen Bach, Amit Bahree, Arash Bakhtiari, Harkirat Behl, et al. Phi-3 technical report: A highly capable language model locally on your phone. arXiv preprint arXiv:2404.14219, 2024. 15 +[2] Julius A Adebayo et al. *FairML: ToolBox for diagnosing bias in predictive modeling*. PhD thesis, Massachusetts Institute of Technology, 2016. 1 +[3] Chirag Agarwal, Daniel D'souza, and Sara Hooker. Estimating example difficulty using variance of gradients. In CVPR, 2022. 2 +[4] Maximilian Augustin, Valentyn Boreiko, Francesco Croce, and Matthias Hein. Diffusion visual counterfactual explanations. NeurIPS, 2022. 2 +[5] Razvan Azamfirei, Sapna R. Kudchadkar, and James Fackler. Large language models and the perils of their hallucinations. Critical Care, 27(1):120, 2023. 2 +[6] Yujia Bao and Regina Barzilay. Learning to split for automatic bias detection. arXiv preprint arXiv:2204.13749, 2022. 1 +[7] Romain Beaumont. Clip retrieval: Easily compute clip embeddings and build a clip retrieval system with them. https://github.com/rom1504/clip-retrieval, 2022.5 +[8] R. K. E. Bellamy, K. Dey, M. Hind, S. C. Hoffman, S. Houde, K. Kannan, P. Lohia, J. Martino, S. Mehta, A. Mojsilovic, S. Nagar, K. Natesan Ramamurthy, J. Richards, D. Saha, P. Sattigeri, M. Singh, K. R. Varshney, and Y. Zhang. Ai fairness 360: An extensible toolkit for detecting and mitigating algorithmic bias. IBM Journal of Research and Development, 63(4/5):4:1-4:15, 2019. 1 +[9] Tolga Bolukbasi, Kai-Wei Chang, James Zou, Venkatesh Saligrama, and Adam Kalai. Man is to computer programmer as woman is to homemaker? debiasing word embeddings. In NeurIPS, 2016. 1 +[10] Joy Buolamwini and Timnit Gebru. Gender shades: Intersectional accuracy disparities in commercial gender classification. In Proceedings of the 1st Conference on Fairness, Accountability and Transparency, 2018. 1 +[11] Angel Alexander Cabrera, Will Epperson, Fred Hohman, Minsuk Kahng, Jamie Morgenstern, and Duen Horng Chau. Fairvis: Visual analytics for discovering intersectional bias in machine learning. In 2019 IEEE Conference on Visual Analytics Science and Technology (VAST), pages 46-56, 2019. +[12] Soravit Changpinyo, Piyush Sharma, Nan Ding, and Radu Soricut. Conceptual 12M: Pushing web-scale image-text + +pre-training to recognize long-tail visual concepts. In CVPR, 2021.5 +[13] Davide Chicco, Niklas Totsch, and Giuseppe Jurman. The matthews correlation coefficient (mcc) is more reliable than balanced accuracy, bookmaker informedness, and markedness in two-class confusion matrix evaluation. BioData Mining, 14(1):13, 2021. 11 +[14] Massimiliano Ciranni, Luca Molinaro, Carlo Alberto Barbano, Attilio Fiandrotti, Vittorio Murino, Vito Paolo Pastore, and Enzo Tartaglione. Say my name: a model's bias discovery framework. arXiv preprint arXiv:2408.09570, 2024. 2 +[15] Gabriela Csurka, Tyler L Hayes, Diane Larlus, and Riccardo Volpi. What could go wrong? discovering and describing failure modes in computer vision. In ECCV-WS, 2024. 3 +[16] Greg d'Eon, Jason d'Eon, James R Wright, and Kevin Leyton-Brown. The spotlight: A general method for discovering systematic errors in deep learning models. In Proceedings of the 2022 ACM Conference on Fairness, Accountability, and Transparency, pages 1962-1981, 2022. 1, 2 +[17] Moreno D'Inca, Elia Peruzzo, Massimiliano Mancini, Dejia Xu, Vedit Goel, Xingqian Xu, Zhangyang Wang, Humphrey Shi, and Nicu Sebe. Openbias: Open-set bias detection in text-to-image generative models. In CVPR, 2024. 1, 2 +[18] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 5, 15 +[19] Sabri Eyuboglu, Maya Varma, Khaled Kamal Saab, Jean-Benoit Delbrouck, Christopher Lee-Messer, Jared Dunnmon, James Zou, and Christopher Re. Domino: Discovering systematic errors with cross-modal embeddings. In ICLR, 2022. 1, 2 +[20] Samir Yitzhak Gadre, Gabriel Ilharco, Alex Fang, Jonathan Hayase, Georgios Smyrnis, Thao Nguyen, Ryan Marten, Mitchell Wortsman, Dhruba Ghosh, Jieyu Zhang, Eyal Orgad, Rahim Entezari, Giannis Daras, Sarah Pratt, Vivek Ramanujan, Yonatan Bitton, Kalyani Marathe, Stephen Mussmann, Richard Vencu, Mehdi Cherti, Ranjay Krishna, Pang Wei W Koh, Olga Saukh, Alexander J Ratner, Shuran Song, Hannaneh Hajishirzi, Ali Farhadi, Romain Beaumont, Sewoong Oh, Alex Dimakis, Jenia Jitsev, Yair Carmon, Vaishaal Shankar, and Ludwig Schmidt. Datacomp: In search of the next generation of multimodal datasets. In Advances in Neural Information Processing Systems, pages 27092-27112, 2023. 2 +[21] Isabel O. Gallegos, Ryan A. Rossi, Joe Barrow, Md Mehrab Tanjim, Sungchul Kim, Franck Dernoncourt, Tong Yu, Ruiyi Zhang, and Nesreen K. Ahmed. Bias and fairness in large language models: A survey. Computational Linguistics, 50 (3):1097-1179, 2024. 2 +[22] Irena Gao, Gabriel Ilharco, Scott Lundberg, and Marco Tulio Ribeiro. Adaptive testing of computer vision models. In ICCV, 2023. 1, 3 +[23] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 5 + +[24] Lisa Anne Hendricks, Kaylee Burns, Kate Saenko, Trevor Darrell, and Anna Rohrbach. Women also snowboard: Overcoming bias in captioning models. In ECCV, 2018. 1 +[25] Lisa Anne Hendricks, Ronghang Hu, Trevor Darrell, and Zeynep Akata. Generating counterfactual explanations with natural language. In ICML-WS, 2018. 2 +[26] Cheng-Yu Hsieh, Jieyu Zhang, Zixian Ma, Aniruddha Kembhavi, and Ranjay Krishna. Sugarcrepe: Fixing hackable benchmarks for vision-language compositionality. NeurIPS, 2024. 2 +[27] https://huggingface.co/. 1 +[28] Badr Youbi Idrissi, Diane Bouchacourt, Randall Balestriero, Ivan Evtimov, Caner Hazirbas, Nicolas Ballas, Pascal Vincent, Michal Drozdjal, David Lopez-Paz, and Mark Ibrahim. Imagenet-x: Understanding model mistakes with factor of variation annotations. In ICLR, 2023. 1, 2, 6 +[29] Saachi Jain, Hadi Salman, Eric Wong, Pengchuan Zhang, Vibhav Vineet, Sai Vermprala, and Aleksander Madry. Missingness bias in model debugging. In ICLR, 2022. 2 +[30] Saachi Jain, Hannah Lawrence, Ankur Moitra, and Aleksander Madry. Distilling model failures as directions in latent space. In ICLR, 2023. 1, 2 +[31] Ziwei Ji, Nayeon Lee, Rita Frieske, Tiezheng Yu, Dan Su, Yan Xu, Etsuko Ishii, Ye Jin Bang, Andrea Madotto, and Pascale Fung. Survey of hallucination in natural language generation. ACM Comput. Surv., 55(12), 2023. 2 +[32] Krish Kabra, Kathleen M. Lewis, and Guha Balakrishnan. Gelda: A generative language annotation framework to reveal visual biases in image generators. In CVPRW, 2024. 2 +[33] Younghyun Kim, Sangwoo Mo, Minkyu Kim, Kyungmin Lee, Jaeho Lee, and Jinwoo Shin. Discovering and mitigating visual biases through keyword explanation. In CVPR, 2024. 1, 2, 5, 6, 7, 4 +[34] Arvindkumar Krishnakumar, Viraj Prabhu, Sruthi Sudhakar, and Judy Hoffman. Udis: Unsupervised discovery of bias in deep visual recognition models. In BMVC, 2021. 1, 2 +[35] Zhiheng Li and Chenliang Xu. Discover the unknown biased attribute of an image classifier. In ICCV, 2021. 2 +[36] Zhiheng Li, Anthony Hoogs, and Chenliang Xu. Discover and mitigate unknown biases with debiasing alternate networks. In ECCV, 2022. 2 +[37] Evan Z Liu, Behzad Haghgoo, Annie S Chen, Aditi Raghunathan, Pang Wei Koh, Shiori Sagawa, Percy Liang, and Chelsea Finn. Just train twice: Improving group robustness without training group information. In ICML, 2021. 2 +[38] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, 2023. 7, 11 +[39] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In CVPR, 2024. 7, 11 +[40] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava next: Improved reasoning,OCR, and world knowledge, 2024. 11 +[41] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaou Tang. Deep learning face attributes in the wild. In ICCV, 2015. 2, 5 + +[42] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, 2021. 5 +[43] TorchVision maintainers and contributors. Torchvision: Pytorch's computer vision library. https://github.com/pytorch/vision, 2016.5 +[44] Ninareh Mehrabi, Fred Morstatter, Nripsuta Saxena, Kristina Lerman, and Aram Galstyan. A survey on bias and fairness in machine learning. ACM computing surveys (CSUR), 54 (6):1-35, 2021. 3 +[45] Thomas Mesnard, Cassidy Hardin, Robert Dadashi, Surya Bhupatiraju, Shreya Pathak, Laurent Sifre, Morgane Rivière, Mihir Sanjay Kale, Juliette Love, et al. Gemma: Open models based on gemini research and technology. arXiv preprint arXiv:2403.08295, 2024. 15 +[46] Microsoft. Chat markup language chatml (preview), 2024. https://learn.microsoft.com/en-us/azure/ai-services/openai/how-to/chat-markup-language, Last accessed on 2024-11-21. 13 +[47] Margaret Mitchell, Simone Wu, Andrew Zaldivar, Parker Barnes, Lucy Vasserman, Ben Hutchinson, Elena Spitzer, Inioluwa Deborah Raji, and Timnit Gebru. Model cards for model reporting. In Proceedings of the conference on fairness, accountability, and transparency, pages 220-229, 2019. 1 +[48] Kartik Narayan, Vibashan VS, Rama Chellappa, and Vishal M Patel. Facexformer: A unified transformer for facial analysis. arXiv preprint arXiv:2403.12960, 2024. 5 +[49] Ziad Obermeyer, Brian Powers, Christine Vogeli, and Sendhil Mullainathan. Dissecting racial bias in an algorithm used to manage the health of populations. Science, 366(6464): 447-453, 2019. 1 +[50] Jesutofunmi A. Omiye, Jenna C. Lester, Simon Spichak, Veronica Rotemberg, and Roxana Daneshjou. Large language models propagate race-based medicine. npj Digital Medicine, 6(1):195, 2023. 2 +[51] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 4 +[52] Nils Reimers and Iryna Gurevych. Sentence-bert: Sentence embeddings using siamese bert-networks. In ENMLP, 2019. 6 +[53] Keivan Rezaei, Mehrdad Saberi, Mazda Moayeri, and Soheil Feizi. PRIME: Prioritizing interpretability in failure mode extraction. In ICLR, 2024. 1, 3 +[54] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, Alexander C. Berg, and Li Fei-Fei. ImageNet Large Scale Visual Recognition Challenge. IJCV, 115(3):211-252, 2015. 5 +[55] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortman, Patrick Schramowski, Srivatsa Kundurthy, Katherine + +Crowson, Ludwig Schmidt, Robert Kaczmarczyk, and Jenia Jitsev. Laion-5b: An open large-scale dataset for training next generation image-text models. In Advances in Neural Information Processing Systems, pages 25278-25294, 2022. 2 +[56] Tamar Rott Shaham, Sarah Schwettmann, Franklin Wang, Achyuta Rajaram, Evan Hernandez, Jacob Andreas, and Antonio Torralba. A multimodal automated interpretability agent. In ICML, 2024. 3 +[57] Robik Shrestha, Yang Zou, Qiuyu Chen, Zhiheng Li, Yusheng Xie, and Siqi Deng. Fairrag: Fair human generation via fair retrieval augmentation. In CVPR, 2024. 2 +[58] Mannat Singh, Laura Gustafson, Aaron Adcock, Vinicius de Freitas Reis, Bugra Gedik, Raj Prateek Kosaraju, Dhruv Mahajan, Ross Girshick, Piotr Dólár, and Laurens van der Maaten. Revisiting weakly supervised pre-training of visual perception models. In CVPR, 2022. 5 +[59] Sahil Singla and Soheil Feizi. Salient imagenet: How to discover spurious features in deep learning? In ICLR, 2022. 2 +[60] Sahil Singla, Besmira Nushi, Shital Shah, Ece Kamar, and Eric Horvitz. Understanding failures of deep networks via robust feature extraction. In CVPR, 2021. 1 +[61] Nimit Sohoni, Jared Dunnmon, Geoffrey Angus, Albert Gu, and Christopher Ré. No subclass left behind: Fine-grained robustness in coarse-grained classification problems. NeurIPS, 2020. 2 +[62] Tristan Thrush, Ryan Jiang, Max Bartolo, Amanpreet Singh, Adina Williams, Douwe Kiela, and Candace Ross. Winoground: Probing vision and language models for visio-linguistic compositionality. In CVPR, 2022. 2 +[63] Vijay Vasudevan, Benjamin Caine, Raphael Gontijo Lopes, Sara Fridovich-Keil, and Rebecca Roelofs. When does dough become a bagel? analyzing the remaining mistakes on imagenet. In NeurIPS, 2022. 1 +[64] Angelina Wang, Alexander Liu, Ryan Zhang, Anat Kleiman, Leslie Kim, Dora Zhao, Iroha Shirai, Arvind Narayanan, and Olga Russakovsky. Revise: A tool for measuring and mitigating bias in visual datasets. *IJCV*, 130(7):1790–1810, 2022. 1 +[65] Olivia Wiles, Isabela Albuquerque, and Sven Gowal. Discovering bugs in vision models using off-the-shelf image generation and captioning. In NeurIPS-WS, 2022. 2 +[66] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 11975-11986, 2023. 2 +[67] Yue Zhang, Yafu Li, Leyang Cui, Deng Cai, Lemao Liu, Tingchen Fu, Xinting Huang, Enbo Zhao, Yu Zhang, Yu-long Chen, et al. Siren's song in the ai ocean: a survey on hallucination in large language models. arXiv preprint arXiv:2309.01219, 2023. 2 +[68] Zeliang Zhang, Mingqian Feng, Zhiheng Li, and Chenliang Xu. Discover and mitigate multiple biased subgroups in image classifiers. In CVPR, 2024. 1, 2 +[69] Dora Zhao, Angelina Wang, and Olga Russakovsky. Understanding and evaluating racial biases in image captioning. In ICCV, 2021. 1 + +[70] Jieyu Zhao, Tianlu Wang, Mark Yatskar, Vicente Ordonez, and Kai-Wei Chang. Men also like shopping: Reducing gender bias amplification using corpus-level constraints. In EMNLP, 2017. 1 +[71] Zaiying Zhao, Soichiro Kumano, and Toshihiko Yamasaki. Language-guided detection and mitigation of unknown dataset bias. arXiv preprint arXiv:2406.02889, 2024. 2 \ No newline at end of file diff --git a/2025/Classifier-to-Bias_ Toward Unsupervised Automatic Bias Detection for Visual Classifiers/images.zip b/2025/Classifier-to-Bias_ Toward Unsupervised Automatic Bias Detection for Visual Classifiers/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..81cf07c590d6403d0c04ec04ff65ec49f41689a2 --- /dev/null +++ b/2025/Classifier-to-Bias_ Toward Unsupervised Automatic Bias Detection for Visual Classifiers/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3d2bfa4baa03a8f499b28a4e5b4227947b4c10db1323dbcdccf37504d1c9cd10 +size 427942 diff --git a/2025/Classifier-to-Bias_ Toward Unsupervised Automatic Bias Detection for Visual Classifiers/layout.json b/2025/Classifier-to-Bias_ Toward Unsupervised Automatic Bias Detection for Visual Classifiers/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..3d9d5320368da0ee34a6b0f85196da8198a49077 --- /dev/null +++ b/2025/Classifier-to-Bias_ Toward Unsupervised Automatic Bias Detection for Visual Classifiers/layout.json @@ -0,0 +1,8933 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 129, + 103, + 481, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 103, + 481, + 138 + ], + "spans": [ + { + "bbox": [ + 129, + 103, + 481, + 138 + ], + "type": "text", + "content": "CLASSIFIER-TO-BIAS: Toward Unsupervised Automatic Bias Detection for Visual Classifiers" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 109, + 160, + 500, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 109, + 160, + 500, + 190 + ], + "spans": [ + { + "bbox": [ + 109, + 160, + 500, + 190 + ], + "type": "text", + "content": "Quentin Guimard" + }, + { + "bbox": [ + 109, + 160, + 500, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 109, + 160, + 500, + 190 + ], + "type": "text", + "content": " Moreno D'Incà" + }, + { + "bbox": [ + 109, + 160, + 500, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 109, + 160, + 500, + 190 + ], + "type": "text", + "content": " Massimiliano Mancini" + }, + { + "bbox": [ + 109, + 160, + 500, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 109, + 160, + 500, + 190 + ], + "type": "text", + "content": " Elisa Ricci" + }, + { + "bbox": [ + 109, + 160, + 500, + 190 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 109, + 160, + 500, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 109, + 160, + 500, + 190 + ], + "type": "text", + "content": "University of Trento" + }, + { + "bbox": [ + 109, + 160, + 500, + 190 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 109, + 160, + 500, + 190 + ], + "type": "text", + "content": " Fondazione Bruno Kessler" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 221, + 191, + 386, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 221, + 191, + 386, + 202 + ], + "spans": [ + { + "bbox": [ + 221, + 191, + 386, + 202 + ], + "type": "text", + "content": "https://github.com/mardgui/C2B" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "spans": [ + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 258, + 296, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 258, + 296, + 545 + ], + "spans": [ + { + "bbox": [ + 55, + 258, + 296, + 545 + ], + "type": "text", + "content": "A person downloading a pre-trained model from the web should be aware of its biases. Existing approaches for bias identification rely on datasets containing labels for the task of interest, something that a non-expert may not have access to, or may not have the necessary resources to collect: this greatly limits the number of tasks where model biases can be identified. In this work, we present CLASSIFIER-TO-BIAS (C2B), the first bias discovery framework that works without access to any labeled data: it only relies on a textual description of the classification task to identify biases in the target classification model. This description is fed to a large language model to generate bias proposals and corresponding captions depicting biases together with task-specific target labels. A retrieval model collects images for those captions, which are then used to assess the accuracy of the model w.r.t. the given biases. C2B is training-free, does not require any annotations, has no constraints on the list of biases, and can be applied to any pre-trained model on any classification task. Experiments on two publicly available datasets show that C2B discovers biases beyond those of the original datasets and outperforms a recent state-of-the-art bias detection baseline that relies on task-specific annotations, being a promising first step toward addressing task-agnostic unsupervised bias detection." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 572, + 135, + 585 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 572, + 135, + 585 + ], + "spans": [ + { + "bbox": [ + 55, + 572, + 135, + 585 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "content": "The wide availability of pre-trained machine learning models allows anyone to access them. A user may search in online libraries (e.g., HuggingFace [27]) for models addressing their task of interest, download the most suitable one, and use it for their specific use case. As the user did not develop the model, they may not be aware of potential failure modes [28, 60, 63] or harmful biases [9, 10, 24, 49, 69, 70]. These are usually reported on model cards [47], describing how the model has been developed, in which cases it has been tested, its recommended use, and its potential biases." + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 316, + 228, + 553, + 396 + ], + "blocks": [ + { + "bbox": [ + 316, + 228, + 553, + 396 + ], + "lines": [ + { + "bbox": [ + 316, + 228, + 553, + 396 + ], + "spans": [ + { + "bbox": [ + 316, + 228, + 553, + 396 + ], + "type": "image", + "image_path": "1ad9d05a0618ab645e66313f37c59c01ce03526b801871aae1a393c66b8ba488.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 403, + 555, + 460 + ], + "lines": [ + { + "bbox": [ + 313, + 403, + 555, + 460 + ], + "spans": [ + { + "bbox": [ + 313, + 403, + 555, + 460 + ], + "type": "text", + "content": "Figure 1. We explore the novel task of bias discovery when we are only given a specific classification task and a pre-trained model. We propose CLASSIFIER-TO-BIAS (C2B), which automatically detects potential biases in the model, categorizes these biases, and assigns scores to each category." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 483, + 555, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 483, + 555, + 555 + ], + "spans": [ + { + "bbox": [ + 313, + 483, + 555, + 555 + ], + "type": "text", + "content": "While model cards are a useful tool to describe a model and its shortcomings, (i) a lot of models still come without them; (ii) even if they are present, they may not describe specific cases or biases of interest to the users. This means that a user would need to assess the weaknesses of the downloaded model independently." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 556, + 556, + 699 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 556, + 556, + 699 + ], + "spans": [ + { + "bbox": [ + 313, + 556, + 556, + 699 + ], + "type": "text", + "content": "Various tools have already been proposed to measure and mitigate biases in datasets [8, 64] and models [2, 8, 11]. While these tools rely on bias labels, recent work has made progress toward open-set failure mode extraction [16, 19, 22, 30, 53] and bias detection [6, 17, 33, 34, 68] without bias labels, even removing the need for a predefined list of biases. However, in the context of discriminative models such as visual classifiers, all existing methods rely on task-specific labeled data to be able to identify failure modes and measure the biases, limiting the number of tasks where model biases can be identified, and preventing the automation of bias detection." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 326, + 701, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 701, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 326, + 701, + 553, + 713 + ], + "type": "text", + "content": "To address this problem, we present CLASSIFIER-TO" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15151" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 287 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 287 + ], + "type": "text", + "content": "BIAS (C2B), the first framework that, given a pre-trained visual classifier and its task specification, automatically identifies its biases (Fig. 1) in an unsupervised manner, removing the need for task-specific labeled data. We build C2B on two components: a large language model (LLM) and a text-to-image retrieval engine. First, the task specifications are fed to the LLM to produce a series of candidate bias attributes for the task (e.g., gender, ethnicity), and their corresponding bias classes (e.g., male, female). The list of candidate bias classes is fed again to the LLM to generate a set of captions describing both the biases and the task-specific target classes they are associated to. These captions are used by the retrieval module to get images from a large database: the images will constitute our dataset, where, for each image, both the target and bias labels are given by the constructed caption itself. By testing the pre-trained model on this dataset, we can compute the standard bias metrics, verifying which bias class the model is most biased against." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 288, + 294, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 288, + 294, + 443 + ], + "spans": [ + { + "bbox": [ + 55, + 288, + 294, + 443 + ], + "type": "text", + "content": "Importantly, C2B is fully unsupervised and does not suffer from the need of costly annotations, thus it can be applied to automatically detect biases for virtually any classification task. Moreover, the list of biases is class-specific, as different classification tasks and categories (e.g., birds, deers) may exhibit different biases (e.g., tail color, horns size). Finally, the model is modular and can readily incorporate more powerful LLMs, retrieval approaches, or databases as they become available. We test our framework on two datasets, i.e., CelebA [41], and ImageNet-X [28]. Results indicate that C2B can uncover biases beyond those annotated in the original datasets, outperforming a recent supervised open-set bias detector [33]." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 444, + 295, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 444, + 295, + 598 + ], + "spans": [ + { + "bbox": [ + 55, + 444, + 295, + 598 + ], + "type": "text", + "content": "Contributions. To summarize, our contribution is threefold: (i) We propose the new task of unsupervised bias detection in a pre-trained model given only a textual description of the classification task without requiring any annotated dataset or manual intervention, and define a protocol to evaluate methods on this setting; (ii) We present C2B, the first framework for this task that uses an LLM to propose a list of biases and generate captions, and multimodal retrieval to produce a dataset from which biases can be estimated; (iii) We show that C2B can discover biases beyond those annotated in the original datasets, being the first step toward addressing this challenging problem and democratizing bias discovery." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 609, + 142, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 609, + 142, + 621 + ], + "spans": [ + { + "bbox": [ + 55, + 609, + 142, + 621 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 629, + 294, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 294, + 713 + ], + "type": "text", + "content": "Bias Detection and Mitigation. Bias detection and mitigation are essential to ensure fairness and robustness in machine learning models. Several approaches tackle hidden subgroup imbalances without relying on explicit group annotations [35, 37, 59, 61], e.g., by analyzing the mistakes on annotated task data and re-weighting misclassified samples [37]. Other methods focus on open-set or annotation-" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 553, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 287 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 287 + ], + "type": "text", + "content": "free bias discovery [34, 36, 68]. UDIS [34] uses hierarchical clustering to surface systematic biases without predefined subgroup definitions. DebiAN [36] alternates training between a bias discoverer and a classifier to identify and mitigate multiple biases. DIM [68] decomposes feature spaces using Partial Least Squares to uncover multiple unknown biased subgroups and describes them using vision-language models. More recently, language-driven methods have also gained prominence, using captions and keyword extraction to detect and name potential biases learned by models [14, 33, 71]. B2T [33] and Zhao et al. [71] relies on captions derived from annotated datasets or misclassified samples with known labels, validating these biases using, e.g. vision-language similarity. SaMyNa [14] emphasizes interpretability by extracting semantic bias descriptors from large-scale captioning, supporting bias analysis and debiasing. However, all these methods require access to annotated, task-specific datasets." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 288, + 553, + 359 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 288, + 553, + 359 + ], + "spans": [ + { + "bbox": [ + 313, + 288, + 553, + 359 + ], + "type": "text", + "content": "In contrast, our approach, C2B, operates without any kind of annotated data. It relies solely on a natural language description of the classification task and a pre-trained classifier. Using LLM-generated, class-specific bias proposals combined with retrieval-based evaluation, C2B enables fully unsupervised, task-agnostic bias discovery." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 360, + 553, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 360, + 553, + 443 + ], + "spans": [ + { + "bbox": [ + 313, + 360, + 553, + 443 + ], + "type": "text", + "content": "Other related methods focus on bias detection or fairness in generative models [17, 32, 57]. For instance, Open-Bias [17] uses LLMs and VQA models to detect and quantify open-set biases in text-to-image generative models, while GELDA [32] reveals visual biases in image generators by generating domain-specific attributes with LLMs and annotating them using selected VLMs." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 445, + 553, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 445, + 553, + 515 + ], + "spans": [ + { + "bbox": [ + 313, + 445, + 553, + 515 + ], + "type": "text", + "content": "Inspired by these works, we choose to leverage an LLM as well as a text-to-image retrieval engine to tackle the new task of unsupervised bias detection, where only a pretrained model and a textual description of the classification task are available, thus eliminating the need for annotated datasets or manual intervention." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 521, + 553, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 553, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 553, + 628 + ], + "type": "text", + "content": "Explainability and Failure Mode Detection. A parallel research direction focuses on systematic failure mode discovery and interpretability [3, 4, 16, 25, 29]. VoG [3] highlights difficult examples by analyzing gradient variance, and Spotlight [16] identifies contiguous underperforming regions in representation space. Saliency-based methods [29] provide local visual explanations, while counterfactual explanations are generated in either natural language [25] or visual form using diffusion models [4]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 630, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 553, + 713 + ], + "type": "text", + "content": "Recent work has scaled these efforts using embeddings and generative models. Wiles et al. [65] leverages large-scale generative models to automatically find and describe failure clusters. Domino [19] uses cross-modal embeddings and mixture models to identify and describe coherent data slices where models underperform. Jain et al. [30] extracts latent directions corresponding to consistent model errors" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15152" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 203 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 203 + ], + "type": "text", + "content": "and generates challenging counterfactual examples for data augmentation. AdaVision [22] enables interactive discovery of failure modes by combining CLIP-based retrieval with GPT-generated prompts. PRIME [53] prioritizes interpretability by starting from human-defined tags and identifying minimal descriptive combinations that explain failures. MAIA [56] automates interpretability workflows using vision-language models for feature analysis and failure mode discovery. Csurka et al. [15] formalizes language-based error explainability and proposes evaluation metrics to assess the quality of generated explanations." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 204, + 294, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 204, + 294, + 300 + ], + "spans": [ + { + "bbox": [ + 55, + 204, + 294, + 300 + ], + "type": "text", + "content": "While these methods analyze model behavior and failure patterns through model outputs or datasets, our work introduces a complementary approach: starting from a high-level task description, C2B uses LLM-based bias proposals and retrieval-based validation to automatically identify, validate, and quantify biases in pre-trained classifiers, enabling large-scale, unsupervised bias discovery across arbitrary classification tasks." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 312, + 111, + 323 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 312, + 111, + 323 + ], + "spans": [ + { + "bbox": [ + 55, + 312, + 111, + 323 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 331, + 296, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 331, + 296, + 415 + ], + "spans": [ + { + "bbox": [ + 55, + 331, + 296, + 415 + ], + "type": "text", + "content": "In this section, we first define bias attributes and bias classes (Sec. 3.1), and formulate the problem of identifying biases of pre-trained models from their task specifications (Sec. 3.2). We then describe C2B, that proposes a list of biases (Sec. 3.3) and collects data for testing them (Sec. 3.4). Finally, we describe how we measure model's biases from the collected data (Sec. 3.5)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 423, + 145, + 435 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 423, + 145, + 435 + ], + "spans": [ + { + "bbox": [ + 55, + 423, + 145, + 435 + ], + "type": "text", + "content": "3.1. Bias definition" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 442, + 295, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 442, + 295, + 598 + ], + "spans": [ + { + "bbox": [ + 55, + 442, + 295, + 598 + ], + "type": "text", + "content": "Formally, we denote the target classification model as " + }, + { + "bbox": [ + 55, + 442, + 295, + 598 + ], + "type": "inline_equation", + "content": "f: \\mathcal{X} \\to \\mathcal{Y}" + }, + { + "bbox": [ + 55, + 442, + 295, + 598 + ], + "type": "text", + "content": ", mapping an image in the space " + }, + { + "bbox": [ + 55, + 442, + 295, + 598 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 55, + 442, + 295, + 598 + ], + "type": "text", + "content": " to labels in the space " + }, + { + "bbox": [ + 55, + 442, + 295, + 598 + ], + "type": "inline_equation", + "content": "\\mathcal{Y}" + }, + { + "bbox": [ + 55, + 442, + 295, + 598 + ], + "type": "text", + "content": ". We define as bias attributes " + }, + { + "bbox": [ + 55, + 442, + 295, + 598 + ], + "type": "inline_equation", + "content": "\\mathcal{B} = \\{\\mathbf{B}_1, \\dots, \\mathbf{B}_N\\}" + }, + { + "bbox": [ + 55, + 442, + 295, + 598 + ], + "type": "text", + "content": " the different categories of bias, i.e., factors that can impact its performance [44]. For each bias attribute, we define a list of bias classes " + }, + { + "bbox": [ + 55, + 442, + 295, + 598 + ], + "type": "inline_equation", + "content": "\\mathbf{B}_i = \\{b_{i,1}, \\dots, b_{i,n_i}\\}" + }, + { + "bbox": [ + 55, + 442, + 295, + 598 + ], + "type": "text", + "content": ", the different values that each bias attribute can take. For a given target class, we define the model to be biased or unfair w.r.t specific bias classes depending on the per-class accuracy difference of the model. Let us denote the per-class accuracy of a model " + }, + { + "bbox": [ + 55, + 442, + 295, + 598 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 55, + 442, + 295, + 598 + ], + "type": "text", + "content": " on a target class " + }, + { + "bbox": [ + 55, + 442, + 295, + 598 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 55, + 442, + 295, + 598 + ], + "type": "text", + "content": " on data that matches a bias class " + }, + { + "bbox": [ + 55, + 442, + 295, + 598 + ], + "type": "inline_equation", + "content": "b_{i,j}" + }, + { + "bbox": [ + 55, + 442, + 295, + 598 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 55, + 442, + 295, + 598 + ], + "type": "inline_equation", + "content": "A_y(f, b_{i,j})" + }, + { + "bbox": [ + 55, + 442, + 295, + 598 + ], + "type": "text", + "content": ". We define the bias score " + }, + { + "bbox": [ + 55, + 442, + 295, + 598 + ], + "type": "inline_equation", + "content": "\\phi_{y,i,j}" + }, + { + "bbox": [ + 55, + 442, + 295, + 598 + ], + "type": "text", + "content": " for each bias class " + }, + { + "bbox": [ + 55, + 442, + 295, + 598 + ], + "type": "inline_equation", + "content": "b_{i,j}" + }, + { + "bbox": [ + 55, + 442, + 295, + 598 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 61, + 607, + 295, + 637 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 607, + 295, + 637 + ], + "spans": [ + { + "bbox": [ + 61, + 607, + 295, + 637 + ], + "type": "interline_equation", + "content": "\\phi_ {y, i, j} = A _ {y} \\left(f, b _ {i, j}\\right) - \\frac {1}{n _ {i} - 1} \\sum_ {b _ {i, k} \\in \\mathbf {B} _ {i}, k \\neq j} A _ {y} \\left(f, b _ {i, k}\\right). \\tag {1}", + "image_path": "4746ee03137065cc2648ec0678c9a1e5caa1937b011787efeac381ba19d99ecb.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 647, + 176, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 647, + 176, + 658 + ], + "spans": [ + { + "bbox": [ + 55, + 647, + 176, + 658 + ], + "type": "text", + "content": "3.2. Problem formulation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "content": "Given a pre-trained model " + }, + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "content": ", our goal is to identify its biases. We assume " + }, + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "content": " to be a generic classification model that we did not develop and that we aim to use (e.g., downloaded from the web). Thus, we also assume to have no access to" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "content": "additional data " + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "content": " had access to (e.g., the dataset set used for training). As we want to avoid costly annotation procedures, we also assume to have no access to any dataset containing labeled images of the classes in " + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 121, + 554, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 121, + 554, + 205 + ], + "spans": [ + { + "bbox": [ + 313, + 121, + 554, + 205 + ], + "type": "text", + "content": "To address this problem, we can rely only on what we know about the model: i.e., the task it has been designed for, the input domain it expects, and the output it predicts. This information can all be expressed using simple, natural language. Specifically, we assume to have a general description of the model's task " + }, + { + "bbox": [ + 313, + 121, + 554, + 205 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 121, + 554, + 205 + ], + "type": "text", + "content": ", the input the model expects to process, and its output (i.e., classes in " + }, + { + "bbox": [ + 313, + 121, + 554, + 205 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 313, + 121, + 554, + 205 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 207, + 555, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 207, + 555, + 327 + ], + "spans": [ + { + "bbox": [ + 313, + 207, + 555, + 327 + ], + "type": "text", + "content": "This task is challenging as (i) detecting biases requires measuring the performance of the model on a labeled dataset, potentially containing annotations for the factors of interest and (ii) we have no pre-defined list of biases to start with. This implies that we have to both generate a list of potential biases and to collect a dataset to test them. In the following, we will describe our framework, C2B, that exploits LLMs for proposing biases, and a combination of LLMs and multimodal retrieval for scoring them. An overview of C2B is illustrated in Fig. 2." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 339, + 459, + 353 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 339, + 459, + 353 + ], + "spans": [ + { + "bbox": [ + 313, + 339, + 459, + 353 + ], + "type": "text", + "content": "3.3. Generating bias proposals" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 359, + 554, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 359, + 554, + 419 + ], + "spans": [ + { + "bbox": [ + 313, + 359, + 554, + 419 + ], + "type": "text", + "content": "The first step of C2B is to generate a list of bias attributes and classes that we can explicitly test for. To generate the list, we cannot rely on attribute annotations, as we have no access to labeled data. Moreover, we cannot use lists available online, as biases depend on the task of interest." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 421, + 554, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 421, + 554, + 540 + ], + "spans": [ + { + "bbox": [ + 313, + 421, + 554, + 540 + ], + "type": "text", + "content": "To address this problem, we exploit the reasoning capabilities of LLMs. Specifically, let us denote as " + }, + { + "bbox": [ + 313, + 421, + 554, + 540 + ], + "type": "inline_equation", + "content": "\\Phi_{\\mathrm{LLM}}" + }, + { + "bbox": [ + 313, + 421, + 554, + 540 + ], + "type": "text", + "content": " a generic LLM, mapping textual inputs to textual outputs, i.e., " + }, + { + "bbox": [ + 313, + 421, + 554, + 540 + ], + "type": "inline_equation", + "content": "\\Phi_{\\mathrm{LLM}}: \\mathcal{T} \\to \\mathcal{T}" + }, + { + "bbox": [ + 313, + 421, + 554, + 540 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 313, + 421, + 554, + 540 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 313, + 421, + 554, + 540 + ], + "type": "text", + "content": " being the space of natural language. To generate the candidate set of biases, we prompt the LLM to produce a list of bias attributes and classes for the given model and task description. Because different classes may have different biases, we query the LLM to give us a list of potential biases " + }, + { + "bbox": [ + 313, + 421, + 554, + 540 + ], + "type": "inline_equation", + "content": "\\mathcal{B}^y" + }, + { + "bbox": [ + 313, + 421, + 554, + 540 + ], + "type": "text", + "content": " for each class " + }, + { + "bbox": [ + 313, + 421, + 554, + 540 + ], + "type": "inline_equation", + "content": "y \\in \\mathcal{Y}" + }, + { + "bbox": [ + 313, + 421, + 554, + 540 + ], + "type": "text", + "content": ". Formally, we have:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 380, + 553, + 553, + 567 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 380, + 553, + 553, + 567 + ], + "spans": [ + { + "bbox": [ + 380, + 553, + 553, + 567 + ], + "type": "interline_equation", + "content": "\\mathcal {B} ^ {y} = \\Phi_ {\\mathrm {L L M}} \\left(\\pi_ {\\text {b i a s}} \\circ T \\circ y\\right) \\tag {2}", + "image_path": "3c7da4b7df9b9c784803a7d03ed6da0793c313aec8f69d46422d98d31591561a.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 579, + 554, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 579, + 554, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 579, + 554, + 628 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 579, + 554, + 628 + ], + "type": "inline_equation", + "content": "\\circ" + }, + { + "bbox": [ + 313, + 579, + 554, + 628 + ], + "type": "text", + "content": " is the string concatenation operation, " + }, + { + "bbox": [ + 313, + 579, + 554, + 628 + ], + "type": "inline_equation", + "content": "\\pi_{\\mathrm{bias}}" + }, + { + "bbox": [ + 313, + 579, + 554, + 628 + ], + "type": "text", + "content": " is the system prompt that instructs the LLM about the task it should perform (Supp. Mat., Appx. I), and " + }, + { + "bbox": [ + 313, + 579, + 554, + 628 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 579, + 554, + 628 + ], + "type": "text", + "content": " is the model's task description, including input/output specification." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 628, + 553, + 679 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 628, + 553, + 679 + ], + "spans": [ + { + "bbox": [ + 313, + 628, + 553, + 679 + ], + "type": "text", + "content": "The output of Eq. (2), " + }, + { + "bbox": [ + 313, + 628, + 553, + 679 + ], + "type": "inline_equation", + "content": "\\mathcal{B}^y" + }, + { + "bbox": [ + 313, + 628, + 553, + 679 + ], + "type": "text", + "content": ", is a list1 of bias attributes " + }, + { + "bbox": [ + 313, + 628, + 553, + 679 + ], + "type": "inline_equation", + "content": "\\mathcal{B}^y = \\{\\mathbf{B}_1^y, \\dots, \\mathbf{B}_{N_y}^y\\}" + }, + { + "bbox": [ + 313, + 628, + 553, + 679 + ], + "type": "text", + "content": ", of size " + }, + { + "bbox": [ + 313, + 628, + 553, + 679 + ], + "type": "inline_equation", + "content": "N_y" + }, + { + "bbox": [ + 313, + 628, + 553, + 679 + ], + "type": "text", + "content": ". We prompt the LLM so that each bias attribute is associated to a list of " + }, + { + "bbox": [ + 313, + 628, + 553, + 679 + ], + "type": "inline_equation", + "content": "n_{y,i}" + }, + { + "bbox": [ + 313, + 628, + 553, + 679 + ], + "type": "text", + "content": " bias classes " + }, + { + "bbox": [ + 313, + 628, + 553, + 679 + ], + "type": "inline_equation", + "content": "\\mathbf{B}_i^y = \\{b_{i,1}^y, \\dots, b_{i,n_{y,i}}^y\\}" + }, + { + "bbox": [ + 313, + 628, + 553, + 679 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 313, + 693, + 553, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 693, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 693, + 553, + 713 + ], + "type": "text", + "content": "1For simplicity, we assume that the natural language output of the LLM is already formatted as a list." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15153" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 55, + 68, + 558, + 274 + ], + "blocks": [ + { + "bbox": [ + 55, + 68, + 558, + 274 + ], + "lines": [ + { + "bbox": [ + 55, + 68, + 558, + 274 + ], + "spans": [ + { + "bbox": [ + 55, + 68, + 558, + 274 + ], + "type": "image", + "image_path": "5f6973cc196021fad07be63facd83e17b2783b016e6560a09857c8222cbcf2c5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 280, + 555, + 326 + ], + "lines": [ + { + "bbox": [ + 54, + 280, + 555, + 326 + ], + "spans": [ + { + "bbox": [ + 54, + 280, + 555, + 326 + ], + "type": "text", + "content": "Figure 2. Overview of C2B. Given a specific task with an associated description and a pre-trained classification model, our approach leverages an LLM to identify potential candidate bias attributes and corresponding classes. These candidate biases are then used to prompt the LLM to generate captions, which are subsequently utilized by a retrieval module to collect a dataset of images for bias testing. This dataset enables evaluation of the target classification model, producing scores for each identified bias category." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 345, + 149, + 357 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 345, + 149, + 357 + ], + "spans": [ + { + "bbox": [ + 55, + 345, + 149, + 357 + ], + "type": "text", + "content": "3.4. Collecting data" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 366, + 296, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 366, + 296, + 486 + ], + "spans": [ + { + "bbox": [ + 55, + 366, + 296, + 486 + ], + "type": "text", + "content": "Given the list of candidate bias attributes and classes, we still need a dataset to score them. Ideally, for each target class " + }, + { + "bbox": [ + 55, + 366, + 296, + 486 + ], + "type": "inline_equation", + "content": "y \\in \\mathcal{V}" + }, + { + "bbox": [ + 55, + 366, + 296, + 486 + ], + "type": "text", + "content": " and bias class " + }, + { + "bbox": [ + 55, + 366, + 296, + 486 + ], + "type": "inline_equation", + "content": "b^{y} \\in \\mathbf{B}^{y}" + }, + { + "bbox": [ + 55, + 366, + 296, + 486 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 55, + 366, + 296, + 486 + ], + "type": "inline_equation", + "content": "\\mathbf{B}^{y} \\in \\mathcal{B}^{y}" + }, + { + "bbox": [ + 55, + 366, + 296, + 486 + ], + "type": "text", + "content": ", we would like to have a dataset " + }, + { + "bbox": [ + 55, + 366, + 296, + 486 + ], + "type": "inline_equation", + "content": "D(b^{y}, y)" + }, + { + "bbox": [ + 55, + 366, + 296, + 486 + ], + "type": "text", + "content": " depicting images of class " + }, + { + "bbox": [ + 55, + 366, + 296, + 486 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 55, + 366, + 296, + 486 + ], + "type": "text", + "content": " exhibiting the bias class " + }, + { + "bbox": [ + 55, + 366, + 296, + 486 + ], + "type": "inline_equation", + "content": "b^{y}" + }, + { + "bbox": [ + 55, + 366, + 296, + 486 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 55, + 366, + 296, + 486 + ], + "type": "inline_equation", + "content": "D(b^{y}, y) = \\{(x_{i}, y_{i}, b_{i}^{y})\\}_{i=1}^{M}" + }, + { + "bbox": [ + 55, + 366, + 296, + 486 + ], + "type": "text", + "content": " of size " + }, + { + "bbox": [ + 55, + 366, + 296, + 486 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 366, + 296, + 486 + ], + "type": "text", + "content": ". Unfortunately, we cannot rely on labeled data for this task as (i) we do not have access to it; (ii) biases change w.r.t. the particular application, thus it is very hard to collect beforehand a dataset containing all possible target and bias classes of interest." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 490, + 296, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 490, + 296, + 586 + ], + "spans": [ + { + "bbox": [ + 55, + 490, + 296, + 586 + ], + "type": "text", + "content": "To address this problem, we resort to external sources. In particular, we assume to have access to a large-scale database " + }, + { + "bbox": [ + 55, + 490, + 296, + 586 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 55, + 490, + 296, + 586 + ], + "type": "text", + "content": " containing images. Note that " + }, + { + "bbox": [ + 55, + 490, + 296, + 586 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 55, + 490, + 296, + 586 + ], + "type": "text", + "content": " has no annotations: we just assume it to be large-scale (i.e., potentially web-scale). From this database, we can collect images for each target-bias combination of interest using a text-to-image retrieval model. This is achieved in two steps: captions generation and cross-modal retrieval." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "text", + "content": "Caption generation. A straightforward solution to create captions is to use a fixed template (e.g., an image of [target class] with [bias class]). However, this would not consider the specificity of the considered task and thus may lead to retrieving erroneous images. For instance, the biases and captions associated to a facial attribute classification task should be different from those in a natural scene classification task. To avoid this issue, we resort to using the LLM, querying it in two steps: first to create a template for the specific task, and then to gener" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 346, + 555, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 346, + 555, + 369 + ], + "spans": [ + { + "bbox": [ + 313, + 346, + 555, + 369 + ], + "type": "text", + "content": "ate captions specific for target and bias class. Formally, we obtain the initial template as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 375, + 379, + 555, + 392 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 375, + 379, + 555, + 392 + ], + "spans": [ + { + "bbox": [ + 375, + 379, + 555, + 392 + ], + "type": "interline_equation", + "content": "\\operatorname {t e m p} = \\Phi_ {\\mathrm {L L M}} \\left(\\pi_ {\\text {t e m p l a t e}} \\circ T\\right) \\tag {3}", + "image_path": "e54db3478a8bbea265a62f9edae19af847ddb97c7aa359c8653675e5a9905829.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 401, + 554, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 401, + 554, + 449 + ], + "spans": [ + { + "bbox": [ + 313, + 401, + 554, + 449 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 401, + 554, + 449 + ], + "type": "inline_equation", + "content": "\\pi_{\\mathrm{template}}" + }, + { + "bbox": [ + 313, + 401, + 554, + 449 + ], + "type": "text", + "content": " is the system prompt that instructs the LLM about the task it should perform (Supp. Mat., Appx. 1). Then, we ask the LLM to refine this template for each specific target class " + }, + { + "bbox": [ + 313, + 401, + 554, + 449 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 313, + 401, + 554, + 449 + ], + "type": "text", + "content": " and bias class " + }, + { + "bbox": [ + 313, + 401, + 554, + 449 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 313, + 401, + 554, + 449 + ], + "type": "text", + "content": ", as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 359, + 457, + 553, + 472 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 457, + 553, + 472 + ], + "spans": [ + { + "bbox": [ + 359, + 457, + 553, + 472 + ], + "type": "interline_equation", + "content": "c _ {b} ^ {y} = \\Phi_ {\\mathrm {L L M}} \\left(\\pi_ {\\text {c a p t i o n}} \\circ \\operatorname {t e m p} \\circ b \\circ y\\right). \\tag {4}", + "image_path": "a546b83e5d62cc7dc647285a769c01f0348505c77eae4ad85cbf60e98b7b1ff8.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 479, + 554, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 479, + 554, + 528 + ], + "spans": [ + { + "bbox": [ + 313, + 479, + 554, + 528 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 479, + 554, + 528 + ], + "type": "inline_equation", + "content": "c_b^y" + }, + { + "bbox": [ + 313, + 479, + 554, + 528 + ], + "type": "text", + "content": " is the resulting caption. Note that, while with this strategy we obtain one caption per target-bias class pair, in principle we can also generate multiple captions per pair to have more variability, beneficial for more complex tasks." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 533, + 555, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 555, + 652 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 555, + 652 + ], + "type": "text", + "content": "Image retrieval. The obtained captions can be used to retrieve images from a large scale image database, and we implement the retrieval by using a contrastive vision-language model (VLM) that can score the similarity between a text and image. Formally, we can decompose the VLM in three elements: a visual encoder " + }, + { + "bbox": [ + 313, + 533, + 555, + 652 + ], + "type": "inline_equation", + "content": "\\varphi_{\\mathrm{vis}}:\\mathcal{X}\\to \\mathbb{R}^d" + }, + { + "bbox": [ + 313, + 533, + 555, + 652 + ], + "type": "text", + "content": " a text encoder " + }, + { + "bbox": [ + 313, + 533, + 555, + 652 + ], + "type": "inline_equation", + "content": "\\varphi_{\\mathrm{txt}}:\\mathcal{T}\\rightarrow \\mathbb{R}^d" + }, + { + "bbox": [ + 313, + 533, + 555, + 652 + ], + "type": "text", + "content": " , and a similarity function " + }, + { + "bbox": [ + 313, + 533, + 555, + 652 + ], + "type": "inline_equation", + "content": "\\sin :\\mathbb{R}^d\\times \\mathbb{R}^d\\to \\mathbb{R}" + }, + { + "bbox": [ + 313, + 533, + 555, + 652 + ], + "type": "text", + "content": " (e.g., cosine similarity for CLIP [51]). Given a caption, we can retrieve a set of images " + }, + { + "bbox": [ + 313, + 533, + 555, + 652 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_b^y" + }, + { + "bbox": [ + 313, + 533, + 555, + 652 + ], + "type": "text", + "content": " for each class label " + }, + { + "bbox": [ + 313, + 533, + 555, + 652 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 313, + 533, + 555, + 652 + ], + "type": "text", + "content": " and bias " + }, + { + "bbox": [ + 313, + 533, + 555, + 652 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 313, + 533, + 555, + 652 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 355, + 661, + 553, + 681 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 355, + 661, + 553, + 681 + ], + "spans": [ + { + "bbox": [ + 355, + 661, + 553, + 681 + ], + "type": "interline_equation", + "content": "\\mathbf {X} _ {b} ^ {y} = \\underset {x \\in \\mathcal {D}} {\\operatorname {t o p - k}} \\operatorname {s i m} \\left(\\varphi_ {\\text {v i s}} (x), \\varphi_ {\\text {t x t}} \\left(c _ {b} ^ {y}\\right)\\right), \\tag {5}", + "image_path": "0b1e79c28a11ecfe9f5a6360f8827cca7f35914fdadcde82727f90b707127aae.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "content": "where we retrieve the " + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "content": " images with the highest matching score. While these images do not have a ground-truth," + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15154" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "content": "they can still help to estimate the biases of the model, with the fidelity of the latter depending on the retrieval accuracy (Sec. 4.4). We also experiment with a different type of retrieval system, a Web search engine (details in Sec. 4.1)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 129, + 160, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 129, + 160, + 140 + ], + "spans": [ + { + "bbox": [ + 55, + 129, + 160, + 140 + ], + "type": "text", + "content": "3.5. Measuring biases" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 147, + 295, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 147, + 295, + 301 + ], + "spans": [ + { + "bbox": [ + 55, + 147, + 295, + 301 + ], + "type": "text", + "content": "In the previous section, we have collected a dataset with pseudo-annotations for all bias values and classes of interest. We can now test the target model on this dataset to compute its performance, using the variation across bias classes to score each bias. We use the formula defined in Eq. (1) to score each target-bias pair relative to other target-bias pairs. For each bias class, we obtain a bias score between -1 and 1. If the per-class accuracy is higher for a given bias class, the classifier is biased toward the bias class (positive bias score). Reciprocally, if the per-class accuracy is lower for a given bias class, the classifier is biased against the bias class (negative bias score). Intuitively, larger absolute bias score values also correspond to stronger biases." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 303, + 295, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 303, + 295, + 422 + ], + "spans": [ + { + "bbox": [ + 55, + 303, + 295, + 422 + ], + "type": "text", + "content": "To quantitatively evaluate the quality of the biases detected by C2B, we propose two novel evaluation settings that complement each other. First, in Sec. 4.2.1, we propose to match the detected biases with ground-truth annotations using SBERT embeddings for tasks where datasets with additional bias annotations exist. Second, in Sec. 4.2.2, we propose to assign bias labels (based on the initial proposed biases) to ground-truth task-specific data with a VQA model, to check the agreement between the bias detection method and an expert VQA model." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 423, + 295, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 423, + 295, + 506 + ], + "spans": [ + { + "bbox": [ + 55, + 423, + 295, + 506 + ], + "type": "text", + "content": "Collecting scores for all bias attributes and classes enables various types of analyses to explore the biases. Biases can be ranked by score, to assess which attributes are more critical than others. Bias class scores can also be aggregated across bias attributes, to measure which bias attributes exhibit a larger bias. We provide examples of qualitative analyses in Sec. 4.3, and more in the Supp. Mat., Appx. D." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 518, + 135, + 532 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 518, + 135, + 532 + ], + "spans": [ + { + "bbox": [ + 55, + 518, + 135, + 532 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 538, + 295, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 538, + 295, + 634 + ], + "spans": [ + { + "bbox": [ + 55, + 538, + 295, + 634 + ], + "type": "text", + "content": "In this section, we present the results of our experiments, where we apply C2B to detect biases for different models on two different image classification tasks. We first describe our experimental protocol (Sec. 4.1), we then quantitatively compare the biases detected by C2B with a state-of-the-art competitor (Sec. 4.2), and we provide qualitative examples of detected biases (Sec. 4.3). Finally, we analyze the core component of our approach (Sec. 4.4)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 641, + 183, + 654 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 183, + 654 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 183, + 654 + ], + "type": "text", + "content": "4.1. Experimental protocol" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 665, + 295, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 295, + 689 + ], + "type": "text", + "content": "Tasks and models. We test C2B on two different tasks: facial attribute classification and image classification." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": "Facial attribute classification. We consider the CelebA dataset [41], which is composed of 200k images of celebrity" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "content": "faces annotated with 40 facial attributes, such as male, blond hair, young, or high cheekbones. To evaluate C2B, we choose to detect the biases of a recent state-of-the-art model, FaceXFormer [48]. This model relies on a pretrained Swin/B [42] backbone, coupled with an MLP-based fusion module and an attention-based decoder, which are co-trained on several face datasets." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 157, + 553, + 253 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 157, + 553, + 253 + ], + "spans": [ + { + "bbox": [ + 313, + 157, + 553, + 253 + ], + "type": "text", + "content": "Image classification. We focus on the widely used ImageNet-1K [54] dataset, where each image is associated to one of the 1000 ImageNet classes. We choose to detect the biases of four widely used models, easily accessible through TorchVision [43], comprising both convolutional networks and transformers: ResNet50_V2 [23], ResNet101_V2 [23], ResNet152_V2 [23], ViT_B_16_SWAG_E2E_V1 [58]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 258, + 555, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 258, + 555, + 462 + ], + "spans": [ + { + "bbox": [ + 313, + 258, + 555, + 462 + ], + "type": "text", + "content": "Competitors. As we introduce the new problem of unsupervised bias detection relying solely on a textual description of the classification task, there are no existing baseline methods designed for this specific setting. Nevertheless, we still compare C2B to a recent supervised open-set bias detection method, Bias-to-Text (B2T) [33]. However, unlike C2B which does not require task-specific labeled data to identify biases in an image classifier, B2T depends on ground-truth task labels to identify model failures. It then employs a pre-trained captioning model to generate descriptions of misclassified images, extracts common keywords, and computes a bias score based on the difference between (i) the similarity (CLIP score) between the discovered bias keyword and misclassified images, and (ii) the similarity between the discovered bias keyword and well-classified images. Thus, we test B2T using the labeled validation sets of the two datasets (i.e., CelebA and ImageNet)." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 468, + 554, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 468, + 554, + 647 + ], + "spans": [ + { + "bbox": [ + 313, + 468, + 554, + 647 + ], + "type": "text", + "content": "Implementation details. C2B comprises two components: the LLM that proposes biases and captions, and the retrieval module. For the former, we use Llama 3.1 8B [18] (see Supp. Mat., Appx. J for qualitative LLM comparisons). For the retrieval, we explore two strategies: (i) CC12M: we consider the large Conceptual Captions 12M (CC12M) dataset [12] and retrieve images based on CLIP embeddings [7], and (ii) BING: instead of using a VLM to retrieve images from an existing database where some image domains may not be well-represented, we propose to retrieve images from the Web using a publicly available search engine, Bing. For both datasets, we retrieve 20 images per caption, and we measure biases as in Eq. (1), comparing the difference in the model's per-class accuracy between images that match different candidate bias classes." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 658, + 447, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 658, + 447, + 671 + ], + "spans": [ + { + "bbox": [ + 313, + 658, + 447, + 671 + ], + "type": "text", + "content": "4.2. Quantitative evaluation" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 677, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 553, + 714 + ], + "type": "text", + "content": "In this section, we provide a quantitative evaluation to assess the quality of biases identified by C2B. First, we present results based on ground-truth bias annotations, i.e.," + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15155" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "text", + "content": "exploiting attributes annotated in the validation sets of the datasets. For biases not in the ground-truth annotations, we introduce an evaluation based on Visual Question Answering (VQA) to measure the relevance of the proposed biases." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 138, + 268, + 150 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 138, + 268, + 150 + ], + "spans": [ + { + "bbox": [ + 55, + 138, + 268, + 150 + ], + "type": "text", + "content": "4.2.1 Evaluation with ground-truth annotations" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 163, + 295, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 163, + 295, + 270 + ], + "spans": [ + { + "bbox": [ + 54, + 163, + 295, + 270 + ], + "type": "text", + "content": "Setting. In this setting, we exploit the available attribute annotations, checking if models can detect and quantify those biases correctly. In the case of CelebA, when we focus on a specific attribute classification task (e.g., blonde hair) we consider all other 39 attributes (e.g., young) as biases measurable with the ground-truth. For ImageNet-1K, we use ImageNet-X [28], i.e., the validation set of ImageNet-1K with 16 labeled \"factors\" distinguishing images from typical prototypes (e.g., orientation), as ground-truth biases." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 271, + 296, + 379 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 271, + 296, + 379 + ], + "spans": [ + { + "bbox": [ + 55, + 271, + 296, + 379 + ], + "type": "text", + "content": "To obtain ground-truth biases, we test the model on each validation set and check the difference in per-class accuracy among images where the bias attribute is present and images where it is absent. We consider a ground-truth bias to be present if this difference is higher than a threshold (i.e., " + }, + { + "bbox": [ + 55, + 271, + 296, + 379 + ], + "type": "inline_equation", + "content": "\\tau = 0.05" + }, + { + "bbox": [ + 55, + 271, + 296, + 379 + ], + "type": "text", + "content": ") using the same threshold to assess bias identification on both C2B and B2T scores. We refer to the sign of the bias score as the direction of the bias. We provide more results with a varying threshold in the Supp. Mat., Appx. C." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 380, + 296, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 380, + 296, + 475 + ], + "spans": [ + { + "bbox": [ + 55, + 380, + 296, + 475 + ], + "type": "text", + "content": "Due to the open-set nature of the detected biases for both C2B and B2T, there are almost no exact matches between them and ground-truth biases. To compare the two sets, we compute the cosine similarity between SBERT embeddings [52] of the ground-truth and detected biases. We empirically set the similarity threshold to 0.9 to avoid having too many false positives. We provide more details and qualitative examples of matches in the Supp. Mat., Appx. K." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 480, + 296, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 480, + 296, + 624 + ], + "spans": [ + { + "bbox": [ + 55, + 480, + 296, + 624 + ], + "type": "text", + "content": "Metrics. We report results from two complementary perspectives: GT " + }, + { + "bbox": [ + 55, + 480, + 296, + 624 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 55, + 480, + 296, + 624 + ], + "type": "text", + "content": " Detected and Detected " + }, + { + "bbox": [ + 55, + 480, + 296, + 624 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 55, + 480, + 296, + 624 + ], + "type": "text", + "content": " GT. In GT " + }, + { + "bbox": [ + 55, + 480, + 296, + 624 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 55, + 480, + 296, + 624 + ], + "type": "text", + "content": " Detected, we start from each ground-truth bias and check whether it was correctly detected with the same bias direction (HIT), detected with the opposite direction (false hit, FH), or not detected (M1ss). This perspective focuses on recall. However, methods that propose many biases may artificially reduce misses. Therefore, we also report Detected " + }, + { + "bbox": [ + 55, + 480, + 296, + 624 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 55, + 480, + 296, + 624 + ], + "type": "text", + "content": " GT, where we start from each detected bias and verify whether it matches a ground-truth bias with the correct direction. This second view measures precision and helps assess the quality of the proposed biases." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "content": "Results. In Tabs. 1 and 2, we report the results of our analysis. On CelebA (left side of Tab. 1, GT " + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "content": " Detected), C2B outperforms B2T in detecting ground-truth biases, regardless of the retrieval source. We obtain the best results with Bing, detecting between " + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "inline_equation", + "content": "10.8\\%" + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "inline_equation", + "content": "12.3\\%" + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "content": " of the ground-truth biases, compared to only " + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "inline_equation", + "content": "4.5\\%" + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "content": " for B2T. However, B2T exhibits a lower proportion of false hits " + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "inline_equation", + "content": "(3.1\\%)" + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "content": ", whereas" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 318, + 113, + 552, + 174 + ], + "blocks": [ + { + "bbox": [ + 313, + 71, + 555, + 104 + ], + "lines": [ + { + "bbox": [ + 313, + 71, + 555, + 104 + ], + "spans": [ + { + "bbox": [ + 313, + 71, + 555, + 104 + ], + "type": "text", + "content": "Table 1. Proportion " + }, + { + "bbox": [ + 313, + 71, + 555, + 104 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 313, + 71, + 555, + 104 + ], + "type": "text", + "content": " of ground-truth biases detected on CelebA " + }, + { + "bbox": [ + 313, + 71, + 555, + 104 + ], + "type": "inline_equation", + "content": "\\mathrm{(GT\\to Detected)}" + }, + { + "bbox": [ + 313, + 71, + 555, + 104 + ], + "type": "text", + "content": " and of detected biases corresponding to groundtruth ones (Detected " + }, + { + "bbox": [ + 313, + 71, + 555, + 104 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 313, + 71, + 555, + 104 + ], + "type": "text", + "content": " GT). FH=False Hit." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 318, + 113, + 552, + 174 + ], + "lines": [ + { + "bbox": [ + 318, + 113, + 552, + 174 + ], + "spans": [ + { + "bbox": [ + 318, + 113, + 552, + 174 + ], + "type": "table", + "html": "
METHODGT → DETECTEDDETECTED → GT
HIT(↑)FH(↓)MISS (↓)HIT(↑)FH(↓)MISS(↓)
FaceXFormer
B2T [33]4.533.1492.326.594.1089.29
C2B (BING)12.296.8880.8314.187.3478.48
C2B (CC12M)10.767.7581.4912.758.1579.09
", + "image_path": "9fa8d96b5edb8bcc4fe12537b00ffb3b2ccff3a4141fb73ed243cd60ce7d1d8c.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 318, + 227, + 553, + 388 + ], + "blocks": [ + { + "bbox": [ + 313, + 184, + 555, + 217 + ], + "lines": [ + { + "bbox": [ + 313, + 184, + 555, + 217 + ], + "spans": [ + { + "bbox": [ + 313, + 184, + 555, + 217 + ], + "type": "text", + "content": "Table 2. Proportion (\\%) of ground-truth biases detected on ImageNet-X (GT " + }, + { + "bbox": [ + 313, + 184, + 555, + 217 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 313, + 184, + 555, + 217 + ], + "type": "text", + "content": " Detected) and of detected biases corresponding to ground-truth ones (Detected " + }, + { + "bbox": [ + 313, + 184, + 555, + 217 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 313, + 184, + 555, + 217 + ], + "type": "text", + "content": " GT). FH=False Hit." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 318, + 227, + 553, + 388 + ], + "lines": [ + { + "bbox": [ + 318, + 227, + 553, + 388 + ], + "spans": [ + { + "bbox": [ + 318, + 227, + 553, + 388 + ], + "type": "table", + "html": "
METHODGT → DETECTEDDETECTED → GT
HIT(↑)FH(↓)Miss (↓)HIT(↑)FH (↓)Miss(↓)
ResNet50_V2
B2T [33]2.402.1995.410.800.7198.49
C2B (BING)7.808.5383.662.602.9894.42
C2B (CC12M)11.1811.3077.522.992.8794.14
ResNet101_V2
B2T [33]2.552.1195.330.840.6998.47
C2B (BING)7.918.4183.682.562.7794.67
C2B (CC12M)11.2111.6877.112.762.9694.28
ResNet152_V2
B2T [33]2.531.7295.750.830.5898.59
C2B (BING)7.537.9384.542.752.7994.46
C2B (CC12M)11.1012.0176.902.812.9594.24
ViT_B_16_SWAG
B2T [33]2.111.9895.900.720.6798.61
C2B (BING)7.727.1885.102.482.4895.05
C2B (CC12M)10.8511.0778.082.642.7094.66
", + "image_path": "7b5e0cb2fb6bbc9550aa18decfdcaa8c789b486c7c57652a33ba0dee9cbc1858.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 408, + 555, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 408, + 555, + 516 + ], + "spans": [ + { + "bbox": [ + 313, + 408, + 555, + 516 + ], + "type": "text", + "content": "C2B shows slightly higher false hit rates (6.9–7.8%). This is largely due to retrieval noise and the broader search space introduced by LLM-generated proposals, which can sometimes lead to less precise bias candidates. Nevertheless, C2B misses fewer ground-truth biases (80.8–81.5% vs. B2T's 92.3%). These trends remain consistent when considering the proportion of detected biases that correspond to ground-truth biases (Detected → GT), where C2B achieves 12.8–14.2%, compared to 6.6% for B2T." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 517, + 556, + 660 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 517, + 556, + 660 + ], + "spans": [ + { + "bbox": [ + 313, + 517, + 556, + 660 + ], + "type": "text", + "content": "On ImageNet-X (Tab. 1), across four classifiers, C2B again detects more ground-truth biases than B2T (7.5–11.2% vs. 2.1–2.6%), at the cost of higher false hit rates (7.2–12.0% vs. 1.7–2.2%). As before, these false hits are attributable to noisy retrieval and the open-ended nature of bias proposals. Still, C2B misses significantly fewer ground-truth biases (76.9–85.1% vs. 95.3–95.9%). Interestingly, on ImageNet-X, retrieving from CC12M provides better results than Bing. The Detected → GT evaluation confirms these findings, with a larger fraction of C2B-detected biases aligning with ground-truth ones (2.5–3.0% vs. 0.6–0.7% for B2T)." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "content": "Discussion. Due to the limited human annotations available in CelebA and ImageNet-X, the proportion of matching biases is low across all methods, particularly on ImageNet-X. Furthermore, since we are using text embeddings to" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "15156" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 79, + 91, + 272, + 118 + ], + "blocks": [ + { + "bbox": [ + 70, + 71, + 281, + 82 + ], + "lines": [ + { + "bbox": [ + 70, + 71, + 281, + 82 + ], + "spans": [ + { + "bbox": [ + 70, + 71, + 281, + 82 + ], + "type": "text", + "content": "Table 3. Detected biases and VQA agreement on CelebA." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 79, + 91, + 272, + 118 + ], + "lines": [ + { + "bbox": [ + 79, + 91, + 272, + 118 + ], + "spans": [ + { + "bbox": [ + 79, + 91, + 272, + 118 + ], + "type": "table", + "html": "
MODELB2T [33]C2B (BING)C2B (CC12M)
FaceXFormer0.100.220.25
", + "image_path": "940d460151245fe75fad876c287a1a9bf795a2e1a396f8848581e8fe281fc8a3.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 78, + 149, + 273, + 201 + ], + "blocks": [ + { + "bbox": [ + 61, + 129, + 288, + 140 + ], + "lines": [ + { + "bbox": [ + 61, + 129, + 288, + 140 + ], + "spans": [ + { + "bbox": [ + 61, + 129, + 288, + 140 + ], + "type": "text", + "content": "Table 4. Detected biases and VQA agreement on ImageNet-X." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 78, + 149, + 273, + 201 + ], + "lines": [ + { + "bbox": [ + 78, + 149, + 273, + 201 + ], + "spans": [ + { + "bbox": [ + 78, + 149, + 273, + 201 + ], + "type": "table", + "html": "
MODELB2T [33]C2B (BING)C2B (CC12M)
ResNet50_V20.200.260.20
ResNet101_V20.190.280.20
ResNet152_V20.200.280.20
ViT_B_16_SWAG0.200.320.21
", + "image_path": "662c77f6608c654a85818b79a540e8045e9e2fee69393b800ae973222e84a073.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 223, + 296, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 223, + 296, + 319 + ], + "spans": [ + { + "bbox": [ + 55, + 223, + 296, + 319 + ], + "type": "text", + "content": "match the biases, the numbers are also influenced by the semantic proximity between LLM-proposed biases, caption-extracted keywords, and ground-truth annotations. Moreover, a high number of Miss in the Detected " + }, + { + "bbox": [ + 55, + 223, + 296, + 319 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 55, + 223, + 296, + 319 + ], + "type": "text", + "content": " GT case is not necessarily a bad thing, as C2B and B2T may be detecting accurate biases that are not present in the annotation. For these reasons, in the next section we introduce a different evaluation based on visual question answering." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 335, + 185, + 346 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 335, + 185, + 346 + ], + "spans": [ + { + "bbox": [ + 55, + 335, + 185, + 346 + ], + "type": "text", + "content": "4.2.2 VQA-based evaluation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 355, + 296, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 355, + 296, + 403 + ], + "spans": [ + { + "bbox": [ + 55, + 355, + 296, + 403 + ], + "type": "text", + "content": "As ground-truth bias annotations may not always be available, we also propose an alternative evaluation scheme based on VQA, measuring the quality of the proposed biases that are not present in ground-truth annotations." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 408, + 296, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 408, + 296, + 564 + ], + "spans": [ + { + "bbox": [ + 55, + 408, + 296, + 564 + ], + "type": "text", + "content": "Metric. We use a VQA model (LLaVA-1.5-13B [38, 39]) to pseudo-label each image of the validation sets with each of the proposed bias by C2B or B2T. Additional details can be found in the Supp. Mat., Appx. G. From these pseudolabels, we can use the ground-truth task labels (i.e., class annotations) to check how the per-class accuracy varies over different bias classes, using the same threshold " + }, + { + "bbox": [ + 55, + 408, + 296, + 564 + ], + "type": "inline_equation", + "content": "\\tau = 0.05" + }, + { + "bbox": [ + 55, + 408, + 296, + 564 + ], + "type": "text", + "content": " of Sec. 4.2.1. For each bias class " + }, + { + "bbox": [ + 55, + 408, + 296, + 564 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 55, + 408, + 296, + 564 + ], + "type": "text", + "content": " proposed by the model, there are three cases: a positive bias is detected " + }, + { + "bbox": [ + 55, + 408, + 296, + 564 + ], + "type": "inline_equation", + "content": "(\\phi_b \\geq \\tau)" + }, + { + "bbox": [ + 55, + 408, + 296, + 564 + ], + "type": "text", + "content": ", a negative bias is detected " + }, + { + "bbox": [ + 55, + 408, + 296, + 564 + ], + "type": "inline_equation", + "content": "(\\phi_b \\leq -\\tau)" + }, + { + "bbox": [ + 55, + 408, + 296, + 564 + ], + "type": "text", + "content": ", or the bias is not detected " + }, + { + "bbox": [ + 55, + 408, + 296, + 564 + ], + "type": "inline_equation", + "content": "(|\\phi_b| < \\tau)" + }, + { + "bbox": [ + 55, + 408, + 296, + 564 + ], + "type": "text", + "content": ". For each bias class, we define the agreement score to be 1 if both scores fall in the same case, -1 if they detect opposite biases, and 0 otherwise." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "content": "Results. Tabs. 3 and 4 present the results of our VQA-based evaluation on CelebA and ImageNet-X. On CelebA, C2B achieves higher agreement with VQA pseudo-labels (0.22 for Bing and 0.25 for CC12M) than B2T (0.10), indicating that the biases discovered by C2B are more aligned with visually detectable and semantically coherent bias attributes. On ImageNet-X, while C2B often achieves higher agreement scores than B2T with Bing (i.e., +0.09 on average), results with CC12M are either the same or comparable to those of B2T (i.e., +0.005 on average). We attribute this to the challenge of relying on retrieval-based data collection, without access to annotated failure cases or" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 554, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 156 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 156 + ], + "type": "text", + "content": "task-specific captions. In contrast, B2T benefits from captions derived from known misclassifications. Despite this more challenging and fully unsupervised setting, C2B still achieves agreements that are competitive with or superior to B2T. This confirms that C2B's proposals are not only meaningful and visually grounded but also relevant for bias detection in settings where no annotations are available." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 162, + 554, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 162, + 554, + 281 + ], + "spans": [ + { + "bbox": [ + 313, + 162, + 554, + 281 + ], + "type": "text", + "content": "Discussion. The results presented in this section are influenced by both the types of questions posed to the VQA model and the model's own inherent biases. Given these limitations, evaluations based on ground-truth annotations and those derived from VQA outputs serve as complementary approaches. Overall, biases proposed by the LLM tend to offer more relevant bias indicators than keywords extracted from captions as in B2T [33], allowing C2B to detect a greater proportion of ground-truth biases and achieve higher agreement with the VQA model than B2T." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 292, + 440, + 304 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 292, + 440, + 304 + ], + "spans": [ + { + "bbox": [ + 313, + 292, + 440, + 304 + ], + "type": "text", + "content": "4.3. Qualitative evaluation" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 311, + 553, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 311, + 553, + 334 + ], + "spans": [ + { + "bbox": [ + 313, + 311, + 553, + 334 + ], + "type": "text", + "content": "In this section, we show examples of biases discovered by C2B for different models on different tasks." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 335, + 554, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 335, + 554, + 419 + ], + "spans": [ + { + "bbox": [ + 313, + 335, + 554, + 419 + ], + "type": "text", + "content": "In Fig. 3, we show biases that C2B detects for the high cheekbones attribute on face attribute classification, with bias scores from different retrieval methods. We can see that a strong bias was detected on the expression attribute: the classifier appears to perform better when people are smiling and worse when they appear angry. Other smaller biases were detected, such as young, and male." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 421, + 554, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 421, + 554, + 576 + ], + "spans": [ + { + "bbox": [ + 313, + 421, + 554, + 576 + ], + "type": "text", + "content": "Fig. 4 shows a similar analysis for biases detected by C2B for the minibus class on image classification for the ResNet50_V2 model. Strong biases were detected on the lighting, orientation, and presence of reflection bias attribute: the classifier appears stronger on daytime images, portrait-oriented images, and images without reflections, while performing worse on nighttime images, landscape-oriented images, and images with reflections. Finally, Fig. 5 depicts biases for the birdhouse class again for image classification, but for the ViT_B_16_SWAG model. A strong bias is found on the shape bias attribute: the classifier seems to be performing better on images of rectangular birdhouses, while it underperforms on images of spherical birdhouses." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 586, + 501, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 586, + 501, + 597 + ], + "spans": [ + { + "bbox": [ + 313, + 586, + 501, + 597 + ], + "type": "text", + "content": "4.4. Evaluation of VLM-based retrieval" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "type": "text", + "content": "Retrieving images matching the target classes and biases present in the captions is essential for the detected biases to be trustworthy. Therefore, in this section, we analyze the CLIP-based retrieval approach used by C2B, by trying to retrieve images for which we have annotations. For a VQA-based analysis of both the Bing-based retrieval and CLIP-based retrieval on CC12M, please refer to Supp. Mat., Appx. F. To perform our analysis, we use CLIP to retrieve images from the validation sets of CelebA and ImageNet-" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15157" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 72, + 553, + 202 + ], + "blocks": [ + { + "bbox": [ + 60, + 72, + 553, + 202 + ], + "lines": [ + { + "bbox": [ + 60, + 72, + 553, + 202 + ], + "spans": [ + { + "bbox": [ + 60, + 72, + 553, + 202 + ], + "type": "image", + "image_path": "517b6ddb58ba735e614704926cceb229630dc2224ccf7d35b1a7646f0150073c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 209, + 479, + 221 + ], + "lines": [ + { + "bbox": [ + 130, + 209, + 479, + 221 + ], + "spans": [ + { + "bbox": [ + 130, + 209, + 479, + 221 + ], + "type": "text", + "content": "Figure 3. Examples of biases found by C2B for the high cheekbones target attribute on CelebA." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 59, + 243, + 294, + 357 + ], + "blocks": [ + { + "bbox": [ + 59, + 243, + 294, + 357 + ], + "lines": [ + { + "bbox": [ + 59, + 243, + 294, + 357 + ], + "spans": [ + { + "bbox": [ + 59, + 243, + 294, + 357 + ], + "type": "image", + "image_path": "1005d94e80770f02c61266a1f5476eed61590f5c79e0f5812e1d8c7849ac584b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 368, + 295, + 391 + ], + "lines": [ + { + "bbox": [ + 55, + 368, + 295, + 391 + ], + "spans": [ + { + "bbox": [ + 55, + 368, + 295, + 391 + ], + "type": "text", + "content": "Figure 4. Examples of biases found by C2B for the minibus class on ImageNet-X (ResNet50_V2)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 58, + 407, + 294, + 524 + ], + "blocks": [ + { + "bbox": [ + 58, + 407, + 294, + 524 + ], + "lines": [ + { + "bbox": [ + 58, + 407, + 294, + 524 + ], + "spans": [ + { + "bbox": [ + 58, + 407, + 294, + 524 + ], + "type": "image", + "image_path": "40b395fe3a813bd492853383acccf65cabe9055b7085a593f2b01416b297e14a.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 534, + 295, + 556 + ], + "lines": [ + { + "bbox": [ + 55, + 534, + 295, + 556 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 295, + 556 + ], + "type": "text", + "content": "Figure 5. Examples of biases found by C2B for the birdhouse class on ImageNet-X (ViT_B_16_SWAG)." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 581, + 295, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 581, + 295, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 581, + 295, + 605 + ], + "type": "text", + "content": "X, using handcrafted captions as input: combinations of attributes on CelebA, and classes and factors on ImageNet-X." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 611, + 295, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 611, + 295, + 658 + ], + "spans": [ + { + "bbox": [ + 55, + 611, + 295, + 658 + ], + "type": "text", + "content": "Metric. We use recall@ " + }, + { + "bbox": [ + 55, + 611, + 295, + 658 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 611, + 295, + 658 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 611, + 295, + 658 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 611, + 295, + 658 + ], + "type": "text", + "content": " varies for each caption to match the number of corresponding ground-truth samples in the dataset. We show results using recall @ " + }, + { + "bbox": [ + 55, + 611, + 295, + 658 + ], + "type": "inline_equation", + "content": "A \\cdot K" + }, + { + "bbox": [ + 55, + 611, + 295, + 658 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 611, + 295, + 658 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 55, + 611, + 295, + 658 + ], + "type": "text", + "content": " is a coefficient that varies between 0.01 and 1.0." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "Results. We present the results in Fig. 6. For CelebA (left), the recall is the same for both target and bias attributes, as expected, since both are drawn from the same set of attributes. On the same dataset, CLIP struggles to" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 316, + 241, + 438, + 352 + ], + "blocks": [ + { + "bbox": [ + 316, + 241, + 438, + 352 + ], + "lines": [ + { + "bbox": [ + 316, + 241, + 438, + 352 + ], + "spans": [ + { + "bbox": [ + 316, + 241, + 438, + 352 + ], + "type": "image", + "image_path": "69ace24a8a203f0e7f6e92391a298ddf676ffe6c72f6e9d9589910b3f69c440c.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 353, + 361, + 515, + 372 + ], + "lines": [ + { + "bbox": [ + 353, + 361, + 515, + 372 + ], + "spans": [ + { + "bbox": [ + 353, + 361, + 515, + 372 + ], + "type": "text", + "content": "Figure 6. Accuracy of VLM-based retrieval." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 439, + 241, + 552, + 351 + ], + "blocks": [ + { + "bbox": [ + 439, + 241, + 552, + 351 + ], + "lines": [ + { + "bbox": [ + 439, + 241, + 552, + 351 + ], + "spans": [ + { + "bbox": [ + 439, + 241, + 552, + 351 + ], + "type": "image", + "image_path": "5e0928000fd6dc910aaa55266bbb7847b560712dce1ead30e13f9b66bad03d5f.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 398, + 555, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 398, + 555, + 506 + ], + "spans": [ + { + "bbox": [ + 313, + 398, + 555, + 506 + ], + "type": "text", + "content": "retrieve images with combinations of attributes, with a recall @ " + }, + { + "bbox": [ + 313, + 398, + 555, + 506 + ], + "type": "inline_equation", + "content": "0.01 \\cdot K" + }, + { + "bbox": [ + 313, + 398, + 555, + 506 + ], + "type": "text", + "content": " just above " + }, + { + "bbox": [ + 313, + 398, + 555, + 506 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 313, + 398, + 555, + 506 + ], + "type": "text", + "content": ". For ImageNet-X (right), CLIP is fairly accurate, especially for low values of " + }, + { + "bbox": [ + 313, + 398, + 555, + 506 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 313, + 398, + 555, + 506 + ], + "type": "text", + "content": ", with more than " + }, + { + "bbox": [ + 313, + 398, + 555, + 506 + ], + "type": "inline_equation", + "content": "80\\%" + }, + { + "bbox": [ + 313, + 398, + 555, + 506 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 398, + 555, + 506 + ], + "type": "inline_equation", + "content": "70\\%" + }, + { + "bbox": [ + 313, + 398, + 555, + 506 + ], + "type": "text", + "content": " recall at " + }, + { + "bbox": [ + 313, + 398, + 555, + 506 + ], + "type": "inline_equation", + "content": "0.05 \\cdot K" + }, + { + "bbox": [ + 313, + 398, + 555, + 506 + ], + "type": "text", + "content": " to retrieve images with specific biases and classes, respectively. Generally, the recall is higher when the number of retrieved images decreases relative to the number of potential matching images present in the dataset, which suggests that, the larger the dataset, the more accurate the retrieval is." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 523, + 388, + 536 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 523, + 388, + 536 + ], + "spans": [ + { + "bbox": [ + 313, + 523, + 388, + 536 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "type": "text", + "content": "We introduced CLASSIFIER-TO-BIAS (C2B), a novel method for automatically detecting biases in visual classifiers entirely in an unsupervised manner, without the need for task-specific labeled data. Starting from a simple textual description of the classification task and a pre-trained model, C2B generates bias hypotheses, retrieves relevant images to test them, and produces a list of biases with their scores. Although C2B is the first approach to operate in such challenging setting, our results show that it surpasses a recent supervised open-set bias detector across two evaluation settings: one based on ground-truth annotations and another on VQA-based pseudo-labels. We believe our work marks an important step toward automated, open-set, and task-agnostic bias detection." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15158" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 163 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 163 + ], + "type": "text", + "content": "This work was sponsored by the MUR PNRR project FAIR - Future AI Research (PE00000013), funded by NextGeneration, and the Italian Ministerial grants PRIN 2022: \"B-FAIR: Bias-Free Artificial Intelligence methods for automated visual Recognition\" (CUP E53D23008010006), and the EU project ELIAS (No.01120237)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 172, + 115, + 185 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 172, + 115, + 185 + ], + "spans": [ + { + "bbox": [ + 56, + 172, + 115, + 185 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 193, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 193, + 295, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 193, + 295, + 247 + ], + "spans": [ + { + "bbox": [ + 61, + 193, + 295, + 247 + ], + "type": "text", + "content": "[1] Marah Abdin, Sam Ade Jacobs, Ammar Ahmad Awan, Jyoti Aneja, Ahmed Awadallah, Hany Awadalla, Nguyen Bach, Amit Bahree, Arash Bakhtiari, Harkirat Behl, et al. Phi-3 technical report: A highly capable language model locally on your phone. arXiv preprint arXiv:2404.14219, 2024. 15" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 248, + 295, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 248, + 295, + 281 + ], + "spans": [ + { + "bbox": [ + 62, + 248, + 295, + 281 + ], + "type": "text", + "content": "[2] Julius A Adebayo et al. *FairML: ToolBox for diagnosing bias in predictive modeling*. PhD thesis, Massachusetts Institute of Technology, 2016. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 282, + 295, + 313 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 282, + 295, + 313 + ], + "spans": [ + { + "bbox": [ + 62, + 282, + 295, + 313 + ], + "type": "text", + "content": "[3] Chirag Agarwal, Daniel D'souza, and Sara Hooker. Estimating example difficulty using variance of gradients. In CVPR, 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 316, + 294, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 316, + 294, + 346 + ], + "spans": [ + { + "bbox": [ + 62, + 316, + 294, + 346 + ], + "type": "text", + "content": "[4] Maximilian Augustin, Valentyn Boreiko, Francesco Croce, and Matthias Hein. Diffusion visual counterfactual explanations. NeurIPS, 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 349, + 294, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 349, + 294, + 380 + ], + "spans": [ + { + "bbox": [ + 62, + 349, + 294, + 380 + ], + "type": "text", + "content": "[5] Razvan Azamfirei, Sapna R. Kudchadkar, and James Fackler. Large language models and the perils of their hallucinations. Critical Care, 27(1):120, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 382, + 294, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 382, + 294, + 413 + ], + "spans": [ + { + "bbox": [ + 62, + 382, + 294, + 413 + ], + "type": "text", + "content": "[6] Yujia Bao and Regina Barzilay. Learning to split for automatic bias detection. arXiv preprint arXiv:2204.13749, 2022. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 415, + 294, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 415, + 294, + 458 + ], + "spans": [ + { + "bbox": [ + 62, + 415, + 294, + 458 + ], + "type": "text", + "content": "[7] Romain Beaumont. Clip retrieval: Easily compute clip embeddings and build a clip retrieval system with them. https://github.com/rom1504/clip-retrieval, 2022.5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 460, + 294, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 460, + 294, + 535 + ], + "spans": [ + { + "bbox": [ + 62, + 460, + 294, + 535 + ], + "type": "text", + "content": "[8] R. K. E. Bellamy, K. Dey, M. Hind, S. C. Hoffman, S. Houde, K. Kannan, P. Lohia, J. Martino, S. Mehta, A. Mojsilovic, S. Nagar, K. Natesan Ramamurthy, J. Richards, D. Saha, P. Sattigeri, M. Singh, K. R. Varshney, and Y. Zhang. Ai fairness 360: An extensible toolkit for detecting and mitigating algorithmic bias. IBM Journal of Research and Development, 63(4/5):4:1-4:15, 2019. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 536, + 294, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 536, + 294, + 581 + ], + "spans": [ + { + "bbox": [ + 62, + 536, + 294, + 581 + ], + "type": "text", + "content": "[9] Tolga Bolukbasi, Kai-Wei Chang, James Zou, Venkatesh Saligrama, and Adam Kalai. Man is to computer programmer as woman is to homemaker? debiasing word embeddings. In NeurIPS, 2016. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 582, + 294, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 582, + 294, + 624 + ], + "spans": [ + { + "bbox": [ + 57, + 582, + 294, + 624 + ], + "type": "text", + "content": "[10] Joy Buolamwini and Timnit Gebru. Gender shades: Intersectional accuracy disparities in commercial gender classification. In Proceedings of the 1st Conference on Fairness, Accountability and Transparency, 2018. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 625, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 625, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 57, + 625, + 294, + 689 + ], + "type": "text", + "content": "[11] Angel Alexander Cabrera, Will Epperson, Fred Hohman, Minsuk Kahng, Jamie Morgenstern, and Duen Horng Chau. Fairvis: Visual analytics for discovering intersectional bias in machine learning. In 2019 IEEE Conference on Visual Analytics Science and Technology (VAST), pages 46-56, 2019." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 692, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 692, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 692, + 294, + 713 + ], + "type": "text", + "content": "[12] Soravit Changpinyo, Piyush Sharma, Nan Ding, and Radu Soricut. Conceptual 12M: Pushing web-scale image-text" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 74, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 335, + 74, + 553, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 74, + 553, + 94 + ], + "spans": [ + { + "bbox": [ + 335, + 74, + 553, + 94 + ], + "type": "text", + "content": "pre-training to recognize long-tail visual concepts. In CVPR, 2021.5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 96, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 96, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 317, + 96, + 553, + 150 + ], + "type": "text", + "content": "[13] Davide Chicco, Niklas Totsch, and Giuseppe Jurman. The matthews correlation coefficient (mcc) is more reliable than balanced accuracy, bookmaker informedness, and markedness in two-class confusion matrix evaluation. BioData Mining, 14(1):13, 2021. 11" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 152, + 553, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 152, + 553, + 204 + ], + "spans": [ + { + "bbox": [ + 317, + 152, + 553, + 204 + ], + "type": "text", + "content": "[14] Massimiliano Ciranni, Luca Molinaro, Carlo Alberto Barbano, Attilio Fiandrotti, Vittorio Murino, Vito Paolo Pastore, and Enzo Tartaglione. Say my name: a model's bias discovery framework. arXiv preprint arXiv:2408.09570, 2024. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 206, + 553, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 206, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 316, + 206, + 553, + 239 + ], + "type": "text", + "content": "[15] Gabriela Csurka, Tyler L Hayes, Diane Larlus, and Riccardo Volpi. What could go wrong? discovering and describing failure modes in computer vision. In ECCV-WS, 2024. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 240, + 553, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 240, + 553, + 294 + ], + "spans": [ + { + "bbox": [ + 317, + 240, + 553, + 294 + ], + "type": "text", + "content": "[16] Greg d'Eon, Jason d'Eon, James R Wright, and Kevin Leyton-Brown. The spotlight: A general method for discovering systematic errors in deep learning models. In Proceedings of the 2022 ACM Conference on Fairness, Accountability, and Transparency, pages 1962-1981, 2022. 1, 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 295, + 553, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 295, + 553, + 338 + ], + "spans": [ + { + "bbox": [ + 317, + 295, + 553, + 338 + ], + "type": "text", + "content": "[17] Moreno D'Inca, Elia Peruzzo, Massimiliano Mancini, Dejia Xu, Vedit Goel, Xingqian Xu, Zhangyang Wang, Humphrey Shi, and Nicu Sebe. Openbias: Open-set bias detection in text-to-image generative models. In CVPR, 2024. 1, 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 339, + 553, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 339, + 553, + 392 + ], + "spans": [ + { + "bbox": [ + 317, + 339, + 553, + 392 + ], + "type": "text", + "content": "[18] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 5, 15" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 394, + 553, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 394, + 553, + 448 + ], + "spans": [ + { + "bbox": [ + 317, + 394, + 553, + 448 + ], + "type": "text", + "content": "[19] Sabri Eyuboglu, Maya Varma, Khaled Kamal Saab, Jean-Benoit Delbrouck, Christopher Lee-Messer, Jared Dunnmon, James Zou, and Christopher Re. Domino: Discovering systematic errors with cross-modal embeddings. In ICLR, 2022. 1, 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 449, + 553, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 449, + 553, + 590 + ], + "spans": [ + { + "bbox": [ + 317, + 449, + 553, + 590 + ], + "type": "text", + "content": "[20] Samir Yitzhak Gadre, Gabriel Ilharco, Alex Fang, Jonathan Hayase, Georgios Smyrnis, Thao Nguyen, Ryan Marten, Mitchell Wortsman, Dhruba Ghosh, Jieyu Zhang, Eyal Orgad, Rahim Entezari, Giannis Daras, Sarah Pratt, Vivek Ramanujan, Yonatan Bitton, Kalyani Marathe, Stephen Mussmann, Richard Vencu, Mehdi Cherti, Ranjay Krishna, Pang Wei W Koh, Olga Saukh, Alexander J Ratner, Shuran Song, Hannaneh Hajishirzi, Ali Farhadi, Romain Beaumont, Sewoong Oh, Alex Dimakis, Jenia Jitsev, Yair Carmon, Vaishaal Shankar, and Ludwig Schmidt. Datacomp: In search of the next generation of multimodal datasets. In Advances in Neural Information Processing Systems, pages 27092-27112, 2023. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 592, + 553, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 592, + 553, + 646 + ], + "spans": [ + { + "bbox": [ + 317, + 592, + 553, + 646 + ], + "type": "text", + "content": "[21] Isabel O. Gallegos, Ryan A. Rossi, Joe Barrow, Md Mehrab Tanjim, Sungchul Kim, Franck Dernoncourt, Tong Yu, Ruiyi Zhang, and Nesreen K. Ahmed. Bias and fairness in large language models: A survey. Computational Linguistics, 50 (3):1097-1179, 2024. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 647, + 553, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 647, + 553, + 680 + ], + "spans": [ + { + "bbox": [ + 317, + 647, + 553, + 680 + ], + "type": "text", + "content": "[22] Irena Gao, Gabriel Ilharco, Scott Lundberg, and Marco Tulio Ribeiro. Adaptive testing of computer vision models. In ICCV, 2023. 1, 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 681, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 553, + 712 + ], + "type": "text", + "content": "[23] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 5" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15159" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 712 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "type": "text", + "content": "[24] Lisa Anne Hendricks, Kaylee Burns, Kate Saenko, Trevor Darrell, and Anna Rohrbach. Women also snowboard: Overcoming bias in captioning models. In ECCV, 2018. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 294, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 294, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 294, + 140 + ], + "type": "text", + "content": "[25] Lisa Anne Hendricks, Ronghang Hu, Trevor Darrell, and Zeynep Akata. Generating counterfactual explanations with natural language. In ICML-WS, 2018. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 141, + 294, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 141, + 294, + 183 + ], + "spans": [ + { + "bbox": [ + 56, + 141, + 294, + 183 + ], + "type": "text", + "content": "[26] Cheng-Yu Hsieh, Jieyu Zhang, Zixian Ma, Aniruddha Kembhavi, and Ranjay Krishna. Sugarcrepe: Fixing hackable benchmarks for vision-language compositionality. NeurIPS, 2024. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 186, + 170, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 186, + 170, + 196 + ], + "spans": [ + { + "bbox": [ + 56, + 186, + 170, + 196 + ], + "type": "text", + "content": "[27] https://huggingface.co/. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 198, + 294, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 198, + 294, + 251 + ], + "spans": [ + { + "bbox": [ + 56, + 198, + 294, + 251 + ], + "type": "text", + "content": "[28] Badr Youbi Idrissi, Diane Bouchacourt, Randall Balestriero, Ivan Evtimov, Caner Hazirbas, Nicolas Ballas, Pascal Vincent, Michal Drozdjal, David Lopez-Paz, and Mark Ibrahim. Imagenet-x: Understanding model mistakes with factor of variation annotations. In ICLR, 2023. 1, 2, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 254, + 294, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 254, + 294, + 285 + ], + "spans": [ + { + "bbox": [ + 56, + 254, + 294, + 285 + ], + "type": "text", + "content": "[29] Saachi Jain, Hadi Salman, Eric Wong, Pengchuan Zhang, Vibhav Vineet, Sai Vermprala, and Aleksander Madry. Missingness bias in model debugging. In ICLR, 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 287, + 294, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 287, + 294, + 319 + ], + "spans": [ + { + "bbox": [ + 56, + 287, + 294, + 319 + ], + "type": "text", + "content": "[30] Saachi Jain, Hannah Lawrence, Ankur Moitra, and Aleksander Madry. Distilling model failures as directions in latent space. In ICLR, 2023. 1, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 321, + 294, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 321, + 294, + 364 + ], + "spans": [ + { + "bbox": [ + 56, + 321, + 294, + 364 + ], + "type": "text", + "content": "[31] Ziwei Ji, Nayeon Lee, Rita Frieske, Tiezheng Yu, Dan Su, Yan Xu, Etsuko Ishii, Ye Jin Bang, Andrea Madotto, and Pascale Fung. Survey of hallucination in natural language generation. ACM Comput. Surv., 55(12), 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 365, + 294, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 365, + 294, + 408 + ], + "spans": [ + { + "bbox": [ + 56, + 365, + 294, + 408 + ], + "type": "text", + "content": "[32] Krish Kabra, Kathleen M. Lewis, and Guha Balakrishnan. Gelda: A generative language annotation framework to reveal visual biases in image generators. In CVPRW, 2024. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 411, + 294, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 411, + 294, + 453 + ], + "spans": [ + { + "bbox": [ + 56, + 411, + 294, + 453 + ], + "type": "text", + "content": "[33] Younghyun Kim, Sangwoo Mo, Minkyu Kim, Kyungmin Lee, Jaeho Lee, and Jinwoo Shin. Discovering and mitigating visual biases through keyword explanation. In CVPR, 2024. 1, 2, 5, 6, 7, 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 455, + 294, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 455, + 294, + 487 + ], + "spans": [ + { + "bbox": [ + 56, + 455, + 294, + 487 + ], + "type": "text", + "content": "[34] Arvindkumar Krishnakumar, Viraj Prabhu, Sruthi Sudhakar, and Judy Hoffman. Udis: Unsupervised discovery of bias in deep visual recognition models. In BMVC, 2021. 1, 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 489, + 294, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 489, + 294, + 510 + ], + "spans": [ + { + "bbox": [ + 56, + 489, + 294, + 510 + ], + "type": "text", + "content": "[35] Zhiheng Li and Chenliang Xu. Discover the unknown biased attribute of an image classifier. In ICCV, 2021. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 512, + 294, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 512, + 294, + 543 + ], + "spans": [ + { + "bbox": [ + 56, + 512, + 294, + 543 + ], + "type": "text", + "content": "[36] Zhiheng Li, Anthony Hoogs, and Chenliang Xu. Discover and mitigate unknown biases with debiasing alternate networks. In ECCV, 2022. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 545, + 294, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 545, + 294, + 588 + ], + "spans": [ + { + "bbox": [ + 56, + 545, + 294, + 588 + ], + "type": "text", + "content": "[37] Evan Z Liu, Behzad Haghgoo, Annie S Chen, Aditi Raghunathan, Pang Wei Koh, Shiori Sagawa, Percy Liang, and Chelsea Finn. Just train twice: Improving group robustness without training group information. In ICML, 2021. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 590, + 294, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 590, + 294, + 611 + ], + "spans": [ + { + "bbox": [ + 56, + 590, + 294, + 611 + ], + "type": "text", + "content": "[38] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, 2023. 7, 11" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 56, + 613, + 294, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 613, + 294, + 644 + ], + "spans": [ + { + "bbox": [ + 56, + 613, + 294, + 644 + ], + "type": "text", + "content": "[39] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In CVPR, 2024. 7, 11" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 56, + 647, + 294, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 294, + 679 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 294, + 679 + ], + "type": "text", + "content": "[40] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava next: Improved reasoning,OCR, and world knowledge, 2024. 11" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 56, + 681, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 681, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 681, + 294, + 712 + ], + "type": "text", + "content": "[41] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaou Tang. Deep learning face attributes in the wild. In ICCV, 2015. 2, 5" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "text", + "content": "[42] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, 2021. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 118, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 553, + 150 + ], + "type": "text", + "content": "[43] TorchVision maintainers and contributors. Torchvision: Pytorch's computer vision library. https://github.com/pytorch/vision, 2016.5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 152, + 553, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 553, + 195 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 553, + 195 + ], + "type": "text", + "content": "[44] Ninareh Mehrabi, Fred Morstatter, Nripsuta Saxena, Kristina Lerman, and Aram Galstyan. A survey on bias and fairness in machine learning. ACM computing surveys (CSUR), 54 (6):1-35, 2021. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 198, + 553, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 198, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 316, + 198, + 553, + 251 + ], + "type": "text", + "content": "[45] Thomas Mesnard, Cassidy Hardin, Robert Dadashi, Surya Bhupatiraju, Shreya Pathak, Laurent Sifre, Morgane Rivière, Mihir Sanjay Kale, Juliette Love, et al. Gemma: Open models based on gemini research and technology. arXiv preprint arXiv:2403.08295, 2024. 15" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 254, + 553, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 254, + 553, + 297 + ], + "spans": [ + { + "bbox": [ + 316, + 254, + 553, + 297 + ], + "type": "text", + "content": "[46] Microsoft. Chat markup language chatml (preview), 2024. https://learn.microsoft.com/en-us/azure/ai-services/openai/how-to/chat-markup-language, Last accessed on 2024-11-21. 13" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 299, + 553, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 299, + 553, + 363 + ], + "spans": [ + { + "bbox": [ + 316, + 299, + 553, + 363 + ], + "type": "text", + "content": "[47] Margaret Mitchell, Simone Wu, Andrew Zaldivar, Parker Barnes, Lucy Vasserman, Ben Hutchinson, Elena Spitzer, Inioluwa Deborah Raji, and Timnit Gebru. Model cards for model reporting. In Proceedings of the conference on fairness, accountability, and transparency, pages 220-229, 2019. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 365, + 553, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 365, + 553, + 398 + ], + "spans": [ + { + "bbox": [ + 316, + 365, + 553, + 398 + ], + "type": "text", + "content": "[48] Kartik Narayan, Vibashan VS, Rama Chellappa, and Vishal M Patel. Facexformer: A unified transformer for facial analysis. arXiv preprint arXiv:2403.12960, 2024. 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 399, + 553, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 399, + 553, + 442 + ], + "spans": [ + { + "bbox": [ + 316, + 399, + 553, + 442 + ], + "type": "text", + "content": "[49] Ziad Obermeyer, Brian Powers, Christine Vogeli, and Sendhil Mullainathan. Dissecting racial bias in an algorithm used to manage the health of populations. Science, 366(6464): 447-453, 2019. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 445, + 553, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 445, + 553, + 487 + ], + "spans": [ + { + "bbox": [ + 316, + 445, + 553, + 487 + ], + "type": "text", + "content": "[50] Jesutofunmi A. Omiye, Jenna C. Lester, Simon Spichak, Veronica Rotemberg, and Roxana Daneshjou. Large language models propagate race-based medicine. npj Digital Medicine, 6(1):195, 2023. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 490, + 553, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 490, + 553, + 543 + ], + "spans": [ + { + "bbox": [ + 316, + 490, + 553, + 543 + ], + "type": "text", + "content": "[51] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 4" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 545, + 553, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 545, + 553, + 576 + ], + "spans": [ + { + "bbox": [ + 316, + 545, + 553, + 576 + ], + "type": "text", + "content": "[52] Nils Reimers and Iryna Gurevych. Sentence-bert: Sentence embeddings using siamese bert-networks. In ENMLP, 2019. 6" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 580, + 553, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 580, + 553, + 611 + ], + "spans": [ + { + "bbox": [ + 316, + 580, + 553, + 611 + ], + "type": "text", + "content": "[53] Keivan Rezaei, Mehrdad Saberi, Mazda Moayeri, and Soheil Feizi. PRIME: Prioritizing interpretability in failure mode extraction. In ICLR, 2024. 1, 3" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 613, + 553, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 613, + 553, + 667 + ], + "spans": [ + { + "bbox": [ + 316, + 613, + 553, + 667 + ], + "type": "text", + "content": "[54] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, Alexander C. Berg, and Li Fei-Fei. ImageNet Large Scale Visual Recognition Challenge. IJCV, 115(3):211-252, 2015. 5" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "text", + "content": "[55] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortman, Patrick Schramowski, Srivatsa Kundurthy, Katherine" + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15160" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 76, + 72, + 296, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 296, + 127 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 296, + 127 + ], + "type": "text", + "content": "Crowson, Ludwig Schmidt, Robert Kaczmarczyk, and Jenia Jitsev. Laion-5b: An open large-scale dataset for training next generation image-text models. In Advances in Neural Information Processing Systems, pages 25278-25294, 2022. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 294, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 294, + 172 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 294, + 172 + ], + "type": "text", + "content": "[56] Tamar Rott Shaham, Sarah Schwettmann, Franklin Wang, Achyuta Rajaram, Evan Hernandez, Jacob Andreas, and Antonio Torralba. A multimodal automated interpretability agent. In ICML, 2024. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 173, + 294, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 173, + 294, + 205 + ], + "spans": [ + { + "bbox": [ + 56, + 173, + 294, + 205 + ], + "type": "text", + "content": "[57] Robik Shrestha, Yang Zou, Qiuyu Chen, Zhiheng Li, Yusheng Xie, and Siqi Deng. Fairrag: Fair human generation via fair retrieval augmentation. In CVPR, 2024. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 206, + 294, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 206, + 294, + 259 + ], + "spans": [ + { + "bbox": [ + 56, + 206, + 294, + 259 + ], + "type": "text", + "content": "[58] Mannat Singh, Laura Gustafson, Aaron Adcock, Vinicius de Freitas Reis, Bugra Gedik, Raj Prateek Kosaraju, Dhruv Mahajan, Ross Girshick, Piotr Dólár, and Laurens van der Maaten. Revisiting weakly supervised pre-training of visual perception models. In CVPR, 2022. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 261, + 294, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 261, + 294, + 293 + ], + "spans": [ + { + "bbox": [ + 56, + 261, + 294, + 293 + ], + "type": "text", + "content": "[59] Sahil Singla and Soheil Feizi. Salient imagenet: How to discover spurious features in deep learning? In ICLR, 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 294, + 294, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 294, + 294, + 326 + ], + "spans": [ + { + "bbox": [ + 56, + 294, + 294, + 326 + ], + "type": "text", + "content": "[60] Sahil Singla, Besmira Nushi, Shital Shah, Ece Kamar, and Eric Horvitz. Understanding failures of deep networks via robust feature extraction. In CVPR, 2021. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 327, + 294, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 327, + 294, + 369 + ], + "spans": [ + { + "bbox": [ + 56, + 327, + 294, + 369 + ], + "type": "text", + "content": "[61] Nimit Sohoni, Jared Dunnmon, Geoffrey Angus, Albert Gu, and Christopher Ré. No subclass left behind: Fine-grained robustness in coarse-grained classification problems. NeurIPS, 2020. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 372, + 294, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 372, + 294, + 415 + ], + "spans": [ + { + "bbox": [ + 56, + 372, + 294, + 415 + ], + "type": "text", + "content": "[62] Tristan Thrush, Ryan Jiang, Max Bartolo, Amanpreet Singh, Adina Williams, Douwe Kiela, and Candace Ross. Winoground: Probing vision and language models for visio-linguistic compositionality. In CVPR, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 416, + 294, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 416, + 294, + 459 + ], + "spans": [ + { + "bbox": [ + 56, + 416, + 294, + 459 + ], + "type": "text", + "content": "[63] Vijay Vasudevan, Benjamin Caine, Raphael Gontijo Lopes, Sara Fridovich-Keil, and Rebecca Roelofs. When does dough become a bagel? analyzing the remaining mistakes on imagenet. In NeurIPS, 2022. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 460, + 294, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 460, + 294, + 513 + ], + "spans": [ + { + "bbox": [ + 56, + 460, + 294, + 513 + ], + "type": "text", + "content": "[64] Angelina Wang, Alexander Liu, Ryan Zhang, Anat Kleiman, Leslie Kim, Dora Zhao, Iroha Shirai, Arvind Narayanan, and Olga Russakovsky. Revise: A tool for measuring and mitigating bias in visual datasets. *IJCV*, 130(7):1790–1810, 2022. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 515, + 294, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 515, + 294, + 548 + ], + "spans": [ + { + "bbox": [ + 56, + 515, + 294, + 548 + ], + "type": "text", + "content": "[65] Olivia Wiles, Isabela Albuquerque, and Sven Gowal. Discovering bugs in vision models using off-the-shelf image generation and captioning. In NeurIPS-WS, 2022. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 548, + 294, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 548, + 294, + 592 + ], + "spans": [ + { + "bbox": [ + 56, + 548, + 294, + 592 + ], + "type": "text", + "content": "[66] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 11975-11986, 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 593, + 294, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 593, + 294, + 646 + ], + "spans": [ + { + "bbox": [ + 56, + 593, + 294, + 646 + ], + "type": "text", + "content": "[67] Yue Zhang, Yafu Li, Leyang Cui, Deng Cai, Lemao Liu, Tingchen Fu, Xinting Huang, Enbo Zhao, Yu Zhang, Yu-long Chen, et al. Siren's song in the ai ocean: a survey on hallucination in large language models. arXiv preprint arXiv:2309.01219, 2023. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 647, + 294, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 294, + 680 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 294, + 680 + ], + "type": "text", + "content": "[68] Zeliang Zhang, Mingqian Feng, Zhiheng Li, and Chenliang Xu. Discover and mitigate multiple biased subgroups in image classifiers. In CVPR, 2024. 1, 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 681, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 681, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 681, + 294, + 712 + ], + "type": "text", + "content": "[69] Dora Zhao, Angelina Wang, and Olga Russakovsky. Understanding and evaluating racial biases in image captioning. In ICCV, 2021. 1" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 72, + 553, + 151 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 315, + 72, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 72, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 315, + 72, + 553, + 116 + ], + "type": "text", + "content": "[70] Jieyu Zhao, Tianlu Wang, Mark Yatskar, Vicente Ordonez, and Kai-Wei Chang. Men also like shopping: Reducing gender bias amplification using corpus-level constraints. In EMNLP, 2017. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 118, + 553, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 553, + 151 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 553, + 151 + ], + "type": "text", + "content": "[71] Zaiying Zhao, Soichiro Kumano, and Toshihiko Yamasaki. Language-guided detection and mitigation of unknown dataset bias. arXiv preprint arXiv:2406.02889, 2024. 2" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "15161" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/CleanDIFT_ Diffusion Features without Noise/17f652ff-055a-4bf1-a266-9144370345a9_content_list.json b/2025/CleanDIFT_ Diffusion Features without Noise/17f652ff-055a-4bf1-a266-9144370345a9_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..ac7f7eed9808086481d607320449704aa1fb9663 --- /dev/null +++ b/2025/CleanDIFT_ Diffusion Features without Noise/17f652ff-055a-4bf1-a266-9144370345a9_content_list.json @@ -0,0 +1,1652 @@ +[ + { + "type": "text", + "text": "$\\checkmark$ CleanDIFT: Diffusion Features without Noise", + "text_level": 1, + "bbox": [ + 256, + 130, + 743, + 151 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Nick Stracke*", + "text_level": 1, + "bbox": [ + 102, + 180, + 215, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Stefan Andreas Baumann*", + "text_level": 1, + "bbox": [ + 248, + 180, + 459, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Kolja Bauer*", + "text_level": 1, + "bbox": [ + 493, + 181, + 599, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Frank Fundel", + "text_level": 1, + "bbox": [ + 630, + 181, + 738, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Björn Ommer", + "text_level": 1, + "bbox": [ + 779, + 181, + 893, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "CompVis @ LMU Munich, MCML", + "bbox": [ + 356, + 199, + 640, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{nick.stracke,b.ommer}@lmu.de", + "bbox": [ + 367, + 218, + 624, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "compvis.github.io/cleandift", + "bbox": [ + 364, + 239, + 630, + 255 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Normal Diffusion Features 30 min Finetuning", + "text_level": 1, + "bbox": [ + 251, + 268, + 549, + 285 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "CleanDIFT Features", + "text_level": 1, + "bbox": [ + 563, + 273, + 712, + 287 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "noisy images noisy features", + "bbox": [ + 210, + 291, + 473, + 304 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0e4b63c0ac6e7712c160ff1e5389229e3308f90165be123a628a1cc446623839.jpg", + "image_caption": [ + "$t = 300$" + ], + "image_footnote": [], + "bbox": [ + 214, + 306, + 287, + 357 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/48084df88f96a2606168a20000ccdfddaa3939408eb4a2567f1860df4377d05c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 292, + 306, + 382, + 364 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/70715d719fbd02a6850f60f5dc2a5c015492a320ce7444325f7fe1b36d3362f0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 383, + 306, + 470, + 364 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8668d2c527e376f5ca9b196f7f0446b78ecb983ce5f96f05ee26c5bbfd5d1c61.jpg", + "image_caption": [ + "timestep tuning required $X$" + ], + "image_footnote": [], + "bbox": [ + 215, + 388, + 285, + 450 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1d4b2a24cd398a9dc718f4fc53e2f2b24123651a37619ae1a68d30cb93d8c0ff.jpg", + "image_caption": [ + "Figure 1. Our proposed CleanDIFT feature extraction method yields noise-free, timestep-independent, general-purpose features that significantly outperform standard diffusion features. CleanDIFT operates on clean images, while extracting diffusion features with existing approaches requires adding noise to an image before passing it through the model. Adding noise reduces the information present in the image and requires tuning a timestep per downstream task." + ], + "image_footnote": [], + "bbox": [ + 290, + 391, + 380, + 450 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e13d2c508add19b770a90f41496be17677f7b19da282ff4cac391dd9ed3a657f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 383, + 392, + 470, + 450 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "clean images clean features", + "bbox": [ + 511, + 291, + 777, + 304 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1f19a52e101070d30fa1990b76ad79fdb74c69c081b4dbcee6d73ccb04bff8b2.jpg", + "image_caption": [ + "no timestep tuning required $\\checkmark$" + ], + "image_footnote": [], + "bbox": [ + 509, + 306, + 593, + 364 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/dcd916cdd0eae739e33f27d488bdd8f878eb87ad0e03ca7f0ad8be73d8191639.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 594, + 306, + 687, + 364 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8a8e4a04cac091f870cb498ebd96098ebed58a18e19e928a854153fe7ee50abf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 691, + 306, + 774, + 364 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/168d3f58973fdd6f652a8b910e50df8424c684385c73890753205aa80b492256.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 511, + 391, + 593, + 450 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/32fa88c7b0404cc885204005645bc2fe76b1fabf6b70ae7765c4500e5bb8e2b8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 594, + 392, + 687, + 450 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0f923256c1d2d1ea88e5f63c61e591c1f564b805aec913466b509a6ce0c40d97.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 691, + 392, + 774, + 450 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 537, + 326, + 553 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Internal features from large-scale pre-trained diffusion models have recently been established as powerful semantic descriptors for a wide range of downstream tasks. Works that use these features generally need to add noise to images before passing them through the model to obtain the semantic features, as the models do not offer the most useful features when given images with little to no noise. We show that this noise has a critical impact on the usefulness of these features that cannot be remedied by ensembling with different random noises. We address this issue by introducing a lightweight, unsupervised fine-tuning method that enables diffusion backbones to provide high-quality, noise-free semantic features. We show that these features readily outperform previous diffusion features by a wide margin in a wide variety of extraction setups and downstream tasks, offering better performance than even ensemble-based methods at a fraction of the cost.", + "bbox": [ + 86, + 575, + 485, + 834 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 514, + 537, + 643, + 553 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Learning meaningful visual representations that capture a vast amount of world knowledge remains a key problem in the field of computer vision. Diffusion models can be trained at scale in a self-supervised manner and have rapidly advanced the state of the art in image [12, 41, 51] and video generation [13, 25, 58], making them a good candidate to learn visual representations. Many early works have already achieved impressive results using internal features from large-scale pretrained diffusion models for a wide variety of tasks, such as semantic correspondence detection [62, 66, 67], semantic segmentation [2, 39, 63], panoptic segmentation [64], object detection [8], and classification [31]. However, the optimal approach to extract this world knowledge from a diffusion model remains uncertain.", + "bbox": [ + 509, + 566, + 906, + 777 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To understand why that is the case, we take a look at how diffusion models are trained: a varying amount of noise is added to a clean input image (forward process) and the model is tasked to remove the noise from the image (backward process). The amount of added noise is dependent on the diffusion timestep. As a result, the model learns to operate on noisy images and also becomes dependent on the noise timestep as different noise levels require the model to", + "bbox": [ + 509, + 779, + 908, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal Contribution", + "bbox": [ + 109, + 886, + 220, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "117", + "bbox": [ + 486, + 944, + 509, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "perform different tasks [1, 4]. Since noisy images inherently contain less information than clean images (cf. Figure 2), we hypothesize that this harms the internal feature representation of diffusion models [9] and, thus, the extractable world knowledge. Furthermore, the timestep acts as a hyperparameter that influences the internal feature representation and needs to be picked independently for every downstream application (cf. Figure 14).", + "bbox": [ + 89, + 90, + 480, + 210 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We propose a novel feature extraction method that (1) eliminates the need to destroy information by adding noise to the input; and (2) produces timestep-independent generic diffusion features useful for a wide range of down-stream tasks, alleviating the need to tune a noisig timestep per down-stream task. We show how to adapt an off-the-shelf large-scale pre-trained diffusion backbone to provide these features at minimal cost (approximately 30 minutes of fine-tuning on a single A100 GPU) and demonstrate improved performance across a wide range of downstream tasks.", + "bbox": [ + 89, + 213, + 482, + 364 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We achieve this by viewing a diffusion model as a family of $T$ feature extractors that operate on images with different noise levels and provide features with different characteristics. We consolidate all $T$ feature extraction functions in our feature extractor by aligning their internal representations. Specifically, we initialize our feature extractor as a trainable copy of the diffusion model; fine-tune it with clean images and no timestep input; and align its features with all $T$ time-dependent feature extractors of the diffusion model.", + "bbox": [ + 89, + 366, + 482, + 501 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We evaluate our improved features across a wide variety of downstream tasks, such as semantic correspondence matching, monocular depth estimation, semantic segmentation, and classification, and find that they consistently improve upon approaches based on standard diffusion features. These improvements are most evident for dense visual tasks such as semantic correspondence matching, where our features show substantial performance gains across a wide variety of setups [62, 66, 67] and set a new state-of-the-art for unsupervised semantic correspondence matching. Additionally, our proposed method eliminates the need for noise or timestep ensembling [62], offering substantial speed gains (e.g., $8 \\times$ over DIFT [62]), on top of improved quality. Our method is generic and integrates easily with established methods, such as fusing diffusion and DINOv2 features [66, 67].", + "bbox": [ + 89, + 505, + 482, + 746 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our main contributions are as follows:", + "bbox": [ + 89, + 748, + 346, + 762 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. We propose CleanDIFT, a finetuning approach for diffusion models that enables them to operate on clean images and makes the inherent world knowledge of these models more accessible.", + "2. We show how to consolidate information from all diffusion timesteps into a single feature prediction, removing the need for task-specific timestep tuning.", + "3. We demonstrate significant performance gains of our diffusion feature extraction technique across a wide range" + ], + "bbox": [ + 89, + 763, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/fd330a828668af62cc1322a621550038d9f262fe12bb7f536c044c42730d8b13.jpg", + "image_caption": [ + "(a) Reconstruction without Noise", + "Figure 2. Deterioration of Diffusion Features. As current methods need to pass noisy images to the model to obtain useful features, they significantly reduce the information available. We alleviate this problem by obtaining useful features without noise, improving the performance of downstream tasks." + ], + "image_footnote": [], + "bbox": [ + 519, + 90, + 707, + 237 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/12eea81cc5e551bfb499eda8e08eaaa662e3618dbd69cab2eeafb7d8e11d3116.jpg", + "image_caption": [ + "(b) Diffusion Model Reconstruction with added noise $(t = 261$ [62])" + ], + "image_footnote": [], + "bbox": [ + 712, + 90, + 901, + 237 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "of down-stream tasks, notably surpassing the current state of the art in zero-shot unsupervised semantic correspondence detection. We further demonstrate the generality of our enhanced features by showing that these performance gains transfer to advanced methods that fuse diffusion features or operate in a supervised setting.", + "bbox": [ + 531, + 371, + 903, + 462 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "4. Our proposed approach is significantly more efficient than previous methods that tried to address this problem by noise ensembling or supervised training.", + "bbox": [ + 511, + 462, + 903, + 507 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 511, + 521, + 653, + 537 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Self-Supervised Representation Learning Features from large, pre-trained foundation models have been shown to yield competitive performance to supervised models for a variety of downstream tasks, both in zero-shot and fine-tuning settings [20, 43, 48]. These foundation models are trained on different pre-text tasks like inpainting [20], predicting transformations [19], patch reordering [38, 42], and discriminative tasks [6, 43]. DINOv2 [43] uses a discriminative objective combined with self-distillation to learn general-purpose visual features that have proven useful for a variety of downstream tasks [10]. CLIP [48] learns such features by employing a contrastive objective on text-image pairs. Masked Autoencoders [20] (MAEs) are trained to reconstruct masked out patches of the input, also resulting in general-purpose visual features.", + "bbox": [ + 511, + 547, + 906, + 773 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Diffusion Models as Self-Supervised Learners Diffusion models [24, 60, 61] are generative models that have defined the state-of-the-art in image generation [12, 14, 41, 46, 51, 53], video generation [13, 47], and audio generation [15, 30] in recent years. Their primary purpose is to generate high-quality samples (images, videos, etc.). However, generation can also be interpreted as a pretext task for", + "bbox": [ + 511, + 794, + 906, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "118", + "bbox": [ + 486, + 945, + 511, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "learning expressive features, since the model has to build up comprehensive world knowledge in order to generate plausible samples [9, 17, 27, 32, 62]. Features from diffusion models (typically referred to as diffusion features) are obtained by passing a noised image through the diffusion model and extracting intermediate feature representations. They have been shown to be useful for a variety of tasks such as finding semantic correspondences [21, 35, 62], semantic and panoptic segmentation [2, 39, 63, 64], classification [31], and object detection [8].", + "bbox": [ + 89, + 90, + 480, + 241 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For semantic correspondence matching, features are leveraged to identify semantically matching regions across images. Existing approaches utilize diffusion features either in a zero-shot setting [21, 62] or fine-tune them for the semantic correspondence task [33, 35, 67]. Some zero-shot approaches do not fine-tune on semantic correspondence but still require tuning a prompt to activate attention maps at the query location of the correspondence [21]. In contrast, our approach aims to provide universal features usable for various down-stream tasks in a true zero-shot manner.", + "bbox": [ + 89, + 243, + 480, + 393 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Further, diffusion features have been shown to complement features from other self-supervised learning methods such as DINOv2 [18, 43, 66]. DINOv2 features provide sparse but accurate semantic information, while Diffusion features provide dense spatial information albeit with sometimes inaccurate semantics. State-of-the-art approaches for semantic correspondence detection exploit this and fuse features [66, 67]. Compared to DINOv2 features, diffusion features yield smoother, spatially more coherent correspondences [66].", + "bbox": [ + 89, + 395, + 480, + 546 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Distillation for Diffusion Models Knowledge Distillation [22] is a technique used to distill knowledge from a teacher model into a student model. In the context of diffusion models, this is typically applied either to reduce the required denoising steps [54] or to distill a classifier-free guided [23] model into one without CFG [36]. While our approach is inspired by distillation, we consolidate features from $T$ different models, with $T$ being the number of discrete diffusion timesteps, because the features are different at every timestep.", + "bbox": [ + 89, + 563, + 482, + 715 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 729, + 181, + 746 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminaries", + "text_level": 1, + "bbox": [ + 89, + 756, + 230, + 771 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Diffusion Models Diffusion models are trained to predict a clean image $\\mathbf{x}_0$ given a noisy image $\\mathbf{x}_t$ , either explicitly or implicitly. The noisy image $\\mathbf{x}_t$ is a weighted sum with random Gaussian noise $\\mathbf{x}_t = \\sqrt{\\alpha_t}\\mathbf{x}_0 + \\sqrt{1 - \\alpha_t}\\epsilon$ with timestep-dependent coefficients $\\alpha_{t}$ and noise $\\epsilon \\sim \\mathcal{N}(0,\\mathbf{I})$ . $t\\in [0,T]$ denotes the time step of the diffusion process, with $t = 0$ corresponding to the clean image and $t = T$ corresponding to pure noise.", + "bbox": [ + 89, + 779, + 482, + 900 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/e5341a1d6c39dabfda046b3f20d68749e7446c43b0d651440d9de50170ee264d.jpg", + "image_caption": [ + "Figure 3. Fraction of variance of diffusion features explained by 1) encoding the clean image at $t = 0$ (no additive noise), and 2) encoding just the added noise $\\epsilon$ at $t = 999$ . Even at relatively low timesteps such as $t = 261$ as used by DIFT [62], a substantial part of the features directly depends only on the added noise." + ], + "image_footnote": [], + "bbox": [ + 519, + 88, + 895, + 178 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Intuitively, the model faces different objectives for different noise levels [1, 4]: for very high noise, there is little information in the input, and the model is first generating the coarse structure of the image [50]. At lower noise levels, more high- and medium-frequency information is available and the task shifts to generating finer details and intricate structures. This multi-objective nature intuitively explains why previous methods found diffusion features extracted from different timesteps to provide information with differing semantics.", + "bbox": [ + 511, + 284, + 906, + 435 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Diffusion Feature Extraction Typically, diffusion feature extraction happens after first adding noise to an image and passing the resulting $\\mathbf{x}_t$ to a U-Net [52] denoiser. Features are then extracted at multiple hand-picked locations of the U-Net decoder [2, 39, 62, 66, 67]. Different levels of noise added to the input image result in features beneficial for different downstream applications. Typical diffusion timesteps are $t = 261$ [62], $t = 100$ [66] or $t = 50$ [67]. By adding noise to the input image, these methods bottleneck the perceptual information the model can extract. To illustrate this, we show Stable Diffusion 2.1's reconstruction of an image noised at $t = 261$ [62] in Figure 2.", + "bbox": [ + 511, + 455, + 908, + 638 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Diffusion Features Encode Noise We hypothesize that diffusion features extracted from noisy images $\\mathbf{x}_t$ encode the noise $\\epsilon$ in addition to information from the image. We investigate this hypothesis in a very simple setting by examining how well features $\\mathrm{feat}(\\epsilon; T)$ extracted from only the noise $\\epsilon$ approximate the features $\\mathrm{feat}(\\mathbf{x}_t = \\sqrt{\\alpha_t} \\mathbf{x}_0 + \\sqrt{1 - \\alpha_t} \\epsilon; t)$ . Using least squares, we fit a single scalar approximation coefficient to obtain the optimal reconstruction. We then quantify how much of the variance of the overall features is explained by this approximation (cf. Figure 3). Even at relatively low timesteps, such as $t = 261$ used by DIFT [62], encoding pure noise explains a substantial fraction of the features' variance. Current diffusion feature methods extract this information jointly with the image information. Our proposed method addresses this issue by eliminating the noise from the feature extraction process.", + "bbox": [ + 511, + 657, + 908, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "119", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/8908f26484b3c0741935d98d5d3b70f96a8d408bdff2ef9ca61262240b645678.jpg", + "image_caption": [ + "Figure 4. Our training setup. We train our model to predict features from a clean input image, while the frozen diffusion model is fed the noisy image. The projection heads project our model's features onto the noisy diffusion model features, given the noising timestep $t$ . For downstream tasks, we discard the projection heads and directly use our model's internal representations as features." + ], + "image_footnote": [], + "bbox": [ + 91, + 93, + 486, + 250 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We further analyze the residual and similarly decompose it via the features predicted for the clean image $\\mathrm{feat}(x_t = x_0; t = 0)$ . We find that they do not fully explain the remainder of the features either. Instead, a substantial part of the feature variance at medium noise timesteps is timestep dependent and cannot be attributed to components present at $t = 0$ or $t = T$ . This matches observations by previous works [62] that found diffusion features at higher timesteps offer better semantics, despite the added noise at the image input.", + "bbox": [ + 89, + 361, + 483, + 512 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. CleanDIFT: Noise-Free Diffusion Features", + "text_level": 1, + "bbox": [ + 89, + 523, + 465, + 539 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We present CleanDIFT, our method to address the problem of noisy and time-dependent diffusion features. CleanDIFT extracts clean Diffusion FeaTures from a pretrained diffusion backbone through a lightweight fine-tuning process. An overview of our setup is shown in Figure 4.", + "bbox": [ + 89, + 546, + 483, + 623 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Extraction Setup We train our feature extraction model to match the diffusion model's internal representations. We initialize the feature extraction model as a trainable copy of the diffusion model. Crucially, the feature extraction model is given the clean input image, while the diffusion model receives the noisy image and the corresponding timestep as input. Our goal is to obtain a single, noise-free feature map from the feature extraction model that consolidates the information of the diffusion model's timestep-dependent internal representations into a single one. To align our model's representations with the timestep-dependent diffusion model features during training, we introduce pointwise timestep-conditioned feature projection heads. The feature maps predicted by these projection heads are then aligned to the diffusion model's features. For feature extraction at inference time, we usually discard the projection heads and directly use the feature extraction model's", + "bbox": [ + 89, + 643, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "internal representations. However, the projection heads can also be used to efficiently obtain feature maps for specific timesteps by reusing the feature extraction model's internal representations and passing them through the projection heads for different $t$ values.", + "bbox": [ + 511, + 90, + 905, + 167 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Training Objective We regard the diffusion model as a family of feature extraction functions $\\operatorname{feat}(\\cdot, \\epsilon, t)$ for timestep $t \\in [1,999]$ and noise $\\epsilon \\sim \\mathcal{N}(0, \\mathbf{I})$ . Each of these functions maps an image $\\mathbf{x}$ to a feature vector $\\operatorname{feat}(\\mathbf{x}, \\epsilon, t)$ . We aim to consolidate the information provided by all feature extraction functions into a single joint function $\\operatorname{feat}_c(\\cdot)$ with the same dimensionality:", + "bbox": [ + 511, + 198, + 906, + 304 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{c} \\underline {{\\text {S t a b l e D i f f u s i o n}}} \\\\ \\left. \\begin{array}{l} \\operatorname {f e a t} (\\mathbf {x}, \\boldsymbol {\\epsilon}, t = 1) \\\\ \\operatorname {f e a t} (\\mathbf {x}, \\boldsymbol {\\epsilon}, t = 2) \\\\ \\vdots \\\\ \\operatorname {f e a t} (\\mathbf {x}, \\boldsymbol {\\epsilon}, t = 9 9 9) \\end{array} \\right\\} \\xrightarrow [ \\text {i n f o r m a t i o n} ]{\\text {c o n s o l i d a t e}} \\operatorname {f e a t} _ {\\mathrm {c}} (\\mathbf {x}) \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 570, + 306, + 859, + 398 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To align our model's features with the diffusion model's features, we maximize the similarity between the diffusion model's features and the projected features of our feature extraction model:", + "bbox": [ + 511, + 407, + 906, + 467 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {f e a t} _ {\\mathrm {c}} (\\mathbf {x}) \\to \\left\\{ \\begin{array}{l} \\operatorname {p r o j} (\\cdot , t = 1) \\longrightarrow \\mathcal {L} \\longleftarrow \\operatorname {f e a t} (\\mathbf {x}, \\epsilon , t = 1) \\\\ \\operatorname {p r o j} (\\cdot , t = 2) \\longrightarrow \\mathcal {L} \\longleftarrow \\operatorname {f e a t} (\\mathbf {x}, \\epsilon , t = 2) \\\\ \\vdots \\\\ \\operatorname {p r o j} (\\cdot , t = 9 9 9) \\longrightarrow \\mathcal {L} \\longleftarrow \\operatorname {f e a t} (\\mathbf {x}, \\epsilon , t = 9 9 9) \\end{array} \\right.\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 469, + 888, + 534 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Specifically, we minimize the negative cosine similarity between the diffusion model's features and our model's features extracted at stages $k = \\{1,\\dots,K\\}$ in the network. Given a clean image $\\mathbf{x}_0$ , the feature extraction model's output for feature map $k$ is denoted as $\\mathrm{feat}_{\\mathrm{c}}^{(k)}(\\mathbf{x}_0)$ . Our Clean-DIFT feature map is then adapted by the learned projection heads $\\mathrm{proj}^{(k)}(\\mathrm{feat}_{\\mathrm{c}}^{(k)}(\\mathbf{x}_0),t)$ , where $\\mathrm{proj}^{(k)}(\\cdot ,\\cdot)$ is the projection head for feature map $k$ . The diffusion model receives the noisy image $\\mathbf{x}_t$ corresponding to the same $\\mathbf{x}_0$ and timestep $t$ . The projection head then learns a timestep-dependent alignment from CleanDIFT features to the diffusion model's features $\\mathrm{feat}^{(k)}(\\mathbf{x}_t;t)$ . Putting it all together, our loss function is defined as:", + "bbox": [ + 511, + 544, + 906, + 741 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = - \\sum_ {k = 1} ^ {K} \\operatorname {s i m} \\left(\\operatorname {p r o j} ^ {(k)} \\left(\\operatorname {f e a t} _ {\\mathrm {c}} ^ {(k)} \\left(\\mathbf {x} _ {0}\\right); t\\right), \\operatorname {f e a t} ^ {(k)} \\left(\\mathbf {x} _ {t}; t\\right)\\right). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 521, + 753, + 906, + 795 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For each training image $\\mathbf{x}_0$ , we sample $I$ different nois-ing timesteps $t_i$ in a stratified manner, with each timestep $t_i \\sim \\mathcal{U}(\\frac{i}{I} T, \\frac{i+1}{I} T)$ , where $T$ is the maximum timestep. By sampling multiple timesteps per image we incentivize the feature extraction model to match the diffusion model's features across the entire noise spectrum.", + "bbox": [ + 511, + 810, + 908, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "120", + "bbox": [ + 486, + 945, + 511, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 89, + 223, + 107 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We test our hypothesis that the proposed extraction setup enables us to leverage more of the world knowledge inherent in diffusion models compared to existing diffusion feature extraction methods while being task-agnostic and timestep-independent. To that end, we evaluate our features on a wide range of downstream tasks: unsupervised zero-shot semantic correspondence, monocular depth estimation, semantic segmentation, and classification. We compare our features against standard diffusion features, methods that combine diffusion features with additional features, and non-diffusion-based approaches.", + "bbox": [ + 89, + 116, + 483, + 282 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 89, + 294, + 279, + 311 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details Following previous works [62, 66, 67], we evaluate our method on a Stable Diffusion (SD) backbone [51]. We apply our method to SD 1.5 and SD 2.1 to enable fair comparisons with existing methods that use either. We fully fine-tune our feature extraction model on image-caption pairs for only 400 steps, taking 30 minutes on a single A100 GPU, which was sufficient for our strong performance and further training did not yield any significant gains. We extract features after the U-Net's middle block and after each of the U-Net's decoder blocks, except the two final blocks. A detailed visualization of where we extract features is provided in Figure 11. This yields a total of $K = 11$ feature maps that we align between the diffusion model and the feature extraction model. Our point-wise feature projection heads consist of three stacked Feed Forward Networks (FFNs) that are zero-initialized such that initially they act as identity mappings due to their residual connections. Since every aligned feature map has its own projection head, this results in 45M additional trainable parameters for SD 2.1. We study the effect of different projection head architectures in Sec. E. We train using Adam [29] with a batch size of 8 and a learning rate of 2e-6 with a linear warmup. For stratified timestep sampling, we utilize $I = 3$ stratification bins across all our experiments, i.e. three different noise levels per training image.", + "bbox": [ + 89, + 316, + 483, + 696 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets We fine-tune our feature extraction model on a random subset of COYO-700M [5], which is similar to the LAION [56], the dataset that Stable Diffusion 1.5 and 2.1 were trained on originally. That way, we ensure that all performance improvements originate from the feature extraction model consolidating the diffusion model's internal feature representations over time, not from choosing a different dataset that matches the test dataset distribution more closely. The subset selects images with a minimum size of $512^2$ . We crop and resize them to match the corresponding input resolution of the underlying diffusion model. We analyze the effect of using different datasets in Sec. G.", + "bbox": [ + 89, + 719, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/7fbfb9bdf92425f5ea83b84afada92433835a9568da4e6ec14b62de7ad65dd8d.jpg", + "image_caption": [ + "Figure 5. Semantic correspondence results using DIFT [62] features with the standard SD 2.1 ( $t = 261$ ) and our CleanDIFT features. Our clean features show significantly less incorrect matches than the base diffusion model." + ], + "image_footnote": [], + "bbox": [ + 514, + 87, + 906, + 290 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Unsupervised Semantic Correspondence", + "text_level": 1, + "bbox": [ + 511, + 368, + 859, + 383 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As many previous diffusion feature methods focus on (unsupervised) semantic correspondence matching [35, 62, 66, 67], we perform an extensive evaluation of our method on this task. Following previous works on semantic correspondence matching [62, 66, 67], we measure our performance in Percentage of Correct Keypoints (PCK). We average PCK directly across all keypoints, not over images. We use $\\alpha = 0.1$ as a threshold and report both PCK values with error margins relative to the image size and to the bounding box size, denoted as $\\mathrm{PCK}_{\\mathrm{img}}$ and $\\mathrm{PCK}_{\\mathrm{bbox}}$ respectively. We evaluate the performance on the test split of the SPair-71k dataset, which consists of approximately $12\\mathrm{k}$ image pairs from 18 categories. Some existing works [62, 66, 67] evaluate on additional datasets but find SPair-71k to be the most challenging and therefore the most informative benchmark. For the text prompt we use \"A photo of a category.\", with category being the corresponding category of the SPair image. We experiment with distilling the text conditioning in Sec. B.", + "bbox": [ + 511, + 390, + 906, + 676 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Results We first compare our extracted features to DIFT [62], an approach that detects semantic correspondences using standard diffusion features. Substituting these with our CleanDIFT features yields a performance increase of 1.79 absolute percentage points for $\\mathrm{PCK}_{\\mathrm{img}}$ and 1.86 percentage points for $\\mathrm{PCK}_{\\mathrm{bbox}}$ . Notably, DIFT averages the extracted feature maps across 8 different noise samples. Without this averaging over noise samples, our performance gain is even larger (2.81 PCK@ $\\alpha_{\\mathrm{img}}$ gain). This indicates that our feature extraction model learns more than a mere averaging over the noise in the diffusion model's feature maps (see Figure 5 and Sec. C for examples). We present an extended version of the time-step dependent performance analysis conducted by [62] in Figure 6: We evaluate the", + "bbox": [ + 511, + 689, + 908, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "121", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/b04caf2667ee539939992132bc059f285b3dc5ff58fc142abc4636fae9c086b6.jpg", + "image_caption": [ + "Figure 6. Following [62], we evaluate semantic correspondence matching accuracy for different noise levels. Our feature extractor outperforms the standard noisy diffusion features across all timesteps $t$ . We additionally demonstrate that simply providing the diffusion model with a clean image and a non-zero timestep does not result in improved performance." + ], + "image_footnote": [], + "bbox": [ + 107, + 88, + 460, + 172 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "diffusion model's performance for different timesteps $t$ in two settings. In the first setting, we provide the diffusion model with a noisy input image $\\mathbf{x}_t$ as usual. In the second setting, we demonstrate that feeding the clean image along with a non-zero timestep is not a viable solution to obtain meaningful features: We provide the diffusion model with the clean input image $\\mathbf{x}_0$ for all timesteps $t$ . We observe that the model's performance for the clean input image degrades faster and has a lower peak than for the noisy input. This is to be expected, as the diffusion model was trained on noisy images, not clean images. Importantly, our CleanDIFT features are timestep-independent and consistently outperform standard diffusion features, even when the latter are optimized for the best-performing timestep. We further observe that this advantage generalizes effectively to other backbones, such as DiTs [44] (see Tab. 7).", + "bbox": [ + 88, + 279, + 482, + 520 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "A Tale of Two Features [66] extends the approach of DIFT by combining diffusion features with DINOv2 [43] features. Again, we replace the standard diffusion features with our CleanDIFT features and observe that the performance gain transfers when combining our features with DINOv2 features. Telling Left from Right [67] further improves upon the results of A Tale of Two Features by introducing a test-time adaptive pose alignment strategy. We observe that the performance gain transfers to this setting as well. To the best of our knowledge, Telling Left from Right combined with our CleanDIFT features sets a new state-of-the-art in unsupervised zero-shot semantic correspondence matching. In summary, replacing standard diffusion features with our CleanDIFT features consistently results in a significant performance improvement across all three methods. We show an overview of the results in Tab. 1 and a more extensive evaluation per category in Sec. B.", + "bbox": [ + 89, + 521, + 482, + 779 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We also investigate the performance of our features in a supervised fine-tuning setting for semantic correspondence matching. Following [35], we train an aggregation network that uses all extracted feature maps and learns to aggregate them into a single task-specific feature map for semantic correspondence matching. In contrast to [35], we do not have to perform costly DDIM inversion [59] to obtain a matching noisy image for every timestep. Instead,", + "bbox": [ + 89, + 780, + 482, + 902 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/490a10983a65771719747da613bfc460abf075446cc427544c1f35a54cdc06f7.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodOur FeaturesPCK@α(↑)
αimg=0.1αbbox=0.1
General Approaches
DINOv2+NN--55.6
Diff. Feat.-based Approaches
DIFT [62]X66.5359.57
✓68.32▲1.7961.43▲1.86
A Tale of Two Features [66]X72.3163.73
✓73.35▲1.0464.81▲1.08
Telling Left from Right [67]X77.0768.64
✓78.40▲1.3369.99▲1.35
", + "bbox": [ + 517, + 88, + 903, + 250 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Zero-shot unsupervised semantic correspondence matching performance comparison on SPair71k [37]. Our improved features consistently lead to substantial improvements in matching performance. We report PCK on the test split of SPair71k, aggregated per point. Numbers are reproduced, for a discussion and comparison to reported numbers view Tab. 5.", + "bbox": [ + 511, + 258, + 906, + 342 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "we directly feed the clean image to our feature extraction model. Therefore, extracting features with our CleanDIFT approach is $50\\mathrm{x}$ faster, since we perform a single denoiser forward pass while [35] performs 50 for the inversion. Our model achieves a $\\mathrm{PCK}_{\\mathrm{img}}$ value of 72.48 vs their 72.75 and a $\\mathrm{PCK}_{\\mathrm{bbox}}$ value of 64.37 vs their 64.53. We observe a slight performance regression compared to their approach, however, at a speedup of $50\\times$ . Luo et al. [35] also present a single-step ablation of their full method that only requires a single forward pass which makes it more comparable to ours. We outperform this single-step version by a wide margin of 9.0 percentage points for $\\mathrm{PCK}_{\\mathrm{img}}$ and 9.1 percentage points for $\\mathrm{PCK}_{\\mathrm{bbox}}$ .", + "bbox": [ + 511, + 354, + 906, + 551 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Depth Estimation", + "text_level": 1, + "bbox": [ + 511, + 561, + 686, + 575 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We also investigate monocular depth estimation on NYUv2 [40]. Similar to [43], we follow the evaluation protocol from [34]. We use SD 2.1 as the base model and resize the input to the model's native resolution of $768^2$ . We extract features from the same location as [62] and obtain a feature map of dimension $48^2$ . Unlike [43], we do not up-sample the features and directly apply the linear probe. The probe predicts depth in 256 uniform bins which we combine with a classification loss after a linear normalization following [3]. We train one probe for our CleanDIFT features and one for standard diffusion features at $t = 299$ , as that timestep minimizes the error in our settings. Our qualitative results (see Figure 7) show a substantial fidelity gap in the estimated depth maps between the features from the standard SD 2.1 backbone and the features from our feature extraction model. This is reflected in a substantial improvement in quantitative metrics over the baseline as seen in Tab. 2. Lastly, we reuse the probe trained on standard diffusion features and apply it on the CleanDIFT features. While this does not match the performance of the CleanDIFT probe, it still achieves significantly better re", + "bbox": [ + 511, + 583, + 906, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "122", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4b075a4acf8c4fdd30068b4c36b81a1fdb99db7d64c70819c392abcf76c7c770.jpg", + "image_caption": [ + "Figure 7. Qualitative results for depth estimation using a linear probe on diffusion features on NYUv2 [40]. Our CleanDIFT features enable substantially better depth estimation than standard diffusion features. Note how the CleanDIFT features are far less noisy when compared to the standard diffusion features." + ], + "image_footnote": [], + "bbox": [ + 99, + 85, + 475, + 262 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/6a17becec707b4c8ececa1edf230bd8099592d33341847691b41903c52c95c15.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodBackboneRMSE (↓)
Self-Supervised Methods
OpenCLIP [28]ViT-G/140.541
MAE [20]ViT-H/140.517
DINO [6]ViT-B/80.555
iBOT [68]ViT-L/160.417
DINOv2 [43]ViT-g/140.344
Diffusion Features
SD 2.1 [51]0.469
DIFF- like [62]Ours0.444,0.025
+ Probes from noisy features0.453,0.016
", + "bbox": [ + 94, + 361, + 478, + 529 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Monocular Depth Estimation. Following [43], we evaluate metric depth prediction on NYUv2 [40] using a linear probe. Our clean features outperform the noisy features by a significant margin. Probes trained on the noisy features can be reused for the clean features, but incur a smaller performance gain.", + "bbox": [ + 89, + 540, + 482, + 609 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "sults when compared to using standard diffusion features. This indicates that our features can be used as a drop-in replacement for the original diffusion features and offer improved performance on downstream applications.", + "bbox": [ + 89, + 625, + 483, + 686 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Semantic Segmentation", + "text_level": 1, + "bbox": [ + 89, + 696, + 307, + 713 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To further investigate the difference between standard noisy diffusion features and our CleanDIFT features, we evaluate on the semantic segmentation task by training linear probes on our CleanDIFT features and on standard diffusion features. We utilize SD2.1 as the diffusion backbone and extract features at the same location as [62]. This procedure yields feature maps of size $48^2$ . We train our linear probe on the $48^2$ feature maps and upscale the obtained segmentation masks using nearest neighbor upsampling. We train and evaluate on the PASCAL VOC dataset [16]. Following common practice [39], we use mean Intersection over Union (mIoU) as the evaluation metric. Qualitative re", + "bbox": [ + 89, + 719, + 483, + 900 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/703e20298c04cd3462fa6eef64a71f464ec16a7a012ff999757a2546f93a18bf.jpg", + "image_caption": [ + "Figure 8. Qualitative results for semantic segmentation from diffusion features on Pascal VOC [16]. Standard SD features use $t = 100$ as the timestep, which we found to perform best quantitatively (cf. Figure 9). Note how the CleanDIFT segmentation maps are far less noisy than those of the standard diffusion features." + ], + "image_footnote": [], + "bbox": [ + 519, + 85, + 903, + 247 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/751c2eee31ecec5a62b102e93fb85293f183c5fb6cdb447a0e8be90bdee80fe1.jpg", + "image_caption": [ + "Figure 9. Performance on semantic segmentation using linear probes. Our clean features outperform the noisy diffusion features for the best noising timestep $t$ . Semantic segmentation performance of a standard diffusion model heavily depends on the used noising timestep. Unlike for semantic correspondence matching, the optimal $t$ value appears to be around $t = 100$ ." + ], + "image_footnote": [], + "bbox": [ + 553, + 342, + 862, + 429 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "sults are shown in Figure 8. Using our features, we observe significantly less noisy segmentations than with standard diffusion features. We show a quantitative comparison of our CleanDIFT feature's performance against standard diffusion features across timesteps in Figure 9. Notably, the optimal timestep appears to be around $t = 100$ , in contrast to the optimal timestep for semantic correspondences, which [62] found to be $t = 261$ . This highlights the need for tuning a timestep individually per downstream task. Our method both alleviates the need for such a timestep tuning and outperforms the standard diffusion features for the optimal timestep.", + "bbox": [ + 511, + 536, + 908, + 717 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Classification", + "text_level": 1, + "bbox": [ + 511, + 727, + 653, + 742 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To assess the impact of our method on non-spatial tasks, we evaluate classification performance using pooled features. Pooling mitigates the influence of localized noise, so we anticipate classification performance to remain on par with standard diffusion features unless our setup introduces detrimental effects. We perform k-Nearest Neighbor (kNN) classification with $k = 10$ on ImageNet1k [11], using SD 1.5 as the diffusion backbone. We sweep across feature maps and timesteps $t$ for the base model, with results presented in Figure 10. Our analysis shows that the fea", + "bbox": [ + 511, + 750, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "123", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b72a4eccdb33442134d8ae17ce703235bd652130868923e4999f81db0ef6b8f9.jpg", + "image_caption": [ + "Figure 10. Classification performance on ImageNet1k [11], using kNN classifier with $k = 10$ and cosine similarity as the distance metric. We sweep over different timesteps and feature maps. We find that the feature map with the lowest spatial resolution (feature map #0) yields the highest classification accuracy." + ], + "image_footnote": [], + "bbox": [ + 101, + 88, + 467, + 281 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ture map with the lowest spatial resolution, i.e., feature map #0 (see Figure 11), achieves the highest classification accuracy. Furthermore, the optimal timestep $t$ for the base model varies between feature maps. For the best-performing feature map, $t = 100$ yields the highest classification accuracy.", + "bbox": [ + 89, + 371, + 482, + 446 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Importantly, CleanDIFT features slightly outperform the standard diffusion features even when using an optimal timestep $t$ for the base model showing that it does not introduce any detrimental effects.", + "bbox": [ + 89, + 446, + 483, + 508 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.6. Ablation Studies", + "text_level": 1, + "bbox": [ + 89, + 518, + 254, + 534 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "For simplicity, we perform our ablation studies using DIFT [62] and evaluate the performance for unsupervised zero-shot semantic correspondence matching on a subset of the SPair71k [37] test split.", + "bbox": [ + 89, + 541, + 483, + 602 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Training Objective During training, we maximize the cosine similarity between projected outputs of our feature extraction model and standard diffusion features to align them. To investigate the influence of the employed similarity metric, we compare feature extraction models trained on three different alignment objectives commonly used in similar contexts: mean absolute error $(L_{1})$ , mean squared error $(L_{2})$ , and cosine similarity. Quantitative results are provided in Tab. 3. While all objectives result in feature extraction models that outperform standard diffusion features, cosine similarity consistently performs the best across the alignment objectives by a significant margin, confirming our choice of similarity metric.", + "bbox": [ + 89, + 614, + 482, + 811 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Projection Heads We investigate the influence of our proposed projection heads that are used to project Clean-DIFT features onto standard diffusion features. The alignment of feature extraction model and diffusion model is determined by the utilized similarity metric and the pro", + "bbox": [ + 89, + 824, + 483, + 902 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/3760c8264721af75dc2b87439e32644434564cdba61ea4c58b038c6a6a5db62e.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ObjectiveProjection HeadsPCK@α(↑)
αimg=0.1αbbox=0.1
Cosine Sim.✓68.3261.43
×68.1661.29
L2✓66.2359.13
×66.4959.43
L1✓66.9160.00
×66.8759.91
SD 2.1-63.4155.92
", + "bbox": [ + 517, + 88, + 903, + 268 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3. Ablation Study Results. We evaluate the feature extraction models' performance for zero-shot semantic correspondence matching on the SPair71k test split. PCK is aggregated per point.", + "bbox": [ + 511, + 279, + 903, + 321 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "jection heads. Therefore, we evaluate our feature extraction model's performance with and without the projection heads in combination with all three similarity metrics. An overview of the comparison is given in Tab. 3. In our main configuration that uses cosine similarity, using the projection heads yields slight performance improvements of 0.24 percentage points for $\\mathrm{PCK}_{\\mathrm{img}}$ and 0.06 percentage points for $\\mathrm{PCK}_{\\mathrm{bbox}}$ compared to fine-tuning without the projection heads. As the projection heads are typically not used for inference, they add computational overhead only during the lightweight fine-tuning. Therefore, we argue that it is worthwhile to include them and leverage the small performance gain. Additionally, they can be reused to efficiently obtain feature maps for specific timesteps.", + "bbox": [ + 511, + 333, + 906, + 545 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 558, + 633, + 573 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we introduced CleanDIFT, a novel approach for extracting diffusion features. CleanDIFT produces noise-free, timestep-independent, general-purpose diffusion features by consolidating timestep-dependent representations from a pre-trained diffusion backbone into a unified feature representation. We achieve this alignment between our feature extraction model and the pre-trained diffusion backbone through a lightweight fine-tuning procedure that takes approximately 30 minutes on a single A100 GPU. Operating directly on clean images, our method eliminates the information loss associated with adding noise to input images. Furthermore, CleanDIFT removes the requirement for tuning timesteps for each downstream task and avoids the computational overhead of ensembling over noise levels or timesteps. Instead, our method efficiently extracts features with just a single forward pass at inference time, substantially reducing inference costs compared to methods relying on ensembling or inversion. Extensive evaluations of CleanDIFT across diverse downstream tasks demonstrate significant performance improvements over conventional diffusion features.", + "bbox": [ + 511, + 583, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "124", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 91, + 90, + 250, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This project has been supported by the German Federal Ministry for Economic Affairs and Climate Action within the project \"NXT GEN AI METHODS - Generative Methoden für Perzeption, Prädiktion und Planung\", the project \"GeniusRobot\" (01IS24083), funded by the Federal Ministry of Education and Research (BMBF), the bidt project KLIMA-MEMES, and the German Research Foundation (DFG) project 421703927. The authors gratefully acknowledge the Gauss Center for Supercomputing for providing compute through the NIC on JUWELS at JSC and the HPC resources supplied by the Erlangen National High Performance Computing Center (NHR@FAU funded by DFG project 440719683) under the NHR project JA-22883.", + "bbox": [ + 89, + 114, + 483, + 311 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Further, we would like to thank Owen Vincent for continuous technical support.", + "bbox": [ + 89, + 313, + 483, + 344 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 359, + 187, + 376 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Ji-aming Song, Qinsheng Zhang, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, et al. ediff-i: Text-to-image diffusion models with an ensemble of expert denoisers. arXiv preprint arXiv:2211.01324, 2022. 2, 3", + "[2] Dmitry Baranchuk, Andrey Voynov, Ivan Rubachev, Valentin Khrulkov, and Artem Babenko. Label-efficient semantic segmentation with diffusion models. In International Conference on Learning Representations, 2022. 1, 3", + "[3] Shariq Farooq Bhat, Ibrahim Alhashim, and Peter Wonka. Adabins: Depth estimation using adaptive bins. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4009-4018, 2021. 6", + "[4] Giulio Biroli, Tony Bonnaire, Valentin De Bortoli, and Marc Mézard. Dynamical regimes of diffusion models. arXiv preprint arXiv:2402.18491, 2024. 2, 3", + "[5] Minwoo Byeon, Beomhee Park, Haecheon Kim, Sungjun Lee, Woonhyuk Baek, and Saehoon Kim. Coyo-700m: Image-text pair dataset. https://github.com/kakaobrain/coyo-dataset, 2022.5", + "[6] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 2, 7", + "[7] Junsong Chen, Jincheng YU, Chongjian GE, Lewei Yao, Enze Xie, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart- $\\mathbb{S}\\backslash \\mathbb{alpha}$ : Fast training of diffusion transformer for photorealistic text-to-image synthesis. In The Twelfth International Conference on Learning Representations, 2024. 14", + "[8] Shoufa Chen, Peize Sun, Yibing Song, and Ping Luo. Diffusionnet: Diffusion model for object detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 19830-19843, 2023. 1, 3", + "[9] Xinlei Chen, Zhuang Liu, Saining Xie, and Kaiming He. De" + ], + "bbox": [ + 99, + 383, + 483, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "constructing denoising diffusion models for self-supervised learning. arXiv preprint arXiv:2401.14404, 2024. 2, 3", + "[10] Timothee Darcet, Maxime Oquab, Julien Mairal, and Piotr Bojanowski. Vision transformers need registers. In The Twelfth International Conference on Learning Representations, 2024. 2", + "[11] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 248–255, 2009. 7, 8", + "[12] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 1, 2", + "[13] Patrick Esser, Johnathan Chiu, Parmida Atighehchian, Jonathan Granskog, and Anastasis Germanidis. Structure and content-guided video synthesis with diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7346-7356, 2023. 1, 2", + "[14] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In *Forty-first International Conference on Machine Learning*, 2024. 2, 14", + "[15] Zach Evans, CJ Carr, Josiah Taylor, Scott H Hawley, and Jordi Pons. Fast timing-conditioned latent audio diffusion. In *Forty-first International Conference on Machine Learning*, 2024. 2", + "[16] M. Everingham, L. Van Gool, C. K. I. Williams, J. Winn, and A. Zisserman. The Pascal visual object classes (voc) challenge. International Journal of Computer Vision, 88(2): 303-338, 2010. 7, 13", + "[17] Michael Fuest, Pingchuan Ma, Ming Gui, Johannes S. Fischer, Vincent Tao Hu, and Björn Ommer. Diffusion models and representation learning: A survey. CoRR, abs/2407.00783, 2024. 3", + "[18] Frank Fundel, Johannes Schusterbauer, Vincent Tao Hu, and Björn Ommer. Distillation of diffusion features for semantic correspondence. WACV, 2025. 3", + "[19] Spyros Gidaris, Praveer Singh, and Nikos Komodakis. Unsupervised representation learning by predicting image rotations. In International Conference on Learning Representations, 2018. 2", + "[20] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dólár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 2, 7", + "[21] Eric Hedlin, Gopal Sharma, Shweta Mahajan, Hossam Isack, Abhishek Kar, Andrea Tagliasacchi, and Kwang Moo Yi. Unsupervised semantic correspondence using stable diffusion. Advances in Neural Information Processing Systems, 36, 2023. 3", + "[22] Geoffrey Hinton, Oriol Vinyls, and Jeffrey Dean. Distilling the knowledge in a neural network. In NIPS Deep Learning and Representation Learning Workshop, 2015. 3", + "[23] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 3" + ], + "bbox": [ + 516, + 92, + 906, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "125", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[24] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2", + "[25] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. Advances in Neural Information Processing Systems, 35:8633-8646, 2022. 1", + "[26] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. *ICLR*, 1(2):3, 2022. 14", + "[27] Drew A Hudson, Daniel Zoran, Mateusz Malinowski, Andrew K Lampinen, Andrew Jaegle, James L McClelland, Loic Matthey, Felix Hill, and Alexander Lerchner. Soda: Bottleneck diffusion models for representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23115-23127, 2024. 3", + "[28] Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. Openclip, 2021. 7", + "[29] Diederik P. Kingma and Jimmy Lei Ba. Adam: A method for stochastic optimization. In Proceedings of the 3rd International Conference on Learning Representations (ICLR), 2015. 5", + "[30] Zhifeng Kong, Wei Ping, Jiaji Huang, Kexin Zhao, and Bryan Catanzaro. Diffwave: A versatile diffusion model for audio synthesis. In International Conference on Learning Representations, 2021. 2", + "[31] Alexander C Li, Mihir Prabhudesai, Shivam Duggal, Ellis Brown, and Deepak Pathak. Your diffusion model is secretly a zero-shot classifier. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2206-2217, 2023. 1, 3", + "[32] Daiqing Li, Huan Ling, Amlan Kar, David Acuna, Seung Wook Kim, Karsten Kreis, Antonio Torralba, and Sanja Fidler. Dreamteacher: Pretraining image backbones with deep generative models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16698-16708, 2023. 3", + "[33] Xinghui Li, Jingyi Lu, Kai Han, and Victor Adrian Prisacariu. Sd4match: Learning to prompt stable diffusion model for semantic matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 27558-27568, 2024. 3", + "[34] Zhenyu Li, Xuyang Wang, Xianming Liu, and Junjun Jiang. Binsformer: Revisiting adaptive bins for monocular depth estimation. IEEE Transactions on Image Processing, 2024. 6", + "[35] Grace Luo, Lisa Dunlap, Dong Huk Park, Aleksander Holynski, and Trevor Darrell. Diffusion hyperfeatures: Searching through time and space for semantic correspondence. In Advances in Neural Information Processing Systems, pages 47500-47510. Curran Associates, Inc., 2023. 3, 5, 6, 13", + "[36] Chenlin Meng, Robin Rombach, Ruiqi Gao, Diederik Kingma, Stefano Ermon, Jonathan Ho, and Tim Salimans." + ], + "bbox": [ + 91, + 90, + 480, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "On distillation of guided diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14297-14306, 2023. 3", + "[37] Juhong Min, Jongmin Lee, Jean Ponce, and Minsu Cho. Spair-71k: A large-scale benchmark for semantic correspondence. arXiv preprint arXiv:1908.10543, 2019. 6, 8, 12, 13", + "[38] Ishan Misra and Laurens van der Maaten. Self-supervised learning of pretext-invariant representations. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6707-6717, 2020. 2", + "[39] Koichi Namekata, Amirmojtaba Sabour, Sanja Fidler, and Seung Wook Kim. Emerdiff: Emerging pixel-level semantic knowledge in diffusion models. In The Twelfth International Conference on Learning Representations, 2023. 1, 3, 7", + "[40] Pushmeet Kohli Nathan Silberman, Derek Hoiem and Rob Fergus. Indoor segmentation and support inference from rgbd images. In ECCV, 2012. 6, 7, 14", + "[41] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 1, 2", + "[42] Mehdi Noroozi and Paolo Favaro. Unsupervised learning of visual representations by solving jigsaw puzzles. In European conference on computer vision, pages 69-84. Springer, 2016. 2", + "[43] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel HAZIZA, Francisco Massa, Alaaeldin El-Nouby, Mido Assran, Nicolas Ballas, Wojciech Galuba, Russell Howes, Po-Yao Huang, Shang-Wen Li, Ishan Misra, Michael Rabbat, Vasu Sharma, Gabriel Synnaeve, Hu Xu, Herve Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. DINOv2: Learning robust visual features without supervision. Transactions on Machine Learning Research, 2024. 2, 3, 6, 7", + "[44] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4195-4205, 2023. 6", + "[45] Ethan Perez, Florian Strub, Harm De Vries, Vincent Dumoulin, and Aaron Courville. Film: Visual reasoning with a general conditioning layer. In Proceedings of the AAAI conference on artificial intelligence, 2018. 14", + "[46] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. In The Twelfth International Conference on Learning Representations, 2024. 2, 14", + "[47] Adam Polyak, Amit Zohar, Andrew Brown, Andros Tjandra, Animesh Sinha, Ann Lee, Apoorv Vyas, Bowen Shi, Chih-Yao Ma, Ching-Yao Chuang, et al. Movie gen: A cast of media foundation models. arXiv preprint arXiv:2410.13720, 2024. 2", + "[48] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry," + ], + "bbox": [ + 516, + 90, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "126", + "bbox": [ + 486, + 945, + 511, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2", + "[49] Prajit Ramachandran, Barret Zoph, and Quoc V Le. Searching for activation functions. arXiv preprint arXiv:1710.05941, 2017. 14", + "[50] Severi Rissanen, Markus Heinonen, and Arno Solin. Generative modelling with inverse heat dissipation. arXiv preprint arXiv:2206.13397, 2022. 3", + "[51] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 2, 5, 7", + "[52] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention - MICCAI 2015, pages 234-241, Cham, 2015. Springer International Publishing. 3", + "[53] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural information processing systems, 35:36479-36494, 2022. 2", + "[54] Tim Salimans and Jonathan Ho. Progressive distillation for fast sampling of diffusion models. arXiv preprint arXiv:2202.00512, 2022. 3", + "[55] Axel Sauer, Dominik Lorenz, Andreas Blattmann, and Robin Rombach. Adversarial diffusion distillation. In European Conference on Computer Vision, pages 87-103. Springer, 2024. 14", + "[56] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278-25294, 2022. 5", + "[57] Noam Shazeer. Glu variants improve transformer. arXiv preprint arXiv:2002.05202, 2020. 14", + "[58] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. In The Eleventh International Conference on Learning Representations, 2023. 1", + "[59] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In International Conference on Learning Representations, 2021. 6", + "[60] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. Advances in neural information processing systems, 32, 2019. 2", + "[61] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 2", + "[62] Luming Tang, Menglin Jia, Qianqian Wang, Cheng Perng Phoo, and Bharath Hariharan. Emergent correspondence" + ], + "bbox": [ + 91, + 90, + 482, + 901 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "from image diffusion. In Advances in Neural Information Processing Systems, pages 1363-1389. Curran Associates, Inc., 2023. 1, 2, 3, 4, 5, 6, 7, 8, 12, 13, 14, 15", + "[63] Weijia Wu, Yuzhong Zhao, Mike Zheng Shou, Hong Zhou, and Chunhua Shen. Diffumask: Synthesizing images with pixel-level annotations for semantic segmentation using diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1206-1217, 2023. 1, 3", + "[64] Jiarui Xu, Sifei Liu, Arash Vahdat, Wonmin Byeon, Xiaolong Wang, and Shalini De Mello. Open-vocabulary panoptic segmentation with text-to-image diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2955-2966, 2023. 1, 3, 13, 15", + "[65] Biao Zhang and Rico Sennrich. Root mean square layer normalization. Advances in Neural Information Processing Systems, 32, 2019. 14", + "[66] Junyi Zhang, Charles Herrmann, Junhwa Hur, Luisa Polania Cabrera, Varun Jampani, Deqing Sun, and Ming-Hsuan Yang. A tale of two features: Stable diffusion complements dino for zero-shot semantic correspondence. In Advances in Neural Information Processing Systems, pages 45533-45547. Curran Associates, Inc., 2023. 1, 2, 3, 5, 6, 12, 13, 15", + "[67] Junyi Zhang, Charles Herrmann, Junhwa Hur, Eric Chen, Varun Jampani, Deqing Sun, and Ming-Hsuan Yang. Telling left from right: Identifying geometry-aware semantic correspondence. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3076-3085, 2024. 1, 2, 3, 5, 6, 12, 13, 15", + "[68] Jinghao Zhou, Chen Wei, Huiyu Wang, Wei Shen, Cihang Xie, Alan Yuille, and Tao Kong. Image BERT pre-training with online tokenizer. In International Conference on Learning Representations, 2022. 7" + ], + "bbox": [ + 516, + 92, + 903, + 584 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "127", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/CleanDIFT_ Diffusion Features without Noise/17f652ff-055a-4bf1-a266-9144370345a9_model.json b/2025/CleanDIFT_ Diffusion Features without Noise/17f652ff-055a-4bf1-a266-9144370345a9_model.json new file mode 100644 index 0000000000000000000000000000000000000000..010b75067d265c77571393c6f8673c30424aea24 --- /dev/null +++ b/2025/CleanDIFT_ Diffusion Features without Noise/17f652ff-055a-4bf1-a266-9144370345a9_model.json @@ -0,0 +1,2466 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.257, + 0.131, + 0.744, + 0.152 + ], + "angle": 0, + "content": "\\(\\checkmark\\) CleanDIFT: Diffusion Features without Noise" + }, + { + "type": "title", + "bbox": [ + 0.104, + 0.181, + 0.216, + 0.198 + ], + "angle": 0, + "content": "Nick Stracke*" + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.181, + 0.46, + 0.198 + ], + "angle": 0, + "content": "Stefan Andreas Baumann*" + }, + { + "type": "title", + "bbox": [ + 0.495, + 0.182, + 0.6, + 0.199 + ], + "angle": 0, + "content": "Kolja Bauer*" + }, + { + "type": "title", + "bbox": [ + 0.632, + 0.182, + 0.74, + 0.198 + ], + "angle": 0, + "content": "Frank Fundel" + }, + { + "type": "title", + "bbox": [ + 0.78, + 0.182, + 0.895, + 0.199 + ], + "angle": 0, + "content": "Björn Ommer" + }, + { + "type": "text", + "bbox": [ + 0.357, + 0.2, + 0.642, + 0.217 + ], + "angle": 0, + "content": "CompVis @ LMU Munich, MCML" + }, + { + "type": "text", + "bbox": [ + 0.368, + 0.219, + 0.625, + 0.234 + ], + "angle": 0, + "content": "{nick.stracke,b.ommer}@lmu.de" + }, + { + "type": "text", + "bbox": [ + 0.365, + 0.241, + 0.632, + 0.256 + ], + "angle": 0, + "content": "compvis.github.io/cleandift" + }, + { + "type": "title", + "bbox": [ + 0.252, + 0.269, + 0.55, + 0.286 + ], + "angle": 0, + "content": "Normal Diffusion Features 30 min Finetuning" + }, + { + "type": "title", + "bbox": [ + 0.564, + 0.274, + 0.713, + 0.288 + ], + "angle": 0, + "content": "CleanDIFT Features" + }, + { + "type": "text", + "bbox": [ + 0.212, + 0.292, + 0.474, + 0.305 + ], + "angle": 0, + "content": "noisy images noisy features" + }, + { + "type": "image", + "bbox": [ + 0.215, + 0.307, + 0.288, + 0.358 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.232, + 0.359, + 0.275, + 0.368 + ], + "angle": 0, + "content": "\\(t = 300\\)" + }, + { + "type": "image", + "bbox": [ + 0.293, + 0.307, + 0.383, + 0.366 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.385, + 0.307, + 0.472, + 0.366 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.269, + 0.372, + 0.426, + 0.387 + ], + "angle": 0, + "content": "timestep tuning required \\(X\\)" + }, + { + "type": "image", + "bbox": [ + 0.216, + 0.39, + 0.286, + 0.451 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.291, + 0.392, + 0.382, + 0.451 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.385, + 0.393, + 0.472, + 0.451 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.292, + 0.778, + 0.305 + ], + "angle": 0, + "content": "clean images clean features" + }, + { + "type": "image", + "bbox": [ + 0.511, + 0.307, + 0.594, + 0.366 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.595, + 0.308, + 0.688, + 0.366 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.692, + 0.308, + 0.776, + 0.366 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.555, + 0.373, + 0.731, + 0.387 + ], + "angle": 0, + "content": "no timestep tuning required \\(\\checkmark\\)" + }, + { + "type": "image", + "bbox": [ + 0.512, + 0.392, + 0.594, + 0.451 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.595, + 0.393, + 0.688, + 0.451 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.692, + 0.393, + 0.776, + 0.451 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.469, + 0.907, + 0.526 + ], + "angle": 0, + "content": "Figure 1. Our proposed CleanDIFT feature extraction method yields noise-free, timestep-independent, general-purpose features that significantly outperform standard diffusion features. CleanDIFT operates on clean images, while extracting diffusion features with existing approaches requires adding noise to an image before passing it through the model. Adding noise reduces the information present in the image and requires tuning a timestep per downstream task." + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.538, + 0.327, + 0.554 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.577, + 0.486, + 0.835 + ], + "angle": 0, + "content": "Internal features from large-scale pre-trained diffusion models have recently been established as powerful semantic descriptors for a wide range of downstream tasks. Works that use these features generally need to add noise to images before passing them through the model to obtain the semantic features, as the models do not offer the most useful features when given images with little to no noise. We show that this noise has a critical impact on the usefulness of these features that cannot be remedied by ensembling with different random noises. We address this issue by introducing a lightweight, unsupervised fine-tuning method that enables diffusion backbones to provide high-quality, noise-free semantic features. We show that these features readily outperform previous diffusion features by a wide margin in a wide variety of extraction setups and downstream tasks, offering better performance than even ensemble-based methods at a fraction of the cost." + }, + { + "type": "title", + "bbox": [ + 0.515, + 0.538, + 0.645, + 0.554 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.567, + 0.907, + 0.779 + ], + "angle": 0, + "content": "Learning meaningful visual representations that capture a vast amount of world knowledge remains a key problem in the field of computer vision. Diffusion models can be trained at scale in a self-supervised manner and have rapidly advanced the state of the art in image [12, 41, 51] and video generation [13, 25, 58], making them a good candidate to learn visual representations. Many early works have already achieved impressive results using internal features from large-scale pretrained diffusion models for a wide variety of tasks, such as semantic correspondence detection [62, 66, 67], semantic segmentation [2, 39, 63], panoptic segmentation [64], object detection [8], and classification [31]. However, the optimal approach to extract this world knowledge from a diffusion model remains uncertain." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.78, + 0.909, + 0.902 + ], + "angle": 0, + "content": "To understand why that is the case, we take a look at how diffusion models are trained: a varying amount of noise is added to a clean input image (forward process) and the model is tasked to remove the noise from the image (backward process). The amount of added noise is dependent on the diffusion timestep. As a result, the model learns to operate on noisy images and also becomes dependent on the noise timestep as different noise levels require the model to" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.887, + 0.221, + 0.9 + ], + "angle": 0, + "content": "*Equal Contribution" + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.945, + 0.511, + 0.957 + ], + "angle": 0, + "content": "117" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.212 + ], + "angle": 0, + "content": "perform different tasks [1, 4]. Since noisy images inherently contain less information than clean images (cf. Figure 2), we hypothesize that this harms the internal feature representation of diffusion models [9] and, thus, the extractable world knowledge. Furthermore, the timestep acts as a hyperparameter that influences the internal feature representation and needs to be picked independently for every downstream application (cf. Figure 14)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.214, + 0.483, + 0.365 + ], + "angle": 0, + "content": "We propose a novel feature extraction method that (1) eliminates the need to destroy information by adding noise to the input; and (2) produces timestep-independent generic diffusion features useful for a wide range of down-stream tasks, alleviating the need to tune a noisig timestep per down-stream task. We show how to adapt an off-the-shelf large-scale pre-trained diffusion backbone to provide these features at minimal cost (approximately 30 minutes of fine-tuning on a single A100 GPU) and demonstrate improved performance across a wide range of downstream tasks." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.367, + 0.483, + 0.502 + ], + "angle": 0, + "content": "We achieve this by viewing a diffusion model as a family of \\( T \\) feature extractors that operate on images with different noise levels and provide features with different characteristics. We consolidate all \\( T \\) feature extraction functions in our feature extractor by aligning their internal representations. Specifically, we initialize our feature extractor as a trainable copy of the diffusion model; fine-tune it with clean images and no timestep input; and align its features with all \\( T \\) time-dependent feature extractors of the diffusion model." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.506, + 0.483, + 0.747 + ], + "angle": 0, + "content": "We evaluate our improved features across a wide variety of downstream tasks, such as semantic correspondence matching, monocular depth estimation, semantic segmentation, and classification, and find that they consistently improve upon approaches based on standard diffusion features. These improvements are most evident for dense visual tasks such as semantic correspondence matching, where our features show substantial performance gains across a wide variety of setups [62, 66, 67] and set a new state-of-the-art for unsupervised semantic correspondence matching. Additionally, our proposed method eliminates the need for noise or timestep ensembling [62], offering substantial speed gains (e.g., \\(8 \\times\\) over DIFT [62]), on top of improved quality. Our method is generic and integrates easily with established methods, such as fusing diffusion and DINOv2 features [66, 67]." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.749, + 0.347, + 0.763 + ], + "angle": 0, + "content": "Our main contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.765, + 0.483, + 0.825 + ], + "angle": 0, + "content": "1. We propose CleanDIFT, a finetuning approach for diffusion models that enables them to operate on clean images and makes the inherent world knowledge of these models more accessible." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.826, + 0.483, + 0.871 + ], + "angle": 0, + "content": "2. We show how to consolidate information from all diffusion timesteps into a single feature prediction, removing the need for task-specific timestep tuning." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.483, + 0.902 + ], + "angle": 0, + "content": "3. We demonstrate significant performance gains of our diffusion feature extraction technique across a wide range" + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.765, + 0.483, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.092, + 0.709, + 0.238 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.538, + 0.246, + 0.693, + 0.257 + ], + "angle": 0, + "content": "(a) Reconstruction without Noise" + }, + { + "type": "image", + "bbox": [ + 0.714, + 0.092, + 0.902, + 0.238 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.721, + 0.24, + 0.89, + 0.263 + ], + "angle": 0, + "content": "(b) Diffusion Model Reconstruction with added noise \\((t = 261\\) [62])" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.275, + 0.905, + 0.345 + ], + "angle": 0, + "content": "Figure 2. Deterioration of Diffusion Features. As current methods need to pass noisy images to the model to obtain useful features, they significantly reduce the information available. We alleviate this problem by obtaining useful features without noise, improving the performance of downstream tasks." + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.372, + 0.905, + 0.463 + ], + "angle": 0, + "content": "of down-stream tasks, notably surpassing the current state of the art in zero-shot unsupervised semantic correspondence detection. We further demonstrate the generality of our enhanced features by showing that these performance gains transfer to advanced methods that fuse diffusion features or operate in a supervised setting." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.463, + 0.905, + 0.508 + ], + "angle": 0, + "content": "4. Our proposed approach is significantly more efficient than previous methods that tried to address this problem by noise ensembling or supervised training." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.522, + 0.655, + 0.538 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.548, + 0.907, + 0.775 + ], + "angle": 0, + "content": "Self-Supervised Representation Learning Features from large, pre-trained foundation models have been shown to yield competitive performance to supervised models for a variety of downstream tasks, both in zero-shot and fine-tuning settings [20, 43, 48]. These foundation models are trained on different pre-text tasks like inpainting [20], predicting transformations [19], patch reordering [38, 42], and discriminative tasks [6, 43]. DINOv2 [43] uses a discriminative objective combined with self-distillation to learn general-purpose visual features that have proven useful for a variety of downstream tasks [10]. CLIP [48] learns such features by employing a contrastive objective on text-image pairs. Masked Autoencoders [20] (MAEs) are trained to reconstruct masked out patches of the input, also resulting in general-purpose visual features." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Diffusion Models as Self-Supervised Learners Diffusion models [24, 60, 61] are generative models that have defined the state-of-the-art in image generation [12, 14, 41, 46, 51, 53], video generation [13, 47], and audio generation [15, 30] in recent years. Their primary purpose is to generate high-quality samples (images, videos, etc.). However, generation can also be interpreted as a pretext task for" + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "118" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.242 + ], + "angle": 0, + "content": "learning expressive features, since the model has to build up comprehensive world knowledge in order to generate plausible samples [9, 17, 27, 32, 62]. Features from diffusion models (typically referred to as diffusion features) are obtained by passing a noised image through the diffusion model and extracting intermediate feature representations. They have been shown to be useful for a variety of tasks such as finding semantic correspondences [21, 35, 62], semantic and panoptic segmentation [2, 39, 63, 64], classification [31], and object detection [8]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.244, + 0.482, + 0.394 + ], + "angle": 0, + "content": "For semantic correspondence matching, features are leveraged to identify semantically matching regions across images. Existing approaches utilize diffusion features either in a zero-shot setting [21, 62] or fine-tune them for the semantic correspondence task [33, 35, 67]. Some zero-shot approaches do not fine-tune on semantic correspondence but still require tuning a prompt to activate attention maps at the query location of the correspondence [21]. In contrast, our approach aims to provide universal features usable for various down-stream tasks in a true zero-shot manner." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.396, + 0.482, + 0.547 + ], + "angle": 0, + "content": "Further, diffusion features have been shown to complement features from other self-supervised learning methods such as DINOv2 [18, 43, 66]. DINOv2 features provide sparse but accurate semantic information, while Diffusion features provide dense spatial information albeit with sometimes inaccurate semantics. State-of-the-art approaches for semantic correspondence detection exploit this and fuse features [66, 67]. Compared to DINOv2 features, diffusion features yield smoother, spatially more coherent correspondences [66]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.564, + 0.483, + 0.716 + ], + "angle": 0, + "content": "Distillation for Diffusion Models Knowledge Distillation [22] is a technique used to distill knowledge from a teacher model into a student model. In the context of diffusion models, this is typically applied either to reduce the required denoising steps [54] or to distill a classifier-free guided [23] model into one without CFG [36]. While our approach is inspired by distillation, we consolidate features from \\( T \\) different models, with \\( T \\) being the number of discrete diffusion timesteps, because the features are different at every timestep." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.731, + 0.182, + 0.747 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.757, + 0.231, + 0.772 + ], + "angle": 0, + "content": "3.1. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Diffusion Models Diffusion models are trained to predict a clean image \\(\\mathbf{x}_0\\) given a noisy image \\(\\mathbf{x}_t\\), either explicitly or implicitly. The noisy image \\(\\mathbf{x}_t\\) is a weighted sum with random Gaussian noise \\(\\mathbf{x}_t = \\sqrt{\\alpha_t}\\mathbf{x}_0 + \\sqrt{1 - \\alpha_t}\\epsilon\\) with timestep-dependent coefficients \\(\\alpha_{t}\\) and noise \\(\\epsilon \\sim \\mathcal{N}(0,\\mathbf{I})\\). \\(t\\in [0,T]\\) denotes the time step of the diffusion process, with \\(t = 0\\) corresponding to the clean image and \\(t = T\\) corresponding to pure noise." + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.089, + 0.896, + 0.179 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.188, + 0.907, + 0.259 + ], + "angle": 0, + "content": "Figure 3. Fraction of variance of diffusion features explained by 1) encoding the clean image at \\( t = 0 \\) (no additive noise), and 2) encoding just the added noise \\( \\epsilon \\) at \\( t = 999 \\). Even at relatively low timesteps such as \\( t = 261 \\) as used by DIFT [62], a substantial part of the features directly depends only on the added noise." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.285, + 0.907, + 0.436 + ], + "angle": 0, + "content": "Intuitively, the model faces different objectives for different noise levels [1, 4]: for very high noise, there is little information in the input, and the model is first generating the coarse structure of the image [50]. At lower noise levels, more high- and medium-frequency information is available and the task shifts to generating finer details and intricate structures. This multi-objective nature intuitively explains why previous methods found diffusion features extracted from different timesteps to provide information with differing semantics." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.457, + 0.909, + 0.64 + ], + "angle": 0, + "content": "Diffusion Feature Extraction Typically, diffusion feature extraction happens after first adding noise to an image and passing the resulting \\(\\mathbf{x}_t\\) to a U-Net [52] denoiser. Features are then extracted at multiple hand-picked locations of the U-Net decoder [2, 39, 62, 66, 67]. Different levels of noise added to the input image result in features beneficial for different downstream applications. Typical diffusion timesteps are \\(t = 261\\) [62], \\(t = 100\\) [66] or \\(t = 50\\) [67]. By adding noise to the input image, these methods bottleneck the perceptual information the model can extract. To illustrate this, we show Stable Diffusion 2.1's reconstruction of an image noised at \\(t = 261\\) [62] in Figure 2." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.659, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Diffusion Features Encode Noise We hypothesize that diffusion features extracted from noisy images \\(\\mathbf{x}_t\\) encode the noise \\(\\epsilon\\) in addition to information from the image. We investigate this hypothesis in a very simple setting by examining how well features \\(\\mathrm{feat}(\\epsilon; T)\\) extracted from only the noise \\(\\epsilon\\) approximate the features \\(\\mathrm{feat}(\\mathbf{x}_t = \\sqrt{\\alpha_t} \\mathbf{x}_0 + \\sqrt{1 - \\alpha_t} \\epsilon; t)\\). Using least squares, we fit a single scalar approximation coefficient to obtain the optimal reconstruction. We then quantify how much of the variance of the overall features is explained by this approximation (cf. Figure 3). Even at relatively low timesteps, such as \\(t = 261\\) used by DIFT [62], encoding pure noise explains a substantial fraction of the features' variance. Current diffusion feature methods extract this information jointly with the image information. Our proposed method addresses this issue by eliminating the noise from the feature extraction process." + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "119" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.094, + 0.487, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.264, + 0.485, + 0.349 + ], + "angle": 0, + "content": "Figure 4. Our training setup. We train our model to predict features from a clean input image, while the frozen diffusion model is fed the noisy image. The projection heads project our model's features onto the noisy diffusion model features, given the noising timestep \\( t \\). For downstream tasks, we discard the projection heads and directly use our model's internal representations as features." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.362, + 0.484, + 0.513 + ], + "angle": 0, + "content": "We further analyze the residual and similarly decompose it via the features predicted for the clean image \\(\\mathrm{feat}(x_t = x_0; t = 0)\\). We find that they do not fully explain the remainder of the features either. Instead, a substantial part of the feature variance at medium noise timesteps is timestep dependent and cannot be attributed to components present at \\(t = 0\\) or \\(t = T\\). This matches observations by previous works [62] that found diffusion features at higher timesteps offer better semantics, despite the added noise at the image input." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.524, + 0.466, + 0.54 + ], + "angle": 0, + "content": "3.2. CleanDIFT: Noise-Free Diffusion Features" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.547, + 0.485, + 0.624 + ], + "angle": 0, + "content": "We present CleanDIFT, our method to address the problem of noisy and time-dependent diffusion features. CleanDIFT extracts clean Diffusion FeaTures from a pretrained diffusion backbone through a lightweight fine-tuning process. An overview of our setup is shown in Figure 4." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.645, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Extraction Setup We train our feature extraction model to match the diffusion model's internal representations. We initialize the feature extraction model as a trainable copy of the diffusion model. Crucially, the feature extraction model is given the clean input image, while the diffusion model receives the noisy image and the corresponding timestep as input. Our goal is to obtain a single, noise-free feature map from the feature extraction model that consolidates the information of the diffusion model's timestep-dependent internal representations into a single one. To align our model's representations with the timestep-dependent diffusion model features during training, we introduce pointwise timestep-conditioned feature projection heads. The feature maps predicted by these projection heads are then aligned to the diffusion model's features. For feature extraction at inference time, we usually discard the projection heads and directly use the feature extraction model's" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.168 + ], + "angle": 0, + "content": "internal representations. However, the projection heads can also be used to efficiently obtain feature maps for specific timesteps by reusing the feature extraction model's internal representations and passing them through the projection heads for different \\(t\\) values." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.199, + 0.907, + 0.305 + ], + "angle": 0, + "content": "Training Objective We regard the diffusion model as a family of feature extraction functions \\(\\operatorname{feat}(\\cdot, \\epsilon, t)\\) for timestep \\(t \\in [1,999]\\) and noise \\(\\epsilon \\sim \\mathcal{N}(0, \\mathbf{I})\\). Each of these functions maps an image \\(\\mathbf{x}\\) to a feature vector \\(\\operatorname{feat}(\\mathbf{x}, \\epsilon, t)\\). We aim to consolidate the information provided by all feature extraction functions into a single joint function \\(\\operatorname{feat}_c(\\cdot)\\) with the same dimensionality:" + }, + { + "type": "equation", + "bbox": [ + 0.571, + 0.307, + 0.86, + 0.4 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{c} \\underline {{\\text {S t a b l e D i f f u s i o n}}} \\\\ \\left. \\begin{array}{l} \\operatorname {f e a t} (\\mathbf {x}, \\boldsymbol {\\epsilon}, t = 1) \\\\ \\operatorname {f e a t} (\\mathbf {x}, \\boldsymbol {\\epsilon}, t = 2) \\\\ \\vdots \\\\ \\operatorname {f e a t} (\\mathbf {x}, \\boldsymbol {\\epsilon}, t = 9 9 9) \\end{array} \\right\\} \\xrightarrow [ \\text {i n f o r m a t i o n} ]{\\text {c o n s o l i d a t e}} \\operatorname {f e a t} _ {\\mathrm {c}} (\\mathbf {x}) \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.409, + 0.907, + 0.468 + ], + "angle": 0, + "content": "To align our model's features with the diffusion model's features, we maximize the similarity between the diffusion model's features and the projected features of our feature extraction model:" + }, + { + "type": "equation", + "bbox": [ + 0.521, + 0.47, + 0.89, + 0.535 + ], + "angle": 0, + "content": "\\[\n\\operatorname {f e a t} _ {\\mathrm {c}} (\\mathbf {x}) \\to \\left\\{ \\begin{array}{l} \\operatorname {p r o j} (\\cdot , t = 1) \\longrightarrow \\mathcal {L} \\longleftarrow \\operatorname {f e a t} (\\mathbf {x}, \\epsilon , t = 1) \\\\ \\operatorname {p r o j} (\\cdot , t = 2) \\longrightarrow \\mathcal {L} \\longleftarrow \\operatorname {f e a t} (\\mathbf {x}, \\epsilon , t = 2) \\\\ \\vdots \\\\ \\operatorname {p r o j} (\\cdot , t = 9 9 9) \\longrightarrow \\mathcal {L} \\longleftarrow \\operatorname {f e a t} (\\mathbf {x}, \\epsilon , t = 9 9 9) \\end{array} \\right.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.545, + 0.907, + 0.742 + ], + "angle": 0, + "content": "Specifically, we minimize the negative cosine similarity between the diffusion model's features and our model's features extracted at stages \\( k = \\{1,\\dots,K\\} \\) in the network. Given a clean image \\( \\mathbf{x}_0 \\), the feature extraction model's output for feature map \\( k \\) is denoted as \\( \\mathrm{feat}_{\\mathrm{c}}^{(k)}(\\mathbf{x}_0) \\). Our Clean-DIFT feature map is then adapted by the learned projection heads \\( \\mathrm{proj}^{(k)}(\\mathrm{feat}_{\\mathrm{c}}^{(k)}(\\mathbf{x}_0),t) \\), where \\( \\mathrm{proj}^{(k)}(\\cdot ,\\cdot) \\) is the projection head for feature map \\( k \\). The diffusion model receives the noisy image \\( \\mathbf{x}_t \\) corresponding to the same \\( \\mathbf{x}_0 \\) and timestep \\( t \\). The projection head then learns a timestep-dependent alignment from CleanDIFT features to the diffusion model's features \\( \\mathrm{feat}^{(k)}(\\mathbf{x}_t;t) \\). Putting it all together, our loss function is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.522, + 0.755, + 0.907, + 0.796 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = - \\sum_ {k = 1} ^ {K} \\operatorname {s i m} \\left(\\operatorname {p r o j} ^ {(k)} \\left(\\operatorname {f e a t} _ {\\mathrm {c}} ^ {(k)} \\left(\\mathbf {x} _ {0}\\right); t\\right), \\operatorname {f e a t} ^ {(k)} \\left(\\mathbf {x} _ {t}; t\\right)\\right). \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.909, + 0.903 + ], + "angle": 0, + "content": "For each training image \\(\\mathbf{x}_0\\), we sample \\(I\\) different nois-ing timesteps \\(t_i\\) in a stratified manner, with each timestep \\(t_i \\sim \\mathcal{U}(\\frac{i}{I} T, \\frac{i+1}{I} T)\\), where \\(T\\) is the maximum timestep. By sampling multiple timesteps per image we incentivize the feature extraction model to match the diffusion model's features across the entire noise spectrum." + }, + { + "type": "page_number", + "bbox": [ + 0.488, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "120" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.09, + 0.224, + 0.108 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.117, + 0.484, + 0.284 + ], + "angle": 0, + "content": "We test our hypothesis that the proposed extraction setup enables us to leverage more of the world knowledge inherent in diffusion models compared to existing diffusion feature extraction methods while being task-agnostic and timestep-independent. To that end, we evaluate our features on a wide range of downstream tasks: unsupervised zero-shot semantic correspondence, monocular depth estimation, semantic segmentation, and classification. We compare our features against standard diffusion features, methods that combine diffusion features with additional features, and non-diffusion-based approaches." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.295, + 0.281, + 0.312 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.318, + 0.484, + 0.697 + ], + "angle": 0, + "content": "Implementation Details Following previous works [62, 66, 67], we evaluate our method on a Stable Diffusion (SD) backbone [51]. We apply our method to SD 1.5 and SD 2.1 to enable fair comparisons with existing methods that use either. We fully fine-tune our feature extraction model on image-caption pairs for only 400 steps, taking 30 minutes on a single A100 GPU, which was sufficient for our strong performance and further training did not yield any significant gains. We extract features after the U-Net's middle block and after each of the U-Net's decoder blocks, except the two final blocks. A detailed visualization of where we extract features is provided in Figure 11. This yields a total of \\( K = 11 \\) feature maps that we align between the diffusion model and the feature extraction model. Our point-wise feature projection heads consist of three stacked Feed Forward Networks (FFNs) that are zero-initialized such that initially they act as identity mappings due to their residual connections. Since every aligned feature map has its own projection head, this results in 45M additional trainable parameters for SD 2.1. We study the effect of different projection head architectures in Sec. E. We train using Adam [29] with a batch size of 8 and a learning rate of 2e-6 with a linear warmup. For stratified timestep sampling, we utilize \\( I = 3 \\) stratification bins across all our experiments, i.e. three different noise levels per training image." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Datasets We fine-tune our feature extraction model on a random subset of COYO-700M [5], which is similar to the LAION [56], the dataset that Stable Diffusion 1.5 and 2.1 were trained on originally. That way, we ensure that all performance improvements originate from the feature extraction model consolidating the diffusion model's internal feature representations over time, not from choosing a different dataset that matches the test dataset distribution more closely. The subset selects images with a minimum size of \\(512^2\\). We crop and resize them to match the corresponding input resolution of the underlying diffusion model. We analyze the effect of using different datasets in Sec. G." + }, + { + "type": "image", + "bbox": [ + 0.515, + 0.088, + 0.907, + 0.291 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.301, + 0.907, + 0.356 + ], + "angle": 0, + "content": "Figure 5. Semantic correspondence results using DIFT [62] features with the standard SD 2.1 (\\(t = 261\\)) and our CleanDIFT features. Our clean features show significantly less incorrect matches than the base diffusion model." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.369, + 0.861, + 0.385 + ], + "angle": 0, + "content": "4.2. Unsupervised Semantic Correspondence" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.391, + 0.907, + 0.678 + ], + "angle": 0, + "content": "As many previous diffusion feature methods focus on (unsupervised) semantic correspondence matching [35, 62, 66, 67], we perform an extensive evaluation of our method on this task. Following previous works on semantic correspondence matching [62, 66, 67], we measure our performance in Percentage of Correct Keypoints (PCK). We average PCK directly across all keypoints, not over images. We use \\(\\alpha = 0.1\\) as a threshold and report both PCK values with error margins relative to the image size and to the bounding box size, denoted as \\(\\mathrm{PCK}_{\\mathrm{img}}\\) and \\(\\mathrm{PCK}_{\\mathrm{bbox}}\\) respectively. We evaluate the performance on the test split of the SPair-71k dataset, which consists of approximately \\(12\\mathrm{k}\\) image pairs from 18 categories. Some existing works [62, 66, 67] evaluate on additional datasets but find SPair-71k to be the most challenging and therefore the most informative benchmark. For the text prompt we use \"A photo of a category.\", with category being the corresponding category of the SPair image. We experiment with distilling the text conditioning in Sec. B." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.69, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Results We first compare our extracted features to DIFT [62], an approach that detects semantic correspondences using standard diffusion features. Substituting these with our CleanDIFT features yields a performance increase of 1.79 absolute percentage points for \\(\\mathrm{PCK}_{\\mathrm{img}}\\) and 1.86 percentage points for \\(\\mathrm{PCK}_{\\mathrm{bbox}}\\). Notably, DIFT averages the extracted feature maps across 8 different noise samples. Without this averaging over noise samples, our performance gain is even larger (2.81 PCK@ \\(\\alpha_{\\mathrm{img}}\\) gain). This indicates that our feature extraction model learns more than a mere averaging over the noise in the diffusion model's feature maps (see Figure 5 and Sec. C for examples). We present an extended version of the time-step dependent performance analysis conducted by [62] in Figure 6: We evaluate the" + }, + { + "type": "page_number", + "bbox": [ + 0.488, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "121" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.109, + 0.089, + 0.462, + 0.174 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.185, + 0.483, + 0.269 + ], + "angle": 0, + "content": "Figure 6. Following [62], we evaluate semantic correspondence matching accuracy for different noise levels. Our feature extractor outperforms the standard noisy diffusion features across all timesteps \\(t\\). We additionally demonstrate that simply providing the diffusion model with a clean image and a non-zero timestep does not result in improved performance." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.28, + 0.483, + 0.521 + ], + "angle": 0, + "content": "diffusion model's performance for different timesteps \\(t\\) in two settings. In the first setting, we provide the diffusion model with a noisy input image \\(\\mathbf{x}_t\\) as usual. In the second setting, we demonstrate that feeding the clean image along with a non-zero timestep is not a viable solution to obtain meaningful features: We provide the diffusion model with the clean input image \\(\\mathbf{x}_0\\) for all timesteps \\(t\\). We observe that the model's performance for the clean input image degrades faster and has a lower peak than for the noisy input. This is to be expected, as the diffusion model was trained on noisy images, not clean images. Importantly, our CleanDIFT features are timestep-independent and consistently outperform standard diffusion features, even when the latter are optimized for the best-performing timestep. We further observe that this advantage generalizes effectively to other backbones, such as DiTs [44] (see Tab. 7)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.522, + 0.483, + 0.78 + ], + "angle": 0, + "content": "A Tale of Two Features [66] extends the approach of DIFT by combining diffusion features with DINOv2 [43] features. Again, we replace the standard diffusion features with our CleanDIFT features and observe that the performance gain transfers when combining our features with DINOv2 features. Telling Left from Right [67] further improves upon the results of A Tale of Two Features by introducing a test-time adaptive pose alignment strategy. We observe that the performance gain transfers to this setting as well. To the best of our knowledge, Telling Left from Right combined with our CleanDIFT features sets a new state-of-the-art in unsupervised zero-shot semantic correspondence matching. In summary, replacing standard diffusion features with our CleanDIFT features consistently results in a significant performance improvement across all three methods. We show an overview of the results in Tab. 1 and a more extensive evaluation per category in Sec. B." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.483, + 0.903 + ], + "angle": 0, + "content": "We also investigate the performance of our features in a supervised fine-tuning setting for semantic correspondence matching. Following [35], we train an aggregation network that uses all extracted feature maps and learns to aggregate them into a single task-specific feature map for semantic correspondence matching. In contrast to [35], we do not have to perform costly DDIM inversion [59] to obtain a matching noisy image for every timestep. Instead," + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.089, + 0.905, + 0.25 + ], + "angle": 0, + "content": "
MethodOur FeaturesPCK@α(↑)
αimg=0.1αbbox=0.1
General Approaches
DINOv2+NN--55.6
Diff. Feat.-based Approaches
DIFT [62]X66.5359.57
✓68.32▲1.7961.43▲1.86
A Tale of Two Features [66]X72.3163.73
✓73.35▲1.0464.81▲1.08
Telling Left from Right [67]X77.0768.64
✓78.40▲1.3369.99▲1.35
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.259, + 0.907, + 0.343 + ], + "angle": 0, + "content": "Table 1. Zero-shot unsupervised semantic correspondence matching performance comparison on SPair71k [37]. Our improved features consistently lead to substantial improvements in matching performance. We report PCK on the test split of SPair71k, aggregated per point. Numbers are reproduced, for a discussion and comparison to reported numbers view Tab. 5." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.356, + 0.907, + 0.553 + ], + "angle": 0, + "content": "we directly feed the clean image to our feature extraction model. Therefore, extracting features with our CleanDIFT approach is \\(50\\mathrm{x}\\) faster, since we perform a single denoiser forward pass while [35] performs 50 for the inversion. Our model achieves a \\(\\mathrm{PCK}_{\\mathrm{img}}\\) value of 72.48 vs their 72.75 and a \\(\\mathrm{PCK}_{\\mathrm{bbox}}\\) value of 64.37 vs their 64.53. We observe a slight performance regression compared to their approach, however, at a speedup of \\(50\\times\\). Luo et al. [35] also present a single-step ablation of their full method that only requires a single forward pass which makes it more comparable to ours. We outperform this single-step version by a wide margin of 9.0 percentage points for \\(\\mathrm{PCK}_{\\mathrm{img}}\\) and 9.1 percentage points for \\(\\mathrm{PCK}_{\\mathrm{bbox}}\\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.562, + 0.687, + 0.577 + ], + "angle": 0, + "content": "4.3. Depth Estimation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.584, + 0.907, + 0.903 + ], + "angle": 0, + "content": "We also investigate monocular depth estimation on NYUv2 [40]. Similar to [43], we follow the evaluation protocol from [34]. We use SD 2.1 as the base model and resize the input to the model's native resolution of \\(768^2\\). We extract features from the same location as [62] and obtain a feature map of dimension \\(48^2\\). Unlike [43], we do not up-sample the features and directly apply the linear probe. The probe predicts depth in 256 uniform bins which we combine with a classification loss after a linear normalization following [3]. We train one probe for our CleanDIFT features and one for standard diffusion features at \\(t = 299\\), as that timestep minimizes the error in our settings. Our qualitative results (see Figure 7) show a substantial fidelity gap in the estimated depth maps between the features from the standard SD 2.1 backbone and the features from our feature extraction model. This is reflected in a substantial improvement in quantitative metrics over the baseline as seen in Tab. 2. Lastly, we reuse the probe trained on standard diffusion features and apply it on the CleanDIFT features. While this does not match the performance of the CleanDIFT probe, it still achieves significantly better re" + }, + { + "type": "page_number", + "bbox": [ + 0.488, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "122" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.101, + 0.087, + 0.477, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.278, + 0.485, + 0.348 + ], + "angle": 0, + "content": "Figure 7. Qualitative results for depth estimation using a linear probe on diffusion features on NYUv2 [40]. Our CleanDIFT features enable substantially better depth estimation than standard diffusion features. Note how the CleanDIFT features are far less noisy when compared to the standard diffusion features." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.362, + 0.48, + 0.53 + ], + "angle": 0, + "content": "
MethodBackboneRMSE (↓)
Self-Supervised Methods
OpenCLIP [28]ViT-G/140.541
MAE [20]ViT-H/140.517
DINO [6]ViT-B/80.555
iBOT [68]ViT-L/160.417
DINOv2 [43]ViT-g/140.344
Diffusion Features
SD 2.1 [51]0.469
DIFF- like [62]Ours0.444,0.025
+ Probes from noisy features0.453,0.016
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.541, + 0.483, + 0.611 + ], + "angle": 0, + "content": "Table 2. Monocular Depth Estimation. Following [43], we evaluate metric depth prediction on NYUv2 [40] using a linear probe. Our clean features outperform the noisy features by a significant margin. Probes trained on the noisy features can be reused for the clean features, but incur a smaller performance gain." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.625, + 0.484, + 0.687 + ], + "angle": 0, + "content": "sults when compared to using standard diffusion features. This indicates that our features can be used as a drop-in replacement for the original diffusion features and offer improved performance on downstream applications." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.697, + 0.308, + 0.714 + ], + "angle": 0, + "content": "4.4. Semantic Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.484, + 0.901 + ], + "angle": 0, + "content": "To further investigate the difference between standard noisy diffusion features and our CleanDIFT features, we evaluate on the semantic segmentation task by training linear probes on our CleanDIFT features and on standard diffusion features. We utilize SD2.1 as the diffusion backbone and extract features at the same location as [62]. This procedure yields feature maps of size \\(48^2\\). We train our linear probe on the \\(48^2\\) feature maps and upscale the obtained segmentation masks using nearest neighbor upsampling. We train and evaluate on the PASCAL VOC dataset [16]. Following common practice [39], we use mean Intersection over Union (mIoU) as the evaluation metric. Qualitative re" + }, + { + "type": "image", + "bbox": [ + 0.521, + 0.087, + 0.905, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.259, + 0.908, + 0.33 + ], + "angle": 0, + "content": "Figure 8. Qualitative results for semantic segmentation from diffusion features on Pascal VOC [16]. Standard SD features use \\( t = 100 \\) as the timestep, which we found to perform best quantitatively (cf. Figure 9). Note how the CleanDIFT segmentation maps are far less noisy than those of the standard diffusion features." + }, + { + "type": "image", + "bbox": [ + 0.554, + 0.343, + 0.864, + 0.43 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.44, + 0.907, + 0.525 + ], + "angle": 0, + "content": "Figure 9. Performance on semantic segmentation using linear probes. Our clean features outperform the noisy diffusion features for the best noising timestep \\( t \\). Semantic segmentation performance of a standard diffusion model heavily depends on the used noising timestep. Unlike for semantic correspondence matching, the optimal \\( t \\) value appears to be around \\( t = 100 \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.537, + 0.909, + 0.718 + ], + "angle": 0, + "content": "sults are shown in Figure 8. Using our features, we observe significantly less noisy segmentations than with standard diffusion features. We show a quantitative comparison of our CleanDIFT feature's performance against standard diffusion features across timesteps in Figure 9. Notably, the optimal timestep appears to be around \\(t = 100\\), in contrast to the optimal timestep for semantic correspondences, which [62] found to be \\(t = 261\\). This highlights the need for tuning a timestep individually per downstream task. Our method both alleviates the need for such a timestep tuning and outperforms the standard diffusion features for the optimal timestep." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.728, + 0.655, + 0.743 + ], + "angle": 0, + "content": "4.5. Classification" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.909, + 0.903 + ], + "angle": 0, + "content": "To assess the impact of our method on non-spatial tasks, we evaluate classification performance using pooled features. Pooling mitigates the influence of localized noise, so we anticipate classification performance to remain on par with standard diffusion features unless our setup introduces detrimental effects. We perform k-Nearest Neighbor (kNN) classification with \\( k = 10 \\) on ImageNet1k [11], using SD 1.5 as the diffusion backbone. We sweep across feature maps and timesteps \\( t \\) for the base model, with results presented in Figure 10. Our analysis shows that the fea" + }, + { + "type": "page_number", + "bbox": [ + 0.488, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "123" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.102, + 0.089, + 0.468, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.289, + 0.483, + 0.361 + ], + "angle": 0, + "content": "Figure 10. Classification performance on ImageNet1k [11], using kNN classifier with \\( k = 10 \\) and cosine similarity as the distance metric. We sweep over different timesteps and feature maps. We find that the feature map with the lowest spatial resolution (feature map #0) yields the highest classification accuracy." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.372, + 0.483, + 0.448 + ], + "angle": 0, + "content": "ture map with the lowest spatial resolution, i.e., feature map #0 (see Figure 11), achieves the highest classification accuracy. Furthermore, the optimal timestep \\( t \\) for the base model varies between feature maps. For the best-performing feature map, \\( t = 100 \\) yields the highest classification accuracy." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.448, + 0.484, + 0.509 + ], + "angle": 0, + "content": "Importantly, CleanDIFT features slightly outperform the standard diffusion features even when using an optimal timestep \\( t \\) for the base model showing that it does not introduce any detrimental effects." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.52, + 0.255, + 0.535 + ], + "angle": 0, + "content": "4.6. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.542, + 0.484, + 0.603 + ], + "angle": 0, + "content": "For simplicity, we perform our ablation studies using DIFT [62] and evaluate the performance for unsupervised zero-shot semantic correspondence matching on a subset of the SPair71k [37] test split." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.616, + 0.483, + 0.812 + ], + "angle": 0, + "content": "Training Objective During training, we maximize the cosine similarity between projected outputs of our feature extraction model and standard diffusion features to align them. To investigate the influence of the employed similarity metric, we compare feature extraction models trained on three different alignment objectives commonly used in similar contexts: mean absolute error \\((L_{1})\\), mean squared error \\((L_{2})\\), and cosine similarity. Quantitative results are provided in Tab. 3. While all objectives result in feature extraction models that outperform standard diffusion features, cosine similarity consistently performs the best across the alignment objectives by a significant margin, confirming our choice of similarity metric." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Projection Heads We investigate the influence of our proposed projection heads that are used to project Clean-DIFT features onto standard diffusion features. The alignment of feature extraction model and diffusion model is determined by the utilized similarity metric and the pro" + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.089, + 0.905, + 0.27 + ], + "angle": 0, + "content": "
ObjectiveProjection HeadsPCK@α(↑)
αimg=0.1αbbox=0.1
Cosine Sim.✓68.3261.43
×68.1661.29
L2✓66.2359.13
×66.4959.43
L1✓66.9160.00
×66.8759.91
SD 2.1-63.4155.92
" + }, + { + "type": "table_caption", + "bbox": [ + 0.512, + 0.28, + 0.905, + 0.322 + ], + "angle": 0, + "content": "Table 3. Ablation Study Results. We evaluate the feature extraction models' performance for zero-shot semantic correspondence matching on the SPair71k test split. PCK is aggregated per point." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.334, + 0.907, + 0.546 + ], + "angle": 0, + "content": "jection heads. Therefore, we evaluate our feature extraction model's performance with and without the projection heads in combination with all three similarity metrics. An overview of the comparison is given in Tab. 3. In our main configuration that uses cosine similarity, using the projection heads yields slight performance improvements of 0.24 percentage points for \\(\\mathrm{PCK}_{\\mathrm{img}}\\) and 0.06 percentage points for \\(\\mathrm{PCK}_{\\mathrm{bbox}}\\) compared to fine-tuning without the projection heads. As the projection heads are typically not used for inference, they add computational overhead only during the lightweight fine-tuning. Therefore, we argue that it is worthwhile to include them and leverage the small performance gain. Additionally, they can be reused to efficiently obtain feature maps for specific timesteps." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.559, + 0.634, + 0.574 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.584, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In this paper, we introduced CleanDIFT, a novel approach for extracting diffusion features. CleanDIFT produces noise-free, timestep-independent, general-purpose diffusion features by consolidating timestep-dependent representations from a pre-trained diffusion backbone into a unified feature representation. We achieve this alignment between our feature extraction model and the pre-trained diffusion backbone through a lightweight fine-tuning procedure that takes approximately 30 minutes on a single A100 GPU. Operating directly on clean images, our method eliminates the information loss associated with adding noise to input images. Furthermore, CleanDIFT removes the requirement for tuning timesteps for each downstream task and avoids the computational overhead of ensembling over noise levels or timesteps. Instead, our method efficiently extracts features with just a single forward pass at inference time, substantially reducing inference costs compared to methods relying on ensembling or inversion. Extensive evaluations of CleanDIFT across diverse downstream tasks demonstrate significant performance improvements over conventional diffusion features." + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "124" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.251, + 0.108 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.116, + 0.484, + 0.313 + ], + "angle": 0, + "content": "This project has been supported by the German Federal Ministry for Economic Affairs and Climate Action within the project \"NXT GEN AI METHODS - Generative Methoden für Perzeption, Prädiktion und Planung\", the project \"GeniusRobot\" (01IS24083), funded by the Federal Ministry of Education and Research (BMBF), the bidt project KLIMA-MEMES, and the German Research Foundation (DFG) project 421703927. The authors gratefully acknowledge the Gauss Center for Supercomputing for providing compute through the NIC on JUWELS at JSC and the HPC resources supplied by the Erlangen National High Performance Computing Center (NHR@FAU funded by DFG project 440719683) under the NHR project JA-22883." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.314, + 0.484, + 0.345 + ], + "angle": 0, + "content": "Further, we would like to thank Owen Vincent for continuous technical support." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.36, + 0.188, + 0.377 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.385, + 0.484, + 0.455 + ], + "angle": 0, + "content": "[1] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Ji-aming Song, Qinsheng Zhang, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, et al. ediff-i: Text-to-image diffusion models with an ensemble of expert denoisers. arXiv preprint arXiv:2211.01324, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.457, + 0.484, + 0.513 + ], + "angle": 0, + "content": "[2] Dmitry Baranchuk, Andrey Voynov, Ivan Rubachev, Valentin Khrulkov, and Artem Babenko. Label-efficient semantic segmentation with diffusion models. In International Conference on Learning Representations, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.515, + 0.484, + 0.57 + ], + "angle": 0, + "content": "[3] Shariq Farooq Bhat, Ibrahim Alhashim, and Peter Wonka. Adabins: Depth estimation using adaptive bins. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4009-4018, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.572, + 0.484, + 0.613 + ], + "angle": 0, + "content": "[4] Giulio Biroli, Tony Bonnaire, Valentin De Bortoli, and Marc Mézard. Dynamical regimes of diffusion models. arXiv preprint arXiv:2402.18491, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.616, + 0.484, + 0.671 + ], + "angle": 0, + "content": "[5] Minwoo Byeon, Beomhee Park, Haecheon Kim, Sungjun Lee, Woonhyuk Baek, and Saehoon Kim. Coyo-700m: Image-text pair dataset. https://github.com/kakaobrain/coyo-dataset, 2022.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.673, + 0.484, + 0.743 + ], + "angle": 0, + "content": "[6] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.745, + 0.484, + 0.828 + ], + "angle": 0, + "content": "[7] Junsong Chen, Jincheng YU, Chongjian GE, Lewei Yao, Enze Xie, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart-\\(\\mathbb{S}\\backslash \\mathbb{alpha}\\): Fast training of diffusion transformer for photorealistic text-to-image synthesis. In The Twelfth International Conference on Learning Representations, 2024. 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.83, + 0.484, + 0.885 + ], + "angle": 0, + "content": "[8] Shoufa Chen, Peize Sun, Yibing Song, and Ping Luo. Diffusionnet: Diffusion model for object detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 19830-19843, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.888, + 0.484, + 0.902 + ], + "angle": 0, + "content": "[9] Xinlei Chen, Zhuang Liu, Saining Xie, and Kaiming He. De" + }, + { + "type": "list", + "bbox": [ + 0.1, + 0.385, + 0.484, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.907, + 0.121 + ], + "angle": 0, + "content": "constructing denoising diffusion models for self-supervised learning. arXiv preprint arXiv:2401.14404, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.907, + 0.176 + ], + "angle": 0, + "content": "[10] Timothee Darcet, Maxime Oquab, Julien Mairal, and Piotr Bojanowski. Vision transformers need registers. In The Twelfth International Conference on Learning Representations, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.178, + 0.906, + 0.232 + ], + "angle": 0, + "content": "[11] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 248–255, 2009. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.234, + 0.906, + 0.274 + ], + "angle": 0, + "content": "[12] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.275, + 0.906, + 0.344 + ], + "angle": 0, + "content": "[13] Patrick Esser, Johnathan Chiu, Parmida Atighehchian, Jonathan Granskog, and Anastasis Germanidis. Structure and content-guided video synthesis with diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7346-7356, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.345, + 0.906, + 0.426 + ], + "angle": 0, + "content": "[14] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In *Forty-first International Conference on Machine Learning*, 2024. 2, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.428, + 0.906, + 0.481 + ], + "angle": 0, + "content": "[15] Zach Evans, CJ Carr, Josiah Taylor, Scott H Hawley, and Jordi Pons. Fast timing-conditioned latent audio diffusion. In *Forty-first International Conference on Machine Learning*, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.484, + 0.906, + 0.537 + ], + "angle": 0, + "content": "[16] M. Everingham, L. Van Gool, C. K. I. Williams, J. Winn, and A. Zisserman. The Pascal visual object classes (voc) challenge. International Journal of Computer Vision, 88(2): 303-338, 2010. 7, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.539, + 0.906, + 0.593 + ], + "angle": 0, + "content": "[17] Michael Fuest, Pingchuan Ma, Ming Gui, Johannes S. Fischer, Vincent Tao Hu, and Björn Ommer. Diffusion models and representation learning: A survey. CoRR, abs/2407.00783, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.595, + 0.906, + 0.636 + ], + "angle": 0, + "content": "[18] Frank Fundel, Johannes Schusterbauer, Vincent Tao Hu, and Björn Ommer. Distillation of diffusion features for semantic correspondence. WACV, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.638, + 0.906, + 0.691 + ], + "angle": 0, + "content": "[19] Spyros Gidaris, Praveer Singh, and Nikos Komodakis. Unsupervised representation learning by predicting image rotations. In International Conference on Learning Representations, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.693, + 0.906, + 0.761 + ], + "angle": 0, + "content": "[20] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dólár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.763, + 0.906, + 0.83 + ], + "angle": 0, + "content": "[21] Eric Hedlin, Gopal Sharma, Shweta Mahajan, Hossam Isack, Abhishek Kar, Andrea Tagliasacchi, and Kwang Moo Yi. Unsupervised semantic correspondence using stable diffusion. Advances in Neural Information Processing Systems, 36, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.832, + 0.906, + 0.873 + ], + "angle": 0, + "content": "[22] Geoffrey Hinton, Oriol Vinyls, and Jeffrey Dean. Distilling the knowledge in a neural network. In NIPS Deep Learning and Representation Learning Workshop, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.906, + 0.902 + ], + "angle": 0, + "content": "[23] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.488, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "125" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.135 + ], + "angle": 0, + "content": "[24] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.482, + 0.19 + ], + "angle": 0, + "content": "[25] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. Advances in Neural Information Processing Systems, 35:8633-8646, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.192, + 0.482, + 0.247 + ], + "angle": 0, + "content": "[26] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. *ICLR*, 1(2):3, 2022. 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.25, + 0.482, + 0.332 + ], + "angle": 0, + "content": "[27] Drew A Hudson, Daniel Zoran, Mateusz Malinowski, Andrew K Lampinen, Andrew Jaegle, James L McClelland, Loic Matthey, Felix Hill, and Alexander Lerchner. Soda: Bottleneck diffusion models for representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23115-23127, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.334, + 0.482, + 0.403 + ], + "angle": 0, + "content": "[28] Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. Openclip, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.405, + 0.482, + 0.459 + ], + "angle": 0, + "content": "[29] Diederik P. Kingma and Jimmy Lei Ba. Adam: A method for stochastic optimization. In Proceedings of the 3rd International Conference on Learning Representations (ICLR), 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.462, + 0.482, + 0.517 + ], + "angle": 0, + "content": "[30] Zhifeng Kong, Wei Ping, Jiaji Huang, Kexin Zhao, and Bryan Catanzaro. Diffwave: A versatile diffusion model for audio synthesis. In International Conference on Learning Representations, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.519, + 0.482, + 0.587 + ], + "angle": 0, + "content": "[31] Alexander C Li, Mihir Prabhudesai, Shivam Duggal, Ellis Brown, and Deepak Pathak. Your diffusion model is secretly a zero-shot classifier. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2206-2217, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.59, + 0.482, + 0.672 + ], + "angle": 0, + "content": "[32] Daiqing Li, Huan Ling, Amlan Kar, David Acuna, Seung Wook Kim, Karsten Kreis, Antonio Torralba, and Sanja Fidler. Dreamteacher: Pretraining image backbones with deep generative models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16698-16708, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.675, + 0.482, + 0.743 + ], + "angle": 0, + "content": "[33] Xinghui Li, Jingyi Lu, Kai Han, and Victor Adrian Prisacariu. Sd4match: Learning to prompt stable diffusion model for semantic matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 27558-27568, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.746, + 0.482, + 0.8 + ], + "angle": 0, + "content": "[34] Zhenyu Li, Xuyang Wang, Xianming Liu, and Junjun Jiang. Binsformer: Revisiting adaptive bins for monocular depth estimation. IEEE Transactions on Image Processing, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.803, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[35] Grace Luo, Lisa Dunlap, Dong Huk Park, Aleksander Holynski, and Trevor Darrell. Diffusion hyperfeatures: Searching through time and space for semantic correspondence. In Advances in Neural Information Processing Systems, pages 47500-47510. Curran Associates, Inc., 2023. 3, 5, 6, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[36] Chenlin Meng, Robin Rombach, Ruiqi Gao, Diederik Kingma, Stefano Ermon, Jonathan Ho, and Tim Salimans." + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.092, + 0.905, + 0.135 + ], + "angle": 0, + "content": "On distillation of guided diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14297-14306, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.178 + ], + "angle": 0, + "content": "[37] Juhong Min, Jongmin Lee, Jean Ponce, and Minsu Cho. Spair-71k: A large-scale benchmark for semantic correspondence. arXiv preprint arXiv:1908.10543, 2019. 6, 8, 12, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.905, + 0.234 + ], + "angle": 0, + "content": "[38] Ishan Misra and Laurens van der Maaten. Self-supervised learning of pretext-invariant representations. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6707-6717, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.236, + 0.905, + 0.291 + ], + "angle": 0, + "content": "[39] Koichi Namekata, Amirmojtaba Sabour, Sanja Fidler, and Seung Wook Kim. Emerdiff: Emerging pixel-level semantic knowledge in diffusion models. In The Twelfth International Conference on Learning Representations, 2023. 1, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.293, + 0.905, + 0.334 + ], + "angle": 0, + "content": "[40] Pushmeet Kohli Nathan Silberman, Derek Hoiem and Rob Fergus. Indoor segmentation and support inference from rgbd images. In ECCV, 2012. 6, 7, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.336, + 0.905, + 0.404 + ], + "angle": 0, + "content": "[41] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.407, + 0.905, + 0.461 + ], + "angle": 0, + "content": "[42] Mehdi Noroozi and Paolo Favaro. Unsupervised learning of visual representations by solving jigsaw puzzles. In European conference on computer vision, pages 69-84. Springer, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.463, + 0.905, + 0.602 + ], + "angle": 0, + "content": "[43] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel HAZIZA, Francisco Massa, Alaaeldin El-Nouby, Mido Assran, Nicolas Ballas, Wojciech Galuba, Russell Howes, Po-Yao Huang, Shang-Wen Li, Ishan Misra, Michael Rabbat, Vasu Sharma, Gabriel Synnaeve, Hu Xu, Herve Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. DINOv2: Learning robust visual features without supervision. Transactions on Machine Learning Research, 2024. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.604, + 0.905, + 0.657 + ], + "angle": 0, + "content": "[44] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4195-4205, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.66, + 0.905, + 0.716 + ], + "angle": 0, + "content": "[45] Ethan Perez, Florian Strub, Harm De Vries, Vincent Dumoulin, and Aaron Courville. Film: Visual reasoning with a general conditioning layer. In Proceedings of the AAAI conference on artificial intelligence, 2018. 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.718, + 0.905, + 0.799 + ], + "angle": 0, + "content": "[46] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. In The Twelfth International Conference on Learning Representations, 2024. 2, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.803, + 0.905, + 0.87 + ], + "angle": 0, + "content": "[47] Adam Polyak, Amit Zohar, Andrew Brown, Andros Tjandra, Animesh Sinha, Ann Lee, Apoorv Vyas, Bowen Shi, Chih-Yao Ma, Ching-Yao Chuang, et al. Movie gen: A cast of media foundation models. arXiv preprint arXiv:2410.13720, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[48] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry," + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.488, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "126" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.483, + 0.147 + ], + "angle": 0, + "content": "Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.483, + 0.189 + ], + "angle": 0, + "content": "[49] Prajit Ramachandran, Barret Zoph, and Quoc V Le. Searching for activation functions. arXiv preprint arXiv:1710.05941, 2017. 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.191, + 0.482, + 0.232 + ], + "angle": 0, + "content": "[50] Severi Rissanen, Markus Heinonen, and Arno Solin. Generative modelling with inverse heat dissipation. arXiv preprint arXiv:2206.13397, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.233, + 0.483, + 0.302 + ], + "angle": 0, + "content": "[51] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 2, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.302, + 0.483, + 0.372 + ], + "angle": 0, + "content": "[52] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention - MICCAI 2015, pages 234-241, Cham, 2015. Springer International Publishing. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.372, + 0.483, + 0.455 + ], + "angle": 0, + "content": "[53] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural information processing systems, 35:36479-36494, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.455, + 0.483, + 0.496 + ], + "angle": 0, + "content": "[54] Tim Salimans and Jonathan Ho. Progressive distillation for fast sampling of diffusion models. arXiv preprint arXiv:2202.00512, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.497, + 0.483, + 0.552 + ], + "angle": 0, + "content": "[55] Axel Sauer, Dominik Lorenz, Andreas Blattmann, and Robin Rombach. Adversarial diffusion distillation. In European Conference on Computer Vision, pages 87-103. Springer, 2024. 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.553, + 0.483, + 0.636 + ], + "angle": 0, + "content": "[56] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278-25294, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.636, + 0.483, + 0.664 + ], + "angle": 0, + "content": "[57] Noam Shazeer. Glu variants improve transformer. arXiv preprint arXiv:2002.05202, 2020. 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.664, + 0.483, + 0.734 + ], + "angle": 0, + "content": "[58] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. In The Eleventh International Conference on Learning Representations, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.735, + 0.483, + 0.775 + ], + "angle": 0, + "content": "[59] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In International Conference on Learning Representations, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.776, + 0.483, + 0.817 + ], + "angle": 0, + "content": "[60] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. Advances in neural information processing systems, 32, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.818, + 0.483, + 0.873 + ], + "angle": 0, + "content": "[61] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.483, + 0.902 + ], + "angle": 0, + "content": "[62] Luming Tang, Menglin Jia, Qianqian Wang, Cheng Perng Phoo, and Bharath Hariharan. Emergent correspondence" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "from image diffusion. In Advances in Neural Information Processing Systems, pages 1363-1389. Curran Associates, Inc., 2023. 1, 2, 3, 4, 5, 6, 7, 8, 12, 13, 14, 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.219 + ], + "angle": 0, + "content": "[63] Weijia Wu, Yuzhong Zhao, Mike Zheng Shou, Hong Zhou, and Chunhua Shen. Diffumask: Synthesizing images with pixel-level annotations for semantic segmentation using diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1206-1217, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.22, + 0.905, + 0.302 + ], + "angle": 0, + "content": "[64] Jiarui Xu, Sifei Liu, Arash Vahdat, Wonmin Byeon, Xiaolong Wang, and Shalini De Mello. Open-vocabulary panoptic segmentation with text-to-image diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2955-2966, 2023. 1, 3, 13, 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.304, + 0.905, + 0.346 + ], + "angle": 0, + "content": "[65] Biao Zhang and Rico Sennrich. Root mean square layer normalization. Advances in Neural Information Processing Systems, 32, 2019. 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.346, + 0.905, + 0.443 + ], + "angle": 0, + "content": "[66] Junyi Zhang, Charles Herrmann, Junhwa Hur, Luisa Polania Cabrera, Varun Jampani, Deqing Sun, and Ming-Hsuan Yang. A tale of two features: Stable diffusion complements dino for zero-shot semantic correspondence. In Advances in Neural Information Processing Systems, pages 45533-45547. Curran Associates, Inc., 2023. 1, 2, 3, 5, 6, 12, 13, 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.445, + 0.905, + 0.528 + ], + "angle": 0, + "content": "[67] Junyi Zhang, Charles Herrmann, Junhwa Hur, Eric Chen, Varun Jampani, Deqing Sun, and Ming-Hsuan Yang. Telling left from right: Identifying geometry-aware semantic correspondence. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3076-3085, 2024. 1, 2, 3, 5, 6, 12, 13, 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.529, + 0.905, + 0.585 + ], + "angle": 0, + "content": "[68] Jinghao Zhou, Chen Wei, Huiyu Wang, Wei Shen, Cihang Xie, Alan Yuille, and Tao Kong. Image BERT pre-training with online tokenizer. In International Conference on Learning Representations, 2022. 7" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.585 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.488, + 0.946, + 0.511, + 0.956 + ], + "angle": 0, + "content": "127" + } + ] +] \ No newline at end of file diff --git a/2025/CleanDIFT_ Diffusion Features without Noise/17f652ff-055a-4bf1-a266-9144370345a9_origin.pdf b/2025/CleanDIFT_ Diffusion Features without Noise/17f652ff-055a-4bf1-a266-9144370345a9_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..177fdf708c812f7647f4f54b9621488a82d7dc3a --- /dev/null +++ b/2025/CleanDIFT_ Diffusion Features without Noise/17f652ff-055a-4bf1-a266-9144370345a9_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:406bf3f9a770d1ca2f42ad0d673b9e71713edf3041e6631a2527026ef88e198e +size 6845552 diff --git a/2025/CleanDIFT_ Diffusion Features without Noise/full.md b/2025/CleanDIFT_ Diffusion Features without Noise/full.md new file mode 100644 index 0000000000000000000000000000000000000000..aa429e86f359b4bb7b0739c4e7f01d8fd155c9f6 --- /dev/null +++ b/2025/CleanDIFT_ Diffusion Features without Noise/full.md @@ -0,0 +1,324 @@ +# $\checkmark$ CleanDIFT: Diffusion Features without Noise + +# Nick Stracke* + +# Stefan Andreas Baumann* + +# Kolja Bauer* + +# Frank Fundel + +# Björn Ommer + +CompVis @ LMU Munich, MCML + +{nick.stracke,b.ommer}@lmu.de + +compvis.github.io/cleandift + +# Normal Diffusion Features 30 min Finetuning + +# CleanDIFT Features + +noisy images noisy features + +![](images/0e4b63c0ac6e7712c160ff1e5389229e3308f90165be123a628a1cc446623839.jpg) +$t = 300$ + +![](images/48084df88f96a2606168a20000ccdfddaa3939408eb4a2567f1860df4377d05c.jpg) + +![](images/70715d719fbd02a6850f60f5dc2a5c015492a320ce7444325f7fe1b36d3362f0.jpg) + +![](images/8668d2c527e376f5ca9b196f7f0446b78ecb983ce5f96f05ee26c5bbfd5d1c61.jpg) +timestep tuning required $X$ + +![](images/1d4b2a24cd398a9dc718f4fc53e2f2b24123651a37619ae1a68d30cb93d8c0ff.jpg) +Figure 1. Our proposed CleanDIFT feature extraction method yields noise-free, timestep-independent, general-purpose features that significantly outperform standard diffusion features. CleanDIFT operates on clean images, while extracting diffusion features with existing approaches requires adding noise to an image before passing it through the model. Adding noise reduces the information present in the image and requires tuning a timestep per downstream task. + +![](images/e13d2c508add19b770a90f41496be17677f7b19da282ff4cac391dd9ed3a657f.jpg) + +clean images clean features + +![](images/1f19a52e101070d30fa1990b76ad79fdb74c69c081b4dbcee6d73ccb04bff8b2.jpg) +no timestep tuning required $\checkmark$ + +![](images/dcd916cdd0eae739e33f27d488bdd8f878eb87ad0e03ca7f0ad8be73d8191639.jpg) + +![](images/8a8e4a04cac091f870cb498ebd96098ebed58a18e19e928a854153fe7ee50abf.jpg) + +![](images/168d3f58973fdd6f652a8b910e50df8424c684385c73890753205aa80b492256.jpg) + +![](images/32fa88c7b0404cc885204005645bc2fe76b1fabf6b70ae7765c4500e5bb8e2b8.jpg) + +![](images/0f923256c1d2d1ea88e5f63c61e591c1f564b805aec913466b509a6ce0c40d97.jpg) + +# Abstract + +Internal features from large-scale pre-trained diffusion models have recently been established as powerful semantic descriptors for a wide range of downstream tasks. Works that use these features generally need to add noise to images before passing them through the model to obtain the semantic features, as the models do not offer the most useful features when given images with little to no noise. We show that this noise has a critical impact on the usefulness of these features that cannot be remedied by ensembling with different random noises. We address this issue by introducing a lightweight, unsupervised fine-tuning method that enables diffusion backbones to provide high-quality, noise-free semantic features. We show that these features readily outperform previous diffusion features by a wide margin in a wide variety of extraction setups and downstream tasks, offering better performance than even ensemble-based methods at a fraction of the cost. + +# 1. Introduction + +Learning meaningful visual representations that capture a vast amount of world knowledge remains a key problem in the field of computer vision. Diffusion models can be trained at scale in a self-supervised manner and have rapidly advanced the state of the art in image [12, 41, 51] and video generation [13, 25, 58], making them a good candidate to learn visual representations. Many early works have already achieved impressive results using internal features from large-scale pretrained diffusion models for a wide variety of tasks, such as semantic correspondence detection [62, 66, 67], semantic segmentation [2, 39, 63], panoptic segmentation [64], object detection [8], and classification [31]. However, the optimal approach to extract this world knowledge from a diffusion model remains uncertain. + +To understand why that is the case, we take a look at how diffusion models are trained: a varying amount of noise is added to a clean input image (forward process) and the model is tasked to remove the noise from the image (backward process). The amount of added noise is dependent on the diffusion timestep. As a result, the model learns to operate on noisy images and also becomes dependent on the noise timestep as different noise levels require the model to + +perform different tasks [1, 4]. Since noisy images inherently contain less information than clean images (cf. Figure 2), we hypothesize that this harms the internal feature representation of diffusion models [9] and, thus, the extractable world knowledge. Furthermore, the timestep acts as a hyperparameter that influences the internal feature representation and needs to be picked independently for every downstream application (cf. Figure 14). + +We propose a novel feature extraction method that (1) eliminates the need to destroy information by adding noise to the input; and (2) produces timestep-independent generic diffusion features useful for a wide range of down-stream tasks, alleviating the need to tune a noisig timestep per down-stream task. We show how to adapt an off-the-shelf large-scale pre-trained diffusion backbone to provide these features at minimal cost (approximately 30 minutes of fine-tuning on a single A100 GPU) and demonstrate improved performance across a wide range of downstream tasks. + +We achieve this by viewing a diffusion model as a family of $T$ feature extractors that operate on images with different noise levels and provide features with different characteristics. We consolidate all $T$ feature extraction functions in our feature extractor by aligning their internal representations. Specifically, we initialize our feature extractor as a trainable copy of the diffusion model; fine-tune it with clean images and no timestep input; and align its features with all $T$ time-dependent feature extractors of the diffusion model. + +We evaluate our improved features across a wide variety of downstream tasks, such as semantic correspondence matching, monocular depth estimation, semantic segmentation, and classification, and find that they consistently improve upon approaches based on standard diffusion features. These improvements are most evident for dense visual tasks such as semantic correspondence matching, where our features show substantial performance gains across a wide variety of setups [62, 66, 67] and set a new state-of-the-art for unsupervised semantic correspondence matching. Additionally, our proposed method eliminates the need for noise or timestep ensembling [62], offering substantial speed gains (e.g., $8 \times$ over DIFT [62]), on top of improved quality. Our method is generic and integrates easily with established methods, such as fusing diffusion and DINOv2 features [66, 67]. + +Our main contributions are as follows: + +1. We propose CleanDIFT, a finetuning approach for diffusion models that enables them to operate on clean images and makes the inherent world knowledge of these models more accessible. +2. We show how to consolidate information from all diffusion timesteps into a single feature prediction, removing the need for task-specific timestep tuning. +3. We demonstrate significant performance gains of our diffusion feature extraction technique across a wide range + +![](images/fd330a828668af62cc1322a621550038d9f262fe12bb7f536c044c42730d8b13.jpg) +(a) Reconstruction without Noise +Figure 2. Deterioration of Diffusion Features. As current methods need to pass noisy images to the model to obtain useful features, they significantly reduce the information available. We alleviate this problem by obtaining useful features without noise, improving the performance of downstream tasks. + +![](images/12eea81cc5e551bfb499eda8e08eaaa662e3618dbd69cab2eeafb7d8e11d3116.jpg) +(b) Diffusion Model Reconstruction with added noise $(t = 261$ [62]) + +of down-stream tasks, notably surpassing the current state of the art in zero-shot unsupervised semantic correspondence detection. We further demonstrate the generality of our enhanced features by showing that these performance gains transfer to advanced methods that fuse diffusion features or operate in a supervised setting. + +4. Our proposed approach is significantly more efficient than previous methods that tried to address this problem by noise ensembling or supervised training. + +# 2. Related Work + +Self-Supervised Representation Learning Features from large, pre-trained foundation models have been shown to yield competitive performance to supervised models for a variety of downstream tasks, both in zero-shot and fine-tuning settings [20, 43, 48]. These foundation models are trained on different pre-text tasks like inpainting [20], predicting transformations [19], patch reordering [38, 42], and discriminative tasks [6, 43]. DINOv2 [43] uses a discriminative objective combined with self-distillation to learn general-purpose visual features that have proven useful for a variety of downstream tasks [10]. CLIP [48] learns such features by employing a contrastive objective on text-image pairs. Masked Autoencoders [20] (MAEs) are trained to reconstruct masked out patches of the input, also resulting in general-purpose visual features. + +Diffusion Models as Self-Supervised Learners Diffusion models [24, 60, 61] are generative models that have defined the state-of-the-art in image generation [12, 14, 41, 46, 51, 53], video generation [13, 47], and audio generation [15, 30] in recent years. Their primary purpose is to generate high-quality samples (images, videos, etc.). However, generation can also be interpreted as a pretext task for + +learning expressive features, since the model has to build up comprehensive world knowledge in order to generate plausible samples [9, 17, 27, 32, 62]. Features from diffusion models (typically referred to as diffusion features) are obtained by passing a noised image through the diffusion model and extracting intermediate feature representations. They have been shown to be useful for a variety of tasks such as finding semantic correspondences [21, 35, 62], semantic and panoptic segmentation [2, 39, 63, 64], classification [31], and object detection [8]. + +For semantic correspondence matching, features are leveraged to identify semantically matching regions across images. Existing approaches utilize diffusion features either in a zero-shot setting [21, 62] or fine-tune them for the semantic correspondence task [33, 35, 67]. Some zero-shot approaches do not fine-tune on semantic correspondence but still require tuning a prompt to activate attention maps at the query location of the correspondence [21]. In contrast, our approach aims to provide universal features usable for various down-stream tasks in a true zero-shot manner. + +Further, diffusion features have been shown to complement features from other self-supervised learning methods such as DINOv2 [18, 43, 66]. DINOv2 features provide sparse but accurate semantic information, while Diffusion features provide dense spatial information albeit with sometimes inaccurate semantics. State-of-the-art approaches for semantic correspondence detection exploit this and fuse features [66, 67]. Compared to DINOv2 features, diffusion features yield smoother, spatially more coherent correspondences [66]. + +Distillation for Diffusion Models Knowledge Distillation [22] is a technique used to distill knowledge from a teacher model into a student model. In the context of diffusion models, this is typically applied either to reduce the required denoising steps [54] or to distill a classifier-free guided [23] model into one without CFG [36]. While our approach is inspired by distillation, we consolidate features from $T$ different models, with $T$ being the number of discrete diffusion timesteps, because the features are different at every timestep. + +# 3. Method + +# 3.1. Preliminaries + +Diffusion Models Diffusion models are trained to predict a clean image $\mathbf{x}_0$ given a noisy image $\mathbf{x}_t$ , either explicitly or implicitly. The noisy image $\mathbf{x}_t$ is a weighted sum with random Gaussian noise $\mathbf{x}_t = \sqrt{\alpha_t}\mathbf{x}_0 + \sqrt{1 - \alpha_t}\epsilon$ with timestep-dependent coefficients $\alpha_{t}$ and noise $\epsilon \sim \mathcal{N}(0,\mathbf{I})$ . $t\in [0,T]$ denotes the time step of the diffusion process, with $t = 0$ corresponding to the clean image and $t = T$ corresponding to pure noise. + +![](images/e5341a1d6c39dabfda046b3f20d68749e7446c43b0d651440d9de50170ee264d.jpg) +Figure 3. Fraction of variance of diffusion features explained by 1) encoding the clean image at $t = 0$ (no additive noise), and 2) encoding just the added noise $\epsilon$ at $t = 999$ . Even at relatively low timesteps such as $t = 261$ as used by DIFT [62], a substantial part of the features directly depends only on the added noise. + +Intuitively, the model faces different objectives for different noise levels [1, 4]: for very high noise, there is little information in the input, and the model is first generating the coarse structure of the image [50]. At lower noise levels, more high- and medium-frequency information is available and the task shifts to generating finer details and intricate structures. This multi-objective nature intuitively explains why previous methods found diffusion features extracted from different timesteps to provide information with differing semantics. + +Diffusion Feature Extraction Typically, diffusion feature extraction happens after first adding noise to an image and passing the resulting $\mathbf{x}_t$ to a U-Net [52] denoiser. Features are then extracted at multiple hand-picked locations of the U-Net decoder [2, 39, 62, 66, 67]. Different levels of noise added to the input image result in features beneficial for different downstream applications. Typical diffusion timesteps are $t = 261$ [62], $t = 100$ [66] or $t = 50$ [67]. By adding noise to the input image, these methods bottleneck the perceptual information the model can extract. To illustrate this, we show Stable Diffusion 2.1's reconstruction of an image noised at $t = 261$ [62] in Figure 2. + +Diffusion Features Encode Noise We hypothesize that diffusion features extracted from noisy images $\mathbf{x}_t$ encode the noise $\epsilon$ in addition to information from the image. We investigate this hypothesis in a very simple setting by examining how well features $\mathrm{feat}(\epsilon; T)$ extracted from only the noise $\epsilon$ approximate the features $\mathrm{feat}(\mathbf{x}_t = \sqrt{\alpha_t} \mathbf{x}_0 + \sqrt{1 - \alpha_t} \epsilon; t)$ . Using least squares, we fit a single scalar approximation coefficient to obtain the optimal reconstruction. We then quantify how much of the variance of the overall features is explained by this approximation (cf. Figure 3). Even at relatively low timesteps, such as $t = 261$ used by DIFT [62], encoding pure noise explains a substantial fraction of the features' variance. Current diffusion feature methods extract this information jointly with the image information. Our proposed method addresses this issue by eliminating the noise from the feature extraction process. + +![](images/8908f26484b3c0741935d98d5d3b70f96a8d408bdff2ef9ca61262240b645678.jpg) +Figure 4. Our training setup. We train our model to predict features from a clean input image, while the frozen diffusion model is fed the noisy image. The projection heads project our model's features onto the noisy diffusion model features, given the noising timestep $t$ . For downstream tasks, we discard the projection heads and directly use our model's internal representations as features. + +We further analyze the residual and similarly decompose it via the features predicted for the clean image $\mathrm{feat}(x_t = x_0; t = 0)$ . We find that they do not fully explain the remainder of the features either. Instead, a substantial part of the feature variance at medium noise timesteps is timestep dependent and cannot be attributed to components present at $t = 0$ or $t = T$ . This matches observations by previous works [62] that found diffusion features at higher timesteps offer better semantics, despite the added noise at the image input. + +# 3.2. CleanDIFT: Noise-Free Diffusion Features + +We present CleanDIFT, our method to address the problem of noisy and time-dependent diffusion features. CleanDIFT extracts clean Diffusion FeaTures from a pretrained diffusion backbone through a lightweight fine-tuning process. An overview of our setup is shown in Figure 4. + +Extraction Setup We train our feature extraction model to match the diffusion model's internal representations. We initialize the feature extraction model as a trainable copy of the diffusion model. Crucially, the feature extraction model is given the clean input image, while the diffusion model receives the noisy image and the corresponding timestep as input. Our goal is to obtain a single, noise-free feature map from the feature extraction model that consolidates the information of the diffusion model's timestep-dependent internal representations into a single one. To align our model's representations with the timestep-dependent diffusion model features during training, we introduce pointwise timestep-conditioned feature projection heads. The feature maps predicted by these projection heads are then aligned to the diffusion model's features. For feature extraction at inference time, we usually discard the projection heads and directly use the feature extraction model's + +internal representations. However, the projection heads can also be used to efficiently obtain feature maps for specific timesteps by reusing the feature extraction model's internal representations and passing them through the projection heads for different $t$ values. + +Training Objective We regard the diffusion model as a family of feature extraction functions $\operatorname{feat}(\cdot, \epsilon, t)$ for timestep $t \in [1,999]$ and noise $\epsilon \sim \mathcal{N}(0, \mathbf{I})$ . Each of these functions maps an image $\mathbf{x}$ to a feature vector $\operatorname{feat}(\mathbf{x}, \epsilon, t)$ . We aim to consolidate the information provided by all feature extraction functions into a single joint function $\operatorname{feat}_c(\cdot)$ with the same dimensionality: + +$$ +\begin{array}{c} \underline {{\text {S t a b l e D i f f u s i o n}}} \\ \left. \begin{array}{l} \operatorname {f e a t} (\mathbf {x}, \boldsymbol {\epsilon}, t = 1) \\ \operatorname {f e a t} (\mathbf {x}, \boldsymbol {\epsilon}, t = 2) \\ \vdots \\ \operatorname {f e a t} (\mathbf {x}, \boldsymbol {\epsilon}, t = 9 9 9) \end{array} \right\} \xrightarrow [ \text {i n f o r m a t i o n} ]{\text {c o n s o l i d a t e}} \operatorname {f e a t} _ {\mathrm {c}} (\mathbf {x}) \end{array} +$$ + +To align our model's features with the diffusion model's features, we maximize the similarity between the diffusion model's features and the projected features of our feature extraction model: + +$$ +\operatorname {f e a t} _ {\mathrm {c}} (\mathbf {x}) \to \left\{ \begin{array}{l} \operatorname {p r o j} (\cdot , t = 1) \longrightarrow \mathcal {L} \longleftarrow \operatorname {f e a t} (\mathbf {x}, \epsilon , t = 1) \\ \operatorname {p r o j} (\cdot , t = 2) \longrightarrow \mathcal {L} \longleftarrow \operatorname {f e a t} (\mathbf {x}, \epsilon , t = 2) \\ \vdots \\ \operatorname {p r o j} (\cdot , t = 9 9 9) \longrightarrow \mathcal {L} \longleftarrow \operatorname {f e a t} (\mathbf {x}, \epsilon , t = 9 9 9) \end{array} \right. +$$ + +Specifically, we minimize the negative cosine similarity between the diffusion model's features and our model's features extracted at stages $k = \{1,\dots,K\}$ in the network. Given a clean image $\mathbf{x}_0$ , the feature extraction model's output for feature map $k$ is denoted as $\mathrm{feat}_{\mathrm{c}}^{(k)}(\mathbf{x}_0)$ . Our Clean-DIFT feature map is then adapted by the learned projection heads $\mathrm{proj}^{(k)}(\mathrm{feat}_{\mathrm{c}}^{(k)}(\mathbf{x}_0),t)$ , where $\mathrm{proj}^{(k)}(\cdot ,\cdot)$ is the projection head for feature map $k$ . The diffusion model receives the noisy image $\mathbf{x}_t$ corresponding to the same $\mathbf{x}_0$ and timestep $t$ . The projection head then learns a timestep-dependent alignment from CleanDIFT features to the diffusion model's features $\mathrm{feat}^{(k)}(\mathbf{x}_t;t)$ . Putting it all together, our loss function is defined as: + +$$ +\mathcal {L} = - \sum_ {k = 1} ^ {K} \operatorname {s i m} \left(\operatorname {p r o j} ^ {(k)} \left(\operatorname {f e a t} _ {\mathrm {c}} ^ {(k)} \left(\mathbf {x} _ {0}\right); t\right), \operatorname {f e a t} ^ {(k)} \left(\mathbf {x} _ {t}; t\right)\right). \tag {1} +$$ + +For each training image $\mathbf{x}_0$ , we sample $I$ different nois-ing timesteps $t_i$ in a stratified manner, with each timestep $t_i \sim \mathcal{U}(\frac{i}{I} T, \frac{i+1}{I} T)$ , where $T$ is the maximum timestep. By sampling multiple timesteps per image we incentivize the feature extraction model to match the diffusion model's features across the entire noise spectrum. + +# 4. Experiments + +We test our hypothesis that the proposed extraction setup enables us to leverage more of the world knowledge inherent in diffusion models compared to existing diffusion feature extraction methods while being task-agnostic and timestep-independent. To that end, we evaluate our features on a wide range of downstream tasks: unsupervised zero-shot semantic correspondence, monocular depth estimation, semantic segmentation, and classification. We compare our features against standard diffusion features, methods that combine diffusion features with additional features, and non-diffusion-based approaches. + +# 4.1. Experimental Setup + +Implementation Details Following previous works [62, 66, 67], we evaluate our method on a Stable Diffusion (SD) backbone [51]. We apply our method to SD 1.5 and SD 2.1 to enable fair comparisons with existing methods that use either. We fully fine-tune our feature extraction model on image-caption pairs for only 400 steps, taking 30 minutes on a single A100 GPU, which was sufficient for our strong performance and further training did not yield any significant gains. We extract features after the U-Net's middle block and after each of the U-Net's decoder blocks, except the two final blocks. A detailed visualization of where we extract features is provided in Figure 11. This yields a total of $K = 11$ feature maps that we align between the diffusion model and the feature extraction model. Our point-wise feature projection heads consist of three stacked Feed Forward Networks (FFNs) that are zero-initialized such that initially they act as identity mappings due to their residual connections. Since every aligned feature map has its own projection head, this results in 45M additional trainable parameters for SD 2.1. We study the effect of different projection head architectures in Sec. E. We train using Adam [29] with a batch size of 8 and a learning rate of 2e-6 with a linear warmup. For stratified timestep sampling, we utilize $I = 3$ stratification bins across all our experiments, i.e. three different noise levels per training image. + +Datasets We fine-tune our feature extraction model on a random subset of COYO-700M [5], which is similar to the LAION [56], the dataset that Stable Diffusion 1.5 and 2.1 were trained on originally. That way, we ensure that all performance improvements originate from the feature extraction model consolidating the diffusion model's internal feature representations over time, not from choosing a different dataset that matches the test dataset distribution more closely. The subset selects images with a minimum size of $512^2$ . We crop and resize them to match the corresponding input resolution of the underlying diffusion model. We analyze the effect of using different datasets in Sec. G. + +![](images/7fbfb9bdf92425f5ea83b84afada92433835a9568da4e6ec14b62de7ad65dd8d.jpg) +Figure 5. Semantic correspondence results using DIFT [62] features with the standard SD 2.1 ( $t = 261$ ) and our CleanDIFT features. Our clean features show significantly less incorrect matches than the base diffusion model. + +# 4.2. Unsupervised Semantic Correspondence + +As many previous diffusion feature methods focus on (unsupervised) semantic correspondence matching [35, 62, 66, 67], we perform an extensive evaluation of our method on this task. Following previous works on semantic correspondence matching [62, 66, 67], we measure our performance in Percentage of Correct Keypoints (PCK). We average PCK directly across all keypoints, not over images. We use $\alpha = 0.1$ as a threshold and report both PCK values with error margins relative to the image size and to the bounding box size, denoted as $\mathrm{PCK}_{\mathrm{img}}$ and $\mathrm{PCK}_{\mathrm{bbox}}$ respectively. We evaluate the performance on the test split of the SPair-71k dataset, which consists of approximately $12\mathrm{k}$ image pairs from 18 categories. Some existing works [62, 66, 67] evaluate on additional datasets but find SPair-71k to be the most challenging and therefore the most informative benchmark. For the text prompt we use "A photo of a category.", with category being the corresponding category of the SPair image. We experiment with distilling the text conditioning in Sec. B. + +Results We first compare our extracted features to DIFT [62], an approach that detects semantic correspondences using standard diffusion features. Substituting these with our CleanDIFT features yields a performance increase of 1.79 absolute percentage points for $\mathrm{PCK}_{\mathrm{img}}$ and 1.86 percentage points for $\mathrm{PCK}_{\mathrm{bbox}}$ . Notably, DIFT averages the extracted feature maps across 8 different noise samples. Without this averaging over noise samples, our performance gain is even larger (2.81 PCK@ $\alpha_{\mathrm{img}}$ gain). This indicates that our feature extraction model learns more than a mere averaging over the noise in the diffusion model's feature maps (see Figure 5 and Sec. C for examples). We present an extended version of the time-step dependent performance analysis conducted by [62] in Figure 6: We evaluate the + +![](images/b04caf2667ee539939992132bc059f285b3dc5ff58fc142abc4636fae9c086b6.jpg) +Figure 6. Following [62], we evaluate semantic correspondence matching accuracy for different noise levels. Our feature extractor outperforms the standard noisy diffusion features across all timesteps $t$ . We additionally demonstrate that simply providing the diffusion model with a clean image and a non-zero timestep does not result in improved performance. + +diffusion model's performance for different timesteps $t$ in two settings. In the first setting, we provide the diffusion model with a noisy input image $\mathbf{x}_t$ as usual. In the second setting, we demonstrate that feeding the clean image along with a non-zero timestep is not a viable solution to obtain meaningful features: We provide the diffusion model with the clean input image $\mathbf{x}_0$ for all timesteps $t$ . We observe that the model's performance for the clean input image degrades faster and has a lower peak than for the noisy input. This is to be expected, as the diffusion model was trained on noisy images, not clean images. Importantly, our CleanDIFT features are timestep-independent and consistently outperform standard diffusion features, even when the latter are optimized for the best-performing timestep. We further observe that this advantage generalizes effectively to other backbones, such as DiTs [44] (see Tab. 7). + +A Tale of Two Features [66] extends the approach of DIFT by combining diffusion features with DINOv2 [43] features. Again, we replace the standard diffusion features with our CleanDIFT features and observe that the performance gain transfers when combining our features with DINOv2 features. Telling Left from Right [67] further improves upon the results of A Tale of Two Features by introducing a test-time adaptive pose alignment strategy. We observe that the performance gain transfers to this setting as well. To the best of our knowledge, Telling Left from Right combined with our CleanDIFT features sets a new state-of-the-art in unsupervised zero-shot semantic correspondence matching. In summary, replacing standard diffusion features with our CleanDIFT features consistently results in a significant performance improvement across all three methods. We show an overview of the results in Tab. 1 and a more extensive evaluation per category in Sec. B. + +We also investigate the performance of our features in a supervised fine-tuning setting for semantic correspondence matching. Following [35], we train an aggregation network that uses all extracted feature maps and learns to aggregate them into a single task-specific feature map for semantic correspondence matching. In contrast to [35], we do not have to perform costly DDIM inversion [59] to obtain a matching noisy image for every timestep. Instead, + +
MethodOur FeaturesPCK@α(↑)
αimg=0.1αbbox=0.1
General Approaches
DINOv2+NN--55.6
Diff. Feat.-based Approaches
DIFT [62]X66.5359.57
✓68.32▲1.7961.43▲1.86
A Tale of Two Features [66]X72.3163.73
✓73.35▲1.0464.81▲1.08
Telling Left from Right [67]X77.0768.64
✓78.40▲1.3369.99▲1.35
+ +Table 1. Zero-shot unsupervised semantic correspondence matching performance comparison on SPair71k [37]. Our improved features consistently lead to substantial improvements in matching performance. We report PCK on the test split of SPair71k, aggregated per point. Numbers are reproduced, for a discussion and comparison to reported numbers view Tab. 5. + +we directly feed the clean image to our feature extraction model. Therefore, extracting features with our CleanDIFT approach is $50\mathrm{x}$ faster, since we perform a single denoiser forward pass while [35] performs 50 for the inversion. Our model achieves a $\mathrm{PCK}_{\mathrm{img}}$ value of 72.48 vs their 72.75 and a $\mathrm{PCK}_{\mathrm{bbox}}$ value of 64.37 vs their 64.53. We observe a slight performance regression compared to their approach, however, at a speedup of $50\times$ . Luo et al. [35] also present a single-step ablation of their full method that only requires a single forward pass which makes it more comparable to ours. We outperform this single-step version by a wide margin of 9.0 percentage points for $\mathrm{PCK}_{\mathrm{img}}$ and 9.1 percentage points for $\mathrm{PCK}_{\mathrm{bbox}}$ . + +# 4.3. Depth Estimation + +We also investigate monocular depth estimation on NYUv2 [40]. Similar to [43], we follow the evaluation protocol from [34]. We use SD 2.1 as the base model and resize the input to the model's native resolution of $768^2$ . We extract features from the same location as [62] and obtain a feature map of dimension $48^2$ . Unlike [43], we do not up-sample the features and directly apply the linear probe. The probe predicts depth in 256 uniform bins which we combine with a classification loss after a linear normalization following [3]. We train one probe for our CleanDIFT features and one for standard diffusion features at $t = 299$ , as that timestep minimizes the error in our settings. Our qualitative results (see Figure 7) show a substantial fidelity gap in the estimated depth maps between the features from the standard SD 2.1 backbone and the features from our feature extraction model. This is reflected in a substantial improvement in quantitative metrics over the baseline as seen in Tab. 2. Lastly, we reuse the probe trained on standard diffusion features and apply it on the CleanDIFT features. While this does not match the performance of the CleanDIFT probe, it still achieves significantly better re + +![](images/4b075a4acf8c4fdd30068b4c36b81a1fdb99db7d64c70819c392abcf76c7c770.jpg) +Figure 7. Qualitative results for depth estimation using a linear probe on diffusion features on NYUv2 [40]. Our CleanDIFT features enable substantially better depth estimation than standard diffusion features. Note how the CleanDIFT features are far less noisy when compared to the standard diffusion features. + +
MethodBackboneRMSE (↓)
Self-Supervised Methods
OpenCLIP [28]ViT-G/140.541
MAE [20]ViT-H/140.517
DINO [6]ViT-B/80.555
iBOT [68]ViT-L/160.417
DINOv2 [43]ViT-g/140.344
Diffusion Features
SD 2.1 [51]0.469
DIFF- like [62]Ours0.444,0.025
+ Probes from noisy features0.453,0.016
+ +Table 2. Monocular Depth Estimation. Following [43], we evaluate metric depth prediction on NYUv2 [40] using a linear probe. Our clean features outperform the noisy features by a significant margin. Probes trained on the noisy features can be reused for the clean features, but incur a smaller performance gain. + +sults when compared to using standard diffusion features. This indicates that our features can be used as a drop-in replacement for the original diffusion features and offer improved performance on downstream applications. + +# 4.4. Semantic Segmentation + +To further investigate the difference between standard noisy diffusion features and our CleanDIFT features, we evaluate on the semantic segmentation task by training linear probes on our CleanDIFT features and on standard diffusion features. We utilize SD2.1 as the diffusion backbone and extract features at the same location as [62]. This procedure yields feature maps of size $48^2$ . We train our linear probe on the $48^2$ feature maps and upscale the obtained segmentation masks using nearest neighbor upsampling. We train and evaluate on the PASCAL VOC dataset [16]. Following common practice [39], we use mean Intersection over Union (mIoU) as the evaluation metric. Qualitative re + +![](images/703e20298c04cd3462fa6eef64a71f464ec16a7a012ff999757a2546f93a18bf.jpg) +Figure 8. Qualitative results for semantic segmentation from diffusion features on Pascal VOC [16]. Standard SD features use $t = 100$ as the timestep, which we found to perform best quantitatively (cf. Figure 9). Note how the CleanDIFT segmentation maps are far less noisy than those of the standard diffusion features. + +![](images/751c2eee31ecec5a62b102e93fb85293f183c5fb6cdb447a0e8be90bdee80fe1.jpg) +Figure 9. Performance on semantic segmentation using linear probes. Our clean features outperform the noisy diffusion features for the best noising timestep $t$ . Semantic segmentation performance of a standard diffusion model heavily depends on the used noising timestep. Unlike for semantic correspondence matching, the optimal $t$ value appears to be around $t = 100$ . + +sults are shown in Figure 8. Using our features, we observe significantly less noisy segmentations than with standard diffusion features. We show a quantitative comparison of our CleanDIFT feature's performance against standard diffusion features across timesteps in Figure 9. Notably, the optimal timestep appears to be around $t = 100$ , in contrast to the optimal timestep for semantic correspondences, which [62] found to be $t = 261$ . This highlights the need for tuning a timestep individually per downstream task. Our method both alleviates the need for such a timestep tuning and outperforms the standard diffusion features for the optimal timestep. + +# 4.5. Classification + +To assess the impact of our method on non-spatial tasks, we evaluate classification performance using pooled features. Pooling mitigates the influence of localized noise, so we anticipate classification performance to remain on par with standard diffusion features unless our setup introduces detrimental effects. We perform k-Nearest Neighbor (kNN) classification with $k = 10$ on ImageNet1k [11], using SD 1.5 as the diffusion backbone. We sweep across feature maps and timesteps $t$ for the base model, with results presented in Figure 10. Our analysis shows that the fea + +![](images/b72a4eccdb33442134d8ae17ce703235bd652130868923e4999f81db0ef6b8f9.jpg) +Figure 10. Classification performance on ImageNet1k [11], using kNN classifier with $k = 10$ and cosine similarity as the distance metric. We sweep over different timesteps and feature maps. We find that the feature map with the lowest spatial resolution (feature map #0) yields the highest classification accuracy. + +ture map with the lowest spatial resolution, i.e., feature map #0 (see Figure 11), achieves the highest classification accuracy. Furthermore, the optimal timestep $t$ for the base model varies between feature maps. For the best-performing feature map, $t = 100$ yields the highest classification accuracy. + +Importantly, CleanDIFT features slightly outperform the standard diffusion features even when using an optimal timestep $t$ for the base model showing that it does not introduce any detrimental effects. + +# 4.6. Ablation Studies + +For simplicity, we perform our ablation studies using DIFT [62] and evaluate the performance for unsupervised zero-shot semantic correspondence matching on a subset of the SPair71k [37] test split. + +Training Objective During training, we maximize the cosine similarity between projected outputs of our feature extraction model and standard diffusion features to align them. To investigate the influence of the employed similarity metric, we compare feature extraction models trained on three different alignment objectives commonly used in similar contexts: mean absolute error $(L_{1})$ , mean squared error $(L_{2})$ , and cosine similarity. Quantitative results are provided in Tab. 3. While all objectives result in feature extraction models that outperform standard diffusion features, cosine similarity consistently performs the best across the alignment objectives by a significant margin, confirming our choice of similarity metric. + +Projection Heads We investigate the influence of our proposed projection heads that are used to project Clean-DIFT features onto standard diffusion features. The alignment of feature extraction model and diffusion model is determined by the utilized similarity metric and the pro + +
ObjectiveProjection HeadsPCK@α(↑)
αimg=0.1αbbox=0.1
Cosine Sim.✓68.3261.43
×68.1661.29
L2✓66.2359.13
×66.4959.43
L1✓66.9160.00
×66.8759.91
SD 2.1-63.4155.92
+ +Table 3. Ablation Study Results. We evaluate the feature extraction models' performance for zero-shot semantic correspondence matching on the SPair71k test split. PCK is aggregated per point. + +jection heads. Therefore, we evaluate our feature extraction model's performance with and without the projection heads in combination with all three similarity metrics. An overview of the comparison is given in Tab. 3. In our main configuration that uses cosine similarity, using the projection heads yields slight performance improvements of 0.24 percentage points for $\mathrm{PCK}_{\mathrm{img}}$ and 0.06 percentage points for $\mathrm{PCK}_{\mathrm{bbox}}$ compared to fine-tuning without the projection heads. As the projection heads are typically not used for inference, they add computational overhead only during the lightweight fine-tuning. Therefore, we argue that it is worthwhile to include them and leverage the small performance gain. Additionally, they can be reused to efficiently obtain feature maps for specific timesteps. + +# 5. Conclusion + +In this paper, we introduced CleanDIFT, a novel approach for extracting diffusion features. CleanDIFT produces noise-free, timestep-independent, general-purpose diffusion features by consolidating timestep-dependent representations from a pre-trained diffusion backbone into a unified feature representation. We achieve this alignment between our feature extraction model and the pre-trained diffusion backbone through a lightweight fine-tuning procedure that takes approximately 30 minutes on a single A100 GPU. Operating directly on clean images, our method eliminates the information loss associated with adding noise to input images. Furthermore, CleanDIFT removes the requirement for tuning timesteps for each downstream task and avoids the computational overhead of ensembling over noise levels or timesteps. Instead, our method efficiently extracts features with just a single forward pass at inference time, substantially reducing inference costs compared to methods relying on ensembling or inversion. Extensive evaluations of CleanDIFT across diverse downstream tasks demonstrate significant performance improvements over conventional diffusion features. + +# Acknowledgments + +This project has been supported by the German Federal Ministry for Economic Affairs and Climate Action within the project "NXT GEN AI METHODS - Generative Methoden für Perzeption, Prädiktion und Planung", the project "GeniusRobot" (01IS24083), funded by the Federal Ministry of Education and Research (BMBF), the bidt project KLIMA-MEMES, and the German Research Foundation (DFG) project 421703927. The authors gratefully acknowledge the Gauss Center for Supercomputing for providing compute through the NIC on JUWELS at JSC and the HPC resources supplied by the Erlangen National High Performance Computing Center (NHR@FAU funded by DFG project 440719683) under the NHR project JA-22883. + +Further, we would like to thank Owen Vincent for continuous technical support. + +# References + +[1] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Ji-aming Song, Qinsheng Zhang, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, et al. ediff-i: Text-to-image diffusion models with an ensemble of expert denoisers. arXiv preprint arXiv:2211.01324, 2022. 2, 3 +[2] Dmitry Baranchuk, Andrey Voynov, Ivan Rubachev, Valentin Khrulkov, and Artem Babenko. Label-efficient semantic segmentation with diffusion models. In International Conference on Learning Representations, 2022. 1, 3 +[3] Shariq Farooq Bhat, Ibrahim Alhashim, and Peter Wonka. Adabins: Depth estimation using adaptive bins. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4009-4018, 2021. 6 +[4] Giulio Biroli, Tony Bonnaire, Valentin De Bortoli, and Marc Mézard. Dynamical regimes of diffusion models. arXiv preprint arXiv:2402.18491, 2024. 2, 3 +[5] Minwoo Byeon, Beomhee Park, Haecheon Kim, Sungjun Lee, Woonhyuk Baek, and Saehoon Kim. Coyo-700m: Image-text pair dataset. https://github.com/kakaobrain/coyo-dataset, 2022.5 +[6] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 2, 7 +[7] Junsong Chen, Jincheng YU, Chongjian GE, Lewei Yao, Enze Xie, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart- $\mathbb{S}\backslash \mathbb{alpha}$ : Fast training of diffusion transformer for photorealistic text-to-image synthesis. In The Twelfth International Conference on Learning Representations, 2024. 14 +[8] Shoufa Chen, Peize Sun, Yibing Song, and Ping Luo. Diffusionnet: Diffusion model for object detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 19830-19843, 2023. 1, 3 +[9] Xinlei Chen, Zhuang Liu, Saining Xie, and Kaiming He. De + +constructing denoising diffusion models for self-supervised learning. arXiv preprint arXiv:2401.14404, 2024. 2, 3 +[10] Timothee Darcet, Maxime Oquab, Julien Mairal, and Piotr Bojanowski. Vision transformers need registers. In The Twelfth International Conference on Learning Representations, 2024. 2 +[11] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 248–255, 2009. 7, 8 +[12] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 1, 2 +[13] Patrick Esser, Johnathan Chiu, Parmida Atighehchian, Jonathan Granskog, and Anastasis Germanidis. Structure and content-guided video synthesis with diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7346-7356, 2023. 1, 2 +[14] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In *Forty-first International Conference on Machine Learning*, 2024. 2, 14 +[15] Zach Evans, CJ Carr, Josiah Taylor, Scott H Hawley, and Jordi Pons. Fast timing-conditioned latent audio diffusion. In *Forty-first International Conference on Machine Learning*, 2024. 2 +[16] M. Everingham, L. Van Gool, C. K. I. Williams, J. Winn, and A. Zisserman. The Pascal visual object classes (voc) challenge. International Journal of Computer Vision, 88(2): 303-338, 2010. 7, 13 +[17] Michael Fuest, Pingchuan Ma, Ming Gui, Johannes S. Fischer, Vincent Tao Hu, and Björn Ommer. Diffusion models and representation learning: A survey. CoRR, abs/2407.00783, 2024. 3 +[18] Frank Fundel, Johannes Schusterbauer, Vincent Tao Hu, and Björn Ommer. Distillation of diffusion features for semantic correspondence. WACV, 2025. 3 +[19] Spyros Gidaris, Praveer Singh, and Nikos Komodakis. Unsupervised representation learning by predicting image rotations. In International Conference on Learning Representations, 2018. 2 +[20] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dólár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 2, 7 +[21] Eric Hedlin, Gopal Sharma, Shweta Mahajan, Hossam Isack, Abhishek Kar, Andrea Tagliasacchi, and Kwang Moo Yi. Unsupervised semantic correspondence using stable diffusion. Advances in Neural Information Processing Systems, 36, 2023. 3 +[22] Geoffrey Hinton, Oriol Vinyls, and Jeffrey Dean. Distilling the knowledge in a neural network. In NIPS Deep Learning and Representation Learning Workshop, 2015. 3 +[23] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 3 + +[24] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2 +[25] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. Advances in Neural Information Processing Systems, 35:8633-8646, 2022. 1 +[26] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. *ICLR*, 1(2):3, 2022. 14 +[27] Drew A Hudson, Daniel Zoran, Mateusz Malinowski, Andrew K Lampinen, Andrew Jaegle, James L McClelland, Loic Matthey, Felix Hill, and Alexander Lerchner. Soda: Bottleneck diffusion models for representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23115-23127, 2024. 3 +[28] Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. Openclip, 2021. 7 +[29] Diederik P. Kingma and Jimmy Lei Ba. Adam: A method for stochastic optimization. In Proceedings of the 3rd International Conference on Learning Representations (ICLR), 2015. 5 +[30] Zhifeng Kong, Wei Ping, Jiaji Huang, Kexin Zhao, and Bryan Catanzaro. Diffwave: A versatile diffusion model for audio synthesis. In International Conference on Learning Representations, 2021. 2 +[31] Alexander C Li, Mihir Prabhudesai, Shivam Duggal, Ellis Brown, and Deepak Pathak. Your diffusion model is secretly a zero-shot classifier. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2206-2217, 2023. 1, 3 +[32] Daiqing Li, Huan Ling, Amlan Kar, David Acuna, Seung Wook Kim, Karsten Kreis, Antonio Torralba, and Sanja Fidler. Dreamteacher: Pretraining image backbones with deep generative models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16698-16708, 2023. 3 +[33] Xinghui Li, Jingyi Lu, Kai Han, and Victor Adrian Prisacariu. Sd4match: Learning to prompt stable diffusion model for semantic matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 27558-27568, 2024. 3 +[34] Zhenyu Li, Xuyang Wang, Xianming Liu, and Junjun Jiang. Binsformer: Revisiting adaptive bins for monocular depth estimation. IEEE Transactions on Image Processing, 2024. 6 +[35] Grace Luo, Lisa Dunlap, Dong Huk Park, Aleksander Holynski, and Trevor Darrell. Diffusion hyperfeatures: Searching through time and space for semantic correspondence. In Advances in Neural Information Processing Systems, pages 47500-47510. Curran Associates, Inc., 2023. 3, 5, 6, 13 +[36] Chenlin Meng, Robin Rombach, Ruiqi Gao, Diederik Kingma, Stefano Ermon, Jonathan Ho, and Tim Salimans. + +On distillation of guided diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14297-14306, 2023. 3 +[37] Juhong Min, Jongmin Lee, Jean Ponce, and Minsu Cho. Spair-71k: A large-scale benchmark for semantic correspondence. arXiv preprint arXiv:1908.10543, 2019. 6, 8, 12, 13 +[38] Ishan Misra and Laurens van der Maaten. Self-supervised learning of pretext-invariant representations. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6707-6717, 2020. 2 +[39] Koichi Namekata, Amirmojtaba Sabour, Sanja Fidler, and Seung Wook Kim. Emerdiff: Emerging pixel-level semantic knowledge in diffusion models. In The Twelfth International Conference on Learning Representations, 2023. 1, 3, 7 +[40] Pushmeet Kohli Nathan Silberman, Derek Hoiem and Rob Fergus. Indoor segmentation and support inference from rgbd images. In ECCV, 2012. 6, 7, 14 +[41] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 1, 2 +[42] Mehdi Noroozi and Paolo Favaro. Unsupervised learning of visual representations by solving jigsaw puzzles. In European conference on computer vision, pages 69-84. Springer, 2016. 2 +[43] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel HAZIZA, Francisco Massa, Alaaeldin El-Nouby, Mido Assran, Nicolas Ballas, Wojciech Galuba, Russell Howes, Po-Yao Huang, Shang-Wen Li, Ishan Misra, Michael Rabbat, Vasu Sharma, Gabriel Synnaeve, Hu Xu, Herve Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. DINOv2: Learning robust visual features without supervision. Transactions on Machine Learning Research, 2024. 2, 3, 6, 7 +[44] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4195-4205, 2023. 6 +[45] Ethan Perez, Florian Strub, Harm De Vries, Vincent Dumoulin, and Aaron Courville. Film: Visual reasoning with a general conditioning layer. In Proceedings of the AAAI conference on artificial intelligence, 2018. 14 +[46] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. In The Twelfth International Conference on Learning Representations, 2024. 2, 14 +[47] Adam Polyak, Amit Zohar, Andrew Brown, Andros Tjandra, Animesh Sinha, Ann Lee, Apoorv Vyas, Bowen Shi, Chih-Yao Ma, Ching-Yao Chuang, et al. Movie gen: A cast of media foundation models. arXiv preprint arXiv:2410.13720, 2024. 2 +[48] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, + +Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2 +[49] Prajit Ramachandran, Barret Zoph, and Quoc V Le. Searching for activation functions. arXiv preprint arXiv:1710.05941, 2017. 14 +[50] Severi Rissanen, Markus Heinonen, and Arno Solin. Generative modelling with inverse heat dissipation. arXiv preprint arXiv:2206.13397, 2022. 3 +[51] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 2, 5, 7 +[52] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention - MICCAI 2015, pages 234-241, Cham, 2015. Springer International Publishing. 3 +[53] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural information processing systems, 35:36479-36494, 2022. 2 +[54] Tim Salimans and Jonathan Ho. Progressive distillation for fast sampling of diffusion models. arXiv preprint arXiv:2202.00512, 2022. 3 +[55] Axel Sauer, Dominik Lorenz, Andreas Blattmann, and Robin Rombach. Adversarial diffusion distillation. In European Conference on Computer Vision, pages 87-103. Springer, 2024. 14 +[56] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278-25294, 2022. 5 +[57] Noam Shazeer. Glu variants improve transformer. arXiv preprint arXiv:2002.05202, 2020. 14 +[58] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. In The Eleventh International Conference on Learning Representations, 2023. 1 +[59] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In International Conference on Learning Representations, 2021. 6 +[60] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. Advances in neural information processing systems, 32, 2019. 2 +[61] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 2 +[62] Luming Tang, Menglin Jia, Qianqian Wang, Cheng Perng Phoo, and Bharath Hariharan. Emergent correspondence + +from image diffusion. In Advances in Neural Information Processing Systems, pages 1363-1389. Curran Associates, Inc., 2023. 1, 2, 3, 4, 5, 6, 7, 8, 12, 13, 14, 15 +[63] Weijia Wu, Yuzhong Zhao, Mike Zheng Shou, Hong Zhou, and Chunhua Shen. Diffumask: Synthesizing images with pixel-level annotations for semantic segmentation using diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1206-1217, 2023. 1, 3 +[64] Jiarui Xu, Sifei Liu, Arash Vahdat, Wonmin Byeon, Xiaolong Wang, and Shalini De Mello. Open-vocabulary panoptic segmentation with text-to-image diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2955-2966, 2023. 1, 3, 13, 15 +[65] Biao Zhang and Rico Sennrich. Root mean square layer normalization. Advances in Neural Information Processing Systems, 32, 2019. 14 +[66] Junyi Zhang, Charles Herrmann, Junhwa Hur, Luisa Polania Cabrera, Varun Jampani, Deqing Sun, and Ming-Hsuan Yang. A tale of two features: Stable diffusion complements dino for zero-shot semantic correspondence. In Advances in Neural Information Processing Systems, pages 45533-45547. Curran Associates, Inc., 2023. 1, 2, 3, 5, 6, 12, 13, 15 +[67] Junyi Zhang, Charles Herrmann, Junhwa Hur, Eric Chen, Varun Jampani, Deqing Sun, and Ming-Hsuan Yang. Telling left from right: Identifying geometry-aware semantic correspondence. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3076-3085, 2024. 1, 2, 3, 5, 6, 12, 13, 15 +[68] Jinghao Zhou, Chen Wei, Huiyu Wang, Wei Shen, Cihang Xie, Alan Yuille, and Tao Kong. Image BERT pre-training with online tokenizer. In International Conference on Learning Representations, 2022. 7 \ No newline at end of file diff --git a/2025/CleanDIFT_ Diffusion Features without Noise/images.zip b/2025/CleanDIFT_ Diffusion Features without Noise/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..0402817c98078e4697dee03416d98ab037813cda --- /dev/null +++ b/2025/CleanDIFT_ Diffusion Features without Noise/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7bbb24a552b3687d12dbc072418a374a675c2981b313f3e755e8da6d48a1e051 +size 462281 diff --git a/2025/CleanDIFT_ Diffusion Features without Noise/layout.json b/2025/CleanDIFT_ Diffusion Features without Noise/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..95acade1bafe718df2462d1dea33370c2e3d8b67 --- /dev/null +++ b/2025/CleanDIFT_ Diffusion Features without Noise/layout.json @@ -0,0 +1,9696 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 157, + 103, + 455, + 120 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 103, + 455, + 120 + ], + "spans": [ + { + "bbox": [ + 157, + 103, + 455, + 120 + ], + "type": "inline_equation", + "content": "\\checkmark" + }, + { + "bbox": [ + 157, + 103, + 455, + 120 + ], + "type": "text", + "content": " CleanDIFT: Diffusion Features without Noise" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 63, + 143, + 132, + 156 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 143, + 132, + 156 + ], + "spans": [ + { + "bbox": [ + 63, + 143, + 132, + 156 + ], + "type": "text", + "content": "Nick Stracke*" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 152, + 143, + 281, + 156 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 143, + 281, + 156 + ], + "spans": [ + { + "bbox": [ + 152, + 143, + 281, + 156 + ], + "type": "text", + "content": "Stefan Andreas Baumann*" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 302, + 144, + 367, + 157 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 144, + 367, + 157 + ], + "spans": [ + { + "bbox": [ + 302, + 144, + 367, + 157 + ], + "type": "text", + "content": "Kolja Bauer*" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 386, + 144, + 452, + 156 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 386, + 144, + 452, + 156 + ], + "spans": [ + { + "bbox": [ + 386, + 144, + 452, + 156 + ], + "type": "text", + "content": "Frank Fundel" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 477, + 144, + 547, + 157 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 477, + 144, + 547, + 157 + ], + "spans": [ + { + "bbox": [ + 477, + 144, + 547, + 157 + ], + "type": "text", + "content": "Björn Ommer" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 218, + 158, + 392, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 218, + 158, + 392, + 171 + ], + "spans": [ + { + "bbox": [ + 218, + 158, + 392, + 171 + ], + "type": "text", + "content": "CompVis @ LMU Munich, MCML" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 225, + 173, + 382, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 225, + 173, + 382, + 185 + ], + "spans": [ + { + "bbox": [ + 225, + 173, + 382, + 185 + ], + "type": "text", + "content": "{nick.stracke,b.ommer}@lmu.de" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 223, + 190, + 386, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 223, + 190, + 386, + 202 + ], + "spans": [ + { + "bbox": [ + 223, + 190, + 386, + 202 + ], + "type": "text", + "content": "compvis.github.io/cleandift" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 154, + 213, + 336, + 226 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 154, + 213, + 336, + 226 + ], + "spans": [ + { + "bbox": [ + 154, + 213, + 336, + 226 + ], + "type": "text", + "content": "Normal Diffusion Features 30 min Finetuning" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 345, + 217, + 436, + 228 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 217, + 436, + 228 + ], + "spans": [ + { + "bbox": [ + 345, + 217, + 436, + 228 + ], + "type": "text", + "content": "CleanDIFT Features" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 129, + 231, + 290, + 241 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 231, + 290, + 241 + ], + "spans": [ + { + "bbox": [ + 129, + 231, + 290, + 241 + ], + "type": "text", + "content": "noisy images noisy features" + } + ] + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 131, + 243, + 176, + 283 + ], + "blocks": [ + { + "bbox": [ + 131, + 243, + 176, + 283 + ], + "lines": [ + { + "bbox": [ + 131, + 243, + 176, + 283 + ], + "spans": [ + { + "bbox": [ + 131, + 243, + 176, + 283 + ], + "type": "image", + "image_path": "0e4b63c0ac6e7712c160ff1e5389229e3308f90165be123a628a1cc446623839.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 141, + 284, + 168, + 291 + ], + "lines": [ + { + "bbox": [ + 141, + 284, + 168, + 291 + ], + "spans": [ + { + "bbox": [ + 141, + 284, + 168, + 291 + ], + "type": "inline_equation", + "content": "t = 300" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 179, + 243, + 234, + 289 + ], + "blocks": [ + { + "bbox": [ + 179, + 243, + 234, + 289 + ], + "lines": [ + { + "bbox": [ + 179, + 243, + 234, + 289 + ], + "spans": [ + { + "bbox": [ + 179, + 243, + 234, + 289 + ], + "type": "image", + "image_path": "48084df88f96a2606168a20000ccdfddaa3939408eb4a2567f1860df4377d05c.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 235, + 243, + 288, + 289 + ], + "blocks": [ + { + "bbox": [ + 235, + 243, + 288, + 289 + ], + "lines": [ + { + "bbox": [ + 235, + 243, + 288, + 289 + ], + "spans": [ + { + "bbox": [ + 235, + 243, + 288, + 289 + ], + "type": "image", + "image_path": "70715d719fbd02a6850f60f5dc2a5c015492a320ce7444325f7fe1b36d3362f0.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 132, + 308, + 175, + 357 + ], + "blocks": [ + { + "bbox": [ + 164, + 294, + 260, + 306 + ], + "lines": [ + { + "bbox": [ + 164, + 294, + 260, + 306 + ], + "spans": [ + { + "bbox": [ + 164, + 294, + 260, + 306 + ], + "type": "text", + "content": "timestep tuning required " + }, + { + "bbox": [ + 164, + 294, + 260, + 306 + ], + "type": "inline_equation", + "content": "X" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 132, + 308, + 175, + 357 + ], + "lines": [ + { + "bbox": [ + 132, + 308, + 175, + 357 + ], + "spans": [ + { + "bbox": [ + 132, + 308, + 175, + 357 + ], + "type": "image", + "image_path": "8668d2c527e376f5ca9b196f7f0446b78ecb983ce5f96f05ee26c5bbfd5d1c61.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 178, + 310, + 233, + 357 + ], + "blocks": [ + { + "bbox": [ + 178, + 310, + 233, + 357 + ], + "lines": [ + { + "bbox": [ + 178, + 310, + 233, + 357 + ], + "spans": [ + { + "bbox": [ + 178, + 310, + 233, + 357 + ], + "type": "image", + "image_path": "1d4b2a24cd398a9dc718f4fc53e2f2b24123651a37619ae1a68d30cb93d8c0ff.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 371, + 555, + 416 + ], + "lines": [ + { + "bbox": [ + 54, + 371, + 555, + 416 + ], + "spans": [ + { + "bbox": [ + 54, + 371, + 555, + 416 + ], + "type": "text", + "content": "Figure 1. Our proposed CleanDIFT feature extraction method yields noise-free, timestep-independent, general-purpose features that significantly outperform standard diffusion features. CleanDIFT operates on clean images, while extracting diffusion features with existing approaches requires adding noise to an image before passing it through the model. Adding noise reduces the information present in the image and requires tuning a timestep per downstream task." + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 235, + 311, + 288, + 357 + ], + "blocks": [ + { + "bbox": [ + 235, + 311, + 288, + 357 + ], + "lines": [ + { + "bbox": [ + 235, + 311, + 288, + 357 + ], + "spans": [ + { + "bbox": [ + 235, + 311, + 288, + 357 + ], + "type": "image", + "image_path": "e13d2c508add19b770a90f41496be17677f7b19da282ff4cac391dd9ed3a657f.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 231, + 476, + 241 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 231, + 476, + 241 + ], + "spans": [ + { + "bbox": [ + 313, + 231, + 476, + 241 + ], + "type": "text", + "content": "clean images clean features" + } + ] + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 312, + 243, + 363, + 289 + ], + "blocks": [ + { + "bbox": [ + 312, + 243, + 363, + 289 + ], + "lines": [ + { + "bbox": [ + 312, + 243, + 363, + 289 + ], + "spans": [ + { + "bbox": [ + 312, + 243, + 363, + 289 + ], + "type": "image", + "image_path": "1f19a52e101070d30fa1990b76ad79fdb74c69c081b4dbcee6d73ccb04bff8b2.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 339, + 295, + 447, + 306 + ], + "lines": [ + { + "bbox": [ + 339, + 295, + 447, + 306 + ], + "spans": [ + { + "bbox": [ + 339, + 295, + 447, + 306 + ], + "type": "text", + "content": "no timestep tuning required " + }, + { + "bbox": [ + 339, + 295, + 447, + 306 + ], + "type": "inline_equation", + "content": "\\checkmark" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 364, + 243, + 421, + 289 + ], + "blocks": [ + { + "bbox": [ + 364, + 243, + 421, + 289 + ], + "lines": [ + { + "bbox": [ + 364, + 243, + 421, + 289 + ], + "spans": [ + { + "bbox": [ + 364, + 243, + 421, + 289 + ], + "type": "image", + "image_path": "dcd916cdd0eae739e33f27d488bdd8f878eb87ad0e03ca7f0ad8be73d8191639.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 423, + 243, + 474, + 289 + ], + "blocks": [ + { + "bbox": [ + 423, + 243, + 474, + 289 + ], + "lines": [ + { + "bbox": [ + 423, + 243, + 474, + 289 + ], + "spans": [ + { + "bbox": [ + 423, + 243, + 474, + 289 + ], + "type": "image", + "image_path": "8a8e4a04cac091f870cb498ebd96098ebed58a18e19e928a854153fe7ee50abf.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 313, + 310, + 363, + 357 + ], + "blocks": [ + { + "bbox": [ + 313, + 310, + 363, + 357 + ], + "lines": [ + { + "bbox": [ + 313, + 310, + 363, + 357 + ], + "spans": [ + { + "bbox": [ + 313, + 310, + 363, + 357 + ], + "type": "image", + "image_path": "168d3f58973fdd6f652a8b910e50df8424c684385c73890753205aa80b492256.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 364, + 311, + 421, + 357 + ], + "blocks": [ + { + "bbox": [ + 364, + 311, + 421, + 357 + ], + "lines": [ + { + "bbox": [ + 364, + 311, + 421, + 357 + ], + "spans": [ + { + "bbox": [ + 364, + 311, + 421, + 357 + ], + "type": "image", + "image_path": "32fa88c7b0404cc885204005645bc2fe76b1fabf6b70ae7765c4500e5bb8e2b8.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 423, + 311, + 474, + 357 + ], + "blocks": [ + { + "bbox": [ + 423, + 311, + 474, + 357 + ], + "lines": [ + { + "bbox": [ + 423, + 311, + 474, + 357 + ], + "spans": [ + { + "bbox": [ + 423, + 311, + 474, + 357 + ], + "type": "image", + "image_path": "0f923256c1d2d1ea88e5f63c61e591c1f564b805aec913466b509a6ce0c40d97.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + } + ], + "index": 29 + }, + { + "bbox": [ + 152, + 426, + 200, + 438 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 426, + 200, + 438 + ], + "spans": [ + { + "bbox": [ + 152, + 426, + 200, + 438 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 53, + 456, + 297, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 456, + 297, + 661 + ], + "spans": [ + { + "bbox": [ + 53, + 456, + 297, + 661 + ], + "type": "text", + "content": "Internal features from large-scale pre-trained diffusion models have recently been established as powerful semantic descriptors for a wide range of downstream tasks. Works that use these features generally need to add noise to images before passing them through the model to obtain the semantic features, as the models do not offer the most useful features when given images with little to no noise. We show that this noise has a critical impact on the usefulness of these features that cannot be remedied by ensembling with different random noises. We address this issue by introducing a lightweight, unsupervised fine-tuning method that enables diffusion backbones to provide high-quality, noise-free semantic features. We show that these features readily outperform previous diffusion features by a wide margin in a wide variety of extraction setups and downstream tasks, offering better performance than even ensemble-based methods at a fraction of the cost." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 315, + 426, + 394, + 438 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 426, + 394, + 438 + ], + "spans": [ + { + "bbox": [ + 315, + 426, + 394, + 438 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 312, + 449, + 555, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 449, + 555, + 616 + ], + "spans": [ + { + "bbox": [ + 312, + 449, + 555, + 616 + ], + "type": "text", + "content": "Learning meaningful visual representations that capture a vast amount of world knowledge remains a key problem in the field of computer vision. Diffusion models can be trained at scale in a self-supervised manner and have rapidly advanced the state of the art in image [12, 41, 51] and video generation [13, 25, 58], making them a good candidate to learn visual representations. Many early works have already achieved impressive results using internal features from large-scale pretrained diffusion models for a wide variety of tasks, such as semantic correspondence detection [62, 66, 67], semantic segmentation [2, 39, 63], panoptic segmentation [64], object detection [8], and classification [31]. However, the optimal approach to extract this world knowledge from a diffusion model remains uncertain." + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 312, + 617, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 617, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 312, + 617, + 556, + 714 + ], + "type": "text", + "content": "To understand why that is the case, we take a look at how diffusion models are trained: a varying amount of noise is added to a clean input image (forward process) and the model is tasked to remove the noise from the image (backward process). The amount of added noise is dependent on the diffusion timestep. As a result, the model learns to operate on noisy images and also becomes dependent on the noise timestep as different noise levels require the model to" + } + ] + } + ], + "index": 35 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 702, + 135, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 702, + 135, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 702, + 135, + 712 + ], + "type": "text", + "content": "*Equal Contribution" + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 298, + 748, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 748, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 748, + 312, + 757 + ], + "type": "text", + "content": "117" + } + ] + } + ], + "index": 37 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "content": "perform different tasks [1, 4]. Since noisy images inherently contain less information than clean images (cf. Figure 2), we hypothesize that this harms the internal feature representation of diffusion models [9] and, thus, the extractable world knowledge. Furthermore, the timestep acts as a hyperparameter that influences the internal feature representation and needs to be picked independently for every downstream application (cf. Figure 14)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 169, + 295, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 169, + 295, + 289 + ], + "spans": [ + { + "bbox": [ + 55, + 169, + 295, + 289 + ], + "type": "text", + "content": "We propose a novel feature extraction method that (1) eliminates the need to destroy information by adding noise to the input; and (2) produces timestep-independent generic diffusion features useful for a wide range of down-stream tasks, alleviating the need to tune a noisig timestep per down-stream task. We show how to adapt an off-the-shelf large-scale pre-trained diffusion backbone to provide these features at minimal cost (approximately 30 minutes of fine-tuning on a single A100 GPU) and demonstrate improved performance across a wide range of downstream tasks." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 290, + 295, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 290, + 295, + 397 + ], + "spans": [ + { + "bbox": [ + 55, + 290, + 295, + 397 + ], + "type": "text", + "content": "We achieve this by viewing a diffusion model as a family of " + }, + { + "bbox": [ + 55, + 290, + 295, + 397 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 290, + 295, + 397 + ], + "type": "text", + "content": " feature extractors that operate on images with different noise levels and provide features with different characteristics. We consolidate all " + }, + { + "bbox": [ + 55, + 290, + 295, + 397 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 290, + 295, + 397 + ], + "type": "text", + "content": " feature extraction functions in our feature extractor by aligning their internal representations. Specifically, we initialize our feature extractor as a trainable copy of the diffusion model; fine-tune it with clean images and no timestep input; and align its features with all " + }, + { + "bbox": [ + 55, + 290, + 295, + 397 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 290, + 295, + 397 + ], + "type": "text", + "content": " time-dependent feature extractors of the diffusion model." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 400, + 295, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 400, + 295, + 591 + ], + "spans": [ + { + "bbox": [ + 55, + 400, + 295, + 591 + ], + "type": "text", + "content": "We evaluate our improved features across a wide variety of downstream tasks, such as semantic correspondence matching, monocular depth estimation, semantic segmentation, and classification, and find that they consistently improve upon approaches based on standard diffusion features. These improvements are most evident for dense visual tasks such as semantic correspondence matching, where our features show substantial performance gains across a wide variety of setups [62, 66, 67] and set a new state-of-the-art for unsupervised semantic correspondence matching. Additionally, our proposed method eliminates the need for noise or timestep ensembling [62], offering substantial speed gains (e.g., " + }, + { + "bbox": [ + 55, + 400, + 295, + 591 + ], + "type": "inline_equation", + "content": "8 \\times" + }, + { + "bbox": [ + 55, + 400, + 295, + 591 + ], + "type": "text", + "content": " over DIFT [62]), on top of improved quality. Our method is generic and integrates easily with established methods, such as fusing diffusion and DINOv2 features [66, 67]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 593, + 212, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 593, + 212, + 604 + ], + "spans": [ + { + "bbox": [ + 55, + 593, + 212, + 604 + ], + "type": "text", + "content": "Our main contributions are as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 56, + 605, + 295, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 605, + 295, + 653 + ], + "spans": [ + { + "bbox": [ + 56, + 605, + 295, + 653 + ], + "type": "text", + "content": "1. We propose CleanDIFT, a finetuning approach for diffusion models that enables them to operate on clean images and makes the inherent world knowledge of these models more accessible." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 654, + 295, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 295, + 689 + ], + "type": "text", + "content": "2. We show how to consolidate information from all diffusion timesteps into a single feature prediction, removing the need for task-specific timestep tuning." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": "3. We demonstrate significant performance gains of our diffusion feature extraction technique across a wide range" + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "type": "image", + "bbox": [ + 318, + 72, + 433, + 188 + ], + "blocks": [ + { + "bbox": [ + 318, + 72, + 433, + 188 + ], + "lines": [ + { + "bbox": [ + 318, + 72, + 433, + 188 + ], + "spans": [ + { + "bbox": [ + 318, + 72, + 433, + 188 + ], + "type": "image", + "image_path": "fd330a828668af62cc1322a621550038d9f262fe12bb7f536c044c42730d8b13.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 329, + 194, + 424, + 203 + ], + "lines": [ + { + "bbox": [ + 329, + 194, + 424, + 203 + ], + "spans": [ + { + "bbox": [ + 329, + 194, + 424, + 203 + ], + "type": "text", + "content": "(a) Reconstruction without Noise" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 217, + 553, + 273 + ], + "lines": [ + { + "bbox": [ + 313, + 217, + 553, + 273 + ], + "spans": [ + { + "bbox": [ + 313, + 217, + 553, + 273 + ], + "type": "text", + "content": "Figure 2. Deterioration of Diffusion Features. As current methods need to pass noisy images to the model to obtain useful features, they significantly reduce the information available. We alleviate this problem by obtaining useful features without noise, improving the performance of downstream tasks." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 436, + 72, + 552, + 188 + ], + "blocks": [ + { + "bbox": [ + 436, + 72, + 552, + 188 + ], + "lines": [ + { + "bbox": [ + 436, + 72, + 552, + 188 + ], + "spans": [ + { + "bbox": [ + 436, + 72, + 552, + 188 + ], + "type": "image", + "image_path": "12eea81cc5e551bfb499eda8e08eaaa662e3618dbd69cab2eeafb7d8e11d3116.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 441, + 190, + 544, + 208 + ], + "lines": [ + { + "bbox": [ + 441, + 190, + 544, + 208 + ], + "spans": [ + { + "bbox": [ + 441, + 190, + 544, + 208 + ], + "type": "text", + "content": "(b) Diffusion Model Reconstruction with added noise " + }, + { + "bbox": [ + 441, + 190, + 544, + 208 + ], + "type": "inline_equation", + "content": "(t = 261" + }, + { + "bbox": [ + 441, + 190, + 544, + 208 + ], + "type": "text", + "content": " [62])" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 325, + 294, + 553, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 294, + 553, + 366 + ], + "spans": [ + { + "bbox": [ + 325, + 294, + 553, + 366 + ], + "type": "text", + "content": "of down-stream tasks, notably surpassing the current state of the art in zero-shot unsupervised semantic correspondence detection. We further demonstrate the generality of our enhanced features by showing that these performance gains transfer to advanced methods that fuse diffusion features or operate in a supervised setting." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 366, + 553, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 366, + 553, + 402 + ], + "spans": [ + { + "bbox": [ + 313, + 366, + 553, + 402 + ], + "type": "text", + "content": "4. Our proposed approach is significantly more efficient than previous methods that tried to address this problem by noise ensembling or supervised training." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 413, + 400, + 426 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 413, + 400, + 426 + ], + "spans": [ + { + "bbox": [ + 313, + 413, + 400, + 426 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 434, + 555, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 434, + 555, + 613 + ], + "spans": [ + { + "bbox": [ + 313, + 434, + 555, + 613 + ], + "type": "text", + "content": "Self-Supervised Representation Learning Features from large, pre-trained foundation models have been shown to yield competitive performance to supervised models for a variety of downstream tasks, both in zero-shot and fine-tuning settings [20, 43, 48]. These foundation models are trained on different pre-text tasks like inpainting [20], predicting transformations [19], patch reordering [38, 42], and discriminative tasks [6, 43]. DINOv2 [43] uses a discriminative objective combined with self-distillation to learn general-purpose visual features that have proven useful for a variety of downstream tasks [10]. CLIP [48] learns such features by employing a contrastive objective on text-image pairs. Masked Autoencoders [20] (MAEs) are trained to reconstruct masked out patches of the input, also resulting in general-purpose visual features." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "type": "text", + "content": "Diffusion Models as Self-Supervised Learners Diffusion models [24, 60, 61] are generative models that have defined the state-of-the-art in image generation [12, 14, 41, 46, 51, 53], video generation [13, 47], and audio generation [15, 30] in recent years. Their primary purpose is to generate high-quality samples (images, videos, etc.). However, generation can also be interpreted as a pretext task for" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "text", + "content": "118" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "type": "text", + "content": "learning expressive features, since the model has to build up comprehensive world knowledge in order to generate plausible samples [9, 17, 27, 32, 62]. Features from diffusion models (typically referred to as diffusion features) are obtained by passing a noised image through the diffusion model and extracting intermediate feature representations. They have been shown to be useful for a variety of tasks such as finding semantic correspondences [21, 35, 62], semantic and panoptic segmentation [2, 39, 63, 64], classification [31], and object detection [8]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 193, + 294, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 193, + 294, + 312 + ], + "spans": [ + { + "bbox": [ + 55, + 193, + 294, + 312 + ], + "type": "text", + "content": "For semantic correspondence matching, features are leveraged to identify semantically matching regions across images. Existing approaches utilize diffusion features either in a zero-shot setting [21, 62] or fine-tune them for the semantic correspondence task [33, 35, 67]. Some zero-shot approaches do not fine-tune on semantic correspondence but still require tuning a prompt to activate attention maps at the query location of the correspondence [21]. In contrast, our approach aims to provide universal features usable for various down-stream tasks in a true zero-shot manner." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 313, + 294, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 313, + 294, + 433 + ], + "spans": [ + { + "bbox": [ + 55, + 313, + 294, + 433 + ], + "type": "text", + "content": "Further, diffusion features have been shown to complement features from other self-supervised learning methods such as DINOv2 [18, 43, 66]. DINOv2 features provide sparse but accurate semantic information, while Diffusion features provide dense spatial information albeit with sometimes inaccurate semantics. State-of-the-art approaches for semantic correspondence detection exploit this and fuse features [66, 67]. Compared to DINOv2 features, diffusion features yield smoother, spatially more coherent correspondences [66]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 446, + 295, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 446, + 295, + 567 + ], + "spans": [ + { + "bbox": [ + 55, + 446, + 295, + 567 + ], + "type": "text", + "content": "Distillation for Diffusion Models Knowledge Distillation [22] is a technique used to distill knowledge from a teacher model into a student model. In the context of diffusion models, this is typically applied either to reduce the required denoising steps [54] or to distill a classifier-free guided [23] model into one without CFG [36]. While our approach is inspired by distillation, we consolidate features from " + }, + { + "bbox": [ + 55, + 446, + 295, + 567 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 446, + 295, + 567 + ], + "type": "text", + "content": " different models, with " + }, + { + "bbox": [ + 55, + 446, + 295, + 567 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 446, + 295, + 567 + ], + "type": "text", + "content": " being the number of discrete diffusion timesteps, because the features are different at every timestep." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 578, + 111, + 591 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 578, + 111, + 591 + ], + "spans": [ + { + "bbox": [ + 55, + 578, + 111, + 591 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 599, + 141, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 599, + 141, + 611 + ], + "spans": [ + { + "bbox": [ + 55, + 599, + 141, + 611 + ], + "type": "text", + "content": "3.1. Preliminaries" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "text", + "content": "Diffusion Models Diffusion models are trained to predict a clean image " + }, + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0" + }, + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "text", + "content": " given a noisy image " + }, + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t" + }, + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "text", + "content": ", either explicitly or implicitly. The noisy image " + }, + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t" + }, + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "text", + "content": " is a weighted sum with random Gaussian noise " + }, + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t = \\sqrt{\\alpha_t}\\mathbf{x}_0 + \\sqrt{1 - \\alpha_t}\\epsilon" + }, + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "text", + "content": " with timestep-dependent coefficients " + }, + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\alpha_{t}" + }, + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "text", + "content": " and noise " + }, + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\epsilon \\sim \\mathcal{N}(0,\\mathbf{I})" + }, + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "inline_equation", + "content": "t\\in [0,T]" + }, + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "text", + "content": " denotes the time step of the diffusion process, with " + }, + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "inline_equation", + "content": "t = 0" + }, + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "text", + "content": " corresponding to the clean image and " + }, + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "inline_equation", + "content": "t = T" + }, + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "text", + "content": " corresponding to pure noise." + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 318, + 70, + 548, + 141 + ], + "blocks": [ + { + "bbox": [ + 318, + 70, + 548, + 141 + ], + "lines": [ + { + "bbox": [ + 318, + 70, + 548, + 141 + ], + "spans": [ + { + "bbox": [ + 318, + 70, + 548, + 141 + ], + "type": "image", + "image_path": "e5341a1d6c39dabfda046b3f20d68749e7446c43b0d651440d9de50170ee264d.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 148, + 555, + 205 + ], + "lines": [ + { + "bbox": [ + 313, + 148, + 555, + 205 + ], + "spans": [ + { + "bbox": [ + 313, + 148, + 555, + 205 + ], + "type": "text", + "content": "Figure 3. Fraction of variance of diffusion features explained by 1) encoding the clean image at " + }, + { + "bbox": [ + 313, + 148, + 555, + 205 + ], + "type": "inline_equation", + "content": "t = 0" + }, + { + "bbox": [ + 313, + 148, + 555, + 205 + ], + "type": "text", + "content": " (no additive noise), and 2) encoding just the added noise " + }, + { + "bbox": [ + 313, + 148, + 555, + 205 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 313, + 148, + 555, + 205 + ], + "type": "text", + "content": " at " + }, + { + "bbox": [ + 313, + 148, + 555, + 205 + ], + "type": "inline_equation", + "content": "t = 999" + }, + { + "bbox": [ + 313, + 148, + 555, + 205 + ], + "type": "text", + "content": ". Even at relatively low timesteps such as " + }, + { + "bbox": [ + 313, + 148, + 555, + 205 + ], + "type": "inline_equation", + "content": "t = 261" + }, + { + "bbox": [ + 313, + 148, + 555, + 205 + ], + "type": "text", + "content": " as used by DIFT [62], a substantial part of the features directly depends only on the added noise." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 225, + 555, + 345 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 225, + 555, + 345 + ], + "spans": [ + { + "bbox": [ + 313, + 225, + 555, + 345 + ], + "type": "text", + "content": "Intuitively, the model faces different objectives for different noise levels [1, 4]: for very high noise, there is little information in the input, and the model is first generating the coarse structure of the image [50]. At lower noise levels, more high- and medium-frequency information is available and the task shifts to generating finer details and intricate structures. This multi-objective nature intuitively explains why previous methods found diffusion features extracted from different timesteps to provide information with differing semantics." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 361, + 556, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 361, + 556, + 506 + ], + "spans": [ + { + "bbox": [ + 313, + 361, + 556, + 506 + ], + "type": "text", + "content": "Diffusion Feature Extraction Typically, diffusion feature extraction happens after first adding noise to an image and passing the resulting " + }, + { + "bbox": [ + 313, + 361, + 556, + 506 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t" + }, + { + "bbox": [ + 313, + 361, + 556, + 506 + ], + "type": "text", + "content": " to a U-Net [52] denoiser. Features are then extracted at multiple hand-picked locations of the U-Net decoder [2, 39, 62, 66, 67]. Different levels of noise added to the input image result in features beneficial for different downstream applications. Typical diffusion timesteps are " + }, + { + "bbox": [ + 313, + 361, + 556, + 506 + ], + "type": "inline_equation", + "content": "t = 261" + }, + { + "bbox": [ + 313, + 361, + 556, + 506 + ], + "type": "text", + "content": " [62], " + }, + { + "bbox": [ + 313, + 361, + 556, + 506 + ], + "type": "inline_equation", + "content": "t = 100" + }, + { + "bbox": [ + 313, + 361, + 556, + 506 + ], + "type": "text", + "content": " [66] or " + }, + { + "bbox": [ + 313, + 361, + 556, + 506 + ], + "type": "inline_equation", + "content": "t = 50" + }, + { + "bbox": [ + 313, + 361, + 556, + 506 + ], + "type": "text", + "content": " [67]. By adding noise to the input image, these methods bottleneck the perceptual information the model can extract. To illustrate this, we show Stable Diffusion 2.1's reconstruction of an image noised at " + }, + { + "bbox": [ + 313, + 361, + 556, + 506 + ], + "type": "inline_equation", + "content": "t = 261" + }, + { + "bbox": [ + 313, + 361, + 556, + 506 + ], + "type": "text", + "content": " [62] in Figure 2." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 521, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 556, + 713 + ], + "type": "text", + "content": "Diffusion Features Encode Noise We hypothesize that diffusion features extracted from noisy images " + }, + { + "bbox": [ + 313, + 521, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t" + }, + { + "bbox": [ + 313, + 521, + 556, + 713 + ], + "type": "text", + "content": " encode the noise " + }, + { + "bbox": [ + 313, + 521, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 313, + 521, + 556, + 713 + ], + "type": "text", + "content": " in addition to information from the image. We investigate this hypothesis in a very simple setting by examining how well features " + }, + { + "bbox": [ + 313, + 521, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{feat}(\\epsilon; T)" + }, + { + "bbox": [ + 313, + 521, + 556, + 713 + ], + "type": "text", + "content": " extracted from only the noise " + }, + { + "bbox": [ + 313, + 521, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 313, + 521, + 556, + 713 + ], + "type": "text", + "content": " approximate the features " + }, + { + "bbox": [ + 313, + 521, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{feat}(\\mathbf{x}_t = \\sqrt{\\alpha_t} \\mathbf{x}_0 + \\sqrt{1 - \\alpha_t} \\epsilon; t)" + }, + { + "bbox": [ + 313, + 521, + 556, + 713 + ], + "type": "text", + "content": ". Using least squares, we fit a single scalar approximation coefficient to obtain the optimal reconstruction. We then quantify how much of the variance of the overall features is explained by this approximation (cf. Figure 3). Even at relatively low timesteps, such as " + }, + { + "bbox": [ + 313, + 521, + 556, + 713 + ], + "type": "inline_equation", + "content": "t = 261" + }, + { + "bbox": [ + 313, + 521, + 556, + 713 + ], + "type": "text", + "content": " used by DIFT [62], encoding pure noise explains a substantial fraction of the features' variance. Current diffusion feature methods extract this information jointly with the image information. Our proposed method addresses this issue by eliminating the noise from the feature extraction process." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "119" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 74, + 298, + 198 + ], + "blocks": [ + { + "bbox": [ + 56, + 74, + 298, + 198 + ], + "lines": [ + { + "bbox": [ + 56, + 74, + 298, + 198 + ], + "spans": [ + { + "bbox": [ + 56, + 74, + 298, + 198 + ], + "type": "image", + "image_path": "8908f26484b3c0741935d98d5d3b70f96a8d408bdff2ef9ca61262240b645678.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 209, + 296, + 276 + ], + "lines": [ + { + "bbox": [ + 55, + 209, + 296, + 276 + ], + "spans": [ + { + "bbox": [ + 55, + 209, + 296, + 276 + ], + "type": "text", + "content": "Figure 4. Our training setup. We train our model to predict features from a clean input image, while the frozen diffusion model is fed the noisy image. The projection heads project our model's features onto the noisy diffusion model features, given the noising timestep " + }, + { + "bbox": [ + 55, + 209, + 296, + 276 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 209, + 296, + 276 + ], + "type": "text", + "content": ". For downstream tasks, we discard the projection heads and directly use our model's internal representations as features." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 286, + 296, + 406 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 286, + 296, + 406 + ], + "spans": [ + { + "bbox": [ + 55, + 286, + 296, + 406 + ], + "type": "text", + "content": "We further analyze the residual and similarly decompose it via the features predicted for the clean image " + }, + { + "bbox": [ + 55, + 286, + 296, + 406 + ], + "type": "inline_equation", + "content": "\\mathrm{feat}(x_t = x_0; t = 0)" + }, + { + "bbox": [ + 55, + 286, + 296, + 406 + ], + "type": "text", + "content": ". We find that they do not fully explain the remainder of the features either. Instead, a substantial part of the feature variance at medium noise timesteps is timestep dependent and cannot be attributed to components present at " + }, + { + "bbox": [ + 55, + 286, + 296, + 406 + ], + "type": "inline_equation", + "content": "t = 0" + }, + { + "bbox": [ + 55, + 286, + 296, + 406 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 55, + 286, + 296, + 406 + ], + "type": "inline_equation", + "content": "t = T" + }, + { + "bbox": [ + 55, + 286, + 296, + 406 + ], + "type": "text", + "content": ". This matches observations by previous works [62] that found diffusion features at higher timesteps offer better semantics, despite the added noise at the image input." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 415, + 285, + 427 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 415, + 285, + 427 + ], + "spans": [ + { + "bbox": [ + 55, + 415, + 285, + 427 + ], + "type": "text", + "content": "3.2. CleanDIFT: Noise-Free Diffusion Features" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 433, + 296, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 433, + 296, + 494 + ], + "spans": [ + { + "bbox": [ + 55, + 433, + 296, + 494 + ], + "type": "text", + "content": "We present CleanDIFT, our method to address the problem of noisy and time-dependent diffusion features. CleanDIFT extracts clean Diffusion FeaTures from a pretrained diffusion backbone through a lightweight fine-tuning process. An overview of our setup is shown in Figure 4." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 510, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 296, + 713 + ], + "type": "text", + "content": "Extraction Setup We train our feature extraction model to match the diffusion model's internal representations. We initialize the feature extraction model as a trainable copy of the diffusion model. Crucially, the feature extraction model is given the clean input image, while the diffusion model receives the noisy image and the corresponding timestep as input. Our goal is to obtain a single, noise-free feature map from the feature extraction model that consolidates the information of the diffusion model's timestep-dependent internal representations into a single one. To align our model's representations with the timestep-dependent diffusion model features during training, we introduce pointwise timestep-conditioned feature projection heads. The feature maps predicted by these projection heads are then aligned to the diffusion model's features. For feature extraction at inference time, we usually discard the projection heads and directly use the feature extraction model's" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "text", + "content": "internal representations. However, the projection heads can also be used to efficiently obtain feature maps for specific timesteps by reusing the feature extraction model's internal representations and passing them through the projection heads for different " + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "text", + "content": " values." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 157, + 555, + 241 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 157, + 555, + 241 + ], + "spans": [ + { + "bbox": [ + 313, + 157, + 555, + 241 + ], + "type": "text", + "content": "Training Objective We regard the diffusion model as a family of feature extraction functions " + }, + { + "bbox": [ + 313, + 157, + 555, + 241 + ], + "type": "inline_equation", + "content": "\\operatorname{feat}(\\cdot, \\epsilon, t)" + }, + { + "bbox": [ + 313, + 157, + 555, + 241 + ], + "type": "text", + "content": " for timestep " + }, + { + "bbox": [ + 313, + 157, + 555, + 241 + ], + "type": "inline_equation", + "content": "t \\in [1,999]" + }, + { + "bbox": [ + 313, + 157, + 555, + 241 + ], + "type": "text", + "content": " and noise " + }, + { + "bbox": [ + 313, + 157, + 555, + 241 + ], + "type": "inline_equation", + "content": "\\epsilon \\sim \\mathcal{N}(0, \\mathbf{I})" + }, + { + "bbox": [ + 313, + 157, + 555, + 241 + ], + "type": "text", + "content": ". Each of these functions maps an image " + }, + { + "bbox": [ + 313, + 157, + 555, + 241 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 313, + 157, + 555, + 241 + ], + "type": "text", + "content": " to a feature vector " + }, + { + "bbox": [ + 313, + 157, + 555, + 241 + ], + "type": "inline_equation", + "content": "\\operatorname{feat}(\\mathbf{x}, \\epsilon, t)" + }, + { + "bbox": [ + 313, + 157, + 555, + 241 + ], + "type": "text", + "content": ". We aim to consolidate the information provided by all feature extraction functions into a single joint function " + }, + { + "bbox": [ + 313, + 157, + 555, + 241 + ], + "type": "inline_equation", + "content": "\\operatorname{feat}_c(\\cdot)" + }, + { + "bbox": [ + 313, + 157, + 555, + 241 + ], + "type": "text", + "content": " with the same dimensionality:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 349, + 243, + 526, + 316 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 349, + 243, + 526, + 316 + ], + "spans": [ + { + "bbox": [ + 349, + 243, + 526, + 316 + ], + "type": "interline_equation", + "content": "\\begin{array}{c} \\underline {{\\text {S t a b l e D i f f u s i o n}}} \\\\ \\left. \\begin{array}{l} \\operatorname {f e a t} (\\mathbf {x}, \\boldsymbol {\\epsilon}, t = 1) \\\\ \\operatorname {f e a t} (\\mathbf {x}, \\boldsymbol {\\epsilon}, t = 2) \\\\ \\vdots \\\\ \\operatorname {f e a t} (\\mathbf {x}, \\boldsymbol {\\epsilon}, t = 9 9 9) \\end{array} \\right\\} \\xrightarrow [ \\text {i n f o r m a t i o n} ]{\\text {c o n s o l i d a t e}} \\operatorname {f e a t} _ {\\mathrm {c}} (\\mathbf {x}) \\end{array}", + "image_path": "fefa0c4f97815574b9aa94f805bc68e29a567ff25ed3cc1019e6b0252485b1c5.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 323, + 555, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 323, + 555, + 370 + ], + "spans": [ + { + "bbox": [ + 313, + 323, + 555, + 370 + ], + "type": "text", + "content": "To align our model's features with the diffusion model's features, we maximize the similarity between the diffusion model's features and the projected features of our feature extraction model:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 318, + 372, + 544, + 423 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 372, + 544, + 423 + ], + "spans": [ + { + "bbox": [ + 318, + 372, + 544, + 423 + ], + "type": "interline_equation", + "content": "\\operatorname {f e a t} _ {\\mathrm {c}} (\\mathbf {x}) \\to \\left\\{ \\begin{array}{l} \\operatorname {p r o j} (\\cdot , t = 1) \\longrightarrow \\mathcal {L} \\longleftarrow \\operatorname {f e a t} (\\mathbf {x}, \\epsilon , t = 1) \\\\ \\operatorname {p r o j} (\\cdot , t = 2) \\longrightarrow \\mathcal {L} \\longleftarrow \\operatorname {f e a t} (\\mathbf {x}, \\epsilon , t = 2) \\\\ \\vdots \\\\ \\operatorname {p r o j} (\\cdot , t = 9 9 9) \\longrightarrow \\mathcal {L} \\longleftarrow \\operatorname {f e a t} (\\mathbf {x}, \\epsilon , t = 9 9 9) \\end{array} \\right.", + "image_path": "ede9dd98fbd0a1916eb76f53e62b82695ca6fef7d7722a476ff09eb620287654.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 431, + 555, + 587 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 431, + 555, + 587 + ], + "spans": [ + { + "bbox": [ + 313, + 431, + 555, + 587 + ], + "type": "text", + "content": "Specifically, we minimize the negative cosine similarity between the diffusion model's features and our model's features extracted at stages " + }, + { + "bbox": [ + 313, + 431, + 555, + 587 + ], + "type": "inline_equation", + "content": "k = \\{1,\\dots,K\\}" + }, + { + "bbox": [ + 313, + 431, + 555, + 587 + ], + "type": "text", + "content": " in the network. Given a clean image " + }, + { + "bbox": [ + 313, + 431, + 555, + 587 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0" + }, + { + "bbox": [ + 313, + 431, + 555, + 587 + ], + "type": "text", + "content": ", the feature extraction model's output for feature map " + }, + { + "bbox": [ + 313, + 431, + 555, + 587 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 431, + 555, + 587 + ], + "type": "text", + "content": " is denoted as " + }, + { + "bbox": [ + 313, + 431, + 555, + 587 + ], + "type": "inline_equation", + "content": "\\mathrm{feat}_{\\mathrm{c}}^{(k)}(\\mathbf{x}_0)" + }, + { + "bbox": [ + 313, + 431, + 555, + 587 + ], + "type": "text", + "content": ". Our Clean-DIFT feature map is then adapted by the learned projection heads " + }, + { + "bbox": [ + 313, + 431, + 555, + 587 + ], + "type": "inline_equation", + "content": "\\mathrm{proj}^{(k)}(\\mathrm{feat}_{\\mathrm{c}}^{(k)}(\\mathbf{x}_0),t)" + }, + { + "bbox": [ + 313, + 431, + 555, + 587 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 431, + 555, + 587 + ], + "type": "inline_equation", + "content": "\\mathrm{proj}^{(k)}(\\cdot ,\\cdot)" + }, + { + "bbox": [ + 313, + 431, + 555, + 587 + ], + "type": "text", + "content": " is the projection head for feature map " + }, + { + "bbox": [ + 313, + 431, + 555, + 587 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 431, + 555, + 587 + ], + "type": "text", + "content": ". The diffusion model receives the noisy image " + }, + { + "bbox": [ + 313, + 431, + 555, + 587 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t" + }, + { + "bbox": [ + 313, + 431, + 555, + 587 + ], + "type": "text", + "content": " corresponding to the same " + }, + { + "bbox": [ + 313, + 431, + 555, + 587 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0" + }, + { + "bbox": [ + 313, + 431, + 555, + 587 + ], + "type": "text", + "content": " and timestep " + }, + { + "bbox": [ + 313, + 431, + 555, + 587 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 431, + 555, + 587 + ], + "type": "text", + "content": ". The projection head then learns a timestep-dependent alignment from CleanDIFT features to the diffusion model's features " + }, + { + "bbox": [ + 313, + 431, + 555, + 587 + ], + "type": "inline_equation", + "content": "\\mathrm{feat}^{(k)}(\\mathbf{x}_t;t)" + }, + { + "bbox": [ + 313, + 431, + 555, + 587 + ], + "type": "text", + "content": ". Putting it all together, our loss function is defined as:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 319, + 597, + 555, + 630 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 597, + 555, + 630 + ], + "spans": [ + { + "bbox": [ + 319, + 597, + 555, + 630 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = - \\sum_ {k = 1} ^ {K} \\operatorname {s i m} \\left(\\operatorname {p r o j} ^ {(k)} \\left(\\operatorname {f e a t} _ {\\mathrm {c}} ^ {(k)} \\left(\\mathbf {x} _ {0}\\right); t\\right), \\operatorname {f e a t} ^ {(k)} \\left(\\mathbf {x} _ {t}; t\\right)\\right). \\tag {1}", + "image_path": "ddb767340a0044a518a46958a3b2ae45f26a12ff70d9049b07b08e9e2c126821.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "text", + "content": "For each training image " + }, + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0" + }, + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "text", + "content": ", we sample " + }, + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "text", + "content": " different nois-ing timesteps " + }, + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "inline_equation", + "content": "t_i" + }, + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "text", + "content": " in a stratified manner, with each timestep " + }, + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "inline_equation", + "content": "t_i \\sim \\mathcal{U}(\\frac{i}{I} T, \\frac{i+1}{I} T)" + }, + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "text", + "content": " is the maximum timestep. By sampling multiple timesteps per image we incentivize the feature extraction model to match the diffusion model's features across the entire noise spectrum." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "text", + "content": "120" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 71, + 137, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 71, + 137, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 137, + 85 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 92, + 296, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 92, + 296, + 224 + ], + "spans": [ + { + "bbox": [ + 55, + 92, + 296, + 224 + ], + "type": "text", + "content": "We test our hypothesis that the proposed extraction setup enables us to leverage more of the world knowledge inherent in diffusion models compared to existing diffusion feature extraction methods while being task-agnostic and timestep-independent. To that end, we evaluate our features on a wide range of downstream tasks: unsupervised zero-shot semantic correspondence, monocular depth estimation, semantic segmentation, and classification. We compare our features against standard diffusion features, methods that combine diffusion features with additional features, and non-diffusion-based approaches." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 233, + 171, + 247 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 233, + 171, + 247 + ], + "spans": [ + { + "bbox": [ + 55, + 233, + 171, + 247 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 251, + 296, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 251, + 296, + 552 + ], + "spans": [ + { + "bbox": [ + 55, + 251, + 296, + 552 + ], + "type": "text", + "content": "Implementation Details Following previous works [62, 66, 67], we evaluate our method on a Stable Diffusion (SD) backbone [51]. We apply our method to SD 1.5 and SD 2.1 to enable fair comparisons with existing methods that use either. We fully fine-tune our feature extraction model on image-caption pairs for only 400 steps, taking 30 minutes on a single A100 GPU, which was sufficient for our strong performance and further training did not yield any significant gains. We extract features after the U-Net's middle block and after each of the U-Net's decoder blocks, except the two final blocks. A detailed visualization of where we extract features is provided in Figure 11. This yields a total of " + }, + { + "bbox": [ + 55, + 251, + 296, + 552 + ], + "type": "inline_equation", + "content": "K = 11" + }, + { + "bbox": [ + 55, + 251, + 296, + 552 + ], + "type": "text", + "content": " feature maps that we align between the diffusion model and the feature extraction model. Our point-wise feature projection heads consist of three stacked Feed Forward Networks (FFNs) that are zero-initialized such that initially they act as identity mappings due to their residual connections. Since every aligned feature map has its own projection head, this results in 45M additional trainable parameters for SD 2.1. We study the effect of different projection head architectures in Sec. E. We train using Adam [29] with a batch size of 8 and a learning rate of 2e-6 with a linear warmup. For stratified timestep sampling, we utilize " + }, + { + "bbox": [ + 55, + 251, + 296, + 552 + ], + "type": "inline_equation", + "content": "I = 3" + }, + { + "bbox": [ + 55, + 251, + 296, + 552 + ], + "type": "text", + "content": " stratification bins across all our experiments, i.e. three different noise levels per training image." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "content": "Datasets We fine-tune our feature extraction model on a random subset of COYO-700M [5], which is similar to the LAION [56], the dataset that Stable Diffusion 1.5 and 2.1 were trained on originally. That way, we ensure that all performance improvements originate from the feature extraction model consolidating the diffusion model's internal feature representations over time, not from choosing a different dataset that matches the test dataset distribution more closely. The subset selects images with a minimum size of " + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "inline_equation", + "content": "512^2" + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "content": ". We crop and resize them to match the corresponding input resolution of the underlying diffusion model. We analyze the effect of using different datasets in Sec. G." + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 315, + 69, + 555, + 230 + ], + "blocks": [ + { + "bbox": [ + 315, + 69, + 555, + 230 + ], + "lines": [ + { + "bbox": [ + 315, + 69, + 555, + 230 + ], + "spans": [ + { + "bbox": [ + 315, + 69, + 555, + 230 + ], + "type": "image", + "image_path": "7fbfb9bdf92425f5ea83b84afada92433835a9568da4e6ec14b62de7ad65dd8d.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 238, + 555, + 281 + ], + "lines": [ + { + "bbox": [ + 313, + 238, + 555, + 281 + ], + "spans": [ + { + "bbox": [ + 313, + 238, + 555, + 281 + ], + "type": "text", + "content": "Figure 5. Semantic correspondence results using DIFT [62] features with the standard SD 2.1 (" + }, + { + "bbox": [ + 313, + 238, + 555, + 281 + ], + "type": "inline_equation", + "content": "t = 261" + }, + { + "bbox": [ + 313, + 238, + 555, + 281 + ], + "type": "text", + "content": ") and our CleanDIFT features. Our clean features show significantly less incorrect matches than the base diffusion model." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 292, + 526, + 304 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 292, + 526, + 304 + ], + "spans": [ + { + "bbox": [ + 313, + 292, + 526, + 304 + ], + "type": "text", + "content": "4.2. Unsupervised Semantic Correspondence" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 309, + 555, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 309, + 555, + 536 + ], + "spans": [ + { + "bbox": [ + 313, + 309, + 555, + 536 + ], + "type": "text", + "content": "As many previous diffusion feature methods focus on (unsupervised) semantic correspondence matching [35, 62, 66, 67], we perform an extensive evaluation of our method on this task. Following previous works on semantic correspondence matching [62, 66, 67], we measure our performance in Percentage of Correct Keypoints (PCK). We average PCK directly across all keypoints, not over images. We use " + }, + { + "bbox": [ + 313, + 309, + 555, + 536 + ], + "type": "inline_equation", + "content": "\\alpha = 0.1" + }, + { + "bbox": [ + 313, + 309, + 555, + 536 + ], + "type": "text", + "content": " as a threshold and report both PCK values with error margins relative to the image size and to the bounding box size, denoted as " + }, + { + "bbox": [ + 313, + 309, + 555, + 536 + ], + "type": "inline_equation", + "content": "\\mathrm{PCK}_{\\mathrm{img}}" + }, + { + "bbox": [ + 313, + 309, + 555, + 536 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 309, + 555, + 536 + ], + "type": "inline_equation", + "content": "\\mathrm{PCK}_{\\mathrm{bbox}}" + }, + { + "bbox": [ + 313, + 309, + 555, + 536 + ], + "type": "text", + "content": " respectively. We evaluate the performance on the test split of the SPair-71k dataset, which consists of approximately " + }, + { + "bbox": [ + 313, + 309, + 555, + 536 + ], + "type": "inline_equation", + "content": "12\\mathrm{k}" + }, + { + "bbox": [ + 313, + 309, + 555, + 536 + ], + "type": "text", + "content": " image pairs from 18 categories. Some existing works [62, 66, 67] evaluate on additional datasets but find SPair-71k to be the most challenging and therefore the most informative benchmark. For the text prompt we use \"A photo of a category.\", with category being the corresponding category of the SPair image. We experiment with distilling the text conditioning in Sec. B." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 546, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 546, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 546, + 556, + 714 + ], + "type": "text", + "content": "Results We first compare our extracted features to DIFT [62], an approach that detects semantic correspondences using standard diffusion features. Substituting these with our CleanDIFT features yields a performance increase of 1.79 absolute percentage points for " + }, + { + "bbox": [ + 313, + 546, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{PCK}_{\\mathrm{img}}" + }, + { + "bbox": [ + 313, + 546, + 556, + 714 + ], + "type": "text", + "content": " and 1.86 percentage points for " + }, + { + "bbox": [ + 313, + 546, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{PCK}_{\\mathrm{bbox}}" + }, + { + "bbox": [ + 313, + 546, + 556, + 714 + ], + "type": "text", + "content": ". Notably, DIFT averages the extracted feature maps across 8 different noise samples. Without this averaging over noise samples, our performance gain is even larger (2.81 PCK@ " + }, + { + "bbox": [ + 313, + 546, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\alpha_{\\mathrm{img}}" + }, + { + "bbox": [ + 313, + 546, + 556, + 714 + ], + "type": "text", + "content": " gain). This indicates that our feature extraction model learns more than a mere averaging over the noise in the diffusion model's feature maps (see Figure 5 and Sec. C for examples). We present an extended version of the time-step dependent performance analysis conducted by [62] in Figure 6: We evaluate the" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "121" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 66, + 70, + 282, + 137 + ], + "blocks": [ + { + "bbox": [ + 66, + 70, + 282, + 137 + ], + "lines": [ + { + "bbox": [ + 66, + 70, + 282, + 137 + ], + "spans": [ + { + "bbox": [ + 66, + 70, + 282, + 137 + ], + "type": "image", + "image_path": "b04caf2667ee539939992132bc059f285b3dc5ff58fc142abc4636fae9c086b6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 146, + 295, + 213 + ], + "lines": [ + { + "bbox": [ + 55, + 146, + 295, + 213 + ], + "spans": [ + { + "bbox": [ + 55, + 146, + 295, + 213 + ], + "type": "text", + "content": "Figure 6. Following [62], we evaluate semantic correspondence matching accuracy for different noise levels. Our feature extractor outperforms the standard noisy diffusion features across all timesteps " + }, + { + "bbox": [ + 55, + 146, + 295, + 213 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 146, + 295, + 213 + ], + "type": "text", + "content": ". We additionally demonstrate that simply providing the diffusion model with a clean image and a non-zero timestep does not result in improved performance." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 221, + 295, + 412 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 221, + 295, + 412 + ], + "spans": [ + { + "bbox": [ + 54, + 221, + 295, + 412 + ], + "type": "text", + "content": "diffusion model's performance for different timesteps " + }, + { + "bbox": [ + 54, + 221, + 295, + 412 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 54, + 221, + 295, + 412 + ], + "type": "text", + "content": " in two settings. In the first setting, we provide the diffusion model with a noisy input image " + }, + { + "bbox": [ + 54, + 221, + 295, + 412 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t" + }, + { + "bbox": [ + 54, + 221, + 295, + 412 + ], + "type": "text", + "content": " as usual. In the second setting, we demonstrate that feeding the clean image along with a non-zero timestep is not a viable solution to obtain meaningful features: We provide the diffusion model with the clean input image " + }, + { + "bbox": [ + 54, + 221, + 295, + 412 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0" + }, + { + "bbox": [ + 54, + 221, + 295, + 412 + ], + "type": "text", + "content": " for all timesteps " + }, + { + "bbox": [ + 54, + 221, + 295, + 412 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 54, + 221, + 295, + 412 + ], + "type": "text", + "content": ". We observe that the model's performance for the clean input image degrades faster and has a lower peak than for the noisy input. This is to be expected, as the diffusion model was trained on noisy images, not clean images. Importantly, our CleanDIFT features are timestep-independent and consistently outperform standard diffusion features, even when the latter are optimized for the best-performing timestep. We further observe that this advantage generalizes effectively to other backbones, such as DiTs [44] (see Tab. 7)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 413, + 295, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 413, + 295, + 617 + ], + "spans": [ + { + "bbox": [ + 55, + 413, + 295, + 617 + ], + "type": "text", + "content": "A Tale of Two Features [66] extends the approach of DIFT by combining diffusion features with DINOv2 [43] features. Again, we replace the standard diffusion features with our CleanDIFT features and observe that the performance gain transfers when combining our features with DINOv2 features. Telling Left from Right [67] further improves upon the results of A Tale of Two Features by introducing a test-time adaptive pose alignment strategy. We observe that the performance gain transfers to this setting as well. To the best of our knowledge, Telling Left from Right combined with our CleanDIFT features sets a new state-of-the-art in unsupervised zero-shot semantic correspondence matching. In summary, replacing standard diffusion features with our CleanDIFT features consistently results in a significant performance improvement across all three methods. We show an overview of the results in Tab. 1 and a more extensive evaluation per category in Sec. B." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 618, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 295, + 715 + ], + "type": "text", + "content": "We also investigate the performance of our features in a supervised fine-tuning setting for semantic correspondence matching. Following [35], we train an aggregation network that uses all extracted feature maps and learns to aggregate them into a single task-specific feature map for semantic correspondence matching. In contrast to [35], we do not have to perform costly DDIM inversion [59] to obtain a matching noisy image for every timestep. Instead," + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 317, + 70, + 553, + 198 + ], + "blocks": [ + { + "bbox": [ + 317, + 70, + 553, + 198 + ], + "lines": [ + { + "bbox": [ + 317, + 70, + 553, + 198 + ], + "spans": [ + { + "bbox": [ + 317, + 70, + 553, + 198 + ], + "type": "table", + "html": "
MethodOur FeaturesPCK@α(↑)
αimg=0.1αbbox=0.1
General Approaches
DINOv2+NN--55.6
Diff. Feat.-based Approaches
DIFT [62]X66.5359.57
✓68.32▲1.7961.43▲1.86
A Tale of Two Features [66]X72.3163.73
✓73.35▲1.0464.81▲1.08
Telling Left from Right [67]X77.0768.64
✓78.40▲1.3369.99▲1.35
", + "image_path": "490a10983a65771719747da613bfc460abf075446cc427544c1f35a54cdc06f7.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 205, + 555, + 271 + ], + "lines": [ + { + "bbox": [ + 313, + 205, + 555, + 271 + ], + "spans": [ + { + "bbox": [ + 313, + 205, + 555, + 271 + ], + "type": "text", + "content": "Table 1. Zero-shot unsupervised semantic correspondence matching performance comparison on SPair71k [37]. Our improved features consistently lead to substantial improvements in matching performance. We report PCK on the test split of SPair71k, aggregated per point. Numbers are reproduced, for a discussion and comparison to reported numbers view Tab. 5." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 281, + 555, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 281, + 555, + 437 + ], + "spans": [ + { + "bbox": [ + 313, + 281, + 555, + 437 + ], + "type": "text", + "content": "we directly feed the clean image to our feature extraction model. Therefore, extracting features with our CleanDIFT approach is " + }, + { + "bbox": [ + 313, + 281, + 555, + 437 + ], + "type": "inline_equation", + "content": "50\\mathrm{x}" + }, + { + "bbox": [ + 313, + 281, + 555, + 437 + ], + "type": "text", + "content": " faster, since we perform a single denoiser forward pass while [35] performs 50 for the inversion. Our model achieves a " + }, + { + "bbox": [ + 313, + 281, + 555, + 437 + ], + "type": "inline_equation", + "content": "\\mathrm{PCK}_{\\mathrm{img}}" + }, + { + "bbox": [ + 313, + 281, + 555, + 437 + ], + "type": "text", + "content": " value of 72.48 vs their 72.75 and a " + }, + { + "bbox": [ + 313, + 281, + 555, + 437 + ], + "type": "inline_equation", + "content": "\\mathrm{PCK}_{\\mathrm{bbox}}" + }, + { + "bbox": [ + 313, + 281, + 555, + 437 + ], + "type": "text", + "content": " value of 64.37 vs their 64.53. We observe a slight performance regression compared to their approach, however, at a speedup of " + }, + { + "bbox": [ + 313, + 281, + 555, + 437 + ], + "type": "inline_equation", + "content": "50\\times" + }, + { + "bbox": [ + 313, + 281, + 555, + 437 + ], + "type": "text", + "content": ". Luo et al. [35] also present a single-step ablation of their full method that only requires a single forward pass which makes it more comparable to ours. We outperform this single-step version by a wide margin of 9.0 percentage points for " + }, + { + "bbox": [ + 313, + 281, + 555, + 437 + ], + "type": "inline_equation", + "content": "\\mathrm{PCK}_{\\mathrm{img}}" + }, + { + "bbox": [ + 313, + 281, + 555, + 437 + ], + "type": "text", + "content": " and 9.1 percentage points for " + }, + { + "bbox": [ + 313, + 281, + 555, + 437 + ], + "type": "inline_equation", + "content": "\\mathrm{PCK}_{\\mathrm{bbox}}" + }, + { + "bbox": [ + 313, + 281, + 555, + 437 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 445, + 420, + 456 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 445, + 420, + 456 + ], + "spans": [ + { + "bbox": [ + 313, + 445, + 420, + 456 + ], + "type": "text", + "content": "4.3. Depth Estimation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 462, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 462, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 462, + 555, + 715 + ], + "type": "text", + "content": "We also investigate monocular depth estimation on NYUv2 [40]. Similar to [43], we follow the evaluation protocol from [34]. We use SD 2.1 as the base model and resize the input to the model's native resolution of " + }, + { + "bbox": [ + 313, + 462, + 555, + 715 + ], + "type": "inline_equation", + "content": "768^2" + }, + { + "bbox": [ + 313, + 462, + 555, + 715 + ], + "type": "text", + "content": ". We extract features from the same location as [62] and obtain a feature map of dimension " + }, + { + "bbox": [ + 313, + 462, + 555, + 715 + ], + "type": "inline_equation", + "content": "48^2" + }, + { + "bbox": [ + 313, + 462, + 555, + 715 + ], + "type": "text", + "content": ". Unlike [43], we do not up-sample the features and directly apply the linear probe. The probe predicts depth in 256 uniform bins which we combine with a classification loss after a linear normalization following [3]. We train one probe for our CleanDIFT features and one for standard diffusion features at " + }, + { + "bbox": [ + 313, + 462, + 555, + 715 + ], + "type": "inline_equation", + "content": "t = 299" + }, + { + "bbox": [ + 313, + 462, + 555, + 715 + ], + "type": "text", + "content": ", as that timestep minimizes the error in our settings. Our qualitative results (see Figure 7) show a substantial fidelity gap in the estimated depth maps between the features from the standard SD 2.1 backbone and the features from our feature extraction model. This is reflected in a substantial improvement in quantitative metrics over the baseline as seen in Tab. 2. Lastly, we reuse the probe trained on standard diffusion features and apply it on the CleanDIFT features. While this does not match the performance of the CleanDIFT probe, it still achieves significantly better re" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "122" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 68, + 291, + 208 + ], + "blocks": [ + { + "bbox": [ + 61, + 68, + 291, + 208 + ], + "lines": [ + { + "bbox": [ + 61, + 68, + 291, + 208 + ], + "spans": [ + { + "bbox": [ + 61, + 68, + 291, + 208 + ], + "type": "image", + "image_path": "4b075a4acf8c4fdd30068b4c36b81a1fdb99db7d64c70819c392abcf76c7c770.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 220, + 296, + 275 + ], + "lines": [ + { + "bbox": [ + 55, + 220, + 296, + 275 + ], + "spans": [ + { + "bbox": [ + 55, + 220, + 296, + 275 + ], + "type": "text", + "content": "Figure 7. Qualitative results for depth estimation using a linear probe on diffusion features on NYUv2 [40]. Our CleanDIFT features enable substantially better depth estimation than standard diffusion features. Note how the CleanDIFT features are far less noisy when compared to the standard diffusion features." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 58, + 286, + 293, + 419 + ], + "blocks": [ + { + "bbox": [ + 58, + 286, + 293, + 419 + ], + "lines": [ + { + "bbox": [ + 58, + 286, + 293, + 419 + ], + "spans": [ + { + "bbox": [ + 58, + 286, + 293, + 419 + ], + "type": "table", + "html": "
MethodBackboneRMSE (↓)
Self-Supervised Methods
OpenCLIP [28]ViT-G/140.541
MAE [20]ViT-H/140.517
DINO [6]ViT-B/80.555
iBOT [68]ViT-L/160.417
DINOv2 [43]ViT-g/140.344
Diffusion Features
SD 2.1 [51]0.469
DIFF- like [62]Ours0.444,0.025
+ Probes from noisy features0.453,0.016
", + "image_path": "6a17becec707b4c8ececa1edf230bd8099592d33341847691b41903c52c95c15.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 428, + 295, + 483 + ], + "lines": [ + { + "bbox": [ + 55, + 428, + 295, + 483 + ], + "spans": [ + { + "bbox": [ + 55, + 428, + 295, + 483 + ], + "type": "text", + "content": "Table 2. Monocular Depth Estimation. Following [43], we evaluate metric depth prediction on NYUv2 [40] using a linear probe. Our clean features outperform the noisy features by a significant margin. Probes trained on the noisy features can be reused for the clean features, but incur a smaller performance gain." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 495, + 296, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 495, + 296, + 544 + ], + "spans": [ + { + "bbox": [ + 55, + 495, + 296, + 544 + ], + "type": "text", + "content": "sults when compared to using standard diffusion features. This indicates that our features can be used as a drop-in replacement for the original diffusion features and offer improved performance on downstream applications." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 552, + 188, + 565 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 552, + 188, + 565 + ], + "spans": [ + { + "bbox": [ + 55, + 552, + 188, + 565 + ], + "type": "text", + "content": "4.4. Semantic Segmentation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "content": "To further investigate the difference between standard noisy diffusion features and our CleanDIFT features, we evaluate on the semantic segmentation task by training linear probes on our CleanDIFT features and on standard diffusion features. We utilize SD2.1 as the diffusion backbone and extract features at the same location as [62]. This procedure yields feature maps of size " + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "inline_equation", + "content": "48^2" + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "content": ". We train our linear probe on the " + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "inline_equation", + "content": "48^2" + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "content": " feature maps and upscale the obtained segmentation masks using nearest neighbor upsampling. We train and evaluate on the PASCAL VOC dataset [16]. Following common practice [39], we use mean Intersection over Union (mIoU) as the evaluation metric. Qualitative re" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 318, + 68, + 553, + 196 + ], + "blocks": [ + { + "bbox": [ + 318, + 68, + 553, + 196 + ], + "lines": [ + { + "bbox": [ + 318, + 68, + 553, + 196 + ], + "spans": [ + { + "bbox": [ + 318, + 68, + 553, + 196 + ], + "type": "image", + "image_path": "703e20298c04cd3462fa6eef64a71f464ec16a7a012ff999757a2546f93a18bf.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 205, + 555, + 261 + ], + "lines": [ + { + "bbox": [ + 313, + 205, + 555, + 261 + ], + "spans": [ + { + "bbox": [ + 313, + 205, + 555, + 261 + ], + "type": "text", + "content": "Figure 8. Qualitative results for semantic segmentation from diffusion features on Pascal VOC [16]. Standard SD features use " + }, + { + "bbox": [ + 313, + 205, + 555, + 261 + ], + "type": "inline_equation", + "content": "t = 100" + }, + { + "bbox": [ + 313, + 205, + 555, + 261 + ], + "type": "text", + "content": " as the timestep, which we found to perform best quantitatively (cf. Figure 9). Note how the CleanDIFT segmentation maps are far less noisy than those of the standard diffusion features." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 339, + 271, + 528, + 340 + ], + "blocks": [ + { + "bbox": [ + 339, + 271, + 528, + 340 + ], + "lines": [ + { + "bbox": [ + 339, + 271, + 528, + 340 + ], + "spans": [ + { + "bbox": [ + 339, + 271, + 528, + 340 + ], + "type": "image", + "image_path": "751c2eee31ecec5a62b102e93fb85293f183c5fb6cdb447a0e8be90bdee80fe1.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 348, + 555, + 415 + ], + "lines": [ + { + "bbox": [ + 313, + 348, + 555, + 415 + ], + "spans": [ + { + "bbox": [ + 313, + 348, + 555, + 415 + ], + "type": "text", + "content": "Figure 9. Performance on semantic segmentation using linear probes. Our clean features outperform the noisy diffusion features for the best noising timestep " + }, + { + "bbox": [ + 313, + 348, + 555, + 415 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 348, + 555, + 415 + ], + "type": "text", + "content": ". Semantic segmentation performance of a standard diffusion model heavily depends on the used noising timestep. Unlike for semantic correspondence matching, the optimal " + }, + { + "bbox": [ + 313, + 348, + 555, + 415 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 348, + 555, + 415 + ], + "type": "text", + "content": " value appears to be around " + }, + { + "bbox": [ + 313, + 348, + 555, + 415 + ], + "type": "inline_equation", + "content": "t = 100" + }, + { + "bbox": [ + 313, + 348, + 555, + 415 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 425, + 556, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 425, + 556, + 568 + ], + "spans": [ + { + "bbox": [ + 313, + 425, + 556, + 568 + ], + "type": "text", + "content": "sults are shown in Figure 8. Using our features, we observe significantly less noisy segmentations than with standard diffusion features. We show a quantitative comparison of our CleanDIFT feature's performance against standard diffusion features across timesteps in Figure 9. Notably, the optimal timestep appears to be around " + }, + { + "bbox": [ + 313, + 425, + 556, + 568 + ], + "type": "inline_equation", + "content": "t = 100" + }, + { + "bbox": [ + 313, + 425, + 556, + 568 + ], + "type": "text", + "content": ", in contrast to the optimal timestep for semantic correspondences, which [62] found to be " + }, + { + "bbox": [ + 313, + 425, + 556, + 568 + ], + "type": "inline_equation", + "content": "t = 261" + }, + { + "bbox": [ + 313, + 425, + 556, + 568 + ], + "type": "text", + "content": ". This highlights the need for tuning a timestep individually per downstream task. Our method both alleviates the need for such a timestep tuning and outperforms the standard diffusion features for the optimal timestep." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 576, + 400, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 576, + 400, + 588 + ], + "spans": [ + { + "bbox": [ + 313, + 576, + 400, + 588 + ], + "type": "text", + "content": "4.5. Classification" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "content": "To assess the impact of our method on non-spatial tasks, we evaluate classification performance using pooled features. Pooling mitigates the influence of localized noise, so we anticipate classification performance to remain on par with standard diffusion features unless our setup introduces detrimental effects. We perform k-Nearest Neighbor (kNN) classification with " + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "inline_equation", + "content": "k = 10" + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "content": " on ImageNet1k [11], using SD 1.5 as the diffusion backbone. We sweep across feature maps and timesteps " + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "content": " for the base model, with results presented in Figure 10. Our analysis shows that the fea" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "123" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 70, + 286, + 223 + ], + "blocks": [ + { + "bbox": [ + 62, + 70, + 286, + 223 + ], + "lines": [ + { + "bbox": [ + 62, + 70, + 286, + 223 + ], + "spans": [ + { + "bbox": [ + 62, + 70, + 286, + 223 + ], + "type": "image", + "image_path": "b72a4eccdb33442134d8ae17ce703235bd652130868923e4999f81db0ef6b8f9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 228, + 295, + 285 + ], + "lines": [ + { + "bbox": [ + 55, + 228, + 295, + 285 + ], + "spans": [ + { + "bbox": [ + 55, + 228, + 295, + 285 + ], + "type": "text", + "content": "Figure 10. Classification performance on ImageNet1k [11], using kNN classifier with " + }, + { + "bbox": [ + 55, + 228, + 295, + 285 + ], + "type": "inline_equation", + "content": "k = 10" + }, + { + "bbox": [ + 55, + 228, + 295, + 285 + ], + "type": "text", + "content": " and cosine similarity as the distance metric. We sweep over different timesteps and feature maps. We find that the feature map with the lowest spatial resolution (feature map #0) yields the highest classification accuracy." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 294, + 295, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 294, + 295, + 354 + ], + "spans": [ + { + "bbox": [ + 55, + 294, + 295, + 354 + ], + "type": "text", + "content": "ture map with the lowest spatial resolution, i.e., feature map #0 (see Figure 11), achieves the highest classification accuracy. Furthermore, the optimal timestep " + }, + { + "bbox": [ + 55, + 294, + 295, + 354 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 294, + 295, + 354 + ], + "type": "text", + "content": " for the base model varies between feature maps. For the best-performing feature map, " + }, + { + "bbox": [ + 55, + 294, + 295, + 354 + ], + "type": "inline_equation", + "content": "t = 100" + }, + { + "bbox": [ + 55, + 294, + 295, + 354 + ], + "type": "text", + "content": " yields the highest classification accuracy." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 354, + 296, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 354, + 296, + 403 + ], + "spans": [ + { + "bbox": [ + 55, + 354, + 296, + 403 + ], + "type": "text", + "content": "Importantly, CleanDIFT features slightly outperform the standard diffusion features even when using an optimal timestep " + }, + { + "bbox": [ + 55, + 354, + 296, + 403 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 354, + 296, + 403 + ], + "type": "text", + "content": " for the base model showing that it does not introduce any detrimental effects." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 411, + 156, + 423 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 411, + 156, + 423 + ], + "spans": [ + { + "bbox": [ + 55, + 411, + 156, + 423 + ], + "type": "text", + "content": "4.6. Ablation Studies" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 429, + 296, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 429, + 296, + 477 + ], + "spans": [ + { + "bbox": [ + 55, + 429, + 296, + 477 + ], + "type": "text", + "content": "For simplicity, we perform our ablation studies using DIFT [62] and evaluate the performance for unsupervised zero-shot semantic correspondence matching on a subset of the SPair71k [37] test split." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 487, + 295, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 487, + 295, + 643 + ], + "spans": [ + { + "bbox": [ + 55, + 487, + 295, + 643 + ], + "type": "text", + "content": "Training Objective During training, we maximize the cosine similarity between projected outputs of our feature extraction model and standard diffusion features to align them. To investigate the influence of the employed similarity metric, we compare feature extraction models trained on three different alignment objectives commonly used in similar contexts: mean absolute error " + }, + { + "bbox": [ + 55, + 487, + 295, + 643 + ], + "type": "inline_equation", + "content": "(L_{1})" + }, + { + "bbox": [ + 55, + 487, + 295, + 643 + ], + "type": "text", + "content": ", mean squared error " + }, + { + "bbox": [ + 55, + 487, + 295, + 643 + ], + "type": "inline_equation", + "content": "(L_{2})" + }, + { + "bbox": [ + 55, + 487, + 295, + 643 + ], + "type": "text", + "content": ", and cosine similarity. Quantitative results are provided in Tab. 3. While all objectives result in feature extraction models that outperform standard diffusion features, cosine similarity consistently performs the best across the alignment objectives by a significant margin, confirming our choice of similarity metric." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "content": "Projection Heads We investigate the influence of our proposed projection heads that are used to project Clean-DIFT features onto standard diffusion features. The alignment of feature extraction model and diffusion model is determined by the utilized similarity metric and the pro" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 317, + 70, + 553, + 213 + ], + "blocks": [ + { + "bbox": [ + 317, + 70, + 553, + 213 + ], + "lines": [ + { + "bbox": [ + 317, + 70, + 553, + 213 + ], + "spans": [ + { + "bbox": [ + 317, + 70, + 553, + 213 + ], + "type": "table", + "html": "
ObjectiveProjection HeadsPCK@α(↑)
αimg=0.1αbbox=0.1
Cosine Sim.✓68.3261.43
×68.1661.29
L2✓66.2359.13
×66.4959.43
L1✓66.9160.00
×66.8759.91
SD 2.1-63.4155.92
", + "image_path": "3760c8264721af75dc2b87439e32644434564cdba61ea4c58b038c6a6a5db62e.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 221, + 553, + 255 + ], + "lines": [ + { + "bbox": [ + 313, + 221, + 553, + 255 + ], + "spans": [ + { + "bbox": [ + 313, + 221, + 553, + 255 + ], + "type": "text", + "content": "Table 3. Ablation Study Results. We evaluate the feature extraction models' performance for zero-shot semantic correspondence matching on the SPair71k test split. PCK is aggregated per point." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 264, + 555, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 264, + 555, + 432 + ], + "spans": [ + { + "bbox": [ + 313, + 264, + 555, + 432 + ], + "type": "text", + "content": "jection heads. Therefore, we evaluate our feature extraction model's performance with and without the projection heads in combination with all three similarity metrics. An overview of the comparison is given in Tab. 3. In our main configuration that uses cosine similarity, using the projection heads yields slight performance improvements of 0.24 percentage points for " + }, + { + "bbox": [ + 313, + 264, + 555, + 432 + ], + "type": "inline_equation", + "content": "\\mathrm{PCK}_{\\mathrm{img}}" + }, + { + "bbox": [ + 313, + 264, + 555, + 432 + ], + "type": "text", + "content": " and 0.06 percentage points for " + }, + { + "bbox": [ + 313, + 264, + 555, + 432 + ], + "type": "inline_equation", + "content": "\\mathrm{PCK}_{\\mathrm{bbox}}" + }, + { + "bbox": [ + 313, + 264, + 555, + 432 + ], + "type": "text", + "content": " compared to fine-tuning without the projection heads. As the projection heads are typically not used for inference, they add computational overhead only during the lightweight fine-tuning. Therefore, we argue that it is worthwhile to include them and leverage the small performance gain. Additionally, they can be reused to efficiently obtain feature maps for specific timesteps." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 442, + 388, + 454 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 442, + 388, + 454 + ], + "spans": [ + { + "bbox": [ + 313, + 442, + 388, + 454 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 462, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 462, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 462, + 555, + 713 + ], + "type": "text", + "content": "In this paper, we introduced CleanDIFT, a novel approach for extracting diffusion features. CleanDIFT produces noise-free, timestep-independent, general-purpose diffusion features by consolidating timestep-dependent representations from a pre-trained diffusion backbone into a unified feature representation. We achieve this alignment between our feature extraction model and the pre-trained diffusion backbone through a lightweight fine-tuning procedure that takes approximately 30 minutes on a single A100 GPU. Operating directly on clean images, our method eliminates the information loss associated with adding noise to input images. Furthermore, CleanDIFT removes the requirement for tuning timesteps for each downstream task and avoids the computational overhead of ensembling over noise levels or timesteps. Instead, our method efficiently extracts features with just a single forward pass at inference time, substantially reducing inference costs compared to methods relying on ensembling or inversion. Extensive evaluations of CleanDIFT across diverse downstream tasks demonstrate significant performance improvements over conventional diffusion features." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "124" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 247 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 247 + ], + "type": "text", + "content": "This project has been supported by the German Federal Ministry for Economic Affairs and Climate Action within the project \"NXT GEN AI METHODS - Generative Methoden für Perzeption, Prädiktion und Planung\", the project \"GeniusRobot\" (01IS24083), funded by the Federal Ministry of Education and Research (BMBF), the bidt project KLIMA-MEMES, and the German Research Foundation (DFG) project 421703927. The authors gratefully acknowledge the Gauss Center for Supercomputing for providing compute through the NIC on JUWELS at JSC and the HPC resources supplied by the Erlangen National High Performance Computing Center (NHR@FAU funded by DFG project 440719683) under the NHR project JA-22883." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 248, + 296, + 273 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 248, + 296, + 273 + ], + "spans": [ + { + "bbox": [ + 55, + 248, + 296, + 273 + ], + "type": "text", + "content": "Further, we would like to thank Owen Vincent for continuous technical support." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 285, + 115, + 298 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 285, + 115, + 298 + ], + "spans": [ + { + "bbox": [ + 56, + 285, + 115, + 298 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 304, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 304, + 296, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 304, + 296, + 360 + ], + "spans": [ + { + "bbox": [ + 61, + 304, + 296, + 360 + ], + "type": "text", + "content": "[1] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Ji-aming Song, Qinsheng Zhang, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, et al. ediff-i: Text-to-image diffusion models with an ensemble of expert denoisers. arXiv preprint arXiv:2211.01324, 2022. 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 361, + 296, + 406 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 361, + 296, + 406 + ], + "spans": [ + { + "bbox": [ + 61, + 361, + 296, + 406 + ], + "type": "text", + "content": "[2] Dmitry Baranchuk, Andrey Voynov, Ivan Rubachev, Valentin Khrulkov, and Artem Babenko. Label-efficient semantic segmentation with diffusion models. In International Conference on Learning Representations, 2022. 1, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 61, + 407, + 296, + 451 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 407, + 296, + 451 + ], + "spans": [ + { + "bbox": [ + 61, + 407, + 296, + 451 + ], + "type": "text", + "content": "[3] Shariq Farooq Bhat, Ibrahim Alhashim, and Peter Wonka. Adabins: Depth estimation using adaptive bins. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4009-4018, 2021. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 453, + 296, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 453, + 296, + 485 + ], + "spans": [ + { + "bbox": [ + 62, + 453, + 296, + 485 + ], + "type": "text", + "content": "[4] Giulio Biroli, Tony Bonnaire, Valentin De Bortoli, and Marc Mézard. Dynamical regimes of diffusion models. arXiv preprint arXiv:2402.18491, 2024. 2, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 487, + 296, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 487, + 296, + 531 + ], + "spans": [ + { + "bbox": [ + 62, + 487, + 296, + 531 + ], + "type": "text", + "content": "[5] Minwoo Byeon, Beomhee Park, Haecheon Kim, Sungjun Lee, Woonhyuk Baek, and Saehoon Kim. Coyo-700m: Image-text pair dataset. https://github.com/kakaobrain/coyo-dataset, 2022.5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 533, + 296, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 533, + 296, + 588 + ], + "spans": [ + { + "bbox": [ + 62, + 533, + 296, + 588 + ], + "type": "text", + "content": "[6] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 2, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 590, + 296, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 590, + 296, + 655 + ], + "spans": [ + { + "bbox": [ + 62, + 590, + 296, + 655 + ], + "type": "text", + "content": "[7] Junsong Chen, Jincheng YU, Chongjian GE, Lewei Yao, Enze Xie, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart-" + }, + { + "bbox": [ + 62, + 590, + 296, + 655 + ], + "type": "inline_equation", + "content": "\\mathbb{S}\\backslash \\mathbb{alpha}" + }, + { + "bbox": [ + 62, + 590, + 296, + 655 + ], + "type": "text", + "content": ": Fast training of diffusion transformer for photorealistic text-to-image synthesis. In The Twelfth International Conference on Learning Representations, 2024. 14" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 657, + 296, + 700 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 657, + 296, + 700 + ], + "spans": [ + { + "bbox": [ + 62, + 657, + 296, + 700 + ], + "type": "text", + "content": "[8] Shoufa Chen, Peize Sun, Yibing Song, and Ping Luo. Diffusionnet: Diffusion model for object detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 19830-19843, 2023. 1, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 62, + 703, + 296, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 703, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 62, + 703, + 296, + 714 + ], + "type": "text", + "content": "[9] Xinlei Chen, Zhuang Liu, Saining Xie, and Kaiming He. De" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 714 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 333, + 73, + 555, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 555, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 555, + 95 + ], + "type": "text", + "content": "constructing denoising diffusion models for self-supervised learning. arXiv preprint arXiv:2401.14404, 2024. 2, 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 96, + 555, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 555, + 139 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 555, + 139 + ], + "type": "text", + "content": "[10] Timothee Darcet, Maxime Oquab, Julien Mairal, and Piotr Bojanowski. Vision transformers need registers. In The Twelfth International Conference on Learning Representations, 2024. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 140, + 554, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 140, + 554, + 183 + ], + "spans": [ + { + "bbox": [ + 316, + 140, + 554, + 183 + ], + "type": "text", + "content": "[11] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 248–255, 2009. 7, 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 185, + 554, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 185, + 554, + 217 + ], + "spans": [ + { + "bbox": [ + 316, + 185, + 554, + 217 + ], + "type": "text", + "content": "[12] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 1, 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 217, + 554, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 217, + 554, + 272 + ], + "spans": [ + { + "bbox": [ + 316, + 217, + 554, + 272 + ], + "type": "text", + "content": "[13] Patrick Esser, Johnathan Chiu, Parmida Atighehchian, Jonathan Granskog, and Anastasis Germanidis. Structure and content-guided video synthesis with diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7346-7356, 2023. 1, 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 273, + 554, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 273, + 554, + 337 + ], + "spans": [ + { + "bbox": [ + 317, + 273, + 554, + 337 + ], + "type": "text", + "content": "[14] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In *Forty-first International Conference on Machine Learning*, 2024. 2, 14" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 338, + 554, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 338, + 554, + 380 + ], + "spans": [ + { + "bbox": [ + 317, + 338, + 554, + 380 + ], + "type": "text", + "content": "[15] Zach Evans, CJ Carr, Josiah Taylor, Scott H Hawley, and Jordi Pons. Fast timing-conditioned latent audio diffusion. In *Forty-first International Conference on Machine Learning*, 2024. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 383, + 554, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 383, + 554, + 425 + ], + "spans": [ + { + "bbox": [ + 317, + 383, + 554, + 425 + ], + "type": "text", + "content": "[16] M. Everingham, L. Van Gool, C. K. I. Williams, J. Winn, and A. Zisserman. The Pascal visual object classes (voc) challenge. International Journal of Computer Vision, 88(2): 303-338, 2010. 7, 13" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 426, + 554, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 426, + 554, + 469 + ], + "spans": [ + { + "bbox": [ + 317, + 426, + 554, + 469 + ], + "type": "text", + "content": "[17] Michael Fuest, Pingchuan Ma, Ming Gui, Johannes S. Fischer, Vincent Tao Hu, and Björn Ommer. Diffusion models and representation learning: A survey. CoRR, abs/2407.00783, 2024. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 471, + 554, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 471, + 554, + 503 + ], + "spans": [ + { + "bbox": [ + 317, + 471, + 554, + 503 + ], + "type": "text", + "content": "[18] Frank Fundel, Johannes Schusterbauer, Vincent Tao Hu, and Björn Ommer. Distillation of diffusion features for semantic correspondence. WACV, 2025. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 505, + 554, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 505, + 554, + 547 + ], + "spans": [ + { + "bbox": [ + 317, + 505, + 554, + 547 + ], + "type": "text", + "content": "[19] Spyros Gidaris, Praveer Singh, and Nikos Komodakis. Unsupervised representation learning by predicting image rotations. In International Conference on Learning Representations, 2018. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 548, + 554, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 548, + 554, + 602 + ], + "spans": [ + { + "bbox": [ + 317, + 548, + 554, + 602 + ], + "type": "text", + "content": "[20] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dólár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 2, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 604, + 554, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 604, + 554, + 657 + ], + "spans": [ + { + "bbox": [ + 316, + 604, + 554, + 657 + ], + "type": "text", + "content": "[21] Eric Hedlin, Gopal Sharma, Shweta Mahajan, Hossam Isack, Abhishek Kar, Andrea Tagliasacchi, and Kwang Moo Yi. Unsupervised semantic correspondence using stable diffusion. Advances in Neural Information Processing Systems, 36, 2023. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 658, + 554, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 554, + 691 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 554, + 691 + ], + "type": "text", + "content": "[22] Geoffrey Hinton, Oriol Vinyls, and Jeffrey Dean. Distilling the knowledge in a neural network. In NIPS Deep Learning and Representation Learning Workshop, 2015. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 692, + 554, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 554, + 714 + ], + "type": "text", + "content": "[23] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 3" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "125" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "type": "text", + "content": "[24] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 294, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 294, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 294, + 150 + ], + "type": "text", + "content": "[25] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. Advances in Neural Information Processing Systems, 35:8633-8646, 2022. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 294, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 294, + 195 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 294, + 195 + ], + "type": "text", + "content": "[26] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. *ICLR*, 1(2):3, 2022. 14" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 198, + 294, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 198, + 294, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 198, + 294, + 262 + ], + "type": "text", + "content": "[27] Drew A Hudson, Daniel Zoran, Mateusz Malinowski, Andrew K Lampinen, Andrew Jaegle, James L McClelland, Loic Matthey, Felix Hill, and Alexander Lerchner. Soda: Bottleneck diffusion models for representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23115-23127, 2024. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 264, + 294, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 264, + 294, + 319 + ], + "spans": [ + { + "bbox": [ + 56, + 264, + 294, + 319 + ], + "type": "text", + "content": "[28] Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. Openclip, 2021. 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 320, + 294, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 320, + 294, + 363 + ], + "spans": [ + { + "bbox": [ + 56, + 320, + 294, + 363 + ], + "type": "text", + "content": "[29] Diederik P. Kingma and Jimmy Lei Ba. Adam: A method for stochastic optimization. In Proceedings of the 3rd International Conference on Learning Representations (ICLR), 2015. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 365, + 294, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 365, + 294, + 409 + ], + "spans": [ + { + "bbox": [ + 56, + 365, + 294, + 409 + ], + "type": "text", + "content": "[30] Zhifeng Kong, Wei Ping, Jiaji Huang, Kexin Zhao, and Bryan Catanzaro. Diffwave: A versatile diffusion model for audio synthesis. In International Conference on Learning Representations, 2021. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 411, + 294, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 411, + 294, + 464 + ], + "spans": [ + { + "bbox": [ + 56, + 411, + 294, + 464 + ], + "type": "text", + "content": "[31] Alexander C Li, Mihir Prabhudesai, Shivam Duggal, Ellis Brown, and Deepak Pathak. Your diffusion model is secretly a zero-shot classifier. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2206-2217, 2023. 1, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 467, + 294, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 467, + 294, + 532 + ], + "spans": [ + { + "bbox": [ + 56, + 467, + 294, + 532 + ], + "type": "text", + "content": "[32] Daiqing Li, Huan Ling, Amlan Kar, David Acuna, Seung Wook Kim, Karsten Kreis, Antonio Torralba, and Sanja Fidler. Dreamteacher: Pretraining image backbones with deep generative models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16698-16708, 2023. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 534, + 294, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 534, + 294, + 588 + ], + "spans": [ + { + "bbox": [ + 56, + 534, + 294, + 588 + ], + "type": "text", + "content": "[33] Xinghui Li, Jingyi Lu, Kai Han, and Victor Adrian Prisacariu. Sd4match: Learning to prompt stable diffusion model for semantic matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 27558-27568, 2024. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 590, + 294, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 590, + 294, + 633 + ], + "spans": [ + { + "bbox": [ + 56, + 590, + 294, + 633 + ], + "type": "text", + "content": "[34] Zhenyu Li, Xuyang Wang, Xianming Liu, and Junjun Jiang. Binsformer: Revisiting adaptive bins for monocular depth estimation. IEEE Transactions on Image Processing, 2024. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 635, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 294, + 689 + ], + "type": "text", + "content": "[35] Grace Luo, Lisa Dunlap, Dong Huk Park, Aleksander Holynski, and Trevor Darrell. Diffusion hyperfeatures: Searching through time and space for semantic correspondence. In Advances in Neural Information Processing Systems, pages 47500-47510. Curran Associates, Inc., 2023. 3, 5, 6, 13" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "text", + "content": "[36] Chenlin Meng, Robin Rombach, Ruiqi Gao, Diederik Kingma, Stefano Ermon, Jonathan Ho, and Tim Salimans." + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 333, + 72, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 72, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 333, + 72, + 553, + 106 + ], + "type": "text", + "content": "On distillation of guided diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14297-14306, 2023. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 107, + 553, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 140 + ], + "type": "text", + "content": "[37] Juhong Min, Jongmin Lee, Jean Ponce, and Minsu Cho. Spair-71k: A large-scale benchmark for semantic correspondence. arXiv preprint arXiv:1908.10543, 2019. 6, 8, 12, 13" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 141, + 553, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 553, + 185 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 553, + 185 + ], + "type": "text", + "content": "[38] Ishan Misra and Laurens van der Maaten. Self-supervised learning of pretext-invariant representations. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6707-6717, 2020. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 186, + 553, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 186, + 553, + 230 + ], + "spans": [ + { + "bbox": [ + 316, + 186, + 553, + 230 + ], + "type": "text", + "content": "[39] Koichi Namekata, Amirmojtaba Sabour, Sanja Fidler, and Seung Wook Kim. Emerdiff: Emerging pixel-level semantic knowledge in diffusion models. In The Twelfth International Conference on Learning Representations, 2023. 1, 3, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 232, + 553, + 264 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 232, + 553, + 264 + ], + "spans": [ + { + "bbox": [ + 316, + 232, + 553, + 264 + ], + "type": "text", + "content": "[40] Pushmeet Kohli Nathan Silberman, Derek Hoiem and Rob Fergus. Indoor segmentation and support inference from rgbd images. In ECCV, 2012. 6, 7, 14" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 266, + 553, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 266, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 316, + 266, + 553, + 319 + ], + "type": "text", + "content": "[41] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 1, 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 322, + 553, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 322, + 553, + 365 + ], + "spans": [ + { + "bbox": [ + 316, + 322, + 553, + 365 + ], + "type": "text", + "content": "[42] Mehdi Noroozi and Paolo Favaro. Unsupervised learning of visual representations by solving jigsaw puzzles. In European conference on computer vision, pages 69-84. Springer, 2016. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 366, + 553, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 366, + 553, + 476 + ], + "spans": [ + { + "bbox": [ + 316, + 366, + 553, + 476 + ], + "type": "text", + "content": "[43] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel HAZIZA, Francisco Massa, Alaaeldin El-Nouby, Mido Assran, Nicolas Ballas, Wojciech Galuba, Russell Howes, Po-Yao Huang, Shang-Wen Li, Ishan Misra, Michael Rabbat, Vasu Sharma, Gabriel Synnaeve, Hu Xu, Herve Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. DINOv2: Learning robust visual features without supervision. Transactions on Machine Learning Research, 2024. 2, 3, 6, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 478, + 553, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 478, + 553, + 520 + ], + "spans": [ + { + "bbox": [ + 316, + 478, + 553, + 520 + ], + "type": "text", + "content": "[44] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4195-4205, 2023. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 522, + 553, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 522, + 553, + 567 + ], + "spans": [ + { + "bbox": [ + 316, + 522, + 553, + 567 + ], + "type": "text", + "content": "[45] Ethan Perez, Florian Strub, Harm De Vries, Vincent Dumoulin, and Aaron Courville. Film: Visual reasoning with a general conditioning layer. In Proceedings of the AAAI conference on artificial intelligence, 2018. 14" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 568, + 553, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 568, + 553, + 632 + ], + "spans": [ + { + "bbox": [ + 316, + 568, + 553, + 632 + ], + "type": "text", + "content": "[46] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. In The Twelfth International Conference on Learning Representations, 2024. 2, 14" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 635, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 553, + 689 + ], + "type": "text", + "content": "[47] Adam Polyak, Amit Zohar, Andrew Brown, Andros Tjandra, Animesh Sinha, Ann Lee, Apoorv Vyas, Bowen Shi, Chih-Yao Ma, Ching-Yao Chuang, et al. Movie gen: A cast of media foundation models. arXiv preprint arXiv:2410.13720, 2024. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "text", + "content": "[48] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry," + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "text", + "content": "126" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 714 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 76, + 72, + 295, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 295, + 116 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 295, + 116 + ], + "type": "text", + "content": "Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 295, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 295, + 149 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 295, + 149 + ], + "type": "text", + "content": "[49] Prajit Ramachandran, Barret Zoph, and Quoc V Le. Searching for activation functions. arXiv preprint arXiv:1710.05941, 2017. 14" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 151, + 294, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 151, + 294, + 183 + ], + "spans": [ + { + "bbox": [ + 56, + 151, + 294, + 183 + ], + "type": "text", + "content": "[50] Severi Rissanen, Markus Heinonen, and Arno Solin. Generative modelling with inverse heat dissipation. arXiv preprint arXiv:2206.13397, 2022. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 184, + 295, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 184, + 295, + 239 + ], + "spans": [ + { + "bbox": [ + 56, + 184, + 295, + 239 + ], + "type": "text", + "content": "[51] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 2, 5, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 239, + 295, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 239, + 295, + 294 + ], + "spans": [ + { + "bbox": [ + 56, + 239, + 295, + 294 + ], + "type": "text", + "content": "[52] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention - MICCAI 2015, pages 234-241, Cham, 2015. Springer International Publishing. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 294, + 295, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 294, + 295, + 360 + ], + "spans": [ + { + "bbox": [ + 56, + 294, + 295, + 360 + ], + "type": "text", + "content": "[53] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural information processing systems, 35:36479-36494, 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 360, + 295, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 360, + 295, + 392 + ], + "spans": [ + { + "bbox": [ + 56, + 360, + 295, + 392 + ], + "type": "text", + "content": "[54] Tim Salimans and Jonathan Ho. Progressive distillation for fast sampling of diffusion models. arXiv preprint arXiv:2202.00512, 2022. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 393, + 295, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 393, + 295, + 437 + ], + "spans": [ + { + "bbox": [ + 56, + 393, + 295, + 437 + ], + "type": "text", + "content": "[55] Axel Sauer, Dominik Lorenz, Andreas Blattmann, and Robin Rombach. Adversarial diffusion distillation. In European Conference on Computer Vision, pages 87-103. Springer, 2024. 14" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 437, + 295, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 437, + 295, + 503 + ], + "spans": [ + { + "bbox": [ + 56, + 437, + 295, + 503 + ], + "type": "text", + "content": "[56] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278-25294, 2022. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 503, + 295, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 503, + 295, + 525 + ], + "spans": [ + { + "bbox": [ + 56, + 503, + 295, + 525 + ], + "type": "text", + "content": "[57] Noam Shazeer. Glu variants improve transformer. arXiv preprint arXiv:2002.05202, 2020. 14" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 525, + 295, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 525, + 295, + 581 + ], + "spans": [ + { + "bbox": [ + 56, + 525, + 295, + 581 + ], + "type": "text", + "content": "[58] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. In The Eleventh International Conference on Learning Representations, 2023. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 582, + 295, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 582, + 295, + 613 + ], + "spans": [ + { + "bbox": [ + 56, + 582, + 295, + 613 + ], + "type": "text", + "content": "[59] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In International Conference on Learning Representations, 2021. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 614, + 295, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 614, + 295, + 647 + ], + "spans": [ + { + "bbox": [ + 56, + 614, + 295, + 647 + ], + "type": "text", + "content": "[60] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. Advances in neural information processing systems, 32, 2019. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 647, + 295, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 295, + 691 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 295, + 691 + ], + "type": "text", + "content": "[61] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 691, + 295, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 295, + 714 + ], + "type": "text", + "content": "[62] Luming Tang, Menglin Jia, Qianqian Wang, Cheng Perng Phoo, and Bharath Hariharan. Emergent correspondence" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 463 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "text", + "content": "from image diffusion. In Advances in Neural Information Processing Systems, pages 1363-1389. Curran Associates, Inc., 2023. 1, 2, 3, 4, 5, 6, 7, 8, 12, 13, 14, 15" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 107, + 553, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 173 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 173 + ], + "type": "text", + "content": "[63] Weijia Wu, Yuzhong Zhao, Mike Zheng Shou, Hong Zhou, and Chunhua Shen. Diffumask: Synthesizing images with pixel-level annotations for semantic segmentation using diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1206-1217, 2023. 1, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 174, + 553, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 174, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 316, + 174, + 553, + 239 + ], + "type": "text", + "content": "[64] Jiarui Xu, Sifei Liu, Arash Vahdat, Wonmin Byeon, Xiaolong Wang, and Shalini De Mello. Open-vocabulary panoptic segmentation with text-to-image diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2955-2966, 2023. 1, 3, 13, 15" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 240, + 553, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 240, + 553, + 274 + ], + "spans": [ + { + "bbox": [ + 316, + 240, + 553, + 274 + ], + "type": "text", + "content": "[65] Biao Zhang and Rico Sennrich. Root mean square layer normalization. Advances in Neural Information Processing Systems, 32, 2019. 14" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 274, + 553, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 274, + 553, + 350 + ], + "spans": [ + { + "bbox": [ + 316, + 274, + 553, + 350 + ], + "type": "text", + "content": "[66] Junyi Zhang, Charles Herrmann, Junhwa Hur, Luisa Polania Cabrera, Varun Jampani, Deqing Sun, and Ming-Hsuan Yang. A tale of two features: Stable diffusion complements dino for zero-shot semantic correspondence. In Advances in Neural Information Processing Systems, pages 45533-45547. Curran Associates, Inc., 2023. 1, 2, 3, 5, 6, 12, 13, 15" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 352, + 553, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 352, + 553, + 418 + ], + "spans": [ + { + "bbox": [ + 316, + 352, + 553, + 418 + ], + "type": "text", + "content": "[67] Junyi Zhang, Charles Herrmann, Junhwa Hur, Eric Chen, Varun Jampani, Deqing Sun, and Ming-Hsuan Yang. Telling left from right: Identifying geometry-aware semantic correspondence. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3076-3085, 2024. 1, 2, 3, 5, 6, 12, 13, 15" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 418, + 553, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 418, + 553, + 463 + ], + "spans": [ + { + "bbox": [ + 316, + 418, + 553, + 463 + ], + "type": "text", + "content": "[68] Jinghao Zhou, Chen Wei, Huiyu Wang, Wei Shen, Cihang Xie, Alan Yuille, and Tao Kong. Image BERT pre-training with online tokenizer. In International Conference on Learning Representations, 2022. 7" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "127" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/ClearSight_ Visual Signal Enhancement for Object Hallucination Mitigation in Multimodal Large Language Models/cc596546-7afd-468d-96ec-564a706a95d3_content_list.json b/2025/ClearSight_ Visual Signal Enhancement for Object Hallucination Mitigation in Multimodal Large Language Models/cc596546-7afd-468d-96ec-564a706a95d3_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..52181fedf0157480d4d36e1476548ff21257442a --- /dev/null +++ b/2025/ClearSight_ Visual Signal Enhancement for Object Hallucination Mitigation in Multimodal Large Language Models/cc596546-7afd-468d-96ec-564a706a95d3_content_list.json @@ -0,0 +1,1844 @@ +[ + { + "type": "text", + "text": "ClearSight: Visual Signal Enhancement for Object Hallucination Mitigation in Multimodal Large Language Models", + "text_level": 1, + "bbox": [ + 102, + 128, + 893, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hao Yin Gunagzong Si Zilei Wang\\*", + "bbox": [ + 310, + 203, + 687, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "University of Science and Technology of China", + "bbox": [ + 310, + 222, + 684, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{yinhnavi, guangzongsi}@mail.ustc.edu.cn, zlwang@ustc.edu.cn", + "bbox": [ + 230, + 241, + 759, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 291, + 326, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Contrastive decoding strategies are widely used to mitigate object hallucinations in multimodal large language models (MLLMs). By reducing over-reliance on language priors, these strategies ensure that generated content remains closely grounded in visual inputs, producing contextually accurate outputs. Since contrastive decoding requires no additional training or external tools, it offers both computational efficiency and versatility, making it highly attractive. However, these methods present two main limitations: (1) bluntly suppressing language priors can compromise coherence and accuracy of generated content, and (2) processing contrastive inputs adds computational load, significantly slowing inference speed. To address these challenges, we propose Visual Amplification Fusion (VAF), a plug-and-play technique that enhances attention to visual signals within the model's middle layers, where modality fusion predominantly occurs. This approach enables more effective capture of visual features, reducing the model's bias toward language modality. Experimental results demonstrate that VAF significantly reduces hallucinations across various MLLMs without affecting inference speed, while maintaining coherence and accuracy in generated outputs. The code is available at https://github.com/ustc-hyin/ClearSight.", + "bbox": [ + 89, + 323, + 485, + 686 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 715, + 220, + 729 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In recent years, MLLMs [8, 27, 33, 35, 54, 55] have achieved remarkable progress in the intersecting fields of computer vision and natural language processing, and have been widely applied in tasks such as image captioning and visual question answering. However, these models often encounter the issue of \"object hallucination\" [15, 29, 32, 37] in practical applications, where the generated textual descriptions do not match the actual objects in the image. This problem highlights an over-reliance on unimodal pri", + "bbox": [ + 89, + 739, + 483, + 878 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ors (especially language priors) [17, 48, 50, 53] during inference, posing potential risks in high-precision applications such as medical diagnosis [18, 46] and autonomous driving [9, 34, 39, 49].", + "bbox": [ + 511, + 292, + 903, + 353 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To address object hallucination [1, 2, 7, 14], several Contrastive Decoding strategies have been introduced in recent years. Among these, the Visual Contrastive Decoding (VCD) method has shown promise in reducing hallucinations by contrasting output distributions from both original and perturbed visual inputs, thus mitigating the model's excessive reliance on language priors [16, 41]. Notably, contrastive decoding methods do not require additional training or external tools, offering both computational efficiency and versatility, which has garnered them significant attention. However, these methods present two main limitations:", + "bbox": [ + 511, + 354, + 906, + 521 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Limitations of Contrastive Decoding", + "text_level": 1, + "bbox": [ + 539, + 534, + 782, + 550 + ], + "page_idx": 0 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- While reducing over-reliance on language priors, these methods may compromise the coherence and accuracy of generated content.", + "- Contrastive decoding necessitates separate processing of the original and contrastive inputs, which considerably increases inference time." + ], + "bbox": [ + 539, + 561, + 880, + 652 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To address these shortcomings, we hope to propose a training-free method that can effectively reduces hallucinations without compromising content quality or inference speed. Our saliency analysis of the model's attention maps reveals that biases toward language in generated content do not arise from an overemphasis on language signals but rather from insufficient attention on visual information during modality fusion. Based on this insight, we introduce a novel, plug-and-play technique to mitigate hallucinations: Visual Amplification Fusion (VAF).", + "bbox": [ + 511, + 672, + 906, + 823 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Our analysis indicates that modality fusion in MLLMs primarily occurs within the middle layers. VAF specifically amplifies visual signals at these middle layers, enabling the model to capture more distinctive visual features during fusion, which in turn reduces false descriptions in generated", + "bbox": [ + 511, + 824, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding Author", + "bbox": [ + 109, + 887, + 238, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "14625", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "text. This technique not only strengthens the model's visual representations but also retains the beneficial influence of language priors, thus preserving content quality. Furthermore, by eliminating the need to process contrastive samples, VAF maintains inference speed.", + "bbox": [ + 89, + 90, + 482, + 167 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Experimental results validate the effectiveness of the VAF method. Across multiple object hallucination benchmarks, VAF demonstrated notable performance gains, with improvements of approximately $3\\%$ on POPE and $7\\%$ on MME. In terms of coherence and accuracy of generated responses, VCD caused a roughly $19\\%$ decrease on NoCaps, while VAF maintained content quality without negative impacts. Additionally, VCD reduced inference speed by $50\\%$ , whereas VAF had virtually no effect on inference speed.", + "bbox": [ + 89, + 167, + 482, + 303 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, the main contributions are as follows:", + "bbox": [ + 109, + 304, + 444, + 318 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We identify the negative impacts of contrastive decoding methods on both the quality of generated content and model inference speed.", + "- We analyze the modality fusion mechanism in MLLMs, highlighting its insufficient attention to visual information.", + "- We introduce the VAF method, which effectively mitigates the object hallucination problem while maintaining inference speed, coherence, and accuracy.", + "- We demonstrate the significant performance improvements of the VAF method across multiple object hallucination benchmarks." + ], + "bbox": [ + 89, + 319, + 482, + 500 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 89, + 516, + 227, + 532 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Multimodal Large Language Models", + "text_level": 1, + "bbox": [ + 89, + 541, + 408, + 558 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The development of MLLMs [26, 36, 51, 52] has advanced from BERT-based decoders to LLM-based architectures [4, 11, 40, 43-45], enabling improved multimodal relationship capture [6, 10, 24, 25]. Models like BLIP-2 [27] and miniGPT-4 [55] incorporate a Q-Former mechanism, which enhances the alignment between visual and textual inputs, allowing for more precise interactions across modalities. InstructBLIP [12] builds on this approach by adding task-specific instructions, which improve the model's understanding of context-sensitive visual semantics. LLaVA [33] and Qwen-VL [5] utilize simpler linear projection techniques that streamline the alignment process, resulting in improved overall performance on vision-language tasks. However, hallucination issues persist across MLLMs, posing a significant challenge that requires further research.", + "bbox": [ + 89, + 564, + 482, + 791 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Contrastive Decoding Strategies", + "text_level": 1, + "bbox": [ + 89, + 801, + 375, + 819 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In recent years, Contrastive Decoding [19, 21, 22, 28] has emerged as a technique to improve generative model accuracy through contrastive judgment, widely employed to address hallucinations in generated content. For instance, Visual Contrastive Decoding (VCD) [23] contrasts output", + "bbox": [ + 89, + 824, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "distributions derived from original and distorted visual inputs, effectively reducing the over-reliance on statistical bias and unimodal priors, two essential causes of object hallucinations. Similarly, Instruction Contrastive Decoding (ICD) [47] works by comparing distributions derived from standard and disrupted instructions, thereby removing hallucinated concepts from the original distribution. These contrastive methods help ground generated content closely to visual inputs, resulting in contextually accurate outputs. However, despite these advancements, contrastive decoding faces two primary limitations: slower inference speed and reduced coherence in generated content. To overcome these limitations, we propose the VAF method, which achieves effective hallucination reduction while preserving both inference speed and content coherence.", + "bbox": [ + 511, + 90, + 906, + 318 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Preliminary and Motivation", + "text_level": 1, + "bbox": [ + 511, + 329, + 772, + 345 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In Sec. 3.1, we illustrate the working mechanism of contrastive decoding to mitigate hallucinations, using Visual Contrastive Decoding as an example. In Sec. 3.2, we analysis two main drawbacks of this approach: its potential to disrupt the coherence and accuracy of generated content, and its tendency to slow down model inference.", + "bbox": [ + 511, + 354, + 906, + 446 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3.1. Contrastive Decoding", + "text_level": 1, + "bbox": [ + 511, + 453, + 718, + 469 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We consider a MLLM parametrized by $\\theta$ . The model takes as input a textual query $x$ and a visual input $v$ , where $v$ provides contextual visual information to assist the model in generating a relevant response $y$ to the textual query. The response $y$ is sampled auto-regressively from the probability distribution conditioned on the query $x$ and the visual context $v$ . Mathematically, this can be formulated as:", + "bbox": [ + 511, + 474, + 905, + 580 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} y _ {t} \\sim p _ {\\theta} \\left(y _ {t} \\mid v, x, y _ {< t}\\right) \\tag {1} \\\\ \\propto \\exp \\operatorname {l o g i t} _ {\\theta} \\left(y _ {t} \\mid v, x, y _ {< t}\\right) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 604, + 589, + 903, + 625 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "where $y_{t}$ denotes the token at time step $t$ , and $y_{< t}$ represents the sequence of generated tokens up to the time step $(t - 1)$ .", + "bbox": [ + 511, + 632, + 905, + 662 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To mitigate the issue of object hallucination in MLLMs, contrastive decoding techniques can be applied. Here, we present Visual Contrastive Decoding (VCD) as a representative approach, shown in Fig. 1. Specifically, given a textual query $x$ and a visual input $v$ , the model generates two distinct output distributions: one conditioned on the original $v$ and the other on the distorted visual input $v'$ , which is derived by applying pre-defined distortions (i.e., Gaussian noise mask) to $v$ . Then, a new contrastive probability distribution is computed by exploiting the differences between the two initially obtained distributions. The new contrastive distribution $p_{vcd}$ is formulated as:", + "bbox": [ + 511, + 662, + 906, + 844 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} p _ {v c d} (y \\mid v, v ^ {\\prime}, x) = \\operatorname {s o f t m a x} \\left[ \\log \\mathrm {i t} _ {\\theta} (y \\mid v, x) + \\right. \\tag {2} \\\\ \\left. \\alpha \\cdot \\left(\\operatorname {l o g i t} _ {\\theta} (y \\mid v, x) - \\operatorname {l o g i t} _ {\\theta} (y \\mid v ^ {\\prime}, x)\\right) \\right], \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 535, + 849, + 903, + 904 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "14626", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "where larger $\\alpha$ values indicate a stronger amplification of differences between the two distributions ( $\\alpha = 0$ reduces to regular decoding). Essentially, VCD serves as a corrective mechanism, reducing hallucinations by contrasting against a distribution predisposed to favoring them.", + "bbox": [ + 89, + 90, + 482, + 167 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/3530f7ec9c14eee44cc85b156e9cc3c505521d148588eb94b12249782b6dc709.jpg", + "image_caption": [ + "Figure 1. Illustration of Visual Contrastive Decoding. The hallucinated object \"Teacher\" is suppressed by contrasting with an output distribution prone to hallucinations. This method has two main drawbacks: (1) additional processing of distorted visual inputs greatly increases inference time; (2) subtracting the language prior disrupts content coherence." + ], + "image_footnote": [], + "bbox": [ + 96, + 179, + 478, + 323 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Limitations of Contrastive Decoding", + "text_level": 1, + "bbox": [ + 89, + 443, + 405, + 459 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As contrastive decoding methods do not require training or external tools, they offer high computational efficiency and generalizability, attracting significant attention in academia. However, these methods still have two major drawbacks: a reduction in the quality of generated content and slower inference speed.", + "bbox": [ + 89, + 465, + 483, + 556 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/cc50a62940a91f14346d7d0da9631496964595dc457f1bbe791655986ea4a3bd.jpg", + "image_caption": [ + "Figure 2. Impact of VCD on Model Performance. CIDEr scores are reported on the Nocaps benchmark, while Accuracy is presented for the ScienceQA benchmark. The use of VCD leads to a significant decline in model performance." + ], + "image_footnote": [], + "bbox": [ + 120, + 571, + 444, + 741 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "While contrasting logits of $p_{\\theta}(y \\mid v, x)$ and $p_{\\theta}(y \\mid v', x)$ can help reduce over-reliance on language priors and mitigate hallucination in MLLMs-as evidenced by a $4\\%$ performance gain on the POPE benchmark using the VCD method-merely decreasing the influence of the language", + "bbox": [ + 89, + 824, + 483, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "modality on the output distribution may undermine the coherence of the generated content, potentially leading to prediction errors. This issue is less pronounced in straightforward object hallucination tasks, where responses are limited to binary options, such as \"yes\" or \"no\". However, in more complex tasks, including multiple-choice question answering and image caption, the impact of contrastive learning methods on content quality becomes more significant.", + "bbox": [ + 511, + 90, + 903, + 210 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To verify this, we applied VCD method to LLaVA-v1.5-7B and LLaVA-v1.5-13B models, assessing their performance on the ScienceQA [38] and NoCaps benchmarks. As illustrated in Fig. 2, our findings reveal that, following the application of VCD, model performance decreased by $5\\%$ on ScienceQA and by a considerable $45\\%$ on NoCaps. These results suggest that in tasks requiring nuanced natural language generation, contrastive decoding methods can substantially impair content quality.", + "bbox": [ + 511, + 210, + 906, + 348 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/caeeaa2dc2b3706b4a3a110348583e3a34db6fcf764cee126d37bb969e0a74e7.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelMethodScienceQANocaps
LLaVA-v1.5-7BRegular0.141s0.456s
VCD0.293s1.086s
LLaVA-v1.5-13BRegular0.222s0.602s
VCD0.459s1.372s
", + "bbox": [ + 529, + 358, + 890, + 465 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Table 1. Impact of VCD on Model Inference Speed. The table shows the average inference time per sample (in seconds) on the ScienceQA and Nocaps benchmarks. Applying the VCD method nearly doubled the inference time of the model.", + "bbox": [ + 511, + 474, + 906, + 531 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Contrastive decoding methods notably reduce inference speed because they require calculating the output distribution for additional contrastive samples. For instance, in VCD method, each visual input $v$ necessitates computing the logits of both $p_{\\theta}(y \\mid v, x)$ and $p_{\\theta}(y \\mid v', x)$ separately. This doubles the computation load during inference compared to vanilla decoding. We evaluated the inference speed of VCD versus vanilla decoding on ScienceQA. The experimental results, shown in Tab. 1, reveal that VCD's inference time is almost double that of vanilla decoding.", + "bbox": [ + 511, + 546, + 905, + 696 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Visual Neglect in Modal Fusion", + "text_level": 1, + "bbox": [ + 511, + 709, + 802, + 726 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The primary objective of this section is to examine why MLLMs tend to rely excessively on language priors in their predictions. In Sec. 4.1, saliency analysis reveals that image tokens influence prediction outcomes mainly through interactions with instruction tokens within the middle layers. Sec. 4.2 then compares attention weights across different modalities, showing that the attention given to visual features is notably lower than that allocated to system prompts and user instructions. These findings indicate that visual information is often underutilized in the modality fusion process, resulting in an over-reliance on language priors.", + "bbox": [ + 511, + 734, + 906, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "14627", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1. Mid-layer: Visual-Language Fusion", + "text_level": 1, + "bbox": [ + 89, + 90, + 403, + 107 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To uncover why MLLMs tend to overly rely on language priors and overlook visual content in prediction, it is necessary first to clarify how the model utilizes visual modality information. This section explores the influence of the visual modality on prediction outcomes from the perspective of visual information interaction.", + "bbox": [ + 89, + 113, + 482, + 203 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We employ the saliency technique, a widely used interpretability tool, to highlight key token interactions within the attention mechanism. Following established practices, we utilize Taylor expansion to compute saliency scores for each element of the attention matrix:", + "bbox": [ + 89, + 205, + 483, + 280 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nI _ {l} = \\left| \\sum_ {h} A _ {h, l} \\odot \\frac {\\partial \\mathcal {L} (x)}{\\partial A _ {h , l}} \\right|. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 196, + 291, + 482, + 333 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here, $A_{h,l}$ represents the attention matrix value for the $h$ -th attention head in the $l$ -th layer, $x$ denotes the input, and $\\mathcal{L}(x)$ is the loss function of the task, e.g., the cross-entropy objective for question-answering tasks. The saliency matrix $I_l$ for the $l$ -th layer is obtained by averaging across all attention heads. The significance of information flow from the $j$ -th token to the $i$ -th token in MLLMs is represented by $I_l(i,j)$ .", + "bbox": [ + 89, + 345, + 483, + 465 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To draw a clearer picture of visual information flow in MLLMs, we introduce two quantitative metrics based on $I_{l}(i,j)$ , with a particular focus on the information interaction involving image tokens. The definitions of the two quantitative metrics follow below.", + "bbox": [ + 89, + 467, + 483, + 542 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$S_{vv}$ , measuring the importance of information flow among image tokens:", + "bbox": [ + 89, + 542, + 483, + 573 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nS _ {v v} = \\frac {\\sum_ {(i , j) \\in C _ {v v}} I _ {l} (i , j)}{\\left| C _ {v v} \\right|} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 179, + 585, + 480, + 621 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nC _ {v v} = \\{(i, j): i, j \\in \\mathcal {V}, i \\geq j \\}.\n$$\n", + "text_format": "latex", + "bbox": [ + 179, + 625, + 393, + 641 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$S_{vt}$ , measuring the importance of information flow from image tokens to instruction tokens:", + "bbox": [ + 89, + 652, + 483, + 683 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nS _ {v t} = \\frac {\\sum_ {(i , j) \\in C _ {v t}} I _ {l} (i , j)}{\\left| C _ {v t} \\right|} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 184, + 696, + 482, + 732 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nC _ {v t} = \\{(i, j): i \\in \\mathcal {T}, j \\in \\mathcal {V} \\}.\n$$\n", + "text_format": "latex", + "bbox": [ + 184, + 734, + 388, + 751 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here, $\\mathcal{V}$ represents the index set of image tokens, derived from features learned by pre-trained visual encoders, while $\\mathcal{T}$ denotes the index set of instruction tokens, specifying requests or questions related to the images. $S_{vv}$ and $S_{vt}$ are utilized to analyze the mechanisms of visual information processing in MLLMs. We define attention interactions among image tokens as intra-visual information flow and those between image and instruction tokens as visual-textual information flow.", + "bbox": [ + 89, + 763, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We conducted experiments with the LLaVA-v1.5-7B model on the MS COCO dataset under the POPE benchmark, sampling 500 examples for evaluation. Fig. 3 underscores the critical role of the visual-textual information flow within the model's middle layers, specifically from the 8-th to the 15-th layer. This observation indicates that in these layers, visual information interacts intensively with textual information via attention mechanisms, which substantially influences the prediction outcomes.", + "bbox": [ + 511, + 90, + 906, + 227 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/d6388fbbfadbe56b6f02102aad69fc5b9143c2371f156258edace2c19395f97c.jpg", + "image_caption": [ + "Figure 3. The importance of intra-visual flow and visual-textual flow across various layers. The visual-textual information flow in the middle layers has a significant impact on prediction outcomes." + ], + "image_footnote": [], + "bbox": [ + 537, + 244, + 879, + 411 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Attention Imbalance Across Modalities", + "text_level": 1, + "bbox": [ + 511, + 476, + 849, + 491 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Sec. 4.1 reveals that the middle layers facilitate crucial fusion, integrating visual and textual inputs into cross-modal semantic representations that drive final predictions. Accordingly, this section will delve deeper into the attention to visual inputs throughout the modality fusion process.", + "bbox": [ + 511, + 498, + 905, + 574 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/2d3a106b79a394187f660d99430a78354ea5c9f784ed3b32aedd1ce4c42cd54d.jpg", + "image_caption": [ + "Figure 4. Attention Distribution of Modal Information Across Model Layers. In the middle layers, the model allocates insufficient attention to visual features while disproportionately focusing on system prompts." + ], + "image_footnote": [], + "bbox": [ + 524, + 590, + 880, + 773 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We define the attention allocation, $\\lambda$ , as the aggregate attention score assigned to a specific type of token within a single layer. Accordingly, the attention allocation for sys", + "bbox": [ + 511, + 854, + 905, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "14628", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "tem prompts, visual features, and user instructions in the $l$ -th layer can be computed as follows:", + "bbox": [ + 89, + 90, + 482, + 121 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\lambda_ {s y s} ^ {l} = \\sum_ {i \\in \\mathcal {T}} \\sum_ {j \\in \\mathcal {S}} A _ {l} (i, j),\n$$\n", + "text_format": "latex", + "bbox": [ + 207, + 133, + 367, + 166 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\lambda_ {v i s} ^ {l} = \\sum_ {i \\in \\mathcal {T}} \\sum_ {j \\in \\mathcal {V}} A _ {l} (i, j), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 207, + 170, + 482, + 202 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\lambda_ {i n s} ^ {l} = \\sum_ {i \\in \\mathcal {T}} \\sum_ {j \\in \\mathcal {T}} A _ {l} (i, j).\n$$\n", + "text_format": "latex", + "bbox": [ + 207, + 205, + 367, + 239 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this context, $A_{l}$ represents the attention matrix averaged across all attention heads, while $\\mathcal{S}$ represents the indices of system tokens. The measures $\\lambda_{sys}^{l},\\lambda_{vis}^{l}$ , and $\\lambda_{ins}^{l}$ provide insight into the distribution of attention to different modalities across various layers, aiding in understanding the reasons for the underutilization of visual information during the modality fusion process.", + "bbox": [ + 89, + 244, + 483, + 349 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The experimental setup aligns with that described in Sec. 4.1. Fig. 4 illustrates the allocation of attention to different modalities across the model's layers. In the middle layers, attention to visual features is markedly lower than that given to system prompts and user instructions. This suggests that during the critical process of modality fusion, the model's focus on visual input is insufficient. As a result, visual information is underutilized, leading to an output distribution skewed toward language priors.", + "bbox": [ + 89, + 351, + 483, + 487 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Insights", + "text_level": 1, + "bbox": [ + 89, + 497, + 189, + 513 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Based on the experimental results presented in Sec. 4.1 and Sec. 4.2, two significant conclusions can be drawn:", + "bbox": [ + 89, + 520, + 482, + 549 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- The model performs the crucial fusion of visual and textual modalities in the middle layers, creating cross-modal semantic representations that drive the final predictions.", + "- During this critical fusion process, the model demonstrates inadequate attention to the visual modality." + ], + "bbox": [ + 89, + 553, + 482, + 628 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "These findings indicate that models fail to fully utilize visual information, resulting in an excessive dependence on language priors and, subsequently, the occurrence of hallucination phenomena.", + "bbox": [ + 89, + 631, + 483, + 691 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Visual Amplification Fusion", + "text_level": 1, + "bbox": [ + 89, + 705, + 348, + 723 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Building on the insights presented in Sec. 4, we introduce a hallucination mitigation method called Visual Amplification Fusion (VAF). As illustrated in Fig. 5, This approach heightens attention to visual information during modality fusion, effectively reducing the excessive dependency on language priors and ensuring that the generated content is closely grounded to visual inputs.", + "bbox": [ + 89, + 731, + 483, + 838 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Attention Redistribution", + "text_level": 1, + "bbox": [ + 89, + 848, + 316, + 862 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As outlined in Sec. 4, the model performs crucial fusion of visual and textual modalities within the middle layers.", + "bbox": [ + 89, + 869, + 482, + 900 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/4276ac80e92a8e4d964523fdd4d403e279bcfb08f04d187e5d1222c1e38e3c7a.jpg", + "image_caption": [ + "Figure 5. Illustration of the Visual Amplification Fusion Method. In the middle layers, we select attention heads highly responsive to visual information, amplifying their focus on visual features while reducing unnecessary attention to system prompts." + ], + "image_footnote": [], + "bbox": [ + 532, + 88, + 883, + 239 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "However, the attention allocated to visual modality information during this process remains insufficient. To address this, we adjust the attention weights in these layers to achieve a more balanced focus.", + "bbox": [ + 511, + 323, + 905, + 382 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/f1a5946587d4e3f726094e6779b36975910feb9152052a38d0d7fc97e10b0959.jpg", + "image_caption": [ + "Figure 6. Effect of Enhanced Visual Attention on Hallucination Suppression. Increasing attention to visual features in the fusion process of the model's middle layers successfully reduces hallucinations, enabling the model to correct its grape color prediction from \"green\" to \"red\"." + ], + "image_footnote": [], + "bbox": [ + 532, + 392, + 885, + 555 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Let $A_{l,h}$ denote the attention matrix of the $h$ -th attention head in the $l$ -th layer, and $Z_{l,h}$ represent its corresponding attention score matrix, defined as:", + "bbox": [ + 511, + 652, + 905, + 698 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nA _ {l, h} = \\operatorname {s o f t m a x} \\left(Z _ {l, h}\\right). \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 627, + 707, + 903, + 723 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our objective during the modality fusion process is to amplify the model's attention to visual features while curbing an overemphasis on system prompts. This adjustment facilitates improved integration of visual information and reduces over-reliance on language priors. To achieve this, we modify the attention score matrix in the middle layers (i.e., $8 < l < 15$ ) as follows:", + "bbox": [ + 511, + 731, + 906, + 837 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {Z} _ {l, h} = Z _ {l, h} + \\alpha \\cdot M _ {l, h} ^ {\\text {e n h}} \\circ Z _ {l, h} - \\beta \\cdot M _ {l, h} ^ {\\text {s u p}} \\circ Z _ {l, h}. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 531, + 844, + 903, + 864 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here, $\\alpha$ is the enhancement coefficient $(\\alpha >0)$ , where larger values indicate stronger amplification of visual attenuate", + "bbox": [ + 511, + 869, + 905, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "14629", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "tion. The suppression coefficient $\\beta$ ( $0 < \\beta < 1$ ) determines the extent of attention suppression directed at system prompts. The enhancement and suppression mask matrices, $M_{l,h}^{enh}$ and $M_{l,h}^{sup}$ , respectively, are defined to guide the modulation of attention elements:", + "bbox": [ + 89, + 90, + 482, + 165 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nM _ {l, h} ^ {e n h} (i, j) = \\mathbb {I} (i \\in \\mathcal {T}, j \\in \\mathcal {V}), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 181, + 172, + 482, + 196 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nM _ {l, h} ^ {s u p} (i, j) = \\mathbb {I} (i \\in \\mathcal {T}, j \\in \\mathcal {S}).\n$$\n", + "text_format": "latex", + "bbox": [ + 186, + 193, + 388, + 212 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "These modifications optimize attention allocation by enhancing the model's focus on visual features during modality fusion and minimizing superfluous attention to system prompts. As illustrated in Fig. 6, preliminary analysis indicates that this approach effectively mitigates hallucination issues by promoting greater attention to visual information.", + "bbox": [ + 89, + 220, + 483, + 311 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Visual Perception Restriction", + "text_level": 1, + "bbox": [ + 89, + 319, + 352, + 335 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Enhancing visual attention across all attention heads in the middle layers can be overly aggressive and may negatively impact content generation. To address this, we propose a selective enhancement strategy. Specifically, we identify and isolate the attention heads that exhibit higher sensitivity to visual information, which we term visual perception heads. We then restrict the visual attention enhancement to these visual perception heads, ensuring better utilization of visual information while maintaining overall model performance.", + "bbox": [ + 89, + 340, + 483, + 477 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In the model, attention heads that allocate more attention to visual features demonstrate heightened sensitivity to visual information. Let $A_{l,h}$ represent the attention matrix of the $h$ -th attention head in the $l$ -th layer of the model, with its corresponding visual attention allocation denoted by $\\lambda_{\\mathrm{vis}}^{l,h}$ . In each attention layer, we identify the attention heads whose visual attention allocation fall within the top", + "bbox": [ + 89, + 478, + 483, + 583 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "$50\\%$ and designate them as visual perception heads, subsequently redistributing their attention. The attention matrices of the remaining attention heads are kept unchanged.", + "bbox": [ + 511, + 90, + 905, + 137 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6. Experiment", + "text_level": 1, + "bbox": [ + 511, + 150, + 638, + 167 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "This section demonstrates the effectiveness of the proposed VAF method in mitigating hallucinations. Sec. 6.1 outlines the experimental setup, detailing the evaluation benchmarks and VAF parameter configurations. Sec. 6.2 then presents the experimental results from three perspectives: reduction of hallucinations, coherence of generated content, and inference speed. Finally, Sec. 6.3 further verifies the contribution of each VAF component through ablation studies.", + "bbox": [ + 511, + 176, + 906, + 297 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6.1. Experimental Settings", + "text_level": 1, + "bbox": [ + 511, + 306, + 720, + 323 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In Sec. 6.1.1, we present the selected datasets and evaluation metrics. Sec. 6.1.2 details the chosen MLLM backbone models, and Sec. 6.1.3 outlines the baseline settings.", + "bbox": [ + 511, + 329, + 905, + 375 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6.1.1. Datasets & Evaluation Metrics", + "text_level": 1, + "bbox": [ + 511, + 383, + 772, + 397 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Polling-based Object Probing Evaluation (POPE). POPE [30] is a novel framework designed to evaluate object hallucinations in MLLMs. Departing from traditional caption-based approaches, POPE frames hallucination detection as a binary task by posing straightforward yes-or-no questions regarding the presence of specific objects in an image (e.g., \"Is there a chair in the image?\"). Performance on POPE is measured across four metrics: Accuracy, Precision, Recall, and F1 score, allowing for a thorough evaluation of hallucinations in MLLMs.", + "bbox": [ + 511, + 402, + 905, + 551 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Multimodal Model Evaluation (MME). MME [13] benchmark provides a comprehensive framework for evalu", + "bbox": [ + 511, + 551, + 905, + 584 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/8d459acb9b44953bf440c455acef2cbaa8db11da66027968c375775beab5b9ba.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
CategoryMethodLLaVA-v1.5-7BLLaVA-v1.5-13BQwen-VL-Chat-7B
AccuracyF1-scoreAccuracyF1-scoreAccuracyF1-score
RandomRegular87.8 ↑ 0.087.5 ↑ 0.087.6 ↑ 0.087.4 ↑ 0.088.2 ↑ 0.087.9 ↑ 0.0
VCD88.4 ↑ 0.687.7 ↑ 0.288.9 ↑ 1.387.8 ↑ 0.489.1 ↑ 0.988.4 ↑ 0.5
ICD88.1 ↑ 0.387.6 ↑ 0.188.1 ↑ 0.587.6 ↑ 0.288.9 ↑ 0.788.1 ↑ 0.2
VAF89.6 ↑ 1.889.3 ↑ 1.890.1 ↑ 2.589.9 ↑ 2.590.0 ↑ 1.889.7 ↑ 1.8
PopularRegular82.5 ↑ 0.083.2 ↑ 0.082.7 ↑ 0.084.1 ↑ 0.082.4 ↑ 0.083.1 ↑ 0.0
VCD83.1 ↑ 0.684.1 ↑ 0.983.7 ↑ 1.085.1 ↑ 1.083.0 ↑ 0.684.1 ↑ 1.0
ICD82.1 ↓ 0.482.9 ↓ 0.382.9 ↑ 0.284.3 ↑ 0.283.2 ↑ 0.884.5 ↑ 1.4
VAF84.5 ↑ 2.084.9 ↑ 1.785.2 ↑ 2.586.4 ↑ 2.384.9 ↑ 2.585.1 ↑ 2.0
AdversarialRegular77.6 ↑ 0.079.4 ↑ 0.077.8 ↑ 0.079.5 ↑ 0.077.2 ↑ 0.078.9 ↑ 0.0
VCD78.1 ↑ 0.579.6 ↑ 0.278.2 ↑ 0.479.7 ↑ 0.278.8 ↑ 1.680.1 ↑ 1.2
ICD78.5 ↑ 0.979.9 ↑ 0.579.1 ↑ 1.380.1 ↑ 0.678.1 ↑ 0.979.2 ↑ 0.3
VAF80.1 ↑ 2.581.0 ↑ 1.680.7 ↑ 2.981.7 ↑ 2.280.4 ↑ 3.281.2 ↑ 2.3
", + "bbox": [ + 163, + 604, + 841, + 857 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Performance on POPE. Results are averaged across the MS-COCO, A-OKVQA, and GQA datasets. The VAF method demonstrates superior hallucination suppression across all three MLLMs. The best performance for each setting is highlighted in red.", + "bbox": [ + 89, + 867, + 905, + 897 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "14630", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/b83f59448363e37e83379459d404902b696ee19afef140d8b901b37a9a94927b.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelMethodObject-levelAttribute-levelTotal Score
ExistenceCountPositionColor
LLaVA-v1.5-7BRegular185.00↑0.00146.67↑0.00128.33↑0.00150.00↑0.00610.00↑0.00
VCD185.00↑0.00141.33↓5.34128.33↑0.00153.00↑3.00607.66↓2.34
ICD185.00↑0.00148.33↑1.66126.66↓1.67148.33↓1.67608.32↓1.68
VAF195.00↑10.00158.33↑11.66128.33↑0.00155.00↑5.00636.67↑26.67
LLaVA-v1.5-13BRegular185.00↑0.00155.00↑0.00133.33↑0.00165.00↑0.00638.33↑0.00
VCD185.00↑0.00155.00↑0.00130.00↓3.33168.33↑3.33638.33↑0.00
ICD183.33↓1.67153.33↓1.67131.67↓1.66165.00↑0.00633.33↓5.00
VAF195.00↑10.00160.00↑5.00136.67↑3.34170.00↑5.00661.67↑23.34
Qwen-VL-7BRegular158.33↑0.00150.00↑0.00128.33↑0.00170.00↑0.00606.66↑0.00
VCD158.33↑0.00150.00↑0.00133.33↑5.00175.00↑5.00616.66↑10.00
ICD128.33↓30.00151.67↑1.67128.33↑0.00170.00↑0.00578.33↓28.33
VAF165.00↑6.67155.00↑5.00133.33↑5.00175.00↑5.00628.33↑21.67
", + "bbox": [ + 112, + 88, + 890, + 343 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Results on the MME subset. Across three MLLMs, the VAF method achieved the most effective suppression of both object-level and attribute-level hallucinations. The highest scores in each setting are highlighted in red.", + "bbox": [ + 89, + 352, + 906, + 383 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ating MLLMs across both perceptual and cognitive dimensions. It consists of ten perception-oriented tasks and four cognition-oriented tasks, with model performance assessed through accuracy metrics. In addition to the full dataset, we leverage specific subsets, such as object existence and counting to analyze object-level hallucinations, while position and color subsets are employed to examine attributelevel hallucinations.", + "bbox": [ + 88, + 407, + 480, + 527 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Novel Object Captioning at Scale (Nocaps). NoCaps [3] benchmark is designed to evaluate image captioning models on their ability to describe novel objects absent from standard datasets like COCO. Model performance is quantified using the CIDEr score, providing a basis to assess the coherence and accuracy of generated captions in response to images containing unfamiliar objects.", + "bbox": [ + 88, + 527, + 482, + 635 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.1.2. MLLM Backbones", + "text_level": 1, + "bbox": [ + 89, + 643, + 269, + 657 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In comparison to the Q-former structure, linear projection demonstrates greater efficiency in aligning visual and textual features. This advantage is evident in MLLMs with linear projection architectures, such as LLaVA and Qwen-VL, which outperform Q-former-based MLLMs like Instruct-BLIP and MiniGPT4. Based on these findings, we selected three linear-projection-based MLLMs, specifically LLaVA-v1.5-7B, LLaVA-v1.5-13B [35], and Qwen-VL-7B [5], to evaluate the effectiveness of our proposed VAF method. Detailed prompt templates for each model across various benchmarks are included in Sec. 10.", + "bbox": [ + 88, + 661, + 482, + 828 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.1.3. Baseline Settings.", + "text_level": 1, + "bbox": [ + 89, + 835, + 259, + 852 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We primarily compared our approach to the VCD [23] and ICD [47] methods. VCD mitigates hallucinations by contrasting output distributions derived from original and dis", + "bbox": [ + 89, + 854, + 483, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "torted visual inputs, while ICD reduces hallucinated concepts by comparing distributions generated with standard versus disrupted instructions. To ensure consistency and reproducibility in our comparisons, all methods use greedy search. Unless specified otherwise, our experiments set $\\beta = 0.1$ and $\\alpha = 0.15$ .", + "bbox": [ + 511, + 407, + 906, + 500 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.2. Results and Analysis", + "text_level": 1, + "bbox": [ + 511, + 508, + 709, + 523 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Sec. 6.2.1 examines the effectiveness of various methods in mitigating hallucinations, while Sec. 6.2.2 assesses their impact on the quality of generated content. Sec. 6.2.3 then analyzes the influence of each method on inference speed. Additional experimental results are provided in Sec. 8.", + "bbox": [ + 511, + 530, + 905, + 606 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.2.1. Hallucination Mitigation", + "text_level": 1, + "bbox": [ + 511, + 613, + 732, + 628 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Tab. 2 presents the experimental results of the VAF method on the POPE benchmark, with results averaged across the MSCOCO [31], A-OKVQA [42], and GQA [20] datasets. Applied to both the LLaVA-v1.5 model family and the Qwen-VL model, the VAF method consistently surpasses the VCD and ICD methods in reducing hallucinations. Tab. 3 further highlights the performance of VAF on the MME benchmark, demonstrating its effectiveness in suppressing both object-level and attribute-level hallucinations.", + "bbox": [ + 511, + 632, + 905, + 768 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.2.2. Coherence of Generated Content", + "text_level": 1, + "bbox": [ + 511, + 776, + 790, + 789 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Tab. 4 presents the experimental results for various methods on the Nocaps and ScienceQA datasets. It is evident that VCD and ICD substantially degrade the quality of the generated content. Specifically, on the Nocaps dataset, VCD and ICD reduce CIDEr scores by $18\\%$ and $27\\%$ , respectively. This degradation primarily arises from the crude disruption of language priors by contrastive decoding methods,", + "bbox": [ + 511, + 794, + 905, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "14631", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/0b9e784b26f293f21ef2cd61b90fdbb90125e2e44a7f7f3a3b426a35298e190f.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelDecodingScienceQANocaps
AccuracyCIDEr
LLaVA-v1.5-7BRegular68.078.7
VCD64.565.7
ICD62.462.3
VAF68.578.8
LLaVA-v1.5-13BRegular71.682.6
VCD70.068.9
ICD69.260.3
VAF71.782.3
", + "bbox": [ + 101, + 89, + 478, + 275 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "which leads to generated content lacking coherence and accuracy. By contrast, our method demonstrates minimal negative impact on prediction results, maintaining both coherence and accuracy effectively.", + "bbox": [ + 89, + 340, + 482, + 402 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6.2.3. Inference Speed", + "text_level": 1, + "bbox": [ + 89, + 409, + 250, + 425 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Fig. 7 illustrates the impact of different strategies on model inference speed within the Nocaps dataset. In comparison, the VCD and ICD methods nearly double the inference time due to the need to process contrastive input samples, whereas the VAF method has minimal impact on the inference speed of multimodal large language models.", + "bbox": [ + 89, + 428, + 483, + 518 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b27fbe02e4ad6e6fd802441e5658fbec11550e4ce18c5748aac63639a42b5f9d.jpg", + "image_caption": [ + "Figure 7. Comparison of different strategies on inference speed. The VCD and ICD methods reduce inference speed by $50\\%$ , whereas the VAF method shows minimal impact." + ], + "image_footnote": [], + "bbox": [ + 112, + 537, + 442, + 714 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6.3. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 787, + 243, + 804 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ablation studies on the enhancement coefficient $\\alpha$ were conducted using the COCO-Random dataset within the POPE benchmark to understand its influence on model performance. Fig. 8 demonstrates that when $0 < \\alpha < 0.25$ , model hallucinations are effectively suppressed. However, when $\\alpha$ surpasses 0.25, performance starts to degrade. We", + "bbox": [ + 89, + 810, + 483, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "propose that this reduction in performance may stem from an excessive focus on visual features, disrupting the balanced integration of language information and diminishing overall model effectiveness.", + "bbox": [ + 511, + 90, + 905, + 151 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a46d5daaa6e27ef0a8238450f81f2345dcd1ab7d14cb92d2d08010dacc032f81.jpg", + "image_caption": [ + "Figure 8. Ablation study of $\\alpha$ on the POPE benchmark." + ], + "image_footnote": [], + "bbox": [ + 537, + 162, + 879, + 342 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We performed ablation studies on the visual perception restriction mechanism, evaluating its impact on the POPE and Nocaps benchmarks. Tab. 5 highlights the effects of restricting attention reallocation to visual perception heads. Increasing attention to visual features alone reduces model hallucinations, while confining this reallocation strategy to visual perception heads minimizes adverse effects on content quality. More ablation studies can be found in Sec. 9.", + "bbox": [ + 511, + 371, + 906, + 492 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/9151df19ce8ee99f5549b71b1bbb4e4591341c7229e476bc904b1ad988e237de.jpg", + "table_caption": [ + "Table 4. Results on SQA and Nocaps datasets. The highest and second-highest scores are marked in red and blue, respectively." + ], + "table_footnote": [], + "table_body": "
ModelVisual RestrictionPOPENocaps
LLaVA-7B✓89.878.8
X89.976.4
LLaVA-13B✓90.282.3
X90.081.1
", + "bbox": [ + 527, + 503, + 893, + 608 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. Ablation Study of Visual Perception Restriction Mechanism. Restricting attention redistribution to the visual perception heads more effectively preserves the quality of generated content.", + "bbox": [ + 511, + 619, + 906, + 662 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. conclusion", + "text_level": 1, + "bbox": [ + 513, + 679, + 627, + 694 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we identify two key drawbacks of using contrastive decoding to mitigate hallucinations in MLLMs: reduced quality of generated content and slower inference speed. To address these challenges, we propose a novel approach, Visual Amplification Fusion, which effectively mitigates hallucinations while preserving both inference speed and content generation quality. By enhancing the attention to visual features during modality fusion, VAF minimizes the over-reliance on language priors, ensuring a high degree of consistency between generated content and visual inputs. Extensive experiments across multiple benchmarks and MLLMs demonstrate that VAF provides a clear advantage in hallucination mitigation.", + "bbox": [ + 511, + 704, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "14632", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 91, + 90, + 258, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work is supported by the National Natural Science Foundation of China under Grant 62176246. This work is also supported by Anhui Province Key Research and Development Plan (202304a05020045), Anhui Province Natural Science Foundation (2208085UD17) and National Natural Science Foundation of China under Grant 62406098.", + "bbox": [ + 89, + 114, + 483, + 205 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 217, + 187, + 233 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Vedika Agarwal, Rakshith Shetty, and Mario Fritz. Towards causal vqa: Revealing and reducing spurious correlations by invariant and covariant semantic editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9690-9698, 2020. 1", + "[2] Aishwarya Agrawal, Dhruv Batra, and Devi Parikh. Analyzing the behavior of visual question answering models. arXiv preprint arXiv:1606.07356, 2016. 1", + "[3] Harsh Agrawal, Karan Desai, Yufei Wang, Xinlei Chen, Rishabh Jain, Mark Johnson, Dhruv Batra, Devi Parikh, Stefan Lee, and Peter Anderson. nocaps: novel object captioning at scale. In 2019 IEEE/CVF International Conference on Computer Vision (ICCV). IEEE, 2019. 7", + "[4] Jinze Bai, Shuai Bai, and et al. Qwen technical report. arXiv preprint arXiv:2309.16609, 2023. 2", + "[5] Jinze Bai, Shuai Bai, and et al. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966, 2023. 2, 7", + "[6] Rohan Bavishi, Erich Elsen, and et al. Introducing our multimodal models, 2023. 2", + "[7] Ali Furkan Biten, Lluís Gómez, and Dimosthenis Karatzas. Let there be a clock on the beach: Reducing object hallucination in image captioning. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1381-1390, 2022. 1", + "[8] Keqin Chen, Zhao Zhang, and et al. Shikra: Unleashing multimodal llm's referential dialogue magic. arXiv preprint arXiv:2306.15195, 2023. 1", + "[9] Long Chen, Oleg Sinavski, Jan Hunermann, Alice Karnsund, Andrew James Willmott, Danny Birch, Daniel Maund, and Jamie Shotton. Driving with llms: Fusing object-level vector modality for explainable autonomous driving. arXiv preprint arXiv:2310.01957, 2023. 1", + "[10] Zhe Chen, Weiyun Wang, and et al. How far are we to gpt-4v? closing the gap to commercial multimodal models with open-source suites. arXiv preprint arXiv:2404.16821, 2024. 2", + "[11] Wei-Lin Chiang and Zhuohan et al Li. Vicuna: An opensource chatbot impressing gpt-4 with $90\\%$ chatgpt quality. See https://vicuna.lmsys.org (accessed 14 April 2023), 2023. 2", + "[12] Wenliang Dai and Junnan Li et al. Instructlip: Towards general-purpose vision-language models with instruction tuning, 2023. 2", + "[13] Chaoyou Fu, Peixian Chen, and et al. Mme: A comprehensive evaluation benchmark for multimodal large language models. arXiv preprint arXiv:2306.13394, 2023. 6" + ], + "bbox": [ + 93, + 243, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Bartra, and Devi Parikh. Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6904-6913, 2017. 1", + "[15] Anisha Gunjal, Jihan Yin, and Erhan Bas. Detecting and preventing hallucinations in large vision language models. arXiv preprint arXiv:2308.06394, 2023. 1", + "[16] Vipul Gupta, Zhuowan Li, Adam Kortylewski, Chenyu Zhang, Yingwei Li, and Alan Yuille. Swapmix: Diagnosing and regularizing the over-reliance on visual context in visual question answering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5078-5088, 2022. 1", + "[17] Yudong Han, Liqiang Nie, Jianhua Yin, Jianlong Wu, and Yan Yan. Visual perturbation-aware collaborative learning for overcoming the language prior problem. arXiv preprint arXiv:2207.11850, 2022. 1", + "[18] Mingzhe Hu, Shaoyan Pan, Yuheng Li, and Xiaofeng Yang. Advancing medical imaging with language models: A journey from n-grams to chatgpt. arXiv preprint arXiv:2304.04920, 2023. 1", + "[19] Qidong Huang, Xiaoyi Dong, Pan Zhang, Bin Wang, Conghui He, Jiaqi Wang, Dahua Lin, Weiming Zhang, and Nenghai Yu. Opera: Alleviating hallucination in multimodal large language models via over-trust penalty and retrospection-allocation. In CVPR, pages 13418-13427, 2024. 2", + "[20] Drew A Hudson and Christopher D Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering. In CVPR, pages 6700-6709, 2019. 7", + "[21] Fushuo Huo, Wenchao Xu, Zhong Zhang, Haozhao Wang, Zhicheng Chen, and Peilin Zhao. Self-introspective decoding: Alleviating hallucinations for large vision-language models, 2024. 2", + "[22] Chaoya Jiang, Haiyang Xu, and et al. Hallucination augmented contrastive learning for multimodal large language model. In CVPR, pages 27036-27046, 2024. 2", + "[23] Sicong Leng, Hang Zhang, and et al. Mitigating object hallucinations in large vision-language models through visual contrastive decoding. In CVPR, pages 13872-13882, 2024. 2, 7", + "[24] Bo Li, Yuanhan Zhang, and et al. Mimic-it: Multi-modal in-context instruction tuning. arXiv preprint arXiv:2306.05425, 2023. 2", + "[25] Bo Li, Kaichen Zhang, and et al. Llava next: Stronger llms supercharge multimodal capabilities in the wild, 2024. 2", + "[26] Chunyuan Li, Cliff Wong, and et al. Llava-med: Training a large language-and-vision assistant for biomedicine in one day. In NeurIPS, pages 28541-28564, 2023. 2", + "[27] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. BLIP-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML, 2023. 1, 2", + "[28] Xiang Lisa Li, Ari Holtzman, and et al. Contrastive decoding: Open-ended text generation as optimization. arXiv preprint arXiv:2210.15097, 2022. 2" + ], + "bbox": [ + 516, + 92, + 905, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "14633", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[29] Yifan Li, Yifan Du, Kun Zhou, Jinpeng Wang, Wayne Xin Zhao, and Ji-Rong Wen. Evaluating object hallucination in large vision-language models. arXiv preprint arXiv:2305.10355, 2023. 1", + "[30] Yifan Li, Yifan Du, Kun Zhou, Jinpeng Wang, Xin Zhao, and Ji-Rong Wen. Evaluating object hallucination in large vision-language models. In EMNLP, pages 292-305, 2023. 6", + "[31] Tsung-Yi Lin, Michael Maire, and et al. Microsoft coco: Common objects in context. In ECCV, pages 740-755, 2014. 7", + "[32] Fuxiao Liu, Kevin Lin, Linjie Li, Jianfeng Wang, Yaser Ya-coob, and Lijuan Wang. Mitigating hallucination in large multi-modal models via robust instruction tuning. arXiv preprint arXiv:2306.14565, 2023. 1", + "[33] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, pages 34892-34916, 2023. 1, 2", + "[34] Haokun Liu, Yaonan Zhu, Kenji Kato, Izumi Kondo, Tadayoshi Aoyama, and Yasuhisa Hasegawa. Lm-based human-robot collaboration framework for manipulation tasks. arXiv preprint arXiv:2308.14972, 2023. 1", + "[35] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In CVPR, pages 26296-26306, 2024. 1, 7", + "[36] Zhi-Song Liu, Robin Courant, and Vicky Kalogeiton. Funnynet-w: Multimodal learning of funny moments in videos in the wild. International Journal of Computer Vision, pages 1-22, 2024. 2", + "[37] Holy Lvenia, Wenliang Dai, Samuel Cahyawijaya, Ziwei Ji, and Pascale Fung. Negative object presence evaluation (nope) to measure object hallucination in vision-language models. arXiv preprint arXiv:2310.05338, 2023. 1", + "[38] Pan Lu, Swaroop Mishra, Tony Xia, Liang Qiu, Kai-Wei Chang, Song-Chun Zhu, Oyvind Tafjord, Peter Clark, and Ashwin Kalyan. Learn to explain: Multimodal reasoning via thought chains for science question answering. In The 36th Conference on Neural Information Processing Systems (NeurIPS), 2022. 3", + "[39] Jinjie Mai, Jun Chen, Bing Li, Guocheng Qian, Mohamed Elhoseiny, and Bernard Ghanem. Llm as a robotic brain: Unifying egocentric memory and control. arXiv preprint arXiv:2304.09349, 2023. 1", + "[40] AI Meta. Introducing meta llama 3: The most capable openly available llm to date. Meta AI, 2024. 2", + "[41] Yulei Niu, Kaihua Tang, Hanwang Zhang, Zhiwu Lu, XianSheng Hua, and Ji-Rong Wen. Counterfactual vqa: A cause-effect look at language bias. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12700-12710, 2021. 1", + "[42] Dustin Schwenk, Apoorv Khandelwal, and et al. A-okvqa: A benchmark for visual question answering using world knowledge. In ECCV, pages 146–162, 2022. 7", + "[43] Rohan Taori, Ishaan Gulrajani, and et al. Stanford alpaca: an instruction-following llama model (2023). URL https://github.com/tatsu-lab/stanford_alpaca, 1(9), 2023. 2" + ], + "bbox": [ + 91, + 90, + 480, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[44] Hugo Touvron, Thibaut Lavril, and et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023.", + "[45] Hugo Touvron, Louis Martin, and et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 2", + "[46] Sheng Wang, Zihao Zhao, Xi Ouyang, Qian Wang, and Dinggang Shen. Chatcad: Interactive computer-aided diagnosis on medical image using large language models. arXiv preprint arXiv:2302.07257, 2023. 1", + "[47] Xintong Wang, Jingheng Pan, and et al. Mitigating hallucinations in large vision-language models with instruction contrastive decoding. arXiv preprint arXiv:2403.18715, 2024.2, 7", + "[48] Yike Wu, Yu Zhao, Shiwan Zhao, Ying Zhang, Xiaojie Yuan, Guoqing Zhao, and Ning Jiang. Overcoming language priors in visual question answering via distinguishing superficially similar instances. arXiv preprint arXiv:2209.08529, 2022. 1", + "[49] Zhenyu Wu, Ziwei Wang, Xiuwei Xu, Jiwen Lu, and Haibin Yan. Embodied task planning with large language models. arXiv preprint arXiv:2307.01848, 2023. 1", + "[50] Hong Yan, Lijun Liu, Xupeng Feng, and Qingsong Huang. Overcoming language priors with self-contrastive learning for visual question answering. Multimedia Tools and Applications, 82(11):16343–16358, 2023. 1", + "[51] Qinghao Ye, Haiyang Xu, and et al. mplug-owl: Modularization empowers large language models with multimodality. arXiv preprint arXiv:2304.14178, 2023. 2", + "[52] Shilong Zhang, Peize Sun, and et al. Gpt4roi: Instruction tuning large language model on region-of-interest. arXiv preprint arXiv:2307.03601, 2023. 2", + "[53] Ren Zhibo, Wang Huizhen, Zhu Muhua, Wang Yichao, Xiao Tong, and Zhu Jingbo. Overcoming language priors with counterfactual inference for visual question answering. In Proceedings of the 22nd Chinese National Conference on Computational Linguistics, pages 600-610, 2023. 1", + "[54] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. IJCV, 130(9):2337-2348, 2022. 1", + "[55] Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 1, 2" + ], + "bbox": [ + 514, + 90, + 905, + 700 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "14634", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/ClearSight_ Visual Signal Enhancement for Object Hallucination Mitigation in Multimodal Large Language Models/cc596546-7afd-468d-96ec-564a706a95d3_model.json b/2025/ClearSight_ Visual Signal Enhancement for Object Hallucination Mitigation in Multimodal Large Language Models/cc596546-7afd-468d-96ec-564a706a95d3_model.json new file mode 100644 index 0000000000000000000000000000000000000000..25cb375e72c294e55d3d759d3f7cbc5242500bed --- /dev/null +++ b/2025/ClearSight_ Visual Signal Enhancement for Object Hallucination Mitigation in Multimodal Large Language Models/cc596546-7afd-468d-96ec-564a706a95d3_model.json @@ -0,0 +1,2486 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.104, + 0.13, + 0.895, + 0.177 + ], + "angle": 0, + "content": "ClearSight: Visual Signal Enhancement for Object Hallucination Mitigation in Multimodal Large Language Models" + }, + { + "type": "text", + "bbox": [ + 0.312, + 0.204, + 0.689, + 0.221 + ], + "angle": 0, + "content": "Hao Yin Gunagzong Si Zilei Wang\\*" + }, + { + "type": "text", + "bbox": [ + 0.312, + 0.223, + 0.686, + 0.24 + ], + "angle": 0, + "content": "University of Science and Technology of China" + }, + { + "type": "text", + "bbox": [ + 0.232, + 0.242, + 0.76, + 0.257 + ], + "angle": 0, + "content": "{yinhnavi, guangzongsi}@mail.ustc.edu.cn, zlwang@ustc.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.292, + 0.327, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.324, + 0.486, + 0.687 + ], + "angle": 0, + "content": "Contrastive decoding strategies are widely used to mitigate object hallucinations in multimodal large language models (MLLMs). By reducing over-reliance on language priors, these strategies ensure that generated content remains closely grounded in visual inputs, producing contextually accurate outputs. Since contrastive decoding requires no additional training or external tools, it offers both computational efficiency and versatility, making it highly attractive. However, these methods present two main limitations: (1) bluntly suppressing language priors can compromise coherence and accuracy of generated content, and (2) processing contrastive inputs adds computational load, significantly slowing inference speed. To address these challenges, we propose Visual Amplification Fusion (VAF), a plug-and-play technique that enhances attention to visual signals within the model's middle layers, where modality fusion predominantly occurs. This approach enables more effective capture of visual features, reducing the model's bias toward language modality. Experimental results demonstrate that VAF significantly reduces hallucinations across various MLLMs without affecting inference speed, while maintaining coherence and accuracy in generated outputs. The code is available at https://github.com/ustc-hyin/ClearSight." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.716, + 0.222, + 0.731 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.741, + 0.484, + 0.879 + ], + "angle": 0, + "content": "In recent years, MLLMs [8, 27, 33, 35, 54, 55] have achieved remarkable progress in the intersecting fields of computer vision and natural language processing, and have been widely applied in tasks such as image captioning and visual question answering. However, these models often encounter the issue of \"object hallucination\" [15, 29, 32, 37] in practical applications, where the generated textual descriptions do not match the actual objects in the image. This problem highlights an over-reliance on unimodal pri" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.293, + 0.905, + 0.354 + ], + "angle": 0, + "content": "ors (especially language priors) [17, 48, 50, 53] during inference, posing potential risks in high-precision applications such as medical diagnosis [18, 46] and autonomous driving [9, 34, 39, 49]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.355, + 0.907, + 0.522 + ], + "angle": 0, + "content": "To address object hallucination [1, 2, 7, 14], several Contrastive Decoding strategies have been introduced in recent years. Among these, the Visual Contrastive Decoding (VCD) method has shown promise in reducing hallucinations by contrasting output distributions from both original and perturbed visual inputs, thus mitigating the model's excessive reliance on language priors [16, 41]. Notably, contrastive decoding methods do not require additional training or external tools, offering both computational efficiency and versatility, which has garnered them significant attention. However, these methods present two main limitations:" + }, + { + "type": "title", + "bbox": [ + 0.54, + 0.535, + 0.783, + 0.551 + ], + "angle": 0, + "content": "Limitations of Contrastive Decoding" + }, + { + "type": "text", + "bbox": [ + 0.54, + 0.562, + 0.88, + 0.607 + ], + "angle": 0, + "content": "- While reducing over-reliance on language priors, these methods may compromise the coherence and accuracy of generated content." + }, + { + "type": "text", + "bbox": [ + 0.54, + 0.608, + 0.881, + 0.653 + ], + "angle": 0, + "content": "- Contrastive decoding necessitates separate processing of the original and contrastive inputs, which considerably increases inference time." + }, + { + "type": "list", + "bbox": [ + 0.54, + 0.562, + 0.881, + 0.653 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.673, + 0.907, + 0.824 + ], + "angle": 0, + "content": "To address these shortcomings, we hope to propose a training-free method that can effectively reduces hallucinations without compromising content quality or inference speed. Our saliency analysis of the model's attention maps reveals that biases toward language in generated content do not arise from an overemphasis on language signals but rather from insufficient attention on visual information during modality fusion. Based on this insight, we introduce a novel, plug-and-play technique to mitigate hallucinations: Visual Amplification Fusion (VAF)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Our analysis indicates that modality fusion in MLLMs primarily occurs within the middle layers. VAF specifically amplifies visual signals at these middle layers, enabling the model to capture more distinctive visual features during fusion, which in turn reduces false descriptions in generated" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.239, + 0.9 + ], + "angle": 0, + "content": "*Corresponding Author" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14625" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.168 + ], + "angle": 0, + "content": "text. This technique not only strengthens the model's visual representations but also retains the beneficial influence of language priors, thus preserving content quality. Furthermore, by eliminating the need to process contrastive samples, VAF maintains inference speed." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.168, + 0.483, + 0.304 + ], + "angle": 0, + "content": "Experimental results validate the effectiveness of the VAF method. Across multiple object hallucination benchmarks, VAF demonstrated notable performance gains, with improvements of approximately \\(3\\%\\) on POPE and \\(7\\%\\) on MME. In terms of coherence and accuracy of generated responses, VCD caused a roughly \\(19\\%\\) decrease on NoCaps, while VAF maintained content quality without negative impacts. Additionally, VCD reduced inference speed by \\(50\\%\\), whereas VAF had virtually no effect on inference speed." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.305, + 0.446, + 0.319 + ], + "angle": 0, + "content": "In summary, the main contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.32, + 0.483, + 0.365 + ], + "angle": 0, + "content": "- We identify the negative impacts of contrastive decoding methods on both the quality of generated content and model inference speed." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.366, + 0.483, + 0.409 + ], + "angle": 0, + "content": "- We analyze the modality fusion mechanism in MLLMs, highlighting its insufficient attention to visual information." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.411, + 0.483, + 0.456 + ], + "angle": 0, + "content": "- We introduce the VAF method, which effectively mitigates the object hallucination problem while maintaining inference speed, coherence, and accuracy." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.457, + 0.483, + 0.501 + ], + "angle": 0, + "content": "- We demonstrate the significant performance improvements of the VAF method across multiple object hallucination benchmarks." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.32, + 0.483, + 0.501 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.517, + 0.228, + 0.533 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.542, + 0.41, + 0.559 + ], + "angle": 0, + "content": "2.1. Multimodal Large Language Models" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.565, + 0.483, + 0.792 + ], + "angle": 0, + "content": "The development of MLLMs [26, 36, 51, 52] has advanced from BERT-based decoders to LLM-based architectures [4, 11, 40, 43-45], enabling improved multimodal relationship capture [6, 10, 24, 25]. Models like BLIP-2 [27] and miniGPT-4 [55] incorporate a Q-Former mechanism, which enhances the alignment between visual and textual inputs, allowing for more precise interactions across modalities. InstructBLIP [12] builds on this approach by adding task-specific instructions, which improve the model's understanding of context-sensitive visual semantics. LLaVA [33] and Qwen-VL [5] utilize simpler linear projection techniques that streamline the alignment process, resulting in improved overall performance on vision-language tasks. However, hallucination issues persist across MLLMs, posing a significant challenge that requires further research." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.803, + 0.376, + 0.82 + ], + "angle": 0, + "content": "2.2. Contrastive Decoding Strategies" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.483, + 0.902 + ], + "angle": 0, + "content": "In recent years, Contrastive Decoding [19, 21, 22, 28] has emerged as a technique to improve generative model accuracy through contrastive judgment, widely employed to address hallucinations in generated content. For instance, Visual Contrastive Decoding (VCD) [23] contrasts output" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.319 + ], + "angle": 0, + "content": "distributions derived from original and distorted visual inputs, effectively reducing the over-reliance on statistical bias and unimodal priors, two essential causes of object hallucinations. Similarly, Instruction Contrastive Decoding (ICD) [47] works by comparing distributions derived from standard and disrupted instructions, thereby removing hallucinated concepts from the original distribution. These contrastive methods help ground generated content closely to visual inputs, resulting in contextually accurate outputs. However, despite these advancements, contrastive decoding faces two primary limitations: slower inference speed and reduced coherence in generated content. To overcome these limitations, we propose the VAF method, which achieves effective hallucination reduction while preserving both inference speed and content coherence." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.33, + 0.774, + 0.347 + ], + "angle": 0, + "content": "3. Preliminary and Motivation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.356, + 0.907, + 0.448 + ], + "angle": 0, + "content": "In Sec. 3.1, we illustrate the working mechanism of contrastive decoding to mitigate hallucinations, using Visual Contrastive Decoding as an example. In Sec. 3.2, we analysis two main drawbacks of this approach: its potential to disrupt the coherence and accuracy of generated content, and its tendency to slow down model inference." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.454, + 0.719, + 0.47 + ], + "angle": 0, + "content": "3.1. Contrastive Decoding" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.476, + 0.906, + 0.582 + ], + "angle": 0, + "content": "We consider a MLLM parametrized by \\(\\theta\\). The model takes as input a textual query \\(x\\) and a visual input \\(v\\), where \\(v\\) provides contextual visual information to assist the model in generating a relevant response \\(y\\) to the textual query. The response \\(y\\) is sampled auto-regressively from the probability distribution conditioned on the query \\(x\\) and the visual context \\(v\\). Mathematically, this can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.605, + 0.59, + 0.905, + 0.626 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} y _ {t} \\sim p _ {\\theta} \\left(y _ {t} \\mid v, x, y _ {< t}\\right) \\tag {1} \\\\ \\propto \\exp \\operatorname {l o g i t} _ {\\theta} \\left(y _ {t} \\mid v, x, y _ {< t}\\right) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.633, + 0.906, + 0.663 + ], + "angle": 0, + "content": "where \\(y_{t}\\) denotes the token at time step \\(t\\), and \\(y_{< t}\\) represents the sequence of generated tokens up to the time step \\((t - 1)\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.664, + 0.907, + 0.845 + ], + "angle": 0, + "content": "To mitigate the issue of object hallucination in MLLMs, contrastive decoding techniques can be applied. Here, we present Visual Contrastive Decoding (VCD) as a representative approach, shown in Fig. 1. Specifically, given a textual query \\( x \\) and a visual input \\( v \\), the model generates two distinct output distributions: one conditioned on the original \\( v \\) and the other on the distorted visual input \\( v' \\), which is derived by applying pre-defined distortions (i.e., Gaussian noise mask) to \\( v \\). Then, a new contrastive probability distribution is computed by exploiting the differences between the two initially obtained distributions. The new contrastive distribution \\( p_{vcd} \\) is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.536, + 0.851, + 0.905, + 0.905 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} p _ {v c d} (y \\mid v, v ^ {\\prime}, x) = \\operatorname {s o f t m a x} \\left[ \\log \\mathrm {i t} _ {\\theta} (y \\mid v, x) + \\right. \\tag {2} \\\\ \\left. \\alpha \\cdot \\left(\\operatorname {l o g i t} _ {\\theta} (y \\mid v, x) - \\operatorname {l o g i t} _ {\\theta} (y \\mid v ^ {\\prime}, x)\\right) \\right], \\\\ \\end{array}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14626" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.168 + ], + "angle": 0, + "content": "where larger \\(\\alpha\\) values indicate a stronger amplification of differences between the two distributions (\\(\\alpha = 0\\) reduces to regular decoding). Essentially, VCD serves as a corrective mechanism, reducing hallucinations by contrasting against a distribution predisposed to favoring them." + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.18, + 0.48, + 0.324 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.334, + 0.484, + 0.418 + ], + "angle": 0, + "content": "Figure 1. Illustration of Visual Contrastive Decoding. The hallucinated object \"Teacher\" is suppressed by contrasting with an output distribution prone to hallucinations. This method has two main drawbacks: (1) additional processing of distorted visual inputs greatly increases inference time; (2) subtracting the language prior disrupts content coherence." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.444, + 0.406, + 0.46 + ], + "angle": 0, + "content": "3.2. Limitations of Contrastive Decoding" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.467, + 0.484, + 0.557 + ], + "angle": 0, + "content": "As contrastive decoding methods do not require training or external tools, they offer high computational efficiency and generalizability, attracting significant attention in academia. However, these methods still have two major drawbacks: a reduction in the quality of generated content and slower inference speed." + }, + { + "type": "image", + "bbox": [ + 0.122, + 0.572, + 0.445, + 0.742 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.753, + 0.483, + 0.809 + ], + "angle": 0, + "content": "Figure 2. Impact of VCD on Model Performance. CIDEr scores are reported on the Nocaps benchmark, while Accuracy is presented for the ScienceQA benchmark. The use of VCD leads to a significant decline in model performance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.484, + 0.903 + ], + "angle": 0, + "content": "While contrasting logits of \\( p_{\\theta}(y \\mid v, x) \\) and \\( p_{\\theta}(y \\mid v', x) \\) can help reduce over-reliance on language priors and mitigate hallucination in MLLMs-as evidenced by a \\( 4\\% \\) performance gain on the POPE benchmark using the VCD method-merely decreasing the influence of the language" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.212 + ], + "angle": 0, + "content": "modality on the output distribution may undermine the coherence of the generated content, potentially leading to prediction errors. This issue is less pronounced in straightforward object hallucination tasks, where responses are limited to binary options, such as \"yes\" or \"no\". However, in more complex tasks, including multiple-choice question answering and image caption, the impact of contrastive learning methods on content quality becomes more significant." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.212, + 0.907, + 0.349 + ], + "angle": 0, + "content": "To verify this, we applied VCD method to LLaVA-v1.5-7B and LLaVA-v1.5-13B models, assessing their performance on the ScienceQA [38] and NoCaps benchmarks. As illustrated in Fig. 2, our findings reveal that, following the application of VCD, model performance decreased by \\(5\\%\\) on ScienceQA and by a considerable \\(45\\%\\) on NoCaps. These results suggest that in tasks requiring nuanced natural language generation, contrastive decoding methods can substantially impair content quality." + }, + { + "type": "table", + "bbox": [ + 0.53, + 0.359, + 0.892, + 0.466 + ], + "angle": 0, + "content": "
ModelMethodScienceQANocaps
LLaVA-v1.5-7BRegular0.141s0.456s
VCD0.293s1.086s
LLaVA-v1.5-13BRegular0.222s0.602s
VCD0.459s1.372s
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.475, + 0.907, + 0.532 + ], + "angle": 0, + "content": "Table 1. Impact of VCD on Model Inference Speed. The table shows the average inference time per sample (in seconds) on the ScienceQA and Nocaps benchmarks. Applying the VCD method nearly doubled the inference time of the model." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.547, + 0.906, + 0.698 + ], + "angle": 0, + "content": "Contrastive decoding methods notably reduce inference speed because they require calculating the output distribution for additional contrastive samples. For instance, in VCD method, each visual input \\( v \\) necessitates computing the logits of both \\( p_{\\theta}(y \\mid v, x) \\) and \\( p_{\\theta}(y \\mid v', x) \\) separately. This doubles the computation load during inference compared to vanilla decoding. We evaluated the inference speed of VCD versus vanilla decoding on ScienceQA. The experimental results, shown in Tab. 1, reveal that VCD's inference time is almost double that of vanilla decoding." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.71, + 0.803, + 0.727 + ], + "angle": 0, + "content": "4. Visual Neglect in Modal Fusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.907, + 0.902 + ], + "angle": 0, + "content": "The primary objective of this section is to examine why MLLMs tend to rely excessively on language priors in their predictions. In Sec. 4.1, saliency analysis reveals that image tokens influence prediction outcomes mainly through interactions with instruction tokens within the middle layers. Sec. 4.2 then compares attention weights across different modalities, showing that the attention given to visual features is notably lower than that allocated to system prompts and user instructions. These findings indicate that visual information is often underutilized in the modality fusion process, resulting in an over-reliance on language priors." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14627" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.405, + 0.108 + ], + "angle": 0, + "content": "4.1. Mid-layer: Visual-Language Fusion" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.114, + 0.483, + 0.204 + ], + "angle": 0, + "content": "To uncover why MLLMs tend to overly rely on language priors and overlook visual content in prediction, it is necessary first to clarify how the model utilizes visual modality information. This section explores the influence of the visual modality on prediction outcomes from the perspective of visual information interaction." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.206, + 0.484, + 0.281 + ], + "angle": 0, + "content": "We employ the saliency technique, a widely used interpretability tool, to highlight key token interactions within the attention mechanism. Following established practices, we utilize Taylor expansion to compute saliency scores for each element of the attention matrix:" + }, + { + "type": "equation", + "bbox": [ + 0.197, + 0.292, + 0.483, + 0.334 + ], + "angle": 0, + "content": "\\[\nI _ {l} = \\left| \\sum_ {h} A _ {h, l} \\odot \\frac {\\partial \\mathcal {L} (x)}{\\partial A _ {h , l}} \\right|. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.346, + 0.484, + 0.466 + ], + "angle": 0, + "content": "Here, \\(A_{h,l}\\) represents the attention matrix value for the \\(h\\)-th attention head in the \\(l\\)-th layer, \\(x\\) denotes the input, and \\(\\mathcal{L}(x)\\) is the loss function of the task, e.g., the cross-entropy objective for question-answering tasks. The saliency matrix \\(I_l\\) for the \\(l\\)-th layer is obtained by averaging across all attention heads. The significance of information flow from the \\(j\\)-th token to the \\(i\\)-th token in MLLMs is represented by \\(I_l(i,j)\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.468, + 0.484, + 0.543 + ], + "angle": 0, + "content": "To draw a clearer picture of visual information flow in MLLMs, we introduce two quantitative metrics based on \\( I_{l}(i,j) \\), with a particular focus on the information interaction involving image tokens. The definitions of the two quantitative metrics follow below." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.544, + 0.484, + 0.574 + ], + "angle": 0, + "content": "\\(S_{vv}\\), measuring the importance of information flow among image tokens:" + }, + { + "type": "equation", + "bbox": [ + 0.18, + 0.586, + 0.482, + 0.622 + ], + "angle": 0, + "content": "\\[\nS _ {v v} = \\frac {\\sum_ {(i , j) \\in C _ {v v}} I _ {l} (i , j)}{\\left| C _ {v v} \\right|} \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.18, + 0.625, + 0.395, + 0.642 + ], + "angle": 0, + "content": "\\[\nC _ {v v} = \\{(i, j): i, j \\in \\mathcal {V}, i \\geq j \\}.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.654, + 0.484, + 0.684 + ], + "angle": 0, + "content": "\\(S_{vt}\\), measuring the importance of information flow from image tokens to instruction tokens:" + }, + { + "type": "equation", + "bbox": [ + 0.185, + 0.697, + 0.483, + 0.733 + ], + "angle": 0, + "content": "\\[\nS _ {v t} = \\frac {\\sum_ {(i , j) \\in C _ {v t}} I _ {l} (i , j)}{\\left| C _ {v t} \\right|} \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.185, + 0.735, + 0.39, + 0.752 + ], + "angle": 0, + "content": "\\[\nC _ {v t} = \\{(i, j): i \\in \\mathcal {T}, j \\in \\mathcal {V} \\}.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Here, \\(\\mathcal{V}\\) represents the index set of image tokens, derived from features learned by pre-trained visual encoders, while \\(\\mathcal{T}\\) denotes the index set of instruction tokens, specifying requests or questions related to the images. \\(S_{vv}\\) and \\(S_{vt}\\) are utilized to analyze the mechanisms of visual information processing in MLLMs. We define attention interactions among image tokens as intra-visual information flow and those between image and instruction tokens as visual-textual information flow." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.228 + ], + "angle": 0, + "content": "We conducted experiments with the LLaVA-v1.5-7B model on the MS COCO dataset under the POPE benchmark, sampling 500 examples for evaluation. Fig. 3 underscores the critical role of the visual-textual information flow within the model's middle layers, specifically from the 8-th to the 15-th layer. This observation indicates that in these layers, visual information interacts intensively with textual information via attention mechanisms, which substantially influences the prediction outcomes." + }, + { + "type": "image", + "bbox": [ + 0.539, + 0.245, + 0.88, + 0.412 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.419, + 0.907, + 0.462 + ], + "angle": 0, + "content": "Figure 3. The importance of intra-visual flow and visual-textual flow across various layers. The visual-textual information flow in the middle layers has a significant impact on prediction outcomes." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.477, + 0.851, + 0.492 + ], + "angle": 0, + "content": "4.2. Attention Imbalance Across Modalities" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.499, + 0.906, + 0.575 + ], + "angle": 0, + "content": "Sec. 4.1 reveals that the middle layers facilitate crucial fusion, integrating visual and textual inputs into cross-modal semantic representations that drive final predictions. Accordingly, this section will delve deeper into the attention to visual inputs throughout the modality fusion process." + }, + { + "type": "image", + "bbox": [ + 0.526, + 0.592, + 0.882, + 0.774 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.782, + 0.907, + 0.838 + ], + "angle": 0, + "content": "Figure 4. Attention Distribution of Modal Information Across Model Layers. In the middle layers, the model allocates insufficient attention to visual features while disproportionately focusing on system prompts." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.906, + 0.902 + ], + "angle": 0, + "content": "We define the attention allocation, \\(\\lambda\\), as the aggregate attention score assigned to a specific type of token within a single layer. Accordingly, the attention allocation for sys" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14628" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.122 + ], + "angle": 0, + "content": "tem prompts, visual features, and user instructions in the \\(l\\)-th layer can be computed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.208, + 0.135, + 0.369, + 0.167 + ], + "angle": 0, + "content": "\\[\n\\lambda_ {s y s} ^ {l} = \\sum_ {i \\in \\mathcal {T}} \\sum_ {j \\in \\mathcal {S}} A _ {l} (i, j),\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.208, + 0.171, + 0.483, + 0.203 + ], + "angle": 0, + "content": "\\[\n\\lambda_ {v i s} ^ {l} = \\sum_ {i \\in \\mathcal {T}} \\sum_ {j \\in \\mathcal {V}} A _ {l} (i, j), \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.208, + 0.207, + 0.369, + 0.24 + ], + "angle": 0, + "content": "\\[\n\\lambda_ {i n s} ^ {l} = \\sum_ {i \\in \\mathcal {T}} \\sum_ {j \\in \\mathcal {T}} A _ {l} (i, j).\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.246, + 0.484, + 0.351 + ], + "angle": 0, + "content": "In this context, \\(A_{l}\\) represents the attention matrix averaged across all attention heads, while \\(\\mathcal{S}\\) represents the indices of system tokens. The measures \\(\\lambda_{sys}^{l},\\lambda_{vis}^{l}\\), and \\(\\lambda_{ins}^{l}\\) provide insight into the distribution of attention to different modalities across various layers, aiding in understanding the reasons for the underutilization of visual information during the modality fusion process." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.352, + 0.484, + 0.488 + ], + "angle": 0, + "content": "The experimental setup aligns with that described in Sec. 4.1. Fig. 4 illustrates the allocation of attention to different modalities across the model's layers. In the middle layers, attention to visual features is markedly lower than that given to system prompts and user instructions. This suggests that during the critical process of modality fusion, the model's focus on visual input is insufficient. As a result, visual information is underutilized, leading to an output distribution skewed toward language priors." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.498, + 0.19, + 0.514 + ], + "angle": 0, + "content": "4.3. Insights" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.521, + 0.483, + 0.55 + ], + "angle": 0, + "content": "Based on the experimental results presented in Sec. 4.1 and Sec. 4.2, two significant conclusions can be drawn:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.554, + 0.483, + 0.598 + ], + "angle": 0, + "content": "- The model performs the crucial fusion of visual and textual modalities in the middle layers, creating cross-modal semantic representations that drive the final predictions." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.6, + 0.483, + 0.629 + ], + "angle": 0, + "content": "- During this critical fusion process, the model demonstrates inadequate attention to the visual modality." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.554, + 0.483, + 0.629 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.632, + 0.484, + 0.692 + ], + "angle": 0, + "content": "These findings indicate that models fail to fully utilize visual information, resulting in an excessive dependence on language priors and, subsequently, the occurrence of hallucination phenomena." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.707, + 0.349, + 0.724 + ], + "angle": 0, + "content": "5. Visual Amplification Fusion" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.732, + 0.484, + 0.839 + ], + "angle": 0, + "content": "Building on the insights presented in Sec. 4, we introduce a hallucination mitigation method called Visual Amplification Fusion (VAF). As illustrated in Fig. 5, This approach heightens attention to visual information during modality fusion, effectively reducing the excessive dependency on language priors and ensuring that the generated content is closely grounded to visual inputs." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.849, + 0.317, + 0.863 + ], + "angle": 0, + "content": "5.1. Attention Redistribution" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.901 + ], + "angle": 0, + "content": "As outlined in Sec. 4, the model performs crucial fusion of visual and textual modalities within the middle layers." + }, + { + "type": "image", + "bbox": [ + 0.533, + 0.089, + 0.885, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.244, + 0.907, + 0.302 + ], + "angle": 0, + "content": "Figure 5. Illustration of the Visual Amplification Fusion Method. In the middle layers, we select attention heads highly responsive to visual information, amplifying their focus on visual features while reducing unnecessary attention to system prompts." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.324, + 0.906, + 0.383 + ], + "angle": 0, + "content": "However, the attention allocated to visual modality information during this process remains insufficient. To address this, we adjust the attention weights in these layers to achieve a more balanced focus." + }, + { + "type": "image", + "bbox": [ + 0.534, + 0.393, + 0.887, + 0.556 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.567, + 0.907, + 0.636 + ], + "angle": 0, + "content": "Figure 6. Effect of Enhanced Visual Attention on Hallucination Suppression. Increasing attention to visual features in the fusion process of the model's middle layers successfully reduces hallucinations, enabling the model to correct its grape color prediction from \"green\" to \"red\"." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.654, + 0.906, + 0.699 + ], + "angle": 0, + "content": "Let \\(A_{l,h}\\) denote the attention matrix of the \\(h\\)-th attention head in the \\(l\\)-th layer, and \\(Z_{l,h}\\) represent its corresponding attention score matrix, defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.629, + 0.708, + 0.905, + 0.724 + ], + "angle": 0, + "content": "\\[\nA _ {l, h} = \\operatorname {s o f t m a x} \\left(Z _ {l, h}\\right). \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.732, + 0.907, + 0.838 + ], + "angle": 0, + "content": "Our objective during the modality fusion process is to amplify the model's attention to visual features while curbing an overemphasis on system prompts. This adjustment facilitates improved integration of visual information and reduces over-reliance on language priors. To achieve this, we modify the attention score matrix in the middle layers (i.e., \\(8 < l < 15\\)) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.532, + 0.845, + 0.905, + 0.865 + ], + "angle": 0, + "content": "\\[\n\\hat {Z} _ {l, h} = Z _ {l, h} + \\alpha \\cdot M _ {l, h} ^ {\\text {e n h}} \\circ Z _ {l, h} - \\beta \\cdot M _ {l, h} ^ {\\text {s u p}} \\circ Z _ {l, h}. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Here, \\(\\alpha\\) is the enhancement coefficient \\((\\alpha >0)\\) , where larger values indicate stronger amplification of visual attenuate" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14629" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.483, + 0.166 + ], + "angle": 0, + "content": "tion. The suppression coefficient \\(\\beta\\) (\\(0 < \\beta < 1\\)) determines the extent of attention suppression directed at system prompts. The enhancement and suppression mask matrices, \\(M_{l,h}^{enh}\\) and \\(M_{l,h}^{sup}\\), respectively, are defined to guide the modulation of attention elements:" + }, + { + "type": "equation", + "bbox": [ + 0.182, + 0.173, + 0.483, + 0.198 + ], + "angle": 0, + "content": "\\[\nM _ {l, h} ^ {e n h} (i, j) = \\mathbb {I} (i \\in \\mathcal {T}, j \\in \\mathcal {V}), \\tag {9}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.187, + 0.194, + 0.39, + 0.213 + ], + "angle": 0, + "content": "\\[\nM _ {l, h} ^ {s u p} (i, j) = \\mathbb {I} (i \\in \\mathcal {T}, j \\in \\mathcal {S}).\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.222, + 0.484, + 0.313 + ], + "angle": 0, + "content": "These modifications optimize attention allocation by enhancing the model's focus on visual features during modality fusion and minimizing superfluous attention to system prompts. As illustrated in Fig. 6, preliminary analysis indicates that this approach effectively mitigates hallucination issues by promoting greater attention to visual information." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.32, + 0.353, + 0.336 + ], + "angle": 0, + "content": "5.2. Visual Perception Restriction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.342, + 0.484, + 0.478 + ], + "angle": 0, + "content": "Enhancing visual attention across all attention heads in the middle layers can be overly aggressive and may negatively impact content generation. To address this, we propose a selective enhancement strategy. Specifically, we identify and isolate the attention heads that exhibit higher sensitivity to visual information, which we term visual perception heads. We then restrict the visual attention enhancement to these visual perception heads, ensuring better utilization of visual information while maintaining overall model performance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.479, + 0.484, + 0.584 + ], + "angle": 0, + "content": "In the model, attention heads that allocate more attention to visual features demonstrate heightened sensitivity to visual information. Let \\( A_{l,h} \\) represent the attention matrix of the \\( h \\)-th attention head in the \\( l \\)-th layer of the model, with its corresponding visual attention allocation denoted by \\( \\lambda_{\\mathrm{vis}}^{l,h} \\). In each attention layer, we identify the attention heads whose visual attention allocation fall within the top" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.138 + ], + "angle": 0, + "content": "\\(50\\%\\) and designate them as visual perception heads, subsequently redistributing their attention. The attention matrices of the remaining attention heads are kept unchanged." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.151, + 0.64, + 0.168 + ], + "angle": 0, + "content": "6. Experiment" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.177, + 0.907, + 0.299 + ], + "angle": 0, + "content": "This section demonstrates the effectiveness of the proposed VAF method in mitigating hallucinations. Sec. 6.1 outlines the experimental setup, detailing the evaluation benchmarks and VAF parameter configurations. Sec. 6.2 then presents the experimental results from three perspectives: reduction of hallucinations, coherence of generated content, and inference speed. Finally, Sec. 6.3 further verifies the contribution of each VAF component through ablation studies." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.308, + 0.722, + 0.324 + ], + "angle": 0, + "content": "6.1. Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.33, + 0.906, + 0.376 + ], + "angle": 0, + "content": "In Sec. 6.1.1, we present the selected datasets and evaluation metrics. Sec. 6.1.2 details the chosen MLLM backbone models, and Sec. 6.1.3 outlines the baseline settings." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.384, + 0.774, + 0.398 + ], + "angle": 0, + "content": "6.1.1. Datasets & Evaluation Metrics" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.403, + 0.906, + 0.552 + ], + "angle": 0, + "content": "Polling-based Object Probing Evaluation (POPE). POPE [30] is a novel framework designed to evaluate object hallucinations in MLLMs. Departing from traditional caption-based approaches, POPE frames hallucination detection as a binary task by posing straightforward yes-or-no questions regarding the presence of specific objects in an image (e.g., \"Is there a chair in the image?\"). Performance on POPE is measured across four metrics: Accuracy, Precision, Recall, and F1 score, allowing for a thorough evaluation of hallucinations in MLLMs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.553, + 0.906, + 0.585 + ], + "angle": 0, + "content": "Multimodal Model Evaluation (MME). MME [13] benchmark provides a comprehensive framework for evalu" + }, + { + "type": "table", + "bbox": [ + 0.165, + 0.605, + 0.843, + 0.858 + ], + "angle": 0, + "content": "
CategoryMethodLLaVA-v1.5-7BLLaVA-v1.5-13BQwen-VL-Chat-7B
AccuracyF1-scoreAccuracyF1-scoreAccuracyF1-score
RandomRegular87.8 ↑ 0.087.5 ↑ 0.087.6 ↑ 0.087.4 ↑ 0.088.2 ↑ 0.087.9 ↑ 0.0
VCD88.4 ↑ 0.687.7 ↑ 0.288.9 ↑ 1.387.8 ↑ 0.489.1 ↑ 0.988.4 ↑ 0.5
ICD88.1 ↑ 0.387.6 ↑ 0.188.1 ↑ 0.587.6 ↑ 0.288.9 ↑ 0.788.1 ↑ 0.2
VAF89.6 ↑ 1.889.3 ↑ 1.890.1 ↑ 2.589.9 ↑ 2.590.0 ↑ 1.889.7 ↑ 1.8
PopularRegular82.5 ↑ 0.083.2 ↑ 0.082.7 ↑ 0.084.1 ↑ 0.082.4 ↑ 0.083.1 ↑ 0.0
VCD83.1 ↑ 0.684.1 ↑ 0.983.7 ↑ 1.085.1 ↑ 1.083.0 ↑ 0.684.1 ↑ 1.0
ICD82.1 ↓ 0.482.9 ↓ 0.382.9 ↑ 0.284.3 ↑ 0.283.2 ↑ 0.884.5 ↑ 1.4
VAF84.5 ↑ 2.084.9 ↑ 1.785.2 ↑ 2.586.4 ↑ 2.384.9 ↑ 2.585.1 ↑ 2.0
AdversarialRegular77.6 ↑ 0.079.4 ↑ 0.077.8 ↑ 0.079.5 ↑ 0.077.2 ↑ 0.078.9 ↑ 0.0
VCD78.1 ↑ 0.579.6 ↑ 0.278.2 ↑ 0.479.7 ↑ 0.278.8 ↑ 1.680.1 ↑ 1.2
ICD78.5 ↑ 0.979.9 ↑ 0.579.1 ↑ 1.380.1 ↑ 0.678.1 ↑ 0.979.2 ↑ 0.3
VAF80.1 ↑ 2.581.0 ↑ 1.680.7 ↑ 2.981.7 ↑ 2.280.4 ↑ 3.281.2 ↑ 2.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.868, + 0.906, + 0.898 + ], + "angle": 0, + "content": "Table 2. Performance on POPE. Results are averaged across the MS-COCO, A-OKVQA, and GQA datasets. The VAF method demonstrates superior hallucination suppression across all three MLLMs. The best performance for each setting is highlighted in red." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14630" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.114, + 0.089, + 0.891, + 0.344 + ], + "angle": 0, + "content": "
ModelMethodObject-levelAttribute-levelTotal Score
ExistenceCountPositionColor
LLaVA-v1.5-7BRegular185.00↑0.00146.67↑0.00128.33↑0.00150.00↑0.00610.00↑0.00
VCD185.00↑0.00141.33↓5.34128.33↑0.00153.00↑3.00607.66↓2.34
ICD185.00↑0.00148.33↑1.66126.66↓1.67148.33↓1.67608.32↓1.68
VAF195.00↑10.00158.33↑11.66128.33↑0.00155.00↑5.00636.67↑26.67
LLaVA-v1.5-13BRegular185.00↑0.00155.00↑0.00133.33↑0.00165.00↑0.00638.33↑0.00
VCD185.00↑0.00155.00↑0.00130.00↓3.33168.33↑3.33638.33↑0.00
ICD183.33↓1.67153.33↓1.67131.67↓1.66165.00↑0.00633.33↓5.00
VAF195.00↑10.00160.00↑5.00136.67↑3.34170.00↑5.00661.67↑23.34
Qwen-VL-7BRegular158.33↑0.00150.00↑0.00128.33↑0.00170.00↑0.00606.66↑0.00
VCD158.33↑0.00150.00↑0.00133.33↑5.00175.00↑5.00616.66↑10.00
ICD128.33↓30.00151.67↑1.67128.33↑0.00170.00↑0.00578.33↓28.33
VAF165.00↑6.67155.00↑5.00133.33↑5.00175.00↑5.00628.33↑21.67
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.353, + 0.908, + 0.385 + ], + "angle": 0, + "content": "Table 3. Results on the MME subset. Across three MLLMs, the VAF method achieved the most effective suppression of both object-level and attribute-level hallucinations. The highest scores in each setting are highlighted in red." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.409, + 0.482, + 0.528 + ], + "angle": 0, + "content": "ating MLLMs across both perceptual and cognitive dimensions. It consists of ten perception-oriented tasks and four cognition-oriented tasks, with model performance assessed through accuracy metrics. In addition to the full dataset, we leverage specific subsets, such as object existence and counting to analyze object-level hallucinations, while position and color subsets are employed to examine attributelevel hallucinations." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.529, + 0.483, + 0.636 + ], + "angle": 0, + "content": "Novel Object Captioning at Scale (Nocaps). NoCaps [3] benchmark is designed to evaluate image captioning models on their ability to describe novel objects absent from standard datasets like COCO. Model performance is quantified using the CIDEr score, providing a basis to assess the coherence and accuracy of generated captions in response to images containing unfamiliar objects." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.644, + 0.27, + 0.658 + ], + "angle": 0, + "content": "6.1.2. MLLM Backbones" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.662, + 0.483, + 0.829 + ], + "angle": 0, + "content": "In comparison to the Q-former structure, linear projection demonstrates greater efficiency in aligning visual and textual features. This advantage is evident in MLLMs with linear projection architectures, such as LLaVA and Qwen-VL, which outperform Q-former-based MLLMs like Instruct-BLIP and MiniGPT4. Based on these findings, we selected three linear-projection-based MLLMs, specifically LLaVA-v1.5-7B, LLaVA-v1.5-13B [35], and Qwen-VL-7B [5], to evaluate the effectiveness of our proposed VAF method. Detailed prompt templates for each model across various benchmarks are included in Sec. 10." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.837, + 0.26, + 0.853 + ], + "angle": 0, + "content": "6.1.3. Baseline Settings." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.485, + 0.903 + ], + "angle": 0, + "content": "We primarily compared our approach to the VCD [23] and ICD [47] methods. VCD mitigates hallucinations by contrasting output distributions derived from original and dis" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.409, + 0.907, + 0.501 + ], + "angle": 0, + "content": "torted visual inputs, while ICD reduces hallucinated concepts by comparing distributions generated with standard versus disrupted instructions. To ensure consistency and reproducibility in our comparisons, all methods use greedy search. Unless specified otherwise, our experiments set \\(\\beta = 0.1\\) and \\(\\alpha = 0.15\\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.509, + 0.71, + 0.525 + ], + "angle": 0, + "content": "6.2. Results and Analysis" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.531, + 0.906, + 0.607 + ], + "angle": 0, + "content": "Sec. 6.2.1 examines the effectiveness of various methods in mitigating hallucinations, while Sec. 6.2.2 assesses their impact on the quality of generated content. Sec. 6.2.3 then analyzes the influence of each method on inference speed. Additional experimental results are provided in Sec. 8." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.614, + 0.733, + 0.629 + ], + "angle": 0, + "content": "6.2.1. Hallucination Mitigation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.633, + 0.906, + 0.77 + ], + "angle": 0, + "content": "Tab. 2 presents the experimental results of the VAF method on the POPE benchmark, with results averaged across the MSCOCO [31], A-OKVQA [42], and GQA [20] datasets. Applied to both the LLaVA-v1.5 model family and the Qwen-VL model, the VAF method consistently surpasses the VCD and ICD methods in reducing hallucinations. Tab. 3 further highlights the performance of VAF on the MME benchmark, demonstrating its effectiveness in suppressing both object-level and attribute-level hallucinations." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.777, + 0.791, + 0.79 + ], + "angle": 0, + "content": "6.2.2. Coherence of Generated Content" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.906, + 0.903 + ], + "angle": 0, + "content": "Tab. 4 presents the experimental results for various methods on the Nocaps and ScienceQA datasets. It is evident that VCD and ICD substantially degrade the quality of the generated content. Specifically, on the Nocaps dataset, VCD and ICD reduce CIDEr scores by \\(18\\%\\) and \\(27\\%\\), respectively. This degradation primarily arises from the crude disruption of language priors by contrastive decoding methods," + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "14631" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.102, + 0.09, + 0.48, + 0.276 + ], + "angle": 0, + "content": "
ModelDecodingScienceQANocaps
AccuracyCIDEr
LLaVA-v1.5-7BRegular68.078.7
VCD64.565.7
ICD62.462.3
VAF68.578.8
LLaVA-v1.5-13BRegular71.682.6
VCD70.068.9
ICD69.260.3
VAF71.782.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.287, + 0.483, + 0.318 + ], + "angle": 0, + "content": "Table 4. Results on SQA and Nocaps datasets. The highest and second-highest scores are marked in red and blue, respectively." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.342, + 0.483, + 0.404 + ], + "angle": 0, + "content": "which leads to generated content lacking coherence and accuracy. By contrast, our method demonstrates minimal negative impact on prediction results, maintaining both coherence and accuracy effectively." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.41, + 0.25, + 0.426 + ], + "angle": 0, + "content": "6.2.3. Inference Speed" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.429, + 0.484, + 0.52 + ], + "angle": 0, + "content": "Fig. 7 illustrates the impact of different strategies on model inference speed within the Nocaps dataset. In comparison, the VCD and ICD methods nearly double the inference time due to the need to process contrastive input samples, whereas the VAF method has minimal impact on the inference speed of multimodal large language models." + }, + { + "type": "image", + "bbox": [ + 0.114, + 0.539, + 0.443, + 0.715 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.724, + 0.483, + 0.768 + ], + "angle": 0, + "content": "Figure 7. Comparison of different strategies on inference speed. The VCD and ICD methods reduce inference speed by \\(50\\%\\), whereas the VAF method shows minimal impact." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.789, + 0.245, + 0.805 + ], + "angle": 0, + "content": "6.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Ablation studies on the enhancement coefficient \\(\\alpha\\) were conducted using the COCO-Random dataset within the POPE benchmark to understand its influence on model performance. Fig. 8 demonstrates that when \\(0 < \\alpha < 0.25\\), model hallucinations are effectively suppressed. However, when \\(\\alpha\\) surpasses 0.25, performance starts to degrade. We" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.152 + ], + "angle": 0, + "content": "propose that this reduction in performance may stem from an excessive focus on visual features, disrupting the balanced integration of language information and diminishing overall model effectiveness." + }, + { + "type": "image", + "bbox": [ + 0.539, + 0.164, + 0.88, + 0.343 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.535, + 0.353, + 0.884, + 0.368 + ], + "angle": 0, + "content": "Figure 8. Ablation study of \\(\\alpha\\) on the POPE benchmark." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.372, + 0.907, + 0.493 + ], + "angle": 0, + "content": "We performed ablation studies on the visual perception restriction mechanism, evaluating its impact on the POPE and Nocaps benchmarks. Tab. 5 highlights the effects of restricting attention reallocation to visual perception heads. Increasing attention to visual features alone reduces model hallucinations, while confining this reallocation strategy to visual perception heads minimizes adverse effects on content quality. More ablation studies can be found in Sec. 9." + }, + { + "type": "table", + "bbox": [ + 0.528, + 0.504, + 0.895, + 0.609 + ], + "angle": 0, + "content": "
ModelVisual RestrictionPOPENocaps
LLaVA-7B✓89.878.8
X89.976.4
LLaVA-13B✓90.282.3
X90.081.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.62, + 0.907, + 0.663 + ], + "angle": 0, + "content": "Table 5. Ablation Study of Visual Perception Restriction Mechanism. Restricting attention redistribution to the visual perception heads more effectively preserves the quality of generated content." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.68, + 0.628, + 0.695 + ], + "angle": 0, + "content": "7. conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In this paper, we identify two key drawbacks of using contrastive decoding to mitigate hallucinations in MLLMs: reduced quality of generated content and slower inference speed. To address these challenges, we propose a novel approach, Visual Amplification Fusion, which effectively mitigates hallucinations while preserving both inference speed and content generation quality. By enhancing the attention to visual features during modality fusion, VAF minimizes the over-reliance on language priors, ensuring a high degree of consistency between generated content and visual inputs. Extensive experiments across multiple benchmarks and MLLMs demonstrate that VAF provides a clear advantage in hallucination mitigation." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14632" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.259, + 0.108 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.484, + 0.206 + ], + "angle": 0, + "content": "This work is supported by the National Natural Science Foundation of China under Grant 62176246. This work is also supported by Anhui Province Key Research and Development Plan (202304a05020045), Anhui Province Natural Science Foundation (2208085UD17) and National Natural Science Foundation of China under Grant 62406098." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.218, + 0.188, + 0.234 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.244, + 0.484, + 0.312 + ], + "angle": 0, + "content": "[1] Vedika Agarwal, Rakshith Shetty, and Mario Fritz. Towards causal vqa: Revealing and reducing spurious correlations by invariant and covariant semantic editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9690-9698, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.314, + 0.484, + 0.354 + ], + "angle": 0, + "content": "[2] Aishwarya Agrawal, Dhruv Batra, and Devi Parikh. Analyzing the behavior of visual question answering models. arXiv preprint arXiv:1606.07356, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.357, + 0.484, + 0.424 + ], + "angle": 0, + "content": "[3] Harsh Agrawal, Karan Desai, Yufei Wang, Xinlei Chen, Rishabh Jain, Mark Johnson, Dhruv Batra, Devi Parikh, Stefan Lee, and Peter Anderson. nocaps: novel object captioning at scale. In 2019 IEEE/CVF International Conference on Computer Vision (ICCV). IEEE, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.426, + 0.483, + 0.452 + ], + "angle": 0, + "content": "[4] Jinze Bai, Shuai Bai, and et al. Qwen technical report. arXiv preprint arXiv:2309.16609, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.454, + 0.483, + 0.493 + ], + "angle": 0, + "content": "[5] Jinze Bai, Shuai Bai, and et al. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966, 2023. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.496, + 0.483, + 0.522 + ], + "angle": 0, + "content": "[6] Rohan Bavishi, Erich Elsen, and et al. Introducing our multimodal models, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.525, + 0.483, + 0.593 + ], + "angle": 0, + "content": "[7] Ali Furkan Biten, Lluís Gómez, and Dimosthenis Karatzas. Let there be a clock on the beach: Reducing object hallucination in image captioning. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1381-1390, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.595, + 0.483, + 0.634 + ], + "angle": 0, + "content": "[8] Keqin Chen, Zhao Zhang, and et al. Shikra: Unleashing multimodal llm's referential dialogue magic. arXiv preprint arXiv:2306.15195, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.637, + 0.483, + 0.704 + ], + "angle": 0, + "content": "[9] Long Chen, Oleg Sinavski, Jan Hunermann, Alice Karnsund, Andrew James Willmott, Danny Birch, Daniel Maund, and Jamie Shotton. Driving with llms: Fusing object-level vector modality for explainable autonomous driving. arXiv preprint arXiv:2310.01957, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.706, + 0.483, + 0.759 + ], + "angle": 0, + "content": "[10] Zhe Chen, Weiyun Wang, and et al. How far are we to gpt-4v? closing the gap to commercial multimodal models with open-source suites. arXiv preprint arXiv:2404.16821, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.762, + 0.483, + 0.815 + ], + "angle": 0, + "content": "[11] Wei-Lin Chiang and Zhuohan et al Li. Vicuna: An opensource chatbot impressing gpt-4 with \\(90\\%\\) chatgpt quality. See https://vicuna.lmsys.org (accessed 14 April 2023), 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.818, + 0.483, + 0.858 + ], + "angle": 0, + "content": "[12] Wenliang Dai and Junnan Li et al. Instructlip: Towards general-purpose vision-language models with instruction tuning, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.86, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[13] Chaoyou Fu, Peixian Chen, and et al. Mme: A comprehensive evaluation benchmark for multimodal large language models. arXiv preprint arXiv:2306.13394, 2023. 6" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.244, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.162 + ], + "angle": 0, + "content": "[14] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Bartra, and Devi Parikh. Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6904-6913, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.164, + 0.905, + 0.204 + ], + "angle": 0, + "content": "[15] Anisha Gunjal, Jihan Yin, and Erhan Bas. Detecting and preventing hallucinations in large vision language models. arXiv preprint arXiv:2308.06394, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.207, + 0.905, + 0.288 + ], + "angle": 0, + "content": "[16] Vipul Gupta, Zhuowan Li, Adam Kortylewski, Chenyu Zhang, Yingwei Li, and Alan Yuille. Swapmix: Diagnosing and regularizing the over-reliance on visual context in visual question answering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5078-5088, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.291, + 0.905, + 0.345 + ], + "angle": 0, + "content": "[17] Yudong Han, Liqiang Nie, Jianhua Yin, Jianlong Wu, and Yan Yan. Visual perturbation-aware collaborative learning for overcoming the language prior problem. arXiv preprint arXiv:2207.11850, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.348, + 0.905, + 0.401 + ], + "angle": 0, + "content": "[18] Mingzhe Hu, Shaoyan Pan, Yuheng Li, and Xiaofeng Yang. Advancing medical imaging with language models: A journey from n-grams to chatgpt. arXiv preprint arXiv:2304.04920, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.405, + 0.905, + 0.486 + ], + "angle": 0, + "content": "[19] Qidong Huang, Xiaoyi Dong, Pan Zhang, Bin Wang, Conghui He, Jiaqi Wang, Dahua Lin, Weiming Zhang, and Nenghai Yu. Opera: Alleviating hallucination in multimodal large language models via over-trust penalty and retrospection-allocation. In CVPR, pages 13418-13427, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.489, + 0.905, + 0.53 + ], + "angle": 0, + "content": "[20] Drew A Hudson and Christopher D Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering. In CVPR, pages 6700-6709, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.532, + 0.905, + 0.585 + ], + "angle": 0, + "content": "[21] Fushuo Huo, Wenchao Xu, Zhong Zhang, Haozhao Wang, Zhicheng Chen, and Peilin Zhao. Self-introspective decoding: Alleviating hallucinations for large vision-language models, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.588, + 0.905, + 0.629 + ], + "angle": 0, + "content": "[22] Chaoya Jiang, Haiyang Xu, and et al. Hallucination augmented contrastive learning for multimodal large language model. In CVPR, pages 27036-27046, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.632, + 0.905, + 0.685 + ], + "angle": 0, + "content": "[23] Sicong Leng, Hang Zhang, and et al. Mitigating object hallucinations in large vision-language models through visual contrastive decoding. In CVPR, pages 13872-13882, 2024. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.688, + 0.905, + 0.728 + ], + "angle": 0, + "content": "[24] Bo Li, Yuanhan Zhang, and et al. Mimic-it: Multi-modal in-context instruction tuning. arXiv preprint arXiv:2306.05425, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.731, + 0.905, + 0.758 + ], + "angle": 0, + "content": "[25] Bo Li, Kaichen Zhang, and et al. Llava next: Stronger llms supercharge multimodal capabilities in the wild, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.761, + 0.905, + 0.801 + ], + "angle": 0, + "content": "[26] Chunyuan Li, Cliff Wong, and et al. Llava-med: Training a large language-and-vision assistant for biomedicine in one day. In NeurIPS, pages 28541-28564, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.803, + 0.905, + 0.857 + ], + "angle": 0, + "content": "[27] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. BLIP-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[28] Xiang Lisa Li, Ari Holtzman, and et al. Contrastive decoding: Open-ended text generation as optimization. arXiv preprint arXiv:2210.15097, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14633" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[29] Yifan Li, Yifan Du, Kun Zhou, Jinpeng Wang, Wayne Xin Zhao, and Ji-Rong Wen. Evaluating object hallucination in large vision-language models. arXiv preprint arXiv:2305.10355, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.15, + 0.482, + 0.207 + ], + "angle": 0, + "content": "[30] Yifan Li, Yifan Du, Kun Zhou, Jinpeng Wang, Xin Zhao, and Ji-Rong Wen. Evaluating object hallucination in large vision-language models. In EMNLP, pages 292-305, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.209, + 0.482, + 0.248 + ], + "angle": 0, + "content": "[31] Tsung-Yi Lin, Michael Maire, and et al. Microsoft coco: Common objects in context. In ECCV, pages 740-755, 2014. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.252, + 0.482, + 0.308 + ], + "angle": 0, + "content": "[32] Fuxiao Liu, Kevin Lin, Linjie Li, Jianfeng Wang, Yaser Ya-coob, and Lijuan Wang. Mitigating hallucination in large multi-modal models via robust instruction tuning. arXiv preprint arXiv:2306.14565, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.31, + 0.482, + 0.351 + ], + "angle": 0, + "content": "[33] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, pages 34892-34916, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.354, + 0.482, + 0.409 + ], + "angle": 0, + "content": "[34] Haokun Liu, Yaonan Zhu, Kenji Kato, Izumi Kondo, Tadayoshi Aoyama, and Yasuhisa Hasegawa. Lm-based human-robot collaboration framework for manipulation tasks. arXiv preprint arXiv:2308.14972, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.412, + 0.482, + 0.453 + ], + "angle": 0, + "content": "[35] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In CVPR, pages 26296-26306, 2024. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.456, + 0.482, + 0.511 + ], + "angle": 0, + "content": "[36] Zhi-Song Liu, Robin Courant, and Vicky Kalogeiton. Funnynet-w: Multimodal learning of funny moments in videos in the wild. International Journal of Computer Vision, pages 1-22, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.513, + 0.482, + 0.568 + ], + "angle": 0, + "content": "[37] Holy Lvenia, Wenliang Dai, Samuel Cahyawijaya, Ziwei Ji, and Pascale Fung. Negative object presence evaluation (nope) to measure object hallucination in vision-language models. arXiv preprint arXiv:2310.05338, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.571, + 0.482, + 0.653 + ], + "angle": 0, + "content": "[38] Pan Lu, Swaroop Mishra, Tony Xia, Liang Qiu, Kai-Wei Chang, Song-Chun Zhu, Oyvind Tafjord, Peter Clark, and Ashwin Kalyan. Learn to explain: Multimodal reasoning via thought chains for science question answering. In The 36th Conference on Neural Information Processing Systems (NeurIPS), 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.656, + 0.482, + 0.711 + ], + "angle": 0, + "content": "[39] Jinjie Mai, Jun Chen, Bing Li, Guocheng Qian, Mohamed Elhoseiny, and Bernard Ghanem. Llm as a robotic brain: Unifying egocentric memory and control. arXiv preprint arXiv:2304.09349, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.714, + 0.482, + 0.741 + ], + "angle": 0, + "content": "[40] AI Meta. Introducing meta llama 3: The most capable openly available llm to date. Meta AI, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.744, + 0.482, + 0.813 + ], + "angle": 0, + "content": "[41] Yulei Niu, Kaihua Tang, Hanwang Zhang, Zhiwu Lu, XianSheng Hua, and Ji-Rong Wen. Counterfactual vqa: A cause-effect look at language bias. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12700-12710, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.816, + 0.482, + 0.857 + ], + "angle": 0, + "content": "[42] Dustin Schwenk, Apoorv Khandelwal, and et al. A-okvqa: A benchmark for visual question answering using world knowledge. In ECCV, pages 146–162, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.859, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[43] Rohan Taori, Ishaan Gulrajani, and et al. Stanford alpaca: an instruction-following llama model (2023). URL https://github.com/tatsu-lab/stanford_alpaca, 1(9), 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.092, + 0.906, + 0.134 + ], + "angle": 0, + "content": "[44] Hugo Touvron, Thibaut Lavril, and et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.137, + 0.906, + 0.176 + ], + "angle": 0, + "content": "[45] Hugo Touvron, Louis Martin, and et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.906, + 0.234 + ], + "angle": 0, + "content": "[46] Sheng Wang, Zihao Zhao, Xi Ouyang, Qian Wang, and Dinggang Shen. Chatcad: Interactive computer-aided diagnosis on medical image using large language models. arXiv preprint arXiv:2302.07257, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.236, + 0.906, + 0.288 + ], + "angle": 0, + "content": "[47] Xintong Wang, Jingheng Pan, and et al. Mitigating hallucinations in large vision-language models with instruction contrastive decoding. arXiv preprint arXiv:2403.18715, 2024.2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.291, + 0.906, + 0.347 + ], + "angle": 0, + "content": "[48] Yike Wu, Yu Zhao, Shiwan Zhao, Ying Zhang, Xiaojie Yuan, Guoqing Zhao, and Ning Jiang. Overcoming language priors in visual question answering via distinguishing superficially similar instances. arXiv preprint arXiv:2209.08529, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.349, + 0.906, + 0.389 + ], + "angle": 0, + "content": "[49] Zhenyu Wu, Ziwei Wang, Xiuwei Xu, Jiwen Lu, and Haibin Yan. Embodied task planning with large language models. arXiv preprint arXiv:2307.01848, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.391, + 0.906, + 0.446 + ], + "angle": 0, + "content": "[50] Hong Yan, Lijun Liu, Xupeng Feng, and Qingsong Huang. Overcoming language priors with self-contrastive learning for visual question answering. Multimedia Tools and Applications, 82(11):16343–16358, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.448, + 0.906, + 0.489 + ], + "angle": 0, + "content": "[51] Qinghao Ye, Haiyang Xu, and et al. mplug-owl: Modularization empowers large language models with multimodality. arXiv preprint arXiv:2304.14178, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.491, + 0.906, + 0.532 + ], + "angle": 0, + "content": "[52] Shilong Zhang, Peize Sun, and et al. Gpt4roi: Instruction tuning large language model on region-of-interest. arXiv preprint arXiv:2307.03601, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.534, + 0.906, + 0.603 + ], + "angle": 0, + "content": "[53] Ren Zhibo, Wang Huizhen, Zhu Muhua, Wang Yichao, Xiao Tong, and Zhu Jingbo. Overcoming language priors with counterfactual inference for visual question answering. In Proceedings of the 22nd Chinese National Conference on Computational Linguistics, pages 600-610, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.604, + 0.906, + 0.644 + ], + "angle": 0, + "content": "[54] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. IJCV, 130(9):2337-2348, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.646, + 0.906, + 0.702 + ], + "angle": 0, + "content": "[55] Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.092, + 0.906, + 0.702 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14634" + } + ] +] \ No newline at end of file diff --git a/2025/ClearSight_ Visual Signal Enhancement for Object Hallucination Mitigation in Multimodal Large Language Models/cc596546-7afd-468d-96ec-564a706a95d3_origin.pdf b/2025/ClearSight_ Visual Signal Enhancement for Object Hallucination Mitigation in Multimodal Large Language Models/cc596546-7afd-468d-96ec-564a706a95d3_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..b6479ebd0291ec1f1b2492f32898cc6d593fb354 --- /dev/null +++ b/2025/ClearSight_ Visual Signal Enhancement for Object Hallucination Mitigation in Multimodal Large Language Models/cc596546-7afd-468d-96ec-564a706a95d3_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3e73526cd86baf83b2141e04d50e3455fe4c7bbc7f98237e3d697461d6a5bfa2 +size 567753 diff --git a/2025/ClearSight_ Visual Signal Enhancement for Object Hallucination Mitigation in Multimodal Large Language Models/full.md b/2025/ClearSight_ Visual Signal Enhancement for Object Hallucination Mitigation in Multimodal Large Language Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..8060ea86a8bdb94f24241b5e964af0a408db8e55 --- /dev/null +++ b/2025/ClearSight_ Visual Signal Enhancement for Object Hallucination Mitigation in Multimodal Large Language Models/full.md @@ -0,0 +1,371 @@ +# ClearSight: Visual Signal Enhancement for Object Hallucination Mitigation in Multimodal Large Language Models + +Hao Yin Gunagzong Si Zilei Wang\* + +University of Science and Technology of China + +{yinhnavi, guangzongsi}@mail.ustc.edu.cn, zlwang@ustc.edu.cn + +# Abstract + +Contrastive decoding strategies are widely used to mitigate object hallucinations in multimodal large language models (MLLMs). By reducing over-reliance on language priors, these strategies ensure that generated content remains closely grounded in visual inputs, producing contextually accurate outputs. Since contrastive decoding requires no additional training or external tools, it offers both computational efficiency and versatility, making it highly attractive. However, these methods present two main limitations: (1) bluntly suppressing language priors can compromise coherence and accuracy of generated content, and (2) processing contrastive inputs adds computational load, significantly slowing inference speed. To address these challenges, we propose Visual Amplification Fusion (VAF), a plug-and-play technique that enhances attention to visual signals within the model's middle layers, where modality fusion predominantly occurs. This approach enables more effective capture of visual features, reducing the model's bias toward language modality. Experimental results demonstrate that VAF significantly reduces hallucinations across various MLLMs without affecting inference speed, while maintaining coherence and accuracy in generated outputs. The code is available at https://github.com/ustc-hyin/ClearSight. + +# 1. Introduction + +In recent years, MLLMs [8, 27, 33, 35, 54, 55] have achieved remarkable progress in the intersecting fields of computer vision and natural language processing, and have been widely applied in tasks such as image captioning and visual question answering. However, these models often encounter the issue of "object hallucination" [15, 29, 32, 37] in practical applications, where the generated textual descriptions do not match the actual objects in the image. This problem highlights an over-reliance on unimodal pri + +ors (especially language priors) [17, 48, 50, 53] during inference, posing potential risks in high-precision applications such as medical diagnosis [18, 46] and autonomous driving [9, 34, 39, 49]. + +To address object hallucination [1, 2, 7, 14], several Contrastive Decoding strategies have been introduced in recent years. Among these, the Visual Contrastive Decoding (VCD) method has shown promise in reducing hallucinations by contrasting output distributions from both original and perturbed visual inputs, thus mitigating the model's excessive reliance on language priors [16, 41]. Notably, contrastive decoding methods do not require additional training or external tools, offering both computational efficiency and versatility, which has garnered them significant attention. However, these methods present two main limitations: + +# Limitations of Contrastive Decoding + +- While reducing over-reliance on language priors, these methods may compromise the coherence and accuracy of generated content. +- Contrastive decoding necessitates separate processing of the original and contrastive inputs, which considerably increases inference time. + +To address these shortcomings, we hope to propose a training-free method that can effectively reduces hallucinations without compromising content quality or inference speed. Our saliency analysis of the model's attention maps reveals that biases toward language in generated content do not arise from an overemphasis on language signals but rather from insufficient attention on visual information during modality fusion. Based on this insight, we introduce a novel, plug-and-play technique to mitigate hallucinations: Visual Amplification Fusion (VAF). + +Our analysis indicates that modality fusion in MLLMs primarily occurs within the middle layers. VAF specifically amplifies visual signals at these middle layers, enabling the model to capture more distinctive visual features during fusion, which in turn reduces false descriptions in generated + +text. This technique not only strengthens the model's visual representations but also retains the beneficial influence of language priors, thus preserving content quality. Furthermore, by eliminating the need to process contrastive samples, VAF maintains inference speed. + +Experimental results validate the effectiveness of the VAF method. Across multiple object hallucination benchmarks, VAF demonstrated notable performance gains, with improvements of approximately $3\%$ on POPE and $7\%$ on MME. In terms of coherence and accuracy of generated responses, VCD caused a roughly $19\%$ decrease on NoCaps, while VAF maintained content quality without negative impacts. Additionally, VCD reduced inference speed by $50\%$ , whereas VAF had virtually no effect on inference speed. + +In summary, the main contributions are as follows: + +- We identify the negative impacts of contrastive decoding methods on both the quality of generated content and model inference speed. +- We analyze the modality fusion mechanism in MLLMs, highlighting its insufficient attention to visual information. +- We introduce the VAF method, which effectively mitigates the object hallucination problem while maintaining inference speed, coherence, and accuracy. +- We demonstrate the significant performance improvements of the VAF method across multiple object hallucination benchmarks. + +# 2. Related work + +# 2.1. Multimodal Large Language Models + +The development of MLLMs [26, 36, 51, 52] has advanced from BERT-based decoders to LLM-based architectures [4, 11, 40, 43-45], enabling improved multimodal relationship capture [6, 10, 24, 25]. Models like BLIP-2 [27] and miniGPT-4 [55] incorporate a Q-Former mechanism, which enhances the alignment between visual and textual inputs, allowing for more precise interactions across modalities. InstructBLIP [12] builds on this approach by adding task-specific instructions, which improve the model's understanding of context-sensitive visual semantics. LLaVA [33] and Qwen-VL [5] utilize simpler linear projection techniques that streamline the alignment process, resulting in improved overall performance on vision-language tasks. However, hallucination issues persist across MLLMs, posing a significant challenge that requires further research. + +# 2.2. Contrastive Decoding Strategies + +In recent years, Contrastive Decoding [19, 21, 22, 28] has emerged as a technique to improve generative model accuracy through contrastive judgment, widely employed to address hallucinations in generated content. For instance, Visual Contrastive Decoding (VCD) [23] contrasts output + +distributions derived from original and distorted visual inputs, effectively reducing the over-reliance on statistical bias and unimodal priors, two essential causes of object hallucinations. Similarly, Instruction Contrastive Decoding (ICD) [47] works by comparing distributions derived from standard and disrupted instructions, thereby removing hallucinated concepts from the original distribution. These contrastive methods help ground generated content closely to visual inputs, resulting in contextually accurate outputs. However, despite these advancements, contrastive decoding faces two primary limitations: slower inference speed and reduced coherence in generated content. To overcome these limitations, we propose the VAF method, which achieves effective hallucination reduction while preserving both inference speed and content coherence. + +# 3. Preliminary and Motivation + +In Sec. 3.1, we illustrate the working mechanism of contrastive decoding to mitigate hallucinations, using Visual Contrastive Decoding as an example. In Sec. 3.2, we analysis two main drawbacks of this approach: its potential to disrupt the coherence and accuracy of generated content, and its tendency to slow down model inference. + +# 3.1. Contrastive Decoding + +We consider a MLLM parametrized by $\theta$ . The model takes as input a textual query $x$ and a visual input $v$ , where $v$ provides contextual visual information to assist the model in generating a relevant response $y$ to the textual query. The response $y$ is sampled auto-regressively from the probability distribution conditioned on the query $x$ and the visual context $v$ . Mathematically, this can be formulated as: + +$$ +\begin{array}{l} y _ {t} \sim p _ {\theta} \left(y _ {t} \mid v, x, y _ {< t}\right) \tag {1} \\ \propto \exp \operatorname {l o g i t} _ {\theta} \left(y _ {t} \mid v, x, y _ {< t}\right) \\ \end{array} +$$ + +where $y_{t}$ denotes the token at time step $t$ , and $y_{< t}$ represents the sequence of generated tokens up to the time step $(t - 1)$ . + +To mitigate the issue of object hallucination in MLLMs, contrastive decoding techniques can be applied. Here, we present Visual Contrastive Decoding (VCD) as a representative approach, shown in Fig. 1. Specifically, given a textual query $x$ and a visual input $v$ , the model generates two distinct output distributions: one conditioned on the original $v$ and the other on the distorted visual input $v'$ , which is derived by applying pre-defined distortions (i.e., Gaussian noise mask) to $v$ . Then, a new contrastive probability distribution is computed by exploiting the differences between the two initially obtained distributions. The new contrastive distribution $p_{vcd}$ is formulated as: + +$$ +\begin{array}{l} p _ {v c d} (y \mid v, v ^ {\prime}, x) = \operatorname {s o f t m a x} \left[ \log \mathrm {i t} _ {\theta} (y \mid v, x) + \right. \tag {2} \\ \left. \alpha \cdot \left(\operatorname {l o g i t} _ {\theta} (y \mid v, x) - \operatorname {l o g i t} _ {\theta} (y \mid v ^ {\prime}, x)\right) \right], \\ \end{array} +$$ + +where larger $\alpha$ values indicate a stronger amplification of differences between the two distributions ( $\alpha = 0$ reduces to regular decoding). Essentially, VCD serves as a corrective mechanism, reducing hallucinations by contrasting against a distribution predisposed to favoring them. + +![](images/3530f7ec9c14eee44cc85b156e9cc3c505521d148588eb94b12249782b6dc709.jpg) +Figure 1. Illustration of Visual Contrastive Decoding. The hallucinated object "Teacher" is suppressed by contrasting with an output distribution prone to hallucinations. This method has two main drawbacks: (1) additional processing of distorted visual inputs greatly increases inference time; (2) subtracting the language prior disrupts content coherence. + +# 3.2. Limitations of Contrastive Decoding + +As contrastive decoding methods do not require training or external tools, they offer high computational efficiency and generalizability, attracting significant attention in academia. However, these methods still have two major drawbacks: a reduction in the quality of generated content and slower inference speed. + +![](images/cc50a62940a91f14346d7d0da9631496964595dc457f1bbe791655986ea4a3bd.jpg) +Figure 2. Impact of VCD on Model Performance. CIDEr scores are reported on the Nocaps benchmark, while Accuracy is presented for the ScienceQA benchmark. The use of VCD leads to a significant decline in model performance. + +While contrasting logits of $p_{\theta}(y \mid v, x)$ and $p_{\theta}(y \mid v', x)$ can help reduce over-reliance on language priors and mitigate hallucination in MLLMs-as evidenced by a $4\%$ performance gain on the POPE benchmark using the VCD method-merely decreasing the influence of the language + +modality on the output distribution may undermine the coherence of the generated content, potentially leading to prediction errors. This issue is less pronounced in straightforward object hallucination tasks, where responses are limited to binary options, such as "yes" or "no". However, in more complex tasks, including multiple-choice question answering and image caption, the impact of contrastive learning methods on content quality becomes more significant. + +To verify this, we applied VCD method to LLaVA-v1.5-7B and LLaVA-v1.5-13B models, assessing their performance on the ScienceQA [38] and NoCaps benchmarks. As illustrated in Fig. 2, our findings reveal that, following the application of VCD, model performance decreased by $5\%$ on ScienceQA and by a considerable $45\%$ on NoCaps. These results suggest that in tasks requiring nuanced natural language generation, contrastive decoding methods can substantially impair content quality. + +
ModelMethodScienceQANocaps
LLaVA-v1.5-7BRegular0.141s0.456s
VCD0.293s1.086s
LLaVA-v1.5-13BRegular0.222s0.602s
VCD0.459s1.372s
+ +Table 1. Impact of VCD on Model Inference Speed. The table shows the average inference time per sample (in seconds) on the ScienceQA and Nocaps benchmarks. Applying the VCD method nearly doubled the inference time of the model. + +Contrastive decoding methods notably reduce inference speed because they require calculating the output distribution for additional contrastive samples. For instance, in VCD method, each visual input $v$ necessitates computing the logits of both $p_{\theta}(y \mid v, x)$ and $p_{\theta}(y \mid v', x)$ separately. This doubles the computation load during inference compared to vanilla decoding. We evaluated the inference speed of VCD versus vanilla decoding on ScienceQA. The experimental results, shown in Tab. 1, reveal that VCD's inference time is almost double that of vanilla decoding. + +# 4. Visual Neglect in Modal Fusion + +The primary objective of this section is to examine why MLLMs tend to rely excessively on language priors in their predictions. In Sec. 4.1, saliency analysis reveals that image tokens influence prediction outcomes mainly through interactions with instruction tokens within the middle layers. Sec. 4.2 then compares attention weights across different modalities, showing that the attention given to visual features is notably lower than that allocated to system prompts and user instructions. These findings indicate that visual information is often underutilized in the modality fusion process, resulting in an over-reliance on language priors. + +# 4.1. Mid-layer: Visual-Language Fusion + +To uncover why MLLMs tend to overly rely on language priors and overlook visual content in prediction, it is necessary first to clarify how the model utilizes visual modality information. This section explores the influence of the visual modality on prediction outcomes from the perspective of visual information interaction. + +We employ the saliency technique, a widely used interpretability tool, to highlight key token interactions within the attention mechanism. Following established practices, we utilize Taylor expansion to compute saliency scores for each element of the attention matrix: + +$$ +I _ {l} = \left| \sum_ {h} A _ {h, l} \odot \frac {\partial \mathcal {L} (x)}{\partial A _ {h , l}} \right|. \tag {3} +$$ + +Here, $A_{h,l}$ represents the attention matrix value for the $h$ -th attention head in the $l$ -th layer, $x$ denotes the input, and $\mathcal{L}(x)$ is the loss function of the task, e.g., the cross-entropy objective for question-answering tasks. The saliency matrix $I_l$ for the $l$ -th layer is obtained by averaging across all attention heads. The significance of information flow from the $j$ -th token to the $i$ -th token in MLLMs is represented by $I_l(i,j)$ . + +To draw a clearer picture of visual information flow in MLLMs, we introduce two quantitative metrics based on $I_{l}(i,j)$ , with a particular focus on the information interaction involving image tokens. The definitions of the two quantitative metrics follow below. + +$S_{vv}$ , measuring the importance of information flow among image tokens: + +$$ +S _ {v v} = \frac {\sum_ {(i , j) \in C _ {v v}} I _ {l} (i , j)}{\left| C _ {v v} \right|} \tag {4} +$$ + +$$ +C _ {v v} = \{(i, j): i, j \in \mathcal {V}, i \geq j \}. +$$ + +$S_{vt}$ , measuring the importance of information flow from image tokens to instruction tokens: + +$$ +S _ {v t} = \frac {\sum_ {(i , j) \in C _ {v t}} I _ {l} (i , j)}{\left| C _ {v t} \right|} \tag {5} +$$ + +$$ +C _ {v t} = \{(i, j): i \in \mathcal {T}, j \in \mathcal {V} \}. +$$ + +Here, $\mathcal{V}$ represents the index set of image tokens, derived from features learned by pre-trained visual encoders, while $\mathcal{T}$ denotes the index set of instruction tokens, specifying requests or questions related to the images. $S_{vv}$ and $S_{vt}$ are utilized to analyze the mechanisms of visual information processing in MLLMs. We define attention interactions among image tokens as intra-visual information flow and those between image and instruction tokens as visual-textual information flow. + +We conducted experiments with the LLaVA-v1.5-7B model on the MS COCO dataset under the POPE benchmark, sampling 500 examples for evaluation. Fig. 3 underscores the critical role of the visual-textual information flow within the model's middle layers, specifically from the 8-th to the 15-th layer. This observation indicates that in these layers, visual information interacts intensively with textual information via attention mechanisms, which substantially influences the prediction outcomes. + +![](images/d6388fbbfadbe56b6f02102aad69fc5b9143c2371f156258edace2c19395f97c.jpg) +Figure 3. The importance of intra-visual flow and visual-textual flow across various layers. The visual-textual information flow in the middle layers has a significant impact on prediction outcomes. + +# 4.2. Attention Imbalance Across Modalities + +Sec. 4.1 reveals that the middle layers facilitate crucial fusion, integrating visual and textual inputs into cross-modal semantic representations that drive final predictions. Accordingly, this section will delve deeper into the attention to visual inputs throughout the modality fusion process. + +![](images/2d3a106b79a394187f660d99430a78354ea5c9f784ed3b32aedd1ce4c42cd54d.jpg) +Figure 4. Attention Distribution of Modal Information Across Model Layers. In the middle layers, the model allocates insufficient attention to visual features while disproportionately focusing on system prompts. + +We define the attention allocation, $\lambda$ , as the aggregate attention score assigned to a specific type of token within a single layer. Accordingly, the attention allocation for sys + +tem prompts, visual features, and user instructions in the $l$ -th layer can be computed as follows: + +$$ +\lambda_ {s y s} ^ {l} = \sum_ {i \in \mathcal {T}} \sum_ {j \in \mathcal {S}} A _ {l} (i, j), +$$ + +$$ +\lambda_ {v i s} ^ {l} = \sum_ {i \in \mathcal {T}} \sum_ {j \in \mathcal {V}} A _ {l} (i, j), \tag {6} +$$ + +$$ +\lambda_ {i n s} ^ {l} = \sum_ {i \in \mathcal {T}} \sum_ {j \in \mathcal {T}} A _ {l} (i, j). +$$ + +In this context, $A_{l}$ represents the attention matrix averaged across all attention heads, while $\mathcal{S}$ represents the indices of system tokens. The measures $\lambda_{sys}^{l},\lambda_{vis}^{l}$ , and $\lambda_{ins}^{l}$ provide insight into the distribution of attention to different modalities across various layers, aiding in understanding the reasons for the underutilization of visual information during the modality fusion process. + +The experimental setup aligns with that described in Sec. 4.1. Fig. 4 illustrates the allocation of attention to different modalities across the model's layers. In the middle layers, attention to visual features is markedly lower than that given to system prompts and user instructions. This suggests that during the critical process of modality fusion, the model's focus on visual input is insufficient. As a result, visual information is underutilized, leading to an output distribution skewed toward language priors. + +# 4.3. Insights + +Based on the experimental results presented in Sec. 4.1 and Sec. 4.2, two significant conclusions can be drawn: + +- The model performs the crucial fusion of visual and textual modalities in the middle layers, creating cross-modal semantic representations that drive the final predictions. +- During this critical fusion process, the model demonstrates inadequate attention to the visual modality. + +These findings indicate that models fail to fully utilize visual information, resulting in an excessive dependence on language priors and, subsequently, the occurrence of hallucination phenomena. + +# 5. Visual Amplification Fusion + +Building on the insights presented in Sec. 4, we introduce a hallucination mitigation method called Visual Amplification Fusion (VAF). As illustrated in Fig. 5, This approach heightens attention to visual information during modality fusion, effectively reducing the excessive dependency on language priors and ensuring that the generated content is closely grounded to visual inputs. + +# 5.1. Attention Redistribution + +As outlined in Sec. 4, the model performs crucial fusion of visual and textual modalities within the middle layers. + +![](images/4276ac80e92a8e4d964523fdd4d403e279bcfb08f04d187e5d1222c1e38e3c7a.jpg) +Figure 5. Illustration of the Visual Amplification Fusion Method. In the middle layers, we select attention heads highly responsive to visual information, amplifying their focus on visual features while reducing unnecessary attention to system prompts. + +However, the attention allocated to visual modality information during this process remains insufficient. To address this, we adjust the attention weights in these layers to achieve a more balanced focus. + +![](images/f1a5946587d4e3f726094e6779b36975910feb9152052a38d0d7fc97e10b0959.jpg) +Figure 6. Effect of Enhanced Visual Attention on Hallucination Suppression. Increasing attention to visual features in the fusion process of the model's middle layers successfully reduces hallucinations, enabling the model to correct its grape color prediction from "green" to "red". + +Let $A_{l,h}$ denote the attention matrix of the $h$ -th attention head in the $l$ -th layer, and $Z_{l,h}$ represent its corresponding attention score matrix, defined as: + +$$ +A _ {l, h} = \operatorname {s o f t m a x} \left(Z _ {l, h}\right). \tag {7} +$$ + +Our objective during the modality fusion process is to amplify the model's attention to visual features while curbing an overemphasis on system prompts. This adjustment facilitates improved integration of visual information and reduces over-reliance on language priors. To achieve this, we modify the attention score matrix in the middle layers (i.e., $8 < l < 15$ ) as follows: + +$$ +\hat {Z} _ {l, h} = Z _ {l, h} + \alpha \cdot M _ {l, h} ^ {\text {e n h}} \circ Z _ {l, h} - \beta \cdot M _ {l, h} ^ {\text {s u p}} \circ Z _ {l, h}. \tag {8} +$$ + +Here, $\alpha$ is the enhancement coefficient $(\alpha >0)$ , where larger values indicate stronger amplification of visual attenuate + +tion. The suppression coefficient $\beta$ ( $0 < \beta < 1$ ) determines the extent of attention suppression directed at system prompts. The enhancement and suppression mask matrices, $M_{l,h}^{enh}$ and $M_{l,h}^{sup}$ , respectively, are defined to guide the modulation of attention elements: + +$$ +M _ {l, h} ^ {e n h} (i, j) = \mathbb {I} (i \in \mathcal {T}, j \in \mathcal {V}), \tag {9} +$$ + +$$ +M _ {l, h} ^ {s u p} (i, j) = \mathbb {I} (i \in \mathcal {T}, j \in \mathcal {S}). +$$ + +These modifications optimize attention allocation by enhancing the model's focus on visual features during modality fusion and minimizing superfluous attention to system prompts. As illustrated in Fig. 6, preliminary analysis indicates that this approach effectively mitigates hallucination issues by promoting greater attention to visual information. + +# 5.2. Visual Perception Restriction + +Enhancing visual attention across all attention heads in the middle layers can be overly aggressive and may negatively impact content generation. To address this, we propose a selective enhancement strategy. Specifically, we identify and isolate the attention heads that exhibit higher sensitivity to visual information, which we term visual perception heads. We then restrict the visual attention enhancement to these visual perception heads, ensuring better utilization of visual information while maintaining overall model performance. + +In the model, attention heads that allocate more attention to visual features demonstrate heightened sensitivity to visual information. Let $A_{l,h}$ represent the attention matrix of the $h$ -th attention head in the $l$ -th layer of the model, with its corresponding visual attention allocation denoted by $\lambda_{\mathrm{vis}}^{l,h}$ . In each attention layer, we identify the attention heads whose visual attention allocation fall within the top + +$50\%$ and designate them as visual perception heads, subsequently redistributing their attention. The attention matrices of the remaining attention heads are kept unchanged. + +# 6. Experiment + +This section demonstrates the effectiveness of the proposed VAF method in mitigating hallucinations. Sec. 6.1 outlines the experimental setup, detailing the evaluation benchmarks and VAF parameter configurations. Sec. 6.2 then presents the experimental results from three perspectives: reduction of hallucinations, coherence of generated content, and inference speed. Finally, Sec. 6.3 further verifies the contribution of each VAF component through ablation studies. + +# 6.1. Experimental Settings + +In Sec. 6.1.1, we present the selected datasets and evaluation metrics. Sec. 6.1.2 details the chosen MLLM backbone models, and Sec. 6.1.3 outlines the baseline settings. + +# 6.1.1. Datasets & Evaluation Metrics + +Polling-based Object Probing Evaluation (POPE). POPE [30] is a novel framework designed to evaluate object hallucinations in MLLMs. Departing from traditional caption-based approaches, POPE frames hallucination detection as a binary task by posing straightforward yes-or-no questions regarding the presence of specific objects in an image (e.g., "Is there a chair in the image?"). Performance on POPE is measured across four metrics: Accuracy, Precision, Recall, and F1 score, allowing for a thorough evaluation of hallucinations in MLLMs. + +Multimodal Model Evaluation (MME). MME [13] benchmark provides a comprehensive framework for evalu + +
CategoryMethodLLaVA-v1.5-7BLLaVA-v1.5-13BQwen-VL-Chat-7B
AccuracyF1-scoreAccuracyF1-scoreAccuracyF1-score
RandomRegular87.8 ↑ 0.087.5 ↑ 0.087.6 ↑ 0.087.4 ↑ 0.088.2 ↑ 0.087.9 ↑ 0.0
VCD88.4 ↑ 0.687.7 ↑ 0.288.9 ↑ 1.387.8 ↑ 0.489.1 ↑ 0.988.4 ↑ 0.5
ICD88.1 ↑ 0.387.6 ↑ 0.188.1 ↑ 0.587.6 ↑ 0.288.9 ↑ 0.788.1 ↑ 0.2
VAF89.6 ↑ 1.889.3 ↑ 1.890.1 ↑ 2.589.9 ↑ 2.590.0 ↑ 1.889.7 ↑ 1.8
PopularRegular82.5 ↑ 0.083.2 ↑ 0.082.7 ↑ 0.084.1 ↑ 0.082.4 ↑ 0.083.1 ↑ 0.0
VCD83.1 ↑ 0.684.1 ↑ 0.983.7 ↑ 1.085.1 ↑ 1.083.0 ↑ 0.684.1 ↑ 1.0
ICD82.1 ↓ 0.482.9 ↓ 0.382.9 ↑ 0.284.3 ↑ 0.283.2 ↑ 0.884.5 ↑ 1.4
VAF84.5 ↑ 2.084.9 ↑ 1.785.2 ↑ 2.586.4 ↑ 2.384.9 ↑ 2.585.1 ↑ 2.0
AdversarialRegular77.6 ↑ 0.079.4 ↑ 0.077.8 ↑ 0.079.5 ↑ 0.077.2 ↑ 0.078.9 ↑ 0.0
VCD78.1 ↑ 0.579.6 ↑ 0.278.2 ↑ 0.479.7 ↑ 0.278.8 ↑ 1.680.1 ↑ 1.2
ICD78.5 ↑ 0.979.9 ↑ 0.579.1 ↑ 1.380.1 ↑ 0.678.1 ↑ 0.979.2 ↑ 0.3
VAF80.1 ↑ 2.581.0 ↑ 1.680.7 ↑ 2.981.7 ↑ 2.280.4 ↑ 3.281.2 ↑ 2.3
+ +Table 2. Performance on POPE. Results are averaged across the MS-COCO, A-OKVQA, and GQA datasets. The VAF method demonstrates superior hallucination suppression across all three MLLMs. The best performance for each setting is highlighted in red. + +
ModelMethodObject-levelAttribute-levelTotal Score
ExistenceCountPositionColor
LLaVA-v1.5-7BRegular185.00↑0.00146.67↑0.00128.33↑0.00150.00↑0.00610.00↑0.00
VCD185.00↑0.00141.33↓5.34128.33↑0.00153.00↑3.00607.66↓2.34
ICD185.00↑0.00148.33↑1.66126.66↓1.67148.33↓1.67608.32↓1.68
VAF195.00↑10.00158.33↑11.66128.33↑0.00155.00↑5.00636.67↑26.67
LLaVA-v1.5-13BRegular185.00↑0.00155.00↑0.00133.33↑0.00165.00↑0.00638.33↑0.00
VCD185.00↑0.00155.00↑0.00130.00↓3.33168.33↑3.33638.33↑0.00
ICD183.33↓1.67153.33↓1.67131.67↓1.66165.00↑0.00633.33↓5.00
VAF195.00↑10.00160.00↑5.00136.67↑3.34170.00↑5.00661.67↑23.34
Qwen-VL-7BRegular158.33↑0.00150.00↑0.00128.33↑0.00170.00↑0.00606.66↑0.00
VCD158.33↑0.00150.00↑0.00133.33↑5.00175.00↑5.00616.66↑10.00
ICD128.33↓30.00151.67↑1.67128.33↑0.00170.00↑0.00578.33↓28.33
VAF165.00↑6.67155.00↑5.00133.33↑5.00175.00↑5.00628.33↑21.67
+ +Table 3. Results on the MME subset. Across three MLLMs, the VAF method achieved the most effective suppression of both object-level and attribute-level hallucinations. The highest scores in each setting are highlighted in red. + +ating MLLMs across both perceptual and cognitive dimensions. It consists of ten perception-oriented tasks and four cognition-oriented tasks, with model performance assessed through accuracy metrics. In addition to the full dataset, we leverage specific subsets, such as object existence and counting to analyze object-level hallucinations, while position and color subsets are employed to examine attributelevel hallucinations. + +Novel Object Captioning at Scale (Nocaps). NoCaps [3] benchmark is designed to evaluate image captioning models on their ability to describe novel objects absent from standard datasets like COCO. Model performance is quantified using the CIDEr score, providing a basis to assess the coherence and accuracy of generated captions in response to images containing unfamiliar objects. + +# 6.1.2. MLLM Backbones + +In comparison to the Q-former structure, linear projection demonstrates greater efficiency in aligning visual and textual features. This advantage is evident in MLLMs with linear projection architectures, such as LLaVA and Qwen-VL, which outperform Q-former-based MLLMs like Instruct-BLIP and MiniGPT4. Based on these findings, we selected three linear-projection-based MLLMs, specifically LLaVA-v1.5-7B, LLaVA-v1.5-13B [35], and Qwen-VL-7B [5], to evaluate the effectiveness of our proposed VAF method. Detailed prompt templates for each model across various benchmarks are included in Sec. 10. + +# 6.1.3. Baseline Settings. + +We primarily compared our approach to the VCD [23] and ICD [47] methods. VCD mitigates hallucinations by contrasting output distributions derived from original and dis + +torted visual inputs, while ICD reduces hallucinated concepts by comparing distributions generated with standard versus disrupted instructions. To ensure consistency and reproducibility in our comparisons, all methods use greedy search. Unless specified otherwise, our experiments set $\beta = 0.1$ and $\alpha = 0.15$ . + +# 6.2. Results and Analysis + +Sec. 6.2.1 examines the effectiveness of various methods in mitigating hallucinations, while Sec. 6.2.2 assesses their impact on the quality of generated content. Sec. 6.2.3 then analyzes the influence of each method on inference speed. Additional experimental results are provided in Sec. 8. + +# 6.2.1. Hallucination Mitigation + +Tab. 2 presents the experimental results of the VAF method on the POPE benchmark, with results averaged across the MSCOCO [31], A-OKVQA [42], and GQA [20] datasets. Applied to both the LLaVA-v1.5 model family and the Qwen-VL model, the VAF method consistently surpasses the VCD and ICD methods in reducing hallucinations. Tab. 3 further highlights the performance of VAF on the MME benchmark, demonstrating its effectiveness in suppressing both object-level and attribute-level hallucinations. + +# 6.2.2. Coherence of Generated Content + +Tab. 4 presents the experimental results for various methods on the Nocaps and ScienceQA datasets. It is evident that VCD and ICD substantially degrade the quality of the generated content. Specifically, on the Nocaps dataset, VCD and ICD reduce CIDEr scores by $18\%$ and $27\%$ , respectively. This degradation primarily arises from the crude disruption of language priors by contrastive decoding methods, + +
ModelDecodingScienceQANocaps
AccuracyCIDEr
LLaVA-v1.5-7BRegular68.078.7
VCD64.565.7
ICD62.462.3
VAF68.578.8
LLaVA-v1.5-13BRegular71.682.6
VCD70.068.9
ICD69.260.3
VAF71.782.3
+ +which leads to generated content lacking coherence and accuracy. By contrast, our method demonstrates minimal negative impact on prediction results, maintaining both coherence and accuracy effectively. + +# 6.2.3. Inference Speed + +Fig. 7 illustrates the impact of different strategies on model inference speed within the Nocaps dataset. In comparison, the VCD and ICD methods nearly double the inference time due to the need to process contrastive input samples, whereas the VAF method has minimal impact on the inference speed of multimodal large language models. + +![](images/b27fbe02e4ad6e6fd802441e5658fbec11550e4ce18c5748aac63639a42b5f9d.jpg) +Figure 7. Comparison of different strategies on inference speed. The VCD and ICD methods reduce inference speed by $50\%$ , whereas the VAF method shows minimal impact. + +# 6.3. Ablation Study + +Ablation studies on the enhancement coefficient $\alpha$ were conducted using the COCO-Random dataset within the POPE benchmark to understand its influence on model performance. Fig. 8 demonstrates that when $0 < \alpha < 0.25$ , model hallucinations are effectively suppressed. However, when $\alpha$ surpasses 0.25, performance starts to degrade. We + +propose that this reduction in performance may stem from an excessive focus on visual features, disrupting the balanced integration of language information and diminishing overall model effectiveness. + +![](images/a46d5daaa6e27ef0a8238450f81f2345dcd1ab7d14cb92d2d08010dacc032f81.jpg) +Figure 8. Ablation study of $\alpha$ on the POPE benchmark. + +We performed ablation studies on the visual perception restriction mechanism, evaluating its impact on the POPE and Nocaps benchmarks. Tab. 5 highlights the effects of restricting attention reallocation to visual perception heads. Increasing attention to visual features alone reduces model hallucinations, while confining this reallocation strategy to visual perception heads minimizes adverse effects on content quality. More ablation studies can be found in Sec. 9. + +Table 4. Results on SQA and Nocaps datasets. The highest and second-highest scores are marked in red and blue, respectively. + +
ModelVisual RestrictionPOPENocaps
LLaVA-7B✓89.878.8
X89.976.4
LLaVA-13B✓90.282.3
X90.081.1
+ +Table 5. Ablation Study of Visual Perception Restriction Mechanism. Restricting attention redistribution to the visual perception heads more effectively preserves the quality of generated content. + +# 7. conclusion + +In this paper, we identify two key drawbacks of using contrastive decoding to mitigate hallucinations in MLLMs: reduced quality of generated content and slower inference speed. To address these challenges, we propose a novel approach, Visual Amplification Fusion, which effectively mitigates hallucinations while preserving both inference speed and content generation quality. By enhancing the attention to visual features during modality fusion, VAF minimizes the over-reliance on language priors, ensuring a high degree of consistency between generated content and visual inputs. Extensive experiments across multiple benchmarks and MLLMs demonstrate that VAF provides a clear advantage in hallucination mitigation. + +# Acknowledgements + +This work is supported by the National Natural Science Foundation of China under Grant 62176246. This work is also supported by Anhui Province Key Research and Development Plan (202304a05020045), Anhui Province Natural Science Foundation (2208085UD17) and National Natural Science Foundation of China under Grant 62406098. + +# References + +[1] Vedika Agarwal, Rakshith Shetty, and Mario Fritz. Towards causal vqa: Revealing and reducing spurious correlations by invariant and covariant semantic editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9690-9698, 2020. 1 +[2] Aishwarya Agrawal, Dhruv Batra, and Devi Parikh. Analyzing the behavior of visual question answering models. arXiv preprint arXiv:1606.07356, 2016. 1 +[3] Harsh Agrawal, Karan Desai, Yufei Wang, Xinlei Chen, Rishabh Jain, Mark Johnson, Dhruv Batra, Devi Parikh, Stefan Lee, and Peter Anderson. nocaps: novel object captioning at scale. In 2019 IEEE/CVF International Conference on Computer Vision (ICCV). IEEE, 2019. 7 +[4] Jinze Bai, Shuai Bai, and et al. Qwen technical report. arXiv preprint arXiv:2309.16609, 2023. 2 +[5] Jinze Bai, Shuai Bai, and et al. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966, 2023. 2, 7 +[6] Rohan Bavishi, Erich Elsen, and et al. Introducing our multimodal models, 2023. 2 +[7] Ali Furkan Biten, Lluís Gómez, and Dimosthenis Karatzas. Let there be a clock on the beach: Reducing object hallucination in image captioning. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1381-1390, 2022. 1 +[8] Keqin Chen, Zhao Zhang, and et al. Shikra: Unleashing multimodal llm's referential dialogue magic. arXiv preprint arXiv:2306.15195, 2023. 1 +[9] Long Chen, Oleg Sinavski, Jan Hunermann, Alice Karnsund, Andrew James Willmott, Danny Birch, Daniel Maund, and Jamie Shotton. Driving with llms: Fusing object-level vector modality for explainable autonomous driving. arXiv preprint arXiv:2310.01957, 2023. 1 +[10] Zhe Chen, Weiyun Wang, and et al. How far are we to gpt-4v? closing the gap to commercial multimodal models with open-source suites. arXiv preprint arXiv:2404.16821, 2024. 2 +[11] Wei-Lin Chiang and Zhuohan et al Li. Vicuna: An opensource chatbot impressing gpt-4 with $90\%$ chatgpt quality. See https://vicuna.lmsys.org (accessed 14 April 2023), 2023. 2 +[12] Wenliang Dai and Junnan Li et al. Instructlip: Towards general-purpose vision-language models with instruction tuning, 2023. 2 +[13] Chaoyou Fu, Peixian Chen, and et al. Mme: A comprehensive evaluation benchmark for multimodal large language models. arXiv preprint arXiv:2306.13394, 2023. 6 + +[14] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Bartra, and Devi Parikh. Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6904-6913, 2017. 1 +[15] Anisha Gunjal, Jihan Yin, and Erhan Bas. Detecting and preventing hallucinations in large vision language models. arXiv preprint arXiv:2308.06394, 2023. 1 +[16] Vipul Gupta, Zhuowan Li, Adam Kortylewski, Chenyu Zhang, Yingwei Li, and Alan Yuille. Swapmix: Diagnosing and regularizing the over-reliance on visual context in visual question answering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5078-5088, 2022. 1 +[17] Yudong Han, Liqiang Nie, Jianhua Yin, Jianlong Wu, and Yan Yan. Visual perturbation-aware collaborative learning for overcoming the language prior problem. arXiv preprint arXiv:2207.11850, 2022. 1 +[18] Mingzhe Hu, Shaoyan Pan, Yuheng Li, and Xiaofeng Yang. Advancing medical imaging with language models: A journey from n-grams to chatgpt. arXiv preprint arXiv:2304.04920, 2023. 1 +[19] Qidong Huang, Xiaoyi Dong, Pan Zhang, Bin Wang, Conghui He, Jiaqi Wang, Dahua Lin, Weiming Zhang, and Nenghai Yu. Opera: Alleviating hallucination in multimodal large language models via over-trust penalty and retrospection-allocation. In CVPR, pages 13418-13427, 2024. 2 +[20] Drew A Hudson and Christopher D Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering. In CVPR, pages 6700-6709, 2019. 7 +[21] Fushuo Huo, Wenchao Xu, Zhong Zhang, Haozhao Wang, Zhicheng Chen, and Peilin Zhao. Self-introspective decoding: Alleviating hallucinations for large vision-language models, 2024. 2 +[22] Chaoya Jiang, Haiyang Xu, and et al. Hallucination augmented contrastive learning for multimodal large language model. In CVPR, pages 27036-27046, 2024. 2 +[23] Sicong Leng, Hang Zhang, and et al. Mitigating object hallucinations in large vision-language models through visual contrastive decoding. In CVPR, pages 13872-13882, 2024. 2, 7 +[24] Bo Li, Yuanhan Zhang, and et al. Mimic-it: Multi-modal in-context instruction tuning. arXiv preprint arXiv:2306.05425, 2023. 2 +[25] Bo Li, Kaichen Zhang, and et al. Llava next: Stronger llms supercharge multimodal capabilities in the wild, 2024. 2 +[26] Chunyuan Li, Cliff Wong, and et al. Llava-med: Training a large language-and-vision assistant for biomedicine in one day. In NeurIPS, pages 28541-28564, 2023. 2 +[27] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. BLIP-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML, 2023. 1, 2 +[28] Xiang Lisa Li, Ari Holtzman, and et al. Contrastive decoding: Open-ended text generation as optimization. arXiv preprint arXiv:2210.15097, 2022. 2 + +[29] Yifan Li, Yifan Du, Kun Zhou, Jinpeng Wang, Wayne Xin Zhao, and Ji-Rong Wen. Evaluating object hallucination in large vision-language models. arXiv preprint arXiv:2305.10355, 2023. 1 +[30] Yifan Li, Yifan Du, Kun Zhou, Jinpeng Wang, Xin Zhao, and Ji-Rong Wen. Evaluating object hallucination in large vision-language models. In EMNLP, pages 292-305, 2023. 6 +[31] Tsung-Yi Lin, Michael Maire, and et al. Microsoft coco: Common objects in context. In ECCV, pages 740-755, 2014. 7 +[32] Fuxiao Liu, Kevin Lin, Linjie Li, Jianfeng Wang, Yaser Ya-coob, and Lijuan Wang. Mitigating hallucination in large multi-modal models via robust instruction tuning. arXiv preprint arXiv:2306.14565, 2023. 1 +[33] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, pages 34892-34916, 2023. 1, 2 +[34] Haokun Liu, Yaonan Zhu, Kenji Kato, Izumi Kondo, Tadayoshi Aoyama, and Yasuhisa Hasegawa. Lm-based human-robot collaboration framework for manipulation tasks. arXiv preprint arXiv:2308.14972, 2023. 1 +[35] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In CVPR, pages 26296-26306, 2024. 1, 7 +[36] Zhi-Song Liu, Robin Courant, and Vicky Kalogeiton. Funnynet-w: Multimodal learning of funny moments in videos in the wild. International Journal of Computer Vision, pages 1-22, 2024. 2 +[37] Holy Lvenia, Wenliang Dai, Samuel Cahyawijaya, Ziwei Ji, and Pascale Fung. Negative object presence evaluation (nope) to measure object hallucination in vision-language models. arXiv preprint arXiv:2310.05338, 2023. 1 +[38] Pan Lu, Swaroop Mishra, Tony Xia, Liang Qiu, Kai-Wei Chang, Song-Chun Zhu, Oyvind Tafjord, Peter Clark, and Ashwin Kalyan. Learn to explain: Multimodal reasoning via thought chains for science question answering. In The 36th Conference on Neural Information Processing Systems (NeurIPS), 2022. 3 +[39] Jinjie Mai, Jun Chen, Bing Li, Guocheng Qian, Mohamed Elhoseiny, and Bernard Ghanem. Llm as a robotic brain: Unifying egocentric memory and control. arXiv preprint arXiv:2304.09349, 2023. 1 +[40] AI Meta. Introducing meta llama 3: The most capable openly available llm to date. Meta AI, 2024. 2 +[41] Yulei Niu, Kaihua Tang, Hanwang Zhang, Zhiwu Lu, XianSheng Hua, and Ji-Rong Wen. Counterfactual vqa: A cause-effect look at language bias. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12700-12710, 2021. 1 +[42] Dustin Schwenk, Apoorv Khandelwal, and et al. A-okvqa: A benchmark for visual question answering using world knowledge. In ECCV, pages 146–162, 2022. 7 +[43] Rohan Taori, Ishaan Gulrajani, and et al. Stanford alpaca: an instruction-following llama model (2023). URL https://github.com/tatsu-lab/stanford_alpaca, 1(9), 2023. 2 + +[44] Hugo Touvron, Thibaut Lavril, and et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. +[45] Hugo Touvron, Louis Martin, and et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 2 +[46] Sheng Wang, Zihao Zhao, Xi Ouyang, Qian Wang, and Dinggang Shen. Chatcad: Interactive computer-aided diagnosis on medical image using large language models. arXiv preprint arXiv:2302.07257, 2023. 1 +[47] Xintong Wang, Jingheng Pan, and et al. Mitigating hallucinations in large vision-language models with instruction contrastive decoding. arXiv preprint arXiv:2403.18715, 2024.2, 7 +[48] Yike Wu, Yu Zhao, Shiwan Zhao, Ying Zhang, Xiaojie Yuan, Guoqing Zhao, and Ning Jiang. Overcoming language priors in visual question answering via distinguishing superficially similar instances. arXiv preprint arXiv:2209.08529, 2022. 1 +[49] Zhenyu Wu, Ziwei Wang, Xiuwei Xu, Jiwen Lu, and Haibin Yan. Embodied task planning with large language models. arXiv preprint arXiv:2307.01848, 2023. 1 +[50] Hong Yan, Lijun Liu, Xupeng Feng, and Qingsong Huang. Overcoming language priors with self-contrastive learning for visual question answering. Multimedia Tools and Applications, 82(11):16343–16358, 2023. 1 +[51] Qinghao Ye, Haiyang Xu, and et al. mplug-owl: Modularization empowers large language models with multimodality. arXiv preprint arXiv:2304.14178, 2023. 2 +[52] Shilong Zhang, Peize Sun, and et al. Gpt4roi: Instruction tuning large language model on region-of-interest. arXiv preprint arXiv:2307.03601, 2023. 2 +[53] Ren Zhibo, Wang Huizhen, Zhu Muhua, Wang Yichao, Xiao Tong, and Zhu Jingbo. Overcoming language priors with counterfactual inference for visual question answering. In Proceedings of the 22nd Chinese National Conference on Computational Linguistics, pages 600-610, 2023. 1 +[54] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. IJCV, 130(9):2337-2348, 2022. 1 +[55] Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 1, 2 \ No newline at end of file diff --git a/2025/ClearSight_ Visual Signal Enhancement for Object Hallucination Mitigation in Multimodal Large Language Models/images.zip b/2025/ClearSight_ Visual Signal Enhancement for Object Hallucination Mitigation in Multimodal Large Language Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..84b363072958da5dad7c8a745f72999adab64bc8 --- /dev/null +++ b/2025/ClearSight_ Visual Signal Enhancement for Object Hallucination Mitigation in Multimodal Large Language Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dd095dd8cefd020afa7ec418e34c1f1f889fe77a6e76d4ee7de6c63d4dd63956 +size 590777 diff --git a/2025/ClearSight_ Visual Signal Enhancement for Object Hallucination Mitigation in Multimodal Large Language Models/layout.json b/2025/ClearSight_ Visual Signal Enhancement for Object Hallucination Mitigation in Multimodal Large Language Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..30e44175d14dc0891e229a132ccf8432bfcb22f0 --- /dev/null +++ b/2025/ClearSight_ Visual Signal Enhancement for Object Hallucination Mitigation in Multimodal Large Language Models/layout.json @@ -0,0 +1,9140 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 63, + 102, + 547, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 102, + 547, + 140 + ], + "spans": [ + { + "bbox": [ + 63, + 102, + 547, + 140 + ], + "type": "text", + "content": "ClearSight: Visual Signal Enhancement for Object Hallucination Mitigation in Multimodal Large Language Models" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 190, + 161, + 421, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 190, + 161, + 421, + 175 + ], + "spans": [ + { + "bbox": [ + 190, + 161, + 421, + 175 + ], + "type": "text", + "content": "Hao Yin Gunagzong Si Zilei Wang\\*" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 190, + 176, + 419, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 190, + 176, + 419, + 190 + ], + "spans": [ + { + "bbox": [ + 190, + 176, + 419, + 190 + ], + "type": "text", + "content": "University of Science and Technology of China" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 141, + 191, + 465, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 191, + 465, + 203 + ], + "spans": [ + { + "bbox": [ + 141, + 191, + 465, + 203 + ], + "type": "text", + "content": "{yinhnavi, guangzongsi}@mail.ustc.edu.cn, zlwang@ustc.edu.cn" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "spans": [ + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 256, + 297, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 256, + 297, + 544 + ], + "spans": [ + { + "bbox": [ + 55, + 256, + 297, + 544 + ], + "type": "text", + "content": "Contrastive decoding strategies are widely used to mitigate object hallucinations in multimodal large language models (MLLMs). By reducing over-reliance on language priors, these strategies ensure that generated content remains closely grounded in visual inputs, producing contextually accurate outputs. Since contrastive decoding requires no additional training or external tools, it offers both computational efficiency and versatility, making it highly attractive. However, these methods present two main limitations: (1) bluntly suppressing language priors can compromise coherence and accuracy of generated content, and (2) processing contrastive inputs adds computational load, significantly slowing inference speed. To address these challenges, we propose Visual Amplification Fusion (VAF), a plug-and-play technique that enhances attention to visual signals within the model's middle layers, where modality fusion predominantly occurs. This approach enables more effective capture of visual features, reducing the model's bias toward language modality. Experimental results demonstrate that VAF significantly reduces hallucinations across various MLLMs without affecting inference speed, while maintaining coherence and accuracy in generated outputs. The code is available at https://github.com/ustc-hyin/ClearSight." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 567, + 135, + 578 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 567, + 135, + 578 + ], + "spans": [ + { + "bbox": [ + 56, + 567, + 135, + 578 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 586, + 296, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 586, + 296, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 586, + 296, + 696 + ], + "type": "text", + "content": "In recent years, MLLMs [8, 27, 33, 35, 54, 55] have achieved remarkable progress in the intersecting fields of computer vision and natural language processing, and have been widely applied in tasks such as image captioning and visual question answering. However, these models often encounter the issue of \"object hallucination\" [15, 29, 32, 37] in practical applications, where the generated textual descriptions do not match the actual objects in the image. This problem highlights an over-reliance on unimodal pri" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 232, + 553, + 280 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 232, + 553, + 280 + ], + "spans": [ + { + "bbox": [ + 313, + 232, + 553, + 280 + ], + "type": "text", + "content": "ors (especially language priors) [17, 48, 50, 53] during inference, posing potential risks in high-precision applications such as medical diagnosis [18, 46] and autonomous driving [9, 34, 39, 49]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 281, + 555, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 281, + 555, + 413 + ], + "spans": [ + { + "bbox": [ + 313, + 281, + 555, + 413 + ], + "type": "text", + "content": "To address object hallucination [1, 2, 7, 14], several Contrastive Decoding strategies have been introduced in recent years. Among these, the Visual Contrastive Decoding (VCD) method has shown promise in reducing hallucinations by contrasting output distributions from both original and perturbed visual inputs, thus mitigating the model's excessive reliance on language priors [16, 41]. Notably, contrastive decoding methods do not require additional training or external tools, offering both computational efficiency and versatility, which has garnered them significant attention. However, these methods present two main limitations:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 330, + 423, + 479, + 436 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 330, + 423, + 479, + 436 + ], + "spans": [ + { + "bbox": [ + 330, + 423, + 479, + 436 + ], + "type": "text", + "content": "Limitations of Contrastive Decoding" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 330, + 445, + 539, + 517 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 330, + 445, + 538, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 330, + 445, + 538, + 480 + ], + "spans": [ + { + "bbox": [ + 330, + 445, + 538, + 480 + ], + "type": "text", + "content": "- While reducing over-reliance on language priors, these methods may compromise the coherence and accuracy of generated content." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 330, + 481, + 539, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 330, + 481, + 539, + 517 + ], + "spans": [ + { + "bbox": [ + 330, + 481, + 539, + 517 + ], + "type": "text", + "content": "- Contrastive decoding necessitates separate processing of the original and contrastive inputs, which considerably increases inference time." + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 533, + 555, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 555, + 652 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 555, + 652 + ], + "type": "text", + "content": "To address these shortcomings, we hope to propose a training-free method that can effectively reduces hallucinations without compromising content quality or inference speed. Our saliency analysis of the model's attention maps reveals that biases toward language in generated content do not arise from an overemphasis on language signals but rather from insufficient attention on visual information during modality fusion. Based on this insight, we introduce a novel, plug-and-play technique to mitigate hallucinations: Visual Amplification Fusion (VAF)." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 653, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 556, + 715 + ], + "type": "text", + "content": "Our analysis indicates that modality fusion in MLLMs primarily occurs within the middle layers. VAF specifically amplifies visual signals at these middle layers, enabling the model to capture more distinctive visual features during fusion, which in turn reduces false descriptions in generated" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 703, + 146, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 146, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 146, + 712 + ], + "type": "text", + "content": "*Corresponding Author" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14625" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 133 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 133 + ], + "type": "text", + "content": "text. This technique not only strengthens the model's visual representations but also retains the beneficial influence of language priors, thus preserving content quality. Furthermore, by eliminating the need to process contrastive samples, VAF maintains inference speed." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 133, + 295, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 133, + 295, + 240 + ], + "spans": [ + { + "bbox": [ + 55, + 133, + 295, + 240 + ], + "type": "text", + "content": "Experimental results validate the effectiveness of the VAF method. Across multiple object hallucination benchmarks, VAF demonstrated notable performance gains, with improvements of approximately " + }, + { + "bbox": [ + 55, + 133, + 295, + 240 + ], + "type": "inline_equation", + "content": "3\\%" + }, + { + "bbox": [ + 55, + 133, + 295, + 240 + ], + "type": "text", + "content": " on POPE and " + }, + { + "bbox": [ + 55, + 133, + 295, + 240 + ], + "type": "inline_equation", + "content": "7\\%" + }, + { + "bbox": [ + 55, + 133, + 295, + 240 + ], + "type": "text", + "content": " on MME. In terms of coherence and accuracy of generated responses, VCD caused a roughly " + }, + { + "bbox": [ + 55, + 133, + 295, + 240 + ], + "type": "inline_equation", + "content": "19\\%" + }, + { + "bbox": [ + 55, + 133, + 295, + 240 + ], + "type": "text", + "content": " decrease on NoCaps, while VAF maintained content quality without negative impacts. Additionally, VCD reduced inference speed by " + }, + { + "bbox": [ + 55, + 133, + 295, + 240 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 55, + 133, + 295, + 240 + ], + "type": "text", + "content": ", whereas VAF had virtually no effect on inference speed." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 241, + 272, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 241, + 272, + 252 + ], + "spans": [ + { + "bbox": [ + 67, + 241, + 272, + 252 + ], + "type": "text", + "content": "In summary, the main contributions are as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 253, + 295, + 396 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 55, + 253, + 295, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 253, + 295, + 289 + ], + "spans": [ + { + "bbox": [ + 55, + 253, + 295, + 289 + ], + "type": "text", + "content": "- We identify the negative impacts of contrastive decoding methods on both the quality of generated content and model inference speed." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 289, + 295, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 289, + 295, + 323 + ], + "spans": [ + { + "bbox": [ + 56, + 289, + 295, + 323 + ], + "type": "text", + "content": "- We analyze the modality fusion mechanism in MLLMs, highlighting its insufficient attention to visual information." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 325, + 295, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 325, + 295, + 361 + ], + "spans": [ + { + "bbox": [ + 56, + 325, + 295, + 361 + ], + "type": "text", + "content": "- We introduce the VAF method, which effectively mitigates the object hallucination problem while maintaining inference speed, coherence, and accuracy." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 361, + 295, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 361, + 295, + 396 + ], + "spans": [ + { + "bbox": [ + 56, + 361, + 295, + 396 + ], + "type": "text", + "content": "- We demonstrate the significant performance improvements of the VAF method across multiple object hallucination benchmarks." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 409, + 139, + 422 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 409, + 139, + 422 + ], + "spans": [ + { + "bbox": [ + 55, + 409, + 139, + 422 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 429, + 250, + 442 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 429, + 250, + 442 + ], + "spans": [ + { + "bbox": [ + 55, + 429, + 250, + 442 + ], + "type": "text", + "content": "2.1. Multimodal Large Language Models" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 447, + 295, + 627 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 447, + 295, + 627 + ], + "spans": [ + { + "bbox": [ + 55, + 447, + 295, + 627 + ], + "type": "text", + "content": "The development of MLLMs [26, 36, 51, 52] has advanced from BERT-based decoders to LLM-based architectures [4, 11, 40, 43-45], enabling improved multimodal relationship capture [6, 10, 24, 25]. Models like BLIP-2 [27] and miniGPT-4 [55] incorporate a Q-Former mechanism, which enhances the alignment between visual and textual inputs, allowing for more precise interactions across modalities. InstructBLIP [12] builds on this approach by adding task-specific instructions, which improve the model's understanding of context-sensitive visual semantics. LLaVA [33] and Qwen-VL [5] utilize simpler linear projection techniques that streamline the alignment process, resulting in improved overall performance on vision-language tasks. However, hallucination issues persist across MLLMs, posing a significant challenge that requires further research." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 635, + 230, + 649 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 635, + 230, + 649 + ], + "spans": [ + { + "bbox": [ + 55, + 635, + 230, + 649 + ], + "type": "text", + "content": "2.2. Contrastive Decoding Strategies" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 653, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 295, + 714 + ], + "type": "text", + "content": "In recent years, Contrastive Decoding [19, 21, 22, 28] has emerged as a technique to improve generative model accuracy through contrastive judgment, widely employed to address hallucinations in generated content. For instance, Visual Contrastive Decoding (VCD) [23] contrasts output" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 72, + 555, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 252 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 252 + ], + "type": "text", + "content": "distributions derived from original and distorted visual inputs, effectively reducing the over-reliance on statistical bias and unimodal priors, two essential causes of object hallucinations. Similarly, Instruction Contrastive Decoding (ICD) [47] works by comparing distributions derived from standard and disrupted instructions, thereby removing hallucinated concepts from the original distribution. These contrastive methods help ground generated content closely to visual inputs, resulting in contextually accurate outputs. However, despite these advancements, contrastive decoding faces two primary limitations: slower inference speed and reduced coherence in generated content. To overcome these limitations, we propose the VAF method, which achieves effective hallucination reduction while preserving both inference speed and content coherence." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 261, + 473, + 274 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 261, + 473, + 274 + ], + "spans": [ + { + "bbox": [ + 313, + 261, + 473, + 274 + ], + "type": "text", + "content": "3. Preliminary and Motivation" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 281, + 555, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 281, + 555, + 354 + ], + "spans": [ + { + "bbox": [ + 313, + 281, + 555, + 354 + ], + "type": "text", + "content": "In Sec. 3.1, we illustrate the working mechanism of contrastive decoding to mitigate hallucinations, using Visual Contrastive Decoding as an example. In Sec. 3.2, we analysis two main drawbacks of this approach: its potential to disrupt the coherence and accuracy of generated content, and its tendency to slow down model inference." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 359, + 440, + 372 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 359, + 440, + 372 + ], + "spans": [ + { + "bbox": [ + 313, + 359, + 440, + 372 + ], + "type": "text", + "content": "3.1. Contrastive Decoding" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 376, + 554, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 376, + 554, + 460 + ], + "spans": [ + { + "bbox": [ + 313, + 376, + 554, + 460 + ], + "type": "text", + "content": "We consider a MLLM parametrized by " + }, + { + "bbox": [ + 313, + 376, + 554, + 460 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 376, + 554, + 460 + ], + "type": "text", + "content": ". The model takes as input a textual query " + }, + { + "bbox": [ + 313, + 376, + 554, + 460 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 376, + 554, + 460 + ], + "type": "text", + "content": " and a visual input " + }, + { + "bbox": [ + 313, + 376, + 554, + 460 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 313, + 376, + 554, + 460 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 376, + 554, + 460 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 313, + 376, + 554, + 460 + ], + "type": "text", + "content": " provides contextual visual information to assist the model in generating a relevant response " + }, + { + "bbox": [ + 313, + 376, + 554, + 460 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 313, + 376, + 554, + 460 + ], + "type": "text", + "content": " to the textual query. The response " + }, + { + "bbox": [ + 313, + 376, + 554, + 460 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 313, + 376, + 554, + 460 + ], + "type": "text", + "content": " is sampled auto-regressively from the probability distribution conditioned on the query " + }, + { + "bbox": [ + 313, + 376, + 554, + 460 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 376, + 554, + 460 + ], + "type": "text", + "content": " and the visual context " + }, + { + "bbox": [ + 313, + 376, + 554, + 460 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 313, + 376, + 554, + 460 + ], + "type": "text", + "content": ". Mathematically, this can be formulated as:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 370, + 467, + 553, + 495 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 467, + 553, + 495 + ], + "spans": [ + { + "bbox": [ + 370, + 467, + 553, + 495 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} y _ {t} \\sim p _ {\\theta} \\left(y _ {t} \\mid v, x, y _ {< t}\\right) \\tag {1} \\\\ \\propto \\exp \\operatorname {l o g i t} _ {\\theta} \\left(y _ {t} \\mid v, x, y _ {< t}\\right) \\\\ \\end{array}", + "image_path": "25eede69b08a8a91a6732b35eae10b3950af8dcef1ec8856d269a84fe7ffa677.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 501, + 554, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 501, + 554, + 525 + ], + "spans": [ + { + "bbox": [ + 313, + 501, + 554, + 525 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 501, + 554, + 525 + ], + "type": "inline_equation", + "content": "y_{t}" + }, + { + "bbox": [ + 313, + 501, + 554, + 525 + ], + "type": "text", + "content": " denotes the token at time step " + }, + { + "bbox": [ + 313, + 501, + 554, + 525 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 501, + 554, + 525 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 501, + 554, + 525 + ], + "type": "inline_equation", + "content": "y_{< t}" + }, + { + "bbox": [ + 313, + 501, + 554, + 525 + ], + "type": "text", + "content": " represents the sequence of generated tokens up to the time step " + }, + { + "bbox": [ + 313, + 501, + 554, + 525 + ], + "type": "inline_equation", + "content": "(t - 1)" + }, + { + "bbox": [ + 313, + 501, + 554, + 525 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 525, + 555, + 669 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 525, + 555, + 669 + ], + "spans": [ + { + "bbox": [ + 313, + 525, + 555, + 669 + ], + "type": "text", + "content": "To mitigate the issue of object hallucination in MLLMs, contrastive decoding techniques can be applied. Here, we present Visual Contrastive Decoding (VCD) as a representative approach, shown in Fig. 1. Specifically, given a textual query " + }, + { + "bbox": [ + 313, + 525, + 555, + 669 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 525, + 555, + 669 + ], + "type": "text", + "content": " and a visual input " + }, + { + "bbox": [ + 313, + 525, + 555, + 669 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 313, + 525, + 555, + 669 + ], + "type": "text", + "content": ", the model generates two distinct output distributions: one conditioned on the original " + }, + { + "bbox": [ + 313, + 525, + 555, + 669 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 313, + 525, + 555, + 669 + ], + "type": "text", + "content": " and the other on the distorted visual input " + }, + { + "bbox": [ + 313, + 525, + 555, + 669 + ], + "type": "inline_equation", + "content": "v'" + }, + { + "bbox": [ + 313, + 525, + 555, + 669 + ], + "type": "text", + "content": ", which is derived by applying pre-defined distortions (i.e., Gaussian noise mask) to " + }, + { + "bbox": [ + 313, + 525, + 555, + 669 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 313, + 525, + 555, + 669 + ], + "type": "text", + "content": ". Then, a new contrastive probability distribution is computed by exploiting the differences between the two initially obtained distributions. The new contrastive distribution " + }, + { + "bbox": [ + 313, + 525, + 555, + 669 + ], + "type": "inline_equation", + "content": "p_{vcd}" + }, + { + "bbox": [ + 313, + 525, + 555, + 669 + ], + "type": "text", + "content": " is formulated as:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 328, + 673, + 553, + 716 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 673, + 553, + 716 + ], + "spans": [ + { + "bbox": [ + 328, + 673, + 553, + 716 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} p _ {v c d} (y \\mid v, v ^ {\\prime}, x) = \\operatorname {s o f t m a x} \\left[ \\log \\mathrm {i t} _ {\\theta} (y \\mid v, x) + \\right. \\tag {2} \\\\ \\left. \\alpha \\cdot \\left(\\operatorname {l o g i t} _ {\\theta} (y \\mid v, x) - \\operatorname {l o g i t} _ {\\theta} (y \\mid v ^ {\\prime}, x)\\right) \\right], \\\\ \\end{array}", + "image_path": "e57cc2d95b38e69bc5aefacf9e577b3dcc1dcf604d91c0ede13bf4e9168dc1a6.jpg" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "14626" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 133 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 133 + ], + "type": "text", + "content": "where larger " + }, + { + "bbox": [ + 55, + 72, + 295, + 133 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 72, + 295, + 133 + ], + "type": "text", + "content": " values indicate a stronger amplification of differences between the two distributions (" + }, + { + "bbox": [ + 55, + 72, + 295, + 133 + ], + "type": "inline_equation", + "content": "\\alpha = 0" + }, + { + "bbox": [ + 55, + 72, + 295, + 133 + ], + "type": "text", + "content": " reduces to regular decoding). Essentially, VCD serves as a corrective mechanism, reducing hallucinations by contrasting against a distribution predisposed to favoring them." + } + ] + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 59, + 142, + 293, + 256 + ], + "blocks": [ + { + "bbox": [ + 59, + 142, + 293, + 256 + ], + "lines": [ + { + "bbox": [ + 59, + 142, + 293, + 256 + ], + "spans": [ + { + "bbox": [ + 59, + 142, + 293, + 256 + ], + "type": "image", + "image_path": "3530f7ec9c14eee44cc85b156e9cc3c505521d148588eb94b12249782b6dc709.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 264, + 296, + 331 + ], + "lines": [ + { + "bbox": [ + 55, + 264, + 296, + 331 + ], + "spans": [ + { + "bbox": [ + 55, + 264, + 296, + 331 + ], + "type": "text", + "content": "Figure 1. Illustration of Visual Contrastive Decoding. The hallucinated object \"Teacher\" is suppressed by contrasting with an output distribution prone to hallucinations. This method has two main drawbacks: (1) additional processing of distorted visual inputs greatly increases inference time; (2) subtracting the language prior disrupts content coherence." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 351, + 248, + 364 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 351, + 248, + 364 + ], + "spans": [ + { + "bbox": [ + 55, + 351, + 248, + 364 + ], + "type": "text", + "content": "3.2. Limitations of Contrastive Decoding" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 369, + 296, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 369, + 296, + 441 + ], + "spans": [ + { + "bbox": [ + 55, + 369, + 296, + 441 + ], + "type": "text", + "content": "As contrastive decoding methods do not require training or external tools, they offer high computational efficiency and generalizability, attracting significant attention in academia. However, these methods still have two major drawbacks: a reduction in the quality of generated content and slower inference speed." + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 74, + 453, + 272, + 587 + ], + "blocks": [ + { + "bbox": [ + 74, + 453, + 272, + 587 + ], + "lines": [ + { + "bbox": [ + 74, + 453, + 272, + 587 + ], + "spans": [ + { + "bbox": [ + 74, + 453, + 272, + 587 + ], + "type": "image", + "image_path": "cc50a62940a91f14346d7d0da9631496964595dc457f1bbe791655986ea4a3bd.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 596, + 295, + 640 + ], + "lines": [ + { + "bbox": [ + 55, + 596, + 295, + 640 + ], + "spans": [ + { + "bbox": [ + 55, + 596, + 295, + 640 + ], + "type": "text", + "content": "Figure 2. Impact of VCD on Model Performance. CIDEr scores are reported on the Nocaps benchmark, while Accuracy is presented for the ScienceQA benchmark. The use of VCD leads to a significant decline in model performance." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "content": "While contrasting logits of " + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "inline_equation", + "content": "p_{\\theta}(y \\mid v, x)" + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "inline_equation", + "content": "p_{\\theta}(y \\mid v', x)" + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "content": " can help reduce over-reliance on language priors and mitigate hallucination in MLLMs-as evidenced by a " + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "inline_equation", + "content": "4\\%" + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "content": " performance gain on the POPE benchmark using the VCD method-merely decreasing the influence of the language" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "text", + "content": "modality on the output distribution may undermine the coherence of the generated content, potentially leading to prediction errors. This issue is less pronounced in straightforward object hallucination tasks, where responses are limited to binary options, such as \"yes\" or \"no\". However, in more complex tasks, including multiple-choice question answering and image caption, the impact of contrastive learning methods on content quality becomes more significant." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 167, + 555, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 167, + 555, + 276 + ], + "spans": [ + { + "bbox": [ + 313, + 167, + 555, + 276 + ], + "type": "text", + "content": "To verify this, we applied VCD method to LLaVA-v1.5-7B and LLaVA-v1.5-13B models, assessing their performance on the ScienceQA [38] and NoCaps benchmarks. As illustrated in Fig. 2, our findings reveal that, following the application of VCD, model performance decreased by " + }, + { + "bbox": [ + 313, + 167, + 555, + 276 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 313, + 167, + 555, + 276 + ], + "type": "text", + "content": " on ScienceQA and by a considerable " + }, + { + "bbox": [ + 313, + 167, + 555, + 276 + ], + "type": "inline_equation", + "content": "45\\%" + }, + { + "bbox": [ + 313, + 167, + 555, + 276 + ], + "type": "text", + "content": " on NoCaps. These results suggest that in tasks requiring nuanced natural language generation, contrastive decoding methods can substantially impair content quality." + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 324, + 284, + 545, + 369 + ], + "blocks": [ + { + "bbox": [ + 324, + 284, + 545, + 369 + ], + "lines": [ + { + "bbox": [ + 324, + 284, + 545, + 369 + ], + "spans": [ + { + "bbox": [ + 324, + 284, + 545, + 369 + ], + "type": "table", + "html": "
ModelMethodScienceQANocaps
LLaVA-v1.5-7BRegular0.141s0.456s
VCD0.293s1.086s
LLaVA-v1.5-13BRegular0.222s0.602s
VCD0.459s1.372s
", + "image_path": "caeeaa2dc2b3706b4a3a110348583e3a34db6fcf764cee126d37bb969e0a74e7.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 376, + 555, + 421 + ], + "lines": [ + { + "bbox": [ + 313, + 376, + 555, + 421 + ], + "spans": [ + { + "bbox": [ + 313, + 376, + 555, + 421 + ], + "type": "text", + "content": "Table 1. Impact of VCD on Model Inference Speed. The table shows the average inference time per sample (in seconds) on the ScienceQA and Nocaps benchmarks. Applying the VCD method nearly doubled the inference time of the model." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 433, + 554, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 433, + 554, + 552 + ], + "spans": [ + { + "bbox": [ + 313, + 433, + 554, + 552 + ], + "type": "text", + "content": "Contrastive decoding methods notably reduce inference speed because they require calculating the output distribution for additional contrastive samples. For instance, in VCD method, each visual input " + }, + { + "bbox": [ + 313, + 433, + 554, + 552 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 313, + 433, + 554, + 552 + ], + "type": "text", + "content": " necessitates computing the logits of both " + }, + { + "bbox": [ + 313, + 433, + 554, + 552 + ], + "type": "inline_equation", + "content": "p_{\\theta}(y \\mid v, x)" + }, + { + "bbox": [ + 313, + 433, + 554, + 552 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 433, + 554, + 552 + ], + "type": "inline_equation", + "content": "p_{\\theta}(y \\mid v', x)" + }, + { + "bbox": [ + 313, + 433, + 554, + 552 + ], + "type": "text", + "content": " separately. This doubles the computation load during inference compared to vanilla decoding. We evaluated the inference speed of VCD versus vanilla decoding on ScienceQA. The experimental results, shown in Tab. 1, reveal that VCD's inference time is almost double that of vanilla decoding." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 562, + 491, + 575 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 562, + 491, + 575 + ], + "spans": [ + { + "bbox": [ + 313, + 562, + 491, + 575 + ], + "type": "text", + "content": "4. Visual Neglect in Modal Fusion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 582, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 555, + 714 + ], + "type": "text", + "content": "The primary objective of this section is to examine why MLLMs tend to rely excessively on language priors in their predictions. In Sec. 4.1, saliency analysis reveals that image tokens influence prediction outcomes mainly through interactions with instruction tokens within the middle layers. Sec. 4.2 then compares attention weights across different modalities, showing that the attention given to visual features is notably lower than that allocated to system prompts and user instructions. These findings indicate that visual information is often underutilized in the modality fusion process, resulting in an over-reliance on language priors." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14627" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 247, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 247, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 247, + 85 + ], + "type": "text", + "content": "4.1. Mid-layer: Visual-Language Fusion" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 90, + 295, + 161 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 90, + 295, + 161 + ], + "spans": [ + { + "bbox": [ + 55, + 90, + 295, + 161 + ], + "type": "text", + "content": "To uncover why MLLMs tend to overly rely on language priors and overlook visual content in prediction, it is necessary first to clarify how the model utilizes visual modality information. This section explores the influence of the visual modality on prediction outcomes from the perspective of visual information interaction." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 163, + 296, + 222 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 163, + 296, + 222 + ], + "spans": [ + { + "bbox": [ + 55, + 163, + 296, + 222 + ], + "type": "text", + "content": "We employ the saliency technique, a widely used interpretability tool, to highlight key token interactions within the attention mechanism. Following established practices, we utilize Taylor expansion to compute saliency scores for each element of the attention matrix:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 120, + 231, + 295, + 264 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 120, + 231, + 295, + 264 + ], + "spans": [ + { + "bbox": [ + 120, + 231, + 295, + 264 + ], + "type": "interline_equation", + "content": "I _ {l} = \\left| \\sum_ {h} A _ {h, l} \\odot \\frac {\\partial \\mathcal {L} (x)}{\\partial A _ {h , l}} \\right|. \\tag {3}", + "image_path": "2bb102268be2df4c174992da1d09e14552735be094d9188059a86662ef382efd.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 274, + 296, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 274, + 296, + 369 + ], + "spans": [ + { + "bbox": [ + 55, + 274, + 296, + 369 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 55, + 274, + 296, + 369 + ], + "type": "inline_equation", + "content": "A_{h,l}" + }, + { + "bbox": [ + 55, + 274, + 296, + 369 + ], + "type": "text", + "content": " represents the attention matrix value for the " + }, + { + "bbox": [ + 55, + 274, + 296, + 369 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 55, + 274, + 296, + 369 + ], + "type": "text", + "content": "-th attention head in the " + }, + { + "bbox": [ + 55, + 274, + 296, + 369 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 55, + 274, + 296, + 369 + ], + "type": "text", + "content": "-th layer, " + }, + { + "bbox": [ + 55, + 274, + 296, + 369 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 274, + 296, + 369 + ], + "type": "text", + "content": " denotes the input, and " + }, + { + "bbox": [ + 55, + 274, + 296, + 369 + ], + "type": "inline_equation", + "content": "\\mathcal{L}(x)" + }, + { + "bbox": [ + 55, + 274, + 296, + 369 + ], + "type": "text", + "content": " is the loss function of the task, e.g., the cross-entropy objective for question-answering tasks. The saliency matrix " + }, + { + "bbox": [ + 55, + 274, + 296, + 369 + ], + "type": "inline_equation", + "content": "I_l" + }, + { + "bbox": [ + 55, + 274, + 296, + 369 + ], + "type": "text", + "content": " for the " + }, + { + "bbox": [ + 55, + 274, + 296, + 369 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 55, + 274, + 296, + 369 + ], + "type": "text", + "content": "-th layer is obtained by averaging across all attention heads. The significance of information flow from the " + }, + { + "bbox": [ + 55, + 274, + 296, + 369 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 55, + 274, + 296, + 369 + ], + "type": "text", + "content": "-th token to the " + }, + { + "bbox": [ + 55, + 274, + 296, + 369 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 274, + 296, + 369 + ], + "type": "text", + "content": "-th token in MLLMs is represented by " + }, + { + "bbox": [ + 55, + 274, + 296, + 369 + ], + "type": "inline_equation", + "content": "I_l(i,j)" + }, + { + "bbox": [ + 55, + 274, + 296, + 369 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 370, + 296, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 370, + 296, + 430 + ], + "spans": [ + { + "bbox": [ + 55, + 370, + 296, + 430 + ], + "type": "text", + "content": "To draw a clearer picture of visual information flow in MLLMs, we introduce two quantitative metrics based on " + }, + { + "bbox": [ + 55, + 370, + 296, + 430 + ], + "type": "inline_equation", + "content": "I_{l}(i,j)" + }, + { + "bbox": [ + 55, + 370, + 296, + 430 + ], + "type": "text", + "content": ", with a particular focus on the information interaction involving image tokens. The definitions of the two quantitative metrics follow below." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 430, + 296, + 454 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 430, + 296, + 454 + ], + "spans": [ + { + "bbox": [ + 55, + 430, + 296, + 454 + ], + "type": "inline_equation", + "content": "S_{vv}" + }, + { + "bbox": [ + 55, + 430, + 296, + 454 + ], + "type": "text", + "content": ", measuring the importance of information flow among image tokens:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 110, + 464, + 294, + 492 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 464, + 294, + 492 + ], + "spans": [ + { + "bbox": [ + 110, + 464, + 294, + 492 + ], + "type": "interline_equation", + "content": "S _ {v v} = \\frac {\\sum_ {(i , j) \\in C _ {v v}} I _ {l} (i , j)}{\\left| C _ {v v} \\right|} \\tag {4}", + "image_path": "32bd17678a581b4a3c88a2d6f8b7231f9af2abfa5ab269e916152ffaf304d8bb.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 110, + 495, + 241, + 508 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 495, + 241, + 508 + ], + "spans": [ + { + "bbox": [ + 110, + 495, + 241, + 508 + ], + "type": "interline_equation", + "content": "C _ {v v} = \\{(i, j): i, j \\in \\mathcal {V}, i \\geq j \\}.", + "image_path": "44392e3b7155058c47798887ef9e404d89768d1ff8817d6634cb6f58be0505b2.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 517, + 296, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 517, + 296, + 541 + ], + "spans": [ + { + "bbox": [ + 55, + 517, + 296, + 541 + ], + "type": "inline_equation", + "content": "S_{vt}" + }, + { + "bbox": [ + 55, + 517, + 296, + 541 + ], + "type": "text", + "content": ", measuring the importance of information flow from image tokens to instruction tokens:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 113, + 552, + 295, + 580 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 552, + 295, + 580 + ], + "spans": [ + { + "bbox": [ + 113, + 552, + 295, + 580 + ], + "type": "interline_equation", + "content": "S _ {v t} = \\frac {\\sum_ {(i , j) \\in C _ {v t}} I _ {l} (i , j)}{\\left| C _ {v t} \\right|} \\tag {5}", + "image_path": "9d25a79d9ead44b061168924ce01d77c02b5b77c59b2868a337a384bd336a0b6.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 113, + 582, + 238, + 595 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 582, + 238, + 595 + ], + "spans": [ + { + "bbox": [ + 113, + 582, + 238, + 595 + ], + "type": "interline_equation", + "content": "C _ {v t} = \\{(i, j): i \\in \\mathcal {T}, j \\in \\mathcal {V} \\}.", + "image_path": "6828a3415653a88657fb6ab10dd1e048c92ebccc70d3ba2bf081658e4a993472.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "text", + "content": " represents the index set of image tokens, derived from features learned by pre-trained visual encoders, while " + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "text", + "content": " denotes the index set of instruction tokens, specifying requests or questions related to the images. " + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "inline_equation", + "content": "S_{vv}" + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "inline_equation", + "content": "S_{vt}" + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "text", + "content": " are utilized to analyze the mechanisms of visual information processing in MLLMs. We define attention interactions among image tokens as intra-visual information flow and those between image and instruction tokens as visual-textual information flow." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "text", + "content": "We conducted experiments with the LLaVA-v1.5-7B model on the MS COCO dataset under the POPE benchmark, sampling 500 examples for evaluation. Fig. 3 underscores the critical role of the visual-textual information flow within the model's middle layers, specifically from the 8-th to the 15-th layer. This observation indicates that in these layers, visual information interacts intensively with textual information via attention mechanisms, which substantially influences the prediction outcomes." + } + ] + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 329, + 194, + 538, + 326 + ], + "blocks": [ + { + "bbox": [ + 329, + 194, + 538, + 326 + ], + "lines": [ + { + "bbox": [ + 329, + 194, + 538, + 326 + ], + "spans": [ + { + "bbox": [ + 329, + 194, + 538, + 326 + ], + "type": "image", + "image_path": "d6388fbbfadbe56b6f02102aad69fc5b9143c2371f156258edace2c19395f97c.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 331, + 555, + 365 + ], + "lines": [ + { + "bbox": [ + 313, + 331, + 555, + 365 + ], + "spans": [ + { + "bbox": [ + 313, + 331, + 555, + 365 + ], + "type": "text", + "content": "Figure 3. The importance of intra-visual flow and visual-textual flow across various layers. The visual-textual information flow in the middle layers has a significant impact on prediction outcomes." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 377, + 520, + 389 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 377, + 520, + 389 + ], + "spans": [ + { + "bbox": [ + 313, + 377, + 520, + 389 + ], + "type": "text", + "content": "4.2. Attention Imbalance Across Modalities" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 395, + 554, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 395, + 554, + 455 + ], + "spans": [ + { + "bbox": [ + 313, + 395, + 554, + 455 + ], + "type": "text", + "content": "Sec. 4.1 reveals that the middle layers facilitate crucial fusion, integrating visual and textual inputs into cross-modal semantic representations that drive final predictions. Accordingly, this section will delve deeper into the attention to visual inputs throughout the modality fusion process." + } + ] + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 321, + 468, + 539, + 613 + ], + "blocks": [ + { + "bbox": [ + 321, + 468, + 539, + 613 + ], + "lines": [ + { + "bbox": [ + 321, + 468, + 539, + 613 + ], + "spans": [ + { + "bbox": [ + 321, + 468, + 539, + 613 + ], + "type": "image", + "image_path": "2d3a106b79a394187f660d99430a78354ea5c9f784ed3b32aedd1ce4c42cd54d.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 619, + 555, + 663 + ], + "lines": [ + { + "bbox": [ + 313, + 619, + 555, + 663 + ], + "spans": [ + { + "bbox": [ + 313, + 619, + 555, + 663 + ], + "type": "text", + "content": "Figure 4. Attention Distribution of Modal Information Across Model Layers. In the middle layers, the model allocates insufficient attention to visual features while disproportionately focusing on system prompts." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "text", + "content": "We define the attention allocation, " + }, + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "text", + "content": ", as the aggregate attention score assigned to a specific type of token within a single layer. Accordingly, the attention allocation for sys" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14628" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "type": "text", + "content": "tem prompts, visual features, and user instructions in the " + }, + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "type": "text", + "content": "-th layer can be computed as follows:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 127, + 106, + 225, + 132 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 106, + 225, + 132 + ], + "spans": [ + { + "bbox": [ + 127, + 106, + 225, + 132 + ], + "type": "interline_equation", + "content": "\\lambda_ {s y s} ^ {l} = \\sum_ {i \\in \\mathcal {T}} \\sum_ {j \\in \\mathcal {S}} A _ {l} (i, j),", + "image_path": "9fa21b0fa02e354e9a460860c23d462f86a77e935df639f47f028f1a490a0729.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 127, + 135, + 295, + 160 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 135, + 295, + 160 + ], + "spans": [ + { + "bbox": [ + 127, + 135, + 295, + 160 + ], + "type": "interline_equation", + "content": "\\lambda_ {v i s} ^ {l} = \\sum_ {i \\in \\mathcal {T}} \\sum_ {j \\in \\mathcal {V}} A _ {l} (i, j), \\tag {6}", + "image_path": "d62dc0111aea8b6abffaf56327bc6c054e12778fefea16bfde173d0e3f0a7173.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 127, + 163, + 225, + 190 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 163, + 225, + 190 + ], + "spans": [ + { + "bbox": [ + 127, + 163, + 225, + 190 + ], + "type": "interline_equation", + "content": "\\lambda_ {i n s} ^ {l} = \\sum_ {i \\in \\mathcal {T}} \\sum_ {j \\in \\mathcal {T}} A _ {l} (i, j).", + "image_path": "36e4cf6a3a6bd6ca40ec2ea6b52e7727b34f7350434c8e9c760bb8b83e1ad115.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 194, + 296, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 194, + 296, + 277 + ], + "spans": [ + { + "bbox": [ + 55, + 194, + 296, + 277 + ], + "type": "text", + "content": "In this context, " + }, + { + "bbox": [ + 55, + 194, + 296, + 277 + ], + "type": "inline_equation", + "content": "A_{l}" + }, + { + "bbox": [ + 55, + 194, + 296, + 277 + ], + "type": "text", + "content": " represents the attention matrix averaged across all attention heads, while " + }, + { + "bbox": [ + 55, + 194, + 296, + 277 + ], + "type": "inline_equation", + "content": "\\mathcal{S}" + }, + { + "bbox": [ + 55, + 194, + 296, + 277 + ], + "type": "text", + "content": " represents the indices of system tokens. The measures " + }, + { + "bbox": [ + 55, + 194, + 296, + 277 + ], + "type": "inline_equation", + "content": "\\lambda_{sys}^{l},\\lambda_{vis}^{l}" + }, + { + "bbox": [ + 55, + 194, + 296, + 277 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 194, + 296, + 277 + ], + "type": "inline_equation", + "content": "\\lambda_{ins}^{l}" + }, + { + "bbox": [ + 55, + 194, + 296, + 277 + ], + "type": "text", + "content": " provide insight into the distribution of attention to different modalities across various layers, aiding in understanding the reasons for the underutilization of visual information during the modality fusion process." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 278, + 296, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 278, + 296, + 386 + ], + "spans": [ + { + "bbox": [ + 55, + 278, + 296, + 386 + ], + "type": "text", + "content": "The experimental setup aligns with that described in Sec. 4.1. Fig. 4 illustrates the allocation of attention to different modalities across the model's layers. In the middle layers, attention to visual features is markedly lower than that given to system prompts and user instructions. This suggests that during the critical process of modality fusion, the model's focus on visual input is insufficient. As a result, visual information is underutilized, leading to an output distribution skewed toward language priors." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 394, + 116, + 407 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 394, + 116, + 407 + ], + "spans": [ + { + "bbox": [ + 55, + 394, + 116, + 407 + ], + "type": "text", + "content": "4.3. Insights" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 412, + 295, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 412, + 295, + 435 + ], + "spans": [ + { + "bbox": [ + 55, + 412, + 295, + 435 + ], + "type": "text", + "content": "Based on the experimental results presented in Sec. 4.1 and Sec. 4.2, two significant conclusions can be drawn:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 438, + 295, + 498 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 55, + 438, + 295, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 438, + 295, + 473 + ], + "spans": [ + { + "bbox": [ + 55, + 438, + 295, + 473 + ], + "type": "text", + "content": "- The model performs the crucial fusion of visual and textual modalities in the middle layers, creating cross-modal semantic representations that drive the final predictions." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 475, + 295, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 475, + 295, + 498 + ], + "spans": [ + { + "bbox": [ + 55, + 475, + 295, + 498 + ], + "type": "text", + "content": "- During this critical fusion process, the model demonstrates inadequate attention to the visual modality." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 500, + 296, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 500, + 296, + 548 + ], + "spans": [ + { + "bbox": [ + 55, + 500, + 296, + 548 + ], + "type": "text", + "content": "These findings indicate that models fail to fully utilize visual information, resulting in an excessive dependence on language priors and, subsequently, the occurrence of hallucination phenomena." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 559, + 213, + 573 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 559, + 213, + 573 + ], + "spans": [ + { + "bbox": [ + 55, + 559, + 213, + 573 + ], + "type": "text", + "content": "5. Visual Amplification Fusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 579, + 296, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 579, + 296, + 664 + ], + "spans": [ + { + "bbox": [ + 55, + 579, + 296, + 664 + ], + "type": "text", + "content": "Building on the insights presented in Sec. 4, we introduce a hallucination mitigation method called Visual Amplification Fusion (VAF). As illustrated in Fig. 5, This approach heightens attention to visual information during modality fusion, effectively reducing the excessive dependency on language priors and ensuring that the generated content is closely grounded to visual inputs." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 672, + 194, + 683 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 672, + 194, + 683 + ], + "spans": [ + { + "bbox": [ + 55, + 672, + 194, + 683 + ], + "type": "text", + "content": "5.1. Attention Redistribution" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "content": "As outlined in Sec. 4, the model performs crucial fusion of visual and textual modalities within the middle layers." + } + ] + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 326, + 70, + 541, + 190 + ], + "blocks": [ + { + "bbox": [ + 326, + 70, + 541, + 190 + ], + "lines": [ + { + "bbox": [ + 326, + 70, + 541, + 190 + ], + "spans": [ + { + "bbox": [ + 326, + 70, + 541, + 190 + ], + "type": "image", + "image_path": "4276ac80e92a8e4d964523fdd4d403e279bcfb08f04d187e5d1222c1e38e3c7a.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 193, + 555, + 239 + ], + "lines": [ + { + "bbox": [ + 313, + 193, + 555, + 239 + ], + "spans": [ + { + "bbox": [ + 313, + 193, + 555, + 239 + ], + "type": "text", + "content": "Figure 5. Illustration of the Visual Amplification Fusion Method. In the middle layers, we select attention heads highly responsive to visual information, amplifying their focus on visual features while reducing unnecessary attention to system prompts." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 256, + 554, + 303 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 256, + 554, + 303 + ], + "spans": [ + { + "bbox": [ + 313, + 256, + 554, + 303 + ], + "type": "text", + "content": "However, the attention allocated to visual modality information during this process remains insufficient. To address this, we adjust the attention weights in these layers to achieve a more balanced focus." + } + ] + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 326, + 311, + 542, + 440 + ], + "blocks": [ + { + "bbox": [ + 326, + 311, + 542, + 440 + ], + "lines": [ + { + "bbox": [ + 326, + 311, + 542, + 440 + ], + "spans": [ + { + "bbox": [ + 326, + 311, + 542, + 440 + ], + "type": "image", + "image_path": "f1a5946587d4e3f726094e6779b36975910feb9152052a38d0d7fc97e10b0959.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 449, + 555, + 503 + ], + "lines": [ + { + "bbox": [ + 313, + 449, + 555, + 503 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 555, + 503 + ], + "type": "text", + "content": "Figure 6. Effect of Enhanced Visual Attention on Hallucination Suppression. Increasing attention to visual features in the fusion process of the model's middle layers successfully reduces hallucinations, enabling the model to correct its grape color prediction from \"green\" to \"red\"." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 517, + 554, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 517, + 554, + 553 + ], + "spans": [ + { + "bbox": [ + 313, + 517, + 554, + 553 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 313, + 517, + 554, + 553 + ], + "type": "inline_equation", + "content": "A_{l,h}" + }, + { + "bbox": [ + 313, + 517, + 554, + 553 + ], + "type": "text", + "content": " denote the attention matrix of the " + }, + { + "bbox": [ + 313, + 517, + 554, + 553 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 313, + 517, + 554, + 553 + ], + "type": "text", + "content": "-th attention head in the " + }, + { + "bbox": [ + 313, + 517, + 554, + 553 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 313, + 517, + 554, + 553 + ], + "type": "text", + "content": "-th layer, and " + }, + { + "bbox": [ + 313, + 517, + 554, + 553 + ], + "type": "inline_equation", + "content": "Z_{l,h}" + }, + { + "bbox": [ + 313, + 517, + 554, + 553 + ], + "type": "text", + "content": " represent its corresponding attention score matrix, defined as:" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 384, + 560, + 553, + 573 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 384, + 560, + 553, + 573 + ], + "spans": [ + { + "bbox": [ + 384, + 560, + 553, + 573 + ], + "type": "interline_equation", + "content": "A _ {l, h} = \\operatorname {s o f t m a x} \\left(Z _ {l, h}\\right). \\tag {7}", + "image_path": "19648ce9cc3425808887793a637b83d8f8850b1470446da263ef5177b57ba373.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 579, + 555, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 579, + 555, + 663 + ], + "spans": [ + { + "bbox": [ + 313, + 579, + 555, + 663 + ], + "type": "text", + "content": "Our objective during the modality fusion process is to amplify the model's attention to visual features while curbing an overemphasis on system prompts. This adjustment facilitates improved integration of visual information and reduces over-reliance on language priors. To achieve this, we modify the attention score matrix in the middle layers (i.e., " + }, + { + "bbox": [ + 313, + 579, + 555, + 663 + ], + "type": "inline_equation", + "content": "8 < l < 15" + }, + { + "bbox": [ + 313, + 579, + 555, + 663 + ], + "type": "text", + "content": ") as follows:" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 325, + 669, + 553, + 685 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 669, + 553, + 685 + ], + "spans": [ + { + "bbox": [ + 325, + 669, + 553, + 685 + ], + "type": "interline_equation", + "content": "\\hat {Z} _ {l, h} = Z _ {l, h} + \\alpha \\cdot M _ {l, h} ^ {\\text {e n h}} \\circ Z _ {l, h} - \\beta \\cdot M _ {l, h} ^ {\\text {s u p}} \\circ Z _ {l, h}. \\tag {8}", + "image_path": "ac10dc3011db69474b4b9c2785f02ac76668536d3579db5ba870e9f60c41cff0.jpg" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "content": " is the enhancement coefficient " + }, + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "inline_equation", + "content": "(\\alpha >0)" + }, + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "content": " , where larger values indicate stronger amplification of visual attenuate" + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14629" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 131 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 131 + ], + "type": "text", + "content": "tion. The suppression coefficient " + }, + { + "bbox": [ + 55, + 72, + 295, + 131 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 55, + 72, + 295, + 131 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 55, + 72, + 295, + 131 + ], + "type": "inline_equation", + "content": "0 < \\beta < 1" + }, + { + "bbox": [ + 55, + 72, + 295, + 131 + ], + "type": "text", + "content": ") determines the extent of attention suppression directed at system prompts. The enhancement and suppression mask matrices, " + }, + { + "bbox": [ + 55, + 72, + 295, + 131 + ], + "type": "inline_equation", + "content": "M_{l,h}^{enh}" + }, + { + "bbox": [ + 55, + 72, + 295, + 131 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 72, + 295, + 131 + ], + "type": "inline_equation", + "content": "M_{l,h}^{sup}" + }, + { + "bbox": [ + 55, + 72, + 295, + 131 + ], + "type": "text", + "content": ", respectively, are defined to guide the modulation of attention elements:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 111, + 137, + 295, + 156 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 137, + 295, + 156 + ], + "spans": [ + { + "bbox": [ + 111, + 137, + 295, + 156 + ], + "type": "interline_equation", + "content": "M _ {l, h} ^ {e n h} (i, j) = \\mathbb {I} (i \\in \\mathcal {T}, j \\in \\mathcal {V}), \\tag {9}", + "image_path": "30623c3a122f4563c7023513f223453b6c2f4754e28f332fabf42351e97cd8d2.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 114, + 153, + 238, + 168 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 114, + 153, + 238, + 168 + ], + "spans": [ + { + "bbox": [ + 114, + 153, + 238, + 168 + ], + "type": "interline_equation", + "content": "M _ {l, h} ^ {s u p} (i, j) = \\mathbb {I} (i \\in \\mathcal {T}, j \\in \\mathcal {S}).", + "image_path": "6de344daac623231b003fc869ede69ca786f16a1d47ae0f7f61ef2c91bd7182d.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 175, + 296, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 175, + 296, + 247 + ], + "spans": [ + { + "bbox": [ + 55, + 175, + 296, + 247 + ], + "type": "text", + "content": "These modifications optimize attention allocation by enhancing the model's focus on visual features during modality fusion and minimizing superfluous attention to system prompts. As illustrated in Fig. 6, preliminary analysis indicates that this approach effectively mitigates hallucination issues by promoting greater attention to visual information." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 253, + 216, + 266 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 253, + 216, + 266 + ], + "spans": [ + { + "bbox": [ + 55, + 253, + 216, + 266 + ], + "type": "text", + "content": "5.2. Visual Perception Restriction" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 270, + 296, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 270, + 296, + 378 + ], + "spans": [ + { + "bbox": [ + 55, + 270, + 296, + 378 + ], + "type": "text", + "content": "Enhancing visual attention across all attention heads in the middle layers can be overly aggressive and may negatively impact content generation. To address this, we propose a selective enhancement strategy. Specifically, we identify and isolate the attention heads that exhibit higher sensitivity to visual information, which we term visual perception heads. We then restrict the visual attention enhancement to these visual perception heads, ensuring better utilization of visual information while maintaining overall model performance." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 379, + 296, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 379, + 296, + 462 + ], + "spans": [ + { + "bbox": [ + 55, + 379, + 296, + 462 + ], + "type": "text", + "content": "In the model, attention heads that allocate more attention to visual features demonstrate heightened sensitivity to visual information. Let " + }, + { + "bbox": [ + 55, + 379, + 296, + 462 + ], + "type": "inline_equation", + "content": "A_{l,h}" + }, + { + "bbox": [ + 55, + 379, + 296, + 462 + ], + "type": "text", + "content": " represent the attention matrix of the " + }, + { + "bbox": [ + 55, + 379, + 296, + 462 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 55, + 379, + 296, + 462 + ], + "type": "text", + "content": "-th attention head in the " + }, + { + "bbox": [ + 55, + 379, + 296, + 462 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 55, + 379, + 296, + 462 + ], + "type": "text", + "content": "-th layer of the model, with its corresponding visual attention allocation denoted by " + }, + { + "bbox": [ + 55, + 379, + 296, + 462 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{vis}}^{l,h}" + }, + { + "bbox": [ + 55, + 379, + 296, + 462 + ], + "type": "text", + "content": ". In each attention layer, we identify the attention heads whose visual attention allocation fall within the top" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "type": "text", + "content": " and designate them as visual perception heads, subsequently redistributing their attention. The attention matrices of the remaining attention heads are kept unchanged." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 119, + 391, + 133 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 119, + 391, + 133 + ], + "spans": [ + { + "bbox": [ + 313, + 119, + 391, + 133 + ], + "type": "text", + "content": "6. Experiment" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 140, + 555, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 140, + 555, + 236 + ], + "spans": [ + { + "bbox": [ + 313, + 140, + 555, + 236 + ], + "type": "text", + "content": "This section demonstrates the effectiveness of the proposed VAF method in mitigating hallucinations. Sec. 6.1 outlines the experimental setup, detailing the evaluation benchmarks and VAF parameter configurations. Sec. 6.2 then presents the experimental results from three perspectives: reduction of hallucinations, coherence of generated content, and inference speed. Finally, Sec. 6.3 further verifies the contribution of each VAF component through ablation studies." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 243, + 441, + 256 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 243, + 441, + 256 + ], + "spans": [ + { + "bbox": [ + 313, + 243, + 441, + 256 + ], + "type": "text", + "content": "6.1. Experimental Settings" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 261, + 554, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 261, + 554, + 297 + ], + "spans": [ + { + "bbox": [ + 313, + 261, + 554, + 297 + ], + "type": "text", + "content": "In Sec. 6.1.1, we present the selected datasets and evaluation metrics. Sec. 6.1.2 details the chosen MLLM backbone models, and Sec. 6.1.3 outlines the baseline settings." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 304, + 473, + 315 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 304, + 473, + 315 + ], + "spans": [ + { + "bbox": [ + 313, + 304, + 473, + 315 + ], + "type": "text", + "content": "6.1.1. Datasets & Evaluation Metrics" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 319, + 554, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 319, + 554, + 437 + ], + "spans": [ + { + "bbox": [ + 313, + 319, + 554, + 437 + ], + "type": "text", + "content": "Polling-based Object Probing Evaluation (POPE). POPE [30] is a novel framework designed to evaluate object hallucinations in MLLMs. Departing from traditional caption-based approaches, POPE frames hallucination detection as a binary task by posing straightforward yes-or-no questions regarding the presence of specific objects in an image (e.g., \"Is there a chair in the image?\"). Performance on POPE is measured across four metrics: Accuracy, Precision, Recall, and F1 score, allowing for a thorough evaluation of hallucinations in MLLMs." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 437, + 554, + 463 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 554, + 463 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 554, + 463 + ], + "type": "text", + "content": "Multimodal Model Evaluation (MME). MME [13] benchmark provides a comprehensive framework for evalu" + } + ] + } + ], + "index": 14 + }, + { + "type": "table", + "bbox": [ + 100, + 479, + 515, + 679 + ], + "blocks": [ + { + "bbox": [ + 100, + 479, + 515, + 679 + ], + "lines": [ + { + "bbox": [ + 100, + 479, + 515, + 679 + ], + "spans": [ + { + "bbox": [ + 100, + 479, + 515, + 679 + ], + "type": "table", + "html": "
CategoryMethodLLaVA-v1.5-7BLLaVA-v1.5-13BQwen-VL-Chat-7B
AccuracyF1-scoreAccuracyF1-scoreAccuracyF1-score
RandomRegular87.8 ↑ 0.087.5 ↑ 0.087.6 ↑ 0.087.4 ↑ 0.088.2 ↑ 0.087.9 ↑ 0.0
VCD88.4 ↑ 0.687.7 ↑ 0.288.9 ↑ 1.387.8 ↑ 0.489.1 ↑ 0.988.4 ↑ 0.5
ICD88.1 ↑ 0.387.6 ↑ 0.188.1 ↑ 0.587.6 ↑ 0.288.9 ↑ 0.788.1 ↑ 0.2
VAF89.6 ↑ 1.889.3 ↑ 1.890.1 ↑ 2.589.9 ↑ 2.590.0 ↑ 1.889.7 ↑ 1.8
PopularRegular82.5 ↑ 0.083.2 ↑ 0.082.7 ↑ 0.084.1 ↑ 0.082.4 ↑ 0.083.1 ↑ 0.0
VCD83.1 ↑ 0.684.1 ↑ 0.983.7 ↑ 1.085.1 ↑ 1.083.0 ↑ 0.684.1 ↑ 1.0
ICD82.1 ↓ 0.482.9 ↓ 0.382.9 ↑ 0.284.3 ↑ 0.283.2 ↑ 0.884.5 ↑ 1.4
VAF84.5 ↑ 2.084.9 ↑ 1.785.2 ↑ 2.586.4 ↑ 2.384.9 ↑ 2.585.1 ↑ 2.0
AdversarialRegular77.6 ↑ 0.079.4 ↑ 0.077.8 ↑ 0.079.5 ↑ 0.077.2 ↑ 0.078.9 ↑ 0.0
VCD78.1 ↑ 0.579.6 ↑ 0.278.2 ↑ 0.479.7 ↑ 0.278.8 ↑ 1.680.1 ↑ 1.2
ICD78.5 ↑ 0.979.9 ↑ 0.579.1 ↑ 1.380.1 ↑ 0.678.1 ↑ 0.979.2 ↑ 0.3
VAF80.1 ↑ 2.581.0 ↑ 1.680.7 ↑ 2.981.7 ↑ 2.280.4 ↑ 3.281.2 ↑ 2.3
", + "image_path": "8d459acb9b44953bf440c455acef2cbaa8db11da66027968c375775beab5b9ba.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_body" + } + ], + "index": 15 + }, + { + "bbox": [ + 55, + 687, + 554, + 711 + ], + "lines": [ + { + "bbox": [ + 55, + 687, + 554, + 711 + ], + "spans": [ + { + "bbox": [ + 55, + 687, + 554, + 711 + ], + "type": "text", + "content": "Table 2. Performance on POPE. Results are averaged across the MS-COCO, A-OKVQA, and GQA datasets. The VAF method demonstrates superior hallucination suppression across all three MLLMs. The best performance for each setting is highlighted in red." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14630" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 69, + 70, + 545, + 272 + ], + "blocks": [ + { + "bbox": [ + 69, + 70, + 545, + 272 + ], + "lines": [ + { + "bbox": [ + 69, + 70, + 545, + 272 + ], + "spans": [ + { + "bbox": [ + 69, + 70, + 545, + 272 + ], + "type": "table", + "html": "
ModelMethodObject-levelAttribute-levelTotal Score
ExistenceCountPositionColor
LLaVA-v1.5-7BRegular185.00↑0.00146.67↑0.00128.33↑0.00150.00↑0.00610.00↑0.00
VCD185.00↑0.00141.33↓5.34128.33↑0.00153.00↑3.00607.66↓2.34
ICD185.00↑0.00148.33↑1.66126.66↓1.67148.33↓1.67608.32↓1.68
VAF195.00↑10.00158.33↑11.66128.33↑0.00155.00↑5.00636.67↑26.67
LLaVA-v1.5-13BRegular185.00↑0.00155.00↑0.00133.33↑0.00165.00↑0.00638.33↑0.00
VCD185.00↑0.00155.00↑0.00130.00↓3.33168.33↑3.33638.33↑0.00
ICD183.33↓1.67153.33↓1.67131.67↓1.66165.00↑0.00633.33↓5.00
VAF195.00↑10.00160.00↑5.00136.67↑3.34170.00↑5.00661.67↑23.34
Qwen-VL-7BRegular158.33↑0.00150.00↑0.00128.33↑0.00170.00↑0.00606.66↑0.00
VCD158.33↑0.00150.00↑0.00133.33↑5.00175.00↑5.00616.66↑10.00
ICD128.33↓30.00151.67↑1.67128.33↑0.00170.00↑0.00578.33↓28.33
VAF165.00↑6.67155.00↑5.00133.33↑5.00175.00↑5.00628.33↑21.67
", + "image_path": "b83f59448363e37e83379459d404902b696ee19afef140d8b901b37a9a94927b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 279, + 555, + 304 + ], + "lines": [ + { + "bbox": [ + 55, + 279, + 555, + 304 + ], + "spans": [ + { + "bbox": [ + 55, + 279, + 555, + 304 + ], + "type": "text", + "content": "Table 3. Results on the MME subset. Across three MLLMs, the VAF method achieved the most effective suppression of both object-level and attribute-level hallucinations. The highest scores in each setting are highlighted in red." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 323, + 294, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 323, + 294, + 418 + ], + "spans": [ + { + "bbox": [ + 54, + 323, + 294, + 418 + ], + "type": "text", + "content": "ating MLLMs across both perceptual and cognitive dimensions. It consists of ten perception-oriented tasks and four cognition-oriented tasks, with model performance assessed through accuracy metrics. In addition to the full dataset, we leverage specific subsets, such as object existence and counting to analyze object-level hallucinations, while position and color subsets are employed to examine attributelevel hallucinations." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 418, + 295, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 418, + 295, + 503 + ], + "spans": [ + { + "bbox": [ + 54, + 418, + 295, + 503 + ], + "type": "text", + "content": "Novel Object Captioning at Scale (Nocaps). NoCaps [3] benchmark is designed to evaluate image captioning models on their ability to describe novel objects absent from standard datasets like COCO. Model performance is quantified using the CIDEr score, providing a basis to assess the coherence and accuracy of generated captions in response to images containing unfamiliar objects." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 510, + 165, + 521 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 165, + 521 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 165, + 521 + ], + "type": "text", + "content": "6.1.2. MLLM Backbones" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 524, + 295, + 656 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 524, + 295, + 656 + ], + "spans": [ + { + "bbox": [ + 54, + 524, + 295, + 656 + ], + "type": "text", + "content": "In comparison to the Q-former structure, linear projection demonstrates greater efficiency in aligning visual and textual features. This advantage is evident in MLLMs with linear projection architectures, such as LLaVA and Qwen-VL, which outperform Q-former-based MLLMs like Instruct-BLIP and MiniGPT4. Based on these findings, we selected three linear-projection-based MLLMs, specifically LLaVA-v1.5-7B, LLaVA-v1.5-13B [35], and Qwen-VL-7B [5], to evaluate the effectiveness of our proposed VAF method. Detailed prompt templates for each model across various benchmarks are included in Sec. 10." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 662, + 159, + 675 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 662, + 159, + 675 + ], + "spans": [ + { + "bbox": [ + 55, + 662, + 159, + 675 + ], + "type": "text", + "content": "6.1.3. Baseline Settings." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": "We primarily compared our approach to the VCD [23] and ICD [47] methods. VCD mitigates hallucinations by contrasting output distributions derived from original and dis" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 323, + 555, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 323, + 555, + 396 + ], + "spans": [ + { + "bbox": [ + 313, + 323, + 555, + 396 + ], + "type": "text", + "content": "torted visual inputs, while ICD reduces hallucinated concepts by comparing distributions generated with standard versus disrupted instructions. To ensure consistency and reproducibility in our comparisons, all methods use greedy search. Unless specified otherwise, our experiments set " + }, + { + "bbox": [ + 313, + 323, + 555, + 396 + ], + "type": "inline_equation", + "content": "\\beta = 0.1" + }, + { + "bbox": [ + 313, + 323, + 555, + 396 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 323, + 555, + 396 + ], + "type": "inline_equation", + "content": "\\alpha = 0.15" + }, + { + "bbox": [ + 313, + 323, + 555, + 396 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 403, + 434, + 415 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 403, + 434, + 415 + ], + "spans": [ + { + "bbox": [ + 313, + 403, + 434, + 415 + ], + "type": "text", + "content": "6.2. Results and Analysis" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 420, + 554, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 420, + 554, + 480 + ], + "spans": [ + { + "bbox": [ + 313, + 420, + 554, + 480 + ], + "type": "text", + "content": "Sec. 6.2.1 examines the effectiveness of various methods in mitigating hallucinations, while Sec. 6.2.2 assesses their impact on the quality of generated content. Sec. 6.2.3 then analyzes the influence of each method on inference speed. Additional experimental results are provided in Sec. 8." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 486, + 448, + 498 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 486, + 448, + 498 + ], + "spans": [ + { + "bbox": [ + 313, + 486, + 448, + 498 + ], + "type": "text", + "content": "6.2.1. Hallucination Mitigation" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 501, + 554, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 501, + 554, + 609 + ], + "spans": [ + { + "bbox": [ + 313, + 501, + 554, + 609 + ], + "type": "text", + "content": "Tab. 2 presents the experimental results of the VAF method on the POPE benchmark, with results averaged across the MSCOCO [31], A-OKVQA [42], and GQA [20] datasets. Applied to both the LLaVA-v1.5 model family and the Qwen-VL model, the VAF method consistently surpasses the VCD and ICD methods in reducing hallucinations. Tab. 3 further highlights the performance of VAF on the MME benchmark, demonstrating its effectiveness in suppressing both object-level and attribute-level hallucinations." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 615, + 484, + 625 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 615, + 484, + 625 + ], + "spans": [ + { + "bbox": [ + 313, + 615, + 484, + 625 + ], + "type": "text", + "content": "6.2.2. Coherence of Generated Content" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 629, + 554, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 554, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 554, + 715 + ], + "type": "text", + "content": "Tab. 4 presents the experimental results for various methods on the Nocaps and ScienceQA datasets. It is evident that VCD and ICD substantially degrade the quality of the generated content. Specifically, on the Nocaps dataset, VCD and ICD reduce CIDEr scores by " + }, + { + "bbox": [ + 313, + 629, + 554, + 715 + ], + "type": "inline_equation", + "content": "18\\%" + }, + { + "bbox": [ + 313, + 629, + 554, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 629, + 554, + 715 + ], + "type": "inline_equation", + "content": "27\\%" + }, + { + "bbox": [ + 313, + 629, + 554, + 715 + ], + "type": "text", + "content": ", respectively. This degradation primarily arises from the crude disruption of language priors by contrastive decoding methods," + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "14631" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 62, + 71, + 293, + 218 + ], + "blocks": [ + { + "bbox": [ + 62, + 71, + 293, + 218 + ], + "lines": [ + { + "bbox": [ + 62, + 71, + 293, + 218 + ], + "spans": [ + { + "bbox": [ + 62, + 71, + 293, + 218 + ], + "type": "table", + "html": "
ModelDecodingScienceQANocaps
AccuracyCIDEr
LLaVA-v1.5-7BRegular68.078.7
VCD64.565.7
ICD62.462.3
VAF68.578.8
LLaVA-v1.5-13BRegular71.682.6
VCD70.068.9
ICD69.260.3
VAF71.782.3
", + "image_path": "0b9e784b26f293f21ef2cd61b90fdbb90125e2e44a7f7f3a3b426a35298e190f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 270, + 295, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 270, + 295, + 319 + ], + "spans": [ + { + "bbox": [ + 55, + 270, + 295, + 319 + ], + "type": "text", + "content": "which leads to generated content lacking coherence and accuracy. By contrast, our method demonstrates minimal negative impact on prediction results, maintaining both coherence and accuracy effectively." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 324, + 153, + 337 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 324, + 153, + 337 + ], + "spans": [ + { + "bbox": [ + 55, + 324, + 153, + 337 + ], + "type": "text", + "content": "6.2.3. Inference Speed" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 339, + 296, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 339, + 296, + 411 + ], + "spans": [ + { + "bbox": [ + 55, + 339, + 296, + 411 + ], + "type": "text", + "content": "Fig. 7 illustrates the impact of different strategies on model inference speed within the Nocaps dataset. In comparison, the VCD and ICD methods nearly double the inference time due to the need to process contrastive input samples, whereas the VAF method has minimal impact on the inference speed of multimodal large language models." + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 69, + 426, + 271, + 566 + ], + "blocks": [ + { + "bbox": [ + 69, + 426, + 271, + 566 + ], + "lines": [ + { + "bbox": [ + 69, + 426, + 271, + 566 + ], + "spans": [ + { + "bbox": [ + 69, + 426, + 271, + 566 + ], + "type": "image", + "image_path": "b27fbe02e4ad6e6fd802441e5658fbec11550e4ce18c5748aac63639a42b5f9d.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 573, + 295, + 608 + ], + "lines": [ + { + "bbox": [ + 55, + 573, + 295, + 608 + ], + "spans": [ + { + "bbox": [ + 55, + 573, + 295, + 608 + ], + "type": "text", + "content": "Figure 7. Comparison of different strategies on inference speed. The VCD and ICD methods reduce inference speed by " + }, + { + "bbox": [ + 55, + 573, + 295, + 608 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 55, + 573, + 295, + 608 + ], + "type": "text", + "content": ", whereas the VAF method shows minimal impact." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 624, + 149, + 637 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 624, + 149, + 637 + ], + "spans": [ + { + "bbox": [ + 55, + 624, + 149, + 637 + ], + "type": "text", + "content": "6.3. Ablation Study" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": "Ablation studies on the enhancement coefficient " + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": " were conducted using the COCO-Random dataset within the POPE benchmark to understand its influence on model performance. Fig. 8 demonstrates that when " + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "inline_equation", + "content": "0 < \\alpha < 0.25" + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": ", model hallucinations are effectively suppressed. However, when " + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": " surpasses 0.25, performance starts to degrade. We" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "text", + "content": "propose that this reduction in performance may stem from an excessive focus on visual features, disrupting the balanced integration of language information and diminishing overall model effectiveness." + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 329, + 129, + 538, + 271 + ], + "blocks": [ + { + "bbox": [ + 329, + 129, + 538, + 271 + ], + "lines": [ + { + "bbox": [ + 329, + 129, + 538, + 271 + ], + "spans": [ + { + "bbox": [ + 329, + 129, + 538, + 271 + ], + "type": "image", + "image_path": "a46d5daaa6e27ef0a8238450f81f2345dcd1ab7d14cb92d2d08010dacc032f81.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 327, + 279, + 541, + 291 + ], + "lines": [ + { + "bbox": [ + 327, + 279, + 541, + 291 + ], + "spans": [ + { + "bbox": [ + 327, + 279, + 541, + 291 + ], + "type": "text", + "content": "Figure 8. Ablation study of " + }, + { + "bbox": [ + 327, + 279, + 541, + 291 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 327, + 279, + 541, + 291 + ], + "type": "text", + "content": " on the POPE benchmark." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 294, + 555, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 294, + 555, + 390 + ], + "spans": [ + { + "bbox": [ + 313, + 294, + 555, + 390 + ], + "type": "text", + "content": "We performed ablation studies on the visual perception restriction mechanism, evaluating its impact on the POPE and Nocaps benchmarks. Tab. 5 highlights the effects of restricting attention reallocation to visual perception heads. Increasing attention to visual features alone reduces model hallucinations, while confining this reallocation strategy to visual perception heads minimizes adverse effects on content quality. More ablation studies can be found in Sec. 9." + } + ] + } + ], + "index": 12 + }, + { + "type": "table", + "bbox": [ + 323, + 399, + 547, + 482 + ], + "blocks": [ + { + "bbox": [ + 55, + 227, + 295, + 251 + ], + "lines": [ + { + "bbox": [ + 55, + 227, + 295, + 251 + ], + "spans": [ + { + "bbox": [ + 55, + 227, + 295, + 251 + ], + "type": "text", + "content": "Table 4. Results on SQA and Nocaps datasets. The highest and second-highest scores are marked in red and blue, respectively." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 323, + 399, + 547, + 482 + ], + "lines": [ + { + "bbox": [ + 323, + 399, + 547, + 482 + ], + "spans": [ + { + "bbox": [ + 323, + 399, + 547, + 482 + ], + "type": "table", + "html": "
ModelVisual RestrictionPOPENocaps
LLaVA-7B✓89.878.8
X89.976.4
LLaVA-13B✓90.282.3
X90.081.1
", + "image_path": "9151df19ce8ee99f5549b71b1bbb4e4591341c7229e476bc904b1ad988e237de.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 491, + 555, + 525 + ], + "lines": [ + { + "bbox": [ + 313, + 491, + 555, + 525 + ], + "spans": [ + { + "bbox": [ + 313, + 491, + 555, + 525 + ], + "type": "text", + "content": "Table 5. Ablation Study of Visual Perception Restriction Mechanism. Restricting attention redistribution to the visual perception heads more effectively preserves the quality of generated content." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 314, + 538, + 384, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 538, + 384, + 550 + ], + "spans": [ + { + "bbox": [ + 314, + 538, + 384, + 550 + ], + "type": "text", + "content": "7. conclusion" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "text", + "content": "In this paper, we identify two key drawbacks of using contrastive decoding to mitigate hallucinations in MLLMs: reduced quality of generated content and slower inference speed. To address these challenges, we propose a novel approach, Visual Amplification Fusion, which effectively mitigates hallucinations while preserving both inference speed and content generation quality. By enhancing the attention to visual features during modality fusion, VAF minimizes the over-reliance on language priors, ensuring a high degree of consistency between generated content and visual inputs. Extensive experiments across multiple benchmarks and MLLMs demonstrate that VAF provides a clear advantage in hallucination mitigation." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "14632" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 163 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 163 + ], + "type": "text", + "content": "This work is supported by the National Natural Science Foundation of China under Grant 62176246. This work is also supported by Anhui Province Key Research and Development Plan (202304a05020045), Anhui Province Natural Science Foundation (2208085UD17) and National Natural Science Foundation of China under Grant 62406098." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 172, + 115, + 185 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 172, + 115, + 185 + ], + "spans": [ + { + "bbox": [ + 56, + 172, + 115, + 185 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 193, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 61, + 193, + 296, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 193, + 296, + 247 + ], + "spans": [ + { + "bbox": [ + 61, + 193, + 296, + 247 + ], + "type": "text", + "content": "[1] Vedika Agarwal, Rakshith Shetty, and Mario Fritz. Towards causal vqa: Revealing and reducing spurious correlations by invariant and covariant semantic editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9690-9698, 2020. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 248, + 296, + 280 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 248, + 296, + 280 + ], + "spans": [ + { + "bbox": [ + 62, + 248, + 296, + 280 + ], + "type": "text", + "content": "[2] Aishwarya Agrawal, Dhruv Batra, and Devi Parikh. Analyzing the behavior of visual question answering models. arXiv preprint arXiv:1606.07356, 2016. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 282, + 296, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 282, + 296, + 335 + ], + "spans": [ + { + "bbox": [ + 62, + 282, + 296, + 335 + ], + "type": "text", + "content": "[3] Harsh Agrawal, Karan Desai, Yufei Wang, Xinlei Chen, Rishabh Jain, Mark Johnson, Dhruv Batra, Devi Parikh, Stefan Lee, and Peter Anderson. nocaps: novel object captioning at scale. In 2019 IEEE/CVF International Conference on Computer Vision (ICCV). IEEE, 2019. 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 337, + 295, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 337, + 295, + 357 + ], + "spans": [ + { + "bbox": [ + 62, + 337, + 295, + 357 + ], + "type": "text", + "content": "[4] Jinze Bai, Shuai Bai, and et al. Qwen technical report. arXiv preprint arXiv:2309.16609, 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 359, + 295, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 359, + 295, + 390 + ], + "spans": [ + { + "bbox": [ + 62, + 359, + 295, + 390 + ], + "type": "text", + "content": "[5] Jinze Bai, Shuai Bai, and et al. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966, 2023. 2, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 392, + 295, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 392, + 295, + 413 + ], + "spans": [ + { + "bbox": [ + 62, + 392, + 295, + 413 + ], + "type": "text", + "content": "[6] Rohan Bavishi, Erich Elsen, and et al. Introducing our multimodal models, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 415, + 295, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 415, + 295, + 469 + ], + "spans": [ + { + "bbox": [ + 62, + 415, + 295, + 469 + ], + "type": "text", + "content": "[7] Ali Furkan Biten, Lluís Gómez, and Dimosthenis Karatzas. Let there be a clock on the beach: Reducing object hallucination in image captioning. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1381-1390, 2022. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 471, + 295, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 471, + 295, + 502 + ], + "spans": [ + { + "bbox": [ + 62, + 471, + 295, + 502 + ], + "type": "text", + "content": "[8] Keqin Chen, Zhao Zhang, and et al. Shikra: Unleashing multimodal llm's referential dialogue magic. arXiv preprint arXiv:2306.15195, 2023. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 504, + 295, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 504, + 295, + 557 + ], + "spans": [ + { + "bbox": [ + 62, + 504, + 295, + 557 + ], + "type": "text", + "content": "[9] Long Chen, Oleg Sinavski, Jan Hunermann, Alice Karnsund, Andrew James Willmott, Danny Birch, Daniel Maund, and Jamie Shotton. Driving with llms: Fusing object-level vector modality for explainable autonomous driving. arXiv preprint arXiv:2310.01957, 2023. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 559, + 295, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 559, + 295, + 601 + ], + "spans": [ + { + "bbox": [ + 57, + 559, + 295, + 601 + ], + "type": "text", + "content": "[10] Zhe Chen, Weiyun Wang, and et al. How far are we to gpt-4v? closing the gap to commercial multimodal models with open-source suites. arXiv preprint arXiv:2404.16821, 2024. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 603, + 295, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 603, + 295, + 645 + ], + "spans": [ + { + "bbox": [ + 57, + 603, + 295, + 645 + ], + "type": "text", + "content": "[11] Wei-Lin Chiang and Zhuohan et al Li. Vicuna: An opensource chatbot impressing gpt-4 with " + }, + { + "bbox": [ + 57, + 603, + 295, + 645 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 57, + 603, + 295, + 645 + ], + "type": "text", + "content": " chatgpt quality. See https://vicuna.lmsys.org (accessed 14 April 2023), 2023. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 647, + 295, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 295, + 679 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 295, + 679 + ], + "type": "text", + "content": "[12] Wenliang Dai and Junnan Li et al. Instructlip: Towards general-purpose vision-language models with instruction tuning, 2023. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 681, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 681, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 681, + 295, + 713 + ], + "type": "text", + "content": "[13] Chaoyou Fu, Peixian Chen, and et al. Mme: A comprehensive evaluation benchmark for multimodal large language models. arXiv preprint arXiv:2306.13394, 2023. 6" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 554, + 713 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 128 + ], + "type": "text", + "content": "[14] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Bartra, and Devi Parikh. Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6904-6913, 2017. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 129, + 553, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 129, + 553, + 161 + ], + "spans": [ + { + "bbox": [ + 317, + 129, + 553, + 161 + ], + "type": "text", + "content": "[15] Anisha Gunjal, Jihan Yin, and Erhan Bas. Detecting and preventing hallucinations in large vision language models. arXiv preprint arXiv:2308.06394, 2023. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 163, + 553, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 163, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 317, + 163, + 553, + 228 + ], + "type": "text", + "content": "[16] Vipul Gupta, Zhuowan Li, Adam Kortylewski, Chenyu Zhang, Yingwei Li, and Alan Yuille. Swapmix: Diagnosing and regularizing the over-reliance on visual context in visual question answering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5078-5088, 2022. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 230, + 553, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 230, + 553, + 273 + ], + "spans": [ + { + "bbox": [ + 317, + 230, + 553, + 273 + ], + "type": "text", + "content": "[17] Yudong Han, Liqiang Nie, Jianhua Yin, Jianlong Wu, and Yan Yan. Visual perturbation-aware collaborative learning for overcoming the language prior problem. arXiv preprint arXiv:2207.11850, 2022. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 275, + 553, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 275, + 553, + 317 + ], + "spans": [ + { + "bbox": [ + 317, + 275, + 553, + 317 + ], + "type": "text", + "content": "[18] Mingzhe Hu, Shaoyan Pan, Yuheng Li, and Xiaofeng Yang. Advancing medical imaging with language models: A journey from n-grams to chatgpt. arXiv preprint arXiv:2304.04920, 2023. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 320, + 553, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 320, + 553, + 384 + ], + "spans": [ + { + "bbox": [ + 317, + 320, + 553, + 384 + ], + "type": "text", + "content": "[19] Qidong Huang, Xiaoyi Dong, Pan Zhang, Bin Wang, Conghui He, Jiaqi Wang, Dahua Lin, Weiming Zhang, and Nenghai Yu. Opera: Alleviating hallucination in multimodal large language models via over-trust penalty and retrospection-allocation. In CVPR, pages 13418-13427, 2024. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 387, + 553, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 387, + 553, + 419 + ], + "spans": [ + { + "bbox": [ + 317, + 387, + 553, + 419 + ], + "type": "text", + "content": "[20] Drew A Hudson and Christopher D Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering. In CVPR, pages 6700-6709, 2019. 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 421, + 553, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 421, + 553, + 463 + ], + "spans": [ + { + "bbox": [ + 317, + 421, + 553, + 463 + ], + "type": "text", + "content": "[21] Fushuo Huo, Wenchao Xu, Zhong Zhang, Haozhao Wang, Zhicheng Chen, and Peilin Zhao. Self-introspective decoding: Alleviating hallucinations for large vision-language models, 2024. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 465, + 553, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 465, + 553, + 498 + ], + "spans": [ + { + "bbox": [ + 317, + 465, + 553, + 498 + ], + "type": "text", + "content": "[22] Chaoya Jiang, Haiyang Xu, and et al. Hallucination augmented contrastive learning for multimodal large language model. In CVPR, pages 27036-27046, 2024. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 500, + 553, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 500, + 553, + 542 + ], + "spans": [ + { + "bbox": [ + 317, + 500, + 553, + 542 + ], + "type": "text", + "content": "[23] Sicong Leng, Hang Zhang, and et al. Mitigating object hallucinations in large vision-language models through visual contrastive decoding. In CVPR, pages 13872-13882, 2024. 2, 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 544, + 553, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 544, + 553, + 576 + ], + "spans": [ + { + "bbox": [ + 317, + 544, + 553, + 576 + ], + "type": "text", + "content": "[24] Bo Li, Yuanhan Zhang, and et al. Mimic-it: Multi-modal in-context instruction tuning. arXiv preprint arXiv:2306.05425, 2023. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 578, + 553, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 578, + 553, + 600 + ], + "spans": [ + { + "bbox": [ + 317, + 578, + 553, + 600 + ], + "type": "text", + "content": "[25] Bo Li, Kaichen Zhang, and et al. Llava next: Stronger llms supercharge multimodal capabilities in the wild, 2024. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 317, + 602, + 553, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 602, + 553, + 634 + ], + "spans": [ + { + "bbox": [ + 317, + 602, + 553, + 634 + ], + "type": "text", + "content": "[26] Chunyuan Li, Cliff Wong, and et al. Llava-med: Training a large language-and-vision assistant for biomedicine in one day. In NeurIPS, pages 28541-28564, 2023. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 317, + 635, + 553, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 635, + 553, + 678 + ], + "spans": [ + { + "bbox": [ + 317, + 635, + 553, + 678 + ], + "type": "text", + "content": "[27] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. BLIP-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML, 2023. 1, 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "text", + "content": "[28] Xiang Lisa Li, Ari Holtzman, and et al. Contrastive decoding: Open-ended text generation as optimization. arXiv preprint arXiv:2210.15097, 2022. 2" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14633" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "type": "text", + "content": "[29] Yifan Li, Yifan Du, Kun Zhou, Jinpeng Wang, Wayne Xin Zhao, and Ji-Rong Wen. Evaluating object hallucination in large vision-language models. arXiv preprint arXiv:2305.10355, 2023. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 163 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 163 + ], + "type": "text", + "content": "[30] Yifan Li, Yifan Du, Kun Zhou, Jinpeng Wang, Xin Zhao, and Ji-Rong Wen. Evaluating object hallucination in large vision-language models. In EMNLP, pages 292-305, 2023. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 165, + 294, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 165, + 294, + 196 + ], + "spans": [ + { + "bbox": [ + 56, + 165, + 294, + 196 + ], + "type": "text", + "content": "[31] Tsung-Yi Lin, Michael Maire, and et al. Microsoft coco: Common objects in context. In ECCV, pages 740-755, 2014. 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 199, + 294, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 199, + 294, + 243 + ], + "spans": [ + { + "bbox": [ + 56, + 199, + 294, + 243 + ], + "type": "text", + "content": "[32] Fuxiao Liu, Kevin Lin, Linjie Li, Jianfeng Wang, Yaser Ya-coob, and Lijuan Wang. Mitigating hallucination in large multi-modal models via robust instruction tuning. arXiv preprint arXiv:2306.14565, 2023. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 245, + 294, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 245, + 294, + 277 + ], + "spans": [ + { + "bbox": [ + 56, + 245, + 294, + 277 + ], + "type": "text", + "content": "[33] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, pages 34892-34916, 2023. 1, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 280, + 294, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 280, + 294, + 323 + ], + "spans": [ + { + "bbox": [ + 56, + 280, + 294, + 323 + ], + "type": "text", + "content": "[34] Haokun Liu, Yaonan Zhu, Kenji Kato, Izumi Kondo, Tadayoshi Aoyama, and Yasuhisa Hasegawa. Lm-based human-robot collaboration framework for manipulation tasks. arXiv preprint arXiv:2308.14972, 2023. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 326, + 294, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 326, + 294, + 358 + ], + "spans": [ + { + "bbox": [ + 56, + 326, + 294, + 358 + ], + "type": "text", + "content": "[35] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In CVPR, pages 26296-26306, 2024. 1, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 361, + 294, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 361, + 294, + 404 + ], + "spans": [ + { + "bbox": [ + 56, + 361, + 294, + 404 + ], + "type": "text", + "content": "[36] Zhi-Song Liu, Robin Courant, and Vicky Kalogeiton. Funnynet-w: Multimodal learning of funny moments in videos in the wild. International Journal of Computer Vision, pages 1-22, 2024. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 406, + 294, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 406, + 294, + 449 + ], + "spans": [ + { + "bbox": [ + 56, + 406, + 294, + 449 + ], + "type": "text", + "content": "[37] Holy Lvenia, Wenliang Dai, Samuel Cahyawijaya, Ziwei Ji, and Pascale Fung. Negative object presence evaluation (nope) to measure object hallucination in vision-language models. arXiv preprint arXiv:2310.05338, 2023. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 452, + 294, + 517 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 452, + 294, + 517 + ], + "spans": [ + { + "bbox": [ + 56, + 452, + 294, + 517 + ], + "type": "text", + "content": "[38] Pan Lu, Swaroop Mishra, Tony Xia, Liang Qiu, Kai-Wei Chang, Song-Chun Zhu, Oyvind Tafjord, Peter Clark, and Ashwin Kalyan. Learn to explain: Multimodal reasoning via thought chains for science question answering. In The 36th Conference on Neural Information Processing Systems (NeurIPS), 2022. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 519, + 294, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 519, + 294, + 563 + ], + "spans": [ + { + "bbox": [ + 56, + 519, + 294, + 563 + ], + "type": "text", + "content": "[39] Jinjie Mai, Jun Chen, Bing Li, Guocheng Qian, Mohamed Elhoseiny, and Bernard Ghanem. Llm as a robotic brain: Unifying egocentric memory and control. arXiv preprint arXiv:2304.09349, 2023. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 565, + 294, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 565, + 294, + 586 + ], + "spans": [ + { + "bbox": [ + 56, + 565, + 294, + 586 + ], + "type": "text", + "content": "[40] AI Meta. Introducing meta llama 3: The most capable openly available llm to date. Meta AI, 2024. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 589, + 294, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 589, + 294, + 643 + ], + "spans": [ + { + "bbox": [ + 56, + 589, + 294, + 643 + ], + "type": "text", + "content": "[41] Yulei Niu, Kaihua Tang, Hanwang Zhang, Zhiwu Lu, XianSheng Hua, and Ji-Rong Wen. Counterfactual vqa: A cause-effect look at language bias. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12700-12710, 2021. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 646, + 294, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 646, + 294, + 678 + ], + "spans": [ + { + "bbox": [ + 56, + 646, + 294, + 678 + ], + "type": "text", + "content": "[42] Dustin Schwenk, Apoorv Khandelwal, and et al. A-okvqa: A benchmark for visual question answering using world knowledge. In ECCV, pages 146–162, 2022. 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 680, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 680, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 680, + 294, + 713 + ], + "type": "text", + "content": "[43] Rohan Taori, Ishaan Gulrajani, and et al. Stanford alpaca: an instruction-following llama model (2023). URL https://github.com/tatsu-lab/stanford_alpaca, 1(9), 2023. 2" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 72, + 554, + 555 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 315, + 72, + 554, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 72, + 554, + 106 + ], + "spans": [ + { + "bbox": [ + 315, + 72, + 554, + 106 + ], + "type": "text", + "content": "[44] Hugo Touvron, Thibaut Lavril, and et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 108, + 554, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 108, + 554, + 139 + ], + "spans": [ + { + "bbox": [ + 316, + 108, + 554, + 139 + ], + "type": "text", + "content": "[45] Hugo Touvron, Louis Martin, and et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 141, + 554, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 554, + 185 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 554, + 185 + ], + "type": "text", + "content": "[46] Sheng Wang, Zihao Zhao, Xi Ouyang, Qian Wang, and Dinggang Shen. Chatcad: Interactive computer-aided diagnosis on medical image using large language models. arXiv preprint arXiv:2302.07257, 2023. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 186, + 554, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 186, + 554, + 228 + ], + "spans": [ + { + "bbox": [ + 316, + 186, + 554, + 228 + ], + "type": "text", + "content": "[47] Xintong Wang, Jingheng Pan, and et al. Mitigating hallucinations in large vision-language models with instruction contrastive decoding. arXiv preprint arXiv:2403.18715, 2024.2, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 230, + 554, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 230, + 554, + 274 + ], + "spans": [ + { + "bbox": [ + 316, + 230, + 554, + 274 + ], + "type": "text", + "content": "[48] Yike Wu, Yu Zhao, Shiwan Zhao, Ying Zhang, Xiaojie Yuan, Guoqing Zhao, and Ning Jiang. Overcoming language priors in visual question answering via distinguishing superficially similar instances. arXiv preprint arXiv:2209.08529, 2022. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 276, + 554, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 276, + 554, + 308 + ], + "spans": [ + { + "bbox": [ + 316, + 276, + 554, + 308 + ], + "type": "text", + "content": "[49] Zhenyu Wu, Ziwei Wang, Xiuwei Xu, Jiwen Lu, and Haibin Yan. Embodied task planning with large language models. arXiv preprint arXiv:2307.01848, 2023. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 309, + 554, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 309, + 554, + 353 + ], + "spans": [ + { + "bbox": [ + 316, + 309, + 554, + 353 + ], + "type": "text", + "content": "[50] Hong Yan, Lijun Liu, Xupeng Feng, and Qingsong Huang. Overcoming language priors with self-contrastive learning for visual question answering. Multimedia Tools and Applications, 82(11):16343–16358, 2023. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 354, + 554, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 354, + 554, + 387 + ], + "spans": [ + { + "bbox": [ + 316, + 354, + 554, + 387 + ], + "type": "text", + "content": "[51] Qinghao Ye, Haiyang Xu, and et al. mplug-owl: Modularization empowers large language models with multimodality. arXiv preprint arXiv:2304.14178, 2023. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 388, + 554, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 388, + 554, + 421 + ], + "spans": [ + { + "bbox": [ + 316, + 388, + 554, + 421 + ], + "type": "text", + "content": "[52] Shilong Zhang, Peize Sun, and et al. Gpt4roi: Instruction tuning large language model on region-of-interest. arXiv preprint arXiv:2307.03601, 2023. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 422, + 554, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 422, + 554, + 477 + ], + "spans": [ + { + "bbox": [ + 316, + 422, + 554, + 477 + ], + "type": "text", + "content": "[53] Ren Zhibo, Wang Huizhen, Zhu Muhua, Wang Yichao, Xiao Tong, and Zhu Jingbo. Overcoming language priors with counterfactual inference for visual question answering. In Proceedings of the 22nd Chinese National Conference on Computational Linguistics, pages 600-610, 2023. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 478, + 554, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 478, + 554, + 510 + ], + "spans": [ + { + "bbox": [ + 316, + 478, + 554, + 510 + ], + "type": "text", + "content": "[54] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. IJCV, 130(9):2337-2348, 2022. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 511, + 554, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 511, + 554, + 555 + ], + "spans": [ + { + "bbox": [ + 316, + 511, + 554, + 555 + ], + "type": "text", + "content": "[55] Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 1, 2" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "14634" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/ClimbingCap_ Multi-Modal Dataset and Method for Rock Climbing in World Coordinate/56189f24-e199-449b-838e-e830c93475ae_content_list.json b/2025/ClimbingCap_ Multi-Modal Dataset and Method for Rock Climbing in World Coordinate/56189f24-e199-449b-838e-e830c93475ae_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..1f44dea89906160031060b8f99c4dd272db34552 --- /dev/null +++ b/2025/ClimbingCap_ Multi-Modal Dataset and Method for Rock Climbing in World Coordinate/56189f24-e199-449b-838e-e830c93475ae_content_list.json @@ -0,0 +1,1608 @@ +[ + { + "type": "text", + "text": "ClimbingCap: Multi-Modal Dataset and Method for Rock Climbing in World Coordinate", + "text_level": 1, + "bbox": [ + 142, + 130, + 828, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ming Yan $^{1,2,3*}$ Xincheng Lin $^{1,3*}$ Yuhua Luo $^{1,3}$ Shuqi Fan $^{1,3}$ Yudi Dai $^{7}$ Qixin Zhong $^{4}$ Lincai Zhong $^{5}$ Yuexin Ma $^{6}$ Lan Xu $^{6}$ Chenglu Wen $^{1,3}$ Siqi Shen $^{1,3\\dagger}$ Cheng Wang $^{1,3}$", + "bbox": [ + 96, + 186, + 854, + 224 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Fujian Key Laboratory of Sensing and Computing for Smart Cities, Xiamen University", + "bbox": [ + 138, + 227, + 828, + 246 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ National Institute for Data Science in Health and Medicine, Xiamen University", + "bbox": [ + 166, + 246, + 799, + 263 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3Key Laboratory of Multimedia Trusted Perception and Efficient Computing,", + "bbox": [ + 181, + 263, + 787, + 281 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ministry of Education of China, School of Informatics, Xiamen University", + "bbox": [ + 191, + 282, + 777, + 297 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "4China National Climbing Team 5Ningbo Sports Work Training Team", + "bbox": [ + 205, + 299, + 764, + 316 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{6}$ ShanghaiTech University ETH AI Center, ETH Zürich", + "bbox": [ + 259, + 316, + 709, + 334 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c3ad0d134fecc3ddf3c7919012bf63efd650c3a94edc4a9feb043c105f06e1fa.jpg", + "image_caption": [ + "Figure 1. Overview. To address the challenging problem of global climbing motion recovery, we collect the dataset AscendMotion, using LiDAR, RGB camera and Inertial Measurement Unit (IMU) motion capture system with accurate motion labels and global trajectories (the blue and orange human bodies in right side of the figure represent labeled motions, and the orange curve represents the motion trajectory in the world coordinate.). Meanwhile, we propose ClimbingCap, a global climbing motion capturing method in world coordinate. As shown in the left part of this figure, it uses both image and LiDAR point cloud to recover human motions." + ], + "image_footnote": [], + "bbox": [ + 78, + 340, + 888, + 537 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 623, + 313, + 638 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Human Motion Recovery (HMR) research mainly focuses on ground-based motions such as running. The study on capturing climbing motion, an off-ground motion, is sparse. This is partly due to the limited availability of climbing motion datasets, especially large-scale and challenging 3D labeled datasets. To address the insufficiency of climbing motion datasets, we collect AscendMotion, a large-scale well-annotated, and challenging climbing motion dataset. It consists of $412k$ RGB, LiDAR frames, and IMU measurements, including the challenging climbing motions of 22 skilled climbing coaches across 12 different rock walls. Capturing the climbing motions is challenging as it requires precise recovery of not only the complex pose but also the global position of climbers. Although multi", + "bbox": [ + 75, + 654, + 470, + 866 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ple global HMR methods have been proposed, they cannot faithfully capture climbing motions. To address the limitations of HMR methods for climbing, we propose ClimbingCap, a motion recovery method that reconstructs continuous 3D human climbing motion in a global coordinate system. One key insight is to use the RGB and LiDAR modalities to separately reconstruct motions in camera coordinates and global coordinates and to optimize them jointly. We demonstrate the quality of the AscendMotion dataset and present promising results from ClimbingCap. The AscendMotion dataset and source code release publicly at http://www.lidarhumanmotion.net/climbingcap/", + "bbox": [ + 500, + 625, + 892, + 806 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 500, + 814, + 632, + 830 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "It is challenging for global Human Motion Recovery (HMR) [1, 9, 40, 42, 80] due to the complexity of human poses and dynamic interaction between humans and their environments. Moreover, global HMR must main", + "bbox": [ + 500, + 839, + 892, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Equal contribution.", + "bbox": [ + 101, + 875, + 215, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author.", + "bbox": [ + 102, + 887, + 230, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "12312", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tain consistency in world coordinates to ensure authenticity and physical feasibility. The research community has proposed various methods to estimate human poses from different sensors such as RGB images [21-23, 33, 71], LiDAR point clouds [36, 49, 78], and inertial measurement units (IMUs) [25, 76]. Most of the studies [8, 15, 16, 24, 28, 39, 49, 62, 68, 73, 78] focus on recovering ground-based motions such as running, interacting, and dancing.", + "bbox": [ + 75, + 90, + 472, + 210 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Unlike ground-based motions such as running, climbing is an activity performed off-ground, where climbers use hands and feet to ascend holds or walls. As an important sport and recreation, the climbing race has become an official event in the Olympic Games. Yet, research on climbing motion capture [26, 34, 35, 54, 61] is sparse, partly due to the limited availability of climbing datasets. There are only two publicly available climbing datasets: SPEEDD21 [19] and CIMI4D [72]. SPEEDD21 is a 2D dataset. Although CIMI4D is 3D, it consists of many trivial climbing motions from casual climbers. Moreover, the size of the two datasets is small. As a result, the research community lacks an in-depth understanding of challenging climbing motions.", + "bbox": [ + 75, + 214, + 472, + 411 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We collect AscendMotion, a large-scale and challenging climbing motion multi-modal dataset, to address the data insufficiency. AscendMotion dataset includes the motions of 22 skilled climbers climbing 12 different rock walls. It includes rich modalities from hardware timesynchronized RGB, LiDAR point clouds, and Inertial Measurement Unit(IMU) MoCap system. Moreover, it consists of the global trajectories of climbers, which are important for understanding human motions in world coordinates. We also combine automatic annotation with manual refinement to ensure the accuracy of the motion annotation.", + "bbox": [ + 75, + 412, + 472, + 579 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Not only does climbing motion challenge humans, but capturing climbing motion is also challenging for researchers. Capturing climbing motion requires precise estimation of the complex pose resulting from human-scene interactions, as well as the global localization of climbers within scenes. Existing research dedicated to climbing [8, 62, 68] uses standard 2D HMR (e.g. OpenPose [9]) for climbing motions, which is suboptimal. Most existing HMR methods [20, 37, 55, 56, 60, 75] focus on recovering ground-based human motions through RGB imagery. The estimation results of these methods cannot be rigidly transformed among camera coordinates and global coordinate systems, due to the inherent ambiguity between coordinate systems. Moreover, global HMR methods often accumulate errors during long-term sequence recovery, affecting the accuracy of global trajectories and postures, especially in climbing movements, as we show through experiments.", + "bbox": [ + 75, + 580, + 472, + 838 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To tackle the above challenges and fill the gap in the motion capture community for climbing motions, we introduce ClimbingCap, an HMR method that reconstructs continuous 3D human climbing motion in both camera and", + "bbox": [ + 75, + 839, + 472, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "global coordinate systems (see Fig.1). We adopt a trilogy in ClimbingCap: separate coordinate decoding, postprocessing, and semi-supervised training. For the separate coordinate decoding stage, ClimbingCap uses image and point cloud modality to estimate poses in the camera and the global coordinate systems, respectively. The postprocessing stage accurately ensures the consistency of motions between the two coordinate systems. In the semi-supervised training stage, a teacher-student training method is adopted to make use of easily obtainable unlabeled climbing data. Through these innovations, we can handle complex climbing movements effectively. To summarize, our main contributions include:", + "bbox": [ + 496, + 90, + 893, + 286 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We collect the dataset AscendMotion, which is more comprehensive than existing datasets. We demonstrate the quality of the dataset and its poses are challenging.", + "- We propose ClimbingCap, which is a multimodality global HMR method for rock climbing.", + "- We conduct extensive experiments on evaluate the ClimbingCap in multiple datasets with various state-of-the-art methods. The experimental results show ClimbingCap perform better than all the methods for climbing motions." + ], + "bbox": [ + 500, + 287, + 890, + 424 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 439, + 640, + 455 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Camera-Space HMR. Camera-space HMR methods [7, 12, 13, 18, 27, 30, 31, 38, 44, 46, 48, 52, 53, 57-59, 63, 65, 70, 74, 81] predict human motion in independent camera coordinate system for each frame of the input video. Although camera-space methods can predict accurate human poses, they cannot effectively reflect the orientation and trajectory of the human body in world coordinates.", + "bbox": [ + 496, + 468, + 893, + 575 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "World-Grounded HMR. Recent studies [32, 37, 55, 56, 60, 75, 77] have achieved remarkable progress in world coordinates. Most of them are two-stage methods. The first stage estimates human motion in camera coordinates, and the second stage optimizes the motion in world coordinates with the scene information. Most global HMR methods focus on ground-based motions (e.g., walking) that are located on the ground. They cannot accurately model climbing motions, which is an off-ground motion. Our method, ClimbingCap, is a global HMR method developed for recovering rock climbing motions.", + "bbox": [ + 496, + 578, + 893, + 744 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "HMR Dataset. We have compared the AscendMotion dataset with multiple recent datasets in Table Tab. 1. Most of the datasets focus on ground-based human daily activities or sports. There exist only two public available climbing motion datasets. SPEED21 [19] collects 2D motions of climbers from sports videos. CIMI4D [72] contains 3D motions and global trajectories of climbers. Our climbing dataset, AscendMotion (344 minutes labeled data and 441 minutes unlabeled data, 412k frames), is significantly larger than both SPEED21 (21 minutes, 46k frames) and CIMI4D", + "bbox": [ + 496, + 750, + 895, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "12313", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "table", + "img_path": "images/dbc930d8732376cdca91189e9da57bec65c520c9b378a07cf4b9dff33d21341d.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetSensor ModalitiesGlobal TrajectoryFrames3D SceneMotionReal/SyntheticSeqsSubjectsCareer
RGBMoCapLiDAR
SLOPER4D [15]✓IMU✓✓100k✓DaliyReal1512Normal
EMDB[28]✓EM-✓105k-DaliyReal8110Normal
BEHAVE [5]✓--✓15k-InteractionsReal-8Normal
RICH [24]✓--✓577k✓InteractionsReal14222Normal
AGORA [45]✓--✓106.7k✓DaliySynthetic---
BEDLAM [6]✓---1M✓DaliySynthetic---
FreeMotion [49]✓IMU✓✓578k-DaliyReal--Normal
HmPEAR [39]✓-✓-300k-DaliyReal6k25Normal
RELI11D [73]✓IMU✓✓239k✓SportReal4810Normal
HiSC4D [16]✓IMU✓✓36k✓SportReal8-Normal
LiDARHuman51M [78]✓IMU✓-374k-DailyReal5210Normal
CIMI4D [72]✓IMU✓✓180k✓ClimbingReal4212Normal
SPEED21 [19]✓--46kClimbingReal95--
AscendMotion(Ours)✓IMU✓✓412k✓(12)ClimbingReal22022Skilled
", + "bbox": [ + 81, + 88, + 893, + 330 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Table 1. Comparisons with related HMR datasets. The ”-” symbol indicates that it is not included in the dataset.", + "bbox": [ + 151, + 340, + 815, + 354 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "(120 minutes, 180k frames). Moreover, the motions in our dataset are more challenging than those in CIMI4D, as they come from climbing coaches who are more skilled than the casual climbers in CIMI4D.", + "bbox": [ + 75, + 375, + 468, + 436 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Climbing Motion Recovery. In recent years, the increasing popularity of climbing, coupled with the inclusion of rock climbing races in the Olympic Games, has led to growing interest in the capture and analysis of climbing motions [2, 3, 10, 51, 67]. Research on climbing MoCap methods remains limited. [43] employs OpenPose [9] for climbing pose estimation. [50] performs RGB-based method for local pose and estimates climbing position through a marker-based method.", + "bbox": [ + 75, + 440, + 470, + 575 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method: ClimbingCap", + "text_level": 1, + "bbox": [ + 76, + 594, + 292, + 612 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Climbing motion capture is challenging, involving poses with extreme limb extension and full-body exertion in camera coordinates. Moreover, it requires precise alignment with the rock wall in world coordinate as climbers are ascending. The ClimbingCap pipeline consists of three parts: separate coordinate decoding, post-processing, and semi-supervised training. An overview of the proposed pipeline is shown in Fig. 2. The separate coordinate decoding and post-processing parts take into account the unique challenges posed by climbing motion, which involves complex off-ground dynamics and interactions with scenes. The semi-supervised training part makes uses of the large-scale unlabeled climbing motion data to learn better a HMR model.", + "bbox": [ + 75, + 621, + 468, + 832 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Separate Coordinate Decoding", + "text_level": 1, + "bbox": [ + 76, + 845, + 349, + 862 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The separate coordinate decoding (SCD) stage extracts features from the RGB sequence and LiDAR point clouds, and", + "bbox": [ + 75, + 869, + 468, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "predicts the poses in camera coordinates and the positions in global coordinates.", + "bbox": [ + 496, + 375, + 890, + 405 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Input and Feature Extraction. The overall network structure is illustrated in Fig. 2. The input includes RGB images and point cloud data. First, the point cloud data is transformed from the world coordinate system to the camera coordinate system via an extrinsic matrix, represented as $\\mathcal{P}_{\\mathrm{c}} = \\Omega_{w2c} \\cdot \\mathcal{P}_{\\mathrm{w}}$ . Subsequently, the RGB images and transformed point cloud data are passed through feature extraction modules, RGB Extract and PC Extract, to obtain visual and geometric features. We build the feature extraction modules based on ViT [17] and PointNet++ [47]. These features are then fed into the following two decoder modules, which regress the SMPL parameters and global motion parameters of the human body, respectively.", + "bbox": [ + 496, + 409, + 892, + 606 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Camera Coordinate Decoder. This module decodes the SMPL parameters in the camera coordinate system. The RGB and point cloud features serve as inputs to the Camera Coordinate Decoder (denoted as $\\mathcal{T}_{\\mathrm{Decoder}}$ ), which processes the inputs with contextual information $\\mathbf{f}_{\\mathrm{backbone}}$ , generating an output token $\\mathbf{t}_{\\mathrm{out}}$ . This output token is then used to iteratively optimize the SMPL parameters, including the pose $\\theta$ , shape $\\beta$ , and camera translation $\\Delta c$ . The iterative decoding approach allows the model to gradually approximate the true pose and shape in the camera coordinate system. In each iteration, the decoder updates the current SMPL parameters $\\theta_{i}, \\beta_{i}$ , and $\\Delta c_{t}$ as follows:", + "bbox": [ + 496, + 609, + 893, + 790 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {t} _ {\\text {o u t}} = \\mathcal {T} _ {\\text {D e c o d e r}} (\\mathbf {t}, \\mathbf {f} _ {\\text {b a c k b o n e}}), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 604, + 800, + 890, + 816 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\theta_{i + 1} = \\Phi_{\\theta}\\cdot \\mathbf{t}_{\\mathrm{out}} + \\theta_{i},\\beta_{i + 1} = \\Phi_{\\beta}\\cdot \\mathbf{t}_{\\mathrm{out}} + \\beta_{i}$ and $\\Delta c_{i + 1} = \\Phi_c\\cdot \\mathbf{t}_{\\mathrm{out}} + \\Delta c_i$ are the update equations, with $\\Phi_{\\theta}$ , $\\Phi_{\\beta}$ , and $\\Phi_c$ representing the respective weight matrices for each parameter. Here, the input token $\\mathbf{t}$ can include initialized pose, shape, and camera parameters as needed.", + "bbox": [ + 498, + 824, + 893, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "12314", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/b67825fb626a0f553f2c89c4583a420fc7d2f223b3664a0b47b0d3f1c9a777d1.jpg", + "image_caption": [ + "Figure 2. Overview of ClimbingCap. The arrows indicate the three stages of the ClimbingCap framework: separate coordinate decoding (the green box), post-processing (the blue box), and semi-supervised training (the red box)." + ], + "image_footnote": [], + "bbox": [ + 81, + 89, + 537, + 292 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/f2055e3ca958be05ee17c14dc0e93d409a5e8bc3b45085b1fc921b9637d45885.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 540, + 88, + 890, + 292 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Global Coordinate Decoder. To fully capture the human motion trajectory in the world coordinate, we design the Global Translation Decoder to predict the global translation parameters $\\Gamma^{\\mathrm{trans}}$ of the human body. In this module, the decoder processes the features $\\mathbf{f}_{\\mathrm{backbone}}$ as contextual input, iteratively updating the global translation parameters. The update formula in each iteration is given by:", + "bbox": [ + 75, + 354, + 468, + 460 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\Gamma_ {i + 1} ^ {\\text {t r a n s}} = \\Psi \\cdot \\mathbf {t} _ {\\text {o u t}} + \\Gamma_ {i} ^ {\\text {t r a n s}}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 189, + 473, + 468, + 491 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\Gamma_{i}^{\\mathrm{trans}}$ represents the global translation parameters at time step $i$ , and $\\Psi$ is the weight matrix for the update. This decoding process enables the model to capture a complete motion trajectory in the global coordinate system.", + "bbox": [ + 75, + 506, + 468, + 568 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Loss. The total loss function not only includes the 3D keypoint loss $\\mathcal{L}_{\\mathrm{kp3d}}$ and 2D keypoint loss $\\mathcal{L}_{\\mathrm{kp2d}}$ but also incorporates the SMPL parameter loss $\\mathcal{L}^{\\mathrm{spl}}$ and the global trajectory loss $\\mathcal{L}_{\\mathrm{traj}}$ . Specifically, the 3D keypoint loss $\\mathcal{L}_{\\mathrm{kp3d}}$ measures the 3D error of the predicted keypoints, while the 2D keypoint loss $\\mathcal{L}_{\\mathrm{kp2d}}$ measures the 2D projection error. The SMPL parameter loss $\\mathcal{L}^{\\mathrm{spl}}$ supervises the decoded pose and shape parameters, and the global trajectory loss $\\mathcal{L}_{\\mathrm{traj}}$ constrains the translation parameters to within close distance of ground truth positions. Please see the supplementary for details. The final total loss is formulated as:", + "bbox": [ + 75, + 571, + 468, + 738 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {k p 3 d} + \\mathcal {L} _ {k p 2 d} + \\mathcal {L} _ {\\theta} ^ {s m p l} + \\mathcal {L} _ {\\beta} ^ {s m p l} + \\mathcal {L} _ {t r a j}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 102, + 753, + 468, + 772 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Post-processing", + "text_level": 1, + "bbox": [ + 76, + 800, + 233, + 816 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Researches [55, 56, 75] have shown that a post-processing stage can be used to improve the output motion recovery results. Following these approaches, we employ a post-processing stage to optimize the output pose from SCD 3.1 stage.", + "bbox": [ + 75, + 824, + 468, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "One distinct advantage of ClimbingCap is that the output results for the pose decoding stage can be rigidly transformed between the camera coordinate system and the world coordinate system. Thanks to the LiDAR modality, the point cloud contains 3D information in the world coordinate system. The poses obtained from the SCD stage are converted from the camera coordinate system to the world coordinate system through the inverse extrinsic matrix $\\Omega_{w2c}^{-1}$ .", + "bbox": [ + 496, + 354, + 890, + 491 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The post-processing stage consists of three losses: $\\mathcal{L}_{LWD}$ , $\\mathcal{L}_{SDS}$ , and $\\mathcal{L}_{VLR}$ . $\\mathcal{L}_{LWD}$ assigns different weights to the vertices of different parts of the climbing human SMPL, and optimizes the position of the human body in the world coordinate system according to these weights. $\\mathcal{L}_{SDS}$ optimizes joint positions by smoothing the changes in the direction of the human body's velocity on the scene (i.e., rock wall). $\\mathcal{L}_{VLR}$ optimizes the difficult-to-estimate limb end poses by using the position of the point cloud in space. In this stage, the global poses are optimized through using the Adam [29] optimizer. Please refer to the supplementary for detailed definition of these losses.", + "bbox": [ + 496, + 491, + 892, + 672 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Semi-supervised Training", + "text_level": 1, + "bbox": [ + 500, + 680, + 733, + 696 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Compared to ground-based motions [41], the size of labeled climbing motions is small. Simply using labeled climbing motion data may not be enough to train a robust model. Different from labeled climbing motion data, collecting unlabeled climbing motion data is cheaper. The AscendMotion dataset contains more unlabeled data than label data. They can be used to further improve the HMR model.", + "bbox": [ + 496, + 703, + 890, + 809 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "It has been shown by researches [64, 69, 79] from the object detection community that through using a teacher-student semi-supervised training framework, the performance of object detection models can be improved. We adopt such semi-supervised training frame for HMR. In this work, we refer the model trained after the SCD and the post", + "bbox": [ + 496, + 810, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "12315", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/2f5bb39936921b1bc236f306567782bea86427b6529ebdb6bcc2199ba0be9fb2.jpg", + "image_caption": [ + "Figure 3. Dataset collection hardware system." + ], + "image_footnote": [], + "bbox": [ + 81, + 89, + 467, + 287 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "processing stage as the teacher model (green box in Fig. 2). The student model (red box in Fig. 2) clones the parameters of the teacher model. During semi-supervised training, the teacher model estimate the pose labels from unlabeled motion data, and the pose label is used as pseudo-label to further train the student model. We show in the experiment section that via utilizing the semi-supervised training framework, the performance of HMR can be further improved. Please refer to the supplementary for the detailed training process.", + "bbox": [ + 75, + 330, + 472, + 483 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. The AscendMotion Dataset", + "text_level": 1, + "bbox": [ + 76, + 494, + 328, + 510 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To advance research in Human Mesh Recovery (HMR) for climbing motion within a global coordinate system, we present the AscendMotion dataset. This dataset specifically captures complex multi-directional rock climbing motions on non-planar surfaces. It provides a unique opportunity for challenging HMR research. AscendMotion includes multi-modal motion data from skilled climbers performing various climbing styles, such as bouldering, speed climbing, and lead climbing, with synchronized recordings from IMU, RGB cameras, and LiDAR. The dataset comprises 344 minutes (6 hours) of labeled data and 441 minutes of unlabeled data, featuring 22 adult climbers across 12 real-world climbing scenes. All participants agreed to use their recorded data for scientific purposes.", + "bbox": [ + 75, + 521, + 468, + 733 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Hardware and Configuration", + "text_level": 1, + "bbox": [ + 76, + 741, + 336, + 758 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The AscendMotion data collection system integrates multiple sensors and captures motion data in both indoor and outdoor environments. As shown in Fig. 3, the LiDAR (OusterOS1, 128 beams) captures 3D dynamic point clouds at 20 frames per second (FPS), while the Hik 1080P RGB camera records RGB video at 20 FPS. For each climbing environment, we use the Trimble X7 3D laser scanner to reconstruct high-resolution RGB point cloud scenes, each containing approximately 80 million points. For the labeled", + "bbox": [ + 75, + 763, + 468, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "dataset, each climber wears an Xsens MVN inertial motion capture system with 17 IMUs, recording at 60 FPS. [4] and a handheld point cloud scanner to ensure accurate body shape representation. For the unlabeled dataset, climbers do not wear a MoCap outfit, and we record their climbing motions using a RGB camera and a LiDAR. A human motion is represented by $M = (T, \\theta, \\beta)$ , where $T$ represents global translation, $\\theta$ is the SMPL [40] pose parameters, and $\\beta$ is the SMPL shape parameter. We obtain each climber's shape parameters $\\beta$ using IPNet [4].", + "bbox": [ + 496, + 90, + 890, + 243 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Annotation Pipeline", + "text_level": 1, + "bbox": [ + 498, + 253, + 689, + 270 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The pose $\\theta$ and the translation $T$ obtained from the IMU measurements may be inaccurate. Especially the translation $T$ may significantly drift for long-duration capture. The annotation pipeline of AscendMotion is used to find the accurate translate $T$ and pose $\\theta$ as labels.", + "bbox": [ + 496, + 277, + 890, + 353 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The annotation pipeline of AscendMotion is depicted in Fig. 4. It consists of three parts: preprocessing, multistage global optimization, and manual annotation. The preprocessing part is used for time synchronization and spatial calibration. The multi-stage global optimization part is developed for automatic generation of annotation labels. The manual annotation part is used to further improve the quality of generated annotations.", + "bbox": [ + 496, + 354, + 890, + 474 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2.1 Multi-modal Data Preprocessing Stage", + "text_level": 1, + "bbox": [ + 498, + 497, + 823, + 513 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Time Synchronization. The time among RGB camera and LiDAR is synchronized via Precision Time Protocol (PTP). We employ CollShark Auto 66 unit as the master clock, and it sends PTP slave clocks to the RGB camera and LiDAR. The time of IMU MoCap is post-synchronized with the LiDAR and RGB through anchor frames.", + "bbox": [ + 496, + 526, + 890, + 617 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Calibration. First, the LiDAR point cloud are registered with high-precision scanned-scenes. Next, the coordinate of LiDAR is treated as the world coordinate. The IMU measurements are transformed into the world coordinate through a calibration matrix. Finally, we perform frame-level calibration among RGB, LiDAR and IMU. Please refer to the appendix for details.", + "bbox": [ + 496, + 621, + 890, + 727 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2.2 Multi-stage Global Optimization", + "text_level": 1, + "bbox": [ + 498, + 750, + 781, + 765 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "AscendMotion uses the translation $T$ and pose $\\theta$ provided by the IMU MoCap as the initialization of annotation labels, and performs multi-stage global optimization. To achieve accurate and natural human motion data consistent with the scene, we apply two loss functions: the Global Refit Loss $\\mathcal{L}_{\\mathrm{GR}}$ and the Scene Touch Loss $\\mathcal{L}_{\\mathrm{ST}}$ .", + "bbox": [ + 496, + 775, + 890, + 866 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Global Refit Loss $\\mathcal{L}_{\\mathrm{GR}}$ . The global refit loss promotes precise global alignment between the SMPL model and the", + "bbox": [ + 496, + 869, + 890, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "12316", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/cb659f075ccab75a8b7ef24ee341b2eee0dfa8721f3287452241e0c2a727fc21.jpg", + "image_caption": [ + "Figure 4. AscendMotion Annotation Pipeline. From left to right, this pipeline represents the three stages of dataset annotation: time-space synchronous input pre-processing(the blue box), multi-stage global optimization(the green box), and manual repair(the red box)." + ], + "image_footnote": [], + "bbox": [ + 81, + 101, + 277, + 250 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4eaebb54defeeede5abb455c3ee28a2042c923d3e256fb4ba42597f6c3737d9e.jpg", + "image_caption": [ + "Multi-modal data preprocessing" + ], + "image_footnote": [], + "bbox": [ + 277, + 101, + 444, + 250 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f0efa8d3143d7e6e2957752f95e095641c12c19e46df544446a3ed86939e99c2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 446, + 101, + 614, + 250 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/eaa8e9134b89f229f15dd2e6151ce8dd1bcdec8de604fb7faf664f190e59899a.jpg", + "image_caption": [ + "Multi-stage global optimization" + ], + "image_footnote": [], + "bbox": [ + 617, + 101, + 784, + 250 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/194057910d2d7f7e7f177bed9f78643651b2d946d684688b6aab09ff446a2aee.jpg", + "image_caption": [ + "Manual Repair" + ], + "image_footnote": [], + "bbox": [ + 785, + 101, + 888, + 250 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "point clouds. $\\mathcal{L}_{\\mathrm{GR}}$ calculates the geometric discrepancy between the SMPL model vertices $V$ and the human body point cloud $P$ using a modified Chamfer distance. It applies different loss terms ( $f(|v_{i} - p_{j}|^{2})$ or $g(|v_{i} - p_{j}|^{2}))$ to different body parts to make body regions, such as the torso and limbs, match the point clouds of the human. We calculate the Euclidean distance between a vertex $v_{i}\\in V$ and a point $p_j\\in P$ as $d_{v_i,p_j}$ . A loss term $f(|v_i - p_j|^2)$ is applied to a vertex $v_{i}$ if $d_{v_i,p_j}\\leq d_{\\mathrm{torso}}$ , and $g(|p_j - v_i|^2)$ is applied to $v_{i}$ if $d_{v_i,p_j}\\leq d_{\\mathrm{limb}}$ , where $d_{\\mathrm{torso}}$ and $d_{\\mathrm{limb}}$ are the distance thresholds for torso and limbs, respectively. During climbing, humans use limbs to climb up, the distance among limbs and scenes should smaller than the distance among torso and scenes, $d_{\\mathrm{limb}}\\leq d_{\\mathrm{torso}}$ .", + "bbox": [ + 75, + 318, + 472, + 530 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Scene Touch Loss $\\mathcal{L}_{\\mathrm{ST}}$ . This loss prevents unrealistic intersections between the SMPL model and the scene mesh by measuring the penetration depth between the SMPL vertices $v_{i} \\in V$ and the scene mesh vertices $q_{j} \\in Q$ . Given the scene mesh normal vectors $n_{j}$ , the penetration depth for each vertex is calculated as the dot product $\\eta(v_{i}) = (v_{i} - q_{j}) \\cdot n_{j}$ , where $q_{j}$ is the closest mesh vertex to $v_{i}$ . If the penetration depth $\\eta(v_{i})$ is negative, it indicates that the vertex $v_{i}$ has penetrated the scene mesh, and this value contributes to the loss $\\mathcal{L}_{\\mathrm{ST}}$ . This loss encourages the SMPL model to avoid intersecting with the scene mesh, promoting more realistic interactions between the model and the environment.", + "bbox": [ + 75, + 535, + 468, + 731 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2.3 Manual Annotation and Verification", + "text_level": 1, + "bbox": [ + 76, + 753, + 382, + 768 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To further improve the dataset quality, we use the SMPL annotation tool [66] for enhanced labeling of sequences. We engage four observers to review the automatically annotated results. For any ambiguous limb movements observed in RGB views, we perform additional manual adjustments and mark these frames as key frames. Finally, we propagate poses from multiple key frames within each sequence to ensure global consistency in both smoothness and accuracy.", + "bbox": [ + 75, + 779, + 468, + 902 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/26e84c208b3581fef37bcd1ab1f9de38c0053493141d6489982ff09da2e707ce.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Constraint termHorizontal SceneVertical Scene
\\( \\mathcal{L}_{ST} \\)\\( \\mathcal{L}_{GR} \\)ACCEL↓MPJPE↓PA-MPJPE↓ACCEL↓MPJPE↓PA-MPJPE↓
XX1.8337.5732.526.7951.5146.24
✓X1.6735.4531.813.7340.3635.37
X✓1.5929.0425.951.6629.2424.99
✓✓1.5428.2023.251.6029.0724.37
", + "bbox": [ + 503, + 315, + 890, + 386 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Quantitative Evaluation of Dataset Quality.", + "bbox": [ + 540, + 397, + 849, + 411 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. EXPERIMENTS", + "text_level": 1, + "bbox": [ + 500, + 441, + 668, + 458 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we demonstrate the high-quality of the AscendMotion optimization process in Sec. 5.1, show that the AscendMotion dataset is challenging to existing HMR methods, and evaluate ClimbingCap and multiple HMR methods in climbing motion datasets in Sec. 5.2. We demonstrate the promising performance of ClimbingCap and shows that each component of ClimbingCap is useful through ablation study in Sec. 5.3. Please refer to the detailed experimental setup and metrics in the supplementary.", + "bbox": [ + 496, + 470, + 890, + 608 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Datasets Evaluation", + "text_level": 1, + "bbox": [ + 500, + 626, + 689, + 641 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To quantitatively evaluate the annotation quality of AscendMotion, we divide the scene into horizontal and vertical rock walls and evaluate the performance of the global optimization stage (in Sec. 4.2.2) by comparing the generated annotations with manual annotations.", + "bbox": [ + 496, + 652, + 890, + 727 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The evaluation metrics include widely used MPJPE, Procrustes Aligned MPJPE (PA-MPJPE), per vertex error (PVE), acceleration error (Accel, in units of $m / s^2$ ). These metrics are used for evaluating HMR methods as well.", + "bbox": [ + 496, + 731, + 890, + 792 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To understand the impact of different constraints used in the combined optimization stage, we perform an ablation study on two different losses: $\\mathcal{L}_{ST}$ and $\\mathcal{L}_{GR}$ . Tab. 2 shows the error metrics using different loss combinations for two scenes. The error metrics are small, demonstrating the effectiveness of the global optimization pipeline and the high quality of AscendMotion.", + "bbox": [ + 496, + 794, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "12317", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/bd996503b0d5a7b37b23d2873470e13553fdf37ea61aac769822259a48d7d949.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModalityMethodCamera CoordinateWorld Coordinate
ACCEL↓MPJPE↓PA-MPJPE↓PVE↓PCK0.3↑WA-MPJPE↓W-MPJPE↓RTE↓Jitter↓T-Error↓
RGBTRACE [60]18.68/76.79875.56/577.6069.21/85.81951.52/619.930.06/0.09144.33/385.71254.38/703.3514.73/26.17115.96/521.402.56/6.62
SLAHMR [75]5.46/96.98232.46/467.8884.13/285.15283.24/552.630.36/0.12277.47/447.68804.85/613.603.64/39.394.91/201.332.81/6.54
WHAM [56]4.59/35.01110.92/143.1776.09/73.36124.2/164.910.76/0.62229.42/1125.77647.70/1499.855.16/9.043.58/40.691.77/2.49
GVHMR [55]4.50/26.22107.09/124.6060.06/80.30118.89/151.100.77/0.71105.15/1002.11202.45/1442.504.09/7.916.85/32.711.48/2.54
LiDARLiDARCapV2 [78]87.99/119.62244.6/234.52192.17/156.39326.45/283.270.53/0.50282.12/1396.42442.12/1518.2916.42/10.85176.95/165.551.65/2.89
LiveHPS [49]157.87/195.23156.5/147.31142.19/121.76191.87/189.300.64/0.70235.4/1369.89392.34/1506.5013.94/10.45279.96/358.542.1/6.73
LiDAR +RGBImmFusion [11]108.76/74.20473.18/464.83254.07/179.51533.5/529.710.17/0.14324.4/1446.01487.92/1532.8816.52/10.8627.03/14.491.94/3.99
FusionPose [14]112.08/86.44256.81/315.93198.55/193.83306.22/359.680.36/0.42275.02/1445.32444.47/1532.2816.29/10.8592.97/80.792.02/6.48
LEIR [73]110.18/94.57297.95/299.62187.26/150.56340.61/351.520.41/0.37266.82/1313.09282.31/1435.929.78/9.9773.38/85.031.1/1.20
Ours5.17/17.2575.45/88.9261.73/74.5094.89/106.420.91/0.7862.95/85.2678.99/106.951.57/3.128.3/27.751.07/1.29
", + "bbox": [ + 81, + 88, + 893, + 220 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/fb417e75282438d86db760f2ec8b761a1f8ec5ce1fda126065d88a1eacd198a6.jpg", + "table_caption": [ + "Table 3. HMR Comparison in the AscendMotion dataset (Horizontal Scene/Vertical Scene). The results for the horizontal and the vertical scene are separated by “/” in each cell. ClimbingCap performs significantly than others." + ], + "table_footnote": [], + "table_body": "
ModalityMethodCamera CoordinateWorld Coordinate
ACCEL↓MPJPE↓PA-MPJPE↓PVE↓PCK0.3↑WA-MPJPE↓W-MPJPE↓RTE↓Jitter↓T-Error↓
RGBTRACE [60]57.65488.6398.83685.290.17365.11608.3828.9827.304.58
SLAHMR [75]6.10228.8282.39201.420.61391.24445.3813.8812.632.47
WHAM [56]9.11184.4394.40191.270.76321.58430.3532.959.263.78
GVHMR [55]6.98152.5691.13140.190.60349.17476.9021.5411.632.26
LiDARLiDARCapV2 [78]70.72389.97267.76364.260.50468.00520.0613.5654.253.12
LiveHPS [49]72.57190.86148.00225.650.59412.22524.4712.1150.773.52
LiDAR+RGBImmFusion [11]68.82322.21232.44435.720.45465.46586.3613.5914.553.96
FusionPose [14]58.54242.20189.92330.090.53373.46437.8813.5448.302.01
LEIR [73]16.46206.69117.52269.120.58228.1370.7410.638.421.84
Ours5.1884.0360.6999.890.86228.28261.068.178.721.26
", + "bbox": [ + 81, + 270, + 892, + 431 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. HMR Comparison in the CIMI4D dataset. ClimbingCap demonstrates remarkable generalization and robustness on other climbing datasets.", + "bbox": [ + 75, + 440, + 892, + 468 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Comparison on Global Motion Recovery", + "text_level": 1, + "bbox": [ + 76, + 494, + 423, + 510 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To evaluate the performance of global HMR, we follow the motion evaluation protocols of [55, 56] and the global trajectory evaluation protocol of [15, 73]. Following WHAM [56], to calculate the world coordinate MPJPE metrics, the predicted global sequence is divided into segments of 100 frames, and each segment is aligned with the ground truth segment either to the entire segment (WA-MPJPE) or to the first two frames (W-MPJPE). Besides WA-MPJPE and W-MPJPE, the root translation error (RTE), motion jitter (Jitter, in $m / s^3$ ), relative global translation error (T-Error in $m$ ) of the entire sequence are reported.", + "bbox": [ + 73, + 518, + 468, + 684 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We evaluate ClimbingCap against nine methods with different modalities. The global RGB methods are TRACE [60], SLAHMR [75], WHAM [56], and GVHMR [55]. The LiDAR-based methods are LiDAR-CapV2 [78] and LiveHPS [49]. The LiDAR+RGB methods are ImmFusion [11], FusionPose [14], and LEIR [73].", + "bbox": [ + 75, + 685, + 468, + 776 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results in AscendMotion. The scenes in AscendMotion are categories into horizontal and vertical scenes, based on the major direction of human motions. Vertical motions are more challenging to human and to HMR methods than horizontal motions. The experimental results are depicted in Tab. 3. The top rows depicted the results for multiple state-of-the-art RGB-based HRM methods, and the bottom rows depicted the results for state-of-the-art LiDAR and Li-", + "bbox": [ + 75, + 779, + 470, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "DAR+RGB based methods. In each cell of the table, the results for the horizontal and the vertical scenes are separated by $l$ . For horizontal scenes, ClimbingCap performs the best in most of the metrics, especially in terms of MPJPE, WA-MPJPE and W-MPJPE. WHAM and GVHMR performs well for the camera coordinate metrics, but not in the world coordinate metrics. For the vertical scenes, ClimbingCap performs the best and significantly better than the second-best methods. GVHMR, a representative global HMR method, performs poorly in vertical scenes. It estimates the direction of movement by predicting the horizontal velocity. However, the major movement on the vertical scene of AscendMotion is an upward climbing movement. The LiDAR-based, and other LiDAR+RGB-based methods perform inferior to ClimbingCap. These methods do not fully consider global trajectories. The results demonstrate the importance of considering the relationship among camera coordinates and global coordinates.", + "bbox": [ + 496, + 494, + 893, + 767 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results in CIMI4D. To test the generalization ability of all the HMR methods, we evaluate their performance in the CIMI4D climbing dataset without retraining or fine-tuning. The CIMI4D dataset consists mostly horizontal scenes, and the motions are less challenging than AscendMotion. As it is reported in Tab. 4, ClimbingCap performs the best in the CIMI4D dataset for most the metrics, except the WA-MPJPE metric.", + "bbox": [ + 496, + 779, + 893, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "12318", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/78f3baf5b8fcbef56deeffe2cd3146b27bff70865b84098acd4ccbc897346237.jpg", + "image_caption": [ + "Figure 5. Qualitative Evaluation in the AscendMotion and CIMI4D dataset. The left and right areas show the results of Camera Coordinate and World Coordinate respectively. The red circles indicate obvious errors. The last row shows the results for CIMI4D dataset. Our method ClimbingCap performs best qualitatively by comparison." + ], + "image_footnote": [], + "bbox": [ + 80, + 89, + 547, + 409 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e06950a81260058782cbeb2c7d1a301de62fb9c1dab1d56f44a4089fd50ab2e5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 550, + 89, + 890, + 409 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/d95eb10522f16a18ba3de677580ac37fe5b50bcc0cb3189a776b0eea69773164.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
VariantMPIJEPA-MPJPEPCK0.3WA-MPJPEW-MPJPERTE
(1) RGB Input105.6763.050.78117.17174.537.64
(2) w/o LwD80.4652.150.8970.0491.232.02
(3) w/o LSDS99.1360.660.81109.35164.117.03
(4) w/o VLRR91.1061.850.8388.59120.113.34
(5) w/o SS77.4352.570.9065.3082.091.83
RGB+LiDAR Input Full Model75.4550.510.9162.9578.991.57
", + "bbox": [ + 81, + 484, + 465, + 575 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. Ablation Experiment for ClimbingCap in AscendMotion.", + "bbox": [ + 76, + 585, + 467, + 599 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.3. Ablation Experiment", + "text_level": 1, + "bbox": [ + 76, + 617, + 276, + 633 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To understand the impact of each global loss and the semi-supervised module on ClimbingCap's performance, we evaluate five variants of ClimbingCap on the AscendMotion dataset following the same training and evaluation protocol as the previous section. (1) RGB Input: Using only RGB imagery without LiDAR, this variant performs worst overall due to missing global position data required for post-processing. (2) w/o $\\mathcal{L}_{LWD}$ : Excluding the limb weight differentiation loss, which guides optimization using global point clouds, has a limited impact, indicating a beneficial but non-essential role. (3) w/o $\\mathcal{L}_{SDS}$ : Removing the velocity direction smoothing loss, which ensures motion consistency, significantly reduces global metrics, highlighting its importance. (4) w/o $\\mathcal{L}_{VLR}$ : The velocity direction smoothing loss constrains limb movements and corrects unexpected mis-predicted movements in a motion sequence; without this loss, ClimbingCap's global metrics drop sig", + "bbox": [ + 75, + 643, + 468, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "nificantly, highlighting its importance for consistency and accuracy. (5) w/o $SS$ : Removing the semi-supervised module results in a slight performance decrease, indicating its contribution to improving network training, though with a modest effect. The full ClimbingCap model, with RGB and LiDAR inputs, all loss terms, and the semi-supervised framework, achieves the best results in Tab. 5, validating the design's effectiveness for climbing motion analysis on AscendMotion.", + "bbox": [ + 496, + 488, + 892, + 625 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 642, + 617, + 657 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we propose AscendMotion, a large-scale and multi-modal climbing motion dataset, and ClimbingCap, a human motion recovery framework for climbing motions. AscendMotion provides a rich, high-quality dataset that surpasses previous climbing datasets in both scale and complexity. ClimbingCap effectively recovers 3D climbing motions in the global coordinate system, ensuring accurate pose estimation and global localization. Experimental results demonstrate the quality of AscendMotion dataset, and show that ClimbingCap outperforms existing methods in terms of accuracy and robustness.", + "bbox": [ + 496, + 669, + 890, + 835 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgment. This work was partially supported by the Fundamental Research Funds for the Central Universities (No. 20720230033); by PDL (2022-PDL-12); by Xiaomi Young Talents Program.", + "bbox": [ + 496, + 839, + 890, + 901 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "12319", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Thiemo Alldieck, Marc Kassubeck, Bastian Wandt, Bodo Rosenhahn, and Marcus Magnor. Optical flow-based 3d human motion estimation from monocular video. In German Conference on Pattern Recognition, pages 347-360. Springer, 2017. 1", + "[2] Marina Andric, Francesco Ricci, and Floriano Zini. Sensor-based activity recognition and performance assessment in climbing: A review. IEEE Access, 10:108583-108603, 2022. 3", + "[3] Raul Beltrán Beltrán, Julia Richter, Guido Köstermeyer, and Ulrich Heinkel. Climbing technique evaluation by means of skeleton video stream analysis. Sensors, 23(19):8216, 2023. 3", + "[4] Bharat Lal Bhatnagar, Cristian Sminchisescu, Christian Theobalt, and Gerard Pons-Moll. Combining implicit function learning and parametric models for 3d human reconstruction. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16, pages 311-329. Springer, 2020. 5", + "[5] Bharat Lal Bhatnagar, Xianghui Xie, Ilya A Petrov, Cristian Sminchisescu, Christian Theobalt, and Gerard Pons-Moll. Behave: Dataset and method for tracking human object interactions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15935-15946, 2022. 3", + "[6] Michael J Black, Priyanka Patel, Joachim Tesch, and Jinlong Yang. Bedlam: A synthetic dataset of bodies exhibiting detailed lifelike animated motion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8726-8737, 2023. 3", + "[7] Federica Bogo, Angjoo Kanazawa, Christoph Lassner, Peter V. Gehler, Javier Romero, and Michael J. Black. Keep it smpl: Automatic estimation of 3d human pose and shape from a single image. In ECCV, 2016. 2", + "[8] Christoph Bregler and Jitendra Malik. Tracking people with twists and exponential maps. Proceedings. 1998 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (Cat. No.98CB36231), pages 8-15, 1998. 2", + "[9] Zhe Cao, Tomas Simon, Shih-En Wei, and Yaser Sheikh. Realtime multi-person 2d pose estimation using part affinity fields. In Computer Vision and Pattern Recognition (CVPR), 2017. 1, 2, 3", + "[10] Kyungsik Cha, Eun-Young Lee, Myeong-Hyeon Heo, Kyu-Cheol Shin, Jonghee Son, and Dongho Kim. Analysis of climbing postures and movements in sport climbing for realistic 3d climbing animations. Procedia Engineering, 112: 52-57, 2015. 3", + "[11] Anjun Chen, Xiangyu Wang, Kun Shi, Shaohao Zhu, Bin Fang, Yingfeng Chen, Jiming Chen, Yuchi Huo, and Qi Ye. Immfusion: Robust mmwave-rgb fusion for 3d human body reconstruction in all weather conditions. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 2752-2758. IEEE, 2023. 7", + "[12] Xin Chen, Anqi Pang, Wei Yang, Yuexin Ma, Lan Xu, and Jingyi Yu. Sportscap: Monocular 3d human motion capture and fine-grained understanding in challenging sports videos." + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "International Journal of Computer Vision, 129:2846-2864, 2021. 2", + "[13] Hai Ci, Mingdong Wu, Wentao Zhu, Xiaoxuan Ma, Hao Dong, Fangwei Zhong, and Yizhou Wang. Gfpose: Learning 3d human pose prior with gradient fields. In CVPR, pages 4800-4810, 2023. 2", + "[14] Peishan Cong, Yiteng Xu, Yiming Ren, Juze Zhang, Lan Xu, Jingya Wang, Jingyi Yu, and Yuexin Ma. Weakly supervised 3d multi-person pose estimation for large-scale scenes based on monocular camera and single lidar. In AAAI, pages 461-469. AAAI Press, 2023. 7", + "[15] Yudi Dai, Yitai Lin, XiPing Lin, Chenglu Wen, Lan Xu, Hongwei Yi, Siqi Shen, Yuexin Ma, and Cheng Wang. Sloper4d: A scene-aware dataset for global 4d human pose estimation in urban environments. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 682-692, 2023. 2, 3, 7", + "[16] Yudi Dai, Zhiyong Wang, Xiping Lin, Chenglu Wen, Lan Xu, Siqi Shen, Yuexin Ma, and Cheng Wang. Hisc4d: Human-centered interaction and 4d scene capture in large-scale space using wearable imus and lidar. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 2, 3", + "[17] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In 9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7, 2021. OpenReview.net, 2021. 3", + "[18] Zhiyang Dou, Qingxuan Wu, Cheng Lin, Zeyu Cao, Qiangqiang Wu, Weilin Wan, Taku Komura, and Wenping Wang. TORE: token reduction for efficient human mesh recovery with transformer. In ICCV, 2023. 2", + "[19] Petr Elias, Veronika Skvarlova, and Pavel Zezula. Speed21: Speed climbing motion dataset. In Proceedings of the 4th International Workshop on Multimedia Content Analysis in Sports, pages 43-50, 2021. 2, 3", + "[20] Shubham Goel, Georgios Pavlakos, Jathushan Rajasegaran, Angjoo Kanazawa, and Jitendra Malik. Humans in 4d: Reconstructing and tracking humans with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14783-14794, 2023. 2", + "[21] Marc Habermann, Weipeng Xu, Michael Zollhöfer, Gerard Pons-Moll, and Christian Theobalt. Livecap: Real-time human performance capture from monocular video. ACM Transactions on Graphics (TOG), 38(2):14:1-14:17, 2019. 2", + "[22] Marc Habermann, Weipeng Xu, Michael Zollhofer, Gerard Pons-Moll, and Christian Theobalt. Deepcap: Monocular human performance capture using weak supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020.", + "[23] Yannan He, Anqi Pang, Xin Chen, Han Liang, Minye Wu, Yuexin Ma, and Lan Xu. Challencap: Monocular 3d capture of challenging human performances using multi-modal refer" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "12320", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ences. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11400-11411, 2021. 2", + "[24] Chun-Hao P Huang, Hongwei Yi, Markus Höschle, Matvey Safroshkin, Tsvetelina Alexiadis, Senya Polikovsky, Daniel Scharstein, and Michael J Black. Capturing and inferring dense full-body human-scene contact. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13274-13285, 2022. 2, 3", + "[25] Yinghao Huang, Manuel Kaufmann, Emre Aksan, Michael J. Black, Otmar Hilliges, and Gerard Pons-Moll. Deep inertial poser: Learning to reconstruct human pose from sparse inertial measurements in real time. ACM Transactions on Graphics, (Proc. SIGGRAPH Asia), 37(6):185:1-185:15, 2018. 2", + "[26] Hitomi Iguma, Akihiro Kawamura, and Ryo Kurazume. A new 3d motion and force measurement system for sport climbing. In 2020 IEEE/SICE International Symposium on System Integration, SII 2020, Honolulu, HI, USA, January 12-15, 2020, pages 1002-1007. IEEE, 2020. 2", + "[27] Angjoo Kanazawa, Michael J. Black, David W. Jacobs, and Jitendra Malik. End-to-end recovery of human shape and pose. In Computer Vision and Pattern Recognition (CVPR), 2018. 2", + "[28] Manuel Kaufmann, Jie Song, Chen Guo, Kaiyue Shen, Tianjian Jiang, Chengcheng Tang, Juan José Zárate, and Otmar Hilliges. Emdb: The electromagnetic database of global 3d human pose and shape in the wild. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14632-14643, 2023. 2, 3", + "[29] Diederik P Kingma and Jimmy Lei Ba. Adam: A method for stochastic gradient descent. In ICLR: international conference on learning representations, pages 1-15. ICLR US., 2015. 4", + "[30] Muhammed Kocabas, Nikos Athanasiou, and Michael J. Black. Vibe: Video inference for human body pose and shape estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2", + "[31] Muhammed Kocabas, Chun-Hao P. Huang, Otmar Hilliges, and Michael J. Black. Pare: Part attention regressor for 3d human body estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 11127-11137, 2021. 2", + "[32] Muhammed Kocabas, Ye Yuan, Pavlo Molchanov, Yunrong Guo, Michael J Black, Otmar Hilliges, Jan Kautz, and Umar Iqbal. Pace: Human and camera motion estimation from inthe-wild videos. In 2024 International Conference on 3D Vision (3DV), pages 397-408. IEEE, 2024. 2", + "[33] Nikos Kolotouros, Georgios Pavlakos, Michael J Black, and Kostas Daniilidis. Learning to reconstruct 3d human pose and shape via model-fitting in the loop. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2252-2261, 2019. 2", + "[34] Felix Kosmalla, André Zenner, Marco Speicher, Florian Daiber, Nico Herbig, and Antonio Krüger. Exploring rock climbing in mixed reality environments. In Proceedings of the 2017 CHI Conference Extended Abstracts on Human" + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Factors in Computing Systems, page 1787-1793, New York, NY, USA, 2017. Association for Computing Machinery. 2", + "[35] Felix Kosmalla, André Zenner, Corinna Tasch, Florian Daiber, and Antonio Krüger. The importance of virtual hands and feet for virtual reality climbing. In Extended Abstracts of the 2020 CHI Conference on Human Factors in Computing Systems, CHI 2020, Honolulu, HI, USA, April 25-30, 2020, pages 1-8. ACM, 2020. 2", + "[36] Jialian Li, Jingyi Zhang, Zhiyong Wang, Siqi Shen, Chenglu Wen, Yuexin Ma, Lan Xu, Jingyi Yu, and Cheng Wang. Lidarcap: Long-range marker-less 3d human motion capture with lidar point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20502-20512, 2022. 2", + "[37] Jiefeng Li, Ye Yuan, Davis Rempe, Haotian Zhang, Pavlo Molchanov, Cewu Lu, Jan Kautz, and Umar Iqbal. Coin: Control-inpainting diffusion prior for human and camera motion estimation. In European Conference on Computer Vision, pages 426-446. Springer, 2025. 2", + "[38] Zhihao Li, Jianzhuang Liu, Zhensong Zhang, Songcen Xu, and Youliang Yan. Cliff: Carrying location information in full frames into human pose and shape estimation. In ECCV, 2022. 2", + "[39] Yitai Lin, Zhijie Wei, Wanfa Zhang, Xiping Lin, Yudi Dai, Chenglu Wen, Siqi Shen, Lan Xu, and Cheng Wang. Hmpear: A dataset for human pose estimation and action recognition. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 2069-2078, 2024. 2, 3", + "[40] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J. Black. Smpl: A skinned multiperson linear model. ACM Trans. Graph., 34(6):248:1-248:16, 2015. 1, 5", + "[41] Naureen Mahmood, Nima Ghorbani, Nikolaus F. Troje, Gerard Pons-Moll, and Michael J. Black. Amass: Archive of motion capture as surface shapes. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2019. 4", + "[42] Julieta Martinez, Rayat Hossain, Javier Romero, and James J Little. A simple yet effective baseline for 3d human pose estimation. In ICCV, 2017. 1", + "[43] Dominik Pandurevic, Paweł Draga, Alexander Sutor, and Klaus Hochradel. Analysis of competition and training videos of speed climbing athletes using feature and human body keypoint detection algorithms. Sensors, 22(6):2251, 2022. 3", + "[44] Chaitanya Patel, Zhouyingcheng Liao, and Gerard Pons-Moll. Tailornet: Predicting clothing in 3d as a function of human pose, shape and garment style. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7363-7373, 2020. 2", + "[45] Priyanka Patel, Chun-Hao P. Huang, Joachim Tesch, David T. Hoffmann, Shashank Tripathi, and Michael J. Black. AGORA: Avatars in geography optimized for regression analysis. In Proceedings IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2021. 3", + "[46] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed A. A. Osman, Dimitrios Tzionas, and" + ], + "bbox": [ + 501, + 92, + 893, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "12321", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Michael J. Black. Expressive body capture: 3d hands, face, and body from a single image. In Proceedings IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 10975-10985, 2019. 2", + "[47] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems, 30, 2017. 3", + "[48] Davis Rempe, Tolga Birdal, Aaron Hertzmann, Jimei Yang, Srinath Sridhar, and Leonidas J. Guibas. Humor: 3d human motion model for robust pose estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 11488-11499, 2021. 2", + "[49] Yiming Ren, Xiao Han, Chengfeng Zhao, Jingya Wang, Lan Xu, Jingyi Yu, and Yuexin Ma. Livehps: Lidar-based scene-level human pose and shape estimation in free environment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1281-1291, 2024. 2, 3, 7", + "[50] Lionel Reveret, Sylvain Chapelle, Franck Quaine, and Pierre Legreneur. 3d visualization of body motion in speed climbing. Frontiers in Psychology, 11:2188, 2020. 3", + "[51] Julia Richter, Raul Beltrán Beltrán, Guido Köstermeyer, and Ulrich Heinkel. Human climbing and bouldering motion analysis: A survey on sensors, motion capture, analysis algorithms, recent advances and applications. VISIGRAPP (5: VISAPP), pages 751-758, 2020. 3", + "[52] Shunsuke Saito, Zeng Huang, Ryota Natsume, Shigeo Morishima, Hao Li, and Angjoo Kanazawa. Pifu: Pixel-aligned implicit function for high-resolution clothed human digitization. In 2019 IEEE/CVF International Conference on Computer Vision, ICCV 2019, Seoul, Korea (South), October 27 - November 2, 2019, pages 2304-2314. IEEE, 2019. 2", + "[53] Shunsuke Saito, Tomas Simon, Jason M. Saragih, and Hanbyul Joo. Pifuhd: Multi-level pixel-aligned implicit function for high-resolution 3d human digitization. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2020, Seattle, WA, USA, June 13-19, 2020, pages 81-90. Computer Vision Foundation / IEEE, 2020. 2", + "[54] Katsuhito Sasaki, Keisuke Shiro, and Jun Rekimoto. Exempliser: Predicting poses of experts as examples for beginners in climbing using a neural network. In Proceedings of the Augmented Humans International Conference, AHs 2020, Kaiserslautern, Germany, 16-17 March, 2020, pages 18:1-18:9. ACM, 2020. 2", + "[55] Zehong Shen, Huaijin Pi, Yan Xia, Zhi Cen, Sida Peng, Zechen Hu, Hujun Bao, Ruizhen Hu, and Xiaowei Zhou. World-grounded human motion recovery via gravity-view coordinates. In SIGGRAPH Asia Conference Proceedings, 2024. 2, 4, 7", + "[56] Soyong Shin, Juyong Kim, Eni Halilaj, and Michael J Black. Wham: Reconstructing world-grounded humans with accurate 3d motion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2070-2080, 2024. 2, 4, 7", + "[57] Zhuo Su, Lan Xu, Zerong Zheng, Tao Yu, Yebin Liu, and Lu" + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Fang. Robustfusion: Human volumetric capture with data-driven visual cues using a rgbd camera. In ECCV, 2020. 2", + "[58] Yu Sun, Qian Bao, Wu Liu, Yili Fu, Black Michael J., and Tao Mei. Monocular, One-stage, Regression of Multiple 3D People. In ICCV, 2021.", + "[59] Yu Sun, Wu Liu, Qian Bao, Yili Fu, Tao Mei, and Michael J Black. Putting People in their Place: Monocular Regression of 3D People in Depth. In CVPR, 2022. 2", + "[60] Yu Sun, Qian Bao, Wu Liu, Tao Mei, and Michael J Black. Trace: 5d temporal regression of avatars with dynamic cameras in 3d environments. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8856-8866, 2023. 2, 7", + "[61] Marcel Tiator, Christian Geiger, Bastian Dewitz, Ben Fischer, Laurin Gerhardt, David Nowottnik, and Hendrik Preu. Venga!: climbing in mixed reality. In Proceedings of the First Superhuman Sports Design Challenge: First International Symposium on Amplifying Capabilities and Competing in Mixed Realities, July 2-5, 2018, Delft, The Netherlands, pages 9:1-9:8. ACM, 2018. 2", + "[62] Daniel Vlasic, Rolf Adelsberger, Giovanni Vannucci, John C. Barnwell, Markus H. Gross, Wojciech Matusik, and Jovan Popovic. Practical motion capture in everyday surroundings. ACM SIGGRAPH 2007 papers, 2007. 2", + "[63] Ziniu Wan, Zhengjia Li, Maoqing Tian, Jianbo Liu, Shuai Yi, and Hongsheng Li. Encoder-decoder with multi-level attention for 3d human shape and pose estimation. In 2021 IEEE/CVF International Conference on Computer Vision, ICCV 2021, Montreal, QC, Canada, October 10-17, 2021, pages 13013-13022. IEEE, 2021. 2", + "[64] He Wang, Yezhen Cong, Or Litany, Yue Gao, and Leonidas J Guibas. 3dioumatch: Leveraging iou prediction for semi-supervised 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14615-14624, 2021. 4", + "[65] Kuan-Chieh Wang, Zhenzhen Weng, Maria Xenochristou, João Pedro Araújo, Jeffrey Gu, Karen Liu, and Serena Yeung. Nemo: Learning 3d neural motion fields from multiple video instances of the same action. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22129-22138, 2023. 2", + "[66] MeiJun Wang, Yu Meng, Zhongwei Qiu, Chao Zheng, Yan Xu, Jian Gao, et al. Pedestrian-centric 3d pre-collision pose and shape estimation from dashcam perspective. In The Thirty-eighth Annual Conference on Neural Information Processing Systems. 6", + "[67] Emily Whiting, Nada Ouf, Liane Makatura, Christos Mousas, Zhenyu Shu, and Ladislav Kavan. Environment-scale fabrication: Replicating outdoor climbing experiences. In Proceedings of the 2017 CHI Conference on Human Factors in Computing Systems, pages 1794-1804. ACM, 2017. 3", + "[68] HJ Wolring. New possibilities for human motion studies by real-time light spot position measurement. Biotelemetry, 1 (3):132, 1974. 2", + "[69] Qiming Xia, Wei Ye, Hai Wu, Shijia Zhao, Leyuan Xing, Xun Huang, Jinhao Deng, Xin Li, Chenglu Wen, and Cheng" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "12322", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Wang. Hinted: Hard instance enhanced detector with mixed-density feature fusion for sparsely-supervised 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15321-15330, 2024. 4", + "[70] Yuliang Xiu, Jinlong Yang, Dimitrios Tzionas, and Michael J. Black. ICON: implicit clothed humans obtained from normals. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 13286-13296. IEEE, 2022. 2", + "[71] Weipeng Xu, Avishek Chatterjee, Michael Zollhöfer, Helge Rhodin, Dushyant Mehta, Hans-Peter Seidel, and Christian Theobalt. Monoperfcap: Human performance capture from monocular video. ACM Transactions on Graphics (TOG), 37 (2):27:1-27:15, 2018. 2", + "[72] Ming Yan, Xin Wang, Yudi Dai, Siqi Shen, Chenglu Wen, Lan Xu, Yuexin Ma, and Cheng Wang. Cimi4d: A large multimodal climbing motion dataset under human-scene interactions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12977-12988, 2023. 2, 3", + "[73] Ming Yan, Yan Zhang, Shuqiang Cai, Shuqi Fan, Xincheng Lin, Yudi Dai, Siqi Shen, Chenglu Wen, Lan Xu, Yuexin Ma, et al. Reli11d: A comprehensive multimodal human motion dataset and method. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2250-2262, 2024. 2, 3, 7", + "[74] Ze Yang, Shenlong Wang, Siva Manivasagam, Zeng Huang, Wei-Chiu Ma, Xinchen Yan, Ersin Yumer, and Raquel Urtasun. S3: Neural shape, skeleton, and skinning fields for 3d human modeling. In CVPR, 2021. 2", + "[75] Vickie Ye, Georgios Pavlakos, Jitendra Malik, and Angjoo Kanazawa. Decoupling human and camera motion from videos in the wild. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2, 4, 7", + "[76] Xinyu Yi, Yuxiao Zhou, Marc Habermann, Soshi Shimada, Vladislav Golyanik, Christian Theobalt, and Feng Xu. Physical inertial poser (PIP): physics-aware real-time human motion tracking from sparse inertial sensors. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 13157-13168. IEEE, 2022. 2", + "[77] Ye Yuan, Umar Iqbal, Pavlo Molchanov, Kris Kitani, and Jan Kautz. Glamr: Global occlusion-aware human mesh recovery with dynamic cameras. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2", + "[78] Jingyi Zhang, Qihong Mao, Siqi Shen, Chenglu Wen, Lan Xu, and Cheng Wang. Lidarcapv2: 3d human pose estimation with human-object interaction from lidar point clouds. Pattern Recognition, 156:110848, 2024. 2, 3, 7", + "[79] Wu Zheng, Weiliang Tang, Li Jiang, and Chi-Wing Fu. Sessd: Self-ensembling single-stage object detector from point cloud. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14494–14503, 2021. 4", + "[80] Xiaowei Zhou, Menglong Zhu, Spyridon Leonardos, Konstantinos G Derpanis, and Kostas Daniilidis. Sparseness" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Meets Deepness: 3D Human Pose Estimation from Monocular Video. In Computer Vision and Pattern Recognition (CVPR), 2016. 1", + "[81] Wentao Zhu, Xiaoxuan Ma, Zhaoyang Liu, Libin Liu, Wayne Wu, and Yizhou Wang. Motionbert: Unified pretraining for human motion analysis. In ICCV, 2023. 2" + ], + "bbox": [ + 501, + 92, + 890, + 176 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12323", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2025/ClimbingCap_ Multi-Modal Dataset and Method for Rock Climbing in World Coordinate/56189f24-e199-449b-838e-e830c93475ae_model.json b/2025/ClimbingCap_ Multi-Modal Dataset and Method for Rock Climbing in World Coordinate/56189f24-e199-449b-838e-e830c93475ae_model.json new file mode 100644 index 0000000000000000000000000000000000000000..03112f5f021262d844a2a7eb0152eb52b590541e --- /dev/null +++ b/2025/ClimbingCap_ Multi-Modal Dataset and Method for Rock Climbing in World Coordinate/56189f24-e199-449b-838e-e830c93475ae_model.json @@ -0,0 +1,2556 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.143, + 0.131, + 0.83, + 0.175 + ], + "angle": 0, + "content": "ClimbingCap: Multi-Modal Dataset and Method for Rock Climbing in World Coordinate" + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.188, + 0.855, + 0.226 + ], + "angle": 0, + "content": "Ming Yan\\(^{1,2,3*}\\) Xincheng Lin\\(^{1,3*}\\) Yuhua Luo\\(^{1,3}\\) Shuqi Fan\\(^{1,3}\\) Yudi Dai\\(^{7}\\) Qixin Zhong\\(^{4}\\) Lincai Zhong\\(^{5}\\) Yuexin Ma\\(^{6}\\) Lan Xu\\(^{6}\\) Chenglu Wen\\(^{1,3}\\) Siqi Shen\\(^{1,3\\dagger}\\) Cheng Wang\\(^{1,3}\\)" + }, + { + "type": "text", + "bbox": [ + 0.14, + 0.228, + 0.83, + 0.247 + ], + "angle": 0, + "content": "\\(^{1}\\)Fujian Key Laboratory of Sensing and Computing for Smart Cities, Xiamen University" + }, + { + "type": "text", + "bbox": [ + 0.168, + 0.247, + 0.8, + 0.265 + ], + "angle": 0, + "content": "\\(^{2}\\)National Institute for Data Science in Health and Medicine, Xiamen University" + }, + { + "type": "text", + "bbox": [ + 0.182, + 0.265, + 0.788, + 0.282 + ], + "angle": 0, + "content": "3Key Laboratory of Multimedia Trusted Perception and Efficient Computing," + }, + { + "type": "text", + "bbox": [ + 0.192, + 0.283, + 0.779, + 0.299 + ], + "angle": 0, + "content": "Ministry of Education of China, School of Informatics, Xiamen University" + }, + { + "type": "text", + "bbox": [ + 0.207, + 0.3, + 0.765, + 0.317 + ], + "angle": 0, + "content": "4China National Climbing Team 5Ningbo Sports Work Training Team" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.318, + 0.71, + 0.335 + ], + "angle": 0, + "content": "\\(^{6}\\)ShanghaiTech University ETH AI Center, ETH Zürich" + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.342, + 0.889, + 0.539 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.542, + 0.893, + 0.611 + ], + "angle": 0, + "content": "Figure 1. Overview. To address the challenging problem of global climbing motion recovery, we collect the dataset AscendMotion, using LiDAR, RGB camera and Inertial Measurement Unit (IMU) motion capture system with accurate motion labels and global trajectories (the blue and orange human bodies in right side of the figure represent labeled motions, and the orange curve represents the motion trajectory in the world coordinate.). Meanwhile, we propose ClimbingCap, a global climbing motion capturing method in world coordinate. As shown in the left part of this figure, it uses both image and LiDAR point cloud to recover human motions." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.624, + 0.314, + 0.639 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.655, + 0.471, + 0.867 + ], + "angle": 0, + "content": "Human Motion Recovery (HMR) research mainly focuses on ground-based motions such as running. The study on capturing climbing motion, an off-ground motion, is sparse. This is partly due to the limited availability of climbing motion datasets, especially large-scale and challenging 3D labeled datasets. To address the insufficiency of climbing motion datasets, we collect AscendMotion, a large-scale well-annotated, and challenging climbing motion dataset. It consists of \\(412k\\) RGB, LiDAR frames, and IMU measurements, including the challenging climbing motions of 22 skilled climbing coaches across 12 different rock walls. Capturing the climbing motions is challenging as it requires precise recovery of not only the complex pose but also the global position of climbers. Although multi" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.625, + 0.893, + 0.807 + ], + "angle": 0, + "content": "ple global HMR methods have been proposed, they cannot faithfully capture climbing motions. To address the limitations of HMR methods for climbing, we propose ClimbingCap, a motion recovery method that reconstructs continuous 3D human climbing motion in a global coordinate system. One key insight is to use the RGB and LiDAR modalities to separately reconstruct motions in camera coordinates and global coordinates and to optimize them jointly. We demonstrate the quality of the AscendMotion dataset and present promising results from ClimbingCap. The AscendMotion dataset and source code release publicly at http://www.lidarhumanmotion.net/climbingcap/" + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.815, + 0.633, + 0.831 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.84, + 0.893, + 0.901 + ], + "angle": 0, + "content": "It is challenging for global Human Motion Recovery (HMR) [1, 9, 40, 42, 80] due to the complexity of human poses and dynamic interaction between humans and their environments. Moreover, global HMR must main" + }, + { + "type": "page_footnote", + "bbox": [ + 0.102, + 0.875, + 0.216, + 0.888 + ], + "angle": 0, + "content": "* Equal contribution." + }, + { + "type": "page_footnote", + "bbox": [ + 0.103, + 0.888, + 0.232, + 0.901 + ], + "angle": 0, + "content": "† Corresponding author." + }, + { + "type": "list", + "bbox": [ + 0.102, + 0.875, + 0.232, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "12312" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.212 + ], + "angle": 0, + "content": "tain consistency in world coordinates to ensure authenticity and physical feasibility. The research community has proposed various methods to estimate human poses from different sensors such as RGB images [21-23, 33, 71], LiDAR point clouds [36, 49, 78], and inertial measurement units (IMUs) [25, 76]. Most of the studies [8, 15, 16, 24, 28, 39, 49, 62, 68, 73, 78] focus on recovering ground-based motions such as running, interacting, and dancing." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.215, + 0.473, + 0.412 + ], + "angle": 0, + "content": "Unlike ground-based motions such as running, climbing is an activity performed off-ground, where climbers use hands and feet to ascend holds or walls. As an important sport and recreation, the climbing race has become an official event in the Olympic Games. Yet, research on climbing motion capture [26, 34, 35, 54, 61] is sparse, partly due to the limited availability of climbing datasets. There are only two publicly available climbing datasets: SPEEDD21 [19] and CIMI4D [72]. SPEEDD21 is a 2D dataset. Although CIMI4D is 3D, it consists of many trivial climbing motions from casual climbers. Moreover, the size of the two datasets is small. As a result, the research community lacks an in-depth understanding of challenging climbing motions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.413, + 0.473, + 0.58 + ], + "angle": 0, + "content": "We collect AscendMotion, a large-scale and challenging climbing motion multi-modal dataset, to address the data insufficiency. AscendMotion dataset includes the motions of 22 skilled climbers climbing 12 different rock walls. It includes rich modalities from hardware timesynchronized RGB, LiDAR point clouds, and Inertial Measurement Unit(IMU) MoCap system. Moreover, it consists of the global trajectories of climbers, which are important for understanding human motions in world coordinates. We also combine automatic annotation with manual refinement to ensure the accuracy of the motion annotation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.582, + 0.473, + 0.839 + ], + "angle": 0, + "content": "Not only does climbing motion challenge humans, but capturing climbing motion is also challenging for researchers. Capturing climbing motion requires precise estimation of the complex pose resulting from human-scene interactions, as well as the global localization of climbers within scenes. Existing research dedicated to climbing [8, 62, 68] uses standard 2D HMR (e.g. OpenPose [9]) for climbing motions, which is suboptimal. Most existing HMR methods [20, 37, 55, 56, 60, 75] focus on recovering ground-based human motions through RGB imagery. The estimation results of these methods cannot be rigidly transformed among camera coordinates and global coordinate systems, due to the inherent ambiguity between coordinate systems. Moreover, global HMR methods often accumulate errors during long-term sequence recovery, affecting the accuracy of global trajectories and postures, especially in climbing movements, as we show through experiments." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.473, + 0.901 + ], + "angle": 0, + "content": "To tackle the above challenges and fill the gap in the motion capture community for climbing motions, we introduce ClimbingCap, an HMR method that reconstructs continuous 3D human climbing motion in both camera and" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.287 + ], + "angle": 0, + "content": "global coordinate systems (see Fig.1). We adopt a trilogy in ClimbingCap: separate coordinate decoding, postprocessing, and semi-supervised training. For the separate coordinate decoding stage, ClimbingCap uses image and point cloud modality to estimate poses in the camera and the global coordinate systems, respectively. The postprocessing stage accurately ensures the consistency of motions between the two coordinate systems. In the semi-supervised training stage, a teacher-student training method is adopted to make use of easily obtainable unlabeled climbing data. Through these innovations, we can handle complex climbing movements effectively. To summarize, our main contributions include:" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.289, + 0.892, + 0.334 + ], + "angle": 0, + "content": "- We collect the dataset AscendMotion, which is more comprehensive than existing datasets. We demonstrate the quality of the dataset and its poses are challenging." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.334, + 0.892, + 0.365 + ], + "angle": 0, + "content": "- We propose ClimbingCap, which is a multimodality global HMR method for rock climbing." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.365, + 0.892, + 0.425 + ], + "angle": 0, + "content": "- We conduct extensive experiments on evaluate the ClimbingCap in multiple datasets with various state-of-the-art methods. The experimental results show ClimbingCap perform better than all the methods for climbing motions." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.289, + 0.892, + 0.425 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.44, + 0.642, + 0.456 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.469, + 0.895, + 0.577 + ], + "angle": 0, + "content": "Camera-Space HMR. Camera-space HMR methods [7, 12, 13, 18, 27, 30, 31, 38, 44, 46, 48, 52, 53, 57-59, 63, 65, 70, 74, 81] predict human motion in independent camera coordinate system for each frame of the input video. Although camera-space methods can predict accurate human poses, they cannot effectively reflect the orientation and trajectory of the human body in world coordinates." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.579, + 0.895, + 0.746 + ], + "angle": 0, + "content": "World-Grounded HMR. Recent studies [32, 37, 55, 56, 60, 75, 77] have achieved remarkable progress in world coordinates. Most of them are two-stage methods. The first stage estimates human motion in camera coordinates, and the second stage optimizes the motion in world coordinates with the scene information. Most global HMR methods focus on ground-based motions (e.g., walking) that are located on the ground. They cannot accurately model climbing motions, which is an off-ground motion. Our method, ClimbingCap, is a global HMR method developed for recovering rock climbing motions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.897, + 0.903 + ], + "angle": 0, + "content": "HMR Dataset. We have compared the AscendMotion dataset with multiple recent datasets in Table Tab. 1. Most of the datasets focus on ground-based human daily activities or sports. There exist only two public available climbing motion datasets. SPEED21 [19] collects 2D motions of climbers from sports videos. CIMI4D [72] contains 3D motions and global trajectories of climbers. Our climbing dataset, AscendMotion (344 minutes labeled data and 441 minutes unlabeled data, 412k frames), is significantly larger than both SPEED21 (21 minutes, 46k frames) and CIMI4D" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "12313" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.089, + 0.895, + 0.332 + ], + "angle": 0, + "content": "
DatasetSensor ModalitiesGlobal TrajectoryFrames3D SceneMotionReal/SyntheticSeqsSubjectsCareer
RGBMoCapLiDAR
SLOPER4D [15]✓IMU✓✓100k✓DaliyReal1512Normal
EMDB[28]✓EM-✓105k-DaliyReal8110Normal
BEHAVE [5]✓--✓15k-InteractionsReal-8Normal
RICH [24]✓--✓577k✓InteractionsReal14222Normal
AGORA [45]✓--✓106.7k✓DaliySynthetic---
BEDLAM [6]✓---1M✓DaliySynthetic---
FreeMotion [49]✓IMU✓✓578k-DaliyReal--Normal
HmPEAR [39]✓-✓-300k-DaliyReal6k25Normal
RELI11D [73]✓IMU✓✓239k✓SportReal4810Normal
HiSC4D [16]✓IMU✓✓36k✓SportReal8-Normal
LiDARHuman51M [78]✓IMU✓-374k-DailyReal5210Normal
CIMI4D [72]✓IMU✓✓180k✓ClimbingReal4212Normal
SPEED21 [19]✓--46kClimbingReal95--
AscendMotion(Ours)✓IMU✓✓412k✓(12)ClimbingReal22022Skilled
" + }, + { + "type": "table_caption", + "bbox": [ + 0.153, + 0.342, + 0.816, + 0.356 + ], + "angle": 0, + "content": "Table 1. Comparisons with related HMR datasets. The ”-” symbol indicates that it is not included in the dataset." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.376, + 0.47, + 0.437 + ], + "angle": 0, + "content": "(120 minutes, 180k frames). Moreover, the motions in our dataset are more challenging than those in CIMI4D, as they come from climbing coaches who are more skilled than the casual climbers in CIMI4D." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.441, + 0.471, + 0.577 + ], + "angle": 0, + "content": "Climbing Motion Recovery. In recent years, the increasing popularity of climbing, coupled with the inclusion of rock climbing races in the Olympic Games, has led to growing interest in the capture and analysis of climbing motions [2, 3, 10, 51, 67]. Research on climbing MoCap methods remains limited. [43] employs OpenPose [9] for climbing pose estimation. [50] performs RGB-based method for local pose and estimates climbing position through a marker-based method." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.595, + 0.294, + 0.613 + ], + "angle": 0, + "content": "3. Method: ClimbingCap" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.622, + 0.47, + 0.833 + ], + "angle": 0, + "content": "Climbing motion capture is challenging, involving poses with extreme limb extension and full-body exertion in camera coordinates. Moreover, it requires precise alignment with the rock wall in world coordinate as climbers are ascending. The ClimbingCap pipeline consists of three parts: separate coordinate decoding, post-processing, and semi-supervised training. An overview of the proposed pipeline is shown in Fig. 2. The separate coordinate decoding and post-processing parts take into account the unique challenges posed by climbing motion, which involves complex off-ground dynamics and interactions with scenes. The semi-supervised training part makes uses of the large-scale unlabeled climbing motion data to learn better a HMR model." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.846, + 0.351, + 0.863 + ], + "angle": 0, + "content": "3.1. Separate Coordinate Decoding" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.47, + 0.903 + ], + "angle": 0, + "content": "The separate coordinate decoding (SCD) stage extracts features from the RGB sequence and LiDAR point clouds, and" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.376, + 0.892, + 0.406 + ], + "angle": 0, + "content": "predicts the poses in camera coordinates and the positions in global coordinates." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.41, + 0.893, + 0.607 + ], + "angle": 0, + "content": "Input and Feature Extraction. The overall network structure is illustrated in Fig. 2. The input includes RGB images and point cloud data. First, the point cloud data is transformed from the world coordinate system to the camera coordinate system via an extrinsic matrix, represented as \\(\\mathcal{P}_{\\mathrm{c}} = \\Omega_{w2c} \\cdot \\mathcal{P}_{\\mathrm{w}}\\). Subsequently, the RGB images and transformed point cloud data are passed through feature extraction modules, RGB Extract and PC Extract, to obtain visual and geometric features. We build the feature extraction modules based on ViT [17] and PointNet++ [47]. These features are then fed into the following two decoder modules, which regress the SMPL parameters and global motion parameters of the human body, respectively." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.61, + 0.894, + 0.791 + ], + "angle": 0, + "content": "Camera Coordinate Decoder. This module decodes the SMPL parameters in the camera coordinate system. The RGB and point cloud features serve as inputs to the Camera Coordinate Decoder (denoted as \\(\\mathcal{T}_{\\mathrm{Decoder}}\\)), which processes the inputs with contextual information \\(\\mathbf{f}_{\\mathrm{backbone}}\\), generating an output token \\(\\mathbf{t}_{\\mathrm{out}}\\). This output token is then used to iteratively optimize the SMPL parameters, including the pose \\(\\theta\\), shape \\(\\beta\\), and camera translation \\(\\Delta c\\). The iterative decoding approach allows the model to gradually approximate the true pose and shape in the camera coordinate system. In each iteration, the decoder updates the current SMPL parameters \\(\\theta_{i}, \\beta_{i}\\), and \\(\\Delta c_{t}\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.605, + 0.801, + 0.891, + 0.817 + ], + "angle": 0, + "content": "\\[\n\\mathbf {t} _ {\\text {o u t}} = \\mathcal {T} _ {\\text {D e c o d e r}} (\\mathbf {t}, \\mathbf {f} _ {\\text {b a c k b o n e}}), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.825, + 0.894, + 0.901 + ], + "angle": 0, + "content": "where \\(\\theta_{i + 1} = \\Phi_{\\theta}\\cdot \\mathbf{t}_{\\mathrm{out}} + \\theta_{i},\\beta_{i + 1} = \\Phi_{\\beta}\\cdot \\mathbf{t}_{\\mathrm{out}} + \\beta_{i}\\) and \\(\\Delta c_{i + 1} = \\Phi_c\\cdot \\mathbf{t}_{\\mathrm{out}} + \\Delta c_i\\) are the update equations, with \\(\\Phi_{\\theta}\\), \\(\\Phi_{\\beta}\\), and \\(\\Phi_c\\) representing the respective weight matrices for each parameter. Here, the input token \\(\\mathbf{t}\\) can include initialized pose, shape, and camera parameters as needed." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "12314" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.09, + 0.538, + 0.294 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.542, + 0.089, + 0.891, + 0.294 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.307, + 0.893, + 0.337 + ], + "angle": 0, + "content": "Figure 2. Overview of ClimbingCap. The arrows indicate the three stages of the ClimbingCap framework: separate coordinate decoding (the green box), post-processing (the blue box), and semi-supervised training (the red box)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.355, + 0.47, + 0.462 + ], + "angle": 0, + "content": "Global Coordinate Decoder. To fully capture the human motion trajectory in the world coordinate, we design the Global Translation Decoder to predict the global translation parameters \\(\\Gamma^{\\mathrm{trans}}\\) of the human body. In this module, the decoder processes the features \\(\\mathbf{f}_{\\mathrm{backbone}}\\) as contextual input, iteratively updating the global translation parameters. The update formula in each iteration is given by:" + }, + { + "type": "equation", + "bbox": [ + 0.19, + 0.474, + 0.469, + 0.492 + ], + "angle": 0, + "content": "\\[\n\\Gamma_ {i + 1} ^ {\\text {t r a n s}} = \\Psi \\cdot \\mathbf {t} _ {\\text {o u t}} + \\Gamma_ {i} ^ {\\text {t r a n s}}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.507, + 0.47, + 0.569 + ], + "angle": 0, + "content": "where \\(\\Gamma_{i}^{\\mathrm{trans}}\\) represents the global translation parameters at time step \\(i\\), and \\(\\Psi\\) is the weight matrix for the update. This decoding process enables the model to capture a complete motion trajectory in the global coordinate system." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.572, + 0.47, + 0.739 + ], + "angle": 0, + "content": "Loss. The total loss function not only includes the 3D keypoint loss \\(\\mathcal{L}_{\\mathrm{kp3d}}\\) and 2D keypoint loss \\(\\mathcal{L}_{\\mathrm{kp2d}}\\) but also incorporates the SMPL parameter loss \\(\\mathcal{L}^{\\mathrm{spl}}\\) and the global trajectory loss \\(\\mathcal{L}_{\\mathrm{traj}}\\). Specifically, the 3D keypoint loss \\(\\mathcal{L}_{\\mathrm{kp3d}}\\) measures the 3D error of the predicted keypoints, while the 2D keypoint loss \\(\\mathcal{L}_{\\mathrm{kp2d}}\\) measures the 2D projection error. The SMPL parameter loss \\(\\mathcal{L}^{\\mathrm{spl}}\\) supervises the decoded pose and shape parameters, and the global trajectory loss \\(\\mathcal{L}_{\\mathrm{traj}}\\) constrains the translation parameters to within close distance of ground truth positions. Please see the supplementary for details. The final total loss is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.104, + 0.754, + 0.469, + 0.773 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {k p 3 d} + \\mathcal {L} _ {k p 2 d} + \\mathcal {L} _ {\\theta} ^ {s m p l} + \\mathcal {L} _ {\\beta} ^ {s m p l} + \\mathcal {L} _ {t r a j}. \\tag {3}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.801, + 0.235, + 0.818 + ], + "angle": 0, + "content": "3.2. Post-processing" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.469, + 0.902 + ], + "angle": 0, + "content": "Researches [55, 56, 75] have shown that a post-processing stage can be used to improve the output motion recovery results. Following these approaches, we employ a post-processing stage to optimize the output pose from SCD 3.1 stage." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.355, + 0.892, + 0.492 + ], + "angle": 0, + "content": "One distinct advantage of ClimbingCap is that the output results for the pose decoding stage can be rigidly transformed between the camera coordinate system and the world coordinate system. Thanks to the LiDAR modality, the point cloud contains 3D information in the world coordinate system. The poses obtained from the SCD stage are converted from the camera coordinate system to the world coordinate system through the inverse extrinsic matrix \\(\\Omega_{w2c}^{-1}\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.492, + 0.893, + 0.673 + ], + "angle": 0, + "content": "The post-processing stage consists of three losses: \\(\\mathcal{L}_{LWD}\\), \\(\\mathcal{L}_{SDS}\\), and \\(\\mathcal{L}_{VLR}\\). \\(\\mathcal{L}_{LWD}\\) assigns different weights to the vertices of different parts of the climbing human SMPL, and optimizes the position of the human body in the world coordinate system according to these weights. \\(\\mathcal{L}_{SDS}\\) optimizes joint positions by smoothing the changes in the direction of the human body's velocity on the scene (i.e., rock wall). \\(\\mathcal{L}_{VLR}\\) optimizes the difficult-to-estimate limb end poses by using the position of the point cloud in space. In this stage, the global poses are optimized through using the Adam [29] optimizer. Please refer to the supplementary for detailed definition of these losses." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.681, + 0.734, + 0.698 + ], + "angle": 0, + "content": "3.3. Semi-supervised Training" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.704, + 0.892, + 0.81 + ], + "angle": 0, + "content": "Compared to ground-based motions [41], the size of labeled climbing motions is small. Simply using labeled climbing motion data may not be enough to train a robust model. Different from labeled climbing motion data, collecting unlabeled climbing motion data is cheaper. The AscendMotion dataset contains more unlabeled data than label data. They can be used to further improve the HMR model." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.892, + 0.901 + ], + "angle": 0, + "content": "It has been shown by researches [64, 69, 79] from the object detection community that through using a teacher-student semi-supervised training framework, the performance of object detection models can be improved. We adopt such semi-supervised training frame for HMR. In this work, we refer the model trained after the SCD and the post" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12315" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.09, + 0.468, + 0.288 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.136, + 0.3, + 0.411, + 0.314 + ], + "angle": 0, + "content": "Figure 3. Dataset collection hardware system." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.332, + 0.473, + 0.484 + ], + "angle": 0, + "content": "processing stage as the teacher model (green box in Fig. 2). The student model (red box in Fig. 2) clones the parameters of the teacher model. During semi-supervised training, the teacher model estimate the pose labels from unlabeled motion data, and the pose label is used as pseudo-label to further train the student model. We show in the experiment section that via utilizing the semi-supervised training framework, the performance of HMR can be further improved. Please refer to the supplementary for the detailed training process." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.496, + 0.33, + 0.511 + ], + "angle": 0, + "content": "4. The AscendMotion Dataset" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.522, + 0.47, + 0.734 + ], + "angle": 0, + "content": "To advance research in Human Mesh Recovery (HMR) for climbing motion within a global coordinate system, we present the AscendMotion dataset. This dataset specifically captures complex multi-directional rock climbing motions on non-planar surfaces. It provides a unique opportunity for challenging HMR research. AscendMotion includes multi-modal motion data from skilled climbers performing various climbing styles, such as bouldering, speed climbing, and lead climbing, with synchronized recordings from IMU, RGB cameras, and LiDAR. The dataset comprises 344 minutes (6 hours) of labeled data and 441 minutes of unlabeled data, featuring 22 adult climbers across 12 real-world climbing scenes. All participants agreed to use their recorded data for scientific purposes." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.742, + 0.338, + 0.759 + ], + "angle": 0, + "content": "4.1. Hardware and Configuration" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.47, + 0.902 + ], + "angle": 0, + "content": "The AscendMotion data collection system integrates multiple sensors and captures motion data in both indoor and outdoor environments. As shown in Fig. 3, the LiDAR (OusterOS1, 128 beams) captures 3D dynamic point clouds at 20 frames per second (FPS), while the Hik 1080P RGB camera records RGB video at 20 FPS. For each climbing environment, we use the Trimble X7 3D laser scanner to reconstruct high-resolution RGB point cloud scenes, each containing approximately 80 million points. For the labeled" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.244 + ], + "angle": 0, + "content": "dataset, each climber wears an Xsens MVN inertial motion capture system with 17 IMUs, recording at 60 FPS. [4] and a handheld point cloud scanner to ensure accurate body shape representation. For the unlabeled dataset, climbers do not wear a MoCap outfit, and we record their climbing motions using a RGB camera and a LiDAR. A human motion is represented by \\( M = (T, \\theta, \\beta) \\), where \\( T \\) represents global translation, \\( \\theta \\) is the SMPL [40] pose parameters, and \\( \\beta \\) is the SMPL shape parameter. We obtain each climber's shape parameters \\( \\beta \\) using IPNet [4]." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.255, + 0.691, + 0.271 + ], + "angle": 0, + "content": "4.2. Annotation Pipeline" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.279, + 0.892, + 0.354 + ], + "angle": 0, + "content": "The pose \\(\\theta\\) and the translation \\(T\\) obtained from the IMU measurements may be inaccurate. Especially the translation \\(T\\) may significantly drift for long-duration capture. The annotation pipeline of AscendMotion is used to find the accurate translate \\(T\\) and pose \\(\\theta\\) as labels." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.355, + 0.892, + 0.476 + ], + "angle": 0, + "content": "The annotation pipeline of AscendMotion is depicted in Fig. 4. It consists of three parts: preprocessing, multistage global optimization, and manual annotation. The preprocessing part is used for time synchronization and spatial calibration. The multi-stage global optimization part is developed for automatic generation of annotation labels. The manual annotation part is used to further improve the quality of generated annotations." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.498, + 0.824, + 0.514 + ], + "angle": 0, + "content": "4.2.1 Multi-modal Data Preprocessing Stage" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.527, + 0.892, + 0.618 + ], + "angle": 0, + "content": "Time Synchronization. The time among RGB camera and LiDAR is synchronized via Precision Time Protocol (PTP). We employ CollShark Auto 66 unit as the master clock, and it sends PTP slave clocks to the RGB camera and LiDAR. The time of IMU MoCap is post-synchronized with the LiDAR and RGB through anchor frames." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.622, + 0.892, + 0.728 + ], + "angle": 0, + "content": "Calibration. First, the LiDAR point cloud are registered with high-precision scanned-scenes. Next, the coordinate of LiDAR is treated as the world coordinate. The IMU measurements are transformed into the world coordinate through a calibration matrix. Finally, we perform frame-level calibration among RGB, LiDAR and IMU. Please refer to the appendix for details." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.751, + 0.782, + 0.766 + ], + "angle": 0, + "content": "4.2.2 Multi-stage Global Optimization" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.776, + 0.892, + 0.867 + ], + "angle": 0, + "content": "AscendMotion uses the translation \\(T\\) and pose \\(\\theta\\) provided by the IMU MoCap as the initialization of annotation labels, and performs multi-stage global optimization. To achieve accurate and natural human motion data consistent with the scene, we apply two loss functions: the Global Refit Loss \\(\\mathcal{L}_{\\mathrm{GR}}\\) and the Scene Touch Loss \\(\\mathcal{L}_{\\mathrm{ST}}\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.871, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Global Refit Loss \\(\\mathcal{L}_{\\mathrm{GR}}\\). The global refit loss promotes precise global alignment between the SMPL model and the" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "12316" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.291, + 0.089, + 0.464, + 0.102 + ], + "angle": 0, + "content": "Multi-modal data preprocessing" + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.102, + 0.278, + 0.251 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.279, + 0.102, + 0.446, + 0.251 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.447, + 0.102, + 0.616, + 0.251 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.618, + 0.089, + 0.788, + 0.102 + ], + "angle": 0, + "content": "Multi-stage global optimization" + }, + { + "type": "image", + "bbox": [ + 0.618, + 0.102, + 0.785, + 0.251 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.79, + 0.09, + 0.882, + 0.102 + ], + "angle": 0, + "content": "Manual Repair" + }, + { + "type": "image", + "bbox": [ + 0.787, + 0.102, + 0.89, + 0.251 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.264, + 0.893, + 0.294 + ], + "angle": 0, + "content": "Figure 4. AscendMotion Annotation Pipeline. From left to right, this pipeline represents the three stages of dataset annotation: time-space synchronous input pre-processing(the blue box), multi-stage global optimization(the green box), and manual repair(the red box)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.319, + 0.473, + 0.531 + ], + "angle": 0, + "content": "point clouds. \\(\\mathcal{L}_{\\mathrm{GR}}\\) calculates the geometric discrepancy between the SMPL model vertices \\(V\\) and the human body point cloud \\(P\\) using a modified Chamfer distance. It applies different loss terms (\\(f(|v_{i} - p_{j}|^{2})\\) or \\(g(|v_{i} - p_{j}|^{2}))\\) to different body parts to make body regions, such as the torso and limbs, match the point clouds of the human. We calculate the Euclidean distance between a vertex \\(v_{i}\\in V\\) and a point \\(p_j\\in P\\) as \\(d_{v_i,p_j}\\). A loss term \\(f(|v_i - p_j|^2)\\) is applied to a vertex \\(v_{i}\\) if \\(d_{v_i,p_j}\\leq d_{\\mathrm{torso}}\\), and \\(g(|p_j - v_i|^2)\\) is applied to \\(v_{i}\\) if \\(d_{v_i,p_j}\\leq d_{\\mathrm{limb}}\\), where \\(d_{\\mathrm{torso}}\\) and \\(d_{\\mathrm{limb}}\\) are the distance thresholds for torso and limbs, respectively. During climbing, humans use limbs to climb up, the distance among limbs and scenes should smaller than the distance among torso and scenes, \\(d_{\\mathrm{limb}}\\leq d_{\\mathrm{torso}}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.536, + 0.47, + 0.732 + ], + "angle": 0, + "content": "Scene Touch Loss \\(\\mathcal{L}_{\\mathrm{ST}}\\). This loss prevents unrealistic intersections between the SMPL model and the scene mesh by measuring the penetration depth between the SMPL vertices \\(v_{i} \\in V\\) and the scene mesh vertices \\(q_{j} \\in Q\\). Given the scene mesh normal vectors \\(n_{j}\\), the penetration depth for each vertex is calculated as the dot product \\(\\eta(v_{i}) = (v_{i} - q_{j}) \\cdot n_{j}\\), where \\(q_{j}\\) is the closest mesh vertex to \\(v_{i}\\). If the penetration depth \\(\\eta(v_{i})\\) is negative, it indicates that the vertex \\(v_{i}\\) has penetrated the scene mesh, and this value contributes to the loss \\(\\mathcal{L}_{\\mathrm{ST}}\\). This loss encourages the SMPL model to avoid intersecting with the scene mesh, promoting more realistic interactions between the model and the environment." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.755, + 0.383, + 0.769 + ], + "angle": 0, + "content": "4.2.3 Manual Annotation and Verification" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.47, + 0.903 + ], + "angle": 0, + "content": "To further improve the dataset quality, we use the SMPL annotation tool [66] for enhanced labeling of sequences. We engage four observers to review the automatically annotated results. For any ambiguous limb movements observed in RGB views, we perform additional manual adjustments and mark these frames as key frames. Finally, we propagate poses from multiple key frames within each sequence to ensure global consistency in both smoothness and accuracy." + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.316, + 0.892, + 0.387 + ], + "angle": 0, + "content": "
Constraint termHorizontal SceneVertical Scene
\\( \\mathcal{L}_{ST} \\)\\( \\mathcal{L}_{GR} \\)ACCEL↓MPJPE↓PA-MPJPE↓ACCEL↓MPJPE↓PA-MPJPE↓
XX1.8337.5732.526.7951.5146.24
✓X1.6735.4531.813.7340.3635.37
X✓1.5929.0425.951.6629.2424.99
✓✓1.5428.2023.251.6029.0724.37
" + }, + { + "type": "table_caption", + "bbox": [ + 0.541, + 0.398, + 0.851, + 0.412 + ], + "angle": 0, + "content": "Table 2. Quantitative Evaluation of Dataset Quality." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.443, + 0.669, + 0.459 + ], + "angle": 0, + "content": "5. EXPERIMENTS" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.472, + 0.892, + 0.609 + ], + "angle": 0, + "content": "In this section, we demonstrate the high-quality of the AscendMotion optimization process in Sec. 5.1, show that the AscendMotion dataset is challenging to existing HMR methods, and evaluate ClimbingCap and multiple HMR methods in climbing motion datasets in Sec. 5.2. We demonstrate the promising performance of ClimbingCap and shows that each component of ClimbingCap is useful through ablation study in Sec. 5.3. Please refer to the detailed experimental setup and metrics in the supplementary." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.627, + 0.691, + 0.642 + ], + "angle": 0, + "content": "5.1. Datasets Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.653, + 0.892, + 0.728 + ], + "angle": 0, + "content": "To quantitatively evaluate the annotation quality of AscendMotion, we divide the scene into horizontal and vertical rock walls and evaluate the performance of the global optimization stage (in Sec. 4.2.2) by comparing the generated annotations with manual annotations." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.732, + 0.892, + 0.793 + ], + "angle": 0, + "content": "The evaluation metrics include widely used MPJPE, Procrustes Aligned MPJPE (PA-MPJPE), per vertex error (PVE), acceleration error (Accel, in units of \\( m / s^2 \\)). These metrics are used for evaluating HMR methods as well." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.892, + 0.901 + ], + "angle": 0, + "content": "To understand the impact of different constraints used in the combined optimization stage, we perform an ablation study on two different losses: \\(\\mathcal{L}_{ST}\\) and \\(\\mathcal{L}_{GR}\\). Tab. 2 shows the error metrics using different loss combinations for two scenes. The error metrics are small, demonstrating the effectiveness of the global optimization pipeline and the high quality of AscendMotion." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12317" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.089, + 0.894, + 0.221 + ], + "angle": 0, + "content": "
ModalityMethodCamera CoordinateWorld Coordinate
ACCEL↓MPJPE↓PA-MPJPE↓PVE↓PCK0.3↑WA-MPJPE↓W-MPJPE↓RTE↓Jitter↓T-Error↓
RGBTRACE [60]18.68/76.79875.56/577.6069.21/85.81951.52/619.930.06/0.09144.33/385.71254.38/703.3514.73/26.17115.96/521.402.56/6.62
SLAHMR [75]5.46/96.98232.46/467.8884.13/285.15283.24/552.630.36/0.12277.47/447.68804.85/613.603.64/39.394.91/201.332.81/6.54
WHAM [56]4.59/35.01110.92/143.1776.09/73.36124.2/164.910.76/0.62229.42/1125.77647.70/1499.855.16/9.043.58/40.691.77/2.49
GVHMR [55]4.50/26.22107.09/124.6060.06/80.30118.89/151.100.77/0.71105.15/1002.11202.45/1442.504.09/7.916.85/32.711.48/2.54
LiDARLiDARCapV2 [78]87.99/119.62244.6/234.52192.17/156.39326.45/283.270.53/0.50282.12/1396.42442.12/1518.2916.42/10.85176.95/165.551.65/2.89
LiveHPS [49]157.87/195.23156.5/147.31142.19/121.76191.87/189.300.64/0.70235.4/1369.89392.34/1506.5013.94/10.45279.96/358.542.1/6.73
LiDAR +RGBImmFusion [11]108.76/74.20473.18/464.83254.07/179.51533.5/529.710.17/0.14324.4/1446.01487.92/1532.8816.52/10.8627.03/14.491.94/3.99
FusionPose [14]112.08/86.44256.81/315.93198.55/193.83306.22/359.680.36/0.42275.02/1445.32444.47/1532.2816.29/10.8592.97/80.792.02/6.48
LEIR [73]110.18/94.57297.95/299.62187.26/150.56340.61/351.520.41/0.37266.82/1313.09282.31/1435.929.78/9.9773.38/85.031.1/1.20
Ours5.17/17.2575.45/88.9261.73/74.5094.89/106.420.91/0.7862.95/85.2678.99/106.951.57/3.128.3/27.751.07/1.29
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.229, + 0.893, + 0.259 + ], + "angle": 0, + "content": "Table 3. HMR Comparison in the AscendMotion dataset (Horizontal Scene/Vertical Scene). The results for the horizontal and the vertical scene are separated by “/” in each cell. ClimbingCap performs significantly than others." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.271, + 0.893, + 0.433 + ], + "angle": 0, + "content": "
ModalityMethodCamera CoordinateWorld Coordinate
ACCEL↓MPJPE↓PA-MPJPE↓PVE↓PCK0.3↑WA-MPJPE↓W-MPJPE↓RTE↓Jitter↓T-Error↓
RGBTRACE [60]57.65488.6398.83685.290.17365.11608.3828.9827.304.58
SLAHMR [75]6.10228.8282.39201.420.61391.24445.3813.8812.632.47
WHAM [56]9.11184.4394.40191.270.76321.58430.3532.959.263.78
GVHMR [55]6.98152.5691.13140.190.60349.17476.9021.5411.632.26
LiDARLiDARCapV2 [78]70.72389.97267.76364.260.50468.00520.0613.5654.253.12
LiveHPS [49]72.57190.86148.00225.650.59412.22524.4712.1150.773.52
LiDAR+RGBImmFusion [11]68.82322.21232.44435.720.45465.46586.3613.5914.553.96
FusionPose [14]58.54242.20189.92330.090.53373.46437.8813.5448.302.01
LEIR [73]16.46206.69117.52269.120.58228.1370.7410.638.421.84
Ours5.1884.0360.6999.890.86228.28261.068.178.721.26
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.441, + 0.893, + 0.469 + ], + "angle": 0, + "content": "Table 4. HMR Comparison in the CIMI4D dataset. ClimbingCap demonstrates remarkable generalization and robustness on other climbing datasets." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.495, + 0.424, + 0.511 + ], + "angle": 0, + "content": "5.2. Comparison on Global Motion Recovery" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.519, + 0.47, + 0.685 + ], + "angle": 0, + "content": "To evaluate the performance of global HMR, we follow the motion evaluation protocols of [55, 56] and the global trajectory evaluation protocol of [15, 73]. Following WHAM [56], to calculate the world coordinate MPJPE metrics, the predicted global sequence is divided into segments of 100 frames, and each segment is aligned with the ground truth segment either to the entire segment (WA-MPJPE) or to the first two frames (W-MPJPE). Besides WA-MPJPE and W-MPJPE, the root translation error (RTE), motion jitter (Jitter, in \\(m / s^3\\)), relative global translation error (T-Error in \\(m\\)) of the entire sequence are reported." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.686, + 0.47, + 0.777 + ], + "angle": 0, + "content": "We evaluate ClimbingCap against nine methods with different modalities. The global RGB methods are TRACE [60], SLAHMR [75], WHAM [56], and GVHMR [55]. The LiDAR-based methods are LiDAR-CapV2 [78] and LiveHPS [49]. The LiDAR+RGB methods are ImmFusion [11], FusionPose [14], and LEIR [73]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Results in AscendMotion. The scenes in AscendMotion are categories into horizontal and vertical scenes, based on the major direction of human motions. Vertical motions are more challenging to human and to HMR methods than horizontal motions. The experimental results are depicted in Tab. 3. The top rows depicted the results for multiple state-of-the-art RGB-based HRM methods, and the bottom rows depicted the results for state-of-the-art LiDAR and Li-" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.495, + 0.895, + 0.768 + ], + "angle": 0, + "content": "DAR+RGB based methods. In each cell of the table, the results for the horizontal and the vertical scenes are separated by \\( l \\). For horizontal scenes, ClimbingCap performs the best in most of the metrics, especially in terms of MPJPE, WA-MPJPE and W-MPJPE. WHAM and GVHMR performs well for the camera coordinate metrics, but not in the world coordinate metrics. For the vertical scenes, ClimbingCap performs the best and significantly better than the second-best methods. GVHMR, a representative global HMR method, performs poorly in vertical scenes. It estimates the direction of movement by predicting the horizontal velocity. However, the major movement on the vertical scene of AscendMotion is an upward climbing movement. The LiDAR-based, and other LiDAR+RGB-based methods perform inferior to ClimbingCap. These methods do not fully consider global trajectories. The results demonstrate the importance of considering the relationship among camera coordinates and global coordinates." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Results in CIMI4D. To test the generalization ability of all the HMR methods, we evaluate their performance in the CIMI4D climbing dataset without retraining or fine-tuning. The CIMI4D dataset consists mostly horizontal scenes, and the motions are less challenging than AscendMotion. As it is reported in Tab. 4, ClimbingCap performs the best in the CIMI4D dataset for most the metrics, except the WA-MPJPE metric." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "12318" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.09, + 0.548, + 0.41 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.552, + 0.09, + 0.891, + 0.41 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.42, + 0.893, + 0.464 + ], + "angle": 0, + "content": "Figure 5. Qualitative Evaluation in the AscendMotion and CIMI4D dataset. The left and right areas show the results of Camera Coordinate and World Coordinate respectively. The red circles indicate obvious errors. The last row shows the results for CIMI4D dataset. Our method ClimbingCap performs best qualitatively by comparison." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.486, + 0.466, + 0.576 + ], + "angle": 0, + "content": "
VariantMPIJEPA-MPJPEPCK0.3WA-MPJPEW-MPJPERTE
(1) RGB Input105.6763.050.78117.17174.537.64
(2) w/o LwD80.4652.150.8970.0491.232.02
(3) w/o LSDS99.1360.660.81109.35164.117.03
(4) w/o VLRR91.1061.850.8388.59120.113.34
(5) w/o SS77.4352.570.9065.3082.091.83
RGB+LiDAR Input Full Model75.4550.510.9162.9578.991.57
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.587, + 0.468, + 0.601 + ], + "angle": 0, + "content": "Table 5. Ablation Experiment for ClimbingCap in AscendMotion." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.618, + 0.277, + 0.635 + ], + "angle": 0, + "content": "5.3. Ablation Experiment" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.644, + 0.47, + 0.903 + ], + "angle": 0, + "content": "To understand the impact of each global loss and the semi-supervised module on ClimbingCap's performance, we evaluate five variants of ClimbingCap on the AscendMotion dataset following the same training and evaluation protocol as the previous section. (1) RGB Input: Using only RGB imagery without LiDAR, this variant performs worst overall due to missing global position data required for post-processing. (2) w/o \\(\\mathcal{L}_{LWD}\\): Excluding the limb weight differentiation loss, which guides optimization using global point clouds, has a limited impact, indicating a beneficial but non-essential role. (3) w/o \\(\\mathcal{L}_{SDS}\\): Removing the velocity direction smoothing loss, which ensures motion consistency, significantly reduces global metrics, highlighting its importance. (4) w/o \\(\\mathcal{L}_{VLR}\\): The velocity direction smoothing loss constrains limb movements and corrects unexpected mis-predicted movements in a motion sequence; without this loss, ClimbingCap's global metrics drop sig" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.489, + 0.893, + 0.625 + ], + "angle": 0, + "content": "nificantly, highlighting its importance for consistency and accuracy. (5) w/o \\( SS \\): Removing the semi-supervised module results in a slight performance decrease, indicating its contribution to improving network training, though with a modest effect. The full ClimbingCap model, with RGB and LiDAR inputs, all loss terms, and the semi-supervised framework, achieves the best results in Tab. 5, validating the design's effectiveness for climbing motion analysis on AscendMotion." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.643, + 0.619, + 0.658 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.67, + 0.892, + 0.836 + ], + "angle": 0, + "content": "In this work, we propose AscendMotion, a large-scale and multi-modal climbing motion dataset, and ClimbingCap, a human motion recovery framework for climbing motions. AscendMotion provides a rich, high-quality dataset that surpasses previous climbing datasets in both scale and complexity. ClimbingCap effectively recovers 3D climbing motions in the global coordinate system, ensuring accurate pose estimation and global localization. Experimental results demonstrate the quality of AscendMotion dataset, and show that ClimbingCap outperforms existing methods in terms of accuracy and robustness." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Acknowledgment. This work was partially supported by the Fundamental Research Funds for the Central Universities (No. 20720230033); by PDL (2022-PDL-12); by Xiaomi Young Talents Program." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12319" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.105 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.184 + ], + "angle": 0, + "content": "[1] Thiemo Alldieck, Marc Kassubeck, Bastian Wandt, Bodo Rosenhahn, and Marcus Magnor. Optical flow-based 3d human motion estimation from monocular video. In German Conference on Pattern Recognition, pages 347-360. Springer, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.185, + 0.472, + 0.239 + ], + "angle": 0, + "content": "[2] Marina Andric, Francesco Ricci, and Floriano Zini. Sensor-based activity recognition and performance assessment in climbing: A review. IEEE Access, 10:108583-108603, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.242, + 0.471, + 0.295 + ], + "angle": 0, + "content": "[3] Raul Beltrán Beltrán, Julia Richter, Guido Köstermeyer, and Ulrich Heinkel. Climbing technique evaluation by means of skeleton video stream analysis. Sensors, 23(19):8216, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.298, + 0.47, + 0.382 + ], + "angle": 0, + "content": "[4] Bharat Lal Bhatnagar, Cristian Sminchisescu, Christian Theobalt, and Gerard Pons-Moll. Combining implicit function learning and parametric models for 3d human reconstruction. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16, pages 311-329. Springer, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.382, + 0.47, + 0.464 + ], + "angle": 0, + "content": "[5] Bharat Lal Bhatnagar, Xianghui Xie, Ilya A Petrov, Cristian Sminchisescu, Christian Theobalt, and Gerard Pons-Moll. Behave: Dataset and method for tracking human object interactions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15935-15946, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.466, + 0.471, + 0.535 + ], + "angle": 0, + "content": "[6] Michael J Black, Priyanka Patel, Joachim Tesch, and Jinlong Yang. Bedlam: A synthetic dataset of bodies exhibiting detailed lifelike animated motion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8726-8737, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.536, + 0.47, + 0.592 + ], + "angle": 0, + "content": "[7] Federica Bogo, Angjoo Kanazawa, Christoph Lassner, Peter V. Gehler, Javier Romero, and Michael J. Black. Keep it smpl: Automatic estimation of 3d human pose and shape from a single image. In ECCV, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.592, + 0.47, + 0.648 + ], + "angle": 0, + "content": "[8] Christoph Bregler and Jitendra Malik. Tracking people with twists and exponential maps. Proceedings. 1998 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (Cat. No.98CB36231), pages 8-15, 1998. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.649, + 0.469, + 0.703 + ], + "angle": 0, + "content": "[9] Zhe Cao, Tomas Simon, Shih-En Wei, and Yaser Sheikh. Realtime multi-person 2d pose estimation using part affinity fields. In Computer Vision and Pattern Recognition (CVPR), 2017. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.705, + 0.47, + 0.773 + ], + "angle": 0, + "content": "[10] Kyungsik Cha, Eun-Young Lee, Myeong-Hyeon Heo, Kyu-Cheol Shin, Jonghee Son, and Dongho Kim. Analysis of climbing postures and movements in sport climbing for realistic 3d climbing animations. Procedia Engineering, 112: 52-57, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.775, + 0.469, + 0.858 + ], + "angle": 0, + "content": "[11] Anjun Chen, Xiangyu Wang, Kun Shi, Shaohao Zhu, Bin Fang, Yingfeng Chen, Jiming Chen, Yuchi Huo, and Qi Ye. Immfusion: Robust mmwave-rgb fusion for 3d human body reconstruction in all weather conditions. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 2752-2758. IEEE, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[12] Xin Chen, Anqi Pang, Wei Yang, Yuexin Ma, Lan Xu, and Jingyi Yu. Sportscap: Monocular 3d human motion capture and fine-grained understanding in challenging sports videos." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.119 + ], + "angle": 0, + "content": "International Journal of Computer Vision, 129:2846-2864, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.892, + 0.177 + ], + "angle": 0, + "content": "[13] Hai Ci, Mingdong Wu, Wentao Zhu, Xiaoxuan Ma, Hao Dong, Fangwei Zhong, and Yizhou Wang. Gfpose: Learning 3d human pose prior with gradient fields. In CVPR, pages 4800-4810, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.892, + 0.248 + ], + "angle": 0, + "content": "[14] Peishan Cong, Yiteng Xu, Yiming Ren, Juze Zhang, Lan Xu, Jingya Wang, Jingyi Yu, and Yuexin Ma. Weakly supervised 3d multi-person pose estimation for large-scale scenes based on monocular camera and single lidar. In AAAI, pages 461-469. AAAI Press, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.25, + 0.892, + 0.333 + ], + "angle": 0, + "content": "[15] Yudi Dai, Yitai Lin, XiPing Lin, Chenglu Wen, Lan Xu, Hongwei Yi, Siqi Shen, Yuexin Ma, and Cheng Wang. Sloper4d: A scene-aware dataset for global 4d human pose estimation in urban environments. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 682-692, 2023. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.334, + 0.892, + 0.417 + ], + "angle": 0, + "content": "[16] Yudi Dai, Zhiyong Wang, Xiping Lin, Chenglu Wen, Lan Xu, Siqi Shen, Yuexin Ma, and Cheng Wang. Hisc4d: Human-centered interaction and 4d scene capture in large-scale space using wearable imus and lidar. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.419, + 0.892, + 0.531 + ], + "angle": 0, + "content": "[17] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In 9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7, 2021. OpenReview.net, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.532, + 0.892, + 0.587 + ], + "angle": 0, + "content": "[18] Zhiyang Dou, Qingxuan Wu, Cheng Lin, Zeyu Cao, Qiangqiang Wu, Weilin Wan, Taku Komura, and Wenping Wang. TORE: token reduction for efficient human mesh recovery with transformer. In ICCV, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.589, + 0.892, + 0.645 + ], + "angle": 0, + "content": "[19] Petr Elias, Veronika Skvarlova, and Pavel Zezula. Speed21: Speed climbing motion dataset. In Proceedings of the 4th International Workshop on Multimedia Content Analysis in Sports, pages 43-50, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.646, + 0.892, + 0.716 + ], + "angle": 0, + "content": "[20] Shubham Goel, Georgios Pavlakos, Jathushan Rajasegaran, Angjoo Kanazawa, and Jitendra Malik. Humans in 4d: Reconstructing and tracking humans with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14783-14794, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.717, + 0.892, + 0.785 + ], + "angle": 0, + "content": "[21] Marc Habermann, Weipeng Xu, Michael Zollhöfer, Gerard Pons-Moll, and Christian Theobalt. Livecap: Real-time human performance capture from monocular video. ACM Transactions on Graphics (TOG), 38(2):14:1-14:17, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.788, + 0.892, + 0.858 + ], + "angle": 0, + "content": "[22] Marc Habermann, Weipeng Xu, Michael Zollhofer, Gerard Pons-Moll, and Christian Theobalt. Deepcap: Monocular human performance capture using weak supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.859, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[23] Yannan He, Anqi Pang, Xin Chen, Han Liang, Minye Wu, Yuexin Ma, and Lan Xu. Challencap: Monocular 3d capture of challenging human performances using multi-modal refer" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "12320" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.47, + 0.133 + ], + "angle": 0, + "content": "ences. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11400-11411, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.472, + 0.219 + ], + "angle": 0, + "content": "[24] Chun-Hao P Huang, Hongwei Yi, Markus Höschle, Matvey Safroshkin, Tsvetelina Alexiadis, Senya Polikovsky, Daniel Scharstein, and Michael J Black. Capturing and inferring dense full-body human-scene contact. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13274-13285, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.221, + 0.47, + 0.29 + ], + "angle": 0, + "content": "[25] Yinghao Huang, Manuel Kaufmann, Emre Aksan, Michael J. Black, Otmar Hilliges, and Gerard Pons-Moll. Deep inertial poser: Learning to reconstruct human pose from sparse inertial measurements in real time. ACM Transactions on Graphics, (Proc. SIGGRAPH Asia), 37(6):185:1-185:15, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.292, + 0.47, + 0.361 + ], + "angle": 0, + "content": "[26] Hitomi Iguma, Akihiro Kawamura, and Ryo Kurazume. A new 3d motion and force measurement system for sport climbing. In 2020 IEEE/SICE International Symposium on System Integration, SII 2020, Honolulu, HI, USA, January 12-15, 2020, pages 1002-1007. IEEE, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.363, + 0.469, + 0.417 + ], + "angle": 0, + "content": "[27] Angjoo Kanazawa, Michael J. Black, David W. Jacobs, and Jitendra Malik. End-to-end recovery of human shape and pose. In Computer Vision and Pattern Recognition (CVPR), 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.42, + 0.469, + 0.502 + ], + "angle": 0, + "content": "[28] Manuel Kaufmann, Jie Song, Chen Guo, Kaiyue Shen, Tianjian Jiang, Chengcheng Tang, Juan José Zárate, and Otmar Hilliges. Emdb: The electromagnetic database of global 3d human pose and shape in the wild. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14632-14643, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.505, + 0.469, + 0.558 + ], + "angle": 0, + "content": "[29] Diederik P Kingma and Jimmy Lei Ba. Adam: A method for stochastic gradient descent. In ICLR: international conference on learning representations, pages 1-15. ICLR US., 2015. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.562, + 0.469, + 0.63 + ], + "angle": 0, + "content": "[30] Muhammed Kocabas, Nikos Athanasiou, and Michael J. Black. Vibe: Video inference for human body pose and shape estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.633, + 0.469, + 0.701 + ], + "angle": 0, + "content": "[31] Muhammed Kocabas, Chun-Hao P. Huang, Otmar Hilliges, and Michael J. Black. Pare: Part attention regressor for 3d human body estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 11127-11137, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.704, + 0.469, + 0.773 + ], + "angle": 0, + "content": "[32] Muhammed Kocabas, Ye Yuan, Pavlo Molchanov, Yunrong Guo, Michael J Black, Otmar Hilliges, Jan Kautz, and Umar Iqbal. Pace: Human and camera motion estimation from inthe-wild videos. In 2024 International Conference on 3D Vision (3DV), pages 397-408. IEEE, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.776, + 0.469, + 0.844 + ], + "angle": 0, + "content": "[33] Nikos Kolotouros, Georgios Pavlakos, Michael J Black, and Kostas Daniilidis. Learning to reconstruct 3d human pose and shape via model-fitting in the loop. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2252-2261, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[34] Felix Kosmalla, André Zenner, Marco Speicher, Florian Daiber, Nico Herbig, and Antonio Krüger. Exploring rock climbing in mixed reality environments. In Proceedings of the 2017 CHI Conference Extended Abstracts on Human" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.121 + ], + "angle": 0, + "content": "Factors in Computing Systems, page 1787-1793, New York, NY, USA, 2017. Association for Computing Machinery. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.123, + 0.894, + 0.205 + ], + "angle": 0, + "content": "[35] Felix Kosmalla, André Zenner, Corinna Tasch, Florian Daiber, and Antonio Krüger. The importance of virtual hands and feet for virtual reality climbing. In Extended Abstracts of the 2020 CHI Conference on Human Factors in Computing Systems, CHI 2020, Honolulu, HI, USA, April 25-30, 2020, pages 1-8. ACM, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.207, + 0.892, + 0.29 + ], + "angle": 0, + "content": "[36] Jialian Li, Jingyi Zhang, Zhiyong Wang, Siqi Shen, Chenglu Wen, Yuexin Ma, Lan Xu, Jingyi Yu, and Cheng Wang. Lidarcap: Long-range marker-less 3d human motion capture with lidar point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20502-20512, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.292, + 0.892, + 0.36 + ], + "angle": 0, + "content": "[37] Jiefeng Li, Ye Yuan, Davis Rempe, Haotian Zhang, Pavlo Molchanov, Cewu Lu, Jan Kautz, and Umar Iqbal. Coin: Control-inpainting diffusion prior for human and camera motion estimation. In European Conference on Computer Vision, pages 426-446. Springer, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.362, + 0.892, + 0.416 + ], + "angle": 0, + "content": "[38] Zhihao Li, Jianzhuang Liu, Zhensong Zhang, Songcen Xu, and Youliang Yan. Cliff: Carrying location information in full frames into human pose and shape estimation. In ECCV, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.42, + 0.892, + 0.489 + ], + "angle": 0, + "content": "[39] Yitai Lin, Zhijie Wei, Wanfa Zhang, Xiping Lin, Yudi Dai, Chenglu Wen, Siqi Shen, Lan Xu, and Cheng Wang. Hmpear: A dataset for human pose estimation and action recognition. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 2069-2078, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.49, + 0.892, + 0.544 + ], + "angle": 0, + "content": "[40] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J. Black. Smpl: A skinned multiperson linear model. ACM Trans. Graph., 34(6):248:1-248:16, 2015. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.547, + 0.892, + 0.616 + ], + "angle": 0, + "content": "[41] Naureen Mahmood, Nima Ghorbani, Nikolaus F. Troje, Gerard Pons-Moll, and Michael J. Black. Amass: Archive of motion capture as surface shapes. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.618, + 0.892, + 0.657 + ], + "angle": 0, + "content": "[42] Julieta Martinez, Rayat Hossain, Javier Romero, and James J Little. A simple yet effective baseline for 3d human pose estimation. In ICCV, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.661, + 0.892, + 0.729 + ], + "angle": 0, + "content": "[43] Dominik Pandurevic, Paweł Draga, Alexander Sutor, and Klaus Hochradel. Analysis of competition and training videos of speed climbing athletes using feature and human body keypoint detection algorithms. Sensors, 22(6):2251, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.732, + 0.892, + 0.8 + ], + "angle": 0, + "content": "[44] Chaitanya Patel, Zhouyingcheng Liao, and Gerard Pons-Moll. Tailornet: Predicting clothing in 3d as a function of human pose, shape and garment style. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7363-7373, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.803, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[45] Priyanka Patel, Chun-Hao P. Huang, Joachim Tesch, David T. Hoffmann, Shashank Tripathi, and Michael J. Black. AGORA: Avatars in geography optimized for regression analysis. In Proceedings IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.874, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[46] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed A. A. Osman, Dimitrios Tzionas, and" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "12321" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.469, + 0.147 + ], + "angle": 0, + "content": "Michael J. Black. Expressive body capture: 3d hands, face, and body from a single image. In Proceedings IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 10975-10985, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.15, + 0.47, + 0.207 + ], + "angle": 0, + "content": "[47] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems, 30, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.208, + 0.471, + 0.277 + ], + "angle": 0, + "content": "[48] Davis Rempe, Tolga Birdal, Aaron Hertzmann, Jimei Yang, Srinath Sridhar, and Leonidas J. Guibas. Humor: 3d human motion model for robust pose estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 11488-11499, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.278, + 0.469, + 0.36 + ], + "angle": 0, + "content": "[49] Yiming Ren, Xiao Han, Chengfeng Zhao, Jingya Wang, Lan Xu, Jingyi Yu, and Yuexin Ma. Livehps: Lidar-based scene-level human pose and shape estimation in free environment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1281-1291, 2024. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.362, + 0.469, + 0.404 + ], + "angle": 0, + "content": "[50] Lionel Reveret, Sylvain Chapelle, Franck Quaine, and Pierre Legreneur. 3d visualization of body motion in speed climbing. Frontiers in Psychology, 11:2188, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.406, + 0.469, + 0.476 + ], + "angle": 0, + "content": "[51] Julia Richter, Raul Beltrán Beltrán, Guido Köstermeyer, and Ulrich Heinkel. Human climbing and bouldering motion analysis: A survey on sensors, motion capture, analysis algorithms, recent advances and applications. VISIGRAPP (5: VISAPP), pages 751-758, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.478, + 0.469, + 0.56 + ], + "angle": 0, + "content": "[52] Shunsuke Saito, Zeng Huang, Ryota Natsume, Shigeo Morishima, Hao Li, and Angjoo Kanazawa. Pifu: Pixel-aligned implicit function for high-resolution clothed human digitization. In 2019 IEEE/CVF International Conference on Computer Vision, ICCV 2019, Seoul, Korea (South), October 27 - November 2, 2019, pages 2304-2314. IEEE, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.562, + 0.469, + 0.657 + ], + "angle": 0, + "content": "[53] Shunsuke Saito, Tomas Simon, Jason M. Saragih, and Hanbyul Joo. Pifuhd: Multi-level pixel-aligned implicit function for high-resolution 3d human digitization. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2020, Seattle, WA, USA, June 13-19, 2020, pages 81-90. Computer Vision Foundation / IEEE, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.66, + 0.469, + 0.743 + ], + "angle": 0, + "content": "[54] Katsuhito Sasaki, Keisuke Shiro, and Jun Rekimoto. Exempliser: Predicting poses of experts as examples for beginners in climbing using a neural network. In Proceedings of the Augmented Humans International Conference, AHs 2020, Kaiserslautern, Germany, 16-17 March, 2020, pages 18:1-18:9. ACM, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.745, + 0.469, + 0.814 + ], + "angle": 0, + "content": "[55] Zehong Shen, Huaijin Pi, Yan Xia, Zhi Cen, Sida Peng, Zechen Hu, Hujun Bao, Ruizhen Hu, and Xiaowei Zhou. World-grounded human motion recovery via gravity-view coordinates. In SIGGRAPH Asia Conference Proceedings, 2024. 2, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.816, + 0.469, + 0.884 + ], + "angle": 0, + "content": "[56] Soyong Shin, Juyong Kim, Eni Halilaj, and Michael J Black. Wham: Reconstructing world-grounded humans with accurate 3d motion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2070-2080, 2024. 2, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.887, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[57] Zhuo Su, Lan Xu, Zerong Zheng, Tao Yu, Yebin Liu, and Lu" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "Fang. Robustfusion: Human volumetric capture with data-driven visual cues using a rgbd camera. In ECCV, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.892, + 0.163 + ], + "angle": 0, + "content": "[58] Yu Sun, Qian Bao, Wu Liu, Yili Fu, Black Michael J., and Tao Mei. Monocular, One-stage, Regression of Multiple 3D People. In ICCV, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.165, + 0.892, + 0.207 + ], + "angle": 0, + "content": "[59] Yu Sun, Wu Liu, Qian Bao, Yili Fu, Tao Mei, and Michael J Black. Putting People in their Place: Monocular Regression of 3D People in Depth. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.208, + 0.892, + 0.277 + ], + "angle": 0, + "content": "[60] Yu Sun, Qian Bao, Wu Liu, Tao Mei, and Michael J Black. Trace: 5d temporal regression of avatars with dynamic cameras in 3d environments. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8856-8866, 2023. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.279, + 0.892, + 0.376 + ], + "angle": 0, + "content": "[61] Marcel Tiator, Christian Geiger, Bastian Dewitz, Ben Fischer, Laurin Gerhardt, David Nowottnik, and Hendrik Preu. Venga!: climbing in mixed reality. In Proceedings of the First Superhuman Sports Design Challenge: First International Symposium on Amplifying Capabilities and Competing in Mixed Realities, July 2-5, 2018, Delft, The Netherlands, pages 9:1-9:8. ACM, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.377, + 0.892, + 0.434 + ], + "angle": 0, + "content": "[62] Daniel Vlasic, Rolf Adelsberger, Giovanni Vannucci, John C. Barnwell, Markus H. Gross, Wojciech Matusik, and Jovan Popovic. Practical motion capture in everyday surroundings. ACM SIGGRAPH 2007 papers, 2007. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.435, + 0.892, + 0.517 + ], + "angle": 0, + "content": "[63] Ziniu Wan, Zhengjia Li, Maoqing Tian, Jianbo Liu, Shuai Yi, and Hongsheng Li. Encoder-decoder with multi-level attention for 3d human shape and pose estimation. In 2021 IEEE/CVF International Conference on Computer Vision, ICCV 2021, Montreal, QC, Canada, October 10-17, 2021, pages 13013-13022. IEEE, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.519, + 0.892, + 0.589 + ], + "angle": 0, + "content": "[64] He Wang, Yezhen Cong, Or Litany, Yue Gao, and Leonidas J Guibas. 3dioumatch: Leveraging iou prediction for semi-supervised 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14615-14624, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.591, + 0.892, + 0.673 + ], + "angle": 0, + "content": "[65] Kuan-Chieh Wang, Zhenzhen Weng, Maria Xenochristou, João Pedro Araújo, Jeffrey Gu, Karen Liu, and Serena Yeung. Nemo: Learning 3d neural motion fields from multiple video instances of the same action. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22129-22138, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.675, + 0.892, + 0.743 + ], + "angle": 0, + "content": "[66] MeiJun Wang, Yu Meng, Zhongwei Qiu, Chao Zheng, Yan Xu, Jian Gao, et al. Pedestrian-centric 3d pre-collision pose and shape estimation from dashcam perspective. In The Thirty-eighth Annual Conference on Neural Information Processing Systems. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.745, + 0.892, + 0.828 + ], + "angle": 0, + "content": "[67] Emily Whiting, Nada Ouf, Liane Makatura, Christos Mousas, Zhenyu Shu, and Ladislav Kavan. Environment-scale fabrication: Replicating outdoor climbing experiences. In Proceedings of the 2017 CHI Conference on Human Factors in Computing Systems, pages 1794-1804. ACM, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.83, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[68] HJ Wolring. New possibilities for human motion studies by real-time light spot position measurement. Biotelemetry, 1 (3):132, 1974. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.873, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[69] Qiming Xia, Wei Ye, Hai Wu, Shijia Zhao, Leyuan Xing, Xun Huang, Jinhao Deng, Xin Li, Chenglu Wen, and Cheng" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "12322" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "Wang. Hinted: Hard instance enhanced detector with mixed-density feature fusion for sparsely-supervised 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15321-15330, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.163, + 0.47, + 0.232 + ], + "angle": 0, + "content": "[70] Yuliang Xiu, Jinlong Yang, Dimitrios Tzionas, and Michael J. Black. ICON: implicit clothed humans obtained from normals. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 13286-13296. IEEE, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.233, + 0.47, + 0.3 + ], + "angle": 0, + "content": "[71] Weipeng Xu, Avishek Chatterjee, Michael Zollhöfer, Helge Rhodin, Dushyant Mehta, Hans-Peter Seidel, and Christian Theobalt. Monoperfcap: Human performance capture from monocular video. ACM Transactions on Graphics (TOG), 37 (2):27:1-27:15, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.302, + 0.47, + 0.383 + ], + "angle": 0, + "content": "[72] Ming Yan, Xin Wang, Yudi Dai, Siqi Shen, Chenglu Wen, Lan Xu, Yuexin Ma, and Cheng Wang. Cimi4d: A large multimodal climbing motion dataset under human-scene interactions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12977-12988, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.385, + 0.47, + 0.468 + ], + "angle": 0, + "content": "[73] Ming Yan, Yan Zhang, Shuqiang Cai, Shuqi Fan, Xincheng Lin, Yudi Dai, Siqi Shen, Chenglu Wen, Lan Xu, Yuexin Ma, et al. Reli11d: A comprehensive multimodal human motion dataset and method. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2250-2262, 2024. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.469, + 0.47, + 0.524 + ], + "angle": 0, + "content": "[74] Ze Yang, Shenlong Wang, Siva Manivasagam, Zeng Huang, Wei-Chiu Ma, Xinchen Yan, Ersin Yumer, and Raquel Urtasun. S3: Neural shape, skeleton, and skinning fields for 3d human modeling. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.525, + 0.47, + 0.58 + ], + "angle": 0, + "content": "[75] Vickie Ye, Georgios Pavlakos, Jitendra Malik, and Angjoo Kanazawa. Decoupling human and camera motion from videos in the wild. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.581, + 0.47, + 0.677 + ], + "angle": 0, + "content": "[76] Xinyu Yi, Yuxiao Zhou, Marc Habermann, Soshi Shimada, Vladislav Golyanik, Christian Theobalt, and Feng Xu. Physical inertial poser (PIP): physics-aware real-time human motion tracking from sparse inertial sensors. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 13157-13168. IEEE, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.679, + 0.47, + 0.746 + ], + "angle": 0, + "content": "[77] Ye Yuan, Umar Iqbal, Pavlo Molchanov, Kris Kitani, and Jan Kautz. Glamr: Global occlusion-aware human mesh recovery with dynamic cameras. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.748, + 0.47, + 0.803 + ], + "angle": 0, + "content": "[78] Jingyi Zhang, Qihong Mao, Siqi Shen, Chenglu Wen, Lan Xu, and Cheng Wang. Lidarcapv2: 3d human pose estimation with human-object interaction from lidar point clouds. Pattern Recognition, 156:110848, 2024. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.804, + 0.47, + 0.871 + ], + "angle": 0, + "content": "[79] Wu Zheng, Weiliang Tang, Li Jiang, and Chi-Wing Fu. Sessd: Self-ensembling single-stage object detector from point cloud. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14494–14503, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[80] Xiaowei Zhou, Menglong Zhu, Spyridon Leonardos, Konstantinos G Derpanis, and Kostas Daniilidis. Sparseness" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "Meets Deepness: 3D Human Pose Estimation from Monocular Video. In Computer Vision and Pattern Recognition (CVPR), 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.177 + ], + "angle": 0, + "content": "[81] Wentao Zhu, Xiaoxuan Ma, Zhaoyang Liu, Libin Liu, Wayne Wu, and Yizhou Wang. Motionbert: Unified pretraining for human motion analysis. In ICCV, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.177 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12323" + } + ] +] \ No newline at end of file diff --git a/2025/ClimbingCap_ Multi-Modal Dataset and Method for Rock Climbing in World Coordinate/56189f24-e199-449b-838e-e830c93475ae_origin.pdf b/2025/ClimbingCap_ Multi-Modal Dataset and Method for Rock Climbing in World Coordinate/56189f24-e199-449b-838e-e830c93475ae_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..6f3d5220dab6a9dc0ce6dbac6cbfbd256b3e8aac --- /dev/null +++ b/2025/ClimbingCap_ Multi-Modal Dataset and Method for Rock Climbing in World Coordinate/56189f24-e199-449b-838e-e830c93475ae_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f1cdad3de05c7f231e12cbbc638c6747618240e96bd824369d7134b69f768ff0 +size 3106933 diff --git a/2025/ClimbingCap_ Multi-Modal Dataset and Method for Rock Climbing in World Coordinate/full.md b/2025/ClimbingCap_ Multi-Modal Dataset and Method for Rock Climbing in World Coordinate/full.md new file mode 100644 index 0000000000000000000000000000000000000000..542529ff8ecebb9cd22ae2c83878db4c71281437 --- /dev/null +++ b/2025/ClimbingCap_ Multi-Modal Dataset and Method for Rock Climbing in World Coordinate/full.md @@ -0,0 +1,323 @@ +# ClimbingCap: Multi-Modal Dataset and Method for Rock Climbing in World Coordinate + +Ming Yan $^{1,2,3*}$ Xincheng Lin $^{1,3*}$ Yuhua Luo $^{1,3}$ Shuqi Fan $^{1,3}$ Yudi Dai $^{7}$ Qixin Zhong $^{4}$ Lincai Zhong $^{5}$ Yuexin Ma $^{6}$ Lan Xu $^{6}$ Chenglu Wen $^{1,3}$ Siqi Shen $^{1,3\dagger}$ Cheng Wang $^{1,3}$ + +$^{1}$ Fujian Key Laboratory of Sensing and Computing for Smart Cities, Xiamen University + +$^{2}$ National Institute for Data Science in Health and Medicine, Xiamen University + +3Key Laboratory of Multimedia Trusted Perception and Efficient Computing, + +Ministry of Education of China, School of Informatics, Xiamen University + +4China National Climbing Team 5Ningbo Sports Work Training Team + +$^{6}$ ShanghaiTech University ETH AI Center, ETH Zürich + +![](images/c3ad0d134fecc3ddf3c7919012bf63efd650c3a94edc4a9feb043c105f06e1fa.jpg) +Figure 1. Overview. To address the challenging problem of global climbing motion recovery, we collect the dataset AscendMotion, using LiDAR, RGB camera and Inertial Measurement Unit (IMU) motion capture system with accurate motion labels and global trajectories (the blue and orange human bodies in right side of the figure represent labeled motions, and the orange curve represents the motion trajectory in the world coordinate.). Meanwhile, we propose ClimbingCap, a global climbing motion capturing method in world coordinate. As shown in the left part of this figure, it uses both image and LiDAR point cloud to recover human motions. + +# Abstract + +Human Motion Recovery (HMR) research mainly focuses on ground-based motions such as running. The study on capturing climbing motion, an off-ground motion, is sparse. This is partly due to the limited availability of climbing motion datasets, especially large-scale and challenging 3D labeled datasets. To address the insufficiency of climbing motion datasets, we collect AscendMotion, a large-scale well-annotated, and challenging climbing motion dataset. It consists of $412k$ RGB, LiDAR frames, and IMU measurements, including the challenging climbing motions of 22 skilled climbing coaches across 12 different rock walls. Capturing the climbing motions is challenging as it requires precise recovery of not only the complex pose but also the global position of climbers. Although multi + +ple global HMR methods have been proposed, they cannot faithfully capture climbing motions. To address the limitations of HMR methods for climbing, we propose ClimbingCap, a motion recovery method that reconstructs continuous 3D human climbing motion in a global coordinate system. One key insight is to use the RGB and LiDAR modalities to separately reconstruct motions in camera coordinates and global coordinates and to optimize them jointly. We demonstrate the quality of the AscendMotion dataset and present promising results from ClimbingCap. The AscendMotion dataset and source code release publicly at http://www.lidarhumanmotion.net/climbingcap/ + +# 1. Introduction + +It is challenging for global Human Motion Recovery (HMR) [1, 9, 40, 42, 80] due to the complexity of human poses and dynamic interaction between humans and their environments. Moreover, global HMR must main + +tain consistency in world coordinates to ensure authenticity and physical feasibility. The research community has proposed various methods to estimate human poses from different sensors such as RGB images [21-23, 33, 71], LiDAR point clouds [36, 49, 78], and inertial measurement units (IMUs) [25, 76]. Most of the studies [8, 15, 16, 24, 28, 39, 49, 62, 68, 73, 78] focus on recovering ground-based motions such as running, interacting, and dancing. + +Unlike ground-based motions such as running, climbing is an activity performed off-ground, where climbers use hands and feet to ascend holds or walls. As an important sport and recreation, the climbing race has become an official event in the Olympic Games. Yet, research on climbing motion capture [26, 34, 35, 54, 61] is sparse, partly due to the limited availability of climbing datasets. There are only two publicly available climbing datasets: SPEEDD21 [19] and CIMI4D [72]. SPEEDD21 is a 2D dataset. Although CIMI4D is 3D, it consists of many trivial climbing motions from casual climbers. Moreover, the size of the two datasets is small. As a result, the research community lacks an in-depth understanding of challenging climbing motions. + +We collect AscendMotion, a large-scale and challenging climbing motion multi-modal dataset, to address the data insufficiency. AscendMotion dataset includes the motions of 22 skilled climbers climbing 12 different rock walls. It includes rich modalities from hardware timesynchronized RGB, LiDAR point clouds, and Inertial Measurement Unit(IMU) MoCap system. Moreover, it consists of the global trajectories of climbers, which are important for understanding human motions in world coordinates. We also combine automatic annotation with manual refinement to ensure the accuracy of the motion annotation. + +Not only does climbing motion challenge humans, but capturing climbing motion is also challenging for researchers. Capturing climbing motion requires precise estimation of the complex pose resulting from human-scene interactions, as well as the global localization of climbers within scenes. Existing research dedicated to climbing [8, 62, 68] uses standard 2D HMR (e.g. OpenPose [9]) for climbing motions, which is suboptimal. Most existing HMR methods [20, 37, 55, 56, 60, 75] focus on recovering ground-based human motions through RGB imagery. The estimation results of these methods cannot be rigidly transformed among camera coordinates and global coordinate systems, due to the inherent ambiguity between coordinate systems. Moreover, global HMR methods often accumulate errors during long-term sequence recovery, affecting the accuracy of global trajectories and postures, especially in climbing movements, as we show through experiments. + +To tackle the above challenges and fill the gap in the motion capture community for climbing motions, we introduce ClimbingCap, an HMR method that reconstructs continuous 3D human climbing motion in both camera and + +global coordinate systems (see Fig.1). We adopt a trilogy in ClimbingCap: separate coordinate decoding, postprocessing, and semi-supervised training. For the separate coordinate decoding stage, ClimbingCap uses image and point cloud modality to estimate poses in the camera and the global coordinate systems, respectively. The postprocessing stage accurately ensures the consistency of motions between the two coordinate systems. In the semi-supervised training stage, a teacher-student training method is adopted to make use of easily obtainable unlabeled climbing data. Through these innovations, we can handle complex climbing movements effectively. To summarize, our main contributions include: + +- We collect the dataset AscendMotion, which is more comprehensive than existing datasets. We demonstrate the quality of the dataset and its poses are challenging. +- We propose ClimbingCap, which is a multimodality global HMR method for rock climbing. +- We conduct extensive experiments on evaluate the ClimbingCap in multiple datasets with various state-of-the-art methods. The experimental results show ClimbingCap perform better than all the methods for climbing motions. + +# 2. Related Work + +Camera-Space HMR. Camera-space HMR methods [7, 12, 13, 18, 27, 30, 31, 38, 44, 46, 48, 52, 53, 57-59, 63, 65, 70, 74, 81] predict human motion in independent camera coordinate system for each frame of the input video. Although camera-space methods can predict accurate human poses, they cannot effectively reflect the orientation and trajectory of the human body in world coordinates. + +World-Grounded HMR. Recent studies [32, 37, 55, 56, 60, 75, 77] have achieved remarkable progress in world coordinates. Most of them are two-stage methods. The first stage estimates human motion in camera coordinates, and the second stage optimizes the motion in world coordinates with the scene information. Most global HMR methods focus on ground-based motions (e.g., walking) that are located on the ground. They cannot accurately model climbing motions, which is an off-ground motion. Our method, ClimbingCap, is a global HMR method developed for recovering rock climbing motions. + +HMR Dataset. We have compared the AscendMotion dataset with multiple recent datasets in Table Tab. 1. Most of the datasets focus on ground-based human daily activities or sports. There exist only two public available climbing motion datasets. SPEED21 [19] collects 2D motions of climbers from sports videos. CIMI4D [72] contains 3D motions and global trajectories of climbers. Our climbing dataset, AscendMotion (344 minutes labeled data and 441 minutes unlabeled data, 412k frames), is significantly larger than both SPEED21 (21 minutes, 46k frames) and CIMI4D + +
DatasetSensor ModalitiesGlobal TrajectoryFrames3D SceneMotionReal/SyntheticSeqsSubjectsCareer
RGBMoCapLiDAR
SLOPER4D [15]✓IMU✓✓100k✓DaliyReal1512Normal
EMDB[28]✓EM-✓105k-DaliyReal8110Normal
BEHAVE [5]✓--✓15k-InteractionsReal-8Normal
RICH [24]✓--✓577k✓InteractionsReal14222Normal
AGORA [45]✓--✓106.7k✓DaliySynthetic---
BEDLAM [6]✓---1M✓DaliySynthetic---
FreeMotion [49]✓IMU✓✓578k-DaliyReal--Normal
HmPEAR [39]✓-✓-300k-DaliyReal6k25Normal
RELI11D [73]✓IMU✓✓239k✓SportReal4810Normal
HiSC4D [16]✓IMU✓✓36k✓SportReal8-Normal
LiDARHuman51M [78]✓IMU✓-374k-DailyReal5210Normal
CIMI4D [72]✓IMU✓✓180k✓ClimbingReal4212Normal
SPEED21 [19]✓--46kClimbingReal95--
AscendMotion(Ours)✓IMU✓✓412k✓(12)ClimbingReal22022Skilled
+ +Table 1. Comparisons with related HMR datasets. The ”-” symbol indicates that it is not included in the dataset. + +(120 minutes, 180k frames). Moreover, the motions in our dataset are more challenging than those in CIMI4D, as they come from climbing coaches who are more skilled than the casual climbers in CIMI4D. + +Climbing Motion Recovery. In recent years, the increasing popularity of climbing, coupled with the inclusion of rock climbing races in the Olympic Games, has led to growing interest in the capture and analysis of climbing motions [2, 3, 10, 51, 67]. Research on climbing MoCap methods remains limited. [43] employs OpenPose [9] for climbing pose estimation. [50] performs RGB-based method for local pose and estimates climbing position through a marker-based method. + +# 3. Method: ClimbingCap + +Climbing motion capture is challenging, involving poses with extreme limb extension and full-body exertion in camera coordinates. Moreover, it requires precise alignment with the rock wall in world coordinate as climbers are ascending. The ClimbingCap pipeline consists of three parts: separate coordinate decoding, post-processing, and semi-supervised training. An overview of the proposed pipeline is shown in Fig. 2. The separate coordinate decoding and post-processing parts take into account the unique challenges posed by climbing motion, which involves complex off-ground dynamics and interactions with scenes. The semi-supervised training part makes uses of the large-scale unlabeled climbing motion data to learn better a HMR model. + +# 3.1. Separate Coordinate Decoding + +The separate coordinate decoding (SCD) stage extracts features from the RGB sequence and LiDAR point clouds, and + +predicts the poses in camera coordinates and the positions in global coordinates. + +Input and Feature Extraction. The overall network structure is illustrated in Fig. 2. The input includes RGB images and point cloud data. First, the point cloud data is transformed from the world coordinate system to the camera coordinate system via an extrinsic matrix, represented as $\mathcal{P}_{\mathrm{c}} = \Omega_{w2c} \cdot \mathcal{P}_{\mathrm{w}}$ . Subsequently, the RGB images and transformed point cloud data are passed through feature extraction modules, RGB Extract and PC Extract, to obtain visual and geometric features. We build the feature extraction modules based on ViT [17] and PointNet++ [47]. These features are then fed into the following two decoder modules, which regress the SMPL parameters and global motion parameters of the human body, respectively. + +Camera Coordinate Decoder. This module decodes the SMPL parameters in the camera coordinate system. The RGB and point cloud features serve as inputs to the Camera Coordinate Decoder (denoted as $\mathcal{T}_{\mathrm{Decoder}}$ ), which processes the inputs with contextual information $\mathbf{f}_{\mathrm{backbone}}$ , generating an output token $\mathbf{t}_{\mathrm{out}}$ . This output token is then used to iteratively optimize the SMPL parameters, including the pose $\theta$ , shape $\beta$ , and camera translation $\Delta c$ . The iterative decoding approach allows the model to gradually approximate the true pose and shape in the camera coordinate system. In each iteration, the decoder updates the current SMPL parameters $\theta_{i}, \beta_{i}$ , and $\Delta c_{t}$ as follows: + +$$ +\mathbf {t} _ {\text {o u t}} = \mathcal {T} _ {\text {D e c o d e r}} (\mathbf {t}, \mathbf {f} _ {\text {b a c k b o n e}}), \tag {1} +$$ + +where $\theta_{i + 1} = \Phi_{\theta}\cdot \mathbf{t}_{\mathrm{out}} + \theta_{i},\beta_{i + 1} = \Phi_{\beta}\cdot \mathbf{t}_{\mathrm{out}} + \beta_{i}$ and $\Delta c_{i + 1} = \Phi_c\cdot \mathbf{t}_{\mathrm{out}} + \Delta c_i$ are the update equations, with $\Phi_{\theta}$ , $\Phi_{\beta}$ , and $\Phi_c$ representing the respective weight matrices for each parameter. Here, the input token $\mathbf{t}$ can include initialized pose, shape, and camera parameters as needed. + +![](images/b67825fb626a0f553f2c89c4583a420fc7d2f223b3664a0b47b0d3f1c9a777d1.jpg) +Figure 2. Overview of ClimbingCap. The arrows indicate the three stages of the ClimbingCap framework: separate coordinate decoding (the green box), post-processing (the blue box), and semi-supervised training (the red box). + +![](images/f2055e3ca958be05ee17c14dc0e93d409a5e8bc3b45085b1fc921b9637d45885.jpg) + +Global Coordinate Decoder. To fully capture the human motion trajectory in the world coordinate, we design the Global Translation Decoder to predict the global translation parameters $\Gamma^{\mathrm{trans}}$ of the human body. In this module, the decoder processes the features $\mathbf{f}_{\mathrm{backbone}}$ as contextual input, iteratively updating the global translation parameters. The update formula in each iteration is given by: + +$$ +\Gamma_ {i + 1} ^ {\text {t r a n s}} = \Psi \cdot \mathbf {t} _ {\text {o u t}} + \Gamma_ {i} ^ {\text {t r a n s}}, \tag {2} +$$ + +where $\Gamma_{i}^{\mathrm{trans}}$ represents the global translation parameters at time step $i$ , and $\Psi$ is the weight matrix for the update. This decoding process enables the model to capture a complete motion trajectory in the global coordinate system. + +Loss. The total loss function not only includes the 3D keypoint loss $\mathcal{L}_{\mathrm{kp3d}}$ and 2D keypoint loss $\mathcal{L}_{\mathrm{kp2d}}$ but also incorporates the SMPL parameter loss $\mathcal{L}^{\mathrm{spl}}$ and the global trajectory loss $\mathcal{L}_{\mathrm{traj}}$ . Specifically, the 3D keypoint loss $\mathcal{L}_{\mathrm{kp3d}}$ measures the 3D error of the predicted keypoints, while the 2D keypoint loss $\mathcal{L}_{\mathrm{kp2d}}$ measures the 2D projection error. The SMPL parameter loss $\mathcal{L}^{\mathrm{spl}}$ supervises the decoded pose and shape parameters, and the global trajectory loss $\mathcal{L}_{\mathrm{traj}}$ constrains the translation parameters to within close distance of ground truth positions. Please see the supplementary for details. The final total loss is formulated as: + +$$ +\mathcal {L} = \mathcal {L} _ {k p 3 d} + \mathcal {L} _ {k p 2 d} + \mathcal {L} _ {\theta} ^ {s m p l} + \mathcal {L} _ {\beta} ^ {s m p l} + \mathcal {L} _ {t r a j}. \tag {3} +$$ + +# 3.2. Post-processing + +Researches [55, 56, 75] have shown that a post-processing stage can be used to improve the output motion recovery results. Following these approaches, we employ a post-processing stage to optimize the output pose from SCD 3.1 stage. + +One distinct advantage of ClimbingCap is that the output results for the pose decoding stage can be rigidly transformed between the camera coordinate system and the world coordinate system. Thanks to the LiDAR modality, the point cloud contains 3D information in the world coordinate system. The poses obtained from the SCD stage are converted from the camera coordinate system to the world coordinate system through the inverse extrinsic matrix $\Omega_{w2c}^{-1}$ . + +The post-processing stage consists of three losses: $\mathcal{L}_{LWD}$ , $\mathcal{L}_{SDS}$ , and $\mathcal{L}_{VLR}$ . $\mathcal{L}_{LWD}$ assigns different weights to the vertices of different parts of the climbing human SMPL, and optimizes the position of the human body in the world coordinate system according to these weights. $\mathcal{L}_{SDS}$ optimizes joint positions by smoothing the changes in the direction of the human body's velocity on the scene (i.e., rock wall). $\mathcal{L}_{VLR}$ optimizes the difficult-to-estimate limb end poses by using the position of the point cloud in space. In this stage, the global poses are optimized through using the Adam [29] optimizer. Please refer to the supplementary for detailed definition of these losses. + +# 3.3. Semi-supervised Training + +Compared to ground-based motions [41], the size of labeled climbing motions is small. Simply using labeled climbing motion data may not be enough to train a robust model. Different from labeled climbing motion data, collecting unlabeled climbing motion data is cheaper. The AscendMotion dataset contains more unlabeled data than label data. They can be used to further improve the HMR model. + +It has been shown by researches [64, 69, 79] from the object detection community that through using a teacher-student semi-supervised training framework, the performance of object detection models can be improved. We adopt such semi-supervised training frame for HMR. In this work, we refer the model trained after the SCD and the post + +![](images/2f5bb39936921b1bc236f306567782bea86427b6529ebdb6bcc2199ba0be9fb2.jpg) +Figure 3. Dataset collection hardware system. + +processing stage as the teacher model (green box in Fig. 2). The student model (red box in Fig. 2) clones the parameters of the teacher model. During semi-supervised training, the teacher model estimate the pose labels from unlabeled motion data, and the pose label is used as pseudo-label to further train the student model. We show in the experiment section that via utilizing the semi-supervised training framework, the performance of HMR can be further improved. Please refer to the supplementary for the detailed training process. + +# 4. The AscendMotion Dataset + +To advance research in Human Mesh Recovery (HMR) for climbing motion within a global coordinate system, we present the AscendMotion dataset. This dataset specifically captures complex multi-directional rock climbing motions on non-planar surfaces. It provides a unique opportunity for challenging HMR research. AscendMotion includes multi-modal motion data from skilled climbers performing various climbing styles, such as bouldering, speed climbing, and lead climbing, with synchronized recordings from IMU, RGB cameras, and LiDAR. The dataset comprises 344 minutes (6 hours) of labeled data and 441 minutes of unlabeled data, featuring 22 adult climbers across 12 real-world climbing scenes. All participants agreed to use their recorded data for scientific purposes. + +# 4.1. Hardware and Configuration + +The AscendMotion data collection system integrates multiple sensors and captures motion data in both indoor and outdoor environments. As shown in Fig. 3, the LiDAR (OusterOS1, 128 beams) captures 3D dynamic point clouds at 20 frames per second (FPS), while the Hik 1080P RGB camera records RGB video at 20 FPS. For each climbing environment, we use the Trimble X7 3D laser scanner to reconstruct high-resolution RGB point cloud scenes, each containing approximately 80 million points. For the labeled + +dataset, each climber wears an Xsens MVN inertial motion capture system with 17 IMUs, recording at 60 FPS. [4] and a handheld point cloud scanner to ensure accurate body shape representation. For the unlabeled dataset, climbers do not wear a MoCap outfit, and we record their climbing motions using a RGB camera and a LiDAR. A human motion is represented by $M = (T, \theta, \beta)$ , where $T$ represents global translation, $\theta$ is the SMPL [40] pose parameters, and $\beta$ is the SMPL shape parameter. We obtain each climber's shape parameters $\beta$ using IPNet [4]. + +# 4.2. Annotation Pipeline + +The pose $\theta$ and the translation $T$ obtained from the IMU measurements may be inaccurate. Especially the translation $T$ may significantly drift for long-duration capture. The annotation pipeline of AscendMotion is used to find the accurate translate $T$ and pose $\theta$ as labels. + +The annotation pipeline of AscendMotion is depicted in Fig. 4. It consists of three parts: preprocessing, multistage global optimization, and manual annotation. The preprocessing part is used for time synchronization and spatial calibration. The multi-stage global optimization part is developed for automatic generation of annotation labels. The manual annotation part is used to further improve the quality of generated annotations. + +# 4.2.1 Multi-modal Data Preprocessing Stage + +Time Synchronization. The time among RGB camera and LiDAR is synchronized via Precision Time Protocol (PTP). We employ CollShark Auto 66 unit as the master clock, and it sends PTP slave clocks to the RGB camera and LiDAR. The time of IMU MoCap is post-synchronized with the LiDAR and RGB through anchor frames. + +Calibration. First, the LiDAR point cloud are registered with high-precision scanned-scenes. Next, the coordinate of LiDAR is treated as the world coordinate. The IMU measurements are transformed into the world coordinate through a calibration matrix. Finally, we perform frame-level calibration among RGB, LiDAR and IMU. Please refer to the appendix for details. + +# 4.2.2 Multi-stage Global Optimization + +AscendMotion uses the translation $T$ and pose $\theta$ provided by the IMU MoCap as the initialization of annotation labels, and performs multi-stage global optimization. To achieve accurate and natural human motion data consistent with the scene, we apply two loss functions: the Global Refit Loss $\mathcal{L}_{\mathrm{GR}}$ and the Scene Touch Loss $\mathcal{L}_{\mathrm{ST}}$ . + +Global Refit Loss $\mathcal{L}_{\mathrm{GR}}$ . The global refit loss promotes precise global alignment between the SMPL model and the + +![](images/cb659f075ccab75a8b7ef24ee341b2eee0dfa8721f3287452241e0c2a727fc21.jpg) +Figure 4. AscendMotion Annotation Pipeline. From left to right, this pipeline represents the three stages of dataset annotation: time-space synchronous input pre-processing(the blue box), multi-stage global optimization(the green box), and manual repair(the red box). + +![](images/4eaebb54defeeede5abb455c3ee28a2042c923d3e256fb4ba42597f6c3737d9e.jpg) +Multi-modal data preprocessing + +![](images/f0efa8d3143d7e6e2957752f95e095641c12c19e46df544446a3ed86939e99c2.jpg) + +![](images/eaa8e9134b89f229f15dd2e6151ce8dd1bcdec8de604fb7faf664f190e59899a.jpg) +Multi-stage global optimization + +![](images/194057910d2d7f7e7f177bed9f78643651b2d946d684688b6aab09ff446a2aee.jpg) +Manual Repair + +point clouds. $\mathcal{L}_{\mathrm{GR}}$ calculates the geometric discrepancy between the SMPL model vertices $V$ and the human body point cloud $P$ using a modified Chamfer distance. It applies different loss terms ( $f(|v_{i} - p_{j}|^{2})$ or $g(|v_{i} - p_{j}|^{2}))$ to different body parts to make body regions, such as the torso and limbs, match the point clouds of the human. We calculate the Euclidean distance between a vertex $v_{i}\in V$ and a point $p_j\in P$ as $d_{v_i,p_j}$ . A loss term $f(|v_i - p_j|^2)$ is applied to a vertex $v_{i}$ if $d_{v_i,p_j}\leq d_{\mathrm{torso}}$ , and $g(|p_j - v_i|^2)$ is applied to $v_{i}$ if $d_{v_i,p_j}\leq d_{\mathrm{limb}}$ , where $d_{\mathrm{torso}}$ and $d_{\mathrm{limb}}$ are the distance thresholds for torso and limbs, respectively. During climbing, humans use limbs to climb up, the distance among limbs and scenes should smaller than the distance among torso and scenes, $d_{\mathrm{limb}}\leq d_{\mathrm{torso}}$ . + +Scene Touch Loss $\mathcal{L}_{\mathrm{ST}}$ . This loss prevents unrealistic intersections between the SMPL model and the scene mesh by measuring the penetration depth between the SMPL vertices $v_{i} \in V$ and the scene mesh vertices $q_{j} \in Q$ . Given the scene mesh normal vectors $n_{j}$ , the penetration depth for each vertex is calculated as the dot product $\eta(v_{i}) = (v_{i} - q_{j}) \cdot n_{j}$ , where $q_{j}$ is the closest mesh vertex to $v_{i}$ . If the penetration depth $\eta(v_{i})$ is negative, it indicates that the vertex $v_{i}$ has penetrated the scene mesh, and this value contributes to the loss $\mathcal{L}_{\mathrm{ST}}$ . This loss encourages the SMPL model to avoid intersecting with the scene mesh, promoting more realistic interactions between the model and the environment. + +# 4.2.3 Manual Annotation and Verification + +To further improve the dataset quality, we use the SMPL annotation tool [66] for enhanced labeling of sequences. We engage four observers to review the automatically annotated results. For any ambiguous limb movements observed in RGB views, we perform additional manual adjustments and mark these frames as key frames. Finally, we propagate poses from multiple key frames within each sequence to ensure global consistency in both smoothness and accuracy. + +
Constraint termHorizontal SceneVertical Scene
\( \mathcal{L}_{ST} \)\( \mathcal{L}_{GR} \)ACCEL↓MPJPE↓PA-MPJPE↓ACCEL↓MPJPE↓PA-MPJPE↓
XX1.8337.5732.526.7951.5146.24
✓X1.6735.4531.813.7340.3635.37
X✓1.5929.0425.951.6629.2424.99
✓✓1.5428.2023.251.6029.0724.37
+ +Table 2. Quantitative Evaluation of Dataset Quality. + +# 5. EXPERIMENTS + +In this section, we demonstrate the high-quality of the AscendMotion optimization process in Sec. 5.1, show that the AscendMotion dataset is challenging to existing HMR methods, and evaluate ClimbingCap and multiple HMR methods in climbing motion datasets in Sec. 5.2. We demonstrate the promising performance of ClimbingCap and shows that each component of ClimbingCap is useful through ablation study in Sec. 5.3. Please refer to the detailed experimental setup and metrics in the supplementary. + +# 5.1. Datasets Evaluation + +To quantitatively evaluate the annotation quality of AscendMotion, we divide the scene into horizontal and vertical rock walls and evaluate the performance of the global optimization stage (in Sec. 4.2.2) by comparing the generated annotations with manual annotations. + +The evaluation metrics include widely used MPJPE, Procrustes Aligned MPJPE (PA-MPJPE), per vertex error (PVE), acceleration error (Accel, in units of $m / s^2$ ). These metrics are used for evaluating HMR methods as well. + +To understand the impact of different constraints used in the combined optimization stage, we perform an ablation study on two different losses: $\mathcal{L}_{ST}$ and $\mathcal{L}_{GR}$ . Tab. 2 shows the error metrics using different loss combinations for two scenes. The error metrics are small, demonstrating the effectiveness of the global optimization pipeline and the high quality of AscendMotion. + +
ModalityMethodCamera CoordinateWorld Coordinate
ACCEL↓MPJPE↓PA-MPJPE↓PVE↓PCK0.3↑WA-MPJPE↓W-MPJPE↓RTE↓Jitter↓T-Error↓
RGBTRACE [60]18.68/76.79875.56/577.6069.21/85.81951.52/619.930.06/0.09144.33/385.71254.38/703.3514.73/26.17115.96/521.402.56/6.62
SLAHMR [75]5.46/96.98232.46/467.8884.13/285.15283.24/552.630.36/0.12277.47/447.68804.85/613.603.64/39.394.91/201.332.81/6.54
WHAM [56]4.59/35.01110.92/143.1776.09/73.36124.2/164.910.76/0.62229.42/1125.77647.70/1499.855.16/9.043.58/40.691.77/2.49
GVHMR [55]4.50/26.22107.09/124.6060.06/80.30118.89/151.100.77/0.71105.15/1002.11202.45/1442.504.09/7.916.85/32.711.48/2.54
LiDARLiDARCapV2 [78]87.99/119.62244.6/234.52192.17/156.39326.45/283.270.53/0.50282.12/1396.42442.12/1518.2916.42/10.85176.95/165.551.65/2.89
LiveHPS [49]157.87/195.23156.5/147.31142.19/121.76191.87/189.300.64/0.70235.4/1369.89392.34/1506.5013.94/10.45279.96/358.542.1/6.73
LiDAR +RGBImmFusion [11]108.76/74.20473.18/464.83254.07/179.51533.5/529.710.17/0.14324.4/1446.01487.92/1532.8816.52/10.8627.03/14.491.94/3.99
FusionPose [14]112.08/86.44256.81/315.93198.55/193.83306.22/359.680.36/0.42275.02/1445.32444.47/1532.2816.29/10.8592.97/80.792.02/6.48
LEIR [73]110.18/94.57297.95/299.62187.26/150.56340.61/351.520.41/0.37266.82/1313.09282.31/1435.929.78/9.9773.38/85.031.1/1.20
Ours5.17/17.2575.45/88.9261.73/74.5094.89/106.420.91/0.7862.95/85.2678.99/106.951.57/3.128.3/27.751.07/1.29
+ +Table 3. HMR Comparison in the AscendMotion dataset (Horizontal Scene/Vertical Scene). The results for the horizontal and the vertical scene are separated by “/” in each cell. ClimbingCap performs significantly than others. + +
ModalityMethodCamera CoordinateWorld Coordinate
ACCEL↓MPJPE↓PA-MPJPE↓PVE↓PCK0.3↑WA-MPJPE↓W-MPJPE↓RTE↓Jitter↓T-Error↓
RGBTRACE [60]57.65488.6398.83685.290.17365.11608.3828.9827.304.58
SLAHMR [75]6.10228.8282.39201.420.61391.24445.3813.8812.632.47
WHAM [56]9.11184.4394.40191.270.76321.58430.3532.959.263.78
GVHMR [55]6.98152.5691.13140.190.60349.17476.9021.5411.632.26
LiDARLiDARCapV2 [78]70.72389.97267.76364.260.50468.00520.0613.5654.253.12
LiveHPS [49]72.57190.86148.00225.650.59412.22524.4712.1150.773.52
LiDAR+RGBImmFusion [11]68.82322.21232.44435.720.45465.46586.3613.5914.553.96
FusionPose [14]58.54242.20189.92330.090.53373.46437.8813.5448.302.01
LEIR [73]16.46206.69117.52269.120.58228.1370.7410.638.421.84
Ours5.1884.0360.6999.890.86228.28261.068.178.721.26
+ +Table 4. HMR Comparison in the CIMI4D dataset. ClimbingCap demonstrates remarkable generalization and robustness on other climbing datasets. + +# 5.2. Comparison on Global Motion Recovery + +To evaluate the performance of global HMR, we follow the motion evaluation protocols of [55, 56] and the global trajectory evaluation protocol of [15, 73]. Following WHAM [56], to calculate the world coordinate MPJPE metrics, the predicted global sequence is divided into segments of 100 frames, and each segment is aligned with the ground truth segment either to the entire segment (WA-MPJPE) or to the first two frames (W-MPJPE). Besides WA-MPJPE and W-MPJPE, the root translation error (RTE), motion jitter (Jitter, in $m / s^3$ ), relative global translation error (T-Error in $m$ ) of the entire sequence are reported. + +We evaluate ClimbingCap against nine methods with different modalities. The global RGB methods are TRACE [60], SLAHMR [75], WHAM [56], and GVHMR [55]. The LiDAR-based methods are LiDAR-CapV2 [78] and LiveHPS [49]. The LiDAR+RGB methods are ImmFusion [11], FusionPose [14], and LEIR [73]. + +Results in AscendMotion. The scenes in AscendMotion are categories into horizontal and vertical scenes, based on the major direction of human motions. Vertical motions are more challenging to human and to HMR methods than horizontal motions. The experimental results are depicted in Tab. 3. The top rows depicted the results for multiple state-of-the-art RGB-based HRM methods, and the bottom rows depicted the results for state-of-the-art LiDAR and Li- + +DAR+RGB based methods. In each cell of the table, the results for the horizontal and the vertical scenes are separated by $l$ . For horizontal scenes, ClimbingCap performs the best in most of the metrics, especially in terms of MPJPE, WA-MPJPE and W-MPJPE. WHAM and GVHMR performs well for the camera coordinate metrics, but not in the world coordinate metrics. For the vertical scenes, ClimbingCap performs the best and significantly better than the second-best methods. GVHMR, a representative global HMR method, performs poorly in vertical scenes. It estimates the direction of movement by predicting the horizontal velocity. However, the major movement on the vertical scene of AscendMotion is an upward climbing movement. The LiDAR-based, and other LiDAR+RGB-based methods perform inferior to ClimbingCap. These methods do not fully consider global trajectories. The results demonstrate the importance of considering the relationship among camera coordinates and global coordinates. + +Results in CIMI4D. To test the generalization ability of all the HMR methods, we evaluate their performance in the CIMI4D climbing dataset without retraining or fine-tuning. The CIMI4D dataset consists mostly horizontal scenes, and the motions are less challenging than AscendMotion. As it is reported in Tab. 4, ClimbingCap performs the best in the CIMI4D dataset for most the metrics, except the WA-MPJPE metric. + +![](images/78f3baf5b8fcbef56deeffe2cd3146b27bff70865b84098acd4ccbc897346237.jpg) +Figure 5. Qualitative Evaluation in the AscendMotion and CIMI4D dataset. The left and right areas show the results of Camera Coordinate and World Coordinate respectively. The red circles indicate obvious errors. The last row shows the results for CIMI4D dataset. Our method ClimbingCap performs best qualitatively by comparison. + +![](images/e06950a81260058782cbeb2c7d1a301de62fb9c1dab1d56f44a4089fd50ab2e5.jpg) + +
VariantMPIJEPA-MPJPEPCK0.3WA-MPJPEW-MPJPERTE
(1) RGB Input105.6763.050.78117.17174.537.64
(2) w/o LwD80.4652.150.8970.0491.232.02
(3) w/o LSDS99.1360.660.81109.35164.117.03
(4) w/o VLRR91.1061.850.8388.59120.113.34
(5) w/o SS77.4352.570.9065.3082.091.83
RGB+LiDAR Input Full Model75.4550.510.9162.9578.991.57
+ +Table 5. Ablation Experiment for ClimbingCap in AscendMotion. + +# 5.3. Ablation Experiment + +To understand the impact of each global loss and the semi-supervised module on ClimbingCap's performance, we evaluate five variants of ClimbingCap on the AscendMotion dataset following the same training and evaluation protocol as the previous section. (1) RGB Input: Using only RGB imagery without LiDAR, this variant performs worst overall due to missing global position data required for post-processing. (2) w/o $\mathcal{L}_{LWD}$ : Excluding the limb weight differentiation loss, which guides optimization using global point clouds, has a limited impact, indicating a beneficial but non-essential role. (3) w/o $\mathcal{L}_{SDS}$ : Removing the velocity direction smoothing loss, which ensures motion consistency, significantly reduces global metrics, highlighting its importance. (4) w/o $\mathcal{L}_{VLR}$ : The velocity direction smoothing loss constrains limb movements and corrects unexpected mis-predicted movements in a motion sequence; without this loss, ClimbingCap's global metrics drop sig + +nificantly, highlighting its importance for consistency and accuracy. (5) w/o $SS$ : Removing the semi-supervised module results in a slight performance decrease, indicating its contribution to improving network training, though with a modest effect. The full ClimbingCap model, with RGB and LiDAR inputs, all loss terms, and the semi-supervised framework, achieves the best results in Tab. 5, validating the design's effectiveness for climbing motion analysis on AscendMotion. + +# 6. Conclusion + +In this work, we propose AscendMotion, a large-scale and multi-modal climbing motion dataset, and ClimbingCap, a human motion recovery framework for climbing motions. AscendMotion provides a rich, high-quality dataset that surpasses previous climbing datasets in both scale and complexity. ClimbingCap effectively recovers 3D climbing motions in the global coordinate system, ensuring accurate pose estimation and global localization. Experimental results demonstrate the quality of AscendMotion dataset, and show that ClimbingCap outperforms existing methods in terms of accuracy and robustness. + +Acknowledgment. This work was partially supported by the Fundamental Research Funds for the Central Universities (No. 20720230033); by PDL (2022-PDL-12); by Xiaomi Young Talents Program. + +# References + +[1] Thiemo Alldieck, Marc Kassubeck, Bastian Wandt, Bodo Rosenhahn, and Marcus Magnor. Optical flow-based 3d human motion estimation from monocular video. In German Conference on Pattern Recognition, pages 347-360. Springer, 2017. 1 +[2] Marina Andric, Francesco Ricci, and Floriano Zini. Sensor-based activity recognition and performance assessment in climbing: A review. IEEE Access, 10:108583-108603, 2022. 3 +[3] Raul Beltrán Beltrán, Julia Richter, Guido Köstermeyer, and Ulrich Heinkel. Climbing technique evaluation by means of skeleton video stream analysis. Sensors, 23(19):8216, 2023. 3 +[4] Bharat Lal Bhatnagar, Cristian Sminchisescu, Christian Theobalt, and Gerard Pons-Moll. Combining implicit function learning and parametric models for 3d human reconstruction. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16, pages 311-329. Springer, 2020. 5 +[5] Bharat Lal Bhatnagar, Xianghui Xie, Ilya A Petrov, Cristian Sminchisescu, Christian Theobalt, and Gerard Pons-Moll. Behave: Dataset and method for tracking human object interactions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15935-15946, 2022. 3 +[6] Michael J Black, Priyanka Patel, Joachim Tesch, and Jinlong Yang. Bedlam: A synthetic dataset of bodies exhibiting detailed lifelike animated motion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8726-8737, 2023. 3 +[7] Federica Bogo, Angjoo Kanazawa, Christoph Lassner, Peter V. Gehler, Javier Romero, and Michael J. Black. Keep it smpl: Automatic estimation of 3d human pose and shape from a single image. In ECCV, 2016. 2 +[8] Christoph Bregler and Jitendra Malik. Tracking people with twists and exponential maps. Proceedings. 1998 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (Cat. No.98CB36231), pages 8-15, 1998. 2 +[9] Zhe Cao, Tomas Simon, Shih-En Wei, and Yaser Sheikh. Realtime multi-person 2d pose estimation using part affinity fields. In Computer Vision and Pattern Recognition (CVPR), 2017. 1, 2, 3 +[10] Kyungsik Cha, Eun-Young Lee, Myeong-Hyeon Heo, Kyu-Cheol Shin, Jonghee Son, and Dongho Kim. Analysis of climbing postures and movements in sport climbing for realistic 3d climbing animations. Procedia Engineering, 112: 52-57, 2015. 3 +[11] Anjun Chen, Xiangyu Wang, Kun Shi, Shaohao Zhu, Bin Fang, Yingfeng Chen, Jiming Chen, Yuchi Huo, and Qi Ye. Immfusion: Robust mmwave-rgb fusion for 3d human body reconstruction in all weather conditions. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 2752-2758. IEEE, 2023. 7 +[12] Xin Chen, Anqi Pang, Wei Yang, Yuexin Ma, Lan Xu, and Jingyi Yu. Sportscap: Monocular 3d human motion capture and fine-grained understanding in challenging sports videos. + +International Journal of Computer Vision, 129:2846-2864, 2021. 2 +[13] Hai Ci, Mingdong Wu, Wentao Zhu, Xiaoxuan Ma, Hao Dong, Fangwei Zhong, and Yizhou Wang. Gfpose: Learning 3d human pose prior with gradient fields. In CVPR, pages 4800-4810, 2023. 2 +[14] Peishan Cong, Yiteng Xu, Yiming Ren, Juze Zhang, Lan Xu, Jingya Wang, Jingyi Yu, and Yuexin Ma. Weakly supervised 3d multi-person pose estimation for large-scale scenes based on monocular camera and single lidar. In AAAI, pages 461-469. AAAI Press, 2023. 7 +[15] Yudi Dai, Yitai Lin, XiPing Lin, Chenglu Wen, Lan Xu, Hongwei Yi, Siqi Shen, Yuexin Ma, and Cheng Wang. Sloper4d: A scene-aware dataset for global 4d human pose estimation in urban environments. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 682-692, 2023. 2, 3, 7 +[16] Yudi Dai, Zhiyong Wang, Xiping Lin, Chenglu Wen, Lan Xu, Siqi Shen, Yuexin Ma, and Cheng Wang. Hisc4d: Human-centered interaction and 4d scene capture in large-scale space using wearable imus and lidar. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 2, 3 +[17] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In 9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7, 2021. OpenReview.net, 2021. 3 +[18] Zhiyang Dou, Qingxuan Wu, Cheng Lin, Zeyu Cao, Qiangqiang Wu, Weilin Wan, Taku Komura, and Wenping Wang. TORE: token reduction for efficient human mesh recovery with transformer. In ICCV, 2023. 2 +[19] Petr Elias, Veronika Skvarlova, and Pavel Zezula. Speed21: Speed climbing motion dataset. In Proceedings of the 4th International Workshop on Multimedia Content Analysis in Sports, pages 43-50, 2021. 2, 3 +[20] Shubham Goel, Georgios Pavlakos, Jathushan Rajasegaran, Angjoo Kanazawa, and Jitendra Malik. Humans in 4d: Reconstructing and tracking humans with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14783-14794, 2023. 2 +[21] Marc Habermann, Weipeng Xu, Michael Zollhöfer, Gerard Pons-Moll, and Christian Theobalt. Livecap: Real-time human performance capture from monocular video. ACM Transactions on Graphics (TOG), 38(2):14:1-14:17, 2019. 2 +[22] Marc Habermann, Weipeng Xu, Michael Zollhofer, Gerard Pons-Moll, and Christian Theobalt. Deepcap: Monocular human performance capture using weak supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. +[23] Yannan He, Anqi Pang, Xin Chen, Han Liang, Minye Wu, Yuexin Ma, and Lan Xu. Challencap: Monocular 3d capture of challenging human performances using multi-modal refer + +ences. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11400-11411, 2021. 2 +[24] Chun-Hao P Huang, Hongwei Yi, Markus Höschle, Matvey Safroshkin, Tsvetelina Alexiadis, Senya Polikovsky, Daniel Scharstein, and Michael J Black. Capturing and inferring dense full-body human-scene contact. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13274-13285, 2022. 2, 3 +[25] Yinghao Huang, Manuel Kaufmann, Emre Aksan, Michael J. Black, Otmar Hilliges, and Gerard Pons-Moll. Deep inertial poser: Learning to reconstruct human pose from sparse inertial measurements in real time. ACM Transactions on Graphics, (Proc. SIGGRAPH Asia), 37(6):185:1-185:15, 2018. 2 +[26] Hitomi Iguma, Akihiro Kawamura, and Ryo Kurazume. A new 3d motion and force measurement system for sport climbing. In 2020 IEEE/SICE International Symposium on System Integration, SII 2020, Honolulu, HI, USA, January 12-15, 2020, pages 1002-1007. IEEE, 2020. 2 +[27] Angjoo Kanazawa, Michael J. Black, David W. Jacobs, and Jitendra Malik. End-to-end recovery of human shape and pose. In Computer Vision and Pattern Recognition (CVPR), 2018. 2 +[28] Manuel Kaufmann, Jie Song, Chen Guo, Kaiyue Shen, Tianjian Jiang, Chengcheng Tang, Juan José Zárate, and Otmar Hilliges. Emdb: The electromagnetic database of global 3d human pose and shape in the wild. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14632-14643, 2023. 2, 3 +[29] Diederik P Kingma and Jimmy Lei Ba. Adam: A method for stochastic gradient descent. In ICLR: international conference on learning representations, pages 1-15. ICLR US., 2015. 4 +[30] Muhammed Kocabas, Nikos Athanasiou, and Michael J. Black. Vibe: Video inference for human body pose and shape estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2 +[31] Muhammed Kocabas, Chun-Hao P. Huang, Otmar Hilliges, and Michael J. Black. Pare: Part attention regressor for 3d human body estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 11127-11137, 2021. 2 +[32] Muhammed Kocabas, Ye Yuan, Pavlo Molchanov, Yunrong Guo, Michael J Black, Otmar Hilliges, Jan Kautz, and Umar Iqbal. Pace: Human and camera motion estimation from inthe-wild videos. In 2024 International Conference on 3D Vision (3DV), pages 397-408. IEEE, 2024. 2 +[33] Nikos Kolotouros, Georgios Pavlakos, Michael J Black, and Kostas Daniilidis. Learning to reconstruct 3d human pose and shape via model-fitting in the loop. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2252-2261, 2019. 2 +[34] Felix Kosmalla, André Zenner, Marco Speicher, Florian Daiber, Nico Herbig, and Antonio Krüger. Exploring rock climbing in mixed reality environments. In Proceedings of the 2017 CHI Conference Extended Abstracts on Human + +Factors in Computing Systems, page 1787-1793, New York, NY, USA, 2017. Association for Computing Machinery. 2 +[35] Felix Kosmalla, André Zenner, Corinna Tasch, Florian Daiber, and Antonio Krüger. The importance of virtual hands and feet for virtual reality climbing. In Extended Abstracts of the 2020 CHI Conference on Human Factors in Computing Systems, CHI 2020, Honolulu, HI, USA, April 25-30, 2020, pages 1-8. ACM, 2020. 2 +[36] Jialian Li, Jingyi Zhang, Zhiyong Wang, Siqi Shen, Chenglu Wen, Yuexin Ma, Lan Xu, Jingyi Yu, and Cheng Wang. Lidarcap: Long-range marker-less 3d human motion capture with lidar point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20502-20512, 2022. 2 +[37] Jiefeng Li, Ye Yuan, Davis Rempe, Haotian Zhang, Pavlo Molchanov, Cewu Lu, Jan Kautz, and Umar Iqbal. Coin: Control-inpainting diffusion prior for human and camera motion estimation. In European Conference on Computer Vision, pages 426-446. Springer, 2025. 2 +[38] Zhihao Li, Jianzhuang Liu, Zhensong Zhang, Songcen Xu, and Youliang Yan. Cliff: Carrying location information in full frames into human pose and shape estimation. In ECCV, 2022. 2 +[39] Yitai Lin, Zhijie Wei, Wanfa Zhang, Xiping Lin, Yudi Dai, Chenglu Wen, Siqi Shen, Lan Xu, and Cheng Wang. Hmpear: A dataset for human pose estimation and action recognition. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 2069-2078, 2024. 2, 3 +[40] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J. Black. Smpl: A skinned multiperson linear model. ACM Trans. Graph., 34(6):248:1-248:16, 2015. 1, 5 +[41] Naureen Mahmood, Nima Ghorbani, Nikolaus F. Troje, Gerard Pons-Moll, and Michael J. Black. Amass: Archive of motion capture as surface shapes. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2019. 4 +[42] Julieta Martinez, Rayat Hossain, Javier Romero, and James J Little. A simple yet effective baseline for 3d human pose estimation. In ICCV, 2017. 1 +[43] Dominik Pandurevic, Paweł Draga, Alexander Sutor, and Klaus Hochradel. Analysis of competition and training videos of speed climbing athletes using feature and human body keypoint detection algorithms. Sensors, 22(6):2251, 2022. 3 +[44] Chaitanya Patel, Zhouyingcheng Liao, and Gerard Pons-Moll. Tailornet: Predicting clothing in 3d as a function of human pose, shape and garment style. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7363-7373, 2020. 2 +[45] Priyanka Patel, Chun-Hao P. Huang, Joachim Tesch, David T. Hoffmann, Shashank Tripathi, and Michael J. Black. AGORA: Avatars in geography optimized for regression analysis. In Proceedings IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2021. 3 +[46] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed A. A. Osman, Dimitrios Tzionas, and + +Michael J. Black. Expressive body capture: 3d hands, face, and body from a single image. In Proceedings IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 10975-10985, 2019. 2 +[47] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems, 30, 2017. 3 +[48] Davis Rempe, Tolga Birdal, Aaron Hertzmann, Jimei Yang, Srinath Sridhar, and Leonidas J. Guibas. Humor: 3d human motion model for robust pose estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 11488-11499, 2021. 2 +[49] Yiming Ren, Xiao Han, Chengfeng Zhao, Jingya Wang, Lan Xu, Jingyi Yu, and Yuexin Ma. Livehps: Lidar-based scene-level human pose and shape estimation in free environment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1281-1291, 2024. 2, 3, 7 +[50] Lionel Reveret, Sylvain Chapelle, Franck Quaine, and Pierre Legreneur. 3d visualization of body motion in speed climbing. Frontiers in Psychology, 11:2188, 2020. 3 +[51] Julia Richter, Raul Beltrán Beltrán, Guido Köstermeyer, and Ulrich Heinkel. Human climbing and bouldering motion analysis: A survey on sensors, motion capture, analysis algorithms, recent advances and applications. VISIGRAPP (5: VISAPP), pages 751-758, 2020. 3 +[52] Shunsuke Saito, Zeng Huang, Ryota Natsume, Shigeo Morishima, Hao Li, and Angjoo Kanazawa. Pifu: Pixel-aligned implicit function for high-resolution clothed human digitization. In 2019 IEEE/CVF International Conference on Computer Vision, ICCV 2019, Seoul, Korea (South), October 27 - November 2, 2019, pages 2304-2314. IEEE, 2019. 2 +[53] Shunsuke Saito, Tomas Simon, Jason M. Saragih, and Hanbyul Joo. Pifuhd: Multi-level pixel-aligned implicit function for high-resolution 3d human digitization. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2020, Seattle, WA, USA, June 13-19, 2020, pages 81-90. Computer Vision Foundation / IEEE, 2020. 2 +[54] Katsuhito Sasaki, Keisuke Shiro, and Jun Rekimoto. Exempliser: Predicting poses of experts as examples for beginners in climbing using a neural network. In Proceedings of the Augmented Humans International Conference, AHs 2020, Kaiserslautern, Germany, 16-17 March, 2020, pages 18:1-18:9. ACM, 2020. 2 +[55] Zehong Shen, Huaijin Pi, Yan Xia, Zhi Cen, Sida Peng, Zechen Hu, Hujun Bao, Ruizhen Hu, and Xiaowei Zhou. World-grounded human motion recovery via gravity-view coordinates. In SIGGRAPH Asia Conference Proceedings, 2024. 2, 4, 7 +[56] Soyong Shin, Juyong Kim, Eni Halilaj, and Michael J Black. Wham: Reconstructing world-grounded humans with accurate 3d motion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2070-2080, 2024. 2, 4, 7 +[57] Zhuo Su, Lan Xu, Zerong Zheng, Tao Yu, Yebin Liu, and Lu + +Fang. Robustfusion: Human volumetric capture with data-driven visual cues using a rgbd camera. In ECCV, 2020. 2 +[58] Yu Sun, Qian Bao, Wu Liu, Yili Fu, Black Michael J., and Tao Mei. Monocular, One-stage, Regression of Multiple 3D People. In ICCV, 2021. +[59] Yu Sun, Wu Liu, Qian Bao, Yili Fu, Tao Mei, and Michael J Black. Putting People in their Place: Monocular Regression of 3D People in Depth. In CVPR, 2022. 2 +[60] Yu Sun, Qian Bao, Wu Liu, Tao Mei, and Michael J Black. Trace: 5d temporal regression of avatars with dynamic cameras in 3d environments. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8856-8866, 2023. 2, 7 +[61] Marcel Tiator, Christian Geiger, Bastian Dewitz, Ben Fischer, Laurin Gerhardt, David Nowottnik, and Hendrik Preu. Venga!: climbing in mixed reality. In Proceedings of the First Superhuman Sports Design Challenge: First International Symposium on Amplifying Capabilities and Competing in Mixed Realities, July 2-5, 2018, Delft, The Netherlands, pages 9:1-9:8. ACM, 2018. 2 +[62] Daniel Vlasic, Rolf Adelsberger, Giovanni Vannucci, John C. Barnwell, Markus H. Gross, Wojciech Matusik, and Jovan Popovic. Practical motion capture in everyday surroundings. ACM SIGGRAPH 2007 papers, 2007. 2 +[63] Ziniu Wan, Zhengjia Li, Maoqing Tian, Jianbo Liu, Shuai Yi, and Hongsheng Li. Encoder-decoder with multi-level attention for 3d human shape and pose estimation. In 2021 IEEE/CVF International Conference on Computer Vision, ICCV 2021, Montreal, QC, Canada, October 10-17, 2021, pages 13013-13022. IEEE, 2021. 2 +[64] He Wang, Yezhen Cong, Or Litany, Yue Gao, and Leonidas J Guibas. 3dioumatch: Leveraging iou prediction for semi-supervised 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14615-14624, 2021. 4 +[65] Kuan-Chieh Wang, Zhenzhen Weng, Maria Xenochristou, João Pedro Araújo, Jeffrey Gu, Karen Liu, and Serena Yeung. Nemo: Learning 3d neural motion fields from multiple video instances of the same action. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22129-22138, 2023. 2 +[66] MeiJun Wang, Yu Meng, Zhongwei Qiu, Chao Zheng, Yan Xu, Jian Gao, et al. Pedestrian-centric 3d pre-collision pose and shape estimation from dashcam perspective. In The Thirty-eighth Annual Conference on Neural Information Processing Systems. 6 +[67] Emily Whiting, Nada Ouf, Liane Makatura, Christos Mousas, Zhenyu Shu, and Ladislav Kavan. Environment-scale fabrication: Replicating outdoor climbing experiences. In Proceedings of the 2017 CHI Conference on Human Factors in Computing Systems, pages 1794-1804. ACM, 2017. 3 +[68] HJ Wolring. New possibilities for human motion studies by real-time light spot position measurement. Biotelemetry, 1 (3):132, 1974. 2 +[69] Qiming Xia, Wei Ye, Hai Wu, Shijia Zhao, Leyuan Xing, Xun Huang, Jinhao Deng, Xin Li, Chenglu Wen, and Cheng + +Wang. Hinted: Hard instance enhanced detector with mixed-density feature fusion for sparsely-supervised 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15321-15330, 2024. 4 +[70] Yuliang Xiu, Jinlong Yang, Dimitrios Tzionas, and Michael J. Black. ICON: implicit clothed humans obtained from normals. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 13286-13296. IEEE, 2022. 2 +[71] Weipeng Xu, Avishek Chatterjee, Michael Zollhöfer, Helge Rhodin, Dushyant Mehta, Hans-Peter Seidel, and Christian Theobalt. Monoperfcap: Human performance capture from monocular video. ACM Transactions on Graphics (TOG), 37 (2):27:1-27:15, 2018. 2 +[72] Ming Yan, Xin Wang, Yudi Dai, Siqi Shen, Chenglu Wen, Lan Xu, Yuexin Ma, and Cheng Wang. Cimi4d: A large multimodal climbing motion dataset under human-scene interactions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12977-12988, 2023. 2, 3 +[73] Ming Yan, Yan Zhang, Shuqiang Cai, Shuqi Fan, Xincheng Lin, Yudi Dai, Siqi Shen, Chenglu Wen, Lan Xu, Yuexin Ma, et al. Reli11d: A comprehensive multimodal human motion dataset and method. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2250-2262, 2024. 2, 3, 7 +[74] Ze Yang, Shenlong Wang, Siva Manivasagam, Zeng Huang, Wei-Chiu Ma, Xinchen Yan, Ersin Yumer, and Raquel Urtasun. S3: Neural shape, skeleton, and skinning fields for 3d human modeling. In CVPR, 2021. 2 +[75] Vickie Ye, Georgios Pavlakos, Jitendra Malik, and Angjoo Kanazawa. Decoupling human and camera motion from videos in the wild. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2, 4, 7 +[76] Xinyu Yi, Yuxiao Zhou, Marc Habermann, Soshi Shimada, Vladislav Golyanik, Christian Theobalt, and Feng Xu. Physical inertial poser (PIP): physics-aware real-time human motion tracking from sparse inertial sensors. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 13157-13168. IEEE, 2022. 2 +[77] Ye Yuan, Umar Iqbal, Pavlo Molchanov, Kris Kitani, and Jan Kautz. Glamr: Global occlusion-aware human mesh recovery with dynamic cameras. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2 +[78] Jingyi Zhang, Qihong Mao, Siqi Shen, Chenglu Wen, Lan Xu, and Cheng Wang. Lidarcapv2: 3d human pose estimation with human-object interaction from lidar point clouds. Pattern Recognition, 156:110848, 2024. 2, 3, 7 +[79] Wu Zheng, Weiliang Tang, Li Jiang, and Chi-Wing Fu. Sessd: Self-ensembling single-stage object detector from point cloud. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14494–14503, 2021. 4 +[80] Xiaowei Zhou, Menglong Zhu, Spyridon Leonardos, Konstantinos G Derpanis, and Kostas Daniilidis. Sparseness + +Meets Deepness: 3D Human Pose Estimation from Monocular Video. In Computer Vision and Pattern Recognition (CVPR), 2016. 1 +[81] Wentao Zhu, Xiaoxuan Ma, Zhaoyang Liu, Libin Liu, Wayne Wu, and Yizhou Wang. Motionbert: Unified pretraining for human motion analysis. In ICCV, 2023. 2 \ No newline at end of file diff --git a/2025/ClimbingCap_ Multi-Modal Dataset and Method for Rock Climbing in World Coordinate/images.zip b/2025/ClimbingCap_ Multi-Modal Dataset and Method for Rock Climbing in World Coordinate/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..0ae701033352283b67fa31d91bb2695d1fd9d813 --- /dev/null +++ b/2025/ClimbingCap_ Multi-Modal Dataset and Method for Rock Climbing in World Coordinate/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8057ec1632870cc2966041788c59852ddc4169e351c9ad2d196f2464b2ff7fd8 +size 916519 diff --git a/2025/ClimbingCap_ Multi-Modal Dataset and Method for Rock Climbing in World Coordinate/layout.json b/2025/ClimbingCap_ Multi-Modal Dataset and Method for Rock Climbing in World Coordinate/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..1cdc4f8dbeb4e9aac269a4144e0dc61a72a07e49 --- /dev/null +++ b/2025/ClimbingCap_ Multi-Modal Dataset and Method for Rock Climbing in World Coordinate/layout.json @@ -0,0 +1,9744 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 87, + 103, + 507, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 87, + 103, + 507, + 138 + ], + "spans": [ + { + "bbox": [ + 87, + 103, + 507, + 138 + ], + "type": "text", + "content": "ClimbingCap: Multi-Modal Dataset and Method for Rock Climbing in World Coordinate" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 59, + 148, + 523, + 178 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 148, + 523, + 178 + ], + "spans": [ + { + "bbox": [ + 59, + 148, + 523, + 178 + ], + "type": "text", + "content": "Ming Yan" + }, + { + "bbox": [ + 59, + 148, + 523, + 178 + ], + "type": "inline_equation", + "content": "^{1,2,3*}" + }, + { + "bbox": [ + 59, + 148, + 523, + 178 + ], + "type": "text", + "content": " Xincheng Lin" + }, + { + "bbox": [ + 59, + 148, + 523, + 178 + ], + "type": "inline_equation", + "content": "^{1,3*}" + }, + { + "bbox": [ + 59, + 148, + 523, + 178 + ], + "type": "text", + "content": " Yuhua Luo" + }, + { + "bbox": [ + 59, + 148, + 523, + 178 + ], + "type": "inline_equation", + "content": "^{1,3}" + }, + { + "bbox": [ + 59, + 148, + 523, + 178 + ], + "type": "text", + "content": " Shuqi Fan" + }, + { + "bbox": [ + 59, + 148, + 523, + 178 + ], + "type": "inline_equation", + "content": "^{1,3}" + }, + { + "bbox": [ + 59, + 148, + 523, + 178 + ], + "type": "text", + "content": " Yudi Dai" + }, + { + "bbox": [ + 59, + 148, + 523, + 178 + ], + "type": "inline_equation", + "content": "^{7}" + }, + { + "bbox": [ + 59, + 148, + 523, + 178 + ], + "type": "text", + "content": " Qixin Zhong" + }, + { + "bbox": [ + 59, + 148, + 523, + 178 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 59, + 148, + 523, + 178 + ], + "type": "text", + "content": " Lincai Zhong" + }, + { + "bbox": [ + 59, + 148, + 523, + 178 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 59, + 148, + 523, + 178 + ], + "type": "text", + "content": " Yuexin Ma" + }, + { + "bbox": [ + 59, + 148, + 523, + 178 + ], + "type": "inline_equation", + "content": "^{6}" + }, + { + "bbox": [ + 59, + 148, + 523, + 178 + ], + "type": "text", + "content": " Lan Xu" + }, + { + "bbox": [ + 59, + 148, + 523, + 178 + ], + "type": "inline_equation", + "content": "^{6}" + }, + { + "bbox": [ + 59, + 148, + 523, + 178 + ], + "type": "text", + "content": " Chenglu Wen" + }, + { + "bbox": [ + 59, + 148, + 523, + 178 + ], + "type": "inline_equation", + "content": "^{1,3}" + }, + { + "bbox": [ + 59, + 148, + 523, + 178 + ], + "type": "text", + "content": " Siqi Shen" + }, + { + "bbox": [ + 59, + 148, + 523, + 178 + ], + "type": "inline_equation", + "content": "^{1,3\\dagger}" + }, + { + "bbox": [ + 59, + 148, + 523, + 178 + ], + "type": "text", + "content": " Cheng Wang" + }, + { + "bbox": [ + 59, + 148, + 523, + 178 + ], + "type": "inline_equation", + "content": "^{1,3}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 85, + 180, + 507, + 195 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 180, + 507, + 195 + ], + "spans": [ + { + "bbox": [ + 85, + 180, + 507, + 195 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 85, + 180, + 507, + 195 + ], + "type": "text", + "content": "Fujian Key Laboratory of Sensing and Computing for Smart Cities, Xiamen University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 102, + 195, + 489, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 195, + 489, + 209 + ], + "spans": [ + { + "bbox": [ + 102, + 195, + 489, + 209 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 102, + 195, + 489, + 209 + ], + "type": "text", + "content": "National Institute for Data Science in Health and Medicine, Xiamen University" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 111, + 209, + 482, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 209, + 482, + 223 + ], + "spans": [ + { + "bbox": [ + 111, + 209, + 482, + 223 + ], + "type": "text", + "content": "3Key Laboratory of Multimedia Trusted Perception and Efficient Computing," + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 117, + 224, + 476, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 117, + 224, + 476, + 236 + ], + "spans": [ + { + "bbox": [ + 117, + 224, + 476, + 236 + ], + "type": "text", + "content": "Ministry of Education of China, School of Informatics, Xiamen University" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 126, + 237, + 468, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 126, + 237, + 468, + 251 + ], + "spans": [ + { + "bbox": [ + 126, + 237, + 468, + 251 + ], + "type": "text", + "content": "4China National Climbing Team 5Ningbo Sports Work Training Team" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 159, + 251, + 434, + 265 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 251, + 434, + 265 + ], + "spans": [ + { + "bbox": [ + 159, + 251, + 434, + 265 + ], + "type": "inline_equation", + "content": "^{6}" + }, + { + "bbox": [ + 159, + 251, + 434, + 265 + ], + "type": "text", + "content": "ShanghaiTech University ETH AI Center, ETH Zürich" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 48, + 270, + 544, + 426 + ], + "blocks": [ + { + "bbox": [ + 48, + 270, + 544, + 426 + ], + "lines": [ + { + "bbox": [ + 48, + 270, + 544, + 426 + ], + "spans": [ + { + "bbox": [ + 48, + 270, + 544, + 426 + ], + "type": "image", + "image_path": "c3ad0d134fecc3ddf3c7919012bf63efd650c3a94edc4a9feb043c105f06e1fa.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 429, + 546, + 483 + ], + "lines": [ + { + "bbox": [ + 46, + 429, + 546, + 483 + ], + "spans": [ + { + "bbox": [ + 46, + 429, + 546, + 483 + ], + "type": "text", + "content": "Figure 1. Overview. To address the challenging problem of global climbing motion recovery, we collect the dataset AscendMotion, using LiDAR, RGB camera and Inertial Measurement Unit (IMU) motion capture system with accurate motion labels and global trajectories (the blue and orange human bodies in right side of the figure represent labeled motions, and the orange curve represents the motion trajectory in the world coordinate.). Meanwhile, we propose ClimbingCap, a global climbing motion capturing method in world coordinate. As shown in the left part of this figure, it uses both image and LiDAR point cloud to recover human motions." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 143, + 494, + 192, + 506 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 494, + 192, + 506 + ], + "spans": [ + { + "bbox": [ + 143, + 494, + 192, + 506 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 518, + 288, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 518, + 288, + 686 + ], + "spans": [ + { + "bbox": [ + 46, + 518, + 288, + 686 + ], + "type": "text", + "content": "Human Motion Recovery (HMR) research mainly focuses on ground-based motions such as running. The study on capturing climbing motion, an off-ground motion, is sparse. This is partly due to the limited availability of climbing motion datasets, especially large-scale and challenging 3D labeled datasets. To address the insufficiency of climbing motion datasets, we collect AscendMotion, a large-scale well-annotated, and challenging climbing motion dataset. It consists of " + }, + { + "bbox": [ + 46, + 518, + 288, + 686 + ], + "type": "inline_equation", + "content": "412k" + }, + { + "bbox": [ + 46, + 518, + 288, + 686 + ], + "type": "text", + "content": " RGB, LiDAR frames, and IMU measurements, including the challenging climbing motions of 22 skilled climbing coaches across 12 different rock walls. Capturing the climbing motions is challenging as it requires precise recovery of not only the complex pose but also the global position of climbers. Although multi" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 495, + 546, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 495, + 546, + 639 + ], + "spans": [ + { + "bbox": [ + 306, + 495, + 546, + 639 + ], + "type": "text", + "content": "ple global HMR methods have been proposed, they cannot faithfully capture climbing motions. To address the limitations of HMR methods for climbing, we propose ClimbingCap, a motion recovery method that reconstructs continuous 3D human climbing motion in a global coordinate system. One key insight is to use the RGB and LiDAR modalities to separately reconstruct motions in camera coordinates and global coordinates and to optimize them jointly. We demonstrate the quality of the AscendMotion dataset and present promising results from ClimbingCap. The AscendMotion dataset and source code release publicly at http://www.lidarhumanmotion.net/climbingcap/" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 645, + 387, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 645, + 387, + 658 + ], + "spans": [ + { + "bbox": [ + 306, + 645, + 387, + 658 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 665, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 665, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 665, + 546, + 713 + ], + "type": "text", + "content": "It is challenging for global Human Motion Recovery (HMR) [1, 9, 40, 42, 80] due to the complexity of human poses and dynamic interaction between humans and their environments. Moreover, global HMR must main" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 693, + 132, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 693, + 132, + 703 + ], + "spans": [ + { + "bbox": [ + 62, + 693, + 132, + 703 + ], + "type": "text", + "content": "* Equal contribution." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 63, + 703, + 141, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 703, + 141, + 713 + ], + "spans": [ + { + "bbox": [ + 63, + 703, + 141, + 713 + ], + "type": "text", + "content": "† Corresponding author." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "12312" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 167 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 167 + ], + "type": "text", + "content": "tain consistency in world coordinates to ensure authenticity and physical feasibility. The research community has proposed various methods to estimate human poses from different sensors such as RGB images [21-23, 33, 71], LiDAR point clouds [36, 49, 78], and inertial measurement units (IMUs) [25, 76]. Most of the studies [8, 15, 16, 24, 28, 39, 49, 62, 68, 73, 78] focus on recovering ground-based motions such as running, interacting, and dancing." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 170, + 289, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 170, + 289, + 326 + ], + "spans": [ + { + "bbox": [ + 46, + 170, + 289, + 326 + ], + "type": "text", + "content": "Unlike ground-based motions such as running, climbing is an activity performed off-ground, where climbers use hands and feet to ascend holds or walls. As an important sport and recreation, the climbing race has become an official event in the Olympic Games. Yet, research on climbing motion capture [26, 34, 35, 54, 61] is sparse, partly due to the limited availability of climbing datasets. There are only two publicly available climbing datasets: SPEEDD21 [19] and CIMI4D [72]. SPEEDD21 is a 2D dataset. Although CIMI4D is 3D, it consists of many trivial climbing motions from casual climbers. Moreover, the size of the two datasets is small. As a result, the research community lacks an in-depth understanding of challenging climbing motions." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 327, + 289, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 327, + 289, + 459 + ], + "spans": [ + { + "bbox": [ + 46, + 327, + 289, + 459 + ], + "type": "text", + "content": "We collect AscendMotion, a large-scale and challenging climbing motion multi-modal dataset, to address the data insufficiency. AscendMotion dataset includes the motions of 22 skilled climbers climbing 12 different rock walls. It includes rich modalities from hardware timesynchronized RGB, LiDAR point clouds, and Inertial Measurement Unit(IMU) MoCap system. Moreover, it consists of the global trajectories of climbers, which are important for understanding human motions in world coordinates. We also combine automatic annotation with manual refinement to ensure the accuracy of the motion annotation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 460, + 289, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 460, + 289, + 664 + ], + "spans": [ + { + "bbox": [ + 46, + 460, + 289, + 664 + ], + "type": "text", + "content": "Not only does climbing motion challenge humans, but capturing climbing motion is also challenging for researchers. Capturing climbing motion requires precise estimation of the complex pose resulting from human-scene interactions, as well as the global localization of climbers within scenes. Existing research dedicated to climbing [8, 62, 68] uses standard 2D HMR (e.g. OpenPose [9]) for climbing motions, which is suboptimal. Most existing HMR methods [20, 37, 55, 56, 60, 75] focus on recovering ground-based human motions through RGB imagery. The estimation results of these methods cannot be rigidly transformed among camera coordinates and global coordinate systems, due to the inherent ambiguity between coordinate systems. Moreover, global HMR methods often accumulate errors during long-term sequence recovery, affecting the accuracy of global trajectories and postures, especially in climbing movements, as we show through experiments." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 665, + 289, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 289, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 289, + 713 + ], + "type": "text", + "content": "To tackle the above challenges and fill the gap in the motion capture community for climbing motions, we introduce ClimbingCap, an HMR method that reconstructs continuous 3D human climbing motion in both camera and" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 72, + 547, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 227 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 227 + ], + "type": "text", + "content": "global coordinate systems (see Fig.1). We adopt a trilogy in ClimbingCap: separate coordinate decoding, postprocessing, and semi-supervised training. For the separate coordinate decoding stage, ClimbingCap uses image and point cloud modality to estimate poses in the camera and the global coordinate systems, respectively. The postprocessing stage accurately ensures the consistency of motions between the two coordinate systems. In the semi-supervised training stage, a teacher-student training method is adopted to make use of easily obtainable unlabeled climbing data. Through these innovations, we can handle complex climbing movements effectively. To summarize, our main contributions include:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 228, + 545, + 336 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 306, + 228, + 545, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 228, + 545, + 264 + ], + "spans": [ + { + "bbox": [ + 306, + 228, + 545, + 264 + ], + "type": "text", + "content": "- We collect the dataset AscendMotion, which is more comprehensive than existing datasets. We demonstrate the quality of the dataset and its poses are challenging." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 264, + 545, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 264, + 545, + 289 + ], + "spans": [ + { + "bbox": [ + 306, + 264, + 545, + 289 + ], + "type": "text", + "content": "- We propose ClimbingCap, which is a multimodality global HMR method for rock climbing." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 289, + 545, + 336 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 289, + 545, + 336 + ], + "spans": [ + { + "bbox": [ + 306, + 289, + 545, + 336 + ], + "type": "text", + "content": "- We conduct extensive experiments on evaluate the ClimbingCap in multiple datasets with various state-of-the-art methods. The experimental results show ClimbingCap perform better than all the methods for climbing motions." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 348, + 392, + 361 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 348, + 392, + 361 + ], + "spans": [ + { + "bbox": [ + 306, + 348, + 392, + 361 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 371, + 547, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 371, + 547, + 456 + ], + "spans": [ + { + "bbox": [ + 304, + 371, + 547, + 456 + ], + "type": "text", + "content": "Camera-Space HMR. Camera-space HMR methods [7, 12, 13, 18, 27, 30, 31, 38, 44, 46, 48, 52, 53, 57-59, 63, 65, 70, 74, 81] predict human motion in independent camera coordinate system for each frame of the input video. Although camera-space methods can predict accurate human poses, they cannot effectively reflect the orientation and trajectory of the human body in world coordinates." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 458, + 547, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 458, + 547, + 590 + ], + "spans": [ + { + "bbox": [ + 304, + 458, + 547, + 590 + ], + "type": "text", + "content": "World-Grounded HMR. Recent studies [32, 37, 55, 56, 60, 75, 77] have achieved remarkable progress in world coordinates. Most of them are two-stage methods. The first stage estimates human motion in camera coordinates, and the second stage optimizes the motion in world coordinates with the scene information. Most global HMR methods focus on ground-based motions (e.g., walking) that are located on the ground. They cannot accurately model climbing motions, which is an off-ground motion. Our method, ClimbingCap, is a global HMR method developed for recovering rock climbing motions." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 594, + 548, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 548, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 548, + 715 + ], + "type": "text", + "content": "HMR Dataset. We have compared the AscendMotion dataset with multiple recent datasets in Table Tab. 1. Most of the datasets focus on ground-based human daily activities or sports. There exist only two public available climbing motion datasets. SPEED21 [19] collects 2D motions of climbers from sports videos. CIMI4D [72] contains 3D motions and global trajectories of climbers. Our climbing dataset, AscendMotion (344 minutes labeled data and 441 minutes unlabeled data, 412k frames), is significantly larger than both SPEED21 (21 minutes, 46k frames) and CIMI4D" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "12313" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 70, + 547, + 262 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 547, + 262 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 547, + 262 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 547, + 262 + ], + "type": "table", + "html": "
DatasetSensor ModalitiesGlobal TrajectoryFrames3D SceneMotionReal/SyntheticSeqsSubjectsCareer
RGBMoCapLiDAR
SLOPER4D [15]✓IMU✓✓100k✓DaliyReal1512Normal
EMDB[28]✓EM-✓105k-DaliyReal8110Normal
BEHAVE [5]✓--✓15k-InteractionsReal-8Normal
RICH [24]✓--✓577k✓InteractionsReal14222Normal
AGORA [45]✓--✓106.7k✓DaliySynthetic---
BEDLAM [6]✓---1M✓DaliySynthetic---
FreeMotion [49]✓IMU✓✓578k-DaliyReal--Normal
HmPEAR [39]✓-✓-300k-DaliyReal6k25Normal
RELI11D [73]✓IMU✓✓239k✓SportReal4810Normal
HiSC4D [16]✓IMU✓✓36k✓SportReal8-Normal
LiDARHuman51M [78]✓IMU✓-374k-DailyReal5210Normal
CIMI4D [72]✓IMU✓✓180k✓ClimbingReal4212Normal
SPEED21 [19]✓--46kClimbingReal95--
AscendMotion(Ours)✓IMU✓✓412k✓(12)ClimbingReal22022Skilled
", + "image_path": "dbc930d8732376cdca91189e9da57bec65c520c9b378a07cf4b9dff33d21341d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 93, + 270, + 499, + 281 + ], + "lines": [ + { + "bbox": [ + 93, + 270, + 499, + 281 + ], + "spans": [ + { + "bbox": [ + 93, + 270, + 499, + 281 + ], + "type": "text", + "content": "Table 1. Comparisons with related HMR datasets. The ”-” symbol indicates that it is not included in the dataset." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 297, + 287, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 297, + 287, + 346 + ], + "spans": [ + { + "bbox": [ + 46, + 297, + 287, + 346 + ], + "type": "text", + "content": "(120 minutes, 180k frames). Moreover, the motions in our dataset are more challenging than those in CIMI4D, as they come from climbing coaches who are more skilled than the casual climbers in CIMI4D." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 349, + 288, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 349, + 288, + 456 + ], + "spans": [ + { + "bbox": [ + 46, + 349, + 288, + 456 + ], + "type": "text", + "content": "Climbing Motion Recovery. In recent years, the increasing popularity of climbing, coupled with the inclusion of rock climbing races in the Olympic Games, has led to growing interest in the capture and analysis of climbing motions [2, 3, 10, 51, 67]. Research on climbing MoCap methods remains limited. [43] employs OpenPose [9] for climbing pose estimation. [50] performs RGB-based method for local pose and estimates climbing position through a marker-based method." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 471, + 179, + 485 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 471, + 179, + 485 + ], + "spans": [ + { + "bbox": [ + 47, + 471, + 179, + 485 + ], + "type": "text", + "content": "3. Method: ClimbingCap" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 492, + 287, + 659 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 492, + 287, + 659 + ], + "spans": [ + { + "bbox": [ + 46, + 492, + 287, + 659 + ], + "type": "text", + "content": "Climbing motion capture is challenging, involving poses with extreme limb extension and full-body exertion in camera coordinates. Moreover, it requires precise alignment with the rock wall in world coordinate as climbers are ascending. The ClimbingCap pipeline consists of three parts: separate coordinate decoding, post-processing, and semi-supervised training. An overview of the proposed pipeline is shown in Fig. 2. The separate coordinate decoding and post-processing parts take into account the unique challenges posed by climbing motion, which involves complex off-ground dynamics and interactions with scenes. The semi-supervised training part makes uses of the large-scale unlabeled climbing motion data to learn better a HMR model." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 670, + 214, + 683 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 670, + 214, + 683 + ], + "spans": [ + { + "bbox": [ + 47, + 670, + 214, + 683 + ], + "type": "text", + "content": "3.1. Separate Coordinate Decoding" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 689, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 715 + ], + "type": "text", + "content": "The separate coordinate decoding (SCD) stage extracts features from the RGB sequence and LiDAR point clouds, and" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 297, + 545, + 321 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 297, + 545, + 321 + ], + "spans": [ + { + "bbox": [ + 304, + 297, + 545, + 321 + ], + "type": "text", + "content": "predicts the poses in camera coordinates and the positions in global coordinates." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 324, + 546, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 324, + 546, + 480 + ], + "spans": [ + { + "bbox": [ + 304, + 324, + 546, + 480 + ], + "type": "text", + "content": "Input and Feature Extraction. The overall network structure is illustrated in Fig. 2. The input includes RGB images and point cloud data. First, the point cloud data is transformed from the world coordinate system to the camera coordinate system via an extrinsic matrix, represented as " + }, + { + "bbox": [ + 304, + 324, + 546, + 480 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{\\mathrm{c}} = \\Omega_{w2c} \\cdot \\mathcal{P}_{\\mathrm{w}}" + }, + { + "bbox": [ + 304, + 324, + 546, + 480 + ], + "type": "text", + "content": ". Subsequently, the RGB images and transformed point cloud data are passed through feature extraction modules, RGB Extract and PC Extract, to obtain visual and geometric features. We build the feature extraction modules based on ViT [17] and PointNet++ [47]. These features are then fed into the following two decoder modules, which regress the SMPL parameters and global motion parameters of the human body, respectively." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 483, + 547, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 483, + 547, + 626 + ], + "spans": [ + { + "bbox": [ + 304, + 483, + 547, + 626 + ], + "type": "text", + "content": "Camera Coordinate Decoder. This module decodes the SMPL parameters in the camera coordinate system. The RGB and point cloud features serve as inputs to the Camera Coordinate Decoder (denoted as " + }, + { + "bbox": [ + 304, + 483, + 547, + 626 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{\\mathrm{Decoder}}" + }, + { + "bbox": [ + 304, + 483, + 547, + 626 + ], + "type": "text", + "content": "), which processes the inputs with contextual information " + }, + { + "bbox": [ + 304, + 483, + 547, + 626 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{\\mathrm{backbone}}" + }, + { + "bbox": [ + 304, + 483, + 547, + 626 + ], + "type": "text", + "content": ", generating an output token " + }, + { + "bbox": [ + 304, + 483, + 547, + 626 + ], + "type": "inline_equation", + "content": "\\mathbf{t}_{\\mathrm{out}}" + }, + { + "bbox": [ + 304, + 483, + 547, + 626 + ], + "type": "text", + "content": ". This output token is then used to iteratively optimize the SMPL parameters, including the pose " + }, + { + "bbox": [ + 304, + 483, + 547, + 626 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 304, + 483, + 547, + 626 + ], + "type": "text", + "content": ", shape " + }, + { + "bbox": [ + 304, + 483, + 547, + 626 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 304, + 483, + 547, + 626 + ], + "type": "text", + "content": ", and camera translation " + }, + { + "bbox": [ + 304, + 483, + 547, + 626 + ], + "type": "inline_equation", + "content": "\\Delta c" + }, + { + "bbox": [ + 304, + 483, + 547, + 626 + ], + "type": "text", + "content": ". The iterative decoding approach allows the model to gradually approximate the true pose and shape in the camera coordinate system. In each iteration, the decoder updates the current SMPL parameters " + }, + { + "bbox": [ + 304, + 483, + 547, + 626 + ], + "type": "inline_equation", + "content": "\\theta_{i}, \\beta_{i}" + }, + { + "bbox": [ + 304, + 483, + 547, + 626 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 483, + 547, + 626 + ], + "type": "inline_equation", + "content": "\\Delta c_{t}" + }, + { + "bbox": [ + 304, + 483, + 547, + 626 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 370, + 634, + 545, + 647 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 634, + 545, + 647 + ], + "spans": [ + { + "bbox": [ + 370, + 634, + 545, + 647 + ], + "type": "interline_equation", + "content": "\\mathbf {t} _ {\\text {o u t}} = \\mathcal {T} _ {\\text {D e c o d e r}} (\\mathbf {t}, \\mathbf {f} _ {\\text {b a c k b o n e}}), \\tag {1}", + "image_path": "7b6b41270dd9fdd60d2ad3ccb008fc724dfa5b1f7b15f833b3656800d3e8b8ea.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 653, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 653, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 653, + 547, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 653, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\theta_{i + 1} = \\Phi_{\\theta}\\cdot \\mathbf{t}_{\\mathrm{out}} + \\theta_{i},\\beta_{i + 1} = \\Phi_{\\beta}\\cdot \\mathbf{t}_{\\mathrm{out}} + \\beta_{i}" + }, + { + "bbox": [ + 305, + 653, + 547, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 653, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\Delta c_{i + 1} = \\Phi_c\\cdot \\mathbf{t}_{\\mathrm{out}} + \\Delta c_i" + }, + { + "bbox": [ + 305, + 653, + 547, + 713 + ], + "type": "text", + "content": " are the update equations, with " + }, + { + "bbox": [ + 305, + 653, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\Phi_{\\theta}" + }, + { + "bbox": [ + 305, + 653, + 547, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 653, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\Phi_{\\beta}" + }, + { + "bbox": [ + 305, + 653, + 547, + 713 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 305, + 653, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\Phi_c" + }, + { + "bbox": [ + 305, + 653, + 547, + 713 + ], + "type": "text", + "content": " representing the respective weight matrices for each parameter. Here, the input token " + }, + { + "bbox": [ + 305, + 653, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{t}" + }, + { + "bbox": [ + 305, + 653, + 547, + 713 + ], + "type": "text", + "content": " can include initialized pose, shape, and camera parameters as needed." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "12314" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 71, + 329, + 232 + ], + "blocks": [ + { + "bbox": [ + 50, + 71, + 329, + 232 + ], + "lines": [ + { + "bbox": [ + 50, + 71, + 329, + 232 + ], + "spans": [ + { + "bbox": [ + 50, + 71, + 329, + 232 + ], + "type": "image", + "image_path": "b67825fb626a0f553f2c89c4583a420fc7d2f223b3664a0b47b0d3f1c9a777d1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 243, + 546, + 266 + ], + "lines": [ + { + "bbox": [ + 46, + 243, + 546, + 266 + ], + "spans": [ + { + "bbox": [ + 46, + 243, + 546, + 266 + ], + "type": "text", + "content": "Figure 2. Overview of ClimbingCap. The arrows indicate the three stages of the ClimbingCap framework: separate coordinate decoding (the green box), post-processing (the blue box), and semi-supervised training (the red box)." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 331, + 70, + 545, + 232 + ], + "blocks": [ + { + "bbox": [ + 331, + 70, + 545, + 232 + ], + "lines": [ + { + "bbox": [ + 331, + 70, + 545, + 232 + ], + "spans": [ + { + "bbox": [ + 331, + 70, + 545, + 232 + ], + "type": "image", + "image_path": "f2055e3ca958be05ee17c14dc0e93d409a5e8bc3b45085b1fc921b9637d45885.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 281, + 287, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 281, + 287, + 365 + ], + "spans": [ + { + "bbox": [ + 46, + 281, + 287, + 365 + ], + "type": "text", + "content": "Global Coordinate Decoder. To fully capture the human motion trajectory in the world coordinate, we design the Global Translation Decoder to predict the global translation parameters " + }, + { + "bbox": [ + 46, + 281, + 287, + 365 + ], + "type": "inline_equation", + "content": "\\Gamma^{\\mathrm{trans}}" + }, + { + "bbox": [ + 46, + 281, + 287, + 365 + ], + "type": "text", + "content": " of the human body. In this module, the decoder processes the features " + }, + { + "bbox": [ + 46, + 281, + 287, + 365 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{\\mathrm{backbone}}" + }, + { + "bbox": [ + 46, + 281, + 287, + 365 + ], + "type": "text", + "content": " as contextual input, iteratively updating the global translation parameters. The update formula in each iteration is given by:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 116, + 375, + 287, + 389 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 375, + 287, + 389 + ], + "spans": [ + { + "bbox": [ + 116, + 375, + 287, + 389 + ], + "type": "interline_equation", + "content": "\\Gamma_ {i + 1} ^ {\\text {t r a n s}} = \\Psi \\cdot \\mathbf {t} _ {\\text {o u t}} + \\Gamma_ {i} ^ {\\text {t r a n s}}, \\tag {2}", + "image_path": "61dbdb8cd20ee8dfeb1661a3921decb36cf75701fefda4b78fdc744db94f1133.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 401, + 287, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 401, + 287, + 450 + ], + "spans": [ + { + "bbox": [ + 46, + 401, + 287, + 450 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 401, + 287, + 450 + ], + "type": "inline_equation", + "content": "\\Gamma_{i}^{\\mathrm{trans}}" + }, + { + "bbox": [ + 46, + 401, + 287, + 450 + ], + "type": "text", + "content": " represents the global translation parameters at time step " + }, + { + "bbox": [ + 46, + 401, + 287, + 450 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 401, + 287, + 450 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 401, + 287, + 450 + ], + "type": "inline_equation", + "content": "\\Psi" + }, + { + "bbox": [ + 46, + 401, + 287, + 450 + ], + "type": "text", + "content": " is the weight matrix for the update. This decoding process enables the model to capture a complete motion trajectory in the global coordinate system." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 453, + 287, + 585 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 453, + 287, + 585 + ], + "spans": [ + { + "bbox": [ + 46, + 453, + 287, + 585 + ], + "type": "text", + "content": "Loss. The total loss function not only includes the 3D keypoint loss " + }, + { + "bbox": [ + 46, + 453, + 287, + 585 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{kp3d}}" + }, + { + "bbox": [ + 46, + 453, + 287, + 585 + ], + "type": "text", + "content": " and 2D keypoint loss " + }, + { + "bbox": [ + 46, + 453, + 287, + 585 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{kp2d}}" + }, + { + "bbox": [ + 46, + 453, + 287, + 585 + ], + "type": "text", + "content": " but also incorporates the SMPL parameter loss " + }, + { + "bbox": [ + 46, + 453, + 287, + 585 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^{\\mathrm{spl}}" + }, + { + "bbox": [ + 46, + 453, + 287, + 585 + ], + "type": "text", + "content": " and the global trajectory loss " + }, + { + "bbox": [ + 46, + 453, + 287, + 585 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{traj}}" + }, + { + "bbox": [ + 46, + 453, + 287, + 585 + ], + "type": "text", + "content": ". Specifically, the 3D keypoint loss " + }, + { + "bbox": [ + 46, + 453, + 287, + 585 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{kp3d}}" + }, + { + "bbox": [ + 46, + 453, + 287, + 585 + ], + "type": "text", + "content": " measures the 3D error of the predicted keypoints, while the 2D keypoint loss " + }, + { + "bbox": [ + 46, + 453, + 287, + 585 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{kp2d}}" + }, + { + "bbox": [ + 46, + 453, + 287, + 585 + ], + "type": "text", + "content": " measures the 2D projection error. The SMPL parameter loss " + }, + { + "bbox": [ + 46, + 453, + 287, + 585 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^{\\mathrm{spl}}" + }, + { + "bbox": [ + 46, + 453, + 287, + 585 + ], + "type": "text", + "content": " supervises the decoded pose and shape parameters, and the global trajectory loss " + }, + { + "bbox": [ + 46, + 453, + 287, + 585 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{traj}}" + }, + { + "bbox": [ + 46, + 453, + 287, + 585 + ], + "type": "text", + "content": " constrains the translation parameters to within close distance of ground truth positions. Please see the supplementary for details. The final total loss is formulated as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 63, + 597, + 287, + 612 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 597, + 287, + 612 + ], + "spans": [ + { + "bbox": [ + 63, + 597, + 287, + 612 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {k p 3 d} + \\mathcal {L} _ {k p 2 d} + \\mathcal {L} _ {\\theta} ^ {s m p l} + \\mathcal {L} _ {\\beta} ^ {s m p l} + \\mathcal {L} _ {t r a j}. \\tag {3}", + "image_path": "cff8c1c37ea06abf32e6668f3979c5fef77721cfad4b3cd584e34e2c9147bb68.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 634, + 143, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 634, + 143, + 647 + ], + "spans": [ + { + "bbox": [ + 47, + 634, + 143, + 647 + ], + "type": "text", + "content": "3.2. Post-processing" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 653, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 714 + ], + "type": "text", + "content": "Researches [55, 56, 75] have shown that a post-processing stage can be used to improve the output motion recovery results. Following these approaches, we employ a post-processing stage to optimize the output pose from SCD 3.1 stage." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 281, + 545, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 281, + 545, + 389 + ], + "spans": [ + { + "bbox": [ + 304, + 281, + 545, + 389 + ], + "type": "text", + "content": "One distinct advantage of ClimbingCap is that the output results for the pose decoding stage can be rigidly transformed between the camera coordinate system and the world coordinate system. Thanks to the LiDAR modality, the point cloud contains 3D information in the world coordinate system. The poses obtained from the SCD stage are converted from the camera coordinate system to the world coordinate system through the inverse extrinsic matrix " + }, + { + "bbox": [ + 304, + 281, + 545, + 389 + ], + "type": "inline_equation", + "content": "\\Omega_{w2c}^{-1}" + }, + { + "bbox": [ + 304, + 281, + 545, + 389 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 389, + 546, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 389, + 546, + 533 + ], + "spans": [ + { + "bbox": [ + 304, + 389, + 546, + 533 + ], + "type": "text", + "content": "The post-processing stage consists of three losses: " + }, + { + "bbox": [ + 304, + 389, + 546, + 533 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{LWD}" + }, + { + "bbox": [ + 304, + 389, + 546, + 533 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 389, + 546, + 533 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{SDS}" + }, + { + "bbox": [ + 304, + 389, + 546, + 533 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 389, + 546, + 533 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{VLR}" + }, + { + "bbox": [ + 304, + 389, + 546, + 533 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 304, + 389, + 546, + 533 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{LWD}" + }, + { + "bbox": [ + 304, + 389, + 546, + 533 + ], + "type": "text", + "content": " assigns different weights to the vertices of different parts of the climbing human SMPL, and optimizes the position of the human body in the world coordinate system according to these weights. " + }, + { + "bbox": [ + 304, + 389, + 546, + 533 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{SDS}" + }, + { + "bbox": [ + 304, + 389, + 546, + 533 + ], + "type": "text", + "content": " optimizes joint positions by smoothing the changes in the direction of the human body's velocity on the scene (i.e., rock wall). " + }, + { + "bbox": [ + 304, + 389, + 546, + 533 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{VLR}" + }, + { + "bbox": [ + 304, + 389, + 546, + 533 + ], + "type": "text", + "content": " optimizes the difficult-to-estimate limb end poses by using the position of the point cloud in space. In this stage, the global poses are optimized through using the Adam [29] optimizer. Please refer to the supplementary for detailed definition of these losses." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 539, + 449, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 539, + 449, + 552 + ], + "spans": [ + { + "bbox": [ + 306, + 539, + 449, + 552 + ], + "type": "text", + "content": "3.3. Semi-supervised Training" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 557, + 545, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 557, + 545, + 641 + ], + "spans": [ + { + "bbox": [ + 304, + 557, + 545, + 641 + ], + "type": "text", + "content": "Compared to ground-based motions [41], the size of labeled climbing motions is small. Simply using labeled climbing motion data may not be enough to train a robust model. Different from labeled climbing motion data, collecting unlabeled climbing motion data is cheaper. The AscendMotion dataset contains more unlabeled data than label data. They can be used to further improve the HMR model." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "content": "It has been shown by researches [64, 69, 79] from the object detection community that through using a teacher-student semi-supervised training framework, the performance of object detection models can be improved. We adopt such semi-supervised training frame for HMR. In this work, we refer the model trained after the SCD and the post" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12315" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 71, + 286, + 228 + ], + "blocks": [ + { + "bbox": [ + 50, + 71, + 286, + 228 + ], + "lines": [ + { + "bbox": [ + 50, + 71, + 286, + 228 + ], + "spans": [ + { + "bbox": [ + 50, + 71, + 286, + 228 + ], + "type": "image", + "image_path": "2f5bb39936921b1bc236f306567782bea86427b6529ebdb6bcc2199ba0be9fb2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 83, + 237, + 251, + 248 + ], + "lines": [ + { + "bbox": [ + 83, + 237, + 251, + 248 + ], + "spans": [ + { + "bbox": [ + 83, + 237, + 251, + 248 + ], + "type": "text", + "content": "Figure 3. Dataset collection hardware system." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 262, + 289, + 383 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 262, + 289, + 383 + ], + "spans": [ + { + "bbox": [ + 46, + 262, + 289, + 383 + ], + "type": "text", + "content": "processing stage as the teacher model (green box in Fig. 2). The student model (red box in Fig. 2) clones the parameters of the teacher model. During semi-supervised training, the teacher model estimate the pose labels from unlabeled motion data, and the pose label is used as pseudo-label to further train the student model. We show in the experiment section that via utilizing the semi-supervised training framework, the performance of HMR can be further improved. Please refer to the supplementary for the detailed training process." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 392, + 201, + 404 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 392, + 201, + 404 + ], + "spans": [ + { + "bbox": [ + 47, + 392, + 201, + 404 + ], + "type": "text", + "content": "4. The AscendMotion Dataset" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 413, + 287, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 413, + 287, + 581 + ], + "spans": [ + { + "bbox": [ + 46, + 413, + 287, + 581 + ], + "type": "text", + "content": "To advance research in Human Mesh Recovery (HMR) for climbing motion within a global coordinate system, we present the AscendMotion dataset. This dataset specifically captures complex multi-directional rock climbing motions on non-planar surfaces. It provides a unique opportunity for challenging HMR research. AscendMotion includes multi-modal motion data from skilled climbers performing various climbing styles, such as bouldering, speed climbing, and lead climbing, with synchronized recordings from IMU, RGB cameras, and LiDAR. The dataset comprises 344 minutes (6 hours) of labeled data and 441 minutes of unlabeled data, featuring 22 adult climbers across 12 real-world climbing scenes. All participants agreed to use their recorded data for scientific purposes." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 587, + 206, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 587, + 206, + 601 + ], + "spans": [ + { + "bbox": [ + 47, + 587, + 206, + 601 + ], + "type": "text", + "content": "4.1. Hardware and Configuration" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "text", + "content": "The AscendMotion data collection system integrates multiple sensors and captures motion data in both indoor and outdoor environments. As shown in Fig. 3, the LiDAR (OusterOS1, 128 beams) captures 3D dynamic point clouds at 20 frames per second (FPS), while the Hik 1080P RGB camera records RGB video at 20 FPS. For each climbing environment, we use the Trimble X7 3D laser scanner to reconstruct high-resolution RGB point cloud scenes, each containing approximately 80 million points. For the labeled" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 72, + 545, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 193 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 193 + ], + "type": "text", + "content": "dataset, each climber wears an Xsens MVN inertial motion capture system with 17 IMUs, recording at 60 FPS. [4] and a handheld point cloud scanner to ensure accurate body shape representation. For the unlabeled dataset, climbers do not wear a MoCap outfit, and we record their climbing motions using a RGB camera and a LiDAR. A human motion is represented by " + }, + { + "bbox": [ + 304, + 72, + 545, + 193 + ], + "type": "inline_equation", + "content": "M = (T, \\theta, \\beta)" + }, + { + "bbox": [ + 304, + 72, + 545, + 193 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 72, + 545, + 193 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 72, + 545, + 193 + ], + "type": "text", + "content": " represents global translation, " + }, + { + "bbox": [ + 304, + 72, + 545, + 193 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 304, + 72, + 545, + 193 + ], + "type": "text", + "content": " is the SMPL [40] pose parameters, and " + }, + { + "bbox": [ + 304, + 72, + 545, + 193 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 304, + 72, + 545, + 193 + ], + "type": "text", + "content": " is the SMPL shape parameter. We obtain each climber's shape parameters " + }, + { + "bbox": [ + 304, + 72, + 545, + 193 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 304, + 72, + 545, + 193 + ], + "type": "text", + "content": " using IPNet [4]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 201, + 422, + 214 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 201, + 422, + 214 + ], + "spans": [ + { + "bbox": [ + 305, + 201, + 422, + 214 + ], + "type": "text", + "content": "4.2. Annotation Pipeline" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 220, + 545, + 280 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 220, + 545, + 280 + ], + "spans": [ + { + "bbox": [ + 304, + 220, + 545, + 280 + ], + "type": "text", + "content": "The pose " + }, + { + "bbox": [ + 304, + 220, + 545, + 280 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 304, + 220, + 545, + 280 + ], + "type": "text", + "content": " and the translation " + }, + { + "bbox": [ + 304, + 220, + 545, + 280 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 220, + 545, + 280 + ], + "type": "text", + "content": " obtained from the IMU measurements may be inaccurate. Especially the translation " + }, + { + "bbox": [ + 304, + 220, + 545, + 280 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 220, + 545, + 280 + ], + "type": "text", + "content": " may significantly drift for long-duration capture. The annotation pipeline of AscendMotion is used to find the accurate translate " + }, + { + "bbox": [ + 304, + 220, + 545, + 280 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 220, + 545, + 280 + ], + "type": "text", + "content": " and pose " + }, + { + "bbox": [ + 304, + 220, + 545, + 280 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 304, + 220, + 545, + 280 + ], + "type": "text", + "content": " as labels." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 281, + 545, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 281, + 545, + 376 + ], + "spans": [ + { + "bbox": [ + 304, + 281, + 545, + 376 + ], + "type": "text", + "content": "The annotation pipeline of AscendMotion is depicted in Fig. 4. It consists of three parts: preprocessing, multistage global optimization, and manual annotation. The preprocessing part is used for time synchronization and spatial calibration. The multi-stage global optimization part is developed for automatic generation of annotation labels. The manual annotation part is used to further improve the quality of generated annotations." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 394, + 504, + 407 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 394, + 504, + 407 + ], + "spans": [ + { + "bbox": [ + 305, + 394, + 504, + 407 + ], + "type": "text", + "content": "4.2.1 Multi-modal Data Preprocessing Stage" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 417, + 545, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 417, + 545, + 489 + ], + "spans": [ + { + "bbox": [ + 304, + 417, + 545, + 489 + ], + "type": "text", + "content": "Time Synchronization. The time among RGB camera and LiDAR is synchronized via Precision Time Protocol (PTP). We employ CollShark Auto 66 unit as the master clock, and it sends PTP slave clocks to the RGB camera and LiDAR. The time of IMU MoCap is post-synchronized with the LiDAR and RGB through anchor frames." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 492, + 545, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 492, + 545, + 576 + ], + "spans": [ + { + "bbox": [ + 304, + 492, + 545, + 576 + ], + "type": "text", + "content": "Calibration. First, the LiDAR point cloud are registered with high-precision scanned-scenes. Next, the coordinate of LiDAR is treated as the world coordinate. The IMU measurements are transformed into the world coordinate through a calibration matrix. Finally, we perform frame-level calibration among RGB, LiDAR and IMU. Please refer to the appendix for details." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 594, + 478, + 606 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 594, + 478, + 606 + ], + "spans": [ + { + "bbox": [ + 305, + 594, + 478, + 606 + ], + "type": "text", + "content": "4.2.2 Multi-stage Global Optimization" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 614, + 545, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 614, + 545, + 686 + ], + "spans": [ + { + "bbox": [ + 304, + 614, + 545, + 686 + ], + "type": "text", + "content": "AscendMotion uses the translation " + }, + { + "bbox": [ + 304, + 614, + 545, + 686 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 614, + 545, + 686 + ], + "type": "text", + "content": " and pose " + }, + { + "bbox": [ + 304, + 614, + 545, + 686 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 304, + 614, + 545, + 686 + ], + "type": "text", + "content": " provided by the IMU MoCap as the initialization of annotation labels, and performs multi-stage global optimization. To achieve accurate and natural human motion data consistent with the scene, we apply two loss functions: the Global Refit Loss " + }, + { + "bbox": [ + 304, + 614, + 545, + 686 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{GR}}" + }, + { + "bbox": [ + 304, + 614, + 545, + 686 + ], + "type": "text", + "content": " and the Scene Touch Loss " + }, + { + "bbox": [ + 304, + 614, + 545, + 686 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{ST}}" + }, + { + "bbox": [ + 304, + 614, + 545, + 686 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "text", + "content": "Global Refit Loss " + }, + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{GR}}" + }, + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "text", + "content": ". The global refit loss promotes precise global alignment between the SMPL model and the" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "12316" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 80, + 170, + 198 + ], + "blocks": [ + { + "bbox": [ + 50, + 80, + 170, + 198 + ], + "lines": [ + { + "bbox": [ + 50, + 80, + 170, + 198 + ], + "spans": [ + { + "bbox": [ + 50, + 80, + 170, + 198 + ], + "type": "image", + "image_path": "cb659f075ccab75a8b7ef24ee341b2eee0dfa8721f3287452241e0c2a727fc21.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 209, + 546, + 232 + ], + "lines": [ + { + "bbox": [ + 46, + 209, + 546, + 232 + ], + "spans": [ + { + "bbox": [ + 46, + 209, + 546, + 232 + ], + "type": "text", + "content": "Figure 4. AscendMotion Annotation Pipeline. From left to right, this pipeline represents the three stages of dataset annotation: time-space synchronous input pre-processing(the blue box), multi-stage global optimization(the green box), and manual repair(the red box)." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 170, + 80, + 272, + 198 + ], + "blocks": [ + { + "bbox": [ + 178, + 70, + 283, + 80 + ], + "lines": [ + { + "bbox": [ + 178, + 70, + 283, + 80 + ], + "spans": [ + { + "bbox": [ + 178, + 70, + 283, + 80 + ], + "type": "text", + "content": "Multi-modal data preprocessing" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 170, + 80, + 272, + 198 + ], + "lines": [ + { + "bbox": [ + 170, + 80, + 272, + 198 + ], + "spans": [ + { + "bbox": [ + 170, + 80, + 272, + 198 + ], + "type": "image", + "image_path": "4eaebb54defeeede5abb455c3ee28a2042c923d3e256fb4ba42597f6c3737d9e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 273, + 80, + 376, + 198 + ], + "blocks": [ + { + "bbox": [ + 273, + 80, + 376, + 198 + ], + "lines": [ + { + "bbox": [ + 273, + 80, + 376, + 198 + ], + "spans": [ + { + "bbox": [ + 273, + 80, + 376, + 198 + ], + "type": "image", + "image_path": "f0efa8d3143d7e6e2957752f95e095641c12c19e46df544446a3ed86939e99c2.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 378, + 80, + 480, + 198 + ], + "blocks": [ + { + "bbox": [ + 378, + 70, + 482, + 80 + ], + "lines": [ + { + "bbox": [ + 378, + 70, + 482, + 80 + ], + "spans": [ + { + "bbox": [ + 378, + 70, + 482, + 80 + ], + "type": "text", + "content": "Multi-stage global optimization" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 378, + 80, + 480, + 198 + ], + "lines": [ + { + "bbox": [ + 378, + 80, + 480, + 198 + ], + "spans": [ + { + "bbox": [ + 378, + 80, + 480, + 198 + ], + "type": "image", + "image_path": "eaa8e9134b89f229f15dd2e6151ce8dd1bcdec8de604fb7faf664f190e59899a.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 481, + 80, + 544, + 198 + ], + "blocks": [ + { + "bbox": [ + 483, + 71, + 539, + 80 + ], + "lines": [ + { + "bbox": [ + 483, + 71, + 539, + 80 + ], + "spans": [ + { + "bbox": [ + 483, + 71, + 539, + 80 + ], + "type": "text", + "content": "Manual Repair" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 481, + 80, + 544, + 198 + ], + "lines": [ + { + "bbox": [ + 481, + 80, + 544, + 198 + ], + "spans": [ + { + "bbox": [ + 481, + 80, + 544, + 198 + ], + "type": "image", + "image_path": "194057910d2d7f7e7f177bed9f78643651b2d946d684688b6aab09ff446a2aee.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 252, + 289, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 252, + 289, + 420 + ], + "spans": [ + { + "bbox": [ + 46, + 252, + 289, + 420 + ], + "type": "text", + "content": "point clouds. " + }, + { + "bbox": [ + 46, + 252, + 289, + 420 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{GR}}" + }, + { + "bbox": [ + 46, + 252, + 289, + 420 + ], + "type": "text", + "content": " calculates the geometric discrepancy between the SMPL model vertices " + }, + { + "bbox": [ + 46, + 252, + 289, + 420 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 46, + 252, + 289, + 420 + ], + "type": "text", + "content": " and the human body point cloud " + }, + { + "bbox": [ + 46, + 252, + 289, + 420 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 46, + 252, + 289, + 420 + ], + "type": "text", + "content": " using a modified Chamfer distance. It applies different loss terms (" + }, + { + "bbox": [ + 46, + 252, + 289, + 420 + ], + "type": "inline_equation", + "content": "f(|v_{i} - p_{j}|^{2})" + }, + { + "bbox": [ + 46, + 252, + 289, + 420 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 46, + 252, + 289, + 420 + ], + "type": "inline_equation", + "content": "g(|v_{i} - p_{j}|^{2}))" + }, + { + "bbox": [ + 46, + 252, + 289, + 420 + ], + "type": "text", + "content": " to different body parts to make body regions, such as the torso and limbs, match the point clouds of the human. We calculate the Euclidean distance between a vertex " + }, + { + "bbox": [ + 46, + 252, + 289, + 420 + ], + "type": "inline_equation", + "content": "v_{i}\\in V" + }, + { + "bbox": [ + 46, + 252, + 289, + 420 + ], + "type": "text", + "content": " and a point " + }, + { + "bbox": [ + 46, + 252, + 289, + 420 + ], + "type": "inline_equation", + "content": "p_j\\in P" + }, + { + "bbox": [ + 46, + 252, + 289, + 420 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 46, + 252, + 289, + 420 + ], + "type": "inline_equation", + "content": "d_{v_i,p_j}" + }, + { + "bbox": [ + 46, + 252, + 289, + 420 + ], + "type": "text", + "content": ". A loss term " + }, + { + "bbox": [ + 46, + 252, + 289, + 420 + ], + "type": "inline_equation", + "content": "f(|v_i - p_j|^2)" + }, + { + "bbox": [ + 46, + 252, + 289, + 420 + ], + "type": "text", + "content": " is applied to a vertex " + }, + { + "bbox": [ + 46, + 252, + 289, + 420 + ], + "type": "inline_equation", + "content": "v_{i}" + }, + { + "bbox": [ + 46, + 252, + 289, + 420 + ], + "type": "text", + "content": " if " + }, + { + "bbox": [ + 46, + 252, + 289, + 420 + ], + "type": "inline_equation", + "content": "d_{v_i,p_j}\\leq d_{\\mathrm{torso}}" + }, + { + "bbox": [ + 46, + 252, + 289, + 420 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 252, + 289, + 420 + ], + "type": "inline_equation", + "content": "g(|p_j - v_i|^2)" + }, + { + "bbox": [ + 46, + 252, + 289, + 420 + ], + "type": "text", + "content": " is applied to " + }, + { + "bbox": [ + 46, + 252, + 289, + 420 + ], + "type": "inline_equation", + "content": "v_{i}" + }, + { + "bbox": [ + 46, + 252, + 289, + 420 + ], + "type": "text", + "content": " if " + }, + { + "bbox": [ + 46, + 252, + 289, + 420 + ], + "type": "inline_equation", + "content": "d_{v_i,p_j}\\leq d_{\\mathrm{limb}}" + }, + { + "bbox": [ + 46, + 252, + 289, + 420 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 252, + 289, + 420 + ], + "type": "inline_equation", + "content": "d_{\\mathrm{torso}}" + }, + { + "bbox": [ + 46, + 252, + 289, + 420 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 252, + 289, + 420 + ], + "type": "inline_equation", + "content": "d_{\\mathrm{limb}}" + }, + { + "bbox": [ + 46, + 252, + 289, + 420 + ], + "type": "text", + "content": " are the distance thresholds for torso and limbs, respectively. During climbing, humans use limbs to climb up, the distance among limbs and scenes should smaller than the distance among torso and scenes, " + }, + { + "bbox": [ + 46, + 252, + 289, + 420 + ], + "type": "inline_equation", + "content": "d_{\\mathrm{limb}}\\leq d_{\\mathrm{torso}}" + }, + { + "bbox": [ + 46, + 252, + 289, + 420 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 424, + 287, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 424, + 287, + 579 + ], + "spans": [ + { + "bbox": [ + 46, + 424, + 287, + 579 + ], + "type": "text", + "content": "Scene Touch Loss " + }, + { + "bbox": [ + 46, + 424, + 287, + 579 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{ST}}" + }, + { + "bbox": [ + 46, + 424, + 287, + 579 + ], + "type": "text", + "content": ". This loss prevents unrealistic intersections between the SMPL model and the scene mesh by measuring the penetration depth between the SMPL vertices " + }, + { + "bbox": [ + 46, + 424, + 287, + 579 + ], + "type": "inline_equation", + "content": "v_{i} \\in V" + }, + { + "bbox": [ + 46, + 424, + 287, + 579 + ], + "type": "text", + "content": " and the scene mesh vertices " + }, + { + "bbox": [ + 46, + 424, + 287, + 579 + ], + "type": "inline_equation", + "content": "q_{j} \\in Q" + }, + { + "bbox": [ + 46, + 424, + 287, + 579 + ], + "type": "text", + "content": ". Given the scene mesh normal vectors " + }, + { + "bbox": [ + 46, + 424, + 287, + 579 + ], + "type": "inline_equation", + "content": "n_{j}" + }, + { + "bbox": [ + 46, + 424, + 287, + 579 + ], + "type": "text", + "content": ", the penetration depth for each vertex is calculated as the dot product " + }, + { + "bbox": [ + 46, + 424, + 287, + 579 + ], + "type": "inline_equation", + "content": "\\eta(v_{i}) = (v_{i} - q_{j}) \\cdot n_{j}" + }, + { + "bbox": [ + 46, + 424, + 287, + 579 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 424, + 287, + 579 + ], + "type": "inline_equation", + "content": "q_{j}" + }, + { + "bbox": [ + 46, + 424, + 287, + 579 + ], + "type": "text", + "content": " is the closest mesh vertex to " + }, + { + "bbox": [ + 46, + 424, + 287, + 579 + ], + "type": "inline_equation", + "content": "v_{i}" + }, + { + "bbox": [ + 46, + 424, + 287, + 579 + ], + "type": "text", + "content": ". If the penetration depth " + }, + { + "bbox": [ + 46, + 424, + 287, + 579 + ], + "type": "inline_equation", + "content": "\\eta(v_{i})" + }, + { + "bbox": [ + 46, + 424, + 287, + 579 + ], + "type": "text", + "content": " is negative, it indicates that the vertex " + }, + { + "bbox": [ + 46, + 424, + 287, + 579 + ], + "type": "inline_equation", + "content": "v_{i}" + }, + { + "bbox": [ + 46, + 424, + 287, + 579 + ], + "type": "text", + "content": " has penetrated the scene mesh, and this value contributes to the loss " + }, + { + "bbox": [ + 46, + 424, + 287, + 579 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{ST}}" + }, + { + "bbox": [ + 46, + 424, + 287, + 579 + ], + "type": "text", + "content": ". This loss encourages the SMPL model to avoid intersecting with the scene mesh, promoting more realistic interactions between the model and the environment." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 597, + 234, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 597, + 234, + 609 + ], + "spans": [ + { + "bbox": [ + 47, + 597, + 234, + 609 + ], + "type": "text", + "content": "4.2.3 Manual Annotation and Verification" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 617, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 287, + 715 + ], + "type": "text", + "content": "To further improve the dataset quality, we use the SMPL annotation tool [66] for enhanced labeling of sequences. We engage four observers to review the automatically annotated results. For any ambiguous limb movements observed in RGB views, we perform additional manual adjustments and mark these frames as key frames. Finally, we propagate poses from multiple key frames within each sequence to ensure global consistency in both smoothness and accuracy." + } + ] + } + ], + "index": 12 + }, + { + "type": "table", + "bbox": [ + 308, + 250, + 545, + 306 + ], + "blocks": [ + { + "bbox": [ + 308, + 250, + 545, + 306 + ], + "lines": [ + { + "bbox": [ + 308, + 250, + 545, + 306 + ], + "spans": [ + { + "bbox": [ + 308, + 250, + 545, + 306 + ], + "type": "table", + "html": "
Constraint termHorizontal SceneVertical Scene
\\( \\mathcal{L}_{ST} \\)\\( \\mathcal{L}_{GR} \\)ACCEL↓MPJPE↓PA-MPJPE↓ACCEL↓MPJPE↓PA-MPJPE↓
XX1.8337.5732.526.7951.5146.24
✓X1.6735.4531.813.7340.3635.37
X✓1.5929.0425.951.6629.2424.99
✓✓1.5428.2023.251.6029.0724.37
", + "image_path": "26e84c208b3581fef37bcd1ab1f9de38c0053493141d6489982ff09da2e707ce.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 331, + 315, + 520, + 326 + ], + "lines": [ + { + "bbox": [ + 331, + 315, + 520, + 326 + ], + "spans": [ + { + "bbox": [ + 331, + 315, + 520, + 326 + ], + "type": "text", + "content": "Table 2. Quantitative Evaluation of Dataset Quality." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 306, + 350, + 409, + 363 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 350, + 409, + 363 + ], + "spans": [ + { + "bbox": [ + 306, + 350, + 409, + 363 + ], + "type": "text", + "content": "5. EXPERIMENTS" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 373, + 545, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 373, + 545, + 482 + ], + "spans": [ + { + "bbox": [ + 304, + 373, + 545, + 482 + ], + "type": "text", + "content": "In this section, we demonstrate the high-quality of the AscendMotion optimization process in Sec. 5.1, show that the AscendMotion dataset is challenging to existing HMR methods, and evaluate ClimbingCap and multiple HMR methods in climbing motion datasets in Sec. 5.2. We demonstrate the promising performance of ClimbingCap and shows that each component of ClimbingCap is useful through ablation study in Sec. 5.3. Please refer to the detailed experimental setup and metrics in the supplementary." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 496, + 422, + 508 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 496, + 422, + 508 + ], + "spans": [ + { + "bbox": [ + 306, + 496, + 422, + 508 + ], + "type": "text", + "content": "5.1. Datasets Evaluation" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 517, + 545, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 517, + 545, + 576 + ], + "spans": [ + { + "bbox": [ + 304, + 517, + 545, + 576 + ], + "type": "text", + "content": "To quantitatively evaluate the annotation quality of AscendMotion, we divide the scene into horizontal and vertical rock walls and evaluate the performance of the global optimization stage (in Sec. 4.2.2) by comparing the generated annotations with manual annotations." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 579, + 545, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 579, + 545, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 579, + 545, + 628 + ], + "type": "text", + "content": "The evaluation metrics include widely used MPJPE, Procrustes Aligned MPJPE (PA-MPJPE), per vertex error (PVE), acceleration error (Accel, in units of " + }, + { + "bbox": [ + 304, + 579, + 545, + 628 + ], + "type": "inline_equation", + "content": "m / s^2" + }, + { + "bbox": [ + 304, + 579, + 545, + 628 + ], + "type": "text", + "content": "). These metrics are used for evaluating HMR methods as well." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": "To understand the impact of different constraints used in the combined optimization stage, we perform an ablation study on two different losses: " + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{ST}" + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{GR}" + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": ". Tab. 2 shows the error metrics using different loss combinations for two scenes. The error metrics are small, demonstrating the effectiveness of the global optimization pipeline and the high quality of AscendMotion." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12317" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 70, + 547, + 175 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 547, + 175 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 547, + 175 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 547, + 175 + ], + "type": "table", + "html": "
ModalityMethodCamera CoordinateWorld Coordinate
ACCEL↓MPJPE↓PA-MPJPE↓PVE↓PCK0.3↑WA-MPJPE↓W-MPJPE↓RTE↓Jitter↓T-Error↓
RGBTRACE [60]18.68/76.79875.56/577.6069.21/85.81951.52/619.930.06/0.09144.33/385.71254.38/703.3514.73/26.17115.96/521.402.56/6.62
SLAHMR [75]5.46/96.98232.46/467.8884.13/285.15283.24/552.630.36/0.12277.47/447.68804.85/613.603.64/39.394.91/201.332.81/6.54
WHAM [56]4.59/35.01110.92/143.1776.09/73.36124.2/164.910.76/0.62229.42/1125.77647.70/1499.855.16/9.043.58/40.691.77/2.49
GVHMR [55]4.50/26.22107.09/124.6060.06/80.30118.89/151.100.77/0.71105.15/1002.11202.45/1442.504.09/7.916.85/32.711.48/2.54
LiDARLiDARCapV2 [78]87.99/119.62244.6/234.52192.17/156.39326.45/283.270.53/0.50282.12/1396.42442.12/1518.2916.42/10.85176.95/165.551.65/2.89
LiveHPS [49]157.87/195.23156.5/147.31142.19/121.76191.87/189.300.64/0.70235.4/1369.89392.34/1506.5013.94/10.45279.96/358.542.1/6.73
LiDAR +RGBImmFusion [11]108.76/74.20473.18/464.83254.07/179.51533.5/529.710.17/0.14324.4/1446.01487.92/1532.8816.52/10.8627.03/14.491.94/3.99
FusionPose [14]112.08/86.44256.81/315.93198.55/193.83306.22/359.680.36/0.42275.02/1445.32444.47/1532.2816.29/10.8592.97/80.792.02/6.48
LEIR [73]110.18/94.57297.95/299.62187.26/150.56340.61/351.520.41/0.37266.82/1313.09282.31/1435.929.78/9.9773.38/85.031.1/1.20
Ours5.17/17.2575.45/88.9261.73/74.5094.89/106.420.91/0.7862.95/85.2678.99/106.951.57/3.128.3/27.751.07/1.29
", + "image_path": "bd996503b0d5a7b37b23d2873470e13553fdf37ea61aac769822259a48d7d949.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 50, + 214, + 546, + 342 + ], + "blocks": [ + { + "bbox": [ + 47, + 181, + 546, + 205 + ], + "lines": [ + { + "bbox": [ + 47, + 181, + 546, + 205 + ], + "spans": [ + { + "bbox": [ + 47, + 181, + 546, + 205 + ], + "type": "text", + "content": "Table 3. HMR Comparison in the AscendMotion dataset (Horizontal Scene/Vertical Scene). The results for the horizontal and the vertical scene are separated by “/” in each cell. ClimbingCap performs significantly than others." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 214, + 546, + 342 + ], + "lines": [ + { + "bbox": [ + 50, + 214, + 546, + 342 + ], + "spans": [ + { + "bbox": [ + 50, + 214, + 546, + 342 + ], + "type": "table", + "html": "
ModalityMethodCamera CoordinateWorld Coordinate
ACCEL↓MPJPE↓PA-MPJPE↓PVE↓PCK0.3↑WA-MPJPE↓W-MPJPE↓RTE↓Jitter↓T-Error↓
RGBTRACE [60]57.65488.6398.83685.290.17365.11608.3828.9827.304.58
SLAHMR [75]6.10228.8282.39201.420.61391.24445.3813.8812.632.47
WHAM [56]9.11184.4394.40191.270.76321.58430.3532.959.263.78
GVHMR [55]6.98152.5691.13140.190.60349.17476.9021.5411.632.26
LiDARLiDARCapV2 [78]70.72389.97267.76364.260.50468.00520.0613.5654.253.12
LiveHPS [49]72.57190.86148.00225.650.59412.22524.4712.1150.773.52
LiDAR+RGBImmFusion [11]68.82322.21232.44435.720.45465.46586.3613.5914.553.96
FusionPose [14]58.54242.20189.92330.090.53373.46437.8813.5448.302.01
LEIR [73]16.46206.69117.52269.120.58228.1370.7410.638.421.84
Ours5.1884.0360.6999.890.86228.28261.068.178.721.26
", + "image_path": "fb417e75282438d86db760f2ec8b761a1f8ec5ce1fda126065d88a1eacd198a6.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 349, + 546, + 371 + ], + "lines": [ + { + "bbox": [ + 46, + 349, + 546, + 371 + ], + "spans": [ + { + "bbox": [ + 46, + 349, + 546, + 371 + ], + "type": "text", + "content": "Table 4. HMR Comparison in the CIMI4D dataset. ClimbingCap demonstrates remarkable generalization and robustness on other climbing datasets." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 47, + 392, + 259, + 404 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 392, + 259, + 404 + ], + "spans": [ + { + "bbox": [ + 47, + 392, + 259, + 404 + ], + "type": "text", + "content": "5.2. Comparison on Global Motion Recovery" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 45, + 411, + 287, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 411, + 287, + 542 + ], + "spans": [ + { + "bbox": [ + 45, + 411, + 287, + 542 + ], + "type": "text", + "content": "To evaluate the performance of global HMR, we follow the motion evaluation protocols of [55, 56] and the global trajectory evaluation protocol of [15, 73]. Following WHAM [56], to calculate the world coordinate MPJPE metrics, the predicted global sequence is divided into segments of 100 frames, and each segment is aligned with the ground truth segment either to the entire segment (WA-MPJPE) or to the first two frames (W-MPJPE). Besides WA-MPJPE and W-MPJPE, the root translation error (RTE), motion jitter (Jitter, in " + }, + { + "bbox": [ + 45, + 411, + 287, + 542 + ], + "type": "inline_equation", + "content": "m / s^3" + }, + { + "bbox": [ + 45, + 411, + 287, + 542 + ], + "type": "text", + "content": "), relative global translation error (T-Error in " + }, + { + "bbox": [ + 45, + 411, + 287, + 542 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 45, + 411, + 287, + 542 + ], + "type": "text", + "content": ") of the entire sequence are reported." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 543, + 287, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 543, + 287, + 615 + ], + "spans": [ + { + "bbox": [ + 46, + 543, + 287, + 615 + ], + "type": "text", + "content": "We evaluate ClimbingCap against nine methods with different modalities. The global RGB methods are TRACE [60], SLAHMR [75], WHAM [56], and GVHMR [55]. The LiDAR-based methods are LiDAR-CapV2 [78] and LiveHPS [49]. The LiDAR+RGB methods are ImmFusion [11], FusionPose [14], and LEIR [73]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "text", + "content": "Results in AscendMotion. The scenes in AscendMotion are categories into horizontal and vertical scenes, based on the major direction of human motions. Vertical motions are more challenging to human and to HMR methods than horizontal motions. The experimental results are depicted in Tab. 3. The top rows depicted the results for multiple state-of-the-art RGB-based HRM methods, and the bottom rows depicted the results for state-of-the-art LiDAR and Li-" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 392, + 547, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 392, + 547, + 608 + ], + "spans": [ + { + "bbox": [ + 304, + 392, + 547, + 608 + ], + "type": "text", + "content": "DAR+RGB based methods. In each cell of the table, the results for the horizontal and the vertical scenes are separated by " + }, + { + "bbox": [ + 304, + 392, + 547, + 608 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 304, + 392, + 547, + 608 + ], + "type": "text", + "content": ". For horizontal scenes, ClimbingCap performs the best in most of the metrics, especially in terms of MPJPE, WA-MPJPE and W-MPJPE. WHAM and GVHMR performs well for the camera coordinate metrics, but not in the world coordinate metrics. For the vertical scenes, ClimbingCap performs the best and significantly better than the second-best methods. GVHMR, a representative global HMR method, performs poorly in vertical scenes. It estimates the direction of movement by predicting the horizontal velocity. However, the major movement on the vertical scene of AscendMotion is an upward climbing movement. The LiDAR-based, and other LiDAR+RGB-based methods perform inferior to ClimbingCap. These methods do not fully consider global trajectories. The results demonstrate the importance of considering the relationship among camera coordinates and global coordinates." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 617, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 547, + 713 + ], + "type": "text", + "content": "Results in CIMI4D. To test the generalization ability of all the HMR methods, we evaluate their performance in the CIMI4D climbing dataset without retraining or fine-tuning. The CIMI4D dataset consists mostly horizontal scenes, and the motions are less challenging than AscendMotion. As it is reported in Tab. 4, ClimbingCap performs the best in the CIMI4D dataset for most the metrics, except the WA-MPJPE metric." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "12318" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 71, + 335, + 324 + ], + "blocks": [ + { + "bbox": [ + 49, + 71, + 335, + 324 + ], + "lines": [ + { + "bbox": [ + 49, + 71, + 335, + 324 + ], + "spans": [ + { + "bbox": [ + 49, + 71, + 335, + 324 + ], + "type": "image", + "image_path": "78f3baf5b8fcbef56deeffe2cd3146b27bff70865b84098acd4ccbc897346237.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 332, + 546, + 367 + ], + "lines": [ + { + "bbox": [ + 46, + 332, + 546, + 367 + ], + "spans": [ + { + "bbox": [ + 46, + 332, + 546, + 367 + ], + "type": "text", + "content": "Figure 5. Qualitative Evaluation in the AscendMotion and CIMI4D dataset. The left and right areas show the results of Camera Coordinate and World Coordinate respectively. The red circles indicate obvious errors. The last row shows the results for CIMI4D dataset. Our method ClimbingCap performs best qualitatively by comparison." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 337, + 71, + 545, + 324 + ], + "blocks": [ + { + "bbox": [ + 337, + 71, + 545, + 324 + ], + "lines": [ + { + "bbox": [ + 337, + 71, + 545, + 324 + ], + "spans": [ + { + "bbox": [ + 337, + 71, + 545, + 324 + ], + "type": "image", + "image_path": "e06950a81260058782cbeb2c7d1a301de62fb9c1dab1d56f44a4089fd50ab2e5.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 50, + 384, + 285, + 456 + ], + "blocks": [ + { + "bbox": [ + 50, + 384, + 285, + 456 + ], + "lines": [ + { + "bbox": [ + 50, + 384, + 285, + 456 + ], + "spans": [ + { + "bbox": [ + 50, + 384, + 285, + 456 + ], + "type": "table", + "html": "
VariantMPIJEPA-MPJPEPCK0.3WA-MPJPEW-MPJPERTE
(1) RGB Input105.6763.050.78117.17174.537.64
(2) w/o LwD80.4652.150.8970.0491.232.02
(3) w/o LSDS99.1360.660.81109.35164.117.03
(4) w/o VLRR91.1061.850.8388.59120.113.34
(5) w/o SS77.4352.570.9065.3082.091.83
RGB+LiDAR Input Full Model75.4550.510.9162.9578.991.57
", + "image_path": "d95eb10522f16a18ba3de677580ac37fe5b50bcc0cb3189a776b0eea69773164.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 464, + 286, + 475 + ], + "lines": [ + { + "bbox": [ + 47, + 464, + 286, + 475 + ], + "spans": [ + { + "bbox": [ + 47, + 464, + 286, + 475 + ], + "type": "text", + "content": "Table 5. Ablation Experiment for ClimbingCap in AscendMotion." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 47, + 489, + 169, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 489, + 169, + 502 + ], + "spans": [ + { + "bbox": [ + 47, + 489, + 169, + 502 + ], + "type": "text", + "content": "5.3. Ablation Experiment" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 510, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 287, + 715 + ], + "type": "text", + "content": "To understand the impact of each global loss and the semi-supervised module on ClimbingCap's performance, we evaluate five variants of ClimbingCap on the AscendMotion dataset following the same training and evaluation protocol as the previous section. (1) RGB Input: Using only RGB imagery without LiDAR, this variant performs worst overall due to missing global position data required for post-processing. (2) w/o " + }, + { + "bbox": [ + 46, + 510, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{LWD}" + }, + { + "bbox": [ + 46, + 510, + 287, + 715 + ], + "type": "text", + "content": ": Excluding the limb weight differentiation loss, which guides optimization using global point clouds, has a limited impact, indicating a beneficial but non-essential role. (3) w/o " + }, + { + "bbox": [ + 46, + 510, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{SDS}" + }, + { + "bbox": [ + 46, + 510, + 287, + 715 + ], + "type": "text", + "content": ": Removing the velocity direction smoothing loss, which ensures motion consistency, significantly reduces global metrics, highlighting its importance. (4) w/o " + }, + { + "bbox": [ + 46, + 510, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{VLR}" + }, + { + "bbox": [ + 46, + 510, + 287, + 715 + ], + "type": "text", + "content": ": The velocity direction smoothing loss constrains limb movements and corrects unexpected mis-predicted movements in a motion sequence; without this loss, ClimbingCap's global metrics drop sig" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 387, + 546, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 387, + 546, + 495 + ], + "spans": [ + { + "bbox": [ + 304, + 387, + 546, + 495 + ], + "type": "text", + "content": "nificantly, highlighting its importance for consistency and accuracy. (5) w/o " + }, + { + "bbox": [ + 304, + 387, + 546, + 495 + ], + "type": "inline_equation", + "content": "SS" + }, + { + "bbox": [ + 304, + 387, + 546, + 495 + ], + "type": "text", + "content": ": Removing the semi-supervised module results in a slight performance decrease, indicating its contribution to improving network training, though with a modest effect. The full ClimbingCap model, with RGB and LiDAR inputs, all loss terms, and the semi-supervised framework, achieves the best results in Tab. 5, validating the design's effectiveness for climbing motion analysis on AscendMotion." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 509, + 378, + 521 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 509, + 378, + 521 + ], + "spans": [ + { + "bbox": [ + 306, + 509, + 378, + 521 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 530, + 545, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 530, + 545, + 662 + ], + "spans": [ + { + "bbox": [ + 304, + 530, + 545, + 662 + ], + "type": "text", + "content": "In this work, we propose AscendMotion, a large-scale and multi-modal climbing motion dataset, and ClimbingCap, a human motion recovery framework for climbing motions. AscendMotion provides a rich, high-quality dataset that surpasses previous climbing datasets in both scale and complexity. ClimbingCap effectively recovers 3D climbing motions in the global coordinate system, ensuring accurate pose estimation and global localization. Experimental results demonstrate the quality of AscendMotion dataset, and show that ClimbingCap outperforms existing methods in terms of accuracy and robustness." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 665, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 714 + ], + "type": "text", + "content": "Acknowledgment. This work was partially supported by the Fundamental Research Funds for the Central Universities (No. 20720230033); by PDL (2022-PDL-12); by Xiaomi Young Talents Program." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12319" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "text", + "content": "[1] Thiemo Alldieck, Marc Kassubeck, Bastian Wandt, Bodo Rosenhahn, and Marcus Magnor. Optical flow-based 3d human motion estimation from monocular video. In German Conference on Pattern Recognition, pages 347-360. Springer, 2017. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 146, + 288, + 189 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 146, + 288, + 189 + ], + "spans": [ + { + "bbox": [ + 53, + 146, + 288, + 189 + ], + "type": "text", + "content": "[2] Marina Andric, Francesco Ricci, and Floriano Zini. Sensor-based activity recognition and performance assessment in climbing: A review. IEEE Access, 10:108583-108603, 2022. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 191, + 288, + 233 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 191, + 288, + 233 + ], + "spans": [ + { + "bbox": [ + 53, + 191, + 288, + 233 + ], + "type": "text", + "content": "[3] Raul Beltrán Beltrán, Julia Richter, Guido Köstermeyer, and Ulrich Heinkel. Climbing technique evaluation by means of skeleton video stream analysis. Sensors, 23(19):8216, 2023. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 236, + 287, + 302 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 236, + 287, + 302 + ], + "spans": [ + { + "bbox": [ + 53, + 236, + 287, + 302 + ], + "type": "text", + "content": "[4] Bharat Lal Bhatnagar, Cristian Sminchisescu, Christian Theobalt, and Gerard Pons-Moll. Combining implicit function learning and parametric models for 3d human reconstruction. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16, pages 311-329. Springer, 2020. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 302, + 287, + 367 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 302, + 287, + 367 + ], + "spans": [ + { + "bbox": [ + 53, + 302, + 287, + 367 + ], + "type": "text", + "content": "[5] Bharat Lal Bhatnagar, Xianghui Xie, Ilya A Petrov, Cristian Sminchisescu, Christian Theobalt, and Gerard Pons-Moll. Behave: Dataset and method for tracking human object interactions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15935-15946, 2022. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 369, + 288, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 369, + 288, + 423 + ], + "spans": [ + { + "bbox": [ + 53, + 369, + 288, + 423 + ], + "type": "text", + "content": "[6] Michael J Black, Priyanka Patel, Joachim Tesch, and Jinlong Yang. Bedlam: A synthetic dataset of bodies exhibiting detailed lifelike animated motion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8726-8737, 2023. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 424, + 287, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 424, + 287, + 468 + ], + "spans": [ + { + "bbox": [ + 53, + 424, + 287, + 468 + ], + "type": "text", + "content": "[7] Federica Bogo, Angjoo Kanazawa, Christoph Lassner, Peter V. Gehler, Javier Romero, and Michael J. Black. Keep it smpl: Automatic estimation of 3d human pose and shape from a single image. In ECCV, 2016. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 468, + 287, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 468, + 287, + 513 + ], + "spans": [ + { + "bbox": [ + 53, + 468, + 287, + 513 + ], + "type": "text", + "content": "[8] Christoph Bregler and Jitendra Malik. Tracking people with twists and exponential maps. Proceedings. 1998 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (Cat. No.98CB36231), pages 8-15, 1998. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 514, + 287, + 556 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 514, + 287, + 556 + ], + "spans": [ + { + "bbox": [ + 53, + 514, + 287, + 556 + ], + "type": "text", + "content": "[9] Zhe Cao, Tomas Simon, Shih-En Wei, and Yaser Sheikh. Realtime multi-person 2d pose estimation using part affinity fields. In Computer Vision and Pattern Recognition (CVPR), 2017. 1, 2, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 558, + 287, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 558, + 287, + 612 + ], + "spans": [ + { + "bbox": [ + 48, + 558, + 287, + 612 + ], + "type": "text", + "content": "[10] Kyungsik Cha, Eun-Young Lee, Myeong-Hyeon Heo, Kyu-Cheol Shin, Jonghee Son, and Dongho Kim. Analysis of climbing postures and movements in sport climbing for realistic 3d climbing animations. Procedia Engineering, 112: 52-57, 2015. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 613, + 287, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 287, + 679 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 287, + 679 + ], + "type": "text", + "content": "[11] Anjun Chen, Xiangyu Wang, Kun Shi, Shaohao Zhu, Bin Fang, Yingfeng Chen, Jiming Chen, Yuchi Huo, and Qi Ye. Immfusion: Robust mmwave-rgb fusion for 3d human body reconstruction in all weather conditions. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 2752-2758. IEEE, 2023. 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "text", + "content": "[12] Xin Chen, Anqi Pang, Wei Yang, Yuexin Ma, Lan Xu, and Jingyi Yu. Sportscap: Monocular 3d human motion capture and fine-grained understanding in challenging sports videos." + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 94 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 94 + ], + "type": "text", + "content": "International Journal of Computer Vision, 129:2846-2864, 2021. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 96, + 545, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 545, + 140 + ], + "type": "text", + "content": "[13] Hai Ci, Mingdong Wu, Wentao Zhu, Xiaoxuan Ma, Hao Dong, Fangwei Zhong, and Yizhou Wang. Gfpose: Learning 3d human pose prior with gradient fields. In CVPR, pages 4800-4810, 2023. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 141, + 545, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 141, + 545, + 196 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 545, + 196 + ], + "type": "text", + "content": "[14] Peishan Cong, Yiteng Xu, Yiming Ren, Juze Zhang, Lan Xu, Jingya Wang, Jingyi Yu, and Yuexin Ma. Weakly supervised 3d multi-person pose estimation for large-scale scenes based on monocular camera and single lidar. In AAAI, pages 461-469. AAAI Press, 2023. 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 198, + 545, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 198, + 545, + 263 + ], + "spans": [ + { + "bbox": [ + 308, + 198, + 545, + 263 + ], + "type": "text", + "content": "[15] Yudi Dai, Yitai Lin, XiPing Lin, Chenglu Wen, Lan Xu, Hongwei Yi, Siqi Shen, Yuexin Ma, and Cheng Wang. Sloper4d: A scene-aware dataset for global 4d human pose estimation in urban environments. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 682-692, 2023. 2, 3, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 264, + 545, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 264, + 545, + 330 + ], + "spans": [ + { + "bbox": [ + 308, + 264, + 545, + 330 + ], + "type": "text", + "content": "[16] Yudi Dai, Zhiyong Wang, Xiping Lin, Chenglu Wen, Lan Xu, Siqi Shen, Yuexin Ma, and Cheng Wang. Hisc4d: Human-centered interaction and 4d scene capture in large-scale space using wearable imus and lidar. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 2, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 331, + 545, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 331, + 545, + 420 + ], + "spans": [ + { + "bbox": [ + 308, + 331, + 545, + 420 + ], + "type": "text", + "content": "[17] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In 9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7, 2021. OpenReview.net, 2021. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 421, + 545, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 421, + 545, + 464 + ], + "spans": [ + { + "bbox": [ + 308, + 421, + 545, + 464 + ], + "type": "text", + "content": "[18] Zhiyang Dou, Qingxuan Wu, Cheng Lin, Zeyu Cao, Qiangqiang Wu, Weilin Wan, Taku Komura, and Wenping Wang. TORE: token reduction for efficient human mesh recovery with transformer. In ICCV, 2023. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 466, + 545, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 466, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 308, + 466, + 545, + 510 + ], + "type": "text", + "content": "[19] Petr Elias, Veronika Skvarlova, and Pavel Zezula. Speed21: Speed climbing motion dataset. In Proceedings of the 4th International Workshop on Multimedia Content Analysis in Sports, pages 43-50, 2021. 2, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 511, + 545, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 511, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 308, + 511, + 545, + 567 + ], + "type": "text", + "content": "[20] Shubham Goel, Georgios Pavlakos, Jathushan Rajasegaran, Angjoo Kanazawa, and Jitendra Malik. Humans in 4d: Reconstructing and tracking humans with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14783-14794, 2023. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 567, + 545, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 567, + 545, + 621 + ], + "spans": [ + { + "bbox": [ + 308, + 567, + 545, + 621 + ], + "type": "text", + "content": "[21] Marc Habermann, Weipeng Xu, Michael Zollhöfer, Gerard Pons-Moll, and Christian Theobalt. Livecap: Real-time human performance capture from monocular video. ACM Transactions on Graphics (TOG), 38(2):14:1-14:17, 2019. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 624, + 545, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 624, + 545, + 679 + ], + "spans": [ + { + "bbox": [ + 308, + 624, + 545, + 679 + ], + "type": "text", + "content": "[22] Marc Habermann, Weipeng Xu, Michael Zollhofer, Gerard Pons-Moll, and Christian Theobalt. Deepcap: Monocular human performance capture using weak supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 680, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 680, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 680, + 545, + 713 + ], + "type": "text", + "content": "[23] Yannan He, Anqi Pang, Xin Chen, Han Liang, Minye Wu, Yuexin Ma, and Lan Xu. Challencap: Monocular 3d capture of challenging human performances using multi-modal refer" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "12320" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 105 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 105 + ], + "type": "text", + "content": "ences. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11400-11411, 2021. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 288, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 288, + 173 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 288, + 173 + ], + "type": "text", + "content": "[24] Chun-Hao P Huang, Hongwei Yi, Markus Höschle, Matvey Safroshkin, Tsvetelina Alexiadis, Senya Polikovsky, Daniel Scharstein, and Michael J Black. Capturing and inferring dense full-body human-scene contact. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13274-13285, 2022. 2, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 175, + 287, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 175, + 287, + 229 + ], + "spans": [ + { + "bbox": [ + 48, + 175, + 287, + 229 + ], + "type": "text", + "content": "[25] Yinghao Huang, Manuel Kaufmann, Emre Aksan, Michael J. Black, Otmar Hilliges, and Gerard Pons-Moll. Deep inertial poser: Learning to reconstruct human pose from sparse inertial measurements in real time. ACM Transactions on Graphics, (Proc. SIGGRAPH Asia), 37(6):185:1-185:15, 2018. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 231, + 287, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 231, + 287, + 285 + ], + "spans": [ + { + "bbox": [ + 48, + 231, + 287, + 285 + ], + "type": "text", + "content": "[26] Hitomi Iguma, Akihiro Kawamura, and Ryo Kurazume. A new 3d motion and force measurement system for sport climbing. In 2020 IEEE/SICE International Symposium on System Integration, SII 2020, Honolulu, HI, USA, January 12-15, 2020, pages 1002-1007. IEEE, 2020. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 287, + 287, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 287, + 287, + 330 + ], + "spans": [ + { + "bbox": [ + 48, + 287, + 287, + 330 + ], + "type": "text", + "content": "[27] Angjoo Kanazawa, Michael J. Black, David W. Jacobs, and Jitendra Malik. End-to-end recovery of human shape and pose. In Computer Vision and Pattern Recognition (CVPR), 2018. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 332, + 287, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 332, + 287, + 397 + ], + "spans": [ + { + "bbox": [ + 48, + 332, + 287, + 397 + ], + "type": "text", + "content": "[28] Manuel Kaufmann, Jie Song, Chen Guo, Kaiyue Shen, Tianjian Jiang, Chengcheng Tang, Juan José Zárate, and Otmar Hilliges. Emdb: The electromagnetic database of global 3d human pose and shape in the wild. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14632-14643, 2023. 2, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 399, + 287, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 399, + 287, + 441 + ], + "spans": [ + { + "bbox": [ + 48, + 399, + 287, + 441 + ], + "type": "text", + "content": "[29] Diederik P Kingma and Jimmy Lei Ba. Adam: A method for stochastic gradient descent. In ICLR: international conference on learning representations, pages 1-15. ICLR US., 2015. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 445, + 287, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 445, + 287, + 498 + ], + "spans": [ + { + "bbox": [ + 48, + 445, + 287, + 498 + ], + "type": "text", + "content": "[30] Muhammed Kocabas, Nikos Athanasiou, and Michael J. Black. Vibe: Video inference for human body pose and shape estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 501, + 287, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 501, + 287, + 555 + ], + "spans": [ + { + "bbox": [ + 48, + 501, + 287, + 555 + ], + "type": "text", + "content": "[31] Muhammed Kocabas, Chun-Hao P. Huang, Otmar Hilliges, and Michael J. Black. Pare: Part attention regressor for 3d human body estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 11127-11137, 2021. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 557, + 287, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 557, + 287, + 612 + ], + "spans": [ + { + "bbox": [ + 48, + 557, + 287, + 612 + ], + "type": "text", + "content": "[32] Muhammed Kocabas, Ye Yuan, Pavlo Molchanov, Yunrong Guo, Michael J Black, Otmar Hilliges, Jan Kautz, and Umar Iqbal. Pace: Human and camera motion estimation from inthe-wild videos. In 2024 International Conference on 3D Vision (3DV), pages 397-408. IEEE, 2024. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 614, + 287, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 614, + 287, + 668 + ], + "spans": [ + { + "bbox": [ + 48, + 614, + 287, + 668 + ], + "type": "text", + "content": "[33] Nikos Kolotouros, Georgios Pavlakos, Michael J Black, and Kostas Daniilidis. Learning to reconstruct 3d human pose and shape via model-fitting in the loop. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2252-2261, 2019. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "type": "text", + "content": "[34] Felix Kosmalla, André Zenner, Marco Speicher, Florian Daiber, Nico Herbig, and Antonio Krüger. Exploring rock climbing in mixed reality environments. In Proceedings of the 2017 CHI Conference Extended Abstracts on Human" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "Factors in Computing Systems, page 1787-1793, New York, NY, USA, 2017. Association for Computing Machinery. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 97, + 547, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 97, + 547, + 162 + ], + "spans": [ + { + "bbox": [ + 307, + 97, + 547, + 162 + ], + "type": "text", + "content": "[35] Felix Kosmalla, André Zenner, Corinna Tasch, Florian Daiber, and Antonio Krüger. The importance of virtual hands and feet for virtual reality climbing. In Extended Abstracts of the 2020 CHI Conference on Human Factors in Computing Systems, CHI 2020, Honolulu, HI, USA, April 25-30, 2020, pages 1-8. ACM, 2020. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 163, + 545, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 163, + 545, + 229 + ], + "spans": [ + { + "bbox": [ + 307, + 163, + 545, + 229 + ], + "type": "text", + "content": "[36] Jialian Li, Jingyi Zhang, Zhiyong Wang, Siqi Shen, Chenglu Wen, Yuexin Ma, Lan Xu, Jingyi Yu, and Cheng Wang. Lidarcap: Long-range marker-less 3d human motion capture with lidar point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20502-20512, 2022. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 231, + 545, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 231, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 307, + 231, + 545, + 285 + ], + "type": "text", + "content": "[37] Jiefeng Li, Ye Yuan, Davis Rempe, Haotian Zhang, Pavlo Molchanov, Cewu Lu, Jan Kautz, and Umar Iqbal. Coin: Control-inpainting diffusion prior for human and camera motion estimation. In European Conference on Computer Vision, pages 426-446. Springer, 2025. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 286, + 545, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 286, + 545, + 329 + ], + "spans": [ + { + "bbox": [ + 307, + 286, + 545, + 329 + ], + "type": "text", + "content": "[38] Zhihao Li, Jianzhuang Liu, Zhensong Zhang, Songcen Xu, and Youliang Yan. Cliff: Carrying location information in full frames into human pose and shape estimation. In ECCV, 2022. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 332, + 545, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 332, + 545, + 387 + ], + "spans": [ + { + "bbox": [ + 307, + 332, + 545, + 387 + ], + "type": "text", + "content": "[39] Yitai Lin, Zhijie Wei, Wanfa Zhang, Xiping Lin, Yudi Dai, Chenglu Wen, Siqi Shen, Lan Xu, and Cheng Wang. Hmpear: A dataset for human pose estimation and action recognition. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 2069-2078, 2024. 2, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 388, + 545, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 388, + 545, + 430 + ], + "spans": [ + { + "bbox": [ + 307, + 388, + 545, + 430 + ], + "type": "text", + "content": "[40] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J. Black. Smpl: A skinned multiperson linear model. ACM Trans. Graph., 34(6):248:1-248:16, 2015. 1, 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 433, + 545, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 433, + 545, + 487 + ], + "spans": [ + { + "bbox": [ + 307, + 433, + 545, + 487 + ], + "type": "text", + "content": "[41] Naureen Mahmood, Nima Ghorbani, Nikolaus F. Troje, Gerard Pons-Moll, and Michael J. Black. Amass: Archive of motion capture as surface shapes. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2019. 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 489, + 545, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 489, + 545, + 520 + ], + "spans": [ + { + "bbox": [ + 307, + 489, + 545, + 520 + ], + "type": "text", + "content": "[42] Julieta Martinez, Rayat Hossain, Javier Romero, and James J Little. A simple yet effective baseline for 3d human pose estimation. In ICCV, 2017. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 523, + 545, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 523, + 545, + 577 + ], + "spans": [ + { + "bbox": [ + 307, + 523, + 545, + 577 + ], + "type": "text", + "content": "[43] Dominik Pandurevic, Paweł Draga, Alexander Sutor, and Klaus Hochradel. Analysis of competition and training videos of speed climbing athletes using feature and human body keypoint detection algorithms. Sensors, 22(6):2251, 2022. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 579, + 545, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 579, + 545, + 633 + ], + "spans": [ + { + "bbox": [ + 307, + 579, + 545, + 633 + ], + "type": "text", + "content": "[44] Chaitanya Patel, Zhouyingcheng Liao, and Gerard Pons-Moll. Tailornet: Predicting clothing in 3d as a function of human pose, shape and garment style. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7363-7373, 2020. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 635, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 635, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 307, + 635, + 545, + 689 + ], + "type": "text", + "content": "[45] Priyanka Patel, Chun-Hao P. Huang, Joachim Tesch, David T. Hoffmann, Shashank Tripathi, and Michael J. Black. AGORA: Avatars in geography optimized for regression analysis. In Proceedings IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), 2021. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 692, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 692, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 692, + 545, + 713 + ], + "type": "text", + "content": "[46] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed A. A. Osman, Dimitrios Tzionas, and" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "12321" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 67, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 287, + 116 + ], + "type": "text", + "content": "Michael J. Black. Expressive body capture: 3d hands, face, and body from a single image. In Proceedings IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 10975-10985, 2019. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 118, + 287, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 118, + 287, + 163 + ], + "spans": [ + { + "bbox": [ + 48, + 118, + 287, + 163 + ], + "type": "text", + "content": "[47] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems, 30, 2017. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 164, + 288, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 164, + 288, + 219 + ], + "spans": [ + { + "bbox": [ + 48, + 164, + 288, + 219 + ], + "type": "text", + "content": "[48] Davis Rempe, Tolga Birdal, Aaron Hertzmann, Jimei Yang, Srinath Sridhar, and Leonidas J. Guibas. Humor: 3d human motion model for robust pose estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 11488-11499, 2021. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 220, + 287, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 220, + 287, + 285 + ], + "spans": [ + { + "bbox": [ + 48, + 220, + 287, + 285 + ], + "type": "text", + "content": "[49] Yiming Ren, Xiao Han, Chengfeng Zhao, Jingya Wang, Lan Xu, Jingyi Yu, and Yuexin Ma. Livehps: Lidar-based scene-level human pose and shape estimation in free environment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1281-1291, 2024. 2, 3, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 286, + 287, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 286, + 287, + 319 + ], + "spans": [ + { + "bbox": [ + 48, + 286, + 287, + 319 + ], + "type": "text", + "content": "[50] Lionel Reveret, Sylvain Chapelle, Franck Quaine, and Pierre Legreneur. 3d visualization of body motion in speed climbing. Frontiers in Psychology, 11:2188, 2020. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 321, + 287, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 321, + 287, + 376 + ], + "spans": [ + { + "bbox": [ + 48, + 321, + 287, + 376 + ], + "type": "text", + "content": "[51] Julia Richter, Raul Beltrán Beltrán, Guido Köstermeyer, and Ulrich Heinkel. Human climbing and bouldering motion analysis: A survey on sensors, motion capture, analysis algorithms, recent advances and applications. VISIGRAPP (5: VISAPP), pages 751-758, 2020. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 378, + 287, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 378, + 287, + 443 + ], + "spans": [ + { + "bbox": [ + 48, + 378, + 287, + 443 + ], + "type": "text", + "content": "[52] Shunsuke Saito, Zeng Huang, Ryota Natsume, Shigeo Morishima, Hao Li, and Angjoo Kanazawa. Pifu: Pixel-aligned implicit function for high-resolution clothed human digitization. In 2019 IEEE/CVF International Conference on Computer Vision, ICCV 2019, Seoul, Korea (South), October 27 - November 2, 2019, pages 2304-2314. IEEE, 2019. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 445, + 287, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 445, + 287, + 520 + ], + "spans": [ + { + "bbox": [ + 48, + 445, + 287, + 520 + ], + "type": "text", + "content": "[53] Shunsuke Saito, Tomas Simon, Jason M. Saragih, and Hanbyul Joo. Pifuhd: Multi-level pixel-aligned implicit function for high-resolution 3d human digitization. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2020, Seattle, WA, USA, June 13-19, 2020, pages 81-90. Computer Vision Foundation / IEEE, 2020. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 522, + 287, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 522, + 287, + 588 + ], + "spans": [ + { + "bbox": [ + 48, + 522, + 287, + 588 + ], + "type": "text", + "content": "[54] Katsuhito Sasaki, Keisuke Shiro, and Jun Rekimoto. Exempliser: Predicting poses of experts as examples for beginners in climbing using a neural network. In Proceedings of the Augmented Humans International Conference, AHs 2020, Kaiserslautern, Germany, 16-17 March, 2020, pages 18:1-18:9. ACM, 2020. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 590, + 287, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 590, + 287, + 644 + ], + "spans": [ + { + "bbox": [ + 48, + 590, + 287, + 644 + ], + "type": "text", + "content": "[55] Zehong Shen, Huaijin Pi, Yan Xia, Zhi Cen, Sida Peng, Zechen Hu, Hujun Bao, Ruizhen Hu, and Xiaowei Zhou. World-grounded human motion recovery via gravity-view coordinates. In SIGGRAPH Asia Conference Proceedings, 2024. 2, 4, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 646, + 287, + 700 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 646, + 287, + 700 + ], + "spans": [ + { + "bbox": [ + 48, + 646, + 287, + 700 + ], + "type": "text", + "content": "[56] Soyong Shin, Juyong Kim, Eni Halilaj, and Michael J Black. Wham: Reconstructing world-grounded humans with accurate 3d motion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2070-2080, 2024. 2, 4, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 702, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 702, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 702, + 287, + 713 + ], + "type": "text", + "content": "[57] Zhuo Su, Lan Xu, Zerong Zheng, Tao Yu, Yebin Liu, and Lu" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "Fang. Robustfusion: Human volumetric capture with data-driven visual cues using a rgbd camera. In ECCV, 2020. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 308, + 96, + 545, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 545, + 129 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 545, + 129 + ], + "type": "text", + "content": "[58] Yu Sun, Qian Bao, Wu Liu, Yili Fu, Black Michael J., and Tao Mei. Monocular, One-stage, Regression of Multiple 3D People. In ICCV, 2021." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 130, + 545, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 130, + 545, + 163 + ], + "spans": [ + { + "bbox": [ + 307, + 130, + 545, + 163 + ], + "type": "text", + "content": "[59] Yu Sun, Wu Liu, Qian Bao, Yili Fu, Tao Mei, and Michael J Black. Putting People in their Place: Monocular Regression of 3D People in Depth. In CVPR, 2022. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 164, + 545, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 164, + 545, + 219 + ], + "spans": [ + { + "bbox": [ + 308, + 164, + 545, + 219 + ], + "type": "text", + "content": "[60] Yu Sun, Qian Bao, Wu Liu, Tao Mei, and Michael J Black. Trace: 5d temporal regression of avatars with dynamic cameras in 3d environments. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8856-8866, 2023. 2, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 220, + 545, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 220, + 545, + 297 + ], + "spans": [ + { + "bbox": [ + 308, + 220, + 545, + 297 + ], + "type": "text", + "content": "[61] Marcel Tiator, Christian Geiger, Bastian Dewitz, Ben Fischer, Laurin Gerhardt, David Nowottnik, and Hendrik Preu. Venga!: climbing in mixed reality. In Proceedings of the First Superhuman Sports Design Challenge: First International Symposium on Amplifying Capabilities and Competing in Mixed Realities, July 2-5, 2018, Delft, The Netherlands, pages 9:1-9:8. ACM, 2018. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 298, + 545, + 343 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 298, + 545, + 343 + ], + "spans": [ + { + "bbox": [ + 308, + 298, + 545, + 343 + ], + "type": "text", + "content": "[62] Daniel Vlasic, Rolf Adelsberger, Giovanni Vannucci, John C. Barnwell, Markus H. Gross, Wojciech Matusik, and Jovan Popovic. Practical motion capture in everyday surroundings. ACM SIGGRAPH 2007 papers, 2007. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 344, + 545, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 344, + 545, + 409 + ], + "spans": [ + { + "bbox": [ + 308, + 344, + 545, + 409 + ], + "type": "text", + "content": "[63] Ziniu Wan, Zhengjia Li, Maoqing Tian, Jianbo Liu, Shuai Yi, and Hongsheng Li. Encoder-decoder with multi-level attention for 3d human shape and pose estimation. In 2021 IEEE/CVF International Conference on Computer Vision, ICCV 2021, Montreal, QC, Canada, October 10-17, 2021, pages 13013-13022. IEEE, 2021. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 411, + 545, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 411, + 545, + 466 + ], + "spans": [ + { + "bbox": [ + 308, + 411, + 545, + 466 + ], + "type": "text", + "content": "[64] He Wang, Yezhen Cong, Or Litany, Yue Gao, and Leonidas J Guibas. 3dioumatch: Leveraging iou prediction for semi-supervised 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14615-14624, 2021. 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 468, + 545, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 468, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 308, + 468, + 545, + 533 + ], + "type": "text", + "content": "[65] Kuan-Chieh Wang, Zhenzhen Weng, Maria Xenochristou, João Pedro Araújo, Jeffrey Gu, Karen Liu, and Serena Yeung. Nemo: Learning 3d neural motion fields from multiple video instances of the same action. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22129-22138, 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 534, + 545, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 534, + 545, + 588 + ], + "spans": [ + { + "bbox": [ + 308, + 534, + 545, + 588 + ], + "type": "text", + "content": "[66] MeiJun Wang, Yu Meng, Zhongwei Qiu, Chao Zheng, Yan Xu, Jian Gao, et al. Pedestrian-centric 3d pre-collision pose and shape estimation from dashcam perspective. In The Thirty-eighth Annual Conference on Neural Information Processing Systems. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 590, + 545, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 590, + 545, + 655 + ], + "spans": [ + { + "bbox": [ + 308, + 590, + 545, + 655 + ], + "type": "text", + "content": "[67] Emily Whiting, Nada Ouf, Liane Makatura, Christos Mousas, Zhenyu Shu, and Ladislav Kavan. Environment-scale fabrication: Replicating outdoor climbing experiences. In Proceedings of the 2017 CHI Conference on Human Factors in Computing Systems, pages 1794-1804. ACM, 2017. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 657, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 657, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 308, + 657, + 545, + 689 + ], + "type": "text", + "content": "[68] HJ Wolring. New possibilities for human motion studies by real-time light spot position measurement. Biotelemetry, 1 (3):132, 1974. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "text", + "content": "[69] Qiming Xia, Wei Ye, Hai Wu, Shijia Zhao, Leyuan Xing, Xun Huang, Jinhao Deng, Xin Li, Chenglu Wen, and Cheng" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "12322" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 67, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 287, + 127 + ], + "type": "text", + "content": "Wang. Hinted: Hard instance enhanced detector with mixed-density feature fusion for sparsely-supervised 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15321-15330, 2024. 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 129, + 287, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 129, + 287, + 183 + ], + "spans": [ + { + "bbox": [ + 48, + 129, + 287, + 183 + ], + "type": "text", + "content": "[70] Yuliang Xiu, Jinlong Yang, Dimitrios Tzionas, and Michael J. Black. ICON: implicit clothed humans obtained from normals. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 13286-13296. IEEE, 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 184, + 287, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 184, + 287, + 237 + ], + "spans": [ + { + "bbox": [ + 48, + 184, + 287, + 237 + ], + "type": "text", + "content": "[71] Weipeng Xu, Avishek Chatterjee, Michael Zollhöfer, Helge Rhodin, Dushyant Mehta, Hans-Peter Seidel, and Christian Theobalt. Monoperfcap: Human performance capture from monocular video. ACM Transactions on Graphics (TOG), 37 (2):27:1-27:15, 2018. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 239, + 287, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 239, + 287, + 303 + ], + "spans": [ + { + "bbox": [ + 48, + 239, + 287, + 303 + ], + "type": "text", + "content": "[72] Ming Yan, Xin Wang, Yudi Dai, Siqi Shen, Chenglu Wen, Lan Xu, Yuexin Ma, and Cheng Wang. Cimi4d: A large multimodal climbing motion dataset under human-scene interactions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12977-12988, 2023. 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 304, + 287, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 304, + 287, + 370 + ], + "spans": [ + { + "bbox": [ + 48, + 304, + 287, + 370 + ], + "type": "text", + "content": "[73] Ming Yan, Yan Zhang, Shuqiang Cai, Shuqi Fan, Xincheng Lin, Yudi Dai, Siqi Shen, Chenglu Wen, Lan Xu, Yuexin Ma, et al. Reli11d: A comprehensive multimodal human motion dataset and method. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2250-2262, 2024. 2, 3, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 371, + 287, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 371, + 287, + 415 + ], + "spans": [ + { + "bbox": [ + 48, + 371, + 287, + 415 + ], + "type": "text", + "content": "[74] Ze Yang, Shenlong Wang, Siva Manivasagam, Zeng Huang, Wei-Chiu Ma, Xinchen Yan, Ersin Yumer, and Raquel Urtasun. S3: Neural shape, skeleton, and skinning fields for 3d human modeling. In CVPR, 2021. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 415, + 287, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 415, + 287, + 459 + ], + "spans": [ + { + "bbox": [ + 48, + 415, + 287, + 459 + ], + "type": "text", + "content": "[75] Vickie Ye, Georgios Pavlakos, Jitendra Malik, and Angjoo Kanazawa. Decoupling human and camera motion from videos in the wild. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2, 4, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 460, + 287, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 460, + 287, + 536 + ], + "spans": [ + { + "bbox": [ + 48, + 460, + 287, + 536 + ], + "type": "text", + "content": "[76] Xinyu Yi, Yuxiao Zhou, Marc Habermann, Soshi Shimada, Vladislav Golyanik, Christian Theobalt, and Feng Xu. Physical inertial poser (PIP): physics-aware real-time human motion tracking from sparse inertial sensors. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 13157-13168. IEEE, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 537, + 287, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 537, + 287, + 590 + ], + "spans": [ + { + "bbox": [ + 48, + 537, + 287, + 590 + ], + "type": "text", + "content": "[77] Ye Yuan, Umar Iqbal, Pavlo Molchanov, Kris Kitani, and Jan Kautz. Glamr: Global occlusion-aware human mesh recovery with dynamic cameras. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 592, + 287, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 592, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 592, + 287, + 635 + ], + "type": "text", + "content": "[78] Jingyi Zhang, Qihong Mao, Siqi Shen, Chenglu Wen, Lan Xu, and Cheng Wang. Lidarcapv2: 3d human pose estimation with human-object interaction from lidar point clouds. Pattern Recognition, 156:110848, 2024. 2, 3, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 636, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 636, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 636, + 287, + 689 + ], + "type": "text", + "content": "[79] Wu Zheng, Weiliang Tang, Li Jiang, and Chi-Wing Fu. Sessd: Self-ensembling single-stage object detector from point cloud. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14494–14503, 2021. 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "text", + "content": "[80] Xiaowei Zhou, Menglong Zhu, Spyridon Leonardos, Konstantinos G Derpanis, and Kostas Daniilidis. Sparseness" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 140 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 106 + ], + "type": "text", + "content": "Meets Deepness: 3D Human Pose Estimation from Monocular Video. In Computer Vision and Pattern Recognition (CVPR), 2016. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 107, + 545, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 140 + ], + "type": "text", + "content": "[81] Wentao Zhu, Xiaoxuan Ma, Zhaoyang Liu, Libin Liu, Wayne Wu, and Yizhou Wang. Motionbert: Unified pretraining for human motion analysis. In ICCV, 2023. 2" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "12323" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Closed-Loop Supervised Fine-Tuning of Tokenized Traffic Models/2dc3d2cc-243e-4ca0-9955-c4c727b1c79c_content_list.json b/2025/Closed-Loop Supervised Fine-Tuning of Tokenized Traffic Models/2dc3d2cc-243e-4ca0-9955-c4c727b1c79c_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..bd3f235ab7a0ecd2ea53918f37dcd2f5fcc34df9 --- /dev/null +++ b/2025/Closed-Loop Supervised Fine-Tuning of Tokenized Traffic Models/2dc3d2cc-243e-4ca0-9955-c4c727b1c79c_content_list.json @@ -0,0 +1,1536 @@ +[ + { + "type": "text", + "text": "Closed-Loop Supervised Fine-Tuning of Tokenized Traffic Models", + "text_level": 1, + "bbox": [ + 166, + 130, + 828, + 151 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zhejun Zhang $^{1}$ Peter Karkus $^{1}$ Maximilian Igl $^{1}$", + "bbox": [ + 295, + 180, + 699, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Wenhao Ding1 Yuxiao Chen1 Boris Ivanovic1 Marco Pavone1,2", + "bbox": [ + 222, + 198, + 774, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ NVIDIA Research $^{2}$ Stanford University", + "bbox": [ + 328, + 215, + 669, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{zhejunz, pkarkus, migl, wenhaod, yuxiaoc, bivanovic, mpavone}@nvidia.com", + "bbox": [ + 176, + 236, + 820, + 252 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "pavone@stanford.edu", + "bbox": [ + 410, + 255, + 581, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 304, + 326, + 319 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Traffic simulation aims to learn a policy for traffic agents that, when unrolled in closed-loop, faithfully recovers the joint distribution of trajectories observed in the real world. Inspired by large language models, tokenized multi-agent policies have recently become the state-of-the-art in traffic simulation. However, they are typically trained through open-loop behavior cloning, and thus suffer from covariate shift when executed in closed-loop during simulation. In this work, we present Closest Among Top-K (CAT-K) rollouts, a simple yet effective closed-loop fine-tuning strategy to mitigate covariate shift. CAT-K fine-tuning only requires existing trajectory data, without reinforcement learning or generative adversarial imitation. Concretely, CAT-K fine-tuning enables a small 7M-parameter tokenized traffic simulation policy to outperform a 102M-parameter model from the same model family, achieving the top spot on the Waymo Sim Agent Challenge leaderboard at the time of submission. The code is available at https://github.com/NVlabs/catk.", + "bbox": [ + 86, + 335, + 483, + 621 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 648, + 220, + 664 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Traffic modeling is a cornerstone of autonomous driving simulation and evaluation, typically formulated as learning a multi-agent policy that imitates the behavior of traffic participants in the real world. Given a set of historical agent trajectories and scene context (map, traffic light states, etc.), the policy generates actions for all simulated agents. The task gives rise to an imitation learning (IL) problem, with two key challenges: multimodality and covariate shift.", + "bbox": [ + 89, + 674, + 482, + 795 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Traffic agent behavior is highly multimodal, and faithfully recovering accurate behavior distributions is a key challenge in the field. Inspired by large language models [2, 32], recent works introduce next-token-prediction (NTP) models where the policy reduces to a classifier over a discrete set of trajectory tokens, which makes it easier to represent highly-multimodal distributions. Accordingly,", + "bbox": [ + 89, + 795, + 482, + 901 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5826def3f53e4acfaab5592110aca1388ebaf147c39a499454332abd02c6fcd6.jpg", + "image_caption": [ + "Figure 1. Closest Among Top-K (CAT-K) rollouts. The key idea of our approach is to unroll the policy during fine-tuning in a way that visited states remain close to the GT. At each time step, CAT-K first takes the top-K most likely action tokens according to the policy, then chooses the one leading to the state closest to the GT. As a result, CAT-K rollouts follow the mode of the GT (e.g., turning left), while random or top-K rollouts can lead to large deviations (e.g., going straight or right). Since the policy is essentially trained to minimize the distance between the rollout states and the GT states, the GT-based supervision remains effective for CAT-K rollouts, but not for random or top-K rollouts." + ], + "image_footnote": [], + "bbox": [ + 529, + 301, + 877, + 419 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "the Waymo Open Sim Agent Challenge (WOSAC) leaderboard [21] is heavily populated by tokenized traffic models [9, 35, 43, 47].", + "bbox": [ + 511, + 598, + 903, + 642 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Covariate shift is a well-known challenge of IL arising from the gap between open-loop training and closed-loop deployment. When a model is trained on a fixed dataset of expert demonstrations, it can face a distribution mismatch between the states seen during training and those encountered during deployment, as small errors compound and lead to unseen states where the policy performs poorly. A classic approach to tackle this problem is Dataset Aggregation (DAgger) [27], which unrolls the policy and queries an expert to generate new demonstrations, but querying experts is not readily available for traffic simulation. Prior work has proposed closed-loop training using hand-crafted recovery controllers [1] or reinforcement learning (RL) [19, 23, 37]. However, it is inherently difficult to design rewards with high behavioral realism or recovery controllers robust to divergent modes. Consequently, such approaches are not currently competitive on WOSAC realism metrics.", + "bbox": [ + 509, + 643, + 906, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "5422", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Contributions: We introduce Closest Among Top-K (CAT-K) rollouts, a simple yet highly efficient fine-tuning strategy to address the open-loop to closed-loop gap. The key idea, illustrated in Fig. 1, is to unroll the multimodal policy during training in a way that the policy-visited states remain close to the ground-truth (GT) demonstration. CAT-K achieves this by first finding the K most likely modes of the policy and then choosing the mode closest to the GT. At inference time, actions are sampled from the policy according to the predicted likelihoods. During training, however, random sampling would lead to large deviations from the demonstrations, making them invalid and degrading the final policy performance. Our CAT-K rollout strategy balances being on-policy and staying close to GT demonstrations, such that they remain a valid supervision signal.", + "bbox": [ + 89, + 90, + 480, + 316 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Experiments on the Waymo Open Motion Dataset (WOMD) demonstrate the efficacy of CAT-K. Notably, fine-tuning the SMART-7M [35] next-token-prediction traffic model enables it to outperform the 14x larger State-of-The-Art (SoTA) SMART-102M from the same model family, achieving the #1 spot on the public WOSAC leaderboard at the time of submission. To further demonstrate the potential of employing CAT-K fine-tuning for different tasks and policy representations, we apply it to an ego-motion planning task using a Gaussian Mixture Model (GMM) policy, yielding significant gains in closed-loop behavior, reducing collisions by $25.7\\%$ and off-road driving by $33.9\\%$ .", + "bbox": [ + 89, + 319, + 482, + 501 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 89, + 518, + 227, + 534 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Traffic simulation", + "text_level": 1, + "bbox": [ + 89, + 544, + 264, + 559 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Prior work explored various architectures for traffic models, including conditional variational autoencoders [13, 31, 36], motion forecasting Transformers [6, 22, 30, 46], and diffusion models [11, 15, 18, 26, 44, 45]. However, long-term stability is an open challenge for these models due to covariate shift when transitioning from open-loop training to closed-loop deployment. Various ways have been proposed to mitigate covariate shift. BITS [36] and Symphony [12] introduce hierarchy, with high-level intent and low-level behavior prediction. TrafficBots [40, 42] incorporates configurable behaviors through destination and personality. Most diffusion models use guiding to generate rule-abiding behavior; however, they often struggle with computational efficiency and long horizons. The latest advances come from NTP models, which predict the next action as a token, e.g., Trajeglish [24], GUMP [9], KiGRAS [43], MotionLM [28], and SMART [35]. Notably, SMART is the current SoTA on the WOSAC leaderboard. In addition to the strong scalability and flexibility, NTP models also show better closed-loop stability than regression-based models, thanks to their discrete action space. However, achieving generalization and reducing compounding errors continue to be challenges.", + "bbox": [ + 89, + 568, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Data augmentation for behavior cloning", + "text_level": 1, + "bbox": [ + 513, + 90, + 856, + 107 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Data augmentation is a simple yet effective way to improve the generalizability of traffic simulation models. ChauffeurNet [1] showed that carefully perturbing the vehicle trajectory and designing a recovery trajectory could alleviate the covariate shift suffered by behavior cloning (BC). However, this technique is difficult to apply to traffic simulation with complicated scenarios, including pedestrians and cyclists; and adding handcrafted recovery trajectories may negatively impact behavioral realism. Recently, NTP works such as Trajeglish [24] and SMART [35] have explored similar ideas by using noisy tokenization to perturb trajectories during training, but their data augmentation did not lead to significant improvements in performance. Our method is related to Trajeglish's noisy tokenization, but importantly, instead of blindly sampling tokens close to the GT without considering the policy, our CAT-K rollout selects the token from the most likely K tokens predicted by the policy that is closest to the GT. It can be considered as a closed-loop variants of the \"winner-takes-all\" training strategy [20] widely applied to open-loop motion prediction models [22, 29, 41].", + "bbox": [ + 511, + 112, + 903, + 415 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.3. Closed-loop fine-tuning", + "text_level": 1, + "bbox": [ + 513, + 424, + 728, + 440 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "As mentioned above, covariate shift is a major challenge faced by traffic simulation models, since most of them are trained in open-loop and evaluated in closed-loop. Even with data augmentation, the issue is not fully resolved as the augmented noisy data does not reflect the compounded error during closed-loop rollout. Therefore, some existing works explored the use of closed-loop fine-tuning.", + "bbox": [ + 511, + 446, + 903, + 551 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "As a classic remedy for the covariate shift, DAgger has been applied to end-to-end driving [25, 38, 39], yet its application in traffic simulation is limited, as it requires interactive demonstrations from human drivers or an expert policy. As an expert-free variant of DAgger, Data As Demonstrator (DAD) [34] obtains recovery actions using a pair of trajectories: a forward simulation (i.e., rollout) trajectory and a GT trajectory. At each time step, the policy is trained to return to the GT next state from the simulated current state. However, DAD fails when the rollout deviates from the GT trajectory. Our CAT-K rollout directly addresses this problem, enabling the use of DAD for closed-loop fine-tuning of traffic simulation and ego-vehicle policies.", + "bbox": [ + 511, + 553, + 905, + 748 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Another popular approach is to use RL. BC is combined with RL to improve the robustness of a policy in [19]. Yet, it also exposes RL's weakness in improving realism as it is difficult to handcraft a reward that promotes realism. Its follow-up work [23] learns a joint traffic model capable of rolling out the entire scenario by itself. The reward/loss are typically handcrafted, and can be distilled from explicit traffic rules [37]. With human preference data, RL from human feedback (RLHF) has also been applied to traffic model training for better user alignment [4, 10]. Finally, Sym-", + "bbox": [ + 511, + 750, + 905, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "5423", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "phony [12] adds a Generative Adversarial Imitation Learning (GAIL) loss to encourage the rollout states to stay in distribution; however, the well-known issue of training stability and discriminator overfitting remains a challenge.", + "bbox": [ + 89, + 90, + 483, + 154 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Background", + "text_level": 1, + "bbox": [ + 89, + 165, + 220, + 183 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Problem formulation", + "text_level": 1, + "bbox": [ + 89, + 190, + 289, + 205 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "A multi-agent traffic simulation policy can be typically formulated as $\\pi_{\\theta}(\\mathbf{a}_t|\\mathbf{h}_t,\\mathcal{M})$ where $\\theta$ denotes the trainable model parameters, $\\mathbf{h}_t = \\mathbf{s}_{t - H:t}$ is the state history of length $H$ $\\mathcal{M}$ is the context, including for example highdefinition (HD) maps and traffic light states, $t$ is the current time step, $\\mathbf{a}_t = [a_t^1,\\dots,a_t^N ]$ and $\\mathbf{s}_t = [s_t^1,\\dots,s_t^N ]$ are respectively the actions and states of $N$ agents at the current time step. The dimensions of actions $a$ and states $S$ are respectively denoted as $D_{a}$ and $D_{s}$ , i.e., $a\\in \\mathbb{R}^{D_a}$ and $s\\in \\mathbb{R}^{D_s}$ . From the current states and actions at step $t$ , the next states are computed using the per-agent forward dynamics $\\mathbf{s}_{t + 1} = f(\\mathbf{s}_t,\\mathbf{a}_t) = \\left[f(s_t^i,a_t^i)\\right]_{i = 1}^N$ . We assume that $f(s_{t}^{i},a_{t}^{i})$ is deterministic, and can be queried during training, which is the case for traffic simulation. Extensions to stochastic dynamics would be possible in future work. We define a rollout of $T$ steps starting at $t = 0$ as a sequence of states $\\mathbf{s}_{0:T} = [\\mathbf{s}_0,\\dots,\\mathbf{s}_T]$ , while the GT trajectories of all agents are denoted as $\\hat{\\mathbf{s}}_{0:T}$ . For the training, we are given a dataset $\\mathcal{D} = \\{\\hat{\\mathbf{s}}_{0:T}^j,\\mathcal{M}^j\\}_{j = 1}^{|D|}$ of such real-world trajectories that we want to emulate with their corresponding contexts.", + "bbox": [ + 89, + 212, + 483, + 521 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Next Token Prediction (NTP) policies", + "text_level": 1, + "bbox": [ + 89, + 531, + 413, + 547 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "NTP policies, such as SMART [35] and Trajeglish [24], are parameterized as a probability distribution over a vocabulary of action tokens denoted as $V = \\{x_{c} \\mid c = 1,2,\\ldots ,|V|\\}$ , where $|V|$ is the size of the vocabulary, $x_{c} \\in \\mathbb{R}^{D_{a}}$ are template actions and $c \\in \\mathbb{N}$ is the token index. Hence, an autoregressive NTP policy for traffic simulation can equivalently be written as an agent-factorized categorical distribution at each timestep $t$ , i.e.,", + "bbox": [ + 89, + 553, + 483, + 675 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\pi_ {\\theta} (\\mathbf {c} _ {t} \\mid \\mathbf {h} _ {t}, \\mathcal {M}) = \\prod_ {i = 1} ^ {N} \\pi_ {\\theta} (c _ {t} ^ {i} \\mid \\mathbf {h} _ {t}, \\mathcal {M}) = \\prod_ {i = 1} ^ {N} \\mathrm {C a t} (c _ {t} ^ {i}),\n$$\n", + "text_format": "latex", + "bbox": [ + 109, + 685, + 465, + 729 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathrm{Cat}(c_t^i)$ is the categorical distribution over the action token index for agent $i$ (and not to be confused with our method CAT-K). Given the sampled output $\\mathbf{c}_t = [c_t^1, \\dots, c_t^N]$ , the actions $\\mathbf{a}_t = [x_{c_t^1}, \\dots, x_{c_t^N}]$ are obtained using the token vocabulary $V$ .", + "bbox": [ + 89, + 739, + 483, + 816 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Method", + "text_level": 1, + "bbox": [ + 89, + 829, + 181, + 845 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Two key challenges in learning a policy from real-world trajectories are the multimodal nature of the trajectory distribution and the problem of covariate shift when policies are", + "bbox": [ + 89, + 854, + 483, + 902 + ], + "page_idx": 2 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1 CAT-K fine-tuning" + ], + "code_body": "1: Input: Policy $\\pi_{\\theta}$ , action token vocabulary $V$ , dataset $\\mathcal{D}$ \n2: Pre-train $\\pi_{\\theta}(\\mathbf{c}_t \\mid \\hat{\\mathbf{h}}_t, \\mathcal{M})$ with BC until convergence \n3: repeat ▷ Closed-loop supervised fine-tuning \n4: Sample a traffic scenario $\\{\\hat{s}_{0:T}, \\mathcal{M}\\}$ \n5: Init rollout state $\\mathbf{s}_0 = \\hat{\\mathbf{s}}_0$ ▷ CAT-K Rollout \n6: for $t$ in $[0, \\dots, T-1]$ do ▷ $T$ steps \n7: for $i$ in $[1, \\dots, N]$ do ▷ $N$ agents \n8: Get action index for rollout $c_t^i$ . (Eq. 1) \n9: Get next rollout state $s_{t+1}^i$ . (Eq. 3) \n10: Compute target $\\hat{c}_t^i$ . (Eq. 4) \n11: end for \n12: end for \n13: Update $\\theta$ by minimizing $\\mathcal{L}_{\\theta}(\\mathbf{s}_{0:T}, \\hat{\\mathbf{c}}_{1:T}, \\mathcal{M})$ . (Eq. 5) \n14: until convergence", + "bbox": [ + 516, + 108, + 908, + 324 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "trained open-loop, resulting in a distribution mismatch between expert states seen during training and states visited during policy deployment. Covariate shift can be overcome by closed-loop training, i.e., by training on trajectories sampled from the learned policy. However, this requires the generation of expert actions (or other notions of optimality) to be used as training targets along those trajectories [27]. Querying a human expert is infeasible at scale, RL-based methods require hard-to-define rewards, and methods such as GAIL [7] are prone to mode collapse. Consequently, they fail to achieve good performance in the WOSAC challenge.", + "bbox": [ + 511, + 351, + 906, + 518 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "An alternative strategy for generating \"expert\" actions is to construct recovery actions that bring the agent back to the available GT trajectory. However, this is complicated by the multimodal nature of the data, as the available GT trajectory might not be a valid recovery target for the generated trajectory. For example, as shown in Fig. 1, the GT trajectory $\\hat{s}_{0:T}$ turns left at the intersection, while the sampled trajectory $s_{0:T} \\sim \\pi_{\\theta}$ might go straight or turn right. As a result, while some SoTA traffic models [24, 35] augment the training data with recovery actions to reduce the covariate shift, they do so only from states that were reached by injecting small amounts of noise into the GT trajectory. This does guarantee that the GT trajectory remains a valid recovery target, but it completely ignores the learned policy and the state distribution induced by it. Instead, our method, Closest Among Top-K (CAT-K) rollout, informs the sampling process by the learned policy, but biases it towards the GT trajectory to guarantee the validity of the recovery actions. While simple to implement, it yields significant performance improvements in our experiments.", + "bbox": [ + 511, + 518, + 908, + 820 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1. Closest Among Top-K (CAT-K) rollout", + "text_level": 1, + "bbox": [ + 511, + 828, + 846, + 845 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To facilitate the formulation, we define a top $K$ operator:", + "bbox": [ + 511, + 849, + 885, + 866 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\{\\xi_ {1}, \\dots , \\xi_ {K} \\} = \\operatorname * {t o p} _ {c \\in \\{1, \\dots , | V | \\}} ^ {K} (\\operatorname {C a t} (c)),\n$$\n", + "text_format": "latex", + "bbox": [ + 584, + 875, + 834, + 904 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "5424", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/8f4cd23f6138102ff7538ee0289e9f1e9d3dff057449df6728ddcac3bc964411.jpg", + "image_caption": [ + "(a) CAT-K rollout" + ], + "image_footnote": [], + "bbox": [ + 91, + 87, + 310, + 239 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/6b2b69c710da47b2cbcd063c122ba307e476abdae9d37f9601fa10f9f61a9ea5.jpg", + "image_caption": [ + "(b) Top-K sampling", + "Figure 2. Schematic comparison of CAT-K rollout, top-K sampling, and data augmentation techniques of Trajeglish and SMART. In this example, the token vocabulary has a size of 5. We rollout three steps from $t = 0$ to $t = 3$ . For CAT-K rollout and top-K sampling, the top-K is w.r.t the probabilities $p$ of tokens predicted by the policy. For the data augmentations used by Trajeglish and SMART, the policy is unavailable, and the top-K selection is based on the negative distances between tokens and GT." + ], + "image_footnote": [], + "bbox": [ + 316, + 88, + 511, + 242 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/d2d24500cd37190dcf9257f87d47dcb6a2c6a2f95bd9a5f670f394bdd1b11fc5.jpg", + "image_caption": [ + "(c) Trajeglish noisy tokenization" + ], + "image_footnote": [], + "bbox": [ + 516, + 95, + 705, + 242 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/cca8454ae73ec2cf6cf84fd5c8ab361955d017b58f6c19b1d868395ec13f50ed.jpg", + "image_caption": [ + "(d) SMART trajectory perturbation" + ], + "image_footnote": [], + "bbox": [ + 712, + 95, + 903, + 242 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\operatorname{Cat}:\\mathbb{N}\\to \\mathbb{R}$ is the probability density of a categorical distribution on the vocabulary index, and $\\{\\xi_1,\\dots ,\\xi_K\\}$ are the $K$ most probable indices. The top operator can be considered as a variation of the arg max operator that returns multiple indices, with top equivalent to arg max.", + "bbox": [ + 89, + 334, + 482, + 409 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "At time step $t$ , the policy $\\pi_{\\theta}(\\mathbf{c}_t \\mid \\mathbf{h}_t, \\mathcal{M})$ outputs independent categorical distributions over the token vocabulary for each agent. Our method, CAT-K rollout, deterministically rolls out the policy by selecting, at each time step and for each agent, the one action among the top-K likeliest according to $\\pi_{\\theta}$ that brings the agent closest to the GT next state. Using a distance metric $d(\\cdot, \\cdot)$ on the states, this is formally expressed as follows:", + "bbox": [ + 89, + 409, + 483, + 530 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nc _ {t} ^ {i} = \\underset {c \\in \\left\\{\\xi_ {1}, \\dots , \\xi_ {K} \\right\\}} {\\arg \\min } d \\left(f \\left(s _ {t} ^ {i}, x _ {c}\\right), \\hat {s} _ {t + 1} ^ {i}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 189, + 540, + 480, + 569 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{\\xi_ {1}, \\dots , \\xi_ {K} \\right\\} = \\underset {c _ {t} ^ {i} \\in \\{1, \\dots , | V | \\}} {\\operatorname {t o p} ^ {K}} \\left[ \\pi \\left(c _ {t} ^ {i} \\mid \\mathbf {h} _ {t}, \\mathcal {M}\\right) \\right], \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 117, + 571, + 480, + 601 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $c_{t}^{i}$ is the action token indices of agent $i$ at step $t$ for the CAT-K rollout, and $\\{\\xi_1,\\dots ,\\xi_K\\}$ are the top-K likeliest token index according to the policy. Given $c_{t}^{i}$ , the next state is obtained using the vocabulary $V$ and the dynamics as:", + "bbox": [ + 89, + 612, + 483, + 674 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ns _ {t + 1} ^ {i} = f \\left(s _ {t} ^ {i}, x _ {c _ {t} ^ {i}}\\right). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 223, + 683, + 480, + 703 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "These rollout states will be used as the input $\\mathbf{h}$ to the policy at the next time step. By doing this sequentially from $t = 0$ to $t = T - 1$ and repeating for all $N$ agents, we obtain the CAT-K rollout trajectories $\\mathbf{s}_{0:T}$ .", + "bbox": [ + 89, + 712, + 483, + 773 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Closed-loop supervised fine-tuning", + "text_level": 1, + "bbox": [ + 89, + 782, + 393, + 799 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given the CAT-K rollout trajectory $\\mathbf{s}_{0:T}$ , we can apply the idea of DAD [34] and construct the recovery action indices $\\hat{\\mathbf{c}}_t$ from the GT trajectories $\\hat{\\mathbf{s}}_{0:T}$ by finding the action token that brings each agent closest to its original trajectory:", + "bbox": [ + 89, + 804, + 483, + 866 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {c} _ {t} ^ {i} = \\underset {c \\in \\{1, \\dots , | V | \\}} {\\arg \\min } d \\left(f \\left(s _ {t} ^ {i}, x _ {c}\\right), \\hat {s} _ {t + 1} ^ {i}\\right). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 163, + 875, + 480, + 904 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given these indices $\\hat{\\mathbf{c}}_t$ , the NTP policy is trained using the cross-entropy loss", + "bbox": [ + 511, + 334, + 905, + 364 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\theta} = - \\frac {1}{N T} \\sum_ {i = 1} ^ {N} \\sum_ {t = 0} ^ {T - 1} \\log \\pi_ {\\theta} \\left(\\hat {c} _ {t} ^ {i} \\mid \\mathbf {h} _ {t}, \\mathcal {M}\\right). \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 376, + 905, + 417 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Since CAT-K rollout is effective only when the top-K rollouts of the policy cover the GT mode, we adopt a two-stage training procedure, summarized in Algorithm 1. First, we obtain a reasonably well-trained policy through BC pretraining, then fine-tune it using CAT-K rollouts.", + "bbox": [ + 511, + 428, + 905, + 503 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.3. Comparison to previous methods", + "text_level": 1, + "bbox": [ + 511, + 515, + 803, + 531 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In Fig. 2, we compare our method with previous data augmentation approaches that alleviate the covariate shift for NTP traffic simulation policies.", + "bbox": [ + 511, + 537, + 905, + 582 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Top-K sampling is a common approach used by NTP models for generating sequences during inference as it improves the sample quality. However, it is unsuitable for generating trajectories during training as it does not consider the distance to the GT trajectory, and hence the validity of generated recovery actions (c.f. Fig. 2b). This can be partially addressed by post-hoc filtering of trajectories that deviated too far from the GT, but the resulting method is sample-inefficient, as most rollouts will be discarded, and the distance threshold hyperparameter is difficult to tune, since its optimal value varies across scenarios (e.g., high-speed vs. low-speed situations). In contrast, CAT-K rollout (c.f. Fig. 2a) is sample efficient by choosing the closest among top-K actions at each timestep and also removes the need to tune a distance-based hyperparameter. As shown in Tab. 2, our hyperparameter $K$ provides strong results across a large range of values and is hence much easier to tune.", + "bbox": [ + 511, + 583, + 906, + 839 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Instead of sampling top-K trajectories from the policy, current SoTA methods such as Trajeglish [24] and SMART [35] use forms of trajectory noising to address the issue of covariate shift. All trajectory noising approaches rely", + "bbox": [ + 511, + 839, + 905, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "5425", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "on the injection of small perturbations into the tokenization of the GT trajectory, but implementations can vary from each other in details. For example, in Trajeglish, the likelihood of each noised token is a function of the resulting distance to the ground-truth trajectory, i.e. $q_{t}^{i}(c) \\sim \\exp(-d(f(s_{t}^{i}, x_{c}), \\hat{s}_{t+1}^{i}) / \\tau)$ , while for SMART, tokens are sampled uniformly from those $K$ tokens with the highest likelihood $q_{t}^{i}(c)$ . Additionally, Trajeglish, similar to our methods, uses the DAD [34] recovery target that finds the token that would bring the agent back to the GT trajectory (c.f. Fig. 2c), while SMART uses the next, also noised, token as target - effectively treating the noise injection as data augmentation on the training trajectories (c.f. Fig. 2d).", + "bbox": [ + 89, + 90, + 480, + 287 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "However, Trajeglish's noisy tokenization does not yield a significant improvement (see Fig. 9 in [24]). Similarly, while SMART's trajectory perturbation enhances the performance of a zero-shot policy trained on NuPlan [3] and evaluated on WOSAC, it does not improve performance on WOSAC itself (see Tab. 4 in [35]). We believe this is because the state distributions generated by their sampling strategies are sub-optimal as they completely ignore what state distribution would be induced by the learned policy, hence likely oversampling irrelevant states and undersampling states the learned policy would actually encounter. By incorporating the learned policy into the sampling strategy, CAT-K generates a state distribution more like the policy's, and is hence better able to reduce the covariate shift between training and inference. Although sampling from the policy incurs higher data generation costs, these can be effectively mitigated through the two-phase training strategy with BC pre-training and closed-loop fine-tuning.", + "bbox": [ + 91, + 287, + 482, + 559 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Lastly, note that CAT-K with $K = |V|$ is equivalent to noise-free BC, since in this case rollouts follow the GT as closely as the available token book allows. On the other hand, for $K = 1$ , CAT-K is equivalent to deterministically rolling out the policy by always choosing the most likely token. Consequently, for CAT-K, the hyperparameter $K$ trades off following the policy (for $K = 1$ ) vs. following the GT (for $K = |V|$ ). In Fig. 5 we show how various choices of $K$ impact the Average Displacement Error (ADE) between the CAT-K rollouts and the GT. The intuition behind CAT-K is to find, among all the top-K T-step rollouts (with a total of $K^T$ possibilities), the one with the minimum ADE (minADE) to the GT. However, since directly solving this optimization problem is difficult, we use CAT-K as a greedy approximation to solve it sequentially.", + "bbox": [ + 89, + 560, + 482, + 785 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 89, + 799, + 220, + 816 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To show the broad relevance of CAT-K fine-tuning, we evaluate its performance on two different tasks using two different types of policy architecture, namely a traffic simulation task using a NTP policy and an ego-motion planning task using a policy parameterized as GMM. We base both tasks", + "bbox": [ + 89, + 824, + 480, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "on the widely used WOMD [5] and follow the simulation setup proposed in WOSAC [21], namely providing 1 second of history and generating rollouts of 8 seconds length.", + "bbox": [ + 511, + 90, + 903, + 137 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Traffic Simulation", + "text_level": 1, + "bbox": [ + 513, + 147, + 689, + 162 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Metrics. In the traffic simulation task, we follow the WOSAC protocol: for each scenario, we generate 32 simulated rollouts for all agents in the scene, at $10\\mathrm{Hz}$ , and evaluate how well their distribution matches that of the human demonstrations in the data. This distributional matching is evaluated along three dimensions, namely kinematics (e.g., velocities and accelerations), interactions (e.g., collisions), and map alignment (e.g., off-road). All three are summarized as weighted averages in the \"Realism Meta-Metric\" (RMM), the key performance indicator of the WOSAC leaderboard. For more details please see [21]. We also report the minADE, which is not used in the RMM, but widely applied for motion prediction and policy evaluation.", + "bbox": [ + 511, + 170, + 903, + 367 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Policy. We use SMART [35] as our policy architecture due to its strong performance and open-source implementation. Specifically, we use the SMART-tiny (7M) model and its open-sourced token vocabularies. The map polyline token vocabulary has a size of 1024, with each token representing a polyline consisting of 10 segments, each 0.5 meters long. The agent trajectory token vocabulary has a size of 2048, with each token representing a 0.5-second trajectory at $10\\mathrm{Hz}$ . SMART re-plans at $2\\mathrm{Hz}$ , hence the final output is at $10\\mathrm{Hz}$ when submitted to the WOSAC leaderboard. However, our experiments show that SMART's open-sourced agent token vocabulary results in lower kinematic metrics. To address this issue, we reran the K-disk clustering [24] with a larger number of trajectories. By using our own agent token vocabulary, the performance on the WOSAC leaderboard is significantly improved (c.f. Tab. 1). Please note that all other experiments in this paper (i.e., except those in Tab. 1) still use SMART's open-sourced token vocabularies. SMART uses Transformers [33] with query-centric representations [41, 46], and the distance metric $d(\\cdot, \\cdot)$ is the average Euclidean distance between the four pairs of corners of two bounding boxes. We disable SMART's trajectory perturbation as it deteriorates performance on the WOSAC leaderboard (see [35] Tab. 4). We run BC pretraining for 32 epochs, each taking 1.7 hour, then continue with closed-loop supervised fine-tuning with CAT-32 for 10 epochs, each taking 2.6 hours. During fine-tuning, the map encoder is frozen to save GPU memory, allowing for a larger training batch size with negligible impact on performance. For both pre-training and fine-tuning, we use $8 \\times$ A100 80GB GPUs with a total batch size of 80.", + "bbox": [ + 511, + 367, + 903, + 835 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.2. Ego-motion planning", + "text_level": 1, + "bbox": [ + 513, + 847, + 712, + 864 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Metrics. Unlike traffic simulation, in ego-motion planning only the ego-vehicle is controlled and evaluated. Moreover,", + "bbox": [ + 511, + 869, + 903, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5426", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/ab9e642723d8d2a53cbd6479d2c14307af227d441278716e9cbfd99a598a07b4.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Leaderboard, test split\nMethod# model\nparamsRMM\n↑RMM diff. to\nSMART-largeKinematic\nmetrics ↑Interactive\nmetrics ↑Map-based\nmetrics ↑min\nADE ↓
SMART-tiny fine-tuned w. CAT-K (ours)7 M0.7702+0.00880.49310.81190.87491.3068
SMART-large [35]102 M0.7614+0.00000.47860.80660.86481.3728
KiGRAS [43]0.7 M0.7597-0.00140.46910.80640.86581.4384
SMART-tiny [35]7 M0.7591-0.00230.47590.80390.86321.4062
FDriver-tiny7 M0.7584-0.00300.46140.80690.86581.4475
SMART [35]8 M0.7511-0.01030.44450.80500.85711.5447
BehaviorGPT [47]3 M0.7473-0.01410.43330.79970.85931.4147
GUMP [9]523 M0.7431-0.01830.47800.78870.83591.6041
", + "bbox": [ + 99, + 88, + 895, + 229 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Results on the WOSAC 2024 leaderboard [17]. RMM stands for Realism Meta Metric, the key metric used for ranking. Note that on the public leaderboard [17] our method appears under the name \"SMART-tiny-CLSFT\" (Closed-Loop Supervised Fine-Tuning).", + "bbox": [ + 89, + 232, + 903, + 262 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/558b1ca35205381da0e84a506a8d3b693a40841a3649ddb6830ab7615582e586.jpg", + "image_caption": [ + "Figure 3. Influence of $K_{\\mathrm{infer}}$ for inference-time top-K sampling." + ], + "image_footnote": [], + "bbox": [ + 96, + 272, + 473, + 378 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "the optimal policy is not the one that perfectly matches the GT trajectory distribution, but which minimizes planning metrics, such as collision rate, off-road rate, and ADE. For completeness, we also report RMM, applied only to ego trajectories, and minADE over 32 sampled rollouts. With the exception of minADE, and in contrast to the traffic simulation task, the rollouts for evaluation are sampled deterministically, by choosing the most-likely mode at each timestep.", + "bbox": [ + 88, + 414, + 482, + 537 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Policy. To show that our fine-tuning approach can be applied to a more general class of policies beyond NTP, we parameterized the output of the ego-policy as GMM. Specifically, we replace the final layer of the SMART network with two heads, a classification head that predicts the mixture densities and a regression head that predicts the means of the Gaussian distributions with fixed standard deviations (similar to, e.g., [14]). Specifically, the GMM predicts 16 modes, each representing a 3-dimensional Gaussian distribution over the changes in x- and y-position and yaw heading. The final model comprises a total of 3.2M parameters. In contrast to the SMART traffic simulation policy that uses cross-entropy loss on all agents, the GMM-based egopolicy is trained using negative log-likelihood loss and only predicts actions for the ego-vehicle. When applying CAT-K rollout or top-K sampling to the GMM, we select the K most-likely modes of the GMM, and within each mode deterministically use the mean of the Gaussian. Please note that ego-motion planning is a separate experiment from traffic simulation; hence, the ego-policy is not fine-tuned on the traffic simulation policy. Also note that the ego-policy operates in a mid-to-end manner, taking tracking results and HD maps as inputs, rather than in a fully end-to-end fashion that maps raw sensor observations to actions.", + "bbox": [ + 91, + 537, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.3. Results", + "text_level": 1, + "bbox": [ + 511, + 270, + 607, + 284 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.3.1. WOSAC leaderboard for traffic simulation", + "text_level": 1, + "bbox": [ + 511, + 292, + 857, + 306 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In Tab. 1 we compare our approach with other traffic simulation policies on the WOSAC leaderboard. Notably, all the top-ranking methods on the leaderboard are NTP policies trained via BC. Our SMART-tiny model using CAT-K fine-tuning outperforms the previous SoTA, SMART-large with 102M parameters, by a significant margin of $+0.0088$ ; and improves on SMART-tiny by $+0.0111$ . As the first to perform closed-loop fine-tuning on the leaderboard, our approach improves all metrics and sets a new SoTA.", + "bbox": [ + 511, + 311, + 903, + 446 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For fine-tuning, we chose CAT-32 after preliminary hyperparameter explorations, a choice that was later confirmed to perform well in our ablation studies (see Tab. 2), though CAT-K improved performance for a wide range of values of $K$ . We also found that choosing a sufficiently high $K_{\\mathrm{infer}}$ for top-K sampling during inference is important for a high RMM [9], and we chose $K_{\\mathrm{infer}} = 48$ for our leaderboard submission based on local validation results (see Fig. 3). In all of our experiments, we fixed the inference time sampling temperature to 1.0 and did not use top-p (nucleus) sampling [8], though we expect that tuning these hyperparameters can lead to further performance gains.", + "bbox": [ + 511, + 448, + 905, + 628 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To ensure that the observed performance gains with CAT-K are not due to improved hyperparameter tuning, we also conduct a large-scale hyperparameter grid search for the baseline method SMART-tiny. This includes training for more epochs, adjusting the learning rate and learning rate scheduler, experimenting with various data augmentation and data pre-processing strategies, and using various values of $K_{\\mathrm{infer}}$ for the inference-time top-K sampling. This allowed us to push the performance of the BC baseline to 0.7671 RMM on the test split (not shown in Tab. 1), but still falling significantly short of the performance of our CAT-K fine-tuning method (0.7702 RMM). Note that even comparing against this improved baseline, our method improves the performance more (+0.0031) than scaling up the model size by a factor of 14 to 102M parameters (+0.0023).", + "bbox": [ + 511, + 628, + 905, + 854 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Qualitatively, our method can generate diverse and realistic behavior over long periods of time. As shown in Fig. 6, our method can handle the subtle interactions between traf", + "bbox": [ + 511, + 854, + 903, + 898 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "5427", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/a2225f2b46f1e60bb4a79875848802bb2a087bb65384536fcdd262a80d741e3c.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Local val. split\nMethodCriterion\nof topKK for topKSampled fromNext targetRMM\n↑Kinematic metrics ↑Interactive metrics ↑Map-based metrics ↑min ADE ↓
BC pre-training---GT0.75810.45120.80760.86971.3152
BC fine-tuning---GT0.75900.45140.80960.87001.3039
Trajeglish's noisy\ntokenizationneg. dist.5†neg. dist.GT0.75620.44690.80740.86731.3459
neg. dist.5†uniformGT0.75540.44670.80690.86551.3404
neg. dist.32neg. dist.GT0.74010.41740.79850.84931.6669
SMART's trajectory\nperturbationneg. dist.5†uniformRO0.75560.44400.80820.86611.3177
neg. dist.5†neg. dist.RO0.75600.44690.80690.86731.3514
neg. dist.32uniformRO0.73140.41580.79490.83001.5380
Top-5prob5probGT0.64780.33130.68470.75281.8802
Top-5 + distance filterprob5probGT0.68600.33560.74660.80831.7627
Top-5 + distance based samplingprob5neg. dist.GT0.70580.35360.75790.84001.5848
Deterministic rollout--max-probGT0.63610.32910.68450.74921.8695
CAT-5prob-closestGT0.74230.42510.79170.86011.4677
CAT-16prob-closestGT0.76040.45920.80820.87091.3372
CAT-32 (leaderboard)prob-closestGT0.76160.45830.81050.87201.3105
CAT-40prob-closestGT0.76170.45670.81010.87381.2998
CAT-64prob-closestGT0.76020.45520.80980.87071.3028
", + "bbox": [ + 96, + 88, + 898, + 377 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Ablation study on WOSAC $2\\%$ validation split. We compare different ways to fine-tune the same base mode (BC pre-training). \"Sampled from\" indicates how the action is sampled during fine-tuning, either based on the distance to the GT (\"neg. dist\", \"uniform\", \"closest\") or based on the model outputs (\"prob\", \"max-prob\"). Here dist. is the abbreviation of distance. RO stands for rollout, i.e., the next target action is computed based on the rollout, not the GT state. $\\dagger$ indicates original hyperparameter choices of baseline algorithms.", + "bbox": [ + 89, + 378, + 906, + 434 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c7f34f8b1aaf5c7e0377b71e667b22c87ce2f03e92b5c9d311fa88c28deaa85f.jpg", + "image_caption": [ + "Figure 4. On server vs. local evaluation of SMART-tiny." + ], + "image_footnote": [], + "bbox": [ + 96, + 439, + 473, + 545 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "fic participants in a dense parking lot, which is arguably a more challenging scenario than intersections and highways for traffic simulation. Moreover, the behavior remains realistic at the end of the required simulation time of 8 seconds. Additional examples are in the supplementary videos.", + "bbox": [ + 89, + 571, + 482, + 650 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3.2. Ablation studies on WOSAC", + "text_level": 1, + "bbox": [ + 89, + 655, + 336, + 669 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In Tab. 2 we provide a thorough ablation study. Due to the high cost of evaluation we use $2\\%$ of the validation split (880 out of 44097 scenarios). To verify the fidelity of this evaluation setting, we compare results with those on the full validation set in Fig. 4. We observe consistent differences indicating that our evaluation setting is reasonable. We begin with a SMART-tiny model trained with BC for 32 epochs (BC pre-training, row 1). All other models fine-tune the BC pre-training model for 5 epochs. During inference, we use top-K sampling with $K_{\\mathrm{infer}} = 40$ for all methods (the best $K_{\\mathrm{infer}}$ for SMART-tiny according to Fig. 3).", + "bbox": [ + 89, + 672, + 482, + 839 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In Tab. 2, CAT-K fine-tuning is the only method that significantly outperforms the BC pre-training model. Further fine-tuning with BC (row 2), using Trajeglish's noisy tokenization (rows 3-5), or SMART's trajectory perturbation", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7fd159d160ec228322ae3be9c8fa4dd8a4f9f2866f069ef960c13031264f4f4b.jpg", + "image_caption": [ + "Figure 5. ADE between CAT-K rollouts and GT trajectories." + ], + "image_footnote": [], + "bbox": [ + 532, + 439, + 882, + 545 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "rows 6-8) remain on par with the original model or even reduce its performance. Closed-loop fine-tuning with top-K rollouts instead of CAT-K rollouts during training significantly reduces performance (rows 9-12), even when rollouts close to the GT are selected for training by either filtering them based on the distance to the GT or sampling among them from a distance-dependent distribution. The observation that data augmentation and fine-tuning with top-K sampling cannot improve the RMM is consistent with prior works [19, 23, 24, 35] and the WOSAC leaderboard. Ablations on different values of $K$ for top-K sampling and other hyperparameters are provided in the appendix.", + "bbox": [ + 511, + 580, + 906, + 762 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Next, we ablate the value of $K$ used for CAT-K rollout during fine-tuning (rows 13-17). Results indicate that the performance improvement is robust to the choice of $K$ after a reasonable minimum value. As discussed in Sec. 4.3, the hyperparameter $K$ in CAT-K determines how closely the policy follows the GT trajectory, in a way that is more robust than a distance based threshold, for which the optimal value varies strongly based on the situation (vehicle speed, proximity to other cars, etc.). To give more insight into how", + "bbox": [ + 511, + 763, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "5428", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/142ca603bb68ea37ecd251c565617b76663bddccdb672d82fc9c708d6966a61b.jpg", + "image_caption": [ + "Figure 6. Simulation results on WOSAC. Our fine-tuned policy generates interesting and diverse behaviors rarely seen in prior works. Each row represents a different rollout of our model in the same scene. The transparent boxes show the GT agents in the dataset, while the solid boxes show the agents generated by our model. We highlight the agents within the red, green, and orange rectangles, across time steps and rollouts. The red rectangle shows different interactive negotiations emerging between a pedestrian and two vehicles. The green rectangle shows an initially parked vehicle, that leaves (row 1) or stays parked (row 2, 3). The orange rectangle shows a vehicle waiting (row 1) in front of a speed bump (visualized as two light-blue lines), proceeding (row 2), or entering a parking space (row 3)." + ], + "image_footnote": [], + "bbox": [ + 91, + 88, + 251, + 282 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/03f8c0e01b5cecfa9e9deaff410f72e69e919f5baffe32152c5736896e22acca.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 254, + 88, + 413, + 282 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f523e846d9100a9e7d1bd3460223686115386bfc54f2d6ce803dbdd846d616bb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 416, + 88, + 575, + 282 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f1253b0be5349f0499b6eb5fbdb964683e7d2932e2886746342da5b8fc82b0d0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 578, + 88, + 736, + 282 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a21d42867d7ea82036d4a5422e8dac4d3690d87e9db4ae3e6938a448cb2b2548.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 740, + 88, + 898, + 282 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/faad84e4f890a2fae3ddaa70b8d0eebb60fafe7a60d5c55a09cd0da51e2dc003.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Method (Local val. split)Collision rate ↓Off-road rate ↓RMM ↑ADE ↓minADE32 ↓
BC pre-training0.05680.00530.81081.36231.3537
BC fine-tuning0.05990.00580.81051.35201.3509
Deterministic rollout0.04330.01380.80811.17990.7962
CAT-30.04220.00350.81691.30960.6912
", + "bbox": [ + 181, + 376, + 815, + 460 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3. Performance of ego policies on WOSAC with local evaluation on $2\\%$ validation split. All models are fine-tuned for 5 epochs based on the BC pre-training model, which is trained for 32 epochs. We use deterministic rollout during inference and compute all metrics, except for the $\\min\\mathrm{ADE}^{32}$ . For $\\min\\mathrm{ADE}^{32}$ , we generate 32 rollouts by using top-3 sampling with a temperature of 1.0 to first sample the categorical distribution over the mixtures, then selecting the mean of the sampled Gaussian mixture.", + "bbox": [ + 89, + 462, + 906, + 518 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "$K$ impacts rollouts, in Fig. 5, we inspect the average ADE between rollouts and GT trajectories over training epochs, for different $K$ values. As expected, as $K \\to |V|$ , the ADE decreases towards the level of quantization error, induced by tokenization with a finite vocabulary size (dashed line). With more fine-tuning epochs the average ADE slightly reduces, highlighting how CAT-K fine-tuning improves the policy to follow all behavior modes more closely.", + "bbox": [ + 89, + 539, + 482, + 661 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.3.3. Fine-tuning a GMM-based ego-policy", + "text_level": 1, + "bbox": [ + 89, + 669, + 398, + 686 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Besides NTP traffic simulation polices, CAT-K fine-tuning also improves the performance of a GMM-based ego-policy with continuous action space. In Tab. 3 we compare our approach with fine-tuning using deterministic rollouts, as well as with continued BC. Our CAT-3 fine-tuning improves all metrics except for ADE, where deterministic rollout performs better. This is expected, as deterministic rollout aligns all modes towards the GT, resulting in mode averaging. While this reduces ADE, it negatively impacts other metrics. Additionally, CAT-K fine-tuning helps overcome the limitation of GMM trained with negative log-likelihood loss, which struggles to capture multimodality as effectively as NTP policies trained with cross-entropy loss. This is supported by the significant drop in minADE after fine-tuning.", + "bbox": [ + 89, + 688, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 537, + 633, + 554 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we introduce CAT-K rollouts, a closed-loop supervised fine-tuning technique for IL problems with highly multimodal demonstrations, such as traffic simulation. The CAT-K rollout approximately finds the rollout closest to the GT among likely rollouts of a policy, ensuring adherence to the policy while maintaining GT as a reliable reference for supervised learning. As the first method using closed-loop fine-tuning, it achieves the top spot on the WOSAC leaderboard.", + "bbox": [ + 511, + 563, + 906, + 698 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In the future, we aim to incorporate modern sampling techniques into CAT-K rollouts, such as top-p sampling. Furthermore, instead of always selecting the closest token, we could loosen the \"winner-takes-all\" approach and sample from a group of nearby tokens. We also plan to explore CAT-K fine-tuning for a broader range of policy classes, such as variational autoencoders and diffusion models. Additionally, we want to apply it to other multimodal IL tasks, including end-to-end driving, motion generation for animation, and robot navigation and manipulation. Importantly, our results show that closed-loop supervised fine-tuning is a promising area of future research for policies trained in open-loop, such as the widely used NTP policies.", + "bbox": [ + 511, + 699, + 908, + 896 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "5429", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 90, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Mayank Bansal, Alex Krizhevsky, and Abhijit Ogale. Chauffeurnet: Learning to drive by imitating the best and synthesizing the worst. arXiv preprint arXiv:1812.03079, 2018. 1, 2", + "[2] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel Ziegler, Jeffrey Wu, Clemens Winter, Chris Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. Language models are few-shot learners. In Advances in Neural Information Processing Systems (NeurIPS), pages 1877-1901. Curran Associates, Inc., 2020. 1", + "[3] Holger Caesar, Juraj Kabzan, Kok Seang Tan, Whye Kit Fong, Eric M. Wolff, Alex H. Lang, Luke Fletcher, Oscar Beijbom, and Sammy Omari. NuPlan: A closed-loop ml-based planning benchmark for autonomous vehicles. In Conference on Computer Vision and Pattern Recognition (CVPR) ADP3 Workshop, 2021. 5", + "[4] Yulong Cao, Boris Ivanovic, Chaowei Xiao, and Marco Pavone. Reinforcement learning with human feedback for realistic traffic simulation. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 14428-14434. IEEE, 2024. 2", + "[5] Scott Ettinger, Shuyang Cheng, Benjamin Caine, Chenxi Liu, Hang Zhao, Sabeek Pradhan, Yuning Chai, Ben Sapp, Charles R Qi, Yin Zhou, et al. Large scale interactive motion forecasting for autonomous driving: The waymo open motion dataset. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 9710-9719, 2021. 5", + "[6] Roger Girgis, Florian Golemo, Felipe Codevilla, Martin Weiss, Jim Aldon D'Souza, Samira Ebrahimi Kahou, Felix Heide, and Christopher Pal. Latent variable sequential set transformers for joint multi-agent motion prediction. arXiv preprint arXiv:2104.00563, 2021. 2", + "[7] Jonathan Ho and Stefano Ermon. Generative adversarial imitation learning. Advances in Neural Information Processing Systems (NeurIPS), 29, 2016. 3", + "[8] Ari Holtzman, Jan Buys, Li Du, Maxwell Forbes, and Yejin Choi. The curious case of neural text degeneration. In International Conference on Learning Representations (ICLR), 2020. 6", + "[9] Yihan Hu, Siqi Chai, Zhening Yang, Jingyu Qian, Kun Li, Wenxin Shao, Haichao Zhang, Wei Xu, and Qiang Liu. Solving motion planning tasks with a scalable generative model. In European Conference on Computer Vision (ECCV), 2025. 1, 2, 6", + "[10] Zhiyu Huang, Xinshuo Weng, Maximilian Igl, Yuxiao Chen, Yulong Cao, Boris Ivanovic, Marco Pavone, and Chen Lv. Gen-drive: Enhancing diffusion generative driving policies with reward modeling and reinforcement learning finetuning. arXiv preprint arXiv:2410.05582, 2024. 2" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[11] Zhiyu Huang, Zixu Zhang, Ameya Vaidya, Yuxiao Chen, Chen Lv, and Jaime Fernandez Fisac. Versatile scene-consistent traffic scenario generation as optimization with diffusion. arXiv preprint arXiv:2404.02524, 2024. 2", + "[12] Maximilian Igl, Daewoo Kim, Alex Kuefler, Paul Mougin, Punit Shah, Kyriacos Shiarlis, Dragomir Anguelov, Mark Palatucci, Brandyn White, and Shimon Whiteson. Symphony: Learning realistic and diverse agents for autonomous driving simulation. In 2022 International Conference on Robotics and Automation (ICRA), pages 2445-2451. IEEE, 2022. 2, 3", + "[13] Maximilian Igl, Punit Shah, Paul Mougin, Sirish Srinivasan, Tarun Gupta, Brandyn White, Kyriacos Shiarlis, and Shimon Whiteson. Hierarchical imitation learning for stochastic environments. In 2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 1697-1704. IEEE, 2023. 2", + "[14] Boris Ivanovic and Marco Pavone. The trajectory: Probabilistic multi-agent trajectory modeling with dynamic spatiotemporal graphs. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 2375-2384, 2019. 6", + "[15] Chiyu Jiang, Andre Cornman, Cheolho Park, Benjamin Sapp, Yin Zhou, Dragomir Anguelov, et al. Motiondiffuser: Controllable multi-agent motion prediction using diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9644-9653, 2023. 2", + "[16] Longzhong Lin, Xuewu Lin, Kechun Xu, Haojian Lu, Lichao Huang, Rong Xiong, and Yue Wang. Revisit mixture models for multi-agent simulation: Experimental study within a unified framework, 2025. 2", + "[17] Waymo LLC. Waymo open sim agent challenge (wosac) 2024 leaderboard. https://waymo.com/open/challenges/2024/sim-agents/, 2024. Accessed: 2025-03-14. 6, 2", + "[18] Jack Lu, Kelvin Wong, Chris Zhang, Simon Suo, and Raquel Urtasun. Scenecontrol: Diffusion for controllable traffic scene generation. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 16908-16914. IEEE, 2024. 2", + "[19] Yiren Lu, Justin Fu, George Tucker, Xinlei Pan, Eli Bronstein, Rebecca Roelofs, Benjamin Sapp, Brandyn White, Aleksandra Faust, Shimon Whiteson, et al. Imitation is not enough: Robustifying imitation with reinforcement learning for challenging driving scenarios. In 2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 7553-7560. IEEE, 2023. 1, 2, 7, 3", + "[20] Osama Makansi, Eddy Ilg, Ozgun Cicek, and Thomas Brox. Overcoming limitations of mixture density networks: A sampling and fitting framework for multimodal future prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7144-7153, 2019. 2", + "[21] Nico Montali, John Lambert, Paul Mougin, Alex Kuefler, Nicholas Rhinehart, Michelle Li, Cole Gulino, Tristan Emrich, Zoey Yang, Shimon Whiteson, et al. The waymo open" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "5430", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "sim agents challenge. Advances in Neural Information Processing Systems (NeurIPS), 36, 2024. 1, 5", + "[22] Jiquan Ngiam, Benjamin Caine, Vijay Vasudevan, Zhengdong Zhang, Hao-Tien Lewis Chiang, Jeffrey Ling, Rebecca Roelofs, Alex Bewley, Chenxi Liu, Ashish Venugopal, et al. Scene transformer: A unified architecture for predicting multiple agent trajectories. arXiv preprint arXiv:2106.08417, 2021. 2", + "[23] Zhenghao Peng, Wenjie Luo, Yiren Lu, Tianyi Shen, Cole Gulino, Ari Seff, and Justin Fu. Improving agent behaviors with rl fine-tuning for autonomous driving. In European Conference on Computer Vision (ECCV), pages 165-181. Springer, 2025. 1, 2, 7", + "[24] Jonah Philion, Xue Bin Peng, and Sanja Fidler. Trajeglish: Traffic modeling as next-token prediction. In International Conference on Learning Representations (ICLR), 2024. 2, 3, 4, 5, 7", + "[25] Aditya Prakash, Aseem Behl, Eshed Ohn-Bar, Kashyap Chitta, and Andreas Geiger. Exploring data aggregation in policy learning for vision-based urban autonomous driving. In Conference on Computer Vision and Pattern Recognition (CVPR), pages 11763-11773, 2020. 2", + "[26] Ethan Pronovost, Meghan Reddy Ganesina, Noureldin Hendy, Zeyu Wang, Andres Morales, Kai Wang, and Nick Roy. Scenario diffusion: Controllable driving scenario generation with diffusion. Advances in Neural Information Processing Systems (NeurIPS), 36:68873-68894, 2023. 2", + "[27] Stéphane Ross, Geoffrey Gordon, and Drew Bagnell. A reduction of imitation learning and structured prediction to no-regret online learning. In International Conference on Artificial Intelligence and Statistics (AISTATS), pages 627-635. JMLR Workshop and Conference Proceedings, 2011. 1, 3", + "[28] Ari Seff, Brian Cera, Dian Chen, Mason Ng, Aurick Zhou, Nigamaa Nayakanti, Khaled S Refaat, Rami Al-Rfou, and Benjamin Sapp. Motionlm: Multi-agent motion forecasting as language modeling. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 8579-8590, 2023. 2", + "[29] Shaoshuai Shi, Li Jiang, Dengxin Dai, and Bernt Schiele. Motion transformer with global intention localization and local movement refinement. Advances in Neural Information Processing Systems (NeurIPS), 35:6531-6543, 2022. 2", + "[30] Shaoshuai Shi, Li Jiang, Dengxin Dai, and Bernt Schiele. Mtr++: Multi-agent motion prediction with symmetric scene modeling and guided intention querying. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 2024. 2", + "[31] Simon Suo, Sebastian Regalado, Sergio Casas, and Raquel Urtasun. TrafficSim: Learning to simulate realistic multi-agent behaviors. In Conference on Computer Vision and Pattern Recognition (CVPR), pages 10400-10409, 2021. 2", + "[32] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 1", + "[33] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia" + ], + "bbox": [ + 91, + 92, + 482, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems (NeurIPS), pages 5998-6008, 2017. 5", + "[34] Arun Venkatraman, Martial Hebert, and J Bagnell. Improving multi-step prediction of learned time series models. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), 2015. 2, 4, 5", + "[35] Wei Wu, Xiaoxin Feng, Ziyan Gao, and Yuheng Kan. Smart: Scalable multi-agent real-time simulation via next-token prediction. Advances in Neural Information Processing Systems (NeurIPS), 2025. 1, 2, 3, 4, 5, 6, 7", + "[36] Danfei Xu, Yuxiao Chen, Boris Ivanovic, and Marco Pavone. Bits: Bi-level imitation for traffic simulation. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 2929-2936. IEEE, 2023. 2", + "[37] Chris Zhang, James Tu, Lunjun Zhang, Kelvin Wong, Simon Suo, and Raquel Urtasun. Learning realistic traffic agents in closed-loop. In Conference on Robot Learning (CoRL), 2023. 1, 2, 3", + "[38] Jiakai Zhang and Kyunghyun Cho. Query-efficient imitation learning for end-to-end autonomous driving. arXiv preprint arXiv:1605.06450, 2016. 2", + "[39] Zhejun Zhang, Alexander Liniger, Dengxin Dai, Fisher Yu, and Luc Van Gool. End-to-end urban driving by imitating a reinforcement learning coach. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 2", + "[40] Zhejun Zhang, Alexander Liniger, Dengxin Dai, Fisher Yu, and Luc Van Gool. Trafficbots: Towards world models for autonomous driving simulation and motion prediction. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 1522-1529. IEEE, 2023. 2", + "[41] Zhejun Zhang, Alexander Liniger, Christos Sakaridis, Fisher Yu, and Luc V Gool. Real-time motion prediction via heterogeneous polyline transformer with relative pose encoding. Advances in Neural Information Processing Systems (NeurIPS), 2024. 2, 5", + "[42] Zhejun Zhang, Christos Sakaridis, and Luc Van Gool. Trafficbots v1. 5: Traffic simulation via conditional vaes and transformers with relative pose encoding. arXiv preprint arXiv:2406.10898, 2024. 2", + "[43] Jianbo Zhao, Jiaheng Zhuang, Qibin Zhou, Taiyu Ban, Ziyao Xu, Hangning Zhou, Junhe Wang, Guoan Wang, Zhiheng Li, and Bin Li. Kigras: Kinematic-driven generative model for realistic agent simulation. arXiv preprint arXiv:2407.12940, 2024. 1, 2, 6", + "[44] Ziyuan Zhong, Davis Rempe, Yuxiao Chen, Boris Ivanovic, Yulong Cao, Danfei Xu, Marco Pavone, and Baishakhi Ray. Language-guided traffic simulation via scene-level diffusion. arXiv preprint arXiv:2306.06344, 2023. 2", + "[45] Ziyuan Zhong, Davis Rempe, Danfei Xu, Yuxiao Chen, Sushant Veer, Tong Che, Baishakhi Ray, and Marco Pavone. Guided conditional diffusion for controllable traffic simulation. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 3560-3566. IEEE, 2023. 2", + "[46] Zikang Zhou, Jianping Wang, Yung-Hui Li, and Yu-Kai Huang. Query-centric trajectory prediction. In Proceedings" + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "5431", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17863-17873, 2023. 2, 5", + "[47] Zikang Zhou, Haibo Hu, Xinhong Chen, Jianping Wang, Nan Guan, Kui Wu, Yung-Hui Li, Yu-Kai Huang, and Chun Jason Xue. Behaviorgpt: Smart agent simulation for autonomous driving with next-patch prediction. arXiv preprint arXiv:2405.17372, 2024. 1, 6, 2" + ], + "bbox": [ + 91, + 90, + 482, + 191 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "5432", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Closed-Loop Supervised Fine-Tuning of Tokenized Traffic Models/2dc3d2cc-243e-4ca0-9955-c4c727b1c79c_model.json b/2025/Closed-Loop Supervised Fine-Tuning of Tokenized Traffic Models/2dc3d2cc-243e-4ca0-9955-c4c727b1c79c_model.json new file mode 100644 index 0000000000000000000000000000000000000000..c093630bf2ebb7841458e438a14638dc3e756dfd --- /dev/null +++ b/2025/Closed-Loop Supervised Fine-Tuning of Tokenized Traffic Models/2dc3d2cc-243e-4ca0-9955-c4c727b1c79c_model.json @@ -0,0 +1,2081 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.168, + 0.131, + 0.83, + 0.152 + ], + "angle": 0, + "content": "Closed-Loop Supervised Fine-Tuning of Tokenized Traffic Models" + }, + { + "type": "text", + "bbox": [ + 0.297, + 0.181, + 0.7, + 0.199 + ], + "angle": 0, + "content": "Zhejun Zhang\\(^{1}\\) Peter Karkus\\(^{1}\\) Maximilian Igl\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.199, + 0.776, + 0.216 + ], + "angle": 0, + "content": "Wenhao Ding1 Yuxiao Chen1 Boris Ivanovic1 Marco Pavone1,2" + }, + { + "type": "text", + "bbox": [ + 0.329, + 0.217, + 0.67, + 0.235 + ], + "angle": 0, + "content": "\\(^{1}\\)NVIDIA Research \\(^{2}\\)Stanford University" + }, + { + "type": "text", + "bbox": [ + 0.178, + 0.237, + 0.821, + 0.253 + ], + "angle": 0, + "content": "{zhejunz, pkarkus, migl, wenhaod, yuxiaoc, bivanovic, mpavone}@nvidia.com" + }, + { + "type": "text", + "bbox": [ + 0.411, + 0.256, + 0.582, + 0.269 + ], + "angle": 0, + "content": "pavone@stanford.edu" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.305, + 0.327, + 0.32 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.336, + 0.485, + 0.622 + ], + "angle": 0, + "content": "Traffic simulation aims to learn a policy for traffic agents that, when unrolled in closed-loop, faithfully recovers the joint distribution of trajectories observed in the real world. Inspired by large language models, tokenized multi-agent policies have recently become the state-of-the-art in traffic simulation. However, they are typically trained through open-loop behavior cloning, and thus suffer from covariate shift when executed in closed-loop during simulation. In this work, we present Closest Among Top-K (CAT-K) rollouts, a simple yet effective closed-loop fine-tuning strategy to mitigate covariate shift. CAT-K fine-tuning only requires existing trajectory data, without reinforcement learning or generative adversarial imitation. Concretely, CAT-K fine-tuning enables a small 7M-parameter tokenized traffic simulation policy to outperform a 102M-parameter model from the same model family, achieving the top spot on the Waymo Sim Agent Challenge leaderboard at the time of submission. The code is available at https://github.com/NVlabs/catk." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.649, + 0.222, + 0.665 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.675, + 0.483, + 0.796 + ], + "angle": 0, + "content": "Traffic modeling is a cornerstone of autonomous driving simulation and evaluation, typically formulated as learning a multi-agent policy that imitates the behavior of traffic participants in the real world. Given a set of historical agent trajectories and scene context (map, traffic light states, etc.), the policy generates actions for all simulated agents. The task gives rise to an imitation learning (IL) problem, with two key challenges: multimodality and covariate shift." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Traffic agent behavior is highly multimodal, and faithfully recovering accurate behavior distributions is a key challenge in the field. Inspired by large language models [2, 32], recent works introduce next-token-prediction (NTP) models where the policy reduces to a classifier over a discrete set of trajectory tokens, which makes it easier to represent highly-multimodal distributions. Accordingly," + }, + { + "type": "image", + "bbox": [ + 0.53, + 0.303, + 0.878, + 0.42 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.424, + 0.908, + 0.577 + ], + "angle": 0, + "content": "Figure 1. Closest Among Top-K (CAT-K) rollouts. The key idea of our approach is to unroll the policy during fine-tuning in a way that visited states remain close to the GT. At each time step, CAT-K first takes the top-K most likely action tokens according to the policy, then chooses the one leading to the state closest to the GT. As a result, CAT-K rollouts follow the mode of the GT (e.g., turning left), while random or top-K rollouts can lead to large deviations (e.g., going straight or right). Since the policy is essentially trained to minimize the distance between the rollout states and the GT states, the GT-based supervision remains effective for CAT-K rollouts, but not for random or top-K rollouts." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.599, + 0.905, + 0.643 + ], + "angle": 0, + "content": "the Waymo Open Sim Agent Challenge (WOSAC) leaderboard [21] is heavily populated by tokenized traffic models [9, 35, 43, 47]." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.645, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Covariate shift is a well-known challenge of IL arising from the gap between open-loop training and closed-loop deployment. When a model is trained on a fixed dataset of expert demonstrations, it can face a distribution mismatch between the states seen during training and those encountered during deployment, as small errors compound and lead to unseen states where the policy performs poorly. A classic approach to tackle this problem is Dataset Aggregation (DAgger) [27], which unrolls the policy and queries an expert to generate new demonstrations, but querying experts is not readily available for traffic simulation. Prior work has proposed closed-loop training using hand-crafted recovery controllers [1] or reinforcement learning (RL) [19, 23, 37]. However, it is inherently difficult to design rewards with high behavioral realism or recovery controllers robust to divergent modes. Consequently, such approaches are not currently competitive on WOSAC realism metrics." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5422" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.318 + ], + "angle": 0, + "content": "Contributions: We introduce Closest Among Top-K (CAT-K) rollouts, a simple yet highly efficient fine-tuning strategy to address the open-loop to closed-loop gap. The key idea, illustrated in Fig. 1, is to unroll the multimodal policy during training in a way that the policy-visited states remain close to the ground-truth (GT) demonstration. CAT-K achieves this by first finding the K most likely modes of the policy and then choosing the mode closest to the GT. At inference time, actions are sampled from the policy according to the predicted likelihoods. During training, however, random sampling would lead to large deviations from the demonstrations, making them invalid and degrading the final policy performance. Our CAT-K rollout strategy balances being on-policy and staying close to GT demonstrations, such that they remain a valid supervision signal." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.32, + 0.483, + 0.502 + ], + "angle": 0, + "content": "Experiments on the Waymo Open Motion Dataset (WOMD) demonstrate the efficacy of CAT-K. Notably, fine-tuning the SMART-7M [35] next-token-prediction traffic model enables it to outperform the 14x larger State-of-The-Art (SoTA) SMART-102M from the same model family, achieving the #1 spot on the public WOSAC leaderboard at the time of submission. To further demonstrate the potential of employing CAT-K fine-tuning for different tasks and policy representations, we apply it to an ego-motion planning task using a Gaussian Mixture Model (GMM) policy, yielding significant gains in closed-loop behavior, reducing collisions by \\(25.7\\%\\) and off-road driving by \\(33.9\\%\\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.519, + 0.228, + 0.535 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.545, + 0.265, + 0.56 + ], + "angle": 0, + "content": "2.1. Traffic simulation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.569, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Prior work explored various architectures for traffic models, including conditional variational autoencoders [13, 31, 36], motion forecasting Transformers [6, 22, 30, 46], and diffusion models [11, 15, 18, 26, 44, 45]. However, long-term stability is an open challenge for these models due to covariate shift when transitioning from open-loop training to closed-loop deployment. Various ways have been proposed to mitigate covariate shift. BITS [36] and Symphony [12] introduce hierarchy, with high-level intent and low-level behavior prediction. TrafficBots [40, 42] incorporates configurable behaviors through destination and personality. Most diffusion models use guiding to generate rule-abiding behavior; however, they often struggle with computational efficiency and long horizons. The latest advances come from NTP models, which predict the next action as a token, e.g., Trajeglish [24], GUMP [9], KiGRAS [43], MotionLM [28], and SMART [35]. Notably, SMART is the current SoTA on the WOSAC leaderboard. In addition to the strong scalability and flexibility, NTP models also show better closed-loop stability than regression-based models, thanks to their discrete action space. However, achieving generalization and reducing compounding errors continue to be challenges." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.091, + 0.857, + 0.108 + ], + "angle": 0, + "content": "2.2. Data augmentation for behavior cloning" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.113, + 0.905, + 0.416 + ], + "angle": 0, + "content": "Data augmentation is a simple yet effective way to improve the generalizability of traffic simulation models. ChauffeurNet [1] showed that carefully perturbing the vehicle trajectory and designing a recovery trajectory could alleviate the covariate shift suffered by behavior cloning (BC). However, this technique is difficult to apply to traffic simulation with complicated scenarios, including pedestrians and cyclists; and adding handcrafted recovery trajectories may negatively impact behavioral realism. Recently, NTP works such as Trajeglish [24] and SMART [35] have explored similar ideas by using noisy tokenization to perturb trajectories during training, but their data augmentation did not lead to significant improvements in performance. Our method is related to Trajeglish's noisy tokenization, but importantly, instead of blindly sampling tokens close to the GT without considering the policy, our CAT-K rollout selects the token from the most likely K tokens predicted by the policy that is closest to the GT. It can be considered as a closed-loop variants of the \"winner-takes-all\" training strategy [20] widely applied to open-loop motion prediction models [22, 29, 41]." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.425, + 0.73, + 0.441 + ], + "angle": 0, + "content": "2.3. Closed-loop fine-tuning" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.448, + 0.905, + 0.553 + ], + "angle": 0, + "content": "As mentioned above, covariate shift is a major challenge faced by traffic simulation models, since most of them are trained in open-loop and evaluated in closed-loop. Even with data augmentation, the issue is not fully resolved as the augmented noisy data does not reflect the compounded error during closed-loop rollout. Therefore, some existing works explored the use of closed-loop fine-tuning." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.554, + 0.906, + 0.749 + ], + "angle": 0, + "content": "As a classic remedy for the covariate shift, DAgger has been applied to end-to-end driving [25, 38, 39], yet its application in traffic simulation is limited, as it requires interactive demonstrations from human drivers or an expert policy. As an expert-free variant of DAgger, Data As Demonstrator (DAD) [34] obtains recovery actions using a pair of trajectories: a forward simulation (i.e., rollout) trajectory and a GT trajectory. At each time step, the policy is trained to return to the GT next state from the simulated current state. However, DAD fails when the rollout deviates from the GT trajectory. Our CAT-K rollout directly addresses this problem, enabling the use of DAD for closed-loop fine-tuning of traffic simulation and ego-vehicle policies." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.75, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Another popular approach is to use RL. BC is combined with RL to improve the robustness of a policy in [19]. Yet, it also exposes RL's weakness in improving realism as it is difficult to handcraft a reward that promotes realism. Its follow-up work [23] learns a joint traffic model capable of rolling out the entire scenario by itself. The reward/loss are typically handcrafted, and can be distilled from explicit traffic rules [37]. With human preference data, RL from human feedback (RLHF) has also been applied to traffic model training for better user alignment [4, 10]. Finally, Sym-" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5423" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.155 + ], + "angle": 0, + "content": "phony [12] adds a Generative Adversarial Imitation Learning (GAIL) loss to encourage the rollout states to stay in distribution; however, the well-known issue of training stability and discriminator overfitting remains a challenge." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.166, + 0.221, + 0.184 + ], + "angle": 0, + "content": "3. Background" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.191, + 0.29, + 0.206 + ], + "angle": 0, + "content": "3.1. Problem formulation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.213, + 0.485, + 0.522 + ], + "angle": 0, + "content": "A multi-agent traffic simulation policy can be typically formulated as \\(\\pi_{\\theta}(\\mathbf{a}_t|\\mathbf{h}_t,\\mathcal{M})\\) where \\(\\theta\\) denotes the trainable model parameters, \\(\\mathbf{h}_t = \\mathbf{s}_{t - H:t}\\) is the state history of length \\(H\\) \\(\\mathcal{M}\\) is the context, including for example highdefinition (HD) maps and traffic light states, \\(t\\) is the current time step, \\(\\mathbf{a}_t = [a_t^1,\\dots,a_t^N ]\\) and \\(\\mathbf{s}_t = [s_t^1,\\dots,s_t^N ]\\) are respectively the actions and states of \\(N\\) agents at the current time step. The dimensions of actions \\(a\\) and states \\(S\\) are respectively denoted as \\(D_{a}\\) and \\(D_{s}\\), i.e., \\(a\\in \\mathbb{R}^{D_a}\\) and \\(s\\in \\mathbb{R}^{D_s}\\). From the current states and actions at step \\(t\\), the next states are computed using the per-agent forward dynamics \\(\\mathbf{s}_{t + 1} = f(\\mathbf{s}_t,\\mathbf{a}_t) = \\left[f(s_t^i,a_t^i)\\right]_{i = 1}^N\\). We assume that \\(f(s_{t}^{i},a_{t}^{i})\\) is deterministic, and can be queried during training, which is the case for traffic simulation. Extensions to stochastic dynamics would be possible in future work. We define a rollout of \\(T\\) steps starting at \\(t = 0\\) as a sequence of states \\(\\mathbf{s}_{0:T} = [\\mathbf{s}_0,\\dots,\\mathbf{s}_T]\\), while the GT trajectories of all agents are denoted as \\(\\hat{\\mathbf{s}}_{0:T}\\). For the training, we are given a dataset \\(\\mathcal{D} = \\{\\hat{\\mathbf{s}}_{0:T}^j,\\mathcal{M}^j\\}_{j = 1}^{|D|}\\) of such real-world trajectories that we want to emulate with their corresponding contexts." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.532, + 0.414, + 0.549 + ], + "angle": 0, + "content": "3.2. Next Token Prediction (NTP) policies" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.554, + 0.484, + 0.676 + ], + "angle": 0, + "content": "NTP policies, such as SMART [35] and Trajeglish [24], are parameterized as a probability distribution over a vocabulary of action tokens denoted as \\( V = \\{x_{c} \\mid c = 1,2,\\ldots ,|V|\\} \\), where \\( |V| \\) is the size of the vocabulary, \\( x_{c} \\in \\mathbb{R}^{D_{a}} \\) are template actions and \\( c \\in \\mathbb{N} \\) is the token index. Hence, an autoregressive NTP policy for traffic simulation can equivalently be written as an agent-factorized categorical distribution at each timestep \\( t \\), i.e.," + }, + { + "type": "equation", + "bbox": [ + 0.111, + 0.686, + 0.466, + 0.73 + ], + "angle": 0, + "content": "\\[\n\\pi_ {\\theta} (\\mathbf {c} _ {t} \\mid \\mathbf {h} _ {t}, \\mathcal {M}) = \\prod_ {i = 1} ^ {N} \\pi_ {\\theta} (c _ {t} ^ {i} \\mid \\mathbf {h} _ {t}, \\mathcal {M}) = \\prod_ {i = 1} ^ {N} \\mathrm {C a t} (c _ {t} ^ {i}),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.74, + 0.485, + 0.817 + ], + "angle": 0, + "content": "where \\(\\mathrm{Cat}(c_t^i)\\) is the categorical distribution over the action token index for agent \\(i\\) (and not to be confused with our method CAT-K). Given the sampled output \\(\\mathbf{c}_t = [c_t^1, \\dots, c_t^N]\\), the actions \\(\\mathbf{a}_t = [x_{c_t^1}, \\dots, x_{c_t^N}]\\) are obtained using the token vocabulary \\(V\\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.83, + 0.183, + 0.847 + ], + "angle": 0, + "content": "4. Method" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Two key challenges in learning a policy from real-world trajectories are the multimodal nature of the trajectory distribution and the problem of covariate shift when policies are" + }, + { + "type": "code_caption", + "bbox": [ + 0.516, + 0.091, + 0.731, + 0.107 + ], + "angle": 0, + "content": "Algorithm 1 CAT-K fine-tuning" + }, + { + "type": "algorithm", + "bbox": [ + 0.517, + 0.109, + 0.909, + 0.325 + ], + "angle": 0, + "content": "1: Input: Policy \\(\\pi_{\\theta}\\), action token vocabulary \\(V\\), dataset \\(\\mathcal{D}\\) \n2: Pre-train \\(\\pi_{\\theta}(\\mathbf{c}_t \\mid \\hat{\\mathbf{h}}_t, \\mathcal{M})\\) with BC until convergence \n3: repeat ▷ Closed-loop supervised fine-tuning \n4: Sample a traffic scenario \\(\\{\\hat{s}_{0:T}, \\mathcal{M}\\}\\) \n5: Init rollout state \\(\\mathbf{s}_0 = \\hat{\\mathbf{s}}_0\\) ▷ CAT-K Rollout \n6: for \\(t\\) in \\([0, \\dots, T-1]\\) do ▷ \\(T\\) steps \n7: for \\(i\\) in \\([1, \\dots, N]\\) do ▷ \\(N\\) agents \n8: Get action index for rollout \\(c_t^i\\). (Eq. 1) \n9: Get next rollout state \\(s_{t+1}^i\\). (Eq. 3) \n10: Compute target \\(\\hat{c}_t^i\\). (Eq. 4) \n11: end for \n12: end for \n13: Update \\(\\theta\\) by minimizing \\(\\mathcal{L}_{\\theta}(\\mathbf{s}_{0:T}, \\hat{\\mathbf{c}}_{1:T}, \\mathcal{M})\\). (Eq. 5) \n14: until convergence" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.352, + 0.907, + 0.519 + ], + "angle": 0, + "content": "trained open-loop, resulting in a distribution mismatch between expert states seen during training and states visited during policy deployment. Covariate shift can be overcome by closed-loop training, i.e., by training on trajectories sampled from the learned policy. However, this requires the generation of expert actions (or other notions of optimality) to be used as training targets along those trajectories [27]. Querying a human expert is infeasible at scale, RL-based methods require hard-to-define rewards, and methods such as GAIL [7] are prone to mode collapse. Consequently, they fail to achieve good performance in the WOSAC challenge." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.519, + 0.909, + 0.821 + ], + "angle": 0, + "content": "An alternative strategy for generating \"expert\" actions is to construct recovery actions that bring the agent back to the available GT trajectory. However, this is complicated by the multimodal nature of the data, as the available GT trajectory might not be a valid recovery target for the generated trajectory. For example, as shown in Fig. 1, the GT trajectory \\(\\hat{s}_{0:T}\\) turns left at the intersection, while the sampled trajectory \\(s_{0:T} \\sim \\pi_{\\theta}\\) might go straight or turn right. As a result, while some SoTA traffic models [24, 35] augment the training data with recovery actions to reduce the covariate shift, they do so only from states that were reached by injecting small amounts of noise into the GT trajectory. This does guarantee that the GT trajectory remains a valid recovery target, but it completely ignores the learned policy and the state distribution induced by it. Instead, our method, Closest Among Top-K (CAT-K) rollout, informs the sampling process by the learned policy, but biases it towards the GT trajectory to guarantee the validity of the recovery actions. While simple to implement, it yields significant performance improvements in our experiments." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.829, + 0.848, + 0.846 + ], + "angle": 0, + "content": "4.1. Closest Among Top-K (CAT-K) rollout" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.85, + 0.887, + 0.867 + ], + "angle": 0, + "content": "To facilitate the formulation, we define a top \\(K\\) operator:" + }, + { + "type": "equation", + "bbox": [ + 0.585, + 0.875, + 0.836, + 0.905 + ], + "angle": 0, + "content": "\\[\n\\{\\xi_ {1}, \\dots , \\xi_ {K} \\} = \\operatorname * {t o p} _ {c \\in \\{1, \\dots , | V | \\}} ^ {K} (\\operatorname {C a t} (c)),\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5424" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.088, + 0.312, + 0.241 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.155, + 0.245, + 0.25, + 0.256 + ], + "angle": 0, + "content": "(a) CAT-K rollout" + }, + { + "type": "image", + "bbox": [ + 0.318, + 0.089, + 0.512, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.361, + 0.245, + 0.467, + 0.258 + ], + "angle": 0, + "content": "(b) Top-K sampling" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.097, + 0.707, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.525, + 0.245, + 0.697, + 0.257 + ], + "angle": 0, + "content": "(c) Trajeglish noisy tokenization" + }, + { + "type": "image", + "bbox": [ + 0.713, + 0.097, + 0.904, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.715, + 0.245, + 0.899, + 0.257 + ], + "angle": 0, + "content": "(d) SMART trajectory perturbation" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.261, + 0.907, + 0.318 + ], + "angle": 0, + "content": "Figure 2. Schematic comparison of CAT-K rollout, top-K sampling, and data augmentation techniques of Trajeglish and SMART. In this example, the token vocabulary has a size of 5. We rollout three steps from \\( t = 0 \\) to \\( t = 3 \\). For CAT-K rollout and top-K sampling, the top-K is w.r.t the probabilities \\( p \\) of tokens predicted by the policy. For the data augmentations used by Trajeglish and SMART, the policy is unavailable, and the top-K selection is based on the negative distances between tokens and GT." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.335, + 0.483, + 0.41 + ], + "angle": 0, + "content": "where \\(\\operatorname{Cat}:\\mathbb{N}\\to \\mathbb{R}\\) is the probability density of a categorical distribution on the vocabulary index, and \\(\\{\\xi_1,\\dots ,\\xi_K\\}\\) are the \\(K\\) most probable indices. The top operator can be considered as a variation of the arg max operator that returns multiple indices, with top equivalent to arg max." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.41, + 0.484, + 0.531 + ], + "angle": 0, + "content": "At time step \\( t \\), the policy \\( \\pi_{\\theta}(\\mathbf{c}_t \\mid \\mathbf{h}_t, \\mathcal{M}) \\) outputs independent categorical distributions over the token vocabulary for each agent. Our method, CAT-K rollout, deterministically rolls out the policy by selecting, at each time step and for each agent, the one action among the top-K likeliest according to \\( \\pi_{\\theta} \\) that brings the agent closest to the GT next state. Using a distance metric \\( d(\\cdot, \\cdot) \\) on the states, this is formally expressed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.19, + 0.541, + 0.482, + 0.57 + ], + "angle": 0, + "content": "\\[\nc _ {t} ^ {i} = \\underset {c \\in \\left\\{\\xi_ {1}, \\dots , \\xi_ {K} \\right\\}} {\\arg \\min } d \\left(f \\left(s _ {t} ^ {i}, x _ {c}\\right), \\hat {s} _ {t + 1} ^ {i}\\right), \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.119, + 0.572, + 0.482, + 0.602 + ], + "angle": 0, + "content": "\\[\n\\left\\{\\xi_ {1}, \\dots , \\xi_ {K} \\right\\} = \\underset {c _ {t} ^ {i} \\in \\{1, \\dots , | V | \\}} {\\operatorname {t o p} ^ {K}} \\left[ \\pi \\left(c _ {t} ^ {i} \\mid \\mathbf {h} _ {t}, \\mathcal {M}\\right) \\right], \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.613, + 0.484, + 0.675 + ], + "angle": 0, + "content": "where \\( c_{t}^{i} \\) is the action token indices of agent \\( i \\) at step \\( t \\) for the CAT-K rollout, and \\( \\{\\xi_1,\\dots ,\\xi_K\\} \\) are the top-K likeliest token index according to the policy. Given \\( c_{t}^{i} \\), the next state is obtained using the vocabulary \\( V \\) and the dynamics as:" + }, + { + "type": "equation", + "bbox": [ + 0.224, + 0.684, + 0.482, + 0.704 + ], + "angle": 0, + "content": "\\[\ns _ {t + 1} ^ {i} = f \\left(s _ {t} ^ {i}, x _ {c _ {t} ^ {i}}\\right). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.713, + 0.484, + 0.775 + ], + "angle": 0, + "content": "These rollout states will be used as the input \\(\\mathbf{h}\\) to the policy at the next time step. By doing this sequentially from \\(t = 0\\) to \\(t = T - 1\\) and repeating for all \\(N\\) agents, we obtain the CAT-K rollout trajectories \\(\\mathbf{s}_{0:T}\\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.783, + 0.394, + 0.8 + ], + "angle": 0, + "content": "4.2. Closed-loop supervised fine-tuning" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.805, + 0.484, + 0.867 + ], + "angle": 0, + "content": "Given the CAT-K rollout trajectory \\(\\mathbf{s}_{0:T}\\), we can apply the idea of DAD [34] and construct the recovery action indices \\(\\hat{\\mathbf{c}}_t\\) from the GT trajectories \\(\\hat{\\mathbf{s}}_{0:T}\\) by finding the action token that brings each agent closest to its original trajectory:" + }, + { + "type": "equation", + "bbox": [ + 0.164, + 0.876, + 0.482, + 0.905 + ], + "angle": 0, + "content": "\\[\n\\hat {c} _ {t} ^ {i} = \\underset {c \\in \\{1, \\dots , | V | \\}} {\\arg \\min } d \\left(f \\left(s _ {t} ^ {i}, x _ {c}\\right), \\hat {s} _ {t + 1} ^ {i}\\right). \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.335, + 0.906, + 0.365 + ], + "angle": 0, + "content": "Given these indices \\(\\hat{\\mathbf{c}}_t\\), the NTP policy is trained using the cross-entropy loss" + }, + { + "type": "equation", + "bbox": [ + 0.568, + 0.377, + 0.906, + 0.419 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\theta} = - \\frac {1}{N T} \\sum_ {i = 1} ^ {N} \\sum_ {t = 0} ^ {T - 1} \\log \\pi_ {\\theta} \\left(\\hat {c} _ {t} ^ {i} \\mid \\mathbf {h} _ {t}, \\mathcal {M}\\right). \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.429, + 0.906, + 0.505 + ], + "angle": 0, + "content": "Since CAT-K rollout is effective only when the top-K rollouts of the policy cover the GT mode, we adopt a two-stage training procedure, summarized in Algorithm 1. First, we obtain a reasonably well-trained policy through BC pretraining, then fine-tune it using CAT-K rollouts." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.516, + 0.805, + 0.532 + ], + "angle": 0, + "content": "4.3. Comparison to previous methods" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.538, + 0.906, + 0.583 + ], + "angle": 0, + "content": "In Fig. 2, we compare our method with previous data augmentation approaches that alleviate the covariate shift for NTP traffic simulation policies." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.584, + 0.907, + 0.84 + ], + "angle": 0, + "content": "Top-K sampling is a common approach used by NTP models for generating sequences during inference as it improves the sample quality. However, it is unsuitable for generating trajectories during training as it does not consider the distance to the GT trajectory, and hence the validity of generated recovery actions (c.f. Fig. 2b). This can be partially addressed by post-hoc filtering of trajectories that deviated too far from the GT, but the resulting method is sample-inefficient, as most rollouts will be discarded, and the distance threshold hyperparameter is difficult to tune, since its optimal value varies across scenarios (e.g., high-speed vs. low-speed situations). In contrast, CAT-K rollout (c.f. Fig. 2a) is sample efficient by choosing the closest among top-K actions at each timestep and also removes the need to tune a distance-based hyperparameter. As shown in Tab. 2, our hyperparameter \\( K \\) provides strong results across a large range of values and is hence much easier to tune." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.906, + 0.902 + ], + "angle": 0, + "content": "Instead of sampling top-K trajectories from the policy, current SoTA methods such as Trajeglish [24] and SMART [35] use forms of trajectory noising to address the issue of covariate shift. All trajectory noising approaches rely" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5425" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.288 + ], + "angle": 0, + "content": "on the injection of small perturbations into the tokenization of the GT trajectory, but implementations can vary from each other in details. For example, in Trajeglish, the likelihood of each noised token is a function of the resulting distance to the ground-truth trajectory, i.e. \\( q_{t}^{i}(c) \\sim \\exp(-d(f(s_{t}^{i}, x_{c}), \\hat{s}_{t+1}^{i}) / \\tau) \\), while for SMART, tokens are sampled uniformly from those \\( K \\) tokens with the highest likelihood \\( q_{t}^{i}(c) \\). Additionally, Trajeglish, similar to our methods, uses the DAD [34] recovery target that finds the token that would bring the agent back to the GT trajectory (c.f. Fig. 2c), while SMART uses the next, also noised, token as target - effectively treating the noise injection as data augmentation on the training trajectories (c.f. Fig. 2d)." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.289, + 0.483, + 0.56 + ], + "angle": 0, + "content": "However, Trajeglish's noisy tokenization does not yield a significant improvement (see Fig. 9 in [24]). Similarly, while SMART's trajectory perturbation enhances the performance of a zero-shot policy trained on NuPlan [3] and evaluated on WOSAC, it does not improve performance on WOSAC itself (see Tab. 4 in [35]). We believe this is because the state distributions generated by their sampling strategies are sub-optimal as they completely ignore what state distribution would be induced by the learned policy, hence likely oversampling irrelevant states and undersampling states the learned policy would actually encounter. By incorporating the learned policy into the sampling strategy, CAT-K generates a state distribution more like the policy's, and is hence better able to reduce the covariate shift between training and inference. Although sampling from the policy incurs higher data generation costs, these can be effectively mitigated through the two-phase training strategy with BC pre-training and closed-loop fine-tuning." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.561, + 0.483, + 0.786 + ], + "angle": 0, + "content": "Lastly, note that CAT-K with \\( K = |V| \\) is equivalent to noise-free BC, since in this case rollouts follow the GT as closely as the available token book allows. On the other hand, for \\( K = 1 \\), CAT-K is equivalent to deterministically rolling out the policy by always choosing the most likely token. Consequently, for CAT-K, the hyperparameter \\( K \\) trades off following the policy (for \\( K = 1 \\)) vs. following the GT (for \\( K = |V| \\)). In Fig. 5 we show how various choices of \\( K \\) impact the Average Displacement Error (ADE) between the CAT-K rollouts and the GT. The intuition behind CAT-K is to find, among all the top-K T-step rollouts (with a total of \\( K^T \\) possibilities), the one with the minimum ADE (minADE) to the GT. However, since directly solving this optimization problem is difficult, we use CAT-K as a greedy approximation to solve it sequentially." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.8, + 0.222, + 0.817 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.482, + 0.901 + ], + "angle": 0, + "content": "To show the broad relevance of CAT-K fine-tuning, we evaluate its performance on two different tasks using two different types of policy architecture, namely a traffic simulation task using a NTP policy and an ego-motion planning task using a policy parameterized as GMM. We base both tasks" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.138 + ], + "angle": 0, + "content": "on the widely used WOMD [5] and follow the simulation setup proposed in WOSAC [21], namely providing 1 second of history and generating rollouts of 8 seconds length." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.148, + 0.691, + 0.164 + ], + "angle": 0, + "content": "5.1. Traffic Simulation" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.171, + 0.905, + 0.368 + ], + "angle": 0, + "content": "Metrics. In the traffic simulation task, we follow the WOSAC protocol: for each scenario, we generate 32 simulated rollouts for all agents in the scene, at \\(10\\mathrm{Hz}\\), and evaluate how well their distribution matches that of the human demonstrations in the data. This distributional matching is evaluated along three dimensions, namely kinematics (e.g., velocities and accelerations), interactions (e.g., collisions), and map alignment (e.g., off-road). All three are summarized as weighted averages in the \"Realism Meta-Metric\" (RMM), the key performance indicator of the WOSAC leaderboard. For more details please see [21]. We also report the minADE, which is not used in the RMM, but widely applied for motion prediction and policy evaluation." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.368, + 0.905, + 0.837 + ], + "angle": 0, + "content": "Policy. We use SMART [35] as our policy architecture due to its strong performance and open-source implementation. Specifically, we use the SMART-tiny (7M) model and its open-sourced token vocabularies. The map polyline token vocabulary has a size of 1024, with each token representing a polyline consisting of 10 segments, each 0.5 meters long. The agent trajectory token vocabulary has a size of 2048, with each token representing a 0.5-second trajectory at \\(10\\mathrm{Hz}\\). SMART re-plans at \\(2\\mathrm{Hz}\\), hence the final output is at \\(10\\mathrm{Hz}\\) when submitted to the WOSAC leaderboard. However, our experiments show that SMART's open-sourced agent token vocabulary results in lower kinematic metrics. To address this issue, we reran the K-disk clustering [24] with a larger number of trajectories. By using our own agent token vocabulary, the performance on the WOSAC leaderboard is significantly improved (c.f. Tab. 1). Please note that all other experiments in this paper (i.e., except those in Tab. 1) still use SMART's open-sourced token vocabularies. SMART uses Transformers [33] with query-centric representations [41, 46], and the distance metric \\(d(\\cdot, \\cdot)\\) is the average Euclidean distance between the four pairs of corners of two bounding boxes. We disable SMART's trajectory perturbation as it deteriorates performance on the WOSAC leaderboard (see [35] Tab. 4). We run BC pretraining for 32 epochs, each taking 1.7 hour, then continue with closed-loop supervised fine-tuning with CAT-32 for 10 epochs, each taking 2.6 hours. During fine-tuning, the map encoder is frozen to save GPU memory, allowing for a larger training batch size with negligible impact on performance. For both pre-training and fine-tuning, we use \\(8 \\times\\) A100 80GB GPUs with a total batch size of 80." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.848, + 0.714, + 0.865 + ], + "angle": 0, + "content": "5.2. Ego-motion planning" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Metrics. Unlike traffic simulation, in ego-motion planning only the ego-vehicle is controlled and evaluated. Moreover," + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "5426" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.1, + 0.089, + 0.897, + 0.23 + ], + "angle": 0, + "content": "
Leaderboard, test split\nMethod# model\nparamsRMM\n↑RMM diff. to\nSMART-largeKinematic\nmetrics ↑Interactive\nmetrics ↑Map-based\nmetrics ↑min\nADE ↓
SMART-tiny fine-tuned w. CAT-K (ours)7 M0.7702+0.00880.49310.81190.87491.3068
SMART-large [35]102 M0.7614+0.00000.47860.80660.86481.3728
KiGRAS [43]0.7 M0.7597-0.00140.46910.80640.86581.4384
SMART-tiny [35]7 M0.7591-0.00230.47590.80390.86321.4062
FDriver-tiny7 M0.7584-0.00300.46140.80690.86581.4475
SMART [35]8 M0.7511-0.01030.44450.80500.85711.5447
BehaviorGPT [47]3 M0.7473-0.01410.43330.79970.85931.4147
GUMP [9]523 M0.7431-0.01830.47800.78870.83591.6041
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.233, + 0.905, + 0.263 + ], + "angle": 0, + "content": "Table 1. Results on the WOSAC 2024 leaderboard [17]. RMM stands for Realism Meta Metric, the key metric used for ranking. Note that on the public leaderboard [17] our method appears under the name \"SMART-tiny-CLSFT\" (Closed-Loop Supervised Fine-Tuning)." + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.273, + 0.475, + 0.379 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.098, + 0.381, + 0.475, + 0.396 + ], + "angle": 0, + "content": "Figure 3. Influence of \\( K_{\\mathrm{infer}} \\) for inference-time top-K sampling." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.415, + 0.483, + 0.538 + ], + "angle": 0, + "content": "the optimal policy is not the one that perfectly matches the GT trajectory distribution, but which minimizes planning metrics, such as collision rate, off-road rate, and ADE. For completeness, we also report RMM, applied only to ego trajectories, and minADE over 32 sampled rollouts. With the exception of minADE, and in contrast to the traffic simulation task, the rollouts for evaluation are sampled deterministically, by choosing the most-likely mode at each timestep." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.539, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Policy. To show that our fine-tuning approach can be applied to a more general class of policies beyond NTP, we parameterized the output of the ego-policy as GMM. Specifically, we replace the final layer of the SMART network with two heads, a classification head that predicts the mixture densities and a regression head that predicts the means of the Gaussian distributions with fixed standard deviations (similar to, e.g., [14]). Specifically, the GMM predicts 16 modes, each representing a 3-dimensional Gaussian distribution over the changes in x- and y-position and yaw heading. The final model comprises a total of 3.2M parameters. In contrast to the SMART traffic simulation policy that uses cross-entropy loss on all agents, the GMM-based egopolicy is trained using negative log-likelihood loss and only predicts actions for the ego-vehicle. When applying CAT-K rollout or top-K sampling to the GMM, we select the K most-likely modes of the GMM, and within each mode deterministically use the mean of the Gaussian. Please note that ego-motion planning is a separate experiment from traffic simulation; hence, the ego-policy is not fine-tuned on the traffic simulation policy. Also note that the ego-policy operates in a mid-to-end manner, taking tracking results and HD maps as inputs, rather than in a fully end-to-end fashion that maps raw sensor observations to actions." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.271, + 0.608, + 0.285 + ], + "angle": 0, + "content": "5.3. Results" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.293, + 0.859, + 0.307 + ], + "angle": 0, + "content": "5.3.1. WOSAC leaderboard for traffic simulation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.312, + 0.905, + 0.447 + ], + "angle": 0, + "content": "In Tab. 1 we compare our approach with other traffic simulation policies on the WOSAC leaderboard. Notably, all the top-ranking methods on the leaderboard are NTP policies trained via BC. Our SMART-tiny model using CAT-K fine-tuning outperforms the previous SoTA, SMART-large with 102M parameters, by a significant margin of \\(+0.0088\\); and improves on SMART-tiny by \\(+0.0111\\). As the first to perform closed-loop fine-tuning on the leaderboard, our approach improves all metrics and sets a new SoTA." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.449, + 0.906, + 0.629 + ], + "angle": 0, + "content": "For fine-tuning, we chose CAT-32 after preliminary hyperparameter explorations, a choice that was later confirmed to perform well in our ablation studies (see Tab. 2), though CAT-K improved performance for a wide range of values of \\( K \\). We also found that choosing a sufficiently high \\( K_{\\mathrm{infer}} \\) for top-K sampling during inference is important for a high RMM [9], and we chose \\( K_{\\mathrm{infer}} = 48 \\) for our leaderboard submission based on local validation results (see Fig. 3). In all of our experiments, we fixed the inference time sampling temperature to 1.0 and did not use top-p (nucleus) sampling [8], though we expect that tuning these hyperparameters can lead to further performance gains." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.63, + 0.906, + 0.855 + ], + "angle": 0, + "content": "To ensure that the observed performance gains with CAT-K are not due to improved hyperparameter tuning, we also conduct a large-scale hyperparameter grid search for the baseline method SMART-tiny. This includes training for more epochs, adjusting the learning rate and learning rate scheduler, experimenting with various data augmentation and data pre-processing strategies, and using various values of \\( K_{\\mathrm{infer}} \\) for the inference-time top-K sampling. This allowed us to push the performance of the BC baseline to 0.7671 RMM on the test split (not shown in Tab. 1), but still falling significantly short of the performance of our CAT-K fine-tuning method (0.7702 RMM). Note that even comparing against this improved baseline, our method improves the performance more (+0.0031) than scaling up the model size by a factor of 14 to 102M parameters (+0.0023)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.905, + 0.9 + ], + "angle": 0, + "content": "Qualitatively, our method can generate diverse and realistic behavior over long periods of time. As shown in Fig. 6, our method can handle the subtle interactions between traf" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "5427" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.098, + 0.089, + 0.9, + 0.378 + ], + "angle": 0, + "content": "
Local val. split\nMethodCriterion\nof topKK for topKSampled fromNext targetRMM\n↑Kinematic metrics ↑Interactive metrics ↑Map-based metrics ↑min ADE ↓
BC pre-training---GT0.75810.45120.80760.86971.3152
BC fine-tuning---GT0.75900.45140.80960.87001.3039
Trajeglish's noisy\ntokenizationneg. dist.5†neg. dist.GT0.75620.44690.80740.86731.3459
neg. dist.5†uniformGT0.75540.44670.80690.86551.3404
neg. dist.32neg. dist.GT0.74010.41740.79850.84931.6669
SMART's trajectory\nperturbationneg. dist.5†uniformRO0.75560.44400.80820.86611.3177
neg. dist.5†neg. dist.RO0.75600.44690.80690.86731.3514
neg. dist.32uniformRO0.73140.41580.79490.83001.5380
Top-5prob5probGT0.64780.33130.68470.75281.8802
Top-5 + distance filterprob5probGT0.68600.33560.74660.80831.7627
Top-5 + distance based samplingprob5neg. dist.GT0.70580.35360.75790.84001.5848
Deterministic rollout--max-probGT0.63610.32910.68450.74921.8695
CAT-5prob-closestGT0.74230.42510.79170.86011.4677
CAT-16prob-closestGT0.76040.45920.80820.87091.3372
CAT-32 (leaderboard)prob-closestGT0.76160.45830.81050.87201.3105
CAT-40prob-closestGT0.76170.45670.81010.87381.2998
CAT-64prob-closestGT0.76020.45520.80980.87071.3028
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.379, + 0.907, + 0.435 + ], + "angle": 0, + "content": "Table 2. Ablation study on WOSAC \\(2\\%\\) validation split. We compare different ways to fine-tune the same base mode (BC pre-training). \"Sampled from\" indicates how the action is sampled during fine-tuning, either based on the distance to the GT (\"neg. dist\", \"uniform\", \"closest\") or based on the model outputs (\"prob\", \"max-prob\"). Here dist. is the abbreviation of distance. RO stands for rollout, i.e., the next target action is computed based on the rollout, not the GT state. \\(\\dagger\\) indicates original hyperparameter choices of baseline algorithms." + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.44, + 0.475, + 0.546 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.119, + 0.546, + 0.455, + 0.56 + ], + "angle": 0, + "content": "Figure 4. On server vs. local evaluation of SMART-tiny." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.573, + 0.483, + 0.651 + ], + "angle": 0, + "content": "fic participants in a dense parking lot, which is arguably a more challenging scenario than intersections and highways for traffic simulation. Moreover, the behavior remains realistic at the end of the required simulation time of 8 seconds. Additional examples are in the supplementary videos." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.656, + 0.337, + 0.67 + ], + "angle": 0, + "content": "5.3.2. Ablation studies on WOSAC" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.674, + 0.483, + 0.84 + ], + "angle": 0, + "content": "In Tab. 2 we provide a thorough ablation study. Due to the high cost of evaluation we use \\(2\\%\\) of the validation split (880 out of 44097 scenarios). To verify the fidelity of this evaluation setting, we compare results with those on the full validation set in Fig. 4. We observe consistent differences indicating that our evaluation setting is reasonable. We begin with a SMART-tiny model trained with BC for 32 epochs (BC pre-training, row 1). All other models fine-tune the BC pre-training model for 5 epochs. During inference, we use top-K sampling with \\(K_{\\mathrm{infer}} = 40\\) for all methods (the best \\(K_{\\mathrm{infer}}\\) for SMART-tiny according to Fig. 3)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "In Tab. 2, CAT-K fine-tuning is the only method that significantly outperforms the BC pre-training model. Further fine-tuning with BC (row 2), using Trajeglish's noisy tokenization (rows 3-5), or SMART's trajectory perturbation" + }, + { + "type": "image", + "bbox": [ + 0.534, + 0.44, + 0.883, + 0.546 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.53, + 0.546, + 0.889, + 0.56 + ], + "angle": 0, + "content": "Figure 5. ADE between CAT-K rollouts and GT trajectories." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.581, + 0.907, + 0.763 + ], + "angle": 0, + "content": "rows 6-8) remain on par with the original model or even reduce its performance. Closed-loop fine-tuning with top-K rollouts instead of CAT-K rollouts during training significantly reduces performance (rows 9-12), even when rollouts close to the GT are selected for training by either filtering them based on the distance to the GT or sampling among them from a distance-dependent distribution. The observation that data augmentation and fine-tuning with top-K sampling cannot improve the RMM is consistent with prior works [19, 23, 24, 35] and the WOSAC leaderboard. Ablations on different values of \\( K \\) for top-K sampling and other hyperparameters are provided in the appendix." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Next, we ablate the value of \\( K \\) used for CAT-K rollout during fine-tuning (rows 13-17). Results indicate that the performance improvement is robust to the choice of \\( K \\) after a reasonable minimum value. As discussed in Sec. 4.3, the hyperparameter \\( K \\) in CAT-K determines how closely the policy follows the GT trajectory, in a way that is more robust than a distance based threshold, for which the optimal value varies strongly based on the situation (vehicle speed, proximity to other cars, etc.). To give more insight into how" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "5428" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.093, + 0.089, + 0.252, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.255, + 0.089, + 0.414, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.418, + 0.089, + 0.576, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.58, + 0.089, + 0.738, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.741, + 0.089, + 0.9, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.286, + 0.907, + 0.37 + ], + "angle": 0, + "content": "Figure 6. Simulation results on WOSAC. Our fine-tuned policy generates interesting and diverse behaviors rarely seen in prior works. Each row represents a different rollout of our model in the same scene. The transparent boxes show the GT agents in the dataset, while the solid boxes show the agents generated by our model. We highlight the agents within the red, green, and orange rectangles, across time steps and rollouts. The red rectangle shows different interactive negotiations emerging between a pedestrian and two vehicles. The green rectangle shows an initially parked vehicle, that leaves (row 1) or stays parked (row 2, 3). The orange rectangle shows a vehicle waiting (row 1) in front of a speed bump (visualized as two light-blue lines), proceeding (row 2), or entering a parking space (row 3)." + }, + { + "type": "table", + "bbox": [ + 0.182, + 0.377, + 0.816, + 0.461 + ], + "angle": 0, + "content": "
Method (Local val. split)Collision rate ↓Off-road rate ↓RMM ↑ADE ↓minADE32 ↓
BC pre-training0.05680.00530.81081.36231.3537
BC fine-tuning0.05990.00580.81051.35201.3509
Deterministic rollout0.04330.01380.80811.17990.7962
CAT-30.04220.00350.81691.30960.6912
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.463, + 0.907, + 0.52 + ], + "angle": 0, + "content": "Table 3. Performance of ego policies on WOSAC with local evaluation on \\(2\\%\\) validation split. All models are fine-tuned for 5 epochs based on the BC pre-training model, which is trained for 32 epochs. We use deterministic rollout during inference and compute all metrics, except for the \\(\\min\\mathrm{ADE}^{32}\\). For \\(\\min\\mathrm{ADE}^{32}\\), we generate 32 rollouts by using top-3 sampling with a temperature of 1.0 to first sample the categorical distribution over the mixtures, then selecting the mean of the sampled Gaussian mixture." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.54, + 0.483, + 0.662 + ], + "angle": 0, + "content": "\\(K\\) impacts rollouts, in Fig. 5, we inspect the average ADE between rollouts and GT trajectories over training epochs, for different \\(K\\) values. As expected, as \\(K \\to |V|\\), the ADE decreases towards the level of quantization error, induced by tokenization with a finite vocabulary size (dashed line). With more fine-tuning epochs the average ADE slightly reduces, highlighting how CAT-K fine-tuning improves the policy to follow all behavior modes more closely." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.67, + 0.399, + 0.687 + ], + "angle": 0, + "content": "5.3.3. Fine-tuning a GMM-based ego-policy" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Besides NTP traffic simulation polices, CAT-K fine-tuning also improves the performance of a GMM-based ego-policy with continuous action space. In Tab. 3 we compare our approach with fine-tuning using deterministic rollouts, as well as with continued BC. Our CAT-3 fine-tuning improves all metrics except for ADE, where deterministic rollout performs better. This is expected, as deterministic rollout aligns all modes towards the GT, resulting in mode averaging. While this reduces ADE, it negatively impacts other metrics. Additionally, CAT-K fine-tuning helps overcome the limitation of GMM trained with negative log-likelihood loss, which struggles to capture multimodality as effectively as NTP policies trained with cross-entropy loss. This is supported by the significant drop in minADE after fine-tuning." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.539, + 0.634, + 0.555 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.564, + 0.907, + 0.699 + ], + "angle": 0, + "content": "In this paper, we introduce CAT-K rollouts, a closed-loop supervised fine-tuning technique for IL problems with highly multimodal demonstrations, such as traffic simulation. The CAT-K rollout approximately finds the rollout closest to the GT among likely rollouts of a policy, ensuring adherence to the policy while maintaining GT as a reliable reference for supervised learning. As the first method using closed-loop fine-tuning, it achieves the top spot on the WOSAC leaderboard." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.7, + 0.909, + 0.897 + ], + "angle": 0, + "content": "In the future, we aim to incorporate modern sampling techniques into CAT-K rollouts, such as top-p sampling. Furthermore, instead of always selecting the closest token, we could loosen the \"winner-takes-all\" approach and sample from a group of nearby tokens. We also plan to explore CAT-K fine-tuning for a broader range of policy classes, such as variational autoencoders and diffusion models. Additionally, we want to apply it to other multimodal IL tasks, including end-to-end driving, motion generation for animation, and robot navigation and manipulation. Importantly, our results show that closed-loop supervised fine-tuning is a promising area of future research for policies trained in open-loop, such as the widely used NTP policies." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5429" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.091, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.169 + ], + "angle": 0, + "content": "[1] Mayank Bansal, Alex Krizhevsky, and Abhijit Ogale. Chauffeurnet: Learning to drive by imitating the best and synthesizing the worst. arXiv preprint arXiv:1812.03079, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.173, + 0.484, + 0.337 + ], + "angle": 0, + "content": "[2] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel Ziegler, Jeffrey Wu, Clemens Winter, Chris Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. Language models are few-shot learners. In Advances in Neural Information Processing Systems (NeurIPS), pages 1877-1901. Curran Associates, Inc., 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.34, + 0.483, + 0.422 + ], + "angle": 0, + "content": "[3] Holger Caesar, Juraj Kabzan, Kok Seang Tan, Whye Kit Fong, Eric M. Wolff, Alex H. Lang, Luke Fletcher, Oscar Beijbom, and Sammy Omari. NuPlan: A closed-loop ml-based planning benchmark for autonomous vehicles. In Conference on Computer Vision and Pattern Recognition (CVPR) ADP3 Workshop, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.424, + 0.483, + 0.491 + ], + "angle": 0, + "content": "[4] Yulong Cao, Boris Ivanovic, Chaowei Xiao, and Marco Pavone. Reinforcement learning with human feedback for realistic traffic simulation. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 14428-14434. IEEE, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.494, + 0.483, + 0.589 + ], + "angle": 0, + "content": "[5] Scott Ettinger, Shuyang Cheng, Benjamin Caine, Chenxi Liu, Hang Zhao, Sabeek Pradhan, Yuning Chai, Ben Sapp, Charles R Qi, Yin Zhou, et al. Large scale interactive motion forecasting for autonomous driving: The waymo open motion dataset. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 9710-9719, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.592, + 0.483, + 0.661 + ], + "angle": 0, + "content": "[6] Roger Girgis, Florian Golemo, Felipe Codevilla, Martin Weiss, Jim Aldon D'Souza, Samira Ebrahimi Kahou, Felix Heide, and Christopher Pal. Latent variable sequential set transformers for joint multi-agent motion prediction. arXiv preprint arXiv:2104.00563, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.663, + 0.483, + 0.702 + ], + "angle": 0, + "content": "[7] Jonathan Ho and Stefano Ermon. Generative adversarial imitation learning. Advances in Neural Information Processing Systems (NeurIPS), 29, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.705, + 0.483, + 0.759 + ], + "angle": 0, + "content": "[8] Ari Holtzman, Jan Buys, Li Du, Maxwell Forbes, and Yejin Choi. The curious case of neural text degeneration. In International Conference on Learning Representations (ICLR), 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.762, + 0.483, + 0.829 + ], + "angle": 0, + "content": "[9] Yihan Hu, Siqi Chai, Zhening Yang, Jingyu Qian, Kun Li, Wenxin Shao, Haichao Zhang, Wei Xu, and Qiang Liu. Solving motion planning tasks with a scalable generative model. In European Conference on Computer Vision (ECCV), 2025. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.832, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[10] Zhiyu Huang, Xinshuo Weng, Maximilian Igl, Yuxiao Chen, Yulong Cao, Boris Ivanovic, Marco Pavone, and Chen Lv. Gen-drive: Enhancing diffusion generative driving policies with reward modeling and reinforcement learning finetuning. arXiv preprint arXiv:2410.05582, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[11] Zhiyu Huang, Zixu Zhang, Ameya Vaidya, Yuxiao Chen, Chen Lv, and Jaime Fernandez Fisac. Versatile scene-consistent traffic scenario generation as optimization with diffusion. arXiv preprint arXiv:2404.02524, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.151, + 0.905, + 0.245 + ], + "angle": 0, + "content": "[12] Maximilian Igl, Daewoo Kim, Alex Kuefler, Paul Mougin, Punit Shah, Kyriacos Shiarlis, Dragomir Anguelov, Mark Palatucci, Brandyn White, and Shimon Whiteson. Symphony: Learning realistic and diverse agents for autonomous driving simulation. In 2022 International Conference on Robotics and Automation (ICRA), pages 2445-2451. IEEE, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.249, + 0.905, + 0.33 + ], + "angle": 0, + "content": "[13] Maximilian Igl, Punit Shah, Paul Mougin, Sirish Srinivasan, Tarun Gupta, Brandyn White, Kyriacos Shiarlis, and Shimon Whiteson. Hierarchical imitation learning for stochastic environments. In 2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 1697-1704. IEEE, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.334, + 0.905, + 0.402 + ], + "angle": 0, + "content": "[14] Boris Ivanovic and Marco Pavone. The trajectory: Probabilistic multi-agent trajectory modeling with dynamic spatiotemporal graphs. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 2375-2384, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.406, + 0.905, + 0.487 + ], + "angle": 0, + "content": "[15] Chiyu Jiang, Andre Cornman, Cheolho Park, Benjamin Sapp, Yin Zhou, Dragomir Anguelov, et al. Motiondiffuser: Controllable multi-agent motion prediction using diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9644-9653, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.49, + 0.905, + 0.544 + ], + "angle": 0, + "content": "[16] Longzhong Lin, Xuewu Lin, Kechun Xu, Haojian Lu, Lichao Huang, Rong Xiong, and Yue Wang. Revisit mixture models for multi-agent simulation: Experimental study within a unified framework, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.548, + 0.905, + 0.601 + ], + "angle": 0, + "content": "[17] Waymo LLC. Waymo open sim agent challenge (wosac) 2024 leaderboard. https://waymo.com/open/challenges/2024/sim-agents/, 2024. Accessed: 2025-03-14. 6, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.605, + 0.905, + 0.672 + ], + "angle": 0, + "content": "[18] Jack Lu, Kelvin Wong, Chris Zhang, Simon Suo, and Raquel Urtasun. Scenecontrol: Diffusion for controllable traffic scene generation. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 16908-16914. IEEE, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.676, + 0.905, + 0.772 + ], + "angle": 0, + "content": "[19] Yiren Lu, Justin Fu, George Tucker, Xinlei Pan, Eli Bronstein, Rebecca Roelofs, Benjamin Sapp, Brandyn White, Aleksandra Faust, Shimon Whiteson, et al. Imitation is not enough: Robustifying imitation with reinforcement learning for challenging driving scenarios. In 2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 7553-7560. IEEE, 2023. 1, 2, 7, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.775, + 0.905, + 0.856 + ], + "angle": 0, + "content": "[20] Osama Makansi, Eddy Ilg, Ozgun Cicek, and Thomas Brox. Overcoming limitations of mixture density networks: A sampling and fitting framework for multimodal future prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7144-7153, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.86, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[21] Nico Montali, John Lambert, Paul Mougin, Alex Kuefler, Nicholas Rhinehart, Michelle Li, Cole Gulino, Tristan Emrich, Zoey Yang, Shimon Whiteson, et al. The waymo open" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.956 + ], + "angle": 0, + "content": "5430" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.126, + 0.093, + 0.482, + 0.12 + ], + "angle": 0, + "content": "sim agents challenge. Advances in Neural Information Processing Systems (NeurIPS), 36, 2024. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.202 + ], + "angle": 0, + "content": "[22] Jiquan Ngiam, Benjamin Caine, Vijay Vasudevan, Zhengdong Zhang, Hao-Tien Lewis Chiang, Jeffrey Ling, Rebecca Roelofs, Alex Bewley, Chenxi Liu, Ashish Venugopal, et al. Scene transformer: A unified architecture for predicting multiple agent trajectories. arXiv preprint arXiv:2106.08417, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.204, + 0.482, + 0.273 + ], + "angle": 0, + "content": "[23] Zhenghao Peng, Wenjie Luo, Yiren Lu, Tianyi Shen, Cole Gulino, Ari Seff, and Justin Fu. Improving agent behaviors with rl fine-tuning for autonomous driving. In European Conference on Computer Vision (ECCV), pages 165-181. Springer, 2025. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.274, + 0.482, + 0.328 + ], + "angle": 0, + "content": "[24] Jonah Philion, Xue Bin Peng, and Sanja Fidler. Trajeglish: Traffic modeling as next-token prediction. In International Conference on Learning Representations (ICLR), 2024. 2, 3, 4, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.33, + 0.482, + 0.399 + ], + "angle": 0, + "content": "[25] Aditya Prakash, Aseem Behl, Eshed Ohn-Bar, Kashyap Chitta, and Andreas Geiger. Exploring data aggregation in policy learning for vision-based urban autonomous driving. In Conference on Computer Vision and Pattern Recognition (CVPR), pages 11763-11773, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.4, + 0.482, + 0.468 + ], + "angle": 0, + "content": "[26] Ethan Pronovost, Meghan Reddy Ganesina, Noureldin Hendy, Zeyu Wang, Andres Morales, Kai Wang, and Nick Roy. Scenario diffusion: Controllable driving scenario generation with diffusion. Advances in Neural Information Processing Systems (NeurIPS), 36:68873-68894, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.469, + 0.482, + 0.538 + ], + "angle": 0, + "content": "[27] Stéphane Ross, Geoffrey Gordon, and Drew Bagnell. A reduction of imitation learning and structured prediction to no-regret online learning. In International Conference on Artificial Intelligence and Statistics (AISTATS), pages 627-635. JMLR Workshop and Conference Proceedings, 2011. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.539, + 0.482, + 0.621 + ], + "angle": 0, + "content": "[28] Ari Seff, Brian Cera, Dian Chen, Mason Ng, Aurick Zhou, Nigamaa Nayakanti, Khaled S Refaat, Rami Al-Rfou, and Benjamin Sapp. Motionlm: Multi-agent motion forecasting as language modeling. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 8579-8590, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.622, + 0.482, + 0.678 + ], + "angle": 0, + "content": "[29] Shaoshuai Shi, Li Jiang, Dengxin Dai, and Bernt Schiele. Motion transformer with global intention localization and local movement refinement. Advances in Neural Information Processing Systems (NeurIPS), 35:6531-6543, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.679, + 0.482, + 0.746 + ], + "angle": 0, + "content": "[30] Shaoshuai Shi, Li Jiang, Dengxin Dai, and Bernt Schiele. Mtr++: Multi-agent motion prediction with symmetric scene modeling and guided intention querying. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.748, + 0.482, + 0.803 + ], + "angle": 0, + "content": "[31] Simon Suo, Sebastian Regalado, Sergio Casas, and Raquel Urtasun. TrafficSim: Learning to simulate realistic multi-agent behaviors. In Conference on Computer Vision and Pattern Recognition (CVPR), pages 10400-10409, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.804, + 0.482, + 0.872 + ], + "angle": 0, + "content": "[32] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[33] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems (NeurIPS), pages 5998-6008, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[34] Arun Venkatraman, Martial Hebert, and J Bagnell. Improving multi-step prediction of learned time series models. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), 2015. 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.192, + 0.905, + 0.247 + ], + "angle": 0, + "content": "[35] Wei Wu, Xiaoxin Feng, Ziyan Gao, and Yuheng Kan. Smart: Scalable multi-agent real-time simulation via next-token prediction. Advances in Neural Information Processing Systems (NeurIPS), 2025. 1, 2, 3, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.249, + 0.905, + 0.305 + ], + "angle": 0, + "content": "[36] Danfei Xu, Yuxiao Chen, Boris Ivanovic, and Marco Pavone. Bits: Bi-level imitation for traffic simulation. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 2929-2936. IEEE, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.307, + 0.905, + 0.36 + ], + "angle": 0, + "content": "[37] Chris Zhang, James Tu, Lunjun Zhang, Kelvin Wong, Simon Suo, and Raquel Urtasun. Learning realistic traffic agents in closed-loop. In Conference on Robot Learning (CoRL), 2023. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.362, + 0.905, + 0.404 + ], + "angle": 0, + "content": "[38] Jiakai Zhang and Kyunghyun Cho. Query-efficient imitation learning for end-to-end autonomous driving. arXiv preprint arXiv:1605.06450, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.406, + 0.905, + 0.475 + ], + "angle": 0, + "content": "[39] Zhejun Zhang, Alexander Liniger, Dengxin Dai, Fisher Yu, and Luc Van Gool. End-to-end urban driving by imitating a reinforcement learning coach. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.477, + 0.905, + 0.546 + ], + "angle": 0, + "content": "[40] Zhejun Zhang, Alexander Liniger, Dengxin Dai, Fisher Yu, and Luc Van Gool. Trafficbots: Towards world models for autonomous driving simulation and motion prediction. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 1522-1529. IEEE, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.548, + 0.905, + 0.616 + ], + "angle": 0, + "content": "[41] Zhejun Zhang, Alexander Liniger, Christos Sakaridis, Fisher Yu, and Luc V Gool. Real-time motion prediction via heterogeneous polyline transformer with relative pose encoding. Advances in Neural Information Processing Systems (NeurIPS), 2024. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.618, + 0.905, + 0.673 + ], + "angle": 0, + "content": "[42] Zhejun Zhang, Christos Sakaridis, and Luc Van Gool. Trafficbots v1. 5: Traffic simulation via conditional vaes and transformers with relative pose encoding. arXiv preprint arXiv:2406.10898, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.675, + 0.905, + 0.743 + ], + "angle": 0, + "content": "[43] Jianbo Zhao, Jiaheng Zhuang, Qibin Zhou, Taiyu Ban, Ziyao Xu, Hangning Zhou, Junhe Wang, Guoan Wang, Zhiheng Li, and Bin Li. Kigras: Kinematic-driven generative model for realistic agent simulation. arXiv preprint arXiv:2407.12940, 2024. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.746, + 0.905, + 0.801 + ], + "angle": 0, + "content": "[44] Ziyuan Zhong, Davis Rempe, Yuxiao Chen, Boris Ivanovic, Yulong Cao, Danfei Xu, Marco Pavone, and Baishakhi Ray. Language-guided traffic simulation via scene-level diffusion. arXiv preprint arXiv:2306.06344, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.802, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[45] Ziyuan Zhong, Davis Rempe, Danfei Xu, Yuxiao Chen, Sushant Veer, Tong Che, Baishakhi Ray, and Marco Pavone. Guided conditional diffusion for controllable traffic simulation. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 3560-3566. IEEE, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[46] Zikang Zhou, Jianping Wang, Yung-Hui Li, and Yu-Kai Huang. Query-centric trajectory prediction. In Proceedings" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.956 + ], + "angle": 0, + "content": "5431" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.12 + ], + "angle": 0, + "content": "of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17863-17873, 2023. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.192 + ], + "angle": 0, + "content": "[47] Zikang Zhou, Haibo Hu, Xinhong Chen, Jianping Wang, Nan Guan, Kui Wu, Yung-Hui Li, Yu-Kai Huang, and Chun Jason Xue. Behaviorgpt: Smart agent simulation for autonomous driving with next-patch prediction. arXiv preprint arXiv:2405.17372, 2024. 1, 6, 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.192 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.956 + ], + "angle": 0, + "content": "5432" + } + ] +] \ No newline at end of file diff --git a/2025/Closed-Loop Supervised Fine-Tuning of Tokenized Traffic Models/2dc3d2cc-243e-4ca0-9955-c4c727b1c79c_origin.pdf b/2025/Closed-Loop Supervised Fine-Tuning of Tokenized Traffic Models/2dc3d2cc-243e-4ca0-9955-c4c727b1c79c_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..2d7706deb89e16fd37dd0c7f15d7cc2f8eccca82 --- /dev/null +++ b/2025/Closed-Loop Supervised Fine-Tuning of Tokenized Traffic Models/2dc3d2cc-243e-4ca0-9955-c4c727b1c79c_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:792ab53aee26ccaeb3fff9692f2a0358c84e46ff2630215d42e06baa471f03fa +size 2382120 diff --git a/2025/Closed-Loop Supervised Fine-Tuning of Tokenized Traffic Models/full.md b/2025/Closed-Loop Supervised Fine-Tuning of Tokenized Traffic Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..c967dea0db7c855bbd231d06c78069de5936c0e0 --- /dev/null +++ b/2025/Closed-Loop Supervised Fine-Tuning of Tokenized Traffic Models/full.md @@ -0,0 +1,321 @@ +# Closed-Loop Supervised Fine-Tuning of Tokenized Traffic Models + +Zhejun Zhang $^{1}$ Peter Karkus $^{1}$ Maximilian Igl $^{1}$ + +Wenhao Ding1 Yuxiao Chen1 Boris Ivanovic1 Marco Pavone1,2 + +$^{1}$ NVIDIA Research $^{2}$ Stanford University + +{zhejunz, pkarkus, migl, wenhaod, yuxiaoc, bivanovic, mpavone}@nvidia.com + +pavone@stanford.edu + +# Abstract + +Traffic simulation aims to learn a policy for traffic agents that, when unrolled in closed-loop, faithfully recovers the joint distribution of trajectories observed in the real world. Inspired by large language models, tokenized multi-agent policies have recently become the state-of-the-art in traffic simulation. However, they are typically trained through open-loop behavior cloning, and thus suffer from covariate shift when executed in closed-loop during simulation. In this work, we present Closest Among Top-K (CAT-K) rollouts, a simple yet effective closed-loop fine-tuning strategy to mitigate covariate shift. CAT-K fine-tuning only requires existing trajectory data, without reinforcement learning or generative adversarial imitation. Concretely, CAT-K fine-tuning enables a small 7M-parameter tokenized traffic simulation policy to outperform a 102M-parameter model from the same model family, achieving the top spot on the Waymo Sim Agent Challenge leaderboard at the time of submission. The code is available at https://github.com/NVlabs/catk. + +# 1. Introduction + +Traffic modeling is a cornerstone of autonomous driving simulation and evaluation, typically formulated as learning a multi-agent policy that imitates the behavior of traffic participants in the real world. Given a set of historical agent trajectories and scene context (map, traffic light states, etc.), the policy generates actions for all simulated agents. The task gives rise to an imitation learning (IL) problem, with two key challenges: multimodality and covariate shift. + +Traffic agent behavior is highly multimodal, and faithfully recovering accurate behavior distributions is a key challenge in the field. Inspired by large language models [2, 32], recent works introduce next-token-prediction (NTP) models where the policy reduces to a classifier over a discrete set of trajectory tokens, which makes it easier to represent highly-multimodal distributions. Accordingly, + +![](images/5826def3f53e4acfaab5592110aca1388ebaf147c39a499454332abd02c6fcd6.jpg) +Figure 1. Closest Among Top-K (CAT-K) rollouts. The key idea of our approach is to unroll the policy during fine-tuning in a way that visited states remain close to the GT. At each time step, CAT-K first takes the top-K most likely action tokens according to the policy, then chooses the one leading to the state closest to the GT. As a result, CAT-K rollouts follow the mode of the GT (e.g., turning left), while random or top-K rollouts can lead to large deviations (e.g., going straight or right). Since the policy is essentially trained to minimize the distance between the rollout states and the GT states, the GT-based supervision remains effective for CAT-K rollouts, but not for random or top-K rollouts. + +the Waymo Open Sim Agent Challenge (WOSAC) leaderboard [21] is heavily populated by tokenized traffic models [9, 35, 43, 47]. + +Covariate shift is a well-known challenge of IL arising from the gap between open-loop training and closed-loop deployment. When a model is trained on a fixed dataset of expert demonstrations, it can face a distribution mismatch between the states seen during training and those encountered during deployment, as small errors compound and lead to unseen states where the policy performs poorly. A classic approach to tackle this problem is Dataset Aggregation (DAgger) [27], which unrolls the policy and queries an expert to generate new demonstrations, but querying experts is not readily available for traffic simulation. Prior work has proposed closed-loop training using hand-crafted recovery controllers [1] or reinforcement learning (RL) [19, 23, 37]. However, it is inherently difficult to design rewards with high behavioral realism or recovery controllers robust to divergent modes. Consequently, such approaches are not currently competitive on WOSAC realism metrics. + +Contributions: We introduce Closest Among Top-K (CAT-K) rollouts, a simple yet highly efficient fine-tuning strategy to address the open-loop to closed-loop gap. The key idea, illustrated in Fig. 1, is to unroll the multimodal policy during training in a way that the policy-visited states remain close to the ground-truth (GT) demonstration. CAT-K achieves this by first finding the K most likely modes of the policy and then choosing the mode closest to the GT. At inference time, actions are sampled from the policy according to the predicted likelihoods. During training, however, random sampling would lead to large deviations from the demonstrations, making them invalid and degrading the final policy performance. Our CAT-K rollout strategy balances being on-policy and staying close to GT demonstrations, such that they remain a valid supervision signal. + +Experiments on the Waymo Open Motion Dataset (WOMD) demonstrate the efficacy of CAT-K. Notably, fine-tuning the SMART-7M [35] next-token-prediction traffic model enables it to outperform the 14x larger State-of-The-Art (SoTA) SMART-102M from the same model family, achieving the #1 spot on the public WOSAC leaderboard at the time of submission. To further demonstrate the potential of employing CAT-K fine-tuning for different tasks and policy representations, we apply it to an ego-motion planning task using a Gaussian Mixture Model (GMM) policy, yielding significant gains in closed-loop behavior, reducing collisions by $25.7\%$ and off-road driving by $33.9\%$ . + +# 2. Related work + +# 2.1. Traffic simulation + +Prior work explored various architectures for traffic models, including conditional variational autoencoders [13, 31, 36], motion forecasting Transformers [6, 22, 30, 46], and diffusion models [11, 15, 18, 26, 44, 45]. However, long-term stability is an open challenge for these models due to covariate shift when transitioning from open-loop training to closed-loop deployment. Various ways have been proposed to mitigate covariate shift. BITS [36] and Symphony [12] introduce hierarchy, with high-level intent and low-level behavior prediction. TrafficBots [40, 42] incorporates configurable behaviors through destination and personality. Most diffusion models use guiding to generate rule-abiding behavior; however, they often struggle with computational efficiency and long horizons. The latest advances come from NTP models, which predict the next action as a token, e.g., Trajeglish [24], GUMP [9], KiGRAS [43], MotionLM [28], and SMART [35]. Notably, SMART is the current SoTA on the WOSAC leaderboard. In addition to the strong scalability and flexibility, NTP models also show better closed-loop stability than regression-based models, thanks to their discrete action space. However, achieving generalization and reducing compounding errors continue to be challenges. + +# 2.2. Data augmentation for behavior cloning + +Data augmentation is a simple yet effective way to improve the generalizability of traffic simulation models. ChauffeurNet [1] showed that carefully perturbing the vehicle trajectory and designing a recovery trajectory could alleviate the covariate shift suffered by behavior cloning (BC). However, this technique is difficult to apply to traffic simulation with complicated scenarios, including pedestrians and cyclists; and adding handcrafted recovery trajectories may negatively impact behavioral realism. Recently, NTP works such as Trajeglish [24] and SMART [35] have explored similar ideas by using noisy tokenization to perturb trajectories during training, but their data augmentation did not lead to significant improvements in performance. Our method is related to Trajeglish's noisy tokenization, but importantly, instead of blindly sampling tokens close to the GT without considering the policy, our CAT-K rollout selects the token from the most likely K tokens predicted by the policy that is closest to the GT. It can be considered as a closed-loop variants of the "winner-takes-all" training strategy [20] widely applied to open-loop motion prediction models [22, 29, 41]. + +# 2.3. Closed-loop fine-tuning + +As mentioned above, covariate shift is a major challenge faced by traffic simulation models, since most of them are trained in open-loop and evaluated in closed-loop. Even with data augmentation, the issue is not fully resolved as the augmented noisy data does not reflect the compounded error during closed-loop rollout. Therefore, some existing works explored the use of closed-loop fine-tuning. + +As a classic remedy for the covariate shift, DAgger has been applied to end-to-end driving [25, 38, 39], yet its application in traffic simulation is limited, as it requires interactive demonstrations from human drivers or an expert policy. As an expert-free variant of DAgger, Data As Demonstrator (DAD) [34] obtains recovery actions using a pair of trajectories: a forward simulation (i.e., rollout) trajectory and a GT trajectory. At each time step, the policy is trained to return to the GT next state from the simulated current state. However, DAD fails when the rollout deviates from the GT trajectory. Our CAT-K rollout directly addresses this problem, enabling the use of DAD for closed-loop fine-tuning of traffic simulation and ego-vehicle policies. + +Another popular approach is to use RL. BC is combined with RL to improve the robustness of a policy in [19]. Yet, it also exposes RL's weakness in improving realism as it is difficult to handcraft a reward that promotes realism. Its follow-up work [23] learns a joint traffic model capable of rolling out the entire scenario by itself. The reward/loss are typically handcrafted, and can be distilled from explicit traffic rules [37]. With human preference data, RL from human feedback (RLHF) has also been applied to traffic model training for better user alignment [4, 10]. Finally, Sym- + +phony [12] adds a Generative Adversarial Imitation Learning (GAIL) loss to encourage the rollout states to stay in distribution; however, the well-known issue of training stability and discriminator overfitting remains a challenge. + +# 3. Background + +# 3.1. Problem formulation + +A multi-agent traffic simulation policy can be typically formulated as $\pi_{\theta}(\mathbf{a}_t|\mathbf{h}_t,\mathcal{M})$ where $\theta$ denotes the trainable model parameters, $\mathbf{h}_t = \mathbf{s}_{t - H:t}$ is the state history of length $H$ $\mathcal{M}$ is the context, including for example highdefinition (HD) maps and traffic light states, $t$ is the current time step, $\mathbf{a}_t = [a_t^1,\dots,a_t^N ]$ and $\mathbf{s}_t = [s_t^1,\dots,s_t^N ]$ are respectively the actions and states of $N$ agents at the current time step. The dimensions of actions $a$ and states $S$ are respectively denoted as $D_{a}$ and $D_{s}$ , i.e., $a\in \mathbb{R}^{D_a}$ and $s\in \mathbb{R}^{D_s}$ . From the current states and actions at step $t$ , the next states are computed using the per-agent forward dynamics $\mathbf{s}_{t + 1} = f(\mathbf{s}_t,\mathbf{a}_t) = \left[f(s_t^i,a_t^i)\right]_{i = 1}^N$ . We assume that $f(s_{t}^{i},a_{t}^{i})$ is deterministic, and can be queried during training, which is the case for traffic simulation. Extensions to stochastic dynamics would be possible in future work. We define a rollout of $T$ steps starting at $t = 0$ as a sequence of states $\mathbf{s}_{0:T} = [\mathbf{s}_0,\dots,\mathbf{s}_T]$ , while the GT trajectories of all agents are denoted as $\hat{\mathbf{s}}_{0:T}$ . For the training, we are given a dataset $\mathcal{D} = \{\hat{\mathbf{s}}_{0:T}^j,\mathcal{M}^j\}_{j = 1}^{|D|}$ of such real-world trajectories that we want to emulate with their corresponding contexts. + +# 3.2. Next Token Prediction (NTP) policies + +NTP policies, such as SMART [35] and Trajeglish [24], are parameterized as a probability distribution over a vocabulary of action tokens denoted as $V = \{x_{c} \mid c = 1,2,\ldots ,|V|\}$ , where $|V|$ is the size of the vocabulary, $x_{c} \in \mathbb{R}^{D_{a}}$ are template actions and $c \in \mathbb{N}$ is the token index. Hence, an autoregressive NTP policy for traffic simulation can equivalently be written as an agent-factorized categorical distribution at each timestep $t$ , i.e., + +$$ +\pi_ {\theta} (\mathbf {c} _ {t} \mid \mathbf {h} _ {t}, \mathcal {M}) = \prod_ {i = 1} ^ {N} \pi_ {\theta} (c _ {t} ^ {i} \mid \mathbf {h} _ {t}, \mathcal {M}) = \prod_ {i = 1} ^ {N} \mathrm {C a t} (c _ {t} ^ {i}), +$$ + +where $\mathrm{Cat}(c_t^i)$ is the categorical distribution over the action token index for agent $i$ (and not to be confused with our method CAT-K). Given the sampled output $\mathbf{c}_t = [c_t^1, \dots, c_t^N]$ , the actions $\mathbf{a}_t = [x_{c_t^1}, \dots, x_{c_t^N}]$ are obtained using the token vocabulary $V$ . + +# 4. Method + +Two key challenges in learning a policy from real-world trajectories are the multimodal nature of the trajectory distribution and the problem of covariate shift when policies are + +Algorithm 1 CAT-K fine-tuning +1: Input: Policy $\pi_{\theta}$ , action token vocabulary $V$ , dataset $\mathcal{D}$ +2: Pre-train $\pi_{\theta}(\mathbf{c}_t \mid \hat{\mathbf{h}}_t, \mathcal{M})$ with BC until convergence +3: repeat ▷ Closed-loop supervised fine-tuning +4: Sample a traffic scenario $\{\hat{s}_{0:T}, \mathcal{M}\}$ +5: Init rollout state $\mathbf{s}_0 = \hat{\mathbf{s}}_0$ ▷ CAT-K Rollout +6: for $t$ in $[0, \dots, T-1]$ do ▷ $T$ steps +7: for $i$ in $[1, \dots, N]$ do ▷ $N$ agents +8: Get action index for rollout $c_t^i$ . (Eq. 1) +9: Get next rollout state $s_{t+1}^i$ . (Eq. 3) +10: Compute target $\hat{c}_t^i$ . (Eq. 4) +11: end for +12: end for +13: Update $\theta$ by minimizing $\mathcal{L}_{\theta}(\mathbf{s}_{0:T}, \hat{\mathbf{c}}_{1:T}, \mathcal{M})$ . (Eq. 5) +14: until convergence + +trained open-loop, resulting in a distribution mismatch between expert states seen during training and states visited during policy deployment. Covariate shift can be overcome by closed-loop training, i.e., by training on trajectories sampled from the learned policy. However, this requires the generation of expert actions (or other notions of optimality) to be used as training targets along those trajectories [27]. Querying a human expert is infeasible at scale, RL-based methods require hard-to-define rewards, and methods such as GAIL [7] are prone to mode collapse. Consequently, they fail to achieve good performance in the WOSAC challenge. + +An alternative strategy for generating "expert" actions is to construct recovery actions that bring the agent back to the available GT trajectory. However, this is complicated by the multimodal nature of the data, as the available GT trajectory might not be a valid recovery target for the generated trajectory. For example, as shown in Fig. 1, the GT trajectory $\hat{s}_{0:T}$ turns left at the intersection, while the sampled trajectory $s_{0:T} \sim \pi_{\theta}$ might go straight or turn right. As a result, while some SoTA traffic models [24, 35] augment the training data with recovery actions to reduce the covariate shift, they do so only from states that were reached by injecting small amounts of noise into the GT trajectory. This does guarantee that the GT trajectory remains a valid recovery target, but it completely ignores the learned policy and the state distribution induced by it. Instead, our method, Closest Among Top-K (CAT-K) rollout, informs the sampling process by the learned policy, but biases it towards the GT trajectory to guarantee the validity of the recovery actions. While simple to implement, it yields significant performance improvements in our experiments. + +# 4.1. Closest Among Top-K (CAT-K) rollout + +To facilitate the formulation, we define a top $K$ operator: + +$$ +\{\xi_ {1}, \dots , \xi_ {K} \} = \operatorname * {t o p} _ {c \in \{1, \dots , | V | \}} ^ {K} (\operatorname {C a t} (c)), +$$ + +![](images/8f4cd23f6138102ff7538ee0289e9f1e9d3dff057449df6728ddcac3bc964411.jpg) +(a) CAT-K rollout + +![](images/6b2b69c710da47b2cbcd063c122ba307e476abdae9d37f9601fa10f9f61a9ea5.jpg) +(b) Top-K sampling +Figure 2. Schematic comparison of CAT-K rollout, top-K sampling, and data augmentation techniques of Trajeglish and SMART. In this example, the token vocabulary has a size of 5. We rollout three steps from $t = 0$ to $t = 3$ . For CAT-K rollout and top-K sampling, the top-K is w.r.t the probabilities $p$ of tokens predicted by the policy. For the data augmentations used by Trajeglish and SMART, the policy is unavailable, and the top-K selection is based on the negative distances between tokens and GT. + +![](images/d2d24500cd37190dcf9257f87d47dcb6a2c6a2f95bd9a5f670f394bdd1b11fc5.jpg) +(c) Trajeglish noisy tokenization + +![](images/cca8454ae73ec2cf6cf84fd5c8ab361955d017b58f6c19b1d868395ec13f50ed.jpg) +(d) SMART trajectory perturbation + +where $\operatorname{Cat}:\mathbb{N}\to \mathbb{R}$ is the probability density of a categorical distribution on the vocabulary index, and $\{\xi_1,\dots ,\xi_K\}$ are the $K$ most probable indices. The top operator can be considered as a variation of the arg max operator that returns multiple indices, with top equivalent to arg max. + +At time step $t$ , the policy $\pi_{\theta}(\mathbf{c}_t \mid \mathbf{h}_t, \mathcal{M})$ outputs independent categorical distributions over the token vocabulary for each agent. Our method, CAT-K rollout, deterministically rolls out the policy by selecting, at each time step and for each agent, the one action among the top-K likeliest according to $\pi_{\theta}$ that brings the agent closest to the GT next state. Using a distance metric $d(\cdot, \cdot)$ on the states, this is formally expressed as follows: + +$$ +c _ {t} ^ {i} = \underset {c \in \left\{\xi_ {1}, \dots , \xi_ {K} \right\}} {\arg \min } d \left(f \left(s _ {t} ^ {i}, x _ {c}\right), \hat {s} _ {t + 1} ^ {i}\right), \tag {1} +$$ + +$$ +\left\{\xi_ {1}, \dots , \xi_ {K} \right\} = \underset {c _ {t} ^ {i} \in \{1, \dots , | V | \}} {\operatorname {t o p} ^ {K}} \left[ \pi \left(c _ {t} ^ {i} \mid \mathbf {h} _ {t}, \mathcal {M}\right) \right], \tag {2} +$$ + +where $c_{t}^{i}$ is the action token indices of agent $i$ at step $t$ for the CAT-K rollout, and $\{\xi_1,\dots ,\xi_K\}$ are the top-K likeliest token index according to the policy. Given $c_{t}^{i}$ , the next state is obtained using the vocabulary $V$ and the dynamics as: + +$$ +s _ {t + 1} ^ {i} = f \left(s _ {t} ^ {i}, x _ {c _ {t} ^ {i}}\right). \tag {3} +$$ + +These rollout states will be used as the input $\mathbf{h}$ to the policy at the next time step. By doing this sequentially from $t = 0$ to $t = T - 1$ and repeating for all $N$ agents, we obtain the CAT-K rollout trajectories $\mathbf{s}_{0:T}$ . + +# 4.2. Closed-loop supervised fine-tuning + +Given the CAT-K rollout trajectory $\mathbf{s}_{0:T}$ , we can apply the idea of DAD [34] and construct the recovery action indices $\hat{\mathbf{c}}_t$ from the GT trajectories $\hat{\mathbf{s}}_{0:T}$ by finding the action token that brings each agent closest to its original trajectory: + +$$ +\hat {c} _ {t} ^ {i} = \underset {c \in \{1, \dots , | V | \}} {\arg \min } d \left(f \left(s _ {t} ^ {i}, x _ {c}\right), \hat {s} _ {t + 1} ^ {i}\right). \tag {4} +$$ + +Given these indices $\hat{\mathbf{c}}_t$ , the NTP policy is trained using the cross-entropy loss + +$$ +\mathcal {L} _ {\theta} = - \frac {1}{N T} \sum_ {i = 1} ^ {N} \sum_ {t = 0} ^ {T - 1} \log \pi_ {\theta} \left(\hat {c} _ {t} ^ {i} \mid \mathbf {h} _ {t}, \mathcal {M}\right). \tag {5} +$$ + +Since CAT-K rollout is effective only when the top-K rollouts of the policy cover the GT mode, we adopt a two-stage training procedure, summarized in Algorithm 1. First, we obtain a reasonably well-trained policy through BC pretraining, then fine-tune it using CAT-K rollouts. + +# 4.3. Comparison to previous methods + +In Fig. 2, we compare our method with previous data augmentation approaches that alleviate the covariate shift for NTP traffic simulation policies. + +Top-K sampling is a common approach used by NTP models for generating sequences during inference as it improves the sample quality. However, it is unsuitable for generating trajectories during training as it does not consider the distance to the GT trajectory, and hence the validity of generated recovery actions (c.f. Fig. 2b). This can be partially addressed by post-hoc filtering of trajectories that deviated too far from the GT, but the resulting method is sample-inefficient, as most rollouts will be discarded, and the distance threshold hyperparameter is difficult to tune, since its optimal value varies across scenarios (e.g., high-speed vs. low-speed situations). In contrast, CAT-K rollout (c.f. Fig. 2a) is sample efficient by choosing the closest among top-K actions at each timestep and also removes the need to tune a distance-based hyperparameter. As shown in Tab. 2, our hyperparameter $K$ provides strong results across a large range of values and is hence much easier to tune. + +Instead of sampling top-K trajectories from the policy, current SoTA methods such as Trajeglish [24] and SMART [35] use forms of trajectory noising to address the issue of covariate shift. All trajectory noising approaches rely + +on the injection of small perturbations into the tokenization of the GT trajectory, but implementations can vary from each other in details. For example, in Trajeglish, the likelihood of each noised token is a function of the resulting distance to the ground-truth trajectory, i.e. $q_{t}^{i}(c) \sim \exp(-d(f(s_{t}^{i}, x_{c}), \hat{s}_{t+1}^{i}) / \tau)$ , while for SMART, tokens are sampled uniformly from those $K$ tokens with the highest likelihood $q_{t}^{i}(c)$ . Additionally, Trajeglish, similar to our methods, uses the DAD [34] recovery target that finds the token that would bring the agent back to the GT trajectory (c.f. Fig. 2c), while SMART uses the next, also noised, token as target - effectively treating the noise injection as data augmentation on the training trajectories (c.f. Fig. 2d). + +However, Trajeglish's noisy tokenization does not yield a significant improvement (see Fig. 9 in [24]). Similarly, while SMART's trajectory perturbation enhances the performance of a zero-shot policy trained on NuPlan [3] and evaluated on WOSAC, it does not improve performance on WOSAC itself (see Tab. 4 in [35]). We believe this is because the state distributions generated by their sampling strategies are sub-optimal as they completely ignore what state distribution would be induced by the learned policy, hence likely oversampling irrelevant states and undersampling states the learned policy would actually encounter. By incorporating the learned policy into the sampling strategy, CAT-K generates a state distribution more like the policy's, and is hence better able to reduce the covariate shift between training and inference. Although sampling from the policy incurs higher data generation costs, these can be effectively mitigated through the two-phase training strategy with BC pre-training and closed-loop fine-tuning. + +Lastly, note that CAT-K with $K = |V|$ is equivalent to noise-free BC, since in this case rollouts follow the GT as closely as the available token book allows. On the other hand, for $K = 1$ , CAT-K is equivalent to deterministically rolling out the policy by always choosing the most likely token. Consequently, for CAT-K, the hyperparameter $K$ trades off following the policy (for $K = 1$ ) vs. following the GT (for $K = |V|$ ). In Fig. 5 we show how various choices of $K$ impact the Average Displacement Error (ADE) between the CAT-K rollouts and the GT. The intuition behind CAT-K is to find, among all the top-K T-step rollouts (with a total of $K^T$ possibilities), the one with the minimum ADE (minADE) to the GT. However, since directly solving this optimization problem is difficult, we use CAT-K as a greedy approximation to solve it sequentially. + +# 5. Experiments + +To show the broad relevance of CAT-K fine-tuning, we evaluate its performance on two different tasks using two different types of policy architecture, namely a traffic simulation task using a NTP policy and an ego-motion planning task using a policy parameterized as GMM. We base both tasks + +on the widely used WOMD [5] and follow the simulation setup proposed in WOSAC [21], namely providing 1 second of history and generating rollouts of 8 seconds length. + +# 5.1. Traffic Simulation + +Metrics. In the traffic simulation task, we follow the WOSAC protocol: for each scenario, we generate 32 simulated rollouts for all agents in the scene, at $10\mathrm{Hz}$ , and evaluate how well their distribution matches that of the human demonstrations in the data. This distributional matching is evaluated along three dimensions, namely kinematics (e.g., velocities and accelerations), interactions (e.g., collisions), and map alignment (e.g., off-road). All three are summarized as weighted averages in the "Realism Meta-Metric" (RMM), the key performance indicator of the WOSAC leaderboard. For more details please see [21]. We also report the minADE, which is not used in the RMM, but widely applied for motion prediction and policy evaluation. + +Policy. We use SMART [35] as our policy architecture due to its strong performance and open-source implementation. Specifically, we use the SMART-tiny (7M) model and its open-sourced token vocabularies. The map polyline token vocabulary has a size of 1024, with each token representing a polyline consisting of 10 segments, each 0.5 meters long. The agent trajectory token vocabulary has a size of 2048, with each token representing a 0.5-second trajectory at $10\mathrm{Hz}$ . SMART re-plans at $2\mathrm{Hz}$ , hence the final output is at $10\mathrm{Hz}$ when submitted to the WOSAC leaderboard. However, our experiments show that SMART's open-sourced agent token vocabulary results in lower kinematic metrics. To address this issue, we reran the K-disk clustering [24] with a larger number of trajectories. By using our own agent token vocabulary, the performance on the WOSAC leaderboard is significantly improved (c.f. Tab. 1). Please note that all other experiments in this paper (i.e., except those in Tab. 1) still use SMART's open-sourced token vocabularies. SMART uses Transformers [33] with query-centric representations [41, 46], and the distance metric $d(\cdot, \cdot)$ is the average Euclidean distance between the four pairs of corners of two bounding boxes. We disable SMART's trajectory perturbation as it deteriorates performance on the WOSAC leaderboard (see [35] Tab. 4). We run BC pretraining for 32 epochs, each taking 1.7 hour, then continue with closed-loop supervised fine-tuning with CAT-32 for 10 epochs, each taking 2.6 hours. During fine-tuning, the map encoder is frozen to save GPU memory, allowing for a larger training batch size with negligible impact on performance. For both pre-training and fine-tuning, we use $8 \times$ A100 80GB GPUs with a total batch size of 80. + +# 5.2. Ego-motion planning + +Metrics. Unlike traffic simulation, in ego-motion planning only the ego-vehicle is controlled and evaluated. Moreover, + +
Leaderboard, test split +Method# model +paramsRMM +↑RMM diff. to +SMART-largeKinematic +metrics ↑Interactive +metrics ↑Map-based +metrics ↑min +ADE ↓
SMART-tiny fine-tuned w. CAT-K (ours)7 M0.7702+0.00880.49310.81190.87491.3068
SMART-large [35]102 M0.7614+0.00000.47860.80660.86481.3728
KiGRAS [43]0.7 M0.7597-0.00140.46910.80640.86581.4384
SMART-tiny [35]7 M0.7591-0.00230.47590.80390.86321.4062
FDriver-tiny7 M0.7584-0.00300.46140.80690.86581.4475
SMART [35]8 M0.7511-0.01030.44450.80500.85711.5447
BehaviorGPT [47]3 M0.7473-0.01410.43330.79970.85931.4147
GUMP [9]523 M0.7431-0.01830.47800.78870.83591.6041
+ +Table 1. Results on the WOSAC 2024 leaderboard [17]. RMM stands for Realism Meta Metric, the key metric used for ranking. Note that on the public leaderboard [17] our method appears under the name "SMART-tiny-CLSFT" (Closed-Loop Supervised Fine-Tuning). + +![](images/558b1ca35205381da0e84a506a8d3b693a40841a3649ddb6830ab7615582e586.jpg) +Figure 3. Influence of $K_{\mathrm{infer}}$ for inference-time top-K sampling. + +the optimal policy is not the one that perfectly matches the GT trajectory distribution, but which minimizes planning metrics, such as collision rate, off-road rate, and ADE. For completeness, we also report RMM, applied only to ego trajectories, and minADE over 32 sampled rollouts. With the exception of minADE, and in contrast to the traffic simulation task, the rollouts for evaluation are sampled deterministically, by choosing the most-likely mode at each timestep. + +Policy. To show that our fine-tuning approach can be applied to a more general class of policies beyond NTP, we parameterized the output of the ego-policy as GMM. Specifically, we replace the final layer of the SMART network with two heads, a classification head that predicts the mixture densities and a regression head that predicts the means of the Gaussian distributions with fixed standard deviations (similar to, e.g., [14]). Specifically, the GMM predicts 16 modes, each representing a 3-dimensional Gaussian distribution over the changes in x- and y-position and yaw heading. The final model comprises a total of 3.2M parameters. In contrast to the SMART traffic simulation policy that uses cross-entropy loss on all agents, the GMM-based egopolicy is trained using negative log-likelihood loss and only predicts actions for the ego-vehicle. When applying CAT-K rollout or top-K sampling to the GMM, we select the K most-likely modes of the GMM, and within each mode deterministically use the mean of the Gaussian. Please note that ego-motion planning is a separate experiment from traffic simulation; hence, the ego-policy is not fine-tuned on the traffic simulation policy. Also note that the ego-policy operates in a mid-to-end manner, taking tracking results and HD maps as inputs, rather than in a fully end-to-end fashion that maps raw sensor observations to actions. + +# 5.3. Results + +# 5.3.1. WOSAC leaderboard for traffic simulation + +In Tab. 1 we compare our approach with other traffic simulation policies on the WOSAC leaderboard. Notably, all the top-ranking methods on the leaderboard are NTP policies trained via BC. Our SMART-tiny model using CAT-K fine-tuning outperforms the previous SoTA, SMART-large with 102M parameters, by a significant margin of $+0.0088$ ; and improves on SMART-tiny by $+0.0111$ . As the first to perform closed-loop fine-tuning on the leaderboard, our approach improves all metrics and sets a new SoTA. + +For fine-tuning, we chose CAT-32 after preliminary hyperparameter explorations, a choice that was later confirmed to perform well in our ablation studies (see Tab. 2), though CAT-K improved performance for a wide range of values of $K$ . We also found that choosing a sufficiently high $K_{\mathrm{infer}}$ for top-K sampling during inference is important for a high RMM [9], and we chose $K_{\mathrm{infer}} = 48$ for our leaderboard submission based on local validation results (see Fig. 3). In all of our experiments, we fixed the inference time sampling temperature to 1.0 and did not use top-p (nucleus) sampling [8], though we expect that tuning these hyperparameters can lead to further performance gains. + +To ensure that the observed performance gains with CAT-K are not due to improved hyperparameter tuning, we also conduct a large-scale hyperparameter grid search for the baseline method SMART-tiny. This includes training for more epochs, adjusting the learning rate and learning rate scheduler, experimenting with various data augmentation and data pre-processing strategies, and using various values of $K_{\mathrm{infer}}$ for the inference-time top-K sampling. This allowed us to push the performance of the BC baseline to 0.7671 RMM on the test split (not shown in Tab. 1), but still falling significantly short of the performance of our CAT-K fine-tuning method (0.7702 RMM). Note that even comparing against this improved baseline, our method improves the performance more (+0.0031) than scaling up the model size by a factor of 14 to 102M parameters (+0.0023). + +Qualitatively, our method can generate diverse and realistic behavior over long periods of time. As shown in Fig. 6, our method can handle the subtle interactions between traf + +
Local val. split +MethodCriterion +of topKK for topKSampled fromNext targetRMM +↑Kinematic metrics ↑Interactive metrics ↑Map-based metrics ↑min ADE ↓
BC pre-training---GT0.75810.45120.80760.86971.3152
BC fine-tuning---GT0.75900.45140.80960.87001.3039
Trajeglish's noisy +tokenizationneg. dist.5†neg. dist.GT0.75620.44690.80740.86731.3459
neg. dist.5†uniformGT0.75540.44670.80690.86551.3404
neg. dist.32neg. dist.GT0.74010.41740.79850.84931.6669
SMART's trajectory +perturbationneg. dist.5†uniformRO0.75560.44400.80820.86611.3177
neg. dist.5†neg. dist.RO0.75600.44690.80690.86731.3514
neg. dist.32uniformRO0.73140.41580.79490.83001.5380
Top-5prob5probGT0.64780.33130.68470.75281.8802
Top-5 + distance filterprob5probGT0.68600.33560.74660.80831.7627
Top-5 + distance based samplingprob5neg. dist.GT0.70580.35360.75790.84001.5848
Deterministic rollout--max-probGT0.63610.32910.68450.74921.8695
CAT-5prob-closestGT0.74230.42510.79170.86011.4677
CAT-16prob-closestGT0.76040.45920.80820.87091.3372
CAT-32 (leaderboard)prob-closestGT0.76160.45830.81050.87201.3105
CAT-40prob-closestGT0.76170.45670.81010.87381.2998
CAT-64prob-closestGT0.76020.45520.80980.87071.3028
+ +Table 2. Ablation study on WOSAC $2\%$ validation split. We compare different ways to fine-tune the same base mode (BC pre-training). "Sampled from" indicates how the action is sampled during fine-tuning, either based on the distance to the GT ("neg. dist", "uniform", "closest") or based on the model outputs ("prob", "max-prob"). Here dist. is the abbreviation of distance. RO stands for rollout, i.e., the next target action is computed based on the rollout, not the GT state. $\dagger$ indicates original hyperparameter choices of baseline algorithms. + +![](images/c7f34f8b1aaf5c7e0377b71e667b22c87ce2f03e92b5c9d311fa88c28deaa85f.jpg) +Figure 4. On server vs. local evaluation of SMART-tiny. + +fic participants in a dense parking lot, which is arguably a more challenging scenario than intersections and highways for traffic simulation. Moreover, the behavior remains realistic at the end of the required simulation time of 8 seconds. Additional examples are in the supplementary videos. + +# 5.3.2. Ablation studies on WOSAC + +In Tab. 2 we provide a thorough ablation study. Due to the high cost of evaluation we use $2\%$ of the validation split (880 out of 44097 scenarios). To verify the fidelity of this evaluation setting, we compare results with those on the full validation set in Fig. 4. We observe consistent differences indicating that our evaluation setting is reasonable. We begin with a SMART-tiny model trained with BC for 32 epochs (BC pre-training, row 1). All other models fine-tune the BC pre-training model for 5 epochs. During inference, we use top-K sampling with $K_{\mathrm{infer}} = 40$ for all methods (the best $K_{\mathrm{infer}}$ for SMART-tiny according to Fig. 3). + +In Tab. 2, CAT-K fine-tuning is the only method that significantly outperforms the BC pre-training model. Further fine-tuning with BC (row 2), using Trajeglish's noisy tokenization (rows 3-5), or SMART's trajectory perturbation + +![](images/7fd159d160ec228322ae3be9c8fa4dd8a4f9f2866f069ef960c13031264f4f4b.jpg) +Figure 5. ADE between CAT-K rollouts and GT trajectories. + +rows 6-8) remain on par with the original model or even reduce its performance. Closed-loop fine-tuning with top-K rollouts instead of CAT-K rollouts during training significantly reduces performance (rows 9-12), even when rollouts close to the GT are selected for training by either filtering them based on the distance to the GT or sampling among them from a distance-dependent distribution. The observation that data augmentation and fine-tuning with top-K sampling cannot improve the RMM is consistent with prior works [19, 23, 24, 35] and the WOSAC leaderboard. Ablations on different values of $K$ for top-K sampling and other hyperparameters are provided in the appendix. + +Next, we ablate the value of $K$ used for CAT-K rollout during fine-tuning (rows 13-17). Results indicate that the performance improvement is robust to the choice of $K$ after a reasonable minimum value. As discussed in Sec. 4.3, the hyperparameter $K$ in CAT-K determines how closely the policy follows the GT trajectory, in a way that is more robust than a distance based threshold, for which the optimal value varies strongly based on the situation (vehicle speed, proximity to other cars, etc.). To give more insight into how + +![](images/142ca603bb68ea37ecd251c565617b76663bddccdb672d82fc9c708d6966a61b.jpg) +Figure 6. Simulation results on WOSAC. Our fine-tuned policy generates interesting and diverse behaviors rarely seen in prior works. Each row represents a different rollout of our model in the same scene. The transparent boxes show the GT agents in the dataset, while the solid boxes show the agents generated by our model. We highlight the agents within the red, green, and orange rectangles, across time steps and rollouts. The red rectangle shows different interactive negotiations emerging between a pedestrian and two vehicles. The green rectangle shows an initially parked vehicle, that leaves (row 1) or stays parked (row 2, 3). The orange rectangle shows a vehicle waiting (row 1) in front of a speed bump (visualized as two light-blue lines), proceeding (row 2), or entering a parking space (row 3). + +![](images/03f8c0e01b5cecfa9e9deaff410f72e69e919f5baffe32152c5736896e22acca.jpg) + +![](images/f523e846d9100a9e7d1bd3460223686115386bfc54f2d6ce803dbdd846d616bb.jpg) + +![](images/f1253b0be5349f0499b6eb5fbdb964683e7d2932e2886746342da5b8fc82b0d0.jpg) + +![](images/a21d42867d7ea82036d4a5422e8dac4d3690d87e9db4ae3e6938a448cb2b2548.jpg) + +
Method (Local val. split)Collision rate ↓Off-road rate ↓RMM ↑ADE ↓minADE32 ↓
BC pre-training0.05680.00530.81081.36231.3537
BC fine-tuning0.05990.00580.81051.35201.3509
Deterministic rollout0.04330.01380.80811.17990.7962
CAT-30.04220.00350.81691.30960.6912
+ +Table 3. Performance of ego policies on WOSAC with local evaluation on $2\%$ validation split. All models are fine-tuned for 5 epochs based on the BC pre-training model, which is trained for 32 epochs. We use deterministic rollout during inference and compute all metrics, except for the $\min\mathrm{ADE}^{32}$ . For $\min\mathrm{ADE}^{32}$ , we generate 32 rollouts by using top-3 sampling with a temperature of 1.0 to first sample the categorical distribution over the mixtures, then selecting the mean of the sampled Gaussian mixture. + +$K$ impacts rollouts, in Fig. 5, we inspect the average ADE between rollouts and GT trajectories over training epochs, for different $K$ values. As expected, as $K \to |V|$ , the ADE decreases towards the level of quantization error, induced by tokenization with a finite vocabulary size (dashed line). With more fine-tuning epochs the average ADE slightly reduces, highlighting how CAT-K fine-tuning improves the policy to follow all behavior modes more closely. + +# 5.3.3. Fine-tuning a GMM-based ego-policy + +Besides NTP traffic simulation polices, CAT-K fine-tuning also improves the performance of a GMM-based ego-policy with continuous action space. In Tab. 3 we compare our approach with fine-tuning using deterministic rollouts, as well as with continued BC. Our CAT-3 fine-tuning improves all metrics except for ADE, where deterministic rollout performs better. This is expected, as deterministic rollout aligns all modes towards the GT, resulting in mode averaging. While this reduces ADE, it negatively impacts other metrics. Additionally, CAT-K fine-tuning helps overcome the limitation of GMM trained with negative log-likelihood loss, which struggles to capture multimodality as effectively as NTP policies trained with cross-entropy loss. This is supported by the significant drop in minADE after fine-tuning. + +# 6. Conclusion + +In this paper, we introduce CAT-K rollouts, a closed-loop supervised fine-tuning technique for IL problems with highly multimodal demonstrations, such as traffic simulation. The CAT-K rollout approximately finds the rollout closest to the GT among likely rollouts of a policy, ensuring adherence to the policy while maintaining GT as a reliable reference for supervised learning. As the first method using closed-loop fine-tuning, it achieves the top spot on the WOSAC leaderboard. + +In the future, we aim to incorporate modern sampling techniques into CAT-K rollouts, such as top-p sampling. Furthermore, instead of always selecting the closest token, we could loosen the "winner-takes-all" approach and sample from a group of nearby tokens. We also plan to explore CAT-K fine-tuning for a broader range of policy classes, such as variational autoencoders and diffusion models. Additionally, we want to apply it to other multimodal IL tasks, including end-to-end driving, motion generation for animation, and robot navigation and manipulation. Importantly, our results show that closed-loop supervised fine-tuning is a promising area of future research for policies trained in open-loop, such as the widely used NTP policies. + +# References + +[1] Mayank Bansal, Alex Krizhevsky, and Abhijit Ogale. Chauffeurnet: Learning to drive by imitating the best and synthesizing the worst. arXiv preprint arXiv:1812.03079, 2018. 1, 2 +[2] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel Ziegler, Jeffrey Wu, Clemens Winter, Chris Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. Language models are few-shot learners. In Advances in Neural Information Processing Systems (NeurIPS), pages 1877-1901. Curran Associates, Inc., 2020. 1 +[3] Holger Caesar, Juraj Kabzan, Kok Seang Tan, Whye Kit Fong, Eric M. Wolff, Alex H. Lang, Luke Fletcher, Oscar Beijbom, and Sammy Omari. NuPlan: A closed-loop ml-based planning benchmark for autonomous vehicles. In Conference on Computer Vision and Pattern Recognition (CVPR) ADP3 Workshop, 2021. 5 +[4] Yulong Cao, Boris Ivanovic, Chaowei Xiao, and Marco Pavone. Reinforcement learning with human feedback for realistic traffic simulation. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 14428-14434. IEEE, 2024. 2 +[5] Scott Ettinger, Shuyang Cheng, Benjamin Caine, Chenxi Liu, Hang Zhao, Sabeek Pradhan, Yuning Chai, Ben Sapp, Charles R Qi, Yin Zhou, et al. Large scale interactive motion forecasting for autonomous driving: The waymo open motion dataset. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 9710-9719, 2021. 5 +[6] Roger Girgis, Florian Golemo, Felipe Codevilla, Martin Weiss, Jim Aldon D'Souza, Samira Ebrahimi Kahou, Felix Heide, and Christopher Pal. Latent variable sequential set transformers for joint multi-agent motion prediction. arXiv preprint arXiv:2104.00563, 2021. 2 +[7] Jonathan Ho and Stefano Ermon. Generative adversarial imitation learning. Advances in Neural Information Processing Systems (NeurIPS), 29, 2016. 3 +[8] Ari Holtzman, Jan Buys, Li Du, Maxwell Forbes, and Yejin Choi. The curious case of neural text degeneration. In International Conference on Learning Representations (ICLR), 2020. 6 +[9] Yihan Hu, Siqi Chai, Zhening Yang, Jingyu Qian, Kun Li, Wenxin Shao, Haichao Zhang, Wei Xu, and Qiang Liu. Solving motion planning tasks with a scalable generative model. In European Conference on Computer Vision (ECCV), 2025. 1, 2, 6 +[10] Zhiyu Huang, Xinshuo Weng, Maximilian Igl, Yuxiao Chen, Yulong Cao, Boris Ivanovic, Marco Pavone, and Chen Lv. Gen-drive: Enhancing diffusion generative driving policies with reward modeling and reinforcement learning finetuning. arXiv preprint arXiv:2410.05582, 2024. 2 + +[11] Zhiyu Huang, Zixu Zhang, Ameya Vaidya, Yuxiao Chen, Chen Lv, and Jaime Fernandez Fisac. Versatile scene-consistent traffic scenario generation as optimization with diffusion. arXiv preprint arXiv:2404.02524, 2024. 2 +[12] Maximilian Igl, Daewoo Kim, Alex Kuefler, Paul Mougin, Punit Shah, Kyriacos Shiarlis, Dragomir Anguelov, Mark Palatucci, Brandyn White, and Shimon Whiteson. Symphony: Learning realistic and diverse agents for autonomous driving simulation. In 2022 International Conference on Robotics and Automation (ICRA), pages 2445-2451. IEEE, 2022. 2, 3 +[13] Maximilian Igl, Punit Shah, Paul Mougin, Sirish Srinivasan, Tarun Gupta, Brandyn White, Kyriacos Shiarlis, and Shimon Whiteson. Hierarchical imitation learning for stochastic environments. In 2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 1697-1704. IEEE, 2023. 2 +[14] Boris Ivanovic and Marco Pavone. The trajectory: Probabilistic multi-agent trajectory modeling with dynamic spatiotemporal graphs. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 2375-2384, 2019. 6 +[15] Chiyu Jiang, Andre Cornman, Cheolho Park, Benjamin Sapp, Yin Zhou, Dragomir Anguelov, et al. Motiondiffuser: Controllable multi-agent motion prediction using diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9644-9653, 2023. 2 +[16] Longzhong Lin, Xuewu Lin, Kechun Xu, Haojian Lu, Lichao Huang, Rong Xiong, and Yue Wang. Revisit mixture models for multi-agent simulation: Experimental study within a unified framework, 2025. 2 +[17] Waymo LLC. Waymo open sim agent challenge (wosac) 2024 leaderboard. https://waymo.com/open/challenges/2024/sim-agents/, 2024. Accessed: 2025-03-14. 6, 2 +[18] Jack Lu, Kelvin Wong, Chris Zhang, Simon Suo, and Raquel Urtasun. Scenecontrol: Diffusion for controllable traffic scene generation. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 16908-16914. IEEE, 2024. 2 +[19] Yiren Lu, Justin Fu, George Tucker, Xinlei Pan, Eli Bronstein, Rebecca Roelofs, Benjamin Sapp, Brandyn White, Aleksandra Faust, Shimon Whiteson, et al. Imitation is not enough: Robustifying imitation with reinforcement learning for challenging driving scenarios. In 2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 7553-7560. IEEE, 2023. 1, 2, 7, 3 +[20] Osama Makansi, Eddy Ilg, Ozgun Cicek, and Thomas Brox. Overcoming limitations of mixture density networks: A sampling and fitting framework for multimodal future prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7144-7153, 2019. 2 +[21] Nico Montali, John Lambert, Paul Mougin, Alex Kuefler, Nicholas Rhinehart, Michelle Li, Cole Gulino, Tristan Emrich, Zoey Yang, Shimon Whiteson, et al. The waymo open + +sim agents challenge. Advances in Neural Information Processing Systems (NeurIPS), 36, 2024. 1, 5 +[22] Jiquan Ngiam, Benjamin Caine, Vijay Vasudevan, Zhengdong Zhang, Hao-Tien Lewis Chiang, Jeffrey Ling, Rebecca Roelofs, Alex Bewley, Chenxi Liu, Ashish Venugopal, et al. Scene transformer: A unified architecture for predicting multiple agent trajectories. arXiv preprint arXiv:2106.08417, 2021. 2 +[23] Zhenghao Peng, Wenjie Luo, Yiren Lu, Tianyi Shen, Cole Gulino, Ari Seff, and Justin Fu. Improving agent behaviors with rl fine-tuning for autonomous driving. In European Conference on Computer Vision (ECCV), pages 165-181. Springer, 2025. 1, 2, 7 +[24] Jonah Philion, Xue Bin Peng, and Sanja Fidler. Trajeglish: Traffic modeling as next-token prediction. In International Conference on Learning Representations (ICLR), 2024. 2, 3, 4, 5, 7 +[25] Aditya Prakash, Aseem Behl, Eshed Ohn-Bar, Kashyap Chitta, and Andreas Geiger. Exploring data aggregation in policy learning for vision-based urban autonomous driving. In Conference on Computer Vision and Pattern Recognition (CVPR), pages 11763-11773, 2020. 2 +[26] Ethan Pronovost, Meghan Reddy Ganesina, Noureldin Hendy, Zeyu Wang, Andres Morales, Kai Wang, and Nick Roy. Scenario diffusion: Controllable driving scenario generation with diffusion. Advances in Neural Information Processing Systems (NeurIPS), 36:68873-68894, 2023. 2 +[27] Stéphane Ross, Geoffrey Gordon, and Drew Bagnell. A reduction of imitation learning and structured prediction to no-regret online learning. In International Conference on Artificial Intelligence and Statistics (AISTATS), pages 627-635. JMLR Workshop and Conference Proceedings, 2011. 1, 3 +[28] Ari Seff, Brian Cera, Dian Chen, Mason Ng, Aurick Zhou, Nigamaa Nayakanti, Khaled S Refaat, Rami Al-Rfou, and Benjamin Sapp. Motionlm: Multi-agent motion forecasting as language modeling. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 8579-8590, 2023. 2 +[29] Shaoshuai Shi, Li Jiang, Dengxin Dai, and Bernt Schiele. Motion transformer with global intention localization and local movement refinement. Advances in Neural Information Processing Systems (NeurIPS), 35:6531-6543, 2022. 2 +[30] Shaoshuai Shi, Li Jiang, Dengxin Dai, and Bernt Schiele. Mtr++: Multi-agent motion prediction with symmetric scene modeling and guided intention querying. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 2024. 2 +[31] Simon Suo, Sebastian Regalado, Sergio Casas, and Raquel Urtasun. TrafficSim: Learning to simulate realistic multi-agent behaviors. In Conference on Computer Vision and Pattern Recognition (CVPR), pages 10400-10409, 2021. 2 +[32] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 1 +[33] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia + +Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems (NeurIPS), pages 5998-6008, 2017. 5 +[34] Arun Venkatraman, Martial Hebert, and J Bagnell. Improving multi-step prediction of learned time series models. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), 2015. 2, 4, 5 +[35] Wei Wu, Xiaoxin Feng, Ziyan Gao, and Yuheng Kan. Smart: Scalable multi-agent real-time simulation via next-token prediction. Advances in Neural Information Processing Systems (NeurIPS), 2025. 1, 2, 3, 4, 5, 6, 7 +[36] Danfei Xu, Yuxiao Chen, Boris Ivanovic, and Marco Pavone. Bits: Bi-level imitation for traffic simulation. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 2929-2936. IEEE, 2023. 2 +[37] Chris Zhang, James Tu, Lunjun Zhang, Kelvin Wong, Simon Suo, and Raquel Urtasun. Learning realistic traffic agents in closed-loop. In Conference on Robot Learning (CoRL), 2023. 1, 2, 3 +[38] Jiakai Zhang and Kyunghyun Cho. Query-efficient imitation learning for end-to-end autonomous driving. arXiv preprint arXiv:1605.06450, 2016. 2 +[39] Zhejun Zhang, Alexander Liniger, Dengxin Dai, Fisher Yu, and Luc Van Gool. End-to-end urban driving by imitating a reinforcement learning coach. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 2 +[40] Zhejun Zhang, Alexander Liniger, Dengxin Dai, Fisher Yu, and Luc Van Gool. Trafficbots: Towards world models for autonomous driving simulation and motion prediction. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 1522-1529. IEEE, 2023. 2 +[41] Zhejun Zhang, Alexander Liniger, Christos Sakaridis, Fisher Yu, and Luc V Gool. Real-time motion prediction via heterogeneous polyline transformer with relative pose encoding. Advances in Neural Information Processing Systems (NeurIPS), 2024. 2, 5 +[42] Zhejun Zhang, Christos Sakaridis, and Luc Van Gool. Trafficbots v1. 5: Traffic simulation via conditional vaes and transformers with relative pose encoding. arXiv preprint arXiv:2406.10898, 2024. 2 +[43] Jianbo Zhao, Jiaheng Zhuang, Qibin Zhou, Taiyu Ban, Ziyao Xu, Hangning Zhou, Junhe Wang, Guoan Wang, Zhiheng Li, and Bin Li. Kigras: Kinematic-driven generative model for realistic agent simulation. arXiv preprint arXiv:2407.12940, 2024. 1, 2, 6 +[44] Ziyuan Zhong, Davis Rempe, Yuxiao Chen, Boris Ivanovic, Yulong Cao, Danfei Xu, Marco Pavone, and Baishakhi Ray. Language-guided traffic simulation via scene-level diffusion. arXiv preprint arXiv:2306.06344, 2023. 2 +[45] Ziyuan Zhong, Davis Rempe, Danfei Xu, Yuxiao Chen, Sushant Veer, Tong Che, Baishakhi Ray, and Marco Pavone. Guided conditional diffusion for controllable traffic simulation. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 3560-3566. IEEE, 2023. 2 +[46] Zikang Zhou, Jianping Wang, Yung-Hui Li, and Yu-Kai Huang. Query-centric trajectory prediction. In Proceedings + +of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17863-17873, 2023. 2, 5 +[47] Zikang Zhou, Haibo Hu, Xinhong Chen, Jianping Wang, Nan Guan, Kui Wu, Yung-Hui Li, Yu-Kai Huang, and Chun Jason Xue. Behaviorgpt: Smart agent simulation for autonomous driving with next-patch prediction. arXiv preprint arXiv:2405.17372, 2024. 1, 6, 2 \ No newline at end of file diff --git a/2025/Closed-Loop Supervised Fine-Tuning of Tokenized Traffic Models/images.zip b/2025/Closed-Loop Supervised Fine-Tuning of Tokenized Traffic Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..28154098c80e3445213c2a8653d9f5a15043316f --- /dev/null +++ b/2025/Closed-Loop Supervised Fine-Tuning of Tokenized Traffic Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c288453f73114a82a7f305c876f47371a73fa4d421baabe6a850446f8e489524 +size 539583 diff --git a/2025/Closed-Loop Supervised Fine-Tuning of Tokenized Traffic Models/layout.json b/2025/Closed-Loop Supervised Fine-Tuning of Tokenized Traffic Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..5bf6fe56ebd9412cc7dedc787d086c316375f021 --- /dev/null +++ b/2025/Closed-Loop Supervised Fine-Tuning of Tokenized Traffic Models/layout.json @@ -0,0 +1,8858 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 102, + 103, + 507, + 120 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 103, + 507, + 120 + ], + "spans": [ + { + "bbox": [ + 102, + 103, + 507, + 120 + ], + "type": "text", + "content": "Closed-Loop Supervised Fine-Tuning of Tokenized Traffic Models" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 181, + 143, + 428, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 181, + 143, + 428, + 157 + ], + "spans": [ + { + "bbox": [ + 181, + 143, + 428, + 157 + ], + "type": "text", + "content": "Zhejun Zhang" + }, + { + "bbox": [ + 181, + 143, + 428, + 157 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 181, + 143, + 428, + 157 + ], + "type": "text", + "content": " Peter Karkus" + }, + { + "bbox": [ + 181, + 143, + 428, + 157 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 181, + 143, + 428, + 157 + ], + "type": "text", + "content": " Maximilian Igl" + }, + { + "bbox": [ + 181, + 143, + 428, + 157 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 136, + 157, + 474, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 157, + 474, + 171 + ], + "spans": [ + { + "bbox": [ + 136, + 157, + 474, + 171 + ], + "type": "text", + "content": "Wenhao Ding1 Yuxiao Chen1 Boris Ivanovic1 Marco Pavone1,2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 201, + 171, + 410, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 201, + 171, + 410, + 186 + ], + "spans": [ + { + "bbox": [ + 201, + 171, + 410, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 201, + 171, + 410, + 186 + ], + "type": "text", + "content": "NVIDIA Research " + }, + { + "bbox": [ + 201, + 171, + 410, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 201, + 171, + 410, + 186 + ], + "type": "text", + "content": "Stanford University" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 108, + 187, + 502, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 187, + 502, + 200 + ], + "spans": [ + { + "bbox": [ + 108, + 187, + 502, + 200 + ], + "type": "text", + "content": "{zhejunz, pkarkus, migl, wenhaod, yuxiaoc, bivanovic, mpavone}@nvidia.com" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 251, + 202, + 356, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 251, + 202, + 356, + 213 + ], + "spans": [ + { + "bbox": [ + 251, + 202, + 356, + 213 + ], + "type": "text", + "content": "pavone@stanford.edu" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 151, + 241, + 200, + 253 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 241, + 200, + 253 + ], + "spans": [ + { + "bbox": [ + 151, + 241, + 200, + 253 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 266, + 296, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 266, + 296, + 492 + ], + "spans": [ + { + "bbox": [ + 53, + 266, + 296, + 492 + ], + "type": "text", + "content": "Traffic simulation aims to learn a policy for traffic agents that, when unrolled in closed-loop, faithfully recovers the joint distribution of trajectories observed in the real world. Inspired by large language models, tokenized multi-agent policies have recently become the state-of-the-art in traffic simulation. However, they are typically trained through open-loop behavior cloning, and thus suffer from covariate shift when executed in closed-loop during simulation. In this work, we present Closest Among Top-K (CAT-K) rollouts, a simple yet effective closed-loop fine-tuning strategy to mitigate covariate shift. CAT-K fine-tuning only requires existing trajectory data, without reinforcement learning or generative adversarial imitation. Concretely, CAT-K fine-tuning enables a small 7M-parameter tokenized traffic simulation policy to outperform a 102M-parameter model from the same model family, achieving the top spot on the Waymo Sim Agent Challenge leaderboard at the time of submission. The code is available at https://github.com/NVlabs/catk." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 514, + 135, + 526 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 514, + 135, + 526 + ], + "spans": [ + { + "bbox": [ + 55, + 514, + 135, + 526 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 534, + 295, + 630 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 295, + 630 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 295, + 630 + ], + "type": "text", + "content": "Traffic modeling is a cornerstone of autonomous driving simulation and evaluation, typically formulated as learning a multi-agent policy that imitates the behavior of traffic participants in the real world. Given a set of historical agent trajectories and scene context (map, traffic light states, etc.), the policy generates actions for all simulated agents. The task gives rise to an imitation learning (IL) problem, with two key challenges: multimodality and covariate shift." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 630, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 295, + 714 + ], + "type": "text", + "content": "Traffic agent behavior is highly multimodal, and faithfully recovering accurate behavior distributions is a key challenge in the field. Inspired by large language models [2, 32], recent works introduce next-token-prediction (NTP) models where the policy reduces to a classifier over a discrete set of trajectory tokens, which makes it easier to represent highly-multimodal distributions. Accordingly," + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 324, + 239, + 537, + 332 + ], + "blocks": [ + { + "bbox": [ + 324, + 239, + 537, + 332 + ], + "lines": [ + { + "bbox": [ + 324, + 239, + 537, + 332 + ], + "spans": [ + { + "bbox": [ + 324, + 239, + 537, + 332 + ], + "type": "image", + "image_path": "5826def3f53e4acfaab5592110aca1388ebaf147c39a499454332abd02c6fcd6.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 335, + 555, + 456 + ], + "lines": [ + { + "bbox": [ + 313, + 335, + 555, + 456 + ], + "spans": [ + { + "bbox": [ + 313, + 335, + 555, + 456 + ], + "type": "text", + "content": "Figure 1. Closest Among Top-K (CAT-K) rollouts. The key idea of our approach is to unroll the policy during fine-tuning in a way that visited states remain close to the GT. At each time step, CAT-K first takes the top-K most likely action tokens according to the policy, then chooses the one leading to the state closest to the GT. As a result, CAT-K rollouts follow the mode of the GT (e.g., turning left), while random or top-K rollouts can lead to large deviations (e.g., going straight or right). Since the policy is essentially trained to minimize the distance between the rollout states and the GT states, the GT-based supervision remains effective for CAT-K rollouts, but not for random or top-K rollouts." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 474, + 553, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 474, + 553, + 509 + ], + "spans": [ + { + "bbox": [ + 313, + 474, + 553, + 509 + ], + "type": "text", + "content": "the Waymo Open Sim Agent Challenge (WOSAC) leaderboard [21] is heavily populated by tokenized traffic models [9, 35, 43, 47]." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 312, + 510, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 510, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 312, + 510, + 555, + 713 + ], + "type": "text", + "content": "Covariate shift is a well-known challenge of IL arising from the gap between open-loop training and closed-loop deployment. When a model is trained on a fixed dataset of expert demonstrations, it can face a distribution mismatch between the states seen during training and those encountered during deployment, as small errors compound and lead to unseen states where the policy performs poorly. A classic approach to tackle this problem is Dataset Aggregation (DAgger) [27], which unrolls the policy and queries an expert to generate new demonstrations, but querying experts is not readily available for traffic simulation. Prior work has proposed closed-loop training using hand-crafted recovery controllers [1] or reinforcement learning (RL) [19, 23, 37]. However, it is inherently difficult to design rewards with high behavioral realism or recovery controllers robust to divergent modes. Consequently, such approaches are not currently competitive on WOSAC realism metrics." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5422" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 251 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 251 + ], + "type": "text", + "content": "Contributions: We introduce Closest Among Top-K (CAT-K) rollouts, a simple yet highly efficient fine-tuning strategy to address the open-loop to closed-loop gap. The key idea, illustrated in Fig. 1, is to unroll the multimodal policy during training in a way that the policy-visited states remain close to the ground-truth (GT) demonstration. CAT-K achieves this by first finding the K most likely modes of the policy and then choosing the mode closest to the GT. At inference time, actions are sampled from the policy according to the predicted likelihoods. During training, however, random sampling would lead to large deviations from the demonstrations, making them invalid and degrading the final policy performance. Our CAT-K rollout strategy balances being on-policy and staying close to GT demonstrations, such that they remain a valid supervision signal." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 253, + 295, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 253, + 295, + 397 + ], + "spans": [ + { + "bbox": [ + 55, + 253, + 295, + 397 + ], + "type": "text", + "content": "Experiments on the Waymo Open Motion Dataset (WOMD) demonstrate the efficacy of CAT-K. Notably, fine-tuning the SMART-7M [35] next-token-prediction traffic model enables it to outperform the 14x larger State-of-The-Art (SoTA) SMART-102M from the same model family, achieving the #1 spot on the public WOSAC leaderboard at the time of submission. To further demonstrate the potential of employing CAT-K fine-tuning for different tasks and policy representations, we apply it to an ego-motion planning task using a Gaussian Mixture Model (GMM) policy, yielding significant gains in closed-loop behavior, reducing collisions by " + }, + { + "bbox": [ + 55, + 253, + 295, + 397 + ], + "type": "inline_equation", + "content": "25.7\\%" + }, + { + "bbox": [ + 55, + 253, + 295, + 397 + ], + "type": "text", + "content": " and off-road driving by " + }, + { + "bbox": [ + 55, + 253, + 295, + 397 + ], + "type": "inline_equation", + "content": "33.9\\%" + }, + { + "bbox": [ + 55, + 253, + 295, + 397 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 411, + 139, + 423 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 411, + 139, + 423 + ], + "spans": [ + { + "bbox": [ + 55, + 411, + 139, + 423 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 431, + 162, + 443 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 431, + 162, + 443 + ], + "spans": [ + { + "bbox": [ + 55, + 431, + 162, + 443 + ], + "type": "text", + "content": "2.1. Traffic simulation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 450, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 450, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 450, + 295, + 713 + ], + "type": "text", + "content": "Prior work explored various architectures for traffic models, including conditional variational autoencoders [13, 31, 36], motion forecasting Transformers [6, 22, 30, 46], and diffusion models [11, 15, 18, 26, 44, 45]. However, long-term stability is an open challenge for these models due to covariate shift when transitioning from open-loop training to closed-loop deployment. Various ways have been proposed to mitigate covariate shift. BITS [36] and Symphony [12] introduce hierarchy, with high-level intent and low-level behavior prediction. TrafficBots [40, 42] incorporates configurable behaviors through destination and personality. Most diffusion models use guiding to generate rule-abiding behavior; however, they often struggle with computational efficiency and long horizons. The latest advances come from NTP models, which predict the next action as a token, e.g., Trajeglish [24], GUMP [9], KiGRAS [43], MotionLM [28], and SMART [35]. Notably, SMART is the current SoTA on the WOSAC leaderboard. In addition to the strong scalability and flexibility, NTP models also show better closed-loop stability than regression-based models, thanks to their discrete action space. However, achieving generalization and reducing compounding errors continue to be challenges." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 314, + 72, + 524, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 524, + 85 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 524, + 85 + ], + "type": "text", + "content": "2.2. Data augmentation for behavior cloning" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 89, + 553, + 329 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 89, + 553, + 329 + ], + "spans": [ + { + "bbox": [ + 313, + 89, + 553, + 329 + ], + "type": "text", + "content": "Data augmentation is a simple yet effective way to improve the generalizability of traffic simulation models. ChauffeurNet [1] showed that carefully perturbing the vehicle trajectory and designing a recovery trajectory could alleviate the covariate shift suffered by behavior cloning (BC). However, this technique is difficult to apply to traffic simulation with complicated scenarios, including pedestrians and cyclists; and adding handcrafted recovery trajectories may negatively impact behavioral realism. Recently, NTP works such as Trajeglish [24] and SMART [35] have explored similar ideas by using noisy tokenization to perturb trajectories during training, but their data augmentation did not lead to significant improvements in performance. Our method is related to Trajeglish's noisy tokenization, but importantly, instead of blindly sampling tokens close to the GT without considering the policy, our CAT-K rollout selects the token from the most likely K tokens predicted by the policy that is closest to the GT. It can be considered as a closed-loop variants of the \"winner-takes-all\" training strategy [20] widely applied to open-loop motion prediction models [22, 29, 41]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 336, + 446, + 349 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 336, + 446, + 349 + ], + "spans": [ + { + "bbox": [ + 314, + 336, + 446, + 349 + ], + "type": "text", + "content": "2.3. Closed-loop fine-tuning" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 354, + 553, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 354, + 553, + 437 + ], + "spans": [ + { + "bbox": [ + 313, + 354, + 553, + 437 + ], + "type": "text", + "content": "As mentioned above, covariate shift is a major challenge faced by traffic simulation models, since most of them are trained in open-loop and evaluated in closed-loop. Even with data augmentation, the issue is not fully resolved as the augmented noisy data does not reflect the compounded error during closed-loop rollout. Therefore, some existing works explored the use of closed-loop fine-tuning." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 438, + 554, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 438, + 554, + 593 + ], + "spans": [ + { + "bbox": [ + 313, + 438, + 554, + 593 + ], + "type": "text", + "content": "As a classic remedy for the covariate shift, DAgger has been applied to end-to-end driving [25, 38, 39], yet its application in traffic simulation is limited, as it requires interactive demonstrations from human drivers or an expert policy. As an expert-free variant of DAgger, Data As Demonstrator (DAD) [34] obtains recovery actions using a pair of trajectories: a forward simulation (i.e., rollout) trajectory and a GT trajectory. At each time step, the policy is trained to return to the GT next state from the simulated current state. However, DAD fails when the rollout deviates from the GT trajectory. Our CAT-K rollout directly addresses this problem, enabling the use of DAD for closed-loop fine-tuning of traffic simulation and ego-vehicle policies." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 594, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 554, + 713 + ], + "type": "text", + "content": "Another popular approach is to use RL. BC is combined with RL to improve the robustness of a policy in [19]. Yet, it also exposes RL's weakness in improving realism as it is difficult to handcraft a reward that promotes realism. Its follow-up work [23] learns a joint traffic model capable of rolling out the entire scenario by itself. The reward/loss are typically handcrafted, and can be distilled from explicit traffic rules [37]. With human preference data, RL from human feedback (RLHF) has also been applied to traffic model training for better user alignment [4, 10]. Finally, Sym-" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5423" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 122 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 122 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 122 + ], + "type": "text", + "content": "phony [12] adds a Generative Adversarial Imitation Learning (GAIL) loss to encourage the rollout states to stay in distribution; however, the well-known issue of training stability and discriminator overfitting remains a challenge." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 131, + 135, + 145 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 131, + 135, + 145 + ], + "spans": [ + { + "bbox": [ + 55, + 131, + 135, + 145 + ], + "type": "text", + "content": "3. Background" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 151, + 177, + 163 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 151, + 177, + 163 + ], + "spans": [ + { + "bbox": [ + 55, + 151, + 177, + 163 + ], + "type": "text", + "content": "3.1. Problem formulation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 168, + 296, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 168, + 296, + 413 + ], + "spans": [ + { + "bbox": [ + 55, + 168, + 296, + 413 + ], + "type": "text", + "content": "A multi-agent traffic simulation policy can be typically formulated as " + }, + { + "bbox": [ + 55, + 168, + 296, + 413 + ], + "type": "inline_equation", + "content": "\\pi_{\\theta}(\\mathbf{a}_t|\\mathbf{h}_t,\\mathcal{M})" + }, + { + "bbox": [ + 55, + 168, + 296, + 413 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 55, + 168, + 296, + 413 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 55, + 168, + 296, + 413 + ], + "type": "text", + "content": " denotes the trainable model parameters, " + }, + { + "bbox": [ + 55, + 168, + 296, + 413 + ], + "type": "inline_equation", + "content": "\\mathbf{h}_t = \\mathbf{s}_{t - H:t}" + }, + { + "bbox": [ + 55, + 168, + 296, + 413 + ], + "type": "text", + "content": " is the state history of length " + }, + { + "bbox": [ + 55, + 168, + 296, + 413 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 55, + 168, + 296, + 413 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 55, + 168, + 296, + 413 + ], + "type": "text", + "content": " is the context, including for example highdefinition (HD) maps and traffic light states, " + }, + { + "bbox": [ + 55, + 168, + 296, + 413 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 168, + 296, + 413 + ], + "type": "text", + "content": " is the current time step, " + }, + { + "bbox": [ + 55, + 168, + 296, + 413 + ], + "type": "inline_equation", + "content": "\\mathbf{a}_t = [a_t^1,\\dots,a_t^N ]" + }, + { + "bbox": [ + 55, + 168, + 296, + 413 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 168, + 296, + 413 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_t = [s_t^1,\\dots,s_t^N ]" + }, + { + "bbox": [ + 55, + 168, + 296, + 413 + ], + "type": "text", + "content": " are respectively the actions and states of " + }, + { + "bbox": [ + 55, + 168, + 296, + 413 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 168, + 296, + 413 + ], + "type": "text", + "content": " agents at the current time step. The dimensions of actions " + }, + { + "bbox": [ + 55, + 168, + 296, + 413 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 55, + 168, + 296, + 413 + ], + "type": "text", + "content": " and states " + }, + { + "bbox": [ + 55, + 168, + 296, + 413 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 55, + 168, + 296, + 413 + ], + "type": "text", + "content": " are respectively denoted as " + }, + { + "bbox": [ + 55, + 168, + 296, + 413 + ], + "type": "inline_equation", + "content": "D_{a}" + }, + { + "bbox": [ + 55, + 168, + 296, + 413 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 168, + 296, + 413 + ], + "type": "inline_equation", + "content": "D_{s}" + }, + { + "bbox": [ + 55, + 168, + 296, + 413 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 55, + 168, + 296, + 413 + ], + "type": "inline_equation", + "content": "a\\in \\mathbb{R}^{D_a}" + }, + { + "bbox": [ + 55, + 168, + 296, + 413 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 168, + 296, + 413 + ], + "type": "inline_equation", + "content": "s\\in \\mathbb{R}^{D_s}" + }, + { + "bbox": [ + 55, + 168, + 296, + 413 + ], + "type": "text", + "content": ". From the current states and actions at step " + }, + { + "bbox": [ + 55, + 168, + 296, + 413 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 168, + 296, + 413 + ], + "type": "text", + "content": ", the next states are computed using the per-agent forward dynamics " + }, + { + "bbox": [ + 55, + 168, + 296, + 413 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_{t + 1} = f(\\mathbf{s}_t,\\mathbf{a}_t) = \\left[f(s_t^i,a_t^i)\\right]_{i = 1}^N" + }, + { + "bbox": [ + 55, + 168, + 296, + 413 + ], + "type": "text", + "content": ". We assume that " + }, + { + "bbox": [ + 55, + 168, + 296, + 413 + ], + "type": "inline_equation", + "content": "f(s_{t}^{i},a_{t}^{i})" + }, + { + "bbox": [ + 55, + 168, + 296, + 413 + ], + "type": "text", + "content": " is deterministic, and can be queried during training, which is the case for traffic simulation. Extensions to stochastic dynamics would be possible in future work. We define a rollout of " + }, + { + "bbox": [ + 55, + 168, + 296, + 413 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 168, + 296, + 413 + ], + "type": "text", + "content": " steps starting at " + }, + { + "bbox": [ + 55, + 168, + 296, + 413 + ], + "type": "inline_equation", + "content": "t = 0" + }, + { + "bbox": [ + 55, + 168, + 296, + 413 + ], + "type": "text", + "content": " as a sequence of states " + }, + { + "bbox": [ + 55, + 168, + 296, + 413 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_{0:T} = [\\mathbf{s}_0,\\dots,\\mathbf{s}_T]" + }, + { + "bbox": [ + 55, + 168, + 296, + 413 + ], + "type": "text", + "content": ", while the GT trajectories of all agents are denoted as " + }, + { + "bbox": [ + 55, + 168, + 296, + 413 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{s}}_{0:T}" + }, + { + "bbox": [ + 55, + 168, + 296, + 413 + ], + "type": "text", + "content": ". For the training, we are given a dataset " + }, + { + "bbox": [ + 55, + 168, + 296, + 413 + ], + "type": "inline_equation", + "content": "\\mathcal{D} = \\{\\hat{\\mathbf{s}}_{0:T}^j,\\mathcal{M}^j\\}_{j = 1}^{|D|}" + }, + { + "bbox": [ + 55, + 168, + 296, + 413 + ], + "type": "text", + "content": " of such real-world trajectories that we want to emulate with their corresponding contexts." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 421, + 253, + 434 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 421, + 253, + 434 + ], + "spans": [ + { + "bbox": [ + 55, + 421, + 253, + 434 + ], + "type": "text", + "content": "3.2. Next Token Prediction (NTP) policies" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 438, + 296, + 535 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 438, + 296, + 535 + ], + "spans": [ + { + "bbox": [ + 55, + 438, + 296, + 535 + ], + "type": "text", + "content": "NTP policies, such as SMART [35] and Trajeglish [24], are parameterized as a probability distribution over a vocabulary of action tokens denoted as " + }, + { + "bbox": [ + 55, + 438, + 296, + 535 + ], + "type": "inline_equation", + "content": "V = \\{x_{c} \\mid c = 1,2,\\ldots ,|V|\\}" + }, + { + "bbox": [ + 55, + 438, + 296, + 535 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 438, + 296, + 535 + ], + "type": "inline_equation", + "content": "|V|" + }, + { + "bbox": [ + 55, + 438, + 296, + 535 + ], + "type": "text", + "content": " is the size of the vocabulary, " + }, + { + "bbox": [ + 55, + 438, + 296, + 535 + ], + "type": "inline_equation", + "content": "x_{c} \\in \\mathbb{R}^{D_{a}}" + }, + { + "bbox": [ + 55, + 438, + 296, + 535 + ], + "type": "text", + "content": " are template actions and " + }, + { + "bbox": [ + 55, + 438, + 296, + 535 + ], + "type": "inline_equation", + "content": "c \\in \\mathbb{N}" + }, + { + "bbox": [ + 55, + 438, + 296, + 535 + ], + "type": "text", + "content": " is the token index. Hence, an autoregressive NTP policy for traffic simulation can equivalently be written as an agent-factorized categorical distribution at each timestep " + }, + { + "bbox": [ + 55, + 438, + 296, + 535 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 438, + 296, + 535 + ], + "type": "text", + "content": ", i.e.," + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 543, + 285, + 578 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 543, + 285, + 578 + ], + "spans": [ + { + "bbox": [ + 67, + 543, + 285, + 578 + ], + "type": "interline_equation", + "content": "\\pi_ {\\theta} (\\mathbf {c} _ {t} \\mid \\mathbf {h} _ {t}, \\mathcal {M}) = \\prod_ {i = 1} ^ {N} \\pi_ {\\theta} (c _ {t} ^ {i} \\mid \\mathbf {h} _ {t}, \\mathcal {M}) = \\prod_ {i = 1} ^ {N} \\mathrm {C a t} (c _ {t} ^ {i}),", + "image_path": "379b217a9439018ffd5bf009a0d6d7d760637b80c58ed642ac6acccddcb9f65b.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 586, + 296, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 586, + 296, + 647 + ], + "spans": [ + { + "bbox": [ + 55, + 586, + 296, + 647 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 586, + 296, + 647 + ], + "type": "inline_equation", + "content": "\\mathrm{Cat}(c_t^i)" + }, + { + "bbox": [ + 55, + 586, + 296, + 647 + ], + "type": "text", + "content": " is the categorical distribution over the action token index for agent " + }, + { + "bbox": [ + 55, + 586, + 296, + 647 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 586, + 296, + 647 + ], + "type": "text", + "content": " (and not to be confused with our method CAT-K). Given the sampled output " + }, + { + "bbox": [ + 55, + 586, + 296, + 647 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_t = [c_t^1, \\dots, c_t^N]" + }, + { + "bbox": [ + 55, + 586, + 296, + 647 + ], + "type": "text", + "content": ", the actions " + }, + { + "bbox": [ + 55, + 586, + 296, + 647 + ], + "type": "inline_equation", + "content": "\\mathbf{a}_t = [x_{c_t^1}, \\dots, x_{c_t^N}]" + }, + { + "bbox": [ + 55, + 586, + 296, + 647 + ], + "type": "text", + "content": " are obtained using the token vocabulary " + }, + { + "bbox": [ + 55, + 586, + 296, + 647 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 55, + 586, + 296, + 647 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 657, + 111, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 657, + 111, + 670 + ], + "spans": [ + { + "bbox": [ + 55, + 657, + 111, + 670 + ], + "type": "text", + "content": "4. Method" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": "Two key challenges in learning a policy from real-world trajectories are the multimodal nature of the trajectory distribution and the problem of covariate shift when policies are" + } + ] + } + ], + "index": 9 + }, + { + "type": "code", + "bbox": [ + 316, + 86, + 556, + 257 + ], + "blocks": [ + { + "bbox": [ + 315, + 72, + 447, + 84 + ], + "lines": [ + { + "bbox": [ + 315, + 72, + 447, + 84 + ], + "spans": [ + { + "bbox": [ + 315, + 72, + 447, + 84 + ], + "type": "text", + "content": "Algorithm 1 CAT-K fine-tuning" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 316, + 86, + 556, + 257 + ], + "lines": [ + { + "bbox": [ + 316, + 86, + 556, + 257 + ], + "spans": [ + { + "bbox": [ + 316, + 86, + 556, + 257 + ], + "type": "text", + "content": "1: Input: Policy " + }, + { + "bbox": [ + 316, + 86, + 556, + 257 + ], + "type": "inline_equation", + "content": "\\pi_{\\theta}" + }, + { + "bbox": [ + 316, + 86, + 556, + 257 + ], + "type": "text", + "content": ", action token vocabulary " + }, + { + "bbox": [ + 316, + 86, + 556, + 257 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 316, + 86, + 556, + 257 + ], + "type": "text", + "content": ", dataset " + }, + { + "bbox": [ + 316, + 86, + 556, + 257 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 316, + 86, + 556, + 257 + ], + "type": "text", + "content": " \n2: Pre-train " + }, + { + "bbox": [ + 316, + 86, + 556, + 257 + ], + "type": "inline_equation", + "content": "\\pi_{\\theta}(\\mathbf{c}_t \\mid \\hat{\\mathbf{h}}_t, \\mathcal{M})" + }, + { + "bbox": [ + 316, + 86, + 556, + 257 + ], + "type": "text", + "content": " with BC until convergence \n3: repeat ▷ Closed-loop supervised fine-tuning \n4: Sample a traffic scenario " + }, + { + "bbox": [ + 316, + 86, + 556, + 257 + ], + "type": "inline_equation", + "content": "\\{\\hat{s}_{0:T}, \\mathcal{M}\\}" + }, + { + "bbox": [ + 316, + 86, + 556, + 257 + ], + "type": "text", + "content": " \n5: Init rollout state " + }, + { + "bbox": [ + 316, + 86, + 556, + 257 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_0 = \\hat{\\mathbf{s}}_0" + }, + { + "bbox": [ + 316, + 86, + 556, + 257 + ], + "type": "text", + "content": " ▷ CAT-K Rollout \n6: for " + }, + { + "bbox": [ + 316, + 86, + 556, + 257 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 316, + 86, + 556, + 257 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 316, + 86, + 556, + 257 + ], + "type": "inline_equation", + "content": "[0, \\dots, T-1]" + }, + { + "bbox": [ + 316, + 86, + 556, + 257 + ], + "type": "text", + "content": " do ▷ " + }, + { + "bbox": [ + 316, + 86, + 556, + 257 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 316, + 86, + 556, + 257 + ], + "type": "text", + "content": " steps \n7: for " + }, + { + "bbox": [ + 316, + 86, + 556, + 257 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 316, + 86, + 556, + 257 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 316, + 86, + 556, + 257 + ], + "type": "inline_equation", + "content": "[1, \\dots, N]" + }, + { + "bbox": [ + 316, + 86, + 556, + 257 + ], + "type": "text", + "content": " do ▷ " + }, + { + "bbox": [ + 316, + 86, + 556, + 257 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 316, + 86, + 556, + 257 + ], + "type": "text", + "content": " agents \n8: Get action index for rollout " + }, + { + "bbox": [ + 316, + 86, + 556, + 257 + ], + "type": "inline_equation", + "content": "c_t^i" + }, + { + "bbox": [ + 316, + 86, + 556, + 257 + ], + "type": "text", + "content": ". (Eq. 1) \n9: Get next rollout state " + }, + { + "bbox": [ + 316, + 86, + 556, + 257 + ], + "type": "inline_equation", + "content": "s_{t+1}^i" + }, + { + "bbox": [ + 316, + 86, + 556, + 257 + ], + "type": "text", + "content": ". (Eq. 3) \n10: Compute target " + }, + { + "bbox": [ + 316, + 86, + 556, + 257 + ], + "type": "inline_equation", + "content": "\\hat{c}_t^i" + }, + { + "bbox": [ + 316, + 86, + 556, + 257 + ], + "type": "text", + "content": ". (Eq. 4) \n11: end for \n12: end for \n13: Update " + }, + { + "bbox": [ + 316, + 86, + 556, + 257 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 316, + 86, + 556, + 257 + ], + "type": "text", + "content": " by minimizing " + }, + { + "bbox": [ + 316, + 86, + 556, + 257 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\theta}(\\mathbf{s}_{0:T}, \\hat{\\mathbf{c}}_{1:T}, \\mathcal{M})" + }, + { + "bbox": [ + 316, + 86, + 556, + 257 + ], + "type": "text", + "content": ". (Eq. 5) \n14: until convergence" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "code_body" + } + ], + "index": 11, + "sub_type": "algorithm" + }, + { + "bbox": [ + 313, + 278, + 555, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 278, + 555, + 411 + ], + "spans": [ + { + "bbox": [ + 313, + 278, + 555, + 411 + ], + "type": "text", + "content": "trained open-loop, resulting in a distribution mismatch between expert states seen during training and states visited during policy deployment. Covariate shift can be overcome by closed-loop training, i.e., by training on trajectories sampled from the learned policy. However, this requires the generation of expert actions (or other notions of optimality) to be used as training targets along those trajectories [27]. Querying a human expert is infeasible at scale, RL-based methods require hard-to-define rewards, and methods such as GAIL [7] are prone to mode collapse. Consequently, they fail to achieve good performance in the WOSAC challenge." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 411, + 556, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 411, + 556, + 650 + ], + "spans": [ + { + "bbox": [ + 313, + 411, + 556, + 650 + ], + "type": "text", + "content": "An alternative strategy for generating \"expert\" actions is to construct recovery actions that bring the agent back to the available GT trajectory. However, this is complicated by the multimodal nature of the data, as the available GT trajectory might not be a valid recovery target for the generated trajectory. For example, as shown in Fig. 1, the GT trajectory " + }, + { + "bbox": [ + 313, + 411, + 556, + 650 + ], + "type": "inline_equation", + "content": "\\hat{s}_{0:T}" + }, + { + "bbox": [ + 313, + 411, + 556, + 650 + ], + "type": "text", + "content": " turns left at the intersection, while the sampled trajectory " + }, + { + "bbox": [ + 313, + 411, + 556, + 650 + ], + "type": "inline_equation", + "content": "s_{0:T} \\sim \\pi_{\\theta}" + }, + { + "bbox": [ + 313, + 411, + 556, + 650 + ], + "type": "text", + "content": " might go straight or turn right. As a result, while some SoTA traffic models [24, 35] augment the training data with recovery actions to reduce the covariate shift, they do so only from states that were reached by injecting small amounts of noise into the GT trajectory. This does guarantee that the GT trajectory remains a valid recovery target, but it completely ignores the learned policy and the state distribution induced by it. Instead, our method, Closest Among Top-K (CAT-K) rollout, informs the sampling process by the learned policy, but biases it towards the GT trajectory to guarantee the validity of the recovery actions. While simple to implement, it yields significant performance improvements in our experiments." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 656, + 518, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 656, + 518, + 670 + ], + "spans": [ + { + "bbox": [ + 313, + 656, + 518, + 670 + ], + "type": "text", + "content": "4.1. Closest Among Top-K (CAT-K) rollout" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 673, + 542, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 673, + 542, + 686 + ], + "spans": [ + { + "bbox": [ + 313, + 673, + 542, + 686 + ], + "type": "text", + "content": "To facilitate the formulation, we define a top " + }, + { + "bbox": [ + 313, + 673, + 542, + 686 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 673, + 542, + 686 + ], + "type": "text", + "content": " operator:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 358, + 693, + 511, + 716 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 693, + 511, + 716 + ], + "spans": [ + { + "bbox": [ + 358, + 693, + 511, + 716 + ], + "type": "interline_equation", + "content": "\\{\\xi_ {1}, \\dots , \\xi_ {K} \\} = \\operatorname * {t o p} _ {c \\in \\{1, \\dots , | V | \\}} ^ {K} (\\operatorname {C a t} (c)),", + "image_path": "66162732dccec6adc92288bcdc7e55de6c4f71124c404b57ce5ba9b92bc99c6c.jpg" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5424" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 69, + 190, + 190 + ], + "blocks": [ + { + "bbox": [ + 56, + 69, + 190, + 190 + ], + "lines": [ + { + "bbox": [ + 56, + 69, + 190, + 190 + ], + "spans": [ + { + "bbox": [ + 56, + 69, + 190, + 190 + ], + "type": "image", + "image_path": "8f4cd23f6138102ff7538ee0289e9f1e9d3dff057449df6728ddcac3bc964411.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 94, + 194, + 153, + 202 + ], + "lines": [ + { + "bbox": [ + 94, + 194, + 153, + 202 + ], + "spans": [ + { + "bbox": [ + 94, + 194, + 153, + 202 + ], + "type": "text", + "content": "(a) CAT-K rollout" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 194, + 70, + 313, + 192 + ], + "blocks": [ + { + "bbox": [ + 194, + 70, + 313, + 192 + ], + "lines": [ + { + "bbox": [ + 194, + 70, + 313, + 192 + ], + "spans": [ + { + "bbox": [ + 194, + 70, + 313, + 192 + ], + "type": "image", + "image_path": "6b2b69c710da47b2cbcd063c122ba307e476abdae9d37f9601fa10f9f61a9ea5.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 220, + 194, + 285, + 204 + ], + "lines": [ + { + "bbox": [ + 220, + 194, + 285, + 204 + ], + "spans": [ + { + "bbox": [ + 220, + 194, + 285, + 204 + ], + "type": "text", + "content": "(b) Top-K sampling" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 206, + 555, + 251 + ], + "lines": [ + { + "bbox": [ + 55, + 206, + 555, + 251 + ], + "spans": [ + { + "bbox": [ + 55, + 206, + 555, + 251 + ], + "type": "text", + "content": "Figure 2. Schematic comparison of CAT-K rollout, top-K sampling, and data augmentation techniques of Trajeglish and SMART. In this example, the token vocabulary has a size of 5. We rollout three steps from " + }, + { + "bbox": [ + 55, + 206, + 555, + 251 + ], + "type": "inline_equation", + "content": "t = 0" + }, + { + "bbox": [ + 55, + 206, + 555, + 251 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 206, + 555, + 251 + ], + "type": "inline_equation", + "content": "t = 3" + }, + { + "bbox": [ + 55, + 206, + 555, + 251 + ], + "type": "text", + "content": ". For CAT-K rollout and top-K sampling, the top-K is w.r.t the probabilities " + }, + { + "bbox": [ + 55, + 206, + 555, + 251 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 55, + 206, + 555, + 251 + ], + "type": "text", + "content": " of tokens predicted by the policy. For the data augmentations used by Trajeglish and SMART, the policy is unavailable, and the top-K selection is based on the negative distances between tokens and GT." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 316, + 76, + 432, + 192 + ], + "blocks": [ + { + "bbox": [ + 316, + 76, + 432, + 192 + ], + "lines": [ + { + "bbox": [ + 316, + 76, + 432, + 192 + ], + "spans": [ + { + "bbox": [ + 316, + 76, + 432, + 192 + ], + "type": "image", + "image_path": "d2d24500cd37190dcf9257f87d47dcb6a2c6a2f95bd9a5f670f394bdd1b11fc5.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 321, + 194, + 426, + 203 + ], + "lines": [ + { + "bbox": [ + 321, + 194, + 426, + 203 + ], + "spans": [ + { + "bbox": [ + 321, + 194, + 426, + 203 + ], + "type": "text", + "content": "(c) Trajeglish noisy tokenization" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 436, + 76, + 553, + 192 + ], + "blocks": [ + { + "bbox": [ + 436, + 76, + 553, + 192 + ], + "lines": [ + { + "bbox": [ + 436, + 76, + 553, + 192 + ], + "spans": [ + { + "bbox": [ + 436, + 76, + 553, + 192 + ], + "type": "image", + "image_path": "cca8454ae73ec2cf6cf84fd5c8ab361955d017b58f6c19b1d868395ec13f50ed.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 437, + 194, + 550, + 203 + ], + "lines": [ + { + "bbox": [ + 437, + 194, + 550, + 203 + ], + "spans": [ + { + "bbox": [ + 437, + 194, + 550, + 203 + ], + "type": "text", + "content": "(d) SMART trajectory perturbation" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 265, + 295, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 265, + 295, + 324 + ], + "spans": [ + { + "bbox": [ + 55, + 265, + 295, + 324 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 265, + 295, + 324 + ], + "type": "inline_equation", + "content": "\\operatorname{Cat}:\\mathbb{N}\\to \\mathbb{R}" + }, + { + "bbox": [ + 55, + 265, + 295, + 324 + ], + "type": "text", + "content": " is the probability density of a categorical distribution on the vocabulary index, and " + }, + { + "bbox": [ + 55, + 265, + 295, + 324 + ], + "type": "inline_equation", + "content": "\\{\\xi_1,\\dots ,\\xi_K\\}" + }, + { + "bbox": [ + 55, + 265, + 295, + 324 + ], + "type": "text", + "content": " are the " + }, + { + "bbox": [ + 55, + 265, + 295, + 324 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 265, + 295, + 324 + ], + "type": "text", + "content": " most probable indices. The top operator can be considered as a variation of the arg max operator that returns multiple indices, with top equivalent to arg max." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 324, + 296, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 324, + 296, + 420 + ], + "spans": [ + { + "bbox": [ + 55, + 324, + 296, + 420 + ], + "type": "text", + "content": "At time step " + }, + { + "bbox": [ + 55, + 324, + 296, + 420 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 324, + 296, + 420 + ], + "type": "text", + "content": ", the policy " + }, + { + "bbox": [ + 55, + 324, + 296, + 420 + ], + "type": "inline_equation", + "content": "\\pi_{\\theta}(\\mathbf{c}_t \\mid \\mathbf{h}_t, \\mathcal{M})" + }, + { + "bbox": [ + 55, + 324, + 296, + 420 + ], + "type": "text", + "content": " outputs independent categorical distributions over the token vocabulary for each agent. Our method, CAT-K rollout, deterministically rolls out the policy by selecting, at each time step and for each agent, the one action among the top-K likeliest according to " + }, + { + "bbox": [ + 55, + 324, + 296, + 420 + ], + "type": "inline_equation", + "content": "\\pi_{\\theta}" + }, + { + "bbox": [ + 55, + 324, + 296, + 420 + ], + "type": "text", + "content": " that brings the agent closest to the GT next state. Using a distance metric " + }, + { + "bbox": [ + 55, + 324, + 296, + 420 + ], + "type": "inline_equation", + "content": "d(\\cdot, \\cdot)" + }, + { + "bbox": [ + 55, + 324, + 296, + 420 + ], + "type": "text", + "content": " on the states, this is formally expressed as follows:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 116, + 428, + 294, + 451 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 428, + 294, + 451 + ], + "spans": [ + { + "bbox": [ + 116, + 428, + 294, + 451 + ], + "type": "interline_equation", + "content": "c _ {t} ^ {i} = \\underset {c \\in \\left\\{\\xi_ {1}, \\dots , \\xi_ {K} \\right\\}} {\\arg \\min } d \\left(f \\left(s _ {t} ^ {i}, x _ {c}\\right), \\hat {s} _ {t + 1} ^ {i}\\right), \\tag {1}", + "image_path": "8cabd02e474af68545409b23edb6a217844b3c15c2c017a7eec87a7ae340bfc6.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 72, + 453, + 294, + 476 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 453, + 294, + 476 + ], + "spans": [ + { + "bbox": [ + 72, + 453, + 294, + 476 + ], + "type": "interline_equation", + "content": "\\left\\{\\xi_ {1}, \\dots , \\xi_ {K} \\right\\} = \\underset {c _ {t} ^ {i} \\in \\{1, \\dots , | V | \\}} {\\operatorname {t o p} ^ {K}} \\left[ \\pi \\left(c _ {t} ^ {i} \\mid \\mathbf {h} _ {t}, \\mathcal {M}\\right) \\right], \\tag {2}", + "image_path": "3b8cbfc25391f1abe504e4c890b643e43e58dd46ea18e741fb7baea66867f5f7.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 485, + 296, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 485, + 296, + 534 + ], + "spans": [ + { + "bbox": [ + 55, + 485, + 296, + 534 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 485, + 296, + 534 + ], + "type": "inline_equation", + "content": "c_{t}^{i}" + }, + { + "bbox": [ + 55, + 485, + 296, + 534 + ], + "type": "text", + "content": " is the action token indices of agent " + }, + { + "bbox": [ + 55, + 485, + 296, + 534 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 485, + 296, + 534 + ], + "type": "text", + "content": " at step " + }, + { + "bbox": [ + 55, + 485, + 296, + 534 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 485, + 296, + 534 + ], + "type": "text", + "content": " for the CAT-K rollout, and " + }, + { + "bbox": [ + 55, + 485, + 296, + 534 + ], + "type": "inline_equation", + "content": "\\{\\xi_1,\\dots ,\\xi_K\\}" + }, + { + "bbox": [ + 55, + 485, + 296, + 534 + ], + "type": "text", + "content": " are the top-K likeliest token index according to the policy. Given " + }, + { + "bbox": [ + 55, + 485, + 296, + 534 + ], + "type": "inline_equation", + "content": "c_{t}^{i}" + }, + { + "bbox": [ + 55, + 485, + 296, + 534 + ], + "type": "text", + "content": ", the next state is obtained using the vocabulary " + }, + { + "bbox": [ + 55, + 485, + 296, + 534 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 55, + 485, + 296, + 534 + ], + "type": "text", + "content": " and the dynamics as:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 137, + 541, + 294, + 557 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 541, + 294, + 557 + ], + "spans": [ + { + "bbox": [ + 137, + 541, + 294, + 557 + ], + "type": "interline_equation", + "content": "s _ {t + 1} ^ {i} = f \\left(s _ {t} ^ {i}, x _ {c _ {t} ^ {i}}\\right). \\tag {3}", + "image_path": "1daec19ddb5fec502f6b8a1c94944b827f641280404b03b5d83029630f8381a7.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 564, + 296, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 564, + 296, + 613 + ], + "spans": [ + { + "bbox": [ + 55, + 564, + 296, + 613 + ], + "type": "text", + "content": "These rollout states will be used as the input " + }, + { + "bbox": [ + 55, + 564, + 296, + 613 + ], + "type": "inline_equation", + "content": "\\mathbf{h}" + }, + { + "bbox": [ + 55, + 564, + 296, + 613 + ], + "type": "text", + "content": " to the policy at the next time step. By doing this sequentially from " + }, + { + "bbox": [ + 55, + 564, + 296, + 613 + ], + "type": "inline_equation", + "content": "t = 0" + }, + { + "bbox": [ + 55, + 564, + 296, + 613 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 564, + 296, + 613 + ], + "type": "inline_equation", + "content": "t = T - 1" + }, + { + "bbox": [ + 55, + 564, + 296, + 613 + ], + "type": "text", + "content": " and repeating for all " + }, + { + "bbox": [ + 55, + 564, + 296, + 613 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 564, + 296, + 613 + ], + "type": "text", + "content": " agents, we obtain the CAT-K rollout trajectories " + }, + { + "bbox": [ + 55, + 564, + 296, + 613 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_{0:T}" + }, + { + "bbox": [ + 55, + 564, + 296, + 613 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 55, + 620, + 241, + 633 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 620, + 241, + 633 + ], + "spans": [ + { + "bbox": [ + 55, + 620, + 241, + 633 + ], + "type": "text", + "content": "4.2. Closed-loop supervised fine-tuning" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 55, + 637, + 296, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 637, + 296, + 686 + ], + "spans": [ + { + "bbox": [ + 55, + 637, + 296, + 686 + ], + "type": "text", + "content": "Given the CAT-K rollout trajectory " + }, + { + "bbox": [ + 55, + 637, + 296, + 686 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_{0:T}" + }, + { + "bbox": [ + 55, + 637, + 296, + 686 + ], + "type": "text", + "content": ", we can apply the idea of DAD [34] and construct the recovery action indices " + }, + { + "bbox": [ + 55, + 637, + 296, + 686 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{c}}_t" + }, + { + "bbox": [ + 55, + 637, + 296, + 686 + ], + "type": "text", + "content": " from the GT trajectories " + }, + { + "bbox": [ + 55, + 637, + 296, + 686 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{s}}_{0:T}" + }, + { + "bbox": [ + 55, + 637, + 296, + 686 + ], + "type": "text", + "content": " by finding the action token that brings each agent closest to its original trajectory:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 100, + 693, + 294, + 716 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 693, + 294, + 716 + ], + "spans": [ + { + "bbox": [ + 100, + 693, + 294, + 716 + ], + "type": "interline_equation", + "content": "\\hat {c} _ {t} ^ {i} = \\underset {c \\in \\{1, \\dots , | V | \\}} {\\arg \\min } d \\left(f \\left(s _ {t} ^ {i}, x _ {c}\\right), \\hat {s} _ {t + 1} ^ {i}\\right). \\tag {4}", + "image_path": "5f761ba566d8f7eeaa464f3dc871f80ca6a104d9d793e78f8ca36ccc8933c558.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 265, + 554, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 265, + 554, + 289 + ], + "spans": [ + { + "bbox": [ + 313, + 265, + 554, + 289 + ], + "type": "text", + "content": "Given these indices " + }, + { + "bbox": [ + 313, + 265, + 554, + 289 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{c}}_t" + }, + { + "bbox": [ + 313, + 265, + 554, + 289 + ], + "type": "text", + "content": ", the NTP policy is trained using the cross-entropy loss" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 347, + 298, + 554, + 331 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 298, + 554, + 331 + ], + "spans": [ + { + "bbox": [ + 347, + 298, + 554, + 331 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\theta} = - \\frac {1}{N T} \\sum_ {i = 1} ^ {N} \\sum_ {t = 0} ^ {T - 1} \\log \\pi_ {\\theta} \\left(\\hat {c} _ {t} ^ {i} \\mid \\mathbf {h} _ {t}, \\mathcal {M}\\right). \\tag {5}", + "image_path": "cc0c2fe5e10259125cd4751a4b51bde294e5df17c5c1940ea121f7dd59bd2a14.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 339, + 554, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 339, + 554, + 399 + ], + "spans": [ + { + "bbox": [ + 313, + 339, + 554, + 399 + ], + "type": "text", + "content": "Since CAT-K rollout is effective only when the top-K rollouts of the policy cover the GT mode, we adopt a two-stage training procedure, summarized in Algorithm 1. First, we obtain a reasonably well-trained policy through BC pretraining, then fine-tune it using CAT-K rollouts." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 408, + 492, + 421 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 408, + 492, + 421 + ], + "spans": [ + { + "bbox": [ + 313, + 408, + 492, + 421 + ], + "type": "text", + "content": "4.3. Comparison to previous methods" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 426, + 554, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 554, + 461 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 554, + 461 + ], + "type": "text", + "content": "In Fig. 2, we compare our method with previous data augmentation approaches that alleviate the covariate shift for NTP traffic simulation policies." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 462, + 555, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 462, + 555, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 462, + 555, + 665 + ], + "type": "text", + "content": "Top-K sampling is a common approach used by NTP models for generating sequences during inference as it improves the sample quality. However, it is unsuitable for generating trajectories during training as it does not consider the distance to the GT trajectory, and hence the validity of generated recovery actions (c.f. Fig. 2b). This can be partially addressed by post-hoc filtering of trajectories that deviated too far from the GT, but the resulting method is sample-inefficient, as most rollouts will be discarded, and the distance threshold hyperparameter is difficult to tune, since its optimal value varies across scenarios (e.g., high-speed vs. low-speed situations). In contrast, CAT-K rollout (c.f. Fig. 2a) is sample efficient by choosing the closest among top-K actions at each timestep and also removes the need to tune a distance-based hyperparameter. As shown in Tab. 2, our hyperparameter " + }, + { + "bbox": [ + 313, + 462, + 555, + 665 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 462, + 555, + 665 + ], + "type": "text", + "content": " provides strong results across a large range of values and is hence much easier to tune." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "type": "text", + "content": "Instead of sampling top-K trajectories from the policy, current SoTA methods such as Trajeglish [24] and SMART [35] use forms of trajectory noising to address the issue of covariate shift. All trajectory noising approaches rely" + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5425" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 228 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 228 + ], + "type": "text", + "content": "on the injection of small perturbations into the tokenization of the GT trajectory, but implementations can vary from each other in details. For example, in Trajeglish, the likelihood of each noised token is a function of the resulting distance to the ground-truth trajectory, i.e. " + }, + { + "bbox": [ + 55, + 72, + 294, + 228 + ], + "type": "inline_equation", + "content": "q_{t}^{i}(c) \\sim \\exp(-d(f(s_{t}^{i}, x_{c}), \\hat{s}_{t+1}^{i}) / \\tau)" + }, + { + "bbox": [ + 55, + 72, + 294, + 228 + ], + "type": "text", + "content": ", while for SMART, tokens are sampled uniformly from those " + }, + { + "bbox": [ + 55, + 72, + 294, + 228 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 72, + 294, + 228 + ], + "type": "text", + "content": " tokens with the highest likelihood " + }, + { + "bbox": [ + 55, + 72, + 294, + 228 + ], + "type": "inline_equation", + "content": "q_{t}^{i}(c)" + }, + { + "bbox": [ + 55, + 72, + 294, + 228 + ], + "type": "text", + "content": ". Additionally, Trajeglish, similar to our methods, uses the DAD [34] recovery target that finds the token that would bring the agent back to the GT trajectory (c.f. Fig. 2c), while SMART uses the next, also noised, token as target - effectively treating the noise injection as data augmentation on the training trajectories (c.f. Fig. 2d)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 228, + 295, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 228, + 295, + 443 + ], + "spans": [ + { + "bbox": [ + 56, + 228, + 295, + 443 + ], + "type": "text", + "content": "However, Trajeglish's noisy tokenization does not yield a significant improvement (see Fig. 9 in [24]). Similarly, while SMART's trajectory perturbation enhances the performance of a zero-shot policy trained on NuPlan [3] and evaluated on WOSAC, it does not improve performance on WOSAC itself (see Tab. 4 in [35]). We believe this is because the state distributions generated by their sampling strategies are sub-optimal as they completely ignore what state distribution would be induced by the learned policy, hence likely oversampling irrelevant states and undersampling states the learned policy would actually encounter. By incorporating the learned policy into the sampling strategy, CAT-K generates a state distribution more like the policy's, and is hence better able to reduce the covariate shift between training and inference. Although sampling from the policy incurs higher data generation costs, these can be effectively mitigated through the two-phase training strategy with BC pre-training and closed-loop fine-tuning." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 444, + 295, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 444, + 295, + 622 + ], + "spans": [ + { + "bbox": [ + 55, + 444, + 295, + 622 + ], + "type": "text", + "content": "Lastly, note that CAT-K with " + }, + { + "bbox": [ + 55, + 444, + 295, + 622 + ], + "type": "inline_equation", + "content": "K = |V|" + }, + { + "bbox": [ + 55, + 444, + 295, + 622 + ], + "type": "text", + "content": " is equivalent to noise-free BC, since in this case rollouts follow the GT as closely as the available token book allows. On the other hand, for " + }, + { + "bbox": [ + 55, + 444, + 295, + 622 + ], + "type": "inline_equation", + "content": "K = 1" + }, + { + "bbox": [ + 55, + 444, + 295, + 622 + ], + "type": "text", + "content": ", CAT-K is equivalent to deterministically rolling out the policy by always choosing the most likely token. Consequently, for CAT-K, the hyperparameter " + }, + { + "bbox": [ + 55, + 444, + 295, + 622 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 444, + 295, + 622 + ], + "type": "text", + "content": " trades off following the policy (for " + }, + { + "bbox": [ + 55, + 444, + 295, + 622 + ], + "type": "inline_equation", + "content": "K = 1" + }, + { + "bbox": [ + 55, + 444, + 295, + 622 + ], + "type": "text", + "content": ") vs. following the GT (for " + }, + { + "bbox": [ + 55, + 444, + 295, + 622 + ], + "type": "inline_equation", + "content": "K = |V|" + }, + { + "bbox": [ + 55, + 444, + 295, + 622 + ], + "type": "text", + "content": "). In Fig. 5 we show how various choices of " + }, + { + "bbox": [ + 55, + 444, + 295, + 622 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 444, + 295, + 622 + ], + "type": "text", + "content": " impact the Average Displacement Error (ADE) between the CAT-K rollouts and the GT. The intuition behind CAT-K is to find, among all the top-K T-step rollouts (with a total of " + }, + { + "bbox": [ + 55, + 444, + 295, + 622 + ], + "type": "inline_equation", + "content": "K^T" + }, + { + "bbox": [ + 55, + 444, + 295, + 622 + ], + "type": "text", + "content": " possibilities), the one with the minimum ADE (minADE) to the GT. However, since directly solving this optimization problem is difficult, we use CAT-K as a greedy approximation to solve it sequentially." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 633, + 135, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 633, + 135, + 647 + ], + "spans": [ + { + "bbox": [ + 55, + 633, + 135, + 647 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 653, + 294, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 294, + 713 + ], + "type": "text", + "content": "To show the broad relevance of CAT-K fine-tuning, we evaluate its performance on two different tasks using two different types of policy architecture, namely a traffic simulation task using a NTP policy and an ego-motion planning task using a policy parameterized as GMM. We base both tasks" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 553, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 109 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 109 + ], + "type": "text", + "content": "on the widely used WOMD [5] and follow the simulation setup proposed in WOSAC [21], namely providing 1 second of history and generating rollouts of 8 seconds length." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 117, + 422, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 117, + 422, + 129 + ], + "spans": [ + { + "bbox": [ + 314, + 117, + 422, + 129 + ], + "type": "text", + "content": "5.1. Traffic Simulation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 135, + 553, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 135, + 553, + 291 + ], + "spans": [ + { + "bbox": [ + 313, + 135, + 553, + 291 + ], + "type": "text", + "content": "Metrics. In the traffic simulation task, we follow the WOSAC protocol: for each scenario, we generate 32 simulated rollouts for all agents in the scene, at " + }, + { + "bbox": [ + 313, + 135, + 553, + 291 + ], + "type": "inline_equation", + "content": "10\\mathrm{Hz}" + }, + { + "bbox": [ + 313, + 135, + 553, + 291 + ], + "type": "text", + "content": ", and evaluate how well their distribution matches that of the human demonstrations in the data. This distributional matching is evaluated along three dimensions, namely kinematics (e.g., velocities and accelerations), interactions (e.g., collisions), and map alignment (e.g., off-road). All three are summarized as weighted averages in the \"Realism Meta-Metric\" (RMM), the key performance indicator of the WOSAC leaderboard. For more details please see [21]. We also report the minADE, which is not used in the RMM, but widely applied for motion prediction and policy evaluation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 291, + 553, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 291, + 553, + 662 + ], + "spans": [ + { + "bbox": [ + 313, + 291, + 553, + 662 + ], + "type": "text", + "content": "Policy. We use SMART [35] as our policy architecture due to its strong performance and open-source implementation. Specifically, we use the SMART-tiny (7M) model and its open-sourced token vocabularies. The map polyline token vocabulary has a size of 1024, with each token representing a polyline consisting of 10 segments, each 0.5 meters long. The agent trajectory token vocabulary has a size of 2048, with each token representing a 0.5-second trajectory at " + }, + { + "bbox": [ + 313, + 291, + 553, + 662 + ], + "type": "inline_equation", + "content": "10\\mathrm{Hz}" + }, + { + "bbox": [ + 313, + 291, + 553, + 662 + ], + "type": "text", + "content": ". SMART re-plans at " + }, + { + "bbox": [ + 313, + 291, + 553, + 662 + ], + "type": "inline_equation", + "content": "2\\mathrm{Hz}" + }, + { + "bbox": [ + 313, + 291, + 553, + 662 + ], + "type": "text", + "content": ", hence the final output is at " + }, + { + "bbox": [ + 313, + 291, + 553, + 662 + ], + "type": "inline_equation", + "content": "10\\mathrm{Hz}" + }, + { + "bbox": [ + 313, + 291, + 553, + 662 + ], + "type": "text", + "content": " when submitted to the WOSAC leaderboard. However, our experiments show that SMART's open-sourced agent token vocabulary results in lower kinematic metrics. To address this issue, we reran the K-disk clustering [24] with a larger number of trajectories. By using our own agent token vocabulary, the performance on the WOSAC leaderboard is significantly improved (c.f. Tab. 1). Please note that all other experiments in this paper (i.e., except those in Tab. 1) still use SMART's open-sourced token vocabularies. SMART uses Transformers [33] with query-centric representations [41, 46], and the distance metric " + }, + { + "bbox": [ + 313, + 291, + 553, + 662 + ], + "type": "inline_equation", + "content": "d(\\cdot, \\cdot)" + }, + { + "bbox": [ + 313, + 291, + 553, + 662 + ], + "type": "text", + "content": " is the average Euclidean distance between the four pairs of corners of two bounding boxes. We disable SMART's trajectory perturbation as it deteriorates performance on the WOSAC leaderboard (see [35] Tab. 4). We run BC pretraining for 32 epochs, each taking 1.7 hour, then continue with closed-loop supervised fine-tuning with CAT-32 for 10 epochs, each taking 2.6 hours. During fine-tuning, the map encoder is frozen to save GPU memory, allowing for a larger training batch size with negligible impact on performance. For both pre-training and fine-tuning, we use " + }, + { + "bbox": [ + 313, + 291, + 553, + 662 + ], + "type": "inline_equation", + "content": "8 \\times" + }, + { + "bbox": [ + 313, + 291, + 553, + 662 + ], + "type": "text", + "content": " A100 80GB GPUs with a total batch size of 80." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 671, + 436, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 671, + 436, + 685 + ], + "spans": [ + { + "bbox": [ + 314, + 671, + 436, + 685 + ], + "type": "text", + "content": "5.2. Ego-motion planning" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "content": "Metrics. Unlike traffic simulation, in ego-motion planning only the ego-vehicle is controlled and evaluated. Moreover," + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5426" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 61, + 70, + 548, + 182 + ], + "blocks": [ + { + "bbox": [ + 61, + 70, + 548, + 182 + ], + "lines": [ + { + "bbox": [ + 61, + 70, + 548, + 182 + ], + "spans": [ + { + "bbox": [ + 61, + 70, + 548, + 182 + ], + "type": "table", + "html": "
Leaderboard, test split\nMethod# model\nparamsRMM\n↑RMM diff. to\nSMART-largeKinematic\nmetrics ↑Interactive\nmetrics ↑Map-based\nmetrics ↑min\nADE ↓
SMART-tiny fine-tuned w. CAT-K (ours)7 M0.7702+0.00880.49310.81190.87491.3068
SMART-large [35]102 M0.7614+0.00000.47860.80660.86481.3728
KiGRAS [43]0.7 M0.7597-0.00140.46910.80640.86581.4384
SMART-tiny [35]7 M0.7591-0.00230.47590.80390.86321.4062
FDriver-tiny7 M0.7584-0.00300.46140.80690.86581.4475
SMART [35]8 M0.7511-0.01030.44450.80500.85711.5447
BehaviorGPT [47]3 M0.7473-0.01410.43330.79970.85931.4147
GUMP [9]523 M0.7431-0.01830.47800.78870.83591.6041
", + "image_path": "ab9e642723d8d2a53cbd6479d2c14307af227d441278716e9cbfd99a598a07b4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 184, + 553, + 208 + ], + "lines": [ + { + "bbox": [ + 55, + 184, + 553, + 208 + ], + "spans": [ + { + "bbox": [ + 55, + 184, + 553, + 208 + ], + "type": "text", + "content": "Table 1. Results on the WOSAC 2024 leaderboard [17]. RMM stands for Realism Meta Metric, the key metric used for ranking. Note that on the public leaderboard [17] our method appears under the name \"SMART-tiny-CLSFT\" (Closed-Loop Supervised Fine-Tuning)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 59, + 216, + 290, + 300 + ], + "blocks": [ + { + "bbox": [ + 59, + 216, + 290, + 300 + ], + "lines": [ + { + "bbox": [ + 59, + 216, + 290, + 300 + ], + "spans": [ + { + "bbox": [ + 59, + 216, + 290, + 300 + ], + "type": "image", + "image_path": "558b1ca35205381da0e84a506a8d3b693a40841a3649ddb6830ab7615582e586.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 59, + 301, + 290, + 313 + ], + "lines": [ + { + "bbox": [ + 59, + 301, + 290, + 313 + ], + "spans": [ + { + "bbox": [ + 59, + 301, + 290, + 313 + ], + "type": "text", + "content": "Figure 3. Influence of " + }, + { + "bbox": [ + 59, + 301, + 290, + 313 + ], + "type": "inline_equation", + "content": "K_{\\mathrm{infer}}" + }, + { + "bbox": [ + 59, + 301, + 290, + 313 + ], + "type": "text", + "content": " for inference-time top-K sampling." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 328, + 295, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 328, + 295, + 426 + ], + "spans": [ + { + "bbox": [ + 54, + 328, + 295, + 426 + ], + "type": "text", + "content": "the optimal policy is not the one that perfectly matches the GT trajectory distribution, but which minimizes planning metrics, such as collision rate, off-road rate, and ADE. For completeness, we also report RMM, applied only to ego trajectories, and minADE over 32 sampled rollouts. With the exception of minADE, and in contrast to the traffic simulation task, the rollouts for evaluation are sampled deterministically, by choosing the most-likely mode at each timestep." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 426, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 426, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 426, + 295, + 713 + ], + "type": "text", + "content": "Policy. To show that our fine-tuning approach can be applied to a more general class of policies beyond NTP, we parameterized the output of the ego-policy as GMM. Specifically, we replace the final layer of the SMART network with two heads, a classification head that predicts the mixture densities and a regression head that predicts the means of the Gaussian distributions with fixed standard deviations (similar to, e.g., [14]). Specifically, the GMM predicts 16 modes, each representing a 3-dimensional Gaussian distribution over the changes in x- and y-position and yaw heading. The final model comprises a total of 3.2M parameters. In contrast to the SMART traffic simulation policy that uses cross-entropy loss on all agents, the GMM-based egopolicy is trained using negative log-likelihood loss and only predicts actions for the ego-vehicle. When applying CAT-K rollout or top-K sampling to the GMM, we select the K most-likely modes of the GMM, and within each mode deterministically use the mean of the Gaussian. Please note that ego-motion planning is a separate experiment from traffic simulation; hence, the ego-policy is not fine-tuned on the traffic simulation policy. Also note that the ego-policy operates in a mid-to-end manner, taking tracking results and HD maps as inputs, rather than in a fully end-to-end fashion that maps raw sensor observations to actions." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 214, + 372, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 214, + 372, + 225 + ], + "spans": [ + { + "bbox": [ + 313, + 214, + 372, + 225 + ], + "type": "text", + "content": "5.3. Results" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 232, + 525, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 232, + 525, + 243 + ], + "spans": [ + { + "bbox": [ + 313, + 232, + 525, + 243 + ], + "type": "text", + "content": "5.3.1. WOSAC leaderboard for traffic simulation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 247, + 553, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 247, + 553, + 354 + ], + "spans": [ + { + "bbox": [ + 313, + 247, + 553, + 354 + ], + "type": "text", + "content": "In Tab. 1 we compare our approach with other traffic simulation policies on the WOSAC leaderboard. Notably, all the top-ranking methods on the leaderboard are NTP policies trained via BC. Our SMART-tiny model using CAT-K fine-tuning outperforms the previous SoTA, SMART-large with 102M parameters, by a significant margin of " + }, + { + "bbox": [ + 313, + 247, + 553, + 354 + ], + "type": "inline_equation", + "content": "+0.0088" + }, + { + "bbox": [ + 313, + 247, + 553, + 354 + ], + "type": "text", + "content": "; and improves on SMART-tiny by " + }, + { + "bbox": [ + 313, + 247, + 553, + 354 + ], + "type": "inline_equation", + "content": "+0.0111" + }, + { + "bbox": [ + 313, + 247, + 553, + 354 + ], + "type": "text", + "content": ". As the first to perform closed-loop fine-tuning on the leaderboard, our approach improves all metrics and sets a new SoTA." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 355, + 554, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 355, + 554, + 498 + ], + "spans": [ + { + "bbox": [ + 313, + 355, + 554, + 498 + ], + "type": "text", + "content": "For fine-tuning, we chose CAT-32 after preliminary hyperparameter explorations, a choice that was later confirmed to perform well in our ablation studies (see Tab. 2), though CAT-K improved performance for a wide range of values of " + }, + { + "bbox": [ + 313, + 355, + 554, + 498 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 355, + 554, + 498 + ], + "type": "text", + "content": ". We also found that choosing a sufficiently high " + }, + { + "bbox": [ + 313, + 355, + 554, + 498 + ], + "type": "inline_equation", + "content": "K_{\\mathrm{infer}}" + }, + { + "bbox": [ + 313, + 355, + 554, + 498 + ], + "type": "text", + "content": " for top-K sampling during inference is important for a high RMM [9], and we chose " + }, + { + "bbox": [ + 313, + 355, + 554, + 498 + ], + "type": "inline_equation", + "content": "K_{\\mathrm{infer}} = 48" + }, + { + "bbox": [ + 313, + 355, + 554, + 498 + ], + "type": "text", + "content": " for our leaderboard submission based on local validation results (see Fig. 3). In all of our experiments, we fixed the inference time sampling temperature to 1.0 and did not use top-p (nucleus) sampling [8], though we expect that tuning these hyperparameters can lead to further performance gains." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 498, + 554, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 554, + 677 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 554, + 677 + ], + "type": "text", + "content": "To ensure that the observed performance gains with CAT-K are not due to improved hyperparameter tuning, we also conduct a large-scale hyperparameter grid search for the baseline method SMART-tiny. This includes training for more epochs, adjusting the learning rate and learning rate scheduler, experimenting with various data augmentation and data pre-processing strategies, and using various values of " + }, + { + "bbox": [ + 313, + 498, + 554, + 677 + ], + "type": "inline_equation", + "content": "K_{\\mathrm{infer}}" + }, + { + "bbox": [ + 313, + 498, + 554, + 677 + ], + "type": "text", + "content": " for the inference-time top-K sampling. This allowed us to push the performance of the BC baseline to 0.7671 RMM on the test split (not shown in Tab. 1), but still falling significantly short of the performance of our CAT-K fine-tuning method (0.7702 RMM). Note that even comparing against this improved baseline, our method improves the performance more (+0.0031) than scaling up the model size by a factor of 14 to 102M parameters (+0.0023)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 677, + 553, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 553, + 712 + ], + "type": "text", + "content": "Qualitatively, our method can generate diverse and realistic behavior over long periods of time. As shown in Fig. 6, our method can handle the subtle interactions between traf" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5427" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 59, + 70, + 550, + 299 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 550, + 299 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 550, + 299 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 550, + 299 + ], + "type": "table", + "html": "
Local val. split\nMethodCriterion\nof topKK for topKSampled fromNext targetRMM\n↑Kinematic metrics ↑Interactive metrics ↑Map-based metrics ↑min ADE ↓
BC pre-training---GT0.75810.45120.80760.86971.3152
BC fine-tuning---GT0.75900.45140.80960.87001.3039
Trajeglish's noisy\ntokenizationneg. dist.5†neg. dist.GT0.75620.44690.80740.86731.3459
neg. dist.5†uniformGT0.75540.44670.80690.86551.3404
neg. dist.32neg. dist.GT0.74010.41740.79850.84931.6669
SMART's trajectory\nperturbationneg. dist.5†uniformRO0.75560.44400.80820.86611.3177
neg. dist.5†neg. dist.RO0.75600.44690.80690.86731.3514
neg. dist.32uniformRO0.73140.41580.79490.83001.5380
Top-5prob5probGT0.64780.33130.68470.75281.8802
Top-5 + distance filterprob5probGT0.68600.33560.74660.80831.7627
Top-5 + distance based samplingprob5neg. dist.GT0.70580.35360.75790.84001.5848
Deterministic rollout--max-probGT0.63610.32910.68450.74921.8695
CAT-5prob-closestGT0.74230.42510.79170.86011.4677
CAT-16prob-closestGT0.76040.45920.80820.87091.3372
CAT-32 (leaderboard)prob-closestGT0.76160.45830.81050.87201.3105
CAT-40prob-closestGT0.76170.45670.81010.87381.2998
CAT-64prob-closestGT0.76020.45520.80980.87071.3028
", + "image_path": "a2225f2b46f1e60bb4a79875848802bb2a087bb65384536fcdd262a80d741e3c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 300, + 555, + 344 + ], + "lines": [ + { + "bbox": [ + 55, + 300, + 555, + 344 + ], + "spans": [ + { + "bbox": [ + 55, + 300, + 555, + 344 + ], + "type": "text", + "content": "Table 2. Ablation study on WOSAC " + }, + { + "bbox": [ + 55, + 300, + 555, + 344 + ], + "type": "inline_equation", + "content": "2\\%" + }, + { + "bbox": [ + 55, + 300, + 555, + 344 + ], + "type": "text", + "content": " validation split. We compare different ways to fine-tune the same base mode (BC pre-training). \"Sampled from\" indicates how the action is sampled during fine-tuning, either based on the distance to the GT (\"neg. dist\", \"uniform\", \"closest\") or based on the model outputs (\"prob\", \"max-prob\"). Here dist. is the abbreviation of distance. RO stands for rollout, i.e., the next target action is computed based on the rollout, not the GT state. " + }, + { + "bbox": [ + 55, + 300, + 555, + 344 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 55, + 300, + 555, + 344 + ], + "type": "text", + "content": " indicates original hyperparameter choices of baseline algorithms." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 59, + 348, + 290, + 432 + ], + "blocks": [ + { + "bbox": [ + 59, + 348, + 290, + 432 + ], + "lines": [ + { + "bbox": [ + 59, + 348, + 290, + 432 + ], + "spans": [ + { + "bbox": [ + 59, + 348, + 290, + 432 + ], + "type": "image", + "image_path": "c7f34f8b1aaf5c7e0377b71e667b22c87ce2f03e92b5c9d311fa88c28deaa85f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 72, + 432, + 278, + 443 + ], + "lines": [ + { + "bbox": [ + 72, + 432, + 278, + 443 + ], + "spans": [ + { + "bbox": [ + 72, + 432, + 278, + 443 + ], + "type": "text", + "content": "Figure 4. On server vs. local evaluation of SMART-tiny." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 453, + 295, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 453, + 295, + 515 + ], + "spans": [ + { + "bbox": [ + 55, + 453, + 295, + 515 + ], + "type": "text", + "content": "fic participants in a dense parking lot, which is arguably a more challenging scenario than intersections and highways for traffic simulation. Moreover, the behavior remains realistic at the end of the required simulation time of 8 seconds. Additional examples are in the supplementary videos." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 519, + 206, + 530 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 519, + 206, + 530 + ], + "spans": [ + { + "bbox": [ + 55, + 519, + 206, + 530 + ], + "type": "text", + "content": "5.3.2. Ablation studies on WOSAC" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 533, + 295, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 533, + 295, + 665 + ], + "spans": [ + { + "bbox": [ + 55, + 533, + 295, + 665 + ], + "type": "text", + "content": "In Tab. 2 we provide a thorough ablation study. Due to the high cost of evaluation we use " + }, + { + "bbox": [ + 55, + 533, + 295, + 665 + ], + "type": "inline_equation", + "content": "2\\%" + }, + { + "bbox": [ + 55, + 533, + 295, + 665 + ], + "type": "text", + "content": " of the validation split (880 out of 44097 scenarios). To verify the fidelity of this evaluation setting, we compare results with those on the full validation set in Fig. 4. We observe consistent differences indicating that our evaluation setting is reasonable. We begin with a SMART-tiny model trained with BC for 32 epochs (BC pre-training, row 1). All other models fine-tune the BC pre-training model for 5 epochs. During inference, we use top-K sampling with " + }, + { + "bbox": [ + 55, + 533, + 295, + 665 + ], + "type": "inline_equation", + "content": "K_{\\mathrm{infer}} = 40" + }, + { + "bbox": [ + 55, + 533, + 295, + 665 + ], + "type": "text", + "content": " for all methods (the best " + }, + { + "bbox": [ + 55, + 533, + 295, + 665 + ], + "type": "inline_equation", + "content": "K_{\\mathrm{infer}}" + }, + { + "bbox": [ + 55, + 533, + 295, + 665 + ], + "type": "text", + "content": " for SMART-tiny according to Fig. 3)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "In Tab. 2, CAT-K fine-tuning is the only method that significantly outperforms the BC pre-training model. Further fine-tuning with BC (row 2), using Trajeglish's noisy tokenization (rows 3-5), or SMART's trajectory perturbation" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 326, + 348, + 540, + 432 + ], + "blocks": [ + { + "bbox": [ + 326, + 348, + 540, + 432 + ], + "lines": [ + { + "bbox": [ + 326, + 348, + 540, + 432 + ], + "spans": [ + { + "bbox": [ + 326, + 348, + 540, + 432 + ], + "type": "image", + "image_path": "7fd159d160ec228322ae3be9c8fa4dd8a4f9f2866f069ef960c13031264f4f4b.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 324, + 432, + 544, + 443 + ], + "lines": [ + { + "bbox": [ + 324, + 432, + 544, + 443 + ], + "spans": [ + { + "bbox": [ + 324, + 432, + 544, + 443 + ], + "type": "text", + "content": "Figure 5. ADE between CAT-K rollouts and GT trajectories." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 460, + 555, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 460, + 555, + 604 + ], + "spans": [ + { + "bbox": [ + 313, + 460, + 555, + 604 + ], + "type": "text", + "content": "rows 6-8) remain on par with the original model or even reduce its performance. Closed-loop fine-tuning with top-K rollouts instead of CAT-K rollouts during training significantly reduces performance (rows 9-12), even when rollouts close to the GT are selected for training by either filtering them based on the distance to the GT or sampling among them from a distance-dependent distribution. The observation that data augmentation and fine-tuning with top-K sampling cannot improve the RMM is consistent with prior works [19, 23, 24, 35] and the WOSAC leaderboard. Ablations on different values of " + }, + { + "bbox": [ + 313, + 460, + 555, + 604 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 460, + 555, + 604 + ], + "type": "text", + "content": " for top-K sampling and other hyperparameters are provided in the appendix." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "content": "Next, we ablate the value of " + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "content": " used for CAT-K rollout during fine-tuning (rows 13-17). Results indicate that the performance improvement is robust to the choice of " + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "content": " after a reasonable minimum value. As discussed in Sec. 4.3, the hyperparameter " + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "content": " in CAT-K determines how closely the policy follows the GT trajectory, in a way that is more robust than a distance based threshold, for which the optimal value varies strongly based on the situation (vehicle speed, proximity to other cars, etc.). To give more insight into how" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5428" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 70, + 154, + 224 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 154, + 224 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 154, + 224 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 154, + 224 + ], + "type": "image", + "image_path": "142ca603bb68ea37ecd251c565617b76663bddccdb672d82fc9c708d6966a61b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 226, + 555, + 293 + ], + "lines": [ + { + "bbox": [ + 55, + 226, + 555, + 293 + ], + "spans": [ + { + "bbox": [ + 55, + 226, + 555, + 293 + ], + "type": "text", + "content": "Figure 6. Simulation results on WOSAC. Our fine-tuned policy generates interesting and diverse behaviors rarely seen in prior works. Each row represents a different rollout of our model in the same scene. The transparent boxes show the GT agents in the dataset, while the solid boxes show the agents generated by our model. We highlight the agents within the red, green, and orange rectangles, across time steps and rollouts. The red rectangle shows different interactive negotiations emerging between a pedestrian and two vehicles. The green rectangle shows an initially parked vehicle, that leaves (row 1) or stays parked (row 2, 3). The orange rectangle shows a vehicle waiting (row 1) in front of a speed bump (visualized as two light-blue lines), proceeding (row 2), or entering a parking space (row 3)." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 156, + 70, + 253, + 224 + ], + "blocks": [ + { + "bbox": [ + 156, + 70, + 253, + 224 + ], + "lines": [ + { + "bbox": [ + 156, + 70, + 253, + 224 + ], + "spans": [ + { + "bbox": [ + 156, + 70, + 253, + 224 + ], + "type": "image", + "image_path": "03f8c0e01b5cecfa9e9deaff410f72e69e919f5baffe32152c5736896e22acca.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 255, + 70, + 352, + 224 + ], + "blocks": [ + { + "bbox": [ + 255, + 70, + 352, + 224 + ], + "lines": [ + { + "bbox": [ + 255, + 70, + 352, + 224 + ], + "spans": [ + { + "bbox": [ + 255, + 70, + 352, + 224 + ], + "type": "image", + "image_path": "f523e846d9100a9e7d1bd3460223686115386bfc54f2d6ce803dbdd846d616bb.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 354, + 70, + 451, + 224 + ], + "blocks": [ + { + "bbox": [ + 354, + 70, + 451, + 224 + ], + "lines": [ + { + "bbox": [ + 354, + 70, + 451, + 224 + ], + "spans": [ + { + "bbox": [ + 354, + 70, + 451, + 224 + ], + "type": "image", + "image_path": "f1253b0be5349f0499b6eb5fbdb964683e7d2932e2886746342da5b8fc82b0d0.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 453, + 70, + 550, + 224 + ], + "blocks": [ + { + "bbox": [ + 453, + 70, + 550, + 224 + ], + "lines": [ + { + "bbox": [ + 453, + 70, + 550, + 224 + ], + "spans": [ + { + "bbox": [ + 453, + 70, + 550, + 224 + ], + "type": "image", + "image_path": "a21d42867d7ea82036d4a5422e8dac4d3690d87e9db4ae3e6938a448cb2b2548.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 111, + 298, + 499, + 365 + ], + "blocks": [ + { + "bbox": [ + 111, + 298, + 499, + 365 + ], + "lines": [ + { + "bbox": [ + 111, + 298, + 499, + 365 + ], + "spans": [ + { + "bbox": [ + 111, + 298, + 499, + 365 + ], + "type": "table", + "html": "
Method (Local val. split)Collision rate ↓Off-road rate ↓RMM ↑ADE ↓minADE32 ↓
BC pre-training0.05680.00530.81081.36231.3537
BC fine-tuning0.05990.00580.81051.35201.3509
Deterministic rollout0.04330.01380.80811.17990.7962
CAT-30.04220.00350.81691.30960.6912
", + "image_path": "faad84e4f890a2fae3ddaa70b8d0eebb60fafe7a60d5c55a09cd0da51e2dc003.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 366, + 555, + 411 + ], + "lines": [ + { + "bbox": [ + 55, + 366, + 555, + 411 + ], + "spans": [ + { + "bbox": [ + 55, + 366, + 555, + 411 + ], + "type": "text", + "content": "Table 3. Performance of ego policies on WOSAC with local evaluation on " + }, + { + "bbox": [ + 55, + 366, + 555, + 411 + ], + "type": "inline_equation", + "content": "2\\%" + }, + { + "bbox": [ + 55, + 366, + 555, + 411 + ], + "type": "text", + "content": " validation split. All models are fine-tuned for 5 epochs based on the BC pre-training model, which is trained for 32 epochs. We use deterministic rollout during inference and compute all metrics, except for the " + }, + { + "bbox": [ + 55, + 366, + 555, + 411 + ], + "type": "inline_equation", + "content": "\\min\\mathrm{ADE}^{32}" + }, + { + "bbox": [ + 55, + 366, + 555, + 411 + ], + "type": "text", + "content": ". For " + }, + { + "bbox": [ + 55, + 366, + 555, + 411 + ], + "type": "inline_equation", + "content": "\\min\\mathrm{ADE}^{32}" + }, + { + "bbox": [ + 55, + 366, + 555, + 411 + ], + "type": "text", + "content": ", we generate 32 rollouts by using top-3 sampling with a temperature of 1.0 to first sample the categorical distribution over the mixtures, then selecting the mean of the sampled Gaussian mixture." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 427, + 295, + 524 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 427, + 295, + 524 + ], + "spans": [ + { + "bbox": [ + 55, + 427, + 295, + 524 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 427, + 295, + 524 + ], + "type": "text", + "content": " impacts rollouts, in Fig. 5, we inspect the average ADE between rollouts and GT trajectories over training epochs, for different " + }, + { + "bbox": [ + 55, + 427, + 295, + 524 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 427, + 295, + 524 + ], + "type": "text", + "content": " values. As expected, as " + }, + { + "bbox": [ + 55, + 427, + 295, + 524 + ], + "type": "inline_equation", + "content": "K \\to |V|" + }, + { + "bbox": [ + 55, + 427, + 295, + 524 + ], + "type": "text", + "content": ", the ADE decreases towards the level of quantization error, induced by tokenization with a finite vocabulary size (dashed line). With more fine-tuning epochs the average ADE slightly reduces, highlighting how CAT-K fine-tuning improves the policy to follow all behavior modes more closely." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 530, + 244, + 544 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 530, + 244, + 544 + ], + "spans": [ + { + "bbox": [ + 55, + 530, + 244, + 544 + ], + "type": "text", + "content": "5.3.3. Fine-tuning a GMM-based ego-policy" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "text", + "content": "Besides NTP traffic simulation polices, CAT-K fine-tuning also improves the performance of a GMM-based ego-policy with continuous action space. In Tab. 3 we compare our approach with fine-tuning using deterministic rollouts, as well as with continued BC. Our CAT-3 fine-tuning improves all metrics except for ADE, where deterministic rollout performs better. This is expected, as deterministic rollout aligns all modes towards the GT, resulting in mode averaging. While this reduces ADE, it negatively impacts other metrics. Additionally, CAT-K fine-tuning helps overcome the limitation of GMM trained with negative log-likelihood loss, which struggles to capture multimodality as effectively as NTP policies trained with cross-entropy loss. This is supported by the significant drop in minADE after fine-tuning." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 426, + 388, + 439 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 426, + 388, + 439 + ], + "spans": [ + { + "bbox": [ + 314, + 426, + 388, + 439 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 446, + 555, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 446, + 555, + 553 + ], + "spans": [ + { + "bbox": [ + 313, + 446, + 555, + 553 + ], + "type": "text", + "content": "In this paper, we introduce CAT-K rollouts, a closed-loop supervised fine-tuning technique for IL problems with highly multimodal demonstrations, such as traffic simulation. The CAT-K rollout approximately finds the rollout closest to the GT among likely rollouts of a policy, ensuring adherence to the policy while maintaining GT as a reliable reference for supervised learning. As the first method using closed-loop fine-tuning, it achieves the top spot on the WOSAC leaderboard." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 554, + 556, + 710 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 554, + 556, + 710 + ], + "spans": [ + { + "bbox": [ + 313, + 554, + 556, + 710 + ], + "type": "text", + "content": "In the future, we aim to incorporate modern sampling techniques into CAT-K rollouts, such as top-p sampling. Furthermore, instead of always selecting the closest token, we could loosen the \"winner-takes-all\" approach and sample from a group of nearby tokens. We also plan to explore CAT-K fine-tuning for a broader range of policy classes, such as variational autoencoders and diffusion models. Additionally, we want to apply it to other multimodal IL tasks, including end-to-end driving, motion generation for animation, and robot navigation and manipulation. Importantly, our results show that closed-loop supervised fine-tuning is a promising area of future research for policies trained in open-loop, such as the widely used NTP policies." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5429" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 133 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 133 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 133 + ], + "type": "text", + "content": "[1] Mayank Bansal, Alex Krizhevsky, and Abhijit Ogale. Chauffeurnet: Learning to drive by imitating the best and synthesizing the worst. arXiv preprint arXiv:1812.03079, 2018. 1, 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 137, + 296, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 137, + 296, + 266 + ], + "spans": [ + { + "bbox": [ + 61, + 137, + 296, + 266 + ], + "type": "text", + "content": "[2] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel Ziegler, Jeffrey Wu, Clemens Winter, Chris Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. Language models are few-shot learners. In Advances in Neural Information Processing Systems (NeurIPS), pages 1877-1901. Curran Associates, Inc., 2020. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 269, + 295, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 269, + 295, + 334 + ], + "spans": [ + { + "bbox": [ + 62, + 269, + 295, + 334 + ], + "type": "text", + "content": "[3] Holger Caesar, Juraj Kabzan, Kok Seang Tan, Whye Kit Fong, Eric M. Wolff, Alex H. Lang, Luke Fletcher, Oscar Beijbom, and Sammy Omari. NuPlan: A closed-loop ml-based planning benchmark for autonomous vehicles. In Conference on Computer Vision and Pattern Recognition (CVPR) ADP3 Workshop, 2021. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 335, + 295, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 335, + 295, + 388 + ], + "spans": [ + { + "bbox": [ + 62, + 335, + 295, + 388 + ], + "type": "text", + "content": "[4] Yulong Cao, Boris Ivanovic, Chaowei Xiao, and Marco Pavone. Reinforcement learning with human feedback for realistic traffic simulation. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 14428-14434. IEEE, 2024. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 391, + 295, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 391, + 295, + 466 + ], + "spans": [ + { + "bbox": [ + 62, + 391, + 295, + 466 + ], + "type": "text", + "content": "[5] Scott Ettinger, Shuyang Cheng, Benjamin Caine, Chenxi Liu, Hang Zhao, Sabeek Pradhan, Yuning Chai, Ben Sapp, Charles R Qi, Yin Zhou, et al. Large scale interactive motion forecasting for autonomous driving: The waymo open motion dataset. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 9710-9719, 2021. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 468, + 295, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 468, + 295, + 523 + ], + "spans": [ + { + "bbox": [ + 62, + 468, + 295, + 523 + ], + "type": "text", + "content": "[6] Roger Girgis, Florian Golemo, Felipe Codevilla, Martin Weiss, Jim Aldon D'Souza, Samira Ebrahimi Kahou, Felix Heide, and Christopher Pal. Latent variable sequential set transformers for joint multi-agent motion prediction. arXiv preprint arXiv:2104.00563, 2021. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 525, + 295, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 525, + 295, + 555 + ], + "spans": [ + { + "bbox": [ + 62, + 525, + 295, + 555 + ], + "type": "text", + "content": "[7] Jonathan Ho and Stefano Ermon. Generative adversarial imitation learning. Advances in Neural Information Processing Systems (NeurIPS), 29, 2016. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 558, + 295, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 558, + 295, + 601 + ], + "spans": [ + { + "bbox": [ + 62, + 558, + 295, + 601 + ], + "type": "text", + "content": "[8] Ari Holtzman, Jan Buys, Li Du, Maxwell Forbes, and Yejin Choi. The curious case of neural text degeneration. In International Conference on Learning Representations (ICLR), 2020. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 603, + 295, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 603, + 295, + 656 + ], + "spans": [ + { + "bbox": [ + 62, + 603, + 295, + 656 + ], + "type": "text", + "content": "[9] Yihan Hu, Siqi Chai, Zhening Yang, Jingyu Qian, Kun Li, Wenxin Shao, Haichao Zhang, Wei Xu, and Qiang Liu. Solving motion planning tasks with a scalable generative model. In European Conference on Computer Vision (ECCV), 2025. 1, 2, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 658, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 658, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 658, + 295, + 713 + ], + "type": "text", + "content": "[10] Zhiyu Huang, Xinshuo Weng, Maximilian Igl, Yuxiao Chen, Yulong Cao, Boris Ivanovic, Marco Pavone, and Chen Lv. Gen-drive: Enhancing diffusion generative driving policies with reward modeling and reinforcement learning finetuning. arXiv preprint arXiv:2410.05582, 2024. 2" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "text", + "content": "[11] Zhiyu Huang, Zixu Zhang, Ameya Vaidya, Yuxiao Chen, Chen Lv, and Jaime Fernandez Fisac. Versatile scene-consistent traffic scenario generation as optimization with diffusion. arXiv preprint arXiv:2404.02524, 2024. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 317, + 119, + 553, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 119, + 553, + 194 + ], + "spans": [ + { + "bbox": [ + 317, + 119, + 553, + 194 + ], + "type": "text", + "content": "[12] Maximilian Igl, Daewoo Kim, Alex Kuefler, Paul Mougin, Punit Shah, Kyriacos Shiarlis, Dragomir Anguelov, Mark Palatucci, Brandyn White, and Shimon Whiteson. Symphony: Learning realistic and diverse agents for autonomous driving simulation. In 2022 International Conference on Robotics and Automation (ICRA), pages 2445-2451. IEEE, 2022. 2, 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 317, + 197, + 553, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 197, + 553, + 261 + ], + "spans": [ + { + "bbox": [ + 317, + 197, + 553, + 261 + ], + "type": "text", + "content": "[13] Maximilian Igl, Punit Shah, Paul Mougin, Sirish Srinivasan, Tarun Gupta, Brandyn White, Kyriacos Shiarlis, and Shimon Whiteson. Hierarchical imitation learning for stochastic environments. In 2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 1697-1704. IEEE, 2023. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 264, + 553, + 318 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 264, + 553, + 318 + ], + "spans": [ + { + "bbox": [ + 317, + 264, + 553, + 318 + ], + "type": "text", + "content": "[14] Boris Ivanovic and Marco Pavone. The trajectory: Probabilistic multi-agent trajectory modeling with dynamic spatiotemporal graphs. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 2375-2384, 2019. 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 321, + 553, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 321, + 553, + 385 + ], + "spans": [ + { + "bbox": [ + 317, + 321, + 553, + 385 + ], + "type": "text", + "content": "[15] Chiyu Jiang, Andre Cornman, Cheolho Park, Benjamin Sapp, Yin Zhou, Dragomir Anguelov, et al. Motiondiffuser: Controllable multi-agent motion prediction using diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9644-9653, 2023. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 388, + 553, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 388, + 553, + 430 + ], + "spans": [ + { + "bbox": [ + 317, + 388, + 553, + 430 + ], + "type": "text", + "content": "[16] Longzhong Lin, Xuewu Lin, Kechun Xu, Haojian Lu, Lichao Huang, Rong Xiong, and Yue Wang. Revisit mixture models for multi-agent simulation: Experimental study within a unified framework, 2025. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 434, + 553, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 434, + 553, + 475 + ], + "spans": [ + { + "bbox": [ + 317, + 434, + 553, + 475 + ], + "type": "text", + "content": "[17] Waymo LLC. Waymo open sim agent challenge (wosac) 2024 leaderboard. https://waymo.com/open/challenges/2024/sim-agents/, 2024. Accessed: 2025-03-14. 6, 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 479, + 553, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 479, + 553, + 532 + ], + "spans": [ + { + "bbox": [ + 317, + 479, + 553, + 532 + ], + "type": "text", + "content": "[18] Jack Lu, Kelvin Wong, Chris Zhang, Simon Suo, and Raquel Urtasun. Scenecontrol: Diffusion for controllable traffic scene generation. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 16908-16914. IEEE, 2024. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 535, + 553, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 535, + 553, + 611 + ], + "spans": [ + { + "bbox": [ + 317, + 535, + 553, + 611 + ], + "type": "text", + "content": "[19] Yiren Lu, Justin Fu, George Tucker, Xinlei Pan, Eli Bronstein, Rebecca Roelofs, Benjamin Sapp, Brandyn White, Aleksandra Faust, Shimon Whiteson, et al. Imitation is not enough: Robustifying imitation with reinforcement learning for challenging driving scenarios. In 2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 7553-7560. IEEE, 2023. 1, 2, 7, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 613, + 553, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 613, + 553, + 677 + ], + "spans": [ + { + "bbox": [ + 317, + 613, + 553, + 677 + ], + "type": "text", + "content": "[20] Osama Makansi, Eddy Ilg, Ozgun Cicek, and Thomas Brox. Overcoming limitations of mixture density networks: A sampling and fitting framework for multimodal future prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7144-7153, 2019. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 681, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 681, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 681, + 553, + 713 + ], + "type": "text", + "content": "[21] Nico Montali, John Lambert, Paul Mougin, Alex Kuefler, Nicholas Rhinehart, Michelle Li, Cole Gulino, Tristan Emrich, Zoey Yang, Shimon Whiteson, et al. The waymo open" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "5430" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 77, + 73, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 73, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 77, + 73, + 294, + 95 + ], + "type": "text", + "content": "sim agents challenge. Advances in Neural Information Processing Systems (NeurIPS), 36, 2024. 1, 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 159 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 159 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 159 + ], + "type": "text", + "content": "[22] Jiquan Ngiam, Benjamin Caine, Vijay Vasudevan, Zhengdong Zhang, Hao-Tien Lewis Chiang, Jeffrey Ling, Rebecca Roelofs, Alex Bewley, Chenxi Liu, Ashish Venugopal, et al. Scene transformer: A unified architecture for predicting multiple agent trajectories. arXiv preprint arXiv:2106.08417, 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 161, + 294, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 161, + 294, + 216 + ], + "spans": [ + { + "bbox": [ + 56, + 161, + 294, + 216 + ], + "type": "text", + "content": "[23] Zhenghao Peng, Wenjie Luo, Yiren Lu, Tianyi Shen, Cole Gulino, Ari Seff, and Justin Fu. Improving agent behaviors with rl fine-tuning for autonomous driving. In European Conference on Computer Vision (ECCV), pages 165-181. Springer, 2025. 1, 2, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 217, + 294, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 217, + 294, + 259 + ], + "spans": [ + { + "bbox": [ + 56, + 217, + 294, + 259 + ], + "type": "text", + "content": "[24] Jonah Philion, Xue Bin Peng, and Sanja Fidler. Trajeglish: Traffic modeling as next-token prediction. In International Conference on Learning Representations (ICLR), 2024. 2, 3, 4, 5, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 261, + 294, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 261, + 294, + 316 + ], + "spans": [ + { + "bbox": [ + 56, + 261, + 294, + 316 + ], + "type": "text", + "content": "[25] Aditya Prakash, Aseem Behl, Eshed Ohn-Bar, Kashyap Chitta, and Andreas Geiger. Exploring data aggregation in policy learning for vision-based urban autonomous driving. In Conference on Computer Vision and Pattern Recognition (CVPR), pages 11763-11773, 2020. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 316, + 294, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 316, + 294, + 370 + ], + "spans": [ + { + "bbox": [ + 56, + 316, + 294, + 370 + ], + "type": "text", + "content": "[26] Ethan Pronovost, Meghan Reddy Ganesina, Noureldin Hendy, Zeyu Wang, Andres Morales, Kai Wang, and Nick Roy. Scenario diffusion: Controllable driving scenario generation with diffusion. Advances in Neural Information Processing Systems (NeurIPS), 36:68873-68894, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 371, + 294, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 371, + 294, + 426 + ], + "spans": [ + { + "bbox": [ + 56, + 371, + 294, + 426 + ], + "type": "text", + "content": "[27] Stéphane Ross, Geoffrey Gordon, and Drew Bagnell. A reduction of imitation learning and structured prediction to no-regret online learning. In International Conference on Artificial Intelligence and Statistics (AISTATS), pages 627-635. JMLR Workshop and Conference Proceedings, 2011. 1, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 426, + 294, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 426, + 294, + 491 + ], + "spans": [ + { + "bbox": [ + 56, + 426, + 294, + 491 + ], + "type": "text", + "content": "[28] Ari Seff, Brian Cera, Dian Chen, Mason Ng, Aurick Zhou, Nigamaa Nayakanti, Khaled S Refaat, Rami Al-Rfou, and Benjamin Sapp. Motionlm: Multi-agent motion forecasting as language modeling. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 8579-8590, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 492, + 294, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 492, + 294, + 536 + ], + "spans": [ + { + "bbox": [ + 56, + 492, + 294, + 536 + ], + "type": "text", + "content": "[29] Shaoshuai Shi, Li Jiang, Dengxin Dai, and Bernt Schiele. Motion transformer with global intention localization and local movement refinement. Advances in Neural Information Processing Systems (NeurIPS), 35:6531-6543, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 537, + 294, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 537, + 294, + 590 + ], + "spans": [ + { + "bbox": [ + 56, + 537, + 294, + 590 + ], + "type": "text", + "content": "[30] Shaoshuai Shi, Li Jiang, Dengxin Dai, and Bernt Schiele. Mtr++: Multi-agent motion prediction with symmetric scene modeling and guided intention querying. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 2024. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 592, + 294, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 592, + 294, + 635 + ], + "spans": [ + { + "bbox": [ + 56, + 592, + 294, + 635 + ], + "type": "text", + "content": "[31] Simon Suo, Sebastian Regalado, Sergio Casas, and Raquel Urtasun. TrafficSim: Learning to simulate realistic multi-agent behaviors. In Conference on Computer Vision and Pattern Recognition (CVPR), pages 10400-10409, 2021. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 636, + 294, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 636, + 294, + 690 + ], + "spans": [ + { + "bbox": [ + 56, + 636, + 294, + 690 + ], + "type": "text", + "content": "[32] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 691, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 712 + ], + "type": "text", + "content": "[33] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "text", + "content": "Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems (NeurIPS), pages 5998-6008, 2017. 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 107, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 150 + ], + "type": "text", + "content": "[34] Arun Venkatraman, Martial Hebert, and J Bagnell. Improving multi-step prediction of learned time series models. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), 2015. 2, 4, 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 152, + 553, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 553, + 195 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 553, + 195 + ], + "type": "text", + "content": "[35] Wei Wu, Xiaoxin Feng, Ziyan Gao, and Yuheng Kan. Smart: Scalable multi-agent real-time simulation via next-token prediction. Advances in Neural Information Processing Systems (NeurIPS), 2025. 1, 2, 3, 4, 5, 6, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 197, + 553, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 197, + 553, + 241 + ], + "spans": [ + { + "bbox": [ + 316, + 197, + 553, + 241 + ], + "type": "text", + "content": "[36] Danfei Xu, Yuxiao Chen, Boris Ivanovic, and Marco Pavone. Bits: Bi-level imitation for traffic simulation. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 2929-2936. IEEE, 2023. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 243, + 553, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 243, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 316, + 243, + 553, + 285 + ], + "type": "text", + "content": "[37] Chris Zhang, James Tu, Lunjun Zhang, Kelvin Wong, Simon Suo, and Raquel Urtasun. Learning realistic traffic agents in closed-loop. In Conference on Robot Learning (CoRL), 2023. 1, 2, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 286, + 553, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 286, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 316, + 286, + 553, + 319 + ], + "type": "text", + "content": "[38] Jiakai Zhang and Kyunghyun Cho. Query-efficient imitation learning for end-to-end autonomous driving. arXiv preprint arXiv:1605.06450, 2016. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 321, + 553, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 321, + 553, + 376 + ], + "spans": [ + { + "bbox": [ + 316, + 321, + 553, + 376 + ], + "type": "text", + "content": "[39] Zhejun Zhang, Alexander Liniger, Dengxin Dai, Fisher Yu, and Luc Van Gool. End-to-end urban driving by imitating a reinforcement learning coach. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 377, + 553, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 377, + 553, + 432 + ], + "spans": [ + { + "bbox": [ + 316, + 377, + 553, + 432 + ], + "type": "text", + "content": "[40] Zhejun Zhang, Alexander Liniger, Dengxin Dai, Fisher Yu, and Luc Van Gool. Trafficbots: Towards world models for autonomous driving simulation and motion prediction. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 1522-1529. IEEE, 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 434, + 553, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 434, + 553, + 487 + ], + "spans": [ + { + "bbox": [ + 316, + 434, + 553, + 487 + ], + "type": "text", + "content": "[41] Zhejun Zhang, Alexander Liniger, Christos Sakaridis, Fisher Yu, and Luc V Gool. Real-time motion prediction via heterogeneous polyline transformer with relative pose encoding. Advances in Neural Information Processing Systems (NeurIPS), 2024. 2, 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 489, + 553, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 489, + 553, + 533 + ], + "spans": [ + { + "bbox": [ + 316, + 489, + 553, + 533 + ], + "type": "text", + "content": "[42] Zhejun Zhang, Christos Sakaridis, and Luc Van Gool. Trafficbots v1. 5: Traffic simulation via conditional vaes and transformers with relative pose encoding. arXiv preprint arXiv:2406.10898, 2024. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 534, + 553, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 534, + 553, + 588 + ], + "spans": [ + { + "bbox": [ + 316, + 534, + 553, + 588 + ], + "type": "text", + "content": "[43] Jianbo Zhao, Jiaheng Zhuang, Qibin Zhou, Taiyu Ban, Ziyao Xu, Hangning Zhou, Junhe Wang, Guoan Wang, Zhiheng Li, and Bin Li. Kigras: Kinematic-driven generative model for realistic agent simulation. arXiv preprint arXiv:2407.12940, 2024. 1, 2, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 590, + 553, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 590, + 553, + 634 + ], + "spans": [ + { + "bbox": [ + 316, + 590, + 553, + 634 + ], + "type": "text", + "content": "[44] Ziyuan Zhong, Davis Rempe, Yuxiao Chen, Boris Ivanovic, Yulong Cao, Danfei Xu, Marco Pavone, and Baishakhi Ray. Language-guided traffic simulation via scene-level diffusion. arXiv preprint arXiv:2306.06344, 2023. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 635, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 553, + 689 + ], + "type": "text", + "content": "[45] Ziyuan Zhong, Davis Rempe, Danfei Xu, Yuxiao Chen, Sushant Veer, Tong Che, Baishakhi Ray, and Marco Pavone. Guided conditional diffusion for controllable traffic simulation. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 3560-3566. IEEE, 2023. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 691, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 712 + ], + "type": "text", + "content": "[46] Zikang Zhou, Jianping Wang, Yung-Hui Li, and Yu-Kai Huang. Query-centric trajectory prediction. In Proceedings" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "5431" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 152 + ], + "type": "list", + "angle": 0, + "index": 2, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "text", + "content": "of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17863-17873, 2023. 2, 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 152 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 152 + ], + "type": "text", + "content": "[47] Zikang Zhou, Haibo Hu, Xinhong Chen, Jianping Wang, Nan Guan, Kui Wu, Yung-Hui Li, Yu-Kai Huang, and Chun Jason Xue. Behaviorgpt: Smart agent simulation for autonomous driving with next-patch prediction. arXiv preprint arXiv:2405.17372, 2024. 1, 6, 2" + } + ] + } + ], + "index": 1 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "5432" + } + ] + } + ], + "index": 3 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Closest Neighbors are Harmful for Lightweight Masked Auto-encoders/ce070f35-5bfa-41af-9465-b1f84c742005_content_list.json b/2025/Closest Neighbors are Harmful for Lightweight Masked Auto-encoders/ce070f35-5bfa-41af-9465-b1f84c742005_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..7ac9a4f8c69ff02dcfcc859655288ba430af9d10 --- /dev/null +++ b/2025/Closest Neighbors are Harmful for Lightweight Masked Auto-encoders/ce070f35-5bfa-41af-9465-b1f84c742005_content_list.json @@ -0,0 +1,1368 @@ +[ + { + "type": "text", + "text": "Closest Neighbors are Harmful for Lightweight Masked Auto-encoders", + "text_level": 1, + "bbox": [ + 140, + 130, + 854, + 151 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Jian Meng, $^{1}$ Ahmed Hassssan, $^{2}$ Li Yang, $^{3}$ Deliang Fan, $^{4}$ Jinwoo Shin, and $^{1}$ Jae-sun Seo \n $^{1}$ Cornell University, $^{2}$ University of North Carolina at Charlotte, $^{3}$ Arizona State University, $^{4}$ KAIST \n $^{1}\\{jm2787, ah2288, js3528\\} @cornell.edu, ^{2}lyang50@uncc.edu, ^{3}dfan12@asu.edu, ^{4}jinwoos@kaist.ac.kr$", + "bbox": [ + 99, + 179, + 931, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 268, + 326, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Learning the visual representation via masked autoencoder (MAE) training has been proven to be a powerful technique. Transferring the pre-trained vision transformer (ViT) to downstream tasks leads to superior performance compared to conventional task-by-task supervised learning. Recent research works on MAE focus on large-sized vision transformers ( $>50$ million parameters) with outstanding performance. However, improving the generality of the under-parametrized lightweight model has been widely ignored. In practice, downstream applications are commonly intended for resource-constrained platforms, where large-scale ViT cannot easily meet the resource budget. Current lightweight MAE training heavily relies on knowledge distillation with a pre-trained teacher, whereas the root cause behind the poor performance remains underexplored. Motivated by that, this paper first introduces the concept of \"closest neighbor patch\" to characterize the local semantics among the input tokens. Our discovery shows that the lightweight model failed to distinguish different local information, leading to biased understanding and poor accuracy. Motivated by this finding, we propose NoR-MAE, a novel MAE training algorithm for lightweight vision transformers. NoR-MAE elegantly repels the semantic aliasing between patches and their closest neighboring patch (semantic centroid) with negligible training cost overhead. With the ViT-Tiny model, NoR-MAE achieves up to $7.22\\% / 3.64\\%$ accuracy improvements on ImageNet100/ImageNet-1K datasets, as well as up to $5.13\\%$ accuracy improvements in tested downstream tasks. https://github.com/SeoLabCornell/NoR-MAE", + "bbox": [ + 86, + 300, + 485, + 753 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 782, + 220, + 799 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Starting from the early exploration with contrastive learning [6, 18], learning powerful and generic vision or semantic representation has been the focal point of all prior self-supervised learning (SSL) algorithms across various machine learning domains [2, 8, 28, 30] and vision-language model [15, 26]. The superiority of SSL is demonstrated", + "bbox": [ + 89, + 809, + 483, + 900 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/74bc6eb3dbc3e2a1a4f6f4bf024bd78d64fd939e071375606bdf94fc097022c7.jpg", + "image_caption": [ + "Figure 1. Top: Closest Neighbor Patches (CNP) characterizes the \"Local Semantic Centroid\" among the input tokens. Bottom: Low-dimensional projection (via PCA) of the transformer-encoded patches. Unlike the large-sized transformer model (e.g., ViT-Base), lightweight models (e.g., ViT-Tiny) cannot separate the \"semantic difference\" between different CNPs." + ], + "image_footnote": [], + "bbox": [ + 516, + 265, + 906, + 452 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "by the capability of extracting transferable knowledge from unlabeled datasets [26, 33]. Recent efforts on masked autoencoder (MAE) [19] train the vision transformer [11] by encouraging the model to understand the semantic knowledge of the masked patches throughout the reconstruction-based self-supervised learning. Particularly, the evolution from contrastive learning [16, 18, 36] to MAE-based training [17, 19, 34, 37] shows consistent dominance by large-scale models to ensure on-par or even better accuracy compared to supervised learning, along with superior performance in various downstream applications. In practice, downstream small-scale tasks are commonly deployed on edge devices with limited resources. Therefore, large-sized vision transformers (ViTs) pre-trained by MAE [19] are suboptimal candidates for resource-constrained downstream deployment and hardware platforms. Unfortunately, recent state-of-the-art (SoTA) methods that train lightweight ViTs with MAE from scratch lead to poor performance and insufficient transferability. For example, the MAE-trained lightweight ViT-Tiny model only achieves $66.6\\%$ top-1 accuracy (with 200-epoch pre-training [3, 19, 37]) on the", + "bbox": [ + 511, + 583, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "25230", + "bbox": [ + 478, + 924, + 519, + 936 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ImageNet-1K dataset [3], which is $\\sim 10\\%$ lower than the supervised learning counterpart. Evidently, the superiority of MAE [19] with the large model failed to be maintained in the lightweight ViTs [11]. Given the diverse downstream tasks and urgent need for energy-efficient foundation models, unleashing the power of MAE and enhancing the versatility of lightweight models is desired.", + "bbox": [ + 89, + 90, + 480, + 196 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Prior works enhance the performance of the lightweight transformer models by heavily relying on knowledge distillation (KD) [20] for both supervised [29] and self-supervised learning paradigms [3, 21, 38]. For instance, the recent DMAE work [3] combines MAE [19] with KD and transfers the pre-trained knowledge from a ViT-Base teacher down to the lightweight ViT model. DMAE [3] chooses the 3/4 depth of both student and teacher ViT models as the bonding portal for feature alignment and knowledge distillation. The heuristic design of DMAE cannot be easily extended to other encoder models. Furthermore, distilling the knowledge on the fly introduces an additional projector [3] or logits mask [31], which further elevates the complexity and computation cost of the entire training process. In addition to DMAE, other recent works further increase the complexity of the KD-based MAE by employing multi-layer distillation [38] or complex generic-to-specific distillation [21]. All of these prior works [3, 21, 31, 38] rely on the brute force approach to enhance the model performance with increasingly complex training schemes, which results in inflated GPU memory and training time.", + "bbox": [ + 91, + 198, + 486, + 513 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, none of these prior works can collectively achieve 1) the capability of training lightweight model from scratch without a large-sized teacher model, 2) negligible training cost overhead compared to the vanilla MAE, and 3) clear insights that cause the poor performance of lightweight MAE. To overcome this, the key questions to address are:", + "bbox": [ + 89, + 515, + 483, + 606 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "① What is the bottleneck of MAE with under parametrized small vision transformer models?", + "② How to efficiently train a high performance lightweight model from scratch via MAE?" + ], + "bbox": [ + 89, + 609, + 482, + 667 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To answer ①, our investigation shows that the lightweight model failed to understand different local semantic information among the input tokens and lead to poor performance. We characterize different local semantics by introducing the concept of closest neighborhood patch (CNP) (Figure 1), which carries the most similar semantics of a given input patch. Since multiple patches can share one CNP (e.g., Background Forrest), different CNPs represent various local semantics across the entire input sequence. Unlike large-sized models, the under-parameterized model falls into the semantic aliasing between different CNPs with poor accuracy, as indicated in Figure 1 and Section 3.3.", + "bbox": [ + 89, + 672, + 482, + 852 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Motivated by the findings in ①, we propose the Neighbor Repelling Masked Autoencoder (NoR-MAE) to resolve ②, a novel self-supervised learning algorithm designed for train", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ing powerful lightweight vision transformers. NoR-MAE trains the powerful small vision transformer by penalizing the similarity between the unmasked patches and CNPs. Different from prior works, NoR-MAE directly trains the vision transformer from scratch without introducing the largesized pre-trained teacher or heuristically designed distillation scheme. The proposed method achieves up to $7.22\\%$ and $3.64\\%$ accuracy improvements on ImageNet-100 and ImageNet-1K datasets, with negligible training time and memory overhead compared to the vanilla MAE. The major contributions of our work are:", + "bbox": [ + 511, + 90, + 906, + 256 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Simplicity: NoR-MAE does not require teacher pretraining or heuristic knowledge distillation to train a powerful lightweight ViT.", + "- High Performance: Compared to the baseline MAE, NoR-MAE achieves $7.22\\%$ and $3.64\\%$ accuracy improvements on ImageNet-100 and ImageNet-1K datasets with the ViT-Tiny model, achieving the new SoTA performance on the lightweight MAE training.", + "- Transferrability: Different from prior works, which mainly focus on fine-tuning the performance on the pretraining dataset (e.g., ImageNet), NoR-MAE is also evaluated with the downstream fine-tuning and linear evaluation benchmarks, with significantly improved performance compared to prior SSL baselines and the vanilla MAE.", + "- Rationality: Besides the superior performance, for the first time, NoR-MAE reveals the semantic aliasing issue caused by the neighboring patches, which is proved to be the reason that hinders efficient training and representation learning of lightweight ViT models." + ], + "bbox": [ + 511, + 257, + 906, + 544 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 558, + 653, + 571 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Empowering a deep neural network model with strong and transferable knowledge has been widely investigated, especially in self-supervised learning. Early research works in contrastive learning (CL) introduce the learning paradigm with \"positive\" and \"negative\" sample pairs [6, 18], together with the alignment-repellent strategy embedded into InfoNCE [7, 18, 25] and NT-Xent loss [6] as the learning objective. The early success of CL motivated the successors to investigate the potential of different training and sampling strategies, inducing the standpoint of asymmetrical learning [16, 24], knowledge distillation [13], and the latent-dimension correlation [36].", + "bbox": [ + 511, + 583, + 906, + 763 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "While the core idea of CL is training the model to understand the similarities and differences between different augmentations, the emergence of masked autoencoder (MAE) [19] training reconsiders SSL from a different perspective, which is perfectly suitable for vision transformers. Encouraging the encoder-decoder to reconstruct the randomly sparsified patches allows the model to understand the semantics. Due to the powerful visual representation, MAE-based pretaining exhibits strong performance on", + "bbox": [ + 511, + 765, + 908, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "25231", + "bbox": [ + 478, + 924, + 517, + 936 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/aa9725bd0ee949afb6e6d813dfebc077932fe2a507d0e1ad34fa4413310f62ff.jpg", + "image_caption": [ + "Figure 2. Closest Neighbor Patches (CNP) within the image." + ], + "image_footnote": [], + "bbox": [ + 96, + 89, + 477, + 170 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "both backbone datasets (e.g., ImageNet-1K) and downstream transfer learning tasks. Subsequent research works further improved the performance of MAE from the perspectives of contrastive learning [27, 37], cross-attention [14, 17], mixed feature learning [23], or intricate tokenizer [12].", + "bbox": [ + 89, + 233, + 483, + 309 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Across all the different MAE training methods, elevating the performance of the large-sized models (e.g., ViT-Large) has been the primary focus of exploration, while improving the poor performance of the small-sized model (e.g., ViTTiny) has been largely ignored. Given that the downstream tasks are well-suited for resource-constrained edge devices, the imbalanced research focus of the prior works largely hinders the efficiency and transferability of MAE [19]. The practical needs and the imperfection of lightweight MAE motivate the recent research works to investigate MAE training strategies with small models. Following the protocol from supervised distillation [20], knowledge distillation-based MAE [3, 21] is proposed to enhance the performance of the lightweight student. However, the learning scheme of the single-layer or multi-layer distillation is established based on the heuristically designed distillation between intermediate features. Compared to the supervised knowledge distillation, recent studies largely complicate the overall cost of training lightweight ViT models with either heuristic design [3], multi-phased distillation [21], or distillation with extensive fine-tuning [31]. In particular, employing the pre-trained large ViT model as the teacher introduces additional GPU memory usage and time cost toward the total training efforts. Furthermore, training the interconnect projectors between the student and teacher exacerbates the training complexity even further. More importantly, the knowledge distillation-based MAE inherits the essence of KD from supervised learning, but the insights of training lightweight ViT from scratch via MAE remain largely unexplored.", + "bbox": [ + 91, + 314, + 483, + 753 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Proposed Method", + "text_level": 1, + "bbox": [ + 89, + 779, + 264, + 796 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We propose the Neighbor-Repelling MAE (NoR-MAE). In this section, we will 1) introduce the concept of the Closest Neighbor Patch into MAE training, 2) present the proposed NoR-MAE training, and 3) unravel the rationality of NoR-MAE by revealing the insights that cause the poor performance of the directly-trained lightweight ViT model.", + "bbox": [ + 89, + 809, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/fac5c4699b24f06121d93a3717d4ec1f9a6ccdf8d196fae1fd81ffb59723d3dc.jpg", + "image_caption": [ + "Figure 3. Cosine Similarity between each patch and their Closest Neighbor Patch across a subset of ImageNet-1K with 4096 images." + ], + "image_footnote": [], + "bbox": [ + 516, + 89, + 903, + 261 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Closest Neighbor Patch (CNP)", + "text_level": 1, + "bbox": [ + 511, + 330, + 782, + 347 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Let $X \\in \\mathbf{R}^{H \\times W \\times C}$ be an input image to the model, where $H$ , $W$ , and $C$ represent the height, width, and channel of the image. Vision transformer (ViT) reformulates the input by decomposing the image $X$ into a sequence of non-overlapped patches with the pre-defined patch size $P \\times P$ and the total number of patches (tokens) $K$ . All the patches are further encoded into the embeddings with dimension $D$ , which is the input of the subsequent transformer blocks.", + "bbox": [ + 511, + 354, + 906, + 477 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For a given input patch $X_P$ , there are naturally 8 neighbors that share a common edge with $X_P$ , as shown in Figure 2(a). We define the Closest Neighbor Patch (CNP) $X_{CNP}$ by selecting the neighbor that has the Top-1 cosine similarity with $X_P$ . In other words, $X_P$ and $X_{CNP}$ are sharing the highest degree of semantically meaningful relations.", + "bbox": [ + 511, + 479, + 908, + 570 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Shared Closest Neighbor Patch. In the example of Figure 2, two patches $X_P^i$ and $X_P^j$ partially share the neighborhoods in between, while the CNP of $X_P^i$ and $X_P^j$ happen to be identical as $X_{CNP}^{ij}$ . Naturally, $X_{CNP}^{ij}$ carries the shared semantic information (\"Background Forest\") for both $X_P^i$ and $X_P^j$ , but the degree of similarity between $(X_P^i, X_{CNP}^{ij})$ and $(X_P^j, X_{CNP}^{ij})$ are different.", + "bbox": [ + 511, + 571, + 908, + 686 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Closest Neighbor Patch $\\neq$ Identical Patch. Although $X_{P}$ and $X_{CNP}$ are sharing the highest degree of similarity, the semantics between $X_{P}$ and $X_{CNP}$ are not identical. As shown in Figure 3, the cosine similarity between all the patches and their closest neighbor in ImageNet-1K images varies from 0.07 to 0.999, with an average of 0.83. In other words, the highly similar neighboring patches also contain the implicit semantical difference.", + "bbox": [ + 511, + 686, + 908, + 806 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Difficulty of Lightweight Models to Understand the Different Semantics Carried by CNP. With MAE, the input patches of the transformer models are randomly masked with a pre-defined sparsity ratio $\\tau$ (e.g., $75\\%$ ), as illustrated in Figure 4, left. Naturally, each unmasked visible patch has a Closest Neighbor Patch (CNP). As a result, the CNP", + "bbox": [ + 511, + 809, + 908, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "25232", + "bbox": [ + 478, + 925, + 519, + 936 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/fe7c07b0b0d88cc7ead5211fbdf2d18f93ea04eb765971248ccbd407a769717e.jpg", + "image_caption": [ + "Figure 4. The shared Closest Neighbor Patch (CNP) characterizes the estimated local semantic centroids of the input sequence." + ], + "image_footnote": [], + "bbox": [ + 94, + 89, + 480, + 200 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "of the unmasked patches formulates another sparse input image (Figure 4, right). From the perspective of the entire image, visible patches and their CNPs are close neighbors, while different visible patches may share the same closest neighbor, which estimates the local and regional semantics of the input sequence. However, as described in Figure 1 and Section 1, understanding the semantic difference between different CNPs is the bottleneck for the under-parameterized models. Furthermore, our extended experiments in Figure 7 show that the key difference between the big (e.g., ViT-Base) and small (e.g., ViT-Tiny) model is understanding and separating the semantic differences between CNPs. Therefore, the challenge of lightweight MAE is clear:", + "bbox": [ + 88, + 271, + 483, + 467 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In MAE training, how to facilitate the lightweight model to understand different local semantics?", + "bbox": [ + 89, + 468, + 483, + 500 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Proposed Method: NoR-MAE", + "text_level": 1, + "bbox": [ + 89, + 510, + 359, + 526 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We propose the Neighbor-Repelling training for self-supervised MAE learning (NoR-MAE). The overall objective is to train a powerful, lightweight vision transformer by encouraging the model to avoid the biased relation between different local semantics of the input. We follow the standard protocol of the vanilla MAE with the encoder $f_{\\theta}$ and decoder $g_{\\phi}$ , where both $f_{\\theta}$ and $g_{\\phi}$ are vision transformers with the same embedding dimension $D$ .", + "bbox": [ + 89, + 535, + 483, + 655 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given the masking ratio $\\tau$ , we first generate the masked input sequence $x^{\\tau}$ and $x_{CNP}^{\\tau}$ that are embedded from the unmasked patches with the corresponding closest neighbor patches, respectively. During the forward pass of each iteration, we parallelly encode and decode the masked sequence $x^{\\tau}$ and CNP sequence $x_{CNP}^{\\tau}$ by $f_{\\theta}$ and $g_{\\phi}$ , generating the reconstructed embedding $Z$ and $Z_{CNP}$ , respectively. In particular, $f_{\\theta}$ and $g_{\\phi}$ are served as the siamese encoder and decoder (with stop gradient) during the forward pass of $Z_{CNP}$ .", + "bbox": [ + 89, + 657, + 483, + 792 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Overall, each forward pass will generate two pairs of decoded tokens resulting from $x^{\\tau}$ and $x_{\\mathrm{CNP}}^{\\tau}$ .", + "bbox": [ + 89, + 794, + 483, + 824 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Following the standard training protocol of MAE, we first compute the $\\mathcal{L}_2$ reconstruction loss between the decoded sequence $Z$ and the ground truth. Subsequently, we compute the matrix multiplication between the normalized $Z$ and $Z_{CNP}$ . As shown in Figure 5, the encoder $f_{\\theta}$ and decoder", + "bbox": [ + 89, + 825, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$g_{\\phi}$ are shared for both branches, while the stop gradient is enabled for the CNP input. Given the size of $[N,K,D]$ , the matrix multiplication between $\\bar{Z}$ and transposed $\\bar{Z}_{\\mathrm{CNP}}$ is performed along the token dimension $K$ . The resultant matrix is further scaled down by $K$ and takes the average along the batch dimension $N$ . Mathematically, we have:", + "bbox": [ + 511, + 90, + 905, + 181 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nC _ {\\text {n e i g h b o r}} = \\frac {1}{N} \\sum_ {N} \\left(\\frac {1}{K} (\\bar {Z}) ^ {T} \\cdot \\bar {Z} _ {\\mathrm {C N P}}\\right) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 584, + 191, + 906, + 227 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Where $C_{\\mathrm{neighbor}}$ represents the resultant relation matrix with the size of $D \\times D$ . Each entry $c_{i,j}$ of $C_{\\mathrm{neighbor}}$ characterizes the correlation between the randomly-unmasked input sequence (output logits $Z$ ) and the corresponding CNP sequence (output logits $Z_{\\mathrm{CNP}}$ ) between embedding dimension $i$ and $j$ . To minimize the semantic aliasing between the unmasked tokens and their CNPs, we decorate different embeddings by penalizing the off-diagonal terms of $C_{\\mathrm{neighbor}}$ :", + "bbox": [ + 511, + 232, + 906, + 354 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {N o R}} = \\sum_ {i} \\sum_ {i \\neq j} C _ {\\text {n e i g h b o r}} ^ {i, j} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 625, + 364, + 906, + 398 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Essentially, the proposed NoR-Loss $(\\mathcal{L}_{\\mathrm{NoR}})$ minimizes semantic aliasing by preventing the model encoding the visible patches and CNP as aliased embeddings. It has been shown in Figure 3 that patches and their CNPs share similar but non-identical semantics. We balance such a tradeoff of \"similarity-discrepancy\" by decorrelating the off-diagonal embeddings only. The overall loss is the combination of the standard $\\mathcal{L}_2$ reconstruction loss and the proposed NeighborRepelling Loss $\\mathcal{L}_{\\mathrm{NoR}}$ scaled by a hyperparameter $\\lambda$ :", + "bbox": [ + 511, + 407, + 908, + 546 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {t o t a l}} = \\left| \\left| \\text {G r o u n d T r u t h} - \\bar {Z} \\right| \\right| _ {2} + \\lambda \\cdot \\mathcal {L} _ {\\mathrm {N o R}} (\\bar {Z}, \\bar {Z} _ {\\mathrm {C N P}}) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 526, + 554, + 906, + 573 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "More importantly, decorrelating embeddings eventually separate the aliased semantics between different CNPs and their corresponding \"local semantic centroids\" (Section 3.3). In other words, the embedding-level anti-aliasing regularization of the proposed NoR-Loss facilitates the understanding on the representation (token) level, leading to the largely improved model performance.", + "bbox": [ + 511, + 583, + 908, + 688 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Decorrelating Embeddings vs. Decorrelating Tokens. Although the proposed NoR-MAE selects the Top-1 neighbor (the closest neighbor patch (CNP)) as the candidate for decorrelation, the CNP patches can be overlapped or shared between each other. For instance, the closest neighbor patch of patch $X_{i}$ could also be the Top-2/3/4 ... 8 neighbor of patch $X_{j}$ . Minimizing the semantic aliasing by naively decorrelating tokens leads to collapsed performance and unsuccessful training, as shown in Figure 6.", + "bbox": [ + 511, + 689, + 908, + 824 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "NoR-MAE vs. Contrastive Learning. Empowering representation learning via invariance-covariance optimization has been investigated in Barlow Twins [36]. Recently proposed U-MAE [37] combines the contrastive learning and MAE [19] by introducing the random asymmetry masks", + "bbox": [ + 511, + 825, + 908, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "25233", + "bbox": [ + 478, + 924, + 517, + 936 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/d3373eb9e2747902089d8a19e3c05b23dc98b5f1630e933e554681ef24e57089.jpg", + "image_caption": [ + "Figure 5. Overview of the proposed NoR-MAE algorithm." + ], + "image_footnote": [], + "bbox": [ + 155, + 88, + 848, + 315 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/b60a5977a31156d6ab39a487c36b30b811d9a96cd9f5c03d8ed2d552827182f7.jpg", + "image_caption": [ + "Figure 6. Different relation penalty of the NoR-Loss $(\\lambda = 5\\mathrm{e} - 5)$" + ], + "image_footnote": [], + "bbox": [ + 112, + 362, + 460, + 486 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "during the forward pass of the training. On the contrary, the proposed NoR-MAE exploits the semantic aliasing issue by constructing the asymmetry between the visible patches and the corresponding closest neighbors, instead of relying on the randomness. As a result, NoR-MAE outperforms U-MAE [37] in various sizes of models (Section 4).", + "bbox": [ + 89, + 537, + 483, + 628 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the context of contrastive learning, the similar samples (patches) are considered as \"positive samples\" which are intended to be aligned by the algorithms [35]. However, the important discovery of NoR-MAE points out that semantic aliasing among the closest neighbors is harmful for the lightweight masked autoencoder training, which is not properly addressed by prior works. Moreover, unlike Barlow Twins [36], NoR-MAE does not need the embedding alignment along the diagonal terms due to the well-established similarity between the visible patches and their CNPs.", + "bbox": [ + 88, + 628, + 485, + 779 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "NoR-MAE Does Not Need a Teacher Model. We would like to highlight that the proposed NoR-MAE is entirely different from knowledge distillation-based MAE. NoR-MAE does not require a pre-trained teacher [3] or multistage distillation [21] during training. The lightweight encoder is trained from scratch following the standard MAE protocol [19]. Although the CNPs and the original unmasked patches are separately generated, the encoding-decoding pro", + "bbox": [ + 89, + 780, + 485, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "cess remains the same as the vanilla MAE. For example, given the $75\\%$ mask ratio with $25\\%$ visible patches, the equivalent input sparsity of NoR-MAE is $50\\%$ . The overhead of the NoR-MAE is minimal compared to the vanilla MAE, as profiled in the supplementary.", + "bbox": [ + 511, + 366, + 906, + 441 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Rationality of NoR-MAE", + "text_level": 1, + "bbox": [ + 511, + 452, + 743, + 468 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As shown in Figure 2 and Figure 4, the proposed concept of Closest Neighbor Patch (CNP) characterizes the local semantics of the image. For instance, the patches that contain \"Background Forest\" or \"Back of the Fox\" are represented by separate CNP patches. However, the lightweight model failed to learn the semantic difference between CNPs.", + "bbox": [ + 511, + 476, + 906, + 566 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We prove the significance of the CNP semantics by analyzing the original unmasked features and CNP features encoded by the trained encoder $f_{\\theta}$ . Specifically, we first normalize the encoded features and then compress the feature dimension down to 2-D via the Principal Component Analysis (PCA) [1]. As shown in Figure 7 and Figure 8, the lightweight models (ViT-Tiny and ViT-Small) trained by the vanilla MAE [19] failed to distinguish the semantical difference between different closest neighbor patches (CNP, blue dots), regardless of the patch sizes. Different semantics of CNP are heavily mixed together in the low dimensional space. Moreover, CNP is critical for MAE training regardless of the input patch sizes and the corresponding sequence length. The distorted and aliased understanding of the local semantics leads to the sub-optimal performance of MAE, as shown in Table 1 with the empirical verification on ImageNet-100 ( $224 \\times 224$ ) with ViT-Tiny.", + "bbox": [ + 511, + 566, + 908, + 824 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "On the contrary, the model trained by NoR-MAE successfully understands the semantic difference between different CNP features, leading to largely improved accuracy on the lightweight ViT models across different patch sizes. In the meantime, the large-sized model (e.g., ViT-Base) exhibits", + "bbox": [ + 511, + 825, + 908, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "25234", + "bbox": [ + 478, + 925, + 519, + 936 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/48ece0badf6f42df429dff5acb665a13e75d4b78025050dd218266046eece57a.jpg", + "image_caption": [ + "ViT-Tiny (# of params = 5.7 Million) MAE [19]: Aliased Semantics! NoR-MAE: Clearly Separated CNP!", + "Accuracy $= 66.60$" + ], + "image_footnote": [], + "bbox": [ + 125, + 114, + 233, + 183 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a3c7f37475c3840661a41f651fc6cdd3f5c2ce776f044424d7bd7d13fdd8c037.jpg", + "image_caption": [ + "Accuracy $= 70.24$" + ], + "image_footnote": [], + "bbox": [ + 250, + 114, + 359, + 181 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b4d68fa8ebda505a9e55496391aba91dc406ddcfd9f979af87b1fbba76ee70fa.jpg", + "image_caption": [ + "Latent Feature of Unmasked Patches", + "Latent Feature of Closest Neighbor Patches" + ], + "image_footnote": [], + "bbox": [ + 380, + 90, + 630, + 191 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c8d1d286339bd12eb261ad440299bfac49a011b79b73297a9177b3873587d88a.jpg", + "image_caption": [ + "Patch size $= 16^{2}$ Token length $= 196$" + ], + "image_footnote": [], + "bbox": [ + 643, + 90, + 888, + 191 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e261fe14e69b24b254ab14b55e69a4de170ae8965021bae505c05c3edeec4168.jpg", + "image_caption": [ + "Figure 7. Low dimensional features of the unmasked patches (red dots) and the closest neighbor patches (CNP) (blue dots) with the same input from ImageNet-1K dataset. NoR-MAE clearly separates the semantic information among different CNP, with improved accuracy.", + "Figure 8. Low dimensional features of the unmasked patches (red dots) and the CNP (blue dots) encoded by ViT-Tiny with patch size $= 4 \\times 4$ (sequence length $= 3136$ ), evaluated on ImageNet-100." + ], + "image_footnote": [], + "bbox": [ + 107, + 273, + 472, + 386 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/f0303f5faaf2dc8ac86402fb0441eaa2035140937299cd8d188a7fbeb96b7f2a.jpg", + "table_caption": [ + "Table 1. NoR-MAE consistently outperforms the MAE [19] across different input patch sizes (evaluated on the ImageNet-100 dataset)." + ], + "table_footnote": [], + "table_body": "
Masking RatioInput Patch SizeMAE [19]NoR-MAE (This work)
75%2²71.08%83.56% (+12.48%)
75%4²71.16%83.58% (+12.42%)
75%8²71.27%79.24% (+7.97%)
75%16²71.04%78.26% (+7.22%)
75%32²68.42%74.73% (+6.31%)
", + "bbox": [ + 91, + 498, + 483, + 598 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "a better understanding and stronger separation of different semantics with the vanilla MAE [19].", + "bbox": [ + 89, + 625, + 482, + 654 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In summary, the proposed NoR-MAE successfully mitigates the performance gap between model sizes, facilitating the learnability of the transformer model with a stronger understanding of the semantical relations from the input, leading to improved accuracy with both lightweight and large-sized transformer models.", + "bbox": [ + 89, + 655, + 483, + 746 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experimental Results", + "text_level": 1, + "bbox": [ + 89, + 761, + 294, + 777 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. ImageNet Classification", + "text_level": 1, + "bbox": [ + 89, + 786, + 308, + 801 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For image classification tasks, we follow the standard protocol of MAE [19] by training the lightweight encoder on the ImageNet-1K and ImageNet-100 datasets with the batch size of 4,096 and 2,048, and subsequently performing the standard fine-tuning [19]. Specifically, the model is pre-trained by 200 and 400 epochs with the initial learning rate set to", + "bbox": [ + 89, + 810, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "1.5e-4, together with the weight decay value of 0.05. The penalty coefficient $\\lambda$ of the proposed NoR Loss is set to 5e-5, and the optimality of the selected penalty level is demonstrated in Table 5. The mask ratio is fixed at $75\\%$ based on the findings in the vanilla MAE [19]. We implement the standard fine-tuning protocol with 100 epochs as the vanilla MAE [19]. The detailed experimental setup (pre-train and downstream) are summarized in the supplementary, together with the report of GPU memory usage and training cost.", + "bbox": [ + 511, + 273, + 906, + 410 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2 and Table 3 summarize the performance of the NoR-MAE on ImageNet-100 and ImageNet-1K datasets. Compared to the vanilla MAE [19], NoR-MAE largely improves the performance on top of the MAE baseline [19] with up to $7.22\\%$ and $3.64\\%$ accuracy on ImageNet-100 and ImageNet-1K datasets, respectively. For the large-sized vision transformer training (e.g., ViT-Base), NoR-MAE still maintains the superiority compared to the recent high-performance MAE training (Table 3).", + "bbox": [ + 511, + 411, + 908, + 547 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "More importantly, NoR-MAE improves the model performance marginal training cost overhead. NoR-MAE trains the model with the standard MAE protocol without introducing 1) a pre-trained teacher [3] or 2) other additional architectures [27] or clustering modules [12]. As profiled in the supplementary, the performance boost reward by the NoR-MAE only exhibits up to $8\\%$ memory overhead and $2\\%$ additional training time. Compared to the distillation-based MAE [3], NoR-MAE outperforms DMAE [3] with $\\sim 1\\%$ accuracy improvement and $50\\%$ less training cost.", + "bbox": [ + 511, + 549, + 908, + 700 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We further validate the NoR-MAE on the larger ViT-Base model. The proposed algorithm consistently achieves better performance compared to the MAE baseline and the recent ImageNet-100 SoTA method without introducing any additional clustering efforts or special tokenizer [12].", + "bbox": [ + 511, + 700, + 908, + 777 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Finally, NoR-MAE is also suitable for the MAE with supervision (SupMAE) [22]. Introducing supervised learning into MAE does not impact the effectiveness of the proposed NoR-MAE, while the NoR-MAE further enhances the performance of SupMAE [22], as shown in Table 2.", + "bbox": [ + 511, + 777, + 908, + 853 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Comparison with the Distillation-based MAE. We would like to highlight that the proposed NoR-MAE can achieve on-par or even better performance compared to the", + "bbox": [ + 511, + 854, + 906, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "25235", + "bbox": [ + 478, + 925, + 517, + 936 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/fceec31f7dddf08af45711f5cbeb78b68de0c63db0b52e8af63952b09279c510.jpg", + "table_caption": [ + "Table 2. Fine-tuning and semi-supervised learning accuracy (%) of the ViT encoder pre-trained by MAE [19] and recent SoTA lightweight MAE on ImageNet-1K dataset (with 200 epochs and 400 epochs pre-training)." + ], + "table_footnote": [], + "table_body": "
ViT Model# of ParamsMethodTeacher ModelPre-training EpochSupervised Fine-tuning Accuracy (%)
ViT-Tiny5.7MMAE [19]-20066.60
MAE [19]-40067.63
Sup MAE [22]-20067.88
Sup MAE [22]-40068.91
MoCo-V3 [18]-20070.09
DINO [5]-30068.17
SimMIM [34]-20066.08
DMAE [3]ViT-B20070.00
Sup NoR-MAE (This work)-20070.63 (+2.75)
NoR-MAE (This work)-20070.24 (+3.64)
NoR-MAE (This work)-40071.18 (+3.55)
ViT-Small21.7MMAE [19]-20079.00
MAE [19]-40080.11
Sup MAE [22]-20079.27
CrossMAE [14]-40079.30
MoCo-V3 [18]-20079.46
DMAE [3]ViT-B20079.30
Sup NoR-MAE (This work)-20080.16 (+0.89)
NoR-MAE (This work)-20080.13 (+1.03)
NoR-MAE (This work)-40081.04 (+0.93)
ViT-Base88.6MMAE [19]-20083.30
Sup MAE [22]-20083.60
DMAE [3]-20084.00
CrossMAE [14]-20083.60
MoCo-V3 [3]-20083.14
DINO [5]-20082.80
SimMIM [34]-20083.50
U-MAE [37]-20083.00
NoR-MAE (This work)-20083.42 (+0.12)
Sup NoR-MAE (This work)-20083.70 (+0.10)
NoR-DMAE (This work)-20084.22 (+0.22)
", + "bbox": [ + 130, + 126, + 864, + 645 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "distillation-based MAE, while significantly reducing the training resources, in terms of GPU training time and memory usage. As shown in Table 2, NoR-MAE directly trains the lightweight model from scratch, achieving on-par or even better performance compared to the recent distillation-based MAE training, which yet consumes massive GPU resources.", + "bbox": [ + 89, + 667, + 485, + 761 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Transfer Learning on the Downstream Tasks", + "text_level": 1, + "bbox": [ + 89, + 770, + 470, + 787 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We fine-tune the pre-trained ViT-Tiny and ViT-Small models (from Table 2) on the downstream vision tasks with 100 epochs. As shown in Table 4, the proposed NoR-MAE algorithm outperforms the MAE baseline [19] among all the downstream tasks with up to $5.13\\%$ accuracy improvements. The outstanding downstream performance indicates the capability of learning strong representation through NoR-MAE.", + "bbox": [ + 89, + 794, + 485, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 513, + 667, + 666, + 683 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Intensity of Neighbor-Repelling $(\\lambda)$ . We evaluate the impact of the penalty level $\\lambda$ with different values. As shown in Table 5, the proposed NoR-MAE algorithm achieves the best performance with $\\lambda = 5\\mathrm{e} - 5$ . In other words, the repelling between the embedding features of the visible patches and the CNP should be properly controlled. Over-penalized semantic relation leads to sub-optimal performance.", + "bbox": [ + 511, + 696, + 908, + 803 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Impact of Masking Ratio. In addition to the $75\\%$ masking ratio that has been used in all prior works as the default setting, we evaluate the impact of different input masking ratios, varying from $50\\%$ to $85\\%$ sparsity. Among different input masking ratios, the proposed NoR-MAE consistently outperforms the vanilla MAE [19] with lightweight vision", + "bbox": [ + 511, + 809, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "25236", + "bbox": [ + 478, + 925, + 519, + 936 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/3360bdb22e4e5203b11368cbb0a97f615dba953c55afca0e2ada8a189ed4cdc8.jpg", + "table_caption": [ + "Table 3. Fine-tuning accuracy (%) of the vision transformer encoder pre-trained by MAE and recent SoTA MAE methods on ImageNet-100." + ], + "table_footnote": [], + "table_body": "
ViT Model# of ParametersPre-training MethodPre-training EpochsSupervised Fine-tuning Accuracy (%)
ViT-Tiny5.7MMAE [19]20071.04
DiNO-Tiny [5, 9]20063.04
CrossMAE [14]20070.82
NoR-MAE (This work)20078.26 (+7.22)
ViT-Small21.7MMAE [19]20081.82
MaskFeat HOG [32]20082.80
PeCo [10]20083.60
NoR-MAE (This work)20084.28 (+2.46)
ViT-Base88.6MMAE [19]20086.80
CrossMAE [14]20086.29
U-MAE [37]20086.80
BEiT [4]20086.10
NoR-MAE (This work)20087.56 (+0.76)
", + "bbox": [ + 130, + 114, + 864, + 375 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/2d1bac84e481a2f4088286a0dfa059f056dfea473085ccb0ff72e5d948fc72db.jpg", + "table_caption": [ + "Table 4. Fine-tuning accuracy (%) of the NoR-MAE on the downstream tasks with the lightweight ViT pre-trained on the ImageNet-1K." + ], + "table_footnote": [], + "table_body": "
ViT ModelMethodCIFAR-10CIFAR-100FlowersPetsDTDFoodCaltech-101AircraftADE20K (mIoU)
ViT-TinyMAE [19]95.0178.7485.8082.4755.0775.6368.0564.6026.54
NoR-MAE (This work)95.9880.1388.7185.2960.1878.7173.1867.0733.54
ViT-SmallMAE [19]97.5084.8391.4991.1462.3181.5780.8555.9041.14
NoR-MAE (This work)98.1785.5791.8591.7064.7983.1582.0857.4142.99
", + "bbox": [ + 109, + 412, + 883, + 497 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/b5cf72c9eb31b49ba95bd8e21b3e65347d4a84491360aa39d04c260391834cea.jpg", + "table_caption": [ + "Table 5. Impact of different penalty levels $(\\lambda)$ on Neighbor Repelling on the ImageNet-100 dataset." + ], + "table_footnote": [], + "table_body": "
Penalty Level (λ)0.0 (Baseline)1e-55e-51e-45e-4
ViT-Tiny70.2476.2078.2677.8972.17
ViT-Small81.8282.1484.2884.2181.83
", + "bbox": [ + 91, + 559, + 483, + 606 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/21e231951e5b1b7d9bd179e3f822d7cd39a7a599106be2d60c4e0ff88a2f41a7.jpg", + "table_caption": [ + "Table 6. Impact of different input masking ratios. NoR-MAE shows consistent performance improvements on the ImageNet-1K dataset with ViT-Tiny model." + ], + "table_footnote": [], + "table_body": "
Patch-wise Masking RatioBaseline MAE [19]NoR-MAE (This work)
50%65.43%69.27%
75%66.60%70.24%
85%66.37%70.16%
", + "bbox": [ + 91, + 667, + 483, + 734 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "transformers, as shown in Table 6. In other words, semantic aliasing is persistent in lightweight transformer models regardless of the information density in the input sequence.", + "bbox": [ + 89, + 763, + 482, + 808 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "NoR-MAE with extended training effort We further report the performance of the NoR-MAE algorithm with 1,000 and 1,600 epoch training on the ImageNet-100. As shown in Table 7, the extended training effort can further boost up the performance of the NoR-MAE-trained ViT-Tiny (5.2M) model to $83.96\\%$ , while consistently outperforming MAE.", + "bbox": [ + 89, + 810, + 483, + 901 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/340e97d5a5a6aeab9e3242f5bdc38a47693307c74c21bef70d58ac9f3a8a0f5a.jpg", + "table_caption": [ + "Table 7. Performance of NoR-MAE with extended training effort." + ], + "table_footnote": [], + "table_body": "
Training EpochsMAE on ImageNet-100 (%)NoR-MAE on ImageNet-100 (%)
20071.0478.26
100073.2282.97
160074.2983.96
", + "bbox": [ + 514, + 534, + 903, + 595 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 648, + 633, + 666 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we propose NoR-MAE, a novel self-supervised learning algorithm designed for lightweight vision transformers via masked autoencoder training. We first introduce the concept of closest neighbor patches (CNP) into the MAE training, which is a critical concept in lightweight SSL. On top of that, we propose Neighbor-Repelling Loss for MAE training. The proposed method trains the lightweight model from scratch without using a pre-trained teacher model. NoR-MAE achieves the new state-of-the-art performance on direct lightweight MAE training with largely improved performance on both pre-training and downstream vision tasks. Finally, our discovery of the biased semantic relations provides valuable insights regarding lightweight masked autoencoder learning.", + "bbox": [ + 509, + 688, + 908, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "25237", + "bbox": [ + 478, + 925, + 519, + 936 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Acknowledgment", + "text_level": 1, + "bbox": [ + 91, + 90, + 263, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work is supported in part by the National Science Foundation under grant 2403723, and CoCoSys Center in JUMP 2.0, an SRC program sponsored by DARPA.", + "bbox": [ + 89, + 114, + 485, + 161 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 176, + 187, + 193 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Herve Abdi and Lynne J Williams. Principal component analysis. Wiley interdisciplinary reviews: computational statistics, 2(4):433-459, 2010. 5", + "[2] Alexei Baevski, Wei-Ning Hsu, Qiantong Xu, Arun Babu, Jiatao Gu, and Michael Auli. Data2vec: A general framework for self-supervised learning in speech, vision and language. In International Conference on Machine Learning (ICML), 2022. 1", + "[3] Yutong Bai, Zeyu Wang, Junfei Xiao, Chen Wei, Huiyu Wang, Alan L Yuille, Yuyin Zhou, and Cihang Xie. Masked autoencoders enable efficient knowledge distillers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1, 2, 3, 5, 6, 7", + "[4] Hangbo Bao, Li Dong, Songhao Piao, and Furu Wei. BEit: BERT pre-training of image transformers. In International Conference on Learning Representations (ICLR), 2022. 8", + "[5] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the International Conference on Computer Vision (ICCV), 2021. 7, 8", + "[6] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International Conference on Machine Learning (ICML), 2020. 1, 2", + "[7] Xinlei Chen, Haoqi Fan, Ross Girshick, and Kaiming He. Improved Baselines with Momentum Contrastive learning. arXiv preprint arXiv:2003.04297, 2020. 2", + "[8] Yu-An Chung, Yu Zhang, Wei Han, Chung-Cheng Chiu, James Qin, Ruoming Pang, and Yonghui Wu. W2v-bert: Combining contrastive learning and masked language modeling for self-supervised speech pre-training. In 2021 IEEE Automatic Speech Recognition and Understanding Workshop (ASRU), 2021. 1", + "[9] Victor Guilherme Turrisi da Costa, Enrico Fini, Moin Nabi, Nicu Sebe, and Elisa Ricci. solo-learn: A library of self-supervised methods for visual representation learning. Journal of Machine Learning Research, 2022. 8", + "[10] Xiaoyi Dong, Jianmin Bao, Ting Zhang, Dongdong Chen, Weiming Zhang, Lu Yuan, Dong Chen, Fang Wen, Nenghai Yu, and Baining Guo. Peco: Perceptual codebook for bert pre-training of vision transformers. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), 2023. 8", + "[11] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 1, 2" + ], + "bbox": [ + 93, + 202, + 485, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[12] Tianqi Du, Yifei Wang, and Yisen Wang. On the role of discrete tokenization in visual representation learning. In International Conference on Learning Representations (ICLR), 2023. 3, 6", + "[13] Zhiyuan Fang, Jianfeng Wang, Lijuan Wang, Lei Zhang, Yezhou Yang, and Zicheng Liu. Seed: Self-supervised distillation for visual representation. In International Conference on Learning Representations (ICLR), 2021. 2", + "[14] Letian Fu, Long Lian, Renhao Wang, Baifeng Shi, Xudong Wang, Adam Yala, Trevor Darrell, Alexei A Efros, and Ken Goldberg. Rethinking patch dependence for masked autoencoders. arXiv preprint arXiv:2401.14391, 2024. 3, 7, 8", + "[15] Peng Gao, Shijie Geng, Renrui Zhang, Teli Ma, Rongyao Fang, Yongfeng Zhang, Hongsheng Li, and Yu Qiao. Clip-adapter: Better vision-language models with feature adapters. International Journal of Computer Vision, 132(2):581-595, 2024. 1", + "[16] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. Advances in Neural Information Processing Systems (NeurIPS), 2020. 1, 2", + "[17] Agrim Gupta, Jiajun Wu, Jia Deng, and Li Fei-Fei. Siamese masked autoencoders. In Thirty-seventh Conference on Neural Information Processing Systems (NeurIPS), 2023. 1, 3", + "[18] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 1, 2, 7", + "[19] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dólár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 2, 3, 4, 5, 6, 7, 8", + "[20] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 2, 3", + "[21] Wei Huang, Zhiliang Peng, Li Dong, Furu Wei, Jianbin Jiao, and Qixiang Ye. Generic-to-specific distillation of masked autoencoders. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2, 3, 5", + "[22] Feng Liang, Yangguang Li, and Diana Marculescu. Supmae: Supervised masked autoencoders are efficient vision learners. arXiv preprint arXiv:2205.14540, 2022. 6, 7", + "[23] Jihao Liu, Xin Huang, Jinliang Zheng, Yu Liu, and Hongsheng Li. Mixmae: Mixed and masked autoencoder for efficient pretraining of hierarchical vision transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 3", + "[24] Jian Meng, Li Yang, Kyungmin Lee, Jinwoo Shin, Deliang Fan, and Jae sun Seo. Slimmed asymmetrical contrastive learning and cross distillation for lightweight model training. In Thirty-seventh Conference on Neural Information Processing Systems (NeurIPS), 2023. 2" + ], + "bbox": [ + 516, + 92, + 906, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "25238", + "bbox": [ + 478, + 926, + 519, + 936 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[25] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation Learning with Contrastive Predictive Coding. arXiv preprint arXiv:1807.03748, 2018. 2", + "[26] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning (ICML), 2021. 1", + "[27] Bowen Shi, Xiaopeng Zhang, Yaoming Wang, Jin Li, Wenrui Dai, Junni Zou, Hongkai Xiong, and Qi Tian. Hybrid distillation: Connecting masked autoencoders with contrastive learners. In The Twelfth International Conference on Learning Representations (ICLR), 2024. 3, 6", + "[28] Yucheng Tang, Dong Yang, Wenqi Li, Holger R Roth, Bennett Landman, Daguang Xu, Vishwesh Nath, and Ali Hatamizadeh. Self-supervised pre-training of swim transformers for 3d medical image analysis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1", + "[29] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In International Conference on Machine Learning (ICML), 2021. 2", + "[30] Jue Wang, Gedas Bertasius, Du Tran, and Lorenzo Torresani. Long-short temporal contrastive learning of video transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (NeurIPS), 2022. 1", + "[31] Shaoru Wang, Jin Gao, Zeming Li, Xiaoqin Zhang, and Weiming Hu. A closer look at self-supervised lightweight vision transformers. In International Conference on Machine Learning (ICML). PMLR, 2023. 2, 3", + "[32] Chen Wei, Haoqi Fan, Saining Xie, Chao-Yuan Wu, Alan Yuille, and Christoph Feichtenhofer. Masked feature prediction for self-supervised visual pre-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 8", + "[33] Kan Wu, Houwen Peng, Zhenghong Zhou, Bin Xiao, Mengchen Liu, Lu Yuan, Hong Xuan, Michael Valenzuela, Xi Stephen Chen, Xinggang Wang, et al. Tinyclip: Clip distillation via affinity mimicking and weight inheritance. In Proceedings of the IEEE/CVF International Conference on Computer Vision (CVPR), 2023. 1", + "[34] Zhenda Xie, Zheng Zhang, Yue Cao, Yutong Lin, Jianmin Bao, Zhuliang Yao, Qi Dai, and Han Hu. Simmim: A simple framework for masked image modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 7", + "[35] Sukmin Yun, Hankook Lee, Jaehung Kim, and Jinwoo Shin. Patch-level representation learning for self-supervised vision transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 5", + "[36] Jure Zbontar, Li Jing, Ishan Misra, Yann LeCun, and Stéphane Deny. Barlow Twins: Self-supervised Learning via Redundancy Reduction. In International Conference on Machine Learning (ICML), pages 12310-12320, 2021. 1, 2, 4, 5" + ], + "bbox": [ + 91, + 92, + 483, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[37] Qi Zhang, Yifei Wang, and Yisen Wang. How mask matters: Towards theoretical understandings of masked autoencoders. Advances in Neural Information Processing Systems (NeurIPS), 2022. 1, 3, 4, 5, 7, 8", + "[38] Zhiyu Zhao, Bingkun Huang, Sen Xing, Gangshan Wu, Yu Qiao, and Limin Wang. Asymmetric masked distillation for pre-training small foundation models. arXiv preprint arXiv:2311.03149, 2023. 2" + ], + "bbox": [ + 516, + 92, + 906, + 203 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "25239", + "bbox": [ + 478, + 926, + 517, + 936 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Closest Neighbors are Harmful for Lightweight Masked Auto-encoders/ce070f35-5bfa-41af-9465-b1f84c742005_model.json b/2025/Closest Neighbors are Harmful for Lightweight Masked Auto-encoders/ce070f35-5bfa-41af-9465-b1f84c742005_model.json new file mode 100644 index 0000000000000000000000000000000000000000..e96c26372cbb96f1fad0ae7748af6f0f5e28eadf --- /dev/null +++ b/2025/Closest Neighbors are Harmful for Lightweight Masked Auto-encoders/ce070f35-5bfa-41af-9465-b1f84c742005_model.json @@ -0,0 +1,1947 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.141, + 0.131, + 0.856, + 0.152 + ], + "angle": 0, + "content": "Closest Neighbors are Harmful for Lightweight Masked Auto-encoders" + }, + { + "type": "text", + "bbox": [ + 0.1, + 0.18, + 0.932, + 0.234 + ], + "angle": 0, + "content": "\\(^{1}\\)Jian Meng, \\(^{1}\\)Ahmed Hassssan, \\(^{2}\\)Li Yang, \\(^{3}\\)Deliang Fan, \\(^{4}\\)Jinwoo Shin, and \\(^{1}\\)Jae-sun Seo \n\\(^{1}\\)Cornell University, \\(^{2}\\)University of North Carolina at Charlotte, \\(^{3}\\)Arizona State University, \\(^{4}\\)KAIST \n\\(^{1}\\{jm2787, ah2288, js3528\\} @cornell.edu, ^{2}lyang50@uncc.edu, ^{3}dfan12@asu.edu, ^{4}jinwoos@kaist.ac.kr\\)" + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.269, + 0.327, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.301, + 0.486, + 0.754 + ], + "angle": 0, + "content": "Learning the visual representation via masked autoencoder (MAE) training has been proven to be a powerful technique. Transferring the pre-trained vision transformer (ViT) to downstream tasks leads to superior performance compared to conventional task-by-task supervised learning. Recent research works on MAE focus on large-sized vision transformers (\\(>50\\) million parameters) with outstanding performance. However, improving the generality of the under-parametrized lightweight model has been widely ignored. In practice, downstream applications are commonly intended for resource-constrained platforms, where large-scale ViT cannot easily meet the resource budget. Current lightweight MAE training heavily relies on knowledge distillation with a pre-trained teacher, whereas the root cause behind the poor performance remains underexplored. Motivated by that, this paper first introduces the concept of \"closest neighbor patch\" to characterize the local semantics among the input tokens. Our discovery shows that the lightweight model failed to distinguish different local information, leading to biased understanding and poor accuracy. Motivated by this finding, we propose NoR-MAE, a novel MAE training algorithm for lightweight vision transformers. NoR-MAE elegantly repels the semantic aliasing between patches and their closest neighboring patch (semantic centroid) with negligible training cost overhead. With the ViT-Tiny model, NoR-MAE achieves up to \\(7.22\\% / 3.64\\%\\) accuracy improvements on ImageNet100/ImageNet-1K datasets, as well as up to \\(5.13\\%\\) accuracy improvements in tested downstream tasks. https://github.com/SeoLabCornell/NoR-MAE" + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.784, + 0.222, + 0.8 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.485, + 0.901 + ], + "angle": 0, + "content": "Starting from the early exploration with contrastive learning [6, 18], learning powerful and generic vision or semantic representation has been the focal point of all prior self-supervised learning (SSL) algorithms across various machine learning domains [2, 8, 28, 30] and vision-language model [15, 26]. The superiority of SSL is demonstrated" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.266, + 0.907, + 0.453 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.463, + 0.909, + 0.547 + ], + "angle": 0, + "content": "Figure 1. Top: Closest Neighbor Patches (CNP) characterizes the \"Local Semantic Centroid\" among the input tokens. Bottom: Low-dimensional projection (via PCA) of the transformer-encoded patches. Unlike the large-sized transformer model (e.g., ViT-Base), lightweight models (e.g., ViT-Tiny) cannot separate the \"semantic difference\" between different CNPs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.584, + 0.909, + 0.903 + ], + "angle": 0, + "content": "by the capability of extracting transferable knowledge from unlabeled datasets [26, 33]. Recent efforts on masked autoencoder (MAE) [19] train the vision transformer [11] by encouraging the model to understand the semantic knowledge of the masked patches throughout the reconstruction-based self-supervised learning. Particularly, the evolution from contrastive learning [16, 18, 36] to MAE-based training [17, 19, 34, 37] shows consistent dominance by large-scale models to ensure on-par or even better accuracy compared to supervised learning, along with superior performance in various downstream applications. In practice, downstream small-scale tasks are commonly deployed on edge devices with limited resources. Therefore, large-sized vision transformers (ViTs) pre-trained by MAE [19] are suboptimal candidates for resource-constrained downstream deployment and hardware platforms. Unfortunately, recent state-of-the-art (SoTA) methods that train lightweight ViTs with MAE from scratch lead to poor performance and insufficient transferability. For example, the MAE-trained lightweight ViT-Tiny model only achieves \\(66.6\\%\\) top-1 accuracy (with 200-epoch pre-training [3, 19, 37]) on the" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.925, + 0.521, + 0.937 + ], + "angle": 0, + "content": "25230" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.198 + ], + "angle": 0, + "content": "ImageNet-1K dataset [3], which is \\(\\sim 10\\%\\) lower than the supervised learning counterpart. Evidently, the superiority of MAE [19] with the large model failed to be maintained in the lightweight ViTs [11]. Given the diverse downstream tasks and urgent need for energy-efficient foundation models, unleashing the power of MAE and enhancing the versatility of lightweight models is desired." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.199, + 0.487, + 0.515 + ], + "angle": 0, + "content": "Prior works enhance the performance of the lightweight transformer models by heavily relying on knowledge distillation (KD) [20] for both supervised [29] and self-supervised learning paradigms [3, 21, 38]. For instance, the recent DMAE work [3] combines MAE [19] with KD and transfers the pre-trained knowledge from a ViT-Base teacher down to the lightweight ViT model. DMAE [3] chooses the 3/4 depth of both student and teacher ViT models as the bonding portal for feature alignment and knowledge distillation. The heuristic design of DMAE cannot be easily extended to other encoder models. Furthermore, distilling the knowledge on the fly introduces an additional projector [3] or logits mask [31], which further elevates the complexity and computation cost of the entire training process. In addition to DMAE, other recent works further increase the complexity of the KD-based MAE by employing multi-layer distillation [38] or complex generic-to-specific distillation [21]. All of these prior works [3, 21, 31, 38] rely on the brute force approach to enhance the model performance with increasingly complex training schemes, which results in inflated GPU memory and training time." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.516, + 0.485, + 0.607 + ], + "angle": 0, + "content": "However, none of these prior works can collectively achieve 1) the capability of training lightweight model from scratch without a large-sized teacher model, 2) negligible training cost overhead compared to the vanilla MAE, and 3) clear insights that cause the poor performance of lightweight MAE. To overcome this, the key questions to address are:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.61, + 0.483, + 0.639 + ], + "angle": 0, + "content": "① What is the bottleneck of MAE with under parametrized small vision transformer models?" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.64, + 0.483, + 0.669 + ], + "angle": 0, + "content": "② How to efficiently train a high performance lightweight model from scratch via MAE?" + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.61, + 0.483, + 0.669 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.674, + 0.483, + 0.853 + ], + "angle": 0, + "content": "To answer ①, our investigation shows that the lightweight model failed to understand different local semantic information among the input tokens and lead to poor performance. We characterize different local semantics by introducing the concept of closest neighborhood patch (CNP) (Figure 1), which carries the most similar semantics of a given input patch. Since multiple patches can share one CNP (e.g., Background Forrest), different CNPs represent various local semantics across the entire input sequence. Unlike large-sized models, the under-parameterized model falls into the semantic aliasing between different CNPs with poor accuracy, as indicated in Figure 1 and Section 3.3." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.485, + 0.901 + ], + "angle": 0, + "content": "Motivated by the findings in ①, we propose the Neighbor Repelling Masked Autoencoder (NoR-MAE) to resolve ②, a novel self-supervised learning algorithm designed for train" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.908, + 0.257 + ], + "angle": 0, + "content": "ing powerful lightweight vision transformers. NoR-MAE trains the powerful small vision transformer by penalizing the similarity between the unmasked patches and CNPs. Different from prior works, NoR-MAE directly trains the vision transformer from scratch without introducing the largesized pre-trained teacher or heuristically designed distillation scheme. The proposed method achieves up to \\(7.22\\%\\) and \\(3.64\\%\\) accuracy improvements on ImageNet-100 and ImageNet-1K datasets, with negligible training time and memory overhead compared to the vanilla MAE. The major contributions of our work are:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.258, + 0.907, + 0.302 + ], + "angle": 0, + "content": "- Simplicity: NoR-MAE does not require teacher pretraining or heuristic knowledge distillation to train a powerful lightweight ViT." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.303, + 0.907, + 0.378 + ], + "angle": 0, + "content": "- High Performance: Compared to the baseline MAE, NoR-MAE achieves \\(7.22\\%\\) and \\(3.64\\%\\) accuracy improvements on ImageNet-100 and ImageNet-1K datasets with the ViT-Tiny model, achieving the new SoTA performance on the lightweight MAE training." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.379, + 0.907, + 0.468 + ], + "angle": 0, + "content": "- Transferrability: Different from prior works, which mainly focus on fine-tuning the performance on the pretraining dataset (e.g., ImageNet), NoR-MAE is also evaluated with the downstream fine-tuning and linear evaluation benchmarks, with significantly improved performance compared to prior SSL baselines and the vanilla MAE." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.47, + 0.905, + 0.545 + ], + "angle": 0, + "content": "- Rationality: Besides the superior performance, for the first time, NoR-MAE reveals the semantic aliasing issue caused by the neighboring patches, which is proved to be the reason that hinders efficient training and representation learning of lightweight ViT models." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.258, + 0.907, + 0.545 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.559, + 0.655, + 0.573 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.584, + 0.908, + 0.765 + ], + "angle": 0, + "content": "Empowering a deep neural network model with strong and transferable knowledge has been widely investigated, especially in self-supervised learning. Early research works in contrastive learning (CL) introduce the learning paradigm with \"positive\" and \"negative\" sample pairs [6, 18], together with the alignment-repellent strategy embedded into InfoNCE [7, 18, 25] and NT-Xent loss [6] as the learning objective. The early success of CL motivated the successors to investigate the potential of different training and sampling strategies, inducing the standpoint of asymmetrical learning [16, 24], knowledge distillation [13], and the latent-dimension correlation [36]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.766, + 0.909, + 0.902 + ], + "angle": 0, + "content": "While the core idea of CL is training the model to understand the similarities and differences between different augmentations, the emergence of masked autoencoder (MAE) [19] training reconsiders SSL from a different perspective, which is perfectly suitable for vision transformers. Encouraging the encoder-decoder to reconstruct the randomly sparsified patches allows the model to understand the semantics. Due to the powerful visual representation, MAE-based pretaining exhibits strong performance on" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.925, + 0.518, + 0.937 + ], + "angle": 0, + "content": "25231" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.098, + 0.09, + 0.478, + 0.171 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.105, + 0.184, + 0.468, + 0.199 + ], + "angle": 0, + "content": "Figure 2. Closest Neighbor Patches (CNP) within the image." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.234, + 0.484, + 0.31 + ], + "angle": 0, + "content": "both backbone datasets (e.g., ImageNet-1K) and downstream transfer learning tasks. Subsequent research works further improved the performance of MAE from the perspectives of contrastive learning [27, 37], cross-attention [14, 17], mixed feature learning [23], or intricate tokenizer [12]." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.315, + 0.485, + 0.754 + ], + "angle": 0, + "content": "Across all the different MAE training methods, elevating the performance of the large-sized models (e.g., ViT-Large) has been the primary focus of exploration, while improving the poor performance of the small-sized model (e.g., ViTTiny) has been largely ignored. Given that the downstream tasks are well-suited for resource-constrained edge devices, the imbalanced research focus of the prior works largely hinders the efficiency and transferability of MAE [19]. The practical needs and the imperfection of lightweight MAE motivate the recent research works to investigate MAE training strategies with small models. Following the protocol from supervised distillation [20], knowledge distillation-based MAE [3, 21] is proposed to enhance the performance of the lightweight student. However, the learning scheme of the single-layer or multi-layer distillation is established based on the heuristically designed distillation between intermediate features. Compared to the supervised knowledge distillation, recent studies largely complicate the overall cost of training lightweight ViT models with either heuristic design [3], multi-phased distillation [21], or distillation with extensive fine-tuning [31]. In particular, employing the pre-trained large ViT model as the teacher introduces additional GPU memory usage and time cost toward the total training efforts. Furthermore, training the interconnect projectors between the student and teacher exacerbates the training complexity even further. More importantly, the knowledge distillation-based MAE inherits the essence of KD from supervised learning, but the insights of training lightweight ViT from scratch via MAE remain largely unexplored." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.78, + 0.265, + 0.797 + ], + "angle": 0, + "content": "3. Proposed Method" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.484, + 0.902 + ], + "angle": 0, + "content": "We propose the Neighbor-Repelling MAE (NoR-MAE). In this section, we will 1) introduce the concept of the Closest Neighbor Patch into MAE training, 2) present the proposed NoR-MAE training, and 3) unravel the rationality of NoR-MAE by revealing the insights that cause the poor performance of the directly-trained lightweight ViT model." + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.09, + 0.905, + 0.262 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.273, + 0.908, + 0.303 + ], + "angle": 0, + "content": "Figure 3. Cosine Similarity between each patch and their Closest Neighbor Patch across a subset of ImageNet-1K with 4096 images." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.331, + 0.783, + 0.348 + ], + "angle": 0, + "content": "3.1. Closest Neighbor Patch (CNP)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.356, + 0.908, + 0.478 + ], + "angle": 0, + "content": "Let \\( X \\in \\mathbf{R}^{H \\times W \\times C} \\) be an input image to the model, where \\( H \\), \\( W \\), and \\( C \\) represent the height, width, and channel of the image. Vision transformer (ViT) reformulates the input by decomposing the image \\( X \\) into a sequence of non-overlapped patches with the pre-defined patch size \\( P \\times P \\) and the total number of patches (tokens) \\( K \\). All the patches are further encoded into the embeddings with dimension \\( D \\), which is the input of the subsequent transformer blocks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.48, + 0.909, + 0.571 + ], + "angle": 0, + "content": "For a given input patch \\( X_P \\), there are naturally 8 neighbors that share a common edge with \\( X_P \\), as shown in Figure 2(a). We define the Closest Neighbor Patch (CNP) \\( X_{CNP} \\) by selecting the neighbor that has the Top-1 cosine similarity with \\( X_P \\). In other words, \\( X_P \\) and \\( X_{CNP} \\) are sharing the highest degree of semantically meaningful relations." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.573, + 0.909, + 0.687 + ], + "angle": 0, + "content": "Shared Closest Neighbor Patch. In the example of Figure 2, two patches \\( X_P^i \\) and \\( X_P^j \\) partially share the neighborhoods in between, while the CNP of \\( X_P^i \\) and \\( X_P^j \\) happen to be identical as \\( X_{CNP}^{ij} \\). Naturally, \\( X_{CNP}^{ij} \\) carries the shared semantic information (\"Background Forest\") for both \\( X_P^i \\) and \\( X_P^j \\), but the degree of similarity between \\( (X_P^i, X_{CNP}^{ij}) \\) and \\( (X_P^j, X_{CNP}^{ij}) \\) are different." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.688, + 0.909, + 0.808 + ], + "angle": 0, + "content": "Closest Neighbor Patch \\(\\neq\\) Identical Patch. Although \\(X_{P}\\) and \\(X_{CNP}\\) are sharing the highest degree of similarity, the semantics between \\(X_{P}\\) and \\(X_{CNP}\\) are not identical. As shown in Figure 3, the cosine similarity between all the patches and their closest neighbor in ImageNet-1K images varies from 0.07 to 0.999, with an average of 0.83. In other words, the highly similar neighboring patches also contain the implicit semantical difference." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Difficulty of Lightweight Models to Understand the Different Semantics Carried by CNP. With MAE, the input patches of the transformer models are randomly masked with a pre-defined sparsity ratio \\(\\tau\\) (e.g., \\(75\\%\\)), as illustrated in Figure 4, left. Naturally, each unmasked visible patch has a Closest Neighbor Patch (CNP). As a result, the CNP" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.926, + 0.521, + 0.938 + ], + "angle": 0, + "content": "25232" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.09, + 0.482, + 0.202 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.215, + 0.483, + 0.245 + ], + "angle": 0, + "content": "Figure 4. The shared Closest Neighbor Patch (CNP) characterizes the estimated local semantic centroids of the input sequence." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.272, + 0.484, + 0.468 + ], + "angle": 0, + "content": "of the unmasked patches formulates another sparse input image (Figure 4, right). From the perspective of the entire image, visible patches and their CNPs are close neighbors, while different visible patches may share the same closest neighbor, which estimates the local and regional semantics of the input sequence. However, as described in Figure 1 and Section 1, understanding the semantic difference between different CNPs is the bottleneck for the under-parameterized models. Furthermore, our extended experiments in Figure 7 show that the key difference between the big (e.g., ViT-Base) and small (e.g., ViT-Tiny) model is understanding and separating the semantic differences between CNPs. Therefore, the challenge of lightweight MAE is clear:" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.469, + 0.484, + 0.501 + ], + "angle": 0, + "content": "In MAE training, how to facilitate the lightweight model to understand different local semantics?" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.511, + 0.36, + 0.527 + ], + "angle": 0, + "content": "3.2. Proposed Method: NoR-MAE" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.536, + 0.484, + 0.656 + ], + "angle": 0, + "content": "We propose the Neighbor-Repelling training for self-supervised MAE learning (NoR-MAE). The overall objective is to train a powerful, lightweight vision transformer by encouraging the model to avoid the biased relation between different local semantics of the input. We follow the standard protocol of the vanilla MAE with the encoder \\( f_{\\theta} \\) and decoder \\( g_{\\phi} \\), where both \\( f_{\\theta} \\) and \\( g_{\\phi} \\) are vision transformers with the same embedding dimension \\( D \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.658, + 0.484, + 0.794 + ], + "angle": 0, + "content": "Given the masking ratio \\(\\tau\\), we first generate the masked input sequence \\(x^{\\tau}\\) and \\(x_{CNP}^{\\tau}\\) that are embedded from the unmasked patches with the corresponding closest neighbor patches, respectively. During the forward pass of each iteration, we parallelly encode and decode the masked sequence \\(x^{\\tau}\\) and CNP sequence \\(x_{CNP}^{\\tau}\\) by \\(f_{\\theta}\\) and \\(g_{\\phi}\\), generating the reconstructed embedding \\(Z\\) and \\(Z_{CNP}\\), respectively. In particular, \\(f_{\\theta}\\) and \\(g_{\\phi}\\) are served as the siamese encoder and decoder (with stop gradient) during the forward pass of \\(Z_{CNP}\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.484, + 0.825 + ], + "angle": 0, + "content": "Overall, each forward pass will generate two pairs of decoded tokens resulting from \\( x^{\\tau} \\) and \\( x_{\\mathrm{CNP}}^{\\tau} \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Following the standard training protocol of MAE, we first compute the \\(\\mathcal{L}_2\\) reconstruction loss between the decoded sequence \\(Z\\) and the ground truth. Subsequently, we compute the matrix multiplication between the normalized \\(Z\\) and \\(Z_{CNP}\\). As shown in Figure 5, the encoder \\(f_{\\theta}\\) and decoder" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.183 + ], + "angle": 0, + "content": "\\(g_{\\phi}\\) are shared for both branches, while the stop gradient is enabled for the CNP input. Given the size of \\([N,K,D]\\), the matrix multiplication between \\(\\bar{Z}\\) and transposed \\(\\bar{Z}_{\\mathrm{CNP}}\\) is performed along the token dimension \\(K\\). The resultant matrix is further scaled down by \\(K\\) and takes the average along the batch dimension \\(N\\). Mathematically, we have:" + }, + { + "type": "equation", + "bbox": [ + 0.585, + 0.192, + 0.907, + 0.228 + ], + "angle": 0, + "content": "\\[\nC _ {\\text {n e i g h b o r}} = \\frac {1}{N} \\sum_ {N} \\left(\\frac {1}{K} (\\bar {Z}) ^ {T} \\cdot \\bar {Z} _ {\\mathrm {C N P}}\\right) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.233, + 0.907, + 0.355 + ], + "angle": 0, + "content": "Where \\( C_{\\mathrm{neighbor}} \\) represents the resultant relation matrix with the size of \\( D \\times D \\). Each entry \\( c_{i,j} \\) of \\( C_{\\mathrm{neighbor}} \\) characterizes the correlation between the randomly-unmasked input sequence (output logits \\( Z \\)) and the corresponding CNP sequence (output logits \\( Z_{\\mathrm{CNP}} \\)) between embedding dimension \\( i \\) and \\( j \\). To minimize the semantic aliasing between the unmasked tokens and their CNPs, we decorate different embeddings by penalizing the off-diagonal terms of \\( C_{\\mathrm{neighbor}} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.626, + 0.366, + 0.907, + 0.4 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {N o R}} = \\sum_ {i} \\sum_ {i \\neq j} C _ {\\text {n e i g h b o r}} ^ {i, j} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.409, + 0.909, + 0.547 + ], + "angle": 0, + "content": "Essentially, the proposed NoR-Loss \\((\\mathcal{L}_{\\mathrm{NoR}})\\) minimizes semantic aliasing by preventing the model encoding the visible patches and CNP as aliased embeddings. It has been shown in Figure 3 that patches and their CNPs share similar but non-identical semantics. We balance such a tradeoff of \"similarity-discrepancy\" by decorrelating the off-diagonal embeddings only. The overall loss is the combination of the standard \\(\\mathcal{L}_2\\) reconstruction loss and the proposed NeighborRepelling Loss \\(\\mathcal{L}_{\\mathrm{NoR}}\\) scaled by a hyperparameter \\(\\lambda\\):" + }, + { + "type": "equation", + "bbox": [ + 0.527, + 0.555, + 0.907, + 0.574 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {t o t a l}} = \\left| \\left| \\text {G r o u n d T r u t h} - \\bar {Z} \\right| \\right| _ {2} + \\lambda \\cdot \\mathcal {L} _ {\\mathrm {N o R}} (\\bar {Z}, \\bar {Z} _ {\\mathrm {C N P}}) \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.584, + 0.909, + 0.689 + ], + "angle": 0, + "content": "More importantly, decorrelating embeddings eventually separate the aliased semantics between different CNPs and their corresponding \"local semantic centroids\" (Section 3.3). In other words, the embedding-level anti-aliasing regularization of the proposed NoR-Loss facilitates the understanding on the representation (token) level, leading to the largely improved model performance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.69, + 0.909, + 0.825 + ], + "angle": 0, + "content": "Decorrelating Embeddings vs. Decorrelating Tokens. Although the proposed NoR-MAE selects the Top-1 neighbor (the closest neighbor patch (CNP)) as the candidate for decorrelation, the CNP patches can be overlapped or shared between each other. For instance, the closest neighbor patch of patch \\( X_{i} \\) could also be the Top-2/3/4 ... 8 neighbor of patch \\( X_{j} \\). Minimizing the semantic aliasing by naively decorrelating tokens leads to collapsed performance and unsuccessful training, as shown in Figure 6." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.909, + 0.901 + ], + "angle": 0, + "content": "NoR-MAE vs. Contrastive Learning. Empowering representation learning via invariance-covariance optimization has been investigated in Barlow Twins [36]. Recently proposed U-MAE [37] combines the contrastive learning and MAE [19] by introducing the random asymmetry masks" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.925, + 0.519, + 0.937 + ], + "angle": 0, + "content": "25233" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.156, + 0.089, + 0.849, + 0.316 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.324, + 0.326, + 0.673, + 0.34 + ], + "angle": 0, + "content": "Figure 5. Overview of the proposed NoR-MAE algorithm." + }, + { + "type": "image", + "bbox": [ + 0.114, + 0.363, + 0.462, + 0.487 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.1, + 0.498, + 0.473, + 0.512 + ], + "angle": 0, + "content": "Figure 6. Different relation penalty of the NoR-Loss \\((\\lambda = 5\\mathrm{e} - 5)\\)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.539, + 0.484, + 0.629 + ], + "angle": 0, + "content": "during the forward pass of the training. On the contrary, the proposed NoR-MAE exploits the semantic aliasing issue by constructing the asymmetry between the visible patches and the corresponding closest neighbors, instead of relying on the randomness. As a result, NoR-MAE outperforms U-MAE [37] in various sizes of models (Section 4)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.63, + 0.486, + 0.78 + ], + "angle": 0, + "content": "In the context of contrastive learning, the similar samples (patches) are considered as \"positive samples\" which are intended to be aligned by the algorithms [35]. However, the important discovery of NoR-MAE points out that semantic aliasing among the closest neighbors is harmful for the lightweight masked autoencoder training, which is not properly addressed by prior works. Moreover, unlike Barlow Twins [36], NoR-MAE does not need the embedding alignment along the diagonal terms due to the well-established similarity between the visible patches and their CNPs." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.486, + 0.902 + ], + "angle": 0, + "content": "NoR-MAE Does Not Need a Teacher Model. We would like to highlight that the proposed NoR-MAE is entirely different from knowledge distillation-based MAE. NoR-MAE does not require a pre-trained teacher [3] or multistage distillation [21] during training. The lightweight encoder is trained from scratch following the standard MAE protocol [19]. Although the CNPs and the original unmasked patches are separately generated, the encoding-decoding pro" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.367, + 0.908, + 0.443 + ], + "angle": 0, + "content": "cess remains the same as the vanilla MAE. For example, given the \\(75\\%\\) mask ratio with \\(25\\%\\) visible patches, the equivalent input sparsity of NoR-MAE is \\(50\\%\\). The overhead of the NoR-MAE is minimal compared to the vanilla MAE, as profiled in the supplementary." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.453, + 0.744, + 0.469 + ], + "angle": 0, + "content": "3.3. Rationality of NoR-MAE" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.477, + 0.908, + 0.567 + ], + "angle": 0, + "content": "As shown in Figure 2 and Figure 4, the proposed concept of Closest Neighbor Patch (CNP) characterizes the local semantics of the image. For instance, the patches that contain \"Background Forest\" or \"Back of the Fox\" are represented by separate CNP patches. However, the lightweight model failed to learn the semantic difference between CNPs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.568, + 0.909, + 0.825 + ], + "angle": 0, + "content": "We prove the significance of the CNP semantics by analyzing the original unmasked features and CNP features encoded by the trained encoder \\( f_{\\theta} \\). Specifically, we first normalize the encoded features and then compress the feature dimension down to 2-D via the Principal Component Analysis (PCA) [1]. As shown in Figure 7 and Figure 8, the lightweight models (ViT-Tiny and ViT-Small) trained by the vanilla MAE [19] failed to distinguish the semantical difference between different closest neighbor patches (CNP, blue dots), regardless of the patch sizes. Different semantics of CNP are heavily mixed together in the low dimensional space. Moreover, CNP is critical for MAE training regardless of the input patch sizes and the corresponding sequence length. The distorted and aliased understanding of the local semantics leads to the sub-optimal performance of MAE, as shown in Table 1 with the empirical verification on ImageNet-100 (\\( 224 \\times 224 \\)) with ViT-Tiny." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.909, + 0.901 + ], + "angle": 0, + "content": "On the contrary, the model trained by NoR-MAE successfully understands the semantic difference between different CNP features, leading to largely improved accuracy on the lightweight ViT models across different patch sizes. In the meantime, the large-sized model (e.g., ViT-Base) exhibits" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.926, + 0.521, + 0.938 + ], + "angle": 0, + "content": "25234" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.128, + 0.092, + 0.374, + 0.113 + ], + "angle": 0, + "content": "ViT-Tiny (# of params = 5.7 Million) MAE [19]: Aliased Semantics! NoR-MAE: Clearly Separated CNP!" + }, + { + "type": "image", + "bbox": [ + 0.126, + 0.115, + 0.235, + 0.184 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.134, + 0.182, + 0.218, + 0.192 + ], + "angle": 0, + "content": "Accuracy \\(= 66.60\\)" + }, + { + "type": "image", + "bbox": [ + 0.25, + 0.115, + 0.361, + 0.183 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.258, + 0.182, + 0.347, + 0.192 + ], + "angle": 0, + "content": "Accuracy \\(= 70.24\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.267, + 0.195, + 0.432, + 0.204 + ], + "angle": 0, + "content": "Latent Feature of Unmasked Patches" + }, + { + "type": "image", + "bbox": [ + 0.382, + 0.092, + 0.631, + 0.193 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.453, + 0.195, + 0.645, + 0.205 + ], + "angle": 0, + "content": "Latent Feature of Closest Neighbor Patches" + }, + { + "type": "image", + "bbox": [ + 0.644, + 0.092, + 0.89, + 0.193 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.665, + 0.195, + 0.832, + 0.205 + ], + "angle": 0, + "content": "Patch size \\(= 16^{2}\\) Token length \\(= 196\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.22, + 0.907, + 0.248 + ], + "angle": 0, + "content": "Figure 7. Low dimensional features of the unmasked patches (red dots) and the closest neighbor patches (CNP) (blue dots) with the same input from ImageNet-1K dataset. NoR-MAE clearly separates the semantic information among different CNP, with improved accuracy." + }, + { + "type": "image", + "bbox": [ + 0.108, + 0.275, + 0.473, + 0.387 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.403, + 0.483, + 0.445 + ], + "angle": 0, + "content": "Figure 8. Low dimensional features of the unmasked patches (red dots) and the CNP (blue dots) encoded by ViT-Tiny with patch size \\(= 4 \\times 4\\) (sequence length \\(= 3136\\)), evaluated on ImageNet-100." + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.46, + 0.483, + 0.489 + ], + "angle": 0, + "content": "Table 1. NoR-MAE consistently outperforms the MAE [19] across different input patch sizes (evaluated on the ImageNet-100 dataset)." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.499, + 0.485, + 0.599 + ], + "angle": 0, + "content": "
Masking RatioInput Patch SizeMAE [19]NoR-MAE (This work)
75%2²71.08%83.56% (+12.48%)
75%4²71.16%83.58% (+12.42%)
75%8²71.27%79.24% (+7.97%)
75%16²71.04%78.26% (+7.22%)
75%32²68.42%74.73% (+6.31%)
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.625, + 0.483, + 0.655 + ], + "angle": 0, + "content": "a better understanding and stronger separation of different semantics with the vanilla MAE [19]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.656, + 0.484, + 0.747 + ], + "angle": 0, + "content": "In summary, the proposed NoR-MAE successfully mitigates the performance gap between model sizes, facilitating the learnability of the transformer model with a stronger understanding of the semantical relations from the input, leading to improved accuracy with both lightweight and large-sized transformer models." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.762, + 0.295, + 0.779 + ], + "angle": 0, + "content": "4. Experimental Results" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.787, + 0.31, + 0.803 + ], + "angle": 0, + "content": "4.1. ImageNet Classification" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.484, + 0.901 + ], + "angle": 0, + "content": "For image classification tasks, we follow the standard protocol of MAE [19] by training the lightweight encoder on the ImageNet-1K and ImageNet-100 datasets with the batch size of 4,096 and 2,048, and subsequently performing the standard fine-tuning [19]. Specifically, the model is pre-trained by 200 and 400 epochs with the initial learning rate set to" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.275, + 0.907, + 0.411 + ], + "angle": 0, + "content": "1.5e-4, together with the weight decay value of 0.05. The penalty coefficient \\(\\lambda\\) of the proposed NoR Loss is set to 5e-5, and the optimality of the selected penalty level is demonstrated in Table 5. The mask ratio is fixed at \\(75\\%\\) based on the findings in the vanilla MAE [19]. We implement the standard fine-tuning protocol with 100 epochs as the vanilla MAE [19]. The detailed experimental setup (pre-train and downstream) are summarized in the supplementary, together with the report of GPU memory usage and training cost." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.412, + 0.909, + 0.548 + ], + "angle": 0, + "content": "Table 2 and Table 3 summarize the performance of the NoR-MAE on ImageNet-100 and ImageNet-1K datasets. Compared to the vanilla MAE [19], NoR-MAE largely improves the performance on top of the MAE baseline [19] with up to \\(7.22\\%\\) and \\(3.64\\%\\) accuracy on ImageNet-100 and ImageNet-1K datasets, respectively. For the large-sized vision transformer training (e.g., ViT-Base), NoR-MAE still maintains the superiority compared to the recent high-performance MAE training (Table 3)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.55, + 0.909, + 0.701 + ], + "angle": 0, + "content": "More importantly, NoR-MAE improves the model performance marginal training cost overhead. NoR-MAE trains the model with the standard MAE protocol without introducing 1) a pre-trained teacher [3] or 2) other additional architectures [27] or clustering modules [12]. As profiled in the supplementary, the performance boost reward by the NoR-MAE only exhibits up to \\(8\\%\\) memory overhead and \\(2\\%\\) additional training time. Compared to the distillation-based MAE [3], NoR-MAE outperforms DMAE [3] with \\(\\sim 1\\%\\) accuracy improvement and \\(50\\%\\) less training cost." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.702, + 0.909, + 0.778 + ], + "angle": 0, + "content": "We further validate the NoR-MAE on the larger ViT-Base model. The proposed algorithm consistently achieves better performance compared to the MAE baseline and the recent ImageNet-100 SoTA method without introducing any additional clustering efforts or special tokenizer [12]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.779, + 0.909, + 0.854 + ], + "angle": 0, + "content": "Finally, NoR-MAE is also suitable for the MAE with supervision (SupMAE) [22]. Introducing supervised learning into MAE does not impact the effectiveness of the proposed NoR-MAE, while the NoR-MAE further enhances the performance of SupMAE [22], as shown in Table 2." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Comparison with the Distillation-based MAE. We would like to highlight that the proposed NoR-MAE can achieve on-par or even better performance compared to the" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.926, + 0.518, + 0.938 + ], + "angle": 0, + "content": "25235" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.908, + 0.119 + ], + "angle": 0, + "content": "Table 2. Fine-tuning and semi-supervised learning accuracy (%) of the ViT encoder pre-trained by MAE [19] and recent SoTA lightweight MAE on ImageNet-1K dataset (with 200 epochs and 400 epochs pre-training)." + }, + { + "type": "table", + "bbox": [ + 0.131, + 0.127, + 0.865, + 0.646 + ], + "angle": 0, + "content": "
ViT Model# of ParamsMethodTeacher ModelPre-training EpochSupervised Fine-tuning Accuracy (%)
ViT-Tiny5.7MMAE [19]-20066.60
MAE [19]-40067.63
Sup MAE [22]-20067.88
Sup MAE [22]-40068.91
MoCo-V3 [18]-20070.09
DINO [5]-30068.17
SimMIM [34]-20066.08
DMAE [3]ViT-B20070.00
Sup NoR-MAE (This work)-20070.63 (+2.75)
NoR-MAE (This work)-20070.24 (+3.64)
NoR-MAE (This work)-40071.18 (+3.55)
ViT-Small21.7MMAE [19]-20079.00
MAE [19]-40080.11
Sup MAE [22]-20079.27
CrossMAE [14]-40079.30
MoCo-V3 [18]-20079.46
DMAE [3]ViT-B20079.30
Sup NoR-MAE (This work)-20080.16 (+0.89)
NoR-MAE (This work)-20080.13 (+1.03)
NoR-MAE (This work)-40081.04 (+0.93)
ViT-Base88.6MMAE [19]-20083.30
Sup MAE [22]-20083.60
DMAE [3]-20084.00
CrossMAE [14]-20083.60
MoCo-V3 [3]-20083.14
DINO [5]-20082.80
SimMIM [34]-20083.50
U-MAE [37]-20083.00
NoR-MAE (This work)-20083.42 (+0.12)
Sup NoR-MAE (This work)-20083.70 (+0.10)
NoR-DMAE (This work)-20084.22 (+0.22)
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.668, + 0.486, + 0.762 + ], + "angle": 0, + "content": "distillation-based MAE, while significantly reducing the training resources, in terms of GPU training time and memory usage. As shown in Table 2, NoR-MAE directly trains the lightweight model from scratch, achieving on-par or even better performance compared to the recent distillation-based MAE training, which yet consumes massive GPU resources." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.771, + 0.472, + 0.789 + ], + "angle": 0, + "content": "4.2. Transfer Learning on the Downstream Tasks" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.486, + 0.903 + ], + "angle": 0, + "content": "We fine-tune the pre-trained ViT-Tiny and ViT-Small models (from Table 2) on the downstream vision tasks with 100 epochs. As shown in Table 4, the proposed NoR-MAE algorithm outperforms the MAE baseline [19] among all the downstream tasks with up to \\(5.13\\%\\) accuracy improvements. The outstanding downstream performance indicates the capability of learning strong representation through NoR-MAE." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.668, + 0.667, + 0.684 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.698, + 0.909, + 0.804 + ], + "angle": 0, + "content": "Intensity of Neighbor-Repelling \\((\\lambda)\\). We evaluate the impact of the penalty level \\(\\lambda\\) with different values. As shown in Table 5, the proposed NoR-MAE algorithm achieves the best performance with \\(\\lambda = 5\\mathrm{e} - 5\\). In other words, the repelling between the embedding features of the visible patches and the CNP should be properly controlled. Over-penalized semantic relation leads to sub-optimal performance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Impact of Masking Ratio. In addition to the \\(75\\%\\) masking ratio that has been used in all prior works as the default setting, we evaluate the impact of different input masking ratios, varying from \\(50\\%\\) to \\(85\\%\\) sparsity. Among different input masking ratios, the proposed NoR-MAE consistently outperforms the vanilla MAE [19] with lightweight vision" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.926, + 0.521, + 0.938 + ], + "angle": 0, + "content": "25236" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.089, + 0.905, + 0.105 + ], + "angle": 0, + "content": "Table 3. Fine-tuning accuracy (%) of the vision transformer encoder pre-trained by MAE and recent SoTA MAE methods on ImageNet-100." + }, + { + "type": "table", + "bbox": [ + 0.132, + 0.115, + 0.865, + 0.375 + ], + "angle": 0, + "content": "
ViT Model# of ParametersPre-training MethodPre-training EpochsSupervised Fine-tuning Accuracy (%)
ViT-Tiny5.7MMAE [19]20071.04
DiNO-Tiny [5, 9]20063.04
CrossMAE [14]20070.82
NoR-MAE (This work)20078.26 (+7.22)
ViT-Small21.7MMAE [19]20081.82
MaskFeat HOG [32]20082.80
PeCo [10]20083.60
NoR-MAE (This work)20084.28 (+2.46)
ViT-Base88.6MMAE [19]20086.80
CrossMAE [14]20086.29
U-MAE [37]20086.80
BEiT [4]20086.10
NoR-MAE (This work)20087.56 (+0.76)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.096, + 0.388, + 0.9, + 0.403 + ], + "angle": 0, + "content": "Table 4. Fine-tuning accuracy (%) of the NoR-MAE on the downstream tasks with the lightweight ViT pre-trained on the ImageNet-1K." + }, + { + "type": "table", + "bbox": [ + 0.111, + 0.413, + 0.885, + 0.498 + ], + "angle": 0, + "content": "
ViT ModelMethodCIFAR-10CIFAR-100FlowersPetsDTDFoodCaltech-101AircraftADE20K (mIoU)
ViT-TinyMAE [19]95.0178.7485.8082.4755.0775.6368.0564.6026.54
NoR-MAE (This work)95.9880.1388.7185.2960.1878.7173.1867.0733.54
ViT-SmallMAE [19]97.5084.8391.4991.1462.3181.5780.8555.9041.14
NoR-MAE (This work)98.1785.5791.8591.7064.7983.1582.0857.4142.99
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.52, + 0.486, + 0.55 + ], + "angle": 0, + "content": "Table 5. Impact of different penalty levels \\((\\lambda)\\) on Neighbor Repelling on the ImageNet-100 dataset." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.56, + 0.484, + 0.607 + ], + "angle": 0, + "content": "
Penalty Level (λ)0.0 (Baseline)1e-55e-51e-45e-4
ViT-Tiny70.2476.2078.2677.8972.17
ViT-Small81.8282.1484.2884.2181.83
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.624, + 0.483, + 0.666 + ], + "angle": 0, + "content": "Table 6. Impact of different input masking ratios. NoR-MAE shows consistent performance improvements on the ImageNet-1K dataset with ViT-Tiny model." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.668, + 0.484, + 0.735 + ], + "angle": 0, + "content": "
Patch-wise Masking RatioBaseline MAE [19]NoR-MAE (This work)
50%65.43%69.27%
75%66.60%70.24%
85%66.37%70.16%
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.764, + 0.483, + 0.809 + ], + "angle": 0, + "content": "transformers, as shown in Table 6. In other words, semantic aliasing is persistent in lightweight transformer models regardless of the information density in the input sequence." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.484, + 0.902 + ], + "angle": 0, + "content": "NoR-MAE with extended training effort We further report the performance of the NoR-MAE algorithm with 1,000 and 1,600 epoch training on the ImageNet-100. As shown in Table 7, the extended training effort can further boost up the performance of the NoR-MAE-trained ViT-Tiny (5.2M) model to \\(83.96\\%\\), while consistently outperforming MAE." + }, + { + "type": "table_caption", + "bbox": [ + 0.514, + 0.52, + 0.905, + 0.534 + ], + "angle": 0, + "content": "Table 7. Performance of NoR-MAE with extended training effort." + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.535, + 0.905, + 0.597 + ], + "angle": 0, + "content": "
Training EpochsMAE on ImageNet-100 (%)NoR-MAE on ImageNet-100 (%)
20071.0478.26
100073.2282.97
160074.2983.96
" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.65, + 0.634, + 0.667 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.689, + 0.909, + 0.901 + ], + "angle": 0, + "content": "In this paper, we propose NoR-MAE, a novel self-supervised learning algorithm designed for lightweight vision transformers via masked autoencoder training. We first introduce the concept of closest neighbor patches (CNP) into the MAE training, which is a critical concept in lightweight SSL. On top of that, we propose Neighbor-Repelling Loss for MAE training. The proposed method trains the lightweight model from scratch without using a pre-trained teacher model. NoR-MAE achieves the new state-of-the-art performance on direct lightweight MAE training with largely improved performance on both pre-training and downstream vision tasks. Finally, our discovery of the biased semantic relations provides valuable insights regarding lightweight masked autoencoder learning." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.926, + 0.521, + 0.938 + ], + "angle": 0, + "content": "25237" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.264, + 0.108 + ], + "angle": 0, + "content": "6. Acknowledgment" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.116, + 0.486, + 0.162 + ], + "angle": 0, + "content": "This work is supported in part by the National Science Foundation under grant 2403723, and CoCoSys Center in JUMP 2.0, an SRC program sponsored by DARPA." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.178, + 0.188, + 0.194 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.203, + 0.486, + 0.244 + ], + "angle": 0, + "content": "[1] Herve Abdi and Lynne J Williams. Principal component analysis. Wiley interdisciplinary reviews: computational statistics, 2(4):433-459, 2010. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.246, + 0.485, + 0.316 + ], + "angle": 0, + "content": "[2] Alexei Baevski, Wei-Ning Hsu, Qiantong Xu, Arun Babu, Jiatao Gu, and Michael Auli. Data2vec: A general framework for self-supervised learning in speech, vision and language. In International Conference on Machine Learning (ICML), 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.318, + 0.484, + 0.387 + ], + "angle": 0, + "content": "[3] Yutong Bai, Zeyu Wang, Junfei Xiao, Chen Wei, Huiyu Wang, Alan L Yuille, Yuyin Zhou, and Cihang Xie. Masked autoencoders enable efficient knowledge distillers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1, 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.389, + 0.484, + 0.431 + ], + "angle": 0, + "content": "[4] Hangbo Bao, Li Dong, Songhao Piao, and Furu Wei. BEit: BERT pre-training of image transformers. In International Conference on Learning Representations (ICLR), 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.433, + 0.485, + 0.501 + ], + "angle": 0, + "content": "[5] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the International Conference on Computer Vision (ICCV), 2021. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.504, + 0.484, + 0.559 + ], + "angle": 0, + "content": "[6] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International Conference on Machine Learning (ICML), 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.562, + 0.484, + 0.602 + ], + "angle": 0, + "content": "[7] Xinlei Chen, Haoqi Fan, Ross Girshick, and Kaiming He. Improved Baselines with Momentum Contrastive learning. arXiv preprint arXiv:2003.04297, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.605, + 0.484, + 0.687 + ], + "angle": 0, + "content": "[8] Yu-An Chung, Yu Zhang, Wei Han, Chung-Cheng Chiu, James Qin, Ruoming Pang, and Yonghui Wu. W2v-bert: Combining contrastive learning and masked language modeling for self-supervised speech pre-training. In 2021 IEEE Automatic Speech Recognition and Understanding Workshop (ASRU), 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.69, + 0.484, + 0.745 + ], + "angle": 0, + "content": "[9] Victor Guilherme Turrisi da Costa, Enrico Fini, Moin Nabi, Nicu Sebe, and Elisa Ricci. solo-learn: A library of self-supervised methods for visual representation learning. Journal of Machine Learning Research, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.747, + 0.484, + 0.816 + ], + "angle": 0, + "content": "[10] Xiaoyi Dong, Jianmin Bao, Ting Zhang, Dongdong Chen, Weiming Zhang, Lu Yuan, Dong Chen, Fang Wen, Nenghai Yu, and Baining Guo. Peco: Perceptual codebook for bert pre-training of vision transformers. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.818, + 0.484, + 0.9 + ], + "angle": 0, + "content": "[11] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.203, + 0.486, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.147 + ], + "angle": 0, + "content": "[12] Tianqi Du, Yifei Wang, and Yisen Wang. On the role of discrete tokenization in visual representation learning. In International Conference on Learning Representations (ICLR), 2023. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.151, + 0.907, + 0.205 + ], + "angle": 0, + "content": "[13] Zhiyuan Fang, Jianfeng Wang, Lijuan Wang, Lei Zhang, Yezhou Yang, and Zicheng Liu. Seed: Self-supervised distillation for visual representation. In International Conference on Learning Representations (ICLR), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.208, + 0.907, + 0.262 + ], + "angle": 0, + "content": "[14] Letian Fu, Long Lian, Renhao Wang, Baifeng Shi, Xudong Wang, Adam Yala, Trevor Darrell, Alexei A Efros, and Ken Goldberg. Rethinking patch dependence for masked autoencoders. arXiv preprint arXiv:2401.14391, 2024. 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.264, + 0.907, + 0.331 + ], + "angle": 0, + "content": "[15] Peng Gao, Shijie Geng, Renrui Zhang, Teli Ma, Rongyao Fang, Yongfeng Zhang, Hongsheng Li, and Yu Qiao. Clip-adapter: Better vision-language models with feature adapters. International Journal of Computer Vision, 132(2):581-595, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.335, + 0.907, + 0.417 + ], + "angle": 0, + "content": "[16] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. Advances in Neural Information Processing Systems (NeurIPS), 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.42, + 0.907, + 0.461 + ], + "angle": 0, + "content": "[17] Agrim Gupta, Jiajun Wu, Jia Deng, and Li Fei-Fei. Siamese masked autoencoders. In Thirty-seventh Conference on Neural Information Processing Systems (NeurIPS), 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.463, + 0.907, + 0.53 + ], + "angle": 0, + "content": "[18] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.533, + 0.907, + 0.601 + ], + "angle": 0, + "content": "[19] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dólár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 2, 3, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.604, + 0.907, + 0.645 + ], + "angle": 0, + "content": "[20] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.648, + 0.907, + 0.716 + ], + "angle": 0, + "content": "[21] Wei Huang, Zhiliang Peng, Li Dong, Furu Wei, Jianbin Jiao, and Qixiang Ye. Generic-to-specific distillation of masked autoencoders. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.719, + 0.907, + 0.759 + ], + "angle": 0, + "content": "[22] Feng Liang, Yangguang Li, and Diana Marculescu. Supmae: Supervised masked autoencoders are efficient vision learners. arXiv preprint arXiv:2205.14540, 2022. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.761, + 0.907, + 0.829 + ], + "angle": 0, + "content": "[23] Jihao Liu, Xin Huang, Jinliang Zheng, Yu Liu, and Hongsheng Li. Mixmae: Mixed and masked autoencoder for efficient pretraining of hierarchical vision transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.832, + 0.907, + 0.9 + ], + "angle": 0, + "content": "[24] Jian Meng, Li Yang, Kyungmin Lee, Jinwoo Shin, Deliang Fan, and Jae sun Seo. Slimmed asymmetrical contrastive learning and cross distillation for lightweight model training. In Thirty-seventh Conference on Neural Information Processing Systems (NeurIPS), 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.927, + 0.521, + 0.938 + ], + "angle": 0, + "content": "25238" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.484, + 0.134 + ], + "angle": 0, + "content": "[25] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation Learning with Contrastive Predictive Coding. arXiv preprint arXiv:1807.03748, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.134, + 0.484, + 0.217 + ], + "angle": 0, + "content": "[26] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning (ICML), 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.218, + 0.484, + 0.288 + ], + "angle": 0, + "content": "[27] Bowen Shi, Xiaopeng Zhang, Yaoming Wang, Jin Li, Wenrui Dai, Junni Zou, Hongkai Xiong, and Qi Tian. Hybrid distillation: Connecting masked autoencoders with contrastive learners. In The Twelfth International Conference on Learning Representations (ICLR), 2024. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.289, + 0.484, + 0.37 + ], + "angle": 0, + "content": "[28] Yucheng Tang, Dong Yang, Wenqi Li, Holger R Roth, Bennett Landman, Daguang Xu, Vishwesh Nath, and Ali Hatamizadeh. Self-supervised pre-training of swim transformers for 3d medical image analysis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.371, + 0.484, + 0.44 + ], + "angle": 0, + "content": "[29] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In International Conference on Machine Learning (ICML), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.441, + 0.484, + 0.497 + ], + "angle": 0, + "content": "[30] Jue Wang, Gedas Bertasius, Du Tran, and Lorenzo Torresani. Long-short temporal contrastive learning of video transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (NeurIPS), 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.497, + 0.484, + 0.552 + ], + "angle": 0, + "content": "[31] Shaoru Wang, Jin Gao, Zeming Li, Xiaoqin Zhang, and Weiming Hu. A closer look at self-supervised lightweight vision transformers. In International Conference on Machine Learning (ICML). PMLR, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.553, + 0.484, + 0.621 + ], + "angle": 0, + "content": "[32] Chen Wei, Haoqi Fan, Saining Xie, Chao-Yuan Wu, Alan Yuille, and Christoph Feichtenhofer. Masked feature prediction for self-supervised visual pre-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.622, + 0.484, + 0.705 + ], + "angle": 0, + "content": "[33] Kan Wu, Houwen Peng, Zhenghong Zhou, Bin Xiao, Mengchen Liu, Lu Yuan, Hong Xuan, Michael Valenzuela, Xi Stephen Chen, Xinggang Wang, et al. Tinyclip: Clip distillation via affinity mimicking and weight inheritance. In Proceedings of the IEEE/CVF International Conference on Computer Vision (CVPR), 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.706, + 0.484, + 0.775 + ], + "angle": 0, + "content": "[34] Zhenda Xie, Zheng Zhang, Yue Cao, Yutong Lin, Jianmin Bao, Zhuliang Yao, Qi Dai, and Han Hu. Simmim: A simple framework for masked image modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.776, + 0.484, + 0.843 + ], + "angle": 0, + "content": "[35] Sukmin Yun, Hankook Lee, Jaehung Kim, and Jinwoo Shin. Patch-level representation learning for self-supervised vision transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.845, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[36] Jure Zbontar, Li Jing, Ishan Misra, Yann LeCun, and Stéphane Deny. Barlow Twins: Self-supervised Learning via Redundancy Reduction. In International Conference on Machine Learning (ICML), pages 12310-12320, 2021. 1, 2, 4, 5" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.147 + ], + "angle": 0, + "content": "[37] Qi Zhang, Yifei Wang, and Yisen Wang. How mask matters: Towards theoretical understandings of masked autoencoders. Advances in Neural Information Processing Systems (NeurIPS), 2022. 1, 3, 4, 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.149, + 0.906, + 0.204 + ], + "angle": 0, + "content": "[38] Zhiyu Zhao, Bingkun Huang, Sen Xing, Gangshan Wu, Yu Qiao, and Limin Wang. Asymmetric masked distillation for pre-training small foundation models. arXiv preprint arXiv:2311.03149, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.204 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.927, + 0.519, + 0.938 + ], + "angle": 0, + "content": "25239" + } + ] +] \ No newline at end of file diff --git a/2025/Closest Neighbors are Harmful for Lightweight Masked Auto-encoders/ce070f35-5bfa-41af-9465-b1f84c742005_origin.pdf b/2025/Closest Neighbors are Harmful for Lightweight Masked Auto-encoders/ce070f35-5bfa-41af-9465-b1f84c742005_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..0f25ed6e6f8860baa3d12cab1bf8a0228491e7c7 --- /dev/null +++ b/2025/Closest Neighbors are Harmful for Lightweight Masked Auto-encoders/ce070f35-5bfa-41af-9465-b1f84c742005_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2aadbe57b03a262efd334e33555085216c71ae4d35cdb19bb805a283685e9262 +size 7094433 diff --git a/2025/Closest Neighbors are Harmful for Lightweight Masked Auto-encoders/full.md b/2025/Closest Neighbors are Harmful for Lightweight Masked Auto-encoders/full.md new file mode 100644 index 0000000000000000000000000000000000000000..9c551bba451e07f671c9a2573793105061d69136 --- /dev/null +++ b/2025/Closest Neighbors are Harmful for Lightweight Masked Auto-encoders/full.md @@ -0,0 +1,269 @@ +# Closest Neighbors are Harmful for Lightweight Masked Auto-encoders + +$^{1}$ Jian Meng, $^{1}$ Ahmed Hassssan, $^{2}$ Li Yang, $^{3}$ Deliang Fan, $^{4}$ Jinwoo Shin, and $^{1}$ Jae-sun Seo + $^{1}$ Cornell University, $^{2}$ University of North Carolina at Charlotte, $^{3}$ Arizona State University, $^{4}$ KAIST + $^{1}\{jm2787, ah2288, js3528\} @cornell.edu, ^{2}lyang50@uncc.edu, ^{3}dfan12@asu.edu, ^{4}jinwoos@kaist.ac.kr$ + +# Abstract + +Learning the visual representation via masked autoencoder (MAE) training has been proven to be a powerful technique. Transferring the pre-trained vision transformer (ViT) to downstream tasks leads to superior performance compared to conventional task-by-task supervised learning. Recent research works on MAE focus on large-sized vision transformers ( $>50$ million parameters) with outstanding performance. However, improving the generality of the under-parametrized lightweight model has been widely ignored. In practice, downstream applications are commonly intended for resource-constrained platforms, where large-scale ViT cannot easily meet the resource budget. Current lightweight MAE training heavily relies on knowledge distillation with a pre-trained teacher, whereas the root cause behind the poor performance remains underexplored. Motivated by that, this paper first introduces the concept of "closest neighbor patch" to characterize the local semantics among the input tokens. Our discovery shows that the lightweight model failed to distinguish different local information, leading to biased understanding and poor accuracy. Motivated by this finding, we propose NoR-MAE, a novel MAE training algorithm for lightweight vision transformers. NoR-MAE elegantly repels the semantic aliasing between patches and their closest neighboring patch (semantic centroid) with negligible training cost overhead. With the ViT-Tiny model, NoR-MAE achieves up to $7.22\% / 3.64\%$ accuracy improvements on ImageNet100/ImageNet-1K datasets, as well as up to $5.13\%$ accuracy improvements in tested downstream tasks. https://github.com/SeoLabCornell/NoR-MAE + +# 1. Introduction + +Starting from the early exploration with contrastive learning [6, 18], learning powerful and generic vision or semantic representation has been the focal point of all prior self-supervised learning (SSL) algorithms across various machine learning domains [2, 8, 28, 30] and vision-language model [15, 26]. The superiority of SSL is demonstrated + +![](images/74bc6eb3dbc3e2a1a4f6f4bf024bd78d64fd939e071375606bdf94fc097022c7.jpg) +Figure 1. Top: Closest Neighbor Patches (CNP) characterizes the "Local Semantic Centroid" among the input tokens. Bottom: Low-dimensional projection (via PCA) of the transformer-encoded patches. Unlike the large-sized transformer model (e.g., ViT-Base), lightweight models (e.g., ViT-Tiny) cannot separate the "semantic difference" between different CNPs. + +by the capability of extracting transferable knowledge from unlabeled datasets [26, 33]. Recent efforts on masked autoencoder (MAE) [19] train the vision transformer [11] by encouraging the model to understand the semantic knowledge of the masked patches throughout the reconstruction-based self-supervised learning. Particularly, the evolution from contrastive learning [16, 18, 36] to MAE-based training [17, 19, 34, 37] shows consistent dominance by large-scale models to ensure on-par or even better accuracy compared to supervised learning, along with superior performance in various downstream applications. In practice, downstream small-scale tasks are commonly deployed on edge devices with limited resources. Therefore, large-sized vision transformers (ViTs) pre-trained by MAE [19] are suboptimal candidates for resource-constrained downstream deployment and hardware platforms. Unfortunately, recent state-of-the-art (SoTA) methods that train lightweight ViTs with MAE from scratch lead to poor performance and insufficient transferability. For example, the MAE-trained lightweight ViT-Tiny model only achieves $66.6\%$ top-1 accuracy (with 200-epoch pre-training [3, 19, 37]) on the + +ImageNet-1K dataset [3], which is $\sim 10\%$ lower than the supervised learning counterpart. Evidently, the superiority of MAE [19] with the large model failed to be maintained in the lightweight ViTs [11]. Given the diverse downstream tasks and urgent need for energy-efficient foundation models, unleashing the power of MAE and enhancing the versatility of lightweight models is desired. + +Prior works enhance the performance of the lightweight transformer models by heavily relying on knowledge distillation (KD) [20] for both supervised [29] and self-supervised learning paradigms [3, 21, 38]. For instance, the recent DMAE work [3] combines MAE [19] with KD and transfers the pre-trained knowledge from a ViT-Base teacher down to the lightweight ViT model. DMAE [3] chooses the 3/4 depth of both student and teacher ViT models as the bonding portal for feature alignment and knowledge distillation. The heuristic design of DMAE cannot be easily extended to other encoder models. Furthermore, distilling the knowledge on the fly introduces an additional projector [3] or logits mask [31], which further elevates the complexity and computation cost of the entire training process. In addition to DMAE, other recent works further increase the complexity of the KD-based MAE by employing multi-layer distillation [38] or complex generic-to-specific distillation [21]. All of these prior works [3, 21, 31, 38] rely on the brute force approach to enhance the model performance with increasingly complex training schemes, which results in inflated GPU memory and training time. + +However, none of these prior works can collectively achieve 1) the capability of training lightweight model from scratch without a large-sized teacher model, 2) negligible training cost overhead compared to the vanilla MAE, and 3) clear insights that cause the poor performance of lightweight MAE. To overcome this, the key questions to address are: + +① What is the bottleneck of MAE with under parametrized small vision transformer models? +② How to efficiently train a high performance lightweight model from scratch via MAE? + +To answer ①, our investigation shows that the lightweight model failed to understand different local semantic information among the input tokens and lead to poor performance. We characterize different local semantics by introducing the concept of closest neighborhood patch (CNP) (Figure 1), which carries the most similar semantics of a given input patch. Since multiple patches can share one CNP (e.g., Background Forrest), different CNPs represent various local semantics across the entire input sequence. Unlike large-sized models, the under-parameterized model falls into the semantic aliasing between different CNPs with poor accuracy, as indicated in Figure 1 and Section 3.3. + +Motivated by the findings in ①, we propose the Neighbor Repelling Masked Autoencoder (NoR-MAE) to resolve ②, a novel self-supervised learning algorithm designed for train + +ing powerful lightweight vision transformers. NoR-MAE trains the powerful small vision transformer by penalizing the similarity between the unmasked patches and CNPs. Different from prior works, NoR-MAE directly trains the vision transformer from scratch without introducing the largesized pre-trained teacher or heuristically designed distillation scheme. The proposed method achieves up to $7.22\%$ and $3.64\%$ accuracy improvements on ImageNet-100 and ImageNet-1K datasets, with negligible training time and memory overhead compared to the vanilla MAE. The major contributions of our work are: + +- Simplicity: NoR-MAE does not require teacher pretraining or heuristic knowledge distillation to train a powerful lightweight ViT. +- High Performance: Compared to the baseline MAE, NoR-MAE achieves $7.22\%$ and $3.64\%$ accuracy improvements on ImageNet-100 and ImageNet-1K datasets with the ViT-Tiny model, achieving the new SoTA performance on the lightweight MAE training. +- Transferrability: Different from prior works, which mainly focus on fine-tuning the performance on the pretraining dataset (e.g., ImageNet), NoR-MAE is also evaluated with the downstream fine-tuning and linear evaluation benchmarks, with significantly improved performance compared to prior SSL baselines and the vanilla MAE. +- Rationality: Besides the superior performance, for the first time, NoR-MAE reveals the semantic aliasing issue caused by the neighboring patches, which is proved to be the reason that hinders efficient training and representation learning of lightweight ViT models. + +# 2. Related Work + +Empowering a deep neural network model with strong and transferable knowledge has been widely investigated, especially in self-supervised learning. Early research works in contrastive learning (CL) introduce the learning paradigm with "positive" and "negative" sample pairs [6, 18], together with the alignment-repellent strategy embedded into InfoNCE [7, 18, 25] and NT-Xent loss [6] as the learning objective. The early success of CL motivated the successors to investigate the potential of different training and sampling strategies, inducing the standpoint of asymmetrical learning [16, 24], knowledge distillation [13], and the latent-dimension correlation [36]. + +While the core idea of CL is training the model to understand the similarities and differences between different augmentations, the emergence of masked autoencoder (MAE) [19] training reconsiders SSL from a different perspective, which is perfectly suitable for vision transformers. Encouraging the encoder-decoder to reconstruct the randomly sparsified patches allows the model to understand the semantics. Due to the powerful visual representation, MAE-based pretaining exhibits strong performance on + +![](images/aa9725bd0ee949afb6e6d813dfebc077932fe2a507d0e1ad34fa4413310f62ff.jpg) +Figure 2. Closest Neighbor Patches (CNP) within the image. + +both backbone datasets (e.g., ImageNet-1K) and downstream transfer learning tasks. Subsequent research works further improved the performance of MAE from the perspectives of contrastive learning [27, 37], cross-attention [14, 17], mixed feature learning [23], or intricate tokenizer [12]. + +Across all the different MAE training methods, elevating the performance of the large-sized models (e.g., ViT-Large) has been the primary focus of exploration, while improving the poor performance of the small-sized model (e.g., ViTTiny) has been largely ignored. Given that the downstream tasks are well-suited for resource-constrained edge devices, the imbalanced research focus of the prior works largely hinders the efficiency and transferability of MAE [19]. The practical needs and the imperfection of lightweight MAE motivate the recent research works to investigate MAE training strategies with small models. Following the protocol from supervised distillation [20], knowledge distillation-based MAE [3, 21] is proposed to enhance the performance of the lightweight student. However, the learning scheme of the single-layer or multi-layer distillation is established based on the heuristically designed distillation between intermediate features. Compared to the supervised knowledge distillation, recent studies largely complicate the overall cost of training lightweight ViT models with either heuristic design [3], multi-phased distillation [21], or distillation with extensive fine-tuning [31]. In particular, employing the pre-trained large ViT model as the teacher introduces additional GPU memory usage and time cost toward the total training efforts. Furthermore, training the interconnect projectors between the student and teacher exacerbates the training complexity even further. More importantly, the knowledge distillation-based MAE inherits the essence of KD from supervised learning, but the insights of training lightweight ViT from scratch via MAE remain largely unexplored. + +# 3. Proposed Method + +We propose the Neighbor-Repelling MAE (NoR-MAE). In this section, we will 1) introduce the concept of the Closest Neighbor Patch into MAE training, 2) present the proposed NoR-MAE training, and 3) unravel the rationality of NoR-MAE by revealing the insights that cause the poor performance of the directly-trained lightweight ViT model. + +![](images/fac5c4699b24f06121d93a3717d4ec1f9a6ccdf8d196fae1fd81ffb59723d3dc.jpg) +Figure 3. Cosine Similarity between each patch and their Closest Neighbor Patch across a subset of ImageNet-1K with 4096 images. + +# 3.1. Closest Neighbor Patch (CNP) + +Let $X \in \mathbf{R}^{H \times W \times C}$ be an input image to the model, where $H$ , $W$ , and $C$ represent the height, width, and channel of the image. Vision transformer (ViT) reformulates the input by decomposing the image $X$ into a sequence of non-overlapped patches with the pre-defined patch size $P \times P$ and the total number of patches (tokens) $K$ . All the patches are further encoded into the embeddings with dimension $D$ , which is the input of the subsequent transformer blocks. + +For a given input patch $X_P$ , there are naturally 8 neighbors that share a common edge with $X_P$ , as shown in Figure 2(a). We define the Closest Neighbor Patch (CNP) $X_{CNP}$ by selecting the neighbor that has the Top-1 cosine similarity with $X_P$ . In other words, $X_P$ and $X_{CNP}$ are sharing the highest degree of semantically meaningful relations. + +Shared Closest Neighbor Patch. In the example of Figure 2, two patches $X_P^i$ and $X_P^j$ partially share the neighborhoods in between, while the CNP of $X_P^i$ and $X_P^j$ happen to be identical as $X_{CNP}^{ij}$ . Naturally, $X_{CNP}^{ij}$ carries the shared semantic information ("Background Forest") for both $X_P^i$ and $X_P^j$ , but the degree of similarity between $(X_P^i, X_{CNP}^{ij})$ and $(X_P^j, X_{CNP}^{ij})$ are different. + +Closest Neighbor Patch $\neq$ Identical Patch. Although $X_{P}$ and $X_{CNP}$ are sharing the highest degree of similarity, the semantics between $X_{P}$ and $X_{CNP}$ are not identical. As shown in Figure 3, the cosine similarity between all the patches and their closest neighbor in ImageNet-1K images varies from 0.07 to 0.999, with an average of 0.83. In other words, the highly similar neighboring patches also contain the implicit semantical difference. + +Difficulty of Lightweight Models to Understand the Different Semantics Carried by CNP. With MAE, the input patches of the transformer models are randomly masked with a pre-defined sparsity ratio $\tau$ (e.g., $75\%$ ), as illustrated in Figure 4, left. Naturally, each unmasked visible patch has a Closest Neighbor Patch (CNP). As a result, the CNP + +![](images/fe7c07b0b0d88cc7ead5211fbdf2d18f93ea04eb765971248ccbd407a769717e.jpg) +Figure 4. The shared Closest Neighbor Patch (CNP) characterizes the estimated local semantic centroids of the input sequence. + +of the unmasked patches formulates another sparse input image (Figure 4, right). From the perspective of the entire image, visible patches and their CNPs are close neighbors, while different visible patches may share the same closest neighbor, which estimates the local and regional semantics of the input sequence. However, as described in Figure 1 and Section 1, understanding the semantic difference between different CNPs is the bottleneck for the under-parameterized models. Furthermore, our extended experiments in Figure 7 show that the key difference between the big (e.g., ViT-Base) and small (e.g., ViT-Tiny) model is understanding and separating the semantic differences between CNPs. Therefore, the challenge of lightweight MAE is clear: + +In MAE training, how to facilitate the lightweight model to understand different local semantics? + +# 3.2. Proposed Method: NoR-MAE + +We propose the Neighbor-Repelling training for self-supervised MAE learning (NoR-MAE). The overall objective is to train a powerful, lightweight vision transformer by encouraging the model to avoid the biased relation between different local semantics of the input. We follow the standard protocol of the vanilla MAE with the encoder $f_{\theta}$ and decoder $g_{\phi}$ , where both $f_{\theta}$ and $g_{\phi}$ are vision transformers with the same embedding dimension $D$ . + +Given the masking ratio $\tau$ , we first generate the masked input sequence $x^{\tau}$ and $x_{CNP}^{\tau}$ that are embedded from the unmasked patches with the corresponding closest neighbor patches, respectively. During the forward pass of each iteration, we parallelly encode and decode the masked sequence $x^{\tau}$ and CNP sequence $x_{CNP}^{\tau}$ by $f_{\theta}$ and $g_{\phi}$ , generating the reconstructed embedding $Z$ and $Z_{CNP}$ , respectively. In particular, $f_{\theta}$ and $g_{\phi}$ are served as the siamese encoder and decoder (with stop gradient) during the forward pass of $Z_{CNP}$ . + +Overall, each forward pass will generate two pairs of decoded tokens resulting from $x^{\tau}$ and $x_{\mathrm{CNP}}^{\tau}$ . + +Following the standard training protocol of MAE, we first compute the $\mathcal{L}_2$ reconstruction loss between the decoded sequence $Z$ and the ground truth. Subsequently, we compute the matrix multiplication between the normalized $Z$ and $Z_{CNP}$ . As shown in Figure 5, the encoder $f_{\theta}$ and decoder + +$g_{\phi}$ are shared for both branches, while the stop gradient is enabled for the CNP input. Given the size of $[N,K,D]$ , the matrix multiplication between $\bar{Z}$ and transposed $\bar{Z}_{\mathrm{CNP}}$ is performed along the token dimension $K$ . The resultant matrix is further scaled down by $K$ and takes the average along the batch dimension $N$ . Mathematically, we have: + +$$ +C _ {\text {n e i g h b o r}} = \frac {1}{N} \sum_ {N} \left(\frac {1}{K} (\bar {Z}) ^ {T} \cdot \bar {Z} _ {\mathrm {C N P}}\right) \tag {1} +$$ + +Where $C_{\mathrm{neighbor}}$ represents the resultant relation matrix with the size of $D \times D$ . Each entry $c_{i,j}$ of $C_{\mathrm{neighbor}}$ characterizes the correlation between the randomly-unmasked input sequence (output logits $Z$ ) and the corresponding CNP sequence (output logits $Z_{\mathrm{CNP}}$ ) between embedding dimension $i$ and $j$ . To minimize the semantic aliasing between the unmasked tokens and their CNPs, we decorate different embeddings by penalizing the off-diagonal terms of $C_{\mathrm{neighbor}}$ : + +$$ +\mathcal {L} _ {\mathrm {N o R}} = \sum_ {i} \sum_ {i \neq j} C _ {\text {n e i g h b o r}} ^ {i, j} \tag {2} +$$ + +Essentially, the proposed NoR-Loss $(\mathcal{L}_{\mathrm{NoR}})$ minimizes semantic aliasing by preventing the model encoding the visible patches and CNP as aliased embeddings. It has been shown in Figure 3 that patches and their CNPs share similar but non-identical semantics. We balance such a tradeoff of "similarity-discrepancy" by decorrelating the off-diagonal embeddings only. The overall loss is the combination of the standard $\mathcal{L}_2$ reconstruction loss and the proposed NeighborRepelling Loss $\mathcal{L}_{\mathrm{NoR}}$ scaled by a hyperparameter $\lambda$ : + +$$ +\mathcal {L} _ {\text {t o t a l}} = \left| \left| \text {G r o u n d T r u t h} - \bar {Z} \right| \right| _ {2} + \lambda \cdot \mathcal {L} _ {\mathrm {N o R}} (\bar {Z}, \bar {Z} _ {\mathrm {C N P}}) \tag {3} +$$ + +More importantly, decorrelating embeddings eventually separate the aliased semantics between different CNPs and their corresponding "local semantic centroids" (Section 3.3). In other words, the embedding-level anti-aliasing regularization of the proposed NoR-Loss facilitates the understanding on the representation (token) level, leading to the largely improved model performance. + +Decorrelating Embeddings vs. Decorrelating Tokens. Although the proposed NoR-MAE selects the Top-1 neighbor (the closest neighbor patch (CNP)) as the candidate for decorrelation, the CNP patches can be overlapped or shared between each other. For instance, the closest neighbor patch of patch $X_{i}$ could also be the Top-2/3/4 ... 8 neighbor of patch $X_{j}$ . Minimizing the semantic aliasing by naively decorrelating tokens leads to collapsed performance and unsuccessful training, as shown in Figure 6. + +NoR-MAE vs. Contrastive Learning. Empowering representation learning via invariance-covariance optimization has been investigated in Barlow Twins [36]. Recently proposed U-MAE [37] combines the contrastive learning and MAE [19] by introducing the random asymmetry masks + +![](images/d3373eb9e2747902089d8a19e3c05b23dc98b5f1630e933e554681ef24e57089.jpg) +Figure 5. Overview of the proposed NoR-MAE algorithm. + +![](images/b60a5977a31156d6ab39a487c36b30b811d9a96cd9f5c03d8ed2d552827182f7.jpg) +Figure 6. Different relation penalty of the NoR-Loss $(\lambda = 5\mathrm{e} - 5)$ + +during the forward pass of the training. On the contrary, the proposed NoR-MAE exploits the semantic aliasing issue by constructing the asymmetry between the visible patches and the corresponding closest neighbors, instead of relying on the randomness. As a result, NoR-MAE outperforms U-MAE [37] in various sizes of models (Section 4). + +In the context of contrastive learning, the similar samples (patches) are considered as "positive samples" which are intended to be aligned by the algorithms [35]. However, the important discovery of NoR-MAE points out that semantic aliasing among the closest neighbors is harmful for the lightweight masked autoencoder training, which is not properly addressed by prior works. Moreover, unlike Barlow Twins [36], NoR-MAE does not need the embedding alignment along the diagonal terms due to the well-established similarity between the visible patches and their CNPs. + +NoR-MAE Does Not Need a Teacher Model. We would like to highlight that the proposed NoR-MAE is entirely different from knowledge distillation-based MAE. NoR-MAE does not require a pre-trained teacher [3] or multistage distillation [21] during training. The lightweight encoder is trained from scratch following the standard MAE protocol [19]. Although the CNPs and the original unmasked patches are separately generated, the encoding-decoding pro + +cess remains the same as the vanilla MAE. For example, given the $75\%$ mask ratio with $25\%$ visible patches, the equivalent input sparsity of NoR-MAE is $50\%$ . The overhead of the NoR-MAE is minimal compared to the vanilla MAE, as profiled in the supplementary. + +# 3.3. Rationality of NoR-MAE + +As shown in Figure 2 and Figure 4, the proposed concept of Closest Neighbor Patch (CNP) characterizes the local semantics of the image. For instance, the patches that contain "Background Forest" or "Back of the Fox" are represented by separate CNP patches. However, the lightweight model failed to learn the semantic difference between CNPs. + +We prove the significance of the CNP semantics by analyzing the original unmasked features and CNP features encoded by the trained encoder $f_{\theta}$ . Specifically, we first normalize the encoded features and then compress the feature dimension down to 2-D via the Principal Component Analysis (PCA) [1]. As shown in Figure 7 and Figure 8, the lightweight models (ViT-Tiny and ViT-Small) trained by the vanilla MAE [19] failed to distinguish the semantical difference between different closest neighbor patches (CNP, blue dots), regardless of the patch sizes. Different semantics of CNP are heavily mixed together in the low dimensional space. Moreover, CNP is critical for MAE training regardless of the input patch sizes and the corresponding sequence length. The distorted and aliased understanding of the local semantics leads to the sub-optimal performance of MAE, as shown in Table 1 with the empirical verification on ImageNet-100 ( $224 \times 224$ ) with ViT-Tiny. + +On the contrary, the model trained by NoR-MAE successfully understands the semantic difference between different CNP features, leading to largely improved accuracy on the lightweight ViT models across different patch sizes. In the meantime, the large-sized model (e.g., ViT-Base) exhibits + +![](images/48ece0badf6f42df429dff5acb665a13e75d4b78025050dd218266046eece57a.jpg) +ViT-Tiny (# of params = 5.7 Million) MAE [19]: Aliased Semantics! NoR-MAE: Clearly Separated CNP! +Accuracy $= 66.60$ + +![](images/a3c7f37475c3840661a41f651fc6cdd3f5c2ce776f044424d7bd7d13fdd8c037.jpg) +Accuracy $= 70.24$ + +![](images/b4d68fa8ebda505a9e55496391aba91dc406ddcfd9f979af87b1fbba76ee70fa.jpg) +Latent Feature of Unmasked Patches +Latent Feature of Closest Neighbor Patches + +![](images/c8d1d286339bd12eb261ad440299bfac49a011b79b73297a9177b3873587d88a.jpg) +Patch size $= 16^{2}$ Token length $= 196$ + +![](images/e261fe14e69b24b254ab14b55e69a4de170ae8965021bae505c05c3edeec4168.jpg) +Figure 7. Low dimensional features of the unmasked patches (red dots) and the closest neighbor patches (CNP) (blue dots) with the same input from ImageNet-1K dataset. NoR-MAE clearly separates the semantic information among different CNP, with improved accuracy. +Figure 8. Low dimensional features of the unmasked patches (red dots) and the CNP (blue dots) encoded by ViT-Tiny with patch size $= 4 \times 4$ (sequence length $= 3136$ ), evaluated on ImageNet-100. + +Table 1. NoR-MAE consistently outperforms the MAE [19] across different input patch sizes (evaluated on the ImageNet-100 dataset). + +
Masking RatioInput Patch SizeMAE [19]NoR-MAE (This work)
75%2²71.08%83.56% (+12.48%)
75%4²71.16%83.58% (+12.42%)
75%8²71.27%79.24% (+7.97%)
75%16²71.04%78.26% (+7.22%)
75%32²68.42%74.73% (+6.31%)
+ +a better understanding and stronger separation of different semantics with the vanilla MAE [19]. + +In summary, the proposed NoR-MAE successfully mitigates the performance gap between model sizes, facilitating the learnability of the transformer model with a stronger understanding of the semantical relations from the input, leading to improved accuracy with both lightweight and large-sized transformer models. + +# 4. Experimental Results + +# 4.1. ImageNet Classification + +For image classification tasks, we follow the standard protocol of MAE [19] by training the lightweight encoder on the ImageNet-1K and ImageNet-100 datasets with the batch size of 4,096 and 2,048, and subsequently performing the standard fine-tuning [19]. Specifically, the model is pre-trained by 200 and 400 epochs with the initial learning rate set to + +1.5e-4, together with the weight decay value of 0.05. The penalty coefficient $\lambda$ of the proposed NoR Loss is set to 5e-5, and the optimality of the selected penalty level is demonstrated in Table 5. The mask ratio is fixed at $75\%$ based on the findings in the vanilla MAE [19]. We implement the standard fine-tuning protocol with 100 epochs as the vanilla MAE [19]. The detailed experimental setup (pre-train and downstream) are summarized in the supplementary, together with the report of GPU memory usage and training cost. + +Table 2 and Table 3 summarize the performance of the NoR-MAE on ImageNet-100 and ImageNet-1K datasets. Compared to the vanilla MAE [19], NoR-MAE largely improves the performance on top of the MAE baseline [19] with up to $7.22\%$ and $3.64\%$ accuracy on ImageNet-100 and ImageNet-1K datasets, respectively. For the large-sized vision transformer training (e.g., ViT-Base), NoR-MAE still maintains the superiority compared to the recent high-performance MAE training (Table 3). + +More importantly, NoR-MAE improves the model performance marginal training cost overhead. NoR-MAE trains the model with the standard MAE protocol without introducing 1) a pre-trained teacher [3] or 2) other additional architectures [27] or clustering modules [12]. As profiled in the supplementary, the performance boost reward by the NoR-MAE only exhibits up to $8\%$ memory overhead and $2\%$ additional training time. Compared to the distillation-based MAE [3], NoR-MAE outperforms DMAE [3] with $\sim 1\%$ accuracy improvement and $50\%$ less training cost. + +We further validate the NoR-MAE on the larger ViT-Base model. The proposed algorithm consistently achieves better performance compared to the MAE baseline and the recent ImageNet-100 SoTA method without introducing any additional clustering efforts or special tokenizer [12]. + +Finally, NoR-MAE is also suitable for the MAE with supervision (SupMAE) [22]. Introducing supervised learning into MAE does not impact the effectiveness of the proposed NoR-MAE, while the NoR-MAE further enhances the performance of SupMAE [22], as shown in Table 2. + +Comparison with the Distillation-based MAE. We would like to highlight that the proposed NoR-MAE can achieve on-par or even better performance compared to the + +Table 2. Fine-tuning and semi-supervised learning accuracy (%) of the ViT encoder pre-trained by MAE [19] and recent SoTA lightweight MAE on ImageNet-1K dataset (with 200 epochs and 400 epochs pre-training). + +
ViT Model# of ParamsMethodTeacher ModelPre-training EpochSupervised Fine-tuning Accuracy (%)
ViT-Tiny5.7MMAE [19]-20066.60
MAE [19]-40067.63
Sup MAE [22]-20067.88
Sup MAE [22]-40068.91
MoCo-V3 [18]-20070.09
DINO [5]-30068.17
SimMIM [34]-20066.08
DMAE [3]ViT-B20070.00
Sup NoR-MAE (This work)-20070.63 (+2.75)
NoR-MAE (This work)-20070.24 (+3.64)
NoR-MAE (This work)-40071.18 (+3.55)
ViT-Small21.7MMAE [19]-20079.00
MAE [19]-40080.11
Sup MAE [22]-20079.27
CrossMAE [14]-40079.30
MoCo-V3 [18]-20079.46
DMAE [3]ViT-B20079.30
Sup NoR-MAE (This work)-20080.16 (+0.89)
NoR-MAE (This work)-20080.13 (+1.03)
NoR-MAE (This work)-40081.04 (+0.93)
ViT-Base88.6MMAE [19]-20083.30
Sup MAE [22]-20083.60
DMAE [3]-20084.00
CrossMAE [14]-20083.60
MoCo-V3 [3]-20083.14
DINO [5]-20082.80
SimMIM [34]-20083.50
U-MAE [37]-20083.00
NoR-MAE (This work)-20083.42 (+0.12)
Sup NoR-MAE (This work)-20083.70 (+0.10)
NoR-DMAE (This work)-20084.22 (+0.22)
+ +distillation-based MAE, while significantly reducing the training resources, in terms of GPU training time and memory usage. As shown in Table 2, NoR-MAE directly trains the lightweight model from scratch, achieving on-par or even better performance compared to the recent distillation-based MAE training, which yet consumes massive GPU resources. + +# 4.2. Transfer Learning on the Downstream Tasks + +We fine-tune the pre-trained ViT-Tiny and ViT-Small models (from Table 2) on the downstream vision tasks with 100 epochs. As shown in Table 4, the proposed NoR-MAE algorithm outperforms the MAE baseline [19] among all the downstream tasks with up to $5.13\%$ accuracy improvements. The outstanding downstream performance indicates the capability of learning strong representation through NoR-MAE. + +# 4.3. Ablation Study + +Intensity of Neighbor-Repelling $(\lambda)$ . We evaluate the impact of the penalty level $\lambda$ with different values. As shown in Table 5, the proposed NoR-MAE algorithm achieves the best performance with $\lambda = 5\mathrm{e} - 5$ . In other words, the repelling between the embedding features of the visible patches and the CNP should be properly controlled. Over-penalized semantic relation leads to sub-optimal performance. + +Impact of Masking Ratio. In addition to the $75\%$ masking ratio that has been used in all prior works as the default setting, we evaluate the impact of different input masking ratios, varying from $50\%$ to $85\%$ sparsity. Among different input masking ratios, the proposed NoR-MAE consistently outperforms the vanilla MAE [19] with lightweight vision + +Table 3. Fine-tuning accuracy (%) of the vision transformer encoder pre-trained by MAE and recent SoTA MAE methods on ImageNet-100. + +
ViT Model# of ParametersPre-training MethodPre-training EpochsSupervised Fine-tuning Accuracy (%)
ViT-Tiny5.7MMAE [19]20071.04
DiNO-Tiny [5, 9]20063.04
CrossMAE [14]20070.82
NoR-MAE (This work)20078.26 (+7.22)
ViT-Small21.7MMAE [19]20081.82
MaskFeat HOG [32]20082.80
PeCo [10]20083.60
NoR-MAE (This work)20084.28 (+2.46)
ViT-Base88.6MMAE [19]20086.80
CrossMAE [14]20086.29
U-MAE [37]20086.80
BEiT [4]20086.10
NoR-MAE (This work)20087.56 (+0.76)
+ +Table 4. Fine-tuning accuracy (%) of the NoR-MAE on the downstream tasks with the lightweight ViT pre-trained on the ImageNet-1K. + +
ViT ModelMethodCIFAR-10CIFAR-100FlowersPetsDTDFoodCaltech-101AircraftADE20K (mIoU)
ViT-TinyMAE [19]95.0178.7485.8082.4755.0775.6368.0564.6026.54
NoR-MAE (This work)95.9880.1388.7185.2960.1878.7173.1867.0733.54
ViT-SmallMAE [19]97.5084.8391.4991.1462.3181.5780.8555.9041.14
NoR-MAE (This work)98.1785.5791.8591.7064.7983.1582.0857.4142.99
+ +Table 5. Impact of different penalty levels $(\lambda)$ on Neighbor Repelling on the ImageNet-100 dataset. + +
Penalty Level (λ)0.0 (Baseline)1e-55e-51e-45e-4
ViT-Tiny70.2476.2078.2677.8972.17
ViT-Small81.8282.1484.2884.2181.83
+ +Table 6. Impact of different input masking ratios. NoR-MAE shows consistent performance improvements on the ImageNet-1K dataset with ViT-Tiny model. + +
Patch-wise Masking RatioBaseline MAE [19]NoR-MAE (This work)
50%65.43%69.27%
75%66.60%70.24%
85%66.37%70.16%
+ +transformers, as shown in Table 6. In other words, semantic aliasing is persistent in lightweight transformer models regardless of the information density in the input sequence. + +NoR-MAE with extended training effort We further report the performance of the NoR-MAE algorithm with 1,000 and 1,600 epoch training on the ImageNet-100. As shown in Table 7, the extended training effort can further boost up the performance of the NoR-MAE-trained ViT-Tiny (5.2M) model to $83.96\%$ , while consistently outperforming MAE. + +Table 7. Performance of NoR-MAE with extended training effort. + +
Training EpochsMAE on ImageNet-100 (%)NoR-MAE on ImageNet-100 (%)
20071.0478.26
100073.2282.97
160074.2983.96
+ +# 5. Conclusion + +In this paper, we propose NoR-MAE, a novel self-supervised learning algorithm designed for lightweight vision transformers via masked autoencoder training. We first introduce the concept of closest neighbor patches (CNP) into the MAE training, which is a critical concept in lightweight SSL. On top of that, we propose Neighbor-Repelling Loss for MAE training. The proposed method trains the lightweight model from scratch without using a pre-trained teacher model. NoR-MAE achieves the new state-of-the-art performance on direct lightweight MAE training with largely improved performance on both pre-training and downstream vision tasks. Finally, our discovery of the biased semantic relations provides valuable insights regarding lightweight masked autoencoder learning. + +# 6. Acknowledgment + +This work is supported in part by the National Science Foundation under grant 2403723, and CoCoSys Center in JUMP 2.0, an SRC program sponsored by DARPA. + +# References + +[1] Herve Abdi and Lynne J Williams. Principal component analysis. Wiley interdisciplinary reviews: computational statistics, 2(4):433-459, 2010. 5 +[2] Alexei Baevski, Wei-Ning Hsu, Qiantong Xu, Arun Babu, Jiatao Gu, and Michael Auli. Data2vec: A general framework for self-supervised learning in speech, vision and language. In International Conference on Machine Learning (ICML), 2022. 1 +[3] Yutong Bai, Zeyu Wang, Junfei Xiao, Chen Wei, Huiyu Wang, Alan L Yuille, Yuyin Zhou, and Cihang Xie. Masked autoencoders enable efficient knowledge distillers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1, 2, 3, 5, 6, 7 +[4] Hangbo Bao, Li Dong, Songhao Piao, and Furu Wei. BEit: BERT pre-training of image transformers. In International Conference on Learning Representations (ICLR), 2022. 8 +[5] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the International Conference on Computer Vision (ICCV), 2021. 7, 8 +[6] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International Conference on Machine Learning (ICML), 2020. 1, 2 +[7] Xinlei Chen, Haoqi Fan, Ross Girshick, and Kaiming He. Improved Baselines with Momentum Contrastive learning. arXiv preprint arXiv:2003.04297, 2020. 2 +[8] Yu-An Chung, Yu Zhang, Wei Han, Chung-Cheng Chiu, James Qin, Ruoming Pang, and Yonghui Wu. W2v-bert: Combining contrastive learning and masked language modeling for self-supervised speech pre-training. In 2021 IEEE Automatic Speech Recognition and Understanding Workshop (ASRU), 2021. 1 +[9] Victor Guilherme Turrisi da Costa, Enrico Fini, Moin Nabi, Nicu Sebe, and Elisa Ricci. solo-learn: A library of self-supervised methods for visual representation learning. Journal of Machine Learning Research, 2022. 8 +[10] Xiaoyi Dong, Jianmin Bao, Ting Zhang, Dongdong Chen, Weiming Zhang, Lu Yuan, Dong Chen, Fang Wen, Nenghai Yu, and Baining Guo. Peco: Perceptual codebook for bert pre-training of vision transformers. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), 2023. 8 +[11] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 1, 2 + +[12] Tianqi Du, Yifei Wang, and Yisen Wang. On the role of discrete tokenization in visual representation learning. In International Conference on Learning Representations (ICLR), 2023. 3, 6 +[13] Zhiyuan Fang, Jianfeng Wang, Lijuan Wang, Lei Zhang, Yezhou Yang, and Zicheng Liu. Seed: Self-supervised distillation for visual representation. In International Conference on Learning Representations (ICLR), 2021. 2 +[14] Letian Fu, Long Lian, Renhao Wang, Baifeng Shi, Xudong Wang, Adam Yala, Trevor Darrell, Alexei A Efros, and Ken Goldberg. Rethinking patch dependence for masked autoencoders. arXiv preprint arXiv:2401.14391, 2024. 3, 7, 8 +[15] Peng Gao, Shijie Geng, Renrui Zhang, Teli Ma, Rongyao Fang, Yongfeng Zhang, Hongsheng Li, and Yu Qiao. Clip-adapter: Better vision-language models with feature adapters. International Journal of Computer Vision, 132(2):581-595, 2024. 1 +[16] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. Advances in Neural Information Processing Systems (NeurIPS), 2020. 1, 2 +[17] Agrim Gupta, Jiajun Wu, Jia Deng, and Li Fei-Fei. Siamese masked autoencoders. In Thirty-seventh Conference on Neural Information Processing Systems (NeurIPS), 2023. 1, 3 +[18] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 1, 2, 7 +[19] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dólár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 2, 3, 4, 5, 6, 7, 8 +[20] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 2, 3 +[21] Wei Huang, Zhiliang Peng, Li Dong, Furu Wei, Jianbin Jiao, and Qixiang Ye. Generic-to-specific distillation of masked autoencoders. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2, 3, 5 +[22] Feng Liang, Yangguang Li, and Diana Marculescu. Supmae: Supervised masked autoencoders are efficient vision learners. arXiv preprint arXiv:2205.14540, 2022. 6, 7 +[23] Jihao Liu, Xin Huang, Jinliang Zheng, Yu Liu, and Hongsheng Li. Mixmae: Mixed and masked autoencoder for efficient pretraining of hierarchical vision transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 3 +[24] Jian Meng, Li Yang, Kyungmin Lee, Jinwoo Shin, Deliang Fan, and Jae sun Seo. Slimmed asymmetrical contrastive learning and cross distillation for lightweight model training. In Thirty-seventh Conference on Neural Information Processing Systems (NeurIPS), 2023. 2 + +[25] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation Learning with Contrastive Predictive Coding. arXiv preprint arXiv:1807.03748, 2018. 2 +[26] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning (ICML), 2021. 1 +[27] Bowen Shi, Xiaopeng Zhang, Yaoming Wang, Jin Li, Wenrui Dai, Junni Zou, Hongkai Xiong, and Qi Tian. Hybrid distillation: Connecting masked autoencoders with contrastive learners. In The Twelfth International Conference on Learning Representations (ICLR), 2024. 3, 6 +[28] Yucheng Tang, Dong Yang, Wenqi Li, Holger R Roth, Bennett Landman, Daguang Xu, Vishwesh Nath, and Ali Hatamizadeh. Self-supervised pre-training of swim transformers for 3d medical image analysis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1 +[29] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In International Conference on Machine Learning (ICML), 2021. 2 +[30] Jue Wang, Gedas Bertasius, Du Tran, and Lorenzo Torresani. Long-short temporal contrastive learning of video transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (NeurIPS), 2022. 1 +[31] Shaoru Wang, Jin Gao, Zeming Li, Xiaoqin Zhang, and Weiming Hu. A closer look at self-supervised lightweight vision transformers. In International Conference on Machine Learning (ICML). PMLR, 2023. 2, 3 +[32] Chen Wei, Haoqi Fan, Saining Xie, Chao-Yuan Wu, Alan Yuille, and Christoph Feichtenhofer. Masked feature prediction for self-supervised visual pre-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 8 +[33] Kan Wu, Houwen Peng, Zhenghong Zhou, Bin Xiao, Mengchen Liu, Lu Yuan, Hong Xuan, Michael Valenzuela, Xi Stephen Chen, Xinggang Wang, et al. Tinyclip: Clip distillation via affinity mimicking and weight inheritance. In Proceedings of the IEEE/CVF International Conference on Computer Vision (CVPR), 2023. 1 +[34] Zhenda Xie, Zheng Zhang, Yue Cao, Yutong Lin, Jianmin Bao, Zhuliang Yao, Qi Dai, and Han Hu. Simmim: A simple framework for masked image modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 7 +[35] Sukmin Yun, Hankook Lee, Jaehung Kim, and Jinwoo Shin. Patch-level representation learning for self-supervised vision transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 5 +[36] Jure Zbontar, Li Jing, Ishan Misra, Yann LeCun, and Stéphane Deny. Barlow Twins: Self-supervised Learning via Redundancy Reduction. In International Conference on Machine Learning (ICML), pages 12310-12320, 2021. 1, 2, 4, 5 + +[37] Qi Zhang, Yifei Wang, and Yisen Wang. How mask matters: Towards theoretical understandings of masked autoencoders. Advances in Neural Information Processing Systems (NeurIPS), 2022. 1, 3, 4, 5, 7, 8 +[38] Zhiyu Zhao, Bingkun Huang, Sen Xing, Gangshan Wu, Yu Qiao, and Limin Wang. Asymmetric masked distillation for pre-training small foundation models. arXiv preprint arXiv:2311.03149, 2023. 2 \ No newline at end of file diff --git a/2025/Closest Neighbors are Harmful for Lightweight Masked Auto-encoders/images.zip b/2025/Closest Neighbors are Harmful for Lightweight Masked Auto-encoders/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..fa03b5cbc9b20da779037bf5e93b5ef59db9a31a --- /dev/null +++ b/2025/Closest Neighbors are Harmful for Lightweight Masked Auto-encoders/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2fb0f03776a92f35f1879a1ca9f07c013157f80b95d6414daddcd2b5c1f7cd7f +size 607560 diff --git a/2025/Closest Neighbors are Harmful for Lightweight Masked Auto-encoders/layout.json b/2025/Closest Neighbors are Harmful for Lightweight Masked Auto-encoders/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..b8d719a24436a6cce1f57ada4400a165d0374aae --- /dev/null +++ b/2025/Closest Neighbors are Harmful for Lightweight Masked Auto-encoders/layout.json @@ -0,0 +1,8453 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 86, + 103, + 523, + 120 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 103, + 523, + 120 + ], + "spans": [ + { + "bbox": [ + 86, + 103, + 523, + 120 + ], + "type": "text", + "content": "Closest Neighbors are Harmful for Lightweight Masked Auto-encoders" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 142, + 570, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 142, + 570, + 185 + ], + "spans": [ + { + "bbox": [ + 61, + 142, + 570, + 185 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 61, + 142, + 570, + 185 + ], + "type": "text", + "content": "Jian Meng, " + }, + { + "bbox": [ + 61, + 142, + 570, + 185 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 61, + 142, + 570, + 185 + ], + "type": "text", + "content": "Ahmed Hassssan, " + }, + { + "bbox": [ + 61, + 142, + 570, + 185 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 61, + 142, + 570, + 185 + ], + "type": "text", + "content": "Li Yang, " + }, + { + "bbox": [ + 61, + 142, + 570, + 185 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 61, + 142, + 570, + 185 + ], + "type": "text", + "content": "Deliang Fan, " + }, + { + "bbox": [ + 61, + 142, + 570, + 185 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 61, + 142, + 570, + 185 + ], + "type": "text", + "content": "Jinwoo Shin, and " + }, + { + "bbox": [ + 61, + 142, + 570, + 185 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 61, + 142, + 570, + 185 + ], + "type": "text", + "content": "Jae-sun Seo \n" + }, + { + "bbox": [ + 61, + 142, + 570, + 185 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 61, + 142, + 570, + 185 + ], + "type": "text", + "content": "Cornell University, " + }, + { + "bbox": [ + 61, + 142, + 570, + 185 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 61, + 142, + 570, + 185 + ], + "type": "text", + "content": "University of North Carolina at Charlotte, " + }, + { + "bbox": [ + 61, + 142, + 570, + 185 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 61, + 142, + 570, + 185 + ], + "type": "text", + "content": "Arizona State University, " + }, + { + "bbox": [ + 61, + 142, + 570, + 185 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 61, + 142, + 570, + 185 + ], + "type": "text", + "content": "KAIST \n" + }, + { + "bbox": [ + 61, + 142, + 570, + 185 + ], + "type": "inline_equation", + "content": "^{1}\\{jm2787, ah2288, js3528\\} @cornell.edu, ^{2}lyang50@uncc.edu, ^{3}dfan12@asu.edu, ^{4}jinwoos@kaist.ac.kr" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 152, + 213, + 200, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 213, + 200, + 225 + ], + "spans": [ + { + "bbox": [ + 152, + 213, + 200, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 238, + 297, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 238, + 297, + 597 + ], + "spans": [ + { + "bbox": [ + 53, + 238, + 297, + 597 + ], + "type": "text", + "content": "Learning the visual representation via masked autoencoder (MAE) training has been proven to be a powerful technique. Transferring the pre-trained vision transformer (ViT) to downstream tasks leads to superior performance compared to conventional task-by-task supervised learning. Recent research works on MAE focus on large-sized vision transformers (" + }, + { + "bbox": [ + 53, + 238, + 297, + 597 + ], + "type": "inline_equation", + "content": ">50" + }, + { + "bbox": [ + 53, + 238, + 297, + 597 + ], + "type": "text", + "content": " million parameters) with outstanding performance. However, improving the generality of the under-parametrized lightweight model has been widely ignored. In practice, downstream applications are commonly intended for resource-constrained platforms, where large-scale ViT cannot easily meet the resource budget. Current lightweight MAE training heavily relies on knowledge distillation with a pre-trained teacher, whereas the root cause behind the poor performance remains underexplored. Motivated by that, this paper first introduces the concept of \"closest neighbor patch\" to characterize the local semantics among the input tokens. Our discovery shows that the lightweight model failed to distinguish different local information, leading to biased understanding and poor accuracy. Motivated by this finding, we propose NoR-MAE, a novel MAE training algorithm for lightweight vision transformers. NoR-MAE elegantly repels the semantic aliasing between patches and their closest neighboring patch (semantic centroid) with negligible training cost overhead. With the ViT-Tiny model, NoR-MAE achieves up to " + }, + { + "bbox": [ + 53, + 238, + 297, + 597 + ], + "type": "inline_equation", + "content": "7.22\\% / 3.64\\%" + }, + { + "bbox": [ + 53, + 238, + 297, + 597 + ], + "type": "text", + "content": " accuracy improvements on ImageNet100/ImageNet-1K datasets, as well as up to " + }, + { + "bbox": [ + 53, + 238, + 297, + 597 + ], + "type": "inline_equation", + "content": "5.13\\%" + }, + { + "bbox": [ + 53, + 238, + 297, + 597 + ], + "type": "text", + "content": " accuracy improvements in tested downstream tasks. https://github.com/SeoLabCornell/NoR-MAE" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 620, + 135, + 633 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 620, + 135, + 633 + ], + "spans": [ + { + "bbox": [ + 56, + 620, + 135, + 633 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 641, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 296, + 713 + ], + "type": "text", + "content": "Starting from the early exploration with contrastive learning [6, 18], learning powerful and generic vision or semantic representation has been the focal point of all prior self-supervised learning (SSL) algorithms across various machine learning domains [2, 8, 28, 30] and vision-language model [15, 26]. The superiority of SSL is demonstrated" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 316, + 210, + 555, + 358 + ], + "blocks": [ + { + "bbox": [ + 316, + 210, + 555, + 358 + ], + "lines": [ + { + "bbox": [ + 316, + 210, + 555, + 358 + ], + "spans": [ + { + "bbox": [ + 316, + 210, + 555, + 358 + ], + "type": "image", + "image_path": "74bc6eb3dbc3e2a1a4f6f4bf024bd78d64fd939e071375606bdf94fc097022c7.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 366, + 556, + 433 + ], + "lines": [ + { + "bbox": [ + 313, + 366, + 556, + 433 + ], + "spans": [ + { + "bbox": [ + 313, + 366, + 556, + 433 + ], + "type": "text", + "content": "Figure 1. Top: Closest Neighbor Patches (CNP) characterizes the \"Local Semantic Centroid\" among the input tokens. Bottom: Low-dimensional projection (via PCA) of the transformer-encoded patches. Unlike the large-sized transformer model (e.g., ViT-Base), lightweight models (e.g., ViT-Tiny) cannot separate the \"semantic difference\" between different CNPs." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 462, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 462, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 462, + 556, + 715 + ], + "type": "text", + "content": "by the capability of extracting transferable knowledge from unlabeled datasets [26, 33]. Recent efforts on masked autoencoder (MAE) [19] train the vision transformer [11] by encouraging the model to understand the semantic knowledge of the masked patches throughout the reconstruction-based self-supervised learning. Particularly, the evolution from contrastive learning [16, 18, 36] to MAE-based training [17, 19, 34, 37] shows consistent dominance by large-scale models to ensure on-par or even better accuracy compared to supervised learning, along with superior performance in various downstream applications. In practice, downstream small-scale tasks are commonly deployed on edge devices with limited resources. Therefore, large-sized vision transformers (ViTs) pre-trained by MAE [19] are suboptimal candidates for resource-constrained downstream deployment and hardware platforms. Unfortunately, recent state-of-the-art (SoTA) methods that train lightweight ViTs with MAE from scratch lead to poor performance and insufficient transferability. For example, the MAE-trained lightweight ViT-Tiny model only achieves " + }, + { + "bbox": [ + 313, + 462, + 556, + 715 + ], + "type": "inline_equation", + "content": "66.6\\%" + }, + { + "bbox": [ + 313, + 462, + 556, + 715 + ], + "type": "text", + "content": " top-1 accuracy (with 200-epoch pre-training [3, 19, 37]) on the" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "type": "text", + "content": "25230" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": "ImageNet-1K dataset [3], which is " + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "inline_equation", + "content": "\\sim 10\\%" + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": " lower than the supervised learning counterpart. Evidently, the superiority of MAE [19] with the large model failed to be maintained in the lightweight ViTs [11]. Given the diverse downstream tasks and urgent need for energy-efficient foundation models, unleashing the power of MAE and enhancing the versatility of lightweight models is desired." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 157, + 298, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 157, + 298, + 407 + ], + "spans": [ + { + "bbox": [ + 56, + 157, + 298, + 407 + ], + "type": "text", + "content": "Prior works enhance the performance of the lightweight transformer models by heavily relying on knowledge distillation (KD) [20] for both supervised [29] and self-supervised learning paradigms [3, 21, 38]. For instance, the recent DMAE work [3] combines MAE [19] with KD and transfers the pre-trained knowledge from a ViT-Base teacher down to the lightweight ViT model. DMAE [3] chooses the 3/4 depth of both student and teacher ViT models as the bonding portal for feature alignment and knowledge distillation. The heuristic design of DMAE cannot be easily extended to other encoder models. Furthermore, distilling the knowledge on the fly introduces an additional projector [3] or logits mask [31], which further elevates the complexity and computation cost of the entire training process. In addition to DMAE, other recent works further increase the complexity of the KD-based MAE by employing multi-layer distillation [38] or complex generic-to-specific distillation [21]. All of these prior works [3, 21, 31, 38] rely on the brute force approach to enhance the model performance with increasingly complex training schemes, which results in inflated GPU memory and training time." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 408, + 296, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 408, + 296, + 480 + ], + "spans": [ + { + "bbox": [ + 55, + 408, + 296, + 480 + ], + "type": "text", + "content": "However, none of these prior works can collectively achieve 1) the capability of training lightweight model from scratch without a large-sized teacher model, 2) negligible training cost overhead compared to the vanilla MAE, and 3) clear insights that cause the poor performance of lightweight MAE. To overcome this, the key questions to address are:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 483, + 295, + 529 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 55, + 483, + 295, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 483, + 295, + 506 + ], + "spans": [ + { + "bbox": [ + 55, + 483, + 295, + 506 + ], + "type": "text", + "content": "① What is the bottleneck of MAE with under parametrized small vision transformer models?" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 506, + 295, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 506, + 295, + 529 + ], + "spans": [ + { + "bbox": [ + 55, + 506, + 295, + 529 + ], + "type": "text", + "content": "② How to efficiently train a high performance lightweight model from scratch via MAE?" + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 533, + 295, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 533, + 295, + 675 + ], + "spans": [ + { + "bbox": [ + 55, + 533, + 295, + 675 + ], + "type": "text", + "content": "To answer ①, our investigation shows that the lightweight model failed to understand different local semantic information among the input tokens and lead to poor performance. We characterize different local semantics by introducing the concept of closest neighborhood patch (CNP) (Figure 1), which carries the most similar semantics of a given input patch. Since multiple patches can share one CNP (e.g., Background Forrest), different CNPs represent various local semantics across the entire input sequence. Unlike large-sized models, the under-parameterized model falls into the semantic aliasing between different CNPs with poor accuracy, as indicated in Figure 1 and Section 3.3." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "Motivated by the findings in ①, we propose the Neighbor Repelling Masked Autoencoder (NoR-MAE) to resolve ②, a novel self-supervised learning algorithm designed for train" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 555, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 203 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 203 + ], + "type": "text", + "content": "ing powerful lightweight vision transformers. NoR-MAE trains the powerful small vision transformer by penalizing the similarity between the unmasked patches and CNPs. Different from prior works, NoR-MAE directly trains the vision transformer from scratch without introducing the largesized pre-trained teacher or heuristically designed distillation scheme. The proposed method achieves up to " + }, + { + "bbox": [ + 313, + 72, + 555, + 203 + ], + "type": "inline_equation", + "content": "7.22\\%" + }, + { + "bbox": [ + 313, + 72, + 555, + 203 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 72, + 555, + 203 + ], + "type": "inline_equation", + "content": "3.64\\%" + }, + { + "bbox": [ + 313, + 72, + 555, + 203 + ], + "type": "text", + "content": " accuracy improvements on ImageNet-100 and ImageNet-1K datasets, with negligible training time and memory overhead compared to the vanilla MAE. The major contributions of our work are:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 204, + 555, + 431 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 314, + 204, + 555, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 204, + 555, + 239 + ], + "spans": [ + { + "bbox": [ + 314, + 204, + 555, + 239 + ], + "type": "text", + "content": "- Simplicity: NoR-MAE does not require teacher pretraining or heuristic knowledge distillation to train a powerful lightweight ViT." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 239, + 555, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 239, + 555, + 299 + ], + "spans": [ + { + "bbox": [ + 313, + 239, + 555, + 299 + ], + "type": "text", + "content": "- High Performance: Compared to the baseline MAE, NoR-MAE achieves " + }, + { + "bbox": [ + 313, + 239, + 555, + 299 + ], + "type": "inline_equation", + "content": "7.22\\%" + }, + { + "bbox": [ + 313, + 239, + 555, + 299 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 239, + 555, + 299 + ], + "type": "inline_equation", + "content": "3.64\\%" + }, + { + "bbox": [ + 313, + 239, + 555, + 299 + ], + "type": "text", + "content": " accuracy improvements on ImageNet-100 and ImageNet-1K datasets with the ViT-Tiny model, achieving the new SoTA performance on the lightweight MAE training." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 300, + 555, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 300, + 555, + 370 + ], + "spans": [ + { + "bbox": [ + 313, + 300, + 555, + 370 + ], + "type": "text", + "content": "- Transferrability: Different from prior works, which mainly focus on fine-tuning the performance on the pretraining dataset (e.g., ImageNet), NoR-MAE is also evaluated with the downstream fine-tuning and linear evaluation benchmarks, with significantly improved performance compared to prior SSL baselines and the vanilla MAE." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 372, + 553, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 372, + 553, + 431 + ], + "spans": [ + { + "bbox": [ + 313, + 372, + 553, + 431 + ], + "type": "text", + "content": "- Rationality: Besides the superior performance, for the first time, NoR-MAE reveals the semantic aliasing issue caused by the neighboring patches, which is proved to be the reason that hinders efficient training and representation learning of lightweight ViT models." + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 442, + 400, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 442, + 400, + 453 + ], + "spans": [ + { + "bbox": [ + 314, + 442, + 400, + 453 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 462, + 555, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 462, + 555, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 462, + 555, + 605 + ], + "type": "text", + "content": "Empowering a deep neural network model with strong and transferable knowledge has been widely investigated, especially in self-supervised learning. Early research works in contrastive learning (CL) introduce the learning paradigm with \"positive\" and \"negative\" sample pairs [6, 18], together with the alignment-repellent strategy embedded into InfoNCE [7, 18, 25] and NT-Xent loss [6] as the learning objective. The early success of CL motivated the successors to investigate the potential of different training and sampling strategies, inducing the standpoint of asymmetrical learning [16, 24], knowledge distillation [13], and the latent-dimension correlation [36]." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 606, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 606, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 606, + 556, + 714 + ], + "type": "text", + "content": "While the core idea of CL is training the model to understand the similarities and differences between different augmentations, the emergence of masked autoencoder (MAE) [19] training reconsiders SSL from a different perspective, which is perfectly suitable for vision transformers. Encouraging the encoder-decoder to reconstruct the randomly sparsified patches allows the model to understand the semantics. Due to the powerful visual representation, MAE-based pretaining exhibits strong performance on" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "type": "text", + "content": "25231" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 71, + 292, + 135 + ], + "blocks": [ + { + "bbox": [ + 59, + 71, + 292, + 135 + ], + "lines": [ + { + "bbox": [ + 59, + 71, + 292, + 135 + ], + "spans": [ + { + "bbox": [ + 59, + 71, + 292, + 135 + ], + "type": "image", + "image_path": "aa9725bd0ee949afb6e6d813dfebc077932fe2a507d0e1ad34fa4413310f62ff.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 64, + 145, + 286, + 157 + ], + "lines": [ + { + "bbox": [ + 64, + 145, + 286, + 157 + ], + "spans": [ + { + "bbox": [ + 64, + 145, + 286, + 157 + ], + "type": "text", + "content": "Figure 2. Closest Neighbor Patches (CNP) within the image." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 185, + 296, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 185, + 296, + 245 + ], + "spans": [ + { + "bbox": [ + 55, + 185, + 296, + 245 + ], + "type": "text", + "content": "both backbone datasets (e.g., ImageNet-1K) and downstream transfer learning tasks. Subsequent research works further improved the performance of MAE from the perspectives of contrastive learning [27, 37], cross-attention [14, 17], mixed feature learning [23], or intricate tokenizer [12]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 249, + 296, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 249, + 296, + 597 + ], + "spans": [ + { + "bbox": [ + 56, + 249, + 296, + 597 + ], + "type": "text", + "content": "Across all the different MAE training methods, elevating the performance of the large-sized models (e.g., ViT-Large) has been the primary focus of exploration, while improving the poor performance of the small-sized model (e.g., ViTTiny) has been largely ignored. Given that the downstream tasks are well-suited for resource-constrained edge devices, the imbalanced research focus of the prior works largely hinders the efficiency and transferability of MAE [19]. The practical needs and the imperfection of lightweight MAE motivate the recent research works to investigate MAE training strategies with small models. Following the protocol from supervised distillation [20], knowledge distillation-based MAE [3, 21] is proposed to enhance the performance of the lightweight student. However, the learning scheme of the single-layer or multi-layer distillation is established based on the heuristically designed distillation between intermediate features. Compared to the supervised knowledge distillation, recent studies largely complicate the overall cost of training lightweight ViT models with either heuristic design [3], multi-phased distillation [21], or distillation with extensive fine-tuning [31]. In particular, employing the pre-trained large ViT model as the teacher introduces additional GPU memory usage and time cost toward the total training efforts. Furthermore, training the interconnect projectors between the student and teacher exacerbates the training complexity even further. More importantly, the knowledge distillation-based MAE inherits the essence of KD from supervised learning, but the insights of training lightweight ViT from scratch via MAE remain largely unexplored." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 617, + 162, + 631 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 162, + 631 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 162, + 631 + ], + "type": "text", + "content": "3. Proposed Method" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": "We propose the Neighbor-Repelling MAE (NoR-MAE). In this section, we will 1) introduce the concept of the Closest Neighbor Patch into MAE training, 2) present the proposed NoR-MAE training, and 3) unravel the rationality of NoR-MAE by revealing the insights that cause the poor performance of the directly-trained lightweight ViT model." + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 316, + 71, + 553, + 207 + ], + "blocks": [ + { + "bbox": [ + 316, + 71, + 553, + 207 + ], + "lines": [ + { + "bbox": [ + 316, + 71, + 553, + 207 + ], + "spans": [ + { + "bbox": [ + 316, + 71, + 553, + 207 + ], + "type": "image", + "image_path": "fac5c4699b24f06121d93a3717d4ec1f9a6ccdf8d196fae1fd81ffb59723d3dc.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 216, + 555, + 239 + ], + "lines": [ + { + "bbox": [ + 313, + 216, + 555, + 239 + ], + "spans": [ + { + "bbox": [ + 313, + 216, + 555, + 239 + ], + "type": "text", + "content": "Figure 3. Cosine Similarity between each patch and their Closest Neighbor Patch across a subset of ImageNet-1K with 4096 images." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 262, + 479, + 275 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 262, + 479, + 275 + ], + "spans": [ + { + "bbox": [ + 313, + 262, + 479, + 275 + ], + "type": "text", + "content": "3.1. Closest Neighbor Patch (CNP)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 281, + 555, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 281, + 555, + 378 + ], + "spans": [ + { + "bbox": [ + 313, + 281, + 555, + 378 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 313, + 281, + 555, + 378 + ], + "type": "inline_equation", + "content": "X \\in \\mathbf{R}^{H \\times W \\times C}" + }, + { + "bbox": [ + 313, + 281, + 555, + 378 + ], + "type": "text", + "content": " be an input image to the model, where " + }, + { + "bbox": [ + 313, + 281, + 555, + 378 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 313, + 281, + 555, + 378 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 281, + 555, + 378 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 313, + 281, + 555, + 378 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 281, + 555, + 378 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 313, + 281, + 555, + 378 + ], + "type": "text", + "content": " represent the height, width, and channel of the image. Vision transformer (ViT) reformulates the input by decomposing the image " + }, + { + "bbox": [ + 313, + 281, + 555, + 378 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 313, + 281, + 555, + 378 + ], + "type": "text", + "content": " into a sequence of non-overlapped patches with the pre-defined patch size " + }, + { + "bbox": [ + 313, + 281, + 555, + 378 + ], + "type": "inline_equation", + "content": "P \\times P" + }, + { + "bbox": [ + 313, + 281, + 555, + 378 + ], + "type": "text", + "content": " and the total number of patches (tokens) " + }, + { + "bbox": [ + 313, + 281, + 555, + 378 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 281, + 555, + 378 + ], + "type": "text", + "content": ". All the patches are further encoded into the embeddings with dimension " + }, + { + "bbox": [ + 313, + 281, + 555, + 378 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 313, + 281, + 555, + 378 + ], + "type": "text", + "content": ", which is the input of the subsequent transformer blocks." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 380, + 556, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 380, + 556, + 452 + ], + "spans": [ + { + "bbox": [ + 313, + 380, + 556, + 452 + ], + "type": "text", + "content": "For a given input patch " + }, + { + "bbox": [ + 313, + 380, + 556, + 452 + ], + "type": "inline_equation", + "content": "X_P" + }, + { + "bbox": [ + 313, + 380, + 556, + 452 + ], + "type": "text", + "content": ", there are naturally 8 neighbors that share a common edge with " + }, + { + "bbox": [ + 313, + 380, + 556, + 452 + ], + "type": "inline_equation", + "content": "X_P" + }, + { + "bbox": [ + 313, + 380, + 556, + 452 + ], + "type": "text", + "content": ", as shown in Figure 2(a). We define the Closest Neighbor Patch (CNP) " + }, + { + "bbox": [ + 313, + 380, + 556, + 452 + ], + "type": "inline_equation", + "content": "X_{CNP}" + }, + { + "bbox": [ + 313, + 380, + 556, + 452 + ], + "type": "text", + "content": " by selecting the neighbor that has the Top-1 cosine similarity with " + }, + { + "bbox": [ + 313, + 380, + 556, + 452 + ], + "type": "inline_equation", + "content": "X_P" + }, + { + "bbox": [ + 313, + 380, + 556, + 452 + ], + "type": "text", + "content": ". In other words, " + }, + { + "bbox": [ + 313, + 380, + 556, + 452 + ], + "type": "inline_equation", + "content": "X_P" + }, + { + "bbox": [ + 313, + 380, + 556, + 452 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 380, + 556, + 452 + ], + "type": "inline_equation", + "content": "X_{CNP}" + }, + { + "bbox": [ + 313, + 380, + 556, + 452 + ], + "type": "text", + "content": " are sharing the highest degree of semantically meaningful relations." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 453, + 556, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 453, + 556, + 544 + ], + "spans": [ + { + "bbox": [ + 313, + 453, + 556, + 544 + ], + "type": "text", + "content": "Shared Closest Neighbor Patch. In the example of Figure 2, two patches " + }, + { + "bbox": [ + 313, + 453, + 556, + 544 + ], + "type": "inline_equation", + "content": "X_P^i" + }, + { + "bbox": [ + 313, + 453, + 556, + 544 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 453, + 556, + 544 + ], + "type": "inline_equation", + "content": "X_P^j" + }, + { + "bbox": [ + 313, + 453, + 556, + 544 + ], + "type": "text", + "content": " partially share the neighborhoods in between, while the CNP of " + }, + { + "bbox": [ + 313, + 453, + 556, + 544 + ], + "type": "inline_equation", + "content": "X_P^i" + }, + { + "bbox": [ + 313, + 453, + 556, + 544 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 453, + 556, + 544 + ], + "type": "inline_equation", + "content": "X_P^j" + }, + { + "bbox": [ + 313, + 453, + 556, + 544 + ], + "type": "text", + "content": " happen to be identical as " + }, + { + "bbox": [ + 313, + 453, + 556, + 544 + ], + "type": "inline_equation", + "content": "X_{CNP}^{ij}" + }, + { + "bbox": [ + 313, + 453, + 556, + 544 + ], + "type": "text", + "content": ". Naturally, " + }, + { + "bbox": [ + 313, + 453, + 556, + 544 + ], + "type": "inline_equation", + "content": "X_{CNP}^{ij}" + }, + { + "bbox": [ + 313, + 453, + 556, + 544 + ], + "type": "text", + "content": " carries the shared semantic information (\"Background Forest\") for both " + }, + { + "bbox": [ + 313, + 453, + 556, + 544 + ], + "type": "inline_equation", + "content": "X_P^i" + }, + { + "bbox": [ + 313, + 453, + 556, + 544 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 453, + 556, + 544 + ], + "type": "inline_equation", + "content": "X_P^j" + }, + { + "bbox": [ + 313, + 453, + 556, + 544 + ], + "type": "text", + "content": ", but the degree of similarity between " + }, + { + "bbox": [ + 313, + 453, + 556, + 544 + ], + "type": "inline_equation", + "content": "(X_P^i, X_{CNP}^{ij})" + }, + { + "bbox": [ + 313, + 453, + 556, + 544 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 453, + 556, + 544 + ], + "type": "inline_equation", + "content": "(X_P^j, X_{CNP}^{ij})" + }, + { + "bbox": [ + 313, + 453, + 556, + 544 + ], + "type": "text", + "content": " are different." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 544, + 556, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 544, + 556, + 639 + ], + "spans": [ + { + "bbox": [ + 313, + 544, + 556, + 639 + ], + "type": "text", + "content": "Closest Neighbor Patch " + }, + { + "bbox": [ + 313, + 544, + 556, + 639 + ], + "type": "inline_equation", + "content": "\\neq" + }, + { + "bbox": [ + 313, + 544, + 556, + 639 + ], + "type": "text", + "content": " Identical Patch. Although " + }, + { + "bbox": [ + 313, + 544, + 556, + 639 + ], + "type": "inline_equation", + "content": "X_{P}" + }, + { + "bbox": [ + 313, + 544, + 556, + 639 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 544, + 556, + 639 + ], + "type": "inline_equation", + "content": "X_{CNP}" + }, + { + "bbox": [ + 313, + 544, + 556, + 639 + ], + "type": "text", + "content": " are sharing the highest degree of similarity, the semantics between " + }, + { + "bbox": [ + 313, + 544, + 556, + 639 + ], + "type": "inline_equation", + "content": "X_{P}" + }, + { + "bbox": [ + 313, + 544, + 556, + 639 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 544, + 556, + 639 + ], + "type": "inline_equation", + "content": "X_{CNP}" + }, + { + "bbox": [ + 313, + 544, + 556, + 639 + ], + "type": "text", + "content": " are not identical. As shown in Figure 3, the cosine similarity between all the patches and their closest neighbor in ImageNet-1K images varies from 0.07 to 0.999, with an average of 0.83. In other words, the highly similar neighboring patches also contain the implicit semantical difference." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 641, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 556, + 713 + ], + "type": "text", + "content": "Difficulty of Lightweight Models to Understand the Different Semantics Carried by CNP. With MAE, the input patches of the transformer models are randomly masked with a pre-defined sparsity ratio " + }, + { + "bbox": [ + 313, + 641, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 641, + 556, + 713 + ], + "type": "text", + "content": " (e.g., " + }, + { + "bbox": [ + 313, + 641, + 556, + 713 + ], + "type": "inline_equation", + "content": "75\\%" + }, + { + "bbox": [ + 313, + 641, + 556, + 713 + ], + "type": "text", + "content": "), as illustrated in Figure 4, left. Naturally, each unmasked visible patch has a Closest Neighbor Patch (CNP). As a result, the CNP" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 733, + 318, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 733, + 318, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 733, + 318, + 742 + ], + "type": "text", + "content": "25232" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 71, + 294, + 159 + ], + "blocks": [ + { + "bbox": [ + 58, + 71, + 294, + 159 + ], + "lines": [ + { + "bbox": [ + 58, + 71, + 294, + 159 + ], + "spans": [ + { + "bbox": [ + 58, + 71, + 294, + 159 + ], + "type": "image", + "image_path": "fe7c07b0b0d88cc7ead5211fbdf2d18f93ea04eb765971248ccbd407a769717e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 170, + 295, + 194 + ], + "lines": [ + { + "bbox": [ + 55, + 170, + 295, + 194 + ], + "spans": [ + { + "bbox": [ + 55, + 170, + 295, + 194 + ], + "type": "text", + "content": "Figure 4. The shared Closest Neighbor Patch (CNP) characterizes the estimated local semantic centroids of the input sequence." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 215, + 296, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 215, + 296, + 370 + ], + "spans": [ + { + "bbox": [ + 54, + 215, + 296, + 370 + ], + "type": "text", + "content": "of the unmasked patches formulates another sparse input image (Figure 4, right). From the perspective of the entire image, visible patches and their CNPs are close neighbors, while different visible patches may share the same closest neighbor, which estimates the local and regional semantics of the input sequence. However, as described in Figure 1 and Section 1, understanding the semantic difference between different CNPs is the bottleneck for the under-parameterized models. Furthermore, our extended experiments in Figure 7 show that the key difference between the big (e.g., ViT-Base) and small (e.g., ViT-Tiny) model is understanding and separating the semantic differences between CNPs. Therefore, the challenge of lightweight MAE is clear:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 371, + 296, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 371, + 296, + 396 + ], + "spans": [ + { + "bbox": [ + 55, + 371, + 296, + 396 + ], + "type": "text", + "content": "In MAE training, how to facilitate the lightweight model to understand different local semantics?" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 404, + 220, + 417 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 404, + 220, + 417 + ], + "spans": [ + { + "bbox": [ + 55, + 404, + 220, + 417 + ], + "type": "text", + "content": "3.2. Proposed Method: NoR-MAE" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 424, + 296, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 424, + 296, + 519 + ], + "spans": [ + { + "bbox": [ + 55, + 424, + 296, + 519 + ], + "type": "text", + "content": "We propose the Neighbor-Repelling training for self-supervised MAE learning (NoR-MAE). The overall objective is to train a powerful, lightweight vision transformer by encouraging the model to avoid the biased relation between different local semantics of the input. We follow the standard protocol of the vanilla MAE with the encoder " + }, + { + "bbox": [ + 55, + 424, + 296, + 519 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 55, + 424, + 296, + 519 + ], + "type": "text", + "content": " and decoder " + }, + { + "bbox": [ + 55, + 424, + 296, + 519 + ], + "type": "inline_equation", + "content": "g_{\\phi}" + }, + { + "bbox": [ + 55, + 424, + 296, + 519 + ], + "type": "text", + "content": ", where both " + }, + { + "bbox": [ + 55, + 424, + 296, + 519 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 55, + 424, + 296, + 519 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 424, + 296, + 519 + ], + "type": "inline_equation", + "content": "g_{\\phi}" + }, + { + "bbox": [ + 55, + 424, + 296, + 519 + ], + "type": "text", + "content": " are vision transformers with the same embedding dimension " + }, + { + "bbox": [ + 55, + 424, + 296, + 519 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 55, + 424, + 296, + 519 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 521, + 296, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 296, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 296, + 628 + ], + "type": "text", + "content": "Given the masking ratio " + }, + { + "bbox": [ + 55, + 521, + 296, + 628 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 55, + 521, + 296, + 628 + ], + "type": "text", + "content": ", we first generate the masked input sequence " + }, + { + "bbox": [ + 55, + 521, + 296, + 628 + ], + "type": "inline_equation", + "content": "x^{\\tau}" + }, + { + "bbox": [ + 55, + 521, + 296, + 628 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 521, + 296, + 628 + ], + "type": "inline_equation", + "content": "x_{CNP}^{\\tau}" + }, + { + "bbox": [ + 55, + 521, + 296, + 628 + ], + "type": "text", + "content": " that are embedded from the unmasked patches with the corresponding closest neighbor patches, respectively. During the forward pass of each iteration, we parallelly encode and decode the masked sequence " + }, + { + "bbox": [ + 55, + 521, + 296, + 628 + ], + "type": "inline_equation", + "content": "x^{\\tau}" + }, + { + "bbox": [ + 55, + 521, + 296, + 628 + ], + "type": "text", + "content": " and CNP sequence " + }, + { + "bbox": [ + 55, + 521, + 296, + 628 + ], + "type": "inline_equation", + "content": "x_{CNP}^{\\tau}" + }, + { + "bbox": [ + 55, + 521, + 296, + 628 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 55, + 521, + 296, + 628 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 55, + 521, + 296, + 628 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 521, + 296, + 628 + ], + "type": "inline_equation", + "content": "g_{\\phi}" + }, + { + "bbox": [ + 55, + 521, + 296, + 628 + ], + "type": "text", + "content": ", generating the reconstructed embedding " + }, + { + "bbox": [ + 55, + 521, + 296, + 628 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 55, + 521, + 296, + 628 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 521, + 296, + 628 + ], + "type": "inline_equation", + "content": "Z_{CNP}" + }, + { + "bbox": [ + 55, + 521, + 296, + 628 + ], + "type": "text", + "content": ", respectively. In particular, " + }, + { + "bbox": [ + 55, + 521, + 296, + 628 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 55, + 521, + 296, + 628 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 521, + 296, + 628 + ], + "type": "inline_equation", + "content": "g_{\\phi}" + }, + { + "bbox": [ + 55, + 521, + 296, + 628 + ], + "type": "text", + "content": " are served as the siamese encoder and decoder (with stop gradient) during the forward pass of " + }, + { + "bbox": [ + 55, + 521, + 296, + 628 + ], + "type": "inline_equation", + "content": "Z_{CNP}" + }, + { + "bbox": [ + 55, + 521, + 296, + 628 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 629, + 296, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 296, + 653 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 296, + 653 + ], + "type": "text", + "content": "Overall, each forward pass will generate two pairs of decoded tokens resulting from " + }, + { + "bbox": [ + 55, + 629, + 296, + 653 + ], + "type": "inline_equation", + "content": "x^{\\tau}" + }, + { + "bbox": [ + 55, + 629, + 296, + 653 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 629, + 296, + 653 + ], + "type": "inline_equation", + "content": "x_{\\mathrm{CNP}}^{\\tau}" + }, + { + "bbox": [ + 55, + 629, + 296, + 653 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "content": "Following the standard training protocol of MAE, we first compute the " + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_2" + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "content": " reconstruction loss between the decoded sequence " + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "content": " and the ground truth. Subsequently, we compute the matrix multiplication between the normalized " + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "inline_equation", + "content": "Z_{CNP}" + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "content": ". As shown in Figure 5, the encoder " + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "content": " and decoder" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "inline_equation", + "content": "g_{\\phi}" + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "text", + "content": " are shared for both branches, while the stop gradient is enabled for the CNP input. Given the size of " + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "inline_equation", + "content": "[N,K,D]" + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "text", + "content": ", the matrix multiplication between " + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "inline_equation", + "content": "\\bar{Z}" + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "text", + "content": " and transposed " + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "inline_equation", + "content": "\\bar{Z}_{\\mathrm{CNP}}" + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "text", + "content": " is performed along the token dimension " + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "text", + "content": ". The resultant matrix is further scaled down by " + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "text", + "content": " and takes the average along the batch dimension " + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "text", + "content": ". Mathematically, we have:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 358, + 152, + 555, + 180 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 152, + 555, + 180 + ], + "spans": [ + { + "bbox": [ + 358, + 152, + 555, + 180 + ], + "type": "interline_equation", + "content": "C _ {\\text {n e i g h b o r}} = \\frac {1}{N} \\sum_ {N} \\left(\\frac {1}{K} (\\bar {Z}) ^ {T} \\cdot \\bar {Z} _ {\\mathrm {C N P}}\\right) \\tag {1}", + "image_path": "99877e3333551e82789823e86f271b8332f40745340022b7bf7b51a909c1771f.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 184, + 555, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 184, + 555, + 281 + ], + "spans": [ + { + "bbox": [ + 313, + 184, + 555, + 281 + ], + "type": "text", + "content": "Where " + }, + { + "bbox": [ + 313, + 184, + 555, + 281 + ], + "type": "inline_equation", + "content": "C_{\\mathrm{neighbor}}" + }, + { + "bbox": [ + 313, + 184, + 555, + 281 + ], + "type": "text", + "content": " represents the resultant relation matrix with the size of " + }, + { + "bbox": [ + 313, + 184, + 555, + 281 + ], + "type": "inline_equation", + "content": "D \\times D" + }, + { + "bbox": [ + 313, + 184, + 555, + 281 + ], + "type": "text", + "content": ". Each entry " + }, + { + "bbox": [ + 313, + 184, + 555, + 281 + ], + "type": "inline_equation", + "content": "c_{i,j}" + }, + { + "bbox": [ + 313, + 184, + 555, + 281 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 313, + 184, + 555, + 281 + ], + "type": "inline_equation", + "content": "C_{\\mathrm{neighbor}}" + }, + { + "bbox": [ + 313, + 184, + 555, + 281 + ], + "type": "text", + "content": " characterizes the correlation between the randomly-unmasked input sequence (output logits " + }, + { + "bbox": [ + 313, + 184, + 555, + 281 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 313, + 184, + 555, + 281 + ], + "type": "text", + "content": ") and the corresponding CNP sequence (output logits " + }, + { + "bbox": [ + 313, + 184, + 555, + 281 + ], + "type": "inline_equation", + "content": "Z_{\\mathrm{CNP}}" + }, + { + "bbox": [ + 313, + 184, + 555, + 281 + ], + "type": "text", + "content": ") between embedding dimension " + }, + { + "bbox": [ + 313, + 184, + 555, + 281 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 184, + 555, + 281 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 184, + 555, + 281 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 184, + 555, + 281 + ], + "type": "text", + "content": ". To minimize the semantic aliasing between the unmasked tokens and their CNPs, we decorate different embeddings by penalizing the off-diagonal terms of " + }, + { + "bbox": [ + 313, + 184, + 555, + 281 + ], + "type": "inline_equation", + "content": "C_{\\mathrm{neighbor}}" + }, + { + "bbox": [ + 313, + 184, + 555, + 281 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 383, + 289, + 555, + 316 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 383, + 289, + 555, + 316 + ], + "spans": [ + { + "bbox": [ + 383, + 289, + 555, + 316 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {N o R}} = \\sum_ {i} \\sum_ {i \\neq j} C _ {\\text {n e i g h b o r}} ^ {i, j} \\tag {2}", + "image_path": "c7ec491f90a0258a6de91003d3ba6d3b666ff386cdc6003bd5124bcd4e0027a0.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 323, + 556, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 323, + 556, + 433 + ], + "spans": [ + { + "bbox": [ + 313, + 323, + 556, + 433 + ], + "type": "text", + "content": "Essentially, the proposed NoR-Loss " + }, + { + "bbox": [ + 313, + 323, + 556, + 433 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{\\mathrm{NoR}})" + }, + { + "bbox": [ + 313, + 323, + 556, + 433 + ], + "type": "text", + "content": " minimizes semantic aliasing by preventing the model encoding the visible patches and CNP as aliased embeddings. It has been shown in Figure 3 that patches and their CNPs share similar but non-identical semantics. We balance such a tradeoff of \"similarity-discrepancy\" by decorrelating the off-diagonal embeddings only. The overall loss is the combination of the standard " + }, + { + "bbox": [ + 313, + 323, + 556, + 433 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_2" + }, + { + "bbox": [ + 313, + 323, + 556, + 433 + ], + "type": "text", + "content": " reconstruction loss and the proposed NeighborRepelling Loss " + }, + { + "bbox": [ + 313, + 323, + 556, + 433 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{NoR}}" + }, + { + "bbox": [ + 313, + 323, + 556, + 433 + ], + "type": "text", + "content": " scaled by a hyperparameter " + }, + { + "bbox": [ + 313, + 323, + 556, + 433 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 323, + 556, + 433 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 322, + 439, + 555, + 454 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 439, + 555, + 454 + ], + "spans": [ + { + "bbox": [ + 322, + 439, + 555, + 454 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {t o t a l}} = \\left| \\left| \\text {G r o u n d T r u t h} - \\bar {Z} \\right| \\right| _ {2} + \\lambda \\cdot \\mathcal {L} _ {\\mathrm {N o R}} (\\bar {Z}, \\bar {Z} _ {\\mathrm {C N P}}) \\tag {3}", + "image_path": "eb2f8fd0eac3f92af8cbd64d2d0cdb6a98d389b25602f5573feaf3a07ed3e356.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 462, + 556, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 462, + 556, + 545 + ], + "spans": [ + { + "bbox": [ + 313, + 462, + 556, + 545 + ], + "type": "text", + "content": "More importantly, decorrelating embeddings eventually separate the aliased semantics between different CNPs and their corresponding \"local semantic centroids\" (Section 3.3). In other words, the embedding-level anti-aliasing regularization of the proposed NoR-Loss facilitates the understanding on the representation (token) level, leading to the largely improved model performance." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 546, + 556, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 546, + 556, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 546, + 556, + 653 + ], + "type": "text", + "content": "Decorrelating Embeddings vs. Decorrelating Tokens. Although the proposed NoR-MAE selects the Top-1 neighbor (the closest neighbor patch (CNP)) as the candidate for decorrelation, the CNP patches can be overlapped or shared between each other. For instance, the closest neighbor patch of patch " + }, + { + "bbox": [ + 313, + 546, + 556, + 653 + ], + "type": "inline_equation", + "content": "X_{i}" + }, + { + "bbox": [ + 313, + 546, + 556, + 653 + ], + "type": "text", + "content": " could also be the Top-2/3/4 ... 8 neighbor of patch " + }, + { + "bbox": [ + 313, + 546, + 556, + 653 + ], + "type": "inline_equation", + "content": "X_{j}" + }, + { + "bbox": [ + 313, + 546, + 556, + 653 + ], + "type": "text", + "content": ". Minimizing the semantic aliasing by naively decorrelating tokens leads to collapsed performance and unsuccessful training, as shown in Figure 6." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 654, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 556, + 713 + ], + "type": "text", + "content": "NoR-MAE vs. Contrastive Learning. Empowering representation learning via invariance-covariance optimization has been investigated in Barlow Twins [36]. Recently proposed U-MAE [37] combines the contrastive learning and MAE [19] by introducing the random asymmetry masks" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "type": "text", + "content": "25233" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 95, + 70, + 519, + 250 + ], + "blocks": [ + { + "bbox": [ + 95, + 70, + 519, + 250 + ], + "lines": [ + { + "bbox": [ + 95, + 70, + 519, + 250 + ], + "spans": [ + { + "bbox": [ + 95, + 70, + 519, + 250 + ], + "type": "image", + "image_path": "d3373eb9e2747902089d8a19e3c05b23dc98b5f1630e933e554681ef24e57089.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 198, + 258, + 411, + 269 + ], + "lines": [ + { + "bbox": [ + 198, + 258, + 411, + 269 + ], + "spans": [ + { + "bbox": [ + 198, + 258, + 411, + 269 + ], + "type": "text", + "content": "Figure 5. Overview of the proposed NoR-MAE algorithm." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 69, + 287, + 282, + 385 + ], + "blocks": [ + { + "bbox": [ + 69, + 287, + 282, + 385 + ], + "lines": [ + { + "bbox": [ + 69, + 287, + 282, + 385 + ], + "spans": [ + { + "bbox": [ + 69, + 287, + 282, + 385 + ], + "type": "image", + "image_path": "b60a5977a31156d6ab39a487c36b30b811d9a96cd9f5c03d8ed2d552827182f7.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 61, + 394, + 289, + 405 + ], + "lines": [ + { + "bbox": [ + 61, + 394, + 289, + 405 + ], + "spans": [ + { + "bbox": [ + 61, + 394, + 289, + 405 + ], + "type": "text", + "content": "Figure 6. Different relation penalty of the NoR-Loss " + }, + { + "bbox": [ + 61, + 394, + 289, + 405 + ], + "type": "inline_equation", + "content": "(\\lambda = 5\\mathrm{e} - 5)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 426, + 296, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 426, + 296, + 498 + ], + "spans": [ + { + "bbox": [ + 55, + 426, + 296, + 498 + ], + "type": "text", + "content": "during the forward pass of the training. On the contrary, the proposed NoR-MAE exploits the semantic aliasing issue by constructing the asymmetry between the visible patches and the corresponding closest neighbors, instead of relying on the randomness. As a result, NoR-MAE outperforms U-MAE [37] in various sizes of models (Section 4)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 498, + 297, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 498, + 297, + 617 + ], + "spans": [ + { + "bbox": [ + 54, + 498, + 297, + 617 + ], + "type": "text", + "content": "In the context of contrastive learning, the similar samples (patches) are considered as \"positive samples\" which are intended to be aligned by the algorithms [35]. However, the important discovery of NoR-MAE points out that semantic aliasing among the closest neighbors is harmful for the lightweight masked autoencoder training, which is not properly addressed by prior works. Moreover, unlike Barlow Twins [36], NoR-MAE does not need the embedding alignment along the diagonal terms due to the well-established similarity between the visible patches and their CNPs." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 618, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 297, + 714 + ], + "type": "text", + "content": "NoR-MAE Does Not Need a Teacher Model. We would like to highlight that the proposed NoR-MAE is entirely different from knowledge distillation-based MAE. NoR-MAE does not require a pre-trained teacher [3] or multistage distillation [21] during training. The lightweight encoder is trained from scratch following the standard MAE protocol [19]. Although the CNPs and the original unmasked patches are separately generated, the encoding-decoding pro" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 290, + 555, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 290, + 555, + 350 + ], + "spans": [ + { + "bbox": [ + 313, + 290, + 555, + 350 + ], + "type": "text", + "content": "cess remains the same as the vanilla MAE. For example, given the " + }, + { + "bbox": [ + 313, + 290, + 555, + 350 + ], + "type": "inline_equation", + "content": "75\\%" + }, + { + "bbox": [ + 313, + 290, + 555, + 350 + ], + "type": "text", + "content": " mask ratio with " + }, + { + "bbox": [ + 313, + 290, + 555, + 350 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 313, + 290, + 555, + 350 + ], + "type": "text", + "content": " visible patches, the equivalent input sparsity of NoR-MAE is " + }, + { + "bbox": [ + 313, + 290, + 555, + 350 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 313, + 290, + 555, + 350 + ], + "type": "text", + "content": ". The overhead of the NoR-MAE is minimal compared to the vanilla MAE, as profiled in the supplementary." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 358, + 455, + 371 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 358, + 455, + 371 + ], + "spans": [ + { + "bbox": [ + 313, + 358, + 455, + 371 + ], + "type": "text", + "content": "3.3. Rationality of NoR-MAE" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 377, + 555, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 377, + 555, + 449 + ], + "spans": [ + { + "bbox": [ + 313, + 377, + 555, + 449 + ], + "type": "text", + "content": "As shown in Figure 2 and Figure 4, the proposed concept of Closest Neighbor Patch (CNP) characterizes the local semantics of the image. For instance, the patches that contain \"Background Forest\" or \"Back of the Fox\" are represented by separate CNP patches. However, the lightweight model failed to learn the semantic difference between CNPs." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 449, + 556, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 449, + 556, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 556, + 653 + ], + "type": "text", + "content": "We prove the significance of the CNP semantics by analyzing the original unmasked features and CNP features encoded by the trained encoder " + }, + { + "bbox": [ + 313, + 449, + 556, + 653 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 313, + 449, + 556, + 653 + ], + "type": "text", + "content": ". Specifically, we first normalize the encoded features and then compress the feature dimension down to 2-D via the Principal Component Analysis (PCA) [1]. As shown in Figure 7 and Figure 8, the lightweight models (ViT-Tiny and ViT-Small) trained by the vanilla MAE [19] failed to distinguish the semantical difference between different closest neighbor patches (CNP, blue dots), regardless of the patch sizes. Different semantics of CNP are heavily mixed together in the low dimensional space. Moreover, CNP is critical for MAE training regardless of the input patch sizes and the corresponding sequence length. The distorted and aliased understanding of the local semantics leads to the sub-optimal performance of MAE, as shown in Table 1 with the empirical verification on ImageNet-100 (" + }, + { + "bbox": [ + 313, + 449, + 556, + 653 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 313, + 449, + 556, + 653 + ], + "type": "text", + "content": ") with ViT-Tiny." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 654, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 556, + 713 + ], + "type": "text", + "content": "On the contrary, the model trained by NoR-MAE successfully understands the semantic difference between different CNP features, leading to largely improved accuracy on the lightweight ViT models across different patch sizes. In the meantime, the large-sized model (e.g., ViT-Base) exhibits" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 733, + 318, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 733, + 318, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 733, + 318, + 742 + ], + "type": "text", + "content": "25234" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 77, + 91, + 143, + 145 + ], + "blocks": [ + { + "bbox": [ + 78, + 72, + 228, + 89 + ], + "lines": [ + { + "bbox": [ + 78, + 72, + 228, + 89 + ], + "spans": [ + { + "bbox": [ + 78, + 72, + 228, + 89 + ], + "type": "text", + "content": "ViT-Tiny (# of params = 5.7 Million) MAE [19]: Aliased Semantics! NoR-MAE: Clearly Separated CNP!" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 77, + 91, + 143, + 145 + ], + "lines": [ + { + "bbox": [ + 77, + 91, + 143, + 145 + ], + "spans": [ + { + "bbox": [ + 77, + 91, + 143, + 145 + ], + "type": "image", + "image_path": "48ece0badf6f42df429dff5acb665a13e75d4b78025050dd218266046eece57a.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 82, + 144, + 133, + 152 + ], + "lines": [ + { + "bbox": [ + 82, + 144, + 133, + 152 + ], + "spans": [ + { + "bbox": [ + 82, + 144, + 133, + 152 + ], + "type": "text", + "content": "Accuracy " + }, + { + "bbox": [ + 82, + 144, + 133, + 152 + ], + "type": "inline_equation", + "content": "= 66.60" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 153, + 91, + 220, + 144 + ], + "blocks": [ + { + "bbox": [ + 153, + 91, + 220, + 144 + ], + "lines": [ + { + "bbox": [ + 153, + 91, + 220, + 144 + ], + "spans": [ + { + "bbox": [ + 153, + 91, + 220, + 144 + ], + "type": "image", + "image_path": "a3c7f37475c3840661a41f651fc6cdd3f5c2ce776f044424d7bd7d13fdd8c037.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 157, + 144, + 212, + 152 + ], + "lines": [ + { + "bbox": [ + 157, + 144, + 212, + 152 + ], + "spans": [ + { + "bbox": [ + 157, + 144, + 212, + 152 + ], + "type": "text", + "content": "Accuracy " + }, + { + "bbox": [ + 157, + 144, + 212, + 152 + ], + "type": "inline_equation", + "content": "= 70.24" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 233, + 72, + 386, + 152 + ], + "blocks": [ + { + "bbox": [ + 163, + 154, + 264, + 161 + ], + "lines": [ + { + "bbox": [ + 163, + 154, + 264, + 161 + ], + "spans": [ + { + "bbox": [ + 163, + 154, + 264, + 161 + ], + "type": "text", + "content": "Latent Feature of Unmasked Patches" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 233, + 72, + 386, + 152 + ], + "lines": [ + { + "bbox": [ + 233, + 72, + 386, + 152 + ], + "spans": [ + { + "bbox": [ + 233, + 72, + 386, + 152 + ], + "type": "image", + "image_path": "b4d68fa8ebda505a9e55496391aba91dc406ddcfd9f979af87b1fbba76ee70fa.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 277, + 154, + 394, + 162 + ], + "lines": [ + { + "bbox": [ + 277, + 154, + 394, + 162 + ], + "spans": [ + { + "bbox": [ + 277, + 154, + 394, + 162 + ], + "type": "text", + "content": "Latent Feature of Closest Neighbor Patches" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 394, + 72, + 544, + 152 + ], + "blocks": [ + { + "bbox": [ + 394, + 72, + 544, + 152 + ], + "lines": [ + { + "bbox": [ + 394, + 72, + 544, + 152 + ], + "spans": [ + { + "bbox": [ + 394, + 72, + 544, + 152 + ], + "type": "image", + "image_path": "c8d1d286339bd12eb261ad440299bfac49a011b79b73297a9177b3873587d88a.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 406, + 154, + 509, + 162 + ], + "lines": [ + { + "bbox": [ + 406, + 154, + 509, + 162 + ], + "spans": [ + { + "bbox": [ + 406, + 154, + 509, + 162 + ], + "type": "text", + "content": "Patch size " + }, + { + "bbox": [ + 406, + 154, + 509, + 162 + ], + "type": "inline_equation", + "content": "= 16^{2}" + }, + { + "bbox": [ + 406, + 154, + 509, + 162 + ], + "type": "text", + "content": " Token length " + }, + { + "bbox": [ + 406, + 154, + 509, + 162 + ], + "type": "inline_equation", + "content": "= 196" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 66, + 217, + 289, + 306 + ], + "blocks": [ + { + "bbox": [ + 55, + 174, + 555, + 196 + ], + "lines": [ + { + "bbox": [ + 55, + 174, + 555, + 196 + ], + "spans": [ + { + "bbox": [ + 55, + 174, + 555, + 196 + ], + "type": "text", + "content": "Figure 7. Low dimensional features of the unmasked patches (red dots) and the closest neighbor patches (CNP) (blue dots) with the same input from ImageNet-1K dataset. NoR-MAE clearly separates the semantic information among different CNP, with improved accuracy." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 66, + 217, + 289, + 306 + ], + "lines": [ + { + "bbox": [ + 66, + 217, + 289, + 306 + ], + "spans": [ + { + "bbox": [ + 66, + 217, + 289, + 306 + ], + "type": "image", + "image_path": "e261fe14e69b24b254ab14b55e69a4de170ae8965021bae505c05c3edeec4168.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 319, + 295, + 352 + ], + "lines": [ + { + "bbox": [ + 55, + 319, + 295, + 352 + ], + "spans": [ + { + "bbox": [ + 55, + 319, + 295, + 352 + ], + "type": "text", + "content": "Figure 8. Low dimensional features of the unmasked patches (red dots) and the CNP (blue dots) encoded by ViT-Tiny with patch size " + }, + { + "bbox": [ + 55, + 319, + 295, + 352 + ], + "type": "inline_equation", + "content": "= 4 \\times 4" + }, + { + "bbox": [ + 55, + 319, + 295, + 352 + ], + "type": "text", + "content": " (sequence length " + }, + { + "bbox": [ + 55, + 319, + 295, + 352 + ], + "type": "inline_equation", + "content": "= 3136" + }, + { + "bbox": [ + 55, + 319, + 295, + 352 + ], + "type": "text", + "content": "), evaluated on ImageNet-100." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 56, + 395, + 296, + 474 + ], + "blocks": [ + { + "bbox": [ + 55, + 364, + 295, + 387 + ], + "lines": [ + { + "bbox": [ + 55, + 364, + 295, + 387 + ], + "spans": [ + { + "bbox": [ + 55, + 364, + 295, + 387 + ], + "type": "text", + "content": "Table 1. NoR-MAE consistently outperforms the MAE [19] across different input patch sizes (evaluated on the ImageNet-100 dataset)." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 395, + 296, + 474 + ], + "lines": [ + { + "bbox": [ + 56, + 395, + 296, + 474 + ], + "spans": [ + { + "bbox": [ + 56, + 395, + 296, + 474 + ], + "type": "table", + "html": "
Masking RatioInput Patch SizeMAE [19]NoR-MAE (This work)
75%2²71.08%83.56% (+12.48%)
75%4²71.16%83.58% (+12.42%)
75%8²71.27%79.24% (+7.97%)
75%16²71.04%78.26% (+7.22%)
75%32²68.42%74.73% (+6.31%)
", + "image_path": "f0303f5faaf2dc8ac86402fb0441eaa2035140937299cd8d188a7fbeb96b7f2a.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 495, + 295, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 495, + 295, + 518 + ], + "spans": [ + { + "bbox": [ + 55, + 495, + 295, + 518 + ], + "type": "text", + "content": "a better understanding and stronger separation of different semantics with the vanilla MAE [19]." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 55, + 519, + 296, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 519, + 296, + 591 + ], + "spans": [ + { + "bbox": [ + 55, + 519, + 296, + 591 + ], + "type": "text", + "content": "In summary, the proposed NoR-MAE successfully mitigates the performance gap between model sizes, facilitating the learnability of the transformer model with a stronger understanding of the semantical relations from the input, leading to improved accuracy with both lightweight and large-sized transformer models." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 55, + 603, + 180, + 616 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 603, + 180, + 616 + ], + "spans": [ + { + "bbox": [ + 55, + 603, + 180, + 616 + ], + "type": "text", + "content": "4. Experimental Results" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 55, + 623, + 189, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 623, + 189, + 635 + ], + "spans": [ + { + "bbox": [ + 55, + 623, + 189, + 635 + ], + "type": "text", + "content": "4.1. ImageNet Classification" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "type": "text", + "content": "For image classification tasks, we follow the standard protocol of MAE [19] by training the lightweight encoder on the ImageNet-1K and ImageNet-100 datasets with the batch size of 4,096 and 2,048, and subsequently performing the standard fine-tuning [19]. Specifically, the model is pre-trained by 200 and 400 epochs with the initial learning rate set to" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 217, + 555, + 325 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 217, + 555, + 325 + ], + "spans": [ + { + "bbox": [ + 313, + 217, + 555, + 325 + ], + "type": "text", + "content": "1.5e-4, together with the weight decay value of 0.05. The penalty coefficient " + }, + { + "bbox": [ + 313, + 217, + 555, + 325 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 217, + 555, + 325 + ], + "type": "text", + "content": " of the proposed NoR Loss is set to 5e-5, and the optimality of the selected penalty level is demonstrated in Table 5. The mask ratio is fixed at " + }, + { + "bbox": [ + 313, + 217, + 555, + 325 + ], + "type": "inline_equation", + "content": "75\\%" + }, + { + "bbox": [ + 313, + 217, + 555, + 325 + ], + "type": "text", + "content": " based on the findings in the vanilla MAE [19]. We implement the standard fine-tuning protocol with 100 epochs as the vanilla MAE [19]. The detailed experimental setup (pre-train and downstream) are summarized in the supplementary, together with the report of GPU memory usage and training cost." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 326, + 556, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 326, + 556, + 434 + ], + "spans": [ + { + "bbox": [ + 313, + 326, + 556, + 434 + ], + "type": "text", + "content": "Table 2 and Table 3 summarize the performance of the NoR-MAE on ImageNet-100 and ImageNet-1K datasets. Compared to the vanilla MAE [19], NoR-MAE largely improves the performance on top of the MAE baseline [19] with up to " + }, + { + "bbox": [ + 313, + 326, + 556, + 434 + ], + "type": "inline_equation", + "content": "7.22\\%" + }, + { + "bbox": [ + 313, + 326, + 556, + 434 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 326, + 556, + 434 + ], + "type": "inline_equation", + "content": "3.64\\%" + }, + { + "bbox": [ + 313, + 326, + 556, + 434 + ], + "type": "text", + "content": " accuracy on ImageNet-100 and ImageNet-1K datasets, respectively. For the large-sized vision transformer training (e.g., ViT-Base), NoR-MAE still maintains the superiority compared to the recent high-performance MAE training (Table 3)." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 435, + 556, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 435, + 556, + 555 + ], + "spans": [ + { + "bbox": [ + 313, + 435, + 556, + 555 + ], + "type": "text", + "content": "More importantly, NoR-MAE improves the model performance marginal training cost overhead. NoR-MAE trains the model with the standard MAE protocol without introducing 1) a pre-trained teacher [3] or 2) other additional architectures [27] or clustering modules [12]. As profiled in the supplementary, the performance boost reward by the NoR-MAE only exhibits up to " + }, + { + "bbox": [ + 313, + 435, + 556, + 555 + ], + "type": "inline_equation", + "content": "8\\%" + }, + { + "bbox": [ + 313, + 435, + 556, + 555 + ], + "type": "text", + "content": " memory overhead and " + }, + { + "bbox": [ + 313, + 435, + 556, + 555 + ], + "type": "inline_equation", + "content": "2\\%" + }, + { + "bbox": [ + 313, + 435, + 556, + 555 + ], + "type": "text", + "content": " additional training time. Compared to the distillation-based MAE [3], NoR-MAE outperforms DMAE [3] with " + }, + { + "bbox": [ + 313, + 435, + 556, + 555 + ], + "type": "inline_equation", + "content": "\\sim 1\\%" + }, + { + "bbox": [ + 313, + 435, + 556, + 555 + ], + "type": "text", + "content": " accuracy improvement and " + }, + { + "bbox": [ + 313, + 435, + 556, + 555 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 313, + 435, + 556, + 555 + ], + "type": "text", + "content": " less training cost." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 555, + 556, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 555, + 556, + 616 + ], + "spans": [ + { + "bbox": [ + 313, + 555, + 556, + 616 + ], + "type": "text", + "content": "We further validate the NoR-MAE on the larger ViT-Base model. The proposed algorithm consistently achieves better performance compared to the MAE baseline and the recent ImageNet-100 SoTA method without introducing any additional clustering efforts or special tokenizer [12]." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 616, + 556, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 616, + 556, + 676 + ], + "spans": [ + { + "bbox": [ + 313, + 616, + 556, + 676 + ], + "type": "text", + "content": "Finally, NoR-MAE is also suitable for the MAE with supervision (SupMAE) [22]. Introducing supervised learning into MAE does not impact the effectiveness of the proposed NoR-MAE, while the NoR-MAE further enhances the performance of SupMAE [22], as shown in Table 2." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "content": "Comparison with the Distillation-based MAE. We would like to highlight that the proposed NoR-MAE can achieve on-par or even better performance compared to the" + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 733, + 317, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 733, + 317, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 733, + 317, + 742 + ], + "type": "text", + "content": "25235" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 80, + 100, + 529, + 511 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "text", + "content": "Table 2. Fine-tuning and semi-supervised learning accuracy (%) of the ViT encoder pre-trained by MAE [19] and recent SoTA lightweight MAE on ImageNet-1K dataset (with 200 epochs and 400 epochs pre-training)." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 80, + 100, + 529, + 511 + ], + "lines": [ + { + "bbox": [ + 80, + 100, + 529, + 511 + ], + "spans": [ + { + "bbox": [ + 80, + 100, + 529, + 511 + ], + "type": "table", + "html": "
ViT Model# of ParamsMethodTeacher ModelPre-training EpochSupervised Fine-tuning Accuracy (%)
ViT-Tiny5.7MMAE [19]-20066.60
MAE [19]-40067.63
Sup MAE [22]-20067.88
Sup MAE [22]-40068.91
MoCo-V3 [18]-20070.09
DINO [5]-30068.17
SimMIM [34]-20066.08
DMAE [3]ViT-B20070.00
Sup NoR-MAE (This work)-20070.63 (+2.75)
NoR-MAE (This work)-20070.24 (+3.64)
NoR-MAE (This work)-40071.18 (+3.55)
ViT-Small21.7MMAE [19]-20079.00
MAE [19]-40080.11
Sup MAE [22]-20079.27
CrossMAE [14]-40079.30
MoCo-V3 [18]-20079.46
DMAE [3]ViT-B20079.30
Sup NoR-MAE (This work)-20080.16 (+0.89)
NoR-MAE (This work)-20080.13 (+1.03)
NoR-MAE (This work)-40081.04 (+0.93)
ViT-Base88.6MMAE [19]-20083.30
Sup MAE [22]-20083.60
DMAE [3]-20084.00
CrossMAE [14]-20083.60
MoCo-V3 [3]-20083.14
DINO [5]-20082.80
SimMIM [34]-20083.50
U-MAE [37]-20083.00
NoR-MAE (This work)-20083.42 (+0.12)
Sup NoR-MAE (This work)-20083.70 (+0.10)
NoR-DMAE (This work)-20084.22 (+0.22)
", + "image_path": "fceec31f7dddf08af45711f5cbeb78b68de0c63db0b52e8af63952b09279c510.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 529, + 297, + 603 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 529, + 297, + 603 + ], + "spans": [ + { + "bbox": [ + 55, + 529, + 297, + 603 + ], + "type": "text", + "content": "distillation-based MAE, while significantly reducing the training resources, in terms of GPU training time and memory usage. As shown in Table 2, NoR-MAE directly trains the lightweight model from scratch, achieving on-par or even better performance compared to the recent distillation-based MAE training, which yet consumes massive GPU resources." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 610, + 288, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 610, + 288, + 624 + ], + "spans": [ + { + "bbox": [ + 55, + 610, + 288, + 624 + ], + "type": "text", + "content": "4.2. Transfer Learning on the Downstream Tasks" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 629, + 297, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 297, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 297, + 715 + ], + "type": "text", + "content": "We fine-tune the pre-trained ViT-Tiny and ViT-Small models (from Table 2) on the downstream vision tasks with 100 epochs. As shown in Table 4, the proposed NoR-MAE algorithm outperforms the MAE baseline [19] among all the downstream tasks with up to " + }, + { + "bbox": [ + 55, + 629, + 297, + 715 + ], + "type": "inline_equation", + "content": "5.13\\%" + }, + { + "bbox": [ + 55, + 629, + 297, + 715 + ], + "type": "text", + "content": " accuracy improvements. The outstanding downstream performance indicates the capability of learning strong representation through NoR-MAE." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 314, + 529, + 408, + 541 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 529, + 408, + 541 + ], + "spans": [ + { + "bbox": [ + 314, + 529, + 408, + 541 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 552, + 556, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 552, + 556, + 636 + ], + "spans": [ + { + "bbox": [ + 313, + 552, + 556, + 636 + ], + "type": "text", + "content": "Intensity of Neighbor-Repelling " + }, + { + "bbox": [ + 313, + 552, + 556, + 636 + ], + "type": "inline_equation", + "content": "(\\lambda)" + }, + { + "bbox": [ + 313, + 552, + 556, + 636 + ], + "type": "text", + "content": ". We evaluate the impact of the penalty level " + }, + { + "bbox": [ + 313, + 552, + 556, + 636 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 552, + 556, + 636 + ], + "type": "text", + "content": " with different values. As shown in Table 5, the proposed NoR-MAE algorithm achieves the best performance with " + }, + { + "bbox": [ + 313, + 552, + 556, + 636 + ], + "type": "inline_equation", + "content": "\\lambda = 5\\mathrm{e} - 5" + }, + { + "bbox": [ + 313, + 552, + 556, + 636 + ], + "type": "text", + "content": ". In other words, the repelling between the embedding features of the visible patches and the CNP should be properly controlled. Over-penalized semantic relation leads to sub-optimal performance." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "text", + "content": "Impact of Masking Ratio. In addition to the " + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "inline_equation", + "content": "75\\%" + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "text", + "content": " masking ratio that has been used in all prior works as the default setting, we evaluate the impact of different input masking ratios, varying from " + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "inline_equation", + "content": "85\\%" + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "text", + "content": " sparsity. Among different input masking ratios, the proposed NoR-MAE consistently outperforms the vanilla MAE [19] with lightweight vision" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 733, + 318, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 733, + 318, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 733, + 318, + 742 + ], + "type": "text", + "content": "25236" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 80, + 91, + 529, + 297 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 553, + 83 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 553, + 83 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 553, + 83 + ], + "type": "text", + "content": "Table 3. Fine-tuning accuracy (%) of the vision transformer encoder pre-trained by MAE and recent SoTA MAE methods on ImageNet-100." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 80, + 91, + 529, + 297 + ], + "lines": [ + { + "bbox": [ + 80, + 91, + 529, + 297 + ], + "spans": [ + { + "bbox": [ + 80, + 91, + 529, + 297 + ], + "type": "table", + "html": "
ViT Model# of ParametersPre-training MethodPre-training EpochsSupervised Fine-tuning Accuracy (%)
ViT-Tiny5.7MMAE [19]20071.04
DiNO-Tiny [5, 9]20063.04
CrossMAE [14]20070.82
NoR-MAE (This work)20078.26 (+7.22)
ViT-Small21.7MMAE [19]20081.82
MaskFeat HOG [32]20082.80
PeCo [10]20083.60
NoR-MAE (This work)20084.28 (+2.46)
ViT-Base88.6MMAE [19]20086.80
CrossMAE [14]20086.29
U-MAE [37]20086.80
BEiT [4]20086.10
NoR-MAE (This work)20087.56 (+0.76)
", + "image_path": "3360bdb22e4e5203b11368cbb0a97f615dba953c55afca0e2ada8a189ed4cdc8.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 67, + 327, + 541, + 394 + ], + "blocks": [ + { + "bbox": [ + 58, + 307, + 550, + 319 + ], + "lines": [ + { + "bbox": [ + 58, + 307, + 550, + 319 + ], + "spans": [ + { + "bbox": [ + 58, + 307, + 550, + 319 + ], + "type": "text", + "content": "Table 4. Fine-tuning accuracy (%) of the NoR-MAE on the downstream tasks with the lightweight ViT pre-trained on the ImageNet-1K." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 67, + 327, + 541, + 394 + ], + "lines": [ + { + "bbox": [ + 67, + 327, + 541, + 394 + ], + "spans": [ + { + "bbox": [ + 67, + 327, + 541, + 394 + ], + "type": "table", + "html": "
ViT ModelMethodCIFAR-10CIFAR-100FlowersPetsDTDFoodCaltech-101AircraftADE20K (mIoU)
ViT-TinyMAE [19]95.0178.7485.8082.4755.0775.6368.0564.6026.54
NoR-MAE (This work)95.9880.1388.7185.2960.1878.7173.1867.0733.54
ViT-SmallMAE [19]97.5084.8391.4991.1462.3181.5780.8555.9041.14
NoR-MAE (This work)98.1785.5791.8591.7064.7983.1582.0857.4142.99
", + "image_path": "2d1bac84e481a2f4088286a0dfa059f056dfea473085ccb0ff72e5d948fc72db.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 56, + 443, + 296, + 480 + ], + "blocks": [ + { + "bbox": [ + 55, + 411, + 297, + 435 + ], + "lines": [ + { + "bbox": [ + 55, + 411, + 297, + 435 + ], + "spans": [ + { + "bbox": [ + 55, + 411, + 297, + 435 + ], + "type": "text", + "content": "Table 5. Impact of different penalty levels " + }, + { + "bbox": [ + 55, + 411, + 297, + 435 + ], + "type": "inline_equation", + "content": "(\\lambda)" + }, + { + "bbox": [ + 55, + 411, + 297, + 435 + ], + "type": "text", + "content": " on Neighbor Repelling on the ImageNet-100 dataset." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 443, + 296, + 480 + ], + "lines": [ + { + "bbox": [ + 56, + 443, + 296, + 480 + ], + "spans": [ + { + "bbox": [ + 56, + 443, + 296, + 480 + ], + "type": "table", + "html": "
Penalty Level (λ)0.0 (Baseline)1e-55e-51e-45e-4
ViT-Tiny70.2476.2078.2677.8972.17
ViT-Small81.8282.1484.2884.2181.83
", + "image_path": "b5cf72c9eb31b49ba95bd8e21b3e65347d4a84491360aa39d04c260391834cea.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 56, + 529, + 296, + 582 + ], + "blocks": [ + { + "bbox": [ + 55, + 494, + 295, + 527 + ], + "lines": [ + { + "bbox": [ + 55, + 494, + 295, + 527 + ], + "spans": [ + { + "bbox": [ + 55, + 494, + 295, + 527 + ], + "type": "text", + "content": "Table 6. Impact of different input masking ratios. NoR-MAE shows consistent performance improvements on the ImageNet-1K dataset with ViT-Tiny model." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 529, + 296, + 582 + ], + "lines": [ + { + "bbox": [ + 56, + 529, + 296, + 582 + ], + "spans": [ + { + "bbox": [ + 56, + 529, + 296, + 582 + ], + "type": "table", + "html": "
Patch-wise Masking RatioBaseline MAE [19]NoR-MAE (This work)
50%65.43%69.27%
75%66.60%70.24%
85%66.37%70.16%
", + "image_path": "21e231951e5b1b7d9bd179e3f822d7cd39a7a599106be2d60c4e0ff88a2f41a7.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 605, + 295, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 640 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 640 + ], + "type": "text", + "content": "transformers, as shown in Table 6. In other words, semantic aliasing is persistent in lightweight transformer models regardless of the information density in the input sequence." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": "NoR-MAE with extended training effort We further report the performance of the NoR-MAE algorithm with 1,000 and 1,600 epoch training on the ImageNet-100. As shown in Table 7, the extended training effort can further boost up the performance of the NoR-MAE-trained ViT-Tiny (5.2M) model to " + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "inline_equation", + "content": "83.96\\%" + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": ", while consistently outperforming MAE." + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 315, + 423, + 553, + 472 + ], + "blocks": [ + { + "bbox": [ + 314, + 411, + 553, + 422 + ], + "lines": [ + { + "bbox": [ + 314, + 411, + 553, + 422 + ], + "spans": [ + { + "bbox": [ + 314, + 411, + 553, + 422 + ], + "type": "text", + "content": "Table 7. Performance of NoR-MAE with extended training effort." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 423, + 553, + 472 + ], + "lines": [ + { + "bbox": [ + 315, + 423, + 553, + 472 + ], + "spans": [ + { + "bbox": [ + 315, + 423, + 553, + 472 + ], + "type": "table", + "html": "
Training EpochsMAE on ImageNet-100 (%)NoR-MAE on ImageNet-100 (%)
20071.0478.26
100073.2282.97
160074.2983.96
", + "image_path": "340e97d5a5a6aeab9e3242f5bdc38a47693307c74c21bef70d58ac9f3a8a0f5a.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 514, + 388, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 514, + 388, + 528 + ], + "spans": [ + { + "bbox": [ + 314, + 514, + 388, + 528 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 312, + 545, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 545, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 312, + 545, + 556, + 713 + ], + "type": "text", + "content": "In this paper, we propose NoR-MAE, a novel self-supervised learning algorithm designed for lightweight vision transformers via masked autoencoder training. We first introduce the concept of closest neighbor patches (CNP) into the MAE training, which is a critical concept in lightweight SSL. On top of that, we propose Neighbor-Repelling Loss for MAE training. The proposed method trains the lightweight model from scratch without using a pre-trained teacher model. NoR-MAE achieves the new state-of-the-art performance on direct lightweight MAE training with largely improved performance on both pre-training and downstream vision tasks. Finally, our discovery of the biased semantic relations provides valuable insights regarding lightweight masked autoencoder learning." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 733, + 318, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 733, + 318, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 733, + 318, + 742 + ], + "type": "text", + "content": "25237" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 161, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 161, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 161, + 85 + ], + "type": "text", + "content": "6. Acknowledgment" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 297, + 128 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 297, + 128 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 297, + 128 + ], + "type": "text", + "content": "This work is supported in part by the National Science Foundation under grant 2403723, and CoCoSys Center in JUMP 2.0, an SRC program sponsored by DARPA." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 140, + 115, + 153 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 140, + 115, + 153 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 115, + 153 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 160, + 297, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 160, + 297, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 160, + 297, + 193 + ], + "spans": [ + { + "bbox": [ + 61, + 160, + 297, + 193 + ], + "type": "text", + "content": "[1] Herve Abdi and Lynne J Williams. Principal component analysis. Wiley interdisciplinary reviews: computational statistics, 2(4):433-459, 2010. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 194, + 296, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 194, + 296, + 250 + ], + "spans": [ + { + "bbox": [ + 61, + 194, + 296, + 250 + ], + "type": "text", + "content": "[2] Alexei Baevski, Wei-Ning Hsu, Qiantong Xu, Arun Babu, Jiatao Gu, and Michael Auli. Data2vec: A general framework for self-supervised learning in speech, vision and language. In International Conference on Machine Learning (ICML), 2022. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 251, + 296, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 251, + 296, + 306 + ], + "spans": [ + { + "bbox": [ + 61, + 251, + 296, + 306 + ], + "type": "text", + "content": "[3] Yutong Bai, Zeyu Wang, Junfei Xiao, Chen Wei, Huiyu Wang, Alan L Yuille, Yuyin Zhou, and Cihang Xie. Masked autoencoders enable efficient knowledge distillers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1, 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 308, + 296, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 308, + 296, + 341 + ], + "spans": [ + { + "bbox": [ + 62, + 308, + 296, + 341 + ], + "type": "text", + "content": "[4] Hangbo Bao, Li Dong, Songhao Piao, and Furu Wei. BEit: BERT pre-training of image transformers. In International Conference on Learning Representations (ICLR), 2022. 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 342, + 296, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 342, + 296, + 396 + ], + "spans": [ + { + "bbox": [ + 62, + 342, + 296, + 396 + ], + "type": "text", + "content": "[5] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the International Conference on Computer Vision (ICCV), 2021. 7, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 399, + 296, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 399, + 296, + 442 + ], + "spans": [ + { + "bbox": [ + 62, + 399, + 296, + 442 + ], + "type": "text", + "content": "[6] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International Conference on Machine Learning (ICML), 2020. 1, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 445, + 296, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 445, + 296, + 476 + ], + "spans": [ + { + "bbox": [ + 62, + 445, + 296, + 476 + ], + "type": "text", + "content": "[7] Xinlei Chen, Haoqi Fan, Ross Girshick, and Kaiming He. Improved Baselines with Momentum Contrastive learning. arXiv preprint arXiv:2003.04297, 2020. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 479, + 296, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 479, + 296, + 544 + ], + "spans": [ + { + "bbox": [ + 62, + 479, + 296, + 544 + ], + "type": "text", + "content": "[8] Yu-An Chung, Yu Zhang, Wei Han, Chung-Cheng Chiu, James Qin, Ruoming Pang, and Yonghui Wu. W2v-bert: Combining contrastive learning and masked language modeling for self-supervised speech pre-training. In 2021 IEEE Automatic Speech Recognition and Understanding Workshop (ASRU), 2021. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 546, + 296, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 546, + 296, + 590 + ], + "spans": [ + { + "bbox": [ + 62, + 546, + 296, + 590 + ], + "type": "text", + "content": "[9] Victor Guilherme Turrisi da Costa, Enrico Fini, Moin Nabi, Nicu Sebe, and Elisa Ricci. solo-learn: A library of self-supervised methods for visual representation learning. Journal of Machine Learning Research, 2022. 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 591, + 296, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 591, + 296, + 646 + ], + "spans": [ + { + "bbox": [ + 57, + 591, + 296, + 646 + ], + "type": "text", + "content": "[10] Xiaoyi Dong, Jianmin Bao, Ting Zhang, Dongdong Chen, Weiming Zhang, Lu Yuan, Dong Chen, Fang Wen, Nenghai Yu, and Baining Guo. Peco: Perceptual codebook for bert pre-training of vision transformers. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), 2023. 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 647, + 296, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 296, + 712 + ], + "type": "text", + "content": "[11] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 1, 2" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 316, + 73, + 555, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 555, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 555, + 116 + ], + "type": "text", + "content": "[12] Tianqi Du, Yifei Wang, and Yisen Wang. On the role of discrete tokenization in visual representation learning. In International Conference on Learning Representations (ICLR), 2023. 3, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 119, + 555, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 119, + 555, + 162 + ], + "spans": [ + { + "bbox": [ + 317, + 119, + 555, + 162 + ], + "type": "text", + "content": "[13] Zhiyuan Fang, Jianfeng Wang, Lijuan Wang, Lei Zhang, Yezhou Yang, and Zicheng Liu. Seed: Self-supervised distillation for visual representation. In International Conference on Learning Representations (ICLR), 2021. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 164, + 555, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 164, + 555, + 207 + ], + "spans": [ + { + "bbox": [ + 317, + 164, + 555, + 207 + ], + "type": "text", + "content": "[14] Letian Fu, Long Lian, Renhao Wang, Baifeng Shi, Xudong Wang, Adam Yala, Trevor Darrell, Alexei A Efros, and Ken Goldberg. Rethinking patch dependence for masked autoencoders. arXiv preprint arXiv:2401.14391, 2024. 3, 7, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 209, + 555, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 209, + 555, + 262 + ], + "spans": [ + { + "bbox": [ + 316, + 209, + 555, + 262 + ], + "type": "text", + "content": "[15] Peng Gao, Shijie Geng, Renrui Zhang, Teli Ma, Rongyao Fang, Yongfeng Zhang, Hongsheng Li, and Yu Qiao. Clip-adapter: Better vision-language models with feature adapters. International Journal of Computer Vision, 132(2):581-595, 2024. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 265, + 555, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 265, + 555, + 330 + ], + "spans": [ + { + "bbox": [ + 316, + 265, + 555, + 330 + ], + "type": "text", + "content": "[16] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. Advances in Neural Information Processing Systems (NeurIPS), 2020. 1, 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 332, + 555, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 332, + 555, + 365 + ], + "spans": [ + { + "bbox": [ + 316, + 332, + 555, + 365 + ], + "type": "text", + "content": "[17] Agrim Gupta, Jiajun Wu, Jia Deng, and Li Fei-Fei. Siamese masked autoencoders. In Thirty-seventh Conference on Neural Information Processing Systems (NeurIPS), 2023. 1, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 366, + 555, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 366, + 555, + 419 + ], + "spans": [ + { + "bbox": [ + 317, + 366, + 555, + 419 + ], + "type": "text", + "content": "[18] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 1, 2, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 422, + 555, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 422, + 555, + 475 + ], + "spans": [ + { + "bbox": [ + 317, + 422, + 555, + 475 + ], + "type": "text", + "content": "[19] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dólár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 2, 3, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 478, + 555, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 478, + 555, + 510 + ], + "spans": [ + { + "bbox": [ + 316, + 478, + 555, + 510 + ], + "type": "text", + "content": "[20] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 2, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 513, + 555, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 513, + 555, + 567 + ], + "spans": [ + { + "bbox": [ + 317, + 513, + 555, + 567 + ], + "type": "text", + "content": "[21] Wei Huang, Zhiliang Peng, Li Dong, Furu Wei, Jianbin Jiao, and Qixiang Ye. Generic-to-specific distillation of masked autoencoders. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2, 3, 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 569, + 555, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 569, + 555, + 601 + ], + "spans": [ + { + "bbox": [ + 316, + 569, + 555, + 601 + ], + "type": "text", + "content": "[22] Feng Liang, Yangguang Li, and Diana Marculescu. Supmae: Supervised masked autoencoders are efficient vision learners. arXiv preprint arXiv:2205.14540, 2022. 6, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 602, + 555, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 602, + 555, + 656 + ], + "spans": [ + { + "bbox": [ + 316, + 602, + 555, + 656 + ], + "type": "text", + "content": "[23] Jihao Liu, Xin Huang, Jinliang Zheng, Yu Liu, and Hongsheng Li. Mixmae: Mixed and masked autoencoder for efficient pretraining of hierarchical vision transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 658, + 555, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 555, + 712 + ], + "type": "text", + "content": "[24] Jian Meng, Li Yang, Kyungmin Lee, Jinwoo Shin, Deliang Fan, and Jae sun Seo. Slimmed asymmetrical contrastive learning and cross distillation for lightweight model training. In Thirty-seventh Conference on Neural Information Processing Systems (NeurIPS), 2023. 2" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 734, + 318, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 734, + 318, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 734, + 318, + 742 + ], + "type": "text", + "content": "25238" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 56, + 73, + 296, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 296, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 296, + 106 + ], + "type": "text", + "content": "[25] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation Learning with Contrastive Predictive Coding. arXiv preprint arXiv:1807.03748, 2018. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 106, + 296, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 106, + 296, + 171 + ], + "spans": [ + { + "bbox": [ + 56, + 106, + 296, + 171 + ], + "type": "text", + "content": "[26] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning (ICML), 2021. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 172, + 296, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 172, + 296, + 228 + ], + "spans": [ + { + "bbox": [ + 56, + 172, + 296, + 228 + ], + "type": "text", + "content": "[27] Bowen Shi, Xiaopeng Zhang, Yaoming Wang, Jin Li, Wenrui Dai, Junni Zou, Hongkai Xiong, and Qi Tian. Hybrid distillation: Connecting masked autoencoders with contrastive learners. In The Twelfth International Conference on Learning Representations (ICLR), 2024. 3, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 228, + 296, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 228, + 296, + 293 + ], + "spans": [ + { + "bbox": [ + 56, + 228, + 296, + 293 + ], + "type": "text", + "content": "[28] Yucheng Tang, Dong Yang, Wenqi Li, Holger R Roth, Bennett Landman, Daguang Xu, Vishwesh Nath, and Ali Hatamizadeh. Self-supervised pre-training of swim transformers for 3d medical image analysis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 293, + 296, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 293, + 296, + 348 + ], + "spans": [ + { + "bbox": [ + 56, + 293, + 296, + 348 + ], + "type": "text", + "content": "[29] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In International Conference on Machine Learning (ICML), 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 349, + 296, + 393 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 349, + 296, + 393 + ], + "spans": [ + { + "bbox": [ + 56, + 349, + 296, + 393 + ], + "type": "text", + "content": "[30] Jue Wang, Gedas Bertasius, Du Tran, and Lorenzo Torresani. Long-short temporal contrastive learning of video transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (NeurIPS), 2022. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 393, + 296, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 393, + 296, + 437 + ], + "spans": [ + { + "bbox": [ + 56, + 393, + 296, + 437 + ], + "type": "text", + "content": "[31] Shaoru Wang, Jin Gao, Zeming Li, Xiaoqin Zhang, and Weiming Hu. A closer look at self-supervised lightweight vision transformers. In International Conference on Machine Learning (ICML). PMLR, 2023. 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 437, + 296, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 437, + 296, + 491 + ], + "spans": [ + { + "bbox": [ + 56, + 437, + 296, + 491 + ], + "type": "text", + "content": "[32] Chen Wei, Haoqi Fan, Saining Xie, Chao-Yuan Wu, Alan Yuille, and Christoph Feichtenhofer. Masked feature prediction for self-supervised visual pre-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 492, + 296, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 492, + 296, + 558 + ], + "spans": [ + { + "bbox": [ + 56, + 492, + 296, + 558 + ], + "type": "text", + "content": "[33] Kan Wu, Houwen Peng, Zhenghong Zhou, Bin Xiao, Mengchen Liu, Lu Yuan, Hong Xuan, Michael Valenzuela, Xi Stephen Chen, Xinggang Wang, et al. Tinyclip: Clip distillation via affinity mimicking and weight inheritance. In Proceedings of the IEEE/CVF International Conference on Computer Vision (CVPR), 2023. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 559, + 296, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 559, + 296, + 613 + ], + "spans": [ + { + "bbox": [ + 56, + 559, + 296, + 613 + ], + "type": "text", + "content": "[34] Zhenda Xie, Zheng Zhang, Yue Cao, Yutong Lin, Jianmin Bao, Zhuliang Yao, Qi Dai, and Han Hu. Simmim: A simple framework for masked image modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 614, + 296, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 614, + 296, + 667 + ], + "spans": [ + { + "bbox": [ + 56, + 614, + 296, + 667 + ], + "type": "text", + "content": "[35] Sukmin Yun, Hankook Lee, Jaehung Kim, and Jinwoo Shin. Patch-level representation learning for self-supervised vision transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 669, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 669, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 669, + 296, + 713 + ], + "type": "text", + "content": "[36] Jure Zbontar, Li Jing, Ishan Misra, Yann LeCun, and Stéphane Deny. Barlow Twins: Self-supervised Learning via Redundancy Reduction. In International Conference on Machine Learning (ICML), pages 12310-12320, 2021. 1, 2, 4, 5" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 161 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 316, + 73, + 555, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 555, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 555, + 116 + ], + "type": "text", + "content": "[37] Qi Zhang, Yifei Wang, and Yisen Wang. How mask matters: Towards theoretical understandings of masked autoencoders. Advances in Neural Information Processing Systems (NeurIPS), 2022. 1, 3, 4, 5, 7, 8" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 118, + 554, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 554, + 161 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 554, + 161 + ], + "type": "text", + "content": "[38] Zhiyu Zhao, Bingkun Huang, Sen Xing, Gangshan Wu, Yu Qiao, and Limin Wang. Asymmetric masked distillation for pre-training small foundation models. arXiv preprint arXiv:2311.03149, 2023. 2" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 734, + 317, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 734, + 317, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 734, + 317, + 742 + ], + "type": "text", + "content": "25239" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Co-op_ Correspondence-based Novel Object Pose Estimation/5a137c7c-5457-4cd6-bb8c-541c7e90dee7_content_list.json b/2025/Co-op_ Correspondence-based Novel Object Pose Estimation/5a137c7c-5457-4cd6-bb8c-541c7e90dee7_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..f56c8f096aa91566dc990dec4bfc434cc61c4cd5 --- /dev/null +++ b/2025/Co-op_ Correspondence-based Novel Object Pose Estimation/5a137c7c-5457-4cd6-bb8c-541c7e90dee7_content_list.json @@ -0,0 +1,1631 @@ +[ + { + "type": "text", + "text": "Co-op: Correspondence-based Novel Object Pose Estimation", + "text_level": 1, + "bbox": [ + 191, + 130, + 805, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Sunghill Moon Hyeontae Son Dongcheol Hur Sangwook Kim Naver LABS", + "bbox": [ + 189, + 180, + 807, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{sungphill.moon, son.ht, dongcheol.hur, o.s.w.a.l.k}@naverlabs.com", + "bbox": [ + 204, + 218, + 787, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 268, + 326, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We propose Co-op, a novel method for accurately and robustly estimating the 6DoF pose of objects unseen during training from a single RGB image. Our method requires only the CAD model of the target object and can precisely estimate its pose without any additional fine-tuning. While existing model-based methods suffer from inefficiency due to using a large number of templates, our method enables fast and accurate estimation with a small number of templates. This improvement is achieved by finding semidense correspondences between the input image and the pre-rendered templates. Our method achieves strong generalization performance by leveraging a hybrid representation that combines patch-level classification and offset regression. Additionally, our pose refinement model estimates probabilistic flow between the input image and the rendered image, refining the initial estimate to an accurate pose using a differentiable PnP layer. We demonstrate that our method not only estimates object poses rapidly but also outperforms existing methods by a large margin on the seven core datasets of the BOP Challenge, achieving state-of-the-art accuracy.", + "bbox": [ + 89, + 301, + 483, + 619 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 648, + 220, + 662 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Over the past decade, significant improvements in the accuracy of 6D object pose estimation have been achieved through advancements in deep learning methods [33, 35, 46, 47, 54, 58, 63]. However, despite their remarkable accuracy, these methods are challenging to apply in diverse, real-world environments because handling new objects requires acquiring large amounts of data and retraining the models for several hours or even days.", + "bbox": [ + 89, + 672, + 482, + 794 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To address these challenges, recent studies have explored various methods. One approach is the category-level methods [33, 59], which estimate the pose of unseen objects within known categories but struggle with entirely new ones. Another approach aims for generalization by training on large-scale pose estimation datasets and is divided into two types: model-free and model-based methods. However,", + "bbox": [ + 89, + 795, + 482, + 900 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9649e065d52e5f42b7de043e7d4c0d9765a89bc3bb28df18c1d665105552c145.jpg", + "image_caption": [ + "Query", + "Figure 1. Examples of 6D pose estimation of novel objects. Our method estimates semi-dense or dense correspondences between the input image and rendered images and uses them to estimate the pose." + ], + "image_footnote": [], + "bbox": [ + 516, + 270, + 571, + 446 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8fe44d0dfc52bd9cd8631a13969fe849a884dc2c73df306e28a366f53dbb2907.jpg", + "image_caption": [ + "Semi-dense Correspondences Coarse Pose" + ], + "image_footnote": [], + "bbox": [ + 575, + 271, + 736, + 446 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a98177e5321631133ffe9f14ac0f0e3bafaf8feb182d880b06fe245be6654863.jpg", + "image_caption": [ + "Initialization Dense Flow Refined Pose" + ], + "image_footnote": [], + "bbox": [ + 743, + 271, + 901, + 446 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "model-free methods [18, 29, 36, 53] often face challenges with low-texture object and report results only on limited datasets or those without occlusions.", + "bbox": [ + 511, + 551, + 905, + 597 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Given these limitations, model-based unseen object pose estimation methods [27, 38, 41, 44, 62] that utilize 3D CAD models have been gaining attention. These model-based methods typically consist of three processes—object detection, coarse pose estimation, and pose refinement—or sometimes include a fourth process, pose selection. Since previous works like CNOS [40] and SAM-6D [32] have already addressed object detection with excellent generalization performance, we concentrate our research on the subsequent pose estimation process.", + "bbox": [ + 511, + 599, + 905, + 750 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we propose Co-op (Correspondence-based Novel Object Pose Estimation), a novel method for accurately and robustly estimating the 6DoF pose of objects unseen during training from a single RGB image. Our research begins with the consideration of how to effectively define and find correspondences between two images to achieve object pose estimation with excellent generalization performance.", + "bbox": [ + 511, + 750, + 905, + 869 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "By setting object pose estimation as finding correspondences between two images, our model is trained at the", + "bbox": [ + 511, + 869, + 905, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "11622", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "patch or pixel level rather than the image level. This enables the model to learn low-level information, making it robust to domain shifts. Furthermore, by directly leveraging the geometric and structural information of the 3D model, it enables more accurate pose estimation. In our research, to achieve superior generalization performance, we estimate poses based on correspondences in both stages of pose estimation: coarse estimation and pose refinement.", + "bbox": [ + 89, + 90, + 480, + 210 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In the first stage, coarse estimation, we estimate the initial pose of the object by estimating semi-dense correspondences between multiple templates and the input image. In this stage, we use a hybrid representation that combines patch-level classification and offset regression to estimate the semi-dense correspondences. Then, we estimate the pose using the Perspective-n-Point (PnP) algorithm [28] based on these correspondences. By formulating correspondence estimation not as a direct regression problem but by combining it with classification, we achieve outstanding generalization performance. Thanks to this representation, we achieve high accuracy while using a much smaller number of templates compared to other existing coarse estimation methods.", + "bbox": [ + 89, + 212, + 480, + 421 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In the second stage, pose refinement, we refine the initial pose using a render-and-compare approach. To achieve precise refinement, we estimate probabilistic dense correspondences and learn the confidence in pose estimation by training end-to-end through a differentiable PnP layer. This allows us to robustly estimate poses even in challenging situations such as occlusions and texture-less objects.", + "bbox": [ + 89, + 422, + 480, + 527 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions can be summarized as follows:", + "bbox": [ + 109, + 529, + 434, + 542 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We present Co-op, a novel framework for unseen object pose estimation in RGB-only cases. Co-op does not require additional training or fine-tuning for new objects and outperforms existing methods by a large margin on the seven core datasets of the BOP Challenge.", + "- We propose a method for fast and accurate coarse pose estimation using a hybrid representation that combines patch-level classification and offset regression.", + "- Additionally, we propose a precise object pose refinement method. It estimates dense correspondences defined by probabilistic flow and learns confidence end-to-end through a differentiable PnP layer." + ], + "bbox": [ + 91, + 544, + 480, + 724 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 739, + 232, + 753 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Correspondence estimation Traditionally, finding correspondences between two images has been handled using detector-based methods involving three steps: keypoint detection, feature description, and feature matching. These methods have been extensively researched, transitioning from handcrafted approaches [2, 4, 37, 50] to deep learning-based approaches [9, 13, 49, 57]. Recently, detector-free methods [6, 52, 60] have integrated feature detection and matching into a unified process, processing entire images", + "bbox": [ + 89, + 763, + 480, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "without predefined keypoints and improving matching performance even in low-texture or repetitive areas. Furthermore, dense matching methods [14, 15, 55, 56] find matches for every pixel, enhancing accuracy in tasks such as image-based localization, 3D reconstruction, and pose estimation. Our method follows the detector-free approach in coarse estimation and pose refinement. In the coarse estimation stage, we estimate semi-dense correspondences to achieve textureless-resilient and accurate matching, followed by the pose refinement stage, where the model estimates dense correspondences and their confidence to obtain an exact 6D pose.", + "bbox": [ + 511, + 90, + 903, + 271 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Seen Object Pose Estimation Deep learning-based methods have shown remarkable performance improvements with the growth of high-quality datasets and advances in GPU computing power. Some of these methods regress sparse [25, 46, 47, 54] or dense [17, 31, 45] features of the object, such as keypoints or correspondences and separately recover the object's pose by applying RANSAC [16] and PnP [28] on the regressed features. To achieve higher accuracy, other methods use render-and-compare based refinement models [23, 26, 30, 34]. Because these methods are designed to encode specific object's features into the network weights, they tend to perform poorly when encountering unseen objects.", + "bbox": [ + 511, + 275, + 903, + 470 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Unseen Object Pose Estimation To tackle problems in more practical situations, some studies assume that objects are unknown in advance. In particular, there has been active research on model-based unseen object pose estimation.", + "bbox": [ + 511, + 474, + 903, + 535 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Some methods aim to solve the problem using template matching approaches. These methods [27, 38, 39, 42, 51, 62] render the 3D model to generate templates and estimate poses by selecting the most similar template to the query image. Other methods use feature matching approaches, These methods [1, 32, 39, 41, 44] find 2D-3D or 3D-3D correspondences between the 3D model and the query image, recovering poses using the PnP [28] algorithm. To enhance accuracy, several methods [27, 38, 62] apply render-and compare pose refinement, iteratively minimizing discrepancies between the rendered model and the query image. Combined approaches like GenFlow [38] and MegaPose [27] integrate template matching with refinement, achieving remarkable performance but incurring significant computational costs due to the need for numerous template comparisons.", + "bbox": [ + 511, + 537, + 903, + 777 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address this issue, FoundPose [44] and GigaPose [41] shift from exhaustive template comparison to finding correspondences between the template and the query image, enabling faster and more robust initial pose estimation. These methods depend on the segmentation masks from CNOS [40], to separate the object regions, and perform feature matching between the query image and the template based on DINOv2 [43]. These methods can be", + "bbox": [ + 511, + 779, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "11623", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/15495b07cbb318c090bb642b70f594120fb290197fa489f70fac85ac2cfe50c1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 96, + 92, + 903, + 292 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/c8a1f008e442db6c8672505aa31dd42c622b51d53b0663b83de0d9b24ef78c5d.jpg", + "image_caption": [ + "Figure 2. Overview. We estimate object pose through two main stages. In the Coarse Pose Estimation stage (Sec 3.1), we estimate semi-dense correspondences between the query image and templates and compute the initial pose using PnP. In the Pose Refinement stage (Sec 3.2), we refine the initial pose by estimating dense flow between the query and rendered images. Both stages utilize transformer encoders and decoders with identical structures, with the Pose Refinement stage additionally incorporating a DPT module after the decoder for dense prediction." + ], + "image_footnote": [], + "bbox": [ + 94, + 301, + 903, + 508 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "considered a detect-and-describe framework using the segmentation mask as a feature detector and the DINOv2 patch descriptor as a feature descriptor.", + "bbox": [ + 89, + 614, + 482, + 660 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "On the other hand, we propose a detector-free correspondence estimation framework for the coarse estimation stage. By not relying on segmentation, our method is robust to the noisy or partially detected regions and improves the pose accuracy.", + "bbox": [ + 89, + 667, + 483, + 744 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 777, + 181, + 792 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we describe Co-op, our proposed method for unseen object pose estimation. It consists of two main components: coarse pose estimation and pose refinement. Additionally, it optionally incorporates pose selection to estimate the pose more accurately. An overview of our method is presented in Fig. 2.", + "bbox": [ + 89, + 810, + 483, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Coarse Pose Estimation", + "text_level": 1, + "bbox": [ + 513, + 614, + 732, + 628 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Template Generation In the coarse estimation stage, we use a set of pre-rendered templates from the 3D model. Similar to GigaPose [41], we minimize the number of templates required for pose estimation by generating those that include only out-of-plane rotations.", + "bbox": [ + 511, + 638, + 906, + 714 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Specifically, we render the CAD model from 42 viewpoints. These viewpoints are defined by subdividing each triangle of Blender's icosphere primitive into four smaller triangles, resulting in a uniform distribution of viewpoints over the sphere, following the method used in previous works [39, 40].", + "bbox": [ + 511, + 715, + 905, + 806 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "**Hybrid Representation** The role of the coarse estimation model is to estimate semi-dense correspondences between the query image and the pre-rendered template, denoted by $\\mathcal{I}_Q, \\mathcal{I}_T \\in \\mathbb{R}^{H \\times W \\times 3}$ . To enhance the generalization performance of our model, we use a hybrid representation that combines patch-level classification and offset regression.", + "bbox": [ + 511, + 810, + 906, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "11624", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/43e1d6b54781f5ef8ae7a42d8a3485f8b1b7b12cbbf0699caf5b9a28d63f4fae.jpg", + "image_caption": [ + "Figure 3. Visualization of Our Hybrid Representation. Left: Patch-level classification results; matching patches are highlighted with the same color. Right: Offset regression within template patches to refine correspondences; red arrows represent the estimated offsets." + ], + "image_footnote": [], + "bbox": [ + 101, + 95, + 472, + 262 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Fig. 3 illustrates an example of our hybrid representation. On the left, we show partial results of the patch-level classification; patches highlighted with the same color in the query image (left) and the template (right) indicate matched patches. On the right, we show the results of offset regression within the template patches after classification, demonstrating the estimated offset from the center of the template patch that best aligns with the center of the corresponding query patch.", + "bbox": [ + 89, + 378, + 483, + 515 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As shown in Fig. 2, our model consists of a Vision Transformer (ViT) encoder for extracting image features and a transformer decoder for interpreting these features. Here, the subscript $i,j$ represents the 2D location in the feature map. Because downsizing is performed by the encoder, each location $i,j$ in the feature map corresponds to a $16\\times 16$ patch in the input image.", + "bbox": [ + 89, + 516, + 483, + 622 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The encoder takes $\\mathcal{I}_Q$ and $\\mathcal{I}_T$ as inputs and extracts the feature maps $\\mathcal{F}_Q, \\mathcal{F}_T \\in \\mathbb{R}^{\\frac{H}{16} \\times \\frac{W}{16} \\times 1024}$ , respectively. The decoder and following heads then processes $\\mathcal{F}_Q$ and $\\mathcal{F}_T$ to compute both a tensor for patch-level classification $\\mathcal{C} \\in \\mathbb{R}^{\\frac{H}{16} \\times \\frac{W}{16} \\times K}$ and the xy-offsets $\\mathcal{U} \\in \\mathbb{R}^{\\frac{H}{16} \\times \\frac{W}{16} \\times 2}$ .", + "bbox": [ + 89, + 623, + 483, + 700 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here, $K = \\frac{H}{16} \\times \\frac{W}{16} + 1$ represents the number of classes for the patch-level classification. The first $\\frac{H}{16} \\times \\frac{W}{16}$ classes correspond to the template patches, while the last class represents the case where there is no matching part, such as occlusion. For each position $(i,j)$ in the feature map, $\\mathcal{C}_{i,j} \\in \\mathbb{R}^K$ contains classification scores indicating the degree of matchability between the $\\frac{H}{16} \\times \\frac{W}{16}$ template patches and the query patch at that position.", + "bbox": [ + 89, + 700, + 482, + 824 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Following this, we define the patch index as $c_{i,j} = \\arg \\max_{k\\in 1,2,\\dots,K}\\mathcal{C}_{i,j}^{k}$ and the offset $\\mathcal{U}_{i,j}$ has a range of $[-0.5,0.5]$ . At location $(i,j)$ where $c_{i,j} \\neq K$ , the position $M_{i,j}^{T}$ in the template corresponding to the query patch is given by:", + "bbox": [ + 89, + 825, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {M} _ {i, j} ^ {T} = (\\left[ \\begin{array}{c} c _ {i, j} \\bmod 1 6 + 0. 5 \\\\ \\lfloor c _ {i, j} / 1 6 \\rfloor + 0. 5 \\end{array} \\right] + \\mathcal {U} _ {i, j}) \\times 1 6. \\qquad (1)\n$$\n", + "text_format": "latex", + "bbox": [ + 560, + 112, + 906, + 147 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Adding $\\mathcal{U}_{i,j}$ refines the position within the feature map grid. Multiplying by 16 maps this position back to the coordinate system of the original images $\\mathcal{I}_Q$ and $\\mathcal{I}_T$ , since each location in the feature map corresponds to a $16\\times 16$ patch in the original images. The corresponding query image position $\\mathcal{M}_{i,j}^{Q}$ is $((i + 0.5)\\times 16,(j + 0.5)\\times 16)$ , which is the center position of each patch in $\\mathcal{I}_Q$ . The correspondence between the $\\mathcal{I}_Q$ and $\\mathcal{I}_T$ is then defined as $\\mathcal{M}_{i,j} = (\\mathcal{M}_{i,j}^{Q},\\mathcal{M}_{i,j}^{T})$ .", + "bbox": [ + 511, + 154, + 906, + 294 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Pose Fitting To estimate the pose in practice, we compute the correspondences $\\mathcal{M}$ for $\\mathcal{I}_Q$ and all $\\mathcal{I}_T$ , and select the template $k$ most similar to the query image. The similarity score $S_{t}$ for each template $t$ is defined as follows:", + "bbox": [ + 511, + 295, + 906, + 354 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nS _ {t} = \\sum_ {i, j} \\left\\{ \\begin{array}{l l} \\max \\left(\\mathcal {C} _ {i, j}\\right) & \\text {i f} c _ {i, j} \\neq K \\\\ 0 & \\text {o t h e r w i s e} \\end{array} . \\right. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 584, + 366, + 905, + 407 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "If the predicted class $c_{i,j} = K$ (representing occlusion or unmatched patches), we exclude $\\max(\\mathcal{C}_{i,j})$ from the sum. By computing $S_t$ for each template, we select the template with the highest similarity score as the best match to $\\mathcal{I}_Q$ . Once the most similar template is determined, we construct 2D-3D correspondences using the correspondences between the query image and the selected template along with the depth information from the template. We estimate the initial pose by applying RANSAC [16] with the EPnP algorithm [28].", + "bbox": [ + 511, + 412, + 905, + 565 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Pose Refinement", + "text_level": 1, + "bbox": [ + 513, + 573, + 678, + 587 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the refinement stage, we accurately refine the pose estimated from the coarse estimation at the pixel-level. To achieve this, we estimate pixel-wise dense correspondences. As shown in Fig. 2, the structure of the refinement model is the same as the coarse estimation model with an added Dense Prediction Transformer (DPT) [48]. The DPT enables pixel-wise prediction, allowing for precise pose refinement. Additionally, by iteratively refining the pose using a render-and-compare approach, this stage overcomes limitations that arise from using pre-rendered templates in the coarse estimation stage, such as self-occlusion.", + "bbox": [ + 511, + 595, + 905, + 762 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Probabilistic Flow Regression Similar to previous works [23, 38], our refinement model estimates the flow between the query image $\\mathcal{I}_Q$ and the rendered image $\\mathcal{I}_R$ to refine the pose. To recover the pose accurately from the flow, it is essential to prevent inaccurate flows from significantly influencing the pose calculation. To this end, PFA [23] relies on RANSAC [16] to probabilistically exclude inaccurate flows from pose estimation. However, because RANSAC is sensitive to outlier distribution, it is insufficient when outliers", + "bbox": [ + 511, + 763, + 906, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "11625", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "are prevalent or unevenly distributed. GenFlow [38] focuses on reducing the influence of inaccurate flows by learning to estimate a confidence map through an end-to-end differentiable system and 6D pose loss. Although they show accurate and robust pose estimation performance, there is still room for improvement since learning confidence does not enhance the accuracy of the flow prediction itself.", + "bbox": [ + 88, + 90, + 480, + 196 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Unlike previous flow-based refinement methods [23, 38], we aim to learn the conditional probability of the flow. In line with several correspondence estimation studies [6, 56], we define this conditional probability as $p(Y|\\mathcal{I}_Q,\\mathcal{I}_R;\\theta)$ , where $Y$ represents the flow between the image pair $(\\mathcal{I}_Q,\\mathcal{I}_R)$ , and $\\theta$ denotes the network parameters. Many methods achieve this by learning the variance of the prediction $Y$ , and the predictive density is modeled using Gaussian or Laplace distributions. We model it as a univariate Laplace distribution to simplify the problem. Specifically, $p(Y|\\mathcal{I}_Q,\\mathcal{I}_R;\\theta)$ is modeled as a Laplace distribution with mean $\\mu \\in \\mathbb{R}^{H\\times W\\times 2}$ and scale $b\\in \\mathbb{R}^{H\\times W\\times 1}$ , both predicted by the network. Formulating flow estimation as a probabilistic regression makes our model concentrate on highly reliable correspondences by adapting the scale parameter $b$ .", + "bbox": [ + 93, + 198, + 482, + 436 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Flow Confidence To compute the pose using the flow estimated by the model, we need a confidence $\\mathcal{W} \\in \\mathbb{R}^{H \\times W \\times 1}$ . This value of $\\mathcal{W}$ determines how much weight each flow error will have in calculating the pose in the differentiable PnP layer. $\\mathcal{W}$ is calculated from the certainty, sensitivity, and flow probability, which are learned through different loss functions. Certainty estimates whether the flow from $\\mathcal{I}_R$ to $\\mathcal{I}_Q$ is occluded or not. Sensitivity is learned from the pose loss and highlights areas with rich textures or extremities of the object. Similar to PDC-Net+ [56], we define the flow probability $P_R$ , which represents the probability that the true flow $y$ is within a radius $R$ from the estimated mean flow $\\mu$ . This can be calculated as follows:", + "bbox": [ + 89, + 440, + 483, + 638 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nP _ {R} = P \\left(\\| y - \\mu \\| _ {1} < R\\right) = 1 - \\exp \\left(- \\frac {R}{b}\\right). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 132, + 659, + 482, + 694 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$P_{R}$ is an interpretable measure of reliability, representing the accuracy of the flow with a threshold $R$ . The flow confidence $\\mathcal{W}$ is calculated as the element-wise multiplication of certainty, sensitivity, and $P_{R}$ . This means that $\\mathcal{W}$ will have higher values when there is no occlusion (high certainty), discriminative information is available to solve the pose (high sensitivity), and the flow is accurate (high $P_{R}$ ).", + "bbox": [ + 89, + 700, + 482, + 821 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Pose Update To compute the refined 6D pose $\\mathbf{P}_{\\mathrm{refined}}$ using the flow $Y$ and flow confidence $\\mathcal{W}$ , we employ a Levenberg-Marquardt (LM) algorithm-based PnP solver. Given the input pose $\\mathbf{P}_{\\mathrm{input}} = [\\mathbf{R}_{\\mathrm{input}} \\mid \\mathbf{t}_{\\mathrm{input}}]$ , the camera intrinsic matrix $\\mathbf{K}$ , and the depth map $\\mathcal{D}_R$ corresponding to $\\mathcal{I}_R$ , we", + "bbox": [ + 89, + 824, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/085c02889f82c2b52104fa7b864cda011a2ea9376379532d59c0c37cd7414e74.jpg", + "image_caption": [ + "Figure 4. Pose Selection. To achieve more precise pose estimation using a multiple hypothesis strategy, we introduce a pose selection stage (Sec 3.3)." + ], + "image_footnote": [], + "bbox": [ + 517, + 89, + 903, + 234 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "calculate the 3D spatial coordinates $\\mathbf{x}_{u,v}^{3\\mathrm{D}}$ corresponding to $\\mathcal{I}_R$ as follows:", + "bbox": [ + 511, + 313, + 906, + 345 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {x} _ {u, v} ^ {\\mathrm {3 D}} = \\mathbf {R} _ {\\text {i n p u t}} ^ {- 1} \\left(\\mathbf {K} ^ {- 1} \\mathcal {D} _ {R} (u, v) \\mathbf {x} _ {u, v} ^ {\\mathrm {2 D}} - \\mathbf {t} _ {\\text {i n p u t}}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 565, + 361, + 906, + 382 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $(u,v)$ are the pixel coordinates in $\\mathcal{I}_R$ , $\\mathcal{D}_R(u,v)$ is the depth value at each pixel, and $\\mathbf{x}_{u,v}^{2\\mathrm{D}} = (u,v,1)^T$ . We optimize the 6D pose by minimizing the sum of squared weighted reprojection errors as follows:", + "bbox": [ + 511, + 388, + 906, + 450 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\underset {\\mathrm {R}, \\mathrm {t}} {\\operatorname {a r g m i n}} \\frac {1}{2} \\sum_ {u} \\sum_ {v} \\| \\mathcal {W} (u, v) \\times (\\pi (\\mathrm {R x} _ {u, v} ^ {\\mathrm {3 D}} + \\mathrm {t}) \\tag {5} \\\\ - \\left((u, v) ^ {T} + Y (u, v))\\right) \\| ^ {2}. \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 557, + 476, + 905, + 532 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here, $\\pi$ is the reprojection function that maps 3D points in camera coordinates to 2D image points using the intrinsic camera parameters $\\mathbf{K}$ .", + "bbox": [ + 511, + 545, + 905, + 590 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Following previous work [7, 38], we update the pose in three iterations using the LM algorithm and further refine it to the final pose using the Gauss-Newton algorithm.", + "bbox": [ + 511, + 592, + 905, + 638 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Pose Selection", + "text_level": 1, + "bbox": [ + 511, + 648, + 658, + 664 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the coarse estimation stage, the best scoring template may not provide the optimal initial pose for refinement. For example, selecting a template rotated by 180 degrees relative to the ground truth (see Fig. 4) makes refinement challenging. To address this issue, many methods [27, 38, 41, 44, 62] employ a multiple hypothesis strategy: generating $N$ coarse pose estimates, refining each, and selecting the best match by comparing rendered results to the query image. Similar to the coarse estimation model, our pose selection model takes the query image and template as inputs but incorporates a score head to estimate a score for each pose hypothesis. Although this increases inference time, considering multiple refined poses helps avoid cases where a difficult-to-refine template leads to a poor final pose.", + "bbox": [ + 511, + 674, + 906, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "11626", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/00e0e8f0a6251e9ce02eeb56aca1c5cb676371e176d8eab60efac44be5741155.jpg", + "image_caption": [ + "Query", + "Figure 5. In-plane Rotation Invariant Matching Example. From left to right: Query image, semi-dense correspondences between the query image and the best scoring template, and the coarse pose recovered using the PnP algorithm." + ], + "image_footnote": [], + "bbox": [ + 99, + 97, + 194, + 305 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/8fe42afed8cf11b18a494bdb2c6a78cc2be0c3ca475d311fabe9801e3e012bbc.jpg", + "image_caption": [ + "Correspondences" + ], + "image_footnote": [], + "bbox": [ + 205, + 98, + 383, + 305 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d4cc4ea731ab12b80c8e7ad17e6a22bc5d63b37f1c0ecc4335f4cbc0abce59eb.jpg", + "image_caption": [ + "Coarse Pose" + ], + "image_footnote": [], + "bbox": [ + 383, + 98, + 473, + 305 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.4. Training", + "text_level": 1, + "bbox": [ + 89, + 417, + 194, + 434 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets To train our three models—the coarse estimator, the pose refiner, and the pose selector—we need RGB-D images with ground-truth 6D object pose annotations. We use the large-scale dataset provided by MegaPose [27]. This dataset consists of synthetic data generated using BlenderProc [8], featuring a diverse set of objects from ShapeNet [5] and Google Scanned Objects (GSO) [11], including comprehensive ground-truth 6D pose annotations and object masks.", + "bbox": [ + 89, + 441, + 483, + 575 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Coarse Model Our coarse estimation model is trained to estimate the correspondence between $\\mathcal{I}_Q$ and $\\mathcal{I}_T$ . Similar to GigaPose [41], we select templates that have similar out-of-plane rotations but different in-plane rotations compared to the training images cropped around the target object. As shown in Fig. 5, our model is designed to estimate correspondences between the query image and the template that are invariant to in-plane rotations. This invariance means that the templates only need to consider out-of-plane rotations, significantly reducing the number of templates required. Since we have the 3D model of the target object, we can generate these correspondences by projecting the 3D model onto the image planes of both the query and template images. Our model is trained to estimate these generated 2D-2D correspondences.", + "bbox": [ + 89, + 580, + 483, + 806 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Refiner Model Our refinement model is trained similarly to previous works [26, 27, 38] that use the render-and-compass are approach. We generate noisy input poses $\\mathbf{P}_{\\mathrm{input}}$ by adding zero-mean Gaussian noise to the ground truth poses $\\mathbf{P}_{\\mathrm{gt}}$ , with translation noise standard deviations of (0.01, 0.01, 0.05) along the x, y, and z axes, and rotation noise", + "bbox": [ + 89, + 810, + 483, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "standard deviations of 15 degrees per axis in Euler angles. The model is trained to predict $\\mathbf{P}_{\\mathrm{gt}}$ from $\\mathbf{P}_{\\mathrm{input}}$ .", + "bbox": [ + 511, + 90, + 903, + 122 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Selection Model Our selection model is trained to assess the similarity between $\\mathcal{I}_Q$ and $\\mathcal{I}_R$ based on $\\mathbf{P}_{\\mathrm{input}}$ to select the most accurate pose from multiple hypotheses. We train the model using binary cross-entropy loss with six poses per $\\mathbf{P}_{\\mathrm{gt}}$ : one positive, five negatives. Positives have translation differences within (0.01, 0.01, 0.05) and rotation differences within 5 degrees of $\\mathbf{P}_{\\mathrm{gt}}$ . By setting small rotation thresholds for positives, we enhance the model's ability to distinguish poses close to the ground truth, improving its discriminative capability.", + "bbox": [ + 511, + 123, + 906, + 275 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.5. Implementation Details", + "text_level": 1, + "bbox": [ + 513, + 285, + 730, + 301 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The encoder and decoder architecture of Co-op are based on CroCo [61], a vision foundation model trained on large-scale datasets for 3D vision tasks. This allows us to significantly leverage the benefits of CroCo pre-training. The coarse model processes input images of resolution $224 \\times 224$ , whereas the refinement and selection models process images of resolution $256 \\times 256$ . For detailed information about the model configurations, learning rates, and training schedules, please refer to the supplementary material.", + "bbox": [ + 511, + 309, + 905, + 445 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Coarse Model In our coarse model, we define the semidense correspondence as a hybrid representation combining patch-level classification and offset regression. Therefore, the coarse model is trained using two loss functions: the classification loss $\\mathcal{L}_{\\mathrm{cls}}$ and the regression loss $\\mathcal{L}_{\\mathrm{reg}}$ .", + "bbox": [ + 511, + 446, + 905, + 523 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "When defining the ground-truth match from the patch center $(i,j)$ in the query image to the template as $\\mathcal{M}_{gt}^T = (\\bar{u},\\bar{v})$ , the ground-truth index for $\\mathcal{L}_{\\mathrm{cls}}$ is defined as $\\left\\lfloor \\frac{\\bar{v}}{16}\\right\\rfloor \\times \\left(\\frac{W}{16}\\right) + \\left\\lfloor \\frac{\\bar{u}}{16}\\right\\rfloor$ when a match exists. Here, $W$ is the width of the template image, and dividing by 16 accounts for the downsizing due to the patch size. When there is no match, the ground-truth index is defined as $\\left(\\frac{H}{16} \\times \\frac{W}{16}\\right) + 1$ , representing an additional class for \"no match.\" Therefore, $\\mathcal{L}_{\\mathrm{cls}}$ is defined as the cross-entropy loss between the ground-truth index and the model's output patch-level probability vector.", + "bbox": [ + 511, + 523, + 905, + 676 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The offset ground truth $\\mathcal{U}_{gt}$ for $\\mathcal{L}_{\\mathrm{reg}}$ is defined as:", + "bbox": [ + 532, + 678, + 859, + 694 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {U} _ {g t} = \\frac {\\mathcal {M} _ {g t} ^ {T}}{1 6} - \\left\\lfloor \\frac {\\mathcal {M} _ {g t} ^ {T}}{1 6} \\right\\rfloor - 0. 5, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 602, + 703, + 905, + 744 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathcal{M}_{gt}^T$ is the ground-truth match in the template. Subtracting $\\left\\lfloor \\mathcal{M}_{gt}^T /16\\right\\rfloor$ and 0.5 centers the offset within the patch, ranging from -0.5 to 0.5. Here, $\\mathcal{U}$ denotes the predicted offset output by the model. The regression loss $\\mathcal{L}_{\\mathrm{reg}}$ is then defined as the L1 loss between $\\mathcal{U}$ and $\\mathcal{U}_{gt}$ :", + "bbox": [ + 511, + 752, + 906, + 830 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {r e g}} = \\left\\| \\mathcal {U} - \\mathcal {U} _ {g t} \\right\\| _ {1}. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 640, + 845, + 903, + 864 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Therefore, the overall loss for training the coarse model, $\\mathcal{L}_{\\mathrm{coarse}}$ , is defined as:", + "bbox": [ + 513, + 869, + 905, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "11627", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {c o a r s e}} = \\mathcal {L} _ {\\text {c l s}} + \\alpha \\mathcal {L} _ {\\text {r e g}}. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 209, + 109, + 482, + 126 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In the above equation, $\\alpha$ is set to 2.", + "bbox": [ + 109, + 138, + 344, + 152 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Refiner Model Our refiner model is trained using flow loss, certainty loss, and pose loss, similar to GenFlow [38]. Since we parameterize the flow output of the network with a Laplace distribution, we train the flow by minimizing the negative log-likelihood. Therefore, the flow loss is defined as follows:", + "bbox": [ + 89, + 156, + 483, + 247 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {f l o w} = \\sum_ {u} \\sum_ {v} \\left[ \\frac {\\left| \\mu_ {u , v} - \\bar {\\mu} _ {u , v} \\right|}{b _ {u , v}} + 2 \\log b _ {u, v} \\right]. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 117, + 273, + 483, + 311 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this equation, $\\mu_{u,v}$ is the flow at pixel position $(u,v)$ within the rendered mask, and $\\bar{\\mu}_{u,v}$ is the ground-truth flow. $b_{u,v}$ is the scale (uncertainty) at pixel position $(u,v)$ . Additionally, the certainty loss $\\mathcal{L}_{\\text{cert}}$ is defined as the binary cross-entropy loss that determines whether the flow from the rendered image $\\mathcal{I}_R$ to the query image $\\mathcal{I}_Q$ falls within the ground-truth mask in $\\mathcal{I}_Q$ or not. The pose loss $\\mathcal{L}_{\\text{pose}}$ which quantifies the difference between the refined pose and the ground-truth pose, is defined as the distance between corresponding 3D points on the 3D model, following the approach used in previous works [27, 38]. Detailed information about $\\mathcal{L}_{\\text{pose}}$ is provided in the supplementary material.", + "bbox": [ + 89, + 325, + 483, + 521 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The overall loss for refiner model is defined as follows:", + "bbox": [ + 109, + 523, + 475, + 537 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {r e f i n e r}} = \\mathcal {L} _ {\\text {f l o w}} + \\beta \\mathcal {L} _ {\\text {c e r t}} + \\gamma \\mathcal {L} _ {\\text {p o s e}}. \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 165, + 559, + 482, + 575 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The weights $\\beta$ and $\\gamma$ for each loss are set to 5 and 20, respectively.", + "bbox": [ + 89, + 587, + 482, + 617 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 633, + 223, + 652 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 89, + 660, + 281, + 678 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Datasets We evaluate our method on the seven core datasets of the BOP Challenge [22]: LineMod Occlusion(LM-O) [3], T-LESS [19], TUD-L [20], IC-BIN [10], ITODD [12], HomebrewedDB(HB) [24], and YCB-Video(YCB-V) [63]. Since our method does not require training for each dataset, we did not use any training images; only the 3D object models and test images were used.", + "bbox": [ + 89, + 685, + 483, + 790 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Metrics We follow the protocol of the BOP Challenge [21], which consists of three pose-error functions: Visible Surface Discrepancy (VSD), Maximum Symmetry-Aware Surface Distance (MSSD), and Maximum Symmetry-Aware Projection Distance (MSPD). The overall accuracy is measured by the Average Recall (AR), calculated as: AR = $(\\mathrm{AR}_{\\mathrm{VSD}} + \\mathrm{AR}_{\\mathrm{MSSD}} + \\mathrm{AR}_{\\mathrm{MSPD}}) / 3$", + "bbox": [ + 89, + 794, + 483, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. BOP Benchmark Results", + "text_level": 1, + "bbox": [ + 513, + 90, + 741, + 104 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 1 shows the pose estimation results on the BOP dataset. In our method's experiments, we use a single RTX 3090 Ti GPU. For a fair comparison with previous works [1, 27, 38, 41, 44], we adopt the default CNOS [40] detections provided by the BOP challenge. We also set the number of hypotheses for the multiple hypotheses strategy to 5, consistent with the previous methods. Under these standardized conditions, our method demonstrates improved pose estimation accuracy compared to prior approaches. Although our method is proposed to solve the RGB-only object pose estimation problem, it is also available for RGBD input as a correspondence-based method. The results for this can be found in the supplementary material.", + "bbox": [ + 511, + 114, + 906, + 325 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2.1 Coarse Estimation", + "text_level": 1, + "bbox": [ + 511, + 349, + 697, + 364 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Rows 1 to 6 show the performance of various methods using only coarse estimation, without any refinement stage. Our method, Co-op (row 6), significantly outperforms previous approaches in this setting. Co-op achieves the highest AR score across all datasets, which is approximately $56.6\\%$ higher than the second-best method, FoundPose [44] (row 5). Although GigaPose [41] is the fastest coarse estimation method with an inference time of 0.384 seconds per image (row 4), Co-op's coarse estimation alone achieves a higher mean AR score than GigaPose combined with the refinement model GenFlow [38] (row 11). This demonstrates that our proposed coarse estimation method is highly effective.", + "bbox": [ + 511, + 376, + 906, + 556 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2.2 Pose Refinement", + "text_level": 1, + "bbox": [ + 511, + 580, + 684, + 594 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Rows 7 to 13 present the results of combining coarse estimation and refinement using a single hypothesis. In this setting, our method, Co-op (row 13), outperforms previous works [27, 38]. To focus on the performance of the refinement model independently of the coarse estimation method, we conducted experiments by combining our refinement model with GigaPose [41] (rows 9, 11, and 12). This allows for a direct comparison of the refinement models using the same coarse estimation. The results show that our refinement model consistently outperforms others, achieving higher AR scores even when paired with the same coarse estimation method. In the multiple hypothesis experiments (rows 14 to 18), our method outperforms the previous best method, FoundPose [44] (row 17), by more than 6 points in AR score.", + "bbox": [ + 511, + 606, + 906, + 832 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 511, + 845, + 666, + 861 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2 presents various ablation evaluations on seven core datasets, as described in Section 4.1. The results for", + "bbox": [ + 511, + 869, + 905, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "11628", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/46836c9aaf47554680d04f2c3ef0f2f14781421139fd55b426b511f0f47bae7c.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
#Coarse estimationPose refinementLM-OT-LESSTUD-LIC-BINITODDHBYCB-VMeanTime(sec)
Coarse estimation only:
1ZS6D [1]-29.821.0----32.4--
2MegaPose [27]-22.917.725.815.210.825.128.120.815.465
3GenFlow [38]-25.021.530.016.815.428.327.723.53.839
4GigaPose [41]-29.927.330.223.118.834.829.027.60.384
5FoundPose [44]-39.733.846.923.920.450.845.237.31.690
6Co-op-59.759.264.245.839.178.162.658.40.979
With pose refinement (a single hypothesis):
7MegaPose [27]MegaPose [27]49.947.765.336.731.565.460.150.931.724
8GenFlow [38]GenFlow [38]54.751.467.043.738.473.061.955.710.553
9GigaPose [41]MegaPose [27]55.754.158.045.037.669.363.254.72.301
10FoundPose [44]MegaPose [27]55.751.063.343.335.769.766.155.06.385
11GigaPose [41]GenFlow [38]59.555.060.747.841.372.260.856.82.213
12GigaPose [41]Co-op60.256.062.850.243.473.961.958.31.149
13Co-opCo-op64.263.571.751.247.383.267.064.01.852
With pose refinement (5 hypotheses):
14MegaPose [27]MegaPose [27]56.050.768.441.433.870.462.154.747.386
15GenFlow [38]GenFlow [38]56.352.368.445.339.573.963.357.020.890
16GigaPose [41]MegaPose [27]59.957.064.546.739.772.266.357.97.682
17FoundPose [44]MegaPose [27]61.057.069.447.940.772.369.059.620.523
18Co-opCo-op65.564.872.954.249.084.968.965.74.186
", + "bbox": [ + 89, + 87, + 906, + 359 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/c4d67d0e6dc8da2b88daa294e67245338945761dc039712eb40a508fb17110da.jpg", + "table_caption": [ + "Table 1. Results on the BOP challenge datasets. The table shows the Average Recall (AR) score for each dataset, the mean AR score across all datasets, and the time required to estimate the poses of all objects within an image. From top to bottom, the results are ordered as follows: coarse estimation only, refinement applied to a single hypothesis, and refinement results with multiple hypotheses." + ], + "table_footnote": [], + "table_body": "
#MethodLM-OT-LESSTUD-LIC-BINITODDHBYCB-VMean
Coarse Estimation:
1Ours (Proposed)59.759.264.245.839.178.162.658.4
2W/o Classification52.656.144.140.331.071.755.850.2
3W/o CroCo pretrained47.248.946.738.127.061.655.046.4
Pose Refinement:
4Ours (Proposed)64.263.571.751.247.383.267.064.0
5W/o Probabilistic Flow63.663.271.649.546.482.865.863.3
6W/o CroCo pretrained62.361.266.949.042.479.667.161.2
Pose Selection:
7Ours (Proposed)65.564.872.954.249.084.968.965.7
8W/o CroCo pretrained65.164.472.853.147.384.268.365.0
", + "bbox": [ + 91, + 446, + 483, + 638 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 2. Ablation Study. We present the results for each stage under various settings. Rows 1 to 3 correspond to Coarse Estimation, rows 4 to 5 to Pose Refinement, and rows 7 and 8 to Pose Selection. Please refer to Section 4.3 for more details.", + "bbox": [ + 89, + 659, + 482, + 715 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "each of our stages are shown in rows 1, 4, 7, respectively.", + "bbox": [ + 89, + 743, + 470, + 758 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Fine-tuning vs. Training from Scratch We assessed the impact of initializing our models with CroCo [61] weights versus random initialization. Initializing with CroCo significantly improved performance in both the coarse estimation (rows 1 vs. 3) and pose refinement stages (rows 4 vs. 6). Although the performance improvement in the pose selection stage (rows 7, 8) is relatively small than others, using CroCo pretraining enhances performance across all stages.", + "bbox": [ + 89, + 761, + 482, + 883 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Hybrid Representation vs. Direct Regression Comparing", + "bbox": [ + 89, + 885, + 482, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "rows 1 and 2, we evaluated the impact of our hybrid representation (combining patch-level classification and offset regression) against direct regression of correspondences. Removing patch-level classification reduced the model's generalization ability, demonstrating the effectiveness of our hybrid approach.", + "bbox": [ + 511, + 449, + 906, + 541 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Probabilistic Flow Regression We evaluated the effectiveness of defining flow as a conditional probability by training a model with the same architecture as Co-op but using the loss function and prediction head from GenFlow [38]. The results (rows 4 vs. 5) show that probabilistic flow regression improves flow learning and pose accuracy by estimating flow reliability, enabling the model to better handle ambiguous or unreliable correspondences.", + "bbox": [ + 511, + 542, + 908, + 666 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 690, + 633, + 705 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We propose Co-op, a novel method for 6D pose estimation of unseen objects. Co-op can estimate the poses of novel objects using CAD models that are not provided during training but are available only during testing. We introduce a coarse estimation method that achieves strong generalization performance by leveraging a hybrid representation. Furthermore, we propose a render-and-compare refiner that employs probabilistic flow. Using these components, we demonstrate that our method significantly outperforms existing methods by a large margin on the BOP benchmark, which comprises hundreds of diverse objects and cluttered scenes.", + "bbox": [ + 511, + 719, + 908, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "11629", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments. This work was supported by the Technology Innovation Program(00417108, Technologies of Rapid Digital Twin Development Tools for Robotic Service Environments) funded By the Ministry of Trade Industry & Energy(MOTIE, Korea)", + "bbox": [ + 89, + 90, + 485, + 167 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 178, + 187, + 194 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Philipp Ausserlechner, David Haberger, Stefan Thalhammer, Jean-Baptiste Weibel, and Markus Vincze. Zs6d: Zero-shot 6d object pose estimation using vision transformers. In ICRA, pages 463-469. IEEE, 2024. 2, 7, 8", + "[2] Herbert Bay, Tinne Tuytelaars, and Luc Van Gool. Surf: Speeded up robust features. In ECCV, pages 404-417, Berlin, Heidelberg, 2006. Springer Berlin Heidelberg. 2", + "[3] Eric Brachmann, Alexander Krull, Frank Michel, Stefan Gumhold, Jamie Shotton, and Carsten Rother. Learning 6d object pose estimation using 3d object coordinates. In ECCV, 2014. 7", + "[4] Michael Calonder, Vincent Lepetit, Mustafa Ozuysal, Tomasz Trzcinski, Christoph Strecha, and Pascal Fua. Brief: Computing a local binary descriptor very fast. IEEE Transactions on Pattern Analysis and Machine Intelligence, 34(7): 1281-1298, 2012. 2", + "[5] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 6", + "[6] Hongkai Chen, Zixin Luo, Lei Zhou, Yurun Tian, Mingmin Zhen, Tian Fang, David McKinnon, Yanghai Tsin, and Long Quan. Aspanformer: Detector-free image matching with adaptive span transformer. In ECCV, 2022. 2, 5", + "[7] Hansheng Chen, Pichao Wang, Fan Wang, Wei Tian, Lu Xiong, and Hao Li. Epro-pnp: Generalized end-to-end probabilistic perspective-n-points for monocular object pose estimation. In CVPR, pages 2781-2790, 2022. 5", + "[8] Maximilian Denninger, Dominik Winkelbauer, Martin Sundermeyer, Wout Boerdijk, Markus Knauer, Klaus H. Strobl, Matthias Humt, and Rudolph Triebel. Blenderproc2: A procedural pipeline for photorealistic rendering. Journal of Open Source Software, 8(82):4901, 2023. 6", + "[9] Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superpoint: Self-supervised interest point detection and description. In CVPRW, pages 337-33712, 2018. 2", + "[10] Andreas Doumanoglou, Rigas Kouskouridas, Sotiris Malassiotis, and Tae-Kyun Kim. Recovering 6d object pose and predicting next-best-view in the crowd. In CVPR, pages 3583-3592, 2016. 7", + "[11] Laura Downs, Anthony Francis, Nate Koenig, Brandon Kinman, Ryan Hickman, Krista Reymann, Thomas B McHugh, and Vincent Vanhoucke. Google scanned objects: A high-quality dataset of 3d scanned household items. In ICRA, pages 2553-2560. IEEE, 2022. 6", + "[12] Bertram Drost, Markus Ulrich, Paul Bergmann, Philipp Hartinger, and Carsten Steger. Introducing mvtec itodd — a dataset for 3d object recognition in industry. In ICCVW, pages 2200-2208, 2017. 7" + ], + "bbox": [ + 93, + 203, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Mihai Dusmanu, Ignacio Rocco, Tomas Pajdla, Marc Pollefeys, Josef Sivic, Akihiko Torii, and Torsten Sattler. D2-Net: A Trainable CNN for Joint Detection and Description of Local Features. In CVPR, 2019. 2", + "[14] Johan Edstedt, Ioannis Athanasiadis, Mårten Wadenback, and Michael Felsberg. DKM: Dense kernelized feature matching for geometry estimation. In CVPR, 2023. 2", + "[15] Johan Edstedt, Qiyu Sun, Georg Bokman, Marten Wadenback, and Michael Felsberg. RoMa: Robust Dense Feature Matching. In CVPR, 2024. 2", + "[16] Martin A. Fischler and Robert C. Bolles. Random sample consensus: A paradigm for model fitting with applications to image analysis and automated cartography. Commun. ACM, 24(6):381-395, 1981. 2, 4", + "[17] Rasmus Laurvig Haugaard and Anders Glent Buch. Surfemb: Dense and continuous correspondence distributions for object pose estimation with learnt surface embeddings. In CVPR, pages 6749-6758, 2022. 2", + "[18] Xingyi He, Jiaming Sun, Yuang Wang, Di Huang, Hujun Bao, and Xiaowei Zhou. Onepose++: Keypoint-free one-shot object pose estimation without CAD models. In NeurIPS, 2022. 1", + "[19] Tomás Hodań, Pavel Haluza, Štepan Obdržálek, Jiří Matas, Manolis Lourakis, and Xenophon Zabulis. T-LESS: An RGB-D dataset for 6D pose estimation of texture-less objects. WACV, 2017. 7", + "[20] Tomás Hodań, Frank Michel, Eric Brachmann, Wadim Kehl, Anders Glent Buch, Dirk Kraft, Bertram Drost, Joel Vidal, Stephan Ihrke, Xenophon Zabulis, Caner Sahin, Fabian Manhardt, Federico Tombari, Tae-Kyun Kim, Jiri Matas, and Carsten Rother. BOP: Benchmark for 6D object pose estimation. ECCV, 2018. 7", + "[21] Tomáš Hodań, Martin Sundermeyer, Bertram Drost, Yann Labbé, Eric Brachmann, Frank Michel, Carsten Rother, and Jiri Matas. Bop challenge 2020 on 6d object localization. In Computer Vision-ECCV 2020 Workshops: Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16, pages 577-594. Springer, 2020. 7", + "[22] Tomas Hodan, Martin Sundermeyer, Yann Labbe, Van Nguyen Nguyen, Gu Wang, Eric Brachmann, Bertram Drost, Vincent Lepetit, Carsten Rother, and Jiri Matas. Bop challenge 2023 on detection segmentation and pose estimation of seen and unseen rigid objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5610-5619, 2024. 7", + "[23] Yinlin Hu, Pascal Fua, and Mathieu Salzmann. Perspective flow aggregation for data-limited 6d object pose estimation. In ECCV, pages 89-106. Springer, 2022. 2, 4, 5", + "[24] Roman Kaskman, Sergey Zakharov, Ivan Shugurov, and Slobodan Ilic. Homebreweddb: Rgb-d dataset for 6d pose estimation of 3d objects. In ICCVW, pages 2767-2776, 2019. 7", + "[25] Wadim Kehl, Fabian Manhardt, Federico Tombari, Slobodan Ilic, and Nassir Navab. Ssd-6d: Making rgb-based 3d detection and 6d pose estimation great again. In ICCV, pages 1521–1529, 2017. 2" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "11630", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[26] Y. Labbe, J. Carpentier, M. Aubry, and J. Sivic. Cosyposse: Consistent multi-view multi-object 6d pose estimation. In ECCV, 2020. 2, 6", + "[27] Yann Labbe, Lucas Manuelli, Arsalan Mousavian, Stephen Tyree, Stan Birchfield, Jonathan Tremblay, Justin Carpentier, Mathieu Aubry, Dieter Fox, and Josef Sivic. MegaPose: 6D Pose Estimation of Novel Objects via Render & Compare. In CoRL, 2022. 1, 2, 5, 6, 7, 8", + "[28] Vincent Lepetit, Francesc Moreno-Noguer, and Pascal Fua. Epnp: An accurate o(n) solution to the pnp problem. International Journal Of Computer Vision, 81:155-166, 2009. 2, 4", + "[29] Fu Li, Shishir Reddy Vutukur, Hao Yu, Ivan Shugurov, Benjamin Busam, Shaowu Yang, and Slobodan Ilic. Nerfpose: A first-reconstruct-then-regress approach for weakly-supervised 6d object pose estimation. In ICCV, pages 2123–2133, 2023. 1", + "[30] Yi Li, Gu Wang, Xiangyang Ji, Yu Xiang, and Dieter Fox. DeepIM: Deep iterative matching for 6D pose estimation. In ECCV, 2018. 2", + "[31] Zhigang Li, Gu Wang, and Xiangyang Ji. Cdpn: Coordinates-based disentangled pose network for real-time rgb-based 6-dof object pose estimation. In ICCV, pages 7678-7687, 2019. 2", + "[32] Jiehong Lin, Lihua Liu, Dekun Lu, and Kui Jia. Sam-6d: Segment anything model meets zero-shot 6d object pose estimation. In CVPR, pages 27906–27916, 2024. 1, 2", + "[33] Yunzhi Lin, Jonathan Tremblay, Stephen Tyree, Patricio A. Vela, and Stan Birchfield. Single-stage keypoint-based category-level object pose estimation from an RGB image. In ICRA, 2022. 1", + "[34] Lahav Lipson, Zachary Teed, Ankit Goyal, and Jia Deng. Coupled iterative refinement for 6d multi-object pose estimation. In CVPR, pages 6728-6737, 2022. 2", + "[35] Xingyu Liu, Ruida Zhang, Chenyangguang Zhang, Bowen Fu, Jiwen Tang, Xiquan Liang, Jingyi Tang, Xiaotian Cheng, Yukang Zhang, Gu Wang, and Xiangyang Ji. Gdrnpp. https://github.com/shanice-1/gdrnpp_bop2022, 2022.1", + "[36] Yuan Liu, Yilin Wen, Sida Peng, Cheng Lin, Xiaoxiao Long, Taku Komura, and Wenping Wang. Gen6d: Generalizable model-free 6-dof object pose estimation from rgb images. In ECCV, 2022. 1", + "[37] David G. Lowe. Distinctive image features from scale-invariant keypoints. International Journal of Computer Vision, 60:91-110, 2004. 2", + "[38] Sungphill Moon, Hyeontae Son, Dongcheol Hur, and Sangwook Kim. Genflow: Generalizable recurrent flow for 6d pose refinement of novel objects. In CVPR, pages 10039-10049, 2024. 1, 2, 4, 5, 6, 7, 8", + "[39] Van Nguyen Nguyen, Yinlin Hu, Yang Xiao, Mathieu Salzmann, and Vincent Lepetit. Templates for 3d object pose estimation revisited: Generalization to new objects and robustness to occlusions. In CVPR, pages 6771-6780, 2022. 2, 3", + "[40] Van Nguyen Nguyen, Thibault Groueix, Georgy Ponimatkin, Vincent Lepetit, and Tomas Hodan. Cnos: A strong baseline" + ], + "bbox": [ + 91, + 90, + 480, + 902 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "for cad-based novel object segmentation. In ICCVW, pages 2134-2140, 2023. 1, 2, 3, 7", + "[41] Van Nguyen Nguyen, Thibault Groueix, Mathieu Salzmann, and Vincent Lepetit. Gigapose: Fast and robust novel object pose estimation via one correspondence. In CVPR, pages 9903-9913, 2024. 1, 2, 3, 5, 6, 7, 8", + "[42] Brian Okorn, Qiao Gu, Martial Hebert, and David Held. Zephyr: Zero-shot pose hypothesis rating. In ICRA, pages 14141-14148. IEEE, 2021. 2", + "[43] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 2", + "[44] Evin Pinar Örnek, Yann Labbe, Bugra Tekin, Lingni Ma, Cem Keskin, Christian Forster, and Tomáš Hodač. Found- pose: Unseen object pose estimation with foundation features. In ECCV, 2024. 1, 2, 5, 7, 8", + "[45] Kiru Park, Timothy Patten, and Markus Vincze. Pix2pose: Pixel-wise coordinate regression of objects for 6d pose estimation. In ICCV, pages 7668-7677, 2019. 2", + "[46] Sida Peng, Yuan Liu, Qixing Huang, Xiaowei Zhou, and Hujun Bao. Pvnet: Pixel-wise voting network for 6dof pose estimation. In CVPR, 2019. 1, 2", + "[47] Mahdi Rad and Vincent Lepetit. Bb8: A scalable, accurate, robust to partial occlusion method for predicting the 3d poses of challenging objects without using depth. In ICCV, 2017. 1, 2", + "[48] René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. In ICCV, pages 12179-12188, 2021. 4", + "[49] Jerome Revaud, Philippe Weinzaepfel, César Roberto de Souza, and Martin Humenberger. R2D2: repeatable and reliable detector and descriptor. In NeurIPS, 2019. 2", + "[50] Ethan Rublee, Vincent Rabaud, Kurt Konolige, and Gary Bradski. Orb: An efficient alternative to sift or surf. In ICCV, pages 2564-2571, 2011. 2", + "[51] Ivan Shugurov, Fu Li, Benjamin Busam, and Slobodan Ilic. Osop: A multi-stage one shot object pose estimation framework. In CVPR, pages 6835-6844, 2022. 2", + "[52] Jiaming Sun, Zehong Shen, Yuang Wang, Hujun Bao, and Xiaowei Zhou. LoFTR: Detector-free local feature matching with transformers. In CVPR, 2021. 2", + "[53] Jiaming Sun, Zihao Wang, Siyu Zhang, Xingyi He, Hongcheng Zhao, Guofeng Zhang, and Xiaowei Zhou. OnePose: One-shot object pose estimation without CAD models. In CVPR, 2022. 1", + "[54] Bugra Tekin, Sudipta N. Sinha, and Pascal Fua. Real-Time Seamless Single Shot 6D Object Pose Prediction. In CVPR, 2018. 1, 2", + "[55] Prune Truong, Martin Danelljan, and Radu Timofte. Glunet: Global-local universal network for dense flow and correspondences. In CVPR, pages 6258-6268, 2020. 2", + "[56] Prune Truong, Martin Danelljan, Radu Timofte, and Luc Van Gool. Pdc-net+: Enhanced probabilistic dense correspondence network. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(8):10247-10266, 2023. 2, 5" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "11631", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[57] Michal Tyszkiiewicz, Pascal Fua, and Eduard Trulls. Disk: Learning local features with policy gradient. In NeurIPS, 2020. 2", + "[58] Gu Wang, Fabian Manhardt, Federico Tombari, and Xi-angyang Ji. GDR-Net: Geometry-guided direct regression network for monocular 6d object pose estimation. In CVPR, pages 16611-16621, 2021. 1", + "[59] He Wang, Srinath Sridhar, Jingwei Huang, Julien Valentin, Shuran Song, and Leonidas J. Guibas. Normalized object coordinate space for category-level 6d object pose and size estimation. In CVPR, 2019. 1", + "[60] Yifan Wang, Xingyi He, Sida Peng, Dongli Tan, and Xiaowei Zhou. Efficient LoFTR: Semi-dense local feature matching with sparse-like speed. In CVPR, 2024. 2", + "[61] Philippe Weinzaepfel, Thomas Lucas, Vincent Leroy, Yohann Cabon, Vaibhav Arora, Romain Brégier, Gabriela Csurka, Leonid Antsfeld, Boris Chidlovskii, and Jérôme Revaud. Croco v2: Improved cross-view completion pretraining for stereo matching and optical flow. In ICCV, pages 17969-17980, 2023. 6, 8", + "[62] Bowen Wen, Wei Yang, Jan Kautz, and Stan Birchfield. Foundationpose: Unified 6d pose estimation and tracking of novel objects. In CVPR, pages 17868-17879, 2024. 1, 2, 5", + "[63] Yu Xiang, Tanner Schmidt, Venkatraman Narayanan, and Dieter Fox. Posecnn: A convolutional neural network for 6d object pose estimation in cluttered scenes. In RSS, 2018. 1, 7" + ], + "bbox": [ + 91, + 92, + 480, + 472 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11632", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Co-op_ Correspondence-based Novel Object Pose Estimation/5a137c7c-5457-4cd6-bb8c-541c7e90dee7_model.json b/2025/Co-op_ Correspondence-based Novel Object Pose Estimation/5a137c7c-5457-4cd6-bb8c-541c7e90dee7_model.json new file mode 100644 index 0000000000000000000000000000000000000000..7eb1f9e2c34230198a08a474296797c26834ba51 --- /dev/null +++ b/2025/Co-op_ Correspondence-based Novel Object Pose Estimation/5a137c7c-5457-4cd6-bb8c-541c7e90dee7_model.json @@ -0,0 +1,2367 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.192, + 0.131, + 0.806, + 0.153 + ], + "angle": 0, + "content": "Co-op: Correspondence-based Novel Object Pose Estimation" + }, + { + "type": "text", + "bbox": [ + 0.19, + 0.181, + 0.808, + 0.216 + ], + "angle": 0, + "content": "Sunghill Moon Hyeontae Son Dongcheol Hur Sangwook Kim Naver LABS" + }, + { + "type": "text", + "bbox": [ + 0.205, + 0.219, + 0.788, + 0.234 + ], + "angle": 0, + "content": "{sungphill.moon, son.ht, dongcheol.hur, o.s.w.a.l.k}@naverlabs.com" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.269, + 0.327, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.302, + 0.485, + 0.62 + ], + "angle": 0, + "content": "We propose Co-op, a novel method for accurately and robustly estimating the 6DoF pose of objects unseen during training from a single RGB image. Our method requires only the CAD model of the target object and can precisely estimate its pose without any additional fine-tuning. While existing model-based methods suffer from inefficiency due to using a large number of templates, our method enables fast and accurate estimation with a small number of templates. This improvement is achieved by finding semidense correspondences between the input image and the pre-rendered templates. Our method achieves strong generalization performance by leveraging a hybrid representation that combines patch-level classification and offset regression. Additionally, our pose refinement model estimates probabilistic flow between the input image and the rendered image, refining the initial estimate to an accurate pose using a differentiable PnP layer. We demonstrate that our method not only estimates object poses rapidly but also outperforms existing methods by a large margin on the seven core datasets of the BOP Challenge, achieving state-of-the-art accuracy." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.649, + 0.222, + 0.663 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.674, + 0.483, + 0.795 + ], + "angle": 0, + "content": "Over the past decade, significant improvements in the accuracy of 6D object pose estimation have been achieved through advancements in deep learning methods [33, 35, 46, 47, 54, 58, 63]. However, despite their remarkable accuracy, these methods are challenging to apply in diverse, real-world environments because handling new objects requires acquiring large amounts of data and retraining the models for several hours or even days." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.483, + 0.901 + ], + "angle": 0, + "content": "To address these challenges, recent studies have explored various methods. One approach is the category-level methods [33, 59], which estimate the pose of unseen objects within known categories but struggle with entirely new ones. Another approach aims for generalization by training on large-scale pose estimation datasets and is divided into two types: model-free and model-based methods. However," + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.271, + 0.573, + 0.447 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.532, + 0.448, + 0.555, + 0.457 + ], + "angle": 0, + "content": "Query" + }, + { + "type": "image", + "bbox": [ + 0.576, + 0.272, + 0.737, + 0.447 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.577, + 0.448, + 0.734, + 0.456 + ], + "angle": 0, + "content": "Semi-dense Correspondences Coarse Pose" + }, + { + "type": "image", + "bbox": [ + 0.744, + 0.272, + 0.902, + 0.447 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.747, + 0.448, + 0.901, + 0.456 + ], + "angle": 0, + "content": "Initialization Dense Flow Refined Pose" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.471, + 0.907, + 0.527 + ], + "angle": 0, + "content": "Figure 1. Examples of 6D pose estimation of novel objects. Our method estimates semi-dense or dense correspondences between the input image and rendered images and uses them to estimate the pose." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.553, + 0.906, + 0.598 + ], + "angle": 0, + "content": "model-free methods [18, 29, 36, 53] often face challenges with low-texture object and report results only on limited datasets or those without occlusions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.6, + 0.906, + 0.75 + ], + "angle": 0, + "content": "Given these limitations, model-based unseen object pose estimation methods [27, 38, 41, 44, 62] that utilize 3D CAD models have been gaining attention. These model-based methods typically consist of three processes—object detection, coarse pose estimation, and pose refinement—or sometimes include a fourth process, pose selection. Since previous works like CNOS [40] and SAM-6D [32] have already addressed object detection with excellent generalization performance, we concentrate our research on the subsequent pose estimation process." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.906, + 0.87 + ], + "angle": 0, + "content": "In this paper, we propose Co-op (Correspondence-based Novel Object Pose Estimation), a novel method for accurately and robustly estimating the 6DoF pose of objects unseen during training from a single RGB image. Our research begins with the consideration of how to effectively define and find correspondences between two images to achieve object pose estimation with excellent generalization performance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.906, + 0.901 + ], + "angle": 0, + "content": "By setting object pose estimation as finding correspondences between two images, our model is trained at the" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11622" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.212 + ], + "angle": 0, + "content": "patch or pixel level rather than the image level. This enables the model to learn low-level information, making it robust to domain shifts. Furthermore, by directly leveraging the geometric and structural information of the 3D model, it enables more accurate pose estimation. In our research, to achieve superior generalization performance, we estimate poses based on correspondences in both stages of pose estimation: coarse estimation and pose refinement." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.213, + 0.482, + 0.422 + ], + "angle": 0, + "content": "In the first stage, coarse estimation, we estimate the initial pose of the object by estimating semi-dense correspondences between multiple templates and the input image. In this stage, we use a hybrid representation that combines patch-level classification and offset regression to estimate the semi-dense correspondences. Then, we estimate the pose using the Perspective-n-Point (PnP) algorithm [28] based on these correspondences. By formulating correspondence estimation not as a direct regression problem but by combining it with classification, we achieve outstanding generalization performance. Thanks to this representation, we achieve high accuracy while using a much smaller number of templates compared to other existing coarse estimation methods." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.424, + 0.482, + 0.529 + ], + "angle": 0, + "content": "In the second stage, pose refinement, we refine the initial pose using a render-and-compare approach. To achieve precise refinement, we estimate probabilistic dense correspondences and learn the confidence in pose estimation by training end-to-end through a differentiable PnP layer. This allows us to robustly estimate poses even in challenging situations such as occlusions and texture-less objects." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.53, + 0.435, + 0.544 + ], + "angle": 0, + "content": "Our contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.545, + 0.481, + 0.619 + ], + "angle": 0, + "content": "- We present Co-op, a novel framework for unseen object pose estimation in RGB-only cases. Co-op does not require additional training or fine-tuning for new objects and outperforms existing methods by a large margin on the seven core datasets of the BOP Challenge." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.621, + 0.481, + 0.664 + ], + "angle": 0, + "content": "- We propose a method for fast and accurate coarse pose estimation using a hybrid representation that combines patch-level classification and offset regression." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.666, + 0.481, + 0.725 + ], + "angle": 0, + "content": "- Additionally, we propose a precise object pose refinement method. It estimates dense correspondences defined by probabilistic flow and learns confidence end-to-end through a differentiable PnP layer." + }, + { + "type": "list", + "bbox": [ + 0.092, + 0.545, + 0.481, + 0.725 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.74, + 0.233, + 0.755 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.482, + 0.901 + ], + "angle": 0, + "content": "Correspondence estimation Traditionally, finding correspondences between two images has been handled using detector-based methods involving three steps: keypoint detection, feature description, and feature matching. These methods have been extensively researched, transitioning from handcrafted approaches [2, 4, 37, 50] to deep learning-based approaches [9, 13, 49, 57]. Recently, detector-free methods [6, 52, 60] have integrated feature detection and matching into a unified process, processing entire images" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.272 + ], + "angle": 0, + "content": "without predefined keypoints and improving matching performance even in low-texture or repetitive areas. Furthermore, dense matching methods [14, 15, 55, 56] find matches for every pixel, enhancing accuracy in tasks such as image-based localization, 3D reconstruction, and pose estimation. Our method follows the detector-free approach in coarse estimation and pose refinement. In the coarse estimation stage, we estimate semi-dense correspondences to achieve textureless-resilient and accurate matching, followed by the pose refinement stage, where the model estimates dense correspondences and their confidence to obtain an exact 6D pose." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.276, + 0.905, + 0.472 + ], + "angle": 0, + "content": "Seen Object Pose Estimation Deep learning-based methods have shown remarkable performance improvements with the growth of high-quality datasets and advances in GPU computing power. Some of these methods regress sparse [25, 46, 47, 54] or dense [17, 31, 45] features of the object, such as keypoints or correspondences and separately recover the object's pose by applying RANSAC [16] and PnP [28] on the regressed features. To achieve higher accuracy, other methods use render-and-compare based refinement models [23, 26, 30, 34]. Because these methods are designed to encode specific object's features into the network weights, they tend to perform poorly when encountering unseen objects." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.476, + 0.905, + 0.536 + ], + "angle": 0, + "content": "Unseen Object Pose Estimation To tackle problems in more practical situations, some studies assume that objects are unknown in advance. In particular, there has been active research on model-based unseen object pose estimation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.538, + 0.905, + 0.778 + ], + "angle": 0, + "content": "Some methods aim to solve the problem using template matching approaches. These methods [27, 38, 39, 42, 51, 62] render the 3D model to generate templates and estimate poses by selecting the most similar template to the query image. Other methods use feature matching approaches, These methods [1, 32, 39, 41, 44] find 2D-3D or 3D-3D correspondences between the 3D model and the query image, recovering poses using the PnP [28] algorithm. To enhance accuracy, several methods [27, 38, 62] apply render-and compare pose refinement, iteratively minimizing discrepancies between the rendered model and the query image. Combined approaches like GenFlow [38] and MegaPose [27] integrate template matching with refinement, achieving remarkable performance but incurring significant computational costs due to the need for numerous template comparisons." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.905, + 0.901 + ], + "angle": 0, + "content": "To address this issue, FoundPose [44] and GigaPose [41] shift from exhaustive template comparison to finding correspondences between the template and the query image, enabling faster and more robust initial pose estimation. These methods depend on the segmentation masks from CNOS [40], to separate the object regions, and perform feature matching between the query image and the template based on DINOv2 [43]. These methods can be" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.956 + ], + "angle": 0, + "content": "11623" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.093, + 0.904, + 0.294 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.096, + 0.302, + 0.904, + 0.509 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.519, + 0.907, + 0.59 + ], + "angle": 0, + "content": "Figure 2. Overview. We estimate object pose through two main stages. In the Coarse Pose Estimation stage (Sec 3.1), we estimate semi-dense correspondences between the query image and templates and compute the initial pose using PnP. In the Pose Refinement stage (Sec 3.2), we refine the initial pose by estimating dense flow between the query and rendered images. Both stages utilize transformer encoders and decoders with identical structures, with the Pose Refinement stage additionally incorporating a DPT module after the decoder for dense prediction." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.615, + 0.483, + 0.661 + ], + "angle": 0, + "content": "considered a detect-and-describe framework using the segmentation mask as a feature detector and the DINOv2 patch descriptor as a feature descriptor." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.668, + 0.484, + 0.745 + ], + "angle": 0, + "content": "On the other hand, we propose a detector-free correspondence estimation framework for the coarse estimation stage. By not relying on segmentation, our method is robust to the noisy or partially detected regions and improves the pose accuracy." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.778, + 0.182, + 0.793 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.484, + 0.903 + ], + "angle": 0, + "content": "In this section, we describe Co-op, our proposed method for unseen object pose estimation. It consists of two main components: coarse pose estimation and pose refinement. Additionally, it optionally incorporates pose selection to estimate the pose more accurately. An overview of our method is presented in Fig. 2." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.615, + 0.733, + 0.63 + ], + "angle": 0, + "content": "3.1. Coarse Pose Estimation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.639, + 0.907, + 0.715 + ], + "angle": 0, + "content": "Template Generation In the coarse estimation stage, we use a set of pre-rendered templates from the 3D model. Similar to GigaPose [41], we minimize the number of templates required for pose estimation by generating those that include only out-of-plane rotations." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.716, + 0.906, + 0.807 + ], + "angle": 0, + "content": "Specifically, we render the CAD model from 42 viewpoints. These viewpoints are defined by subdividing each triangle of Blender's icosphere primitive into four smaller triangles, resulting in a uniform distribution of viewpoints over the sphere, following the method used in previous works [39, 40]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.907, + 0.902 + ], + "angle": 0, + "content": "**Hybrid Representation** The role of the coarse estimation model is to estimate semi-dense correspondences between the query image and the pre-rendered template, denoted by \\(\\mathcal{I}_Q, \\mathcal{I}_T \\in \\mathbb{R}^{H \\times W \\times 3}\\). To enhance the generalization performance of our model, we use a hybrid representation that combines patch-level classification and offset regression." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11624" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.102, + 0.097, + 0.473, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.281, + 0.483, + 0.351 + ], + "angle": 0, + "content": "Figure 3. Visualization of Our Hybrid Representation. Left: Patch-level classification results; matching patches are highlighted with the same color. Right: Offset regression within template patches to refine correspondences; red arrows represent the estimated offsets." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.38, + 0.484, + 0.516 + ], + "angle": 0, + "content": "Fig. 3 illustrates an example of our hybrid representation. On the left, we show partial results of the patch-level classification; patches highlighted with the same color in the query image (left) and the template (right) indicate matched patches. On the right, we show the results of offset regression within the template patches after classification, demonstrating the estimated offset from the center of the template patch that best aligns with the center of the corresponding query patch." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.517, + 0.484, + 0.623 + ], + "angle": 0, + "content": "As shown in Fig. 2, our model consists of a Vision Transformer (ViT) encoder for extracting image features and a transformer decoder for interpreting these features. Here, the subscript \\( i,j \\) represents the 2D location in the feature map. Because downsizing is performed by the encoder, each location \\( i,j \\) in the feature map corresponds to a \\( 16\\times 16 \\) patch in the input image." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.624, + 0.484, + 0.701 + ], + "angle": 0, + "content": "The encoder takes \\(\\mathcal{I}_Q\\) and \\(\\mathcal{I}_T\\) as inputs and extracts the feature maps \\(\\mathcal{F}_Q, \\mathcal{F}_T \\in \\mathbb{R}^{\\frac{H}{16} \\times \\frac{W}{16} \\times 1024}\\), respectively. The decoder and following heads then processes \\(\\mathcal{F}_Q\\) and \\(\\mathcal{F}_T\\) to compute both a tensor for patch-level classification \\(\\mathcal{C} \\in \\mathbb{R}^{\\frac{H}{16} \\times \\frac{W}{16} \\times K}\\) and the xy-offsets \\(\\mathcal{U} \\in \\mathbb{R}^{\\frac{H}{16} \\times \\frac{W}{16} \\times 2}\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.701, + 0.483, + 0.825 + ], + "angle": 0, + "content": "Here, \\( K = \\frac{H}{16} \\times \\frac{W}{16} + 1 \\) represents the number of classes for the patch-level classification. The first \\( \\frac{H}{16} \\times \\frac{W}{16} \\) classes correspond to the template patches, while the last class represents the case where there is no matching part, such as occlusion. For each position \\( (i,j) \\) in the feature map, \\( \\mathcal{C}_{i,j} \\in \\mathbb{R}^K \\) contains classification scores indicating the degree of matchability between the \\( \\frac{H}{16} \\times \\frac{W}{16} \\) template patches and the query patch at that position." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Following this, we define the patch index as \\( c_{i,j} = \\arg \\max_{k\\in 1,2,\\dots,K}\\mathcal{C}_{i,j}^{k} \\) and the offset \\( \\mathcal{U}_{i,j} \\) has a range of \\([-0.5,0.5]\\). At location \\((i,j)\\) where \\( c_{i,j} \\neq K \\), the position \\( M_{i,j}^{T} \\) in the template corresponding to the query patch is given by:" + }, + { + "type": "equation", + "bbox": [ + 0.561, + 0.113, + 0.907, + 0.148 + ], + "angle": 0, + "content": "\\[\n\\mathcal {M} _ {i, j} ^ {T} = (\\left[ \\begin{array}{c} c _ {i, j} \\bmod 1 6 + 0. 5 \\\\ \\lfloor c _ {i, j} / 1 6 \\rfloor + 0. 5 \\end{array} \\right] + \\mathcal {U} _ {i, j}) \\times 1 6. \\qquad (1)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.155, + 0.907, + 0.295 + ], + "angle": 0, + "content": "Adding \\(\\mathcal{U}_{i,j}\\) refines the position within the feature map grid. Multiplying by 16 maps this position back to the coordinate system of the original images \\(\\mathcal{I}_Q\\) and \\(\\mathcal{I}_T\\), since each location in the feature map corresponds to a \\(16\\times 16\\) patch in the original images. The corresponding query image position \\(\\mathcal{M}_{i,j}^{Q}\\) is \\(((i + 0.5)\\times 16,(j + 0.5)\\times 16)\\), which is the center position of each patch in \\(\\mathcal{I}_Q\\). The correspondence between the \\(\\mathcal{I}_Q\\) and \\(\\mathcal{I}_T\\) is then defined as \\(\\mathcal{M}_{i,j} = (\\mathcal{M}_{i,j}^{Q},\\mathcal{M}_{i,j}^{T})\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.296, + 0.907, + 0.356 + ], + "angle": 0, + "content": "Pose Fitting To estimate the pose in practice, we compute the correspondences \\(\\mathcal{M}\\) for \\(\\mathcal{I}_Q\\) and all \\(\\mathcal{I}_T\\), and select the template \\(k\\) most similar to the query image. The similarity score \\(S_{t}\\) for each template \\(t\\) is defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.585, + 0.367, + 0.906, + 0.409 + ], + "angle": 0, + "content": "\\[\nS _ {t} = \\sum_ {i, j} \\left\\{ \\begin{array}{l l} \\max \\left(\\mathcal {C} _ {i, j}\\right) & \\text {i f} c _ {i, j} \\neq K \\\\ 0 & \\text {o t h e r w i s e} \\end{array} . \\right. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.414, + 0.906, + 0.566 + ], + "angle": 0, + "content": "If the predicted class \\( c_{i,j} = K \\) (representing occlusion or unmatched patches), we exclude \\( \\max(\\mathcal{C}_{i,j}) \\) from the sum. By computing \\( S_t \\) for each template, we select the template with the highest similarity score as the best match to \\( \\mathcal{I}_Q \\). Once the most similar template is determined, we construct 2D-3D correspondences using the correspondences between the query image and the selected template along with the depth information from the template. We estimate the initial pose by applying RANSAC [16] with the EPnP algorithm [28]." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.574, + 0.679, + 0.588 + ], + "angle": 0, + "content": "3.2. Pose Refinement" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.597, + 0.906, + 0.763 + ], + "angle": 0, + "content": "In the refinement stage, we accurately refine the pose estimated from the coarse estimation at the pixel-level. To achieve this, we estimate pixel-wise dense correspondences. As shown in Fig. 2, the structure of the refinement model is the same as the coarse estimation model with an added Dense Prediction Transformer (DPT) [48]. The DPT enables pixel-wise prediction, allowing for precise pose refinement. Additionally, by iteratively refining the pose using a render-and-compare approach, this stage overcomes limitations that arise from using pre-rendered templates in the coarse estimation stage, such as self-occlusion." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Probabilistic Flow Regression Similar to previous works [23, 38], our refinement model estimates the flow between the query image \\(\\mathcal{I}_Q\\) and the rendered image \\(\\mathcal{I}_R\\) to refine the pose. To recover the pose accurately from the flow, it is essential to prevent inaccurate flows from significantly influencing the pose calculation. To this end, PFA [23] relies on RANSAC [16] to probabilistically exclude inaccurate flows from pose estimation. However, because RANSAC is sensitive to outlier distribution, it is insufficient when outliers" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11625" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.089, + 0.092, + 0.482, + 0.198 + ], + "angle": 0, + "content": "are prevalent or unevenly distributed. GenFlow [38] focuses on reducing the influence of inaccurate flows by learning to estimate a confidence map through an end-to-end differentiable system and 6D pose loss. Although they show accurate and robust pose estimation performance, there is still room for improvement since learning confidence does not enhance the accuracy of the flow prediction itself." + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.199, + 0.483, + 0.438 + ], + "angle": 0, + "content": "Unlike previous flow-based refinement methods [23, 38], we aim to learn the conditional probability of the flow. In line with several correspondence estimation studies [6, 56], we define this conditional probability as \\( p(Y|\\mathcal{I}_Q,\\mathcal{I}_R;\\theta) \\), where \\( Y \\) represents the flow between the image pair \\( (\\mathcal{I}_Q,\\mathcal{I}_R) \\), and \\( \\theta \\) denotes the network parameters. Many methods achieve this by learning the variance of the prediction \\( Y \\), and the predictive density is modeled using Gaussian or Laplace distributions. We model it as a univariate Laplace distribution to simplify the problem. Specifically, \\( p(Y|\\mathcal{I}_Q,\\mathcal{I}_R;\\theta) \\) is modeled as a Laplace distribution with mean \\( \\mu \\in \\mathbb{R}^{H\\times W\\times 2} \\) and scale \\( b\\in \\mathbb{R}^{H\\times W\\times 1} \\), both predicted by the network. Formulating flow estimation as a probabilistic regression makes our model concentrate on highly reliable correspondences by adapting the scale parameter \\( b \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.441, + 0.484, + 0.639 + ], + "angle": 0, + "content": "Flow Confidence To compute the pose using the flow estimated by the model, we need a confidence \\(\\mathcal{W} \\in \\mathbb{R}^{H \\times W \\times 1}\\). This value of \\(\\mathcal{W}\\) determines how much weight each flow error will have in calculating the pose in the differentiable PnP layer. \\(\\mathcal{W}\\) is calculated from the certainty, sensitivity, and flow probability, which are learned through different loss functions. Certainty estimates whether the flow from \\(\\mathcal{I}_R\\) to \\(\\mathcal{I}_Q\\) is occluded or not. Sensitivity is learned from the pose loss and highlights areas with rich textures or extremities of the object. Similar to PDC-Net+ [56], we define the flow probability \\(P_R\\), which represents the probability that the true flow \\(y\\) is within a radius \\(R\\) from the estimated mean flow \\(\\mu\\). This can be calculated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.133, + 0.66, + 0.483, + 0.695 + ], + "angle": 0, + "content": "\\[\nP _ {R} = P \\left(\\| y - \\mu \\| _ {1} < R\\right) = 1 - \\exp \\left(- \\frac {R}{b}\\right). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.702, + 0.483, + 0.822 + ], + "angle": 0, + "content": "\\(P_{R}\\) is an interpretable measure of reliability, representing the accuracy of the flow with a threshold \\(R\\). The flow confidence \\(\\mathcal{W}\\) is calculated as the element-wise multiplication of certainty, sensitivity, and \\(P_{R}\\). This means that \\(\\mathcal{W}\\) will have higher values when there is no occlusion (high certainty), discriminative information is available to solve the pose (high sensitivity), and the flow is accurate (high \\(P_{R}\\))." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Pose Update To compute the refined 6D pose \\(\\mathbf{P}_{\\mathrm{refined}}\\) using the flow \\(Y\\) and flow confidence \\(\\mathcal{W}\\), we employ a Levenberg-Marquardt (LM) algorithm-based PnP solver. Given the input pose \\(\\mathbf{P}_{\\mathrm{input}} = [\\mathbf{R}_{\\mathrm{input}} \\mid \\mathbf{t}_{\\mathrm{input}}]\\), the camera intrinsic matrix \\(\\mathbf{K}\\), and the depth map \\(\\mathcal{D}_R\\) corresponding to \\(\\mathcal{I}_R\\), we" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.09, + 0.905, + 0.235 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.244, + 0.907, + 0.287 + ], + "angle": 0, + "content": "Figure 4. Pose Selection. To achieve more precise pose estimation using a multiple hypothesis strategy, we introduce a pose selection stage (Sec 3.3)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.314, + 0.907, + 0.346 + ], + "angle": 0, + "content": "calculate the 3D spatial coordinates \\(\\mathbf{x}_{u,v}^{3\\mathrm{D}}\\) corresponding to \\(\\mathcal{I}_R\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.566, + 0.362, + 0.907, + 0.383 + ], + "angle": 0, + "content": "\\[\n\\mathbf {x} _ {u, v} ^ {\\mathrm {3 D}} = \\mathbf {R} _ {\\text {i n p u t}} ^ {- 1} \\left(\\mathbf {K} ^ {- 1} \\mathcal {D} _ {R} (u, v) \\mathbf {x} _ {u, v} ^ {\\mathrm {2 D}} - \\mathbf {t} _ {\\text {i n p u t}}\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.39, + 0.907, + 0.452 + ], + "angle": 0, + "content": "where \\((u,v)\\) are the pixel coordinates in \\(\\mathcal{I}_R\\), \\(\\mathcal{D}_R(u,v)\\) is the depth value at each pixel, and \\(\\mathbf{x}_{u,v}^{2\\mathrm{D}} = (u,v,1)^T\\). We optimize the 6D pose by minimizing the sum of squared weighted reprojection errors as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.558, + 0.477, + 0.906, + 0.533 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\underset {\\mathrm {R}, \\mathrm {t}} {\\operatorname {a r g m i n}} \\frac {1}{2} \\sum_ {u} \\sum_ {v} \\| \\mathcal {W} (u, v) \\times (\\pi (\\mathrm {R x} _ {u, v} ^ {\\mathrm {3 D}} + \\mathrm {t}) \\tag {5} \\\\ - \\left((u, v) ^ {T} + Y (u, v))\\right) \\| ^ {2}. \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.546, + 0.906, + 0.592 + ], + "angle": 0, + "content": "Here, \\(\\pi\\) is the reprojection function that maps 3D points in camera coordinates to 2D image points using the intrinsic camera parameters \\(\\mathbf{K}\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.593, + 0.906, + 0.639 + ], + "angle": 0, + "content": "Following previous work [7, 38], we update the pose in three iterations using the LM algorithm and further refine it to the final pose using the Gauss-Newton algorithm." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.65, + 0.66, + 0.665 + ], + "angle": 0, + "content": "3.3. Pose Selection" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In the coarse estimation stage, the best scoring template may not provide the optimal initial pose for refinement. For example, selecting a template rotated by 180 degrees relative to the ground truth (see Fig. 4) makes refinement challenging. To address this issue, many methods [27, 38, 41, 44, 62] employ a multiple hypothesis strategy: generating \\( N \\) coarse pose estimates, refining each, and selecting the best match by comparing rendered results to the query image. Similar to the coarse estimation model, our pose selection model takes the query image and template as inputs but incorporates a score head to estimate a score for each pose hypothesis. Although this increases inference time, considering multiple refined poses helps avoid cases where a difficult-to-refine template leads to a poor final pose." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11626" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.101, + 0.098, + 0.195, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.128, + 0.308, + 0.166, + 0.32 + ], + "angle": 0, + "content": "Query" + }, + { + "type": "image", + "bbox": [ + 0.206, + 0.099, + 0.385, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.246, + 0.307, + 0.351, + 0.319 + ], + "angle": 0, + "content": "Correspondences" + }, + { + "type": "image", + "bbox": [ + 0.385, + 0.099, + 0.475, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.392, + 0.308, + 0.465, + 0.319 + ], + "angle": 0, + "content": "Coarse Pose" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.335, + 0.483, + 0.392 + ], + "angle": 0, + "content": "Figure 5. In-plane Rotation Invariant Matching Example. From left to right: Query image, semi-dense correspondences between the query image and the best scoring template, and the coarse pose recovered using the PnP algorithm." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.418, + 0.196, + 0.435 + ], + "angle": 0, + "content": "3.4. Training" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.442, + 0.484, + 0.577 + ], + "angle": 0, + "content": "Datasets To train our three models—the coarse estimator, the pose refiner, and the pose selector—we need RGB-D images with ground-truth 6D object pose annotations. We use the large-scale dataset provided by MegaPose [27]. This dataset consists of synthetic data generated using BlenderProc [8], featuring a diverse set of objects from ShapeNet [5] and Google Scanned Objects (GSO) [11], including comprehensive ground-truth 6D pose annotations and object masks." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.581, + 0.484, + 0.807 + ], + "angle": 0, + "content": "Coarse Model Our coarse estimation model is trained to estimate the correspondence between \\(\\mathcal{I}_Q\\) and \\(\\mathcal{I}_T\\). Similar to GigaPose [41], we select templates that have similar out-of-plane rotations but different in-plane rotations compared to the training images cropped around the target object. As shown in Fig. 5, our model is designed to estimate correspondences between the query image and the template that are invariant to in-plane rotations. This invariance means that the templates only need to consider out-of-plane rotations, significantly reducing the number of templates required. Since we have the 3D model of the target object, we can generate these correspondences by projecting the 3D model onto the image planes of both the query and template images. Our model is trained to estimate these generated 2D-2D correspondences." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Refiner Model Our refinement model is trained similarly to previous works [26, 27, 38] that use the render-and-compass are approach. We generate noisy input poses \\(\\mathbf{P}_{\\mathrm{input}}\\) by adding zero-mean Gaussian noise to the ground truth poses \\(\\mathbf{P}_{\\mathrm{gt}}\\), with translation noise standard deviations of (0.01, 0.01, 0.05) along the x, y, and z axes, and rotation noise" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.123 + ], + "angle": 0, + "content": "standard deviations of 15 degrees per axis in Euler angles. The model is trained to predict \\(\\mathbf{P}_{\\mathrm{gt}}\\) from \\(\\mathbf{P}_{\\mathrm{input}}\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.124, + 0.907, + 0.276 + ], + "angle": 0, + "content": "Selection Model Our selection model is trained to assess the similarity between \\(\\mathcal{I}_Q\\) and \\(\\mathcal{I}_R\\) based on \\(\\mathbf{P}_{\\mathrm{input}}\\) to select the most accurate pose from multiple hypotheses. We train the model using binary cross-entropy loss with six poses per \\(\\mathbf{P}_{\\mathrm{gt}}\\): one positive, five negatives. Positives have translation differences within (0.01, 0.01, 0.05) and rotation differences within 5 degrees of \\(\\mathbf{P}_{\\mathrm{gt}}\\). By setting small rotation thresholds for positives, we enhance the model's ability to distinguish poses close to the ground truth, improving its discriminative capability." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.286, + 0.731, + 0.302 + ], + "angle": 0, + "content": "3.5. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.31, + 0.906, + 0.446 + ], + "angle": 0, + "content": "The encoder and decoder architecture of Co-op are based on CroCo [61], a vision foundation model trained on large-scale datasets for 3D vision tasks. This allows us to significantly leverage the benefits of CroCo pre-training. The coarse model processes input images of resolution \\(224 \\times 224\\), whereas the refinement and selection models process images of resolution \\(256 \\times 256\\). For detailed information about the model configurations, learning rates, and training schedules, please refer to the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.448, + 0.906, + 0.524 + ], + "angle": 0, + "content": "Coarse Model In our coarse model, we define the semidense correspondence as a hybrid representation combining patch-level classification and offset regression. Therefore, the coarse model is trained using two loss functions: the classification loss \\(\\mathcal{L}_{\\mathrm{cls}}\\) and the regression loss \\(\\mathcal{L}_{\\mathrm{reg}}\\)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.524, + 0.906, + 0.678 + ], + "angle": 0, + "content": "When defining the ground-truth match from the patch center \\((i,j)\\) in the query image to the template as \\(\\mathcal{M}_{gt}^T = (\\bar{u},\\bar{v})\\), the ground-truth index for \\(\\mathcal{L}_{\\mathrm{cls}}\\) is defined as \\(\\left\\lfloor \\frac{\\bar{v}}{16}\\right\\rfloor \\times \\left(\\frac{W}{16}\\right) + \\left\\lfloor \\frac{\\bar{u}}{16}\\right\\rfloor\\) when a match exists. Here, \\(W\\) is the width of the template image, and dividing by 16 accounts for the downsizing due to the patch size. When there is no match, the ground-truth index is defined as \\(\\left(\\frac{H}{16} \\times \\frac{W}{16}\\right) + 1\\), representing an additional class for \"no match.\" Therefore, \\(\\mathcal{L}_{\\mathrm{cls}}\\) is defined as the cross-entropy loss between the ground-truth index and the model's output patch-level probability vector." + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.679, + 0.86, + 0.695 + ], + "angle": 0, + "content": "The offset ground truth \\(\\mathcal{U}_{gt}\\) for \\(\\mathcal{L}_{\\mathrm{reg}}\\) is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.604, + 0.704, + 0.906, + 0.746 + ], + "angle": 0, + "content": "\\[\n\\mathcal {U} _ {g t} = \\frac {\\mathcal {M} _ {g t} ^ {T}}{1 6} - \\left\\lfloor \\frac {\\mathcal {M} _ {g t} ^ {T}}{1 6} \\right\\rfloor - 0. 5, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.753, + 0.907, + 0.832 + ], + "angle": 0, + "content": "where \\(\\mathcal{M}_{gt}^T\\) is the ground-truth match in the template. Subtracting \\(\\left\\lfloor \\mathcal{M}_{gt}^T /16\\right\\rfloor\\) and 0.5 centers the offset within the patch, ranging from -0.5 to 0.5. Here, \\(\\mathcal{U}\\) denotes the predicted offset output by the model. The regression loss \\(\\mathcal{L}_{\\mathrm{reg}}\\) is then defined as the L1 loss between \\(\\mathcal{U}\\) and \\(\\mathcal{U}_{gt}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.641, + 0.847, + 0.905, + 0.865 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {r e g}} = \\left\\| \\mathcal {U} - \\mathcal {U} _ {g t} \\right\\| _ {1}. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.871, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Therefore, the overall loss for training the coarse model, \\(\\mathcal{L}_{\\mathrm{coarse}}\\), is defined as:" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11627" + } + ], + [ + { + "type": "equation", + "bbox": [ + 0.21, + 0.111, + 0.483, + 0.127 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {c o a r s e}} = \\mathcal {L} _ {\\text {c l s}} + \\alpha \\mathcal {L} _ {\\text {r e g}}. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.139, + 0.345, + 0.154 + ], + "angle": 0, + "content": "In the above equation, \\(\\alpha\\) is set to 2." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.157, + 0.484, + 0.248 + ], + "angle": 0, + "content": "Refiner Model Our refiner model is trained using flow loss, certainty loss, and pose loss, similar to GenFlow [38]. Since we parameterize the flow output of the network with a Laplace distribution, we train the flow by minimizing the negative log-likelihood. Therefore, the flow loss is defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.119, + 0.275, + 0.484, + 0.312 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {f l o w} = \\sum_ {u} \\sum_ {v} \\left[ \\frac {\\left| \\mu_ {u , v} - \\bar {\\mu} _ {u , v} \\right|}{b _ {u , v}} + 2 \\log b _ {u, v} \\right]. \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.327, + 0.484, + 0.522 + ], + "angle": 0, + "content": "In this equation, \\(\\mu_{u,v}\\) is the flow at pixel position \\((u,v)\\) within the rendered mask, and \\(\\bar{\\mu}_{u,v}\\) is the ground-truth flow. \\(b_{u,v}\\) is the scale (uncertainty) at pixel position \\((u,v)\\). Additionally, the certainty loss \\(\\mathcal{L}_{\\text{cert}}\\) is defined as the binary cross-entropy loss that determines whether the flow from the rendered image \\(\\mathcal{I}_R\\) to the query image \\(\\mathcal{I}_Q\\) falls within the ground-truth mask in \\(\\mathcal{I}_Q\\) or not. The pose loss \\(\\mathcal{L}_{\\text{pose}}\\) which quantifies the difference between the refined pose and the ground-truth pose, is defined as the distance between corresponding 3D points on the 3D model, following the approach used in previous works [27, 38]. Detailed information about \\(\\mathcal{L}_{\\text{pose}}\\) is provided in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.111, + 0.524, + 0.476, + 0.539 + ], + "angle": 0, + "content": "The overall loss for refiner model is defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.166, + 0.56, + 0.483, + 0.577 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {r e f i n e r}} = \\mathcal {L} _ {\\text {f l o w}} + \\beta \\mathcal {L} _ {\\text {c e r t}} + \\gamma \\mathcal {L} _ {\\text {p o s e}}. \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.588, + 0.483, + 0.618 + ], + "angle": 0, + "content": "The weights \\(\\beta\\) and \\(\\gamma\\) for each loss are set to 5 and 20, respectively." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.635, + 0.224, + 0.653 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.661, + 0.282, + 0.679 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.686, + 0.484, + 0.791 + ], + "angle": 0, + "content": "Datasets We evaluate our method on the seven core datasets of the BOP Challenge [22]: LineMod Occlusion(LM-O) [3], T-LESS [19], TUD-L [20], IC-BIN [10], ITODD [12], HomebrewedDB(HB) [24], and YCB-Video(YCB-V) [63]. Since our method does not require training for each dataset, we did not use any training images; only the 3D object models and test images were used." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Metrics We follow the protocol of the BOP Challenge [21], which consists of three pose-error functions: Visible Surface Discrepancy (VSD), Maximum Symmetry-Aware Surface Distance (MSSD), and Maximum Symmetry-Aware Projection Distance (MSPD). The overall accuracy is measured by the Average Recall (AR), calculated as: AR = \\((\\mathrm{AR}_{\\mathrm{VSD}} + \\mathrm{AR}_{\\mathrm{MSSD}} + \\mathrm{AR}_{\\mathrm{MSPD}}) / 3\\)" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.091, + 0.743, + 0.106 + ], + "angle": 0, + "content": "4.2. BOP Benchmark Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.115, + 0.907, + 0.326 + ], + "angle": 0, + "content": "Table 1 shows the pose estimation results on the BOP dataset. In our method's experiments, we use a single RTX 3090 Ti GPU. For a fair comparison with previous works [1, 27, 38, 41, 44], we adopt the default CNOS [40] detections provided by the BOP challenge. We also set the number of hypotheses for the multiple hypotheses strategy to 5, consistent with the previous methods. Under these standardized conditions, our method demonstrates improved pose estimation accuracy compared to prior approaches. Although our method is proposed to solve the RGB-only object pose estimation problem, it is also available for RGBD input as a correspondence-based method. The results for this can be found in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.351, + 0.698, + 0.365 + ], + "angle": 0, + "content": "4.2.1 Coarse Estimation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.377, + 0.907, + 0.558 + ], + "angle": 0, + "content": "Rows 1 to 6 show the performance of various methods using only coarse estimation, without any refinement stage. Our method, Co-op (row 6), significantly outperforms previous approaches in this setting. Co-op achieves the highest AR score across all datasets, which is approximately \\(56.6\\%\\) higher than the second-best method, FoundPose [44] (row 5). Although GigaPose [41] is the fastest coarse estimation method with an inference time of 0.384 seconds per image (row 4), Co-op's coarse estimation alone achieves a higher mean AR score than GigaPose combined with the refinement model GenFlow [38] (row 11). This demonstrates that our proposed coarse estimation method is highly effective." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.582, + 0.685, + 0.595 + ], + "angle": 0, + "content": "4.2.2 Pose Refinement" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.607, + 0.907, + 0.833 + ], + "angle": 0, + "content": "Rows 7 to 13 present the results of combining coarse estimation and refinement using a single hypothesis. In this setting, our method, Co-op (row 13), outperforms previous works [27, 38]. To focus on the performance of the refinement model independently of the coarse estimation method, we conducted experiments by combining our refinement model with GigaPose [41] (rows 9, 11, and 12). This allows for a direct comparison of the refinement models using the same coarse estimation. The results show that our refinement model consistently outperforms others, achieving higher AR scores even when paired with the same coarse estimation method. In the multiple hypothesis experiments (rows 14 to 18), our method outperforms the previous best method, FoundPose [44] (row 17), by more than 6 points in AR score." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.847, + 0.667, + 0.862 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Table 2 presents various ablation evaluations on seven core datasets, as described in Section 4.1. The results for" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11628" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.091, + 0.088, + 0.907, + 0.361 + ], + "angle": 0, + "content": "
#Coarse estimationPose refinementLM-OT-LESSTUD-LIC-BINITODDHBYCB-VMeanTime(sec)
Coarse estimation only:
1ZS6D [1]-29.821.0----32.4--
2MegaPose [27]-22.917.725.815.210.825.128.120.815.465
3GenFlow [38]-25.021.530.016.815.428.327.723.53.839
4GigaPose [41]-29.927.330.223.118.834.829.027.60.384
5FoundPose [44]-39.733.846.923.920.450.845.237.31.690
6Co-op-59.759.264.245.839.178.162.658.40.979
With pose refinement (a single hypothesis):
7MegaPose [27]MegaPose [27]49.947.765.336.731.565.460.150.931.724
8GenFlow [38]GenFlow [38]54.751.467.043.738.473.061.955.710.553
9GigaPose [41]MegaPose [27]55.754.158.045.037.669.363.254.72.301
10FoundPose [44]MegaPose [27]55.751.063.343.335.769.766.155.06.385
11GigaPose [41]GenFlow [38]59.555.060.747.841.372.260.856.82.213
12GigaPose [41]Co-op60.256.062.850.243.473.961.958.31.149
13Co-opCo-op64.263.571.751.247.383.267.064.01.852
With pose refinement (5 hypotheses):
14MegaPose [27]MegaPose [27]56.050.768.441.433.870.462.154.747.386
15GenFlow [38]GenFlow [38]56.352.368.445.339.573.963.357.020.890
16GigaPose [41]MegaPose [27]59.957.064.546.739.772.266.357.97.682
17FoundPose [44]MegaPose [27]61.057.069.447.940.772.369.059.620.523
18Co-opCo-op65.564.872.954.249.084.968.965.74.186
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.381, + 0.907, + 0.424 + ], + "angle": 0, + "content": "Table 1. Results on the BOP challenge datasets. The table shows the Average Recall (AR) score for each dataset, the mean AR score across all datasets, and the time required to estimate the poses of all objects within an image. From top to bottom, the results are ordered as follows: coarse estimation only, refinement applied to a single hypothesis, and refinement results with multiple hypotheses." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.448, + 0.485, + 0.64 + ], + "angle": 0, + "content": "
#MethodLM-OT-LESSTUD-LIC-BINITODDHBYCB-VMean
Coarse Estimation:
1Ours (Proposed)59.759.264.245.839.178.162.658.4
2W/o Classification52.656.144.140.331.071.755.850.2
3W/o CroCo pretrained47.248.946.738.127.061.655.046.4
Pose Refinement:
4Ours (Proposed)64.263.571.751.247.383.267.064.0
5W/o Probabilistic Flow63.663.271.649.546.482.865.863.3
6W/o CroCo pretrained62.361.266.949.042.479.667.161.2
Pose Selection:
7Ours (Proposed)65.564.872.954.249.084.968.965.7
8W/o CroCo pretrained65.164.472.853.147.384.268.365.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.66, + 0.483, + 0.716 + ], + "angle": 0, + "content": "Table 2. Ablation Study. We present the results for each stage under various settings. Rows 1 to 3 correspond to Coarse Estimation, rows 4 to 5 to Pose Refinement, and rows 7 and 8 to Pose Selection. Please refer to Section 4.3 for more details." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.744, + 0.471, + 0.76 + ], + "angle": 0, + "content": "each of our stages are shown in rows 1, 4, 7, respectively." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.762, + 0.483, + 0.884 + ], + "angle": 0, + "content": "Fine-tuning vs. Training from Scratch We assessed the impact of initializing our models with CroCo [61] weights versus random initialization. Initializing with CroCo significantly improved performance in both the coarse estimation (rows 1 vs. 3) and pose refinement stages (rows 4 vs. 6). Although the performance improvement in the pose selection stage (rows 7, 8) is relatively small than others, using CroCo pretraining enhances performance across all stages." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.886, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Hybrid Representation vs. Direct Regression Comparing" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.45, + 0.907, + 0.542 + ], + "angle": 0, + "content": "rows 1 and 2, we evaluated the impact of our hybrid representation (combining patch-level classification and offset regression) against direct regression of correspondences. Removing patch-level classification reduced the model's generalization ability, demonstrating the effectiveness of our hybrid approach." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.544, + 0.909, + 0.667 + ], + "angle": 0, + "content": "Probabilistic Flow Regression We evaluated the effectiveness of defining flow as a conditional probability by training a model with the same architecture as Co-op but using the loss function and prediction head from GenFlow [38]. The results (rows 4 vs. 5) show that probabilistic flow regression improves flow learning and pose accuracy by estimating flow reliability, enabling the model to better handle ambiguous or unreliable correspondences." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.691, + 0.634, + 0.707 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.909, + 0.901 + ], + "angle": 0, + "content": "We propose Co-op, a novel method for 6D pose estimation of unseen objects. Co-op can estimate the poses of novel objects using CAD models that are not provided during training but are available only during testing. We introduce a coarse estimation method that achieves strong generalization performance by leveraging a hybrid representation. Furthermore, we propose a render-and-compare refiner that employs probabilistic flow. Using these components, we demonstrate that our method significantly outperforms existing methods by a large margin on the BOP benchmark, which comprises hundreds of diverse objects and cluttered scenes." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "11629" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.168 + ], + "angle": 0, + "content": "Acknowledgments. This work was supported by the Technology Innovation Program(00417108, Technologies of Rapid Digital Twin Development Tools for Robotic Service Environments) funded By the Ministry of Trade Industry & Energy(MOTIE, Korea)" + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.179, + 0.188, + 0.195 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.204, + 0.484, + 0.259 + ], + "angle": 0, + "content": "[1] Philipp Ausserlechner, David Haberger, Stefan Thalhammer, Jean-Baptiste Weibel, and Markus Vincze. Zs6d: Zero-shot 6d object pose estimation using vision transformers. In ICRA, pages 463-469. IEEE, 2024. 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.261, + 0.484, + 0.302 + ], + "angle": 0, + "content": "[2] Herbert Bay, Tinne Tuytelaars, and Luc Van Gool. Surf: Speeded up robust features. In ECCV, pages 404-417, Berlin, Heidelberg, 2006. Springer Berlin Heidelberg. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.303, + 0.484, + 0.357 + ], + "angle": 0, + "content": "[3] Eric Brachmann, Alexander Krull, Frank Michel, Stefan Gumhold, Jamie Shotton, and Carsten Rother. Learning 6d object pose estimation using 3d object coordinates. In ECCV, 2014. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.359, + 0.484, + 0.426 + ], + "angle": 0, + "content": "[4] Michael Calonder, Vincent Lepetit, Mustafa Ozuysal, Tomasz Trzcinski, Christoph Strecha, and Pascal Fua. Brief: Computing a local binary descriptor very fast. IEEE Transactions on Pattern Analysis and Machine Intelligence, 34(7): 1281-1298, 2012. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.428, + 0.484, + 0.496 + ], + "angle": 0, + "content": "[5] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.498, + 0.484, + 0.552 + ], + "angle": 0, + "content": "[6] Hongkai Chen, Zixin Luo, Lei Zhou, Yurun Tian, Mingmin Zhen, Tian Fang, David McKinnon, Yanghai Tsin, and Long Quan. Aspanformer: Detector-free image matching with adaptive span transformer. In ECCV, 2022. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.554, + 0.484, + 0.608 + ], + "angle": 0, + "content": "[7] Hansheng Chen, Pichao Wang, Fan Wang, Wei Tian, Lu Xiong, and Hao Li. Epro-pnp: Generalized end-to-end probabilistic perspective-n-points for monocular object pose estimation. In CVPR, pages 2781-2790, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.609, + 0.484, + 0.678 + ], + "angle": 0, + "content": "[8] Maximilian Denninger, Dominik Winkelbauer, Martin Sundermeyer, Wout Boerdijk, Markus Knauer, Klaus H. Strobl, Matthias Humt, and Rudolph Triebel. Blenderproc2: A procedural pipeline for photorealistic rendering. Journal of Open Source Software, 8(82):4901, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.679, + 0.484, + 0.72 + ], + "angle": 0, + "content": "[9] Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superpoint: Self-supervised interest point detection and description. In CVPRW, pages 337-33712, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.721, + 0.484, + 0.775 + ], + "angle": 0, + "content": "[10] Andreas Doumanoglou, Rigas Kouskouridas, Sotiris Malassiotis, and Tae-Kyun Kim. Recovering 6d object pose and predicting next-best-view in the crowd. In CVPR, pages 3583-3592, 2016. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.777, + 0.484, + 0.845 + ], + "angle": 0, + "content": "[11] Laura Downs, Anthony Francis, Nate Koenig, Brandon Kinman, Ryan Hickman, Krista Reymann, Thomas B McHugh, and Vincent Vanhoucke. Google scanned objects: A high-quality dataset of 3d scanned household items. In ICRA, pages 2553-2560. IEEE, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[12] Bertram Drost, Markus Ulrich, Paul Bergmann, Philipp Hartinger, and Carsten Steger. Introducing mvtec itodd — a dataset for 3d object recognition in industry. In ICCVW, pages 2200-2208, 2017. 7" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.204, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[13] Mihai Dusmanu, Ignacio Rocco, Tomas Pajdla, Marc Pollefeys, Josef Sivic, Akihiko Torii, and Torsten Sattler. D2-Net: A Trainable CNN for Joint Detection and Description of Local Features. In CVPR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.151, + 0.905, + 0.192 + ], + "angle": 0, + "content": "[14] Johan Edstedt, Ioannis Athanasiadis, Mårten Wadenback, and Michael Felsberg. DKM: Dense kernelized feature matching for geometry estimation. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.196, + 0.905, + 0.237 + ], + "angle": 0, + "content": "[15] Johan Edstedt, Qiyu Sun, Georg Bokman, Marten Wadenback, and Michael Felsberg. RoMa: Robust Dense Feature Matching. In CVPR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.24, + 0.905, + 0.295 + ], + "angle": 0, + "content": "[16] Martin A. Fischler and Robert C. Bolles. Random sample consensus: A paradigm for model fitting with applications to image analysis and automated cartography. Commun. ACM, 24(6):381-395, 1981. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.298, + 0.905, + 0.354 + ], + "angle": 0, + "content": "[17] Rasmus Laurvig Haugaard and Anders Glent Buch. Surfemb: Dense and continuous correspondence distributions for object pose estimation with learnt surface embeddings. In CVPR, pages 6749-6758, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.355, + 0.905, + 0.411 + ], + "angle": 0, + "content": "[18] Xingyi He, Jiaming Sun, Yuang Wang, Di Huang, Hujun Bao, and Xiaowei Zhou. Onepose++: Keypoint-free one-shot object pose estimation without CAD models. In NeurIPS, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.413, + 0.905, + 0.469 + ], + "angle": 0, + "content": "[19] Tomás Hodań, Pavel Haluza, Štepan Obdržálek, Jiří Matas, Manolis Lourakis, and Xenophon Zabulis. T-LESS: An RGB-D dataset for 6D pose estimation of texture-less objects. WACV, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.472, + 0.905, + 0.555 + ], + "angle": 0, + "content": "[20] Tomás Hodań, Frank Michel, Eric Brachmann, Wadim Kehl, Anders Glent Buch, Dirk Kraft, Bertram Drost, Joel Vidal, Stephan Ihrke, Xenophon Zabulis, Caner Sahin, Fabian Manhardt, Federico Tombari, Tae-Kyun Kim, Jiri Matas, and Carsten Rother. BOP: Benchmark for 6D object pose estimation. ECCV, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.557, + 0.905, + 0.641 + ], + "angle": 0, + "content": "[21] Tomáš Hodań, Martin Sundermeyer, Bertram Drost, Yann Labbé, Eric Brachmann, Frank Michel, Carsten Rother, and Jiri Matas. Bop challenge 2020 on 6d object localization. In Computer Vision-ECCV 2020 Workshops: Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16, pages 577-594. Springer, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.643, + 0.905, + 0.74 + ], + "angle": 0, + "content": "[22] Tomas Hodan, Martin Sundermeyer, Yann Labbe, Van Nguyen Nguyen, Gu Wang, Eric Brachmann, Bertram Drost, Vincent Lepetit, Carsten Rother, and Jiri Matas. Bop challenge 2023 on detection segmentation and pose estimation of seen and unseen rigid objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5610-5619, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.743, + 0.905, + 0.785 + ], + "angle": 0, + "content": "[23] Yinlin Hu, Pascal Fua, and Mathieu Salzmann. Perspective flow aggregation for data-limited 6d object pose estimation. In ECCV, pages 89-106. Springer, 2022. 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.788, + 0.905, + 0.842 + ], + "angle": 0, + "content": "[24] Roman Kaskman, Sergey Zakharov, Ivan Shugurov, and Slobodan Ilic. Homebreweddb: Rgb-d dataset for 6d pose estimation of 3d objects. In ICCVW, pages 2767-2776, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.845, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[25] Wadim Kehl, Fabian Manhardt, Federico Tombari, Slobodan Ilic, and Nassir Navab. Ssd-6d: Making rgb-based 3d detection and 6d pose estimation great again. In ICCV, pages 1521–1529, 2017. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11630" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.134 + ], + "angle": 0, + "content": "[26] Y. Labbe, J. Carpentier, M. Aubry, and J. Sivic. Cosyposse: Consistent multi-view multi-object 6d pose estimation. In ECCV, 2020. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.482, + 0.205 + ], + "angle": 0, + "content": "[27] Yann Labbe, Lucas Manuelli, Arsalan Mousavian, Stephen Tyree, Stan Birchfield, Jonathan Tremblay, Justin Carpentier, Mathieu Aubry, Dieter Fox, and Josef Sivic. MegaPose: 6D Pose Estimation of Novel Objects via Render & Compare. In CoRL, 2022. 1, 2, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.207, + 0.482, + 0.261 + ], + "angle": 0, + "content": "[28] Vincent Lepetit, Francesc Moreno-Noguer, and Pascal Fua. Epnp: An accurate o(n) solution to the pnp problem. International Journal Of Computer Vision, 81:155-166, 2009. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.263, + 0.482, + 0.331 + ], + "angle": 0, + "content": "[29] Fu Li, Shishir Reddy Vutukur, Hao Yu, Ivan Shugurov, Benjamin Busam, Shaowu Yang, and Slobodan Ilic. Nerfpose: A first-reconstruct-then-regress approach for weakly-supervised 6d object pose estimation. In ICCV, pages 2123–2133, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.333, + 0.482, + 0.374 + ], + "angle": 0, + "content": "[30] Yi Li, Gu Wang, Xiangyang Ji, Yu Xiang, and Dieter Fox. DeepIM: Deep iterative matching for 6D pose estimation. In ECCV, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.376, + 0.482, + 0.432 + ], + "angle": 0, + "content": "[31] Zhigang Li, Gu Wang, and Xiangyang Ji. Cdpn: Coordinates-based disentangled pose network for real-time rgb-based 6-dof object pose estimation. In ICCV, pages 7678-7687, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.434, + 0.482, + 0.475 + ], + "angle": 0, + "content": "[32] Jiehong Lin, Lihua Liu, Dekun Lu, and Kui Jia. Sam-6d: Segment anything model meets zero-shot 6d object pose estimation. In CVPR, pages 27906–27916, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.476, + 0.482, + 0.531 + ], + "angle": 0, + "content": "[33] Yunzhi Lin, Jonathan Tremblay, Stephen Tyree, Patricio A. Vela, and Stan Birchfield. Single-stage keypoint-based category-level object pose estimation from an RGB image. In ICRA, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.533, + 0.482, + 0.575 + ], + "angle": 0, + "content": "[34] Lahav Lipson, Zachary Teed, Ankit Goyal, and Jia Deng. Coupled iterative refinement for 6d multi-object pose estimation. In CVPR, pages 6728-6737, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.576, + 0.482, + 0.645 + ], + "angle": 0, + "content": "[35] Xingyu Liu, Ruida Zhang, Chenyangguang Zhang, Bowen Fu, Jiwen Tang, Xiquan Liang, Jingyi Tang, Xiaotian Cheng, Yukang Zhang, Gu Wang, and Xiangyang Ji. Gdrnpp. https://github.com/shanice-1/gdrnpp_bop2022, 2022.1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.647, + 0.482, + 0.701 + ], + "angle": 0, + "content": "[36] Yuan Liu, Yilin Wen, Sida Peng, Cheng Lin, Xiaoxiao Long, Taku Komura, and Wenping Wang. Gen6d: Generalizable model-free 6-dof object pose estimation from rgb images. In ECCV, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.703, + 0.482, + 0.744 + ], + "angle": 0, + "content": "[37] David G. Lowe. Distinctive image features from scale-invariant keypoints. International Journal of Computer Vision, 60:91-110, 2004. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.746, + 0.482, + 0.801 + ], + "angle": 0, + "content": "[38] Sungphill Moon, Hyeontae Son, Dongcheol Hur, and Sangwook Kim. Genflow: Generalizable recurrent flow for 6d pose refinement of novel objects. In CVPR, pages 10039-10049, 2024. 1, 2, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.803, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[39] Van Nguyen Nguyen, Yinlin Hu, Yang Xiao, Mathieu Salzmann, and Vincent Lepetit. Templates for 3d object pose estimation revisited: Generalization to new objects and robustness to occlusions. In CVPR, pages 6771-6780, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.903 + ], + "angle": 0, + "content": "[40] Van Nguyen Nguyen, Thibault Groueix, Georgy Ponimatkin, Vincent Lepetit, and Tomas Hodan. Cnos: A strong baseline" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "for cad-based novel object segmentation. In ICCVW, pages 2134-2140, 2023. 1, 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.177 + ], + "angle": 0, + "content": "[41] Van Nguyen Nguyen, Thibault Groueix, Mathieu Salzmann, and Vincent Lepetit. Gigapose: Fast and robust novel object pose estimation via one correspondence. In CVPR, pages 9903-9913, 2024. 1, 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.905, + 0.219 + ], + "angle": 0, + "content": "[42] Brian Okorn, Qiao Gu, Martial Hebert, and David Held. Zephyr: Zero-shot pose hypothesis rating. In ICRA, pages 14141-14148. IEEE, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.221, + 0.905, + 0.29 + ], + "angle": 0, + "content": "[43] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.292, + 0.905, + 0.347 + ], + "angle": 0, + "content": "[44] Evin Pinar Örnek, Yann Labbe, Bugra Tekin, Lingni Ma, Cem Keskin, Christian Forster, and Tomáš Hodač. Found- pose: Unseen object pose estimation with foundation features. In ECCV, 2024. 1, 2, 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.349, + 0.905, + 0.389 + ], + "angle": 0, + "content": "[45] Kiru Park, Timothy Patten, and Markus Vincze. Pix2pose: Pixel-wise coordinate regression of objects for 6d pose estimation. In ICCV, pages 7668-7677, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.391, + 0.905, + 0.432 + ], + "angle": 0, + "content": "[46] Sida Peng, Yuan Liu, Qixing Huang, Xiaowei Zhou, and Hujun Bao. Pvnet: Pixel-wise voting network for 6dof pose estimation. In CVPR, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.434, + 0.905, + 0.488 + ], + "angle": 0, + "content": "[47] Mahdi Rad and Vincent Lepetit. Bb8: A scalable, accurate, robust to partial occlusion method for predicting the 3d poses of challenging objects without using depth. In ICCV, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.49, + 0.905, + 0.531 + ], + "angle": 0, + "content": "[48] René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. In ICCV, pages 12179-12188, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.533, + 0.905, + 0.574 + ], + "angle": 0, + "content": "[49] Jerome Revaud, Philippe Weinzaepfel, César Roberto de Souza, and Martin Humenberger. R2D2: repeatable and reliable detector and descriptor. In NeurIPS, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.576, + 0.905, + 0.617 + ], + "angle": 0, + "content": "[50] Ethan Rublee, Vincent Rabaud, Kurt Konolige, and Gary Bradski. Orb: An efficient alternative to sift or surf. In ICCV, pages 2564-2571, 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.619, + 0.905, + 0.659 + ], + "angle": 0, + "content": "[51] Ivan Shugurov, Fu Li, Benjamin Busam, and Slobodan Ilic. Osop: A multi-stage one shot object pose estimation framework. In CVPR, pages 6835-6844, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.661, + 0.905, + 0.702 + ], + "angle": 0, + "content": "[52] Jiaming Sun, Zehong Shen, Yuang Wang, Hujun Bao, and Xiaowei Zhou. LoFTR: Detector-free local feature matching with transformers. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.704, + 0.905, + 0.758 + ], + "angle": 0, + "content": "[53] Jiaming Sun, Zihao Wang, Siyu Zhang, Xingyi He, Hongcheng Zhao, Guofeng Zhang, and Xiaowei Zhou. OnePose: One-shot object pose estimation without CAD models. In CVPR, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.76, + 0.905, + 0.801 + ], + "angle": 0, + "content": "[54] Bugra Tekin, Sudipta N. Sinha, and Pascal Fua. Real-Time Seamless Single Shot 6D Object Pose Prediction. In CVPR, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.803, + 0.905, + 0.845 + ], + "angle": 0, + "content": "[55] Prune Truong, Martin Danelljan, and Radu Timofte. Glunet: Global-local universal network for dense flow and correspondences. In CVPR, pages 6258-6268, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.847, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[56] Prune Truong, Martin Danelljan, Radu Timofte, and Luc Van Gool. Pdc-net+: Enhanced probabilistic dense correspondence network. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(8):10247-10266, 2023. 2, 5" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "11631" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.133 + ], + "angle": 0, + "content": "[57] Michal Tyszkiiewicz, Pascal Fua, and Eduard Trulls. Disk: Learning local features with policy gradient. In NeurIPS, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.482, + 0.19 + ], + "angle": 0, + "content": "[58] Gu Wang, Fabian Manhardt, Federico Tombari, and Xi-angyang Ji. GDR-Net: Geometry-guided direct regression network for monocular 6d object pose estimation. In CVPR, pages 16611-16621, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.193, + 0.482, + 0.247 + ], + "angle": 0, + "content": "[59] He Wang, Srinath Sridhar, Jingwei Huang, Julien Valentin, Shuran Song, and Leonidas J. Guibas. Normalized object coordinate space for category-level 6d object pose and size estimation. In CVPR, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.249, + 0.482, + 0.29 + ], + "angle": 0, + "content": "[60] Yifan Wang, Xingyi He, Sida Peng, Dongli Tan, and Xiaowei Zhou. Efficient LoFTR: Semi-dense local feature matching with sparse-like speed. In CVPR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.292, + 0.482, + 0.373 + ], + "angle": 0, + "content": "[61] Philippe Weinzaepfel, Thomas Lucas, Vincent Leroy, Yohann Cabon, Vaibhav Arora, Romain Brégier, Gabriela Csurka, Leonid Antsfeld, Boris Chidlovskii, and Jérôme Revaud. Croco v2: Improved cross-view completion pretraining for stereo matching and optical flow. In ICCV, pages 17969-17980, 2023. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.376, + 0.482, + 0.417 + ], + "angle": 0, + "content": "[62] Bowen Wen, Wei Yang, Jan Kautz, and Stan Birchfield. Foundationpose: Unified 6d pose estimation and tracking of novel objects. In CVPR, pages 17868-17879, 2024. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.419, + 0.482, + 0.473 + ], + "angle": 0, + "content": "[63] Yu Xiang, Tanner Schmidt, Venkatraman Narayanan, and Dieter Fox. Posecnn: A convolutional neural network for 6d object pose estimation in cluttered scenes. In RSS, 2018. 1, 7" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.473 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.956 + ], + "angle": 0, + "content": "11632" + } + ] +] \ No newline at end of file diff --git a/2025/Co-op_ Correspondence-based Novel Object Pose Estimation/5a137c7c-5457-4cd6-bb8c-541c7e90dee7_origin.pdf b/2025/Co-op_ Correspondence-based Novel Object Pose Estimation/5a137c7c-5457-4cd6-bb8c-541c7e90dee7_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..3a897ed09d330fbb494d6c0de6cce80343c5a332 --- /dev/null +++ b/2025/Co-op_ Correspondence-based Novel Object Pose Estimation/5a137c7c-5457-4cd6-bb8c-541c7e90dee7_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:263bd4f997e01b99337a1b9ef1a561e68e9fc6601501916203c7c68f7b887cf2 +size 9603762 diff --git a/2025/Co-op_ Correspondence-based Novel Object Pose Estimation/full.md b/2025/Co-op_ Correspondence-based Novel Object Pose Estimation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..19641678767df5ee5dc8ef20950960f9ae59b0bf --- /dev/null +++ b/2025/Co-op_ Correspondence-based Novel Object Pose Estimation/full.md @@ -0,0 +1,335 @@ +# Co-op: Correspondence-based Novel Object Pose Estimation + +Sunghill Moon Hyeontae Son Dongcheol Hur Sangwook Kim Naver LABS + +{sungphill.moon, son.ht, dongcheol.hur, o.s.w.a.l.k}@naverlabs.com + +# Abstract + +We propose Co-op, a novel method for accurately and robustly estimating the 6DoF pose of objects unseen during training from a single RGB image. Our method requires only the CAD model of the target object and can precisely estimate its pose without any additional fine-tuning. While existing model-based methods suffer from inefficiency due to using a large number of templates, our method enables fast and accurate estimation with a small number of templates. This improvement is achieved by finding semidense correspondences between the input image and the pre-rendered templates. Our method achieves strong generalization performance by leveraging a hybrid representation that combines patch-level classification and offset regression. Additionally, our pose refinement model estimates probabilistic flow between the input image and the rendered image, refining the initial estimate to an accurate pose using a differentiable PnP layer. We demonstrate that our method not only estimates object poses rapidly but also outperforms existing methods by a large margin on the seven core datasets of the BOP Challenge, achieving state-of-the-art accuracy. + +# 1. Introduction + +Over the past decade, significant improvements in the accuracy of 6D object pose estimation have been achieved through advancements in deep learning methods [33, 35, 46, 47, 54, 58, 63]. However, despite their remarkable accuracy, these methods are challenging to apply in diverse, real-world environments because handling new objects requires acquiring large amounts of data and retraining the models for several hours or even days. + +To address these challenges, recent studies have explored various methods. One approach is the category-level methods [33, 59], which estimate the pose of unseen objects within known categories but struggle with entirely new ones. Another approach aims for generalization by training on large-scale pose estimation datasets and is divided into two types: model-free and model-based methods. However, + +![](images/9649e065d52e5f42b7de043e7d4c0d9765a89bc3bb28df18c1d665105552c145.jpg) +Query +Figure 1. Examples of 6D pose estimation of novel objects. Our method estimates semi-dense or dense correspondences between the input image and rendered images and uses them to estimate the pose. + +![](images/8fe44d0dfc52bd9cd8631a13969fe849a884dc2c73df306e28a366f53dbb2907.jpg) +Semi-dense Correspondences Coarse Pose + +![](images/a98177e5321631133ffe9f14ac0f0e3bafaf8feb182d880b06fe245be6654863.jpg) +Initialization Dense Flow Refined Pose + +model-free methods [18, 29, 36, 53] often face challenges with low-texture object and report results only on limited datasets or those without occlusions. + +Given these limitations, model-based unseen object pose estimation methods [27, 38, 41, 44, 62] that utilize 3D CAD models have been gaining attention. These model-based methods typically consist of three processes—object detection, coarse pose estimation, and pose refinement—or sometimes include a fourth process, pose selection. Since previous works like CNOS [40] and SAM-6D [32] have already addressed object detection with excellent generalization performance, we concentrate our research on the subsequent pose estimation process. + +In this paper, we propose Co-op (Correspondence-based Novel Object Pose Estimation), a novel method for accurately and robustly estimating the 6DoF pose of objects unseen during training from a single RGB image. Our research begins with the consideration of how to effectively define and find correspondences between two images to achieve object pose estimation with excellent generalization performance. + +By setting object pose estimation as finding correspondences between two images, our model is trained at the + +patch or pixel level rather than the image level. This enables the model to learn low-level information, making it robust to domain shifts. Furthermore, by directly leveraging the geometric and structural information of the 3D model, it enables more accurate pose estimation. In our research, to achieve superior generalization performance, we estimate poses based on correspondences in both stages of pose estimation: coarse estimation and pose refinement. + +In the first stage, coarse estimation, we estimate the initial pose of the object by estimating semi-dense correspondences between multiple templates and the input image. In this stage, we use a hybrid representation that combines patch-level classification and offset regression to estimate the semi-dense correspondences. Then, we estimate the pose using the Perspective-n-Point (PnP) algorithm [28] based on these correspondences. By formulating correspondence estimation not as a direct regression problem but by combining it with classification, we achieve outstanding generalization performance. Thanks to this representation, we achieve high accuracy while using a much smaller number of templates compared to other existing coarse estimation methods. + +In the second stage, pose refinement, we refine the initial pose using a render-and-compare approach. To achieve precise refinement, we estimate probabilistic dense correspondences and learn the confidence in pose estimation by training end-to-end through a differentiable PnP layer. This allows us to robustly estimate poses even in challenging situations such as occlusions and texture-less objects. + +Our contributions can be summarized as follows: + +- We present Co-op, a novel framework for unseen object pose estimation in RGB-only cases. Co-op does not require additional training or fine-tuning for new objects and outperforms existing methods by a large margin on the seven core datasets of the BOP Challenge. +- We propose a method for fast and accurate coarse pose estimation using a hybrid representation that combines patch-level classification and offset regression. +- Additionally, we propose a precise object pose refinement method. It estimates dense correspondences defined by probabilistic flow and learns confidence end-to-end through a differentiable PnP layer. + +# 2. Related Work + +Correspondence estimation Traditionally, finding correspondences between two images has been handled using detector-based methods involving three steps: keypoint detection, feature description, and feature matching. These methods have been extensively researched, transitioning from handcrafted approaches [2, 4, 37, 50] to deep learning-based approaches [9, 13, 49, 57]. Recently, detector-free methods [6, 52, 60] have integrated feature detection and matching into a unified process, processing entire images + +without predefined keypoints and improving matching performance even in low-texture or repetitive areas. Furthermore, dense matching methods [14, 15, 55, 56] find matches for every pixel, enhancing accuracy in tasks such as image-based localization, 3D reconstruction, and pose estimation. Our method follows the detector-free approach in coarse estimation and pose refinement. In the coarse estimation stage, we estimate semi-dense correspondences to achieve textureless-resilient and accurate matching, followed by the pose refinement stage, where the model estimates dense correspondences and their confidence to obtain an exact 6D pose. + +Seen Object Pose Estimation Deep learning-based methods have shown remarkable performance improvements with the growth of high-quality datasets and advances in GPU computing power. Some of these methods regress sparse [25, 46, 47, 54] or dense [17, 31, 45] features of the object, such as keypoints or correspondences and separately recover the object's pose by applying RANSAC [16] and PnP [28] on the regressed features. To achieve higher accuracy, other methods use render-and-compare based refinement models [23, 26, 30, 34]. Because these methods are designed to encode specific object's features into the network weights, they tend to perform poorly when encountering unseen objects. + +Unseen Object Pose Estimation To tackle problems in more practical situations, some studies assume that objects are unknown in advance. In particular, there has been active research on model-based unseen object pose estimation. + +Some methods aim to solve the problem using template matching approaches. These methods [27, 38, 39, 42, 51, 62] render the 3D model to generate templates and estimate poses by selecting the most similar template to the query image. Other methods use feature matching approaches, These methods [1, 32, 39, 41, 44] find 2D-3D or 3D-3D correspondences between the 3D model and the query image, recovering poses using the PnP [28] algorithm. To enhance accuracy, several methods [27, 38, 62] apply render-and compare pose refinement, iteratively minimizing discrepancies between the rendered model and the query image. Combined approaches like GenFlow [38] and MegaPose [27] integrate template matching with refinement, achieving remarkable performance but incurring significant computational costs due to the need for numerous template comparisons. + +To address this issue, FoundPose [44] and GigaPose [41] shift from exhaustive template comparison to finding correspondences between the template and the query image, enabling faster and more robust initial pose estimation. These methods depend on the segmentation masks from CNOS [40], to separate the object regions, and perform feature matching between the query image and the template based on DINOv2 [43]. These methods can be + +![](images/15495b07cbb318c090bb642b70f594120fb290197fa489f70fac85ac2cfe50c1.jpg) + +![](images/c8a1f008e442db6c8672505aa31dd42c622b51d53b0663b83de0d9b24ef78c5d.jpg) +Figure 2. Overview. We estimate object pose through two main stages. In the Coarse Pose Estimation stage (Sec 3.1), we estimate semi-dense correspondences between the query image and templates and compute the initial pose using PnP. In the Pose Refinement stage (Sec 3.2), we refine the initial pose by estimating dense flow between the query and rendered images. Both stages utilize transformer encoders and decoders with identical structures, with the Pose Refinement stage additionally incorporating a DPT module after the decoder for dense prediction. + +considered a detect-and-describe framework using the segmentation mask as a feature detector and the DINOv2 patch descriptor as a feature descriptor. + +On the other hand, we propose a detector-free correspondence estimation framework for the coarse estimation stage. By not relying on segmentation, our method is robust to the noisy or partially detected regions and improves the pose accuracy. + +# 3. Method + +In this section, we describe Co-op, our proposed method for unseen object pose estimation. It consists of two main components: coarse pose estimation and pose refinement. Additionally, it optionally incorporates pose selection to estimate the pose more accurately. An overview of our method is presented in Fig. 2. + +# 3.1. Coarse Pose Estimation + +Template Generation In the coarse estimation stage, we use a set of pre-rendered templates from the 3D model. Similar to GigaPose [41], we minimize the number of templates required for pose estimation by generating those that include only out-of-plane rotations. + +Specifically, we render the CAD model from 42 viewpoints. These viewpoints are defined by subdividing each triangle of Blender's icosphere primitive into four smaller triangles, resulting in a uniform distribution of viewpoints over the sphere, following the method used in previous works [39, 40]. + +**Hybrid Representation** The role of the coarse estimation model is to estimate semi-dense correspondences between the query image and the pre-rendered template, denoted by $\mathcal{I}_Q, \mathcal{I}_T \in \mathbb{R}^{H \times W \times 3}$ . To enhance the generalization performance of our model, we use a hybrid representation that combines patch-level classification and offset regression. + +![](images/43e1d6b54781f5ef8ae7a42d8a3485f8b1b7b12cbbf0699caf5b9a28d63f4fae.jpg) +Figure 3. Visualization of Our Hybrid Representation. Left: Patch-level classification results; matching patches are highlighted with the same color. Right: Offset regression within template patches to refine correspondences; red arrows represent the estimated offsets. + +Fig. 3 illustrates an example of our hybrid representation. On the left, we show partial results of the patch-level classification; patches highlighted with the same color in the query image (left) and the template (right) indicate matched patches. On the right, we show the results of offset regression within the template patches after classification, demonstrating the estimated offset from the center of the template patch that best aligns with the center of the corresponding query patch. + +As shown in Fig. 2, our model consists of a Vision Transformer (ViT) encoder for extracting image features and a transformer decoder for interpreting these features. Here, the subscript $i,j$ represents the 2D location in the feature map. Because downsizing is performed by the encoder, each location $i,j$ in the feature map corresponds to a $16\times 16$ patch in the input image. + +The encoder takes $\mathcal{I}_Q$ and $\mathcal{I}_T$ as inputs and extracts the feature maps $\mathcal{F}_Q, \mathcal{F}_T \in \mathbb{R}^{\frac{H}{16} \times \frac{W}{16} \times 1024}$ , respectively. The decoder and following heads then processes $\mathcal{F}_Q$ and $\mathcal{F}_T$ to compute both a tensor for patch-level classification $\mathcal{C} \in \mathbb{R}^{\frac{H}{16} \times \frac{W}{16} \times K}$ and the xy-offsets $\mathcal{U} \in \mathbb{R}^{\frac{H}{16} \times \frac{W}{16} \times 2}$ . + +Here, $K = \frac{H}{16} \times \frac{W}{16} + 1$ represents the number of classes for the patch-level classification. The first $\frac{H}{16} \times \frac{W}{16}$ classes correspond to the template patches, while the last class represents the case where there is no matching part, such as occlusion. For each position $(i,j)$ in the feature map, $\mathcal{C}_{i,j} \in \mathbb{R}^K$ contains classification scores indicating the degree of matchability between the $\frac{H}{16} \times \frac{W}{16}$ template patches and the query patch at that position. + +Following this, we define the patch index as $c_{i,j} = \arg \max_{k\in 1,2,\dots,K}\mathcal{C}_{i,j}^{k}$ and the offset $\mathcal{U}_{i,j}$ has a range of $[-0.5,0.5]$ . At location $(i,j)$ where $c_{i,j} \neq K$ , the position $M_{i,j}^{T}$ in the template corresponding to the query patch is given by: + +$$ +\mathcal {M} _ {i, j} ^ {T} = (\left[ \begin{array}{c} c _ {i, j} \bmod 1 6 + 0. 5 \\ \lfloor c _ {i, j} / 1 6 \rfloor + 0. 5 \end{array} \right] + \mathcal {U} _ {i, j}) \times 1 6. \qquad (1) +$$ + +Adding $\mathcal{U}_{i,j}$ refines the position within the feature map grid. Multiplying by 16 maps this position back to the coordinate system of the original images $\mathcal{I}_Q$ and $\mathcal{I}_T$ , since each location in the feature map corresponds to a $16\times 16$ patch in the original images. The corresponding query image position $\mathcal{M}_{i,j}^{Q}$ is $((i + 0.5)\times 16,(j + 0.5)\times 16)$ , which is the center position of each patch in $\mathcal{I}_Q$ . The correspondence between the $\mathcal{I}_Q$ and $\mathcal{I}_T$ is then defined as $\mathcal{M}_{i,j} = (\mathcal{M}_{i,j}^{Q},\mathcal{M}_{i,j}^{T})$ . + +Pose Fitting To estimate the pose in practice, we compute the correspondences $\mathcal{M}$ for $\mathcal{I}_Q$ and all $\mathcal{I}_T$ , and select the template $k$ most similar to the query image. The similarity score $S_{t}$ for each template $t$ is defined as follows: + +$$ +S _ {t} = \sum_ {i, j} \left\{ \begin{array}{l l} \max \left(\mathcal {C} _ {i, j}\right) & \text {i f} c _ {i, j} \neq K \\ 0 & \text {o t h e r w i s e} \end{array} . \right. \tag {2} +$$ + +If the predicted class $c_{i,j} = K$ (representing occlusion or unmatched patches), we exclude $\max(\mathcal{C}_{i,j})$ from the sum. By computing $S_t$ for each template, we select the template with the highest similarity score as the best match to $\mathcal{I}_Q$ . Once the most similar template is determined, we construct 2D-3D correspondences using the correspondences between the query image and the selected template along with the depth information from the template. We estimate the initial pose by applying RANSAC [16] with the EPnP algorithm [28]. + +# 3.2. Pose Refinement + +In the refinement stage, we accurately refine the pose estimated from the coarse estimation at the pixel-level. To achieve this, we estimate pixel-wise dense correspondences. As shown in Fig. 2, the structure of the refinement model is the same as the coarse estimation model with an added Dense Prediction Transformer (DPT) [48]. The DPT enables pixel-wise prediction, allowing for precise pose refinement. Additionally, by iteratively refining the pose using a render-and-compare approach, this stage overcomes limitations that arise from using pre-rendered templates in the coarse estimation stage, such as self-occlusion. + +Probabilistic Flow Regression Similar to previous works [23, 38], our refinement model estimates the flow between the query image $\mathcal{I}_Q$ and the rendered image $\mathcal{I}_R$ to refine the pose. To recover the pose accurately from the flow, it is essential to prevent inaccurate flows from significantly influencing the pose calculation. To this end, PFA [23] relies on RANSAC [16] to probabilistically exclude inaccurate flows from pose estimation. However, because RANSAC is sensitive to outlier distribution, it is insufficient when outliers + +are prevalent or unevenly distributed. GenFlow [38] focuses on reducing the influence of inaccurate flows by learning to estimate a confidence map through an end-to-end differentiable system and 6D pose loss. Although they show accurate and robust pose estimation performance, there is still room for improvement since learning confidence does not enhance the accuracy of the flow prediction itself. + +Unlike previous flow-based refinement methods [23, 38], we aim to learn the conditional probability of the flow. In line with several correspondence estimation studies [6, 56], we define this conditional probability as $p(Y|\mathcal{I}_Q,\mathcal{I}_R;\theta)$ , where $Y$ represents the flow between the image pair $(\mathcal{I}_Q,\mathcal{I}_R)$ , and $\theta$ denotes the network parameters. Many methods achieve this by learning the variance of the prediction $Y$ , and the predictive density is modeled using Gaussian or Laplace distributions. We model it as a univariate Laplace distribution to simplify the problem. Specifically, $p(Y|\mathcal{I}_Q,\mathcal{I}_R;\theta)$ is modeled as a Laplace distribution with mean $\mu \in \mathbb{R}^{H\times W\times 2}$ and scale $b\in \mathbb{R}^{H\times W\times 1}$ , both predicted by the network. Formulating flow estimation as a probabilistic regression makes our model concentrate on highly reliable correspondences by adapting the scale parameter $b$ . + +Flow Confidence To compute the pose using the flow estimated by the model, we need a confidence $\mathcal{W} \in \mathbb{R}^{H \times W \times 1}$ . This value of $\mathcal{W}$ determines how much weight each flow error will have in calculating the pose in the differentiable PnP layer. $\mathcal{W}$ is calculated from the certainty, sensitivity, and flow probability, which are learned through different loss functions. Certainty estimates whether the flow from $\mathcal{I}_R$ to $\mathcal{I}_Q$ is occluded or not. Sensitivity is learned from the pose loss and highlights areas with rich textures or extremities of the object. Similar to PDC-Net+ [56], we define the flow probability $P_R$ , which represents the probability that the true flow $y$ is within a radius $R$ from the estimated mean flow $\mu$ . This can be calculated as follows: + +$$ +P _ {R} = P \left(\| y - \mu \| _ {1} < R\right) = 1 - \exp \left(- \frac {R}{b}\right). \tag {3} +$$ + +$P_{R}$ is an interpretable measure of reliability, representing the accuracy of the flow with a threshold $R$ . The flow confidence $\mathcal{W}$ is calculated as the element-wise multiplication of certainty, sensitivity, and $P_{R}$ . This means that $\mathcal{W}$ will have higher values when there is no occlusion (high certainty), discriminative information is available to solve the pose (high sensitivity), and the flow is accurate (high $P_{R}$ ). + +Pose Update To compute the refined 6D pose $\mathbf{P}_{\mathrm{refined}}$ using the flow $Y$ and flow confidence $\mathcal{W}$ , we employ a Levenberg-Marquardt (LM) algorithm-based PnP solver. Given the input pose $\mathbf{P}_{\mathrm{input}} = [\mathbf{R}_{\mathrm{input}} \mid \mathbf{t}_{\mathrm{input}}]$ , the camera intrinsic matrix $\mathbf{K}$ , and the depth map $\mathcal{D}_R$ corresponding to $\mathcal{I}_R$ , we + +![](images/085c02889f82c2b52104fa7b864cda011a2ea9376379532d59c0c37cd7414e74.jpg) +Figure 4. Pose Selection. To achieve more precise pose estimation using a multiple hypothesis strategy, we introduce a pose selection stage (Sec 3.3). + +calculate the 3D spatial coordinates $\mathbf{x}_{u,v}^{3\mathrm{D}}$ corresponding to $\mathcal{I}_R$ as follows: + +$$ +\mathbf {x} _ {u, v} ^ {\mathrm {3 D}} = \mathbf {R} _ {\text {i n p u t}} ^ {- 1} \left(\mathbf {K} ^ {- 1} \mathcal {D} _ {R} (u, v) \mathbf {x} _ {u, v} ^ {\mathrm {2 D}} - \mathbf {t} _ {\text {i n p u t}}\right), \tag {4} +$$ + +where $(u,v)$ are the pixel coordinates in $\mathcal{I}_R$ , $\mathcal{D}_R(u,v)$ is the depth value at each pixel, and $\mathbf{x}_{u,v}^{2\mathrm{D}} = (u,v,1)^T$ . We optimize the 6D pose by minimizing the sum of squared weighted reprojection errors as follows: + +$$ +\begin{array}{l} \underset {\mathrm {R}, \mathrm {t}} {\operatorname {a r g m i n}} \frac {1}{2} \sum_ {u} \sum_ {v} \| \mathcal {W} (u, v) \times (\pi (\mathrm {R x} _ {u, v} ^ {\mathrm {3 D}} + \mathrm {t}) \tag {5} \\ - \left((u, v) ^ {T} + Y (u, v))\right) \| ^ {2}. \\ \end{array} +$$ + +Here, $\pi$ is the reprojection function that maps 3D points in camera coordinates to 2D image points using the intrinsic camera parameters $\mathbf{K}$ . + +Following previous work [7, 38], we update the pose in three iterations using the LM algorithm and further refine it to the final pose using the Gauss-Newton algorithm. + +# 3.3. Pose Selection + +In the coarse estimation stage, the best scoring template may not provide the optimal initial pose for refinement. For example, selecting a template rotated by 180 degrees relative to the ground truth (see Fig. 4) makes refinement challenging. To address this issue, many methods [27, 38, 41, 44, 62] employ a multiple hypothesis strategy: generating $N$ coarse pose estimates, refining each, and selecting the best match by comparing rendered results to the query image. Similar to the coarse estimation model, our pose selection model takes the query image and template as inputs but incorporates a score head to estimate a score for each pose hypothesis. Although this increases inference time, considering multiple refined poses helps avoid cases where a difficult-to-refine template leads to a poor final pose. + +![](images/00e0e8f0a6251e9ce02eeb56aca1c5cb676371e176d8eab60efac44be5741155.jpg) +Query +Figure 5. In-plane Rotation Invariant Matching Example. From left to right: Query image, semi-dense correspondences between the query image and the best scoring template, and the coarse pose recovered using the PnP algorithm. + +![](images/8fe42afed8cf11b18a494bdb2c6a78cc2be0c3ca475d311fabe9801e3e012bbc.jpg) +Correspondences + +![](images/d4cc4ea731ab12b80c8e7ad17e6a22bc5d63b37f1c0ecc4335f4cbc0abce59eb.jpg) +Coarse Pose + +# 3.4. Training + +Datasets To train our three models—the coarse estimator, the pose refiner, and the pose selector—we need RGB-D images with ground-truth 6D object pose annotations. We use the large-scale dataset provided by MegaPose [27]. This dataset consists of synthetic data generated using BlenderProc [8], featuring a diverse set of objects from ShapeNet [5] and Google Scanned Objects (GSO) [11], including comprehensive ground-truth 6D pose annotations and object masks. + +Coarse Model Our coarse estimation model is trained to estimate the correspondence between $\mathcal{I}_Q$ and $\mathcal{I}_T$ . Similar to GigaPose [41], we select templates that have similar out-of-plane rotations but different in-plane rotations compared to the training images cropped around the target object. As shown in Fig. 5, our model is designed to estimate correspondences between the query image and the template that are invariant to in-plane rotations. This invariance means that the templates only need to consider out-of-plane rotations, significantly reducing the number of templates required. Since we have the 3D model of the target object, we can generate these correspondences by projecting the 3D model onto the image planes of both the query and template images. Our model is trained to estimate these generated 2D-2D correspondences. + +Refiner Model Our refinement model is trained similarly to previous works [26, 27, 38] that use the render-and-compass are approach. We generate noisy input poses $\mathbf{P}_{\mathrm{input}}$ by adding zero-mean Gaussian noise to the ground truth poses $\mathbf{P}_{\mathrm{gt}}$ , with translation noise standard deviations of (0.01, 0.01, 0.05) along the x, y, and z axes, and rotation noise + +standard deviations of 15 degrees per axis in Euler angles. The model is trained to predict $\mathbf{P}_{\mathrm{gt}}$ from $\mathbf{P}_{\mathrm{input}}$ . + +Selection Model Our selection model is trained to assess the similarity between $\mathcal{I}_Q$ and $\mathcal{I}_R$ based on $\mathbf{P}_{\mathrm{input}}$ to select the most accurate pose from multiple hypotheses. We train the model using binary cross-entropy loss with six poses per $\mathbf{P}_{\mathrm{gt}}$ : one positive, five negatives. Positives have translation differences within (0.01, 0.01, 0.05) and rotation differences within 5 degrees of $\mathbf{P}_{\mathrm{gt}}$ . By setting small rotation thresholds for positives, we enhance the model's ability to distinguish poses close to the ground truth, improving its discriminative capability. + +# 3.5. Implementation Details + +The encoder and decoder architecture of Co-op are based on CroCo [61], a vision foundation model trained on large-scale datasets for 3D vision tasks. This allows us to significantly leverage the benefits of CroCo pre-training. The coarse model processes input images of resolution $224 \times 224$ , whereas the refinement and selection models process images of resolution $256 \times 256$ . For detailed information about the model configurations, learning rates, and training schedules, please refer to the supplementary material. + +Coarse Model In our coarse model, we define the semidense correspondence as a hybrid representation combining patch-level classification and offset regression. Therefore, the coarse model is trained using two loss functions: the classification loss $\mathcal{L}_{\mathrm{cls}}$ and the regression loss $\mathcal{L}_{\mathrm{reg}}$ . + +When defining the ground-truth match from the patch center $(i,j)$ in the query image to the template as $\mathcal{M}_{gt}^T = (\bar{u},\bar{v})$ , the ground-truth index for $\mathcal{L}_{\mathrm{cls}}$ is defined as $\left\lfloor \frac{\bar{v}}{16}\right\rfloor \times \left(\frac{W}{16}\right) + \left\lfloor \frac{\bar{u}}{16}\right\rfloor$ when a match exists. Here, $W$ is the width of the template image, and dividing by 16 accounts for the downsizing due to the patch size. When there is no match, the ground-truth index is defined as $\left(\frac{H}{16} \times \frac{W}{16}\right) + 1$ , representing an additional class for "no match." Therefore, $\mathcal{L}_{\mathrm{cls}}$ is defined as the cross-entropy loss between the ground-truth index and the model's output patch-level probability vector. + +The offset ground truth $\mathcal{U}_{gt}$ for $\mathcal{L}_{\mathrm{reg}}$ is defined as: + +$$ +\mathcal {U} _ {g t} = \frac {\mathcal {M} _ {g t} ^ {T}}{1 6} - \left\lfloor \frac {\mathcal {M} _ {g t} ^ {T}}{1 6} \right\rfloor - 0. 5, \tag {6} +$$ + +where $\mathcal{M}_{gt}^T$ is the ground-truth match in the template. Subtracting $\left\lfloor \mathcal{M}_{gt}^T /16\right\rfloor$ and 0.5 centers the offset within the patch, ranging from -0.5 to 0.5. Here, $\mathcal{U}$ denotes the predicted offset output by the model. The regression loss $\mathcal{L}_{\mathrm{reg}}$ is then defined as the L1 loss between $\mathcal{U}$ and $\mathcal{U}_{gt}$ : + +$$ +\mathcal {L} _ {\text {r e g}} = \left\| \mathcal {U} - \mathcal {U} _ {g t} \right\| _ {1}. \tag {7} +$$ + +Therefore, the overall loss for training the coarse model, $\mathcal{L}_{\mathrm{coarse}}$ , is defined as: + +$$ +\mathcal {L} _ {\text {c o a r s e}} = \mathcal {L} _ {\text {c l s}} + \alpha \mathcal {L} _ {\text {r e g}}. \tag {8} +$$ + +In the above equation, $\alpha$ is set to 2. + +Refiner Model Our refiner model is trained using flow loss, certainty loss, and pose loss, similar to GenFlow [38]. Since we parameterize the flow output of the network with a Laplace distribution, we train the flow by minimizing the negative log-likelihood. Therefore, the flow loss is defined as follows: + +$$ +\mathcal {L} _ {f l o w} = \sum_ {u} \sum_ {v} \left[ \frac {\left| \mu_ {u , v} - \bar {\mu} _ {u , v} \right|}{b _ {u , v}} + 2 \log b _ {u, v} \right]. \tag {9} +$$ + +In this equation, $\mu_{u,v}$ is the flow at pixel position $(u,v)$ within the rendered mask, and $\bar{\mu}_{u,v}$ is the ground-truth flow. $b_{u,v}$ is the scale (uncertainty) at pixel position $(u,v)$ . Additionally, the certainty loss $\mathcal{L}_{\text{cert}}$ is defined as the binary cross-entropy loss that determines whether the flow from the rendered image $\mathcal{I}_R$ to the query image $\mathcal{I}_Q$ falls within the ground-truth mask in $\mathcal{I}_Q$ or not. The pose loss $\mathcal{L}_{\text{pose}}$ which quantifies the difference between the refined pose and the ground-truth pose, is defined as the distance between corresponding 3D points on the 3D model, following the approach used in previous works [27, 38]. Detailed information about $\mathcal{L}_{\text{pose}}$ is provided in the supplementary material. + +The overall loss for refiner model is defined as follows: + +$$ +\mathcal {L} _ {\text {r e f i n e r}} = \mathcal {L} _ {\text {f l o w}} + \beta \mathcal {L} _ {\text {c e r t}} + \gamma \mathcal {L} _ {\text {p o s e}}. \tag {10} +$$ + +The weights $\beta$ and $\gamma$ for each loss are set to 5 and 20, respectively. + +# 4. Experiments + +# 4.1. Experimental Setup + +Datasets We evaluate our method on the seven core datasets of the BOP Challenge [22]: LineMod Occlusion(LM-O) [3], T-LESS [19], TUD-L [20], IC-BIN [10], ITODD [12], HomebrewedDB(HB) [24], and YCB-Video(YCB-V) [63]. Since our method does not require training for each dataset, we did not use any training images; only the 3D object models and test images were used. + +Metrics We follow the protocol of the BOP Challenge [21], which consists of three pose-error functions: Visible Surface Discrepancy (VSD), Maximum Symmetry-Aware Surface Distance (MSSD), and Maximum Symmetry-Aware Projection Distance (MSPD). The overall accuracy is measured by the Average Recall (AR), calculated as: AR = $(\mathrm{AR}_{\mathrm{VSD}} + \mathrm{AR}_{\mathrm{MSSD}} + \mathrm{AR}_{\mathrm{MSPD}}) / 3$ + +# 4.2. BOP Benchmark Results + +Table 1 shows the pose estimation results on the BOP dataset. In our method's experiments, we use a single RTX 3090 Ti GPU. For a fair comparison with previous works [1, 27, 38, 41, 44], we adopt the default CNOS [40] detections provided by the BOP challenge. We also set the number of hypotheses for the multiple hypotheses strategy to 5, consistent with the previous methods. Under these standardized conditions, our method demonstrates improved pose estimation accuracy compared to prior approaches. Although our method is proposed to solve the RGB-only object pose estimation problem, it is also available for RGBD input as a correspondence-based method. The results for this can be found in the supplementary material. + +# 4.2.1 Coarse Estimation + +Rows 1 to 6 show the performance of various methods using only coarse estimation, without any refinement stage. Our method, Co-op (row 6), significantly outperforms previous approaches in this setting. Co-op achieves the highest AR score across all datasets, which is approximately $56.6\%$ higher than the second-best method, FoundPose [44] (row 5). Although GigaPose [41] is the fastest coarse estimation method with an inference time of 0.384 seconds per image (row 4), Co-op's coarse estimation alone achieves a higher mean AR score than GigaPose combined with the refinement model GenFlow [38] (row 11). This demonstrates that our proposed coarse estimation method is highly effective. + +# 4.2.2 Pose Refinement + +Rows 7 to 13 present the results of combining coarse estimation and refinement using a single hypothesis. In this setting, our method, Co-op (row 13), outperforms previous works [27, 38]. To focus on the performance of the refinement model independently of the coarse estimation method, we conducted experiments by combining our refinement model with GigaPose [41] (rows 9, 11, and 12). This allows for a direct comparison of the refinement models using the same coarse estimation. The results show that our refinement model consistently outperforms others, achieving higher AR scores even when paired with the same coarse estimation method. In the multiple hypothesis experiments (rows 14 to 18), our method outperforms the previous best method, FoundPose [44] (row 17), by more than 6 points in AR score. + +# 4.3. Ablation Study + +Table 2 presents various ablation evaluations on seven core datasets, as described in Section 4.1. The results for + +
#Coarse estimationPose refinementLM-OT-LESSTUD-LIC-BINITODDHBYCB-VMeanTime(sec)
Coarse estimation only:
1ZS6D [1]-29.821.0----32.4--
2MegaPose [27]-22.917.725.815.210.825.128.120.815.465
3GenFlow [38]-25.021.530.016.815.428.327.723.53.839
4GigaPose [41]-29.927.330.223.118.834.829.027.60.384
5FoundPose [44]-39.733.846.923.920.450.845.237.31.690
6Co-op-59.759.264.245.839.178.162.658.40.979
With pose refinement (a single hypothesis):
7MegaPose [27]MegaPose [27]49.947.765.336.731.565.460.150.931.724
8GenFlow [38]GenFlow [38]54.751.467.043.738.473.061.955.710.553
9GigaPose [41]MegaPose [27]55.754.158.045.037.669.363.254.72.301
10FoundPose [44]MegaPose [27]55.751.063.343.335.769.766.155.06.385
11GigaPose [41]GenFlow [38]59.555.060.747.841.372.260.856.82.213
12GigaPose [41]Co-op60.256.062.850.243.473.961.958.31.149
13Co-opCo-op64.263.571.751.247.383.267.064.01.852
With pose refinement (5 hypotheses):
14MegaPose [27]MegaPose [27]56.050.768.441.433.870.462.154.747.386
15GenFlow [38]GenFlow [38]56.352.368.445.339.573.963.357.020.890
16GigaPose [41]MegaPose [27]59.957.064.546.739.772.266.357.97.682
17FoundPose [44]MegaPose [27]61.057.069.447.940.772.369.059.620.523
18Co-opCo-op65.564.872.954.249.084.968.965.74.186
+ +Table 1. Results on the BOP challenge datasets. The table shows the Average Recall (AR) score for each dataset, the mean AR score across all datasets, and the time required to estimate the poses of all objects within an image. From top to bottom, the results are ordered as follows: coarse estimation only, refinement applied to a single hypothesis, and refinement results with multiple hypotheses. + +
#MethodLM-OT-LESSTUD-LIC-BINITODDHBYCB-VMean
Coarse Estimation:
1Ours (Proposed)59.759.264.245.839.178.162.658.4
2W/o Classification52.656.144.140.331.071.755.850.2
3W/o CroCo pretrained47.248.946.738.127.061.655.046.4
Pose Refinement:
4Ours (Proposed)64.263.571.751.247.383.267.064.0
5W/o Probabilistic Flow63.663.271.649.546.482.865.863.3
6W/o CroCo pretrained62.361.266.949.042.479.667.161.2
Pose Selection:
7Ours (Proposed)65.564.872.954.249.084.968.965.7
8W/o CroCo pretrained65.164.472.853.147.384.268.365.0
+ +Table 2. Ablation Study. We present the results for each stage under various settings. Rows 1 to 3 correspond to Coarse Estimation, rows 4 to 5 to Pose Refinement, and rows 7 and 8 to Pose Selection. Please refer to Section 4.3 for more details. + +each of our stages are shown in rows 1, 4, 7, respectively. + +Fine-tuning vs. Training from Scratch We assessed the impact of initializing our models with CroCo [61] weights versus random initialization. Initializing with CroCo significantly improved performance in both the coarse estimation (rows 1 vs. 3) and pose refinement stages (rows 4 vs. 6). Although the performance improvement in the pose selection stage (rows 7, 8) is relatively small than others, using CroCo pretraining enhances performance across all stages. + +Hybrid Representation vs. Direct Regression Comparing + +rows 1 and 2, we evaluated the impact of our hybrid representation (combining patch-level classification and offset regression) against direct regression of correspondences. Removing patch-level classification reduced the model's generalization ability, demonstrating the effectiveness of our hybrid approach. + +Probabilistic Flow Regression We evaluated the effectiveness of defining flow as a conditional probability by training a model with the same architecture as Co-op but using the loss function and prediction head from GenFlow [38]. The results (rows 4 vs. 5) show that probabilistic flow regression improves flow learning and pose accuracy by estimating flow reliability, enabling the model to better handle ambiguous or unreliable correspondences. + +# 5. Conclusion + +We propose Co-op, a novel method for 6D pose estimation of unseen objects. Co-op can estimate the poses of novel objects using CAD models that are not provided during training but are available only during testing. We introduce a coarse estimation method that achieves strong generalization performance by leveraging a hybrid representation. Furthermore, we propose a render-and-compare refiner that employs probabilistic flow. Using these components, we demonstrate that our method significantly outperforms existing methods by a large margin on the BOP benchmark, which comprises hundreds of diverse objects and cluttered scenes. + +Acknowledgments. This work was supported by the Technology Innovation Program(00417108, Technologies of Rapid Digital Twin Development Tools for Robotic Service Environments) funded By the Ministry of Trade Industry & Energy(MOTIE, Korea) + +# References + +[1] Philipp Ausserlechner, David Haberger, Stefan Thalhammer, Jean-Baptiste Weibel, and Markus Vincze. Zs6d: Zero-shot 6d object pose estimation using vision transformers. In ICRA, pages 463-469. IEEE, 2024. 2, 7, 8 +[2] Herbert Bay, Tinne Tuytelaars, and Luc Van Gool. Surf: Speeded up robust features. In ECCV, pages 404-417, Berlin, Heidelberg, 2006. Springer Berlin Heidelberg. 2 +[3] Eric Brachmann, Alexander Krull, Frank Michel, Stefan Gumhold, Jamie Shotton, and Carsten Rother. Learning 6d object pose estimation using 3d object coordinates. In ECCV, 2014. 7 +[4] Michael Calonder, Vincent Lepetit, Mustafa Ozuysal, Tomasz Trzcinski, Christoph Strecha, and Pascal Fua. Brief: Computing a local binary descriptor very fast. IEEE Transactions on Pattern Analysis and Machine Intelligence, 34(7): 1281-1298, 2012. 2 +[5] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 6 +[6] Hongkai Chen, Zixin Luo, Lei Zhou, Yurun Tian, Mingmin Zhen, Tian Fang, David McKinnon, Yanghai Tsin, and Long Quan. Aspanformer: Detector-free image matching with adaptive span transformer. In ECCV, 2022. 2, 5 +[7] Hansheng Chen, Pichao Wang, Fan Wang, Wei Tian, Lu Xiong, and Hao Li. Epro-pnp: Generalized end-to-end probabilistic perspective-n-points for monocular object pose estimation. In CVPR, pages 2781-2790, 2022. 5 +[8] Maximilian Denninger, Dominik Winkelbauer, Martin Sundermeyer, Wout Boerdijk, Markus Knauer, Klaus H. Strobl, Matthias Humt, and Rudolph Triebel. Blenderproc2: A procedural pipeline for photorealistic rendering. Journal of Open Source Software, 8(82):4901, 2023. 6 +[9] Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superpoint: Self-supervised interest point detection and description. In CVPRW, pages 337-33712, 2018. 2 +[10] Andreas Doumanoglou, Rigas Kouskouridas, Sotiris Malassiotis, and Tae-Kyun Kim. Recovering 6d object pose and predicting next-best-view in the crowd. In CVPR, pages 3583-3592, 2016. 7 +[11] Laura Downs, Anthony Francis, Nate Koenig, Brandon Kinman, Ryan Hickman, Krista Reymann, Thomas B McHugh, and Vincent Vanhoucke. Google scanned objects: A high-quality dataset of 3d scanned household items. In ICRA, pages 2553-2560. IEEE, 2022. 6 +[12] Bertram Drost, Markus Ulrich, Paul Bergmann, Philipp Hartinger, and Carsten Steger. Introducing mvtec itodd — a dataset for 3d object recognition in industry. In ICCVW, pages 2200-2208, 2017. 7 + +[13] Mihai Dusmanu, Ignacio Rocco, Tomas Pajdla, Marc Pollefeys, Josef Sivic, Akihiko Torii, and Torsten Sattler. D2-Net: A Trainable CNN for Joint Detection and Description of Local Features. In CVPR, 2019. 2 +[14] Johan Edstedt, Ioannis Athanasiadis, Mårten Wadenback, and Michael Felsberg. DKM: Dense kernelized feature matching for geometry estimation. In CVPR, 2023. 2 +[15] Johan Edstedt, Qiyu Sun, Georg Bokman, Marten Wadenback, and Michael Felsberg. RoMa: Robust Dense Feature Matching. In CVPR, 2024. 2 +[16] Martin A. Fischler and Robert C. Bolles. Random sample consensus: A paradigm for model fitting with applications to image analysis and automated cartography. Commun. ACM, 24(6):381-395, 1981. 2, 4 +[17] Rasmus Laurvig Haugaard and Anders Glent Buch. Surfemb: Dense and continuous correspondence distributions for object pose estimation with learnt surface embeddings. In CVPR, pages 6749-6758, 2022. 2 +[18] Xingyi He, Jiaming Sun, Yuang Wang, Di Huang, Hujun Bao, and Xiaowei Zhou. Onepose++: Keypoint-free one-shot object pose estimation without CAD models. In NeurIPS, 2022. 1 +[19] Tomás Hodań, Pavel Haluza, Štepan Obdržálek, Jiří Matas, Manolis Lourakis, and Xenophon Zabulis. T-LESS: An RGB-D dataset for 6D pose estimation of texture-less objects. WACV, 2017. 7 +[20] Tomás Hodań, Frank Michel, Eric Brachmann, Wadim Kehl, Anders Glent Buch, Dirk Kraft, Bertram Drost, Joel Vidal, Stephan Ihrke, Xenophon Zabulis, Caner Sahin, Fabian Manhardt, Federico Tombari, Tae-Kyun Kim, Jiri Matas, and Carsten Rother. BOP: Benchmark for 6D object pose estimation. ECCV, 2018. 7 +[21] Tomáš Hodań, Martin Sundermeyer, Bertram Drost, Yann Labbé, Eric Brachmann, Frank Michel, Carsten Rother, and Jiri Matas. Bop challenge 2020 on 6d object localization. In Computer Vision-ECCV 2020 Workshops: Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16, pages 577-594. Springer, 2020. 7 +[22] Tomas Hodan, Martin Sundermeyer, Yann Labbe, Van Nguyen Nguyen, Gu Wang, Eric Brachmann, Bertram Drost, Vincent Lepetit, Carsten Rother, and Jiri Matas. Bop challenge 2023 on detection segmentation and pose estimation of seen and unseen rigid objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5610-5619, 2024. 7 +[23] Yinlin Hu, Pascal Fua, and Mathieu Salzmann. Perspective flow aggregation for data-limited 6d object pose estimation. In ECCV, pages 89-106. Springer, 2022. 2, 4, 5 +[24] Roman Kaskman, Sergey Zakharov, Ivan Shugurov, and Slobodan Ilic. Homebreweddb: Rgb-d dataset for 6d pose estimation of 3d objects. In ICCVW, pages 2767-2776, 2019. 7 +[25] Wadim Kehl, Fabian Manhardt, Federico Tombari, Slobodan Ilic, and Nassir Navab. Ssd-6d: Making rgb-based 3d detection and 6d pose estimation great again. In ICCV, pages 1521–1529, 2017. 2 + +[26] Y. Labbe, J. Carpentier, M. Aubry, and J. Sivic. Cosyposse: Consistent multi-view multi-object 6d pose estimation. In ECCV, 2020. 2, 6 +[27] Yann Labbe, Lucas Manuelli, Arsalan Mousavian, Stephen Tyree, Stan Birchfield, Jonathan Tremblay, Justin Carpentier, Mathieu Aubry, Dieter Fox, and Josef Sivic. MegaPose: 6D Pose Estimation of Novel Objects via Render & Compare. In CoRL, 2022. 1, 2, 5, 6, 7, 8 +[28] Vincent Lepetit, Francesc Moreno-Noguer, and Pascal Fua. Epnp: An accurate o(n) solution to the pnp problem. International Journal Of Computer Vision, 81:155-166, 2009. 2, 4 +[29] Fu Li, Shishir Reddy Vutukur, Hao Yu, Ivan Shugurov, Benjamin Busam, Shaowu Yang, and Slobodan Ilic. Nerfpose: A first-reconstruct-then-regress approach for weakly-supervised 6d object pose estimation. In ICCV, pages 2123–2133, 2023. 1 +[30] Yi Li, Gu Wang, Xiangyang Ji, Yu Xiang, and Dieter Fox. DeepIM: Deep iterative matching for 6D pose estimation. In ECCV, 2018. 2 +[31] Zhigang Li, Gu Wang, and Xiangyang Ji. Cdpn: Coordinates-based disentangled pose network for real-time rgb-based 6-dof object pose estimation. In ICCV, pages 7678-7687, 2019. 2 +[32] Jiehong Lin, Lihua Liu, Dekun Lu, and Kui Jia. Sam-6d: Segment anything model meets zero-shot 6d object pose estimation. In CVPR, pages 27906–27916, 2024. 1, 2 +[33] Yunzhi Lin, Jonathan Tremblay, Stephen Tyree, Patricio A. Vela, and Stan Birchfield. Single-stage keypoint-based category-level object pose estimation from an RGB image. In ICRA, 2022. 1 +[34] Lahav Lipson, Zachary Teed, Ankit Goyal, and Jia Deng. Coupled iterative refinement for 6d multi-object pose estimation. In CVPR, pages 6728-6737, 2022. 2 +[35] Xingyu Liu, Ruida Zhang, Chenyangguang Zhang, Bowen Fu, Jiwen Tang, Xiquan Liang, Jingyi Tang, Xiaotian Cheng, Yukang Zhang, Gu Wang, and Xiangyang Ji. Gdrnpp. https://github.com/shanice-1/gdrnpp_bop2022, 2022.1 +[36] Yuan Liu, Yilin Wen, Sida Peng, Cheng Lin, Xiaoxiao Long, Taku Komura, and Wenping Wang. Gen6d: Generalizable model-free 6-dof object pose estimation from rgb images. In ECCV, 2022. 1 +[37] David G. Lowe. Distinctive image features from scale-invariant keypoints. International Journal of Computer Vision, 60:91-110, 2004. 2 +[38] Sungphill Moon, Hyeontae Son, Dongcheol Hur, and Sangwook Kim. Genflow: Generalizable recurrent flow for 6d pose refinement of novel objects. In CVPR, pages 10039-10049, 2024. 1, 2, 4, 5, 6, 7, 8 +[39] Van Nguyen Nguyen, Yinlin Hu, Yang Xiao, Mathieu Salzmann, and Vincent Lepetit. Templates for 3d object pose estimation revisited: Generalization to new objects and robustness to occlusions. In CVPR, pages 6771-6780, 2022. 2, 3 +[40] Van Nguyen Nguyen, Thibault Groueix, Georgy Ponimatkin, Vincent Lepetit, and Tomas Hodan. Cnos: A strong baseline + +for cad-based novel object segmentation. In ICCVW, pages 2134-2140, 2023. 1, 2, 3, 7 +[41] Van Nguyen Nguyen, Thibault Groueix, Mathieu Salzmann, and Vincent Lepetit. Gigapose: Fast and robust novel object pose estimation via one correspondence. In CVPR, pages 9903-9913, 2024. 1, 2, 3, 5, 6, 7, 8 +[42] Brian Okorn, Qiao Gu, Martial Hebert, and David Held. Zephyr: Zero-shot pose hypothesis rating. In ICRA, pages 14141-14148. IEEE, 2021. 2 +[43] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 2 +[44] Evin Pinar Örnek, Yann Labbe, Bugra Tekin, Lingni Ma, Cem Keskin, Christian Forster, and Tomáš Hodač. Found- pose: Unseen object pose estimation with foundation features. In ECCV, 2024. 1, 2, 5, 7, 8 +[45] Kiru Park, Timothy Patten, and Markus Vincze. Pix2pose: Pixel-wise coordinate regression of objects for 6d pose estimation. In ICCV, pages 7668-7677, 2019. 2 +[46] Sida Peng, Yuan Liu, Qixing Huang, Xiaowei Zhou, and Hujun Bao. Pvnet: Pixel-wise voting network for 6dof pose estimation. In CVPR, 2019. 1, 2 +[47] Mahdi Rad and Vincent Lepetit. Bb8: A scalable, accurate, robust to partial occlusion method for predicting the 3d poses of challenging objects without using depth. In ICCV, 2017. 1, 2 +[48] René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. In ICCV, pages 12179-12188, 2021. 4 +[49] Jerome Revaud, Philippe Weinzaepfel, César Roberto de Souza, and Martin Humenberger. R2D2: repeatable and reliable detector and descriptor. In NeurIPS, 2019. 2 +[50] Ethan Rublee, Vincent Rabaud, Kurt Konolige, and Gary Bradski. Orb: An efficient alternative to sift or surf. In ICCV, pages 2564-2571, 2011. 2 +[51] Ivan Shugurov, Fu Li, Benjamin Busam, and Slobodan Ilic. Osop: A multi-stage one shot object pose estimation framework. In CVPR, pages 6835-6844, 2022. 2 +[52] Jiaming Sun, Zehong Shen, Yuang Wang, Hujun Bao, and Xiaowei Zhou. LoFTR: Detector-free local feature matching with transformers. In CVPR, 2021. 2 +[53] Jiaming Sun, Zihao Wang, Siyu Zhang, Xingyi He, Hongcheng Zhao, Guofeng Zhang, and Xiaowei Zhou. OnePose: One-shot object pose estimation without CAD models. In CVPR, 2022. 1 +[54] Bugra Tekin, Sudipta N. Sinha, and Pascal Fua. Real-Time Seamless Single Shot 6D Object Pose Prediction. In CVPR, 2018. 1, 2 +[55] Prune Truong, Martin Danelljan, and Radu Timofte. Glunet: Global-local universal network for dense flow and correspondences. In CVPR, pages 6258-6268, 2020. 2 +[56] Prune Truong, Martin Danelljan, Radu Timofte, and Luc Van Gool. Pdc-net+: Enhanced probabilistic dense correspondence network. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(8):10247-10266, 2023. 2, 5 + +[57] Michal Tyszkiiewicz, Pascal Fua, and Eduard Trulls. Disk: Learning local features with policy gradient. In NeurIPS, 2020. 2 +[58] Gu Wang, Fabian Manhardt, Federico Tombari, and Xi-angyang Ji. GDR-Net: Geometry-guided direct regression network for monocular 6d object pose estimation. In CVPR, pages 16611-16621, 2021. 1 +[59] He Wang, Srinath Sridhar, Jingwei Huang, Julien Valentin, Shuran Song, and Leonidas J. Guibas. Normalized object coordinate space for category-level 6d object pose and size estimation. In CVPR, 2019. 1 +[60] Yifan Wang, Xingyi He, Sida Peng, Dongli Tan, and Xiaowei Zhou. Efficient LoFTR: Semi-dense local feature matching with sparse-like speed. In CVPR, 2024. 2 +[61] Philippe Weinzaepfel, Thomas Lucas, Vincent Leroy, Yohann Cabon, Vaibhav Arora, Romain Brégier, Gabriela Csurka, Leonid Antsfeld, Boris Chidlovskii, and Jérôme Revaud. Croco v2: Improved cross-view completion pretraining for stereo matching and optical flow. In ICCV, pages 17969-17980, 2023. 6, 8 +[62] Bowen Wen, Wei Yang, Jan Kautz, and Stan Birchfield. Foundationpose: Unified 6d pose estimation and tracking of novel objects. In CVPR, pages 17868-17879, 2024. 1, 2, 5 +[63] Yu Xiang, Tanner Schmidt, Venkatraman Narayanan, and Dieter Fox. Posecnn: A convolutional neural network for 6d object pose estimation in cluttered scenes. In RSS, 2018. 1, 7 \ No newline at end of file diff --git a/2025/Co-op_ Correspondence-based Novel Object Pose Estimation/images.zip b/2025/Co-op_ Correspondence-based Novel Object Pose Estimation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..0c40081f2734ca8daed48e55da0455b3e883faa2 --- /dev/null +++ b/2025/Co-op_ Correspondence-based Novel Object Pose Estimation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:141a48d5253aa9f62172720f3b97c43a0326416bfc0dcf0bdb3b8fdb37128b86 +size 553128 diff --git a/2025/Co-op_ Correspondence-based Novel Object Pose Estimation/layout.json b/2025/Co-op_ Correspondence-based Novel Object Pose Estimation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..a11c9eeb786d6ebfdbb824382eb7cb8f56347e51 --- /dev/null +++ b/2025/Co-op_ Correspondence-based Novel Object Pose Estimation/layout.json @@ -0,0 +1,9798 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 117, + 103, + 493, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 117, + 103, + 493, + 121 + ], + "spans": [ + { + "bbox": [ + 117, + 103, + 493, + 121 + ], + "type": "text", + "content": "Co-op: Correspondence-based Novel Object Pose Estimation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 116, + 143, + 494, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 143, + 494, + 171 + ], + "spans": [ + { + "bbox": [ + 116, + 143, + 494, + 171 + ], + "type": "text", + "content": "Sunghill Moon Hyeontae Son Dongcheol Hur Sangwook Kim Naver LABS" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 125, + 173, + 482, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 125, + 173, + 482, + 185 + ], + "spans": [ + { + "bbox": [ + 125, + 173, + 482, + 185 + ], + "type": "text", + "content": "{sungphill.moon, son.ht, dongcheol.hur, o.s.w.a.l.k}@naverlabs.com" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 213, + 200, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 213, + 200, + 225 + ], + "spans": [ + { + "bbox": [ + 151, + 213, + 200, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 239, + 296, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 239, + 296, + 491 + ], + "spans": [ + { + "bbox": [ + 55, + 239, + 296, + 491 + ], + "type": "text", + "content": "We propose Co-op, a novel method for accurately and robustly estimating the 6DoF pose of objects unseen during training from a single RGB image. Our method requires only the CAD model of the target object and can precisely estimate its pose without any additional fine-tuning. While existing model-based methods suffer from inefficiency due to using a large number of templates, our method enables fast and accurate estimation with a small number of templates. This improvement is achieved by finding semidense correspondences between the input image and the pre-rendered templates. Our method achieves strong generalization performance by leveraging a hybrid representation that combines patch-level classification and offset regression. Additionally, our pose refinement model estimates probabilistic flow between the input image and the rendered image, refining the initial estimate to an accurate pose using a differentiable PnP layer. We demonstrate that our method not only estimates object poses rapidly but also outperforms existing methods by a large margin on the seven core datasets of the BOP Challenge, achieving state-of-the-art accuracy." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 514, + 135, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 514, + 135, + 525 + ], + "spans": [ + { + "bbox": [ + 56, + 514, + 135, + 525 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 533, + 295, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 533, + 295, + 629 + ], + "spans": [ + { + "bbox": [ + 55, + 533, + 295, + 629 + ], + "type": "text", + "content": "Over the past decade, significant improvements in the accuracy of 6D object pose estimation have been achieved through advancements in deep learning methods [33, 35, 46, 47, 54, 58, 63]. However, despite their remarkable accuracy, these methods are challenging to apply in diverse, real-world environments because handling new objects requires acquiring large amounts of data and retraining the models for several hours or even days." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "type": "text", + "content": "To address these challenges, recent studies have explored various methods. One approach is the category-level methods [33, 59], which estimate the pose of unseen objects within known categories but struggle with entirely new ones. Another approach aims for generalization by training on large-scale pose estimation datasets and is divided into two types: model-free and model-based methods. However," + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 316, + 214, + 350, + 354 + ], + "blocks": [ + { + "bbox": [ + 316, + 214, + 350, + 354 + ], + "lines": [ + { + "bbox": [ + 316, + 214, + 350, + 354 + ], + "spans": [ + { + "bbox": [ + 316, + 214, + 350, + 354 + ], + "type": "image", + "image_path": "9649e065d52e5f42b7de043e7d4c0d9765a89bc3bb28df18c1d665105552c145.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 325, + 354, + 339, + 361 + ], + "lines": [ + { + "bbox": [ + 325, + 354, + 339, + 361 + ], + "spans": [ + { + "bbox": [ + 325, + 354, + 339, + 361 + ], + "type": "text", + "content": "Query" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 373, + 555, + 417 + ], + "lines": [ + { + "bbox": [ + 313, + 373, + 555, + 417 + ], + "spans": [ + { + "bbox": [ + 313, + 373, + 555, + 417 + ], + "type": "text", + "content": "Figure 1. Examples of 6D pose estimation of novel objects. Our method estimates semi-dense or dense correspondences between the input image and rendered images and uses them to estimate the pose." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 352, + 215, + 451, + 354 + ], + "blocks": [ + { + "bbox": [ + 352, + 215, + 451, + 354 + ], + "lines": [ + { + "bbox": [ + 352, + 215, + 451, + 354 + ], + "spans": [ + { + "bbox": [ + 352, + 215, + 451, + 354 + ], + "type": "image", + "image_path": "8fe44d0dfc52bd9cd8631a13969fe849a884dc2c73df306e28a366f53dbb2907.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 353, + 354, + 449, + 361 + ], + "lines": [ + { + "bbox": [ + 353, + 354, + 449, + 361 + ], + "spans": [ + { + "bbox": [ + 353, + 354, + 449, + 361 + ], + "type": "text", + "content": "Semi-dense Correspondences Coarse Pose" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 455, + 215, + 552, + 354 + ], + "blocks": [ + { + "bbox": [ + 455, + 215, + 552, + 354 + ], + "lines": [ + { + "bbox": [ + 455, + 215, + 552, + 354 + ], + "spans": [ + { + "bbox": [ + 455, + 215, + 552, + 354 + ], + "type": "image", + "image_path": "a98177e5321631133ffe9f14ac0f0e3bafaf8feb182d880b06fe245be6654863.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 457, + 354, + 551, + 361 + ], + "lines": [ + { + "bbox": [ + 457, + 354, + 551, + 361 + ], + "spans": [ + { + "bbox": [ + 457, + 354, + 551, + 361 + ], + "type": "text", + "content": "Initialization Dense Flow Refined Pose" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 437, + 554, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 554, + 473 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 554, + 473 + ], + "type": "text", + "content": "model-free methods [18, 29, 36, 53] often face challenges with low-texture object and report results only on limited datasets or those without occlusions." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 475, + 554, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 475, + 554, + 594 + ], + "spans": [ + { + "bbox": [ + 313, + 475, + 554, + 594 + ], + "type": "text", + "content": "Given these limitations, model-based unseen object pose estimation methods [27, 38, 41, 44, 62] that utilize 3D CAD models have been gaining attention. These model-based methods typically consist of three processes—object detection, coarse pose estimation, and pose refinement—or sometimes include a fourth process, pose selection. Since previous works like CNOS [40] and SAM-6D [32] have already addressed object detection with excellent generalization performance, we concentrate our research on the subsequent pose estimation process." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 594, + 554, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 554, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 554, + 689 + ], + "type": "text", + "content": "In this paper, we propose Co-op (Correspondence-based Novel Object Pose Estimation), a novel method for accurately and robustly estimating the 6DoF pose of objects unseen during training from a single RGB image. Our research begins with the consideration of how to effectively define and find correspondences between two images to achieve object pose estimation with excellent generalization performance." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "content": "By setting object pose estimation as finding correspondences between two images, our model is trained at the" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11622" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "content": "patch or pixel level rather than the image level. This enables the model to learn low-level information, making it robust to domain shifts. Furthermore, by directly leveraging the geometric and structural information of the 3D model, it enables more accurate pose estimation. In our research, to achieve superior generalization performance, we estimate poses based on correspondences in both stages of pose estimation: coarse estimation and pose refinement." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 168, + 294, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 168, + 294, + 334 + ], + "spans": [ + { + "bbox": [ + 55, + 168, + 294, + 334 + ], + "type": "text", + "content": "In the first stage, coarse estimation, we estimate the initial pose of the object by estimating semi-dense correspondences between multiple templates and the input image. In this stage, we use a hybrid representation that combines patch-level classification and offset regression to estimate the semi-dense correspondences. Then, we estimate the pose using the Perspective-n-Point (PnP) algorithm [28] based on these correspondences. By formulating correspondence estimation not as a direct regression problem but by combining it with classification, we achieve outstanding generalization performance. Thanks to this representation, we achieve high accuracy while using a much smaller number of templates compared to other existing coarse estimation methods." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 335, + 294, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 335, + 294, + 418 + ], + "spans": [ + { + "bbox": [ + 55, + 335, + 294, + 418 + ], + "type": "text", + "content": "In the second stage, pose refinement, we refine the initial pose using a render-and-compare approach. To achieve precise refinement, we estimate probabilistic dense correspondences and learn the confidence in pose estimation by training end-to-end through a differentiable PnP layer. This allows us to robustly estimate poses even in challenging situations such as occlusions and texture-less objects." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 419, + 266, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 419, + 266, + 430 + ], + "spans": [ + { + "bbox": [ + 67, + 419, + 266, + 430 + ], + "type": "text", + "content": "Our contributions can be summarized as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 431, + 294, + 574 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 56, + 431, + 294, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 431, + 294, + 490 + ], + "spans": [ + { + "bbox": [ + 56, + 431, + 294, + 490 + ], + "type": "text", + "content": "- We present Co-op, a novel framework for unseen object pose estimation in RGB-only cases. Co-op does not require additional training or fine-tuning for new objects and outperforms existing methods by a large margin on the seven core datasets of the BOP Challenge." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 491, + 294, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 491, + 294, + 525 + ], + "spans": [ + { + "bbox": [ + 56, + 491, + 294, + 525 + ], + "type": "text", + "content": "- We propose a method for fast and accurate coarse pose estimation using a hybrid representation that combines patch-level classification and offset regression." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 527, + 294, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 527, + 294, + 574 + ], + "spans": [ + { + "bbox": [ + 56, + 527, + 294, + 574 + ], + "type": "text", + "content": "- Additionally, we propose a precise object pose refinement method. It estimates dense correspondences defined by probabilistic flow and learns confidence end-to-end through a differentiable PnP layer." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 586, + 142, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 586, + 142, + 597 + ], + "spans": [ + { + "bbox": [ + 55, + 586, + 142, + 597 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 605, + 294, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 294, + 713 + ], + "type": "text", + "content": "Correspondence estimation Traditionally, finding correspondences between two images has been handled using detector-based methods involving three steps: keypoint detection, feature description, and feature matching. These methods have been extensively researched, transitioning from handcrafted approaches [2, 4, 37, 50] to deep learning-based approaches [9, 13, 49, 57]. Recently, detector-free methods [6, 52, 60] have integrated feature detection and matching into a unified process, processing entire images" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 553, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 215 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 215 + ], + "type": "text", + "content": "without predefined keypoints and improving matching performance even in low-texture or repetitive areas. Furthermore, dense matching methods [14, 15, 55, 56] find matches for every pixel, enhancing accuracy in tasks such as image-based localization, 3D reconstruction, and pose estimation. Our method follows the detector-free approach in coarse estimation and pose refinement. In the coarse estimation stage, we estimate semi-dense correspondences to achieve textureless-resilient and accurate matching, followed by the pose refinement stage, where the model estimates dense correspondences and their confidence to obtain an exact 6D pose." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 218, + 553, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 218, + 553, + 373 + ], + "spans": [ + { + "bbox": [ + 313, + 218, + 553, + 373 + ], + "type": "text", + "content": "Seen Object Pose Estimation Deep learning-based methods have shown remarkable performance improvements with the growth of high-quality datasets and advances in GPU computing power. Some of these methods regress sparse [25, 46, 47, 54] or dense [17, 31, 45] features of the object, such as keypoints or correspondences and separately recover the object's pose by applying RANSAC [16] and PnP [28] on the regressed features. To achieve higher accuracy, other methods use render-and-compare based refinement models [23, 26, 30, 34]. Because these methods are designed to encode specific object's features into the network weights, they tend to perform poorly when encountering unseen objects." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 376, + 553, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 376, + 553, + 424 + ], + "spans": [ + { + "bbox": [ + 313, + 376, + 553, + 424 + ], + "type": "text", + "content": "Unseen Object Pose Estimation To tackle problems in more practical situations, some studies assume that objects are unknown in advance. In particular, there has been active research on model-based unseen object pose estimation." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 426, + 553, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 553, + 616 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 553, + 616 + ], + "type": "text", + "content": "Some methods aim to solve the problem using template matching approaches. These methods [27, 38, 39, 42, 51, 62] render the 3D model to generate templates and estimate poses by selecting the most similar template to the query image. Other methods use feature matching approaches, These methods [1, 32, 39, 41, 44] find 2D-3D or 3D-3D correspondences between the 3D model and the query image, recovering poses using the PnP [28] algorithm. To enhance accuracy, several methods [27, 38, 62] apply render-and compare pose refinement, iteratively minimizing discrepancies between the rendered model and the query image. Combined approaches like GenFlow [38] and MegaPose [27] integrate template matching with refinement, achieving remarkable performance but incurring significant computational costs due to the need for numerous template comparisons." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 617, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 553, + 713 + ], + "type": "text", + "content": "To address this issue, FoundPose [44] and GigaPose [41] shift from exhaustive template comparison to finding correspondences between the template and the query image, enabling faster and more robust initial pose estimation. These methods depend on the segmentation masks from CNOS [40], to separate the object regions, and perform feature matching between the query image and the template based on DINOv2 [43]. These methods can be" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11623" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 73, + 553, + 232 + ], + "blocks": [ + { + "bbox": [ + 59, + 73, + 553, + 232 + ], + "lines": [ + { + "bbox": [ + 59, + 73, + 553, + 232 + ], + "spans": [ + { + "bbox": [ + 59, + 73, + 553, + 232 + ], + "type": "image", + "image_path": "15495b07cbb318c090bb642b70f594120fb290197fa489f70fac85ac2cfe50c1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 58, + 239, + 553, + 403 + ], + "blocks": [ + { + "bbox": [ + 58, + 239, + 553, + 403 + ], + "lines": [ + { + "bbox": [ + 58, + 239, + 553, + 403 + ], + "spans": [ + { + "bbox": [ + 58, + 239, + 553, + 403 + ], + "type": "image", + "image_path": "c8a1f008e442db6c8672505aa31dd42c622b51d53b0663b83de0d9b24ef78c5d.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 411, + 555, + 467 + ], + "lines": [ + { + "bbox": [ + 55, + 411, + 555, + 467 + ], + "spans": [ + { + "bbox": [ + 55, + 411, + 555, + 467 + ], + "type": "text", + "content": "Figure 2. Overview. We estimate object pose through two main stages. In the Coarse Pose Estimation stage (Sec 3.1), we estimate semi-dense correspondences between the query image and templates and compute the initial pose using PnP. In the Pose Refinement stage (Sec 3.2), we refine the initial pose by estimating dense flow between the query and rendered images. Both stages utilize transformer encoders and decoders with identical structures, with the Pose Refinement stage additionally incorporating a DPT module after the decoder for dense prediction." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 487, + 295, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 487, + 295, + 523 + ], + "spans": [ + { + "bbox": [ + 55, + 487, + 295, + 523 + ], + "type": "text", + "content": "considered a detect-and-describe framework using the segmentation mask as a feature detector and the DINOv2 patch descriptor as a feature descriptor." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 529, + 296, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 529, + 296, + 590 + ], + "spans": [ + { + "bbox": [ + 55, + 529, + 296, + 590 + ], + "type": "text", + "content": "On the other hand, we propose a detector-free correspondence estimation framework for the coarse estimation stage. By not relying on segmentation, our method is robust to the noisy or partially detected regions and improves the pose accuracy." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 616, + 111, + 628 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 616, + 111, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 616, + 111, + 628 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 642, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 715 + ], + "type": "text", + "content": "In this section, we describe Co-op, our proposed method for unseen object pose estimation. It consists of two main components: coarse pose estimation and pose refinement. Additionally, it optionally incorporates pose selection to estimate the pose more accurately. An overview of our method is presented in Fig. 2." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 487, + 448, + 498 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 487, + 448, + 498 + ], + "spans": [ + { + "bbox": [ + 314, + 487, + 448, + 498 + ], + "type": "text", + "content": "3.1. Coarse Pose Estimation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 506, + 555, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 506, + 555, + 566 + ], + "spans": [ + { + "bbox": [ + 313, + 506, + 555, + 566 + ], + "type": "text", + "content": "Template Generation In the coarse estimation stage, we use a set of pre-rendered templates from the 3D model. Similar to GigaPose [41], we minimize the number of templates required for pose estimation by generating those that include only out-of-plane rotations." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 567, + 554, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 567, + 554, + 639 + ], + "spans": [ + { + "bbox": [ + 313, + 567, + 554, + 639 + ], + "type": "text", + "content": "Specifically, we render the CAD model from 42 viewpoints. These viewpoints are defined by subdividing each triangle of Blender's icosphere primitive into four smaller triangles, resulting in a uniform distribution of viewpoints over the sphere, following the method used in previous works [39, 40]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 642, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 555, + 714 + ], + "type": "text", + "content": "**Hybrid Representation** The role of the coarse estimation model is to estimate semi-dense correspondences between the query image and the pre-rendered template, denoted by " + }, + { + "bbox": [ + 313, + 642, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_Q, \\mathcal{I}_T \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 313, + 642, + 555, + 714 + ], + "type": "text", + "content": ". To enhance the generalization performance of our model, we use a hybrid representation that combines patch-level classification and offset regression." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "11624" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 76, + 289, + 208 + ], + "blocks": [ + { + "bbox": [ + 62, + 76, + 289, + 208 + ], + "lines": [ + { + "bbox": [ + 62, + 76, + 289, + 208 + ], + "spans": [ + { + "bbox": [ + 62, + 76, + 289, + 208 + ], + "type": "image", + "image_path": "43e1d6b54781f5ef8ae7a42d8a3485f8b1b7b12cbbf0699caf5b9a28d63f4fae.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 222, + 295, + 277 + ], + "lines": [ + { + "bbox": [ + 55, + 222, + 295, + 277 + ], + "spans": [ + { + "bbox": [ + 55, + 222, + 295, + 277 + ], + "type": "text", + "content": "Figure 3. Visualization of Our Hybrid Representation. Left: Patch-level classification results; matching patches are highlighted with the same color. Right: Offset regression within template patches to refine correspondences; red arrows represent the estimated offsets." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 300, + 296, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 300, + 296, + 408 + ], + "spans": [ + { + "bbox": [ + 55, + 300, + 296, + 408 + ], + "type": "text", + "content": "Fig. 3 illustrates an example of our hybrid representation. On the left, we show partial results of the patch-level classification; patches highlighted with the same color in the query image (left) and the template (right) indicate matched patches. On the right, we show the results of offset regression within the template patches after classification, demonstrating the estimated offset from the center of the template patch that best aligns with the center of the corresponding query patch." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 409, + 296, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 409, + 296, + 493 + ], + "spans": [ + { + "bbox": [ + 55, + 409, + 296, + 493 + ], + "type": "text", + "content": "As shown in Fig. 2, our model consists of a Vision Transformer (ViT) encoder for extracting image features and a transformer decoder for interpreting these features. Here, the subscript " + }, + { + "bbox": [ + 55, + 409, + 296, + 493 + ], + "type": "inline_equation", + "content": "i,j" + }, + { + "bbox": [ + 55, + 409, + 296, + 493 + ], + "type": "text", + "content": " represents the 2D location in the feature map. Because downsizing is performed by the encoder, each location " + }, + { + "bbox": [ + 55, + 409, + 296, + 493 + ], + "type": "inline_equation", + "content": "i,j" + }, + { + "bbox": [ + 55, + 409, + 296, + 493 + ], + "type": "text", + "content": " in the feature map corresponds to a " + }, + { + "bbox": [ + 55, + 409, + 296, + 493 + ], + "type": "inline_equation", + "content": "16\\times 16" + }, + { + "bbox": [ + 55, + 409, + 296, + 493 + ], + "type": "text", + "content": " patch in the input image." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 494, + 296, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 494, + 296, + 555 + ], + "spans": [ + { + "bbox": [ + 55, + 494, + 296, + 555 + ], + "type": "text", + "content": "The encoder takes " + }, + { + "bbox": [ + 55, + 494, + 296, + 555 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_Q" + }, + { + "bbox": [ + 55, + 494, + 296, + 555 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 494, + 296, + 555 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_T" + }, + { + "bbox": [ + 55, + 494, + 296, + 555 + ], + "type": "text", + "content": " as inputs and extracts the feature maps " + }, + { + "bbox": [ + 55, + 494, + 296, + 555 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_Q, \\mathcal{F}_T \\in \\mathbb{R}^{\\frac{H}{16} \\times \\frac{W}{16} \\times 1024}" + }, + { + "bbox": [ + 55, + 494, + 296, + 555 + ], + "type": "text", + "content": ", respectively. The decoder and following heads then processes " + }, + { + "bbox": [ + 55, + 494, + 296, + 555 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_Q" + }, + { + "bbox": [ + 55, + 494, + 296, + 555 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 494, + 296, + 555 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_T" + }, + { + "bbox": [ + 55, + 494, + 296, + 555 + ], + "type": "text", + "content": " to compute both a tensor for patch-level classification " + }, + { + "bbox": [ + 55, + 494, + 296, + 555 + ], + "type": "inline_equation", + "content": "\\mathcal{C} \\in \\mathbb{R}^{\\frac{H}{16} \\times \\frac{W}{16} \\times K}" + }, + { + "bbox": [ + 55, + 494, + 296, + 555 + ], + "type": "text", + "content": " and the xy-offsets " + }, + { + "bbox": [ + 55, + 494, + 296, + 555 + ], + "type": "inline_equation", + "content": "\\mathcal{U} \\in \\mathbb{R}^{\\frac{H}{16} \\times \\frac{W}{16} \\times 2}" + }, + { + "bbox": [ + 55, + 494, + 296, + 555 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 555, + 295, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 555, + 295, + 653 + ], + "spans": [ + { + "bbox": [ + 55, + 555, + 295, + 653 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 55, + 555, + 295, + 653 + ], + "type": "inline_equation", + "content": "K = \\frac{H}{16} \\times \\frac{W}{16} + 1" + }, + { + "bbox": [ + 55, + 555, + 295, + 653 + ], + "type": "text", + "content": " represents the number of classes for the patch-level classification. The first " + }, + { + "bbox": [ + 55, + 555, + 295, + 653 + ], + "type": "inline_equation", + "content": "\\frac{H}{16} \\times \\frac{W}{16}" + }, + { + "bbox": [ + 55, + 555, + 295, + 653 + ], + "type": "text", + "content": " classes correspond to the template patches, while the last class represents the case where there is no matching part, such as occlusion. For each position " + }, + { + "bbox": [ + 55, + 555, + 295, + 653 + ], + "type": "inline_equation", + "content": "(i,j)" + }, + { + "bbox": [ + 55, + 555, + 295, + 653 + ], + "type": "text", + "content": " in the feature map, " + }, + { + "bbox": [ + 55, + 555, + 295, + 653 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_{i,j} \\in \\mathbb{R}^K" + }, + { + "bbox": [ + 55, + 555, + 295, + 653 + ], + "type": "text", + "content": " contains classification scores indicating the degree of matchability between the " + }, + { + "bbox": [ + 55, + 555, + 295, + 653 + ], + "type": "inline_equation", + "content": "\\frac{H}{16} \\times \\frac{W}{16}" + }, + { + "bbox": [ + 55, + 555, + 295, + 653 + ], + "type": "text", + "content": " template patches and the query patch at that position." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": "Following this, we define the patch index as " + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "inline_equation", + "content": "c_{i,j} = \\arg \\max_{k\\in 1,2,\\dots,K}\\mathcal{C}_{i,j}^{k}" + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": " and the offset " + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{U}_{i,j}" + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": " has a range of " + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "inline_equation", + "content": "[-0.5,0.5]" + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": ". At location " + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "inline_equation", + "content": "(i,j)" + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "inline_equation", + "content": "c_{i,j} \\neq K" + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": ", the position " + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "inline_equation", + "content": "M_{i,j}^{T}" + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": " in the template corresponding to the query patch is given by:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 343, + 89, + 555, + 117 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 343, + 89, + 555, + 117 + ], + "spans": [ + { + "bbox": [ + 343, + 89, + 555, + 117 + ], + "type": "interline_equation", + "content": "\\mathcal {M} _ {i, j} ^ {T} = (\\left[ \\begin{array}{c} c _ {i, j} \\bmod 1 6 + 0. 5 \\\\ \\lfloor c _ {i, j} / 1 6 \\rfloor + 0. 5 \\end{array} \\right] + \\mathcal {U} _ {i, j}) \\times 1 6. \\qquad (1)", + "image_path": "4305db8420ad7782c1d0272a1fb81a027dec0d0e905eae8e1d765e68864c122b.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 122, + 555, + 233 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 122, + 555, + 233 + ], + "spans": [ + { + "bbox": [ + 313, + 122, + 555, + 233 + ], + "type": "text", + "content": "Adding " + }, + { + "bbox": [ + 313, + 122, + 555, + 233 + ], + "type": "inline_equation", + "content": "\\mathcal{U}_{i,j}" + }, + { + "bbox": [ + 313, + 122, + 555, + 233 + ], + "type": "text", + "content": " refines the position within the feature map grid. Multiplying by 16 maps this position back to the coordinate system of the original images " + }, + { + "bbox": [ + 313, + 122, + 555, + 233 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_Q" + }, + { + "bbox": [ + 313, + 122, + 555, + 233 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 122, + 555, + 233 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_T" + }, + { + "bbox": [ + 313, + 122, + 555, + 233 + ], + "type": "text", + "content": ", since each location in the feature map corresponds to a " + }, + { + "bbox": [ + 313, + 122, + 555, + 233 + ], + "type": "inline_equation", + "content": "16\\times 16" + }, + { + "bbox": [ + 313, + 122, + 555, + 233 + ], + "type": "text", + "content": " patch in the original images. The corresponding query image position " + }, + { + "bbox": [ + 313, + 122, + 555, + 233 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{i,j}^{Q}" + }, + { + "bbox": [ + 313, + 122, + 555, + 233 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 313, + 122, + 555, + 233 + ], + "type": "inline_equation", + "content": "((i + 0.5)\\times 16,(j + 0.5)\\times 16)" + }, + { + "bbox": [ + 313, + 122, + 555, + 233 + ], + "type": "text", + "content": ", which is the center position of each patch in " + }, + { + "bbox": [ + 313, + 122, + 555, + 233 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_Q" + }, + { + "bbox": [ + 313, + 122, + 555, + 233 + ], + "type": "text", + "content": ". The correspondence between the " + }, + { + "bbox": [ + 313, + 122, + 555, + 233 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_Q" + }, + { + "bbox": [ + 313, + 122, + 555, + 233 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 122, + 555, + 233 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_T" + }, + { + "bbox": [ + 313, + 122, + 555, + 233 + ], + "type": "text", + "content": " is then defined as " + }, + { + "bbox": [ + 313, + 122, + 555, + 233 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{i,j} = (\\mathcal{M}_{i,j}^{Q},\\mathcal{M}_{i,j}^{T})" + }, + { + "bbox": [ + 313, + 122, + 555, + 233 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 234, + 555, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 234, + 555, + 281 + ], + "spans": [ + { + "bbox": [ + 313, + 234, + 555, + 281 + ], + "type": "text", + "content": "Pose Fitting To estimate the pose in practice, we compute the correspondences " + }, + { + "bbox": [ + 313, + 234, + 555, + 281 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 313, + 234, + 555, + 281 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 313, + 234, + 555, + 281 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_Q" + }, + { + "bbox": [ + 313, + 234, + 555, + 281 + ], + "type": "text", + "content": " and all " + }, + { + "bbox": [ + 313, + 234, + 555, + 281 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_T" + }, + { + "bbox": [ + 313, + 234, + 555, + 281 + ], + "type": "text", + "content": ", and select the template " + }, + { + "bbox": [ + 313, + 234, + 555, + 281 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 234, + 555, + 281 + ], + "type": "text", + "content": " most similar to the query image. The similarity score " + }, + { + "bbox": [ + 313, + 234, + 555, + 281 + ], + "type": "inline_equation", + "content": "S_{t}" + }, + { + "bbox": [ + 313, + 234, + 555, + 281 + ], + "type": "text", + "content": " for each template " + }, + { + "bbox": [ + 313, + 234, + 555, + 281 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 234, + 555, + 281 + ], + "type": "text", + "content": " is defined as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 358, + 290, + 554, + 323 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 290, + 554, + 323 + ], + "spans": [ + { + "bbox": [ + 358, + 290, + 554, + 323 + ], + "type": "interline_equation", + "content": "S _ {t} = \\sum_ {i, j} \\left\\{ \\begin{array}{l l} \\max \\left(\\mathcal {C} _ {i, j}\\right) & \\text {i f} c _ {i, j} \\neq K \\\\ 0 & \\text {o t h e r w i s e} \\end{array} . \\right. \\tag {2}", + "image_path": "102df7e4f5a88260aeeccd78a558c945c7e6828b3760234f56aa09ace3a43d5f.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 327, + 554, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 327, + 554, + 448 + ], + "spans": [ + { + "bbox": [ + 313, + 327, + 554, + 448 + ], + "type": "text", + "content": "If the predicted class " + }, + { + "bbox": [ + 313, + 327, + 554, + 448 + ], + "type": "inline_equation", + "content": "c_{i,j} = K" + }, + { + "bbox": [ + 313, + 327, + 554, + 448 + ], + "type": "text", + "content": " (representing occlusion or unmatched patches), we exclude " + }, + { + "bbox": [ + 313, + 327, + 554, + 448 + ], + "type": "inline_equation", + "content": "\\max(\\mathcal{C}_{i,j})" + }, + { + "bbox": [ + 313, + 327, + 554, + 448 + ], + "type": "text", + "content": " from the sum. By computing " + }, + { + "bbox": [ + 313, + 327, + 554, + 448 + ], + "type": "inline_equation", + "content": "S_t" + }, + { + "bbox": [ + 313, + 327, + 554, + 448 + ], + "type": "text", + "content": " for each template, we select the template with the highest similarity score as the best match to " + }, + { + "bbox": [ + 313, + 327, + 554, + 448 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_Q" + }, + { + "bbox": [ + 313, + 327, + 554, + 448 + ], + "type": "text", + "content": ". Once the most similar template is determined, we construct 2D-3D correspondences using the correspondences between the query image and the selected template along with the depth information from the template. We estimate the initial pose by applying RANSAC [16] with the EPnP algorithm [28]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 454, + 415, + 465 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 454, + 415, + 465 + ], + "spans": [ + { + "bbox": [ + 314, + 454, + 415, + 465 + ], + "type": "text", + "content": "3.2. Pose Refinement" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 472, + 554, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 472, + 554, + 604 + ], + "spans": [ + { + "bbox": [ + 313, + 472, + 554, + 604 + ], + "type": "text", + "content": "In the refinement stage, we accurately refine the pose estimated from the coarse estimation at the pixel-level. To achieve this, we estimate pixel-wise dense correspondences. As shown in Fig. 2, the structure of the refinement model is the same as the coarse estimation model with an added Dense Prediction Transformer (DPT) [48]. The DPT enables pixel-wise prediction, allowing for precise pose refinement. Additionally, by iteratively refining the pose using a render-and-compare approach, this stage overcomes limitations that arise from using pre-rendered templates in the coarse estimation stage, such as self-occlusion." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "content": "Probabilistic Flow Regression Similar to previous works [23, 38], our refinement model estimates the flow between the query image " + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_Q" + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "content": " and the rendered image " + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_R" + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "content": " to refine the pose. To recover the pose accurately from the flow, it is essential to prevent inaccurate flows from significantly influencing the pose calculation. To this end, PFA [23] relies on RANSAC [16] to probabilistically exclude inaccurate flows from pose estimation. However, because RANSAC is sensitive to outlier distribution, it is insufficient when outliers" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11625" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 54, + 72, + 294, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 72, + 294, + 156 + ], + "spans": [ + { + "bbox": [ + 54, + 72, + 294, + 156 + ], + "type": "text", + "content": "are prevalent or unevenly distributed. GenFlow [38] focuses on reducing the influence of inaccurate flows by learning to estimate a confidence map through an end-to-end differentiable system and 6D pose loss. Although they show accurate and robust pose estimation performance, there is still room for improvement since learning confidence does not enhance the accuracy of the flow prediction itself." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 157, + 295, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 157, + 295, + 346 + ], + "spans": [ + { + "bbox": [ + 57, + 157, + 295, + 346 + ], + "type": "text", + "content": "Unlike previous flow-based refinement methods [23, 38], we aim to learn the conditional probability of the flow. In line with several correspondence estimation studies [6, 56], we define this conditional probability as " + }, + { + "bbox": [ + 57, + 157, + 295, + 346 + ], + "type": "inline_equation", + "content": "p(Y|\\mathcal{I}_Q,\\mathcal{I}_R;\\theta)" + }, + { + "bbox": [ + 57, + 157, + 295, + 346 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 57, + 157, + 295, + 346 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 57, + 157, + 295, + 346 + ], + "type": "text", + "content": " represents the flow between the image pair " + }, + { + "bbox": [ + 57, + 157, + 295, + 346 + ], + "type": "inline_equation", + "content": "(\\mathcal{I}_Q,\\mathcal{I}_R)" + }, + { + "bbox": [ + 57, + 157, + 295, + 346 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 57, + 157, + 295, + 346 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 57, + 157, + 295, + 346 + ], + "type": "text", + "content": " denotes the network parameters. Many methods achieve this by learning the variance of the prediction " + }, + { + "bbox": [ + 57, + 157, + 295, + 346 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 57, + 157, + 295, + 346 + ], + "type": "text", + "content": ", and the predictive density is modeled using Gaussian or Laplace distributions. We model it as a univariate Laplace distribution to simplify the problem. Specifically, " + }, + { + "bbox": [ + 57, + 157, + 295, + 346 + ], + "type": "inline_equation", + "content": "p(Y|\\mathcal{I}_Q,\\mathcal{I}_R;\\theta)" + }, + { + "bbox": [ + 57, + 157, + 295, + 346 + ], + "type": "text", + "content": " is modeled as a Laplace distribution with mean " + }, + { + "bbox": [ + 57, + 157, + 295, + 346 + ], + "type": "inline_equation", + "content": "\\mu \\in \\mathbb{R}^{H\\times W\\times 2}" + }, + { + "bbox": [ + 57, + 157, + 295, + 346 + ], + "type": "text", + "content": " and scale " + }, + { + "bbox": [ + 57, + 157, + 295, + 346 + ], + "type": "inline_equation", + "content": "b\\in \\mathbb{R}^{H\\times W\\times 1}" + }, + { + "bbox": [ + 57, + 157, + 295, + 346 + ], + "type": "text", + "content": ", both predicted by the network. Formulating flow estimation as a probabilistic regression makes our model concentrate on highly reliable correspondences by adapting the scale parameter " + }, + { + "bbox": [ + 57, + 157, + 295, + 346 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 57, + 157, + 295, + 346 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 349, + 296, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 349, + 296, + 506 + ], + "spans": [ + { + "bbox": [ + 55, + 349, + 296, + 506 + ], + "type": "text", + "content": "Flow Confidence To compute the pose using the flow estimated by the model, we need a confidence " + }, + { + "bbox": [ + 55, + 349, + 296, + 506 + ], + "type": "inline_equation", + "content": "\\mathcal{W} \\in \\mathbb{R}^{H \\times W \\times 1}" + }, + { + "bbox": [ + 55, + 349, + 296, + 506 + ], + "type": "text", + "content": ". This value of " + }, + { + "bbox": [ + 55, + 349, + 296, + 506 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 55, + 349, + 296, + 506 + ], + "type": "text", + "content": " determines how much weight each flow error will have in calculating the pose in the differentiable PnP layer. " + }, + { + "bbox": [ + 55, + 349, + 296, + 506 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 55, + 349, + 296, + 506 + ], + "type": "text", + "content": " is calculated from the certainty, sensitivity, and flow probability, which are learned through different loss functions. Certainty estimates whether the flow from " + }, + { + "bbox": [ + 55, + 349, + 296, + 506 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_R" + }, + { + "bbox": [ + 55, + 349, + 296, + 506 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 349, + 296, + 506 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_Q" + }, + { + "bbox": [ + 55, + 349, + 296, + 506 + ], + "type": "text", + "content": " is occluded or not. Sensitivity is learned from the pose loss and highlights areas with rich textures or extremities of the object. Similar to PDC-Net+ [56], we define the flow probability " + }, + { + "bbox": [ + 55, + 349, + 296, + 506 + ], + "type": "inline_equation", + "content": "P_R" + }, + { + "bbox": [ + 55, + 349, + 296, + 506 + ], + "type": "text", + "content": ", which represents the probability that the true flow " + }, + { + "bbox": [ + 55, + 349, + 296, + 506 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 55, + 349, + 296, + 506 + ], + "type": "text", + "content": " is within a radius " + }, + { + "bbox": [ + 55, + 349, + 296, + 506 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 55, + 349, + 296, + 506 + ], + "type": "text", + "content": " from the estimated mean flow " + }, + { + "bbox": [ + 55, + 349, + 296, + 506 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 55, + 349, + 296, + 506 + ], + "type": "text", + "content": ". This can be calculated as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 81, + 522, + 295, + 550 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 522, + 295, + 550 + ], + "spans": [ + { + "bbox": [ + 81, + 522, + 295, + 550 + ], + "type": "interline_equation", + "content": "P _ {R} = P \\left(\\| y - \\mu \\| _ {1} < R\\right) = 1 - \\exp \\left(- \\frac {R}{b}\\right). \\tag {3}", + "image_path": "0a32bd1e57be852ee63803b954ff620aed1b024dd514792afdb92e5dc6f28340.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 555, + 295, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 555, + 295, + 651 + ], + "spans": [ + { + "bbox": [ + 55, + 555, + 295, + 651 + ], + "type": "inline_equation", + "content": "P_{R}" + }, + { + "bbox": [ + 55, + 555, + 295, + 651 + ], + "type": "text", + "content": " is an interpretable measure of reliability, representing the accuracy of the flow with a threshold " + }, + { + "bbox": [ + 55, + 555, + 295, + 651 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 55, + 555, + 295, + 651 + ], + "type": "text", + "content": ". The flow confidence " + }, + { + "bbox": [ + 55, + 555, + 295, + 651 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 55, + 555, + 295, + 651 + ], + "type": "text", + "content": " is calculated as the element-wise multiplication of certainty, sensitivity, and " + }, + { + "bbox": [ + 55, + 555, + 295, + 651 + ], + "type": "inline_equation", + "content": "P_{R}" + }, + { + "bbox": [ + 55, + 555, + 295, + 651 + ], + "type": "text", + "content": ". This means that " + }, + { + "bbox": [ + 55, + 555, + 295, + 651 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 55, + 555, + 295, + 651 + ], + "type": "text", + "content": " will have higher values when there is no occlusion (high certainty), discriminative information is available to solve the pose (high sensitivity), and the flow is accurate (high " + }, + { + "bbox": [ + 55, + 555, + 295, + 651 + ], + "type": "inline_equation", + "content": "P_{R}" + }, + { + "bbox": [ + 55, + 555, + 295, + 651 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "content": "Pose Update To compute the refined 6D pose " + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{\\mathrm{refined}}" + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "content": " using the flow " + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "content": " and flow confidence " + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "content": ", we employ a Levenberg-Marquardt (LM) algorithm-based PnP solver. Given the input pose " + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{\\mathrm{input}} = [\\mathbf{R}_{\\mathrm{input}} \\mid \\mathbf{t}_{\\mathrm{input}}]" + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "content": ", the camera intrinsic matrix " + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{K}" + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "content": ", and the depth map " + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_R" + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "content": " corresponding to " + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_R" + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "content": ", we" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 317, + 71, + 553, + 186 + ], + "blocks": [ + { + "bbox": [ + 317, + 71, + 553, + 186 + ], + "lines": [ + { + "bbox": [ + 317, + 71, + 553, + 186 + ], + "spans": [ + { + "bbox": [ + 317, + 71, + 553, + 186 + ], + "type": "image", + "image_path": "085c02889f82c2b52104fa7b864cda011a2ea9376379532d59c0c37cd7414e74.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 193, + 555, + 227 + ], + "lines": [ + { + "bbox": [ + 313, + 193, + 555, + 227 + ], + "spans": [ + { + "bbox": [ + 313, + 193, + 555, + 227 + ], + "type": "text", + "content": "Figure 4. Pose Selection. To achieve more precise pose estimation using a multiple hypothesis strategy, we introduce a pose selection stage (Sec 3.3)." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 248, + 555, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 248, + 555, + 274 + ], + "spans": [ + { + "bbox": [ + 313, + 248, + 555, + 274 + ], + "type": "text", + "content": "calculate the 3D spatial coordinates " + }, + { + "bbox": [ + 313, + 248, + 555, + 274 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{u,v}^{3\\mathrm{D}}" + }, + { + "bbox": [ + 313, + 248, + 555, + 274 + ], + "type": "text", + "content": " corresponding to " + }, + { + "bbox": [ + 313, + 248, + 555, + 274 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_R" + }, + { + "bbox": [ + 313, + 248, + 555, + 274 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 346, + 286, + 555, + 303 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 346, + 286, + 555, + 303 + ], + "spans": [ + { + "bbox": [ + 346, + 286, + 555, + 303 + ], + "type": "interline_equation", + "content": "\\mathbf {x} _ {u, v} ^ {\\mathrm {3 D}} = \\mathbf {R} _ {\\text {i n p u t}} ^ {- 1} \\left(\\mathbf {K} ^ {- 1} \\mathcal {D} _ {R} (u, v) \\mathbf {x} _ {u, v} ^ {\\mathrm {2 D}} - \\mathbf {t} _ {\\text {i n p u t}}\\right), \\tag {4}", + "image_path": "64853a818c2281154383d086301d53bf29437a11f23ac4e52d276102b980e8de.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 308, + 555, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 308, + 555, + 357 + ], + "spans": [ + { + "bbox": [ + 313, + 308, + 555, + 357 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 308, + 555, + 357 + ], + "type": "inline_equation", + "content": "(u,v)" + }, + { + "bbox": [ + 313, + 308, + 555, + 357 + ], + "type": "text", + "content": " are the pixel coordinates in " + }, + { + "bbox": [ + 313, + 308, + 555, + 357 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_R" + }, + { + "bbox": [ + 313, + 308, + 555, + 357 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 308, + 555, + 357 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_R(u,v)" + }, + { + "bbox": [ + 313, + 308, + 555, + 357 + ], + "type": "text", + "content": " is the depth value at each pixel, and " + }, + { + "bbox": [ + 313, + 308, + 555, + 357 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{u,v}^{2\\mathrm{D}} = (u,v,1)^T" + }, + { + "bbox": [ + 313, + 308, + 555, + 357 + ], + "type": "text", + "content": ". We optimize the 6D pose by minimizing the sum of squared weighted reprojection errors as follows:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 341, + 377, + 554, + 422 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 341, + 377, + 554, + 422 + ], + "spans": [ + { + "bbox": [ + 341, + 377, + 554, + 422 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\underset {\\mathrm {R}, \\mathrm {t}} {\\operatorname {a r g m i n}} \\frac {1}{2} \\sum_ {u} \\sum_ {v} \\| \\mathcal {W} (u, v) \\times (\\pi (\\mathrm {R x} _ {u, v} ^ {\\mathrm {3 D}} + \\mathrm {t}) \\tag {5} \\\\ - \\left((u, v) ^ {T} + Y (u, v))\\right) \\| ^ {2}. \\\\ \\end{array}", + "image_path": "7e77ad38974377243f6f07801ee7229a2f247e00ef131c68bf1ebadfa114aacc.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 432, + 554, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 432, + 554, + 468 + ], + "spans": [ + { + "bbox": [ + 313, + 432, + 554, + 468 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 313, + 432, + 554, + 468 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 313, + 432, + 554, + 468 + ], + "type": "text", + "content": " is the reprojection function that maps 3D points in camera coordinates to 2D image points using the intrinsic camera parameters " + }, + { + "bbox": [ + 313, + 432, + 554, + 468 + ], + "type": "inline_equation", + "content": "\\mathbf{K}" + }, + { + "bbox": [ + 313, + 432, + 554, + 468 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 469, + 554, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 469, + 554, + 506 + ], + "spans": [ + { + "bbox": [ + 313, + 469, + 554, + 506 + ], + "type": "text", + "content": "Following previous work [7, 38], we update the pose in three iterations using the LM algorithm and further refine it to the final pose using the Gauss-Newton algorithm." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 514, + 403, + 526 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 514, + 403, + 526 + ], + "spans": [ + { + "bbox": [ + 313, + 514, + 403, + 526 + ], + "type": "text", + "content": "3.3. Pose Selection" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "text", + "content": "In the coarse estimation stage, the best scoring template may not provide the optimal initial pose for refinement. For example, selecting a template rotated by 180 degrees relative to the ground truth (see Fig. 4) makes refinement challenging. To address this issue, many methods [27, 38, 41, 44, 62] employ a multiple hypothesis strategy: generating " + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "text", + "content": " coarse pose estimates, refining each, and selecting the best match by comparing rendered results to the query image. Similar to the coarse estimation model, our pose selection model takes the query image and template as inputs but incorporates a score head to estimate a score for each pose hypothesis. Although this increases inference time, considering multiple refined poses helps avoid cases where a difficult-to-refine template leads to a poor final pose." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "11626" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 77, + 119, + 242 + ], + "blocks": [ + { + "bbox": [ + 61, + 77, + 119, + 242 + ], + "lines": [ + { + "bbox": [ + 61, + 77, + 119, + 242 + ], + "spans": [ + { + "bbox": [ + 61, + 77, + 119, + 242 + ], + "type": "image", + "image_path": "00e0e8f0a6251e9ce02eeb56aca1c5cb676371e176d8eab60efac44be5741155.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 78, + 243, + 101, + 253 + ], + "lines": [ + { + "bbox": [ + 78, + 243, + 101, + 253 + ], + "spans": [ + { + "bbox": [ + 78, + 243, + 101, + 253 + ], + "type": "text", + "content": "Query" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 265, + 295, + 310 + ], + "lines": [ + { + "bbox": [ + 55, + 265, + 295, + 310 + ], + "spans": [ + { + "bbox": [ + 55, + 265, + 295, + 310 + ], + "type": "text", + "content": "Figure 5. In-plane Rotation Invariant Matching Example. From left to right: Query image, semi-dense correspondences between the query image and the best scoring template, and the coarse pose recovered using the PnP algorithm." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 126, + 78, + 235, + 242 + ], + "blocks": [ + { + "bbox": [ + 126, + 78, + 235, + 242 + ], + "lines": [ + { + "bbox": [ + 126, + 78, + 235, + 242 + ], + "spans": [ + { + "bbox": [ + 126, + 78, + 235, + 242 + ], + "type": "image", + "image_path": "8fe42afed8cf11b18a494bdb2c6a78cc2be0c3ca475d311fabe9801e3e012bbc.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 150, + 243, + 214, + 252 + ], + "lines": [ + { + "bbox": [ + 150, + 243, + 214, + 252 + ], + "spans": [ + { + "bbox": [ + 150, + 243, + 214, + 252 + ], + "type": "text", + "content": "Correspondences" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 235, + 78, + 290, + 242 + ], + "blocks": [ + { + "bbox": [ + 235, + 78, + 290, + 242 + ], + "lines": [ + { + "bbox": [ + 235, + 78, + 290, + 242 + ], + "spans": [ + { + "bbox": [ + 235, + 78, + 290, + 242 + ], + "type": "image", + "image_path": "d4cc4ea731ab12b80c8e7ad17e6a22bc5d63b37f1c0ecc4335f4cbc0abce59eb.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 239, + 243, + 284, + 252 + ], + "lines": [ + { + "bbox": [ + 239, + 243, + 284, + 252 + ], + "spans": [ + { + "bbox": [ + 239, + 243, + 284, + 252 + ], + "type": "text", + "content": "Coarse Pose" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 331, + 119, + 344 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 331, + 119, + 344 + ], + "spans": [ + { + "bbox": [ + 55, + 331, + 119, + 344 + ], + "type": "text", + "content": "3.4. Training" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 350, + 296, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 350, + 296, + 456 + ], + "spans": [ + { + "bbox": [ + 55, + 350, + 296, + 456 + ], + "type": "text", + "content": "Datasets To train our three models—the coarse estimator, the pose refiner, and the pose selector—we need RGB-D images with ground-truth 6D object pose annotations. We use the large-scale dataset provided by MegaPose [27]. This dataset consists of synthetic data generated using BlenderProc [8], featuring a diverse set of objects from ShapeNet [5] and Google Scanned Objects (GSO) [11], including comprehensive ground-truth 6D pose annotations and object masks." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 460, + 296, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 460, + 296, + 639 + ], + "spans": [ + { + "bbox": [ + 55, + 460, + 296, + 639 + ], + "type": "text", + "content": "Coarse Model Our coarse estimation model is trained to estimate the correspondence between " + }, + { + "bbox": [ + 55, + 460, + 296, + 639 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_Q" + }, + { + "bbox": [ + 55, + 460, + 296, + 639 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 460, + 296, + 639 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_T" + }, + { + "bbox": [ + 55, + 460, + 296, + 639 + ], + "type": "text", + "content": ". Similar to GigaPose [41], we select templates that have similar out-of-plane rotations but different in-plane rotations compared to the training images cropped around the target object. As shown in Fig. 5, our model is designed to estimate correspondences between the query image and the template that are invariant to in-plane rotations. This invariance means that the templates only need to consider out-of-plane rotations, significantly reducing the number of templates required. Since we have the 3D model of the target object, we can generate these correspondences by projecting the 3D model onto the image planes of both the query and template images. Our model is trained to estimate these generated 2D-2D correspondences." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": "Refiner Model Our refinement model is trained similarly to previous works [26, 27, 38] that use the render-and-compass are approach. We generate noisy input poses " + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{\\mathrm{input}}" + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": " by adding zero-mean Gaussian noise to the ground truth poses " + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{\\mathrm{gt}}" + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": ", with translation noise standard deviations of (0.01, 0.01, 0.05) along the x, y, and z axes, and rotation noise" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 72, + 553, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 97 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 97 + ], + "type": "text", + "content": "standard deviations of 15 degrees per axis in Euler angles. The model is trained to predict " + }, + { + "bbox": [ + 313, + 72, + 553, + 97 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{\\mathrm{gt}}" + }, + { + "bbox": [ + 313, + 72, + 553, + 97 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 313, + 72, + 553, + 97 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{\\mathrm{input}}" + }, + { + "bbox": [ + 313, + 72, + 553, + 97 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 98, + 555, + 218 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 98, + 555, + 218 + ], + "spans": [ + { + "bbox": [ + 313, + 98, + 555, + 218 + ], + "type": "text", + "content": "Selection Model Our selection model is trained to assess the similarity between " + }, + { + "bbox": [ + 313, + 98, + 555, + 218 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_Q" + }, + { + "bbox": [ + 313, + 98, + 555, + 218 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 98, + 555, + 218 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_R" + }, + { + "bbox": [ + 313, + 98, + 555, + 218 + ], + "type": "text", + "content": " based on " + }, + { + "bbox": [ + 313, + 98, + 555, + 218 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{\\mathrm{input}}" + }, + { + "bbox": [ + 313, + 98, + 555, + 218 + ], + "type": "text", + "content": " to select the most accurate pose from multiple hypotheses. We train the model using binary cross-entropy loss with six poses per " + }, + { + "bbox": [ + 313, + 98, + 555, + 218 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{\\mathrm{gt}}" + }, + { + "bbox": [ + 313, + 98, + 555, + 218 + ], + "type": "text", + "content": ": one positive, five negatives. Positives have translation differences within (0.01, 0.01, 0.05) and rotation differences within 5 degrees of " + }, + { + "bbox": [ + 313, + 98, + 555, + 218 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{\\mathrm{gt}}" + }, + { + "bbox": [ + 313, + 98, + 555, + 218 + ], + "type": "text", + "content": ". By setting small rotation thresholds for positives, we enhance the model's ability to distinguish poses close to the ground truth, improving its discriminative capability." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 226, + 447, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 226, + 447, + 239 + ], + "spans": [ + { + "bbox": [ + 314, + 226, + 447, + 239 + ], + "type": "text", + "content": "3.5. Implementation Details" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 245, + 554, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 245, + 554, + 353 + ], + "spans": [ + { + "bbox": [ + 313, + 245, + 554, + 353 + ], + "type": "text", + "content": "The encoder and decoder architecture of Co-op are based on CroCo [61], a vision foundation model trained on large-scale datasets for 3D vision tasks. This allows us to significantly leverage the benefits of CroCo pre-training. The coarse model processes input images of resolution " + }, + { + "bbox": [ + 313, + 245, + 554, + 353 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 313, + 245, + 554, + 353 + ], + "type": "text", + "content": ", whereas the refinement and selection models process images of resolution " + }, + { + "bbox": [ + 313, + 245, + 554, + 353 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 313, + 245, + 554, + 353 + ], + "type": "text", + "content": ". For detailed information about the model configurations, learning rates, and training schedules, please refer to the supplementary material." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 354, + 554, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 354, + 554, + 415 + ], + "spans": [ + { + "bbox": [ + 313, + 354, + 554, + 415 + ], + "type": "text", + "content": "Coarse Model In our coarse model, we define the semidense correspondence as a hybrid representation combining patch-level classification and offset regression. Therefore, the coarse model is trained using two loss functions: the classification loss " + }, + { + "bbox": [ + 313, + 354, + 554, + 415 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{cls}}" + }, + { + "bbox": [ + 313, + 354, + 554, + 415 + ], + "type": "text", + "content": " and the regression loss " + }, + { + "bbox": [ + 313, + 354, + 554, + 415 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{reg}}" + }, + { + "bbox": [ + 313, + 354, + 554, + 415 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 415, + 554, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 415, + 554, + 536 + ], + "spans": [ + { + "bbox": [ + 313, + 415, + 554, + 536 + ], + "type": "text", + "content": "When defining the ground-truth match from the patch center " + }, + { + "bbox": [ + 313, + 415, + 554, + 536 + ], + "type": "inline_equation", + "content": "(i,j)" + }, + { + "bbox": [ + 313, + 415, + 554, + 536 + ], + "type": "text", + "content": " in the query image to the template as " + }, + { + "bbox": [ + 313, + 415, + 554, + 536 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{gt}^T = (\\bar{u},\\bar{v})" + }, + { + "bbox": [ + 313, + 415, + 554, + 536 + ], + "type": "text", + "content": ", the ground-truth index for " + }, + { + "bbox": [ + 313, + 415, + 554, + 536 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{cls}}" + }, + { + "bbox": [ + 313, + 415, + 554, + 536 + ], + "type": "text", + "content": " is defined as " + }, + { + "bbox": [ + 313, + 415, + 554, + 536 + ], + "type": "inline_equation", + "content": "\\left\\lfloor \\frac{\\bar{v}}{16}\\right\\rfloor \\times \\left(\\frac{W}{16}\\right) + \\left\\lfloor \\frac{\\bar{u}}{16}\\right\\rfloor" + }, + { + "bbox": [ + 313, + 415, + 554, + 536 + ], + "type": "text", + "content": " when a match exists. Here, " + }, + { + "bbox": [ + 313, + 415, + 554, + 536 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 313, + 415, + 554, + 536 + ], + "type": "text", + "content": " is the width of the template image, and dividing by 16 accounts for the downsizing due to the patch size. When there is no match, the ground-truth index is defined as " + }, + { + "bbox": [ + 313, + 415, + 554, + 536 + ], + "type": "inline_equation", + "content": "\\left(\\frac{H}{16} \\times \\frac{W}{16}\\right) + 1" + }, + { + "bbox": [ + 313, + 415, + 554, + 536 + ], + "type": "text", + "content": ", representing an additional class for \"no match.\" Therefore, " + }, + { + "bbox": [ + 313, + 415, + 554, + 536 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{cls}}" + }, + { + "bbox": [ + 313, + 415, + 554, + 536 + ], + "type": "text", + "content": " is defined as the cross-entropy loss between the ground-truth index and the model's output patch-level probability vector." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 326, + 537, + 526, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 537, + 526, + 550 + ], + "spans": [ + { + "bbox": [ + 326, + 537, + 526, + 550 + ], + "type": "text", + "content": "The offset ground truth " + }, + { + "bbox": [ + 326, + 537, + 526, + 550 + ], + "type": "inline_equation", + "content": "\\mathcal{U}_{gt}" + }, + { + "bbox": [ + 326, + 537, + 526, + 550 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 326, + 537, + 526, + 550 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{reg}}" + }, + { + "bbox": [ + 326, + 537, + 526, + 550 + ], + "type": "text", + "content": " is defined as:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 369, + 557, + 554, + 590 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 369, + 557, + 554, + 590 + ], + "spans": [ + { + "bbox": [ + 369, + 557, + 554, + 590 + ], + "type": "interline_equation", + "content": "\\mathcal {U} _ {g t} = \\frac {\\mathcal {M} _ {g t} ^ {T}}{1 6} - \\left\\lfloor \\frac {\\mathcal {M} _ {g t} ^ {T}}{1 6} \\right\\rfloor - 0. 5, \\tag {6}", + "image_path": "2e69075bc196f29eb57f4594b4610f15a5d58e62ca854312fe584519ae118a17.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 596, + 555, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 596, + 555, + 658 + ], + "spans": [ + { + "bbox": [ + 313, + 596, + 555, + 658 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 596, + 555, + 658 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{gt}^T" + }, + { + "bbox": [ + 313, + 596, + 555, + 658 + ], + "type": "text", + "content": " is the ground-truth match in the template. Subtracting " + }, + { + "bbox": [ + 313, + 596, + 555, + 658 + ], + "type": "inline_equation", + "content": "\\left\\lfloor \\mathcal{M}_{gt}^T /16\\right\\rfloor" + }, + { + "bbox": [ + 313, + 596, + 555, + 658 + ], + "type": "text", + "content": " and 0.5 centers the offset within the patch, ranging from -0.5 to 0.5. Here, " + }, + { + "bbox": [ + 313, + 596, + 555, + 658 + ], + "type": "inline_equation", + "content": "\\mathcal{U}" + }, + { + "bbox": [ + 313, + 596, + 555, + 658 + ], + "type": "text", + "content": " denotes the predicted offset output by the model. The regression loss " + }, + { + "bbox": [ + 313, + 596, + 555, + 658 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{reg}}" + }, + { + "bbox": [ + 313, + 596, + 555, + 658 + ], + "type": "text", + "content": " is then defined as the L1 loss between " + }, + { + "bbox": [ + 313, + 596, + 555, + 658 + ], + "type": "inline_equation", + "content": "\\mathcal{U}" + }, + { + "bbox": [ + 313, + 596, + 555, + 658 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 596, + 555, + 658 + ], + "type": "inline_equation", + "content": "\\mathcal{U}_{gt}" + }, + { + "bbox": [ + 313, + 596, + 555, + 658 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 392, + 670, + 553, + 685 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 392, + 670, + 553, + 685 + ], + "spans": [ + { + "bbox": [ + 392, + 670, + 553, + 685 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {r e g}} = \\left\\| \\mathcal {U} - \\mathcal {U} _ {g t} \\right\\| _ {1}. \\tag {7}", + "image_path": "0a1d844d491685ce02607d4a105efdc9d738f005b2c179b7e919b48caa909a4d.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 314, + 689, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 689, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 314, + 689, + 554, + 713 + ], + "type": "text", + "content": "Therefore, the overall loss for training the coarse model, " + }, + { + "bbox": [ + 314, + 689, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{coarse}}" + }, + { + "bbox": [ + 314, + 689, + 554, + 713 + ], + "type": "text", + "content": ", is defined as:" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11627" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 128, + 87, + 295, + 100 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 128, + 87, + 295, + 100 + ], + "spans": [ + { + "bbox": [ + 128, + 87, + 295, + 100 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {c o a r s e}} = \\mathcal {L} _ {\\text {c l s}} + \\alpha \\mathcal {L} _ {\\text {r e g}}. \\tag {8}", + "image_path": "eaaf3ec982cea38896077a6d5f222753f1ee40d5044bdb6274da59e902eb9371.jpg" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 110, + 211, + 121 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 110, + 211, + 121 + ], + "spans": [ + { + "bbox": [ + 67, + 110, + 211, + 121 + ], + "type": "text", + "content": "In the above equation, " + }, + { + "bbox": [ + 67, + 110, + 211, + 121 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 67, + 110, + 211, + 121 + ], + "type": "text", + "content": " is set to 2." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 124, + 296, + 196 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 124, + 296, + 196 + ], + "spans": [ + { + "bbox": [ + 55, + 124, + 296, + 196 + ], + "type": "text", + "content": "Refiner Model Our refiner model is trained using flow loss, certainty loss, and pose loss, similar to GenFlow [38]. Since we parameterize the flow output of the network with a Laplace distribution, we train the flow by minimizing the negative log-likelihood. Therefore, the flow loss is defined as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 72, + 217, + 296, + 247 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 217, + 296, + 247 + ], + "spans": [ + { + "bbox": [ + 72, + 217, + 296, + 247 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {f l o w} = \\sum_ {u} \\sum_ {v} \\left[ \\frac {\\left| \\mu_ {u , v} - \\bar {\\mu} _ {u , v} \\right|}{b _ {u , v}} + 2 \\log b _ {u, v} \\right]. \\tag {9}", + "image_path": "9731759d5a02ee05761c4e29276784f42423ee518eb2d5daa4114045316ed742.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 258, + 296, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 258, + 296, + 413 + ], + "spans": [ + { + "bbox": [ + 55, + 258, + 296, + 413 + ], + "type": "text", + "content": "In this equation, " + }, + { + "bbox": [ + 55, + 258, + 296, + 413 + ], + "type": "inline_equation", + "content": "\\mu_{u,v}" + }, + { + "bbox": [ + 55, + 258, + 296, + 413 + ], + "type": "text", + "content": " is the flow at pixel position " + }, + { + "bbox": [ + 55, + 258, + 296, + 413 + ], + "type": "inline_equation", + "content": "(u,v)" + }, + { + "bbox": [ + 55, + 258, + 296, + 413 + ], + "type": "text", + "content": " within the rendered mask, and " + }, + { + "bbox": [ + 55, + 258, + 296, + 413 + ], + "type": "inline_equation", + "content": "\\bar{\\mu}_{u,v}" + }, + { + "bbox": [ + 55, + 258, + 296, + 413 + ], + "type": "text", + "content": " is the ground-truth flow. " + }, + { + "bbox": [ + 55, + 258, + 296, + 413 + ], + "type": "inline_equation", + "content": "b_{u,v}" + }, + { + "bbox": [ + 55, + 258, + 296, + 413 + ], + "type": "text", + "content": " is the scale (uncertainty) at pixel position " + }, + { + "bbox": [ + 55, + 258, + 296, + 413 + ], + "type": "inline_equation", + "content": "(u,v)" + }, + { + "bbox": [ + 55, + 258, + 296, + 413 + ], + "type": "text", + "content": ". Additionally, the certainty loss " + }, + { + "bbox": [ + 55, + 258, + 296, + 413 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\text{cert}}" + }, + { + "bbox": [ + 55, + 258, + 296, + 413 + ], + "type": "text", + "content": " is defined as the binary cross-entropy loss that determines whether the flow from the rendered image " + }, + { + "bbox": [ + 55, + 258, + 296, + 413 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_R" + }, + { + "bbox": [ + 55, + 258, + 296, + 413 + ], + "type": "text", + "content": " to the query image " + }, + { + "bbox": [ + 55, + 258, + 296, + 413 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_Q" + }, + { + "bbox": [ + 55, + 258, + 296, + 413 + ], + "type": "text", + "content": " falls within the ground-truth mask in " + }, + { + "bbox": [ + 55, + 258, + 296, + 413 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_Q" + }, + { + "bbox": [ + 55, + 258, + 296, + 413 + ], + "type": "text", + "content": " or not. The pose loss " + }, + { + "bbox": [ + 55, + 258, + 296, + 413 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\text{pose}}" + }, + { + "bbox": [ + 55, + 258, + 296, + 413 + ], + "type": "text", + "content": " which quantifies the difference between the refined pose and the ground-truth pose, is defined as the distance between corresponding 3D points on the 3D model, following the approach used in previous works [27, 38]. Detailed information about " + }, + { + "bbox": [ + 55, + 258, + 296, + 413 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\text{pose}}" + }, + { + "bbox": [ + 55, + 258, + 296, + 413 + ], + "type": "text", + "content": " is provided in the supplementary material." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 415, + 291, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 415, + 291, + 426 + ], + "spans": [ + { + "bbox": [ + 67, + 415, + 291, + 426 + ], + "type": "text", + "content": "The overall loss for refiner model is defined as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 101, + 443, + 295, + 456 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 101, + 443, + 295, + 456 + ], + "spans": [ + { + "bbox": [ + 101, + 443, + 295, + 456 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {r e f i n e r}} = \\mathcal {L} _ {\\text {f l o w}} + \\beta \\mathcal {L} _ {\\text {c e r t}} + \\gamma \\mathcal {L} _ {\\text {p o s e}}. \\tag {10}", + "image_path": "367f28fb2d88ca2f70a9f6f862e955ce1752549a67baf36bcba9109b55ac368e.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 465, + 295, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 465, + 295, + 489 + ], + "spans": [ + { + "bbox": [ + 55, + 465, + 295, + 489 + ], + "type": "text", + "content": "The weights " + }, + { + "bbox": [ + 55, + 465, + 295, + 489 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 55, + 465, + 295, + 489 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 465, + 295, + 489 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 55, + 465, + 295, + 489 + ], + "type": "text", + "content": " for each loss are set to 5 and 20, respectively." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 502, + 137, + 517 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 502, + 137, + 517 + ], + "spans": [ + { + "bbox": [ + 55, + 502, + 137, + 517 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 523, + 172, + 537 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 523, + 172, + 537 + ], + "spans": [ + { + "bbox": [ + 55, + 523, + 172, + 537 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 543, + 296, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 543, + 296, + 626 + ], + "spans": [ + { + "bbox": [ + 55, + 543, + 296, + 626 + ], + "type": "text", + "content": "Datasets We evaluate our method on the seven core datasets of the BOP Challenge [22]: LineMod Occlusion(LM-O) [3], T-LESS [19], TUD-L [20], IC-BIN [10], ITODD [12], HomebrewedDB(HB) [24], and YCB-Video(YCB-V) [63]. Since our method does not require training for each dataset, we did not use any training images; only the 3D object models and test images were used." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "text", + "content": "Metrics We follow the protocol of the BOP Challenge [21], which consists of three pose-error functions: Visible Surface Discrepancy (VSD), Maximum Symmetry-Aware Surface Distance (MSSD), and Maximum Symmetry-Aware Projection Distance (MSPD). The overall accuracy is measured by the Average Recall (AR), calculated as: AR = " + }, + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "inline_equation", + "content": "(\\mathrm{AR}_{\\mathrm{VSD}} + \\mathrm{AR}_{\\mathrm{MSSD}} + \\mathrm{AR}_{\\mathrm{MSPD}}) / 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 72, + 454, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 454, + 83 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 454, + 83 + ], + "type": "text", + "content": "4.2. BOP Benchmark Results" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 91, + 555, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 91, + 555, + 258 + ], + "spans": [ + { + "bbox": [ + 313, + 91, + 555, + 258 + ], + "type": "text", + "content": "Table 1 shows the pose estimation results on the BOP dataset. In our method's experiments, we use a single RTX 3090 Ti GPU. For a fair comparison with previous works [1, 27, 38, 41, 44], we adopt the default CNOS [40] detections provided by the BOP challenge. We also set the number of hypotheses for the multiple hypotheses strategy to 5, consistent with the previous methods. Under these standardized conditions, our method demonstrates improved pose estimation accuracy compared to prior approaches. Although our method is proposed to solve the RGB-only object pose estimation problem, it is also available for RGBD input as a correspondence-based method. The results for this can be found in the supplementary material." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 277, + 427, + 289 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 277, + 427, + 289 + ], + "spans": [ + { + "bbox": [ + 313, + 277, + 427, + 289 + ], + "type": "text", + "content": "4.2.1 Coarse Estimation" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 298, + 555, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 298, + 555, + 441 + ], + "spans": [ + { + "bbox": [ + 313, + 298, + 555, + 441 + ], + "type": "text", + "content": "Rows 1 to 6 show the performance of various methods using only coarse estimation, without any refinement stage. Our method, Co-op (row 6), significantly outperforms previous approaches in this setting. Co-op achieves the highest AR score across all datasets, which is approximately " + }, + { + "bbox": [ + 313, + 298, + 555, + 441 + ], + "type": "inline_equation", + "content": "56.6\\%" + }, + { + "bbox": [ + 313, + 298, + 555, + 441 + ], + "type": "text", + "content": " higher than the second-best method, FoundPose [44] (row 5). Although GigaPose [41] is the fastest coarse estimation method with an inference time of 0.384 seconds per image (row 4), Co-op's coarse estimation alone achieves a higher mean AR score than GigaPose combined with the refinement model GenFlow [38] (row 11). This demonstrates that our proposed coarse estimation method is highly effective." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 460, + 419, + 471 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 460, + 419, + 471 + ], + "spans": [ + { + "bbox": [ + 313, + 460, + 419, + 471 + ], + "type": "text", + "content": "4.2.2 Pose Refinement" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 480, + 555, + 659 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 480, + 555, + 659 + ], + "spans": [ + { + "bbox": [ + 313, + 480, + 555, + 659 + ], + "type": "text", + "content": "Rows 7 to 13 present the results of combining coarse estimation and refinement using a single hypothesis. In this setting, our method, Co-op (row 13), outperforms previous works [27, 38]. To focus on the performance of the refinement model independently of the coarse estimation method, we conducted experiments by combining our refinement model with GigaPose [41] (rows 9, 11, and 12). This allows for a direct comparison of the refinement models using the same coarse estimation. The results show that our refinement model consistently outperforms others, achieving higher AR scores even when paired with the same coarse estimation method. In the multiple hypothesis experiments (rows 14 to 18), our method outperforms the previous best method, FoundPose [44] (row 17), by more than 6 points in AR score." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 670, + 408, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 670, + 408, + 682 + ], + "spans": [ + { + "bbox": [ + 313, + 670, + 408, + 682 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "content": "Table 2 presents various ablation evaluations on seven core datasets, as described in Section 4.1. The results for" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11628" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 55, + 69, + 555, + 285 + ], + "blocks": [ + { + "bbox": [ + 55, + 69, + 555, + 285 + ], + "lines": [ + { + "bbox": [ + 55, + 69, + 555, + 285 + ], + "spans": [ + { + "bbox": [ + 55, + 69, + 555, + 285 + ], + "type": "table", + "html": "
#Coarse estimationPose refinementLM-OT-LESSTUD-LIC-BINITODDHBYCB-VMeanTime(sec)
Coarse estimation only:
1ZS6D [1]-29.821.0----32.4--
2MegaPose [27]-22.917.725.815.210.825.128.120.815.465
3GenFlow [38]-25.021.530.016.815.428.327.723.53.839
4GigaPose [41]-29.927.330.223.118.834.829.027.60.384
5FoundPose [44]-39.733.846.923.920.450.845.237.31.690
6Co-op-59.759.264.245.839.178.162.658.40.979
With pose refinement (a single hypothesis):
7MegaPose [27]MegaPose [27]49.947.765.336.731.565.460.150.931.724
8GenFlow [38]GenFlow [38]54.751.467.043.738.473.061.955.710.553
9GigaPose [41]MegaPose [27]55.754.158.045.037.669.363.254.72.301
10FoundPose [44]MegaPose [27]55.751.063.343.335.769.766.155.06.385
11GigaPose [41]GenFlow [38]59.555.060.747.841.372.260.856.82.213
12GigaPose [41]Co-op60.256.062.850.243.473.961.958.31.149
13Co-opCo-op64.263.571.751.247.383.267.064.01.852
With pose refinement (5 hypotheses):
14MegaPose [27]MegaPose [27]56.050.768.441.433.870.462.154.747.386
15GenFlow [38]GenFlow [38]56.352.368.445.339.573.963.357.020.890
16GigaPose [41]MegaPose [27]59.957.064.546.739.772.266.357.97.682
17FoundPose [44]MegaPose [27]61.057.069.447.940.772.369.059.620.523
18Co-opCo-op65.564.872.954.249.084.968.965.74.186
", + "image_path": "46836c9aaf47554680d04f2c3ef0f2f14781421139fd55b426b511f0f47bae7c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 56, + 354, + 296, + 506 + ], + "blocks": [ + { + "bbox": [ + 55, + 301, + 555, + 335 + ], + "lines": [ + { + "bbox": [ + 55, + 301, + 555, + 335 + ], + "spans": [ + { + "bbox": [ + 55, + 301, + 555, + 335 + ], + "type": "text", + "content": "Table 1. Results on the BOP challenge datasets. The table shows the Average Recall (AR) score for each dataset, the mean AR score across all datasets, and the time required to estimate the poses of all objects within an image. From top to bottom, the results are ordered as follows: coarse estimation only, refinement applied to a single hypothesis, and refinement results with multiple hypotheses." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 354, + 296, + 506 + ], + "lines": [ + { + "bbox": [ + 56, + 354, + 296, + 506 + ], + "spans": [ + { + "bbox": [ + 56, + 354, + 296, + 506 + ], + "type": "table", + "html": "
#MethodLM-OT-LESSTUD-LIC-BINITODDHBYCB-VMean
Coarse Estimation:
1Ours (Proposed)59.759.264.245.839.178.162.658.4
2W/o Classification52.656.144.140.331.071.755.850.2
3W/o CroCo pretrained47.248.946.738.127.061.655.046.4
Pose Refinement:
4Ours (Proposed)64.263.571.751.247.383.267.064.0
5W/o Probabilistic Flow63.663.271.649.546.482.865.863.3
6W/o CroCo pretrained62.361.266.949.042.479.667.161.2
Pose Selection:
7Ours (Proposed)65.564.872.954.249.084.968.965.7
8W/o CroCo pretrained65.164.472.853.147.384.268.365.0
", + "image_path": "c4d67d0e6dc8da2b88daa294e67245338945761dc039712eb40a508fb17110da.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 522, + 295, + 567 + ], + "lines": [ + { + "bbox": [ + 55, + 522, + 295, + 567 + ], + "spans": [ + { + "bbox": [ + 55, + 522, + 295, + 567 + ], + "type": "text", + "content": "Table 2. Ablation Study. We present the results for each stage under various settings. Rows 1 to 3 correspond to Coarse Estimation, rows 4 to 5 to Pose Refinement, and rows 7 and 8 to Pose Selection. Please refer to Section 4.3 for more details." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 589, + 288, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 589, + 288, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 589, + 288, + 601 + ], + "type": "text", + "content": "each of our stages are shown in rows 1, 4, 7, respectively." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 603, + 295, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 603, + 295, + 700 + ], + "spans": [ + { + "bbox": [ + 55, + 603, + 295, + 700 + ], + "type": "text", + "content": "Fine-tuning vs. Training from Scratch We assessed the impact of initializing our models with CroCo [61] weights versus random initialization. Initializing with CroCo significantly improved performance in both the coarse estimation (rows 1 vs. 3) and pose refinement stages (rows 4 vs. 6). Although the performance improvement in the pose selection stage (rows 7, 8) is relatively small than others, using CroCo pretraining enhances performance across all stages." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 701, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 701, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 701, + 295, + 714 + ], + "type": "text", + "content": "Hybrid Representation vs. Direct Regression Comparing" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 356, + 555, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 356, + 555, + 429 + ], + "spans": [ + { + "bbox": [ + 313, + 356, + 555, + 429 + ], + "type": "text", + "content": "rows 1 and 2, we evaluated the impact of our hybrid representation (combining patch-level classification and offset regression) against direct regression of correspondences. Removing patch-level classification reduced the model's generalization ability, demonstrating the effectiveness of our hybrid approach." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 430, + 556, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 430, + 556, + 528 + ], + "spans": [ + { + "bbox": [ + 313, + 430, + 556, + 528 + ], + "type": "text", + "content": "Probabilistic Flow Regression We evaluated the effectiveness of defining flow as a conditional probability by training a model with the same architecture as Co-op but using the loss function and prediction head from GenFlow [38]. The results (rows 4 vs. 5) show that probabilistic flow regression improves flow learning and pose accuracy by estimating flow reliability, enabling the model to better handle ambiguous or unreliable correspondences." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 547, + 388, + 559 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 547, + 388, + 559 + ], + "spans": [ + { + "bbox": [ + 313, + 547, + 388, + 559 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "text", + "content": "We propose Co-op, a novel method for 6D pose estimation of unseen objects. Co-op can estimate the poses of novel objects using CAD models that are not provided during training but are available only during testing. We introduce a coarse estimation method that achieves strong generalization performance by leveraging a hybrid representation. Furthermore, we propose a render-and-compare refiner that employs probabilistic flow. Using these components, we demonstrate that our method significantly outperforms existing methods by a large margin on the BOP benchmark, which comprises hundreds of diverse objects and cluttered scenes." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "11629" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 133 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 133 + ], + "type": "text", + "content": "Acknowledgments. This work was supported by the Technology Innovation Program(00417108, Technologies of Rapid Digital Twin Development Tools for Robotic Service Environments) funded By the Ministry of Trade Industry & Energy(MOTIE, Korea)" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 141, + 115, + 154 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 141, + 115, + 154 + ], + "spans": [ + { + "bbox": [ + 56, + 141, + 115, + 154 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 161, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 161, + 296, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 161, + 296, + 205 + ], + "spans": [ + { + "bbox": [ + 61, + 161, + 296, + 205 + ], + "type": "text", + "content": "[1] Philipp Ausserlechner, David Haberger, Stefan Thalhammer, Jean-Baptiste Weibel, and Markus Vincze. Zs6d: Zero-shot 6d object pose estimation using vision transformers. In ICRA, pages 463-469. IEEE, 2024. 2, 7, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 206, + 296, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 206, + 296, + 239 + ], + "spans": [ + { + "bbox": [ + 62, + 206, + 296, + 239 + ], + "type": "text", + "content": "[2] Herbert Bay, Tinne Tuytelaars, and Luc Van Gool. Surf: Speeded up robust features. In ECCV, pages 404-417, Berlin, Heidelberg, 2006. Springer Berlin Heidelberg. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 239, + 296, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 239, + 296, + 282 + ], + "spans": [ + { + "bbox": [ + 62, + 239, + 296, + 282 + ], + "type": "text", + "content": "[3] Eric Brachmann, Alexander Krull, Frank Michel, Stefan Gumhold, Jamie Shotton, and Carsten Rother. Learning 6d object pose estimation using 3d object coordinates. In ECCV, 2014. 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 284, + 296, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 284, + 296, + 337 + ], + "spans": [ + { + "bbox": [ + 62, + 284, + 296, + 337 + ], + "type": "text", + "content": "[4] Michael Calonder, Vincent Lepetit, Mustafa Ozuysal, Tomasz Trzcinski, Christoph Strecha, and Pascal Fua. Brief: Computing a local binary descriptor very fast. IEEE Transactions on Pattern Analysis and Machine Intelligence, 34(7): 1281-1298, 2012. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 338, + 296, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 338, + 296, + 392 + ], + "spans": [ + { + "bbox": [ + 62, + 338, + 296, + 392 + ], + "type": "text", + "content": "[5] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 394, + 296, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 394, + 296, + 437 + ], + "spans": [ + { + "bbox": [ + 62, + 394, + 296, + 437 + ], + "type": "text", + "content": "[6] Hongkai Chen, Zixin Luo, Lei Zhou, Yurun Tian, Mingmin Zhen, Tian Fang, David McKinnon, Yanghai Tsin, and Long Quan. Aspanformer: Detector-free image matching with adaptive span transformer. In ECCV, 2022. 2, 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 438, + 296, + 481 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 438, + 296, + 481 + ], + "spans": [ + { + "bbox": [ + 62, + 438, + 296, + 481 + ], + "type": "text", + "content": "[7] Hansheng Chen, Pichao Wang, Fan Wang, Wei Tian, Lu Xiong, and Hao Li. Epro-pnp: Generalized end-to-end probabilistic perspective-n-points for monocular object pose estimation. In CVPR, pages 2781-2790, 2022. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 482, + 296, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 482, + 296, + 536 + ], + "spans": [ + { + "bbox": [ + 62, + 482, + 296, + 536 + ], + "type": "text", + "content": "[8] Maximilian Denninger, Dominik Winkelbauer, Martin Sundermeyer, Wout Boerdijk, Markus Knauer, Klaus H. Strobl, Matthias Humt, and Rudolph Triebel. Blenderproc2: A procedural pipeline for photorealistic rendering. Journal of Open Source Software, 8(82):4901, 2023. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 537, + 296, + 570 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 537, + 296, + 570 + ], + "spans": [ + { + "bbox": [ + 62, + 537, + 296, + 570 + ], + "type": "text", + "content": "[9] Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superpoint: Self-supervised interest point detection and description. In CVPRW, pages 337-33712, 2018. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 571, + 296, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 571, + 296, + 613 + ], + "spans": [ + { + "bbox": [ + 57, + 571, + 296, + 613 + ], + "type": "text", + "content": "[10] Andreas Doumanoglou, Rigas Kouskouridas, Sotiris Malassiotis, and Tae-Kyun Kim. Recovering 6d object pose and predicting next-best-view in the crowd. In CVPR, pages 3583-3592, 2016. 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 615, + 296, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 615, + 296, + 669 + ], + "spans": [ + { + "bbox": [ + 57, + 615, + 296, + 669 + ], + "type": "text", + "content": "[11] Laura Downs, Anthony Francis, Nate Koenig, Brandon Kinman, Ryan Hickman, Krista Reymann, Thomas B McHugh, and Vincent Vanhoucke. Google scanned objects: A high-quality dataset of 3d scanned household items. In ICRA, pages 2553-2560. IEEE, 2022. 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 670, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 296, + 713 + ], + "type": "text", + "content": "[12] Bertram Drost, Markus Ulrich, Paul Bergmann, Philipp Hartinger, and Carsten Steger. Introducing mvtec itodd — a dataset for 3d object recognition in industry. In ICCVW, pages 2200-2208, 2017. 7" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "text", + "content": "[13] Mihai Dusmanu, Ignacio Rocco, Tomas Pajdla, Marc Pollefeys, Josef Sivic, Akihiko Torii, and Torsten Sattler. D2-Net: A Trainable CNN for Joint Detection and Description of Local Features. In CVPR, 2019. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 119, + 553, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 119, + 553, + 152 + ], + "spans": [ + { + "bbox": [ + 317, + 119, + 553, + 152 + ], + "type": "text", + "content": "[14] Johan Edstedt, Ioannis Athanasiadis, Mårten Wadenback, and Michael Felsberg. DKM: Dense kernelized feature matching for geometry estimation. In CVPR, 2023. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 155, + 553, + 187 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 155, + 553, + 187 + ], + "spans": [ + { + "bbox": [ + 317, + 155, + 553, + 187 + ], + "type": "text", + "content": "[15] Johan Edstedt, Qiyu Sun, Georg Bokman, Marten Wadenback, and Michael Felsberg. RoMa: Robust Dense Feature Matching. In CVPR, 2024. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 190, + 553, + 233 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 190, + 553, + 233 + ], + "spans": [ + { + "bbox": [ + 317, + 190, + 553, + 233 + ], + "type": "text", + "content": "[16] Martin A. Fischler and Robert C. Bolles. Random sample consensus: A paradigm for model fitting with applications to image analysis and automated cartography. Commun. ACM, 24(6):381-395, 1981. 2, 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 236, + 553, + 280 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 236, + 553, + 280 + ], + "spans": [ + { + "bbox": [ + 317, + 236, + 553, + 280 + ], + "type": "text", + "content": "[17] Rasmus Laurvig Haugaard and Anders Glent Buch. Surfemb: Dense and continuous correspondence distributions for object pose estimation with learnt surface embeddings. In CVPR, pages 6749-6758, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 281, + 553, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 281, + 553, + 325 + ], + "spans": [ + { + "bbox": [ + 316, + 281, + 553, + 325 + ], + "type": "text", + "content": "[18] Xingyi He, Jiaming Sun, Yuang Wang, Di Huang, Hujun Bao, and Xiaowei Zhou. Onepose++: Keypoint-free one-shot object pose estimation without CAD models. In NeurIPS, 2022. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 327, + 553, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 327, + 553, + 371 + ], + "spans": [ + { + "bbox": [ + 316, + 327, + 553, + 371 + ], + "type": "text", + "content": "[19] Tomás Hodań, Pavel Haluza, Štepan Obdržálek, Jiří Matas, Manolis Lourakis, and Xenophon Zabulis. T-LESS: An RGB-D dataset for 6D pose estimation of texture-less objects. WACV, 2017. 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 373, + 553, + 439 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 373, + 553, + 439 + ], + "spans": [ + { + "bbox": [ + 316, + 373, + 553, + 439 + ], + "type": "text", + "content": "[20] Tomás Hodań, Frank Michel, Eric Brachmann, Wadim Kehl, Anders Glent Buch, Dirk Kraft, Bertram Drost, Joel Vidal, Stephan Ihrke, Xenophon Zabulis, Caner Sahin, Fabian Manhardt, Federico Tombari, Tae-Kyun Kim, Jiri Matas, and Carsten Rother. BOP: Benchmark for 6D object pose estimation. ECCV, 2018. 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 441, + 553, + 507 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 441, + 553, + 507 + ], + "spans": [ + { + "bbox": [ + 317, + 441, + 553, + 507 + ], + "type": "text", + "content": "[21] Tomáš Hodań, Martin Sundermeyer, Bertram Drost, Yann Labbé, Eric Brachmann, Frank Michel, Carsten Rother, and Jiri Matas. Bop challenge 2020 on 6d object localization. In Computer Vision-ECCV 2020 Workshops: Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16, pages 577-594. Springer, 2020. 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 509, + 553, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 509, + 553, + 586 + ], + "spans": [ + { + "bbox": [ + 317, + 509, + 553, + 586 + ], + "type": "text", + "content": "[22] Tomas Hodan, Martin Sundermeyer, Yann Labbe, Van Nguyen Nguyen, Gu Wang, Eric Brachmann, Bertram Drost, Vincent Lepetit, Carsten Rother, and Jiri Matas. Bop challenge 2023 on detection segmentation and pose estimation of seen and unseen rigid objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5610-5619, 2024. 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 588, + 553, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 588, + 553, + 621 + ], + "spans": [ + { + "bbox": [ + 316, + 588, + 553, + 621 + ], + "type": "text", + "content": "[23] Yinlin Hu, Pascal Fua, and Mathieu Salzmann. Perspective flow aggregation for data-limited 6d object pose estimation. In ECCV, pages 89-106. Springer, 2022. 2, 4, 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 624, + 553, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 553, + 666 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 553, + 666 + ], + "type": "text", + "content": "[24] Roman Kaskman, Sergey Zakharov, Ivan Shugurov, and Slobodan Ilic. Homebreweddb: Rgb-d dataset for 6d pose estimation of 3d objects. In ICCVW, pages 2767-2776, 2019. 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 669, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 669, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 669, + 553, + 713 + ], + "type": "text", + "content": "[25] Wadim Kehl, Fabian Manhardt, Federico Tombari, Slobodan Ilic, and Nassir Navab. Ssd-6d: Making rgb-based 3d detection and 6d pose estimation great again. In ICCV, pages 1521–1529, 2017. 2" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "11630" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 715 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "type": "text", + "content": "[26] Y. Labbe, J. Carpentier, M. Aubry, and J. Sivic. Cosyposse: Consistent multi-view multi-object 6d pose estimation. In ECCV, 2020. 2, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 294, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 294, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 294, + 162 + ], + "type": "text", + "content": "[27] Yann Labbe, Lucas Manuelli, Arsalan Mousavian, Stephen Tyree, Stan Birchfield, Jonathan Tremblay, Justin Carpentier, Mathieu Aubry, Dieter Fox, and Josef Sivic. MegaPose: 6D Pose Estimation of Novel Objects via Render & Compare. In CoRL, 2022. 1, 2, 5, 6, 7, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 294, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 294, + 206 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 294, + 206 + ], + "type": "text", + "content": "[28] Vincent Lepetit, Francesc Moreno-Noguer, and Pascal Fua. Epnp: An accurate o(n) solution to the pnp problem. International Journal Of Computer Vision, 81:155-166, 2009. 2, 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 208, + 294, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 208, + 294, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 208, + 294, + 262 + ], + "type": "text", + "content": "[29] Fu Li, Shishir Reddy Vutukur, Hao Yu, Ivan Shugurov, Benjamin Busam, Shaowu Yang, and Slobodan Ilic. Nerfpose: A first-reconstruct-then-regress approach for weakly-supervised 6d object pose estimation. In ICCV, pages 2123–2133, 2023. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 263, + 294, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 263, + 294, + 296 + ], + "spans": [ + { + "bbox": [ + 56, + 263, + 294, + 296 + ], + "type": "text", + "content": "[30] Yi Li, Gu Wang, Xiangyang Ji, Yu Xiang, and Dieter Fox. DeepIM: Deep iterative matching for 6D pose estimation. In ECCV, 2018. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 297, + 294, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 297, + 294, + 342 + ], + "spans": [ + { + "bbox": [ + 56, + 297, + 294, + 342 + ], + "type": "text", + "content": "[31] Zhigang Li, Gu Wang, and Xiangyang Ji. Cdpn: Coordinates-based disentangled pose network for real-time rgb-based 6-dof object pose estimation. In ICCV, pages 7678-7687, 2019. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 343, + 294, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 343, + 294, + 376 + ], + "spans": [ + { + "bbox": [ + 56, + 343, + 294, + 376 + ], + "type": "text", + "content": "[32] Jiehong Lin, Lihua Liu, Dekun Lu, and Kui Jia. Sam-6d: Segment anything model meets zero-shot 6d object pose estimation. In CVPR, pages 27906–27916, 2024. 1, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 376, + 294, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 376, + 294, + 420 + ], + "spans": [ + { + "bbox": [ + 56, + 376, + 294, + 420 + ], + "type": "text", + "content": "[33] Yunzhi Lin, Jonathan Tremblay, Stephen Tyree, Patricio A. Vela, and Stan Birchfield. Single-stage keypoint-based category-level object pose estimation from an RGB image. In ICRA, 2022. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 422, + 294, + 455 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 422, + 294, + 455 + ], + "spans": [ + { + "bbox": [ + 56, + 422, + 294, + 455 + ], + "type": "text", + "content": "[34] Lahav Lipson, Zachary Teed, Ankit Goyal, and Jia Deng. Coupled iterative refinement for 6d multi-object pose estimation. In CVPR, pages 6728-6737, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 456, + 294, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 456, + 294, + 510 + ], + "spans": [ + { + "bbox": [ + 56, + 456, + 294, + 510 + ], + "type": "text", + "content": "[35] Xingyu Liu, Ruida Zhang, Chenyangguang Zhang, Bowen Fu, Jiwen Tang, Xiquan Liang, Jingyi Tang, Xiaotian Cheng, Yukang Zhang, Gu Wang, and Xiangyang Ji. Gdrnpp. https://github.com/shanice-1/gdrnpp_bop2022, 2022.1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 512, + 294, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 512, + 294, + 555 + ], + "spans": [ + { + "bbox": [ + 56, + 512, + 294, + 555 + ], + "type": "text", + "content": "[36] Yuan Liu, Yilin Wen, Sida Peng, Cheng Lin, Xiaoxiao Long, Taku Komura, and Wenping Wang. Gen6d: Generalizable model-free 6-dof object pose estimation from rgb images. In ECCV, 2022. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 556, + 294, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 556, + 294, + 589 + ], + "spans": [ + { + "bbox": [ + 56, + 556, + 294, + 589 + ], + "type": "text", + "content": "[37] David G. Lowe. Distinctive image features from scale-invariant keypoints. International Journal of Computer Vision, 60:91-110, 2004. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 590, + 294, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 590, + 294, + 634 + ], + "spans": [ + { + "bbox": [ + 56, + 590, + 294, + 634 + ], + "type": "text", + "content": "[38] Sungphill Moon, Hyeontae Son, Dongcheol Hur, and Sangwook Kim. Genflow: Generalizable recurrent flow for 6d pose refinement of novel objects. In CVPR, pages 10039-10049, 2024. 1, 2, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 635, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 294, + 689 + ], + "type": "text", + "content": "[39] Van Nguyen Nguyen, Yinlin Hu, Yang Xiao, Mathieu Salzmann, and Vincent Lepetit. Templates for 3d object pose estimation revisited: Generalization to new objects and robustness to occlusions. In CVPR, pages 6771-6780, 2022. 2, 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 691, + 294, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 715 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 715 + ], + "type": "text", + "content": "[40] Van Nguyen Nguyen, Thibault Groueix, Georgy Ponimatkin, Vincent Lepetit, and Tomas Hodan. Cnos: A strong baseline" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "for cad-based novel object segmentation. In ICCVW, pages 2134-2140, 2023. 1, 2, 3, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "type": "text", + "content": "[41] Van Nguyen Nguyen, Thibault Groueix, Mathieu Salzmann, and Vincent Lepetit. Gigapose: Fast and robust novel object pose estimation via one correspondence. In CVPR, pages 9903-9913, 2024. 1, 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 141, + 553, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 553, + 173 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 553, + 173 + ], + "type": "text", + "content": "[42] Brian Okorn, Qiao Gu, Martial Hebert, and David Held. Zephyr: Zero-shot pose hypothesis rating. In ICRA, pages 14141-14148. IEEE, 2021. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 175, + 553, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 553, + 229 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 553, + 229 + ], + "type": "text", + "content": "[43] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 231, + 553, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 231, + 553, + 274 + ], + "spans": [ + { + "bbox": [ + 316, + 231, + 553, + 274 + ], + "type": "text", + "content": "[44] Evin Pinar Örnek, Yann Labbe, Bugra Tekin, Lingni Ma, Cem Keskin, Christian Forster, and Tomáš Hodač. Found- pose: Unseen object pose estimation with foundation features. In ECCV, 2024. 1, 2, 5, 7, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 276, + 553, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 276, + 553, + 308 + ], + "spans": [ + { + "bbox": [ + 316, + 276, + 553, + 308 + ], + "type": "text", + "content": "[45] Kiru Park, Timothy Patten, and Markus Vincze. Pix2pose: Pixel-wise coordinate regression of objects for 6d pose estimation. In ICCV, pages 7668-7677, 2019. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 309, + 553, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 309, + 553, + 342 + ], + "spans": [ + { + "bbox": [ + 316, + 309, + 553, + 342 + ], + "type": "text", + "content": "[46] Sida Peng, Yuan Liu, Qixing Huang, Xiaowei Zhou, and Hujun Bao. Pvnet: Pixel-wise voting network for 6dof pose estimation. In CVPR, 2019. 1, 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 343, + 553, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 343, + 553, + 386 + ], + "spans": [ + { + "bbox": [ + 316, + 343, + 553, + 386 + ], + "type": "text", + "content": "[47] Mahdi Rad and Vincent Lepetit. Bb8: A scalable, accurate, robust to partial occlusion method for predicting the 3d poses of challenging objects without using depth. In ICCV, 2017. 1, 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 388, + 553, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 388, + 553, + 420 + ], + "spans": [ + { + "bbox": [ + 316, + 388, + 553, + 420 + ], + "type": "text", + "content": "[48] René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. In ICCV, pages 12179-12188, 2021. 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 422, + 553, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 422, + 553, + 454 + ], + "spans": [ + { + "bbox": [ + 316, + 422, + 553, + 454 + ], + "type": "text", + "content": "[49] Jerome Revaud, Philippe Weinzaepfel, César Roberto de Souza, and Martin Humenberger. R2D2: repeatable and reliable detector and descriptor. In NeurIPS, 2019. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 456, + 553, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 456, + 553, + 488 + ], + "spans": [ + { + "bbox": [ + 316, + 456, + 553, + 488 + ], + "type": "text", + "content": "[50] Ethan Rublee, Vincent Rabaud, Kurt Konolige, and Gary Bradski. Orb: An efficient alternative to sift or surf. In ICCV, pages 2564-2571, 2011. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 490, + 553, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 490, + 553, + 521 + ], + "spans": [ + { + "bbox": [ + 316, + 490, + 553, + 521 + ], + "type": "text", + "content": "[51] Ivan Shugurov, Fu Li, Benjamin Busam, and Slobodan Ilic. Osop: A multi-stage one shot object pose estimation framework. In CVPR, pages 6835-6844, 2022. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 523, + 553, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 523, + 553, + 555 + ], + "spans": [ + { + "bbox": [ + 316, + 523, + 553, + 555 + ], + "type": "text", + "content": "[52] Jiaming Sun, Zehong Shen, Yuang Wang, Hujun Bao, and Xiaowei Zhou. LoFTR: Detector-free local feature matching with transformers. In CVPR, 2021. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 557, + 553, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 557, + 553, + 600 + ], + "spans": [ + { + "bbox": [ + 316, + 557, + 553, + 600 + ], + "type": "text", + "content": "[53] Jiaming Sun, Zihao Wang, Siyu Zhang, Xingyi He, Hongcheng Zhao, Guofeng Zhang, and Xiaowei Zhou. OnePose: One-shot object pose estimation without CAD models. In CVPR, 2022. 1" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 601, + 553, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 601, + 553, + 634 + ], + "spans": [ + { + "bbox": [ + 316, + 601, + 553, + 634 + ], + "type": "text", + "content": "[54] Bugra Tekin, Sudipta N. Sinha, and Pascal Fua. Real-Time Seamless Single Shot 6D Object Pose Prediction. In CVPR, 2018. 1, 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 635, + 553, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 553, + 669 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 553, + 669 + ], + "type": "text", + "content": "[55] Prune Truong, Martin Danelljan, and Radu Timofte. Glunet: Global-local universal network for dense flow and correspondences. In CVPR, pages 6258-6268, 2020. 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "text", + "content": "[56] Prune Truong, Martin Danelljan, Radu Timofte, and Luc Van Gool. Pdc-net+: Enhanced probabilistic dense correspondence network. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(8):10247-10266, 2023. 2, 5" + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "11631" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 374 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 105 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 105 + ], + "type": "text", + "content": "[57] Michal Tyszkiiewicz, Pascal Fua, and Eduard Trulls. Disk: Learning local features with policy gradient. In NeurIPS, 2020. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 294, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 294, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 294, + 150 + ], + "type": "text", + "content": "[58] Gu Wang, Fabian Manhardt, Federico Tombari, and Xi-angyang Ji. GDR-Net: Geometry-guided direct regression network for monocular 6d object pose estimation. In CVPR, pages 16611-16621, 2021. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 294, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 294, + 195 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 294, + 195 + ], + "type": "text", + "content": "[59] He Wang, Srinath Sridhar, Jingwei Huang, Julien Valentin, Shuran Song, and Leonidas J. Guibas. Normalized object coordinate space for category-level 6d object pose and size estimation. In CVPR, 2019. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 197, + 294, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 197, + 294, + 229 + ], + "spans": [ + { + "bbox": [ + 56, + 197, + 294, + 229 + ], + "type": "text", + "content": "[60] Yifan Wang, Xingyi He, Sida Peng, Dongli Tan, and Xiaowei Zhou. Efficient LoFTR: Semi-dense local feature matching with sparse-like speed. In CVPR, 2024. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 231, + 294, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 231, + 294, + 295 + ], + "spans": [ + { + "bbox": [ + 56, + 231, + 294, + 295 + ], + "type": "text", + "content": "[61] Philippe Weinzaepfel, Thomas Lucas, Vincent Leroy, Yohann Cabon, Vaibhav Arora, Romain Brégier, Gabriela Csurka, Leonid Antsfeld, Boris Chidlovskii, and Jérôme Revaud. Croco v2: Improved cross-view completion pretraining for stereo matching and optical flow. In ICCV, pages 17969-17980, 2023. 6, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 297, + 294, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 297, + 294, + 330 + ], + "spans": [ + { + "bbox": [ + 56, + 297, + 294, + 330 + ], + "type": "text", + "content": "[62] Bowen Wen, Wei Yang, Jan Kautz, and Stan Birchfield. Foundationpose: Unified 6d pose estimation and tracking of novel objects. In CVPR, pages 17868-17879, 2024. 1, 2, 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 331, + 294, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 331, + 294, + 374 + ], + "spans": [ + { + "bbox": [ + 56, + 331, + 294, + 374 + ], + "type": "text", + "content": "[63] Yu Xiang, Tanner Schmidt, Venkatraman Narayanan, and Dieter Fox. Posecnn: A convolutional neural network for 6d object pose estimation in cluttered scenes. In RSS, 2018. 1, 7" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "11632" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/CoA_ Towards Real Image Dehazing via Compression-and-Adaptation/06a0b6a5-a641-41a6-99e5-3f3c8a8c1a70_content_list.json b/2025/CoA_ Towards Real Image Dehazing via Compression-and-Adaptation/06a0b6a5-a641-41a6-99e5-3f3c8a8c1a70_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..7968eef6f45111f96bb09ec833c3ededc131484c --- /dev/null +++ b/2025/CoA_ Towards Real Image Dehazing via Compression-and-Adaptation/06a0b6a5-a641-41a6-99e5-3f3c8a8c1a70_content_list.json @@ -0,0 +1,2223 @@ +[ + { + "type": "text", + "text": "CoA: Towards Real Image Dehazing via Compression-and-Adaptation", + "text_level": 1, + "bbox": [ + 142, + 130, + 854, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Long Ma", + "bbox": [ + 153, + 179, + 238, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yuxin Feng $^{2}$ \nJiang-Yong Chou", + "bbox": [ + 282, + 180, + 385, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yan Zhang", + "bbox": [ + 429, + 180, + 524, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jinyuan Liu1", + "bbox": [ + 571, + 180, + 673, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Weimin Wang", + "bbox": [ + 720, + 180, + 839, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Chengpei $\\mathrm{Xu}^{4,*}$", + "bbox": [ + 478, + 199, + 604, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zhuo Su", + "bbox": [ + 658, + 199, + 736, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Dalian University of Technology $^{2}$ Sun Yat-sen University $^{3}$ Fuzhou University $^{4}$ University of New South Wales malone94319@gmail.com, {fengyx26, zhangy2779} $@$ mail2.sysu.edu.cn, atlantis918@hotmail.com, chengpei.Xu@unsw.edu.au, wangweimin@dlut.edu.cn, gychen@fzu.edu.cn, suzhuo3@mail.sysu.edu.cn", + "bbox": [ + 116, + 217, + 883, + 268 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/bfa2a24e2f256479b04856035ca0b551b569bc05431acec6d9768ac1352674b5.jpg", + "image_caption": [ + "(a) Evaluating efficiency across various metrics" + ], + "image_footnote": [], + "bbox": [ + 102, + 289, + 374, + 405 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/cf30f643d4fe7f77f26fdc98d832ff5cc0fe48bd8885d7eb6b5126dc2f79663f.jpg", + "image_caption": [ + "(b) The computational process of CoA" + ], + "image_footnote": [], + "bbox": [ + 374, + 289, + 598, + 404 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/67d32c0a664118b8f2f44a755cffe18c2ed788d072e7141723b0bcbd931155e6.jpg", + "image_caption": [ + "(c) Evaluating adaptability across different scenes", + "Figure 1. Performance evaluation. The proposed CoA incorporates model compression in synthetic domain for efficiency and bilevel adaptation in real domain for adaptability, as illustrated in the central sub-figure (b). The left sub-figure (a) presents efficiency across various metrics, clearly showing that our CoA outperforms others by a significant margin. The right sub-figure (c) shows adaptability across different scenes, where it is evident that our CoA consistently performs excellently." + ], + "image_footnote": [], + "bbox": [ + 598, + 289, + 898, + 405 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 498, + 326, + 513 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Learning-based image dehazing algorithms have shown remarkable success in synthetic domains. However, real image dehazing is still in suspense due to computational resource constraints and the diversity of real-world scenes. Therefore, there is an urgent need for an algorithm that excels in both efficiency and adaptability to address real image dehazing effectively. This work proposes a Compression-and-Adaptation (CoA) computational flow to tackle these challenges from a divide-and-conquer perspective. First, model compression is performed in the synthetic domain to develop a compact dehazing parameter space, satisfying efficiency demands. Then, a bilevel adaptation in the real domain is introduced to be fearless in unknown real environments by aggregating the synthetic dehazing capabilities during the learning process. Leveraging a succinct design free from additional constraints, our CoA exhibits domain-irrelevant stability and model-agnostic flexibility, effectively bridging the model chasm between synthetic and real domains to further improve its practical utility. Extensive evaluations and analyses underscore the approach's superiority and effectiveness. The code is publicly available at https://github.com/fyxn1/CoA.", + "bbox": [ + 89, + 531, + 483, + 863 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 514, + 498, + 643, + 513 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Image dehazing tasks aim to utilize the atmospheric optical model and advanced computing theory to achieve efficient restoration of scene information and image details through inverse imaging solution or data-driven model learning. It has drawn much attention in multiple emerging computer vision areas recently [46] [27] [28]. Similar to other low-level vision tasks [32] [11], image dehazing technology has evolved from an early stage focused on improving synthetic data metrics to a direction aimed at effectively generalizing across various real-world haze scenarios. Current approaches face dual challenges: efficiency-oriented methods achieve real-time processing but lack scene adaptability, while adaptability-focused techniques dynamically adjust strategies at the cost of high computational complexity that hinders deployment in resource-constrained environments. This paper addresses these limitations through novel architectural innovations in both computational efficiency and adaptive scene understanding.", + "bbox": [ + 511, + 523, + 906, + 796 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1.1. Related Works", + "text_level": 1, + "bbox": [ + 513, + 804, + 665, + 819 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Efficiency-oriented Dehazing Methods. This work focuses on enhancing computational efficiency through optimization algorithms, model simplification, or machine learning methods to enable faster processing under limited resources. Broadly, dehazing methods can be divided", + "bbox": [ + 511, + 825, + 906, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author.", + "bbox": [ + 109, + 888, + 235, + 901 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "11197", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "into image enhancement-based, physics-based, and image layering-based approaches. Image enhancement-based dehazing methods [24] treat the task as an image improvement problem, increasing contrast and color accuracy with techniques like histogram equalization and color correction. Physics-based methods [17], [28] are more specific, leveraging hazy image priors and constraints to aid inverse problem solving. Image layering-based approaches, such as that of Zhang et al. [41], increase efficiency by breaking down the image into scales or layers, progressively refining transmittance maps and enhancing detail using multilevel pyramid structures and dense connectivity. Despite their computational and real-time advantages, they struggle with stability and consistent quality, especially in complex real-world haze conditions, often resulting in partial dehazing and insufficient color clarity.", + "bbox": [ + 89, + 90, + 483, + 332 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Adaptability-focused Dehazing Methods. Adaptive dehazing methods tailor their strategies based on specific image attributes to improve dehazing effectiveness. These methods encompass three main types: local feature-based, atmospheric scattering model-based, and domain adaptation-based approaches. Local feature-based methods [15] adaptively adjust dehazing intensity according to local contrast and brightness variations, which helps restore image details effectively. Atmospheric scattering model-based adaptive dehazing methods [20] dynamically adjust the transmittance estimation by analyzing different regions of the image to achieve more accurate restoration. These methods can, to some extent, balance both efficiency and adaptability. Domain adaptation-based dehazing methods [26, 37] have gradually emerged in the field of image dehazing in recent years. These methods learn a shared feature representation, enabling the model to simultaneously adapt to the features of both the source domain and the target domain. Although these methods yield superior dehazing results, their complex models and dynamic imagespecific adjustments often increase computational costs, potentially impacting real-time applicability.", + "bbox": [ + 91, + 333, + 483, + 667 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1.2. Our Contributions", + "text_level": 1, + "bbox": [ + 89, + 679, + 272, + 695 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address the challenges of efficiency and adaptability in real image dehazing, this work proposes a new Compression-and-Adaptation (CoA) scheme as shown in Fig. 1 (b). The predefined model is compressed in the synthetic domain to ensure computational efficiency (refer to Fig. 1 (a)). Further, a bilevel adaptive learning is developed to leverage the dehazing capability acquired in the synthetic domain to improve adaptability (see Fig. 1 (c)) in the real domain. Our key contributions are summarized as", + "bbox": [ + 89, + 702, + 482, + 838 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- Following the divide-and-conquer paradigm, we propose a new CoA learning strategy that first compresses the predefined model and then adapts it to unseen scenes, offering an efficient solution for real image dehazing. To the", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "best of our knowledge, this is the first work to successfully integrate both efficiency and adaptability, achieving the best of both worlds in real image dehazing.", + "bbox": [ + 524, + 90, + 903, + 136 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- To adapt to unlabeled, diverse real-world scenes, we propose a new cross-domain bilevel model designed to learn dehazing parameters for the real domain constrained by synthetic dehazing learning. Additionally, we derive a bilevel adaptive learning scheme that effectively harnesses the synthetic dehazing capacity, preventing instability and enhancing adaptability.", + "- Benefiting from succinct design free from additional constraints, CoA shows two key properties on three standard real benchmarks. It shows stability across synthetic domains, improving $14.2\\%$ on average across four metrics (evaluated on three domains). CoA also offers flexibility with dehazing models, reducing computational cost by at least $73\\%$ in parameters (tested on three methods)." + ], + "bbox": [ + 511, + 136, + 906, + 347 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. The Proposed Method", + "text_level": 1, + "bbox": [ + 511, + 358, + 723, + 375 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this section, we first rethinking real image dehazing, highlighting two key practical needs. We then propose model compression and bilevel adaptation to meet these needs, followed by an overview of the network architecture and training loss.", + "bbox": [ + 511, + 383, + 905, + 460 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Rethinking Real Image Dehazing", + "text_level": 1, + "bbox": [ + 511, + 465, + 803, + 483 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Despite the strong performance of advanced image dehazing techniques [31, 36, 44] on synthetic datasets [8, 21, 30], these datasets often rely on simplifying assumptions, resulting in a substantial gap between synthetic and real-world domains. Addressing the core challenges in real-world image dehazing involves two key aspects:", + "bbox": [ + 511, + 488, + 905, + 578 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Efficiency: Edge devices often have limited computational resources compared to systems with ample capacity, requiring dehazing models to be highly efficient.", + "- Adaptability: Real-world imaging conditions vary dynamically. Dehazing models must therefore be adaptable to maintain reliable performance." + ], + "bbox": [ + 513, + 579, + 903, + 669 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The latter is more challenging, relying on effective learning mechanisms like unsupervised learning [12]. However, current methods struggle with unsupervised learning due to difficulty in capturing statistical regularities, making transferring dehazing capabilities from the synthetic to real domain a viable alternative. To provide an intuitive presentation of the goal upon the above demands, we build", + "bbox": [ + 511, + 670, + 905, + 775 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\min _ {\\boldsymbol {\\theta} _ {\\mathrm {r e a l}}} f \\Big (\\boldsymbol {\\theta} _ {\\mathrm {r e a l}} \\left(\\boldsymbol {\\theta} _ {\\mathrm {s y n}}\\right) \\Big), \\\\ s. t., \\left\\{ \\begin{array}{l} \\kappa (\\boldsymbol {\\theta} _ {\\text {r e a l}}) < \\kappa (\\boldsymbol {\\theta} _ {\\text {s y n}}), (\\text {e f f i c i e n c y}) \\\\ \\zeta (\\boldsymbol {\\theta} _ {\\text {r e a l}}) > \\zeta (\\boldsymbol {\\theta} _ {\\text {s y n}}), (\\text {a d a p t a b i l i t y}) \\end{array} \\right. \\tag {1} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 560, + 781, + 903, + 849 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "where $\\theta_{\\mathrm{syn}}$ and $\\theta_{\\mathrm{real}}$ represent parameters in synthetic and real domain, respectively. The function $\\kappa (\\cdot)$ and $\\zeta (\\cdot)$ denote the evaluator for efficiency and adaptability, respectively.", + "bbox": [ + 511, + 854, + 905, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "11198", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/9bd811b0ce8fe713b7738e967a4989bb964291c9f9c69583f440ffc245559744.jpg", + "image_caption": [ + "Figure 2. Qualitative comparisons of algorithmic properties. Arrows indicate specific regions that highlight visible differences." + ], + "image_footnote": [], + "bbox": [ + 102, + 90, + 893, + 215 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To achieve this, we propose a divide-and-conquer strategy: the first phase focuses on model compression for efficiency, while the second phase employs bilevel adaptation for adaptability. Details are provided below.", + "bbox": [ + 89, + 261, + 483, + 323 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. MoC: Model Compression in Synthetic Domain", + "text_level": 1, + "bbox": [ + 89, + 330, + 483, + 347 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The main challenge in model compression is transferring dehazing capabilities from a large-scale parameter space to a smaller one. Here we propose a composite loss function with context alignment to realize it.", + "bbox": [ + 89, + 353, + 483, + 412 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\boldsymbol {\\theta} _ {\\mathrm {s y n}} ^ {\\mathrm {s}}} \\mathcal {L} _ {\\mathrm {s y n}} \\left(\\boldsymbol {\\theta} _ {\\mathrm {s y n}} ^ {\\mathrm {s}}\\right) + \\mathcal {L} _ {\\mathrm {a}} \\left(\\boldsymbol {\\theta} _ {\\mathrm {s y n}} ^ {\\mathrm {s}}, \\boldsymbol {\\theta} _ {\\mathrm {s y n}} ^ {\\mathrm {t}}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 171, + 422, + 482, + 449 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\theta_{\\mathrm{syn}}^{\\mathrm{t}}$ and $\\theta_{\\mathrm{syn}}^{\\mathrm{s}}$ are the pre-trained teacher model and desired student model in the synthetic domain, respectively. The function $\\mathcal{L}_{\\mathrm{syn}}$ and $\\mathcal{L}_{\\mathrm{a}}$ represent a group of supervised losses and the context alignment loss, respectively.", + "bbox": [ + 89, + 458, + 483, + 518 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "During model compression, the student model follows the teacher model's training strategy on paired synthetic data. To maximize the retention of the teacher model's capabilities, we introduce a context alignment loss $(\\mathcal{L}_{\\mathrm{a}})$ that aligns the outputs of the teacher and student encoders at corresponding positions. By weighting the loss functions across layers, this approach enables fine-grained feature alignment, improving knowledge transfer and model optimization. The layered alignment ensures both global feature distributions and layer-specific details are accurately matched, resulting in more precise compression of the teacher model into the student model.", + "bbox": [ + 89, + 520, + 483, + 700 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. BiA:Bilevel Adaptation to Real Domain", + "text_level": 1, + "bbox": [ + 89, + 709, + 429, + 724 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "After compressing the model for efficiency, we address adaptation to real-world scenes by proposing a cross-domain bilevel model. This is followed by a bilevel adaptive learning scheme to enable smooth transition from the synthetic to real domain.", + "bbox": [ + 89, + 731, + 483, + 806 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3.1. Cross-Domain Bilevel Modeling", + "text_level": 1, + "bbox": [ + 89, + 814, + 357, + 829 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The gap between the synthetic and real domains makes adaptation challenging. The dehazing parameters in the synthetic and real domains $(\\theta_{\\mathrm{rea}}^s)$ have a nested subordinate", + "bbox": [ + 89, + 832, + 483, + 878 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Algorithm 1 Learning via Compression-and-Adaptation", + "text_level": 1, + "bbox": [ + 514, + 261, + 885, + 276 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Require: The pre-trained teacher model $\\theta_{\\mathrm{syn}}^{\\mathrm{t}}$ , the initial parameter $\\theta_{\\mathrm{syn}}^{\\mathrm{s}}$ , step-sizes $\\eta_{\\mathrm{syn}}^{m}$ , $\\eta_{\\mathrm{syn}}^{b}$ , coefficient $\\alpha$ , iteration numbers $N$ in MoC phase and $T$ in BiA phase.", + "bbox": [ + 517, + 281, + 905, + 328 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Ensure: The optimal parameters $\\theta_{\\mathrm{rea}}^{\\mathrm{s}}$", + "bbox": [ + 517, + 329, + 776, + 345 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1: % The MoC Phase", + "2: for $n = 0:N - 1$ do", + "3: Calculate the gradient $g$ of Eq. (2).", + "4: $\\pmb{\\theta}_{\\mathrm{syn}}^{\\mathrm{s}}(n + 1) = \\pmb{\\theta}_{\\mathrm{syn}}^{\\mathrm{s}}(n) - \\eta_{\\mathrm{syn}}^{m}\\pmb{g}(\\pmb{\\theta}_{\\mathrm{syn}}^{\\mathrm{s}},n),\\pmb{\\theta}_{\\mathrm{syn}}^{\\mathrm{t}}).$", + "5: end for", + "6: Initialize the dehazing model's parameter in the real domain by $\\pmb{\\theta}_{\\mathrm{rea}}^{\\mathrm{s}} = \\pmb{\\theta}_{\\mathrm{syn}}^{\\mathrm{s}}(N)$ .", + "7: % The BiA Phase", + "8: for $t = 0 : T - 1$ do", + "9: $\\%$ Update the lower-level parameter $\\theta_{\\mathrm{syn}}^{\\mathrm{s}}$ in $\\mathcal{D}_{\\mathrm{rea}}$ .", + "10: $\\pmb{\\theta}_{\\mathrm{syn}}^{\\mathrm{s}}(t + 1) = \\pmb{\\theta}_{\\mathrm{syn}}^{\\mathrm{s}}(t) - \\eta_{\\mathrm{syn}}^{b}\\frac{\\partial(\\mathcal{L}_{\\mathrm{rea}} + \\mathcal{L}_{1})}{\\partial\\pmb{\\theta}_{\\mathrm{syn}}^{\\mathrm{s}}}$ .", + "11: $\\%$ Update the upper-level parameter $\\theta_{\\mathrm{rea}}^{\\mathrm{s}}$ by EMA.", + "12: $\\pmb{\\theta}_{\\mathrm{rea}}^{\\mathrm{s}}(t + 1) = \\alpha \\pmb{\\theta}_{\\mathrm{rea}}^{\\mathrm{s}}(t) + (1 - \\alpha)\\pmb{\\theta}_{\\mathrm{syn}}^{\\mathrm{s}}(t + 1).$", + "13: end for", + "14: return $\\theta_{\\mathrm{rea}}^{\\mathrm{s}}(T)$" + ], + "bbox": [ + 519, + 345, + 903, + 578 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "relationship. This coupling can be effectively modeled using bilevel programming techniques [9, 29]. From the perspective of hyperparameter optimization, our cross-domain bilevel model is written as:", + "bbox": [ + 511, + 606, + 905, + 666 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\boldsymbol {\\theta} _ {\\mathrm {r e a}} ^ {\\mathrm {s}}} \\mathcal {L} _ {\\mathrm {r e a}} \\big (\\boldsymbol {\\theta} _ {\\mathrm {r e a}} ^ {\\mathrm {s}}, \\boldsymbol {\\theta} _ {\\mathrm {s y n}} ^ {\\mathrm {s}} \\big (\\boldsymbol {\\theta} _ {\\mathrm {r e a}} ^ {\\mathrm {s}} \\big); \\mathcal {D} _ {\\mathrm {r e a}} \\big),\n$$\n", + "text_format": "latex", + "bbox": [ + 571, + 672, + 808, + 698 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\ns. t., \\theta_ {\\text {s y n}} ^ {\\mathrm {s}} \\left(\\theta_ {\\text {r e a}} ^ {\\mathrm {s}}\\right) \\in \\arg \\min _ {\\theta_ {\\text {s y n}} ^ {\\mathrm {s}}} \\Psi \\left(\\theta_ {\\text {r e a}} ^ {\\mathrm {s}}, \\theta_ {\\text {s y n}} ^ {\\mathrm {s}}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 571, + 694, + 903, + 727 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where", + "bbox": [ + 513, + 736, + 558, + 750 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\Psi \\left(\\boldsymbol {\\theta} _ {\\mathrm {r e a}} ^ {\\mathrm {s}}, \\boldsymbol {\\theta} _ {\\mathrm {s y n}} ^ {\\mathrm {s}}\\right) = \\mathcal {L} _ {\\mathrm {r e a}} \\left(\\boldsymbol {\\theta} _ {\\mathrm {s y n}} ^ {\\mathrm {s}}; \\mathcal {D} _ {\\mathrm {r e a}}\\right) + \\mathcal {L} _ {1} \\left(\\boldsymbol {\\theta} _ {\\mathrm {s y n}} ^ {\\mathrm {s}}, \\boldsymbol {\\theta} _ {\\mathrm {r e a}} ^ {\\mathrm {s}}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 527, + 758, + 903, + 779 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_{\\mathrm{rea}}$ and $\\mathcal{L}_1$ represent the loss constraints for real domain and alignment. The datasets $\\mathcal{D}_{\\mathrm{rea}}$ and $\\mathcal{D}_{\\mathrm{syn}}$ are captured in the real and synthetic domains, respectively.", + "bbox": [ + 511, + 787, + 905, + 833 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Actually, suppose an effective loss $\\mathcal{L}_{\\mathrm{rea}}$ could be designed to constrain the output in the real domain, the adaptation might be achieved solely through the upper-level model", + "bbox": [ + 511, + 834, + 905, + 878 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "1Details of the specific loss functions can be found in Sec. 2.4.", + "bbox": [ + 107, + 886, + 439, + 898 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "2Their specific forms can be found in Sec. 2.4.", + "bbox": [ + 529, + 886, + 777, + 898 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "11199", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/a77dd7427336c634b727164766cb68fe65fbacc85c73507f6dba7f903f300046.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Testing BenchmarkMetricsStability across various synthetic domainFlexibility with different dehazing model
RESIDEHaze4KTHazeMSBDNDehazeFormerDEA
BaseCoABaseCoABaseCoABaseCoABaseCoABaseCoA
RTTS [22]FADE↓1.47531.26871.73681.08501.20780.85941.33541.15561.18660.93251.21170.9332
PM2.5↓163.77139.04170.03110.84131.7788.871135.6895.793126.77115.00121.9889.813
Entropy↑7.29637.34587.34467.46377.43197.57987.39687.53827.48157.55117.46237.5426
BIQME↑0.55620.56390.55110.57480.57730.59320.56960.58190.57330.59280.57890.5860
URHI [22]FADE↓1.44631.30131.72631.02781.31460.92721.34880.95491.32631.17261.27961.0321
PM2.5↓164.36145.01172.98119.97140.3598.793146.4399.856141.11125.87133.97108.94
Entropy↑7.28677.31527.37117.46577.46527.59287.41447.55747.50097.58177.49207.5658
BIQME↑0.54990.55420.54810.56810.58210.59610.57270.58320.57570.59480.57860.5863
FATTAL [6]FADE↓0.51360.47780.58280.37340.42030.31420.43150.33270.42710.38750.42700.3976
PM2.5↓81.73880.28981.401110.9762.31291.63969.272104.5678.09360.69381.38780.723
Entropy↑7.38477.45757.39997.50067.44827.58587.41217.54607.45667.56107.43157.5569
BIQME↑0.56450.58020.55540.59190.59450.61890.58070.60550.59270.61170.59030.6097
", + "bbox": [ + 106, + 88, + 890, + 309 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Table 1. Quantitative comparison of algorithmic properties. In the left portion, our designed architecture serves as the baseline teacher model. In the right portion, the THaze dataset is used as the synthetic domain.", + "bbox": [ + 89, + 311, + 906, + 340 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "(initialized with $\\theta_{\\mathrm{syn}}^{\\mathrm{s}}$ ). However, as discussed in Sec. 2.1, the challenge of accurately modeling real haze distributions makes this process difficult to implement. To this end, we introduce lower-level sub-optimization to ensure a feasible solution and maintain the correct optimization trajectory. Experimental validation can be seen in Sec. 5.2.", + "bbox": [ + 89, + 352, + 483, + 444 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.3.2. Bilevel Adaptive Learning", + "text_level": 1, + "bbox": [ + 89, + 450, + 320, + 465 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Without compromising the model's dehazing capability in the synthetic domain, narrowing the performance gap between the synthetic and real domains is a central focus. Inspired by the Exponential Moving Average (EMA) method [2], we draw on its smoothing characteristics during the training process and introduce a dynamic weight adjustment mechanism to achieve a balance in learning between the synthetic and real domains, formulated as", + "bbox": [ + 89, + 468, + 483, + 590 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{ \\begin{array}{l} \\boldsymbol {\\theta} _ {\\text {s y n}} ^ {\\mathrm {s}} (t + 1) = \\boldsymbol {\\theta} _ {\\text {s y n}} ^ {\\mathrm {s}} (t) - \\eta_ {\\text {s y n}} ^ {b} \\frac {\\partial \\left(\\mathcal {L} _ {\\text {r e a}} + \\mathcal {L} _ {1}\\right)}{\\partial \\boldsymbol {\\theta} _ {\\text {s y n}} ^ {\\mathrm {s}}}, \\\\ \\boldsymbol {\\theta} _ {\\text {r e a}} ^ {\\mathrm {s}} (t + 1) = \\alpha \\boldsymbol {\\theta} _ {\\text {r e a}} ^ {\\mathrm {s}} (t) + (1 - \\alpha) \\boldsymbol {\\theta} _ {\\text {s y n}} ^ {\\mathrm {s}} (t + 1), \\end{array} \\right. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 114, + 598, + 483, + 655 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\alpha$ represents the moving average coefficient, and $\\mathcal{L}_1$ denotes the $\\ell_1$ -norm supervised loss used to reduce the risk of the model overfitting to the real domain. Specifically, during each update, $\\theta_{\\mathrm{rea}}^{\\mathrm{s}}$ is adjusted by $\\theta_{\\mathrm{syn}}^{\\mathrm{s}}$ , allowing the model to strike a balance between the synthetic and real domains throughout the learning process. The overall algorithm can be seen in Alg. 1.", + "bbox": [ + 89, + 662, + 483, + 770 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.4. Network Architecture and Training Loss", + "text_level": 1, + "bbox": [ + 89, + 777, + 439, + 792 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This section introduces the details about network architecture and training loss adopted in this work.", + "bbox": [ + 89, + 799, + 482, + 830 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.4.1. Network Architecture", + "text_level": 1, + "bbox": [ + 89, + 835, + 290, + 849 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We emphasize that our primary focus is on developing an effective learning mechanism to enhance the model's adaptability to the real domain. Notably, our approach offers", + "bbox": [ + 89, + 854, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "flexibility3 with different dehazing models. To this end, we adopt a hybrid architectural design, incorporating a pretrained Res2Net encoder [10] with the decoder of multiscale boosted dehazing network [5].", + "bbox": [ + 511, + 352, + 905, + 414 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.4.2. Loss Functions in MoC Phase", + "text_level": 1, + "bbox": [ + 511, + 421, + 766, + 435 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the MoC process, to enable the student model to learn the features extracted by the teacher model more effectively and to enhance the quality the generated images. We define", + "bbox": [ + 511, + 441, + 905, + 487 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {s y n}} = \\lambda_ {s u} \\ell_ {s u} + \\lambda_ {s s} \\ell_ {s s} + \\lambda_ {p e} \\ell_ {p e}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 591, + 500, + 906, + 516 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\ell_{su},\\ell_{ss},\\ell_{pe}$ denote $\\ell_1$ -norm supervised loss, SSIM loss, and perceptual loss, respectively. The coefficient $\\lambda_{su},\\lambda_{ss},\\lambda_{pe}$ are positive balancing parameters.", + "bbox": [ + 511, + 527, + 905, + 575 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To more effectively enhance consistency between the student and teacher models in the feature space, the design of $\\mathcal{L}_{\\mathrm{a}}$ incorporates a similarity measure of feature distributions. Its formulation can be written as follows:", + "bbox": [ + 511, + 575, + 905, + 633 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {a}} = \\sum_ {i = 0} ^ {L - 1} w _ {i} \\cdot \\left(\\left(T _ {i} - S _ {i}\\right) ^ {2}\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 612, + 643, + 906, + 685 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $w_{i}$ represents the weights of $i$ -th layer in the teacher feature map $T_{i}$ and the student feature map $S_{i}$ .", + "bbox": [ + 511, + 696, + 905, + 727 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.4.3. Loss Functions in BiA Phase", + "text_level": 1, + "bbox": [ + 511, + 734, + 758, + 750 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here we incorporate the CLIP model to construct the training loss $\\mathcal{L}_{\\mathrm{rea}}$ to constrain the output in the real domain. We select haze images $I_{\\mathcal{H}}$ and clear images $I_{\\mathcal{C}}$ as references, initialize corresponding prompts $T_{\\mathcal{H}}$ and $T_{\\mathcal{C}}$ , and input them into the CLIP text encoder, while feeding the images into the image encoder. The sample categories are predicted using the text-image similarity in the CLIP space. To minimize the classification error between positive and negative", + "bbox": [ + 511, + 755, + 906, + 876 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "3For further details, please refer to Sec. 3.", + "bbox": [ + 529, + 886, + 753, + 898 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "11200", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/2081aea959a9ec7a42004fca43e2baedc424f704352c01f25fdce19a114dbb45.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 107, + 88, + 488, + 205 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/406dd7eb2f328f1e40df05ccea96ebd025e8459fb0ac9c6ba3d84e976a4acd49.jpg", + "image_caption": [ + "Figure 3. Quantitative comparison on three real-world datasets. Four no-reference image quality assessments were calculated." + ], + "image_footnote": [], + "bbox": [ + 107, + 210, + 488, + 329 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/c5ad0c68ced8b4d7419a2ca3c17f3d729871e68bef62a575ae48fb4fdf736a35.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 87, + 887, + 205 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/5b78fef2c6ae07a68ace097c64d2dfa157f6d39b79d1b314eb87e32e6e0c3ea2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 210, + 887, + 329 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/aea6bcee39fdf4c6cae19a7ab6caba61f1b5a7e8d3a342116e57f4a544a50f42.jpg", + "image_caption": [ + "Figure 4. Qualitative comparisons on daytime haze and dusty scenes. All these observations come from RTTS and URHI datasets." + ], + "image_footnote": [], + "bbox": [ + 102, + 358, + 898, + 500 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "samples, we employ the binary cross entropy loss to optimize the text prompt pair. After training the text prompts to effectively distinguish between real haze and clear images, we can derive $\\mathcal{L}_{\\mathrm{rea}}$ based on the contrastive similarity loss:", + "bbox": [ + 89, + 539, + 482, + 599 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {r e a}} = \\frac {e ^ {\\cos \\left(\\Phi_ {\\text {i m a g e}} \\left(I _ {\\mathcal {R}}\\right) , \\Phi_ {\\text {t e x t}} \\left(T _ {\\mathcal {H}}\\right)\\right)}}{\\sum_ {i \\in \\{\\mathcal {H} , \\mathcal {C} \\}} e ^ {\\cos \\left(\\Phi_ {\\text {i m a g e}} \\left(I _ {\\mathcal {R}}\\right) , \\Phi_ {\\text {t e x t}} \\left(T _ {i}\\right)\\right)}}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 133, + 611, + 482, + 661 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $I_{\\mathcal{R}}$ corresponds to the real haze image processed by the model obtained through the model compression process.", + "bbox": [ + 89, + 672, + 482, + 703 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3. Exploring Algorithmic Property", + "text_level": 1, + "bbox": [ + 89, + 734, + 383, + 752 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "CoA presents an effective learning strategy to improve efficiency and adaptability for real-world scenes, without assumptions on the synthetic domain or teacher model. We analyze CoA's properties, highlighting its stability across synthetic domains and flexibility with different models.", + "bbox": [ + 89, + 758, + 483, + 835 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1. Stability across Various Synthetic Domains", + "text_level": 1, + "bbox": [ + 89, + 847, + 457, + 864 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We perform three synthetic benchmarks (RESIDE [21] (the outdoor subset), Haze4K [30], and THaze [8]) and", + "bbox": [ + 89, + 869, + 482, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "three real-world datasets (RTTS [21], URHI [21], and FAT-TAL [6]) to assess our stability. As shown in the left section of Table 1, CoA consistently improves scores across four no-reference image quality metrics. Notably, due to the diverse scenes in THaze, CoA achieves near-optimal performance within the synthetic domain. We use THaze as the synthetic domain in our comparative experiments. Fig. 2 (left) compares visual results. The dehazing effects of CoA is significant, especially in distant regions highlighted by arrows. Overall, CoA improves the original teacher model across synthetic domains, demonstrating its stability.", + "bbox": [ + 511, + 537, + 906, + 705 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2. Flexibility with Different Dehazing Models", + "text_level": 1, + "bbox": [ + 511, + 712, + 877, + 729 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "CoA's teacher model-irrelevant nature allows us to enhance the adaptability of existing dehazing models across various scenes. We applied CoA to three dehazing models (MSBDN [5], DehazeFormer [34], and DEA [3]) as teacher models. The right section of Table 1 shows significant performance improvements before and after applying CoA across different models and test settings. Fig. 2 (right) compares visual results using DehazeFormer and DEA, highlighting areas where CoA enhances dehazing for more visually appealing outcomes. In summary, CoA demonstrates notable flexibility with various dehazing models. The intro", + "bbox": [ + 511, + 734, + 906, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "11201", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/834400768a84b4bb0a3db0d0017dcf9c98d7255ce0716f77731805c37479cac5.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodRTTSURHIFATTAL
FADE↓PM2.5↓Entropy↑BIQME↑FADE↓PM2.5↓Entropy↑BIQME↑FADE↓PM2.5↓Entropy↑BIQME↑
AirNet1.3342155.847.29590.54711.6591143.026.91430.52580.406376.1147.35450.5315
WeatherDiff2.4102193.277.14430.51552.1429180.267.14190.51020.8749118.797.36210.4937
DiffUIR2.1305189.617.17220.53031.9879180.567.18110.52050.8405116.687.36090.5049
Ours0.859488.8717.57980.59320.927298.7937.59280.59610.314291.6397.58580.6189
", + "bbox": [ + 99, + 88, + 895, + 179 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Quantitative comparison on daytime haze scenes. All compared methods are designed for multi-weather image restoration.", + "bbox": [ + 101, + 181, + 893, + 198 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/30411f15b0370bc90d65016d9c1b5c44442ddd87a457233362cb5ff6dd4b5c00.jpg", + "image_caption": [ + "Figure 5. Qualitative comparisons corresponding to Table 2. All examples are sourced from RTTS and URHI datasets." + ], + "image_footnote": [], + "bbox": [ + 102, + 210, + 220, + 328 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a70559dd05378698edbd70ac1e86f036a0145c937eeb9828fd3de22e923e6cdd.jpg", + "image_caption": [ + "Figure 6. Performance evaluation on nighttime haze scenes." + ], + "image_footnote": [], + "bbox": [ + 222, + 210, + 300, + 328 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/19e662b11062569afea83a8edb90457032053c125f2f4898ec0f42baadf153fd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 302, + 210, + 419, + 328 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/3035cc059c694102b63d5d70d74235479e6ba43fc2097af432ed5bc25ce5d557.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 421, + 210, + 500, + 328 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5fcfec06c607a62685a9b3172fef701a4dd0ed5ff6b6a9f4f2c0e07eaaac1f93.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 210, + 617, + 328 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/81a7dfa115e4fa3f75b6a7a4a702ec04bf732c6603e099aa29aea1879f57fb36.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 620, + 210, + 697, + 328 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/97ce5c805f5f9ec1948ca002453019e8069debd24f09b3884649d92842f1f4f0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 697, + 210, + 816, + 328 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d252059606f4b48f4dd313da11ac0a31d2070d68cff07a4939cf038b2e0818a1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 818, + 210, + 897, + 328 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/7f322d82690e90b3e3eac3628041ad5f6ffb57b1c2328b42c6b09a9c2bf369f5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 102, + 366, + 214, + 494 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/78e6a40c771cc43fe8cd5a56c2fb732bf409eb96dfed9b90b5478faed2a0a1b5.jpg", + "image_caption": [ + "Figure 7. Performance evaluation on underwater scenes." + ], + "image_footnote": [], + "bbox": [ + 215, + 366, + 316, + 494 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/cd1b00f5a4f8fe4360c66e558947ef230eea6139eed5716eb06f6adc06b9c708.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 316, + 366, + 418, + 494 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e99ff60e553377d8a29086f4983bb9e9648adfe096697ebacf22c9abd9cf72b1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 419, + 366, + 472, + 493 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/6dfd92c63faa6fa5534baf87fc3c6cb25b67cd02d14a6a56fe37d91ae4174197.jpg", + "table_caption": [ + "(a) Qualitative comparison" + ], + "table_footnote": [], + "table_body": "
MetricsGS (ICCV'15)MRP (CVPR'17)OSFD (ACM MM'20)GAPSF (ACM MM'23)CoA
PM2.5↓102.43106.20146.93110.5981.141
Entropy↑6.70257.04886.94126.37696.9907
BIQME↑0.45520.47220.46840.38950.5112
", + "bbox": [ + 99, + 508, + 470, + 578 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "(b) Quantitative comparison", + "bbox": [ + 210, + 579, + 359, + 590 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "duction of CoA significantly reduces the model's parameter count, improving computational efficiency. Specifically, the parameters in the three dehazing models decrease by $94.06\\%$ , $86.74\\%$ , and $73.89\\%$ , respectively. This reduction accelerates inference speed and lowers resource requirements, making the models more lightweight and better suited for practical applications.", + "bbox": [ + 89, + 628, + 483, + 734 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experimental Results", + "text_level": 1, + "bbox": [ + 89, + 748, + 294, + 763 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Implementation Details", + "text_level": 1, + "bbox": [ + 89, + 772, + 305, + 787 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Training settings. The CoA model was implemented using the PyTorch framework on a single NVIDIA RTX 3090 GPU. The Adam optimizer was employed to update each module's parameters, with $\\beta_{1},\\beta_{2}$ , and $\\varepsilon$ set to 0.9, 0.999, and $1e^{-8}$ , respectively. The initial learning rate was set to $1e^{-4}$ and gradually decayed to $1e^{-6}$ using a cosine annealing schedule. The algorithm was trained on RGB channels", + "bbox": [ + 89, + 794, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/fa221ef9745c88ba0e7b0ead84dd29ab5c15f46e99ae8dbbcf2a945d85ea5f28.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 524, + 366, + 599, + 508 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/afc883fc039d3cde309afa825c065fa6982c8f1be04068a76e65b01db0d3a52d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 599, + 366, + 673, + 507 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/23289f9a394973120e1a00b487f3675d2feb399dcd4a4b7d31b0bf4afc39081b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 674, + 366, + 746, + 508 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e3d84f23a41c6053066797c2c1542c3d6bb79843dea24c776536fa9c12243988.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 750, + 366, + 821, + 507 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d97f6bbe3dae7d93afaa4b1f98f0db494772f139c2fe14a6c6028233d5261e35.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 821, + 366, + 895, + 507 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/b92fb11db9d447991bb9d6ed472379a836075183bcc54a383523d29b1463ee21.jpg", + "table_caption": [ + "(a) Qualitative comparison" + ], + "table_footnote": [], + "table_body": "
MetricsDM(ACM MM'23)SUIR(CVPR'23)SMDR(AAAI'24)CoA
UIQM↑4.09653.98534.03144.1284
CCF↑26.65131.41728.43930.878
", + "bbox": [ + 524, + 523, + 890, + 579 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "(b) Quantitative comparison", + "bbox": [ + 633, + 579, + 782, + 590 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "with data augmentation via random $90^{\\circ}$ , $180^{\\circ}$ , and $270^{\\circ}$ rotations and horizontal flipping. We cropped random $256 \\times 256$ sub-regions from images, further expanding the training set. Using a pre-trained Res2Net encoder, the teacher model achieved strong performance after just 20 epochs.", + "bbox": [ + 511, + 628, + 903, + 702 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Benchmarks and metrics. Three synthetic datasets (Haze4K, RESIDE (the outdoor subset), and THaze served as the synthetic domain, while three real-world datasets (RTTS, Fattal, and URHI were used for testing in the real domain. We employed four objective metrics for realworld haze scenes: FADE [4], PM2.5 [14], Entropy, and BIQME [13]. For underwater enhancement, we used UIQM and CCF to assess image quality.", + "bbox": [ + 511, + 704, + 905, + 824 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Compared methods. We evaluated our CoA method against state-of-the-art techniques across daytime, dusty, nighttime, and underwater conditions. For daytime and dusty conditions, we compared CoA with SGID [1], Dehamer [16], C2P [46], KANet [7], DEA, D4 [38], $\\mathrm{D4^{+}}$ [39],", + "bbox": [ + 511, + 825, + 905, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "11202", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/74cc649886a45a5395e15768aae5935d3f614924104c3575b3acb060378c9a0f.jpg", + "table_caption": [], + "table_footnote": [ + "This work requires excessive video memory for high-resolution images, resulting in the failure to produce results." + ], + "table_body": "
MetricsSGID (TIP'22)Dehamer (CVPR'22)C2P (CVPR'23)\\( \\text{RIDCP}^\\dagger \\)(CVPR'23)D4 (CVPR'22)D4+ (IJCV'24)KANet (TPAMI'24)DEA (TIP'24)Ours
SIZE (M)13.87132.407.1728.7210.7010.7055.253.651.69
FLOPs (G)108.4048.91352.90144.432.822.824.4232.202.67
Time (ms)1280×720240.01136.361169.18630.6451.2551.2522.9386.3324.33
1920×1080878.94295.982531.431588.26105.31105.31205.40190.7852.52
2560×14401514.82519.524409.12—185.55185.5510631334.7892.13
", + "bbox": [ + 122, + 88, + 874, + 200 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/db2bef2bd5bf7c8402e3bf843f0fd406367d66c8e355bd294db11a6e195922cd.jpg", + "image_caption": [ + "(a) Qualitative comparison" + ], + "image_footnote": [], + "bbox": [ + 112, + 247, + 460, + 353 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/43d08bb725f84c18288e156bcb0f0155ec81130d9009550512e591161b4c9e63.jpg", + "table_caption": [ + "Table 3. Evaluating the computational efficiency. Note that the size of the testing image for FLOPs calculation is ${224} \\times {224}$ ." + ], + "table_footnote": [], + "table_body": "
ModelFADE↓PM2.5↓BIQME↑SIZE (M)FLOPs (G)
Naive for Teacher1.21131.770.5852.8318.56
Naive for Student1.23135.760.571.692.67
MoC for Student1.13124.430.581.692.67
", + "bbox": [ + 112, + 369, + 457, + 421 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "and RIDCP [37]. Additionally, we assessed CoA against multi-weather restoration models: AirNet [23], WeatherDiff [33], and DiffUIR [45]. For nighttime dehazing, we benchmarked against GS [25], MRP [42], OSFD [43], and GAPSF [19], and for underwater scenes, we compared with DM [35], SUIR [18], and SMDR [40].", + "bbox": [ + 89, + 470, + 483, + 561 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Evaluation on Regular Scenes", + "text_level": 1, + "bbox": [ + 89, + 574, + 356, + 589 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Quantitative comparisons. In Fig. 3, we presented a comparison of eight state-of-the-art dehazing algorithms with our proposed method, using four authoritative unsupervised evaluation metrics. It is evident that our method ranked among the top in almost all metrics, achieving optimal or near-optimal performance in most cases.", + "bbox": [ + 89, + 597, + 483, + 686 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Qualitative comparisons. As shown in Fig. 4, the supervised methods (SGID, C2P, Dehamer, DEA) and unsupervised methods $(\\mathrm{D}4,\\mathrm{D}4^{+})$ struggle with haze removal and generalization. RIDCP causes color distortion and overbehazing, especially in sandstorm and colorful haze scenes. KANet has limitations in detail and color restoration. In contrast, CoA excels at haze removal while preserving fine textures, resulting in more natural and realistic images.", + "bbox": [ + 89, + 688, + 483, + 809 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparisons with multi-weather restorers. Here we compared CoA with state-of-the-art multi-weather degradation methods. As shown in Table 2, CoA ranks second in one metric, outperforming others in all others, highlighting its superiority. Fig. 5 reveals that while existing methods excel in specific scenarios, they show limitations in handling", + "bbox": [ + 89, + 810, + 483, + 901 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e97872433ad8806d3362f267cc5bdd774ca4e0cafe2ec5f0b8f53a2144ea6e36.jpg", + "image_caption": [ + "Figure 8. Effects of MoC phase.", + "Input" + ], + "image_footnote": [], + "bbox": [ + 532, + 244, + 705, + 324 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b6eda8b4b2a5426283b294eaacd223ad60d967ea47049356749234569e004eaf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 709, + 244, + 882, + 323 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/797617484687b57bcb3a6bf742ca04f67b17ddcc3554044447065b22123ce452.jpg", + "image_caption": [ + "(b) Quantitative comparison", + "Only Upper-Level Modeling", + "Figure 9. Necessity of bilevel modeling." + ], + "image_footnote": [], + "bbox": [ + 532, + 337, + 704, + 415 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/092eaff8c72d7d7dd4461cdd5075cf160a79586127fcf9e3a93ef8e22a20ba27.jpg", + "image_caption": [ + "Only Lower-Level Modeling", + "Bilevel Modeling" + ], + "image_footnote": [], + "bbox": [ + 709, + 337, + 882, + 415 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "the full spectrum of haze conditions.", + "bbox": [ + 511, + 468, + 754, + 483 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Efficiency comparisons. Table 3 presents a comprehensive comparison of our proposed CoA method with existing approaches, evaluating parameters, floating-point operations per second (FLOPs), and processing time across various image resolutions. The results clearly show that our method has the smallest number of parameters and FLOPs among all the evaluated methods. Notably, as the image resolution increases, our method demonstrates a significant advantage in processing time, indicating that our model achieves a favorable balance between performance and complexity.", + "bbox": [ + 511, + 484, + 906, + 636 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Adaptability Verification", + "text_level": 1, + "bbox": [ + 511, + 648, + 740, + 666 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Nighttime haze scene. Compared to state-of-the-art nighttime dehazing algorithms on the NHRW dataset (Fig. 6 (a)), many existing methods fail to handle overexposure or low-light conditions, resulting in overly dark or distorted images due to noise and artificial light scattering. In contrast, CoA excels in real-world nighttime haze scenes, effectively removing haze, suppressing noise, and preserving details. As shown in Fig. 6 (b), our method outperforms others across three authoritative objective metrics", + "bbox": [ + 511, + 672, + 906, + 808 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Underwater scene. Compared to state-of-the-art underwater dehazing algorithms, DM results in reddish and blurry images, failing to correct color distortion effectively, while SUIR and SMDR struggle to fully eliminate haze. As shown in Fig. 7, our method significantly reduces haze impact and enhances clarity and detail visibility, closely align-", + "bbox": [ + 511, + 810, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "11203", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/285705a7df2e711b0a1d2b4402b7e9d0e3e84db1b75a0842ba0989e1de9ed8c6.jpg", + "image_caption": [ + "Figure 10. Effects of BiA phase." + ], + "image_footnote": [], + "bbox": [ + 102, + 85, + 477, + 250 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/59f9f86a4a9e8e1aad6b178f26df9270b7af675b4e39fd50ebfaf0886a0db105.jpg", + "image_caption": [ + "Figure 11. Parameters Analysis. The RTTS dataset is adopted." + ], + "image_footnote": [], + "bbox": [ + 102, + 279, + 475, + 380 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ing with human visual perception standards.", + "bbox": [ + 89, + 425, + 383, + 440 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Analytical Experiments", + "text_level": 1, + "bbox": [ + 89, + 454, + 313, + 470 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.1. Effects of MoC Phase", + "text_level": 1, + "bbox": [ + 89, + 479, + 292, + 494 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Without feature transfer from the \"naive for teacher\" (naive: end-to-end training, the same below) model via compression, the \"naive for student\" model struggles to capture core data patterns, particularly in complex or extreme weather scenarios, resulting in limited dehazing. As illustrated in Fig. 8, while \"MoC for student\" may not restore details as effectively as \"naive for teacher\", it successfully inherits high-quality features. Even under resource constraints, it surpasses \"naive for student\" in dehazing and performs well on unsupervised metrics.", + "bbox": [ + 89, + 502, + 483, + 654 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.2. Necessity of Bilevel Modeling", + "text_level": 1, + "bbox": [ + 89, + 662, + 352, + 680 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Here we highlight the critical role of bilevel modeling by training the same network model on the THaze dataset. As shown in Fig. 9, the results on the real test set show marked contrasts: employing only the lower-level modeling strategy results in suboptimal dehazing, while relying solely on the upper-level strategy with CLIP fine-tuning produces unnatural effects with notable color distortion. Different from them, the bilevel modeling strategy enables both effective dehazing and natural visual quality, achieving a balance between dehazing performance and perceptual realism.", + "bbox": [ + 89, + 686, + 483, + 838 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.3. Effects of BiA Phase", + "text_level": 1, + "bbox": [ + 89, + 847, + 284, + 862 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In Fig. 10, we evaluate effects of the BiA phase. The results clearly show significant improvements in both FADE and", + "bbox": [ + 89, + 869, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/cc5cca6b38d12f1b1ca8f5e1721640f10e329dde0a7cbbf10acba18de888ccfa.jpg", + "image_caption": [ + "Figure 12. Limitations. The example is from RS-Haze [34]." + ], + "image_footnote": [], + "bbox": [ + 527, + 88, + 897, + 172 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "PM2.5 metrics following the BiA phase. Notably, in images with complex backgrounds or high noise levels, our model outperforms the baseline, demonstrating the BiA phase's effectiveness in capturing key features and improving robustness in challenging scenarios.", + "bbox": [ + 511, + 219, + 905, + 294 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.4. Parameters Analyses", + "text_level": 1, + "bbox": [ + 511, + 306, + 710, + 321 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We test different values of the parameter $\\alpha$ during the BiA phase, as shown in Fig. 11. The results indicate that $\\alpha = 0.95$ yields almost optimal performance. Notably, this setting achieves the fastest convergence and exhibits superior stability and accuracy across multiple evaluation metrics, clearly outperforming other parameter settings.", + "bbox": [ + 511, + 329, + 905, + 420 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.5. Limitations", + "text_level": 1, + "bbox": [ + 511, + 431, + 638, + 445 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "While effective in many scenarios, our method struggles with remote sensing haze images due to additional imaging constraints. As depicted in Fig. 12, although it improves dehazing (outperforming DEA), non-uniform fog occlusion remains. This stems from varying haze effects across spectral bands in hyperspectral images, requiring multi-band integration for effective dehazing.", + "bbox": [ + 511, + 454, + 905, + 560 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 575, + 633, + 590 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work proposes a novel compression-and-adaptation scheme to tackle the problem of unbalance between dehazing quality and computational efficiency in real image dehazing. Following the divide-and-conquer paradigm, we design a new CoA learning strategy that first compresses the predefined model and then adapts it to unlabeled, diverse real-world scenes. We not only make a thorough exploration to take on the excellent properties of CoA, but also we perform extensive experiments to indicate our CoA method's superiority in terms of image dehazing quality and computational efficiency.", + "bbox": [ + 511, + 601, + 906, + 768 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 511, + 784, + 671, + 801 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work was supported by the Postdoctoral Fellowship Program of CPSF (No. GZB20240098), China Postdoctoral Science Foundation (No. 2023M740491), Dalian Science and Technology Innovation Fund-Young Tech Star (No. 2023RQ017), Fundamental Research Funds for the Central Universities (No. DUT24BS011).", + "bbox": [ + 511, + 810, + 905, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "11204", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Haoran Bai, Jinshan Pan, Xinguang Xiang, and Jinhui Tang. Self-guided image dehazing using progressive feature fusion. IEEE Transactions on Image Processing, 31:1217-1229, 2022. 6", + "[2] Zhaowei Cai, Avinash Ravichandran, Subhransu Maji, Charless Fowlkes, Zhuowen Tu, and Stefano Soatto. Exponential moving average normalization for self-supervised and semi-supervised learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 194-203, 2021. 4", + "[3] Zixuan Chen, Zewei He, and Zhe-Ming Lu. Dea-net: Single image dehazing based on detail-enhanced convolution and content-guided attention. IEEE Transactions on Image Processing, 2024. 5", + "[4] Lark Kwon Choi, Jaehee You, and Alan C. Bovik. Referenceless prediction of perceptual fog density and perceptual image defogging. IEEE Transactions on Image Processing, 24(11):3888-3901, 2015. 6", + "[5] Hang Dong, Jinshan Pan, Lei Xiang, Zhe Hu, Xinyi Zhang, Fei Wang, and Ming-Hsuan Yang. Multi-scale boosted dehazing network with dense feature fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2157-2167, 2020. 4, 5", + "[6] Raanan Fattal. Dehazing using color-lines. ACM transactions on graphics, 34(1):1-14, 2014. 4, 5", + "[7] Yuxin Feng, Long Ma, Xiaozhe Meng, Fan Zhou, Risheng Liu, and Zhuo Su. Advancing real-world image dehazing: Perspective, modules, and training. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(12):9303-9320, 2024. 6", + "[8] Yuxin Feng, Zhuo Su, Long Ma, Xin Li, Risheng Liu, and Fan Zhou. Bridging the gap between haze scenarios: A unified image dehazing model. IEEE Transactions on Circuits and Systems for Video Technology, 34(11):11070-11085, 2024. 2, 5", + "[9] Luca Franceschi, Paolo Frasconi, Saverio Salzo, Riccardo Grazzi, and Massimiliano Pontil. Bilevel programming for hyperparameter optimization and meta-learning. In Proceedings of the International Conference on Machine Learning, pages 1568-1577, 2018. 3", + "[10] Minghan Fu, Huan Liu, Yankun Yu, Jun Chen, and Keyan Wang. DW-GAN: A discrete wavelet transform GAN for nonhomogeneous dehazing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 203-212, 2021. 4", + "[11] Xueyang Fu, Jie Xiao, Yurui Zhu, Aiping Liu, Feng Wu, and Zheng-Jun Zha. Continual image deraining with hypergraph convolutional networks. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(8):9534-9551, 2023. 1", + "[12] Alona Golts, Daniel Freedman, and Michael Elad. Unsupervised single image dehazing using dark channel prior loss. IEEE Transactions on Image Processing, 29:2692-2701, 2019. 2", + "[13] Ke Gu, Dacheng Tao, Junfei Qiao, and Weisi Lin. Learning a no-reference quality assessment model of enhanced images" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "with big data. IEEE Transactions on Neural Networks and Learning Systems, 29(4):1301-1313, 2018. 6", + "[14] Ke Gu, Junfei Qiao, and Xiaoli Li. Highly efficient picture-based prediction of PM2.5 concentration. IEEE Transactions on Industrial Electronics, 66(4):3176-3184, 2019. 6", + "[15] Jie Gui, Xiaofeng Cong, Lei He, Yuan Yan Tang, and James Tin-Yau Kwok. Illumination controllable dehazing network based on unsupervised retinae embedding. IEEE Transactions on Multimedia, 26:4819-4830, 2024. 2", + "[16] Chunle Guo, Qixin Yan, Saeed Anwar, Runmin Cong, Wenqi Ren, and Chongyi Li. Image dehazing transformer with transmission-aware 3d position embedding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5802-5810, 2022. 6", + "[17] Kaiming He, Jian Sun, and Xiaou Tang. Single image haze removal using dark channel prior. IEEE Transactions on Pattern Analysis and Machine Intelligence, 33(12):2341-2353, 2010. 2", + "[18] Shirui Huang, Keyan Wang, Huan Liu, Jun Chen, and Yun-song Li. Contrastive semi-supervised learning for underwater image restoration via reliable bank. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18145-18155, 2023. 7", + "[19] Yeying Jin, Beibei Lin, Wending Yan, Yuan Yuan, Wei Ye, and Robby T. Tan. Enhancing visibility in nighttime haze images using guided APSF and gradient adaptive convolution. In Proceedings of the ACM International Conference on Multimedia, pages 2446-2457, 2023. 7", + "[20] Mingye Ju, Can Ding, Charles A. Guo, Wenqi Ren, and Dacheng Tao. IDRLP: image dehazing using region line prior. IEEE Transactions on Image Processing, 30:9043-9057, 2021. 2", + "[21] Boyi Li, Wenqi Ren, Dengpan Fu, Dacheng Tao, Dan Feng, Wenjun Zeng, and Zhangyang Wang. Benchmarking single-image dehazing and beyond. IEEE Transactions on Image Processing, 28(1):492-505, 2018. 2, 5", + "[22] Boyi Li, Wenqi Ren, Dengpan Fu, Dacheng Tao, Dan Feng, Wenjun Zeng, and Zhangyang Wang. Benchmarking single-image dehazing and beyond. IEEE Transactions on Image Processing, 28(1):492-505, 2019. 4", + "[23] Boyun Li, Xiao Liu, Peng Hu, Zhongqin Wu, Jiancheng Lv, and Xi Peng. All-in-one image restoration for unknown corruption. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17431-17441, 2022. 7", + "[24] Pengyue Li, Jiandong Tian, Yandong Tang, Guolin Wang, and Chengdong Wu. Deep retina network for single image dehazing. IEEE Transactions on Image Processing, 30: 1100-1115, 2021. 2", + "[25] Yu Li, Robby T. Tan, and Michael S. Brown. Nighttime haze removal with glow and multiple light colors. In Proceedings of the International Conference on Computer Vision, pages 226-234, 2015. 7", + "[26] Yudong Liang, Bin Wang, Wangmeng Zuo, Jiaying Liu, and Wenqi Ren. Self-supervised learning and adaptation for single image dehazing. In Proceedings of the Thirty-First International Joint Conference on Artificial Intelligence, pages 1137–1143, 2022. 2" + ], + "bbox": [ + 517, + 92, + 906, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "11205", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[27] Pengyang Ling, Huaian Chen, Xiao Tan, Yi Jin, and Enhong Chen. Single image dehazing using saturation line prior. IEEE Transactions on Image Processing, 32:3238-3253, 2023. 1", + "[28] Jun Liu, Ryan Wen Liu, Jianing Sun, and Tieyong Zeng. Rank-one prior: Real-time scene recovery. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(7): 8845-8860, 2023. 1, 2", + "[29] Risheng Liu, Jiaxin Gao, Jin Zhang, Deyu Meng, and Zhouchen Lin. Investigating bi-level optimization for learning and vision from a unified perspective: A survey and beyond. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(12):10045-10067, 2021. 3", + "[30] Ye Liu, Lei Zhu, Shunda Pei, Huazhu Fu, Jing Qin, Qing Zhang, Liang Wan, and Wei Feng. From synthetic to real: Image dehazing collaborating with unlabeled real data. In Proceedings of the ACM International Conference on Multimedia, pages 50-58, 2021. 2, 5", + "[31] Yidi Liu, Xingbo Wang, Yurui Zhu, Xueyang Fu, and Zheng-Jun Zha. Sdcnet: Spatially-adaptive deformable convolution networks for hr nonhomogeneous dehazing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6682-6691, 2024. 2", + "[32] Long Ma, Tengyu Ma, Risheng Liu, Xin Fan, and Zhongxuan Luo. Toward fast, flexible, and robust low-light image enhancement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5637-5646, 2022. 1", + "[33] Ozan Özdenizci and Robert Legenstein. Restoring vision in adverse weather conditions with patch-based denoising diffusion models. IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 1-12, 2023. 7", + "[34] Yuda Song, Zhuqing He, Hui Qian, and Xin Du. Vision transformers for single image dehazing. IEEE Transactions on Image Processing, 32:1927-1941, 2023. 5, 8", + "[35] Yi Tang, Hiroshi Kawasaki, and Takafumi Iwaguchi. Underwater image enhancement by transformer-based diffusion model with non-uniform sampling for skip strategy. In Proceedings of the ACM International Conference on Multimedia, pages 5419-5427, 2023. 7", + "[36] Zhongze Wang, Haitao Zhao, Jingchao Peng, Lujian Yao, and Kaijie Zhao. Odcr: Orthogonal decoupling contrastive regularization for unpaired image dehazing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25479-25489, 2024. 2", + "[37] Ruiqi Wu, Zheng-Peng Duan, Chun-Le Guo, Zhi Chai, and Chongyi Li. RIDCP: revitalizing real image dehazing via high-quality codebook priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22282-22291, 2023. 2, 7", + "[38] Yang Yang, Chaoyue Wang, Risheng Liu, Lin Zhang, Xiaojie Guo, and Dacheng Tao. Self-augmented unpaired image dehazing via density and depth decomposition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2037-2046, 2022. 6", + "[39] Yang Yang, Chaoyue Wang, Xiaojie Guo, and Dacheng Tao. Robust unpaired image dehazing via density and depth de" + ], + "bbox": [ + 91, + 92, + 480, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "composition. International Journal of Computer Vision, 132 (5):1557-1577, 2024. 6", + "[40] Dehuan Zhang, Jingchun Zhou, Chunle Guo, Weishi Zhang, and Chongyi Li. Synergistic multiscale detail refinement via intrinsic supervision for underwater image enhancement. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 7033-7041, 2024. 7", + "[41] He Zhang and Vishal M Patel. Densely connected pyramid dehazing network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3194-3203, 2018. 2", + "[42] Jing Zhang, Yang Cao, and et al. Fast haze removal for nighttime image using maximum reflectance prior. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7016-7024, 2017. 7", + "[43] Jing Zhang, Yang Cao, and et al. Nighttime dehazing with a synthetic benchmark. In Proceedings of the 28th ACM International Conference on Multimedia, pages 2355-2363, 2020. 7", + "[44] Yafei Zhang, Shen Zhou, and Huafeng Li. Depth information assisted collaborative mutual promotion network for single image dehazing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2846-2855, 2024. 2", + "[45] Dian Zheng, Xiao-Ming Wu, Shuzhou Yang, Jian Zhang, Jian-Fang Hu, and Wei-Shi Zheng. Selective hourglass mapping for universal image restoration based on diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 7", + "[46] Yu Zheng, Jiahui Zhan, and et al. Curricular contrastive regularization for physics-aware single image dehazing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition., 1, 6" + ], + "bbox": [ + 516, + 92, + 903, + 556 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "11206", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/CoA_ Towards Real Image Dehazing via Compression-and-Adaptation/06a0b6a5-a641-41a6-99e5-3f3c8a8c1a70_model.json b/2025/CoA_ Towards Real Image Dehazing via Compression-and-Adaptation/06a0b6a5-a641-41a6-99e5-3f3c8a8c1a70_model.json new file mode 100644 index 0000000000000000000000000000000000000000..7248da4207e2ebae2c8f1d2e5c588acef7303d5e --- /dev/null +++ b/2025/CoA_ Towards Real Image Dehazing via Compression-and-Adaptation/06a0b6a5-a641-41a6-99e5-3f3c8a8c1a70_model.json @@ -0,0 +1,2992 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.143, + 0.131, + 0.856, + 0.154 + ], + "angle": 0, + "content": "CoA: Towards Real Image Dehazing via Compression-and-Adaptation" + }, + { + "type": "text", + "bbox": [ + 0.154, + 0.18, + 0.239, + 0.199 + ], + "angle": 0, + "content": "Long Ma" + }, + { + "type": "text", + "bbox": [ + 0.283, + 0.181, + 0.386, + 0.217 + ], + "angle": 0, + "content": "Yuxin Feng\\(^{2}\\) \nJiang-Yong Chou" + }, + { + "type": "text", + "bbox": [ + 0.43, + 0.181, + 0.526, + 0.199 + ], + "angle": 0, + "content": "Yan Zhang" + }, + { + "type": "text", + "bbox": [ + 0.572, + 0.181, + 0.674, + 0.199 + ], + "angle": 0, + "content": "Jinyuan Liu1" + }, + { + "type": "text", + "bbox": [ + 0.721, + 0.181, + 0.84, + 0.199 + ], + "angle": 0, + "content": "Weimin Wang" + }, + { + "type": "text", + "bbox": [ + 0.48, + 0.2, + 0.605, + 0.218 + ], + "angle": 0, + "content": "Chengpei \\(\\mathrm{Xu}^{4,*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.66, + 0.2, + 0.738, + 0.216 + ], + "angle": 0, + "content": "Zhuo Su" + }, + { + "type": "text", + "bbox": [ + 0.117, + 0.218, + 0.884, + 0.27 + ], + "angle": 0, + "content": "\\(^{1}\\)Dalian University of Technology \\(^{2}\\)Sun Yat-sen University \\(^{3}\\)Fuzhou University \\(^{4}\\)University of New South Wales malone94319@gmail.com, {fengyx26, zhangy2779} \\(@\\) mail2.sysu.edu.cn, atlantis918@hotmail.com, chengpei.Xu@unsw.edu.au, wangweimin@dlut.edu.cn, gychen@fzu.edu.cn, suzhuo3@mail.sysu.edu.cn" + }, + { + "type": "image", + "bbox": [ + 0.104, + 0.29, + 0.375, + 0.406 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.136, + 0.408, + 0.34, + 0.418 + ], + "angle": 0, + "content": "(a) Evaluating efficiency across various metrics" + }, + { + "type": "image", + "bbox": [ + 0.375, + 0.29, + 0.599, + 0.405 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.413, + 0.408, + 0.575, + 0.418 + ], + "angle": 0, + "content": "(b) The computational process of CoA" + }, + { + "type": "image", + "bbox": [ + 0.599, + 0.29, + 0.9, + 0.406 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.651, + 0.408, + 0.862, + 0.418 + ], + "angle": 0, + "content": "(c) Evaluating adaptability across different scenes" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.423, + 0.907, + 0.48 + ], + "angle": 0, + "content": "Figure 1. Performance evaluation. The proposed CoA incorporates model compression in synthetic domain for efficiency and bilevel adaptation in real domain for adaptability, as illustrated in the central sub-figure (b). The left sub-figure (a) presents efficiency across various metrics, clearly showing that our CoA outperforms others by a significant margin. The right sub-figure (c) shows adaptability across different scenes, where it is evident that our CoA consistently performs excellently." + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.499, + 0.327, + 0.515 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.532, + 0.485, + 0.864 + ], + "angle": 0, + "content": "Learning-based image dehazing algorithms have shown remarkable success in synthetic domains. However, real image dehazing is still in suspense due to computational resource constraints and the diversity of real-world scenes. Therefore, there is an urgent need for an algorithm that excels in both efficiency and adaptability to address real image dehazing effectively. This work proposes a Compression-and-Adaptation (CoA) computational flow to tackle these challenges from a divide-and-conquer perspective. First, model compression is performed in the synthetic domain to develop a compact dehazing parameter space, satisfying efficiency demands. Then, a bilevel adaptation in the real domain is introduced to be fearless in unknown real environments by aggregating the synthetic dehazing capabilities during the learning process. Leveraging a succinct design free from additional constraints, our CoA exhibits domain-irrelevant stability and model-agnostic flexibility, effectively bridging the model chasm between synthetic and real domains to further improve its practical utility. Extensive evaluations and analyses underscore the approach's superiority and effectiveness. The code is publicly available at https://github.com/fyxn1/CoA." + }, + { + "type": "title", + "bbox": [ + 0.515, + 0.499, + 0.645, + 0.515 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.524, + 0.907, + 0.797 + ], + "angle": 0, + "content": "Image dehazing tasks aim to utilize the atmospheric optical model and advanced computing theory to achieve efficient restoration of scene information and image details through inverse imaging solution or data-driven model learning. It has drawn much attention in multiple emerging computer vision areas recently [46] [27] [28]. Similar to other low-level vision tasks [32] [11], image dehazing technology has evolved from an early stage focused on improving synthetic data metrics to a direction aimed at effectively generalizing across various real-world haze scenarios. Current approaches face dual challenges: efficiency-oriented methods achieve real-time processing but lack scene adaptability, while adaptability-focused techniques dynamically adjust strategies at the cost of high computational complexity that hinders deployment in resource-constrained environments. This paper addresses these limitations through novel architectural innovations in both computational efficiency and adaptive scene understanding." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.805, + 0.666, + 0.82 + ], + "angle": 0, + "content": "1.1. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.827, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Efficiency-oriented Dehazing Methods. This work focuses on enhancing computational efficiency through optimization algorithms, model simplification, or machine learning methods to enable faster processing under limited resources. Broadly, dehazing methods can be divided" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.889, + 0.236, + 0.902 + ], + "angle": 0, + "content": "*Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11197" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.484, + 0.333 + ], + "angle": 0, + "content": "into image enhancement-based, physics-based, and image layering-based approaches. Image enhancement-based dehazing methods [24] treat the task as an image improvement problem, increasing contrast and color accuracy with techniques like histogram equalization and color correction. Physics-based methods [17], [28] are more specific, leveraging hazy image priors and constraints to aid inverse problem solving. Image layering-based approaches, such as that of Zhang et al. [41], increase efficiency by breaking down the image into scales or layers, progressively refining transmittance maps and enhancing detail using multilevel pyramid structures and dense connectivity. Despite their computational and real-time advantages, they struggle with stability and consistent quality, especially in complex real-world haze conditions, often resulting in partial dehazing and insufficient color clarity." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.334, + 0.485, + 0.668 + ], + "angle": 0, + "content": "Adaptability-focused Dehazing Methods. Adaptive dehazing methods tailor their strategies based on specific image attributes to improve dehazing effectiveness. These methods encompass three main types: local feature-based, atmospheric scattering model-based, and domain adaptation-based approaches. Local feature-based methods [15] adaptively adjust dehazing intensity according to local contrast and brightness variations, which helps restore image details effectively. Atmospheric scattering model-based adaptive dehazing methods [20] dynamically adjust the transmittance estimation by analyzing different regions of the image to achieve more accurate restoration. These methods can, to some extent, balance both efficiency and adaptability. Domain adaptation-based dehazing methods [26, 37] have gradually emerged in the field of image dehazing in recent years. These methods learn a shared feature representation, enabling the model to simultaneously adapt to the features of both the source domain and the target domain. Although these methods yield superior dehazing results, their complex models and dynamic imagespecific adjustments often increase computational costs, potentially impacting real-time applicability." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.68, + 0.273, + 0.696 + ], + "angle": 0, + "content": "1.2. Our Contributions" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.703, + 0.483, + 0.839 + ], + "angle": 0, + "content": "To address the challenges of efficiency and adaptability in real image dehazing, this work proposes a new Compression-and-Adaptation (CoA) scheme as shown in Fig. 1 (b). The predefined model is compressed in the synthetic domain to ensure computational efficiency (refer to Fig. 1 (a)). Further, a bilevel adaptive learning is developed to leverage the dehazing capability acquired in the synthetic domain to improve adaptability (see Fig. 1 (c)) in the real domain. Our key contributions are summarized as" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "- Following the divide-and-conquer paradigm, we propose a new CoA learning strategy that first compresses the predefined model and then adapts it to unseen scenes, offering an efficient solution for real image dehazing. To the" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.092, + 0.905, + 0.137 + ], + "angle": 0, + "content": "best of our knowledge, this is the first work to successfully integrate both efficiency and adaptability, achieving the best of both worlds in real image dehazing." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.137, + 0.905, + 0.242 + ], + "angle": 0, + "content": "- To adapt to unlabeled, diverse real-world scenes, we propose a new cross-domain bilevel model designed to learn dehazing parameters for the real domain constrained by synthetic dehazing learning. Additionally, we derive a bilevel adaptive learning scheme that effectively harnesses the synthetic dehazing capacity, preventing instability and enhancing adaptability." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.243, + 0.907, + 0.348 + ], + "angle": 0, + "content": "- Benefiting from succinct design free from additional constraints, CoA shows two key properties on three standard real benchmarks. It shows stability across synthetic domains, improving \\(14.2\\%\\) on average across four metrics (evaluated on three domains). CoA also offers flexibility with dehazing models, reducing computational cost by at least \\(73\\%\\) in parameters (tested on three methods)." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.137, + 0.907, + 0.348 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.359, + 0.724, + 0.376 + ], + "angle": 0, + "content": "2. The Proposed Method" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.385, + 0.906, + 0.461 + ], + "angle": 0, + "content": "In this section, we first rethinking real image dehazing, highlighting two key practical needs. We then propose model compression and bilevel adaptation to meet these needs, followed by an overview of the network architecture and training loss." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.467, + 0.805, + 0.484 + ], + "angle": 0, + "content": "2.1. Rethinking Real Image Dehazing" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.489, + 0.906, + 0.579 + ], + "angle": 0, + "content": "Despite the strong performance of advanced image dehazing techniques [31, 36, 44] on synthetic datasets [8, 21, 30], these datasets often rely on simplifying assumptions, resulting in a substantial gap between synthetic and real-world domains. Addressing the core challenges in real-world image dehazing involves two key aspects:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.58, + 0.905, + 0.625 + ], + "angle": 0, + "content": "- Efficiency: Edge devices often have limited computational resources compared to systems with ample capacity, requiring dehazing models to be highly efficient." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.625, + 0.905, + 0.67 + ], + "angle": 0, + "content": "- Adaptability: Real-world imaging conditions vary dynamically. Dehazing models must therefore be adaptable to maintain reliable performance." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.58, + 0.905, + 0.67 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.671, + 0.906, + 0.776 + ], + "angle": 0, + "content": "The latter is more challenging, relying on effective learning mechanisms like unsupervised learning [12]. However, current methods struggle with unsupervised learning due to difficulty in capturing statistical regularities, making transferring dehazing capabilities from the synthetic to real domain a viable alternative. To provide an intuitive presentation of the goal upon the above demands, we build" + }, + { + "type": "equation", + "bbox": [ + 0.561, + 0.782, + 0.905, + 0.851 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\min _ {\\boldsymbol {\\theta} _ {\\mathrm {r e a l}}} f \\Big (\\boldsymbol {\\theta} _ {\\mathrm {r e a l}} \\left(\\boldsymbol {\\theta} _ {\\mathrm {s y n}}\\right) \\Big), \\\\ s. t., \\left\\{ \\begin{array}{l} \\kappa (\\boldsymbol {\\theta} _ {\\text {r e a l}}) < \\kappa (\\boldsymbol {\\theta} _ {\\text {s y n}}), (\\text {e f f i c i e n c y}) \\\\ \\zeta (\\boldsymbol {\\theta} _ {\\text {r e a l}}) > \\zeta (\\boldsymbol {\\theta} _ {\\text {s y n}}), (\\text {a d a p t a b i l i t y}) \\end{array} \\right. \\tag {1} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.906, + 0.902 + ], + "angle": 0, + "content": "where \\(\\theta_{\\mathrm{syn}}\\) and \\(\\theta_{\\mathrm{real}}\\) represent parameters in synthetic and real domain, respectively. The function \\(\\kappa (\\cdot)\\) and \\(\\zeta (\\cdot)\\) denote the evaluator for efficiency and adaptability, respectively." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11198" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.104, + 0.092, + 0.895, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.113, + 0.221, + 0.882, + 0.236 + ], + "angle": 0, + "content": "Figure 2. Qualitative comparisons of algorithmic properties. Arrows indicate specific regions that highlight visible differences." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.262, + 0.484, + 0.324 + ], + "angle": 0, + "content": "To achieve this, we propose a divide-and-conquer strategy: the first phase focuses on model compression for efficiency, while the second phase employs bilevel adaptation for adaptability. Details are provided below." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.331, + 0.484, + 0.348 + ], + "angle": 0, + "content": "2.2. MoC: Model Compression in Synthetic Domain" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.354, + 0.484, + 0.414 + ], + "angle": 0, + "content": "The main challenge in model compression is transferring dehazing capabilities from a large-scale parameter space to a smaller one. Here we propose a composite loss function with context alignment to realize it." + }, + { + "type": "equation", + "bbox": [ + 0.172, + 0.424, + 0.483, + 0.45 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\boldsymbol {\\theta} _ {\\mathrm {s y n}} ^ {\\mathrm {s}}} \\mathcal {L} _ {\\mathrm {s y n}} \\left(\\boldsymbol {\\theta} _ {\\mathrm {s y n}} ^ {\\mathrm {s}}\\right) + \\mathcal {L} _ {\\mathrm {a}} \\left(\\boldsymbol {\\theta} _ {\\mathrm {s y n}} ^ {\\mathrm {s}}, \\boldsymbol {\\theta} _ {\\mathrm {s y n}} ^ {\\mathrm {t}}\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.459, + 0.484, + 0.52 + ], + "angle": 0, + "content": "where \\(\\theta_{\\mathrm{syn}}^{\\mathrm{t}}\\) and \\(\\theta_{\\mathrm{syn}}^{\\mathrm{s}}\\) are the pre-trained teacher model and desired student model in the synthetic domain, respectively. The function \\(\\mathcal{L}_{\\mathrm{syn}}\\) and \\(\\mathcal{L}_{\\mathrm{a}}\\) represent a group of supervised losses and the context alignment loss, respectively." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.521, + 0.484, + 0.701 + ], + "angle": 0, + "content": "During model compression, the student model follows the teacher model's training strategy on paired synthetic data. To maximize the retention of the teacher model's capabilities, we introduce a context alignment loss \\((\\mathcal{L}_{\\mathrm{a}})\\) that aligns the outputs of the teacher and student encoders at corresponding positions. By weighting the loss functions across layers, this approach enables fine-grained feature alignment, improving knowledge transfer and model optimization. The layered alignment ensures both global feature distributions and layer-specific details are accurately matched, resulting in more precise compression of the teacher model into the student model." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.71, + 0.431, + 0.726 + ], + "angle": 0, + "content": "2.3. BiA:Bilevel Adaptation to Real Domain" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.732, + 0.484, + 0.807 + ], + "angle": 0, + "content": "After compressing the model for efficiency, we address adaptation to real-world scenes by proposing a cross-domain bilevel model. This is followed by a bilevel adaptive learning scheme to enable smooth transition from the synthetic to real domain." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.815, + 0.359, + 0.83 + ], + "angle": 0, + "content": "2.3.1. Cross-Domain Bilevel Modeling" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.833, + 0.484, + 0.88 + ], + "angle": 0, + "content": "The gap between the synthetic and real domains makes adaptation challenging. The dehazing parameters in the synthetic and real domains \\((\\theta_{\\mathrm{rea}}^s)\\) have a nested subordinate" + }, + { + "type": "title", + "bbox": [ + 0.516, + 0.262, + 0.887, + 0.277 + ], + "angle": 0, + "content": "Algorithm 1 Learning via Compression-and-Adaptation" + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.282, + 0.906, + 0.329 + ], + "angle": 0, + "content": "Require: The pre-trained teacher model \\(\\theta_{\\mathrm{syn}}^{\\mathrm{t}}\\), the initial parameter \\(\\theta_{\\mathrm{syn}}^{\\mathrm{s}}\\), step-sizes \\(\\eta_{\\mathrm{syn}}^{m}\\), \\(\\eta_{\\mathrm{syn}}^{b}\\), coefficient \\(\\alpha\\), iteration numbers \\(N\\) in MoC phase and \\(T\\) in BiA phase." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.33, + 0.777, + 0.346 + ], + "angle": 0, + "content": "Ensure: The optimal parameters \\(\\theta_{\\mathrm{rea}}^{\\mathrm{s}}\\)" + }, + { + "type": "text", + "bbox": [ + 0.525, + 0.346, + 0.671, + 0.358 + ], + "angle": 0, + "content": "1: % The MoC Phase" + }, + { + "type": "text", + "bbox": [ + 0.525, + 0.36, + 0.69, + 0.373 + ], + "angle": 0, + "content": "2: for \\( n = 0:N - 1 \\) do" + }, + { + "type": "text", + "bbox": [ + 0.524, + 0.375, + 0.791, + 0.389 + ], + "angle": 0, + "content": "3: Calculate the gradient \\( g \\) of Eq. (2)." + }, + { + "type": "text", + "bbox": [ + 0.525, + 0.389, + 0.872, + 0.407 + ], + "angle": 0, + "content": "4: \\(\\pmb{\\theta}_{\\mathrm{syn}}^{\\mathrm{s}}(n + 1) = \\pmb{\\theta}_{\\mathrm{syn}}^{\\mathrm{s}}(n) - \\eta_{\\mathrm{syn}}^{m}\\pmb{g}(\\pmb{\\theta}_{\\mathrm{syn}}^{\\mathrm{s}},n),\\pmb{\\theta}_{\\mathrm{syn}}^{\\mathrm{t}}).\\)" + }, + { + "type": "text", + "bbox": [ + 0.525, + 0.407, + 0.597, + 0.419 + ], + "angle": 0, + "content": "5: end for" + }, + { + "type": "text", + "bbox": [ + 0.524, + 0.42, + 0.905, + 0.452 + ], + "angle": 0, + "content": "6: Initialize the dehazing model's parameter in the real domain by \\(\\pmb{\\theta}_{\\mathrm{rea}}^{\\mathrm{s}} = \\pmb{\\theta}_{\\mathrm{syn}}^{\\mathrm{s}}(N)\\)." + }, + { + "type": "text", + "bbox": [ + 0.525, + 0.452, + 0.664, + 0.464 + ], + "angle": 0, + "content": "7: % The BiA Phase" + }, + { + "type": "text", + "bbox": [ + 0.525, + 0.466, + 0.682, + 0.479 + ], + "angle": 0, + "content": "8: for \\( t = 0 : T - 1 \\) do" + }, + { + "type": "text", + "bbox": [ + 0.524, + 0.481, + 0.886, + 0.498 + ], + "angle": 0, + "content": "9: \\(\\%\\) Update the lower-level parameter \\(\\theta_{\\mathrm{syn}}^{\\mathrm{s}}\\) in \\(\\mathcal{D}_{\\mathrm{rea}}\\)." + }, + { + "type": "text", + "bbox": [ + 0.524, + 0.498, + 0.825, + 0.518 + ], + "angle": 0, + "content": "10: \\(\\pmb{\\theta}_{\\mathrm{syn}}^{\\mathrm{s}}(t + 1) = \\pmb{\\theta}_{\\mathrm{syn}}^{\\mathrm{s}}(t) - \\eta_{\\mathrm{syn}}^{b}\\frac{\\partial(\\mathcal{L}_{\\mathrm{rea}} + \\mathcal{L}_{1})}{\\partial\\pmb{\\theta}_{\\mathrm{syn}}^{\\mathrm{s}}}\\)." + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.518, + 0.894, + 0.534 + ], + "angle": 0, + "content": "11: \\(\\%\\) Update the upper-level parameter \\(\\theta_{\\mathrm{rea}}^{\\mathrm{s}}\\) by EMA." + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.534, + 0.866, + 0.55 + ], + "angle": 0, + "content": "12: \\(\\pmb{\\theta}_{\\mathrm{rea}}^{\\mathrm{s}}(t + 1) = \\alpha \\pmb{\\theta}_{\\mathrm{rea}}^{\\mathrm{s}}(t) + (1 - \\alpha)\\pmb{\\theta}_{\\mathrm{syn}}^{\\mathrm{s}}(t + 1).\\)" + }, + { + "type": "text", + "bbox": [ + 0.521, + 0.55, + 0.597, + 0.561 + ], + "angle": 0, + "content": "13: end for" + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.563, + 0.655, + 0.579 + ], + "angle": 0, + "content": "14: return \\(\\theta_{\\mathrm{rea}}^{\\mathrm{s}}(T)\\)" + }, + { + "type": "list", + "bbox": [ + 0.52, + 0.346, + 0.905, + 0.579 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.607, + 0.906, + 0.667 + ], + "angle": 0, + "content": "relationship. This coupling can be effectively modeled using bilevel programming techniques [9, 29]. From the perspective of hyperparameter optimization, our cross-domain bilevel model is written as:" + }, + { + "type": "equation", + "bbox": [ + 0.573, + 0.674, + 0.809, + 0.699 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\boldsymbol {\\theta} _ {\\mathrm {r e a}} ^ {\\mathrm {s}}} \\mathcal {L} _ {\\mathrm {r e a}} \\big (\\boldsymbol {\\theta} _ {\\mathrm {r e a}} ^ {\\mathrm {s}}, \\boldsymbol {\\theta} _ {\\mathrm {s y n}} ^ {\\mathrm {s}} \\big (\\boldsymbol {\\theta} _ {\\mathrm {r e a}} ^ {\\mathrm {s}} \\big); \\mathcal {D} _ {\\mathrm {r e a}} \\big),\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.573, + 0.695, + 0.905, + 0.728 + ], + "angle": 0, + "content": "\\[\ns. t., \\theta_ {\\text {s y n}} ^ {\\mathrm {s}} \\left(\\theta_ {\\text {r e a}} ^ {\\mathrm {s}}\\right) \\in \\arg \\min _ {\\theta_ {\\text {s y n}} ^ {\\mathrm {s}}} \\Psi \\left(\\theta_ {\\text {r e a}} ^ {\\mathrm {s}}, \\theta_ {\\text {s y n}} ^ {\\mathrm {s}}\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.737, + 0.56, + 0.75 + ], + "angle": 0, + "content": "where" + }, + { + "type": "equation", + "bbox": [ + 0.528, + 0.76, + 0.905, + 0.78 + ], + "angle": 0, + "content": "\\[\n\\Psi \\left(\\boldsymbol {\\theta} _ {\\mathrm {r e a}} ^ {\\mathrm {s}}, \\boldsymbol {\\theta} _ {\\mathrm {s y n}} ^ {\\mathrm {s}}\\right) = \\mathcal {L} _ {\\mathrm {r e a}} \\left(\\boldsymbol {\\theta} _ {\\mathrm {s y n}} ^ {\\mathrm {s}}; \\mathcal {D} _ {\\mathrm {r e a}}\\right) + \\mathcal {L} _ {1} \\left(\\boldsymbol {\\theta} _ {\\mathrm {s y n}} ^ {\\mathrm {s}}, \\boldsymbol {\\theta} _ {\\mathrm {r e a}} ^ {\\mathrm {s}}\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.789, + 0.906, + 0.834 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_{\\mathrm{rea}}\\) and \\(\\mathcal{L}_1\\) represent the loss constraints for real domain and alignment. The datasets \\(\\mathcal{D}_{\\mathrm{rea}}\\) and \\(\\mathcal{D}_{\\mathrm{syn}}\\) are captured in the real and synthetic domains, respectively." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.835, + 0.906, + 0.88 + ], + "angle": 0, + "content": "Actually, suppose an effective loss \\(\\mathcal{L}_{\\mathrm{rea}}\\) could be designed to constrain the output in the real domain, the adaptation might be achieved solely through the upper-level model" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.887, + 0.441, + 0.9 + ], + "angle": 0, + "content": "1Details of the specific loss functions can be found in Sec. 2.4." + }, + { + "type": "page_footnote", + "bbox": [ + 0.531, + 0.887, + 0.779, + 0.9 + ], + "angle": 0, + "content": "2Their specific forms can be found in Sec. 2.4." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11199" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.107, + 0.089, + 0.891, + 0.31 + ], + "angle": 0, + "content": "
Testing BenchmarkMetricsStability across various synthetic domainFlexibility with different dehazing model
RESIDEHaze4KTHazeMSBDNDehazeFormerDEA
BaseCoABaseCoABaseCoABaseCoABaseCoABaseCoA
RTTS [22]FADE↓1.47531.26871.73681.08501.20780.85941.33541.15561.18660.93251.21170.9332
PM2.5↓163.77139.04170.03110.84131.7788.871135.6895.793126.77115.00121.9889.813
Entropy↑7.29637.34587.34467.46377.43197.57987.39687.53827.48157.55117.46237.5426
BIQME↑0.55620.56390.55110.57480.57730.59320.56960.58190.57330.59280.57890.5860
URHI [22]FADE↓1.44631.30131.72631.02781.31460.92721.34880.95491.32631.17261.27961.0321
PM2.5↓164.36145.01172.98119.97140.3598.793146.4399.856141.11125.87133.97108.94
Entropy↑7.28677.31527.37117.46577.46527.59287.41447.55747.50097.58177.49207.5658
BIQME↑0.54990.55420.54810.56810.58210.59610.57270.58320.57570.59480.57860.5863
FATTAL [6]FADE↓0.51360.47780.58280.37340.42030.31420.43150.33270.42710.38750.42700.3976
PM2.5↓81.73880.28981.401110.9762.31291.63969.272104.5678.09360.69381.38780.723
Entropy↑7.38477.45757.39997.50067.44827.58587.41217.54607.45667.56107.43157.5569
BIQME↑0.56450.58020.55540.59190.59450.61890.58070.60550.59270.61170.59030.6097
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.313, + 0.908, + 0.342 + ], + "angle": 0, + "content": "Table 1. Quantitative comparison of algorithmic properties. In the left portion, our designed architecture serves as the baseline teacher model. In the right portion, the THaze dataset is used as the synthetic domain." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.353, + 0.484, + 0.445 + ], + "angle": 0, + "content": "(initialized with \\(\\theta_{\\mathrm{syn}}^{\\mathrm{s}}\\)). However, as discussed in Sec. 2.1, the challenge of accurately modeling real haze distributions makes this process difficult to implement. To this end, we introduce lower-level sub-optimization to ensure a feasible solution and maintain the correct optimization trajectory. Experimental validation can be seen in Sec. 5.2." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.451, + 0.321, + 0.466 + ], + "angle": 0, + "content": "2.3.2. Bilevel Adaptive Learning" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.469, + 0.484, + 0.591 + ], + "angle": 0, + "content": "Without compromising the model's dehazing capability in the synthetic domain, narrowing the performance gap between the synthetic and real domains is a central focus. Inspired by the Exponential Moving Average (EMA) method [2], we draw on its smoothing characteristics during the training process and introduce a dynamic weight adjustment mechanism to achieve a balance in learning between the synthetic and real domains, formulated as" + }, + { + "type": "equation", + "bbox": [ + 0.116, + 0.599, + 0.484, + 0.656 + ], + "angle": 0, + "content": "\\[\n\\left\\{ \\begin{array}{l} \\boldsymbol {\\theta} _ {\\text {s y n}} ^ {\\mathrm {s}} (t + 1) = \\boldsymbol {\\theta} _ {\\text {s y n}} ^ {\\mathrm {s}} (t) - \\eta_ {\\text {s y n}} ^ {b} \\frac {\\partial \\left(\\mathcal {L} _ {\\text {r e a}} + \\mathcal {L} _ {1}\\right)}{\\partial \\boldsymbol {\\theta} _ {\\text {s y n}} ^ {\\mathrm {s}}}, \\\\ \\boldsymbol {\\theta} _ {\\text {r e a}} ^ {\\mathrm {s}} (t + 1) = \\alpha \\boldsymbol {\\theta} _ {\\text {r e a}} ^ {\\mathrm {s}} (t) + (1 - \\alpha) \\boldsymbol {\\theta} _ {\\text {s y n}} ^ {\\mathrm {s}} (t + 1), \\end{array} \\right. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.664, + 0.484, + 0.771 + ], + "angle": 0, + "content": "where \\(\\alpha\\) represents the moving average coefficient, and \\(\\mathcal{L}_1\\) denotes the \\(\\ell_1\\)-norm supervised loss used to reduce the risk of the model overfitting to the real domain. Specifically, during each update, \\(\\theta_{\\mathrm{rea}}^{\\mathrm{s}}\\) is adjusted by \\(\\theta_{\\mathrm{syn}}^{\\mathrm{s}}\\), allowing the model to strike a balance between the synthetic and real domains throughout the learning process. The overall algorithm can be seen in Alg. 1." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.778, + 0.441, + 0.794 + ], + "angle": 0, + "content": "2.4. Network Architecture and Training Loss" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.8, + 0.483, + 0.831 + ], + "angle": 0, + "content": "This section introduces the details about network architecture and training loss adopted in this work." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.837, + 0.291, + 0.851 + ], + "angle": 0, + "content": "2.4.1. Network Architecture" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.902 + ], + "angle": 0, + "content": "We emphasize that our primary focus is on developing an effective learning mechanism to enhance the model's adaptability to the real domain. Notably, our approach offers" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.353, + 0.906, + 0.415 + ], + "angle": 0, + "content": "flexibility3 with different dehazing models. To this end, we adopt a hybrid architectural design, incorporating a pretrained Res2Net encoder [10] with the decoder of multiscale boosted dehazing network [5]." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.422, + 0.767, + 0.436 + ], + "angle": 0, + "content": "2.4.2. Loss Functions in MoC Phase" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.442, + 0.906, + 0.488 + ], + "angle": 0, + "content": "In the MoC process, to enable the student model to learn the features extracted by the teacher model more effectively and to enhance the quality the generated images. We define" + }, + { + "type": "equation", + "bbox": [ + 0.593, + 0.501, + 0.907, + 0.517 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {s y n}} = \\lambda_ {s u} \\ell_ {s u} + \\lambda_ {s s} \\ell_ {s s} + \\lambda_ {p e} \\ell_ {p e}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.529, + 0.906, + 0.576 + ], + "angle": 0, + "content": "where \\(\\ell_{su},\\ell_{ss},\\ell_{pe}\\) denote \\(\\ell_1\\) -norm supervised loss, SSIM loss, and perceptual loss, respectively. The coefficient \\(\\lambda_{su},\\lambda_{ss},\\lambda_{pe}\\) are positive balancing parameters." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.576, + 0.906, + 0.635 + ], + "angle": 0, + "content": "To more effectively enhance consistency between the student and teacher models in the feature space, the design of \\(\\mathcal{L}_{\\mathrm{a}}\\) incorporates a similarity measure of feature distributions. Its formulation can be written as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.614, + 0.645, + 0.907, + 0.686 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {a}} = \\sum_ {i = 0} ^ {L - 1} w _ {i} \\cdot \\left(\\left(T _ {i} - S _ {i}\\right) ^ {2}\\right), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.697, + 0.906, + 0.728 + ], + "angle": 0, + "content": "where \\(w_{i}\\) represents the weights of \\(i\\)-th layer in the teacher feature map \\(T_{i}\\) and the student feature map \\(S_{i}\\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.736, + 0.759, + 0.75 + ], + "angle": 0, + "content": "2.4.3. Loss Functions in BiA Phase" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.756, + 0.907, + 0.877 + ], + "angle": 0, + "content": "Here we incorporate the CLIP model to construct the training loss \\(\\mathcal{L}_{\\mathrm{rea}}\\) to constrain the output in the real domain. We select haze images \\(I_{\\mathcal{H}}\\) and clear images \\(I_{\\mathcal{C}}\\) as references, initialize corresponding prompts \\(T_{\\mathcal{H}}\\) and \\(T_{\\mathcal{C}}\\), and input them into the CLIP text encoder, while feeding the images into the image encoder. The sample categories are predicted using the text-image similarity in the CLIP space. To minimize the classification error between positive and negative" + }, + { + "type": "page_footnote", + "bbox": [ + 0.531, + 0.887, + 0.754, + 0.9 + ], + "angle": 0, + "content": "3For further details, please refer to Sec. 3." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11200" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.108, + 0.089, + 0.49, + 0.207 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.108, + 0.211, + 0.49, + 0.33 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.088, + 0.888, + 0.207 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.211, + 0.888, + 0.33 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.114, + 0.338, + 0.881, + 0.352 + ], + "angle": 0, + "content": "Figure 3. Quantitative comparison on three real-world datasets. Four no-reference image quality assessments were calculated." + }, + { + "type": "image", + "bbox": [ + 0.104, + 0.359, + 0.9, + 0.501 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.105, + 0.509, + 0.89, + 0.523 + ], + "angle": 0, + "content": "Figure 4. Qualitative comparisons on daytime haze and dusty scenes. All these observations come from RTTS and URHI datasets." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.54, + 0.483, + 0.601 + ], + "angle": 0, + "content": "samples, we employ the binary cross entropy loss to optimize the text prompt pair. After training the text prompts to effectively distinguish between real haze and clear images, we can derive \\(\\mathcal{L}_{\\mathrm{rea}}\\) based on the contrastive similarity loss:" + }, + { + "type": "equation", + "bbox": [ + 0.135, + 0.612, + 0.483, + 0.662 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {r e a}} = \\frac {e ^ {\\cos \\left(\\Phi_ {\\text {i m a g e}} \\left(I _ {\\mathcal {R}}\\right) , \\Phi_ {\\text {t e x t}} \\left(T _ {\\mathcal {H}}\\right)\\right)}}{\\sum_ {i \\in \\{\\mathcal {H} , \\mathcal {C} \\}} e ^ {\\cos \\left(\\Phi_ {\\text {i m a g e}} \\left(I _ {\\mathcal {R}}\\right) , \\Phi_ {\\text {t e x t}} \\left(T _ {i}\\right)\\right)}}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.673, + 0.483, + 0.704 + ], + "angle": 0, + "content": "where \\(I_{\\mathcal{R}}\\) corresponds to the real haze image processed by the model obtained through the model compression process." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.735, + 0.385, + 0.753 + ], + "angle": 0, + "content": "3. Exploring Algorithmic Property" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.76, + 0.484, + 0.837 + ], + "angle": 0, + "content": "CoA presents an effective learning strategy to improve efficiency and adaptability for real-world scenes, without assumptions on the synthetic domain or teacher model. We analyze CoA's properties, highlighting its stability across synthetic domains and flexibility with different models." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.848, + 0.458, + 0.865 + ], + "angle": 0, + "content": "3.1. Stability across Various Synthetic Domains" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.902 + ], + "angle": 0, + "content": "We perform three synthetic benchmarks (RESIDE [21] (the outdoor subset), Haze4K [30], and THaze [8]) and" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.539, + 0.907, + 0.706 + ], + "angle": 0, + "content": "three real-world datasets (RTTS [21], URHI [21], and FAT-TAL [6]) to assess our stability. As shown in the left section of Table 1, CoA consistently improves scores across four no-reference image quality metrics. Notably, due to the diverse scenes in THaze, CoA achieves near-optimal performance within the synthetic domain. We use THaze as the synthetic domain in our comparative experiments. Fig. 2 (left) compares visual results. The dehazing effects of CoA is significant, especially in distant regions highlighted by arrows. Overall, CoA improves the original teacher model across synthetic domains, demonstrating its stability." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.713, + 0.878, + 0.73 + ], + "angle": 0, + "content": "3.2. Flexibility with Different Dehazing Models" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.907, + 0.901 + ], + "angle": 0, + "content": "CoA's teacher model-irrelevant nature allows us to enhance the adaptability of existing dehazing models across various scenes. We applied CoA to three dehazing models (MSBDN [5], DehazeFormer [34], and DEA [3]) as teacher models. The right section of Table 1 shows significant performance improvements before and after applying CoA across different models and test settings. Fig. 2 (right) compares visual results using DehazeFormer and DEA, highlighting areas where CoA enhances dehazing for more visually appealing outcomes. In summary, CoA demonstrates notable flexibility with various dehazing models. The intro" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "11201" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.1, + 0.089, + 0.897, + 0.18 + ], + "angle": 0, + "content": "
MethodRTTSURHIFATTAL
FADE↓PM2.5↓Entropy↑BIQME↑FADE↓PM2.5↓Entropy↑BIQME↑FADE↓PM2.5↓Entropy↑BIQME↑
AirNet1.3342155.847.29590.54711.6591143.026.91430.52580.406376.1147.35450.5315
WeatherDiff2.4102193.277.14430.51552.1429180.267.14190.51020.8749118.797.36210.4937
DiffUIR2.1305189.617.17220.53031.9879180.567.18110.52050.8405116.687.36090.5049
Ours0.859488.8717.57980.59320.927298.7937.59280.59610.314291.6397.58580.6189
" + }, + { + "type": "table_caption", + "bbox": [ + 0.102, + 0.183, + 0.894, + 0.199 + ], + "angle": 0, + "content": "Table 2. Quantitative comparison on daytime haze scenes. All compared methods are designed for multi-weather image restoration." + }, + { + "type": "image", + "bbox": [ + 0.104, + 0.212, + 0.222, + 0.329 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.212, + 0.301, + 0.329 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.303, + 0.212, + 0.421, + 0.329 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.422, + 0.212, + 0.5, + 0.329 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.212, + 0.619, + 0.329 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.621, + 0.212, + 0.698, + 0.329 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.699, + 0.212, + 0.818, + 0.329 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.82, + 0.212, + 0.898, + 0.329 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.138, + 0.339, + 0.856, + 0.353 + ], + "angle": 0, + "content": "Figure 5. Qualitative comparisons corresponding to Table 2. All examples are sourced from RTTS and URHI datasets." + }, + { + "type": "image", + "bbox": [ + 0.104, + 0.367, + 0.215, + 0.495 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.216, + 0.367, + 0.317, + 0.495 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.318, + 0.367, + 0.419, + 0.495 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.421, + 0.367, + 0.473, + 0.494 + ], + "angle": 0, + "content": null + }, + { + "type": "table_caption", + "bbox": [ + 0.218, + 0.497, + 0.36, + 0.509 + ], + "angle": 0, + "content": "(a) Qualitative comparison" + }, + { + "type": "table", + "bbox": [ + 0.1, + 0.51, + 0.472, + 0.579 + ], + "angle": 0, + "content": "
MetricsGS (ICCV'15)MRP (CVPR'17)OSFD (ACM MM'20)GAPSF (ACM MM'23)CoA
PM2.5↓102.43106.20146.93110.5981.141
Entropy↑6.70257.04886.94126.37696.9907
BIQME↑0.45520.47220.46840.38950.5112
" + }, + { + "type": "table_caption", + "bbox": [ + 0.212, + 0.58, + 0.36, + 0.591 + ], + "angle": 0, + "content": "(b) Quantitative comparison" + }, + { + "type": "image_caption", + "bbox": [ + 0.102, + 0.597, + 0.471, + 0.611 + ], + "angle": 0, + "content": "Figure 6. Performance evaluation on nighttime haze scenes." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.63, + 0.484, + 0.736 + ], + "angle": 0, + "content": "duction of CoA significantly reduces the model's parameter count, improving computational efficiency. Specifically, the parameters in the three dehazing models decrease by \\(94.06\\%\\), \\(86.74\\%\\), and \\(73.89\\%\\), respectively. This reduction accelerates inference speed and lowers resource requirements, making the models more lightweight and better suited for practical applications." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.749, + 0.295, + 0.765 + ], + "angle": 0, + "content": "4. Experimental Results" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.773, + 0.307, + 0.789 + ], + "angle": 0, + "content": "4.1. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Training settings. The CoA model was implemented using the PyTorch framework on a single NVIDIA RTX 3090 GPU. The Adam optimizer was employed to update each module's parameters, with \\(\\beta_{1},\\beta_{2}\\), and \\(\\varepsilon\\) set to 0.9, 0.999, and \\(1e^{-8}\\), respectively. The initial learning rate was set to \\(1e^{-4}\\) and gradually decayed to \\(1e^{-6}\\) using a cosine annealing schedule. The algorithm was trained on RGB channels" + }, + { + "type": "image", + "bbox": [ + 0.526, + 0.367, + 0.6, + 0.509 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.601, + 0.367, + 0.674, + 0.508 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.675, + 0.367, + 0.748, + 0.509 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.75, + 0.367, + 0.822, + 0.508 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.823, + 0.367, + 0.897, + 0.508 + ], + "angle": 0, + "content": null + }, + { + "type": "table_caption", + "bbox": [ + 0.641, + 0.512, + 0.782, + 0.523 + ], + "angle": 0, + "content": "(a) Qualitative comparison" + }, + { + "type": "table", + "bbox": [ + 0.526, + 0.525, + 0.892, + 0.58 + ], + "angle": 0, + "content": "
MetricsDM(ACM MM'23)SUIR(CVPR'23)SMDR(AAAI'24)CoA
UIQM↑4.09653.98534.03144.1284
CCF↑26.65131.41728.43930.878
" + }, + { + "type": "table_caption", + "bbox": [ + 0.635, + 0.58, + 0.783, + 0.592 + ], + "angle": 0, + "content": "(b) Quantitative comparison" + }, + { + "type": "image_caption", + "bbox": [ + 0.534, + 0.598, + 0.883, + 0.612 + ], + "angle": 0, + "content": "Figure 7. Performance evaluation on underwater scenes." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.629, + 0.905, + 0.703 + ], + "angle": 0, + "content": "with data augmentation via random \\(90^{\\circ}\\), \\(180^{\\circ}\\), and \\(270^{\\circ}\\) rotations and horizontal flipping. We cropped random \\(256 \\times 256\\) sub-regions from images, further expanding the training set. Using a pre-trained Res2Net encoder, the teacher model achieved strong performance after just 20 epochs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.906, + 0.825 + ], + "angle": 0, + "content": "Benchmarks and metrics. Three synthetic datasets (Haze4K, RESIDE (the outdoor subset), and THaze served as the synthetic domain, while three real-world datasets (RTTS, Fattal, and URHI were used for testing in the real domain. We employed four objective metrics for realworld haze scenes: FADE [4], PM2.5 [14], Entropy, and BIQME [13]. For underwater enhancement, we used UIQM and CCF to assess image quality." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Compared methods. We evaluated our CoA method against state-of-the-art techniques across daytime, dusty, nighttime, and underwater conditions. For daytime and dusty conditions, we compared CoA with SGID [1], Dehamer [16], C2P [46], KANet [7], DEA, D4 [38], \\(\\mathrm{D4^{+}}\\) [39]," + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11202" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.123, + 0.089, + 0.875, + 0.202 + ], + "angle": 0, + "content": "
MetricsSGID (TIP'22)Dehamer (CVPR'22)C2P (CVPR'23)\\( \\text{RIDCP}^\\dagger \\)(CVPR'23)D4 (CVPR'22)D4+ (IJCV'24)KANet (TPAMI'24)DEA (TIP'24)Ours
SIZE (M)13.87132.407.1728.7210.7010.7055.253.651.69
FLOPs (G)108.4048.91352.90144.432.822.824.4232.202.67
Time (ms)1280×720240.01136.361169.18630.6451.2551.2522.9386.3324.33
1920×1080878.94295.982531.431588.26105.31105.31205.40190.7852.52
2560×14401514.82519.524409.12—185.55185.5510631334.7892.13
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.133, + 0.202, + 0.515, + 0.212 + ], + "angle": 0, + "content": "This work requires excessive video memory for high-resolution images, resulting in the failure to produce results." + }, + { + "type": "table_caption", + "bbox": [ + 0.126, + 0.215, + 0.87, + 0.23 + ], + "angle": 0, + "content": "Table 3. Evaluating the computational efficiency. Note that the size of the testing image for FLOPs calculation is \\( {224} \\times {224} \\) ." + }, + { + "type": "image", + "bbox": [ + 0.114, + 0.248, + 0.462, + 0.354 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.358, + 0.359, + 0.37 + ], + "angle": 0, + "content": "(a) Qualitative comparison" + }, + { + "type": "table", + "bbox": [ + 0.114, + 0.371, + 0.459, + 0.422 + ], + "angle": 0, + "content": "
ModelFADE↓PM2.5↓BIQME↑SIZE (M)FLOPs (G)
Naive for Teacher1.21131.770.5852.8318.56
Naive for Student1.23135.760.571.692.67
MoC for Student1.13124.430.581.692.67
" + }, + { + "type": "image_caption", + "bbox": [ + 0.211, + 0.423, + 0.361, + 0.434 + ], + "angle": 0, + "content": "(b) Quantitative comparison" + }, + { + "type": "image_caption", + "bbox": [ + 0.187, + 0.439, + 0.386, + 0.453 + ], + "angle": 0, + "content": "Figure 8. Effects of MoC phase." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.471, + 0.484, + 0.562 + ], + "angle": 0, + "content": "and RIDCP [37]. Additionally, we assessed CoA against multi-weather restoration models: AirNet [23], WeatherDiff [33], and DiffUIR [45]. For nighttime dehazing, we benchmarked against GS [25], MRP [42], OSFD [43], and GAPSF [19], and for underwater scenes, we compared with DM [35], SUIR [18], and SMDR [40]." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.575, + 0.357, + 0.59 + ], + "angle": 0, + "content": "4.2. Evaluation on Regular Scenes" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.598, + 0.484, + 0.687 + ], + "angle": 0, + "content": "Quantitative comparisons. In Fig. 3, we presented a comparison of eight state-of-the-art dehazing algorithms with our proposed method, using four authoritative unsupervised evaluation metrics. It is evident that our method ranked among the top in almost all metrics, achieving optimal or near-optimal performance in most cases." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.484, + 0.81 + ], + "angle": 0, + "content": "Qualitative comparisons. As shown in Fig. 4, the supervised methods (SGID, C2P, Dehamer, DEA) and unsupervised methods \\((\\mathrm{D}4,\\mathrm{D}4^{+})\\) struggle with haze removal and generalization. RIDCP causes color distortion and overbehazing, especially in sandstorm and colorful haze scenes. KANet has limitations in detail and color restoration. In contrast, CoA excels at haze removal while preserving fine textures, resulting in more natural and realistic images." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Comparisons with multi-weather restorers. Here we compared CoA with state-of-the-art multi-weather degradation methods. As shown in Table 2, CoA ranks second in one metric, outperforming others in all others, highlighting its superiority. Fig. 5 reveals that while existing methods excel in specific scenarios, they show limitations in handling" + }, + { + "type": "image", + "bbox": [ + 0.533, + 0.246, + 0.706, + 0.325 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.606, + 0.326, + 0.631, + 0.335 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image", + "bbox": [ + 0.71, + 0.246, + 0.883, + 0.324 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.731, + 0.326, + 0.859, + 0.336 + ], + "angle": 0, + "content": "Only Lower-Level Modeling" + }, + { + "type": "image", + "bbox": [ + 0.533, + 0.338, + 0.705, + 0.416 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.555, + 0.417, + 0.681, + 0.428 + ], + "angle": 0, + "content": "Only Upper-Level Modeling" + }, + { + "type": "image", + "bbox": [ + 0.71, + 0.338, + 0.883, + 0.416 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.756, + 0.417, + 0.834, + 0.428 + ], + "angle": 0, + "content": "Bilevel Modeling" + }, + { + "type": "image_caption", + "bbox": [ + 0.587, + 0.436, + 0.831, + 0.451 + ], + "angle": 0, + "content": "Figure 9. Necessity of bilevel modeling." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.469, + 0.756, + 0.484 + ], + "angle": 0, + "content": "the full spectrum of haze conditions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.486, + 0.907, + 0.637 + ], + "angle": 0, + "content": "Efficiency comparisons. Table 3 presents a comprehensive comparison of our proposed CoA method with existing approaches, evaluating parameters, floating-point operations per second (FLOPs), and processing time across various image resolutions. The results clearly show that our method has the smallest number of parameters and FLOPs among all the evaluated methods. Notably, as the image resolution increases, our method demonstrates a significant advantage in processing time, indicating that our model achieves a favorable balance between performance and complexity." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.65, + 0.741, + 0.667 + ], + "angle": 0, + "content": "4.3. Adaptability Verification" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.673, + 0.907, + 0.809 + ], + "angle": 0, + "content": "Nighttime haze scene. Compared to state-of-the-art nighttime dehazing algorithms on the NHRW dataset (Fig. 6 (a)), many existing methods fail to handle overexposure or low-light conditions, resulting in overly dark or distorted images due to noise and artificial light scattering. In contrast, CoA excels in real-world nighttime haze scenes, effectively removing haze, suppressing noise, and preserving details. As shown in Fig. 6 (b), our method outperforms others across three authoritative objective metrics" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Underwater scene. Compared to state-of-the-art underwater dehazing algorithms, DM results in reddish and blurry images, failing to correct color distortion effectively, while SUIR and SMDR struggle to fully eliminate haze. As shown in Fig. 7, our method significantly reduces haze impact and enhances clarity and detail visibility, closely align-" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11203" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.104, + 0.087, + 0.478, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.186, + 0.258, + 0.386, + 0.273 + ], + "angle": 0, + "content": "Figure 10. Effects of BiA phase." + }, + { + "type": "image", + "bbox": [ + 0.104, + 0.28, + 0.476, + 0.381 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.096, + 0.391, + 0.476, + 0.407 + ], + "angle": 0, + "content": "Figure 11. Parameters Analysis. The RTTS dataset is adopted." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.426, + 0.384, + 0.441 + ], + "angle": 0, + "content": "ing with human visual perception standards." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.455, + 0.314, + 0.472 + ], + "angle": 0, + "content": "5. Analytical Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.481, + 0.294, + 0.495 + ], + "angle": 0, + "content": "5.1. Effects of MoC Phase" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.503, + 0.484, + 0.655 + ], + "angle": 0, + "content": "Without feature transfer from the \"naive for teacher\" (naive: end-to-end training, the same below) model via compression, the \"naive for student\" model struggles to capture core data patterns, particularly in complex or extreme weather scenarios, resulting in limited dehazing. As illustrated in Fig. 8, while \"MoC for student\" may not restore details as effectively as \"naive for teacher\", it successfully inherits high-quality features. Even under resource constraints, it surpasses \"naive for student\" in dehazing and performs well on unsupervised metrics." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.664, + 0.353, + 0.681 + ], + "angle": 0, + "content": "5.2. Necessity of Bilevel Modeling" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.687, + 0.484, + 0.839 + ], + "angle": 0, + "content": "Here we highlight the critical role of bilevel modeling by training the same network model on the THaze dataset. As shown in Fig. 9, the results on the real test set show marked contrasts: employing only the lower-level modeling strategy results in suboptimal dehazing, while relying solely on the upper-level strategy with CLIP fine-tuning produces unnatural effects with notable color distortion. Different from them, the bilevel modeling strategy enables both effective dehazing and natural visual quality, achieving a balance between dehazing performance and perceptual realism." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.848, + 0.285, + 0.863 + ], + "angle": 0, + "content": "5.3. Effects of BiA Phase" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.901 + ], + "angle": 0, + "content": "In Fig. 10, we evaluate effects of the BiA phase. The results clearly show significant improvements in both FADE and" + }, + { + "type": "image", + "bbox": [ + 0.528, + 0.089, + 0.898, + 0.173 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.529, + 0.184, + 0.89, + 0.199 + ], + "angle": 0, + "content": "Figure 12. Limitations. The example is from RS-Haze [34]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.22, + 0.906, + 0.295 + ], + "angle": 0, + "content": "PM2.5 metrics following the BiA phase. Notably, in images with complex backgrounds or high noise levels, our model outperforms the baseline, demonstrating the BiA phase's effectiveness in capturing key features and improving robustness in challenging scenarios." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.307, + 0.712, + 0.323 + ], + "angle": 0, + "content": "5.4. Parameters Analyses" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.33, + 0.906, + 0.421 + ], + "angle": 0, + "content": "We test different values of the parameter \\(\\alpha\\) during the BiA phase, as shown in Fig. 11. The results indicate that \\(\\alpha = 0.95\\) yields almost optimal performance. Notably, this setting achieves the fastest convergence and exhibits superior stability and accuracy across multiple evaluation metrics, clearly outperforming other parameter settings." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.432, + 0.64, + 0.446 + ], + "angle": 0, + "content": "5.5. Limitations" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.455, + 0.906, + 0.561 + ], + "angle": 0, + "content": "While effective in many scenarios, our method struggles with remote sensing haze images due to additional imaging constraints. As depicted in Fig. 12, although it improves dehazing (outperforming DEA), non-uniform fog occlusion remains. This stems from varying haze effects across spectral bands in hyperspectral images, requiring multi-band integration for effective dehazing." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.577, + 0.634, + 0.592 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.602, + 0.907, + 0.769 + ], + "angle": 0, + "content": "This work proposes a novel compression-and-adaptation scheme to tackle the problem of unbalance between dehazing quality and computational efficiency in real image dehazing. Following the divide-and-conquer paradigm, we design a new CoA learning strategy that first compresses the predefined model and then adapts it to unlabeled, diverse real-world scenes. We not only make a thorough exploration to take on the excellent properties of CoA, but also we perform extensive experiments to indicate our CoA method's superiority in terms of image dehazing quality and computational efficiency." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.785, + 0.673, + 0.802 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.906, + 0.9 + ], + "angle": 0, + "content": "This work was supported by the Postdoctoral Fellowship Program of CPSF (No. GZB20240098), China Postdoctoral Science Foundation (No. 2023M740491), Dalian Science and Technology Innovation Fund-Young Tech Star (No. 2023RQ017), Fundamental Research Funds for the Central Universities (No. DUT24BS011)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11204" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.17 + ], + "angle": 0, + "content": "[1] Haoran Bai, Jinshan Pan, Xinguang Xiang, and Jinhui Tang. Self-guided image dehazing using progressive feature fusion. IEEE Transactions on Image Processing, 31:1217-1229, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.173, + 0.484, + 0.255 + ], + "angle": 0, + "content": "[2] Zhaowei Cai, Avinash Ravichandran, Subhransu Maji, Charless Fowlkes, Zhuowen Tu, and Stefano Soatto. Exponential moving average normalization for self-supervised and semi-supervised learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 194-203, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.258, + 0.483, + 0.313 + ], + "angle": 0, + "content": "[3] Zixuan Chen, Zewei He, and Zhe-Ming Lu. Dea-net: Single image dehazing based on detail-enhanced convolution and content-guided attention. IEEE Transactions on Image Processing, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.316, + 0.483, + 0.37 + ], + "angle": 0, + "content": "[4] Lark Kwon Choi, Jaehee You, and Alan C. Bovik. Referenceless prediction of perceptual fog density and perceptual image defogging. IEEE Transactions on Image Processing, 24(11):3888-3901, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.373, + 0.484, + 0.442 + ], + "angle": 0, + "content": "[5] Hang Dong, Jinshan Pan, Lei Xiang, Zhe Hu, Xinyi Zhang, Fei Wang, and Ming-Hsuan Yang. Multi-scale boosted dehazing network with dense feature fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2157-2167, 2020. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.444, + 0.483, + 0.471 + ], + "angle": 0, + "content": "[6] Raanan Fattal. Dehazing using color-lines. ACM transactions on graphics, 34(1):1-14, 2014. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.474, + 0.483, + 0.542 + ], + "angle": 0, + "content": "[7] Yuxin Feng, Long Ma, Xiaozhe Meng, Fan Zhou, Risheng Liu, and Zhuo Su. Advancing real-world image dehazing: Perspective, modules, and training. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(12):9303-9320, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.545, + 0.483, + 0.613 + ], + "angle": 0, + "content": "[8] Yuxin Feng, Zhuo Su, Long Ma, Xin Li, Risheng Liu, and Fan Zhou. Bridging the gap between haze scenarios: A unified image dehazing model. IEEE Transactions on Circuits and Systems for Video Technology, 34(11):11070-11085, 2024. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.616, + 0.483, + 0.685 + ], + "angle": 0, + "content": "[9] Luca Franceschi, Paolo Frasconi, Saverio Salzo, Riccardo Grazzi, and Massimiliano Pontil. Bilevel programming for hyperparameter optimization and meta-learning. In Proceedings of the International Conference on Machine Learning, pages 1568-1577, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.687, + 0.483, + 0.757 + ], + "angle": 0, + "content": "[10] Minghan Fu, Huan Liu, Yankun Yu, Jun Chen, and Keyan Wang. DW-GAN: A discrete wavelet transform GAN for nonhomogeneous dehazing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 203-212, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.758, + 0.483, + 0.814 + ], + "angle": 0, + "content": "[11] Xueyang Fu, Jie Xiao, Yurui Zhu, Aiping Liu, Feng Wu, and Zheng-Jun Zha. Continual image deraining with hypergraph convolutional networks. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(8):9534-9551, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.816, + 0.483, + 0.87 + ], + "angle": 0, + "content": "[12] Alona Golts, Daniel Freedman, and Michael Elad. Unsupervised single image dehazing using dark channel prior loss. IEEE Transactions on Image Processing, 29:2692-2701, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[13] Ke Gu, Dacheng Tao, Junfei Qiao, and Weisi Lin. Learning a no-reference quality assessment model of enhanced images" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.548, + 0.093, + 0.906, + 0.12 + ], + "angle": 0, + "content": "with big data. IEEE Transactions on Neural Networks and Learning Systems, 29(4):1301-1313, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.122, + 0.906, + 0.162 + ], + "angle": 0, + "content": "[14] Ke Gu, Junfei Qiao, and Xiaoli Li. Highly efficient picture-based prediction of PM2.5 concentration. IEEE Transactions on Industrial Electronics, 66(4):3176-3184, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.164, + 0.906, + 0.218 + ], + "angle": 0, + "content": "[15] Jie Gui, Xiaofeng Cong, Lei He, Yuan Yan Tang, and James Tin-Yau Kwok. Illumination controllable dehazing network based on unsupervised retinae embedding. IEEE Transactions on Multimedia, 26:4819-4830, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.219, + 0.906, + 0.288 + ], + "angle": 0, + "content": "[16] Chunle Guo, Qixin Yan, Saeed Anwar, Runmin Cong, Wenqi Ren, and Chongyi Li. Image dehazing transformer with transmission-aware 3d position embedding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5802-5810, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.289, + 0.906, + 0.342 + ], + "angle": 0, + "content": "[17] Kaiming He, Jian Sun, and Xiaou Tang. Single image haze removal using dark channel prior. IEEE Transactions on Pattern Analysis and Machine Intelligence, 33(12):2341-2353, 2010. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.344, + 0.908, + 0.413 + ], + "angle": 0, + "content": "[18] Shirui Huang, Keyan Wang, Huan Liu, Jun Chen, and Yun-song Li. Contrastive semi-supervised learning for underwater image restoration via reliable bank. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18145-18155, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.414, + 0.906, + 0.483 + ], + "angle": 0, + "content": "[19] Yeying Jin, Beibei Lin, Wending Yan, Yuan Yuan, Wei Ye, and Robby T. Tan. Enhancing visibility in nighttime haze images using guided APSF and gradient adaptive convolution. In Proceedings of the ACM International Conference on Multimedia, pages 2446-2457, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.484, + 0.906, + 0.537 + ], + "angle": 0, + "content": "[20] Mingye Ju, Can Ding, Charles A. Guo, Wenqi Ren, and Dacheng Tao. IDRLP: image dehazing using region line prior. IEEE Transactions on Image Processing, 30:9043-9057, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.539, + 0.906, + 0.594 + ], + "angle": 0, + "content": "[21] Boyi Li, Wenqi Ren, Dengpan Fu, Dacheng Tao, Dan Feng, Wenjun Zeng, and Zhangyang Wang. Benchmarking single-image dehazing and beyond. IEEE Transactions on Image Processing, 28(1):492-505, 2018. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.596, + 0.906, + 0.65 + ], + "angle": 0, + "content": "[22] Boyi Li, Wenqi Ren, Dengpan Fu, Dacheng Tao, Dan Feng, Wenjun Zeng, and Zhangyang Wang. Benchmarking single-image dehazing and beyond. IEEE Transactions on Image Processing, 28(1):492-505, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.651, + 0.906, + 0.718 + ], + "angle": 0, + "content": "[23] Boyun Li, Xiao Liu, Peng Hu, Zhongqin Wu, Jiancheng Lv, and Xi Peng. All-in-one image restoration for unknown corruption. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17431-17441, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.72, + 0.906, + 0.774 + ], + "angle": 0, + "content": "[24] Pengyue Li, Jiandong Tian, Yandong Tang, Guolin Wang, and Chengdong Wu. Deep retina network for single image dehazing. IEEE Transactions on Image Processing, 30: 1100-1115, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.776, + 0.906, + 0.83 + ], + "angle": 0, + "content": "[25] Yu Li, Robby T. Tan, and Michael S. Brown. Nighttime haze removal with glow and multiple light colors. In Proceedings of the International Conference on Computer Vision, pages 226-234, 2015. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.832, + 0.906, + 0.9 + ], + "angle": 0, + "content": "[26] Yudong Liang, Bin Wang, Wangmeng Zuo, Jiaying Liu, and Wenqi Ren. Self-supervised learning and adaptation for single image dehazing. In Proceedings of the Thirty-First International Joint Conference on Artificial Intelligence, pages 1137–1143, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.093, + 0.908, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11205" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[27] Pengyang Ling, Huaian Chen, Xiao Tan, Yi Jin, and Enhong Chen. Single image dehazing using saturation line prior. IEEE Transactions on Image Processing, 32:3238-3253, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.15, + 0.482, + 0.205 + ], + "angle": 0, + "content": "[28] Jun Liu, Ryan Wen Liu, Jianing Sun, and Tieyong Zeng. Rank-one prior: Real-time scene recovery. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(7): 8845-8860, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.207, + 0.482, + 0.275 + ], + "angle": 0, + "content": "[29] Risheng Liu, Jiaxin Gao, Jin Zhang, Deyu Meng, and Zhouchen Lin. Investigating bi-level optimization for learning and vision from a unified perspective: A survey and beyond. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(12):10045-10067, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.278, + 0.482, + 0.346 + ], + "angle": 0, + "content": "[30] Ye Liu, Lei Zhu, Shunda Pei, Huazhu Fu, Jing Qin, Qing Zhang, Liang Wan, and Wei Feng. From synthetic to real: Image dehazing collaborating with unlabeled real data. In Proceedings of the ACM International Conference on Multimedia, pages 50-58, 2021. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.349, + 0.482, + 0.417 + ], + "angle": 0, + "content": "[31] Yidi Liu, Xingbo Wang, Yurui Zhu, Xueyang Fu, and Zheng-Jun Zha. Sdcnet: Spatially-adaptive deformable convolution networks for hr nonhomogeneous dehazing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6682-6691, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.42, + 0.482, + 0.487 + ], + "angle": 0, + "content": "[32] Long Ma, Tengyu Ma, Risheng Liu, Xin Fan, and Zhongxuan Luo. Toward fast, flexible, and robust low-light image enhancement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5637-5646, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.49, + 0.482, + 0.545 + ], + "angle": 0, + "content": "[33] Ozan Özdenizci and Robert Legenstein. Restoring vision in adverse weather conditions with patch-based denoising diffusion models. IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 1-12, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.548, + 0.482, + 0.588 + ], + "angle": 0, + "content": "[34] Yuda Song, Zhuqing He, Hui Qian, and Xin Du. Vision transformers for single image dehazing. IEEE Transactions on Image Processing, 32:1927-1941, 2023. 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.591, + 0.482, + 0.659 + ], + "angle": 0, + "content": "[35] Yi Tang, Hiroshi Kawasaki, and Takafumi Iwaguchi. Underwater image enhancement by transformer-based diffusion model with non-uniform sampling for skip strategy. In Proceedings of the ACM International Conference on Multimedia, pages 5419-5427, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.661, + 0.482, + 0.73 + ], + "angle": 0, + "content": "[36] Zhongze Wang, Haitao Zhao, Jingchao Peng, Lujian Yao, and Kaijie Zhao. Odcr: Orthogonal decoupling contrastive regularization for unpaired image dehazing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25479-25489, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.732, + 0.482, + 0.801 + ], + "angle": 0, + "content": "[37] Ruiqi Wu, Zheng-Peng Duan, Chun-Le Guo, Zhi Chai, and Chongyi Li. RIDCP: revitalizing real image dehazing via high-quality codebook priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22282-22291, 2023. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.803, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[38] Yang Yang, Chaoyue Wang, Risheng Liu, Lin Zhang, Xiaojie Guo, and Dacheng Tao. Self-augmented unpaired image dehazing via density and depth decomposition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2037-2046, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.874, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[39] Yang Yang, Chaoyue Wang, Xiaojie Guo, and Dacheng Tao. Robust unpaired image dehazing via density and depth de" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "composition. International Journal of Computer Vision, 132 (5):1557-1577, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[40] Dehuan Zhang, Jingchun Zhou, Chunle Guo, Weishi Zhang, and Chongyi Li. Synergistic multiscale detail refinement via intrinsic supervision for underwater image enhancement. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 7033-7041, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.192, + 0.905, + 0.247 + ], + "angle": 0, + "content": "[41] He Zhang and Vishal M Patel. Densely connected pyramid dehazing network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3194-3203, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.249, + 0.905, + 0.304 + ], + "angle": 0, + "content": "[42] Jing Zhang, Yang Cao, and et al. Fast haze removal for nighttime image using maximum reflectance prior. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7016-7024, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.305, + 0.905, + 0.359 + ], + "angle": 0, + "content": "[43] Jing Zhang, Yang Cao, and et al. Nighttime dehazing with a synthetic benchmark. In Proceedings of the 28th ACM International Conference on Multimedia, pages 2355-2363, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.362, + 0.905, + 0.431 + ], + "angle": 0, + "content": "[44] Yafei Zhang, Shen Zhou, and Huafeng Li. Depth information assisted collaborative mutual promotion network for single image dehazing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2846-2855, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.433, + 0.905, + 0.502 + ], + "angle": 0, + "content": "[45] Dian Zheng, Xiao-Ming Wu, Shuzhou Yang, Jian Zhang, Jian-Fang Hu, and Wei-Shi Zheng. Selective hourglass mapping for universal image restoration based on diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.504, + 0.905, + 0.558 + ], + "angle": 0, + "content": "[46] Yu Zheng, Jiahui Zhan, and et al. Curricular contrastive regularization for physics-aware single image dehazing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition., 1, 6" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.558 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11206" + } + ] +] \ No newline at end of file diff --git a/2025/CoA_ Towards Real Image Dehazing via Compression-and-Adaptation/06a0b6a5-a641-41a6-99e5-3f3c8a8c1a70_origin.pdf b/2025/CoA_ Towards Real Image Dehazing via Compression-and-Adaptation/06a0b6a5-a641-41a6-99e5-3f3c8a8c1a70_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..108898dc3bd37bc653f5ce45ce1bb0167db37d16 --- /dev/null +++ b/2025/CoA_ Towards Real Image Dehazing via Compression-and-Adaptation/06a0b6a5-a641-41a6-99e5-3f3c8a8c1a70_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:627e8e91f89ea9a410883c06dccfc9e1bc5514499d705b8ddd07e0aab8b2ed53 +size 4190735 diff --git a/2025/CoA_ Towards Real Image Dehazing via Compression-and-Adaptation/full.md b/2025/CoA_ Towards Real Image Dehazing via Compression-and-Adaptation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..18a39856533d8b72d924a8b5dc3c46f90618c31d --- /dev/null +++ b/2025/CoA_ Towards Real Image Dehazing via Compression-and-Adaptation/full.md @@ -0,0 +1,432 @@ +# CoA: Towards Real Image Dehazing via Compression-and-Adaptation + +Long Ma + +Yuxin Feng $^{2}$ +Jiang-Yong Chou + +Yan Zhang + +Jinyuan Liu1 + +Weimin Wang + +Chengpei $\mathrm{Xu}^{4,*}$ + +Zhuo Su + +$^{1}$ Dalian University of Technology $^{2}$ Sun Yat-sen University $^{3}$ Fuzhou University $^{4}$ University of New South Wales malone94319@gmail.com, {fengyx26, zhangy2779} $@$ mail2.sysu.edu.cn, atlantis918@hotmail.com, chengpei.Xu@unsw.edu.au, wangweimin@dlut.edu.cn, gychen@fzu.edu.cn, suzhuo3@mail.sysu.edu.cn + +![](images/bfa2a24e2f256479b04856035ca0b551b569bc05431acec6d9768ac1352674b5.jpg) +(a) Evaluating efficiency across various metrics + +![](images/cf30f643d4fe7f77f26fdc98d832ff5cc0fe48bd8885d7eb6b5126dc2f79663f.jpg) +(b) The computational process of CoA + +![](images/67d32c0a664118b8f2f44a755cffe18c2ed788d072e7141723b0bcbd931155e6.jpg) +(c) Evaluating adaptability across different scenes +Figure 1. Performance evaluation. The proposed CoA incorporates model compression in synthetic domain for efficiency and bilevel adaptation in real domain for adaptability, as illustrated in the central sub-figure (b). The left sub-figure (a) presents efficiency across various metrics, clearly showing that our CoA outperforms others by a significant margin. The right sub-figure (c) shows adaptability across different scenes, where it is evident that our CoA consistently performs excellently. + +# Abstract + +Learning-based image dehazing algorithms have shown remarkable success in synthetic domains. However, real image dehazing is still in suspense due to computational resource constraints and the diversity of real-world scenes. Therefore, there is an urgent need for an algorithm that excels in both efficiency and adaptability to address real image dehazing effectively. This work proposes a Compression-and-Adaptation (CoA) computational flow to tackle these challenges from a divide-and-conquer perspective. First, model compression is performed in the synthetic domain to develop a compact dehazing parameter space, satisfying efficiency demands. Then, a bilevel adaptation in the real domain is introduced to be fearless in unknown real environments by aggregating the synthetic dehazing capabilities during the learning process. Leveraging a succinct design free from additional constraints, our CoA exhibits domain-irrelevant stability and model-agnostic flexibility, effectively bridging the model chasm between synthetic and real domains to further improve its practical utility. Extensive evaluations and analyses underscore the approach's superiority and effectiveness. The code is publicly available at https://github.com/fyxn1/CoA. + +# 1. Introduction + +Image dehazing tasks aim to utilize the atmospheric optical model and advanced computing theory to achieve efficient restoration of scene information and image details through inverse imaging solution or data-driven model learning. It has drawn much attention in multiple emerging computer vision areas recently [46] [27] [28]. Similar to other low-level vision tasks [32] [11], image dehazing technology has evolved from an early stage focused on improving synthetic data metrics to a direction aimed at effectively generalizing across various real-world haze scenarios. Current approaches face dual challenges: efficiency-oriented methods achieve real-time processing but lack scene adaptability, while adaptability-focused techniques dynamically adjust strategies at the cost of high computational complexity that hinders deployment in resource-constrained environments. This paper addresses these limitations through novel architectural innovations in both computational efficiency and adaptive scene understanding. + +# 1.1. Related Works + +Efficiency-oriented Dehazing Methods. This work focuses on enhancing computational efficiency through optimization algorithms, model simplification, or machine learning methods to enable faster processing under limited resources. Broadly, dehazing methods can be divided + +into image enhancement-based, physics-based, and image layering-based approaches. Image enhancement-based dehazing methods [24] treat the task as an image improvement problem, increasing contrast and color accuracy with techniques like histogram equalization and color correction. Physics-based methods [17], [28] are more specific, leveraging hazy image priors and constraints to aid inverse problem solving. Image layering-based approaches, such as that of Zhang et al. [41], increase efficiency by breaking down the image into scales or layers, progressively refining transmittance maps and enhancing detail using multilevel pyramid structures and dense connectivity. Despite their computational and real-time advantages, they struggle with stability and consistent quality, especially in complex real-world haze conditions, often resulting in partial dehazing and insufficient color clarity. + +Adaptability-focused Dehazing Methods. Adaptive dehazing methods tailor their strategies based on specific image attributes to improve dehazing effectiveness. These methods encompass three main types: local feature-based, atmospheric scattering model-based, and domain adaptation-based approaches. Local feature-based methods [15] adaptively adjust dehazing intensity according to local contrast and brightness variations, which helps restore image details effectively. Atmospheric scattering model-based adaptive dehazing methods [20] dynamically adjust the transmittance estimation by analyzing different regions of the image to achieve more accurate restoration. These methods can, to some extent, balance both efficiency and adaptability. Domain adaptation-based dehazing methods [26, 37] have gradually emerged in the field of image dehazing in recent years. These methods learn a shared feature representation, enabling the model to simultaneously adapt to the features of both the source domain and the target domain. Although these methods yield superior dehazing results, their complex models and dynamic imagespecific adjustments often increase computational costs, potentially impacting real-time applicability. + +# 1.2. Our Contributions + +To address the challenges of efficiency and adaptability in real image dehazing, this work proposes a new Compression-and-Adaptation (CoA) scheme as shown in Fig. 1 (b). The predefined model is compressed in the synthetic domain to ensure computational efficiency (refer to Fig. 1 (a)). Further, a bilevel adaptive learning is developed to leverage the dehazing capability acquired in the synthetic domain to improve adaptability (see Fig. 1 (c)) in the real domain. Our key contributions are summarized as + +- Following the divide-and-conquer paradigm, we propose a new CoA learning strategy that first compresses the predefined model and then adapts it to unseen scenes, offering an efficient solution for real image dehazing. To the + +best of our knowledge, this is the first work to successfully integrate both efficiency and adaptability, achieving the best of both worlds in real image dehazing. + +- To adapt to unlabeled, diverse real-world scenes, we propose a new cross-domain bilevel model designed to learn dehazing parameters for the real domain constrained by synthetic dehazing learning. Additionally, we derive a bilevel adaptive learning scheme that effectively harnesses the synthetic dehazing capacity, preventing instability and enhancing adaptability. +- Benefiting from succinct design free from additional constraints, CoA shows two key properties on three standard real benchmarks. It shows stability across synthetic domains, improving $14.2\%$ on average across four metrics (evaluated on three domains). CoA also offers flexibility with dehazing models, reducing computational cost by at least $73\%$ in parameters (tested on three methods). + +# 2. The Proposed Method + +In this section, we first rethinking real image dehazing, highlighting two key practical needs. We then propose model compression and bilevel adaptation to meet these needs, followed by an overview of the network architecture and training loss. + +# 2.1. Rethinking Real Image Dehazing + +Despite the strong performance of advanced image dehazing techniques [31, 36, 44] on synthetic datasets [8, 21, 30], these datasets often rely on simplifying assumptions, resulting in a substantial gap between synthetic and real-world domains. Addressing the core challenges in real-world image dehazing involves two key aspects: + +- Efficiency: Edge devices often have limited computational resources compared to systems with ample capacity, requiring dehazing models to be highly efficient. +- Adaptability: Real-world imaging conditions vary dynamically. Dehazing models must therefore be adaptable to maintain reliable performance. + +The latter is more challenging, relying on effective learning mechanisms like unsupervised learning [12]. However, current methods struggle with unsupervised learning due to difficulty in capturing statistical regularities, making transferring dehazing capabilities from the synthetic to real domain a viable alternative. To provide an intuitive presentation of the goal upon the above demands, we build + +$$ +\begin{array}{l} \min _ {\boldsymbol {\theta} _ {\mathrm {r e a l}}} f \Big (\boldsymbol {\theta} _ {\mathrm {r e a l}} \left(\boldsymbol {\theta} _ {\mathrm {s y n}}\right) \Big), \\ s. t., \left\{ \begin{array}{l} \kappa (\boldsymbol {\theta} _ {\text {r e a l}}) < \kappa (\boldsymbol {\theta} _ {\text {s y n}}), (\text {e f f i c i e n c y}) \\ \zeta (\boldsymbol {\theta} _ {\text {r e a l}}) > \zeta (\boldsymbol {\theta} _ {\text {s y n}}), (\text {a d a p t a b i l i t y}) \end{array} \right. \tag {1} \\ \end{array} +$$ + +where $\theta_{\mathrm{syn}}$ and $\theta_{\mathrm{real}}$ represent parameters in synthetic and real domain, respectively. The function $\kappa (\cdot)$ and $\zeta (\cdot)$ denote the evaluator for efficiency and adaptability, respectively. + +![](images/9bd811b0ce8fe713b7738e967a4989bb964291c9f9c69583f440ffc245559744.jpg) +Figure 2. Qualitative comparisons of algorithmic properties. Arrows indicate specific regions that highlight visible differences. + +To achieve this, we propose a divide-and-conquer strategy: the first phase focuses on model compression for efficiency, while the second phase employs bilevel adaptation for adaptability. Details are provided below. + +# 2.2. MoC: Model Compression in Synthetic Domain + +The main challenge in model compression is transferring dehazing capabilities from a large-scale parameter space to a smaller one. Here we propose a composite loss function with context alignment to realize it. + +$$ +\min _ {\boldsymbol {\theta} _ {\mathrm {s y n}} ^ {\mathrm {s}}} \mathcal {L} _ {\mathrm {s y n}} \left(\boldsymbol {\theta} _ {\mathrm {s y n}} ^ {\mathrm {s}}\right) + \mathcal {L} _ {\mathrm {a}} \left(\boldsymbol {\theta} _ {\mathrm {s y n}} ^ {\mathrm {s}}, \boldsymbol {\theta} _ {\mathrm {s y n}} ^ {\mathrm {t}}\right), \tag {2} +$$ + +where $\theta_{\mathrm{syn}}^{\mathrm{t}}$ and $\theta_{\mathrm{syn}}^{\mathrm{s}}$ are the pre-trained teacher model and desired student model in the synthetic domain, respectively. The function $\mathcal{L}_{\mathrm{syn}}$ and $\mathcal{L}_{\mathrm{a}}$ represent a group of supervised losses and the context alignment loss, respectively. + +During model compression, the student model follows the teacher model's training strategy on paired synthetic data. To maximize the retention of the teacher model's capabilities, we introduce a context alignment loss $(\mathcal{L}_{\mathrm{a}})$ that aligns the outputs of the teacher and student encoders at corresponding positions. By weighting the loss functions across layers, this approach enables fine-grained feature alignment, improving knowledge transfer and model optimization. The layered alignment ensures both global feature distributions and layer-specific details are accurately matched, resulting in more precise compression of the teacher model into the student model. + +# 2.3. BiA:Bilevel Adaptation to Real Domain + +After compressing the model for efficiency, we address adaptation to real-world scenes by proposing a cross-domain bilevel model. This is followed by a bilevel adaptive learning scheme to enable smooth transition from the synthetic to real domain. + +# 2.3.1. Cross-Domain Bilevel Modeling + +The gap between the synthetic and real domains makes adaptation challenging. The dehazing parameters in the synthetic and real domains $(\theta_{\mathrm{rea}}^s)$ have a nested subordinate + +# Algorithm 1 Learning via Compression-and-Adaptation + +Require: The pre-trained teacher model $\theta_{\mathrm{syn}}^{\mathrm{t}}$ , the initial parameter $\theta_{\mathrm{syn}}^{\mathrm{s}}$ , step-sizes $\eta_{\mathrm{syn}}^{m}$ , $\eta_{\mathrm{syn}}^{b}$ , coefficient $\alpha$ , iteration numbers $N$ in MoC phase and $T$ in BiA phase. + +Ensure: The optimal parameters $\theta_{\mathrm{rea}}^{\mathrm{s}}$ + +1: % The MoC Phase +2: for $n = 0:N - 1$ do +3: Calculate the gradient $g$ of Eq. (2). +4: $\pmb{\theta}_{\mathrm{syn}}^{\mathrm{s}}(n + 1) = \pmb{\theta}_{\mathrm{syn}}^{\mathrm{s}}(n) - \eta_{\mathrm{syn}}^{m}\pmb{g}(\pmb{\theta}_{\mathrm{syn}}^{\mathrm{s}},n),\pmb{\theta}_{\mathrm{syn}}^{\mathrm{t}}).$ +5: end for +6: Initialize the dehazing model's parameter in the real domain by $\pmb{\theta}_{\mathrm{rea}}^{\mathrm{s}} = \pmb{\theta}_{\mathrm{syn}}^{\mathrm{s}}(N)$ . +7: % The BiA Phase +8: for $t = 0 : T - 1$ do +9: $\%$ Update the lower-level parameter $\theta_{\mathrm{syn}}^{\mathrm{s}}$ in $\mathcal{D}_{\mathrm{rea}}$ . +10: $\pmb{\theta}_{\mathrm{syn}}^{\mathrm{s}}(t + 1) = \pmb{\theta}_{\mathrm{syn}}^{\mathrm{s}}(t) - \eta_{\mathrm{syn}}^{b}\frac{\partial(\mathcal{L}_{\mathrm{rea}} + \mathcal{L}_{1})}{\partial\pmb{\theta}_{\mathrm{syn}}^{\mathrm{s}}}$ . +11: $\%$ Update the upper-level parameter $\theta_{\mathrm{rea}}^{\mathrm{s}}$ by EMA. +12: $\pmb{\theta}_{\mathrm{rea}}^{\mathrm{s}}(t + 1) = \alpha \pmb{\theta}_{\mathrm{rea}}^{\mathrm{s}}(t) + (1 - \alpha)\pmb{\theta}_{\mathrm{syn}}^{\mathrm{s}}(t + 1).$ +13: end for +14: return $\theta_{\mathrm{rea}}^{\mathrm{s}}(T)$ + +relationship. This coupling can be effectively modeled using bilevel programming techniques [9, 29]. From the perspective of hyperparameter optimization, our cross-domain bilevel model is written as: + +$$ +\min _ {\boldsymbol {\theta} _ {\mathrm {r e a}} ^ {\mathrm {s}}} \mathcal {L} _ {\mathrm {r e a}} \big (\boldsymbol {\theta} _ {\mathrm {r e a}} ^ {\mathrm {s}}, \boldsymbol {\theta} _ {\mathrm {s y n}} ^ {\mathrm {s}} \big (\boldsymbol {\theta} _ {\mathrm {r e a}} ^ {\mathrm {s}} \big); \mathcal {D} _ {\mathrm {r e a}} \big), +$$ + +$$ +s. t., \theta_ {\text {s y n}} ^ {\mathrm {s}} \left(\theta_ {\text {r e a}} ^ {\mathrm {s}}\right) \in \arg \min _ {\theta_ {\text {s y n}} ^ {\mathrm {s}}} \Psi \left(\theta_ {\text {r e a}} ^ {\mathrm {s}}, \theta_ {\text {s y n}} ^ {\mathrm {s}}\right), \tag {3} +$$ + +where + +$$ +\Psi \left(\boldsymbol {\theta} _ {\mathrm {r e a}} ^ {\mathrm {s}}, \boldsymbol {\theta} _ {\mathrm {s y n}} ^ {\mathrm {s}}\right) = \mathcal {L} _ {\mathrm {r e a}} \left(\boldsymbol {\theta} _ {\mathrm {s y n}} ^ {\mathrm {s}}; \mathcal {D} _ {\mathrm {r e a}}\right) + \mathcal {L} _ {1} \left(\boldsymbol {\theta} _ {\mathrm {s y n}} ^ {\mathrm {s}}, \boldsymbol {\theta} _ {\mathrm {r e a}} ^ {\mathrm {s}}\right), \tag {4} +$$ + +where $\mathcal{L}_{\mathrm{rea}}$ and $\mathcal{L}_1$ represent the loss constraints for real domain and alignment. The datasets $\mathcal{D}_{\mathrm{rea}}$ and $\mathcal{D}_{\mathrm{syn}}$ are captured in the real and synthetic domains, respectively. + +Actually, suppose an effective loss $\mathcal{L}_{\mathrm{rea}}$ could be designed to constrain the output in the real domain, the adaptation might be achieved solely through the upper-level model + +
Testing BenchmarkMetricsStability across various synthetic domainFlexibility with different dehazing model
RESIDEHaze4KTHazeMSBDNDehazeFormerDEA
BaseCoABaseCoABaseCoABaseCoABaseCoABaseCoA
RTTS [22]FADE↓1.47531.26871.73681.08501.20780.85941.33541.15561.18660.93251.21170.9332
PM2.5↓163.77139.04170.03110.84131.7788.871135.6895.793126.77115.00121.9889.813
Entropy↑7.29637.34587.34467.46377.43197.57987.39687.53827.48157.55117.46237.5426
BIQME↑0.55620.56390.55110.57480.57730.59320.56960.58190.57330.59280.57890.5860
URHI [22]FADE↓1.44631.30131.72631.02781.31460.92721.34880.95491.32631.17261.27961.0321
PM2.5↓164.36145.01172.98119.97140.3598.793146.4399.856141.11125.87133.97108.94
Entropy↑7.28677.31527.37117.46577.46527.59287.41447.55747.50097.58177.49207.5658
BIQME↑0.54990.55420.54810.56810.58210.59610.57270.58320.57570.59480.57860.5863
FATTAL [6]FADE↓0.51360.47780.58280.37340.42030.31420.43150.33270.42710.38750.42700.3976
PM2.5↓81.73880.28981.401110.9762.31291.63969.272104.5678.09360.69381.38780.723
Entropy↑7.38477.45757.39997.50067.44827.58587.41217.54607.45667.56107.43157.5569
BIQME↑0.56450.58020.55540.59190.59450.61890.58070.60550.59270.61170.59030.6097
+ +Table 1. Quantitative comparison of algorithmic properties. In the left portion, our designed architecture serves as the baseline teacher model. In the right portion, the THaze dataset is used as the synthetic domain. + +(initialized with $\theta_{\mathrm{syn}}^{\mathrm{s}}$ ). However, as discussed in Sec. 2.1, the challenge of accurately modeling real haze distributions makes this process difficult to implement. To this end, we introduce lower-level sub-optimization to ensure a feasible solution and maintain the correct optimization trajectory. Experimental validation can be seen in Sec. 5.2. + +# 2.3.2. Bilevel Adaptive Learning + +Without compromising the model's dehazing capability in the synthetic domain, narrowing the performance gap between the synthetic and real domains is a central focus. Inspired by the Exponential Moving Average (EMA) method [2], we draw on its smoothing characteristics during the training process and introduce a dynamic weight adjustment mechanism to achieve a balance in learning between the synthetic and real domains, formulated as + +$$ +\left\{ \begin{array}{l} \boldsymbol {\theta} _ {\text {s y n}} ^ {\mathrm {s}} (t + 1) = \boldsymbol {\theta} _ {\text {s y n}} ^ {\mathrm {s}} (t) - \eta_ {\text {s y n}} ^ {b} \frac {\partial \left(\mathcal {L} _ {\text {r e a}} + \mathcal {L} _ {1}\right)}{\partial \boldsymbol {\theta} _ {\text {s y n}} ^ {\mathrm {s}}}, \\ \boldsymbol {\theta} _ {\text {r e a}} ^ {\mathrm {s}} (t + 1) = \alpha \boldsymbol {\theta} _ {\text {r e a}} ^ {\mathrm {s}} (t) + (1 - \alpha) \boldsymbol {\theta} _ {\text {s y n}} ^ {\mathrm {s}} (t + 1), \end{array} \right. \tag {5} +$$ + +where $\alpha$ represents the moving average coefficient, and $\mathcal{L}_1$ denotes the $\ell_1$ -norm supervised loss used to reduce the risk of the model overfitting to the real domain. Specifically, during each update, $\theta_{\mathrm{rea}}^{\mathrm{s}}$ is adjusted by $\theta_{\mathrm{syn}}^{\mathrm{s}}$ , allowing the model to strike a balance between the synthetic and real domains throughout the learning process. The overall algorithm can be seen in Alg. 1. + +# 2.4. Network Architecture and Training Loss + +This section introduces the details about network architecture and training loss adopted in this work. + +# 2.4.1. Network Architecture + +We emphasize that our primary focus is on developing an effective learning mechanism to enhance the model's adaptability to the real domain. Notably, our approach offers + +flexibility3 with different dehazing models. To this end, we adopt a hybrid architectural design, incorporating a pretrained Res2Net encoder [10] with the decoder of multiscale boosted dehazing network [5]. + +# 2.4.2. Loss Functions in MoC Phase + +In the MoC process, to enable the student model to learn the features extracted by the teacher model more effectively and to enhance the quality the generated images. We define + +$$ +\mathcal {L} _ {\mathrm {s y n}} = \lambda_ {s u} \ell_ {s u} + \lambda_ {s s} \ell_ {s s} + \lambda_ {p e} \ell_ {p e}, \tag {6} +$$ + +where $\ell_{su},\ell_{ss},\ell_{pe}$ denote $\ell_1$ -norm supervised loss, SSIM loss, and perceptual loss, respectively. The coefficient $\lambda_{su},\lambda_{ss},\lambda_{pe}$ are positive balancing parameters. + +To more effectively enhance consistency between the student and teacher models in the feature space, the design of $\mathcal{L}_{\mathrm{a}}$ incorporates a similarity measure of feature distributions. Its formulation can be written as follows: + +$$ +\mathcal {L} _ {\mathrm {a}} = \sum_ {i = 0} ^ {L - 1} w _ {i} \cdot \left(\left(T _ {i} - S _ {i}\right) ^ {2}\right), \tag {7} +$$ + +where $w_{i}$ represents the weights of $i$ -th layer in the teacher feature map $T_{i}$ and the student feature map $S_{i}$ . + +# 2.4.3. Loss Functions in BiA Phase + +Here we incorporate the CLIP model to construct the training loss $\mathcal{L}_{\mathrm{rea}}$ to constrain the output in the real domain. We select haze images $I_{\mathcal{H}}$ and clear images $I_{\mathcal{C}}$ as references, initialize corresponding prompts $T_{\mathcal{H}}$ and $T_{\mathcal{C}}$ , and input them into the CLIP text encoder, while feeding the images into the image encoder. The sample categories are predicted using the text-image similarity in the CLIP space. To minimize the classification error between positive and negative + +![](images/2081aea959a9ec7a42004fca43e2baedc424f704352c01f25fdce19a114dbb45.jpg) + +![](images/406dd7eb2f328f1e40df05ccea96ebd025e8459fb0ac9c6ba3d84e976a4acd49.jpg) +Figure 3. Quantitative comparison on three real-world datasets. Four no-reference image quality assessments were calculated. + +![](images/c5ad0c68ced8b4d7419a2ca3c17f3d729871e68bef62a575ae48fb4fdf736a35.jpg) + +![](images/5b78fef2c6ae07a68ace097c64d2dfa157f6d39b79d1b314eb87e32e6e0c3ea2.jpg) + +![](images/aea6bcee39fdf4c6cae19a7ab6caba61f1b5a7e8d3a342116e57f4a544a50f42.jpg) +Figure 4. Qualitative comparisons on daytime haze and dusty scenes. All these observations come from RTTS and URHI datasets. + +samples, we employ the binary cross entropy loss to optimize the text prompt pair. After training the text prompts to effectively distinguish between real haze and clear images, we can derive $\mathcal{L}_{\mathrm{rea}}$ based on the contrastive similarity loss: + +$$ +\mathcal {L} _ {\mathrm {r e a}} = \frac {e ^ {\cos \left(\Phi_ {\text {i m a g e}} \left(I _ {\mathcal {R}}\right) , \Phi_ {\text {t e x t}} \left(T _ {\mathcal {H}}\right)\right)}}{\sum_ {i \in \{\mathcal {H} , \mathcal {C} \}} e ^ {\cos \left(\Phi_ {\text {i m a g e}} \left(I _ {\mathcal {R}}\right) , \Phi_ {\text {t e x t}} \left(T _ {i}\right)\right)}}, \tag {8} +$$ + +where $I_{\mathcal{R}}$ corresponds to the real haze image processed by the model obtained through the model compression process. + +# 3. Exploring Algorithmic Property + +CoA presents an effective learning strategy to improve efficiency and adaptability for real-world scenes, without assumptions on the synthetic domain or teacher model. We analyze CoA's properties, highlighting its stability across synthetic domains and flexibility with different models. + +# 3.1. Stability across Various Synthetic Domains + +We perform three synthetic benchmarks (RESIDE [21] (the outdoor subset), Haze4K [30], and THaze [8]) and + +three real-world datasets (RTTS [21], URHI [21], and FAT-TAL [6]) to assess our stability. As shown in the left section of Table 1, CoA consistently improves scores across four no-reference image quality metrics. Notably, due to the diverse scenes in THaze, CoA achieves near-optimal performance within the synthetic domain. We use THaze as the synthetic domain in our comparative experiments. Fig. 2 (left) compares visual results. The dehazing effects of CoA is significant, especially in distant regions highlighted by arrows. Overall, CoA improves the original teacher model across synthetic domains, demonstrating its stability. + +# 3.2. Flexibility with Different Dehazing Models + +CoA's teacher model-irrelevant nature allows us to enhance the adaptability of existing dehazing models across various scenes. We applied CoA to three dehazing models (MSBDN [5], DehazeFormer [34], and DEA [3]) as teacher models. The right section of Table 1 shows significant performance improvements before and after applying CoA across different models and test settings. Fig. 2 (right) compares visual results using DehazeFormer and DEA, highlighting areas where CoA enhances dehazing for more visually appealing outcomes. In summary, CoA demonstrates notable flexibility with various dehazing models. The intro + +
MethodRTTSURHIFATTAL
FADE↓PM2.5↓Entropy↑BIQME↑FADE↓PM2.5↓Entropy↑BIQME↑FADE↓PM2.5↓Entropy↑BIQME↑
AirNet1.3342155.847.29590.54711.6591143.026.91430.52580.406376.1147.35450.5315
WeatherDiff2.4102193.277.14430.51552.1429180.267.14190.51020.8749118.797.36210.4937
DiffUIR2.1305189.617.17220.53031.9879180.567.18110.52050.8405116.687.36090.5049
Ours0.859488.8717.57980.59320.927298.7937.59280.59610.314291.6397.58580.6189
+ +Table 2. Quantitative comparison on daytime haze scenes. All compared methods are designed for multi-weather image restoration. + +![](images/30411f15b0370bc90d65016d9c1b5c44442ddd87a457233362cb5ff6dd4b5c00.jpg) +Figure 5. Qualitative comparisons corresponding to Table 2. All examples are sourced from RTTS and URHI datasets. + +![](images/a70559dd05378698edbd70ac1e86f036a0145c937eeb9828fd3de22e923e6cdd.jpg) +Figure 6. Performance evaluation on nighttime haze scenes. + +![](images/19e662b11062569afea83a8edb90457032053c125f2f4898ec0f42baadf153fd.jpg) + +![](images/3035cc059c694102b63d5d70d74235479e6ba43fc2097af432ed5bc25ce5d557.jpg) + +![](images/5fcfec06c607a62685a9b3172fef701a4dd0ed5ff6b6a9f4f2c0e07eaaac1f93.jpg) + +![](images/81a7dfa115e4fa3f75b6a7a4a702ec04bf732c6603e099aa29aea1879f57fb36.jpg) + +![](images/97ce5c805f5f9ec1948ca002453019e8069debd24f09b3884649d92842f1f4f0.jpg) + +![](images/d252059606f4b48f4dd313da11ac0a31d2070d68cff07a4939cf038b2e0818a1.jpg) + +![](images/7f322d82690e90b3e3eac3628041ad5f6ffb57b1c2328b42c6b09a9c2bf369f5.jpg) + +![](images/78e6a40c771cc43fe8cd5a56c2fb732bf409eb96dfed9b90b5478faed2a0a1b5.jpg) +Figure 7. Performance evaluation on underwater scenes. + +![](images/cd1b00f5a4f8fe4360c66e558947ef230eea6139eed5716eb06f6adc06b9c708.jpg) + +![](images/e99ff60e553377d8a29086f4983bb9e9648adfe096697ebacf22c9abd9cf72b1.jpg) + +(a) Qualitative comparison + +
MetricsGS (ICCV'15)MRP (CVPR'17)OSFD (ACM MM'20)GAPSF (ACM MM'23)CoA
PM2.5↓102.43106.20146.93110.5981.141
Entropy↑6.70257.04886.94126.37696.9907
BIQME↑0.45520.47220.46840.38950.5112
+ +(b) Quantitative comparison + +duction of CoA significantly reduces the model's parameter count, improving computational efficiency. Specifically, the parameters in the three dehazing models decrease by $94.06\%$ , $86.74\%$ , and $73.89\%$ , respectively. This reduction accelerates inference speed and lowers resource requirements, making the models more lightweight and better suited for practical applications. + +# 4. Experimental Results + +# 4.1. Implementation Details + +Training settings. The CoA model was implemented using the PyTorch framework on a single NVIDIA RTX 3090 GPU. The Adam optimizer was employed to update each module's parameters, with $\beta_{1},\beta_{2}$ , and $\varepsilon$ set to 0.9, 0.999, and $1e^{-8}$ , respectively. The initial learning rate was set to $1e^{-4}$ and gradually decayed to $1e^{-6}$ using a cosine annealing schedule. The algorithm was trained on RGB channels + +![](images/fa221ef9745c88ba0e7b0ead84dd29ab5c15f46e99ae8dbbcf2a945d85ea5f28.jpg) + +![](images/afc883fc039d3cde309afa825c065fa6982c8f1be04068a76e65b01db0d3a52d.jpg) + +![](images/23289f9a394973120e1a00b487f3675d2feb399dcd4a4b7d31b0bf4afc39081b.jpg) + +![](images/e3d84f23a41c6053066797c2c1542c3d6bb79843dea24c776536fa9c12243988.jpg) + +![](images/d97f6bbe3dae7d93afaa4b1f98f0db494772f139c2fe14a6c6028233d5261e35.jpg) + +(a) Qualitative comparison + +
MetricsDM(ACM MM'23)SUIR(CVPR'23)SMDR(AAAI'24)CoA
UIQM↑4.09653.98534.03144.1284
CCF↑26.65131.41728.43930.878
+ +(b) Quantitative comparison + +with data augmentation via random $90^{\circ}$ , $180^{\circ}$ , and $270^{\circ}$ rotations and horizontal flipping. We cropped random $256 \times 256$ sub-regions from images, further expanding the training set. Using a pre-trained Res2Net encoder, the teacher model achieved strong performance after just 20 epochs. + +Benchmarks and metrics. Three synthetic datasets (Haze4K, RESIDE (the outdoor subset), and THaze served as the synthetic domain, while three real-world datasets (RTTS, Fattal, and URHI were used for testing in the real domain. We employed four objective metrics for realworld haze scenes: FADE [4], PM2.5 [14], Entropy, and BIQME [13]. For underwater enhancement, we used UIQM and CCF to assess image quality. + +Compared methods. We evaluated our CoA method against state-of-the-art techniques across daytime, dusty, nighttime, and underwater conditions. For daytime and dusty conditions, we compared CoA with SGID [1], Dehamer [16], C2P [46], KANet [7], DEA, D4 [38], $\mathrm{D4^{+}}$ [39], + +
MetricsSGID (TIP'22)Dehamer (CVPR'22)C2P (CVPR'23)\( \text{RIDCP}^\dagger \)(CVPR'23)D4 (CVPR'22)D4+ (IJCV'24)KANet (TPAMI'24)DEA (TIP'24)Ours
SIZE (M)13.87132.407.1728.7210.7010.7055.253.651.69
FLOPs (G)108.4048.91352.90144.432.822.824.4232.202.67
Time (ms)1280×720240.01136.361169.18630.6451.2551.2522.9386.3324.33
1920×1080878.94295.982531.431588.26105.31105.31205.40190.7852.52
2560×14401514.82519.524409.12—185.55185.5510631334.7892.13
+ +This work requires excessive video memory for high-resolution images, resulting in the failure to produce results. + +![](images/db2bef2bd5bf7c8402e3bf843f0fd406367d66c8e355bd294db11a6e195922cd.jpg) +(a) Qualitative comparison + +Table 3. Evaluating the computational efficiency. Note that the size of the testing image for FLOPs calculation is ${224} \times {224}$ . + +
ModelFADE↓PM2.5↓BIQME↑SIZE (M)FLOPs (G)
Naive for Teacher1.21131.770.5852.8318.56
Naive for Student1.23135.760.571.692.67
MoC for Student1.13124.430.581.692.67
+ +and RIDCP [37]. Additionally, we assessed CoA against multi-weather restoration models: AirNet [23], WeatherDiff [33], and DiffUIR [45]. For nighttime dehazing, we benchmarked against GS [25], MRP [42], OSFD [43], and GAPSF [19], and for underwater scenes, we compared with DM [35], SUIR [18], and SMDR [40]. + +# 4.2. Evaluation on Regular Scenes + +Quantitative comparisons. In Fig. 3, we presented a comparison of eight state-of-the-art dehazing algorithms with our proposed method, using four authoritative unsupervised evaluation metrics. It is evident that our method ranked among the top in almost all metrics, achieving optimal or near-optimal performance in most cases. + +Qualitative comparisons. As shown in Fig. 4, the supervised methods (SGID, C2P, Dehamer, DEA) and unsupervised methods $(\mathrm{D}4,\mathrm{D}4^{+})$ struggle with haze removal and generalization. RIDCP causes color distortion and overbehazing, especially in sandstorm and colorful haze scenes. KANet has limitations in detail and color restoration. In contrast, CoA excels at haze removal while preserving fine textures, resulting in more natural and realistic images. + +Comparisons with multi-weather restorers. Here we compared CoA with state-of-the-art multi-weather degradation methods. As shown in Table 2, CoA ranks second in one metric, outperforming others in all others, highlighting its superiority. Fig. 5 reveals that while existing methods excel in specific scenarios, they show limitations in handling + +![](images/e97872433ad8806d3362f267cc5bdd774ca4e0cafe2ec5f0b8f53a2144ea6e36.jpg) +Figure 8. Effects of MoC phase. +Input + +![](images/b6eda8b4b2a5426283b294eaacd223ad60d967ea47049356749234569e004eaf.jpg) + +![](images/797617484687b57bcb3a6bf742ca04f67b17ddcc3554044447065b22123ce452.jpg) +(b) Quantitative comparison +Only Upper-Level Modeling +Figure 9. Necessity of bilevel modeling. + +![](images/092eaff8c72d7d7dd4461cdd5075cf160a79586127fcf9e3a93ef8e22a20ba27.jpg) +Only Lower-Level Modeling +Bilevel Modeling + +the full spectrum of haze conditions. + +Efficiency comparisons. Table 3 presents a comprehensive comparison of our proposed CoA method with existing approaches, evaluating parameters, floating-point operations per second (FLOPs), and processing time across various image resolutions. The results clearly show that our method has the smallest number of parameters and FLOPs among all the evaluated methods. Notably, as the image resolution increases, our method demonstrates a significant advantage in processing time, indicating that our model achieves a favorable balance between performance and complexity. + +# 4.3. Adaptability Verification + +Nighttime haze scene. Compared to state-of-the-art nighttime dehazing algorithms on the NHRW dataset (Fig. 6 (a)), many existing methods fail to handle overexposure or low-light conditions, resulting in overly dark or distorted images due to noise and artificial light scattering. In contrast, CoA excels in real-world nighttime haze scenes, effectively removing haze, suppressing noise, and preserving details. As shown in Fig. 6 (b), our method outperforms others across three authoritative objective metrics + +Underwater scene. Compared to state-of-the-art underwater dehazing algorithms, DM results in reddish and blurry images, failing to correct color distortion effectively, while SUIR and SMDR struggle to fully eliminate haze. As shown in Fig. 7, our method significantly reduces haze impact and enhances clarity and detail visibility, closely align- + +![](images/285705a7df2e711b0a1d2b4402b7e9d0e3e84db1b75a0842ba0989e1de9ed8c6.jpg) +Figure 10. Effects of BiA phase. + +![](images/59f9f86a4a9e8e1aad6b178f26df9270b7af675b4e39fd50ebfaf0886a0db105.jpg) +Figure 11. Parameters Analysis. The RTTS dataset is adopted. + +ing with human visual perception standards. + +# 5. Analytical Experiments + +# 5.1. Effects of MoC Phase + +Without feature transfer from the "naive for teacher" (naive: end-to-end training, the same below) model via compression, the "naive for student" model struggles to capture core data patterns, particularly in complex or extreme weather scenarios, resulting in limited dehazing. As illustrated in Fig. 8, while "MoC for student" may not restore details as effectively as "naive for teacher", it successfully inherits high-quality features. Even under resource constraints, it surpasses "naive for student" in dehazing and performs well on unsupervised metrics. + +# 5.2. Necessity of Bilevel Modeling + +Here we highlight the critical role of bilevel modeling by training the same network model on the THaze dataset. As shown in Fig. 9, the results on the real test set show marked contrasts: employing only the lower-level modeling strategy results in suboptimal dehazing, while relying solely on the upper-level strategy with CLIP fine-tuning produces unnatural effects with notable color distortion. Different from them, the bilevel modeling strategy enables both effective dehazing and natural visual quality, achieving a balance between dehazing performance and perceptual realism. + +# 5.3. Effects of BiA Phase + +In Fig. 10, we evaluate effects of the BiA phase. The results clearly show significant improvements in both FADE and + +![](images/cc5cca6b38d12f1b1ca8f5e1721640f10e329dde0a7cbbf10acba18de888ccfa.jpg) +Figure 12. Limitations. The example is from RS-Haze [34]. + +PM2.5 metrics following the BiA phase. Notably, in images with complex backgrounds or high noise levels, our model outperforms the baseline, demonstrating the BiA phase's effectiveness in capturing key features and improving robustness in challenging scenarios. + +# 5.4. Parameters Analyses + +We test different values of the parameter $\alpha$ during the BiA phase, as shown in Fig. 11. The results indicate that $\alpha = 0.95$ yields almost optimal performance. Notably, this setting achieves the fastest convergence and exhibits superior stability and accuracy across multiple evaluation metrics, clearly outperforming other parameter settings. + +# 5.5. Limitations + +While effective in many scenarios, our method struggles with remote sensing haze images due to additional imaging constraints. As depicted in Fig. 12, although it improves dehazing (outperforming DEA), non-uniform fog occlusion remains. This stems from varying haze effects across spectral bands in hyperspectral images, requiring multi-band integration for effective dehazing. + +# 6. Conclusion + +This work proposes a novel compression-and-adaptation scheme to tackle the problem of unbalance between dehazing quality and computational efficiency in real image dehazing. Following the divide-and-conquer paradigm, we design a new CoA learning strategy that first compresses the predefined model and then adapts it to unlabeled, diverse real-world scenes. We not only make a thorough exploration to take on the excellent properties of CoA, but also we perform extensive experiments to indicate our CoA method's superiority in terms of image dehazing quality and computational efficiency. + +# Acknowledgments + +This work was supported by the Postdoctoral Fellowship Program of CPSF (No. GZB20240098), China Postdoctoral Science Foundation (No. 2023M740491), Dalian Science and Technology Innovation Fund-Young Tech Star (No. 2023RQ017), Fundamental Research Funds for the Central Universities (No. DUT24BS011). + +# References + +[1] Haoran Bai, Jinshan Pan, Xinguang Xiang, and Jinhui Tang. Self-guided image dehazing using progressive feature fusion. IEEE Transactions on Image Processing, 31:1217-1229, 2022. 6 +[2] Zhaowei Cai, Avinash Ravichandran, Subhransu Maji, Charless Fowlkes, Zhuowen Tu, and Stefano Soatto. Exponential moving average normalization for self-supervised and semi-supervised learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 194-203, 2021. 4 +[3] Zixuan Chen, Zewei He, and Zhe-Ming Lu. Dea-net: Single image dehazing based on detail-enhanced convolution and content-guided attention. IEEE Transactions on Image Processing, 2024. 5 +[4] Lark Kwon Choi, Jaehee You, and Alan C. Bovik. Referenceless prediction of perceptual fog density and perceptual image defogging. IEEE Transactions on Image Processing, 24(11):3888-3901, 2015. 6 +[5] Hang Dong, Jinshan Pan, Lei Xiang, Zhe Hu, Xinyi Zhang, Fei Wang, and Ming-Hsuan Yang. Multi-scale boosted dehazing network with dense feature fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2157-2167, 2020. 4, 5 +[6] Raanan Fattal. Dehazing using color-lines. ACM transactions on graphics, 34(1):1-14, 2014. 4, 5 +[7] Yuxin Feng, Long Ma, Xiaozhe Meng, Fan Zhou, Risheng Liu, and Zhuo Su. Advancing real-world image dehazing: Perspective, modules, and training. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(12):9303-9320, 2024. 6 +[8] Yuxin Feng, Zhuo Su, Long Ma, Xin Li, Risheng Liu, and Fan Zhou. Bridging the gap between haze scenarios: A unified image dehazing model. IEEE Transactions on Circuits and Systems for Video Technology, 34(11):11070-11085, 2024. 2, 5 +[9] Luca Franceschi, Paolo Frasconi, Saverio Salzo, Riccardo Grazzi, and Massimiliano Pontil. Bilevel programming for hyperparameter optimization and meta-learning. In Proceedings of the International Conference on Machine Learning, pages 1568-1577, 2018. 3 +[10] Minghan Fu, Huan Liu, Yankun Yu, Jun Chen, and Keyan Wang. DW-GAN: A discrete wavelet transform GAN for nonhomogeneous dehazing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 203-212, 2021. 4 +[11] Xueyang Fu, Jie Xiao, Yurui Zhu, Aiping Liu, Feng Wu, and Zheng-Jun Zha. Continual image deraining with hypergraph convolutional networks. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(8):9534-9551, 2023. 1 +[12] Alona Golts, Daniel Freedman, and Michael Elad. Unsupervised single image dehazing using dark channel prior loss. IEEE Transactions on Image Processing, 29:2692-2701, 2019. 2 +[13] Ke Gu, Dacheng Tao, Junfei Qiao, and Weisi Lin. Learning a no-reference quality assessment model of enhanced images + +with big data. IEEE Transactions on Neural Networks and Learning Systems, 29(4):1301-1313, 2018. 6 +[14] Ke Gu, Junfei Qiao, and Xiaoli Li. Highly efficient picture-based prediction of PM2.5 concentration. IEEE Transactions on Industrial Electronics, 66(4):3176-3184, 2019. 6 +[15] Jie Gui, Xiaofeng Cong, Lei He, Yuan Yan Tang, and James Tin-Yau Kwok. Illumination controllable dehazing network based on unsupervised retinae embedding. IEEE Transactions on Multimedia, 26:4819-4830, 2024. 2 +[16] Chunle Guo, Qixin Yan, Saeed Anwar, Runmin Cong, Wenqi Ren, and Chongyi Li. Image dehazing transformer with transmission-aware 3d position embedding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5802-5810, 2022. 6 +[17] Kaiming He, Jian Sun, and Xiaou Tang. Single image haze removal using dark channel prior. IEEE Transactions on Pattern Analysis and Machine Intelligence, 33(12):2341-2353, 2010. 2 +[18] Shirui Huang, Keyan Wang, Huan Liu, Jun Chen, and Yun-song Li. Contrastive semi-supervised learning for underwater image restoration via reliable bank. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18145-18155, 2023. 7 +[19] Yeying Jin, Beibei Lin, Wending Yan, Yuan Yuan, Wei Ye, and Robby T. Tan. Enhancing visibility in nighttime haze images using guided APSF and gradient adaptive convolution. In Proceedings of the ACM International Conference on Multimedia, pages 2446-2457, 2023. 7 +[20] Mingye Ju, Can Ding, Charles A. Guo, Wenqi Ren, and Dacheng Tao. IDRLP: image dehazing using region line prior. IEEE Transactions on Image Processing, 30:9043-9057, 2021. 2 +[21] Boyi Li, Wenqi Ren, Dengpan Fu, Dacheng Tao, Dan Feng, Wenjun Zeng, and Zhangyang Wang. Benchmarking single-image dehazing and beyond. IEEE Transactions on Image Processing, 28(1):492-505, 2018. 2, 5 +[22] Boyi Li, Wenqi Ren, Dengpan Fu, Dacheng Tao, Dan Feng, Wenjun Zeng, and Zhangyang Wang. Benchmarking single-image dehazing and beyond. IEEE Transactions on Image Processing, 28(1):492-505, 2019. 4 +[23] Boyun Li, Xiao Liu, Peng Hu, Zhongqin Wu, Jiancheng Lv, and Xi Peng. All-in-one image restoration for unknown corruption. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17431-17441, 2022. 7 +[24] Pengyue Li, Jiandong Tian, Yandong Tang, Guolin Wang, and Chengdong Wu. Deep retina network for single image dehazing. IEEE Transactions on Image Processing, 30: 1100-1115, 2021. 2 +[25] Yu Li, Robby T. Tan, and Michael S. Brown. Nighttime haze removal with glow and multiple light colors. In Proceedings of the International Conference on Computer Vision, pages 226-234, 2015. 7 +[26] Yudong Liang, Bin Wang, Wangmeng Zuo, Jiaying Liu, and Wenqi Ren. Self-supervised learning and adaptation for single image dehazing. In Proceedings of the Thirty-First International Joint Conference on Artificial Intelligence, pages 1137–1143, 2022. 2 + +[27] Pengyang Ling, Huaian Chen, Xiao Tan, Yi Jin, and Enhong Chen. Single image dehazing using saturation line prior. IEEE Transactions on Image Processing, 32:3238-3253, 2023. 1 +[28] Jun Liu, Ryan Wen Liu, Jianing Sun, and Tieyong Zeng. Rank-one prior: Real-time scene recovery. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(7): 8845-8860, 2023. 1, 2 +[29] Risheng Liu, Jiaxin Gao, Jin Zhang, Deyu Meng, and Zhouchen Lin. Investigating bi-level optimization for learning and vision from a unified perspective: A survey and beyond. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(12):10045-10067, 2021. 3 +[30] Ye Liu, Lei Zhu, Shunda Pei, Huazhu Fu, Jing Qin, Qing Zhang, Liang Wan, and Wei Feng. From synthetic to real: Image dehazing collaborating with unlabeled real data. In Proceedings of the ACM International Conference on Multimedia, pages 50-58, 2021. 2, 5 +[31] Yidi Liu, Xingbo Wang, Yurui Zhu, Xueyang Fu, and Zheng-Jun Zha. Sdcnet: Spatially-adaptive deformable convolution networks for hr nonhomogeneous dehazing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6682-6691, 2024. 2 +[32] Long Ma, Tengyu Ma, Risheng Liu, Xin Fan, and Zhongxuan Luo. Toward fast, flexible, and robust low-light image enhancement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5637-5646, 2022. 1 +[33] Ozan Özdenizci and Robert Legenstein. Restoring vision in adverse weather conditions with patch-based denoising diffusion models. IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 1-12, 2023. 7 +[34] Yuda Song, Zhuqing He, Hui Qian, and Xin Du. Vision transformers for single image dehazing. IEEE Transactions on Image Processing, 32:1927-1941, 2023. 5, 8 +[35] Yi Tang, Hiroshi Kawasaki, and Takafumi Iwaguchi. Underwater image enhancement by transformer-based diffusion model with non-uniform sampling for skip strategy. In Proceedings of the ACM International Conference on Multimedia, pages 5419-5427, 2023. 7 +[36] Zhongze Wang, Haitao Zhao, Jingchao Peng, Lujian Yao, and Kaijie Zhao. Odcr: Orthogonal decoupling contrastive regularization for unpaired image dehazing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25479-25489, 2024. 2 +[37] Ruiqi Wu, Zheng-Peng Duan, Chun-Le Guo, Zhi Chai, and Chongyi Li. RIDCP: revitalizing real image dehazing via high-quality codebook priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22282-22291, 2023. 2, 7 +[38] Yang Yang, Chaoyue Wang, Risheng Liu, Lin Zhang, Xiaojie Guo, and Dacheng Tao. Self-augmented unpaired image dehazing via density and depth decomposition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2037-2046, 2022. 6 +[39] Yang Yang, Chaoyue Wang, Xiaojie Guo, and Dacheng Tao. Robust unpaired image dehazing via density and depth de + +composition. International Journal of Computer Vision, 132 (5):1557-1577, 2024. 6 +[40] Dehuan Zhang, Jingchun Zhou, Chunle Guo, Weishi Zhang, and Chongyi Li. Synergistic multiscale detail refinement via intrinsic supervision for underwater image enhancement. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 7033-7041, 2024. 7 +[41] He Zhang and Vishal M Patel. Densely connected pyramid dehazing network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3194-3203, 2018. 2 +[42] Jing Zhang, Yang Cao, and et al. Fast haze removal for nighttime image using maximum reflectance prior. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7016-7024, 2017. 7 +[43] Jing Zhang, Yang Cao, and et al. Nighttime dehazing with a synthetic benchmark. In Proceedings of the 28th ACM International Conference on Multimedia, pages 2355-2363, 2020. 7 +[44] Yafei Zhang, Shen Zhou, and Huafeng Li. Depth information assisted collaborative mutual promotion network for single image dehazing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2846-2855, 2024. 2 +[45] Dian Zheng, Xiao-Ming Wu, Shuzhou Yang, Jian Zhang, Jian-Fang Hu, and Wei-Shi Zheng. Selective hourglass mapping for universal image restoration based on diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 7 +[46] Yu Zheng, Jiahui Zhan, and et al. Curricular contrastive regularization for physics-aware single image dehazing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition., 1, 6 \ No newline at end of file diff --git a/2025/CoA_ Towards Real Image Dehazing via Compression-and-Adaptation/images.zip b/2025/CoA_ Towards Real Image Dehazing via Compression-and-Adaptation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..ab566649a79e9d7e63503318467b4f0401b60195 --- /dev/null +++ b/2025/CoA_ Towards Real Image Dehazing via Compression-and-Adaptation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8beed3303f46e762f9e841b4949cc897eba2db4edcad6a74ea85816cf59776d7 +size 979939 diff --git a/2025/CoA_ Towards Real Image Dehazing via Compression-and-Adaptation/layout.json b/2025/CoA_ Towards Real Image Dehazing via Compression-and-Adaptation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..0ff575244b60bf8d78ddec1919511717990f083f --- /dev/null +++ b/2025/CoA_ Towards Real Image Dehazing via Compression-and-Adaptation/layout.json @@ -0,0 +1,10988 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 87, + 103, + 523, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 87, + 103, + 523, + 121 + ], + "spans": [ + { + "bbox": [ + 87, + 103, + 523, + 121 + ], + "type": "text", + "content": "CoA: Towards Real Image Dehazing via Compression-and-Adaptation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 94, + 142, + 146, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 142, + 146, + 157 + ], + "spans": [ + { + "bbox": [ + 94, + 142, + 146, + 157 + ], + "type": "text", + "content": "Long Ma" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 173, + 143, + 236, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 173, + 143, + 236, + 171 + ], + "spans": [ + { + "bbox": [ + 173, + 143, + 236, + 171 + ], + "type": "text", + "content": "Yuxin Feng" + }, + { + "bbox": [ + 173, + 143, + 236, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 173, + 143, + 236, + 171 + ], + "type": "text", + "content": " \nJiang-Yong Chou" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 263, + 143, + 321, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 263, + 143, + 321, + 157 + ], + "spans": [ + { + "bbox": [ + 263, + 143, + 321, + 157 + ], + "type": "text", + "content": "Yan Zhang" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 350, + 143, + 412, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 143, + 412, + 157 + ], + "spans": [ + { + "bbox": [ + 350, + 143, + 412, + 157 + ], + "type": "text", + "content": "Jinyuan Liu1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 441, + 143, + 514, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 441, + 143, + 514, + 157 + ], + "spans": [ + { + "bbox": [ + 441, + 143, + 514, + 157 + ], + "type": "text", + "content": "Weimin Wang" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 293, + 158, + 370, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 158, + 370, + 172 + ], + "spans": [ + { + "bbox": [ + 293, + 158, + 370, + 172 + ], + "type": "text", + "content": "Chengpei " + }, + { + "bbox": [ + 293, + 158, + 370, + 172 + ], + "type": "inline_equation", + "content": "\\mathrm{Xu}^{4,*}" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 403, + 158, + 451, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 403, + 158, + 451, + 171 + ], + "spans": [ + { + "bbox": [ + 403, + 158, + 451, + 171 + ], + "type": "text", + "content": "Zhuo Su" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 71, + 172, + 541, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 172, + 541, + 213 + ], + "spans": [ + { + "bbox": [ + 71, + 172, + 541, + 213 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 71, + 172, + 541, + 213 + ], + "type": "text", + "content": "Dalian University of Technology " + }, + { + "bbox": [ + 71, + 172, + 541, + 213 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 71, + 172, + 541, + 213 + ], + "type": "text", + "content": "Sun Yat-sen University " + }, + { + "bbox": [ + 71, + 172, + 541, + 213 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 71, + 172, + 541, + 213 + ], + "type": "text", + "content": "Fuzhou University " + }, + { + "bbox": [ + 71, + 172, + 541, + 213 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 71, + 172, + 541, + 213 + ], + "type": "text", + "content": "University of New South Wales malone94319@gmail.com, {fengyx26, zhangy2779} " + }, + { + "bbox": [ + 71, + 172, + 541, + 213 + ], + "type": "inline_equation", + "content": "@" + }, + { + "bbox": [ + 71, + 172, + 541, + 213 + ], + "type": "text", + "content": " mail2.sysu.edu.cn, atlantis918@hotmail.com, chengpei.Xu@unsw.edu.au, wangweimin@dlut.edu.cn, gychen@fzu.edu.cn, suzhuo3@mail.sysu.edu.cn" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 63, + 229, + 229, + 321 + ], + "blocks": [ + { + "bbox": [ + 63, + 229, + 229, + 321 + ], + "lines": [ + { + "bbox": [ + 63, + 229, + 229, + 321 + ], + "spans": [ + { + "bbox": [ + 63, + 229, + 229, + 321 + ], + "type": "image", + "image_path": "bfa2a24e2f256479b04856035ca0b551b569bc05431acec6d9768ac1352674b5.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 83, + 323, + 208, + 331 + ], + "lines": [ + { + "bbox": [ + 83, + 323, + 208, + 331 + ], + "spans": [ + { + "bbox": [ + 83, + 323, + 208, + 331 + ], + "type": "text", + "content": "(a) Evaluating efficiency across various metrics" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 229, + 229, + 366, + 320 + ], + "blocks": [ + { + "bbox": [ + 229, + 229, + 366, + 320 + ], + "lines": [ + { + "bbox": [ + 229, + 229, + 366, + 320 + ], + "spans": [ + { + "bbox": [ + 229, + 229, + 366, + 320 + ], + "type": "image", + "image_path": "cf30f643d4fe7f77f26fdc98d832ff5cc0fe48bd8885d7eb6b5126dc2f79663f.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 252, + 323, + 351, + 331 + ], + "lines": [ + { + "bbox": [ + 252, + 323, + 351, + 331 + ], + "spans": [ + { + "bbox": [ + 252, + 323, + 351, + 331 + ], + "type": "text", + "content": "(b) The computational process of CoA" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 366, + 229, + 550, + 321 + ], + "blocks": [ + { + "bbox": [ + 366, + 229, + 550, + 321 + ], + "lines": [ + { + "bbox": [ + 366, + 229, + 550, + 321 + ], + "spans": [ + { + "bbox": [ + 366, + 229, + 550, + 321 + ], + "type": "image", + "image_path": "67d32c0a664118b8f2f44a755cffe18c2ed788d072e7141723b0bcbd931155e6.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 398, + 323, + 527, + 331 + ], + "lines": [ + { + "bbox": [ + 398, + 323, + 527, + 331 + ], + "spans": [ + { + "bbox": [ + 398, + 323, + 527, + 331 + ], + "type": "text", + "content": "(c) Evaluating adaptability across different scenes" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 335, + 555, + 380 + ], + "lines": [ + { + "bbox": [ + 55, + 335, + 555, + 380 + ], + "spans": [ + { + "bbox": [ + 55, + 335, + 555, + 380 + ], + "type": "text", + "content": "Figure 1. Performance evaluation. The proposed CoA incorporates model compression in synthetic domain for efficiency and bilevel adaptation in real domain for adaptability, as illustrated in the central sub-figure (b). The left sub-figure (a) presents efficiency across various metrics, clearly showing that our CoA outperforms others by a significant margin. The right sub-figure (c) shows adaptability across different scenes, where it is evident that our CoA consistently performs excellently." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 152, + 395, + 200, + 407 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 395, + 200, + 407 + ], + "spans": [ + { + "bbox": [ + 152, + 395, + 200, + 407 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 55, + 421, + 296, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 421, + 296, + 684 + ], + "spans": [ + { + "bbox": [ + 55, + 421, + 296, + 684 + ], + "type": "text", + "content": "Learning-based image dehazing algorithms have shown remarkable success in synthetic domains. However, real image dehazing is still in suspense due to computational resource constraints and the diversity of real-world scenes. Therefore, there is an urgent need for an algorithm that excels in both efficiency and adaptability to address real image dehazing effectively. This work proposes a Compression-and-Adaptation (CoA) computational flow to tackle these challenges from a divide-and-conquer perspective. First, model compression is performed in the synthetic domain to develop a compact dehazing parameter space, satisfying efficiency demands. Then, a bilevel adaptation in the real domain is introduced to be fearless in unknown real environments by aggregating the synthetic dehazing capabilities during the learning process. Leveraging a succinct design free from additional constraints, our CoA exhibits domain-irrelevant stability and model-agnostic flexibility, effectively bridging the model chasm between synthetic and real domains to further improve its practical utility. Extensive evaluations and analyses underscore the approach's superiority and effectiveness. The code is publicly available at https://github.com/fyxn1/CoA." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 315, + 395, + 394, + 407 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 395, + 394, + 407 + ], + "spans": [ + { + "bbox": [ + 315, + 395, + 394, + 407 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 415, + 555, + 631 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 415, + 555, + 631 + ], + "spans": [ + { + "bbox": [ + 313, + 415, + 555, + 631 + ], + "type": "text", + "content": "Image dehazing tasks aim to utilize the atmospheric optical model and advanced computing theory to achieve efficient restoration of scene information and image details through inverse imaging solution or data-driven model learning. It has drawn much attention in multiple emerging computer vision areas recently [46] [27] [28]. Similar to other low-level vision tasks [32] [11], image dehazing technology has evolved from an early stage focused on improving synthetic data metrics to a direction aimed at effectively generalizing across various real-world haze scenarios. Current approaches face dual challenges: efficiency-oriented methods achieve real-time processing but lack scene adaptability, while adaptability-focused techniques dynamically adjust strategies at the cost of high computational complexity that hinders deployment in resource-constrained environments. This paper addresses these limitations through novel architectural innovations in both computational efficiency and adaptive scene understanding." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 314, + 637, + 407, + 649 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 637, + 407, + 649 + ], + "spans": [ + { + "bbox": [ + 314, + 637, + 407, + 649 + ], + "type": "text", + "content": "1.1. Related Works" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 654, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 555, + 715 + ], + "type": "text", + "content": "Efficiency-oriented Dehazing Methods. This work focuses on enhancing computational efficiency through optimization algorithms, model simplification, or machine learning methods to enable faster processing under limited resources. Broadly, dehazing methods can be divided" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 704, + 144, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 704, + 144, + 714 + ], + "spans": [ + { + "bbox": [ + 67, + 704, + 144, + 714 + ], + "type": "text", + "content": "*Corresponding author." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11197" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 263 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 263 + ], + "type": "text", + "content": "into image enhancement-based, physics-based, and image layering-based approaches. Image enhancement-based dehazing methods [24] treat the task as an image improvement problem, increasing contrast and color accuracy with techniques like histogram equalization and color correction. Physics-based methods [17], [28] are more specific, leveraging hazy image priors and constraints to aid inverse problem solving. Image layering-based approaches, such as that of Zhang et al. [41], increase efficiency by breaking down the image into scales or layers, progressively refining transmittance maps and enhancing detail using multilevel pyramid structures and dense connectivity. Despite their computational and real-time advantages, they struggle with stability and consistent quality, especially in complex real-world haze conditions, often resulting in partial dehazing and insufficient color clarity." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 264, + 296, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 264, + 296, + 529 + ], + "spans": [ + { + "bbox": [ + 56, + 264, + 296, + 529 + ], + "type": "text", + "content": "Adaptability-focused Dehazing Methods. Adaptive dehazing methods tailor their strategies based on specific image attributes to improve dehazing effectiveness. These methods encompass three main types: local feature-based, atmospheric scattering model-based, and domain adaptation-based approaches. Local feature-based methods [15] adaptively adjust dehazing intensity according to local contrast and brightness variations, which helps restore image details effectively. Atmospheric scattering model-based adaptive dehazing methods [20] dynamically adjust the transmittance estimation by analyzing different regions of the image to achieve more accurate restoration. These methods can, to some extent, balance both efficiency and adaptability. Domain adaptation-based dehazing methods [26, 37] have gradually emerged in the field of image dehazing in recent years. These methods learn a shared feature representation, enabling the model to simultaneously adapt to the features of both the source domain and the target domain. Although these methods yield superior dehazing results, their complex models and dynamic imagespecific adjustments often increase computational costs, potentially impacting real-time applicability." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 538, + 167, + 551 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 538, + 167, + 551 + ], + "spans": [ + { + "bbox": [ + 55, + 538, + 167, + 551 + ], + "type": "text", + "content": "1.2. Our Contributions" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 556, + 295, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 556, + 295, + 664 + ], + "spans": [ + { + "bbox": [ + 55, + 556, + 295, + 664 + ], + "type": "text", + "content": "To address the challenges of efficiency and adaptability in real image dehazing, this work proposes a new Compression-and-Adaptation (CoA) scheme as shown in Fig. 1 (b). The predefined model is compressed in the synthetic domain to ensure computational efficiency (refer to Fig. 1 (a)). Further, a bilevel adaptive learning is developed to leverage the dehazing capability acquired in the synthetic domain to improve adaptability (see Fig. 1 (c)) in the real domain. Our key contributions are summarized as" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "- Following the divide-and-conquer paradigm, we propose a new CoA learning strategy that first compresses the predefined model and then adapts it to unseen scenes, offering an efficient solution for real image dehazing. To the" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 321, + 72, + 553, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 72, + 553, + 108 + ], + "spans": [ + { + "bbox": [ + 321, + 72, + 553, + 108 + ], + "type": "text", + "content": "best of our knowledge, this is the first work to successfully integrate both efficiency and adaptability, achieving the best of both worlds in real image dehazing." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 108, + 555, + 275 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 314, + 108, + 553, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 108, + 553, + 191 + ], + "spans": [ + { + "bbox": [ + 314, + 108, + 553, + 191 + ], + "type": "text", + "content": "- To adapt to unlabeled, diverse real-world scenes, we propose a new cross-domain bilevel model designed to learn dehazing parameters for the real domain constrained by synthetic dehazing learning. Additionally, we derive a bilevel adaptive learning scheme that effectively harnesses the synthetic dehazing capacity, preventing instability and enhancing adaptability." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 192, + 555, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 192, + 555, + 275 + ], + "spans": [ + { + "bbox": [ + 313, + 192, + 555, + 275 + ], + "type": "text", + "content": "- Benefiting from succinct design free from additional constraints, CoA shows two key properties on three standard real benchmarks. It shows stability across synthetic domains, improving " + }, + { + "bbox": [ + 313, + 192, + 555, + 275 + ], + "type": "inline_equation", + "content": "14.2\\%" + }, + { + "bbox": [ + 313, + 192, + 555, + 275 + ], + "type": "text", + "content": " on average across four metrics (evaluated on three domains). CoA also offers flexibility with dehazing models, reducing computational cost by at least " + }, + { + "bbox": [ + 313, + 192, + 555, + 275 + ], + "type": "inline_equation", + "content": "73\\%" + }, + { + "bbox": [ + 313, + 192, + 555, + 275 + ], + "type": "text", + "content": " in parameters (tested on three methods)." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 284, + 443, + 297 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 284, + 443, + 297 + ], + "spans": [ + { + "bbox": [ + 313, + 284, + 443, + 297 + ], + "type": "text", + "content": "2. The Proposed Method" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 304, + 554, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 304, + 554, + 365 + ], + "spans": [ + { + "bbox": [ + 313, + 304, + 554, + 365 + ], + "type": "text", + "content": "In this section, we first rethinking real image dehazing, highlighting two key practical needs. We then propose model compression and bilevel adaptation to meet these needs, followed by an overview of the network architecture and training loss." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 369, + 492, + 383 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 369, + 492, + 383 + ], + "spans": [ + { + "bbox": [ + 313, + 369, + 492, + 383 + ], + "type": "text", + "content": "2.1. Rethinking Real Image Dehazing" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 387, + 554, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 387, + 554, + 458 + ], + "spans": [ + { + "bbox": [ + 313, + 387, + 554, + 458 + ], + "type": "text", + "content": "Despite the strong performance of advanced image dehazing techniques [31, 36, 44] on synthetic datasets [8, 21, 30], these datasets often rely on simplifying assumptions, resulting in a substantial gap between synthetic and real-world domains. Addressing the core challenges in real-world image dehazing involves two key aspects:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 459, + 553, + 530 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 314, + 459, + 553, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 459, + 553, + 495 + ], + "spans": [ + { + "bbox": [ + 314, + 459, + 553, + 495 + ], + "type": "text", + "content": "- Efficiency: Edge devices often have limited computational resources compared to systems with ample capacity, requiring dehazing models to be highly efficient." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 314, + 495, + 553, + 530 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 495, + 553, + 530 + ], + "spans": [ + { + "bbox": [ + 314, + 495, + 553, + 530 + ], + "type": "text", + "content": "- Adaptability: Real-world imaging conditions vary dynamically. Dehazing models must therefore be adaptable to maintain reliable performance." + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 531, + 554, + 614 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 531, + 554, + 614 + ], + "spans": [ + { + "bbox": [ + 313, + 531, + 554, + 614 + ], + "type": "text", + "content": "The latter is more challenging, relying on effective learning mechanisms like unsupervised learning [12]. However, current methods struggle with unsupervised learning due to difficulty in capturing statistical regularities, making transferring dehazing capabilities from the synthetic to real domain a viable alternative. To provide an intuitive presentation of the goal upon the above demands, we build" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 343, + 619, + 553, + 673 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 343, + 619, + 553, + 673 + ], + "spans": [ + { + "bbox": [ + 343, + 619, + 553, + 673 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\min _ {\\boldsymbol {\\theta} _ {\\mathrm {r e a l}}} f \\Big (\\boldsymbol {\\theta} _ {\\mathrm {r e a l}} \\left(\\boldsymbol {\\theta} _ {\\mathrm {s y n}}\\right) \\Big), \\\\ s. t., \\left\\{ \\begin{array}{l} \\kappa (\\boldsymbol {\\theta} _ {\\text {r e a l}}) < \\kappa (\\boldsymbol {\\theta} _ {\\text {s y n}}), (\\text {e f f i c i e n c y}) \\\\ \\zeta (\\boldsymbol {\\theta} _ {\\text {r e a l}}) > \\zeta (\\boldsymbol {\\theta} _ {\\text {s y n}}), (\\text {a d a p t a b i l i t y}) \\end{array} \\right. \\tag {1} \\\\ \\end{array}", + "image_path": "6b708f64aa8677066832aed2251f6e4768fd8fed7f531e51d3b96e0dd5e53a67.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "inline_equation", + "content": "\\theta_{\\mathrm{syn}}" + }, + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "inline_equation", + "content": "\\theta_{\\mathrm{real}}" + }, + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "text", + "content": " represent parameters in synthetic and real domain, respectively. The function " + }, + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "inline_equation", + "content": "\\kappa (\\cdot)" + }, + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "inline_equation", + "content": "\\zeta (\\cdot)" + }, + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "text", + "content": " denote the evaluator for efficiency and adaptability, respectively." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "11198" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 63, + 72, + 547, + 171 + ], + "blocks": [ + { + "bbox": [ + 63, + 72, + 547, + 171 + ], + "lines": [ + { + "bbox": [ + 63, + 72, + 547, + 171 + ], + "spans": [ + { + "bbox": [ + 63, + 72, + 547, + 171 + ], + "type": "image", + "image_path": "9bd811b0ce8fe713b7738e967a4989bb964291c9f9c69583f440ffc245559744.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 69, + 175, + 539, + 186 + ], + "lines": [ + { + "bbox": [ + 69, + 175, + 539, + 186 + ], + "spans": [ + { + "bbox": [ + 69, + 175, + 539, + 186 + ], + "type": "text", + "content": "Figure 2. Qualitative comparisons of algorithmic properties. Arrows indicate specific regions that highlight visible differences." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 207, + 296, + 256 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 207, + 296, + 256 + ], + "spans": [ + { + "bbox": [ + 55, + 207, + 296, + 256 + ], + "type": "text", + "content": "To achieve this, we propose a divide-and-conquer strategy: the first phase focuses on model compression for efficiency, while the second phase employs bilevel adaptation for adaptability. Details are provided below." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 262, + 296, + 275 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 262, + 296, + 275 + ], + "spans": [ + { + "bbox": [ + 55, + 262, + 296, + 275 + ], + "type": "text", + "content": "2.2. MoC: Model Compression in Synthetic Domain" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 280, + 296, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 280, + 296, + 327 + ], + "spans": [ + { + "bbox": [ + 55, + 280, + 296, + 327 + ], + "type": "text", + "content": "The main challenge in model compression is transferring dehazing capabilities from a large-scale parameter space to a smaller one. Here we propose a composite loss function with context alignment to realize it." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 335, + 295, + 356 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 335, + 295, + 356 + ], + "spans": [ + { + "bbox": [ + 105, + 335, + 295, + 356 + ], + "type": "interline_equation", + "content": "\\min _ {\\boldsymbol {\\theta} _ {\\mathrm {s y n}} ^ {\\mathrm {s}}} \\mathcal {L} _ {\\mathrm {s y n}} \\left(\\boldsymbol {\\theta} _ {\\mathrm {s y n}} ^ {\\mathrm {s}}\\right) + \\mathcal {L} _ {\\mathrm {a}} \\left(\\boldsymbol {\\theta} _ {\\mathrm {s y n}} ^ {\\mathrm {s}}, \\boldsymbol {\\theta} _ {\\mathrm {s y n}} ^ {\\mathrm {t}}\\right), \\tag {2}", + "image_path": "da03508c0f412d2baa10b91be76c13aafc6e3a50d3b4f21daa0e12588291dc05.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 363, + 296, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 363, + 296, + 411 + ], + "spans": [ + { + "bbox": [ + 55, + 363, + 296, + 411 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 363, + 296, + 411 + ], + "type": "inline_equation", + "content": "\\theta_{\\mathrm{syn}}^{\\mathrm{t}}" + }, + { + "bbox": [ + 55, + 363, + 296, + 411 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 363, + 296, + 411 + ], + "type": "inline_equation", + "content": "\\theta_{\\mathrm{syn}}^{\\mathrm{s}}" + }, + { + "bbox": [ + 55, + 363, + 296, + 411 + ], + "type": "text", + "content": " are the pre-trained teacher model and desired student model in the synthetic domain, respectively. The function " + }, + { + "bbox": [ + 55, + 363, + 296, + 411 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{syn}}" + }, + { + "bbox": [ + 55, + 363, + 296, + 411 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 363, + 296, + 411 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{a}}" + }, + { + "bbox": [ + 55, + 363, + 296, + 411 + ], + "type": "text", + "content": " represent a group of supervised losses and the context alignment loss, respectively." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 412, + 296, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 412, + 296, + 555 + ], + "spans": [ + { + "bbox": [ + 55, + 412, + 296, + 555 + ], + "type": "text", + "content": "During model compression, the student model follows the teacher model's training strategy on paired synthetic data. To maximize the retention of the teacher model's capabilities, we introduce a context alignment loss " + }, + { + "bbox": [ + 55, + 412, + 296, + 555 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{\\mathrm{a}})" + }, + { + "bbox": [ + 55, + 412, + 296, + 555 + ], + "type": "text", + "content": " that aligns the outputs of the teacher and student encoders at corresponding positions. By weighting the loss functions across layers, this approach enables fine-grained feature alignment, improving knowledge transfer and model optimization. The layered alignment ensures both global feature distributions and layer-specific details are accurately matched, resulting in more precise compression of the teacher model into the student model." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 562, + 263, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 562, + 263, + 574 + ], + "spans": [ + { + "bbox": [ + 55, + 562, + 263, + 574 + ], + "type": "text", + "content": "2.3. BiA:Bilevel Adaptation to Real Domain" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 579, + 296, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 579, + 296, + 639 + ], + "spans": [ + { + "bbox": [ + 55, + 579, + 296, + 639 + ], + "type": "text", + "content": "After compressing the model for efficiency, we address adaptation to real-world scenes by proposing a cross-domain bilevel model. This is followed by a bilevel adaptive learning scheme to enable smooth transition from the synthetic to real domain." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 645, + 219, + 657 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 645, + 219, + 657 + ], + "spans": [ + { + "bbox": [ + 55, + 645, + 219, + 657 + ], + "type": "text", + "content": "2.3.1. Cross-Domain Bilevel Modeling" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 659, + 296, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 659, + 296, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 659, + 296, + 696 + ], + "type": "text", + "content": "The gap between the synthetic and real domains makes adaptation challenging. The dehazing parameters in the synthetic and real domains " + }, + { + "bbox": [ + 55, + 659, + 296, + 696 + ], + "type": "inline_equation", + "content": "(\\theta_{\\mathrm{rea}}^s)" + }, + { + "bbox": [ + 55, + 659, + 296, + 696 + ], + "type": "text", + "content": " have a nested subordinate" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 315, + 207, + 542, + 219 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 207, + 542, + 219 + ], + "spans": [ + { + "bbox": [ + 315, + 207, + 542, + 219 + ], + "type": "text", + "content": "Algorithm 1 Learning via Compression-and-Adaptation" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 317, + 223, + 554, + 260 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 223, + 554, + 260 + ], + "spans": [ + { + "bbox": [ + 317, + 223, + 554, + 260 + ], + "type": "text", + "content": "Require: The pre-trained teacher model " + }, + { + "bbox": [ + 317, + 223, + 554, + 260 + ], + "type": "inline_equation", + "content": "\\theta_{\\mathrm{syn}}^{\\mathrm{t}}" + }, + { + "bbox": [ + 317, + 223, + 554, + 260 + ], + "type": "text", + "content": ", the initial parameter " + }, + { + "bbox": [ + 317, + 223, + 554, + 260 + ], + "type": "inline_equation", + "content": "\\theta_{\\mathrm{syn}}^{\\mathrm{s}}" + }, + { + "bbox": [ + 317, + 223, + 554, + 260 + ], + "type": "text", + "content": ", step-sizes " + }, + { + "bbox": [ + 317, + 223, + 554, + 260 + ], + "type": "inline_equation", + "content": "\\eta_{\\mathrm{syn}}^{m}" + }, + { + "bbox": [ + 317, + 223, + 554, + 260 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 317, + 223, + 554, + 260 + ], + "type": "inline_equation", + "content": "\\eta_{\\mathrm{syn}}^{b}" + }, + { + "bbox": [ + 317, + 223, + 554, + 260 + ], + "type": "text", + "content": ", coefficient " + }, + { + "bbox": [ + 317, + 223, + 554, + 260 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 317, + 223, + 554, + 260 + ], + "type": "text", + "content": ", iteration numbers " + }, + { + "bbox": [ + 317, + 223, + 554, + 260 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 317, + 223, + 554, + 260 + ], + "type": "text", + "content": " in MoC phase and " + }, + { + "bbox": [ + 317, + 223, + 554, + 260 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 317, + 223, + 554, + 260 + ], + "type": "text", + "content": " in BiA phase." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 317, + 261, + 475, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 261, + 475, + 274 + ], + "spans": [ + { + "bbox": [ + 317, + 261, + 475, + 274 + ], + "type": "text", + "content": "Ensure: The optimal parameters " + }, + { + "bbox": [ + 317, + 261, + 475, + 274 + ], + "type": "inline_equation", + "content": "\\theta_{\\mathrm{rea}}^{\\mathrm{s}}" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 318, + 274, + 553, + 458 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 321, + 274, + 410, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 274, + 410, + 283 + ], + "spans": [ + { + "bbox": [ + 321, + 274, + 410, + 283 + ], + "type": "text", + "content": "1: % The MoC Phase" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 321, + 285, + 422, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 285, + 422, + 295 + ], + "spans": [ + { + "bbox": [ + 321, + 285, + 422, + 295 + ], + "type": "text", + "content": "2: for " + }, + { + "bbox": [ + 321, + 285, + 422, + 295 + ], + "type": "inline_equation", + "content": "n = 0:N - 1" + }, + { + "bbox": [ + 321, + 285, + 422, + 295 + ], + "type": "text", + "content": " do" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 320, + 297, + 484, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 297, + 484, + 308 + ], + "spans": [ + { + "bbox": [ + 320, + 297, + 484, + 308 + ], + "type": "text", + "content": "3: Calculate the gradient " + }, + { + "bbox": [ + 320, + 297, + 484, + 308 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 320, + 297, + 484, + 308 + ], + "type": "text", + "content": " of Eq. (2)." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 321, + 308, + 533, + 322 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 308, + 533, + 322 + ], + "spans": [ + { + "bbox": [ + 321, + 308, + 533, + 322 + ], + "type": "text", + "content": "4: " + }, + { + "bbox": [ + 321, + 308, + 533, + 322 + ], + "type": "inline_equation", + "content": "\\pmb{\\theta}_{\\mathrm{syn}}^{\\mathrm{s}}(n + 1) = \\pmb{\\theta}_{\\mathrm{syn}}^{\\mathrm{s}}(n) - \\eta_{\\mathrm{syn}}^{m}\\pmb{g}(\\pmb{\\theta}_{\\mathrm{syn}}^{\\mathrm{s}},n),\\pmb{\\theta}_{\\mathrm{syn}}^{\\mathrm{t}})." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 321, + 322, + 365, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 322, + 365, + 331 + ], + "spans": [ + { + "bbox": [ + 321, + 322, + 365, + 331 + ], + "type": "text", + "content": "5: end for" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 320, + 332, + 553, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 332, + 553, + 357 + ], + "spans": [ + { + "bbox": [ + 320, + 332, + 553, + 357 + ], + "type": "text", + "content": "6: Initialize the dehazing model's parameter in the real domain by " + }, + { + "bbox": [ + 320, + 332, + 553, + 357 + ], + "type": "inline_equation", + "content": "\\pmb{\\theta}_{\\mathrm{rea}}^{\\mathrm{s}} = \\pmb{\\theta}_{\\mathrm{syn}}^{\\mathrm{s}}(N)" + }, + { + "bbox": [ + 320, + 332, + 553, + 357 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 321, + 357, + 406, + 367 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 357, + 406, + 367 + ], + "spans": [ + { + "bbox": [ + 321, + 357, + 406, + 367 + ], + "type": "text", + "content": "7: % The BiA Phase" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 321, + 369, + 417, + 379 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 369, + 417, + 379 + ], + "spans": [ + { + "bbox": [ + 321, + 369, + 417, + 379 + ], + "type": "text", + "content": "8: for " + }, + { + "bbox": [ + 321, + 369, + 417, + 379 + ], + "type": "inline_equation", + "content": "t = 0 : T - 1" + }, + { + "bbox": [ + 321, + 369, + 417, + 379 + ], + "type": "text", + "content": " do" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 320, + 380, + 542, + 394 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 380, + 542, + 394 + ], + "spans": [ + { + "bbox": [ + 320, + 380, + 542, + 394 + ], + "type": "text", + "content": "9: " + }, + { + "bbox": [ + 320, + 380, + 542, + 394 + ], + "type": "inline_equation", + "content": "\\%" + }, + { + "bbox": [ + 320, + 380, + 542, + 394 + ], + "type": "text", + "content": " Update the lower-level parameter " + }, + { + "bbox": [ + 320, + 380, + 542, + 394 + ], + "type": "inline_equation", + "content": "\\theta_{\\mathrm{syn}}^{\\mathrm{s}}" + }, + { + "bbox": [ + 320, + 380, + 542, + 394 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 320, + 380, + 542, + 394 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathrm{rea}}" + }, + { + "bbox": [ + 320, + 380, + 542, + 394 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 320, + 394, + 504, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 394, + 504, + 410 + ], + "spans": [ + { + "bbox": [ + 320, + 394, + 504, + 410 + ], + "type": "text", + "content": "10: " + }, + { + "bbox": [ + 320, + 394, + 504, + 410 + ], + "type": "inline_equation", + "content": "\\pmb{\\theta}_{\\mathrm{syn}}^{\\mathrm{s}}(t + 1) = \\pmb{\\theta}_{\\mathrm{syn}}^{\\mathrm{s}}(t) - \\eta_{\\mathrm{syn}}^{b}\\frac{\\partial(\\mathcal{L}_{\\mathrm{rea}} + \\mathcal{L}_{1})}{\\partial\\pmb{\\theta}_{\\mathrm{syn}}^{\\mathrm{s}}}" + }, + { + "bbox": [ + 320, + 394, + 504, + 410 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 318, + 410, + 547, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 410, + 547, + 422 + ], + "spans": [ + { + "bbox": [ + 318, + 410, + 547, + 422 + ], + "type": "text", + "content": "11: " + }, + { + "bbox": [ + 318, + 410, + 547, + 422 + ], + "type": "inline_equation", + "content": "\\%" + }, + { + "bbox": [ + 318, + 410, + 547, + 422 + ], + "type": "text", + "content": " Update the upper-level parameter " + }, + { + "bbox": [ + 318, + 410, + 547, + 422 + ], + "type": "inline_equation", + "content": "\\theta_{\\mathrm{rea}}^{\\mathrm{s}}" + }, + { + "bbox": [ + 318, + 410, + 547, + 422 + ], + "type": "text", + "content": " by EMA." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 318, + 422, + 529, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 422, + 529, + 435 + ], + "spans": [ + { + "bbox": [ + 318, + 422, + 529, + 435 + ], + "type": "text", + "content": "12: " + }, + { + "bbox": [ + 318, + 422, + 529, + 435 + ], + "type": "inline_equation", + "content": "\\pmb{\\theta}_{\\mathrm{rea}}^{\\mathrm{s}}(t + 1) = \\alpha \\pmb{\\theta}_{\\mathrm{rea}}^{\\mathrm{s}}(t) + (1 - \\alpha)\\pmb{\\theta}_{\\mathrm{syn}}^{\\mathrm{s}}(t + 1)." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 318, + 435, + 365, + 444 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 435, + 365, + 444 + ], + "spans": [ + { + "bbox": [ + 318, + 435, + 365, + 444 + ], + "type": "text", + "content": "13: end for" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 318, + 445, + 400, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 445, + 400, + 458 + ], + "spans": [ + { + "bbox": [ + 318, + 445, + 400, + 458 + ], + "type": "text", + "content": "14: return " + }, + { + "bbox": [ + 318, + 445, + 400, + 458 + ], + "type": "inline_equation", + "content": "\\theta_{\\mathrm{rea}}^{\\mathrm{s}}(T)" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 480, + 554, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 480, + 554, + 528 + ], + "spans": [ + { + "bbox": [ + 313, + 480, + 554, + 528 + ], + "type": "text", + "content": "relationship. This coupling can be effectively modeled using bilevel programming techniques [9, 29]. From the perspective of hyperparameter optimization, our cross-domain bilevel model is written as:" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 350, + 533, + 495, + 553 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 533, + 495, + 553 + ], + "spans": [ + { + "bbox": [ + 350, + 533, + 495, + 553 + ], + "type": "interline_equation", + "content": "\\min _ {\\boldsymbol {\\theta} _ {\\mathrm {r e a}} ^ {\\mathrm {s}}} \\mathcal {L} _ {\\mathrm {r e a}} \\big (\\boldsymbol {\\theta} _ {\\mathrm {r e a}} ^ {\\mathrm {s}}, \\boldsymbol {\\theta} _ {\\mathrm {s y n}} ^ {\\mathrm {s}} \\big (\\boldsymbol {\\theta} _ {\\mathrm {r e a}} ^ {\\mathrm {s}} \\big); \\mathcal {D} _ {\\mathrm {r e a}} \\big),", + "image_path": "b1c004a447a12a1e1be9abc93d2c84a68a080767a09b0204cc98b3284a2c7533.jpg" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 350, + 550, + 553, + 576 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 550, + 553, + 576 + ], + "spans": [ + { + "bbox": [ + 350, + 550, + 553, + 576 + ], + "type": "interline_equation", + "content": "s. t., \\theta_ {\\text {s y n}} ^ {\\mathrm {s}} \\left(\\theta_ {\\text {r e a}} ^ {\\mathrm {s}}\\right) \\in \\arg \\min _ {\\theta_ {\\text {s y n}} ^ {\\mathrm {s}}} \\Psi \\left(\\theta_ {\\text {r e a}} ^ {\\mathrm {s}}, \\theta_ {\\text {s y n}} ^ {\\mathrm {s}}\\right), \\tag {3}", + "image_path": "a3ba25f5a05ce56ed41d8a3cd7a21812255609c1f8a84588652d3e090a8d4a39.jpg" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 314, + 583, + 342, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 583, + 342, + 594 + ], + "spans": [ + { + "bbox": [ + 314, + 583, + 342, + 594 + ], + "type": "text", + "content": "where" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 323, + 601, + 553, + 617 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 601, + 553, + 617 + ], + "spans": [ + { + "bbox": [ + 323, + 601, + 553, + 617 + ], + "type": "interline_equation", + "content": "\\Psi \\left(\\boldsymbol {\\theta} _ {\\mathrm {r e a}} ^ {\\mathrm {s}}, \\boldsymbol {\\theta} _ {\\mathrm {s y n}} ^ {\\mathrm {s}}\\right) = \\mathcal {L} _ {\\mathrm {r e a}} \\left(\\boldsymbol {\\theta} _ {\\mathrm {s y n}} ^ {\\mathrm {s}}; \\mathcal {D} _ {\\mathrm {r e a}}\\right) + \\mathcal {L} _ {1} \\left(\\boldsymbol {\\theta} _ {\\mathrm {s y n}} ^ {\\mathrm {s}}, \\boldsymbol {\\theta} _ {\\mathrm {r e a}} ^ {\\mathrm {s}}\\right), \\tag {4}", + "image_path": "0ad895cea92fa03b1aa5014b5b9ea8a576659f39068c0d1d40ab87409c21fa4f.jpg" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 313, + 624, + 554, + 660 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 624, + 554, + 660 + ], + "spans": [ + { + "bbox": [ + 313, + 624, + 554, + 660 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 624, + 554, + 660 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{rea}}" + }, + { + "bbox": [ + 313, + 624, + 554, + 660 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 624, + 554, + 660 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_1" + }, + { + "bbox": [ + 313, + 624, + 554, + 660 + ], + "type": "text", + "content": " represent the loss constraints for real domain and alignment. The datasets " + }, + { + "bbox": [ + 313, + 624, + 554, + 660 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathrm{rea}}" + }, + { + "bbox": [ + 313, + 624, + 554, + 660 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 624, + 554, + 660 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathrm{syn}}" + }, + { + "bbox": [ + 313, + 624, + 554, + 660 + ], + "type": "text", + "content": " are captured in the real and synthetic domains, respectively." + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 313, + 661, + 554, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 661, + 554, + 696 + ], + "spans": [ + { + "bbox": [ + 313, + 661, + 554, + 696 + ], + "type": "text", + "content": "Actually, suppose an effective loss " + }, + { + "bbox": [ + 313, + 661, + 554, + 696 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{rea}}" + }, + { + "bbox": [ + 313, + 661, + 554, + 696 + ], + "type": "text", + "content": " could be designed to constrain the output in the real domain, the adaptation might be achieved solely through the upper-level model" + } + ] + } + ], + "index": 36 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 66, + 702, + 269, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 702, + 269, + 712 + ], + "spans": [ + { + "bbox": [ + 66, + 702, + 269, + 712 + ], + "type": "text", + "content": "1Details of the specific loss functions can be found in Sec. 2.4." + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 324, + 702, + 476, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 702, + 476, + 712 + ], + "spans": [ + { + "bbox": [ + 324, + 702, + 476, + 712 + ], + "type": "text", + "content": "2Their specific forms can be found in Sec. 2.4." + } + ] + } + ], + "index": 38 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11199" + } + ] + } + ], + "index": 39 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 65, + 70, + 545, + 245 + ], + "blocks": [ + { + "bbox": [ + 65, + 70, + 545, + 245 + ], + "lines": [ + { + "bbox": [ + 65, + 70, + 545, + 245 + ], + "spans": [ + { + "bbox": [ + 65, + 70, + 545, + 245 + ], + "type": "table", + "html": "
Testing BenchmarkMetricsStability across various synthetic domainFlexibility with different dehazing model
RESIDEHaze4KTHazeMSBDNDehazeFormerDEA
BaseCoABaseCoABaseCoABaseCoABaseCoABaseCoA
RTTS [22]FADE↓1.47531.26871.73681.08501.20780.85941.33541.15561.18660.93251.21170.9332
PM2.5↓163.77139.04170.03110.84131.7788.871135.6895.793126.77115.00121.9889.813
Entropy↑7.29637.34587.34467.46377.43197.57987.39687.53827.48157.55117.46237.5426
BIQME↑0.55620.56390.55110.57480.57730.59320.56960.58190.57330.59280.57890.5860
URHI [22]FADE↓1.44631.30131.72631.02781.31460.92721.34880.95491.32631.17261.27961.0321
PM2.5↓164.36145.01172.98119.97140.3598.793146.4399.856141.11125.87133.97108.94
Entropy↑7.28677.31527.37117.46577.46527.59287.41447.55747.50097.58177.49207.5658
BIQME↑0.54990.55420.54810.56810.58210.59610.57270.58320.57570.59480.57860.5863
FATTAL [6]FADE↓0.51360.47780.58280.37340.42030.31420.43150.33270.42710.38750.42700.3976
PM2.5↓81.73880.28981.401110.9762.31291.63969.272104.5678.09360.69381.38780.723
Entropy↑7.38477.45757.39997.50067.44827.58587.41217.54607.45667.56107.43157.5569
BIQME↑0.56450.58020.55540.59190.59450.61890.58070.60550.59270.61170.59030.6097
", + "image_path": "a77dd7427336c634b727164766cb68fe65fbacc85c73507f6dba7f903f300046.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 247, + 555, + 270 + ], + "lines": [ + { + "bbox": [ + 55, + 247, + 555, + 270 + ], + "spans": [ + { + "bbox": [ + 55, + 247, + 555, + 270 + ], + "type": "text", + "content": "Table 1. Quantitative comparison of algorithmic properties. In the left portion, our designed architecture serves as the baseline teacher model. In the right portion, the THaze dataset is used as the synthetic domain." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 279, + 296, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 279, + 296, + 352 + ], + "spans": [ + { + "bbox": [ + 55, + 279, + 296, + 352 + ], + "type": "text", + "content": "(initialized with " + }, + { + "bbox": [ + 55, + 279, + 296, + 352 + ], + "type": "inline_equation", + "content": "\\theta_{\\mathrm{syn}}^{\\mathrm{s}}" + }, + { + "bbox": [ + 55, + 279, + 296, + 352 + ], + "type": "text", + "content": "). However, as discussed in Sec. 2.1, the challenge of accurately modeling real haze distributions makes this process difficult to implement. To this end, we introduce lower-level sub-optimization to ensure a feasible solution and maintain the correct optimization trajectory. Experimental validation can be seen in Sec. 5.2." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 357, + 196, + 369 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 357, + 196, + 369 + ], + "spans": [ + { + "bbox": [ + 55, + 357, + 196, + 369 + ], + "type": "text", + "content": "2.3.2. Bilevel Adaptive Learning" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 371, + 296, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 371, + 296, + 468 + ], + "spans": [ + { + "bbox": [ + 55, + 371, + 296, + 468 + ], + "type": "text", + "content": "Without compromising the model's dehazing capability in the synthetic domain, narrowing the performance gap between the synthetic and real domains is a central focus. Inspired by the Exponential Moving Average (EMA) method [2], we draw on its smoothing characteristics during the training process and introduce a dynamic weight adjustment mechanism to achieve a balance in learning between the synthetic and real domains, formulated as" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 70, + 474, + 296, + 519 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 474, + 296, + 519 + ], + "spans": [ + { + "bbox": [ + 70, + 474, + 296, + 519 + ], + "type": "interline_equation", + "content": "\\left\\{ \\begin{array}{l} \\boldsymbol {\\theta} _ {\\text {s y n}} ^ {\\mathrm {s}} (t + 1) = \\boldsymbol {\\theta} _ {\\text {s y n}} ^ {\\mathrm {s}} (t) - \\eta_ {\\text {s y n}} ^ {b} \\frac {\\partial \\left(\\mathcal {L} _ {\\text {r e a}} + \\mathcal {L} _ {1}\\right)}{\\partial \\boldsymbol {\\theta} _ {\\text {s y n}} ^ {\\mathrm {s}}}, \\\\ \\boldsymbol {\\theta} _ {\\text {r e a}} ^ {\\mathrm {s}} (t + 1) = \\alpha \\boldsymbol {\\theta} _ {\\text {r e a}} ^ {\\mathrm {s}} (t) + (1 - \\alpha) \\boldsymbol {\\theta} _ {\\text {s y n}} ^ {\\mathrm {s}} (t + 1), \\end{array} \\right. \\tag {5}", + "image_path": "30a8c0242cf9fe2aeea1898c214e1095449d7bf983cf818c8bbe262c6ec23eb0.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 525, + 296, + 610 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 525, + 296, + 610 + ], + "spans": [ + { + "bbox": [ + 55, + 525, + 296, + 610 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 525, + 296, + 610 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 525, + 296, + 610 + ], + "type": "text", + "content": " represents the moving average coefficient, and " + }, + { + "bbox": [ + 55, + 525, + 296, + 610 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_1" + }, + { + "bbox": [ + 55, + 525, + 296, + 610 + ], + "type": "text", + "content": " denotes the " + }, + { + "bbox": [ + 55, + 525, + 296, + 610 + ], + "type": "inline_equation", + "content": "\\ell_1" + }, + { + "bbox": [ + 55, + 525, + 296, + 610 + ], + "type": "text", + "content": "-norm supervised loss used to reduce the risk of the model overfitting to the real domain. Specifically, during each update, " + }, + { + "bbox": [ + 55, + 525, + 296, + 610 + ], + "type": "inline_equation", + "content": "\\theta_{\\mathrm{rea}}^{\\mathrm{s}}" + }, + { + "bbox": [ + 55, + 525, + 296, + 610 + ], + "type": "text", + "content": " is adjusted by " + }, + { + "bbox": [ + 55, + 525, + 296, + 610 + ], + "type": "inline_equation", + "content": "\\theta_{\\mathrm{syn}}^{\\mathrm{s}}" + }, + { + "bbox": [ + 55, + 525, + 296, + 610 + ], + "type": "text", + "content": ", allowing the model to strike a balance between the synthetic and real domains throughout the learning process. The overall algorithm can be seen in Alg. 1." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 616, + 269, + 628 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 616, + 269, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 616, + 269, + 628 + ], + "type": "text", + "content": "2.4. Network Architecture and Training Loss" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 633, + 295, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 633, + 295, + 658 + ], + "spans": [ + { + "bbox": [ + 55, + 633, + 295, + 658 + ], + "type": "text", + "content": "This section introduces the details about network architecture and training loss adopted in this work." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 662, + 178, + 673 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 662, + 178, + 673 + ], + "spans": [ + { + "bbox": [ + 55, + 662, + 178, + 673 + ], + "type": "text", + "content": "2.4.1. Network Architecture" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": "We emphasize that our primary focus is on developing an effective learning mechanism to enhance the model's adaptability to the real domain. Notably, our approach offers" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 279, + 554, + 328 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 279, + 554, + 328 + ], + "spans": [ + { + "bbox": [ + 313, + 279, + 554, + 328 + ], + "type": "text", + "content": "flexibility3 with different dehazing models. To this end, we adopt a hybrid architectural design, incorporating a pretrained Res2Net encoder [10] with the decoder of multiscale boosted dehazing network [5]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 334, + 469, + 345 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 334, + 469, + 345 + ], + "spans": [ + { + "bbox": [ + 313, + 334, + 469, + 345 + ], + "type": "text", + "content": "2.4.2. Loss Functions in MoC Phase" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 350, + 554, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 350, + 554, + 386 + ], + "spans": [ + { + "bbox": [ + 313, + 350, + 554, + 386 + ], + "type": "text", + "content": "In the MoC process, to enable the student model to learn the features extracted by the teacher model more effectively and to enhance the quality the generated images. We define" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 362, + 396, + 555, + 409 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 362, + 396, + 555, + 409 + ], + "spans": [ + { + "bbox": [ + 362, + 396, + 555, + 409 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {s y n}} = \\lambda_ {s u} \\ell_ {s u} + \\lambda_ {s s} \\ell_ {s s} + \\lambda_ {p e} \\ell_ {p e}, \\tag {6}", + "image_path": "92bce06aabfae505baeefb7769827b46da99ba305161a0586ea254e6a6aa79b6.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 418, + 554, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 418, + 554, + 456 + ], + "spans": [ + { + "bbox": [ + 313, + 418, + 554, + 456 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 418, + 554, + 456 + ], + "type": "inline_equation", + "content": "\\ell_{su},\\ell_{ss},\\ell_{pe}" + }, + { + "bbox": [ + 313, + 418, + 554, + 456 + ], + "type": "text", + "content": " denote " + }, + { + "bbox": [ + 313, + 418, + 554, + 456 + ], + "type": "inline_equation", + "content": "\\ell_1" + }, + { + "bbox": [ + 313, + 418, + 554, + 456 + ], + "type": "text", + "content": " -norm supervised loss, SSIM loss, and perceptual loss, respectively. The coefficient " + }, + { + "bbox": [ + 313, + 418, + 554, + 456 + ], + "type": "inline_equation", + "content": "\\lambda_{su},\\lambda_{ss},\\lambda_{pe}" + }, + { + "bbox": [ + 313, + 418, + 554, + 456 + ], + "type": "text", + "content": " are positive balancing parameters." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 456, + 554, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 456, + 554, + 502 + ], + "spans": [ + { + "bbox": [ + 313, + 456, + 554, + 502 + ], + "type": "text", + "content": "To more effectively enhance consistency between the student and teacher models in the feature space, the design of " + }, + { + "bbox": [ + 313, + 456, + 554, + 502 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{a}}" + }, + { + "bbox": [ + 313, + 456, + 554, + 502 + ], + "type": "text", + "content": " incorporates a similarity measure of feature distributions. Its formulation can be written as follows:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 375, + 510, + 555, + 543 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 375, + 510, + 555, + 543 + ], + "spans": [ + { + "bbox": [ + 375, + 510, + 555, + 543 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {a}} = \\sum_ {i = 0} ^ {L - 1} w _ {i} \\cdot \\left(\\left(T _ {i} - S _ {i}\\right) ^ {2}\\right), \\tag {7}", + "image_path": "46fafd722cc6190ba0901d0c0ebc7e70b3cc6a8fe2739ef1d8d22662f798a45c.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 552, + 554, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 552, + 554, + 576 + ], + "spans": [ + { + "bbox": [ + 313, + 552, + 554, + 576 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 552, + 554, + 576 + ], + "type": "inline_equation", + "content": "w_{i}" + }, + { + "bbox": [ + 313, + 552, + 554, + 576 + ], + "type": "text", + "content": " represents the weights of " + }, + { + "bbox": [ + 313, + 552, + 554, + 576 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 552, + 554, + 576 + ], + "type": "text", + "content": "-th layer in the teacher feature map " + }, + { + "bbox": [ + 313, + 552, + 554, + 576 + ], + "type": "inline_equation", + "content": "T_{i}" + }, + { + "bbox": [ + 313, + 552, + 554, + 576 + ], + "type": "text", + "content": " and the student feature map " + }, + { + "bbox": [ + 313, + 552, + 554, + 576 + ], + "type": "inline_equation", + "content": "S_{i}" + }, + { + "bbox": [ + 313, + 552, + 554, + 576 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 582, + 464, + 594 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 464, + 594 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 464, + 594 + ], + "type": "text", + "content": "2.4.3. Loss Functions in BiA Phase" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 598, + 555, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 598, + 555, + 694 + ], + "spans": [ + { + "bbox": [ + 313, + 598, + 555, + 694 + ], + "type": "text", + "content": "Here we incorporate the CLIP model to construct the training loss " + }, + { + "bbox": [ + 313, + 598, + 555, + 694 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{rea}}" + }, + { + "bbox": [ + 313, + 598, + 555, + 694 + ], + "type": "text", + "content": " to constrain the output in the real domain. We select haze images " + }, + { + "bbox": [ + 313, + 598, + 555, + 694 + ], + "type": "inline_equation", + "content": "I_{\\mathcal{H}}" + }, + { + "bbox": [ + 313, + 598, + 555, + 694 + ], + "type": "text", + "content": " and clear images " + }, + { + "bbox": [ + 313, + 598, + 555, + 694 + ], + "type": "inline_equation", + "content": "I_{\\mathcal{C}}" + }, + { + "bbox": [ + 313, + 598, + 555, + 694 + ], + "type": "text", + "content": " as references, initialize corresponding prompts " + }, + { + "bbox": [ + 313, + 598, + 555, + 694 + ], + "type": "inline_equation", + "content": "T_{\\mathcal{H}}" + }, + { + "bbox": [ + 313, + 598, + 555, + 694 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 598, + 555, + 694 + ], + "type": "inline_equation", + "content": "T_{\\mathcal{C}}" + }, + { + "bbox": [ + 313, + 598, + 555, + 694 + ], + "type": "text", + "content": ", and input them into the CLIP text encoder, while feeding the images into the image encoder. The sample categories are predicted using the text-image similarity in the CLIP space. To minimize the classification error between positive and negative" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 324, + 702, + 461, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 702, + 461, + 712 + ], + "spans": [ + { + "bbox": [ + 324, + 702, + 461, + 712 + ], + "type": "text", + "content": "3For further details, please refer to Sec. 3." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11200" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 66, + 70, + 299, + 163 + ], + "blocks": [ + { + "bbox": [ + 66, + 70, + 299, + 163 + ], + "lines": [ + { + "bbox": [ + 66, + 70, + 299, + 163 + ], + "spans": [ + { + "bbox": [ + 66, + 70, + 299, + 163 + ], + "type": "image", + "image_path": "2081aea959a9ec7a42004fca43e2baedc424f704352c01f25fdce19a114dbb45.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 66, + 167, + 299, + 261 + ], + "blocks": [ + { + "bbox": [ + 66, + 167, + 299, + 261 + ], + "lines": [ + { + "bbox": [ + 66, + 167, + 299, + 261 + ], + "spans": [ + { + "bbox": [ + 66, + 167, + 299, + 261 + ], + "type": "image", + "image_path": "406dd7eb2f328f1e40df05ccea96ebd025e8459fb0ac9c6ba3d84e976a4acd49.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 69, + 267, + 539, + 278 + ], + "lines": [ + { + "bbox": [ + 69, + 267, + 539, + 278 + ], + "spans": [ + { + "bbox": [ + 69, + 267, + 539, + 278 + ], + "type": "text", + "content": "Figure 3. Quantitative comparison on three real-world datasets. Four no-reference image quality assessments were calculated." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 309, + 69, + 543, + 163 + ], + "blocks": [ + { + "bbox": [ + 309, + 69, + 543, + 163 + ], + "lines": [ + { + "bbox": [ + 309, + 69, + 543, + 163 + ], + "spans": [ + { + "bbox": [ + 309, + 69, + 543, + 163 + ], + "type": "image", + "image_path": "c5ad0c68ced8b4d7419a2ca3c17f3d729871e68bef62a575ae48fb4fdf736a35.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 309, + 167, + 543, + 261 + ], + "blocks": [ + { + "bbox": [ + 309, + 167, + 543, + 261 + ], + "lines": [ + { + "bbox": [ + 309, + 167, + 543, + 261 + ], + "spans": [ + { + "bbox": [ + 309, + 167, + 543, + 261 + ], + "type": "image", + "image_path": "5b78fef2c6ae07a68ace097c64d2dfa157f6d39b79d1b314eb87e32e6e0c3ea2.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 63, + 284, + 550, + 396 + ], + "blocks": [ + { + "bbox": [ + 63, + 284, + 550, + 396 + ], + "lines": [ + { + "bbox": [ + 63, + 284, + 550, + 396 + ], + "spans": [ + { + "bbox": [ + 63, + 284, + 550, + 396 + ], + "type": "image", + "image_path": "aea6bcee39fdf4c6cae19a7ab6caba61f1b5a7e8d3a342116e57f4a544a50f42.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 64, + 403, + 544, + 414 + ], + "lines": [ + { + "bbox": [ + 64, + 403, + 544, + 414 + ], + "spans": [ + { + "bbox": [ + 64, + 403, + 544, + 414 + ], + "type": "text", + "content": "Figure 4. Qualitative comparisons on daytime haze and dusty scenes. All these observations come from RTTS and URHI datasets." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 427, + 295, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 427, + 295, + 475 + ], + "spans": [ + { + "bbox": [ + 55, + 427, + 295, + 475 + ], + "type": "text", + "content": "samples, we employ the binary cross entropy loss to optimize the text prompt pair. After training the text prompts to effectively distinguish between real haze and clear images, we can derive " + }, + { + "bbox": [ + 55, + 427, + 295, + 475 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{rea}}" + }, + { + "bbox": [ + 55, + 427, + 295, + 475 + ], + "type": "text", + "content": " based on the contrastive similarity loss:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 82, + 484, + 295, + 524 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 484, + 295, + 524 + ], + "spans": [ + { + "bbox": [ + 82, + 484, + 295, + 524 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {r e a}} = \\frac {e ^ {\\cos \\left(\\Phi_ {\\text {i m a g e}} \\left(I _ {\\mathcal {R}}\\right) , \\Phi_ {\\text {t e x t}} \\left(T _ {\\mathcal {H}}\\right)\\right)}}{\\sum_ {i \\in \\{\\mathcal {H} , \\mathcal {C} \\}} e ^ {\\cos \\left(\\Phi_ {\\text {i m a g e}} \\left(I _ {\\mathcal {R}}\\right) , \\Phi_ {\\text {t e x t}} \\left(T _ {i}\\right)\\right)}}, \\tag {8}", + "image_path": "f11e8f084bab9bda34db857b15d9c611f3548205283dc2bb9c6879e91fce0c1d.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 533, + 295, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 533, + 295, + 557 + ], + "spans": [ + { + "bbox": [ + 55, + 533, + 295, + 557 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 533, + 295, + 557 + ], + "type": "inline_equation", + "content": "I_{\\mathcal{R}}" + }, + { + "bbox": [ + 55, + 533, + 295, + 557 + ], + "type": "text", + "content": " corresponds to the real haze image processed by the model obtained through the model compression process." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 582, + 235, + 596 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 235, + 596 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 235, + 596 + ], + "type": "text", + "content": "3. Exploring Algorithmic Property" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 601, + 296, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 601, + 296, + 662 + ], + "spans": [ + { + "bbox": [ + 55, + 601, + 296, + 662 + ], + "type": "text", + "content": "CoA presents an effective learning strategy to improve efficiency and adaptability for real-world scenes, without assumptions on the synthetic domain or teacher model. We analyze CoA's properties, highlighting its stability across synthetic domains and flexibility with different models." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 671, + 280, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 671, + 280, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 671, + 280, + 685 + ], + "type": "text", + "content": "3.1. Stability across Various Synthetic Domains" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": "We perform three synthetic benchmarks (RESIDE [21] (the outdoor subset), Haze4K [30], and THaze [8]) and" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 426, + 555, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 555, + 559 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 555, + 559 + ], + "type": "text", + "content": "three real-world datasets (RTTS [21], URHI [21], and FAT-TAL [6]) to assess our stability. As shown in the left section of Table 1, CoA consistently improves scores across four no-reference image quality metrics. Notably, due to the diverse scenes in THaze, CoA achieves near-optimal performance within the synthetic domain. We use THaze as the synthetic domain in our comparative experiments. Fig. 2 (left) compares visual results. The dehazing effects of CoA is significant, especially in distant regions highlighted by arrows. Overall, CoA improves the original teacher model across synthetic domains, demonstrating its stability." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 564, + 537, + 578 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 564, + 537, + 578 + ], + "spans": [ + { + "bbox": [ + 313, + 564, + 537, + 578 + ], + "type": "text", + "content": "3.2. Flexibility with Different Dehazing Models" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "content": "CoA's teacher model-irrelevant nature allows us to enhance the adaptability of existing dehazing models across various scenes. We applied CoA to three dehazing models (MSBDN [5], DehazeFormer [34], and DEA [3]) as teacher models. The right section of Table 1 shows significant performance improvements before and after applying CoA across different models and test settings. Fig. 2 (right) compares visual results using DehazeFormer and DEA, highlighting areas where CoA enhances dehazing for more visually appealing outcomes. In summary, CoA demonstrates notable flexibility with various dehazing models. The intro" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11201" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 61, + 70, + 548, + 142 + ], + "blocks": [ + { + "bbox": [ + 61, + 70, + 548, + 142 + ], + "lines": [ + { + "bbox": [ + 61, + 70, + 548, + 142 + ], + "spans": [ + { + "bbox": [ + 61, + 70, + 548, + 142 + ], + "type": "table", + "html": "
MethodRTTSURHIFATTAL
FADE↓PM2.5↓Entropy↑BIQME↑FADE↓PM2.5↓Entropy↑BIQME↑FADE↓PM2.5↓Entropy↑BIQME↑
AirNet1.3342155.847.29590.54711.6591143.026.91430.52580.406376.1147.35450.5315
WeatherDiff2.4102193.277.14430.51552.1429180.267.14190.51020.8749118.797.36210.4937
DiffUIR2.1305189.617.17220.53031.9879180.567.18110.52050.8405116.687.36090.5049
Ours0.859488.8717.57980.59320.927298.7937.59280.59610.314291.6397.58580.6189
", + "image_path": "834400768a84b4bb0a3db0d0017dcf9c98d7255ce0716f77731805c37479cac5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 62, + 144, + 547, + 157 + ], + "lines": [ + { + "bbox": [ + 62, + 144, + 547, + 157 + ], + "spans": [ + { + "bbox": [ + 62, + 144, + 547, + 157 + ], + "type": "text", + "content": "Table 2. Quantitative comparison on daytime haze scenes. All compared methods are designed for multi-weather image restoration." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 63, + 167, + 135, + 260 + ], + "blocks": [ + { + "bbox": [ + 63, + 167, + 135, + 260 + ], + "lines": [ + { + "bbox": [ + 63, + 167, + 135, + 260 + ], + "spans": [ + { + "bbox": [ + 63, + 167, + 135, + 260 + ], + "type": "image", + "image_path": "30411f15b0370bc90d65016d9c1b5c44442ddd87a457233362cb5ff6dd4b5c00.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 84, + 268, + 523, + 279 + ], + "lines": [ + { + "bbox": [ + 84, + 268, + 523, + 279 + ], + "spans": [ + { + "bbox": [ + 84, + 268, + 523, + 279 + ], + "type": "text", + "content": "Figure 5. Qualitative comparisons corresponding to Table 2. All examples are sourced from RTTS and URHI datasets." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 136, + 167, + 184, + 260 + ], + "blocks": [ + { + "bbox": [ + 136, + 167, + 184, + 260 + ], + "lines": [ + { + "bbox": [ + 136, + 167, + 184, + 260 + ], + "spans": [ + { + "bbox": [ + 136, + 167, + 184, + 260 + ], + "type": "image", + "image_path": "a70559dd05378698edbd70ac1e86f036a0145c937eeb9828fd3de22e923e6cdd.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 62, + 472, + 288, + 483 + ], + "lines": [ + { + "bbox": [ + 62, + 472, + 288, + 483 + ], + "spans": [ + { + "bbox": [ + 62, + 472, + 288, + 483 + ], + "type": "text", + "content": "Figure 6. Performance evaluation on nighttime haze scenes." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 185, + 167, + 257, + 260 + ], + "blocks": [ + { + "bbox": [ + 185, + 167, + 257, + 260 + ], + "lines": [ + { + "bbox": [ + 185, + 167, + 257, + 260 + ], + "spans": [ + { + "bbox": [ + 185, + 167, + 257, + 260 + ], + "type": "image", + "image_path": "19e662b11062569afea83a8edb90457032053c125f2f4898ec0f42baadf153fd.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 258, + 167, + 306, + 260 + ], + "blocks": [ + { + "bbox": [ + 258, + 167, + 306, + 260 + ], + "lines": [ + { + "bbox": [ + 258, + 167, + 306, + 260 + ], + "spans": [ + { + "bbox": [ + 258, + 167, + 306, + 260 + ], + "type": "image", + "image_path": "3035cc059c694102b63d5d70d74235479e6ba43fc2097af432ed5bc25ce5d557.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 307, + 167, + 378, + 260 + ], + "blocks": [ + { + "bbox": [ + 307, + 167, + 378, + 260 + ], + "lines": [ + { + "bbox": [ + 307, + 167, + 378, + 260 + ], + "spans": [ + { + "bbox": [ + 307, + 167, + 378, + 260 + ], + "type": "image", + "image_path": "5fcfec06c607a62685a9b3172fef701a4dd0ed5ff6b6a9f4f2c0e07eaaac1f93.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 380, + 167, + 427, + 260 + ], + "blocks": [ + { + "bbox": [ + 380, + 167, + 427, + 260 + ], + "lines": [ + { + "bbox": [ + 380, + 167, + 427, + 260 + ], + "spans": [ + { + "bbox": [ + 380, + 167, + 427, + 260 + ], + "type": "image", + "image_path": "81a7dfa115e4fa3f75b6a7a4a702ec04bf732c6603e099aa29aea1879f57fb36.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 427, + 167, + 500, + 260 + ], + "blocks": [ + { + "bbox": [ + 427, + 167, + 500, + 260 + ], + "lines": [ + { + "bbox": [ + 427, + 167, + 500, + 260 + ], + "spans": [ + { + "bbox": [ + 427, + 167, + 500, + 260 + ], + "type": "image", + "image_path": "97ce5c805f5f9ec1948ca002453019e8069debd24f09b3884649d92842f1f4f0.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 501, + 167, + 549, + 260 + ], + "blocks": [ + { + "bbox": [ + 501, + 167, + 549, + 260 + ], + "lines": [ + { + "bbox": [ + 501, + 167, + 549, + 260 + ], + "spans": [ + { + "bbox": [ + 501, + 167, + 549, + 260 + ], + "type": "image", + "image_path": "d252059606f4b48f4dd313da11ac0a31d2070d68cff07a4939cf038b2e0818a1.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 63, + 290, + 131, + 392 + ], + "blocks": [ + { + "bbox": [ + 63, + 290, + 131, + 392 + ], + "lines": [ + { + "bbox": [ + 63, + 290, + 131, + 392 + ], + "spans": [ + { + "bbox": [ + 63, + 290, + 131, + 392 + ], + "type": "image", + "image_path": "7f322d82690e90b3e3eac3628041ad5f6ffb57b1c2328b42c6b09a9c2bf369f5.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 132, + 290, + 194, + 392 + ], + "blocks": [ + { + "bbox": [ + 132, + 290, + 194, + 392 + ], + "lines": [ + { + "bbox": [ + 132, + 290, + 194, + 392 + ], + "spans": [ + { + "bbox": [ + 132, + 290, + 194, + 392 + ], + "type": "image", + "image_path": "78e6a40c771cc43fe8cd5a56c2fb732bf409eb96dfed9b90b5478faed2a0a1b5.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 326, + 473, + 540, + 484 + ], + "lines": [ + { + "bbox": [ + 326, + 473, + 540, + 484 + ], + "spans": [ + { + "bbox": [ + 326, + 473, + 540, + 484 + ], + "type": "text", + "content": "Figure 7. Performance evaluation on underwater scenes." + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 194, + 290, + 256, + 392 + ], + "blocks": [ + { + "bbox": [ + 194, + 290, + 256, + 392 + ], + "lines": [ + { + "bbox": [ + 194, + 290, + 256, + 392 + ], + "spans": [ + { + "bbox": [ + 194, + 290, + 256, + 392 + ], + "type": "image", + "image_path": "cd1b00f5a4f8fe4360c66e558947ef230eea6139eed5716eb06f6adc06b9c708.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 257, + 290, + 289, + 391 + ], + "blocks": [ + { + "bbox": [ + 257, + 290, + 289, + 391 + ], + "lines": [ + { + "bbox": [ + 257, + 290, + 289, + 391 + ], + "spans": [ + { + "bbox": [ + 257, + 290, + 289, + 391 + ], + "type": "image", + "image_path": "e99ff60e553377d8a29086f4983bb9e9648adfe096697ebacf22c9abd9cf72b1.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "table", + "bbox": [ + 61, + 403, + 288, + 458 + ], + "blocks": [ + { + "bbox": [ + 133, + 393, + 220, + 403 + ], + "lines": [ + { + "bbox": [ + 133, + 393, + 220, + 403 + ], + "spans": [ + { + "bbox": [ + 133, + 393, + 220, + 403 + ], + "type": "text", + "content": "(a) Qualitative comparison" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 61, + 403, + 288, + 458 + ], + "lines": [ + { + "bbox": [ + 61, + 403, + 288, + 458 + ], + "spans": [ + { + "bbox": [ + 61, + 403, + 288, + 458 + ], + "type": "table", + "html": "
MetricsGS (ICCV'15)MRP (CVPR'17)OSFD (ACM MM'20)GAPSF (ACM MM'23)CoA
PM2.5↓102.43106.20146.93110.5981.141
Entropy↑6.70257.04886.94126.37696.9907
BIQME↑0.45520.47220.46840.38950.5112
", + "image_path": "6dfd92c63faa6fa5534baf87fc3c6cb25b67cd02d14a6a56fe37d91ae4174197.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "table_body" + } + ], + "index": 16 + }, + { + "bbox": [ + 129, + 459, + 220, + 468 + ], + "lines": [ + { + "bbox": [ + 129, + 459, + 220, + 468 + ], + "spans": [ + { + "bbox": [ + 129, + 459, + 220, + 468 + ], + "type": "text", + "content": "(b) Quantitative comparison" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 498, + 296, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 296, + 582 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 296, + 582 + ], + "type": "text", + "content": "duction of CoA significantly reduces the model's parameter count, improving computational efficiency. Specifically, the parameters in the three dehazing models decrease by " + }, + { + "bbox": [ + 55, + 498, + 296, + 582 + ], + "type": "inline_equation", + "content": "94.06\\%" + }, + { + "bbox": [ + 55, + 498, + 296, + 582 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 498, + 296, + 582 + ], + "type": "inline_equation", + "content": "86.74\\%" + }, + { + "bbox": [ + 55, + 498, + 296, + 582 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 498, + 296, + 582 + ], + "type": "inline_equation", + "content": "73.89\\%" + }, + { + "bbox": [ + 55, + 498, + 296, + 582 + ], + "type": "text", + "content": ", respectively. This reduction accelerates inference speed and lowers resource requirements, making the models more lightweight and better suited for practical applications." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 55, + 593, + 180, + 605 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 593, + 180, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 593, + 180, + 605 + ], + "type": "text", + "content": "4. Experimental Results" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 55, + 612, + 187, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 612, + 187, + 624 + ], + "spans": [ + { + "bbox": [ + 55, + 612, + 187, + 624 + ], + "type": "text", + "content": "4.1. Implementation Details" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "content": "Training settings. The CoA model was implemented using the PyTorch framework on a single NVIDIA RTX 3090 GPU. The Adam optimizer was employed to update each module's parameters, with " + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\beta_{1},\\beta_{2}" + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\varepsilon" + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "content": " set to 0.9, 0.999, and " + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "inline_equation", + "content": "1e^{-8}" + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "content": ", respectively. The initial learning rate was set to " + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "inline_equation", + "content": "1e^{-4}" + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "content": " and gradually decayed to " + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "inline_equation", + "content": "1e^{-6}" + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "content": " using a cosine annealing schedule. The algorithm was trained on RGB channels" + } + ] + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 321, + 290, + 367, + 403 + ], + "blocks": [ + { + "bbox": [ + 321, + 290, + 367, + 403 + ], + "lines": [ + { + "bbox": [ + 321, + 290, + 367, + 403 + ], + "spans": [ + { + "bbox": [ + 321, + 290, + 367, + 403 + ], + "type": "image", + "image_path": "fa221ef9745c88ba0e7b0ead84dd29ab5c15f46e99ae8dbbcf2a945d85ea5f28.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 367, + 290, + 412, + 402 + ], + "blocks": [ + { + "bbox": [ + 367, + 290, + 412, + 402 + ], + "lines": [ + { + "bbox": [ + 367, + 290, + 412, + 402 + ], + "spans": [ + { + "bbox": [ + 367, + 290, + 412, + 402 + ], + "type": "image", + "image_path": "afc883fc039d3cde309afa825c065fa6982c8f1be04068a76e65b01db0d3a52d.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 413, + 290, + 457, + 403 + ], + "blocks": [ + { + "bbox": [ + 413, + 290, + 457, + 403 + ], + "lines": [ + { + "bbox": [ + 413, + 290, + 457, + 403 + ], + "spans": [ + { + "bbox": [ + 413, + 290, + 457, + 403 + ], + "type": "image", + "image_path": "23289f9a394973120e1a00b487f3675d2feb399dcd4a4b7d31b0bf4afc39081b.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 459, + 290, + 503, + 402 + ], + "blocks": [ + { + "bbox": [ + 459, + 290, + 503, + 402 + ], + "lines": [ + { + "bbox": [ + 459, + 290, + 503, + 402 + ], + "spans": [ + { + "bbox": [ + 459, + 290, + 503, + 402 + ], + "type": "image", + "image_path": "e3d84f23a41c6053066797c2c1542c3d6bb79843dea24c776536fa9c12243988.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 503, + 290, + 548, + 402 + ], + "blocks": [ + { + "bbox": [ + 503, + 290, + 548, + 402 + ], + "lines": [ + { + "bbox": [ + 503, + 290, + 548, + 402 + ], + "spans": [ + { + "bbox": [ + 503, + 290, + 548, + 402 + ], + "type": "image", + "image_path": "d97f6bbe3dae7d93afaa4b1f98f0db494772f139c2fe14a6c6028233d5261e35.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "type": "table", + "bbox": [ + 321, + 415, + 545, + 459 + ], + "blocks": [ + { + "bbox": [ + 392, + 405, + 478, + 414 + ], + "lines": [ + { + "bbox": [ + 392, + 405, + 478, + 414 + ], + "spans": [ + { + "bbox": [ + 392, + 405, + 478, + 414 + ], + "type": "text", + "content": "(a) Qualitative comparison" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 321, + 415, + 545, + 459 + ], + "lines": [ + { + "bbox": [ + 321, + 415, + 545, + 459 + ], + "spans": [ + { + "bbox": [ + 321, + 415, + 545, + 459 + ], + "type": "table", + "html": "
MetricsDM(ACM MM'23)SUIR(CVPR'23)SMDR(AAAI'24)CoA
UIQM↑4.09653.98534.03144.1284
CCF↑26.65131.41728.43930.878
", + "image_path": "b92fb11db9d447991bb9d6ed472379a836075183bcc54a383523d29b1463ee21.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "table_body" + } + ], + "index": 29 + }, + { + "bbox": [ + 388, + 459, + 479, + 468 + ], + "lines": [ + { + "bbox": [ + 388, + 459, + 479, + 468 + ], + "spans": [ + { + "bbox": [ + 388, + 459, + 479, + 468 + ], + "type": "text", + "content": "(b) Quantitative comparison" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 498, + 553, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 553, + 556 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 553, + 556 + ], + "type": "text", + "content": "with data augmentation via random " + }, + { + "bbox": [ + 313, + 498, + 553, + 556 + ], + "type": "inline_equation", + "content": "90^{\\circ}" + }, + { + "bbox": [ + 313, + 498, + 553, + 556 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 498, + 553, + 556 + ], + "type": "inline_equation", + "content": "180^{\\circ}" + }, + { + "bbox": [ + 313, + 498, + 553, + 556 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 498, + 553, + 556 + ], + "type": "inline_equation", + "content": "270^{\\circ}" + }, + { + "bbox": [ + 313, + 498, + 553, + 556 + ], + "type": "text", + "content": " rotations and horizontal flipping. We cropped random " + }, + { + "bbox": [ + 313, + 498, + 553, + 556 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 313, + 498, + 553, + 556 + ], + "type": "text", + "content": " sub-regions from images, further expanding the training set. Using a pre-trained Res2Net encoder, the teacher model achieved strong performance after just 20 epochs." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 313, + 558, + 554, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 554, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 554, + 653 + ], + "type": "text", + "content": "Benchmarks and metrics. Three synthetic datasets (Haze4K, RESIDE (the outdoor subset), and THaze served as the synthetic domain, while three real-world datasets (RTTS, Fattal, and URHI were used for testing in the real domain. We employed four objective metrics for realworld haze scenes: FADE [4], PM2.5 [14], Entropy, and BIQME [13]. For underwater enhancement, we used UIQM and CCF to assess image quality." + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "text", + "content": "Compared methods. We evaluated our CoA method against state-of-the-art techniques across daytime, dusty, nighttime, and underwater conditions. For daytime and dusty conditions, we compared CoA with SGID [1], Dehamer [16], C2P [46], KANet [7], DEA, D4 [38], " + }, + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{D4^{+}}" + }, + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "text", + "content": " [39]," + } + ] + } + ], + "index": 34 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11202" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 75, + 70, + 535, + 159 + ], + "blocks": [ + { + "bbox": [ + 75, + 70, + 535, + 159 + ], + "lines": [ + { + "bbox": [ + 75, + 70, + 535, + 159 + ], + "spans": [ + { + "bbox": [ + 75, + 70, + 535, + 159 + ], + "type": "table", + "html": "
MetricsSGID (TIP'22)Dehamer (CVPR'22)C2P (CVPR'23)\\( \\text{RIDCP}^\\dagger \\)(CVPR'23)D4 (CVPR'22)D4+ (IJCV'24)KANet (TPAMI'24)DEA (TIP'24)Ours
SIZE (M)13.87132.407.1728.7210.7010.7055.253.651.69
FLOPs (G)108.4048.91352.90144.432.822.824.4232.202.67
Time (ms)1280×720240.01136.361169.18630.6451.2551.2522.9386.3324.33
1920×1080878.94295.982531.431588.26105.31105.31205.40190.7852.52
2560×14401514.82519.524409.12—185.55185.5510631334.7892.13
", + "image_path": "74cc649886a45a5395e15768aae5935d3f614924104c3575b3acb060378c9a0f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 81, + 159, + 315, + 167 + ], + "lines": [ + { + "bbox": [ + 81, + 159, + 315, + 167 + ], + "spans": [ + { + "bbox": [ + 81, + 159, + 315, + 167 + ], + "type": "text", + "content": "This work requires excessive video memory for high-resolution images, resulting in the failure to produce results." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 69, + 196, + 282, + 280 + ], + "blocks": [ + { + "bbox": [ + 69, + 196, + 282, + 280 + ], + "lines": [ + { + "bbox": [ + 69, + 196, + 282, + 280 + ], + "spans": [ + { + "bbox": [ + 69, + 196, + 282, + 280 + ], + "type": "image", + "image_path": "db2bef2bd5bf7c8402e3bf843f0fd406367d66c8e355bd294db11a6e195922cd.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 283, + 219, + 293 + ], + "lines": [ + { + "bbox": [ + 132, + 283, + 219, + 293 + ], + "spans": [ + { + "bbox": [ + 132, + 283, + 219, + 293 + ], + "type": "text", + "content": "(a) Qualitative comparison" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 69, + 293, + 280, + 334 + ], + "blocks": [ + { + "bbox": [ + 77, + 170, + 532, + 182 + ], + "lines": [ + { + "bbox": [ + 77, + 170, + 532, + 182 + ], + "spans": [ + { + "bbox": [ + 77, + 170, + 532, + 182 + ], + "type": "text", + "content": "Table 3. Evaluating the computational efficiency. Note that the size of the testing image for FLOPs calculation is " + }, + { + "bbox": [ + 77, + 170, + 532, + 182 + ], + "type": "inline_equation", + "content": "{224} \\times {224}" + }, + { + "bbox": [ + 77, + 170, + 532, + 182 + ], + "type": "text", + "content": " ." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 69, + 293, + 280, + 334 + ], + "lines": [ + { + "bbox": [ + 69, + 293, + 280, + 334 + ], + "spans": [ + { + "bbox": [ + 69, + 293, + 280, + 334 + ], + "type": "table", + "html": "
ModelFADE↓PM2.5↓BIQME↑SIZE (M)FLOPs (G)
Naive for Teacher1.21131.770.5852.8318.56
Naive for Student1.23135.760.571.692.67
MoC for Student1.13124.430.581.692.67
", + "image_path": "43d08bb725f84c18288e156bcb0f0155ec81130d9009550512e591161b4c9e63.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 373, + 296, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 373, + 296, + 445 + ], + "spans": [ + { + "bbox": [ + 55, + 373, + 296, + 445 + ], + "type": "text", + "content": "and RIDCP [37]. Additionally, we assessed CoA against multi-weather restoration models: AirNet [23], WeatherDiff [33], and DiffUIR [45]. For nighttime dehazing, we benchmarked against GS [25], MRP [42], OSFD [43], and GAPSF [19], and for underwater scenes, we compared with DM [35], SUIR [18], and SMDR [40]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 455, + 218, + 467 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 455, + 218, + 467 + ], + "spans": [ + { + "bbox": [ + 55, + 455, + 218, + 467 + ], + "type": "text", + "content": "4.2. Evaluation on Regular Scenes" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 473, + 296, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 473, + 296, + 544 + ], + "spans": [ + { + "bbox": [ + 55, + 473, + 296, + 544 + ], + "type": "text", + "content": "Quantitative comparisons. In Fig. 3, we presented a comparison of eight state-of-the-art dehazing algorithms with our proposed method, using four authoritative unsupervised evaluation metrics. It is evident that our method ranked among the top in almost all metrics, achieving optimal or near-optimal performance in most cases." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 545, + 296, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 296, + 641 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 296, + 641 + ], + "type": "text", + "content": "Qualitative comparisons. As shown in Fig. 4, the supervised methods (SGID, C2P, Dehamer, DEA) and unsupervised methods " + }, + { + "bbox": [ + 55, + 545, + 296, + 641 + ], + "type": "inline_equation", + "content": "(\\mathrm{D}4,\\mathrm{D}4^{+})" + }, + { + "bbox": [ + 55, + 545, + 296, + 641 + ], + "type": "text", + "content": " struggle with haze removal and generalization. RIDCP causes color distortion and overbehazing, especially in sandstorm and colorful haze scenes. KANet has limitations in detail and color restoration. In contrast, CoA excels at haze removal while preserving fine textures, resulting in more natural and realistic images." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": "Comparisons with multi-weather restorers. Here we compared CoA with state-of-the-art multi-weather degradation methods. As shown in Table 2, CoA ranks second in one metric, outperforming others in all others, highlighting its superiority. Fig. 5 reveals that while existing methods excel in specific scenarios, they show limitations in handling" + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 326, + 194, + 432, + 257 + ], + "blocks": [ + { + "bbox": [ + 114, + 347, + 236, + 358 + ], + "lines": [ + { + "bbox": [ + 114, + 347, + 236, + 358 + ], + "spans": [ + { + "bbox": [ + 114, + 347, + 236, + 358 + ], + "type": "text", + "content": "Figure 8. Effects of MoC phase." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 326, + 194, + 432, + 257 + ], + "lines": [ + { + "bbox": [ + 326, + 194, + 432, + 257 + ], + "spans": [ + { + "bbox": [ + 326, + 194, + 432, + 257 + ], + "type": "image", + "image_path": "e97872433ad8806d3362f267cc5bdd774ca4e0cafe2ec5f0b8f53a2144ea6e36.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 370, + 258, + 386, + 265 + ], + "lines": [ + { + "bbox": [ + 370, + 258, + 386, + 265 + ], + "spans": [ + { + "bbox": [ + 370, + 258, + 386, + 265 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 434, + 194, + 540, + 256 + ], + "blocks": [ + { + "bbox": [ + 434, + 194, + 540, + 256 + ], + "lines": [ + { + "bbox": [ + 434, + 194, + 540, + 256 + ], + "spans": [ + { + "bbox": [ + 434, + 194, + 540, + 256 + ], + "type": "image", + "image_path": "b6eda8b4b2a5426283b294eaacd223ad60d967ea47049356749234569e004eaf.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 326, + 267, + 431, + 329 + ], + "blocks": [ + { + "bbox": [ + 129, + 335, + 220, + 343 + ], + "lines": [ + { + "bbox": [ + 129, + 335, + 220, + 343 + ], + "spans": [ + { + "bbox": [ + 129, + 335, + 220, + 343 + ], + "type": "text", + "content": "(b) Quantitative comparison" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 326, + 267, + 431, + 329 + ], + "lines": [ + { + "bbox": [ + 326, + 267, + 431, + 329 + ], + "spans": [ + { + "bbox": [ + 326, + 267, + 431, + 329 + ], + "type": "image", + "image_path": "797617484687b57bcb3a6bf742ca04f67b17ddcc3554044447065b22123ce452.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 339, + 330, + 416, + 338 + ], + "lines": [ + { + "bbox": [ + 339, + 330, + 416, + 338 + ], + "spans": [ + { + "bbox": [ + 339, + 330, + 416, + 338 + ], + "type": "text", + "content": "Only Upper-Level Modeling" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 359, + 345, + 508, + 357 + ], + "lines": [ + { + "bbox": [ + 359, + 345, + 508, + 357 + ], + "spans": [ + { + "bbox": [ + 359, + 345, + 508, + 357 + ], + "type": "text", + "content": "Figure 9. Necessity of bilevel modeling." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 434, + 267, + 540, + 329 + ], + "blocks": [ + { + "bbox": [ + 447, + 258, + 525, + 266 + ], + "lines": [ + { + "bbox": [ + 447, + 258, + 525, + 266 + ], + "spans": [ + { + "bbox": [ + 447, + 258, + 525, + 266 + ], + "type": "text", + "content": "Only Lower-Level Modeling" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 434, + 267, + 540, + 329 + ], + "lines": [ + { + "bbox": [ + 434, + 267, + 540, + 329 + ], + "spans": [ + { + "bbox": [ + 434, + 267, + 540, + 329 + ], + "type": "image", + "image_path": "092eaff8c72d7d7dd4461cdd5075cf160a79586127fcf9e3a93ef8e22a20ba27.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 462, + 330, + 510, + 338 + ], + "lines": [ + { + "bbox": [ + 462, + 330, + 510, + 338 + ], + "spans": [ + { + "bbox": [ + 462, + 330, + 510, + 338 + ], + "type": "text", + "content": "Bilevel Modeling" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 371, + 462, + 383 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 371, + 462, + 383 + ], + "spans": [ + { + "bbox": [ + 313, + 371, + 462, + 383 + ], + "type": "text", + "content": "the full spectrum of haze conditions." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 384, + 555, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 384, + 555, + 504 + ], + "spans": [ + { + "bbox": [ + 313, + 384, + 555, + 504 + ], + "type": "text", + "content": "Efficiency comparisons. Table 3 presents a comprehensive comparison of our proposed CoA method with existing approaches, evaluating parameters, floating-point operations per second (FLOPs), and processing time across various image resolutions. The results clearly show that our method has the smallest number of parameters and FLOPs among all the evaluated methods. Notably, as the image resolution increases, our method demonstrates a significant advantage in processing time, indicating that our model achieves a favorable balance between performance and complexity." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 514, + 453, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 514, + 453, + 528 + ], + "spans": [ + { + "bbox": [ + 313, + 514, + 453, + 528 + ], + "type": "text", + "content": "4.3. Adaptability Verification" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 533, + 555, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 555, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 555, + 640 + ], + "type": "text", + "content": "Nighttime haze scene. Compared to state-of-the-art nighttime dehazing algorithms on the NHRW dataset (Fig. 6 (a)), many existing methods fail to handle overexposure or low-light conditions, resulting in overly dark or distorted images due to noise and artificial light scattering. In contrast, CoA excels in real-world nighttime haze scenes, effectively removing haze, suppressing noise, and preserving details. As shown in Fig. 6 (b), our method outperforms others across three authoritative objective metrics" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "text", + "content": "Underwater scene. Compared to state-of-the-art underwater dehazing algorithms, DM results in reddish and blurry images, failing to correct color distortion effectively, while SUIR and SMDR struggle to fully eliminate haze. As shown in Fig. 7, our method significantly reduces haze impact and enhances clarity and detail visibility, closely align-" + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11203" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 63, + 68, + 292, + 198 + ], + "blocks": [ + { + "bbox": [ + 63, + 68, + 292, + 198 + ], + "lines": [ + { + "bbox": [ + 63, + 68, + 292, + 198 + ], + "spans": [ + { + "bbox": [ + 63, + 68, + 292, + 198 + ], + "type": "image", + "image_path": "285705a7df2e711b0a1d2b4402b7e9d0e3e84db1b75a0842ba0989e1de9ed8c6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 113, + 204, + 236, + 216 + ], + "lines": [ + { + "bbox": [ + 113, + 204, + 236, + 216 + ], + "spans": [ + { + "bbox": [ + 113, + 204, + 236, + 216 + ], + "type": "text", + "content": "Figure 10. Effects of BiA phase." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 63, + 221, + 291, + 301 + ], + "blocks": [ + { + "bbox": [ + 63, + 221, + 291, + 301 + ], + "lines": [ + { + "bbox": [ + 63, + 221, + 291, + 301 + ], + "spans": [ + { + "bbox": [ + 63, + 221, + 291, + 301 + ], + "type": "image", + "image_path": "59f9f86a4a9e8e1aad6b178f26df9270b7af675b4e39fd50ebfaf0886a0db105.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 58, + 309, + 291, + 322 + ], + "lines": [ + { + "bbox": [ + 58, + 309, + 291, + 322 + ], + "spans": [ + { + "bbox": [ + 58, + 309, + 291, + 322 + ], + "type": "text", + "content": "Figure 11. Parameters Analysis. The RTTS dataset is adopted." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 337, + 235, + 349 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 337, + 235, + 349 + ], + "spans": [ + { + "bbox": [ + 55, + 337, + 235, + 349 + ], + "type": "text", + "content": "ing with human visual perception standards." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 360, + 192, + 373 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 360, + 192, + 373 + ], + "spans": [ + { + "bbox": [ + 55, + 360, + 192, + 373 + ], + "type": "text", + "content": "5. Analytical Experiments" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 380, + 179, + 392 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 380, + 179, + 392 + ], + "spans": [ + { + "bbox": [ + 55, + 380, + 179, + 392 + ], + "type": "text", + "content": "5.1. Effects of MoC Phase" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 398, + 296, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 398, + 296, + 518 + ], + "spans": [ + { + "bbox": [ + 55, + 398, + 296, + 518 + ], + "type": "text", + "content": "Without feature transfer from the \"naive for teacher\" (naive: end-to-end training, the same below) model via compression, the \"naive for student\" model struggles to capture core data patterns, particularly in complex or extreme weather scenarios, resulting in limited dehazing. As illustrated in Fig. 8, while \"MoC for student\" may not restore details as effectively as \"naive for teacher\", it successfully inherits high-quality features. Even under resource constraints, it surpasses \"naive for student\" in dehazing and performs well on unsupervised metrics." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 525, + 216, + 539 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 525, + 216, + 539 + ], + "spans": [ + { + "bbox": [ + 55, + 525, + 216, + 539 + ], + "type": "text", + "content": "5.2. Necessity of Bilevel Modeling" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 544, + 296, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 544, + 296, + 664 + ], + "spans": [ + { + "bbox": [ + 55, + 544, + 296, + 664 + ], + "type": "text", + "content": "Here we highlight the critical role of bilevel modeling by training the same network model on the THaze dataset. As shown in Fig. 9, the results on the real test set show marked contrasts: employing only the lower-level modeling strategy results in suboptimal dehazing, while relying solely on the upper-level strategy with CLIP fine-tuning produces unnatural effects with notable color distortion. Different from them, the bilevel modeling strategy enables both effective dehazing and natural visual quality, achieving a balance between dehazing performance and perceptual realism." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 671, + 174, + 683 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 671, + 174, + 683 + ], + "spans": [ + { + "bbox": [ + 55, + 671, + 174, + 683 + ], + "type": "text", + "content": "5.3. Effects of BiA Phase" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": "In Fig. 10, we evaluate effects of the BiA phase. The results clearly show significant improvements in both FADE and" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 323, + 70, + 549, + 137 + ], + "blocks": [ + { + "bbox": [ + 323, + 70, + 549, + 137 + ], + "lines": [ + { + "bbox": [ + 323, + 70, + 549, + 137 + ], + "spans": [ + { + "bbox": [ + 323, + 70, + 549, + 137 + ], + "type": "image", + "image_path": "cc5cca6b38d12f1b1ca8f5e1721640f10e329dde0a7cbbf10acba18de888ccfa.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 323, + 145, + 544, + 157 + ], + "lines": [ + { + "bbox": [ + 323, + 145, + 544, + 157 + ], + "spans": [ + { + "bbox": [ + 323, + 145, + 544, + 157 + ], + "type": "text", + "content": "Figure 12. Limitations. The example is from RS-Haze [34]." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 174, + 554, + 233 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 174, + 554, + 233 + ], + "spans": [ + { + "bbox": [ + 313, + 174, + 554, + 233 + ], + "type": "text", + "content": "PM2.5 metrics following the BiA phase. Notably, in images with complex backgrounds or high noise levels, our model outperforms the baseline, demonstrating the BiA phase's effectiveness in capturing key features and improving robustness in challenging scenarios." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 243, + 435, + 255 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 243, + 435, + 255 + ], + "spans": [ + { + "bbox": [ + 313, + 243, + 435, + 255 + ], + "type": "text", + "content": "5.4. Parameters Analyses" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 261, + 554, + 333 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 261, + 554, + 333 + ], + "spans": [ + { + "bbox": [ + 313, + 261, + 554, + 333 + ], + "type": "text", + "content": "We test different values of the parameter " + }, + { + "bbox": [ + 313, + 261, + 554, + 333 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 261, + 554, + 333 + ], + "type": "text", + "content": " during the BiA phase, as shown in Fig. 11. The results indicate that " + }, + { + "bbox": [ + 313, + 261, + 554, + 333 + ], + "type": "inline_equation", + "content": "\\alpha = 0.95" + }, + { + "bbox": [ + 313, + 261, + 554, + 333 + ], + "type": "text", + "content": " yields almost optimal performance. Notably, this setting achieves the fastest convergence and exhibits superior stability and accuracy across multiple evaluation metrics, clearly outperforming other parameter settings." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 342, + 391, + 353 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 342, + 391, + 353 + ], + "spans": [ + { + "bbox": [ + 313, + 342, + 391, + 353 + ], + "type": "text", + "content": "5.5. Limitations" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 360, + 554, + 444 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 360, + 554, + 444 + ], + "spans": [ + { + "bbox": [ + 313, + 360, + 554, + 444 + ], + "type": "text", + "content": "While effective in many scenarios, our method struggles with remote sensing haze images due to additional imaging constraints. As depicted in Fig. 12, although it improves dehazing (outperforming DEA), non-uniform fog occlusion remains. This stems from varying haze effects across spectral bands in hyperspectral images, requiring multi-band integration for effective dehazing." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 456, + 388, + 468 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 456, + 388, + 468 + ], + "spans": [ + { + "bbox": [ + 313, + 456, + 388, + 468 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 476, + 555, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 476, + 555, + 609 + ], + "spans": [ + { + "bbox": [ + 313, + 476, + 555, + 609 + ], + "type": "text", + "content": "This work proposes a novel compression-and-adaptation scheme to tackle the problem of unbalance between dehazing quality and computational efficiency in real image dehazing. Following the divide-and-conquer paradigm, we design a new CoA learning strategy that first compresses the predefined model and then adapts it to unlabeled, diverse real-world scenes. We not only make a thorough exploration to take on the excellent properties of CoA, but also we perform extensive experiments to indicate our CoA method's superiority in terms of image dehazing quality and computational efficiency." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 621, + 411, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 621, + 411, + 635 + ], + "spans": [ + { + "bbox": [ + 313, + 621, + 411, + 635 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 642, + 554, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 554, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 554, + 712 + ], + "type": "text", + "content": "This work was supported by the Postdoctoral Fellowship Program of CPSF (No. GZB20240098), China Postdoctoral Science Foundation (No. 2023M740491), Dalian Science and Technology Innovation Fund-Young Tech Star (No. 2023RQ017), Fundamental Research Funds for the Central Universities (No. DUT24BS011)." + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11204" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 134 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 134 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 134 + ], + "type": "text", + "content": "[1] Haoran Bai, Jinshan Pan, Xinguang Xiang, and Jinhui Tang. Self-guided image dehazing using progressive feature fusion. IEEE Transactions on Image Processing, 31:1217-1229, 2022. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 137, + 296, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 137, + 296, + 201 + ], + "spans": [ + { + "bbox": [ + 61, + 137, + 296, + 201 + ], + "type": "text", + "content": "[2] Zhaowei Cai, Avinash Ravichandran, Subhransu Maji, Charless Fowlkes, Zhuowen Tu, and Stefano Soatto. Exponential moving average normalization for self-supervised and semi-supervised learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 194-203, 2021. 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 204, + 295, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 204, + 295, + 247 + ], + "spans": [ + { + "bbox": [ + 62, + 204, + 295, + 247 + ], + "type": "text", + "content": "[3] Zixuan Chen, Zewei He, and Zhe-Ming Lu. Dea-net: Single image dehazing based on detail-enhanced convolution and content-guided attention. IEEE Transactions on Image Processing, 2024. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 250, + 295, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 250, + 295, + 293 + ], + "spans": [ + { + "bbox": [ + 62, + 250, + 295, + 293 + ], + "type": "text", + "content": "[4] Lark Kwon Choi, Jaehee You, and Alan C. Bovik. Referenceless prediction of perceptual fog density and perceptual image defogging. IEEE Transactions on Image Processing, 24(11):3888-3901, 2015. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 295, + 296, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 295, + 296, + 350 + ], + "spans": [ + { + "bbox": [ + 62, + 295, + 296, + 350 + ], + "type": "text", + "content": "[5] Hang Dong, Jinshan Pan, Lei Xiang, Zhe Hu, Xinyi Zhang, Fei Wang, and Ming-Hsuan Yang. Multi-scale boosted dehazing network with dense feature fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2157-2167, 2020. 4, 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 351, + 295, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 351, + 295, + 373 + ], + "spans": [ + { + "bbox": [ + 62, + 351, + 295, + 373 + ], + "type": "text", + "content": "[6] Raanan Fattal. Dehazing using color-lines. ACM transactions on graphics, 34(1):1-14, 2014. 4, 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 375, + 295, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 375, + 295, + 429 + ], + "spans": [ + { + "bbox": [ + 62, + 375, + 295, + 429 + ], + "type": "text", + "content": "[7] Yuxin Feng, Long Ma, Xiaozhe Meng, Fan Zhou, Risheng Liu, and Zhuo Su. Advancing real-world image dehazing: Perspective, modules, and training. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(12):9303-9320, 2024. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 431, + 295, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 431, + 295, + 485 + ], + "spans": [ + { + "bbox": [ + 62, + 431, + 295, + 485 + ], + "type": "text", + "content": "[8] Yuxin Feng, Zhuo Su, Long Ma, Xin Li, Risheng Liu, and Fan Zhou. Bridging the gap between haze scenarios: A unified image dehazing model. IEEE Transactions on Circuits and Systems for Video Technology, 34(11):11070-11085, 2024. 2, 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 487, + 295, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 487, + 295, + 542 + ], + "spans": [ + { + "bbox": [ + 62, + 487, + 295, + 542 + ], + "type": "text", + "content": "[9] Luca Franceschi, Paolo Frasconi, Saverio Salzo, Riccardo Grazzi, and Massimiliano Pontil. Bilevel programming for hyperparameter optimization and meta-learning. In Proceedings of the International Conference on Machine Learning, pages 1568-1577, 2018. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 544, + 295, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 544, + 295, + 599 + ], + "spans": [ + { + "bbox": [ + 57, + 544, + 295, + 599 + ], + "type": "text", + "content": "[10] Minghan Fu, Huan Liu, Yankun Yu, Jun Chen, and Keyan Wang. DW-GAN: A discrete wavelet transform GAN for nonhomogeneous dehazing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 203-212, 2021. 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 600, + 295, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 600, + 295, + 644 + ], + "spans": [ + { + "bbox": [ + 57, + 600, + 295, + 644 + ], + "type": "text", + "content": "[11] Xueyang Fu, Jie Xiao, Yurui Zhu, Aiping Liu, Feng Wu, and Zheng-Jun Zha. Continual image deraining with hypergraph convolutional networks. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(8):9534-9551, 2023. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 646, + 295, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 646, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 57, + 646, + 295, + 689 + ], + "type": "text", + "content": "[12] Alona Golts, Daniel Freedman, and Michael Elad. Unsupervised single image dehazing using dark channel prior loss. IEEE Transactions on Image Processing, 29:2692-2701, 2019. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "type": "text", + "content": "[13] Ke Gu, Dacheng Tao, Junfei Qiao, and Weisi Lin. Learning a no-reference quality assessment model of enhanced images" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 317, + 73, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 335, + 73, + 554, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 554, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 554, + 95 + ], + "type": "text", + "content": "with big data. IEEE Transactions on Neural Networks and Learning Systems, 29(4):1301-1313, 2018. 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 96, + 554, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 96, + 554, + 128 + ], + "spans": [ + { + "bbox": [ + 317, + 96, + 554, + 128 + ], + "type": "text", + "content": "[14] Ke Gu, Junfei Qiao, and Xiaoli Li. Highly efficient picture-based prediction of PM2.5 concentration. IEEE Transactions on Industrial Electronics, 66(4):3176-3184, 2019. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 129, + 554, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 129, + 554, + 172 + ], + "spans": [ + { + "bbox": [ + 317, + 129, + 554, + 172 + ], + "type": "text", + "content": "[15] Jie Gui, Xiaofeng Cong, Lei He, Yuan Yan Tang, and James Tin-Yau Kwok. Illumination controllable dehazing network based on unsupervised retinae embedding. IEEE Transactions on Multimedia, 26:4819-4830, 2024. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 173, + 554, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 173, + 554, + 228 + ], + "spans": [ + { + "bbox": [ + 317, + 173, + 554, + 228 + ], + "type": "text", + "content": "[16] Chunle Guo, Qixin Yan, Saeed Anwar, Runmin Cong, Wenqi Ren, and Chongyi Li. Image dehazing transformer with transmission-aware 3d position embedding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5802-5810, 2022. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 228, + 554, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 228, + 554, + 270 + ], + "spans": [ + { + "bbox": [ + 317, + 228, + 554, + 270 + ], + "type": "text", + "content": "[17] Kaiming He, Jian Sun, and Xiaou Tang. Single image haze removal using dark channel prior. IEEE Transactions on Pattern Analysis and Machine Intelligence, 33(12):2341-2353, 2010. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 272, + 555, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 272, + 555, + 327 + ], + "spans": [ + { + "bbox": [ + 317, + 272, + 555, + 327 + ], + "type": "text", + "content": "[18] Shirui Huang, Keyan Wang, Huan Liu, Jun Chen, and Yun-song Li. Contrastive semi-supervised learning for underwater image restoration via reliable bank. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18145-18155, 2023. 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 327, + 554, + 382 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 327, + 554, + 382 + ], + "spans": [ + { + "bbox": [ + 317, + 327, + 554, + 382 + ], + "type": "text", + "content": "[19] Yeying Jin, Beibei Lin, Wending Yan, Yuan Yuan, Wei Ye, and Robby T. Tan. Enhancing visibility in nighttime haze images using guided APSF and gradient adaptive convolution. In Proceedings of the ACM International Conference on Multimedia, pages 2446-2457, 2023. 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 383, + 554, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 383, + 554, + 425 + ], + "spans": [ + { + "bbox": [ + 317, + 383, + 554, + 425 + ], + "type": "text", + "content": "[20] Mingye Ju, Can Ding, Charles A. Guo, Wenqi Ren, and Dacheng Tao. IDRLP: image dehazing using region line prior. IEEE Transactions on Image Processing, 30:9043-9057, 2021. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 426, + 554, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 426, + 554, + 470 + ], + "spans": [ + { + "bbox": [ + 317, + 426, + 554, + 470 + ], + "type": "text", + "content": "[21] Boyi Li, Wenqi Ren, Dengpan Fu, Dacheng Tao, Dan Feng, Wenjun Zeng, and Zhangyang Wang. Benchmarking single-image dehazing and beyond. IEEE Transactions on Image Processing, 28(1):492-505, 2018. 2, 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 472, + 554, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 472, + 554, + 514 + ], + "spans": [ + { + "bbox": [ + 317, + 472, + 554, + 514 + ], + "type": "text", + "content": "[22] Boyi Li, Wenqi Ren, Dengpan Fu, Dacheng Tao, Dan Feng, Wenjun Zeng, and Zhangyang Wang. Benchmarking single-image dehazing and beyond. IEEE Transactions on Image Processing, 28(1):492-505, 2019. 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 515, + 554, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 515, + 554, + 568 + ], + "spans": [ + { + "bbox": [ + 317, + 515, + 554, + 568 + ], + "type": "text", + "content": "[23] Boyun Li, Xiao Liu, Peng Hu, Zhongqin Wu, Jiancheng Lv, and Xi Peng. All-in-one image restoration for unknown corruption. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17431-17441, 2022. 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 570, + 554, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 570, + 554, + 613 + ], + "spans": [ + { + "bbox": [ + 317, + 570, + 554, + 613 + ], + "type": "text", + "content": "[24] Pengyue Li, Jiandong Tian, Yandong Tang, Guolin Wang, and Chengdong Wu. Deep retina network for single image dehazing. IEEE Transactions on Image Processing, 30: 1100-1115, 2021. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 614, + 554, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 614, + 554, + 657 + ], + "spans": [ + { + "bbox": [ + 317, + 614, + 554, + 657 + ], + "type": "text", + "content": "[25] Yu Li, Robby T. Tan, and Michael S. Brown. Nighttime haze removal with glow and multiple light colors. In Proceedings of the International Conference on Computer Vision, pages 226-234, 2015. 7" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 658, + 554, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 658, + 554, + 712 + ], + "spans": [ + { + "bbox": [ + 317, + 658, + 554, + 712 + ], + "type": "text", + "content": "[26] Yudong Liang, Bin Wang, Wangmeng Zuo, Jiaying Liu, and Wenqi Ren. Self-supervised learning and adaptation for single image dehazing. In Proceedings of the Thirty-First International Joint Conference on Artificial Intelligence, pages 1137–1143, 2022. 2" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11205" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "text", + "content": "[27] Pengyang Ling, Huaian Chen, Xiao Tan, Yi Jin, and Enhong Chen. Single image dehazing using saturation line prior. IEEE Transactions on Image Processing, 32:3238-3253, 2023. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 162 + ], + "type": "text", + "content": "[28] Jun Liu, Ryan Wen Liu, Jianing Sun, and Tieyong Zeng. Rank-one prior: Real-time scene recovery. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(7): 8845-8860, 2023. 1, 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 294, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 294, + 217 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 294, + 217 + ], + "type": "text", + "content": "[29] Risheng Liu, Jiaxin Gao, Jin Zhang, Deyu Meng, and Zhouchen Lin. Investigating bi-level optimization for learning and vision from a unified perspective: A survey and beyond. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(12):10045-10067, 2021. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 220, + 294, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 220, + 294, + 274 + ], + "spans": [ + { + "bbox": [ + 56, + 220, + 294, + 274 + ], + "type": "text", + "content": "[30] Ye Liu, Lei Zhu, Shunda Pei, Huazhu Fu, Jing Qin, Qing Zhang, Liang Wan, and Wei Feng. From synthetic to real: Image dehazing collaborating with unlabeled real data. In Proceedings of the ACM International Conference on Multimedia, pages 50-58, 2021. 2, 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 276, + 294, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 276, + 294, + 330 + ], + "spans": [ + { + "bbox": [ + 56, + 276, + 294, + 330 + ], + "type": "text", + "content": "[31] Yidi Liu, Xingbo Wang, Yurui Zhu, Xueyang Fu, and Zheng-Jun Zha. Sdcnet: Spatially-adaptive deformable convolution networks for hr nonhomogeneous dehazing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6682-6691, 2024. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 332, + 294, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 332, + 294, + 385 + ], + "spans": [ + { + "bbox": [ + 56, + 332, + 294, + 385 + ], + "type": "text", + "content": "[32] Long Ma, Tengyu Ma, Risheng Liu, Xin Fan, and Zhongxuan Luo. Toward fast, flexible, and robust low-light image enhancement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5637-5646, 2022. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 388, + 294, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 388, + 294, + 431 + ], + "spans": [ + { + "bbox": [ + 56, + 388, + 294, + 431 + ], + "type": "text", + "content": "[33] Ozan Özdenizci and Robert Legenstein. Restoring vision in adverse weather conditions with patch-based denoising diffusion models. IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 1-12, 2023. 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 434, + 294, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 434, + 294, + 465 + ], + "spans": [ + { + "bbox": [ + 56, + 434, + 294, + 465 + ], + "type": "text", + "content": "[34] Yuda Song, Zhuqing He, Hui Qian, and Xin Du. Vision transformers for single image dehazing. IEEE Transactions on Image Processing, 32:1927-1941, 2023. 5, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 468, + 294, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 468, + 294, + 521 + ], + "spans": [ + { + "bbox": [ + 56, + 468, + 294, + 521 + ], + "type": "text", + "content": "[35] Yi Tang, Hiroshi Kawasaki, and Takafumi Iwaguchi. Underwater image enhancement by transformer-based diffusion model with non-uniform sampling for skip strategy. In Proceedings of the ACM International Conference on Multimedia, pages 5419-5427, 2023. 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 523, + 294, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 523, + 294, + 578 + ], + "spans": [ + { + "bbox": [ + 56, + 523, + 294, + 578 + ], + "type": "text", + "content": "[36] Zhongze Wang, Haitao Zhao, Jingchao Peng, Lujian Yao, and Kaijie Zhao. Odcr: Orthogonal decoupling contrastive regularization for unpaired image dehazing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25479-25489, 2024. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 579, + 294, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 579, + 294, + 634 + ], + "spans": [ + { + "bbox": [ + 56, + 579, + 294, + 634 + ], + "type": "text", + "content": "[37] Ruiqi Wu, Zheng-Peng Duan, Chun-Le Guo, Zhi Chai, and Chongyi Li. RIDCP: revitalizing real image dehazing via high-quality codebook priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22282-22291, 2023. 2, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 635, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 294, + 689 + ], + "type": "text", + "content": "[38] Yang Yang, Chaoyue Wang, Risheng Liu, Lin Zhang, Xiaojie Guo, and Dacheng Tao. Self-augmented unpaired image dehazing via density and depth decomposition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2037-2046, 2022. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 692, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 692, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 692, + 294, + 713 + ], + "type": "text", + "content": "[39] Yang Yang, Chaoyue Wang, Xiaojie Guo, and Dacheng Tao. Robust unpaired image dehazing via density and depth de" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 441 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "text", + "content": "composition. International Journal of Computer Vision, 132 (5):1557-1577, 2024. 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 96, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 150 + ], + "type": "text", + "content": "[40] Dehuan Zhang, Jingchun Zhou, Chunle Guo, Weishi Zhang, and Chongyi Li. Synergistic multiscale detail refinement via intrinsic supervision for underwater image enhancement. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 7033-7041, 2024. 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 152, + 553, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 553, + 195 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 553, + 195 + ], + "type": "text", + "content": "[41] He Zhang and Vishal M Patel. Densely connected pyramid dehazing network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3194-3203, 2018. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 197, + 553, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 197, + 553, + 240 + ], + "spans": [ + { + "bbox": [ + 316, + 197, + 553, + 240 + ], + "type": "text", + "content": "[42] Jing Zhang, Yang Cao, and et al. Fast haze removal for nighttime image using maximum reflectance prior. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7016-7024, 2017. 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 241, + 553, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 241, + 553, + 284 + ], + "spans": [ + { + "bbox": [ + 316, + 241, + 553, + 284 + ], + "type": "text", + "content": "[43] Jing Zhang, Yang Cao, and et al. Nighttime dehazing with a synthetic benchmark. In Proceedings of the 28th ACM International Conference on Multimedia, pages 2355-2363, 2020. 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 286, + 553, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 286, + 553, + 341 + ], + "spans": [ + { + "bbox": [ + 316, + 286, + 553, + 341 + ], + "type": "text", + "content": "[44] Yafei Zhang, Shen Zhou, and Huafeng Li. Depth information assisted collaborative mutual promotion network for single image dehazing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2846-2855, 2024. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 342, + 553, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 342, + 553, + 397 + ], + "spans": [ + { + "bbox": [ + 316, + 342, + 553, + 397 + ], + "type": "text", + "content": "[45] Dian Zheng, Xiao-Ming Wu, Shuzhou Yang, Jian Zhang, Jian-Fang Hu, and Wei-Shi Zheng. Selective hourglass mapping for universal image restoration based on diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 399, + 553, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 399, + 553, + 441 + ], + "spans": [ + { + "bbox": [ + 316, + 399, + 553, + 441 + ], + "type": "text", + "content": "[46] Yu Zheng, Jiahui Zhan, and et al. Curricular contrastive regularization for physics-aware single image dehazing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition., 1, 6" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "11206" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/CoCoGaussian_ Leveraging Circle of Confusion for Gaussian Splatting from Defocused Images/c9f7dcb9-6d18-430a-9426-253bcaf437ef_content_list.json b/2025/CoCoGaussian_ Leveraging Circle of Confusion for Gaussian Splatting from Defocused Images/c9f7dcb9-6d18-430a-9426-253bcaf437ef_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..dba2cec8e9da068a971e694a993f58f0e5277203 --- /dev/null +++ b/2025/CoCoGaussian_ Leveraging Circle of Confusion for Gaussian Splatting from Defocused Images/c9f7dcb9-6d18-430a-9426-253bcaf437ef_content_list.json @@ -0,0 +1,2014 @@ +[ + { + "type": "text", + "text": "CoCoGaussian: Leveraging Circle of Confusion for Gaussian Splatting from Defocused Images", + "text_level": 1, + "bbox": [ + 238, + 130, + 759, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jungho Lee $^{1*}$ Suhwan Cho $^{1}$ Taeoh Kim $^{2}$ Ho-Deok Jang $^{2}$ Minhyeok Lee $^{1}$ Geonho Cha $^{2}$ Dongyoon Wee $^{2}$ Dogyoon Lee $^{1}$ Sangyoun Lee $^{1}$", + "bbox": [ + 196, + 195, + 799, + 232 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ School of Electrical and Electronic Engineering, Yonsei University $^{2}$ NAVER Cloud https://Jho-Yonsei.github.io/CoCoGaussian", + "bbox": [ + 183, + 238, + 815, + 273 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6372921ee99eef124a57bdb275cb484704d28be09f6167b5e9ae6d6aab16c7fc.jpg", + "image_caption": [ + "(a) Defocused Image Acquisition" + ], + "image_footnote": [], + "bbox": [ + 129, + 284, + 354, + 523 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/209813a7d770da27e57b642b8969f2261761f3a17b2fb0a2ec7343a2402da039.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 372, + 286, + 591, + 400 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8365538702c8e3e10961f876106bfabd8df9d0464e33d1cef9a46205c55397c3.jpg", + "image_caption": [ + "(b) Reconstructed Gaussian Points" + ], + "image_footnote": [], + "bbox": [ + 374, + 411, + 591, + 522 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/972cfdd3b13a32857f3418c0a1d90d6d6e2da4e6a0362463315c24f998ed614d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 602, + 323, + 633, + 364 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a0dfa3ea76ab11e9d7b02e1ac7d0a8ead6cf0eb7294973e9e2ff490e930b63fe.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 643, + 287, + 867, + 405 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/fdb771382f32b5abcb68f80cc05014b4105a772e4f61b24dfecad55370309602.jpg", + "image_caption": [ + "(c) Rendered Images", + "Figure 1. We propose CoCoGaussian, a novel framework for 3D scene reconstruction from defocused images. (a) With a large aperture, radiance from the focus plane appears as a small circle of confusion (green) on the image sensor, while radiance from greater depths results in larger circles (pink). (b) We visualize reconstructed Gaussians of CoCoGaussian, where the black dots indicate Gaussian means, while the gray dots represent Gaussian means forming the circle of confusion, which decreases in size near the focus plane. (c) The defocused image focuses on shallow depths, with circle size increasing with depth, as shown in (b). CoCoGaussian allows customization of defocused images by adjusting the depth of field or focus plane, while sharp images can be obtained by rendering without the circle of confusion." + ], + "image_footnote": [], + "bbox": [ + 643, + 410, + 866, + 523 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 638, + 326, + 655 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3D Gaussian Splatting (3DGS) has attracted significant attention for its high-quality novel view rendering, inspiring research to address real-world challenges. While conventional methods depend on sharp images for accurate scene reconstruction, real-world scenarios are often affected by defocus blur due to finite depth of field, making it essential to account for realistic 3D scene representation. In this study, we propose CoCoGaussian, a Circle of Confusion-aware Gaussian Splatting that enables precise 3D scene representation using only defocused images. CoCoGaussian addresses the challenge of defocus blur by modeling the Circle of Confusion (CoC) through a physically grounded approach based on the principles of photo", + "bbox": [ + 88, + 672, + 485, + 869 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "graphic defocus. Exploiting 3D Gaussians, we compute the CoC diameter from depth and learnable aperture information, generating multiple Gaussians to precisely capture the CoC shape. Furthermore, we introduce a learnable scaling factor to enhance robustness and provide more flexibility in handling unreliable depth in scenes with reflective or refractive surfaces. Experiments on both synthetic and real-world datasets demonstrate that CoCoGaussian achieves state-of-the-art performance across multiple benchmarks.", + "bbox": [ + 511, + 641, + 908, + 779 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 811, + 645, + 828 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The emergence of the Neural Radiance Field (NeRF) [20] has brought significant attention to 3D scene representation for photo-realistic novel view synthesis. They reconstruct 3D scenes from images captured from multiple views", + "bbox": [ + 511, + 839, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*This work was done during an internship at Naver Cloud.", + "bbox": [ + 107, + 886, + 431, + 901 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "16101", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "and render images of unseen views. However, since NeRF models 3D scenes as implicit neural representations through a ray tracing-based approach, they suffer from inefficient memory usage. In contrast, a rasterization-based method, 3D Gaussian Splatting (3DGS) [9] explicitly models 3D scenes using independent 3D Gaussians with different 3D means, covariances, opacities, and spherical harmonic coefficients. The tile-based rasterizer of 3DGS applies alpha blending to Gaussian splats sorted by visibility order, enabling fast training, rendering, and efficient memory usage.", + "bbox": [ + 89, + 90, + 480, + 242 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Both NeRF [20] and 3DGS [9] rely on sharp images to represent 3D scenes accurately, which is a highly ideal assumption. In real-world settings, various factors (e.g., finite depth of field and camera motion blur) can hinder the capture of sharp images, leading to image degradation. Achieving a perfectly sharp image requires every part of the scene to be in focus, namely all-in-focus, which calls for a large depth of field. However, obtaining a large depth of field necessitates using a small aperture, which in turn requires a longer exposure time to allow enough light to enter the camera. During this extended exposure, even slight movements of a handheld camera can introduce motion blur. To prevent motion blur, the aperture needs to be widened, but this comes at the cost of a shallower depth of field. As a result, areas outside the focus plane appear blurred due to defocus. Defocus blur is closely related to the concept of the Circle of Confusion (CoC); as shown in Fig. 1, when a subject is positioned away from the focus plane, the radiances from the subject pass through the aperture and are mapped onto the image sensor in a circular pattern, resulting in shallow depth of field and defocus blur [7]. However, despite these challenges, recent research in 3D scene representation has rarely focused on addressing defocus blur directly. In this paper, we propose a method to represent 3D scenes using only images with defocus blur, tackling the CoC-related challenges posed by real-world image acquisition scenarios.", + "bbox": [ + 91, + 244, + 483, + 637 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recently, there has been increasing interest in reconstructing 3D scenes using only degraded images, with a focus on addressing challenges caused by camera motion blur and defocus blur. Deblur-NeRF [18] is the first study to take on this task, adopting a blind image deblurring approaches [2, 31, 40]. Follow-up studies [10-13, 25, 26, 37] have expanded this approach beyond raytracing to rasterization-based methods. However, many of these methods depend heavily on learning-based strategies and fail to incorporate photographic principles that accurately capture how defocus occurs in real-world scenarios. While DoF-NeRF [41] adopts the concept of CoC to model defocus blur in 3D scene representations, it relies on an implicit representation, which leads to slow rendering speeds, making it less suitable for real-time applications.", + "bbox": [ + 89, + 641, + 482, + 867 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we propose CoCoGaussian, which leverages a physically grounded photographic defocus princi", + "bbox": [ + 89, + 869, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "plies to represent 3D scenes. Our method incorporates an aperture to calculate the CoC diameter based on depth and learnable aperture information using 3D Gaussians, accurately modeling shallow depth of field images. Using the CoC diameter and the 3D Gaussian, we generate multiple Gaussians to form the CoC shape, as illustrated in Fig. 1. However, when objects with refraction or reflection exist in the scene, the depth obtained through the Gaussians may become unreliable. To address this, we propose a method to keep the set of generated Gaussians within the CoC radius, introducing a learnable scaling factor to increase flexibility and reduce dependence on estimated depth in the CoC modeling. As a result, our model combines physically grounded defocus principles with an adaptive CoC modeling that effectively handles uncertain depths, allowing it to reconstruct sharp 3D scenes using only defocused images. Additionally, our approach has an advantage of learning aperture and focus plane information, enabling dynamic control over depth of field and flexible adjustment of the focus plane. This feature allows for highly customizable scene visualization, adapting to various focus and depth requirements.", + "bbox": [ + 511, + 90, + 903, + 407 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To demonstrate the effectiveness of our model, we conduct extensive experiments on the benchmarks: the Deblur-NeRF [18] dataset and the DoF-NeRF [41] real-world dataset. Our contributions can be summarized as follows:", + "bbox": [ + 511, + 409, + 903, + 468 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- CoCoGaussian models the CoC at the 3D Gaussian level, reconstructing the precise 3D scene and enabling sharp novel view synthesis from defocused images.", + "- We propose an adaptive learning approach that robustly models the CoC even with unreliable depth information.", + "- CoCoGaussian enables the customizable depth of field and flexible focus plane adjustment during rendering by learning aperture and focus plane information.", + "- CoCoGaussian achieves state-of-the-art performance, both quantitatively and qualitatively." + ], + "bbox": [ + 511, + 474, + 903, + 633 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 511, + 652, + 653, + 669 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Scene Representations for Novel View Synthesis", + "text_level": 1, + "bbox": [ + 511, + 679, + 903, + 695 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3D scene representation for novel view synthesis has advanced significantly with the advent of neural radiance fields (NeRF) [20], a ray tracing-based volumetric rendering method that generates photo-realistic novel view images from multi-view images. However, as an implicit representation method using deep MLPs, NeRF suffers from slow training and rendering speeds. To address this limitation, explicit representation methods such as Plenoxel [6], TensoRF [3], and Instant-NGP [21] have been introduced. More recently, 3D Gaussian Splatting (3DGS) [9], a rasterization-based method, has emerged as an alternative to ray tracing, significantly mitigating the speed and memory efficiency constraints of ray tracing", + "bbox": [ + 511, + 704, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "16102", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "based approaches. NeRF and 3DGS has enabled a wide range of related research, including dynamic scene representation [14, 15, 23, 24, 28, 33], human avatars [8, 27, 39], 3D mesh reconstruction [16, 32, 35, 36], 3D scene representation from sparse-view images [22, 34, 42, 43], and 3D scene representation from blurry images [4, 10-13, 18, 25, 26, 37, 46]. In this paper, we propose a method for 3D scene representation from defocused images by exploiting 3DGS.", + "bbox": [ + 89, + 90, + 483, + 212 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Novel View Synthesis from Blurry Images", + "text_level": 1, + "bbox": [ + 89, + 223, + 447, + 239 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Scene representation methods like NeRF [20] and 3DGS [9] require sharp images as input to render photo-realistic images. However, in real-world scenarios, capturing sharp images is challenging, often suffering from image degradation such as camera motion blur and defocus blur. To address this issue, Deblur-NeRF [18] introduced a 3D ray-based blurring kernel inspired by image blind deblurring [2, 31, 40], which intentionally generates blurry images during training. After training, it renders sharp novel-view images by excluding the trained kernel. Following Deblur-NeRF, various approaches have been proposed to tackle the degradation issue. DP-NeRF [11] introduces a kernel based on the rigid body transformation [17] with prior knowledge that object shape remains consistent in static scenes, and PDRF [25] proposes a blur estimation method with 2-stage efficient rendering. Recently, methods utilizing 3DGS have been proposed to enable faster rendering. For instance, Deblurring 3DGS [10] adjusts Gaussian parameters to generate blurry images during training, and BAGS [26] proposes a blur-agnostic kernel and a blur mask using convolutional neural networks (CNNs). DoF-NeRF [41] has explored modeling defocus blur by leveraging the circle of confusion (CoC) within an implicit neural representation. In this paper, we propose a method that also leverages CoC to model defocus, which enables real-time rendering, while still maintaining high-quality depth-of-field effects.", + "bbox": [ + 91, + 247, + 483, + 640 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Preliminary", + "text_level": 1, + "bbox": [ + 89, + 655, + 217, + 672 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. 3D Gaussian Splatting", + "text_level": 1, + "bbox": [ + 89, + 680, + 300, + 696 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Unlike ray tracing-based methods [1, 3, 20], 3DGS [9] is built on a rasterization-based approach with differentiable 3D Gaussians. These 3D Gaussians are initialized from a sparse point cloud obtained via a Structure-from-Motion (SfM) [29, 30] algorithm and are defined as follows:", + "bbox": [ + 89, + 704, + 483, + 780 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nG (\\mathbf {x}) = e ^ {- \\frac {1}{2} (\\mathbf {x} - \\mu) ^ {\\top} \\boldsymbol {\\Sigma} ^ {- 1} (\\mathbf {x} - \\mu)}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 184, + 791, + 482, + 811 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbf{x} \\in \\mathbb{R}^3$ is a point on the Gaussian $G$ centered at the mean vector $\\mu \\in \\mathbb{R}^3$ with an covariance matrix $\\boldsymbol{\\Sigma} \\in \\mathbb{R}^{3 \\times 3}$ . The 3D covariance matrix $\\boldsymbol{\\Sigma}$ is derived from a learnable scaling vector $\\mathbf{s} \\in \\mathbb{R}^3$ and rotation quaternion $\\mathbf{q} \\in \\mathbb{R}^4$ , from which the scaling matrix $\\mathbf{S} \\in \\mathbb{R}^{3 \\times 3}$ and rotation matrix $\\mathbf{R} \\in$", + "bbox": [ + 89, + 824, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "$\\mathbb{R}^{3\\times 3}$ are obtained and represented as follows:", + "bbox": [ + 513, + 90, + 821, + 106 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\Sigma} = \\mathbf {R} \\mathbf {S} \\mathbf {S} ^ {\\top} \\mathbf {R} ^ {\\top}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 651, + 111, + 903, + 128 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For differentiable splatting [44], the Gaussians in the 3D world coordinate system are projected into the 2D camera coordinate system. This projection uses the viewing transformation $\\mathbf{W} \\in \\mathbb{R}^{3 \\times 3}$ and the Jacobian $\\mathbf{J} \\in \\mathbb{R}^{2 \\times 3}$ of the affine approximation of the projective transformation to derive the 2D covariance $\\Sigma^{2\\mathrm{D}} \\in \\mathbb{R}^{2 \\times 2}$ :", + "bbox": [ + 511, + 135, + 906, + 224 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\Sigma} ^ {2 \\mathrm {D}} = \\mathbf {J} \\mathbf {W} \\boldsymbol {\\Sigma} \\mathbf {W} ^ {\\top} \\mathbf {J} ^ {\\top}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 632, + 229, + 903, + 250 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Each Gaussian includes a set of spherical harmonics (SH) coefficients and an opacity value $\\alpha$ to represent view-dependent color $\\mathbf{c}$ . The pixel color $\\mathbf{c}_p$ is then obtained by applying alpha blending to $\\mathcal{N}$ ordered Gaussians:", + "bbox": [ + 511, + 255, + 905, + 316 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {c} _ {p} = \\sum_ {i \\in \\mathcal {N}} \\mathbf {c} _ {i} \\alpha_ {i} \\prod_ {j = 1} ^ {i - 1} (1 - \\alpha_ {j}). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 611, + 321, + 903, + 364 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our approach, as illustrated in Fig. 1, exploits 3DGS and aims to render defocused images by generating multiple Gaussians to form the CoC for each Gaussian.", + "bbox": [ + 511, + 369, + 905, + 415 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. 3D Scene Blind Deblurring", + "text_level": 1, + "bbox": [ + 511, + 422, + 758, + 438 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Image blind deblurring is a technique for learning an image blurring kernel without the supervision of sharp images. Blurry images are obtained by applying convolution with a learned blurring kernel to sharp images, where the kernel size is fixed as a grid structure at each pixel position. Inspired by this approach, Deblur-NeRF [18] applies this technique to NeRF, presenting a method to represent a clean 3D scene using only blurry images. Specifically, Deblur-NeRF warps the input ray into multiple rays that constitute the blur, modeling a ray-based sparse kernel. A blurry pixel is obtained by combining the pixel colors of each ray:", + "bbox": [ + 511, + 445, + 906, + 611 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {c} _ {b l u r} = \\sum w _ {p} ^ {i} \\mathbf {c} _ {p} ^ {i}, w. r. t., \\sum w _ {p} ^ {i} = 1, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 618, + 903, + 640 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $w_{p}$ and $\\mathbf{c}_p$ denote the weight and pixel color for each corresponding ray. Deblur-NeRF enhances learning efficiency by setting the number of warped rays smaller than the kernel size of 2D convolution, and it designs a deformable kernel by adaptive origins and directions of rays.", + "bbox": [ + 511, + 645, + 905, + 719 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As we adopt a rasterization method based on 3DGS [9], we propose an alternative approach: instead of warping rays, we generate 3D Gaussians in the shape of the CoC and render them. In other words, the defocused image is estimated by setting the generated 3D Gaussians, namely CoC Gaussians, as the blurring kernel.", + "bbox": [ + 511, + 720, + 905, + 811 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Method", + "text_level": 1, + "bbox": [ + 511, + 821, + 604, + 837 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1. CoCoGaussian Framework", + "text_level": 1, + "bbox": [ + 511, + 847, + 759, + 861 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our goal is to represent a 3D scene using only defocused multi-view images, leveraging fundamental photographic", + "bbox": [ + 511, + 869, + 905, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "16103", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/4169d199f407b6ec6e4cce324d5394acd0ea85e11d85e29f803b274456542d6c.jpg", + "image_caption": [ + "Figure 2. We take the camera position $\\mathbf{x}_{cam}$ and the base Gaussian parameters $\\mu_B$ , $\\mathbf{s}_B$ , and $\\mathbf{q}_B$ as inputs to the MLP $h_\\theta$ , which produces five outputs in total. (a) We set the depth $d(\\mu_B)$ as the Euclidean distance between $\\mathbf{x}_{cam}$ and $\\mu_B$ . Using the output $K$ from $h_\\theta$ , the $d(\\mu_B)$ and the learnable focus plane $d_F$ , we apply Eq. (7) to determine the CoC diameter. The diameter, combined with the outputs $\\beta$ and $\\mathbf{d}$ from $h_\\theta$ , is then used in Eq. (9) to compute the offset values for $\\mu_{\\mathrm{CoC}}$ . By adding these offsets to $\\mu_B$ , we obtain the $\\mathbf{G}_{\\mathrm{CoC}}$ means. (b) We obtain $\\mathbf{s}_{\\mathrm{CoC}}$ and $\\mathbf{q}_{\\mathrm{CoC}}$ by applying Eqs. (10) and (11) to $\\delta \\mathbf{s}_{\\mathrm{CoC}}$ and $\\delta \\mathbf{q}_{\\mathrm{CoC}}$ , the outputs of $h_\\theta$ . (c) Using $\\mu_{\\mathrm{CoC}}$ , $\\mathbf{s}_{\\mathrm{CoC}}$ , and $\\mathbf{q}_{\\mathrm{CoC}}$ , we get the $\\mathbf{G}_{\\mathrm{CoC}}$ . Finally, we rasterize $\\mathbf{G}_{\\mathrm{CoC}}$ along with $\\mathbf{G}_B$ to produce $(M + 1)$ images, and apply a weighted sum to generate the final defocused image." + ], + "image_footnote": [], + "bbox": [ + 98, + 88, + 903, + 402 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "principles to guide our method. We generate 3D Gaussians to capture a blurring kernel shaped as the CoC. The whole framework of CoCoGaussian is on Fig. 2. First, we compute the CoC diameter using the Gaussian depth, derived from the means of the given base 3D Gaussian set $\\mathbf{G}_B$ , which consists of $N$ Gaussians, along with the camera position and aperture information (Sec. 4.2). Next, for the base Gaussian set $\\mathbf{G}_B$ , we generate $M$ CoC Gaussian sets $\\mathbf{G}_{\\mathrm{CoC}}$ , resulting in a total of $(M\\times N)$ Gaussians to capture the CoC shape. However, when the scene contains objects with refraction or reflection, the Gaussian depth may be unreliable, which can lead to suboptimal CoC diameter. To address this issue, we propose an adaptive CoC Gaussian generation method that minimizes the dependence on depth for $\\mathbf{G}_{\\mathrm{CoC}}$ (Sec. 4.3). Finally, we incorporate adaptive rotation quaternion and scaling vector adjustments, inspired by the Deblurring 3DGS [10] approach, to allow flexible learning of $\\mathbf{G}_{\\mathrm{CoC}}$ . Furthermore, we apply Eq. (5) to images rendered from $\\mathbf{G}_B$ and $\\mathbf{G}_{\\mathrm{CoC}}$ , using a weighted sum approach [12, 26] to create the final blurry image. This photography-prior-based approach allows us to accurately represent 3D scenes using only defocused images while adhering to realistic defocus principles.", + "bbox": [ + 91, + 523, + 483, + 872 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Circle of Confusion from 3D Gaussians", + "text_level": 1, + "bbox": [ + 511, + 523, + 852, + 539 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To model the CoC, we assume an ideal system that ignores distortions such as lens aberrations. Before generating the CoC Gaussian sets $\\mathbf{G}_{\\mathrm{CoC}}$ , we first need to obtain the CoC diameter, which depends on several factors: the aperture diameter $D$ , focal length $f$ , focus plane $d_{F}$ , and depth $p$ . The focus plane represents the distance from the camera position to the plane in focus. We define the depth $d(\\mu_B)$ as the Euclidean distance between the camera position $\\mathbf{x}_{cam} \\in \\mathbb{R}^3$ and the means of the base Gaussian set $\\mu_B \\in \\mathbb{R}^{N \\times 3}$ . Exploiting these values, we express the CoC diameter $\\sigma(\\mu_B)$ as follows [7]:", + "bbox": [ + 511, + 547, + 906, + 714 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\sigma (\\mu_ {B}) = f D \\times \\frac {\\left| d (\\mu_ {B}) - d _ {F} \\right|}{d (\\mu_ {B}) (d _ {F} - f)}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 594, + 724, + 906, + 758 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $d_F$ is set to a learnable parameter and initialized as the average distance between $\\mathbf{x}_{cam}$ and the points in the SfM point cloud to ensure stable training. Note that the focus plane $d_F$ is specific to each input image. Additionally, because the focal length $f$ is usually much smaller than the focus plane $d_F$ , we modify Eq. (6) as follows:", + "bbox": [ + 511, + 768, + 908, + 859 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\sigma \\left(\\mu_ {B}\\right) \\approx K \\times \\left| \\frac {1}{d \\left(\\mu_ {B}\\right)} - \\frac {1}{d _ {F}} \\right|, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 601, + 869, + 906, + 905 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "16104", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where the product of the focal length and aperture diameter is represented as a single learnable scalar $K = f \\times D$ [41].", + "bbox": [ + 89, + 90, + 482, + 119 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We design a simple MLP $h_\\theta$ parameterized by $\\theta$ to estimate the scalar $K$ . Then, the aperture parameter $K$ is obtained by passing these features as inputs to $h_\\theta$ , along with the scaling vectors $\\mathbf{s}_B \\in \\mathbb{R}^{N \\times 3}$ and rotation quaternions $\\mathbf{q}_B \\in \\mathbb{R}^{N \\times 4}$ of $\\mathbf{G}_B$ . This approach allows us to obtain the CoC diameter $\\sigma(\\mu_B)$ for the $\\mathbf{G}_B$ , marking the first step in modeling $\\mathbf{G}_{\\mathrm{CoC}}$ .", + "bbox": [ + 89, + 121, + 483, + 227 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Adaptive CoC Gaussian Generation", + "text_level": 1, + "bbox": [ + 89, + 237, + 403, + 253 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After obtaining the CoC diameters, we use these diameters to generate $M$ CoC Gaussian sets $\\mathbf{G}_{\\mathrm{CoC}}$ . For the Gaussian parameters of $\\mathbf{G}_{\\mathrm{CoC}}$ , we set the spherical harmonics (SH) coefficients and opacity values identical to those of $\\mathbf{G}_B$ , while varying only the means $\\mu$ , scaling vectors $\\mathbf{s}$ , and rotation quaternions $\\mathbf{q}$ . The means of each CoC Gaussian set are created by adding offsets $\\Delta \\mu_{\\mathrm{CoC}} \\in \\mathbb{R}^{M \\times N \\times 3}$ to $\\mu_B$ , defined as $\\mu_{\\mathrm{CoC}} = \\mu_B + \\Delta \\mu_{\\mathrm{CoC}}$ . Since $\\Delta \\mu_{\\mathrm{CoC}}$ are 3D vector sets, we obtain unit vector sets $\\mathbf{d} \\in \\mathbb{R}^{M \\times N \\times 3}$ representing the directions from $\\mathbf{G}_B$ to $\\mathbf{G}_{\\mathrm{CoC}}$ , which are additional outputs from the MLP $h_\\theta$ discussed in the previous section. Moreover, since $\\mathbf{G}_{\\mathrm{CoC}}$ consists of a total of $M$ Gaussian sets, The direction vectors $\\mathbf{d}$ consist of $M$ instances:", + "bbox": [ + 89, + 260, + 483, + 455 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\Delta \\mu_ {\\mathrm {C o C}; m} = \\frac {\\sigma (\\mu_ {B})}{2} \\mathbf {d} _ {m}, \\text {w h e r e} m \\in M. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 145, + 465, + 482, + 496 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "However, theses offsets $\\Delta \\mu_{\\mathrm{CoC}}$ presents two potential issues. First, this approach places all $\\mathbf{G}_{\\mathrm{CoC}}$ only on the boundary of the CoC, which limits its ability to fully capture the defocus effect. Modeling the entire CoC, including its interior, would allow for a more accurate and flexible representation of defocus blur. The second issue arises when there are refractive or reflective surfaces in the scene, as these subjects can lead to incorrect optimization of the Gaussian means, making the CoC diameter obtained through Eq. (7) unreliable. To address these problems, we introduce simple learnable CoC scaling parameters $\\beta \\in (0,1]$ , which are additional outputs of $h_\\theta$ . The parameter $\\beta$ is learned within the range of 0 to 1, ensuring that $\\mu_{\\mathrm{CoC}}$ resides within the CoC boundary. Additionally, even when depth is inaccurately measured and the predicted CoC diameter is larger than the actual size, $\\beta$ adaptively scales it down. Thus, the offset $\\Delta \\mu_{\\mathrm{CoC}}$ is modified by applying $\\beta$ as follows:", + "bbox": [ + 89, + 506, + 483, + 775 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\Delta \\mu_ {\\mathrm {C o C}; m} = \\frac {\\sigma (\\mu_ {B})}{2} \\beta_ {m} \\mathbf {d} _ {m}. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 194, + 776, + 480, + 804 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Additionally, similar to Deblurring 3DGS [10], we introduce scaling factors $\\delta \\mathbf{s}_{\\mathrm{CoC}} \\in \\mathbb{R}^{M \\times N \\times 3}$ and $\\delta \\mathbf{q}_{\\mathrm{CoC}} \\in \\mathbb{R}^{M \\times N \\times 4}$ , which allow flexible learning of the scaling vectors $\\mathbf{s}_{\\mathrm{CoC}}$ and rotation quaternions $\\mathbf{q}_{\\mathrm{CoC}}$ of $\\mathbf{G}_{\\mathrm{CoC}}$ . These factors are also the outputs of $h_{\\theta}$ . The parameters $\\delta \\mathbf{s}_{\\mathrm{CoC}}$ and $\\delta \\mathbf{q}_{\\mathrm{CoC}}$ are multiplied by the corresponding parameters", + "bbox": [ + 89, + 809, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "of the base Gaussian set $\\mathbf{G}_B$ , specifically $\\mathbf{s}_B$ and $\\mathbf{q}_B$ , and are expressed as follows:", + "bbox": [ + 511, + 90, + 906, + 121 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {s} _ {\\mathrm {C o C}; m} = \\mathbf {s} _ {B} \\times \\delta \\mathbf {s} _ {\\mathrm {C o C}; m}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 624, + 135, + 906, + 152 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {q} _ {\\mathrm {C o C}; m} = \\mathbf {q} _ {B} \\times \\delta \\mathbf {q} _ {\\mathrm {C o C}; m}, \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 622, + 165, + 906, + 180 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where the scaling parameters are constrained by the fixed values $\\delta \\mathbf{s}_{\\mathrm{max}}$ and $\\delta \\mathbf{q}_{\\mathrm{max}}$ : $\\delta \\mathbf{s}_{\\mathrm{CoC}} \\in [1, \\delta \\mathbf{s}_{\\mathrm{max}}]$ and $\\delta \\mathbf{q}_{\\mathrm{CoC}} \\in [1, \\delta \\mathbf{q}_{\\mathrm{max}}]$ .", + "bbox": [ + 511, + 186, + 906, + 233 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We obtain a total of $M$ means of CoC Gaussian sets $\\mu_{\\mathrm{CoC}}$ , scaling vectors $\\mathbf{s}_{\\mathrm{CoC}}$ , and rotation quaternion $\\mathbf{q}_{\\mathrm{CoC}}$ . They are combined with the opacity and SH coefficients of $\\mathbf{G}_B$ , generate the $M$ Gaussian sets $\\mathbf{G}_{\\mathrm{CoC}}$ :", + "bbox": [ + 511, + 234, + 906, + 294 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {x} _ {\\mathrm {C o C}; m}, \\mathbf {s} _ {\\mathrm {C o C}; m}, \\mathbf {q} _ {\\mathrm {C o C}; m}, \\alpha_ {B}, \\mathbf {S H} _ {B} \\rightarrow \\mathbf {G} _ {\\mathrm {C o C}; m}. \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 532, + 308, + 906, + 325 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We rasterize the resulting $M$ Gaussian sets along with the base Gaussian set, producing $(M + 1)$ images. Then, using the weighted sum approach from Eq. (5), as detailed in the following section, we obtain the final blurry image.", + "bbox": [ + 511, + 335, + 905, + 397 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.4. Optimization", + "text_level": 1, + "bbox": [ + 511, + 407, + 653, + 422 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Weighted Sum. To apply Eq. (5) to the $(M + 1)$ images $\\mathbf{I}$ rasterized from $\\mathbf{G}_B$ and $\\mathbf{G}_{\\mathrm{CoC}}$ , we adopt the methods of recent studies [12, 26]. We use a shallow CNN $\\mathcal{F}$ to compute pixel-wise weights $\\mathcal{W}$ for these images, applying a softmax function to ensure that the weights for each pixel sum to 1. Then, we multiply each image by its corresponding weight and sum the results to obtain the final defocused image $\\mathcal{I}_{blur}$ :", + "bbox": [ + 511, + 430, + 906, + 551 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {I} _ {\\text {b l u r}} = \\sum_ {m = 1} ^ {M + 1} \\mathcal {I} _ {m} \\mathcal {W} _ {m}, \\text {w h e r e} \\mathcal {W} = \\operatorname {s o f t m a x} (\\mathcal {F} (\\mathbf {I})), \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 521, + 563, + 905, + 604 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{I}_m$ represents the image rasterized from $\\mathbf{G}_{\\mathrm{CoC};m}$ , and $\\mathcal{I}_{M + 1}$ represents the image rasterized from $\\mathbf{G}_B$ .", + "bbox": [ + 511, + 616, + 905, + 646 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Objective. We minimize the $\\mathcal{L}_1$ loss and D-SSIM loss $\\mathcal{L}_{\\mathrm{D - SSIM}}$ between the ground truth defocused image and the output defocused image. The $\\mathcal{L}_1$ loss reduces the pixelwise differences between images, while the D-SSIM loss minimizes structural differences between them. The final objective $\\mathcal{L}_{rgb}$ is defined as follows:", + "bbox": [ + 511, + 666, + 905, + 758 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {r g b} = (1 - \\lambda) \\mathcal {L} _ {1} + \\lambda \\mathcal {L} _ {\\mathrm {D} - \\mathrm {S S I M}}, \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 602, + 770, + 906, + 787 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where we use $\\lambda = 0.3$ for all experiments.", + "bbox": [ + 511, + 799, + 794, + 814 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 511, + 829, + 645, + 845 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. We evaluate our method on three different datasets: the Deblur-NeRF [18] synthetic dataset, the Deblur-NeRF real-world dataset, and the DoF-NeRF [41]", + "bbox": [ + 511, + 854, + 906, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "16105", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/40e329192089d6e354522d1e258dcb776c6db337478a0e09ed574fff99be737f.jpg", + "table_caption": [ + "Table 1. Comparisons on Deblur-NeRF synthetic and real-world scene dataset. We evaluate the performance on three metrics (PSNR, SSIM, LPIPS). “*” denotes the results obtained by reproducing the released code. The orange and yellow cells respectively indicate the highest and second-highest value." + ], + "table_footnote": [], + "table_body": "
MethodsSynthetic Scene [18]Real-World Scene [18]
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
Naive NeRF [20]25.950.77910.230322.400.66610.2310
3DGS [9]25.110.74760.214823.380.66550.3140
Deblur-NeRF [18]28.370.85270.118823.470.71990.1207
PDRF-10 [25]30.080.89310.110123.850.73820.1746
DP-NeRF [11]29.330.87130.098723.670.72990.1082
Deblurring 3DGS* [10]28.900.89120.105223.540.73830.1232
BAGS* [26]30.650.91280.063123.480.74080.0962
Ours30.840.92120.047823.700.75310.0825
", + "bbox": [ + 102, + 172, + 480, + 327 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/ebb3ab6669fad17098da7b494f089b22ec4289573a1de66f91ef99fd138aedee.jpg", + "table_caption": [ + "Table 2. Comparisons on DoF-NeRF real-world scene dataset." + ], + "table_footnote": [], + "table_body": "
MethodsDoF-NeRF Real-World Scene [41]
PSNR↑SSIM↑LPIPS↓
3DGS [9]25.720.82910.1817
DP-NeRF [11]26.800.81450.1790
Deblurring 3DGS [10]26.580.83860.1708
BAGS [26]29.870.88160.1100
Ours30.140.91270.0701
", + "bbox": [ + 120, + 369, + 459, + 494 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "real-world dataset. The Deblur-NeRF synthetic dataset consists of 5 scenes, each containing defocused images generated using the built-in function of Blender [5]. The Deblur-NeRF real-world dataset includes 10 scenes, with images captured by enlarging the aperture on a Canon EOS RP. The DoF-NeRF real-world dataset comprises 7 scenes, each containing 20 to 30 image triplets. Each triplet includes an all-in-focus image, along with two defocused images focused on the background and foreground. Since DoF-NeRF does not provide training code, we use half of the training defocused images with a foreground focus and the other half with a background focus. All camera poses and initial point clouds are obtained through COLMAP [29, 30].", + "bbox": [ + 88, + 518, + 482, + 715 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Rendering Results", + "text_level": 1, + "bbox": [ + 89, + 724, + 269, + 742 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Quantitative Results. We quantitatively evaluate our CoCoGaussian using the following three metrics: peak signal-to-noise ratio (PSNR), structural similarity index measure (SSIM) [38], and learned perceptual image patch similarity (LPIPS) [45]. We compare our approach to several state-of-the-art methods, including both raytracing [11, 18, 25] and rasterization-based methods [10, 26]. Quantitative results are shown in Tab. 1 and Tab. 2, where our approach achieves the best performance across all metrics except for PSNR on the Deblur-NeRF real-world", + "bbox": [ + 89, + 750, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "dataset. Note that we report the reproduced performance scores for Deblurring 3DGS [10] and BAGS [26] using their official codes. The Deblur-NeRF real-world dataset inherently presents challenges, as there are illumination differences between the sharp and defocused images, leading to relatively lower PSNR score. We discuss this issue in the appendix. For the DoF-NeRF dataset, CoCoGaussian achieves outstanding performance across all metrics, with an approximately 0.04 reduction in LPIPS compared to BAGS. Since the DoF-NeRF dataset has a higher resolution than the Deblur-NeRF dataset, this performance drop in BAGS is likely due to its requirement for manual adjustment of blurring kernel size, highlighting an intrinsic limitation. Conversely, Deblurring 3DGS shows minimal improvement over 3DGS, which can also be attributed to its limited kernel size relative to the higher resolution. The details regarding the time complexity of our model are provided in the appendix.", + "bbox": [ + 511, + 90, + 906, + 364 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Qualitative Results. For qualitative evaluation, we present the rendering results of 3DGS [9], Deblurring 3DGS [10], and BAGS [26] in Fig. 3. Our method demonstrates higher-fidelity results compared to others. Specifically, in the fourth row of our results, the white line on the right side is rendered more sharply, and in the fifth row, the detailed quality of the graphic card appears superior to that of BAGS. Additionally, more CoC visualizations similar to Fig. 1 (b), are in the appendix.", + "bbox": [ + 511, + 383, + 908, + 521 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Ablation Study", + "text_level": 1, + "bbox": [ + 511, + 531, + 666, + 547 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To validate the effectiveness of our proposed methods, we perform a series of ablation studies, with all experiments conducted on the DoF-NeRF real-world dataset. We chose this dataset as it is based on real captured images and does not present the illumination issues found in the Deblur-NeRF real-world dataset, ensuring more reliable evaluation. The results are summarized in Tab. 3, where the baseline denotes naive 3DGS [9]. Additional ablative experiments with other factors are in the appendix.", + "bbox": [ + 511, + 555, + 906, + 691 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Circle of Confusion. To demonstrate the impact of our core concept, CoC, we conduct an experiment that excludes the CoC from the model, retaining only $\\beta_{m}$ and $\\mathbf{d}_m$ . This configuration yields significantly lower performance than the full model, indicating an over-reliance on parametric learning. The result suggests that without the CoC, which leverages depth and aperture information, the means of the generated Gaussian sets are not optimally arranged, leading to overfitting.", + "bbox": [ + 511, + 712, + 906, + 849 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "CoC Direction Vector. In this experiment, we retain the CoC concept but exclude the learnable direction vectors", + "bbox": [ + 511, + 869, + 906, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "16106", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/9cb75fc5bbc895292c428596b4c164354427d742a8147dbd1ed06ad66a7319a5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 93, + 89, + 243, + 167 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8b6a4f88c0374675cac50e3bfca5115bdccef49500438b090836a1550bb2eb6f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 245, + 90, + 372, + 167 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/65dc56d33dd4b64ad77759b39f75049b420ea498977ce01edf43bfaf674f05d3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 375, + 90, + 506, + 167 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e1b3d16adc142ec22e3f5ee7c54d10479dc38d17952d9bad8c813afc3a51beff.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 506, + 90, + 638, + 167 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6794c98aa1a6a9a82c103519bcad327ab99e081d3a317e9d74de259022f06f00.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 640, + 90, + 771, + 167 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/557b78154bcfc77d4c2a283206ef51104a693a1d2b25ae2086605bffda617f57.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 772, + 90, + 903, + 167 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ab9e9b898833f7d861a78f8560d9f092cfc5341b88af731176724cfccc0b4a35.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 94, + 170, + 243, + 247 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7d7ef3d8dfd972cd5742c3257e390f318020b206d3f3aff185abd7c86ef54233.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 245, + 170, + 374, + 247 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/44b72f0dab6c517d5c47e13e262da96138a7795bc3c7cc1a9d7418fbc9a07164.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 375, + 170, + 504, + 247 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ab67a5fa763d68ad2c7c96a4587da0f5e7e883034440bf36dae723c1c18aec78.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 506, + 170, + 638, + 247 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/894068c4173df7c1af6ada250a64b63b4fa2855e097b2db1edc5e5e4238440b1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 640, + 170, + 771, + 247 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/053df0d43e802824196a3ed7271d0914b67c30c735b137562cf7724b6f796428.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 772, + 170, + 903, + 247 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c223c05e7a7216926ee1a9c3526b5daa005c1677ee8c7f6c02bb0c541755c8b9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 94, + 250, + 243, + 327 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f95eb3e7ef4c3776bd9fa2afc9644accf8594658601b64a88cb385bb05f5128a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 245, + 250, + 374, + 327 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7a8c55cdd317cc31b6f5ad765767179cd8c2c38570de47482ffe16983c80338a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 375, + 250, + 504, + 327 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/06777173a890b5b4c65c60d3ec37a450cd2a29cb3472edac5a169bf9b801aeab.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 506, + 250, + 638, + 327 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7e568ed2bfd40d10661d49024ae43969507bd25150670c5b66977c4e6bfaef0e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 640, + 250, + 771, + 327 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2a9b2ed0bc8edafdb6fe745809757e19ad83d513b8268a1428faf410426bcd46.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 772, + 250, + 903, + 327 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/54afc9bc38c4814c92f333c69f4d1431b9e6437de2d3cde6a4048556412c4b51.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 94, + 329, + 243, + 407 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2344fb31ed42130a2ecb419b1f16dea64c5648d8f37916789bbc87453b0d106a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 245, + 329, + 374, + 407 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a01d3ec93c34a08ce1eb4ab47d979f1121c887b80d99ffc0c79f148c90c03fa8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 375, + 329, + 504, + 407 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/fe7f4007f721a1d0a96d29153f7b1fa20da7e65dc14ac4a4a48aceca82da3f2a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 506, + 329, + 638, + 407 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7e956a745bfe1aac5440edbf2cb366c122b1995d2175feb2a6235e8e65d58d0a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 640, + 329, + 771, + 407 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a42e95ff77ea49b01929a4d24b052e1d4a1fef45e4d95d21016c9c99ed1cf938.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 772, + 329, + 903, + 407 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/38339c5098d520d020d8df9c6030e4692d2574151d39297acdfe39fd3e6f46eb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 94, + 410, + 243, + 488 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0c0ee419b5ed6fa92e8b5f1f375840694022385dc58916864d4dedb7458178a5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 245, + 410, + 374, + 488 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3eec4c0f8400d57ddd6011d77355fbebed540f77af64564e9b1b805d41aea3be.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 375, + 410, + 504, + 488 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/11b31f0ca21e8765fb539b0dbff3f65261b5764445b1fdf3c952e7ee7158d901.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 506, + 410, + 638, + 488 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a7f313060dbd9e68038eb922a9396478fdf13218bff2f8c1aae4e4446b14eb8a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 640, + 410, + 771, + 488 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e6aeafa7bf2c79d8edb837a2f55eac89efd0e259f1a5eeb26a83c04041d8161d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 772, + 410, + 903, + 488 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/365029911a67e00921e80b9a067c20f4b7242d200cc03b355e07ee4717cb7fe6.jpg", + "image_caption": [ + "Reference" + ], + "image_footnote": [], + "bbox": [ + 94, + 489, + 241, + 568 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0fc3a7f0655d52d3b1befad0954f9806fb9a6e5aa900c78880a896a8269d5057.jpg", + "image_caption": [ + "3DGS" + ], + "image_footnote": [], + "bbox": [ + 243, + 489, + 374, + 568 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/56be89c78cd25de733c32f6ec3046914c77e49f580c82ccb39226f16a8f75ce5.jpg", + "image_caption": [ + "Deblurring 3DGS" + ], + "image_footnote": [], + "bbox": [ + 375, + 489, + 506, + 568 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ba88978bbb98bc401526f99df6b474572ccdb06b74a87ebc9f255b229c7bf6f1.jpg", + "image_caption": [ + "BAGS" + ], + "image_footnote": [], + "bbox": [ + 506, + 489, + 638, + 568 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5b0b786477b09fff51e80bdbf6bf408e683240f8ceccb35c26320049e774cda9.jpg", + "image_caption": [ + "Ours", + "Figure 3. Qualitative comparison on the Deblur-NeRF and DoF-NeRF datasets." + ], + "image_footnote": [], + "bbox": [ + 640, + 489, + 771, + 568 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/015d01065fff52da18631ddc2ed73d69933657e4eaa95ba3470a89cfffef32c7.jpg", + "image_caption": [ + "Ground Truth" + ], + "image_footnote": [], + "bbox": [ + 772, + 489, + 903, + 568 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "$\\mathbf{d}_m$ . Instead of using learnable vectors, we use fixed direction vectors, arranging $M$ CoC Gaussians evenly in a circular pattern on a plane perpendicular to the vector between the camera position and the base Gaussian. This configuration slightly underperforms compared to the full model, as the fixed directions restrict optimal Gaussian position, preventing adaptation to the most reliable configuration.", + "bbox": [ + 89, + 622, + 483, + 729 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "CoC Scaling Factor. In this setup, we include the learnable direction vector while omitting the CoC scaling factor $\\beta_{m}$ . Although the performance is slightly better than when the direction vector is excluded, it remains lower than that of the full model. Without the CoC scaling factor, CoC Gaussian generation becomes overly dependent on depth, leading to errors in CoC diameter calculation when depth values are inaccurate. Including the CoC scaling factor enables the full model to handle uncertain depth more robustly, improving performance. Related visualization re", + "bbox": [ + 89, + 750, + 483, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "sults are in the appendix.", + "bbox": [ + 513, + 623, + 684, + 638 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Aperture Parameter. The aperture parameter $K$ includes the aperture diameter and focal length, and it plays a critical role in determining the size of the CoC. Without this parameter, there is an upper limit to the CoC diameter calculation, making accurate computation challenging. Experimental results indicate that although performance is somewhat lower than the full model, it is improved over the baseline. This suggests that while the model is unable to determine precise CoC size without $K$ , it still learns the focus plane and captures the in-focus region accurately.", + "bbox": [ + 511, + 656, + 906, + 808 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Experiments on All-in-Focus Images.", + "text_level": 1, + "bbox": [ + 511, + 816, + 834, + 834 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We conduct novel view synthesis experiments using all-in-focus images to verify the generalization performance of CoCoGaussian. We utilize the NeRF-LLFF [19, 20] dataset, which contains real-world images. As shown in Tab. 4, our", + "bbox": [ + 511, + 839, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "16107", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/40134f8f7422f6029deea3284d5dfc03cf87ca4d59036fc79d378a9ed595f953.jpg", + "image_caption": [ + "Figure 4. Visualization of Aperture parameter and Focus Plane Customization. The top row of images decreases the aperture parameter $K$ from left to right, while the bottom row moves the focus plane $d_{F}$ further from the camera from left to right." + ], + "image_footnote": [], + "bbox": [ + 94, + 89, + 906, + 277 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/9c0d328d9e45c270441b5b1c99adcbd1befc081ed89e7bdfce11116cfeb8657e.jpg", + "table_caption": [ + "Table 3. Ablation Results of CoCoGaussian." + ], + "table_footnote": [], + "table_body": "
MethodsPSNR↑SSIM↑LPIPS↓
Baseline [9]25.720.82910.1817
w/o CoC28.290.87780.0927
w/o CoC direction vector28.910.88930.0896
w/o CoC scaling29.460.90580.0793
w/o aperture parameter27.370.85100.1113
Ours Full30.140.91270.0701
", + "bbox": [ + 98, + 340, + 478, + 454 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "approach outperforms the baseline 3DGS [9] across all metrics. In practice, when the aperture is very small, radiance from a point on the subject forms a tiny CoC on the image sensor after passing through the aperture, even if the subject is slightly away from the focus plane. Because the CoC is so small, the resulting image is not perceived as defocused. By modeling this small circle, which is difficult to detect with the human eye, our approach achieves higher performance than the baseline 3DGS. Related CoC visualizations are in the appendix.", + "bbox": [ + 89, + 460, + 483, + 612 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.4.Customizable Depth of Field and Focus Plane", + "text_level": 1, + "bbox": [ + 89, + 619, + 475, + 636 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Customizing the depth of field and focus settings is essential in various applications, such as augmented reality (AR) and virtual reality (VR), where control over depth cues and focus effects can greatly enhance visual realism and viewer engagement. Since our approach incorporates the aperture parameter $K$ in CoC modeling, it enables customization of both CoC size and depth of field by adjusting this parameter. The top row in Fig. 4 visualizes increasing depth of field by gradually reducing $K$ from left to right. As the focus plane is positioned at a shallow depth, the model consistently captures focus in this region. However, in the leftmost image, the large aperture causes the CoC size to increase at deeper depths, resulting in defocus blur.", + "bbox": [ + 89, + 643, + 482, + 839 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Additionally, our model can customize the focus plane of defocused images, as shown in the bottom row of Fig. 4, where the focus plane shifts from shallow depth on the left to deeper depth on the right. The shallow regions are in fo", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/e5843e11b0e48c2007e56aa31c41a1adc97c12f019b21bb9ea01321cb137141b.jpg", + "table_caption": [ + "Table 4. Results on NeRF-LLFF [19, 20] Dataset." + ], + "table_footnote": [], + "table_body": "
MethodsPSNR↑SSIM↑LPIPS↓
3DGS [9]27.100.86190.0569
Ours27.760.87380.0372
", + "bbox": [ + 575, + 340, + 841, + 388 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "cus on the leftmost image, while deeper areas remain out of focus, and the opposite holds for the rightmost image. This level of control over focus and depth effects is essential for achieving realistic depth cues. Therefore, CoCoGaussian not only demonstrates superior capability in synthesizing sharp novel view images but also provides enhanced flexibility in adjusting aperture and focus plane settings.", + "bbox": [ + 511, + 398, + 906, + 505 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Limitation and Future Work", + "text_level": 1, + "bbox": [ + 511, + 518, + 777, + 534 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "While our model demonstrates strong qualitative and quantitative performance, there is potential for further optimization in the adaptive CoC scaling factor. To ensure training stability, we currently limit the factor size, enabling adaptive optimization mainly when the CoC diameter is overestimated. However, future improvements could enhance adaptability to handle cases where the CoC diameter is underestimated, further improving depth handling in complex scenes without compromising stability.", + "bbox": [ + 511, + 542, + 906, + 680 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 694, + 633, + 709 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we propose CoCoGaussian, a 3D scene representation method that effectively leverages photographic defocus principles to handle defocused image inputs. CoCoGaussian accurately models defocus blur by constructing the CoC through 3D Gaussians and learnable aperture parameters. We introduce an adaptive CoC Gaussian generation method to address the challenges posed by uncertain depth, making our model robust to reflective or refractive objects. Our approach demonstrates the feasibility and effectiveness of representing 3D scenes from defocused images, advancing possibilities for applications in real-world image synthesis.", + "bbox": [ + 511, + 719, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "16108", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements. This project was supported by the Naver Cloud Corporation and the National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIT) (No. RS-2024-00340745).", + "bbox": [ + 89, + 90, + 485, + 151 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 165, + 187, + 181 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5855–5864, 2021. 3", + "[2] Ayan Chakrabarti. A neural approach to blind motion deblurring. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part III 14, pages 221-235. Springer, 2016. 2, 3", + "[3] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In European Conference on Computer Vision, pages 333-350. Springer, 2022. 2, 3", + "[4] Wenbo Chen and Ligang Liu. Deblur-gs: 3d gaussian splattering from camera motion blurred images. Proceedings of the ACM on Computer Graphics and Interactive Techniques, 7 (1):1-15, 2024. 3", + "[5] Blender Online Community. Blender - a 3D modelling and rendering package. Blender Foundation, Stichting Blender Foundation, Amsterdam, 2018. 6", + "[6] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5501–5510, 2022. 2", + "[7] Eugene Hecht. Optics. Pearson Education India, 2012. 2, 4", + "[8] Wei Jiang, Kwang Moo Yi, Golnoosh Samei, Oncel Tuzel, and Anurag Ranjan. Neuman: Neural human radiance field from a single video. In European Conference on Computer Vision, pages 402-418. Springer, 2022. 3", + "[9] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42 (4):1-14, 2023. 2, 3, 6, 8", + "[10] Byeonghyeon Lee, Howoong Lee, Xiangyu Sun, Usman Ali, and Eunbyung Park. Deblurring 3d gaussian splatting. arXiv preprint arXiv:2401.00834, 2024. 2, 3, 4, 5, 6", + "[11] Dogyoon Lee, Minhyeok Lee, Chajin Shin, and Sangyoun Lee. Dp-nerf: Deblurred neural radiance field with physical scene priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12386–12396, 2023. 3, 6", + "[12] Junghe Lee, Donghyeong Kim, Dogyoon Lee, Suhwan Cho, and Sangyoun Lee. Crime: Continuous rigid motion-aware gaussian splatting from motion blur images. arXiv preprint arXiv:2407.03923, 2024. 4, 5", + "[13] Jungho Lee, Dogyoon Lee, Minhyeok Lee, Donghyung Kim, and Sangyoun Lee. Smurf: Continuous dynam-" + ], + "bbox": [ + 93, + 191, + 480, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ics for motion-deblurring radiance fields. arXiv preprint arXiv:2403.07547, 2024. 2, 3", + "[14] Tianye Li, Mira Slavcheva, Michael Zollhoefer, Simon Green, Christoph Lassner, Changil Kim, Tanner Schmidt, Steven Lovegrove, Michael Goesele, Richard Newcombe, et al. Neural 3d video synthesis from multi-view video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5521-5531, 2022. 3", + "[15] Zhengqi Li, Simon Niklaus, Noah Snavely, and Oliver Wang. Neural scene flow fields for space-time view synthesis of dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6498-6508, 2021. 3", + "[16] Zhaoshuo Li, Thomas Müller, Alex Evans, Russell H Taylor, Mathias Unberath, Ming-Yu Liu, and Chen-Hsuan Lin. Neuralangelo: High-fidelity neural surface reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8456-8465, 2023. 3", + "[17] Kevin M Lynch and Frank C Park. Modern robotics. Cambridge University Press, 2017. 3", + "[18] Li Ma, Xiaoyu Li, Jing Liao, Qi Zhang, Xuan Wang, Jue Wang, and Pedro V Sander. Deblur-nerf: Neural radiance fields from blurry images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12861–12870, 2022. 2, 3, 5, 6", + "[19] Ben Mildenhall, Pratul P Srinivasan, Rodrigo Ortiz-Cayon, Nima Khademi Kalantari, Ravi Ramamoorthi, Ren Ng, and Abhishek Kar. Local light field fusion: Practical view synthesis with prescriptive sampling guidelines. ACM Transactions on Graphics (ToG), 38(4):1-14, 2019. 7, 8", + "[20] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I, pages 405-421, 2020. 1, 2, 3, 6, 7, 8", + "[21] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG), 41(4):1-15, 2022. 2", + "[22] Michael Niemeyer, Jonathan T Barron, Ben Mildenhall, Mehdi SM Sajjadi, Andreas Geiger, and Noha Radwan. Regnerf: Regularizing neural radiance fields for view synthesis from sparse inputs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5480-5490, 2022. 3", + "[23] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5865-5874, 2021. 3", + "[24] Keunhong Park, Utkarsh Sinha, Peter Hedman, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Ricardo MartinBrualla, and Steven M Seitz. Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields. arXiv preprint arXiv:2106.13228, 2021. 3", + "[25] Cheng Peng and Rama Chellappa. Pdrf: progressively deblurring radiance field for fast scene reconstruction from" + ], + "bbox": [ + 516, + 93, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "16109", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "blurry images. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2029-2037, 2023. 2, 3, 6", + "[26] Cheng Peng, Yutao Tang, Yifan Zhou, Nengyu Wang, Xijun Liu, Deming Li, and Rama Chellappa. Bags: Blur agnostic gaussian splatting through multi-scale kernel modeling. arXiv preprint arXiv:2403.04926, 2024. 2, 3, 4, 5, 6", + "[27] Sida Peng, Junting Dong, Qianqian Wang, Shangzhan Zhang, Qing Shuai, Xiaowei Zhou, and Hujun Bao. Animatable neural radiance fields for modeling dynamic human bodies. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14314-14323, 2021. 3", + "[28] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-nerf: Neural radiance fields for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10318–10327, 2021. 3", + "[29] Johannes L Schonberger, Enliang Zheng, Jan-Michael Frahm, and Marc Pollefeys. Pixelwise view selection for unstructured multi-view stereo. In European conference on computer vision, pages 501-518. Springer, 2016. 3, 6", + "[30] Qi Shan, Jiaya Jia, and Aseem Agarwala. High-quality motion deblurring from a single image. Acm transactions on graphics (tog), 27(3):1-10, 2008. 3, 6", + "[31] Pratul P Srinivasan, Ren Ng, and Ravi Ramamoorthi. Light field blind motion deblurring. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3958-3966, 2017. 2, 3", + "[32] Jiaming Sun, Yiming Xie, Linghao Chen, Xiaowei Zhou, and Hujun Bao. Neuralrecon: Real-time coherent 3d reconstruction from monocular video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15598-15607, 2021. 3", + "[33] Edgar Tretschk, Ayush Tewari, Vladislav Golyanik, Michael Zollhöfer, Christoph Lassner, and Christian Theobalt. Non-rigid neural radiance fields: Reconstruction and novel view synthesis of a dynamic scene from monocular video. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12959-12970, 2021. 3", + "[34] Guangcong Wang, Zhaoxi Chen, Chen Change Loy, and Zwei Liu. Sparsenerf: Distilling depth ranking for few-shot novel view synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9065-9076, 2023. 3", + "[35] Jiepeng Wang, Peng Wang, Xiaoxiao Long, Christian Theobalt, Taku Komura, Lingjie Liu, and Wenping Wang. Neuris: Neural reconstruction of indoor scenes using normal priors. In European Conference on Computer Vision, pages 139-155. Springer, 2022. 3", + "[36] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. arXiv preprint arXiv:2106.10689, 2021. 3", + "[37] Peng Wang, Lingzhe Zhao, Ruijie Ma, and Peidong Liu. Bad-nerf: Bundle adjusted deblur neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4170–4179, 2023. 2, 3" + ], + "bbox": [ + 91, + 90, + 482, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[38] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 6", + "[39] Chung-Yi Weng, Brian Curless, Pratul P Srinivasan, Jonathan T Barron, and Ira Kemelmacher-Shlizerman. Humanerf: Free-viewpoint rendering of moving people from monocular video. In Proceedings of the IEEE/CVF conference on computer vision and pattern Recognition, pages 16210-16220, 2022. 3", + "[40] Oliver Whyte, Josef Sivic, Andrew Zisserman, and Jean Ponce. Non-uniform deblurring for shaken images. International journal of computer vision, 98:168-186, 2012. 2, 3", + "[41] Zijin Wu, Xingyi Li, Juewen Peng, Hao Lu, Zhiguo Cao, and Weicai Zhong. Dof-nerf: Depth-of-field meets neural radiance fields. In Proceedings of the 30th ACM International Conference on Multimedia, pages 1718–1729, 2022. 2, 3, 5, 6", + "[42] Jamie Wynn and Daniyar Turmukhambetov. Diffusionerf: Regularizing neural radiance fields with denoising diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4180-4189, 2023. 3", + "[43] Jiawei Yang, Marco Pavone, and Yue Wang. Freenerf: Improving few-shot neural rendering with free frequency regularization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8254-8263, 2023. 3", + "[44] Wang Yifan, Felice Serena, Shihao Wu, Cengiz Özireli, and Olga Sorkine-Hornung. Differentiable surface splatting for point-based geometry processing. ACM Transactions on Graphics (TOG), 38(6):1-14, 2019. 3", + "[45] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 6", + "[46] Lingzhe Zhao, Peng Wang, and Peidong Liu. Bad-gaussians: Bundle adjusted deblur gaussian splatting. arXiv preprint arXiv:2403.11831, 2024. 3" + ], + "bbox": [ + 516, + 90, + 903, + 667 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "16110", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/CoCoGaussian_ Leveraging Circle of Confusion for Gaussian Splatting from Defocused Images/c9f7dcb9-6d18-430a-9426-253bcaf437ef_model.json b/2025/CoCoGaussian_ Leveraging Circle of Confusion for Gaussian Splatting from Defocused Images/c9f7dcb9-6d18-430a-9426-253bcaf437ef_model.json new file mode 100644 index 0000000000000000000000000000000000000000..4f70a09aca4986a9cab61a74a32550d93b6de7e3 --- /dev/null +++ b/2025/CoCoGaussian_ Leveraging Circle of Confusion for Gaussian Splatting from Defocused Images/c9f7dcb9-6d18-430a-9426-253bcaf437ef_model.json @@ -0,0 +1,2563 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.24, + 0.131, + 0.761, + 0.177 + ], + "angle": 0, + "content": "CoCoGaussian: Leveraging Circle of Confusion for Gaussian Splatting from Defocused Images" + }, + { + "type": "text", + "bbox": [ + 0.197, + 0.196, + 0.8, + 0.233 + ], + "angle": 0, + "content": "Jungho Lee\\(^{1*}\\) Suhwan Cho\\(^{1}\\) Taeoh Kim\\(^{2}\\) Ho-Deok Jang\\(^{2}\\) Minhyeok Lee\\(^{1}\\) Geonho Cha\\(^{2}\\) Dongyoon Wee\\(^{2}\\) Dogyoon Lee\\(^{1}\\) Sangyoun Lee\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.184, + 0.239, + 0.816, + 0.275 + ], + "angle": 0, + "content": "\\(^{1}\\)School of Electrical and Electronic Engineering, Yonsei University \\(^{2}\\)NAVER Cloud https://Jho-Yonsei.github.io/CoCoGaussian" + }, + { + "type": "image", + "bbox": [ + 0.13, + 0.285, + 0.355, + 0.524 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.138, + 0.527, + 0.309, + 0.54 + ], + "angle": 0, + "content": "(a) Defocused Image Acquisition" + }, + { + "type": "image", + "bbox": [ + 0.373, + 0.287, + 0.593, + 0.401 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.375, + 0.412, + 0.592, + 0.523 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.394, + 0.527, + 0.572, + 0.539 + ], + "angle": 0, + "content": "(b) Reconstructed Gaussian Points" + }, + { + "type": "image", + "bbox": [ + 0.603, + 0.324, + 0.635, + 0.365 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.644, + 0.289, + 0.868, + 0.406 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.644, + 0.411, + 0.867, + 0.525 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.703, + 0.527, + 0.812, + 0.54 + ], + "angle": 0, + "content": "(c) Rendered Images" + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.544, + 0.907, + 0.629 + ], + "angle": 0, + "content": "Figure 1. We propose CoCoGaussian, a novel framework for 3D scene reconstruction from defocused images. (a) With a large aperture, radiance from the focus plane appears as a small circle of confusion (green) on the image sensor, while radiance from greater depths results in larger circles (pink). (b) We visualize reconstructed Gaussians of CoCoGaussian, where the black dots indicate Gaussian means, while the gray dots represent Gaussian means forming the circle of confusion, which decreases in size near the focus plane. (c) The defocused image focuses on shallow depths, with circle size increasing with depth, as shown in (b). CoCoGaussian allows customization of defocused images by adjusting the depth of field or focus plane, while sharp images can be obtained by rendering without the circle of confusion." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.64, + 0.327, + 0.656 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.674, + 0.486, + 0.871 + ], + "angle": 0, + "content": "3D Gaussian Splatting (3DGS) has attracted significant attention for its high-quality novel view rendering, inspiring research to address real-world challenges. While conventional methods depend on sharp images for accurate scene reconstruction, real-world scenarios are often affected by defocus blur due to finite depth of field, making it essential to account for realistic 3D scene representation. In this study, we propose CoCoGaussian, a Circle of Confusion-aware Gaussian Splatting that enables precise 3D scene representation using only defocused images. CoCoGaussian addresses the challenge of defocus blur by modeling the Circle of Confusion (CoC) through a physically grounded approach based on the principles of photo" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.642, + 0.909, + 0.78 + ], + "angle": 0, + "content": "graphic defocus. Exploiting 3D Gaussians, we compute the CoC diameter from depth and learnable aperture information, generating multiple Gaussians to precisely capture the CoC shape. Furthermore, we introduce a learnable scaling factor to enhance robustness and provide more flexibility in handling unreliable depth in scenes with reflective or refractive surfaces. Experiments on both synthetic and real-world datasets demonstrate that CoCoGaussian achieves state-of-the-art performance across multiple benchmarks." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.813, + 0.646, + 0.829 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.909, + 0.903 + ], + "angle": 0, + "content": "The emergence of the Neural Radiance Field (NeRF) [20] has brought significant attention to 3D scene representation for photo-realistic novel view synthesis. They reconstruct 3D scenes from images captured from multiple views" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.887, + 0.432, + 0.902 + ], + "angle": 0, + "content": "*This work was done during an internship at Naver Cloud." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "16101" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.243 + ], + "angle": 0, + "content": "and render images of unseen views. However, since NeRF models 3D scenes as implicit neural representations through a ray tracing-based approach, they suffer from inefficient memory usage. In contrast, a rasterization-based method, 3D Gaussian Splatting (3DGS) [9] explicitly models 3D scenes using independent 3D Gaussians with different 3D means, covariances, opacities, and spherical harmonic coefficients. The tile-based rasterizer of 3DGS applies alpha blending to Gaussian splats sorted by visibility order, enabling fast training, rendering, and efficient memory usage." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.246, + 0.485, + 0.638 + ], + "angle": 0, + "content": "Both NeRF [20] and 3DGS [9] rely on sharp images to represent 3D scenes accurately, which is a highly ideal assumption. In real-world settings, various factors (e.g., finite depth of field and camera motion blur) can hinder the capture of sharp images, leading to image degradation. Achieving a perfectly sharp image requires every part of the scene to be in focus, namely all-in-focus, which calls for a large depth of field. However, obtaining a large depth of field necessitates using a small aperture, which in turn requires a longer exposure time to allow enough light to enter the camera. During this extended exposure, even slight movements of a handheld camera can introduce motion blur. To prevent motion blur, the aperture needs to be widened, but this comes at the cost of a shallower depth of field. As a result, areas outside the focus plane appear blurred due to defocus. Defocus blur is closely related to the concept of the Circle of Confusion (CoC); as shown in Fig. 1, when a subject is positioned away from the focus plane, the radiances from the subject pass through the aperture and are mapped onto the image sensor in a circular pattern, resulting in shallow depth of field and defocus blur [7]. However, despite these challenges, recent research in 3D scene representation has rarely focused on addressing defocus blur directly. In this paper, we propose a method to represent 3D scenes using only images with defocus blur, tackling the CoC-related challenges posed by real-world image acquisition scenarios." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.642, + 0.483, + 0.868 + ], + "angle": 0, + "content": "Recently, there has been increasing interest in reconstructing 3D scenes using only degraded images, with a focus on addressing challenges caused by camera motion blur and defocus blur. Deblur-NeRF [18] is the first study to take on this task, adopting a blind image deblurring approaches [2, 31, 40]. Follow-up studies [10-13, 25, 26, 37] have expanded this approach beyond raytracing to rasterization-based methods. However, many of these methods depend heavily on learning-based strategies and fail to incorporate photographic principles that accurately capture how defocus occurs in real-world scenarios. While DoF-NeRF [41] adopts the concept of CoC to model defocus blur in 3D scene representations, it relies on an implicit representation, which leads to slow rendering speeds, making it less suitable for real-time applications." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.483, + 0.902 + ], + "angle": 0, + "content": "In this paper, we propose CoCoGaussian, which leverages a physically grounded photographic defocus princi" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.409 + ], + "angle": 0, + "content": "plies to represent 3D scenes. Our method incorporates an aperture to calculate the CoC diameter based on depth and learnable aperture information using 3D Gaussians, accurately modeling shallow depth of field images. Using the CoC diameter and the 3D Gaussian, we generate multiple Gaussians to form the CoC shape, as illustrated in Fig. 1. However, when objects with refraction or reflection exist in the scene, the depth obtained through the Gaussians may become unreliable. To address this, we propose a method to keep the set of generated Gaussians within the CoC radius, introducing a learnable scaling factor to increase flexibility and reduce dependence on estimated depth in the CoC modeling. As a result, our model combines physically grounded defocus principles with an adaptive CoC modeling that effectively handles uncertain depths, allowing it to reconstruct sharp 3D scenes using only defocused images. Additionally, our approach has an advantage of learning aperture and focus plane information, enabling dynamic control over depth of field and flexible adjustment of the focus plane. This feature allows for highly customizable scene visualization, adapting to various focus and depth requirements." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.41, + 0.905, + 0.469 + ], + "angle": 0, + "content": "To demonstrate the effectiveness of our model, we conduct extensive experiments on the benchmarks: the Deblur-NeRF [18] dataset and the DoF-NeRF [41] real-world dataset. Our contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.476, + 0.904, + 0.519 + ], + "angle": 0, + "content": "- CoCoGaussian models the CoC at the 3D Gaussian level, reconstructing the precise 3D scene and enabling sharp novel view synthesis from defocused images." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.524, + 0.904, + 0.552 + ], + "angle": 0, + "content": "- We propose an adaptive learning approach that robustly models the CoC even with unreliable depth information." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.558, + 0.904, + 0.601 + ], + "angle": 0, + "content": "- CoCoGaussian enables the customizable depth of field and flexible focus plane adjustment during rendering by learning aperture and focus plane information." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.607, + 0.903, + 0.635 + ], + "angle": 0, + "content": "- CoCoGaussian achieves state-of-the-art performance, both quantitatively and qualitatively." + }, + { + "type": "list", + "bbox": [ + 0.512, + 0.476, + 0.904, + 0.635 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.654, + 0.655, + 0.67 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.68, + 0.905, + 0.696 + ], + "angle": 0, + "content": "2.1. Scene Representations for Novel View Synthesis" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.905, + 0.901 + ], + "angle": 0, + "content": "3D scene representation for novel view synthesis has advanced significantly with the advent of neural radiance fields (NeRF) [20], a ray tracing-based volumetric rendering method that generates photo-realistic novel view images from multi-view images. However, as an implicit representation method using deep MLPs, NeRF suffers from slow training and rendering speeds. To address this limitation, explicit representation methods such as Plenoxel [6], TensoRF [3], and Instant-NGP [21] have been introduced. More recently, 3D Gaussian Splatting (3DGS) [9], a rasterization-based method, has emerged as an alternative to ray tracing, significantly mitigating the speed and memory efficiency constraints of ray tracing" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16102" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.213 + ], + "angle": 0, + "content": "based approaches. NeRF and 3DGS has enabled a wide range of related research, including dynamic scene representation [14, 15, 23, 24, 28, 33], human avatars [8, 27, 39], 3D mesh reconstruction [16, 32, 35, 36], 3D scene representation from sparse-view images [22, 34, 42, 43], and 3D scene representation from blurry images [4, 10-13, 18, 25, 26, 37, 46]. In this paper, we propose a method for 3D scene representation from defocused images by exploiting 3DGS." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.224, + 0.449, + 0.24 + ], + "angle": 0, + "content": "2.2. Novel View Synthesis from Blurry Images" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.248, + 0.485, + 0.641 + ], + "angle": 0, + "content": "Scene representation methods like NeRF [20] and 3DGS [9] require sharp images as input to render photo-realistic images. However, in real-world scenarios, capturing sharp images is challenging, often suffering from image degradation such as camera motion blur and defocus blur. To address this issue, Deblur-NeRF [18] introduced a 3D ray-based blurring kernel inspired by image blind deblurring [2, 31, 40], which intentionally generates blurry images during training. After training, it renders sharp novel-view images by excluding the trained kernel. Following Deblur-NeRF, various approaches have been proposed to tackle the degradation issue. DP-NeRF [11] introduces a kernel based on the rigid body transformation [17] with prior knowledge that object shape remains consistent in static scenes, and PDRF [25] proposes a blur estimation method with 2-stage efficient rendering. Recently, methods utilizing 3DGS have been proposed to enable faster rendering. For instance, Deblurring 3DGS [10] adjusts Gaussian parameters to generate blurry images during training, and BAGS [26] proposes a blur-agnostic kernel and a blur mask using convolutional neural networks (CNNs). DoF-NeRF [41] has explored modeling defocus blur by leveraging the circle of confusion (CoC) within an implicit neural representation. In this paper, we propose a method that also leverages CoC to model defocus, which enables real-time rendering, while still maintaining high-quality depth-of-field effects." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.656, + 0.218, + 0.673 + ], + "angle": 0, + "content": "3. Preliminary" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.681, + 0.301, + 0.698 + ], + "angle": 0, + "content": "3.1. 3D Gaussian Splatting" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.484, + 0.781 + ], + "angle": 0, + "content": "Unlike ray tracing-based methods [1, 3, 20], 3DGS [9] is built on a rasterization-based approach with differentiable 3D Gaussians. These 3D Gaussians are initialized from a sparse point cloud obtained via a Structure-from-Motion (SfM) [29, 30] algorithm and are defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.185, + 0.792, + 0.483, + 0.813 + ], + "angle": 0, + "content": "\\[\nG (\\mathbf {x}) = e ^ {- \\frac {1}{2} (\\mathbf {x} - \\mu) ^ {\\top} \\boldsymbol {\\Sigma} ^ {- 1} (\\mathbf {x} - \\mu)}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.484, + 0.901 + ], + "angle": 0, + "content": "where \\(\\mathbf{x} \\in \\mathbb{R}^3\\) is a point on the Gaussian \\(G\\) centered at the mean vector \\(\\mu \\in \\mathbb{R}^3\\) with an covariance matrix \\(\\boldsymbol{\\Sigma} \\in \\mathbb{R}^{3 \\times 3}\\). The 3D covariance matrix \\(\\boldsymbol{\\Sigma}\\) is derived from a learnable scaling vector \\(\\mathbf{s} \\in \\mathbb{R}^3\\) and rotation quaternion \\(\\mathbf{q} \\in \\mathbb{R}^4\\), from which the scaling matrix \\(\\mathbf{S} \\in \\mathbb{R}^{3 \\times 3}\\) and rotation matrix \\(\\mathbf{R} \\in\\)" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.091, + 0.822, + 0.107 + ], + "angle": 0, + "content": "\\(\\mathbb{R}^{3\\times 3}\\) are obtained and represented as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.652, + 0.112, + 0.905, + 0.129 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\Sigma} = \\mathbf {R} \\mathbf {S} \\mathbf {S} ^ {\\top} \\mathbf {R} ^ {\\top}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.136, + 0.907, + 0.226 + ], + "angle": 0, + "content": "For differentiable splatting [44], the Gaussians in the 3D world coordinate system are projected into the 2D camera coordinate system. This projection uses the viewing transformation \\(\\mathbf{W} \\in \\mathbb{R}^{3 \\times 3}\\) and the Jacobian \\(\\mathbf{J} \\in \\mathbb{R}^{2 \\times 3}\\) of the affine approximation of the projective transformation to derive the 2D covariance \\(\\Sigma^{2\\mathrm{D}} \\in \\mathbb{R}^{2 \\times 2}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.633, + 0.231, + 0.905, + 0.25 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\Sigma} ^ {2 \\mathrm {D}} = \\mathbf {J} \\mathbf {W} \\boldsymbol {\\Sigma} \\mathbf {W} ^ {\\top} \\mathbf {J} ^ {\\top}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.256, + 0.906, + 0.317 + ], + "angle": 0, + "content": "Each Gaussian includes a set of spherical harmonics (SH) coefficients and an opacity value \\(\\alpha\\) to represent view-dependent color \\(\\mathbf{c}\\). The pixel color \\(\\mathbf{c}_p\\) is then obtained by applying alpha blending to \\(\\mathcal{N}\\) ordered Gaussians:" + }, + { + "type": "equation", + "bbox": [ + 0.612, + 0.323, + 0.905, + 0.365 + ], + "angle": 0, + "content": "\\[\n\\mathbf {c} _ {p} = \\sum_ {i \\in \\mathcal {N}} \\mathbf {c} _ {i} \\alpha_ {i} \\prod_ {j = 1} ^ {i - 1} (1 - \\alpha_ {j}). \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.371, + 0.906, + 0.416 + ], + "angle": 0, + "content": "Our approach, as illustrated in Fig. 1, exploits 3DGS and aims to render defocused images by generating multiple Gaussians to form the CoC for each Gaussian." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.423, + 0.759, + 0.439 + ], + "angle": 0, + "content": "3.2. 3D Scene Blind Deblurring" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.446, + 0.907, + 0.612 + ], + "angle": 0, + "content": "Image blind deblurring is a technique for learning an image blurring kernel without the supervision of sharp images. Blurry images are obtained by applying convolution with a learned blurring kernel to sharp images, where the kernel size is fixed as a grid structure at each pixel position. Inspired by this approach, Deblur-NeRF [18] applies this technique to NeRF, presenting a method to represent a clean 3D scene using only blurry images. Specifically, Deblur-NeRF warps the input ray into multiple rays that constitute the blur, modeling a ray-based sparse kernel. A blurry pixel is obtained by combining the pixel colors of each ray:" + }, + { + "type": "equation", + "bbox": [ + 0.577, + 0.619, + 0.905, + 0.641 + ], + "angle": 0, + "content": "\\[\n\\mathbf {c} _ {b l u r} = \\sum w _ {p} ^ {i} \\mathbf {c} _ {p} ^ {i}, w. r. t., \\sum w _ {p} ^ {i} = 1, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.646, + 0.906, + 0.72 + ], + "angle": 0, + "content": "where \\( w_{p} \\) and \\( \\mathbf{c}_p \\) denote the weight and pixel color for each corresponding ray. Deblur-NeRF enhances learning efficiency by setting the number of warped rays smaller than the kernel size of 2D convolution, and it designs a deformable kernel by adaptive origins and directions of rays." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.721, + 0.906, + 0.812 + ], + "angle": 0, + "content": "As we adopt a rasterization method based on 3DGS [9], we propose an alternative approach: instead of warping rays, we generate 3D Gaussians in the shape of the CoC and render them. In other words, the defocused image is estimated by setting the generated 3D Gaussians, namely CoC Gaussians, as the blurring kernel." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.822, + 0.605, + 0.838 + ], + "angle": 0, + "content": "4. Method" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.848, + 0.761, + 0.862 + ], + "angle": 0, + "content": "4.1. CoCoGaussian Framework" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.906, + 0.902 + ], + "angle": 0, + "content": "Our goal is to represent a 3D scene using only defocused multi-view images, leveraging fundamental photographic" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16103" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.089, + 0.904, + 0.403 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.415, + 0.908, + 0.501 + ], + "angle": 0, + "content": "Figure 2. We take the camera position \\(\\mathbf{x}_{cam}\\) and the base Gaussian parameters \\(\\mu_B\\), \\(\\mathbf{s}_B\\), and \\(\\mathbf{q}_B\\) as inputs to the MLP \\(h_\\theta\\), which produces five outputs in total. (a) We set the depth \\(d(\\mu_B)\\) as the Euclidean distance between \\(\\mathbf{x}_{cam}\\) and \\(\\mu_B\\). Using the output \\(K\\) from \\(h_\\theta\\), the \\(d(\\mu_B)\\) and the learnable focus plane \\(d_F\\), we apply Eq. (7) to determine the CoC diameter. The diameter, combined with the outputs \\(\\beta\\) and \\(\\mathbf{d}\\) from \\(h_\\theta\\), is then used in Eq. (9) to compute the offset values for \\(\\mu_{\\mathrm{CoC}}\\). By adding these offsets to \\(\\mu_B\\), we obtain the \\(\\mathbf{G}_{\\mathrm{CoC}}\\) means. (b) We obtain \\(\\mathbf{s}_{\\mathrm{CoC}}\\) and \\(\\mathbf{q}_{\\mathrm{CoC}}\\) by applying Eqs. (10) and (11) to \\(\\delta \\mathbf{s}_{\\mathrm{CoC}}\\) and \\(\\delta \\mathbf{q}_{\\mathrm{CoC}}\\), the outputs of \\(h_\\theta\\). (c) Using \\(\\mu_{\\mathrm{CoC}}\\), \\(\\mathbf{s}_{\\mathrm{CoC}}\\), and \\(\\mathbf{q}_{\\mathrm{CoC}}\\), we get the \\(\\mathbf{G}_{\\mathrm{CoC}}\\). Finally, we rasterize \\(\\mathbf{G}_{\\mathrm{CoC}}\\) along with \\(\\mathbf{G}_B\\) to produce \\((M + 1)\\) images, and apply a weighted sum to generate the final defocused image." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.525, + 0.485, + 0.873 + ], + "angle": 0, + "content": "principles to guide our method. We generate 3D Gaussians to capture a blurring kernel shaped as the CoC. The whole framework of CoCoGaussian is on Fig. 2. First, we compute the CoC diameter using the Gaussian depth, derived from the means of the given base 3D Gaussian set \\(\\mathbf{G}_B\\), which consists of \\(N\\) Gaussians, along with the camera position and aperture information (Sec. 4.2). Next, for the base Gaussian set \\(\\mathbf{G}_B\\), we generate \\(M\\) CoC Gaussian sets \\(\\mathbf{G}_{\\mathrm{CoC}}\\), resulting in a total of \\((M\\times N)\\) Gaussians to capture the CoC shape. However, when the scene contains objects with refraction or reflection, the Gaussian depth may be unreliable, which can lead to suboptimal CoC diameter. To address this issue, we propose an adaptive CoC Gaussian generation method that minimizes the dependence on depth for \\(\\mathbf{G}_{\\mathrm{CoC}}\\) (Sec. 4.3). Finally, we incorporate adaptive rotation quaternion and scaling vector adjustments, inspired by the Deblurring 3DGS [10] approach, to allow flexible learning of \\(\\mathbf{G}_{\\mathrm{CoC}}\\). Furthermore, we apply Eq. (5) to images rendered from \\(\\mathbf{G}_B\\) and \\(\\mathbf{G}_{\\mathrm{CoC}}\\), using a weighted sum approach [12, 26] to create the final blurry image. This photography-prior-based approach allows us to accurately represent 3D scenes using only defocused images while adhering to realistic defocus principles." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.524, + 0.853, + 0.54 + ], + "angle": 0, + "content": "4.2. Circle of Confusion from 3D Gaussians" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.548, + 0.907, + 0.715 + ], + "angle": 0, + "content": "To model the CoC, we assume an ideal system that ignores distortions such as lens aberrations. Before generating the CoC Gaussian sets \\(\\mathbf{G}_{\\mathrm{CoC}}\\), we first need to obtain the CoC diameter, which depends on several factors: the aperture diameter \\(D\\), focal length \\(f\\), focus plane \\(d_{F}\\), and depth \\(p\\). The focus plane represents the distance from the camera position to the plane in focus. We define the depth \\(d(\\mu_B)\\) as the Euclidean distance between the camera position \\(\\mathbf{x}_{cam} \\in \\mathbb{R}^3\\) and the means of the base Gaussian set \\(\\mu_B \\in \\mathbb{R}^{N \\times 3}\\). Exploiting these values, we express the CoC diameter \\(\\sigma(\\mu_B)\\) as follows [7]:" + }, + { + "type": "equation", + "bbox": [ + 0.596, + 0.725, + 0.907, + 0.76 + ], + "angle": 0, + "content": "\\[\n\\sigma (\\mu_ {B}) = f D \\times \\frac {\\left| d (\\mu_ {B}) - d _ {F} \\right|}{d (\\mu_ {B}) (d _ {F} - f)}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.769, + 0.909, + 0.86 + ], + "angle": 0, + "content": "where \\(d_F\\) is set to a learnable parameter and initialized as the average distance between \\(\\mathbf{x}_{cam}\\) and the points in the SfM point cloud to ensure stable training. Note that the focus plane \\(d_F\\) is specific to each input image. Additionally, because the focal length \\(f\\) is usually much smaller than the focus plane \\(d_F\\), we modify Eq. (6) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.602, + 0.871, + 0.907, + 0.906 + ], + "angle": 0, + "content": "\\[\n\\sigma \\left(\\mu_ {B}\\right) \\approx K \\times \\left| \\frac {1}{d \\left(\\mu_ {B}\\right)} - \\frac {1}{d _ {F}} \\right|, \\tag {7}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "16104" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.121 + ], + "angle": 0, + "content": "where the product of the focal length and aperture diameter is represented as a single learnable scalar \\( K = f \\times D \\) [41]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.122, + 0.484, + 0.228 + ], + "angle": 0, + "content": "We design a simple MLP \\( h_\\theta \\) parameterized by \\( \\theta \\) to estimate the scalar \\( K \\). Then, the aperture parameter \\( K \\) is obtained by passing these features as inputs to \\( h_\\theta \\), along with the scaling vectors \\( \\mathbf{s}_B \\in \\mathbb{R}^{N \\times 3} \\) and rotation quaternions \\( \\mathbf{q}_B \\in \\mathbb{R}^{N \\times 4} \\) of \\( \\mathbf{G}_B \\). This approach allows us to obtain the CoC diameter \\( \\sigma(\\mu_B) \\) for the \\( \\mathbf{G}_B \\), marking the first step in modeling \\( \\mathbf{G}_{\\mathrm{CoC}} \\)." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.238, + 0.405, + 0.254 + ], + "angle": 0, + "content": "4.3. Adaptive CoC Gaussian Generation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.261, + 0.484, + 0.457 + ], + "angle": 0, + "content": "After obtaining the CoC diameters, we use these diameters to generate \\(M\\) CoC Gaussian sets \\(\\mathbf{G}_{\\mathrm{CoC}}\\). For the Gaussian parameters of \\(\\mathbf{G}_{\\mathrm{CoC}}\\), we set the spherical harmonics (SH) coefficients and opacity values identical to those of \\(\\mathbf{G}_B\\), while varying only the means \\(\\mu\\), scaling vectors \\(\\mathbf{s}\\), and rotation quaternions \\(\\mathbf{q}\\). The means of each CoC Gaussian set are created by adding offsets \\(\\Delta \\mu_{\\mathrm{CoC}} \\in \\mathbb{R}^{M \\times N \\times 3}\\) to \\(\\mu_B\\), defined as \\(\\mu_{\\mathrm{CoC}} = \\mu_B + \\Delta \\mu_{\\mathrm{CoC}}\\). Since \\(\\Delta \\mu_{\\mathrm{CoC}}\\) are 3D vector sets, we obtain unit vector sets \\(\\mathbf{d} \\in \\mathbb{R}^{M \\times N \\times 3}\\) representing the directions from \\(\\mathbf{G}_B\\) to \\(\\mathbf{G}_{\\mathrm{CoC}}\\), which are additional outputs from the MLP \\(h_\\theta\\) discussed in the previous section. Moreover, since \\(\\mathbf{G}_{\\mathrm{CoC}}\\) consists of a total of \\(M\\) Gaussian sets, The direction vectors \\(\\mathbf{d}\\) consist of \\(M\\) instances:" + }, + { + "type": "equation", + "bbox": [ + 0.146, + 0.466, + 0.483, + 0.497 + ], + "angle": 0, + "content": "\\[\n\\Delta \\mu_ {\\mathrm {C o C}; m} = \\frac {\\sigma (\\mu_ {B})}{2} \\mathbf {d} _ {m}, \\text {w h e r e} m \\in M. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.507, + 0.484, + 0.776 + ], + "angle": 0, + "content": "However, theses offsets \\(\\Delta \\mu_{\\mathrm{CoC}}\\) presents two potential issues. First, this approach places all \\(\\mathbf{G}_{\\mathrm{CoC}}\\) only on the boundary of the CoC, which limits its ability to fully capture the defocus effect. Modeling the entire CoC, including its interior, would allow for a more accurate and flexible representation of defocus blur. The second issue arises when there are refractive or reflective surfaces in the scene, as these subjects can lead to incorrect optimization of the Gaussian means, making the CoC diameter obtained through Eq. (7) unreliable. To address these problems, we introduce simple learnable CoC scaling parameters \\(\\beta \\in (0,1]\\), which are additional outputs of \\(h_\\theta\\). The parameter \\(\\beta\\) is learned within the range of 0 to 1, ensuring that \\(\\mu_{\\mathrm{CoC}}\\) resides within the CoC boundary. Additionally, even when depth is inaccurately measured and the predicted CoC diameter is larger than the actual size, \\(\\beta\\) adaptively scales it down. Thus, the offset \\(\\Delta \\mu_{\\mathrm{CoC}}\\) is modified by applying \\(\\beta\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.195, + 0.777, + 0.482, + 0.805 + ], + "angle": 0, + "content": "\\[\n\\Delta \\mu_ {\\mathrm {C o C}; m} = \\frac {\\sigma (\\mu_ {B})}{2} \\beta_ {m} \\mathbf {d} _ {m}. \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Additionally, similar to Deblurring 3DGS [10], we introduce scaling factors \\(\\delta \\mathbf{s}_{\\mathrm{CoC}} \\in \\mathbb{R}^{M \\times N \\times 3}\\) and \\(\\delta \\mathbf{q}_{\\mathrm{CoC}} \\in \\mathbb{R}^{M \\times N \\times 4}\\), which allow flexible learning of the scaling vectors \\(\\mathbf{s}_{\\mathrm{CoC}}\\) and rotation quaternions \\(\\mathbf{q}_{\\mathrm{CoC}}\\) of \\(\\mathbf{G}_{\\mathrm{CoC}}\\). These factors are also the outputs of \\(h_{\\theta}\\). The parameters \\(\\delta \\mathbf{s}_{\\mathrm{CoC}}\\) and \\(\\delta \\mathbf{q}_{\\mathrm{CoC}}\\) are multiplied by the corresponding parameters" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.122 + ], + "angle": 0, + "content": "of the base Gaussian set \\(\\mathbf{G}_B\\), specifically \\(\\mathbf{s}_B\\) and \\(\\mathbf{q}_B\\), and are expressed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.625, + 0.136, + 0.907, + 0.153 + ], + "angle": 0, + "content": "\\[\n\\mathbf {s} _ {\\mathrm {C o C}; m} = \\mathbf {s} _ {B} \\times \\delta \\mathbf {s} _ {\\mathrm {C o C}; m}, \\tag {10}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.623, + 0.166, + 0.907, + 0.181 + ], + "angle": 0, + "content": "\\[\n\\mathbf {q} _ {\\mathrm {C o C}; m} = \\mathbf {q} _ {B} \\times \\delta \\mathbf {q} _ {\\mathrm {C o C}; m}, \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.188, + 0.907, + 0.234 + ], + "angle": 0, + "content": "where the scaling parameters are constrained by the fixed values \\(\\delta \\mathbf{s}_{\\mathrm{max}}\\) and \\(\\delta \\mathbf{q}_{\\mathrm{max}}\\): \\(\\delta \\mathbf{s}_{\\mathrm{CoC}} \\in [1, \\delta \\mathbf{s}_{\\mathrm{max}}]\\) and \\(\\delta \\mathbf{q}_{\\mathrm{CoC}} \\in [1, \\delta \\mathbf{q}_{\\mathrm{max}}]\\)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.235, + 0.907, + 0.295 + ], + "angle": 0, + "content": "We obtain a total of \\(M\\) means of CoC Gaussian sets \\(\\mu_{\\mathrm{CoC}}\\), scaling vectors \\(\\mathbf{s}_{\\mathrm{CoC}}\\), and rotation quaternion \\(\\mathbf{q}_{\\mathrm{CoC}}\\). They are combined with the opacity and SH coefficients of \\(\\mathbf{G}_B\\), generate the \\(M\\) Gaussian sets \\(\\mathbf{G}_{\\mathrm{CoC}}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.533, + 0.309, + 0.907, + 0.326 + ], + "angle": 0, + "content": "\\[\n\\mathbf {x} _ {\\mathrm {C o C}; m}, \\mathbf {s} _ {\\mathrm {C o C}; m}, \\mathbf {q} _ {\\mathrm {C o C}; m}, \\alpha_ {B}, \\mathbf {S H} _ {B} \\rightarrow \\mathbf {G} _ {\\mathrm {C o C}; m}. \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.337, + 0.906, + 0.398 + ], + "angle": 0, + "content": "We rasterize the resulting \\(M\\) Gaussian sets along with the base Gaussian set, producing \\((M + 1)\\) images. Then, using the weighted sum approach from Eq. (5), as detailed in the following section, we obtain the final blurry image." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.408, + 0.654, + 0.424 + ], + "angle": 0, + "content": "4.4. Optimization" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.431, + 0.907, + 0.553 + ], + "angle": 0, + "content": "Weighted Sum. To apply Eq. (5) to the \\((M + 1)\\) images \\(\\mathbf{I}\\) rasterized from \\(\\mathbf{G}_B\\) and \\(\\mathbf{G}_{\\mathrm{CoC}}\\), we adopt the methods of recent studies [12, 26]. We use a shallow CNN \\(\\mathcal{F}\\) to compute pixel-wise weights \\(\\mathcal{W}\\) for these images, applying a softmax function to ensure that the weights for each pixel sum to 1. Then, we multiply each image by its corresponding weight and sum the results to obtain the final defocused image \\(\\mathcal{I}_{blur}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.522, + 0.564, + 0.906, + 0.606 + ], + "angle": 0, + "content": "\\[\n\\mathcal {I} _ {\\text {b l u r}} = \\sum_ {m = 1} ^ {M + 1} \\mathcal {I} _ {m} \\mathcal {W} _ {m}, \\text {w h e r e} \\mathcal {W} = \\operatorname {s o f t m a x} (\\mathcal {F} (\\mathbf {I})), \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.617, + 0.906, + 0.647 + ], + "angle": 0, + "content": "where \\(\\mathcal{I}_m\\) represents the image rasterized from \\(\\mathbf{G}_{\\mathrm{CoC};m}\\), and \\(\\mathcal{I}_{M + 1}\\) represents the image rasterized from \\(\\mathbf{G}_B\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.667, + 0.906, + 0.759 + ], + "angle": 0, + "content": "Objective. We minimize the \\(\\mathcal{L}_1\\) loss and D-SSIM loss \\(\\mathcal{L}_{\\mathrm{D - SSIM}}\\) between the ground truth defocused image and the output defocused image. The \\(\\mathcal{L}_1\\) loss reduces the pixelwise differences between images, while the D-SSIM loss minimizes structural differences between them. The final objective \\(\\mathcal{L}_{rgb}\\) is defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.604, + 0.771, + 0.907, + 0.789 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {r g b} = (1 - \\lambda) \\mathcal {L} _ {1} + \\lambda \\mathcal {L} _ {\\mathrm {D} - \\mathrm {S S I M}}, \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.8, + 0.795, + 0.815 + ], + "angle": 0, + "content": "where we use \\(\\lambda = 0.3\\) for all experiments." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.83, + 0.646, + 0.847 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Datasets. We evaluate our method on three different datasets: the Deblur-NeRF [18] synthetic dataset, the Deblur-NeRF real-world dataset, and the DoF-NeRF [41]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16105" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.483, + 0.162 + ], + "angle": 0, + "content": "Table 1. Comparisons on Deblur-NeRF synthetic and real-world scene dataset. We evaluate the performance on three metrics (PSNR, SSIM, LPIPS). “*” denotes the results obtained by reproducing the released code. The orange and yellow cells respectively indicate the highest and second-highest value." + }, + { + "type": "table", + "bbox": [ + 0.104, + 0.174, + 0.481, + 0.328 + ], + "angle": 0, + "content": "
MethodsSynthetic Scene [18]Real-World Scene [18]
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
Naive NeRF [20]25.950.77910.230322.400.66610.2310
3DGS [9]25.110.74760.214823.380.66550.3140
Deblur-NeRF [18]28.370.85270.118823.470.71990.1207
PDRF-10 [25]30.080.89310.110123.850.73820.1746
DP-NeRF [11]29.330.87130.098723.670.72990.1082
Deblurring 3DGS* [10]28.900.89120.105223.540.73830.1232
BAGS* [26]30.650.91280.063123.480.74080.0962
Ours30.840.92120.047823.700.75310.0825
" + }, + { + "type": "table_caption", + "bbox": [ + 0.102, + 0.345, + 0.472, + 0.358 + ], + "angle": 0, + "content": "Table 2. Comparisons on DoF-NeRF real-world scene dataset." + }, + { + "type": "table", + "bbox": [ + 0.122, + 0.37, + 0.46, + 0.495 + ], + "angle": 0, + "content": "
MethodsDoF-NeRF Real-World Scene [41]
PSNR↑SSIM↑LPIPS↓
3DGS [9]25.720.82910.1817
DP-NeRF [11]26.800.81450.1790
Deblurring 3DGS [10]26.580.83860.1708
BAGS [26]29.870.88160.1100
Ours30.140.91270.0701
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.519, + 0.483, + 0.716 + ], + "angle": 0, + "content": "real-world dataset. The Deblur-NeRF synthetic dataset consists of 5 scenes, each containing defocused images generated using the built-in function of Blender [5]. The Deblur-NeRF real-world dataset includes 10 scenes, with images captured by enlarging the aperture on a Canon EOS RP. The DoF-NeRF real-world dataset comprises 7 scenes, each containing 20 to 30 image triplets. Each triplet includes an all-in-focus image, along with two defocused images focused on the background and foreground. Since DoF-NeRF does not provide training code, we use half of the training defocused images with a foreground focus and the other half with a background focus. All camera poses and initial point clouds are obtained through COLMAP [29, 30]." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.726, + 0.27, + 0.743 + ], + "angle": 0, + "content": "5.1. Rendering Results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Quantitative Results. We quantitatively evaluate our CoCoGaussian using the following three metrics: peak signal-to-noise ratio (PSNR), structural similarity index measure (SSIM) [38], and learned perceptual image patch similarity (LPIPS) [45]. We compare our approach to several state-of-the-art methods, including both raytracing [11, 18, 25] and rasterization-based methods [10, 26]. Quantitative results are shown in Tab. 1 and Tab. 2, where our approach achieves the best performance across all metrics except for PSNR on the Deblur-NeRF real-world" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.365 + ], + "angle": 0, + "content": "dataset. Note that we report the reproduced performance scores for Deblurring 3DGS [10] and BAGS [26] using their official codes. The Deblur-NeRF real-world dataset inherently presents challenges, as there are illumination differences between the sharp and defocused images, leading to relatively lower PSNR score. We discuss this issue in the appendix. For the DoF-NeRF dataset, CoCoGaussian achieves outstanding performance across all metrics, with an approximately 0.04 reduction in LPIPS compared to BAGS. Since the DoF-NeRF dataset has a higher resolution than the Deblur-NeRF dataset, this performance drop in BAGS is likely due to its requirement for manual adjustment of blurring kernel size, highlighting an intrinsic limitation. Conversely, Deblurring 3DGS shows minimal improvement over 3DGS, which can also be attributed to its limited kernel size relative to the higher resolution. The details regarding the time complexity of our model are provided in the appendix." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.385, + 0.909, + 0.522 + ], + "angle": 0, + "content": "Qualitative Results. For qualitative evaluation, we present the rendering results of 3DGS [9], Deblurring 3DGS [10], and BAGS [26] in Fig. 3. Our method demonstrates higher-fidelity results compared to others. Specifically, in the fourth row of our results, the white line on the right side is rendered more sharply, and in the fifth row, the detailed quality of the graphic card appears superior to that of BAGS. Additionally, more CoC visualizations similar to Fig. 1 (b), are in the appendix." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.532, + 0.667, + 0.548 + ], + "angle": 0, + "content": "5.2. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.556, + 0.907, + 0.693 + ], + "angle": 0, + "content": "To validate the effectiveness of our proposed methods, we perform a series of ablation studies, with all experiments conducted on the DoF-NeRF real-world dataset. We chose this dataset as it is based on real captured images and does not present the illumination issues found in the Deblur-NeRF real-world dataset, ensuring more reliable evaluation. The results are summarized in Tab. 3, where the baseline denotes naive 3DGS [9]. Additional ablative experiments with other factors are in the appendix." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.713, + 0.907, + 0.85 + ], + "angle": 0, + "content": "Circle of Confusion. To demonstrate the impact of our core concept, CoC, we conduct an experiment that excludes the CoC from the model, retaining only \\(\\beta_{m}\\) and \\(\\mathbf{d}_m\\). This configuration yields significantly lower performance than the full model, indicating an over-reliance on parametric learning. The result suggests that without the CoC, which leverages depth and aperture information, the means of the generated Gaussian sets are not optimally arranged, leading to overfitting." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.907, + 0.901 + ], + "angle": 0, + "content": "CoC Direction Vector. In this experiment, we retain the CoC concept but exclude the learnable direction vectors" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16106" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.09, + 0.245, + 0.169 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.246, + 0.091, + 0.374, + 0.169 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.376, + 0.091, + 0.507, + 0.169 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.091, + 0.64, + 0.169 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.641, + 0.091, + 0.772, + 0.169 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.773, + 0.091, + 0.904, + 0.169 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.171, + 0.244, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.246, + 0.171, + 0.375, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.377, + 0.171, + 0.506, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.171, + 0.639, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.641, + 0.171, + 0.772, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.773, + 0.171, + 0.904, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.25, + 0.244, + 0.328 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.246, + 0.25, + 0.375, + 0.328 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.377, + 0.25, + 0.506, + 0.328 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.25, + 0.639, + 0.328 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.641, + 0.25, + 0.772, + 0.328 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.773, + 0.25, + 0.904, + 0.328 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.33, + 0.244, + 0.409 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.246, + 0.33, + 0.375, + 0.409 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.377, + 0.33, + 0.506, + 0.409 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.33, + 0.639, + 0.409 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.641, + 0.33, + 0.772, + 0.409 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.773, + 0.33, + 0.904, + 0.409 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.411, + 0.244, + 0.489 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.246, + 0.411, + 0.375, + 0.489 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.377, + 0.411, + 0.506, + 0.489 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.411, + 0.639, + 0.489 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.641, + 0.411, + 0.772, + 0.489 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.773, + 0.411, + 0.904, + 0.489 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.491, + 0.243, + 0.569 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.135, + 0.572, + 0.203, + 0.585 + ], + "angle": 0, + "content": "Reference" + }, + { + "type": "image", + "bbox": [ + 0.245, + 0.491, + 0.375, + 0.569 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.288, + 0.572, + 0.331, + 0.585 + ], + "angle": 0, + "content": "3DGS" + }, + { + "type": "image", + "bbox": [ + 0.377, + 0.491, + 0.507, + 0.569 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.384, + 0.572, + 0.501, + 0.586 + ], + "angle": 0, + "content": "Deblurring 3DGS" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.491, + 0.639, + 0.569 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.551, + 0.572, + 0.597, + 0.585 + ], + "angle": 0, + "content": "BAGS" + }, + { + "type": "image", + "bbox": [ + 0.641, + 0.491, + 0.772, + 0.569 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.688, + 0.572, + 0.725, + 0.585 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.773, + 0.491, + 0.904, + 0.569 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.794, + 0.572, + 0.884, + 0.585 + ], + "angle": 0, + "content": "Ground Truth" + }, + { + "type": "image_caption", + "bbox": [ + 0.262, + 0.6, + 0.735, + 0.616 + ], + "angle": 0, + "content": "Figure 3. Qualitative comparison on the Deblur-NeRF and DoF-NeRF datasets." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.623, + 0.484, + 0.73 + ], + "angle": 0, + "content": "\\(\\mathbf{d}_m\\). Instead of using learnable vectors, we use fixed direction vectors, arranging \\(M\\) CoC Gaussians evenly in a circular pattern on a plane perpendicular to the vector between the camera position and the base Gaussian. This configuration slightly underperforms compared to the full model, as the fixed directions restrict optimal Gaussian position, preventing adaptation to the most reliable configuration." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.484, + 0.903 + ], + "angle": 0, + "content": "CoC Scaling Factor. In this setup, we include the learnable direction vector while omitting the CoC scaling factor \\(\\beta_{m}\\). Although the performance is slightly better than when the direction vector is excluded, it remains lower than that of the full model. Without the CoC scaling factor, CoC Gaussian generation becomes overly dependent on depth, leading to errors in CoC diameter calculation when depth values are inaccurate. Including the CoC scaling factor enables the full model to handle uncertain depth more robustly, improving performance. Related visualization re" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.624, + 0.686, + 0.64 + ], + "angle": 0, + "content": "sults are in the appendix." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.657, + 0.907, + 0.809 + ], + "angle": 0, + "content": "Aperture Parameter. The aperture parameter \\( K \\) includes the aperture diameter and focal length, and it plays a critical role in determining the size of the CoC. Without this parameter, there is an upper limit to the CoC diameter calculation, making accurate computation challenging. Experimental results indicate that although performance is somewhat lower than the full model, it is improved over the baseline. This suggests that while the model is unable to determine precise CoC size without \\( K \\), it still learns the focus plane and captures the in-focus region accurately." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.818, + 0.835, + 0.835 + ], + "angle": 0, + "content": "5.3. Experiments on All-in-Focus Images." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.84, + 0.909, + 0.903 + ], + "angle": 0, + "content": "We conduct novel view synthesis experiments using all-in-focus images to verify the generalization performance of CoCoGaussian. We utilize the NeRF-LLFF [19, 20] dataset, which contains real-world images. As shown in Tab. 4, our" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "16107" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.09, + 0.907, + 0.278 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.281, + 0.905, + 0.31 + ], + "angle": 0, + "content": "Figure 4. Visualization of Aperture parameter and Focus Plane Customization. The top row of images decreases the aperture parameter \\( K \\) from left to right, while the bottom row moves the focus plane \\( d_{F} \\) further from the camera from left to right." + }, + { + "type": "table_caption", + "bbox": [ + 0.155, + 0.323, + 0.42, + 0.336 + ], + "angle": 0, + "content": "Table 3. Ablation Results of CoCoGaussian." + }, + { + "type": "table", + "bbox": [ + 0.099, + 0.342, + 0.48, + 0.455 + ], + "angle": 0, + "content": "
MethodsPSNR↑SSIM↑LPIPS↓
Baseline [9]25.720.82910.1817
w/o CoC28.290.87780.0927
w/o CoC direction vector28.910.88930.0896
w/o CoC scaling29.460.90580.0793
w/o aperture parameter27.370.85100.1113
Ours Full30.140.91270.0701
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.461, + 0.484, + 0.613 + ], + "angle": 0, + "content": "approach outperforms the baseline 3DGS [9] across all metrics. In practice, when the aperture is very small, radiance from a point on the subject forms a tiny CoC on the image sensor after passing through the aperture, even if the subject is slightly away from the focus plane. Because the CoC is so small, the resulting image is not perceived as defocused. By modeling this small circle, which is difficult to detect with the human eye, our approach achieves higher performance than the baseline 3DGS. Related CoC visualizations are in the appendix." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.621, + 0.476, + 0.637 + ], + "angle": 0, + "content": "5.4.Customizable Depth of Field and Focus Plane" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.645, + 0.483, + 0.84 + ], + "angle": 0, + "content": "Customizing the depth of field and focus settings is essential in various applications, such as augmented reality (AR) and virtual reality (VR), where control over depth cues and focus effects can greatly enhance visual realism and viewer engagement. Since our approach incorporates the aperture parameter \\( K \\) in CoC modeling, it enables customization of both CoC size and depth of field by adjusting this parameter. The top row in Fig. 4 visualizes increasing depth of field by gradually reducing \\( K \\) from left to right. As the focus plane is positioned at a shallow depth, the model consistently captures focus in this region. However, in the leftmost image, the large aperture causes the CoC size to increase at deeper depths, resulting in defocus blur." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Additionally, our model can customize the focus plane of defocused images, as shown in the bottom row of Fig. 4, where the focus plane shifts from shallow depth on the left to deeper depth on the right. The shallow regions are in fo" + }, + { + "type": "table_caption", + "bbox": [ + 0.561, + 0.323, + 0.857, + 0.336 + ], + "angle": 0, + "content": "Table 4. Results on NeRF-LLFF [19, 20] Dataset." + }, + { + "type": "table", + "bbox": [ + 0.576, + 0.342, + 0.842, + 0.39 + ], + "angle": 0, + "content": "
MethodsPSNR↑SSIM↑LPIPS↓
3DGS [9]27.100.86190.0569
Ours27.760.87380.0372
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.4, + 0.907, + 0.506 + ], + "angle": 0, + "content": "cus on the leftmost image, while deeper areas remain out of focus, and the opposite holds for the rightmost image. This level of control over focus and depth effects is essential for achieving realistic depth cues. Therefore, CoCoGaussian not only demonstrates superior capability in synthesizing sharp novel view images but also provides enhanced flexibility in adjusting aperture and focus plane settings." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.519, + 0.779, + 0.535 + ], + "angle": 0, + "content": "6. Limitation and Future Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.544, + 0.907, + 0.681 + ], + "angle": 0, + "content": "While our model demonstrates strong qualitative and quantitative performance, there is potential for further optimization in the adaptive CoC scaling factor. To ensure training stability, we currently limit the factor size, enabling adaptive optimization mainly when the CoC diameter is overestimated. However, future improvements could enhance adaptability to handle cases where the CoC diameter is underestimated, further improving depth handling in complex scenes without compromising stability." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.695, + 0.634, + 0.71 + ], + "angle": 0, + "content": "7. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In this work, we propose CoCoGaussian, a 3D scene representation method that effectively leverages photographic defocus principles to handle defocused image inputs. CoCoGaussian accurately models defocus blur by constructing the CoC through 3D Gaussians and learnable aperture parameters. We introduce an adaptive CoC Gaussian generation method to address the challenges posed by uncertain depth, making our model robust to reflective or refractive objects. Our approach demonstrates the feasibility and effectiveness of representing 3D scenes from defocused images, advancing possibilities for applications in real-world image synthesis." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16108" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.152 + ], + "angle": 0, + "content": "Acknowledgements. This project was supported by the Naver Cloud Corporation and the National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIT) (No. RS-2024-00340745)." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.166, + 0.188, + 0.182 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.192, + 0.482, + 0.274 + ], + "angle": 0, + "content": "[1] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5855–5864, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.277, + 0.482, + 0.343 + ], + "angle": 0, + "content": "[2] Ayan Chakrabarti. A neural approach to blind motion deblurring. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part III 14, pages 221-235. Springer, 2016. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.346, + 0.482, + 0.401 + ], + "angle": 0, + "content": "[3] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In European Conference on Computer Vision, pages 333-350. Springer, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.404, + 0.482, + 0.457 + ], + "angle": 0, + "content": "[4] Wenbo Chen and Ligang Liu. Deblur-gs: 3d gaussian splattering from camera motion blurred images. Proceedings of the ACM on Computer Graphics and Interactive Techniques, 7 (1):1-15, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.461, + 0.482, + 0.5 + ], + "angle": 0, + "content": "[5] Blender Online Community. Blender - a 3D modelling and rendering package. Blender Foundation, Stichting Blender Foundation, Amsterdam, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.504, + 0.482, + 0.572 + ], + "angle": 0, + "content": "[6] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5501–5510, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.575, + 0.478, + 0.588 + ], + "angle": 0, + "content": "[7] Eugene Hecht. Optics. Pearson Education India, 2012. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.59, + 0.482, + 0.644 + ], + "angle": 0, + "content": "[8] Wei Jiang, Kwang Moo Yi, Golnoosh Samei, Oncel Tuzel, and Anurag Ranjan. Neuman: Neural human radiance field from a single video. In European Conference on Computer Vision, pages 402-418. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.647, + 0.482, + 0.7 + ], + "angle": 0, + "content": "[9] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42 (4):1-14, 2023. 2, 3, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.704, + 0.482, + 0.744 + ], + "angle": 0, + "content": "[10] Byeonghyeon Lee, Howoong Lee, Xiangyu Sun, Usman Ali, and Eunbyung Park. Deblurring 3d gaussian splatting. arXiv preprint arXiv:2401.00834, 2024. 2, 3, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.747, + 0.482, + 0.814 + ], + "angle": 0, + "content": "[11] Dogyoon Lee, Minhyeok Lee, Chajin Shin, and Sangyoun Lee. Dp-nerf: Deblurred neural radiance field with physical scene priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12386–12396, 2023. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.818, + 0.482, + 0.87 + ], + "angle": 0, + "content": "[12] Junghe Lee, Donghyeong Kim, Dogyoon Lee, Suhwan Cho, and Sangyoun Lee. Crime: Continuous rigid motion-aware gaussian splatting from motion blur images. arXiv preprint arXiv:2407.03923, 2024. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.874, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[13] Jungho Lee, Dogyoon Lee, Minhyeok Lee, Donghyung Kim, and Sangyoun Lee. Smurf: Continuous dynam-" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.192, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.094, + 0.905, + 0.119 + ], + "angle": 0, + "content": "ics for motion-deblurring radiance fields. arXiv preprint arXiv:2403.07547, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.122, + 0.905, + 0.204 + ], + "angle": 0, + "content": "[14] Tianye Li, Mira Slavcheva, Michael Zollhoefer, Simon Green, Christoph Lassner, Changil Kim, Tanner Schmidt, Steven Lovegrove, Michael Goesele, Richard Newcombe, et al. Neural 3d video synthesis from multi-view video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5521-5531, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.206, + 0.905, + 0.272 + ], + "angle": 0, + "content": "[15] Zhengqi Li, Simon Niklaus, Noah Snavely, and Oliver Wang. Neural scene flow fields for space-time view synthesis of dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6498-6508, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.274, + 0.904, + 0.343 + ], + "angle": 0, + "content": "[16] Zhaoshuo Li, Thomas Müller, Alex Evans, Russell H Taylor, Mathias Unberath, Ming-Yu Liu, and Chen-Hsuan Lin. Neuralangelo: High-fidelity neural surface reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8456-8465, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.345, + 0.904, + 0.371 + ], + "angle": 0, + "content": "[17] Kevin M Lynch and Frank C Park. Modern robotics. Cambridge University Press, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.373, + 0.905, + 0.44 + ], + "angle": 0, + "content": "[18] Li Ma, Xiaoyu Li, Jing Liao, Qi Zhang, Xuan Wang, Jue Wang, and Pedro V Sander. Deblur-nerf: Neural radiance fields from blurry images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12861–12870, 2022. 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.443, + 0.905, + 0.51 + ], + "angle": 0, + "content": "[19] Ben Mildenhall, Pratul P Srinivasan, Rodrigo Ortiz-Cayon, Nima Khademi Kalantari, Ravi Ramamoorthi, Ren Ng, and Abhishek Kar. Local light field fusion: Practical view synthesis with prescriptive sampling guidelines. ACM Transactions on Graphics (ToG), 38(4):1-14, 2019. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.512, + 0.905, + 0.594 + ], + "angle": 0, + "content": "[20] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I, pages 405-421, 2020. 1, 2, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.597, + 0.905, + 0.649 + ], + "angle": 0, + "content": "[21] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG), 41(4):1-15, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.651, + 0.905, + 0.733 + ], + "angle": 0, + "content": "[22] Michael Niemeyer, Jonathan T Barron, Ben Mildenhall, Mehdi SM Sajjadi, Andreas Geiger, and Noha Radwan. Regnerf: Regularizing neural radiance fields for view synthesis from sparse inputs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5480-5490, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.735, + 0.905, + 0.803 + ], + "angle": 0, + "content": "[23] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5865-5874, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.805, + 0.905, + 0.872 + ], + "angle": 0, + "content": "[24] Keunhong Park, Utkarsh Sinha, Peter Hedman, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Ricardo MartinBrualla, and Steven M Seitz. Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields. arXiv preprint arXiv:2106.13228, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[25] Cheng Peng and Rama Chellappa. Pdrf: progressively deblurring radiance field for fast scene reconstruction from" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.094, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16109" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.12 + ], + "angle": 0, + "content": "blurry images. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2029-2037, 2023. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.177 + ], + "angle": 0, + "content": "[26] Cheng Peng, Yutao Tang, Yifan Zhou, Nengyu Wang, Xijun Liu, Deming Li, and Rama Chellappa. Bags: Blur agnostic gaussian splatting through multi-scale kernel modeling. arXiv preprint arXiv:2403.04926, 2024. 2, 3, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.179, + 0.482, + 0.248 + ], + "angle": 0, + "content": "[27] Sida Peng, Junting Dong, Qianqian Wang, Shangzhan Zhang, Qing Shuai, Xiaowei Zhou, and Hujun Bao. Animatable neural radiance fields for modeling dynamic human bodies. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14314-14323, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.25, + 0.482, + 0.318 + ], + "angle": 0, + "content": "[28] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-nerf: Neural radiance fields for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10318–10327, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.321, + 0.482, + 0.376 + ], + "angle": 0, + "content": "[29] Johannes L Schonberger, Enliang Zheng, Jan-Michael Frahm, and Marc Pollefeys. Pixelwise view selection for unstructured multi-view stereo. In European conference on computer vision, pages 501-518. Springer, 2016. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.378, + 0.482, + 0.419 + ], + "angle": 0, + "content": "[30] Qi Shan, Jiaya Jia, and Aseem Agarwala. High-quality motion deblurring from a single image. Acm transactions on graphics (tog), 27(3):1-10, 2008. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.421, + 0.482, + 0.476 + ], + "angle": 0, + "content": "[31] Pratul P Srinivasan, Ren Ng, and Ravi Ramamoorthi. Light field blind motion deblurring. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3958-3966, 2017. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.478, + 0.482, + 0.547 + ], + "angle": 0, + "content": "[32] Jiaming Sun, Yiming Xie, Linghao Chen, Xiaowei Zhou, and Hujun Bao. Neuralrecon: Real-time coherent 3d reconstruction from monocular video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15598-15607, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.549, + 0.482, + 0.631 + ], + "angle": 0, + "content": "[33] Edgar Tretschk, Ayush Tewari, Vladislav Golyanik, Michael Zollhöfer, Christoph Lassner, and Christian Theobalt. Non-rigid neural radiance fields: Reconstruction and novel view synthesis of a dynamic scene from monocular video. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12959-12970, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.633, + 0.482, + 0.701 + ], + "angle": 0, + "content": "[34] Guangcong Wang, Zhaoxi Chen, Chen Change Loy, and Zwei Liu. Sparsenerf: Distilling depth ranking for few-shot novel view synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9065-9076, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.704, + 0.482, + 0.773 + ], + "angle": 0, + "content": "[35] Jiepeng Wang, Peng Wang, Xiaoxiao Long, Christian Theobalt, Taku Komura, Lingjie Liu, and Wenping Wang. Neuris: Neural reconstruction of indoor scenes using normal priors. In European Conference on Computer Vision, pages 139-155. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.775, + 0.482, + 0.83 + ], + "angle": 0, + "content": "[36] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. arXiv preprint arXiv:2106.10689, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.482, + 0.899 + ], + "angle": 0, + "content": "[37] Peng Wang, Lingzhe Zhao, Ruijie Ma, and Peidong Liu. Bad-nerf: Bundle adjusted deblur neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4170–4179, 2023. 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[38] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.149, + 0.905, + 0.232 + ], + "angle": 0, + "content": "[39] Chung-Yi Weng, Brian Curless, Pratul P Srinivasan, Jonathan T Barron, and Ira Kemelmacher-Shlizerman. Humanerf: Free-viewpoint rendering of moving people from monocular video. In Proceedings of the IEEE/CVF conference on computer vision and pattern Recognition, pages 16210-16220, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.234, + 0.905, + 0.288 + ], + "angle": 0, + "content": "[40] Oliver Whyte, Josef Sivic, Andrew Zisserman, and Jean Ponce. Non-uniform deblurring for shaken images. International journal of computer vision, 98:168-186, 2012. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.29, + 0.905, + 0.357 + ], + "angle": 0, + "content": "[41] Zijin Wu, Xingyi Li, Juewen Peng, Hao Lu, Zhiguo Cao, and Weicai Zhong. Dof-nerf: Depth-of-field meets neural radiance fields. In Proceedings of the 30th ACM International Conference on Multimedia, pages 1718–1729, 2022. 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.36, + 0.905, + 0.429 + ], + "angle": 0, + "content": "[42] Jamie Wynn and Daniyar Turmukhambetov. Diffusionerf: Regularizing neural radiance fields with denoising diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4180-4189, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.431, + 0.905, + 0.5 + ], + "angle": 0, + "content": "[43] Jiawei Yang, Marco Pavone, and Yue Wang. Freenerf: Improving few-shot neural rendering with free frequency regularization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8254-8263, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.501, + 0.905, + 0.557 + ], + "angle": 0, + "content": "[44] Wang Yifan, Felice Serena, Shihao Wu, Cengiz Özireli, and Olga Sorkine-Hornung. Differentiable surface splatting for point-based geometry processing. ACM Transactions on Graphics (TOG), 38(6):1-14, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.558, + 0.905, + 0.627 + ], + "angle": 0, + "content": "[45] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.629, + 0.905, + 0.669 + ], + "angle": 0, + "content": "[46] Lingzhe Zhao, Peng Wang, and Peidong Liu. Bad-gaussians: Bundle adjusted deblur gaussian splatting. arXiv preprint arXiv:2403.11831, 2024. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.669 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16110" + } + ] +] \ No newline at end of file diff --git a/2025/CoCoGaussian_ Leveraging Circle of Confusion for Gaussian Splatting from Defocused Images/c9f7dcb9-6d18-430a-9426-253bcaf437ef_origin.pdf b/2025/CoCoGaussian_ Leveraging Circle of Confusion for Gaussian Splatting from Defocused Images/c9f7dcb9-6d18-430a-9426-253bcaf437ef_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..384067a8e9f7e2cbcff3dfe07679c178aec2ae7c --- /dev/null +++ b/2025/CoCoGaussian_ Leveraging Circle of Confusion for Gaussian Splatting from Defocused Images/c9f7dcb9-6d18-430a-9426-253bcaf437ef_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:87d7ef97b6b915cef7b87878c6ce34928d6d19ef38efd22268a014e3691651dc +size 6378133 diff --git a/2025/CoCoGaussian_ Leveraging Circle of Confusion for Gaussian Splatting from Defocused Images/full.md b/2025/CoCoGaussian_ Leveraging Circle of Confusion for Gaussian Splatting from Defocused Images/full.md new file mode 100644 index 0000000000000000000000000000000000000000..085a7040baaf69e0eb5b1e786a1c9dc1d7d5af74 --- /dev/null +++ b/2025/CoCoGaussian_ Leveraging Circle of Confusion for Gaussian Splatting from Defocused Images/full.md @@ -0,0 +1,393 @@ +# CoCoGaussian: Leveraging Circle of Confusion for Gaussian Splatting from Defocused Images + +Jungho Lee $^{1*}$ Suhwan Cho $^{1}$ Taeoh Kim $^{2}$ Ho-Deok Jang $^{2}$ Minhyeok Lee $^{1}$ Geonho Cha $^{2}$ Dongyoon Wee $^{2}$ Dogyoon Lee $^{1}$ Sangyoun Lee $^{1}$ + +$^{1}$ School of Electrical and Electronic Engineering, Yonsei University $^{2}$ NAVER Cloud https://Jho-Yonsei.github.io/CoCoGaussian + +![](images/6372921ee99eef124a57bdb275cb484704d28be09f6167b5e9ae6d6aab16c7fc.jpg) +(a) Defocused Image Acquisition + +![](images/209813a7d770da27e57b642b8969f2261761f3a17b2fb0a2ec7343a2402da039.jpg) + +![](images/8365538702c8e3e10961f876106bfabd8df9d0464e33d1cef9a46205c55397c3.jpg) +(b) Reconstructed Gaussian Points + +![](images/972cfdd3b13a32857f3418c0a1d90d6d6e2da4e6a0362463315c24f998ed614d.jpg) + +![](images/a0dfa3ea76ab11e9d7b02e1ac7d0a8ead6cf0eb7294973e9e2ff490e930b63fe.jpg) + +![](images/fdb771382f32b5abcb68f80cc05014b4105a772e4f61b24dfecad55370309602.jpg) +(c) Rendered Images +Figure 1. We propose CoCoGaussian, a novel framework for 3D scene reconstruction from defocused images. (a) With a large aperture, radiance from the focus plane appears as a small circle of confusion (green) on the image sensor, while radiance from greater depths results in larger circles (pink). (b) We visualize reconstructed Gaussians of CoCoGaussian, where the black dots indicate Gaussian means, while the gray dots represent Gaussian means forming the circle of confusion, which decreases in size near the focus plane. (c) The defocused image focuses on shallow depths, with circle size increasing with depth, as shown in (b). CoCoGaussian allows customization of defocused images by adjusting the depth of field or focus plane, while sharp images can be obtained by rendering without the circle of confusion. + +# Abstract + +3D Gaussian Splatting (3DGS) has attracted significant attention for its high-quality novel view rendering, inspiring research to address real-world challenges. While conventional methods depend on sharp images for accurate scene reconstruction, real-world scenarios are often affected by defocus blur due to finite depth of field, making it essential to account for realistic 3D scene representation. In this study, we propose CoCoGaussian, a Circle of Confusion-aware Gaussian Splatting that enables precise 3D scene representation using only defocused images. CoCoGaussian addresses the challenge of defocus blur by modeling the Circle of Confusion (CoC) through a physically grounded approach based on the principles of photo + +graphic defocus. Exploiting 3D Gaussians, we compute the CoC diameter from depth and learnable aperture information, generating multiple Gaussians to precisely capture the CoC shape. Furthermore, we introduce a learnable scaling factor to enhance robustness and provide more flexibility in handling unreliable depth in scenes with reflective or refractive surfaces. Experiments on both synthetic and real-world datasets demonstrate that CoCoGaussian achieves state-of-the-art performance across multiple benchmarks. + +# 1. Introduction + +The emergence of the Neural Radiance Field (NeRF) [20] has brought significant attention to 3D scene representation for photo-realistic novel view synthesis. They reconstruct 3D scenes from images captured from multiple views + +and render images of unseen views. However, since NeRF models 3D scenes as implicit neural representations through a ray tracing-based approach, they suffer from inefficient memory usage. In contrast, a rasterization-based method, 3D Gaussian Splatting (3DGS) [9] explicitly models 3D scenes using independent 3D Gaussians with different 3D means, covariances, opacities, and spherical harmonic coefficients. The tile-based rasterizer of 3DGS applies alpha blending to Gaussian splats sorted by visibility order, enabling fast training, rendering, and efficient memory usage. + +Both NeRF [20] and 3DGS [9] rely on sharp images to represent 3D scenes accurately, which is a highly ideal assumption. In real-world settings, various factors (e.g., finite depth of field and camera motion blur) can hinder the capture of sharp images, leading to image degradation. Achieving a perfectly sharp image requires every part of the scene to be in focus, namely all-in-focus, which calls for a large depth of field. However, obtaining a large depth of field necessitates using a small aperture, which in turn requires a longer exposure time to allow enough light to enter the camera. During this extended exposure, even slight movements of a handheld camera can introduce motion blur. To prevent motion blur, the aperture needs to be widened, but this comes at the cost of a shallower depth of field. As a result, areas outside the focus plane appear blurred due to defocus. Defocus blur is closely related to the concept of the Circle of Confusion (CoC); as shown in Fig. 1, when a subject is positioned away from the focus plane, the radiances from the subject pass through the aperture and are mapped onto the image sensor in a circular pattern, resulting in shallow depth of field and defocus blur [7]. However, despite these challenges, recent research in 3D scene representation has rarely focused on addressing defocus blur directly. In this paper, we propose a method to represent 3D scenes using only images with defocus blur, tackling the CoC-related challenges posed by real-world image acquisition scenarios. + +Recently, there has been increasing interest in reconstructing 3D scenes using only degraded images, with a focus on addressing challenges caused by camera motion blur and defocus blur. Deblur-NeRF [18] is the first study to take on this task, adopting a blind image deblurring approaches [2, 31, 40]. Follow-up studies [10-13, 25, 26, 37] have expanded this approach beyond raytracing to rasterization-based methods. However, many of these methods depend heavily on learning-based strategies and fail to incorporate photographic principles that accurately capture how defocus occurs in real-world scenarios. While DoF-NeRF [41] adopts the concept of CoC to model defocus blur in 3D scene representations, it relies on an implicit representation, which leads to slow rendering speeds, making it less suitable for real-time applications. + +In this paper, we propose CoCoGaussian, which leverages a physically grounded photographic defocus princi + +plies to represent 3D scenes. Our method incorporates an aperture to calculate the CoC diameter based on depth and learnable aperture information using 3D Gaussians, accurately modeling shallow depth of field images. Using the CoC diameter and the 3D Gaussian, we generate multiple Gaussians to form the CoC shape, as illustrated in Fig. 1. However, when objects with refraction or reflection exist in the scene, the depth obtained through the Gaussians may become unreliable. To address this, we propose a method to keep the set of generated Gaussians within the CoC radius, introducing a learnable scaling factor to increase flexibility and reduce dependence on estimated depth in the CoC modeling. As a result, our model combines physically grounded defocus principles with an adaptive CoC modeling that effectively handles uncertain depths, allowing it to reconstruct sharp 3D scenes using only defocused images. Additionally, our approach has an advantage of learning aperture and focus plane information, enabling dynamic control over depth of field and flexible adjustment of the focus plane. This feature allows for highly customizable scene visualization, adapting to various focus and depth requirements. + +To demonstrate the effectiveness of our model, we conduct extensive experiments on the benchmarks: the Deblur-NeRF [18] dataset and the DoF-NeRF [41] real-world dataset. Our contributions can be summarized as follows: + +- CoCoGaussian models the CoC at the 3D Gaussian level, reconstructing the precise 3D scene and enabling sharp novel view synthesis from defocused images. +- We propose an adaptive learning approach that robustly models the CoC even with unreliable depth information. +- CoCoGaussian enables the customizable depth of field and flexible focus plane adjustment during rendering by learning aperture and focus plane information. +- CoCoGaussian achieves state-of-the-art performance, both quantitatively and qualitatively. + +# 2. Related Work + +# 2.1. Scene Representations for Novel View Synthesis + +3D scene representation for novel view synthesis has advanced significantly with the advent of neural radiance fields (NeRF) [20], a ray tracing-based volumetric rendering method that generates photo-realistic novel view images from multi-view images. However, as an implicit representation method using deep MLPs, NeRF suffers from slow training and rendering speeds. To address this limitation, explicit representation methods such as Plenoxel [6], TensoRF [3], and Instant-NGP [21] have been introduced. More recently, 3D Gaussian Splatting (3DGS) [9], a rasterization-based method, has emerged as an alternative to ray tracing, significantly mitigating the speed and memory efficiency constraints of ray tracing + +based approaches. NeRF and 3DGS has enabled a wide range of related research, including dynamic scene representation [14, 15, 23, 24, 28, 33], human avatars [8, 27, 39], 3D mesh reconstruction [16, 32, 35, 36], 3D scene representation from sparse-view images [22, 34, 42, 43], and 3D scene representation from blurry images [4, 10-13, 18, 25, 26, 37, 46]. In this paper, we propose a method for 3D scene representation from defocused images by exploiting 3DGS. + +# 2.2. Novel View Synthesis from Blurry Images + +Scene representation methods like NeRF [20] and 3DGS [9] require sharp images as input to render photo-realistic images. However, in real-world scenarios, capturing sharp images is challenging, often suffering from image degradation such as camera motion blur and defocus blur. To address this issue, Deblur-NeRF [18] introduced a 3D ray-based blurring kernel inspired by image blind deblurring [2, 31, 40], which intentionally generates blurry images during training. After training, it renders sharp novel-view images by excluding the trained kernel. Following Deblur-NeRF, various approaches have been proposed to tackle the degradation issue. DP-NeRF [11] introduces a kernel based on the rigid body transformation [17] with prior knowledge that object shape remains consistent in static scenes, and PDRF [25] proposes a blur estimation method with 2-stage efficient rendering. Recently, methods utilizing 3DGS have been proposed to enable faster rendering. For instance, Deblurring 3DGS [10] adjusts Gaussian parameters to generate blurry images during training, and BAGS [26] proposes a blur-agnostic kernel and a blur mask using convolutional neural networks (CNNs). DoF-NeRF [41] has explored modeling defocus blur by leveraging the circle of confusion (CoC) within an implicit neural representation. In this paper, we propose a method that also leverages CoC to model defocus, which enables real-time rendering, while still maintaining high-quality depth-of-field effects. + +# 3. Preliminary + +# 3.1. 3D Gaussian Splatting + +Unlike ray tracing-based methods [1, 3, 20], 3DGS [9] is built on a rasterization-based approach with differentiable 3D Gaussians. These 3D Gaussians are initialized from a sparse point cloud obtained via a Structure-from-Motion (SfM) [29, 30] algorithm and are defined as follows: + +$$ +G (\mathbf {x}) = e ^ {- \frac {1}{2} (\mathbf {x} - \mu) ^ {\top} \boldsymbol {\Sigma} ^ {- 1} (\mathbf {x} - \mu)}, \tag {1} +$$ + +where $\mathbf{x} \in \mathbb{R}^3$ is a point on the Gaussian $G$ centered at the mean vector $\mu \in \mathbb{R}^3$ with an covariance matrix $\boldsymbol{\Sigma} \in \mathbb{R}^{3 \times 3}$ . The 3D covariance matrix $\boldsymbol{\Sigma}$ is derived from a learnable scaling vector $\mathbf{s} \in \mathbb{R}^3$ and rotation quaternion $\mathbf{q} \in \mathbb{R}^4$ , from which the scaling matrix $\mathbf{S} \in \mathbb{R}^{3 \times 3}$ and rotation matrix $\mathbf{R} \in$ + +$\mathbb{R}^{3\times 3}$ are obtained and represented as follows: + +$$ +\boldsymbol {\Sigma} = \mathbf {R} \mathbf {S} \mathbf {S} ^ {\top} \mathbf {R} ^ {\top}. \tag {2} +$$ + +For differentiable splatting [44], the Gaussians in the 3D world coordinate system are projected into the 2D camera coordinate system. This projection uses the viewing transformation $\mathbf{W} \in \mathbb{R}^{3 \times 3}$ and the Jacobian $\mathbf{J} \in \mathbb{R}^{2 \times 3}$ of the affine approximation of the projective transformation to derive the 2D covariance $\Sigma^{2\mathrm{D}} \in \mathbb{R}^{2 \times 2}$ : + +$$ +\boldsymbol {\Sigma} ^ {2 \mathrm {D}} = \mathbf {J} \mathbf {W} \boldsymbol {\Sigma} \mathbf {W} ^ {\top} \mathbf {J} ^ {\top}. \tag {3} +$$ + +Each Gaussian includes a set of spherical harmonics (SH) coefficients and an opacity value $\alpha$ to represent view-dependent color $\mathbf{c}$ . The pixel color $\mathbf{c}_p$ is then obtained by applying alpha blending to $\mathcal{N}$ ordered Gaussians: + +$$ +\mathbf {c} _ {p} = \sum_ {i \in \mathcal {N}} \mathbf {c} _ {i} \alpha_ {i} \prod_ {j = 1} ^ {i - 1} (1 - \alpha_ {j}). \tag {4} +$$ + +Our approach, as illustrated in Fig. 1, exploits 3DGS and aims to render defocused images by generating multiple Gaussians to form the CoC for each Gaussian. + +# 3.2. 3D Scene Blind Deblurring + +Image blind deblurring is a technique for learning an image blurring kernel without the supervision of sharp images. Blurry images are obtained by applying convolution with a learned blurring kernel to sharp images, where the kernel size is fixed as a grid structure at each pixel position. Inspired by this approach, Deblur-NeRF [18] applies this technique to NeRF, presenting a method to represent a clean 3D scene using only blurry images. Specifically, Deblur-NeRF warps the input ray into multiple rays that constitute the blur, modeling a ray-based sparse kernel. A blurry pixel is obtained by combining the pixel colors of each ray: + +$$ +\mathbf {c} _ {b l u r} = \sum w _ {p} ^ {i} \mathbf {c} _ {p} ^ {i}, w. r. t., \sum w _ {p} ^ {i} = 1, \tag {5} +$$ + +where $w_{p}$ and $\mathbf{c}_p$ denote the weight and pixel color for each corresponding ray. Deblur-NeRF enhances learning efficiency by setting the number of warped rays smaller than the kernel size of 2D convolution, and it designs a deformable kernel by adaptive origins and directions of rays. + +As we adopt a rasterization method based on 3DGS [9], we propose an alternative approach: instead of warping rays, we generate 3D Gaussians in the shape of the CoC and render them. In other words, the defocused image is estimated by setting the generated 3D Gaussians, namely CoC Gaussians, as the blurring kernel. + +# 4. Method + +# 4.1. CoCoGaussian Framework + +Our goal is to represent a 3D scene using only defocused multi-view images, leveraging fundamental photographic + +![](images/4169d199f407b6ec6e4cce324d5394acd0ea85e11d85e29f803b274456542d6c.jpg) +Figure 2. We take the camera position $\mathbf{x}_{cam}$ and the base Gaussian parameters $\mu_B$ , $\mathbf{s}_B$ , and $\mathbf{q}_B$ as inputs to the MLP $h_\theta$ , which produces five outputs in total. (a) We set the depth $d(\mu_B)$ as the Euclidean distance between $\mathbf{x}_{cam}$ and $\mu_B$ . Using the output $K$ from $h_\theta$ , the $d(\mu_B)$ and the learnable focus plane $d_F$ , we apply Eq. (7) to determine the CoC diameter. The diameter, combined with the outputs $\beta$ and $\mathbf{d}$ from $h_\theta$ , is then used in Eq. (9) to compute the offset values for $\mu_{\mathrm{CoC}}$ . By adding these offsets to $\mu_B$ , we obtain the $\mathbf{G}_{\mathrm{CoC}}$ means. (b) We obtain $\mathbf{s}_{\mathrm{CoC}}$ and $\mathbf{q}_{\mathrm{CoC}}$ by applying Eqs. (10) and (11) to $\delta \mathbf{s}_{\mathrm{CoC}}$ and $\delta \mathbf{q}_{\mathrm{CoC}}$ , the outputs of $h_\theta$ . (c) Using $\mu_{\mathrm{CoC}}$ , $\mathbf{s}_{\mathrm{CoC}}$ , and $\mathbf{q}_{\mathrm{CoC}}$ , we get the $\mathbf{G}_{\mathrm{CoC}}$ . Finally, we rasterize $\mathbf{G}_{\mathrm{CoC}}$ along with $\mathbf{G}_B$ to produce $(M + 1)$ images, and apply a weighted sum to generate the final defocused image. + +principles to guide our method. We generate 3D Gaussians to capture a blurring kernel shaped as the CoC. The whole framework of CoCoGaussian is on Fig. 2. First, we compute the CoC diameter using the Gaussian depth, derived from the means of the given base 3D Gaussian set $\mathbf{G}_B$ , which consists of $N$ Gaussians, along with the camera position and aperture information (Sec. 4.2). Next, for the base Gaussian set $\mathbf{G}_B$ , we generate $M$ CoC Gaussian sets $\mathbf{G}_{\mathrm{CoC}}$ , resulting in a total of $(M\times N)$ Gaussians to capture the CoC shape. However, when the scene contains objects with refraction or reflection, the Gaussian depth may be unreliable, which can lead to suboptimal CoC diameter. To address this issue, we propose an adaptive CoC Gaussian generation method that minimizes the dependence on depth for $\mathbf{G}_{\mathrm{CoC}}$ (Sec. 4.3). Finally, we incorporate adaptive rotation quaternion and scaling vector adjustments, inspired by the Deblurring 3DGS [10] approach, to allow flexible learning of $\mathbf{G}_{\mathrm{CoC}}$ . Furthermore, we apply Eq. (5) to images rendered from $\mathbf{G}_B$ and $\mathbf{G}_{\mathrm{CoC}}$ , using a weighted sum approach [12, 26] to create the final blurry image. This photography-prior-based approach allows us to accurately represent 3D scenes using only defocused images while adhering to realistic defocus principles. + +# 4.2. Circle of Confusion from 3D Gaussians + +To model the CoC, we assume an ideal system that ignores distortions such as lens aberrations. Before generating the CoC Gaussian sets $\mathbf{G}_{\mathrm{CoC}}$ , we first need to obtain the CoC diameter, which depends on several factors: the aperture diameter $D$ , focal length $f$ , focus plane $d_{F}$ , and depth $p$ . The focus plane represents the distance from the camera position to the plane in focus. We define the depth $d(\mu_B)$ as the Euclidean distance between the camera position $\mathbf{x}_{cam} \in \mathbb{R}^3$ and the means of the base Gaussian set $\mu_B \in \mathbb{R}^{N \times 3}$ . Exploiting these values, we express the CoC diameter $\sigma(\mu_B)$ as follows [7]: + +$$ +\sigma (\mu_ {B}) = f D \times \frac {\left| d (\mu_ {B}) - d _ {F} \right|}{d (\mu_ {B}) (d _ {F} - f)}, \tag {6} +$$ + +where $d_F$ is set to a learnable parameter and initialized as the average distance between $\mathbf{x}_{cam}$ and the points in the SfM point cloud to ensure stable training. Note that the focus plane $d_F$ is specific to each input image. Additionally, because the focal length $f$ is usually much smaller than the focus plane $d_F$ , we modify Eq. (6) as follows: + +$$ +\sigma \left(\mu_ {B}\right) \approx K \times \left| \frac {1}{d \left(\mu_ {B}\right)} - \frac {1}{d _ {F}} \right|, \tag {7} +$$ + +where the product of the focal length and aperture diameter is represented as a single learnable scalar $K = f \times D$ [41]. + +We design a simple MLP $h_\theta$ parameterized by $\theta$ to estimate the scalar $K$ . Then, the aperture parameter $K$ is obtained by passing these features as inputs to $h_\theta$ , along with the scaling vectors $\mathbf{s}_B \in \mathbb{R}^{N \times 3}$ and rotation quaternions $\mathbf{q}_B \in \mathbb{R}^{N \times 4}$ of $\mathbf{G}_B$ . This approach allows us to obtain the CoC diameter $\sigma(\mu_B)$ for the $\mathbf{G}_B$ , marking the first step in modeling $\mathbf{G}_{\mathrm{CoC}}$ . + +# 4.3. Adaptive CoC Gaussian Generation + +After obtaining the CoC diameters, we use these diameters to generate $M$ CoC Gaussian sets $\mathbf{G}_{\mathrm{CoC}}$ . For the Gaussian parameters of $\mathbf{G}_{\mathrm{CoC}}$ , we set the spherical harmonics (SH) coefficients and opacity values identical to those of $\mathbf{G}_B$ , while varying only the means $\mu$ , scaling vectors $\mathbf{s}$ , and rotation quaternions $\mathbf{q}$ . The means of each CoC Gaussian set are created by adding offsets $\Delta \mu_{\mathrm{CoC}} \in \mathbb{R}^{M \times N \times 3}$ to $\mu_B$ , defined as $\mu_{\mathrm{CoC}} = \mu_B + \Delta \mu_{\mathrm{CoC}}$ . Since $\Delta \mu_{\mathrm{CoC}}$ are 3D vector sets, we obtain unit vector sets $\mathbf{d} \in \mathbb{R}^{M \times N \times 3}$ representing the directions from $\mathbf{G}_B$ to $\mathbf{G}_{\mathrm{CoC}}$ , which are additional outputs from the MLP $h_\theta$ discussed in the previous section. Moreover, since $\mathbf{G}_{\mathrm{CoC}}$ consists of a total of $M$ Gaussian sets, The direction vectors $\mathbf{d}$ consist of $M$ instances: + +$$ +\Delta \mu_ {\mathrm {C o C}; m} = \frac {\sigma (\mu_ {B})}{2} \mathbf {d} _ {m}, \text {w h e r e} m \in M. \tag {8} +$$ + +However, theses offsets $\Delta \mu_{\mathrm{CoC}}$ presents two potential issues. First, this approach places all $\mathbf{G}_{\mathrm{CoC}}$ only on the boundary of the CoC, which limits its ability to fully capture the defocus effect. Modeling the entire CoC, including its interior, would allow for a more accurate and flexible representation of defocus blur. The second issue arises when there are refractive or reflective surfaces in the scene, as these subjects can lead to incorrect optimization of the Gaussian means, making the CoC diameter obtained through Eq. (7) unreliable. To address these problems, we introduce simple learnable CoC scaling parameters $\beta \in (0,1]$ , which are additional outputs of $h_\theta$ . The parameter $\beta$ is learned within the range of 0 to 1, ensuring that $\mu_{\mathrm{CoC}}$ resides within the CoC boundary. Additionally, even when depth is inaccurately measured and the predicted CoC diameter is larger than the actual size, $\beta$ adaptively scales it down. Thus, the offset $\Delta \mu_{\mathrm{CoC}}$ is modified by applying $\beta$ as follows: + +$$ +\Delta \mu_ {\mathrm {C o C}; m} = \frac {\sigma (\mu_ {B})}{2} \beta_ {m} \mathbf {d} _ {m}. \tag {9} +$$ + +Additionally, similar to Deblurring 3DGS [10], we introduce scaling factors $\delta \mathbf{s}_{\mathrm{CoC}} \in \mathbb{R}^{M \times N \times 3}$ and $\delta \mathbf{q}_{\mathrm{CoC}} \in \mathbb{R}^{M \times N \times 4}$ , which allow flexible learning of the scaling vectors $\mathbf{s}_{\mathrm{CoC}}$ and rotation quaternions $\mathbf{q}_{\mathrm{CoC}}$ of $\mathbf{G}_{\mathrm{CoC}}$ . These factors are also the outputs of $h_{\theta}$ . The parameters $\delta \mathbf{s}_{\mathrm{CoC}}$ and $\delta \mathbf{q}_{\mathrm{CoC}}$ are multiplied by the corresponding parameters + +of the base Gaussian set $\mathbf{G}_B$ , specifically $\mathbf{s}_B$ and $\mathbf{q}_B$ , and are expressed as follows: + +$$ +\mathbf {s} _ {\mathrm {C o C}; m} = \mathbf {s} _ {B} \times \delta \mathbf {s} _ {\mathrm {C o C}; m}, \tag {10} +$$ + +$$ +\mathbf {q} _ {\mathrm {C o C}; m} = \mathbf {q} _ {B} \times \delta \mathbf {q} _ {\mathrm {C o C}; m}, \tag {11} +$$ + +where the scaling parameters are constrained by the fixed values $\delta \mathbf{s}_{\mathrm{max}}$ and $\delta \mathbf{q}_{\mathrm{max}}$ : $\delta \mathbf{s}_{\mathrm{CoC}} \in [1, \delta \mathbf{s}_{\mathrm{max}}]$ and $\delta \mathbf{q}_{\mathrm{CoC}} \in [1, \delta \mathbf{q}_{\mathrm{max}}]$ . + +We obtain a total of $M$ means of CoC Gaussian sets $\mu_{\mathrm{CoC}}$ , scaling vectors $\mathbf{s}_{\mathrm{CoC}}$ , and rotation quaternion $\mathbf{q}_{\mathrm{CoC}}$ . They are combined with the opacity and SH coefficients of $\mathbf{G}_B$ , generate the $M$ Gaussian sets $\mathbf{G}_{\mathrm{CoC}}$ : + +$$ +\mathbf {x} _ {\mathrm {C o C}; m}, \mathbf {s} _ {\mathrm {C o C}; m}, \mathbf {q} _ {\mathrm {C o C}; m}, \alpha_ {B}, \mathbf {S H} _ {B} \rightarrow \mathbf {G} _ {\mathrm {C o C}; m}. \tag {12} +$$ + +We rasterize the resulting $M$ Gaussian sets along with the base Gaussian set, producing $(M + 1)$ images. Then, using the weighted sum approach from Eq. (5), as detailed in the following section, we obtain the final blurry image. + +# 4.4. Optimization + +Weighted Sum. To apply Eq. (5) to the $(M + 1)$ images $\mathbf{I}$ rasterized from $\mathbf{G}_B$ and $\mathbf{G}_{\mathrm{CoC}}$ , we adopt the methods of recent studies [12, 26]. We use a shallow CNN $\mathcal{F}$ to compute pixel-wise weights $\mathcal{W}$ for these images, applying a softmax function to ensure that the weights for each pixel sum to 1. Then, we multiply each image by its corresponding weight and sum the results to obtain the final defocused image $\mathcal{I}_{blur}$ : + +$$ +\mathcal {I} _ {\text {b l u r}} = \sum_ {m = 1} ^ {M + 1} \mathcal {I} _ {m} \mathcal {W} _ {m}, \text {w h e r e} \mathcal {W} = \operatorname {s o f t m a x} (\mathcal {F} (\mathbf {I})), \tag {13} +$$ + +where $\mathcal{I}_m$ represents the image rasterized from $\mathbf{G}_{\mathrm{CoC};m}$ , and $\mathcal{I}_{M + 1}$ represents the image rasterized from $\mathbf{G}_B$ . + +Objective. We minimize the $\mathcal{L}_1$ loss and D-SSIM loss $\mathcal{L}_{\mathrm{D - SSIM}}$ between the ground truth defocused image and the output defocused image. The $\mathcal{L}_1$ loss reduces the pixelwise differences between images, while the D-SSIM loss minimizes structural differences between them. The final objective $\mathcal{L}_{rgb}$ is defined as follows: + +$$ +\mathcal {L} _ {r g b} = (1 - \lambda) \mathcal {L} _ {1} + \lambda \mathcal {L} _ {\mathrm {D} - \mathrm {S S I M}}, \tag {14} +$$ + +where we use $\lambda = 0.3$ for all experiments. + +# 5. Experiments + +Datasets. We evaluate our method on three different datasets: the Deblur-NeRF [18] synthetic dataset, the Deblur-NeRF real-world dataset, and the DoF-NeRF [41] + +Table 1. Comparisons on Deblur-NeRF synthetic and real-world scene dataset. We evaluate the performance on three metrics (PSNR, SSIM, LPIPS). “*” denotes the results obtained by reproducing the released code. The orange and yellow cells respectively indicate the highest and second-highest value. + +
MethodsSynthetic Scene [18]Real-World Scene [18]
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
Naive NeRF [20]25.950.77910.230322.400.66610.2310
3DGS [9]25.110.74760.214823.380.66550.3140
Deblur-NeRF [18]28.370.85270.118823.470.71990.1207
PDRF-10 [25]30.080.89310.110123.850.73820.1746
DP-NeRF [11]29.330.87130.098723.670.72990.1082
Deblurring 3DGS* [10]28.900.89120.105223.540.73830.1232
BAGS* [26]30.650.91280.063123.480.74080.0962
Ours30.840.92120.047823.700.75310.0825
+ +Table 2. Comparisons on DoF-NeRF real-world scene dataset. + +
MethodsDoF-NeRF Real-World Scene [41]
PSNR↑SSIM↑LPIPS↓
3DGS [9]25.720.82910.1817
DP-NeRF [11]26.800.81450.1790
Deblurring 3DGS [10]26.580.83860.1708
BAGS [26]29.870.88160.1100
Ours30.140.91270.0701
+ +real-world dataset. The Deblur-NeRF synthetic dataset consists of 5 scenes, each containing defocused images generated using the built-in function of Blender [5]. The Deblur-NeRF real-world dataset includes 10 scenes, with images captured by enlarging the aperture on a Canon EOS RP. The DoF-NeRF real-world dataset comprises 7 scenes, each containing 20 to 30 image triplets. Each triplet includes an all-in-focus image, along with two defocused images focused on the background and foreground. Since DoF-NeRF does not provide training code, we use half of the training defocused images with a foreground focus and the other half with a background focus. All camera poses and initial point clouds are obtained through COLMAP [29, 30]. + +# 5.1. Rendering Results + +Quantitative Results. We quantitatively evaluate our CoCoGaussian using the following three metrics: peak signal-to-noise ratio (PSNR), structural similarity index measure (SSIM) [38], and learned perceptual image patch similarity (LPIPS) [45]. We compare our approach to several state-of-the-art methods, including both raytracing [11, 18, 25] and rasterization-based methods [10, 26]. Quantitative results are shown in Tab. 1 and Tab. 2, where our approach achieves the best performance across all metrics except for PSNR on the Deblur-NeRF real-world + +dataset. Note that we report the reproduced performance scores for Deblurring 3DGS [10] and BAGS [26] using their official codes. The Deblur-NeRF real-world dataset inherently presents challenges, as there are illumination differences between the sharp and defocused images, leading to relatively lower PSNR score. We discuss this issue in the appendix. For the DoF-NeRF dataset, CoCoGaussian achieves outstanding performance across all metrics, with an approximately 0.04 reduction in LPIPS compared to BAGS. Since the DoF-NeRF dataset has a higher resolution than the Deblur-NeRF dataset, this performance drop in BAGS is likely due to its requirement for manual adjustment of blurring kernel size, highlighting an intrinsic limitation. Conversely, Deblurring 3DGS shows minimal improvement over 3DGS, which can also be attributed to its limited kernel size relative to the higher resolution. The details regarding the time complexity of our model are provided in the appendix. + +Qualitative Results. For qualitative evaluation, we present the rendering results of 3DGS [9], Deblurring 3DGS [10], and BAGS [26] in Fig. 3. Our method demonstrates higher-fidelity results compared to others. Specifically, in the fourth row of our results, the white line on the right side is rendered more sharply, and in the fifth row, the detailed quality of the graphic card appears superior to that of BAGS. Additionally, more CoC visualizations similar to Fig. 1 (b), are in the appendix. + +# 5.2. Ablation Study + +To validate the effectiveness of our proposed methods, we perform a series of ablation studies, with all experiments conducted on the DoF-NeRF real-world dataset. We chose this dataset as it is based on real captured images and does not present the illumination issues found in the Deblur-NeRF real-world dataset, ensuring more reliable evaluation. The results are summarized in Tab. 3, where the baseline denotes naive 3DGS [9]. Additional ablative experiments with other factors are in the appendix. + +Circle of Confusion. To demonstrate the impact of our core concept, CoC, we conduct an experiment that excludes the CoC from the model, retaining only $\beta_{m}$ and $\mathbf{d}_m$ . This configuration yields significantly lower performance than the full model, indicating an over-reliance on parametric learning. The result suggests that without the CoC, which leverages depth and aperture information, the means of the generated Gaussian sets are not optimally arranged, leading to overfitting. + +CoC Direction Vector. In this experiment, we retain the CoC concept but exclude the learnable direction vectors + +![](images/9cb75fc5bbc895292c428596b4c164354427d742a8147dbd1ed06ad66a7319a5.jpg) + +![](images/8b6a4f88c0374675cac50e3bfca5115bdccef49500438b090836a1550bb2eb6f.jpg) + +![](images/65dc56d33dd4b64ad77759b39f75049b420ea498977ce01edf43bfaf674f05d3.jpg) + +![](images/e1b3d16adc142ec22e3f5ee7c54d10479dc38d17952d9bad8c813afc3a51beff.jpg) + +![](images/6794c98aa1a6a9a82c103519bcad327ab99e081d3a317e9d74de259022f06f00.jpg) + +![](images/557b78154bcfc77d4c2a283206ef51104a693a1d2b25ae2086605bffda617f57.jpg) + +![](images/ab9e9b898833f7d861a78f8560d9f092cfc5341b88af731176724cfccc0b4a35.jpg) + +![](images/7d7ef3d8dfd972cd5742c3257e390f318020b206d3f3aff185abd7c86ef54233.jpg) + +![](images/44b72f0dab6c517d5c47e13e262da96138a7795bc3c7cc1a9d7418fbc9a07164.jpg) + +![](images/ab67a5fa763d68ad2c7c96a4587da0f5e7e883034440bf36dae723c1c18aec78.jpg) + +![](images/894068c4173df7c1af6ada250a64b63b4fa2855e097b2db1edc5e5e4238440b1.jpg) + +![](images/053df0d43e802824196a3ed7271d0914b67c30c735b137562cf7724b6f796428.jpg) + +![](images/c223c05e7a7216926ee1a9c3526b5daa005c1677ee8c7f6c02bb0c541755c8b9.jpg) + +![](images/f95eb3e7ef4c3776bd9fa2afc9644accf8594658601b64a88cb385bb05f5128a.jpg) + +![](images/7a8c55cdd317cc31b6f5ad765767179cd8c2c38570de47482ffe16983c80338a.jpg) + +![](images/06777173a890b5b4c65c60d3ec37a450cd2a29cb3472edac5a169bf9b801aeab.jpg) + +![](images/7e568ed2bfd40d10661d49024ae43969507bd25150670c5b66977c4e6bfaef0e.jpg) + +![](images/2a9b2ed0bc8edafdb6fe745809757e19ad83d513b8268a1428faf410426bcd46.jpg) + +![](images/54afc9bc38c4814c92f333c69f4d1431b9e6437de2d3cde6a4048556412c4b51.jpg) + +![](images/2344fb31ed42130a2ecb419b1f16dea64c5648d8f37916789bbc87453b0d106a.jpg) + +![](images/a01d3ec93c34a08ce1eb4ab47d979f1121c887b80d99ffc0c79f148c90c03fa8.jpg) + +![](images/fe7f4007f721a1d0a96d29153f7b1fa20da7e65dc14ac4a4a48aceca82da3f2a.jpg) + +![](images/7e956a745bfe1aac5440edbf2cb366c122b1995d2175feb2a6235e8e65d58d0a.jpg) + +![](images/a42e95ff77ea49b01929a4d24b052e1d4a1fef45e4d95d21016c9c99ed1cf938.jpg) + +![](images/38339c5098d520d020d8df9c6030e4692d2574151d39297acdfe39fd3e6f46eb.jpg) + +![](images/0c0ee419b5ed6fa92e8b5f1f375840694022385dc58916864d4dedb7458178a5.jpg) + +![](images/3eec4c0f8400d57ddd6011d77355fbebed540f77af64564e9b1b805d41aea3be.jpg) + +![](images/11b31f0ca21e8765fb539b0dbff3f65261b5764445b1fdf3c952e7ee7158d901.jpg) + +![](images/a7f313060dbd9e68038eb922a9396478fdf13218bff2f8c1aae4e4446b14eb8a.jpg) + +![](images/e6aeafa7bf2c79d8edb837a2f55eac89efd0e259f1a5eeb26a83c04041d8161d.jpg) + +![](images/365029911a67e00921e80b9a067c20f4b7242d200cc03b355e07ee4717cb7fe6.jpg) +Reference + +![](images/0fc3a7f0655d52d3b1befad0954f9806fb9a6e5aa900c78880a896a8269d5057.jpg) +3DGS + +![](images/56be89c78cd25de733c32f6ec3046914c77e49f580c82ccb39226f16a8f75ce5.jpg) +Deblurring 3DGS + +![](images/ba88978bbb98bc401526f99df6b474572ccdb06b74a87ebc9f255b229c7bf6f1.jpg) +BAGS + +![](images/5b0b786477b09fff51e80bdbf6bf408e683240f8ceccb35c26320049e774cda9.jpg) +Ours +Figure 3. Qualitative comparison on the Deblur-NeRF and DoF-NeRF datasets. + +![](images/015d01065fff52da18631ddc2ed73d69933657e4eaa95ba3470a89cfffef32c7.jpg) +Ground Truth + +$\mathbf{d}_m$ . Instead of using learnable vectors, we use fixed direction vectors, arranging $M$ CoC Gaussians evenly in a circular pattern on a plane perpendicular to the vector between the camera position and the base Gaussian. This configuration slightly underperforms compared to the full model, as the fixed directions restrict optimal Gaussian position, preventing adaptation to the most reliable configuration. + +CoC Scaling Factor. In this setup, we include the learnable direction vector while omitting the CoC scaling factor $\beta_{m}$ . Although the performance is slightly better than when the direction vector is excluded, it remains lower than that of the full model. Without the CoC scaling factor, CoC Gaussian generation becomes overly dependent on depth, leading to errors in CoC diameter calculation when depth values are inaccurate. Including the CoC scaling factor enables the full model to handle uncertain depth more robustly, improving performance. Related visualization re + +sults are in the appendix. + +Aperture Parameter. The aperture parameter $K$ includes the aperture diameter and focal length, and it plays a critical role in determining the size of the CoC. Without this parameter, there is an upper limit to the CoC diameter calculation, making accurate computation challenging. Experimental results indicate that although performance is somewhat lower than the full model, it is improved over the baseline. This suggests that while the model is unable to determine precise CoC size without $K$ , it still learns the focus plane and captures the in-focus region accurately. + +# 5.3. Experiments on All-in-Focus Images. + +We conduct novel view synthesis experiments using all-in-focus images to verify the generalization performance of CoCoGaussian. We utilize the NeRF-LLFF [19, 20] dataset, which contains real-world images. As shown in Tab. 4, our + +![](images/40134f8f7422f6029deea3284d5dfc03cf87ca4d59036fc79d378a9ed595f953.jpg) +Figure 4. Visualization of Aperture parameter and Focus Plane Customization. The top row of images decreases the aperture parameter $K$ from left to right, while the bottom row moves the focus plane $d_{F}$ further from the camera from left to right. + +Table 3. Ablation Results of CoCoGaussian. + +
MethodsPSNR↑SSIM↑LPIPS↓
Baseline [9]25.720.82910.1817
w/o CoC28.290.87780.0927
w/o CoC direction vector28.910.88930.0896
w/o CoC scaling29.460.90580.0793
w/o aperture parameter27.370.85100.1113
Ours Full30.140.91270.0701
+ +approach outperforms the baseline 3DGS [9] across all metrics. In practice, when the aperture is very small, radiance from a point on the subject forms a tiny CoC on the image sensor after passing through the aperture, even if the subject is slightly away from the focus plane. Because the CoC is so small, the resulting image is not perceived as defocused. By modeling this small circle, which is difficult to detect with the human eye, our approach achieves higher performance than the baseline 3DGS. Related CoC visualizations are in the appendix. + +# 5.4.Customizable Depth of Field and Focus Plane + +Customizing the depth of field and focus settings is essential in various applications, such as augmented reality (AR) and virtual reality (VR), where control over depth cues and focus effects can greatly enhance visual realism and viewer engagement. Since our approach incorporates the aperture parameter $K$ in CoC modeling, it enables customization of both CoC size and depth of field by adjusting this parameter. The top row in Fig. 4 visualizes increasing depth of field by gradually reducing $K$ from left to right. As the focus plane is positioned at a shallow depth, the model consistently captures focus in this region. However, in the leftmost image, the large aperture causes the CoC size to increase at deeper depths, resulting in defocus blur. + +Additionally, our model can customize the focus plane of defocused images, as shown in the bottom row of Fig. 4, where the focus plane shifts from shallow depth on the left to deeper depth on the right. The shallow regions are in fo + +Table 4. Results on NeRF-LLFF [19, 20] Dataset. + +
MethodsPSNR↑SSIM↑LPIPS↓
3DGS [9]27.100.86190.0569
Ours27.760.87380.0372
+ +cus on the leftmost image, while deeper areas remain out of focus, and the opposite holds for the rightmost image. This level of control over focus and depth effects is essential for achieving realistic depth cues. Therefore, CoCoGaussian not only demonstrates superior capability in synthesizing sharp novel view images but also provides enhanced flexibility in adjusting aperture and focus plane settings. + +# 6. Limitation and Future Work + +While our model demonstrates strong qualitative and quantitative performance, there is potential for further optimization in the adaptive CoC scaling factor. To ensure training stability, we currently limit the factor size, enabling adaptive optimization mainly when the CoC diameter is overestimated. However, future improvements could enhance adaptability to handle cases where the CoC diameter is underestimated, further improving depth handling in complex scenes without compromising stability. + +# 7. Conclusion + +In this work, we propose CoCoGaussian, a 3D scene representation method that effectively leverages photographic defocus principles to handle defocused image inputs. CoCoGaussian accurately models defocus blur by constructing the CoC through 3D Gaussians and learnable aperture parameters. We introduce an adaptive CoC Gaussian generation method to address the challenges posed by uncertain depth, making our model robust to reflective or refractive objects. Our approach demonstrates the feasibility and effectiveness of representing 3D scenes from defocused images, advancing possibilities for applications in real-world image synthesis. + +Acknowledgements. This project was supported by the Naver Cloud Corporation and the National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIT) (No. RS-2024-00340745). + +# References + +[1] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5855–5864, 2021. 3 +[2] Ayan Chakrabarti. A neural approach to blind motion deblurring. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part III 14, pages 221-235. Springer, 2016. 2, 3 +[3] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In European Conference on Computer Vision, pages 333-350. Springer, 2022. 2, 3 +[4] Wenbo Chen and Ligang Liu. Deblur-gs: 3d gaussian splattering from camera motion blurred images. Proceedings of the ACM on Computer Graphics and Interactive Techniques, 7 (1):1-15, 2024. 3 +[5] Blender Online Community. Blender - a 3D modelling and rendering package. Blender Foundation, Stichting Blender Foundation, Amsterdam, 2018. 6 +[6] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5501–5510, 2022. 2 +[7] Eugene Hecht. Optics. Pearson Education India, 2012. 2, 4 +[8] Wei Jiang, Kwang Moo Yi, Golnoosh Samei, Oncel Tuzel, and Anurag Ranjan. Neuman: Neural human radiance field from a single video. In European Conference on Computer Vision, pages 402-418. Springer, 2022. 3 +[9] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42 (4):1-14, 2023. 2, 3, 6, 8 +[10] Byeonghyeon Lee, Howoong Lee, Xiangyu Sun, Usman Ali, and Eunbyung Park. Deblurring 3d gaussian splatting. arXiv preprint arXiv:2401.00834, 2024. 2, 3, 4, 5, 6 +[11] Dogyoon Lee, Minhyeok Lee, Chajin Shin, and Sangyoun Lee. Dp-nerf: Deblurred neural radiance field with physical scene priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12386–12396, 2023. 3, 6 +[12] Junghe Lee, Donghyeong Kim, Dogyoon Lee, Suhwan Cho, and Sangyoun Lee. Crime: Continuous rigid motion-aware gaussian splatting from motion blur images. arXiv preprint arXiv:2407.03923, 2024. 4, 5 +[13] Jungho Lee, Dogyoon Lee, Minhyeok Lee, Donghyung Kim, and Sangyoun Lee. Smurf: Continuous dynam- + +ics for motion-deblurring radiance fields. arXiv preprint arXiv:2403.07547, 2024. 2, 3 +[14] Tianye Li, Mira Slavcheva, Michael Zollhoefer, Simon Green, Christoph Lassner, Changil Kim, Tanner Schmidt, Steven Lovegrove, Michael Goesele, Richard Newcombe, et al. Neural 3d video synthesis from multi-view video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5521-5531, 2022. 3 +[15] Zhengqi Li, Simon Niklaus, Noah Snavely, and Oliver Wang. Neural scene flow fields for space-time view synthesis of dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6498-6508, 2021. 3 +[16] Zhaoshuo Li, Thomas Müller, Alex Evans, Russell H Taylor, Mathias Unberath, Ming-Yu Liu, and Chen-Hsuan Lin. Neuralangelo: High-fidelity neural surface reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8456-8465, 2023. 3 +[17] Kevin M Lynch and Frank C Park. Modern robotics. Cambridge University Press, 2017. 3 +[18] Li Ma, Xiaoyu Li, Jing Liao, Qi Zhang, Xuan Wang, Jue Wang, and Pedro V Sander. Deblur-nerf: Neural radiance fields from blurry images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12861–12870, 2022. 2, 3, 5, 6 +[19] Ben Mildenhall, Pratul P Srinivasan, Rodrigo Ortiz-Cayon, Nima Khademi Kalantari, Ravi Ramamoorthi, Ren Ng, and Abhishek Kar. Local light field fusion: Practical view synthesis with prescriptive sampling guidelines. ACM Transactions on Graphics (ToG), 38(4):1-14, 2019. 7, 8 +[20] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I, pages 405-421, 2020. 1, 2, 3, 6, 7, 8 +[21] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG), 41(4):1-15, 2022. 2 +[22] Michael Niemeyer, Jonathan T Barron, Ben Mildenhall, Mehdi SM Sajjadi, Andreas Geiger, and Noha Radwan. Regnerf: Regularizing neural radiance fields for view synthesis from sparse inputs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5480-5490, 2022. 3 +[23] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5865-5874, 2021. 3 +[24] Keunhong Park, Utkarsh Sinha, Peter Hedman, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Ricardo MartinBrualla, and Steven M Seitz. Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields. arXiv preprint arXiv:2106.13228, 2021. 3 +[25] Cheng Peng and Rama Chellappa. Pdrf: progressively deblurring radiance field for fast scene reconstruction from + +blurry images. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2029-2037, 2023. 2, 3, 6 +[26] Cheng Peng, Yutao Tang, Yifan Zhou, Nengyu Wang, Xijun Liu, Deming Li, and Rama Chellappa. Bags: Blur agnostic gaussian splatting through multi-scale kernel modeling. arXiv preprint arXiv:2403.04926, 2024. 2, 3, 4, 5, 6 +[27] Sida Peng, Junting Dong, Qianqian Wang, Shangzhan Zhang, Qing Shuai, Xiaowei Zhou, and Hujun Bao. Animatable neural radiance fields for modeling dynamic human bodies. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14314-14323, 2021. 3 +[28] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-nerf: Neural radiance fields for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10318–10327, 2021. 3 +[29] Johannes L Schonberger, Enliang Zheng, Jan-Michael Frahm, and Marc Pollefeys. Pixelwise view selection for unstructured multi-view stereo. In European conference on computer vision, pages 501-518. Springer, 2016. 3, 6 +[30] Qi Shan, Jiaya Jia, and Aseem Agarwala. High-quality motion deblurring from a single image. Acm transactions on graphics (tog), 27(3):1-10, 2008. 3, 6 +[31] Pratul P Srinivasan, Ren Ng, and Ravi Ramamoorthi. Light field blind motion deblurring. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3958-3966, 2017. 2, 3 +[32] Jiaming Sun, Yiming Xie, Linghao Chen, Xiaowei Zhou, and Hujun Bao. Neuralrecon: Real-time coherent 3d reconstruction from monocular video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15598-15607, 2021. 3 +[33] Edgar Tretschk, Ayush Tewari, Vladislav Golyanik, Michael Zollhöfer, Christoph Lassner, and Christian Theobalt. Non-rigid neural radiance fields: Reconstruction and novel view synthesis of a dynamic scene from monocular video. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12959-12970, 2021. 3 +[34] Guangcong Wang, Zhaoxi Chen, Chen Change Loy, and Zwei Liu. Sparsenerf: Distilling depth ranking for few-shot novel view synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9065-9076, 2023. 3 +[35] Jiepeng Wang, Peng Wang, Xiaoxiao Long, Christian Theobalt, Taku Komura, Lingjie Liu, and Wenping Wang. Neuris: Neural reconstruction of indoor scenes using normal priors. In European Conference on Computer Vision, pages 139-155. Springer, 2022. 3 +[36] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. arXiv preprint arXiv:2106.10689, 2021. 3 +[37] Peng Wang, Lingzhe Zhao, Ruijie Ma, and Peidong Liu. Bad-nerf: Bundle adjusted deblur neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4170–4179, 2023. 2, 3 + +[38] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 6 +[39] Chung-Yi Weng, Brian Curless, Pratul P Srinivasan, Jonathan T Barron, and Ira Kemelmacher-Shlizerman. Humanerf: Free-viewpoint rendering of moving people from monocular video. In Proceedings of the IEEE/CVF conference on computer vision and pattern Recognition, pages 16210-16220, 2022. 3 +[40] Oliver Whyte, Josef Sivic, Andrew Zisserman, and Jean Ponce. Non-uniform deblurring for shaken images. International journal of computer vision, 98:168-186, 2012. 2, 3 +[41] Zijin Wu, Xingyi Li, Juewen Peng, Hao Lu, Zhiguo Cao, and Weicai Zhong. Dof-nerf: Depth-of-field meets neural radiance fields. In Proceedings of the 30th ACM International Conference on Multimedia, pages 1718–1729, 2022. 2, 3, 5, 6 +[42] Jamie Wynn and Daniyar Turmukhambetov. Diffusionerf: Regularizing neural radiance fields with denoising diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4180-4189, 2023. 3 +[43] Jiawei Yang, Marco Pavone, and Yue Wang. Freenerf: Improving few-shot neural rendering with free frequency regularization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8254-8263, 2023. 3 +[44] Wang Yifan, Felice Serena, Shihao Wu, Cengiz Özireli, and Olga Sorkine-Hornung. Differentiable surface splatting for point-based geometry processing. ACM Transactions on Graphics (TOG), 38(6):1-14, 2019. 3 +[45] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 6 +[46] Lingzhe Zhao, Peng Wang, and Peidong Liu. Bad-gaussians: Bundle adjusted deblur gaussian splatting. arXiv preprint arXiv:2403.11831, 2024. 3 \ No newline at end of file diff --git a/2025/CoCoGaussian_ Leveraging Circle of Confusion for Gaussian Splatting from Defocused Images/images.zip b/2025/CoCoGaussian_ Leveraging Circle of Confusion for Gaussian Splatting from Defocused Images/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..7c88a999084d5f92b78775d9f2eaf413edfcc81d --- /dev/null +++ b/2025/CoCoGaussian_ Leveraging Circle of Confusion for Gaussian Splatting from Defocused Images/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5e21a0fbbaadd7df6e60aa6e0f3066d9470f32ca79ce4dad12b4778f5c882362 +size 690130 diff --git a/2025/CoCoGaussian_ Leveraging Circle of Confusion for Gaussian Splatting from Defocused Images/layout.json b/2025/CoCoGaussian_ Leveraging Circle of Confusion for Gaussian Splatting from Defocused Images/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..9775da65c8ec078656391b7b733b8454ce275182 --- /dev/null +++ b/2025/CoCoGaussian_ Leveraging Circle of Confusion for Gaussian Splatting from Defocused Images/layout.json @@ -0,0 +1,11688 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 146, + 103, + 465, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 103, + 465, + 140 + ], + "spans": [ + { + "bbox": [ + 146, + 103, + 465, + 140 + ], + "type": "text", + "content": "CoCoGaussian: Leveraging Circle of Confusion for Gaussian Splatting from Defocused Images" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 120, + 155, + 489, + 184 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 120, + 155, + 489, + 184 + ], + "spans": [ + { + "bbox": [ + 120, + 155, + 489, + 184 + ], + "type": "text", + "content": "Jungho Lee" + }, + { + "bbox": [ + 120, + 155, + 489, + 184 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 120, + 155, + 489, + 184 + ], + "type": "text", + "content": " Suhwan Cho" + }, + { + "bbox": [ + 120, + 155, + 489, + 184 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 120, + 155, + 489, + 184 + ], + "type": "text", + "content": " Taeoh Kim" + }, + { + "bbox": [ + 120, + 155, + 489, + 184 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 120, + 155, + 489, + 184 + ], + "type": "text", + "content": " Ho-Deok Jang" + }, + { + "bbox": [ + 120, + 155, + 489, + 184 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 120, + 155, + 489, + 184 + ], + "type": "text", + "content": " Minhyeok Lee" + }, + { + "bbox": [ + 120, + 155, + 489, + 184 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 120, + 155, + 489, + 184 + ], + "type": "text", + "content": " Geonho Cha" + }, + { + "bbox": [ + 120, + 155, + 489, + 184 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 120, + 155, + 489, + 184 + ], + "type": "text", + "content": " Dongyoon Wee" + }, + { + "bbox": [ + 120, + 155, + 489, + 184 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 120, + 155, + 489, + 184 + ], + "type": "text", + "content": " Dogyoon Lee" + }, + { + "bbox": [ + 120, + 155, + 489, + 184 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 120, + 155, + 489, + 184 + ], + "type": "text", + "content": " Sangyoun Lee" + }, + { + "bbox": [ + 120, + 155, + 489, + 184 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 112, + 189, + 499, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 112, + 189, + 499, + 217 + ], + "spans": [ + { + "bbox": [ + 112, + 189, + 499, + 217 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 112, + 189, + 499, + 217 + ], + "type": "text", + "content": "School of Electrical and Electronic Engineering, Yonsei University " + }, + { + "bbox": [ + 112, + 189, + 499, + 217 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 112, + 189, + 499, + 217 + ], + "type": "text", + "content": "NAVER Cloud https://Jho-Yonsei.github.io/CoCoGaussian" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 79, + 225, + 217, + 415 + ], + "blocks": [ + { + "bbox": [ + 79, + 225, + 217, + 415 + ], + "lines": [ + { + "bbox": [ + 79, + 225, + 217, + 415 + ], + "spans": [ + { + "bbox": [ + 79, + 225, + 217, + 415 + ], + "type": "image", + "image_path": "6372921ee99eef124a57bdb275cb484704d28be09f6167b5e9ae6d6aab16c7fc.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 84, + 417, + 189, + 427 + ], + "lines": [ + { + "bbox": [ + 84, + 417, + 189, + 427 + ], + "spans": [ + { + "bbox": [ + 84, + 417, + 189, + 427 + ], + "type": "text", + "content": "(a) Defocused Image Acquisition" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 228, + 227, + 362, + 317 + ], + "blocks": [ + { + "bbox": [ + 228, + 227, + 362, + 317 + ], + "lines": [ + { + "bbox": [ + 228, + 227, + 362, + 317 + ], + "spans": [ + { + "bbox": [ + 228, + 227, + 362, + 317 + ], + "type": "image", + "image_path": "209813a7d770da27e57b642b8969f2261761f3a17b2fb0a2ec7343a2402da039.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 229, + 326, + 362, + 414 + ], + "blocks": [ + { + "bbox": [ + 229, + 326, + 362, + 414 + ], + "lines": [ + { + "bbox": [ + 229, + 326, + 362, + 414 + ], + "spans": [ + { + "bbox": [ + 229, + 326, + 362, + 414 + ], + "type": "image", + "image_path": "8365538702c8e3e10961f876106bfabd8df9d0464e33d1cef9a46205c55397c3.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 241, + 417, + 350, + 426 + ], + "lines": [ + { + "bbox": [ + 241, + 417, + 350, + 426 + ], + "spans": [ + { + "bbox": [ + 241, + 417, + 350, + 426 + ], + "type": "text", + "content": "(b) Reconstructed Gaussian Points" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 369, + 256, + 388, + 289 + ], + "blocks": [ + { + "bbox": [ + 369, + 256, + 388, + 289 + ], + "lines": [ + { + "bbox": [ + 369, + 256, + 388, + 289 + ], + "spans": [ + { + "bbox": [ + 369, + 256, + 388, + 289 + ], + "type": "image", + "image_path": "972cfdd3b13a32857f3418c0a1d90d6d6e2da4e6a0362463315c24f998ed614d.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 394, + 228, + 531, + 321 + ], + "blocks": [ + { + "bbox": [ + 394, + 228, + 531, + 321 + ], + "lines": [ + { + "bbox": [ + 394, + 228, + 531, + 321 + ], + "spans": [ + { + "bbox": [ + 394, + 228, + 531, + 321 + ], + "type": "image", + "image_path": "a0dfa3ea76ab11e9d7b02e1ac7d0a8ead6cf0eb7294973e9e2ff490e930b63fe.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 394, + 325, + 530, + 415 + ], + "blocks": [ + { + "bbox": [ + 394, + 325, + 530, + 415 + ], + "lines": [ + { + "bbox": [ + 394, + 325, + 530, + 415 + ], + "spans": [ + { + "bbox": [ + 394, + 325, + 530, + 415 + ], + "type": "image", + "image_path": "fdb771382f32b5abcb68f80cc05014b4105a772e4f61b24dfecad55370309602.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 430, + 417, + 496, + 427 + ], + "lines": [ + { + "bbox": [ + 430, + 417, + 496, + 427 + ], + "spans": [ + { + "bbox": [ + 430, + 417, + 496, + 427 + ], + "type": "text", + "content": "(c) Rendered Images" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 54, + 430, + 555, + 498 + ], + "lines": [ + { + "bbox": [ + 54, + 430, + 555, + 498 + ], + "spans": [ + { + "bbox": [ + 54, + 430, + 555, + 498 + ], + "type": "text", + "content": "Figure 1. We propose CoCoGaussian, a novel framework for 3D scene reconstruction from defocused images. (a) With a large aperture, radiance from the focus plane appears as a small circle of confusion (green) on the image sensor, while radiance from greater depths results in larger circles (pink). (b) We visualize reconstructed Gaussians of CoCoGaussian, where the black dots indicate Gaussian means, while the gray dots represent Gaussian means forming the circle of confusion, which decreases in size near the focus plane. (c) The defocused image focuses on shallow depths, with circle size increasing with depth, as shown in (b). CoCoGaussian allows customization of defocused images by adjusting the depth of field or focus plane, while sharp images can be obtained by rendering without the circle of confusion." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 151, + 506, + 200, + 519 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 506, + 200, + 519 + ], + "spans": [ + { + "bbox": [ + 151, + 506, + 200, + 519 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 54, + 533, + 297, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 533, + 297, + 689 + ], + "spans": [ + { + "bbox": [ + 54, + 533, + 297, + 689 + ], + "type": "text", + "content": "3D Gaussian Splatting (3DGS) has attracted significant attention for its high-quality novel view rendering, inspiring research to address real-world challenges. While conventional methods depend on sharp images for accurate scene reconstruction, real-world scenarios are often affected by defocus blur due to finite depth of field, making it essential to account for realistic 3D scene representation. In this study, we propose CoCoGaussian, a Circle of Confusion-aware Gaussian Splatting that enables precise 3D scene representation using only defocused images. CoCoGaussian addresses the challenge of defocus blur by modeling the Circle of Confusion (CoC) through a physically grounded approach based on the principles of photo" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 508, + 556, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 508, + 556, + 617 + ], + "spans": [ + { + "bbox": [ + 313, + 508, + 556, + 617 + ], + "type": "text", + "content": "graphic defocus. Exploiting 3D Gaussians, we compute the CoC diameter from depth and learnable aperture information, generating multiple Gaussians to precisely capture the CoC shape. Furthermore, we introduce a learnable scaling factor to enhance robustness and provide more flexibility in handling unreliable depth in scenes with reflective or refractive surfaces. Experiments on both synthetic and real-world datasets demonstrate that CoCoGaussian achieves state-of-the-art performance across multiple benchmarks." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 314, + 643, + 395, + 656 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 643, + 395, + 656 + ], + "spans": [ + { + "bbox": [ + 314, + 643, + 395, + 656 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "content": "The emergence of the Neural Radiance Field (NeRF) [20] has brought significant attention to 3D scene representation for photo-realistic novel view synthesis. They reconstruct 3D scenes from images captured from multiple views" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 702, + 264, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 702, + 264, + 714 + ], + "spans": [ + { + "bbox": [ + 66, + 702, + 264, + 714 + ], + "type": "text", + "content": "*This work was done during an internship at Naver Cloud." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16101" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 192 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 192 + ], + "type": "text", + "content": "and render images of unseen views. However, since NeRF models 3D scenes as implicit neural representations through a ray tracing-based approach, they suffer from inefficient memory usage. In contrast, a rasterization-based method, 3D Gaussian Splatting (3DGS) [9] explicitly models 3D scenes using independent 3D Gaussians with different 3D means, covariances, opacities, and spherical harmonic coefficients. The tile-based rasterizer of 3DGS applies alpha blending to Gaussian splats sorted by visibility order, enabling fast training, rendering, and efficient memory usage." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 194, + 296, + 505 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 194, + 296, + 505 + ], + "spans": [ + { + "bbox": [ + 56, + 194, + 296, + 505 + ], + "type": "text", + "content": "Both NeRF [20] and 3DGS [9] rely on sharp images to represent 3D scenes accurately, which is a highly ideal assumption. In real-world settings, various factors (e.g., finite depth of field and camera motion blur) can hinder the capture of sharp images, leading to image degradation. Achieving a perfectly sharp image requires every part of the scene to be in focus, namely all-in-focus, which calls for a large depth of field. However, obtaining a large depth of field necessitates using a small aperture, which in turn requires a longer exposure time to allow enough light to enter the camera. During this extended exposure, even slight movements of a handheld camera can introduce motion blur. To prevent motion blur, the aperture needs to be widened, but this comes at the cost of a shallower depth of field. As a result, areas outside the focus plane appear blurred due to defocus. Defocus blur is closely related to the concept of the Circle of Confusion (CoC); as shown in Fig. 1, when a subject is positioned away from the focus plane, the radiances from the subject pass through the aperture and are mapped onto the image sensor in a circular pattern, resulting in shallow depth of field and defocus blur [7]. However, despite these challenges, recent research in 3D scene representation has rarely focused on addressing defocus blur directly. In this paper, we propose a method to represent 3D scenes using only images with defocus blur, tackling the CoC-related challenges posed by real-world image acquisition scenarios." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 508, + 295, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 508, + 295, + 687 + ], + "spans": [ + { + "bbox": [ + 55, + 508, + 295, + 687 + ], + "type": "text", + "content": "Recently, there has been increasing interest in reconstructing 3D scenes using only degraded images, with a focus on addressing challenges caused by camera motion blur and defocus blur. Deblur-NeRF [18] is the first study to take on this task, adopting a blind image deblurring approaches [2, 31, 40]. Follow-up studies [10-13, 25, 26, 37] have expanded this approach beyond raytracing to rasterization-based methods. However, many of these methods depend heavily on learning-based strategies and fail to incorporate photographic principles that accurately capture how defocus occurs in real-world scenarios. While DoF-NeRF [41] adopts the concept of CoC to model defocus blur in 3D scene representations, it relies on an implicit representation, which leads to slow rendering speeds, making it less suitable for real-time applications." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": "In this paper, we propose CoCoGaussian, which leverages a physically grounded photographic defocus princi" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 553, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 323 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 323 + ], + "type": "text", + "content": "plies to represent 3D scenes. Our method incorporates an aperture to calculate the CoC diameter based on depth and learnable aperture information using 3D Gaussians, accurately modeling shallow depth of field images. Using the CoC diameter and the 3D Gaussian, we generate multiple Gaussians to form the CoC shape, as illustrated in Fig. 1. However, when objects with refraction or reflection exist in the scene, the depth obtained through the Gaussians may become unreliable. To address this, we propose a method to keep the set of generated Gaussians within the CoC radius, introducing a learnable scaling factor to increase flexibility and reduce dependence on estimated depth in the CoC modeling. As a result, our model combines physically grounded defocus principles with an adaptive CoC modeling that effectively handles uncertain depths, allowing it to reconstruct sharp 3D scenes using only defocused images. Additionally, our approach has an advantage of learning aperture and focus plane information, enabling dynamic control over depth of field and flexible adjustment of the focus plane. This feature allows for highly customizable scene visualization, adapting to various focus and depth requirements." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 324, + 553, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 324, + 553, + 371 + ], + "spans": [ + { + "bbox": [ + 313, + 324, + 553, + 371 + ], + "type": "text", + "content": "To demonstrate the effectiveness of our model, we conduct extensive experiments on the benchmarks: the Deblur-NeRF [18] dataset and the DoF-NeRF [41] real-world dataset. Our contributions can be summarized as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 376, + 553, + 502 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 313, + 376, + 553, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 376, + 553, + 411 + ], + "spans": [ + { + "bbox": [ + 313, + 376, + 553, + 411 + ], + "type": "text", + "content": "- CoCoGaussian models the CoC at the 3D Gaussian level, reconstructing the precise 3D scene and enabling sharp novel view synthesis from defocused images." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 415, + 553, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 415, + 553, + 437 + ], + "spans": [ + { + "bbox": [ + 313, + 415, + 553, + 437 + ], + "type": "text", + "content": "- We propose an adaptive learning approach that robustly models the CoC even with unreliable depth information." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 441, + 553, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 441, + 553, + 475 + ], + "spans": [ + { + "bbox": [ + 313, + 441, + 553, + 475 + ], + "type": "text", + "content": "- CoCoGaussian enables the customizable depth of field and flexible focus plane adjustment during rendering by learning aperture and focus plane information." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 480, + 552, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 480, + 552, + 502 + ], + "spans": [ + { + "bbox": [ + 313, + 480, + 552, + 502 + ], + "type": "text", + "content": "- CoCoGaussian achieves state-of-the-art performance, both quantitatively and qualitatively." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 517, + 400, + 530 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 517, + 400, + 530 + ], + "spans": [ + { + "bbox": [ + 313, + 517, + 400, + 530 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 538, + 553, + 551 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 538, + 553, + 551 + ], + "spans": [ + { + "bbox": [ + 313, + 538, + 553, + 551 + ], + "type": "text", + "content": "2.1. Scene Representations for Novel View Synthesis" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 558, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 553, + 713 + ], + "type": "text", + "content": "3D scene representation for novel view synthesis has advanced significantly with the advent of neural radiance fields (NeRF) [20], a ray tracing-based volumetric rendering method that generates photo-realistic novel view images from multi-view images. However, as an implicit representation method using deep MLPs, NeRF suffers from slow training and rendering speeds. To address this limitation, explicit representation methods such as Plenoxel [6], TensoRF [3], and Instant-NGP [21] have been introduced. More recently, 3D Gaussian Splatting (3DGS) [9], a rasterization-based method, has emerged as an alternative to ray tracing, significantly mitigating the speed and memory efficiency constraints of ray tracing" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16102" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "text", + "content": "based approaches. NeRF and 3DGS has enabled a wide range of related research, including dynamic scene representation [14, 15, 23, 24, 28, 33], human avatars [8, 27, 39], 3D mesh reconstruction [16, 32, 35, 36], 3D scene representation from sparse-view images [22, 34, 42, 43], and 3D scene representation from blurry images [4, 10-13, 18, 25, 26, 37, 46]. In this paper, we propose a method for 3D scene representation from defocused images by exploiting 3DGS." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 177, + 274, + 190 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 177, + 274, + 190 + ], + "spans": [ + { + "bbox": [ + 55, + 177, + 274, + 190 + ], + "type": "text", + "content": "2.2. Novel View Synthesis from Blurry Images" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 196, + 296, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 196, + 296, + 507 + ], + "spans": [ + { + "bbox": [ + 56, + 196, + 296, + 507 + ], + "type": "text", + "content": "Scene representation methods like NeRF [20] and 3DGS [9] require sharp images as input to render photo-realistic images. However, in real-world scenarios, capturing sharp images is challenging, often suffering from image degradation such as camera motion blur and defocus blur. To address this issue, Deblur-NeRF [18] introduced a 3D ray-based blurring kernel inspired by image blind deblurring [2, 31, 40], which intentionally generates blurry images during training. After training, it renders sharp novel-view images by excluding the trained kernel. Following Deblur-NeRF, various approaches have been proposed to tackle the degradation issue. DP-NeRF [11] introduces a kernel based on the rigid body transformation [17] with prior knowledge that object shape remains consistent in static scenes, and PDRF [25] proposes a blur estimation method with 2-stage efficient rendering. Recently, methods utilizing 3DGS have been proposed to enable faster rendering. For instance, Deblurring 3DGS [10] adjusts Gaussian parameters to generate blurry images during training, and BAGS [26] proposes a blur-agnostic kernel and a blur mask using convolutional neural networks (CNNs). DoF-NeRF [41] has explored modeling defocus blur by leveraging the circle of confusion (CoC) within an implicit neural representation. In this paper, we propose a method that also leverages CoC to model defocus, which enables real-time rendering, while still maintaining high-quality depth-of-field effects." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 519, + 133, + 533 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 519, + 133, + 533 + ], + "spans": [ + { + "bbox": [ + 55, + 519, + 133, + 533 + ], + "type": "text", + "content": "3. Preliminary" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 539, + 184, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 539, + 184, + 552 + ], + "spans": [ + { + "bbox": [ + 55, + 539, + 184, + 552 + ], + "type": "text", + "content": "3.1. 3D Gaussian Splatting" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 558, + 296, + 618 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 296, + 618 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 296, + 618 + ], + "type": "text", + "content": "Unlike ray tracing-based methods [1, 3, 20], 3DGS [9] is built on a rasterization-based approach with differentiable 3D Gaussians. These 3D Gaussians are initialized from a sparse point cloud obtained via a Structure-from-Motion (SfM) [29, 30] algorithm and are defined as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 113, + 627, + 295, + 643 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 627, + 295, + 643 + ], + "spans": [ + { + "bbox": [ + 113, + 627, + 295, + 643 + ], + "type": "interline_equation", + "content": "G (\\mathbf {x}) = e ^ {- \\frac {1}{2} (\\mathbf {x} - \\mu) ^ {\\top} \\boldsymbol {\\Sigma} ^ {- 1} (\\mathbf {x} - \\mu)}, \\tag {1}", + "image_path": "980e967cf80c773a2c634ad08fd5335bd6efa3a75e7936c05e0ed1c266786ace.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{x} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "type": "text", + "content": " is a point on the Gaussian " + }, + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "type": "text", + "content": " centered at the mean vector " + }, + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mu \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "type": "text", + "content": " with an covariance matrix " + }, + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\boldsymbol{\\Sigma} \\in \\mathbb{R}^{3 \\times 3}" + }, + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "type": "text", + "content": ". The 3D covariance matrix " + }, + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\boldsymbol{\\Sigma}" + }, + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "type": "text", + "content": " is derived from a learnable scaling vector " + }, + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{s} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "type": "text", + "content": " and rotation quaternion " + }, + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{q} \\in \\mathbb{R}^4" + }, + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "type": "text", + "content": ", from which the scaling matrix " + }, + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{S} \\in \\mathbb{R}^{3 \\times 3}" + }, + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "type": "text", + "content": " and rotation matrix " + }, + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{R} \\in" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 72, + 503, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 503, + 84 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 503, + 84 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{3\\times 3}" + }, + { + "bbox": [ + 314, + 72, + 503, + 84 + ], + "type": "text", + "content": " are obtained and represented as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 399, + 88, + 553, + 102 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 399, + 88, + 553, + 102 + ], + "spans": [ + { + "bbox": [ + 399, + 88, + 553, + 102 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\Sigma} = \\mathbf {R} \\mathbf {S} \\mathbf {S} ^ {\\top} \\mathbf {R} ^ {\\top}. \\tag {2}", + "image_path": "bab35dcebcf6a748291d1662cc6464f1837b995b0fddea23a2042dc55934f52e.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 107, + 555, + 178 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 107, + 555, + 178 + ], + "spans": [ + { + "bbox": [ + 313, + 107, + 555, + 178 + ], + "type": "text", + "content": "For differentiable splatting [44], the Gaussians in the 3D world coordinate system are projected into the 2D camera coordinate system. This projection uses the viewing transformation " + }, + { + "bbox": [ + 313, + 107, + 555, + 178 + ], + "type": "inline_equation", + "content": "\\mathbf{W} \\in \\mathbb{R}^{3 \\times 3}" + }, + { + "bbox": [ + 313, + 107, + 555, + 178 + ], + "type": "text", + "content": " and the Jacobian " + }, + { + "bbox": [ + 313, + 107, + 555, + 178 + ], + "type": "inline_equation", + "content": "\\mathbf{J} \\in \\mathbb{R}^{2 \\times 3}" + }, + { + "bbox": [ + 313, + 107, + 555, + 178 + ], + "type": "text", + "content": " of the affine approximation of the projective transformation to derive the 2D covariance " + }, + { + "bbox": [ + 313, + 107, + 555, + 178 + ], + "type": "inline_equation", + "content": "\\Sigma^{2\\mathrm{D}} \\in \\mathbb{R}^{2 \\times 2}" + }, + { + "bbox": [ + 313, + 107, + 555, + 178 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 387, + 182, + 553, + 198 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 387, + 182, + 553, + 198 + ], + "spans": [ + { + "bbox": [ + 387, + 182, + 553, + 198 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\Sigma} ^ {2 \\mathrm {D}} = \\mathbf {J} \\mathbf {W} \\boldsymbol {\\Sigma} \\mathbf {W} ^ {\\top} \\mathbf {J} ^ {\\top}. \\tag {3}", + "image_path": "5e7ff3b99720bb464cfce9a8a2c39be11d5afa7e6e04d9583ab45a7724f9d7b4.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 202, + 554, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 202, + 554, + 251 + ], + "spans": [ + { + "bbox": [ + 313, + 202, + 554, + 251 + ], + "type": "text", + "content": "Each Gaussian includes a set of spherical harmonics (SH) coefficients and an opacity value " + }, + { + "bbox": [ + 313, + 202, + 554, + 251 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 202, + 554, + 251 + ], + "type": "text", + "content": " to represent view-dependent color " + }, + { + "bbox": [ + 313, + 202, + 554, + 251 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 313, + 202, + 554, + 251 + ], + "type": "text", + "content": ". The pixel color " + }, + { + "bbox": [ + 313, + 202, + 554, + 251 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_p" + }, + { + "bbox": [ + 313, + 202, + 554, + 251 + ], + "type": "text", + "content": " is then obtained by applying alpha blending to " + }, + { + "bbox": [ + 313, + 202, + 554, + 251 + ], + "type": "inline_equation", + "content": "\\mathcal{N}" + }, + { + "bbox": [ + 313, + 202, + 554, + 251 + ], + "type": "text", + "content": " ordered Gaussians:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 374, + 255, + 553, + 289 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 374, + 255, + 553, + 289 + ], + "spans": [ + { + "bbox": [ + 374, + 255, + 553, + 289 + ], + "type": "interline_equation", + "content": "\\mathbf {c} _ {p} = \\sum_ {i \\in \\mathcal {N}} \\mathbf {c} _ {i} \\alpha_ {i} \\prod_ {j = 1} ^ {i - 1} (1 - \\alpha_ {j}). \\tag {4}", + "image_path": "c86d95204798736f713cefd8ae3f85d610abde18461faa15bd674463ac896401.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 293, + 554, + 329 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 293, + 554, + 329 + ], + "spans": [ + { + "bbox": [ + 313, + 293, + 554, + 329 + ], + "type": "text", + "content": "Our approach, as illustrated in Fig. 1, exploits 3DGS and aims to render defocused images by generating multiple Gaussians to form the CoC for each Gaussian." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 335, + 464, + 347 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 335, + 464, + 347 + ], + "spans": [ + { + "bbox": [ + 313, + 335, + 464, + 347 + ], + "type": "text", + "content": "3.2. 3D Scene Blind Deblurring" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 353, + 555, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 353, + 555, + 484 + ], + "spans": [ + { + "bbox": [ + 313, + 353, + 555, + 484 + ], + "type": "text", + "content": "Image blind deblurring is a technique for learning an image blurring kernel without the supervision of sharp images. Blurry images are obtained by applying convolution with a learned blurring kernel to sharp images, where the kernel size is fixed as a grid structure at each pixel position. Inspired by this approach, Deblur-NeRF [18] applies this technique to NeRF, presenting a method to represent a clean 3D scene using only blurry images. Specifically, Deblur-NeRF warps the input ray into multiple rays that constitute the blur, modeling a ray-based sparse kernel. A blurry pixel is obtained by combining the pixel colors of each ray:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 353, + 490, + 553, + 507 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 490, + 553, + 507 + ], + "spans": [ + { + "bbox": [ + 353, + 490, + 553, + 507 + ], + "type": "interline_equation", + "content": "\\mathbf {c} _ {b l u r} = \\sum w _ {p} ^ {i} \\mathbf {c} _ {p} ^ {i}, w. r. t., \\sum w _ {p} ^ {i} = 1, \\tag {5}", + "image_path": "356cd321daa5dc80df01112c00d0c0371ae4b57a344ae318d43254abbd13965e.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 511, + 554, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 511, + 554, + 570 + ], + "spans": [ + { + "bbox": [ + 313, + 511, + 554, + 570 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 511, + 554, + 570 + ], + "type": "inline_equation", + "content": "w_{p}" + }, + { + "bbox": [ + 313, + 511, + 554, + 570 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 511, + 554, + 570 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_p" + }, + { + "bbox": [ + 313, + 511, + 554, + 570 + ], + "type": "text", + "content": " denote the weight and pixel color for each corresponding ray. Deblur-NeRF enhances learning efficiency by setting the number of warped rays smaller than the kernel size of 2D convolution, and it designs a deformable kernel by adaptive origins and directions of rays." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 571, + 554, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 571, + 554, + 643 + ], + "spans": [ + { + "bbox": [ + 313, + 571, + 554, + 643 + ], + "type": "text", + "content": "As we adopt a rasterization method based on 3DGS [9], we propose an alternative approach: instead of warping rays, we generate 3D Gaussians in the shape of the CoC and render them. In other words, the defocused image is estimated by setting the generated 3D Gaussians, namely CoC Gaussians, as the blurring kernel." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 651, + 370, + 663 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 651, + 370, + 663 + ], + "spans": [ + { + "bbox": [ + 313, + 651, + 370, + 663 + ], + "type": "text", + "content": "4. Method" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 671, + 465, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 671, + 465, + 682 + ], + "spans": [ + { + "bbox": [ + 313, + 671, + 465, + 682 + ], + "type": "text", + "content": "4.1. CoCoGaussian Framework" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "content": "Our goal is to represent a 3D scene using only defocused multi-view images, leveraging fundamental photographic" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16103" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 70, + 553, + 319 + ], + "blocks": [ + { + "bbox": [ + 60, + 70, + 553, + 319 + ], + "lines": [ + { + "bbox": [ + 60, + 70, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 60, + 70, + 553, + 319 + ], + "type": "image", + "image_path": "4169d199f407b6ec6e4cce324d5394acd0ea85e11d85e29f803b274456542d6c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 328, + 555, + 396 + ], + "lines": [ + { + "bbox": [ + 54, + 328, + 555, + 396 + ], + "spans": [ + { + "bbox": [ + 54, + 328, + 555, + 396 + ], + "type": "text", + "content": "Figure 2. We take the camera position " + }, + { + "bbox": [ + 54, + 328, + 555, + 396 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{cam}" + }, + { + "bbox": [ + 54, + 328, + 555, + 396 + ], + "type": "text", + "content": " and the base Gaussian parameters " + }, + { + "bbox": [ + 54, + 328, + 555, + 396 + ], + "type": "inline_equation", + "content": "\\mu_B" + }, + { + "bbox": [ + 54, + 328, + 555, + 396 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 54, + 328, + 555, + 396 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_B" + }, + { + "bbox": [ + 54, + 328, + 555, + 396 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 54, + 328, + 555, + 396 + ], + "type": "inline_equation", + "content": "\\mathbf{q}_B" + }, + { + "bbox": [ + 54, + 328, + 555, + 396 + ], + "type": "text", + "content": " as inputs to the MLP " + }, + { + "bbox": [ + 54, + 328, + 555, + 396 + ], + "type": "inline_equation", + "content": "h_\\theta" + }, + { + "bbox": [ + 54, + 328, + 555, + 396 + ], + "type": "text", + "content": ", which produces five outputs in total. (a) We set the depth " + }, + { + "bbox": [ + 54, + 328, + 555, + 396 + ], + "type": "inline_equation", + "content": "d(\\mu_B)" + }, + { + "bbox": [ + 54, + 328, + 555, + 396 + ], + "type": "text", + "content": " as the Euclidean distance between " + }, + { + "bbox": [ + 54, + 328, + 555, + 396 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{cam}" + }, + { + "bbox": [ + 54, + 328, + 555, + 396 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 328, + 555, + 396 + ], + "type": "inline_equation", + "content": "\\mu_B" + }, + { + "bbox": [ + 54, + 328, + 555, + 396 + ], + "type": "text", + "content": ". Using the output " + }, + { + "bbox": [ + 54, + 328, + 555, + 396 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 54, + 328, + 555, + 396 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 54, + 328, + 555, + 396 + ], + "type": "inline_equation", + "content": "h_\\theta" + }, + { + "bbox": [ + 54, + 328, + 555, + 396 + ], + "type": "text", + "content": ", the " + }, + { + "bbox": [ + 54, + 328, + 555, + 396 + ], + "type": "inline_equation", + "content": "d(\\mu_B)" + }, + { + "bbox": [ + 54, + 328, + 555, + 396 + ], + "type": "text", + "content": " and the learnable focus plane " + }, + { + "bbox": [ + 54, + 328, + 555, + 396 + ], + "type": "inline_equation", + "content": "d_F" + }, + { + "bbox": [ + 54, + 328, + 555, + 396 + ], + "type": "text", + "content": ", we apply Eq. (7) to determine the CoC diameter. The diameter, combined with the outputs " + }, + { + "bbox": [ + 54, + 328, + 555, + 396 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 54, + 328, + 555, + 396 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 328, + 555, + 396 + ], + "type": "inline_equation", + "content": "\\mathbf{d}" + }, + { + "bbox": [ + 54, + 328, + 555, + 396 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 54, + 328, + 555, + 396 + ], + "type": "inline_equation", + "content": "h_\\theta" + }, + { + "bbox": [ + 54, + 328, + 555, + 396 + ], + "type": "text", + "content": ", is then used in Eq. (9) to compute the offset values for " + }, + { + "bbox": [ + 54, + 328, + 555, + 396 + ], + "type": "inline_equation", + "content": "\\mu_{\\mathrm{CoC}}" + }, + { + "bbox": [ + 54, + 328, + 555, + 396 + ], + "type": "text", + "content": ". By adding these offsets to " + }, + { + "bbox": [ + 54, + 328, + 555, + 396 + ], + "type": "inline_equation", + "content": "\\mu_B" + }, + { + "bbox": [ + 54, + 328, + 555, + 396 + ], + "type": "text", + "content": ", we obtain the " + }, + { + "bbox": [ + 54, + 328, + 555, + 396 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_{\\mathrm{CoC}}" + }, + { + "bbox": [ + 54, + 328, + 555, + 396 + ], + "type": "text", + "content": " means. (b) We obtain " + }, + { + "bbox": [ + 54, + 328, + 555, + 396 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_{\\mathrm{CoC}}" + }, + { + "bbox": [ + 54, + 328, + 555, + 396 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 328, + 555, + 396 + ], + "type": "inline_equation", + "content": "\\mathbf{q}_{\\mathrm{CoC}}" + }, + { + "bbox": [ + 54, + 328, + 555, + 396 + ], + "type": "text", + "content": " by applying Eqs. (10) and (11) to " + }, + { + "bbox": [ + 54, + 328, + 555, + 396 + ], + "type": "inline_equation", + "content": "\\delta \\mathbf{s}_{\\mathrm{CoC}}" + }, + { + "bbox": [ + 54, + 328, + 555, + 396 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 328, + 555, + 396 + ], + "type": "inline_equation", + "content": "\\delta \\mathbf{q}_{\\mathrm{CoC}}" + }, + { + "bbox": [ + 54, + 328, + 555, + 396 + ], + "type": "text", + "content": ", the outputs of " + }, + { + "bbox": [ + 54, + 328, + 555, + 396 + ], + "type": "inline_equation", + "content": "h_\\theta" + }, + { + "bbox": [ + 54, + 328, + 555, + 396 + ], + "type": "text", + "content": ". (c) Using " + }, + { + "bbox": [ + 54, + 328, + 555, + 396 + ], + "type": "inline_equation", + "content": "\\mu_{\\mathrm{CoC}}" + }, + { + "bbox": [ + 54, + 328, + 555, + 396 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 54, + 328, + 555, + 396 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_{\\mathrm{CoC}}" + }, + { + "bbox": [ + 54, + 328, + 555, + 396 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 54, + 328, + 555, + 396 + ], + "type": "inline_equation", + "content": "\\mathbf{q}_{\\mathrm{CoC}}" + }, + { + "bbox": [ + 54, + 328, + 555, + 396 + ], + "type": "text", + "content": ", we get the " + }, + { + "bbox": [ + 54, + 328, + 555, + 396 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_{\\mathrm{CoC}}" + }, + { + "bbox": [ + 54, + 328, + 555, + 396 + ], + "type": "text", + "content": ". Finally, we rasterize " + }, + { + "bbox": [ + 54, + 328, + 555, + 396 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_{\\mathrm{CoC}}" + }, + { + "bbox": [ + 54, + 328, + 555, + 396 + ], + "type": "text", + "content": " along with " + }, + { + "bbox": [ + 54, + 328, + 555, + 396 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_B" + }, + { + "bbox": [ + 54, + 328, + 555, + 396 + ], + "type": "text", + "content": " to produce " + }, + { + "bbox": [ + 54, + 328, + 555, + 396 + ], + "type": "inline_equation", + "content": "(M + 1)" + }, + { + "bbox": [ + 54, + 328, + 555, + 396 + ], + "type": "text", + "content": " images, and apply a weighted sum to generate the final defocused image." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 415, + 296, + 691 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 415, + 296, + 691 + ], + "spans": [ + { + "bbox": [ + 56, + 415, + 296, + 691 + ], + "type": "text", + "content": "principles to guide our method. We generate 3D Gaussians to capture a blurring kernel shaped as the CoC. The whole framework of CoCoGaussian is on Fig. 2. First, we compute the CoC diameter using the Gaussian depth, derived from the means of the given base 3D Gaussian set " + }, + { + "bbox": [ + 56, + 415, + 296, + 691 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_B" + }, + { + "bbox": [ + 56, + 415, + 296, + 691 + ], + "type": "text", + "content": ", which consists of " + }, + { + "bbox": [ + 56, + 415, + 296, + 691 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 56, + 415, + 296, + 691 + ], + "type": "text", + "content": " Gaussians, along with the camera position and aperture information (Sec. 4.2). Next, for the base Gaussian set " + }, + { + "bbox": [ + 56, + 415, + 296, + 691 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_B" + }, + { + "bbox": [ + 56, + 415, + 296, + 691 + ], + "type": "text", + "content": ", we generate " + }, + { + "bbox": [ + 56, + 415, + 296, + 691 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 56, + 415, + 296, + 691 + ], + "type": "text", + "content": " CoC Gaussian sets " + }, + { + "bbox": [ + 56, + 415, + 296, + 691 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_{\\mathrm{CoC}}" + }, + { + "bbox": [ + 56, + 415, + 296, + 691 + ], + "type": "text", + "content": ", resulting in a total of " + }, + { + "bbox": [ + 56, + 415, + 296, + 691 + ], + "type": "inline_equation", + "content": "(M\\times N)" + }, + { + "bbox": [ + 56, + 415, + 296, + 691 + ], + "type": "text", + "content": " Gaussians to capture the CoC shape. However, when the scene contains objects with refraction or reflection, the Gaussian depth may be unreliable, which can lead to suboptimal CoC diameter. To address this issue, we propose an adaptive CoC Gaussian generation method that minimizes the dependence on depth for " + }, + { + "bbox": [ + 56, + 415, + 296, + 691 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_{\\mathrm{CoC}}" + }, + { + "bbox": [ + 56, + 415, + 296, + 691 + ], + "type": "text", + "content": " (Sec. 4.3). Finally, we incorporate adaptive rotation quaternion and scaling vector adjustments, inspired by the Deblurring 3DGS [10] approach, to allow flexible learning of " + }, + { + "bbox": [ + 56, + 415, + 296, + 691 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_{\\mathrm{CoC}}" + }, + { + "bbox": [ + 56, + 415, + 296, + 691 + ], + "type": "text", + "content": ". Furthermore, we apply Eq. (5) to images rendered from " + }, + { + "bbox": [ + 56, + 415, + 296, + 691 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_B" + }, + { + "bbox": [ + 56, + 415, + 296, + 691 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 56, + 415, + 296, + 691 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_{\\mathrm{CoC}}" + }, + { + "bbox": [ + 56, + 415, + 296, + 691 + ], + "type": "text", + "content": ", using a weighted sum approach [12, 26] to create the final blurry image. This photography-prior-based approach allows us to accurately represent 3D scenes using only defocused images while adhering to realistic defocus principles." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 313, + 415, + 522, + 427 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 415, + 522, + 427 + ], + "spans": [ + { + "bbox": [ + 313, + 415, + 522, + 427 + ], + "type": "text", + "content": "4.2. Circle of Confusion from 3D Gaussians" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 434, + 555, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 434, + 555, + 566 + ], + "spans": [ + { + "bbox": [ + 313, + 434, + 555, + 566 + ], + "type": "text", + "content": "To model the CoC, we assume an ideal system that ignores distortions such as lens aberrations. Before generating the CoC Gaussian sets " + }, + { + "bbox": [ + 313, + 434, + 555, + 566 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_{\\mathrm{CoC}}" + }, + { + "bbox": [ + 313, + 434, + 555, + 566 + ], + "type": "text", + "content": ", we first need to obtain the CoC diameter, which depends on several factors: the aperture diameter " + }, + { + "bbox": [ + 313, + 434, + 555, + 566 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 313, + 434, + 555, + 566 + ], + "type": "text", + "content": ", focal length " + }, + { + "bbox": [ + 313, + 434, + 555, + 566 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 313, + 434, + 555, + 566 + ], + "type": "text", + "content": ", focus plane " + }, + { + "bbox": [ + 313, + 434, + 555, + 566 + ], + "type": "inline_equation", + "content": "d_{F}" + }, + { + "bbox": [ + 313, + 434, + 555, + 566 + ], + "type": "text", + "content": ", and depth " + }, + { + "bbox": [ + 313, + 434, + 555, + 566 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 313, + 434, + 555, + 566 + ], + "type": "text", + "content": ". The focus plane represents the distance from the camera position to the plane in focus. We define the depth " + }, + { + "bbox": [ + 313, + 434, + 555, + 566 + ], + "type": "inline_equation", + "content": "d(\\mu_B)" + }, + { + "bbox": [ + 313, + 434, + 555, + 566 + ], + "type": "text", + "content": " as the Euclidean distance between the camera position " + }, + { + "bbox": [ + 313, + 434, + 555, + 566 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{cam} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 313, + 434, + 555, + 566 + ], + "type": "text", + "content": " and the means of the base Gaussian set " + }, + { + "bbox": [ + 313, + 434, + 555, + 566 + ], + "type": "inline_equation", + "content": "\\mu_B \\in \\mathbb{R}^{N \\times 3}" + }, + { + "bbox": [ + 313, + 434, + 555, + 566 + ], + "type": "text", + "content": ". Exploiting these values, we express the CoC diameter " + }, + { + "bbox": [ + 313, + 434, + 555, + 566 + ], + "type": "inline_equation", + "content": "\\sigma(\\mu_B)" + }, + { + "bbox": [ + 313, + 434, + 555, + 566 + ], + "type": "text", + "content": " as follows [7]:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 364, + 574, + 555, + 601 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 364, + 574, + 555, + 601 + ], + "spans": [ + { + "bbox": [ + 364, + 574, + 555, + 601 + ], + "type": "interline_equation", + "content": "\\sigma (\\mu_ {B}) = f D \\times \\frac {\\left| d (\\mu_ {B}) - d _ {F} \\right|}{d (\\mu_ {B}) (d _ {F} - f)}, \\tag {6}", + "image_path": "270c15e63d8552669bc3967a3ad2e4b5e4b9cd1817f64083929ef149ef8ea846.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 609, + 556, + 681 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 609, + 556, + 681 + ], + "spans": [ + { + "bbox": [ + 313, + 609, + 556, + 681 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 609, + 556, + 681 + ], + "type": "inline_equation", + "content": "d_F" + }, + { + "bbox": [ + 313, + 609, + 556, + 681 + ], + "type": "text", + "content": " is set to a learnable parameter and initialized as the average distance between " + }, + { + "bbox": [ + 313, + 609, + 556, + 681 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{cam}" + }, + { + "bbox": [ + 313, + 609, + 556, + 681 + ], + "type": "text", + "content": " and the points in the SfM point cloud to ensure stable training. Note that the focus plane " + }, + { + "bbox": [ + 313, + 609, + 556, + 681 + ], + "type": "inline_equation", + "content": "d_F" + }, + { + "bbox": [ + 313, + 609, + 556, + 681 + ], + "type": "text", + "content": " is specific to each input image. Additionally, because the focal length " + }, + { + "bbox": [ + 313, + 609, + 556, + 681 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 313, + 609, + 556, + 681 + ], + "type": "text", + "content": " is usually much smaller than the focus plane " + }, + { + "bbox": [ + 313, + 609, + 556, + 681 + ], + "type": "inline_equation", + "content": "d_F" + }, + { + "bbox": [ + 313, + 609, + 556, + 681 + ], + "type": "text", + "content": ", we modify Eq. (6) as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 368, + 689, + 555, + 717 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 368, + 689, + 555, + 717 + ], + "spans": [ + { + "bbox": [ + 368, + 689, + 555, + 717 + ], + "type": "interline_equation", + "content": "\\sigma \\left(\\mu_ {B}\\right) \\approx K \\times \\left| \\frac {1}{d \\left(\\mu_ {B}\\right)} - \\frac {1}{d _ {F}} \\right|, \\tag {7}", + "image_path": "2f7bd55152838ac9f28afc85aaf6f7bad6b48f2f0fbfd54fa2923a7e6494ddf9.jpg" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "16104" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "type": "text", + "content": "where the product of the focal length and aperture diameter is represented as a single learnable scalar " + }, + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "type": "inline_equation", + "content": "K = f \\times D" + }, + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "type": "text", + "content": " [41]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 96, + 296, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 96, + 296, + 180 + ], + "spans": [ + { + "bbox": [ + 55, + 96, + 296, + 180 + ], + "type": "text", + "content": "We design a simple MLP " + }, + { + "bbox": [ + 55, + 96, + 296, + 180 + ], + "type": "inline_equation", + "content": "h_\\theta" + }, + { + "bbox": [ + 55, + 96, + 296, + 180 + ], + "type": "text", + "content": " parameterized by " + }, + { + "bbox": [ + 55, + 96, + 296, + 180 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 55, + 96, + 296, + 180 + ], + "type": "text", + "content": " to estimate the scalar " + }, + { + "bbox": [ + 55, + 96, + 296, + 180 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 96, + 296, + 180 + ], + "type": "text", + "content": ". Then, the aperture parameter " + }, + { + "bbox": [ + 55, + 96, + 296, + 180 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 96, + 296, + 180 + ], + "type": "text", + "content": " is obtained by passing these features as inputs to " + }, + { + "bbox": [ + 55, + 96, + 296, + 180 + ], + "type": "inline_equation", + "content": "h_\\theta" + }, + { + "bbox": [ + 55, + 96, + 296, + 180 + ], + "type": "text", + "content": ", along with the scaling vectors " + }, + { + "bbox": [ + 55, + 96, + 296, + 180 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_B \\in \\mathbb{R}^{N \\times 3}" + }, + { + "bbox": [ + 55, + 96, + 296, + 180 + ], + "type": "text", + "content": " and rotation quaternions " + }, + { + "bbox": [ + 55, + 96, + 296, + 180 + ], + "type": "inline_equation", + "content": "\\mathbf{q}_B \\in \\mathbb{R}^{N \\times 4}" + }, + { + "bbox": [ + 55, + 96, + 296, + 180 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 55, + 96, + 296, + 180 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_B" + }, + { + "bbox": [ + 55, + 96, + 296, + 180 + ], + "type": "text", + "content": ". This approach allows us to obtain the CoC diameter " + }, + { + "bbox": [ + 55, + 96, + 296, + 180 + ], + "type": "inline_equation", + "content": "\\sigma(\\mu_B)" + }, + { + "bbox": [ + 55, + 96, + 296, + 180 + ], + "type": "text", + "content": " for the " + }, + { + "bbox": [ + 55, + 96, + 296, + 180 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_B" + }, + { + "bbox": [ + 55, + 96, + 296, + 180 + ], + "type": "text", + "content": ", marking the first step in modeling " + }, + { + "bbox": [ + 55, + 96, + 296, + 180 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_{\\mathrm{CoC}}" + }, + { + "bbox": [ + 55, + 96, + 296, + 180 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 188, + 247, + 201 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 188, + 247, + 201 + ], + "spans": [ + { + "bbox": [ + 55, + 188, + 247, + 201 + ], + "type": "text", + "content": "4.3. Adaptive CoC Gaussian Generation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 206, + 296, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 206, + 296, + 361 + ], + "spans": [ + { + "bbox": [ + 55, + 206, + 296, + 361 + ], + "type": "text", + "content": "After obtaining the CoC diameters, we use these diameters to generate " + }, + { + "bbox": [ + 55, + 206, + 296, + 361 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 206, + 296, + 361 + ], + "type": "text", + "content": " CoC Gaussian sets " + }, + { + "bbox": [ + 55, + 206, + 296, + 361 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_{\\mathrm{CoC}}" + }, + { + "bbox": [ + 55, + 206, + 296, + 361 + ], + "type": "text", + "content": ". For the Gaussian parameters of " + }, + { + "bbox": [ + 55, + 206, + 296, + 361 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_{\\mathrm{CoC}}" + }, + { + "bbox": [ + 55, + 206, + 296, + 361 + ], + "type": "text", + "content": ", we set the spherical harmonics (SH) coefficients and opacity values identical to those of " + }, + { + "bbox": [ + 55, + 206, + 296, + 361 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_B" + }, + { + "bbox": [ + 55, + 206, + 296, + 361 + ], + "type": "text", + "content": ", while varying only the means " + }, + { + "bbox": [ + 55, + 206, + 296, + 361 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 55, + 206, + 296, + 361 + ], + "type": "text", + "content": ", scaling vectors " + }, + { + "bbox": [ + 55, + 206, + 296, + 361 + ], + "type": "inline_equation", + "content": "\\mathbf{s}" + }, + { + "bbox": [ + 55, + 206, + 296, + 361 + ], + "type": "text", + "content": ", and rotation quaternions " + }, + { + "bbox": [ + 55, + 206, + 296, + 361 + ], + "type": "inline_equation", + "content": "\\mathbf{q}" + }, + { + "bbox": [ + 55, + 206, + 296, + 361 + ], + "type": "text", + "content": ". The means of each CoC Gaussian set are created by adding offsets " + }, + { + "bbox": [ + 55, + 206, + 296, + 361 + ], + "type": "inline_equation", + "content": "\\Delta \\mu_{\\mathrm{CoC}} \\in \\mathbb{R}^{M \\times N \\times 3}" + }, + { + "bbox": [ + 55, + 206, + 296, + 361 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 206, + 296, + 361 + ], + "type": "inline_equation", + "content": "\\mu_B" + }, + { + "bbox": [ + 55, + 206, + 296, + 361 + ], + "type": "text", + "content": ", defined as " + }, + { + "bbox": [ + 55, + 206, + 296, + 361 + ], + "type": "inline_equation", + "content": "\\mu_{\\mathrm{CoC}} = \\mu_B + \\Delta \\mu_{\\mathrm{CoC}}" + }, + { + "bbox": [ + 55, + 206, + 296, + 361 + ], + "type": "text", + "content": ". Since " + }, + { + "bbox": [ + 55, + 206, + 296, + 361 + ], + "type": "inline_equation", + "content": "\\Delta \\mu_{\\mathrm{CoC}}" + }, + { + "bbox": [ + 55, + 206, + 296, + 361 + ], + "type": "text", + "content": " are 3D vector sets, we obtain unit vector sets " + }, + { + "bbox": [ + 55, + 206, + 296, + 361 + ], + "type": "inline_equation", + "content": "\\mathbf{d} \\in \\mathbb{R}^{M \\times N \\times 3}" + }, + { + "bbox": [ + 55, + 206, + 296, + 361 + ], + "type": "text", + "content": " representing the directions from " + }, + { + "bbox": [ + 55, + 206, + 296, + 361 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_B" + }, + { + "bbox": [ + 55, + 206, + 296, + 361 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 206, + 296, + 361 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_{\\mathrm{CoC}}" + }, + { + "bbox": [ + 55, + 206, + 296, + 361 + ], + "type": "text", + "content": ", which are additional outputs from the MLP " + }, + { + "bbox": [ + 55, + 206, + 296, + 361 + ], + "type": "inline_equation", + "content": "h_\\theta" + }, + { + "bbox": [ + 55, + 206, + 296, + 361 + ], + "type": "text", + "content": " discussed in the previous section. Moreover, since " + }, + { + "bbox": [ + 55, + 206, + 296, + 361 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_{\\mathrm{CoC}}" + }, + { + "bbox": [ + 55, + 206, + 296, + 361 + ], + "type": "text", + "content": " consists of a total of " + }, + { + "bbox": [ + 55, + 206, + 296, + 361 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 206, + 296, + 361 + ], + "type": "text", + "content": " Gaussian sets, The direction vectors " + }, + { + "bbox": [ + 55, + 206, + 296, + 361 + ], + "type": "inline_equation", + "content": "\\mathbf{d}" + }, + { + "bbox": [ + 55, + 206, + 296, + 361 + ], + "type": "text", + "content": " consist of " + }, + { + "bbox": [ + 55, + 206, + 296, + 361 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 206, + 296, + 361 + ], + "type": "text", + "content": " instances:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 89, + 369, + 295, + 393 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 369, + 295, + 393 + ], + "spans": [ + { + "bbox": [ + 89, + 369, + 295, + 393 + ], + "type": "interline_equation", + "content": "\\Delta \\mu_ {\\mathrm {C o C}; m} = \\frac {\\sigma (\\mu_ {B})}{2} \\mathbf {d} _ {m}, \\text {w h e r e} m \\in M. \\tag {8}", + "image_path": "1e1a11bdda4ee68dfe67a31bd8f1c869d825da3e44daa368eb0c0bf1ccf81202.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 401, + 296, + 614 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 401, + 296, + 614 + ], + "spans": [ + { + "bbox": [ + 55, + 401, + 296, + 614 + ], + "type": "text", + "content": "However, theses offsets " + }, + { + "bbox": [ + 55, + 401, + 296, + 614 + ], + "type": "inline_equation", + "content": "\\Delta \\mu_{\\mathrm{CoC}}" + }, + { + "bbox": [ + 55, + 401, + 296, + 614 + ], + "type": "text", + "content": " presents two potential issues. First, this approach places all " + }, + { + "bbox": [ + 55, + 401, + 296, + 614 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_{\\mathrm{CoC}}" + }, + { + "bbox": [ + 55, + 401, + 296, + 614 + ], + "type": "text", + "content": " only on the boundary of the CoC, which limits its ability to fully capture the defocus effect. Modeling the entire CoC, including its interior, would allow for a more accurate and flexible representation of defocus blur. The second issue arises when there are refractive or reflective surfaces in the scene, as these subjects can lead to incorrect optimization of the Gaussian means, making the CoC diameter obtained through Eq. (7) unreliable. To address these problems, we introduce simple learnable CoC scaling parameters " + }, + { + "bbox": [ + 55, + 401, + 296, + 614 + ], + "type": "inline_equation", + "content": "\\beta \\in (0,1]" + }, + { + "bbox": [ + 55, + 401, + 296, + 614 + ], + "type": "text", + "content": ", which are additional outputs of " + }, + { + "bbox": [ + 55, + 401, + 296, + 614 + ], + "type": "inline_equation", + "content": "h_\\theta" + }, + { + "bbox": [ + 55, + 401, + 296, + 614 + ], + "type": "text", + "content": ". The parameter " + }, + { + "bbox": [ + 55, + 401, + 296, + 614 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 55, + 401, + 296, + 614 + ], + "type": "text", + "content": " is learned within the range of 0 to 1, ensuring that " + }, + { + "bbox": [ + 55, + 401, + 296, + 614 + ], + "type": "inline_equation", + "content": "\\mu_{\\mathrm{CoC}}" + }, + { + "bbox": [ + 55, + 401, + 296, + 614 + ], + "type": "text", + "content": " resides within the CoC boundary. Additionally, even when depth is inaccurately measured and the predicted CoC diameter is larger than the actual size, " + }, + { + "bbox": [ + 55, + 401, + 296, + 614 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 55, + 401, + 296, + 614 + ], + "type": "text", + "content": " adaptively scales it down. Thus, the offset " + }, + { + "bbox": [ + 55, + 401, + 296, + 614 + ], + "type": "inline_equation", + "content": "\\Delta \\mu_{\\mathrm{CoC}}" + }, + { + "bbox": [ + 55, + 401, + 296, + 614 + ], + "type": "text", + "content": " is modified by applying " + }, + { + "bbox": [ + 55, + 401, + 296, + 614 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 55, + 401, + 296, + 614 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 119, + 615, + 294, + 637 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 119, + 615, + 294, + 637 + ], + "spans": [ + { + "bbox": [ + 119, + 615, + 294, + 637 + ], + "type": "interline_equation", + "content": "\\Delta \\mu_ {\\mathrm {C o C}; m} = \\frac {\\sigma (\\mu_ {B})}{2} \\beta_ {m} \\mathbf {d} _ {m}. \\tag {9}", + "image_path": "22f2dfef239fb6110d1420cb532c04e0fc119b82129236a62420c51a7d5ffcc1.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": "Additionally, similar to Deblurring 3DGS [10], we introduce scaling factors " + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\delta \\mathbf{s}_{\\mathrm{CoC}} \\in \\mathbb{R}^{M \\times N \\times 3}" + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\delta \\mathbf{q}_{\\mathrm{CoC}} \\in \\mathbb{R}^{M \\times N \\times 4}" + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": ", which allow flexible learning of the scaling vectors " + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_{\\mathrm{CoC}}" + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": " and rotation quaternions " + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{q}_{\\mathrm{CoC}}" + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_{\\mathrm{CoC}}" + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": ". These factors are also the outputs of " + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "inline_equation", + "content": "h_{\\theta}" + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": ". The parameters " + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\delta \\mathbf{s}_{\\mathrm{CoC}}" + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\delta \\mathbf{q}_{\\mathrm{CoC}}" + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": " are multiplied by the corresponding parameters" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "type": "text", + "content": "of the base Gaussian set " + }, + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_B" + }, + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "type": "text", + "content": ", specifically " + }, + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_B" + }, + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "type": "inline_equation", + "content": "\\mathbf{q}_B" + }, + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "type": "text", + "content": ", and are expressed as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 382, + 107, + 555, + 121 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 382, + 107, + 555, + 121 + ], + "spans": [ + { + "bbox": [ + 382, + 107, + 555, + 121 + ], + "type": "interline_equation", + "content": "\\mathbf {s} _ {\\mathrm {C o C}; m} = \\mathbf {s} _ {B} \\times \\delta \\mathbf {s} _ {\\mathrm {C o C}; m}, \\tag {10}", + "image_path": "a9b958d3f9bf86d9c4a46f3338c1ac1443ecee1e0dcfa3552693fe2b3929a8f3.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 381, + 131, + 555, + 143 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 381, + 131, + 555, + 143 + ], + "spans": [ + { + "bbox": [ + 381, + 131, + 555, + 143 + ], + "type": "interline_equation", + "content": "\\mathbf {q} _ {\\mathrm {C o C}; m} = \\mathbf {q} _ {B} \\times \\delta \\mathbf {q} _ {\\mathrm {C o C}; m}, \\tag {11}", + "image_path": "566defaad30e8784d7a9970a01666f9fe79e813bab36773ab163ade116a0fc1c.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 148, + 555, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 148, + 555, + 185 + ], + "spans": [ + { + "bbox": [ + 313, + 148, + 555, + 185 + ], + "type": "text", + "content": "where the scaling parameters are constrained by the fixed values " + }, + { + "bbox": [ + 313, + 148, + 555, + 185 + ], + "type": "inline_equation", + "content": "\\delta \\mathbf{s}_{\\mathrm{max}}" + }, + { + "bbox": [ + 313, + 148, + 555, + 185 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 148, + 555, + 185 + ], + "type": "inline_equation", + "content": "\\delta \\mathbf{q}_{\\mathrm{max}}" + }, + { + "bbox": [ + 313, + 148, + 555, + 185 + ], + "type": "text", + "content": ": " + }, + { + "bbox": [ + 313, + 148, + 555, + 185 + ], + "type": "inline_equation", + "content": "\\delta \\mathbf{s}_{\\mathrm{CoC}} \\in [1, \\delta \\mathbf{s}_{\\mathrm{max}}]" + }, + { + "bbox": [ + 313, + 148, + 555, + 185 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 148, + 555, + 185 + ], + "type": "inline_equation", + "content": "\\delta \\mathbf{q}_{\\mathrm{CoC}} \\in [1, \\delta \\mathbf{q}_{\\mathrm{max}}]" + }, + { + "bbox": [ + 313, + 148, + 555, + 185 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 186, + 555, + 233 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 186, + 555, + 233 + ], + "spans": [ + { + "bbox": [ + 313, + 186, + 555, + 233 + ], + "type": "text", + "content": "We obtain a total of " + }, + { + "bbox": [ + 313, + 186, + 555, + 233 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 186, + 555, + 233 + ], + "type": "text", + "content": " means of CoC Gaussian sets " + }, + { + "bbox": [ + 313, + 186, + 555, + 233 + ], + "type": "inline_equation", + "content": "\\mu_{\\mathrm{CoC}}" + }, + { + "bbox": [ + 313, + 186, + 555, + 233 + ], + "type": "text", + "content": ", scaling vectors " + }, + { + "bbox": [ + 313, + 186, + 555, + 233 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_{\\mathrm{CoC}}" + }, + { + "bbox": [ + 313, + 186, + 555, + 233 + ], + "type": "text", + "content": ", and rotation quaternion " + }, + { + "bbox": [ + 313, + 186, + 555, + 233 + ], + "type": "inline_equation", + "content": "\\mathbf{q}_{\\mathrm{CoC}}" + }, + { + "bbox": [ + 313, + 186, + 555, + 233 + ], + "type": "text", + "content": ". They are combined with the opacity and SH coefficients of " + }, + { + "bbox": [ + 313, + 186, + 555, + 233 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_B" + }, + { + "bbox": [ + 313, + 186, + 555, + 233 + ], + "type": "text", + "content": ", generate the " + }, + { + "bbox": [ + 313, + 186, + 555, + 233 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 186, + 555, + 233 + ], + "type": "text", + "content": " Gaussian sets " + }, + { + "bbox": [ + 313, + 186, + 555, + 233 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_{\\mathrm{CoC}}" + }, + { + "bbox": [ + 313, + 186, + 555, + 233 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 326, + 244, + 555, + 258 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 244, + 555, + 258 + ], + "spans": [ + { + "bbox": [ + 326, + 244, + 555, + 258 + ], + "type": "interline_equation", + "content": "\\mathbf {x} _ {\\mathrm {C o C}; m}, \\mathbf {s} _ {\\mathrm {C o C}; m}, \\mathbf {q} _ {\\mathrm {C o C}; m}, \\alpha_ {B}, \\mathbf {S H} _ {B} \\rightarrow \\mathbf {G} _ {\\mathrm {C o C}; m}. \\tag {12}", + "image_path": "436c3dd6be60244cf9508ade02808a9a64fe83e17c7fa28e90366c8cbae38c59.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 266, + 554, + 315 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 266, + 554, + 315 + ], + "spans": [ + { + "bbox": [ + 313, + 266, + 554, + 315 + ], + "type": "text", + "content": "We rasterize the resulting " + }, + { + "bbox": [ + 313, + 266, + 554, + 315 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 266, + 554, + 315 + ], + "type": "text", + "content": " Gaussian sets along with the base Gaussian set, producing " + }, + { + "bbox": [ + 313, + 266, + 554, + 315 + ], + "type": "inline_equation", + "content": "(M + 1)" + }, + { + "bbox": [ + 313, + 266, + 554, + 315 + ], + "type": "text", + "content": " images. Then, using the weighted sum approach from Eq. (5), as detailed in the following section, we obtain the final blurry image." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 323, + 400, + 335 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 323, + 400, + 335 + ], + "spans": [ + { + "bbox": [ + 313, + 323, + 400, + 335 + ], + "type": "text", + "content": "4.4. Optimization" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 341, + 555, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 341, + 555, + 437 + ], + "spans": [ + { + "bbox": [ + 313, + 341, + 555, + 437 + ], + "type": "text", + "content": "Weighted Sum. To apply Eq. (5) to the " + }, + { + "bbox": [ + 313, + 341, + 555, + 437 + ], + "type": "inline_equation", + "content": "(M + 1)" + }, + { + "bbox": [ + 313, + 341, + 555, + 437 + ], + "type": "text", + "content": " images " + }, + { + "bbox": [ + 313, + 341, + 555, + 437 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 313, + 341, + 555, + 437 + ], + "type": "text", + "content": " rasterized from " + }, + { + "bbox": [ + 313, + 341, + 555, + 437 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_B" + }, + { + "bbox": [ + 313, + 341, + 555, + 437 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 341, + 555, + 437 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_{\\mathrm{CoC}}" + }, + { + "bbox": [ + 313, + 341, + 555, + 437 + ], + "type": "text", + "content": ", we adopt the methods of recent studies [12, 26]. We use a shallow CNN " + }, + { + "bbox": [ + 313, + 341, + 555, + 437 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 313, + 341, + 555, + 437 + ], + "type": "text", + "content": " to compute pixel-wise weights " + }, + { + "bbox": [ + 313, + 341, + 555, + 437 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 313, + 341, + 555, + 437 + ], + "type": "text", + "content": " for these images, applying a softmax function to ensure that the weights for each pixel sum to 1. Then, we multiply each image by its corresponding weight and sum the results to obtain the final defocused image " + }, + { + "bbox": [ + 313, + 341, + 555, + 437 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_{blur}" + }, + { + "bbox": [ + 313, + 341, + 555, + 437 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 319, + 446, + 554, + 479 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 446, + 554, + 479 + ], + "spans": [ + { + "bbox": [ + 319, + 446, + 554, + 479 + ], + "type": "interline_equation", + "content": "\\mathcal {I} _ {\\text {b l u r}} = \\sum_ {m = 1} ^ {M + 1} \\mathcal {I} _ {m} \\mathcal {W} _ {m}, \\text {w h e r e} \\mathcal {W} = \\operatorname {s o f t m a x} (\\mathcal {F} (\\mathbf {I})), \\tag {13}", + "image_path": "3abe9f01bf8c947bd0e210f7e801a37fe3710aa9912b4762f4bfe619efd2a61c.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 488, + 554, + 512 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 488, + 554, + 512 + ], + "spans": [ + { + "bbox": [ + 313, + 488, + 554, + 512 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 488, + 554, + 512 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_m" + }, + { + "bbox": [ + 313, + 488, + 554, + 512 + ], + "type": "text", + "content": " represents the image rasterized from " + }, + { + "bbox": [ + 313, + 488, + 554, + 512 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_{\\mathrm{CoC};m}" + }, + { + "bbox": [ + 313, + 488, + 554, + 512 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 488, + 554, + 512 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_{M + 1}" + }, + { + "bbox": [ + 313, + 488, + 554, + 512 + ], + "type": "text", + "content": " represents the image rasterized from " + }, + { + "bbox": [ + 313, + 488, + 554, + 512 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_B" + }, + { + "bbox": [ + 313, + 488, + 554, + 512 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 528, + 554, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 528, + 554, + 601 + ], + "spans": [ + { + "bbox": [ + 313, + 528, + 554, + 601 + ], + "type": "text", + "content": "Objective. We minimize the " + }, + { + "bbox": [ + 313, + 528, + 554, + 601 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_1" + }, + { + "bbox": [ + 313, + 528, + 554, + 601 + ], + "type": "text", + "content": " loss and D-SSIM loss " + }, + { + "bbox": [ + 313, + 528, + 554, + 601 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{D - SSIM}}" + }, + { + "bbox": [ + 313, + 528, + 554, + 601 + ], + "type": "text", + "content": " between the ground truth defocused image and the output defocused image. The " + }, + { + "bbox": [ + 313, + 528, + 554, + 601 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_1" + }, + { + "bbox": [ + 313, + 528, + 554, + 601 + ], + "type": "text", + "content": " loss reduces the pixelwise differences between images, while the D-SSIM loss minimizes structural differences between them. The final objective " + }, + { + "bbox": [ + 313, + 528, + 554, + 601 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{rgb}" + }, + { + "bbox": [ + 313, + 528, + 554, + 601 + ], + "type": "text", + "content": " is defined as follows:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 369, + 610, + 555, + 624 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 369, + 610, + 555, + 624 + ], + "spans": [ + { + "bbox": [ + 369, + 610, + 555, + 624 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {r g b} = (1 - \\lambda) \\mathcal {L} _ {1} + \\lambda \\mathcal {L} _ {\\mathrm {D} - \\mathrm {S S I M}}, \\tag {14}", + "image_path": "a7292e880cbc971ab79651bbbbe99ef3eff1499fb17ea3cca5a720ad32cfb0cd.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 633, + 486, + 645 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 633, + 486, + 645 + ], + "spans": [ + { + "bbox": [ + 313, + 633, + 486, + 645 + ], + "type": "text", + "content": "where we use " + }, + { + "bbox": [ + 313, + 633, + 486, + 645 + ], + "type": "inline_equation", + "content": "\\lambda = 0.3" + }, + { + "bbox": [ + 313, + 633, + 486, + 645 + ], + "type": "text", + "content": " for all experiments." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 657, + 395, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 657, + 395, + 670 + ], + "spans": [ + { + "bbox": [ + 313, + 657, + 395, + 670 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "content": "Datasets. We evaluate our method on three different datasets: the Deblur-NeRF [18] synthetic dataset, the Deblur-NeRF real-world dataset, and the DoF-NeRF [41]" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16105" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 63, + 137, + 294, + 259 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 295, + 128 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 295, + 128 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 295, + 128 + ], + "type": "text", + "content": "Table 1. Comparisons on Deblur-NeRF synthetic and real-world scene dataset. We evaluate the performance on three metrics (PSNR, SSIM, LPIPS). “*” denotes the results obtained by reproducing the released code. The orange and yellow cells respectively indicate the highest and second-highest value." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 63, + 137, + 294, + 259 + ], + "lines": [ + { + "bbox": [ + 63, + 137, + 294, + 259 + ], + "spans": [ + { + "bbox": [ + 63, + 137, + 294, + 259 + ], + "type": "table", + "html": "
MethodsSynthetic Scene [18]Real-World Scene [18]
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
Naive NeRF [20]25.950.77910.230322.400.66610.2310
3DGS [9]25.110.74760.214823.380.66550.3140
Deblur-NeRF [18]28.370.85270.118823.470.71990.1207
PDRF-10 [25]30.080.89310.110123.850.73820.1746
DP-NeRF [11]29.330.87130.098723.670.72990.1082
Deblurring 3DGS* [10]28.900.89120.105223.540.73830.1232
BAGS* [26]30.650.91280.063123.480.74080.0962
Ours30.840.92120.047823.700.75310.0825
", + "image_path": "40e329192089d6e354522d1e258dcb776c6db337478a0e09ed574fff99be737f.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 74, + 293, + 281, + 392 + ], + "blocks": [ + { + "bbox": [ + 62, + 273, + 288, + 283 + ], + "lines": [ + { + "bbox": [ + 62, + 273, + 288, + 283 + ], + "spans": [ + { + "bbox": [ + 62, + 273, + 288, + 283 + ], + "type": "text", + "content": "Table 2. Comparisons on DoF-NeRF real-world scene dataset." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 74, + 293, + 281, + 392 + ], + "lines": [ + { + "bbox": [ + 74, + 293, + 281, + 392 + ], + "spans": [ + { + "bbox": [ + 74, + 293, + 281, + 392 + ], + "type": "table", + "html": "
MethodsDoF-NeRF Real-World Scene [41]
PSNR↑SSIM↑LPIPS↓
3DGS [9]25.720.82910.1817
DP-NeRF [11]26.800.81450.1790
Deblurring 3DGS [10]26.580.83860.1708
BAGS [26]29.870.88160.1100
Ours30.140.91270.0701
", + "image_path": "ebb3ab6669fad17098da7b494f089b22ec4289573a1de66f91ef99fd138aedee.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 411, + 295, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 411, + 295, + 567 + ], + "spans": [ + { + "bbox": [ + 54, + 411, + 295, + 567 + ], + "type": "text", + "content": "real-world dataset. The Deblur-NeRF synthetic dataset consists of 5 scenes, each containing defocused images generated using the built-in function of Blender [5]. The Deblur-NeRF real-world dataset includes 10 scenes, with images captured by enlarging the aperture on a Canon EOS RP. The DoF-NeRF real-world dataset comprises 7 scenes, each containing 20 to 30 image triplets. Each triplet includes an all-in-focus image, along with two defocused images focused on the background and foreground. Since DoF-NeRF does not provide training code, we use half of the training defocused images with a foreground focus and the other half with a background focus. All camera poses and initial point clouds are obtained through COLMAP [29, 30]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 574, + 165, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 574, + 165, + 588 + ], + "spans": [ + { + "bbox": [ + 55, + 574, + 165, + 588 + ], + "type": "text", + "content": "5.1. Rendering Results" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "text", + "content": "Quantitative Results. We quantitatively evaluate our CoCoGaussian using the following three metrics: peak signal-to-noise ratio (PSNR), structural similarity index measure (SSIM) [38], and learned perceptual image patch similarity (LPIPS) [45]. We compare our approach to several state-of-the-art methods, including both raytracing [11, 18, 25] and rasterization-based methods [10, 26]. Quantitative results are shown in Tab. 1 and Tab. 2, where our approach achieves the best performance across all metrics except for PSNR on the Deblur-NeRF real-world" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 555, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 289 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 289 + ], + "type": "text", + "content": "dataset. Note that we report the reproduced performance scores for Deblurring 3DGS [10] and BAGS [26] using their official codes. The Deblur-NeRF real-world dataset inherently presents challenges, as there are illumination differences between the sharp and defocused images, leading to relatively lower PSNR score. We discuss this issue in the appendix. For the DoF-NeRF dataset, CoCoGaussian achieves outstanding performance across all metrics, with an approximately 0.04 reduction in LPIPS compared to BAGS. Since the DoF-NeRF dataset has a higher resolution than the Deblur-NeRF dataset, this performance drop in BAGS is likely due to its requirement for manual adjustment of blurring kernel size, highlighting an intrinsic limitation. Conversely, Deblurring 3DGS shows minimal improvement over 3DGS, which can also be attributed to its limited kernel size relative to the higher resolution. The details regarding the time complexity of our model are provided in the appendix." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 304, + 556, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 304, + 556, + 413 + ], + "spans": [ + { + "bbox": [ + 313, + 304, + 556, + 413 + ], + "type": "text", + "content": "Qualitative Results. For qualitative evaluation, we present the rendering results of 3DGS [9], Deblurring 3DGS [10], and BAGS [26] in Fig. 3. Our method demonstrates higher-fidelity results compared to others. Specifically, in the fourth row of our results, the white line on the right side is rendered more sharply, and in the fifth row, the detailed quality of the graphic card appears superior to that of BAGS. Additionally, more CoC visualizations similar to Fig. 1 (b), are in the appendix." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 421, + 408, + 434 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 421, + 408, + 434 + ], + "spans": [ + { + "bbox": [ + 313, + 421, + 408, + 434 + ], + "type": "text", + "content": "5.2. Ablation Study" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 440, + 555, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 440, + 555, + 548 + ], + "spans": [ + { + "bbox": [ + 313, + 440, + 555, + 548 + ], + "type": "text", + "content": "To validate the effectiveness of our proposed methods, we perform a series of ablation studies, with all experiments conducted on the DoF-NeRF real-world dataset. We chose this dataset as it is based on real captured images and does not present the illumination issues found in the Deblur-NeRF real-world dataset, ensuring more reliable evaluation. The results are summarized in Tab. 3, where the baseline denotes naive 3DGS [9]. Additional ablative experiments with other factors are in the appendix." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 564, + 555, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 564, + 555, + 673 + ], + "spans": [ + { + "bbox": [ + 313, + 564, + 555, + 673 + ], + "type": "text", + "content": "Circle of Confusion. To demonstrate the impact of our core concept, CoC, we conduct an experiment that excludes the CoC from the model, retaining only " + }, + { + "bbox": [ + 313, + 564, + 555, + 673 + ], + "type": "inline_equation", + "content": "\\beta_{m}" + }, + { + "bbox": [ + 313, + 564, + 555, + 673 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 564, + 555, + 673 + ], + "type": "inline_equation", + "content": "\\mathbf{d}_m" + }, + { + "bbox": [ + 313, + 564, + 555, + 673 + ], + "type": "text", + "content": ". This configuration yields significantly lower performance than the full model, indicating an over-reliance on parametric learning. The result suggests that without the CoC, which leverages depth and aperture information, the means of the generated Gaussian sets are not optimally arranged, leading to overfitting." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "content": "CoC Direction Vector. In this experiment, we retain the CoC concept but exclude the learnable direction vectors" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "16106" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 71, + 149, + 133 + ], + "blocks": [ + { + "bbox": [ + 57, + 71, + 149, + 133 + ], + "lines": [ + { + "bbox": [ + 57, + 71, + 149, + 133 + ], + "spans": [ + { + "bbox": [ + 57, + 71, + 149, + 133 + ], + "type": "image", + "image_path": "9cb75fc5bbc895292c428596b4c164354427d742a8147dbd1ed06ad66a7319a5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 150, + 72, + 228, + 133 + ], + "blocks": [ + { + "bbox": [ + 150, + 72, + 228, + 133 + ], + "lines": [ + { + "bbox": [ + 150, + 72, + 228, + 133 + ], + "spans": [ + { + "bbox": [ + 150, + 72, + 228, + 133 + ], + "type": "image", + "image_path": "8b6a4f88c0374675cac50e3bfca5115bdccef49500438b090836a1550bb2eb6f.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 230, + 72, + 310, + 133 + ], + "blocks": [ + { + "bbox": [ + 230, + 72, + 310, + 133 + ], + "lines": [ + { + "bbox": [ + 230, + 72, + 310, + 133 + ], + "spans": [ + { + "bbox": [ + 230, + 72, + 310, + 133 + ], + "type": "image", + "image_path": "65dc56d33dd4b64ad77759b39f75049b420ea498977ce01edf43bfaf674f05d3.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 310, + 72, + 391, + 133 + ], + "blocks": [ + { + "bbox": [ + 310, + 72, + 391, + 133 + ], + "lines": [ + { + "bbox": [ + 310, + 72, + 391, + 133 + ], + "spans": [ + { + "bbox": [ + 310, + 72, + 391, + 133 + ], + "type": "image", + "image_path": "e1b3d16adc142ec22e3f5ee7c54d10479dc38d17952d9bad8c813afc3a51beff.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 392, + 72, + 472, + 133 + ], + "blocks": [ + { + "bbox": [ + 392, + 72, + 472, + 133 + ], + "lines": [ + { + "bbox": [ + 392, + 72, + 472, + 133 + ], + "spans": [ + { + "bbox": [ + 392, + 72, + 472, + 133 + ], + "type": "image", + "image_path": "6794c98aa1a6a9a82c103519bcad327ab99e081d3a317e9d74de259022f06f00.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 473, + 72, + 553, + 133 + ], + "blocks": [ + { + "bbox": [ + 473, + 72, + 553, + 133 + ], + "lines": [ + { + "bbox": [ + 473, + 72, + 553, + 133 + ], + "spans": [ + { + "bbox": [ + 473, + 72, + 553, + 133 + ], + "type": "image", + "image_path": "557b78154bcfc77d4c2a283206ef51104a693a1d2b25ae2086605bffda617f57.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 58, + 135, + 149, + 196 + ], + "blocks": [ + { + "bbox": [ + 58, + 135, + 149, + 196 + ], + "lines": [ + { + "bbox": [ + 58, + 135, + 149, + 196 + ], + "spans": [ + { + "bbox": [ + 58, + 135, + 149, + 196 + ], + "type": "image", + "image_path": "ab9e9b898833f7d861a78f8560d9f092cfc5341b88af731176724cfccc0b4a35.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 150, + 135, + 229, + 196 + ], + "blocks": [ + { + "bbox": [ + 150, + 135, + 229, + 196 + ], + "lines": [ + { + "bbox": [ + 150, + 135, + 229, + 196 + ], + "spans": [ + { + "bbox": [ + 150, + 135, + 229, + 196 + ], + "type": "image", + "image_path": "7d7ef3d8dfd972cd5742c3257e390f318020b206d3f3aff185abd7c86ef54233.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 230, + 135, + 309, + 196 + ], + "blocks": [ + { + "bbox": [ + 230, + 135, + 309, + 196 + ], + "lines": [ + { + "bbox": [ + 230, + 135, + 309, + 196 + ], + "spans": [ + { + "bbox": [ + 230, + 135, + 309, + 196 + ], + "type": "image", + "image_path": "44b72f0dab6c517d5c47e13e262da96138a7795bc3c7cc1a9d7418fbc9a07164.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 310, + 135, + 391, + 196 + ], + "blocks": [ + { + "bbox": [ + 310, + 135, + 391, + 196 + ], + "lines": [ + { + "bbox": [ + 310, + 135, + 391, + 196 + ], + "spans": [ + { + "bbox": [ + 310, + 135, + 391, + 196 + ], + "type": "image", + "image_path": "ab67a5fa763d68ad2c7c96a4587da0f5e7e883034440bf36dae723c1c18aec78.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 392, + 135, + 472, + 196 + ], + "blocks": [ + { + "bbox": [ + 392, + 135, + 472, + 196 + ], + "lines": [ + { + "bbox": [ + 392, + 135, + 472, + 196 + ], + "spans": [ + { + "bbox": [ + 392, + 135, + 472, + 196 + ], + "type": "image", + "image_path": "894068c4173df7c1af6ada250a64b63b4fa2855e097b2db1edc5e5e4238440b1.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 473, + 135, + 553, + 196 + ], + "blocks": [ + { + "bbox": [ + 473, + 135, + 553, + 196 + ], + "lines": [ + { + "bbox": [ + 473, + 135, + 553, + 196 + ], + "spans": [ + { + "bbox": [ + 473, + 135, + 553, + 196 + ], + "type": "image", + "image_path": "053df0d43e802824196a3ed7271d0914b67c30c735b137562cf7724b6f796428.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 58, + 198, + 149, + 259 + ], + "blocks": [ + { + "bbox": [ + 58, + 198, + 149, + 259 + ], + "lines": [ + { + "bbox": [ + 58, + 198, + 149, + 259 + ], + "spans": [ + { + "bbox": [ + 58, + 198, + 149, + 259 + ], + "type": "image", + "image_path": "c223c05e7a7216926ee1a9c3526b5daa005c1677ee8c7f6c02bb0c541755c8b9.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 150, + 198, + 229, + 259 + ], + "blocks": [ + { + "bbox": [ + 150, + 198, + 229, + 259 + ], + "lines": [ + { + "bbox": [ + 150, + 198, + 229, + 259 + ], + "spans": [ + { + "bbox": [ + 150, + 198, + 229, + 259 + ], + "type": "image", + "image_path": "f95eb3e7ef4c3776bd9fa2afc9644accf8594658601b64a88cb385bb05f5128a.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 230, + 198, + 309, + 259 + ], + "blocks": [ + { + "bbox": [ + 230, + 198, + 309, + 259 + ], + "lines": [ + { + "bbox": [ + 230, + 198, + 309, + 259 + ], + "spans": [ + { + "bbox": [ + 230, + 198, + 309, + 259 + ], + "type": "image", + "image_path": "7a8c55cdd317cc31b6f5ad765767179cd8c2c38570de47482ffe16983c80338a.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 310, + 198, + 391, + 259 + ], + "blocks": [ + { + "bbox": [ + 310, + 198, + 391, + 259 + ], + "lines": [ + { + "bbox": [ + 310, + 198, + 391, + 259 + ], + "spans": [ + { + "bbox": [ + 310, + 198, + 391, + 259 + ], + "type": "image", + "image_path": "06777173a890b5b4c65c60d3ec37a450cd2a29cb3472edac5a169bf9b801aeab.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 392, + 198, + 472, + 259 + ], + "blocks": [ + { + "bbox": [ + 392, + 198, + 472, + 259 + ], + "lines": [ + { + "bbox": [ + 392, + 198, + 472, + 259 + ], + "spans": [ + { + "bbox": [ + 392, + 198, + 472, + 259 + ], + "type": "image", + "image_path": "7e568ed2bfd40d10661d49024ae43969507bd25150670c5b66977c4e6bfaef0e.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 473, + 198, + 553, + 259 + ], + "blocks": [ + { + "bbox": [ + 473, + 198, + 553, + 259 + ], + "lines": [ + { + "bbox": [ + 473, + 198, + 553, + 259 + ], + "spans": [ + { + "bbox": [ + 473, + 198, + 553, + 259 + ], + "type": "image", + "image_path": "2a9b2ed0bc8edafdb6fe745809757e19ad83d513b8268a1428faf410426bcd46.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 58, + 261, + 149, + 323 + ], + "blocks": [ + { + "bbox": [ + 58, + 261, + 149, + 323 + ], + "lines": [ + { + "bbox": [ + 58, + 261, + 149, + 323 + ], + "spans": [ + { + "bbox": [ + 58, + 261, + 149, + 323 + ], + "type": "image", + "image_path": "54afc9bc38c4814c92f333c69f4d1431b9e6437de2d3cde6a4048556412c4b51.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 150, + 261, + 229, + 323 + ], + "blocks": [ + { + "bbox": [ + 150, + 261, + 229, + 323 + ], + "lines": [ + { + "bbox": [ + 150, + 261, + 229, + 323 + ], + "spans": [ + { + "bbox": [ + 150, + 261, + 229, + 323 + ], + "type": "image", + "image_path": "2344fb31ed42130a2ecb419b1f16dea64c5648d8f37916789bbc87453b0d106a.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 230, + 261, + 309, + 323 + ], + "blocks": [ + { + "bbox": [ + 230, + 261, + 309, + 323 + ], + "lines": [ + { + "bbox": [ + 230, + 261, + 309, + 323 + ], + "spans": [ + { + "bbox": [ + 230, + 261, + 309, + 323 + ], + "type": "image", + "image_path": "a01d3ec93c34a08ce1eb4ab47d979f1121c887b80d99ffc0c79f148c90c03fa8.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 310, + 261, + 391, + 323 + ], + "blocks": [ + { + "bbox": [ + 310, + 261, + 391, + 323 + ], + "lines": [ + { + "bbox": [ + 310, + 261, + 391, + 323 + ], + "spans": [ + { + "bbox": [ + 310, + 261, + 391, + 323 + ], + "type": "image", + "image_path": "fe7f4007f721a1d0a96d29153f7b1fa20da7e65dc14ac4a4a48aceca82da3f2a.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 392, + 261, + 472, + 323 + ], + "blocks": [ + { + "bbox": [ + 392, + 261, + 472, + 323 + ], + "lines": [ + { + "bbox": [ + 392, + 261, + 472, + 323 + ], + "spans": [ + { + "bbox": [ + 392, + 261, + 472, + 323 + ], + "type": "image", + "image_path": "7e956a745bfe1aac5440edbf2cb366c122b1995d2175feb2a6235e8e65d58d0a.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 473, + 261, + 553, + 323 + ], + "blocks": [ + { + "bbox": [ + 473, + 261, + 553, + 323 + ], + "lines": [ + { + "bbox": [ + 473, + 261, + 553, + 323 + ], + "spans": [ + { + "bbox": [ + 473, + 261, + 553, + 323 + ], + "type": "image", + "image_path": "a42e95ff77ea49b01929a4d24b052e1d4a1fef45e4d95d21016c9c99ed1cf938.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 58, + 325, + 149, + 387 + ], + "blocks": [ + { + "bbox": [ + 58, + 325, + 149, + 387 + ], + "lines": [ + { + "bbox": [ + 58, + 325, + 149, + 387 + ], + "spans": [ + { + "bbox": [ + 58, + 325, + 149, + 387 + ], + "type": "image", + "image_path": "38339c5098d520d020d8df9c6030e4692d2574151d39297acdfe39fd3e6f46eb.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 150, + 325, + 229, + 387 + ], + "blocks": [ + { + "bbox": [ + 150, + 325, + 229, + 387 + ], + "lines": [ + { + "bbox": [ + 150, + 325, + 229, + 387 + ], + "spans": [ + { + "bbox": [ + 150, + 325, + 229, + 387 + ], + "type": "image", + "image_path": "0c0ee419b5ed6fa92e8b5f1f375840694022385dc58916864d4dedb7458178a5.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 230, + 325, + 309, + 387 + ], + "blocks": [ + { + "bbox": [ + 230, + 325, + 309, + 387 + ], + "lines": [ + { + "bbox": [ + 230, + 325, + 309, + 387 + ], + "spans": [ + { + "bbox": [ + 230, + 325, + 309, + 387 + ], + "type": "image", + "image_path": "3eec4c0f8400d57ddd6011d77355fbebed540f77af64564e9b1b805d41aea3be.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 310, + 325, + 391, + 387 + ], + "blocks": [ + { + "bbox": [ + 310, + 325, + 391, + 387 + ], + "lines": [ + { + "bbox": [ + 310, + 325, + 391, + 387 + ], + "spans": [ + { + "bbox": [ + 310, + 325, + 391, + 387 + ], + "type": "image", + "image_path": "11b31f0ca21e8765fb539b0dbff3f65261b5764445b1fdf3c952e7ee7158d901.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 392, + 325, + 472, + 387 + ], + "blocks": [ + { + "bbox": [ + 392, + 325, + 472, + 387 + ], + "lines": [ + { + "bbox": [ + 392, + 325, + 472, + 387 + ], + "spans": [ + { + "bbox": [ + 392, + 325, + 472, + 387 + ], + "type": "image", + "image_path": "a7f313060dbd9e68038eb922a9396478fdf13218bff2f8c1aae4e4446b14eb8a.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 473, + 325, + 553, + 387 + ], + "blocks": [ + { + "bbox": [ + 473, + 325, + 553, + 387 + ], + "lines": [ + { + "bbox": [ + 473, + 325, + 553, + 387 + ], + "spans": [ + { + "bbox": [ + 473, + 325, + 553, + 387 + ], + "type": "image", + "image_path": "e6aeafa7bf2c79d8edb837a2f55eac89efd0e259f1a5eeb26a83c04041d8161d.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 58, + 388, + 148, + 450 + ], + "blocks": [ + { + "bbox": [ + 58, + 388, + 148, + 450 + ], + "lines": [ + { + "bbox": [ + 58, + 388, + 148, + 450 + ], + "spans": [ + { + "bbox": [ + 58, + 388, + 148, + 450 + ], + "type": "image", + "image_path": "365029911a67e00921e80b9a067c20f4b7242d200cc03b355e07ee4717cb7fe6.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 82, + 453, + 124, + 463 + ], + "lines": [ + { + "bbox": [ + 82, + 453, + 124, + 463 + ], + "spans": [ + { + "bbox": [ + 82, + 453, + 124, + 463 + ], + "type": "text", + "content": "Reference" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 149, + 388, + 229, + 450 + ], + "blocks": [ + { + "bbox": [ + 149, + 388, + 229, + 450 + ], + "lines": [ + { + "bbox": [ + 149, + 388, + 229, + 450 + ], + "spans": [ + { + "bbox": [ + 149, + 388, + 229, + 450 + ], + "type": "image", + "image_path": "0fc3a7f0655d52d3b1befad0954f9806fb9a6e5aa900c78880a896a8269d5057.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 176, + 453, + 202, + 463 + ], + "lines": [ + { + "bbox": [ + 176, + 453, + 202, + 463 + ], + "spans": [ + { + "bbox": [ + 176, + 453, + 202, + 463 + ], + "type": "text", + "content": "3DGS" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 230, + 388, + 310, + 450 + ], + "blocks": [ + { + "bbox": [ + 230, + 388, + 310, + 450 + ], + "lines": [ + { + "bbox": [ + 230, + 388, + 310, + 450 + ], + "spans": [ + { + "bbox": [ + 230, + 388, + 310, + 450 + ], + "type": "image", + "image_path": "56be89c78cd25de733c32f6ec3046914c77e49f580c82ccb39226f16a8f75ce5.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 235, + 453, + 306, + 464 + ], + "lines": [ + { + "bbox": [ + 235, + 453, + 306, + 464 + ], + "spans": [ + { + "bbox": [ + 235, + 453, + 306, + 464 + ], + "type": "text", + "content": "Deblurring 3DGS" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_caption" + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 310, + 388, + 391, + 450 + ], + "blocks": [ + { + "bbox": [ + 310, + 388, + 391, + 450 + ], + "lines": [ + { + "bbox": [ + 310, + 388, + 391, + 450 + ], + "spans": [ + { + "bbox": [ + 310, + 388, + 391, + 450 + ], + "type": "image", + "image_path": "ba88978bbb98bc401526f99df6b474572ccdb06b74a87ebc9f255b229c7bf6f1.jpg" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 337, + 453, + 365, + 463 + ], + "lines": [ + { + "bbox": [ + 337, + 453, + 365, + 463 + ], + "spans": [ + { + "bbox": [ + 337, + 453, + 365, + 463 + ], + "type": "text", + "content": "BAGS" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_caption" + } + ], + "index": 36 + }, + { + "type": "image", + "bbox": [ + 392, + 388, + 472, + 450 + ], + "blocks": [ + { + "bbox": [ + 392, + 388, + 472, + 450 + ], + "lines": [ + { + "bbox": [ + 392, + 388, + 472, + 450 + ], + "spans": [ + { + "bbox": [ + 392, + 388, + 472, + 450 + ], + "type": "image", + "image_path": "5b0b786477b09fff51e80bdbf6bf408e683240f8ceccb35c26320049e774cda9.jpg" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 421, + 453, + 443, + 463 + ], + "lines": [ + { + "bbox": [ + 421, + 453, + 443, + 463 + ], + "spans": [ + { + "bbox": [ + 421, + 453, + 443, + 463 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 160, + 475, + 449, + 487 + ], + "lines": [ + { + "bbox": [ + 160, + 475, + 449, + 487 + ], + "spans": [ + { + "bbox": [ + 160, + 475, + 449, + 487 + ], + "type": "text", + "content": "Figure 3. Qualitative comparison on the Deblur-NeRF and DoF-NeRF datasets." + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_caption" + } + ], + "index": 38 + }, + { + "type": "image", + "bbox": [ + 473, + 388, + 553, + 450 + ], + "blocks": [ + { + "bbox": [ + 473, + 388, + 553, + 450 + ], + "lines": [ + { + "bbox": [ + 473, + 388, + 553, + 450 + ], + "spans": [ + { + "bbox": [ + 473, + 388, + 553, + 450 + ], + "type": "image", + "image_path": "015d01065fff52da18631ddc2ed73d69933657e4eaa95ba3470a89cfffef32c7.jpg" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 485, + 453, + 541, + 463 + ], + "lines": [ + { + "bbox": [ + 485, + 453, + 541, + 463 + ], + "spans": [ + { + "bbox": [ + 485, + 453, + 541, + 463 + ], + "type": "text", + "content": "Ground Truth" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_caption" + } + ], + "index": 40 + }, + { + "bbox": [ + 55, + 493, + 296, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 493, + 296, + 578 + ], + "spans": [ + { + "bbox": [ + 55, + 493, + 296, + 578 + ], + "type": "inline_equation", + "content": "\\mathbf{d}_m" + }, + { + "bbox": [ + 55, + 493, + 296, + 578 + ], + "type": "text", + "content": ". Instead of using learnable vectors, we use fixed direction vectors, arranging " + }, + { + "bbox": [ + 55, + 493, + 296, + 578 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 493, + 296, + 578 + ], + "type": "text", + "content": " CoC Gaussians evenly in a circular pattern on a plane perpendicular to the vector between the camera position and the base Gaussian. This configuration slightly underperforms compared to the full model, as the fixed directions restrict optimal Gaussian position, preventing adaptation to the most reliable configuration." + } + ] + } + ], + "index": 43 + }, + { + "bbox": [ + 55, + 594, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 296, + 715 + ], + "type": "text", + "content": "CoC Scaling Factor. In this setup, we include the learnable direction vector while omitting the CoC scaling factor " + }, + { + "bbox": [ + 55, + 594, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\beta_{m}" + }, + { + "bbox": [ + 55, + 594, + 296, + 715 + ], + "type": "text", + "content": ". Although the performance is slightly better than when the direction vector is excluded, it remains lower than that of the full model. Without the CoC scaling factor, CoC Gaussian generation becomes overly dependent on depth, leading to errors in CoC diameter calculation when depth values are inaccurate. Including the CoC scaling factor enables the full model to handle uncertain depth more robustly, improving performance. Related visualization re" + } + ] + } + ], + "index": 44 + }, + { + "bbox": [ + 314, + 494, + 419, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 494, + 419, + 506 + ], + "spans": [ + { + "bbox": [ + 314, + 494, + 419, + 506 + ], + "type": "text", + "content": "sults are in the appendix." + } + ] + } + ], + "index": 45 + }, + { + "bbox": [ + 313, + 520, + 555, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 520, + 555, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 520, + 555, + 640 + ], + "type": "text", + "content": "Aperture Parameter. The aperture parameter " + }, + { + "bbox": [ + 313, + 520, + 555, + 640 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 520, + 555, + 640 + ], + "type": "text", + "content": " includes the aperture diameter and focal length, and it plays a critical role in determining the size of the CoC. Without this parameter, there is an upper limit to the CoC diameter calculation, making accurate computation challenging. Experimental results indicate that although performance is somewhat lower than the full model, it is improved over the baseline. This suggests that while the model is unable to determine precise CoC size without " + }, + { + "bbox": [ + 313, + 520, + 555, + 640 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 520, + 555, + 640 + ], + "type": "text", + "content": ", it still learns the focus plane and captures the in-focus region accurately." + } + ] + } + ], + "index": 46 + }, + { + "bbox": [ + 313, + 647, + 511, + 661 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 647, + 511, + 661 + ], + "spans": [ + { + "bbox": [ + 313, + 647, + 511, + 661 + ], + "type": "text", + "content": "5.3. Experiments on All-in-Focus Images." + } + ] + } + ], + "index": 47 + }, + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "content": "We conduct novel view synthesis experiments using all-in-focus images to verify the generalization performance of CoCoGaussian. We utilize the NeRF-LLFF [19, 20] dataset, which contains real-world images. As shown in Tab. 4, our" + } + ] + } + ], + "index": 48 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "16107" + } + ] + } + ], + "index": 49 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 71, + 555, + 220 + ], + "blocks": [ + { + "bbox": [ + 58, + 71, + 555, + 220 + ], + "lines": [ + { + "bbox": [ + 58, + 71, + 555, + 220 + ], + "spans": [ + { + "bbox": [ + 58, + 71, + 555, + 220 + ], + "type": "image", + "image_path": "40134f8f7422f6029deea3284d5dfc03cf87ca4d59036fc79d378a9ed595f953.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 222, + 553, + 245 + ], + "lines": [ + { + "bbox": [ + 55, + 222, + 553, + 245 + ], + "spans": [ + { + "bbox": [ + 55, + 222, + 553, + 245 + ], + "type": "text", + "content": "Figure 4. Visualization of Aperture parameter and Focus Plane Customization. The top row of images decreases the aperture parameter " + }, + { + "bbox": [ + 55, + 222, + 553, + 245 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 222, + 553, + 245 + ], + "type": "text", + "content": " from left to right, while the bottom row moves the focus plane " + }, + { + "bbox": [ + 55, + 222, + 553, + 245 + ], + "type": "inline_equation", + "content": "d_{F}" + }, + { + "bbox": [ + 55, + 222, + 553, + 245 + ], + "type": "text", + "content": " further from the camera from left to right." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 60, + 270, + 293, + 360 + ], + "blocks": [ + { + "bbox": [ + 94, + 255, + 257, + 266 + ], + "lines": [ + { + "bbox": [ + 94, + 255, + 257, + 266 + ], + "spans": [ + { + "bbox": [ + 94, + 255, + 257, + 266 + ], + "type": "text", + "content": "Table 3. Ablation Results of CoCoGaussian." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 60, + 270, + 293, + 360 + ], + "lines": [ + { + "bbox": [ + 60, + 270, + 293, + 360 + ], + "spans": [ + { + "bbox": [ + 60, + 270, + 293, + 360 + ], + "type": "table", + "html": "
MethodsPSNR↑SSIM↑LPIPS↓
Baseline [9]25.720.82910.1817
w/o CoC28.290.87780.0927
w/o CoC direction vector28.910.88930.0896
w/o CoC scaling29.460.90580.0793
w/o aperture parameter27.370.85100.1113
Ours Full30.140.91270.0701
", + "image_path": "9c0d328d9e45c270441b5b1c99adcbd1befc081ed89e7bdfce11116cfeb8657e.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 365, + 296, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 365, + 296, + 485 + ], + "spans": [ + { + "bbox": [ + 55, + 365, + 296, + 485 + ], + "type": "text", + "content": "approach outperforms the baseline 3DGS [9] across all metrics. In practice, when the aperture is very small, radiance from a point on the subject forms a tiny CoC on the image sensor after passing through the aperture, even if the subject is slightly away from the focus plane. Because the CoC is so small, the resulting image is not perceived as defocused. By modeling this small circle, which is difficult to detect with the human eye, our approach achieves higher performance than the baseline 3DGS. Related CoC visualizations are in the appendix." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 491, + 291, + 504 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 491, + 291, + 504 + ], + "spans": [ + { + "bbox": [ + 55, + 491, + 291, + 504 + ], + "type": "text", + "content": "5.4.Customizable Depth of Field and Focus Plane" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 510, + 295, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 295, + 665 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 295, + 665 + ], + "type": "text", + "content": "Customizing the depth of field and focus settings is essential in various applications, such as augmented reality (AR) and virtual reality (VR), where control over depth cues and focus effects can greatly enhance visual realism and viewer engagement. Since our approach incorporates the aperture parameter " + }, + { + "bbox": [ + 55, + 510, + 295, + 665 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 510, + 295, + 665 + ], + "type": "text", + "content": " in CoC modeling, it enables customization of both CoC size and depth of field by adjusting this parameter. The top row in Fig. 4 visualizes increasing depth of field by gradually reducing " + }, + { + "bbox": [ + 55, + 510, + 295, + 665 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 510, + 295, + 665 + ], + "type": "text", + "content": " from left to right. As the focus plane is positioned at a shallow depth, the model consistently captures focus in this region. However, in the leftmost image, the large aperture causes the CoC size to increase at deeper depths, resulting in defocus blur." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "Additionally, our model can customize the focus plane of defocused images, as shown in the bottom row of Fig. 4, where the focus plane shifts from shallow depth on the left to deeper depth on the right. The shallow regions are in fo" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 352, + 270, + 515, + 308 + ], + "blocks": [ + { + "bbox": [ + 343, + 255, + 524, + 266 + ], + "lines": [ + { + "bbox": [ + 343, + 255, + 524, + 266 + ], + "spans": [ + { + "bbox": [ + 343, + 255, + 524, + 266 + ], + "type": "text", + "content": "Table 4. Results on NeRF-LLFF [19, 20] Dataset." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 352, + 270, + 515, + 308 + ], + "lines": [ + { + "bbox": [ + 352, + 270, + 515, + 308 + ], + "spans": [ + { + "bbox": [ + 352, + 270, + 515, + 308 + ], + "type": "table", + "html": "
MethodsPSNR↑SSIM↑LPIPS↓
3DGS [9]27.100.86190.0569
Ours27.760.87380.0372
", + "image_path": "e5843e11b0e48c2007e56aa31c41a1adc97c12f019b21bb9ea01321cb137141b.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 316, + 555, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 316, + 555, + 400 + ], + "spans": [ + { + "bbox": [ + 313, + 316, + 555, + 400 + ], + "type": "text", + "content": "cus on the leftmost image, while deeper areas remain out of focus, and the opposite holds for the rightmost image. This level of control over focus and depth effects is essential for achieving realistic depth cues. Therefore, CoCoGaussian not only demonstrates superior capability in synthesizing sharp novel view images but also provides enhanced flexibility in adjusting aperture and focus plane settings." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 411, + 476, + 423 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 411, + 476, + 423 + ], + "spans": [ + { + "bbox": [ + 313, + 411, + 476, + 423 + ], + "type": "text", + "content": "6. Limitation and Future Work" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 430, + 555, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 430, + 555, + 539 + ], + "spans": [ + { + "bbox": [ + 313, + 430, + 555, + 539 + ], + "type": "text", + "content": "While our model demonstrates strong qualitative and quantitative performance, there is potential for further optimization in the adaptive CoC scaling factor. To ensure training stability, we currently limit the factor size, enabling adaptive optimization mainly when the CoC diameter is overestimated. However, future improvements could enhance adaptability to handle cases where the CoC diameter is underestimated, further improving depth handling in complex scenes without compromising stability." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 550, + 388, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 550, + 388, + 562 + ], + "spans": [ + { + "bbox": [ + 313, + 550, + 388, + 562 + ], + "type": "text", + "content": "7. Conclusion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "content": "In this work, we propose CoCoGaussian, a 3D scene representation method that effectively leverages photographic defocus principles to handle defocused image inputs. CoCoGaussian accurately models defocus blur by constructing the CoC through 3D Gaussians and learnable aperture parameters. We introduce an adaptive CoC Gaussian generation method to address the challenges posed by uncertain depth, making our model robust to reflective or refractive objects. Our approach demonstrates the feasibility and effectiveness of representing 3D scenes from defocused images, advancing possibilities for applications in real-world image synthesis." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16108" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 120 + ], + "type": "text", + "content": "Acknowledgements. This project was supported by the Naver Cloud Corporation and the National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIT) (No. RS-2024-00340745)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 131, + 115, + 144 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 131, + 115, + 144 + ], + "spans": [ + { + "bbox": [ + 56, + 131, + 115, + 144 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 152, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 152, + 294, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 152, + 294, + 217 + ], + "spans": [ + { + "bbox": [ + 61, + 152, + 294, + 217 + ], + "type": "text", + "content": "[1] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5855–5864, 2021. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 219, + 294, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 219, + 294, + 271 + ], + "spans": [ + { + "bbox": [ + 61, + 219, + 294, + 271 + ], + "type": "text", + "content": "[2] Ayan Chakrabarti. A neural approach to blind motion deblurring. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part III 14, pages 221-235. Springer, 2016. 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 274, + 294, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 274, + 294, + 317 + ], + "spans": [ + { + "bbox": [ + 62, + 274, + 294, + 317 + ], + "type": "text", + "content": "[3] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In European Conference on Computer Vision, pages 333-350. Springer, 2022. 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 319, + 294, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 319, + 294, + 361 + ], + "spans": [ + { + "bbox": [ + 62, + 319, + 294, + 361 + ], + "type": "text", + "content": "[4] Wenbo Chen and Ligang Liu. Deblur-gs: 3d gaussian splattering from camera motion blurred images. Proceedings of the ACM on Computer Graphics and Interactive Techniques, 7 (1):1-15, 2024. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 365, + 294, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 365, + 294, + 396 + ], + "spans": [ + { + "bbox": [ + 62, + 365, + 294, + 396 + ], + "type": "text", + "content": "[5] Blender Online Community. Blender - a 3D modelling and rendering package. Blender Foundation, Stichting Blender Foundation, Amsterdam, 2018. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 399, + 294, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 399, + 294, + 453 + ], + "spans": [ + { + "bbox": [ + 62, + 399, + 294, + 453 + ], + "type": "text", + "content": "[6] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5501–5510, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 455, + 292, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 455, + 292, + 465 + ], + "spans": [ + { + "bbox": [ + 62, + 455, + 292, + 465 + ], + "type": "text", + "content": "[7] Eugene Hecht. Optics. Pearson Education India, 2012. 2, 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 467, + 294, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 467, + 294, + 510 + ], + "spans": [ + { + "bbox": [ + 62, + 467, + 294, + 510 + ], + "type": "text", + "content": "[8] Wei Jiang, Kwang Moo Yi, Golnoosh Samei, Oncel Tuzel, and Anurag Ranjan. Neuman: Neural human radiance field from a single video. In European Conference on Computer Vision, pages 402-418. Springer, 2022. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 512, + 294, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 512, + 294, + 554 + ], + "spans": [ + { + "bbox": [ + 62, + 512, + 294, + 554 + ], + "type": "text", + "content": "[9] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42 (4):1-14, 2023. 2, 3, 6, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 557, + 294, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 557, + 294, + 589 + ], + "spans": [ + { + "bbox": [ + 57, + 557, + 294, + 589 + ], + "type": "text", + "content": "[10] Byeonghyeon Lee, Howoong Lee, Xiangyu Sun, Usman Ali, and Eunbyung Park. Deblurring 3d gaussian splatting. arXiv preprint arXiv:2401.00834, 2024. 2, 3, 4, 5, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 591, + 294, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 591, + 294, + 644 + ], + "spans": [ + { + "bbox": [ + 57, + 591, + 294, + 644 + ], + "type": "text", + "content": "[11] Dogyoon Lee, Minhyeok Lee, Chajin Shin, and Sangyoun Lee. Dp-nerf: Deblurred neural radiance field with physical scene priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12386–12396, 2023. 3, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 647, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 294, + 689 + ], + "type": "text", + "content": "[12] Junghe Lee, Donghyeong Kim, Dogyoon Lee, Suhwan Cho, and Sangyoun Lee. Crime: Continuous rigid motion-aware gaussian splatting from motion blur images. arXiv preprint arXiv:2407.03923, 2024. 4, 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 692, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 692, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 692, + 294, + 713 + ], + "type": "text", + "content": "[13] Jungho Lee, Dogyoon Lee, Minhyeok Lee, Donghyung Kim, and Sangyoun Lee. Smurf: Continuous dynam-" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 74, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 333, + 74, + 553, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 74, + 553, + 94 + ], + "spans": [ + { + "bbox": [ + 333, + 74, + 553, + 94 + ], + "type": "text", + "content": "ics for motion-deblurring radiance fields. arXiv preprint arXiv:2403.07547, 2024. 2, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 96, + 553, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 96, + 553, + 161 + ], + "spans": [ + { + "bbox": [ + 317, + 96, + 553, + 161 + ], + "type": "text", + "content": "[14] Tianye Li, Mira Slavcheva, Michael Zollhoefer, Simon Green, Christoph Lassner, Changil Kim, Tanner Schmidt, Steven Lovegrove, Michael Goesele, Richard Newcombe, et al. Neural 3d video synthesis from multi-view video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5521-5531, 2022. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 163, + 553, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 163, + 553, + 215 + ], + "spans": [ + { + "bbox": [ + 317, + 163, + 553, + 215 + ], + "type": "text", + "content": "[15] Zhengqi Li, Simon Niklaus, Noah Snavely, and Oliver Wang. Neural scene flow fields for space-time view synthesis of dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6498-6508, 2021. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 217, + 553, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 217, + 553, + 271 + ], + "spans": [ + { + "bbox": [ + 317, + 217, + 553, + 271 + ], + "type": "text", + "content": "[16] Zhaoshuo Li, Thomas Müller, Alex Evans, Russell H Taylor, Mathias Unberath, Ming-Yu Liu, and Chen-Hsuan Lin. Neuralangelo: High-fidelity neural surface reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8456-8465, 2023. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 273, + 553, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 273, + 553, + 293 + ], + "spans": [ + { + "bbox": [ + 316, + 273, + 553, + 293 + ], + "type": "text", + "content": "[17] Kevin M Lynch and Frank C Park. Modern robotics. Cambridge University Press, 2017. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 295, + 553, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 295, + 553, + 348 + ], + "spans": [ + { + "bbox": [ + 317, + 295, + 553, + 348 + ], + "type": "text", + "content": "[18] Li Ma, Xiaoyu Li, Jing Liao, Qi Zhang, Xuan Wang, Jue Wang, and Pedro V Sander. Deblur-nerf: Neural radiance fields from blurry images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12861–12870, 2022. 2, 3, 5, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 350, + 553, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 350, + 553, + 403 + ], + "spans": [ + { + "bbox": [ + 317, + 350, + 553, + 403 + ], + "type": "text", + "content": "[19] Ben Mildenhall, Pratul P Srinivasan, Rodrigo Ortiz-Cayon, Nima Khademi Kalantari, Ravi Ramamoorthi, Ren Ng, and Abhishek Kar. Local light field fusion: Practical view synthesis with prescriptive sampling guidelines. ACM Transactions on Graphics (ToG), 38(4):1-14, 2019. 7, 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 405, + 553, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 405, + 553, + 470 + ], + "spans": [ + { + "bbox": [ + 317, + 405, + 553, + 470 + ], + "type": "text", + "content": "[20] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I, pages 405-421, 2020. 1, 2, 3, 6, 7, 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 472, + 553, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 472, + 553, + 514 + ], + "spans": [ + { + "bbox": [ + 317, + 472, + 553, + 514 + ], + "type": "text", + "content": "[21] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG), 41(4):1-15, 2022. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 515, + 553, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 515, + 553, + 580 + ], + "spans": [ + { + "bbox": [ + 317, + 515, + 553, + 580 + ], + "type": "text", + "content": "[22] Michael Niemeyer, Jonathan T Barron, Ben Mildenhall, Mehdi SM Sajjadi, Andreas Geiger, and Noha Radwan. Regnerf: Regularizing neural radiance fields for view synthesis from sparse inputs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5480-5490, 2022. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 582, + 553, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 582, + 553, + 635 + ], + "spans": [ + { + "bbox": [ + 317, + 582, + 553, + 635 + ], + "type": "text", + "content": "[23] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5865-5874, 2021. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 637, + 553, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 637, + 553, + 690 + ], + "spans": [ + { + "bbox": [ + 317, + 637, + 553, + 690 + ], + "type": "text", + "content": "[24] Keunhong Park, Utkarsh Sinha, Peter Hedman, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Ricardo MartinBrualla, and Steven M Seitz. Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields. arXiv preprint arXiv:2106.13228, 2021. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "text", + "content": "[25] Cheng Peng and Rama Chellappa. Pdrf: progressively deblurring radiance field for fast scene reconstruction from" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16109" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "text", + "content": "blurry images. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2029-2037, 2023. 2, 3, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 140 + ], + "type": "text", + "content": "[26] Cheng Peng, Yutao Tang, Yifan Zhou, Nengyu Wang, Xijun Liu, Deming Li, and Rama Chellappa. Bags: Blur agnostic gaussian splatting through multi-scale kernel modeling. arXiv preprint arXiv:2403.04926, 2024. 2, 3, 4, 5, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 141, + 294, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 141, + 294, + 196 + ], + "spans": [ + { + "bbox": [ + 56, + 141, + 294, + 196 + ], + "type": "text", + "content": "[27] Sida Peng, Junting Dong, Qianqian Wang, Shangzhan Zhang, Qing Shuai, Xiaowei Zhou, and Hujun Bao. Animatable neural radiance fields for modeling dynamic human bodies. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14314-14323, 2021. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 198, + 294, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 198, + 294, + 251 + ], + "spans": [ + { + "bbox": [ + 56, + 198, + 294, + 251 + ], + "type": "text", + "content": "[28] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-nerf: Neural radiance fields for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10318–10327, 2021. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 254, + 294, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 254, + 294, + 297 + ], + "spans": [ + { + "bbox": [ + 56, + 254, + 294, + 297 + ], + "type": "text", + "content": "[29] Johannes L Schonberger, Enliang Zheng, Jan-Michael Frahm, and Marc Pollefeys. Pixelwise view selection for unstructured multi-view stereo. In European conference on computer vision, pages 501-518. Springer, 2016. 3, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 299, + 294, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 299, + 294, + 331 + ], + "spans": [ + { + "bbox": [ + 56, + 299, + 294, + 331 + ], + "type": "text", + "content": "[30] Qi Shan, Jiaya Jia, and Aseem Agarwala. High-quality motion deblurring from a single image. Acm transactions on graphics (tog), 27(3):1-10, 2008. 3, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 333, + 294, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 333, + 294, + 376 + ], + "spans": [ + { + "bbox": [ + 56, + 333, + 294, + 376 + ], + "type": "text", + "content": "[31] Pratul P Srinivasan, Ren Ng, and Ravi Ramamoorthi. Light field blind motion deblurring. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3958-3966, 2017. 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 378, + 294, + 433 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 378, + 294, + 433 + ], + "spans": [ + { + "bbox": [ + 56, + 378, + 294, + 433 + ], + "type": "text", + "content": "[32] Jiaming Sun, Yiming Xie, Linghao Chen, Xiaowei Zhou, and Hujun Bao. Neuralrecon: Real-time coherent 3d reconstruction from monocular video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15598-15607, 2021. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 434, + 294, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 434, + 294, + 499 + ], + "spans": [ + { + "bbox": [ + 56, + 434, + 294, + 499 + ], + "type": "text", + "content": "[33] Edgar Tretschk, Ayush Tewari, Vladislav Golyanik, Michael Zollhöfer, Christoph Lassner, and Christian Theobalt. Non-rigid neural radiance fields: Reconstruction and novel view synthesis of a dynamic scene from monocular video. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12959-12970, 2021. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 501, + 294, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 501, + 294, + 555 + ], + "spans": [ + { + "bbox": [ + 56, + 501, + 294, + 555 + ], + "type": "text", + "content": "[34] Guangcong Wang, Zhaoxi Chen, Chen Change Loy, and Zwei Liu. Sparsenerf: Distilling depth ranking for few-shot novel view synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9065-9076, 2023. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 557, + 294, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 557, + 294, + 612 + ], + "spans": [ + { + "bbox": [ + 56, + 557, + 294, + 612 + ], + "type": "text", + "content": "[35] Jiepeng Wang, Peng Wang, Xiaoxiao Long, Christian Theobalt, Taku Komura, Lingjie Liu, and Wenping Wang. Neuris: Neural reconstruction of indoor scenes using normal priors. In European Conference on Computer Vision, pages 139-155. Springer, 2022. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 613, + 294, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 613, + 294, + 657 + ], + "spans": [ + { + "bbox": [ + 56, + 613, + 294, + 657 + ], + "type": "text", + "content": "[36] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. arXiv preprint arXiv:2106.10689, 2021. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "type": "text", + "content": "[37] Peng Wang, Lingzhe Zhao, Ruijie Ma, and Peidong Liu. Bad-nerf: Bundle adjusted deblur neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4170–4179, 2023. 2, 3" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 529 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 116 + ], + "type": "text", + "content": "[38] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 118, + 553, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 553, + 183 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 553, + 183 + ], + "type": "text", + "content": "[39] Chung-Yi Weng, Brian Curless, Pratul P Srinivasan, Jonathan T Barron, and Ira Kemelmacher-Shlizerman. Humanerf: Free-viewpoint rendering of moving people from monocular video. In Proceedings of the IEEE/CVF conference on computer vision and pattern Recognition, pages 16210-16220, 2022. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 185, + 553, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 185, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 316, + 185, + 553, + 228 + ], + "type": "text", + "content": "[40] Oliver Whyte, Josef Sivic, Andrew Zisserman, and Jean Ponce. Non-uniform deblurring for shaken images. International journal of computer vision, 98:168-186, 2012. 2, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 229, + 553, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 229, + 553, + 282 + ], + "spans": [ + { + "bbox": [ + 316, + 229, + 553, + 282 + ], + "type": "text", + "content": "[41] Zijin Wu, Xingyi Li, Juewen Peng, Hao Lu, Zhiguo Cao, and Weicai Zhong. Dof-nerf: Depth-of-field meets neural radiance fields. In Proceedings of the 30th ACM International Conference on Multimedia, pages 1718–1729, 2022. 2, 3, 5, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 285, + 553, + 339 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 285, + 553, + 339 + ], + "spans": [ + { + "bbox": [ + 316, + 285, + 553, + 339 + ], + "type": "text", + "content": "[42] Jamie Wynn and Daniyar Turmukhambetov. Diffusionerf: Regularizing neural radiance fields with denoising diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4180-4189, 2023. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 341, + 553, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 341, + 553, + 396 + ], + "spans": [ + { + "bbox": [ + 316, + 341, + 553, + 396 + ], + "type": "text", + "content": "[43] Jiawei Yang, Marco Pavone, and Yue Wang. Freenerf: Improving few-shot neural rendering with free frequency regularization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8254-8263, 2023. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 396, + 553, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 396, + 553, + 441 + ], + "spans": [ + { + "bbox": [ + 316, + 396, + 553, + 441 + ], + "type": "text", + "content": "[44] Wang Yifan, Felice Serena, Shihao Wu, Cengiz Özireli, and Olga Sorkine-Hornung. Differentiable surface splatting for point-based geometry processing. ACM Transactions on Graphics (TOG), 38(6):1-14, 2019. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 441, + 553, + 496 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 441, + 553, + 496 + ], + "spans": [ + { + "bbox": [ + 316, + 441, + 553, + 496 + ], + "type": "text", + "content": "[45] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 498, + 553, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 498, + 553, + 529 + ], + "spans": [ + { + "bbox": [ + 316, + 498, + 553, + 529 + ], + "type": "text", + "content": "[46] Lingzhe Zhao, Peng Wang, and Peidong Liu. Bad-gaussians: Bundle adjusted deblur gaussian splatting. arXiv preprint arXiv:2403.11831, 2024. 3" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16110" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/CoE_ Chain-of-Explanation via Automatic Visual Concept Circuit Description and Polysemanticity Quantification/ed717be1-08cc-4195-89dc-b7ee977099ae_content_list.json b/2025/CoE_ Chain-of-Explanation via Automatic Visual Concept Circuit Description and Polysemanticity Quantification/ed717be1-08cc-4195-89dc-b7ee977099ae_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..497997d8751d8622b64c305b4d79c0240a899e3d --- /dev/null +++ b/2025/CoE_ Chain-of-Explanation via Automatic Visual Concept Circuit Description and Polysemanticity Quantification/ed717be1-08cc-4195-89dc-b7ee977099ae_content_list.json @@ -0,0 +1,1527 @@ +[ + { + "type": "text", + "text": "CoE: Chain-of-Explanation via Automatic Visual Concept Circuit Description and Polysemanticity Quantification", + "text_level": 1, + "bbox": [ + 107, + 128, + 890, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Wenlong Yu Qilong Wang* Chuang Liu Dong Li Qinghua Hu Tianjin Key Lab of Machine Learning, College of Intelligence and Computing, Tianjin University {ywl_95, qlwang, liuc_09, li_dong, huqinghua}@tju.edu.cn", + "bbox": [ + 112, + 203, + 883, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 291, + 326, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Explainability is a critical factor influencing the wide deployment of deep vision models (DVMs). Concept-based post-hoc explanation methods can provide both global and local insights into model decisions. However, current methods in this field face challenges in that they are inflexible to automatically construct accurate and sufficient linguistic explanations for global concepts and local circuits. Particularly, the intrinsic polysemanticity in semantic Visual Concepts (VCs) impedes the interpretability of concepts and DVMs, which is underestimated severely. In this paper, we propose a Chain-of-Explanation (CoE) approach to address these issues. Specifically, CoE automates the decoding and description of VCs to construct global concept explanation datasets. Further, to alleviate the effect of polysemanticity on model explainability, we design a concept polysemanticity disentanglement and filtering mechanism to distinguish the most contextually relevant concept atoms. Besides, a Concept Polysemanticity Entropy (CPE), as a measure of model interpretability, is formulated to quantify the degree of concept uncertainty. The modeling of deterministic concepts is upgraded to uncertain concept atom distributions. Finally, CoE automatically enables linguistic local explanations of the decision-making process of DVMs by tracing the concept circuit. GPT-4o and human-based experiments demonstrate the effectiveness of CPE and the superiority of CoE, achieving an average absolute improvement of $36\\%$ in terms of explainability scores.", + "bbox": [ + 89, + 323, + 485, + 731 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 760, + 220, + 776 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Deep learning-based vision models have demonstrated outstanding performance in various visual tasks, owing to the powerful learning and representation capabilities of deep neural networks (DNNs) [29, 50]. However, the extreme complexity of DNNs, while enhancing their performance, also limits their explainability [41]. Humans are often un", + "bbox": [ + 89, + 785, + 483, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/aa96329ea76bccd11daf890f0a4b579f749490e38481e2465b4b91d59f82d235.jpg", + "image_caption": [ + "Figure 1. Illustration of the integrated process of concept-based post-hoc XAI methods [36], along with two challenges they face." + ], + "image_footnote": [], + "bbox": [ + 524, + 290, + 895, + 468 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "aware of the reasons behind the model's conclusions. This lack of explainability significantly hinders the wide application of deep vision models (DVMs) in critical fields such as autonomous driving and medical diagnosis [7, 28].", + "bbox": [ + 511, + 521, + 906, + 582 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Post-hoc eXplainable Artificial Intelligence (XAI) can enhance the transparency of DVMs without causing degradations in performance [34, 43]. Research in this field can roughly be divided into local and global XAI methods [6]. The former explains either specific stages or the entire decision path with respect to a specific sample, identifying key activated regions or significant semantic concepts [3, 45, 46]. The latter decipher the overall behavior of the DVM in a dataset by concept discovery or visualization techniques [22, 27, 52]. Among them, concept-based methods can explain DVMs from both global and local (glocal) perspectives [2, 4], assuming that each channel or neuron in a certain layer is responsible for recognizing essential semantic concepts. As shown in Fig. 1, they primarily perform one or two of the three operations: decoding Visual Concepts (VCs) of DVMs, manually constructing global concept explanation datasets, and providing local concept circuit explanations for individual samples [14, 16, 26].", + "bbox": [ + 511, + 582, + 908, + 853 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite the significant progress made in concept-based explanation research, certain challenges remain, as illustrated in Fig. 1. First, the VCs are decoded by manually", + "bbox": [ + 511, + 854, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author.", + "bbox": [ + 109, + 887, + 236, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "4364", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "identifying and annotating the commonalities of a set of extracted image patches [2, 26]. These manual methods intuitively become ineffective when faced with new scenarios, new models, and new XAI methods, due to high labor costs, low efficiency, and the limited cognitive scope of individual annotators. Second, most methods make a naive assumption: each concept is monosemantic and user-friendly [33]. However, in practice, VCs encapsulate polysemantic information [23] (e.g., an average of 13 concept atoms are coupled per channel in a ResNet152 model), presenting challenges for the model explanation. Particularly, polysemanticity can serve as a metric for evaluating a model's interpretability. A higher level of polysemanticity signifies a broader conceptual space, increasing the risks of selecting incorrect semantics and generating deviated explanations. Arbitrary concept selection or presenting them solely as images without descriptions yields suboptimal explanations [43]. Both visualizations and linguistic explanations become challenging for humans to comprehend, as it is unclear which semantic is actually represented or holds prominence [16]. Although polysemanticity is crucial to XAI, its systematic quantification and resolution remain unaddressed. All of these challenges collectively exacerbate the difficulty of concept-based explanations for DVMs.", + "bbox": [ + 93, + 90, + 480, + 452 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To alleviate this dilemma, we propose a Chain-of-Explanation (CoE) approach via automatic concept decoding, disentanglement, filtering, description, and polysemanticity quantification. Specifically, CoE first proposes an Automatic Concept decoding and Description (ACD) method to construct global concept explanation datasets. It decodes every VC of key layers through arbitrary advanced XAI methods. A Large Vision Language Model (LVLM) is utilized to describe the commonalities of the extracted image patches since the linguistic explanation is more comprehensible than images [43]. As for the polysemantic concepts, we formulate a Concept Polysemanticity Disentanglement and Filtering (CPDF) mechanism to disentangle them into a set of orthogonal concept atoms, along with a Concept Polysemanticity Entropy (CPE) to quantify their degree of polysemanticity. In this paper, 13 distinct semantic directions are prompted, including low-level and high-level semantics, which is much more than previous work (e.g., 5 in [56]). Within the CPDF, a semantic entailment model is applied to mitigate the impact of redundant semantics. After these operations, the modeling of deterministic concepts is upgraded to uncertain concept atom distributions. When elucidating the local decision-making process, a concept circuit method is utilized to form the basis of an explanation chain, with each node representing a hierarchical set of essential concepts. Relevant atoms and their interconnections are contextually filtered and aligned within the chain. Finally, we consolidate all conceptual information along the explanation chain to generate linguistic explanations similar", + "bbox": [ + 93, + 463, + 480, + 898 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "to a Chain-of-Thought (CoT) pipeline [21, 51], leveraging a Large Language Model (LLM), as shown in Fig. 2. Qualitative and quantitative experiments validate the effectiveness and the superiority of the CoE approach. The contributions of this paper are summarized as follows.", + "bbox": [ + 516, + 90, + 903, + 166 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a novel CoE approach to automatically and systematically provide concept-based glocal explanations, along with managing the problem of polysemantics, in a more interpretable natural language format.", + "- A CPDF mechanism is formulated to handle the hard-to-interpret polysemantic concepts. To our knowledge, CPDF is a pioneer tailored to comprehensively disentangle and quantify polysemanticity. It disentangles them into a set of succinct concept atoms, followed by a filtering function to contextually find the suitable atom as the explanation node. A CPE score is defined to quantify the polysemanticity of concepts, which can be utilized as a metric of the interpretability of concepts or models.", + "- Both GPT-4o and human-based experiments validate the superiority of the CoE in automatically constructing global and local explanations and the effectiveness of CPE in quantifying concept polysemanticity." + ], + "bbox": [ + 511, + 167, + 903, + 424 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 516, + 436, + 653, + 452 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Research on XAI can be broadly categorized into ad-hoc model interpretability and post-hoc explanation studies [43, 54]. The former actively refine DVMs to improve interpretability while also necessitating consideration of a trade-off between performance and interpretability [8, 31, 37, 42, 55]. The latter explains DVMs passively without performance decline [5, 11, 22, 45, 46]. Among these post-hoc efforts, concept-based explanations can provide glacial insights into DVMs [36]. Literature [56] annotated every pixel in special datasets with only 5 kinds of concepts. The explanation is derived by analyzing the activation regions of each neuron. MILAN [26] decoded VCs by calculating the maximum mutual information between channels and regions, followed by manually summarizing commonalities into 3 sentences. However, defining concepts solely through activations is insufficient. CRP decoded VCs through the relevance value derived from the deep Taylor decomposition [35], showing more accurate results than the activation-based methods [2]. The commonalities of the VCs were also identified manually, and one of these commonalities was selected randomly. However, these methods encounter significant challenges, as discussed in Sec. 1. In contrast, this paper proposes an automated concept decoding and description method based on LVLM, which offers greater flexibility in providing linguistic explanations. The most suitable concepts for interpretation are contextually selected, showing superior results.", + "bbox": [ + 516, + 463, + 903, + 869 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Circuits, as sub-graphs, are proposed to explain the key decision route of a model [3, 13, 17, 20, 48]. They identify", + "bbox": [ + 516, + 869, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "4365", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/82d54784ccc7eb026d1a531358c99154558cbe048c177f8a8e964c9a6e8e3789.jpg", + "image_caption": [ + "Figure 2. Outline of the Chain-of-Explanation. It consists of three steps (i.e., ACD, CPDF, and local explanation). Each part utilizes a set of channels $\\mathcal{C}$ from $\\mathcal{M}$ . In the concept circuit, $\\mathcal{P}$ , lines of different depths and colors represent the magnitude of the relevance." + ], + "image_footnote": [], + "bbox": [ + 96, + 87, + 898, + 406 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "the highest-activated concepts and then explain the model's decision by linking all concepts across the circuit. However, its interpretability for DVMs is limited due to the lack of a cohesive, language-based explanatory output. Additionally, the polysemanticity of VCs significantly hinders the accurate explanation [18, 33], and research in this area remains limited. Some methods attempted to indirectly mitigate its impact by predefining principal vectors or subspaces [20, 27, 38]. The lack of flexible and accurate concept-level explanations also restricts their adaptability. Literature [33] analyzed the polysemanticity from an information compression perspective. Pure [16] alleviated the polysemanticity by clustering and decoupling the circuit route. These methods, however, lack the ability to simultaneously and comprehensively quantify, disentangle, and describe polysemantic VCs. Their applicability is limited—challenges that this paper seeks to address, as discussed in Sec. 3.", + "bbox": [ + 88, + 457, + 485, + 715 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Chain-of-Explanation Approach", + "text_level": 1, + "bbox": [ + 89, + 736, + 387, + 753 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this paper, a CoE approach is proposed to automatically explain DVMs from both global and local perspectives. It breaks down the complex reasoning path into a chain of concept nodes, within which polysemanticity is effectively managed. As shown in Fig. 2, CoE is composed of three steps. Step 1: ACD is proposed to automatically construct global concept descriptive explanation datasets of a given DVM. Step 2: With the assistance of an LVLM, a CPDF mechanism is designed to disentangle each original VC into", + "bbox": [ + 89, + 763, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "a set of concept atoms and extract the most suitable atom according to the context, making CoE more comprehensible and flexible. Besides, a CPE score is formulated accordingly to quantify the degree of polysemanticity of concepts. Step 3: A local explanation chain in a manner akin to a CoT for the decision-making progress of the DVM is established based on the dataset, CPDF, concept circuit, and LLMs.", + "bbox": [ + 511, + 457, + 906, + 564 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Automatic Concept Decoding and Description", + "text_level": 1, + "bbox": [ + 511, + 571, + 901, + 588 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ACD method is first formulated to automatically and effectively construct databases of global concept descriptions. Specifically, given an image dataset $\\mathcal{T}$ and a DVM $\\mathcal{M}$ requiring evaluation and explanation, we establish an injective function $f$ between channel set $\\mathcal{C}$ and visual semantic concept set $\\mathcal{S}$ by an XAI method $E_{x}, f: \\mathcal{C} \\mapsto \\mathcal{S}$ . Without loss of generality, for $j_{th}$ channel $c_{j}^{l}$ in layer $l$ , $E_{x}$ extracts a set of image patches within $\\mathcal{T}$ that most activate this channel, named VC $s_{j}^{l} = \\{I_{n}\\}_{N}$ , where $N$ is the number of image patches. Each $I_{n}$ highlights certain regions of the original image, representing the focused concept, as exemplified in the bottom left part of Fig. 2.", + "bbox": [ + 511, + 594, + 906, + 775 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "A describing function $r$ is then defined to decipher $\\{I_n\\}_N$ into a set of language descriptions:", + "bbox": [ + 513, + 775, + 906, + 808 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {D} = r \\left(\\left\\{I _ {n} \\right\\} _ {N}\\right) = \\left\\{d _ {m} \\right\\} _ {M}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 614, + 814, + 906, + 832 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $M$ is the number of language descriptions for the current channel. Traditionally, $r$ is implemented by manually and subjectively identifying commonalities within $\\{I_n\\}_N$ [56], while ACD utilizes a powerful LVLM as the descriptor", + "bbox": [ + 511, + 840, + 908, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "4366", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "with well-engineered prompts. After constructing $S$ , $\\mathcal{D}$ and the mapping function $f$ , a concept description dataset $\\mathcal{B}_{\\mathcal{M},\\mathcal{T}} = \\{\\mathcal{C}, S, \\mathcal{D}, f\\}$ is automatically constructed, providing global linguistic explanations for all VCs of $\\mathcal{M}$ .", + "bbox": [ + 89, + 90, + 483, + 154 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Managing Polysemantic Concept", + "text_level": 1, + "bbox": [ + 89, + 159, + 382, + 176 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "VCs of each node within a DVM exhibit polysemantics, introducing certain deviations in the explanation framework, as shown in Fig. 2. Higher polysemantics indicate that the current VC aggregates more semantics, making the intended representation more ambiguous. Such semantic uncertainty impedes human comprehension of the concept, particularly within the CoE, where various concepts must be understood collectively. Quantifying and solving this dilemma is difficult and has been underestimated by previous studies. Under this circumstance, we propose a CPDF mechanism along with a CPE score to mitigate and quantify its detriments to the CoE simultaneously.", + "bbox": [ + 89, + 183, + 483, + 364 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.1 Polysemantic Concept Decoupling and Filtering", + "text_level": 1, + "bbox": [ + 89, + 382, + 478, + 398 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Considering that a polysemous word can be understood as an aggregation of various distinct meanings, we define a function $g$ to disentangle the ambiguous VC into a set of easily understandable linguistic concept atoms:", + "bbox": [ + 89, + 406, + 483, + 467 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {A} = \\left\\{a ^ {1}, \\dots , a ^ {i}, \\dots , a ^ {Q} \\right\\} = g (s), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 165, + 473, + 482, + 492 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $Q$ is the number of the concept atoms. We posit that each atom represents a view of the current VC $s$ , which will be activated upon the appearance of the corresponding semantic in the input. Furthermore, given that concept atoms should accurately represent the commonalities in $\\{I_n\\}_N$ and recognizing that this naive disentanglement function (Eq. 2) does not provide a foundation for quantifying polysemanticity, we refine the function $g$ to disentangle each image patch $I_n$ into a set of atoms:", + "bbox": [ + 89, + 500, + 483, + 638 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {A} _ {n} = \\left\\{a _ {n} ^ {1}, \\dots , a _ {n} ^ {i}, \\dots , a _ {n} ^ {Q} \\right\\} = g (I _ {n}, s), \\tag {3} \\\\ \\mathcal {D} = \\{\\mathcal {A} _ {n} \\} _ {N}. \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 145, + 643, + 482, + 681 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We instantiate $g$ by a powerful LVLM (e.g., GPT-4o [39]), considering that this is quite an open and complex problem. $g$ prioritizes disentangling and describing commonalities within the VC while analyzing each individual image. Various commonalities among any certain number of image patches should be summarized and described by precise terminologies. These common semantics within an open scenario are as diverse as they are abundant, unfolding in countless forms and nuances. Besides, it should generate specific atoms when any commonalities identified above do not appear in the current $I_{n}$ . Therefore, the art of prompt engineering prompt-com becomes paramount. It should provide rules for recognizing commonalities, possible conceptual directions, and output format control. In", + "bbox": [ + 89, + 689, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "this paper, we prompt 13 high-level and low-level semantical directions that cover the normal concepts encountered in daily life, including object category, scene, object part, color, texture, material, position, transparency, brightness, shape, size, edges, and their relationships. Additionally, the CoT techniques [21, 51] are introduced. Details of all sophisticated prompts are shown in Sec. S2 of the Appendix.", + "bbox": [ + 511, + 90, + 903, + 196 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "All $\\mathcal{A}_n$ are then aggregated into a description atom set $\\mathcal{A}$ with all duplicate atoms removed:", + "bbox": [ + 511, + 198, + 903, + 226 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {A} = \\left\\{\\mathcal {A} _ {1} \\cup \\mathcal {A} _ {2} \\cup \\dots \\cup \\mathcal {A} _ {N} \\right\\} \\tag {4} \\\\ = \\left\\{a ^ {1}, \\dots , a ^ {i}, \\dots , a ^ {P} \\right\\}, \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 604, + 234, + 903, + 272 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $P$ is the number of concept atoms after deleting the repeating atoms. Each element in $\\mathcal{A}$ represents a disentangled monosemantic concept and can be utilized as a node of the CoE. However, some atoms are semantically equivalent even if they have variant descriptions [19]. In order to remove semantically redundant atoms, we introduce an entailment function $\\delta$ to detect the relations between atoms:", + "bbox": [ + 511, + 277, + 903, + 383 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\{1, 0, - 1 \\} = \\delta \\left(a ^ {i 1}, a ^ {i 2}\\right), \\forall a ^ {i 1}, a ^ {i 2} \\in \\mathcal {A}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 573, + 388, + 903, + 407 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where the responses 1,0,-1 represent entailment, neutral, or contradiction of the candidates, respectively. We instantiate $\\delta$ by an NLI model (e.g., DeBERTa model [25]) because the atom candidates are relatively short and simple that an NLI model is good enough. Following [19], a bidirectional entailment check is applied, meaning that one of the candidates will be removed when the response is entailment or neutral. The first atom is designated as the monosemantic representative of the semantic equivalence set due to the transitivity between them. The final concept set $A^{*}$ only contains heterogeneous atoms that describe the current VC:", + "bbox": [ + 511, + 414, + 905, + 580 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {A} ^ {*} = \\left\\{a ^ {1}, \\dots , a ^ {i}, \\dots , a ^ {P ^ {*}} \\right\\}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 602, + 585, + 903, + 612 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "After disentangling the concept atoms, we define a filtering function $z$ to select an atom $a^*$ as the most suitable explanation of current VC with respect to the input $x$ :", + "bbox": [ + 511, + 617, + 903, + 662 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\na _ {s} ^ {*} (x) = z \\left(\\mathcal {A} _ {s} ^ {*}, C (x)\\right). \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 632, + 669, + 903, + 686 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Some research ignores this step or selects the concept randomly with an implicit assumption that all concepts in $\\mathcal{A}^*$ are semantically equivalent [14, 26]. In contrast, we filter the most suitable atom $a^*$ based on the context of the CoE. In particular, $z$ is instantiated by an LLM. $C(x)$ is an image captioning function utilized for providing contextual information. It can be instantiated by a relatively small LVLM [9, 10] or a professional model [30, 49].", + "bbox": [ + 511, + 693, + 905, + 814 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.2 Quantification of Concept Polysemanticity", + "text_level": 1, + "bbox": [ + 511, + 830, + 862, + 848 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The polysemanticity of concepts can serve as a metric for the model's interpretability. Lower polysemanticity of concepts indicates reduced semantic uncertainty with a more", + "bbox": [ + 511, + 854, + 905, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4367", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "singular form of activation, leading to greater interpretability. In this paper, we endeavor to quantify polysemanticity from three levels of granularity (i.e., conceptual channels, layers, and models) by defining a CPE score. The larger the CPE, the more pronounced the polysemanticity, which leads to decreased interpretability.", + "bbox": [ + 89, + 90, + 480, + 181 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In particular, the disentanglement function $g$ in Eq. 3 enables us to count the frequency of each atom in $\\mathcal{A}^*$ . Given two fixed numbers $Q$ and $N$ , we calculate the probability of the $i_{th}$ atom within $j_{th}$ concept:", + "bbox": [ + 89, + 181, + 483, + 242 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\np _ {i} = \\frac {\\text {N u m} _ {i}}{Q \\times N + P a d}. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 212, + 247, + 482, + 280 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Then, the CPE of $j_{th}$ concept can be formulated as", + "bbox": [ + 109, + 284, + 447, + 299 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nH _ {j} ^ {*} = \\frac {- \\sum_ {i = 1} ^ {P ^ {*} + P a d} p _ {i} \\log p _ {i}}{\\log (P ^ {*} + P a d)}. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 187, + 305, + 482, + 342 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The denominator normalizes the entropy value to a range between 0 and 1, considering that the size of $P^*$ varies for different VCs. Besides, the quantity of $A^*$ is set to be at least equal to $N$ due to some $D_s$ having a small number of semantical equivalent atoms with equal probabilities. If there are insufficient unique atoms, we supplement the corresponding number of atoms, with each supplementary atom set to a number of 1. The atomic objects are expanded, and the denominator is increased by the number of supplemented atoms (a total amount of $Pad$ ). Please refer to Sec. S3.1 of the Appendix for details. In addition, two averaged CPEs are formulated to enable comparisons of polysemantics between specific layers or models:", + "bbox": [ + 89, + 347, + 483, + 542 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nH _ {l} ^ {*} = \\frac {1}{d _ {l}} \\sum^ {d _ {l}} H _ {j} ^ {*}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 225, + 549, + 482, + 585 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nH _ {\\mathcal {M}} ^ {*} = \\frac {1}{L} \\sum^ {L} H _ {l} ^ {*}, \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 223, + 592, + 482, + 626 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $d_{l}$ is the dimension of the channel and $L$ is the number of layers. Finally, to accommodate the diverse needs of researchers within the community, the automatically constructed concept explanation dataset is extended as ACD $\\mathcal{B}_{\\mathcal{M},\\mathcal{T}} = \\{\\mathcal{C},\\mathcal{S},\\mathcal{D},\\mathcal{A}^{*},\\mathcal{Q},\\mathcal{H},f\\}$ , where $\\mathcal{Q}$ represents the probability set of each atom and $\\mathcal{H}$ is the set of CPE.", + "bbox": [ + 89, + 628, + 483, + 719 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Construction of Explanation Chain", + "text_level": 1, + "bbox": [ + 89, + 726, + 398, + 742 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Concept circuits in DVMs can identify the most activated units in each layer for a specific input [3]. When aggregating these units across all key layers in a bottom-up manner, a corresponding maximum activation path $\\mathcal{P}$ is generated, establishing the foundation for the explanation chain. Each node in this path serves as the explanation for the model's decision at that layer. Traditional concept circuit methods for DVMs identify the activation path without providing thorough overarching natural language-based explanations. Inspired by the CoT explanation technique in LLMs", + "bbox": [ + 89, + 750, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "[12], CoE proposes a local explanation chain to interpret the decision-making process of DVMs in the form of natural language descriptions based on ACD- $\\mathcal{B}_{\\mathcal{M},\\mathcal{T}}$ , CPDF, concept circuit method, and LLM, as shown in Fig. 2.", + "bbox": [ + 511, + 90, + 903, + 151 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Specifically, some user-predefined layers requiring explanations or key layers facilitating CoE (e.g., output layers of 4 stages of a DVM) are listed in a bottom-up manner. Given an image $x$ and layer $l$ , an advanced XAI method $E_{x}$ computes the maximization normalized activation or relevance values for all concepts. Each value represents the importance of the current concept to the prediction $\\mathcal{M}(x)$ . CoE acknowledges top $k^l$ relevant concepts $\\mathcal{C}^l(x) = \\{c_1^l, \\dots, c_{k^l}^l\\}$ and their values $\\mathcal{V}^l(x) = \\{v_1^l, \\dots, v_{k^l}^l\\}$ , determined by the $\\alpha$ quantile of relevance values across all concepts (i.e., $k^l = |\\{v_j^l | v_j^l > \\alpha \\max(v^l), \\forall j \\in d^l\\}|$ , where $|\\mathcal{X}|$ computes the number of elements). After that, the descriptive concept atoms of the current layer $\\widetilde{\\mathcal{A}}_{k^l}^{*,l}(x) = \\{a_1^{*,l}, \\dots, a_{k^l}^{*,l}\\}$ are constructed based on ACD- $\\mathcal{B}_{\\mathcal{M},\\mathcal{T}}$ and the filter Eq. 7, serving as a node of the explanation chain.", + "bbox": [ + 511, + 152, + 906, + 381 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The final CoE for the local decision-making explanation, abstracting and synthesizing all top activated concept atoms along the chain, is formulated as", + "bbox": [ + 511, + 383, + 905, + 428 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nC o E _ {\\mathcal {M}} (x) = e \\left(\\widetilde {\\mathcal {A}} _ {\\mathcal {P}} ^ {*} (x), \\mathcal {V} _ {\\mathcal {P}} (x), C (x), G (x), \\mathcal {M} (x)\\right), \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 517, + 440, + 905, + 459 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $e$ formulates a synthesizing and describing function with inputs obtained from the above sections, and $G(x)$ is the label of $x$ . We instantiate $CoE_{\\mathcal{M}}(x)$ through a powerful LLM (e.g., GPT-4 [1]), with prompt prompt-coe well-designed by combining CoT and few-shot prompting [44].", + "bbox": [ + 511, + 472, + 905, + 547 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Additionally, these local explanations are evaluated by a powerful LVLM and human judges based on three explainability metrics, each utilizing a three-tier scoring system.", + "bbox": [ + 511, + 549, + 905, + 594 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 611, + 645, + 628 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We evaluate the proposed CoE approach on two types of DVMs (i.e., ResNet [24] and CLIP [40]). The relevance-based XAI method CRP is adopted as the primary approach for $E_{x}$ in this paper, owing to its demonstrated superiority in terms of fidelity and reliability [2]. Please refer to the Appendix for more details on the implementation and results. Codes are available at https://github.com/YuWLong666/CoE.", + "bbox": [ + 511, + 637, + 905, + 758 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Concept Explanation Dataset Construction", + "text_level": 1, + "bbox": [ + 511, + 770, + 879, + 787 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the CoE, we first automatically construct global concept description databases $\\mathsf{ACD - B}_{\\mathcal{M},\\mathcal{T}}$ for various DVMs. Taking a ResNet152 model as an example, 4 output layers of 4 stages (with 256, 512, 1024, and 2048 channels $\\mathcal{C}$ , respectively) are selected as the key layer set. CRP is applied as $E_{x}$ to extract the VC set $\\mathcal{S}$ . Following [2], each VC is represented by 15 (i.e., $N$ ) image patches. The mapping function", + "bbox": [ + 511, + 795, + 905, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "4368", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/f65763e0d8dc88cba8549798805f243cf4a1783d6e014205e1854af5a41178fe.jpg", + "table_caption": [ + "Table 1. Examples of VCs and their disentangled concept atoms, along with their concept probability distributions and CPE scores. The first row shows concept 75 of Layer 4 extracted by CRP. The second row shows concept 697 of Layer 4 extracted by MILAN." + ], + "table_footnote": [], + "table_body": "
Visual Concept and Disentangled Concept AtomsConcept DistributionCPE
Image 1: [burger, lettuce, tomato], \nImage 2: [blue, monkey, face], \nImage 3: [shark, blue, water], \nImage 4: [keyboard, buttons, numbers], \nImage 5: [shark, blue, water], \nImage 6: [shark, blue, water], \nImage 7: [shark, blue, water], \nImage 8: [blue, circular, shape], \nImage 9: [shark, blue, water], \nImage 10: [keyboard, buttons, numbers], \nImage 11: [blue, yellow, objects], \nImage 12: [shark, blue, text], \nImage 13: [shark, blue, water], \nImage 14: [shark, blue, water], \nImage 15: [shark, blue, water]0.30\n0.25\n0.20\n0.15\n0.10\n0.05\n0.00\nhurty green, blue, texture, yellow, objects, water, black0.81
Image 1: [hand, blinds, string], \nImage 2: [curtains, room, grayscale], \nImage 3: [fabric, green, closeup], \nImage 4: [curtains, room, light], \nImage 5: [reflection, glass, light], \nImage 6: [hand, fabric, texture], \nImage 7: [water, fountain, greenery], \nImage 8: [feathers, pattern, dark], \nImage 9: [fabric, blue, closeup], \nImage 10: [fabric, green, seatbelt], \nImage 11: [leaf, bark, texture], \nImage 12: [nature, mountain, grayscale], \nImage 13: [skull, ground, dark], \nImage 14: [fabric, colorful, pattern], \nImage 15: [curtains, window, light]0.12\n0.10\n0.08\n0.06\n0.04\n0.02\n0.00\n0.00\n0.00\n0.00\n0.00\n0.00\n0.00\n0.00\n0.00\n0.00\n0.00\n0.00\n0.00\n0.00\n0.00\n0.00\n0.00\n0.00\n0.00\n0.00\n0.02\n0.02\n0.02\n0.02\n0.02\n0.02\n0.02\n0.02\n0.02\n0.02\n0.02\n0.02\n0.02\n0.02\n0.02\n0.02\n0.02\n0.02\n0.02\n0.02\n0.00\n0.00\n0.00\n0.00\n0.00\n0.00\n0.00\n0.00\n0.00\n0.00\n0.00\n0.00\n0.00\n0.00\n0.00\n0.00\n0.00\n0.00\n0.02\n0.00\n0.00\n0.00\n0.00\n0.00\n0.00\n0.00\n0.00\n0.00\n0.00\n0.00\n0.00\n0.00\n0.00\n0.00\n0.00\n0.00\n0.00\n0.01\n0.01\n0.01\n0.01\n0.01\n0.01\n0.01\n0.01\n0.01\n0.01\n0.01\n0.01\n0.01\n0.01\n0.01\n0.01\n0.01\n0.01\n0.01\n0.01\n0.02\n0.02\n0.02\n0.02\n0.02\n0.02\n0.02\n0.02\n0.02\n0.02\n0.02\n0.02\n0.02\n0.02\n0.02\n0.02\n0.02\n0.02\n0.02\n0.03\n0.03\n0.03\n0.03\n0.03\n0.03\n0.03\n0.03\n0.03\n0.03\n0.03\n0.03\n0.03\n0.03\n0.03\n0.03\n0.03\n0.03\n0.03\n0.03\n0.04\n0.04\n0.04\n0.04\n0.04\n0.04\n0.04\n0.04\n0.04\n0.04\n0.04\n0.04\n0.04\n0.04\n0.04\n0.04\n0.04\n0.04\n0.04\n0.04\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.08\n0.08\n0.08\n0.08\n0.08\n0.08\n0.08\n0.08\n0.08\n0.08\n0.08\n0.08\n0.08\n0.08\n0.08\n0.08\n0.08\n0.08\n0.08\n0.08\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.08\n0.08\n0.08\n0.08\n0.08\n0.08\n0.08\n0.08\n0.08\n0.08\n0.08\n0.08\n0.08\n0.08\n0.08\n0.08\n0.08\n0.08\n0.08\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.04\n0.04\n0.04\n0.04\n0.04\n0.04\n0.04\n0.04\n0.04\n0.04\n0.04\n0.04\n0.04\n0.04\n0.04\n0.04\n0.04\n0.04\n0.04\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.08\n0.08\n0.08\n0.08\n0.08\n0.08\n0.08\n0.08\n0.08\n0.08\n0.08\n0.08\n0.08\n0.08\n0.08\n0.08\n0.08\n0.08\n0.08\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.04\n0.04\n0.04\n0.04\n0.04\n0.04\n0.04\n0.04\n0.04\n0.04\n0.04\n0.04\n0.04\n0.04\n0.04\n0.04\n0.04\n0.04\n0.04\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.02\n0.12\n0.12\n0.12\n0.12\n0.12\n0.12\n0.12\n0.12\n0.12\n0.12\n0.12\n0.12\n0.12\n0.12\n0.12\n0.12\n0.12\n0.12\n0.12\n0.12\n0.25\n0.25\n0.25\n0.25\n0.25\n0.25\n0.25\n0.25\n0.25\n0.25\n0.25\n0.25\n0.25\n0.25\n0.25\n0.25\n0.25\n0.25\n0.25\n0.25\n0.26\n0.26\n0.26\n0.26\n0.26\n0.26\n0.26\n0.26\n0.26\n0.26\n0.26\n0.26\n0.26\n0.26\n0.26\n0.26\n0.26\n0.26\n0.26\n0.26\n0.25\n0.25\n0.25\n0.25\n0.25\n0.25\n0.25\n0.25\n0.25\n0.25\n0.25\n0.25\n0.25\n0.25\n0.25\n0.25\n0.25\n0.25\n0.25\n0.2.\n12\n12\n12\n12\n12\n12\n12\n12\n12\n12\n12\n12\n12\n12\n12\n12\n12\n12\n12\n12\n12\n12\n12\n12\n12\n12\n12\n12\n12\n12\n12\n12\n12\n12\n25\n25\n25\n25\n25\n25\n25\n25\n25\n25\n25\n25\n25\n25\n25\n25\n25\n25\n25\n25\n25\n25\n25\n25\n25\n25\n25\n25\n25\n25\n25\n25\n25\n25
", + "bbox": [ + 94, + 127, + 901, + 479 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "$f$ from channels to VCs is built. We then employ a GPT4o-2024-08-06 model as the function $g$ to generate the description set $\\mathcal{D}$ (i.e., Eq. 3). The disentangled and clustered atom set $\\mathcal{A}^*$ (processed by a DeBERTa model), the probability set $\\mathcal{Q}$ , and the CPE set $\\mathcal{H}$ are computed and gathered accordingly. Overall, ACD- $\\mathcal{B}_{\\mathcal{M},\\mathcal{T}}$ containing 3,840 entries is constructed, with each entry comprising a channel, a CPE score, and 15 VC representations—each consisting of 3 concept atoms (i.e., $Q$ )—and multiple orthogonal linguistic atoms, along with their associated probabilities.", + "bbox": [ + 89, + 491, + 483, + 643 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Disentanglement of Polysemantic Concept", + "text_level": 1, + "bbox": [ + 89, + 650, + 450, + 667 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "CPDF automatically disentangles and describes the polysemantic VCs into a set of linguistic concept atoms, $\\mathcal{D}$ . As exemplified by Table 1, two randomly selected VCs are extracted by two advanced XAI methods (CRP and MILAN). They are disentangled and described into atoms, exhibiting different degrees of polysemasticity. We observe that the commonalities within the VCs are almost entirely identified, and most of the atoms describe the commonalities of subsets of image patches. In the first row (CPE=0.81), $g$ successfully identifies all visually discernible commonalities, such as shark and water. The atoms exhibit high consistency, and fewer but more frequently occurring common semantics are identified. Only a few atoms, such as the burger, are not commonalities, as their corresponding images lack sufficient common semantics. The atoms are also repre", + "bbox": [ + 89, + 674, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "sented accurately. For instance, the atom sets for each image featuring the shark semantic all include the word shark. In the second row (CPE=0.95), our method remains effective in disentangling and describing VC with a higher level of polysemantics. $g$ identifies a greater variety of atoms. These results demonstrate the effectiveness of the disentanglement and description operation in the CPDF.", + "bbox": [ + 511, + 491, + 906, + 599 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Concept Distribution and CPE", + "text_level": 1, + "bbox": [ + 511, + 604, + 785, + 622 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "CPE can quantify the polysemanticity and serve as an indicator of a DVM's interpretability. It begins by calculating the probabilities of each orthogonal atom. As the middle column of Table 1 shows, the distribution provides statistics on the extent to which different atoms are associated with each concept. For instance, concept 75 primarily focuses on blue with a probability of 0.27 while also identifying yellow with a smaller probability. The semantics between atoms are non-overlapping. This probabilistic approach enhances explainability by capturing multiple semantics with their occurring probabilities, upgrading the modeling of deterministic concepts to uncertain atom distributions.", + "bbox": [ + 511, + 628, + 906, + 809 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "CPE quantifies polysemanticity from 3 levels of granularity. As shown in Fig. 3a, concepts exhibit varying degrees of polysemanticity, with considerable fluctuations (ranging from a maximum of 1 to a minimum of 0.5). Most of the concepts demonstrate a noticeable level of polysemanticity. When analyzing the polysemanticity of different", + "bbox": [ + 511, + 810, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "4369", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/837cd38290cbb3b4ff70a449f9fa6f7065fa0250c4ccccc3b23df402c253ae85.jpg", + "image_caption": [ + "(a) Sorted CPE along the channel dimension." + ], + "image_footnote": [], + "bbox": [ + 96, + 89, + 361, + 243 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e26469c49c6efe022b123a098cd7322cb4218aa12f003390f9d1044ee078a1e0.jpg", + "image_caption": [ + "(b) Averaged CPE on 4 key layers." + ], + "image_footnote": [], + "bbox": [ + 364, + 92, + 633, + 243 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e8ed054dee5d0031be4da582a14299832de27b55662fe516aa50fa044e9a1663.jpg", + "image_caption": [ + "(c) Averaged CPE on 3 models.", + "Figure 3. Illustrations of CPE scores at 3 levels of granularity. In (a), the channel indices of each layer are max normalized to 1." + ], + "image_footnote": [], + "bbox": [ + 642, + 90, + 898, + 244 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/34ec9f81003292b9fa199d3f1b485a5b42396cac8c156878d62ffb434c890dc9.jpg", + "table_caption": [ + "Table 2. The consistency of the recognition of concept polysemantics between the CPE metric and human scorers." + ], + "table_footnote": [], + "table_body": "
Consistency
CPE Metric and Human Scorers75 %
", + "bbox": [ + 93, + 327, + 473, + 375 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "layers, as shown in Fig. 3b, the averaged CPE peaks in the third stage, while the shallowest and deepest layers exhibit lower values. Typically, deeper layers capture abstract concepts, whereas the shallower layers extract elementary features like textures [53, 56]. DVMs show better monosemanticity when identifying elementary and abstract concepts, while more ambiguous concepts are captured in the intermediate layers. This experiment validates the theoretical and experimental claims of previous work [14, 32]. In Fig. 3c, we observe that as the number of layers increases, the CPE of the model shows a rising trend. This result indicates a direct proportional relationship between the polysemanticity and the complexity of DVMs, which is that the more complex a DVM becomes, the harder it is to interpret.", + "bbox": [ + 88, + 385, + 482, + 595 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The CPE is evaluated in comparison with human assessments. As shown in Table 1, the CPE is consistent with the degree of polysemanticity from the human intuitive observations (the polysemanticity of the first row is smaller than that of the second row). Similarly, we randomly sample 300 pairs of VCs, each containing two VCs with varying levels of polysemanticity. Comparisons are drawn from these 300 pairs by both human evaluators and the CPE metric. As shown in Table 2, results from human assessments and CPE achieve a $75\\%$ agreement rate, demonstrating the effectiveness of CPE. Other results are shown in Sec. S4 of the Appendix. All of these results reveal that CPE is effective in quantifying the polysemanticity of concepts and DVMs.", + "bbox": [ + 89, + 597, + 482, + 792 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. CoE for Local Explanations", + "text_level": 1, + "bbox": [ + 89, + 801, + 341, + 816 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "CoE generates local explanations akin to CoT by automatically describing the concept circuit. We set $\\alpha = 0.001$ to acknowledge the top relevant VCs, resulting in an average of 3 key VCs per layer. As exemplified in Table 3, given an input image, CoE first outputs whether the prediction of the", + "bbox": [ + 89, + 824, + 483, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "DVM is correct and then gives reasons behind this prediction according to the information along the concept circuits. CoE employs an LLM as an engine to infer the logical relationships among concepts, as well as between concepts and the context. It synthesizes all evidence to generate an inductive commentary on the predictions. The explanation chain is presented in a linguistic format, enhancing human understanding. For instance, CoE explains why the DVM incorrectly predicted a tench as a mosquito net: the top activation of mosquito-net-related concepts, rather than those related to fish, misled the decision. Additional instances are shown in Sec. S5 of the Appendix. Overall, CoE effectively captures and elucidates the core decision-making evidence for both correct and incorrect predictions of DVMs.", + "bbox": [ + 511, + 297, + 906, + 510 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Quantitative Evaluation via GPT-4o & Human", + "text_level": 1, + "bbox": [ + 511, + 525, + 905, + 541 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We employ GPT-4o and human evaluators to assess the local explanations for 500 and 100 samples randomly selected from the ImageNet Validation dataset [15], respectively. The GPT-4o evaluation is achieved through a carefully designed prompt prompt-coe-eval. Explanations generated by three methods are rated according to three explainability metrics: Accuracy, Completeness, and User Interpretability [47]. Each is assigned a three-tier scoring system (2, 1, and 0), with specific criteria defined for each tier, resulting in a total explanation score of 6. Please refer to Sec. S2 and S5 of the Appendix for details.", + "bbox": [ + 511, + 551, + 905, + 717 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As shown in Table 4, our CoE approach attains peak performance across all three metrics, achieving total scores of 5.06 and 5.04 in GPT-4o and human evaluations, respectively. The lowest scores (3.14 and 2.62) are obtained on the MILAN ANNOTATION dataset without the ACD and CPDF methods. This dataset was manually annotated, resulting in inaccuracy and insufficiency within concept explanations. Scores increase by 0.11 and 0.72, respectively, with the application of our proposed ACD- $\\mathcal{B}_{\\mathcal{M},\\mathcal{T}}$ dataset. The ACD method not only enables the automated disentanglement and description of concept atoms but also enhances their precision and comprehensibility. When the CoE em", + "bbox": [ + 511, + 719, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "4370", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Cases of local explanations of CoE. The first row shows the explanation of a correct prediction, and the second row is reversed.", + "bbox": [ + 96, + 89, + 898, + 103 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6649d862f2f93427734b895aff043653f785e8b10d9b9ad65eed48d8fc7b2ee9.jpg", + "image_caption": [ + "label:" + ], + "image_footnote": [], + "bbox": [ + 96, + 119, + 285, + 231 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e6e0418a4bb3aa96e7aba5d4fc71145eef7293294ccc89b41a5b018021469042.jpg", + "image_caption": [ + "black-and-tan coonhound", + "predict:", + "black-and-tan coonhound" + ], + "image_footnote": [], + "bbox": [ + 99, + 232, + 145, + 279 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The model outputs a correct result: black-and-tan coonhound. Specifically, in the shallowest layer, layer 1, channel 88 with a relevance value of 1.0 describes the concept curved, which may relate to the curved shapes observed in parts of the dog such as its tail or ears. Moving deeper, in layer 2, channel 506 with a relevance value of 0.85 and channel 391 with a relevance value of 0.75 describe the concept of animal, indicating the model has identified an animal in the image, and channel 229 with a relevance value of 0.77 emphasizes the color black, aligning with the black and tan colors of the coonhound. Further, layer 2 also identifies the presence of grass several times (channels 94 and 103), which corresponds to the grassy field described in the image caption. In layer 3, various channels, such as 985, 174, 422, and 13, with high relevance scores, from 0.72 to 1.0, repeatedly identify the color green and the concept grass, firmly establishing the setting of the image. Additionally, channels 11, 787, and 591 in layer 3 enhance the certainty by repeatedly identifying animal and creature, confirming that the subject is an animal. Finally, layer 4 decisively contributes with channels 1354 and 1661, both signifying dog with high relevance values (1.0 and 0.72 respectively), which correctly identifies the animal in the image as a dog.", + "bbox": [ + 289, + 123, + 901, + 253 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Given that the model's structured concepts strongly match the description and characteristics noted in the caption and the ground truth label, it is clear why the model outputs the correct result of black-and-tan coonhound.", + "bbox": [ + 290, + 255, + 901, + 276 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3f9920565c7c12df94010542ee15f666995976bb18fbb963c0e8f48eb2c8fc38.jpg", + "image_caption": [ + "tench", + "mosquito net" + ], + "image_footnote": [], + "bbox": [ + 96, + 289, + 287, + 400 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1982e38d7a411595f21ae29d6975d4eb3f1390c4182d501fe2578a0618818a49.jpg", + "image_caption": [ + "label:", + "predict:" + ], + "image_footnote": [], + "bbox": [ + 101, + 400, + 148, + 446 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The model outputs an incorrect result mosquito net instead of the correct label tench. Specifically, in the shallowest layer 1 of the model, channel 88 with a relevance value of 1.0 describes the concept of smooth, channel 61 with a relevance value of 0.81 describes the concept of green, and channel 111 with a relevance value of 0.72 describes the concept of brown. In the deeper layer 2 of the model, channel 94 with a relevance value of 1.0 describes the concept of water, channel 229 with a relevance value of 0.84 describes the concept of fish, channel 103 with a relevance value of 0.81 also describes the concept of fish, and channel 467 with a relevance value of 0.73 describes the concept of mesh. At layer 3, channel 985 with a relevance value of 1.0 further describes the concept of fish, while channels 894, 122, and 715 with relevance values of 0.83, 0.73, and 0.72, respectively, describe the concept of green, possibly relating to the color of the net. In the deepest layer 4, channel 697 with a relevance value of 1.0 describes the concept of net, channel 1878 with a relevance value of 0.76 describes the concept of baby, and channel 321 with a relevance value of 0.76 represents the concept of hexagon, which could misconstrue the net structure. The prevalence of concepts related to fish within the decision path suggests a relationship to the tench, but the final prediction of mosquito net was likely influenced by the dominance of net-related concepts. Thus, the model was misled by the presence of a net, resulting in the incorrect classification of mosquito net.", + "bbox": [ + 290, + 292, + 901, + 446 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/09c80edd499595a3a5ba2f76b1b1fcb08b581874b092b043ce6228e05a853191.jpg", + "table_caption": [ + "Table 4. Scores of CoE for the local explanation on three methods evaluated by GPT-4o and human scorers. The first method utilizes MILAN ANNOTATION as the concept explanation dataset, while the rest two work on the ACD- $\\mathcal{B}_{\\mathcal{M},\\mathcal{T}}$ datasets. Each criterion has a maximum score of 2 points. The total explanation score is 6 points. Higher scores indicate better explainability." + ], + "table_footnote": [], + "table_body": "
MethodAccuracyCompletenessUser Interpret.Total Explanation
GPT-4oMILAN A.[26] + Description (Baseline)1.031.071.043.14
CoE w/o Filtering1.071.101.083.25
CoE (ours)1.691.701.675.06
HumanMILAN A.[26] + Description (Baseline)0.870.830.922.62
CoE w/o Filtering1.101.161.073.34
CoE (ours)1.731.721.595.04
", + "bbox": [ + 94, + 517, + 898, + 646 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ploys the full CPDF mechanism, the explanation score rises by 1.92 and 2.42 points (the average absolute improvement is $36\\%$ ), respectively, with the filtering step contributing 1.81 and 1.70 points. This result suggests that by taking polysemanticity and contextual filtering of concepts into account, CoE effectively guides the LLM in producing better linguistic explanations, substantially enhancing their accuracy, completeness, and interpretability for users. The results of the human evaluation are consistent with the GPT-4o results. All of these results strongly affirm the superiority of the proposed CoE approach.", + "bbox": [ + 89, + 664, + 482, + 830 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 89, + 844, + 209, + 859 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we propose a CoE approach to tackle the challenges of the inflexibility to automatically construct con", + "bbox": [ + 89, + 869, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "cept explanation datasets, insufficient linguistic explanations, and the weakness in managing concept polysemantics of concept-based XAI methods. CoE automates the decoding and description of concepts via an LVLM and constructs global concept explanation datasets. A CPDF mechanism is designed to identify the most contextually relevant concept atoms, advancing the modeling of deterministic concepts to uncertain concept atom distributions. The CPE is proposed to quantify the polysemanticsy. Finally, CoE enables local explanations of a DVM's decisionmaking process, represented through natural language. The effectiveness and superiority of CPE and CoE are demonstrated. In summary, CoE establishes a systematical framework for explaining DVMs, and we believe that the findings presented in this paper warrant further exploration.", + "bbox": [ + 511, + 664, + 906, + 891 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "4371", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgment", + "text_level": 1, + "bbox": [ + 91, + 90, + 240, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This research is supported in part by the National Natural Science Foundation of China under Grants U23B2049, 62276186, and 61925602.", + "bbox": [ + 89, + 114, + 482, + 160 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 176, + 186, + 193 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023.5", + "[2] Reduan Achtibat, Maximilian Dreyer, Ilona Eisenbraun, Sebastian Bosse, Thomas Wiegand, Wojciech Samek, and Sebastian Lapuschkin. From attribution maps to human-understandable explanations through concept relevance propagation. Nature Machine Intelligence, 5(9):1006-1019, 2023. 1, 2, 5", + "[3] Rajaram Achyuta, Chowdhury Neil, Torralba Antonio, Andreas Jacob, and Schwettmann Sarah. Automatic discovery of visual circuits. In NeurIPS Workshop on Attributing Model Behavior at Scale, 2023. 1, 2, 5", + "[4] Sebastian Bach, Alexander Binder, Grégoire Montavon, Frederick Klauschen, Klaus-Robert Müller, and Wojciech Samek. On pixel-wise explanations for non-linear classifier decisions by layer-wise relevance propagation. PLoS ONE, 10, 2015. 1", + "[5] David Bau, Jun-Yan Zhu, Hendrik Strobelt, Agata Lapedriza, Bolei Zhou, and Antonio Torralba. Understanding the role of individual units in a deep neural network. Proceedings of the National Academy of Sciences, 117(48):30071-30078, 2020. 2", + "[6] Leonard Bereska and Efstratios Gavves. Mechanistic interpretability for ai safety-a review. arXiv preprint arXiv:2404.14082, 2024. 1", + "[7] Subrato Bharati, M Rubaiyat Hossain Mondal, and Prajoy Podder. A review on explainable artificial intelligence for healthcare: why, how, and when? IEEE Transactions on Artificial Intelligence, 2023. 1", + "[8] Moritz Böhle, Navdeeppal Singh, Mario Fritz, and Bernt Schiele. B-cos alignment for inherently interpretable cnns and vision transformers. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(6):4504-4518, 2024. 2", + "[9] Zhe Chen, Weiyun Wang, Hao Tian, Shenglong Ye, Zhang-wei Gao, Erfei Cui, Wenwen Tong, Kongzhi Hu, Jiapeng Luo, Zheng Ma, et al. How far are we to gpt-4v? closing the gap to commercial multimodal models with open-source suites. arXiv preprint arXiv:2404.16821, 2024. 4", + "[10] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24185–24198, 2024. 4", + "[11] Ming-Ming Cheng, Peng-Tao Jiang, Ling-Hao Han, Liang Wang, and Philip Torr. Deeply explain cnn via hierarchical" + ], + "bbox": [ + 93, + 202, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "decomposition. International Journal of Computer Vision, 131(5):1091-1105, 2023. 2", + "[12] Zheng Chu, Jingchang Chen, Qianglong Chen, Weijiang Yu, Tao He, Haotian Wang, Weihua Peng, Ming Liu, Bing Qin, and Ting Liu. Navigate through enigmatic labyrinth a survey of chain of thought reasoning: Advances, frontiers and future. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1173–1203, 2024. 5", + "[13] Arthur Conmy, Augustine Mavor-Parker, Aengus Lynch, Stefan Heimersheim, and Adrià Garriga-Alonso. Towards automated circuit discovery for mechanistic interpretability. Advances in Neural Information Processing Systems, 36: 16318-16352, 2023. 2", + "[14] Meghal Dani, Isabel Rio-Torto, Stephan Alaniz, and Zeynep Akata. Devil: Decoding vision features into language. In DAGM German Conference on Pattern Recognition, pages 363-377. Springer, 2023. 1, 4, 7", + "[15] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 7", + "[16] Maximilian Dreyer, Erblina Purelku, Johanna Vielhaben, Wojciech Samek, and Sebastian Lapuschkin. Pure: Turning polysemantic neurons into pure features by identifying relevant circuits. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 8212-8217, 2024. 1, 2, 3", + "[17] Nelson Elhage, Neel Nanda, Catherine Olsson, Tom Henighan, Nicholas Joseph, Ben Mann, Amanda Askell, Yuntao Bai, Anna Chen, Tom Conerly, et al. A mathematical framework for transformer circuits. Transformer Circuits Thread, 1(1):12, 2021. 2", + "[18] Nelson Elhage, Tristan Hume, Catherine Olsson, Nicholas Schiefer, Tom Henighan, Shauna Kravec, Zac Hatfield-Dodds, Robert Lasenby, Dawn Drain, Carol Chen, et al. Toy models of superposition. arXiv preprint arXiv:2209.10652, 2022.3", + "[19] Sebastian Farquhar, Jannik Kossen, Lorenz Kuhn, and Yarin Gal. Detecting hallucinations in large language models using semantic entropy. Nature, 630(8017):625-630, 2024. 4", + "[20] Thomas Fel, Victor Boutin, Louis Béthune, Rémi Cadène, Mazda Moayeri, Léo Andéol, Mathieu Chalvidal, and Thomas Serre. A holistic approach to unifying automatic concept extraction and concept importance estimation. Advances in Neural Information Processing Systems, 36, 2024. 2, 3", + "[21] Guhao Feng, Bohang Zhang, Yuntian Gu, Haotian Ye, Di He, and Liwei Wang. Towards revealing the mystery behind chain of thought: a theoretical perspective. Advances in Neural Information Processing Systems, 36, 2024. 2, 4", + "[22] Amirata Ghorbani, James Wexler, James Y Zou, and Been Kim. Towards automatic concept-based explanations. Advances in neural information processing systems, 32, 2019. 1, 2", + "[23] Dan W Grupe and Jack B Nitschke. Uncertainty and anticipation in anxiety: an integrated neurobiological and psy" + ], + "bbox": [ + 516, + 92, + 903, + 902 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "4372", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "chological perspective. Nature Reviews Neuroscience, 14(7): 488-501, 2013. 2", + "[24] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 770-778, 2016. 5", + "[25] Pengcheng He, Xiaodong Liu, Jianfeng Gao, and Weizhu Chen. Deberta: Decoding-enhanced bert with disentangled attention. In International Conference on Learning Representations, 2020. 4", + "[26] Evan Hernandez, Sarah Schwettmann, David Bau, Teona Bagashvili, Antonio Torralba, and Jacob Andreas. Natural language descriptions of deep features. In International Conference on Learning Representations, 2022. 1, 2, 4, 8", + "[27] Been Kim, Martin Wattenberg, Justin Gilmer, Carrie Cai, James Wexler, Fernanda Viegas, et al. Interpretability beyond feature attribution: Quantitative testing with concept activation vectors (tcav). In International conference on machine learning, pages 2668-2677. PMLR, 2018. 1, 3", + "[28] Anton Kuznietsov, Balint Gyevnar, Cheng Wang, Steven Peters, and Stefano V. Albrecht. Explainable ai for safe and trustworthy autonomous driving: A systematic review. IEEE Transactions on Intelligent Transportation Systems, pages 1-23, 2024. 1", + "[29] Yann LeCun, Yoshua Bengio, and Geoffrey Hinton. Deep learning. nature, 521(7553):436-444, 2015. 1", + "[30] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730-19742. PMLR, 2023. 4", + "[31] Haomiao Liu, Ruiping Wang, Shiguang Shan, and Xilin Chen. What is a tabby? interpretable model decisions by learning attribute-based classification criteria. IEEE transactions on pattern analysis and machine intelligence, 43(5): 1791-1807, 2019. 2", + "[32] Francesco Locatello, Stefan Bauer, Mario Lucic, Gunnar Raetsch, Sylvain Gelly, Bernhard Schölkopf, and Olivier Bachem. Challenging common assumptions in the unsupervised learning of disentangled representations. In international conference on machine learning, pages 4114-4124. PMLR, 2019. 7", + "[33] Simon C Marshall and Jan H Kirchner. Understanding polysemantics in neural networks through coding theory. arXiv preprint arXiv:2401.17975, 2024. 2, 3", + "[34] Dang Minh, H Xiang Wang, Y Fen Li, and Tan N Nguyen. Explainable artificial intelligence: a comprehensive review. Artificial Intelligence Review, pages 1-66, 2022. 1", + "[35] Grégoire Montavon, Sebastian Lapuschkin, Alexander Binder, Wojciech Samek, and Klaus-Robert Müller. Explaining nonlinear classification decisions with deep taylor decomposition. Pattern recognition, 65:211-222, 2017. 2", + "[36] Grégoire Montavon, Alexander Binder, Sebastian Lapuschkin, Wojciech Samek, and Klaus-Robert Müller. Layer-wise relevance propagation: an overview. Explainable AI: interpreting, explaining and visualizing deep learning, pages 193-209, 2019. 1, 2" + ], + "bbox": [ + 91, + 90, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[37] Meike Nauta, Ron van Bree, and Christin Seifert. Neural prototype trees for interpretable fine-grained image recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14933-14943, 2021. 2", + "[38] Laura O'Mahony, Vincent Andrearczyk, Henning Müller, and Mara Graziani. Disentangling neuron representations with concept vectors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3770-3775, 2023. 3", + "[39] OpenAI. Hello gpt-4o. https://openai.com/index/hellogpt-4o/, 2024.4", + "[40] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 5", + "[41] Yao Rong, Tobias Leemann, Thai-Trang Nguyen, Lisa Fiedler, Peizhu Qian, Vaibhav Unhelkar, Tina Seidel, Gjergji Kasneci, and Enkelejda Kasneci. Towards human-centered explainable ai: A survey of user studies for model explanations. IEEE transactions on pattern analysis and machine intelligence, 46(4):2104-2122, 2024. 1", + "[42] Cynthia Rudin. Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead. Nature Machine Intelligence, 1(5):206-215, 2019. 2", + "[43] Waddah Saeed and Christian Omlin. Explainable ai (xai): A systematic meta-survey of current challenges and future opportunities. Knowledge-Based Systems, 263:110273, 2023. 1, 2", + "[44] Pranab Sahoo, Ayush Kumar Singh, Sriparna Saha, Vinija Jain, Samrat Mondal, and Aman Chadha. A systematic survey of prompt engineering in large language models: Techniques and applications. arXiv preprint arXiv:2402.07927, 2024.5", + "[45] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE international conference on computer vision (ICCV), pages 618-626, 2017. 1, 2", + "[46] Ao Sun, Pingchuan Ma, Yuanyuan Yuan, and Shuai Wang. Explain any concept: Segment anything meets concept-based explanation. Advances in Neural Information Processing Systems, 36, 2024. 1, 2", + "[47] Giulia Vilone and Luca Longo. Notions of explainability and evaluation approaches for explainable artificial intelligence. Information Fusion, 76:89-106, 2021. 7", + "[48] Kevin Ro Wang, Alexandre Variengien, Arthur Conmy, Buck Shlegeris, and Jacob Steinhardt. Interpretability in the wild: a circuit for indirect object identification in GPT-2 small. In The Eleventh International Conference on Learning Representations, ICLR 2023, Kigali, Rwanda, May 1-5, 2023. 2", + "[49] Teng Wang, Jinrui Zhang, Junjie Fei, Hao Zheng, Yunlong Tang, Zhe Li, Mingqi Gao, and Shanshan Zhao. Caption" + ], + "bbox": [ + 516, + 90, + 905, + 901 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "4373", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "anything: Interactive image description with diverse multimodal controls. arXiv preprint arXiv:2305.02677, 2023. 4", + "[50] Xin Wang, Hong Chen, Si'ao Tang, Zihao Wu, and Wenwu Zhu. Disentangled representation learning. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(12): 9677-9696, 2024. 1", + "[51] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824-24837, 2022. 2, 4", + "[52] Hongxu Yin, Pavlo Molchanov, Jose M Alvarez, Zhizhong Li, Arun Mallya, Derek Hoiem, Niraj K Jha, and Jan Kautz. Dreaming to distill: Data-free knowledge transfer via deep-inversion. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (CVPR), pages 8715-8724, 2020. 1", + "[53] Wenlong Yu, Ruonan Liu, Dongyue Chen, and Qinghua Hu. Explainability enhanced object detection transformer with feature disentanglement. IEEE Transactions on Image Processing, 33:6439-6454, 2024. 7", + "[54] Yu Zhang, Peter Tiño, Aleš Leonardis, and Ke Tang. A survey on neural network interpretability. IEEE Transactions on Emerging Topics in Computational Intelligence, 5(5):726-742, 2021. 2", + "[55] Shitian Zhao, Zhuowan Li, Yadong Lu, Alan Yuille, and Yan Wang. Causal-cog: A causal-effect look at context generation for boosting multi-modal language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13342-13351, 2024. 2", + "[56] Bolei Zhou, David Bau, Aude Oliva, and Antonio Torralba. Interpreting deep visual representations via network dissection. IEEE transactions on pattern analysis and machine intelligence, 41(9):2131-2145, 2018. 2, 3, 7" + ], + "bbox": [ + 91, + 90, + 482, + 571 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "4374", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/CoE_ Chain-of-Explanation via Automatic Visual Concept Circuit Description and Polysemanticity Quantification/ed717be1-08cc-4195-89dc-b7ee977099ae_model.json b/2025/CoE_ Chain-of-Explanation via Automatic Visual Concept Circuit Description and Polysemanticity Quantification/ed717be1-08cc-4195-89dc-b7ee977099ae_model.json new file mode 100644 index 0000000000000000000000000000000000000000..f0b385c3755979fbb32b4275695d14f7cf3a5a2c --- /dev/null +++ b/2025/CoE_ Chain-of-Explanation via Automatic Visual Concept Circuit Description and Polysemanticity Quantification/ed717be1-08cc-4195-89dc-b7ee977099ae_model.json @@ -0,0 +1,2257 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.108, + 0.13, + 0.892, + 0.177 + ], + "angle": 0, + "content": "CoE: Chain-of-Explanation via Automatic Visual Concept Circuit Description and Polysemanticity Quantification" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.204, + 0.885, + 0.258 + ], + "angle": 0, + "content": "Wenlong Yu Qilong Wang* Chuang Liu Dong Li Qinghua Hu Tianjin Key Lab of Machine Learning, College of Intelligence and Computing, Tianjin University {ywl_95, qlwang, liuc_09, li_dong, huqinghua}@tju.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.292, + 0.327, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.324, + 0.486, + 0.732 + ], + "angle": 0, + "content": "Explainability is a critical factor influencing the wide deployment of deep vision models (DVMs). Concept-based post-hoc explanation methods can provide both global and local insights into model decisions. However, current methods in this field face challenges in that they are inflexible to automatically construct accurate and sufficient linguistic explanations for global concepts and local circuits. Particularly, the intrinsic polysemanticity in semantic Visual Concepts (VCs) impedes the interpretability of concepts and DVMs, which is underestimated severely. In this paper, we propose a Chain-of-Explanation (CoE) approach to address these issues. Specifically, CoE automates the decoding and description of VCs to construct global concept explanation datasets. Further, to alleviate the effect of polysemanticity on model explainability, we design a concept polysemanticity disentanglement and filtering mechanism to distinguish the most contextually relevant concept atoms. Besides, a Concept Polysemanticity Entropy (CPE), as a measure of model interpretability, is formulated to quantify the degree of concept uncertainty. The modeling of deterministic concepts is upgraded to uncertain concept atom distributions. Finally, CoE automatically enables linguistic local explanations of the decision-making process of DVMs by tracing the concept circuit. GPT-4o and human-based experiments demonstrate the effectiveness of CPE and the superiority of CoE, achieving an average absolute improvement of \\(36\\%\\) in terms of explainability scores." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.761, + 0.222, + 0.777 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.786, + 0.485, + 0.879 + ], + "angle": 0, + "content": "Deep learning-based vision models have demonstrated outstanding performance in various visual tasks, owing to the powerful learning and representation capabilities of deep neural networks (DNNs) [29, 50]. However, the extreme complexity of DNNs, while enhancing their performance, also limits their explainability [41]. Humans are often un" + }, + { + "type": "image", + "bbox": [ + 0.526, + 0.291, + 0.896, + 0.469 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.48, + 0.908, + 0.509 + ], + "angle": 0, + "content": "Figure 1. Illustration of the integrated process of concept-based post-hoc XAI methods [36], along with two challenges they face." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.522, + 0.907, + 0.583 + ], + "angle": 0, + "content": "aware of the reasons behind the model's conclusions. This lack of explainability significantly hinders the wide application of deep vision models (DVMs) in critical fields such as autonomous driving and medical diagnosis [7, 28]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.583, + 0.909, + 0.854 + ], + "angle": 0, + "content": "Post-hoc eXplainable Artificial Intelligence (XAI) can enhance the transparency of DVMs without causing degradations in performance [34, 43]. Research in this field can roughly be divided into local and global XAI methods [6]. The former explains either specific stages or the entire decision path with respect to a specific sample, identifying key activated regions or significant semantic concepts [3, 45, 46]. The latter decipher the overall behavior of the DVM in a dataset by concept discovery or visualization techniques [22, 27, 52]. Among them, concept-based methods can explain DVMs from both global and local (glocal) perspectives [2, 4], assuming that each channel or neuron in a certain layer is responsible for recognizing essential semantic concepts. As shown in Fig. 1, they primarily perform one or two of the three operations: decoding Visual Concepts (VCs) of DVMs, manually constructing global concept explanation datasets, and providing local concept circuit explanations for individual samples [14, 16, 26]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Despite the significant progress made in concept-based explanation research, certain challenges remain, as illustrated in Fig. 1. First, the VCs are decoded by manually" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.237, + 0.9 + ], + "angle": 0, + "content": "*Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4364" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.094, + 0.092, + 0.482, + 0.453 + ], + "angle": 0, + "content": "identifying and annotating the commonalities of a set of extracted image patches [2, 26]. These manual methods intuitively become ineffective when faced with new scenarios, new models, and new XAI methods, due to high labor costs, low efficiency, and the limited cognitive scope of individual annotators. Second, most methods make a naive assumption: each concept is monosemantic and user-friendly [33]. However, in practice, VCs encapsulate polysemantic information [23] (e.g., an average of 13 concept atoms are coupled per channel in a ResNet152 model), presenting challenges for the model explanation. Particularly, polysemanticity can serve as a metric for evaluating a model's interpretability. A higher level of polysemanticity signifies a broader conceptual space, increasing the risks of selecting incorrect semantics and generating deviated explanations. Arbitrary concept selection or presenting them solely as images without descriptions yields suboptimal explanations [43]. Both visualizations and linguistic explanations become challenging for humans to comprehend, as it is unclear which semantic is actually represented or holds prominence [16]. Although polysemanticity is crucial to XAI, its systematic quantification and resolution remain unaddressed. All of these challenges collectively exacerbate the difficulty of concept-based explanations for DVMs." + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.464, + 0.482, + 0.9 + ], + "angle": 0, + "content": "To alleviate this dilemma, we propose a Chain-of-Explanation (CoE) approach via automatic concept decoding, disentanglement, filtering, description, and polysemanticity quantification. Specifically, CoE first proposes an Automatic Concept decoding and Description (ACD) method to construct global concept explanation datasets. It decodes every VC of key layers through arbitrary advanced XAI methods. A Large Vision Language Model (LVLM) is utilized to describe the commonalities of the extracted image patches since the linguistic explanation is more comprehensible than images [43]. As for the polysemantic concepts, we formulate a Concept Polysemanticity Disentanglement and Filtering (CPDF) mechanism to disentangle them into a set of orthogonal concept atoms, along with a Concept Polysemanticity Entropy (CPE) to quantify their degree of polysemanticity. In this paper, 13 distinct semantic directions are prompted, including low-level and high-level semantics, which is much more than previous work (e.g., 5 in [56]). Within the CPDF, a semantic entailment model is applied to mitigate the impact of redundant semantics. After these operations, the modeling of deterministic concepts is upgraded to uncertain concept atom distributions. When elucidating the local decision-making process, a concept circuit method is utilized to form the basis of an explanation chain, with each node representing a hierarchical set of essential concepts. Relevant atoms and their interconnections are contextually filtered and aligned within the chain. Finally, we consolidate all conceptual information along the explanation chain to generate linguistic explanations similar" + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.167 + ], + "angle": 0, + "content": "to a Chain-of-Thought (CoT) pipeline [21, 51], leveraging a Large Language Model (LLM), as shown in Fig. 2. Qualitative and quantitative experiments validate the effectiveness and the superiority of the CoE approach. The contributions of this paper are summarized as follows." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.169, + 0.904, + 0.228 + ], + "angle": 0, + "content": "- We propose a novel CoE approach to automatically and systematically provide concept-based glocal explanations, along with managing the problem of polysemantics, in a more interpretable natural language format." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.229, + 0.905, + 0.363 + ], + "angle": 0, + "content": "- A CPDF mechanism is formulated to handle the hard-to-interpret polysemantic concepts. To our knowledge, CPDF is a pioneer tailored to comprehensively disentangle and quantify polysemanticity. It disentangles them into a set of succinct concept atoms, followed by a filtering function to contextually find the suitable atom as the explanation node. A CPE score is defined to quantify the polysemanticity of concepts, which can be utilized as a metric of the interpretability of concepts or models." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.365, + 0.904, + 0.425 + ], + "angle": 0, + "content": "- Both GPT-4o and human-based experiments validate the superiority of the CoE in automatically constructing global and local explanations and the effectiveness of CPE in quantifying concept polysemanticity." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.169, + 0.905, + 0.425 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.517, + 0.438, + 0.654, + 0.453 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.464, + 0.905, + 0.87 + ], + "angle": 0, + "content": "Research on XAI can be broadly categorized into ad-hoc model interpretability and post-hoc explanation studies [43, 54]. The former actively refine DVMs to improve interpretability while also necessitating consideration of a trade-off between performance and interpretability [8, 31, 37, 42, 55]. The latter explains DVMs passively without performance decline [5, 11, 22, 45, 46]. Among these post-hoc efforts, concept-based explanations can provide glacial insights into DVMs [36]. Literature [56] annotated every pixel in special datasets with only 5 kinds of concepts. The explanation is derived by analyzing the activation regions of each neuron. MILAN [26] decoded VCs by calculating the maximum mutual information between channels and regions, followed by manually summarizing commonalities into 3 sentences. However, defining concepts solely through activations is insufficient. CRP decoded VCs through the relevance value derived from the deep Taylor decomposition [35], showing more accurate results than the activation-based methods [2]. The commonalities of the VCs were also identified manually, and one of these commonalities was selected randomly. However, these methods encounter significant challenges, as discussed in Sec. 1. In contrast, this paper proposes an automated concept decoding and description method based on LVLM, which offers greater flexibility in providing linguistic explanations. The most suitable concepts for interpretation are contextually selected, showing superior results." + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.871, + 0.904, + 0.901 + ], + "angle": 0, + "content": "Circuits, as sub-graphs, are proposed to explain the key decision route of a model [3, 13, 17, 20, 48]. They identify" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.515, + 0.957 + ], + "angle": 0, + "content": "4365" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.098, + 0.088, + 0.9, + 0.407 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.416, + 0.907, + 0.445 + ], + "angle": 0, + "content": "Figure 2. Outline of the Chain-of-Explanation. It consists of three steps (i.e., ACD, CPDF, and local explanation). Each part utilizes a set of channels \\(\\mathcal{C}\\) from \\(\\mathcal{M}\\). In the concept circuit, \\(\\mathcal{P}\\), lines of different depths and colors represent the magnitude of the relevance." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.458, + 0.486, + 0.716 + ], + "angle": 0, + "content": "the highest-activated concepts and then explain the model's decision by linking all concepts across the circuit. However, its interpretability for DVMs is limited due to the lack of a cohesive, language-based explanatory output. Additionally, the polysemanticity of VCs significantly hinders the accurate explanation [18, 33], and research in this area remains limited. Some methods attempted to indirectly mitigate its impact by predefining principal vectors or subspaces [20, 27, 38]. The lack of flexible and accurate concept-level explanations also restricts their adaptability. Literature [33] analyzed the polysemanticity from an information compression perspective. Pure [16] alleviated the polysemanticity by clustering and decoupling the circuit route. These methods, however, lack the ability to simultaneously and comprehensively quantify, disentangle, and describe polysemantic VCs. Their applicability is limited—challenges that this paper seeks to address, as discussed in Sec. 3." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.737, + 0.388, + 0.755 + ], + "angle": 0, + "content": "3. Chain-of-Explanation Approach" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.485, + 0.902 + ], + "angle": 0, + "content": "In this paper, a CoE approach is proposed to automatically explain DVMs from both global and local perspectives. It breaks down the complex reasoning path into a chain of concept nodes, within which polysemanticity is effectively managed. As shown in Fig. 2, CoE is composed of three steps. Step 1: ACD is proposed to automatically construct global concept descriptive explanation datasets of a given DVM. Step 2: With the assistance of an LVLM, a CPDF mechanism is designed to disentangle each original VC into" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.458, + 0.907, + 0.565 + ], + "angle": 0, + "content": "a set of concept atoms and extract the most suitable atom according to the context, making CoE more comprehensible and flexible. Besides, a CPE score is formulated accordingly to quantify the degree of polysemanticity of concepts. Step 3: A local explanation chain in a manner akin to a CoT for the decision-making progress of the DVM is established based on the dataset, CPDF, concept circuit, and LLMs." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.572, + 0.902, + 0.589 + ], + "angle": 0, + "content": "3.1. Automatic Concept Decoding and Description" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.595, + 0.907, + 0.776 + ], + "angle": 0, + "content": "ACD method is first formulated to automatically and effectively construct databases of global concept descriptions. Specifically, given an image dataset \\(\\mathcal{T}\\) and a DVM \\(\\mathcal{M}\\) requiring evaluation and explanation, we establish an injective function \\(f\\) between channel set \\(\\mathcal{C}\\) and visual semantic concept set \\(\\mathcal{S}\\) by an XAI method \\(E_{x}, f: \\mathcal{C} \\mapsto \\mathcal{S}\\). Without loss of generality, for \\(j_{th}\\) channel \\(c_{j}^{l}\\) in layer \\(l\\), \\(E_{x}\\) extracts a set of image patches within \\(\\mathcal{T}\\) that most activate this channel, named VC \\(s_{j}^{l} = \\{I_{n}\\}_{N}\\), where \\(N\\) is the number of image patches. Each \\(I_{n}\\) highlights certain regions of the original image, representing the focused concept, as exemplified in the bottom left part of Fig. 2." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.776, + 0.907, + 0.809 + ], + "angle": 0, + "content": "A describing function \\( r \\) is then defined to decipher \\( \\{I_n\\}_N \\) into a set of language descriptions:" + }, + { + "type": "equation", + "bbox": [ + 0.616, + 0.815, + 0.907, + 0.833 + ], + "angle": 0, + "content": "\\[\n\\mathcal {D} = r \\left(\\left\\{I _ {n} \\right\\} _ {N}\\right) = \\left\\{d _ {m} \\right\\} _ {M}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.841, + 0.909, + 0.903 + ], + "angle": 0, + "content": "where \\(M\\) is the number of language descriptions for the current channel. Traditionally, \\(r\\) is implemented by manually and subjectively identifying commonalities within \\(\\{I_n\\}_N\\) [56], while ACD utilizes a powerful LVLM as the descriptor" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4366" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.155 + ], + "angle": 0, + "content": "with well-engineered prompts. After constructing \\(S\\), \\(\\mathcal{D}\\) and the mapping function \\(f\\), a concept description dataset \\(\\mathcal{B}_{\\mathcal{M},\\mathcal{T}} = \\{\\mathcal{C}, S, \\mathcal{D}, f\\}\\) is automatically constructed, providing global linguistic explanations for all VCs of \\(\\mathcal{M}\\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.16, + 0.383, + 0.178 + ], + "angle": 0, + "content": "3.2. Managing Polysemantic Concept" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.184, + 0.485, + 0.366 + ], + "angle": 0, + "content": "VCs of each node within a DVM exhibit polysemantics, introducing certain deviations in the explanation framework, as shown in Fig. 2. Higher polysemantics indicate that the current VC aggregates more semantics, making the intended representation more ambiguous. Such semantic uncertainty impedes human comprehension of the concept, particularly within the CoE, where various concepts must be understood collectively. Quantifying and solving this dilemma is difficult and has been underestimated by previous studies. Under this circumstance, we propose a CPDF mechanism along with a CPE score to mitigate and quantify its detriments to the CoE simultaneously." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.383, + 0.48, + 0.4 + ], + "angle": 0, + "content": "3.2.1 Polysemantic Concept Decoupling and Filtering" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.407, + 0.485, + 0.468 + ], + "angle": 0, + "content": "Considering that a polysemous word can be understood as an aggregation of various distinct meanings, we define a function \\( g \\) to disentangle the ambiguous VC into a set of easily understandable linguistic concept atoms:" + }, + { + "type": "equation", + "bbox": [ + 0.166, + 0.474, + 0.483, + 0.493 + ], + "angle": 0, + "content": "\\[\n\\mathcal {A} = \\left\\{a ^ {1}, \\dots , a ^ {i}, \\dots , a ^ {Q} \\right\\} = g (s), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.501, + 0.484, + 0.639 + ], + "angle": 0, + "content": "where \\( Q \\) is the number of the concept atoms. We posit that each atom represents a view of the current VC \\( s \\), which will be activated upon the appearance of the corresponding semantic in the input. Furthermore, given that concept atoms should accurately represent the commonalities in \\( \\{I_n\\}_N \\) and recognizing that this naive disentanglement function (Eq. 2) does not provide a foundation for quantifying polysemanticity, we refine the function \\( g \\) to disentangle each image patch \\( I_n \\) into a set of atoms:" + }, + { + "type": "equation", + "bbox": [ + 0.147, + 0.644, + 0.483, + 0.682 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {A} _ {n} = \\left\\{a _ {n} ^ {1}, \\dots , a _ {n} ^ {i}, \\dots , a _ {n} ^ {Q} \\right\\} = g (I _ {n}, s), \\tag {3} \\\\ \\mathcal {D} = \\{\\mathcal {A} _ {n} \\} _ {N}. \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.69, + 0.484, + 0.901 + ], + "angle": 0, + "content": "We instantiate \\( g \\) by a powerful LVLM (e.g., GPT-4o [39]), considering that this is quite an open and complex problem. \\( g \\) prioritizes disentangling and describing commonalities within the VC while analyzing each individual image. Various commonalities among any certain number of image patches should be summarized and described by precise terminologies. These common semantics within an open scenario are as diverse as they are abundant, unfolding in countless forms and nuances. Besides, it should generate specific atoms when any commonalities identified above do not appear in the current \\( I_{n} \\). Therefore, the art of prompt engineering prompt-com becomes paramount. It should provide rules for recognizing commonalities, possible conceptual directions, and output format control. In" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.198 + ], + "angle": 0, + "content": "this paper, we prompt 13 high-level and low-level semantical directions that cover the normal concepts encountered in daily life, including object category, scene, object part, color, texture, material, position, transparency, brightness, shape, size, edges, and their relationships. Additionally, the CoT techniques [21, 51] are introduced. Details of all sophisticated prompts are shown in Sec. S2 of the Appendix." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.199, + 0.905, + 0.227 + ], + "angle": 0, + "content": "All \\(\\mathcal{A}_n\\) are then aggregated into a description atom set \\(\\mathcal{A}\\) with all duplicate atoms removed:" + }, + { + "type": "equation", + "bbox": [ + 0.605, + 0.235, + 0.905, + 0.273 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {A} = \\left\\{\\mathcal {A} _ {1} \\cup \\mathcal {A} _ {2} \\cup \\dots \\cup \\mathcal {A} _ {N} \\right\\} \\tag {4} \\\\ = \\left\\{a ^ {1}, \\dots , a ^ {i}, \\dots , a ^ {P} \\right\\}, \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.278, + 0.905, + 0.385 + ], + "angle": 0, + "content": "where \\(P\\) is the number of concept atoms after deleting the repeating atoms. Each element in \\(\\mathcal{A}\\) represents a disentangled monosemantic concept and can be utilized as a node of the CoE. However, some atoms are semantically equivalent even if they have variant descriptions [19]. In order to remove semantically redundant atoms, we introduce an entailment function \\(\\delta\\) to detect the relations between atoms:" + }, + { + "type": "equation", + "bbox": [ + 0.575, + 0.39, + 0.905, + 0.408 + ], + "angle": 0, + "content": "\\[\n\\{1, 0, - 1 \\} = \\delta \\left(a ^ {i 1}, a ^ {i 2}\\right), \\forall a ^ {i 1}, a ^ {i 2} \\in \\mathcal {A}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.415, + 0.906, + 0.581 + ], + "angle": 0, + "content": "where the responses 1,0,-1 represent entailment, neutral, or contradiction of the candidates, respectively. We instantiate \\(\\delta\\) by an NLI model (e.g., DeBERTa model [25]) because the atom candidates are relatively short and simple that an NLI model is good enough. Following [19], a bidirectional entailment check is applied, meaning that one of the candidates will be removed when the response is entailment or neutral. The first atom is designated as the monosemantic representative of the semantic equivalence set due to the transitivity between them. The final concept set \\(A^{*}\\) only contains heterogeneous atoms that describe the current VC:" + }, + { + "type": "equation", + "bbox": [ + 0.604, + 0.587, + 0.905, + 0.613 + ], + "angle": 0, + "content": "\\[\n\\mathcal {A} ^ {*} = \\left\\{a ^ {1}, \\dots , a ^ {i}, \\dots , a ^ {P ^ {*}} \\right\\}. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.618, + 0.905, + 0.664 + ], + "angle": 0, + "content": "After disentangling the concept atoms, we define a filtering function \\( z \\) to select an atom \\( a^* \\) as the most suitable explanation of current VC with respect to the input \\( x \\):" + }, + { + "type": "equation", + "bbox": [ + 0.633, + 0.67, + 0.905, + 0.688 + ], + "angle": 0, + "content": "\\[\na _ {s} ^ {*} (x) = z \\left(\\mathcal {A} _ {s} ^ {*}, C (x)\\right). \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.694, + 0.906, + 0.815 + ], + "angle": 0, + "content": "Some research ignores this step or selects the concept randomly with an implicit assumption that all concepts in \\(\\mathcal{A}^*\\) are semantically equivalent [14, 26]. In contrast, we filter the most suitable atom \\(a^*\\) based on the context of the CoE. In particular, \\(z\\) is instantiated by an LLM. \\(C(x)\\) is an image captioning function utilized for providing contextual information. It can be instantiated by a relatively small LVLM [9, 10] or a professional model [30, 49]." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.832, + 0.864, + 0.849 + ], + "angle": 0, + "content": "3.2.2 Quantification of Concept Polysemanticity" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.906, + 0.901 + ], + "angle": 0, + "content": "The polysemanticity of concepts can serve as a metric for the model's interpretability. Lower polysemanticity of concepts indicates reduced semantic uncertainty with a more" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4367" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.182 + ], + "angle": 0, + "content": "singular form of activation, leading to greater interpretability. In this paper, we endeavor to quantify polysemanticity from three levels of granularity (i.e., conceptual channels, layers, and models) by defining a CPE score. The larger the CPE, the more pronounced the polysemanticity, which leads to decreased interpretability." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.182, + 0.485, + 0.243 + ], + "angle": 0, + "content": "In particular, the disentanglement function \\( g \\) in Eq. 3 enables us to count the frequency of each atom in \\( \\mathcal{A}^* \\). Given two fixed numbers \\( Q \\) and \\( N \\), we calculate the probability of the \\( i_{th} \\) atom within \\( j_{th} \\) concept:" + }, + { + "type": "equation", + "bbox": [ + 0.213, + 0.248, + 0.483, + 0.281 + ], + "angle": 0, + "content": "\\[\np _ {i} = \\frac {\\text {N u m} _ {i}}{Q \\times N + P a d}. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.111, + 0.285, + 0.449, + 0.3 + ], + "angle": 0, + "content": "Then, the CPE of \\(j_{th}\\) concept can be formulated as" + }, + { + "type": "equation", + "bbox": [ + 0.188, + 0.306, + 0.483, + 0.343 + ], + "angle": 0, + "content": "\\[\nH _ {j} ^ {*} = \\frac {- \\sum_ {i = 1} ^ {P ^ {*} + P a d} p _ {i} \\log p _ {i}}{\\log (P ^ {*} + P a d)}. \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.348, + 0.484, + 0.543 + ], + "angle": 0, + "content": "The denominator normalizes the entropy value to a range between 0 and 1, considering that the size of \\( P^* \\) varies for different VCs. Besides, the quantity of \\( A^* \\) is set to be at least equal to \\( N \\) due to some \\( D_s \\) having a small number of semantical equivalent atoms with equal probabilities. If there are insufficient unique atoms, we supplement the corresponding number of atoms, with each supplementary atom set to a number of 1. The atomic objects are expanded, and the denominator is increased by the number of supplemented atoms (a total amount of \\( Pad \\)). Please refer to Sec. S3.1 of the Appendix for details. In addition, two averaged CPEs are formulated to enable comparisons of polysemantics between specific layers or models:" + }, + { + "type": "equation", + "bbox": [ + 0.226, + 0.55, + 0.483, + 0.587 + ], + "angle": 0, + "content": "\\[\nH _ {l} ^ {*} = \\frac {1}{d _ {l}} \\sum^ {d _ {l}} H _ {j} ^ {*}, \\tag {10}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.224, + 0.593, + 0.483, + 0.627 + ], + "angle": 0, + "content": "\\[\nH _ {\\mathcal {M}} ^ {*} = \\frac {1}{L} \\sum^ {L} H _ {l} ^ {*}, \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.629, + 0.484, + 0.72 + ], + "angle": 0, + "content": "where \\(d_{l}\\) is the dimension of the channel and \\(L\\) is the number of layers. Finally, to accommodate the diverse needs of researchers within the community, the automatically constructed concept explanation dataset is extended as ACD\\(\\mathcal{B}_{\\mathcal{M},\\mathcal{T}} = \\{\\mathcal{C},\\mathcal{S},\\mathcal{D},\\mathcal{A}^{*},\\mathcal{Q},\\mathcal{H},f\\}\\), where \\(\\mathcal{Q}\\) represents the probability set of each atom and \\(\\mathcal{H}\\) is the set of CPE." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.727, + 0.399, + 0.743 + ], + "angle": 0, + "content": "3.3. Construction of Explanation Chain" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Concept circuits in DVMs can identify the most activated units in each layer for a specific input [3]. When aggregating these units across all key layers in a bottom-up manner, a corresponding maximum activation path \\(\\mathcal{P}\\) is generated, establishing the foundation for the explanation chain. Each node in this path serves as the explanation for the model's decision at that layer. Traditional concept circuit methods for DVMs identify the activation path without providing thorough overarching natural language-based explanations. Inspired by the CoT explanation technique in LLMs" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.152 + ], + "angle": 0, + "content": "[12], CoE proposes a local explanation chain to interpret the decision-making process of DVMs in the form of natural language descriptions based on ACD-\\(\\mathcal{B}_{\\mathcal{M},\\mathcal{T}}\\), CPDF, concept circuit method, and LLM, as shown in Fig. 2." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.153, + 0.907, + 0.382 + ], + "angle": 0, + "content": "Specifically, some user-predefined layers requiring explanations or key layers facilitating CoE (e.g., output layers of 4 stages of a DVM) are listed in a bottom-up manner. Given an image \\(x\\) and layer \\(l\\), an advanced XAI method \\(E_{x}\\) computes the maximization normalized activation or relevance values for all concepts. Each value represents the importance of the current concept to the prediction \\(\\mathcal{M}(x)\\). CoE acknowledges top \\(k^l\\) relevant concepts \\(\\mathcal{C}^l(x) = \\{c_1^l, \\dots, c_{k^l}^l\\}\\) and their values \\(\\mathcal{V}^l(x) = \\{v_1^l, \\dots, v_{k^l}^l\\}\\), determined by the \\(\\alpha\\) quantile of relevance values across all concepts (i.e., \\(k^l = |\\{v_j^l | v_j^l > \\alpha \\max(v^l), \\forall j \\in d^l\\}|\\), where \\(|\\mathcal{X}|\\) computes the number of elements). After that, the descriptive concept atoms of the current layer \\(\\widetilde{\\mathcal{A}}_{k^l}^{*,l}(x) = \\{a_1^{*,l}, \\dots, a_{k^l}^{*,l}\\}\\) are constructed based on ACD-\\(\\mathcal{B}_{\\mathcal{M},\\mathcal{T}}\\) and the filter Eq. 7, serving as a node of the explanation chain." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.384, + 0.906, + 0.429 + ], + "angle": 0, + "content": "The final CoE for the local decision-making explanation, abstracting and synthesizing all top activated concept atoms along the chain, is formulated as" + }, + { + "type": "equation", + "bbox": [ + 0.518, + 0.441, + 0.906, + 0.46 + ], + "angle": 0, + "content": "\\[\nC o E _ {\\mathcal {M}} (x) = e \\left(\\widetilde {\\mathcal {A}} _ {\\mathcal {P}} ^ {*} (x), \\mathcal {V} _ {\\mathcal {P}} (x), C (x), G (x), \\mathcal {M} (x)\\right), \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.473, + 0.906, + 0.548 + ], + "angle": 0, + "content": "where \\( e \\) formulates a synthesizing and describing function with inputs obtained from the above sections, and \\( G(x) \\) is the label of \\( x \\). We instantiate \\( CoE_{\\mathcal{M}}(x) \\) through a powerful LLM (e.g., GPT-4 [1]), with prompt prompt-coe well-designed by combining CoT and few-shot prompting [44]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.55, + 0.906, + 0.595 + ], + "angle": 0, + "content": "Additionally, these local explanations are evaluated by a powerful LVLM and human judges based on three explainability metrics, each utilizing a three-tier scoring system." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.612, + 0.646, + 0.629 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.638, + 0.906, + 0.759 + ], + "angle": 0, + "content": "We evaluate the proposed CoE approach on two types of DVMs (i.e., ResNet [24] and CLIP [40]). The relevance-based XAI method CRP is adopted as the primary approach for \\( E_{x} \\) in this paper, owing to its demonstrated superiority in terms of fidelity and reliability [2]. Please refer to the Appendix for more details on the implementation and results. Codes are available at https://github.com/YuWLong666/CoE." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.771, + 0.88, + 0.788 + ], + "angle": 0, + "content": "4.1. Concept Explanation Dataset Construction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.906, + 0.901 + ], + "angle": 0, + "content": "In the CoE, we first automatically construct global concept description databases \\(\\mathsf{ACD - B}_{\\mathcal{M},\\mathcal{T}}\\) for various DVMs. Taking a ResNet152 model as an example, 4 output layers of 4 stages (with 256, 512, 1024, and 2048 channels \\(\\mathcal{C}\\), respectively) are selected as the key layer set. CRP is applied as \\(E_{x}\\) to extract the VC set \\(\\mathcal{S}\\). Following [2], each VC is represented by 15 (i.e., \\(N\\)) image patches. The mapping function" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4368" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.908, + 0.119 + ], + "angle": 0, + "content": "Table 1. Examples of VCs and their disentangled concept atoms, along with their concept probability distributions and CPE scores. The first row shows concept 75 of Layer 4 extracted by CRP. The second row shows concept 697 of Layer 4 extracted by MILAN." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.128, + 0.902, + 0.48 + ], + "angle": 0, + "content": "
Visual Concept and Disentangled Concept AtomsConcept DistributionCPE
Image 1: [burger, lettuce, tomato], \nImage 2: [blue, monkey, face], \nImage 3: [shark, blue, water], \nImage 4: [keyboard, buttons, numbers], \nImage 5: [shark, blue, water], \nImage 6: [shark, blue, water], \nImage 7: [shark, blue, water], \nImage 8: [blue, circular, shape], \nImage 9: [shark, blue, water], \nImage 10: [keyboard, buttons, numbers], \nImage 11: [blue, yellow, objects], \nImage 12: [shark, blue, text], \nImage 13: [shark, blue, water], \nImage 14: [shark, blue, water], \nImage 15: [shark, blue, water]0.30\n0.25\n0.20\n0.15\n0.10\n0.05\n0.00\nhurty green, blue, texture, yellow, objects, water, black0.81
Image 1: [hand, blinds, string], \nImage 2: [curtains, room, grayscale], \nImage 3: [fabric, green, closeup], \nImage 4: [curtains, room, light], \nImage 5: [reflection, glass, light], \nImage 6: [hand, fabric, texture], \nImage 7: [water, fountain, greenery], \nImage 8: [feathers, pattern, dark], \nImage 9: [fabric, blue, closeup], \nImage 10: [fabric, green, seatbelt], \nImage 11: [leaf, bark, texture], \nImage 12: [nature, mountain, grayscale], \nImage 13: [skull, ground, dark], \nImage 14: [fabric, colorful, pattern], \nImage 15: [curtains, window, light]0.12\n0.10\n0.08\n0.06\n0.04\n0.02\n0.00\n0.00\n0.00\n0.00\n0.00\n0.00\n0.00\n0.00\n0.00\n0.00\n0.00\n0.00\n0.00\n0.00\n0.00\n0.00\n0.00\n0.00\n0.00\n0.00\n0.02\n0.02\n0.02\n0.02\n0.02\n0.02\n0.02\n0.02\n0.02\n0.02\n0.02\n0.02\n0.02\n0.02\n0.02\n0.02\n0.02\n0.02\n0.02\n0.02\n0.00\n0.00\n0.00\n0.00\n0.00\n0.00\n0.00\n0.00\n0.00\n0.00\n0.00\n0.00\n0.00\n0.00\n0.00\n0.00\n0.00\n0.00\n0.02\n0.00\n0.00\n0.00\n0.00\n0.00\n0.00\n0.00\n0.00\n0.00\n0.00\n0.00\n0.00\n0.00\n0.00\n0.00\n0.00\n0.00\n0.00\n0.01\n0.01\n0.01\n0.01\n0.01\n0.01\n0.01\n0.01\n0.01\n0.01\n0.01\n0.01\n0.01\n0.01\n0.01\n0.01\n0.01\n0.01\n0.01\n0.01\n0.02\n0.02\n0.02\n0.02\n0.02\n0.02\n0.02\n0.02\n0.02\n0.02\n0.02\n0.02\n0.02\n0.02\n0.02\n0.02\n0.02\n0.02\n0.02\n0.03\n0.03\n0.03\n0.03\n0.03\n0.03\n0.03\n0.03\n0.03\n0.03\n0.03\n0.03\n0.03\n0.03\n0.03\n0.03\n0.03\n0.03\n0.03\n0.03\n0.04\n0.04\n0.04\n0.04\n0.04\n0.04\n0.04\n0.04\n0.04\n0.04\n0.04\n0.04\n0.04\n0.04\n0.04\n0.04\n0.04\n0.04\n0.04\n0.04\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.08\n0.08\n0.08\n0.08\n0.08\n0.08\n0.08\n0.08\n0.08\n0.08\n0.08\n0.08\n0.08\n0.08\n0.08\n0.08\n0.08\n0.08\n0.08\n0.08\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.08\n0.08\n0.08\n0.08\n0.08\n0.08\n0.08\n0.08\n0.08\n0.08\n0.08\n0.08\n0.08\n0.08\n0.08\n0.08\n0.08\n0.08\n0.08\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.04\n0.04\n0.04\n0.04\n0.04\n0.04\n0.04\n0.04\n0.04\n0.04\n0.04\n0.04\n0.04\n0.04\n0.04\n0.04\n0.04\n0.04\n0.04\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.08\n0.08\n0.08\n0.08\n0.08\n0.08\n0.08\n0.08\n0.08\n0.08\n0.08\n0.08\n0.08\n0.08\n0.08\n0.08\n0.08\n0.08\n0.08\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.04\n0.04\n0.04\n0.04\n0.04\n0.04\n0.04\n0.04\n0.04\n0.04\n0.04\n0.04\n0.04\n0.04\n0.04\n0.04\n0.04\n0.04\n0.04\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.02\n0.12\n0.12\n0.12\n0.12\n0.12\n0.12\n0.12\n0.12\n0.12\n0.12\n0.12\n0.12\n0.12\n0.12\n0.12\n0.12\n0.12\n0.12\n0.12\n0.12\n0.25\n0.25\n0.25\n0.25\n0.25\n0.25\n0.25\n0.25\n0.25\n0.25\n0.25\n0.25\n0.25\n0.25\n0.25\n0.25\n0.25\n0.25\n0.25\n0.25\n0.26\n0.26\n0.26\n0.26\n0.26\n0.26\n0.26\n0.26\n0.26\n0.26\n0.26\n0.26\n0.26\n0.26\n0.26\n0.26\n0.26\n0.26\n0.26\n0.26\n0.25\n0.25\n0.25\n0.25\n0.25\n0.25\n0.25\n0.25\n0.25\n0.25\n0.25\n0.25\n0.25\n0.25\n0.25\n0.25\n0.25\n0.25\n0.25\n0.2.\n12\n12\n12\n12\n12\n12\n12\n12\n12\n12\n12\n12\n12\n12\n12\n12\n12\n12\n12\n12\n12\n12\n12\n12\n12\n12\n12\n12\n12\n12\n12\n12\n12\n12\n25\n25\n25\n25\n25\n25\n25\n25\n25\n25\n25\n25\n25\n25\n25\n25\n25\n25\n25\n25\n25\n25\n25\n25\n25\n25\n25\n25\n25\n25\n25\n25\n25\n25
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.492, + 0.484, + 0.645 + ], + "angle": 0, + "content": "\\(f\\) from channels to VCs is built. We then employ a GPT4o-2024-08-06 model as the function \\(g\\) to generate the description set \\(\\mathcal{D}\\) (i.e., Eq. 3). The disentangled and clustered atom set \\(\\mathcal{A}^*\\) (processed by a DeBERTa model), the probability set \\(\\mathcal{Q}\\), and the CPE set \\(\\mathcal{H}\\) are computed and gathered accordingly. Overall, ACD-\\(\\mathcal{B}_{\\mathcal{M},\\mathcal{T}}\\) containing 3,840 entries is constructed, with each entry comprising a channel, a CPE score, and 15 VC representations—each consisting of 3 concept atoms (i.e., \\(Q\\))—and multiple orthogonal linguistic atoms, along with their associated probabilities." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.651, + 0.452, + 0.669 + ], + "angle": 0, + "content": "4.2. Disentanglement of Polysemantic Concept" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.675, + 0.484, + 0.901 + ], + "angle": 0, + "content": "CPDF automatically disentangles and describes the polysemantic VCs into a set of linguistic concept atoms, \\(\\mathcal{D}\\). As exemplified by Table 1, two randomly selected VCs are extracted by two advanced XAI methods (CRP and MILAN). They are disentangled and described into atoms, exhibiting different degrees of polysemasticity. We observe that the commonalities within the VCs are almost entirely identified, and most of the atoms describe the commonalities of subsets of image patches. In the first row (CPE=0.81), \\(g\\) successfully identifies all visually discernible commonalities, such as shark and water. The atoms exhibit high consistency, and fewer but more frequently occurring common semantics are identified. Only a few atoms, such as the burger, are not commonalities, as their corresponding images lack sufficient common semantics. The atoms are also repre" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.492, + 0.907, + 0.6 + ], + "angle": 0, + "content": "sented accurately. For instance, the atom sets for each image featuring the shark semantic all include the word shark. In the second row (CPE=0.95), our method remains effective in disentangling and describing VC with a higher level of polysemantics. \\( g \\) identifies a greater variety of atoms. These results demonstrate the effectiveness of the disentanglement and description operation in the CPDF." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.606, + 0.787, + 0.623 + ], + "angle": 0, + "content": "4.3. Concept Distribution and CPE" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.629, + 0.907, + 0.81 + ], + "angle": 0, + "content": "CPE can quantify the polysemanticity and serve as an indicator of a DVM's interpretability. It begins by calculating the probabilities of each orthogonal atom. As the middle column of Table 1 shows, the distribution provides statistics on the extent to which different atoms are associated with each concept. For instance, concept 75 primarily focuses on blue with a probability of 0.27 while also identifying yellow with a smaller probability. The semantics between atoms are non-overlapping. This probabilistic approach enhances explainability by capturing multiple semantics with their occurring probabilities, upgrading the modeling of deterministic concepts to uncertain atom distributions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.909, + 0.903 + ], + "angle": 0, + "content": "CPE quantifies polysemanticity from 3 levels of granularity. As shown in Fig. 3a, concepts exhibit varying degrees of polysemanticity, with considerable fluctuations (ranging from a maximum of 1 to a minimum of 0.5). Most of the concepts demonstrate a noticeable level of polysemanticity. When analyzing the polysemanticity of different" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4369" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.09, + 0.362, + 0.244 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.109, + 0.247, + 0.347, + 0.26 + ], + "angle": 0, + "content": "(a) Sorted CPE along the channel dimension." + }, + { + "type": "image", + "bbox": [ + 0.365, + 0.093, + 0.634, + 0.244 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.407, + 0.247, + 0.59, + 0.26 + ], + "angle": 0, + "content": "(b) Averaged CPE on 4 key layers." + }, + { + "type": "image", + "bbox": [ + 0.643, + 0.091, + 0.9, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.686, + 0.247, + 0.853, + 0.26 + ], + "angle": 0, + "content": "(c) Averaged CPE on 3 models." + }, + { + "type": "image_caption", + "bbox": [ + 0.12, + 0.272, + 0.875, + 0.286 + ], + "angle": 0, + "content": "Figure 3. Illustrations of CPE scores at 3 levels of granularity. In (a), the channel indices of each layer are max normalized to 1." + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.297, + 0.483, + 0.325 + ], + "angle": 0, + "content": "Table 2. The consistency of the recognition of concept polysemantics between the CPE metric and human scorers." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.328, + 0.474, + 0.376 + ], + "angle": 0, + "content": "
Consistency
CPE Metric and Human Scorers75 %
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.386, + 0.483, + 0.596 + ], + "angle": 0, + "content": "layers, as shown in Fig. 3b, the averaged CPE peaks in the third stage, while the shallowest and deepest layers exhibit lower values. Typically, deeper layers capture abstract concepts, whereas the shallower layers extract elementary features like textures [53, 56]. DVMs show better monosemanticity when identifying elementary and abstract concepts, while more ambiguous concepts are captured in the intermediate layers. This experiment validates the theoretical and experimental claims of previous work [14, 32]. In Fig. 3c, we observe that as the number of layers increases, the CPE of the model shows a rising trend. This result indicates a direct proportional relationship between the polysemanticity and the complexity of DVMs, which is that the more complex a DVM becomes, the harder it is to interpret." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.598, + 0.483, + 0.794 + ], + "angle": 0, + "content": "The CPE is evaluated in comparison with human assessments. As shown in Table 1, the CPE is consistent with the degree of polysemanticity from the human intuitive observations (the polysemanticity of the first row is smaller than that of the second row). Similarly, we randomly sample 300 pairs of VCs, each containing two VCs with varying levels of polysemanticity. Comparisons are drawn from these 300 pairs by both human evaluators and the CPE metric. As shown in Table 2, results from human assessments and CPE achieve a \\(75\\%\\) agreement rate, demonstrating the effectiveness of CPE. Other results are shown in Sec. S4 of the Appendix. All of these results reveal that CPE is effective in quantifying the polysemanticity of concepts and DVMs." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.802, + 0.342, + 0.818 + ], + "angle": 0, + "content": "4.4. CoE for Local Explanations" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.484, + 0.902 + ], + "angle": 0, + "content": "CoE generates local explanations akin to CoT by automatically describing the concept circuit. We set \\(\\alpha = 0.001\\) to acknowledge the top relevant VCs, resulting in an average of 3 key VCs per layer. As exemplified in Table 3, given an input image, CoE first outputs whether the prediction of the" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.299, + 0.907, + 0.511 + ], + "angle": 0, + "content": "DVM is correct and then gives reasons behind this prediction according to the information along the concept circuits. CoE employs an LLM as an engine to infer the logical relationships among concepts, as well as between concepts and the context. It synthesizes all evidence to generate an inductive commentary on the predictions. The explanation chain is presented in a linguistic format, enhancing human understanding. For instance, CoE explains why the DVM incorrectly predicted a tench as a mosquito net: the top activation of mosquito-net-related concepts, rather than those related to fish, misled the decision. Additional instances are shown in Sec. S5 of the Appendix. Overall, CoE effectively captures and elucidates the core decision-making evidence for both correct and incorrect predictions of DVMs." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.526, + 0.906, + 0.542 + ], + "angle": 0, + "content": "4.5. Quantitative Evaluation via GPT-4o & Human" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.552, + 0.906, + 0.718 + ], + "angle": 0, + "content": "We employ GPT-4o and human evaluators to assess the local explanations for 500 and 100 samples randomly selected from the ImageNet Validation dataset [15], respectively. The GPT-4o evaluation is achieved through a carefully designed prompt prompt-coe-eval. Explanations generated by three methods are rated according to three explainability metrics: Accuracy, Completeness, and User Interpretability [47]. Each is assigned a three-tier scoring system (2, 1, and 0), with specific criteria defined for each tier, resulting in a total explanation score of 6. Please refer to Sec. S2 and S5 of the Appendix for details." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.907, + 0.901 + ], + "angle": 0, + "content": "As shown in Table 4, our CoE approach attains peak performance across all three metrics, achieving total scores of 5.06 and 5.04 in GPT-4o and human evaluations, respectively. The lowest scores (3.14 and 2.62) are obtained on the MILAN ANNOTATION dataset without the ACD and CPDF methods. This dataset was manually annotated, resulting in inaccuracy and insufficiency within concept explanations. Scores increase by 0.11 and 0.72, respectively, with the application of our proposed ACD-\\(\\mathcal{B}_{\\mathcal{M},\\mathcal{T}}\\) dataset. The ACD method not only enables the automated disentanglement and description of concept atoms but also enhances their precision and comprehensibility. When the CoE em" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4370" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.098, + 0.09, + 0.9, + 0.104 + ], + "angle": 0, + "content": "Table 3. Cases of local explanations of CoE. The first row shows the explanation of a correct prediction, and the second row is reversed." + }, + { + "type": "image", + "bbox": [ + 0.098, + 0.12, + 0.287, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.1, + 0.233, + 0.147, + 0.28 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.148, + 0.233, + 0.176, + 0.242 + ], + "angle": 0, + "content": "label:" + }, + { + "type": "image_caption", + "bbox": [ + 0.148, + 0.245, + 0.286, + 0.255 + ], + "angle": 0, + "content": "black-and-tan coonhound" + }, + { + "type": "image_caption", + "bbox": [ + 0.148, + 0.256, + 0.187, + 0.266 + ], + "angle": 0, + "content": "predict:" + }, + { + "type": "image_caption", + "bbox": [ + 0.157, + 0.268, + 0.286, + 0.277 + ], + "angle": 0, + "content": "black-and-tan coonhound" + }, + { + "type": "text", + "bbox": [ + 0.29, + 0.124, + 0.902, + 0.255 + ], + "angle": 0, + "content": "The model outputs a correct result: black-and-tan coonhound. Specifically, in the shallowest layer, layer 1, channel 88 with a relevance value of 1.0 describes the concept curved, which may relate to the curved shapes observed in parts of the dog such as its tail or ears. Moving deeper, in layer 2, channel 506 with a relevance value of 0.85 and channel 391 with a relevance value of 0.75 describe the concept of animal, indicating the model has identified an animal in the image, and channel 229 with a relevance value of 0.77 emphasizes the color black, aligning with the black and tan colors of the coonhound. Further, layer 2 also identifies the presence of grass several times (channels 94 and 103), which corresponds to the grassy field described in the image caption. In layer 3, various channels, such as 985, 174, 422, and 13, with high relevance scores, from 0.72 to 1.0, repeatedly identify the color green and the concept grass, firmly establishing the setting of the image. Additionally, channels 11, 787, and 591 in layer 3 enhance the certainty by repeatedly identifying animal and creature, confirming that the subject is an animal. Finally, layer 4 decisively contributes with channels 1354 and 1661, both signifying dog with high relevance values (1.0 and 0.72 respectively), which correctly identifies the animal in the image as a dog." + }, + { + "type": "text", + "bbox": [ + 0.291, + 0.256, + 0.902, + 0.277 + ], + "angle": 0, + "content": "Given that the model's structured concepts strongly match the description and characteristics noted in the caption and the ground truth label, it is clear why the model outputs the correct result of black-and-tan coonhound." + }, + { + "type": "image", + "bbox": [ + 0.098, + 0.29, + 0.288, + 0.401 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.401, + 0.15, + 0.448 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.152, + 0.412, + 0.185, + 0.422 + ], + "angle": 0, + "content": "label:" + }, + { + "type": "image_caption", + "bbox": [ + 0.153, + 0.425, + 0.197, + 0.435 + ], + "angle": 0, + "content": "predict:" + }, + { + "type": "image_caption", + "bbox": [ + 0.221, + 0.412, + 0.255, + 0.422 + ], + "angle": 0, + "content": "tench" + }, + { + "type": "image_caption", + "bbox": [ + 0.208, + 0.425, + 0.286, + 0.435 + ], + "angle": 0, + "content": "mosquito net" + }, + { + "type": "text", + "bbox": [ + 0.291, + 0.294, + 0.902, + 0.447 + ], + "angle": 0, + "content": "The model outputs an incorrect result mosquito net instead of the correct label tench. Specifically, in the shallowest layer 1 of the model, channel 88 with a relevance value of 1.0 describes the concept of smooth, channel 61 with a relevance value of 0.81 describes the concept of green, and channel 111 with a relevance value of 0.72 describes the concept of brown. In the deeper layer 2 of the model, channel 94 with a relevance value of 1.0 describes the concept of water, channel 229 with a relevance value of 0.84 describes the concept of fish, channel 103 with a relevance value of 0.81 also describes the concept of fish, and channel 467 with a relevance value of 0.73 describes the concept of mesh. At layer 3, channel 985 with a relevance value of 1.0 further describes the concept of fish, while channels 894, 122, and 715 with relevance values of 0.83, 0.73, and 0.72, respectively, describe the concept of green, possibly relating to the color of the net. In the deepest layer 4, channel 697 with a relevance value of 1.0 describes the concept of net, channel 1878 with a relevance value of 0.76 describes the concept of baby, and channel 321 with a relevance value of 0.76 represents the concept of hexagon, which could misconstrue the net structure. The prevalence of concepts related to fish within the decision path suggests a relationship to the tench, but the final prediction of mosquito net was likely influenced by the dominance of net-related concepts. Thus, the model was misled by the presence of a net, resulting in the incorrect classification of mosquito net." + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.465, + 0.907, + 0.508 + ], + "angle": 0, + "content": "Table 4. Scores of CoE for the local explanation on three methods evaluated by GPT-4o and human scorers. The first method utilizes MILAN ANNOTATION as the concept explanation dataset, while the rest two work on the ACD-\\(\\mathcal{B}_{\\mathcal{M},\\mathcal{T}}\\) datasets. Each criterion has a maximum score of 2 points. The total explanation score is 6 points. Higher scores indicate better explainability." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.518, + 0.9, + 0.647 + ], + "angle": 0, + "content": "
MethodAccuracyCompletenessUser Interpret.Total Explanation
GPT-4oMILAN A.[26] + Description (Baseline)1.031.071.043.14
CoE w/o Filtering1.071.101.083.25
CoE (ours)1.691.701.675.06
HumanMILAN A.[26] + Description (Baseline)0.870.830.922.62
CoE w/o Filtering1.101.161.073.34
CoE (ours)1.731.721.595.04
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.665, + 0.483, + 0.831 + ], + "angle": 0, + "content": "ploys the full CPDF mechanism, the explanation score rises by 1.92 and 2.42 points (the average absolute improvement is \\(36\\%\\)), respectively, with the filtering step contributing 1.81 and 1.70 points. This result suggests that by taking polysemanticity and contextual filtering of concepts into account, CoE effectively guides the LLM in producing better linguistic explanations, substantially enhancing their accuracy, completeness, and interpretability for users. The results of the human evaluation are consistent with the GPT-4o results. All of these results strongly affirm the superiority of the proposed CoE approach." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.845, + 0.21, + 0.861 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.901 + ], + "angle": 0, + "content": "In this paper, we propose a CoE approach to tackle the challenges of the inflexibility to automatically construct con" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.665, + 0.907, + 0.892 + ], + "angle": 0, + "content": "cept explanation datasets, insufficient linguistic explanations, and the weakness in managing concept polysemantics of concept-based XAI methods. CoE automates the decoding and description of concepts via an LVLM and constructs global concept explanation datasets. A CPDF mechanism is designed to identify the most contextually relevant concept atoms, advancing the modeling of deterministic concepts to uncertain concept atom distributions. The CPE is proposed to quantify the polysemanticsy. Finally, CoE enables local explanations of a DVM's decisionmaking process, represented through natural language. The effectiveness and superiority of CPE and CoE are demonstrated. In summary, CoE establishes a systematical framework for explaining DVMs, and we believe that the findings presented in this paper warrant further exploration." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "4371" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.091, + 0.241, + 0.108 + ], + "angle": 0, + "content": "Acknowledgment" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.116, + 0.483, + 0.161 + ], + "angle": 0, + "content": "This research is supported in part by the National Natural Science Foundation of China under Grants U23B2049, 62276186, and 61925602." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.178, + 0.187, + 0.194 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.203, + 0.484, + 0.271 + ], + "angle": 0, + "content": "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.274, + 0.484, + 0.356 + ], + "angle": 0, + "content": "[2] Reduan Achtibat, Maximilian Dreyer, Ilona Eisenbraun, Sebastian Bosse, Thomas Wiegand, Wojciech Samek, and Sebastian Lapuschkin. From attribution maps to human-understandable explanations through concept relevance propagation. Nature Machine Intelligence, 5(9):1006-1019, 2023. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.359, + 0.483, + 0.413 + ], + "angle": 0, + "content": "[3] Rajaram Achyuta, Chowdhury Neil, Torralba Antonio, Andreas Jacob, and Schwettmann Sarah. Automatic discovery of visual circuits. In NeurIPS Workshop on Attributing Model Behavior at Scale, 2023. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.416, + 0.483, + 0.485 + ], + "angle": 0, + "content": "[4] Sebastian Bach, Alexander Binder, Grégoire Montavon, Frederick Klauschen, Klaus-Robert Müller, and Wojciech Samek. On pixel-wise explanations for non-linear classifier decisions by layer-wise relevance propagation. PLoS ONE, 10, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.488, + 0.484, + 0.556 + ], + "angle": 0, + "content": "[5] David Bau, Jun-Yan Zhu, Hendrik Strobelt, Agata Lapedriza, Bolei Zhou, and Antonio Torralba. Understanding the role of individual units in a deep neural network. Proceedings of the National Academy of Sciences, 117(48):30071-30078, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.559, + 0.483, + 0.6 + ], + "angle": 0, + "content": "[6] Leonard Bereska and Efstratios Gavves. Mechanistic interpretability for ai safety-a review. arXiv preprint arXiv:2404.14082, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.602, + 0.483, + 0.657 + ], + "angle": 0, + "content": "[7] Subrato Bharati, M Rubaiyat Hossain Mondal, and Prajoy Podder. A review on explainable artificial intelligence for healthcare: why, how, and when? IEEE Transactions on Artificial Intelligence, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.66, + 0.483, + 0.715 + ], + "angle": 0, + "content": "[8] Moritz Böhle, Navdeeppal Singh, Mario Fritz, and Bernt Schiele. B-cos alignment for inherently interpretable cnns and vision transformers. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(6):4504-4518, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.717, + 0.483, + 0.786 + ], + "angle": 0, + "content": "[9] Zhe Chen, Weiyun Wang, Hao Tian, Shenglong Ye, Zhang-wei Gao, Erfei Cui, Wenwen Tong, Kongzhi Hu, Jiapeng Luo, Zheng Ma, et al. How far are we to gpt-4v? closing the gap to commercial multimodal models with open-source suites. arXiv preprint arXiv:2404.16821, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.788, + 0.483, + 0.871 + ], + "angle": 0, + "content": "[10] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24185–24198, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[11] Ming-Ming Cheng, Peng-Tao Jiang, Ling-Hao Han, Liang Wang, and Philip Torr. Deeply explain cnn via hierarchical" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.203, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "decomposition. International Journal of Computer Vision, 131(5):1091-1105, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.219 + ], + "angle": 0, + "content": "[12] Zheng Chu, Jingchang Chen, Qianglong Chen, Weijiang Yu, Tao He, Haotian Wang, Weihua Peng, Ming Liu, Bing Qin, and Ting Liu. Navigate through enigmatic labyrinth a survey of chain of thought reasoning: Advances, frontiers and future. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1173–1203, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.221, + 0.905, + 0.289 + ], + "angle": 0, + "content": "[13] Arthur Conmy, Augustine Mavor-Parker, Aengus Lynch, Stefan Heimersheim, and Adrià Garriga-Alonso. Towards automated circuit discovery for mechanistic interpretability. Advances in Neural Information Processing Systems, 36: 16318-16352, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.291, + 0.905, + 0.347 + ], + "angle": 0, + "content": "[14] Meghal Dani, Isabel Rio-Torto, Stephan Alaniz, and Zeynep Akata. Devil: Decoding vision features into language. In DAGM German Conference on Pattern Recognition, pages 363-377. Springer, 2023. 1, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.349, + 0.905, + 0.404 + ], + "angle": 0, + "content": "[15] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.406, + 0.905, + 0.488 + ], + "angle": 0, + "content": "[16] Maximilian Dreyer, Erblina Purelku, Johanna Vielhaben, Wojciech Samek, and Sebastian Lapuschkin. Pure: Turning polysemantic neurons into pure features by identifying relevant circuits. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 8212-8217, 2024. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.49, + 0.905, + 0.559 + ], + "angle": 0, + "content": "[17] Nelson Elhage, Neel Nanda, Catherine Olsson, Tom Henighan, Nicholas Joseph, Ben Mann, Amanda Askell, Yuntao Bai, Anna Chen, Tom Conerly, et al. A mathematical framework for transformer circuits. Transformer Circuits Thread, 1(1):12, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.561, + 0.905, + 0.628 + ], + "angle": 0, + "content": "[18] Nelson Elhage, Tristan Hume, Catherine Olsson, Nicholas Schiefer, Tom Henighan, Shauna Kravec, Zac Hatfield-Dodds, Robert Lasenby, Dawn Drain, Carol Chen, et al. Toy models of superposition. arXiv preprint arXiv:2209.10652, 2022.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.632, + 0.905, + 0.673 + ], + "angle": 0, + "content": "[19] Sebastian Farquhar, Jannik Kossen, Lorenz Kuhn, and Yarin Gal. Detecting hallucinations in large language models using semantic entropy. Nature, 630(8017):625-630, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.675, + 0.905, + 0.757 + ], + "angle": 0, + "content": "[20] Thomas Fel, Victor Boutin, Louis Béthune, Rémi Cadène, Mazda Moayeri, Léo Andéol, Mathieu Chalvidal, and Thomas Serre. A holistic approach to unifying automatic concept extraction and concept importance estimation. Advances in Neural Information Processing Systems, 36, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.759, + 0.905, + 0.815 + ], + "angle": 0, + "content": "[21] Guhao Feng, Bohang Zhang, Yuntian Gu, Haotian Ye, Di He, and Liwei Wang. Towards revealing the mystery behind chain of thought: a theoretical perspective. Advances in Neural Information Processing Systems, 36, 2024. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.817, + 0.905, + 0.87 + ], + "angle": 0, + "content": "[22] Amirata Ghorbani, James Wexler, James Y Zou, and Been Kim. Towards automatic concept-based explanations. Advances in neural information processing systems, 32, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.903 + ], + "angle": 0, + "content": "[23] Dan W Grupe and Jack B Nitschke. Uncertainty and anticipation in anxiety: an integrated neurobiological and psy" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4372" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.12 + ], + "angle": 0, + "content": "chological perspective. Nature Reviews Neuroscience, 14(7): 488-501, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.178 + ], + "angle": 0, + "content": "[24] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 770-778, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.179, + 0.482, + 0.234 + ], + "angle": 0, + "content": "[25] Pengcheng He, Xiaodong Liu, Jianfeng Gao, and Weizhu Chen. Deberta: Decoding-enhanced bert with disentangled attention. In International Conference on Learning Representations, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.236, + 0.483, + 0.292 + ], + "angle": 0, + "content": "[26] Evan Hernandez, Sarah Schwettmann, David Bau, Teona Bagashvili, Antonio Torralba, and Jacob Andreas. Natural language descriptions of deep features. In International Conference on Learning Representations, 2022. 1, 2, 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.293, + 0.483, + 0.362 + ], + "angle": 0, + "content": "[27] Been Kim, Martin Wattenberg, Justin Gilmer, Carrie Cai, James Wexler, Fernanda Viegas, et al. Interpretability beyond feature attribution: Quantitative testing with concept activation vectors (tcav). In International conference on machine learning, pages 2668-2677. PMLR, 2018. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.363, + 0.483, + 0.432 + ], + "angle": 0, + "content": "[28] Anton Kuznietsov, Balint Gyevnar, Cheng Wang, Steven Peters, and Stefano V. Albrecht. Explainable ai for safe and trustworthy autonomous driving: A systematic review. IEEE Transactions on Intelligent Transportation Systems, pages 1-23, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.434, + 0.483, + 0.462 + ], + "angle": 0, + "content": "[29] Yann LeCun, Yoshua Bengio, and Geoffrey Hinton. Deep learning. nature, 521(7553):436-444, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.463, + 0.483, + 0.532 + ], + "angle": 0, + "content": "[30] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730-19742. PMLR, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.534, + 0.483, + 0.603 + ], + "angle": 0, + "content": "[31] Haomiao Liu, Ruiping Wang, Shiguang Shan, and Xilin Chen. What is a tabby? interpretable model decisions by learning attribute-based classification criteria. IEEE transactions on pattern analysis and machine intelligence, 43(5): 1791-1807, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.604, + 0.483, + 0.687 + ], + "angle": 0, + "content": "[32] Francesco Locatello, Stefan Bauer, Mario Lucic, Gunnar Raetsch, Sylvain Gelly, Bernhard Schölkopf, and Olivier Bachem. Challenging common assumptions in the unsupervised learning of disentangled representations. In international conference on machine learning, pages 4114-4124. PMLR, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.689, + 0.483, + 0.731 + ], + "angle": 0, + "content": "[33] Simon C Marshall and Jan H Kirchner. Understanding polysemantics in neural networks through coding theory. arXiv preprint arXiv:2401.17975, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.732, + 0.483, + 0.773 + ], + "angle": 0, + "content": "[34] Dang Minh, H Xiang Wang, Y Fen Li, and Tan N Nguyen. Explainable artificial intelligence: a comprehensive review. Artificial Intelligence Review, pages 1-66, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.775, + 0.483, + 0.831 + ], + "angle": 0, + "content": "[35] Grégoire Montavon, Sebastian Lapuschkin, Alexander Binder, Wojciech Samek, and Klaus-Robert Müller. Explaining nonlinear classification decisions with deep taylor decomposition. Pattern recognition, 65:211-222, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[36] Grégoire Montavon, Alexander Binder, Sebastian Lapuschkin, Wojciech Samek, and Klaus-Robert Müller. Layer-wise relevance propagation: an overview. Explainable AI: interpreting, explaining and visualizing deep learning, pages 193-209, 2019. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.906, + 0.162 + ], + "angle": 0, + "content": "[37] Meike Nauta, Ron van Bree, and Christin Seifert. Neural prototype trees for interpretable fine-grained image recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14933-14943, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.163, + 0.906, + 0.232 + ], + "angle": 0, + "content": "[38] Laura O'Mahony, Vincent Andrearczyk, Henning Müller, and Mara Graziani. Disentangling neuron representations with concept vectors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3770-3775, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.234, + 0.906, + 0.262 + ], + "angle": 0, + "content": "[39] OpenAI. Hello gpt-4o. https://openai.com/index/hellogpt-4o/, 2024.4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.264, + 0.906, + 0.346 + ], + "angle": 0, + "content": "[40] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.348, + 0.906, + 0.431 + ], + "angle": 0, + "content": "[41] Yao Rong, Tobias Leemann, Thai-Trang Nguyen, Lisa Fiedler, Peizhu Qian, Vaibhav Unhelkar, Tina Seidel, Gjergji Kasneci, and Enkelejda Kasneci. Towards human-centered explainable ai: A survey of user studies for model explanations. IEEE transactions on pattern analysis and machine intelligence, 46(4):2104-2122, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.433, + 0.906, + 0.488 + ], + "angle": 0, + "content": "[42] Cynthia Rudin. Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead. Nature Machine Intelligence, 1(5):206-215, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.49, + 0.906, + 0.545 + ], + "angle": 0, + "content": "[43] Waddah Saeed and Christian Omlin. Explainable ai (xai): A systematic meta-survey of current challenges and future opportunities. Knowledge-Based Systems, 263:110273, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.547, + 0.906, + 0.615 + ], + "angle": 0, + "content": "[44] Pranab Sahoo, Ayush Kumar Singh, Sriparna Saha, Vinija Jain, Samrat Mondal, and Aman Chadha. A systematic survey of prompt engineering in large language models: Techniques and applications. arXiv preprint arXiv:2402.07927, 2024.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.617, + 0.906, + 0.7 + ], + "angle": 0, + "content": "[45] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE international conference on computer vision (ICCV), pages 618-626, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.702, + 0.906, + 0.758 + ], + "angle": 0, + "content": "[46] Ao Sun, Pingchuan Ma, Yuanyuan Yuan, and Shuai Wang. Explain any concept: Segment anything meets concept-based explanation. Advances in Neural Information Processing Systems, 36, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.759, + 0.906, + 0.801 + ], + "angle": 0, + "content": "[47] Giulia Vilone and Luca Longo. Notions of explainability and evaluation approaches for explainable artificial intelligence. Information Fusion, 76:89-106, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.802, + 0.906, + 0.872 + ], + "angle": 0, + "content": "[48] Kevin Ro Wang, Alexandre Variengien, Arthur Conmy, Buck Shlegeris, and Jacob Steinhardt. Interpretability in the wild: a circuit for indirect object identification in GPT-2 small. In The Eleventh International Conference on Learning Representations, ICLR 2023, Kigali, Rwanda, May 1-5, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.906, + 0.902 + ], + "angle": 0, + "content": "[49] Teng Wang, Jinrui Zhang, Junjie Fei, Hao Zheng, Yunlong Tang, Zhe Li, Mingqi Gao, and Shanshan Zhao. Caption" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.906, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4373" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.12 + ], + "angle": 0, + "content": "anything: Interactive image description with diverse multimodal controls. arXiv preprint arXiv:2305.02677, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.177 + ], + "angle": 0, + "content": "[50] Xin Wang, Hong Chen, Si'ao Tang, Zihao Wu, and Wenwu Zhu. Disentangled representation learning. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(12): 9677-9696, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.179, + 0.482, + 0.247 + ], + "angle": 0, + "content": "[51] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824-24837, 2022. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.249, + 0.482, + 0.331 + ], + "angle": 0, + "content": "[52] Hongxu Yin, Pavlo Molchanov, Jose M Alvarez, Zhizhong Li, Arun Mallya, Derek Hoiem, Niraj K Jha, and Jan Kautz. Dreaming to distill: Data-free knowledge transfer via deep-inversion. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (CVPR), pages 8715-8724, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.333, + 0.482, + 0.388 + ], + "angle": 0, + "content": "[53] Wenlong Yu, Ruonan Liu, Dongyue Chen, and Qinghua Hu. Explainability enhanced object detection transformer with feature disentanglement. IEEE Transactions on Image Processing, 33:6439-6454, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.39, + 0.482, + 0.444 + ], + "angle": 0, + "content": "[54] Yu Zhang, Peter Tiño, Aleš Leonardis, and Ke Tang. A survey on neural network interpretability. IEEE Transactions on Emerging Topics in Computational Intelligence, 5(5):726-742, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.446, + 0.482, + 0.515 + ], + "angle": 0, + "content": "[55] Shitian Zhao, Zhuowan Li, Yadong Lu, Alan Yuille, and Yan Wang. Causal-cog: A causal-effect look at context generation for boosting multi-modal language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13342-13351, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.517, + 0.482, + 0.572 + ], + "angle": 0, + "content": "[56] Bolei Zhou, David Bau, Aude Oliva, and Antonio Torralba. Interpreting deep visual representations via network dissection. IEEE transactions on pattern analysis and machine intelligence, 41(9):2131-2145, 2018. 2, 3, 7" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.572 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4374" + } + ] +] \ No newline at end of file diff --git a/2025/CoE_ Chain-of-Explanation via Automatic Visual Concept Circuit Description and Polysemanticity Quantification/ed717be1-08cc-4195-89dc-b7ee977099ae_origin.pdf b/2025/CoE_ Chain-of-Explanation via Automatic Visual Concept Circuit Description and Polysemanticity Quantification/ed717be1-08cc-4195-89dc-b7ee977099ae_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..549a6a4da0a789ea7561767c21240c721e441ecd --- /dev/null +++ b/2025/CoE_ Chain-of-Explanation via Automatic Visual Concept Circuit Description and Polysemanticity Quantification/ed717be1-08cc-4195-89dc-b7ee977099ae_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6540ec68abe743dd3a2b862df24a7eead5b63af0d2beab7038e0ad641a9c7a47 +size 3323306 diff --git a/2025/CoE_ Chain-of-Explanation via Automatic Visual Concept Circuit Description and Polysemanticity Quantification/full.md b/2025/CoE_ Chain-of-Explanation via Automatic Visual Concept Circuit Description and Polysemanticity Quantification/full.md new file mode 100644 index 0000000000000000000000000000000000000000..6152a9700fb4334a557b92c8695669461090acda --- /dev/null +++ b/2025/CoE_ Chain-of-Explanation via Automatic Visual Concept Circuit Description and Polysemanticity Quantification/full.md @@ -0,0 +1,1106 @@ +# CoE: Chain-of-Explanation via Automatic Visual Concept Circuit Description and Polysemanticity Quantification + +Wenlong Yu Qilong Wang* Chuang Liu Dong Li Qinghua Hu Tianjin Key Lab of Machine Learning, College of Intelligence and Computing, Tianjin University {ywl_95, qlwang, liuc_09, li_dong, huqinghua}@tju.edu.cn + +# Abstract + +Explainability is a critical factor influencing the wide deployment of deep vision models (DVMs). Concept-based post-hoc explanation methods can provide both global and local insights into model decisions. However, current methods in this field face challenges in that they are inflexible to automatically construct accurate and sufficient linguistic explanations for global concepts and local circuits. Particularly, the intrinsic polysemanticity in semantic Visual Concepts (VCs) impedes the interpretability of concepts and DVMs, which is underestimated severely. In this paper, we propose a Chain-of-Explanation (CoE) approach to address these issues. Specifically, CoE automates the decoding and description of VCs to construct global concept explanation datasets. Further, to alleviate the effect of polysemanticity on model explainability, we design a concept polysemanticity disentanglement and filtering mechanism to distinguish the most contextually relevant concept atoms. Besides, a Concept Polysemanticity Entropy (CPE), as a measure of model interpretability, is formulated to quantify the degree of concept uncertainty. The modeling of deterministic concepts is upgraded to uncertain concept atom distributions. Finally, CoE automatically enables linguistic local explanations of the decision-making process of DVMs by tracing the concept circuit. GPT-4o and human-based experiments demonstrate the effectiveness of CPE and the superiority of CoE, achieving an average absolute improvement of $36\%$ in terms of explainability scores. + +# 1. Introduction + +Deep learning-based vision models have demonstrated outstanding performance in various visual tasks, owing to the powerful learning and representation capabilities of deep neural networks (DNNs) [29, 50]. However, the extreme complexity of DNNs, while enhancing their performance, also limits their explainability [41]. Humans are often un + +![](images/aa96329ea76bccd11daf890f0a4b579f749490e38481e2465b4b91d59f82d235.jpg) +Figure 1. Illustration of the integrated process of concept-based post-hoc XAI methods [36], along with two challenges they face. + +aware of the reasons behind the model's conclusions. This lack of explainability significantly hinders the wide application of deep vision models (DVMs) in critical fields such as autonomous driving and medical diagnosis [7, 28]. + +Post-hoc eXplainable Artificial Intelligence (XAI) can enhance the transparency of DVMs without causing degradations in performance [34, 43]. Research in this field can roughly be divided into local and global XAI methods [6]. The former explains either specific stages or the entire decision path with respect to a specific sample, identifying key activated regions or significant semantic concepts [3, 45, 46]. The latter decipher the overall behavior of the DVM in a dataset by concept discovery or visualization techniques [22, 27, 52]. Among them, concept-based methods can explain DVMs from both global and local (glocal) perspectives [2, 4], assuming that each channel or neuron in a certain layer is responsible for recognizing essential semantic concepts. As shown in Fig. 1, they primarily perform one or two of the three operations: decoding Visual Concepts (VCs) of DVMs, manually constructing global concept explanation datasets, and providing local concept circuit explanations for individual samples [14, 16, 26]. + +Despite the significant progress made in concept-based explanation research, certain challenges remain, as illustrated in Fig. 1. First, the VCs are decoded by manually + +identifying and annotating the commonalities of a set of extracted image patches [2, 26]. These manual methods intuitively become ineffective when faced with new scenarios, new models, and new XAI methods, due to high labor costs, low efficiency, and the limited cognitive scope of individual annotators. Second, most methods make a naive assumption: each concept is monosemantic and user-friendly [33]. However, in practice, VCs encapsulate polysemantic information [23] (e.g., an average of 13 concept atoms are coupled per channel in a ResNet152 model), presenting challenges for the model explanation. Particularly, polysemanticity can serve as a metric for evaluating a model's interpretability. A higher level of polysemanticity signifies a broader conceptual space, increasing the risks of selecting incorrect semantics and generating deviated explanations. Arbitrary concept selection or presenting them solely as images without descriptions yields suboptimal explanations [43]. Both visualizations and linguistic explanations become challenging for humans to comprehend, as it is unclear which semantic is actually represented or holds prominence [16]. Although polysemanticity is crucial to XAI, its systematic quantification and resolution remain unaddressed. All of these challenges collectively exacerbate the difficulty of concept-based explanations for DVMs. + +To alleviate this dilemma, we propose a Chain-of-Explanation (CoE) approach via automatic concept decoding, disentanglement, filtering, description, and polysemanticity quantification. Specifically, CoE first proposes an Automatic Concept decoding and Description (ACD) method to construct global concept explanation datasets. It decodes every VC of key layers through arbitrary advanced XAI methods. A Large Vision Language Model (LVLM) is utilized to describe the commonalities of the extracted image patches since the linguistic explanation is more comprehensible than images [43]. As for the polysemantic concepts, we formulate a Concept Polysemanticity Disentanglement and Filtering (CPDF) mechanism to disentangle them into a set of orthogonal concept atoms, along with a Concept Polysemanticity Entropy (CPE) to quantify their degree of polysemanticity. In this paper, 13 distinct semantic directions are prompted, including low-level and high-level semantics, which is much more than previous work (e.g., 5 in [56]). Within the CPDF, a semantic entailment model is applied to mitigate the impact of redundant semantics. After these operations, the modeling of deterministic concepts is upgraded to uncertain concept atom distributions. When elucidating the local decision-making process, a concept circuit method is utilized to form the basis of an explanation chain, with each node representing a hierarchical set of essential concepts. Relevant atoms and their interconnections are contextually filtered and aligned within the chain. Finally, we consolidate all conceptual information along the explanation chain to generate linguistic explanations similar + +to a Chain-of-Thought (CoT) pipeline [21, 51], leveraging a Large Language Model (LLM), as shown in Fig. 2. Qualitative and quantitative experiments validate the effectiveness and the superiority of the CoE approach. The contributions of this paper are summarized as follows. + +- We propose a novel CoE approach to automatically and systematically provide concept-based glocal explanations, along with managing the problem of polysemantics, in a more interpretable natural language format. +- A CPDF mechanism is formulated to handle the hard-to-interpret polysemantic concepts. To our knowledge, CPDF is a pioneer tailored to comprehensively disentangle and quantify polysemanticity. It disentangles them into a set of succinct concept atoms, followed by a filtering function to contextually find the suitable atom as the explanation node. A CPE score is defined to quantify the polysemanticity of concepts, which can be utilized as a metric of the interpretability of concepts or models. +- Both GPT-4o and human-based experiments validate the superiority of the CoE in automatically constructing global and local explanations and the effectiveness of CPE in quantifying concept polysemanticity. + +# 2. Related Work + +Research on XAI can be broadly categorized into ad-hoc model interpretability and post-hoc explanation studies [43, 54]. The former actively refine DVMs to improve interpretability while also necessitating consideration of a trade-off between performance and interpretability [8, 31, 37, 42, 55]. The latter explains DVMs passively without performance decline [5, 11, 22, 45, 46]. Among these post-hoc efforts, concept-based explanations can provide glacial insights into DVMs [36]. Literature [56] annotated every pixel in special datasets with only 5 kinds of concepts. The explanation is derived by analyzing the activation regions of each neuron. MILAN [26] decoded VCs by calculating the maximum mutual information between channels and regions, followed by manually summarizing commonalities into 3 sentences. However, defining concepts solely through activations is insufficient. CRP decoded VCs through the relevance value derived from the deep Taylor decomposition [35], showing more accurate results than the activation-based methods [2]. The commonalities of the VCs were also identified manually, and one of these commonalities was selected randomly. However, these methods encounter significant challenges, as discussed in Sec. 1. In contrast, this paper proposes an automated concept decoding and description method based on LVLM, which offers greater flexibility in providing linguistic explanations. The most suitable concepts for interpretation are contextually selected, showing superior results. + +Circuits, as sub-graphs, are proposed to explain the key decision route of a model [3, 13, 17, 20, 48]. They identify + +![](images/82d54784ccc7eb026d1a531358c99154558cbe048c177f8a8e964c9a6e8e3789.jpg) +Figure 2. Outline of the Chain-of-Explanation. It consists of three steps (i.e., ACD, CPDF, and local explanation). Each part utilizes a set of channels $\mathcal{C}$ from $\mathcal{M}$ . In the concept circuit, $\mathcal{P}$ , lines of different depths and colors represent the magnitude of the relevance. + +the highest-activated concepts and then explain the model's decision by linking all concepts across the circuit. However, its interpretability for DVMs is limited due to the lack of a cohesive, language-based explanatory output. Additionally, the polysemanticity of VCs significantly hinders the accurate explanation [18, 33], and research in this area remains limited. Some methods attempted to indirectly mitigate its impact by predefining principal vectors or subspaces [20, 27, 38]. The lack of flexible and accurate concept-level explanations also restricts their adaptability. Literature [33] analyzed the polysemanticity from an information compression perspective. Pure [16] alleviated the polysemanticity by clustering and decoupling the circuit route. These methods, however, lack the ability to simultaneously and comprehensively quantify, disentangle, and describe polysemantic VCs. Their applicability is limited—challenges that this paper seeks to address, as discussed in Sec. 3. + +# 3. Chain-of-Explanation Approach + +In this paper, a CoE approach is proposed to automatically explain DVMs from both global and local perspectives. It breaks down the complex reasoning path into a chain of concept nodes, within which polysemanticity is effectively managed. As shown in Fig. 2, CoE is composed of three steps. Step 1: ACD is proposed to automatically construct global concept descriptive explanation datasets of a given DVM. Step 2: With the assistance of an LVLM, a CPDF mechanism is designed to disentangle each original VC into + +a set of concept atoms and extract the most suitable atom according to the context, making CoE more comprehensible and flexible. Besides, a CPE score is formulated accordingly to quantify the degree of polysemanticity of concepts. Step 3: A local explanation chain in a manner akin to a CoT for the decision-making progress of the DVM is established based on the dataset, CPDF, concept circuit, and LLMs. + +# 3.1. Automatic Concept Decoding and Description + +ACD method is first formulated to automatically and effectively construct databases of global concept descriptions. Specifically, given an image dataset $\mathcal{T}$ and a DVM $\mathcal{M}$ requiring evaluation and explanation, we establish an injective function $f$ between channel set $\mathcal{C}$ and visual semantic concept set $\mathcal{S}$ by an XAI method $E_{x}, f: \mathcal{C} \mapsto \mathcal{S}$ . Without loss of generality, for $j_{th}$ channel $c_{j}^{l}$ in layer $l$ , $E_{x}$ extracts a set of image patches within $\mathcal{T}$ that most activate this channel, named VC $s_{j}^{l} = \{I_{n}\}_{N}$ , where $N$ is the number of image patches. Each $I_{n}$ highlights certain regions of the original image, representing the focused concept, as exemplified in the bottom left part of Fig. 2. + +A describing function $r$ is then defined to decipher $\{I_n\}_N$ into a set of language descriptions: + +$$ +\mathcal {D} = r \left(\left\{I _ {n} \right\} _ {N}\right) = \left\{d _ {m} \right\} _ {M}, \tag {1} +$$ + +where $M$ is the number of language descriptions for the current channel. Traditionally, $r$ is implemented by manually and subjectively identifying commonalities within $\{I_n\}_N$ [56], while ACD utilizes a powerful LVLM as the descriptor + +with well-engineered prompts. After constructing $S$ , $\mathcal{D}$ and the mapping function $f$ , a concept description dataset $\mathcal{B}_{\mathcal{M},\mathcal{T}} = \{\mathcal{C}, S, \mathcal{D}, f\}$ is automatically constructed, providing global linguistic explanations for all VCs of $\mathcal{M}$ . + +# 3.2. Managing Polysemantic Concept + +VCs of each node within a DVM exhibit polysemantics, introducing certain deviations in the explanation framework, as shown in Fig. 2. Higher polysemantics indicate that the current VC aggregates more semantics, making the intended representation more ambiguous. Such semantic uncertainty impedes human comprehension of the concept, particularly within the CoE, where various concepts must be understood collectively. Quantifying and solving this dilemma is difficult and has been underestimated by previous studies. Under this circumstance, we propose a CPDF mechanism along with a CPE score to mitigate and quantify its detriments to the CoE simultaneously. + +# 3.2.1 Polysemantic Concept Decoupling and Filtering + +Considering that a polysemous word can be understood as an aggregation of various distinct meanings, we define a function $g$ to disentangle the ambiguous VC into a set of easily understandable linguistic concept atoms: + +$$ +\mathcal {A} = \left\{a ^ {1}, \dots , a ^ {i}, \dots , a ^ {Q} \right\} = g (s), \tag {2} +$$ + +where $Q$ is the number of the concept atoms. We posit that each atom represents a view of the current VC $s$ , which will be activated upon the appearance of the corresponding semantic in the input. Furthermore, given that concept atoms should accurately represent the commonalities in $\{I_n\}_N$ and recognizing that this naive disentanglement function (Eq. 2) does not provide a foundation for quantifying polysemanticity, we refine the function $g$ to disentangle each image patch $I_n$ into a set of atoms: + +$$ +\begin{array}{l} \mathcal {A} _ {n} = \left\{a _ {n} ^ {1}, \dots , a _ {n} ^ {i}, \dots , a _ {n} ^ {Q} \right\} = g (I _ {n}, s), \tag {3} \\ \mathcal {D} = \{\mathcal {A} _ {n} \} _ {N}. \\ \end{array} +$$ + +We instantiate $g$ by a powerful LVLM (e.g., GPT-4o [39]), considering that this is quite an open and complex problem. $g$ prioritizes disentangling and describing commonalities within the VC while analyzing each individual image. Various commonalities among any certain number of image patches should be summarized and described by precise terminologies. These common semantics within an open scenario are as diverse as they are abundant, unfolding in countless forms and nuances. Besides, it should generate specific atoms when any commonalities identified above do not appear in the current $I_{n}$ . Therefore, the art of prompt engineering prompt-com becomes paramount. It should provide rules for recognizing commonalities, possible conceptual directions, and output format control. In + +this paper, we prompt 13 high-level and low-level semantical directions that cover the normal concepts encountered in daily life, including object category, scene, object part, color, texture, material, position, transparency, brightness, shape, size, edges, and their relationships. Additionally, the CoT techniques [21, 51] are introduced. Details of all sophisticated prompts are shown in Sec. S2 of the Appendix. + +All $\mathcal{A}_n$ are then aggregated into a description atom set $\mathcal{A}$ with all duplicate atoms removed: + +$$ +\begin{array}{l} \mathcal {A} = \left\{\mathcal {A} _ {1} \cup \mathcal {A} _ {2} \cup \dots \cup \mathcal {A} _ {N} \right\} \tag {4} \\ = \left\{a ^ {1}, \dots , a ^ {i}, \dots , a ^ {P} \right\}, \\ \end{array} +$$ + +where $P$ is the number of concept atoms after deleting the repeating atoms. Each element in $\mathcal{A}$ represents a disentangled monosemantic concept and can be utilized as a node of the CoE. However, some atoms are semantically equivalent even if they have variant descriptions [19]. In order to remove semantically redundant atoms, we introduce an entailment function $\delta$ to detect the relations between atoms: + +$$ +\{1, 0, - 1 \} = \delta \left(a ^ {i 1}, a ^ {i 2}\right), \forall a ^ {i 1}, a ^ {i 2} \in \mathcal {A}, \tag {5} +$$ + +where the responses 1,0,-1 represent entailment, neutral, or contradiction of the candidates, respectively. We instantiate $\delta$ by an NLI model (e.g., DeBERTa model [25]) because the atom candidates are relatively short and simple that an NLI model is good enough. Following [19], a bidirectional entailment check is applied, meaning that one of the candidates will be removed when the response is entailment or neutral. The first atom is designated as the monosemantic representative of the semantic equivalence set due to the transitivity between them. The final concept set $A^{*}$ only contains heterogeneous atoms that describe the current VC: + +$$ +\mathcal {A} ^ {*} = \left\{a ^ {1}, \dots , a ^ {i}, \dots , a ^ {P ^ {*}} \right\}. \tag {6} +$$ + +After disentangling the concept atoms, we define a filtering function $z$ to select an atom $a^*$ as the most suitable explanation of current VC with respect to the input $x$ : + +$$ +a _ {s} ^ {*} (x) = z \left(\mathcal {A} _ {s} ^ {*}, C (x)\right). \tag {7} +$$ + +Some research ignores this step or selects the concept randomly with an implicit assumption that all concepts in $\mathcal{A}^*$ are semantically equivalent [14, 26]. In contrast, we filter the most suitable atom $a^*$ based on the context of the CoE. In particular, $z$ is instantiated by an LLM. $C(x)$ is an image captioning function utilized for providing contextual information. It can be instantiated by a relatively small LVLM [9, 10] or a professional model [30, 49]. + +# 3.2.2 Quantification of Concept Polysemanticity + +The polysemanticity of concepts can serve as a metric for the model's interpretability. Lower polysemanticity of concepts indicates reduced semantic uncertainty with a more + +singular form of activation, leading to greater interpretability. In this paper, we endeavor to quantify polysemanticity from three levels of granularity (i.e., conceptual channels, layers, and models) by defining a CPE score. The larger the CPE, the more pronounced the polysemanticity, which leads to decreased interpretability. + +In particular, the disentanglement function $g$ in Eq. 3 enables us to count the frequency of each atom in $\mathcal{A}^*$ . Given two fixed numbers $Q$ and $N$ , we calculate the probability of the $i_{th}$ atom within $j_{th}$ concept: + +$$ +p _ {i} = \frac {\text {N u m} _ {i}}{Q \times N + P a d}. \tag {8} +$$ + +Then, the CPE of $j_{th}$ concept can be formulated as + +$$ +H _ {j} ^ {*} = \frac {- \sum_ {i = 1} ^ {P ^ {*} + P a d} p _ {i} \log p _ {i}}{\log (P ^ {*} + P a d)}. \tag {9} +$$ + +The denominator normalizes the entropy value to a range between 0 and 1, considering that the size of $P^*$ varies for different VCs. Besides, the quantity of $A^*$ is set to be at least equal to $N$ due to some $D_s$ having a small number of semantical equivalent atoms with equal probabilities. If there are insufficient unique atoms, we supplement the corresponding number of atoms, with each supplementary atom set to a number of 1. The atomic objects are expanded, and the denominator is increased by the number of supplemented atoms (a total amount of $Pad$ ). Please refer to Sec. S3.1 of the Appendix for details. In addition, two averaged CPEs are formulated to enable comparisons of polysemantics between specific layers or models: + +$$ +H _ {l} ^ {*} = \frac {1}{d _ {l}} \sum^ {d _ {l}} H _ {j} ^ {*}, \tag {10} +$$ + +$$ +H _ {\mathcal {M}} ^ {*} = \frac {1}{L} \sum^ {L} H _ {l} ^ {*}, \tag {11} +$$ + +where $d_{l}$ is the dimension of the channel and $L$ is the number of layers. Finally, to accommodate the diverse needs of researchers within the community, the automatically constructed concept explanation dataset is extended as ACD $\mathcal{B}_{\mathcal{M},\mathcal{T}} = \{\mathcal{C},\mathcal{S},\mathcal{D},\mathcal{A}^{*},\mathcal{Q},\mathcal{H},f\}$ , where $\mathcal{Q}$ represents the probability set of each atom and $\mathcal{H}$ is the set of CPE. + +# 3.3. Construction of Explanation Chain + +Concept circuits in DVMs can identify the most activated units in each layer for a specific input [3]. When aggregating these units across all key layers in a bottom-up manner, a corresponding maximum activation path $\mathcal{P}$ is generated, establishing the foundation for the explanation chain. Each node in this path serves as the explanation for the model's decision at that layer. Traditional concept circuit methods for DVMs identify the activation path without providing thorough overarching natural language-based explanations. Inspired by the CoT explanation technique in LLMs + +[12], CoE proposes a local explanation chain to interpret the decision-making process of DVMs in the form of natural language descriptions based on ACD- $\mathcal{B}_{\mathcal{M},\mathcal{T}}$ , CPDF, concept circuit method, and LLM, as shown in Fig. 2. + +Specifically, some user-predefined layers requiring explanations or key layers facilitating CoE (e.g., output layers of 4 stages of a DVM) are listed in a bottom-up manner. Given an image $x$ and layer $l$ , an advanced XAI method $E_{x}$ computes the maximization normalized activation or relevance values for all concepts. Each value represents the importance of the current concept to the prediction $\mathcal{M}(x)$ . CoE acknowledges top $k^l$ relevant concepts $\mathcal{C}^l(x) = \{c_1^l, \dots, c_{k^l}^l\}$ and their values $\mathcal{V}^l(x) = \{v_1^l, \dots, v_{k^l}^l\}$ , determined by the $\alpha$ quantile of relevance values across all concepts (i.e., $k^l = |\{v_j^l | v_j^l > \alpha \max(v^l), \forall j \in d^l\}|$ , where $|\mathcal{X}|$ computes the number of elements). After that, the descriptive concept atoms of the current layer $\widetilde{\mathcal{A}}_{k^l}^{*,l}(x) = \{a_1^{*,l}, \dots, a_{k^l}^{*,l}\}$ are constructed based on ACD- $\mathcal{B}_{\mathcal{M},\mathcal{T}}$ and the filter Eq. 7, serving as a node of the explanation chain. + +The final CoE for the local decision-making explanation, abstracting and synthesizing all top activated concept atoms along the chain, is formulated as + +$$ +C o E _ {\mathcal {M}} (x) = e \left(\widetilde {\mathcal {A}} _ {\mathcal {P}} ^ {*} (x), \mathcal {V} _ {\mathcal {P}} (x), C (x), G (x), \mathcal {M} (x)\right), \tag {12} +$$ + +where $e$ formulates a synthesizing and describing function with inputs obtained from the above sections, and $G(x)$ is the label of $x$ . We instantiate $CoE_{\mathcal{M}}(x)$ through a powerful LLM (e.g., GPT-4 [1]), with prompt prompt-coe well-designed by combining CoT and few-shot prompting [44]. + +Additionally, these local explanations are evaluated by a powerful LVLM and human judges based on three explainability metrics, each utilizing a three-tier scoring system. + +# 4. Experiments + +We evaluate the proposed CoE approach on two types of DVMs (i.e., ResNet [24] and CLIP [40]). The relevance-based XAI method CRP is adopted as the primary approach for $E_{x}$ in this paper, owing to its demonstrated superiority in terms of fidelity and reliability [2]. Please refer to the Appendix for more details on the implementation and results. Codes are available at https://github.com/YuWLong666/CoE. + +# 4.1. Concept Explanation Dataset Construction + +In the CoE, we first automatically construct global concept description databases $\mathsf{ACD - B}_{\mathcal{M},\mathcal{T}}$ for various DVMs. Taking a ResNet152 model as an example, 4 output layers of 4 stages (with 256, 512, 1024, and 2048 channels $\mathcal{C}$ , respectively) are selected as the key layer set. CRP is applied as $E_{x}$ to extract the VC set $\mathcal{S}$ . Following [2], each VC is represented by 15 (i.e., $N$ ) image patches. The mapping function + +Table 1. Examples of VCs and their disentangled concept atoms, along with their concept probability distributions and CPE scores. The first row shows concept 75 of Layer 4 extracted by CRP. The second row shows concept 697 of Layer 4 extracted by MILAN. + +
Visual Concept and Disentangled Concept AtomsConcept DistributionCPE
Image 1: [burger, lettuce, tomato], +Image 2: [blue, monkey, face], +Image 3: [shark, blue, water], +Image 4: [keyboard, buttons, numbers], +Image 5: [shark, blue, water], +Image 6: [shark, blue, water], +Image 7: [shark, blue, water], +Image 8: [blue, circular, shape], +Image 9: [shark, blue, water], +Image 10: [keyboard, buttons, numbers], +Image 11: [blue, yellow, objects], +Image 12: [shark, blue, text], +Image 13: [shark, blue, water], +Image 14: [shark, blue, water], +Image 15: [shark, blue, water]0.30 +0.25 +0.20 +0.15 +0.10 +0.05 +0.00 +hurty green, blue, texture, yellow, objects, water, black0.81
Image 1: [hand, blinds, string], +Image 2: [curtains, room, grayscale], +Image 3: [fabric, green, closeup], +Image 4: [curtains, room, light], +Image 5: [reflection, glass, light], +Image 6: [hand, fabric, texture], +Image 7: [water, fountain, greenery], +Image 8: [feathers, pattern, dark], +Image 9: [fabric, blue, closeup], +Image 10: [fabric, green, seatbelt], +Image 11: [leaf, bark, texture], +Image 12: [nature, mountain, grayscale], +Image 13: [skull, ground, dark], +Image 14: [fabric, colorful, pattern], +Image 15: [curtains, window, light]0.12 +0.10 +0.08 +0.06 +0.04 +0.02 +0.00 +0.00 +0.00 +0.00 +0.00 +0.00 +0.00 +0.00 +0.00 +0.00 +0.00 +0.00 +0.00 +0.00 +0.00 +0.00 +0.00 +0.00 +0.00 +0.00 +0.02 +0.02 +0.02 +0.02 +0.02 +0.02 +0.02 +0.02 +0.02 +0.02 +0.02 +0.02 +0.02 +0.02 +0.02 +0.02 +0.02 +0.02 +0.02 +0.02 +0.00 +0.00 +0.00 +0.00 +0.00 +0.00 +0.00 +0.00 +0.00 +0.00 +0.00 +0.00 +0.00 +0.00 +0.00 +0.00 +0.00 +0.00 +0.02 +0.00 +0.00 +0.00 +0.00 +0.00 +0.00 +0.00 +0.00 +0.00 +0.00 +0.00 +0.00 +0.00 +0.00 +0.00 +0.00 +0.00 +0.00 +0.01 +0.01 +0.01 +0.01 +0.01 +0.01 +0.01 +0.01 +0.01 +0.01 +0.01 +0.01 +0.01 +0.01 +0.01 +0.01 +0.01 +0.01 +0.01 +0.01 +0.02 +0.02 +0.02 +0.02 +0.02 +0.02 +0.02 +0.02 +0.02 +0.02 +0.02 +0.02 +0.02 +0.02 +0.02 +0.02 +0.02 +0.02 +0.02 +0.03 +0.03 +0.03 +0.03 +0.03 +0.03 +0.03 +0.03 +0.03 +0.03 +0.03 +0.03 +0.03 +0.03 +0.03 +0.03 +0.03 +0.03 +0.03 +0.03 +0.04 +0.04 +0.04 +0.04 +0.04 +0.04 +0.04 +0.04 +0.04 +0.04 +0.04 +0.04 +0.04 +0.04 +0.04 +0.04 +0.04 +0.04 +0.04 +0.04 +0.05 +0.05 +0.05 +0.05 +0.05 +0.05 +0.05 +0.05 +0.05 +0.05 +0.05 +0.05 +0.05 +0.05 +0.05 +0.05 +0.05 +0.05 +0.05 +0.05 +0.06 +0.06 +0.06 +0.06 +0.06 +0.06 +0.06 +0.06 +0.06 +0.06 +0.06 +0.06 +0.06 +0.06 +0.06 +0.06 +0.06 +0.06 +0.06 +0.06 +0.07 +0.07 +0.07 +0.07 +0.07 +0.07 +0.07 +0.07 +0.07 +0.07 +0.07 +0.07 +0.07 +0.07 +0.07 +0.07 +0.07 +0.07 +0.07 +0.07 +0.08 +0.08 +0.08 +0.08 +0.08 +0.08 +0.08 +0.08 +0.08 +0.08 +0.08 +0.08 +0.08 +0.08 +0.08 +0.08 +0.08 +0.08 +0.08 +0.08 +0.09 +0.09 +0.09 +0.09 +0.09 +0.09 +0.09 +0.09 +0.09 +0.09 +0.09 +0.09 +0.09 +0.09 +0.09 +0.09 +0.09 +0.09 +0.09 +0.09 +0.08 +0.08 +0.08 +0.08 +0.08 +0.08 +0.08 +0.08 +0.08 +0.08 +0.08 +0.08 +0.08 +0.08 +0.08 +0.08 +0.08 +0.08 +0.08 +0.07 +0.07 +0.07 +0.07 +0.07 +0.07 +0.07 +0.07 +0.07 +0.07 +0.07 +0.07 +0.07 +0.07 +0.07 +0.07 +0.07 +0.07 +0.07 +0.06 +0.06 +0.06 +0.06 +0.06 +0.06 +0.06 +0.06 +0.06 +0.06 +0.06 +0.06 +0.06 +0.06 +0.06 +0.06 +0.06 +0.06 +0.06 +0.05 +0.05 +0.05 +0.05 +0.05 +0.05 +0.05 +0.05 +0.05 +0.05 +0.05 +0.05 +0.05 +0.05 +0.05 +0.05 +0.05 +0.05 +0.05 +0.07 +0.07 +0.07 +0.07 +0.07 +0.07 +0.07 +0.07 +0.07 +0.07 +0.07 +0.07 +0.07 +0.07 +0.07 +0.07 +0.07 +0.07 +0.07 +0.09 +0.09 +0.09 +0.09 +0.09 +0.09 +0.09 +0.09 +0.09 +0.09 +0.09 +0.09 +0.09 +0.09 +0.09 +0.09 +0.09 +0.09 +0.09 +0.07 +0.07 +0.07 +0.07 +0.07 +0.07 +0.07 +0.07 +0.07 +0.07 +0.07 +0.07 +0.07 +0.07 +0.07 +0.07 +0.07 +0.07 +0.07 +0.05 +0.05 +0.05 +0.05 +0.05 +0.05 +0.05 +0.05 +0.05 +0.05 +0.05 +0.05 +0.05 +0.05 +0.05 +0.05 +0.05 +0.05 +0.05 +0.04 +0.04 +0.04 +0.04 +0.04 +0.04 +0.04 +0.04 +0.04 +0.04 +0.04 +0.04 +0.04 +0.04 +0.04 +0.04 +0.04 +0.04 +0.04 +0.06 +0.06 +0.06 +0.06 +0.06 +0.06 +0.06 +0.06 +0.06 +0.06 +0.06 +0.06 +0.06 +0.06 +0.06 +0.06 +0.06 +0.06 +0.06 +0.08 +0.08 +0.08 +0.08 +0.08 +0.08 +0.08 +0.08 +0.08 +0.08 +0.08 +0.08 +0.08 +0.08 +0.08 +0.08 +0.08 +0.08 +0.08 +0.06 +0.06 +0.06 +0.06 +0.06 +0.06 +0.06 +0.06 +0.06 +0.06 +0.06 +0.06 +0.06 +0.06 +0.06 +0.06 +0.06 +0.06 +0.06 +0.09 +0.09 +0.09 +0.09 +0.09 +0.09 +0.09 +0.09 +0.09 +0.09 +0.09 +0.09 +0.09 +0.09 +0.09 +0.09 +0.09 +0.09 +0.09 +0.06 +0.06 +0.06 +0.06 +0.06 +0.06 +0.06 +0.06 +0.06 +0.06 +0.06 +0.06 +0.06 +0.06 +0.06 +0.06 +0.06 +0.06 +0.06 +0.04 +0.04 +0.04 +0.04 +0.04 +0.04 +0.04 +0.04 +0.04 +0.04 +0.04 +0.04 +0.04 +0.04 +0.04 +0.04 +0.04 +0.04 +0.04 +0.09 +0.09 +0.09 +0.09 +0.09 +0.09 +0.09 +0.09 +0.09 +0.09 +0.09 +0.09 +0.09 +0.09 +0.09 +0.09 +0.09 +0.09 +0.09 +0.05 +0.05 +0.05 +0.05 +0.05 +0.05 +0.05 +0.05 +0.05 +0.05 +0.05 +0.05 +0.05 +0.05 +0.05 +0.05 +0.05 +0.05 +0.05 +0.09 +0.09 +0.09 +0.09 +0.09 +0.09 +0.09 +0.09 +0.09 +0.09 +0.09 +0.09 +0.09 +0.09 +0.09 +0.09 +0.09 +0.09 +0.09 +0.02 +0.12 +0.12 +0.12 +0.12 +0.12 +0.12 +0.12 +0.12 +0.12 +0.12 +0.12 +0.12 +0.12 +0.12 +0.12 +0.12 +0.12 +0.12 +0.12 +0.12 +0.25 +0.25 +0.25 +0.25 +0.25 +0.25 +0.25 +0.25 +0.25 +0.25 +0.25 +0.25 +0.25 +0.25 +0.25 +0.25 +0.25 +0.25 +0.25 +0.25 +0.26 +0.26 +0.26 +0.26 +0.26 +0.26 +0.26 +0.26 +0.26 +0.26 +0.26 +0.26 +0.26 +0.26 +0.26 +0.26 +0.26 +0.26 +0.26 +0.26 +0.25 +0.25 +0.25 +0.25 +0.25 +0.25 +0.25 +0.25 +0.25 +0.25 +0.25 +0.25 +0.25 +0.25 +0.25 +0.25 +0.25 +0.25 +0.25 +0.2. +12 +12 +12 +12 +12 +12 +12 +12 +12 +12 +12 +12 +12 +12 +12 +12 +12 +12 +12 +12 +12 +12 +12 +12 +12 +12 +12 +12 +12 +12 +12 +12 +12 +12 +25 +25 +25 +25 +25 +25 +25 +25 +25 +25 +25 +25 +25 +25 +25 +25 +25 +25 +25 +25 +25 +25 +25 +25 +25 +25 +25 +25 +25 +25 +25 +25 +25 +25
+ +$f$ from channels to VCs is built. We then employ a GPT4o-2024-08-06 model as the function $g$ to generate the description set $\mathcal{D}$ (i.e., Eq. 3). The disentangled and clustered atom set $\mathcal{A}^*$ (processed by a DeBERTa model), the probability set $\mathcal{Q}$ , and the CPE set $\mathcal{H}$ are computed and gathered accordingly. Overall, ACD- $\mathcal{B}_{\mathcal{M},\mathcal{T}}$ containing 3,840 entries is constructed, with each entry comprising a channel, a CPE score, and 15 VC representations—each consisting of 3 concept atoms (i.e., $Q$ )—and multiple orthogonal linguistic atoms, along with their associated probabilities. + +# 4.2. Disentanglement of Polysemantic Concept + +CPDF automatically disentangles and describes the polysemantic VCs into a set of linguistic concept atoms, $\mathcal{D}$ . As exemplified by Table 1, two randomly selected VCs are extracted by two advanced XAI methods (CRP and MILAN). They are disentangled and described into atoms, exhibiting different degrees of polysemasticity. We observe that the commonalities within the VCs are almost entirely identified, and most of the atoms describe the commonalities of subsets of image patches. In the first row (CPE=0.81), $g$ successfully identifies all visually discernible commonalities, such as shark and water. The atoms exhibit high consistency, and fewer but more frequently occurring common semantics are identified. Only a few atoms, such as the burger, are not commonalities, as their corresponding images lack sufficient common semantics. The atoms are also repre + +sented accurately. For instance, the atom sets for each image featuring the shark semantic all include the word shark. In the second row (CPE=0.95), our method remains effective in disentangling and describing VC with a higher level of polysemantics. $g$ identifies a greater variety of atoms. These results demonstrate the effectiveness of the disentanglement and description operation in the CPDF. + +# 4.3. Concept Distribution and CPE + +CPE can quantify the polysemanticity and serve as an indicator of a DVM's interpretability. It begins by calculating the probabilities of each orthogonal atom. As the middle column of Table 1 shows, the distribution provides statistics on the extent to which different atoms are associated with each concept. For instance, concept 75 primarily focuses on blue with a probability of 0.27 while also identifying yellow with a smaller probability. The semantics between atoms are non-overlapping. This probabilistic approach enhances explainability by capturing multiple semantics with their occurring probabilities, upgrading the modeling of deterministic concepts to uncertain atom distributions. + +CPE quantifies polysemanticity from 3 levels of granularity. As shown in Fig. 3a, concepts exhibit varying degrees of polysemanticity, with considerable fluctuations (ranging from a maximum of 1 to a minimum of 0.5). Most of the concepts demonstrate a noticeable level of polysemanticity. When analyzing the polysemanticity of different + +![](images/837cd38290cbb3b4ff70a449f9fa6f7065fa0250c4ccccc3b23df402c253ae85.jpg) +(a) Sorted CPE along the channel dimension. + +![](images/e26469c49c6efe022b123a098cd7322cb4218aa12f003390f9d1044ee078a1e0.jpg) +(b) Averaged CPE on 4 key layers. + +![](images/e8ed054dee5d0031be4da582a14299832de27b55662fe516aa50fa044e9a1663.jpg) +(c) Averaged CPE on 3 models. +Figure 3. Illustrations of CPE scores at 3 levels of granularity. In (a), the channel indices of each layer are max normalized to 1. + +Table 2. The consistency of the recognition of concept polysemantics between the CPE metric and human scorers. + +
Consistency
CPE Metric and Human Scorers75 %
+ +layers, as shown in Fig. 3b, the averaged CPE peaks in the third stage, while the shallowest and deepest layers exhibit lower values. Typically, deeper layers capture abstract concepts, whereas the shallower layers extract elementary features like textures [53, 56]. DVMs show better monosemanticity when identifying elementary and abstract concepts, while more ambiguous concepts are captured in the intermediate layers. This experiment validates the theoretical and experimental claims of previous work [14, 32]. In Fig. 3c, we observe that as the number of layers increases, the CPE of the model shows a rising trend. This result indicates a direct proportional relationship between the polysemanticity and the complexity of DVMs, which is that the more complex a DVM becomes, the harder it is to interpret. + +The CPE is evaluated in comparison with human assessments. As shown in Table 1, the CPE is consistent with the degree of polysemanticity from the human intuitive observations (the polysemanticity of the first row is smaller than that of the second row). Similarly, we randomly sample 300 pairs of VCs, each containing two VCs with varying levels of polysemanticity. Comparisons are drawn from these 300 pairs by both human evaluators and the CPE metric. As shown in Table 2, results from human assessments and CPE achieve a $75\%$ agreement rate, demonstrating the effectiveness of CPE. Other results are shown in Sec. S4 of the Appendix. All of these results reveal that CPE is effective in quantifying the polysemanticity of concepts and DVMs. + +# 4.4. CoE for Local Explanations + +CoE generates local explanations akin to CoT by automatically describing the concept circuit. We set $\alpha = 0.001$ to acknowledge the top relevant VCs, resulting in an average of 3 key VCs per layer. As exemplified in Table 3, given an input image, CoE first outputs whether the prediction of the + +DVM is correct and then gives reasons behind this prediction according to the information along the concept circuits. CoE employs an LLM as an engine to infer the logical relationships among concepts, as well as between concepts and the context. It synthesizes all evidence to generate an inductive commentary on the predictions. The explanation chain is presented in a linguistic format, enhancing human understanding. For instance, CoE explains why the DVM incorrectly predicted a tench as a mosquito net: the top activation of mosquito-net-related concepts, rather than those related to fish, misled the decision. Additional instances are shown in Sec. S5 of the Appendix. Overall, CoE effectively captures and elucidates the core decision-making evidence for both correct and incorrect predictions of DVMs. + +# 4.5. Quantitative Evaluation via GPT-4o & Human + +We employ GPT-4o and human evaluators to assess the local explanations for 500 and 100 samples randomly selected from the ImageNet Validation dataset [15], respectively. The GPT-4o evaluation is achieved through a carefully designed prompt prompt-coe-eval. Explanations generated by three methods are rated according to three explainability metrics: Accuracy, Completeness, and User Interpretability [47]. Each is assigned a three-tier scoring system (2, 1, and 0), with specific criteria defined for each tier, resulting in a total explanation score of 6. Please refer to Sec. S2 and S5 of the Appendix for details. + +As shown in Table 4, our CoE approach attains peak performance across all three metrics, achieving total scores of 5.06 and 5.04 in GPT-4o and human evaluations, respectively. The lowest scores (3.14 and 2.62) are obtained on the MILAN ANNOTATION dataset without the ACD and CPDF methods. This dataset was manually annotated, resulting in inaccuracy and insufficiency within concept explanations. Scores increase by 0.11 and 0.72, respectively, with the application of our proposed ACD- $\mathcal{B}_{\mathcal{M},\mathcal{T}}$ dataset. The ACD method not only enables the automated disentanglement and description of concept atoms but also enhances their precision and comprehensibility. When the CoE em + +Table 3. Cases of local explanations of CoE. The first row shows the explanation of a correct prediction, and the second row is reversed. + +![](images/6649d862f2f93427734b895aff043653f785e8b10d9b9ad65eed48d8fc7b2ee9.jpg) +label: + +![](images/e6e0418a4bb3aa96e7aba5d4fc71145eef7293294ccc89b41a5b018021469042.jpg) +black-and-tan coonhound +predict: +black-and-tan coonhound + +The model outputs a correct result: black-and-tan coonhound. Specifically, in the shallowest layer, layer 1, channel 88 with a relevance value of 1.0 describes the concept curved, which may relate to the curved shapes observed in parts of the dog such as its tail or ears. Moving deeper, in layer 2, channel 506 with a relevance value of 0.85 and channel 391 with a relevance value of 0.75 describe the concept of animal, indicating the model has identified an animal in the image, and channel 229 with a relevance value of 0.77 emphasizes the color black, aligning with the black and tan colors of the coonhound. Further, layer 2 also identifies the presence of grass several times (channels 94 and 103), which corresponds to the grassy field described in the image caption. In layer 3, various channels, such as 985, 174, 422, and 13, with high relevance scores, from 0.72 to 1.0, repeatedly identify the color green and the concept grass, firmly establishing the setting of the image. Additionally, channels 11, 787, and 591 in layer 3 enhance the certainty by repeatedly identifying animal and creature, confirming that the subject is an animal. Finally, layer 4 decisively contributes with channels 1354 and 1661, both signifying dog with high relevance values (1.0 and 0.72 respectively), which correctly identifies the animal in the image as a dog. + +Given that the model's structured concepts strongly match the description and characteristics noted in the caption and the ground truth label, it is clear why the model outputs the correct result of black-and-tan coonhound. + +![](images/3f9920565c7c12df94010542ee15f666995976bb18fbb963c0e8f48eb2c8fc38.jpg) +tench +mosquito net + +![](images/1982e38d7a411595f21ae29d6975d4eb3f1390c4182d501fe2578a0618818a49.jpg) +label: +predict: + +The model outputs an incorrect result mosquito net instead of the correct label tench. Specifically, in the shallowest layer 1 of the model, channel 88 with a relevance value of 1.0 describes the concept of smooth, channel 61 with a relevance value of 0.81 describes the concept of green, and channel 111 with a relevance value of 0.72 describes the concept of brown. In the deeper layer 2 of the model, channel 94 with a relevance value of 1.0 describes the concept of water, channel 229 with a relevance value of 0.84 describes the concept of fish, channel 103 with a relevance value of 0.81 also describes the concept of fish, and channel 467 with a relevance value of 0.73 describes the concept of mesh. At layer 3, channel 985 with a relevance value of 1.0 further describes the concept of fish, while channels 894, 122, and 715 with relevance values of 0.83, 0.73, and 0.72, respectively, describe the concept of green, possibly relating to the color of the net. In the deepest layer 4, channel 697 with a relevance value of 1.0 describes the concept of net, channel 1878 with a relevance value of 0.76 describes the concept of baby, and channel 321 with a relevance value of 0.76 represents the concept of hexagon, which could misconstrue the net structure. The prevalence of concepts related to fish within the decision path suggests a relationship to the tench, but the final prediction of mosquito net was likely influenced by the dominance of net-related concepts. Thus, the model was misled by the presence of a net, resulting in the incorrect classification of mosquito net. + +Table 4. Scores of CoE for the local explanation on three methods evaluated by GPT-4o and human scorers. The first method utilizes MILAN ANNOTATION as the concept explanation dataset, while the rest two work on the ACD- $\mathcal{B}_{\mathcal{M},\mathcal{T}}$ datasets. Each criterion has a maximum score of 2 points. The total explanation score is 6 points. Higher scores indicate better explainability. + +
MethodAccuracyCompletenessUser Interpret.Total Explanation
GPT-4oMILAN A.[26] + Description (Baseline)1.031.071.043.14
CoE w/o Filtering1.071.101.083.25
CoE (ours)1.691.701.675.06
HumanMILAN A.[26] + Description (Baseline)0.870.830.922.62
CoE w/o Filtering1.101.161.073.34
CoE (ours)1.731.721.595.04
+ +ploys the full CPDF mechanism, the explanation score rises by 1.92 and 2.42 points (the average absolute improvement is $36\%$ ), respectively, with the filtering step contributing 1.81 and 1.70 points. This result suggests that by taking polysemanticity and contextual filtering of concepts into account, CoE effectively guides the LLM in producing better linguistic explanations, substantially enhancing their accuracy, completeness, and interpretability for users. The results of the human evaluation are consistent with the GPT-4o results. All of these results strongly affirm the superiority of the proposed CoE approach. + +# 5. Conclusion + +In this paper, we propose a CoE approach to tackle the challenges of the inflexibility to automatically construct con + +cept explanation datasets, insufficient linguistic explanations, and the weakness in managing concept polysemantics of concept-based XAI methods. CoE automates the decoding and description of concepts via an LVLM and constructs global concept explanation datasets. A CPDF mechanism is designed to identify the most contextually relevant concept atoms, advancing the modeling of deterministic concepts to uncertain concept atom distributions. The CPE is proposed to quantify the polysemanticsy. Finally, CoE enables local explanations of a DVM's decisionmaking process, represented through natural language. The effectiveness and superiority of CPE and CoE are demonstrated. In summary, CoE establishes a systematical framework for explaining DVMs, and we believe that the findings presented in this paper warrant further exploration. + +# Acknowledgment + +This research is supported in part by the National Natural Science Foundation of China under Grants U23B2049, 62276186, and 61925602. + +# References + +[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023.5 +[2] Reduan Achtibat, Maximilian Dreyer, Ilona Eisenbraun, Sebastian Bosse, Thomas Wiegand, Wojciech Samek, and Sebastian Lapuschkin. From attribution maps to human-understandable explanations through concept relevance propagation. Nature Machine Intelligence, 5(9):1006-1019, 2023. 1, 2, 5 +[3] Rajaram Achyuta, Chowdhury Neil, Torralba Antonio, Andreas Jacob, and Schwettmann Sarah. Automatic discovery of visual circuits. In NeurIPS Workshop on Attributing Model Behavior at Scale, 2023. 1, 2, 5 +[4] Sebastian Bach, Alexander Binder, Grégoire Montavon, Frederick Klauschen, Klaus-Robert Müller, and Wojciech Samek. On pixel-wise explanations for non-linear classifier decisions by layer-wise relevance propagation. PLoS ONE, 10, 2015. 1 +[5] David Bau, Jun-Yan Zhu, Hendrik Strobelt, Agata Lapedriza, Bolei Zhou, and Antonio Torralba. Understanding the role of individual units in a deep neural network. Proceedings of the National Academy of Sciences, 117(48):30071-30078, 2020. 2 +[6] Leonard Bereska and Efstratios Gavves. Mechanistic interpretability for ai safety-a review. arXiv preprint arXiv:2404.14082, 2024. 1 +[7] Subrato Bharati, M Rubaiyat Hossain Mondal, and Prajoy Podder. A review on explainable artificial intelligence for healthcare: why, how, and when? IEEE Transactions on Artificial Intelligence, 2023. 1 +[8] Moritz Böhle, Navdeeppal Singh, Mario Fritz, and Bernt Schiele. B-cos alignment for inherently interpretable cnns and vision transformers. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(6):4504-4518, 2024. 2 +[9] Zhe Chen, Weiyun Wang, Hao Tian, Shenglong Ye, Zhang-wei Gao, Erfei Cui, Wenwen Tong, Kongzhi Hu, Jiapeng Luo, Zheng Ma, et al. How far are we to gpt-4v? closing the gap to commercial multimodal models with open-source suites. arXiv preprint arXiv:2404.16821, 2024. 4 +[10] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24185–24198, 2024. 4 +[11] Ming-Ming Cheng, Peng-Tao Jiang, Ling-Hao Han, Liang Wang, and Philip Torr. Deeply explain cnn via hierarchical + +decomposition. International Journal of Computer Vision, 131(5):1091-1105, 2023. 2 +[12] Zheng Chu, Jingchang Chen, Qianglong Chen, Weijiang Yu, Tao He, Haotian Wang, Weihua Peng, Ming Liu, Bing Qin, and Ting Liu. Navigate through enigmatic labyrinth a survey of chain of thought reasoning: Advances, frontiers and future. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1173–1203, 2024. 5 +[13] Arthur Conmy, Augustine Mavor-Parker, Aengus Lynch, Stefan Heimersheim, and Adrià Garriga-Alonso. Towards automated circuit discovery for mechanistic interpretability. Advances in Neural Information Processing Systems, 36: 16318-16352, 2023. 2 +[14] Meghal Dani, Isabel Rio-Torto, Stephan Alaniz, and Zeynep Akata. Devil: Decoding vision features into language. In DAGM German Conference on Pattern Recognition, pages 363-377. Springer, 2023. 1, 4, 7 +[15] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 7 +[16] Maximilian Dreyer, Erblina Purelku, Johanna Vielhaben, Wojciech Samek, and Sebastian Lapuschkin. Pure: Turning polysemantic neurons into pure features by identifying relevant circuits. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 8212-8217, 2024. 1, 2, 3 +[17] Nelson Elhage, Neel Nanda, Catherine Olsson, Tom Henighan, Nicholas Joseph, Ben Mann, Amanda Askell, Yuntao Bai, Anna Chen, Tom Conerly, et al. A mathematical framework for transformer circuits. Transformer Circuits Thread, 1(1):12, 2021. 2 +[18] Nelson Elhage, Tristan Hume, Catherine Olsson, Nicholas Schiefer, Tom Henighan, Shauna Kravec, Zac Hatfield-Dodds, Robert Lasenby, Dawn Drain, Carol Chen, et al. Toy models of superposition. arXiv preprint arXiv:2209.10652, 2022.3 +[19] Sebastian Farquhar, Jannik Kossen, Lorenz Kuhn, and Yarin Gal. Detecting hallucinations in large language models using semantic entropy. Nature, 630(8017):625-630, 2024. 4 +[20] Thomas Fel, Victor Boutin, Louis Béthune, Rémi Cadène, Mazda Moayeri, Léo Andéol, Mathieu Chalvidal, and Thomas Serre. A holistic approach to unifying automatic concept extraction and concept importance estimation. Advances in Neural Information Processing Systems, 36, 2024. 2, 3 +[21] Guhao Feng, Bohang Zhang, Yuntian Gu, Haotian Ye, Di He, and Liwei Wang. Towards revealing the mystery behind chain of thought: a theoretical perspective. Advances in Neural Information Processing Systems, 36, 2024. 2, 4 +[22] Amirata Ghorbani, James Wexler, James Y Zou, and Been Kim. Towards automatic concept-based explanations. Advances in neural information processing systems, 32, 2019. 1, 2 +[23] Dan W Grupe and Jack B Nitschke. Uncertainty and anticipation in anxiety: an integrated neurobiological and psy + +chological perspective. Nature Reviews Neuroscience, 14(7): 488-501, 2013. 2 +[24] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 770-778, 2016. 5 +[25] Pengcheng He, Xiaodong Liu, Jianfeng Gao, and Weizhu Chen. Deberta: Decoding-enhanced bert with disentangled attention. In International Conference on Learning Representations, 2020. 4 +[26] Evan Hernandez, Sarah Schwettmann, David Bau, Teona Bagashvili, Antonio Torralba, and Jacob Andreas. Natural language descriptions of deep features. In International Conference on Learning Representations, 2022. 1, 2, 4, 8 +[27] Been Kim, Martin Wattenberg, Justin Gilmer, Carrie Cai, James Wexler, Fernanda Viegas, et al. Interpretability beyond feature attribution: Quantitative testing with concept activation vectors (tcav). In International conference on machine learning, pages 2668-2677. PMLR, 2018. 1, 3 +[28] Anton Kuznietsov, Balint Gyevnar, Cheng Wang, Steven Peters, and Stefano V. Albrecht. Explainable ai for safe and trustworthy autonomous driving: A systematic review. IEEE Transactions on Intelligent Transportation Systems, pages 1-23, 2024. 1 +[29] Yann LeCun, Yoshua Bengio, and Geoffrey Hinton. Deep learning. nature, 521(7553):436-444, 2015. 1 +[30] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730-19742. PMLR, 2023. 4 +[31] Haomiao Liu, Ruiping Wang, Shiguang Shan, and Xilin Chen. What is a tabby? interpretable model decisions by learning attribute-based classification criteria. IEEE transactions on pattern analysis and machine intelligence, 43(5): 1791-1807, 2019. 2 +[32] Francesco Locatello, Stefan Bauer, Mario Lucic, Gunnar Raetsch, Sylvain Gelly, Bernhard Schölkopf, and Olivier Bachem. Challenging common assumptions in the unsupervised learning of disentangled representations. In international conference on machine learning, pages 4114-4124. PMLR, 2019. 7 +[33] Simon C Marshall and Jan H Kirchner. Understanding polysemantics in neural networks through coding theory. arXiv preprint arXiv:2401.17975, 2024. 2, 3 +[34] Dang Minh, H Xiang Wang, Y Fen Li, and Tan N Nguyen. Explainable artificial intelligence: a comprehensive review. Artificial Intelligence Review, pages 1-66, 2022. 1 +[35] Grégoire Montavon, Sebastian Lapuschkin, Alexander Binder, Wojciech Samek, and Klaus-Robert Müller. Explaining nonlinear classification decisions with deep taylor decomposition. Pattern recognition, 65:211-222, 2017. 2 +[36] Grégoire Montavon, Alexander Binder, Sebastian Lapuschkin, Wojciech Samek, and Klaus-Robert Müller. Layer-wise relevance propagation: an overview. Explainable AI: interpreting, explaining and visualizing deep learning, pages 193-209, 2019. 1, 2 + +[37] Meike Nauta, Ron van Bree, and Christin Seifert. Neural prototype trees for interpretable fine-grained image recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14933-14943, 2021. 2 +[38] Laura O'Mahony, Vincent Andrearczyk, Henning Müller, and Mara Graziani. Disentangling neuron representations with concept vectors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3770-3775, 2023. 3 +[39] OpenAI. Hello gpt-4o. https://openai.com/index/hellogpt-4o/, 2024.4 +[40] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 5 +[41] Yao Rong, Tobias Leemann, Thai-Trang Nguyen, Lisa Fiedler, Peizhu Qian, Vaibhav Unhelkar, Tina Seidel, Gjergji Kasneci, and Enkelejda Kasneci. Towards human-centered explainable ai: A survey of user studies for model explanations. IEEE transactions on pattern analysis and machine intelligence, 46(4):2104-2122, 2024. 1 +[42] Cynthia Rudin. Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead. Nature Machine Intelligence, 1(5):206-215, 2019. 2 +[43] Waddah Saeed and Christian Omlin. Explainable ai (xai): A systematic meta-survey of current challenges and future opportunities. Knowledge-Based Systems, 263:110273, 2023. 1, 2 +[44] Pranab Sahoo, Ayush Kumar Singh, Sriparna Saha, Vinija Jain, Samrat Mondal, and Aman Chadha. A systematic survey of prompt engineering in large language models: Techniques and applications. arXiv preprint arXiv:2402.07927, 2024.5 +[45] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE international conference on computer vision (ICCV), pages 618-626, 2017. 1, 2 +[46] Ao Sun, Pingchuan Ma, Yuanyuan Yuan, and Shuai Wang. Explain any concept: Segment anything meets concept-based explanation. Advances in Neural Information Processing Systems, 36, 2024. 1, 2 +[47] Giulia Vilone and Luca Longo. Notions of explainability and evaluation approaches for explainable artificial intelligence. Information Fusion, 76:89-106, 2021. 7 +[48] Kevin Ro Wang, Alexandre Variengien, Arthur Conmy, Buck Shlegeris, and Jacob Steinhardt. Interpretability in the wild: a circuit for indirect object identification in GPT-2 small. In The Eleventh International Conference on Learning Representations, ICLR 2023, Kigali, Rwanda, May 1-5, 2023. 2 +[49] Teng Wang, Jinrui Zhang, Junjie Fei, Hao Zheng, Yunlong Tang, Zhe Li, Mingqi Gao, and Shanshan Zhao. Caption + +anything: Interactive image description with diverse multimodal controls. arXiv preprint arXiv:2305.02677, 2023. 4 +[50] Xin Wang, Hong Chen, Si'ao Tang, Zihao Wu, and Wenwu Zhu. Disentangled representation learning. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(12): 9677-9696, 2024. 1 +[51] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824-24837, 2022. 2, 4 +[52] Hongxu Yin, Pavlo Molchanov, Jose M Alvarez, Zhizhong Li, Arun Mallya, Derek Hoiem, Niraj K Jha, and Jan Kautz. Dreaming to distill: Data-free knowledge transfer via deep-inversion. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (CVPR), pages 8715-8724, 2020. 1 +[53] Wenlong Yu, Ruonan Liu, Dongyue Chen, and Qinghua Hu. Explainability enhanced object detection transformer with feature disentanglement. IEEE Transactions on Image Processing, 33:6439-6454, 2024. 7 +[54] Yu Zhang, Peter Tiño, Aleš Leonardis, and Ke Tang. A survey on neural network interpretability. IEEE Transactions on Emerging Topics in Computational Intelligence, 5(5):726-742, 2021. 2 +[55] Shitian Zhao, Zhuowan Li, Yadong Lu, Alan Yuille, and Yan Wang. Causal-cog: A causal-effect look at context generation for boosting multi-modal language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13342-13351, 2024. 2 +[56] Bolei Zhou, David Bau, Aude Oliva, and Antonio Torralba. Interpreting deep visual representations via network dissection. IEEE transactions on pattern analysis and machine intelligence, 41(9):2131-2145, 2018. 2, 3, 7 \ No newline at end of file diff --git a/2025/CoE_ Chain-of-Explanation via Automatic Visual Concept Circuit Description and Polysemanticity Quantification/images.zip b/2025/CoE_ Chain-of-Explanation via Automatic Visual Concept Circuit Description and Polysemanticity Quantification/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..3d0af264dc93deff6995ae9a079993a58e198f1f --- /dev/null +++ b/2025/CoE_ Chain-of-Explanation via Automatic Visual Concept Circuit Description and Polysemanticity Quantification/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:588d180a3415e2d7c6f8e99a46b898b2532e31c95d8c587dd2ee659bf6dfd02d +size 574246 diff --git a/2025/CoE_ Chain-of-Explanation via Automatic Visual Concept Circuit Description and Polysemanticity Quantification/layout.json b/2025/CoE_ Chain-of-Explanation via Automatic Visual Concept Circuit Description and Polysemanticity Quantification/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..0ca0d54b7f4f3bb5db54b892f5365e74ee86a8d3 --- /dev/null +++ b/2025/CoE_ Chain-of-Explanation via Automatic Visual Concept Circuit Description and Polysemanticity Quantification/layout.json @@ -0,0 +1,9021 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 66, + 102, + 545, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 102, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 66, + 102, + 545, + 140 + ], + "type": "text", + "content": "CoE: Chain-of-Explanation via Automatic Visual Concept Circuit Description and Polysemanticity Quantification" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 69, + 161, + 541, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 161, + 541, + 204 + ], + "spans": [ + { + "bbox": [ + 69, + 161, + 541, + 204 + ], + "type": "text", + "content": "Wenlong Yu Qilong Wang* Chuang Liu Dong Li Qinghua Hu Tianjin Key Lab of Machine Learning, College of Intelligence and Computing, Tianjin University {ywl_95, qlwang, liuc_09, li_dong, huqinghua}@tju.edu.cn" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "spans": [ + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 256, + 297, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 256, + 297, + 579 + ], + "spans": [ + { + "bbox": [ + 55, + 256, + 297, + 579 + ], + "type": "text", + "content": "Explainability is a critical factor influencing the wide deployment of deep vision models (DVMs). Concept-based post-hoc explanation methods can provide both global and local insights into model decisions. However, current methods in this field face challenges in that they are inflexible to automatically construct accurate and sufficient linguistic explanations for global concepts and local circuits. Particularly, the intrinsic polysemanticity in semantic Visual Concepts (VCs) impedes the interpretability of concepts and DVMs, which is underestimated severely. In this paper, we propose a Chain-of-Explanation (CoE) approach to address these issues. Specifically, CoE automates the decoding and description of VCs to construct global concept explanation datasets. Further, to alleviate the effect of polysemanticity on model explainability, we design a concept polysemanticity disentanglement and filtering mechanism to distinguish the most contextually relevant concept atoms. Besides, a Concept Polysemanticity Entropy (CPE), as a measure of model interpretability, is formulated to quantify the degree of concept uncertainty. The modeling of deterministic concepts is upgraded to uncertain concept atom distributions. Finally, CoE automatically enables linguistic local explanations of the decision-making process of DVMs by tracing the concept circuit. GPT-4o and human-based experiments demonstrate the effectiveness of CPE and the superiority of CoE, achieving an average absolute improvement of " + }, + { + "bbox": [ + 55, + 256, + 297, + 579 + ], + "type": "inline_equation", + "content": "36\\%" + }, + { + "bbox": [ + 55, + 256, + 297, + 579 + ], + "type": "text", + "content": " in terms of explainability scores." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 602, + 135, + 615 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 602, + 135, + 615 + ], + "spans": [ + { + "bbox": [ + 56, + 602, + 135, + 615 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 622, + 296, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 622, + 296, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 622, + 296, + 696 + ], + "type": "text", + "content": "Deep learning-based vision models have demonstrated outstanding performance in various visual tasks, owing to the powerful learning and representation capabilities of deep neural networks (DNNs) [29, 50]. However, the extreme complexity of DNNs, while enhancing their performance, also limits their explainability [41]. Humans are often un" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 321, + 230, + 548, + 371 + ], + "blocks": [ + { + "bbox": [ + 321, + 230, + 548, + 371 + ], + "lines": [ + { + "bbox": [ + 321, + 230, + 548, + 371 + ], + "spans": [ + { + "bbox": [ + 321, + 230, + 548, + 371 + ], + "type": "image", + "image_path": "aa96329ea76bccd11daf890f0a4b579f749490e38481e2465b4b91d59f82d235.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 380, + 555, + 403 + ], + "lines": [ + { + "bbox": [ + 313, + 380, + 555, + 403 + ], + "spans": [ + { + "bbox": [ + 313, + 380, + 555, + 403 + ], + "type": "text", + "content": "Figure 1. Illustration of the integrated process of concept-based post-hoc XAI methods [36], along with two challenges they face." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 413, + 555, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 413, + 555, + 461 + ], + "spans": [ + { + "bbox": [ + 313, + 413, + 555, + 461 + ], + "type": "text", + "content": "aware of the reasons behind the model's conclusions. This lack of explainability significantly hinders the wide application of deep vision models (DVMs) in critical fields such as autonomous driving and medical diagnosis [7, 28]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 461, + 556, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 461, + 556, + 676 + ], + "spans": [ + { + "bbox": [ + 313, + 461, + 556, + 676 + ], + "type": "text", + "content": "Post-hoc eXplainable Artificial Intelligence (XAI) can enhance the transparency of DVMs without causing degradations in performance [34, 43]. Research in this field can roughly be divided into local and global XAI methods [6]. The former explains either specific stages or the entire decision path with respect to a specific sample, identifying key activated regions or significant semantic concepts [3, 45, 46]. The latter decipher the overall behavior of the DVM in a dataset by concept discovery or visualization techniques [22, 27, 52]. Among them, concept-based methods can explain DVMs from both global and local (glocal) perspectives [2, 4], assuming that each channel or neuron in a certain layer is responsible for recognizing essential semantic concepts. As shown in Fig. 1, they primarily perform one or two of the three operations: decoding Visual Concepts (VCs) of DVMs, manually constructing global concept explanation datasets, and providing local concept circuit explanations for individual samples [14, 16, 26]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "text", + "content": "Despite the significant progress made in concept-based explanation research, certain challenges remain, as illustrated in Fig. 1. First, the VCs are decoded by manually" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 703, + 145, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 145, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 145, + 712 + ], + "type": "text", + "content": "*Corresponding author." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4364" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 57, + 72, + 294, + 358 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 72, + 294, + 358 + ], + "spans": [ + { + "bbox": [ + 57, + 72, + 294, + 358 + ], + "type": "text", + "content": "identifying and annotating the commonalities of a set of extracted image patches [2, 26]. These manual methods intuitively become ineffective when faced with new scenarios, new models, and new XAI methods, due to high labor costs, low efficiency, and the limited cognitive scope of individual annotators. Second, most methods make a naive assumption: each concept is monosemantic and user-friendly [33]. However, in practice, VCs encapsulate polysemantic information [23] (e.g., an average of 13 concept atoms are coupled per channel in a ResNet152 model), presenting challenges for the model explanation. Particularly, polysemanticity can serve as a metric for evaluating a model's interpretability. A higher level of polysemanticity signifies a broader conceptual space, increasing the risks of selecting incorrect semantics and generating deviated explanations. Arbitrary concept selection or presenting them solely as images without descriptions yields suboptimal explanations [43]. Both visualizations and linguistic explanations become challenging for humans to comprehend, as it is unclear which semantic is actually represented or holds prominence [16]. Although polysemanticity is crucial to XAI, its systematic quantification and resolution remain unaddressed. All of these challenges collectively exacerbate the difficulty of concept-based explanations for DVMs." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 367, + 294, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 367, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 367, + 294, + 712 + ], + "type": "text", + "content": "To alleviate this dilemma, we propose a Chain-of-Explanation (CoE) approach via automatic concept decoding, disentanglement, filtering, description, and polysemanticity quantification. Specifically, CoE first proposes an Automatic Concept decoding and Description (ACD) method to construct global concept explanation datasets. It decodes every VC of key layers through arbitrary advanced XAI methods. A Large Vision Language Model (LVLM) is utilized to describe the commonalities of the extracted image patches since the linguistic explanation is more comprehensible than images [43]. As for the polysemantic concepts, we formulate a Concept Polysemanticity Disentanglement and Filtering (CPDF) mechanism to disentangle them into a set of orthogonal concept atoms, along with a Concept Polysemanticity Entropy (CPE) to quantify their degree of polysemanticity. In this paper, 13 distinct semantic directions are prompted, including low-level and high-level semantics, which is much more than previous work (e.g., 5 in [56]). Within the CPDF, a semantic entailment model is applied to mitigate the impact of redundant semantics. After these operations, the modeling of deterministic concepts is upgraded to uncertain concept atom distributions. When elucidating the local decision-making process, a concept circuit method is utilized to form the basis of an explanation chain, with each node representing a hierarchical set of essential concepts. Relevant atoms and their interconnections are contextually filtered and aligned within the chain. Finally, we consolidate all conceptual information along the explanation chain to generate linguistic explanations similar" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 316, + 72, + 553, + 132 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 132 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 132 + ], + "type": "text", + "content": "to a Chain-of-Thought (CoT) pipeline [21, 51], leveraging a Large Language Model (LLM), as shown in Fig. 2. Qualitative and quantitative experiments validate the effectiveness and the superiority of the CoE approach. The contributions of this paper are summarized as follows." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 313, + 133, + 553, + 336 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 314, + 133, + 553, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 133, + 553, + 180 + ], + "spans": [ + { + "bbox": [ + 314, + 133, + 553, + 180 + ], + "type": "text", + "content": "- We propose a novel CoE approach to automatically and systematically provide concept-based glocal explanations, along with managing the problem of polysemantics, in a more interpretable natural language format." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 181, + 553, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 181, + 553, + 287 + ], + "spans": [ + { + "bbox": [ + 313, + 181, + 553, + 287 + ], + "type": "text", + "content": "- A CPDF mechanism is formulated to handle the hard-to-interpret polysemantic concepts. To our knowledge, CPDF is a pioneer tailored to comprehensively disentangle and quantify polysemanticity. It disentangles them into a set of succinct concept atoms, followed by a filtering function to contextually find the suitable atom as the explanation node. A CPE score is defined to quantify the polysemanticity of concepts, which can be utilized as a metric of the interpretability of concepts or models." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 289, + 553, + 336 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 289, + 553, + 336 + ], + "spans": [ + { + "bbox": [ + 313, + 289, + 553, + 336 + ], + "type": "text", + "content": "- Both GPT-4o and human-based experiments validate the superiority of the CoE in automatically constructing global and local explanations and the effectiveness of CPE in quantifying concept polysemanticity." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 316, + 346, + 400, + 358 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 346, + 400, + 358 + ], + "spans": [ + { + "bbox": [ + 316, + 346, + 400, + 358 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 316, + 367, + 553, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 367, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 367, + 553, + 689 + ], + "type": "text", + "content": "Research on XAI can be broadly categorized into ad-hoc model interpretability and post-hoc explanation studies [43, 54]. The former actively refine DVMs to improve interpretability while also necessitating consideration of a trade-off between performance and interpretability [8, 31, 37, 42, 55]. The latter explains DVMs passively without performance decline [5, 11, 22, 45, 46]. Among these post-hoc efforts, concept-based explanations can provide glacial insights into DVMs [36]. Literature [56] annotated every pixel in special datasets with only 5 kinds of concepts. The explanation is derived by analyzing the activation regions of each neuron. MILAN [26] decoded VCs by calculating the maximum mutual information between channels and regions, followed by manually summarizing commonalities into 3 sentences. However, defining concepts solely through activations is insufficient. CRP decoded VCs through the relevance value derived from the deep Taylor decomposition [35], showing more accurate results than the activation-based methods [2]. The commonalities of the VCs were also identified manually, and one of these commonalities was selected randomly. However, these methods encounter significant challenges, as discussed in Sec. 1. In contrast, this paper proposes an automated concept decoding and description method based on LVLM, which offers greater flexibility in providing linguistic explanations. The most suitable concepts for interpretation are contextually selected, showing superior results." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 316, + 689, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 689, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 689, + 553, + 713 + ], + "type": "text", + "content": "Circuits, as sub-graphs, are proposed to explain the key decision route of a model [3, 13, 17, 20, 48]. They identify" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "4365" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 69, + 550, + 322 + ], + "blocks": [ + { + "bbox": [ + 59, + 69, + 550, + 322 + ], + "lines": [ + { + "bbox": [ + 59, + 69, + 550, + 322 + ], + "spans": [ + { + "bbox": [ + 59, + 69, + 550, + 322 + ], + "type": "image", + "image_path": "82d54784ccc7eb026d1a531358c99154558cbe048c177f8a8e964c9a6e8e3789.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 329, + 555, + 352 + ], + "lines": [ + { + "bbox": [ + 55, + 329, + 555, + 352 + ], + "spans": [ + { + "bbox": [ + 55, + 329, + 555, + 352 + ], + "type": "text", + "content": "Figure 2. Outline of the Chain-of-Explanation. It consists of three steps (i.e., ACD, CPDF, and local explanation). Each part utilizes a set of channels " + }, + { + "bbox": [ + 55, + 329, + 555, + 352 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 55, + 329, + 555, + 352 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 55, + 329, + 555, + 352 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 55, + 329, + 555, + 352 + ], + "type": "text", + "content": ". In the concept circuit, " + }, + { + "bbox": [ + 55, + 329, + 555, + 352 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 55, + 329, + 555, + 352 + ], + "type": "text", + "content": ", lines of different depths and colors represent the magnitude of the relevance." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 362, + 297, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 362, + 297, + 567 + ], + "spans": [ + { + "bbox": [ + 54, + 362, + 297, + 567 + ], + "type": "text", + "content": "the highest-activated concepts and then explain the model's decision by linking all concepts across the circuit. However, its interpretability for DVMs is limited due to the lack of a cohesive, language-based explanatory output. Additionally, the polysemanticity of VCs significantly hinders the accurate explanation [18, 33], and research in this area remains limited. Some methods attempted to indirectly mitigate its impact by predefining principal vectors or subspaces [20, 27, 38]. The lack of flexible and accurate concept-level explanations also restricts their adaptability. Literature [33] analyzed the polysemanticity from an information compression perspective. Pure [16] alleviated the polysemanticity by clustering and decoupling the circuit route. These methods, however, lack the ability to simultaneously and comprehensively quantify, disentangle, and describe polysemantic VCs. Their applicability is limited—challenges that this paper seeks to address, as discussed in Sec. 3." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 583, + 237, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 583, + 237, + 597 + ], + "spans": [ + { + "bbox": [ + 55, + 583, + 237, + 597 + ], + "type": "text", + "content": "3. Chain-of-Explanation Approach" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "text", + "content": "In this paper, a CoE approach is proposed to automatically explain DVMs from both global and local perspectives. It breaks down the complex reasoning path into a chain of concept nodes, within which polysemanticity is effectively managed. As shown in Fig. 2, CoE is composed of three steps. Step 1: ACD is proposed to automatically construct global concept descriptive explanation datasets of a given DVM. Step 2: With the assistance of an LVLM, a CPDF mechanism is designed to disentangle each original VC into" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 362, + 555, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 362, + 555, + 447 + ], + "spans": [ + { + "bbox": [ + 313, + 362, + 555, + 447 + ], + "type": "text", + "content": "a set of concept atoms and extract the most suitable atom according to the context, making CoE more comprehensible and flexible. Besides, a CPE score is formulated accordingly to quantify the degree of polysemanticity of concepts. Step 3: A local explanation chain in a manner akin to a CoT for the decision-making progress of the DVM is established based on the dataset, CPDF, concept circuit, and LLMs." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 453, + 552, + 466 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 453, + 552, + 466 + ], + "spans": [ + { + "bbox": [ + 313, + 453, + 552, + 466 + ], + "type": "text", + "content": "3.1. Automatic Concept Decoding and Description" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 471, + 555, + 614 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 471, + 555, + 614 + ], + "spans": [ + { + "bbox": [ + 313, + 471, + 555, + 614 + ], + "type": "text", + "content": "ACD method is first formulated to automatically and effectively construct databases of global concept descriptions. Specifically, given an image dataset " + }, + { + "bbox": [ + 313, + 471, + 555, + 614 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 313, + 471, + 555, + 614 + ], + "type": "text", + "content": " and a DVM " + }, + { + "bbox": [ + 313, + 471, + 555, + 614 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 313, + 471, + 555, + 614 + ], + "type": "text", + "content": " requiring evaluation and explanation, we establish an injective function " + }, + { + "bbox": [ + 313, + 471, + 555, + 614 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 313, + 471, + 555, + 614 + ], + "type": "text", + "content": " between channel set " + }, + { + "bbox": [ + 313, + 471, + 555, + 614 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 313, + 471, + 555, + 614 + ], + "type": "text", + "content": " and visual semantic concept set " + }, + { + "bbox": [ + 313, + 471, + 555, + 614 + ], + "type": "inline_equation", + "content": "\\mathcal{S}" + }, + { + "bbox": [ + 313, + 471, + 555, + 614 + ], + "type": "text", + "content": " by an XAI method " + }, + { + "bbox": [ + 313, + 471, + 555, + 614 + ], + "type": "inline_equation", + "content": "E_{x}, f: \\mathcal{C} \\mapsto \\mathcal{S}" + }, + { + "bbox": [ + 313, + 471, + 555, + 614 + ], + "type": "text", + "content": ". Without loss of generality, for " + }, + { + "bbox": [ + 313, + 471, + 555, + 614 + ], + "type": "inline_equation", + "content": "j_{th}" + }, + { + "bbox": [ + 313, + 471, + 555, + 614 + ], + "type": "text", + "content": " channel " + }, + { + "bbox": [ + 313, + 471, + 555, + 614 + ], + "type": "inline_equation", + "content": "c_{j}^{l}" + }, + { + "bbox": [ + 313, + 471, + 555, + 614 + ], + "type": "text", + "content": " in layer " + }, + { + "bbox": [ + 313, + 471, + 555, + 614 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 313, + 471, + 555, + 614 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 471, + 555, + 614 + ], + "type": "inline_equation", + "content": "E_{x}" + }, + { + "bbox": [ + 313, + 471, + 555, + 614 + ], + "type": "text", + "content": " extracts a set of image patches within " + }, + { + "bbox": [ + 313, + 471, + 555, + 614 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 313, + 471, + 555, + 614 + ], + "type": "text", + "content": " that most activate this channel, named VC " + }, + { + "bbox": [ + 313, + 471, + 555, + 614 + ], + "type": "inline_equation", + "content": "s_{j}^{l} = \\{I_{n}\\}_{N}" + }, + { + "bbox": [ + 313, + 471, + 555, + 614 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 471, + 555, + 614 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 471, + 555, + 614 + ], + "type": "text", + "content": " is the number of image patches. Each " + }, + { + "bbox": [ + 313, + 471, + 555, + 614 + ], + "type": "inline_equation", + "content": "I_{n}" + }, + { + "bbox": [ + 313, + 471, + 555, + 614 + ], + "type": "text", + "content": " highlights certain regions of the original image, representing the focused concept, as exemplified in the bottom left part of Fig. 2." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 614, + 555, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 614, + 555, + 640 + ], + "spans": [ + { + "bbox": [ + 314, + 614, + 555, + 640 + ], + "type": "text", + "content": "A describing function " + }, + { + "bbox": [ + 314, + 614, + 555, + 640 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 314, + 614, + 555, + 640 + ], + "type": "text", + "content": " is then defined to decipher " + }, + { + "bbox": [ + 314, + 614, + 555, + 640 + ], + "type": "inline_equation", + "content": "\\{I_n\\}_N" + }, + { + "bbox": [ + 314, + 614, + 555, + 640 + ], + "type": "text", + "content": " into a set of language descriptions:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 376, + 645, + 555, + 659 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 376, + 645, + 555, + 659 + ], + "spans": [ + { + "bbox": [ + 376, + 645, + 555, + 659 + ], + "type": "interline_equation", + "content": "\\mathcal {D} = r \\left(\\left\\{I _ {n} \\right\\} _ {N}\\right) = \\left\\{d _ {m} \\right\\} _ {M}, \\tag {1}", + "image_path": "69456c42eb04feceeb75de630c11c0804eee1cb8252a24dbda17c20d2fc99dd2.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 666, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 666, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 666, + 556, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 666, + 556, + 715 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 666, + 556, + 715 + ], + "type": "text", + "content": " is the number of language descriptions for the current channel. Traditionally, " + }, + { + "bbox": [ + 313, + 666, + 556, + 715 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 313, + 666, + 556, + 715 + ], + "type": "text", + "content": " is implemented by manually and subjectively identifying commonalities within " + }, + { + "bbox": [ + 313, + 666, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\{I_n\\}_N" + }, + { + "bbox": [ + 313, + 666, + 556, + 715 + ], + "type": "text", + "content": " [56], while ACD utilizes a powerful LVLM as the descriptor" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4366" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 122 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 122 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 122 + ], + "type": "text", + "content": "with well-engineered prompts. After constructing " + }, + { + "bbox": [ + 55, + 72, + 296, + 122 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 55, + 72, + 296, + 122 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 72, + 296, + 122 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 55, + 72, + 296, + 122 + ], + "type": "text", + "content": " and the mapping function " + }, + { + "bbox": [ + 55, + 72, + 296, + 122 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 55, + 72, + 296, + 122 + ], + "type": "text", + "content": ", a concept description dataset " + }, + { + "bbox": [ + 55, + 72, + 296, + 122 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_{\\mathcal{M},\\mathcal{T}} = \\{\\mathcal{C}, S, \\mathcal{D}, f\\}" + }, + { + "bbox": [ + 55, + 72, + 296, + 122 + ], + "type": "text", + "content": " is automatically constructed, providing global linguistic explanations for all VCs of " + }, + { + "bbox": [ + 55, + 72, + 296, + 122 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 55, + 72, + 296, + 122 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 126, + 234, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 126, + 234, + 140 + ], + "spans": [ + { + "bbox": [ + 55, + 126, + 234, + 140 + ], + "type": "text", + "content": "3.2. Managing Polysemantic Concept" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 145, + 296, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 145, + 296, + 289 + ], + "spans": [ + { + "bbox": [ + 55, + 145, + 296, + 289 + ], + "type": "text", + "content": "VCs of each node within a DVM exhibit polysemantics, introducing certain deviations in the explanation framework, as shown in Fig. 2. Higher polysemantics indicate that the current VC aggregates more semantics, making the intended representation more ambiguous. Such semantic uncertainty impedes human comprehension of the concept, particularly within the CoE, where various concepts must be understood collectively. Quantifying and solving this dilemma is difficult and has been underestimated by previous studies. Under this circumstance, we propose a CPDF mechanism along with a CPE score to mitigate and quantify its detriments to the CoE simultaneously." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 303, + 293, + 316 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 303, + 293, + 316 + ], + "spans": [ + { + "bbox": [ + 55, + 303, + 293, + 316 + ], + "type": "text", + "content": "3.2.1 Polysemantic Concept Decoupling and Filtering" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 322, + 296, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 322, + 296, + 370 + ], + "spans": [ + { + "bbox": [ + 55, + 322, + 296, + 370 + ], + "type": "text", + "content": "Considering that a polysemous word can be understood as an aggregation of various distinct meanings, we define a function " + }, + { + "bbox": [ + 55, + 322, + 296, + 370 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 55, + 322, + 296, + 370 + ], + "type": "text", + "content": " to disentangle the ambiguous VC into a set of easily understandable linguistic concept atoms:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 101, + 375, + 295, + 390 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 101, + 375, + 295, + 390 + ], + "spans": [ + { + "bbox": [ + 101, + 375, + 295, + 390 + ], + "type": "interline_equation", + "content": "\\mathcal {A} = \\left\\{a ^ {1}, \\dots , a ^ {i}, \\dots , a ^ {Q} \\right\\} = g (s), \\tag {2}", + "image_path": "b40a3c56dbb80127b5b6e21c03eaa6bca11b376c334b8ddc6e5191dff1760e71.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 396, + 296, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 396, + 296, + 506 + ], + "spans": [ + { + "bbox": [ + 55, + 396, + 296, + 506 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 396, + 296, + 506 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 55, + 396, + 296, + 506 + ], + "type": "text", + "content": " is the number of the concept atoms. We posit that each atom represents a view of the current VC " + }, + { + "bbox": [ + 55, + 396, + 296, + 506 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 55, + 396, + 296, + 506 + ], + "type": "text", + "content": ", which will be activated upon the appearance of the corresponding semantic in the input. Furthermore, given that concept atoms should accurately represent the commonalities in " + }, + { + "bbox": [ + 55, + 396, + 296, + 506 + ], + "type": "inline_equation", + "content": "\\{I_n\\}_N" + }, + { + "bbox": [ + 55, + 396, + 296, + 506 + ], + "type": "text", + "content": " and recognizing that this naive disentanglement function (Eq. 2) does not provide a foundation for quantifying polysemanticity, we refine the function " + }, + { + "bbox": [ + 55, + 396, + 296, + 506 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 55, + 396, + 296, + 506 + ], + "type": "text", + "content": " to disentangle each image patch " + }, + { + "bbox": [ + 55, + 396, + 296, + 506 + ], + "type": "inline_equation", + "content": "I_n" + }, + { + "bbox": [ + 55, + 396, + 296, + 506 + ], + "type": "text", + "content": " into a set of atoms:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 89, + 510, + 295, + 540 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 510, + 295, + 540 + ], + "spans": [ + { + "bbox": [ + 89, + 510, + 295, + 540 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {A} _ {n} = \\left\\{a _ {n} ^ {1}, \\dots , a _ {n} ^ {i}, \\dots , a _ {n} ^ {Q} \\right\\} = g (I _ {n}, s), \\tag {3} \\\\ \\mathcal {D} = \\{\\mathcal {A} _ {n} \\} _ {N}. \\\\ \\end{array}", + "image_path": "8511bf51426cdfb44b6819af2353d23e7aa93bb9e37e69ee8333e4aff7ddd60a.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 546, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 546, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 546, + 296, + 713 + ], + "type": "text", + "content": "We instantiate " + }, + { + "bbox": [ + 55, + 546, + 296, + 713 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 55, + 546, + 296, + 713 + ], + "type": "text", + "content": " by a powerful LVLM (e.g., GPT-4o [39]), considering that this is quite an open and complex problem. " + }, + { + "bbox": [ + 55, + 546, + 296, + 713 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 55, + 546, + 296, + 713 + ], + "type": "text", + "content": " prioritizes disentangling and describing commonalities within the VC while analyzing each individual image. Various commonalities among any certain number of image patches should be summarized and described by precise terminologies. These common semantics within an open scenario are as diverse as they are abundant, unfolding in countless forms and nuances. Besides, it should generate specific atoms when any commonalities identified above do not appear in the current " + }, + { + "bbox": [ + 55, + 546, + 296, + 713 + ], + "type": "inline_equation", + "content": "I_{n}" + }, + { + "bbox": [ + 55, + 546, + 296, + 713 + ], + "type": "text", + "content": ". Therefore, the art of prompt engineering prompt-com becomes paramount. It should provide rules for recognizing commonalities, possible conceptual directions, and output format control. In" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "content": "this paper, we prompt 13 high-level and low-level semantical directions that cover the normal concepts encountered in daily life, including object category, scene, object part, color, texture, material, position, transparency, brightness, shape, size, edges, and their relationships. Additionally, the CoT techniques [21, 51] are introduced. Details of all sophisticated prompts are shown in Sec. S2 of the Appendix." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 157, + 553, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 157, + 553, + 179 + ], + "spans": [ + { + "bbox": [ + 313, + 157, + 553, + 179 + ], + "type": "text", + "content": "All " + }, + { + "bbox": [ + 313, + 157, + 553, + 179 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_n" + }, + { + "bbox": [ + 313, + 157, + 553, + 179 + ], + "type": "text", + "content": " are then aggregated into a description atom set " + }, + { + "bbox": [ + 313, + 157, + 553, + 179 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 313, + 157, + 553, + 179 + ], + "type": "text", + "content": " with all duplicate atoms removed:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 370, + 186, + 553, + 216 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 186, + 553, + 216 + ], + "spans": [ + { + "bbox": [ + 370, + 186, + 553, + 216 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {A} = \\left\\{\\mathcal {A} _ {1} \\cup \\mathcal {A} _ {2} \\cup \\dots \\cup \\mathcal {A} _ {N} \\right\\} \\tag {4} \\\\ = \\left\\{a ^ {1}, \\dots , a ^ {i}, \\dots , a ^ {P} \\right\\}, \\\\ \\end{array}", + "image_path": "cbd594ac66e763dcf6944e93b88a3414f9ada8b350403786c1c483889a9c21a5.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 220, + 553, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 220, + 553, + 304 + ], + "spans": [ + { + "bbox": [ + 313, + 220, + 553, + 304 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 220, + 553, + 304 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 313, + 220, + 553, + 304 + ], + "type": "text", + "content": " is the number of concept atoms after deleting the repeating atoms. Each element in " + }, + { + "bbox": [ + 313, + 220, + 553, + 304 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 313, + 220, + 553, + 304 + ], + "type": "text", + "content": " represents a disentangled monosemantic concept and can be utilized as a node of the CoE. However, some atoms are semantically equivalent even if they have variant descriptions [19]. In order to remove semantically redundant atoms, we introduce an entailment function " + }, + { + "bbox": [ + 313, + 220, + 553, + 304 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 313, + 220, + 553, + 304 + ], + "type": "text", + "content": " to detect the relations between atoms:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 351, + 308, + 553, + 323 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 308, + 553, + 323 + ], + "spans": [ + { + "bbox": [ + 351, + 308, + 553, + 323 + ], + "type": "interline_equation", + "content": "\\{1, 0, - 1 \\} = \\delta \\left(a ^ {i 1}, a ^ {i 2}\\right), \\forall a ^ {i 1}, a ^ {i 2} \\in \\mathcal {A}, \\tag {5}", + "image_path": "5647130b2a8052dbdd2b92dc8172167d326a24b439a3e18a0a194c1de6e40844.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 328, + 554, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 328, + 554, + 460 + ], + "spans": [ + { + "bbox": [ + 313, + 328, + 554, + 460 + ], + "type": "text", + "content": "where the responses 1,0,-1 represent entailment, neutral, or contradiction of the candidates, respectively. We instantiate " + }, + { + "bbox": [ + 313, + 328, + 554, + 460 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 313, + 328, + 554, + 460 + ], + "type": "text", + "content": " by an NLI model (e.g., DeBERTa model [25]) because the atom candidates are relatively short and simple that an NLI model is good enough. Following [19], a bidirectional entailment check is applied, meaning that one of the candidates will be removed when the response is entailment or neutral. The first atom is designated as the monosemantic representative of the semantic equivalence set due to the transitivity between them. The final concept set " + }, + { + "bbox": [ + 313, + 328, + 554, + 460 + ], + "type": "inline_equation", + "content": "A^{*}" + }, + { + "bbox": [ + 313, + 328, + 554, + 460 + ], + "type": "text", + "content": " only contains heterogeneous atoms that describe the current VC:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 369, + 464, + 553, + 485 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 369, + 464, + 553, + 485 + ], + "spans": [ + { + "bbox": [ + 369, + 464, + 553, + 485 + ], + "type": "interline_equation", + "content": "\\mathcal {A} ^ {*} = \\left\\{a ^ {1}, \\dots , a ^ {i}, \\dots , a ^ {P ^ {*}} \\right\\}. \\tag {6}", + "image_path": "9660d88921f750f17553a0ca4076a0d05b2c96f553a8ed71bad07eb358674045.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 489, + 553, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 489, + 553, + 525 + ], + "spans": [ + { + "bbox": [ + 313, + 489, + 553, + 525 + ], + "type": "text", + "content": "After disentangling the concept atoms, we define a filtering function " + }, + { + "bbox": [ + 313, + 489, + 553, + 525 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 313, + 489, + 553, + 525 + ], + "type": "text", + "content": " to select an atom " + }, + { + "bbox": [ + 313, + 489, + 553, + 525 + ], + "type": "inline_equation", + "content": "a^*" + }, + { + "bbox": [ + 313, + 489, + 553, + 525 + ], + "type": "text", + "content": " as the most suitable explanation of current VC with respect to the input " + }, + { + "bbox": [ + 313, + 489, + 553, + 525 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 489, + 553, + 525 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 387, + 530, + 553, + 544 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 387, + 530, + 553, + 544 + ], + "spans": [ + { + "bbox": [ + 387, + 530, + 553, + 544 + ], + "type": "interline_equation", + "content": "a _ {s} ^ {*} (x) = z \\left(\\mathcal {A} _ {s} ^ {*}, C (x)\\right). \\tag {7}", + "image_path": "47cc02555941dc340c3b8214286b9a6a02b84472d43dd6837711fadde07459da.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 549, + 554, + 645 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 549, + 554, + 645 + ], + "spans": [ + { + "bbox": [ + 313, + 549, + 554, + 645 + ], + "type": "text", + "content": "Some research ignores this step or selects the concept randomly with an implicit assumption that all concepts in " + }, + { + "bbox": [ + 313, + 549, + 554, + 645 + ], + "type": "inline_equation", + "content": "\\mathcal{A}^*" + }, + { + "bbox": [ + 313, + 549, + 554, + 645 + ], + "type": "text", + "content": " are semantically equivalent [14, 26]. In contrast, we filter the most suitable atom " + }, + { + "bbox": [ + 313, + 549, + 554, + 645 + ], + "type": "inline_equation", + "content": "a^*" + }, + { + "bbox": [ + 313, + 549, + 554, + 645 + ], + "type": "text", + "content": " based on the context of the CoE. In particular, " + }, + { + "bbox": [ + 313, + 549, + 554, + 645 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 313, + 549, + 554, + 645 + ], + "type": "text", + "content": " is instantiated by an LLM. " + }, + { + "bbox": [ + 313, + 549, + 554, + 645 + ], + "type": "inline_equation", + "content": "C(x)" + }, + { + "bbox": [ + 313, + 549, + 554, + 645 + ], + "type": "text", + "content": " is an image captioning function utilized for providing contextual information. It can be instantiated by a relatively small LVLM [9, 10] or a professional model [30, 49]." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 658, + 528, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 658, + 528, + 672 + ], + "spans": [ + { + "bbox": [ + 313, + 658, + 528, + 672 + ], + "type": "text", + "content": "3.2.2 Quantification of Concept Polysemanticity" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "content": "The polysemanticity of concepts can serve as a metric for the model's interpretability. Lower polysemanticity of concepts indicates reduced semantic uncertainty with a more" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4367" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "content": "singular form of activation, leading to greater interpretability. In this paper, we endeavor to quantify polysemanticity from three levels of granularity (i.e., conceptual channels, layers, and models) by defining a CPE score. The larger the CPE, the more pronounced the polysemanticity, which leads to decreased interpretability." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 144, + 296, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 144, + 296, + 192 + ], + "spans": [ + { + "bbox": [ + 55, + 144, + 296, + 192 + ], + "type": "text", + "content": "In particular, the disentanglement function " + }, + { + "bbox": [ + 55, + 144, + 296, + 192 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 55, + 144, + 296, + 192 + ], + "type": "text", + "content": " in Eq. 3 enables us to count the frequency of each atom in " + }, + { + "bbox": [ + 55, + 144, + 296, + 192 + ], + "type": "inline_equation", + "content": "\\mathcal{A}^*" + }, + { + "bbox": [ + 55, + 144, + 296, + 192 + ], + "type": "text", + "content": ". Given two fixed numbers " + }, + { + "bbox": [ + 55, + 144, + 296, + 192 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 55, + 144, + 296, + 192 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 144, + 296, + 192 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 144, + 296, + 192 + ], + "type": "text", + "content": ", we calculate the probability of the " + }, + { + "bbox": [ + 55, + 144, + 296, + 192 + ], + "type": "inline_equation", + "content": "i_{th}" + }, + { + "bbox": [ + 55, + 144, + 296, + 192 + ], + "type": "text", + "content": " atom within " + }, + { + "bbox": [ + 55, + 144, + 296, + 192 + ], + "type": "inline_equation", + "content": "j_{th}" + }, + { + "bbox": [ + 55, + 144, + 296, + 192 + ], + "type": "text", + "content": " concept:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 130, + 196, + 295, + 222 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 196, + 295, + 222 + ], + "spans": [ + { + "bbox": [ + 130, + 196, + 295, + 222 + ], + "type": "interline_equation", + "content": "p _ {i} = \\frac {\\text {N u m} _ {i}}{Q \\times N + P a d}. \\tag {8}", + "image_path": "74b66b2d79e94370cccb6ebf2e5c430c3b3b8a496b8b951fe7b35932b1241d8e.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 225, + 274, + 237 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 225, + 274, + 237 + ], + "spans": [ + { + "bbox": [ + 67, + 225, + 274, + 237 + ], + "type": "text", + "content": "Then, the CPE of " + }, + { + "bbox": [ + 67, + 225, + 274, + 237 + ], + "type": "inline_equation", + "content": "j_{th}" + }, + { + "bbox": [ + 67, + 225, + 274, + 237 + ], + "type": "text", + "content": " concept can be formulated as" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 115, + 242, + 295, + 271 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 242, + 295, + 271 + ], + "spans": [ + { + "bbox": [ + 115, + 242, + 295, + 271 + ], + "type": "interline_equation", + "content": "H _ {j} ^ {*} = \\frac {- \\sum_ {i = 1} ^ {P ^ {*} + P a d} p _ {i} \\log p _ {i}}{\\log (P ^ {*} + P a d)}. \\tag {9}", + "image_path": "9b311d57951ff3d04d72a28d69553182e762648e999a14a0474033b440afac76.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 275, + 296, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 275, + 296, + 430 + ], + "spans": [ + { + "bbox": [ + 55, + 275, + 296, + 430 + ], + "type": "text", + "content": "The denominator normalizes the entropy value to a range between 0 and 1, considering that the size of " + }, + { + "bbox": [ + 55, + 275, + 296, + 430 + ], + "type": "inline_equation", + "content": "P^*" + }, + { + "bbox": [ + 55, + 275, + 296, + 430 + ], + "type": "text", + "content": " varies for different VCs. Besides, the quantity of " + }, + { + "bbox": [ + 55, + 275, + 296, + 430 + ], + "type": "inline_equation", + "content": "A^*" + }, + { + "bbox": [ + 55, + 275, + 296, + 430 + ], + "type": "text", + "content": " is set to be at least equal to " + }, + { + "bbox": [ + 55, + 275, + 296, + 430 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 275, + 296, + 430 + ], + "type": "text", + "content": " due to some " + }, + { + "bbox": [ + 55, + 275, + 296, + 430 + ], + "type": "inline_equation", + "content": "D_s" + }, + { + "bbox": [ + 55, + 275, + 296, + 430 + ], + "type": "text", + "content": " having a small number of semantical equivalent atoms with equal probabilities. If there are insufficient unique atoms, we supplement the corresponding number of atoms, with each supplementary atom set to a number of 1. The atomic objects are expanded, and the denominator is increased by the number of supplemented atoms (a total amount of " + }, + { + "bbox": [ + 55, + 275, + 296, + 430 + ], + "type": "inline_equation", + "content": "Pad" + }, + { + "bbox": [ + 55, + 275, + 296, + 430 + ], + "type": "text", + "content": "). Please refer to Sec. S3.1 of the Appendix for details. In addition, two averaged CPEs are formulated to enable comparisons of polysemantics between specific layers or models:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 435, + 295, + 464 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 435, + 295, + 464 + ], + "spans": [ + { + "bbox": [ + 138, + 435, + 295, + 464 + ], + "type": "interline_equation", + "content": "H _ {l} ^ {*} = \\frac {1}{d _ {l}} \\sum^ {d _ {l}} H _ {j} ^ {*}, \\tag {10}", + "image_path": "ca048f2b99dc3d8eecf529cdc3db89f531d243ee097e01191bb24c54f63a9aef.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 137, + 469, + 295, + 496 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 469, + 295, + 496 + ], + "spans": [ + { + "bbox": [ + 137, + 469, + 295, + 496 + ], + "type": "interline_equation", + "content": "H _ {\\mathcal {M}} ^ {*} = \\frac {1}{L} \\sum^ {L} H _ {l} ^ {*}, \\tag {11}", + "image_path": "1bed507a4a0f9e3b260b45814f34625785011fafd00ad85d85b33532fcbd9a01.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 498, + 296, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 296, + 570 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 296, + 570 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 498, + 296, + 570 + ], + "type": "inline_equation", + "content": "d_{l}" + }, + { + "bbox": [ + 55, + 498, + 296, + 570 + ], + "type": "text", + "content": " is the dimension of the channel and " + }, + { + "bbox": [ + 55, + 498, + 296, + 570 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 55, + 498, + 296, + 570 + ], + "type": "text", + "content": " is the number of layers. Finally, to accommodate the diverse needs of researchers within the community, the automatically constructed concept explanation dataset is extended as ACD" + }, + { + "bbox": [ + 55, + 498, + 296, + 570 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_{\\mathcal{M},\\mathcal{T}} = \\{\\mathcal{C},\\mathcal{S},\\mathcal{D},\\mathcal{A}^{*},\\mathcal{Q},\\mathcal{H},f\\}" + }, + { + "bbox": [ + 55, + 498, + 296, + 570 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 498, + 296, + 570 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}" + }, + { + "bbox": [ + 55, + 498, + 296, + 570 + ], + "type": "text", + "content": " represents the probability set of each atom and " + }, + { + "bbox": [ + 55, + 498, + 296, + 570 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 55, + 498, + 296, + 570 + ], + "type": "text", + "content": " is the set of CPE." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 575, + 244, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 575, + 244, + 588 + ], + "spans": [ + { + "bbox": [ + 55, + 575, + 244, + 588 + ], + "type": "text", + "content": "3.3. Construction of Explanation Chain" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "content": "Concept circuits in DVMs can identify the most activated units in each layer for a specific input [3]. When aggregating these units across all key layers in a bottom-up manner, a corresponding maximum activation path " + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "content": " is generated, establishing the foundation for the explanation chain. Each node in this path serves as the explanation for the model's decision at that layer. Traditional concept circuit methods for DVMs identify the activation path without providing thorough overarching natural language-based explanations. Inspired by the CoT explanation technique in LLMs" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "content": "[12], CoE proposes a local explanation chain to interpret the decision-making process of DVMs in the form of natural language descriptions based on ACD-" + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_{\\mathcal{M},\\mathcal{T}}" + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "content": ", CPDF, concept circuit method, and LLM, as shown in Fig. 2." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 121, + 555, + 302 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 121, + 555, + 302 + ], + "spans": [ + { + "bbox": [ + 313, + 121, + 555, + 302 + ], + "type": "text", + "content": "Specifically, some user-predefined layers requiring explanations or key layers facilitating CoE (e.g., output layers of 4 stages of a DVM) are listed in a bottom-up manner. Given an image " + }, + { + "bbox": [ + 313, + 121, + 555, + 302 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 121, + 555, + 302 + ], + "type": "text", + "content": " and layer " + }, + { + "bbox": [ + 313, + 121, + 555, + 302 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 313, + 121, + 555, + 302 + ], + "type": "text", + "content": ", an advanced XAI method " + }, + { + "bbox": [ + 313, + 121, + 555, + 302 + ], + "type": "inline_equation", + "content": "E_{x}" + }, + { + "bbox": [ + 313, + 121, + 555, + 302 + ], + "type": "text", + "content": " computes the maximization normalized activation or relevance values for all concepts. Each value represents the importance of the current concept to the prediction " + }, + { + "bbox": [ + 313, + 121, + 555, + 302 + ], + "type": "inline_equation", + "content": "\\mathcal{M}(x)" + }, + { + "bbox": [ + 313, + 121, + 555, + 302 + ], + "type": "text", + "content": ". CoE acknowledges top " + }, + { + "bbox": [ + 313, + 121, + 555, + 302 + ], + "type": "inline_equation", + "content": "k^l" + }, + { + "bbox": [ + 313, + 121, + 555, + 302 + ], + "type": "text", + "content": " relevant concepts " + }, + { + "bbox": [ + 313, + 121, + 555, + 302 + ], + "type": "inline_equation", + "content": "\\mathcal{C}^l(x) = \\{c_1^l, \\dots, c_{k^l}^l\\}" + }, + { + "bbox": [ + 313, + 121, + 555, + 302 + ], + "type": "text", + "content": " and their values " + }, + { + "bbox": [ + 313, + 121, + 555, + 302 + ], + "type": "inline_equation", + "content": "\\mathcal{V}^l(x) = \\{v_1^l, \\dots, v_{k^l}^l\\}" + }, + { + "bbox": [ + 313, + 121, + 555, + 302 + ], + "type": "text", + "content": ", determined by the " + }, + { + "bbox": [ + 313, + 121, + 555, + 302 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 121, + 555, + 302 + ], + "type": "text", + "content": " quantile of relevance values across all concepts (i.e., " + }, + { + "bbox": [ + 313, + 121, + 555, + 302 + ], + "type": "inline_equation", + "content": "k^l = |\\{v_j^l | v_j^l > \\alpha \\max(v^l), \\forall j \\in d^l\\}|" + }, + { + "bbox": [ + 313, + 121, + 555, + 302 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 121, + 555, + 302 + ], + "type": "inline_equation", + "content": "|\\mathcal{X}|" + }, + { + "bbox": [ + 313, + 121, + 555, + 302 + ], + "type": "text", + "content": " computes the number of elements). After that, the descriptive concept atoms of the current layer " + }, + { + "bbox": [ + 313, + 121, + 555, + 302 + ], + "type": "inline_equation", + "content": "\\widetilde{\\mathcal{A}}_{k^l}^{*,l}(x) = \\{a_1^{*,l}, \\dots, a_{k^l}^{*,l}\\}" + }, + { + "bbox": [ + 313, + 121, + 555, + 302 + ], + "type": "text", + "content": " are constructed based on ACD-" + }, + { + "bbox": [ + 313, + 121, + 555, + 302 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_{\\mathcal{M},\\mathcal{T}}" + }, + { + "bbox": [ + 313, + 121, + 555, + 302 + ], + "type": "text", + "content": " and the filter Eq. 7, serving as a node of the explanation chain." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 304, + 554, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 304, + 554, + 339 + ], + "spans": [ + { + "bbox": [ + 313, + 304, + 554, + 339 + ], + "type": "text", + "content": "The final CoE for the local decision-making explanation, abstracting and synthesizing all top activated concept atoms along the chain, is formulated as" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 317, + 349, + 554, + 364 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 349, + 554, + 364 + ], + "spans": [ + { + "bbox": [ + 317, + 349, + 554, + 364 + ], + "type": "interline_equation", + "content": "C o E _ {\\mathcal {M}} (x) = e \\left(\\widetilde {\\mathcal {A}} _ {\\mathcal {P}} ^ {*} (x), \\mathcal {V} _ {\\mathcal {P}} (x), C (x), G (x), \\mathcal {M} (x)\\right), \\tag {12}", + "image_path": "b75a87f5ae8de31fdfb06c0a95b1bc4ed0dfe60e7b468e45a0dbc7da288af4d0.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 374, + 554, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 374, + 554, + 434 + ], + "spans": [ + { + "bbox": [ + 313, + 374, + 554, + 434 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 374, + 554, + 434 + ], + "type": "inline_equation", + "content": "e" + }, + { + "bbox": [ + 313, + 374, + 554, + 434 + ], + "type": "text", + "content": " formulates a synthesizing and describing function with inputs obtained from the above sections, and " + }, + { + "bbox": [ + 313, + 374, + 554, + 434 + ], + "type": "inline_equation", + "content": "G(x)" + }, + { + "bbox": [ + 313, + 374, + 554, + 434 + ], + "type": "text", + "content": " is the label of " + }, + { + "bbox": [ + 313, + 374, + 554, + 434 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 374, + 554, + 434 + ], + "type": "text", + "content": ". We instantiate " + }, + { + "bbox": [ + 313, + 374, + 554, + 434 + ], + "type": "inline_equation", + "content": "CoE_{\\mathcal{M}}(x)" + }, + { + "bbox": [ + 313, + 374, + 554, + 434 + ], + "type": "text", + "content": " through a powerful LLM (e.g., GPT-4 [1]), with prompt prompt-coe well-designed by combining CoT and few-shot prompting [44]." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 435, + 554, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 435, + 554, + 471 + ], + "spans": [ + { + "bbox": [ + 313, + 435, + 554, + 471 + ], + "type": "text", + "content": "Additionally, these local explanations are evaluated by a powerful LVLM and human judges based on three explainability metrics, each utilizing a three-tier scoring system." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 484, + 395, + 498 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 484, + 395, + 498 + ], + "spans": [ + { + "bbox": [ + 313, + 484, + 395, + 498 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 505, + 554, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 505, + 554, + 601 + ], + "spans": [ + { + "bbox": [ + 313, + 505, + 554, + 601 + ], + "type": "text", + "content": "We evaluate the proposed CoE approach on two types of DVMs (i.e., ResNet [24] and CLIP [40]). The relevance-based XAI method CRP is adopted as the primary approach for " + }, + { + "bbox": [ + 313, + 505, + 554, + 601 + ], + "type": "inline_equation", + "content": "E_{x}" + }, + { + "bbox": [ + 313, + 505, + 554, + 601 + ], + "type": "text", + "content": " in this paper, owing to its demonstrated superiority in terms of fidelity and reliability [2]. Please refer to the Appendix for more details on the implementation and results. Codes are available at https://github.com/YuWLong666/CoE." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 610, + 538, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 610, + 538, + 624 + ], + "spans": [ + { + "bbox": [ + 313, + 610, + 538, + 624 + ], + "type": "text", + "content": "4.1. Concept Explanation Dataset Construction" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "text", + "content": "In the CoE, we first automatically construct global concept description databases " + }, + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\mathsf{ACD - B}_{\\mathcal{M},\\mathcal{T}}" + }, + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "text", + "content": " for various DVMs. Taking a ResNet152 model as an example, 4 output layers of 4 stages (with 256, 512, 1024, and 2048 channels " + }, + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "text", + "content": ", respectively) are selected as the key layer set. CRP is applied as " + }, + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "inline_equation", + "content": "E_{x}" + }, + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "text", + "content": " to extract the VC set " + }, + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{S}" + }, + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "text", + "content": ". Following [2], each VC is represented by 15 (i.e., " + }, + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "text", + "content": ") image patches. The mapping function" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4368" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 101, + 552, + 380 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "text", + "content": "Table 1. Examples of VCs and their disentangled concept atoms, along with their concept probability distributions and CPE scores. The first row shows concept 75 of Layer 4 extracted by CRP. The second row shows concept 697 of Layer 4 extracted by MILAN." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 101, + 552, + 380 + ], + "lines": [ + { + "bbox": [ + 58, + 101, + 552, + 380 + ], + "spans": [ + { + "bbox": [ + 58, + 101, + 552, + 380 + ], + "type": "table", + "html": "
Visual Concept and Disentangled Concept AtomsConcept DistributionCPE
Image 1: [burger, lettuce, tomato], \nImage 2: [blue, monkey, face], \nImage 3: [shark, blue, water], \nImage 4: [keyboard, buttons, numbers], \nImage 5: [shark, blue, water], \nImage 6: [shark, blue, water], \nImage 7: [shark, blue, water], \nImage 8: [blue, circular, shape], \nImage 9: [shark, blue, water], \nImage 10: [keyboard, buttons, numbers], \nImage 11: [blue, yellow, objects], \nImage 12: [shark, blue, text], \nImage 13: [shark, blue, water], \nImage 14: [shark, blue, water], \nImage 15: [shark, blue, water]0.30\n0.25\n0.20\n0.15\n0.10\n0.05\n0.00\nhurty green, blue, texture, yellow, objects, water, black0.81
Image 1: [hand, blinds, string], \nImage 2: [curtains, room, grayscale], \nImage 3: [fabric, green, closeup], \nImage 4: [curtains, room, light], \nImage 5: [reflection, glass, light], \nImage 6: [hand, fabric, texture], \nImage 7: [water, fountain, greenery], \nImage 8: [feathers, pattern, dark], \nImage 9: [fabric, blue, closeup], \nImage 10: [fabric, green, seatbelt], \nImage 11: [leaf, bark, texture], \nImage 12: [nature, mountain, grayscale], \nImage 13: [skull, ground, dark], \nImage 14: [fabric, colorful, pattern], \nImage 15: [curtains, window, light]0.12\n0.10\n0.08\n0.06\n0.04\n0.02\n0.00\n0.00\n0.00\n0.00\n0.00\n0.00\n0.00\n0.00\n0.00\n0.00\n0.00\n0.00\n0.00\n0.00\n0.00\n0.00\n0.00\n0.00\n0.00\n0.00\n0.02\n0.02\n0.02\n0.02\n0.02\n0.02\n0.02\n0.02\n0.02\n0.02\n0.02\n0.02\n0.02\n0.02\n0.02\n0.02\n0.02\n0.02\n0.02\n0.02\n0.00\n0.00\n0.00\n0.00\n0.00\n0.00\n0.00\n0.00\n0.00\n0.00\n0.00\n0.00\n0.00\n0.00\n0.00\n0.00\n0.00\n0.00\n0.02\n0.00\n0.00\n0.00\n0.00\n0.00\n0.00\n0.00\n0.00\n0.00\n0.00\n0.00\n0.00\n0.00\n0.00\n0.00\n0.00\n0.00\n0.00\n0.01\n0.01\n0.01\n0.01\n0.01\n0.01\n0.01\n0.01\n0.01\n0.01\n0.01\n0.01\n0.01\n0.01\n0.01\n0.01\n0.01\n0.01\n0.01\n0.01\n0.02\n0.02\n0.02\n0.02\n0.02\n0.02\n0.02\n0.02\n0.02\n0.02\n0.02\n0.02\n0.02\n0.02\n0.02\n0.02\n0.02\n0.02\n0.02\n0.03\n0.03\n0.03\n0.03\n0.03\n0.03\n0.03\n0.03\n0.03\n0.03\n0.03\n0.03\n0.03\n0.03\n0.03\n0.03\n0.03\n0.03\n0.03\n0.03\n0.04\n0.04\n0.04\n0.04\n0.04\n0.04\n0.04\n0.04\n0.04\n0.04\n0.04\n0.04\n0.04\n0.04\n0.04\n0.04\n0.04\n0.04\n0.04\n0.04\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.08\n0.08\n0.08\n0.08\n0.08\n0.08\n0.08\n0.08\n0.08\n0.08\n0.08\n0.08\n0.08\n0.08\n0.08\n0.08\n0.08\n0.08\n0.08\n0.08\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.08\n0.08\n0.08\n0.08\n0.08\n0.08\n0.08\n0.08\n0.08\n0.08\n0.08\n0.08\n0.08\n0.08\n0.08\n0.08\n0.08\n0.08\n0.08\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.07\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.04\n0.04\n0.04\n0.04\n0.04\n0.04\n0.04\n0.04\n0.04\n0.04\n0.04\n0.04\n0.04\n0.04\n0.04\n0.04\n0.04\n0.04\n0.04\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.08\n0.08\n0.08\n0.08\n0.08\n0.08\n0.08\n0.08\n0.08\n0.08\n0.08\n0.08\n0.08\n0.08\n0.08\n0.08\n0.08\n0.08\n0.08\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.06\n0.04\n0.04\n0.04\n0.04\n0.04\n0.04\n0.04\n0.04\n0.04\n0.04\n0.04\n0.04\n0.04\n0.04\n0.04\n0.04\n0.04\n0.04\n0.04\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.05\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.09\n0.02\n0.12\n0.12\n0.12\n0.12\n0.12\n0.12\n0.12\n0.12\n0.12\n0.12\n0.12\n0.12\n0.12\n0.12\n0.12\n0.12\n0.12\n0.12\n0.12\n0.12\n0.25\n0.25\n0.25\n0.25\n0.25\n0.25\n0.25\n0.25\n0.25\n0.25\n0.25\n0.25\n0.25\n0.25\n0.25\n0.25\n0.25\n0.25\n0.25\n0.25\n0.26\n0.26\n0.26\n0.26\n0.26\n0.26\n0.26\n0.26\n0.26\n0.26\n0.26\n0.26\n0.26\n0.26\n0.26\n0.26\n0.26\n0.26\n0.26\n0.26\n0.25\n0.25\n0.25\n0.25\n0.25\n0.25\n0.25\n0.25\n0.25\n0.25\n0.25\n0.25\n0.25\n0.25\n0.25\n0.25\n0.25\n0.25\n0.25\n0.2.\n12\n12\n12\n12\n12\n12\n12\n12\n12\n12\n12\n12\n12\n12\n12\n12\n12\n12\n12\n12\n12\n12\n12\n12\n12\n12\n12\n12\n12\n12\n12\n12\n12\n12\n25\n25\n25\n25\n25\n25\n25\n25\n25\n25\n25\n25\n25\n25\n25\n25\n25\n25\n25\n25\n25\n25\n25\n25\n25\n25\n25\n25\n25\n25\n25\n25\n25\n25
", + "image_path": "f65763e0d8dc88cba8549798805f243cf4a1783d6e014205e1854af5a41178fe.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 389, + 296, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 389, + 296, + 510 + ], + "spans": [ + { + "bbox": [ + 55, + 389, + 296, + 510 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 55, + 389, + 296, + 510 + ], + "type": "text", + "content": " from channels to VCs is built. We then employ a GPT4o-2024-08-06 model as the function " + }, + { + "bbox": [ + 55, + 389, + 296, + 510 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 55, + 389, + 296, + 510 + ], + "type": "text", + "content": " to generate the description set " + }, + { + "bbox": [ + 55, + 389, + 296, + 510 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 55, + 389, + 296, + 510 + ], + "type": "text", + "content": " (i.e., Eq. 3). The disentangled and clustered atom set " + }, + { + "bbox": [ + 55, + 389, + 296, + 510 + ], + "type": "inline_equation", + "content": "\\mathcal{A}^*" + }, + { + "bbox": [ + 55, + 389, + 296, + 510 + ], + "type": "text", + "content": " (processed by a DeBERTa model), the probability set " + }, + { + "bbox": [ + 55, + 389, + 296, + 510 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}" + }, + { + "bbox": [ + 55, + 389, + 296, + 510 + ], + "type": "text", + "content": ", and the CPE set " + }, + { + "bbox": [ + 55, + 389, + 296, + 510 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 55, + 389, + 296, + 510 + ], + "type": "text", + "content": " are computed and gathered accordingly. Overall, ACD-" + }, + { + "bbox": [ + 55, + 389, + 296, + 510 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_{\\mathcal{M},\\mathcal{T}}" + }, + { + "bbox": [ + 55, + 389, + 296, + 510 + ], + "type": "text", + "content": " containing 3,840 entries is constructed, with each entry comprising a channel, a CPE score, and 15 VC representations—each consisting of 3 concept atoms (i.e., " + }, + { + "bbox": [ + 55, + 389, + 296, + 510 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 55, + 389, + 296, + 510 + ], + "type": "text", + "content": ")—and multiple orthogonal linguistic atoms, along with their associated probabilities." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 515, + 276, + 529 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 515, + 276, + 529 + ], + "spans": [ + { + "bbox": [ + 55, + 515, + 276, + 529 + ], + "type": "text", + "content": "4.2. Disentanglement of Polysemantic Concept" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 534, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 296, + 713 + ], + "type": "text", + "content": "CPDF automatically disentangles and describes the polysemantic VCs into a set of linguistic concept atoms, " + }, + { + "bbox": [ + 55, + 534, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 55, + 534, + 296, + 713 + ], + "type": "text", + "content": ". As exemplified by Table 1, two randomly selected VCs are extracted by two advanced XAI methods (CRP and MILAN). They are disentangled and described into atoms, exhibiting different degrees of polysemasticity. We observe that the commonalities within the VCs are almost entirely identified, and most of the atoms describe the commonalities of subsets of image patches. In the first row (CPE=0.81), " + }, + { + "bbox": [ + 55, + 534, + 296, + 713 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 55, + 534, + 296, + 713 + ], + "type": "text", + "content": " successfully identifies all visually discernible commonalities, such as shark and water. The atoms exhibit high consistency, and fewer but more frequently occurring common semantics are identified. Only a few atoms, such as the burger, are not commonalities, as their corresponding images lack sufficient common semantics. The atoms are also repre" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 389, + 555, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 389, + 555, + 475 + ], + "spans": [ + { + "bbox": [ + 313, + 389, + 555, + 475 + ], + "type": "text", + "content": "sented accurately. For instance, the atom sets for each image featuring the shark semantic all include the word shark. In the second row (CPE=0.95), our method remains effective in disentangling and describing VC with a higher level of polysemantics. " + }, + { + "bbox": [ + 313, + 389, + 555, + 475 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 313, + 389, + 555, + 475 + ], + "type": "text", + "content": " identifies a greater variety of atoms. These results demonstrate the effectiveness of the disentanglement and description operation in the CPDF." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 479, + 481, + 493 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 479, + 481, + 493 + ], + "spans": [ + { + "bbox": [ + 313, + 479, + 481, + 493 + ], + "type": "text", + "content": "4.3. Concept Distribution and CPE" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 498, + 555, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 555, + 641 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 555, + 641 + ], + "type": "text", + "content": "CPE can quantify the polysemanticity and serve as an indicator of a DVM's interpretability. It begins by calculating the probabilities of each orthogonal atom. As the middle column of Table 1 shows, the distribution provides statistics on the extent to which different atoms are associated with each concept. For instance, concept 75 primarily focuses on blue with a probability of 0.27 while also identifying yellow with a smaller probability. The semantics between atoms are non-overlapping. This probabilistic approach enhances explainability by capturing multiple semantics with their occurring probabilities, upgrading the modeling of deterministic concepts to uncertain atom distributions." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "text", + "content": "CPE quantifies polysemanticity from 3 levels of granularity. As shown in Fig. 3a, concepts exhibit varying degrees of polysemanticity, with considerable fluctuations (ranging from a maximum of 1 to a minimum of 0.5). Most of the concepts demonstrate a noticeable level of polysemanticity. When analyzing the polysemanticity of different" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4369" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 71, + 221, + 193 + ], + "blocks": [ + { + "bbox": [ + 59, + 71, + 221, + 193 + ], + "lines": [ + { + "bbox": [ + 59, + 71, + 221, + 193 + ], + "spans": [ + { + "bbox": [ + 59, + 71, + 221, + 193 + ], + "type": "image", + "image_path": "837cd38290cbb3b4ff70a449f9fa6f7065fa0250c4ccccc3b23df402c253ae85.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 66, + 195, + 212, + 205 + ], + "lines": [ + { + "bbox": [ + 66, + 195, + 212, + 205 + ], + "spans": [ + { + "bbox": [ + 66, + 195, + 212, + 205 + ], + "type": "text", + "content": "(a) Sorted CPE along the channel dimension." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 223, + 73, + 388, + 193 + ], + "blocks": [ + { + "bbox": [ + 223, + 73, + 388, + 193 + ], + "lines": [ + { + "bbox": [ + 223, + 73, + 388, + 193 + ], + "spans": [ + { + "bbox": [ + 223, + 73, + 388, + 193 + ], + "type": "image", + "image_path": "e26469c49c6efe022b123a098cd7322cb4218aa12f003390f9d1044ee078a1e0.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 249, + 195, + 361, + 205 + ], + "lines": [ + { + "bbox": [ + 249, + 195, + 361, + 205 + ], + "spans": [ + { + "bbox": [ + 249, + 195, + 361, + 205 + ], + "type": "text", + "content": "(b) Averaged CPE on 4 key layers." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 393, + 72, + 550, + 194 + ], + "blocks": [ + { + "bbox": [ + 393, + 72, + 550, + 194 + ], + "lines": [ + { + "bbox": [ + 393, + 72, + 550, + 194 + ], + "spans": [ + { + "bbox": [ + 393, + 72, + 550, + 194 + ], + "type": "image", + "image_path": "e8ed054dee5d0031be4da582a14299832de27b55662fe516aa50fa044e9a1663.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 419, + 195, + 522, + 205 + ], + "lines": [ + { + "bbox": [ + 419, + 195, + 522, + 205 + ], + "spans": [ + { + "bbox": [ + 419, + 195, + 522, + 205 + ], + "type": "text", + "content": "(c) Averaged CPE on 3 models." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 73, + 215, + 535, + 226 + ], + "lines": [ + { + "bbox": [ + 73, + 215, + 535, + 226 + ], + "spans": [ + { + "bbox": [ + 73, + 215, + 535, + 226 + ], + "type": "text", + "content": "Figure 3. Illustrations of CPE scores at 3 levels of granularity. In (a), the channel indices of each layer are max normalized to 1." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 57, + 259, + 290, + 297 + ], + "blocks": [ + { + "bbox": [ + 55, + 235, + 295, + 257 + ], + "lines": [ + { + "bbox": [ + 55, + 235, + 295, + 257 + ], + "spans": [ + { + "bbox": [ + 55, + 235, + 295, + 257 + ], + "type": "text", + "content": "Table 2. The consistency of the recognition of concept polysemantics between the CPE metric and human scorers." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 259, + 290, + 297 + ], + "lines": [ + { + "bbox": [ + 57, + 259, + 290, + 297 + ], + "spans": [ + { + "bbox": [ + 57, + 259, + 290, + 297 + ], + "type": "table", + "html": "
Consistency
CPE Metric and Human Scorers75 %
", + "image_path": "34ec9f81003292b9fa199d3f1b485a5b42396cac8c156878d62ffb434c890dc9.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 54, + 305, + 295, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 305, + 295, + 472 + ], + "spans": [ + { + "bbox": [ + 54, + 305, + 295, + 472 + ], + "type": "text", + "content": "layers, as shown in Fig. 3b, the averaged CPE peaks in the third stage, while the shallowest and deepest layers exhibit lower values. Typically, deeper layers capture abstract concepts, whereas the shallower layers extract elementary features like textures [53, 56]. DVMs show better monosemanticity when identifying elementary and abstract concepts, while more ambiguous concepts are captured in the intermediate layers. This experiment validates the theoretical and experimental claims of previous work [14, 32]. In Fig. 3c, we observe that as the number of layers increases, the CPE of the model shows a rising trend. This result indicates a direct proportional relationship between the polysemanticity and the complexity of DVMs, which is that the more complex a DVM becomes, the harder it is to interpret." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 473, + 295, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 473, + 295, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 473, + 295, + 628 + ], + "type": "text", + "content": "The CPE is evaluated in comparison with human assessments. As shown in Table 1, the CPE is consistent with the degree of polysemanticity from the human intuitive observations (the polysemanticity of the first row is smaller than that of the second row). Similarly, we randomly sample 300 pairs of VCs, each containing two VCs with varying levels of polysemanticity. Comparisons are drawn from these 300 pairs by both human evaluators and the CPE metric. As shown in Table 2, results from human assessments and CPE achieve a " + }, + { + "bbox": [ + 55, + 473, + 295, + 628 + ], + "type": "inline_equation", + "content": "75\\%" + }, + { + "bbox": [ + 55, + 473, + 295, + 628 + ], + "type": "text", + "content": " agreement rate, demonstrating the effectiveness of CPE. Other results are shown in Sec. S4 of the Appendix. All of these results reveal that CPE is effective in quantifying the polysemanticity of concepts and DVMs." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 635, + 209, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 635, + 209, + 647 + ], + "spans": [ + { + "bbox": [ + 55, + 635, + 209, + 647 + ], + "type": "text", + "content": "4.4. CoE for Local Explanations" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "content": "CoE generates local explanations akin to CoT by automatically describing the concept circuit. We set " + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\alpha = 0.001" + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "content": " to acknowledge the top relevant VCs, resulting in an average of 3 key VCs per layer. As exemplified in Table 3, given an input image, CoE first outputs whether the prediction of the" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 236, + 555, + 404 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 236, + 555, + 404 + ], + "spans": [ + { + "bbox": [ + 313, + 236, + 555, + 404 + ], + "type": "text", + "content": "DVM is correct and then gives reasons behind this prediction according to the information along the concept circuits. CoE employs an LLM as an engine to infer the logical relationships among concepts, as well as between concepts and the context. It synthesizes all evidence to generate an inductive commentary on the predictions. The explanation chain is presented in a linguistic format, enhancing human understanding. For instance, CoE explains why the DVM incorrectly predicted a tench as a mosquito net: the top activation of mosquito-net-related concepts, rather than those related to fish, misled the decision. Additional instances are shown in Sec. S5 of the Appendix. Overall, CoE effectively captures and elucidates the core decision-making evidence for both correct and incorrect predictions of DVMs." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 416, + 554, + 429 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 416, + 554, + 429 + ], + "spans": [ + { + "bbox": [ + 313, + 416, + 554, + 429 + ], + "type": "text", + "content": "4.5. Quantitative Evaluation via GPT-4o & Human" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 437, + 554, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 554, + 568 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 554, + 568 + ], + "type": "text", + "content": "We employ GPT-4o and human evaluators to assess the local explanations for 500 and 100 samples randomly selected from the ImageNet Validation dataset [15], respectively. The GPT-4o evaluation is achieved through a carefully designed prompt prompt-coe-eval. Explanations generated by three methods are rated according to three explainability metrics: Accuracy, Completeness, and User Interpretability [47]. Each is assigned a three-tier scoring system (2, 1, and 0), with specific criteria defined for each tier, resulting in a total explanation score of 6. Please refer to Sec. S2 and S5 of the Appendix for details." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "content": "As shown in Table 4, our CoE approach attains peak performance across all three metrics, achieving total scores of 5.06 and 5.04 in GPT-4o and human evaluations, respectively. The lowest scores (3.14 and 2.62) are obtained on the MILAN ANNOTATION dataset without the ACD and CPDF methods. This dataset was manually annotated, resulting in inaccuracy and insufficiency within concept explanations. Scores increase by 0.11 and 0.72, respectively, with the application of our proposed ACD-" + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_{\\mathcal{M},\\mathcal{T}}" + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "content": " dataset. The ACD method not only enables the automated disentanglement and description of concept atoms but also enhances their precision and comprehensibility. When the CoE em" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4370" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 59, + 71, + 550, + 82 + ], + "lines": [ + { + "bbox": [ + 59, + 71, + 550, + 82 + ], + "spans": [ + { + "bbox": [ + 59, + 71, + 550, + 82 + ], + "type": "text", + "content": "Table 3. Cases of local explanations of CoE. The first row shows the explanation of a correct prediction, and the second row is reversed." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 59, + 95, + 175, + 183 + ], + "blocks": [ + { + "bbox": [ + 59, + 95, + 175, + 183 + ], + "lines": [ + { + "bbox": [ + 59, + 95, + 175, + 183 + ], + "spans": [ + { + "bbox": [ + 59, + 95, + 175, + 183 + ], + "type": "image", + "image_path": "6649d862f2f93427734b895aff043653f785e8b10d9b9ad65eed48d8fc7b2ee9.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 90, + 184, + 107, + 191 + ], + "lines": [ + { + "bbox": [ + 90, + 184, + 107, + 191 + ], + "spans": [ + { + "bbox": [ + 90, + 184, + 107, + 191 + ], + "type": "text", + "content": "label:" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 61, + 184, + 89, + 221 + ], + "blocks": [ + { + "bbox": [ + 61, + 184, + 89, + 221 + ], + "lines": [ + { + "bbox": [ + 61, + 184, + 89, + 221 + ], + "spans": [ + { + "bbox": [ + 61, + 184, + 89, + 221 + ], + "type": "image", + "image_path": "e6e0418a4bb3aa96e7aba5d4fc71145eef7293294ccc89b41a5b018021469042.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 90, + 194, + 175, + 201 + ], + "lines": [ + { + "bbox": [ + 90, + 194, + 175, + 201 + ], + "spans": [ + { + "bbox": [ + 90, + 194, + 175, + 201 + ], + "type": "text", + "content": "black-and-tan coonhound" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 90, + 202, + 114, + 210 + ], + "lines": [ + { + "bbox": [ + 90, + 202, + 114, + 210 + ], + "spans": [ + { + "bbox": [ + 90, + 202, + 114, + 210 + ], + "type": "text", + "content": "predict:" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 96, + 212, + 175, + 219 + ], + "lines": [ + { + "bbox": [ + 96, + 212, + 175, + 219 + ], + "spans": [ + { + "bbox": [ + 96, + 212, + 175, + 219 + ], + "type": "text", + "content": "black-and-tan coonhound" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 177, + 98, + 552, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 177, + 98, + 552, + 201 + ], + "spans": [ + { + "bbox": [ + 177, + 98, + 552, + 201 + ], + "type": "text", + "content": "The model outputs a correct result: black-and-tan coonhound. Specifically, in the shallowest layer, layer 1, channel 88 with a relevance value of 1.0 describes the concept curved, which may relate to the curved shapes observed in parts of the dog such as its tail or ears. Moving deeper, in layer 2, channel 506 with a relevance value of 0.85 and channel 391 with a relevance value of 0.75 describe the concept of animal, indicating the model has identified an animal in the image, and channel 229 with a relevance value of 0.77 emphasizes the color black, aligning with the black and tan colors of the coonhound. Further, layer 2 also identifies the presence of grass several times (channels 94 and 103), which corresponds to the grassy field described in the image caption. In layer 3, various channels, such as 985, 174, 422, and 13, with high relevance scores, from 0.72 to 1.0, repeatedly identify the color green and the concept grass, firmly establishing the setting of the image. Additionally, channels 11, 787, and 591 in layer 3 enhance the certainty by repeatedly identifying animal and creature, confirming that the subject is an animal. Finally, layer 4 decisively contributes with channels 1354 and 1661, both signifying dog with high relevance values (1.0 and 0.72 respectively), which correctly identifies the animal in the image as a dog." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 178, + 202, + 552, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 178, + 202, + 552, + 219 + ], + "spans": [ + { + "bbox": [ + 178, + 202, + 552, + 219 + ], + "type": "text", + "content": "Given that the model's structured concepts strongly match the description and characteristics noted in the caption and the ground truth label, it is clear why the model outputs the correct result of black-and-tan coonhound." + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 59, + 229, + 176, + 317 + ], + "blocks": [ + { + "bbox": [ + 59, + 229, + 176, + 317 + ], + "lines": [ + { + "bbox": [ + 59, + 229, + 176, + 317 + ], + "spans": [ + { + "bbox": [ + 59, + 229, + 176, + 317 + ], + "type": "image", + "image_path": "3f9920565c7c12df94010542ee15f666995976bb18fbb963c0e8f48eb2c8fc38.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 135, + 326, + 156, + 334 + ], + "lines": [ + { + "bbox": [ + 135, + 326, + 156, + 334 + ], + "spans": [ + { + "bbox": [ + 135, + 326, + 156, + 334 + ], + "type": "text", + "content": "tench" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 127, + 336, + 175, + 344 + ], + "lines": [ + { + "bbox": [ + 127, + 336, + 175, + 344 + ], + "spans": [ + { + "bbox": [ + 127, + 336, + 175, + 344 + ], + "type": "text", + "content": "mosquito net" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 62, + 317, + 91, + 354 + ], + "blocks": [ + { + "bbox": [ + 62, + 317, + 91, + 354 + ], + "lines": [ + { + "bbox": [ + 62, + 317, + 91, + 354 + ], + "spans": [ + { + "bbox": [ + 62, + 317, + 91, + 354 + ], + "type": "image", + "image_path": "1982e38d7a411595f21ae29d6975d4eb3f1390c4182d501fe2578a0618818a49.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 93, + 326, + 113, + 334 + ], + "lines": [ + { + "bbox": [ + 93, + 326, + 113, + 334 + ], + "spans": [ + { + "bbox": [ + 93, + 326, + 113, + 334 + ], + "type": "text", + "content": "label:" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 93, + 336, + 120, + 344 + ], + "lines": [ + { + "bbox": [ + 93, + 336, + 120, + 344 + ], + "spans": [ + { + "bbox": [ + 93, + 336, + 120, + 344 + ], + "type": "text", + "content": "predict:" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 178, + 232, + 552, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 178, + 232, + 552, + 354 + ], + "spans": [ + { + "bbox": [ + 178, + 232, + 552, + 354 + ], + "type": "text", + "content": "The model outputs an incorrect result mosquito net instead of the correct label tench. Specifically, in the shallowest layer 1 of the model, channel 88 with a relevance value of 1.0 describes the concept of smooth, channel 61 with a relevance value of 0.81 describes the concept of green, and channel 111 with a relevance value of 0.72 describes the concept of brown. In the deeper layer 2 of the model, channel 94 with a relevance value of 1.0 describes the concept of water, channel 229 with a relevance value of 0.84 describes the concept of fish, channel 103 with a relevance value of 0.81 also describes the concept of fish, and channel 467 with a relevance value of 0.73 describes the concept of mesh. At layer 3, channel 985 with a relevance value of 1.0 further describes the concept of fish, while channels 894, 122, and 715 with relevance values of 0.83, 0.73, and 0.72, respectively, describe the concept of green, possibly relating to the color of the net. In the deepest layer 4, channel 697 with a relevance value of 1.0 describes the concept of net, channel 1878 with a relevance value of 0.76 describes the concept of baby, and channel 321 with a relevance value of 0.76 represents the concept of hexagon, which could misconstrue the net structure. The prevalence of concepts related to fish within the decision path suggests a relationship to the tench, but the final prediction of mosquito net was likely influenced by the dominance of net-related concepts. Thus, the model was misled by the presence of a net, resulting in the incorrect classification of mosquito net." + } + ] + } + ], + "index": 15 + }, + { + "type": "table", + "bbox": [ + 58, + 410, + 550, + 512 + ], + "blocks": [ + { + "bbox": [ + 55, + 368, + 555, + 402 + ], + "lines": [ + { + "bbox": [ + 55, + 368, + 555, + 402 + ], + "spans": [ + { + "bbox": [ + 55, + 368, + 555, + 402 + ], + "type": "text", + "content": "Table 4. Scores of CoE for the local explanation on three methods evaluated by GPT-4o and human scorers. The first method utilizes MILAN ANNOTATION as the concept explanation dataset, while the rest two work on the ACD-" + }, + { + "bbox": [ + 55, + 368, + 555, + 402 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_{\\mathcal{M},\\mathcal{T}}" + }, + { + "bbox": [ + 55, + 368, + 555, + 402 + ], + "type": "text", + "content": " datasets. Each criterion has a maximum score of 2 points. The total explanation score is 6 points. Higher scores indicate better explainability." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 410, + 550, + 512 + ], + "lines": [ + { + "bbox": [ + 58, + 410, + 550, + 512 + ], + "spans": [ + { + "bbox": [ + 58, + 410, + 550, + 512 + ], + "type": "table", + "html": "
MethodAccuracyCompletenessUser Interpret.Total Explanation
GPT-4oMILAN A.[26] + Description (Baseline)1.031.071.043.14
CoE w/o Filtering1.071.101.083.25
CoE (ours)1.691.701.675.06
HumanMILAN A.[26] + Description (Baseline)0.870.830.922.62
CoE w/o Filtering1.101.161.073.34
CoE (ours)1.731.721.595.04
", + "image_path": "09c80edd499595a3a5ba2f76b1b1fcb08b581874b092b043ce6228e05a853191.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "table_body" + } + ], + "index": 17 + }, + { + "bbox": [ + 55, + 526, + 295, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 526, + 295, + 658 + ], + "spans": [ + { + "bbox": [ + 55, + 526, + 295, + 658 + ], + "type": "text", + "content": "ploys the full CPDF mechanism, the explanation score rises by 1.92 and 2.42 points (the average absolute improvement is " + }, + { + "bbox": [ + 55, + 526, + 295, + 658 + ], + "type": "inline_equation", + "content": "36\\%" + }, + { + "bbox": [ + 55, + 526, + 295, + 658 + ], + "type": "text", + "content": "), respectively, with the filtering step contributing 1.81 and 1.70 points. This result suggests that by taking polysemanticity and contextual filtering of concepts into account, CoE effectively guides the LLM in producing better linguistic explanations, substantially enhancing their accuracy, completeness, and interpretability for users. The results of the human evaluation are consistent with the GPT-4o results. All of these results strongly affirm the superiority of the proposed CoE approach." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 55, + 669, + 128, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 669, + 128, + 681 + ], + "spans": [ + { + "bbox": [ + 55, + 669, + 128, + 681 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "content": "In this paper, we propose a CoE approach to tackle the challenges of the inflexibility to automatically construct con" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 526, + 555, + 706 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 526, + 555, + 706 + ], + "spans": [ + { + "bbox": [ + 313, + 526, + 555, + 706 + ], + "type": "text", + "content": "cept explanation datasets, insufficient linguistic explanations, and the weakness in managing concept polysemantics of concept-based XAI methods. CoE automates the decoding and description of concepts via an LVLM and constructs global concept explanation datasets. A CPDF mechanism is designed to identify the most contextually relevant concept atoms, advancing the modeling of deterministic concepts to uncertain concept atom distributions. The CPE is proposed to quantify the polysemanticsy. Finally, CoE enables local explanations of a DVM's decisionmaking process, represented through natural language. The effectiveness and superiority of CPE and CoE are demonstrated. In summary, CoE establishes a systematical framework for explaining DVMs, and we believe that the findings presented in this paper warrant further exploration." + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "4371" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "type": "text", + "content": "Acknowledgment" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 295, + 127 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 295, + 127 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 295, + 127 + ], + "type": "text", + "content": "This research is supported in part by the National Natural Science Foundation of China under Grants U23B2049, 62276186, and 61925602." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 140, + 114, + 153 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 140, + 114, + 153 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 114, + 153 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 160, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 160, + 296, + 214 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 160, + 296, + 214 + ], + "spans": [ + { + "bbox": [ + 61, + 160, + 296, + 214 + ], + "type": "text", + "content": "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023.5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 217, + 296, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 217, + 296, + 281 + ], + "spans": [ + { + "bbox": [ + 61, + 217, + 296, + 281 + ], + "type": "text", + "content": "[2] Reduan Achtibat, Maximilian Dreyer, Ilona Eisenbraun, Sebastian Bosse, Thomas Wiegand, Wojciech Samek, and Sebastian Lapuschkin. From attribution maps to human-understandable explanations through concept relevance propagation. Nature Machine Intelligence, 5(9):1006-1019, 2023. 1, 2, 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 284, + 295, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 284, + 295, + 327 + ], + "spans": [ + { + "bbox": [ + 62, + 284, + 295, + 327 + ], + "type": "text", + "content": "[3] Rajaram Achyuta, Chowdhury Neil, Torralba Antonio, Andreas Jacob, and Schwettmann Sarah. Automatic discovery of visual circuits. In NeurIPS Workshop on Attributing Model Behavior at Scale, 2023. 1, 2, 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 329, + 295, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 329, + 295, + 384 + ], + "spans": [ + { + "bbox": [ + 62, + 329, + 295, + 384 + ], + "type": "text", + "content": "[4] Sebastian Bach, Alexander Binder, Grégoire Montavon, Frederick Klauschen, Klaus-Robert Müller, and Wojciech Samek. On pixel-wise explanations for non-linear classifier decisions by layer-wise relevance propagation. PLoS ONE, 10, 2015. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 386, + 296, + 440 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 386, + 296, + 440 + ], + "spans": [ + { + "bbox": [ + 62, + 386, + 296, + 440 + ], + "type": "text", + "content": "[5] David Bau, Jun-Yan Zhu, Hendrik Strobelt, Agata Lapedriza, Bolei Zhou, and Antonio Torralba. Understanding the role of individual units in a deep neural network. Proceedings of the National Academy of Sciences, 117(48):30071-30078, 2020. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 442, + 295, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 442, + 295, + 475 + ], + "spans": [ + { + "bbox": [ + 62, + 442, + 295, + 475 + ], + "type": "text", + "content": "[6] Leonard Bereska and Efstratios Gavves. Mechanistic interpretability for ai safety-a review. arXiv preprint arXiv:2404.14082, 2024. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 476, + 295, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 476, + 295, + 520 + ], + "spans": [ + { + "bbox": [ + 62, + 476, + 295, + 520 + ], + "type": "text", + "content": "[7] Subrato Bharati, M Rubaiyat Hossain Mondal, and Prajoy Podder. A review on explainable artificial intelligence for healthcare: why, how, and when? IEEE Transactions on Artificial Intelligence, 2023. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 522, + 295, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 522, + 295, + 566 + ], + "spans": [ + { + "bbox": [ + 62, + 522, + 295, + 566 + ], + "type": "text", + "content": "[8] Moritz Böhle, Navdeeppal Singh, Mario Fritz, and Bernt Schiele. B-cos alignment for inherently interpretable cnns and vision transformers. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(6):4504-4518, 2024. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 567, + 295, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 567, + 295, + 622 + ], + "spans": [ + { + "bbox": [ + 62, + 567, + 295, + 622 + ], + "type": "text", + "content": "[9] Zhe Chen, Weiyun Wang, Hao Tian, Shenglong Ye, Zhang-wei Gao, Erfei Cui, Wenwen Tong, Kongzhi Hu, Jiapeng Luo, Zheng Ma, et al. How far are we to gpt-4v? closing the gap to commercial multimodal models with open-source suites. arXiv preprint arXiv:2404.16821, 2024. 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 624, + 295, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 624, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 57, + 624, + 295, + 689 + ], + "type": "text", + "content": "[10] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24185–24198, 2024. 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "type": "text", + "content": "[11] Ming-Ming Cheng, Peng-Tao Jiang, Ling-Hao Han, Liang Wang, and Philip Torr. Deeply explain cnn via hierarchical" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 715 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "decomposition. International Journal of Computer Vision, 131(5):1091-1105, 2023. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 96, + 553, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 173 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 173 + ], + "type": "text", + "content": "[12] Zheng Chu, Jingchang Chen, Qianglong Chen, Weijiang Yu, Tao He, Haotian Wang, Weihua Peng, Ming Liu, Bing Qin, and Ting Liu. Navigate through enigmatic labyrinth a survey of chain of thought reasoning: Advances, frontiers and future. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1173–1203, 2024. 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 175, + 553, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 553, + 228 + ], + "type": "text", + "content": "[13] Arthur Conmy, Augustine Mavor-Parker, Aengus Lynch, Stefan Heimersheim, and Adrià Garriga-Alonso. Towards automated circuit discovery for mechanistic interpretability. Advances in Neural Information Processing Systems, 36: 16318-16352, 2023. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 230, + 553, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 230, + 553, + 274 + ], + "spans": [ + { + "bbox": [ + 316, + 230, + 553, + 274 + ], + "type": "text", + "content": "[14] Meghal Dani, Isabel Rio-Torto, Stephan Alaniz, and Zeynep Akata. Devil: Decoding vision features into language. In DAGM German Conference on Pattern Recognition, pages 363-377. Springer, 2023. 1, 4, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 276, + 553, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 276, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 316, + 276, + 553, + 319 + ], + "type": "text", + "content": "[15] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 321, + 553, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 321, + 553, + 386 + ], + "spans": [ + { + "bbox": [ + 317, + 321, + 553, + 386 + ], + "type": "text", + "content": "[16] Maximilian Dreyer, Erblina Purelku, Johanna Vielhaben, Wojciech Samek, and Sebastian Lapuschkin. Pure: Turning polysemantic neurons into pure features by identifying relevant circuits. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 8212-8217, 2024. 1, 2, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 388, + 553, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 388, + 553, + 442 + ], + "spans": [ + { + "bbox": [ + 317, + 388, + 553, + 442 + ], + "type": "text", + "content": "[17] Nelson Elhage, Neel Nanda, Catherine Olsson, Tom Henighan, Nicholas Joseph, Ben Mann, Amanda Askell, Yuntao Bai, Anna Chen, Tom Conerly, et al. A mathematical framework for transformer circuits. Transformer Circuits Thread, 1(1):12, 2021. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 444, + 553, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 444, + 553, + 497 + ], + "spans": [ + { + "bbox": [ + 317, + 444, + 553, + 497 + ], + "type": "text", + "content": "[18] Nelson Elhage, Tristan Hume, Catherine Olsson, Nicholas Schiefer, Tom Henighan, Shauna Kravec, Zac Hatfield-Dodds, Robert Lasenby, Dawn Drain, Carol Chen, et al. Toy models of superposition. arXiv preprint arXiv:2209.10652, 2022.3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 500, + 553, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 500, + 553, + 533 + ], + "spans": [ + { + "bbox": [ + 316, + 500, + 553, + 533 + ], + "type": "text", + "content": "[19] Sebastian Farquhar, Jannik Kossen, Lorenz Kuhn, and Yarin Gal. Detecting hallucinations in large language models using semantic entropy. Nature, 630(8017):625-630, 2024. 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 534, + 553, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 534, + 553, + 599 + ], + "spans": [ + { + "bbox": [ + 317, + 534, + 553, + 599 + ], + "type": "text", + "content": "[20] Thomas Fel, Victor Boutin, Louis Béthune, Rémi Cadène, Mazda Moayeri, Léo Andéol, Mathieu Chalvidal, and Thomas Serre. A holistic approach to unifying automatic concept extraction and concept importance estimation. Advances in Neural Information Processing Systems, 36, 2024. 2, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 601, + 553, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 601, + 553, + 645 + ], + "spans": [ + { + "bbox": [ + 316, + 601, + 553, + 645 + ], + "type": "text", + "content": "[21] Guhao Feng, Bohang Zhang, Yuntian Gu, Haotian Ye, Di He, and Liwei Wang. Towards revealing the mystery behind chain of thought: a theoretical perspective. Advances in Neural Information Processing Systems, 36, 2024. 2, 4" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "type": "text", + "content": "[22] Amirata Ghorbani, James Wexler, James Y Zou, and Been Kim. Towards automatic concept-based explanations. Advances in neural information processing systems, 32, 2019. 1, 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 691, + 553, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 715 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 715 + ], + "type": "text", + "content": "[23] Dan W Grupe and Jack B Nitschke. Uncertainty and anticipation in anxiety: an integrated neurobiological and psy" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4372" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "text", + "content": "chological perspective. Nature Reviews Neuroscience, 14(7): 488-501, 2013. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 140 + ], + "type": "text", + "content": "[24] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 770-778, 2016. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 141, + 294, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 141, + 294, + 185 + ], + "spans": [ + { + "bbox": [ + 56, + 141, + 294, + 185 + ], + "type": "text", + "content": "[25] Pengcheng He, Xiaodong Liu, Jianfeng Gao, and Weizhu Chen. Deberta: Decoding-enhanced bert with disentangled attention. In International Conference on Learning Representations, 2020. 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 186, + 295, + 231 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 186, + 295, + 231 + ], + "spans": [ + { + "bbox": [ + 56, + 186, + 295, + 231 + ], + "type": "text", + "content": "[26] Evan Hernandez, Sarah Schwettmann, David Bau, Teona Bagashvili, Antonio Torralba, and Jacob Andreas. Natural language descriptions of deep features. In International Conference on Learning Representations, 2022. 1, 2, 4, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 232, + 295, + 286 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 232, + 295, + 286 + ], + "spans": [ + { + "bbox": [ + 56, + 232, + 295, + 286 + ], + "type": "text", + "content": "[27] Been Kim, Martin Wattenberg, Justin Gilmer, Carrie Cai, James Wexler, Fernanda Viegas, et al. Interpretability beyond feature attribution: Quantitative testing with concept activation vectors (tcav). In International conference on machine learning, pages 2668-2677. PMLR, 2018. 1, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 287, + 295, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 287, + 295, + 342 + ], + "spans": [ + { + "bbox": [ + 56, + 287, + 295, + 342 + ], + "type": "text", + "content": "[28] Anton Kuznietsov, Balint Gyevnar, Cheng Wang, Steven Peters, and Stefano V. Albrecht. Explainable ai for safe and trustworthy autonomous driving: A systematic review. IEEE Transactions on Intelligent Transportation Systems, pages 1-23, 2024. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 343, + 295, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 343, + 295, + 365 + ], + "spans": [ + { + "bbox": [ + 56, + 343, + 295, + 365 + ], + "type": "text", + "content": "[29] Yann LeCun, Yoshua Bengio, and Geoffrey Hinton. Deep learning. nature, 521(7553):436-444, 2015. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 366, + 295, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 366, + 295, + 421 + ], + "spans": [ + { + "bbox": [ + 56, + 366, + 295, + 421 + ], + "type": "text", + "content": "[30] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730-19742. PMLR, 2023. 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 422, + 295, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 422, + 295, + 477 + ], + "spans": [ + { + "bbox": [ + 56, + 422, + 295, + 477 + ], + "type": "text", + "content": "[31] Haomiao Liu, Ruiping Wang, Shiguang Shan, and Xilin Chen. What is a tabby? interpretable model decisions by learning attribute-based classification criteria. IEEE transactions on pattern analysis and machine intelligence, 43(5): 1791-1807, 2019. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 478, + 295, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 478, + 295, + 544 + ], + "spans": [ + { + "bbox": [ + 56, + 478, + 295, + 544 + ], + "type": "text", + "content": "[32] Francesco Locatello, Stefan Bauer, Mario Lucic, Gunnar Raetsch, Sylvain Gelly, Bernhard Schölkopf, and Olivier Bachem. Challenging common assumptions in the unsupervised learning of disentangled representations. In international conference on machine learning, pages 4114-4124. PMLR, 2019. 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 545, + 295, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 545, + 295, + 578 + ], + "spans": [ + { + "bbox": [ + 56, + 545, + 295, + 578 + ], + "type": "text", + "content": "[33] Simon C Marshall and Jan H Kirchner. Understanding polysemantics in neural networks through coding theory. arXiv preprint arXiv:2401.17975, 2024. 2, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 579, + 295, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 579, + 295, + 612 + ], + "spans": [ + { + "bbox": [ + 56, + 579, + 295, + 612 + ], + "type": "text", + "content": "[34] Dang Minh, H Xiang Wang, Y Fen Li, and Tan N Nguyen. Explainable artificial intelligence: a comprehensive review. Artificial Intelligence Review, pages 1-66, 2022. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 613, + 295, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 613, + 295, + 658 + ], + "spans": [ + { + "bbox": [ + 56, + 613, + 295, + 658 + ], + "type": "text", + "content": "[35] Grégoire Montavon, Sebastian Lapuschkin, Alexander Binder, Wojciech Samek, and Klaus-Robert Müller. Explaining nonlinear classification decisions with deep taylor decomposition. Pattern recognition, 65:211-222, 2017. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 658, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 295, + 713 + ], + "type": "text", + "content": "[36] Grégoire Montavon, Alexander Binder, Sebastian Lapuschkin, Wojciech Samek, and Klaus-Robert Müller. Layer-wise relevance propagation: an overview. Explainable AI: interpreting, explaining and visualizing deep learning, pages 193-209, 2019. 1, 2" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 554, + 714 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 316, + 72, + 554, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 554, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 554, + 128 + ], + "type": "text", + "content": "[37] Meike Nauta, Ron van Bree, and Christin Seifert. Neural prototype trees for interpretable fine-grained image recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14933-14943, 2021. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 129, + 554, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 554, + 183 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 554, + 183 + ], + "type": "text", + "content": "[38] Laura O'Mahony, Vincent Andrearczyk, Henning Müller, and Mara Graziani. Disentangling neuron representations with concept vectors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3770-3775, 2023. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 185, + 554, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 185, + 554, + 207 + ], + "spans": [ + { + "bbox": [ + 316, + 185, + 554, + 207 + ], + "type": "text", + "content": "[39] OpenAI. Hello gpt-4o. https://openai.com/index/hellogpt-4o/, 2024.4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 209, + 554, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 209, + 554, + 274 + ], + "spans": [ + { + "bbox": [ + 316, + 209, + 554, + 274 + ], + "type": "text", + "content": "[40] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 275, + 554, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 275, + 554, + 341 + ], + "spans": [ + { + "bbox": [ + 316, + 275, + 554, + 341 + ], + "type": "text", + "content": "[41] Yao Rong, Tobias Leemann, Thai-Trang Nguyen, Lisa Fiedler, Peizhu Qian, Vaibhav Unhelkar, Tina Seidel, Gjergji Kasneci, and Enkelejda Kasneci. Towards human-centered explainable ai: A survey of user studies for model explanations. IEEE transactions on pattern analysis and machine intelligence, 46(4):2104-2122, 2024. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 342, + 554, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 342, + 554, + 386 + ], + "spans": [ + { + "bbox": [ + 316, + 342, + 554, + 386 + ], + "type": "text", + "content": "[42] Cynthia Rudin. Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead. Nature Machine Intelligence, 1(5):206-215, 2019. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 388, + 554, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 388, + 554, + 431 + ], + "spans": [ + { + "bbox": [ + 316, + 388, + 554, + 431 + ], + "type": "text", + "content": "[43] Waddah Saeed and Christian Omlin. Explainable ai (xai): A systematic meta-survey of current challenges and future opportunities. Knowledge-Based Systems, 263:110273, 2023. 1, 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 433, + 554, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 433, + 554, + 487 + ], + "spans": [ + { + "bbox": [ + 316, + 433, + 554, + 487 + ], + "type": "text", + "content": "[44] Pranab Sahoo, Ayush Kumar Singh, Sriparna Saha, Vinija Jain, Samrat Mondal, and Aman Chadha. A systematic survey of prompt engineering in large language models: Techniques and applications. arXiv preprint arXiv:2402.07927, 2024.5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 488, + 554, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 488, + 554, + 554 + ], + "spans": [ + { + "bbox": [ + 316, + 488, + 554, + 554 + ], + "type": "text", + "content": "[45] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE international conference on computer vision (ICCV), pages 618-626, 2017. 1, 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 555, + 554, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 555, + 554, + 600 + ], + "spans": [ + { + "bbox": [ + 316, + 555, + 554, + 600 + ], + "type": "text", + "content": "[46] Ao Sun, Pingchuan Ma, Yuanyuan Yuan, and Shuai Wang. Explain any concept: Segment anything meets concept-based explanation. Advances in Neural Information Processing Systems, 36, 2024. 1, 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 601, + 554, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 601, + 554, + 634 + ], + "spans": [ + { + "bbox": [ + 316, + 601, + 554, + 634 + ], + "type": "text", + "content": "[47] Giulia Vilone and Luca Longo. Notions of explainability and evaluation approaches for explainable artificial intelligence. Information Fusion, 76:89-106, 2021. 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 635, + 554, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 554, + 690 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 554, + 690 + ], + "type": "text", + "content": "[48] Kevin Ro Wang, Alexandre Variengien, Arthur Conmy, Buck Shlegeris, and Jacob Steinhardt. Interpretability in the wild: a circuit for indirect object identification in GPT-2 small. In The Eleventh International Conference on Learning Representations, ICLR 2023, Kigali, Rwanda, May 1-5, 2023. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 691, + 554, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 554, + 714 + ], + "type": "text", + "content": "[49] Teng Wang, Jinrui Zhang, Junjie Fei, Hao Zheng, Yunlong Tang, Zhe Li, Mingqi Gao, and Shanshan Zhao. Caption" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4373" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 453 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "text", + "content": "anything: Interactive image description with diverse multimodal controls. arXiv preprint arXiv:2305.02677, 2023. 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 140 + ], + "type": "text", + "content": "[50] Xin Wang, Hong Chen, Si'ao Tang, Zihao Wu, and Wenwu Zhu. Disentangled representation learning. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(12): 9677-9696, 2024. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 141, + 294, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 141, + 294, + 195 + ], + "spans": [ + { + "bbox": [ + 56, + 141, + 294, + 195 + ], + "type": "text", + "content": "[51] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824-24837, 2022. 2, 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 197, + 294, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 197, + 294, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 197, + 294, + 262 + ], + "type": "text", + "content": "[52] Hongxu Yin, Pavlo Molchanov, Jose M Alvarez, Zhizhong Li, Arun Mallya, Derek Hoiem, Niraj K Jha, and Jan Kautz. Dreaming to distill: Data-free knowledge transfer via deep-inversion. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (CVPR), pages 8715-8724, 2020. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 263, + 294, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 263, + 294, + 307 + ], + "spans": [ + { + "bbox": [ + 56, + 263, + 294, + 307 + ], + "type": "text", + "content": "[53] Wenlong Yu, Ruonan Liu, Dongyue Chen, and Qinghua Hu. Explainability enhanced object detection transformer with feature disentanglement. IEEE Transactions on Image Processing, 33:6439-6454, 2024. 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 308, + 294, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 308, + 294, + 351 + ], + "spans": [ + { + "bbox": [ + 56, + 308, + 294, + 351 + ], + "type": "text", + "content": "[54] Yu Zhang, Peter Tiño, Aleš Leonardis, and Ke Tang. A survey on neural network interpretability. IEEE Transactions on Emerging Topics in Computational Intelligence, 5(5):726-742, 2021. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 353, + 294, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 353, + 294, + 407 + ], + "spans": [ + { + "bbox": [ + 56, + 353, + 294, + 407 + ], + "type": "text", + "content": "[55] Shitian Zhao, Zhuowan Li, Yadong Lu, Alan Yuille, and Yan Wang. Causal-cog: A causal-effect look at context generation for boosting multi-modal language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13342-13351, 2024. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 409, + 294, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 409, + 294, + 453 + ], + "spans": [ + { + "bbox": [ + 56, + 409, + 294, + 453 + ], + "type": "text", + "content": "[56] Bolei Zhou, David Bau, Aude Oliva, and Antonio Torralba. Interpreting deep visual representations via network dissection. IEEE transactions on pattern analysis and machine intelligence, 41(9):2131-2145, 2018. 2, 3, 7" + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4374" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/CoLLM_ A Large Language Model for Composed Image Retrieval/02729911-7a78-4d55-af78-f50b5e45a135_content_list.json b/2025/CoLLM_ A Large Language Model for Composed Image Retrieval/02729911-7a78-4d55-af78-f50b5e45a135_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..5e30af457db7010b4cc69a28168846c535896161 --- /dev/null +++ b/2025/CoLLM_ A Large Language Model for Composed Image Retrieval/02729911-7a78-4d55-af78-f50b5e45a135_content_list.json @@ -0,0 +1,1716 @@ +[ + { + "type": "text", + "text": "CoLLM: A Large Language Model for Composed Image Retrieval", + "text_level": 1, + "bbox": [ + 163, + 130, + 833, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Chuong Huynh $^{1*}$ Jinyu Yang $^{2\\dagger}$ Ashish Tawari $^{2}$ Mubarak Shah $^{2,3}$ Son Tran $^{2}$ Raffay Hamid $^{2}$ Trishul Chilimbi $^{2}$ Abhinav Shrivastava $^{1}$", + "bbox": [ + 161, + 179, + 833, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 University of Maryland, College Park 2 Amazon", + "bbox": [ + 290, + 232, + 705, + 250 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3 Center for Research in Computer Vision, University of Central Florida", + "bbox": [ + 212, + 250, + 782, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}\\{\\mathrm{chuonghm}, \\mathrm{abhinav}\\} @ \\mathrm{cs.umd.edu}$ $^{2}\\{\\mathrm{viyjy}, \\mathrm{atawari}, \\mathrm{sontran}, \\mathrm{raffay}, \\mathrm{trishulc}\\} @ \\mathrm{amazon.com}$ shah@crcv.ucf.edu", + "bbox": [ + 161, + 271, + 828, + 304 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 339, + 326, + 353 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Composed Image Retrieval (CIR) is a complex task that aims to retrieve images based on a multimodal query. Typical training data consists of triplets containing a reference image, a textual description of desired modifications, and the target image, which are expensive and time-consuming to acquire. The scarcity of CIR datasets has led to zero-shot approaches utilizing synthetic triplets or leveraging vision-language models (VLMs) with ubiquitous web-crawled image-caption pairs. However, these methods have significant limitations: synthetic triplets suffer from limited scale, lack of diversity, and unnatural modification text, while image-caption pairs hinder joint embedding learning of the multimodal query due to the absence of triplet data. Moreover, existing approaches struggle with complex and nuanced modification texts that demand sophisticated fusion and understanding of vision and language modalities. We present CoLLM, a one-stop framework that effectively addresses these limitations. Our approach generates triplets on-the-fly from image-caption pairs, enabling supervised training without manual annotation. We leverage Large Language Models (LLMs) to generate joint embeddings of reference images and modification texts, facilitating deeper multimodal fusion. Additionally, we introduce Multi-Text CIR (MTCIR), a large-scale dataset comprising 3.4M samples, and refine existing CIR benchmarks (CIRR and Fashion-IQ) to enhance evaluation reliability. Experimental results demonstrate that CoLLM achieves state-of-the-art performance across multiple CIR benchmarks and settings. MTCIR yields competitive results, with up to $15\\%$ performance improvement. Our refined benchmarks provide more reliable evaluation metrics for CIR models, contributing to the advancement of this important field. Project page is at collm-cvpr25.github.io.", + "bbox": [ + 88, + 372, + 483, + 872 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d857becc33853ef401ca098ec992c3b7de0716d5988b6e9d34475f8aadfd2412.jpg", + "image_caption": [ + "(a) Composed Image Retrieval Example" + ], + "image_footnote": [], + "bbox": [ + 516, + 330, + 906, + 411 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/259b288b4e526d553b3518123107e72c38497de73df2a3acafdc4e85671a6249.jpg", + "image_caption": [ + "(b) Recall Sum of state-of-the-art methods on CIR benchmarks", + "Figure 1. (a) An example of CIR. (b) Recall Sum at $\\{1,10,50\\}$ for CIRR and $\\{10,50\\}$ for Fashion-IQ between CoLLM and state-of-the-art (SoTA) models under zero-shot settings. We evaluate two training scenarios: (i) without triplet data and (ii) with synthetic triplet data." + ], + "image_footnote": [], + "bbox": [ + 516, + 431, + 906, + 579 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 691, + 643, + 705 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Composed Image Retrieval (CIR) enhances traditional image retrieval [42, 62] by combining text and image queries, offering greater flexibility in search systems, with applications in e-commerce, fashion, and design [9, 12, 20, 63, 67]. As illustrated in Fig. 1a, CIR retrieves similar items, such as shirts, where the reference image provides a visual basis, and the modification text specifies desired modifications. By expanding the capabilities of conventional image or text-based searches, CIR allows for more nuanced and specific queries. However, this advanced approach also presents significant challenges compared to traditional image retrieval.", + "bbox": [ + 509, + 717, + 906, + 883 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Supervised CIR approaches face challenges in data ac", + "bbox": [ + 531, + 885, + 903, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*This work was done during internship at Amazon. † Project lead.", + "bbox": [ + 107, + 886, + 459, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "3994", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "quisition. They require high-quality CIR triplets (reference image, target image, and modification text), collected through a labor-intensive and expensive process. Consequently, existing CIR triplet datasets are limited in scale and domain coverage, restricting model generalizability. To overcome these limitations, recent CIR methods [14, 23, 28, 58, 64] have adopted zero-shot approaches [36]. These can be broadly categorized into two strategies: (i) leveraging vision-language models [29, 45, 53] (VLMs) for composed query embedding generation, and (ii) generating synthetic triplets for supervised training. VLM-based approaches utilize pre-trained models that already align vision and language features in latent space. These methods rely on large-scale image-caption pairs and follow two main directions, i.e., textual inversion [8, 15, 55] and direct interpolation [21]. Synthetic triplet generation approaches [23, 28, 64] employ Large Language Models (LLMs) [7, 57, 65] to generate modification text, with CompoDiff [14] further utilizing diffusion models [5, 46] to create synthetic reference-target image pairs. Despite the recent advances in CIR, several critical challenges persist:", + "bbox": [ + 89, + 90, + 480, + 407 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- [Data] Limitation-1: VLM-based methods may hinder efficient and effective composed query embedding learning by relying solely on image-caption pairs.", + "- [Data] Limitation-2: Existing synthetic triplet datasets often suffer from a lack of diversity and unnatural modification text. Moreover, these datasets are either very small or closed-source, impeding research progress in the field.", + "- [Model] Limitation-3: Current methods for composed query embeddings mainly use shallow transformer models or linear interpolation. While these methods are computationally efficient, they lack the ability to capture the full complexity of the composed understanding tasks.", + "- [Evaluation] Limitation-4: Existing CIR benchmarks are often compromised by noise, particularly in the form of ambiguous samples. Ambiguity arises when multiple target images can correctly match a single query, but only one is labeled as the ground truth, ignoring valid matches. Although CIRCO [4] attempts to address this issue, its efforts are limited in scale. This ambiguity in benchmarks hinders meaningful model evaluation and comparison." + ], + "bbox": [ + 89, + 412, + 482, + 715 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We propose CoLLM, an LLM-based CIR approach to address the aforementioned limitations. CoLLM tackles Limitation-1 by dynamically synthesizing triplets from image-caption pairs, introducing two key components: a reference image embedding synthesis and a modification text synthesis module. The former employs Spherical Linear Interpolation [52] (Slerp) to generate an intermediate embedding between a given image and its nearest neighbor within the training batch, serving as a synthesized reference image embedding. The latter utilizes a pre-defined text interpolation template to generate modification text based on the current caption and that of the nearest image neighbor.", + "bbox": [ + 89, + 719, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This strategy effectively leverages the vast amount of readily available image-caption pairs on the Internet, enabling a model's training in a supervised CIR manner. By doing so, CoLLM overcomes the scarcity of labeled triplet data and paves the way for more robust and scalable CIR models.", + "bbox": [ + 511, + 90, + 903, + 166 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address Limitation-2, we introduce Multi-Text CIR (MTCIR), a synthetic dataset of 3.4M image pairs with 17.7M modification texts. MTCIR focuses on two often-overlooked aspects: image diversity and naturalistic modification texts. We curated images from diverse sources to ensure variety. For modification text generation, we employ a two-stage approach using Multi-modal LLM (MLLM) [34, 35] for detailed captioning and LLM for describing interception differences. Uniquely, MTCIR provides multiple short modification texts for each image pair, covering various attributes. This better reflects human query formulation, offering a more realistic, comprehensive training foundation for CIR models.", + "bbox": [ + 511, + 167, + 903, + 361 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To overcome Limitation-3, CoLLM harnesses the power of LLMs for composed query understanding. It is motivated by the extensive world knowledge embedded in pre-trained LLMs. With their deep semantic understanding, we posit that LLMs are superior to shallow transformers and simple embedding interpolation techniques in comprehending the intricate relationships between reference images and modification texts. By leveraging LLMs, CoLLM aims to capture nuanced semantic connections, enhancing the quality and relevance of composed image retrieval results.", + "bbox": [ + 511, + 363, + 903, + 513 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address Limitation-4, we refine two popular CIR benchmarks: CIRR [37] and Fashion-IQ [61]. We use MLLMs to evaluate sample ambiguity in each benchmark and regenerate clear modification text for ambiguous ones. Our pipeline incorporates multiple validation steps to guarantee the enhanced quality of the refined samples.", + "bbox": [ + 511, + 515, + 903, + 604 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our main contributions can be summarized as: (i) We propose a method to synthesize CIR triplets on-the-fly from image-caption pairs, which can even outperform models trained on CIR triplets, eliminating the need for costly CIR datasets. (ii) We collect and will release MTCIR, a new CIR triplet dataset covering 3.4 million image pairs with 17.7 million modification texts. To our knowledge, MTCIR is the largest open-source synthetic CIR dataset. (iii) We introduce an approach to leverage LLMs for composed query understanding, utilizing their instruction-following and embedding generation capabilities. (iv) We refine CIRR and Fashion-IQ to provide more robust evaluation benchmarks for the CIR community. Extensive experiments on popular CIR benchmarks and settings demonstrate the effectiveness of our model innovations and new datasets (Fig. 1b).", + "bbox": [ + 511, + 606, + 903, + 832 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 513, + 840, + 663, + 857 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Composed Image Retrieval. Composed Image Retrieval (CIR) has garnered significant attention due to its flexibil", + "bbox": [ + 511, + 869, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "3995", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ity in search systems [37, 43, 48]. Zero-shot CIR methods have been extensively explored, with textual inversion [4, 8, 15, 49, 55] emerging as a prominent technique. This approach maps image encoder outputs to text encoder inputs, creating pseudo-word tokens. SlerpTAT [21] emphasizes the text encoder's importance by re-aligning visual and textual embeddings. While most zero-shot CIR models utilize image-caption pairs for training, our work introduces an innovative method to synthesize triplets from these pairs during training. This method enables supervised CIR style training, enhancing the model's understanding of composed queries without relying on real CIR datasets.", + "bbox": [ + 89, + 90, + 480, + 271 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recent works enhance enhancing models' language understanding using Large Language Models (LLMs) [26, 30], primarily leveraging their text generation abilities. MCL [30] composes image-text query in the LLM input space and aligns joint embedding with the target image caption. However, this may introduce noise due to limited visual information in the target image captions. CIReVL [26] employs LLMs to generate captions based on the reference image and modification text, subsequently retrieving the target image using a text-to-image retrieval approach. Our model also utilizes LLMs but differs by directly producing a composed query embedding for target image retrieval, potentially reducing intermediate steps and associated errors.", + "bbox": [ + 89, + 273, + 482, + 470 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "CIR with Synthetic Datasets. The scarcity of supervised CIR datasets has prompted recent studies to leverage generative models for synthetic triplet creation. CompoDiff [14] employs image editing pipelines [5, 46] to generate target images, though their nature limits performance. Other approaches [23, 28, 58, 59, 64] utilize LLMs to generate modification text for real image pairs. Our work introduces a novel two-stage approach, combining MLLMs [34, 35] for detailed captioning and LLMs [1, 2] for describing interception differences. Distinctly, our synthetic triplets provide multiple concise modification texts for each image pair, encompassing various attributes, in contrast to the conventional single, lengthy modification text. This enables a more nuanced representation of image modifications, potentially improving CIR model performance and versatility.", + "bbox": [ + 89, + 473, + 482, + 700 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Large Language Models. Recent advancements in LLMs have expanded their applications beyond text generation to include image understanding [34, 35] and text embedding generation [40, 60, 68]. Large Language Embedding Models (LLEMs) are specialized versions trained using contrastive learning, leveraging LLM knowledge for embedding generation while often disabling text generation capabilities. To enhance text embedding quality, [27, 31, 39] propose removing causal attention in LLMs, thereby improving text information encoding efficiency and training these modified LLMs on text retrieval datasets. This text retrieval capability has been further extended to text-image retrieval by aligning visual and LLM text embed", + "bbox": [ + 89, + 704, + 482, + 902 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/9473b742177963d4ac05ea3ed412ecb46f2b36bba425554653d14f6327a0d981.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 539, + 89, + 875, + 212 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/77fbe85028cc23d395fcb33d1ee4270f55402926a2d57f1a408f56357b5a7967.jpg", + "image_caption": [ + "Figure 2. An overview of our model and training strategies when using (a) image-caption pairs and (b) CIR triplets." + ], + "image_footnote": [], + "bbox": [ + 537, + 217, + 880, + 305 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ding spaces [22, 25, 38]. These developments demonstrate the potential of LLMs for composed query understanding in CIR tasks. Our work builds upon this foundation, extending LLMs/LLEMs to generate composed query embeddings by incorporating reference images and modification text. This novel approach leverages the multimodal capabilities of LLMs to enhance the performance of CIR systems, leading to more accurate and context-aware image retrieval.", + "bbox": [ + 511, + 348, + 906, + 470 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 513, + 484, + 604, + 500 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This section outlines our methodology, starting with a brief overview of the model architecture, followed by a formal CIR problem definition. We then detail our triplet synthesis strategy, including reference image embedding and modification text synthesis, using image-caption pairs. Lastly, we describe our LLM-based query composition approach.", + "bbox": [ + 511, + 510, + 905, + 602 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Model Architecture", + "text_level": 1, + "bbox": [ + 511, + 604, + 700, + 619 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As illustrated in Fig. 2, CoLLM consists of several essential components: (1) a vision encoder $f(\\cdot)$ for extracting image features; (2) modules for synthesizing reference image embeddings and modification text; (3) an image adapter $g(\\cdot)$ that maps visual features into the language model's semantic space; (4) a LLM $\\Phi(\\cdot)$ that processes multimodal queries and (5) a projection layer $proj(\\cdot)$ (omitted from the figure for simplicity) that maps the LLM output to a suitable representation for retrieval. It is important to note that Fig. 2 (a) and Fig. 2 (b) illustrate architectures designed for two distinct input formats. The former is tailored for image-caption pairs, while the latter is optimized for CIR triplets.", + "bbox": [ + 511, + 627, + 906, + 809 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. CoLLM with Image-Caption Pairs as Input", + "text_level": 1, + "bbox": [ + 511, + 818, + 883, + 835 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Let $\\mathcal{X} = \\{(v_i, w_i)\\}_{i=1}^N$ denote a set of image-caption pairs, where $v_i$ and $w_i$ represent the image and caption of the $i^{th}$ sample, respectively. To effectively leverage the vast amount of image-caption pairs, we introduce a novel ap", + "bbox": [ + 511, + 839, + 905, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3996", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/9e6e609e23dbeec35a20b5cef8f698bc61dccdd2111547532ea31992ccd53e63.jpg", + "image_caption": [ + "Figure 3. An overview of reference image embedding synthesis and modification text synthesis. The red-framed image represents the nearest neighbor of the augmented image in the training batch." + ], + "image_footnote": [], + "bbox": [ + 94, + 88, + 480, + 247 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "proach that synthesizes a CIR triplet on-the-fly for each image-caption pair $(v_{i},w_{i})$ . In this process, $v_{i}$ serves as the target image, while two distinct modules generate the remaining components: (i) a reference image embedding synthesizer and (ii) a modification text synthesizer (Fig. 2a). Notably, we do not generate an actual image for reference image synthesis. Instead, we synthesize the reference image embedding, which is computationally more efficient and allows seamless integration into the CIR pipeline. This formulation enables CoLLM to exploit the rich information in image-caption pairs, transforming them into CIR triplets that can be used for training. The following subsections detail the specific mechanisms for synthesizing the reference image embedding and the modification text.", + "bbox": [ + 88, + 309, + 483, + 521 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Reference Image Embedding Synthesis. Our reference image embedding synthesis process is illustrated in Fig. 3 (left). We begin by applying an augmentation $t \\sim \\mathcal{T}$ to the input image $v_{i}$ , where $\\mathcal{T}$ represents a family of augmentations. The resulting augmented image is denoted as $v_{i}^{\\prime}$ with its corresponding embedding $\\mathbf{h}_i^\\prime = f(v_i^\\prime)$ . For $\\mathbf{h}_i^\\prime$ , we identify its in-batch nearest neighbor $v_{j}^{\\prime}$ with $j$ defined as:", + "bbox": [ + 89, + 523, + 483, + 632 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nj = \\underset {j \\neq i; j \\in \\{1, \\dots \\mathcal {B} \\}} {\\arg \\max } \\operatorname {s i m} \\left(\\mathbf {h} _ {i} ^ {\\prime}, \\mathbf {h} _ {j} ^ {\\prime}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 181, + 640, + 482, + 669 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{B}$ is the batch size, $\\mathbf{h}_j^{\\prime} = f(v_j^{\\prime})$ , and $\\mathrm{sim}(\\mathbf{u},\\mathbf{v}) = \\mathbf{u}^T\\mathbf{v} / (\\| \\mathbf{u}\\| \\| \\mathbf{v}\\|)$ denotes cosine similarity.", + "bbox": [ + 89, + 676, + 482, + 710 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We then employ Spherical Linear Interpolation (Slerp) [52] to synthesize the reference image embedding by interpolating between $\\mathbf{h}_i^{\\prime}$ and $\\mathbf{h}_j^{\\prime}$ . The synthesized reference image embedding is obtained as:", + "bbox": [ + 89, + 710, + 483, + 771 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\theta = \\operatorname {a r c c o s} \\left(\\mathbf {h} _ {i} ^ {\\prime} \\cdot \\mathbf {h} _ {j} ^ {\\prime}\\right)\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 780, + 303, + 797 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {h} _ {i} ^ {*} = \\frac {\\sin (\\alpha \\theta)}{\\sin \\theta} \\mathbf {h} _ {i} ^ {\\prime} + \\frac {\\sin ((1 - \\alpha) \\theta)}{\\sin \\theta} \\mathbf {h} _ {j} ^ {\\prime}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 158, + 799, + 482, + 832 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\alpha \\in [0,1]$ is a hyper-parameter controlling the interpolation strength. The intuition behind this approach is that an image with embedding $\\mathbf{h}_i^*$ shares certain visual similarities with the target image $v_{i}$ . Specifically, a larger $\\alpha$ results", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "in $\\mathbf{h}_i^*$ more closely resembling $\\mathbf{h}_i'$ , allowing for fine-grained control over the synthesized embedding.", + "bbox": [ + 511, + 90, + 903, + 121 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Modification Text Synthesis. For the augmented image $v_{i}^{\\prime}$ , its nearest neighbor is $(v_{j}^{\\prime}, w_{j})$ . To generate the modification text $w_{i}^{*}$ , we use text interpolation to combine $w_{i}$ and $w_{j}$ using pre-defined templates, as illustrated in Fig. 3 (right). The complete set of templates is provided in the supplementary material. During training, we randomly select a template for each sample to ensure diversity in the synthesized modification text. It is worth noting that we do not simply use $w_{i}$ as the modification text for two primary reasons: (i) $w_{i}$ alone fails to capture the visual differences between the reference image and the target image. (ii) Using only $w_{i}$ could lead to model cheating, where it learns to rely solely on $w_{i}$ for retrieval while ignoring the reference image. Our synthesis approach, in contrast, mimics real-world modification text, which often describes both similarities and differences between the reference and target images. Furthermore, this strategy forces the model to learn composed query embeddings by considering both the reference image and the modification text simultaneously.", + "bbox": [ + 511, + 125, + 905, + 411 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "By combining the synthesized reference embedding $\\mathbf{h}_i^*$ and the interpolated modification text $w_{i}^{*}$ , our method generates diverse CIR triplets from image-caption pairs. This approach enables effective training of the CoLLM model on large-scale image-caption datasets, effectively mitigating the dependency on scarce labeled triplet data.", + "bbox": [ + 511, + 411, + 903, + 502 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Query Composition with LLM. To construct the composed query, we utilize a pre-trained LLM that processes the synthesized reference image embedding $\\mathbf{h}_i^*$ , image caption $w_i$ , and the synthesized modification text $w_i^*$ . We define three distinct composed embeddings:", + "bbox": [ + 511, + 505, + 905, + 580 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {c} _ {i} ^ {v} = p \\left(\\Phi \\left(g \\left(\\mathbf {h} _ {i} ^ {*}\\right)\\right)\\right) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 627, + 587, + 903, + 604 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {c} _ {i} ^ {w} = p \\left(\\Phi \\left(w _ {i}\\right)\\right) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 625, + 606, + 903, + 623 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {c} _ {i} = p \\left(\\Phi \\left([ g \\left(\\mathbf {h} _ {i} ^ {*}\\right); w _ {i} ^ {*} ]\\right)\\right) \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 630, + 625, + 903, + 642 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $[\\cdot ]$ denotes an instruction template that combine two modalities (see supplementary material).", + "bbox": [ + 511, + 647, + 903, + 678 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Training Objective. We employ a contrastive loss $\\mathcal{L}_{cl}$ [44], consistent with previous works [56, 59], to compute the loss between query embeddings $\\{\\mathbf{c}_i^v,\\mathbf{c}_i^w,\\mathbf{c}_i\\}$ and the target image embedding $\\mathbf{z}_i = f(v_i)$ . The final loss for each sample during pre-training is defined as:", + "bbox": [ + 511, + 681, + 905, + 757 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\frac {1}{3} \\left(\\mathcal {L} _ {c l} \\left(\\mathbf {c} _ {i} ^ {v}, \\mathbf {z} _ {i}\\right) + \\mathcal {L} _ {c l} \\left(\\mathbf {c} _ {i} ^ {w}, \\mathbf {z} _ {i}\\right) + \\mathcal {L} _ {c l} \\left(\\mathbf {c} _ {i}, \\mathbf {z} _ {i}\\right)\\right) \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 537, + 762, + 903, + 790 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This formulation encourages the model to learn discriminative representations for different query types while maintaining consistency with the target image embedding. By combining losses from image-to-image $(\\mathbf{c}_i^v)$ , text-to-image $(\\mathbf{c}_i^w)$ , and composed $(\\mathbf{c}_i)$ queries, we ensure that the model learns to effectively process and align various input modalities for compositional image retrieval.", + "bbox": [ + 511, + 794, + 905, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "3997", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/14f36abf0ea4db2e3ced31de7aed85e62ff4976fca677c3fe01b3ec8fe140e7e.jpg", + "table_caption": [ + "Table 1. Comparison of synthetic CIR training datasets." + ], + "table_footnote": [], + "table_body": "
DatasetPublicReference-Target Image PairModification Text Generation
TypeSource# Pairs# EntitiesLLMLLM input# Text# Text/Pair
LaSCo [28]✓imageVQAv2 [13]360K82KGPT-3 [6]question-answer360K1
VDG [23]✓imageNLVR2 [54], COCO [32]467K183KVision-LLaMA2image pair523K1.12
WebCoVR [59]✓videoWebVid2M [3]1.6M131KCustom LLMcaption1.6M1
CC-CoIR [58]✓imageCC3M [51]3.3M357KCustom LLMcaption3.3M1
SynthTriplets [14]synth. imageSD [46], IP2P [5]18.8M37.6MOPT-6.7B [65]-18.8M1
MagicLens [64]imageWeb crawled36.5M-PaLM [7]image metadata36.5M1
MTCIR (Ours)✓imageLLaVA-558k [34]3.4M423KSonnet 3 [2]synth. caption17.7M5.18
", + "bbox": [ + 104, + 88, + 890, + 247 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/393b355a1fc2f9cf39fd0ea505f05b23f01f7cacb1e97084e954d00c07e6619a.jpg", + "image_caption": [ + "Reference Image", + "Category", + "Modification Text" + ], + "image_footnote": [ + "+ Object Added", + "Object Removed", + "Add a person's hand holding the cup.", + "Modify the liquid being poured.", + "Remove the social media text overlay." + ], + "bbox": [ + 120, + 263, + 282, + 337 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/dd58bd807d019a2cdf28e643f4181a82c4c754ec2d6f6cd1f3e93c9442c9b365.jpg", + "image_caption": [ + "Target Image" + ], + "image_footnote": [ + "$\\mathbf{3}$ Attribute Changed" + ], + "bbox": [ + 285, + 263, + 450, + 337 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. CoLLM with CIR Triplet as Input", + "text_level": 1, + "bbox": [ + 89, + 422, + 390, + 439 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For CIR triplet inputs, our model design is shown in Fig. 2b. Let $\\mathcal{X} = \\{(v_i^r,v_i^t,w_i)\\}_{i = 1}^N$ be a set of triplets, where $v_{i}^{r}$ $v_{i}^{t}$ , and $w_{i}$ represent the reference image, target image, and modification text of the $i^{th}$ sample, respectively. The reference embedding is obtained as $\\mathbf{h}_i = f(v_i^r)$ . The composed embedding is then computed as $\\mathbf{c}_i = p(\\Phi ([g(\\mathbf{h}_i);w_i])$ . The training objective is $\\mathcal{L} = \\mathcal{L}_{cl}(\\mathbf{c}_i,\\mathbf{z}_i)$ , where $\\mathbf{z}_i = f(v_i^t)$ denotes the target embedding.", + "bbox": [ + 89, + 444, + 483, + 566 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Dataset Construction", + "text_level": 1, + "bbox": [ + 89, + 573, + 294, + 588 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This section outlines the data construction process for two primary objectives: (1) creating a novel CIR training dataset called Multi-Text CIR (MTCIR) and (2) refining widely used CIR benchmarks, specifically CIRR [37] and FashionIQ [61]. We detail the methodologies for each task, highlighting the improvements and innovations.", + "bbox": [ + 89, + 597, + 483, + 688 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Multi-Text CIR (MTCIR) Dataset", + "text_level": 1, + "bbox": [ + 89, + 696, + 387, + 713 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Existing CIR triplet datasets [23, 28, 59] lack diversity and contain unnatural modification text. We introduce Multi-Text CIR (MTCIR) to address these limitations.", + "bbox": [ + 89, + 719, + 482, + 763 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To enhance image diversity, we source images from the LLaVA-558k dataset [34] filtered from image caption datasets [41, 50, 51] based on noun-phrase frequency for broad concept coverage. We pair images using CLIP [45] visual similarity metrics, following CIRR's grouping approach, yielding 3.4M pairs from approximately 423K images. Leveraging the detailed captions generated by LLaVA-Next-34B [35] for each image in the LLaVA-558k dataset, we employ Claude 3 Sonnet [2]", + "bbox": [ + 89, + 763, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/e47c0a1e12446a6ddd82a5b20250169c508796a04007e3a4b35f9b13485c9e44.jpg", + "image_caption": [ + "Unchanged", + "(a) CIRR" + ], + "image_footnote": [], + "bbox": [ + 519, + 271, + 611, + 347 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/2c16b8e1361659c01de2b3d8d7b5e1c425d74ed5dde92b11fd7245a35a6c2859.jpg", + "image_caption": [ + "Regenerate", + "(b) FIQ-Dress" + ], + "image_footnote": [], + "bbox": [ + 616, + 271, + 707, + 347 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/5b04838b1d5c159594bbb52c8b859c4a3295a1644a686d8a85eea5d0cea863cc.jpg", + "image_caption": [ + "Remove", + "(c) FIQ-Shirt", + "Cannot process" + ], + "image_footnote": [], + "bbox": [ + 712, + 271, + 803, + 347 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/3c3f65416323df20ea4b3345010a918dc2bc39886d72cdee015d046a611b79e9.jpg", + "image_caption": [ + "(d) FIQ-Toptee" + ], + "image_footnote": [], + "bbox": [ + 808, + 271, + 900, + 347 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/44b80bdf9e6caa3bf59217f12cc9164594365c38dbcb6efe490fd782bbfb0bb2.jpg", + "image_caption": [ + "Figure 4. An example from the MTCIR dataset. Each sample contains multiple short texts describing different modifications.", + "Figure 5. Statistical analysis of the refinement process for CIRR and FashionIQ (FIQ) datasets.", + "Reference Image", + "Original Text", + "Regenerated Text" + ], + "image_footnote": [], + "bbox": [ + 545, + 421, + 653, + 483 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/cd682d24ce9cdbb9b9df1a60dadafb80979bc461311654ac5467f738a2f8fbf1.jpg", + "image_caption": [ + "Target Image", + "One puppy looking at the camera." + ], + "image_footnote": [ + "The dog is indoors on a tiled floor with a", + "food bowl nearby." + ], + "bbox": [ + 655, + 421, + 763, + 483 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Figure 6. An example from the refined CIRR dataset. The original modification text yields multiple plausible results (shown as similar images to the target on the right). In contrast, the newly generated text accurately captures the distinctive features of the target image, differentiating it from other candidates.", + "bbox": [ + 511, + 521, + 905, + 590 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "(20240229-v1:0) to generate modification text for each image pair, using these captions as input. To control detail levels in Claude outputs, we define six modification text categories in the prompt, inspired by prior research [4, 37]. This approach maintains diverse aspects and prevents repetition. We also incorporate examples from the CIRR dataset to ensure human-like modification text. Detailed prompting strategies are provided in the supplementary material.", + "bbox": [ + 511, + 598, + 905, + 718 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our method generates multiple brief, focused texts for each image pair, each highlighting a unique aspect. This approach captures all changes between two images without producing overly long sentences, better aligning with human-written modification text. These short texts can be combined for training, enhancing dataset diversity. We obtained 17.7M text samples for the 3.4M image pairs, averaging 5.18 short sentences per pair. Fig. 4 illustrates an example, showcasing modification texts corresponding to specific categories and covering diverse attributes.", + "bbox": [ + 511, + 719, + 905, + 869 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As shown in Table 1, MTCIR offers several advantages over existing synthetic CIR datasets: (i) Unlike [14], which", + "bbox": [ + 511, + 869, + 905, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "3998", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/6ec923a0d09b6a0f170e1e52cf64dddf64c8a1fdd38165c4bca89faa45a9246a.jpg", + "table_caption": [ + "Table 2. Performance comparison of various models and visual encoders on CIR benchmarks, evaluated without training on triplet datasets. Bold and underlined values denote the best and second-best scores within each vision encoder group. Models incorporating LLMs in their architectures are marked with $\\star$ . Results reproduced by our team are indicated by $\\ddagger$ ." + ], + "table_footnote": [], + "table_body": "
MethodCIRCO (mAP↑)CIRR (Rec.↑)FIQ (Rec.↑)
@5@10@50@1@10@50@10@50
OpenAI CLIP-B/32
PALAVRA [8]4.65.36.816.658.584.019.837.3
SEARLE [4]9.49.911.824.066.889.822.942.5
Slerp-TAT [21]9.310.312.328.268.888.523.044.0
CIREVL* [26]14.915.417.823.966.087.028.349.4
CoLLM*12.913.215.028.671.892.724.845.2
OpenAI CLIP-L/14
Pic2World [49]8.79.511.323.965.387.824.743.7
SEARLE [4]11.712.715.124.266.388.825.646.2
LinCIR‡[15]13.013.916.224.666.988.826.446.6
ContextI2W [55]13.013.816.025.768.689.627.848.9
MCL* [30]17.818.421.825.969.491.0--
Slerp-TAT [21]18.519.421.430.970.989.228.347.6
CIREVL* [26]18.619.021.824.664.986.328.648.6
CoLLM*20.320.823.429.772.891.530.149.5
BLIP-L/16
Slerp-TAT [21]17.818.421.134.072.788.932.853.3
CoLLM*19.720.423.135.078.694.234.656.0
", + "bbox": [ + 91, + 157, + 486, + 441 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "uses synthetic images, MTCIR uses real images covering diverse concepts and domains. (ii) MTCIR provides multiple modification texts for each image pair, offering more natural descriptions and training flexibility. (iii) MTCIR is the largest public dataset to advance CIR research.", + "bbox": [ + 89, + 452, + 482, + 527 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To prevent the leakage of biometric information, we further process MTCIR at both the image and text levels. For images, we apply face blurring to all instances. For text, we remove any content containing keywords related to human attributes (e.g., skin, hair, gender, age, race).", + "bbox": [ + 89, + 529, + 483, + 604 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Refined CIRR and Fashion-IQ Datasets", + "text_level": 1, + "bbox": [ + 89, + 619, + 431, + 635 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Current CIR benchmarks often suffer from label ambiguity, where modification text lacks clarity, potentially matching multiple target images. To address this issue, we propose a refinement method using Claude 3 Sonnet [2] (20240229-v1:0) to enhance the validation sets of two well-known benchmarks: CIRR and Fashion-IQ. Each sample in these validation sets comprises a triplet containing a reference image, target image, and modification text.", + "bbox": [ + 89, + 642, + 482, + 763 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our refinement process, applicable to any benchmark, consists of validation, re-generation, and re-validation steps. For each triplet, we first identify the top-3 visually similar images to the target image based on CLIP visual scores, serving as hard negative samples for ambiguity assessment. We employ Claude 3 Sonnet to evaluate each triplet's ambiguity by attempting to identify the target image from these hard negatives. Triplets correctly identified by Claude 3 Sonnet are considered \"good\" samples and re", + "bbox": [ + 89, + 765, + 482, + 901 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/5e988e21aa79d61f3644d8028c86e95cbba952559c19b07b5190623fdf7bbd25.jpg", + "table_caption": [ + "Table 3. Performance of models trained on synthetic triplet datasets. Bold indicates the best method overall, while underline highlights the best method within the same vision encoder. Our CoLLM outperforms comparable methods in most metrics." + ], + "table_footnote": [], + "table_body": "
MethodDatasetCIRR ↑FIQ ↑
@1@10@50@10@50
CoCa-L/18 288 × 288
MagicLens [64]MagicLens [64]33.377.994.438.158.3
EVA-CLIP ViT-G/14 364 × 364
CoVR2 [58]WV-CC-CoVIR [58]43.784.096.138.258.4
OpenAI CLIP-L/14 224 × 224
CompoDiff [14]SynTrip18M [14]18.270.890.336.048.6
MagicLens [64]MagicLens [64]30.174.492.630.752.5
CoLLMMTCIR (ours)34.777.093.132.954.2
BLIP-L/16 384 × 384; fine-tuned on COCO
CASE [28]LaSCo [28]35.478.594.6--
Omkar et al. [56]WebCoVR [59]40.178.994.730.346.5
CoLLMMTCIR (ours)45.884.795.839.160.7
", + "bbox": [ + 511, + 128, + 906, + 328 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "main unchanged. For \"bad\" samples, we use Claude 3 Sonnet to re-generate the modification text. The model produces three new modification texts with increasing detail in a hierarchical structure, with each finer text building upon the coarser one. We then repeat the validation process, evaluating the new texts from coarsest to finest and selecting the best one. Any re-generated modification text that passes the validation process replaces the original text and concludes the process. If all generated texts fail the validation, the triplet is removed from the benchmark. Detailed process information is provided in the supplementary material. Besides, following MTCIR, biometric information is also removed from the refined benchmarks.", + "bbox": [ + 511, + 335, + 906, + 532 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As illustrated in Fig. 5, up to $8\\%$ of samples are omitted due to Claude's processing limitations related to harmful content. Additionally, $4 - 7\\%$ of ambiguous samples that could not be effectively rewritten are removed. More than $40\\%$ of the original samples are retained, and $40 - 55\\%$ of triplets are successfully revised. The refined CIRR and Fashion-IQ datasets exhibit less ambiguity (Fig. 6), offering more robust evaluation benchmarks for the CIR community.", + "bbox": [ + 511, + 534, + 908, + 655 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 511, + 670, + 648, + 688 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our proposed CoLLM framework adopts a two-stage training paradigm: pre-training and fine-tuning phases. In pretraining, CoLLM is exposed to a diverse corpus of image-capture pairs, enabling it to understand composed queries comprehensively. Subsequently, in the fine-tuning stage, we leverage our newly curated MTCIR datasets to verify their effective performance on complex multimodal tasks.", + "bbox": [ + 511, + 696, + 906, + 801 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Pre-Training on Image-Caption Pairs", + "text_level": 1, + "bbox": [ + 511, + 814, + 836, + 830 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Training and Evaluation Datasets. The pre-training dataset consists of 5 million image-caption pairs, compiled from CC3M [51], LAION [50], and LLaVA-558K [34]. We evaluate our model on three CIR benchmarks: CIRCO [4],", + "bbox": [ + 511, + 839, + 906, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "3999", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/2d9db65b9e6b6723183423a9e9bc44b7d5ceeaa209a97b81c301686a737d9379.jpg", + "table_caption": [ + "Table 4. Performance of BLIP-L and CoLLM (pre-trained) trained on synthetic triplet datasets. Bold values indicate the best score within each group. Models trained on MTCIR demonstrate superior performance compared to those trained on other datasets." + ], + "table_footnote": [], + "table_body": "
MethodDatasetCIRR ↑FIQ ↑
@1@10@50@10@50
BLIP-L [29]LaSCo [28]36.678.494.624.844.0
WebCoVR [59]39.378.994.726.743.3
MTCIR (ours)42.483.195.937.959.2
CoLLMLaSCo [28]43.284.195.738.560.1
MTCIR (ours)45.884.795.939.160.7
", + "bbox": [ + 91, + 128, + 483, + 231 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "CIRR [37] test set, and Fashion-IQ [61] validation set. All datasets are filtered to exclude biometric information.", + "bbox": [ + 89, + 236, + 482, + 266 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Evaluation Metrics. We employ recall on the complete image index set as the evaluation metric for CIRR and Fashion-IQ. For CIRCO, we use mean Average Precision (mAP). We exclude the recall on subset for CIRR due to reliability concerns (see supplementary material for details).", + "bbox": [ + 89, + 270, + 483, + 347 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Implementation Details. See supplementary material.", + "bbox": [ + 89, + 349, + 454, + 364 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Quantitative Results. Table 2 compares our pre-trained model with other methods not trained on triplet datasets. We evaluate three variants of our architecture, each employing different vision encoders. CoLLM with the CLIP-B encoder achieves the second-best performance on the CIRCO and Fashion-IQ benchmarks while significantly improving on the CIRR benchmark. Although CIREVL [26] outperforms other methods, its use of the closed-source GPT-4 [1] introduces additional costs and increased inference latency, making it an unfair comparison.", + "bbox": [ + 88, + 368, + 482, + 518 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For larger vision encoders, our model consistently outperforms both non-LLM (e.g., Slerp-TAT [21]) and LLM (CIReVL and MCL [30]) methods across all benchmarks. Notably, our CLIP-L variant achieves state-of-the-art results on the CIRCO benchmark, while BLIP-L significantly improves CIRR and Fashion-IQ scores. The consistent gains across benchmarks and architectures validate the robustness and generalizability of our triplet synthesis approach. By effectively bridging the gap between abundant image-caption data and scarce triplet annotations, our method paves the way for more scalable and efficient training of CIR models. This innovative strategy represents a significant step forward in addressing one of the critical challenges in the field, Paving the way for larger-scale pre-training and more complex multimodal interactions.", + "bbox": [ + 89, + 518, + 482, + 744 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Fine-tuning on MTCIR", + "text_level": 1, + "bbox": [ + 89, + 753, + 307, + 768 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Training Dataset and Benchmarks. Despite the impressive performance achieved using only image-caption datasets, we hypothesize that CoLLM's performance can be further enhanced through fine-tuning on our newly developed MTCIR dataset. For a fair comparison, we also train CoLLM and baseline models on other public datasets, including WebCoVR [59] and LaSCo [28]. Our evaluation focuses on the CIRR [37] test and Fashion-IQ [61] validation.", + "bbox": [ + 89, + 779, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/22f3f52a74dcbbdbddbd53a388bcb2d6d1ee94072de2517a296f2c1274e7e486.jpg", + "table_caption": [ + "Table 5. Performance of models on refined CIRR and Fashion-IQ validation sets. Bold indicates the highest score, while underlined values represent the best metric within the same vision encoder group. CoLLM+MTCIR continues to outperform other configurations on these new benchmarks." + ], + "table_footnote": [], + "table_body": "
MethodDatasetRef. CIRR ↑Ref. FIQ ↑
@1@10@50@10@50
EVA-CLIP ViT-G/14 364 × 364
CoVR2 [58]WV-CC-VIR [58]56.191.197.954.272.9
OpenAI CLIP-L/16 224 × 224
MagicLens [64]MagicLens [64]42.386.397.045.568.1
CoLLMMTCIR (ours)46.587.496.148.368.6
BLIP-L/16 384 × 384; fine-tuned on COCO
BLIP-L [29]MTCIR (ours)53.889.997.754.874.3
CoLLMLaSCo [28]57.392.098.156.975.9
CoLLMMTCIR (ours)60.492.698.257.276.4
", + "bbox": [ + 511, + 142, + 911, + 301 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "tion sets. We exclude CIRCO [4] due to potential domain overlap between COCO [32] images used in CIRCO and those in LaSCo [28], as well as in the pre-training datasets.", + "bbox": [ + 511, + 306, + 906, + 352 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Implementation Details. See supplementary material.", + "bbox": [ + 511, + 354, + 877, + 371 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Quantitative Results. Table 3 presents the evaluation of models trained on publicly available synthetic CIR datasets, including our MTCIR. Fine-tuning CoLLM on MTCIR yields significant gains, improving CIRR@1 by $5\\%$ over CoLLM-OpenAI-CLIP-L/14 in Table 2. This improvement reflects MTCIR's enhanced generalizability and our superior data construction. Our method achieves exceptional results on all benchmarks using the same vision encoder compared to previous approaches and models trained on other synthetic datasets. Notably, our fine-tuned CoLLM outperforms MagicLens CoCa-L [64] and CoVR2 [58], despite these models using larger, more advanced vision encoders and being trained on larger datasets. These results underscore the importance of data quality over quantity.", + "bbox": [ + 511, + 375, + 906, + 585 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "MTCIR vs. Other CIR Datasets. Despite CoLLM's state-of-the-art performance, two questions remain: (i) Can MTCIR benefit other models? (ii) Can CoLLM achieve good performance when fine-tuned on other CIR datasets? To address these questions, we train a baseline BLIP-L [29] model on various CIR datasets, including MTCIR. Both LaSCo and MTCIR experiments are trained for one epoch. As shown in Table 4, we observe that: (i) BLIP-L shows significant improvement when using MTCIR as the training data, and (ii) CoLLM exhibits notable performance degradation when fine-tuned on the LaSCo dataset. These results demonstrate MTCIR's versatility as a plug-and-play dataset, capable of enhancing performance across various models. This underscores MTCIR's potential to make significant contributions to the CIR community.", + "bbox": [ + 511, + 588, + 908, + 816 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Re-evaluation on Refined Benchmarks", + "text_level": 1, + "bbox": [ + 511, + 818, + 844, + 832 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Ambiguity in the CIRR and Fashion-IQ benchmarks potentially skews the metrics presented in previous tables, particularly for high-performing models. While CIRCO mitigates this issue, it still suffers from domain overlap with train-", + "bbox": [ + 511, + 839, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "4000", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/b61da98592cbade6c6b29fa358a6bab67ac8a5c70e157dcfb4369346c347eb69.jpg", + "table_caption": [ + "Table 6. Ablation studies of proposed components.", + "(a) Reference image and modification text interpolation." + ], + "table_footnote": [], + "table_body": "
ImageTextCIRCO (mAP ↑)CIRR (Rec. ↑)
14.7170.1
✓14.5176.4
✓39.6183.3
✓✓52.8194.5
", + "bbox": [ + 120, + 103, + 452, + 172 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/99c752ed9d3bd7ad93bea227978536c61818501ade65d006591627906c2a00e5.jpg", + "table_caption": [ + "(b) Unimodal queries in Eq. (3) and Eq. (4)." + ], + "table_footnote": [], + "table_body": "
Image-onlyText-onlyCIRCO (mAP ↑)CIRR (Rec. ↑)
✓14.599.7
✓47.1196.3
✓✓52.8194.5
", + "bbox": [ + 120, + 195, + 452, + 251 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/19fe09da6cc5e5f5999a410fb5a7d41913b453c5e37ebd6401c3f79800c895c8.jpg", + "table_caption": [ + "(c) Reference image embedding interpolation." + ], + "table_footnote": [], + "table_body": "
CIRCO (mAP ↑)CIRR (Rec. ↑)
Random In-Batch Sample46.7182.2
Nearest In-Batch Neighbor52.8194.5
", + "bbox": [ + 117, + 273, + 449, + 321 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ing images. To assess the impact of ambiguous samples, we re-evaluate the models on our newly refined CIRR and Fashion-IQ benchmarks. Table 5 presents the performance of these models on the revised benchmarks.", + "bbox": [ + 89, + 329, + 480, + 388 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our analysis shows that model rankings remain largely consistent before and after the benchmark refinement. As the modification texts in the refined benchmarks provide additional nuance, we anticipated that models with more robust composed query understanding capabilities would distinguish themselves. Indeed, CoLLM consistently outperforms other models across various settings on the refined benchmarks. Notably, while BLIP-L (MTCIR) trailed CoLLM (MTCIR) by $3.4\\%$ (CIRR@1)(Table 4), this performance gap widened to $6.9\\%$ (CIRR@1) on the refined CIRR benchmark. This suggests CoLLM better captures nuanced modification text and distinguishes visual differences between source and target images.", + "bbox": [ + 89, + 390, + 482, + 587 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.4. Ablation Studies", + "text_level": 1, + "bbox": [ + 89, + 590, + 254, + 604 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We conduct comprehensive ablation studies to evaluate the impact of our proposed components during the pre-training stage. Our experiments use the BLIP-L encoder, trained exclusively on the LLaVA-558k [34] dataset for one epoch. Table 6 presents the performance across various settings. We report the sum of mAP at $k = \\{5,10,25,50\\}$ for CIRCO and Recall@ $k = \\{1,5,10,50\\}$ for CIRR. Image and Text Interpolation are Crucial: As shown in Table 6a, employing Slerp for reference image embedding significantly enhances the model's learning capabilities. Applying interpolation on modification text further improves performance. Unimodal Queries are Beneficial: We investigate the necessity of unimodal queries in Eq. (3) and Eq. (4) and observe that omitting unimodal queries during training, especially text-only queries, substantially degrades performance (Table 6b). Nearest Neighbor is Essential: For reference image embedding interpolation, one might hypothesize that a random in-batch sample could suffice for interpolation. However, our study demonstrates that", + "bbox": [ + 89, + 613, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/222086076f3ae561fcdc1ba9aede1d3758c83e3b6a5dfdaf70df87d9480f21cc.jpg", + "table_caption": [ + "Table 7. Performance of different LLMs with CLIP-L/14 as the visual encoder. **Bold and underline highlight the best and second best score.** indicates the original LLM for text generation; others are Large Language Embedding Models (LLEMs) fine-tuned for text retrieval." + ], + "table_footnote": [], + "table_body": "
LLMCIRCO (mAP↑)CIRR (Rec.↑)FIQ (Rec.↑)
@5@10@50@1@10@50@10@50
Stella-Qwen2-1.5B [11]14.815.317.427.569.188.029.349.0
Mistral-7B* [24]19.820.222.729.672.591.329.549.3
E5-Mistral-7B-Inst [60]19.620.322.829.572.791.129.949.5
SFR-Embedding-2 [47]20.320.823.429.772.891.530.149.5
", + "bbox": [ + 514, + 143, + 911, + 232 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "using the nearest image embedding neighbor yields significantly better performance (Table 6c). LLEMs Outperforms LLMs: We investigate Large Language Embedding Models (LLEMs), which are fine-tuned for text retrieval compared to their base LLMs. As shown in Table 7, both E5-Mistral-7B-Inst [60] and SFR-Embedding-2 outperform their LLM counterpart, Mistral-7B [24]. The superior performance of LLEMs demonstrates that models specifically tailored for embedding and retrieval tasks can offer substantial advantages in CIR applications compared to general-purpose language models.", + "bbox": [ + 511, + 257, + 906, + 424 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Limitations and Conclusion", + "text_level": 1, + "bbox": [ + 511, + 450, + 769, + 465 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "While these advancements significantly improve CIR performance, several areas warrant further investigation. Our work, limited to LLM/LLEMs, could benefit from exploring pre-trained MLLMs for enhanced understanding of CIR tasks. Additionally, our triplet synthesis method generates a single visual token, constraining the model's ability to process detailed image information. Future work should leverage the underutilized text category information in our synthetic datasets to improve model generalization. Lastly, our refined benchmarks, while more reliable, still contain noise from original image pairs, suggesting a need for further refinement of evaluation metrics.", + "bbox": [ + 511, + 484, + 906, + 664 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In conclusion, we present novel approaches to Composed Image Retrieval (CIR) that obviate the need for annotated datasets. Our contributions include: (1) an innovative triplet synthesis method utilizing image-caption pairs, (2) a new architecture leveraging LLM's embedding generation capabilities, (3) MTCIR, a diverse, human-aligned synthetic dataset, and (4) refined versions of CIRR and Fashion-IQ benchmarks, enhancing the reliability of evaluation metrics in the field. Our method consistently outperforms existing LLM and non-LLM baselines across popular benchmarks. Notably, MTCIR achieves superior results with only $10 - 20\\%$ of the size of larger datasets, demonstrating a $1 - 15\\%$ increase in Recall. These advancements collectively push the boundaries of CIR, offering more efficient and effective solutions for real-world applications.", + "bbox": [ + 511, + 674, + 908, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "4001", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 3, 7", + "[2] Anthropic. The claude 3 model family: Opus, sonnet, haiku, 2024. 3, 5, 6", + "[3] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. A clip-hitchhiker's guide to long video retrieval. arXiv preprint arXiv:2205.08508, 2022. 5", + "[4] Alberto Baldrati, Lorenzo Agnolucci, Marco Bertini, and Alberto Del Bimbo. Zero-shot composed image retrieval with textual inversion. In ICCV, 2023. 2, 3, 5, 6, 7, 13, 15", + "[5] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In CVPR, 2023. 2, 3, 5", + "[6] Tom B Brown. Language models are few-shot learners. In NeurIPS, 2020. 5", + "[7] Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, et al. Palm: Scaling language modeling with pathways. Journal of Machine Learning Research, 2023. 2, 5", + "[8] Niv Cohen, Rinon Gal, Eli A. Meirom, Gal Chechik, and Yuval Atzmon. \"this is my unicorn, fluffy\": Personalizing frozen vision-language representations. In ECCV, 2022. 2, 3, 6, 15", + "[9] Yashar Deldjoo, Fatemeh Nazary, Arnau Ramisa, Julian Mcauley, Giovanni Pellegrini, Alejandro Bellogin, and Tommaso Di Noia. A review of modern fashion recommender systems. ACM Computing Surveys, 2023. 1", + "[10] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009. 13", + "[11] Zhang Dun. Stella en qwen2 1.5b v5, 2024.8", + "[12] Sonam Goenka, Zhaoheng Zheng, Ayush Jaiswal, Rakesh Chada, Yue Wu, Varsha Hedau, and Pradeep Natarajan. Fashionvlp: Vision language transformer for fashion retrieval with feedback. In CVPR, 2022. 1", + "[13] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Batra, and Devi Parikh. Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In CVPR, 2017. 5", + "[14] Geonmo Gu, Sanghyuk Chun, Wonjae Kim, HeeJae Jun, Yoohoon Kang, and Sangdoo Yun. Compodiff: Versatile composed image retrieval with latent diffusion. Transactions on Machine Learning Research, 2024. 2, 3, 5, 6, 16", + "[15] Geonmo Gu, Sanghyuk Chun, Wonjae Kim, Yoohoon Kang, and Sangdoo Yun. Language-only training of zero-shot composed image retrieval. In CVPR, 2024. 2, 3, 6, 15", + "[16] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 13" + ], + "bbox": [ + 93, + 114, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[17] Matthew Honnibal, Ines Montani, Sofie Van Landeghem, Adriane Boyd, et al. spacy: Industrial-strength natural language processing in python. 2020. 13", + "[18] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In ICLR, 2022. 15", + "[19] Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Weli-hinda, Alan Hayes, Alec Radford, et al. Gpt-4o system card. arXiv preprint arXiv:2410.21276, 2024. 13", + "[20] Chuong Huynh, Yuqian Zhou, Zhe Lin, Connelly Barnes, Eli Shechtman, Sohrab Amirghodsi, and Abhinav Shrivastava. Simpson: Simplifying photo cleanup with single-click distracting object segmentation network. In CVPR, 2023. 1", + "[21] Young Kyun Jang, Dat Huynh, Ashish Shah, Wen-Kai Chen, and Ser-Nam Lim. Spherical linear interpolation and text-anchoring for zero-shot composed image retrieval. In ECCV, 2024. 2, 3, 6, 7, 14, 15", + "[22] Young Kyun Jang, Junmo Kang, Yong Jae Lee, and Donghyun Kim. Mate: Meet at the embedding-connecting images with long texts. EMNLP Findings, 2024. 3", + "[23] Young Kyun Jang, Donghyun Kim, Zihang Meng, Dat Huynh, and Ser-Nam Lim. Visual delta generator with large multi-modal models for semi-supervised composed image retrieval. In CVPR, 2024. 2, 3, 5", + "[24] Albert Q Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, et al. Mistral 7b. arXiv preprint arXiv:2310.06825, 2023. 8", + "[25] Ting Jiang, Minghui Song, Zihan Zhang, Haizhen Huang, Weiwei Deng, Feng Sun, Qi Zhang, Deqing Wang, and Fuzhen Zhuang. E5-v: Universal embeddings with multimodal large language models. arXiv preprint arXiv:2407.12580, 2024.3", + "[26] Shyamgopal Karthik, Karsten Roth, Massimiliano Mancini, and Zeynep Akata. Vision-by-language for training-free compositional image retrieval. *ICLR*, 2024. 3, 6, 7, 15", + "[27] Chankyu Lee, Rajarshi Roy, Mengyao Xu, Jonathan Raiman, Mohammad Shoeybi, Bryan Catanzaro, and Wei Ping. Nv-embed: Improved techniques for training llms as generalist embedding models. arXiv preprint arXiv:2405.17428, 2024.3", + "[28] Matan Levy, Rami Ben-Ari, Nir Darshan, and Dani Lischinski. Data roaming and quality assessment for composed image retrieval. AAAI, 2024. 2, 3, 5, 6, 7, 14, 15, 17, 18", + "[29] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, 2022. 2, 7, 14, 17, 18", + "[30] Wei Li, Hehe Fan, Yongkang Wong, Yi Yang, and Mohan Kankanhalli. Improving context understanding in multimodal large language models via multimodal composition learning. In ICML, 2024. 3, 6, 7", + "[31] Zehan Li, Xin Zhang, Yanzhao Zhang, Dingkun Long, Pengjun Xie, and Meishan Zhang. Towards general text" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "4002", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "embeddings with multi-stage contrastive learning. arXiv preprint arXiv:2308.03281, 2023. 3", + "[32] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 5, 7, 15", + "[33] Aixin Liu, Bei Feng, Bing Xue, Bingxuan Wang, Bochao Wu, Chengda Lu, Chenggang Zhao, Chengqi Deng, Chenyu Zhang, Chong Ruan, et al. Deepseek-v3 technical report. arXiv preprint arXiv:2412.19437, 2024. 13", + "[34] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, 2023. 2, 3, 5, 6, 8, 14, 15", + "[35] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava-last: Improved reasoning,OCR, and world knowledge, 2024. 2, 3, 5, 13", + "[36] Yikun Liu, Jiangchao Yao, Ya Zhang, Yan-Feng Wang, and Weidi Xie. Zero-shot composed text-image retrieval. In BMVC, 2023. 2", + "[37] Zheyuan Liu, Cristian Rodriguez-Opazo, Damien Teney, and Stephen Gould. Image retrieval on real-life images with pretrained vision-and-language models. In ICCV, 2021. 2, 3, 5, 7, 13, 14, 17", + "[38] Feipeng Ma, Hongwei Xue, Guangting Wang, Yizhou Zhou, Fengyun Rao, Shilin Yan, Yueyi Zhang, Siying Wu, Mike Zheng Shou, and Xiaoyan Sun. Multi-modal generative embedding model. arXiv preprint arXiv:2405.19333, 2024. 3", + "[39] Gabriel de Souza P Moreira, Radek Osmulski, Mengyao Xu, Ronay Ak, Benedikt Schifferer, and Even Oldridge. Nv-retriever: Improving text embedding models with effective hard-negative mining. arXiv preprint arXiv:2407.15831, 2024. 3", + "[40] Niklas Muennighoff, Nouamane Tazi, Loic Magne, and Nils Reimers. Mteb: Massive text embedding benchmark. arXiv preprint arXiv:2210.07316, 2022. 3", + "[41] Vicente Ordonez, Girish Kulkarni, and Tamara Berg. Im2text: Describing images using 1 million captioned photographs. NeurIPS, 2011. 5", + "[42] Khoi Pham, Chuong Huynh, Ser-Nam Lim, and Abhinav Shrivastava. Composing object relations and attributes for image-text matching. In CVPR, 2024. 1", + "[43] Bill Psomas, Ioannis Kakogeorgiou, Nikos Efthymiadis, Giorgos Tolias, Ondrej Chum, Yannis Avrithis, and Konstantinos Karantzalos. Composed image retrieval for remote sensing. In IGARSS 2024-2024 IEEE International Geoscience and Remote Sensing Symposium, 2024. 3", + "[44] Filip Radenovic, Abhimanyu Dubey, Abhishek Kadian, Todor Mihaylov, Simon Vandenhende, Yash Patel, Yi Wen, Vignesh Ramanathan, and Dhruv Mahajan. Filtering, distillation, and hard negatives for vision-language pre-training. In CVPR, 2023. 4", + "[45] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 2, 5, 13, 14" + ], + "bbox": [ + 91, + 92, + 482, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[46] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 2, 3, 5", + "[47] Meng Rui, Liu Ye, Joty Shafiq Rayhan, Xiong Caiming, Zhou Yingbo, and Yavuz Semih. Sfr-embedding-2: Advanced text embedding with multi-stage training, 2024. 8, 14", + "[48] Amrita Saha, Mitesh Khapra, and Karthik Sankaranarayanan. Towards building large scale multimodal domain-aware conversation systems. In AAAI, 2018. 3", + "[49] Kuniaki Saito, Kihyuk Sohn, Xiang Zhang, Chun-Liang Li, Chen-Yu Lee, Kate Saenko, and Tomas Pfister. Pic2word: Mapping pictures to words for zero-shot composed image retrieval. CVPR, 2023. 3, 6, 15", + "[50] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021. 5, 6, 14", + "[51] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In ACL, 2018. 5, 6, 14", + "[52] Ken Shoemake. Animating rotation with quaternion curves. In Proceedings of the 12th annual conference on Computer graphics and interactive techniques, 1985. 2, 4", + "[53] Swetha Sirnam, Jinyu Yang, Tal Neiman, Mamshad Nayeem Rizve, Son Tran, Benjamin Yao, Trishul Chilimbi, and Mubarak Shah. X-former: Unifying contrastive and reconstruction learning for mllms. In European Conference on Computer Vision, pages 146-162. Springer, 2024. 2", + "[54] Alane Suhr, Stephanie Zhou, Ally Zhang, Iris Zhang, Huajun Bai, and Yoav Artzi. A corpus for reasoning about natural language grounded in photographs. In ACL, 2019. 5", + "[55] Yuanmin Tang, Jing Yu, Keke Gai, Jiamin Zhuang, Gang Xiong, Yue Hu, and Qi Wu. Context-i2w: Mapping images to context-dependent words for accurate zero-shot composed image retrieval. In AAAI, 2024. 2, 3, 6, 15", + "[56] Omkar Thawakar, Muzammal Naseer, Rao Muhammad Anwer, Salman Khan, Michael Felsberg, Mubarak Shah, and Fahad Shahbaz Khan. Composed video retrieval via enriched context and discriminative embeddings. CVPR, 2024. 4, 6, 16", + "[57] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 2", + "[58] Lucas Ventura, Antoine Yang, Cordelia Schmid, and Gül Varol. CoVR-2: Automatic data construction for composed video retrieval. IEEE TPAMI, 2024. 2, 3, 5, 6, 7, 14, 16, 18", + "[59] Lucas Ventura, Antoine Yang, Cordelia Schmid, and Güll Varol. CoVR: Learning composed video retrieval from web video captions. AAAI, 2024. 3, 4, 5, 6, 7, 15, 16, 17", + "[60] Liang Wang, Nan Yang, Xiaolong Huang, Linjun Yang, Rangan Majumder, and Furu Wei. Improving text embeddings with large language models. arXiv preprint arXiv:2401.00368, 2023. 3, 8" + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "4003", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[61] Hui Wu, Yupeng Gao, Xiaoxiao Guo, Ziad Al-Halah, Steven Rennie, Kristen Grauman, and Rogerio Feris. The fashion iq dataset: Retrieving images by combining side information and relative natural language feedback. CVPR, 2021. 2, 5, 7, 17", + "[62] Jinyu Yang, Jiali Duan, Son Tran, Yi Xu, Sampath Chanda, Liquun Chen, Belinda Zeng, Trishul Chilimbi, and Junzhou Huang. Vision-language pre-training with triple contrastive learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15671-15680, 2022. 1", + "[63] ZHOU Yuqian, Chuong Huynh, Connelly Barnes, Elya Shechtman, Sohrab Amirghodsi, and Zhe Lin. Machine-learning models for distractor segmentation with reduced user interactions, 2024. US Patent App. 18/307,353. 1", + "[64] Kai Zhang, Yi Luan, Hexiang Hu, Kenton Lee, Siyuan Qiao, Wenhu Chen, Yu Su, and Ming-Wei Chang. Magiclens: Self-supervised image retrieval with open-ended instructions. In ICML, 2024. 2, 3, 5, 6, 7, 16, 18", + "[65] Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen, Christopher Dewan, Mona Diab, Xian Li, Xi Victoria Lin, et al. Opt: Open pre-trained transformer language models. arXiv preprint arXiv:2205.01068, 2022. 2, 5", + "[66] Youcai Zhang, Xinyu Huang, Jinyu Ma, Zhaoyang Li, Zhaochuan Luo, Yanchun Xie, Yuzhuo Qin, Tong Luo, Yaqian Li, Shilong Liu, et al. Recognize anything: A strong image tagging model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1724-1732, 2024. 13", + "[67] Xinliang Zhu, Sheng-Wei Huang, Han Ding, Jinyu Yang, Kelvin Chen, Tao Zhou, Tal Neiman, Ouye Xie, Son Tran, Benjamin Yao, et al. Bringing multimodality to amazon visual search system. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 6390–6399, 2024. 1", + "[68] Yutao Zhu, Huaying Yuan, Shuting Wang, Jiongnan Liu, Wenhan Liu, Chenlong Deng, Haonan Chen, Zhicheng Dou, and Ji-Rong Wen. Large language models for information retrieval: A survey. arXiv preprint arXiv:2308.07107, 2023. 3" + ], + "bbox": [ + 91, + 92, + 480, + 665 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "4004", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/CoLLM_ A Large Language Model for Composed Image Retrieval/02729911-7a78-4d55-af78-f50b5e45a135_model.json b/2025/CoLLM_ A Large Language Model for Composed Image Retrieval/02729911-7a78-4d55-af78-f50b5e45a135_model.json new file mode 100644 index 0000000000000000000000000000000000000000..6f8a945336984615906298d31ea7b4bf97feda07 --- /dev/null +++ b/2025/CoLLM_ A Large Language Model for Composed Image Retrieval/02729911-7a78-4d55-af78-f50b5e45a135_model.json @@ -0,0 +1,2807 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.164, + 0.131, + 0.834, + 0.154 + ], + "angle": 0, + "content": "CoLLM: A Large Language Model for Composed Image Retrieval" + }, + { + "type": "text", + "bbox": [ + 0.163, + 0.18, + 0.834, + 0.218 + ], + "angle": 0, + "content": "Chuong Huynh\\(^{1*}\\) Jinyu Yang\\(^{2\\dagger}\\) Ashish Tawari\\(^{2}\\) Mubarak Shah\\(^{2,3}\\) Son Tran\\(^{2}\\) Raffay Hamid\\(^{2}\\) Trishul Chilimbi\\(^{2}\\) Abhinav Shrivastava\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.291, + 0.233, + 0.707, + 0.251 + ], + "angle": 0, + "content": "1 University of Maryland, College Park 2 Amazon" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.251, + 0.783, + 0.27 + ], + "angle": 0, + "content": "3 Center for Research in Computer Vision, University of Central Florida" + }, + { + "type": "text", + "bbox": [ + 0.163, + 0.272, + 0.829, + 0.305 + ], + "angle": 0, + "content": "\\(^{1}\\{\\mathrm{chuonghm}, \\mathrm{abhinav}\\} @ \\mathrm{cs.umd.edu}\\) \n\\(^{2}\\{\\mathrm{viyjy}, \\mathrm{atawari}, \\mathrm{sontran}, \\mathrm{raffay}, \\mathrm{trishulc}\\} @ \\mathrm{amazon.com}\\) shah@crcv.ucf.edu" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.34, + 0.327, + 0.354 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.373, + 0.485, + 0.873 + ], + "angle": 0, + "content": "Composed Image Retrieval (CIR) is a complex task that aims to retrieve images based on a multimodal query. Typical training data consists of triplets containing a reference image, a textual description of desired modifications, and the target image, which are expensive and time-consuming to acquire. The scarcity of CIR datasets has led to zero-shot approaches utilizing synthetic triplets or leveraging vision-language models (VLMs) with ubiquitous web-crawled image-caption pairs. However, these methods have significant limitations: synthetic triplets suffer from limited scale, lack of diversity, and unnatural modification text, while image-caption pairs hinder joint embedding learning of the multimodal query due to the absence of triplet data. Moreover, existing approaches struggle with complex and nuanced modification texts that demand sophisticated fusion and understanding of vision and language modalities. We present CoLLM, a one-stop framework that effectively addresses these limitations. Our approach generates triplets on-the-fly from image-caption pairs, enabling supervised training without manual annotation. We leverage Large Language Models (LLMs) to generate joint embeddings of reference images and modification texts, facilitating deeper multimodal fusion. Additionally, we introduce Multi-Text CIR (MTCIR), a large-scale dataset comprising 3.4M samples, and refine existing CIR benchmarks (CIRR and Fashion-IQ) to enhance evaluation reliability. Experimental results demonstrate that CoLLM achieves state-of-the-art performance across multiple CIR benchmarks and settings. MTCIR yields competitive results, with up to \\(15\\%\\) performance improvement. Our refined benchmarks provide more reliable evaluation metrics for CIR models, contributing to the advancement of this important field. Project page is at collm-cvpr25.github.io." + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.332, + 0.907, + 0.412 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.604, + 0.414, + 0.817, + 0.427 + ], + "angle": 0, + "content": "(a) Composed Image Retrieval Example" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.433, + 0.907, + 0.58 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.544, + 0.582, + 0.876, + 0.594 + ], + "angle": 0, + "content": "(b) Recall Sum of state-of-the-art methods on CIR benchmarks" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.606, + 0.907, + 0.677 + ], + "angle": 0, + "content": "Figure 1. (a) An example of CIR. (b) Recall Sum at \\(\\{1,10,50\\}\\) for CIRR and \\(\\{10,50\\}\\) for Fashion-IQ between CoLLM and state-of-the-art (SoTA) models under zero-shot settings. We evaluate two training scenarios: (i) without triplet data and (ii) with synthetic triplet data." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.692, + 0.645, + 0.707 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.718, + 0.907, + 0.884 + ], + "angle": 0, + "content": "Composed Image Retrieval (CIR) enhances traditional image retrieval [42, 62] by combining text and image queries, offering greater flexibility in search systems, with applications in e-commerce, fashion, and design [9, 12, 20, 63, 67]. As illustrated in Fig. 1a, CIR retrieves similar items, such as shirts, where the reference image provides a visual basis, and the modification text specifies desired modifications. By expanding the capabilities of conventional image or text-based searches, CIR allows for more nuanced and specific queries. However, this advanced approach also presents significant challenges compared to traditional image retrieval." + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.886, + 0.905, + 0.902 + ], + "angle": 0, + "content": "Supervised CIR approaches face challenges in data ac" + }, + { + "type": "page_footnote", + "bbox": [ + 0.108, + 0.887, + 0.46, + 0.901 + ], + "angle": 0, + "content": "*This work was done during internship at Amazon. † Project lead." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3994" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.409 + ], + "angle": 0, + "content": "quisition. They require high-quality CIR triplets (reference image, target image, and modification text), collected through a labor-intensive and expensive process. Consequently, existing CIR triplet datasets are limited in scale and domain coverage, restricting model generalizability. To overcome these limitations, recent CIR methods [14, 23, 28, 58, 64] have adopted zero-shot approaches [36]. These can be broadly categorized into two strategies: (i) leveraging vision-language models [29, 45, 53] (VLMs) for composed query embedding generation, and (ii) generating synthetic triplets for supervised training. VLM-based approaches utilize pre-trained models that already align vision and language features in latent space. These methods rely on large-scale image-caption pairs and follow two main directions, i.e., textual inversion [8, 15, 55] and direct interpolation [21]. Synthetic triplet generation approaches [23, 28, 64] employ Large Language Models (LLMs) [7, 57, 65] to generate modification text, with CompoDiff [14] further utilizing diffusion models [5, 46] to create synthetic reference-target image pairs. Despite the recent advances in CIR, several critical challenges persist:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.413, + 0.483, + 0.459 + ], + "angle": 0, + "content": "- [Data] Limitation-1: VLM-based methods may hinder efficient and effective composed query embedding learning by relying solely on image-caption pairs." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.459, + 0.483, + 0.519 + ], + "angle": 0, + "content": "- [Data] Limitation-2: Existing synthetic triplet datasets often suffer from a lack of diversity and unnatural modification text. Moreover, these datasets are either very small or closed-source, impeding research progress in the field." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.52, + 0.483, + 0.594 + ], + "angle": 0, + "content": "- [Model] Limitation-3: Current methods for composed query embeddings mainly use shallow transformer models or linear interpolation. While these methods are computationally efficient, they lack the ability to capture the full complexity of the composed understanding tasks." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.595, + 0.483, + 0.716 + ], + "angle": 0, + "content": "- [Evaluation] Limitation-4: Existing CIR benchmarks are often compromised by noise, particularly in the form of ambiguous samples. Ambiguity arises when multiple target images can correctly match a single query, but only one is labeled as the ground truth, ignoring valid matches. Although CIRCO [4] attempts to address this issue, its efforts are limited in scale. This ambiguity in benchmarks hinders meaningful model evaluation and comparison." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.413, + 0.483, + 0.716 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.483, + 0.901 + ], + "angle": 0, + "content": "We propose CoLLM, an LLM-based CIR approach to address the aforementioned limitations. CoLLM tackles Limitation-1 by dynamically synthesizing triplets from image-caption pairs, introducing two key components: a reference image embedding synthesis and a modification text synthesis module. The former employs Spherical Linear Interpolation [52] (Slerp) to generate an intermediate embedding between a given image and its nearest neighbor within the training batch, serving as a synthesized reference image embedding. The latter utilizes a pre-defined text interpolation template to generate modification text based on the current caption and that of the nearest image neighbor." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.167 + ], + "angle": 0, + "content": "This strategy effectively leverages the vast amount of readily available image-caption pairs on the Internet, enabling a model's training in a supervised CIR manner. By doing so, CoLLM overcomes the scarcity of labeled triplet data and paves the way for more robust and scalable CIR models." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.168, + 0.905, + 0.362 + ], + "angle": 0, + "content": "To address Limitation-2, we introduce Multi-Text CIR (MTCIR), a synthetic dataset of 3.4M image pairs with 17.7M modification texts. MTCIR focuses on two often-overlooked aspects: image diversity and naturalistic modification texts. We curated images from diverse sources to ensure variety. For modification text generation, we employ a two-stage approach using Multi-modal LLM (MLLM) [34, 35] for detailed captioning and LLM for describing interception differences. Uniquely, MTCIR provides multiple short modification texts for each image pair, covering various attributes. This better reflects human query formulation, offering a more realistic, comprehensive training foundation for CIR models." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.364, + 0.905, + 0.515 + ], + "angle": 0, + "content": "To overcome Limitation-3, CoLLM harnesses the power of LLMs for composed query understanding. It is motivated by the extensive world knowledge embedded in pre-trained LLMs. With their deep semantic understanding, we posit that LLMs are superior to shallow transformers and simple embedding interpolation techniques in comprehending the intricate relationships between reference images and modification texts. By leveraging LLMs, CoLLM aims to capture nuanced semantic connections, enhancing the quality and relevance of composed image retrieval results." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.516, + 0.905, + 0.606 + ], + "angle": 0, + "content": "To address Limitation-4, we refine two popular CIR benchmarks: CIRR [37] and Fashion-IQ [61]. We use MLLMs to evaluate sample ambiguity in each benchmark and regenerate clear modification text for ambiguous ones. Our pipeline incorporates multiple validation steps to guarantee the enhanced quality of the refined samples." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.607, + 0.905, + 0.833 + ], + "angle": 0, + "content": "Our main contributions can be summarized as: (i) We propose a method to synthesize CIR triplets on-the-fly from image-caption pairs, which can even outperform models trained on CIR triplets, eliminating the need for costly CIR datasets. (ii) We collect and will release MTCIR, a new CIR triplet dataset covering 3.4 million image pairs with 17.7 million modification texts. To our knowledge, MTCIR is the largest open-source synthetic CIR dataset. (iii) We introduce an approach to leverage LLMs for composed query understanding, utilizing their instruction-following and embedding generation capabilities. (iv) We refine CIRR and Fashion-IQ to provide more robust evaluation benchmarks for the CIR community. Extensive experiments on popular CIR benchmarks and settings demonstrate the effectiveness of our model innovations and new datasets (Fig. 1b)." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.841, + 0.665, + 0.858 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Composed Image Retrieval. Composed Image Retrieval (CIR) has garnered significant attention due to its flexibil" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3995" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.272 + ], + "angle": 0, + "content": "ity in search systems [37, 43, 48]. Zero-shot CIR methods have been extensively explored, with textual inversion [4, 8, 15, 49, 55] emerging as a prominent technique. This approach maps image encoder outputs to text encoder inputs, creating pseudo-word tokens. SlerpTAT [21] emphasizes the text encoder's importance by re-aligning visual and textual embeddings. While most zero-shot CIR models utilize image-caption pairs for training, our work introduces an innovative method to synthesize triplets from these pairs during training. This method enables supervised CIR style training, enhancing the model's understanding of composed queries without relying on real CIR datasets." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.274, + 0.483, + 0.472 + ], + "angle": 0, + "content": "Recent works enhance enhancing models' language understanding using Large Language Models (LLMs) [26, 30], primarily leveraging their text generation abilities. MCL [30] composes image-text query in the LLM input space and aligns joint embedding with the target image caption. However, this may introduce noise due to limited visual information in the target image captions. CIReVL [26] employs LLMs to generate captions based on the reference image and modification text, subsequently retrieving the target image using a text-to-image retrieval approach. Our model also utilizes LLMs but differs by directly producing a composed query embedding for target image retrieval, potentially reducing intermediate steps and associated errors." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.474, + 0.483, + 0.702 + ], + "angle": 0, + "content": "CIR with Synthetic Datasets. The scarcity of supervised CIR datasets has prompted recent studies to leverage generative models for synthetic triplet creation. CompoDiff [14] employs image editing pipelines [5, 46] to generate target images, though their nature limits performance. Other approaches [23, 28, 58, 59, 64] utilize LLMs to generate modification text for real image pairs. Our work introduces a novel two-stage approach, combining MLLMs [34, 35] for detailed captioning and LLMs [1, 2] for describing interception differences. Distinctly, our synthetic triplets provide multiple concise modification texts for each image pair, encompassing various attributes, in contrast to the conventional single, lengthy modification text. This enables a more nuanced representation of image modifications, potentially improving CIR model performance and versatility." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.483, + 0.903 + ], + "angle": 0, + "content": "Large Language Models. Recent advancements in LLMs have expanded their applications beyond text generation to include image understanding [34, 35] and text embedding generation [40, 60, 68]. Large Language Embedding Models (LLEMs) are specialized versions trained using contrastive learning, leveraging LLM knowledge for embedding generation while often disabling text generation capabilities. To enhance text embedding quality, [27, 31, 39] propose removing causal attention in LLMs, thereby improving text information encoding efficiency and training these modified LLMs on text retrieval datasets. This text retrieval capability has been further extended to text-image retrieval by aligning visual and LLM text embed" + }, + { + "type": "image", + "bbox": [ + 0.54, + 0.09, + 0.877, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.539, + 0.218, + 0.882, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.313, + 0.905, + 0.342 + ], + "angle": 0, + "content": "Figure 2. An overview of our model and training strategies when using (a) image-caption pairs and (b) CIR triplets." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.349, + 0.907, + 0.471 + ], + "angle": 0, + "content": "ding spaces [22, 25, 38]. These developments demonstrate the potential of LLMs for composed query understanding in CIR tasks. Our work builds upon this foundation, extending LLMs/LLEMs to generate composed query embeddings by incorporating reference images and modification text. This novel approach leverages the multimodal capabilities of LLMs to enhance the performance of CIR systems, leading to more accurate and context-aware image retrieval." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.485, + 0.605, + 0.501 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.511, + 0.906, + 0.603 + ], + "angle": 0, + "content": "This section outlines our methodology, starting with a brief overview of the model architecture, followed by a formal CIR problem definition. We then detail our triplet synthesis strategy, including reference image embedding and modification text synthesis, using image-caption pairs. Lastly, we describe our LLM-based query composition approach." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.605, + 0.702, + 0.62 + ], + "angle": 0, + "content": "3.1. Model Architecture" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.628, + 0.907, + 0.81 + ], + "angle": 0, + "content": "As illustrated in Fig. 2, CoLLM consists of several essential components: (1) a vision encoder \\( f(\\cdot) \\) for extracting image features; (2) modules for synthesizing reference image embeddings and modification text; (3) an image adapter \\( g(\\cdot) \\) that maps visual features into the language model's semantic space; (4) a LLM \\( \\Phi(\\cdot) \\) that processes multimodal queries and (5) a projection layer \\( proj(\\cdot) \\) (omitted from the figure for simplicity) that maps the LLM output to a suitable representation for retrieval. It is important to note that Fig. 2 (a) and Fig. 2 (b) illustrate architectures designed for two distinct input formats. The former is tailored for image-caption pairs, while the latter is optimized for CIR triplets." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.819, + 0.885, + 0.836 + ], + "angle": 0, + "content": "3.2. CoLLM with Image-Caption Pairs as Input" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.84, + 0.906, + 0.903 + ], + "angle": 0, + "content": "Let \\(\\mathcal{X} = \\{(v_i, w_i)\\}_{i=1}^N\\) denote a set of image-caption pairs, where \\(v_i\\) and \\(w_i\\) represent the image and caption of the \\(i^{th}\\) sample, respectively. To effectively leverage the vast amount of image-caption pairs, we introduce a novel ap" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3996" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.089, + 0.482, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.262, + 0.483, + 0.304 + ], + "angle": 0, + "content": "Figure 3. An overview of reference image embedding synthesis and modification text synthesis. The red-framed image represents the nearest neighbor of the augmented image in the training batch." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.31, + 0.484, + 0.522 + ], + "angle": 0, + "content": "proach that synthesizes a CIR triplet on-the-fly for each image-caption pair \\((v_{i},w_{i})\\). In this process, \\(v_{i}\\) serves as the target image, while two distinct modules generate the remaining components: (i) a reference image embedding synthesizer and (ii) a modification text synthesizer (Fig. 2a). Notably, we do not generate an actual image for reference image synthesis. Instead, we synthesize the reference image embedding, which is computationally more efficient and allows seamless integration into the CIR pipeline. This formulation enables CoLLM to exploit the rich information in image-caption pairs, transforming them into CIR triplets that can be used for training. The following subsections detail the specific mechanisms for synthesizing the reference image embedding and the modification text." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.525, + 0.484, + 0.633 + ], + "angle": 0, + "content": "Reference Image Embedding Synthesis. Our reference image embedding synthesis process is illustrated in Fig. 3 (left). We begin by applying an augmentation \\( t \\sim \\mathcal{T} \\) to the input image \\( v_{i} \\), where \\( \\mathcal{T} \\) represents a family of augmentations. The resulting augmented image is denoted as \\( v_{i}^{\\prime} \\) with its corresponding embedding \\( \\mathbf{h}_i^\\prime = f(v_i^\\prime) \\). For \\( \\mathbf{h}_i^\\prime \\), we identify its in-batch nearest neighbor \\( v_{j}^{\\prime} \\) with \\( j \\) defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.182, + 0.641, + 0.483, + 0.67 + ], + "angle": 0, + "content": "\\[\nj = \\underset {j \\neq i; j \\in \\{1, \\dots \\mathcal {B} \\}} {\\arg \\max } \\operatorname {s i m} \\left(\\mathbf {h} _ {i} ^ {\\prime}, \\mathbf {h} _ {j} ^ {\\prime}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.678, + 0.483, + 0.711 + ], + "angle": 0, + "content": "where \\(\\mathcal{B}\\) is the batch size, \\(\\mathbf{h}_j^{\\prime} = f(v_j^{\\prime})\\), and \\(\\mathrm{sim}(\\mathbf{u},\\mathbf{v}) = \\mathbf{u}^T\\mathbf{v} / (\\| \\mathbf{u}\\| \\| \\mathbf{v}\\|)\\) denotes cosine similarity." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.711, + 0.484, + 0.772 + ], + "angle": 0, + "content": "We then employ Spherical Linear Interpolation (Slerp) [52] to synthesize the reference image embedding by interpolating between \\(\\mathbf{h}_i^{\\prime}\\) and \\(\\mathbf{h}_j^{\\prime}\\). The synthesized reference image embedding is obtained as:" + }, + { + "type": "equation", + "bbox": [ + 0.168, + 0.781, + 0.305, + 0.799 + ], + "angle": 0, + "content": "\\[\n\\theta = \\operatorname {a r c c o s} \\left(\\mathbf {h} _ {i} ^ {\\prime} \\cdot \\mathbf {h} _ {j} ^ {\\prime}\\right)\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.16, + 0.8, + 0.483, + 0.833 + ], + "angle": 0, + "content": "\\[\n\\mathbf {h} _ {i} ^ {*} = \\frac {\\sin (\\alpha \\theta)}{\\sin \\theta} \\mathbf {h} _ {i} ^ {\\prime} + \\frac {\\sin ((1 - \\alpha) \\theta)}{\\sin \\theta} \\mathbf {h} _ {j} ^ {\\prime}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "where \\(\\alpha \\in [0,1]\\) is a hyper-parameter controlling the interpolation strength. The intuition behind this approach is that an image with embedding \\(\\mathbf{h}_i^*\\) shares certain visual similarities with the target image \\(v_{i}\\). Specifically, a larger \\(\\alpha\\) results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.122 + ], + "angle": 0, + "content": "in \\(\\mathbf{h}_i^*\\) more closely resembling \\(\\mathbf{h}_i'\\), allowing for fine-grained control over the synthesized embedding." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.125, + 0.906, + 0.412 + ], + "angle": 0, + "content": "Modification Text Synthesis. For the augmented image \\( v_{i}^{\\prime} \\), its nearest neighbor is \\( (v_{j}^{\\prime}, w_{j}) \\). To generate the modification text \\( w_{i}^{*} \\), we use text interpolation to combine \\( w_{i} \\) and \\( w_{j} \\) using pre-defined templates, as illustrated in Fig. 3 (right). The complete set of templates is provided in the supplementary material. During training, we randomly select a template for each sample to ensure diversity in the synthesized modification text. It is worth noting that we do not simply use \\( w_{i} \\) as the modification text for two primary reasons: (i) \\( w_{i} \\) alone fails to capture the visual differences between the reference image and the target image. (ii) Using only \\( w_{i} \\) could lead to model cheating, where it learns to rely solely on \\( w_{i} \\) for retrieval while ignoring the reference image. Our synthesis approach, in contrast, mimics real-world modification text, which often describes both similarities and differences between the reference and target images. Furthermore, this strategy forces the model to learn composed query embeddings by considering both the reference image and the modification text simultaneously." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.412, + 0.905, + 0.503 + ], + "angle": 0, + "content": "By combining the synthesized reference embedding \\(\\mathbf{h}_i^*\\) and the interpolated modification text \\(w_{i}^{*}\\), our method generates diverse CIR triplets from image-caption pairs. This approach enables effective training of the CoLLM model on large-scale image-caption datasets, effectively mitigating the dependency on scarce labeled triplet data." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.506, + 0.906, + 0.582 + ], + "angle": 0, + "content": "Query Composition with LLM. To construct the composed query, we utilize a pre-trained LLM that processes the synthesized reference image embedding \\(\\mathbf{h}_i^*\\), image caption \\(w_i\\), and the synthesized modification text \\(w_i^*\\). We define three distinct composed embeddings:" + }, + { + "type": "equation", + "bbox": [ + 0.628, + 0.588, + 0.905, + 0.605 + ], + "angle": 0, + "content": "\\[\n\\mathbf {c} _ {i} ^ {v} = p \\left(\\Phi \\left(g \\left(\\mathbf {h} _ {i} ^ {*}\\right)\\right)\\right) \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.627, + 0.607, + 0.905, + 0.624 + ], + "angle": 0, + "content": "\\[\n\\mathbf {c} _ {i} ^ {w} = p \\left(\\Phi \\left(w _ {i}\\right)\\right) \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.632, + 0.626, + 0.905, + 0.643 + ], + "angle": 0, + "content": "\\[\n\\mathbf {c} _ {i} = p \\left(\\Phi \\left([ g \\left(\\mathbf {h} _ {i} ^ {*}\\right); w _ {i} ^ {*} ]\\right)\\right) \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.648, + 0.905, + 0.679 + ], + "angle": 0, + "content": "where \\([\\cdot ]\\) denotes an instruction template that combine two modalities (see supplementary material)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.682, + 0.906, + 0.758 + ], + "angle": 0, + "content": "Training Objective. We employ a contrastive loss \\(\\mathcal{L}_{cl}\\) [44], consistent with previous works [56, 59], to compute the loss between query embeddings \\(\\{\\mathbf{c}_i^v,\\mathbf{c}_i^w,\\mathbf{c}_i\\}\\) and the target image embedding \\(\\mathbf{z}_i = f(v_i)\\). The final loss for each sample during pre-training is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.539, + 0.763, + 0.905, + 0.791 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\frac {1}{3} \\left(\\mathcal {L} _ {c l} \\left(\\mathbf {c} _ {i} ^ {v}, \\mathbf {z} _ {i}\\right) + \\mathcal {L} _ {c l} \\left(\\mathbf {c} _ {i} ^ {w}, \\mathbf {z} _ {i}\\right) + \\mathcal {L} _ {c l} \\left(\\mathbf {c} _ {i}, \\mathbf {z} _ {i}\\right)\\right) \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.906, + 0.901 + ], + "angle": 0, + "content": "This formulation encourages the model to learn discriminative representations for different query types while maintaining consistency with the target image embedding. By combining losses from image-to-image \\((\\mathbf{c}_i^v)\\), text-to-image \\((\\mathbf{c}_i^w)\\), and composed \\((\\mathbf{c}_i)\\) queries, we ensure that the model learns to effectively process and align various input modalities for compositional image retrieval." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3997" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.333, + 0.071, + 0.665, + 0.085 + ], + "angle": 0, + "content": "Table 1. Comparison of synthetic CIR training datasets." + }, + { + "type": "table", + "bbox": [ + 0.105, + 0.089, + 0.891, + 0.248 + ], + "angle": 0, + "content": "
DatasetPublicReference-Target Image PairModification Text Generation
TypeSource# Pairs# EntitiesLLMLLM input# Text# Text/Pair
LaSCo [28]✓imageVQAv2 [13]360K82KGPT-3 [6]question-answer360K1
VDG [23]✓imageNLVR2 [54], COCO [32]467K183KVision-LLaMA2image pair523K1.12
WebCoVR [59]✓videoWebVid2M [3]1.6M131KCustom LLMcaption1.6M1
CC-CoIR [58]✓imageCC3M [51]3.3M357KCustom LLMcaption3.3M1
SynthTriplets [14]synth. imageSD [46], IP2P [5]18.8M37.6MOPT-6.7B [65]-18.8M1
MagicLens [64]imageWeb crawled36.5M-PaLM [7]image metadata36.5M1
MTCIR (Ours)✓imageLLaVA-558k [34]3.4M423KSonnet 3 [2]synth. caption17.7M5.18
" + }, + { + "type": "image_caption", + "bbox": [ + 0.162, + 0.254, + 0.246, + 0.264 + ], + "angle": 0, + "content": "Reference Image" + }, + { + "type": "image", + "bbox": [ + 0.121, + 0.265, + 0.284, + 0.338 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.333, + 0.254, + 0.4, + 0.264 + ], + "angle": 0, + "content": "Target Image" + }, + { + "type": "image", + "bbox": [ + 0.287, + 0.265, + 0.451, + 0.338 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.126, + 0.342, + 0.174, + 0.352 + ], + "angle": 0, + "content": "Category" + }, + { + "type": "image_footnote", + "bbox": [ + 0.125, + 0.352, + 0.203, + 0.362 + ], + "angle": 0, + "content": "+ Object Added" + }, + { + "type": "image_footnote", + "bbox": [ + 0.125, + 0.363, + 0.223, + 0.373 + ], + "angle": 0, + "content": "\\(\\mathbf{3}\\) Attribute Changed" + }, + { + "type": "image_footnote", + "bbox": [ + 0.125, + 0.373, + 0.216, + 0.383 + ], + "angle": 0, + "content": "Object Removed" + }, + { + "type": "list", + "bbox": [ + 0.125, + 0.352, + 0.223, + 0.383 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.274, + 0.341, + 0.362, + 0.351 + ], + "angle": 0, + "content": "Modification Text" + }, + { + "type": "image_footnote", + "bbox": [ + 0.274, + 0.352, + 0.444, + 0.362 + ], + "angle": 0, + "content": "Add a person's hand holding the cup." + }, + { + "type": "image_footnote", + "bbox": [ + 0.274, + 0.363, + 0.42, + 0.373 + ], + "angle": 0, + "content": "Modify the liquid being poured." + }, + { + "type": "image_footnote", + "bbox": [ + 0.274, + 0.373, + 0.446, + 0.383 + ], + "angle": 0, + "content": "Remove the social media text overlay." + }, + { + "type": "list", + "bbox": [ + 0.274, + 0.341, + 0.446, + 0.383 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.389, + 0.484, + 0.417 + ], + "angle": 0, + "content": "Figure 4. An example from the MTCIR dataset. Each sample contains multiple short texts describing different modifications." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.424, + 0.392, + 0.44 + ], + "angle": 0, + "content": "3.3. CoLLM with CIR Triplet as Input" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.445, + 0.484, + 0.567 + ], + "angle": 0, + "content": "For CIR triplet inputs, our model design is shown in Fig. 2b. Let \\(\\mathcal{X} = \\{(v_i^r,v_i^t,w_i)\\}_{i = 1}^N\\) be a set of triplets, where \\(v_{i}^{r}\\) \\(v_{i}^{t}\\), and \\(w_{i}\\) represent the reference image, target image, and modification text of the \\(i^{th}\\) sample, respectively. The reference embedding is obtained as \\(\\mathbf{h}_i = f(v_i^r)\\). The composed embedding is then computed as \\(\\mathbf{c}_i = p(\\Phi ([g(\\mathbf{h}_i);w_i])\\). The training objective is \\(\\mathcal{L} = \\mathcal{L}_{cl}(\\mathbf{c}_i,\\mathbf{z}_i)\\), where \\(\\mathbf{z}_i = f(v_i^t)\\) denotes the target embedding." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.574, + 0.295, + 0.589 + ], + "angle": 0, + "content": "4. Dataset Construction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.598, + 0.484, + 0.689 + ], + "angle": 0, + "content": "This section outlines the data construction process for two primary objectives: (1) creating a novel CIR training dataset called Multi-Text CIR (MTCIR) and (2) refining widely used CIR benchmarks, specifically CIRR [37] and FashionIQ [61]. We detail the methodologies for each task, highlighting the improvements and innovations." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.698, + 0.388, + 0.714 + ], + "angle": 0, + "content": "4.1. Multi-Text CIR (MTCIR) Dataset" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.483, + 0.764 + ], + "angle": 0, + "content": "Existing CIR triplet datasets [23, 28, 59] lack diversity and contain unnatural modification text. We introduce Multi-Text CIR (MTCIR) to address these limitations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.484, + 0.902 + ], + "angle": 0, + "content": "To enhance image diversity, we source images from the LLaVA-558k dataset [34] filtered from image caption datasets [41, 50, 51] based on noun-phrase frequency for broad concept coverage. We pair images using CLIP [45] visual similarity metrics, following CIRR's grouping approach, yielding 3.4M pairs from approximately 423K images. Leveraging the detailed captions generated by LLaVA-Next-34B [35] for each image in the LLaVA-558k dataset, we employ Claude 3 Sonnet [2]" + }, + { + "type": "image_caption", + "bbox": [ + 0.522, + 0.254, + 0.603, + 0.268 + ], + "angle": 0, + "content": "Unchanged" + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.272, + 0.612, + 0.348 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.541, + 0.352, + 0.591, + 0.362 + ], + "angle": 0, + "content": "(a) CIRR" + }, + { + "type": "image_caption", + "bbox": [ + 0.622, + 0.254, + 0.694, + 0.268 + ], + "angle": 0, + "content": "Regenerate" + }, + { + "type": "image", + "bbox": [ + 0.617, + 0.272, + 0.708, + 0.348 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.625, + 0.351, + 0.7, + 0.363 + ], + "angle": 0, + "content": "(b) FIQ-Dress" + }, + { + "type": "image_caption", + "bbox": [ + 0.718, + 0.254, + 0.779, + 0.268 + ], + "angle": 0, + "content": "Remove" + }, + { + "type": "image", + "bbox": [ + 0.713, + 0.272, + 0.804, + 0.348 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.725, + 0.351, + 0.795, + 0.363 + ], + "angle": 0, + "content": "(c) FIQ-Shirt" + }, + { + "type": "image_caption", + "bbox": [ + 0.803, + 0.254, + 0.892, + 0.268 + ], + "angle": 0, + "content": "Cannot process" + }, + { + "type": "image", + "bbox": [ + 0.81, + 0.272, + 0.901, + 0.348 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.815, + 0.351, + 0.896, + 0.364 + ], + "angle": 0, + "content": "(d) FIQ-Toptee" + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.376, + 0.906, + 0.404 + ], + "angle": 0, + "content": "Figure 5. Statistical analysis of the refinement process for CIRR and FashionIQ (FIQ) datasets." + }, + { + "type": "image_caption", + "bbox": [ + 0.558, + 0.413, + 0.64, + 0.422 + ], + "angle": 0, + "content": "Reference Image" + }, + { + "type": "image", + "bbox": [ + 0.547, + 0.422, + 0.655, + 0.484 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.55, + 0.484, + 0.631, + 0.495 + ], + "angle": 0, + "content": "Original Text" + }, + { + "type": "image_caption", + "bbox": [ + 0.549, + 0.5, + 0.651, + 0.512 + ], + "angle": 0, + "content": "Regenerated Text" + }, + { + "type": "image_caption", + "bbox": [ + 0.677, + 0.413, + 0.741, + 0.422 + ], + "angle": 0, + "content": "Target Image" + }, + { + "type": "image", + "bbox": [ + 0.656, + 0.422, + 0.764, + 0.484 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.671, + 0.485, + 0.824, + 0.494 + ], + "angle": 0, + "content": "One puppy looking at the camera." + }, + { + "type": "image_footnote", + "bbox": [ + 0.671, + 0.495, + 0.855, + 0.505 + ], + "angle": 0, + "content": "The dog is indoors on a tiled floor with a" + }, + { + "type": "image_footnote", + "bbox": [ + 0.671, + 0.505, + 0.753, + 0.515 + ], + "angle": 0, + "content": "food bowl nearby." + }, + { + "type": "list", + "bbox": [ + 0.671, + 0.485, + 0.855, + 0.515 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.522, + 0.906, + 0.591 + ], + "angle": 0, + "content": "Figure 6. An example from the refined CIRR dataset. The original modification text yields multiple plausible results (shown as similar images to the target on the right). In contrast, the newly generated text accurately captures the distinctive features of the target image, differentiating it from other candidates." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.599, + 0.906, + 0.719 + ], + "angle": 0, + "content": "(20240229-v1:0) to generate modification text for each image pair, using these captions as input. To control detail levels in Claude outputs, we define six modification text categories in the prompt, inspired by prior research [4, 37]. This approach maintains diverse aspects and prevents repetition. We also incorporate examples from the CIRR dataset to ensure human-like modification text. Detailed prompting strategies are provided in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.906, + 0.87 + ], + "angle": 0, + "content": "Our method generates multiple brief, focused texts for each image pair, each highlighting a unique aspect. This approach captures all changes between two images without producing overly long sentences, better aligning with human-written modification text. These short texts can be combined for training, enhancing dataset diversity. We obtained 17.7M text samples for the 3.4M image pairs, averaging 5.18 short sentences per pair. Fig. 4 illustrates an example, showcasing modification texts corresponding to specific categories and covering diverse attributes." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.906, + 0.901 + ], + "angle": 0, + "content": "As shown in Table 1, MTCIR offers several advantages over existing synthetic CIR datasets: (i) Unlike [14], which" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3998" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.071, + 0.483, + 0.154 + ], + "angle": 0, + "content": "Table 2. Performance comparison of various models and visual encoders on CIR benchmarks, evaluated without training on triplet datasets. Bold and underlined values denote the best and second-best scores within each vision encoder group. Models incorporating LLMs in their architectures are marked with \\(\\star\\). Results reproduced by our team are indicated by \\(\\ddagger\\)." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.159, + 0.487, + 0.442 + ], + "angle": 0, + "content": "
MethodCIRCO (mAP↑)CIRR (Rec.↑)FIQ (Rec.↑)
@5@10@50@1@10@50@10@50
OpenAI CLIP-B/32
PALAVRA [8]4.65.36.816.658.584.019.837.3
SEARLE [4]9.49.911.824.066.889.822.942.5
Slerp-TAT [21]9.310.312.328.268.888.523.044.0
CIREVL* [26]14.915.417.823.966.087.028.349.4
CoLLM*12.913.215.028.671.892.724.845.2
OpenAI CLIP-L/14
Pic2World [49]8.79.511.323.965.387.824.743.7
SEARLE [4]11.712.715.124.266.388.825.646.2
LinCIR‡[15]13.013.916.224.666.988.826.446.6
ContextI2W [55]13.013.816.025.768.689.627.848.9
MCL* [30]17.818.421.825.969.491.0--
Slerp-TAT [21]18.519.421.430.970.989.228.347.6
CIREVL* [26]18.619.021.824.664.986.328.648.6
CoLLM*20.320.823.429.772.891.530.149.5
BLIP-L/16
Slerp-TAT [21]17.818.421.134.072.788.932.853.3
CoLLM*19.720.423.135.078.694.234.656.0
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.453, + 0.483, + 0.528 + ], + "angle": 0, + "content": "uses synthetic images, MTCIR uses real images covering diverse concepts and domains. (ii) MTCIR provides multiple modification texts for each image pair, offering more natural descriptions and training flexibility. (iii) MTCIR is the largest public dataset to advance CIR research." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.53, + 0.484, + 0.606 + ], + "angle": 0, + "content": "To prevent the leakage of biometric information, we further process MTCIR at both the image and text levels. For images, we apply face blurring to all instances. For text, we remove any content containing keywords related to human attributes (e.g., skin, hair, gender, age, race)." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.62, + 0.433, + 0.636 + ], + "angle": 0, + "content": "4.2. Refined CIRR and Fashion-IQ Datasets" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.643, + 0.483, + 0.764 + ], + "angle": 0, + "content": "Current CIR benchmarks often suffer from label ambiguity, where modification text lacks clarity, potentially matching multiple target images. To address this issue, we propose a refinement method using Claude 3 Sonnet [2] (20240229-v1:0) to enhance the validation sets of two well-known benchmarks: CIRR and Fashion-IQ. Each sample in these validation sets comprises a triplet containing a reference image, target image, and modification text." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.766, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Our refinement process, applicable to any benchmark, consists of validation, re-generation, and re-validation steps. For each triplet, we first identify the top-3 visually similar images to the target image based on CLIP visual scores, serving as hard negative samples for ambiguity assessment. We employ Claude 3 Sonnet to evaluate each triplet's ambiguity by attempting to identify the target image from these hard negatives. Triplets correctly identified by Claude 3 Sonnet are considered \"good\" samples and re" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.071, + 0.907, + 0.126 + ], + "angle": 0, + "content": "Table 3. Performance of models trained on synthetic triplet datasets. Bold indicates the best method overall, while underline highlights the best method within the same vision encoder. Our CoLLM outperforms comparable methods in most metrics." + }, + { + "type": "table", + "bbox": [ + 0.513, + 0.13, + 0.907, + 0.329 + ], + "angle": 0, + "content": "
MethodDatasetCIRR ↑FIQ ↑
@1@10@50@10@50
CoCa-L/18 288 × 288
MagicLens [64]MagicLens [64]33.377.994.438.158.3
EVA-CLIP ViT-G/14 364 × 364
CoVR2 [58]WV-CC-CoVIR [58]43.784.096.138.258.4
OpenAI CLIP-L/14 224 × 224
CompoDiff [14]SynTrip18M [14]18.270.890.336.048.6
MagicLens [64]MagicLens [64]30.174.492.630.752.5
CoLLMMTCIR (ours)34.777.093.132.954.2
BLIP-L/16 384 × 384; fine-tuned on COCO
CASE [28]LaSCo [28]35.478.594.6--
Omkar et al. [56]WebCoVR [59]40.178.994.730.346.5
CoLLMMTCIR (ours)45.884.795.839.160.7
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.337, + 0.907, + 0.533 + ], + "angle": 0, + "content": "main unchanged. For \"bad\" samples, we use Claude 3 Sonnet to re-generate the modification text. The model produces three new modification texts with increasing detail in a hierarchical structure, with each finer text building upon the coarser one. We then repeat the validation process, evaluating the new texts from coarsest to finest and selecting the best one. Any re-generated modification text that passes the validation process replaces the original text and concludes the process. If all generated texts fail the validation, the triplet is removed from the benchmark. Detailed process information is provided in the supplementary material. Besides, following MTCIR, biometric information is also removed from the refined benchmarks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.535, + 0.909, + 0.656 + ], + "angle": 0, + "content": "As illustrated in Fig. 5, up to \\(8\\%\\) of samples are omitted due to Claude's processing limitations related to harmful content. Additionally, \\(4 - 7\\%\\) of ambiguous samples that could not be effectively rewritten are removed. More than \\(40\\%\\) of the original samples are retained, and \\(40 - 55\\%\\) of triplets are successfully revised. The refined CIRR and Fashion-IQ datasets exhibit less ambiguity (Fig. 6), offering more robust evaluation benchmarks for the CIR community." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.671, + 0.649, + 0.689 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.698, + 0.907, + 0.803 + ], + "angle": 0, + "content": "Our proposed CoLLM framework adopts a two-stage training paradigm: pre-training and fine-tuning phases. In pretraining, CoLLM is exposed to a diverse corpus of image-capture pairs, enabling it to understand composed queries comprehensively. Subsequently, in the fine-tuning stage, we leverage our newly curated MTCIR datasets to verify their effective performance on complex multimodal tasks." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.815, + 0.838, + 0.831 + ], + "angle": 0, + "content": "5.1. Pre-Training on Image-Caption Pairs" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Training and Evaluation Datasets. The pre-training dataset consists of 5 million image-caption pairs, compiled from CC3M [51], LAION [50], and LLaVA-558K [34]. We evaluate our model on three CIR benchmarks: CIRCO [4]," + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3999" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.071, + 0.483, + 0.127 + ], + "angle": 0, + "content": "Table 4. Performance of BLIP-L and CoLLM (pre-trained) trained on synthetic triplet datasets. Bold values indicate the best score within each group. Models trained on MTCIR demonstrate superior performance compared to those trained on other datasets." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.13, + 0.485, + 0.232 + ], + "angle": 0, + "content": "
MethodDatasetCIRR ↑FIQ ↑
@1@10@50@10@50
BLIP-L [29]LaSCo [28]36.678.494.624.844.0
WebCoVR [59]39.378.994.726.743.3
MTCIR (ours)42.483.195.937.959.2
CoLLMLaSCo [28]43.284.195.738.560.1
MTCIR (ours)45.884.795.939.160.7
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.237, + 0.483, + 0.267 + ], + "angle": 0, + "content": "CIRR [37] test set, and Fashion-IQ [61] validation set. All datasets are filtered to exclude biometric information." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.271, + 0.484, + 0.348 + ], + "angle": 0, + "content": "Evaluation Metrics. We employ recall on the complete image index set as the evaluation metric for CIRR and Fashion-IQ. For CIRCO, we use mean Average Precision (mAP). We exclude the recall on subset for CIRR due to reliability concerns (see supplementary material for details)." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.35, + 0.455, + 0.365 + ], + "angle": 0, + "content": "Implementation Details. See supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.369, + 0.483, + 0.519 + ], + "angle": 0, + "content": "Quantitative Results. Table 2 compares our pre-trained model with other methods not trained on triplet datasets. We evaluate three variants of our architecture, each employing different vision encoders. CoLLM with the CLIP-B encoder achieves the second-best performance on the CIRCO and Fashion-IQ benchmarks while significantly improving on the CIRR benchmark. Although CIREVL [26] outperforms other methods, its use of the closed-source GPT-4 [1] introduces additional costs and increased inference latency, making it an unfair comparison." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.52, + 0.483, + 0.746 + ], + "angle": 0, + "content": "For larger vision encoders, our model consistently outperforms both non-LLM (e.g., Slerp-TAT [21]) and LLM (CIReVL and MCL [30]) methods across all benchmarks. Notably, our CLIP-L variant achieves state-of-the-art results on the CIRCO benchmark, while BLIP-L significantly improves CIRR and Fashion-IQ scores. The consistent gains across benchmarks and architectures validate the robustness and generalizability of our triplet synthesis approach. By effectively bridging the gap between abundant image-caption data and scarce triplet annotations, our method paves the way for more scalable and efficient training of CIR models. This innovative strategy represents a significant step forward in addressing one of the critical challenges in the field, Paving the way for larger-scale pre-training and more complex multimodal interactions." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.755, + 0.308, + 0.77 + ], + "angle": 0, + "content": "5.2. Fine-tuning on MTCIR" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Training Dataset and Benchmarks. Despite the impressive performance achieved using only image-caption datasets, we hypothesize that CoLLM's performance can be further enhanced through fine-tuning on our newly developed MTCIR dataset. For a fair comparison, we also train CoLLM and baseline models on other public datasets, including WebCoVR [59] and LaSCo [28]. Our evaluation focuses on the CIRR [37] test and Fashion-IQ [61] validation." + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.071, + 0.907, + 0.14 + ], + "angle": 0, + "content": "Table 5. Performance of models on refined CIRR and Fashion-IQ validation sets. Bold indicates the highest score, while underlined values represent the best metric within the same vision encoder group. CoLLM+MTCIR continues to outperform other configurations on these new benchmarks." + }, + { + "type": "table", + "bbox": [ + 0.513, + 0.143, + 0.912, + 0.303 + ], + "angle": 0, + "content": "
MethodDatasetRef. CIRR ↑Ref. FIQ ↑
@1@10@50@10@50
EVA-CLIP ViT-G/14 364 × 364
CoVR2 [58]WV-CC-VIR [58]56.191.197.954.272.9
OpenAI CLIP-L/16 224 × 224
MagicLens [64]MagicLens [64]42.386.397.045.568.1
CoLLMMTCIR (ours)46.587.496.148.368.6
BLIP-L/16 384 × 384; fine-tuned on COCO
BLIP-L [29]MTCIR (ours)53.889.997.754.874.3
CoLLMLaSCo [28]57.392.098.156.975.9
CoLLMMTCIR (ours)60.492.698.257.276.4
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.307, + 0.907, + 0.353 + ], + "angle": 0, + "content": "tion sets. We exclude CIRCO [4] due to potential domain overlap between COCO [32] images used in CIRCO and those in LaSCo [28], as well as in the pre-training datasets." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.356, + 0.878, + 0.372 + ], + "angle": 0, + "content": "Implementation Details. See supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.375, + 0.908, + 0.587 + ], + "angle": 0, + "content": "Quantitative Results. Table 3 presents the evaluation of models trained on publicly available synthetic CIR datasets, including our MTCIR. Fine-tuning CoLLM on MTCIR yields significant gains, improving CIRR@1 by \\(5\\%\\) over CoLLM-OpenAI-CLIP-L/14 in Table 2. This improvement reflects MTCIR's enhanced generalizability and our superior data construction. Our method achieves exceptional results on all benchmarks using the same vision encoder compared to previous approaches and models trained on other synthetic datasets. Notably, our fine-tuned CoLLM outperforms MagicLens CoCa-L [64] and CoVR2 [58], despite these models using larger, more advanced vision encoders and being trained on larger datasets. These results underscore the importance of data quality over quantity." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.589, + 0.909, + 0.817 + ], + "angle": 0, + "content": "MTCIR vs. Other CIR Datasets. Despite CoLLM's state-of-the-art performance, two questions remain: (i) Can MTCIR benefit other models? (ii) Can CoLLM achieve good performance when fine-tuned on other CIR datasets? To address these questions, we train a baseline BLIP-L [29] model on various CIR datasets, including MTCIR. Both LaSCo and MTCIR experiments are trained for one epoch. As shown in Table 4, we observe that: (i) BLIP-L shows significant improvement when using MTCIR as the training data, and (ii) CoLLM exhibits notable performance degradation when fine-tuned on the LaSCo dataset. These results demonstrate MTCIR's versatility as a plug-and-play dataset, capable of enhancing performance across various models. This underscores MTCIR's potential to make significant contributions to the CIR community." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.819, + 0.845, + 0.833 + ], + "angle": 0, + "content": "5.3. Re-evaluation on Refined Benchmarks" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Ambiguity in the CIRR and Fashion-IQ benchmarks potentially skews the metrics presented in previous tables, particularly for high-performing models. While CIRCO mitigates this issue, it still suffers from domain overlap with train-" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4000" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.135, + 0.071, + 0.438, + 0.085 + ], + "angle": 0, + "content": "Table 6. Ablation studies of proposed components." + }, + { + "type": "table_caption", + "bbox": [ + 0.138, + 0.089, + 0.435, + 0.102 + ], + "angle": 0, + "content": "(a) Reference image and modification text interpolation." + }, + { + "type": "table", + "bbox": [ + 0.122, + 0.104, + 0.454, + 0.173 + ], + "angle": 0, + "content": "
ImageTextCIRCO (mAP ↑)CIRR (Rec. ↑)
14.7170.1
✓14.5176.4
✓39.6183.3
✓✓52.8194.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.171, + 0.18, + 0.404, + 0.194 + ], + "angle": 0, + "content": "(b) Unimodal queries in Eq. (3) and Eq. (4)." + }, + { + "type": "table", + "bbox": [ + 0.122, + 0.196, + 0.454, + 0.252 + ], + "angle": 0, + "content": "
Image-onlyText-onlyCIRCO (mAP ↑)CIRR (Rec. ↑)
✓14.599.7
✓47.1196.3
✓✓52.8194.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.161, + 0.26, + 0.404, + 0.273 + ], + "angle": 0, + "content": "(c) Reference image embedding interpolation." + }, + { + "type": "table", + "bbox": [ + 0.118, + 0.275, + 0.45, + 0.322 + ], + "angle": 0, + "content": "
CIRCO (mAP ↑)CIRR (Rec. ↑)
Random In-Batch Sample46.7182.2
Nearest In-Batch Neighbor52.8194.5
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.33, + 0.482, + 0.389 + ], + "angle": 0, + "content": "ing images. To assess the impact of ambiguous samples, we re-evaluate the models on our newly refined CIRR and Fashion-IQ benchmarks. Table 5 presents the performance of these models on the revised benchmarks." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.391, + 0.483, + 0.588 + ], + "angle": 0, + "content": "Our analysis shows that model rankings remain largely consistent before and after the benchmark refinement. As the modification texts in the refined benchmarks provide additional nuance, we anticipated that models with more robust composed query understanding capabilities would distinguish themselves. Indeed, CoLLM consistently outperforms other models across various settings on the refined benchmarks. Notably, while BLIP-L (MTCIR) trailed CoLLM (MTCIR) by \\(3.4\\%\\) (CIRR@1)(Table 4), this performance gap widened to \\(6.9\\%\\) (CIRR@1) on the refined CIRR benchmark. This suggests CoLLM better captures nuanced modification text and distinguishes visual differences between source and target images." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.592, + 0.255, + 0.606 + ], + "angle": 0, + "content": "5.4. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.614, + 0.483, + 0.901 + ], + "angle": 0, + "content": "We conduct comprehensive ablation studies to evaluate the impact of our proposed components during the pre-training stage. Our experiments use the BLIP-L encoder, trained exclusively on the LLaVA-558k [34] dataset for one epoch. Table 6 presents the performance across various settings. We report the sum of mAP at \\( k = \\{5,10,25,50\\} \\) for CIRCO and Recall@ \\( k = \\{1,5,10,50\\} \\) for CIRR. Image and Text Interpolation are Crucial: As shown in Table 6a, employing Slerp for reference image embedding significantly enhances the model's learning capabilities. Applying interpolation on modification text further improves performance. Unimodal Queries are Beneficial: We investigate the necessity of unimodal queries in Eq. (3) and Eq. (4) and observe that omitting unimodal queries during training, especially text-only queries, substantially degrades performance (Table 6b). Nearest Neighbor is Essential: For reference image embedding interpolation, one might hypothesize that a random in-batch sample could suffice for interpolation. However, our study demonstrates that" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.071, + 0.907, + 0.14 + ], + "angle": 0, + "content": "Table 7. Performance of different LLMs with CLIP-L/14 as the visual encoder. **Bold and underline highlight the best and second best score.** indicates the original LLM for text generation; others are Large Language Embedding Models (LLEMs) fine-tuned for text retrieval." + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.144, + 0.912, + 0.233 + ], + "angle": 0, + "content": "
LLMCIRCO (mAP↑)CIRR (Rec.↑)FIQ (Rec.↑)
@5@10@50@1@10@50@10@50
Stella-Qwen2-1.5B [11]14.815.317.427.569.188.029.349.0
Mistral-7B* [24]19.820.222.729.672.591.329.549.3
E5-Mistral-7B-Inst [60]19.620.322.829.572.791.129.949.5
SFR-Embedding-2 [47]20.320.823.429.772.891.530.149.5
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.258, + 0.907, + 0.425 + ], + "angle": 0, + "content": "using the nearest image embedding neighbor yields significantly better performance (Table 6c). LLEMs Outperforms LLMs: We investigate Large Language Embedding Models (LLEMs), which are fine-tuned for text retrieval compared to their base LLMs. As shown in Table 7, both E5-Mistral-7B-Inst [60] and SFR-Embedding-2 outperform their LLM counterpart, Mistral-7B [24]. The superior performance of LLEMs demonstrates that models specifically tailored for embedding and retrieval tasks can offer substantial advantages in CIR applications compared to general-purpose language models." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.451, + 0.771, + 0.467 + ], + "angle": 0, + "content": "6. Limitations and Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.485, + 0.907, + 0.665 + ], + "angle": 0, + "content": "While these advancements significantly improve CIR performance, several areas warrant further investigation. Our work, limited to LLM/LLEMs, could benefit from exploring pre-trained MLLMs for enhanced understanding of CIR tasks. Additionally, our triplet synthesis method generates a single visual token, constraining the model's ability to process detailed image information. Future work should leverage the underutilized text category information in our synthetic datasets to improve model generalization. Lastly, our refined benchmarks, while more reliable, still contain noise from original image pairs, suggesting a need for further refinement of evaluation metrics." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.909, + 0.901 + ], + "angle": 0, + "content": "In conclusion, we present novel approaches to Composed Image Retrieval (CIR) that obviate the need for annotated datasets. Our contributions include: (1) an innovative triplet synthesis method utilizing image-caption pairs, (2) a new architecture leveraging LLM's embedding generation capabilities, (3) MTCIR, a diverse, human-aligned synthetic dataset, and (4) refined versions of CIRR and Fashion-IQ benchmarks, enhancing the reliability of evaluation metrics in the field. Our method consistently outperforms existing LLM and non-LLM baselines across popular benchmarks. Notably, MTCIR achieves superior results with only \\(10 - 20\\%\\) of the size of larger datasets, demonstrating a \\(1 - 15\\%\\) increase in Recall. These advancements collectively push the boundaries of CIR, offering more efficient and effective solutions for real-world applications." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "4001" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.116, + 0.484, + 0.184 + ], + "angle": 0, + "content": "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.187, + 0.482, + 0.214 + ], + "angle": 0, + "content": "[2] Anthropic. The claude 3 model family: Opus, sonnet, haiku, 2024. 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.218, + 0.483, + 0.259 + ], + "angle": 0, + "content": "[3] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. A clip-hitchhiker's guide to long video retrieval. arXiv preprint arXiv:2205.08508, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.262, + 0.483, + 0.302 + ], + "angle": 0, + "content": "[4] Alberto Baldrati, Lorenzo Agnolucci, Marco Bertini, and Alberto Del Bimbo. Zero-shot composed image retrieval with textual inversion. In ICCV, 2023. 2, 3, 5, 6, 7, 13, 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.306, + 0.482, + 0.347 + ], + "angle": 0, + "content": "[5] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In CVPR, 2023. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.35, + 0.482, + 0.376 + ], + "angle": 0, + "content": "[6] Tom B Brown. Language models are few-shot learners. In NeurIPS, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.38, + 0.482, + 0.461 + ], + "angle": 0, + "content": "[7] Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, et al. Palm: Scaling language modeling with pathways. Journal of Machine Learning Research, 2023. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.465, + 0.482, + 0.52 + ], + "angle": 0, + "content": "[8] Niv Cohen, Rinon Gal, Eli A. Meirom, Gal Chechik, and Yuval Atzmon. \"this is my unicorn, fluffy\": Personalizing frozen vision-language representations. In ECCV, 2022. 2, 3, 6, 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.524, + 0.482, + 0.579 + ], + "angle": 0, + "content": "[9] Yashar Deldjoo, Fatemeh Nazary, Arnau Ramisa, Julian Mcauley, Giovanni Pellegrini, Alejandro Bellogin, and Tommaso Di Noia. A review of modern fashion recommender systems. ACM Computing Surveys, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.582, + 0.482, + 0.622 + ], + "angle": 0, + "content": "[10] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009. 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.626, + 0.398, + 0.639 + ], + "angle": 0, + "content": "[11] Zhang Dun. Stella en qwen2 1.5b v5, 2024.8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.642, + 0.482, + 0.696 + ], + "angle": 0, + "content": "[12] Sonam Goenka, Zhaoheng Zheng, Ayush Jaiswal, Rakesh Chada, Yue Wu, Varsha Hedau, and Pradeep Natarajan. Fashionvlp: Vision language transformer for fashion retrieval with feedback. In CVPR, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.7, + 0.482, + 0.755 + ], + "angle": 0, + "content": "[13] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Batra, and Devi Parikh. Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In CVPR, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.758, + 0.482, + 0.813 + ], + "angle": 0, + "content": "[14] Geonmo Gu, Sanghyuk Chun, Wonjae Kim, HeeJae Jun, Yoohoon Kang, and Sangdoo Yun. Compodiff: Versatile composed image retrieval with latent diffusion. Transactions on Machine Learning Research, 2024. 2, 3, 5, 6, 16" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.816, + 0.482, + 0.857 + ], + "angle": 0, + "content": "[15] Geonmo Gu, Sanghyuk Chun, Wonjae Kim, Yoohoon Kang, and Sangdoo Yun. Language-only training of zero-shot composed image retrieval. In CVPR, 2024. 2, 3, 6, 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.86, + 0.482, + 0.899 + ], + "angle": 0, + "content": "[16] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 13" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.116, + 0.484, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.135 + ], + "angle": 0, + "content": "[17] Matthew Honnibal, Ines Montani, Sofie Van Landeghem, Adriane Boyd, et al. spacy: Industrial-strength natural language processing in python. 2020. 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.137, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[18] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In ICLR, 2022. 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.193, + 0.905, + 0.247 + ], + "angle": 0, + "content": "[19] Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Weli-hinda, Alan Hayes, Alec Radford, et al. Gpt-4o system card. arXiv preprint arXiv:2410.21276, 2024. 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.25, + 0.905, + 0.304 + ], + "angle": 0, + "content": "[20] Chuong Huynh, Yuqian Zhou, Zhe Lin, Connelly Barnes, Eli Shechtman, Sohrab Amirghodsi, and Abhinav Shrivastava. Simpson: Simplifying photo cleanup with single-click distracting object segmentation network. In CVPR, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.307, + 0.905, + 0.36 + ], + "angle": 0, + "content": "[21] Young Kyun Jang, Dat Huynh, Ashish Shah, Wen-Kai Chen, and Ser-Nam Lim. Spherical linear interpolation and text-anchoring for zero-shot composed image retrieval. In ECCV, 2024. 2, 3, 6, 7, 14, 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.363, + 0.905, + 0.404 + ], + "angle": 0, + "content": "[22] Young Kyun Jang, Junmo Kang, Yong Jae Lee, and Donghyun Kim. Mate: Meet at the embedding-connecting images with long texts. EMNLP Findings, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.407, + 0.905, + 0.46 + ], + "angle": 0, + "content": "[23] Young Kyun Jang, Donghyun Kim, Zihang Meng, Dat Huynh, and Ser-Nam Lim. Visual delta generator with large multi-modal models for semi-supervised composed image retrieval. In CVPR, 2024. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.463, + 0.905, + 0.53 + ], + "angle": 0, + "content": "[24] Albert Q Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, et al. Mistral 7b. arXiv preprint arXiv:2310.06825, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.533, + 0.905, + 0.601 + ], + "angle": 0, + "content": "[25] Ting Jiang, Minghui Song, Zihan Zhang, Haizhen Huang, Weiwei Deng, Feng Sun, Qi Zhang, Deqing Wang, and Fuzhen Zhuang. E5-v: Universal embeddings with multimodal large language models. arXiv preprint arXiv:2407.12580, 2024.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.604, + 0.905, + 0.645 + ], + "angle": 0, + "content": "[26] Shyamgopal Karthik, Karsten Roth, Massimiliano Mancini, and Zeynep Akata. Vision-by-language for training-free compositional image retrieval. *ICLR*, 2024. 3, 6, 7, 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.647, + 0.905, + 0.713 + ], + "angle": 0, + "content": "[27] Chankyu Lee, Rajarshi Roy, Mengyao Xu, Jonathan Raiman, Mohammad Shoeybi, Bryan Catanzaro, and Wei Ping. Nv-embed: Improved techniques for training llms as generalist embedding models. arXiv preprint arXiv:2405.17428, 2024.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.718, + 0.905, + 0.758 + ], + "angle": 0, + "content": "[28] Matan Levy, Rami Ben-Ari, Nir Darshan, and Dani Lischinski. Data roaming and quality assessment for composed image retrieval. AAAI, 2024. 2, 3, 5, 6, 7, 14, 15, 17, 18" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.761, + 0.905, + 0.814 + ], + "angle": 0, + "content": "[29] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, 2022. 2, 7, 14, 17, 18" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.817, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[30] Wei Li, Hehe Fan, Yongkang Wong, Yi Yang, and Mohan Kankanhalli. Improving context understanding in multimodal large language models via multimodal composition learning. In ICML, 2024. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[31] Zehan Li, Xin Zhang, Yanzhao Zhang, Dingkun Long, Pengjun Xie, and Meishan Zhang. Towards general text" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4002" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.093, + 0.482, + 0.12 + ], + "angle": 0, + "content": "embeddings with multi-stage contrastive learning. arXiv preprint arXiv:2308.03281, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.121, + 0.483, + 0.176 + ], + "angle": 0, + "content": "[32] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 5, 7, 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.177, + 0.482, + 0.232 + ], + "angle": 0, + "content": "[33] Aixin Liu, Bei Feng, Bing Xue, Bingxuan Wang, Bochao Wu, Chengda Lu, Chenggang Zhao, Chengqi Deng, Chenyu Zhang, Chong Ruan, et al. Deepseek-v3 technical report. arXiv preprint arXiv:2412.19437, 2024. 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.233, + 0.482, + 0.272 + ], + "angle": 0, + "content": "[34] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, 2023. 2, 3, 5, 6, 8, 14, 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.274, + 0.482, + 0.328 + ], + "angle": 0, + "content": "[35] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava-last: Improved reasoning,OCR, and world knowledge, 2024. 2, 3, 5, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.33, + 0.482, + 0.37 + ], + "angle": 0, + "content": "[36] Yikun Liu, Jiangchao Yao, Ya Zhang, Yan-Feng Wang, and Weidi Xie. Zero-shot composed text-image retrieval. In BMVC, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.372, + 0.482, + 0.426 + ], + "angle": 0, + "content": "[37] Zheyuan Liu, Cristian Rodriguez-Opazo, Damien Teney, and Stephen Gould. Image retrieval on real-life images with pretrained vision-and-language models. In ICCV, 2021. 2, 3, 5, 7, 13, 14, 17" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.428, + 0.482, + 0.495 + ], + "angle": 0, + "content": "[38] Feipeng Ma, Hongwei Xue, Guangting Wang, Yizhou Zhou, Fengyun Rao, Shilin Yan, Yueyi Zhang, Siying Wu, Mike Zheng Shou, and Xiaoyan Sun. Multi-modal generative embedding model. arXiv preprint arXiv:2405.19333, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.497, + 0.482, + 0.564 + ], + "angle": 0, + "content": "[39] Gabriel de Souza P Moreira, Radek Osmulski, Mengyao Xu, Ronay Ak, Benedikt Schifferer, and Even Oldridge. Nv-retriever: Improving text embedding models with effective hard-negative mining. arXiv preprint arXiv:2407.15831, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.567, + 0.482, + 0.608 + ], + "angle": 0, + "content": "[40] Niklas Muennighoff, Nouamane Tazi, Loic Magne, and Nils Reimers. Mteb: Massive text embedding benchmark. arXiv preprint arXiv:2210.07316, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.609, + 0.482, + 0.65 + ], + "angle": 0, + "content": "[41] Vicente Ordonez, Girish Kulkarni, and Tamara Berg. Im2text: Describing images using 1 million captioned photographs. NeurIPS, 2011. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.651, + 0.482, + 0.692 + ], + "angle": 0, + "content": "[42] Khoi Pham, Chuong Huynh, Ser-Nam Lim, and Abhinav Shrivastava. Composing object relations and attributes for image-text matching. In CVPR, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.692, + 0.482, + 0.762 + ], + "angle": 0, + "content": "[43] Bill Psomas, Ioannis Kakogeorgiou, Nikos Efthymiadis, Giorgos Tolias, Ondrej Chum, Yannis Avrithis, and Konstantinos Karantzalos. Composed image retrieval for remote sensing. In IGARSS 2024-2024 IEEE International Geoscience and Remote Sensing Symposium, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.763, + 0.482, + 0.83 + ], + "angle": 0, + "content": "[44] Filip Radenovic, Abhimanyu Dubey, Abhishek Kadian, Todor Mihaylov, Simon Vandenhende, Yash Patel, Yi Wen, Vignesh Ramanathan, and Dhruv Mahajan. Filtering, distillation, and hard negatives for vision-language pre-training. In CVPR, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[45] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 2, 5, 13, 14" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "[46] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.188 + ], + "angle": 0, + "content": "[47] Meng Rui, Liu Ye, Joty Shafiq Rayhan, Xiong Caiming, Zhou Yingbo, and Yavuz Semih. Sfr-embedding-2: Advanced text embedding with multi-stage training, 2024. 8, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.19, + 0.905, + 0.232 + ], + "angle": 0, + "content": "[48] Amrita Saha, Mitesh Khapra, and Karthik Sankaranarayanan. Towards building large scale multimodal domain-aware conversation systems. In AAAI, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.233, + 0.905, + 0.287 + ], + "angle": 0, + "content": "[49] Kuniaki Saito, Kihyuk Sohn, Xiang Zhang, Chun-Liang Li, Chen-Yu Lee, Kate Saenko, and Tomas Pfister. Pic2word: Mapping pictures to words for zero-shot composed image retrieval. CVPR, 2023. 3, 6, 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.288, + 0.905, + 0.357 + ], + "angle": 0, + "content": "[50] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021. 5, 6, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.358, + 0.905, + 0.412 + ], + "angle": 0, + "content": "[51] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In ACL, 2018. 5, 6, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.414, + 0.905, + 0.455 + ], + "angle": 0, + "content": "[52] Ken Shoemake. Animating rotation with quaternion curves. In Proceedings of the 12th annual conference on Computer graphics and interactive techniques, 1985. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.456, + 0.905, + 0.524 + ], + "angle": 0, + "content": "[53] Swetha Sirnam, Jinyu Yang, Tal Neiman, Mamshad Nayeem Rizve, Son Tran, Benjamin Yao, Trishul Chilimbi, and Mubarak Shah. X-former: Unifying contrastive and reconstruction learning for mllms. In European Conference on Computer Vision, pages 146-162. Springer, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.525, + 0.905, + 0.567 + ], + "angle": 0, + "content": "[54] Alane Suhr, Stephanie Zhou, Ally Zhang, Iris Zhang, Huajun Bai, and Yoav Artzi. A corpus for reasoning about natural language grounded in photographs. In ACL, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.568, + 0.905, + 0.621 + ], + "angle": 0, + "content": "[55] Yuanmin Tang, Jing Yu, Keke Gai, Jiamin Zhuang, Gang Xiong, Yue Hu, and Qi Wu. Context-i2w: Mapping images to context-dependent words for accurate zero-shot composed image retrieval. In AAAI, 2024. 2, 3, 6, 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.622, + 0.905, + 0.69 + ], + "angle": 0, + "content": "[56] Omkar Thawakar, Muzammal Naseer, Rao Muhammad Anwer, Salman Khan, Michael Felsberg, Mubarak Shah, and Fahad Shahbaz Khan. Composed video retrieval via enriched context and discriminative embeddings. CVPR, 2024. 4, 6, 16" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.692, + 0.905, + 0.761 + ], + "angle": 0, + "content": "[57] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.762, + 0.905, + 0.802 + ], + "angle": 0, + "content": "[58] Lucas Ventura, Antoine Yang, Cordelia Schmid, and Gül Varol. CoVR-2: Automatic data construction for composed video retrieval. IEEE TPAMI, 2024. 2, 3, 5, 6, 7, 14, 16, 18" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.803, + 0.905, + 0.845 + ], + "angle": 0, + "content": "[59] Lucas Ventura, Antoine Yang, Cordelia Schmid, and Güll Varol. CoVR: Learning composed video retrieval from web video captions. AAAI, 2024. 3, 4, 5, 6, 7, 15, 16, 17" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[60] Liang Wang, Nan Yang, Xiaolong Huang, Linjun Yang, Rangan Majumder, and Furu Wei. Improving text embeddings with large language models. arXiv preprint arXiv:2401.00368, 2023. 3, 8" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4003" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.16 + ], + "angle": 0, + "content": "[61] Hui Wu, Yupeng Gao, Xiaoxiao Guo, Ziad Al-Halah, Steven Rennie, Kristen Grauman, and Rogerio Feris. The fashion iq dataset: Retrieving images by combining side information and relative natural language feedback. CVPR, 2021. 2, 5, 7, 17" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.164, + 0.482, + 0.245 + ], + "angle": 0, + "content": "[62] Jinyu Yang, Jiali Duan, Son Tran, Yi Xu, Sampath Chanda, Liquun Chen, Belinda Zeng, Trishul Chilimbi, and Junzhou Huang. Vision-language pre-training with triple contrastive learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15671-15680, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.248, + 0.482, + 0.303 + ], + "angle": 0, + "content": "[63] ZHOU Yuqian, Chuong Huynh, Connelly Barnes, Elya Shechtman, Sohrab Amirghodsi, and Zhe Lin. Machine-learning models for distractor segmentation with reduced user interactions, 2024. US Patent App. 18/307,353. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.305, + 0.482, + 0.358 + ], + "angle": 0, + "content": "[64] Kai Zhang, Yi Luan, Hexiang Hu, Kenton Lee, Siyuan Qiao, Wenhu Chen, Yu Su, and Ming-Wei Chang. Magiclens: Self-supervised image retrieval with open-ended instructions. In ICML, 2024. 2, 3, 5, 6, 7, 16, 18" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.361, + 0.482, + 0.429 + ], + "angle": 0, + "content": "[65] Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen, Christopher Dewan, Mona Diab, Xian Li, Xi Victoria Lin, et al. Opt: Open pre-trained transformer language models. arXiv preprint arXiv:2205.01068, 2022. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.432, + 0.482, + 0.513 + ], + "angle": 0, + "content": "[66] Youcai Zhang, Xinyu Huang, Jinyu Ma, Zhaoyang Li, Zhaochuan Luo, Yanchun Xie, Yuzhuo Qin, Tong Luo, Yaqian Li, Shilong Liu, et al. Recognize anything: A strong image tagging model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1724-1732, 2024. 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.516, + 0.482, + 0.599 + ], + "angle": 0, + "content": "[67] Xinliang Zhu, Sheng-Wei Huang, Han Ding, Jinyu Yang, Kelvin Chen, Tao Zhou, Tal Neiman, Ouye Xie, Son Tran, Benjamin Yao, et al. Bringing multimodality to amazon visual search system. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 6390–6399, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.601, + 0.482, + 0.666 + ], + "angle": 0, + "content": "[68] Yutao Zhu, Huaying Yuan, Shuting Wang, Jiongnan Liu, Wenhan Liu, Chenlong Deng, Haonan Chen, Zhicheng Dou, and Ji-Rong Wen. Large language models for information retrieval: A survey. arXiv preprint arXiv:2308.07107, 2023. 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.666 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4004" + } + ] +] \ No newline at end of file diff --git a/2025/CoLLM_ A Large Language Model for Composed Image Retrieval/02729911-7a78-4d55-af78-f50b5e45a135_origin.pdf b/2025/CoLLM_ A Large Language Model for Composed Image Retrieval/02729911-7a78-4d55-af78-f50b5e45a135_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..2d86b7480038083e9b1242931782b6c260d2ed8e --- /dev/null +++ b/2025/CoLLM_ A Large Language Model for Composed Image Retrieval/02729911-7a78-4d55-af78-f50b5e45a135_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:52f4f7203b9cdb9910f8e178fa2edd3dd743ca6b7259bd6d3176e6e3bf0d0305 +size 1687119 diff --git a/2025/CoLLM_ A Large Language Model for Composed Image Retrieval/full.md b/2025/CoLLM_ A Large Language Model for Composed Image Retrieval/full.md new file mode 100644 index 0000000000000000000000000000000000000000..20988062cee3ca426013168c36f6a0b1219dd495 --- /dev/null +++ b/2025/CoLLM_ A Large Language Model for Composed Image Retrieval/full.md @@ -0,0 +1,373 @@ +# CoLLM: A Large Language Model for Composed Image Retrieval + +Chuong Huynh $^{1*}$ Jinyu Yang $^{2\dagger}$ Ashish Tawari $^{2}$ Mubarak Shah $^{2,3}$ Son Tran $^{2}$ Raffay Hamid $^{2}$ Trishul Chilimbi $^{2}$ Abhinav Shrivastava $^{1}$ + +1 University of Maryland, College Park 2 Amazon + +3 Center for Research in Computer Vision, University of Central Florida + +$^{1}\{\mathrm{chuonghm}, \mathrm{abhinav}\} @ \mathrm{cs.umd.edu}$ $^{2}\{\mathrm{viyjy}, \mathrm{atawari}, \mathrm{sontran}, \mathrm{raffay}, \mathrm{trishulc}\} @ \mathrm{amazon.com}$ shah@crcv.ucf.edu + +# Abstract + +Composed Image Retrieval (CIR) is a complex task that aims to retrieve images based on a multimodal query. Typical training data consists of triplets containing a reference image, a textual description of desired modifications, and the target image, which are expensive and time-consuming to acquire. The scarcity of CIR datasets has led to zero-shot approaches utilizing synthetic triplets or leveraging vision-language models (VLMs) with ubiquitous web-crawled image-caption pairs. However, these methods have significant limitations: synthetic triplets suffer from limited scale, lack of diversity, and unnatural modification text, while image-caption pairs hinder joint embedding learning of the multimodal query due to the absence of triplet data. Moreover, existing approaches struggle with complex and nuanced modification texts that demand sophisticated fusion and understanding of vision and language modalities. We present CoLLM, a one-stop framework that effectively addresses these limitations. Our approach generates triplets on-the-fly from image-caption pairs, enabling supervised training without manual annotation. We leverage Large Language Models (LLMs) to generate joint embeddings of reference images and modification texts, facilitating deeper multimodal fusion. Additionally, we introduce Multi-Text CIR (MTCIR), a large-scale dataset comprising 3.4M samples, and refine existing CIR benchmarks (CIRR and Fashion-IQ) to enhance evaluation reliability. Experimental results demonstrate that CoLLM achieves state-of-the-art performance across multiple CIR benchmarks and settings. MTCIR yields competitive results, with up to $15\%$ performance improvement. Our refined benchmarks provide more reliable evaluation metrics for CIR models, contributing to the advancement of this important field. Project page is at collm-cvpr25.github.io. + +![](images/d857becc33853ef401ca098ec992c3b7de0716d5988b6e9d34475f8aadfd2412.jpg) +(a) Composed Image Retrieval Example + +![](images/259b288b4e526d553b3518123107e72c38497de73df2a3acafdc4e85671a6249.jpg) +(b) Recall Sum of state-of-the-art methods on CIR benchmarks +Figure 1. (a) An example of CIR. (b) Recall Sum at $\{1,10,50\}$ for CIRR and $\{10,50\}$ for Fashion-IQ between CoLLM and state-of-the-art (SoTA) models under zero-shot settings. We evaluate two training scenarios: (i) without triplet data and (ii) with synthetic triplet data. + +# 1. Introduction + +Composed Image Retrieval (CIR) enhances traditional image retrieval [42, 62] by combining text and image queries, offering greater flexibility in search systems, with applications in e-commerce, fashion, and design [9, 12, 20, 63, 67]. As illustrated in Fig. 1a, CIR retrieves similar items, such as shirts, where the reference image provides a visual basis, and the modification text specifies desired modifications. By expanding the capabilities of conventional image or text-based searches, CIR allows for more nuanced and specific queries. However, this advanced approach also presents significant challenges compared to traditional image retrieval. + +Supervised CIR approaches face challenges in data ac + +quisition. They require high-quality CIR triplets (reference image, target image, and modification text), collected through a labor-intensive and expensive process. Consequently, existing CIR triplet datasets are limited in scale and domain coverage, restricting model generalizability. To overcome these limitations, recent CIR methods [14, 23, 28, 58, 64] have adopted zero-shot approaches [36]. These can be broadly categorized into two strategies: (i) leveraging vision-language models [29, 45, 53] (VLMs) for composed query embedding generation, and (ii) generating synthetic triplets for supervised training. VLM-based approaches utilize pre-trained models that already align vision and language features in latent space. These methods rely on large-scale image-caption pairs and follow two main directions, i.e., textual inversion [8, 15, 55] and direct interpolation [21]. Synthetic triplet generation approaches [23, 28, 64] employ Large Language Models (LLMs) [7, 57, 65] to generate modification text, with CompoDiff [14] further utilizing diffusion models [5, 46] to create synthetic reference-target image pairs. Despite the recent advances in CIR, several critical challenges persist: + +- [Data] Limitation-1: VLM-based methods may hinder efficient and effective composed query embedding learning by relying solely on image-caption pairs. +- [Data] Limitation-2: Existing synthetic triplet datasets often suffer from a lack of diversity and unnatural modification text. Moreover, these datasets are either very small or closed-source, impeding research progress in the field. +- [Model] Limitation-3: Current methods for composed query embeddings mainly use shallow transformer models or linear interpolation. While these methods are computationally efficient, they lack the ability to capture the full complexity of the composed understanding tasks. +- [Evaluation] Limitation-4: Existing CIR benchmarks are often compromised by noise, particularly in the form of ambiguous samples. Ambiguity arises when multiple target images can correctly match a single query, but only one is labeled as the ground truth, ignoring valid matches. Although CIRCO [4] attempts to address this issue, its efforts are limited in scale. This ambiguity in benchmarks hinders meaningful model evaluation and comparison. + +We propose CoLLM, an LLM-based CIR approach to address the aforementioned limitations. CoLLM tackles Limitation-1 by dynamically synthesizing triplets from image-caption pairs, introducing two key components: a reference image embedding synthesis and a modification text synthesis module. The former employs Spherical Linear Interpolation [52] (Slerp) to generate an intermediate embedding between a given image and its nearest neighbor within the training batch, serving as a synthesized reference image embedding. The latter utilizes a pre-defined text interpolation template to generate modification text based on the current caption and that of the nearest image neighbor. + +This strategy effectively leverages the vast amount of readily available image-caption pairs on the Internet, enabling a model's training in a supervised CIR manner. By doing so, CoLLM overcomes the scarcity of labeled triplet data and paves the way for more robust and scalable CIR models. + +To address Limitation-2, we introduce Multi-Text CIR (MTCIR), a synthetic dataset of 3.4M image pairs with 17.7M modification texts. MTCIR focuses on two often-overlooked aspects: image diversity and naturalistic modification texts. We curated images from diverse sources to ensure variety. For modification text generation, we employ a two-stage approach using Multi-modal LLM (MLLM) [34, 35] for detailed captioning and LLM for describing interception differences. Uniquely, MTCIR provides multiple short modification texts for each image pair, covering various attributes. This better reflects human query formulation, offering a more realistic, comprehensive training foundation for CIR models. + +To overcome Limitation-3, CoLLM harnesses the power of LLMs for composed query understanding. It is motivated by the extensive world knowledge embedded in pre-trained LLMs. With their deep semantic understanding, we posit that LLMs are superior to shallow transformers and simple embedding interpolation techniques in comprehending the intricate relationships between reference images and modification texts. By leveraging LLMs, CoLLM aims to capture nuanced semantic connections, enhancing the quality and relevance of composed image retrieval results. + +To address Limitation-4, we refine two popular CIR benchmarks: CIRR [37] and Fashion-IQ [61]. We use MLLMs to evaluate sample ambiguity in each benchmark and regenerate clear modification text for ambiguous ones. Our pipeline incorporates multiple validation steps to guarantee the enhanced quality of the refined samples. + +Our main contributions can be summarized as: (i) We propose a method to synthesize CIR triplets on-the-fly from image-caption pairs, which can even outperform models trained on CIR triplets, eliminating the need for costly CIR datasets. (ii) We collect and will release MTCIR, a new CIR triplet dataset covering 3.4 million image pairs with 17.7 million modification texts. To our knowledge, MTCIR is the largest open-source synthetic CIR dataset. (iii) We introduce an approach to leverage LLMs for composed query understanding, utilizing their instruction-following and embedding generation capabilities. (iv) We refine CIRR and Fashion-IQ to provide more robust evaluation benchmarks for the CIR community. Extensive experiments on popular CIR benchmarks and settings demonstrate the effectiveness of our model innovations and new datasets (Fig. 1b). + +# 2. Related Works + +Composed Image Retrieval. Composed Image Retrieval (CIR) has garnered significant attention due to its flexibil + +ity in search systems [37, 43, 48]. Zero-shot CIR methods have been extensively explored, with textual inversion [4, 8, 15, 49, 55] emerging as a prominent technique. This approach maps image encoder outputs to text encoder inputs, creating pseudo-word tokens. SlerpTAT [21] emphasizes the text encoder's importance by re-aligning visual and textual embeddings. While most zero-shot CIR models utilize image-caption pairs for training, our work introduces an innovative method to synthesize triplets from these pairs during training. This method enables supervised CIR style training, enhancing the model's understanding of composed queries without relying on real CIR datasets. + +Recent works enhance enhancing models' language understanding using Large Language Models (LLMs) [26, 30], primarily leveraging their text generation abilities. MCL [30] composes image-text query in the LLM input space and aligns joint embedding with the target image caption. However, this may introduce noise due to limited visual information in the target image captions. CIReVL [26] employs LLMs to generate captions based on the reference image and modification text, subsequently retrieving the target image using a text-to-image retrieval approach. Our model also utilizes LLMs but differs by directly producing a composed query embedding for target image retrieval, potentially reducing intermediate steps and associated errors. + +CIR with Synthetic Datasets. The scarcity of supervised CIR datasets has prompted recent studies to leverage generative models for synthetic triplet creation. CompoDiff [14] employs image editing pipelines [5, 46] to generate target images, though their nature limits performance. Other approaches [23, 28, 58, 59, 64] utilize LLMs to generate modification text for real image pairs. Our work introduces a novel two-stage approach, combining MLLMs [34, 35] for detailed captioning and LLMs [1, 2] for describing interception differences. Distinctly, our synthetic triplets provide multiple concise modification texts for each image pair, encompassing various attributes, in contrast to the conventional single, lengthy modification text. This enables a more nuanced representation of image modifications, potentially improving CIR model performance and versatility. + +Large Language Models. Recent advancements in LLMs have expanded their applications beyond text generation to include image understanding [34, 35] and text embedding generation [40, 60, 68]. Large Language Embedding Models (LLEMs) are specialized versions trained using contrastive learning, leveraging LLM knowledge for embedding generation while often disabling text generation capabilities. To enhance text embedding quality, [27, 31, 39] propose removing causal attention in LLMs, thereby improving text information encoding efficiency and training these modified LLMs on text retrieval datasets. This text retrieval capability has been further extended to text-image retrieval by aligning visual and LLM text embed + +![](images/9473b742177963d4ac05ea3ed412ecb46f2b36bba425554653d14f6327a0d981.jpg) + +![](images/77fbe85028cc23d395fcb33d1ee4270f55402926a2d57f1a408f56357b5a7967.jpg) +Figure 2. An overview of our model and training strategies when using (a) image-caption pairs and (b) CIR triplets. + +ding spaces [22, 25, 38]. These developments demonstrate the potential of LLMs for composed query understanding in CIR tasks. Our work builds upon this foundation, extending LLMs/LLEMs to generate composed query embeddings by incorporating reference images and modification text. This novel approach leverages the multimodal capabilities of LLMs to enhance the performance of CIR systems, leading to more accurate and context-aware image retrieval. + +# 3. Method + +This section outlines our methodology, starting with a brief overview of the model architecture, followed by a formal CIR problem definition. We then detail our triplet synthesis strategy, including reference image embedding and modification text synthesis, using image-caption pairs. Lastly, we describe our LLM-based query composition approach. + +# 3.1. Model Architecture + +As illustrated in Fig. 2, CoLLM consists of several essential components: (1) a vision encoder $f(\cdot)$ for extracting image features; (2) modules for synthesizing reference image embeddings and modification text; (3) an image adapter $g(\cdot)$ that maps visual features into the language model's semantic space; (4) a LLM $\Phi(\cdot)$ that processes multimodal queries and (5) a projection layer $proj(\cdot)$ (omitted from the figure for simplicity) that maps the LLM output to a suitable representation for retrieval. It is important to note that Fig. 2 (a) and Fig. 2 (b) illustrate architectures designed for two distinct input formats. The former is tailored for image-caption pairs, while the latter is optimized for CIR triplets. + +# 3.2. CoLLM with Image-Caption Pairs as Input + +Let $\mathcal{X} = \{(v_i, w_i)\}_{i=1}^N$ denote a set of image-caption pairs, where $v_i$ and $w_i$ represent the image and caption of the $i^{th}$ sample, respectively. To effectively leverage the vast amount of image-caption pairs, we introduce a novel ap + +![](images/9e6e609e23dbeec35a20b5cef8f698bc61dccdd2111547532ea31992ccd53e63.jpg) +Figure 3. An overview of reference image embedding synthesis and modification text synthesis. The red-framed image represents the nearest neighbor of the augmented image in the training batch. + +proach that synthesizes a CIR triplet on-the-fly for each image-caption pair $(v_{i},w_{i})$ . In this process, $v_{i}$ serves as the target image, while two distinct modules generate the remaining components: (i) a reference image embedding synthesizer and (ii) a modification text synthesizer (Fig. 2a). Notably, we do not generate an actual image for reference image synthesis. Instead, we synthesize the reference image embedding, which is computationally more efficient and allows seamless integration into the CIR pipeline. This formulation enables CoLLM to exploit the rich information in image-caption pairs, transforming them into CIR triplets that can be used for training. The following subsections detail the specific mechanisms for synthesizing the reference image embedding and the modification text. + +Reference Image Embedding Synthesis. Our reference image embedding synthesis process is illustrated in Fig. 3 (left). We begin by applying an augmentation $t \sim \mathcal{T}$ to the input image $v_{i}$ , where $\mathcal{T}$ represents a family of augmentations. The resulting augmented image is denoted as $v_{i}^{\prime}$ with its corresponding embedding $\mathbf{h}_i^\prime = f(v_i^\prime)$ . For $\mathbf{h}_i^\prime$ , we identify its in-batch nearest neighbor $v_{j}^{\prime}$ with $j$ defined as: + +$$ +j = \underset {j \neq i; j \in \{1, \dots \mathcal {B} \}} {\arg \max } \operatorname {s i m} \left(\mathbf {h} _ {i} ^ {\prime}, \mathbf {h} _ {j} ^ {\prime}\right), \tag {1} +$$ + +where $\mathcal{B}$ is the batch size, $\mathbf{h}_j^{\prime} = f(v_j^{\prime})$ , and $\mathrm{sim}(\mathbf{u},\mathbf{v}) = \mathbf{u}^T\mathbf{v} / (\| \mathbf{u}\| \| \mathbf{v}\|)$ denotes cosine similarity. + +We then employ Spherical Linear Interpolation (Slerp) [52] to synthesize the reference image embedding by interpolating between $\mathbf{h}_i^{\prime}$ and $\mathbf{h}_j^{\prime}$ . The synthesized reference image embedding is obtained as: + +$$ +\theta = \operatorname {a r c c o s} \left(\mathbf {h} _ {i} ^ {\prime} \cdot \mathbf {h} _ {j} ^ {\prime}\right) +$$ + +$$ +\mathbf {h} _ {i} ^ {*} = \frac {\sin (\alpha \theta)}{\sin \theta} \mathbf {h} _ {i} ^ {\prime} + \frac {\sin ((1 - \alpha) \theta)}{\sin \theta} \mathbf {h} _ {j} ^ {\prime}, \tag {2} +$$ + +where $\alpha \in [0,1]$ is a hyper-parameter controlling the interpolation strength. The intuition behind this approach is that an image with embedding $\mathbf{h}_i^*$ shares certain visual similarities with the target image $v_{i}$ . Specifically, a larger $\alpha$ results + +in $\mathbf{h}_i^*$ more closely resembling $\mathbf{h}_i'$ , allowing for fine-grained control over the synthesized embedding. + +Modification Text Synthesis. For the augmented image $v_{i}^{\prime}$ , its nearest neighbor is $(v_{j}^{\prime}, w_{j})$ . To generate the modification text $w_{i}^{*}$ , we use text interpolation to combine $w_{i}$ and $w_{j}$ using pre-defined templates, as illustrated in Fig. 3 (right). The complete set of templates is provided in the supplementary material. During training, we randomly select a template for each sample to ensure diversity in the synthesized modification text. It is worth noting that we do not simply use $w_{i}$ as the modification text for two primary reasons: (i) $w_{i}$ alone fails to capture the visual differences between the reference image and the target image. (ii) Using only $w_{i}$ could lead to model cheating, where it learns to rely solely on $w_{i}$ for retrieval while ignoring the reference image. Our synthesis approach, in contrast, mimics real-world modification text, which often describes both similarities and differences between the reference and target images. Furthermore, this strategy forces the model to learn composed query embeddings by considering both the reference image and the modification text simultaneously. + +By combining the synthesized reference embedding $\mathbf{h}_i^*$ and the interpolated modification text $w_{i}^{*}$ , our method generates diverse CIR triplets from image-caption pairs. This approach enables effective training of the CoLLM model on large-scale image-caption datasets, effectively mitigating the dependency on scarce labeled triplet data. + +Query Composition with LLM. To construct the composed query, we utilize a pre-trained LLM that processes the synthesized reference image embedding $\mathbf{h}_i^*$ , image caption $w_i$ , and the synthesized modification text $w_i^*$ . We define three distinct composed embeddings: + +$$ +\mathbf {c} _ {i} ^ {v} = p \left(\Phi \left(g \left(\mathbf {h} _ {i} ^ {*}\right)\right)\right) \tag {3} +$$ + +$$ +\mathbf {c} _ {i} ^ {w} = p \left(\Phi \left(w _ {i}\right)\right) \tag {4} +$$ + +$$ +\mathbf {c} _ {i} = p \left(\Phi \left([ g \left(\mathbf {h} _ {i} ^ {*}\right); w _ {i} ^ {*} ]\right)\right) \tag {5} +$$ + +where $[\cdot ]$ denotes an instruction template that combine two modalities (see supplementary material). + +Training Objective. We employ a contrastive loss $\mathcal{L}_{cl}$ [44], consistent with previous works [56, 59], to compute the loss between query embeddings $\{\mathbf{c}_i^v,\mathbf{c}_i^w,\mathbf{c}_i\}$ and the target image embedding $\mathbf{z}_i = f(v_i)$ . The final loss for each sample during pre-training is defined as: + +$$ +\mathcal {L} = \frac {1}{3} \left(\mathcal {L} _ {c l} \left(\mathbf {c} _ {i} ^ {v}, \mathbf {z} _ {i}\right) + \mathcal {L} _ {c l} \left(\mathbf {c} _ {i} ^ {w}, \mathbf {z} _ {i}\right) + \mathcal {L} _ {c l} \left(\mathbf {c} _ {i}, \mathbf {z} _ {i}\right)\right) \tag {6} +$$ + +This formulation encourages the model to learn discriminative representations for different query types while maintaining consistency with the target image embedding. By combining losses from image-to-image $(\mathbf{c}_i^v)$ , text-to-image $(\mathbf{c}_i^w)$ , and composed $(\mathbf{c}_i)$ queries, we ensure that the model learns to effectively process and align various input modalities for compositional image retrieval. + +Table 1. Comparison of synthetic CIR training datasets. + +
DatasetPublicReference-Target Image PairModification Text Generation
TypeSource# Pairs# EntitiesLLMLLM input# Text# Text/Pair
LaSCo [28]✓imageVQAv2 [13]360K82KGPT-3 [6]question-answer360K1
VDG [23]✓imageNLVR2 [54], COCO [32]467K183KVision-LLaMA2image pair523K1.12
WebCoVR [59]✓videoWebVid2M [3]1.6M131KCustom LLMcaption1.6M1
CC-CoIR [58]✓imageCC3M [51]3.3M357KCustom LLMcaption3.3M1
SynthTriplets [14]synth. imageSD [46], IP2P [5]18.8M37.6MOPT-6.7B [65]-18.8M1
MagicLens [64]imageWeb crawled36.5M-PaLM [7]image metadata36.5M1
MTCIR (Ours)✓imageLLaVA-558k [34]3.4M423KSonnet 3 [2]synth. caption17.7M5.18
+ +Reference Image +Category +Modification Text +![](images/393b355a1fc2f9cf39fd0ea505f05b23f01f7cacb1e97084e954d00c07e6619a.jpg) ++ Object Added +Object Removed +Add a person's hand holding the cup. +Modify the liquid being poured. +Remove the social media text overlay. + +Target Image +![](images/dd58bd807d019a2cdf28e643f4181a82c4c754ec2d6f6cd1f3e93c9442c9b365.jpg) +$\mathbf{3}$ Attribute Changed + +# 3.3. CoLLM with CIR Triplet as Input + +For CIR triplet inputs, our model design is shown in Fig. 2b. Let $\mathcal{X} = \{(v_i^r,v_i^t,w_i)\}_{i = 1}^N$ be a set of triplets, where $v_{i}^{r}$ $v_{i}^{t}$ , and $w_{i}$ represent the reference image, target image, and modification text of the $i^{th}$ sample, respectively. The reference embedding is obtained as $\mathbf{h}_i = f(v_i^r)$ . The composed embedding is then computed as $\mathbf{c}_i = p(\Phi ([g(\mathbf{h}_i);w_i])$ . The training objective is $\mathcal{L} = \mathcal{L}_{cl}(\mathbf{c}_i,\mathbf{z}_i)$ , where $\mathbf{z}_i = f(v_i^t)$ denotes the target embedding. + +# 4. Dataset Construction + +This section outlines the data construction process for two primary objectives: (1) creating a novel CIR training dataset called Multi-Text CIR (MTCIR) and (2) refining widely used CIR benchmarks, specifically CIRR [37] and FashionIQ [61]. We detail the methodologies for each task, highlighting the improvements and innovations. + +# 4.1. Multi-Text CIR (MTCIR) Dataset + +Existing CIR triplet datasets [23, 28, 59] lack diversity and contain unnatural modification text. We introduce Multi-Text CIR (MTCIR) to address these limitations. + +To enhance image diversity, we source images from the LLaVA-558k dataset [34] filtered from image caption datasets [41, 50, 51] based on noun-phrase frequency for broad concept coverage. We pair images using CLIP [45] visual similarity metrics, following CIRR's grouping approach, yielding 3.4M pairs from approximately 423K images. Leveraging the detailed captions generated by LLaVA-Next-34B [35] for each image in the LLaVA-558k dataset, we employ Claude 3 Sonnet [2] + +![](images/e47c0a1e12446a6ddd82a5b20250169c508796a04007e3a4b35f9b13485c9e44.jpg) +Unchanged +(a) CIRR + +![](images/2c16b8e1361659c01de2b3d8d7b5e1c425d74ed5dde92b11fd7245a35a6c2859.jpg) +Regenerate +(b) FIQ-Dress + +![](images/5b04838b1d5c159594bbb52c8b859c4a3295a1644a686d8a85eea5d0cea863cc.jpg) +Remove +(c) FIQ-Shirt +Cannot process + +![](images/3c3f65416323df20ea4b3345010a918dc2bc39886d72cdee015d046a611b79e9.jpg) +(d) FIQ-Toptee + +![](images/44b80bdf9e6caa3bf59217f12cc9164594365c38dbcb6efe490fd782bbfb0bb2.jpg) +Figure 4. An example from the MTCIR dataset. Each sample contains multiple short texts describing different modifications. +Figure 5. Statistical analysis of the refinement process for CIRR and FashionIQ (FIQ) datasets. +Reference Image +Original Text +Regenerated Text + +Target Image +One puppy looking at the camera. +![](images/cd682d24ce9cdbb9b9df1a60dadafb80979bc461311654ac5467f738a2f8fbf1.jpg) +The dog is indoors on a tiled floor with a +food bowl nearby. + +Figure 6. An example from the refined CIRR dataset. The original modification text yields multiple plausible results (shown as similar images to the target on the right). In contrast, the newly generated text accurately captures the distinctive features of the target image, differentiating it from other candidates. + +(20240229-v1:0) to generate modification text for each image pair, using these captions as input. To control detail levels in Claude outputs, we define six modification text categories in the prompt, inspired by prior research [4, 37]. This approach maintains diverse aspects and prevents repetition. We also incorporate examples from the CIRR dataset to ensure human-like modification text. Detailed prompting strategies are provided in the supplementary material. + +Our method generates multiple brief, focused texts for each image pair, each highlighting a unique aspect. This approach captures all changes between two images without producing overly long sentences, better aligning with human-written modification text. These short texts can be combined for training, enhancing dataset diversity. We obtained 17.7M text samples for the 3.4M image pairs, averaging 5.18 short sentences per pair. Fig. 4 illustrates an example, showcasing modification texts corresponding to specific categories and covering diverse attributes. + +As shown in Table 1, MTCIR offers several advantages over existing synthetic CIR datasets: (i) Unlike [14], which + +Table 2. Performance comparison of various models and visual encoders on CIR benchmarks, evaluated without training on triplet datasets. Bold and underlined values denote the best and second-best scores within each vision encoder group. Models incorporating LLMs in their architectures are marked with $\star$ . Results reproduced by our team are indicated by $\ddagger$ . + +
MethodCIRCO (mAP↑)CIRR (Rec.↑)FIQ (Rec.↑)
@5@10@50@1@10@50@10@50
OpenAI CLIP-B/32
PALAVRA [8]4.65.36.816.658.584.019.837.3
SEARLE [4]9.49.911.824.066.889.822.942.5
Slerp-TAT [21]9.310.312.328.268.888.523.044.0
CIREVL* [26]14.915.417.823.966.087.028.349.4
CoLLM*12.913.215.028.671.892.724.845.2
OpenAI CLIP-L/14
Pic2World [49]8.79.511.323.965.387.824.743.7
SEARLE [4]11.712.715.124.266.388.825.646.2
LinCIR‡[15]13.013.916.224.666.988.826.446.6
ContextI2W [55]13.013.816.025.768.689.627.848.9
MCL* [30]17.818.421.825.969.491.0--
Slerp-TAT [21]18.519.421.430.970.989.228.347.6
CIREVL* [26]18.619.021.824.664.986.328.648.6
CoLLM*20.320.823.429.772.891.530.149.5
BLIP-L/16
Slerp-TAT [21]17.818.421.134.072.788.932.853.3
CoLLM*19.720.423.135.078.694.234.656.0
+ +uses synthetic images, MTCIR uses real images covering diverse concepts and domains. (ii) MTCIR provides multiple modification texts for each image pair, offering more natural descriptions and training flexibility. (iii) MTCIR is the largest public dataset to advance CIR research. + +To prevent the leakage of biometric information, we further process MTCIR at both the image and text levels. For images, we apply face blurring to all instances. For text, we remove any content containing keywords related to human attributes (e.g., skin, hair, gender, age, race). + +# 4.2. Refined CIRR and Fashion-IQ Datasets + +Current CIR benchmarks often suffer from label ambiguity, where modification text lacks clarity, potentially matching multiple target images. To address this issue, we propose a refinement method using Claude 3 Sonnet [2] (20240229-v1:0) to enhance the validation sets of two well-known benchmarks: CIRR and Fashion-IQ. Each sample in these validation sets comprises a triplet containing a reference image, target image, and modification text. + +Our refinement process, applicable to any benchmark, consists of validation, re-generation, and re-validation steps. For each triplet, we first identify the top-3 visually similar images to the target image based on CLIP visual scores, serving as hard negative samples for ambiguity assessment. We employ Claude 3 Sonnet to evaluate each triplet's ambiguity by attempting to identify the target image from these hard negatives. Triplets correctly identified by Claude 3 Sonnet are considered "good" samples and re + +Table 3. Performance of models trained on synthetic triplet datasets. Bold indicates the best method overall, while underline highlights the best method within the same vision encoder. Our CoLLM outperforms comparable methods in most metrics. + +
MethodDatasetCIRR ↑FIQ ↑
@1@10@50@10@50
CoCa-L/18 288 × 288
MagicLens [64]MagicLens [64]33.377.994.438.158.3
EVA-CLIP ViT-G/14 364 × 364
CoVR2 [58]WV-CC-CoVIR [58]43.784.096.138.258.4
OpenAI CLIP-L/14 224 × 224
CompoDiff [14]SynTrip18M [14]18.270.890.336.048.6
MagicLens [64]MagicLens [64]30.174.492.630.752.5
CoLLMMTCIR (ours)34.777.093.132.954.2
BLIP-L/16 384 × 384; fine-tuned on COCO
CASE [28]LaSCo [28]35.478.594.6--
Omkar et al. [56]WebCoVR [59]40.178.994.730.346.5
CoLLMMTCIR (ours)45.884.795.839.160.7
+ +main unchanged. For "bad" samples, we use Claude 3 Sonnet to re-generate the modification text. The model produces three new modification texts with increasing detail in a hierarchical structure, with each finer text building upon the coarser one. We then repeat the validation process, evaluating the new texts from coarsest to finest and selecting the best one. Any re-generated modification text that passes the validation process replaces the original text and concludes the process. If all generated texts fail the validation, the triplet is removed from the benchmark. Detailed process information is provided in the supplementary material. Besides, following MTCIR, biometric information is also removed from the refined benchmarks. + +As illustrated in Fig. 5, up to $8\%$ of samples are omitted due to Claude's processing limitations related to harmful content. Additionally, $4 - 7\%$ of ambiguous samples that could not be effectively rewritten are removed. More than $40\%$ of the original samples are retained, and $40 - 55\%$ of triplets are successfully revised. The refined CIRR and Fashion-IQ datasets exhibit less ambiguity (Fig. 6), offering more robust evaluation benchmarks for the CIR community. + +# 5. Experiments + +Our proposed CoLLM framework adopts a two-stage training paradigm: pre-training and fine-tuning phases. In pretraining, CoLLM is exposed to a diverse corpus of image-capture pairs, enabling it to understand composed queries comprehensively. Subsequently, in the fine-tuning stage, we leverage our newly curated MTCIR datasets to verify their effective performance on complex multimodal tasks. + +# 5.1. Pre-Training on Image-Caption Pairs + +Training and Evaluation Datasets. The pre-training dataset consists of 5 million image-caption pairs, compiled from CC3M [51], LAION [50], and LLaVA-558K [34]. We evaluate our model on three CIR benchmarks: CIRCO [4], + +Table 4. Performance of BLIP-L and CoLLM (pre-trained) trained on synthetic triplet datasets. Bold values indicate the best score within each group. Models trained on MTCIR demonstrate superior performance compared to those trained on other datasets. + +
MethodDatasetCIRR ↑FIQ ↑
@1@10@50@10@50
BLIP-L [29]LaSCo [28]36.678.494.624.844.0
WebCoVR [59]39.378.994.726.743.3
MTCIR (ours)42.483.195.937.959.2
CoLLMLaSCo [28]43.284.195.738.560.1
MTCIR (ours)45.884.795.939.160.7
+ +CIRR [37] test set, and Fashion-IQ [61] validation set. All datasets are filtered to exclude biometric information. + +Evaluation Metrics. We employ recall on the complete image index set as the evaluation metric for CIRR and Fashion-IQ. For CIRCO, we use mean Average Precision (mAP). We exclude the recall on subset for CIRR due to reliability concerns (see supplementary material for details). + +Implementation Details. See supplementary material. + +Quantitative Results. Table 2 compares our pre-trained model with other methods not trained on triplet datasets. We evaluate three variants of our architecture, each employing different vision encoders. CoLLM with the CLIP-B encoder achieves the second-best performance on the CIRCO and Fashion-IQ benchmarks while significantly improving on the CIRR benchmark. Although CIREVL [26] outperforms other methods, its use of the closed-source GPT-4 [1] introduces additional costs and increased inference latency, making it an unfair comparison. + +For larger vision encoders, our model consistently outperforms both non-LLM (e.g., Slerp-TAT [21]) and LLM (CIReVL and MCL [30]) methods across all benchmarks. Notably, our CLIP-L variant achieves state-of-the-art results on the CIRCO benchmark, while BLIP-L significantly improves CIRR and Fashion-IQ scores. The consistent gains across benchmarks and architectures validate the robustness and generalizability of our triplet synthesis approach. By effectively bridging the gap between abundant image-caption data and scarce triplet annotations, our method paves the way for more scalable and efficient training of CIR models. This innovative strategy represents a significant step forward in addressing one of the critical challenges in the field, Paving the way for larger-scale pre-training and more complex multimodal interactions. + +# 5.2. Fine-tuning on MTCIR + +Training Dataset and Benchmarks. Despite the impressive performance achieved using only image-caption datasets, we hypothesize that CoLLM's performance can be further enhanced through fine-tuning on our newly developed MTCIR dataset. For a fair comparison, we also train CoLLM and baseline models on other public datasets, including WebCoVR [59] and LaSCo [28]. Our evaluation focuses on the CIRR [37] test and Fashion-IQ [61] validation. + +Table 5. Performance of models on refined CIRR and Fashion-IQ validation sets. Bold indicates the highest score, while underlined values represent the best metric within the same vision encoder group. CoLLM+MTCIR continues to outperform other configurations on these new benchmarks. + +
MethodDatasetRef. CIRR ↑Ref. FIQ ↑
@1@10@50@10@50
EVA-CLIP ViT-G/14 364 × 364
CoVR2 [58]WV-CC-VIR [58]56.191.197.954.272.9
OpenAI CLIP-L/16 224 × 224
MagicLens [64]MagicLens [64]42.386.397.045.568.1
CoLLMMTCIR (ours)46.587.496.148.368.6
BLIP-L/16 384 × 384; fine-tuned on COCO
BLIP-L [29]MTCIR (ours)53.889.997.754.874.3
CoLLMLaSCo [28]57.392.098.156.975.9
CoLLMMTCIR (ours)60.492.698.257.276.4
+ +tion sets. We exclude CIRCO [4] due to potential domain overlap between COCO [32] images used in CIRCO and those in LaSCo [28], as well as in the pre-training datasets. + +Implementation Details. See supplementary material. + +Quantitative Results. Table 3 presents the evaluation of models trained on publicly available synthetic CIR datasets, including our MTCIR. Fine-tuning CoLLM on MTCIR yields significant gains, improving CIRR@1 by $5\%$ over CoLLM-OpenAI-CLIP-L/14 in Table 2. This improvement reflects MTCIR's enhanced generalizability and our superior data construction. Our method achieves exceptional results on all benchmarks using the same vision encoder compared to previous approaches and models trained on other synthetic datasets. Notably, our fine-tuned CoLLM outperforms MagicLens CoCa-L [64] and CoVR2 [58], despite these models using larger, more advanced vision encoders and being trained on larger datasets. These results underscore the importance of data quality over quantity. + +MTCIR vs. Other CIR Datasets. Despite CoLLM's state-of-the-art performance, two questions remain: (i) Can MTCIR benefit other models? (ii) Can CoLLM achieve good performance when fine-tuned on other CIR datasets? To address these questions, we train a baseline BLIP-L [29] model on various CIR datasets, including MTCIR. Both LaSCo and MTCIR experiments are trained for one epoch. As shown in Table 4, we observe that: (i) BLIP-L shows significant improvement when using MTCIR as the training data, and (ii) CoLLM exhibits notable performance degradation when fine-tuned on the LaSCo dataset. These results demonstrate MTCIR's versatility as a plug-and-play dataset, capable of enhancing performance across various models. This underscores MTCIR's potential to make significant contributions to the CIR community. + +# 5.3. Re-evaluation on Refined Benchmarks + +Ambiguity in the CIRR and Fashion-IQ benchmarks potentially skews the metrics presented in previous tables, particularly for high-performing models. While CIRCO mitigates this issue, it still suffers from domain overlap with train- + +Table 6. Ablation studies of proposed components. +(a) Reference image and modification text interpolation. + +
ImageTextCIRCO (mAP ↑)CIRR (Rec. ↑)
14.7170.1
✓14.5176.4
✓39.6183.3
✓✓52.8194.5
+ +(b) Unimodal queries in Eq. (3) and Eq. (4). + +
Image-onlyText-onlyCIRCO (mAP ↑)CIRR (Rec. ↑)
✓14.599.7
✓47.1196.3
✓✓52.8194.5
+ +(c) Reference image embedding interpolation. + +
CIRCO (mAP ↑)CIRR (Rec. ↑)
Random In-Batch Sample46.7182.2
Nearest In-Batch Neighbor52.8194.5
+ +ing images. To assess the impact of ambiguous samples, we re-evaluate the models on our newly refined CIRR and Fashion-IQ benchmarks. Table 5 presents the performance of these models on the revised benchmarks. + +Our analysis shows that model rankings remain largely consistent before and after the benchmark refinement. As the modification texts in the refined benchmarks provide additional nuance, we anticipated that models with more robust composed query understanding capabilities would distinguish themselves. Indeed, CoLLM consistently outperforms other models across various settings on the refined benchmarks. Notably, while BLIP-L (MTCIR) trailed CoLLM (MTCIR) by $3.4\%$ (CIRR@1)(Table 4), this performance gap widened to $6.9\%$ (CIRR@1) on the refined CIRR benchmark. This suggests CoLLM better captures nuanced modification text and distinguishes visual differences between source and target images. + +# 5.4. Ablation Studies + +We conduct comprehensive ablation studies to evaluate the impact of our proposed components during the pre-training stage. Our experiments use the BLIP-L encoder, trained exclusively on the LLaVA-558k [34] dataset for one epoch. Table 6 presents the performance across various settings. We report the sum of mAP at $k = \{5,10,25,50\}$ for CIRCO and Recall@ $k = \{1,5,10,50\}$ for CIRR. Image and Text Interpolation are Crucial: As shown in Table 6a, employing Slerp for reference image embedding significantly enhances the model's learning capabilities. Applying interpolation on modification text further improves performance. Unimodal Queries are Beneficial: We investigate the necessity of unimodal queries in Eq. (3) and Eq. (4) and observe that omitting unimodal queries during training, especially text-only queries, substantially degrades performance (Table 6b). Nearest Neighbor is Essential: For reference image embedding interpolation, one might hypothesize that a random in-batch sample could suffice for interpolation. However, our study demonstrates that + +Table 7. Performance of different LLMs with CLIP-L/14 as the visual encoder. **Bold and underline highlight the best and second best score.** indicates the original LLM for text generation; others are Large Language Embedding Models (LLEMs) fine-tuned for text retrieval. + +
LLMCIRCO (mAP↑)CIRR (Rec.↑)FIQ (Rec.↑)
@5@10@50@1@10@50@10@50
Stella-Qwen2-1.5B [11]14.815.317.427.569.188.029.349.0
Mistral-7B* [24]19.820.222.729.672.591.329.549.3
E5-Mistral-7B-Inst [60]19.620.322.829.572.791.129.949.5
SFR-Embedding-2 [47]20.320.823.429.772.891.530.149.5
+ +using the nearest image embedding neighbor yields significantly better performance (Table 6c). LLEMs Outperforms LLMs: We investigate Large Language Embedding Models (LLEMs), which are fine-tuned for text retrieval compared to their base LLMs. As shown in Table 7, both E5-Mistral-7B-Inst [60] and SFR-Embedding-2 outperform their LLM counterpart, Mistral-7B [24]. The superior performance of LLEMs demonstrates that models specifically tailored for embedding and retrieval tasks can offer substantial advantages in CIR applications compared to general-purpose language models. + +# 6. Limitations and Conclusion + +While these advancements significantly improve CIR performance, several areas warrant further investigation. Our work, limited to LLM/LLEMs, could benefit from exploring pre-trained MLLMs for enhanced understanding of CIR tasks. Additionally, our triplet synthesis method generates a single visual token, constraining the model's ability to process detailed image information. Future work should leverage the underutilized text category information in our synthetic datasets to improve model generalization. Lastly, our refined benchmarks, while more reliable, still contain noise from original image pairs, suggesting a need for further refinement of evaluation metrics. + +In conclusion, we present novel approaches to Composed Image Retrieval (CIR) that obviate the need for annotated datasets. Our contributions include: (1) an innovative triplet synthesis method utilizing image-caption pairs, (2) a new architecture leveraging LLM's embedding generation capabilities, (3) MTCIR, a diverse, human-aligned synthetic dataset, and (4) refined versions of CIRR and Fashion-IQ benchmarks, enhancing the reliability of evaluation metrics in the field. Our method consistently outperforms existing LLM and non-LLM baselines across popular benchmarks. Notably, MTCIR achieves superior results with only $10 - 20\%$ of the size of larger datasets, demonstrating a $1 - 15\%$ increase in Recall. These advancements collectively push the boundaries of CIR, offering more efficient and effective solutions for real-world applications. + +# References + +[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 3, 7 +[2] Anthropic. The claude 3 model family: Opus, sonnet, haiku, 2024. 3, 5, 6 +[3] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. A clip-hitchhiker's guide to long video retrieval. arXiv preprint arXiv:2205.08508, 2022. 5 +[4] Alberto Baldrati, Lorenzo Agnolucci, Marco Bertini, and Alberto Del Bimbo. Zero-shot composed image retrieval with textual inversion. In ICCV, 2023. 2, 3, 5, 6, 7, 13, 15 +[5] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In CVPR, 2023. 2, 3, 5 +[6] Tom B Brown. Language models are few-shot learners. In NeurIPS, 2020. 5 +[7] Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, et al. Palm: Scaling language modeling with pathways. Journal of Machine Learning Research, 2023. 2, 5 +[8] Niv Cohen, Rinon Gal, Eli A. Meirom, Gal Chechik, and Yuval Atzmon. "this is my unicorn, fluffy": Personalizing frozen vision-language representations. In ECCV, 2022. 2, 3, 6, 15 +[9] Yashar Deldjoo, Fatemeh Nazary, Arnau Ramisa, Julian Mcauley, Giovanni Pellegrini, Alejandro Bellogin, and Tommaso Di Noia. A review of modern fashion recommender systems. ACM Computing Surveys, 2023. 1 +[10] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009. 13 +[11] Zhang Dun. Stella en qwen2 1.5b v5, 2024.8 +[12] Sonam Goenka, Zhaoheng Zheng, Ayush Jaiswal, Rakesh Chada, Yue Wu, Varsha Hedau, and Pradeep Natarajan. Fashionvlp: Vision language transformer for fashion retrieval with feedback. In CVPR, 2022. 1 +[13] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Batra, and Devi Parikh. Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In CVPR, 2017. 5 +[14] Geonmo Gu, Sanghyuk Chun, Wonjae Kim, HeeJae Jun, Yoohoon Kang, and Sangdoo Yun. Compodiff: Versatile composed image retrieval with latent diffusion. Transactions on Machine Learning Research, 2024. 2, 3, 5, 6, 16 +[15] Geonmo Gu, Sanghyuk Chun, Wonjae Kim, Yoohoon Kang, and Sangdoo Yun. Language-only training of zero-shot composed image retrieval. In CVPR, 2024. 2, 3, 6, 15 +[16] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 13 + +[17] Matthew Honnibal, Ines Montani, Sofie Van Landeghem, Adriane Boyd, et al. spacy: Industrial-strength natural language processing in python. 2020. 13 +[18] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In ICLR, 2022. 15 +[19] Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Weli-hinda, Alan Hayes, Alec Radford, et al. Gpt-4o system card. arXiv preprint arXiv:2410.21276, 2024. 13 +[20] Chuong Huynh, Yuqian Zhou, Zhe Lin, Connelly Barnes, Eli Shechtman, Sohrab Amirghodsi, and Abhinav Shrivastava. Simpson: Simplifying photo cleanup with single-click distracting object segmentation network. In CVPR, 2023. 1 +[21] Young Kyun Jang, Dat Huynh, Ashish Shah, Wen-Kai Chen, and Ser-Nam Lim. Spherical linear interpolation and text-anchoring for zero-shot composed image retrieval. In ECCV, 2024. 2, 3, 6, 7, 14, 15 +[22] Young Kyun Jang, Junmo Kang, Yong Jae Lee, and Donghyun Kim. Mate: Meet at the embedding-connecting images with long texts. EMNLP Findings, 2024. 3 +[23] Young Kyun Jang, Donghyun Kim, Zihang Meng, Dat Huynh, and Ser-Nam Lim. Visual delta generator with large multi-modal models for semi-supervised composed image retrieval. In CVPR, 2024. 2, 3, 5 +[24] Albert Q Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, et al. Mistral 7b. arXiv preprint arXiv:2310.06825, 2023. 8 +[25] Ting Jiang, Minghui Song, Zihan Zhang, Haizhen Huang, Weiwei Deng, Feng Sun, Qi Zhang, Deqing Wang, and Fuzhen Zhuang. E5-v: Universal embeddings with multimodal large language models. arXiv preprint arXiv:2407.12580, 2024.3 +[26] Shyamgopal Karthik, Karsten Roth, Massimiliano Mancini, and Zeynep Akata. Vision-by-language for training-free compositional image retrieval. *ICLR*, 2024. 3, 6, 7, 15 +[27] Chankyu Lee, Rajarshi Roy, Mengyao Xu, Jonathan Raiman, Mohammad Shoeybi, Bryan Catanzaro, and Wei Ping. Nv-embed: Improved techniques for training llms as generalist embedding models. arXiv preprint arXiv:2405.17428, 2024.3 +[28] Matan Levy, Rami Ben-Ari, Nir Darshan, and Dani Lischinski. Data roaming and quality assessment for composed image retrieval. AAAI, 2024. 2, 3, 5, 6, 7, 14, 15, 17, 18 +[29] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, 2022. 2, 7, 14, 17, 18 +[30] Wei Li, Hehe Fan, Yongkang Wong, Yi Yang, and Mohan Kankanhalli. Improving context understanding in multimodal large language models via multimodal composition learning. In ICML, 2024. 3, 6, 7 +[31] Zehan Li, Xin Zhang, Yanzhao Zhang, Dingkun Long, Pengjun Xie, and Meishan Zhang. Towards general text + +embeddings with multi-stage contrastive learning. arXiv preprint arXiv:2308.03281, 2023. 3 +[32] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 5, 7, 15 +[33] Aixin Liu, Bei Feng, Bing Xue, Bingxuan Wang, Bochao Wu, Chengda Lu, Chenggang Zhao, Chengqi Deng, Chenyu Zhang, Chong Ruan, et al. Deepseek-v3 technical report. arXiv preprint arXiv:2412.19437, 2024. 13 +[34] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, 2023. 2, 3, 5, 6, 8, 14, 15 +[35] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava-last: Improved reasoning,OCR, and world knowledge, 2024. 2, 3, 5, 13 +[36] Yikun Liu, Jiangchao Yao, Ya Zhang, Yan-Feng Wang, and Weidi Xie. Zero-shot composed text-image retrieval. In BMVC, 2023. 2 +[37] Zheyuan Liu, Cristian Rodriguez-Opazo, Damien Teney, and Stephen Gould. Image retrieval on real-life images with pretrained vision-and-language models. In ICCV, 2021. 2, 3, 5, 7, 13, 14, 17 +[38] Feipeng Ma, Hongwei Xue, Guangting Wang, Yizhou Zhou, Fengyun Rao, Shilin Yan, Yueyi Zhang, Siying Wu, Mike Zheng Shou, and Xiaoyan Sun. Multi-modal generative embedding model. arXiv preprint arXiv:2405.19333, 2024. 3 +[39] Gabriel de Souza P Moreira, Radek Osmulski, Mengyao Xu, Ronay Ak, Benedikt Schifferer, and Even Oldridge. Nv-retriever: Improving text embedding models with effective hard-negative mining. arXiv preprint arXiv:2407.15831, 2024. 3 +[40] Niklas Muennighoff, Nouamane Tazi, Loic Magne, and Nils Reimers. Mteb: Massive text embedding benchmark. arXiv preprint arXiv:2210.07316, 2022. 3 +[41] Vicente Ordonez, Girish Kulkarni, and Tamara Berg. Im2text: Describing images using 1 million captioned photographs. NeurIPS, 2011. 5 +[42] Khoi Pham, Chuong Huynh, Ser-Nam Lim, and Abhinav Shrivastava. Composing object relations and attributes for image-text matching. In CVPR, 2024. 1 +[43] Bill Psomas, Ioannis Kakogeorgiou, Nikos Efthymiadis, Giorgos Tolias, Ondrej Chum, Yannis Avrithis, and Konstantinos Karantzalos. Composed image retrieval for remote sensing. In IGARSS 2024-2024 IEEE International Geoscience and Remote Sensing Symposium, 2024. 3 +[44] Filip Radenovic, Abhimanyu Dubey, Abhishek Kadian, Todor Mihaylov, Simon Vandenhende, Yash Patel, Yi Wen, Vignesh Ramanathan, and Dhruv Mahajan. Filtering, distillation, and hard negatives for vision-language pre-training. In CVPR, 2023. 4 +[45] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 2, 5, 13, 14 + +[46] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 2, 3, 5 +[47] Meng Rui, Liu Ye, Joty Shafiq Rayhan, Xiong Caiming, Zhou Yingbo, and Yavuz Semih. Sfr-embedding-2: Advanced text embedding with multi-stage training, 2024. 8, 14 +[48] Amrita Saha, Mitesh Khapra, and Karthik Sankaranarayanan. Towards building large scale multimodal domain-aware conversation systems. In AAAI, 2018. 3 +[49] Kuniaki Saito, Kihyuk Sohn, Xiang Zhang, Chun-Liang Li, Chen-Yu Lee, Kate Saenko, and Tomas Pfister. Pic2word: Mapping pictures to words for zero-shot composed image retrieval. CVPR, 2023. 3, 6, 15 +[50] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021. 5, 6, 14 +[51] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In ACL, 2018. 5, 6, 14 +[52] Ken Shoemake. Animating rotation with quaternion curves. In Proceedings of the 12th annual conference on Computer graphics and interactive techniques, 1985. 2, 4 +[53] Swetha Sirnam, Jinyu Yang, Tal Neiman, Mamshad Nayeem Rizve, Son Tran, Benjamin Yao, Trishul Chilimbi, and Mubarak Shah. X-former: Unifying contrastive and reconstruction learning for mllms. In European Conference on Computer Vision, pages 146-162. Springer, 2024. 2 +[54] Alane Suhr, Stephanie Zhou, Ally Zhang, Iris Zhang, Huajun Bai, and Yoav Artzi. A corpus for reasoning about natural language grounded in photographs. In ACL, 2019. 5 +[55] Yuanmin Tang, Jing Yu, Keke Gai, Jiamin Zhuang, Gang Xiong, Yue Hu, and Qi Wu. Context-i2w: Mapping images to context-dependent words for accurate zero-shot composed image retrieval. In AAAI, 2024. 2, 3, 6, 15 +[56] Omkar Thawakar, Muzammal Naseer, Rao Muhammad Anwer, Salman Khan, Michael Felsberg, Mubarak Shah, and Fahad Shahbaz Khan. Composed video retrieval via enriched context and discriminative embeddings. CVPR, 2024. 4, 6, 16 +[57] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 2 +[58] Lucas Ventura, Antoine Yang, Cordelia Schmid, and Gül Varol. CoVR-2: Automatic data construction for composed video retrieval. IEEE TPAMI, 2024. 2, 3, 5, 6, 7, 14, 16, 18 +[59] Lucas Ventura, Antoine Yang, Cordelia Schmid, and Güll Varol. CoVR: Learning composed video retrieval from web video captions. AAAI, 2024. 3, 4, 5, 6, 7, 15, 16, 17 +[60] Liang Wang, Nan Yang, Xiaolong Huang, Linjun Yang, Rangan Majumder, and Furu Wei. Improving text embeddings with large language models. arXiv preprint arXiv:2401.00368, 2023. 3, 8 + +[61] Hui Wu, Yupeng Gao, Xiaoxiao Guo, Ziad Al-Halah, Steven Rennie, Kristen Grauman, and Rogerio Feris. The fashion iq dataset: Retrieving images by combining side information and relative natural language feedback. CVPR, 2021. 2, 5, 7, 17 +[62] Jinyu Yang, Jiali Duan, Son Tran, Yi Xu, Sampath Chanda, Liquun Chen, Belinda Zeng, Trishul Chilimbi, and Junzhou Huang. Vision-language pre-training with triple contrastive learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15671-15680, 2022. 1 +[63] ZHOU Yuqian, Chuong Huynh, Connelly Barnes, Elya Shechtman, Sohrab Amirghodsi, and Zhe Lin. Machine-learning models for distractor segmentation with reduced user interactions, 2024. US Patent App. 18/307,353. 1 +[64] Kai Zhang, Yi Luan, Hexiang Hu, Kenton Lee, Siyuan Qiao, Wenhu Chen, Yu Su, and Ming-Wei Chang. Magiclens: Self-supervised image retrieval with open-ended instructions. In ICML, 2024. 2, 3, 5, 6, 7, 16, 18 +[65] Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen, Christopher Dewan, Mona Diab, Xian Li, Xi Victoria Lin, et al. Opt: Open pre-trained transformer language models. arXiv preprint arXiv:2205.01068, 2022. 2, 5 +[66] Youcai Zhang, Xinyu Huang, Jinyu Ma, Zhaoyang Li, Zhaochuan Luo, Yanchun Xie, Yuzhuo Qin, Tong Luo, Yaqian Li, Shilong Liu, et al. Recognize anything: A strong image tagging model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1724-1732, 2024. 13 +[67] Xinliang Zhu, Sheng-Wei Huang, Han Ding, Jinyu Yang, Kelvin Chen, Tao Zhou, Tal Neiman, Ouye Xie, Son Tran, Benjamin Yao, et al. Bringing multimodality to amazon visual search system. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 6390–6399, 2024. 1 +[68] Yutao Zhu, Huaying Yuan, Shuting Wang, Jiongnan Liu, Wenhan Liu, Chenlong Deng, Haonan Chen, Zhicheng Dou, and Ji-Rong Wen. Large language models for information retrieval: A survey. arXiv preprint arXiv:2308.07107, 2023. 3 \ No newline at end of file diff --git a/2025/CoLLM_ A Large Language Model for Composed Image Retrieval/images.zip b/2025/CoLLM_ A Large Language Model for Composed Image Retrieval/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..c3dfdfef57f2f5c35b93a2abb69c0e8a49e75afb --- /dev/null +++ b/2025/CoLLM_ A Large Language Model for Composed Image Retrieval/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c97031157076021d1fad10fc96ce204760ed6385257b0703252efe61a7b99406 +size 577801 diff --git a/2025/CoLLM_ A Large Language Model for Composed Image Retrieval/layout.json b/2025/CoLLM_ A Large Language Model for Composed Image Retrieval/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..8a073015c988ed15c5c7cae53d3bbd3efbfe761e --- /dev/null +++ b/2025/CoLLM_ A Large Language Model for Composed Image Retrieval/layout.json @@ -0,0 +1,10213 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 100, + 103, + 510, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 103, + 510, + 121 + ], + "spans": [ + { + "bbox": [ + 100, + 103, + 510, + 121 + ], + "type": "text", + "content": "CoLLM: A Large Language Model for Composed Image Retrieval" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 99, + 142, + 510, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 142, + 510, + 172 + ], + "spans": [ + { + "bbox": [ + 99, + 142, + 510, + 172 + ], + "type": "text", + "content": "Chuong Huynh" + }, + { + "bbox": [ + 99, + 142, + 510, + 172 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 99, + 142, + 510, + 172 + ], + "type": "text", + "content": " Jinyu Yang" + }, + { + "bbox": [ + 99, + 142, + 510, + 172 + ], + "type": "inline_equation", + "content": "^{2\\dagger}" + }, + { + "bbox": [ + 99, + 142, + 510, + 172 + ], + "type": "text", + "content": " Ashish Tawari" + }, + { + "bbox": [ + 99, + 142, + 510, + 172 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 99, + 142, + 510, + 172 + ], + "type": "text", + "content": " Mubarak Shah" + }, + { + "bbox": [ + 99, + 142, + 510, + 172 + ], + "type": "inline_equation", + "content": "^{2,3}" + }, + { + "bbox": [ + 99, + 142, + 510, + 172 + ], + "type": "text", + "content": " Son Tran" + }, + { + "bbox": [ + 99, + 142, + 510, + 172 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 99, + 142, + 510, + 172 + ], + "type": "text", + "content": " Raffay Hamid" + }, + { + "bbox": [ + 99, + 142, + 510, + 172 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 99, + 142, + 510, + 172 + ], + "type": "text", + "content": " Trishul Chilimbi" + }, + { + "bbox": [ + 99, + 142, + 510, + 172 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 99, + 142, + 510, + 172 + ], + "type": "text", + "content": " Abhinav Shrivastava" + }, + { + "bbox": [ + 99, + 142, + 510, + 172 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 178, + 184, + 432, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 178, + 184, + 432, + 198 + ], + "spans": [ + { + "bbox": [ + 178, + 184, + 432, + 198 + ], + "type": "text", + "content": "1 University of Maryland, College Park 2 Amazon" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 198, + 479, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 198, + 479, + 213 + ], + "spans": [ + { + "bbox": [ + 130, + 198, + 479, + 213 + ], + "type": "text", + "content": "3 Center for Research in Computer Vision, University of Central Florida" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 99, + 215, + 507, + 241 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 215, + 507, + 241 + ], + "spans": [ + { + "bbox": [ + 99, + 215, + 507, + 241 + ], + "type": "inline_equation", + "content": "^{1}\\{\\mathrm{chuonghm}, \\mathrm{abhinav}\\} @ \\mathrm{cs.umd.edu}" + }, + { + "bbox": [ + 99, + 215, + 507, + 241 + ], + "type": "inline_equation", + "content": "^{2}\\{\\mathrm{viyjy}, \\mathrm{atawari}, \\mathrm{sontran}, \\mathrm{raffay}, \\mathrm{trishulc}\\} @ \\mathrm{amazon.com}" + }, + { + "bbox": [ + 99, + 215, + 507, + 241 + ], + "type": "text", + "content": " shah@crcv.ucf.edu" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 151, + 269, + 200, + 280 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 269, + 200, + 280 + ], + "spans": [ + { + "bbox": [ + 151, + 269, + 200, + 280 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 295, + 296, + 691 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 295, + 296, + 691 + ], + "spans": [ + { + "bbox": [ + 54, + 295, + 296, + 691 + ], + "type": "text", + "content": "Composed Image Retrieval (CIR) is a complex task that aims to retrieve images based on a multimodal query. Typical training data consists of triplets containing a reference image, a textual description of desired modifications, and the target image, which are expensive and time-consuming to acquire. The scarcity of CIR datasets has led to zero-shot approaches utilizing synthetic triplets or leveraging vision-language models (VLMs) with ubiquitous web-crawled image-caption pairs. However, these methods have significant limitations: synthetic triplets suffer from limited scale, lack of diversity, and unnatural modification text, while image-caption pairs hinder joint embedding learning of the multimodal query due to the absence of triplet data. Moreover, existing approaches struggle with complex and nuanced modification texts that demand sophisticated fusion and understanding of vision and language modalities. We present CoLLM, a one-stop framework that effectively addresses these limitations. Our approach generates triplets on-the-fly from image-caption pairs, enabling supervised training without manual annotation. We leverage Large Language Models (LLMs) to generate joint embeddings of reference images and modification texts, facilitating deeper multimodal fusion. Additionally, we introduce Multi-Text CIR (MTCIR), a large-scale dataset comprising 3.4M samples, and refine existing CIR benchmarks (CIRR and Fashion-IQ) to enhance evaluation reliability. Experimental results demonstrate that CoLLM achieves state-of-the-art performance across multiple CIR benchmarks and settings. MTCIR yields competitive results, with up to " + }, + { + "bbox": [ + 54, + 295, + 296, + 691 + ], + "type": "inline_equation", + "content": "15\\%" + }, + { + "bbox": [ + 54, + 295, + 296, + 691 + ], + "type": "text", + "content": " performance improvement. Our refined benchmarks provide more reliable evaluation metrics for CIR models, contributing to the advancement of this important field. Project page is at collm-cvpr25.github.io." + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 316, + 262, + 555, + 326 + ], + "blocks": [ + { + "bbox": [ + 316, + 262, + 555, + 326 + ], + "lines": [ + { + "bbox": [ + 316, + 262, + 555, + 326 + ], + "spans": [ + { + "bbox": [ + 316, + 262, + 555, + 326 + ], + "type": "image", + "image_path": "d857becc33853ef401ca098ec992c3b7de0716d5988b6e9d34475f8aadfd2412.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 369, + 327, + 500, + 338 + ], + "lines": [ + { + "bbox": [ + 369, + 327, + 500, + 338 + ], + "spans": [ + { + "bbox": [ + 369, + 327, + 500, + 338 + ], + "type": "text", + "content": "(a) Composed Image Retrieval Example" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 316, + 342, + 555, + 459 + ], + "blocks": [ + { + "bbox": [ + 316, + 342, + 555, + 459 + ], + "lines": [ + { + "bbox": [ + 316, + 342, + 555, + 459 + ], + "spans": [ + { + "bbox": [ + 316, + 342, + 555, + 459 + ], + "type": "image", + "image_path": "259b288b4e526d553b3518123107e72c38497de73df2a3acafdc4e85671a6249.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 332, + 460, + 536, + 470 + ], + "lines": [ + { + "bbox": [ + 332, + 460, + 536, + 470 + ], + "spans": [ + { + "bbox": [ + 332, + 460, + 536, + 470 + ], + "type": "text", + "content": "(b) Recall Sum of state-of-the-art methods on CIR benchmarks" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 479, + 555, + 536 + ], + "lines": [ + { + "bbox": [ + 313, + 479, + 555, + 536 + ], + "spans": [ + { + "bbox": [ + 313, + 479, + 555, + 536 + ], + "type": "text", + "content": "Figure 1. (a) An example of CIR. (b) Recall Sum at " + }, + { + "bbox": [ + 313, + 479, + 555, + 536 + ], + "type": "inline_equation", + "content": "\\{1,10,50\\}" + }, + { + "bbox": [ + 313, + 479, + 555, + 536 + ], + "type": "text", + "content": " for CIRR and " + }, + { + "bbox": [ + 313, + 479, + 555, + 536 + ], + "type": "inline_equation", + "content": "\\{10,50\\}" + }, + { + "bbox": [ + 313, + 479, + 555, + 536 + ], + "type": "text", + "content": " for Fashion-IQ between CoLLM and state-of-the-art (SoTA) models under zero-shot settings. We evaluate two training scenarios: (i) without triplet data and (ii) with synthetic triplet data." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 548, + 394, + 559 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 548, + 394, + 559 + ], + "spans": [ + { + "bbox": [ + 314, + 548, + 394, + 559 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 312, + 568, + 555, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 568, + 555, + 700 + ], + "spans": [ + { + "bbox": [ + 312, + 568, + 555, + 700 + ], + "type": "text", + "content": "Composed Image Retrieval (CIR) enhances traditional image retrieval [42, 62] by combining text and image queries, offering greater flexibility in search systems, with applications in e-commerce, fashion, and design [9, 12, 20, 63, 67]. As illustrated in Fig. 1a, CIR retrieves similar items, such as shirts, where the reference image provides a visual basis, and the modification text specifies desired modifications. By expanding the capabilities of conventional image or text-based searches, CIR allows for more nuanced and specific queries. However, this advanced approach also presents significant challenges compared to traditional image retrieval." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 325, + 701, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 701, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 325, + 701, + 553, + 714 + ], + "type": "text", + "content": "Supervised CIR approaches face challenges in data ac" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 702, + 281, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 702, + 281, + 713 + ], + "spans": [ + { + "bbox": [ + 66, + 702, + 281, + 713 + ], + "type": "text", + "content": "*This work was done during internship at Amazon. † Project lead." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3994" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 323 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 323 + ], + "type": "text", + "content": "quisition. They require high-quality CIR triplets (reference image, target image, and modification text), collected through a labor-intensive and expensive process. Consequently, existing CIR triplet datasets are limited in scale and domain coverage, restricting model generalizability. To overcome these limitations, recent CIR methods [14, 23, 28, 58, 64] have adopted zero-shot approaches [36]. These can be broadly categorized into two strategies: (i) leveraging vision-language models [29, 45, 53] (VLMs) for composed query embedding generation, and (ii) generating synthetic triplets for supervised training. VLM-based approaches utilize pre-trained models that already align vision and language features in latent space. These methods rely on large-scale image-caption pairs and follow two main directions, i.e., textual inversion [8, 15, 55] and direct interpolation [21]. Synthetic triplet generation approaches [23, 28, 64] employ Large Language Models (LLMs) [7, 57, 65] to generate modification text, with CompoDiff [14] further utilizing diffusion models [5, 46] to create synthetic reference-target image pairs. Despite the recent advances in CIR, several critical challenges persist:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 327, + 295, + 567 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 55, + 327, + 295, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 327, + 295, + 363 + ], + "spans": [ + { + "bbox": [ + 55, + 327, + 295, + 363 + ], + "type": "text", + "content": "- [Data] Limitation-1: VLM-based methods may hinder efficient and effective composed query embedding learning by relying solely on image-caption pairs." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 363, + 295, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 363, + 295, + 411 + ], + "spans": [ + { + "bbox": [ + 56, + 363, + 295, + 411 + ], + "type": "text", + "content": "- [Data] Limitation-2: Existing synthetic triplet datasets often suffer from a lack of diversity and unnatural modification text. Moreover, these datasets are either very small or closed-source, impeding research progress in the field." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 411, + 295, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 411, + 295, + 470 + ], + "spans": [ + { + "bbox": [ + 56, + 411, + 295, + 470 + ], + "type": "text", + "content": "- [Model] Limitation-3: Current methods for composed query embeddings mainly use shallow transformer models or linear interpolation. While these methods are computationally efficient, they lack the ability to capture the full complexity of the composed understanding tasks." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 471, + 295, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 471, + 295, + 567 + ], + "spans": [ + { + "bbox": [ + 56, + 471, + 295, + 567 + ], + "type": "text", + "content": "- [Evaluation] Limitation-4: Existing CIR benchmarks are often compromised by noise, particularly in the form of ambiguous samples. Ambiguity arises when multiple target images can correctly match a single query, but only one is labeled as the ground truth, ignoring valid matches. Although CIRCO [4] attempts to address this issue, its efforts are limited in scale. This ambiguity in benchmarks hinders meaningful model evaluation and comparison." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 570, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 295, + 713 + ], + "type": "text", + "content": "We propose CoLLM, an LLM-based CIR approach to address the aforementioned limitations. CoLLM tackles Limitation-1 by dynamically synthesizing triplets from image-caption pairs, introducing two key components: a reference image embedding synthesis and a modification text synthesis module. The former employs Spherical Linear Interpolation [52] (Slerp) to generate an intermediate embedding between a given image and its nearest neighbor within the training batch, serving as a synthesized reference image embedding. The latter utilizes a pre-defined text interpolation template to generate modification text based on the current caption and that of the nearest image neighbor." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 553, + 132 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 132 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 132 + ], + "type": "text", + "content": "This strategy effectively leverages the vast amount of readily available image-caption pairs on the Internet, enabling a model's training in a supervised CIR manner. By doing so, CoLLM overcomes the scarcity of labeled triplet data and paves the way for more robust and scalable CIR models." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 133, + 553, + 286 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 133, + 553, + 286 + ], + "spans": [ + { + "bbox": [ + 313, + 133, + 553, + 286 + ], + "type": "text", + "content": "To address Limitation-2, we introduce Multi-Text CIR (MTCIR), a synthetic dataset of 3.4M image pairs with 17.7M modification texts. MTCIR focuses on two often-overlooked aspects: image diversity and naturalistic modification texts. We curated images from diverse sources to ensure variety. For modification text generation, we employ a two-stage approach using Multi-modal LLM (MLLM) [34, 35] for detailed captioning and LLM for describing interception differences. Uniquely, MTCIR provides multiple short modification texts for each image pair, covering various attributes. This better reflects human query formulation, offering a more realistic, comprehensive training foundation for CIR models." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 288, + 553, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 288, + 553, + 407 + ], + "spans": [ + { + "bbox": [ + 313, + 288, + 553, + 407 + ], + "type": "text", + "content": "To overcome Limitation-3, CoLLM harnesses the power of LLMs for composed query understanding. It is motivated by the extensive world knowledge embedded in pre-trained LLMs. With their deep semantic understanding, we posit that LLMs are superior to shallow transformers and simple embedding interpolation techniques in comprehending the intricate relationships between reference images and modification texts. By leveraging LLMs, CoLLM aims to capture nuanced semantic connections, enhancing the quality and relevance of composed image retrieval results." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 408, + 553, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 408, + 553, + 479 + ], + "spans": [ + { + "bbox": [ + 313, + 408, + 553, + 479 + ], + "type": "text", + "content": "To address Limitation-4, we refine two popular CIR benchmarks: CIRR [37] and Fashion-IQ [61]. We use MLLMs to evaluate sample ambiguity in each benchmark and regenerate clear modification text for ambiguous ones. Our pipeline incorporates multiple validation steps to guarantee the enhanced quality of the refined samples." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 480, + 553, + 659 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 480, + 553, + 659 + ], + "spans": [ + { + "bbox": [ + 313, + 480, + 553, + 659 + ], + "type": "text", + "content": "Our main contributions can be summarized as: (i) We propose a method to synthesize CIR triplets on-the-fly from image-caption pairs, which can even outperform models trained on CIR triplets, eliminating the need for costly CIR datasets. (ii) We collect and will release MTCIR, a new CIR triplet dataset covering 3.4 million image pairs with 17.7 million modification texts. To our knowledge, MTCIR is the largest open-source synthetic CIR dataset. (iii) We introduce an approach to leverage LLMs for composed query understanding, utilizing their instruction-following and embedding generation capabilities. (iv) We refine CIRR and Fashion-IQ to provide more robust evaluation benchmarks for the CIR community. Extensive experiments on popular CIR benchmarks and settings demonstrate the effectiveness of our model innovations and new datasets (Fig. 1b)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 666, + 406, + 679 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 666, + 406, + 679 + ], + "spans": [ + { + "bbox": [ + 314, + 666, + 406, + 679 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "content": "Composed Image Retrieval. Composed Image Retrieval (CIR) has garnered significant attention due to its flexibil" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3995" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 215 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 215 + ], + "type": "text", + "content": "ity in search systems [37, 43, 48]. Zero-shot CIR methods have been extensively explored, with textual inversion [4, 8, 15, 49, 55] emerging as a prominent technique. This approach maps image encoder outputs to text encoder inputs, creating pseudo-word tokens. SlerpTAT [21] emphasizes the text encoder's importance by re-aligning visual and textual embeddings. While most zero-shot CIR models utilize image-caption pairs for training, our work introduces an innovative method to synthesize triplets from these pairs during training. This method enables supervised CIR style training, enhancing the model's understanding of composed queries without relying on real CIR datasets." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 217, + 295, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 217, + 295, + 373 + ], + "spans": [ + { + "bbox": [ + 55, + 217, + 295, + 373 + ], + "type": "text", + "content": "Recent works enhance enhancing models' language understanding using Large Language Models (LLMs) [26, 30], primarily leveraging their text generation abilities. MCL [30] composes image-text query in the LLM input space and aligns joint embedding with the target image caption. However, this may introduce noise due to limited visual information in the target image captions. CIReVL [26] employs LLMs to generate captions based on the reference image and modification text, subsequently retrieving the target image using a text-to-image retrieval approach. Our model also utilizes LLMs but differs by directly producing a composed query embedding for target image retrieval, potentially reducing intermediate steps and associated errors." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 375, + 295, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 375, + 295, + 555 + ], + "spans": [ + { + "bbox": [ + 55, + 375, + 295, + 555 + ], + "type": "text", + "content": "CIR with Synthetic Datasets. The scarcity of supervised CIR datasets has prompted recent studies to leverage generative models for synthetic triplet creation. CompoDiff [14] employs image editing pipelines [5, 46] to generate target images, though their nature limits performance. Other approaches [23, 28, 58, 59, 64] utilize LLMs to generate modification text for real image pairs. Our work introduces a novel two-stage approach, combining MLLMs [34, 35] for detailed captioning and LLMs [1, 2] for describing interception differences. Distinctly, our synthetic triplets provide multiple concise modification texts for each image pair, encompassing various attributes, in contrast to the conventional single, lengthy modification text. This enables a more nuanced representation of image modifications, potentially improving CIR model performance and versatility." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 558, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 295, + 715 + ], + "type": "text", + "content": "Large Language Models. Recent advancements in LLMs have expanded their applications beyond text generation to include image understanding [34, 35] and text embedding generation [40, 60, 68]. Large Language Embedding Models (LLEMs) are specialized versions trained using contrastive learning, leveraging LLM knowledge for embedding generation while often disabling text generation capabilities. To enhance text embedding quality, [27, 31, 39] propose removing causal attention in LLMs, thereby improving text information encoding efficiency and training these modified LLMs on text retrieval datasets. This text retrieval capability has been further extended to text-image retrieval by aligning visual and LLM text embed" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 330, + 71, + 536, + 168 + ], + "blocks": [ + { + "bbox": [ + 330, + 71, + 536, + 168 + ], + "lines": [ + { + "bbox": [ + 330, + 71, + 536, + 168 + ], + "spans": [ + { + "bbox": [ + 330, + 71, + 536, + 168 + ], + "type": "image", + "image_path": "9473b742177963d4ac05ea3ed412ecb46f2b36bba425554653d14f6327a0d981.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 329, + 172, + 539, + 242 + ], + "blocks": [ + { + "bbox": [ + 329, + 172, + 539, + 242 + ], + "lines": [ + { + "bbox": [ + 329, + 172, + 539, + 242 + ], + "spans": [ + { + "bbox": [ + 329, + 172, + 539, + 242 + ], + "type": "image", + "image_path": "77fbe85028cc23d395fcb33d1ee4270f55402926a2d57f1a408f56357b5a7967.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 247, + 553, + 270 + ], + "lines": [ + { + "bbox": [ + 313, + 247, + 553, + 270 + ], + "spans": [ + { + "bbox": [ + 313, + 247, + 553, + 270 + ], + "type": "text", + "content": "Figure 2. An overview of our model and training strategies when using (a) image-caption pairs and (b) CIR triplets." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 276, + 555, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 276, + 555, + 373 + ], + "spans": [ + { + "bbox": [ + 313, + 276, + 555, + 373 + ], + "type": "text", + "content": "ding spaces [22, 25, 38]. These developments demonstrate the potential of LLMs for composed query understanding in CIR tasks. Our work builds upon this foundation, extending LLMs/LLEMs to generate composed query embeddings by incorporating reference images and modification text. This novel approach leverages the multimodal capabilities of LLMs to enhance the performance of CIR systems, leading to more accurate and context-aware image retrieval." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 384, + 370, + 396 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 384, + 370, + 396 + ], + "spans": [ + { + "bbox": [ + 314, + 384, + 370, + 396 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 404, + 554, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 404, + 554, + 477 + ], + "spans": [ + { + "bbox": [ + 313, + 404, + 554, + 477 + ], + "type": "text", + "content": "This section outlines our methodology, starting with a brief overview of the model architecture, followed by a formal CIR problem definition. We then detail our triplet synthesis strategy, including reference image embedding and modification text synthesis, using image-caption pairs. Lastly, we describe our LLM-based query composition approach." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 479, + 429, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 479, + 429, + 491 + ], + "spans": [ + { + "bbox": [ + 313, + 479, + 429, + 491 + ], + "type": "text", + "content": "3.1. Model Architecture" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 497, + 555, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 497, + 555, + 641 + ], + "spans": [ + { + "bbox": [ + 313, + 497, + 555, + 641 + ], + "type": "text", + "content": "As illustrated in Fig. 2, CoLLM consists of several essential components: (1) a vision encoder " + }, + { + "bbox": [ + 313, + 497, + 555, + 641 + ], + "type": "inline_equation", + "content": "f(\\cdot)" + }, + { + "bbox": [ + 313, + 497, + 555, + 641 + ], + "type": "text", + "content": " for extracting image features; (2) modules for synthesizing reference image embeddings and modification text; (3) an image adapter " + }, + { + "bbox": [ + 313, + 497, + 555, + 641 + ], + "type": "inline_equation", + "content": "g(\\cdot)" + }, + { + "bbox": [ + 313, + 497, + 555, + 641 + ], + "type": "text", + "content": " that maps visual features into the language model's semantic space; (4) a LLM " + }, + { + "bbox": [ + 313, + 497, + 555, + 641 + ], + "type": "inline_equation", + "content": "\\Phi(\\cdot)" + }, + { + "bbox": [ + 313, + 497, + 555, + 641 + ], + "type": "text", + "content": " that processes multimodal queries and (5) a projection layer " + }, + { + "bbox": [ + 313, + 497, + 555, + 641 + ], + "type": "inline_equation", + "content": "proj(\\cdot)" + }, + { + "bbox": [ + 313, + 497, + 555, + 641 + ], + "type": "text", + "content": " (omitted from the figure for simplicity) that maps the LLM output to a suitable representation for retrieval. It is important to note that Fig. 2 (a) and Fig. 2 (b) illustrate architectures designed for two distinct input formats. The former is tailored for image-caption pairs, while the latter is optimized for CIR triplets." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 648, + 541, + 662 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 648, + 541, + 662 + ], + "spans": [ + { + "bbox": [ + 313, + 648, + 541, + 662 + ], + "type": "text", + "content": "3.2. CoLLM with Image-Caption Pairs as Input" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 665, + 554, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 554, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 554, + 715 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 313, + 665, + 554, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{X} = \\{(v_i, w_i)\\}_{i=1}^N" + }, + { + "bbox": [ + 313, + 665, + 554, + 715 + ], + "type": "text", + "content": " denote a set of image-caption pairs, where " + }, + { + "bbox": [ + 313, + 665, + 554, + 715 + ], + "type": "inline_equation", + "content": "v_i" + }, + { + "bbox": [ + 313, + 665, + 554, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 665, + 554, + 715 + ], + "type": "inline_equation", + "content": "w_i" + }, + { + "bbox": [ + 313, + 665, + 554, + 715 + ], + "type": "text", + "content": " represent the image and caption of the " + }, + { + "bbox": [ + 313, + 665, + 554, + 715 + ], + "type": "inline_equation", + "content": "i^{th}" + }, + { + "bbox": [ + 313, + 665, + 554, + 715 + ], + "type": "text", + "content": " sample, respectively. To effectively leverage the vast amount of image-caption pairs, we introduce a novel ap" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3996" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 70, + 294, + 196 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 294, + 196 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 294, + 196 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 294, + 196 + ], + "type": "image", + "image_path": "9e6e609e23dbeec35a20b5cef8f698bc61dccdd2111547532ea31992ccd53e63.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 207, + 295, + 240 + ], + "lines": [ + { + "bbox": [ + 55, + 207, + 295, + 240 + ], + "spans": [ + { + "bbox": [ + 55, + 207, + 295, + 240 + ], + "type": "text", + "content": "Figure 3. An overview of reference image embedding synthesis and modification text synthesis. The red-framed image represents the nearest neighbor of the augmented image in the training batch." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 245, + 296, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 245, + 296, + 413 + ], + "spans": [ + { + "bbox": [ + 54, + 245, + 296, + 413 + ], + "type": "text", + "content": "proach that synthesizes a CIR triplet on-the-fly for each image-caption pair " + }, + { + "bbox": [ + 54, + 245, + 296, + 413 + ], + "type": "inline_equation", + "content": "(v_{i},w_{i})" + }, + { + "bbox": [ + 54, + 245, + 296, + 413 + ], + "type": "text", + "content": ". In this process, " + }, + { + "bbox": [ + 54, + 245, + 296, + 413 + ], + "type": "inline_equation", + "content": "v_{i}" + }, + { + "bbox": [ + 54, + 245, + 296, + 413 + ], + "type": "text", + "content": " serves as the target image, while two distinct modules generate the remaining components: (i) a reference image embedding synthesizer and (ii) a modification text synthesizer (Fig. 2a). Notably, we do not generate an actual image for reference image synthesis. Instead, we synthesize the reference image embedding, which is computationally more efficient and allows seamless integration into the CIR pipeline. This formulation enables CoLLM to exploit the rich information in image-caption pairs, transforming them into CIR triplets that can be used for training. The following subsections detail the specific mechanisms for synthesizing the reference image embedding and the modification text." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 415, + 296, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 415, + 296, + 501 + ], + "spans": [ + { + "bbox": [ + 55, + 415, + 296, + 501 + ], + "type": "text", + "content": "Reference Image Embedding Synthesis. Our reference image embedding synthesis process is illustrated in Fig. 3 (left). We begin by applying an augmentation " + }, + { + "bbox": [ + 55, + 415, + 296, + 501 + ], + "type": "inline_equation", + "content": "t \\sim \\mathcal{T}" + }, + { + "bbox": [ + 55, + 415, + 296, + 501 + ], + "type": "text", + "content": " to the input image " + }, + { + "bbox": [ + 55, + 415, + 296, + 501 + ], + "type": "inline_equation", + "content": "v_{i}" + }, + { + "bbox": [ + 55, + 415, + 296, + 501 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 415, + 296, + 501 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 55, + 415, + 296, + 501 + ], + "type": "text", + "content": " represents a family of augmentations. The resulting augmented image is denoted as " + }, + { + "bbox": [ + 55, + 415, + 296, + 501 + ], + "type": "inline_equation", + "content": "v_{i}^{\\prime}" + }, + { + "bbox": [ + 55, + 415, + 296, + 501 + ], + "type": "text", + "content": " with its corresponding embedding " + }, + { + "bbox": [ + 55, + 415, + 296, + 501 + ], + "type": "inline_equation", + "content": "\\mathbf{h}_i^\\prime = f(v_i^\\prime)" + }, + { + "bbox": [ + 55, + 415, + 296, + 501 + ], + "type": "text", + "content": ". For " + }, + { + "bbox": [ + 55, + 415, + 296, + 501 + ], + "type": "inline_equation", + "content": "\\mathbf{h}_i^\\prime" + }, + { + "bbox": [ + 55, + 415, + 296, + 501 + ], + "type": "text", + "content": ", we identify its in-batch nearest neighbor " + }, + { + "bbox": [ + 55, + 415, + 296, + 501 + ], + "type": "inline_equation", + "content": "v_{j}^{\\prime}" + }, + { + "bbox": [ + 55, + 415, + 296, + 501 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 55, + 415, + 296, + 501 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 55, + 415, + 296, + 501 + ], + "type": "text", + "content": " defined as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 111, + 507, + 295, + 530 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 507, + 295, + 530 + ], + "spans": [ + { + "bbox": [ + 111, + 507, + 295, + 530 + ], + "type": "interline_equation", + "content": "j = \\underset {j \\neq i; j \\in \\{1, \\dots \\mathcal {B} \\}} {\\arg \\max } \\operatorname {s i m} \\left(\\mathbf {h} _ {i} ^ {\\prime}, \\mathbf {h} _ {j} ^ {\\prime}\\right), \\tag {1}", + "image_path": "a83f6bd155637ecda41866af1dfdb6228aebf7acb02115fd3dfe615a37716817.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 536, + 295, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 536, + 295, + 563 + ], + "spans": [ + { + "bbox": [ + 55, + 536, + 295, + 563 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 536, + 295, + 563 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 55, + 536, + 295, + 563 + ], + "type": "text", + "content": " is the batch size, " + }, + { + "bbox": [ + 55, + 536, + 295, + 563 + ], + "type": "inline_equation", + "content": "\\mathbf{h}_j^{\\prime} = f(v_j^{\\prime})" + }, + { + "bbox": [ + 55, + 536, + 295, + 563 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 536, + 295, + 563 + ], + "type": "inline_equation", + "content": "\\mathrm{sim}(\\mathbf{u},\\mathbf{v}) = \\mathbf{u}^T\\mathbf{v} / (\\| \\mathbf{u}\\| \\| \\mathbf{v}\\|)" + }, + { + "bbox": [ + 55, + 536, + 295, + 563 + ], + "type": "text", + "content": " denotes cosine similarity." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 563, + 296, + 611 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 563, + 296, + 611 + ], + "spans": [ + { + "bbox": [ + 55, + 563, + 296, + 611 + ], + "type": "text", + "content": "We then employ Spherical Linear Interpolation (Slerp) [52] to synthesize the reference image embedding by interpolating between " + }, + { + "bbox": [ + 55, + 563, + 296, + 611 + ], + "type": "inline_equation", + "content": "\\mathbf{h}_i^{\\prime}" + }, + { + "bbox": [ + 55, + 563, + 296, + 611 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 563, + 296, + 611 + ], + "type": "inline_equation", + "content": "\\mathbf{h}_j^{\\prime}" + }, + { + "bbox": [ + 55, + 563, + 296, + 611 + ], + "type": "text", + "content": ". The synthesized reference image embedding is obtained as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 102, + 618, + 186, + 632 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 618, + 186, + 632 + ], + "spans": [ + { + "bbox": [ + 102, + 618, + 186, + 632 + ], + "type": "interline_equation", + "content": "\\theta = \\operatorname {a r c c o s} \\left(\\mathbf {h} _ {i} ^ {\\prime} \\cdot \\mathbf {h} _ {j} ^ {\\prime}\\right)", + "image_path": "3622e974883183c652babebb603b7983374808afda807d78bd815f73c8f055c7.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 97, + 633, + 295, + 659 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 633, + 295, + 659 + ], + "spans": [ + { + "bbox": [ + 97, + 633, + 295, + 659 + ], + "type": "interline_equation", + "content": "\\mathbf {h} _ {i} ^ {*} = \\frac {\\sin (\\alpha \\theta)}{\\sin \\theta} \\mathbf {h} _ {i} ^ {\\prime} + \\frac {\\sin ((1 - \\alpha) \\theta)}{\\sin \\theta} \\mathbf {h} _ {j} ^ {\\prime}, \\tag {2}", + "image_path": "5feff9867244fcae742ae85815f65307433fad1d82ee09c881bf127af316fa74.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\alpha \\in [0,1]" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": " is a hyper-parameter controlling the interpolation strength. The intuition behind this approach is that an image with embedding " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{h}_i^*" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": " shares certain visual similarities with the target image " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "v_{i}" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": ". Specifically, a larger " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": " results" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "content": "in " + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "inline_equation", + "content": "\\mathbf{h}_i^*" + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "content": " more closely resembling " + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "inline_equation", + "content": "\\mathbf{h}_i'" + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "content": ", allowing for fine-grained control over the synthesized embedding." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 99, + 554, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 99, + 554, + 326 + ], + "spans": [ + { + "bbox": [ + 313, + 99, + 554, + 326 + ], + "type": "text", + "content": "Modification Text Synthesis. For the augmented image " + }, + { + "bbox": [ + 313, + 99, + 554, + 326 + ], + "type": "inline_equation", + "content": "v_{i}^{\\prime}" + }, + { + "bbox": [ + 313, + 99, + 554, + 326 + ], + "type": "text", + "content": ", its nearest neighbor is " + }, + { + "bbox": [ + 313, + 99, + 554, + 326 + ], + "type": "inline_equation", + "content": "(v_{j}^{\\prime}, w_{j})" + }, + { + "bbox": [ + 313, + 99, + 554, + 326 + ], + "type": "text", + "content": ". To generate the modification text " + }, + { + "bbox": [ + 313, + 99, + 554, + 326 + ], + "type": "inline_equation", + "content": "w_{i}^{*}" + }, + { + "bbox": [ + 313, + 99, + 554, + 326 + ], + "type": "text", + "content": ", we use text interpolation to combine " + }, + { + "bbox": [ + 313, + 99, + 554, + 326 + ], + "type": "inline_equation", + "content": "w_{i}" + }, + { + "bbox": [ + 313, + 99, + 554, + 326 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 99, + 554, + 326 + ], + "type": "inline_equation", + "content": "w_{j}" + }, + { + "bbox": [ + 313, + 99, + 554, + 326 + ], + "type": "text", + "content": " using pre-defined templates, as illustrated in Fig. 3 (right). The complete set of templates is provided in the supplementary material. During training, we randomly select a template for each sample to ensure diversity in the synthesized modification text. It is worth noting that we do not simply use " + }, + { + "bbox": [ + 313, + 99, + 554, + 326 + ], + "type": "inline_equation", + "content": "w_{i}" + }, + { + "bbox": [ + 313, + 99, + 554, + 326 + ], + "type": "text", + "content": " as the modification text for two primary reasons: (i) " + }, + { + "bbox": [ + 313, + 99, + 554, + 326 + ], + "type": "inline_equation", + "content": "w_{i}" + }, + { + "bbox": [ + 313, + 99, + 554, + 326 + ], + "type": "text", + "content": " alone fails to capture the visual differences between the reference image and the target image. (ii) Using only " + }, + { + "bbox": [ + 313, + 99, + 554, + 326 + ], + "type": "inline_equation", + "content": "w_{i}" + }, + { + "bbox": [ + 313, + 99, + 554, + 326 + ], + "type": "text", + "content": " could lead to model cheating, where it learns to rely solely on " + }, + { + "bbox": [ + 313, + 99, + 554, + 326 + ], + "type": "inline_equation", + "content": "w_{i}" + }, + { + "bbox": [ + 313, + 99, + 554, + 326 + ], + "type": "text", + "content": " for retrieval while ignoring the reference image. Our synthesis approach, in contrast, mimics real-world modification text, which often describes both similarities and differences between the reference and target images. Furthermore, this strategy forces the model to learn composed query embeddings by considering both the reference image and the modification text simultaneously." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 326, + 553, + 398 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 326, + 553, + 398 + ], + "spans": [ + { + "bbox": [ + 313, + 326, + 553, + 398 + ], + "type": "text", + "content": "By combining the synthesized reference embedding " + }, + { + "bbox": [ + 313, + 326, + 553, + 398 + ], + "type": "inline_equation", + "content": "\\mathbf{h}_i^*" + }, + { + "bbox": [ + 313, + 326, + 553, + 398 + ], + "type": "text", + "content": " and the interpolated modification text " + }, + { + "bbox": [ + 313, + 326, + 553, + 398 + ], + "type": "inline_equation", + "content": "w_{i}^{*}" + }, + { + "bbox": [ + 313, + 326, + 553, + 398 + ], + "type": "text", + "content": ", our method generates diverse CIR triplets from image-caption pairs. This approach enables effective training of the CoLLM model on large-scale image-caption datasets, effectively mitigating the dependency on scarce labeled triplet data." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 400, + 554, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 400, + 554, + 460 + ], + "spans": [ + { + "bbox": [ + 313, + 400, + 554, + 460 + ], + "type": "text", + "content": "Query Composition with LLM. To construct the composed query, we utilize a pre-trained LLM that processes the synthesized reference image embedding " + }, + { + "bbox": [ + 313, + 400, + 554, + 460 + ], + "type": "inline_equation", + "content": "\\mathbf{h}_i^*" + }, + { + "bbox": [ + 313, + 400, + 554, + 460 + ], + "type": "text", + "content": ", image caption " + }, + { + "bbox": [ + 313, + 400, + 554, + 460 + ], + "type": "inline_equation", + "content": "w_i" + }, + { + "bbox": [ + 313, + 400, + 554, + 460 + ], + "type": "text", + "content": ", and the synthesized modification text " + }, + { + "bbox": [ + 313, + 400, + 554, + 460 + ], + "type": "inline_equation", + "content": "w_i^*" + }, + { + "bbox": [ + 313, + 400, + 554, + 460 + ], + "type": "text", + "content": ". We define three distinct composed embeddings:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 384, + 465, + 553, + 479 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 384, + 465, + 553, + 479 + ], + "spans": [ + { + "bbox": [ + 384, + 465, + 553, + 479 + ], + "type": "interline_equation", + "content": "\\mathbf {c} _ {i} ^ {v} = p \\left(\\Phi \\left(g \\left(\\mathbf {h} _ {i} ^ {*}\\right)\\right)\\right) \\tag {3}", + "image_path": "2cad90c6f871e61c4de8627b63f891516c6b187130ce3200f72ff3a8129ebfa3.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 383, + 480, + 553, + 494 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 383, + 480, + 553, + 494 + ], + "spans": [ + { + "bbox": [ + 383, + 480, + 553, + 494 + ], + "type": "interline_equation", + "content": "\\mathbf {c} _ {i} ^ {w} = p \\left(\\Phi \\left(w _ {i}\\right)\\right) \\tag {4}", + "image_path": "5acf04c3764b9aeaad18c26801a1724d9bb67c796208bbf37c9b3edb8997065b.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 386, + 495, + 553, + 509 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 386, + 495, + 553, + 509 + ], + "spans": [ + { + "bbox": [ + 386, + 495, + 553, + 509 + ], + "type": "interline_equation", + "content": "\\mathbf {c} _ {i} = p \\left(\\Phi \\left([ g \\left(\\mathbf {h} _ {i} ^ {*}\\right); w _ {i} ^ {*} ]\\right)\\right) \\tag {5}", + "image_path": "a260d65435ea45e494b0cb6669a843a2a55a997b939491d5c6a0900782dd9541.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 513, + 553, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 513, + 553, + 537 + ], + "spans": [ + { + "bbox": [ + 313, + 513, + 553, + 537 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 513, + 553, + 537 + ], + "type": "inline_equation", + "content": "[\\cdot ]" + }, + { + "bbox": [ + 313, + 513, + 553, + 537 + ], + "type": "text", + "content": " denotes an instruction template that combine two modalities (see supplementary material)." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 540, + 554, + 600 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 540, + 554, + 600 + ], + "spans": [ + { + "bbox": [ + 313, + 540, + 554, + 600 + ], + "type": "text", + "content": "Training Objective. We employ a contrastive loss " + }, + { + "bbox": [ + 313, + 540, + 554, + 600 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{cl}" + }, + { + "bbox": [ + 313, + 540, + 554, + 600 + ], + "type": "text", + "content": " [44], consistent with previous works [56, 59], to compute the loss between query embeddings " + }, + { + "bbox": [ + 313, + 540, + 554, + 600 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{c}_i^v,\\mathbf{c}_i^w,\\mathbf{c}_i\\}" + }, + { + "bbox": [ + 313, + 540, + 554, + 600 + ], + "type": "text", + "content": " and the target image embedding " + }, + { + "bbox": [ + 313, + 540, + 554, + 600 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_i = f(v_i)" + }, + { + "bbox": [ + 313, + 540, + 554, + 600 + ], + "type": "text", + "content": ". The final loss for each sample during pre-training is defined as:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 329, + 604, + 553, + 626 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 329, + 604, + 553, + 626 + ], + "spans": [ + { + "bbox": [ + 329, + 604, + 553, + 626 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\frac {1}{3} \\left(\\mathcal {L} _ {c l} \\left(\\mathbf {c} _ {i} ^ {v}, \\mathbf {z} _ {i}\\right) + \\mathcal {L} _ {c l} \\left(\\mathbf {c} _ {i} ^ {w}, \\mathbf {z} _ {i}\\right) + \\mathcal {L} _ {c l} \\left(\\mathbf {c} _ {i}, \\mathbf {z} _ {i}\\right)\\right) \\tag {6}", + "image_path": "64596168c6f447b38ecceba33ff10a31e6400778f2d8f3b3f381ff1bdc29b05c.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 629, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 554, + 713 + ], + "type": "text", + "content": "This formulation encourages the model to learn discriminative representations for different query types while maintaining consistency with the target image embedding. By combining losses from image-to-image " + }, + { + "bbox": [ + 313, + 629, + 554, + 713 + ], + "type": "inline_equation", + "content": "(\\mathbf{c}_i^v)" + }, + { + "bbox": [ + 313, + 629, + 554, + 713 + ], + "type": "text", + "content": ", text-to-image " + }, + { + "bbox": [ + 313, + 629, + 554, + 713 + ], + "type": "inline_equation", + "content": "(\\mathbf{c}_i^w)" + }, + { + "bbox": [ + 313, + 629, + 554, + 713 + ], + "type": "text", + "content": ", and composed " + }, + { + "bbox": [ + 313, + 629, + 554, + 713 + ], + "type": "inline_equation", + "content": "(\\mathbf{c}_i)" + }, + { + "bbox": [ + 313, + 629, + 554, + 713 + ], + "type": "text", + "content": " queries, we ensure that the model learns to effectively process and align various input modalities for compositional image retrieval." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3997" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 64, + 70, + 545, + 196 + ], + "blocks": [ + { + "bbox": [ + 203, + 56, + 406, + 67 + ], + "lines": [ + { + "bbox": [ + 203, + 56, + 406, + 67 + ], + "spans": [ + { + "bbox": [ + 203, + 56, + 406, + 67 + ], + "type": "text", + "content": "Table 1. Comparison of synthetic CIR training datasets." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 64, + 70, + 545, + 196 + ], + "lines": [ + { + "bbox": [ + 64, + 70, + 545, + 196 + ], + "spans": [ + { + "bbox": [ + 64, + 70, + 545, + 196 + ], + "type": "table", + "html": "
DatasetPublicReference-Target Image PairModification Text Generation
TypeSource# Pairs# EntitiesLLMLLM input# Text# Text/Pair
LaSCo [28]✓imageVQAv2 [13]360K82KGPT-3 [6]question-answer360K1
VDG [23]✓imageNLVR2 [54], COCO [32]467K183KVision-LLaMA2image pair523K1.12
WebCoVR [59]✓videoWebVid2M [3]1.6M131KCustom LLMcaption1.6M1
CC-CoIR [58]✓imageCC3M [51]3.3M357KCustom LLMcaption3.3M1
SynthTriplets [14]synth. imageSD [46], IP2P [5]18.8M37.6MOPT-6.7B [65]-18.8M1
MagicLens [64]imageWeb crawled36.5M-PaLM [7]image metadata36.5M1
MTCIR (Ours)✓imageLLaVA-558k [34]3.4M423KSonnet 3 [2]synth. caption17.7M5.18
", + "image_path": "14f36abf0ea4db2e3ced31de7aed85e62ff4976fca677c3fe01b3ec8fe140e7e.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 74, + 209, + 173, + 267 + ], + "blocks": [ + { + "bbox": [ + 99, + 201, + 150, + 209 + ], + "lines": [ + { + "bbox": [ + 99, + 201, + 150, + 209 + ], + "spans": [ + { + "bbox": [ + 99, + 201, + 150, + 209 + ], + "type": "text", + "content": "Reference Image" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 74, + 209, + 173, + 267 + ], + "lines": [ + { + "bbox": [ + 74, + 209, + 173, + 267 + ], + "spans": [ + { + "bbox": [ + 74, + 209, + 173, + 267 + ], + "type": "image", + "image_path": "393b355a1fc2f9cf39fd0ea505f05b23f01f7cacb1e97084e954d00c07e6619a.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 77, + 270, + 106, + 278 + ], + "lines": [ + { + "bbox": [ + 77, + 270, + 106, + 278 + ], + "spans": [ + { + "bbox": [ + 77, + 270, + 106, + 278 + ], + "type": "text", + "content": "Category" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 76, + 278, + 124, + 286 + ], + "lines": [ + { + "bbox": [ + 76, + 278, + 124, + 286 + ], + "spans": [ + { + "bbox": [ + 76, + 278, + 124, + 286 + ], + "type": "text", + "content": "+ Object Added" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 76, + 295, + 132, + 303 + ], + "lines": [ + { + "bbox": [ + 76, + 295, + 132, + 303 + ], + "spans": [ + { + "bbox": [ + 76, + 295, + 132, + 303 + ], + "type": "text", + "content": "Object Removed" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 167, + 270, + 221, + 277 + ], + "lines": [ + { + "bbox": [ + 167, + 270, + 221, + 277 + ], + "spans": [ + { + "bbox": [ + 167, + 270, + 221, + 277 + ], + "type": "text", + "content": "Modification Text" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 167, + 278, + 271, + 286 + ], + "lines": [ + { + "bbox": [ + 167, + 278, + 271, + 286 + ], + "spans": [ + { + "bbox": [ + 167, + 278, + 271, + 286 + ], + "type": "text", + "content": "Add a person's hand holding the cup." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 167, + 287, + 257, + 295 + ], + "lines": [ + { + "bbox": [ + 167, + 287, + 257, + 295 + ], + "spans": [ + { + "bbox": [ + 167, + 287, + 257, + 295 + ], + "type": "text", + "content": "Modify the liquid being poured." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 167, + 295, + 272, + 303 + ], + "lines": [ + { + "bbox": [ + 167, + 295, + 272, + 303 + ], + "spans": [ + { + "bbox": [ + 167, + 295, + 272, + 303 + ], + "type": "text", + "content": "Remove the social media text overlay." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 175, + 209, + 276, + 267 + ], + "blocks": [ + { + "bbox": [ + 203, + 201, + 244, + 209 + ], + "lines": [ + { + "bbox": [ + 203, + 201, + 244, + 209 + ], + "spans": [ + { + "bbox": [ + 203, + 201, + 244, + 209 + ], + "type": "text", + "content": "Target Image" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 175, + 209, + 276, + 267 + ], + "lines": [ + { + "bbox": [ + 175, + 209, + 276, + 267 + ], + "spans": [ + { + "bbox": [ + 175, + 209, + 276, + 267 + ], + "type": "image", + "image_path": "dd58bd807d019a2cdf28e643f4181a82c4c754ec2d6f6cd1f3e93c9442c9b365.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 76, + 287, + 136, + 295 + ], + "lines": [ + { + "bbox": [ + 76, + 287, + 136, + 295 + ], + "spans": [ + { + "bbox": [ + 76, + 287, + 136, + 295 + ], + "type": "inline_equation", + "content": "\\mathbf{3}" + }, + { + "bbox": [ + 76, + 287, + 136, + 295 + ], + "type": "text", + "content": " Attribute Changed" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 335, + 239, + 348 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 335, + 239, + 348 + ], + "spans": [ + { + "bbox": [ + 55, + 335, + 239, + 348 + ], + "type": "text", + "content": "3.3. CoLLM with CIR Triplet as Input" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 55, + 352, + 296, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 352, + 296, + 449 + ], + "spans": [ + { + "bbox": [ + 55, + 352, + 296, + 449 + ], + "type": "text", + "content": "For CIR triplet inputs, our model design is shown in Fig. 2b. Let " + }, + { + "bbox": [ + 55, + 352, + 296, + 449 + ], + "type": "inline_equation", + "content": "\\mathcal{X} = \\{(v_i^r,v_i^t,w_i)\\}_{i = 1}^N" + }, + { + "bbox": [ + 55, + 352, + 296, + 449 + ], + "type": "text", + "content": " be a set of triplets, where " + }, + { + "bbox": [ + 55, + 352, + 296, + 449 + ], + "type": "inline_equation", + "content": "v_{i}^{r}" + }, + { + "bbox": [ + 55, + 352, + 296, + 449 + ], + "type": "inline_equation", + "content": "v_{i}^{t}" + }, + { + "bbox": [ + 55, + 352, + 296, + 449 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 352, + 296, + 449 + ], + "type": "inline_equation", + "content": "w_{i}" + }, + { + "bbox": [ + 55, + 352, + 296, + 449 + ], + "type": "text", + "content": " represent the reference image, target image, and modification text of the " + }, + { + "bbox": [ + 55, + 352, + 296, + 449 + ], + "type": "inline_equation", + "content": "i^{th}" + }, + { + "bbox": [ + 55, + 352, + 296, + 449 + ], + "type": "text", + "content": " sample, respectively. The reference embedding is obtained as " + }, + { + "bbox": [ + 55, + 352, + 296, + 449 + ], + "type": "inline_equation", + "content": "\\mathbf{h}_i = f(v_i^r)" + }, + { + "bbox": [ + 55, + 352, + 296, + 449 + ], + "type": "text", + "content": ". The composed embedding is then computed as " + }, + { + "bbox": [ + 55, + 352, + 296, + 449 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_i = p(\\Phi ([g(\\mathbf{h}_i);w_i])" + }, + { + "bbox": [ + 55, + 352, + 296, + 449 + ], + "type": "text", + "content": ". The training objective is " + }, + { + "bbox": [ + 55, + 352, + 296, + 449 + ], + "type": "inline_equation", + "content": "\\mathcal{L} = \\mathcal{L}_{cl}(\\mathbf{c}_i,\\mathbf{z}_i)" + }, + { + "bbox": [ + 55, + 352, + 296, + 449 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 352, + 296, + 449 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_i = f(v_i^t)" + }, + { + "bbox": [ + 55, + 352, + 296, + 449 + ], + "type": "text", + "content": " denotes the target embedding." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 55, + 454, + 180, + 466 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 454, + 180, + 466 + ], + "spans": [ + { + "bbox": [ + 55, + 454, + 180, + 466 + ], + "type": "text", + "content": "4. Dataset Construction" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 55, + 473, + 296, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 473, + 296, + 545 + ], + "spans": [ + { + "bbox": [ + 55, + 473, + 296, + 545 + ], + "type": "text", + "content": "This section outlines the data construction process for two primary objectives: (1) creating a novel CIR training dataset called Multi-Text CIR (MTCIR) and (2) refining widely used CIR benchmarks, specifically CIRR [37] and FashionIQ [61]. We detail the methodologies for each task, highlighting the improvements and innovations." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 55, + 552, + 237, + 565 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 552, + 237, + 565 + ], + "spans": [ + { + "bbox": [ + 55, + 552, + 237, + 565 + ], + "type": "text", + "content": "4.1. Multi-Text CIR (MTCIR) Dataset" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 55, + 570, + 295, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 295, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 295, + 605 + ], + "type": "text", + "content": "Existing CIR triplet datasets [23, 28, 59] lack diversity and contain unnatural modification text. We introduce Multi-Text CIR (MTCIR) to address these limitations." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "text", + "content": "To enhance image diversity, we source images from the LLaVA-558k dataset [34] filtered from image caption datasets [41, 50, 51] based on noun-phrase frequency for broad concept coverage. We pair images using CLIP [45] visual similarity metrics, following CIRR's grouping approach, yielding 3.4M pairs from approximately 423K images. Leveraging the detailed captions generated by LLaVA-Next-34B [35] for each image in the LLaVA-558k dataset, we employ Claude 3 Sonnet [2]" + } + ] + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 318, + 215, + 374, + 275 + ], + "blocks": [ + { + "bbox": [ + 319, + 201, + 369, + 212 + ], + "lines": [ + { + "bbox": [ + 319, + 201, + 369, + 212 + ], + "spans": [ + { + "bbox": [ + 319, + 201, + 369, + 212 + ], + "type": "text", + "content": "Unchanged" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 318, + 215, + 374, + 275 + ], + "lines": [ + { + "bbox": [ + 318, + 215, + 374, + 275 + ], + "spans": [ + { + "bbox": [ + 318, + 215, + 374, + 275 + ], + "type": "image", + "image_path": "e47c0a1e12446a6ddd82a5b20250169c508796a04007e3a4b35f9b13485c9e44.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 331, + 278, + 361, + 286 + ], + "lines": [ + { + "bbox": [ + 331, + 278, + 361, + 286 + ], + "spans": [ + { + "bbox": [ + 331, + 278, + 361, + 286 + ], + "type": "text", + "content": "(a) CIRR" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 377, + 215, + 433, + 275 + ], + "blocks": [ + { + "bbox": [ + 380, + 201, + 424, + 212 + ], + "lines": [ + { + "bbox": [ + 380, + 201, + 424, + 212 + ], + "spans": [ + { + "bbox": [ + 380, + 201, + 424, + 212 + ], + "type": "text", + "content": "Regenerate" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 377, + 215, + 433, + 275 + ], + "lines": [ + { + "bbox": [ + 377, + 215, + 433, + 275 + ], + "spans": [ + { + "bbox": [ + 377, + 215, + 433, + 275 + ], + "type": "image", + "image_path": "2c16b8e1361659c01de2b3d8d7b5e1c425d74ed5dde92b11fd7245a35a6c2859.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 382, + 277, + 428, + 287 + ], + "lines": [ + { + "bbox": [ + 382, + 277, + 428, + 287 + ], + "spans": [ + { + "bbox": [ + 382, + 277, + 428, + 287 + ], + "type": "text", + "content": "(b) FIQ-Dress" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 436, + 215, + 492, + 275 + ], + "blocks": [ + { + "bbox": [ + 439, + 201, + 476, + 212 + ], + "lines": [ + { + "bbox": [ + 439, + 201, + 476, + 212 + ], + "spans": [ + { + "bbox": [ + 439, + 201, + 476, + 212 + ], + "type": "text", + "content": "Remove" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 436, + 215, + 492, + 275 + ], + "lines": [ + { + "bbox": [ + 436, + 215, + 492, + 275 + ], + "spans": [ + { + "bbox": [ + 436, + 215, + 492, + 275 + ], + "type": "image", + "image_path": "5b04838b1d5c159594bbb52c8b859c4a3295a1644a686d8a85eea5d0cea863cc.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 443, + 277, + 486, + 287 + ], + "lines": [ + { + "bbox": [ + 443, + 277, + 486, + 287 + ], + "spans": [ + { + "bbox": [ + 443, + 277, + 486, + 287 + ], + "type": "text", + "content": "(c) FIQ-Shirt" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 491, + 201, + 545, + 212 + ], + "lines": [ + { + "bbox": [ + 491, + 201, + 545, + 212 + ], + "spans": [ + { + "bbox": [ + 491, + 201, + 545, + 212 + ], + "type": "text", + "content": "Cannot process" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 495, + 215, + 551, + 275 + ], + "blocks": [ + { + "bbox": [ + 495, + 215, + 551, + 275 + ], + "lines": [ + { + "bbox": [ + 495, + 215, + 551, + 275 + ], + "spans": [ + { + "bbox": [ + 495, + 215, + 551, + 275 + ], + "type": "image", + "image_path": "3c3f65416323df20ea4b3345010a918dc2bc39886d72cdee015d046a611b79e9.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 498, + 277, + 548, + 288 + ], + "lines": [ + { + "bbox": [ + 498, + 277, + 548, + 288 + ], + "spans": [ + { + "bbox": [ + 498, + 277, + 548, + 288 + ], + "type": "text", + "content": "(d) FIQ-Toptee" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_caption" + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 334, + 334, + 400, + 383 + ], + "blocks": [ + { + "bbox": [ + 55, + 308, + 296, + 330 + ], + "lines": [ + { + "bbox": [ + 55, + 308, + 296, + 330 + ], + "spans": [ + { + "bbox": [ + 55, + 308, + 296, + 330 + ], + "type": "text", + "content": "Figure 4. An example from the MTCIR dataset. Each sample contains multiple short texts describing different modifications." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 297, + 554, + 319 + ], + "lines": [ + { + "bbox": [ + 313, + 297, + 554, + 319 + ], + "spans": [ + { + "bbox": [ + 313, + 297, + 554, + 319 + ], + "type": "text", + "content": "Figure 5. Statistical analysis of the refinement process for CIRR and FashionIQ (FIQ) datasets." + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 341, + 327, + 391, + 334 + ], + "lines": [ + { + "bbox": [ + 341, + 327, + 391, + 334 + ], + "spans": [ + { + "bbox": [ + 341, + 327, + 391, + 334 + ], + "type": "text", + "content": "Reference Image" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 334, + 334, + 400, + 383 + ], + "lines": [ + { + "bbox": [ + 334, + 334, + 400, + 383 + ], + "spans": [ + { + "bbox": [ + 334, + 334, + 400, + 383 + ], + "type": "image", + "image_path": "44b80bdf9e6caa3bf59217f12cc9164594365c38dbcb6efe490fd782bbfb0bb2.jpg" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 336, + 383, + 386, + 392 + ], + "lines": [ + { + "bbox": [ + 336, + 383, + 386, + 392 + ], + "spans": [ + { + "bbox": [ + 336, + 383, + 386, + 392 + ], + "type": "text", + "content": "Original Text" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 335, + 396, + 398, + 405 + ], + "lines": [ + { + "bbox": [ + 335, + 396, + 398, + 405 + ], + "spans": [ + { + "bbox": [ + 335, + 396, + 398, + 405 + ], + "type": "text", + "content": "Regenerated Text" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_caption" + } + ], + "index": 38 + }, + { + "type": "image", + "bbox": [ + 401, + 334, + 467, + 383 + ], + "blocks": [ + { + "bbox": [ + 414, + 327, + 453, + 334 + ], + "lines": [ + { + "bbox": [ + 414, + 327, + 453, + 334 + ], + "spans": [ + { + "bbox": [ + 414, + 327, + 453, + 334 + ], + "type": "text", + "content": "Target Image" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 401, + 334, + 467, + 383 + ], + "lines": [ + { + "bbox": [ + 401, + 334, + 467, + 383 + ], + "spans": [ + { + "bbox": [ + 401, + 334, + 467, + 383 + ], + "type": "image", + "image_path": "cd682d24ce9cdbb9b9df1a60dadafb80979bc461311654ac5467f738a2f8fbf1.jpg" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 410, + 384, + 504, + 391 + ], + "lines": [ + { + "bbox": [ + 410, + 384, + 504, + 391 + ], + "spans": [ + { + "bbox": [ + 410, + 384, + 504, + 391 + ], + "type": "text", + "content": "One puppy looking at the camera." + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 410, + 392, + 523, + 399 + ], + "lines": [ + { + "bbox": [ + 410, + 392, + 523, + 399 + ], + "spans": [ + { + "bbox": [ + 410, + 392, + 523, + 399 + ], + "type": "text", + "content": "The dog is indoors on a tiled floor with a" + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 410, + 399, + 460, + 407 + ], + "lines": [ + { + "bbox": [ + 410, + 399, + 460, + 407 + ], + "spans": [ + { + "bbox": [ + 410, + 399, + 460, + 407 + ], + "type": "text", + "content": "food bowl nearby." + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 42 + }, + { + "bbox": [ + 313, + 413, + 554, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 413, + 554, + 468 + ], + "spans": [ + { + "bbox": [ + 313, + 413, + 554, + 468 + ], + "type": "text", + "content": "Figure 6. An example from the refined CIRR dataset. The original modification text yields multiple plausible results (shown as similar images to the target on the right). In contrast, the newly generated text accurately captures the distinctive features of the target image, differentiating it from other candidates." + } + ] + } + ], + "index": 47 + }, + { + "bbox": [ + 313, + 474, + 554, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 474, + 554, + 569 + ], + "spans": [ + { + "bbox": [ + 313, + 474, + 554, + 569 + ], + "type": "text", + "content": "(20240229-v1:0) to generate modification text for each image pair, using these captions as input. To control detail levels in Claude outputs, we define six modification text categories in the prompt, inspired by prior research [4, 37]. This approach maintains diverse aspects and prevents repetition. We also incorporate examples from the CIRR dataset to ensure human-like modification text. Detailed prompting strategies are provided in the supplementary material." + } + ] + } + ], + "index": 48 + }, + { + "bbox": [ + 313, + 570, + 554, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 554, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 554, + 689 + ], + "type": "text", + "content": "Our method generates multiple brief, focused texts for each image pair, each highlighting a unique aspect. This approach captures all changes between two images without producing overly long sentences, better aligning with human-written modification text. These short texts can be combined for training, enhancing dataset diversity. We obtained 17.7M text samples for the 3.4M image pairs, averaging 5.18 short sentences per pair. Fig. 4 illustrates an example, showcasing modification texts corresponding to specific categories and covering diverse attributes." + } + ] + } + ], + "index": 49 + }, + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "content": "As shown in Table 1, MTCIR offers several advantages over existing synthetic CIR datasets: (i) Unlike [14], which" + } + ] + } + ], + "index": 50 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3998" + } + ] + } + ], + "index": 51 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 125, + 298, + 350 + ], + "blocks": [ + { + "bbox": [ + 55, + 56, + 295, + 121 + ], + "lines": [ + { + "bbox": [ + 55, + 56, + 295, + 121 + ], + "spans": [ + { + "bbox": [ + 55, + 56, + 295, + 121 + ], + "type": "text", + "content": "Table 2. Performance comparison of various models and visual encoders on CIR benchmarks, evaluated without training on triplet datasets. Bold and underlined values denote the best and second-best scores within each vision encoder group. Models incorporating LLMs in their architectures are marked with " + }, + { + "bbox": [ + 55, + 56, + 295, + 121 + ], + "type": "inline_equation", + "content": "\\star" + }, + { + "bbox": [ + 55, + 56, + 295, + 121 + ], + "type": "text", + "content": ". Results reproduced by our team are indicated by " + }, + { + "bbox": [ + 55, + 56, + 295, + 121 + ], + "type": "inline_equation", + "content": "\\ddagger" + }, + { + "bbox": [ + 55, + 56, + 295, + 121 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 125, + 298, + 350 + ], + "lines": [ + { + "bbox": [ + 56, + 125, + 298, + 350 + ], + "spans": [ + { + "bbox": [ + 56, + 125, + 298, + 350 + ], + "type": "table", + "html": "
MethodCIRCO (mAP↑)CIRR (Rec.↑)FIQ (Rec.↑)
@5@10@50@1@10@50@10@50
OpenAI CLIP-B/32
PALAVRA [8]4.65.36.816.658.584.019.837.3
SEARLE [4]9.49.911.824.066.889.822.942.5
Slerp-TAT [21]9.310.312.328.268.888.523.044.0
CIREVL* [26]14.915.417.823.966.087.028.349.4
CoLLM*12.913.215.028.671.892.724.845.2
OpenAI CLIP-L/14
Pic2World [49]8.79.511.323.965.387.824.743.7
SEARLE [4]11.712.715.124.266.388.825.646.2
LinCIR‡[15]13.013.916.224.666.988.826.446.6
ContextI2W [55]13.013.816.025.768.689.627.848.9
MCL* [30]17.818.421.825.969.491.0--
Slerp-TAT [21]18.519.421.430.970.989.228.347.6
CIREVL* [26]18.619.021.824.664.986.328.648.6
CoLLM*20.320.823.429.772.891.530.149.5
BLIP-L/16
Slerp-TAT [21]17.818.421.134.072.788.932.853.3
CoLLM*19.720.423.135.078.694.234.656.0
", + "image_path": "6ec923a0d09b6a0f170e1e52cf64dddf64c8a1fdd38165c4bca89faa45a9246a.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 358, + 295, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 358, + 295, + 418 + ], + "spans": [ + { + "bbox": [ + 55, + 358, + 295, + 418 + ], + "type": "text", + "content": "uses synthetic images, MTCIR uses real images covering diverse concepts and domains. (ii) MTCIR provides multiple modification texts for each image pair, offering more natural descriptions and training flexibility. (iii) MTCIR is the largest public dataset to advance CIR research." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 419, + 296, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 419, + 296, + 479 + ], + "spans": [ + { + "bbox": [ + 55, + 419, + 296, + 479 + ], + "type": "text", + "content": "To prevent the leakage of biometric information, we further process MTCIR at both the image and text levels. For images, we apply face blurring to all instances. For text, we remove any content containing keywords related to human attributes (e.g., skin, hair, gender, age, race)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 491, + 264, + 503 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 491, + 264, + 503 + ], + "spans": [ + { + "bbox": [ + 55, + 491, + 264, + 503 + ], + "type": "text", + "content": "4.2. Refined CIRR and Fashion-IQ Datasets" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 509, + 295, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 509, + 295, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 509, + 295, + 605 + ], + "type": "text", + "content": "Current CIR benchmarks often suffer from label ambiguity, where modification text lacks clarity, potentially matching multiple target images. To address this issue, we propose a refinement method using Claude 3 Sonnet [2] (20240229-v1:0) to enhance the validation sets of two well-known benchmarks: CIRR and Fashion-IQ. Each sample in these validation sets comprises a triplet containing a reference image, target image, and modification text." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 606, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 606, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 606, + 295, + 714 + ], + "type": "text", + "content": "Our refinement process, applicable to any benchmark, consists of validation, re-generation, and re-validation steps. For each triplet, we first identify the top-3 visually similar images to the target image based on CLIP visual scores, serving as hard negative samples for ambiguity assessment. We employ Claude 3 Sonnet to evaluate each triplet's ambiguity by attempting to identify the target image from these hard negatives. Triplets correctly identified by Claude 3 Sonnet are considered \"good\" samples and re" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 313, + 102, + 555, + 260 + ], + "blocks": [ + { + "bbox": [ + 313, + 56, + 555, + 99 + ], + "lines": [ + { + "bbox": [ + 313, + 56, + 555, + 99 + ], + "spans": [ + { + "bbox": [ + 313, + 56, + 555, + 99 + ], + "type": "text", + "content": "Table 3. Performance of models trained on synthetic triplet datasets. Bold indicates the best method overall, while underline highlights the best method within the same vision encoder. Our CoLLM outperforms comparable methods in most metrics." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 313, + 102, + 555, + 260 + ], + "lines": [ + { + "bbox": [ + 313, + 102, + 555, + 260 + ], + "spans": [ + { + "bbox": [ + 313, + 102, + 555, + 260 + ], + "type": "table", + "html": "
MethodDatasetCIRR ↑FIQ ↑
@1@10@50@10@50
CoCa-L/18 288 × 288
MagicLens [64]MagicLens [64]33.377.994.438.158.3
EVA-CLIP ViT-G/14 364 × 364
CoVR2 [58]WV-CC-CoVIR [58]43.784.096.138.258.4
OpenAI CLIP-L/14 224 × 224
CompoDiff [14]SynTrip18M [14]18.270.890.336.048.6
MagicLens [64]MagicLens [64]30.174.492.630.752.5
CoLLMMTCIR (ours)34.777.093.132.954.2
BLIP-L/16 384 × 384; fine-tuned on COCO
CASE [28]LaSCo [28]35.478.594.6--
Omkar et al. [56]WebCoVR [59]40.178.994.730.346.5
CoLLMMTCIR (ours)45.884.795.839.160.7
", + "image_path": "5e988e21aa79d61f3644d8028c86e95cbba952559c19b07b5190623fdf7bbd25.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 266, + 555, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 266, + 555, + 422 + ], + "spans": [ + { + "bbox": [ + 313, + 266, + 555, + 422 + ], + "type": "text", + "content": "main unchanged. For \"bad\" samples, we use Claude 3 Sonnet to re-generate the modification text. The model produces three new modification texts with increasing detail in a hierarchical structure, with each finer text building upon the coarser one. We then repeat the validation process, evaluating the new texts from coarsest to finest and selecting the best one. Any re-generated modification text that passes the validation process replaces the original text and concludes the process. If all generated texts fail the validation, the triplet is removed from the benchmark. Detailed process information is provided in the supplementary material. Besides, following MTCIR, biometric information is also removed from the refined benchmarks." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 423, + 556, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 423, + 556, + 519 + ], + "spans": [ + { + "bbox": [ + 313, + 423, + 556, + 519 + ], + "type": "text", + "content": "As illustrated in Fig. 5, up to " + }, + { + "bbox": [ + 313, + 423, + 556, + 519 + ], + "type": "inline_equation", + "content": "8\\%" + }, + { + "bbox": [ + 313, + 423, + 556, + 519 + ], + "type": "text", + "content": " of samples are omitted due to Claude's processing limitations related to harmful content. Additionally, " + }, + { + "bbox": [ + 313, + 423, + 556, + 519 + ], + "type": "inline_equation", + "content": "4 - 7\\%" + }, + { + "bbox": [ + 313, + 423, + 556, + 519 + ], + "type": "text", + "content": " of ambiguous samples that could not be effectively rewritten are removed. More than " + }, + { + "bbox": [ + 313, + 423, + 556, + 519 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 313, + 423, + 556, + 519 + ], + "type": "text", + "content": " of the original samples are retained, and " + }, + { + "bbox": [ + 313, + 423, + 556, + 519 + ], + "type": "inline_equation", + "content": "40 - 55\\%" + }, + { + "bbox": [ + 313, + 423, + 556, + 519 + ], + "type": "text", + "content": " of triplets are successfully revised. The refined CIRR and Fashion-IQ datasets exhibit less ambiguity (Fig. 6), offering more robust evaluation benchmarks for the CIR community." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 531, + 397, + 545 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 531, + 397, + 545 + ], + "spans": [ + { + "bbox": [ + 313, + 531, + 397, + 545 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 552, + 555, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 552, + 555, + 635 + ], + "spans": [ + { + "bbox": [ + 313, + 552, + 555, + 635 + ], + "type": "text", + "content": "Our proposed CoLLM framework adopts a two-stage training paradigm: pre-training and fine-tuning phases. In pretraining, CoLLM is exposed to a diverse corpus of image-capture pairs, enabling it to understand composed queries comprehensively. Subsequently, in the fine-tuning stage, we leverage our newly curated MTCIR datasets to verify their effective performance on complex multimodal tasks." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 645, + 512, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 645, + 512, + 658 + ], + "spans": [ + { + "bbox": [ + 313, + 645, + 512, + 658 + ], + "type": "text", + "content": "5.1. Pre-Training on Image-Caption Pairs" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "text", + "content": "Training and Evaluation Datasets. The pre-training dataset consists of 5 million image-caption pairs, compiled from CC3M [51], LAION [50], and LLaVA-558K [34]. We evaluate our model on three CIR benchmarks: CIRCO [4]," + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3999" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 102, + 296, + 183 + ], + "blocks": [ + { + "bbox": [ + 55, + 56, + 295, + 100 + ], + "lines": [ + { + "bbox": [ + 55, + 56, + 295, + 100 + ], + "spans": [ + { + "bbox": [ + 55, + 56, + 295, + 100 + ], + "type": "text", + "content": "Table 4. Performance of BLIP-L and CoLLM (pre-trained) trained on synthetic triplet datasets. Bold values indicate the best score within each group. Models trained on MTCIR demonstrate superior performance compared to those trained on other datasets." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 102, + 296, + 183 + ], + "lines": [ + { + "bbox": [ + 56, + 102, + 296, + 183 + ], + "spans": [ + { + "bbox": [ + 56, + 102, + 296, + 183 + ], + "type": "table", + "html": "
MethodDatasetCIRR ↑FIQ ↑
@1@10@50@10@50
BLIP-L [29]LaSCo [28]36.678.494.624.844.0
WebCoVR [59]39.378.994.726.743.3
MTCIR (ours)42.483.195.937.959.2
CoLLMLaSCo [28]43.284.195.738.560.1
MTCIR (ours)45.884.795.939.160.7
", + "image_path": "2d9db65b9e6b6723183423a9e9bc44b7d5ceeaa209a97b81c301686a737d9379.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 187, + 295, + 211 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 187, + 295, + 211 + ], + "spans": [ + { + "bbox": [ + 55, + 187, + 295, + 211 + ], + "type": "text", + "content": "CIRR [37] test set, and Fashion-IQ [61] validation set. All datasets are filtered to exclude biometric information." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 214, + 296, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 214, + 296, + 275 + ], + "spans": [ + { + "bbox": [ + 55, + 214, + 296, + 275 + ], + "type": "text", + "content": "Evaluation Metrics. We employ recall on the complete image index set as the evaluation metric for CIRR and Fashion-IQ. For CIRCO, we use mean Average Precision (mAP). We exclude the recall on subset for CIRR due to reliability concerns (see supplementary material for details)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 277, + 278, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 277, + 278, + 289 + ], + "spans": [ + { + "bbox": [ + 55, + 277, + 278, + 289 + ], + "type": "text", + "content": "Implementation Details. See supplementary material." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 292, + 295, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 292, + 295, + 411 + ], + "spans": [ + { + "bbox": [ + 54, + 292, + 295, + 411 + ], + "type": "text", + "content": "Quantitative Results. Table 2 compares our pre-trained model with other methods not trained on triplet datasets. We evaluate three variants of our architecture, each employing different vision encoders. CoLLM with the CLIP-B encoder achieves the second-best performance on the CIRCO and Fashion-IQ benchmarks while significantly improving on the CIRR benchmark. Although CIREVL [26] outperforms other methods, its use of the closed-source GPT-4 [1] introduces additional costs and increased inference latency, making it an unfair comparison." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 411, + 295, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 411, + 295, + 590 + ], + "spans": [ + { + "bbox": [ + 55, + 411, + 295, + 590 + ], + "type": "text", + "content": "For larger vision encoders, our model consistently outperforms both non-LLM (e.g., Slerp-TAT [21]) and LLM (CIReVL and MCL [30]) methods across all benchmarks. Notably, our CLIP-L variant achieves state-of-the-art results on the CIRCO benchmark, while BLIP-L significantly improves CIRR and Fashion-IQ scores. The consistent gains across benchmarks and architectures validate the robustness and generalizability of our triplet synthesis approach. By effectively bridging the gap between abundant image-caption data and scarce triplet annotations, our method paves the way for more scalable and efficient training of CIR models. This innovative strategy represents a significant step forward in addressing one of the critical challenges in the field, Paving the way for larger-scale pre-training and more complex multimodal interactions." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 597, + 188, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 597, + 188, + 609 + ], + "spans": [ + { + "bbox": [ + 55, + 597, + 188, + 609 + ], + "type": "text", + "content": "5.2. Fine-tuning on MTCIR" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "text", + "content": "Training Dataset and Benchmarks. Despite the impressive performance achieved using only image-caption datasets, we hypothesize that CoLLM's performance can be further enhanced through fine-tuning on our newly developed MTCIR dataset. For a fair comparison, we also train CoLLM and baseline models on other public datasets, including WebCoVR [59] and LaSCo [28]. Our evaluation focuses on the CIRR [37] test and Fashion-IQ [61] validation." + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 313, + 113, + 558, + 239 + ], + "blocks": [ + { + "bbox": [ + 313, + 56, + 555, + 110 + ], + "lines": [ + { + "bbox": [ + 313, + 56, + 555, + 110 + ], + "spans": [ + { + "bbox": [ + 313, + 56, + 555, + 110 + ], + "type": "text", + "content": "Table 5. Performance of models on refined CIRR and Fashion-IQ validation sets. Bold indicates the highest score, while underlined values represent the best metric within the same vision encoder group. CoLLM+MTCIR continues to outperform other configurations on these new benchmarks." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 313, + 113, + 558, + 239 + ], + "lines": [ + { + "bbox": [ + 313, + 113, + 558, + 239 + ], + "spans": [ + { + "bbox": [ + 313, + 113, + 558, + 239 + ], + "type": "table", + "html": "
MethodDatasetRef. CIRR ↑Ref. FIQ ↑
@1@10@50@10@50
EVA-CLIP ViT-G/14 364 × 364
CoVR2 [58]WV-CC-VIR [58]56.191.197.954.272.9
OpenAI CLIP-L/16 224 × 224
MagicLens [64]MagicLens [64]42.386.397.045.568.1
CoLLMMTCIR (ours)46.587.496.148.368.6
BLIP-L/16 384 × 384; fine-tuned on COCO
BLIP-L [29]MTCIR (ours)53.889.997.754.874.3
CoLLMLaSCo [28]57.392.098.156.975.9
CoLLMMTCIR (ours)60.492.698.257.276.4
", + "image_path": "22f3f52a74dcbbdbddbd53a388bcb2d6d1ee94072de2517a296f2c1274e7e486.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 243, + 555, + 279 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 243, + 555, + 279 + ], + "spans": [ + { + "bbox": [ + 313, + 243, + 555, + 279 + ], + "type": "text", + "content": "tion sets. We exclude CIRCO [4] due to potential domain overlap between COCO [32] images used in CIRCO and those in LaSCo [28], as well as in the pre-training datasets." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 281, + 537, + 294 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 281, + 537, + 294 + ], + "spans": [ + { + "bbox": [ + 313, + 281, + 537, + 294 + ], + "type": "text", + "content": "Implementation Details. See supplementary material." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 297, + 555, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 297, + 555, + 464 + ], + "spans": [ + { + "bbox": [ + 313, + 297, + 555, + 464 + ], + "type": "text", + "content": "Quantitative Results. Table 3 presents the evaluation of models trained on publicly available synthetic CIR datasets, including our MTCIR. Fine-tuning CoLLM on MTCIR yields significant gains, improving CIRR@1 by " + }, + { + "bbox": [ + 313, + 297, + 555, + 464 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 313, + 297, + 555, + 464 + ], + "type": "text", + "content": " over CoLLM-OpenAI-CLIP-L/14 in Table 2. This improvement reflects MTCIR's enhanced generalizability and our superior data construction. Our method achieves exceptional results on all benchmarks using the same vision encoder compared to previous approaches and models trained on other synthetic datasets. Notably, our fine-tuned CoLLM outperforms MagicLens CoCa-L [64] and CoVR2 [58], despite these models using larger, more advanced vision encoders and being trained on larger datasets. These results underscore the importance of data quality over quantity." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 466, + 556, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 466, + 556, + 647 + ], + "spans": [ + { + "bbox": [ + 313, + 466, + 556, + 647 + ], + "type": "text", + "content": "MTCIR vs. Other CIR Datasets. Despite CoLLM's state-of-the-art performance, two questions remain: (i) Can MTCIR benefit other models? (ii) Can CoLLM achieve good performance when fine-tuned on other CIR datasets? To address these questions, we train a baseline BLIP-L [29] model on various CIR datasets, including MTCIR. Both LaSCo and MTCIR experiments are trained for one epoch. As shown in Table 4, we observe that: (i) BLIP-L shows significant improvement when using MTCIR as the training data, and (ii) CoLLM exhibits notable performance degradation when fine-tuned on the LaSCo dataset. These results demonstrate MTCIR's versatility as a plug-and-play dataset, capable of enhancing performance across various models. This underscores MTCIR's potential to make significant contributions to the CIR community." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 648, + 517, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 648, + 517, + 659 + ], + "spans": [ + { + "bbox": [ + 313, + 648, + 517, + 659 + ], + "type": "text", + "content": "5.3. Re-evaluation on Refined Benchmarks" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "type": "text", + "content": "Ambiguity in the CIRR and Fashion-IQ benchmarks potentially skews the metrics presented in previous tables, particularly for high-performing models. While CIRCO mitigates this issue, it still suffers from domain overlap with train-" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4000" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 74, + 82, + 277, + 137 + ], + "blocks": [ + { + "bbox": [ + 82, + 56, + 268, + 67 + ], + "lines": [ + { + "bbox": [ + 82, + 56, + 268, + 67 + ], + "spans": [ + { + "bbox": [ + 82, + 56, + 268, + 67 + ], + "type": "text", + "content": "Table 6. Ablation studies of proposed components." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 84, + 70, + 266, + 80 + ], + "lines": [ + { + "bbox": [ + 84, + 70, + 266, + 80 + ], + "spans": [ + { + "bbox": [ + 84, + 70, + 266, + 80 + ], + "type": "text", + "content": "(a) Reference image and modification text interpolation." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 74, + 82, + 277, + 137 + ], + "lines": [ + { + "bbox": [ + 74, + 82, + 277, + 137 + ], + "spans": [ + { + "bbox": [ + 74, + 82, + 277, + 137 + ], + "type": "table", + "html": "
ImageTextCIRCO (mAP ↑)CIRR (Rec. ↑)
14.7170.1
✓14.5176.4
✓39.6183.3
✓✓52.8194.5
", + "image_path": "b61da98592cbade6c6b29fa358a6bab67ac8a5c70e157dcfb4369346c347eb69.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 74, + 155, + 277, + 199 + ], + "blocks": [ + { + "bbox": [ + 104, + 142, + 247, + 153 + ], + "lines": [ + { + "bbox": [ + 104, + 142, + 247, + 153 + ], + "spans": [ + { + "bbox": [ + 104, + 142, + 247, + 153 + ], + "type": "text", + "content": "(b) Unimodal queries in Eq. (3) and Eq. (4)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 74, + 155, + 277, + 199 + ], + "lines": [ + { + "bbox": [ + 74, + 155, + 277, + 199 + ], + "spans": [ + { + "bbox": [ + 74, + 155, + 277, + 199 + ], + "type": "table", + "html": "
Image-onlyText-onlyCIRCO (mAP ↑)CIRR (Rec. ↑)
✓14.599.7
✓47.1196.3
✓✓52.8194.5
", + "image_path": "99c752ed9d3bd7ad93bea227978536c61818501ade65d006591627906c2a00e5.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 72, + 217, + 275, + 255 + ], + "blocks": [ + { + "bbox": [ + 98, + 205, + 247, + 216 + ], + "lines": [ + { + "bbox": [ + 98, + 205, + 247, + 216 + ], + "spans": [ + { + "bbox": [ + 98, + 205, + 247, + 216 + ], + "type": "text", + "content": "(c) Reference image embedding interpolation." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 72, + 217, + 275, + 255 + ], + "lines": [ + { + "bbox": [ + 72, + 217, + 275, + 255 + ], + "spans": [ + { + "bbox": [ + 72, + 217, + 275, + 255 + ], + "type": "table", + "html": "
CIRCO (mAP ↑)CIRR (Rec. ↑)
Random In-Batch Sample46.7182.2
Nearest In-Batch Neighbor52.8194.5
", + "image_path": "19fe09da6cc5e5f5999a410fb5a7d41913b453c5e37ebd6401c3f79800c895c8.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 261, + 294, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 261, + 294, + 308 + ], + "spans": [ + { + "bbox": [ + 55, + 261, + 294, + 308 + ], + "type": "text", + "content": "ing images. To assess the impact of ambiguous samples, we re-evaluate the models on our newly refined CIRR and Fashion-IQ benchmarks. Table 5 presents the performance of these models on the revised benchmarks." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 309, + 295, + 465 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 309, + 295, + 465 + ], + "spans": [ + { + "bbox": [ + 55, + 309, + 295, + 465 + ], + "type": "text", + "content": "Our analysis shows that model rankings remain largely consistent before and after the benchmark refinement. As the modification texts in the refined benchmarks provide additional nuance, we anticipated that models with more robust composed query understanding capabilities would distinguish themselves. Indeed, CoLLM consistently outperforms other models across various settings on the refined benchmarks. Notably, while BLIP-L (MTCIR) trailed CoLLM (MTCIR) by " + }, + { + "bbox": [ + 55, + 309, + 295, + 465 + ], + "type": "inline_equation", + "content": "3.4\\%" + }, + { + "bbox": [ + 55, + 309, + 295, + 465 + ], + "type": "text", + "content": " (CIRR@1)(Table 4), this performance gap widened to " + }, + { + "bbox": [ + 55, + 309, + 295, + 465 + ], + "type": "inline_equation", + "content": "6.9\\%" + }, + { + "bbox": [ + 55, + 309, + 295, + 465 + ], + "type": "text", + "content": " (CIRR@1) on the refined CIRR benchmark. This suggests CoLLM better captures nuanced modification text and distinguishes visual differences between source and target images." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 468, + 156, + 479 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 468, + 156, + 479 + ], + "spans": [ + { + "bbox": [ + 55, + 468, + 156, + 479 + ], + "type": "text", + "content": "5.4. Ablation Studies" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 486, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 486, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 486, + 295, + 713 + ], + "type": "text", + "content": "We conduct comprehensive ablation studies to evaluate the impact of our proposed components during the pre-training stage. Our experiments use the BLIP-L encoder, trained exclusively on the LLaVA-558k [34] dataset for one epoch. Table 6 presents the performance across various settings. We report the sum of mAP at " + }, + { + "bbox": [ + 55, + 486, + 295, + 713 + ], + "type": "inline_equation", + "content": "k = \\{5,10,25,50\\}" + }, + { + "bbox": [ + 55, + 486, + 295, + 713 + ], + "type": "text", + "content": " for CIRCO and Recall@ " + }, + { + "bbox": [ + 55, + 486, + 295, + 713 + ], + "type": "inline_equation", + "content": "k = \\{1,5,10,50\\}" + }, + { + "bbox": [ + 55, + 486, + 295, + 713 + ], + "type": "text", + "content": " for CIRR. Image and Text Interpolation are Crucial: As shown in Table 6a, employing Slerp for reference image embedding significantly enhances the model's learning capabilities. Applying interpolation on modification text further improves performance. Unimodal Queries are Beneficial: We investigate the necessity of unimodal queries in Eq. (3) and Eq. (4) and observe that omitting unimodal queries during training, especially text-only queries, substantially degrades performance (Table 6b). Nearest Neighbor is Essential: For reference image embedding interpolation, one might hypothesize that a random in-batch sample could suffice for interpolation. However, our study demonstrates that" + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 315, + 114, + 558, + 184 + ], + "blocks": [ + { + "bbox": [ + 313, + 56, + 555, + 110 + ], + "lines": [ + { + "bbox": [ + 313, + 56, + 555, + 110 + ], + "spans": [ + { + "bbox": [ + 313, + 56, + 555, + 110 + ], + "type": "text", + "content": "Table 7. Performance of different LLMs with CLIP-L/14 as the visual encoder. **Bold and underline highlight the best and second best score.** indicates the original LLM for text generation; others are Large Language Embedding Models (LLEMs) fine-tuned for text retrieval." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 114, + 558, + 184 + ], + "lines": [ + { + "bbox": [ + 315, + 114, + 558, + 184 + ], + "spans": [ + { + "bbox": [ + 315, + 114, + 558, + 184 + ], + "type": "table", + "html": "
LLMCIRCO (mAP↑)CIRR (Rec.↑)FIQ (Rec.↑)
@5@10@50@1@10@50@10@50
Stella-Qwen2-1.5B [11]14.815.317.427.569.188.029.349.0
Mistral-7B* [24]19.820.222.729.672.591.329.549.3
E5-Mistral-7B-Inst [60]19.620.322.829.572.791.129.949.5
SFR-Embedding-2 [47]20.320.823.429.772.891.530.149.5
", + "image_path": "222086076f3ae561fcdc1ba9aede1d3758c83e3b6a5dfdaf70df87d9480f21cc.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 204, + 555, + 336 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 204, + 555, + 336 + ], + "spans": [ + { + "bbox": [ + 313, + 204, + 555, + 336 + ], + "type": "text", + "content": "using the nearest image embedding neighbor yields significantly better performance (Table 6c). LLEMs Outperforms LLMs: We investigate Large Language Embedding Models (LLEMs), which are fine-tuned for text retrieval compared to their base LLMs. As shown in Table 7, both E5-Mistral-7B-Inst [60] and SFR-Embedding-2 outperform their LLM counterpart, Mistral-7B [24]. The superior performance of LLEMs demonstrates that models specifically tailored for embedding and retrieval tasks can offer substantial advantages in CIR applications compared to general-purpose language models." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 357, + 471, + 369 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 357, + 471, + 369 + ], + "spans": [ + { + "bbox": [ + 313, + 357, + 471, + 369 + ], + "type": "text", + "content": "6. Limitations and Conclusion" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 384, + 555, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 384, + 555, + 526 + ], + "spans": [ + { + "bbox": [ + 313, + 384, + 555, + 526 + ], + "type": "text", + "content": "While these advancements significantly improve CIR performance, several areas warrant further investigation. Our work, limited to LLM/LLEMs, could benefit from exploring pre-trained MLLMs for enhanced understanding of CIR tasks. Additionally, our triplet synthesis method generates a single visual token, constraining the model's ability to process detailed image information. Future work should leverage the underutilized text category information in our synthetic datasets to improve model generalization. Lastly, our refined benchmarks, while more reliable, still contain noise from original image pairs, suggesting a need for further refinement of evaluation metrics." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 534, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 556, + 713 + ], + "type": "text", + "content": "In conclusion, we present novel approaches to Composed Image Retrieval (CIR) that obviate the need for annotated datasets. Our contributions include: (1) an innovative triplet synthesis method utilizing image-caption pairs, (2) a new architecture leveraging LLM's embedding generation capabilities, (3) MTCIR, a diverse, human-aligned synthetic dataset, and (4) refined versions of CIRR and Fashion-IQ benchmarks, enhancing the reliability of evaluation metrics in the field. Our method consistently outperforms existing LLM and non-LLM baselines across popular benchmarks. Notably, MTCIR achieves superior results with only " + }, + { + "bbox": [ + 313, + 534, + 556, + 713 + ], + "type": "inline_equation", + "content": "10 - 20\\%" + }, + { + "bbox": [ + 313, + 534, + 556, + 713 + ], + "type": "text", + "content": " of the size of larger datasets, demonstrating a " + }, + { + "bbox": [ + 313, + 534, + 556, + 713 + ], + "type": "inline_equation", + "content": "1 - 15\\%" + }, + { + "bbox": [ + 313, + 534, + 556, + 713 + ], + "type": "text", + "content": " increase in Recall. These advancements collectively push the boundaries of CIR, offering more efficient and effective solutions for real-world applications." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "4001" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "text", + "content": "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 3, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 148, + 294, + 169 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 148, + 294, + 169 + ], + "spans": [ + { + "bbox": [ + 61, + 148, + 294, + 169 + ], + "type": "text", + "content": "[2] Anthropic. The claude 3 model family: Opus, sonnet, haiku, 2024. 3, 5, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 172, + 295, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 172, + 295, + 205 + ], + "spans": [ + { + "bbox": [ + 62, + 172, + 295, + 205 + ], + "type": "text", + "content": "[3] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. A clip-hitchhiker's guide to long video retrieval. arXiv preprint arXiv:2205.08508, 2022. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 207, + 295, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 207, + 295, + 239 + ], + "spans": [ + { + "bbox": [ + 62, + 207, + 295, + 239 + ], + "type": "text", + "content": "[4] Alberto Baldrati, Lorenzo Agnolucci, Marco Bertini, and Alberto Del Bimbo. Zero-shot composed image retrieval with textual inversion. In ICCV, 2023. 2, 3, 5, 6, 7, 13, 15" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 242, + 294, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 242, + 294, + 274 + ], + "spans": [ + { + "bbox": [ + 62, + 242, + 294, + 274 + ], + "type": "text", + "content": "[5] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In CVPR, 2023. 2, 3, 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 277, + 294, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 277, + 294, + 297 + ], + "spans": [ + { + "bbox": [ + 62, + 277, + 294, + 297 + ], + "type": "text", + "content": "[6] Tom B Brown. Language models are few-shot learners. In NeurIPS, 2020. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 300, + 294, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 300, + 294, + 365 + ], + "spans": [ + { + "bbox": [ + 62, + 300, + 294, + 365 + ], + "type": "text", + "content": "[7] Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, et al. Palm: Scaling language modeling with pathways. Journal of Machine Learning Research, 2023. 2, 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 368, + 294, + 411 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 368, + 294, + 411 + ], + "spans": [ + { + "bbox": [ + 62, + 368, + 294, + 411 + ], + "type": "text", + "content": "[8] Niv Cohen, Rinon Gal, Eli A. Meirom, Gal Chechik, and Yuval Atzmon. \"this is my unicorn, fluffy\": Personalizing frozen vision-language representations. In ECCV, 2022. 2, 3, 6, 15" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 415, + 294, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 415, + 294, + 458 + ], + "spans": [ + { + "bbox": [ + 62, + 415, + 294, + 458 + ], + "type": "text", + "content": "[9] Yashar Deldjoo, Fatemeh Nazary, Arnau Ramisa, Julian Mcauley, Giovanni Pellegrini, Alejandro Bellogin, and Tommaso Di Noia. A review of modern fashion recommender systems. ACM Computing Surveys, 2023. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 460, + 294, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 460, + 294, + 492 + ], + "spans": [ + { + "bbox": [ + 57, + 460, + 294, + 492 + ], + "type": "text", + "content": "[10] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009. 13" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 495, + 243, + 506 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 495, + 243, + 506 + ], + "spans": [ + { + "bbox": [ + 57, + 495, + 243, + 506 + ], + "type": "text", + "content": "[11] Zhang Dun. Stella en qwen2 1.5b v5, 2024.8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 508, + 294, + 551 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 508, + 294, + 551 + ], + "spans": [ + { + "bbox": [ + 57, + 508, + 294, + 551 + ], + "type": "text", + "content": "[12] Sonam Goenka, Zhaoheng Zheng, Ayush Jaiswal, Rakesh Chada, Yue Wu, Varsha Hedau, and Pradeep Natarajan. Fashionvlp: Vision language transformer for fashion retrieval with feedback. In CVPR, 2022. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 554, + 294, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 554, + 294, + 597 + ], + "spans": [ + { + "bbox": [ + 57, + 554, + 294, + 597 + ], + "type": "text", + "content": "[13] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Batra, and Devi Parikh. Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In CVPR, 2017. 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 600, + 294, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 600, + 294, + 643 + ], + "spans": [ + { + "bbox": [ + 57, + 600, + 294, + 643 + ], + "type": "text", + "content": "[14] Geonmo Gu, Sanghyuk Chun, Wonjae Kim, HeeJae Jun, Yoohoon Kang, and Sangdoo Yun. Compodiff: Versatile composed image retrieval with latent diffusion. Transactions on Machine Learning Research, 2024. 2, 3, 5, 6, 16" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 646, + 294, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 646, + 294, + 678 + ], + "spans": [ + { + "bbox": [ + 57, + 646, + 294, + 678 + ], + "type": "text", + "content": "[15] Geonmo Gu, Sanghyuk Chun, Wonjae Kim, Yoohoon Kang, and Sangdoo Yun. Language-only training of zero-shot composed image retrieval. In CVPR, 2024. 2, 3, 6, 15" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 57, + 681, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 681, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 681, + 294, + 712 + ], + "type": "text", + "content": "[16] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 13" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "text", + "content": "[17] Matthew Honnibal, Ines Montani, Sofie Van Landeghem, Adriane Boyd, et al. spacy: Industrial-strength natural language processing in python. 2020. 13" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 108, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 108, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 317, + 108, + 553, + 150 + ], + "type": "text", + "content": "[18] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In ICLR, 2022. 15" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 152, + 553, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 152, + 553, + 195 + ], + "spans": [ + { + "bbox": [ + 317, + 152, + 553, + 195 + ], + "type": "text", + "content": "[19] Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Weli-hinda, Alan Hayes, Alec Radford, et al. Gpt-4o system card. arXiv preprint arXiv:2410.21276, 2024. 13" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 198, + 553, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 198, + 553, + 240 + ], + "spans": [ + { + "bbox": [ + 316, + 198, + 553, + 240 + ], + "type": "text", + "content": "[20] Chuong Huynh, Yuqian Zhou, Zhe Lin, Connelly Barnes, Eli Shechtman, Sohrab Amirghodsi, and Abhinav Shrivastava. Simpson: Simplifying photo cleanup with single-click distracting object segmentation network. In CVPR, 2023. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 243, + 553, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 243, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 316, + 243, + 553, + 285 + ], + "type": "text", + "content": "[21] Young Kyun Jang, Dat Huynh, Ashish Shah, Wen-Kai Chen, and Ser-Nam Lim. Spherical linear interpolation and text-anchoring for zero-shot composed image retrieval. In ECCV, 2024. 2, 3, 6, 7, 14, 15" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 287, + 553, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 287, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 316, + 287, + 553, + 319 + ], + "type": "text", + "content": "[22] Young Kyun Jang, Junmo Kang, Yong Jae Lee, and Donghyun Kim. Mate: Meet at the embedding-connecting images with long texts. EMNLP Findings, 2024. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 322, + 553, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 322, + 553, + 364 + ], + "spans": [ + { + "bbox": [ + 316, + 322, + 553, + 364 + ], + "type": "text", + "content": "[23] Young Kyun Jang, Donghyun Kim, Zihang Meng, Dat Huynh, and Ser-Nam Lim. Visual delta generator with large multi-modal models for semi-supervised composed image retrieval. In CVPR, 2024. 2, 3, 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 366, + 553, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 366, + 553, + 419 + ], + "spans": [ + { + "bbox": [ + 316, + 366, + 553, + 419 + ], + "type": "text", + "content": "[24] Albert Q Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, et al. Mistral 7b. arXiv preprint arXiv:2310.06825, 2023. 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 422, + 553, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 422, + 553, + 475 + ], + "spans": [ + { + "bbox": [ + 316, + 422, + 553, + 475 + ], + "type": "text", + "content": "[25] Ting Jiang, Minghui Song, Zihan Zhang, Haizhen Huang, Weiwei Deng, Feng Sun, Qi Zhang, Deqing Wang, and Fuzhen Zhuang. E5-v: Universal embeddings with multimodal large language models. arXiv preprint arXiv:2407.12580, 2024.3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 478, + 553, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 478, + 553, + 510 + ], + "spans": [ + { + "bbox": [ + 316, + 478, + 553, + 510 + ], + "type": "text", + "content": "[26] Shyamgopal Karthik, Karsten Roth, Massimiliano Mancini, and Zeynep Akata. Vision-by-language for training-free compositional image retrieval. *ICLR*, 2024. 3, 6, 7, 15" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 512, + 553, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 512, + 553, + 564 + ], + "spans": [ + { + "bbox": [ + 316, + 512, + 553, + 564 + ], + "type": "text", + "content": "[27] Chankyu Lee, Rajarshi Roy, Mengyao Xu, Jonathan Raiman, Mohammad Shoeybi, Bryan Catanzaro, and Wei Ping. Nv-embed: Improved techniques for training llms as generalist embedding models. arXiv preprint arXiv:2405.17428, 2024.3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 568, + 553, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 568, + 553, + 600 + ], + "spans": [ + { + "bbox": [ + 316, + 568, + 553, + 600 + ], + "type": "text", + "content": "[28] Matan Levy, Rami Ben-Ari, Nir Darshan, and Dani Lischinski. Data roaming and quality assessment for composed image retrieval. AAAI, 2024. 2, 3, 5, 6, 7, 14, 15, 17, 18" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 602, + 553, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 602, + 553, + 644 + ], + "spans": [ + { + "bbox": [ + 316, + 602, + 553, + 644 + ], + "type": "text", + "content": "[29] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, 2022. 2, 7, 14, 17, 18" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "type": "text", + "content": "[30] Wei Li, Hehe Fan, Yongkang Wong, Yi Yang, and Mohan Kankanhalli. Improving context understanding in multimodal large language models via multimodal composition learning. In ICML, 2024. 3, 6, 7" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "text", + "content": "[31] Zehan Li, Xin Zhang, Yanzhao Zhang, Dingkun Long, Pengjun Xie, and Meishan Zhang. Towards general text" + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4002" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 76, + 73, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 73, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 73, + 294, + 95 + ], + "type": "text", + "content": "embeddings with multi-stage contrastive learning. arXiv preprint arXiv:2308.03281, 2023. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 95, + 295, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 95, + 295, + 139 + ], + "spans": [ + { + "bbox": [ + 56, + 95, + 295, + 139 + ], + "type": "text", + "content": "[32] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 5, 7, 15" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 140, + 294, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 140, + 294, + 183 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 294, + 183 + ], + "type": "text", + "content": "[33] Aixin Liu, Bei Feng, Bing Xue, Bingxuan Wang, Bochao Wu, Chengda Lu, Chenggang Zhao, Chengqi Deng, Chenyu Zhang, Chong Ruan, et al. Deepseek-v3 technical report. arXiv preprint arXiv:2412.19437, 2024. 13" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 184, + 294, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 184, + 294, + 215 + ], + "spans": [ + { + "bbox": [ + 56, + 184, + 294, + 215 + ], + "type": "text", + "content": "[34] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, 2023. 2, 3, 5, 6, 8, 14, 15" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 217, + 294, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 217, + 294, + 259 + ], + "spans": [ + { + "bbox": [ + 56, + 217, + 294, + 259 + ], + "type": "text", + "content": "[35] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava-last: Improved reasoning,OCR, and world knowledge, 2024. 2, 3, 5, 13" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 261, + 294, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 261, + 294, + 293 + ], + "spans": [ + { + "bbox": [ + 56, + 261, + 294, + 293 + ], + "type": "text", + "content": "[36] Yikun Liu, Jiangchao Yao, Ya Zhang, Yan-Feng Wang, and Weidi Xie. Zero-shot composed text-image retrieval. In BMVC, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 294, + 294, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 294, + 294, + 337 + ], + "spans": [ + { + "bbox": [ + 56, + 294, + 294, + 337 + ], + "type": "text", + "content": "[37] Zheyuan Liu, Cristian Rodriguez-Opazo, Damien Teney, and Stephen Gould. Image retrieval on real-life images with pretrained vision-and-language models. In ICCV, 2021. 2, 3, 5, 7, 13, 14, 17" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 338, + 294, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 338, + 294, + 392 + ], + "spans": [ + { + "bbox": [ + 56, + 338, + 294, + 392 + ], + "type": "text", + "content": "[38] Feipeng Ma, Hongwei Xue, Guangting Wang, Yizhou Zhou, Fengyun Rao, Shilin Yan, Yueyi Zhang, Siying Wu, Mike Zheng Shou, and Xiaoyan Sun. Multi-modal generative embedding model. arXiv preprint arXiv:2405.19333, 2024. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 393, + 294, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 393, + 294, + 446 + ], + "spans": [ + { + "bbox": [ + 56, + 393, + 294, + 446 + ], + "type": "text", + "content": "[39] Gabriel de Souza P Moreira, Radek Osmulski, Mengyao Xu, Ronay Ak, Benedikt Schifferer, and Even Oldridge. Nv-retriever: Improving text embedding models with effective hard-negative mining. arXiv preprint arXiv:2407.15831, 2024. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 449, + 294, + 481 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 449, + 294, + 481 + ], + "spans": [ + { + "bbox": [ + 56, + 449, + 294, + 481 + ], + "type": "text", + "content": "[40] Niklas Muennighoff, Nouamane Tazi, Loic Magne, and Nils Reimers. Mteb: Massive text embedding benchmark. arXiv preprint arXiv:2210.07316, 2022. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 482, + 294, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 482, + 294, + 514 + ], + "spans": [ + { + "bbox": [ + 56, + 482, + 294, + 514 + ], + "type": "text", + "content": "[41] Vicente Ordonez, Girish Kulkarni, and Tamara Berg. Im2text: Describing images using 1 million captioned photographs. NeurIPS, 2011. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 515, + 294, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 515, + 294, + 548 + ], + "spans": [ + { + "bbox": [ + 56, + 515, + 294, + 548 + ], + "type": "text", + "content": "[42] Khoi Pham, Chuong Huynh, Ser-Nam Lim, and Abhinav Shrivastava. Composing object relations and attributes for image-text matching. In CVPR, 2024. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 548, + 294, + 603 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 548, + 294, + 603 + ], + "spans": [ + { + "bbox": [ + 56, + 548, + 294, + 603 + ], + "type": "text", + "content": "[43] Bill Psomas, Ioannis Kakogeorgiou, Nikos Efthymiadis, Giorgos Tolias, Ondrej Chum, Yannis Avrithis, and Konstantinos Karantzalos. Composed image retrieval for remote sensing. In IGARSS 2024-2024 IEEE International Geoscience and Remote Sensing Symposium, 2024. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 604, + 294, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 604, + 294, + 657 + ], + "spans": [ + { + "bbox": [ + 56, + 604, + 294, + 657 + ], + "type": "text", + "content": "[44] Filip Radenovic, Abhimanyu Dubey, Abhishek Kadian, Todor Mihaylov, Simon Vandenhende, Yash Patel, Yi Wen, Vignesh Ramanathan, and Dhruv Mahajan. Filtering, distillation, and hard negatives for vision-language pre-training. In CVPR, 2023. 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "type": "text", + "content": "[45] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 2, 5, 13, 14" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "text", + "content": "[46] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 2, 3, 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 107, + 553, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 148 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 148 + ], + "type": "text", + "content": "[47] Meng Rui, Liu Ye, Joty Shafiq Rayhan, Xiong Caiming, Zhou Yingbo, and Yavuz Semih. Sfr-embedding-2: Advanced text embedding with multi-stage training, 2024. 8, 14" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 150, + 553, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 150, + 553, + 183 + ], + "spans": [ + { + "bbox": [ + 316, + 150, + 553, + 183 + ], + "type": "text", + "content": "[48] Amrita Saha, Mitesh Khapra, and Karthik Sankaranarayanan. Towards building large scale multimodal domain-aware conversation systems. In AAAI, 2018. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 184, + 553, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 184, + 553, + 227 + ], + "spans": [ + { + "bbox": [ + 316, + 184, + 553, + 227 + ], + "type": "text", + "content": "[49] Kuniaki Saito, Kihyuk Sohn, Xiang Zhang, Chun-Liang Li, Chen-Yu Lee, Kate Saenko, and Tomas Pfister. Pic2word: Mapping pictures to words for zero-shot composed image retrieval. CVPR, 2023. 3, 6, 15" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 228, + 553, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 228, + 553, + 282 + ], + "spans": [ + { + "bbox": [ + 316, + 228, + 553, + 282 + ], + "type": "text", + "content": "[50] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021. 5, 6, 14" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 283, + 553, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 283, + 553, + 326 + ], + "spans": [ + { + "bbox": [ + 316, + 283, + 553, + 326 + ], + "type": "text", + "content": "[51] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In ACL, 2018. 5, 6, 14" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 327, + 553, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 327, + 553, + 360 + ], + "spans": [ + { + "bbox": [ + 316, + 327, + 553, + 360 + ], + "type": "text", + "content": "[52] Ken Shoemake. Animating rotation with quaternion curves. In Proceedings of the 12th annual conference on Computer graphics and interactive techniques, 1985. 2, 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 361, + 553, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 361, + 553, + 415 + ], + "spans": [ + { + "bbox": [ + 316, + 361, + 553, + 415 + ], + "type": "text", + "content": "[53] Swetha Sirnam, Jinyu Yang, Tal Neiman, Mamshad Nayeem Rizve, Son Tran, Benjamin Yao, Trishul Chilimbi, and Mubarak Shah. X-former: Unifying contrastive and reconstruction learning for mllms. In European Conference on Computer Vision, pages 146-162. Springer, 2024. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 415, + 553, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 415, + 553, + 449 + ], + "spans": [ + { + "bbox": [ + 316, + 415, + 553, + 449 + ], + "type": "text", + "content": "[54] Alane Suhr, Stephanie Zhou, Ally Zhang, Iris Zhang, Huajun Bai, and Yoav Artzi. A corpus for reasoning about natural language grounded in photographs. In ACL, 2019. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 449, + 553, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 449, + 553, + 491 + ], + "spans": [ + { + "bbox": [ + 316, + 449, + 553, + 491 + ], + "type": "text", + "content": "[55] Yuanmin Tang, Jing Yu, Keke Gai, Jiamin Zhuang, Gang Xiong, Yue Hu, and Qi Wu. Context-i2w: Mapping images to context-dependent words for accurate zero-shot composed image retrieval. In AAAI, 2024. 2, 3, 6, 15" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 492, + 553, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 492, + 553, + 546 + ], + "spans": [ + { + "bbox": [ + 316, + 492, + 553, + 546 + ], + "type": "text", + "content": "[56] Omkar Thawakar, Muzammal Naseer, Rao Muhammad Anwer, Salman Khan, Michael Felsberg, Mubarak Shah, and Fahad Shahbaz Khan. Composed video retrieval via enriched context and discriminative embeddings. CVPR, 2024. 4, 6, 16" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 548, + 553, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 548, + 553, + 602 + ], + "spans": [ + { + "bbox": [ + 316, + 548, + 553, + 602 + ], + "type": "text", + "content": "[57] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 603, + 553, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 603, + 553, + 635 + ], + "spans": [ + { + "bbox": [ + 316, + 603, + 553, + 635 + ], + "type": "text", + "content": "[58] Lucas Ventura, Antoine Yang, Cordelia Schmid, and Gül Varol. CoVR-2: Automatic data construction for composed video retrieval. IEEE TPAMI, 2024. 2, 3, 5, 6, 7, 14, 16, 18" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 635, + 553, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 553, + 669 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 553, + 669 + ], + "type": "text", + "content": "[59] Lucas Ventura, Antoine Yang, Cordelia Schmid, and Güll Varol. CoVR: Learning composed video retrieval from web video captions. AAAI, 2024. 3, 4, 5, 6, 7, 15, 16, 17" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "text", + "content": "[60] Liang Wang, Nan Yang, Xiaolong Huang, Linjun Yang, Rangan Majumder, and Furu Wei. Improving text embeddings with large language models. arXiv preprint arXiv:2401.00368, 2023. 3, 8" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4003" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 527 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 126 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 126 + ], + "type": "text", + "content": "[61] Hui Wu, Yupeng Gao, Xiaoxiao Guo, Ziad Al-Halah, Steven Rennie, Kristen Grauman, and Rogerio Feris. The fashion iq dataset: Retrieving images by combining side information and relative natural language feedback. CVPR, 2021. 2, 5, 7, 17" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 294, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 294, + 194 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 294, + 194 + ], + "type": "text", + "content": "[62] Jinyu Yang, Jiali Duan, Son Tran, Yi Xu, Sampath Chanda, Liquun Chen, Belinda Zeng, Trishul Chilimbi, and Junzhou Huang. Vision-language pre-training with triple contrastive learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15671-15680, 2022. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 196, + 294, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 196, + 294, + 239 + ], + "spans": [ + { + "bbox": [ + 56, + 196, + 294, + 239 + ], + "type": "text", + "content": "[63] ZHOU Yuqian, Chuong Huynh, Connelly Barnes, Elya Shechtman, Sohrab Amirghodsi, and Zhe Lin. Machine-learning models for distractor segmentation with reduced user interactions, 2024. US Patent App. 18/307,353. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 241, + 294, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 241, + 294, + 283 + ], + "spans": [ + { + "bbox": [ + 56, + 241, + 294, + 283 + ], + "type": "text", + "content": "[64] Kai Zhang, Yi Luan, Hexiang Hu, Kenton Lee, Siyuan Qiao, Wenhu Chen, Yu Su, and Ming-Wei Chang. Magiclens: Self-supervised image retrieval with open-ended instructions. In ICML, 2024. 2, 3, 5, 6, 7, 16, 18" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 285, + 294, + 339 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 285, + 294, + 339 + ], + "spans": [ + { + "bbox": [ + 56, + 285, + 294, + 339 + ], + "type": "text", + "content": "[65] Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen, Christopher Dewan, Mona Diab, Xian Li, Xi Victoria Lin, et al. Opt: Open pre-trained transformer language models. arXiv preprint arXiv:2205.01068, 2022. 2, 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 342, + 294, + 406 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 342, + 294, + 406 + ], + "spans": [ + { + "bbox": [ + 56, + 342, + 294, + 406 + ], + "type": "text", + "content": "[66] Youcai Zhang, Xinyu Huang, Jinyu Ma, Zhaoyang Li, Zhaochuan Luo, Yanchun Xie, Yuzhuo Qin, Tong Luo, Yaqian Li, Shilong Liu, et al. Recognize anything: A strong image tagging model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1724-1732, 2024. 13" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 408, + 294, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 408, + 294, + 474 + ], + "spans": [ + { + "bbox": [ + 56, + 408, + 294, + 474 + ], + "type": "text", + "content": "[67] Xinliang Zhu, Sheng-Wei Huang, Han Ding, Jinyu Yang, Kelvin Chen, Tao Zhou, Tal Neiman, Ouye Xie, Son Tran, Benjamin Yao, et al. Bringing multimodality to amazon visual search system. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 6390–6399, 2024. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 475, + 294, + 527 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 475, + 294, + 527 + ], + "spans": [ + { + "bbox": [ + 56, + 475, + 294, + 527 + ], + "type": "text", + "content": "[68] Yutao Zhu, Huaying Yuan, Shuting Wang, Jiongnan Liu, Wenhan Liu, Chenlong Deng, Haonan Chen, Zhicheng Dou, and Ji-Rong Wen. Large language models for information retrieval: A survey. arXiv preprint arXiv:2308.07107, 2023. 3" + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4004" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/CoMBO_ Conflict Mitigation via Branched Optimization for Class Incremental Segmentation/a4944725-d45e-4b4f-a404-3fd5f2ee3b6a_content_list.json b/2025/CoMBO_ Conflict Mitigation via Branched Optimization for Class Incremental Segmentation/a4944725-d45e-4b4f-a404-3fd5f2ee3b6a_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..c065ed02fe21c9a80c0628471eca66fbe9dca8ec --- /dev/null +++ b/2025/CoMBO_ Conflict Mitigation via Branched Optimization for Class Incremental Segmentation/a4944725-d45e-4b4f-a404-3fd5f2ee3b6a_content_list.json @@ -0,0 +1,1484 @@ +[ + { + "type": "text", + "text": "CoMBO: Conflict Mitigation via Branched Optimization for Class Incremental Segmentation", + "text_level": 1, + "bbox": [ + 196, + 130, + 802, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Kai Fang $^{1*}$ , Anqi Zhang $^{1*}$ , Guangyu Gao $^{1\\dagger}$ , Jianbo Jiao $^{2}$ , Chi Harold Liu $^{1}$ , Yunchao Wei $^{3}$ $^{1}$ Beijing Institute of Technology $^{2}$ University of Birmingham $^{3}$ Beijing Jiaotong University", + "bbox": [ + 109, + 202, + 880, + 241 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 273, + 326, + 289 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Effective Class Incremental Segmentation (CIS) requires simultaneously mitigating catastrophic forgetting and ensuring sufficient plasticity to integrate new classes. The inherent conflict above often leads to a back-and-forth, which turns the objective into finding the balance between the performance of previous (old) and incremental (new) classes. To address this conflict, we introduce a novel approach, Conflict Mitigation via Branched Optimization (CoMBO). Within this approach, we present the Query Conflict Reduction module, designed to explicitly refine queries for new classes through lightweight, class-specific adapters. This module provides an additional branch for the acquisition of new classes while preserving the original queries for distillation. Moreover, we develop two strategies to further mitigate the conflict following the branched structure, i.e., the Half-Learning Half-Distillation (HDHL) over classification probabilities, and the Importance-Based Knowledge Distillation (IKD) over query features. HDHL selectively engages in learning for classification probabilities of queries that match the ground truth of new classes, while aligning unmatched ones to the corresponding old probabilities, thus ensuring retention of old knowledge while absorbing new classes via learning negative samples. Meanwhile, IKD assesses the importance of queries based on their matching degree to old classes, prioritizing the distillation of important features and allowing less critical features to evolve. Extensive experiments in Class Incremental Panoptic and Semantic Segmentation settings have demonstrated the superior performance of CoMBO. Project page: https://guangyu-ryan.github.io/CoMBO.", + "bbox": [ + 89, + 305, + 485, + 760 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 790, + 220, + 806 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Semantic segmentation, the fundamental task in computer vision, involves classifying each pixel into predefined categories. Panoptic segmentation, the challeng-", + "bbox": [ + 89, + 816, + 483, + 863 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5f9cb9e87dd8c4c509c8e46890c3305acdd634a8bcadfb488ed943b6744c465a.jpg", + "image_caption": [ + "Figure 1. Comparison of our improved distillation and learning strategy (bottom) with previous conflicting strategy (top). Previous strategies impose contradictory supervision on the same target to find a balance, whereas our strategy, including the importance factor for distillation and the QCR module for adaptive learning, enables more compatible, target-specific supervision. The visualization on incremental classes ('Traffic Light') highlights the effectiveness of our strategy." + ], + "image_footnote": [], + "bbox": [ + 519, + 271, + 908, + 623 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ing variant, unifies semantic and instance segmentation to both classify every pixel and simultaneously identify distinct object instances. Recent advancements in mask classification-based segmentation, such as MaskFormer [16] and Mask2Former [17], enable the unification of semantic and panoptic segmentation. However, traditional approaches to segmentation suffer from limitations in dynamic environments where new classes can emerge unpredictably, as they are typically trained on a static set", + "bbox": [ + 511, + 763, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution.", + "bbox": [ + 107, + 875, + 218, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author, guangyugao@bit.edu.cn.", + "bbox": [ + 109, + 887, + 369, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "25667", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "of categories. Therefore, Class Incremental Semantic Segmentation (CISS) has emerged as a pivotal task for continually learning new categories, where models can adapt to new classes without forgetting previously learned ones.", + "bbox": [ + 89, + 90, + 480, + 151 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Historically, studies in CISS [3, 35, 49] mainly focuses on tackling catastrophic forgetting, employing various techniques such as knowledge distillation [5, 20, 47], freezing parameters [9, 26, 48], or generating pseudo-labels from old model [20]. More recent efforts [8, 11] have extended these techniques to the more challenging Class Incremental Panoptic Segmentation (CIPS) based on the advanced Mask2Former. Nonetheless, these methods often fail to adequately reconcile the inherent tension between retaining old knowledge and acquiring new information. Some overemphasize maintaining stability, thus hindering the learning of new classes, while others prioritize plasticity, leading to significant forgetting of old classes. It is necessary to deconstruct these conflicting objectives to facilitate a more compatible incremental learning process, achieving promotion in both the acquisition of new classes and retention of old classes.", + "bbox": [ + 89, + 154, + 482, + 411 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we propose Conflict Mitigation via Branched Optimization (CoMBO), an approach to reduce conflicts between acquiring new classes and retaining old classes. By integrating three specialized knowledge distillation mechanisms within the Mask2Former architecture, CoMBO effectively mitigates catastrophic forgetting while enhancing the assimilation of new classes. Using bipartite matching, Mask2Former optimally aligns each query, complete with its classification logits and mask predictions, with the most fitting ground truth class, ensuring efficient optimization. Therefore, we first introduce the Half-Learning Half-Distillation mechanism, which selectively applies a shared Kullback-Leibler divergence to the classification logits of unmatched queries. This method distinctly separates distillation on unmatched masks from the classification optimization of matched ones. Moreover, rather than applying uniform distillation across all features, we introduce an importance-based knowledge distillation technique. This technique assesses each query's significance in preserving knowledge of old classes and adjusts the distillation intensity accordingly, emphasizing the retention of key queries. Additionally, our Query Conflict Reduction module innovatively adapts features for new classes while maintaining the queries' original characteristics before distillation, using class-specific adapters for each newly learned class. Finally, our approach outperforms previous methods in multiple benchmarks, especially in typical and challenging scenarios such as 100-10 of ADE20K. Our approach achieves remarkable performance of $35.6\\%$ and $41.1\\%$ mIoU on 100-10 of CIPS and CISS, respectively, compared to the previous state-of-the-art.", + "bbox": [ + 89, + 414, + 482, + 883 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Overall, our contributions are summarized as follows:", + "bbox": [ + 109, + 885, + 465, + 898 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a Half-Learning Half-Distillation mechanism that applies soft distillation for old classes and binary-based optimization for new classes to effectively balance retention and acquisition.", + "- An extra Query Conflict Reduction module is proposed to refine queries for new classes while preserving unrefined features, which enables Importance-based Knowledge Distillation for remembering the key features.", + "- Through extensive quantitative and qualitative evaluations on the ADE20K dataset, we demonstrate the state-of-the-art performance of our model in both Class Incremental Semantic and Panoptic Segmentation tasks." + ], + "bbox": [ + 513, + 90, + 903, + 272 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 287, + 653, + 301 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Semantic and Panoptic Segmentation", + "text_level": 1, + "bbox": [ + 513, + 313, + 834, + 329 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Semantic segmentation aims to classify each pixel in an image, which focuses on the complete coverage of object(s) in each category. The pioneering work FCN [30] and U-Net [36] introduces a pixel-wise classification paradigm by removing the fully connected layers. Since then, numerous approaches [2, 12-15, 29, 38, 41] have been proposed following the paradigm, containing effective modules such as ASPP [13] and pyramid structure [50]. In contrast, panoptic segmentation [27] combines semantic and instance segmentation, requiring the classification of every pixel while distinguishing instances of the same category. These two tasks were addressed separately due to the limitation of the pixelwise classification. However, the emergence of the mask classification paradigm [16, 17, 43, 45] introduced a universal, transformer-based framework capable of solving multiple segmentation tasks simultaneously. MaskFormer [16] pioneered this paradigm by combining class-agnostic mask proposal generation and mask classification, enabling both semantic and panoptic segmentation simultaneously. Building upon this, Mask2Former [17] enhanced performance by incorporating multi-scale feature fusion and masked attention techniques. However, when learning new categories beyond the initial data, mask classification methods remain susceptible to catastrophic forgetting.", + "bbox": [ + 511, + 337, + 906, + 700 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Continual Segmentation", + "text_level": 1, + "bbox": [ + 513, + 710, + 736, + 726 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In Class Incremental Learning (CIL), fine-tuning models on new data often results in a performance drop on old categories, i.e., catastrophic forgetting [22, 39]. To address this, various methods [5-7, 20, 21, 32, 33, 51] have been proposed, including retaining representative old samples [1, 4, 9, 10, 31, 32, 35, 53], compensation losses [18, 34, 42, 44], and prompt representations [25, 37, 40]. These advancements in CIL raise interest in more challenging segmentation tasks. Class Incremental Semantic Segmentation (CISS) was first introduced in MiB [5], reconstructing background regions to address background shifting during", + "bbox": [ + 511, + 734, + 906, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "25668", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "incremental steps. Since then, most methods [3, 47] have employed techniques such as knowledge distillation [28], pseudo-labels from previous models [8, 20], and background redefinition [5, 46] to mitigate background shifting. However, these methods primarily rely on pixel-wise classification models like DeepLabV3 [13]. Therefore, CoM-Former [8] first utilizes the mask classification segmentation model Mask2Former [17] enabling solving both CISS and Class Incremental Panoptic Segmentation (CIPS) tasks. ECLIPSE [26] introduces dynamic model structure expansion methods to extend learnable parameters for new classes and freezes the old parameters. CoMasTRe [23] distills queries matched to old queries with high probability on old classes. BalConpas [11] balance the class proportion on selecting representative replay samples. However, these methods often struggle to balance acquisition and preservation. Our approach mitigates this conflict by separating queries to decouple the two objectives.", + "bbox": [ + 89, + 90, + 483, + 363 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Preliminaries", + "text_level": 1, + "bbox": [ + 89, + 377, + 230, + 393 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Problem Definition", + "text_level": 1, + "bbox": [ + 89, + 402, + 272, + 417 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In continual segmentation, the model is tasked with an incremental learning challenge over $T$ steps. At each step $t \\in \\{1, \\ldots, T\\}$ , the model is trained to recognize a unique set of classes $C^t$ , where the intersection across all sets from $i$ to $T$ is empty $\\bigcap_{t=1}^{T} C^t = \\emptyset$ and their union forms the complete set of classes $\\bigcup_{t=1}^{T} C^t = \\mathcal{C}$ . In the current step $t$ , the training dataset $D_{train}^t$ consists of image-label pairs $(x^t, y^t)$ , where $x^t$ represents an image and $y^t$ its corresponding segmentation label. The labels $y^t$ are available only for the classes $C^t$ that need to be learned at this step, while labels for previously learned classes $C^{1:t-1}$ and future classes $C^{t+1:T}$ are inaccessible. After completing training at this step, the model must perform segmentation across all classes learned up to that point, $C^{1:t}$ , thereby preventing catastrophic forgetting of the old classes $C^{1:t-1}$ while effectively acquiring new ones $C^t$ .", + "bbox": [ + 89, + 426, + 483, + 672 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Revisiting Mask2Former and Pseudo-Labeling", + "text_level": 1, + "bbox": [ + 89, + 680, + 483, + 698 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The recent advanced universal segmentation network, Mask2Former [17], has favored the class incremental semantic and panoptic segmentation (CISS and CIPS) tasks for its innovative proposal-based structure. Mask2Former deviates from traditional pixel-wise classification pipelines by introducing a class-agnostic mask prediction and classification mechanism. Specifically, the backbone $\\pmb{f}_b$ and the pixel decoder $\\pmb{f}_p$ generate multi-scale features that interact with $N$ learnable queries through multiple self-attention and cross-attention layers in the transformer decoder $\\pmb{f}_t$ . These $N$ output queries $Q \\in \\mathbb{R}^{C_Q \\times N}$ undergo further processing through two linear layers to form mask embeddings $\\mathcal{E}_{mask} \\in \\mathbb{R}^{D \\times N}$ and class embeddings $\\mathcal{E}_{cls} \\in \\mathbb{R}^{(C + 1) \\times N}$ .", + "bbox": [ + 89, + 704, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The mask embeddings $\\mathcal{E}_{mask}$ yield $N$ mask predictions $M\\in \\mathbb{R}^{N\\times H\\times W}$ via dot product between each mask embedding and per-pixel embeddings $\\mathcal{E}_{pixel}\\in \\mathbb{R}^{D\\times H\\times W}$ from $f_{p}$ , while $\\mathcal{E}_{cls}$ accounts for the classification predictions of these masks. Note that the additional channel in $\\mathcal{E}_{cls}$ represents the no-object class, an auxiliary category utilized during training but excluded during inference.", + "bbox": [ + 511, + 90, + 903, + 196 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the contexts of CISS and CIPS, where ground truth for old classes $C^{1:t-1}$ are inaccessible, pseudo-labeling becomes crucial. PLOP [20] introduces this strategy by generating pixel-level pseudo ground truths for $C^{1:t-1}$ , adapted to the mask-based structure by CoMFormer [8], the first method based on Mask2Former for CISS and CIPS tasks. This adaptation involves weighting the mask predictions $M$ by the corresponding maximum class embedding score after the softmax process over the $C+1$ dimensions, i.e., $\\max_{c=0}^{|C^{0:t-1}|} (\\operatorname{softmax}_{c=0}^{|C^{0:t-1}|} (\\mathcal{E}_{cls}))$ to form $M_{\\omega}$ , with associated categories $C_{\\omega} = \\arg \\max_{c=0}^{|C^{0:t-1}|} \\mathcal{E}_{cls} \\in \\mathbb{R}^{N}$ . The pseudo-labeling process then labels the pixels outside the union of $\\boldsymbol{y}^t$ with the category $c$ that achieves the maximum score across $M_{\\omega}$ and $C_{w}$ :", + "bbox": [ + 511, + 198, + 906, + 416 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {M} _ {c} (h, w) = \\left\\{ \\begin{array}{l l} 1 & \\text {i f} c = C _ {\\omega} [ \\arg \\max _ {n = 1} ^ {N} M _ {w} (n, h, w) ] \\\\ & \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\text {v a r e p s i l o n} \\\\ 0 & \\text {e l s e} \\end{array} \\right.\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 422, + 905, + 500 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $n$ indexes the $N$ proposals, and $h, w$ are the pixel coordinates. This pseudo-labeling strategy has become foundational for CISS and CIPS methods [8].", + "bbox": [ + 511, + 500, + 905, + 544 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Proposed Method", + "text_level": 1, + "bbox": [ + 511, + 556, + 687, + 573 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In our proposed Conflict Mitigation via Branched Optimization (CoMBO), we introduce three distinct strategies: Half-Distillation-Half-Learning Strategy for targeted knowledge retention, Importance-Based Knowledge Distillation to prioritize crucial features, and Query Conflict Reduction for efficient class integration.", + "bbox": [ + 511, + 582, + 905, + 674 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1. Query Conflict Reduction", + "text_level": 1, + "bbox": [ + 511, + 681, + 746, + 696 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As mentioned in Sec. 3.2, both class embeddings $\\mathcal{E}_{cls}$ and mask predictions $M$ are generated from the queries $Q$ . This interface becomes a critical juncture of conflicts between retaining knowledge of old classes and acquiring new classes manifest, which cannot break through the bottleneck of overall performance by simply finding the balance point.", + "bbox": [ + 511, + 703, + 905, + 796 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To deconstruct the crossroad efficiently, we design a Query Conflict Reduction (QCR) module $f_{QCR}(\\cdot)$ as an additional adapter for new classes. As shown in Fig. 2, in most scenarios, queries from the previous layer $Q_{l-1}$ have the same classification results as the queries from the current layer due to the similar distribution and the same clas", + "bbox": [ + 511, + 810, + 905, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "25669", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/127ba02410a6d6f9244f3905c6ea692fdacf93ac171a7a254f79cedd063dc8c9.jpg", + "image_caption": [ + "Figure 2. Illustration of the Query Conflict Reduction (QCR) module. This module refines queries that predict incremental classes, allowing the processes of learning new features and retaining old features to occur separately. Note that the QCR module for previous incremental classes is frozen." + ], + "image_footnote": [], + "bbox": [ + 101, + 90, + 460, + 311 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "sification layer, especially in the latter queries. Due to the shared classifier and mask generator following different layers of the transformer decoder, the classification results on the same position of queries from adjacent layers are approximately the same. Thus, we add the adapter for the queries from the penultimate layer $Q_{L-1}$ with classification results of new classes $C^{2:t}$ , as shown in Fig. 2. To be more specific, the queries $Q_{L-1,\\tilde{c}}$ are selected for the adapter of incremental class $\\tilde{c} \\in C^{2:t}$ via:", + "bbox": [ + 89, + 404, + 483, + 540 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nQ _ {L - 1, \\tilde {c}} = \\left\\{Q _ {L - 1} (n), n \\in N \\wedge \\tilde {c} = \\arg \\max _ {c = 0} ^ {| C ^ {0: t} |} \\mathcal {E} _ {c l s} (n) \\right\\}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 545, + 482, + 566 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $Q_{L-1,\\tilde{c}}$ could contain none, one, or some selected queries that enable recognizing different instances. For each new category, we utilize the corresponding QCR module $f_{QCR,\\tilde{c}}(\\cdot)$ to refine the group of queries $Q_{L-1,\\tilde{c}}$ . Considering the efficiency and effectiveness, we introduce the low-rank two-layer adaptation as the QCR module:", + "bbox": [ + 89, + 571, + 483, + 662 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\hat {Q} _ {L - 1, \\tilde {c}} = f _ {Q C R, \\tilde {c}} \\left(Q _ {L - 1, \\tilde {c}}\\right) \\tag {3} \\\\ = Q _ {L - 1, \\tilde {c}} W _ {1} W _ {2} + Q _ {L - 1, \\tilde {c}}, \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 163, + 670, + 480, + 707 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\hat{Q}_{L - 1,\\tilde{c}}$ represent the adapted queries of class $\\tilde{c}$ , $W_{1} \\in \\mathbb{R}^{D \\times r}$ and $W_{2} \\in \\mathbb{R}^{r \\times D}$ denotes the weights for QCR. The QCR module separates the refined queries $\\hat{Q}_{L,\\tilde{c}}$ from the original queries $Q_{L,\\tilde{c}}$ , enabling a bifurcate structure for simultaneously learning $C^t$ with class-specific adaptation and keep memorizing the features of $Q_{L,\\tilde{c}}$ as well as the logits of $C^{0:t - 1}$ .", + "bbox": [ + 89, + 717, + 483, + 824 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Half-Distillation-Half-Learning Strategy", + "text_level": 1, + "bbox": [ + 89, + 832, + 437, + 849 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Previous approaches based on Mask2Former usually separate learning and distillation processes on class embeddings $\\mathcal{E}_{cls}$ due to its different learning strategies compared", + "bbox": [ + 89, + 854, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/9c5c7aa6edd5c3bd35306df3176cf547cf072e7981c9d23ae82d3c06a9327b9a.jpg", + "image_caption": [ + "Figure 3. Details of the Half-Distillation-Half-Learning strategy on $N$ queries, including classification loss on matched queries and Kullback-Leibler Divergence loss on the other queries. The latter involves the logits of both old and current classes." + ], + "image_footnote": [], + "bbox": [ + 527, + 90, + 883, + 292 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "to pixel-wise segmentation structure (e.g. DeepLabv3 [13]), mentioned in Sec. 3.2. This causes conflicts between distillation and learning, where the same logit could have contradictory optimization directions. Therefore, it is necessary to unify the optimization process towards a fixed target.", + "bbox": [ + 511, + 383, + 905, + 460 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our Half-Distillation-Half-Learning (HDHL) strategy combines distillation and learning for $N$ class embeddings $\\mathcal{E}_{cls}$ , as illustrated in Fig. 3. We follow the bipartite matching of Mask2Former and select embeddings $\\mathcal{E}_{cls}^{C^t}$ matched to labels of current categories $C^t$ and $\\mathcal{E}_{cls}^{C^{1:t-1}}$ matched to pseudo labels of old categories $C^{1:t-1}$ . These embeddings, following the original optimization strategy, are optimized by cross-entropy loss $\\mathcal{L}_{cls}$ . The unmatched embeddings $\\mathcal{E}_{cls}^{\\emptyset^t}$ are grouped under the no-object (no- $obj$ ) category. However, adhering to the original learning strategy leads to catastrophic forgetting. Besides, using the distillation strategy for old categories $C^{1:t-1}$ using $\\mathcal{E}_{cls}^{\\emptyset^{t-1}}$ ignores the learning of current categories $C^t$ . However, identifying regions that do not belong to $C^t$ is just as important as learning those that do. Considering this, we design a strategy to simultaneously optimize the logits of $C^{1:t-1}$ and $C^t$ with a shared Kullback-Leibler Divergence $\\mathcal{L}_{kl}$ :", + "bbox": [ + 511, + 462, + 906, + 727 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {k l} = \\frac {1}{| \\mathcal {E} _ {c l s} ^ {\\varnothing^ {t}} |} \\sum_ {j = 1} ^ {| \\mathcal {E} _ {c l s} ^ {\\varnothing^ {t}} |} \\sum_ {c = 0} ^ {| \\boldsymbol {C} ^ {0: t} |} \\varphi_ {c l s} ^ {\\varnothing^ {t - 1}} (j, c) \\log \\frac {\\varphi_ {c l s} ^ {\\varnothing^ {t - 1}} (j , c)}{\\varphi_ {c l s} ^ {\\varnothing^ {t}} (j , c)}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 744, + 906, + 792 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where", + "bbox": [ + 514, + 811, + 558, + 824 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\varphi_ {c l s} ^ {\\emptyset^ {t}} (j, c) = \\mathcal {S} ^ {t} (j, c), \\\\ \\varphi_ {c l s} ^ {\\emptyset^ {t - 1}} (j, c) = \\left\\{ \\begin{array}{l l} \\mathcal {S} ^ {t - 1} (j, c) & \\text {i f} c \\in C ^ {0: t - 1}, \\\\ 0 & \\text {e l s e}, \\end{array} \\right. \\tag {5} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 563, + 840, + 903, + 904 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "25670", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "and", + "bbox": [ + 89, + 92, + 119, + 104 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {S} ^ {t} (j, c) = \\frac {\\exp \\left(\\mathcal {E} _ {c l s} ^ {\\varnothing^ {t}} (j , c)\\right)}{\\sum_ {\\hat {c} = 0} ^ {| \\mathcal {C} ^ {0 : t} |} \\exp \\left(\\mathcal {E} _ {c l s} ^ {\\varnothing^ {t}} (j , \\hat {c})\\right)} \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 169, + 104, + 482, + 146 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "represent the softmax activation.", + "bbox": [ + 89, + 152, + 312, + 166 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The $\\mathcal{L}_{kl}$ not only distill the previous class embedding $\\mathcal{E}_{cls}^{\\emptyset^{t-1}}$ to the corresponding categories, including the no-object category, but decreasing the logits of current categories to figure out the objects apart from them. The unified HDHL strategy maintains the probability distribution of old logits seamlessly with the integration of new logits. By combining the original $\\mathcal{L}_{cls}$ and the designed $\\mathcal{L}_{kl}$ as $\\mathcal{L}_{DL}$ , the optimization process covers all the class embeddings without contradictory targets:", + "bbox": [ + 89, + 167, + 483, + 306 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {D L} = \\lambda_ {c l s} \\cdot \\mathcal {L} _ {c l s} + \\lambda_ {k l} \\cdot \\mathcal {L} _ {k l}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 181, + 320, + 482, + 338 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda_{cls}$ and $\\lambda_{kl}$ control the balance between two losses.", + "bbox": [ + 89, + 349, + 480, + 364 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Importance-Based Knowledge Distillation", + "text_level": 1, + "bbox": [ + 89, + 377, + 447, + 393 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Previous methods mitigate catastrophic forgetting by applying feature distillation, which may hinder the acquisition of new knowledge. To address this, we propose an Importance-Based Knowledge Distillation (IKD) method, in which stronger distillation is applied to features critical for old classes, while less emphasis is placed on less pertinent features, thereby balancing stability with plasticity.", + "bbox": [ + 89, + 401, + 483, + 507 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The importance measurement requires the cost matrix $\\mathcal{A} \\in \\mathbb{R}^{N \\times S}$ , adopted in the bipartite matching process, where $S$ is the number of image labels. Values in $\\mathcal{A}$ indicate the relationship between queries and classes. For each image, the transformer decoder generates $N$ predicted segments, denoted as $\\mathcal{Z} = \\left\\{(\\mathcal{E}_{cls}(n), M(n))\\right\\}_{n=1}^{N}$ . The cost matrix $\\mathcal{A}$ is then computed based on the output segments $\\mathcal{Z}$ and the image labels $\\boldsymbol{y}^t = \\{(c_s, m_s)\\}_{s=1}^S$ :", + "bbox": [ + 89, + 508, + 483, + 630 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {A} (n, s) = - \\lambda_ {c l s} \\cdot \\varphi_ {c l s} \\left(n, c _ {s}\\right) + \\lambda_ {\\text {m a s k}} \\cdot \\mathcal {L} _ {\\text {m a s k}} \\left(\\mathcal {E} _ {\\text {m a s k}} (n), m _ {s}\\right), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 642, + 495, + 670 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\varphi_{cls}$ is the classification output obtained by applying softmax to $\\mathcal{E}_{cls}$ , and $\\lambda_{cls}$ and $\\lambda_{mask}$ represent the corresponding hyperparameters.", + "bbox": [ + 89, + 672, + 483, + 718 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After each training stage, we compute the cost matrix $\\mathcal{A}$ across every image in the training set $D_{train}^{t}$ of the current incremental step $t$ . The minimum costs among the classes for $N$ segments are accumulated in the corresponding position of buffer $B^{t} \\in \\mathbb{R}^{N}$ , where segments with lower minimum cost values are more likely to recognize and cover the regions of a class $C^{t}$ . By normalizing and reversing the values of $B^{t}$ , we derive the importance matrix $I_{C^{t}}^{t} \\in \\mathbb{R}^{N}$ of current classes $C^{t}$ . The importance matrix $I^{t+1} \\in \\mathbb{R}^{N}$ is further estimated via the weighted accumulation of current importance matrix $I^{t} \\in \\mathbb{R}^{N}$ and $I_{C^{t}}^{t}$ , where the weights refer to the $|C^{1:t-1}|$ and $|C^{t}|$ . The generated $I^{t+1}$ will be", + "bbox": [ + 89, + 719, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "utilized in the next incremental training phase for weighted distillation of the queries $Q^t$ , which is computed as follows:", + "bbox": [ + 511, + 90, + 906, + 121 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {I K D} = \\frac {1}{N} \\sum_ {n = 1} ^ {N} I ^ {t} (n) \\cdot \\| Q ^ {t} (n), Q ^ {t - 1} (n) \\| _ {2} ^ {2}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 558, + 132, + 906, + 174 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\| \\cdot \\|_2$ denotes the Euclidean distance, $Q^t(n)$ and $Q^{t-1}(n)$ denotes the transformer decoder features from the current model $f^t(\\cdot)$ and the old model $f^{t-1}(\\cdot)$ , respectively. Algorithm 1 presents the details of the importance estimation procedure after training $f^t(\\cdot)$ .", + "bbox": [ + 511, + 184, + 906, + 262 + ], + "page_idx": 4 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1 Importance Matrix Estimation" + ], + "code_body": "1: Input: current model $f^{t}(\\cdot)$ , current dataset $D_{train}^{t}$ , current importance matrix $I^{t}$ \n2: Initialization: $I^{t+1} \\gets 0$ , $B^{t} \\gets 0$ \n3: for all $(\\boldsymbol{x}^{t}, \\boldsymbol{y}^{t}) \\in D_{train}^{t}$ do \n4: Compute $A$ following Eq. (8) \n5: Update the buffer: \n6: $B^{t} \\gets B^{t} + \\min(A, \\text{axis} = 1)$ \n7: end for \n8: for $n \\in \\{1, 2, \\dots, N\\}$ do \n9: $I_{C^{t}}^{t}(n) \\gets 1 - \\frac{B^{t}(n) - \\min(B^{t})}{\\max(B^{t}) - \\min(B^{t})}$ \n10: $I^{t+1}(n) \\gets \\frac{|C^{1:t-1}|}{|C^{1:t}|} \\cdot I^{t}(n) + \\frac{|C^{t}|}{|C^{1:t}|} \\cdot I_{C^{t}}^{t}(n)$ \n11: end for \n12: Output: Importance matrix $I^{t+1}$ for step $t + 1$", + "bbox": [ + 521, + 295, + 906, + 508 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.4. Objective Function", + "text_level": 1, + "bbox": [ + 513, + 534, + 697, + 550 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Overall, the objective function of CoMBO is defined as:", + "bbox": [ + 511, + 556, + 883, + 571 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {H} = \\mathcal {L} _ {D L} + \\lambda_ {I K D} \\mathcal {L} _ {I K D}. \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 614, + 585, + 906, + 599 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_{DL}$ represents the Half-Distillation-Half-Learning loss, and $\\mathcal{L}_{IKD}$ means the Importance-Based Knowledge Distillation loss, with $\\lambda_{IKD}$ as its weighting factor.", + "bbox": [ + 511, + 612, + 905, + 657 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Experimental Results", + "text_level": 1, + "bbox": [ + 511, + 671, + 718, + 686 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 511, + 696, + 702, + 712 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets and Evaluation Metrics. We follow the experimental settings of previous works [8, 11, 26] and evaluate our approach on the ADE20K [52] dataset. The ADE20K dataset is specifically designed to support both panoptic and semantic segmentation tasks. ADE20K contains 150 classes, including 100 thing classes and 50 stuff classes. The dataset is composed of 20,210 images for training and 2,000 images for validation.", + "bbox": [ + 511, + 719, + 905, + 839 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For Class Incremental Semantic Segmentation (CISS), we adopt the mean Intersection over Union (mIoU) for evaluation. The Intersection over Union (IoU) is calculated as $\\mathrm{IoU} = \\frac{TP}{TP + FP + FN}$ , where $TP$ , $FP$ , and $FN$ denote the", + "bbox": [ + 511, + 839, + 905, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "25671", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/55b544273ba64af708da489ec1473ddfec0907044a0bd317e0cd39de0a6461c9.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Method100-50 (2 steps)100-10 (6 steps)100-5 (11 steps)50-50 (3 steps)
1-100101-150all1-100101-150all1-100101-150all1-5051-150all
FT0.01.30.40.02.91.00.025.88.60.012.08.1
MiB [5]CVPR2035.119.329.827.110.021.424.06.518.142.415.524.4
PLOP [20]CVPR2140.222.434.330.517.526.128.115.724.045.818.727.7
CoMFormer [8]CVPR2341.127.736.736.017.129.734.415.928.245.019.327.9
ECLIPSE [26]CVPR2441.723.535.641.418.833.941.116.632.946.020.729.2
BalConpas [11]ECCV2442.825.737.140.722.834.736.120.330.851.226.534.7
CoMBO Ours43.925.637.840.825.235.636.120.530.950.728.235.7
Joint43.830.939.543.830.939.543.830.939.550.733.939.5
", + "bbox": [ + 94, + 88, + 903, + 255 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Quantitative comparison under Class Incremental Panoptic Segmentation with state-of-the-art exemplar-free methods on ADE20K in PQ. Scores of novel classes and all classes in bold are the best while underlined are the second best.", + "bbox": [ + 89, + 265, + 903, + 295 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "number of true-positive, false-positive, and false-negative pixels, respectively. The mIoU metric averages the IoU across all classes for a more comprehensive evaluation. For Class Incremental Panoptic Segmentation (CIPS), following previous work [8], we employ Panoptic Quality (PQ) as the evaluation metric. PQ is defined as the product of Recognition Quality (RQ) and Segmentation Quality (SQ). To measure incremental learning capacity, we compute the corresponding metrics for the initial classes $C^1$ , incremental classes $C^{2:T}$ , and the aggregate of all classes $C^{1:T}$ .", + "bbox": [ + 88, + 320, + 480, + 470 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Protocols and Implementation Details. We follow previous incremental protocols and define scenarios as $N_{ini} - N_{inc}$ , where $N_{ini}$ represents the number of initial classes, and $N_{inc}$ denotes the number of new classes introduced at each incremental step. For example, in the 100-10 scenario, the training begins with 100 classes, followed by the addition of 10 new classes per incremental step, without access to annotations of old classes. For both CIPS and CISS, we do evaluations on the following scenarios: 100-10 (6 steps), 100-50 (2 steps), 100-5 (11 steps) and 50-50 (3 steps).", + "bbox": [ + 89, + 489, + 482, + 641 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our approach is based on the Mask2Former structure [17]. Following previous approaches to CISS and CIPS, we utilize ResNet-50 [24] as the backbone for CIPS and ResNet-101 for CISS, with both pre-trained on ImageNet [19]. The input resolution of the images is $640 \\times 640$ with an all-time batch size of 8. We follow the training hyperparameter of Mask2Former in the initial step, with a learning rate of $10^{-4}$ and iterations of 160,000. During the incremental steps, we set 1,000 iterations per class with a learning rate of $5 \\times 10^{-5}$ . The coefficients $r$ , $\\lambda_{cls}$ , $\\lambda_{kl}$ , $\\lambda_{IKD}$ are respectively set to 16, 2, 5, 3. All experiments were conducted on the NVIDIA RTX 4090.", + "bbox": [ + 89, + 641, + 482, + 821 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2.Quantitative Results", + "text_level": 1, + "bbox": [ + 89, + 832, + 284, + 848 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Comparisons in Class Incremental Panoptic Segmentation (CIPS). We evaluated our approach against state-of-the-art exemplar-free methods on the ADE20K dataset", + "bbox": [ + 89, + 854, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "within the CIPS framework, as detailed in Tab. 1. The performance of our CoMBO surpasses other methods in most of the scenarios, particularly in the more challenging 100-10 scenario, where it achieves a remarkable $35.6\\%$ PQ with at least $3.4\\%$ of advantage on incremental classes compared to the previous state-of-the-art methods. This notable performance gain demonstrates the effectiveness of our CoMBO in reducing conflicts. Although our approach still underperforms the state-of-the-art approach ECLIPSE [26] in the 100-5 scenario, primarily due to its strong memorization ability from freezing parameters, our method achieves advanced performance for the new classes. Moreover, in scenarios with fewer initial classes, i.e., 50-50 scenario, we maintain our advantage in learning new classes with a PQ of $28.2\\%$ for incremental classes and $35.7\\%$ for overall performance, widening the gap with strong distillation methods.", + "bbox": [ + 511, + 320, + 906, + 561 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Comparisons in Class Incremental Semantic Segmentation (CISS). We further extended our evaluation to the semantic segmentation benchmark, comparing our approach with previous methods on the ADE20K dataset, as detailed in Tab. 2. Our approach consistently outperforms previous methods in all tested scenarios. Notably, in the 100-10 scenario, our approach achieves a mIoU of $41.1\\%$ , surpassing the previous state-of-the-art by at least $3.5\\%$ for incremental classes, demonstrating its effectiveness in mitigating conflicts. Moreover, in the most challenging 100-5 long-term incremental scenario, our approach not only improves performance on old classes by $2.5\\%$ but also achieves a significant $5.4\\%$ increase in mIoU for new classes, ensuring both model stability and adaptability.", + "bbox": [ + 511, + 580, + 908, + 792 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6. Ablation Study", + "text_level": 1, + "bbox": [ + 511, + 806, + 663, + 824 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6.1. Component Ablations", + "text_level": 1, + "bbox": [ + 511, + 830, + 718, + 847 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We analyze the key components of our proposed framework, which includes the Half-Distillation-Half-Learning Strategy, Importance-based Knowledge Distillation, and the", + "bbox": [ + 511, + 854, + 903, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "25672", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/96b3c37a64fa2ecc2581d4f2b7775056327cb318ad83d268d495e2ec5c8215e1.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Method100-50 (2 steps)100-10 (6 steps)100-5 (11 steps)50-50 (3 steps)
1-100101-150all1-100101-150all1-100101-150all1-5051-150all
FT0.026.78.90.02.30.80.01.10.30.01.71.1
MiB [5]CVPR2037.024.132.623.510.626.621.06.116.145.621.029.3
PLOP [20]CVPR2144.226.238.234.815.928.539.513.630.954.930.238.4
CoMFormer [8]CVPR2344.726.238.440.615.632.339.513.630.949.226.634.1
CoMasTRe [23]CVPR2445.726.039.242.318.434.440.815.832.649.826.634.5
ECLIPSE [26]CVPR2445.021.737.143.417.434.643.316.334.2---
BalConpas [11]ECCV2449.930.143.347.324.238.642.117.233.855.833.340.8
CoMBO Ours50.234.444.947.827.741.144.622.637.355.336.943.0
Joint51.740.247.851.740.247.851.740.247.856.643.547.8
", + "bbox": [ + 94, + 88, + 903, + 268 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Query Conflict Reducing module. Our experiments focus on the 100-10 scenario in the CIPS, as it provides a moderate number of steps to examine the ability of both acquisition and retention. We evaluate various combinations of these components separately and present the results in Tab. 3. The baseline method utilizes the vanilla loss with pseudo-labeling.", + "bbox": [ + 88, + 333, + 480, + 441 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/2a7c60d0d35e7eeaa42b1b21f45f0c99e4a757681e5a8800e5a6c0b3b8a377f2.jpg", + "table_caption": [ + "Table 2. Quantitative comparison under Class Incremental Semantic Segmentation with state-of-the-art exemplar-free methods on ADE20K in mIoU. Scores of novel classes and all classes in bold are the best while underlined are the second best." + ], + "table_footnote": [], + "table_body": "
HDHLIKDQCR100-10 (6 steps)
1-100101-150all
36.323.832.2
✓39.424.234.3
✓✓40.024.634.9
✓✓38.825.434.4
✓✓✓40.825.235.6
", + "bbox": [ + 109, + 454, + 464, + 575 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Ablation study of the main components on the 100-10 task of CIPS. The baseline in the $1^{\\text{st}}$ row employs vanilla losses with pseudo-labeling and excludes the QCR module.", + "bbox": [ + 89, + 587, + 482, + 630 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Under identical experimental conditions, the HDHL Strategy significantly enhances the overall PQ by $2.1\\%$ , demonstrating its ability to unify logits of new classes into the previous logits distribution without contradictory losses. The implementation of the IKD mechanism has a remarkable $3.7\\%$ increase in the PQ for old classes, underscoring its effectiveness in boosting the ability to retain the knowledge of old classes. Furthermore, when combined with the QCR module for branched optimization, CoMBO further boosts both the performance of old and new classes, achieving $0.8\\%$ and $0.6\\%$ gains in PQ, respectively, surpassing previous methods that struggle to balance acquisition and retention. Consequently, CoMBO achieves a $3.4\\%$ PQ improvement over the baseline. The above results demonstrate the effectiveness of the proposed CoMBO and its related modules, showing a noticeable improvement in reducing the conflict on model structures and losses.", + "bbox": [ + 91, + 643, + 482, + 898 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.2. Ablation Study of QCR", + "text_level": 1, + "bbox": [ + 511, + 332, + 730, + 349 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We validate the effectiveness of QCR with different settings of $r$ shown in Fig. 5. Under the estimation of performance enhancement and additional parameters, our QCR module achieves the best result when $r = 16$ , where the quantity of additional parameters comes to $8.2\\mathrm{K}$ per class. Considering the total parameter of Mask2Former [17] is $44.9\\mathrm{M}$ with ResNet-50, the additional parameters only reach up to $2\\%$ of original parameters in total, which is 5 times less than per class additional parameters of ECLIPSE [26] while having better performance on new classes.", + "bbox": [ + 511, + 354, + 906, + 507 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.3. Ablation Study of HDHL Strategy", + "text_level": 1, + "bbox": [ + 511, + 513, + 812, + 531 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To validate the effectiveness of our Half-Distillation-Half-Learning strategy, we conduct a series of experiments among other strategies, as shown in Tab. 4. The $1^{\\mathrm{st}}$ row with $\\mathcal{L}_{cls}$ on all classification embeddings represents the vanilla loss with pseudo-labeling, which is adopted in [8]. The $2^{\\mathrm{nd}}$ row utilizes an auxiliary distillation on the embeddings following [23]. Compared to these strategies with tough supervision on unmatched embeddings $\\mathcal{E}_{cls}^{t,\\emptyset}$ and mostly conflicting losses, our HDHL strategy with soft and concentrated supervision has a better PQ of 34.3. Besides, the HDHL design of $\\mathcal{L}_{kl}$ ensures the integrity while optimizing the unmatched embeddings $\\mathcal{E}_{cls}^{t,\\emptyset}$ , with improvement of $2.1\\%$ compared to distillation-only strategy in the $3^{\\mathrm{rd}}$ row and $0.8\\%$ compared to the simple combination of distillation and $\\mathcal{L}_{cls}$ . The above results show that our HDHL strategy outperforms other learning or distillation strategies on the class embeddings.", + "bbox": [ + 511, + 537, + 906, + 794 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.4. Qualitative Analysis", + "text_level": 1, + "bbox": [ + 511, + 801, + 705, + 816 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We conduct the qualitative analysis (shown in Fig. 4) by comparing our proposed CoMBO to the Baseline and Baseline+HDHL $(1^{\\mathrm{st}}$ and $2^{\\mathrm{nd}}$ rows in Tab. 3). In the $1^{\\mathrm{st}}$ , $2^{\\mathrm{nd}}$ , and $5^{\\mathrm{th}}$ columns of Fig. 4, our CoMBO can indicate the incremental classes, i.e., screen door, fountain, and ottoman,", + "bbox": [ + 511, + 824, + 905, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "25673", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/aa5a3fe79756d31ace26406ce9e8d831a64fa3a8d44c041e001872af8c0736a3.jpg", + "image_caption": [ + "Figure 4. Qualitative results of CoMBO comparing to Baseline, Baseline+HDHL on 100-10 CIPS task of ADE20K." + ], + "image_footnote": [], + "bbox": [ + 122, + 93, + 870, + 441 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c8eee99b786109d66d892a0b6bb7ddb6069f54262cde2270d14dc4c33794e2fe.jpg", + "image_caption": [ + "Figure 5. Ablation study of $r$ in QCR module, including PQ performance (left) and number of parameters (right)." + ], + "image_footnote": [], + "bbox": [ + 99, + 484, + 470, + 660 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/e6650164ecfd5d39c4d80b77764d70d09f27be8153352329b5ad55170914ed70.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
LclsLklPQ
Ectls & Ectls1:t-1EctlsC0:t-1Ct
✓✓32.2
✓✓✓32.6
✓✓32.2
✓✓✓33.5
✓✓✓34.3
", + "bbox": [ + 122, + 696, + 452, + 819 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. Ablation study of the HDHL strategy on task 100-10 of CIPS. The $1^{\\text{st}}$ row represents the vanilla loss with pseudo-labeling. The last row represents our $\\mathcal{L}_{DL}$ .", + "bbox": [ + 89, + 830, + 483, + 875 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "while other methods fail to recognize all these objects. In the $3^{\\mathrm{rd}}$ column, our method successfully remembers the car in the left instead of covering the object with the prediction of the incremental class van. Furthermore, our method allows more precise mask prediction of new class Traffic Light in the $4^{\\mathrm{th}}$ column, demonstrating the effectiveness of refinements on the queries corresponding to new classes.", + "bbox": [ + 511, + 487, + 906, + 595 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Conclusions", + "text_level": 1, + "bbox": [ + 511, + 606, + 640, + 622 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we present CoMBO, a novel Class Incremental Segmentation (CIS) method for mitigating the conflict between acquisition and retention losses. This conflict arises from competing goals of efficiently acquiring knowledge about new classes while preserving knowledge of previously learned ones. Our QCR module branches the conflicting optimization targets via lightweight class-specific adaptation on queries, enabling the coexistence of learning and distillation on separated queries. The HDHL strategy and IKD further reduce the contradictory optimization on classification logits and queries with unified targets, respectively, thereby enhancing the overall model performance. Extensive experimental validation on the ADE20K dataset underscores the superiority of our approach, demonstrating the effectiveness of each component in the CIS task, particularly excelling in performance on new classes.", + "bbox": [ + 511, + 631, + 908, + 875 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "25674", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 91, + 90, + 250, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported by the National Key R&D Program of China (Grant 2022YFC3310200), the National Natural Science Foundation of China (Grant 62472033, 92470203), the Beijing Natural Science Foundation (Grant L242022), the Royal Society grants (SIF\\R1\\231009, IES\\R3\\223050) and an Amazon Research Award.", + "bbox": [ + 89, + 114, + 483, + 205 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 217, + 187, + 233 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Rahaf Aljundi, Min Lin, Baptiste Goujaud, and Yoshua Bengio. Gradient based sample selection for online continual learning. Adv. Neural Inf. Process. Syst., 32, 2019. 2", + "[2] Vijay Badrinarayanan, Alex Kendall, and Roberto Cipolla. Segnet: A deep convolutional encoder-decoder architecture for image segmentation. IEEE Trans. on Pattern Anal. Mach. Intell., 39(12):2481-2495, 2017. 2", + "[3] Donghyeon Baek, Youngmin Oh, Sanghoon Lee, Junghyup Lee, and Bumsub Ham. Decomposed knowledge distillation for class-incremental semantic segmentation. Adv. Neural Inf. Process. Syst., 35:10380-10392, 2022. 2, 3", + "[4] Zalán Borsos, Mojmir Mutny, and Andreas Krause. Coresets via bilevel optimization for continual learning and streaming. Adv. Neural Inf. Process. Syst., 33:14879-14890, 2020. 2", + "[5] Fabio Cermelli, Massimiliano Mancini, Samuel Rota Bulo, Elisa Ricci, and Barbara Caputo. Modeling the background for incremental learning in semantic segmentation. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 9233-9242, 2020. 2, 3, 6, 7", + "[6] Fabio Cermelli, Dario Fontanel, Antonio Tavera, Marco Ciccone, and Barbara Caputo. Incremental learning in semantic segmentation from image labels. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 4371-4381, 2022.", + "[7] Fabio Cermelli, Antonino Geraci, Dario Fontanel, and Barbara Caputo. Modeling missing annotations for incremental learning in object detection. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 3700-3710, 2022. 2", + "[8] Fabio Cermelli, Matthieu Cord, and Arthur Douillard. Comformer: Continual learning in semantic and panoptic segmentation. In Proc. IEEE Conf. Comput. Vis. Pattern Recognition., pages 3010-3020, 2023. 2, 3, 5, 6, 7", + "[9] Sungmin Cha, YoungJoon Yoo, Taesup Moon, et al. Ssul: Semantic segmentation with unknown label for exemplar-based class-incremental learning. Adv. Neural Inf. Process. Syst., 34:10919-10930, 2021. 2", + "[10] Jinpeng Chen, Runmin Cong, Yuxuan Luo, Horace Ip, and Sam Kwong. Saving 100x storage: prototype replay for reconstructing training sample distribution in class incremental semantic segmentation. Adv. Neural Inf. Process. Syst., 36, 2024. 2", + "[11] Jinpeng Chen, Runmin Cong, Yuxuan Luo, Horace Ho Shing Ip, and Sam Kwong. Strike a balance in continual panoptic segmentation. In Proc. Eur. Conf. Comput. Vis., pages 126-142. Springer, 2025. 2, 3, 5, 6, 7", + "[12] Liang-Chieh Chen. Semantic image segmentation with deep convolutional nets and fully connected crfs. arXiv preprint arXiv:1412.7062, 2014. 2" + ], + "bbox": [ + 93, + 243, + 482, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] L. C. Chen. Rethinking atrous convolution for semantic image segmentation. arXiv preprint:1706.05587, 2017. 2, 3, 4", + "[14] Liang-Chieh Chen, George Papandreou, and et al. Deep convolutional nets, atrous convolution, and fully connected crfs. IEEE Trans. on Pattern Anal. Mach. Intell., 40(4):834-848, 2017.", + "[15] Liang-Chieh Chen, Yukun Zhu, George Papandreou, Florian Schroff, and Hartwig Adam. Encoder-decoder with atrous separable convolution for semantic image segmentation. In Proc. Eur. Conf. Comput. Vis., pages 801-818, 2018. 2", + "[16] Bowen Cheng, Alex Schwing, and Alexander Kirillov. Per-pixel classification is not all you need for semantic segmentation. Adv. Neural Inf. Process. Syst., 34:17864-17875, 2021. 1, 2", + "[17] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 1290–1299, 2022. 1, 2, 3, 6, 7", + "[18] Wei Cong, Yang Cong, Yuyang Liu, and Gan Sun. Cs2k: Class-specific and class-shared knowledge guidance for incremental semantic segmentation. In Proc. Eur. Conf. Comput. Vis., pages 244-261, 2025. 2", + "[19] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 248–255, 2009. 6", + "[20] Arthur Douillard, Yifu Chen, Arnaud Dapogny, and Matthieu Cord. Plop: Learning without forgetting for continual semantic segmentation. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 4040-4050, 2021. 2, 3, 6, 7", + "[21] Arthur Douillard, Yifu Chen, Arnaud Dapogny, and Matthieu Cord. Tackling catastrophic forgetting and background shift in continual semantic segmentation. arXiv preprint arXiv:2106.15287, 2021. 2", + "[22] Robert M French. Catastrophic forgetting in connectionist networks. Trends in cognitive sciences, 3(4):128-135, 1999. 2", + "[23] Yizheng Gong, Siyue Yu, Xiaoyang Wang, and Jimin Xiao. Continual segmentation with disentangled objectness learning and class recognition. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 3848-3857, 2024. 3, 7", + "[24] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 770-778, 2016. 6", + "[25] M. G. Z. A. Khan, M. F. Naeem, Luc Van Gool, Didier Stricker, Federico Tombari, and Muhammad Zeshan Afzal. Introducing language guidance in prompt-based continual learning. In Proc. IEEE Int. Conf. Comput. Vis., pages 11463-11473, 2023. 2", + "[26] Beomyoung Kim, Joonsang Yu, and Sung Ju Hwang. Eclipse: Efficient continual learning in panoptic segmentation with visual prompt tuning. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 3346-3356, 2024. 2, 3, 5, 6, 7", + "[27] Alexander Kirillov, Kaiming He, Ross Girshick, Carsten Rother, and Piotr Dollar. Panoptic segmentation. In Proc." + ], + "bbox": [ + 516, + 92, + 905, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "25675", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "IEEE Conf. Comput. Vis. Pattern Recognit., pages 9404-9413, 2019. 2", + "[28] Zhizhong Li and Derek Hoiem. Learning without forgetting. IEEE Trans. on Pattern Anal. Mach. Intell., 40(12):2935-2947, 2017. 3", + "[29] Guosheng Lin, Anton Milan, Chunhua Shen, and Ian Reid. Refinenet: Multi-path refinement networks for high-resolution semantic segmentation. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 1925-1934, 2017. 2", + "[30] Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 3431-3440, 2015. 2", + "[31] David Lopez-Paz and Marc'Aurelio Ranzato. Gradient episodic memory for continual learning. Adv. Neural Inf. Process. Syst., 30, 2017. 2", + "[32] Andrea Maracani, Umberto Michieli, Marco Toldo, and Pietro Zanuttigh. Recall: Replay-based continual learning in semantic segmentation. In Proc. IEEE Int. Conf. Comput. Vis., pages 7026-7035, 2021. 2", + "[33] Umberto Michieli and Pietro Zanuttigh. Continual semantic segmentation via repulsion-attraction of sparse and disentangled latent representations. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 1114-1124, 2021. 2", + "[34] Umberto Michieli and Pietro Zanuttigh. Continual semantic segmentation via repulsion-attraction of sparse and disentangled latent representations. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 1114-1124, 2021. 2", + "[35] Youngmin Oh, Donghyeon Baek, and Bumsub Ham. Alife: Adaptive logit regularizer and feature replay for incremental semantic segmentation. Adv. Neural Inf. Process. Syst., 35: 14516-14528, 2022. 2", + "[36] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Proc. International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 234-241, 2015. 2", + "[37] Chao Shang, Hongliang Li, Fanman Meng, Qingbo Wu, Heqian Qiu, and Lanxiao Wang. Incrementer: Transformer for class-incremental semantic segmentation with knowledge distillation focusing on old class. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 7214-7224, 2023. 2", + "[38] Robin Strudel, Ricardo Garcia, Ivan Laptev, and Cordelia Schmid. Segmenter: Transformer for semantic segmentation. In Proc. IEEE Int. Conf. Comput. Vis., pages 7262-7272, 2021. 2", + "[39] Sebastian Thrun. Lifelong learning algorithms. In Learning to learn, pages 181-209. 1998. 2", + "[40] Zifeng Wang, Zizhao Zhang, Chen-Yu Lee, Han Zhang, Ruoxi Sun, Xiaoqi Ren, Guolong Su, Vincent Perot, Jennifer Dy, and Tomas Pfister. Learning to prompt for continual learning. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 139-149, 2022. 2", + "[41] Yunchao Wei, Jiashi Feng, Xiaodan Liang, Ming-Ming Cheng, Yao Zhao, and Shuicheng Yan. Object region mining with adversarial erasing: A simple classification to semantic segmentation approach. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 1568-1576, 2017. 2" + ], + "bbox": [ + 91, + 90, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[42] Jia-Wen Xiao, Chang-Bin Zhang, Jiekang Feng, Xialei Liu, Joost van de Weijer, and Ming-Ming Cheng. Endpoints weight fusion for class incremental semantic segmentation. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 7204-7213, 2023. 2", + "[43] Shipeng Yan, Jiangwei Xie, and Xuming He. Der: Dynamically expandable representation for class incremental learning. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 3014-3023, 2021. 2", + "[44] Chaohui Yu, Qiang Zhou, Jingliang Li, Jianlong Yuan, Zhibin Wang, and Fan Wang. Foundation model drives weakly incremental learning for semantic segmentation. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 23685-23694, 2023. 2", + "[45] Qihang Yu, Huiyu Wang, Siyuan Qiao, Maxwell Collins, Yukun Zhu, Hartwig Adam, Alan Yuille, and Liang-Chieh Chen. k-means mask transformer. In Proc. Eur. Conf. Comput. Vis., pages 288-307, 2022. 2", + "[46] Anqi Zhang and Guangyu Gao. Background adaptation with residual modeling for exemplar-free class-incremental semantic segmentation. In Proc. Eur. Conf. Comput. Vis., 2024. 3", + "[47] Chang-Bin Zhang, Jia-Wen Xiao, Xialei Liu, Ying-Cong Chen, and Ming-Ming Cheng. Representation compensation networks for continual semantic segmentation. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 7053-7064, 2022. 2, 3", + "[48] Zekang Zhang, Guangyu Gao, Zhiyuan Fang, Jianbo Jiao, and Yunchao Wei. Mining unseen classes via regional objectness: A simple baseline for incremental segmentation. Adv. Neural Inf. Process. Syst., 35:24340-24353, 2022. 2", + "[49] Zekang Zhang, Guangyu Gao, Jianbo Jiao, Chi Harold Liu, and Yunchao Wei. Coinseg: Contrast inter-and intra-class representations for incremental segmentation. In Proc. IEEE Int. Conf. Comput. Vis., pages 843-853, 2023. 2", + "[50] Hengshuang Zhao, Jianping Shi, Xiaojuan Qi, Xiaogang Wang, and Jiaya Jia. Pyramid scene parsing network. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 2881-2890, 2017. 2", + "[51] Hanbin Zhao, Fengyu Yang, Xinghe Fu, and Xi Li. Rbc: Rectifying the biased context in continual semantic segmentation. In Proc. Eur. Conf. Comput. Vis., pages 55-72, 2022. 2", + "[52] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 633-641, 2017. 5", + "[53] Lanyun Zhu, Tianrun Chen, Jianxiong Yin, Simon See, and Jun Liu. Continual semantic segmentation with automatic memory sample selection. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 3082-3092, 2023. 2" + ], + "bbox": [ + 516, + 90, + 903, + 811 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "25676", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/CoMBO_ Conflict Mitigation via Branched Optimization for Class Incremental Segmentation/a4944725-d45e-4b4f-a404-3fd5f2ee3b6a_model.json b/2025/CoMBO_ Conflict Mitigation via Branched Optimization for Class Incremental Segmentation/a4944725-d45e-4b4f-a404-3fd5f2ee3b6a_model.json new file mode 100644 index 0000000000000000000000000000000000000000..e7b5f94c2b0390c99af235968db3513d3f3b179d --- /dev/null +++ b/2025/CoMBO_ Conflict Mitigation via Branched Optimization for Class Incremental Segmentation/a4944725-d45e-4b4f-a404-3fd5f2ee3b6a_model.json @@ -0,0 +1,2079 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.197, + 0.131, + 0.803, + 0.177 + ], + "angle": 0, + "content": "CoMBO: Conflict Mitigation via Branched Optimization for Class Incremental Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.203, + 0.882, + 0.242 + ], + "angle": 0, + "content": "Kai Fang\\(^{1*}\\), Anqi Zhang\\(^{1*}\\), Guangyu Gao\\(^{1\\dagger}\\), Jianbo Jiao\\(^{2}\\), Chi Harold Liu\\(^{1}\\), Yunchao Wei\\(^{3}\\) \n\\(^{1}\\)Beijing Institute of Technology \\(^{2}\\)University of Birmingham \\(^{3}\\)Beijing Jiaotong University" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.274, + 0.327, + 0.29 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.306, + 0.486, + 0.761 + ], + "angle": 0, + "content": "Effective Class Incremental Segmentation (CIS) requires simultaneously mitigating catastrophic forgetting and ensuring sufficient plasticity to integrate new classes. The inherent conflict above often leads to a back-and-forth, which turns the objective into finding the balance between the performance of previous (old) and incremental (new) classes. To address this conflict, we introduce a novel approach, Conflict Mitigation via Branched Optimization (CoMBO). Within this approach, we present the Query Conflict Reduction module, designed to explicitly refine queries for new classes through lightweight, class-specific adapters. This module provides an additional branch for the acquisition of new classes while preserving the original queries for distillation. Moreover, we develop two strategies to further mitigate the conflict following the branched structure, i.e., the Half-Learning Half-Distillation (HDHL) over classification probabilities, and the Importance-Based Knowledge Distillation (IKD) over query features. HDHL selectively engages in learning for classification probabilities of queries that match the ground truth of new classes, while aligning unmatched ones to the corresponding old probabilities, thus ensuring retention of old knowledge while absorbing new classes via learning negative samples. Meanwhile, IKD assesses the importance of queries based on their matching degree to old classes, prioritizing the distillation of important features and allowing less critical features to evolve. Extensive experiments in Class Incremental Panoptic and Semantic Segmentation settings have demonstrated the superior performance of CoMBO. Project page: https://guangyu-ryan.github.io/CoMBO." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.791, + 0.222, + 0.807 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.817, + 0.485, + 0.864 + ], + "angle": 0, + "content": "Semantic segmentation, the fundamental task in computer vision, involves classifying each pixel into predefined categories. Panoptic segmentation, the challeng-" + }, + { + "type": "image", + "bbox": [ + 0.521, + 0.272, + 0.909, + 0.624 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.637, + 0.907, + 0.75 + ], + "angle": 0, + "content": "Figure 1. Comparison of our improved distillation and learning strategy (bottom) with previous conflicting strategy (top). Previous strategies impose contradictory supervision on the same target to find a balance, whereas our strategy, including the importance factor for distillation and the QCR module for adaptive learning, enables more compatible, target-specific supervision. The visualization on incremental classes ('Traffic Light') highlights the effectiveness of our strategy." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.909, + 0.903 + ], + "angle": 0, + "content": "ing variant, unifies semantic and instance segmentation to both classify every pixel and simultaneously identify distinct object instances. Recent advancements in mask classification-based segmentation, such as MaskFormer [16] and Mask2Former [17], enable the unification of semantic and panoptic segmentation. However, traditional approaches to segmentation suffer from limitations in dynamic environments where new classes can emerge unpredictably, as they are typically trained on a static set" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.875, + 0.22, + 0.888 + ], + "angle": 0, + "content": "*Equal contribution." + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.37, + 0.901 + ], + "angle": 0, + "content": "† Corresponding author, guangyugao@bit.edu.cn." + }, + { + "type": "list", + "bbox": [ + 0.109, + 0.875, + 0.37, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "25667" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.152 + ], + "angle": 0, + "content": "of categories. Therefore, Class Incremental Semantic Segmentation (CISS) has emerged as a pivotal task for continually learning new categories, where models can adapt to new classes without forgetting previously learned ones." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.155, + 0.483, + 0.412 + ], + "angle": 0, + "content": "Historically, studies in CISS [3, 35, 49] mainly focuses on tackling catastrophic forgetting, employing various techniques such as knowledge distillation [5, 20, 47], freezing parameters [9, 26, 48], or generating pseudo-labels from old model [20]. More recent efforts [8, 11] have extended these techniques to the more challenging Class Incremental Panoptic Segmentation (CIPS) based on the advanced Mask2Former. Nonetheless, these methods often fail to adequately reconcile the inherent tension between retaining old knowledge and acquiring new information. Some overemphasize maintaining stability, thus hindering the learning of new classes, while others prioritize plasticity, leading to significant forgetting of old classes. It is necessary to deconstruct these conflicting objectives to facilitate a more compatible incremental learning process, achieving promotion in both the acquisition of new classes and retention of old classes." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.415, + 0.483, + 0.884 + ], + "angle": 0, + "content": "In this work, we propose Conflict Mitigation via Branched Optimization (CoMBO), an approach to reduce conflicts between acquiring new classes and retaining old classes. By integrating three specialized knowledge distillation mechanisms within the Mask2Former architecture, CoMBO effectively mitigates catastrophic forgetting while enhancing the assimilation of new classes. Using bipartite matching, Mask2Former optimally aligns each query, complete with its classification logits and mask predictions, with the most fitting ground truth class, ensuring efficient optimization. Therefore, we first introduce the Half-Learning Half-Distillation mechanism, which selectively applies a shared Kullback-Leibler divergence to the classification logits of unmatched queries. This method distinctly separates distillation on unmatched masks from the classification optimization of matched ones. Moreover, rather than applying uniform distillation across all features, we introduce an importance-based knowledge distillation technique. This technique assesses each query's significance in preserving knowledge of old classes and adjusts the distillation intensity accordingly, emphasizing the retention of key queries. Additionally, our Query Conflict Reduction module innovatively adapts features for new classes while maintaining the queries' original characteristics before distillation, using class-specific adapters for each newly learned class. Finally, our approach outperforms previous methods in multiple benchmarks, especially in typical and challenging scenarios such as 100-10 of ADE20K. Our approach achieves remarkable performance of \\(35.6\\%\\) and \\(41.1\\%\\) mIoU on 100-10 of CIPS and CISS, respectively, compared to the previous state-of-the-art." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.886, + 0.466, + 0.9 + ], + "angle": 0, + "content": "Overall, our contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.092, + 0.905, + 0.151 + ], + "angle": 0, + "content": "- We propose a Half-Learning Half-Distillation mechanism that applies soft distillation for old classes and binary-based optimization for new classes to effectively balance retention and acquisition." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.153, + 0.905, + 0.212 + ], + "angle": 0, + "content": "- An extra Query Conflict Reduction module is proposed to refine queries for new classes while preserving unrefined features, which enables Importance-based Knowledge Distillation for remembering the key features." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.213, + 0.905, + 0.273 + ], + "angle": 0, + "content": "- Through extensive quantitative and qualitative evaluations on the ADE20K dataset, we demonstrate the state-of-the-art performance of our model in both Class Incremental Semantic and Panoptic Segmentation tasks." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.092, + 0.905, + 0.273 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.288, + 0.655, + 0.303 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.314, + 0.836, + 0.33 + ], + "angle": 0, + "content": "2.1. Semantic and Panoptic Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.338, + 0.907, + 0.701 + ], + "angle": 0, + "content": "Semantic segmentation aims to classify each pixel in an image, which focuses on the complete coverage of object(s) in each category. The pioneering work FCN [30] and U-Net [36] introduces a pixel-wise classification paradigm by removing the fully connected layers. Since then, numerous approaches [2, 12-15, 29, 38, 41] have been proposed following the paradigm, containing effective modules such as ASPP [13] and pyramid structure [50]. In contrast, panoptic segmentation [27] combines semantic and instance segmentation, requiring the classification of every pixel while distinguishing instances of the same category. These two tasks were addressed separately due to the limitation of the pixelwise classification. However, the emergence of the mask classification paradigm [16, 17, 43, 45] introduced a universal, transformer-based framework capable of solving multiple segmentation tasks simultaneously. MaskFormer [16] pioneered this paradigm by combining class-agnostic mask proposal generation and mask classification, enabling both semantic and panoptic segmentation simultaneously. Building upon this, Mask2Former [17] enhanced performance by incorporating multi-scale feature fusion and masked attention techniques. However, when learning new categories beyond the initial data, mask classification methods remain susceptible to catastrophic forgetting." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.711, + 0.738, + 0.727 + ], + "angle": 0, + "content": "2.2. Continual Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.907, + 0.903 + ], + "angle": 0, + "content": "In Class Incremental Learning (CIL), fine-tuning models on new data often results in a performance drop on old categories, i.e., catastrophic forgetting [22, 39]. To address this, various methods [5-7, 20, 21, 32, 33, 51] have been proposed, including retaining representative old samples [1, 4, 9, 10, 31, 32, 35, 53], compensation losses [18, 34, 42, 44], and prompt representations [25, 37, 40]. These advancements in CIL raise interest in more challenging segmentation tasks. Class Incremental Semantic Segmentation (CISS) was first introduced in MiB [5], reconstructing background regions to address background shifting during" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "25668" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.091, + 0.485, + 0.364 + ], + "angle": 0, + "content": "incremental steps. Since then, most methods [3, 47] have employed techniques such as knowledge distillation [28], pseudo-labels from previous models [8, 20], and background redefinition [5, 46] to mitigate background shifting. However, these methods primarily rely on pixel-wise classification models like DeepLabV3 [13]. Therefore, CoM-Former [8] first utilizes the mask classification segmentation model Mask2Former [17] enabling solving both CISS and Class Incremental Panoptic Segmentation (CIPS) tasks. ECLIPSE [26] introduces dynamic model structure expansion methods to extend learnable parameters for new classes and freezes the old parameters. CoMasTRe [23] distills queries matched to old queries with high probability on old classes. BalConpas [11] balance the class proportion on selecting representative replay samples. However, these methods often struggle to balance acquisition and preservation. Our approach mitigates this conflict by separating queries to decouple the two objectives." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.378, + 0.231, + 0.394 + ], + "angle": 0, + "content": "3. Preliminaries" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.403, + 0.274, + 0.418 + ], + "angle": 0, + "content": "3.1. Problem Definition" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.427, + 0.484, + 0.673 + ], + "angle": 0, + "content": "In continual segmentation, the model is tasked with an incremental learning challenge over \\(T\\) steps. At each step \\(t \\in \\{1, \\ldots, T\\}\\), the model is trained to recognize a unique set of classes \\(C^t\\), where the intersection across all sets from \\(i\\) to \\(T\\) is empty \\(\\bigcap_{t=1}^{T} C^t = \\emptyset\\) and their union forms the complete set of classes \\(\\bigcup_{t=1}^{T} C^t = \\mathcal{C}\\). In the current step \\(t\\), the training dataset \\(D_{train}^t\\) consists of image-label pairs \\((x^t, y^t)\\), where \\(x^t\\) represents an image and \\(y^t\\) its corresponding segmentation label. The labels \\(y^t\\) are available only for the classes \\(C^t\\) that need to be learned at this step, while labels for previously learned classes \\(C^{1:t-1}\\) and future classes \\(C^{t+1:T}\\) are inaccessible. After completing training at this step, the model must perform segmentation across all classes learned up to that point, \\(C^{1:t}\\), thereby preventing catastrophic forgetting of the old classes \\(C^{1:t-1}\\) while effectively acquiring new ones \\(C^t\\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.681, + 0.484, + 0.699 + ], + "angle": 0, + "content": "3.2. Revisiting Mask2Former and Pseudo-Labeling" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.484, + 0.902 + ], + "angle": 0, + "content": "The recent advanced universal segmentation network, Mask2Former [17], has favored the class incremental semantic and panoptic segmentation (CISS and CIPS) tasks for its innovative proposal-based structure. Mask2Former deviates from traditional pixel-wise classification pipelines by introducing a class-agnostic mask prediction and classification mechanism. Specifically, the backbone \\( \\pmb{f}_b \\) and the pixel decoder \\( \\pmb{f}_p \\) generate multi-scale features that interact with \\( N \\) learnable queries through multiple self-attention and cross-attention layers in the transformer decoder \\( \\pmb{f}_t \\). These \\( N \\) output queries \\( Q \\in \\mathbb{R}^{C_Q \\times N} \\) undergo further processing through two linear layers to form mask embeddings \\( \\mathcal{E}_{mask} \\in \\mathbb{R}^{D \\times N} \\) and class embeddings \\( \\mathcal{E}_{cls} \\in \\mathbb{R}^{(C + 1) \\times N} \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.198 + ], + "angle": 0, + "content": "The mask embeddings \\(\\mathcal{E}_{mask}\\) yield \\(N\\) mask predictions \\(M\\in \\mathbb{R}^{N\\times H\\times W}\\) via dot product between each mask embedding and per-pixel embeddings \\(\\mathcal{E}_{pixel}\\in \\mathbb{R}^{D\\times H\\times W}\\) from \\(f_{p}\\), while \\(\\mathcal{E}_{cls}\\) accounts for the classification predictions of these masks. Note that the additional channel in \\(\\mathcal{E}_{cls}\\) represents the no-object class, an auxiliary category utilized during training but excluded during inference." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.199, + 0.907, + 0.417 + ], + "angle": 0, + "content": "In the contexts of CISS and CIPS, where ground truth for old classes \\(C^{1:t-1}\\) are inaccessible, pseudo-labeling becomes crucial. PLOP [20] introduces this strategy by generating pixel-level pseudo ground truths for \\(C^{1:t-1}\\), adapted to the mask-based structure by CoMFormer [8], the first method based on Mask2Former for CISS and CIPS tasks. This adaptation involves weighting the mask predictions \\(M\\) by the corresponding maximum class embedding score after the softmax process over the \\(C+1\\) dimensions, i.e., \\(\\max_{c=0}^{|C^{0:t-1}|} (\\operatorname{softmax}_{c=0}^{|C^{0:t-1}|} (\\mathcal{E}_{cls}))\\) to form \\(M_{\\omega}\\), with associated categories \\(C_{\\omega} = \\arg \\max_{c=0}^{|C^{0:t-1}|} \\mathcal{E}_{cls} \\in \\mathbb{R}^{N}\\). The pseudo-labeling process then labels the pixels outside the union of \\(\\boldsymbol{y}^t\\) with the category \\(c\\) that achieves the maximum score across \\(M_{\\omega}\\) and \\(C_{w}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.52, + 0.424, + 0.906, + 0.5 + ], + "angle": 0, + "content": "\\[\n\\tilde {M} _ {c} (h, w) = \\left\\{ \\begin{array}{l l} 1 & \\text {i f} c = C _ {\\omega} [ \\arg \\max _ {n = 1} ^ {N} M _ {w} (n, h, w) ] \\\\ & \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\text {v a r e p s i l o n} \\\\ 0 & \\text {e l s e} \\end{array} \\right.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.5, + 0.906, + 0.545 + ], + "angle": 0, + "content": "where \\( n \\) indexes the \\( N \\) proposals, and \\( h, w \\) are the pixel coordinates. This pseudo-labeling strategy has become foundational for CISS and CIPS methods [8]." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.558, + 0.688, + 0.574 + ], + "angle": 0, + "content": "4. Proposed Method" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.583, + 0.906, + 0.675 + ], + "angle": 0, + "content": "In our proposed Conflict Mitigation via Branched Optimization (CoMBO), we introduce three distinct strategies: Half-Distillation-Half-Learning Strategy for targeted knowledge retention, Importance-Based Knowledge Distillation to prioritize crucial features, and Query Conflict Reduction for efficient class integration." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.682, + 0.748, + 0.698 + ], + "angle": 0, + "content": "4.1. Query Conflict Reduction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.704, + 0.906, + 0.797 + ], + "angle": 0, + "content": "As mentioned in Sec. 3.2, both class embeddings \\(\\mathcal{E}_{cls}\\) and mask predictions \\(M\\) are generated from the queries \\(Q\\). This interface becomes a critical juncture of conflicts between retaining knowledge of old classes and acquiring new classes manifest, which cannot break through the bottleneck of overall performance by simply finding the balance point." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.906, + 0.901 + ], + "angle": 0, + "content": "To deconstruct the crossroad efficiently, we design a Query Conflict Reduction (QCR) module \\( f_{QCR}(\\cdot) \\) as an additional adapter for new classes. As shown in Fig. 2, in most scenarios, queries from the previous layer \\( Q_{l-1} \\) have the same classification results as the queries from the current layer due to the similar distribution and the same clas" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "25669" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.102, + 0.092, + 0.462, + 0.312 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.322, + 0.484, + 0.392 + ], + "angle": 0, + "content": "Figure 2. Illustration of the Query Conflict Reduction (QCR) module. This module refines queries that predict incremental classes, allowing the processes of learning new features and retaining old features to occur separately. Note that the QCR module for previous incremental classes is frozen." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.405, + 0.484, + 0.541 + ], + "angle": 0, + "content": "sification layer, especially in the latter queries. Due to the shared classifier and mask generator following different layers of the transformer decoder, the classification results on the same position of queries from adjacent layers are approximately the same. Thus, we add the adapter for the queries from the penultimate layer \\( Q_{L-1} \\) with classification results of new classes \\( C^{2:t} \\), as shown in Fig. 2. To be more specific, the queries \\( Q_{L-1,\\tilde{c}} \\) are selected for the adapter of incremental class \\( \\tilde{c} \\in C^{2:t} \\) via:" + }, + { + "type": "equation", + "bbox": [ + 0.101, + 0.546, + 0.483, + 0.567 + ], + "angle": 0, + "content": "\\[\nQ _ {L - 1, \\tilde {c}} = \\left\\{Q _ {L - 1} (n), n \\in N \\wedge \\tilde {c} = \\arg \\max _ {c = 0} ^ {| C ^ {0: t} |} \\mathcal {E} _ {c l s} (n) \\right\\}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.573, + 0.484, + 0.664 + ], + "angle": 0, + "content": "where \\( Q_{L-1,\\tilde{c}} \\) could contain none, one, or some selected queries that enable recognizing different instances. For each new category, we utilize the corresponding QCR module \\( f_{QCR,\\tilde{c}}(\\cdot) \\) to refine the group of queries \\( Q_{L-1,\\tilde{c}} \\). Considering the efficiency and effectiveness, we introduce the low-rank two-layer adaptation as the QCR module:" + }, + { + "type": "equation", + "bbox": [ + 0.164, + 0.671, + 0.482, + 0.708 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\hat {Q} _ {L - 1, \\tilde {c}} = f _ {Q C R, \\tilde {c}} \\left(Q _ {L - 1, \\tilde {c}}\\right) \\tag {3} \\\\ = Q _ {L - 1, \\tilde {c}} W _ {1} W _ {2} + Q _ {L - 1, \\tilde {c}}, \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.718, + 0.484, + 0.825 + ], + "angle": 0, + "content": "where \\(\\hat{Q}_{L - 1,\\tilde{c}}\\) represent the adapted queries of class \\(\\tilde{c}\\), \\(W_{1} \\in \\mathbb{R}^{D \\times r}\\) and \\(W_{2} \\in \\mathbb{R}^{r \\times D}\\) denotes the weights for QCR. The QCR module separates the refined queries \\(\\hat{Q}_{L,\\tilde{c}}\\) from the original queries \\(Q_{L,\\tilde{c}}\\), enabling a bifurcate structure for simultaneously learning \\(C^t\\) with class-specific adaptation and keep memorizing the features of \\(Q_{L,\\tilde{c}}\\) as well as the logits of \\(C^{0:t - 1}\\)." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.833, + 0.439, + 0.85 + ], + "angle": 0, + "content": "4.2. Half-Distillation-Half-Learning Strategy" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Previous approaches based on Mask2Former usually separate learning and distillation processes on class embeddings \\(\\mathcal{E}_{cls}\\) due to its different learning strategies compared" + }, + { + "type": "image", + "bbox": [ + 0.529, + 0.092, + 0.885, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.306, + 0.907, + 0.362 + ], + "angle": 0, + "content": "Figure 3. Details of the Half-Distillation-Half-Learning strategy on \\(N\\) queries, including classification loss on matched queries and Kullback-Leibler Divergence loss on the other queries. The latter involves the logits of both old and current classes." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.385, + 0.906, + 0.461 + ], + "angle": 0, + "content": "to pixel-wise segmentation structure (e.g. DeepLabv3 [13]), mentioned in Sec. 3.2. This causes conflicts between distillation and learning, where the same logit could have contradictory optimization directions. Therefore, it is necessary to unify the optimization process towards a fixed target." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.463, + 0.907, + 0.728 + ], + "angle": 0, + "content": "Our Half-Distillation-Half-Learning (HDHL) strategy combines distillation and learning for \\(N\\) class embeddings \\(\\mathcal{E}_{cls}\\), as illustrated in Fig. 3. We follow the bipartite matching of Mask2Former and select embeddings \\(\\mathcal{E}_{cls}^{C^t}\\) matched to labels of current categories \\(C^t\\) and \\(\\mathcal{E}_{cls}^{C^{1:t-1}}\\) matched to pseudo labels of old categories \\(C^{1:t-1}\\). These embeddings, following the original optimization strategy, are optimized by cross-entropy loss \\(\\mathcal{L}_{cls}\\). The unmatched embeddings \\(\\mathcal{E}_{cls}^{\\emptyset^t}\\) are grouped under the no-object (no-\\(obj\\)) category. However, adhering to the original learning strategy leads to catastrophic forgetting. Besides, using the distillation strategy for old categories \\(C^{1:t-1}\\) using \\(\\mathcal{E}_{cls}^{\\emptyset^{t-1}}\\) ignores the learning of current categories \\(C^t\\). However, identifying regions that do not belong to \\(C^t\\) is just as important as learning those that do. Considering this, we design a strategy to simultaneously optimize the logits of \\(C^{1:t-1}\\) and \\(C^t\\) with a shared Kullback-Leibler Divergence \\(\\mathcal{L}_{kl}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.526, + 0.745, + 0.907, + 0.793 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {k l} = \\frac {1}{| \\mathcal {E} _ {c l s} ^ {\\varnothing^ {t}} |} \\sum_ {j = 1} ^ {| \\mathcal {E} _ {c l s} ^ {\\varnothing^ {t}} |} \\sum_ {c = 0} ^ {| \\boldsymbol {C} ^ {0: t} |} \\varphi_ {c l s} ^ {\\varnothing^ {t - 1}} (j, c) \\log \\frac {\\varphi_ {c l s} ^ {\\varnothing^ {t - 1}} (j , c)}{\\varphi_ {c l s} ^ {\\varnothing^ {t}} (j , c)}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.812, + 0.56, + 0.825 + ], + "angle": 0, + "content": "where" + }, + { + "type": "equation", + "bbox": [ + 0.564, + 0.842, + 0.905, + 0.905 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\varphi_ {c l s} ^ {\\emptyset^ {t}} (j, c) = \\mathcal {S} ^ {t} (j, c), \\\\ \\varphi_ {c l s} ^ {\\emptyset^ {t - 1}} (j, c) = \\left\\{ \\begin{array}{l l} \\mathcal {S} ^ {t - 1} (j, c) & \\text {i f} c \\in C ^ {0: t - 1}, \\\\ 0 & \\text {e l s e}, \\end{array} \\right. \\tag {5} \\\\ \\end{array}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "25670" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.093, + 0.12, + 0.105 + ], + "angle": 0, + "content": "and" + }, + { + "type": "equation", + "bbox": [ + 0.17, + 0.105, + 0.483, + 0.147 + ], + "angle": 0, + "content": "\\[\n\\mathcal {S} ^ {t} (j, c) = \\frac {\\exp \\left(\\mathcal {E} _ {c l s} ^ {\\varnothing^ {t}} (j , c)\\right)}{\\sum_ {\\hat {c} = 0} ^ {| \\mathcal {C} ^ {0 : t} |} \\exp \\left(\\mathcal {E} _ {c l s} ^ {\\varnothing^ {t}} (j , \\hat {c})\\right)} \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.154, + 0.313, + 0.167 + ], + "angle": 0, + "content": "represent the softmax activation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.169, + 0.484, + 0.307 + ], + "angle": 0, + "content": "The \\(\\mathcal{L}_{kl}\\) not only distill the previous class embedding \\(\\mathcal{E}_{cls}^{\\emptyset^{t-1}}\\) to the corresponding categories, including the no-object category, but decreasing the logits of current categories to figure out the objects apart from them. The unified HDHL strategy maintains the probability distribution of old logits seamlessly with the integration of new logits. By combining the original \\(\\mathcal{L}_{cls}\\) and the designed \\(\\mathcal{L}_{kl}\\) as \\(\\mathcal{L}_{DL}\\), the optimization process covers all the class embeddings without contradictory targets:" + }, + { + "type": "equation", + "bbox": [ + 0.183, + 0.321, + 0.483, + 0.339 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {D L} = \\lambda_ {c l s} \\cdot \\mathcal {L} _ {c l s} + \\lambda_ {k l} \\cdot \\mathcal {L} _ {k l}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.351, + 0.482, + 0.366 + ], + "angle": 0, + "content": "where \\(\\lambda_{cls}\\) and \\(\\lambda_{kl}\\) control the balance between two losses." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.378, + 0.449, + 0.394 + ], + "angle": 0, + "content": "4.3. Importance-Based Knowledge Distillation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.402, + 0.484, + 0.508 + ], + "angle": 0, + "content": "Previous methods mitigate catastrophic forgetting by applying feature distillation, which may hinder the acquisition of new knowledge. To address this, we propose an Importance-Based Knowledge Distillation (IKD) method, in which stronger distillation is applied to features critical for old classes, while less emphasis is placed on less pertinent features, thereby balancing stability with plasticity." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.509, + 0.484, + 0.631 + ], + "angle": 0, + "content": "The importance measurement requires the cost matrix \\(\\mathcal{A} \\in \\mathbb{R}^{N \\times S}\\), adopted in the bipartite matching process, where \\(S\\) is the number of image labels. Values in \\(\\mathcal{A}\\) indicate the relationship between queries and classes. For each image, the transformer decoder generates \\(N\\) predicted segments, denoted as \\(\\mathcal{Z} = \\left\\{(\\mathcal{E}_{cls}(n), M(n))\\right\\}_{n=1}^{N}\\). The cost matrix \\(\\mathcal{A}\\) is then computed based on the output segments \\(\\mathcal{Z}\\) and the image labels \\(\\boldsymbol{y}^t = \\{(c_s, m_s)\\}_{s=1}^S\\):" + }, + { + "type": "equation", + "bbox": [ + 0.09, + 0.643, + 0.496, + 0.671 + ], + "angle": 0, + "content": "\\[\n\\mathcal {A} (n, s) = - \\lambda_ {c l s} \\cdot \\varphi_ {c l s} \\left(n, c _ {s}\\right) + \\lambda_ {\\text {m a s k}} \\cdot \\mathcal {L} _ {\\text {m a s k}} \\left(\\mathcal {E} _ {\\text {m a s k}} (n), m _ {s}\\right), \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.674, + 0.484, + 0.719 + ], + "angle": 0, + "content": "where \\(\\varphi_{cls}\\) is the classification output obtained by applying softmax to \\(\\mathcal{E}_{cls}\\), and \\(\\lambda_{cls}\\) and \\(\\lambda_{mask}\\) represent the corresponding hyperparameters." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.484, + 0.902 + ], + "angle": 0, + "content": "After each training stage, we compute the cost matrix \\(\\mathcal{A}\\) across every image in the training set \\(D_{train}^{t}\\) of the current incremental step \\(t\\). The minimum costs among the classes for \\(N\\) segments are accumulated in the corresponding position of buffer \\(B^{t} \\in \\mathbb{R}^{N}\\), where segments with lower minimum cost values are more likely to recognize and cover the regions of a class \\(C^{t}\\). By normalizing and reversing the values of \\(B^{t}\\), we derive the importance matrix \\(I_{C^{t}}^{t} \\in \\mathbb{R}^{N}\\) of current classes \\(C^{t}\\). The importance matrix \\(I^{t+1} \\in \\mathbb{R}^{N}\\) is further estimated via the weighted accumulation of current importance matrix \\(I^{t} \\in \\mathbb{R}^{N}\\) and \\(I_{C^{t}}^{t}\\), where the weights refer to the \\(|C^{1:t-1}|\\) and \\(|C^{t}|\\). The generated \\(I^{t+1}\\) will be" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.907, + 0.122 + ], + "angle": 0, + "content": "utilized in the next incremental training phase for weighted distillation of the queries \\( Q^t \\), which is computed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.56, + 0.133, + 0.907, + 0.175 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {I K D} = \\frac {1}{N} \\sum_ {n = 1} ^ {N} I ^ {t} (n) \\cdot \\| Q ^ {t} (n), Q ^ {t - 1} (n) \\| _ {2} ^ {2}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.185, + 0.907, + 0.263 + ], + "angle": 0, + "content": "where \\( \\| \\cdot \\|_2 \\) denotes the Euclidean distance, \\( Q^t(n) \\) and \\( Q^{t-1}(n) \\) denotes the transformer decoder features from the current model \\( f^t(\\cdot) \\) and the old model \\( f^{t-1}(\\cdot) \\), respectively. Algorithm 1 presents the details of the importance estimation procedure after training \\( f^t(\\cdot) \\)." + }, + { + "type": "code_caption", + "bbox": [ + 0.516, + 0.277, + 0.806, + 0.292 + ], + "angle": 0, + "content": "Algorithm 1 Importance Matrix Estimation" + }, + { + "type": "algorithm", + "bbox": [ + 0.522, + 0.296, + 0.907, + 0.509 + ], + "angle": 0, + "content": "1: Input: current model \\( f^{t}(\\cdot) \\), current dataset \\( D_{train}^{t} \\), current importance matrix \\( I^{t} \\) \n2: Initialization: \\( I^{t+1} \\gets 0 \\), \\( B^{t} \\gets 0 \\) \n3: for all \\( (\\boldsymbol{x}^{t}, \\boldsymbol{y}^{t}) \\in D_{train}^{t} \\) do \n4: Compute \\( A \\) following Eq. (8) \n5: Update the buffer: \n6: \\( B^{t} \\gets B^{t} + \\min(A, \\text{axis} = 1) \\) \n7: end for \n8: for \\( n \\in \\{1, 2, \\dots, N\\} \\) do \n9: \\( I_{C^{t}}^{t}(n) \\gets 1 - \\frac{B^{t}(n) - \\min(B^{t})}{\\max(B^{t}) - \\min(B^{t})} \\) \n10: \\( I^{t+1}(n) \\gets \\frac{|C^{1:t-1}|}{|C^{1:t}|} \\cdot I^{t}(n) + \\frac{|C^{t}|}{|C^{1:t}|} \\cdot I_{C^{t}}^{t}(n) \\) \n11: end for \n12: Output: Importance matrix \\( I^{t+1} \\) for step \\( t + 1 \\)" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.535, + 0.698, + 0.551 + ], + "angle": 0, + "content": "4.4. Objective Function" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.558, + 0.884, + 0.573 + ], + "angle": 0, + "content": "Overall, the objective function of CoMBO is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.616, + 0.586, + 0.907, + 0.601 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {H} = \\mathcal {L} _ {D L} + \\lambda_ {I K D} \\mathcal {L} _ {I K D}. \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.613, + 0.906, + 0.659 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_{DL}\\) represents the Half-Distillation-Half-Learning loss, and \\(\\mathcal{L}_{IKD}\\) means the Importance-Based Knowledge Distillation loss, with \\(\\lambda_{IKD}\\) as its weighting factor." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.672, + 0.719, + 0.688 + ], + "angle": 0, + "content": "5. Experimental Results" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.697, + 0.704, + 0.713 + ], + "angle": 0, + "content": "5.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.906, + 0.84 + ], + "angle": 0, + "content": "Datasets and Evaluation Metrics. We follow the experimental settings of previous works [8, 11, 26] and evaluate our approach on the ADE20K [52] dataset. The ADE20K dataset is specifically designed to support both panoptic and semantic segmentation tasks. ADE20K contains 150 classes, including 100 thing classes and 50 stuff classes. The dataset is composed of 20,210 images for training and 2,000 images for validation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.906, + 0.904 + ], + "angle": 0, + "content": "For Class Incremental Semantic Segmentation (CISS), we adopt the mean Intersection over Union (mIoU) for evaluation. The Intersection over Union (IoU) is calculated as \\(\\mathrm{IoU} = \\frac{TP}{TP + FP + FN}\\), where \\(TP\\), \\(FP\\), and \\(FN\\) denote the" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "25671" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.089, + 0.904, + 0.256 + ], + "angle": 0, + "content": "
Method100-50 (2 steps)100-10 (6 steps)100-5 (11 steps)50-50 (3 steps)
1-100101-150all1-100101-150all1-100101-150all1-5051-150all
FT0.01.30.40.02.91.00.025.88.60.012.08.1
MiB [5]CVPR2035.119.329.827.110.021.424.06.518.142.415.524.4
PLOP [20]CVPR2140.222.434.330.517.526.128.115.724.045.818.727.7
CoMFormer [8]CVPR2341.127.736.736.017.129.734.415.928.245.019.327.9
ECLIPSE [26]CVPR2441.723.535.641.418.833.941.116.632.946.020.729.2
BalConpas [11]ECCV2442.825.737.140.722.834.736.120.330.851.226.534.7
CoMBO Ours43.925.637.840.825.235.636.120.530.950.728.235.7
Joint43.830.939.543.830.939.543.830.939.550.733.939.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.266, + 0.905, + 0.296 + ], + "angle": 0, + "content": "Table 1. Quantitative comparison under Class Incremental Panoptic Segmentation with state-of-the-art exemplar-free methods on ADE20K in PQ. Scores of novel classes and all classes in bold are the best while underlined are the second best." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.321, + 0.482, + 0.471 + ], + "angle": 0, + "content": "number of true-positive, false-positive, and false-negative pixels, respectively. The mIoU metric averages the IoU across all classes for a more comprehensive evaluation. For Class Incremental Panoptic Segmentation (CIPS), following previous work [8], we employ Panoptic Quality (PQ) as the evaluation metric. PQ is defined as the product of Recognition Quality (RQ) and Segmentation Quality (SQ). To measure incremental learning capacity, we compute the corresponding metrics for the initial classes \\( C^1 \\), incremental classes \\( C^{2:T} \\), and the aggregate of all classes \\( C^{1:T} \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.49, + 0.483, + 0.642 + ], + "angle": 0, + "content": "Protocols and Implementation Details. We follow previous incremental protocols and define scenarios as \\( N_{ini} - N_{inc} \\), where \\( N_{ini} \\) represents the number of initial classes, and \\( N_{inc} \\) denotes the number of new classes introduced at each incremental step. For example, in the 100-10 scenario, the training begins with 100 classes, followed by the addition of 10 new classes per incremental step, without access to annotations of old classes. For both CIPS and CISS, we do evaluations on the following scenarios: 100-10 (6 steps), 100-50 (2 steps), 100-5 (11 steps) and 50-50 (3 steps)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.642, + 0.483, + 0.822 + ], + "angle": 0, + "content": "Our approach is based on the Mask2Former structure [17]. Following previous approaches to CISS and CIPS, we utilize ResNet-50 [24] as the backbone for CIPS and ResNet-101 for CISS, with both pre-trained on ImageNet [19]. The input resolution of the images is \\(640 \\times 640\\) with an all-time batch size of 8. We follow the training hyperparameter of Mask2Former in the initial step, with a learning rate of \\(10^{-4}\\) and iterations of 160,000. During the incremental steps, we set 1,000 iterations per class with a learning rate of \\(5 \\times 10^{-5}\\). The coefficients \\(r\\), \\(\\lambda_{cls}\\), \\(\\lambda_{kl}\\), \\(\\lambda_{IKD}\\) are respectively set to 16, 2, 5, 3. All experiments were conducted on the NVIDIA RTX 4090." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.833, + 0.285, + 0.849 + ], + "angle": 0, + "content": "5.2.Quantitative Results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Comparisons in Class Incremental Panoptic Segmentation (CIPS). We evaluated our approach against state-of-the-art exemplar-free methods on the ADE20K dataset" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.321, + 0.908, + 0.563 + ], + "angle": 0, + "content": "within the CIPS framework, as detailed in Tab. 1. The performance of our CoMBO surpasses other methods in most of the scenarios, particularly in the more challenging 100-10 scenario, where it achieves a remarkable \\(35.6\\%\\) PQ with at least \\(3.4\\%\\) of advantage on incremental classes compared to the previous state-of-the-art methods. This notable performance gain demonstrates the effectiveness of our CoMBO in reducing conflicts. Although our approach still underperforms the state-of-the-art approach ECLIPSE [26] in the 100-5 scenario, primarily due to its strong memorization ability from freezing parameters, our method achieves advanced performance for the new classes. Moreover, in scenarios with fewer initial classes, i.e., 50-50 scenario, we maintain our advantage in learning new classes with a PQ of \\(28.2\\%\\) for incremental classes and \\(35.7\\%\\) for overall performance, widening the gap with strong distillation methods." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.582, + 0.909, + 0.794 + ], + "angle": 0, + "content": "Comparisons in Class Incremental Semantic Segmentation (CISS). We further extended our evaluation to the semantic segmentation benchmark, comparing our approach with previous methods on the ADE20K dataset, as detailed in Tab. 2. Our approach consistently outperforms previous methods in all tested scenarios. Notably, in the 100-10 scenario, our approach achieves a mIoU of \\(41.1\\%\\), surpassing the previous state-of-the-art by at least \\(3.5\\%\\) for incremental classes, demonstrating its effectiveness in mitigating conflicts. Moreover, in the most challenging 100-5 long-term incremental scenario, our approach not only improves performance on old classes by \\(2.5\\%\\) but also achieves a significant \\(5.4\\%\\) increase in mIoU for new classes, ensuring both model stability and adaptability." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.808, + 0.665, + 0.825 + ], + "angle": 0, + "content": "6. Ablation Study" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.832, + 0.719, + 0.848 + ], + "angle": 0, + "content": "6.1. Component Ablations" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.905, + 0.901 + ], + "angle": 0, + "content": "We analyze the key components of our proposed framework, which includes the Half-Distillation-Half-Learning Strategy, Importance-based Knowledge Distillation, and the" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "25672" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.089, + 0.904, + 0.269 + ], + "angle": 0, + "content": "
Method100-50 (2 steps)100-10 (6 steps)100-5 (11 steps)50-50 (3 steps)
1-100101-150all1-100101-150all1-100101-150all1-5051-150all
FT0.026.78.90.02.30.80.01.10.30.01.71.1
MiB [5]CVPR2037.024.132.623.510.626.621.06.116.145.621.029.3
PLOP [20]CVPR2144.226.238.234.815.928.539.513.630.954.930.238.4
CoMFormer [8]CVPR2344.726.238.440.615.632.339.513.630.949.226.634.1
CoMasTRe [23]CVPR2445.726.039.242.318.434.440.815.832.649.826.634.5
ECLIPSE [26]CVPR2445.021.737.143.417.434.643.316.334.2---
BalConpas [11]ECCV2449.930.143.347.324.238.642.117.233.855.833.340.8
CoMBO Ours50.234.444.947.827.741.144.622.637.355.336.943.0
Joint51.740.247.851.740.247.851.740.247.856.643.547.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.279, + 0.905, + 0.308 + ], + "angle": 0, + "content": "Table 2. Quantitative comparison under Class Incremental Semantic Segmentation with state-of-the-art exemplar-free methods on ADE20K in mIoU. Scores of novel classes and all classes in bold are the best while underlined are the second best." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.334, + 0.482, + 0.442 + ], + "angle": 0, + "content": "Query Conflict Reducing module. Our experiments focus on the 100-10 scenario in the CIPS, as it provides a moderate number of steps to examine the ability of both acquisition and retention. We evaluate various combinations of these components separately and present the results in Tab. 3. The baseline method utilizes the vanilla loss with pseudo-labeling." + }, + { + "type": "table", + "bbox": [ + 0.111, + 0.455, + 0.465, + 0.577 + ], + "angle": 0, + "content": "
HDHLIKDQCR100-10 (6 steps)
1-100101-150all
36.323.832.2
✓39.424.234.3
✓✓40.024.634.9
✓✓38.825.434.4
✓✓✓40.825.235.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.588, + 0.483, + 0.631 + ], + "angle": 0, + "content": "Table 3. Ablation study of the main components on the 100-10 task of CIPS. The baseline in the \\(1^{\\text{st}}\\) row employs vanilla losses with pseudo-labeling and excludes the QCR module." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.644, + 0.483, + 0.9 + ], + "angle": 0, + "content": "Under identical experimental conditions, the HDHL Strategy significantly enhances the overall PQ by \\(2.1\\%\\), demonstrating its ability to unify logits of new classes into the previous logits distribution without contradictory losses. The implementation of the IKD mechanism has a remarkable \\(3.7\\%\\) increase in the PQ for old classes, underscoring its effectiveness in boosting the ability to retain the knowledge of old classes. Furthermore, when combined with the QCR module for branched optimization, CoMBO further boosts both the performance of old and new classes, achieving \\(0.8\\%\\) and \\(0.6\\%\\) gains in PQ, respectively, surpassing previous methods that struggle to balance acquisition and retention. Consequently, CoMBO achieves a \\(3.4\\%\\) PQ improvement over the baseline. The above results demonstrate the effectiveness of the proposed CoMBO and its related modules, showing a noticeable improvement in reducing the conflict on model structures and losses." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.333, + 0.731, + 0.35 + ], + "angle": 0, + "content": "6.2. Ablation Study of QCR" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.356, + 0.907, + 0.508 + ], + "angle": 0, + "content": "We validate the effectiveness of QCR with different settings of \\( r \\) shown in Fig. 5. Under the estimation of performance enhancement and additional parameters, our QCR module achieves the best result when \\( r = 16 \\), where the quantity of additional parameters comes to \\( 8.2\\mathrm{K} \\) per class. Considering the total parameter of Mask2Former [17] is \\( 44.9\\mathrm{M} \\) with ResNet-50, the additional parameters only reach up to \\( 2\\% \\) of original parameters in total, which is 5 times less than per class additional parameters of ECLIPSE [26] while having better performance on new classes." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.515, + 0.813, + 0.532 + ], + "angle": 0, + "content": "6.3. Ablation Study of HDHL Strategy" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.538, + 0.907, + 0.795 + ], + "angle": 0, + "content": "To validate the effectiveness of our Half-Distillation-Half-Learning strategy, we conduct a series of experiments among other strategies, as shown in Tab. 4. The \\(1^{\\mathrm{st}}\\) row with \\(\\mathcal{L}_{cls}\\) on all classification embeddings represents the vanilla loss with pseudo-labeling, which is adopted in [8]. The \\(2^{\\mathrm{nd}}\\) row utilizes an auxiliary distillation on the embeddings following [23]. Compared to these strategies with tough supervision on unmatched embeddings \\(\\mathcal{E}_{cls}^{t,\\emptyset}\\) and mostly conflicting losses, our HDHL strategy with soft and concentrated supervision has a better PQ of 34.3. Besides, the HDHL design of \\(\\mathcal{L}_{kl}\\) ensures the integrity while optimizing the unmatched embeddings \\(\\mathcal{E}_{cls}^{t,\\emptyset}\\), with improvement of \\(2.1\\%\\) compared to distillation-only strategy in the \\(3^{\\mathrm{rd}}\\) row and \\(0.8\\%\\) compared to the simple combination of distillation and \\(\\mathcal{L}_{cls}\\). The above results show that our HDHL strategy outperforms other learning or distillation strategies on the class embeddings." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.802, + 0.707, + 0.818 + ], + "angle": 0, + "content": "6.4. Qualitative Analysis" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.906, + 0.903 + ], + "angle": 0, + "content": "We conduct the qualitative analysis (shown in Fig. 4) by comparing our proposed CoMBO to the Baseline and Baseline+HDHL \\((1^{\\mathrm{st}}\\) and \\(2^{\\mathrm{nd}}\\) rows in Tab. 3). In the \\(1^{\\mathrm{st}}\\), \\(2^{\\mathrm{nd}}\\), and \\(5^{\\mathrm{th}}\\) columns of Fig. 4, our CoMBO can indicate the incremental classes, i.e., screen door, fountain, and ottoman," + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "25673" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.123, + 0.094, + 0.871, + 0.442 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.155, + 0.45, + 0.841, + 0.466 + ], + "angle": 0, + "content": "Figure 4. Qualitative results of CoMBO comparing to Baseline, Baseline+HDHL on 100-10 CIPS task of ADE20K." + }, + { + "type": "image", + "bbox": [ + 0.101, + 0.485, + 0.472, + 0.661 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.663, + 0.483, + 0.692 + ], + "angle": 0, + "content": "Figure 5. Ablation study of \\( r \\) in QCR module, including PQ performance (left) and number of parameters (right)." + }, + { + "type": "table", + "bbox": [ + 0.124, + 0.697, + 0.454, + 0.82 + ], + "angle": 0, + "content": "
LclsLklPQ
Ectls & Ectls1:t-1EctlsC0:t-1Ct
✓✓32.2
✓✓✓32.6
✓✓32.2
✓✓✓33.5
✓✓✓34.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.831, + 0.485, + 0.875 + ], + "angle": 0, + "content": "Table 4. Ablation study of the HDHL strategy on task 100-10 of CIPS. The \\(1^{\\text{st}}\\) row represents the vanilla loss with pseudo-labeling. The last row represents our \\(\\mathcal{L}_{DL}\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.488, + 0.908, + 0.596 + ], + "angle": 0, + "content": "while other methods fail to recognize all these objects. In the \\(3^{\\mathrm{rd}}\\) column, our method successfully remembers the car in the left instead of covering the object with the prediction of the incremental class van. Furthermore, our method allows more precise mask prediction of new class Traffic Light in the \\(4^{\\mathrm{th}}\\) column, demonstrating the effectiveness of refinements on the queries corresponding to new classes." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.607, + 0.642, + 0.623 + ], + "angle": 0, + "content": "7. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.632, + 0.909, + 0.875 + ], + "angle": 0, + "content": "In this paper, we present CoMBO, a novel Class Incremental Segmentation (CIS) method for mitigating the conflict between acquisition and retention losses. This conflict arises from competing goals of efficiently acquiring knowledge about new classes while preserving knowledge of previously learned ones. Our QCR module branches the conflicting optimization targets via lightweight class-specific adaptation on queries, enabling the coexistence of learning and distillation on separated queries. The HDHL strategy and IKD further reduce the contradictory optimization on classification logits and queries with unified targets, respectively, thereby enhancing the overall model performance. Extensive experimental validation on the ADE20K dataset underscores the superiority of our approach, demonstrating the effectiveness of each component in the CIS task, particularly excelling in performance on new classes." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "25674" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.251, + 0.108 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.484, + 0.207 + ], + "angle": 0, + "content": "This work was supported by the National Key R&D Program of China (Grant 2022YFC3310200), the National Natural Science Foundation of China (Grant 62472033, 92470203), the Beijing Natural Science Foundation (Grant L242022), the Royal Society grants (SIF\\R1\\231009, IES\\R3\\223050) and an Amazon Research Award." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.218, + 0.188, + 0.234 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.244, + 0.483, + 0.285 + ], + "angle": 0, + "content": "[1] Rahaf Aljundi, Min Lin, Baptiste Goujaud, and Yoshua Bengio. Gradient based sample selection for online continual learning. Adv. Neural Inf. Process. Syst., 32, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.286, + 0.483, + 0.342 + ], + "angle": 0, + "content": "[2] Vijay Badrinarayanan, Alex Kendall, and Roberto Cipolla. Segnet: A deep convolutional encoder-decoder architecture for image segmentation. IEEE Trans. on Pattern Anal. Mach. Intell., 39(12):2481-2495, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.343, + 0.483, + 0.397 + ], + "angle": 0, + "content": "[3] Donghyeon Baek, Youngmin Oh, Sanghoon Lee, Junghyup Lee, and Bumsub Ham. Decomposed knowledge distillation for class-incremental semantic segmentation. Adv. Neural Inf. Process. Syst., 35:10380-10392, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.398, + 0.483, + 0.439 + ], + "angle": 0, + "content": "[4] Zalán Borsos, Mojmir Mutny, and Andreas Krause. Coresets via bilevel optimization for continual learning and streaming. Adv. Neural Inf. Process. Syst., 33:14879-14890, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.44, + 0.483, + 0.508 + ], + "angle": 0, + "content": "[5] Fabio Cermelli, Massimiliano Mancini, Samuel Rota Bulo, Elisa Ricci, and Barbara Caputo. Modeling the background for incremental learning in semantic segmentation. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 9233-9242, 2020. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.51, + 0.483, + 0.565 + ], + "angle": 0, + "content": "[6] Fabio Cermelli, Dario Fontanel, Antonio Tavera, Marco Ciccone, and Barbara Caputo. Incremental learning in semantic segmentation from image labels. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 4371-4381, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.567, + 0.483, + 0.621 + ], + "angle": 0, + "content": "[7] Fabio Cermelli, Antonino Geraci, Dario Fontanel, and Barbara Caputo. Modeling missing annotations for incremental learning in object detection. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 3700-3710, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.622, + 0.483, + 0.677 + ], + "angle": 0, + "content": "[8] Fabio Cermelli, Matthieu Cord, and Arthur Douillard. Comformer: Continual learning in semantic and panoptic segmentation. In Proc. IEEE Conf. Comput. Vis. Pattern Recognition., pages 3010-3020, 2023. 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.679, + 0.483, + 0.733 + ], + "angle": 0, + "content": "[9] Sungmin Cha, YoungJoon Yoo, Taesup Moon, et al. Ssul: Semantic segmentation with unknown label for exemplar-based class-incremental learning. Adv. Neural Inf. Process. Syst., 34:10919-10930, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.734, + 0.483, + 0.803 + ], + "angle": 0, + "content": "[10] Jinpeng Chen, Runmin Cong, Yuxuan Luo, Horace Ip, and Sam Kwong. Saving 100x storage: prototype replay for reconstructing training sample distribution in class incremental semantic segmentation. Adv. Neural Inf. Process. Syst., 36, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.804, + 0.483, + 0.859 + ], + "angle": 0, + "content": "[11] Jinpeng Chen, Runmin Cong, Yuxuan Luo, Horace Ho Shing Ip, and Sam Kwong. Strike a balance in continual panoptic segmentation. In Proc. Eur. Conf. Comput. Vis., pages 126-142. Springer, 2025. 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.86, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[12] Liang-Chieh Chen. Semantic image segmentation with deep convolutional nets and fully connected crfs. arXiv preprint arXiv:1412.7062, 2014. 2" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.244, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.133 + ], + "angle": 0, + "content": "[13] L. C. Chen. Rethinking atrous convolution for semantic image segmentation. arXiv preprint:1706.05587, 2017. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.135, + 0.906, + 0.189 + ], + "angle": 0, + "content": "[14] Liang-Chieh Chen, George Papandreou, and et al. Deep convolutional nets, atrous convolution, and fully connected crfs. IEEE Trans. on Pattern Anal. Mach. Intell., 40(4):834-848, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.191, + 0.906, + 0.245 + ], + "angle": 0, + "content": "[15] Liang-Chieh Chen, Yukun Zhu, George Papandreou, Florian Schroff, and Hartwig Adam. Encoder-decoder with atrous separable convolution for semantic image segmentation. In Proc. Eur. Conf. Comput. Vis., pages 801-818, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.247, + 0.906, + 0.3 + ], + "angle": 0, + "content": "[16] Bowen Cheng, Alex Schwing, and Alexander Kirillov. Per-pixel classification is not all you need for semantic segmentation. Adv. Neural Inf. Process. Syst., 34:17864-17875, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.302, + 0.906, + 0.37 + ], + "angle": 0, + "content": "[17] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 1290–1299, 2022. 1, 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.372, + 0.906, + 0.427 + ], + "angle": 0, + "content": "[18] Wei Cong, Yang Cong, Yuyang Liu, and Gan Sun. Cs2k: Class-specific and class-shared knowledge guidance for incremental semantic segmentation. In Proc. Eur. Conf. Comput. Vis., pages 244-261, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.428, + 0.906, + 0.482 + ], + "angle": 0, + "content": "[19] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 248–255, 2009. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.484, + 0.906, + 0.538 + ], + "angle": 0, + "content": "[20] Arthur Douillard, Yifu Chen, Arnaud Dapogny, and Matthieu Cord. Plop: Learning without forgetting for continual semantic segmentation. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 4040-4050, 2021. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.54, + 0.906, + 0.594 + ], + "angle": 0, + "content": "[21] Arthur Douillard, Yifu Chen, Arnaud Dapogny, and Matthieu Cord. Tackling catastrophic forgetting and background shift in continual semantic segmentation. arXiv preprint arXiv:2106.15287, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.596, + 0.906, + 0.634 + ], + "angle": 0, + "content": "[22] Robert M French. Catastrophic forgetting in connectionist networks. Trends in cognitive sciences, 3(4):128-135, 1999. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.637, + 0.906, + 0.692 + ], + "angle": 0, + "content": "[23] Yizheng Gong, Siyue Yu, Xiaoyang Wang, and Jimin Xiao. Continual segmentation with disentangled objectness learning and class recognition. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 3848-3857, 2024. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.693, + 0.906, + 0.747 + ], + "angle": 0, + "content": "[24] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 770-778, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.749, + 0.906, + 0.816 + ], + "angle": 0, + "content": "[25] M. G. Z. A. Khan, M. F. Naeem, Luc Van Gool, Didier Stricker, Federico Tombari, and Muhammad Zeshan Afzal. Introducing language guidance in prompt-based continual learning. In Proc. IEEE Int. Conf. Comput. Vis., pages 11463-11473, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.818, + 0.906, + 0.873 + ], + "angle": 0, + "content": "[26] Beomyoung Kim, Joonsang Yu, and Sung Ju Hwang. Eclipse: Efficient continual learning in panoptic segmentation with visual prompt tuning. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 3346-3356, 2024. 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[27] Alexander Kirillov, Kaiming He, Ross Girshick, Carsten Rother, and Piotr Dollar. Panoptic segmentation. In Proc." + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "25675" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.12 + ], + "angle": 0, + "content": "IEEE Conf. Comput. Vis. Pattern Recognit., pages 9404-9413, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.122, + 0.483, + 0.162 + ], + "angle": 0, + "content": "[28] Zhizhong Li and Derek Hoiem. Learning without forgetting. IEEE Trans. on Pattern Anal. Mach. Intell., 40(12):2935-2947, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.163, + 0.482, + 0.219 + ], + "angle": 0, + "content": "[29] Guosheng Lin, Anton Milan, Chunhua Shen, and Ian Reid. Refinenet: Multi-path refinement networks for high-resolution semantic segmentation. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 1925-1934, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.219, + 0.482, + 0.273 + ], + "angle": 0, + "content": "[30] Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 3431-3440, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.274, + 0.482, + 0.316 + ], + "angle": 0, + "content": "[31] David Lopez-Paz and Marc'Aurelio Ranzato. Gradient episodic memory for continual learning. Adv. Neural Inf. Process. Syst., 30, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.316, + 0.482, + 0.372 + ], + "angle": 0, + "content": "[32] Andrea Maracani, Umberto Michieli, Marco Toldo, and Pietro Zanuttigh. Recall: Replay-based continual learning in semantic segmentation. In Proc. IEEE Int. Conf. Comput. Vis., pages 7026-7035, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.372, + 0.482, + 0.428 + ], + "angle": 0, + "content": "[33] Umberto Michieli and Pietro Zanuttigh. Continual semantic segmentation via repulsion-attraction of sparse and disentangled latent representations. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 1114-1124, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.428, + 0.482, + 0.483 + ], + "angle": 0, + "content": "[34] Umberto Michieli and Pietro Zanuttigh. Continual semantic segmentation via repulsion-attraction of sparse and disentangled latent representations. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 1114-1124, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.484, + 0.482, + 0.538 + ], + "angle": 0, + "content": "[35] Youngmin Oh, Donghyeon Baek, and Bumsub Ham. Alife: Adaptive logit regularizer and feature replay for incremental semantic segmentation. Adv. Neural Inf. Process. Syst., 35: 14516-14528, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.539, + 0.482, + 0.607 + ], + "angle": 0, + "content": "[36] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Proc. International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 234-241, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.609, + 0.482, + 0.678 + ], + "angle": 0, + "content": "[37] Chao Shang, Hongliang Li, Fanman Meng, Qingbo Wu, Heqian Qiu, and Lanxiao Wang. Incrementer: Transformer for class-incremental semantic segmentation with knowledge distillation focusing on old class. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 7214-7224, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.679, + 0.482, + 0.734 + ], + "angle": 0, + "content": "[38] Robin Strudel, Ricardo Garcia, Ivan Laptev, and Cordelia Schmid. Segmenter: Transformer for semantic segmentation. In Proc. IEEE Int. Conf. Comput. Vis., pages 7262-7272, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.735, + 0.482, + 0.762 + ], + "angle": 0, + "content": "[39] Sebastian Thrun. Lifelong learning algorithms. In Learning to learn, pages 181-209. 1998. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.762, + 0.482, + 0.831 + ], + "angle": 0, + "content": "[40] Zifeng Wang, Zizhao Zhang, Chen-Yu Lee, Han Zhang, Ruoxi Sun, Xiaoqi Ren, Guolong Su, Vincent Perot, Jennifer Dy, and Tomas Pfister. Learning to prompt for continual learning. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 139-149, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[41] Yunchao Wei, Jiashi Feng, Xiaodan Liang, Ming-Ming Cheng, Yao Zhao, and Shuicheng Yan. Object region mining with adversarial erasing: A simple classification to semantic segmentation approach. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 1568-1576, 2017. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.162 + ], + "angle": 0, + "content": "[42] Jia-Wen Xiao, Chang-Bin Zhang, Jiekang Feng, Xialei Liu, Joost van de Weijer, and Ming-Ming Cheng. Endpoints weight fusion for class incremental semantic segmentation. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 7204-7213, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.163, + 0.905, + 0.219 + ], + "angle": 0, + "content": "[43] Shipeng Yan, Jiangwei Xie, and Xuming He. Der: Dynamically expandable representation for class incremental learning. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 3014-3023, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.22, + 0.905, + 0.289 + ], + "angle": 0, + "content": "[44] Chaohui Yu, Qiang Zhou, Jingliang Li, Jianlong Yuan, Zhibin Wang, and Fan Wang. Foundation model drives weakly incremental learning for semantic segmentation. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 23685-23694, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.29, + 0.905, + 0.347 + ], + "angle": 0, + "content": "[45] Qihang Yu, Huiyu Wang, Siyuan Qiao, Maxwell Collins, Yukun Zhu, Hartwig Adam, Alan Yuille, and Liang-Chieh Chen. k-means mask transformer. In Proc. Eur. Conf. Comput. Vis., pages 288-307, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.347, + 0.905, + 0.401 + ], + "angle": 0, + "content": "[46] Anqi Zhang and Guangyu Gao. Background adaptation with residual modeling for exemplar-free class-incremental semantic segmentation. In Proc. Eur. Conf. Comput. Vis., 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.403, + 0.905, + 0.472 + ], + "angle": 0, + "content": "[47] Chang-Bin Zhang, Jia-Wen Xiao, Xialei Liu, Ying-Cong Chen, and Ming-Ming Cheng. Representation compensation networks for continual semantic segmentation. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 7053-7064, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.473, + 0.905, + 0.53 + ], + "angle": 0, + "content": "[48] Zekang Zhang, Guangyu Gao, Zhiyuan Fang, Jianbo Jiao, and Yunchao Wei. Mining unseen classes via regional objectness: A simple baseline for incremental segmentation. Adv. Neural Inf. Process. Syst., 35:24340-24353, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.53, + 0.905, + 0.587 + ], + "angle": 0, + "content": "[49] Zekang Zhang, Guangyu Gao, Jianbo Jiao, Chi Harold Liu, and Yunchao Wei. Coinseg: Contrast inter-and intra-class representations for incremental segmentation. In Proc. IEEE Int. Conf. Comput. Vis., pages 843-853, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.587, + 0.905, + 0.643 + ], + "angle": 0, + "content": "[50] Hengshuang Zhao, Jianping Shi, Xiaojuan Qi, Xiaogang Wang, and Jiaya Jia. Pyramid scene parsing network. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 2881-2890, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.643, + 0.905, + 0.698 + ], + "angle": 0, + "content": "[51] Hanbin Zhao, Fengyu Yang, Xinghe Fu, and Xi Li. Rbc: Rectifying the biased context in continual semantic segmentation. In Proc. Eur. Conf. Comput. Vis., pages 55-72, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.7, + 0.905, + 0.757 + ], + "angle": 0, + "content": "[52] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 633-641, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.758, + 0.905, + 0.813 + ], + "angle": 0, + "content": "[53] Lanyun Zhu, Tianrun Chen, Jianxiong Yin, Simon See, and Jun Liu. Continual semantic segmentation with automatic memory sample selection. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 3082-3092, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.813 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "25676" + } + ] +] \ No newline at end of file diff --git a/2025/CoMBO_ Conflict Mitigation via Branched Optimization for Class Incremental Segmentation/a4944725-d45e-4b4f-a404-3fd5f2ee3b6a_origin.pdf b/2025/CoMBO_ Conflict Mitigation via Branched Optimization for Class Incremental Segmentation/a4944725-d45e-4b4f-a404-3fd5f2ee3b6a_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..2ed682999b3a32639b02ad95a589b0ec9e667137 --- /dev/null +++ b/2025/CoMBO_ Conflict Mitigation via Branched Optimization for Class Incremental Segmentation/a4944725-d45e-4b4f-a404-3fd5f2ee3b6a_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:eff89f3d200ba65121c1af0b70e1b610d5ea2eb57588e5135d8cee499cdd20b5 +size 1674858 diff --git a/2025/CoMBO_ Conflict Mitigation via Branched Optimization for Class Incremental Segmentation/full.md b/2025/CoMBO_ Conflict Mitigation via Branched Optimization for Class Incremental Segmentation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..3de1431c37738cb57103851bb5f9a4198ecf3cb4 --- /dev/null +++ b/2025/CoMBO_ Conflict Mitigation via Branched Optimization for Class Incremental Segmentation/full.md @@ -0,0 +1,306 @@ +# CoMBO: Conflict Mitigation via Branched Optimization for Class Incremental Segmentation + +Kai Fang $^{1*}$ , Anqi Zhang $^{1*}$ , Guangyu Gao $^{1\dagger}$ , Jianbo Jiao $^{2}$ , Chi Harold Liu $^{1}$ , Yunchao Wei $^{3}$ $^{1}$ Beijing Institute of Technology $^{2}$ University of Birmingham $^{3}$ Beijing Jiaotong University + +# Abstract + +Effective Class Incremental Segmentation (CIS) requires simultaneously mitigating catastrophic forgetting and ensuring sufficient plasticity to integrate new classes. The inherent conflict above often leads to a back-and-forth, which turns the objective into finding the balance between the performance of previous (old) and incremental (new) classes. To address this conflict, we introduce a novel approach, Conflict Mitigation via Branched Optimization (CoMBO). Within this approach, we present the Query Conflict Reduction module, designed to explicitly refine queries for new classes through lightweight, class-specific adapters. This module provides an additional branch for the acquisition of new classes while preserving the original queries for distillation. Moreover, we develop two strategies to further mitigate the conflict following the branched structure, i.e., the Half-Learning Half-Distillation (HDHL) over classification probabilities, and the Importance-Based Knowledge Distillation (IKD) over query features. HDHL selectively engages in learning for classification probabilities of queries that match the ground truth of new classes, while aligning unmatched ones to the corresponding old probabilities, thus ensuring retention of old knowledge while absorbing new classes via learning negative samples. Meanwhile, IKD assesses the importance of queries based on their matching degree to old classes, prioritizing the distillation of important features and allowing less critical features to evolve. Extensive experiments in Class Incremental Panoptic and Semantic Segmentation settings have demonstrated the superior performance of CoMBO. Project page: https://guangyu-ryan.github.io/CoMBO. + +# 1. Introduction + +Semantic segmentation, the fundamental task in computer vision, involves classifying each pixel into predefined categories. Panoptic segmentation, the challeng- + +![](images/5f9cb9e87dd8c4c509c8e46890c3305acdd634a8bcadfb488ed943b6744c465a.jpg) +Figure 1. Comparison of our improved distillation and learning strategy (bottom) with previous conflicting strategy (top). Previous strategies impose contradictory supervision on the same target to find a balance, whereas our strategy, including the importance factor for distillation and the QCR module for adaptive learning, enables more compatible, target-specific supervision. The visualization on incremental classes ('Traffic Light') highlights the effectiveness of our strategy. + +ing variant, unifies semantic and instance segmentation to both classify every pixel and simultaneously identify distinct object instances. Recent advancements in mask classification-based segmentation, such as MaskFormer [16] and Mask2Former [17], enable the unification of semantic and panoptic segmentation. However, traditional approaches to segmentation suffer from limitations in dynamic environments where new classes can emerge unpredictably, as they are typically trained on a static set + +of categories. Therefore, Class Incremental Semantic Segmentation (CISS) has emerged as a pivotal task for continually learning new categories, where models can adapt to new classes without forgetting previously learned ones. + +Historically, studies in CISS [3, 35, 49] mainly focuses on tackling catastrophic forgetting, employing various techniques such as knowledge distillation [5, 20, 47], freezing parameters [9, 26, 48], or generating pseudo-labels from old model [20]. More recent efforts [8, 11] have extended these techniques to the more challenging Class Incremental Panoptic Segmentation (CIPS) based on the advanced Mask2Former. Nonetheless, these methods often fail to adequately reconcile the inherent tension between retaining old knowledge and acquiring new information. Some overemphasize maintaining stability, thus hindering the learning of new classes, while others prioritize plasticity, leading to significant forgetting of old classes. It is necessary to deconstruct these conflicting objectives to facilitate a more compatible incremental learning process, achieving promotion in both the acquisition of new classes and retention of old classes. + +In this work, we propose Conflict Mitigation via Branched Optimization (CoMBO), an approach to reduce conflicts between acquiring new classes and retaining old classes. By integrating three specialized knowledge distillation mechanisms within the Mask2Former architecture, CoMBO effectively mitigates catastrophic forgetting while enhancing the assimilation of new classes. Using bipartite matching, Mask2Former optimally aligns each query, complete with its classification logits and mask predictions, with the most fitting ground truth class, ensuring efficient optimization. Therefore, we first introduce the Half-Learning Half-Distillation mechanism, which selectively applies a shared Kullback-Leibler divergence to the classification logits of unmatched queries. This method distinctly separates distillation on unmatched masks from the classification optimization of matched ones. Moreover, rather than applying uniform distillation across all features, we introduce an importance-based knowledge distillation technique. This technique assesses each query's significance in preserving knowledge of old classes and adjusts the distillation intensity accordingly, emphasizing the retention of key queries. Additionally, our Query Conflict Reduction module innovatively adapts features for new classes while maintaining the queries' original characteristics before distillation, using class-specific adapters for each newly learned class. Finally, our approach outperforms previous methods in multiple benchmarks, especially in typical and challenging scenarios such as 100-10 of ADE20K. Our approach achieves remarkable performance of $35.6\%$ and $41.1\%$ mIoU on 100-10 of CIPS and CISS, respectively, compared to the previous state-of-the-art. + +Overall, our contributions are summarized as follows: + +- We propose a Half-Learning Half-Distillation mechanism that applies soft distillation for old classes and binary-based optimization for new classes to effectively balance retention and acquisition. +- An extra Query Conflict Reduction module is proposed to refine queries for new classes while preserving unrefined features, which enables Importance-based Knowledge Distillation for remembering the key features. +- Through extensive quantitative and qualitative evaluations on the ADE20K dataset, we demonstrate the state-of-the-art performance of our model in both Class Incremental Semantic and Panoptic Segmentation tasks. + +# 2. Related Work + +# 2.1. Semantic and Panoptic Segmentation + +Semantic segmentation aims to classify each pixel in an image, which focuses on the complete coverage of object(s) in each category. The pioneering work FCN [30] and U-Net [36] introduces a pixel-wise classification paradigm by removing the fully connected layers. Since then, numerous approaches [2, 12-15, 29, 38, 41] have been proposed following the paradigm, containing effective modules such as ASPP [13] and pyramid structure [50]. In contrast, panoptic segmentation [27] combines semantic and instance segmentation, requiring the classification of every pixel while distinguishing instances of the same category. These two tasks were addressed separately due to the limitation of the pixelwise classification. However, the emergence of the mask classification paradigm [16, 17, 43, 45] introduced a universal, transformer-based framework capable of solving multiple segmentation tasks simultaneously. MaskFormer [16] pioneered this paradigm by combining class-agnostic mask proposal generation and mask classification, enabling both semantic and panoptic segmentation simultaneously. Building upon this, Mask2Former [17] enhanced performance by incorporating multi-scale feature fusion and masked attention techniques. However, when learning new categories beyond the initial data, mask classification methods remain susceptible to catastrophic forgetting. + +# 2.2. Continual Segmentation + +In Class Incremental Learning (CIL), fine-tuning models on new data often results in a performance drop on old categories, i.e., catastrophic forgetting [22, 39]. To address this, various methods [5-7, 20, 21, 32, 33, 51] have been proposed, including retaining representative old samples [1, 4, 9, 10, 31, 32, 35, 53], compensation losses [18, 34, 42, 44], and prompt representations [25, 37, 40]. These advancements in CIL raise interest in more challenging segmentation tasks. Class Incremental Semantic Segmentation (CISS) was first introduced in MiB [5], reconstructing background regions to address background shifting during + +incremental steps. Since then, most methods [3, 47] have employed techniques such as knowledge distillation [28], pseudo-labels from previous models [8, 20], and background redefinition [5, 46] to mitigate background shifting. However, these methods primarily rely on pixel-wise classification models like DeepLabV3 [13]. Therefore, CoM-Former [8] first utilizes the mask classification segmentation model Mask2Former [17] enabling solving both CISS and Class Incremental Panoptic Segmentation (CIPS) tasks. ECLIPSE [26] introduces dynamic model structure expansion methods to extend learnable parameters for new classes and freezes the old parameters. CoMasTRe [23] distills queries matched to old queries with high probability on old classes. BalConpas [11] balance the class proportion on selecting representative replay samples. However, these methods often struggle to balance acquisition and preservation. Our approach mitigates this conflict by separating queries to decouple the two objectives. + +# 3. Preliminaries + +# 3.1. Problem Definition + +In continual segmentation, the model is tasked with an incremental learning challenge over $T$ steps. At each step $t \in \{1, \ldots, T\}$ , the model is trained to recognize a unique set of classes $C^t$ , where the intersection across all sets from $i$ to $T$ is empty $\bigcap_{t=1}^{T} C^t = \emptyset$ and their union forms the complete set of classes $\bigcup_{t=1}^{T} C^t = \mathcal{C}$ . In the current step $t$ , the training dataset $D_{train}^t$ consists of image-label pairs $(x^t, y^t)$ , where $x^t$ represents an image and $y^t$ its corresponding segmentation label. The labels $y^t$ are available only for the classes $C^t$ that need to be learned at this step, while labels for previously learned classes $C^{1:t-1}$ and future classes $C^{t+1:T}$ are inaccessible. After completing training at this step, the model must perform segmentation across all classes learned up to that point, $C^{1:t}$ , thereby preventing catastrophic forgetting of the old classes $C^{1:t-1}$ while effectively acquiring new ones $C^t$ . + +# 3.2. Revisiting Mask2Former and Pseudo-Labeling + +The recent advanced universal segmentation network, Mask2Former [17], has favored the class incremental semantic and panoptic segmentation (CISS and CIPS) tasks for its innovative proposal-based structure. Mask2Former deviates from traditional pixel-wise classification pipelines by introducing a class-agnostic mask prediction and classification mechanism. Specifically, the backbone $\pmb{f}_b$ and the pixel decoder $\pmb{f}_p$ generate multi-scale features that interact with $N$ learnable queries through multiple self-attention and cross-attention layers in the transformer decoder $\pmb{f}_t$ . These $N$ output queries $Q \in \mathbb{R}^{C_Q \times N}$ undergo further processing through two linear layers to form mask embeddings $\mathcal{E}_{mask} \in \mathbb{R}^{D \times N}$ and class embeddings $\mathcal{E}_{cls} \in \mathbb{R}^{(C + 1) \times N}$ . + +The mask embeddings $\mathcal{E}_{mask}$ yield $N$ mask predictions $M\in \mathbb{R}^{N\times H\times W}$ via dot product between each mask embedding and per-pixel embeddings $\mathcal{E}_{pixel}\in \mathbb{R}^{D\times H\times W}$ from $f_{p}$ , while $\mathcal{E}_{cls}$ accounts for the classification predictions of these masks. Note that the additional channel in $\mathcal{E}_{cls}$ represents the no-object class, an auxiliary category utilized during training but excluded during inference. + +In the contexts of CISS and CIPS, where ground truth for old classes $C^{1:t-1}$ are inaccessible, pseudo-labeling becomes crucial. PLOP [20] introduces this strategy by generating pixel-level pseudo ground truths for $C^{1:t-1}$ , adapted to the mask-based structure by CoMFormer [8], the first method based on Mask2Former for CISS and CIPS tasks. This adaptation involves weighting the mask predictions $M$ by the corresponding maximum class embedding score after the softmax process over the $C+1$ dimensions, i.e., $\max_{c=0}^{|C^{0:t-1}|} (\operatorname{softmax}_{c=0}^{|C^{0:t-1}|} (\mathcal{E}_{cls}))$ to form $M_{\omega}$ , with associated categories $C_{\omega} = \arg \max_{c=0}^{|C^{0:t-1}|} \mathcal{E}_{cls} \in \mathbb{R}^{N}$ . The pseudo-labeling process then labels the pixels outside the union of $\boldsymbol{y}^t$ with the category $c$ that achieves the maximum score across $M_{\omega}$ and $C_{w}$ : + +$$ +\tilde {M} _ {c} (h, w) = \left\{ \begin{array}{l l} 1 & \text {i f} c = C _ {\omega} [ \arg \max _ {n = 1} ^ {N} M _ {w} (n, h, w) ] \\ & \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \text {v a r e p s i l o n} \\ 0 & \text {e l s e} \end{array} \right. +$$ + +where $n$ indexes the $N$ proposals, and $h, w$ are the pixel coordinates. This pseudo-labeling strategy has become foundational for CISS and CIPS methods [8]. + +# 4. Proposed Method + +In our proposed Conflict Mitigation via Branched Optimization (CoMBO), we introduce three distinct strategies: Half-Distillation-Half-Learning Strategy for targeted knowledge retention, Importance-Based Knowledge Distillation to prioritize crucial features, and Query Conflict Reduction for efficient class integration. + +# 4.1. Query Conflict Reduction + +As mentioned in Sec. 3.2, both class embeddings $\mathcal{E}_{cls}$ and mask predictions $M$ are generated from the queries $Q$ . This interface becomes a critical juncture of conflicts between retaining knowledge of old classes and acquiring new classes manifest, which cannot break through the bottleneck of overall performance by simply finding the balance point. + +To deconstruct the crossroad efficiently, we design a Query Conflict Reduction (QCR) module $f_{QCR}(\cdot)$ as an additional adapter for new classes. As shown in Fig. 2, in most scenarios, queries from the previous layer $Q_{l-1}$ have the same classification results as the queries from the current layer due to the similar distribution and the same clas + +![](images/127ba02410a6d6f9244f3905c6ea692fdacf93ac171a7a254f79cedd063dc8c9.jpg) +Figure 2. Illustration of the Query Conflict Reduction (QCR) module. This module refines queries that predict incremental classes, allowing the processes of learning new features and retaining old features to occur separately. Note that the QCR module for previous incremental classes is frozen. + +sification layer, especially in the latter queries. Due to the shared classifier and mask generator following different layers of the transformer decoder, the classification results on the same position of queries from adjacent layers are approximately the same. Thus, we add the adapter for the queries from the penultimate layer $Q_{L-1}$ with classification results of new classes $C^{2:t}$ , as shown in Fig. 2. To be more specific, the queries $Q_{L-1,\tilde{c}}$ are selected for the adapter of incremental class $\tilde{c} \in C^{2:t}$ via: + +$$ +Q _ {L - 1, \tilde {c}} = \left\{Q _ {L - 1} (n), n \in N \wedge \tilde {c} = \arg \max _ {c = 0} ^ {| C ^ {0: t} |} \mathcal {E} _ {c l s} (n) \right\}, \tag {2} +$$ + +where $Q_{L-1,\tilde{c}}$ could contain none, one, or some selected queries that enable recognizing different instances. For each new category, we utilize the corresponding QCR module $f_{QCR,\tilde{c}}(\cdot)$ to refine the group of queries $Q_{L-1,\tilde{c}}$ . Considering the efficiency and effectiveness, we introduce the low-rank two-layer adaptation as the QCR module: + +$$ +\begin{array}{l} \hat {Q} _ {L - 1, \tilde {c}} = f _ {Q C R, \tilde {c}} \left(Q _ {L - 1, \tilde {c}}\right) \tag {3} \\ = Q _ {L - 1, \tilde {c}} W _ {1} W _ {2} + Q _ {L - 1, \tilde {c}}, \\ \end{array} +$$ + +where $\hat{Q}_{L - 1,\tilde{c}}$ represent the adapted queries of class $\tilde{c}$ , $W_{1} \in \mathbb{R}^{D \times r}$ and $W_{2} \in \mathbb{R}^{r \times D}$ denotes the weights for QCR. The QCR module separates the refined queries $\hat{Q}_{L,\tilde{c}}$ from the original queries $Q_{L,\tilde{c}}$ , enabling a bifurcate structure for simultaneously learning $C^t$ with class-specific adaptation and keep memorizing the features of $Q_{L,\tilde{c}}$ as well as the logits of $C^{0:t - 1}$ . + +# 4.2. Half-Distillation-Half-Learning Strategy + +Previous approaches based on Mask2Former usually separate learning and distillation processes on class embeddings $\mathcal{E}_{cls}$ due to its different learning strategies compared + +![](images/9c5c7aa6edd5c3bd35306df3176cf547cf072e7981c9d23ae82d3c06a9327b9a.jpg) +Figure 3. Details of the Half-Distillation-Half-Learning strategy on $N$ queries, including classification loss on matched queries and Kullback-Leibler Divergence loss on the other queries. The latter involves the logits of both old and current classes. + +to pixel-wise segmentation structure (e.g. DeepLabv3 [13]), mentioned in Sec. 3.2. This causes conflicts between distillation and learning, where the same logit could have contradictory optimization directions. Therefore, it is necessary to unify the optimization process towards a fixed target. + +Our Half-Distillation-Half-Learning (HDHL) strategy combines distillation and learning for $N$ class embeddings $\mathcal{E}_{cls}$ , as illustrated in Fig. 3. We follow the bipartite matching of Mask2Former and select embeddings $\mathcal{E}_{cls}^{C^t}$ matched to labels of current categories $C^t$ and $\mathcal{E}_{cls}^{C^{1:t-1}}$ matched to pseudo labels of old categories $C^{1:t-1}$ . These embeddings, following the original optimization strategy, are optimized by cross-entropy loss $\mathcal{L}_{cls}$ . The unmatched embeddings $\mathcal{E}_{cls}^{\emptyset^t}$ are grouped under the no-object (no- $obj$ ) category. However, adhering to the original learning strategy leads to catastrophic forgetting. Besides, using the distillation strategy for old categories $C^{1:t-1}$ using $\mathcal{E}_{cls}^{\emptyset^{t-1}}$ ignores the learning of current categories $C^t$ . However, identifying regions that do not belong to $C^t$ is just as important as learning those that do. Considering this, we design a strategy to simultaneously optimize the logits of $C^{1:t-1}$ and $C^t$ with a shared Kullback-Leibler Divergence $\mathcal{L}_{kl}$ : + +$$ +\mathcal {L} _ {k l} = \frac {1}{| \mathcal {E} _ {c l s} ^ {\varnothing^ {t}} |} \sum_ {j = 1} ^ {| \mathcal {E} _ {c l s} ^ {\varnothing^ {t}} |} \sum_ {c = 0} ^ {| \boldsymbol {C} ^ {0: t} |} \varphi_ {c l s} ^ {\varnothing^ {t - 1}} (j, c) \log \frac {\varphi_ {c l s} ^ {\varnothing^ {t - 1}} (j , c)}{\varphi_ {c l s} ^ {\varnothing^ {t}} (j , c)}, \tag {4} +$$ + +where + +$$ +\begin{array}{l} \varphi_ {c l s} ^ {\emptyset^ {t}} (j, c) = \mathcal {S} ^ {t} (j, c), \\ \varphi_ {c l s} ^ {\emptyset^ {t - 1}} (j, c) = \left\{ \begin{array}{l l} \mathcal {S} ^ {t - 1} (j, c) & \text {i f} c \in C ^ {0: t - 1}, \\ 0 & \text {e l s e}, \end{array} \right. \tag {5} \\ \end{array} +$$ + +and + +$$ +\mathcal {S} ^ {t} (j, c) = \frac {\exp \left(\mathcal {E} _ {c l s} ^ {\varnothing^ {t}} (j , c)\right)}{\sum_ {\hat {c} = 0} ^ {| \mathcal {C} ^ {0 : t} |} \exp \left(\mathcal {E} _ {c l s} ^ {\varnothing^ {t}} (j , \hat {c})\right)} \tag {6} +$$ + +represent the softmax activation. + +The $\mathcal{L}_{kl}$ not only distill the previous class embedding $\mathcal{E}_{cls}^{\emptyset^{t-1}}$ to the corresponding categories, including the no-object category, but decreasing the logits of current categories to figure out the objects apart from them. The unified HDHL strategy maintains the probability distribution of old logits seamlessly with the integration of new logits. By combining the original $\mathcal{L}_{cls}$ and the designed $\mathcal{L}_{kl}$ as $\mathcal{L}_{DL}$ , the optimization process covers all the class embeddings without contradictory targets: + +$$ +\mathcal {L} _ {D L} = \lambda_ {c l s} \cdot \mathcal {L} _ {c l s} + \lambda_ {k l} \cdot \mathcal {L} _ {k l}, \tag {7} +$$ + +where $\lambda_{cls}$ and $\lambda_{kl}$ control the balance between two losses. + +# 4.3. Importance-Based Knowledge Distillation + +Previous methods mitigate catastrophic forgetting by applying feature distillation, which may hinder the acquisition of new knowledge. To address this, we propose an Importance-Based Knowledge Distillation (IKD) method, in which stronger distillation is applied to features critical for old classes, while less emphasis is placed on less pertinent features, thereby balancing stability with plasticity. + +The importance measurement requires the cost matrix $\mathcal{A} \in \mathbb{R}^{N \times S}$ , adopted in the bipartite matching process, where $S$ is the number of image labels. Values in $\mathcal{A}$ indicate the relationship between queries and classes. For each image, the transformer decoder generates $N$ predicted segments, denoted as $\mathcal{Z} = \left\{(\mathcal{E}_{cls}(n), M(n))\right\}_{n=1}^{N}$ . The cost matrix $\mathcal{A}$ is then computed based on the output segments $\mathcal{Z}$ and the image labels $\boldsymbol{y}^t = \{(c_s, m_s)\}_{s=1}^S$ : + +$$ +\mathcal {A} (n, s) = - \lambda_ {c l s} \cdot \varphi_ {c l s} \left(n, c _ {s}\right) + \lambda_ {\text {m a s k}} \cdot \mathcal {L} _ {\text {m a s k}} \left(\mathcal {E} _ {\text {m a s k}} (n), m _ {s}\right), \tag {8} +$$ + +where $\varphi_{cls}$ is the classification output obtained by applying softmax to $\mathcal{E}_{cls}$ , and $\lambda_{cls}$ and $\lambda_{mask}$ represent the corresponding hyperparameters. + +After each training stage, we compute the cost matrix $\mathcal{A}$ across every image in the training set $D_{train}^{t}$ of the current incremental step $t$ . The minimum costs among the classes for $N$ segments are accumulated in the corresponding position of buffer $B^{t} \in \mathbb{R}^{N}$ , where segments with lower minimum cost values are more likely to recognize and cover the regions of a class $C^{t}$ . By normalizing and reversing the values of $B^{t}$ , we derive the importance matrix $I_{C^{t}}^{t} \in \mathbb{R}^{N}$ of current classes $C^{t}$ . The importance matrix $I^{t+1} \in \mathbb{R}^{N}$ is further estimated via the weighted accumulation of current importance matrix $I^{t} \in \mathbb{R}^{N}$ and $I_{C^{t}}^{t}$ , where the weights refer to the $|C^{1:t-1}|$ and $|C^{t}|$ . The generated $I^{t+1}$ will be + +utilized in the next incremental training phase for weighted distillation of the queries $Q^t$ , which is computed as follows: + +$$ +\mathcal {L} _ {I K D} = \frac {1}{N} \sum_ {n = 1} ^ {N} I ^ {t} (n) \cdot \| Q ^ {t} (n), Q ^ {t - 1} (n) \| _ {2} ^ {2}, \tag {9} +$$ + +where $\| \cdot \|_2$ denotes the Euclidean distance, $Q^t(n)$ and $Q^{t-1}(n)$ denotes the transformer decoder features from the current model $f^t(\cdot)$ and the old model $f^{t-1}(\cdot)$ , respectively. Algorithm 1 presents the details of the importance estimation procedure after training $f^t(\cdot)$ . + +Algorithm 1 Importance Matrix Estimation +1: Input: current model $f^{t}(\cdot)$ , current dataset $D_{train}^{t}$ , current importance matrix $I^{t}$ +2: Initialization: $I^{t+1} \gets 0$ , $B^{t} \gets 0$ +3: for all $(\boldsymbol{x}^{t}, \boldsymbol{y}^{t}) \in D_{train}^{t}$ do +4: Compute $A$ following Eq. (8) +5: Update the buffer: +6: $B^{t} \gets B^{t} + \min(A, \text{axis} = 1)$ +7: end for +8: for $n \in \{1, 2, \dots, N\}$ do +9: $I_{C^{t}}^{t}(n) \gets 1 - \frac{B^{t}(n) - \min(B^{t})}{\max(B^{t}) - \min(B^{t})}$ +10: $I^{t+1}(n) \gets \frac{|C^{1:t-1}|}{|C^{1:t}|} \cdot I^{t}(n) + \frac{|C^{t}|}{|C^{1:t}|} \cdot I_{C^{t}}^{t}(n)$ +11: end for +12: Output: Importance matrix $I^{t+1}$ for step $t + 1$ + +# 4.4. Objective Function + +Overall, the objective function of CoMBO is defined as: + +$$ +\mathcal {L} _ {H} = \mathcal {L} _ {D L} + \lambda_ {I K D} \mathcal {L} _ {I K D}. \tag {10} +$$ + +where $\mathcal{L}_{DL}$ represents the Half-Distillation-Half-Learning loss, and $\mathcal{L}_{IKD}$ means the Importance-Based Knowledge Distillation loss, with $\lambda_{IKD}$ as its weighting factor. + +# 5. Experimental Results + +# 5.1. Experimental Setup + +Datasets and Evaluation Metrics. We follow the experimental settings of previous works [8, 11, 26] and evaluate our approach on the ADE20K [52] dataset. The ADE20K dataset is specifically designed to support both panoptic and semantic segmentation tasks. ADE20K contains 150 classes, including 100 thing classes and 50 stuff classes. The dataset is composed of 20,210 images for training and 2,000 images for validation. + +For Class Incremental Semantic Segmentation (CISS), we adopt the mean Intersection over Union (mIoU) for evaluation. The Intersection over Union (IoU) is calculated as $\mathrm{IoU} = \frac{TP}{TP + FP + FN}$ , where $TP$ , $FP$ , and $FN$ denote the + +
Method100-50 (2 steps)100-10 (6 steps)100-5 (11 steps)50-50 (3 steps)
1-100101-150all1-100101-150all1-100101-150all1-5051-150all
FT0.01.30.40.02.91.00.025.88.60.012.08.1
MiB [5]CVPR2035.119.329.827.110.021.424.06.518.142.415.524.4
PLOP [20]CVPR2140.222.434.330.517.526.128.115.724.045.818.727.7
CoMFormer [8]CVPR2341.127.736.736.017.129.734.415.928.245.019.327.9
ECLIPSE [26]CVPR2441.723.535.641.418.833.941.116.632.946.020.729.2
BalConpas [11]ECCV2442.825.737.140.722.834.736.120.330.851.226.534.7
CoMBO Ours43.925.637.840.825.235.636.120.530.950.728.235.7
Joint43.830.939.543.830.939.543.830.939.550.733.939.5
+ +Table 1. Quantitative comparison under Class Incremental Panoptic Segmentation with state-of-the-art exemplar-free methods on ADE20K in PQ. Scores of novel classes and all classes in bold are the best while underlined are the second best. + +number of true-positive, false-positive, and false-negative pixels, respectively. The mIoU metric averages the IoU across all classes for a more comprehensive evaluation. For Class Incremental Panoptic Segmentation (CIPS), following previous work [8], we employ Panoptic Quality (PQ) as the evaluation metric. PQ is defined as the product of Recognition Quality (RQ) and Segmentation Quality (SQ). To measure incremental learning capacity, we compute the corresponding metrics for the initial classes $C^1$ , incremental classes $C^{2:T}$ , and the aggregate of all classes $C^{1:T}$ . + +Protocols and Implementation Details. We follow previous incremental protocols and define scenarios as $N_{ini} - N_{inc}$ , where $N_{ini}$ represents the number of initial classes, and $N_{inc}$ denotes the number of new classes introduced at each incremental step. For example, in the 100-10 scenario, the training begins with 100 classes, followed by the addition of 10 new classes per incremental step, without access to annotations of old classes. For both CIPS and CISS, we do evaluations on the following scenarios: 100-10 (6 steps), 100-50 (2 steps), 100-5 (11 steps) and 50-50 (3 steps). + +Our approach is based on the Mask2Former structure [17]. Following previous approaches to CISS and CIPS, we utilize ResNet-50 [24] as the backbone for CIPS and ResNet-101 for CISS, with both pre-trained on ImageNet [19]. The input resolution of the images is $640 \times 640$ with an all-time batch size of 8. We follow the training hyperparameter of Mask2Former in the initial step, with a learning rate of $10^{-4}$ and iterations of 160,000. During the incremental steps, we set 1,000 iterations per class with a learning rate of $5 \times 10^{-5}$ . The coefficients $r$ , $\lambda_{cls}$ , $\lambda_{kl}$ , $\lambda_{IKD}$ are respectively set to 16, 2, 5, 3. All experiments were conducted on the NVIDIA RTX 4090. + +# 5.2.Quantitative Results + +Comparisons in Class Incremental Panoptic Segmentation (CIPS). We evaluated our approach against state-of-the-art exemplar-free methods on the ADE20K dataset + +within the CIPS framework, as detailed in Tab. 1. The performance of our CoMBO surpasses other methods in most of the scenarios, particularly in the more challenging 100-10 scenario, where it achieves a remarkable $35.6\%$ PQ with at least $3.4\%$ of advantage on incremental classes compared to the previous state-of-the-art methods. This notable performance gain demonstrates the effectiveness of our CoMBO in reducing conflicts. Although our approach still underperforms the state-of-the-art approach ECLIPSE [26] in the 100-5 scenario, primarily due to its strong memorization ability from freezing parameters, our method achieves advanced performance for the new classes. Moreover, in scenarios with fewer initial classes, i.e., 50-50 scenario, we maintain our advantage in learning new classes with a PQ of $28.2\%$ for incremental classes and $35.7\%$ for overall performance, widening the gap with strong distillation methods. + +Comparisons in Class Incremental Semantic Segmentation (CISS). We further extended our evaluation to the semantic segmentation benchmark, comparing our approach with previous methods on the ADE20K dataset, as detailed in Tab. 2. Our approach consistently outperforms previous methods in all tested scenarios. Notably, in the 100-10 scenario, our approach achieves a mIoU of $41.1\%$ , surpassing the previous state-of-the-art by at least $3.5\%$ for incremental classes, demonstrating its effectiveness in mitigating conflicts. Moreover, in the most challenging 100-5 long-term incremental scenario, our approach not only improves performance on old classes by $2.5\%$ but also achieves a significant $5.4\%$ increase in mIoU for new classes, ensuring both model stability and adaptability. + +# 6. Ablation Study + +# 6.1. Component Ablations + +We analyze the key components of our proposed framework, which includes the Half-Distillation-Half-Learning Strategy, Importance-based Knowledge Distillation, and the + +
Method100-50 (2 steps)100-10 (6 steps)100-5 (11 steps)50-50 (3 steps)
1-100101-150all1-100101-150all1-100101-150all1-5051-150all
FT0.026.78.90.02.30.80.01.10.30.01.71.1
MiB [5]CVPR2037.024.132.623.510.626.621.06.116.145.621.029.3
PLOP [20]CVPR2144.226.238.234.815.928.539.513.630.954.930.238.4
CoMFormer [8]CVPR2344.726.238.440.615.632.339.513.630.949.226.634.1
CoMasTRe [23]CVPR2445.726.039.242.318.434.440.815.832.649.826.634.5
ECLIPSE [26]CVPR2445.021.737.143.417.434.643.316.334.2---
BalConpas [11]ECCV2449.930.143.347.324.238.642.117.233.855.833.340.8
CoMBO Ours50.234.444.947.827.741.144.622.637.355.336.943.0
Joint51.740.247.851.740.247.851.740.247.856.643.547.8
+ +Query Conflict Reducing module. Our experiments focus on the 100-10 scenario in the CIPS, as it provides a moderate number of steps to examine the ability of both acquisition and retention. We evaluate various combinations of these components separately and present the results in Tab. 3. The baseline method utilizes the vanilla loss with pseudo-labeling. + +Table 2. Quantitative comparison under Class Incremental Semantic Segmentation with state-of-the-art exemplar-free methods on ADE20K in mIoU. Scores of novel classes and all classes in bold are the best while underlined are the second best. + +
HDHLIKDQCR100-10 (6 steps)
1-100101-150all
36.323.832.2
✓39.424.234.3
✓✓40.024.634.9
✓✓38.825.434.4
✓✓✓40.825.235.6
+ +Table 3. Ablation study of the main components on the 100-10 task of CIPS. The baseline in the $1^{\text{st}}$ row employs vanilla losses with pseudo-labeling and excludes the QCR module. + +Under identical experimental conditions, the HDHL Strategy significantly enhances the overall PQ by $2.1\%$ , demonstrating its ability to unify logits of new classes into the previous logits distribution without contradictory losses. The implementation of the IKD mechanism has a remarkable $3.7\%$ increase in the PQ for old classes, underscoring its effectiveness in boosting the ability to retain the knowledge of old classes. Furthermore, when combined with the QCR module for branched optimization, CoMBO further boosts both the performance of old and new classes, achieving $0.8\%$ and $0.6\%$ gains in PQ, respectively, surpassing previous methods that struggle to balance acquisition and retention. Consequently, CoMBO achieves a $3.4\%$ PQ improvement over the baseline. The above results demonstrate the effectiveness of the proposed CoMBO and its related modules, showing a noticeable improvement in reducing the conflict on model structures and losses. + +# 6.2. Ablation Study of QCR + +We validate the effectiveness of QCR with different settings of $r$ shown in Fig. 5. Under the estimation of performance enhancement and additional parameters, our QCR module achieves the best result when $r = 16$ , where the quantity of additional parameters comes to $8.2\mathrm{K}$ per class. Considering the total parameter of Mask2Former [17] is $44.9\mathrm{M}$ with ResNet-50, the additional parameters only reach up to $2\%$ of original parameters in total, which is 5 times less than per class additional parameters of ECLIPSE [26] while having better performance on new classes. + +# 6.3. Ablation Study of HDHL Strategy + +To validate the effectiveness of our Half-Distillation-Half-Learning strategy, we conduct a series of experiments among other strategies, as shown in Tab. 4. The $1^{\mathrm{st}}$ row with $\mathcal{L}_{cls}$ on all classification embeddings represents the vanilla loss with pseudo-labeling, which is adopted in [8]. The $2^{\mathrm{nd}}$ row utilizes an auxiliary distillation on the embeddings following [23]. Compared to these strategies with tough supervision on unmatched embeddings $\mathcal{E}_{cls}^{t,\emptyset}$ and mostly conflicting losses, our HDHL strategy with soft and concentrated supervision has a better PQ of 34.3. Besides, the HDHL design of $\mathcal{L}_{kl}$ ensures the integrity while optimizing the unmatched embeddings $\mathcal{E}_{cls}^{t,\emptyset}$ , with improvement of $2.1\%$ compared to distillation-only strategy in the $3^{\mathrm{rd}}$ row and $0.8\%$ compared to the simple combination of distillation and $\mathcal{L}_{cls}$ . The above results show that our HDHL strategy outperforms other learning or distillation strategies on the class embeddings. + +# 6.4. Qualitative Analysis + +We conduct the qualitative analysis (shown in Fig. 4) by comparing our proposed CoMBO to the Baseline and Baseline+HDHL $(1^{\mathrm{st}}$ and $2^{\mathrm{nd}}$ rows in Tab. 3). In the $1^{\mathrm{st}}$ , $2^{\mathrm{nd}}$ , and $5^{\mathrm{th}}$ columns of Fig. 4, our CoMBO can indicate the incremental classes, i.e., screen door, fountain, and ottoman, + +![](images/aa5a3fe79756d31ace26406ce9e8d831a64fa3a8d44c041e001872af8c0736a3.jpg) +Figure 4. Qualitative results of CoMBO comparing to Baseline, Baseline+HDHL on 100-10 CIPS task of ADE20K. + +![](images/c8eee99b786109d66d892a0b6bb7ddb6069f54262cde2270d14dc4c33794e2fe.jpg) +Figure 5. Ablation study of $r$ in QCR module, including PQ performance (left) and number of parameters (right). + +
LclsLklPQ
Ectls & Ectls1:t-1EctlsC0:t-1Ct
✓✓32.2
✓✓✓32.6
✓✓32.2
✓✓✓33.5
✓✓✓34.3
+ +Table 4. Ablation study of the HDHL strategy on task 100-10 of CIPS. The $1^{\text{st}}$ row represents the vanilla loss with pseudo-labeling. The last row represents our $\mathcal{L}_{DL}$ . + +while other methods fail to recognize all these objects. In the $3^{\mathrm{rd}}$ column, our method successfully remembers the car in the left instead of covering the object with the prediction of the incremental class van. Furthermore, our method allows more precise mask prediction of new class Traffic Light in the $4^{\mathrm{th}}$ column, demonstrating the effectiveness of refinements on the queries corresponding to new classes. + +# 7. Conclusions + +In this paper, we present CoMBO, a novel Class Incremental Segmentation (CIS) method for mitigating the conflict between acquisition and retention losses. This conflict arises from competing goals of efficiently acquiring knowledge about new classes while preserving knowledge of previously learned ones. Our QCR module branches the conflicting optimization targets via lightweight class-specific adaptation on queries, enabling the coexistence of learning and distillation on separated queries. The HDHL strategy and IKD further reduce the contradictory optimization on classification logits and queries with unified targets, respectively, thereby enhancing the overall model performance. Extensive experimental validation on the ADE20K dataset underscores the superiority of our approach, demonstrating the effectiveness of each component in the CIS task, particularly excelling in performance on new classes. + +# Acknowledgments + +This work was supported by the National Key R&D Program of China (Grant 2022YFC3310200), the National Natural Science Foundation of China (Grant 62472033, 92470203), the Beijing Natural Science Foundation (Grant L242022), the Royal Society grants (SIF\R1\231009, IES\R3\223050) and an Amazon Research Award. + +# References + +[1] Rahaf Aljundi, Min Lin, Baptiste Goujaud, and Yoshua Bengio. Gradient based sample selection for online continual learning. Adv. Neural Inf. Process. Syst., 32, 2019. 2 +[2] Vijay Badrinarayanan, Alex Kendall, and Roberto Cipolla. Segnet: A deep convolutional encoder-decoder architecture for image segmentation. IEEE Trans. on Pattern Anal. Mach. Intell., 39(12):2481-2495, 2017. 2 +[3] Donghyeon Baek, Youngmin Oh, Sanghoon Lee, Junghyup Lee, and Bumsub Ham. Decomposed knowledge distillation for class-incremental semantic segmentation. Adv. Neural Inf. Process. Syst., 35:10380-10392, 2022. 2, 3 +[4] Zalán Borsos, Mojmir Mutny, and Andreas Krause. Coresets via bilevel optimization for continual learning and streaming. Adv. Neural Inf. Process. Syst., 33:14879-14890, 2020. 2 +[5] Fabio Cermelli, Massimiliano Mancini, Samuel Rota Bulo, Elisa Ricci, and Barbara Caputo. Modeling the background for incremental learning in semantic segmentation. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 9233-9242, 2020. 2, 3, 6, 7 +[6] Fabio Cermelli, Dario Fontanel, Antonio Tavera, Marco Ciccone, and Barbara Caputo. Incremental learning in semantic segmentation from image labels. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 4371-4381, 2022. +[7] Fabio Cermelli, Antonino Geraci, Dario Fontanel, and Barbara Caputo. Modeling missing annotations for incremental learning in object detection. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 3700-3710, 2022. 2 +[8] Fabio Cermelli, Matthieu Cord, and Arthur Douillard. Comformer: Continual learning in semantic and panoptic segmentation. In Proc. IEEE Conf. Comput. Vis. Pattern Recognition., pages 3010-3020, 2023. 2, 3, 5, 6, 7 +[9] Sungmin Cha, YoungJoon Yoo, Taesup Moon, et al. Ssul: Semantic segmentation with unknown label for exemplar-based class-incremental learning. Adv. Neural Inf. Process. Syst., 34:10919-10930, 2021. 2 +[10] Jinpeng Chen, Runmin Cong, Yuxuan Luo, Horace Ip, and Sam Kwong. Saving 100x storage: prototype replay for reconstructing training sample distribution in class incremental semantic segmentation. Adv. Neural Inf. Process. Syst., 36, 2024. 2 +[11] Jinpeng Chen, Runmin Cong, Yuxuan Luo, Horace Ho Shing Ip, and Sam Kwong. Strike a balance in continual panoptic segmentation. In Proc. Eur. Conf. Comput. Vis., pages 126-142. Springer, 2025. 2, 3, 5, 6, 7 +[12] Liang-Chieh Chen. Semantic image segmentation with deep convolutional nets and fully connected crfs. arXiv preprint arXiv:1412.7062, 2014. 2 + +[13] L. C. Chen. Rethinking atrous convolution for semantic image segmentation. arXiv preprint:1706.05587, 2017. 2, 3, 4 +[14] Liang-Chieh Chen, George Papandreou, and et al. Deep convolutional nets, atrous convolution, and fully connected crfs. IEEE Trans. on Pattern Anal. Mach. Intell., 40(4):834-848, 2017. +[15] Liang-Chieh Chen, Yukun Zhu, George Papandreou, Florian Schroff, and Hartwig Adam. Encoder-decoder with atrous separable convolution for semantic image segmentation. In Proc. Eur. Conf. Comput. Vis., pages 801-818, 2018. 2 +[16] Bowen Cheng, Alex Schwing, and Alexander Kirillov. Per-pixel classification is not all you need for semantic segmentation. Adv. Neural Inf. Process. Syst., 34:17864-17875, 2021. 1, 2 +[17] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 1290–1299, 2022. 1, 2, 3, 6, 7 +[18] Wei Cong, Yang Cong, Yuyang Liu, and Gan Sun. Cs2k: Class-specific and class-shared knowledge guidance for incremental semantic segmentation. In Proc. Eur. Conf. Comput. Vis., pages 244-261, 2025. 2 +[19] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 248–255, 2009. 6 +[20] Arthur Douillard, Yifu Chen, Arnaud Dapogny, and Matthieu Cord. Plop: Learning without forgetting for continual semantic segmentation. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 4040-4050, 2021. 2, 3, 6, 7 +[21] Arthur Douillard, Yifu Chen, Arnaud Dapogny, and Matthieu Cord. Tackling catastrophic forgetting and background shift in continual semantic segmentation. arXiv preprint arXiv:2106.15287, 2021. 2 +[22] Robert M French. Catastrophic forgetting in connectionist networks. Trends in cognitive sciences, 3(4):128-135, 1999. 2 +[23] Yizheng Gong, Siyue Yu, Xiaoyang Wang, and Jimin Xiao. Continual segmentation with disentangled objectness learning and class recognition. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 3848-3857, 2024. 3, 7 +[24] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 770-778, 2016. 6 +[25] M. G. Z. A. Khan, M. F. Naeem, Luc Van Gool, Didier Stricker, Federico Tombari, and Muhammad Zeshan Afzal. Introducing language guidance in prompt-based continual learning. In Proc. IEEE Int. Conf. Comput. Vis., pages 11463-11473, 2023. 2 +[26] Beomyoung Kim, Joonsang Yu, and Sung Ju Hwang. Eclipse: Efficient continual learning in panoptic segmentation with visual prompt tuning. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 3346-3356, 2024. 2, 3, 5, 6, 7 +[27] Alexander Kirillov, Kaiming He, Ross Girshick, Carsten Rother, and Piotr Dollar. Panoptic segmentation. In Proc. + +IEEE Conf. Comput. Vis. Pattern Recognit., pages 9404-9413, 2019. 2 +[28] Zhizhong Li and Derek Hoiem. Learning without forgetting. IEEE Trans. on Pattern Anal. Mach. Intell., 40(12):2935-2947, 2017. 3 +[29] Guosheng Lin, Anton Milan, Chunhua Shen, and Ian Reid. Refinenet: Multi-path refinement networks for high-resolution semantic segmentation. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 1925-1934, 2017. 2 +[30] Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 3431-3440, 2015. 2 +[31] David Lopez-Paz and Marc'Aurelio Ranzato. Gradient episodic memory for continual learning. Adv. Neural Inf. Process. Syst., 30, 2017. 2 +[32] Andrea Maracani, Umberto Michieli, Marco Toldo, and Pietro Zanuttigh. Recall: Replay-based continual learning in semantic segmentation. In Proc. IEEE Int. Conf. Comput. Vis., pages 7026-7035, 2021. 2 +[33] Umberto Michieli and Pietro Zanuttigh. Continual semantic segmentation via repulsion-attraction of sparse and disentangled latent representations. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 1114-1124, 2021. 2 +[34] Umberto Michieli and Pietro Zanuttigh. Continual semantic segmentation via repulsion-attraction of sparse and disentangled latent representations. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 1114-1124, 2021. 2 +[35] Youngmin Oh, Donghyeon Baek, and Bumsub Ham. Alife: Adaptive logit regularizer and feature replay for incremental semantic segmentation. Adv. Neural Inf. Process. Syst., 35: 14516-14528, 2022. 2 +[36] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Proc. International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 234-241, 2015. 2 +[37] Chao Shang, Hongliang Li, Fanman Meng, Qingbo Wu, Heqian Qiu, and Lanxiao Wang. Incrementer: Transformer for class-incremental semantic segmentation with knowledge distillation focusing on old class. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 7214-7224, 2023. 2 +[38] Robin Strudel, Ricardo Garcia, Ivan Laptev, and Cordelia Schmid. Segmenter: Transformer for semantic segmentation. In Proc. IEEE Int. Conf. Comput. Vis., pages 7262-7272, 2021. 2 +[39] Sebastian Thrun. Lifelong learning algorithms. In Learning to learn, pages 181-209. 1998. 2 +[40] Zifeng Wang, Zizhao Zhang, Chen-Yu Lee, Han Zhang, Ruoxi Sun, Xiaoqi Ren, Guolong Su, Vincent Perot, Jennifer Dy, and Tomas Pfister. Learning to prompt for continual learning. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 139-149, 2022. 2 +[41] Yunchao Wei, Jiashi Feng, Xiaodan Liang, Ming-Ming Cheng, Yao Zhao, and Shuicheng Yan. Object region mining with adversarial erasing: A simple classification to semantic segmentation approach. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 1568-1576, 2017. 2 + +[42] Jia-Wen Xiao, Chang-Bin Zhang, Jiekang Feng, Xialei Liu, Joost van de Weijer, and Ming-Ming Cheng. Endpoints weight fusion for class incremental semantic segmentation. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 7204-7213, 2023. 2 +[43] Shipeng Yan, Jiangwei Xie, and Xuming He. Der: Dynamically expandable representation for class incremental learning. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 3014-3023, 2021. 2 +[44] Chaohui Yu, Qiang Zhou, Jingliang Li, Jianlong Yuan, Zhibin Wang, and Fan Wang. Foundation model drives weakly incremental learning for semantic segmentation. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 23685-23694, 2023. 2 +[45] Qihang Yu, Huiyu Wang, Siyuan Qiao, Maxwell Collins, Yukun Zhu, Hartwig Adam, Alan Yuille, and Liang-Chieh Chen. k-means mask transformer. In Proc. Eur. Conf. Comput. Vis., pages 288-307, 2022. 2 +[46] Anqi Zhang and Guangyu Gao. Background adaptation with residual modeling for exemplar-free class-incremental semantic segmentation. In Proc. Eur. Conf. Comput. Vis., 2024. 3 +[47] Chang-Bin Zhang, Jia-Wen Xiao, Xialei Liu, Ying-Cong Chen, and Ming-Ming Cheng. Representation compensation networks for continual semantic segmentation. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 7053-7064, 2022. 2, 3 +[48] Zekang Zhang, Guangyu Gao, Zhiyuan Fang, Jianbo Jiao, and Yunchao Wei. Mining unseen classes via regional objectness: A simple baseline for incremental segmentation. Adv. Neural Inf. Process. Syst., 35:24340-24353, 2022. 2 +[49] Zekang Zhang, Guangyu Gao, Jianbo Jiao, Chi Harold Liu, and Yunchao Wei. Coinseg: Contrast inter-and intra-class representations for incremental segmentation. In Proc. IEEE Int. Conf. Comput. Vis., pages 843-853, 2023. 2 +[50] Hengshuang Zhao, Jianping Shi, Xiaojuan Qi, Xiaogang Wang, and Jiaya Jia. Pyramid scene parsing network. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 2881-2890, 2017. 2 +[51] Hanbin Zhao, Fengyu Yang, Xinghe Fu, and Xi Li. Rbc: Rectifying the biased context in continual semantic segmentation. In Proc. Eur. Conf. Comput. Vis., pages 55-72, 2022. 2 +[52] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 633-641, 2017. 5 +[53] Lanyun Zhu, Tianrun Chen, Jianxiong Yin, Simon See, and Jun Liu. Continual semantic segmentation with automatic memory sample selection. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 3082-3092, 2023. 2 \ No newline at end of file diff --git a/2025/CoMBO_ Conflict Mitigation via Branched Optimization for Class Incremental Segmentation/images.zip b/2025/CoMBO_ Conflict Mitigation via Branched Optimization for Class Incremental Segmentation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..11a638a68ba2a274d2b3765871e0f7dbcb2c28de --- /dev/null +++ b/2025/CoMBO_ Conflict Mitigation via Branched Optimization for Class Incremental Segmentation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:57cd9f532b1425238675083018ca75bc1fa443f390b1177d1ff6530ad5455f87 +size 656187 diff --git a/2025/CoMBO_ Conflict Mitigation via Branched Optimization for Class Incremental Segmentation/layout.json b/2025/CoMBO_ Conflict Mitigation via Branched Optimization for Class Incremental Segmentation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..9fdee9aeb8f4d801862595221e94a5a3e41ebf9b --- /dev/null +++ b/2025/CoMBO_ Conflict Mitigation via Branched Optimization for Class Incremental Segmentation/layout.json @@ -0,0 +1,10494 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 120, + 103, + 491, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 120, + 103, + 491, + 140 + ], + "spans": [ + { + "bbox": [ + 120, + 103, + 491, + 140 + ], + "type": "text", + "content": "CoMBO: Conflict Mitigation via Branched Optimization for Class Incremental Segmentation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 160, + 539, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 160, + 539, + 191 + ], + "spans": [ + { + "bbox": [ + 67, + 160, + 539, + 191 + ], + "type": "text", + "content": "Kai Fang" + }, + { + "bbox": [ + 67, + 160, + 539, + 191 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 67, + 160, + 539, + 191 + ], + "type": "text", + "content": ", Anqi Zhang" + }, + { + "bbox": [ + 67, + 160, + 539, + 191 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 67, + 160, + 539, + 191 + ], + "type": "text", + "content": ", Guangyu Gao" + }, + { + "bbox": [ + 67, + 160, + 539, + 191 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 67, + 160, + 539, + 191 + ], + "type": "text", + "content": ", Jianbo Jiao" + }, + { + "bbox": [ + 67, + 160, + 539, + 191 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 67, + 160, + 539, + 191 + ], + "type": "text", + "content": ", Chi Harold Liu" + }, + { + "bbox": [ + 67, + 160, + 539, + 191 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 67, + 160, + 539, + 191 + ], + "type": "text", + "content": ", Yunchao Wei" + }, + { + "bbox": [ + 67, + 160, + 539, + 191 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 67, + 160, + 539, + 191 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 67, + 160, + 539, + 191 + ], + "type": "text", + "content": "Beijing Institute of Technology " + }, + { + "bbox": [ + 67, + 160, + 539, + 191 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 67, + 160, + 539, + 191 + ], + "type": "text", + "content": "University of Birmingham " + }, + { + "bbox": [ + 67, + 160, + 539, + 191 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 67, + 160, + 539, + 191 + ], + "type": "text", + "content": "Beijing Jiaotong University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 151, + 217, + 200, + 229 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 217, + 200, + 229 + ], + "spans": [ + { + "bbox": [ + 151, + 217, + 200, + 229 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 242, + 297, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 242, + 297, + 602 + ], + "spans": [ + { + "bbox": [ + 55, + 242, + 297, + 602 + ], + "type": "text", + "content": "Effective Class Incremental Segmentation (CIS) requires simultaneously mitigating catastrophic forgetting and ensuring sufficient plasticity to integrate new classes. The inherent conflict above often leads to a back-and-forth, which turns the objective into finding the balance between the performance of previous (old) and incremental (new) classes. To address this conflict, we introduce a novel approach, Conflict Mitigation via Branched Optimization (CoMBO). Within this approach, we present the Query Conflict Reduction module, designed to explicitly refine queries for new classes through lightweight, class-specific adapters. This module provides an additional branch for the acquisition of new classes while preserving the original queries for distillation. Moreover, we develop two strategies to further mitigate the conflict following the branched structure, i.e., the Half-Learning Half-Distillation (HDHL) over classification probabilities, and the Importance-Based Knowledge Distillation (IKD) over query features. HDHL selectively engages in learning for classification probabilities of queries that match the ground truth of new classes, while aligning unmatched ones to the corresponding old probabilities, thus ensuring retention of old knowledge while absorbing new classes via learning negative samples. Meanwhile, IKD assesses the importance of queries based on their matching degree to old classes, prioritizing the distillation of important features and allowing less critical features to evolve. Extensive experiments in Class Incremental Panoptic and Semantic Segmentation settings have demonstrated the superior performance of CoMBO. Project page: https://guangyu-ryan.github.io/CoMBO." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 626, + 135, + 639 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 626, + 135, + 639 + ], + "spans": [ + { + "bbox": [ + 56, + 626, + 135, + 639 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 647, + 296, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 647, + 296, + 684 + ], + "spans": [ + { + "bbox": [ + 55, + 647, + 296, + 684 + ], + "type": "text", + "content": "Semantic segmentation, the fundamental task in computer vision, involves classifying each pixel into predefined categories. Panoptic segmentation, the challeng-" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 318, + 215, + 556, + 494 + ], + "blocks": [ + { + "bbox": [ + 318, + 215, + 556, + 494 + ], + "lines": [ + { + "bbox": [ + 318, + 215, + 556, + 494 + ], + "spans": [ + { + "bbox": [ + 318, + 215, + 556, + 494 + ], + "type": "image", + "image_path": "5f9cb9e87dd8c4c509c8e46890c3305acdd634a8bcadfb488ed943b6744c465a.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 504, + 555, + 594 + ], + "lines": [ + { + "bbox": [ + 313, + 504, + 555, + 594 + ], + "spans": [ + { + "bbox": [ + 313, + 504, + 555, + 594 + ], + "type": "text", + "content": "Figure 1. Comparison of our improved distillation and learning strategy (bottom) with previous conflicting strategy (top). Previous strategies impose contradictory supervision on the same target to find a balance, whereas our strategy, including the importance factor for distillation and the QCR module for adaptive learning, enables more compatible, target-specific supervision. The visualization on incremental classes ('Traffic Light') highlights the effectiveness of our strategy." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "content": "ing variant, unifies semantic and instance segmentation to both classify every pixel and simultaneously identify distinct object instances. Recent advancements in mask classification-based segmentation, such as MaskFormer [16] and Mask2Former [17], enable the unification of semantic and panoptic segmentation. However, traditional approaches to segmentation suffer from limitations in dynamic environments where new classes can emerge unpredictably, as they are typically trained on a static set" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 693, + 134, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 693, + 134, + 703 + ], + "spans": [ + { + "bbox": [ + 66, + 693, + 134, + 703 + ], + "type": "text", + "content": "*Equal contribution." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 67, + 703, + 226, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 226, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 226, + 713 + ], + "type": "text", + "content": "† Corresponding author, guangyugao@bit.edu.cn." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25667" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "content": "of categories. Therefore, Class Incremental Semantic Segmentation (CISS) has emerged as a pivotal task for continually learning new categories, where models can adapt to new classes without forgetting previously learned ones." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 122, + 295, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 122, + 295, + 326 + ], + "spans": [ + { + "bbox": [ + 55, + 122, + 295, + 326 + ], + "type": "text", + "content": "Historically, studies in CISS [3, 35, 49] mainly focuses on tackling catastrophic forgetting, employing various techniques such as knowledge distillation [5, 20, 47], freezing parameters [9, 26, 48], or generating pseudo-labels from old model [20]. More recent efforts [8, 11] have extended these techniques to the more challenging Class Incremental Panoptic Segmentation (CIPS) based on the advanced Mask2Former. Nonetheless, these methods often fail to adequately reconcile the inherent tension between retaining old knowledge and acquiring new information. Some overemphasize maintaining stability, thus hindering the learning of new classes, while others prioritize plasticity, leading to significant forgetting of old classes. It is necessary to deconstruct these conflicting objectives to facilitate a more compatible incremental learning process, achieving promotion in both the acquisition of new classes and retention of old classes." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 328, + 295, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 328, + 295, + 700 + ], + "spans": [ + { + "bbox": [ + 55, + 328, + 295, + 700 + ], + "type": "text", + "content": "In this work, we propose Conflict Mitigation via Branched Optimization (CoMBO), an approach to reduce conflicts between acquiring new classes and retaining old classes. By integrating three specialized knowledge distillation mechanisms within the Mask2Former architecture, CoMBO effectively mitigates catastrophic forgetting while enhancing the assimilation of new classes. Using bipartite matching, Mask2Former optimally aligns each query, complete with its classification logits and mask predictions, with the most fitting ground truth class, ensuring efficient optimization. Therefore, we first introduce the Half-Learning Half-Distillation mechanism, which selectively applies a shared Kullback-Leibler divergence to the classification logits of unmatched queries. This method distinctly separates distillation on unmatched masks from the classification optimization of matched ones. Moreover, rather than applying uniform distillation across all features, we introduce an importance-based knowledge distillation technique. This technique assesses each query's significance in preserving knowledge of old classes and adjusts the distillation intensity accordingly, emphasizing the retention of key queries. Additionally, our Query Conflict Reduction module innovatively adapts features for new classes while maintaining the queries' original characteristics before distillation, using class-specific adapters for each newly learned class. Finally, our approach outperforms previous methods in multiple benchmarks, especially in typical and challenging scenarios such as 100-10 of ADE20K. Our approach achieves remarkable performance of " + }, + { + "bbox": [ + 55, + 328, + 295, + 700 + ], + "type": "inline_equation", + "content": "35.6\\%" + }, + { + "bbox": [ + 55, + 328, + 295, + 700 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 328, + 295, + 700 + ], + "type": "inline_equation", + "content": "41.1\\%" + }, + { + "bbox": [ + 55, + 328, + 295, + 700 + ], + "type": "text", + "content": " mIoU on 100-10 of CIPS and CISS, respectively, compared to the previous state-of-the-art." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 701, + 285, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 701, + 285, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 701, + 285, + 712 + ], + "type": "text", + "content": "Overall, our contributions are summarized as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 314, + 72, + 553, + 216 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 314, + 72, + 553, + 119 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 553, + 119 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 553, + 119 + ], + "type": "text", + "content": "- We propose a Half-Learning Half-Distillation mechanism that applies soft distillation for old classes and binary-based optimization for new classes to effectively balance retention and acquisition." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 314, + 121, + 553, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 121, + 553, + 167 + ], + "spans": [ + { + "bbox": [ + 314, + 121, + 553, + 167 + ], + "type": "text", + "content": "- An extra Query Conflict Reduction module is proposed to refine queries for new classes while preserving unrefined features, which enables Importance-based Knowledge Distillation for remembering the key features." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 168, + 553, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 168, + 553, + 216 + ], + "spans": [ + { + "bbox": [ + 314, + 168, + 553, + 216 + ], + "type": "text", + "content": "- Through extensive quantitative and qualitative evaluations on the ADE20K dataset, we demonstrate the state-of-the-art performance of our model in both Class Incremental Semantic and Panoptic Segmentation tasks." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 228, + 400, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 228, + 400, + 239 + ], + "spans": [ + { + "bbox": [ + 314, + 228, + 400, + 239 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 248, + 511, + 261 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 248, + 511, + 261 + ], + "spans": [ + { + "bbox": [ + 314, + 248, + 511, + 261 + ], + "type": "text", + "content": "2.1. Semantic and Panoptic Segmentation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 267, + 555, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 267, + 555, + 555 + ], + "spans": [ + { + "bbox": [ + 313, + 267, + 555, + 555 + ], + "type": "text", + "content": "Semantic segmentation aims to classify each pixel in an image, which focuses on the complete coverage of object(s) in each category. The pioneering work FCN [30] and U-Net [36] introduces a pixel-wise classification paradigm by removing the fully connected layers. Since then, numerous approaches [2, 12-15, 29, 38, 41] have been proposed following the paradigm, containing effective modules such as ASPP [13] and pyramid structure [50]. In contrast, panoptic segmentation [27] combines semantic and instance segmentation, requiring the classification of every pixel while distinguishing instances of the same category. These two tasks were addressed separately due to the limitation of the pixelwise classification. However, the emergence of the mask classification paradigm [16, 17, 43, 45] introduced a universal, transformer-based framework capable of solving multiple segmentation tasks simultaneously. MaskFormer [16] pioneered this paradigm by combining class-agnostic mask proposal generation and mask classification, enabling both semantic and panoptic segmentation simultaneously. Building upon this, Mask2Former [17] enhanced performance by incorporating multi-scale feature fusion and masked attention techniques. However, when learning new categories beyond the initial data, mask classification methods remain susceptible to catastrophic forgetting." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 563, + 451, + 575 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 563, + 451, + 575 + ], + "spans": [ + { + "bbox": [ + 314, + 563, + 451, + 575 + ], + "type": "text", + "content": "2.2. Continual Segmentation" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 582, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 555, + 715 + ], + "type": "text", + "content": "In Class Incremental Learning (CIL), fine-tuning models on new data often results in a performance drop on old categories, i.e., catastrophic forgetting [22, 39]. To address this, various methods [5-7, 20, 21, 32, 33, 51] have been proposed, including retaining representative old samples [1, 4, 9, 10, 31, 32, 35, 53], compensation losses [18, 34, 42, 44], and prompt representations [25, 37, 40]. These advancements in CIL raise interest in more challenging segmentation tasks. Class Incremental Semantic Segmentation (CISS) was first introduced in MiB [5], reconstructing background regions to address background shifting during" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "25668" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 288 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 288 + ], + "type": "text", + "content": "incremental steps. Since then, most methods [3, 47] have employed techniques such as knowledge distillation [28], pseudo-labels from previous models [8, 20], and background redefinition [5, 46] to mitigate background shifting. However, these methods primarily rely on pixel-wise classification models like DeepLabV3 [13]. Therefore, CoM-Former [8] first utilizes the mask classification segmentation model Mask2Former [17] enabling solving both CISS and Class Incremental Panoptic Segmentation (CIPS) tasks. ECLIPSE [26] introduces dynamic model structure expansion methods to extend learnable parameters for new classes and freezes the old parameters. CoMasTRe [23] distills queries matched to old queries with high probability on old classes. BalConpas [11] balance the class proportion on selecting representative replay samples. However, these methods often struggle to balance acquisition and preservation. Our approach mitigates this conflict by separating queries to decouple the two objectives." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 299, + 141, + 312 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 299, + 141, + 312 + ], + "spans": [ + { + "bbox": [ + 55, + 299, + 141, + 312 + ], + "type": "text", + "content": "3. Preliminaries" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 319, + 167, + 331 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 319, + 167, + 331 + ], + "spans": [ + { + "bbox": [ + 55, + 319, + 167, + 331 + ], + "type": "text", + "content": "3.1. Problem Definition" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 338, + 296, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 338, + 296, + 533 + ], + "spans": [ + { + "bbox": [ + 55, + 338, + 296, + 533 + ], + "type": "text", + "content": "In continual segmentation, the model is tasked with an incremental learning challenge over " + }, + { + "bbox": [ + 55, + 338, + 296, + 533 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 338, + 296, + 533 + ], + "type": "text", + "content": " steps. At each step " + }, + { + "bbox": [ + 55, + 338, + 296, + 533 + ], + "type": "inline_equation", + "content": "t \\in \\{1, \\ldots, T\\}" + }, + { + "bbox": [ + 55, + 338, + 296, + 533 + ], + "type": "text", + "content": ", the model is trained to recognize a unique set of classes " + }, + { + "bbox": [ + 55, + 338, + 296, + 533 + ], + "type": "inline_equation", + "content": "C^t" + }, + { + "bbox": [ + 55, + 338, + 296, + 533 + ], + "type": "text", + "content": ", where the intersection across all sets from " + }, + { + "bbox": [ + 55, + 338, + 296, + 533 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 338, + 296, + 533 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 338, + 296, + 533 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 338, + 296, + 533 + ], + "type": "text", + "content": " is empty " + }, + { + "bbox": [ + 55, + 338, + 296, + 533 + ], + "type": "inline_equation", + "content": "\\bigcap_{t=1}^{T} C^t = \\emptyset" + }, + { + "bbox": [ + 55, + 338, + 296, + 533 + ], + "type": "text", + "content": " and their union forms the complete set of classes " + }, + { + "bbox": [ + 55, + 338, + 296, + 533 + ], + "type": "inline_equation", + "content": "\\bigcup_{t=1}^{T} C^t = \\mathcal{C}" + }, + { + "bbox": [ + 55, + 338, + 296, + 533 + ], + "type": "text", + "content": ". In the current step " + }, + { + "bbox": [ + 55, + 338, + 296, + 533 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 338, + 296, + 533 + ], + "type": "text", + "content": ", the training dataset " + }, + { + "bbox": [ + 55, + 338, + 296, + 533 + ], + "type": "inline_equation", + "content": "D_{train}^t" + }, + { + "bbox": [ + 55, + 338, + 296, + 533 + ], + "type": "text", + "content": " consists of image-label pairs " + }, + { + "bbox": [ + 55, + 338, + 296, + 533 + ], + "type": "inline_equation", + "content": "(x^t, y^t)" + }, + { + "bbox": [ + 55, + 338, + 296, + 533 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 338, + 296, + 533 + ], + "type": "inline_equation", + "content": "x^t" + }, + { + "bbox": [ + 55, + 338, + 296, + 533 + ], + "type": "text", + "content": " represents an image and " + }, + { + "bbox": [ + 55, + 338, + 296, + 533 + ], + "type": "inline_equation", + "content": "y^t" + }, + { + "bbox": [ + 55, + 338, + 296, + 533 + ], + "type": "text", + "content": " its corresponding segmentation label. The labels " + }, + { + "bbox": [ + 55, + 338, + 296, + 533 + ], + "type": "inline_equation", + "content": "y^t" + }, + { + "bbox": [ + 55, + 338, + 296, + 533 + ], + "type": "text", + "content": " are available only for the classes " + }, + { + "bbox": [ + 55, + 338, + 296, + 533 + ], + "type": "inline_equation", + "content": "C^t" + }, + { + "bbox": [ + 55, + 338, + 296, + 533 + ], + "type": "text", + "content": " that need to be learned at this step, while labels for previously learned classes " + }, + { + "bbox": [ + 55, + 338, + 296, + 533 + ], + "type": "inline_equation", + "content": "C^{1:t-1}" + }, + { + "bbox": [ + 55, + 338, + 296, + 533 + ], + "type": "text", + "content": " and future classes " + }, + { + "bbox": [ + 55, + 338, + 296, + 533 + ], + "type": "inline_equation", + "content": "C^{t+1:T}" + }, + { + "bbox": [ + 55, + 338, + 296, + 533 + ], + "type": "text", + "content": " are inaccessible. After completing training at this step, the model must perform segmentation across all classes learned up to that point, " + }, + { + "bbox": [ + 55, + 338, + 296, + 533 + ], + "type": "inline_equation", + "content": "C^{1:t}" + }, + { + "bbox": [ + 55, + 338, + 296, + 533 + ], + "type": "text", + "content": ", thereby preventing catastrophic forgetting of the old classes " + }, + { + "bbox": [ + 55, + 338, + 296, + 533 + ], + "type": "inline_equation", + "content": "C^{1:t-1}" + }, + { + "bbox": [ + 55, + 338, + 296, + 533 + ], + "type": "text", + "content": " while effectively acquiring new ones " + }, + { + "bbox": [ + 55, + 338, + 296, + 533 + ], + "type": "inline_equation", + "content": "C^t" + }, + { + "bbox": [ + 55, + 338, + 296, + 533 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 539, + 296, + 553 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 539, + 296, + 553 + ], + "spans": [ + { + "bbox": [ + 55, + 539, + 296, + 553 + ], + "type": "text", + "content": "3.2. Revisiting Mask2Former and Pseudo-Labeling" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "text", + "content": "The recent advanced universal segmentation network, Mask2Former [17], has favored the class incremental semantic and panoptic segmentation (CISS and CIPS) tasks for its innovative proposal-based structure. Mask2Former deviates from traditional pixel-wise classification pipelines by introducing a class-agnostic mask prediction and classification mechanism. Specifically, the backbone " + }, + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\pmb{f}_b" + }, + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "text", + "content": " and the pixel decoder " + }, + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\pmb{f}_p" + }, + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "text", + "content": " generate multi-scale features that interact with " + }, + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "text", + "content": " learnable queries through multiple self-attention and cross-attention layers in the transformer decoder " + }, + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\pmb{f}_t" + }, + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "text", + "content": ". These " + }, + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "text", + "content": " output queries " + }, + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "inline_equation", + "content": "Q \\in \\mathbb{R}^{C_Q \\times N}" + }, + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "text", + "content": " undergo further processing through two linear layers to form mask embeddings " + }, + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{mask} \\in \\mathbb{R}^{D \\times N}" + }, + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "text", + "content": " and class embeddings " + }, + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{cls} \\in \\mathbb{R}^{(C + 1) \\times N}" + }, + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "content": "The mask embeddings " + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{mask}" + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "content": " yield " + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "content": " mask predictions " + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "inline_equation", + "content": "M\\in \\mathbb{R}^{N\\times H\\times W}" + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "content": " via dot product between each mask embedding and per-pixel embeddings " + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{pixel}\\in \\mathbb{R}^{D\\times H\\times W}" + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "inline_equation", + "content": "f_{p}" + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "content": ", while " + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{cls}" + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "content": " accounts for the classification predictions of these masks. Note that the additional channel in " + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{cls}" + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "content": " represents the no-object class, an auxiliary category utilized during training but excluded during inference." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 157, + 555, + 330 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 157, + 555, + 330 + ], + "spans": [ + { + "bbox": [ + 313, + 157, + 555, + 330 + ], + "type": "text", + "content": "In the contexts of CISS and CIPS, where ground truth for old classes " + }, + { + "bbox": [ + 313, + 157, + 555, + 330 + ], + "type": "inline_equation", + "content": "C^{1:t-1}" + }, + { + "bbox": [ + 313, + 157, + 555, + 330 + ], + "type": "text", + "content": " are inaccessible, pseudo-labeling becomes crucial. PLOP [20] introduces this strategy by generating pixel-level pseudo ground truths for " + }, + { + "bbox": [ + 313, + 157, + 555, + 330 + ], + "type": "inline_equation", + "content": "C^{1:t-1}" + }, + { + "bbox": [ + 313, + 157, + 555, + 330 + ], + "type": "text", + "content": ", adapted to the mask-based structure by CoMFormer [8], the first method based on Mask2Former for CISS and CIPS tasks. This adaptation involves weighting the mask predictions " + }, + { + "bbox": [ + 313, + 157, + 555, + 330 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 157, + 555, + 330 + ], + "type": "text", + "content": " by the corresponding maximum class embedding score after the softmax process over the " + }, + { + "bbox": [ + 313, + 157, + 555, + 330 + ], + "type": "inline_equation", + "content": "C+1" + }, + { + "bbox": [ + 313, + 157, + 555, + 330 + ], + "type": "text", + "content": " dimensions, i.e., " + }, + { + "bbox": [ + 313, + 157, + 555, + 330 + ], + "type": "inline_equation", + "content": "\\max_{c=0}^{|C^{0:t-1}|} (\\operatorname{softmax}_{c=0}^{|C^{0:t-1}|} (\\mathcal{E}_{cls}))" + }, + { + "bbox": [ + 313, + 157, + 555, + 330 + ], + "type": "text", + "content": " to form " + }, + { + "bbox": [ + 313, + 157, + 555, + 330 + ], + "type": "inline_equation", + "content": "M_{\\omega}" + }, + { + "bbox": [ + 313, + 157, + 555, + 330 + ], + "type": "text", + "content": ", with associated categories " + }, + { + "bbox": [ + 313, + 157, + 555, + 330 + ], + "type": "inline_equation", + "content": "C_{\\omega} = \\arg \\max_{c=0}^{|C^{0:t-1}|} \\mathcal{E}_{cls} \\in \\mathbb{R}^{N}" + }, + { + "bbox": [ + 313, + 157, + 555, + 330 + ], + "type": "text", + "content": ". The pseudo-labeling process then labels the pixels outside the union of " + }, + { + "bbox": [ + 313, + 157, + 555, + 330 + ], + "type": "inline_equation", + "content": "\\boldsymbol{y}^t" + }, + { + "bbox": [ + 313, + 157, + 555, + 330 + ], + "type": "text", + "content": " with the category " + }, + { + "bbox": [ + 313, + 157, + 555, + 330 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 157, + 555, + 330 + ], + "type": "text", + "content": " that achieves the maximum score across " + }, + { + "bbox": [ + 313, + 157, + 555, + 330 + ], + "type": "inline_equation", + "content": "M_{\\omega}" + }, + { + "bbox": [ + 313, + 157, + 555, + 330 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 157, + 555, + 330 + ], + "type": "inline_equation", + "content": "C_{w}" + }, + { + "bbox": [ + 313, + 157, + 555, + 330 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 318, + 335, + 554, + 396 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 335, + 554, + 396 + ], + "spans": [ + { + "bbox": [ + 318, + 335, + 554, + 396 + ], + "type": "interline_equation", + "content": "\\tilde {M} _ {c} (h, w) = \\left\\{ \\begin{array}{l l} 1 & \\text {i f} c = C _ {\\omega} [ \\arg \\max _ {n = 1} ^ {N} M _ {w} (n, h, w) ] \\\\ & \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\quad \\text {v a r e p s i l o n} \\\\ 0 & \\text {e l s e} \\end{array} \\right.", + "image_path": "1d8d063cb210fc47399dbf50e1a52d3b850d1b1e354e47a656485f3499380b51.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 396, + 554, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 396, + 554, + 431 + ], + "spans": [ + { + "bbox": [ + 313, + 396, + 554, + 431 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 396, + 554, + 431 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 396, + 554, + 431 + ], + "type": "text", + "content": " indexes the " + }, + { + "bbox": [ + 313, + 396, + 554, + 431 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 396, + 554, + 431 + ], + "type": "text", + "content": " proposals, and " + }, + { + "bbox": [ + 313, + 396, + 554, + 431 + ], + "type": "inline_equation", + "content": "h, w" + }, + { + "bbox": [ + 313, + 396, + 554, + 431 + ], + "type": "text", + "content": " are the pixel coordinates. This pseudo-labeling strategy has become foundational for CISS and CIPS methods [8]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 441, + 421, + 454 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 441, + 421, + 454 + ], + "spans": [ + { + "bbox": [ + 313, + 441, + 421, + 454 + ], + "type": "text", + "content": "4. Proposed Method" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 461, + 554, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 461, + 554, + 534 + ], + "spans": [ + { + "bbox": [ + 313, + 461, + 554, + 534 + ], + "type": "text", + "content": "In our proposed Conflict Mitigation via Branched Optimization (CoMBO), we introduce three distinct strategies: Half-Distillation-Half-Learning Strategy for targeted knowledge retention, Importance-Based Knowledge Distillation to prioritize crucial features, and Query Conflict Reduction for efficient class integration." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 540, + 457, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 540, + 457, + 552 + ], + "spans": [ + { + "bbox": [ + 313, + 540, + 457, + 552 + ], + "type": "text", + "content": "4.1. Query Conflict Reduction" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 557, + 554, + 631 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 557, + 554, + 631 + ], + "spans": [ + { + "bbox": [ + 313, + 557, + 554, + 631 + ], + "type": "text", + "content": "As mentioned in Sec. 3.2, both class embeddings " + }, + { + "bbox": [ + 313, + 557, + 554, + 631 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{cls}" + }, + { + "bbox": [ + 313, + 557, + 554, + 631 + ], + "type": "text", + "content": " and mask predictions " + }, + { + "bbox": [ + 313, + 557, + 554, + 631 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 557, + 554, + 631 + ], + "type": "text", + "content": " are generated from the queries " + }, + { + "bbox": [ + 313, + 557, + 554, + 631 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 313, + 557, + 554, + 631 + ], + "type": "text", + "content": ". This interface becomes a critical juncture of conflicts between retaining knowledge of old classes and acquiring new classes manifest, which cannot break through the bottleneck of overall performance by simply finding the balance point." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 642, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 554, + 713 + ], + "type": "text", + "content": "To deconstruct the crossroad efficiently, we design a Query Conflict Reduction (QCR) module " + }, + { + "bbox": [ + 313, + 642, + 554, + 713 + ], + "type": "inline_equation", + "content": "f_{QCR}(\\cdot)" + }, + { + "bbox": [ + 313, + 642, + 554, + 713 + ], + "type": "text", + "content": " as an additional adapter for new classes. As shown in Fig. 2, in most scenarios, queries from the previous layer " + }, + { + "bbox": [ + 313, + 642, + 554, + 713 + ], + "type": "inline_equation", + "content": "Q_{l-1}" + }, + { + "bbox": [ + 313, + 642, + 554, + 713 + ], + "type": "text", + "content": " have the same classification results as the queries from the current layer due to the similar distribution and the same clas" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25669" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 72, + 282, + 247 + ], + "blocks": [ + { + "bbox": [ + 62, + 72, + 282, + 247 + ], + "lines": [ + { + "bbox": [ + 62, + 72, + 282, + 247 + ], + "spans": [ + { + "bbox": [ + 62, + 72, + 282, + 247 + ], + "type": "image", + "image_path": "127ba02410a6d6f9244f3905c6ea692fdacf93ac171a7a254f79cedd063dc8c9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 255, + 296, + 310 + ], + "lines": [ + { + "bbox": [ + 55, + 255, + 296, + 310 + ], + "spans": [ + { + "bbox": [ + 55, + 255, + 296, + 310 + ], + "type": "text", + "content": "Figure 2. Illustration of the Query Conflict Reduction (QCR) module. This module refines queries that predict incremental classes, allowing the processes of learning new features and retaining old features to occur separately. Note that the QCR module for previous incremental classes is frozen." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 320, + 296, + 428 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 320, + 296, + 428 + ], + "spans": [ + { + "bbox": [ + 55, + 320, + 296, + 428 + ], + "type": "text", + "content": "sification layer, especially in the latter queries. Due to the shared classifier and mask generator following different layers of the transformer decoder, the classification results on the same position of queries from adjacent layers are approximately the same. Thus, we add the adapter for the queries from the penultimate layer " + }, + { + "bbox": [ + 55, + 320, + 296, + 428 + ], + "type": "inline_equation", + "content": "Q_{L-1}" + }, + { + "bbox": [ + 55, + 320, + 296, + 428 + ], + "type": "text", + "content": " with classification results of new classes " + }, + { + "bbox": [ + 55, + 320, + 296, + 428 + ], + "type": "inline_equation", + "content": "C^{2:t}" + }, + { + "bbox": [ + 55, + 320, + 296, + 428 + ], + "type": "text", + "content": ", as shown in Fig. 2. To be more specific, the queries " + }, + { + "bbox": [ + 55, + 320, + 296, + 428 + ], + "type": "inline_equation", + "content": "Q_{L-1,\\tilde{c}}" + }, + { + "bbox": [ + 55, + 320, + 296, + 428 + ], + "type": "text", + "content": " are selected for the adapter of incremental class " + }, + { + "bbox": [ + 55, + 320, + 296, + 428 + ], + "type": "inline_equation", + "content": "\\tilde{c} \\in C^{2:t}" + }, + { + "bbox": [ + 55, + 320, + 296, + 428 + ], + "type": "text", + "content": " via:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 432, + 295, + 449 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 432, + 295, + 449 + ], + "spans": [ + { + "bbox": [ + 61, + 432, + 295, + 449 + ], + "type": "interline_equation", + "content": "Q _ {L - 1, \\tilde {c}} = \\left\\{Q _ {L - 1} (n), n \\in N \\wedge \\tilde {c} = \\arg \\max _ {c = 0} ^ {| C ^ {0: t} |} \\mathcal {E} _ {c l s} (n) \\right\\}, \\tag {2}", + "image_path": "e89c8311ce7d3ca1897bf8f30b4c387230391004684bab9e528f6130cae51936.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 453, + 296, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 453, + 296, + 525 + ], + "spans": [ + { + "bbox": [ + 55, + 453, + 296, + 525 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 453, + 296, + 525 + ], + "type": "inline_equation", + "content": "Q_{L-1,\\tilde{c}}" + }, + { + "bbox": [ + 55, + 453, + 296, + 525 + ], + "type": "text", + "content": " could contain none, one, or some selected queries that enable recognizing different instances. For each new category, we utilize the corresponding QCR module " + }, + { + "bbox": [ + 55, + 453, + 296, + 525 + ], + "type": "inline_equation", + "content": "f_{QCR,\\tilde{c}}(\\cdot)" + }, + { + "bbox": [ + 55, + 453, + 296, + 525 + ], + "type": "text", + "content": " to refine the group of queries " + }, + { + "bbox": [ + 55, + 453, + 296, + 525 + ], + "type": "inline_equation", + "content": "Q_{L-1,\\tilde{c}}" + }, + { + "bbox": [ + 55, + 453, + 296, + 525 + ], + "type": "text", + "content": ". Considering the efficiency and effectiveness, we introduce the low-rank two-layer adaptation as the QCR module:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 100, + 531, + 294, + 560 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 531, + 294, + 560 + ], + "spans": [ + { + "bbox": [ + 100, + 531, + 294, + 560 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\hat {Q} _ {L - 1, \\tilde {c}} = f _ {Q C R, \\tilde {c}} \\left(Q _ {L - 1, \\tilde {c}}\\right) \\tag {3} \\\\ = Q _ {L - 1, \\tilde {c}} W _ {1} W _ {2} + Q _ {L - 1, \\tilde {c}}, \\\\ \\end{array}", + "image_path": "662da37a3003651cbb8b436020a71812971df4fa567add01c0d3774babed0f61.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 568, + 296, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 568, + 296, + 653 + ], + "spans": [ + { + "bbox": [ + 55, + 568, + 296, + 653 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 568, + 296, + 653 + ], + "type": "inline_equation", + "content": "\\hat{Q}_{L - 1,\\tilde{c}}" + }, + { + "bbox": [ + 55, + 568, + 296, + 653 + ], + "type": "text", + "content": " represent the adapted queries of class " + }, + { + "bbox": [ + 55, + 568, + 296, + 653 + ], + "type": "inline_equation", + "content": "\\tilde{c}" + }, + { + "bbox": [ + 55, + 568, + 296, + 653 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 568, + 296, + 653 + ], + "type": "inline_equation", + "content": "W_{1} \\in \\mathbb{R}^{D \\times r}" + }, + { + "bbox": [ + 55, + 568, + 296, + 653 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 568, + 296, + 653 + ], + "type": "inline_equation", + "content": "W_{2} \\in \\mathbb{R}^{r \\times D}" + }, + { + "bbox": [ + 55, + 568, + 296, + 653 + ], + "type": "text", + "content": " denotes the weights for QCR. The QCR module separates the refined queries " + }, + { + "bbox": [ + 55, + 568, + 296, + 653 + ], + "type": "inline_equation", + "content": "\\hat{Q}_{L,\\tilde{c}}" + }, + { + "bbox": [ + 55, + 568, + 296, + 653 + ], + "type": "text", + "content": " from the original queries " + }, + { + "bbox": [ + 55, + 568, + 296, + 653 + ], + "type": "inline_equation", + "content": "Q_{L,\\tilde{c}}" + }, + { + "bbox": [ + 55, + 568, + 296, + 653 + ], + "type": "text", + "content": ", enabling a bifurcate structure for simultaneously learning " + }, + { + "bbox": [ + 55, + 568, + 296, + 653 + ], + "type": "inline_equation", + "content": "C^t" + }, + { + "bbox": [ + 55, + 568, + 296, + 653 + ], + "type": "text", + "content": " with class-specific adaptation and keep memorizing the features of " + }, + { + "bbox": [ + 55, + 568, + 296, + 653 + ], + "type": "inline_equation", + "content": "Q_{L,\\tilde{c}}" + }, + { + "bbox": [ + 55, + 568, + 296, + 653 + ], + "type": "text", + "content": " as well as the logits of " + }, + { + "bbox": [ + 55, + 568, + 296, + 653 + ], + "type": "inline_equation", + "content": "C^{0:t - 1}" + }, + { + "bbox": [ + 55, + 568, + 296, + 653 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 659, + 268, + 673 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 659, + 268, + 673 + ], + "spans": [ + { + "bbox": [ + 55, + 659, + 268, + 673 + ], + "type": "text", + "content": "4.2. Half-Distillation-Half-Learning Strategy" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": "Previous approaches based on Mask2Former usually separate learning and distillation processes on class embeddings " + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{cls}" + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": " due to its different learning strategies compared" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 323, + 72, + 541, + 232 + ], + "blocks": [ + { + "bbox": [ + 323, + 72, + 541, + 232 + ], + "lines": [ + { + "bbox": [ + 323, + 72, + 541, + 232 + ], + "spans": [ + { + "bbox": [ + 323, + 72, + 541, + 232 + ], + "type": "image", + "image_path": "9c5c7aa6edd5c3bd35306df3176cf547cf072e7981c9d23ae82d3c06a9327b9a.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 242, + 555, + 286 + ], + "lines": [ + { + "bbox": [ + 313, + 242, + 555, + 286 + ], + "spans": [ + { + "bbox": [ + 313, + 242, + 555, + 286 + ], + "type": "text", + "content": "Figure 3. Details of the Half-Distillation-Half-Learning strategy on " + }, + { + "bbox": [ + 313, + 242, + 555, + 286 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 242, + 555, + 286 + ], + "type": "text", + "content": " queries, including classification loss on matched queries and Kullback-Leibler Divergence loss on the other queries. The latter involves the logits of both old and current classes." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 304, + 554, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 304, + 554, + 365 + ], + "spans": [ + { + "bbox": [ + 313, + 304, + 554, + 365 + ], + "type": "text", + "content": "to pixel-wise segmentation structure (e.g. DeepLabv3 [13]), mentioned in Sec. 3.2. This causes conflicts between distillation and learning, where the same logit could have contradictory optimization directions. Therefore, it is necessary to unify the optimization process towards a fixed target." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 366, + 555, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 366, + 555, + 576 + ], + "spans": [ + { + "bbox": [ + 313, + 366, + 555, + 576 + ], + "type": "text", + "content": "Our Half-Distillation-Half-Learning (HDHL) strategy combines distillation and learning for " + }, + { + "bbox": [ + 313, + 366, + 555, + 576 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 366, + 555, + 576 + ], + "type": "text", + "content": " class embeddings " + }, + { + "bbox": [ + 313, + 366, + 555, + 576 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{cls}" + }, + { + "bbox": [ + 313, + 366, + 555, + 576 + ], + "type": "text", + "content": ", as illustrated in Fig. 3. We follow the bipartite matching of Mask2Former and select embeddings " + }, + { + "bbox": [ + 313, + 366, + 555, + 576 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{cls}^{C^t}" + }, + { + "bbox": [ + 313, + 366, + 555, + 576 + ], + "type": "text", + "content": " matched to labels of current categories " + }, + { + "bbox": [ + 313, + 366, + 555, + 576 + ], + "type": "inline_equation", + "content": "C^t" + }, + { + "bbox": [ + 313, + 366, + 555, + 576 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 366, + 555, + 576 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{cls}^{C^{1:t-1}}" + }, + { + "bbox": [ + 313, + 366, + 555, + 576 + ], + "type": "text", + "content": " matched to pseudo labels of old categories " + }, + { + "bbox": [ + 313, + 366, + 555, + 576 + ], + "type": "inline_equation", + "content": "C^{1:t-1}" + }, + { + "bbox": [ + 313, + 366, + 555, + 576 + ], + "type": "text", + "content": ". These embeddings, following the original optimization strategy, are optimized by cross-entropy loss " + }, + { + "bbox": [ + 313, + 366, + 555, + 576 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{cls}" + }, + { + "bbox": [ + 313, + 366, + 555, + 576 + ], + "type": "text", + "content": ". The unmatched embeddings " + }, + { + "bbox": [ + 313, + 366, + 555, + 576 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{cls}^{\\emptyset^t}" + }, + { + "bbox": [ + 313, + 366, + 555, + 576 + ], + "type": "text", + "content": " are grouped under the no-object (no-" + }, + { + "bbox": [ + 313, + 366, + 555, + 576 + ], + "type": "inline_equation", + "content": "obj" + }, + { + "bbox": [ + 313, + 366, + 555, + 576 + ], + "type": "text", + "content": ") category. However, adhering to the original learning strategy leads to catastrophic forgetting. Besides, using the distillation strategy for old categories " + }, + { + "bbox": [ + 313, + 366, + 555, + 576 + ], + "type": "inline_equation", + "content": "C^{1:t-1}" + }, + { + "bbox": [ + 313, + 366, + 555, + 576 + ], + "type": "text", + "content": " using " + }, + { + "bbox": [ + 313, + 366, + 555, + 576 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{cls}^{\\emptyset^{t-1}}" + }, + { + "bbox": [ + 313, + 366, + 555, + 576 + ], + "type": "text", + "content": " ignores the learning of current categories " + }, + { + "bbox": [ + 313, + 366, + 555, + 576 + ], + "type": "inline_equation", + "content": "C^t" + }, + { + "bbox": [ + 313, + 366, + 555, + 576 + ], + "type": "text", + "content": ". However, identifying regions that do not belong to " + }, + { + "bbox": [ + 313, + 366, + 555, + 576 + ], + "type": "inline_equation", + "content": "C^t" + }, + { + "bbox": [ + 313, + 366, + 555, + 576 + ], + "type": "text", + "content": " is just as important as learning those that do. Considering this, we design a strategy to simultaneously optimize the logits of " + }, + { + "bbox": [ + 313, + 366, + 555, + 576 + ], + "type": "inline_equation", + "content": "C^{1:t-1}" + }, + { + "bbox": [ + 313, + 366, + 555, + 576 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 366, + 555, + 576 + ], + "type": "inline_equation", + "content": "C^t" + }, + { + "bbox": [ + 313, + 366, + 555, + 576 + ], + "type": "text", + "content": " with a shared Kullback-Leibler Divergence " + }, + { + "bbox": [ + 313, + 366, + 555, + 576 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{kl}" + }, + { + "bbox": [ + 313, + 366, + 555, + 576 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 321, + 590, + 555, + 628 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 590, + 555, + 628 + ], + "spans": [ + { + "bbox": [ + 321, + 590, + 555, + 628 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {k l} = \\frac {1}{| \\mathcal {E} _ {c l s} ^ {\\varnothing^ {t}} |} \\sum_ {j = 1} ^ {| \\mathcal {E} _ {c l s} ^ {\\varnothing^ {t}} |} \\sum_ {c = 0} ^ {| \\boldsymbol {C} ^ {0: t} |} \\varphi_ {c l s} ^ {\\varnothing^ {t - 1}} (j, c) \\log \\frac {\\varphi_ {c l s} ^ {\\varnothing^ {t - 1}} (j , c)}{\\varphi_ {c l s} ^ {\\varnothing^ {t}} (j , c)}, \\tag {4}", + "image_path": "a8eb2b86a647debff14e6a110667b2c39c832b800abd5339a85933aa965a52a4.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 315, + 643, + 342, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 643, + 342, + 653 + ], + "spans": [ + { + "bbox": [ + 315, + 643, + 342, + 653 + ], + "type": "text", + "content": "where" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 345, + 666, + 553, + 716 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 666, + 553, + 716 + ], + "spans": [ + { + "bbox": [ + 345, + 666, + 553, + 716 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\varphi_ {c l s} ^ {\\emptyset^ {t}} (j, c) = \\mathcal {S} ^ {t} (j, c), \\\\ \\varphi_ {c l s} ^ {\\emptyset^ {t - 1}} (j, c) = \\left\\{ \\begin{array}{l l} \\mathcal {S} ^ {t - 1} (j, c) & \\text {i f} c \\in C ^ {0: t - 1}, \\\\ 0 & \\text {e l s e}, \\end{array} \\right. \\tag {5} \\\\ \\end{array}", + "image_path": "da4ea7acf4f66d5dfd9c898c5e14785a3a69ca3e4dbde521fe5e3aba82f6bd3f.jpg" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "25670" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 73, + 73, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 73, + 73, + 83 + ], + "spans": [ + { + "bbox": [ + 55, + 73, + 73, + 83 + ], + "type": "text", + "content": "and" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 104, + 83, + 295, + 116 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 83, + 295, + 116 + ], + "spans": [ + { + "bbox": [ + 104, + 83, + 295, + 116 + ], + "type": "interline_equation", + "content": "\\mathcal {S} ^ {t} (j, c) = \\frac {\\exp \\left(\\mathcal {E} _ {c l s} ^ {\\varnothing^ {t}} (j , c)\\right)}{\\sum_ {\\hat {c} = 0} ^ {| \\mathcal {C} ^ {0 : t} |} \\exp \\left(\\mathcal {E} _ {c l s} ^ {\\varnothing^ {t}} (j , \\hat {c})\\right)} \\tag {6}", + "image_path": "a607240177bdb6c6d17987a1a77bb54ee3adbefd20db4270d59ea7c54274242f.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 121, + 191, + 132 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 121, + 191, + 132 + ], + "spans": [ + { + "bbox": [ + 55, + 121, + 191, + 132 + ], + "type": "text", + "content": "represent the softmax activation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 133, + 296, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 133, + 296, + 243 + ], + "spans": [ + { + "bbox": [ + 55, + 133, + 296, + 243 + ], + "type": "text", + "content": "The " + }, + { + "bbox": [ + 55, + 133, + 296, + 243 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{kl}" + }, + { + "bbox": [ + 55, + 133, + 296, + 243 + ], + "type": "text", + "content": " not only distill the previous class embedding " + }, + { + "bbox": [ + 55, + 133, + 296, + 243 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{cls}^{\\emptyset^{t-1}}" + }, + { + "bbox": [ + 55, + 133, + 296, + 243 + ], + "type": "text", + "content": " to the corresponding categories, including the no-object category, but decreasing the logits of current categories to figure out the objects apart from them. The unified HDHL strategy maintains the probability distribution of old logits seamlessly with the integration of new logits. By combining the original " + }, + { + "bbox": [ + 55, + 133, + 296, + 243 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{cls}" + }, + { + "bbox": [ + 55, + 133, + 296, + 243 + ], + "type": "text", + "content": " and the designed " + }, + { + "bbox": [ + 55, + 133, + 296, + 243 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{kl}" + }, + { + "bbox": [ + 55, + 133, + 296, + 243 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 55, + 133, + 296, + 243 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{DL}" + }, + { + "bbox": [ + 55, + 133, + 296, + 243 + ], + "type": "text", + "content": ", the optimization process covers all the class embeddings without contradictory targets:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 111, + 254, + 295, + 268 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 254, + 295, + 268 + ], + "spans": [ + { + "bbox": [ + 111, + 254, + 295, + 268 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {D L} = \\lambda_ {c l s} \\cdot \\mathcal {L} _ {c l s} + \\lambda_ {k l} \\cdot \\mathcal {L} _ {k l}, \\tag {7}", + "image_path": "aadfc76ec2b2b7b9c706d8ef1ce65be14f5beb114cedfb0b93d1597055533e85.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 277, + 294, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 277, + 294, + 289 + ], + "spans": [ + { + "bbox": [ + 55, + 277, + 294, + 289 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 277, + 294, + 289 + ], + "type": "inline_equation", + "content": "\\lambda_{cls}" + }, + { + "bbox": [ + 55, + 277, + 294, + 289 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 277, + 294, + 289 + ], + "type": "inline_equation", + "content": "\\lambda_{kl}" + }, + { + "bbox": [ + 55, + 277, + 294, + 289 + ], + "type": "text", + "content": " control the balance between two losses." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 299, + 274, + 312 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 299, + 274, + 312 + ], + "spans": [ + { + "bbox": [ + 55, + 299, + 274, + 312 + ], + "type": "text", + "content": "4.3. Importance-Based Knowledge Distillation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 318, + 296, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 318, + 296, + 402 + ], + "spans": [ + { + "bbox": [ + 55, + 318, + 296, + 402 + ], + "type": "text", + "content": "Previous methods mitigate catastrophic forgetting by applying feature distillation, which may hinder the acquisition of new knowledge. To address this, we propose an Importance-Based Knowledge Distillation (IKD) method, in which stronger distillation is applied to features critical for old classes, while less emphasis is placed on less pertinent features, thereby balancing stability with plasticity." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 403, + 296, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 403, + 296, + 499 + ], + "spans": [ + { + "bbox": [ + 55, + 403, + 296, + 499 + ], + "type": "text", + "content": "The importance measurement requires the cost matrix " + }, + { + "bbox": [ + 55, + 403, + 296, + 499 + ], + "type": "inline_equation", + "content": "\\mathcal{A} \\in \\mathbb{R}^{N \\times S}" + }, + { + "bbox": [ + 55, + 403, + 296, + 499 + ], + "type": "text", + "content": ", adopted in the bipartite matching process, where " + }, + { + "bbox": [ + 55, + 403, + 296, + 499 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 55, + 403, + 296, + 499 + ], + "type": "text", + "content": " is the number of image labels. Values in " + }, + { + "bbox": [ + 55, + 403, + 296, + 499 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 55, + 403, + 296, + 499 + ], + "type": "text", + "content": " indicate the relationship between queries and classes. For each image, the transformer decoder generates " + }, + { + "bbox": [ + 55, + 403, + 296, + 499 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 403, + 296, + 499 + ], + "type": "text", + "content": " predicted segments, denoted as " + }, + { + "bbox": [ + 55, + 403, + 296, + 499 + ], + "type": "inline_equation", + "content": "\\mathcal{Z} = \\left\\{(\\mathcal{E}_{cls}(n), M(n))\\right\\}_{n=1}^{N}" + }, + { + "bbox": [ + 55, + 403, + 296, + 499 + ], + "type": "text", + "content": ". The cost matrix " + }, + { + "bbox": [ + 55, + 403, + 296, + 499 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 55, + 403, + 296, + 499 + ], + "type": "text", + "content": " is then computed based on the output segments " + }, + { + "bbox": [ + 55, + 403, + 296, + 499 + ], + "type": "inline_equation", + "content": "\\mathcal{Z}" + }, + { + "bbox": [ + 55, + 403, + 296, + 499 + ], + "type": "text", + "content": " and the image labels " + }, + { + "bbox": [ + 55, + 403, + 296, + 499 + ], + "type": "inline_equation", + "content": "\\boldsymbol{y}^t = \\{(c_s, m_s)\\}_{s=1}^S" + }, + { + "bbox": [ + 55, + 403, + 296, + 499 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 509, + 303, + 531 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 509, + 303, + 531 + ], + "spans": [ + { + "bbox": [ + 55, + 509, + 303, + 531 + ], + "type": "interline_equation", + "content": "\\mathcal {A} (n, s) = - \\lambda_ {c l s} \\cdot \\varphi_ {c l s} \\left(n, c _ {s}\\right) + \\lambda_ {\\text {m a s k}} \\cdot \\mathcal {L} _ {\\text {m a s k}} \\left(\\mathcal {E} _ {\\text {m a s k}} (n), m _ {s}\\right), \\tag {8}", + "image_path": "17e08bbb8423c77a2cfe42e149a370283cdcad534e41833adb276e717704d355.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 533, + 296, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 533, + 296, + 569 + ], + "spans": [ + { + "bbox": [ + 55, + 533, + 296, + 569 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 533, + 296, + 569 + ], + "type": "inline_equation", + "content": "\\varphi_{cls}" + }, + { + "bbox": [ + 55, + 533, + 296, + 569 + ], + "type": "text", + "content": " is the classification output obtained by applying softmax to " + }, + { + "bbox": [ + 55, + 533, + 296, + 569 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{cls}" + }, + { + "bbox": [ + 55, + 533, + 296, + 569 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 533, + 296, + 569 + ], + "type": "inline_equation", + "content": "\\lambda_{cls}" + }, + { + "bbox": [ + 55, + 533, + 296, + 569 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 533, + 296, + 569 + ], + "type": "inline_equation", + "content": "\\lambda_{mask}" + }, + { + "bbox": [ + 55, + 533, + 296, + 569 + ], + "type": "text", + "content": " represent the corresponding hyperparameters." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "content": "After each training stage, we compute the cost matrix " + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "content": " across every image in the training set " + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "inline_equation", + "content": "D_{train}^{t}" + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "content": " of the current incremental step " + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "content": ". The minimum costs among the classes for " + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "content": " segments are accumulated in the corresponding position of buffer " + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "inline_equation", + "content": "B^{t} \\in \\mathbb{R}^{N}" + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "content": ", where segments with lower minimum cost values are more likely to recognize and cover the regions of a class " + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "inline_equation", + "content": "C^{t}" + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "content": ". By normalizing and reversing the values of " + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "inline_equation", + "content": "B^{t}" + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "content": ", we derive the importance matrix " + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "inline_equation", + "content": "I_{C^{t}}^{t} \\in \\mathbb{R}^{N}" + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "content": " of current classes " + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "inline_equation", + "content": "C^{t}" + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "content": ". The importance matrix " + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "inline_equation", + "content": "I^{t+1} \\in \\mathbb{R}^{N}" + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "content": " is further estimated via the weighted accumulation of current importance matrix " + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "inline_equation", + "content": "I^{t} \\in \\mathbb{R}^{N}" + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "inline_equation", + "content": "I_{C^{t}}^{t}" + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "content": ", where the weights refer to the " + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "inline_equation", + "content": "|C^{1:t-1}|" + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "inline_equation", + "content": "|C^{t}|" + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "content": ". The generated " + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "inline_equation", + "content": "I^{t+1}" + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "content": " will be" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "type": "text", + "content": "utilized in the next incremental training phase for weighted distillation of the queries " + }, + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "type": "inline_equation", + "content": "Q^t" + }, + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "type": "text", + "content": ", which is computed as follows:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 342, + 105, + 555, + 138 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 342, + 105, + 555, + 138 + ], + "spans": [ + { + "bbox": [ + 342, + 105, + 555, + 138 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {I K D} = \\frac {1}{N} \\sum_ {n = 1} ^ {N} I ^ {t} (n) \\cdot \\| Q ^ {t} (n), Q ^ {t - 1} (n) \\| _ {2} ^ {2}, \\tag {9}", + "image_path": "63f84711f17289052c13561765baec60ee7ba6cacec62b898ba9107f8b0f677c.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 146, + 555, + 208 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 146, + 555, + 208 + ], + "spans": [ + { + "bbox": [ + 313, + 146, + 555, + 208 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 146, + 555, + 208 + ], + "type": "inline_equation", + "content": "\\| \\cdot \\|_2" + }, + { + "bbox": [ + 313, + 146, + 555, + 208 + ], + "type": "text", + "content": " denotes the Euclidean distance, " + }, + { + "bbox": [ + 313, + 146, + 555, + 208 + ], + "type": "inline_equation", + "content": "Q^t(n)" + }, + { + "bbox": [ + 313, + 146, + 555, + 208 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 146, + 555, + 208 + ], + "type": "inline_equation", + "content": "Q^{t-1}(n)" + }, + { + "bbox": [ + 313, + 146, + 555, + 208 + ], + "type": "text", + "content": " denotes the transformer decoder features from the current model " + }, + { + "bbox": [ + 313, + 146, + 555, + 208 + ], + "type": "inline_equation", + "content": "f^t(\\cdot)" + }, + { + "bbox": [ + 313, + 146, + 555, + 208 + ], + "type": "text", + "content": " and the old model " + }, + { + "bbox": [ + 313, + 146, + 555, + 208 + ], + "type": "inline_equation", + "content": "f^{t-1}(\\cdot)" + }, + { + "bbox": [ + 313, + 146, + 555, + 208 + ], + "type": "text", + "content": ", respectively. Algorithm 1 presents the details of the importance estimation procedure after training " + }, + { + "bbox": [ + 313, + 146, + 555, + 208 + ], + "type": "inline_equation", + "content": "f^t(\\cdot)" + }, + { + "bbox": [ + 313, + 146, + 555, + 208 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "type": "code", + "bbox": [ + 319, + 234, + 555, + 403 + ], + "blocks": [ + { + "bbox": [ + 315, + 219, + 493, + 231 + ], + "lines": [ + { + "bbox": [ + 315, + 219, + 493, + 231 + ], + "spans": [ + { + "bbox": [ + 315, + 219, + 493, + 231 + ], + "type": "text", + "content": "Algorithm 1 Importance Matrix Estimation" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 319, + 234, + 555, + 403 + ], + "lines": [ + { + "bbox": [ + 319, + 234, + 555, + 403 + ], + "spans": [ + { + "bbox": [ + 319, + 234, + 555, + 403 + ], + "type": "text", + "content": "1: Input: current model " + }, + { + "bbox": [ + 319, + 234, + 555, + 403 + ], + "type": "inline_equation", + "content": "f^{t}(\\cdot)" + }, + { + "bbox": [ + 319, + 234, + 555, + 403 + ], + "type": "text", + "content": ", current dataset " + }, + { + "bbox": [ + 319, + 234, + 555, + 403 + ], + "type": "inline_equation", + "content": "D_{train}^{t}" + }, + { + "bbox": [ + 319, + 234, + 555, + 403 + ], + "type": "text", + "content": ", current importance matrix " + }, + { + "bbox": [ + 319, + 234, + 555, + 403 + ], + "type": "inline_equation", + "content": "I^{t}" + }, + { + "bbox": [ + 319, + 234, + 555, + 403 + ], + "type": "text", + "content": " \n2: Initialization: " + }, + { + "bbox": [ + 319, + 234, + 555, + 403 + ], + "type": "inline_equation", + "content": "I^{t+1} \\gets 0" + }, + { + "bbox": [ + 319, + 234, + 555, + 403 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 319, + 234, + 555, + 403 + ], + "type": "inline_equation", + "content": "B^{t} \\gets 0" + }, + { + "bbox": [ + 319, + 234, + 555, + 403 + ], + "type": "text", + "content": " \n3: for all " + }, + { + "bbox": [ + 319, + 234, + 555, + 403 + ], + "type": "inline_equation", + "content": "(\\boldsymbol{x}^{t}, \\boldsymbol{y}^{t}) \\in D_{train}^{t}" + }, + { + "bbox": [ + 319, + 234, + 555, + 403 + ], + "type": "text", + "content": " do \n4: Compute " + }, + { + "bbox": [ + 319, + 234, + 555, + 403 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 319, + 234, + 555, + 403 + ], + "type": "text", + "content": " following Eq. (8) \n5: Update the buffer: \n6: " + }, + { + "bbox": [ + 319, + 234, + 555, + 403 + ], + "type": "inline_equation", + "content": "B^{t} \\gets B^{t} + \\min(A, \\text{axis} = 1)" + }, + { + "bbox": [ + 319, + 234, + 555, + 403 + ], + "type": "text", + "content": " \n7: end for \n8: for " + }, + { + "bbox": [ + 319, + 234, + 555, + 403 + ], + "type": "inline_equation", + "content": "n \\in \\{1, 2, \\dots, N\\}" + }, + { + "bbox": [ + 319, + 234, + 555, + 403 + ], + "type": "text", + "content": " do \n9: " + }, + { + "bbox": [ + 319, + 234, + 555, + 403 + ], + "type": "inline_equation", + "content": "I_{C^{t}}^{t}(n) \\gets 1 - \\frac{B^{t}(n) - \\min(B^{t})}{\\max(B^{t}) - \\min(B^{t})}" + }, + { + "bbox": [ + 319, + 234, + 555, + 403 + ], + "type": "text", + "content": " \n10: " + }, + { + "bbox": [ + 319, + 234, + 555, + 403 + ], + "type": "inline_equation", + "content": "I^{t+1}(n) \\gets \\frac{|C^{1:t-1}|}{|C^{1:t}|} \\cdot I^{t}(n) + \\frac{|C^{t}|}{|C^{1:t}|} \\cdot I_{C^{t}}^{t}(n)" + }, + { + "bbox": [ + 319, + 234, + 555, + 403 + ], + "type": "text", + "content": " \n11: end for \n12: Output: Importance matrix " + }, + { + "bbox": [ + 319, + 234, + 555, + 403 + ], + "type": "inline_equation", + "content": "I^{t+1}" + }, + { + "bbox": [ + 319, + 234, + 555, + 403 + ], + "type": "text", + "content": " for step " + }, + { + "bbox": [ + 319, + 234, + 555, + 403 + ], + "type": "inline_equation", + "content": "t + 1" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "code_body" + } + ], + "index": 16, + "sub_type": "algorithm" + }, + { + "bbox": [ + 314, + 423, + 427, + 436 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 423, + 427, + 436 + ], + "spans": [ + { + "bbox": [ + 314, + 423, + 427, + 436 + ], + "type": "text", + "content": "4.4. Objective Function" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 441, + 541, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 441, + 541, + 453 + ], + "spans": [ + { + "bbox": [ + 313, + 441, + 541, + 453 + ], + "type": "text", + "content": "Overall, the objective function of CoMBO is defined as:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 376, + 464, + 555, + 475 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 376, + 464, + 555, + 475 + ], + "spans": [ + { + "bbox": [ + 376, + 464, + 555, + 475 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {H} = \\mathcal {L} _ {D L} + \\lambda_ {I K D} \\mathcal {L} _ {I K D}. \\tag {10}", + "image_path": "0ae1b6d31d26f5330f2751ee0ba2cce082476888dbe2f941b28ec5a77dfe9a5f.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 485, + 554, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 485, + 554, + 521 + ], + "spans": [ + { + "bbox": [ + 313, + 485, + 554, + 521 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 485, + 554, + 521 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{DL}" + }, + { + "bbox": [ + 313, + 485, + 554, + 521 + ], + "type": "text", + "content": " represents the Half-Distillation-Half-Learning loss, and " + }, + { + "bbox": [ + 313, + 485, + 554, + 521 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{IKD}" + }, + { + "bbox": [ + 313, + 485, + 554, + 521 + ], + "type": "text", + "content": " means the Importance-Based Knowledge Distillation loss, with " + }, + { + "bbox": [ + 313, + 485, + 554, + 521 + ], + "type": "inline_equation", + "content": "\\lambda_{IKD}" + }, + { + "bbox": [ + 313, + 485, + 554, + 521 + ], + "type": "text", + "content": " as its weighting factor." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 532, + 440, + 544 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 532, + 440, + 544 + ], + "spans": [ + { + "bbox": [ + 313, + 532, + 440, + 544 + ], + "type": "text", + "content": "5. Experimental Results" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 552, + 430, + 564 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 552, + 430, + 564 + ], + "spans": [ + { + "bbox": [ + 313, + 552, + 430, + 564 + ], + "type": "text", + "content": "5.1. Experimental Setup" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 570, + 554, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 554, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 554, + 665 + ], + "type": "text", + "content": "Datasets and Evaluation Metrics. We follow the experimental settings of previous works [8, 11, 26] and evaluate our approach on the ADE20K [52] dataset. The ADE20K dataset is specifically designed to support both panoptic and semantic segmentation tasks. ADE20K contains 150 classes, including 100 thing classes and 50 stuff classes. The dataset is composed of 20,210 images for training and 2,000 images for validation." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 665, + 554, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 554, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 554, + 715 + ], + "type": "text", + "content": "For Class Incremental Semantic Segmentation (CISS), we adopt the mean Intersection over Union (mIoU) for evaluation. The Intersection over Union (IoU) is calculated as " + }, + { + "bbox": [ + 313, + 665, + 554, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{IoU} = \\frac{TP}{TP + FP + FN}" + }, + { + "bbox": [ + 313, + 665, + 554, + 715 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 665, + 554, + 715 + ], + "type": "inline_equation", + "content": "TP" + }, + { + "bbox": [ + 313, + 665, + 554, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 665, + 554, + 715 + ], + "type": "inline_equation", + "content": "FP" + }, + { + "bbox": [ + 313, + 665, + 554, + 715 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 665, + 554, + 715 + ], + "type": "inline_equation", + "content": "FN" + }, + { + "bbox": [ + 313, + 665, + 554, + 715 + ], + "type": "text", + "content": " denote the" + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25671" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 70, + 553, + 202 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 553, + 202 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 553, + 202 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 553, + 202 + ], + "type": "table", + "html": "
Method100-50 (2 steps)100-10 (6 steps)100-5 (11 steps)50-50 (3 steps)
1-100101-150all1-100101-150all1-100101-150all1-5051-150all
FT0.01.30.40.02.91.00.025.88.60.012.08.1
MiB [5]CVPR2035.119.329.827.110.021.424.06.518.142.415.524.4
PLOP [20]CVPR2140.222.434.330.517.526.128.115.724.045.818.727.7
CoMFormer [8]CVPR2341.127.736.736.017.129.734.415.928.245.019.327.9
ECLIPSE [26]CVPR2441.723.535.641.418.833.941.116.632.946.020.729.2
BalConpas [11]ECCV2442.825.737.140.722.834.736.120.330.851.226.534.7
CoMBO Ours43.925.637.840.825.235.636.120.530.950.728.235.7
Joint43.830.939.543.830.939.543.830.939.550.733.939.5
", + "image_path": "55b544273ba64af708da489ec1473ddfec0907044a0bd317e0cd39de0a6461c9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 210, + 553, + 234 + ], + "lines": [ + { + "bbox": [ + 55, + 210, + 553, + 234 + ], + "spans": [ + { + "bbox": [ + 55, + 210, + 553, + 234 + ], + "type": "text", + "content": "Table 1. Quantitative comparison under Class Incremental Panoptic Segmentation with state-of-the-art exemplar-free methods on ADE20K in PQ. Scores of novel classes and all classes in bold are the best while underlined are the second best." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 254, + 294, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 254, + 294, + 373 + ], + "spans": [ + { + "bbox": [ + 54, + 254, + 294, + 373 + ], + "type": "text", + "content": "number of true-positive, false-positive, and false-negative pixels, respectively. The mIoU metric averages the IoU across all classes for a more comprehensive evaluation. For Class Incremental Panoptic Segmentation (CIPS), following previous work [8], we employ Panoptic Quality (PQ) as the evaluation metric. PQ is defined as the product of Recognition Quality (RQ) and Segmentation Quality (SQ). To measure incremental learning capacity, we compute the corresponding metrics for the initial classes " + }, + { + "bbox": [ + 54, + 254, + 294, + 373 + ], + "type": "inline_equation", + "content": "C^1" + }, + { + "bbox": [ + 54, + 254, + 294, + 373 + ], + "type": "text", + "content": ", incremental classes " + }, + { + "bbox": [ + 54, + 254, + 294, + 373 + ], + "type": "inline_equation", + "content": "C^{2:T}" + }, + { + "bbox": [ + 54, + 254, + 294, + 373 + ], + "type": "text", + "content": ", and the aggregate of all classes " + }, + { + "bbox": [ + 54, + 254, + 294, + 373 + ], + "type": "inline_equation", + "content": "C^{1:T}" + }, + { + "bbox": [ + 54, + 254, + 294, + 373 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 388, + 295, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 388, + 295, + 508 + ], + "spans": [ + { + "bbox": [ + 55, + 388, + 295, + 508 + ], + "type": "text", + "content": "Protocols and Implementation Details. We follow previous incremental protocols and define scenarios as " + }, + { + "bbox": [ + 55, + 388, + 295, + 508 + ], + "type": "inline_equation", + "content": "N_{ini} - N_{inc}" + }, + { + "bbox": [ + 55, + 388, + 295, + 508 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 388, + 295, + 508 + ], + "type": "inline_equation", + "content": "N_{ini}" + }, + { + "bbox": [ + 55, + 388, + 295, + 508 + ], + "type": "text", + "content": " represents the number of initial classes, and " + }, + { + "bbox": [ + 55, + 388, + 295, + 508 + ], + "type": "inline_equation", + "content": "N_{inc}" + }, + { + "bbox": [ + 55, + 388, + 295, + 508 + ], + "type": "text", + "content": " denotes the number of new classes introduced at each incremental step. For example, in the 100-10 scenario, the training begins with 100 classes, followed by the addition of 10 new classes per incremental step, without access to annotations of old classes. For both CIPS and CISS, we do evaluations on the following scenarios: 100-10 (6 steps), 100-50 (2 steps), 100-5 (11 steps) and 50-50 (3 steps)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 508, + 295, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 508, + 295, + 651 + ], + "spans": [ + { + "bbox": [ + 55, + 508, + 295, + 651 + ], + "type": "text", + "content": "Our approach is based on the Mask2Former structure [17]. Following previous approaches to CISS and CIPS, we utilize ResNet-50 [24] as the backbone for CIPS and ResNet-101 for CISS, with both pre-trained on ImageNet [19]. The input resolution of the images is " + }, + { + "bbox": [ + 55, + 508, + 295, + 651 + ], + "type": "inline_equation", + "content": "640 \\times 640" + }, + { + "bbox": [ + 55, + 508, + 295, + 651 + ], + "type": "text", + "content": " with an all-time batch size of 8. We follow the training hyperparameter of Mask2Former in the initial step, with a learning rate of " + }, + { + "bbox": [ + 55, + 508, + 295, + 651 + ], + "type": "inline_equation", + "content": "10^{-4}" + }, + { + "bbox": [ + 55, + 508, + 295, + 651 + ], + "type": "text", + "content": " and iterations of 160,000. During the incremental steps, we set 1,000 iterations per class with a learning rate of " + }, + { + "bbox": [ + 55, + 508, + 295, + 651 + ], + "type": "inline_equation", + "content": "5 \\times 10^{-5}" + }, + { + "bbox": [ + 55, + 508, + 295, + 651 + ], + "type": "text", + "content": ". The coefficients " + }, + { + "bbox": [ + 55, + 508, + 295, + 651 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 55, + 508, + 295, + 651 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 508, + 295, + 651 + ], + "type": "inline_equation", + "content": "\\lambda_{cls}" + }, + { + "bbox": [ + 55, + 508, + 295, + 651 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 508, + 295, + 651 + ], + "type": "inline_equation", + "content": "\\lambda_{kl}" + }, + { + "bbox": [ + 55, + 508, + 295, + 651 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 508, + 295, + 651 + ], + "type": "inline_equation", + "content": "\\lambda_{IKD}" + }, + { + "bbox": [ + 55, + 508, + 295, + 651 + ], + "type": "text", + "content": " are respectively set to 16, 2, 5, 3. All experiments were conducted on the NVIDIA RTX 4090." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 659, + 174, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 659, + 174, + 672 + ], + "spans": [ + { + "bbox": [ + 55, + 659, + 174, + 672 + ], + "type": "text", + "content": "5.2.Quantitative Results" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "content": "Comparisons in Class Incremental Panoptic Segmentation (CIPS). We evaluated our approach against state-of-the-art exemplar-free methods on the ADE20K dataset" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 254, + 555, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 254, + 555, + 445 + ], + "spans": [ + { + "bbox": [ + 313, + 254, + 555, + 445 + ], + "type": "text", + "content": "within the CIPS framework, as detailed in Tab. 1. The performance of our CoMBO surpasses other methods in most of the scenarios, particularly in the more challenging 100-10 scenario, where it achieves a remarkable " + }, + { + "bbox": [ + 313, + 254, + 555, + 445 + ], + "type": "inline_equation", + "content": "35.6\\%" + }, + { + "bbox": [ + 313, + 254, + 555, + 445 + ], + "type": "text", + "content": " PQ with at least " + }, + { + "bbox": [ + 313, + 254, + 555, + 445 + ], + "type": "inline_equation", + "content": "3.4\\%" + }, + { + "bbox": [ + 313, + 254, + 555, + 445 + ], + "type": "text", + "content": " of advantage on incremental classes compared to the previous state-of-the-art methods. This notable performance gain demonstrates the effectiveness of our CoMBO in reducing conflicts. Although our approach still underperforms the state-of-the-art approach ECLIPSE [26] in the 100-5 scenario, primarily due to its strong memorization ability from freezing parameters, our method achieves advanced performance for the new classes. Moreover, in scenarios with fewer initial classes, i.e., 50-50 scenario, we maintain our advantage in learning new classes with a PQ of " + }, + { + "bbox": [ + 313, + 254, + 555, + 445 + ], + "type": "inline_equation", + "content": "28.2\\%" + }, + { + "bbox": [ + 313, + 254, + 555, + 445 + ], + "type": "text", + "content": " for incremental classes and " + }, + { + "bbox": [ + 313, + 254, + 555, + 445 + ], + "type": "inline_equation", + "content": "35.7\\%" + }, + { + "bbox": [ + 313, + 254, + 555, + 445 + ], + "type": "text", + "content": " for overall performance, widening the gap with strong distillation methods." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 460, + 556, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 460, + 556, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 460, + 556, + 628 + ], + "type": "text", + "content": "Comparisons in Class Incremental Semantic Segmentation (CISS). We further extended our evaluation to the semantic segmentation benchmark, comparing our approach with previous methods on the ADE20K dataset, as detailed in Tab. 2. Our approach consistently outperforms previous methods in all tested scenarios. Notably, in the 100-10 scenario, our approach achieves a mIoU of " + }, + { + "bbox": [ + 313, + 460, + 556, + 628 + ], + "type": "inline_equation", + "content": "41.1\\%" + }, + { + "bbox": [ + 313, + 460, + 556, + 628 + ], + "type": "text", + "content": ", surpassing the previous state-of-the-art by at least " + }, + { + "bbox": [ + 313, + 460, + 556, + 628 + ], + "type": "inline_equation", + "content": "3.5\\%" + }, + { + "bbox": [ + 313, + 460, + 556, + 628 + ], + "type": "text", + "content": " for incremental classes, demonstrating its effectiveness in mitigating conflicts. Moreover, in the most challenging 100-5 long-term incremental scenario, our approach not only improves performance on old classes by " + }, + { + "bbox": [ + 313, + 460, + 556, + 628 + ], + "type": "inline_equation", + "content": "2.5\\%" + }, + { + "bbox": [ + 313, + 460, + 556, + 628 + ], + "type": "text", + "content": " but also achieves a significant " + }, + { + "bbox": [ + 313, + 460, + 556, + 628 + ], + "type": "inline_equation", + "content": "5.4\\%" + }, + { + "bbox": [ + 313, + 460, + 556, + 628 + ], + "type": "text", + "content": " increase in mIoU for new classes, ensuring both model stability and adaptability." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 639, + 406, + 653 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 639, + 406, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 639, + 406, + 653 + ], + "type": "text", + "content": "6. Ablation Study" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 658, + 440, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 658, + 440, + 671 + ], + "spans": [ + { + "bbox": [ + 313, + 658, + 440, + 671 + ], + "type": "text", + "content": "6.1. Component Ablations" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "content": "We analyze the key components of our proposed framework, which includes the Half-Distillation-Half-Learning Strategy, Importance-based Knowledge Distillation, and the" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "25672" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 70, + 553, + 213 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 553, + 213 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 553, + 213 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 553, + 213 + ], + "type": "table", + "html": "
Method100-50 (2 steps)100-10 (6 steps)100-5 (11 steps)50-50 (3 steps)
1-100101-150all1-100101-150all1-100101-150all1-5051-150all
FT0.026.78.90.02.30.80.01.10.30.01.71.1
MiB [5]CVPR2037.024.132.623.510.626.621.06.116.145.621.029.3
PLOP [20]CVPR2144.226.238.234.815.928.539.513.630.954.930.238.4
CoMFormer [8]CVPR2344.726.238.440.615.632.339.513.630.949.226.634.1
CoMasTRe [23]CVPR2445.726.039.242.318.434.440.815.832.649.826.634.5
ECLIPSE [26]CVPR2445.021.737.143.417.434.643.316.334.2---
BalConpas [11]ECCV2449.930.143.347.324.238.642.117.233.855.833.340.8
CoMBO Ours50.234.444.947.827.741.144.622.637.355.336.943.0
Joint51.740.247.851.740.247.851.740.247.856.643.547.8
", + "image_path": "96b3c37a64fa2ecc2581d4f2b7775056327cb318ad83d268d495e2ec5c8215e1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 264, + 294, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 264, + 294, + 350 + ], + "spans": [ + { + "bbox": [ + 54, + 264, + 294, + 350 + ], + "type": "text", + "content": "Query Conflict Reducing module. Our experiments focus on the 100-10 scenario in the CIPS, as it provides a moderate number of steps to examine the ability of both acquisition and retention. We evaluate various combinations of these components separately and present the results in Tab. 3. The baseline method utilizes the vanilla loss with pseudo-labeling." + } + ] + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 67, + 360, + 284, + 456 + ], + "blocks": [ + { + "bbox": [ + 55, + 220, + 553, + 243 + ], + "lines": [ + { + "bbox": [ + 55, + 220, + 553, + 243 + ], + "spans": [ + { + "bbox": [ + 55, + 220, + 553, + 243 + ], + "type": "text", + "content": "Table 2. Quantitative comparison under Class Incremental Semantic Segmentation with state-of-the-art exemplar-free methods on ADE20K in mIoU. Scores of novel classes and all classes in bold are the best while underlined are the second best." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 67, + 360, + 284, + 456 + ], + "lines": [ + { + "bbox": [ + 67, + 360, + 284, + 456 + ], + "spans": [ + { + "bbox": [ + 67, + 360, + 284, + 456 + ], + "type": "table", + "html": "
HDHLIKDQCR100-10 (6 steps)
1-100101-150all
36.323.832.2
✓39.424.234.3
✓✓40.024.634.9
✓✓38.825.434.4
✓✓✓40.825.235.6
", + "image_path": "2a7c60d0d35e7eeaa42b1b21f45f0c99e4a757681e5a8800e5a6c0b3b8a377f2.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 465, + 295, + 499 + ], + "lines": [ + { + "bbox": [ + 55, + 465, + 295, + 499 + ], + "spans": [ + { + "bbox": [ + 55, + 465, + 295, + 499 + ], + "type": "text", + "content": "Table 3. Ablation study of the main components on the 100-10 task of CIPS. The baseline in the " + }, + { + "bbox": [ + 55, + 465, + 295, + 499 + ], + "type": "inline_equation", + "content": "1^{\\text{st}}" + }, + { + "bbox": [ + 55, + 465, + 295, + 499 + ], + "type": "text", + "content": " row employs vanilla losses with pseudo-labeling and excludes the QCR module." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 56, + 510, + 295, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 510, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 510, + 295, + 712 + ], + "type": "text", + "content": "Under identical experimental conditions, the HDHL Strategy significantly enhances the overall PQ by " + }, + { + "bbox": [ + 56, + 510, + 295, + 712 + ], + "type": "inline_equation", + "content": "2.1\\%" + }, + { + "bbox": [ + 56, + 510, + 295, + 712 + ], + "type": "text", + "content": ", demonstrating its ability to unify logits of new classes into the previous logits distribution without contradictory losses. The implementation of the IKD mechanism has a remarkable " + }, + { + "bbox": [ + 56, + 510, + 295, + 712 + ], + "type": "inline_equation", + "content": "3.7\\%" + }, + { + "bbox": [ + 56, + 510, + 295, + 712 + ], + "type": "text", + "content": " increase in the PQ for old classes, underscoring its effectiveness in boosting the ability to retain the knowledge of old classes. Furthermore, when combined with the QCR module for branched optimization, CoMBO further boosts both the performance of old and new classes, achieving " + }, + { + "bbox": [ + 56, + 510, + 295, + 712 + ], + "type": "inline_equation", + "content": "0.8\\%" + }, + { + "bbox": [ + 56, + 510, + 295, + 712 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 56, + 510, + 295, + 712 + ], + "type": "inline_equation", + "content": "0.6\\%" + }, + { + "bbox": [ + 56, + 510, + 295, + 712 + ], + "type": "text", + "content": " gains in PQ, respectively, surpassing previous methods that struggle to balance acquisition and retention. Consequently, CoMBO achieves a " + }, + { + "bbox": [ + 56, + 510, + 295, + 712 + ], + "type": "inline_equation", + "content": "3.4\\%" + }, + { + "bbox": [ + 56, + 510, + 295, + 712 + ], + "type": "text", + "content": " PQ improvement over the baseline. The above results demonstrate the effectiveness of the proposed CoMBO and its related modules, showing a noticeable improvement in reducing the conflict on model structures and losses." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 263, + 447, + 277 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 263, + 447, + 277 + ], + "spans": [ + { + "bbox": [ + 313, + 263, + 447, + 277 + ], + "type": "text", + "content": "6.2. Ablation Study of QCR" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 281, + 555, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 281, + 555, + 402 + ], + "spans": [ + { + "bbox": [ + 313, + 281, + 555, + 402 + ], + "type": "text", + "content": "We validate the effectiveness of QCR with different settings of " + }, + { + "bbox": [ + 313, + 281, + 555, + 402 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 313, + 281, + 555, + 402 + ], + "type": "text", + "content": " shown in Fig. 5. Under the estimation of performance enhancement and additional parameters, our QCR module achieves the best result when " + }, + { + "bbox": [ + 313, + 281, + 555, + 402 + ], + "type": "inline_equation", + "content": "r = 16" + }, + { + "bbox": [ + 313, + 281, + 555, + 402 + ], + "type": "text", + "content": ", where the quantity of additional parameters comes to " + }, + { + "bbox": [ + 313, + 281, + 555, + 402 + ], + "type": "inline_equation", + "content": "8.2\\mathrm{K}" + }, + { + "bbox": [ + 313, + 281, + 555, + 402 + ], + "type": "text", + "content": " per class. Considering the total parameter of Mask2Former [17] is " + }, + { + "bbox": [ + 313, + 281, + 555, + 402 + ], + "type": "inline_equation", + "content": "44.9\\mathrm{M}" + }, + { + "bbox": [ + 313, + 281, + 555, + 402 + ], + "type": "text", + "content": " with ResNet-50, the additional parameters only reach up to " + }, + { + "bbox": [ + 313, + 281, + 555, + 402 + ], + "type": "inline_equation", + "content": "2\\%" + }, + { + "bbox": [ + 313, + 281, + 555, + 402 + ], + "type": "text", + "content": " of original parameters in total, which is 5 times less than per class additional parameters of ECLIPSE [26] while having better performance on new classes." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 407, + 497, + 421 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 407, + 497, + 421 + ], + "spans": [ + { + "bbox": [ + 313, + 407, + 497, + 421 + ], + "type": "text", + "content": "6.3. Ablation Study of HDHL Strategy" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 426, + 555, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 555, + 629 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 555, + 629 + ], + "type": "text", + "content": "To validate the effectiveness of our Half-Distillation-Half-Learning strategy, we conduct a series of experiments among other strategies, as shown in Tab. 4. The " + }, + { + "bbox": [ + 313, + 426, + 555, + 629 + ], + "type": "inline_equation", + "content": "1^{\\mathrm{st}}" + }, + { + "bbox": [ + 313, + 426, + 555, + 629 + ], + "type": "text", + "content": " row with " + }, + { + "bbox": [ + 313, + 426, + 555, + 629 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{cls}" + }, + { + "bbox": [ + 313, + 426, + 555, + 629 + ], + "type": "text", + "content": " on all classification embeddings represents the vanilla loss with pseudo-labeling, which is adopted in [8]. The " + }, + { + "bbox": [ + 313, + 426, + 555, + 629 + ], + "type": "inline_equation", + "content": "2^{\\mathrm{nd}}" + }, + { + "bbox": [ + 313, + 426, + 555, + 629 + ], + "type": "text", + "content": " row utilizes an auxiliary distillation on the embeddings following [23]. Compared to these strategies with tough supervision on unmatched embeddings " + }, + { + "bbox": [ + 313, + 426, + 555, + 629 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{cls}^{t,\\emptyset}" + }, + { + "bbox": [ + 313, + 426, + 555, + 629 + ], + "type": "text", + "content": " and mostly conflicting losses, our HDHL strategy with soft and concentrated supervision has a better PQ of 34.3. Besides, the HDHL design of " + }, + { + "bbox": [ + 313, + 426, + 555, + 629 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{kl}" + }, + { + "bbox": [ + 313, + 426, + 555, + 629 + ], + "type": "text", + "content": " ensures the integrity while optimizing the unmatched embeddings " + }, + { + "bbox": [ + 313, + 426, + 555, + 629 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{cls}^{t,\\emptyset}" + }, + { + "bbox": [ + 313, + 426, + 555, + 629 + ], + "type": "text", + "content": ", with improvement of " + }, + { + "bbox": [ + 313, + 426, + 555, + 629 + ], + "type": "inline_equation", + "content": "2.1\\%" + }, + { + "bbox": [ + 313, + 426, + 555, + 629 + ], + "type": "text", + "content": " compared to distillation-only strategy in the " + }, + { + "bbox": [ + 313, + 426, + 555, + 629 + ], + "type": "inline_equation", + "content": "3^{\\mathrm{rd}}" + }, + { + "bbox": [ + 313, + 426, + 555, + 629 + ], + "type": "text", + "content": " row and " + }, + { + "bbox": [ + 313, + 426, + 555, + 629 + ], + "type": "inline_equation", + "content": "0.8\\%" + }, + { + "bbox": [ + 313, + 426, + 555, + 629 + ], + "type": "text", + "content": " compared to the simple combination of distillation and " + }, + { + "bbox": [ + 313, + 426, + 555, + 629 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{cls}" + }, + { + "bbox": [ + 313, + 426, + 555, + 629 + ], + "type": "text", + "content": ". The above results show that our HDHL strategy outperforms other learning or distillation strategies on the class embeddings." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 635, + 432, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 635, + 432, + 647 + ], + "spans": [ + { + "bbox": [ + 313, + 635, + 432, + 647 + ], + "type": "text", + "content": "6.4. Qualitative Analysis" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 653, + 554, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 554, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 554, + 715 + ], + "type": "text", + "content": "We conduct the qualitative analysis (shown in Fig. 4) by comparing our proposed CoMBO to the Baseline and Baseline+HDHL " + }, + { + "bbox": [ + 313, + 653, + 554, + 715 + ], + "type": "inline_equation", + "content": "(1^{\\mathrm{st}}" + }, + { + "bbox": [ + 313, + 653, + 554, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 653, + 554, + 715 + ], + "type": "inline_equation", + "content": "2^{\\mathrm{nd}}" + }, + { + "bbox": [ + 313, + 653, + 554, + 715 + ], + "type": "text", + "content": " rows in Tab. 3). In the " + }, + { + "bbox": [ + 313, + 653, + 554, + 715 + ], + "type": "inline_equation", + "content": "1^{\\mathrm{st}}" + }, + { + "bbox": [ + 313, + 653, + 554, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 653, + 554, + 715 + ], + "type": "inline_equation", + "content": "2^{\\mathrm{nd}}" + }, + { + "bbox": [ + 313, + 653, + 554, + 715 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 653, + 554, + 715 + ], + "type": "inline_equation", + "content": "5^{\\mathrm{th}}" + }, + { + "bbox": [ + 313, + 653, + 554, + 715 + ], + "type": "text", + "content": " columns of Fig. 4, our CoMBO can indicate the incremental classes, i.e., screen door, fountain, and ottoman," + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25673" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 75, + 74, + 533, + 350 + ], + "blocks": [ + { + "bbox": [ + 75, + 74, + 533, + 350 + ], + "lines": [ + { + "bbox": [ + 75, + 74, + 533, + 350 + ], + "spans": [ + { + "bbox": [ + 75, + 74, + 533, + 350 + ], + "type": "image", + "image_path": "aa5a3fe79756d31ace26406ce9e8d831a64fa3a8d44c041e001872af8c0736a3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 94, + 356, + 514, + 369 + ], + "lines": [ + { + "bbox": [ + 94, + 356, + 514, + 369 + ], + "spans": [ + { + "bbox": [ + 94, + 356, + 514, + 369 + ], + "type": "text", + "content": "Figure 4. Qualitative results of CoMBO comparing to Baseline, Baseline+HDHL on 100-10 CIPS task of ADE20K." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 61, + 384, + 288, + 523 + ], + "blocks": [ + { + "bbox": [ + 61, + 384, + 288, + 523 + ], + "lines": [ + { + "bbox": [ + 61, + 384, + 288, + 523 + ], + "spans": [ + { + "bbox": [ + 61, + 384, + 288, + 523 + ], + "type": "image", + "image_path": "c8eee99b786109d66d892a0b6bb7ddb6069f54262cde2270d14dc4c33794e2fe.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 525, + 295, + 548 + ], + "lines": [ + { + "bbox": [ + 55, + 525, + 295, + 548 + ], + "spans": [ + { + "bbox": [ + 55, + 525, + 295, + 548 + ], + "type": "text", + "content": "Figure 5. Ablation study of " + }, + { + "bbox": [ + 55, + 525, + 295, + 548 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 55, + 525, + 295, + 548 + ], + "type": "text", + "content": " in QCR module, including PQ performance (left) and number of parameters (right)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 75, + 552, + 277, + 649 + ], + "blocks": [ + { + "bbox": [ + 75, + 552, + 277, + 649 + ], + "lines": [ + { + "bbox": [ + 75, + 552, + 277, + 649 + ], + "spans": [ + { + "bbox": [ + 75, + 552, + 277, + 649 + ], + "type": "table", + "html": "
LclsLklPQ
Ectls & Ectls1:t-1EctlsC0:t-1Ct
✓✓32.2
✓✓✓32.6
✓✓32.2
✓✓✓33.5
✓✓✓34.3
", + "image_path": "e6650164ecfd5d39c4d80b77764d70d09f27be8153352329b5ad55170914ed70.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 658, + 296, + 693 + ], + "lines": [ + { + "bbox": [ + 55, + 658, + 296, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 658, + 296, + 693 + ], + "type": "text", + "content": "Table 4. Ablation study of the HDHL strategy on task 100-10 of CIPS. The " + }, + { + "bbox": [ + 55, + 658, + 296, + 693 + ], + "type": "inline_equation", + "content": "1^{\\text{st}}" + }, + { + "bbox": [ + 55, + 658, + 296, + 693 + ], + "type": "text", + "content": " row represents the vanilla loss with pseudo-labeling. The last row represents our " + }, + { + "bbox": [ + 55, + 658, + 296, + 693 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{DL}" + }, + { + "bbox": [ + 55, + 658, + 296, + 693 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 386, + 555, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 386, + 555, + 472 + ], + "spans": [ + { + "bbox": [ + 313, + 386, + 555, + 472 + ], + "type": "text", + "content": "while other methods fail to recognize all these objects. In the " + }, + { + "bbox": [ + 313, + 386, + 555, + 472 + ], + "type": "inline_equation", + "content": "3^{\\mathrm{rd}}" + }, + { + "bbox": [ + 313, + 386, + 555, + 472 + ], + "type": "text", + "content": " column, our method successfully remembers the car in the left instead of covering the object with the prediction of the incremental class van. Furthermore, our method allows more precise mask prediction of new class Traffic Light in the " + }, + { + "bbox": [ + 313, + 386, + 555, + 472 + ], + "type": "inline_equation", + "content": "4^{\\mathrm{th}}" + }, + { + "bbox": [ + 313, + 386, + 555, + 472 + ], + "type": "text", + "content": " column, demonstrating the effectiveness of refinements on the queries corresponding to new classes." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 480, + 392, + 493 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 480, + 392, + 493 + ], + "spans": [ + { + "bbox": [ + 313, + 480, + 392, + 493 + ], + "type": "text", + "content": "7. Conclusions" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 500, + 556, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 500, + 556, + 693 + ], + "spans": [ + { + "bbox": [ + 313, + 500, + 556, + 693 + ], + "type": "text", + "content": "In this paper, we present CoMBO, a novel Class Incremental Segmentation (CIS) method for mitigating the conflict between acquisition and retention losses. This conflict arises from competing goals of efficiently acquiring knowledge about new classes while preserving knowledge of previously learned ones. Our QCR module branches the conflicting optimization targets via lightweight class-specific adaptation on queries, enabling the coexistence of learning and distillation on separated queries. The HDHL strategy and IKD further reduce the contradictory optimization on classification logits and queries with unified targets, respectively, thereby enhancing the overall model performance. Extensive experimental validation on the ADE20K dataset underscores the superiority of our approach, demonstrating the effectiveness of each component in the CIS task, particularly excelling in performance on new classes." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "25674" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 163 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 163 + ], + "type": "text", + "content": "This work was supported by the National Key R&D Program of China (Grant 2022YFC3310200), the National Natural Science Foundation of China (Grant 62472033, 92470203), the Beijing Natural Science Foundation (Grant L242022), the Royal Society grants (SIF\\R1\\231009, IES\\R3\\223050) and an Amazon Research Award." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 172, + 115, + 185 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 172, + 115, + 185 + ], + "spans": [ + { + "bbox": [ + 56, + 172, + 115, + 185 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 193, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 193, + 295, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 193, + 295, + 225 + ], + "spans": [ + { + "bbox": [ + 61, + 193, + 295, + 225 + ], + "type": "text", + "content": "[1] Rahaf Aljundi, Min Lin, Baptiste Goujaud, and Yoshua Bengio. Gradient based sample selection for online continual learning. Adv. Neural Inf. Process. Syst., 32, 2019. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 226, + 295, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 226, + 295, + 270 + ], + "spans": [ + { + "bbox": [ + 61, + 226, + 295, + 270 + ], + "type": "text", + "content": "[2] Vijay Badrinarayanan, Alex Kendall, and Roberto Cipolla. Segnet: A deep convolutional encoder-decoder architecture for image segmentation. IEEE Trans. on Pattern Anal. Mach. Intell., 39(12):2481-2495, 2017. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 271, + 295, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 271, + 295, + 314 + ], + "spans": [ + { + "bbox": [ + 62, + 271, + 295, + 314 + ], + "type": "text", + "content": "[3] Donghyeon Baek, Youngmin Oh, Sanghoon Lee, Junghyup Lee, and Bumsub Ham. Decomposed knowledge distillation for class-incremental semantic segmentation. Adv. Neural Inf. Process. Syst., 35:10380-10392, 2022. 2, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 315, + 295, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 315, + 295, + 347 + ], + "spans": [ + { + "bbox": [ + 62, + 315, + 295, + 347 + ], + "type": "text", + "content": "[4] Zalán Borsos, Mojmir Mutny, and Andreas Krause. Coresets via bilevel optimization for continual learning and streaming. Adv. Neural Inf. Process. Syst., 33:14879-14890, 2020. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 348, + 295, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 348, + 295, + 402 + ], + "spans": [ + { + "bbox": [ + 62, + 348, + 295, + 402 + ], + "type": "text", + "content": "[5] Fabio Cermelli, Massimiliano Mancini, Samuel Rota Bulo, Elisa Ricci, and Barbara Caputo. Modeling the background for incremental learning in semantic segmentation. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 9233-9242, 2020. 2, 3, 6, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 403, + 295, + 447 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 403, + 295, + 447 + ], + "spans": [ + { + "bbox": [ + 62, + 403, + 295, + 447 + ], + "type": "text", + "content": "[6] Fabio Cermelli, Dario Fontanel, Antonio Tavera, Marco Ciccone, and Barbara Caputo. Incremental learning in semantic segmentation from image labels. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 4371-4381, 2022." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 449, + 295, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 449, + 295, + 491 + ], + "spans": [ + { + "bbox": [ + 62, + 449, + 295, + 491 + ], + "type": "text", + "content": "[7] Fabio Cermelli, Antonino Geraci, Dario Fontanel, and Barbara Caputo. Modeling missing annotations for incremental learning in object detection. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 3700-3710, 2022. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 492, + 295, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 492, + 295, + 536 + ], + "spans": [ + { + "bbox": [ + 62, + 492, + 295, + 536 + ], + "type": "text", + "content": "[8] Fabio Cermelli, Matthieu Cord, and Arthur Douillard. Comformer: Continual learning in semantic and panoptic segmentation. In Proc. IEEE Conf. Comput. Vis. Pattern Recognition., pages 3010-3020, 2023. 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 537, + 295, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 537, + 295, + 580 + ], + "spans": [ + { + "bbox": [ + 62, + 537, + 295, + 580 + ], + "type": "text", + "content": "[9] Sungmin Cha, YoungJoon Yoo, Taesup Moon, et al. Ssul: Semantic segmentation with unknown label for exemplar-based class-incremental learning. Adv. Neural Inf. Process. Syst., 34:10919-10930, 2021. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 581, + 295, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 581, + 295, + 635 + ], + "spans": [ + { + "bbox": [ + 57, + 581, + 295, + 635 + ], + "type": "text", + "content": "[10] Jinpeng Chen, Runmin Cong, Yuxuan Luo, Horace Ip, and Sam Kwong. Saving 100x storage: prototype replay for reconstructing training sample distribution in class incremental semantic segmentation. Adv. Neural Inf. Process. Syst., 36, 2024. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 636, + 295, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 636, + 295, + 680 + ], + "spans": [ + { + "bbox": [ + 57, + 636, + 295, + 680 + ], + "type": "text", + "content": "[11] Jinpeng Chen, Runmin Cong, Yuxuan Luo, Horace Ho Shing Ip, and Sam Kwong. Strike a balance in continual panoptic segmentation. In Proc. Eur. Conf. Comput. Vis., pages 126-142. Springer, 2025. 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 681, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 681, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 681, + 295, + 713 + ], + "type": "text", + "content": "[12] Liang-Chieh Chen. Semantic image segmentation with deep convolutional nets and fully connected crfs. arXiv preprint arXiv:1412.7062, 2014. 2" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 554, + 713 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 105 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 105 + ], + "type": "text", + "content": "[13] L. C. Chen. Rethinking atrous convolution for semantic image segmentation. arXiv preprint:1706.05587, 2017. 2, 3, 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 106, + 554, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 106, + 554, + 149 + ], + "spans": [ + { + "bbox": [ + 317, + 106, + 554, + 149 + ], + "type": "text", + "content": "[14] Liang-Chieh Chen, George Papandreou, and et al. Deep convolutional nets, atrous convolution, and fully connected crfs. IEEE Trans. on Pattern Anal. Mach. Intell., 40(4):834-848, 2017." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 151, + 554, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 151, + 554, + 194 + ], + "spans": [ + { + "bbox": [ + 317, + 151, + 554, + 194 + ], + "type": "text", + "content": "[15] Liang-Chieh Chen, Yukun Zhu, George Papandreou, Florian Schroff, and Hartwig Adam. Encoder-decoder with atrous separable convolution for semantic image segmentation. In Proc. Eur. Conf. Comput. Vis., pages 801-818, 2018. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 195, + 554, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 195, + 554, + 237 + ], + "spans": [ + { + "bbox": [ + 316, + 195, + 554, + 237 + ], + "type": "text", + "content": "[16] Bowen Cheng, Alex Schwing, and Alexander Kirillov. Per-pixel classification is not all you need for semantic segmentation. Adv. Neural Inf. Process. Syst., 34:17864-17875, 2021. 1, 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 239, + 554, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 239, + 554, + 293 + ], + "spans": [ + { + "bbox": [ + 316, + 239, + 554, + 293 + ], + "type": "text", + "content": "[17] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 1290–1299, 2022. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 294, + 554, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 294, + 554, + 338 + ], + "spans": [ + { + "bbox": [ + 316, + 294, + 554, + 338 + ], + "type": "text", + "content": "[18] Wei Cong, Yang Cong, Yuyang Liu, and Gan Sun. Cs2k: Class-specific and class-shared knowledge guidance for incremental semantic segmentation. In Proc. Eur. Conf. Comput. Vis., pages 244-261, 2025. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 338, + 554, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 338, + 554, + 381 + ], + "spans": [ + { + "bbox": [ + 316, + 338, + 554, + 381 + ], + "type": "text", + "content": "[19] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 248–255, 2009. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 383, + 554, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 383, + 554, + 426 + ], + "spans": [ + { + "bbox": [ + 316, + 383, + 554, + 426 + ], + "type": "text", + "content": "[20] Arthur Douillard, Yifu Chen, Arnaud Dapogny, and Matthieu Cord. Plop: Learning without forgetting for continual semantic segmentation. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 4040-4050, 2021. 2, 3, 6, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 427, + 554, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 427, + 554, + 470 + ], + "spans": [ + { + "bbox": [ + 316, + 427, + 554, + 470 + ], + "type": "text", + "content": "[21] Arthur Douillard, Yifu Chen, Arnaud Dapogny, and Matthieu Cord. Tackling catastrophic forgetting and background shift in continual semantic segmentation. arXiv preprint arXiv:2106.15287, 2021. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 472, + 554, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 472, + 554, + 502 + ], + "spans": [ + { + "bbox": [ + 316, + 472, + 554, + 502 + ], + "type": "text", + "content": "[22] Robert M French. Catastrophic forgetting in connectionist networks. Trends in cognitive sciences, 3(4):128-135, 1999. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 504, + 554, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 504, + 554, + 548 + ], + "spans": [ + { + "bbox": [ + 316, + 504, + 554, + 548 + ], + "type": "text", + "content": "[23] Yizheng Gong, Siyue Yu, Xiaoyang Wang, and Jimin Xiao. Continual segmentation with disentangled objectness learning and class recognition. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 3848-3857, 2024. 3, 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 548, + 554, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 548, + 554, + 591 + ], + "spans": [ + { + "bbox": [ + 316, + 548, + 554, + 591 + ], + "type": "text", + "content": "[24] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 770-778, 2016. 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 593, + 554, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 593, + 554, + 646 + ], + "spans": [ + { + "bbox": [ + 316, + 593, + 554, + 646 + ], + "type": "text", + "content": "[25] M. G. Z. A. Khan, M. F. Naeem, Luc Van Gool, Didier Stricker, Federico Tombari, and Muhammad Zeshan Afzal. Introducing language guidance in prompt-based continual learning. In Proc. IEEE Int. Conf. Comput. Vis., pages 11463-11473, 2023. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 647, + 554, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 554, + 691 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 554, + 691 + ], + "type": "text", + "content": "[26] Beomyoung Kim, Joonsang Yu, and Sung Ju Hwang. Eclipse: Efficient continual learning in panoptic segmentation with visual prompt tuning. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 3346-3356, 2024. 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 692, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 554, + 713 + ], + "type": "text", + "content": "[27] Alexander Kirillov, Kaiming He, Ross Girshick, Carsten Rother, and Piotr Dollar. Panoptic segmentation. In Proc." + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25675" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "text", + "content": "IEEE Conf. Comput. Vis. Pattern Recognit., pages 9404-9413, 2019. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 96, + 295, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 96, + 295, + 128 + ], + "spans": [ + { + "bbox": [ + 57, + 96, + 295, + 128 + ], + "type": "text", + "content": "[28] Zhizhong Li and Derek Hoiem. Learning without forgetting. IEEE Trans. on Pattern Anal. Mach. Intell., 40(12):2935-2947, 2017. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 129, + 294, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 294, + 173 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 294, + 173 + ], + "type": "text", + "content": "[29] Guosheng Lin, Anton Milan, Chunhua Shen, and Ian Reid. Refinenet: Multi-path refinement networks for high-resolution semantic segmentation. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 1925-1934, 2017. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 173, + 294, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 173, + 294, + 216 + ], + "spans": [ + { + "bbox": [ + 56, + 173, + 294, + 216 + ], + "type": "text", + "content": "[30] Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 3431-3440, 2015. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 57, + 217, + 294, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 217, + 294, + 250 + ], + "spans": [ + { + "bbox": [ + 57, + 217, + 294, + 250 + ], + "type": "text", + "content": "[31] David Lopez-Paz and Marc'Aurelio Ranzato. Gradient episodic memory for continual learning. Adv. Neural Inf. Process. Syst., 30, 2017. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 250, + 294, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 250, + 294, + 294 + ], + "spans": [ + { + "bbox": [ + 56, + 250, + 294, + 294 + ], + "type": "text", + "content": "[32] Andrea Maracani, Umberto Michieli, Marco Toldo, and Pietro Zanuttigh. Recall: Replay-based continual learning in semantic segmentation. In Proc. IEEE Int. Conf. Comput. Vis., pages 7026-7035, 2021. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 294, + 294, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 294, + 294, + 338 + ], + "spans": [ + { + "bbox": [ + 56, + 294, + 294, + 338 + ], + "type": "text", + "content": "[33] Umberto Michieli and Pietro Zanuttigh. Continual semantic segmentation via repulsion-attraction of sparse and disentangled latent representations. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 1114-1124, 2021. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 338, + 294, + 382 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 338, + 294, + 382 + ], + "spans": [ + { + "bbox": [ + 56, + 338, + 294, + 382 + ], + "type": "text", + "content": "[34] Umberto Michieli and Pietro Zanuttigh. Continual semantic segmentation via repulsion-attraction of sparse and disentangled latent representations. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 1114-1124, 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 383, + 294, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 383, + 294, + 426 + ], + "spans": [ + { + "bbox": [ + 56, + 383, + 294, + 426 + ], + "type": "text", + "content": "[35] Youngmin Oh, Donghyeon Baek, and Bumsub Ham. Alife: Adaptive logit regularizer and feature replay for incremental semantic segmentation. Adv. Neural Inf. Process. Syst., 35: 14516-14528, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 426, + 294, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 426, + 294, + 480 + ], + "spans": [ + { + "bbox": [ + 56, + 426, + 294, + 480 + ], + "type": "text", + "content": "[36] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Proc. International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 234-241, 2015. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 482, + 294, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 482, + 294, + 536 + ], + "spans": [ + { + "bbox": [ + 56, + 482, + 294, + 536 + ], + "type": "text", + "content": "[37] Chao Shang, Hongliang Li, Fanman Meng, Qingbo Wu, Heqian Qiu, and Lanxiao Wang. Incrementer: Transformer for class-incremental semantic segmentation with knowledge distillation focusing on old class. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 7214-7224, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 537, + 294, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 537, + 294, + 581 + ], + "spans": [ + { + "bbox": [ + 56, + 537, + 294, + 581 + ], + "type": "text", + "content": "[38] Robin Strudel, Ricardo Garcia, Ivan Laptev, and Cordelia Schmid. Segmenter: Transformer for semantic segmentation. In Proc. IEEE Int. Conf. Comput. Vis., pages 7262-7272, 2021. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 582, + 294, + 603 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 582, + 294, + 603 + ], + "spans": [ + { + "bbox": [ + 56, + 582, + 294, + 603 + ], + "type": "text", + "content": "[39] Sebastian Thrun. Lifelong learning algorithms. In Learning to learn, pages 181-209. 1998. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 603, + 294, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 603, + 294, + 658 + ], + "spans": [ + { + "bbox": [ + 56, + 603, + 294, + 658 + ], + "type": "text", + "content": "[40] Zifeng Wang, Zizhao Zhang, Chen-Yu Lee, Han Zhang, Ruoxi Sun, Xiaoqi Ren, Guolong Su, Vincent Perot, Jennifer Dy, and Tomas Pfister. Learning to prompt for continual learning. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 139-149, 2022. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 658, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 294, + 713 + ], + "type": "text", + "content": "[41] Yunchao Wei, Jiashi Feng, Xiaodan Liang, Ming-Ming Cheng, Yao Zhao, and Shuicheng Yan. Object region mining with adversarial erasing: A simple classification to semantic segmentation approach. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 1568-1576, 2017. 2" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 643 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 128 + ], + "type": "text", + "content": "[42] Jia-Wen Xiao, Chang-Bin Zhang, Jiekang Feng, Xialei Liu, Joost van de Weijer, and Ming-Ming Cheng. Endpoints weight fusion for class incremental semantic segmentation. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 7204-7213, 2023. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 129, + 553, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 553, + 173 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 553, + 173 + ], + "type": "text", + "content": "[43] Shipeng Yan, Jiangwei Xie, and Xuming He. Der: Dynamically expandable representation for class incremental learning. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 3014-3023, 2021. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 174, + 553, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 174, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 316, + 174, + 553, + 228 + ], + "type": "text", + "content": "[44] Chaohui Yu, Qiang Zhou, Jingliang Li, Jianlong Yuan, Zhibin Wang, and Fan Wang. Foundation model drives weakly incremental learning for semantic segmentation. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 23685-23694, 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 229, + 553, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 229, + 553, + 274 + ], + "spans": [ + { + "bbox": [ + 316, + 229, + 553, + 274 + ], + "type": "text", + "content": "[45] Qihang Yu, Huiyu Wang, Siyuan Qiao, Maxwell Collins, Yukun Zhu, Hartwig Adam, Alan Yuille, and Liang-Chieh Chen. k-means mask transformer. In Proc. Eur. Conf. Comput. Vis., pages 288-307, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 274, + 553, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 274, + 553, + 317 + ], + "spans": [ + { + "bbox": [ + 316, + 274, + 553, + 317 + ], + "type": "text", + "content": "[46] Anqi Zhang and Guangyu Gao. Background adaptation with residual modeling for exemplar-free class-incremental semantic segmentation. In Proc. Eur. Conf. Comput. Vis., 2024. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 319, + 553, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 319, + 553, + 373 + ], + "spans": [ + { + "bbox": [ + 316, + 319, + 553, + 373 + ], + "type": "text", + "content": "[47] Chang-Bin Zhang, Jia-Wen Xiao, Xialei Liu, Ying-Cong Chen, and Ming-Ming Cheng. Representation compensation networks for continual semantic segmentation. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 7053-7064, 2022. 2, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 374, + 553, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 374, + 553, + 419 + ], + "spans": [ + { + "bbox": [ + 316, + 374, + 553, + 419 + ], + "type": "text", + "content": "[48] Zekang Zhang, Guangyu Gao, Zhiyuan Fang, Jianbo Jiao, and Yunchao Wei. Mining unseen classes via regional objectness: A simple baseline for incremental segmentation. Adv. Neural Inf. Process. Syst., 35:24340-24353, 2022. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 419, + 553, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 419, + 553, + 464 + ], + "spans": [ + { + "bbox": [ + 316, + 419, + 553, + 464 + ], + "type": "text", + "content": "[49] Zekang Zhang, Guangyu Gao, Jianbo Jiao, Chi Harold Liu, and Yunchao Wei. Coinseg: Contrast inter-and intra-class representations for incremental segmentation. In Proc. IEEE Int. Conf. Comput. Vis., pages 843-853, 2023. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 464, + 553, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 464, + 553, + 509 + ], + "spans": [ + { + "bbox": [ + 316, + 464, + 553, + 509 + ], + "type": "text", + "content": "[50] Hengshuang Zhao, Jianping Shi, Xiaojuan Qi, Xiaogang Wang, and Jiaya Jia. Pyramid scene parsing network. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 2881-2890, 2017. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 509, + 553, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 509, + 553, + 552 + ], + "spans": [ + { + "bbox": [ + 316, + 509, + 553, + 552 + ], + "type": "text", + "content": "[51] Hanbin Zhao, Fengyu Yang, Xinghe Fu, and Xi Li. Rbc: Rectifying the biased context in continual semantic segmentation. In Proc. Eur. Conf. Comput. Vis., pages 55-72, 2022. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 554, + 553, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 554, + 553, + 599 + ], + "spans": [ + { + "bbox": [ + 316, + 554, + 553, + 599 + ], + "type": "text", + "content": "[52] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 633-641, 2017. 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 600, + 553, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 600, + 553, + 643 + ], + "spans": [ + { + "bbox": [ + 316, + 600, + 553, + 643 + ], + "type": "text", + "content": "[53] Lanyun Zhu, Tianrun Chen, Jianxiong Yin, Simon See, and Jun Liu. Continual semantic segmentation with automatic memory sample selection. In Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pages 3082-3092, 2023. 2" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "25676" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/CoMM_ A Coherent Interleaved Image-Text Dataset for Multimodal Understanding and Generation/4d3dc3b4-7f37-464f-b503-0ded8f6008de_content_list.json b/2025/CoMM_ A Coherent Interleaved Image-Text Dataset for Multimodal Understanding and Generation/4d3dc3b4-7f37-464f-b503-0ded8f6008de_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..77006ffbcacc2135e4813b3f32a9d7a81f3e8cda --- /dev/null +++ b/2025/CoMM_ A Coherent Interleaved Image-Text Dataset for Multimodal Understanding and Generation/4d3dc3b4-7f37-464f-b503-0ded8f6008de_content_list.json @@ -0,0 +1,1557 @@ +[ + { + "type": "text", + "text": "CoMM: A Coherent Interleaved Image-Text Dataset for Multimodal Understanding and Generation", + "text_level": 1, + "bbox": [ + 233, + 130, + 767, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Wei Chen $^{1,2*}$ , Lin Li $^{1,3*}$ , Yongqi Yang $^{2*}$ , Bin Wen $^{4}$ , Fan Yang $^{4}$ , Tingting Gao $^{4}$ , Yu Wu $^{2\\dagger}$ , Long Chen $^{1\\dagger}$ $^{1}$ The Hong Kong University of Science and Technology, $^{2}$ Wuhan University, $^{3}$ AI Chip Center for Emerging Smart Systems, $^{4}$ Kuaishou Technology", + "bbox": [ + 197, + 202, + 802, + 276 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{wchendb, lllidy, longchen}@ust.hk, {yongqiyang, wuyucs}@whu.edu.cn", + "bbox": [ + 202, + 277, + 795, + 292 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "https://github.com/HKUST-LongGroup/CoMM", + "bbox": [ + 323, + 295, + 673, + 309 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 344, + 326, + 359 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Interleaved image-text generation has emerged as a vital multimodal task aimed at creating sequences of interleaved visual and textual content given a query. Despite notable advancements in recent multimodal large language models (MLLMs), generating integrated image-text sequences that exhibit narrative coherence and entity and style consistency remains challenging due to poor training data quality. To this end, we introduce CoMM, a high-quality Coherent interleaved image-text MultiModal dataset designed to enhance the coherence, consistency, and alignment of generated multimodal content. Initially, CoMM harnesses raw data from diverse sources, focusing on instructional content and visual storytelling, establishing a foundation for coherent and consistent content. To further refine the data quality, we devise a multi-perspective filter strategy that leverages advanced pre-trained models to ensure the development of sentences, consistency of inserted images, and semantic alignment between them. Various quality evaluation metrics are designed to prove the high quality of the filtered dataset. Meanwhile, extensive few-shot experiments on various downstream tasks demonstrate CoMM's effectiveness in significantly enhancing the in-context learning capabilities of MLLMs. Moreover, we propose four new tasks to evaluate MLLMs' interleaved generation abilities, supported by a comprehensive evaluation framework. We believe CoMM opens a new avenue for advanced MLLMs with superior multimodal in-context learning and understanding ability.", + "bbox": [ + 88, + 375, + 485, + 785 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 810, + 223, + 825 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Interleaved image-text generation has begeoned as an emerging multimodal task, striving to imitate the human-", + "bbox": [ + 89, + 835, + 483, + 867 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "like capability to alternately create visual and textual content [3, 45]. Specifically, it aims to generate a sequence of interleaved text descriptions and illustrative images given a query [3]. With its interleaved multimodal generation ability, state-of-the-art models can facilitate a variety of applications, e.g., multimodal instruction generation, tutorial step generation, and visual storytelling.", + "bbox": [ + 511, + 345, + 908, + 450 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent advancements in multimodal large language models (MLLMs) have exhibited exceptional capabilities in cross-modal generation, but they still struggle to generate interleaved image-text sequences coherently [36, 44]. The main reason is that most of them are trained on single imagetext pairs, limited in generating coherent and contextually integrated multimodal content [18, 44]. To this end, some efforts integrate multiple models (e.g., GPT-4 [1] and Stable Diffusion XL [28]) to generate open-domain interleaved image-text contents with arbitrary formats in a training-free manner [3]. However, their results are often dominated by the generated text descriptions, limited by the generation model's understanding of these textual inputs. In contrast, other research efforts [6, 10] have attempted to train models with seamless multimodal understanding and generation capabilities using existing interleaved image-text datasets (e.g., MMC4 [45]).", + "bbox": [ + 511, + 455, + 910, + 714 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Albeit the unprecedented progress, it is becoming increasingly evident that data-centric methods still suffer several limitations: 1) Narrative Coherence: The contextual relevance and coherence of generated image-text sentences are relatively low. In Figure 1(a), the model pre-trained on the MMC4 [45] directly draws the final cooked dessert, missing the intermediate coherent steps of generation. These coherent generation steps not only enhance interpretability but also boost the reasoning capabilities of the model [41]. 2) Entity and Style Consistency: The entity and subject styles of the generated illustrated images are inconsistent. As illustrated in Figure 1(b), the style of the blue-gray sofa image", + "bbox": [ + 511, + 719, + 910, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution. Work was done when Wei Chen interned in Kuaishou. † Corresponding authors.", + "bbox": [ + 89, + 875, + 482, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "8073", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/422e031e86cb9a9247ec9858f00403663e8850d2300d4b483d9e933327b7febe.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 93, + 92, + 898, + 364 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "(c) Sample from MMC4", + "text_level": 1, + "bbox": [ + 209, + 382, + 336, + 393 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "When it comes to making life just a little bit smoother and safer in a wonderful place such as White Bear Lake, ADT Monitored Home Security can be your main solution for keeping your family and belongings safe and secure.", + "bbox": [ + 114, + 396, + 385, + 435 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "An ADT home monitoring system for your home can change your life for the better, letting you explore and enjoy all that you love in White Bear Lake. ADT home monitoring is more than just home security, it raises the bar to better living.", + "bbox": [ + 114, + 440, + 385, + 479 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "It's a chance to live your daily life without all the worry and stress. For over a century, ADT monitoring systems have paved the way in home security.", + "bbox": [ + 114, + 484, + 375, + 513 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "As the #1 home security provider in the country, we know all the vital facts and components that are necessary in keeping our $6+$ million home and business customers safe. It's why so many people in White Bear Lake and the rest of the nation count on us for their safety.", + "bbox": [ + 114, + 516, + 441, + 554 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/3bc55a838fd7d914e38e61a57768a6e3631f8bf27f0b3c4d338e24b8cb03591f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 390, + 400, + 444, + 430 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/fc53a4dd0cfeee1bc185222928bc5b136f1989ed264891ebd792b00625a91a7a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 395, + 441, + 434, + 470 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/600f8c0011ccc3ddf62fd093b815e5d4b11e495b3331c2b1cad99d7be28cc04e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 395, + 486, + 431, + 512 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "generated by the model pre-trained on MMC4 is inconsistent with that in the query. This consistency of entities and style has been proposed as an important metric for evaluating the results of interleaved image-text content generation [3]. These limitations primarily stem from the low quality of training data caused by the poor alignment between image and text content, i.e., the devil is in the data. For example in Figure 1(c), the second image in the training sample has a noticeable gap with the context. Additionally, the styles of all the images are almost completely different. Thus, it raises a natural research question: How to construct a high-quality interleaved image-text dataset?", + "bbox": [ + 89, + 619, + 483, + 800 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we introduce CoMM, a high-quality and coherent interleaved image-text dataset. By \"high-quality\", we mean that the interleaved image-text content should exhibit developed and logically structured text steps, consistent entities and visual styles in images, and strong semantic and contextual alignment between them. To enhance text coher", + "bbox": [ + 89, + 810, + 483, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "(d) Sample from CoMM", + "text_level": 1, + "bbox": [ + 607, + 380, + 738, + 391 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Step1. Remove the tuna steaks from the refrigerator. Allow them to sit out at room temperature for 1/2 to 1 hour before you begin cooking.", + "bbox": [ + 488, + 396, + 826, + 416 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/f96c0ed39e4909f58e8040b510f9c1b5d3e385ff16c3e687d32a48c5d3448eab.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 828, + 395, + 897, + 421 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Step2. Use a paper towel to dry excess moisture from the tuna steaks. Simply pat the paper towel on both sides of the steak to remove marinade or condensation from the meat.", + "bbox": [ + 488, + 419, + 859, + 439 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/eb83bd16c70cc7505f21c4efb375e0ba6b4874dae1a77599aa1043a739777c3a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 861, + 422, + 898, + 479 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Step3. Add 2 tbsp. olive oil or butter to the pan over medium-high heat. Tilt the pan to spread the oil or butter across the surface and coat the pan.", + "bbox": [ + 490, + 441, + 833, + 463 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Step4. Place the tuna steaks on the hot pan. Do not crowd the tuna in the pan. Keep the steaks at least 3 to 4 inches apart to make sure they cook evenly.", + "bbox": [ + 488, + 465, + 846, + 484 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Step5. Sear each side for two minutes. Do not move the meat around too much while it sears. Gently shake the pan after two minutes, and if the tuna steak releases from the surface, you can flip it on its other side. It is important to make sure that the edged are seared, but you want the inside of the tuna steak to remain rare.", + "bbox": [ + 488, + 491, + 839, + 530 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Step6. Make a small cut in the center of the tuna steak with a knife. Inspect the inside to make sure that the middle of the steak stays rare and red in color.", + "bbox": [ + 488, + 534, + 841, + 553 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/6232e448cfc5218dc8fabbed84e8e6f84eec0fea452ff61ff6cbe9bc8e2d8bc5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 861, + 481, + 898, + 510 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/5cbb59f3a24307ac392a50f844c7b9e0b3ac386f61a12df697944ad3bd8e38d9.jpg", + "image_caption": [ + "Figure 1. Illustration of interleaved image-text content generation results and dataset quality. (a) Given the query from CoMM, the interleaved image-text content generation results from the model Emu2 [36] separately trained by MMC4 [45] and CoMM (Ours). (b) The query is from the MMC4. (c) A training sample is from the MMC4. (d) A training sample is from the our CoMM." + ], + "image_footnote": [], + "bbox": [ + 862, + 512, + 898, + 539 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ence and image consistency, we collect raw data from specific websites, e.g., WikiHow [42], known for their instructional content or visual stories. As illustrated in Figure 1(d), instructional content, due to its uniform intention (e.g., cooking steak) and structured presentation (e.g., step 1), exhibits strong text coherence and image consistency. Additionally, we devise a multi-perspective filter strategy consisting of: a text sequence filter, an image sequence filter, and an imagedtext alignment filter. The first two leverage large language models (LLMs), e.g., Llama3 [25], and vision-language models (VLMs), e.g., CLIP [29], to eliminate incoherent text and images, respectively, further enhancing overall coherence. The last one employs both VLMs and MLLMs to comprehensively assess and refine the dataset for better alignment between visual and textual elements. After constructing this dataset, we further enhance it into a preference dataset by generating negative samples through shuffled text and images. This preference dataset can then be used during the", + "bbox": [ + 511, + 619, + 908, + 891 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "8074", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "reinforcement learning stage of post-training [8, 16].", + "bbox": [ + 89, + 90, + 437, + 104 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To evaluate the data quality and make a comprehensive comparison of existing interleaved image-text benchmarks [19, 45], we design four metrics to evaluate the development, completeness, image-text alignment, and consistency of image-text sequences, separately (cf. Table 1). CoMM significantly outperforms previous datasets [19, 45] on all metrics, demonstrating its high quality. To further demonstrate the superiority of CoMM, we conduct few-shot experiments on various downstream tasks (e.g., VQA) [11, 12, 17, 22, 24], comparing with MMC4 [45] and OBELICS [19]. The model trained on our CoMM consistently outperforms others, particularly in long context settings (16 & 32 shots), showcasing CoMM's ability to enhance the in-context learning capacities of MLLMs.", + "bbox": [ + 89, + 107, + 485, + 316 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Furthermore, thanks to our dataset, we introduce four new challenging tasks to evaluate the multimodal understanding and generation abilities of MLLMs: image-to-text sequence generation, text-to-image sequence generation, interleaved image-text content continuation, and question-based interleaved image-text generation. To assess the performance of these newly introduced tasks, we implement a comprehensive evaluation framework considering metrics such as METEOR [5], ROUGE [21] for text quality, and FID [13] and IS [33] for image quality. To accurately evaluate open-ended interleaved image-text generation content, we develop a series of metrics and leverage powerful MLLM (e.g., GPT4o [27]) for assessment. These tasks and evaluations pave the way for novel approaches in evaluating MLLMs by addressing critical gaps in current benchmarks. Conclusively, our contributions are as follows:", + "bbox": [ + 89, + 319, + 486, + 561 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1) We introduce CoMM, a high-quality coherent interleaved image-text dataset designed to address the limitations of existing datasets by ensuring coherent narrative, consistent entity and style, and strong semantic alignment between images and text.", + "2) CoMM introduces four novel benchmark tasks and associated evaluation metrics to comprehensively evaluate multimodal understanding and generation capabilities of MLLMs.", + "3) CoMM can serve as both a supervised fine-tuning dataset and a preference dataset. Extensive ablations across various downstream tasks demonstrate the high-quality of CoMM and its effectiveness in enhancing the understanding and generation capabilities of MLLMs." + ], + "bbox": [ + 83, + 561, + 482, + 767 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 782, + 232, + 799 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Interleaved Image-Text Web Document Datasets. Training on interleaved image-text web documents has demonstrated superior performance over image-description pairs, as evidenced by studies such as Flamingo [2] and KOSMOS-1 [15]. These findings underscore the significant advantages of utilizing the richer and more meaningful correlations in", + "bbox": [ + 89, + 809, + 485, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "herent in interleaved image-text documents[19]. However, the training data in these two studies is not made publicly available.", + "bbox": [ + 511, + 90, + 906, + 133 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The scarcity of interleaved image-text data has been addressed by the introduction of two datasets: MMC4 [45] and OBELICS [19]. MMC4 builds upon the text-only C4 corpus [31] by integrating images into text passages using CLIP features. In contrast, OBELICS emphasizes comprehensive filtering strategies and preserves the original structure of web pages. These datasets present a diverse and extensive interleaved image-text corpus for multimodal language models. Despite their depth and breadth, they exhibit notable deficiencies in document quality, such as weaker completeness and image-text coherence (cf., Table 1). Additionally, as depicted in Figure 2, their documents lack comprehensiveness in the image modality, with most documents containing only one or two images. To address this, we propose CoMM that focuses explicitly on coherence-rich scenarios with more image illustrations. CoMM greatly enhances the baseline's [4] performance (in Table 2), highlighting the critical role of a high-quality coherent interleaved image-text training corpus.", + "bbox": [ + 511, + 136, + 908, + 422 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Modeling of Interleaved Image-Text Data. The impressive performance of language modeling by LLM motivates researchers to delve deeper into the comprehension and generation capabilities of multimodal data. Flamingo [2] introduces image tokens into the language modeling process, facilitating interleaved image-text input and enabling few-shot transfer to tasks (e.g., VQA). Emu [37] leverages Stable Diffusion [32] as the image decoder, enabling the generation of image or text from interleaved image-text input. DreamLLM [6] pioneers the generation of free-form interleaved image-text outputs by modeling text and image within a unified multimodal space.", + "bbox": [ + 511, + 422, + 908, + 604 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recent research has increasingly focused on the intrinsic requirement of interleaved image-text modeling: capturing multimodal coherence while generating interleaved image-text content. OpenLEAF [3] underscores the importance of maintaining semantic and stylistic consistency between generated images. It also proposes a training-free baseline and employs BingChat [26] to evaluate the two types of consistency. MM-Interleaved [38] introduces the multimodal feature synchronizer to extract context-sensitive, multi-scale image features, thereby more effectively capturing multimodal context coherence and visual consistency. Nonetheless, focusing solely on the model level to meet the intrinsic multimodal coherence requirement is insufficient. Therefore, our high-quality data that strictly adhere to this requirement are essential to fully realize the potential of MLLMs.", + "bbox": [ + 511, + 604, + 908, + 830 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. CoMM Dataset", + "text_level": 1, + "bbox": [ + 511, + 844, + 668, + 859 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The CoMM dataset is collected from high-quality interleaved image-text content sourced from various websites, focusing", + "bbox": [ + 511, + 869, + 906, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "8075", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "on instructional steps and visual stories. Besides, we apply multi-perspective quality filter strategies on text sequences, image sequences, and image-text alignment, leveraging advanced models, e.g., CLIP [29] and LLMs [25], to enhance the dataset's coherence and relevance1.", + "bbox": [ + 89, + 90, + 485, + 167 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Dataset Construction", + "text_level": 1, + "bbox": [ + 89, + 198, + 290, + 214 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Collection Process. As mentioned above, we gather specific interleaved image-text data to preliminarily ensure the coherence and consistency of our dataset. To achieve this, we explore websites that are likely to provide high-quality interleaved content, such as instructional steps and visual stories. We collect raw data from these sources1 and download the corresponding text and images. Initially, we apply basic data cleaning techniques, including de-duplication and filtering out images with low resolution, to refine the collected data.", + "bbox": [ + 89, + 229, + 483, + 364 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Discarding NSFW images. To address potential ethical implications, we follow the [45] by utilizing an advanced NSFW binary image classifier [7]. This classifier is devised as a 4-layer multi-layer perceptron (MLP) that has been fully trained on the NSFW dataset from LAION-2B [34]. Specifically, it utilizes features extracted from OpenAI's CLIP ViT-L/14 model[29] as input, achieving an impressive accuracy of $97.4\\%$ on the NSFW test dataset. For each image processed, we employ the classifier to predict the NSFW probability. Images yielding a probability greater than 0.1 are automatically excluded from further consideration.", + "bbox": [ + 89, + 372, + 483, + 539 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Qualities Filter Strategy. Due to the inherent noises in the gathered raw data, the data still contains inconsistent and irrelevant content. Thus, we further employ a multiperspective filter strategy to ensure the quality and coherence of the dataset. Specifically, it involves three following components1:", + "bbox": [ + 89, + 546, + 483, + 637 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1) Text Sequence Filter: For text sequences, the most important aspect is the contextual development and cohesive connection between them. To maintain a smooth and logical flow of the text, we employ LLMs (e.g., Llama3 [25]) to evaluate the development and coherence among text steps in a document (prompt details are in the Appendix). According to evaluation scores, we eliminate sequences that lack coherence and relevance in their content, producing a higher-quality text corpus.", + "2) Image Sequence Filter. As for image sequences, visual consistency and relevance are key factors in maintaining the quality of the content. To assess the coherence and development of image sequences, we devise a metric $\\mathcal{F}(\\cdot)$" + ], + "bbox": [ + 89, + 643, + 483, + 849 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "through the CLIP's visual encoder [29] in Eq. (1):", + "bbox": [ + 511, + 90, + 844, + 107 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {F} \\left(\\left\\{x _ {i} \\mid 1 \\leq i \\leq N \\right\\}\\right) = \\frac {1}{N - 1} \\sum_ {i = 2} ^ {N} \\operatorname {S i m} \\left(x _ {i}, x _ {i - 1}\\right) \\\\ - \\frac {2}{(N - 1) (N - 2)} \\sum_ {i = 2} ^ {N} \\sum_ {j = 1} ^ {i - 1} \\operatorname {S i m} \\left(x _ {i}, x _ {j}\\right), \\tag {1} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 535, + 112, + 906, + 199 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\{x_{i}|1\\leq i\\leq N\\}$ denotes the sequence of images from 1 to $N$ , $\\mathrm{Sim}(\\cdot ,\\cdot)$ denotes the cosine similarity score between two images calculated by the CLIP's visual encoder. The first term denotes the average similarity between consecutive images, promoting smooth transitions and visual coherence within the sequence, while the second term represents the average similarity between all pairs of images, ensuring overall diversity and development. In this way, we can effectively balance coherence and diversity, ensuring that the image sequence is both visually consistent and contextually developed. By applying the image sequence filter, we can identify and eliminate sequences that lack connection or development, resulting in more polished and visually appealing images.", + "bbox": [ + 511, + 205, + 906, + 417 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3) Image-Text Alignment Filter. This filter aims to ensure that the images and text are not only relevant to each other but also maintain a coherent narrative throughout the document. We first utilize CLIP [29] to calculate the CLIP similarity scores between images and text and remove those whose scores are less than 0.1. Since CLIP is trained on individual image-text pairs, where the text often serves as the image's caption, it tends to favor concrete descriptions (i.e., object descriptions and attributes) [20]. However, in an interleaved image-text document, the text for each step describes progress with associated images (e.g., in Figure 1(d), each text focuses on a cooking action linked to previous steps) which is associated with previous steps. Therefore, we also employ advanced model (e.g., GPT-4o [27] or Llama3 [25]) to assess and filter image-text pairs, as these models can better judge current image-text alignment by considering the previous context.", + "bbox": [ + 511, + 417, + 908, + 672 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Preference Dataset Construction. Preference learning [8] has recently gained popularity in the post-training of LLMs, leveraging reinforcement learning [16] to train models with both positive and negative samples. In our dataset, which consists of interleaved image and text sequences, negative samples are easily created by shuffling the order of steps. These shuffles can be categorized into four types: Shuffled Text: Text in different steps is shuffled, while images remain in their original order. Shuffled Image: Images are shuffled, while the text remains in its original order. Both Shuffled Text and Image: Both text and images are shuffled independently. Shuffled Steps: Images and text within each step remain aligned, but the sequence of steps is randomized.", + "bbox": [ + 511, + 674, + 908, + 869 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "These negative samples are combined with original positive samples to form a preference dataset, used for further", + "bbox": [ + 511, + 869, + 906, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "Due to space constraints, more details are left in the Appendix.", + "bbox": [ + 107, + 886, + 446, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "8076", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/f2d152959b3929e02c5645b94f88c1dd435451c980792871ca6f56397cc30311.jpg", + "image_caption": [ + "(a) MMC4" + ], + "image_footnote": [], + "bbox": [ + 91, + 111, + 343, + 294 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/f8394213a912db355d1a5d65a47cfb9f41b2a9f47d950b9343cbe1ca9bece90c.jpg", + "image_caption": [ + "(b) OBELICS", + "Figure 2. Visualization of the image-sentence numbers per document distribution of three datasets. The $\\mu$ and $M$ denote the mean/median number of images/sentences in documents, respectively." + ], + "image_footnote": [], + "bbox": [ + 370, + 111, + 622, + 294 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/914d0e0c5f497a4ea40ee10a12bb6eee8b00f2b865a01df43907138fcd9cb5b3.jpg", + "image_caption": [ + "(c) CoMM (Ours)" + ], + "image_footnote": [], + "bbox": [ + 655, + 111, + 905, + 294 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "model training following supervised fine-tuning.", + "bbox": [ + 89, + 338, + 410, + 354 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2. Dataset Statistic", + "text_level": 1, + "bbox": [ + 89, + 362, + 251, + 376 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After filtering, our dataset contains 227K documents with 2.28M images. Further, Figure 2 illustrates the distribution of image-sentence number per document across MMC4, OBELICS, and CoMM datasets. Darker colors indicate a higher number of documents. The $\\mu$ and $M$ represent the average and the most common number of images/sentences per document, respectively. For MMC4 and OBELICS, the distribution of images per document is highly concentrated around very low values of $M$ (2 and 1 per document, respectively). This indicates a significant scarcity of the image modality in their multimodal data, which directly limits the models' multimodal in-context learning ability.", + "bbox": [ + 88, + 385, + 483, + 566 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Examining Figure 2(a) and Figure 2(b) distributions, it is evident that the number of images does not significantly change with an increase in the number of sentences. However, in practical scenarios, longer documents are more likely to contain a greater number of illustrations, which aids in modeling longer multimodal contexts. In contrast, CoMM features a larger quantity of images per document ( $M = 4$ , $\\mu = 10.1$ ), which significantly enhances the dataset's image modality while focusing on multimodal coherence. Furthermore, CoMM exhibits a trend where longer documents tend to include more image illustrations, thereby facilitating the modeling of longer multimodal contexts.", + "bbox": [ + 89, + 566, + 483, + 748 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Interleaved Generation Task and Benchmark", + "text_level": 1, + "bbox": [ + 89, + 762, + 483, + 779 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Task Formulation", + "text_level": 1, + "bbox": [ + 89, + 786, + 264, + 801 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we propose four benchmark tasks to evaluate MLLMs' interleaved generation capabilities. For ease of presentation, we utilize $\\{x_{i}|1\\leq i\\leq N\\}$ and $\\{t_i|1\\leq i\\leq$ $N\\}$ to represent the image sequences and text sequences, respectively. Besides, we use $t_\\mathrm{prompt}$ as the text prompt, and $\\mathcal{F}_{\\theta}(\\cdot)$ as the model with parameter $\\theta$ . The detailed", + "bbox": [ + 89, + 809, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "formulations of tasks are outlined as follows:", + "bbox": [ + 511, + 338, + 812, + 352 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Task 1: Image-to-Text Sequence Generation. This task evaluates the model's capability to generate coherent and contextually appropriate narratives based on sequential images. The objective is to predict the corresponding text sequences from the given image sequences:", + "bbox": [ + 511, + 353, + 905, + 429 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\{t _ {i} | 1 \\leq i \\leq N \\} = \\mathcal {F} _ {\\theta} \\left(t _ {\\text {p r o m p t}}, x _ {1}, \\dots , x _ {N}\\right). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 578, + 439, + 906, + 455 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Task 2: Text-to-Image Sequence Generation. This task assesses the model's proficiency in generating image sequences that accurately represent provided textual descriptions. The goal is to predict the corresponding image sequences from the given text sequences:", + "bbox": [ + 511, + 465, + 906, + 541 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\{x _ {i} \\mid 1 \\leq i \\leq N \\} = \\mathcal {F} _ {\\theta} \\left(t _ {\\text {p r o m p t}}, t _ {1}, \\dots , t _ {N}\\right). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 578, + 551, + 906, + 566 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Task 3: Interleaved Image-Text Content Continuation. It measures the model's ability to seamlessly continue content in an interleaved image-text format. The model needs to predict the outputs of the remaining image-text steps based on the interleaved image-text content in the previous $k$ steps:", + "bbox": [ + 511, + 577, + 905, + 654 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{t _ {i}, x _ {i} \\mid k < i \\leq N \\right\\} = \\mathcal {F} _ {\\theta} \\left(t _ {\\text {p r o m p t}}, t _ {1}, x _ {1}, \\dots , t _ {k}, x _ {k}\\right). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 550, + 662, + 906, + 679 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Task 4: Question-based Interleaved Image-Text Generation. This challenging task requires the model to generate interleaved content based on a given question, testing its understanding and generative reasoning abilities across modalities. The model needs to predict the outcomes of all image-text steps based on the given question:", + "bbox": [ + 511, + 689, + 906, + 780 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{t _ {i}, x _ {i} \\mid 1 \\leq i \\leq N \\right\\} = \\mathcal {F} _ {\\theta} \\left(t _ {\\text {p r o m p t}}\\right). \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 601, + 790, + 906, + 806 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Evaluation Metric", + "text_level": 1, + "bbox": [ + 511, + 816, + 689, + 830 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This section lists the evaluation metrics for assessing the quality of generated results. Our evaluation framework1 measures textual accuracy, visual fidelity, and the coherence of multi-modal outputs.", + "bbox": [ + 511, + 839, + 905, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "8077", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation for Text Sequence Generation. Generating narratives for a series of images shares similarities with the established task of image captioning. Consequently, we directly apply traditional image captioning metrics like ROUGE [21] and METEOR [5] at both the step and document levels. Additionally, we propose Illustration Relevance Score (IRS) to assess the relevance between image illustration and text context in the final document. Specifically, we simulate the human document creation process by first generating detailed image descriptions from the text context and then assessing the consistency between these descriptions and corresponding images using GPT-4o [27].", + "bbox": [ + 89, + 90, + 483, + 272 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation for Image Sequence Generation. We evaluate the quality of image illustrations on four key assessment aspects including image quality, style consistency between generated images, reconstruction on referenced images, and IRS. To evaluate image quality, we use Fréchet Inception Distance (FID) [13] with ground truth reference and Inception Score (IS) [33]. For assessing style consistency, we employ the Structural Similarity Index Measure (SSIM), and for ground truth reconstruction evaluation, we adopt the Peak Signal-to-Noise Ratio (PSNR).", + "bbox": [ + 89, + 273, + 483, + 422 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation for Interleaved Image-Text Generation. We use MLLM (GPT-4o [27]) for a comprehensive assessment of linguistic and visual coherence. GPT-4o evaluates four aspects: image sequence coherence, image quality, document completeness, and IRS. For image sequence coherence, we check style, entity consistency, and content trend alignment between image and text sequence. Image quality is evaluated on authenticity, integrity, clarity, and aesthetics. Document completeness is scored based on how thoroughly the content covers the topic.", + "bbox": [ + 89, + 425, + 483, + 575 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 89, + 592, + 223, + 609 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Datasets Quality Comparison", + "text_level": 1, + "bbox": [ + 89, + 618, + 354, + 633 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Settings. In this section, we compare the quality of our dataset with previous interleaved image-text datasets (e.g., MMC4 [45] and OBELICS [19]) focusing on text sequences, image sequences, and image-text alignment. We utilize powerful LLMs (e.g., Llama3 [25] and GPT-4o [27]) to assess the text development, text completeness, and image-text alignment of these datasets. Each metric is scored on a scale of 0 to 10, with detailed definitions and prompts provided in the Appendix. For MMC4 and OBELICS, we randomly sample 5,000 documents from their data. We evaluate our entire dataset with Llama3 and randomly sample 5,000 documents for assessment by GPT-4o. To evaluate image sequence quality, we employ the metric $\\mathcal{F}$ as cited in Eq. (1).", + "bbox": [ + 89, + 642, + 482, + 852 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Quantitative Results. As shown in Table 1, our dataset surpasses MMC4 and OBELICS across all evaluated dimensions. Notably, the score difference in development,", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/9d030ec86f499745a453f06fcb81bcd7d6e309e586021bc507143b6efd96a2d0.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetsDLP L.3 / G.4oCPL L.3 / G.4oITA L.3 / G.4oImgS
MMC4 [45]5.56 / 4.756.28 / 5.126.53 / 4.660.21
OBELICS [19]5.60 / 5.976.51 / 5.885.00 / 3.811.00
CoMM (Ours)6.93 / 7.647.44 / 7.077.46 / 8.914.27
", + "bbox": [ + 517, + 88, + 903, + 162 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Quality comparison of interleaved image-text datasets. \"DLP\" stands for Development, \"CPL\" signifies Completeness, \"ITA\" represents Image-Text Alignment, \"ImgS\" refers to Image Sequence, \"L.3\" denotes Llama3 [25], and \"G.4o\" is GPT-4o [27].", + "bbox": [ + 511, + 172, + 906, + 229 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "completeness, and image-text alignment is more pronounced when evaluated by GPT-4o, a more powerful MLLM. Furthermore, the image sequence metric underscores the superior quality of our data, reflecting higher scores compared to MMC4 and OBELICS. These results affirm the high quality of our dataset in interleaved image-text data.", + "bbox": [ + 511, + 257, + 906, + 348 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. In-Context Multimodal Understanding", + "text_level": 1, + "bbox": [ + 511, + 358, + 848, + 375 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Setting. To further demonstrate the high-quality and validate the efficacy of our dataset in promoting in-context understanding, we conduct experiments using few-shot learning settings on several downstream tasks (COCO [22], Flickr30k [43], VQAv2 [11], OKVQA [24], TextVQA [35], VizWiz [12], and Hateful Memes [17]). Few-shot learning is critical for assessing a model's ability of in-context understanding. In our experiments, few-shot scenarios are simulated by providing the model with a small number of example queries and their corresponding answers before it is evaluated on new queries. We consider various shot settings (0, 4, 8, 16, and 32) to understand the model's performance across different levels of supervision.", + "bbox": [ + 511, + 381, + 908, + 578 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Quantitative Results. Table 2 presents the results for the baseline model and its enhancements with MMC4, OBELICS, and our proposed dataset. As observed from Table 2, our method consistently outperforms the performance of MMC4 and OBELICS across all shot settings, demonstrating superior generalization and in-context learning capabilities. Particularly noticeable improvements are seen in the COCO and TextVQA datasets, where our method exceeds the baseline by significant margins in nearly all shot settings. This superior performance underlines the strength of our dataset in enabling models to better learn from context and apply learned knowledge to unseen instances. Furthermore, our dataset surpasses the baseline in all tasks with more shot settings, maintaining its advantage with longer context few-shot data, as shown in the 16-shot and 32-shot scenarios.", + "bbox": [ + 511, + 579, + 908, + 820 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.3. Interleaved Generation Tasks", + "text_level": 1, + "bbox": [ + 511, + 830, + 774, + 845 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines. This section introduces baselines developed from several state-of-the-art approaches, i.e., MiniGPT-5 [44], SEED-Llama [9], and Emu2 [36], to assess their effective", + "bbox": [ + 511, + 854, + 908, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "8078", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/a05b89320f9fe057cd0f96aca50ed3a364ee274df6e565a1504723d7c6f6878a.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ShotCOCOFlickr30kVQAv2OKVQATextVQAVizWizHateful Memes
Baseline79.559.552.737.824.227.551.6
+ MMC4 [45]097.167.045.829.823.415.054.6
+ OBELICS [19]83.551.550.938.925.621.953.8
+ CoMM (Ours)100.369.852.641.333.523.654.8
Baseline89.065.854.840.128.234.154.0
+ MMC4 [45]4103.368.849.536.728.724.455.0
+ OBELICS [19]103.765.253.243.031.429.955.7
+ CoMM (Ours)107.073.253.642.935.730.155.2
Baseline96.362.954.841.129.138.554.7
+ MMC4 [45]8103.667.950.539.028.832.751.3
+ OBELICS [19]101.963.953.744.532.237.251.2
+ CoMM (Ours)108.074.854.244.535.437.952.5
Baseline98.862.854.342.727.342.553.9
+ MMC4 [45]16102.965.550.039.126.436.957.2
+ OBELICS [19]100.858.053.145.330.242.457.8
+ CoMM (Ours)109.371.554.246.935.843.559.9
Baseline99.561.353.342.423.844.053.8
+ MMC4 [45]3294.956.448.335.620.037.749.2
+ OBELICS [19]96.553.152.043.626.844.556.1
+ CoMM (Ours)111.970.754.046.634.945.757.3
", + "bbox": [ + 151, + 87, + 846, + 386 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/5609331f721975aa27d5a22255241d93c31a870655d7881264bedded52d07bf3.jpg", + "table_caption": [ + "Table 2. Performance comparison of different datasets, with the baseline model Open-Flamingo 9B [4]. Evaluations were in an open-ended setting for VQA tasks with random in-context examples. (Task, Metric, Query Split): (COCO [22], CIDEr, test), (Flickr30k [43], CIDEr, test (Karpathy)), (VQAv2 [11], VQA accuracy, testdev), (OKVQA [24], VQA accuracy, val), (TextVQA [35], VQA accuracy, val), (VizWiz [12], VQA accuracy, testdev), (Hateful Memes [17], ROC-AUC, test seen)." + ], + "table_footnote": [], + "table_body": "
MethodsSizeI2T Sequence Generation (Task1)StyleContinuation Generation (Task3)
M.S.R.S.M.W.R.W.IRSEntityTrendCPL.ImgQIRS
MiniGPT-5 [44]7B4.608.649.1010.677.395.585.215.246.336.362.56
SEED-Llama [9]8B2.296.344.207.336.996.285.845.726.286.552.92
SEED-Llama [9]14B3.168.125.718.178.136.686.226.136.666.673.23
Emu2 [36]33B3.658.417.438.625.018.227.997.978.498.622.42
MethodsSizeT2I Sequence Generation (Task2)StyleQuestion-based Generation (Task4)
FID ↓ISSSIMPSNRIRSEntityTrendCPL.ImgQIRS
MiniGPT-5 [44]7B56.517.6020.667.575.485.655.205.255.816.152.71
SEED-Llama [9]8B57.968.4020.537.875.277.556.816.155.136.361.46
SEED-Llama [9]14B66.2311.0620.838.126.247.516.616.306.136.662.50
Emu2 [36]33B60.456.3420.967.914.178.417.567.637.547.592.02
", + "bbox": [ + 107, + 465, + 890, + 638 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Comparison of performance among MiniGPT-5 [44], SEED-Llama [9], and Emu2 [37] on the four generation tasks. \"M.S.\" and \"R.S.\" represent METEOR [5] and ROUGE_L [21] for step-by-step evaluation, while \"M.W.\" and \"R.W.\" correspond to METEOR and ROUGE_L for whole document evaluation. \"CPL.\" stands for Completeness, \"ImgQ\" indicates Image Quality, \"IRS\" means Illustration Relevance Score, and \"↓\" denotes that lower values are better.", + "bbox": [ + 89, + 648, + 906, + 705 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ness in generating interleaved image-text content. Details of these models are thoroughly provided in the Appendix. To construct the baselines, we fine-tune each of these models using our CoMM dataset. We employ a rigorous evaluation, assessing them on benchmark tasks involving interleaved image-text generation and understanding.", + "bbox": [ + 89, + 732, + 482, + 823 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Quantitative Results. As shown in Table 3, the quantitative results clearly illustrate the varying performance of MiniGPT-5, SEED-Llama, and Emu2 across the four interleaved generation tasks. For I2T Sequence Generation (Task1), MiniGPT-5 achieves the highest scores in ME", + "bbox": [ + 89, + 824, + 485, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "TEOR, ROUGE_L, but falls short in Illustration Relevance Score (IRS) where SEED-Llama (14B) leads with a score of 8.13. For T2I Sequence Generation (Task2), MiniGPT-5 excels with the lowest FID score of 56.51, while SEED-Llama (14B) surpasses IS, SSIM, PSNR, and IRS metrics, suggesting stronger visual fidelity and semantic similarity. Emu2 tops in Continuation Generation (Task3) with the highest scores in Style, Entity, Trend, Completeness, and Image Quality. In Question-based Generation (Task4), Emu2 dominates in Style, Entity, Trend, Completeness, and Image Quality, while MiniGPT-5 manages to secure the lead in IRS.", + "bbox": [ + 511, + 732, + 908, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "8079", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/421bd8ca184d99258631bc823a2fba0d2780d95e00f20dd6384697c50f098283.jpg", + "image_caption": [ + "Figure 3. Visualization of interleaved image-text content generation from SEED-Llama [9] (Top) and MiniGPT-5 [44] (Bottom) separately trained by MMC4 [45] and CoMM [Ours]." + ], + "image_footnote": [], + "bbox": [ + 96, + 90, + 893, + 494 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Overall, these results highlight the strengths and weaknesses of each model in handling different aspects of interleaved image-text generation tasks, providing valuable insights for future research and model improvements.", + "bbox": [ + 88, + 537, + 482, + 598 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Preference Dataset Training Results. With SEED-Llama [9] unifying image to discrete token prediction modeling, we can directly apply the mainstream preference learning algorithm DPO [30] to train SEED-Llama on our preference dataset following supervised fine-tuning on our dataset. As shown in Table 4, performance further improves on the image-to-text sequence generation task after training with our preference dataset. Additional results on preference dataset training are available in the Appendix.", + "bbox": [ + 88, + 598, + 482, + 733 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Multimodal Generation Qualitative Analysis. As illustrated in Figure 1 and Figure 3, the interleaved image-text content generated by models [9, 36, 44] trained on MMC4 [45] shows a noticeable difference when compared to the content generated by models trained on our datasets. Our models produce content that is more consistent with the previous input context (e.g., textual coherence and entity consistency). By ensuring higher coherence, consistency, and alignment, CoMM sets a new standard for training datasets in the realm of multimodal large language models. More visualizations are available in the Appendix.", + "bbox": [ + 88, + 734, + 482, + 901 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/54efeb62378bfd9fa11e8e0b32cfa569f4cf31472b5a7d83558e64f404ea848e.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsSizeI2T Sequence Generation (Task1)
M.S.R.S.M.W.R.W.IRS
SEED-Llama [9]8B2.296.344.207.336.99
+ DPO [30]8B4.098.836.9510.247.26
SEED-Llama [9]14B3.168.125.718.178.13
+ DPO [30]14B4.6210.158.0810.308.18
", + "bbox": [ + 517, + 534, + 898, + 623 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. Performance results of SEED-Llama [9] trained by DPO [30] in our preference dataset. \"M.S.\" and \"R.S.\" represent METEOR [5] and ROUGE_L [21] for step-by-step evaluation, while \"M.W.\" and \"R.W.\" correspond to METEOR and ROUGE_L for whole document evaluation.", + "bbox": [ + 511, + 627, + 906, + 696 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 708, + 633, + 724 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we present CoMM, a high-quality interleaved image-text dataset designed to address the limitations of existing multimodal large language models (MLLMs) by ensuring superior text-image alignment and entity consistency. Through diverse interleaved image-text content, CoMM significantly enhances MLLMs' in-context learning capabilities, as demonstrated by extensive few-shot experiments. Additionally, we introduce four novel tasks and a comprehensive evaluation framework to assess interleaved image-text generation abilities, setting new benchmarks for advancing multimodal understanding and generation in MLLMs.", + "bbox": [ + 509, + 734, + 908, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8080", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgment", + "text_level": 1, + "bbox": [ + 91, + 90, + 240, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported by the Hong Kong SAR RGC Early Career Scheme (26208924), the National Natural Science Foundation of China Young Scholar Fund (62402408), and the HKUST Sports Science and Technology Research Grant (SSTRG24EG04), ACCESS - AI Chip Center for Emerging Smart Systems, sponsored by InnoHK funding, Hong Kong SAR.", + "bbox": [ + 89, + 114, + 483, + 219 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 232, + 186, + 248 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florecia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 1", + "[2] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. NeurIPS, 2022. 3", + "[3] Jie An, Zhengyuan Yang, Linjie Li, Jianfeng Wang, Kevin Lin, Zicheng Liu, Lijuan Wang, and Jiebo Luo. Openleaf: Open-domain interleaved image-text generation and evaluation. arXiv preprint arXiv:2310.07749, 2023. 1, 2, 3", + "[4] Anas Awadalla, Irena Gao, Josh Gardner, Jack Hessel, Yusuf Hanafy, Wanrong Zhu, Kalyani Marathe, Yonatan Bitton, Samir Gadre, Shiori Sagawa, et al. Openflamingo: An opensource framework for training large autoregressive vision-language models. arXiv preprint arXiv:2308.01390, 2023. 3, 7", + "[5] Satanjeev Banerjee and Alon Lavie. Meteor: An automatic metric for mt evaluation with improved correlation with human judgments. In ACL Workshop, pages 65-72, 2005. 3, 6, 7, 8", + "[6] Runpei Dong, Chunrui Han, Yuang Peng, Zekun Qi, Zheng Ge, Jinrong Yang, Liang Zhao, Jianjian Sun, Hongyu Zhou, Haoran Wei, et al. Dreamllm: Synergistic multimodal comprehension and creation. arXiv preprint arXiv:2309.11499, 2023. 1, 3", + "[7] Samir Yitzhak Gadre, Gabriel Ilharco, Alex Fang, Jonathan Hayase, Georgios Smyrnis, Thao Nguyen, Ryan Marten, Mitchell Wortsman, Dhruba Ghosh, Jieyu Zhang, et al. Datacomp: In search of the next generation of multimodal datasets. NeurIPS, 36, 2023. 4", + "[8] Bofei Gao, Feifan Song, Yibo Miao, Zefan Cai, Zhe Yang, Liang Chen, Helan Hu, Runxin Xu, Qingxiu Dong, Ce Zheng, et al. Towards a unified view of preference learning for large language models: A survey. arXiv preprint arXiv:2409.02795, 2024. 3, 4", + "[9] Yuying Ge, Sijie Zhao, Ziyun Zeng, Yixiao Ge, Chen Li, Xintao Wang, and Ying Shan. Making llama see and draw with seed tokenizer. arXiv preprint arXiv:2310.01218, 2023. 6, 7, 8, 1, 3, 4, 9, 10, 11, 12, 13", + "[10] Yuying Ge, Sijie Zhao, Jinguo Zhu, Yixiao Ge, Kun Yi, Lin Song, Chen Li, Xiaohan Ding, and Ying Shan. Seed-x: Multimodal models with unified multi-granularity comprehension and generation. arXiv preprint arXiv:2404.14396, 2024. 1" + ], + "bbox": [ + 93, + 257, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[11] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Bartra, and Devi Parikh. Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In CVPR, 2017. 3, 6, 7", + "[12] Danna Gurari, Qing Li, Abigale J Stangl, Anhong Guo, Chi Lin, Kristen Grauman, Jiebo Luo, and Jeffrey P Bigham. Vizwiz grand challenge: Answering visual questions from blind people. In CVPR, 2018. 3, 6, 7", + "[13] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. NeurIPS, 2017. 3, 6", + "[14] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 7", + "[15] S Huang, L Dong, W Wang, Y Hao, S Singhal, S Ma, T Lv, L Cui, OK Mohammed, Q Liu, et al. Language is not all you need: aligning perception with language models. 2023. arXiv preprint arXiv:2302.14045. 3", + "[16] Timo Kaufmann, Paul Weng, Viktor Bengs, and Eyke Hüllermieier. A survey of reinforcement learning from human feedback. arXiv preprint arXiv:2312.14925, 2023. 3, 4", + "[17] Douwe Kiela, Hamed Firooz, Aravind Mohan, Vedanuj Goswami, Amanpreet Singh, Pratik Ringshia, and Davide Testuggine. The hateful memes challenge: Detecting hate speech in multimodal memes. NeurIPS, 2020. 3, 6, 7", + "[18] Jing Yu Koh, Daniel Fried, and Russ R Salakhutdinov. Generating images with multimodal language models. NeurIPS, 2024. 1", + "[19] Hugo Laurençon, Lucile Saulnier, Léo Tronchon, Stas Bekman, Amanpreet Singh, Anton Lozhkov, Thomas Wang, Siddharth Karamcheti, Alexander Rush, Douwe Kiela, et al. Obelics: An open web-scale filtered dataset of interleaved image-text documents. NeurIPS, 2023. 3, 6, 7", + "[20] Samuel Lavoie, Polina Kirichenko, Mark Ibrahim, Mahmoud Assran, Andrew Gordon Wildon, Aaron Courville, and Nicolas Ballas. Modeling caption diversity in contrastive vision-language pretraining. arXiv preprint arXiv:2405.00740, 2024. 4", + "[21] Chin-Yew Lin. Rouge: A package for automatic evaluation of summaries. In Text summarization branches out, 2004. 3, 6, 7, 8", + "[22] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dolkar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 3, 6, 7", + "[23] Chang Liu, Haoning Wu, Yujie Zhong, Xiaoyun Zhang, and Weidi Xie. Intelligent grimm-open-ended visual storytelling via latent diffusion models. arXiv preprint arXiv:2306.00973, 2023. 1, 2", + "[24] Kenneth Marino, Mohammad Rastegari, Ali Farhadi, and Roozbeh Mottaghi. Ok-vqa: A visual question answering benchmark requiring external knowledge. In CVPR, 2019. 3, 6, 7", + "[25] Meta. https://llama.meta.com/llama3, 2024.2, 4,6,1" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "8081", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[26] Microsoft. Bingchat. https://www.microsoft.com/en-us/edge/features/bing-chat, 2023. Accessed: 2023-10-09. 3", + "[27] OpenAI. https://openai.com/index/hello-gpt-4o/, 2024.3,4,6", + "[28] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 1", + "[29] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 2, 4", + "[30] Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. Direct preference optimization: Your language model is secretly a reward model. NeurIPS, 36, 2023. 8, 1, 3", + "[31] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. JMLR, 2020. 3", + "[32] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 3", + "[33] Tim Salimans, Ian Goodfellow, Wojciech Zaremba, Vicki Cheung, Alec Radford, and Xi Chen. Improved techniques for training gans. NeurIPS, 2016. 3, 6", + "[34] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. NeurIPS, 35:25278–25294, 2022. 4", + "[35] Amanpreet Singh, Vivek Natarajan, Meet Shah, Yu Jiang, Xinlei Chen, Dhruv Batra, Devi Parikh, and Marcus Rohrbach. Towards vqa models that can read. In CVPR, 2019. 6, 7", + "[36] Quan Sun, Yufeng Cui, Xiaosong Zhang, Fan Zhang, Qiying Yu, Zhengxiong Luo, Yueze Wang, Yongming Rao, Jingjing Liu, Tiejun Huang, et al. Generative multimodal models are in-context learners. arXiv preprint arXiv:2312.13286, 2023. 1, 2, 6, 7, 8, 4, 9, 10, 11, 12", + "[37] Quan Sun, Qiying Yu, Yufeng Cui, Fan Zhang, Xiaosong Zhang, Yueze Wang, Hongcheng Gao, Jingjing Liu, Tiejun Huang, and Xinlong Wang. Emu: Generative pretraining in multimodality. In ICLR, 2023. 3, 7", + "[38] Changyao Tian, Xizhou Zhu, Yuwen Xiong, Weiyun Wang, Zhe Chen, Wenhai Wang, Yuntao Chen, Lewei Lu, Tong Lu, Jie Zhou, et al. Mm-interleaved: Interleaved image-text generative modeling via multi-modal feature synchronizer. arXiv preprint arXiv:2401.10208, 2024. 3", + "[39] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 7", + "[40] Weihan Wang, Qingsong Lv, Wenmeng Yu, Wenyi Hong, Ji Qi, Yan Wang, Junhui Ji, Zhuoyi Yang, Lei Zhao, Xixuan" + ], + "bbox": [ + 91, + 92, + 485, + 902 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Song, Jiazheng Xu, Bin Xu, Juanzi Li, Yuxiao Dong, Ming Ding, and Jie Tang. CogvIm: Visual expert for pretrained language models, 2023. 4", + "[41] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. NeurIPS, 2022. 1", + "[42] WikiHow. https://www.wikipiow.com. 2,3", + "[43] Peter Young, Alice Lai, Micah Hodosh, and Julia Hockenmaier. From image descriptions to visual denotations: New similarity metrics for semantic inference over event descriptions. TACL, 2014. 6, 7", + "[44] Kaizhi Zheng, Xuehai He, and Xin Eric Wang. Minigpt-5: Interleaved vision-and-language generation via generative vowels. arXiv preprint arXiv:2310.02239, 2023. 1, 6, 7, 8, 4, 9, 10, 11, 12", + "[45] Wanrong Zhu, Jack Hessel, Anas Awadalla, Samir Yitzhak Gadre, Jesse Dodge, Alex Fang, Youngjae Yu, Ludwig Schmidt, William Yang Wang, and Yejin Choi. Multimodal c4: An open, billion-scale corpus of images interleaved with text. NeurIPS, 2023. 1, 2, 3, 4, 6, 7, 8" + ], + "bbox": [ + 516, + 92, + 906, + 388 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "8082", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/CoMM_ A Coherent Interleaved Image-Text Dataset for Multimodal Understanding and Generation/4d3dc3b4-7f37-464f-b503-0ded8f6008de_model.json b/2025/CoMM_ A Coherent Interleaved Image-Text Dataset for Multimodal Understanding and Generation/4d3dc3b4-7f37-464f-b503-0ded8f6008de_model.json new file mode 100644 index 0000000000000000000000000000000000000000..737f159b68fc0f7b16a81a92b145154d85ac4dc3 --- /dev/null +++ b/2025/CoMM_ A Coherent Interleaved Image-Text Dataset for Multimodal Understanding and Generation/4d3dc3b4-7f37-464f-b503-0ded8f6008de_model.json @@ -0,0 +1,2079 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.131, + 0.768, + 0.177 + ], + "angle": 0, + "content": "CoMM: A Coherent Interleaved Image-Text Dataset for Multimodal Understanding and Generation" + }, + { + "type": "text", + "bbox": [ + 0.198, + 0.203, + 0.803, + 0.277 + ], + "angle": 0, + "content": "Wei Chen\\(^{1,2*}\\), Lin Li\\(^{1,3*}\\), Yongqi Yang\\(^{2*}\\), Bin Wen\\(^{4}\\), Fan Yang\\(^{4}\\), Tingting Gao\\(^{4}\\), Yu Wu\\(^{2\\dagger}\\), Long Chen\\(^{1\\dagger}\\) \n\\(^{1}\\)The Hong Kong University of Science and Technology, \\(^{2}\\)Wuhan University, \\(^{3}\\)AI Chip Center for Emerging Smart Systems, \\(^{4}\\)Kuaishou Technology" + }, + { + "type": "text", + "bbox": [ + 0.204, + 0.278, + 0.796, + 0.293 + ], + "angle": 0, + "content": "{wchendb, lllidy, longchen}@ust.hk, {yongqiyang, wuyucs}@whu.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.325, + 0.296, + 0.674, + 0.31 + ], + "angle": 0, + "content": "https://github.com/HKUST-LongGroup/CoMM" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.345, + 0.328, + 0.361 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.376, + 0.486, + 0.786 + ], + "angle": 0, + "content": "Interleaved image-text generation has emerged as a vital multimodal task aimed at creating sequences of interleaved visual and textual content given a query. Despite notable advancements in recent multimodal large language models (MLLMs), generating integrated image-text sequences that exhibit narrative coherence and entity and style consistency remains challenging due to poor training data quality. To this end, we introduce CoMM, a high-quality Coherent interleaved image-text MultiModal dataset designed to enhance the coherence, consistency, and alignment of generated multimodal content. Initially, CoMM harnesses raw data from diverse sources, focusing on instructional content and visual storytelling, establishing a foundation for coherent and consistent content. To further refine the data quality, we devise a multi-perspective filter strategy that leverages advanced pre-trained models to ensure the development of sentences, consistency of inserted images, and semantic alignment between them. Various quality evaluation metrics are designed to prove the high quality of the filtered dataset. Meanwhile, extensive few-shot experiments on various downstream tasks demonstrate CoMM's effectiveness in significantly enhancing the in-context learning capabilities of MLLMs. Moreover, we propose four new tasks to evaluate MLLMs' interleaved generation abilities, supported by a comprehensive evaluation framework. We believe CoMM opens a new avenue for advanced MLLMs with superior multimodal in-context learning and understanding ability." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.811, + 0.224, + 0.827 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.837, + 0.485, + 0.868 + ], + "angle": 0, + "content": "Interleaved image-text generation has begeoned as an emerging multimodal task, striving to imitate the human-" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.346, + 0.909, + 0.452 + ], + "angle": 0, + "content": "like capability to alternately create visual and textual content [3, 45]. Specifically, it aims to generate a sequence of interleaved text descriptions and illustrative images given a query [3]. With its interleaved multimodal generation ability, state-of-the-art models can facilitate a variety of applications, e.g., multimodal instruction generation, tutorial step generation, and visual storytelling." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.457, + 0.911, + 0.715 + ], + "angle": 0, + "content": "Recent advancements in multimodal large language models (MLLMs) have exhibited exceptional capabilities in cross-modal generation, but they still struggle to generate interleaved image-text sequences coherently [36, 44]. The main reason is that most of them are trained on single imagetext pairs, limited in generating coherent and contextually integrated multimodal content [18, 44]. To this end, some efforts integrate multiple models (e.g., GPT-4 [1] and Stable Diffusion XL [28]) to generate open-domain interleaved image-text contents with arbitrary formats in a training-free manner [3]. However, their results are often dominated by the generated text descriptions, limited by the generation model's understanding of these textual inputs. In contrast, other research efforts [6, 10] have attempted to train models with seamless multimodal understanding and generation capabilities using existing interleaved image-text datasets (e.g., MMC4 [45])." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.911, + 0.903 + ], + "angle": 0, + "content": "Albeit the unprecedented progress, it is becoming increasingly evident that data-centric methods still suffer several limitations: 1) Narrative Coherence: The contextual relevance and coherence of generated image-text sentences are relatively low. In Figure 1(a), the model pre-trained on the MMC4 [45] directly draws the final cooked dessert, missing the intermediate coherent steps of generation. These coherent generation steps not only enhance interpretability but also boost the reasoning capabilities of the model [41]. 2) Entity and Style Consistency: The entity and subject styles of the generated illustrated images are inconsistent. As illustrated in Figure 1(b), the style of the blue-gray sofa image" + }, + { + "type": "page_footnote", + "bbox": [ + 0.091, + 0.875, + 0.483, + 0.901 + ], + "angle": 0, + "content": "*Equal contribution. Work was done when Wei Chen interned in Kuaishou. † Corresponding authors." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8073" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.093, + 0.899, + 0.366 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.21, + 0.383, + 0.338, + 0.394 + ], + "angle": 0, + "content": "(c) Sample from MMC4" + }, + { + "type": "text", + "bbox": [ + 0.116, + 0.397, + 0.387, + 0.436 + ], + "angle": 0, + "content": "When it comes to making life just a little bit smoother and safer in a wonderful place such as White Bear Lake, ADT Monitored Home Security can be your main solution for keeping your family and belongings safe and secure." + }, + { + "type": "text", + "bbox": [ + 0.116, + 0.441, + 0.387, + 0.48 + ], + "angle": 0, + "content": "An ADT home monitoring system for your home can change your life for the better, letting you explore and enjoy all that you love in White Bear Lake. ADT home monitoring is more than just home security, it raises the bar to better living." + }, + { + "type": "text", + "bbox": [ + 0.116, + 0.486, + 0.377, + 0.514 + ], + "angle": 0, + "content": "It's a chance to live your daily life without all the worry and stress. For over a century, ADT monitoring systems have paved the way in home security." + }, + { + "type": "text", + "bbox": [ + 0.115, + 0.517, + 0.442, + 0.555 + ], + "angle": 0, + "content": "As the #1 home security provider in the country, we know all the vital facts and components that are necessary in keeping our \\(6+\\) million home and business customers safe. It's why so many people in White Bear Lake and the rest of the nation count on us for their safety." + }, + { + "type": "image", + "bbox": [ + 0.391, + 0.401, + 0.446, + 0.431 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.396, + 0.443, + 0.436, + 0.472 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.396, + 0.487, + 0.432, + 0.513 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.564, + 0.907, + 0.606 + ], + "angle": 0, + "content": "Figure 1. Illustration of interleaved image-text content generation results and dataset quality. (a) Given the query from CoMM, the interleaved image-text content generation results from the model Emu2 [36] separately trained by MMC4 [45] and CoMM (Ours). (b) The query is from the MMC4. (c) A training sample is from the MMC4. (d) A training sample is from the our CoMM." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.62, + 0.484, + 0.801 + ], + "angle": 0, + "content": "generated by the model pre-trained on MMC4 is inconsistent with that in the query. This consistency of entities and style has been proposed as an important metric for evaluating the results of interleaved image-text content generation [3]. These limitations primarily stem from the low quality of training data caused by the poor alignment between image and text content, i.e., the devil is in the data. For example in Figure 1(c), the second image in the training sample has a noticeable gap with the context. Additionally, the styles of all the images are almost completely different. Thus, it raises a natural research question: How to construct a high-quality interleaved image-text dataset?" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.485, + 0.901 + ], + "angle": 0, + "content": "In this paper, we introduce CoMM, a high-quality and coherent interleaved image-text dataset. By \"high-quality\", we mean that the interleaved image-text content should exhibit developed and logically structured text steps, consistent entities and visual styles in images, and strong semantic and contextual alignment between them. To enhance text coher" + }, + { + "type": "title", + "bbox": [ + 0.609, + 0.381, + 0.739, + 0.392 + ], + "angle": 0, + "content": "(d) Sample from CoMM" + }, + { + "type": "text", + "bbox": [ + 0.49, + 0.397, + 0.828, + 0.417 + ], + "angle": 0, + "content": "Step1. Remove the tuna steaks from the refrigerator. Allow them to sit out at room temperature for 1/2 to 1 hour before you begin cooking." + }, + { + "type": "image", + "bbox": [ + 0.829, + 0.396, + 0.898, + 0.422 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.49, + 0.42, + 0.86, + 0.44 + ], + "angle": 0, + "content": "Step2. Use a paper towel to dry excess moisture from the tuna steaks. Simply pat the paper towel on both sides of the steak to remove marinade or condensation from the meat." + }, + { + "type": "image", + "bbox": [ + 0.862, + 0.423, + 0.9, + 0.481 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.491, + 0.443, + 0.834, + 0.464 + ], + "angle": 0, + "content": "Step3. Add 2 tbsp. olive oil or butter to the pan over medium-high heat. Tilt the pan to spread the oil or butter across the surface and coat the pan." + }, + { + "type": "text", + "bbox": [ + 0.49, + 0.466, + 0.847, + 0.486 + ], + "angle": 0, + "content": "Step4. Place the tuna steaks on the hot pan. Do not crowd the tuna in the pan. Keep the steaks at least 3 to 4 inches apart to make sure they cook evenly." + }, + { + "type": "text", + "bbox": [ + 0.49, + 0.492, + 0.84, + 0.531 + ], + "angle": 0, + "content": "Step5. Sear each side for two minutes. Do not move the meat around too much while it sears. Gently shake the pan after two minutes, and if the tuna steak releases from the surface, you can flip it on its other side. It is important to make sure that the edged are seared, but you want the inside of the tuna steak to remain rare." + }, + { + "type": "text", + "bbox": [ + 0.49, + 0.535, + 0.842, + 0.554 + ], + "angle": 0, + "content": "Step6. Make a small cut in the center of the tuna steak with a knife. Inspect the inside to make sure that the middle of the steak stays rare and red in color." + }, + { + "type": "image", + "bbox": [ + 0.862, + 0.482, + 0.9, + 0.511 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.863, + 0.513, + 0.9, + 0.54 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.62, + 0.909, + 0.892 + ], + "angle": 0, + "content": "ence and image consistency, we collect raw data from specific websites, e.g., WikiHow [42], known for their instructional content or visual stories. As illustrated in Figure 1(d), instructional content, due to its uniform intention (e.g., cooking steak) and structured presentation (e.g., step 1), exhibits strong text coherence and image consistency. Additionally, we devise a multi-perspective filter strategy consisting of: a text sequence filter, an image sequence filter, and an imagedtext alignment filter. The first two leverage large language models (LLMs), e.g., Llama3 [25], and vision-language models (VLMs), e.g., CLIP [29], to eliminate incoherent text and images, respectively, further enhancing overall coherence. The last one employs both VLMs and MLLMs to comprehensively assess and refine the dataset for better alignment between visual and textual elements. After constructing this dataset, we further enhance it into a preference dataset by generating negative samples through shuffled text and images. This preference dataset can then be used during the" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8074" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.438, + 0.106 + ], + "angle": 0, + "content": "reinforcement learning stage of post-training [8, 16]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.108, + 0.486, + 0.318 + ], + "angle": 0, + "content": "To evaluate the data quality and make a comprehensive comparison of existing interleaved image-text benchmarks [19, 45], we design four metrics to evaluate the development, completeness, image-text alignment, and consistency of image-text sequences, separately (cf. Table 1). CoMM significantly outperforms previous datasets [19, 45] on all metrics, demonstrating its high quality. To further demonstrate the superiority of CoMM, we conduct few-shot experiments on various downstream tasks (e.g., VQA) [11, 12, 17, 22, 24], comparing with MMC4 [45] and OBELICS [19]. The model trained on our CoMM consistently outperforms others, particularly in long context settings (16 & 32 shots), showcasing CoMM's ability to enhance the in-context learning capacities of MLLMs." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.32, + 0.487, + 0.562 + ], + "angle": 0, + "content": "Furthermore, thanks to our dataset, we introduce four new challenging tasks to evaluate the multimodal understanding and generation abilities of MLLMs: image-to-text sequence generation, text-to-image sequence generation, interleaved image-text content continuation, and question-based interleaved image-text generation. To assess the performance of these newly introduced tasks, we implement a comprehensive evaluation framework considering metrics such as METEOR [5], ROUGE [21] for text quality, and FID [13] and IS [33] for image quality. To accurately evaluate open-ended interleaved image-text generation content, we develop a series of metrics and leverage powerful MLLM (e.g., GPT4o [27]) for assessment. These tasks and evaluations pave the way for novel approaches in evaluating MLLMs by addressing critical gaps in current benchmarks. Conclusively, our contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.084, + 0.563, + 0.483, + 0.636 + ], + "angle": 0, + "content": "1) We introduce CoMM, a high-quality coherent interleaved image-text dataset designed to address the limitations of existing datasets by ensuring coherent narrative, consistent entity and style, and strong semantic alignment between images and text." + }, + { + "type": "text", + "bbox": [ + 0.084, + 0.636, + 0.483, + 0.693 + ], + "angle": 0, + "content": "2) CoMM introduces four novel benchmark tasks and associated evaluation metrics to comprehensively evaluate multimodal understanding and generation capabilities of MLLMs." + }, + { + "type": "text", + "bbox": [ + 0.084, + 0.693, + 0.483, + 0.768 + ], + "angle": 0, + "content": "3) CoMM can serve as both a supervised fine-tuning dataset and a preference dataset. Extensive ablations across various downstream tasks demonstrate the high-quality of CoMM and its effectiveness in enhancing the understanding and generation capabilities of MLLMs." + }, + { + "type": "list", + "bbox": [ + 0.084, + 0.563, + 0.483, + 0.768 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.784, + 0.233, + 0.8 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.486, + 0.901 + ], + "angle": 0, + "content": "Interleaved Image-Text Web Document Datasets. Training on interleaved image-text web documents has demonstrated superior performance over image-description pairs, as evidenced by studies such as Flamingo [2] and KOSMOS-1 [15]. These findings underscore the significant advantages of utilizing the richer and more meaningful correlations in" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.135 + ], + "angle": 0, + "content": "herent in interleaved image-text documents[19]. However, the training data in these two studies is not made publicly available." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.137, + 0.909, + 0.424 + ], + "angle": 0, + "content": "The scarcity of interleaved image-text data has been addressed by the introduction of two datasets: MMC4 [45] and OBELICS [19]. MMC4 builds upon the text-only C4 corpus [31] by integrating images into text passages using CLIP features. In contrast, OBELICS emphasizes comprehensive filtering strategies and preserves the original structure of web pages. These datasets present a diverse and extensive interleaved image-text corpus for multimodal language models. Despite their depth and breadth, they exhibit notable deficiencies in document quality, such as weaker completeness and image-text coherence (cf., Table 1). Additionally, as depicted in Figure 2, their documents lack comprehensiveness in the image modality, with most documents containing only one or two images. To address this, we propose CoMM that focuses explicitly on coherence-rich scenarios with more image illustrations. CoMM greatly enhances the baseline's [4] performance (in Table 2), highlighting the critical role of a high-quality coherent interleaved image-text training corpus." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.424, + 0.909, + 0.605 + ], + "angle": 0, + "content": "Modeling of Interleaved Image-Text Data. The impressive performance of language modeling by LLM motivates researchers to delve deeper into the comprehension and generation capabilities of multimodal data. Flamingo [2] introduces image tokens into the language modeling process, facilitating interleaved image-text input and enabling few-shot transfer to tasks (e.g., VQA). Emu [37] leverages Stable Diffusion [32] as the image decoder, enabling the generation of image or text from interleaved image-text input. DreamLLM [6] pioneers the generation of free-form interleaved image-text outputs by modeling text and image within a unified multimodal space." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.606, + 0.909, + 0.832 + ], + "angle": 0, + "content": "Recent research has increasingly focused on the intrinsic requirement of interleaved image-text modeling: capturing multimodal coherence while generating interleaved image-text content. OpenLEAF [3] underscores the importance of maintaining semantic and stylistic consistency between generated images. It also proposes a training-free baseline and employs BingChat [26] to evaluate the two types of consistency. MM-Interleaved [38] introduces the multimodal feature synchronizer to extract context-sensitive, multi-scale image features, thereby more effectively capturing multimodal context coherence and visual consistency. Nonetheless, focusing solely on the model level to meet the intrinsic multimodal coherence requirement is insufficient. Therefore, our high-quality data that strictly adhere to this requirement are essential to fully realize the potential of MLLMs." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.845, + 0.669, + 0.861 + ], + "angle": 0, + "content": "3. CoMM Dataset" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.908, + 0.903 + ], + "angle": 0, + "content": "The CoMM dataset is collected from high-quality interleaved image-text content sourced from various websites, focusing" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8075" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.168 + ], + "angle": 0, + "content": "on instructional steps and visual stories. Besides, we apply multi-perspective quality filter strategies on text sequences, image sequences, and image-text alignment, leveraging advanced models, e.g., CLIP [29] and LLMs [25], to enhance the dataset's coherence and relevance1." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.199, + 0.292, + 0.215 + ], + "angle": 0, + "content": "3.1. Dataset Construction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.23, + 0.485, + 0.366 + ], + "angle": 0, + "content": "Collection Process. As mentioned above, we gather specific interleaved image-text data to preliminarily ensure the coherence and consistency of our dataset. To achieve this, we explore websites that are likely to provide high-quality interleaved content, such as instructional steps and visual stories. We collect raw data from these sources1 and download the corresponding text and images. Initially, we apply basic data cleaning techniques, including de-duplication and filtering out images with low resolution, to refine the collected data." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.373, + 0.484, + 0.54 + ], + "angle": 0, + "content": "Discarding NSFW images. To address potential ethical implications, we follow the [45] by utilizing an advanced NSFW binary image classifier [7]. This classifier is devised as a 4-layer multi-layer perceptron (MLP) that has been fully trained on the NSFW dataset from LAION-2B [34]. Specifically, it utilizes features extracted from OpenAI's CLIP ViT-L/14 model[29] as input, achieving an impressive accuracy of \\(97.4\\%\\) on the NSFW test dataset. For each image processed, we employ the classifier to predict the NSFW probability. Images yielding a probability greater than 0.1 are automatically excluded from further consideration." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.547, + 0.484, + 0.638 + ], + "angle": 0, + "content": "Qualities Filter Strategy. Due to the inherent noises in the gathered raw data, the data still contains inconsistent and irrelevant content. Thus, we further employ a multiperspective filter strategy to ensure the quality and coherence of the dataset. Specifically, it involves three following components1:" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.645, + 0.484, + 0.781 + ], + "angle": 0, + "content": "1) Text Sequence Filter: For text sequences, the most important aspect is the contextual development and cohesive connection between them. To maintain a smooth and logical flow of the text, we employ LLMs (e.g., Llama3 [25]) to evaluate the development and coherence among text steps in a document (prompt details are in the Appendix). According to evaluation scores, we eliminate sequences that lack coherence and relevance in their content, producing a higher-quality text corpus." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.788, + 0.484, + 0.85 + ], + "angle": 0, + "content": "2) Image Sequence Filter. As for image sequences, visual consistency and relevance are key factors in maintaining the quality of the content. To assess the coherence and development of image sequences, we devise a metric \\(\\mathcal{F}(\\cdot)\\)" + }, + { + "type": "list", + "bbox": [ + 0.09, + 0.645, + 0.484, + 0.85 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.845, + 0.108 + ], + "angle": 0, + "content": "through the CLIP's visual encoder [29] in Eq. (1):" + }, + { + "type": "equation", + "bbox": [ + 0.536, + 0.113, + 0.907, + 0.2 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {F} \\left(\\left\\{x _ {i} \\mid 1 \\leq i \\leq N \\right\\}\\right) = \\frac {1}{N - 1} \\sum_ {i = 2} ^ {N} \\operatorname {S i m} \\left(x _ {i}, x _ {i - 1}\\right) \\\\ - \\frac {2}{(N - 1) (N - 2)} \\sum_ {i = 2} ^ {N} \\sum_ {j = 1} ^ {i - 1} \\operatorname {S i m} \\left(x _ {i}, x _ {j}\\right), \\tag {1} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.207, + 0.908, + 0.418 + ], + "angle": 0, + "content": "where \\(\\{x_{i}|1\\leq i\\leq N\\}\\) denotes the sequence of images from 1 to \\(N\\), \\(\\mathrm{Sim}(\\cdot ,\\cdot)\\) denotes the cosine similarity score between two images calculated by the CLIP's visual encoder. The first term denotes the average similarity between consecutive images, promoting smooth transitions and visual coherence within the sequence, while the second term represents the average similarity between all pairs of images, ensuring overall diversity and development. In this way, we can effectively balance coherence and diversity, ensuring that the image sequence is both visually consistent and contextually developed. By applying the image sequence filter, we can identify and eliminate sequences that lack connection or development, resulting in more polished and visually appealing images." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.419, + 0.909, + 0.674 + ], + "angle": 0, + "content": "3) Image-Text Alignment Filter. This filter aims to ensure that the images and text are not only relevant to each other but also maintain a coherent narrative throughout the document. We first utilize CLIP [29] to calculate the CLIP similarity scores between images and text and remove those whose scores are less than 0.1. Since CLIP is trained on individual image-text pairs, where the text often serves as the image's caption, it tends to favor concrete descriptions (i.e., object descriptions and attributes) [20]. However, in an interleaved image-text document, the text for each step describes progress with associated images (e.g., in Figure 1(d), each text focuses on a cooking action linked to previous steps) which is associated with previous steps. Therefore, we also employ advanced model (e.g., GPT-4o [27] or Llama3 [25]) to assess and filter image-text pairs, as these models can better judge current image-text alignment by considering the previous context." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.909, + 0.87 + ], + "angle": 0, + "content": "Preference Dataset Construction. Preference learning [8] has recently gained popularity in the post-training of LLMs, leveraging reinforcement learning [16] to train models with both positive and negative samples. In our dataset, which consists of interleaved image and text sequences, negative samples are easily created by shuffling the order of steps. These shuffles can be categorized into four types: Shuffled Text: Text in different steps is shuffled, while images remain in their original order. Shuffled Image: Images are shuffled, while the text remains in its original order. Both Shuffled Text and Image: Both text and images are shuffled independently. Shuffled Steps: Images and text within each step remain aligned, but the sequence of steps is randomized." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.908, + 0.901 + ], + "angle": 0, + "content": "These negative samples are combined with original positive samples to form a preference dataset, used for further" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.887, + 0.447, + 0.902 + ], + "angle": 0, + "content": "Due to space constraints, more details are left in the Appendix." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8076" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.206, + 0.097, + 0.277, + 0.111 + ], + "angle": 0, + "content": "(a) MMC4" + }, + { + "type": "image", + "bbox": [ + 0.092, + 0.112, + 0.344, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.475, + 0.097, + 0.567, + 0.112 + ], + "angle": 0, + "content": "(b) OBELICS" + }, + { + "type": "image", + "bbox": [ + 0.372, + 0.112, + 0.624, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.746, + 0.097, + 0.866, + 0.112 + ], + "angle": 0, + "content": "(c) CoMM (Ours)" + }, + { + "type": "image", + "bbox": [ + 0.656, + 0.112, + 0.906, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.297, + 0.908, + 0.325 + ], + "angle": 0, + "content": "Figure 2. Visualization of the image-sentence numbers per document distribution of three datasets. The \\(\\mu\\) and \\(M\\) denote the mean/median number of images/sentences in documents, respectively." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.339, + 0.411, + 0.355 + ], + "angle": 0, + "content": "model training following supervised fine-tuning." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.363, + 0.253, + 0.377 + ], + "angle": 0, + "content": "3.2. Dataset Statistic" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.386, + 0.484, + 0.568 + ], + "angle": 0, + "content": "After filtering, our dataset contains 227K documents with 2.28M images. Further, Figure 2 illustrates the distribution of image-sentence number per document across MMC4, OBELICS, and CoMM datasets. Darker colors indicate a higher number of documents. The \\(\\mu\\) and \\(M\\) represent the average and the most common number of images/sentences per document, respectively. For MMC4 and OBELICS, the distribution of images per document is highly concentrated around very low values of \\(M\\) (2 and 1 per document, respectively). This indicates a significant scarcity of the image modality in their multimodal data, which directly limits the models' multimodal in-context learning ability." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.568, + 0.484, + 0.749 + ], + "angle": 0, + "content": "Examining Figure 2(a) and Figure 2(b) distributions, it is evident that the number of images does not significantly change with an increase in the number of sentences. However, in practical scenarios, longer documents are more likely to contain a greater number of illustrations, which aids in modeling longer multimodal contexts. In contrast, CoMM features a larger quantity of images per document (\\(M = 4\\), \\(\\mu = 10.1\\)), which significantly enhances the dataset's image modality while focusing on multimodal coherence. Furthermore, CoMM exhibits a trend where longer documents tend to include more image illustrations, thereby facilitating the modeling of longer multimodal contexts." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.763, + 0.484, + 0.78 + ], + "angle": 0, + "content": "4. Interleaved Generation Task and Benchmark" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.787, + 0.265, + 0.802 + ], + "angle": 0, + "content": "4.1. Task Formulation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.484, + 0.901 + ], + "angle": 0, + "content": "In this section, we propose four benchmark tasks to evaluate MLLMs' interleaved generation capabilities. For ease of presentation, we utilize \\(\\{x_{i}|1\\leq i\\leq N\\}\\) and \\(\\{t_i|1\\leq i\\leq\\) \\(N\\}\\) to represent the image sequences and text sequences, respectively. Besides, we use \\(t_\\mathrm{prompt}\\) as the text prompt, and \\(\\mathcal{F}_{\\theta}(\\cdot)\\) as the model with parameter \\(\\theta\\). The detailed" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.339, + 0.813, + 0.353 + ], + "angle": 0, + "content": "formulations of tasks are outlined as follows:" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.354, + 0.906, + 0.43 + ], + "angle": 0, + "content": "Task 1: Image-to-Text Sequence Generation. This task evaluates the model's capability to generate coherent and contextually appropriate narratives based on sequential images. The objective is to predict the corresponding text sequences from the given image sequences:" + }, + { + "type": "equation", + "bbox": [ + 0.579, + 0.44, + 0.907, + 0.457 + ], + "angle": 0, + "content": "\\[\n\\{t _ {i} | 1 \\leq i \\leq N \\} = \\mathcal {F} _ {\\theta} \\left(t _ {\\text {p r o m p t}}, x _ {1}, \\dots , x _ {N}\\right). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.466, + 0.907, + 0.542 + ], + "angle": 0, + "content": "Task 2: Text-to-Image Sequence Generation. This task assesses the model's proficiency in generating image sequences that accurately represent provided textual descriptions. The goal is to predict the corresponding image sequences from the given text sequences:" + }, + { + "type": "equation", + "bbox": [ + 0.58, + 0.552, + 0.907, + 0.568 + ], + "angle": 0, + "content": "\\[\n\\{x _ {i} \\mid 1 \\leq i \\leq N \\} = \\mathcal {F} _ {\\theta} \\left(t _ {\\text {p r o m p t}}, t _ {1}, \\dots , t _ {N}\\right). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.578, + 0.906, + 0.655 + ], + "angle": 0, + "content": "Task 3: Interleaved Image-Text Content Continuation. It measures the model's ability to seamlessly continue content in an interleaved image-text format. The model needs to predict the outputs of the remaining image-text steps based on the interleaved image-text content in the previous \\(k\\) steps:" + }, + { + "type": "equation", + "bbox": [ + 0.551, + 0.664, + 0.907, + 0.68 + ], + "angle": 0, + "content": "\\[\n\\left\\{t _ {i}, x _ {i} \\mid k < i \\leq N \\right\\} = \\mathcal {F} _ {\\theta} \\left(t _ {\\text {p r o m p t}}, t _ {1}, x _ {1}, \\dots , t _ {k}, x _ {k}\\right). \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.69, + 0.908, + 0.781 + ], + "angle": 0, + "content": "Task 4: Question-based Interleaved Image-Text Generation. This challenging task requires the model to generate interleaved content based on a given question, testing its understanding and generative reasoning abilities across modalities. The model needs to predict the outcomes of all image-text steps based on the given question:" + }, + { + "type": "equation", + "bbox": [ + 0.602, + 0.791, + 0.907, + 0.807 + ], + "angle": 0, + "content": "\\[\n\\left\\{t _ {i}, x _ {i} \\mid 1 \\leq i \\leq N \\right\\} = \\mathcal {F} _ {\\theta} \\left(t _ {\\text {p r o m p t}}\\right). \\tag {5}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.817, + 0.691, + 0.832 + ], + "angle": 0, + "content": "4.2. Evaluation Metric" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.906, + 0.901 + ], + "angle": 0, + "content": "This section lists the evaluation metrics for assessing the quality of generated results. Our evaluation framework1 measures textual accuracy, visual fidelity, and the coherence of multi-modal outputs." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8077" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.484, + 0.273 + ], + "angle": 0, + "content": "Evaluation for Text Sequence Generation. Generating narratives for a series of images shares similarities with the established task of image captioning. Consequently, we directly apply traditional image captioning metrics like ROUGE [21] and METEOR [5] at both the step and document levels. Additionally, we propose Illustration Relevance Score (IRS) to assess the relevance between image illustration and text context in the final document. Specifically, we simulate the human document creation process by first generating detailed image descriptions from the text context and then assessing the consistency between these descriptions and corresponding images using GPT-4o [27]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.274, + 0.484, + 0.424 + ], + "angle": 0, + "content": "Evaluation for Image Sequence Generation. We evaluate the quality of image illustrations on four key assessment aspects including image quality, style consistency between generated images, reconstruction on referenced images, and IRS. To evaluate image quality, we use Fréchet Inception Distance (FID) [13] with ground truth reference and Inception Score (IS) [33]. For assessing style consistency, we employ the Structural Similarity Index Measure (SSIM), and for ground truth reconstruction evaluation, we adopt the Peak Signal-to-Noise Ratio (PSNR)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.426, + 0.484, + 0.577 + ], + "angle": 0, + "content": "Evaluation for Interleaved Image-Text Generation. We use MLLM (GPT-4o [27]) for a comprehensive assessment of linguistic and visual coherence. GPT-4o evaluates four aspects: image sequence coherence, image quality, document completeness, and IRS. For image sequence coherence, we check style, entity consistency, and content trend alignment between image and text sequence. Image quality is evaluated on authenticity, integrity, clarity, and aesthetics. Document completeness is scored based on how thoroughly the content covers the topic." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.593, + 0.224, + 0.61 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.619, + 0.355, + 0.635 + ], + "angle": 0, + "content": "5.1. Datasets Quality Comparison" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.643, + 0.483, + 0.853 + ], + "angle": 0, + "content": "Settings. In this section, we compare the quality of our dataset with previous interleaved image-text datasets (e.g., MMC4 [45] and OBELICS [19]) focusing on text sequences, image sequences, and image-text alignment. We utilize powerful LLMs (e.g., Llama3 [25] and GPT-4o [27]) to assess the text development, text completeness, and image-text alignment of these datasets. Each metric is scored on a scale of 0 to 10, with detailed definitions and prompts provided in the Appendix. For MMC4 and OBELICS, we randomly sample 5,000 documents from their data. We evaluate our entire dataset with Llama3 and randomly sample 5,000 documents for assessment by GPT-4o. To evaluate image sequence quality, we employ the metric \\(\\mathcal{F}\\) as cited in Eq. (1)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Quantitative Results. As shown in Table 1, our dataset surpasses MMC4 and OBELICS across all evaluated dimensions. Notably, the score difference in development," + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.089, + 0.905, + 0.164 + ], + "angle": 0, + "content": "
DatasetsDLP L.3 / G.4oCPL L.3 / G.4oITA L.3 / G.4oImgS
MMC4 [45]5.56 / 4.756.28 / 5.126.53 / 4.660.21
OBELICS [19]5.60 / 5.976.51 / 5.885.00 / 3.811.00
CoMM (Ours)6.93 / 7.647.44 / 7.077.46 / 8.914.27
" + }, + { + "type": "table_caption", + "bbox": [ + 0.512, + 0.174, + 0.908, + 0.23 + ], + "angle": 0, + "content": "Table 1. Quality comparison of interleaved image-text datasets. \"DLP\" stands for Development, \"CPL\" signifies Completeness, \"ITA\" represents Image-Text Alignment, \"ImgS\" refers to Image Sequence, \"L.3\" denotes Llama3 [25], and \"G.4o\" is GPT-4o [27]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.258, + 0.907, + 0.349 + ], + "angle": 0, + "content": "completeness, and image-text alignment is more pronounced when evaluated by GPT-4o, a more powerful MLLM. Furthermore, the image sequence metric underscores the superior quality of our data, reflecting higher scores compared to MMC4 and OBELICS. These results affirm the high quality of our dataset in interleaved image-text data." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.359, + 0.849, + 0.375 + ], + "angle": 0, + "content": "5.2. In-Context Multimodal Understanding" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.382, + 0.909, + 0.579 + ], + "angle": 0, + "content": "Setting. To further demonstrate the high-quality and validate the efficacy of our dataset in promoting in-context understanding, we conduct experiments using few-shot learning settings on several downstream tasks (COCO [22], Flickr30k [43], VQAv2 [11], OKVQA [24], TextVQA [35], VizWiz [12], and Hateful Memes [17]). Few-shot learning is critical for assessing a model's ability of in-context understanding. In our experiments, few-shot scenarios are simulated by providing the model with a small number of example queries and their corresponding answers before it is evaluated on new queries. We consider various shot settings (0, 4, 8, 16, and 32) to understand the model's performance across different levels of supervision." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.58, + 0.909, + 0.821 + ], + "angle": 0, + "content": "Quantitative Results. Table 2 presents the results for the baseline model and its enhancements with MMC4, OBELICS, and our proposed dataset. As observed from Table 2, our method consistently outperforms the performance of MMC4 and OBELICS across all shot settings, demonstrating superior generalization and in-context learning capabilities. Particularly noticeable improvements are seen in the COCO and TextVQA datasets, where our method exceeds the baseline by significant margins in nearly all shot settings. This superior performance underlines the strength of our dataset in enabling models to better learn from context and apply learned knowledge to unseen instances. Furthermore, our dataset surpasses the baseline in all tasks with more shot settings, maintaining its advantage with longer context few-shot data, as shown in the 16-shot and 32-shot scenarios." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.832, + 0.776, + 0.847 + ], + "angle": 0, + "content": "5.3. Interleaved Generation Tasks" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Baselines. This section introduces baselines developed from several state-of-the-art approaches, i.e., MiniGPT-5 [44], SEED-Llama [9], and Emu2 [36], to assess their effective" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8078" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.153, + 0.088, + 0.847, + 0.387 + ], + "angle": 0, + "content": "
ShotCOCOFlickr30kVQAv2OKVQATextVQAVizWizHateful Memes
Baseline79.559.552.737.824.227.551.6
+ MMC4 [45]097.167.045.829.823.415.054.6
+ OBELICS [19]83.551.550.938.925.621.953.8
+ CoMM (Ours)100.369.852.641.333.523.654.8
Baseline89.065.854.840.128.234.154.0
+ MMC4 [45]4103.368.849.536.728.724.455.0
+ OBELICS [19]103.765.253.243.031.429.955.7
+ CoMM (Ours)107.073.253.642.935.730.155.2
Baseline96.362.954.841.129.138.554.7
+ MMC4 [45]8103.667.950.539.028.832.751.3
+ OBELICS [19]101.963.953.744.532.237.251.2
+ CoMM (Ours)108.074.854.244.535.437.952.5
Baseline98.862.854.342.727.342.553.9
+ MMC4 [45]16102.965.550.039.126.436.957.2
+ OBELICS [19]100.858.053.145.330.242.457.8
+ CoMM (Ours)109.371.554.246.935.843.559.9
Baseline99.561.353.342.423.844.053.8
+ MMC4 [45]3294.956.448.335.620.037.749.2
+ OBELICS [19]96.553.152.043.626.844.556.1
+ CoMM (Ours)111.970.754.046.634.945.757.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.397, + 0.908, + 0.455 + ], + "angle": 0, + "content": "Table 2. Performance comparison of different datasets, with the baseline model Open-Flamingo 9B [4]. Evaluations were in an open-ended setting for VQA tasks with random in-context examples. (Task, Metric, Query Split): (COCO [22], CIDEr, test), (Flickr30k [43], CIDEr, test (Karpathy)), (VQAv2 [11], VQA accuracy, testdev), (OKVQA [24], VQA accuracy, val), (TextVQA [35], VQA accuracy, val), (VizWiz [12], VQA accuracy, testdev), (Hateful Memes [17], ROC-AUC, test seen)." + }, + { + "type": "table", + "bbox": [ + 0.109, + 0.467, + 0.891, + 0.64 + ], + "angle": 0, + "content": "
MethodsSizeI2T Sequence Generation (Task1)StyleContinuation Generation (Task3)
M.S.R.S.M.W.R.W.IRSEntityTrendCPL.ImgQIRS
MiniGPT-5 [44]7B4.608.649.1010.677.395.585.215.246.336.362.56
SEED-Llama [9]8B2.296.344.207.336.996.285.845.726.286.552.92
SEED-Llama [9]14B3.168.125.718.178.136.686.226.136.666.673.23
Emu2 [36]33B3.658.417.438.625.018.227.997.978.498.622.42
MethodsSizeT2I Sequence Generation (Task2)StyleQuestion-based Generation (Task4)
FID ↓ISSSIMPSNRIRSEntityTrendCPL.ImgQIRS
MiniGPT-5 [44]7B56.517.6020.667.575.485.655.205.255.816.152.71
SEED-Llama [9]8B57.968.4020.537.875.277.556.816.155.136.361.46
SEED-Llama [9]14B66.2311.0620.838.126.247.516.616.306.136.662.50
Emu2 [36]33B60.456.3420.967.914.178.417.567.637.547.592.02
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.649, + 0.908, + 0.707 + ], + "angle": 0, + "content": "Table 3. Comparison of performance among MiniGPT-5 [44], SEED-Llama [9], and Emu2 [37] on the four generation tasks. \"M.S.\" and \"R.S.\" represent METEOR [5] and ROUGE_L [21] for step-by-step evaluation, while \"M.W.\" and \"R.W.\" correspond to METEOR and ROUGE_L for whole document evaluation. \"CPL.\" stands for Completeness, \"ImgQ\" indicates Image Quality, \"IRS\" means Illustration Relevance Score, and \"↓\" denotes that lower values are better." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.733, + 0.483, + 0.824 + ], + "angle": 0, + "content": "ness in generating interleaved image-text content. Details of these models are thoroughly provided in the Appendix. To construct the baselines, we fine-tune each of these models using our CoMM dataset. We employ a rigorous evaluation, assessing them on benchmark tasks involving interleaved image-text generation and understanding." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.486, + 0.903 + ], + "angle": 0, + "content": "Quantitative Results. As shown in Table 3, the quantitative results clearly illustrate the varying performance of MiniGPT-5, SEED-Llama, and Emu2 across the four interleaved generation tasks. For I2T Sequence Generation (Task1), MiniGPT-5 achieves the highest scores in ME" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.733, + 0.909, + 0.901 + ], + "angle": 0, + "content": "TEOR, ROUGE_L, but falls short in Illustration Relevance Score (IRS) where SEED-Llama (14B) leads with a score of 8.13. For T2I Sequence Generation (Task2), MiniGPT-5 excels with the lowest FID score of 56.51, while SEED-Llama (14B) surpasses IS, SSIM, PSNR, and IRS metrics, suggesting stronger visual fidelity and semantic similarity. Emu2 tops in Continuation Generation (Task3) with the highest scores in Style, Entity, Trend, Completeness, and Image Quality. In Question-based Generation (Task4), Emu2 dominates in Style, Entity, Trend, Completeness, and Image Quality, while MiniGPT-5 manages to secure the lead in IRS." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8079" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.092, + 0.895, + 0.495 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.496, + 0.908, + 0.525 + ], + "angle": 0, + "content": "Figure 3. Visualization of interleaved image-text content generation from SEED-Llama [9] (Top) and MiniGPT-5 [44] (Bottom) separately trained by MMC4 [45] and CoMM [Ours]." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.539, + 0.483, + 0.599 + ], + "angle": 0, + "content": "Overall, these results highlight the strengths and weaknesses of each model in handling different aspects of interleaved image-text generation tasks, providing valuable insights for future research and model improvements." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.599, + 0.483, + 0.734 + ], + "angle": 0, + "content": "Preference Dataset Training Results. With SEED-Llama [9] unifying image to discrete token prediction modeling, we can directly apply the mainstream preference learning algorithm DPO [30] to train SEED-Llama on our preference dataset following supervised fine-tuning on our dataset. As shown in Table 4, performance further improves on the image-to-text sequence generation task after training with our preference dataset. Additional results on preference dataset training are available in the Appendix." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.735, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Multimodal Generation Qualitative Analysis. As illustrated in Figure 1 and Figure 3, the interleaved image-text content generated by models [9, 36, 44] trained on MMC4 [45] shows a noticeable difference when compared to the content generated by models trained on our datasets. Our models produce content that is more consistent with the previous input context (e.g., textual coherence and entity consistency). By ensuring higher coherence, consistency, and alignment, CoMM sets a new standard for training datasets in the realm of multimodal large language models. More visualizations are available in the Appendix." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.535, + 0.9, + 0.624 + ], + "angle": 0, + "content": "
MethodsSizeI2T Sequence Generation (Task1)
M.S.R.S.M.W.R.W.IRS
SEED-Llama [9]8B2.296.344.207.336.99
+ DPO [30]8B4.098.836.9510.247.26
SEED-Llama [9]14B3.168.125.718.178.13
+ DPO [30]14B4.6210.158.0810.308.18
" + }, + { + "type": "table_caption", + "bbox": [ + 0.512, + 0.628, + 0.908, + 0.698 + ], + "angle": 0, + "content": "Table 4. Performance results of SEED-Llama [9] trained by DPO [30] in our preference dataset. \"M.S.\" and \"R.S.\" represent METEOR [5] and ROUGE_L [21] for step-by-step evaluation, while \"M.W.\" and \"R.W.\" correspond to METEOR and ROUGE_L for whole document evaluation." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.709, + 0.634, + 0.725 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.735, + 0.909, + 0.903 + ], + "angle": 0, + "content": "In this paper, we present CoMM, a high-quality interleaved image-text dataset designed to address the limitations of existing multimodal large language models (MLLMs) by ensuring superior text-image alignment and entity consistency. Through diverse interleaved image-text content, CoMM significantly enhances MLLMs' in-context learning capabilities, as demonstrated by extensive few-shot experiments. Additionally, we introduce four novel tasks and a comprehensive evaluation framework to assess interleaved image-text generation abilities, setting new benchmarks for advancing multimodal understanding and generation in MLLMs." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8080" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.241, + 0.108 + ], + "angle": 0, + "content": "Acknowledgment" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.484, + 0.22 + ], + "angle": 0, + "content": "This work was supported by the Hong Kong SAR RGC Early Career Scheme (26208924), the National Natural Science Foundation of China Young Scholar Fund (62402408), and the HKUST Sports Science and Technology Research Grant (SSTRG24EG04), ACCESS - AI Chip Center for Emerging Smart Systems, sponsored by InnoHK funding, Hong Kong SAR." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.233, + 0.187, + 0.249 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.258, + 0.484, + 0.314 + ], + "angle": 0, + "content": "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florecia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.316, + 0.484, + 0.383 + ], + "angle": 0, + "content": "[2] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. NeurIPS, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.385, + 0.485, + 0.44 + ], + "angle": 0, + "content": "[3] Jie An, Zhengyuan Yang, Linjie Li, Jianfeng Wang, Kevin Lin, Zicheng Liu, Lijuan Wang, and Jiebo Luo. Openleaf: Open-domain interleaved image-text generation and evaluation. arXiv preprint arXiv:2310.07749, 2023. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.442, + 0.484, + 0.522 + ], + "angle": 0, + "content": "[4] Anas Awadalla, Irena Gao, Josh Gardner, Jack Hessel, Yusuf Hanafy, Wanrong Zhu, Kalyani Marathe, Yonatan Bitton, Samir Gadre, Shiori Sagawa, et al. Openflamingo: An opensource framework for training large autoregressive vision-language models. arXiv preprint arXiv:2308.01390, 2023. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.525, + 0.484, + 0.579 + ], + "angle": 0, + "content": "[5] Satanjeev Banerjee and Alon Lavie. Meteor: An automatic metric for mt evaluation with improved correlation with human judgments. In ACL Workshop, pages 65-72, 2005. 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.581, + 0.484, + 0.648 + ], + "angle": 0, + "content": "[6] Runpei Dong, Chunrui Han, Yuang Peng, Zekun Qi, Zheng Ge, Jinrong Yang, Liang Zhao, Jianjian Sun, Hongyu Zhou, Haoran Wei, et al. Dreamllm: Synergistic multimodal comprehension and creation. arXiv preprint arXiv:2309.11499, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.65, + 0.484, + 0.718 + ], + "angle": 0, + "content": "[7] Samir Yitzhak Gadre, Gabriel Ilharco, Alex Fang, Jonathan Hayase, Georgios Smyrnis, Thao Nguyen, Ryan Marten, Mitchell Wortsman, Dhruba Ghosh, Jieyu Zhang, et al. Datacomp: In search of the next generation of multimodal datasets. NeurIPS, 36, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.72, + 0.484, + 0.788 + ], + "angle": 0, + "content": "[8] Bofei Gao, Feifan Song, Yibo Miao, Zefan Cai, Zhe Yang, Liang Chen, Helan Hu, Runxin Xu, Qingxiu Dong, Ce Zheng, et al. Towards a unified view of preference learning for large language models: A survey. arXiv preprint arXiv:2409.02795, 2024. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.79, + 0.484, + 0.844 + ], + "angle": 0, + "content": "[9] Yuying Ge, Sijie Zhao, Ziyun Zeng, Yixiao Ge, Chen Li, Xintao Wang, and Ying Shan. Making llama see and draw with seed tokenizer. arXiv preprint arXiv:2310.01218, 2023. 6, 7, 8, 1, 3, 4, 9, 10, 11, 12, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[10] Yuying Ge, Sijie Zhao, Jinguo Zhu, Yixiao Ge, Kun Yi, Lin Song, Chen Li, Xiaohan Ding, and Ying Shan. Seed-x: Multimodal models with unified multi-granularity comprehension and generation. arXiv preprint arXiv:2404.14396, 2024. 1" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.258, + 0.485, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.147 + ], + "angle": 0, + "content": "[11] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Bartra, and Devi Parikh. Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In CVPR, 2017. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.15, + 0.907, + 0.205 + ], + "angle": 0, + "content": "[12] Danna Gurari, Qing Li, Abigale J Stangl, Anhong Guo, Chi Lin, Kristen Grauman, Jiebo Luo, and Jeffrey P Bigham. Vizwiz grand challenge: Answering visual questions from blind people. In CVPR, 2018. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.207, + 0.908, + 0.261 + ], + "angle": 0, + "content": "[13] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. NeurIPS, 2017. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.263, + 0.907, + 0.317 + ], + "angle": 0, + "content": "[14] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.32, + 0.906, + 0.375 + ], + "angle": 0, + "content": "[15] S Huang, L Dong, W Wang, Y Hao, S Singhal, S Ma, T Lv, L Cui, OK Mohammed, Q Liu, et al. Language is not all you need: aligning perception with language models. 2023. arXiv preprint arXiv:2302.14045. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.377, + 0.907, + 0.418 + ], + "angle": 0, + "content": "[16] Timo Kaufmann, Paul Weng, Viktor Bengs, and Eyke Hüllermieier. A survey of reinforcement learning from human feedback. arXiv preprint arXiv:2312.14925, 2023. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.42, + 0.906, + 0.475 + ], + "angle": 0, + "content": "[17] Douwe Kiela, Hamed Firooz, Aravind Mohan, Vedanuj Goswami, Amanpreet Singh, Pratik Ringshia, and Davide Testuggine. The hateful memes challenge: Detecting hate speech in multimodal memes. NeurIPS, 2020. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.477, + 0.907, + 0.516 + ], + "angle": 0, + "content": "[18] Jing Yu Koh, Daniel Fried, and Russ R Salakhutdinov. Generating images with multimodal language models. NeurIPS, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.519, + 0.907, + 0.587 + ], + "angle": 0, + "content": "[19] Hugo Laurençon, Lucile Saulnier, Léo Tronchon, Stas Bekman, Amanpreet Singh, Anton Lozhkov, Thomas Wang, Siddharth Karamcheti, Alexander Rush, Douwe Kiela, et al. Obelics: An open web-scale filtered dataset of interleaved image-text documents. NeurIPS, 2023. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.59, + 0.908, + 0.657 + ], + "angle": 0, + "content": "[20] Samuel Lavoie, Polina Kirichenko, Mark Ibrahim, Mahmoud Assran, Andrew Gordon Wildon, Aaron Courville, and Nicolas Ballas. Modeling caption diversity in contrastive vision-language pretraining. arXiv preprint arXiv:2405.00740, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.661, + 0.907, + 0.701 + ], + "angle": 0, + "content": "[21] Chin-Yew Lin. Rouge: A package for automatic evaluation of summaries. In Text summarization branches out, 2004. 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.704, + 0.907, + 0.757 + ], + "angle": 0, + "content": "[22] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dolkar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.76, + 0.906, + 0.814 + ], + "angle": 0, + "content": "[23] Chang Liu, Haoning Wu, Yujie Zhong, Xiaoyun Zhang, and Weidi Xie. Intelligent grimm-open-ended visual storytelling via latent diffusion models. arXiv preprint arXiv:2306.00973, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.817, + 0.907, + 0.87 + ], + "angle": 0, + "content": "[24] Kenneth Marino, Mohammad Rastegari, Ali Farhadi, and Roozbeh Mottaghi. Ok-vqa: A visual question answering benchmark requiring external knowledge. In CVPR, 2019. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.907, + 0.901 + ], + "angle": 0, + "content": "[25] Meta. https://llama.meta.com/llama3, 2024.2, 4,6,1" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "8081" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.486, + 0.133 + ], + "angle": 0, + "content": "[26] Microsoft. Bingchat. https://www.microsoft.com/en-us/edge/features/bing-chat, 2023. Accessed: 2023-10-09. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.134, + 0.485, + 0.162 + ], + "angle": 0, + "content": "[27] OpenAI. https://openai.com/index/hello-gpt-4o/, 2024.3,4,6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.163, + 0.484, + 0.218 + ], + "angle": 0, + "content": "[28] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.219, + 0.484, + 0.287 + ], + "angle": 0, + "content": "[29] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.288, + 0.484, + 0.343 + ], + "angle": 0, + "content": "[30] Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. Direct preference optimization: Your language model is secretly a reward model. NeurIPS, 36, 2023. 8, 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.344, + 0.484, + 0.399 + ], + "angle": 0, + "content": "[31] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. JMLR, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.4, + 0.484, + 0.44 + ], + "angle": 0, + "content": "[32] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.442, + 0.484, + 0.483 + ], + "angle": 0, + "content": "[33] Tim Salimans, Ian Goodfellow, Wojciech Zaremba, Vicki Cheung, Alec Radford, and Xi Chen. Improved techniques for training gans. NeurIPS, 2016. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.484, + 0.484, + 0.564 + ], + "angle": 0, + "content": "[34] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. NeurIPS, 35:25278–25294, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.567, + 0.484, + 0.608 + ], + "angle": 0, + "content": "[35] Amanpreet Singh, Vivek Natarajan, Meet Shah, Yu Jiang, Xinlei Chen, Dhruv Batra, Devi Parikh, and Marcus Rohrbach. Towards vqa models that can read. In CVPR, 2019. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.609, + 0.484, + 0.677 + ], + "angle": 0, + "content": "[36] Quan Sun, Yufeng Cui, Xiaosong Zhang, Fan Zhang, Qiying Yu, Zhengxiong Luo, Yueze Wang, Yongming Rao, Jingjing Liu, Tiejun Huang, et al. Generative multimodal models are in-context learners. arXiv preprint arXiv:2312.13286, 2023. 1, 2, 6, 7, 8, 4, 9, 10, 11, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.679, + 0.484, + 0.734 + ], + "angle": 0, + "content": "[37] Quan Sun, Qiying Yu, Yufeng Cui, Fan Zhang, Xiaosong Zhang, Yueze Wang, Hongcheng Gao, Jingjing Liu, Tiejun Huang, and Xinlong Wang. Emu: Generative pretraining in multimodality. In ICLR, 2023. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.735, + 0.484, + 0.803 + ], + "angle": 0, + "content": "[38] Changyao Tian, Xizhou Zhu, Yuwen Xiong, Weiyun Wang, Zhe Chen, Wenhai Wang, Yuntao Chen, Lewei Lu, Tong Lu, Jie Zhou, et al. Mm-interleaved: Interleaved image-text generative modeling via multi-modal feature synchronizer. arXiv preprint arXiv:2401.10208, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.804, + 0.484, + 0.872 + ], + "angle": 0, + "content": "[39] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.484, + 0.903 + ], + "angle": 0, + "content": "[40] Weihan Wang, Qingsong Lv, Wenmeng Yu, Wenyi Hong, Ji Qi, Yan Wang, Junhui Ji, Zhuoyi Yang, Lei Zhao, Xixuan" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.486, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.907, + 0.134 + ], + "angle": 0, + "content": "Song, Jiazheng Xu, Bin Xu, Juanzi Li, Yuxiao Dong, Ming Ding, and Jie Tang. CogvIm: Visual expert for pretrained language models, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.908, + 0.19 + ], + "angle": 0, + "content": "[41] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. NeurIPS, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.192, + 0.852, + 0.206 + ], + "angle": 0, + "content": "[42] WikiHow. https://www.wikipiow.com. 2,3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.208, + 0.908, + 0.261 + ], + "angle": 0, + "content": "[43] Peter Young, Alice Lai, Micah Hodosh, and Julia Hockenmaier. From image descriptions to visual denotations: New similarity metrics for semantic inference over event descriptions. TACL, 2014. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.264, + 0.907, + 0.318 + ], + "angle": 0, + "content": "[44] Kaizhi Zheng, Xuehai He, and Xin Eric Wang. Minigpt-5: Interleaved vision-and-language generation via generative vowels. arXiv preprint arXiv:2310.02239, 2023. 1, 6, 7, 8, 4, 9, 10, 11, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.321, + 0.907, + 0.389 + ], + "angle": 0, + "content": "[45] Wanrong Zhu, Jack Hessel, Anas Awadalla, Samir Yitzhak Gadre, Jesse Dodge, Alex Fang, Youngjae Yu, Ludwig Schmidt, William Yang Wang, and Yejin Choi. Multimodal c4: An open, billion-scale corpus of images interleaved with text. NeurIPS, 2023. 1, 2, 3, 4, 6, 7, 8" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.389 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8082" + } + ] +] \ No newline at end of file diff --git a/2025/CoMM_ A Coherent Interleaved Image-Text Dataset for Multimodal Understanding and Generation/4d3dc3b4-7f37-464f-b503-0ded8f6008de_origin.pdf b/2025/CoMM_ A Coherent Interleaved Image-Text Dataset for Multimodal Understanding and Generation/4d3dc3b4-7f37-464f-b503-0ded8f6008de_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..ff56ff790f39535185cf209f2ef5b425c8510db9 --- /dev/null +++ b/2025/CoMM_ A Coherent Interleaved Image-Text Dataset for Multimodal Understanding and Generation/4d3dc3b4-7f37-464f-b503-0ded8f6008de_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d2c68b18b40095e6d02b296a696b79c14f575fdc9a20eb6cf19ab600eb82093d +size 2394568 diff --git a/2025/CoMM_ A Coherent Interleaved Image-Text Dataset for Multimodal Understanding and Generation/full.md b/2025/CoMM_ A Coherent Interleaved Image-Text Dataset for Multimodal Understanding and Generation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..feae9d87179c601aa026de513d323cdc6652d8c8 --- /dev/null +++ b/2025/CoMM_ A Coherent Interleaved Image-Text Dataset for Multimodal Understanding and Generation/full.md @@ -0,0 +1,292 @@ +# CoMM: A Coherent Interleaved Image-Text Dataset for Multimodal Understanding and Generation + +Wei Chen $^{1,2*}$ , Lin Li $^{1,3*}$ , Yongqi Yang $^{2*}$ , Bin Wen $^{4}$ , Fan Yang $^{4}$ , Tingting Gao $^{4}$ , Yu Wu $^{2\dagger}$ , Long Chen $^{1\dagger}$ $^{1}$ The Hong Kong University of Science and Technology, $^{2}$ Wuhan University, $^{3}$ AI Chip Center for Emerging Smart Systems, $^{4}$ Kuaishou Technology + +{wchendb, lllidy, longchen}@ust.hk, {yongqiyang, wuyucs}@whu.edu.cn + +https://github.com/HKUST-LongGroup/CoMM + +# Abstract + +Interleaved image-text generation has emerged as a vital multimodal task aimed at creating sequences of interleaved visual and textual content given a query. Despite notable advancements in recent multimodal large language models (MLLMs), generating integrated image-text sequences that exhibit narrative coherence and entity and style consistency remains challenging due to poor training data quality. To this end, we introduce CoMM, a high-quality Coherent interleaved image-text MultiModal dataset designed to enhance the coherence, consistency, and alignment of generated multimodal content. Initially, CoMM harnesses raw data from diverse sources, focusing on instructional content and visual storytelling, establishing a foundation for coherent and consistent content. To further refine the data quality, we devise a multi-perspective filter strategy that leverages advanced pre-trained models to ensure the development of sentences, consistency of inserted images, and semantic alignment between them. Various quality evaluation metrics are designed to prove the high quality of the filtered dataset. Meanwhile, extensive few-shot experiments on various downstream tasks demonstrate CoMM's effectiveness in significantly enhancing the in-context learning capabilities of MLLMs. Moreover, we propose four new tasks to evaluate MLLMs' interleaved generation abilities, supported by a comprehensive evaluation framework. We believe CoMM opens a new avenue for advanced MLLMs with superior multimodal in-context learning and understanding ability. + +# 1. Introduction + +Interleaved image-text generation has begeoned as an emerging multimodal task, striving to imitate the human- + +like capability to alternately create visual and textual content [3, 45]. Specifically, it aims to generate a sequence of interleaved text descriptions and illustrative images given a query [3]. With its interleaved multimodal generation ability, state-of-the-art models can facilitate a variety of applications, e.g., multimodal instruction generation, tutorial step generation, and visual storytelling. + +Recent advancements in multimodal large language models (MLLMs) have exhibited exceptional capabilities in cross-modal generation, but they still struggle to generate interleaved image-text sequences coherently [36, 44]. The main reason is that most of them are trained on single imagetext pairs, limited in generating coherent and contextually integrated multimodal content [18, 44]. To this end, some efforts integrate multiple models (e.g., GPT-4 [1] and Stable Diffusion XL [28]) to generate open-domain interleaved image-text contents with arbitrary formats in a training-free manner [3]. However, their results are often dominated by the generated text descriptions, limited by the generation model's understanding of these textual inputs. In contrast, other research efforts [6, 10] have attempted to train models with seamless multimodal understanding and generation capabilities using existing interleaved image-text datasets (e.g., MMC4 [45]). + +Albeit the unprecedented progress, it is becoming increasingly evident that data-centric methods still suffer several limitations: 1) Narrative Coherence: The contextual relevance and coherence of generated image-text sentences are relatively low. In Figure 1(a), the model pre-trained on the MMC4 [45] directly draws the final cooked dessert, missing the intermediate coherent steps of generation. These coherent generation steps not only enhance interpretability but also boost the reasoning capabilities of the model [41]. 2) Entity and Style Consistency: The entity and subject styles of the generated illustrated images are inconsistent. As illustrated in Figure 1(b), the style of the blue-gray sofa image + +![](images/422e031e86cb9a9247ec9858f00403663e8850d2300d4b483d9e933327b7febe.jpg) + +# (c) Sample from MMC4 + +When it comes to making life just a little bit smoother and safer in a wonderful place such as White Bear Lake, ADT Monitored Home Security can be your main solution for keeping your family and belongings safe and secure. + +An ADT home monitoring system for your home can change your life for the better, letting you explore and enjoy all that you love in White Bear Lake. ADT home monitoring is more than just home security, it raises the bar to better living. + +It's a chance to live your daily life without all the worry and stress. For over a century, ADT monitoring systems have paved the way in home security. + +As the #1 home security provider in the country, we know all the vital facts and components that are necessary in keeping our $6+$ million home and business customers safe. It's why so many people in White Bear Lake and the rest of the nation count on us for their safety. + +![](images/3bc55a838fd7d914e38e61a57768a6e3631f8bf27f0b3c4d338e24b8cb03591f.jpg) + +![](images/fc53a4dd0cfeee1bc185222928bc5b136f1989ed264891ebd792b00625a91a7a.jpg) + +![](images/600f8c0011ccc3ddf62fd093b815e5d4b11e495b3331c2b1cad99d7be28cc04e.jpg) + +generated by the model pre-trained on MMC4 is inconsistent with that in the query. This consistency of entities and style has been proposed as an important metric for evaluating the results of interleaved image-text content generation [3]. These limitations primarily stem from the low quality of training data caused by the poor alignment between image and text content, i.e., the devil is in the data. For example in Figure 1(c), the second image in the training sample has a noticeable gap with the context. Additionally, the styles of all the images are almost completely different. Thus, it raises a natural research question: How to construct a high-quality interleaved image-text dataset? + +In this paper, we introduce CoMM, a high-quality and coherent interleaved image-text dataset. By "high-quality", we mean that the interleaved image-text content should exhibit developed and logically structured text steps, consistent entities and visual styles in images, and strong semantic and contextual alignment between them. To enhance text coher + +# (d) Sample from CoMM + +Step1. Remove the tuna steaks from the refrigerator. Allow them to sit out at room temperature for 1/2 to 1 hour before you begin cooking. + +![](images/f96c0ed39e4909f58e8040b510f9c1b5d3e385ff16c3e687d32a48c5d3448eab.jpg) + +Step2. Use a paper towel to dry excess moisture from the tuna steaks. Simply pat the paper towel on both sides of the steak to remove marinade or condensation from the meat. + +![](images/eb83bd16c70cc7505f21c4efb375e0ba6b4874dae1a77599aa1043a739777c3a.jpg) + +Step3. Add 2 tbsp. olive oil or butter to the pan over medium-high heat. Tilt the pan to spread the oil or butter across the surface and coat the pan. + +Step4. Place the tuna steaks on the hot pan. Do not crowd the tuna in the pan. Keep the steaks at least 3 to 4 inches apart to make sure they cook evenly. + +Step5. Sear each side for two minutes. Do not move the meat around too much while it sears. Gently shake the pan after two minutes, and if the tuna steak releases from the surface, you can flip it on its other side. It is important to make sure that the edged are seared, but you want the inside of the tuna steak to remain rare. + +Step6. Make a small cut in the center of the tuna steak with a knife. Inspect the inside to make sure that the middle of the steak stays rare and red in color. + +![](images/6232e448cfc5218dc8fabbed84e8e6f84eec0fea452ff61ff6cbe9bc8e2d8bc5.jpg) + +![](images/5cbb59f3a24307ac392a50f844c7b9e0b3ac386f61a12df697944ad3bd8e38d9.jpg) +Figure 1. Illustration of interleaved image-text content generation results and dataset quality. (a) Given the query from CoMM, the interleaved image-text content generation results from the model Emu2 [36] separately trained by MMC4 [45] and CoMM (Ours). (b) The query is from the MMC4. (c) A training sample is from the MMC4. (d) A training sample is from the our CoMM. + +ence and image consistency, we collect raw data from specific websites, e.g., WikiHow [42], known for their instructional content or visual stories. As illustrated in Figure 1(d), instructional content, due to its uniform intention (e.g., cooking steak) and structured presentation (e.g., step 1), exhibits strong text coherence and image consistency. Additionally, we devise a multi-perspective filter strategy consisting of: a text sequence filter, an image sequence filter, and an imagedtext alignment filter. The first two leverage large language models (LLMs), e.g., Llama3 [25], and vision-language models (VLMs), e.g., CLIP [29], to eliminate incoherent text and images, respectively, further enhancing overall coherence. The last one employs both VLMs and MLLMs to comprehensively assess and refine the dataset for better alignment between visual and textual elements. After constructing this dataset, we further enhance it into a preference dataset by generating negative samples through shuffled text and images. This preference dataset can then be used during the + +reinforcement learning stage of post-training [8, 16]. + +To evaluate the data quality and make a comprehensive comparison of existing interleaved image-text benchmarks [19, 45], we design four metrics to evaluate the development, completeness, image-text alignment, and consistency of image-text sequences, separately (cf. Table 1). CoMM significantly outperforms previous datasets [19, 45] on all metrics, demonstrating its high quality. To further demonstrate the superiority of CoMM, we conduct few-shot experiments on various downstream tasks (e.g., VQA) [11, 12, 17, 22, 24], comparing with MMC4 [45] and OBELICS [19]. The model trained on our CoMM consistently outperforms others, particularly in long context settings (16 & 32 shots), showcasing CoMM's ability to enhance the in-context learning capacities of MLLMs. + +Furthermore, thanks to our dataset, we introduce four new challenging tasks to evaluate the multimodal understanding and generation abilities of MLLMs: image-to-text sequence generation, text-to-image sequence generation, interleaved image-text content continuation, and question-based interleaved image-text generation. To assess the performance of these newly introduced tasks, we implement a comprehensive evaluation framework considering metrics such as METEOR [5], ROUGE [21] for text quality, and FID [13] and IS [33] for image quality. To accurately evaluate open-ended interleaved image-text generation content, we develop a series of metrics and leverage powerful MLLM (e.g., GPT4o [27]) for assessment. These tasks and evaluations pave the way for novel approaches in evaluating MLLMs by addressing critical gaps in current benchmarks. Conclusively, our contributions are as follows: + +1) We introduce CoMM, a high-quality coherent interleaved image-text dataset designed to address the limitations of existing datasets by ensuring coherent narrative, consistent entity and style, and strong semantic alignment between images and text. +2) CoMM introduces four novel benchmark tasks and associated evaluation metrics to comprehensively evaluate multimodal understanding and generation capabilities of MLLMs. +3) CoMM can serve as both a supervised fine-tuning dataset and a preference dataset. Extensive ablations across various downstream tasks demonstrate the high-quality of CoMM and its effectiveness in enhancing the understanding and generation capabilities of MLLMs. + +# 2. Related Work + +Interleaved Image-Text Web Document Datasets. Training on interleaved image-text web documents has demonstrated superior performance over image-description pairs, as evidenced by studies such as Flamingo [2] and KOSMOS-1 [15]. These findings underscore the significant advantages of utilizing the richer and more meaningful correlations in + +herent in interleaved image-text documents[19]. However, the training data in these two studies is not made publicly available. + +The scarcity of interleaved image-text data has been addressed by the introduction of two datasets: MMC4 [45] and OBELICS [19]. MMC4 builds upon the text-only C4 corpus [31] by integrating images into text passages using CLIP features. In contrast, OBELICS emphasizes comprehensive filtering strategies and preserves the original structure of web pages. These datasets present a diverse and extensive interleaved image-text corpus for multimodal language models. Despite their depth and breadth, they exhibit notable deficiencies in document quality, such as weaker completeness and image-text coherence (cf., Table 1). Additionally, as depicted in Figure 2, their documents lack comprehensiveness in the image modality, with most documents containing only one or two images. To address this, we propose CoMM that focuses explicitly on coherence-rich scenarios with more image illustrations. CoMM greatly enhances the baseline's [4] performance (in Table 2), highlighting the critical role of a high-quality coherent interleaved image-text training corpus. + +Modeling of Interleaved Image-Text Data. The impressive performance of language modeling by LLM motivates researchers to delve deeper into the comprehension and generation capabilities of multimodal data. Flamingo [2] introduces image tokens into the language modeling process, facilitating interleaved image-text input and enabling few-shot transfer to tasks (e.g., VQA). Emu [37] leverages Stable Diffusion [32] as the image decoder, enabling the generation of image or text from interleaved image-text input. DreamLLM [6] pioneers the generation of free-form interleaved image-text outputs by modeling text and image within a unified multimodal space. + +Recent research has increasingly focused on the intrinsic requirement of interleaved image-text modeling: capturing multimodal coherence while generating interleaved image-text content. OpenLEAF [3] underscores the importance of maintaining semantic and stylistic consistency between generated images. It also proposes a training-free baseline and employs BingChat [26] to evaluate the two types of consistency. MM-Interleaved [38] introduces the multimodal feature synchronizer to extract context-sensitive, multi-scale image features, thereby more effectively capturing multimodal context coherence and visual consistency. Nonetheless, focusing solely on the model level to meet the intrinsic multimodal coherence requirement is insufficient. Therefore, our high-quality data that strictly adhere to this requirement are essential to fully realize the potential of MLLMs. + +# 3. CoMM Dataset + +The CoMM dataset is collected from high-quality interleaved image-text content sourced from various websites, focusing + +on instructional steps and visual stories. Besides, we apply multi-perspective quality filter strategies on text sequences, image sequences, and image-text alignment, leveraging advanced models, e.g., CLIP [29] and LLMs [25], to enhance the dataset's coherence and relevance1. + +# 3.1. Dataset Construction + +Collection Process. As mentioned above, we gather specific interleaved image-text data to preliminarily ensure the coherence and consistency of our dataset. To achieve this, we explore websites that are likely to provide high-quality interleaved content, such as instructional steps and visual stories. We collect raw data from these sources1 and download the corresponding text and images. Initially, we apply basic data cleaning techniques, including de-duplication and filtering out images with low resolution, to refine the collected data. + +Discarding NSFW images. To address potential ethical implications, we follow the [45] by utilizing an advanced NSFW binary image classifier [7]. This classifier is devised as a 4-layer multi-layer perceptron (MLP) that has been fully trained on the NSFW dataset from LAION-2B [34]. Specifically, it utilizes features extracted from OpenAI's CLIP ViT-L/14 model[29] as input, achieving an impressive accuracy of $97.4\%$ on the NSFW test dataset. For each image processed, we employ the classifier to predict the NSFW probability. Images yielding a probability greater than 0.1 are automatically excluded from further consideration. + +Qualities Filter Strategy. Due to the inherent noises in the gathered raw data, the data still contains inconsistent and irrelevant content. Thus, we further employ a multiperspective filter strategy to ensure the quality and coherence of the dataset. Specifically, it involves three following components1: + +1) Text Sequence Filter: For text sequences, the most important aspect is the contextual development and cohesive connection between them. To maintain a smooth and logical flow of the text, we employ LLMs (e.g., Llama3 [25]) to evaluate the development and coherence among text steps in a document (prompt details are in the Appendix). According to evaluation scores, we eliminate sequences that lack coherence and relevance in their content, producing a higher-quality text corpus. +2) Image Sequence Filter. As for image sequences, visual consistency and relevance are key factors in maintaining the quality of the content. To assess the coherence and development of image sequences, we devise a metric $\mathcal{F}(\cdot)$ + +through the CLIP's visual encoder [29] in Eq. (1): + +$$ +\begin{array}{l} \mathcal {F} \left(\left\{x _ {i} \mid 1 \leq i \leq N \right\}\right) = \frac {1}{N - 1} \sum_ {i = 2} ^ {N} \operatorname {S i m} \left(x _ {i}, x _ {i - 1}\right) \\ - \frac {2}{(N - 1) (N - 2)} \sum_ {i = 2} ^ {N} \sum_ {j = 1} ^ {i - 1} \operatorname {S i m} \left(x _ {i}, x _ {j}\right), \tag {1} \\ \end{array} +$$ + +where $\{x_{i}|1\leq i\leq N\}$ denotes the sequence of images from 1 to $N$ , $\mathrm{Sim}(\cdot ,\cdot)$ denotes the cosine similarity score between two images calculated by the CLIP's visual encoder. The first term denotes the average similarity between consecutive images, promoting smooth transitions and visual coherence within the sequence, while the second term represents the average similarity between all pairs of images, ensuring overall diversity and development. In this way, we can effectively balance coherence and diversity, ensuring that the image sequence is both visually consistent and contextually developed. By applying the image sequence filter, we can identify and eliminate sequences that lack connection or development, resulting in more polished and visually appealing images. + +3) Image-Text Alignment Filter. This filter aims to ensure that the images and text are not only relevant to each other but also maintain a coherent narrative throughout the document. We first utilize CLIP [29] to calculate the CLIP similarity scores between images and text and remove those whose scores are less than 0.1. Since CLIP is trained on individual image-text pairs, where the text often serves as the image's caption, it tends to favor concrete descriptions (i.e., object descriptions and attributes) [20]. However, in an interleaved image-text document, the text for each step describes progress with associated images (e.g., in Figure 1(d), each text focuses on a cooking action linked to previous steps) which is associated with previous steps. Therefore, we also employ advanced model (e.g., GPT-4o [27] or Llama3 [25]) to assess and filter image-text pairs, as these models can better judge current image-text alignment by considering the previous context. + +Preference Dataset Construction. Preference learning [8] has recently gained popularity in the post-training of LLMs, leveraging reinforcement learning [16] to train models with both positive and negative samples. In our dataset, which consists of interleaved image and text sequences, negative samples are easily created by shuffling the order of steps. These shuffles can be categorized into four types: Shuffled Text: Text in different steps is shuffled, while images remain in their original order. Shuffled Image: Images are shuffled, while the text remains in its original order. Both Shuffled Text and Image: Both text and images are shuffled independently. Shuffled Steps: Images and text within each step remain aligned, but the sequence of steps is randomized. + +These negative samples are combined with original positive samples to form a preference dataset, used for further + +![](images/f2d152959b3929e02c5645b94f88c1dd435451c980792871ca6f56397cc30311.jpg) +(a) MMC4 + +![](images/f8394213a912db355d1a5d65a47cfb9f41b2a9f47d950b9343cbe1ca9bece90c.jpg) +(b) OBELICS +Figure 2. Visualization of the image-sentence numbers per document distribution of three datasets. The $\mu$ and $M$ denote the mean/median number of images/sentences in documents, respectively. + +![](images/914d0e0c5f497a4ea40ee10a12bb6eee8b00f2b865a01df43907138fcd9cb5b3.jpg) +(c) CoMM (Ours) + +model training following supervised fine-tuning. + +# 3.2. Dataset Statistic + +After filtering, our dataset contains 227K documents with 2.28M images. Further, Figure 2 illustrates the distribution of image-sentence number per document across MMC4, OBELICS, and CoMM datasets. Darker colors indicate a higher number of documents. The $\mu$ and $M$ represent the average and the most common number of images/sentences per document, respectively. For MMC4 and OBELICS, the distribution of images per document is highly concentrated around very low values of $M$ (2 and 1 per document, respectively). This indicates a significant scarcity of the image modality in their multimodal data, which directly limits the models' multimodal in-context learning ability. + +Examining Figure 2(a) and Figure 2(b) distributions, it is evident that the number of images does not significantly change with an increase in the number of sentences. However, in practical scenarios, longer documents are more likely to contain a greater number of illustrations, which aids in modeling longer multimodal contexts. In contrast, CoMM features a larger quantity of images per document ( $M = 4$ , $\mu = 10.1$ ), which significantly enhances the dataset's image modality while focusing on multimodal coherence. Furthermore, CoMM exhibits a trend where longer documents tend to include more image illustrations, thereby facilitating the modeling of longer multimodal contexts. + +# 4. Interleaved Generation Task and Benchmark + +# 4.1. Task Formulation + +In this section, we propose four benchmark tasks to evaluate MLLMs' interleaved generation capabilities. For ease of presentation, we utilize $\{x_{i}|1\leq i\leq N\}$ and $\{t_i|1\leq i\leq$ $N\}$ to represent the image sequences and text sequences, respectively. Besides, we use $t_\mathrm{prompt}$ as the text prompt, and $\mathcal{F}_{\theta}(\cdot)$ as the model with parameter $\theta$ . The detailed + +formulations of tasks are outlined as follows: + +Task 1: Image-to-Text Sequence Generation. This task evaluates the model's capability to generate coherent and contextually appropriate narratives based on sequential images. The objective is to predict the corresponding text sequences from the given image sequences: + +$$ +\{t _ {i} | 1 \leq i \leq N \} = \mathcal {F} _ {\theta} \left(t _ {\text {p r o m p t}}, x _ {1}, \dots , x _ {N}\right). \tag {2} +$$ + +Task 2: Text-to-Image Sequence Generation. This task assesses the model's proficiency in generating image sequences that accurately represent provided textual descriptions. The goal is to predict the corresponding image sequences from the given text sequences: + +$$ +\{x _ {i} \mid 1 \leq i \leq N \} = \mathcal {F} _ {\theta} \left(t _ {\text {p r o m p t}}, t _ {1}, \dots , t _ {N}\right). \tag {3} +$$ + +Task 3: Interleaved Image-Text Content Continuation. It measures the model's ability to seamlessly continue content in an interleaved image-text format. The model needs to predict the outputs of the remaining image-text steps based on the interleaved image-text content in the previous $k$ steps: + +$$ +\left\{t _ {i}, x _ {i} \mid k < i \leq N \right\} = \mathcal {F} _ {\theta} \left(t _ {\text {p r o m p t}}, t _ {1}, x _ {1}, \dots , t _ {k}, x _ {k}\right). \tag {4} +$$ + +Task 4: Question-based Interleaved Image-Text Generation. This challenging task requires the model to generate interleaved content based on a given question, testing its understanding and generative reasoning abilities across modalities. The model needs to predict the outcomes of all image-text steps based on the given question: + +$$ +\left\{t _ {i}, x _ {i} \mid 1 \leq i \leq N \right\} = \mathcal {F} _ {\theta} \left(t _ {\text {p r o m p t}}\right). \tag {5} +$$ + +# 4.2. Evaluation Metric + +This section lists the evaluation metrics for assessing the quality of generated results. Our evaluation framework1 measures textual accuracy, visual fidelity, and the coherence of multi-modal outputs. + +Evaluation for Text Sequence Generation. Generating narratives for a series of images shares similarities with the established task of image captioning. Consequently, we directly apply traditional image captioning metrics like ROUGE [21] and METEOR [5] at both the step and document levels. Additionally, we propose Illustration Relevance Score (IRS) to assess the relevance between image illustration and text context in the final document. Specifically, we simulate the human document creation process by first generating detailed image descriptions from the text context and then assessing the consistency between these descriptions and corresponding images using GPT-4o [27]. + +Evaluation for Image Sequence Generation. We evaluate the quality of image illustrations on four key assessment aspects including image quality, style consistency between generated images, reconstruction on referenced images, and IRS. To evaluate image quality, we use Fréchet Inception Distance (FID) [13] with ground truth reference and Inception Score (IS) [33]. For assessing style consistency, we employ the Structural Similarity Index Measure (SSIM), and for ground truth reconstruction evaluation, we adopt the Peak Signal-to-Noise Ratio (PSNR). + +Evaluation for Interleaved Image-Text Generation. We use MLLM (GPT-4o [27]) for a comprehensive assessment of linguistic and visual coherence. GPT-4o evaluates four aspects: image sequence coherence, image quality, document completeness, and IRS. For image sequence coherence, we check style, entity consistency, and content trend alignment between image and text sequence. Image quality is evaluated on authenticity, integrity, clarity, and aesthetics. Document completeness is scored based on how thoroughly the content covers the topic. + +# 5. Experiments + +# 5.1. Datasets Quality Comparison + +Settings. In this section, we compare the quality of our dataset with previous interleaved image-text datasets (e.g., MMC4 [45] and OBELICS [19]) focusing on text sequences, image sequences, and image-text alignment. We utilize powerful LLMs (e.g., Llama3 [25] and GPT-4o [27]) to assess the text development, text completeness, and image-text alignment of these datasets. Each metric is scored on a scale of 0 to 10, with detailed definitions and prompts provided in the Appendix. For MMC4 and OBELICS, we randomly sample 5,000 documents from their data. We evaluate our entire dataset with Llama3 and randomly sample 5,000 documents for assessment by GPT-4o. To evaluate image sequence quality, we employ the metric $\mathcal{F}$ as cited in Eq. (1). + +Quantitative Results. As shown in Table 1, our dataset surpasses MMC4 and OBELICS across all evaluated dimensions. Notably, the score difference in development, + +
DatasetsDLP L.3 / G.4oCPL L.3 / G.4oITA L.3 / G.4oImgS
MMC4 [45]5.56 / 4.756.28 / 5.126.53 / 4.660.21
OBELICS [19]5.60 / 5.976.51 / 5.885.00 / 3.811.00
CoMM (Ours)6.93 / 7.647.44 / 7.077.46 / 8.914.27
+ +Table 1. Quality comparison of interleaved image-text datasets. "DLP" stands for Development, "CPL" signifies Completeness, "ITA" represents Image-Text Alignment, "ImgS" refers to Image Sequence, "L.3" denotes Llama3 [25], and "G.4o" is GPT-4o [27]. + +completeness, and image-text alignment is more pronounced when evaluated by GPT-4o, a more powerful MLLM. Furthermore, the image sequence metric underscores the superior quality of our data, reflecting higher scores compared to MMC4 and OBELICS. These results affirm the high quality of our dataset in interleaved image-text data. + +# 5.2. In-Context Multimodal Understanding + +Setting. To further demonstrate the high-quality and validate the efficacy of our dataset in promoting in-context understanding, we conduct experiments using few-shot learning settings on several downstream tasks (COCO [22], Flickr30k [43], VQAv2 [11], OKVQA [24], TextVQA [35], VizWiz [12], and Hateful Memes [17]). Few-shot learning is critical for assessing a model's ability of in-context understanding. In our experiments, few-shot scenarios are simulated by providing the model with a small number of example queries and their corresponding answers before it is evaluated on new queries. We consider various shot settings (0, 4, 8, 16, and 32) to understand the model's performance across different levels of supervision. + +Quantitative Results. Table 2 presents the results for the baseline model and its enhancements with MMC4, OBELICS, and our proposed dataset. As observed from Table 2, our method consistently outperforms the performance of MMC4 and OBELICS across all shot settings, demonstrating superior generalization and in-context learning capabilities. Particularly noticeable improvements are seen in the COCO and TextVQA datasets, where our method exceeds the baseline by significant margins in nearly all shot settings. This superior performance underlines the strength of our dataset in enabling models to better learn from context and apply learned knowledge to unseen instances. Furthermore, our dataset surpasses the baseline in all tasks with more shot settings, maintaining its advantage with longer context few-shot data, as shown in the 16-shot and 32-shot scenarios. + +# 5.3. Interleaved Generation Tasks + +Baselines. This section introduces baselines developed from several state-of-the-art approaches, i.e., MiniGPT-5 [44], SEED-Llama [9], and Emu2 [36], to assess their effective + +
ShotCOCOFlickr30kVQAv2OKVQATextVQAVizWizHateful Memes
Baseline79.559.552.737.824.227.551.6
+ MMC4 [45]097.167.045.829.823.415.054.6
+ OBELICS [19]83.551.550.938.925.621.953.8
+ CoMM (Ours)100.369.852.641.333.523.654.8
Baseline89.065.854.840.128.234.154.0
+ MMC4 [45]4103.368.849.536.728.724.455.0
+ OBELICS [19]103.765.253.243.031.429.955.7
+ CoMM (Ours)107.073.253.642.935.730.155.2
Baseline96.362.954.841.129.138.554.7
+ MMC4 [45]8103.667.950.539.028.832.751.3
+ OBELICS [19]101.963.953.744.532.237.251.2
+ CoMM (Ours)108.074.854.244.535.437.952.5
Baseline98.862.854.342.727.342.553.9
+ MMC4 [45]16102.965.550.039.126.436.957.2
+ OBELICS [19]100.858.053.145.330.242.457.8
+ CoMM (Ours)109.371.554.246.935.843.559.9
Baseline99.561.353.342.423.844.053.8
+ MMC4 [45]3294.956.448.335.620.037.749.2
+ OBELICS [19]96.553.152.043.626.844.556.1
+ CoMM (Ours)111.970.754.046.634.945.757.3
+ +Table 2. Performance comparison of different datasets, with the baseline model Open-Flamingo 9B [4]. Evaluations were in an open-ended setting for VQA tasks with random in-context examples. (Task, Metric, Query Split): (COCO [22], CIDEr, test), (Flickr30k [43], CIDEr, test (Karpathy)), (VQAv2 [11], VQA accuracy, testdev), (OKVQA [24], VQA accuracy, val), (TextVQA [35], VQA accuracy, val), (VizWiz [12], VQA accuracy, testdev), (Hateful Memes [17], ROC-AUC, test seen). + +
MethodsSizeI2T Sequence Generation (Task1)StyleContinuation Generation (Task3)
M.S.R.S.M.W.R.W.IRSEntityTrendCPL.ImgQIRS
MiniGPT-5 [44]7B4.608.649.1010.677.395.585.215.246.336.362.56
SEED-Llama [9]8B2.296.344.207.336.996.285.845.726.286.552.92
SEED-Llama [9]14B3.168.125.718.178.136.686.226.136.666.673.23
Emu2 [36]33B3.658.417.438.625.018.227.997.978.498.622.42
MethodsSizeT2I Sequence Generation (Task2)StyleQuestion-based Generation (Task4)
FID ↓ISSSIMPSNRIRSEntityTrendCPL.ImgQIRS
MiniGPT-5 [44]7B56.517.6020.667.575.485.655.205.255.816.152.71
SEED-Llama [9]8B57.968.4020.537.875.277.556.816.155.136.361.46
SEED-Llama [9]14B66.2311.0620.838.126.247.516.616.306.136.662.50
Emu2 [36]33B60.456.3420.967.914.178.417.567.637.547.592.02
+ +Table 3. Comparison of performance among MiniGPT-5 [44], SEED-Llama [9], and Emu2 [37] on the four generation tasks. "M.S." and "R.S." represent METEOR [5] and ROUGE_L [21] for step-by-step evaluation, while "M.W." and "R.W." correspond to METEOR and ROUGE_L for whole document evaluation. "CPL." stands for Completeness, "ImgQ" indicates Image Quality, "IRS" means Illustration Relevance Score, and "↓" denotes that lower values are better. + +ness in generating interleaved image-text content. Details of these models are thoroughly provided in the Appendix. To construct the baselines, we fine-tune each of these models using our CoMM dataset. We employ a rigorous evaluation, assessing them on benchmark tasks involving interleaved image-text generation and understanding. + +Quantitative Results. As shown in Table 3, the quantitative results clearly illustrate the varying performance of MiniGPT-5, SEED-Llama, and Emu2 across the four interleaved generation tasks. For I2T Sequence Generation (Task1), MiniGPT-5 achieves the highest scores in ME + +TEOR, ROUGE_L, but falls short in Illustration Relevance Score (IRS) where SEED-Llama (14B) leads with a score of 8.13. For T2I Sequence Generation (Task2), MiniGPT-5 excels with the lowest FID score of 56.51, while SEED-Llama (14B) surpasses IS, SSIM, PSNR, and IRS metrics, suggesting stronger visual fidelity and semantic similarity. Emu2 tops in Continuation Generation (Task3) with the highest scores in Style, Entity, Trend, Completeness, and Image Quality. In Question-based Generation (Task4), Emu2 dominates in Style, Entity, Trend, Completeness, and Image Quality, while MiniGPT-5 manages to secure the lead in IRS. + +![](images/421bd8ca184d99258631bc823a2fba0d2780d95e00f20dd6384697c50f098283.jpg) +Figure 3. Visualization of interleaved image-text content generation from SEED-Llama [9] (Top) and MiniGPT-5 [44] (Bottom) separately trained by MMC4 [45] and CoMM [Ours]. + +Overall, these results highlight the strengths and weaknesses of each model in handling different aspects of interleaved image-text generation tasks, providing valuable insights for future research and model improvements. + +Preference Dataset Training Results. With SEED-Llama [9] unifying image to discrete token prediction modeling, we can directly apply the mainstream preference learning algorithm DPO [30] to train SEED-Llama on our preference dataset following supervised fine-tuning on our dataset. As shown in Table 4, performance further improves on the image-to-text sequence generation task after training with our preference dataset. Additional results on preference dataset training are available in the Appendix. + +Multimodal Generation Qualitative Analysis. As illustrated in Figure 1 and Figure 3, the interleaved image-text content generated by models [9, 36, 44] trained on MMC4 [45] shows a noticeable difference when compared to the content generated by models trained on our datasets. Our models produce content that is more consistent with the previous input context (e.g., textual coherence and entity consistency). By ensuring higher coherence, consistency, and alignment, CoMM sets a new standard for training datasets in the realm of multimodal large language models. More visualizations are available in the Appendix. + +
MethodsSizeI2T Sequence Generation (Task1)
M.S.R.S.M.W.R.W.IRS
SEED-Llama [9]8B2.296.344.207.336.99
+ DPO [30]8B4.098.836.9510.247.26
SEED-Llama [9]14B3.168.125.718.178.13
+ DPO [30]14B4.6210.158.0810.308.18
+ +Table 4. Performance results of SEED-Llama [9] trained by DPO [30] in our preference dataset. "M.S." and "R.S." represent METEOR [5] and ROUGE_L [21] for step-by-step evaluation, while "M.W." and "R.W." correspond to METEOR and ROUGE_L for whole document evaluation. + +# 6. Conclusion + +In this paper, we present CoMM, a high-quality interleaved image-text dataset designed to address the limitations of existing multimodal large language models (MLLMs) by ensuring superior text-image alignment and entity consistency. Through diverse interleaved image-text content, CoMM significantly enhances MLLMs' in-context learning capabilities, as demonstrated by extensive few-shot experiments. Additionally, we introduce four novel tasks and a comprehensive evaluation framework to assess interleaved image-text generation abilities, setting new benchmarks for advancing multimodal understanding and generation in MLLMs. + +# Acknowledgment + +This work was supported by the Hong Kong SAR RGC Early Career Scheme (26208924), the National Natural Science Foundation of China Young Scholar Fund (62402408), and the HKUST Sports Science and Technology Research Grant (SSTRG24EG04), ACCESS - AI Chip Center for Emerging Smart Systems, sponsored by InnoHK funding, Hong Kong SAR. + +# References + +[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florecia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 1 +[2] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. NeurIPS, 2022. 3 +[3] Jie An, Zhengyuan Yang, Linjie Li, Jianfeng Wang, Kevin Lin, Zicheng Liu, Lijuan Wang, and Jiebo Luo. Openleaf: Open-domain interleaved image-text generation and evaluation. arXiv preprint arXiv:2310.07749, 2023. 1, 2, 3 +[4] Anas Awadalla, Irena Gao, Josh Gardner, Jack Hessel, Yusuf Hanafy, Wanrong Zhu, Kalyani Marathe, Yonatan Bitton, Samir Gadre, Shiori Sagawa, et al. Openflamingo: An opensource framework for training large autoregressive vision-language models. arXiv preprint arXiv:2308.01390, 2023. 3, 7 +[5] Satanjeev Banerjee and Alon Lavie. Meteor: An automatic metric for mt evaluation with improved correlation with human judgments. In ACL Workshop, pages 65-72, 2005. 3, 6, 7, 8 +[6] Runpei Dong, Chunrui Han, Yuang Peng, Zekun Qi, Zheng Ge, Jinrong Yang, Liang Zhao, Jianjian Sun, Hongyu Zhou, Haoran Wei, et al. Dreamllm: Synergistic multimodal comprehension and creation. arXiv preprint arXiv:2309.11499, 2023. 1, 3 +[7] Samir Yitzhak Gadre, Gabriel Ilharco, Alex Fang, Jonathan Hayase, Georgios Smyrnis, Thao Nguyen, Ryan Marten, Mitchell Wortsman, Dhruba Ghosh, Jieyu Zhang, et al. Datacomp: In search of the next generation of multimodal datasets. NeurIPS, 36, 2023. 4 +[8] Bofei Gao, Feifan Song, Yibo Miao, Zefan Cai, Zhe Yang, Liang Chen, Helan Hu, Runxin Xu, Qingxiu Dong, Ce Zheng, et al. Towards a unified view of preference learning for large language models: A survey. arXiv preprint arXiv:2409.02795, 2024. 3, 4 +[9] Yuying Ge, Sijie Zhao, Ziyun Zeng, Yixiao Ge, Chen Li, Xintao Wang, and Ying Shan. Making llama see and draw with seed tokenizer. arXiv preprint arXiv:2310.01218, 2023. 6, 7, 8, 1, 3, 4, 9, 10, 11, 12, 13 +[10] Yuying Ge, Sijie Zhao, Jinguo Zhu, Yixiao Ge, Kun Yi, Lin Song, Chen Li, Xiaohan Ding, and Ying Shan. Seed-x: Multimodal models with unified multi-granularity comprehension and generation. arXiv preprint arXiv:2404.14396, 2024. 1 + +[11] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Bartra, and Devi Parikh. Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In CVPR, 2017. 3, 6, 7 +[12] Danna Gurari, Qing Li, Abigale J Stangl, Anhong Guo, Chi Lin, Kristen Grauman, Jiebo Luo, and Jeffrey P Bigham. Vizwiz grand challenge: Answering visual questions from blind people. In CVPR, 2018. 3, 6, 7 +[13] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. NeurIPS, 2017. 3, 6 +[14] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 7 +[15] S Huang, L Dong, W Wang, Y Hao, S Singhal, S Ma, T Lv, L Cui, OK Mohammed, Q Liu, et al. Language is not all you need: aligning perception with language models. 2023. arXiv preprint arXiv:2302.14045. 3 +[16] Timo Kaufmann, Paul Weng, Viktor Bengs, and Eyke Hüllermieier. A survey of reinforcement learning from human feedback. arXiv preprint arXiv:2312.14925, 2023. 3, 4 +[17] Douwe Kiela, Hamed Firooz, Aravind Mohan, Vedanuj Goswami, Amanpreet Singh, Pratik Ringshia, and Davide Testuggine. The hateful memes challenge: Detecting hate speech in multimodal memes. NeurIPS, 2020. 3, 6, 7 +[18] Jing Yu Koh, Daniel Fried, and Russ R Salakhutdinov. Generating images with multimodal language models. NeurIPS, 2024. 1 +[19] Hugo Laurençon, Lucile Saulnier, Léo Tronchon, Stas Bekman, Amanpreet Singh, Anton Lozhkov, Thomas Wang, Siddharth Karamcheti, Alexander Rush, Douwe Kiela, et al. Obelics: An open web-scale filtered dataset of interleaved image-text documents. NeurIPS, 2023. 3, 6, 7 +[20] Samuel Lavoie, Polina Kirichenko, Mark Ibrahim, Mahmoud Assran, Andrew Gordon Wildon, Aaron Courville, and Nicolas Ballas. Modeling caption diversity in contrastive vision-language pretraining. arXiv preprint arXiv:2405.00740, 2024. 4 +[21] Chin-Yew Lin. Rouge: A package for automatic evaluation of summaries. In Text summarization branches out, 2004. 3, 6, 7, 8 +[22] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dolkar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 3, 6, 7 +[23] Chang Liu, Haoning Wu, Yujie Zhong, Xiaoyun Zhang, and Weidi Xie. Intelligent grimm-open-ended visual storytelling via latent diffusion models. arXiv preprint arXiv:2306.00973, 2023. 1, 2 +[24] Kenneth Marino, Mohammad Rastegari, Ali Farhadi, and Roozbeh Mottaghi. Ok-vqa: A visual question answering benchmark requiring external knowledge. In CVPR, 2019. 3, 6, 7 +[25] Meta. https://llama.meta.com/llama3, 2024.2, 4,6,1 + +[26] Microsoft. Bingchat. https://www.microsoft.com/en-us/edge/features/bing-chat, 2023. Accessed: 2023-10-09. 3 +[27] OpenAI. https://openai.com/index/hello-gpt-4o/, 2024.3,4,6 +[28] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 1 +[29] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 2, 4 +[30] Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. Direct preference optimization: Your language model is secretly a reward model. NeurIPS, 36, 2023. 8, 1, 3 +[31] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. JMLR, 2020. 3 +[32] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 3 +[33] Tim Salimans, Ian Goodfellow, Wojciech Zaremba, Vicki Cheung, Alec Radford, and Xi Chen. Improved techniques for training gans. NeurIPS, 2016. 3, 6 +[34] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. NeurIPS, 35:25278–25294, 2022. 4 +[35] Amanpreet Singh, Vivek Natarajan, Meet Shah, Yu Jiang, Xinlei Chen, Dhruv Batra, Devi Parikh, and Marcus Rohrbach. Towards vqa models that can read. In CVPR, 2019. 6, 7 +[36] Quan Sun, Yufeng Cui, Xiaosong Zhang, Fan Zhang, Qiying Yu, Zhengxiong Luo, Yueze Wang, Yongming Rao, Jingjing Liu, Tiejun Huang, et al. Generative multimodal models are in-context learners. arXiv preprint arXiv:2312.13286, 2023. 1, 2, 6, 7, 8, 4, 9, 10, 11, 12 +[37] Quan Sun, Qiying Yu, Yufeng Cui, Fan Zhang, Xiaosong Zhang, Yueze Wang, Hongcheng Gao, Jingjing Liu, Tiejun Huang, and Xinlong Wang. Emu: Generative pretraining in multimodality. In ICLR, 2023. 3, 7 +[38] Changyao Tian, Xizhou Zhu, Yuwen Xiong, Weiyun Wang, Zhe Chen, Wenhai Wang, Yuntao Chen, Lewei Lu, Tong Lu, Jie Zhou, et al. Mm-interleaved: Interleaved image-text generative modeling via multi-modal feature synchronizer. arXiv preprint arXiv:2401.10208, 2024. 3 +[39] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 7 +[40] Weihan Wang, Qingsong Lv, Wenmeng Yu, Wenyi Hong, Ji Qi, Yan Wang, Junhui Ji, Zhuoyi Yang, Lei Zhao, Xixuan + +Song, Jiazheng Xu, Bin Xu, Juanzi Li, Yuxiao Dong, Ming Ding, and Jie Tang. CogvIm: Visual expert for pretrained language models, 2023. 4 +[41] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. NeurIPS, 2022. 1 +[42] WikiHow. https://www.wikipiow.com. 2,3 +[43] Peter Young, Alice Lai, Micah Hodosh, and Julia Hockenmaier. From image descriptions to visual denotations: New similarity metrics for semantic inference over event descriptions. TACL, 2014. 6, 7 +[44] Kaizhi Zheng, Xuehai He, and Xin Eric Wang. Minigpt-5: Interleaved vision-and-language generation via generative vowels. arXiv preprint arXiv:2310.02239, 2023. 1, 6, 7, 8, 4, 9, 10, 11, 12 +[45] Wanrong Zhu, Jack Hessel, Anas Awadalla, Samir Yitzhak Gadre, Jesse Dodge, Alex Fang, Youngjae Yu, Ludwig Schmidt, William Yang Wang, and Yejin Choi. Multimodal c4: An open, billion-scale corpus of images interleaved with text. NeurIPS, 2023. 1, 2, 3, 4, 6, 7, 8 \ No newline at end of file diff --git a/2025/CoMM_ A Coherent Interleaved Image-Text Dataset for Multimodal Understanding and Generation/images.zip b/2025/CoMM_ A Coherent Interleaved Image-Text Dataset for Multimodal Understanding and Generation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..a2d847ab8bbf982235b792425414a7530c04ccfe --- /dev/null +++ b/2025/CoMM_ A Coherent Interleaved Image-Text Dataset for Multimodal Understanding and Generation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f9beab7ede35268fb9f0e19f1d65bd95765f6f785f64d7ce066c229baa1adaa1 +size 795529 diff --git a/2025/CoMM_ A Coherent Interleaved Image-Text Dataset for Multimodal Understanding and Generation/layout.json b/2025/CoMM_ A Coherent Interleaved Image-Text Dataset for Multimodal Understanding and Generation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..eb609d1fea733b0b29869330c93fc25a083cb2c0 --- /dev/null +++ b/2025/CoMM_ A Coherent Interleaved Image-Text Dataset for Multimodal Understanding and Generation/layout.json @@ -0,0 +1,7004 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 143, + 103, + 470, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 103, + 470, + 140 + ], + "spans": [ + { + "bbox": [ + 143, + 103, + 470, + 140 + ], + "type": "text", + "content": "CoMM: A Coherent Interleaved Image-Text Dataset for Multimodal Understanding and Generation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 121, + 160, + 491, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 160, + 491, + 219 + ], + "spans": [ + { + "bbox": [ + 121, + 160, + 491, + 219 + ], + "type": "text", + "content": "Wei Chen" + }, + { + "bbox": [ + 121, + 160, + 491, + 219 + ], + "type": "inline_equation", + "content": "^{1,2*}" + }, + { + "bbox": [ + 121, + 160, + 491, + 219 + ], + "type": "text", + "content": ", Lin Li" + }, + { + "bbox": [ + 121, + 160, + 491, + 219 + ], + "type": "inline_equation", + "content": "^{1,3*}" + }, + { + "bbox": [ + 121, + 160, + 491, + 219 + ], + "type": "text", + "content": ", Yongqi Yang" + }, + { + "bbox": [ + 121, + 160, + 491, + 219 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 121, + 160, + 491, + 219 + ], + "type": "text", + "content": ", Bin Wen" + }, + { + "bbox": [ + 121, + 160, + 491, + 219 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 121, + 160, + 491, + 219 + ], + "type": "text", + "content": ", Fan Yang" + }, + { + "bbox": [ + 121, + 160, + 491, + 219 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 121, + 160, + 491, + 219 + ], + "type": "text", + "content": ", Tingting Gao" + }, + { + "bbox": [ + 121, + 160, + 491, + 219 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 121, + 160, + 491, + 219 + ], + "type": "text", + "content": ", Yu Wu" + }, + { + "bbox": [ + 121, + 160, + 491, + 219 + ], + "type": "inline_equation", + "content": "^{2\\dagger}" + }, + { + "bbox": [ + 121, + 160, + 491, + 219 + ], + "type": "text", + "content": ", Long Chen" + }, + { + "bbox": [ + 121, + 160, + 491, + 219 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 121, + 160, + 491, + 219 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 121, + 160, + 491, + 219 + ], + "type": "text", + "content": "The Hong Kong University of Science and Technology, " + }, + { + "bbox": [ + 121, + 160, + 491, + 219 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 121, + 160, + 491, + 219 + ], + "type": "text", + "content": "Wuhan University, " + }, + { + "bbox": [ + 121, + 160, + 491, + 219 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 121, + 160, + 491, + 219 + ], + "type": "text", + "content": "AI Chip Center for Emerging Smart Systems, " + }, + { + "bbox": [ + 121, + 160, + 491, + 219 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 121, + 160, + 491, + 219 + ], + "type": "text", + "content": "Kuaishou Technology" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 124, + 220, + 487, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 124, + 220, + 487, + 232 + ], + "spans": [ + { + "bbox": [ + 124, + 220, + 487, + 232 + ], + "type": "text", + "content": "{wchendb, lllidy, longchen}@ust.hk, {yongqiyang, wuyucs}@whu.edu.cn" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 198, + 234, + 412, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 234, + 412, + 245 + ], + "spans": [ + { + "bbox": [ + 198, + 234, + 412, + 245 + ], + "type": "text", + "content": "https://github.com/HKUST-LongGroup/CoMM" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 151, + 273, + 200, + 285 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 273, + 200, + 285 + ], + "spans": [ + { + "bbox": [ + 151, + 273, + 200, + 285 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 297, + 297, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 297, + 297, + 622 + ], + "spans": [ + { + "bbox": [ + 54, + 297, + 297, + 622 + ], + "type": "text", + "content": "Interleaved image-text generation has emerged as a vital multimodal task aimed at creating sequences of interleaved visual and textual content given a query. Despite notable advancements in recent multimodal large language models (MLLMs), generating integrated image-text sequences that exhibit narrative coherence and entity and style consistency remains challenging due to poor training data quality. To this end, we introduce CoMM, a high-quality Coherent interleaved image-text MultiModal dataset designed to enhance the coherence, consistency, and alignment of generated multimodal content. Initially, CoMM harnesses raw data from diverse sources, focusing on instructional content and visual storytelling, establishing a foundation for coherent and consistent content. To further refine the data quality, we devise a multi-perspective filter strategy that leverages advanced pre-trained models to ensure the development of sentences, consistency of inserted images, and semantic alignment between them. Various quality evaluation metrics are designed to prove the high quality of the filtered dataset. Meanwhile, extensive few-shot experiments on various downstream tasks demonstrate CoMM's effectiveness in significantly enhancing the in-context learning capabilities of MLLMs. Moreover, we propose four new tasks to evaluate MLLMs' interleaved generation abilities, supported by a comprehensive evaluation framework. We believe CoMM opens a new avenue for advanced MLLMs with superior multimodal in-context learning and understanding ability." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 642, + 137, + 654 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 137, + 654 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 137, + 654 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 662, + 296, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 662, + 296, + 687 + ], + "spans": [ + { + "bbox": [ + 55, + 662, + 296, + 687 + ], + "type": "text", + "content": "Interleaved image-text generation has begeoned as an emerging multimodal task, striving to imitate the human-" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 274, + 556, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 274, + 556, + 357 + ], + "spans": [ + { + "bbox": [ + 313, + 274, + 556, + 357 + ], + "type": "text", + "content": "like capability to alternately create visual and textual content [3, 45]. Specifically, it aims to generate a sequence of interleaved text descriptions and illustrative images given a query [3]. With its interleaved multimodal generation ability, state-of-the-art models can facilitate a variety of applications, e.g., multimodal instruction generation, tutorial step generation, and visual storytelling." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 361, + 557, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 361, + 557, + 566 + ], + "spans": [ + { + "bbox": [ + 313, + 361, + 557, + 566 + ], + "type": "text", + "content": "Recent advancements in multimodal large language models (MLLMs) have exhibited exceptional capabilities in cross-modal generation, but they still struggle to generate interleaved image-text sequences coherently [36, 44]. The main reason is that most of them are trained on single imagetext pairs, limited in generating coherent and contextually integrated multimodal content [18, 44]. To this end, some efforts integrate multiple models (e.g., GPT-4 [1] and Stable Diffusion XL [28]) to generate open-domain interleaved image-text contents with arbitrary formats in a training-free manner [3]. However, their results are often dominated by the generated text descriptions, limited by the generation model's understanding of these textual inputs. In contrast, other research efforts [6, 10] have attempted to train models with seamless multimodal understanding and generation capabilities using existing interleaved image-text datasets (e.g., MMC4 [45])." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 570, + 557, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 557, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 557, + 715 + ], + "type": "text", + "content": "Albeit the unprecedented progress, it is becoming increasingly evident that data-centric methods still suffer several limitations: 1) Narrative Coherence: The contextual relevance and coherence of generated image-text sentences are relatively low. In Figure 1(a), the model pre-trained on the MMC4 [45] directly draws the final cooked dessert, missing the intermediate coherent steps of generation. These coherent generation steps not only enhance interpretability but also boost the reasoning capabilities of the model [41]. 2) Entity and Style Consistency: The entity and subject styles of the generated illustrated images are inconsistent. As illustrated in Figure 1(b), the style of the blue-gray sofa image" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "type": "text", + "content": "*Equal contribution. Work was done when Wei Chen interned in Kuaishou. † Corresponding authors." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8073" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 73, + 550, + 289 + ], + "blocks": [ + { + "bbox": [ + 57, + 73, + 550, + 289 + ], + "lines": [ + { + "bbox": [ + 57, + 73, + 550, + 289 + ], + "spans": [ + { + "bbox": [ + 57, + 73, + 550, + 289 + ], + "type": "image", + "image_path": "422e031e86cb9a9247ec9858f00403663e8850d2300d4b483d9e933327b7febe.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 128, + 303, + 206, + 312 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 128, + 303, + 206, + 312 + ], + "spans": [ + { + "bbox": [ + 128, + 303, + 206, + 312 + ], + "type": "text", + "content": "(c) Sample from MMC4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 70, + 314, + 236, + 345 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 314, + 236, + 345 + ], + "spans": [ + { + "bbox": [ + 70, + 314, + 236, + 345 + ], + "type": "text", + "content": "When it comes to making life just a little bit smoother and safer in a wonderful place such as White Bear Lake, ADT Monitored Home Security can be your main solution for keeping your family and belongings safe and secure." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 70, + 349, + 236, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 349, + 236, + 380 + ], + "spans": [ + { + "bbox": [ + 70, + 349, + 236, + 380 + ], + "type": "text", + "content": "An ADT home monitoring system for your home can change your life for the better, letting you explore and enjoy all that you love in White Bear Lake. ADT home monitoring is more than just home security, it raises the bar to better living." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 70, + 384, + 230, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 384, + 230, + 407 + ], + "spans": [ + { + "bbox": [ + 70, + 384, + 230, + 407 + ], + "type": "text", + "content": "It's a chance to live your daily life without all the worry and stress. For over a century, ADT monitoring systems have paved the way in home security." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 70, + 409, + 270, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 409, + 270, + 439 + ], + "spans": [ + { + "bbox": [ + 70, + 409, + 270, + 439 + ], + "type": "text", + "content": "As the #1 home security provider in the country, we know all the vital facts and components that are necessary in keeping our " + }, + { + "bbox": [ + 70, + 409, + 270, + 439 + ], + "type": "inline_equation", + "content": "6+" + }, + { + "bbox": [ + 70, + 409, + 270, + 439 + ], + "type": "text", + "content": " million home and business customers safe. It's why so many people in White Bear Lake and the rest of the nation count on us for their safety." + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 239, + 317, + 272, + 341 + ], + "blocks": [ + { + "bbox": [ + 239, + 317, + 272, + 341 + ], + "lines": [ + { + "bbox": [ + 239, + 317, + 272, + 341 + ], + "spans": [ + { + "bbox": [ + 239, + 317, + 272, + 341 + ], + "type": "image", + "image_path": "3bc55a838fd7d914e38e61a57768a6e3631f8bf27f0b3c4d338e24b8cb03591f.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 242, + 350, + 266, + 373 + ], + "blocks": [ + { + "bbox": [ + 242, + 350, + 266, + 373 + ], + "lines": [ + { + "bbox": [ + 242, + 350, + 266, + 373 + ], + "spans": [ + { + "bbox": [ + 242, + 350, + 266, + 373 + ], + "type": "image", + "image_path": "fc53a4dd0cfeee1bc185222928bc5b136f1989ed264891ebd792b00625a91a7a.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 242, + 385, + 264, + 406 + ], + "blocks": [ + { + "bbox": [ + 242, + 385, + 264, + 406 + ], + "lines": [ + { + "bbox": [ + 242, + 385, + 264, + 406 + ], + "spans": [ + { + "bbox": [ + 242, + 385, + 264, + 406 + ], + "type": "image", + "image_path": "600f8c0011ccc3ddf62fd093b815e5d4b11e495b3331c2b1cad99d7be28cc04e.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 491, + 296, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 491, + 296, + 634 + ], + "spans": [ + { + "bbox": [ + 55, + 491, + 296, + 634 + ], + "type": "text", + "content": "generated by the model pre-trained on MMC4 is inconsistent with that in the query. This consistency of entities and style has been proposed as an important metric for evaluating the results of interleaved image-text content generation [3]. These limitations primarily stem from the low quality of training data caused by the poor alignment between image and text content, i.e., the devil is in the data. For example in Figure 1(c), the second image in the training sample has a noticeable gap with the context. Additionally, the styles of all the images are almost completely different. Thus, it raises a natural research question: How to construct a high-quality interleaved image-text dataset?" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "type": "text", + "content": "In this paper, we introduce CoMM, a high-quality and coherent interleaved image-text dataset. By \"high-quality\", we mean that the interleaved image-text content should exhibit developed and logically structured text steps, consistent entities and visual styles in images, and strong semantic and contextual alignment between them. To enhance text coher" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 372, + 301, + 452, + 310 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 301, + 452, + 310 + ], + "spans": [ + { + "bbox": [ + 372, + 301, + 452, + 310 + ], + "type": "text", + "content": "(d) Sample from CoMM" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 299, + 314, + 506, + 330 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 314, + 506, + 330 + ], + "spans": [ + { + "bbox": [ + 299, + 314, + 506, + 330 + ], + "type": "text", + "content": "Step1. Remove the tuna steaks from the refrigerator. Allow them to sit out at room temperature for 1/2 to 1 hour before you begin cooking." + } + ] + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 507, + 313, + 549, + 334 + ], + "blocks": [ + { + "bbox": [ + 507, + 313, + 549, + 334 + ], + "lines": [ + { + "bbox": [ + 507, + 313, + 549, + 334 + ], + "spans": [ + { + "bbox": [ + 507, + 313, + 549, + 334 + ], + "type": "image", + "image_path": "f96c0ed39e4909f58e8040b510f9c1b5d3e385ff16c3e687d32a48c5d3448eab.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 299, + 332, + 526, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 332, + 526, + 348 + ], + "spans": [ + { + "bbox": [ + 299, + 332, + 526, + 348 + ], + "type": "text", + "content": "Step2. Use a paper towel to dry excess moisture from the tuna steaks. Simply pat the paper towel on both sides of the steak to remove marinade or condensation from the meat." + } + ] + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 527, + 335, + 550, + 380 + ], + "blocks": [ + { + "bbox": [ + 527, + 335, + 550, + 380 + ], + "lines": [ + { + "bbox": [ + 527, + 335, + 550, + 380 + ], + "spans": [ + { + "bbox": [ + 527, + 335, + 550, + 380 + ], + "type": "image", + "image_path": "eb83bd16c70cc7505f21c4efb375e0ba6b4874dae1a77599aa1043a739777c3a.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "bbox": [ + 300, + 350, + 510, + 367 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 300, + 350, + 510, + 367 + ], + "spans": [ + { + "bbox": [ + 300, + 350, + 510, + 367 + ], + "type": "text", + "content": "Step3. Add 2 tbsp. olive oil or butter to the pan over medium-high heat. Tilt the pan to spread the oil or butter across the surface and coat the pan." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 299, + 369, + 518, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 369, + 518, + 384 + ], + "spans": [ + { + "bbox": [ + 299, + 369, + 518, + 384 + ], + "type": "text", + "content": "Step4. Place the tuna steaks on the hot pan. Do not crowd the tuna in the pan. Keep the steaks at least 3 to 4 inches apart to make sure they cook evenly." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 299, + 389, + 514, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 389, + 514, + 420 + ], + "spans": [ + { + "bbox": [ + 299, + 389, + 514, + 420 + ], + "type": "text", + "content": "Step5. Sear each side for two minutes. Do not move the meat around too much while it sears. Gently shake the pan after two minutes, and if the tuna steak releases from the surface, you can flip it on its other side. It is important to make sure that the edged are seared, but you want the inside of the tuna steak to remain rare." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 299, + 423, + 515, + 438 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 423, + 515, + 438 + ], + "spans": [ + { + "bbox": [ + 299, + 423, + 515, + 438 + ], + "type": "text", + "content": "Step6. Make a small cut in the center of the tuna steak with a knife. Inspect the inside to make sure that the middle of the steak stays rare and red in color." + } + ] + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 527, + 381, + 550, + 404 + ], + "blocks": [ + { + "bbox": [ + 527, + 381, + 550, + 404 + ], + "lines": [ + { + "bbox": [ + 527, + 381, + 550, + 404 + ], + "spans": [ + { + "bbox": [ + 527, + 381, + 550, + 404 + ], + "type": "image", + "image_path": "6232e448cfc5218dc8fabbed84e8e6f84eec0fea452ff61ff6cbe9bc8e2d8bc5.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 528, + 406, + 550, + 427 + ], + "blocks": [ + { + "bbox": [ + 55, + 446, + 555, + 479 + ], + "lines": [ + { + "bbox": [ + 55, + 446, + 555, + 479 + ], + "spans": [ + { + "bbox": [ + 55, + 446, + 555, + 479 + ], + "type": "text", + "content": "Figure 1. Illustration of interleaved image-text content generation results and dataset quality. (a) Given the query from CoMM, the interleaved image-text content generation results from the model Emu2 [36] separately trained by MMC4 [45] and CoMM (Ours). (b) The query is from the MMC4. (c) A training sample is from the MMC4. (d) A training sample is from the our CoMM." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 528, + 406, + 550, + 427 + ], + "lines": [ + { + "bbox": [ + 528, + 406, + 550, + 427 + ], + "spans": [ + { + "bbox": [ + 528, + 406, + 550, + 427 + ], + "type": "image", + "image_path": "5cbb59f3a24307ac392a50f844c7b9e0b3ac386f61a12df697944ad3bd8e38d9.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 491, + 556, + 706 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 491, + 556, + 706 + ], + "spans": [ + { + "bbox": [ + 313, + 491, + 556, + 706 + ], + "type": "text", + "content": "ence and image consistency, we collect raw data from specific websites, e.g., WikiHow [42], known for their instructional content or visual stories. As illustrated in Figure 1(d), instructional content, due to its uniform intention (e.g., cooking steak) and structured presentation (e.g., step 1), exhibits strong text coherence and image consistency. Additionally, we devise a multi-perspective filter strategy consisting of: a text sequence filter, an image sequence filter, and an imagedtext alignment filter. The first two leverage large language models (LLMs), e.g., Llama3 [25], and vision-language models (VLMs), e.g., CLIP [29], to eliminate incoherent text and images, respectively, further enhancing overall coherence. The last one employs both VLMs and MLLMs to comprehensively assess and refine the dataset for better alignment between visual and textual elements. After constructing this dataset, we further enhance it into a preference dataset by generating negative samples through shuffled text and images. This preference dataset can then be used during the" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8074" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 268, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 268, + 83 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 268, + 83 + ], + "type": "text", + "content": "reinforcement learning stage of post-training [8, 16]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 85, + 297, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 85, + 297, + 251 + ], + "spans": [ + { + "bbox": [ + 55, + 85, + 297, + 251 + ], + "type": "text", + "content": "To evaluate the data quality and make a comprehensive comparison of existing interleaved image-text benchmarks [19, 45], we design four metrics to evaluate the development, completeness, image-text alignment, and consistency of image-text sequences, separately (cf. Table 1). CoMM significantly outperforms previous datasets [19, 45] on all metrics, demonstrating its high quality. To further demonstrate the superiority of CoMM, we conduct few-shot experiments on various downstream tasks (e.g., VQA) [11, 12, 17, 22, 24], comparing with MMC4 [45] and OBELICS [19]. The model trained on our CoMM consistently outperforms others, particularly in long context settings (16 & 32 shots), showcasing CoMM's ability to enhance the in-context learning capacities of MLLMs." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 253, + 298, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 253, + 298, + 445 + ], + "spans": [ + { + "bbox": [ + 55, + 253, + 298, + 445 + ], + "type": "text", + "content": "Furthermore, thanks to our dataset, we introduce four new challenging tasks to evaluate the multimodal understanding and generation abilities of MLLMs: image-to-text sequence generation, text-to-image sequence generation, interleaved image-text content continuation, and question-based interleaved image-text generation. To assess the performance of these newly introduced tasks, we implement a comprehensive evaluation framework considering metrics such as METEOR [5], ROUGE [21] for text quality, and FID [13] and IS [33] for image quality. To accurately evaluate open-ended interleaved image-text generation content, we develop a series of metrics and leverage powerful MLLM (e.g., GPT4o [27]) for assessment. These tasks and evaluations pave the way for novel approaches in evaluating MLLMs by addressing critical gaps in current benchmarks. Conclusively, our contributions are as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 51, + 445, + 295, + 608 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 51, + 445, + 295, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 445, + 295, + 503 + ], + "spans": [ + { + "bbox": [ + 51, + 445, + 295, + 503 + ], + "type": "text", + "content": "1) We introduce CoMM, a high-quality coherent interleaved image-text dataset designed to address the limitations of existing datasets by ensuring coherent narrative, consistent entity and style, and strong semantic alignment between images and text." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 51, + 503, + 295, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 503, + 295, + 548 + ], + "spans": [ + { + "bbox": [ + 51, + 503, + 295, + 548 + ], + "type": "text", + "content": "2) CoMM introduces four novel benchmark tasks and associated evaluation metrics to comprehensively evaluate multimodal understanding and generation capabilities of MLLMs." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 51, + 548, + 295, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 548, + 295, + 608 + ], + "spans": [ + { + "bbox": [ + 51, + 548, + 295, + 608 + ], + "type": "text", + "content": "3) CoMM can serve as both a supervised fine-tuning dataset and a preference dataset. Extensive ablations across various downstream tasks demonstrate the high-quality of CoMM and its effectiveness in enhancing the understanding and generation capabilities of MLLMs." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 620, + 142, + 633 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 620, + 142, + 633 + ], + "spans": [ + { + "bbox": [ + 55, + 620, + 142, + 633 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 641, + 297, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 297, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 297, + 713 + ], + "type": "text", + "content": "Interleaved Image-Text Web Document Datasets. Training on interleaved image-text web documents has demonstrated superior performance over image-description pairs, as evidenced by studies such as Flamingo [2] and KOSMOS-1 [15]. These findings underscore the significant advantages of utilizing the richer and more meaningful correlations in" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 555, + 106 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 106 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 106 + ], + "type": "text", + "content": "herent in interleaved image-text documents[19]. However, the training data in these two studies is not made publicly available." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 108, + 556, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 108, + 556, + 335 + ], + "spans": [ + { + "bbox": [ + 313, + 108, + 556, + 335 + ], + "type": "text", + "content": "The scarcity of interleaved image-text data has been addressed by the introduction of two datasets: MMC4 [45] and OBELICS [19]. MMC4 builds upon the text-only C4 corpus [31] by integrating images into text passages using CLIP features. In contrast, OBELICS emphasizes comprehensive filtering strategies and preserves the original structure of web pages. These datasets present a diverse and extensive interleaved image-text corpus for multimodal language models. Despite their depth and breadth, they exhibit notable deficiencies in document quality, such as weaker completeness and image-text coherence (cf., Table 1). Additionally, as depicted in Figure 2, their documents lack comprehensiveness in the image modality, with most documents containing only one or two images. To address this, we propose CoMM that focuses explicitly on coherence-rich scenarios with more image illustrations. CoMM greatly enhances the baseline's [4] performance (in Table 2), highlighting the critical role of a high-quality coherent interleaved image-text training corpus." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 335, + 556, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 335, + 556, + 479 + ], + "spans": [ + { + "bbox": [ + 313, + 335, + 556, + 479 + ], + "type": "text", + "content": "Modeling of Interleaved Image-Text Data. The impressive performance of language modeling by LLM motivates researchers to delve deeper into the comprehension and generation capabilities of multimodal data. Flamingo [2] introduces image tokens into the language modeling process, facilitating interleaved image-text input and enabling few-shot transfer to tasks (e.g., VQA). Emu [37] leverages Stable Diffusion [32] as the image decoder, enabling the generation of image or text from interleaved image-text input. DreamLLM [6] pioneers the generation of free-form interleaved image-text outputs by modeling text and image within a unified multimodal space." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 479, + 556, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 479, + 556, + 658 + ], + "spans": [ + { + "bbox": [ + 313, + 479, + 556, + 658 + ], + "type": "text", + "content": "Recent research has increasingly focused on the intrinsic requirement of interleaved image-text modeling: capturing multimodal coherence while generating interleaved image-text content. OpenLEAF [3] underscores the importance of maintaining semantic and stylistic consistency between generated images. It also proposes a training-free baseline and employs BingChat [26] to evaluate the two types of consistency. MM-Interleaved [38] introduces the multimodal feature synchronizer to extract context-sensitive, multi-scale image features, thereby more effectively capturing multimodal context coherence and visual consistency. Nonetheless, focusing solely on the model level to meet the intrinsic multimodal coherence requirement is insufficient. Therefore, our high-quality data that strictly adhere to this requirement are essential to fully realize the potential of MLLMs." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 669, + 409, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 669, + 409, + 681 + ], + "spans": [ + { + "bbox": [ + 313, + 669, + 409, + 681 + ], + "type": "text", + "content": "3. CoMM Dataset" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "content": "The CoMM dataset is collected from high-quality interleaved image-text content sourced from various websites, focusing" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8075" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 133 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 133 + ], + "type": "text", + "content": "on instructional steps and visual stories. Besides, we apply multi-perspective quality filter strategies on text sequences, image sequences, and image-text alignment, leveraging advanced models, e.g., CLIP [29] and LLMs [25], to enhance the dataset's coherence and relevance1." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 157, + 178, + 170 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 157, + 178, + 170 + ], + "spans": [ + { + "bbox": [ + 55, + 157, + 178, + 170 + ], + "type": "text", + "content": "3.1. Dataset Construction" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 182, + 296, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 182, + 296, + 289 + ], + "spans": [ + { + "bbox": [ + 55, + 182, + 296, + 289 + ], + "type": "text", + "content": "Collection Process. As mentioned above, we gather specific interleaved image-text data to preliminarily ensure the coherence and consistency of our dataset. To achieve this, we explore websites that are likely to provide high-quality interleaved content, such as instructional steps and visual stories. We collect raw data from these sources1 and download the corresponding text and images. Initially, we apply basic data cleaning techniques, including de-duplication and filtering out images with low resolution, to refine the collected data." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 295, + 296, + 427 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 295, + 296, + 427 + ], + "spans": [ + { + "bbox": [ + 55, + 295, + 296, + 427 + ], + "type": "text", + "content": "Discarding NSFW images. To address potential ethical implications, we follow the [45] by utilizing an advanced NSFW binary image classifier [7]. This classifier is devised as a 4-layer multi-layer perceptron (MLP) that has been fully trained on the NSFW dataset from LAION-2B [34]. Specifically, it utilizes features extracted from OpenAI's CLIP ViT-L/14 model[29] as input, achieving an impressive accuracy of " + }, + { + "bbox": [ + 55, + 295, + 296, + 427 + ], + "type": "inline_equation", + "content": "97.4\\%" + }, + { + "bbox": [ + 55, + 295, + 296, + 427 + ], + "type": "text", + "content": " on the NSFW test dataset. For each image processed, we employ the classifier to predict the NSFW probability. Images yielding a probability greater than 0.1 are automatically excluded from further consideration." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 433, + 296, + 505 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 433, + 296, + 505 + ], + "spans": [ + { + "bbox": [ + 55, + 433, + 296, + 505 + ], + "type": "text", + "content": "Qualities Filter Strategy. Due to the inherent noises in the gathered raw data, the data still contains inconsistent and irrelevant content. Thus, we further employ a multiperspective filter strategy to ensure the quality and coherence of the dataset. Specifically, it involves three following components1:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 510, + 296, + 673 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 55, + 510, + 296, + 618 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 296, + 618 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 296, + 618 + ], + "type": "text", + "content": "1) Text Sequence Filter: For text sequences, the most important aspect is the contextual development and cohesive connection between them. To maintain a smooth and logical flow of the text, we employ LLMs (e.g., Llama3 [25]) to evaluate the development and coherence among text steps in a document (prompt details are in the Appendix). According to evaluation scores, we eliminate sequences that lack coherence and relevance in their content, producing a higher-quality text corpus." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 624, + 296, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 624, + 296, + 673 + ], + "spans": [ + { + "bbox": [ + 55, + 624, + 296, + 673 + ], + "type": "text", + "content": "2) Image Sequence Filter. As for image sequences, visual consistency and relevance are key factors in maintaining the quality of the content. To assess the coherence and development of image sequences, we devise a metric " + }, + { + "bbox": [ + 55, + 624, + 296, + 673 + ], + "type": "inline_equation", + "content": "\\mathcal{F}(\\cdot)" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 72, + 517, + 85 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 517, + 85 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 517, + 85 + ], + "type": "text", + "content": "through the CLIP's visual encoder [29] in Eq. (1):" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 328, + 89, + 555, + 158 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 89, + 555, + 158 + ], + "spans": [ + { + "bbox": [ + 328, + 89, + 555, + 158 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {F} \\left(\\left\\{x _ {i} \\mid 1 \\leq i \\leq N \\right\\}\\right) = \\frac {1}{N - 1} \\sum_ {i = 2} ^ {N} \\operatorname {S i m} \\left(x _ {i}, x _ {i - 1}\\right) \\\\ - \\frac {2}{(N - 1) (N - 2)} \\sum_ {i = 2} ^ {N} \\sum_ {j = 1} ^ {i - 1} \\operatorname {S i m} \\left(x _ {i}, x _ {j}\\right), \\tag {1} \\\\ \\end{array}", + "image_path": "3d5bc925712f01700e7e9dbb8122ca773bdf05d1b6eff6c93623e9c4e6e3d959.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 163, + 555, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 163, + 555, + 331 + ], + "spans": [ + { + "bbox": [ + 313, + 163, + 555, + 331 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 163, + 555, + 331 + ], + "type": "inline_equation", + "content": "\\{x_{i}|1\\leq i\\leq N\\}" + }, + { + "bbox": [ + 313, + 163, + 555, + 331 + ], + "type": "text", + "content": " denotes the sequence of images from 1 to " + }, + { + "bbox": [ + 313, + 163, + 555, + 331 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 163, + 555, + 331 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 163, + 555, + 331 + ], + "type": "inline_equation", + "content": "\\mathrm{Sim}(\\cdot ,\\cdot)" + }, + { + "bbox": [ + 313, + 163, + 555, + 331 + ], + "type": "text", + "content": " denotes the cosine similarity score between two images calculated by the CLIP's visual encoder. The first term denotes the average similarity between consecutive images, promoting smooth transitions and visual coherence within the sequence, while the second term represents the average similarity between all pairs of images, ensuring overall diversity and development. In this way, we can effectively balance coherence and diversity, ensuring that the image sequence is both visually consistent and contextually developed. By applying the image sequence filter, we can identify and eliminate sequences that lack connection or development, resulting in more polished and visually appealing images." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 331, + 556, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 331, + 556, + 533 + ], + "spans": [ + { + "bbox": [ + 313, + 331, + 556, + 533 + ], + "type": "text", + "content": "3) Image-Text Alignment Filter. This filter aims to ensure that the images and text are not only relevant to each other but also maintain a coherent narrative throughout the document. We first utilize CLIP [29] to calculate the CLIP similarity scores between images and text and remove those whose scores are less than 0.1. Since CLIP is trained on individual image-text pairs, where the text often serves as the image's caption, it tends to favor concrete descriptions (i.e., object descriptions and attributes) [20]. However, in an interleaved image-text document, the text for each step describes progress with associated images (e.g., in Figure 1(d), each text focuses on a cooking action linked to previous steps) which is associated with previous steps. Therefore, we also employ advanced model (e.g., GPT-4o [27] or Llama3 [25]) to assess and filter image-text pairs, as these models can better judge current image-text alignment by considering the previous context." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 534, + 556, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 556, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 556, + 689 + ], + "type": "text", + "content": "Preference Dataset Construction. Preference learning [8] has recently gained popularity in the post-training of LLMs, leveraging reinforcement learning [16] to train models with both positive and negative samples. In our dataset, which consists of interleaved image and text sequences, negative samples are easily created by shuffling the order of steps. These shuffles can be categorized into four types: Shuffled Text: Text in different steps is shuffled, while images remain in their original order. Shuffled Image: Images are shuffled, while the text remains in its original order. Both Shuffled Text and Image: Both text and images are shuffled independently. Shuffled Steps: Images and text within each step remain aligned, but the sequence of steps is randomized." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "content": "These negative samples are combined with original positive samples to form a preference dataset, used for further" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 66, + 702, + 273, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 702, + 273, + 714 + ], + "spans": [ + { + "bbox": [ + 66, + 702, + 273, + 714 + ], + "type": "text", + "content": "Due to space constraints, more details are left in the Appendix." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8076" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 88, + 210, + 233 + ], + "blocks": [ + { + "bbox": [ + 126, + 76, + 169, + 87 + ], + "lines": [ + { + "bbox": [ + 126, + 76, + 169, + 87 + ], + "spans": [ + { + "bbox": [ + 126, + 76, + 169, + 87 + ], + "type": "text", + "content": "(a) MMC4" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 56, + 88, + 210, + 233 + ], + "lines": [ + { + "bbox": [ + 56, + 88, + 210, + 233 + ], + "spans": [ + { + "bbox": [ + 56, + 88, + 210, + 233 + ], + "type": "image", + "image_path": "f2d152959b3929e02c5645b94f88c1dd435451c980792871ca6f56397cc30311.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 227, + 88, + 381, + 233 + ], + "blocks": [ + { + "bbox": [ + 290, + 76, + 347, + 88 + ], + "lines": [ + { + "bbox": [ + 290, + 76, + 347, + 88 + ], + "spans": [ + { + "bbox": [ + 290, + 76, + 347, + 88 + ], + "type": "text", + "content": "(b) OBELICS" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 227, + 88, + 381, + 233 + ], + "lines": [ + { + "bbox": [ + 227, + 88, + 381, + 233 + ], + "spans": [ + { + "bbox": [ + 227, + 88, + 381, + 233 + ], + "type": "image", + "image_path": "f8394213a912db355d1a5d65a47cfb9f41b2a9f47d950b9343cbe1ca9bece90c.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 235, + 555, + 257 + ], + "lines": [ + { + "bbox": [ + 55, + 235, + 555, + 257 + ], + "spans": [ + { + "bbox": [ + 55, + 235, + 555, + 257 + ], + "type": "text", + "content": "Figure 2. Visualization of the image-sentence numbers per document distribution of three datasets. The " + }, + { + "bbox": [ + 55, + 235, + 555, + 257 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 55, + 235, + 555, + 257 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 235, + 555, + 257 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 235, + 555, + 257 + ], + "type": "text", + "content": " denote the mean/median number of images/sentences in documents, respectively." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 401, + 88, + 554, + 233 + ], + "blocks": [ + { + "bbox": [ + 456, + 76, + 529, + 88 + ], + "lines": [ + { + "bbox": [ + 456, + 76, + 529, + 88 + ], + "spans": [ + { + "bbox": [ + 456, + 76, + 529, + 88 + ], + "type": "text", + "content": "(c) CoMM (Ours)" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 401, + 88, + 554, + 233 + ], + "lines": [ + { + "bbox": [ + 401, + 88, + 554, + 233 + ], + "spans": [ + { + "bbox": [ + 401, + 88, + 554, + 233 + ], + "type": "image", + "image_path": "914d0e0c5f497a4ea40ee10a12bb6eee8b00f2b865a01df43907138fcd9cb5b3.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 268, + 251, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 268, + 251, + 281 + ], + "spans": [ + { + "bbox": [ + 55, + 268, + 251, + 281 + ], + "type": "text", + "content": "model training following supervised fine-tuning." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 287, + 154, + 298 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 287, + 154, + 298 + ], + "spans": [ + { + "bbox": [ + 55, + 287, + 154, + 298 + ], + "type": "text", + "content": "3.2. Dataset Statistic" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 54, + 305, + 296, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 305, + 296, + 449 + ], + "spans": [ + { + "bbox": [ + 54, + 305, + 296, + 449 + ], + "type": "text", + "content": "After filtering, our dataset contains 227K documents with 2.28M images. Further, Figure 2 illustrates the distribution of image-sentence number per document across MMC4, OBELICS, and CoMM datasets. Darker colors indicate a higher number of documents. The " + }, + { + "bbox": [ + 54, + 305, + 296, + 449 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 54, + 305, + 296, + 449 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 305, + 296, + 449 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 54, + 305, + 296, + 449 + ], + "type": "text", + "content": " represent the average and the most common number of images/sentences per document, respectively. For MMC4 and OBELICS, the distribution of images per document is highly concentrated around very low values of " + }, + { + "bbox": [ + 54, + 305, + 296, + 449 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 54, + 305, + 296, + 449 + ], + "type": "text", + "content": " (2 and 1 per document, respectively). This indicates a significant scarcity of the image modality in their multimodal data, which directly limits the models' multimodal in-context learning ability." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 449, + 296, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 449, + 296, + 593 + ], + "spans": [ + { + "bbox": [ + 55, + 449, + 296, + 593 + ], + "type": "text", + "content": "Examining Figure 2(a) and Figure 2(b) distributions, it is evident that the number of images does not significantly change with an increase in the number of sentences. However, in practical scenarios, longer documents are more likely to contain a greater number of illustrations, which aids in modeling longer multimodal contexts. In contrast, CoMM features a larger quantity of images per document (" + }, + { + "bbox": [ + 55, + 449, + 296, + 593 + ], + "type": "inline_equation", + "content": "M = 4" + }, + { + "bbox": [ + 55, + 449, + 296, + 593 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 449, + 296, + 593 + ], + "type": "inline_equation", + "content": "\\mu = 10.1" + }, + { + "bbox": [ + 55, + 449, + 296, + 593 + ], + "type": "text", + "content": "), which significantly enhances the dataset's image modality while focusing on multimodal coherence. Furthermore, CoMM exhibits a trend where longer documents tend to include more image illustrations, thereby facilitating the modeling of longer multimodal contexts." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 604, + 296, + 617 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 604, + 296, + 617 + ], + "spans": [ + { + "bbox": [ + 55, + 604, + 296, + 617 + ], + "type": "text", + "content": "4. Interleaved Generation Task and Benchmark" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 623, + 162, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 623, + 162, + 635 + ], + "spans": [ + { + "bbox": [ + 55, + 623, + 162, + 635 + ], + "type": "text", + "content": "4.1. Task Formulation" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 641, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 296, + 713 + ], + "type": "text", + "content": "In this section, we propose four benchmark tasks to evaluate MLLMs' interleaved generation capabilities. For ease of presentation, we utilize " + }, + { + "bbox": [ + 55, + 641, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\{x_{i}|1\\leq i\\leq N\\}" + }, + { + "bbox": [ + 55, + 641, + 296, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 641, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\{t_i|1\\leq i\\leq" + }, + { + "bbox": [ + 55, + 641, + 296, + 713 + ], + "type": "inline_equation", + "content": "N\\}" + }, + { + "bbox": [ + 55, + 641, + 296, + 713 + ], + "type": "text", + "content": " to represent the image sequences and text sequences, respectively. Besides, we use " + }, + { + "bbox": [ + 55, + 641, + 296, + 713 + ], + "type": "inline_equation", + "content": "t_\\mathrm{prompt}" + }, + { + "bbox": [ + 55, + 641, + 296, + 713 + ], + "type": "text", + "content": " as the text prompt, and " + }, + { + "bbox": [ + 55, + 641, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\theta}(\\cdot)" + }, + { + "bbox": [ + 55, + 641, + 296, + 713 + ], + "type": "text", + "content": " as the model with parameter " + }, + { + "bbox": [ + 55, + 641, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 55, + 641, + 296, + 713 + ], + "type": "text", + "content": ". The detailed" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 268, + 497, + 279 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 268, + 497, + 279 + ], + "spans": [ + { + "bbox": [ + 313, + 268, + 497, + 279 + ], + "type": "text", + "content": "formulations of tasks are outlined as follows:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 280, + 554, + 340 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 280, + 554, + 340 + ], + "spans": [ + { + "bbox": [ + 313, + 280, + 554, + 340 + ], + "type": "text", + "content": "Task 1: Image-to-Text Sequence Generation. This task evaluates the model's capability to generate coherent and contextually appropriate narratives based on sequential images. The objective is to predict the corresponding text sequences from the given image sequences:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 354, + 348, + 555, + 361 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 348, + 555, + 361 + ], + "spans": [ + { + "bbox": [ + 354, + 348, + 555, + 361 + ], + "type": "interline_equation", + "content": "\\{t _ {i} | 1 \\leq i \\leq N \\} = \\mathcal {F} _ {\\theta} \\left(t _ {\\text {p r o m p t}}, x _ {1}, \\dots , x _ {N}\\right). \\tag {2}", + "image_path": "24f970338da9292a7b381495f0cd59e4e1261cc9e7fd135d16f2cf83f4024613.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 369, + 555, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 369, + 555, + 429 + ], + "spans": [ + { + "bbox": [ + 313, + 369, + 555, + 429 + ], + "type": "text", + "content": "Task 2: Text-to-Image Sequence Generation. This task assesses the model's proficiency in generating image sequences that accurately represent provided textual descriptions. The goal is to predict the corresponding image sequences from the given text sequences:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 354, + 437, + 555, + 449 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 437, + 555, + 449 + ], + "spans": [ + { + "bbox": [ + 354, + 437, + 555, + 449 + ], + "type": "interline_equation", + "content": "\\{x _ {i} \\mid 1 \\leq i \\leq N \\} = \\mathcal {F} _ {\\theta} \\left(t _ {\\text {p r o m p t}}, t _ {1}, \\dots , t _ {N}\\right). \\tag {3}", + "image_path": "5a6d6ae34f9f5f5f2398a893a342359c0def4c01d19a651fa344d1ec1a571888.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 457, + 554, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 457, + 554, + 518 + ], + "spans": [ + { + "bbox": [ + 313, + 457, + 554, + 518 + ], + "type": "text", + "content": "Task 3: Interleaved Image-Text Content Continuation. It measures the model's ability to seamlessly continue content in an interleaved image-text format. The model needs to predict the outputs of the remaining image-text steps based on the interleaved image-text content in the previous " + }, + { + "bbox": [ + 313, + 457, + 554, + 518 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 457, + 554, + 518 + ], + "type": "text", + "content": " steps:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 337, + 525, + 555, + 538 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 337, + 525, + 555, + 538 + ], + "spans": [ + { + "bbox": [ + 337, + 525, + 555, + 538 + ], + "type": "interline_equation", + "content": "\\left\\{t _ {i}, x _ {i} \\mid k < i \\leq N \\right\\} = \\mathcal {F} _ {\\theta} \\left(t _ {\\text {p r o m p t}}, t _ {1}, x _ {1}, \\dots , t _ {k}, x _ {k}\\right). \\tag {4}", + "image_path": "7b394d871c0fd40d2561e7d2500fac9a94505d9bf169fb6c787943fffdb9102d.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 546, + 555, + 618 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 546, + 555, + 618 + ], + "spans": [ + { + "bbox": [ + 313, + 546, + 555, + 618 + ], + "type": "text", + "content": "Task 4: Question-based Interleaved Image-Text Generation. This challenging task requires the model to generate interleaved content based on a given question, testing its understanding and generative reasoning abilities across modalities. The model needs to predict the outcomes of all image-text steps based on the given question:" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 368, + 626, + 555, + 639 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 368, + 626, + 555, + 639 + ], + "spans": [ + { + "bbox": [ + 368, + 626, + 555, + 639 + ], + "type": "interline_equation", + "content": "\\left\\{t _ {i}, x _ {i} \\mid 1 \\leq i \\leq N \\right\\} = \\mathcal {F} _ {\\theta} \\left(t _ {\\text {p r o m p t}}\\right). \\tag {5}", + "image_path": "6cce7918dc8e1fa359566ed03c638eda4407cf2bfba037061aec7fbf43d55e32.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 647, + 422, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 647, + 422, + 658 + ], + "spans": [ + { + "bbox": [ + 313, + 647, + 422, + 658 + ], + "type": "text", + "content": "4.2. Evaluation Metric" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "text", + "content": "This section lists the evaluation metrics for assessing the quality of generated results. Our evaluation framework1 measures textual accuracy, visual fidelity, and the coherence of multi-modal outputs." + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8077" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 216 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 216 + ], + "type": "text", + "content": "Evaluation for Text Sequence Generation. Generating narratives for a series of images shares similarities with the established task of image captioning. Consequently, we directly apply traditional image captioning metrics like ROUGE [21] and METEOR [5] at both the step and document levels. Additionally, we propose Illustration Relevance Score (IRS) to assess the relevance between image illustration and text context in the final document. Specifically, we simulate the human document creation process by first generating detailed image descriptions from the text context and then assessing the consistency between these descriptions and corresponding images using GPT-4o [27]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 217, + 296, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 217, + 296, + 335 + ], + "spans": [ + { + "bbox": [ + 55, + 217, + 296, + 335 + ], + "type": "text", + "content": "Evaluation for Image Sequence Generation. We evaluate the quality of image illustrations on four key assessment aspects including image quality, style consistency between generated images, reconstruction on referenced images, and IRS. To evaluate image quality, we use Fréchet Inception Distance (FID) [13] with ground truth reference and Inception Score (IS) [33]. For assessing style consistency, we employ the Structural Similarity Index Measure (SSIM), and for ground truth reconstruction evaluation, we adopt the Peak Signal-to-Noise Ratio (PSNR)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 337, + 296, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 337, + 296, + 456 + ], + "spans": [ + { + "bbox": [ + 55, + 337, + 296, + 456 + ], + "type": "text", + "content": "Evaluation for Interleaved Image-Text Generation. We use MLLM (GPT-4o [27]) for a comprehensive assessment of linguistic and visual coherence. GPT-4o evaluates four aspects: image sequence coherence, image quality, document completeness, and IRS. For image sequence coherence, we check style, entity consistency, and content trend alignment between image and text sequence. Image quality is evaluated on authenticity, integrity, clarity, and aesthetics. Document completeness is scored based on how thoroughly the content covers the topic." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 469, + 137, + 483 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 469, + 137, + 483 + ], + "spans": [ + { + "bbox": [ + 55, + 469, + 137, + 483 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 490, + 217, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 490, + 217, + 502 + ], + "spans": [ + { + "bbox": [ + 55, + 490, + 217, + 502 + ], + "type": "text", + "content": "5.1. Datasets Quality Comparison" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 509, + 295, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 509, + 295, + 675 + ], + "spans": [ + { + "bbox": [ + 55, + 509, + 295, + 675 + ], + "type": "text", + "content": "Settings. In this section, we compare the quality of our dataset with previous interleaved image-text datasets (e.g., MMC4 [45] and OBELICS [19]) focusing on text sequences, image sequences, and image-text alignment. We utilize powerful LLMs (e.g., Llama3 [25] and GPT-4o [27]) to assess the text development, text completeness, and image-text alignment of these datasets. Each metric is scored on a scale of 0 to 10, with detailed definitions and prompts provided in the Appendix. For MMC4 and OBELICS, we randomly sample 5,000 documents from their data. We evaluate our entire dataset with Llama3 and randomly sample 5,000 documents for assessment by GPT-4o. To evaluate image sequence quality, we employ the metric " + }, + { + "bbox": [ + 55, + 509, + 295, + 675 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 55, + 509, + 295, + 675 + ], + "type": "text", + "content": " as cited in Eq. (1)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "Quantitative Results. As shown in Table 1, our dataset surpasses MMC4 and OBELICS across all evaluated dimensions. Notably, the score difference in development," + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 317, + 70, + 553, + 129 + ], + "blocks": [ + { + "bbox": [ + 317, + 70, + 553, + 129 + ], + "lines": [ + { + "bbox": [ + 317, + 70, + 553, + 129 + ], + "spans": [ + { + "bbox": [ + 317, + 70, + 553, + 129 + ], + "type": "table", + "html": "
DatasetsDLP L.3 / G.4oCPL L.3 / G.4oITA L.3 / G.4oImgS
MMC4 [45]5.56 / 4.756.28 / 5.126.53 / 4.660.21
OBELICS [19]5.60 / 5.976.51 / 5.885.00 / 3.811.00
CoMM (Ours)6.93 / 7.647.44 / 7.077.46 / 8.914.27
", + "image_path": "9d030ec86f499745a453f06fcb81bcd7d6e309e586021bc507143b6efd96a2d0.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 137, + 555, + 182 + ], + "lines": [ + { + "bbox": [ + 313, + 137, + 555, + 182 + ], + "spans": [ + { + "bbox": [ + 313, + 137, + 555, + 182 + ], + "type": "text", + "content": "Table 1. Quality comparison of interleaved image-text datasets. \"DLP\" stands for Development, \"CPL\" signifies Completeness, \"ITA\" represents Image-Text Alignment, \"ImgS\" refers to Image Sequence, \"L.3\" denotes Llama3 [25], and \"G.4o\" is GPT-4o [27]." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 204, + 555, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 204, + 555, + 276 + ], + "spans": [ + { + "bbox": [ + 313, + 204, + 555, + 276 + ], + "type": "text", + "content": "completeness, and image-text alignment is more pronounced when evaluated by GPT-4o, a more powerful MLLM. Furthermore, the image sequence metric underscores the superior quality of our data, reflecting higher scores compared to MMC4 and OBELICS. These results affirm the high quality of our dataset in interleaved image-text data." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 284, + 519, + 297 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 284, + 519, + 297 + ], + "spans": [ + { + "bbox": [ + 313, + 284, + 519, + 297 + ], + "type": "text", + "content": "5.2. In-Context Multimodal Understanding" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 302, + 556, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 302, + 556, + 458 + ], + "spans": [ + { + "bbox": [ + 313, + 302, + 556, + 458 + ], + "type": "text", + "content": "Setting. To further demonstrate the high-quality and validate the efficacy of our dataset in promoting in-context understanding, we conduct experiments using few-shot learning settings on several downstream tasks (COCO [22], Flickr30k [43], VQAv2 [11], OKVQA [24], TextVQA [35], VizWiz [12], and Hateful Memes [17]). Few-shot learning is critical for assessing a model's ability of in-context understanding. In our experiments, few-shot scenarios are simulated by providing the model with a small number of example queries and their corresponding answers before it is evaluated on new queries. We consider various shot settings (0, 4, 8, 16, and 32) to understand the model's performance across different levels of supervision." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 459, + 556, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 459, + 556, + 650 + ], + "spans": [ + { + "bbox": [ + 313, + 459, + 556, + 650 + ], + "type": "text", + "content": "Quantitative Results. Table 2 presents the results for the baseline model and its enhancements with MMC4, OBELICS, and our proposed dataset. As observed from Table 2, our method consistently outperforms the performance of MMC4 and OBELICS across all shot settings, demonstrating superior generalization and in-context learning capabilities. Particularly noticeable improvements are seen in the COCO and TextVQA datasets, where our method exceeds the baseline by significant margins in nearly all shot settings. This superior performance underlines the strength of our dataset in enabling models to better learn from context and apply learned knowledge to unseen instances. Furthermore, our dataset surpasses the baseline in all tasks with more shot settings, maintaining its advantage with longer context few-shot data, as shown in the 16-shot and 32-shot scenarios." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 658, + 474, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 658, + 474, + 670 + ], + "spans": [ + { + "bbox": [ + 313, + 658, + 474, + 670 + ], + "type": "text", + "content": "5.3. Interleaved Generation Tasks" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 677, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 556, + 713 + ], + "type": "text", + "content": "Baselines. This section introduces baselines developed from several state-of-the-art approaches, i.e., MiniGPT-5 [44], SEED-Llama [9], and Emu2 [36], to assess their effective" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8078" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 93, + 69, + 518, + 306 + ], + "blocks": [ + { + "bbox": [ + 93, + 69, + 518, + 306 + ], + "lines": [ + { + "bbox": [ + 93, + 69, + 518, + 306 + ], + "spans": [ + { + "bbox": [ + 93, + 69, + 518, + 306 + ], + "type": "table", + "html": "
ShotCOCOFlickr30kVQAv2OKVQATextVQAVizWizHateful Memes
Baseline79.559.552.737.824.227.551.6
+ MMC4 [45]097.167.045.829.823.415.054.6
+ OBELICS [19]83.551.550.938.925.621.953.8
+ CoMM (Ours)100.369.852.641.333.523.654.8
Baseline89.065.854.840.128.234.154.0
+ MMC4 [45]4103.368.849.536.728.724.455.0
+ OBELICS [19]103.765.253.243.031.429.955.7
+ CoMM (Ours)107.073.253.642.935.730.155.2
Baseline96.362.954.841.129.138.554.7
+ MMC4 [45]8103.667.950.539.028.832.751.3
+ OBELICS [19]101.963.953.744.532.237.251.2
+ CoMM (Ours)108.074.854.244.535.437.952.5
Baseline98.862.854.342.727.342.553.9
+ MMC4 [45]16102.965.550.039.126.436.957.2
+ OBELICS [19]100.858.053.145.330.242.457.8
+ CoMM (Ours)109.371.554.246.935.843.559.9
Baseline99.561.353.342.423.844.053.8
+ MMC4 [45]3294.956.448.335.620.037.749.2
+ OBELICS [19]96.553.152.043.626.844.556.1
+ CoMM (Ours)111.970.754.046.634.945.757.3
", + "image_path": "a05b89320f9fe057cd0f96aca50ed3a364ee274df6e565a1504723d7c6f6878a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 66, + 369, + 545, + 506 + ], + "blocks": [ + { + "bbox": [ + 55, + 314, + 555, + 360 + ], + "lines": [ + { + "bbox": [ + 55, + 314, + 555, + 360 + ], + "spans": [ + { + "bbox": [ + 55, + 314, + 555, + 360 + ], + "type": "text", + "content": "Table 2. Performance comparison of different datasets, with the baseline model Open-Flamingo 9B [4]. Evaluations were in an open-ended setting for VQA tasks with random in-context examples. (Task, Metric, Query Split): (COCO [22], CIDEr, test), (Flickr30k [43], CIDEr, test (Karpathy)), (VQAv2 [11], VQA accuracy, testdev), (OKVQA [24], VQA accuracy, val), (TextVQA [35], VQA accuracy, val), (VizWiz [12], VQA accuracy, testdev), (Hateful Memes [17], ROC-AUC, test seen)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 66, + 369, + 545, + 506 + ], + "lines": [ + { + "bbox": [ + 66, + 369, + 545, + 506 + ], + "spans": [ + { + "bbox": [ + 66, + 369, + 545, + 506 + ], + "type": "table", + "html": "
MethodsSizeI2T Sequence Generation (Task1)StyleContinuation Generation (Task3)
M.S.R.S.M.W.R.W.IRSEntityTrendCPL.ImgQIRS
MiniGPT-5 [44]7B4.608.649.1010.677.395.585.215.246.336.362.56
SEED-Llama [9]8B2.296.344.207.336.996.285.845.726.286.552.92
SEED-Llama [9]14B3.168.125.718.178.136.686.226.136.666.673.23
Emu2 [36]33B3.658.417.438.625.018.227.997.978.498.622.42
MethodsSizeT2I Sequence Generation (Task2)StyleQuestion-based Generation (Task4)
FID ↓ISSSIMPSNRIRSEntityTrendCPL.ImgQIRS
MiniGPT-5 [44]7B56.517.6020.667.575.485.655.205.255.816.152.71
SEED-Llama [9]8B57.968.4020.537.875.277.556.816.155.136.361.46
SEED-Llama [9]14B66.2311.0620.838.126.247.516.616.306.136.662.50
Emu2 [36]33B60.456.3420.967.914.178.417.567.637.547.592.02
", + "image_path": "5609331f721975aa27d5a22255241d93c31a870655d7881264bedded52d07bf3.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 514, + 555, + 559 + ], + "lines": [ + { + "bbox": [ + 55, + 514, + 555, + 559 + ], + "spans": [ + { + "bbox": [ + 55, + 514, + 555, + 559 + ], + "type": "text", + "content": "Table 3. Comparison of performance among MiniGPT-5 [44], SEED-Llama [9], and Emu2 [37] on the four generation tasks. \"M.S.\" and \"R.S.\" represent METEOR [5] and ROUGE_L [21] for step-by-step evaluation, while \"M.W.\" and \"R.W.\" correspond to METEOR and ROUGE_L for whole document evaluation. \"CPL.\" stands for Completeness, \"ImgQ\" indicates Image Quality, \"IRS\" means Illustration Relevance Score, and \"↓\" denotes that lower values are better." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 580, + 295, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 580, + 295, + 652 + ], + "spans": [ + { + "bbox": [ + 55, + 580, + 295, + 652 + ], + "type": "text", + "content": "ness in generating interleaved image-text content. Details of these models are thoroughly provided in the Appendix. To construct the baselines, we fine-tune each of these models using our CoMM dataset. We employ a rigorous evaluation, assessing them on benchmark tasks involving interleaved image-text generation and understanding." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 653, + 297, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 297, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 297, + 715 + ], + "type": "text", + "content": "Quantitative Results. As shown in Table 3, the quantitative results clearly illustrate the varying performance of MiniGPT-5, SEED-Llama, and Emu2 across the four interleaved generation tasks. For I2T Sequence Generation (Task1), MiniGPT-5 achieves the highest scores in ME" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 580, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 580, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 580, + 556, + 713 + ], + "type": "text", + "content": "TEOR, ROUGE_L, but falls short in Illustration Relevance Score (IRS) where SEED-Llama (14B) leads with a score of 8.13. For T2I Sequence Generation (Task2), MiniGPT-5 excels with the lowest FID score of 56.51, while SEED-Llama (14B) surpasses IS, SSIM, PSNR, and IRS metrics, suggesting stronger visual fidelity and semantic similarity. Emu2 tops in Continuation Generation (Task3) with the highest scores in Style, Entity, Trend, Completeness, and Image Quality. In Question-based Generation (Task4), Emu2 dominates in Style, Entity, Trend, Completeness, and Image Quality, while MiniGPT-5 manages to secure the lead in IRS." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8079" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 72, + 547, + 392 + ], + "blocks": [ + { + "bbox": [ + 59, + 72, + 547, + 392 + ], + "lines": [ + { + "bbox": [ + 59, + 72, + 547, + 392 + ], + "spans": [ + { + "bbox": [ + 59, + 72, + 547, + 392 + ], + "type": "image", + "image_path": "421bd8ca184d99258631bc823a2fba0d2780d95e00f20dd6384697c50f098283.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 392, + 555, + 415 + ], + "lines": [ + { + "bbox": [ + 55, + 392, + 555, + 415 + ], + "spans": [ + { + "bbox": [ + 55, + 392, + 555, + 415 + ], + "type": "text", + "content": "Figure 3. Visualization of interleaved image-text content generation from SEED-Llama [9] (Top) and MiniGPT-5 [44] (Bottom) separately trained by MMC4 [45] and CoMM [Ours]." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 426, + 295, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 426, + 295, + 474 + ], + "spans": [ + { + "bbox": [ + 54, + 426, + 295, + 474 + ], + "type": "text", + "content": "Overall, these results highlight the strengths and weaknesses of each model in handling different aspects of interleaved image-text generation tasks, providing valuable insights for future research and model improvements." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 474, + 295, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 474, + 295, + 581 + ], + "spans": [ + { + "bbox": [ + 54, + 474, + 295, + 581 + ], + "type": "text", + "content": "Preference Dataset Training Results. With SEED-Llama [9] unifying image to discrete token prediction modeling, we can directly apply the mainstream preference learning algorithm DPO [30] to train SEED-Llama on our preference dataset following supervised fine-tuning on our dataset. As shown in Table 4, performance further improves on the image-to-text sequence generation task after training with our preference dataset. Additional results on preference dataset training are available in the Appendix." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 582, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 582, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 582, + 295, + 714 + ], + "type": "text", + "content": "Multimodal Generation Qualitative Analysis. As illustrated in Figure 1 and Figure 3, the interleaved image-text content generated by models [9, 36, 44] trained on MMC4 [45] shows a noticeable difference when compared to the content generated by models trained on our datasets. Our models produce content that is more consistent with the previous input context (e.g., textual coherence and entity consistency). By ensuring higher coherence, consistency, and alignment, CoMM sets a new standard for training datasets in the realm of multimodal large language models. More visualizations are available in the Appendix." + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 317, + 423, + 550, + 494 + ], + "blocks": [ + { + "bbox": [ + 317, + 423, + 550, + 494 + ], + "lines": [ + { + "bbox": [ + 317, + 423, + 550, + 494 + ], + "spans": [ + { + "bbox": [ + 317, + 423, + 550, + 494 + ], + "type": "table", + "html": "
MethodsSizeI2T Sequence Generation (Task1)
M.S.R.S.M.W.R.W.IRS
SEED-Llama [9]8B2.296.344.207.336.99
+ DPO [30]8B4.098.836.9510.247.26
SEED-Llama [9]14B3.168.125.718.178.13
+ DPO [30]14B4.6210.158.0810.308.18
", + "image_path": "54efeb62378bfd9fa11e8e0b32cfa569f4cf31472b5a7d83558e64f404ea848e.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 497, + 555, + 552 + ], + "lines": [ + { + "bbox": [ + 313, + 497, + 555, + 552 + ], + "spans": [ + { + "bbox": [ + 313, + 497, + 555, + 552 + ], + "type": "text", + "content": "Table 4. Performance results of SEED-Llama [9] trained by DPO [30] in our preference dataset. \"M.S.\" and \"R.S.\" represent METEOR [5] and ROUGE_L [21] for step-by-step evaluation, while \"M.W.\" and \"R.W.\" correspond to METEOR and ROUGE_L for whole document evaluation." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 561, + 388, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 561, + 388, + 574 + ], + "spans": [ + { + "bbox": [ + 313, + 561, + 388, + 574 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 312, + 582, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 582, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 312, + 582, + 556, + 715 + ], + "type": "text", + "content": "In this paper, we present CoMM, a high-quality interleaved image-text dataset designed to address the limitations of existing multimodal large language models (MLLMs) by ensuring superior text-image alignment and entity consistency. Through diverse interleaved image-text content, CoMM significantly enhances MLLMs' in-context learning capabilities, as demonstrated by extensive few-shot experiments. Additionally, we introduce four novel tasks and a comprehensive evaluation framework to assess interleaved image-text generation abilities, setting new benchmarks for advancing multimodal understanding and generation in MLLMs." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8080" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "type": "text", + "content": "Acknowledgment" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 174 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 174 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 174 + ], + "type": "text", + "content": "This work was supported by the Hong Kong SAR RGC Early Career Scheme (26208924), the National Natural Science Foundation of China Young Scholar Fund (62402408), and the HKUST Sports Science and Technology Research Grant (SSTRG24EG04), ACCESS - AI Chip Center for Emerging Smart Systems, sponsored by InnoHK funding, Hong Kong SAR." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 184, + 114, + 197 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 184, + 114, + 197 + ], + "spans": [ + { + "bbox": [ + 56, + 184, + 114, + 197 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 204, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 204, + 296, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 204, + 296, + 248 + ], + "spans": [ + { + "bbox": [ + 61, + 204, + 296, + 248 + ], + "type": "text", + "content": "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florecia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 250, + 296, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 250, + 296, + 303 + ], + "spans": [ + { + "bbox": [ + 61, + 250, + 296, + 303 + ], + "type": "text", + "content": "[2] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. NeurIPS, 2022. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 304, + 296, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 304, + 296, + 348 + ], + "spans": [ + { + "bbox": [ + 62, + 304, + 296, + 348 + ], + "type": "text", + "content": "[3] Jie An, Zhengyuan Yang, Linjie Li, Jianfeng Wang, Kevin Lin, Zicheng Liu, Lijuan Wang, and Jiebo Luo. Openleaf: Open-domain interleaved image-text generation and evaluation. arXiv preprint arXiv:2310.07749, 2023. 1, 2, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 350, + 296, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 350, + 296, + 413 + ], + "spans": [ + { + "bbox": [ + 62, + 350, + 296, + 413 + ], + "type": "text", + "content": "[4] Anas Awadalla, Irena Gao, Josh Gardner, Jack Hessel, Yusuf Hanafy, Wanrong Zhu, Kalyani Marathe, Yonatan Bitton, Samir Gadre, Shiori Sagawa, et al. Openflamingo: An opensource framework for training large autoregressive vision-language models. arXiv preprint arXiv:2308.01390, 2023. 3, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 415, + 296, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 415, + 296, + 458 + ], + "spans": [ + { + "bbox": [ + 62, + 415, + 296, + 458 + ], + "type": "text", + "content": "[5] Satanjeev Banerjee and Alon Lavie. Meteor: An automatic metric for mt evaluation with improved correlation with human judgments. In ACL Workshop, pages 65-72, 2005. 3, 6, 7, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 460, + 296, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 460, + 296, + 513 + ], + "spans": [ + { + "bbox": [ + 62, + 460, + 296, + 513 + ], + "type": "text", + "content": "[6] Runpei Dong, Chunrui Han, Yuang Peng, Zekun Qi, Zheng Ge, Jinrong Yang, Liang Zhao, Jianjian Sun, Hongyu Zhou, Haoran Wei, et al. Dreamllm: Synergistic multimodal comprehension and creation. arXiv preprint arXiv:2309.11499, 2023. 1, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 514, + 296, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 514, + 296, + 568 + ], + "spans": [ + { + "bbox": [ + 62, + 514, + 296, + 568 + ], + "type": "text", + "content": "[7] Samir Yitzhak Gadre, Gabriel Ilharco, Alex Fang, Jonathan Hayase, Georgios Smyrnis, Thao Nguyen, Ryan Marten, Mitchell Wortsman, Dhruba Ghosh, Jieyu Zhang, et al. Datacomp: In search of the next generation of multimodal datasets. NeurIPS, 36, 2023. 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 570, + 296, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 570, + 296, + 624 + ], + "spans": [ + { + "bbox": [ + 62, + 570, + 296, + 624 + ], + "type": "text", + "content": "[8] Bofei Gao, Feifan Song, Yibo Miao, Zefan Cai, Zhe Yang, Liang Chen, Helan Hu, Runxin Xu, Qingxiu Dong, Ce Zheng, et al. Towards a unified view of preference learning for large language models: A survey. arXiv preprint arXiv:2409.02795, 2024. 3, 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 625, + 296, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 625, + 296, + 668 + ], + "spans": [ + { + "bbox": [ + 62, + 625, + 296, + 668 + ], + "type": "text", + "content": "[9] Yuying Ge, Sijie Zhao, Ziyun Zeng, Yixiao Ge, Chen Li, Xintao Wang, and Ying Shan. Making llama see and draw with seed tokenizer. arXiv preprint arXiv:2310.01218, 2023. 6, 7, 8, 1, 3, 4, 9, 10, 11, 12, 13" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 670, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 296, + 713 + ], + "type": "text", + "content": "[10] Yuying Ge, Sijie Zhao, Jinguo Zhu, Yixiao Ge, Kun Yi, Lin Song, Chen Li, Xiaohan Ding, and Ying Shan. Seed-x: Multimodal models with unified multi-granularity comprehension and generation. arXiv preprint arXiv:2404.14396, 2024. 1" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 316, + 73, + 555, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 555, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 555, + 116 + ], + "type": "text", + "content": "[11] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Bartra, and Devi Parikh. Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In CVPR, 2017. 3, 6, 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 118, + 555, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 118, + 555, + 162 + ], + "spans": [ + { + "bbox": [ + 317, + 118, + 555, + 162 + ], + "type": "text", + "content": "[12] Danna Gurari, Qing Li, Abigale J Stangl, Anhong Guo, Chi Lin, Kristen Grauman, Jiebo Luo, and Jeffrey P Bigham. Vizwiz grand challenge: Answering visual questions from blind people. In CVPR, 2018. 3, 6, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 163, + 555, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 163, + 555, + 206 + ], + "spans": [ + { + "bbox": [ + 317, + 163, + 555, + 206 + ], + "type": "text", + "content": "[13] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. NeurIPS, 2017. 3, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 208, + 555, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 208, + 555, + 251 + ], + "spans": [ + { + "bbox": [ + 316, + 208, + 555, + 251 + ], + "type": "text", + "content": "[14] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 253, + 554, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 253, + 554, + 297 + ], + "spans": [ + { + "bbox": [ + 316, + 253, + 554, + 297 + ], + "type": "text", + "content": "[15] S Huang, L Dong, W Wang, Y Hao, S Singhal, S Ma, T Lv, L Cui, OK Mohammed, Q Liu, et al. Language is not all you need: aligning perception with language models. 2023. arXiv preprint arXiv:2302.14045. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 298, + 555, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 298, + 555, + 331 + ], + "spans": [ + { + "bbox": [ + 316, + 298, + 555, + 331 + ], + "type": "text", + "content": "[16] Timo Kaufmann, Paul Weng, Viktor Bengs, and Eyke Hüllermieier. A survey of reinforcement learning from human feedback. arXiv preprint arXiv:2312.14925, 2023. 3, 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 332, + 554, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 332, + 554, + 376 + ], + "spans": [ + { + "bbox": [ + 316, + 332, + 554, + 376 + ], + "type": "text", + "content": "[17] Douwe Kiela, Hamed Firooz, Aravind Mohan, Vedanuj Goswami, Amanpreet Singh, Pratik Ringshia, and Davide Testuggine. The hateful memes challenge: Detecting hate speech in multimodal memes. NeurIPS, 2020. 3, 6, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 377, + 555, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 377, + 555, + 408 + ], + "spans": [ + { + "bbox": [ + 316, + 377, + 555, + 408 + ], + "type": "text", + "content": "[18] Jing Yu Koh, Daniel Fried, and Russ R Salakhutdinov. Generating images with multimodal language models. NeurIPS, 2024. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 411, + 555, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 411, + 555, + 464 + ], + "spans": [ + { + "bbox": [ + 316, + 411, + 555, + 464 + ], + "type": "text", + "content": "[19] Hugo Laurençon, Lucile Saulnier, Léo Tronchon, Stas Bekman, Amanpreet Singh, Anton Lozhkov, Thomas Wang, Siddharth Karamcheti, Alexander Rush, Douwe Kiela, et al. Obelics: An open web-scale filtered dataset of interleaved image-text documents. NeurIPS, 2023. 3, 6, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 467, + 555, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 467, + 555, + 520 + ], + "spans": [ + { + "bbox": [ + 317, + 467, + 555, + 520 + ], + "type": "text", + "content": "[20] Samuel Lavoie, Polina Kirichenko, Mark Ibrahim, Mahmoud Assran, Andrew Gordon Wildon, Aaron Courville, and Nicolas Ballas. Modeling caption diversity in contrastive vision-language pretraining. arXiv preprint arXiv:2405.00740, 2024. 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 523, + 555, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 523, + 555, + 555 + ], + "spans": [ + { + "bbox": [ + 316, + 523, + 555, + 555 + ], + "type": "text", + "content": "[21] Chin-Yew Lin. Rouge: A package for automatic evaluation of summaries. In Text summarization branches out, 2004. 3, 6, 7, 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 557, + 555, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 557, + 555, + 599 + ], + "spans": [ + { + "bbox": [ + 317, + 557, + 555, + 599 + ], + "type": "text", + "content": "[22] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dolkar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 3, 6, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 601, + 554, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 601, + 554, + 644 + ], + "spans": [ + { + "bbox": [ + 316, + 601, + 554, + 644 + ], + "type": "text", + "content": "[23] Chang Liu, Haoning Wu, Yujie Zhong, Xiaoyun Zhang, and Weidi Xie. Intelligent grimm-open-ended visual storytelling via latent diffusion models. arXiv preprint arXiv:2306.00973, 2023. 1, 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 647, + 555, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 555, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 555, + 689 + ], + "type": "text", + "content": "[24] Kenneth Marino, Mohammad Rastegari, Ali Farhadi, and Roozbeh Mottaghi. Ok-vqa: A visual question answering benchmark requiring external knowledge. In CVPR, 2019. 3, 6, 7" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 691, + 555, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 555, + 713 + ], + "type": "text", + "content": "[25] Meta. https://llama.meta.com/llama3, 2024.2, 4,6,1" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "8081" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 297, + 715 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 56, + 73, + 297, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 297, + 105 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 297, + 105 + ], + "type": "text", + "content": "[26] Microsoft. Bingchat. https://www.microsoft.com/en-us/edge/features/bing-chat, 2023. Accessed: 2023-10-09. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 106, + 296, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 106, + 296, + 128 + ], + "spans": [ + { + "bbox": [ + 56, + 106, + 296, + 128 + ], + "type": "text", + "content": "[27] OpenAI. https://openai.com/index/hello-gpt-4o/, 2024.3,4,6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 129, + 296, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 296, + 172 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 296, + 172 + ], + "type": "text", + "content": "[28] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 173, + 296, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 173, + 296, + 227 + ], + "spans": [ + { + "bbox": [ + 56, + 173, + 296, + 227 + ], + "type": "text", + "content": "[29] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 2, 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 228, + 296, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 228, + 296, + 271 + ], + "spans": [ + { + "bbox": [ + 56, + 228, + 296, + 271 + ], + "type": "text", + "content": "[30] Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. Direct preference optimization: Your language model is secretly a reward model. NeurIPS, 36, 2023. 8, 1, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 272, + 296, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 272, + 296, + 316 + ], + "spans": [ + { + "bbox": [ + 56, + 272, + 296, + 316 + ], + "type": "text", + "content": "[31] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. JMLR, 2020. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 316, + 296, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 316, + 296, + 348 + ], + "spans": [ + { + "bbox": [ + 56, + 316, + 296, + 348 + ], + "type": "text", + "content": "[32] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 350, + 296, + 382 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 350, + 296, + 382 + ], + "spans": [ + { + "bbox": [ + 56, + 350, + 296, + 382 + ], + "type": "text", + "content": "[33] Tim Salimans, Ian Goodfellow, Wojciech Zaremba, Vicki Cheung, Alec Radford, and Xi Chen. Improved techniques for training gans. NeurIPS, 2016. 3, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 383, + 296, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 383, + 296, + 446 + ], + "spans": [ + { + "bbox": [ + 56, + 383, + 296, + 446 + ], + "type": "text", + "content": "[34] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. NeurIPS, 35:25278–25294, 2022. 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 449, + 296, + 481 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 449, + 296, + 481 + ], + "spans": [ + { + "bbox": [ + 56, + 449, + 296, + 481 + ], + "type": "text", + "content": "[35] Amanpreet Singh, Vivek Natarajan, Meet Shah, Yu Jiang, Xinlei Chen, Dhruv Batra, Devi Parikh, and Marcus Rohrbach. Towards vqa models that can read. In CVPR, 2019. 6, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 482, + 296, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 482, + 296, + 536 + ], + "spans": [ + { + "bbox": [ + 56, + 482, + 296, + 536 + ], + "type": "text", + "content": "[36] Quan Sun, Yufeng Cui, Xiaosong Zhang, Fan Zhang, Qiying Yu, Zhengxiong Luo, Yueze Wang, Yongming Rao, Jingjing Liu, Tiejun Huang, et al. Generative multimodal models are in-context learners. arXiv preprint arXiv:2312.13286, 2023. 1, 2, 6, 7, 8, 4, 9, 10, 11, 12" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 537, + 296, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 537, + 296, + 581 + ], + "spans": [ + { + "bbox": [ + 56, + 537, + 296, + 581 + ], + "type": "text", + "content": "[37] Quan Sun, Qiying Yu, Yufeng Cui, Fan Zhang, Xiaosong Zhang, Yueze Wang, Hongcheng Gao, Jingjing Liu, Tiejun Huang, and Xinlong Wang. Emu: Generative pretraining in multimodality. In ICLR, 2023. 3, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 582, + 296, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 582, + 296, + 635 + ], + "spans": [ + { + "bbox": [ + 56, + 582, + 296, + 635 + ], + "type": "text", + "content": "[38] Changyao Tian, Xizhou Zhu, Yuwen Xiong, Weiyun Wang, Zhe Chen, Wenhai Wang, Yuntao Chen, Lewei Lu, Tong Lu, Jie Zhou, et al. Mm-interleaved: Interleaved image-text generative modeling via multi-modal feature synchronizer. arXiv preprint arXiv:2401.10208, 2024. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 636, + 296, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 636, + 296, + 690 + ], + "spans": [ + { + "bbox": [ + 56, + 636, + 296, + 690 + ], + "type": "text", + "content": "[39] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 691, + 296, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 296, + 715 + ], + "type": "text", + "content": "[40] Weihan Wang, Qingsong Lv, Wenmeng Yu, Wenyi Hong, Ji Qi, Yan Wang, Junhui Ji, Zhuoyi Yang, Lei Zhao, Xixuan" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 308 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 333, + 73, + 555, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 555, + 106 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 555, + 106 + ], + "type": "text", + "content": "Song, Jiazheng Xu, Bin Xu, Juanzi Li, Yuxiao Dong, Ming Ding, and Jie Tang. CogvIm: Visual expert for pretrained language models, 2023. 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 107, + 555, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 555, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 555, + 150 + ], + "type": "text", + "content": "[41] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. NeurIPS, 2022. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 152, + 521, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 521, + 163 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 521, + 163 + ], + "type": "text", + "content": "[42] WikiHow. https://www.wikipiow.com. 2,3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 164, + 555, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 164, + 555, + 206 + ], + "spans": [ + { + "bbox": [ + 316, + 164, + 555, + 206 + ], + "type": "text", + "content": "[43] Peter Young, Alice Lai, Micah Hodosh, and Julia Hockenmaier. From image descriptions to visual denotations: New similarity metrics for semantic inference over event descriptions. TACL, 2014. 6, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 209, + 555, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 209, + 555, + 251 + ], + "spans": [ + { + "bbox": [ + 316, + 209, + 555, + 251 + ], + "type": "text", + "content": "[44] Kaizhi Zheng, Xuehai He, and Xin Eric Wang. Minigpt-5: Interleaved vision-and-language generation via generative vowels. arXiv preprint arXiv:2310.02239, 2023. 1, 6, 7, 8, 4, 9, 10, 11, 12" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 254, + 555, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 254, + 555, + 308 + ], + "spans": [ + { + "bbox": [ + 316, + 254, + 555, + 308 + ], + "type": "text", + "content": "[45] Wanrong Zhu, Jack Hessel, Anas Awadalla, Samir Yitzhak Gadre, Jesse Dodge, Alex Fang, Youngjae Yu, Ludwig Schmidt, William Yang Wang, and Yejin Choi. Multimodal c4: An open, billion-scale corpus of images interleaved with text. NeurIPS, 2023. 1, 2, 3, 4, 6, 7, 8" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8082" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/CoMapGS_ Covisibility Map-based Gaussian Splatting for Sparse Novel View Synthesis/025e1d28-cc10-46ad-8c05-b01e6a4dff3b_content_list.json b/2025/CoMapGS_ Covisibility Map-based Gaussian Splatting for Sparse Novel View Synthesis/025e1d28-cc10-46ad-8c05-b01e6a4dff3b_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..c95658c0af301509de9f28f1daba573bac816f51 --- /dev/null +++ b/2025/CoMapGS_ Covisibility Map-based Gaussian Splatting for Sparse Novel View Synthesis/025e1d28-cc10-46ad-8c05-b01e6a4dff3b_content_list.json @@ -0,0 +1,1595 @@ +[ + { + "type": "text", + "text": "CoMapGS: Covisibility Map-based Gaussian Splating for Sparse Novel View Synthesis", + "text_level": 1, + "bbox": [ + 222, + 130, + 776, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Youngkyoon Jang and Eduardo Pérez-Pellitero Huawei Noah's Ark Lab, London, UK", + "bbox": [ + 313, + 203, + 684, + 239 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8036dd9eb105e2357f9a64c9bf7788af13ca673d5fb63cc490a72b2ddd5d1aa3.jpg", + "image_caption": [ + "Figure 1. Main idea. Conventional sparse view synthesis methods using 3DGS [38] prioritize frequently captured regions, shown as the brightest areas in the covisibility map, resulting in better reconstruction in these regions but missing details in sparsely captured areas. By using a covisibility map, we guide sparse 3DGS methods to focus on underrepresented regions, enhancing sparse novel view synthesis." + ], + "image_footnote": [], + "bbox": [ + 94, + 256, + 903, + 465 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 535, + 326, + 551 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We propose Covisibility Map-based Gaussian Splatting (CoMapGS), designed to recover underrepresented sparse regions in sparse novel view synthesis. CoMapGS addresses both high- and low-uncertainty regions by constructing covisibility maps, enhancing initial point clouds, and applying uncertainty-aware weighted supervision using a proximity classifier. Our contributions are threefold: (1) CoMapGS reframes novel view synthesis by leveraging covisibility maps as a core component to address region-specific uncertainty; (2) Enhanced initial point clouds for both low- and high-uncertainty regions compensate for sparse COLMAP-derived point clouds, improving reconstruction quality and benefiting few-shot 3DGS methods; (3) Adaptive supervision with covisibility-score-based weighting and proximity classification achieves consistent performance gains across scenes with varying sparsity scores derived from covisibility maps. Experimental results demonstrate that CoMapGS outperforms state-of-the-art methods on datasets including Mip-NeRF 360 and LLFF. $^{1}$", + "bbox": [ + 89, + 569, + 485, + 857 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 514, + 535, + 643, + 551 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent advances in novel view synthesis have made significant progress in overcoming various challenges, such as generalized model building [27, 40], view synthesis in unbounded scenes [1], fast training [7, 8], and sparse view [38, 41], following the emergence of seminal methodologies such as Neural Radiance Fields (NeRF) [20] and 3D Gaussian Splatting (3DGS) [12]. While dense-view approaches now achieve photo-realistic fidelity, sparse-view synthesis methods often suffer from unrealistic artifacts due to the fundamental shape-radiance ambiguity—i.e., models can overfit to image observations because of implausible geometry, a problem that is exacerbated when training views are limited. Despite the limitations of available training view images, recent studies have proposed various approaches—such as generalized novel view synthesis methods [3, 5, 14, 27, 37, 40], diffusion-based techniques [30, 39], updated densification strategies [38, 41], and the use of pseudo ground truth [30, 38, 41—to improve performance in sparse view synthesis. While recent approaches, particularly those using 3DGS, have significantly improved performance in sparse view synthesis by leveraging rapid point-cloud-based rasterization, efficient gradient-", + "bbox": [ + 511, + 569, + 908, + 904 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "1 Project page: youngkyoonjang.github.io/projects/comapgs", + "bbox": [ + 107, + 887, + 423, + 902 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "26779", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "descent supervision, and multiview geometry, several open challenges remain:", + "bbox": [ + 89, + 90, + 480, + 121 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Region-wise imbalanced supervision: Conventional novel view synthesis methods prioritize highly covisible regions due to their multiview presence, effectively weighting them more while under-supervising monoview regions that do not have multiview constraints.", + "- Sparse point cloud initialization: Conventional sparse point clouds, generated through feature-based keypoint matching (e.g., SIFT [18]) in COLMAP [24], lack the geometric detail that dense view settings offer, especially when the number of training images is limited.", + "- High-uncertainty regions: Methods that rely heavily on multiview geometry often struggle with mono-view regions where pixels do not appear in other views, posing a critical challenge in sparse view synthesis." + ], + "bbox": [ + 89, + 125, + 482, + 337 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In response to these challenges, we propose CoMapGS, a Covisibility Map-based Gaussian Splating method that reframes sparse view synthesis by utilizing per-image, pixelwise covisibility counts (i.e., the covisibility map) for adaptive supervision and initial point cloud enhancement, effectively addressing both low- and high-uncertainty regions, as shown in Fig. 2. CoMapGS updates initial point clouds and applies adaptive supervision, using a proximity classifier to regularize plausible geometry in high-uncertainty (i.e., mono-view) regions for enhanced supervision. To effectively manage regions with varying covisibility levels, we construct a covisibility map by accumulating correspondence counts for each pixel from the output of the state-of-the-art dense correspondence prediction model, MASt3R [16]. Additionally, we recover initial point clouds across all regions, including mono-view areas, by learning the correlation between the scale of monocular depth estimation and the point clouds registered using already-optimized COLMAP camera poses, enabling precise rescaling and alignment within the same coordinate system. The updated point clouds are then integrated into CoMapGS to initialize the first set of Gaussians, which are rendered and refined throughout the training process.", + "bbox": [ + 89, + 340, + 482, + 688 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our proposed CoMapGS can be applied to most sparse view synthesis methods built on the conventional 3DGS framework. Starting with our updated initial point clouds, we supervise CoMapGS by incorporating an additional proximity loss term into the baseline objective function (e.g., FSGS [41], CoR-GS [38]). This proximity loss is derived from our proposed proximity MLP classifier, which predicts the closeness of each point to the original geometry. Leveraging the covisibility map, we apply adaptive weights to the Gaussians that compose the final proximity term, assigning greater weight to Gaussians projected onto mono-view regions and reduced weight to those in multiview regions. Additionally, we adjust supervision strength for Gaussians outside the view frustum based on scene-", + "bbox": [ + 89, + 688, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "specific可视ability scores, ensuring they contribute effectively to training even if they are not directly represented in rendered images.", + "bbox": [ + 511, + 90, + 903, + 136 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The main contributions of the CoMapGS are threefold:", + "bbox": [ + 532, + 136, + 895, + 151 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. Covisibility Map-based learning: CoMapGS reframes sparse view synthesis by leveraging covisibility maps for adaptive supervision, enabling balanced guidance across different uncertainty levels within the scene.", + "2. Enhanced point cloud initialization: By applying learned scaling parameters between monocular depth estimates and triangulated points registered using the given COLMAP camera poses, CoMapGS densely enhances initial point cloud quality, increasing the potential to densify sparse and underrepresented regions, even in mono-view areas.", + "3. Weighted supervision using covisibility map: This strategy adaptively reinforces high- and low-uncertainty regions, strengthening supervision based on region-specific covisibility levels and stabilizing training with learned geometric constraints that extend beyond the view frustum." + ], + "bbox": [ + 511, + 152, + 903, + 407 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To the best of our knowledge, CoMapGS is the first approach to explicitly address and recover high-uncertainty regions in sparse view synthesis, effectively balancing supervision across regions with varying uncertainty levels.", + "bbox": [ + 511, + 410, + 903, + 472 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 513, + 482, + 661, + 498 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Sparse novel view synthesis. The growing interest in reconstructing spatial photographs from minimal input image sets has driven the development of various methods aimed at generating novel views from just a few shots, utilizing approaches based on NeRF [6, 11, 13, 21, 23, 26, 28, 30, 31, 36] and 3DGS [17, 22, 32, 34, 38, 41]. Due to the inherent challenges of sparse view synthesis, such as the limited number of training views, NeRF-based studies have primarily focused on proposing regularization techniques with depth ranking [6, 15, 23, 26, 28, 31], frequency-based loss [36], semantic consistency with visual encoders [11], between-view smoothness [13], or uncertainty-aware adaptive loss [33] for both seen and unseen views [11, 15, 21, 30]. Some methods have also addressed uncertainty issues [13, 36], particularly for high-uncertainty mono-view or unseen regions. However, these methods either disregard [13] or penalize [36] uncertain areas, rather than fully utilizing the information contained within.", + "bbox": [ + 511, + 508, + 906, + 779 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "While recent 3DGS-based methods continue to rely on regularization techniques, such as depth constraints [17, 22, 32, 34, 41], some have additionally introduced novel densification strategies [38, 41] and explored enhancing initial point clouds [4, 7, 22, 34], capitalizing on the technical characteristics of 3DGS. These methods primarily focus on improving effective multiview regions [9], often biased towards well-represented regions in the training views, rather", + "bbox": [ + 511, + 779, + 908, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "26780", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/6fc020d41154632b362667b859297b38d6e31a8f398eb8d8cc0b0e616e0ed1b4.jpg", + "image_caption": [ + "Figure 2. Overview of the proposed CoMapGS. CoMapGS leverages existing modules, including Structure-from-Motion (SfM), monocular depth, and dense correspondence prediction, to produce preliminary outputs. Covisibility maps are generated and used alongside module outputs to enhance initial point clouds, which initialize Gaussians for CoMapGS. During training, covisibility map-based weighted supervision with the proximity classifier applies varied strengths to Gaussians based on region-specific covisibility." + ], + "image_footnote": [], + "bbox": [ + 94, + 88, + 903, + 310 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "than addressing recovery strategies for high-uncertainty mono-view regions. To overcome the inherent limitations of sparse training views, several prior works have explored generating and utilizing pseudo-ground truths for novel viewpoints via mono-depth estimation [41], forward warping with reversed bilinear sampling [34], or diffusion models with CLIP embedding [30]. While these approaches compensate for the lack of training views, they still face unique challenges in recovering unseen or mono-view regions from the original training view images, potentially misleading the learning process in high-uncertainty areas.", + "bbox": [ + 88, + 402, + 485, + 570 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Covisibility-aware supervision. Beginning with Dy-Check [9], which emphasizes the importance of covisibility in multiview regions to address fair evaluation in dynamic novel view synthesis, recent studies have explored integrating covisibility information into supervision strategies. NeRFVS [35] uses a view coverage map to enforce stricter regularization in low-coverage (covisible) regions, but it is primarily suited for dense view synthesis with video input, where extensive multiview overlap is expected. Similarly, InfoNorm [29] leverages mutual information among surface normals at highly correlated scene points to regularize geometric modeling. However, InfoNorm captures covisible areas within local subregions in structured scenes (e.g., rooms) rather than using a comprehensive covisibility map, which limits its applicability for addressing both multiview and monoview regions. While these methods demonstrate the value of covisibility in supervision, they lack a framework that addresses the challenges of sparse novel view synthesis for natural scenes, including outdoor environments, through a pixel-wise covisibility map.", + "bbox": [ + 89, + 598, + 485, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Covisibility Map-based Gaussian Splitting", + "text_level": 1, + "bbox": [ + 511, + 401, + 897, + 419 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The proposed CoMapGS framework is illustrated in Fig. 2. CoMapGS builds on the latest 3DGS-based sparse novel view synthesis method, CoR-GS [38], with two key improvements: (1) updating initial point clouds specifically for sparse and mono-view regions and (2) applying adaptive supervision across varying covisibility regions using a covisibility map and proximity MLP classifier. We begin by generating covisibility maps using the dense correspondence prediction model MASt3R [16], refining it with morphological operations (e.g., erosion and dilation) to support these strategies effectively.", + "bbox": [ + 511, + 426, + 906, + 594 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Covisibility Map Generation", + "text_level": 1, + "bbox": [ + 511, + 601, + 769, + 618 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We generate a covisibility map by identifying dense correspondences between each training view image and all other training view images. Given a set of training view images $T = \\{I_1, I_2, \\ldots, I_n\\}$ , where $n$ is the total number of training views, we compute a covisibility map $M_i$ for each view $I_i$ . For every pair $(I_i, I_{\\{j\\}})$ with $j \\neq i$ , dense correspondences $C_i^j$ are predicted by a dense correspondence prediction function $f_{\\mathrm{cor}}$ , where $C_i^j = f_{\\mathrm{cor}}(I_i, I_j)$ . The covisibility map $M_i$ is initialized as a zero matrix of size $W \\times H$ and defined as:", + "bbox": [ + 511, + 623, + 908, + 777 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nM _ {i} (x, y) = \\sum_ {j = 1, j \\neq i} ^ {n} \\delta_ {(x, y) \\in \\mathcal {P} \\left(C _ {i} ^ {j}\\right)}, \\quad \\forall (x, y) \\in I _ {i}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 535, + 785, + 906, + 828 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\delta_{(x,y)\\in \\mathcal{P}(C_i^j)}$ is an indicator variable returning 1 if pixel $(x,y)$ in $I_{i}$ has a correspondence match in $I_{j}$ , and 0 otherwise. Here, $\\mathcal{P}(C_i^j)$ represents the set of pixel coordinates in $I_{i}$ that have correspondences in $I_{j}$ .", + "bbox": [ + 511, + 835, + 910, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "26781", + "bbox": [ + 478, + 944, + 517, + 957 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The resulting covisibility map $M_{i}$ captures pixel-wise covisibility counts ranging from 0 to $n - 1$ , indicating the frequency with which a pixel appears across views, as shown in Fig. 1, with values rescaled from 0 (black) to 255 (white) for visualization purposes. Unlike DyCheck [9]'s covisibility mask, which only indicates pixel visibility, our covisibility map provides precise covisibility details (i.e., covisible view count) at the pixel level. This map supports both the initial point cloud update (Sec. 3.2) and the weighted supervision method (Sec. 3.3).", + "bbox": [ + 89, + 90, + 483, + 242 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Enhancing Initial Point Clouds", + "text_level": 1, + "bbox": [ + 89, + 250, + 366, + 265 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this subsection, we explain how to improve initial point clouds for both low-uncertainty multi-view and high-uncertainty mono-view regions. To enhance initial point clouds in low-uncertainty multi-view regions where sparse SfM points are available, we integrate a triangulated set of 3D points, $P_T$ , derived from dense correspondences obtained from MASt3R [16], to complement COLMAP's point cloud, $P_C$ . This approach enhances geometrically plausible structures, especially in regions where COLMAP's keypoint-based matching is sparse. For each view $I_i$ , we generate $P_T$ by triangulating dense correspondences $C_i^{\\{j\\}}$ from view $I_i$ to other views $I_{\\{j\\}}$ (e.g., two or three correspondences in a 3-view case). To avoid redundancy, correspondences already reconstructed from other views are excluded from subsequent triangulations.", + "bbox": [ + 89, + 271, + 483, + 500 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Formally, we define the updated point cloud $P_{\\mathrm{u}}$ as:", + "bbox": [ + 109, + 500, + 444, + 515 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nP _ {\\mathrm {u}} = P _ {C} \\cup \\left\\{p _ {T} \\in P _ {T} \\mid D \\left(p _ {T}, P _ {C}\\right) > \\epsilon \\right\\}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 143, + 521, + 482, + 539 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $D(p_{T}, P_{C})$ is the distance from a point $p_{T}$ in $P_{T}$ to its nearest neighbor in $P_{C}$ , and $\\epsilon$ is a threshold ensuring that only points sufficiently distant from existing points in $P_{C}$ are added. This selective integration allows the triangulated points to supplement COLMAP's sparse initial point cloud in the same coordinate system, filling gaps in low-uncertainty regions and improving overall density in covisible areas, as shown in Fig. 3.", + "bbox": [ + 89, + 545, + 483, + 666 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In high-uncertainty mono-view regions, we use monocular depth estimation [10] to generate a preliminary set of 3D points, $P_{d}$ , by unprojecting depth estimates using the provided camera intrinsics. This set is defined as $P_{d} = P_{d}^{\\mathrm{low}} \\cup P_{d}^{\\mathrm{high}}$ , where $P_{d}^{\\mathrm{low}}$ comprises points unprojected from multiview regions (where $M_{i} \\geq 1$ ), and $P_{d}^{\\mathrm{high}}$ comprises points unprojected from mono-view regions (where $M_{i} = 0$ ). To align these randomly scaled unprojected points with triangulated points from multiview geometry, we learn an anisotropic scaling transformation $f_{\\mathrm{scale}}$ , using covisible points from $P_{u}^{\\mathrm{low}} \\subseteq P_{u}$ as ground truth and $P_{d}^{\\mathrm{low}}$ as input. This transformation applies scaling parameters learned through linear regression to match the multiview-derived geometry:", + "bbox": [ + 89, + 667, + 483, + 878 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nf _ {\\text {s c a l e}} \\left(P _ {d} ^ {\\text {l o w}}\\right) \\approx P _ {u} ^ {\\text {l o w}}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 215, + 883, + 482, + 902 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/7ac89f152a45c3aed602face5f0743f28110c5ad04e7e71c0549e6f47f736e12.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 517, + 87, + 640, + 166 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/cac3204969951a0f3bce964575f0577b29406bcdd4ae4ff3fbbb61a710f1e9b7.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 640, + 87, + 769, + 166 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/0e6f6a6ad6de1689732fb55910d5654658f039daf54df1b211cb8798d412a638.jpg", + "image_caption": [ + "(c)", + "Figure 3. Examples of enhanced initial PCLs from the Fern scene in the LLFF dataset. (a) Initial SfM points are first (b) updated through triangulation using dense correspondences and COLMAP camera poses, and then (c) fully refined to include mono-view regions (mono-view updates are colored in red for visualization)." + ], + "image_footnote": [], + "bbox": [ + 771, + 87, + 900, + 167 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For consistency, we define $P_{u} = P_{u}^{\\mathrm{low}} \\cup P_{u}^{\\mathrm{high}}$ , where $P_{u}^{\\mathrm{low}}$ is selected based on covisibility:", + "bbox": [ + 511, + 285, + 905, + 316 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nP _ {u} ^ {\\text {l o w}} = \\left\\{p \\mid M _ {i} \\left(\\pi \\left(P _ {u}, \\mathbf {H} _ {i}\\right)\\right) \\geq 1 \\right\\}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 325, + 905, + 344 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\pi(P_u, \\mathbf{H}_i)$ denotes the projection of $P_u$ using the camera transformation matrix $\\mathbf{H}_i$ , yielding the image coordinates $(x, y)$ for a given camera viewpoint and ensuring inclusion of only points visible in multiple views.", + "bbox": [ + 511, + 353, + 905, + 414 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "After training, we apply $f_{\\mathrm{scale}}$ to the randomly scaled unprojected points from depth estimates in high-uncertainty mono-view regions:", + "bbox": [ + 511, + 415, + 905, + 459 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nP _ {s} ^ {\\text {h i g h}} = f _ {\\text {s c a l e}} \\left(P _ {d} ^ {\\text {h i g h}}\\right). \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 635, + 467, + 903, + 487 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This transformation aligns the randomly scaled points $P_{d}^{\\mathrm{high}}$ to the scale of COLMAP's triangulated points $P_{s}^{\\mathrm{high}}$ , ensuring spatial consistency across regions. The final point cloud, $P_{\\mathrm{final}}$ , is defined as:", + "bbox": [ + 511, + 497, + 905, + 561 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nP _ {\\text {f i n a l}} = P _ {u} ^ {\\text {l o w}} \\cup P _ {s} ^ {\\text {h i g h}}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 635, + 569, + 903, + 589 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "resulting in enhanced initial point cloud density and accuracy across both multiview and mono-view regions, as shown in Fig. 3, thus supporting cohesive scene reconstruction in sparse view synthesis.", + "bbox": [ + 511, + 598, + 905, + 657 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Covisibility Map-Based Weighted Supervision", + "text_level": 1, + "bbox": [ + 511, + 666, + 901, + 683 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We propose a covisibility map-based supervision strategy that leverages the enhanced initial point cloud, $P_{\\mathrm{final}}$ , and the covisibility map $M_{i}$ to adaptively supervise regions based on uncertainty levels represented by covisibility counts, as shown in Fig. 4. This approach improves conventional sparse view Gaussian splatting methods [38, 41] by introducing region-specific weighting across covisibility levels, $M_{i} = \\{0:n - 1\\}$ , and by supervising Gaussians located outside the training view frustum with a scene-level weight based on the average covisibility score $S = \\mathrm{AVG}(M_{[1:n]})$ .", + "bbox": [ + 511, + 688, + 905, + 840 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our approach integrates seamlessly with the latest 3DGS-based sparse view synthesis methods [38, 41] by adding a proximity loss term, weighted using inverse proximity scores predicted by a three-layer MLP classifier", + "bbox": [ + 511, + 840, + 905, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "26782", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "trained on $P_{\\mathrm{final}}$ (see details in the supplementary materials).", + "bbox": [ + 89, + 90, + 482, + 119 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Proximity loss with adaptive weighting. To implement this approach, we first train an MLP classifier $f_{p}$ on $P_{\\mathrm{final}}$ , labeling these curated scene 3D points as positive (representing the original scene geometry) and randomly distributed points, $P_{\\mathrm{random}}$ , constrained to remain distant from $P_{\\mathrm{final}}$ , as negative. The classifier $f_{p}: \\mathbb{R}^{3} \\to [0,1]$ outputs a proximity score $s$ for each 3D point $p$ :", + "bbox": [ + 89, + 121, + 483, + 227 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\ns = f _ {p} (p) \\in [ 0, 1 ], \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 220, + 236, + 482, + 253 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $s$ approaches 1 for points close to the scene geometry and 0 for distant points. Using batch processing, we compute the proximity scores efficiently for the entire Gaussian set, $G$ , and define the proximity loss $\\mathcal{L}_p$ as:", + "bbox": [ + 89, + 261, + 483, + 323 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {p} = \\frac {1}{| G |} \\sum_ {g \\in G} (\\chi (g) w _ {\\text {i n}} + (1 - \\chi (g)) w _ {\\text {o u t}}) \\cdot (1 - s), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 329, + 482, + 367 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $g$ denotes the 3D position of a Gaussian in the set $G$ , representing the complete Gaussian set. The indicator function $\\chi(g)$ is defined as:", + "bbox": [ + 89, + 375, + 483, + 420 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\chi (g) = \\left\\{ \\begin{array}{l l} 1, & \\text {i f} \\pi (g, \\mathbf {H} _ {i}) \\in [ 0, \\mathsf {W}) \\times [ 0, \\mathsf {H}) \\\\ 0, & \\text {o t h e r w i s e} \\end{array} , \\right. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 133, + 428, + 482, + 468 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\pi(g, \\mathbf{H}_i)$ is the projection of $g$ under camera transformation $\\mathbf{H}_i$ onto the可视ibility map $M_i$ .", + "bbox": [ + 89, + 476, + 482, + 506 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Within the camera-view frustum. We use the weight $w_{in}$ for Gaussians projected onto covisibility mask regions, defined as:", + "bbox": [ + 89, + 506, + 483, + 550 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nw _ {\\mathrm {i n}} = \\frac {1}{M _ {i} \\left(\\pi \\left(g , \\mathbf {H} _ {i}\\right)\\right) + 1}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 194, + 556, + 482, + 588 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $M_{i}(\\pi (g,\\mathbf{H}i))$ is the covisibility count of $g$ in the covisibility map for the corresponding projection. This weighting ensures that high-covisibility regions, which are frequently supervised by other loss terms, receive weaker proximity-related supervision, while mono-view regions benefit from a more direct application of $\\mathcal{L}_p$ .", + "bbox": [ + 89, + 595, + 482, + 686 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Outside the camera-view frustum. In conventional 3DGS-based methods, only Gaussians within the view frustum are supervised per iteration. However, for scenes with high covisibility scores ( $S > 0.7$ ), we extend supervision to Gaussians located outside the view frustum, applying a scene-level weight $w_{out}$ for these points:", + "bbox": [ + 89, + 686, + 483, + 777 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nw _ {o u t} = \\max \\left(0, \\frac {S - 0 . 7}{0 . 3}\\right), \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 187, + 784, + 482, + 818 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "which scales linearly from 1 to 0 as $S$ ranges from 1 to 0.7. This approach improves supervision outside the view frustum in high-covisibility score scenes, ensuring comprehensive guidance while making efficient use of the proximity classifier.", + "bbox": [ + 89, + 824, + 483, + 898 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/7a91f338337b3a8e1d39c6031da96fbb1027443d5d666fac143882504ccb23d4.jpg", + "image_caption": [ + "Figure 4. Alongside (a) the conventional reconstruction loss $\\mathcal{L}1$ , we add (b) a weighted proximity loss $\\mathcal{L}_p$ that applies classifier results with assigned weights based on Gaussian projections onto the可视ability map and their positions relative to the frustum." + ], + "image_footnote": [], + "bbox": [ + 516, + 89, + 901, + 236 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Objective Function", + "text_level": 1, + "bbox": [ + 511, + 329, + 697, + 345 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our proposed CoMapGS enhances the objective function by introducing a proximity loss term, effectively complementing the function while preserving the original strategies of state-of-the-art methods, such as Co-pruning and Pseudoview Co-regularization from CoR-GS [38]. The full objective function is expressed as:", + "bbox": [ + 511, + 352, + 906, + 443 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = (1 - \\lambda) \\mathcal {L} 1 \\left(I, I ^ {*}\\right) + \\lambda \\mathcal {L} _ {\\mathrm {D} - \\mathrm {S S I M}} \\left(I, I ^ {*}\\right) + \\mathcal {L} _ {p}, \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 532, + 455, + 905, + 474 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $I$ is the rendered image, supervised against ground truth $I^{*}$ , $\\mathcal{L}1$ is the L1 reconstruction loss, and $\\mathcal{L}_{\\mathrm{D - SSIM}}$ represents the D-SSIM loss term. The balancing parameter $\\lambda$ controls the weighting between L1 and D-SSIM, in line with standard Gaussian splattering objectives. The proximity loss term $\\mathcal{L}_p$ , defined in Eq. 8, adds adaptive supervision to improve geometry alignment.", + "bbox": [ + 511, + 486, + 905, + 592 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our proposed approach extends seamlessly to other recent sparse view synthesis methods, such as FSGS [41], by incorporating $\\mathcal{L}_p$ into their objective function while preserving their original strategies, as reported in Table 1.", + "bbox": [ + 511, + 593, + 905, + 654 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experimental Results", + "text_level": 1, + "bbox": [ + 511, + 669, + 718, + 686 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Setup", + "text_level": 1, + "bbox": [ + 511, + 695, + 594, + 710 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. We conducted experiments on two benchmark datasets: LLFF [19], using 3, 6, and 9 training views with resolutions downsampled to $8 \\times$ , and Mip-NeRF 360 [2], using 12 and 24 training views per scene with resolutions downsampled to $4 \\times$ . We adhered to the standard training and testing conventions [38] for datasets commonly used in sparse view synthesis.", + "bbox": [ + 511, + 718, + 905, + 823 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Comparison Baselines. We compare CoMapGS with existing few-shot novel view synthesis methods, including FSGS [41] and CoR-GS [38]. As our method primarily enhances the performance of 3DGS-based approaches, we reproduced results for these methods. For other methods, we", + "bbox": [ + 511, + 824, + 905, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "26783", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/af3865340f116d86b37d98feeb210a7dd6bbb618f33e0094fe9b77e8429f00e5.jpg", + "table_caption": [ + "Table 1. Quantitative results on LLFF with 3, 6, and 9 training views. The best, second-best, and third-best entries are marked in red, orange, and yellow, respectively. $\\dagger$ indicates results reproduced by us using the released code base." + ], + "table_footnote": [], + "table_body": "
MethodPSNR ↑SSIM ↑LPIPS ↓
3-view6-view9-view3-view6-view9-view3-view6-view9-view
DietNeRF [11]14.9421.7524.280.3700.7170.8010.4960.2480.183
Mip-NeRF [1]16.1122.9124.880.4010.7560.8260.4600.2130.160
RegNeRF [21]19.0823.1024.860.5870.7600.8200.3360.2060.161
SimpleNeRF [25]19.2423.0523.980.6230.7370.7620.3750.2960.286
FreeNeRF [36]19.6323.7325.130.6120.7790.8270.3080.1950.160
SparseNeRF [28]19.8623.2624.270.7140.7410.7810.2430.2350.228
DiffusioNeRF [31]20.1323.6024.620.6310.7750.8070.3440.2350.216
ReconFusion [30]21.3424.2525.210.7240.8150.8480.2030.1520.134
DNGaussian [17]19.1222.1823.170.5910.7550.7880.2940.1980.180
3DGS [12]19.2223.8025.440.6490.8140.8600.2290.1250.096
FSGS [41]†20.45824.69426.0060.7130.8390.8750.2050.1210.095
CoR-GS [38]†20.47324.77726.4750.7170.8440.8810.1990.1160.086
[41]† + CoMapGS (ours)20.89124.80826.2920.7370.8440.8790.1900.1170.092
[38]† + CoMapGS (ours)21.10525.20426.7310.7470.8540.8860.1820.1080.082
", + "bbox": [ + 106, + 128, + 893, + 400 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/cdc5b80ddb9038cc63817b9cc6b55aed2a723fc2dc829f9963d54573207839b3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 94, + 412, + 256, + 507 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/bb2d96001490401f4cd3b6753cee514f270c6e63b47cc167c83f71bbe56c71a1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 96, + 508, + 256, + 601 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1fd01778b3bda0fd3e2a157c910a5ec072a1c80b53fa32a6c12f730ac44ae21d.jpg", + "image_caption": [ + "FSGS" + ], + "image_footnote": [], + "bbox": [ + 96, + 602, + 256, + 695 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c5391d363b97828f571572ad1bf6b11034cc6075bd3acdf9b09859bff86ce904.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 258, + 412, + 416, + 508 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4473f9fd1920f51392b5c01e00b0b306101e5002633053af0dd6a1d0417e7db4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 258, + 508, + 416, + 601 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/6e0f8f5891ce569ccecfa80422462dacf1b82806f59c8d96ac9b62219cf971ff.jpg", + "image_caption": [ + "CoR-GS" + ], + "image_footnote": [], + "bbox": [ + 258, + 602, + 416, + 695 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/3f9bac9de3ab091468f154e9ed3ab449efdead776ec23e8d4c6803508d45c6fd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 418, + 412, + 578, + 508 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/34195533195e7087d72fdb7d3b8205aa224f5434703239bc3c15c672bcfe88d5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 418, + 508, + 578, + 601 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1d53208fa5e2da576cc6539cd1fd4108fd5102a34b86f71152c6a5a0e9799be5.jpg", + "image_caption": [ + "Ours", + "Figure 5. Qualitative comparison on 3 training views from the LLFF dataset. We compare our results with FSGS [41] and CoR-GS [38], demonstrating that our method achieves superior texture synthesis in novel views, particularly in high-uncertainty mono-view regions that appear only once in the training set." + ], + "image_footnote": [], + "bbox": [ + 418, + 602, + 578, + 695 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/efbe0a61a970aaff676ad996c7956404d11c4bf77998d5df632ff1f7c4b66992.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 580, + 412, + 740, + 508 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/bbb2b25365aec124c991a1f3dc5428e2d9b7fa386ee1f516047eb5ee9cc9a8a7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 580, + 508, + 738, + 601 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/0e53e929cde0538493dcf585612cb5cf4209e4d7d7845cd26eaafb8ac9551e41.jpg", + "image_caption": [ + "Ground Truth" + ], + "image_footnote": [], + "bbox": [ + 580, + 602, + 738, + 695 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/130b119ca75288856f5b6cb312d047e2bbecd3e1c85de00903114c8336554421.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 741, + 412, + 901, + 508 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/16d3dab5adf8d13e3fb94ee6cca4d753674e7eb29eaeb73a9b2e294308a82183.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 741, + 508, + 901, + 601 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f252d99f9e77044a86033fb48eda671de035f1a6441a9b543a08e2cea7e7533d.jpg", + "image_caption": [ + "Covisibility Mask" + ], + "image_footnote": [], + "bbox": [ + 741, + 601, + 901, + 695 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "report the best published scores from prior works, as shown in Tables 1 and 2.", + "bbox": [ + 89, + 789, + 482, + 819 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metrics. We evaluate our results using PSNR, SSIM, and LPIPS metrics.", + "bbox": [ + 89, + 821, + 482, + 852 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation. Training was conducted for 10k iterations on the LLFF dataset and 30k iterations on the MipNeRF 360 dataset, following the setup of 3DGS [12]. Con", + "bbox": [ + 89, + 854, + 483, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "sistent with experimental setups in FSGS [41] and CoR-GS [38], CoMapGS was initialized with point clouds and precomputed camera poses from COLMAP.", + "bbox": [ + 511, + 789, + 906, + 835 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Comparisons", + "text_level": 1, + "bbox": [ + 511, + 847, + 653, + 863 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "LLFF. Quantitative and qualitative results on the LLFF dataset are presented in Table 1 and Fig. 5, respectively.", + "bbox": [ + 511, + 869, + 906, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "26784", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/2ac3ed3e6269f2afab7c9d6e5cdd9ff1fb44528846105a9e835431440a7cad39.jpg", + "table_caption": [ + "Table 2. Quantitative results on Mip-NeRF360 with 12 and 24 training views. The best, second-best, and third-best entries are marked in red, orange, and yellow, respectively. ${}^{ \\dagger }$ indicates results reproduced by us using the released code base." + ], + "table_footnote": [], + "table_body": "
MethodPSNR ↑SSIM ↑LPIPS ↓
12-view24-view12-view24-view12-view24-view
3DGS [12]18.5222.800.5230.7080.4150.276
FSGS [41]18.8023.280.5310.7150.4180.274
CoR-GS [38]†19.16223.3210.5740.7290.4140.271
[38]† + CoMapGS (ours)19.68023.4620.5910.7340.3940.264
", + "bbox": [ + 184, + 128, + 816, + 237 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/657c6f64c99b25fec1064da2a6118654b62ca309a4d69645f3f1770ab20cd348.jpg", + "image_caption": [ + "Figure 6. Qualitative comparison on 12 training views from the Mip-NeRF 360 dataset. Mip-NeRF 360 highlights diverse covisibility levels, with high-uncertainty regions sparsely captured. CoMapGS effectively synthesizes these regions, while CoR-GS struggles to recover textures." + ], + "image_footnote": [], + "bbox": [ + 96, + 247, + 901, + 676 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We reproduced evaluation scores for the two baseline methods, FSGS [41] and CoR-GS [38], to highlight the improvements over our reproduced results. The results show consistent improvements, ranking first across various training view counts, with PSNR placing second only to ReconFusion [30] in the 3-view case. This difference is likely due to diffusion-based sparse view synthesis methods, which are optimized for generating visually seamless colors that blend with surrounding regions. However, this approach may not", + "bbox": [ + 89, + 755, + 485, + 891 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "accurately capture original patterns in mono-view or unseen regions—common in sparse view synthesis datasets, especially with fewer training views. Consequently, while we ranked second in PSNR, our method achieved substantially higher SSIM and lower LPIPS scores under comparable view settings, indicating improved structural fidelity and perceptual quality in the synthesized views.", + "bbox": [ + 511, + 755, + 906, + 861 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Mip-NeRF 360. To validate CoMapGS's applicability to broader scenes that capture large areas with outward-", + "bbox": [ + 511, + 869, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "26785", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "facing cameras, we evaluated our method using the MipNeRF 360 dataset. This dataset, which centers objects of interest with outward-facing cameras, captures the motivation of our approach by featuring distinct low-uncertainty, multiview regions alongside high-uncertainty, less-covisible regions within an image, as shown in Fig. 1. In these scenarios, COLMAP struggles to establish stable correspondences, leading to sparse initial point clouds that fail to adequately recover details in high-uncertainty regions, as shown in Fig. 6.", + "bbox": [ + 89, + 90, + 480, + 241 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In contrast, CoMapGS begins with enhanced initial point clouds, preserving the scene's original structure while benefiting from proximity loss supervision, resulting in improved texture representation over conventional methods, as shown in Fig. 6. Quantitative results in Table 2 confirm significant improvements across metrics, particularly LPIPS, compared to recent state-of-the-art methods, which perform worse than their baselines (e.g., 3DGS [12], FSGS [41]). In the 24-view case, while PSNR improvements may appear modest due to unseen regions in test views, the LPIPS gain highlights our method's effectiveness in recovering textures from both low- and high-uncertainty regions in novel view.", + "bbox": [ + 89, + 242, + 482, + 422 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 435, + 243, + 450 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The ablation study in Table 3 highlights that dense reconstruction in low-uncertainty (covisible) regions already provides significant improvement over baseline methods, underscoring the benefit of increased point density in covisible areas. Furthermore, adding points from mono-view regions to the enhanced point cloud yields additional gains, demonstrating the value of incorporating points from underrepresented, high-uncertainty regions in sparse view synthesis.", + "bbox": [ + 89, + 457, + 482, + 578 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We also evaluated the impact of our proximity loss-based weighted supervision both with and without the initial point cloud enhancement strategy. Interestingly, starting from sparse COLMAP point clouds without updates showed limited improvement, likely because 3DGS-based methods tend to densify Gaussians near their original positions. However, when initialized with our enhanced point clouds, the benefits of weighted supervision became evident, delivering significant performance gains in both partial (multiview region) and full (multi- and mono-view region) update scenarios. This demonstrates that enhancing initial point clouds and incorporating proximity loss-based weighted supervision work synergistically to optimize 3DGS-based sparse view synthesis.", + "bbox": [ + 89, + 580, + 482, + 790 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Discussion", + "text_level": 1, + "bbox": [ + 89, + 801, + 209, + 816 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Definition of sparse view synthesis. Although sparse novel view synthesis is often defined by a specific number of training views (e.g., LLFF with 3, 6, or 9 views; Mip-NeRF with 12 or 24 views), we argue that the true challenge lies in reconstructing underrepresented regions with low covisibility", + "bbox": [ + 89, + 824, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/02c822666b22c4f6cb078096a6930aa92769971ed505ca88dcd07c95e85813de.jpg", + "table_caption": [ + "Table 3. Ablation study of the proposed CoMapGS with six training views. $\\triangle$ indicates enhancing initial point clouds only in the low-uncertainty region (where covisible count $\\geq 2$ ), representing a straightforward application of dense point clouds generated by conventional methods such as MAST3R [16]." + ], + "table_footnote": [], + "table_body": "
Proposed methodLLFF
Sec. 3.2Sec. 3.3PSNR ↑SSIM ↑LPIPS ↓
CoR-GS [38]†24.7770.8440.116
✓24.7870.8450.116
△24.900.8490.112
△✓25.1530.8540.109
✓25.0760.8520.109
✓✓25.2040.8540.108
", + "bbox": [ + 527, + 169, + 893, + 309 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "counts rather than the total view count. This problem is inherently region-wise and imbalanced, where certain areas appear in only a few views, leading to sparse representation even when many views are available. For example, as DyCheck [9] categorizes scene types based on effective multiview factors (EMFs), scenarios with high frame counts but low region-specific可视ibility can still be considered sparse view cases. We address this by applying region-wise adaptive supervision based on可视ibility counts, improving results, especially for imbalanced datasets like MipNeRF 360 (12 views). In such cases, central objects are densely captured, while outer areas show low可视ibility due to the 360-degree, outward-facing capture setup. This issue is common in real-world scene capture, where uneven可视ibility across views motivates balanced,可视ibility-aware supervision.", + "bbox": [ + 511, + 340, + 906, + 583 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusions", + "text_level": 1, + "bbox": [ + 513, + 608, + 638, + 625 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper introduces CoMapGS, a Covisibility Map-based Gaussian Splitting approach designed to enhance sparse view synthesis methods by addressing region-specific uncertainty levels. We generate covisibility maps for a given non-continuous set of distant images, using them to identify low- and high-uncertainty regions and to improve point cloud initialization beyond conventional methods like COLMAP. By introducing covisibility-adaptive weighted supervision, CoMapGS effectively manages varying sparsity across both multiview and mono-view regions, consistently outperforming state-of-the-art 3DGS-based sparse novel view synthesis methods. Experimental results validate CoMapGS's effectiveness across diverse, realistic scenes, including both indoor and outdoor environments and various scales of captured areas. CoMapGS demonstrates superiority in handling complex region-wise covisibility, highlighting its robustness and adaptability in challenging sparse view synthesis tasks.", + "bbox": [ + 511, + 636, + 906, + 888 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "26786", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 90, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jonathan T. Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P. Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In ICCV, pages 5855–5864, 2021. 1, 6", + "[2] Jonathan T. Barron, Ben Mildenhall, Dor Verbin, Pratul P. Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5470-5479, 2022. 5", + "[3] Anpei Chen, Zexiang Xu, Fuqiang Zhao, Xiaoshuai Zhang, Fanbo Xiang, Jingyi Yu, and Hao Su. Mvsnerf: Fast generalizable radiance field reconstruction from multi-view stereo. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 14124-14133, 2021. 1", + "[4] Shan Chen, Jiale Zhou, and Lei Li. Dense point clouds matter: Dust-gs for scene reconstruction from sparse viewpoints, 2024. 2", + "[5] Wenyan Cong, Hanxue Liang, Peihao Wang, Zhiwen Fan, Tianlong Chen, Mukund Varma, Yi Wang, and Zhangyang Wang. Enhancing nerf akin to enhancing llms: Generalizable nerf transformer with mixture-of-view-experts. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 3193-3204, 2023. 1", + "[6] Kangle Deng, Andrew Liu, Jun-Yan Zhu, and Deva Ramanan. Depth-supervised nerf: Fewer views and faster training for free. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12882-12891, 2022. 2", + "[7] Zhiwen Fan, Wenyan Cong, Kairun Wen, Kevin Wang, Jian Zhang, Xinghao Ding, Danfei Xu, Boris Ivanovic, Marco Pavone, Georgios Pavlakos, Zhangyang Wang, and Yue Wang. Instantsplat: Unbounded sparse-view pose-free gaussian splatting in 40 seconds, 2024. 1, 2", + "[8] Guangchi Fang and Bing Wang. Mini-Splatting: Representing scenes with a constrained number of gaussians. In ECCV, Milano, Italy, 2024. 1", + "[9] Hang Gao, Ruilong Li, Shubham Tulsiani, Bryan Russell, and Angjoo Kanazawa. Monocular dynamic view synthesis: A reality check. In NeurIPS, 2022. 2, 3, 4, 8", + "[10] Mu Hu, Wei Yin, Chi Zhang, Zhipeng Cai, Xiaoxiao Long, Hao Chen, Kaixuan Wang, Gang Yu, Chunhua Shen, and Shaojie Shen. Metric3d v2: A versatile monocular geometric foundation model for zero-shot metric depth and surface normal estimation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(12):10579-10596, 2024. 4", + "[11] Ajay Jain, Matthew Tancik, and Pieter Abbeel. Putting nerf on a diet: Semantically consistent few-shot view synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 5885-5894, 2021. 2, 6", + "[12] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42 (4), 2023. 1, 6, 7, 8", + "[13] Mijeong Kim, Seonguk Seo, and Bohyung Han. Infonerf: Ray entropy minimization for few-shot neural volume rendering. In Proceedings of the IEEE/CVF Conference on" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Computer Vision and Pattern Recognition (CVPR), pages 12912-12921, 2022. 2", + "[14] Jonás Kulhánek, Erik Derner, Torsten Sattler, and Robert Babuška. Viewformer: Nerf-free neural rendering from few images using transformers. In European Conference on Computer Vision (ECCV), 2022. 1", + "[15] Min-Seop Kwak, Jiuhn Song, and Seungryong Kim. Geconerf: few-shot neural radiance fields via geometric consistency. In Proceedings of the 40th International Conference on Machine Learning. JMLR.org, 2023. 2", + "[16] Vincent Leroy, Yohann Cabon, and Jerome Revaud. Grounding image matching in 3d with mast3r, 2024. 2, 3, 4, 8", + "[17] Jiahe Li, Jiawei Zhang, Xiao Bai, Jin Zheng, Xin Ning, Jun Zhou, and Lin Gu. Dngaussian: Optimizing sparse-view 3d gaussian radiance fields with global-local depth normalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20775-20785, 2024. 2, 6", + "[18] David G. Lowe. Distinctive image features from scale-invariant keypoints. Int. J. Comput. Vision, 60(2):91-110, 2004. 2", + "[19] Ben Mildenhall, Pratul P. Srinivasan, Rodrigo Ortiz-Cayon, Nima Khademi Kalantari, Ravi Ramamoorthi, Ren Ng, and Abhishek Kar. Local light field fusion: Practical view synthesis with prescriptive sampling guidelines. ACM Transactions on Graphics (TOG), 2019. 5", + "[20] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 1", + "[21] Michael Niemeyer, Jonathan T. Barron, Ben Mildenhall, Mehdi S. M. Sajjadi, Andreas Geiger, and Noha Radwan. Regnerf: Regularizing neural radiance fields for view synthesis from sparse inputs. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 6", + "[22] Avinash Paliwal, Wei Ye, Jinhui Xiong, Dmytro Kotovenko, Rakesh Ranjan, Vikas Chandra, and Nima Khademi Kalantari. CoherentGS: Sparse novel view synthesis with coherent 3d gaussians. In ECCV, Milano, Italy, 2024. 2", + "[23] Barbara Roessle, Jonathan T. Barron, Ben Mildenhall, Pratul P. Srinivasan, and Matthias Nießner. Dense depth priors for neural radiance fields from sparse input views. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12892-12901, 2022. 2", + "[24] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 2", + "[25] Nagabhushan Somraj, Adithyan Karanayil, and Rajiv Soundararajan. SimpleNeRF: Regularizing sparse input neural radiance fields with simpler solutions. In SIGGRAPH Asia, 2023. 6", + "[26] Jiuhn Song, Seonghoon Park, Honggyu An, Seokju Cho, Min-Seop Kwak, Sungjin Cho, and Seungryong Kim. Därf: Boosting radiance fields from sparse input views with monocular depth adaptation. In Advances in Neural Information Processing Systems, pages 68458-68470. Curran Associates, Inc., 2023. 2" + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "26787", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[27] Thomas Tanay and Matteo Maggioni. Global latent neural rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19723-19733, 2024. 1", + "[28] Guangcong Wang, Zhaoxi Chen, Chen Change Loy, and Ziwei Liu. Sparsenerf: Distilling depth ranking for few-shot novel view synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 9065-9076, 2023. 2, 6", + "[29] Xulong Wang, Siyan Dong, Youyi Zheng, and Yanchao Yang. InfoNorm: Mutual information shaping of normals for sparse-view reconstruction. In ECCV, Milano, Italy, 2024. 3", + "[30] Rundi Wu, Ben Mildenhall, Philipp Henzler, Keunhong Park, Ruiqi Gao, Daniel Watson, Pratul P. Srinivasan, Dor Verbin, Jonathan T. Barron, Ben Poole, and Aleksander Ho?y?ski. Reconfusion: 3d reconstruction with diffusion priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 21551-21561, 2024. 1, 2, 3, 6, 7", + "[31] Jamie Wynn and Daniyar Turmukhambetov. Diffusionerf: Regularizing neural radiance fields with denoising diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4180-4189, 2023. 2, 6", + "[32] Haolin Xiong, Sairisheek Muttukuru, Rishi Upadhyay, Pradyumna Chari, and Achuta Kadambi. Sparsegs: Realtime $360^{\\circ}$ sparse view synthesis using gaussian splatting. Arxiv, 2023. 2", + "[33] Qingshan Xu, Xuanyu Yi, Jianyao Xu, Wenbing Tao, Yew Soon Ong, and Hanwang Zhang. Few-shot nerf by adaptive rendering loss regularization. In ECCV, Milano, Italy, 2024. 2", + "[34] Wangze Xu, Huachen Gao, Shihe Shen, Rui Peng, Jianbo Jiao, and Ronggang Wang. Mvpgs: Excavating multi-view priors for gaussian splatting from sparse input views. In ECCV, Milano, Italy, 2024. 2, 3", + "[35] Chen Yang, Peihao Li, Zanwei Zhou, Shanxin Yuan, Bingbing Liu, Xiaokang Yang, Weichao Qiu, and Wei Shen. NeRFVS: Neural Radiance Fields for Free View Synthesis via Geometry Scaffolds. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 3", + "[36] Jiawei Yang, Marco Pavone, and Yue Wang. Freenerf: Improving few-shot neural rendering with free frequency regularization. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2023. 2, 6", + "[37] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelnerf: Neural radiance fields from one or few images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4578-4587, 2021. 1", + "[38] Jiawei Zhang, Jiahe Li, Xiaohan Yu, Lei Huang, Lin Gu, Jin Zheng, and Xiao Bai. Cor-gs: Sparse-view 3d gaussian splatt-ting via co-regularization. In ECCV, Milano, Italy, 2024. 1, 2, 3, 4, 5, 6, 7, 8", + "[39] Zhizhuo Zhou and Shubham Tulsiani. Sparsefusion: Distilling view-conditioned diffusion for 3d reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vi" + ], + "bbox": [ + 91, + 90, + 480, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "sion and Pattern Recognition (CVPR), pages 12588-12597, 2023. 1", + "[40] Haidong Zhu, Tianyu Ding, Tianyi Chen, Ilya Zharkov, Ram Nevatia, and Luming Liang. CaesarNeRF: Calibrated semantic representation for few-shot generalizable neural rendering. In ECCV, Milano, Italy, 2024. 1", + "[41] Zehao Zhu, Zhiwen Fan, Yifan Jiang, and Zhangyang Wang. Fsgs: Real-time few-shot view synthesis using gaussian splatting. In ECCV, Milano, Italy, 2024. 1, 2, 3, 4, 5, 6, 7, 8" + ], + "bbox": [ + 516, + 92, + 903, + 231 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "26788", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/CoMapGS_ Covisibility Map-based Gaussian Splatting for Sparse Novel View Synthesis/025e1d28-cc10-46ad-8c05-b01e6a4dff3b_model.json b/2025/CoMapGS_ Covisibility Map-based Gaussian Splatting for Sparse Novel View Synthesis/025e1d28-cc10-46ad-8c05-b01e6a4dff3b_model.json new file mode 100644 index 0000000000000000000000000000000000000000..c8a8f699e0a01a4e3ce5a7a8a6675d93aa5685fa --- /dev/null +++ b/2025/CoMapGS_ Covisibility Map-based Gaussian Splatting for Sparse Novel View Synthesis/025e1d28-cc10-46ad-8c05-b01e6a4dff3b_model.json @@ -0,0 +1,2167 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.223, + 0.131, + 0.777, + 0.177 + ], + "angle": 0, + "content": "CoMapGS: Covisibility Map-based Gaussian Splating for Sparse Novel View Synthesis" + }, + { + "type": "text", + "bbox": [ + 0.315, + 0.204, + 0.685, + 0.24 + ], + "angle": 0, + "content": "Youngkyoon Jang and Eduardo Pérez-Pellitero Huawei Noah's Ark Lab, London, UK" + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.257, + 0.905, + 0.466 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.474, + 0.908, + 0.517 + ], + "angle": 0, + "content": "Figure 1. Main idea. Conventional sparse view synthesis methods using 3DGS [38] prioritize frequently captured regions, shown as the brightest areas in the covisibility map, resulting in better reconstruction in these regions but missing details in sparsely captured areas. By using a covisibility map, we guide sparse 3DGS methods to focus on underrepresented regions, enhancing sparse novel view synthesis." + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.536, + 0.327, + 0.552 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.57, + 0.486, + 0.858 + ], + "angle": 0, + "content": "We propose Covisibility Map-based Gaussian Splatting (CoMapGS), designed to recover underrepresented sparse regions in sparse novel view synthesis. CoMapGS addresses both high- and low-uncertainty regions by constructing covisibility maps, enhancing initial point clouds, and applying uncertainty-aware weighted supervision using a proximity classifier. Our contributions are threefold: (1) CoMapGS reframes novel view synthesis by leveraging covisibility maps as a core component to address region-specific uncertainty; (2) Enhanced initial point clouds for both low- and high-uncertainty regions compensate for sparse COLMAP-derived point clouds, improving reconstruction quality and benefiting few-shot 3DGS methods; (3) Adaptive supervision with covisibility-score-based weighting and proximity classification achieves consistent performance gains across scenes with varying sparsity scores derived from covisibility maps. Experimental results demonstrate that CoMapGS outperforms state-of-the-art methods on datasets including Mip-NeRF 360 and LLFF.\\(^{1}\\)" + }, + { + "type": "title", + "bbox": [ + 0.515, + 0.536, + 0.645, + 0.552 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.57, + 0.909, + 0.905 + ], + "angle": 0, + "content": "Recent advances in novel view synthesis have made significant progress in overcoming various challenges, such as generalized model building [27, 40], view synthesis in unbounded scenes [1], fast training [7, 8], and sparse view [38, 41], following the emergence of seminal methodologies such as Neural Radiance Fields (NeRF) [20] and 3D Gaussian Splatting (3DGS) [12]. While dense-view approaches now achieve photo-realistic fidelity, sparse-view synthesis methods often suffer from unrealistic artifacts due to the fundamental shape-radiance ambiguity—i.e., models can overfit to image observations because of implausible geometry, a problem that is exacerbated when training views are limited. Despite the limitations of available training view images, recent studies have proposed various approaches—such as generalized novel view synthesis methods [3, 5, 14, 27, 37, 40], diffusion-based techniques [30, 39], updated densification strategies [38, 41], and the use of pseudo ground truth [30, 38, 41—to improve performance in sparse view synthesis. While recent approaches, particularly those using 3DGS, have significantly improved performance in sparse view synthesis by leveraging rapid point-cloud-based rasterization, efficient gradient-" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.888, + 0.424, + 0.904 + ], + "angle": 0, + "content": "1 Project page: youngkyoonjang.github.io/projects/comapgs" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "26779" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.122 + ], + "angle": 0, + "content": "descent supervision, and multiview geometry, several open challenges remain:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.126, + 0.483, + 0.201 + ], + "angle": 0, + "content": "- Region-wise imbalanced supervision: Conventional novel view synthesis methods prioritize highly covisible regions due to their multiview presence, effectively weighting them more while under-supervising monoview regions that do not have multiview constraints." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.202, + 0.483, + 0.277 + ], + "angle": 0, + "content": "- Sparse point cloud initialization: Conventional sparse point clouds, generated through feature-based keypoint matching (e.g., SIFT [18]) in COLMAP [24], lack the geometric detail that dense view settings offer, especially when the number of training images is limited." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.277, + 0.483, + 0.338 + ], + "angle": 0, + "content": "- High-uncertainty regions: Methods that rely heavily on multiview geometry often struggle with mono-view regions where pixels do not appear in other views, posing a critical challenge in sparse view synthesis." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.126, + 0.483, + 0.338 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.342, + 0.483, + 0.689 + ], + "angle": 0, + "content": "In response to these challenges, we propose CoMapGS, a Covisibility Map-based Gaussian Splating method that reframes sparse view synthesis by utilizing per-image, pixelwise covisibility counts (i.e., the covisibility map) for adaptive supervision and initial point cloud enhancement, effectively addressing both low- and high-uncertainty regions, as shown in Fig. 2. CoMapGS updates initial point clouds and applies adaptive supervision, using a proximity classifier to regularize plausible geometry in high-uncertainty (i.e., mono-view) regions for enhanced supervision. To effectively manage regions with varying covisibility levels, we construct a covisibility map by accumulating correspondence counts for each pixel from the output of the state-of-the-art dense correspondence prediction model, MASt3R [16]. Additionally, we recover initial point clouds across all regions, including mono-view areas, by learning the correlation between the scale of monocular depth estimation and the point clouds registered using already-optimized COLMAP camera poses, enabling precise rescaling and alignment within the same coordinate system. The updated point clouds are then integrated into CoMapGS to initialize the first set of Gaussians, which are rendered and refined throughout the training process." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Our proposed CoMapGS can be applied to most sparse view synthesis methods built on the conventional 3DGS framework. Starting with our updated initial point clouds, we supervise CoMapGS by incorporating an additional proximity loss term into the baseline objective function (e.g., FSGS [41], CoR-GS [38]). This proximity loss is derived from our proposed proximity MLP classifier, which predicts the closeness of each point to the original geometry. Leveraging the covisibility map, we apply adaptive weights to the Gaussians that compose the final proximity term, assigning greater weight to Gaussians projected onto mono-view regions and reduced weight to those in multiview regions. Additionally, we adjust supervision strength for Gaussians outside the view frustum based on scene-" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.137 + ], + "angle": 0, + "content": "specific可视ability scores, ensuring they contribute effectively to training even if they are not directly represented in rendered images." + }, + { + "type": "text", + "bbox": [ + 0.534, + 0.137, + 0.897, + 0.152 + ], + "angle": 0, + "content": "The main contributions of the CoMapGS are threefold:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.153, + 0.905, + 0.213 + ], + "angle": 0, + "content": "1. Covisibility Map-based learning: CoMapGS reframes sparse view synthesis by leveraging covisibility maps for adaptive supervision, enabling balanced guidance across different uncertainty levels within the scene." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.214, + 0.905, + 0.319 + ], + "angle": 0, + "content": "2. Enhanced point cloud initialization: By applying learned scaling parameters between monocular depth estimates and triangulated points registered using the given COLMAP camera poses, CoMapGS densely enhances initial point cloud quality, increasing the potential to densify sparse and underrepresented regions, even in mono-view areas." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.32, + 0.905, + 0.409 + ], + "angle": 0, + "content": "3. Weighted supervision using covisibility map: This strategy adaptively reinforces high- and low-uncertainty regions, strengthening supervision based on region-specific covisibility levels and stabilizing training with learned geometric constraints that extend beyond the view frustum." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.153, + 0.905, + 0.409 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.411, + 0.905, + 0.473 + ], + "angle": 0, + "content": "To the best of our knowledge, CoMapGS is the first approach to explicitly address and recover high-uncertainty regions in sparse view synthesis, effectively balancing supervision across regions with varying uncertainty levels." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.483, + 0.663, + 0.499 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.509, + 0.907, + 0.78 + ], + "angle": 0, + "content": "Sparse novel view synthesis. The growing interest in reconstructing spatial photographs from minimal input image sets has driven the development of various methods aimed at generating novel views from just a few shots, utilizing approaches based on NeRF [6, 11, 13, 21, 23, 26, 28, 30, 31, 36] and 3DGS [17, 22, 32, 34, 38, 41]. Due to the inherent challenges of sparse view synthesis, such as the limited number of training views, NeRF-based studies have primarily focused on proposing regularization techniques with depth ranking [6, 15, 23, 26, 28, 31], frequency-based loss [36], semantic consistency with visual encoders [11], between-view smoothness [13], or uncertainty-aware adaptive loss [33] for both seen and unseen views [11, 15, 21, 30]. Some methods have also addressed uncertainty issues [13, 36], particularly for high-uncertainty mono-view or unseen regions. However, these methods either disregard [13] or penalize [36] uncertain areas, rather than fully utilizing the information contained within." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.909, + 0.903 + ], + "angle": 0, + "content": "While recent 3DGS-based methods continue to rely on regularization techniques, such as depth constraints [17, 22, 32, 34, 41], some have additionally introduced novel densification strategies [38, 41] and explored enhancing initial point clouds [4, 7, 22, 34], capitalizing on the technical characteristics of 3DGS. These methods primarily focus on improving effective multiview regions [9], often biased towards well-represented regions in the training views, rather" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "26780" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.096, + 0.089, + 0.904, + 0.311 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.32, + 0.908, + 0.381 + ], + "angle": 0, + "content": "Figure 2. Overview of the proposed CoMapGS. CoMapGS leverages existing modules, including Structure-from-Motion (SfM), monocular depth, and dense correspondence prediction, to produce preliminary outputs. Covisibility maps are generated and used alongside module outputs to enhance initial point clouds, which initialize Gaussians for CoMapGS. During training, covisibility map-based weighted supervision with the proximity classifier applies varied strengths to Gaussians based on region-specific covisibility." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.403, + 0.486, + 0.571 + ], + "angle": 0, + "content": "than addressing recovery strategies for high-uncertainty mono-view regions. To overcome the inherent limitations of sparse training views, several prior works have explored generating and utilizing pseudo-ground truths for novel viewpoints via mono-depth estimation [41], forward warping with reversed bilinear sampling [34], or diffusion models with CLIP embedding [30]. While these approaches compensate for the lack of training views, they still face unique challenges in recovering unseen or mono-view regions from the original training view images, potentially misleading the learning process in high-uncertainty areas." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.599, + 0.486, + 0.903 + ], + "angle": 0, + "content": "Covisibility-aware supervision. Beginning with Dy-Check [9], which emphasizes the importance of covisibility in multiview regions to address fair evaluation in dynamic novel view synthesis, recent studies have explored integrating covisibility information into supervision strategies. NeRFVS [35] uses a view coverage map to enforce stricter regularization in low-coverage (covisible) regions, but it is primarily suited for dense view synthesis with video input, where extensive multiview overlap is expected. Similarly, InfoNorm [29] leverages mutual information among surface normals at highly correlated scene points to regularize geometric modeling. However, InfoNorm captures covisible areas within local subregions in structured scenes (e.g., rooms) rather than using a comprehensive covisibility map, which limits its applicability for addressing both multiview and monoview regions. While these methods demonstrate the value of covisibility in supervision, they lack a framework that addresses the challenges of sparse novel view synthesis for natural scenes, including outdoor environments, through a pixel-wise covisibility map." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.402, + 0.898, + 0.42 + ], + "angle": 0, + "content": "3. Covisibility Map-based Gaussian Splitting" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.427, + 0.907, + 0.595 + ], + "angle": 0, + "content": "The proposed CoMapGS framework is illustrated in Fig. 2. CoMapGS builds on the latest 3DGS-based sparse novel view synthesis method, CoR-GS [38], with two key improvements: (1) updating initial point clouds specifically for sparse and mono-view regions and (2) applying adaptive supervision across varying covisibility regions using a covisibility map and proximity MLP classifier. We begin by generating covisibility maps using the dense correspondence prediction model MASt3R [16], refining it with morphological operations (e.g., erosion and dilation) to support these strategies effectively." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.602, + 0.771, + 0.619 + ], + "angle": 0, + "content": "3.1. Covisibility Map Generation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.624, + 0.909, + 0.779 + ], + "angle": 0, + "content": "We generate a covisibility map by identifying dense correspondences between each training view image and all other training view images. Given a set of training view images \\( T = \\{I_1, I_2, \\ldots, I_n\\} \\), where \\( n \\) is the total number of training views, we compute a covisibility map \\( M_i \\) for each view \\( I_i \\). For every pair \\( (I_i, I_{\\{j\\}}) \\) with \\( j \\neq i \\), dense correspondences \\( C_i^j \\) are predicted by a dense correspondence prediction function \\( f_{\\mathrm{cor}} \\), where \\( C_i^j = f_{\\mathrm{cor}}(I_i, I_j) \\). The covisibility map \\( M_i \\) is initialized as a zero matrix of size \\( W \\times H \\) and defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.536, + 0.786, + 0.907, + 0.829 + ], + "angle": 0, + "content": "\\[\nM _ {i} (x, y) = \\sum_ {j = 1, j \\neq i} ^ {n} \\delta_ {(x, y) \\in \\mathcal {P} \\left(C _ {i} ^ {j}\\right)}, \\quad \\forall (x, y) \\in I _ {i}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.836, + 0.911, + 0.903 + ], + "angle": 0, + "content": "where \\(\\delta_{(x,y)\\in \\mathcal{P}(C_i^j)}\\) is an indicator variable returning 1 if pixel \\((x,y)\\) in \\(I_{i}\\) has a correspondence match in \\(I_{j}\\), and 0 otherwise. Here, \\(\\mathcal{P}(C_i^j)\\) represents the set of pixel coordinates in \\(I_{i}\\) that have correspondences in \\(I_{j}\\)." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.958 + ], + "angle": 0, + "content": "26781" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.243 + ], + "angle": 0, + "content": "The resulting covisibility map \\( M_{i} \\) captures pixel-wise covisibility counts ranging from 0 to \\( n - 1 \\), indicating the frequency with which a pixel appears across views, as shown in Fig. 1, with values rescaled from 0 (black) to 255 (white) for visualization purposes. Unlike DyCheck [9]'s covisibility mask, which only indicates pixel visibility, our covisibility map provides precise covisibility details (i.e., covisible view count) at the pixel level. This map supports both the initial point cloud update (Sec. 3.2) and the weighted supervision method (Sec. 3.3)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.25, + 0.367, + 0.266 + ], + "angle": 0, + "content": "3.2. Enhancing Initial Point Clouds" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.272, + 0.485, + 0.501 + ], + "angle": 0, + "content": "In this subsection, we explain how to improve initial point clouds for both low-uncertainty multi-view and high-uncertainty mono-view regions. To enhance initial point clouds in low-uncertainty multi-view regions where sparse SfM points are available, we integrate a triangulated set of 3D points, \\( P_T \\), derived from dense correspondences obtained from MASt3R [16], to complement COLMAP's point cloud, \\( P_C \\). This approach enhances geometrically plausible structures, especially in regions where COLMAP's keypoint-based matching is sparse. For each view \\( I_i \\), we generate \\( P_T \\) by triangulating dense correspondences \\( C_i^{\\{j\\}} \\) from view \\( I_i \\) to other views \\( I_{\\{j\\}} \\) (e.g., two or three correspondences in a 3-view case). To avoid redundancy, correspondences already reconstructed from other views are excluded from subsequent triangulations." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.501, + 0.446, + 0.516 + ], + "angle": 0, + "content": "Formally, we define the updated point cloud \\(P_{\\mathrm{u}}\\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.145, + 0.522, + 0.483, + 0.54 + ], + "angle": 0, + "content": "\\[\nP _ {\\mathrm {u}} = P _ {C} \\cup \\left\\{p _ {T} \\in P _ {T} \\mid D \\left(p _ {T}, P _ {C}\\right) > \\epsilon \\right\\}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.546, + 0.484, + 0.667 + ], + "angle": 0, + "content": "where \\( D(p_{T}, P_{C}) \\) is the distance from a point \\( p_{T} \\) in \\( P_{T} \\) to its nearest neighbor in \\( P_{C} \\), and \\( \\epsilon \\) is a threshold ensuring that only points sufficiently distant from existing points in \\( P_{C} \\) are added. This selective integration allows the triangulated points to supplement COLMAP's sparse initial point cloud in the same coordinate system, filling gaps in low-uncertainty regions and improving overall density in covisible areas, as shown in Fig. 3." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.668, + 0.484, + 0.879 + ], + "angle": 0, + "content": "In high-uncertainty mono-view regions, we use monocular depth estimation [10] to generate a preliminary set of 3D points, \\( P_{d} \\), by unprojecting depth estimates using the provided camera intrinsics. This set is defined as \\( P_{d} = P_{d}^{\\mathrm{low}} \\cup P_{d}^{\\mathrm{high}} \\), where \\( P_{d}^{\\mathrm{low}} \\) comprises points unprojected from multiview regions (where \\( M_{i} \\geq 1 \\)), and \\( P_{d}^{\\mathrm{high}} \\) comprises points unprojected from mono-view regions (where \\( M_{i} = 0 \\)). To align these randomly scaled unprojected points with triangulated points from multiview geometry, we learn an anisotropic scaling transformation \\( f_{\\mathrm{scale}} \\), using covisible points from \\( P_{u}^{\\mathrm{low}} \\subseteq P_{u} \\) as ground truth and \\( P_{d}^{\\mathrm{low}} \\) as input. This transformation applies scaling parameters learned through linear regression to match the multiview-derived geometry:" + }, + { + "type": "equation", + "bbox": [ + 0.217, + 0.884, + 0.483, + 0.903 + ], + "angle": 0, + "content": "\\[\nf _ {\\text {s c a l e}} \\left(P _ {d} ^ {\\text {l o w}}\\right) \\approx P _ {u} ^ {\\text {l o w}}. \\tag {3}\n\\]" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.088, + 0.641, + 0.167 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.57, + 0.168, + 0.587, + 0.178 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.641, + 0.088, + 0.77, + 0.167 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.699, + 0.168, + 0.715, + 0.178 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.772, + 0.088, + 0.901, + 0.168 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.828, + 0.168, + 0.843, + 0.178 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.193, + 0.906, + 0.263 + ], + "angle": 0, + "content": "Figure 3. Examples of enhanced initial PCLs from the Fern scene in the LLFF dataset. (a) Initial SfM points are first (b) updated through triangulation using dense correspondences and COLMAP camera poses, and then (c) fully refined to include mono-view regions (mono-view updates are colored in red for visualization)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.286, + 0.906, + 0.318 + ], + "angle": 0, + "content": "For consistency, we define \\( P_{u} = P_{u}^{\\mathrm{low}} \\cup P_{u}^{\\mathrm{high}} \\), where \\( P_{u}^{\\mathrm{low}} \\) is selected based on covisibility:" + }, + { + "type": "equation", + "bbox": [ + 0.591, + 0.327, + 0.906, + 0.345 + ], + "angle": 0, + "content": "\\[\nP _ {u} ^ {\\text {l o w}} = \\left\\{p \\mid M _ {i} \\left(\\pi \\left(P _ {u}, \\mathbf {H} _ {i}\\right)\\right) \\geq 1 \\right\\}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.354, + 0.906, + 0.415 + ], + "angle": 0, + "content": "where \\(\\pi(P_u, \\mathbf{H}_i)\\) denotes the projection of \\(P_u\\) using the camera transformation matrix \\(\\mathbf{H}_i\\), yielding the image coordinates \\((x, y)\\) for a given camera viewpoint and ensuring inclusion of only points visible in multiple views." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.416, + 0.906, + 0.46 + ], + "angle": 0, + "content": "After training, we apply \\( f_{\\mathrm{scale}} \\) to the randomly scaled unprojected points from depth estimates in high-uncertainty mono-view regions:" + }, + { + "type": "equation", + "bbox": [ + 0.636, + 0.468, + 0.905, + 0.488 + ], + "angle": 0, + "content": "\\[\nP _ {s} ^ {\\text {h i g h}} = f _ {\\text {s c a l e}} \\left(P _ {d} ^ {\\text {h i g h}}\\right). \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.498, + 0.906, + 0.562 + ], + "angle": 0, + "content": "This transformation aligns the randomly scaled points \\( P_{d}^{\\mathrm{high}} \\) to the scale of COLMAP's triangulated points \\( P_{s}^{\\mathrm{high}} \\), ensuring spatial consistency across regions. The final point cloud, \\( P_{\\mathrm{final}} \\), is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.636, + 0.57, + 0.905, + 0.59 + ], + "angle": 0, + "content": "\\[\nP _ {\\text {f i n a l}} = P _ {u} ^ {\\text {l o w}} \\cup P _ {s} ^ {\\text {h i g h}}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.599, + 0.906, + 0.659 + ], + "angle": 0, + "content": "resulting in enhanced initial point cloud density and accuracy across both multiview and mono-view regions, as shown in Fig. 3, thus supporting cohesive scene reconstruction in sparse view synthesis." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.667, + 0.903, + 0.684 + ], + "angle": 0, + "content": "3.3. Covisibility Map-Based Weighted Supervision" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.906, + 0.841 + ], + "angle": 0, + "content": "We propose a covisibility map-based supervision strategy that leverages the enhanced initial point cloud, \\( P_{\\mathrm{final}} \\), and the covisibility map \\( M_{i} \\) to adaptively supervise regions based on uncertainty levels represented by covisibility counts, as shown in Fig. 4. This approach improves conventional sparse view Gaussian splatting methods [38, 41] by introducing region-specific weighting across covisibility levels, \\( M_{i} = \\{0:n - 1\\} \\), and by supervising Gaussians located outside the training view frustum with a scene-level weight based on the average covisibility score \\( S = \\mathrm{AVG}(M_{[1:n]}) \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.841, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Our approach integrates seamlessly with the latest 3DGS-based sparse view synthesis methods [38, 41] by adding a proximity loss term, weighted using inverse proximity scores predicted by a three-layer MLP classifier" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "26782" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.12 + ], + "angle": 0, + "content": "trained on \\( P_{\\mathrm{final}} \\) (see details in the supplementary materials)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.122, + 0.484, + 0.228 + ], + "angle": 0, + "content": "Proximity loss with adaptive weighting. To implement this approach, we first train an MLP classifier \\( f_{p} \\) on \\( P_{\\mathrm{final}} \\), labeling these curated scene 3D points as positive (representing the original scene geometry) and randomly distributed points, \\( P_{\\mathrm{random}} \\), constrained to remain distant from \\( P_{\\mathrm{final}} \\), as negative. The classifier \\( f_{p}: \\mathbb{R}^{3} \\to [0,1] \\) outputs a proximity score \\( s \\) for each 3D point \\( p \\):" + }, + { + "type": "equation", + "bbox": [ + 0.221, + 0.237, + 0.483, + 0.254 + ], + "angle": 0, + "content": "\\[\ns = f _ {p} (p) \\in [ 0, 1 ], \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.262, + 0.484, + 0.324 + ], + "angle": 0, + "content": "where \\(s\\) approaches 1 for points close to the scene geometry and 0 for distant points. Using batch processing, we compute the proximity scores efficiently for the entire Gaussian set, \\(G\\), and define the proximity loss \\(\\mathcal{L}_p\\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.33, + 0.483, + 0.368 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {p} = \\frac {1}{| G |} \\sum_ {g \\in G} (\\chi (g) w _ {\\text {i n}} + (1 - \\chi (g)) w _ {\\text {o u t}}) \\cdot (1 - s), \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.375, + 0.484, + 0.421 + ], + "angle": 0, + "content": "where \\( g \\) denotes the 3D position of a Gaussian in the set \\( G \\), representing the complete Gaussian set. The indicator function \\( \\chi(g) \\) is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.134, + 0.429, + 0.483, + 0.469 + ], + "angle": 0, + "content": "\\[\n\\chi (g) = \\left\\{ \\begin{array}{l l} 1, & \\text {i f} \\pi (g, \\mathbf {H} _ {i}) \\in [ 0, \\mathsf {W}) \\times [ 0, \\mathsf {H}) \\\\ 0, & \\text {o t h e r w i s e} \\end{array} , \\right. \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.477, + 0.483, + 0.507 + ], + "angle": 0, + "content": "where \\(\\pi(g, \\mathbf{H}_i)\\) is the projection of \\(g\\) under camera transformation \\(\\mathbf{H}_i\\) onto the可视ibility map \\(M_i\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.507, + 0.484, + 0.551 + ], + "angle": 0, + "content": "Within the camera-view frustum. We use the weight \\( w_{in} \\) for Gaussians projected onto covisibility mask regions, defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.196, + 0.557, + 0.483, + 0.589 + ], + "angle": 0, + "content": "\\[\nw _ {\\mathrm {i n}} = \\frac {1}{M _ {i} \\left(\\pi \\left(g , \\mathbf {H} _ {i}\\right)\\right) + 1}, \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.596, + 0.483, + 0.687 + ], + "angle": 0, + "content": "where \\( M_{i}(\\pi (g,\\mathbf{H}i)) \\) is the covisibility count of \\( g \\) in the covisibility map for the corresponding projection. This weighting ensures that high-covisibility regions, which are frequently supervised by other loss terms, receive weaker proximity-related supervision, while mono-view regions benefit from a more direct application of \\( \\mathcal{L}_p \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.687, + 0.484, + 0.779 + ], + "angle": 0, + "content": "Outside the camera-view frustum. In conventional 3DGS-based methods, only Gaussians within the view frustum are supervised per iteration. However, for scenes with high covisibility scores (\\(S > 0.7\\)), we extend supervision to Gaussians located outside the view frustum, applying a scene-level weight \\(w_{out}\\) for these points:" + }, + { + "type": "equation", + "bbox": [ + 0.188, + 0.785, + 0.483, + 0.819 + ], + "angle": 0, + "content": "\\[\nw _ {o u t} = \\max \\left(0, \\frac {S - 0 . 7}{0 . 3}\\right), \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.484, + 0.9 + ], + "angle": 0, + "content": "which scales linearly from 1 to 0 as \\( S \\) ranges from 1 to 0.7. This approach improves supervision outside the view frustum in high-covisibility score scenes, ensuring comprehensive guidance while making efficient use of the proximity classifier." + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.09, + 0.903, + 0.237 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.246, + 0.907, + 0.303 + ], + "angle": 0, + "content": "Figure 4. Alongside (a) the conventional reconstruction loss \\(\\mathcal{L}1\\), we add (b) a weighted proximity loss \\(\\mathcal{L}_p\\) that applies classifier results with assigned weights based on Gaussian projections onto the可视ability map and their positions relative to the frustum." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.33, + 0.698, + 0.347 + ], + "angle": 0, + "content": "3.4. Objective Function" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.353, + 0.907, + 0.444 + ], + "angle": 0, + "content": "Our proposed CoMapGS enhances the objective function by introducing a proximity loss term, effectively complementing the function while preserving the original strategies of state-of-the-art methods, such as Co-pruning and Pseudoview Co-regularization from CoR-GS [38]. The full objective function is expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.534, + 0.457, + 0.906, + 0.475 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = (1 - \\lambda) \\mathcal {L} 1 \\left(I, I ^ {*}\\right) + \\lambda \\mathcal {L} _ {\\mathrm {D} - \\mathrm {S S I M}} \\left(I, I ^ {*}\\right) + \\mathcal {L} _ {p}, \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.487, + 0.906, + 0.593 + ], + "angle": 0, + "content": "where \\(I\\) is the rendered image, supervised against ground truth \\(I^{*}\\), \\(\\mathcal{L}1\\) is the L1 reconstruction loss, and \\(\\mathcal{L}_{\\mathrm{D - SSIM}}\\) represents the D-SSIM loss term. The balancing parameter \\(\\lambda\\) controls the weighting between L1 and D-SSIM, in line with standard Gaussian splattering objectives. The proximity loss term \\(\\mathcal{L}_p\\), defined in Eq. 8, adds adaptive supervision to improve geometry alignment." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.594, + 0.906, + 0.655 + ], + "angle": 0, + "content": "Our proposed approach extends seamlessly to other recent sparse view synthesis methods, such as FSGS [41], by incorporating \\(\\mathcal{L}_p\\) into their objective function while preserving their original strategies, as reported in Table 1." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.67, + 0.719, + 0.687 + ], + "angle": 0, + "content": "4. Experimental Results" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.696, + 0.596, + 0.712 + ], + "angle": 0, + "content": "4.1. Setup" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.719, + 0.906, + 0.824 + ], + "angle": 0, + "content": "Datasets. We conducted experiments on two benchmark datasets: LLFF [19], using 3, 6, and 9 training views with resolutions downsampled to \\(8 \\times\\), and Mip-NeRF 360 [2], using 12 and 24 training views per scene with resolutions downsampled to \\(4 \\times\\). We adhered to the standard training and testing conventions [38] for datasets commonly used in sparse view synthesis." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Comparison Baselines. We compare CoMapGS with existing few-shot novel view synthesis methods, including FSGS [41] and CoR-GS [38]. As our method primarily enhances the performance of 3DGS-based approaches, we reproduced results for these methods. For other methods, we" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "26783" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.908, + 0.119 + ], + "angle": 0, + "content": "Table 1. Quantitative results on LLFF with 3, 6, and 9 training views. The best, second-best, and third-best entries are marked in red, orange, and yellow, respectively. \\(\\dagger\\) indicates results reproduced by us using the released code base." + }, + { + "type": "table", + "bbox": [ + 0.107, + 0.129, + 0.894, + 0.401 + ], + "angle": 0, + "content": "
MethodPSNR ↑SSIM ↑LPIPS ↓
3-view6-view9-view3-view6-view9-view3-view6-view9-view
DietNeRF [11]14.9421.7524.280.3700.7170.8010.4960.2480.183
Mip-NeRF [1]16.1122.9124.880.4010.7560.8260.4600.2130.160
RegNeRF [21]19.0823.1024.860.5870.7600.8200.3360.2060.161
SimpleNeRF [25]19.2423.0523.980.6230.7370.7620.3750.2960.286
FreeNeRF [36]19.6323.7325.130.6120.7790.8270.3080.1950.160
SparseNeRF [28]19.8623.2624.270.7140.7410.7810.2430.2350.228
DiffusioNeRF [31]20.1323.6024.620.6310.7750.8070.3440.2350.216
ReconFusion [30]21.3424.2525.210.7240.8150.8480.2030.1520.134
DNGaussian [17]19.1222.1823.170.5910.7550.7880.2940.1980.180
3DGS [12]19.2223.8025.440.6490.8140.8600.2290.1250.096
FSGS [41]†20.45824.69426.0060.7130.8390.8750.2050.1210.095
CoR-GS [38]†20.47324.77726.4750.7170.8440.8810.1990.1160.086
[41]† + CoMapGS (ours)20.89124.80826.2920.7370.8440.8790.1900.1170.092
[38]† + CoMapGS (ours)21.10525.20426.7310.7470.8540.8860.1820.1080.082
" + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.413, + 0.257, + 0.508 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.509, + 0.257, + 0.602 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.603, + 0.257, + 0.696 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.159, + 0.699, + 0.195, + 0.71 + ], + "angle": 0, + "content": "FSGS" + }, + { + "type": "image", + "bbox": [ + 0.259, + 0.413, + 0.418, + 0.509 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.259, + 0.509, + 0.418, + 0.602 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.259, + 0.603, + 0.418, + 0.696 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.314, + 0.699, + 0.362, + 0.71 + ], + "angle": 0, + "content": "CoR-GS" + }, + { + "type": "image", + "bbox": [ + 0.419, + 0.413, + 0.58, + 0.509 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.419, + 0.509, + 0.579, + 0.602 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.419, + 0.603, + 0.58, + 0.696 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.485, + 0.7, + 0.514, + 0.71 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.581, + 0.413, + 0.741, + 0.509 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.581, + 0.509, + 0.74, + 0.602 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.581, + 0.603, + 0.74, + 0.696 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.623, + 0.7, + 0.7, + 0.71 + ], + "angle": 0, + "content": "Ground Truth" + }, + { + "type": "image", + "bbox": [ + 0.742, + 0.413, + 0.903, + 0.509 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.742, + 0.509, + 0.903, + 0.602 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.742, + 0.602, + 0.903, + 0.696 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.771, + 0.699, + 0.872, + 0.711 + ], + "angle": 0, + "content": "Covisibility Mask" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.721, + 0.908, + 0.765 + ], + "angle": 0, + "content": "Figure 5. Qualitative comparison on 3 training views from the LLFF dataset. We compare our results with FSGS [41] and CoR-GS [38], demonstrating that our method achieves superior texture synthesis in novel views, particularly in high-uncertainty mono-view regions that appear only once in the training set." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.79, + 0.483, + 0.82 + ], + "angle": 0, + "content": "report the best published scores from prior works, as shown in Tables 1 and 2." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.822, + 0.483, + 0.853 + ], + "angle": 0, + "content": "Evaluation Metrics. We evaluate our results using PSNR, SSIM, and LPIPS metrics." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Implementation. Training was conducted for 10k iterations on the LLFF dataset and 30k iterations on the MipNeRF 360 dataset, following the setup of 3DGS [12]. Con" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.79, + 0.907, + 0.837 + ], + "angle": 0, + "content": "sistent with experimental setups in FSGS [41] and CoR-GS [38], CoMapGS was initialized with point clouds and precomputed camera poses from COLMAP." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.848, + 0.655, + 0.864 + ], + "angle": 0, + "content": "4.2. Comparisons" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.908, + 0.902 + ], + "angle": 0, + "content": "LLFF. Quantitative and qualitative results on the LLFF dataset are presented in Table 1 and Fig. 5, respectively." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "26784" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.908, + 0.119 + ], + "angle": 0, + "content": "Table 2. Quantitative results on Mip-NeRF360 with 12 and 24 training views. The best, second-best, and third-best entries are marked in red, orange, and yellow, respectively. \\( {}^{ \\dagger } \\) indicates results reproduced by us using the released code base." + }, + { + "type": "table", + "bbox": [ + 0.185, + 0.129, + 0.817, + 0.238 + ], + "angle": 0, + "content": "
MethodPSNR ↑SSIM ↑LPIPS ↓
12-view24-view12-view24-view12-view24-view
3DGS [12]18.5222.800.5230.7080.4150.276
FSGS [41]18.8023.280.5310.7150.4180.274
CoR-GS [38]†19.16223.3210.5740.7290.4140.271
[38]† + CoMapGS (ours)19.68023.4620.5910.7340.3940.264
" + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.248, + 0.903, + 0.678 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.687, + 0.908, + 0.729 + ], + "angle": 0, + "content": "Figure 6. Qualitative comparison on 12 training views from the Mip-NeRF 360 dataset. Mip-NeRF 360 highlights diverse covisibility levels, with high-uncertainty regions sparsely captured. CoMapGS effectively synthesizes these regions, while CoR-GS struggles to recover textures." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.756, + 0.486, + 0.892 + ], + "angle": 0, + "content": "We reproduced evaluation scores for the two baseline methods, FSGS [41] and CoR-GS [38], to highlight the improvements over our reproduced results. The results show consistent improvements, ranking first across various training view counts, with PSNR placing second only to ReconFusion [30] in the 3-view case. This difference is likely due to diffusion-based sparse view synthesis methods, which are optimized for generating visually seamless colors that blend with surrounding regions. However, this approach may not" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.756, + 0.907, + 0.862 + ], + "angle": 0, + "content": "accurately capture original patterns in mono-view or unseen regions—common in sparse view synthesis datasets, especially with fewer training views. Consequently, while we ranked second in PSNR, our method achieved substantially higher SSIM and lower LPIPS scores under comparable view settings, indicating improved structural fidelity and perceptual quality in the synthesized views." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Mip-NeRF 360. To validate CoMapGS's applicability to broader scenes that capture large areas with outward-" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26785" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.242 + ], + "angle": 0, + "content": "facing cameras, we evaluated our method using the MipNeRF 360 dataset. This dataset, which centers objects of interest with outward-facing cameras, captures the motivation of our approach by featuring distinct low-uncertainty, multiview regions alongside high-uncertainty, less-covisible regions within an image, as shown in Fig. 1. In these scenarios, COLMAP struggles to establish stable correspondences, leading to sparse initial point clouds that fail to adequately recover details in high-uncertainty regions, as shown in Fig. 6." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.243, + 0.483, + 0.424 + ], + "angle": 0, + "content": "In contrast, CoMapGS begins with enhanced initial point clouds, preserving the scene's original structure while benefiting from proximity loss supervision, resulting in improved texture representation over conventional methods, as shown in Fig. 6. Quantitative results in Table 2 confirm significant improvements across metrics, particularly LPIPS, compared to recent state-of-the-art methods, which perform worse than their baselines (e.g., 3DGS [12], FSGS [41]). In the 24-view case, while PSNR improvements may appear modest due to unseen regions in test views, the LPIPS gain highlights our method's effectiveness in recovering textures from both low- and high-uncertainty regions in novel view." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.436, + 0.245, + 0.452 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.458, + 0.483, + 0.579 + ], + "angle": 0, + "content": "The ablation study in Table 3 highlights that dense reconstruction in low-uncertainty (covisible) regions already provides significant improvement over baseline methods, underscoring the benefit of increased point density in covisible areas. Furthermore, adding points from mono-view regions to the enhanced point cloud yields additional gains, demonstrating the value of incorporating points from underrepresented, high-uncertainty regions in sparse view synthesis." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.581, + 0.483, + 0.791 + ], + "angle": 0, + "content": "We also evaluated the impact of our proximity loss-based weighted supervision both with and without the initial point cloud enhancement strategy. Interestingly, starting from sparse COLMAP point clouds without updates showed limited improvement, likely because 3DGS-based methods tend to densify Gaussians near their original positions. However, when initialized with our enhanced point clouds, the benefits of weighted supervision became evident, delivering significant performance gains in both partial (multiview region) and full (multi- and mono-view region) update scenarios. This demonstrates that enhancing initial point clouds and incorporating proximity loss-based weighted supervision work synergistically to optimize 3DGS-based sparse view synthesis." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.803, + 0.21, + 0.818 + ], + "angle": 0, + "content": "4.4. Discussion" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Definition of sparse view synthesis. Although sparse novel view synthesis is often defined by a specific number of training views (e.g., LLFF with 3, 6, or 9 views; Mip-NeRF with 12 or 24 views), we argue that the true challenge lies in reconstructing underrepresented regions with low covisibility" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.089, + 0.905, + 0.16 + ], + "angle": 0, + "content": "Table 3. Ablation study of the proposed CoMapGS with six training views. \\(\\triangle\\) indicates enhancing initial point clouds only in the low-uncertainty region (where covisible count \\(\\geq 2\\)), representing a straightforward application of dense point clouds generated by conventional methods such as MAST3R [16]." + }, + { + "type": "table", + "bbox": [ + 0.528, + 0.17, + 0.895, + 0.31 + ], + "angle": 0, + "content": "
Proposed methodLLFF
Sec. 3.2Sec. 3.3PSNR ↑SSIM ↑LPIPS ↓
CoR-GS [38]†24.7770.8440.116
✓24.7870.8450.116
△24.900.8490.112
△✓25.1530.8540.109
✓25.0760.8520.109
✓✓25.2040.8540.108
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.342, + 0.907, + 0.584 + ], + "angle": 0, + "content": "counts rather than the total view count. This problem is inherently region-wise and imbalanced, where certain areas appear in only a few views, leading to sparse representation even when many views are available. For example, as DyCheck [9] categorizes scene types based on effective multiview factors (EMFs), scenarios with high frame counts but low region-specific可视ibility can still be considered sparse view cases. We address this by applying region-wise adaptive supervision based on可视ibility counts, improving results, especially for imbalanced datasets like MipNeRF 360 (12 views). In such cases, central objects are densely captured, while outer areas show low可视ibility due to the 360-degree, outward-facing capture setup. This issue is common in real-world scene capture, where uneven可视ibility across views motivates balanced,可视ibility-aware supervision." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.609, + 0.64, + 0.625 + ], + "angle": 0, + "content": "5. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.637, + 0.907, + 0.889 + ], + "angle": 0, + "content": "This paper introduces CoMapGS, a Covisibility Map-based Gaussian Splitting approach designed to enhance sparse view synthesis methods by addressing region-specific uncertainty levels. We generate covisibility maps for a given non-continuous set of distant images, using them to identify low- and high-uncertainty regions and to improve point cloud initialization beyond conventional methods like COLMAP. By introducing covisibility-adaptive weighted supervision, CoMapGS effectively manages varying sparsity across both multiview and mono-view regions, consistently outperforming state-of-the-art 3DGS-based sparse novel view synthesis methods. Experimental results validate CoMapGS's effectiveness across diverse, realistic scenes, including both indoor and outdoor environments and various scales of captured areas. CoMapGS demonstrates superiority in handling complex region-wise covisibility, highlighting its robustness and adaptability in challenging sparse view synthesis tasks." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26786" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.091, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.116, + 0.484, + 0.171 + ], + "angle": 0, + "content": "[1] Jonathan T. Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P. Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In ICCV, pages 5855–5864, 2021. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.172, + 0.484, + 0.242 + ], + "angle": 0, + "content": "[2] Jonathan T. Barron, Ben Mildenhall, Dor Verbin, Pratul P. Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5470-5479, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.242, + 0.483, + 0.311 + ], + "angle": 0, + "content": "[3] Anpei Chen, Zexiang Xu, Fuqiang Zhao, Xiaoshuai Zhang, Fanbo Xiang, Jingyi Yu, and Hao Su. Mvsnerf: Fast generalizable radiance field reconstruction from multi-view stereo. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 14124-14133, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.312, + 0.483, + 0.352 + ], + "angle": 0, + "content": "[4] Shan Chen, Jiale Zhou, and Lei Li. Dense point clouds matter: Dust-gs for scene reconstruction from sparse viewpoints, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.354, + 0.483, + 0.437 + ], + "angle": 0, + "content": "[5] Wenyan Cong, Hanxue Liang, Peihao Wang, Zhiwen Fan, Tianlong Chen, Mukund Varma, Yi Wang, and Zhangyang Wang. Enhancing nerf akin to enhancing llms: Generalizable nerf transformer with mixture-of-view-experts. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 3193-3204, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.438, + 0.483, + 0.507 + ], + "angle": 0, + "content": "[6] Kangle Deng, Andrew Liu, Jun-Yan Zhu, and Deva Ramanan. Depth-supervised nerf: Fewer views and faster training for free. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12882-12891, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.508, + 0.483, + 0.577 + ], + "angle": 0, + "content": "[7] Zhiwen Fan, Wenyan Cong, Kairun Wen, Kevin Wang, Jian Zhang, Xinghao Ding, Danfei Xu, Boris Ivanovic, Marco Pavone, Georgios Pavlakos, Zhangyang Wang, and Yue Wang. Instantsplat: Unbounded sparse-view pose-free gaussian splatting in 40 seconds, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.579, + 0.483, + 0.62 + ], + "angle": 0, + "content": "[8] Guangchi Fang and Bing Wang. Mini-Splatting: Representing scenes with a constrained number of gaussians. In ECCV, Milano, Italy, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.621, + 0.483, + 0.662 + ], + "angle": 0, + "content": "[9] Hang Gao, Ruilong Li, Shubham Tulsiani, Bryan Russell, and Angjoo Kanazawa. Monocular dynamic view synthesis: A reality check. In NeurIPS, 2022. 2, 3, 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.663, + 0.483, + 0.746 + ], + "angle": 0, + "content": "[10] Mu Hu, Wei Yin, Chi Zhang, Zhipeng Cai, Xiaoxiao Long, Hao Chen, Kaixuan Wang, Gang Yu, Chunhua Shen, and Shaojie Shen. Metric3d v2: A versatile monocular geometric foundation model for zero-shot metric depth and surface normal estimation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(12):10579-10596, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.747, + 0.483, + 0.803 + ], + "angle": 0, + "content": "[11] Ajay Jain, Matthew Tancik, and Pieter Abbeel. Putting nerf on a diet: Semantically consistent few-shot view synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 5885-5894, 2021. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.804, + 0.483, + 0.858 + ], + "angle": 0, + "content": "[12] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42 (4), 2023. 1, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.86, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[13] Mijeong Kim, Seonguk Seo, and Bohyung Han. Infonerf: Ray entropy minimization for few-shot neural volume rendering. In Proceedings of the IEEE/CVF Conference on" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.116, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.093, + 0.905, + 0.119 + ], + "angle": 0, + "content": "Computer Vision and Pattern Recognition (CVPR), pages 12912-12921, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.121, + 0.905, + 0.176 + ], + "angle": 0, + "content": "[14] Jonás Kulhánek, Erik Derner, Torsten Sattler, and Robert Babuška. Viewformer: Nerf-free neural rendering from few images using transformers. In European Conference on Computer Vision (ECCV), 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.177, + 0.905, + 0.232 + ], + "angle": 0, + "content": "[15] Min-Seop Kwak, Jiuhn Song, and Seungryong Kim. Geconerf: few-shot neural radiance fields via geometric consistency. In Proceedings of the 40th International Conference on Machine Learning. JMLR.org, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.233, + 0.905, + 0.26 + ], + "angle": 0, + "content": "[16] Vincent Leroy, Yohann Cabon, and Jerome Revaud. Grounding image matching in 3d with mast3r, 2024. 2, 3, 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.261, + 0.905, + 0.343 + ], + "angle": 0, + "content": "[17] Jiahe Li, Jiawei Zhang, Xiao Bai, Jin Zheng, Xin Ning, Jun Zhou, and Lin Gu. Dngaussian: Optimizing sparse-view 3d gaussian radiance fields with global-local depth normalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20775-20785, 2024. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.344, + 0.904, + 0.384 + ], + "angle": 0, + "content": "[18] David G. Lowe. Distinctive image features from scale-invariant keypoints. Int. J. Comput. Vision, 60(2):91-110, 2004. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.385, + 0.905, + 0.455 + ], + "angle": 0, + "content": "[19] Ben Mildenhall, Pratul P. Srinivasan, Rodrigo Ortiz-Cayon, Nima Khademi Kalantari, Ravi Ramamoorthi, Ren Ng, and Abhishek Kar. Local light field fusion: Practical view synthesis with prescriptive sampling guidelines. ACM Transactions on Graphics (TOG), 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.456, + 0.905, + 0.511 + ], + "angle": 0, + "content": "[20] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.511, + 0.905, + 0.58 + ], + "angle": 0, + "content": "[21] Michael Niemeyer, Jonathan T. Barron, Ben Mildenhall, Mehdi S. M. Sajjadi, Andreas Geiger, and Noha Radwan. Regnerf: Regularizing neural radiance fields for view synthesis from sparse inputs. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.581, + 0.905, + 0.637 + ], + "angle": 0, + "content": "[22] Avinash Paliwal, Wei Ye, Jinhui Xiong, Dmytro Kotovenko, Rakesh Ranjan, Vikas Chandra, and Nima Khademi Kalantari. CoherentGS: Sparse novel view synthesis with coherent 3d gaussians. In ECCV, Milano, Italy, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.637, + 0.905, + 0.718 + ], + "angle": 0, + "content": "[23] Barbara Roessle, Jonathan T. Barron, Ben Mildenhall, Pratul P. Srinivasan, and Matthias Nießner. Dense depth priors for neural radiance fields from sparse input views. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12892-12901, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.72, + 0.905, + 0.762 + ], + "angle": 0, + "content": "[24] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.762, + 0.905, + 0.816 + ], + "angle": 0, + "content": "[25] Nagabhushan Somraj, Adithyan Karanayil, and Rajiv Soundararajan. SimpleNeRF: Regularizing sparse input neural radiance fields with simpler solutions. In SIGGRAPH Asia, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.818, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[26] Jiuhn Song, Seonghoon Park, Honggyu An, Seokju Cho, Min-Seop Kwak, Sungjin Cho, and Seungryong Kim. Därf: Boosting radiance fields from sparse input views with monocular depth adaptation. In Advances in Neural Information Processing Systems, pages 68458-68470. Curran Associates, Inc., 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26787" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[27] Thomas Tanay and Matteo Maggioni. Global latent neural rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19723-19733, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.219 + ], + "angle": 0, + "content": "[28] Guangcong Wang, Zhaoxi Chen, Chen Change Loy, and Ziwei Liu. Sparsenerf: Distilling depth ranking for few-shot novel view synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 9065-9076, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.221, + 0.482, + 0.262 + ], + "angle": 0, + "content": "[29] Xulong Wang, Siyan Dong, Youyi Zheng, and Yanchao Yang. InfoNorm: Mutual information shaping of normals for sparse-view reconstruction. In ECCV, Milano, Italy, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.264, + 0.482, + 0.359 + ], + "angle": 0, + "content": "[30] Rundi Wu, Ben Mildenhall, Philipp Henzler, Keunhong Park, Ruiqi Gao, Daniel Watson, Pratul P. Srinivasan, Dor Verbin, Jonathan T. Barron, Ben Poole, and Aleksander Ho?y?ski. Reconfusion: 3d reconstruction with diffusion priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 21551-21561, 2024. 1, 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.362, + 0.482, + 0.431 + ], + "angle": 0, + "content": "[31] Jamie Wynn and Daniyar Turmukhambetov. Diffusionerf: Regularizing neural radiance fields with denoising diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4180-4189, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.433, + 0.482, + 0.487 + ], + "angle": 0, + "content": "[32] Haolin Xiong, Sairisheek Muttukuru, Rishi Upadhyay, Pradyumna Chari, and Achuta Kadambi. Sparsegs: Realtime \\(360^{\\circ}\\) sparse view synthesis using gaussian splatting. Arxiv, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.49, + 0.482, + 0.545 + ], + "angle": 0, + "content": "[33] Qingshan Xu, Xuanyu Yi, Jianyao Xu, Wenbing Tao, Yew Soon Ong, and Hanwang Zhang. Few-shot nerf by adaptive rendering loss regularization. In ECCV, Milano, Italy, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.548, + 0.482, + 0.602 + ], + "angle": 0, + "content": "[34] Wangze Xu, Huachen Gao, Shihe Shen, Rui Peng, Jianbo Jiao, and Ronggang Wang. Mvpgs: Excavating multi-view priors for gaussian splatting from sparse input views. In ECCV, Milano, Italy, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.604, + 0.482, + 0.673 + ], + "angle": 0, + "content": "[35] Chen Yang, Peihao Li, Zanwei Zhou, Shanxin Yuan, Bingbing Liu, Xiaokang Yang, Weichao Qiu, and Wei Shen. NeRFVS: Neural Radiance Fields for Free View Synthesis via Geometry Scaffolds. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.675, + 0.482, + 0.73 + ], + "angle": 0, + "content": "[36] Jiawei Yang, Marco Pavone, and Yue Wang. Freenerf: Improving few-shot neural rendering with free frequency regularization. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.732, + 0.482, + 0.8 + ], + "angle": 0, + "content": "[37] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelnerf: Neural radiance fields from one or few images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4578-4587, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.803, + 0.482, + 0.857 + ], + "angle": 0, + "content": "[38] Jiawei Zhang, Jiahe Li, Xiaohan Yu, Lei Huang, Lin Gu, Jin Zheng, and Xiao Bai. Cor-gs: Sparse-view 3d gaussian splatt-ting via co-regularization. In ECCV, Milano, Italy, 2024. 1, 2, 3, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.859, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[39] Zhizhuo Zhou and Shubham Tulsiani. Sparsefusion: Distilling view-conditioned diffusion for 3d reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vi" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.548, + 0.093, + 0.905, + 0.119 + ], + "angle": 0, + "content": "sion and Pattern Recognition (CVPR), pages 12588-12597, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.177 + ], + "angle": 0, + "content": "[40] Haidong Zhu, Tianyu Ding, Tianyi Chen, Ilya Zharkov, Ram Nevatia, and Luming Liang. CaesarNeRF: Calibrated semantic representation for few-shot generalizable neural rendering. In ECCV, Milano, Italy, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.905, + 0.232 + ], + "angle": 0, + "content": "[41] Zehao Zhu, Zhiwen Fan, Yifan Jiang, and Zhangyang Wang. Fsgs: Real-time few-shot view synthesis using gaussian splatting. In ECCV, Milano, Italy, 2024. 1, 2, 3, 4, 5, 6, 7, 8" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26788" + } + ] +] \ No newline at end of file diff --git a/2025/CoMapGS_ Covisibility Map-based Gaussian Splatting for Sparse Novel View Synthesis/025e1d28-cc10-46ad-8c05-b01e6a4dff3b_origin.pdf b/2025/CoMapGS_ Covisibility Map-based Gaussian Splatting for Sparse Novel View Synthesis/025e1d28-cc10-46ad-8c05-b01e6a4dff3b_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..462a8179b4b9c802d28f72c20d2ab2103a722d79 --- /dev/null +++ b/2025/CoMapGS_ Covisibility Map-based Gaussian Splatting for Sparse Novel View Synthesis/025e1d28-cc10-46ad-8c05-b01e6a4dff3b_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8f5753895efdc2cfe5e95d833fa6e4ee2b37d31e5b2a5b9d1e1735672be9b12d +size 3789341 diff --git a/2025/CoMapGS_ Covisibility Map-based Gaussian Splatting for Sparse Novel View Synthesis/full.md b/2025/CoMapGS_ Covisibility Map-based Gaussian Splatting for Sparse Novel View Synthesis/full.md new file mode 100644 index 0000000000000000000000000000000000000000..215d3cca9df9ed633c7a5c2acff218e69a19a045 --- /dev/null +++ b/2025/CoMapGS_ Covisibility Map-based Gaussian Splatting for Sparse Novel View Synthesis/full.md @@ -0,0 +1,318 @@ +# CoMapGS: Covisibility Map-based Gaussian Splating for Sparse Novel View Synthesis + +Youngkyoon Jang and Eduardo Pérez-Pellitero Huawei Noah's Ark Lab, London, UK + +![](images/8036dd9eb105e2357f9a64c9bf7788af13ca673d5fb63cc490a72b2ddd5d1aa3.jpg) +Figure 1. Main idea. Conventional sparse view synthesis methods using 3DGS [38] prioritize frequently captured regions, shown as the brightest areas in the covisibility map, resulting in better reconstruction in these regions but missing details in sparsely captured areas. By using a covisibility map, we guide sparse 3DGS methods to focus on underrepresented regions, enhancing sparse novel view synthesis. + +# Abstract + +We propose Covisibility Map-based Gaussian Splatting (CoMapGS), designed to recover underrepresented sparse regions in sparse novel view synthesis. CoMapGS addresses both high- and low-uncertainty regions by constructing covisibility maps, enhancing initial point clouds, and applying uncertainty-aware weighted supervision using a proximity classifier. Our contributions are threefold: (1) CoMapGS reframes novel view synthesis by leveraging covisibility maps as a core component to address region-specific uncertainty; (2) Enhanced initial point clouds for both low- and high-uncertainty regions compensate for sparse COLMAP-derived point clouds, improving reconstruction quality and benefiting few-shot 3DGS methods; (3) Adaptive supervision with covisibility-score-based weighting and proximity classification achieves consistent performance gains across scenes with varying sparsity scores derived from covisibility maps. Experimental results demonstrate that CoMapGS outperforms state-of-the-art methods on datasets including Mip-NeRF 360 and LLFF. $^{1}$ + +# 1. Introduction + +Recent advances in novel view synthesis have made significant progress in overcoming various challenges, such as generalized model building [27, 40], view synthesis in unbounded scenes [1], fast training [7, 8], and sparse view [38, 41], following the emergence of seminal methodologies such as Neural Radiance Fields (NeRF) [20] and 3D Gaussian Splatting (3DGS) [12]. While dense-view approaches now achieve photo-realistic fidelity, sparse-view synthesis methods often suffer from unrealistic artifacts due to the fundamental shape-radiance ambiguity—i.e., models can overfit to image observations because of implausible geometry, a problem that is exacerbated when training views are limited. Despite the limitations of available training view images, recent studies have proposed various approaches—such as generalized novel view synthesis methods [3, 5, 14, 27, 37, 40], diffusion-based techniques [30, 39], updated densification strategies [38, 41], and the use of pseudo ground truth [30, 38, 41—to improve performance in sparse view synthesis. While recent approaches, particularly those using 3DGS, have significantly improved performance in sparse view synthesis by leveraging rapid point-cloud-based rasterization, efficient gradient- + +descent supervision, and multiview geometry, several open challenges remain: + +- Region-wise imbalanced supervision: Conventional novel view synthesis methods prioritize highly covisible regions due to their multiview presence, effectively weighting them more while under-supervising monoview regions that do not have multiview constraints. +- Sparse point cloud initialization: Conventional sparse point clouds, generated through feature-based keypoint matching (e.g., SIFT [18]) in COLMAP [24], lack the geometric detail that dense view settings offer, especially when the number of training images is limited. +- High-uncertainty regions: Methods that rely heavily on multiview geometry often struggle with mono-view regions where pixels do not appear in other views, posing a critical challenge in sparse view synthesis. + +In response to these challenges, we propose CoMapGS, a Covisibility Map-based Gaussian Splating method that reframes sparse view synthesis by utilizing per-image, pixelwise covisibility counts (i.e., the covisibility map) for adaptive supervision and initial point cloud enhancement, effectively addressing both low- and high-uncertainty regions, as shown in Fig. 2. CoMapGS updates initial point clouds and applies adaptive supervision, using a proximity classifier to regularize plausible geometry in high-uncertainty (i.e., mono-view) regions for enhanced supervision. To effectively manage regions with varying covisibility levels, we construct a covisibility map by accumulating correspondence counts for each pixel from the output of the state-of-the-art dense correspondence prediction model, MASt3R [16]. Additionally, we recover initial point clouds across all regions, including mono-view areas, by learning the correlation between the scale of monocular depth estimation and the point clouds registered using already-optimized COLMAP camera poses, enabling precise rescaling and alignment within the same coordinate system. The updated point clouds are then integrated into CoMapGS to initialize the first set of Gaussians, which are rendered and refined throughout the training process. + +Our proposed CoMapGS can be applied to most sparse view synthesis methods built on the conventional 3DGS framework. Starting with our updated initial point clouds, we supervise CoMapGS by incorporating an additional proximity loss term into the baseline objective function (e.g., FSGS [41], CoR-GS [38]). This proximity loss is derived from our proposed proximity MLP classifier, which predicts the closeness of each point to the original geometry. Leveraging the covisibility map, we apply adaptive weights to the Gaussians that compose the final proximity term, assigning greater weight to Gaussians projected onto mono-view regions and reduced weight to those in multiview regions. Additionally, we adjust supervision strength for Gaussians outside the view frustum based on scene- + +specific可视ability scores, ensuring they contribute effectively to training even if they are not directly represented in rendered images. + +The main contributions of the CoMapGS are threefold: + +1. Covisibility Map-based learning: CoMapGS reframes sparse view synthesis by leveraging covisibility maps for adaptive supervision, enabling balanced guidance across different uncertainty levels within the scene. +2. Enhanced point cloud initialization: By applying learned scaling parameters between monocular depth estimates and triangulated points registered using the given COLMAP camera poses, CoMapGS densely enhances initial point cloud quality, increasing the potential to densify sparse and underrepresented regions, even in mono-view areas. +3. Weighted supervision using covisibility map: This strategy adaptively reinforces high- and low-uncertainty regions, strengthening supervision based on region-specific covisibility levels and stabilizing training with learned geometric constraints that extend beyond the view frustum. + +To the best of our knowledge, CoMapGS is the first approach to explicitly address and recover high-uncertainty regions in sparse view synthesis, effectively balancing supervision across regions with varying uncertainty levels. + +# 2. Related Works + +Sparse novel view synthesis. The growing interest in reconstructing spatial photographs from minimal input image sets has driven the development of various methods aimed at generating novel views from just a few shots, utilizing approaches based on NeRF [6, 11, 13, 21, 23, 26, 28, 30, 31, 36] and 3DGS [17, 22, 32, 34, 38, 41]. Due to the inherent challenges of sparse view synthesis, such as the limited number of training views, NeRF-based studies have primarily focused on proposing regularization techniques with depth ranking [6, 15, 23, 26, 28, 31], frequency-based loss [36], semantic consistency with visual encoders [11], between-view smoothness [13], or uncertainty-aware adaptive loss [33] for both seen and unseen views [11, 15, 21, 30]. Some methods have also addressed uncertainty issues [13, 36], particularly for high-uncertainty mono-view or unseen regions. However, these methods either disregard [13] or penalize [36] uncertain areas, rather than fully utilizing the information contained within. + +While recent 3DGS-based methods continue to rely on regularization techniques, such as depth constraints [17, 22, 32, 34, 41], some have additionally introduced novel densification strategies [38, 41] and explored enhancing initial point clouds [4, 7, 22, 34], capitalizing on the technical characteristics of 3DGS. These methods primarily focus on improving effective multiview regions [9], often biased towards well-represented regions in the training views, rather + +![](images/6fc020d41154632b362667b859297b38d6e31a8f398eb8d8cc0b0e616e0ed1b4.jpg) +Figure 2. Overview of the proposed CoMapGS. CoMapGS leverages existing modules, including Structure-from-Motion (SfM), monocular depth, and dense correspondence prediction, to produce preliminary outputs. Covisibility maps are generated and used alongside module outputs to enhance initial point clouds, which initialize Gaussians for CoMapGS. During training, covisibility map-based weighted supervision with the proximity classifier applies varied strengths to Gaussians based on region-specific covisibility. + +than addressing recovery strategies for high-uncertainty mono-view regions. To overcome the inherent limitations of sparse training views, several prior works have explored generating and utilizing pseudo-ground truths for novel viewpoints via mono-depth estimation [41], forward warping with reversed bilinear sampling [34], or diffusion models with CLIP embedding [30]. While these approaches compensate for the lack of training views, they still face unique challenges in recovering unseen or mono-view regions from the original training view images, potentially misleading the learning process in high-uncertainty areas. + +Covisibility-aware supervision. Beginning with Dy-Check [9], which emphasizes the importance of covisibility in multiview regions to address fair evaluation in dynamic novel view synthesis, recent studies have explored integrating covisibility information into supervision strategies. NeRFVS [35] uses a view coverage map to enforce stricter regularization in low-coverage (covisible) regions, but it is primarily suited for dense view synthesis with video input, where extensive multiview overlap is expected. Similarly, InfoNorm [29] leverages mutual information among surface normals at highly correlated scene points to regularize geometric modeling. However, InfoNorm captures covisible areas within local subregions in structured scenes (e.g., rooms) rather than using a comprehensive covisibility map, which limits its applicability for addressing both multiview and monoview regions. While these methods demonstrate the value of covisibility in supervision, they lack a framework that addresses the challenges of sparse novel view synthesis for natural scenes, including outdoor environments, through a pixel-wise covisibility map. + +# 3. Covisibility Map-based Gaussian Splitting + +The proposed CoMapGS framework is illustrated in Fig. 2. CoMapGS builds on the latest 3DGS-based sparse novel view synthesis method, CoR-GS [38], with two key improvements: (1) updating initial point clouds specifically for sparse and mono-view regions and (2) applying adaptive supervision across varying covisibility regions using a covisibility map and proximity MLP classifier. We begin by generating covisibility maps using the dense correspondence prediction model MASt3R [16], refining it with morphological operations (e.g., erosion and dilation) to support these strategies effectively. + +# 3.1. Covisibility Map Generation + +We generate a covisibility map by identifying dense correspondences between each training view image and all other training view images. Given a set of training view images $T = \{I_1, I_2, \ldots, I_n\}$ , where $n$ is the total number of training views, we compute a covisibility map $M_i$ for each view $I_i$ . For every pair $(I_i, I_{\{j\}})$ with $j \neq i$ , dense correspondences $C_i^j$ are predicted by a dense correspondence prediction function $f_{\mathrm{cor}}$ , where $C_i^j = f_{\mathrm{cor}}(I_i, I_j)$ . The covisibility map $M_i$ is initialized as a zero matrix of size $W \times H$ and defined as: + +$$ +M _ {i} (x, y) = \sum_ {j = 1, j \neq i} ^ {n} \delta_ {(x, y) \in \mathcal {P} \left(C _ {i} ^ {j}\right)}, \quad \forall (x, y) \in I _ {i}, \tag {1} +$$ + +where $\delta_{(x,y)\in \mathcal{P}(C_i^j)}$ is an indicator variable returning 1 if pixel $(x,y)$ in $I_{i}$ has a correspondence match in $I_{j}$ , and 0 otherwise. Here, $\mathcal{P}(C_i^j)$ represents the set of pixel coordinates in $I_{i}$ that have correspondences in $I_{j}$ . + +The resulting covisibility map $M_{i}$ captures pixel-wise covisibility counts ranging from 0 to $n - 1$ , indicating the frequency with which a pixel appears across views, as shown in Fig. 1, with values rescaled from 0 (black) to 255 (white) for visualization purposes. Unlike DyCheck [9]'s covisibility mask, which only indicates pixel visibility, our covisibility map provides precise covisibility details (i.e., covisible view count) at the pixel level. This map supports both the initial point cloud update (Sec. 3.2) and the weighted supervision method (Sec. 3.3). + +# 3.2. Enhancing Initial Point Clouds + +In this subsection, we explain how to improve initial point clouds for both low-uncertainty multi-view and high-uncertainty mono-view regions. To enhance initial point clouds in low-uncertainty multi-view regions where sparse SfM points are available, we integrate a triangulated set of 3D points, $P_T$ , derived from dense correspondences obtained from MASt3R [16], to complement COLMAP's point cloud, $P_C$ . This approach enhances geometrically plausible structures, especially in regions where COLMAP's keypoint-based matching is sparse. For each view $I_i$ , we generate $P_T$ by triangulating dense correspondences $C_i^{\{j\}}$ from view $I_i$ to other views $I_{\{j\}}$ (e.g., two or three correspondences in a 3-view case). To avoid redundancy, correspondences already reconstructed from other views are excluded from subsequent triangulations. + +Formally, we define the updated point cloud $P_{\mathrm{u}}$ as: + +$$ +P _ {\mathrm {u}} = P _ {C} \cup \left\{p _ {T} \in P _ {T} \mid D \left(p _ {T}, P _ {C}\right) > \epsilon \right\}, \tag {2} +$$ + +where $D(p_{T}, P_{C})$ is the distance from a point $p_{T}$ in $P_{T}$ to its nearest neighbor in $P_{C}$ , and $\epsilon$ is a threshold ensuring that only points sufficiently distant from existing points in $P_{C}$ are added. This selective integration allows the triangulated points to supplement COLMAP's sparse initial point cloud in the same coordinate system, filling gaps in low-uncertainty regions and improving overall density in covisible areas, as shown in Fig. 3. + +In high-uncertainty mono-view regions, we use monocular depth estimation [10] to generate a preliminary set of 3D points, $P_{d}$ , by unprojecting depth estimates using the provided camera intrinsics. This set is defined as $P_{d} = P_{d}^{\mathrm{low}} \cup P_{d}^{\mathrm{high}}$ , where $P_{d}^{\mathrm{low}}$ comprises points unprojected from multiview regions (where $M_{i} \geq 1$ ), and $P_{d}^{\mathrm{high}}$ comprises points unprojected from mono-view regions (where $M_{i} = 0$ ). To align these randomly scaled unprojected points with triangulated points from multiview geometry, we learn an anisotropic scaling transformation $f_{\mathrm{scale}}$ , using covisible points from $P_{u}^{\mathrm{low}} \subseteq P_{u}$ as ground truth and $P_{d}^{\mathrm{low}}$ as input. This transformation applies scaling parameters learned through linear regression to match the multiview-derived geometry: + +$$ +f _ {\text {s c a l e}} \left(P _ {d} ^ {\text {l o w}}\right) \approx P _ {u} ^ {\text {l o w}}. \tag {3} +$$ + +![](images/7ac89f152a45c3aed602face5f0743f28110c5ad04e7e71c0549e6f47f736e12.jpg) +(a) + +![](images/cac3204969951a0f3bce964575f0577b29406bcdd4ae4ff3fbbb61a710f1e9b7.jpg) +(b) + +![](images/0e6f6a6ad6de1689732fb55910d5654658f039daf54df1b211cb8798d412a638.jpg) +(c) +Figure 3. Examples of enhanced initial PCLs from the Fern scene in the LLFF dataset. (a) Initial SfM points are first (b) updated through triangulation using dense correspondences and COLMAP camera poses, and then (c) fully refined to include mono-view regions (mono-view updates are colored in red for visualization). + +For consistency, we define $P_{u} = P_{u}^{\mathrm{low}} \cup P_{u}^{\mathrm{high}}$ , where $P_{u}^{\mathrm{low}}$ is selected based on covisibility: + +$$ +P _ {u} ^ {\text {l o w}} = \left\{p \mid M _ {i} \left(\pi \left(P _ {u}, \mathbf {H} _ {i}\right)\right) \geq 1 \right\}, \tag {4} +$$ + +where $\pi(P_u, \mathbf{H}_i)$ denotes the projection of $P_u$ using the camera transformation matrix $\mathbf{H}_i$ , yielding the image coordinates $(x, y)$ for a given camera viewpoint and ensuring inclusion of only points visible in multiple views. + +After training, we apply $f_{\mathrm{scale}}$ to the randomly scaled unprojected points from depth estimates in high-uncertainty mono-view regions: + +$$ +P _ {s} ^ {\text {h i g h}} = f _ {\text {s c a l e}} \left(P _ {d} ^ {\text {h i g h}}\right). \tag {5} +$$ + +This transformation aligns the randomly scaled points $P_{d}^{\mathrm{high}}$ to the scale of COLMAP's triangulated points $P_{s}^{\mathrm{high}}$ , ensuring spatial consistency across regions. The final point cloud, $P_{\mathrm{final}}$ , is defined as: + +$$ +P _ {\text {f i n a l}} = P _ {u} ^ {\text {l o w}} \cup P _ {s} ^ {\text {h i g h}}, \tag {6} +$$ + +resulting in enhanced initial point cloud density and accuracy across both multiview and mono-view regions, as shown in Fig. 3, thus supporting cohesive scene reconstruction in sparse view synthesis. + +# 3.3. Covisibility Map-Based Weighted Supervision + +We propose a covisibility map-based supervision strategy that leverages the enhanced initial point cloud, $P_{\mathrm{final}}$ , and the covisibility map $M_{i}$ to adaptively supervise regions based on uncertainty levels represented by covisibility counts, as shown in Fig. 4. This approach improves conventional sparse view Gaussian splatting methods [38, 41] by introducing region-specific weighting across covisibility levels, $M_{i} = \{0:n - 1\}$ , and by supervising Gaussians located outside the training view frustum with a scene-level weight based on the average covisibility score $S = \mathrm{AVG}(M_{[1:n]})$ . + +Our approach integrates seamlessly with the latest 3DGS-based sparse view synthesis methods [38, 41] by adding a proximity loss term, weighted using inverse proximity scores predicted by a three-layer MLP classifier + +trained on $P_{\mathrm{final}}$ (see details in the supplementary materials). + +Proximity loss with adaptive weighting. To implement this approach, we first train an MLP classifier $f_{p}$ on $P_{\mathrm{final}}$ , labeling these curated scene 3D points as positive (representing the original scene geometry) and randomly distributed points, $P_{\mathrm{random}}$ , constrained to remain distant from $P_{\mathrm{final}}$ , as negative. The classifier $f_{p}: \mathbb{R}^{3} \to [0,1]$ outputs a proximity score $s$ for each 3D point $p$ : + +$$ +s = f _ {p} (p) \in [ 0, 1 ], \tag {7} +$$ + +where $s$ approaches 1 for points close to the scene geometry and 0 for distant points. Using batch processing, we compute the proximity scores efficiently for the entire Gaussian set, $G$ , and define the proximity loss $\mathcal{L}_p$ as: + +$$ +\mathcal {L} _ {p} = \frac {1}{| G |} \sum_ {g \in G} (\chi (g) w _ {\text {i n}} + (1 - \chi (g)) w _ {\text {o u t}}) \cdot (1 - s), \tag {8} +$$ + +where $g$ denotes the 3D position of a Gaussian in the set $G$ , representing the complete Gaussian set. The indicator function $\chi(g)$ is defined as: + +$$ +\chi (g) = \left\{ \begin{array}{l l} 1, & \text {i f} \pi (g, \mathbf {H} _ {i}) \in [ 0, \mathsf {W}) \times [ 0, \mathsf {H}) \\ 0, & \text {o t h e r w i s e} \end{array} , \right. \tag {9} +$$ + +where $\pi(g, \mathbf{H}_i)$ is the projection of $g$ under camera transformation $\mathbf{H}_i$ onto the可视ibility map $M_i$ . + +Within the camera-view frustum. We use the weight $w_{in}$ for Gaussians projected onto covisibility mask regions, defined as: + +$$ +w _ {\mathrm {i n}} = \frac {1}{M _ {i} \left(\pi \left(g , \mathbf {H} _ {i}\right)\right) + 1}, \tag {10} +$$ + +where $M_{i}(\pi (g,\mathbf{H}i))$ is the covisibility count of $g$ in the covisibility map for the corresponding projection. This weighting ensures that high-covisibility regions, which are frequently supervised by other loss terms, receive weaker proximity-related supervision, while mono-view regions benefit from a more direct application of $\mathcal{L}_p$ . + +Outside the camera-view frustum. In conventional 3DGS-based methods, only Gaussians within the view frustum are supervised per iteration. However, for scenes with high covisibility scores ( $S > 0.7$ ), we extend supervision to Gaussians located outside the view frustum, applying a scene-level weight $w_{out}$ for these points: + +$$ +w _ {o u t} = \max \left(0, \frac {S - 0 . 7}{0 . 3}\right), \tag {11} +$$ + +which scales linearly from 1 to 0 as $S$ ranges from 1 to 0.7. This approach improves supervision outside the view frustum in high-covisibility score scenes, ensuring comprehensive guidance while making efficient use of the proximity classifier. + +![](images/7a91f338337b3a8e1d39c6031da96fbb1027443d5d666fac143882504ccb23d4.jpg) +Figure 4. Alongside (a) the conventional reconstruction loss $\mathcal{L}1$ , we add (b) a weighted proximity loss $\mathcal{L}_p$ that applies classifier results with assigned weights based on Gaussian projections onto the可视ability map and their positions relative to the frustum. + +# 3.4. Objective Function + +Our proposed CoMapGS enhances the objective function by introducing a proximity loss term, effectively complementing the function while preserving the original strategies of state-of-the-art methods, such as Co-pruning and Pseudoview Co-regularization from CoR-GS [38]. The full objective function is expressed as: + +$$ +\mathcal {L} = (1 - \lambda) \mathcal {L} 1 \left(I, I ^ {*}\right) + \lambda \mathcal {L} _ {\mathrm {D} - \mathrm {S S I M}} \left(I, I ^ {*}\right) + \mathcal {L} _ {p}, \tag {12} +$$ + +where $I$ is the rendered image, supervised against ground truth $I^{*}$ , $\mathcal{L}1$ is the L1 reconstruction loss, and $\mathcal{L}_{\mathrm{D - SSIM}}$ represents the D-SSIM loss term. The balancing parameter $\lambda$ controls the weighting between L1 and D-SSIM, in line with standard Gaussian splattering objectives. The proximity loss term $\mathcal{L}_p$ , defined in Eq. 8, adds adaptive supervision to improve geometry alignment. + +Our proposed approach extends seamlessly to other recent sparse view synthesis methods, such as FSGS [41], by incorporating $\mathcal{L}_p$ into their objective function while preserving their original strategies, as reported in Table 1. + +# 4. Experimental Results + +# 4.1. Setup + +Datasets. We conducted experiments on two benchmark datasets: LLFF [19], using 3, 6, and 9 training views with resolutions downsampled to $8 \times$ , and Mip-NeRF 360 [2], using 12 and 24 training views per scene with resolutions downsampled to $4 \times$ . We adhered to the standard training and testing conventions [38] for datasets commonly used in sparse view synthesis. + +Comparison Baselines. We compare CoMapGS with existing few-shot novel view synthesis methods, including FSGS [41] and CoR-GS [38]. As our method primarily enhances the performance of 3DGS-based approaches, we reproduced results for these methods. For other methods, we + +Table 1. Quantitative results on LLFF with 3, 6, and 9 training views. The best, second-best, and third-best entries are marked in red, orange, and yellow, respectively. $\dagger$ indicates results reproduced by us using the released code base. + +
MethodPSNR ↑SSIM ↑LPIPS ↓
3-view6-view9-view3-view6-view9-view3-view6-view9-view
DietNeRF [11]14.9421.7524.280.3700.7170.8010.4960.2480.183
Mip-NeRF [1]16.1122.9124.880.4010.7560.8260.4600.2130.160
RegNeRF [21]19.0823.1024.860.5870.7600.8200.3360.2060.161
SimpleNeRF [25]19.2423.0523.980.6230.7370.7620.3750.2960.286
FreeNeRF [36]19.6323.7325.130.6120.7790.8270.3080.1950.160
SparseNeRF [28]19.8623.2624.270.7140.7410.7810.2430.2350.228
DiffusioNeRF [31]20.1323.6024.620.6310.7750.8070.3440.2350.216
ReconFusion [30]21.3424.2525.210.7240.8150.8480.2030.1520.134
DNGaussian [17]19.1222.1823.170.5910.7550.7880.2940.1980.180
3DGS [12]19.2223.8025.440.6490.8140.8600.2290.1250.096
FSGS [41]†20.45824.69426.0060.7130.8390.8750.2050.1210.095
CoR-GS [38]†20.47324.77726.4750.7170.8440.8810.1990.1160.086
[41]† + CoMapGS (ours)20.89124.80826.2920.7370.8440.8790.1900.1170.092
[38]† + CoMapGS (ours)21.10525.20426.7310.7470.8540.8860.1820.1080.082
+ +![](images/cdc5b80ddb9038cc63817b9cc6b55aed2a723fc2dc829f9963d54573207839b3.jpg) + +![](images/bb2d96001490401f4cd3b6753cee514f270c6e63b47cc167c83f71bbe56c71a1.jpg) + +![](images/1fd01778b3bda0fd3e2a157c910a5ec072a1c80b53fa32a6c12f730ac44ae21d.jpg) +FSGS + +![](images/c5391d363b97828f571572ad1bf6b11034cc6075bd3acdf9b09859bff86ce904.jpg) + +![](images/4473f9fd1920f51392b5c01e00b0b306101e5002633053af0dd6a1d0417e7db4.jpg) + +![](images/6e0f8f5891ce569ccecfa80422462dacf1b82806f59c8d96ac9b62219cf971ff.jpg) +CoR-GS + +![](images/3f9bac9de3ab091468f154e9ed3ab449efdead776ec23e8d4c6803508d45c6fd.jpg) + +![](images/34195533195e7087d72fdb7d3b8205aa224f5434703239bc3c15c672bcfe88d5.jpg) + +![](images/1d53208fa5e2da576cc6539cd1fd4108fd5102a34b86f71152c6a5a0e9799be5.jpg) +Ours +Figure 5. Qualitative comparison on 3 training views from the LLFF dataset. We compare our results with FSGS [41] and CoR-GS [38], demonstrating that our method achieves superior texture synthesis in novel views, particularly in high-uncertainty mono-view regions that appear only once in the training set. + +![](images/efbe0a61a970aaff676ad996c7956404d11c4bf77998d5df632ff1f7c4b66992.jpg) + +![](images/bbb2b25365aec124c991a1f3dc5428e2d9b7fa386ee1f516047eb5ee9cc9a8a7.jpg) + +![](images/0e53e929cde0538493dcf585612cb5cf4209e4d7d7845cd26eaafb8ac9551e41.jpg) +Ground Truth + +![](images/130b119ca75288856f5b6cb312d047e2bbecd3e1c85de00903114c8336554421.jpg) + +![](images/16d3dab5adf8d13e3fb94ee6cca4d753674e7eb29eaeb73a9b2e294308a82183.jpg) + +![](images/f252d99f9e77044a86033fb48eda671de035f1a6441a9b543a08e2cea7e7533d.jpg) +Covisibility Mask + +report the best published scores from prior works, as shown in Tables 1 and 2. + +Evaluation Metrics. We evaluate our results using PSNR, SSIM, and LPIPS metrics. + +Implementation. Training was conducted for 10k iterations on the LLFF dataset and 30k iterations on the MipNeRF 360 dataset, following the setup of 3DGS [12]. Con + +sistent with experimental setups in FSGS [41] and CoR-GS [38], CoMapGS was initialized with point clouds and precomputed camera poses from COLMAP. + +# 4.2. Comparisons + +LLFF. Quantitative and qualitative results on the LLFF dataset are presented in Table 1 and Fig. 5, respectively. + +Table 2. Quantitative results on Mip-NeRF360 with 12 and 24 training views. The best, second-best, and third-best entries are marked in red, orange, and yellow, respectively. ${}^{ \dagger }$ indicates results reproduced by us using the released code base. + +
MethodPSNR ↑SSIM ↑LPIPS ↓
12-view24-view12-view24-view12-view24-view
3DGS [12]18.5222.800.5230.7080.4150.276
FSGS [41]18.8023.280.5310.7150.4180.274
CoR-GS [38]†19.16223.3210.5740.7290.4140.271
[38]† + CoMapGS (ours)19.68023.4620.5910.7340.3940.264
+ +![](images/657c6f64c99b25fec1064da2a6118654b62ca309a4d69645f3f1770ab20cd348.jpg) +Figure 6. Qualitative comparison on 12 training views from the Mip-NeRF 360 dataset. Mip-NeRF 360 highlights diverse covisibility levels, with high-uncertainty regions sparsely captured. CoMapGS effectively synthesizes these regions, while CoR-GS struggles to recover textures. + +We reproduced evaluation scores for the two baseline methods, FSGS [41] and CoR-GS [38], to highlight the improvements over our reproduced results. The results show consistent improvements, ranking first across various training view counts, with PSNR placing second only to ReconFusion [30] in the 3-view case. This difference is likely due to diffusion-based sparse view synthesis methods, which are optimized for generating visually seamless colors that blend with surrounding regions. However, this approach may not + +accurately capture original patterns in mono-view or unseen regions—common in sparse view synthesis datasets, especially with fewer training views. Consequently, while we ranked second in PSNR, our method achieved substantially higher SSIM and lower LPIPS scores under comparable view settings, indicating improved structural fidelity and perceptual quality in the synthesized views. + +Mip-NeRF 360. To validate CoMapGS's applicability to broader scenes that capture large areas with outward- + +facing cameras, we evaluated our method using the MipNeRF 360 dataset. This dataset, which centers objects of interest with outward-facing cameras, captures the motivation of our approach by featuring distinct low-uncertainty, multiview regions alongside high-uncertainty, less-covisible regions within an image, as shown in Fig. 1. In these scenarios, COLMAP struggles to establish stable correspondences, leading to sparse initial point clouds that fail to adequately recover details in high-uncertainty regions, as shown in Fig. 6. + +In contrast, CoMapGS begins with enhanced initial point clouds, preserving the scene's original structure while benefiting from proximity loss supervision, resulting in improved texture representation over conventional methods, as shown in Fig. 6. Quantitative results in Table 2 confirm significant improvements across metrics, particularly LPIPS, compared to recent state-of-the-art methods, which perform worse than their baselines (e.g., 3DGS [12], FSGS [41]). In the 24-view case, while PSNR improvements may appear modest due to unseen regions in test views, the LPIPS gain highlights our method's effectiveness in recovering textures from both low- and high-uncertainty regions in novel view. + +# 4.3. Ablation Study + +The ablation study in Table 3 highlights that dense reconstruction in low-uncertainty (covisible) regions already provides significant improvement over baseline methods, underscoring the benefit of increased point density in covisible areas. Furthermore, adding points from mono-view regions to the enhanced point cloud yields additional gains, demonstrating the value of incorporating points from underrepresented, high-uncertainty regions in sparse view synthesis. + +We also evaluated the impact of our proximity loss-based weighted supervision both with and without the initial point cloud enhancement strategy. Interestingly, starting from sparse COLMAP point clouds without updates showed limited improvement, likely because 3DGS-based methods tend to densify Gaussians near their original positions. However, when initialized with our enhanced point clouds, the benefits of weighted supervision became evident, delivering significant performance gains in both partial (multiview region) and full (multi- and mono-view region) update scenarios. This demonstrates that enhancing initial point clouds and incorporating proximity loss-based weighted supervision work synergistically to optimize 3DGS-based sparse view synthesis. + +# 4.4. Discussion + +Definition of sparse view synthesis. Although sparse novel view synthesis is often defined by a specific number of training views (e.g., LLFF with 3, 6, or 9 views; Mip-NeRF with 12 or 24 views), we argue that the true challenge lies in reconstructing underrepresented regions with low covisibility + +Table 3. Ablation study of the proposed CoMapGS with six training views. $\triangle$ indicates enhancing initial point clouds only in the low-uncertainty region (where covisible count $\geq 2$ ), representing a straightforward application of dense point clouds generated by conventional methods such as MAST3R [16]. + +
Proposed methodLLFF
Sec. 3.2Sec. 3.3PSNR ↑SSIM ↑LPIPS ↓
CoR-GS [38]†24.7770.8440.116
✓24.7870.8450.116
△24.900.8490.112
△✓25.1530.8540.109
✓25.0760.8520.109
✓✓25.2040.8540.108
+ +counts rather than the total view count. This problem is inherently region-wise and imbalanced, where certain areas appear in only a few views, leading to sparse representation even when many views are available. For example, as DyCheck [9] categorizes scene types based on effective multiview factors (EMFs), scenarios with high frame counts but low region-specific可视ibility can still be considered sparse view cases. We address this by applying region-wise adaptive supervision based on可视ibility counts, improving results, especially for imbalanced datasets like MipNeRF 360 (12 views). In such cases, central objects are densely captured, while outer areas show low可视ibility due to the 360-degree, outward-facing capture setup. This issue is common in real-world scene capture, where uneven可视ibility across views motivates balanced,可视ibility-aware supervision. + +# 5. Conclusions + +This paper introduces CoMapGS, a Covisibility Map-based Gaussian Splitting approach designed to enhance sparse view synthesis methods by addressing region-specific uncertainty levels. We generate covisibility maps for a given non-continuous set of distant images, using them to identify low- and high-uncertainty regions and to improve point cloud initialization beyond conventional methods like COLMAP. By introducing covisibility-adaptive weighted supervision, CoMapGS effectively manages varying sparsity across both multiview and mono-view regions, consistently outperforming state-of-the-art 3DGS-based sparse novel view synthesis methods. Experimental results validate CoMapGS's effectiveness across diverse, realistic scenes, including both indoor and outdoor environments and various scales of captured areas. CoMapGS demonstrates superiority in handling complex region-wise covisibility, highlighting its robustness and adaptability in challenging sparse view synthesis tasks. + +# References + +[1] Jonathan T. Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P. Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In ICCV, pages 5855–5864, 2021. 1, 6 +[2] Jonathan T. Barron, Ben Mildenhall, Dor Verbin, Pratul P. Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5470-5479, 2022. 5 +[3] Anpei Chen, Zexiang Xu, Fuqiang Zhao, Xiaoshuai Zhang, Fanbo Xiang, Jingyi Yu, and Hao Su. Mvsnerf: Fast generalizable radiance field reconstruction from multi-view stereo. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 14124-14133, 2021. 1 +[4] Shan Chen, Jiale Zhou, and Lei Li. Dense point clouds matter: Dust-gs for scene reconstruction from sparse viewpoints, 2024. 2 +[5] Wenyan Cong, Hanxue Liang, Peihao Wang, Zhiwen Fan, Tianlong Chen, Mukund Varma, Yi Wang, and Zhangyang Wang. Enhancing nerf akin to enhancing llms: Generalizable nerf transformer with mixture-of-view-experts. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 3193-3204, 2023. 1 +[6] Kangle Deng, Andrew Liu, Jun-Yan Zhu, and Deva Ramanan. Depth-supervised nerf: Fewer views and faster training for free. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12882-12891, 2022. 2 +[7] Zhiwen Fan, Wenyan Cong, Kairun Wen, Kevin Wang, Jian Zhang, Xinghao Ding, Danfei Xu, Boris Ivanovic, Marco Pavone, Georgios Pavlakos, Zhangyang Wang, and Yue Wang. Instantsplat: Unbounded sparse-view pose-free gaussian splatting in 40 seconds, 2024. 1, 2 +[8] Guangchi Fang and Bing Wang. Mini-Splatting: Representing scenes with a constrained number of gaussians. In ECCV, Milano, Italy, 2024. 1 +[9] Hang Gao, Ruilong Li, Shubham Tulsiani, Bryan Russell, and Angjoo Kanazawa. Monocular dynamic view synthesis: A reality check. In NeurIPS, 2022. 2, 3, 4, 8 +[10] Mu Hu, Wei Yin, Chi Zhang, Zhipeng Cai, Xiaoxiao Long, Hao Chen, Kaixuan Wang, Gang Yu, Chunhua Shen, and Shaojie Shen. Metric3d v2: A versatile monocular geometric foundation model for zero-shot metric depth and surface normal estimation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(12):10579-10596, 2024. 4 +[11] Ajay Jain, Matthew Tancik, and Pieter Abbeel. Putting nerf on a diet: Semantically consistent few-shot view synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 5885-5894, 2021. 2, 6 +[12] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42 (4), 2023. 1, 6, 7, 8 +[13] Mijeong Kim, Seonguk Seo, and Bohyung Han. Infonerf: Ray entropy minimization for few-shot neural volume rendering. In Proceedings of the IEEE/CVF Conference on + +Computer Vision and Pattern Recognition (CVPR), pages 12912-12921, 2022. 2 +[14] Jonás Kulhánek, Erik Derner, Torsten Sattler, and Robert Babuška. Viewformer: Nerf-free neural rendering from few images using transformers. In European Conference on Computer Vision (ECCV), 2022. 1 +[15] Min-Seop Kwak, Jiuhn Song, and Seungryong Kim. Geconerf: few-shot neural radiance fields via geometric consistency. In Proceedings of the 40th International Conference on Machine Learning. JMLR.org, 2023. 2 +[16] Vincent Leroy, Yohann Cabon, and Jerome Revaud. Grounding image matching in 3d with mast3r, 2024. 2, 3, 4, 8 +[17] Jiahe Li, Jiawei Zhang, Xiao Bai, Jin Zheng, Xin Ning, Jun Zhou, and Lin Gu. Dngaussian: Optimizing sparse-view 3d gaussian radiance fields with global-local depth normalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20775-20785, 2024. 2, 6 +[18] David G. Lowe. Distinctive image features from scale-invariant keypoints. Int. J. Comput. Vision, 60(2):91-110, 2004. 2 +[19] Ben Mildenhall, Pratul P. Srinivasan, Rodrigo Ortiz-Cayon, Nima Khademi Kalantari, Ravi Ramamoorthi, Ren Ng, and Abhishek Kar. Local light field fusion: Practical view synthesis with prescriptive sampling guidelines. ACM Transactions on Graphics (TOG), 2019. 5 +[20] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 1 +[21] Michael Niemeyer, Jonathan T. Barron, Ben Mildenhall, Mehdi S. M. Sajjadi, Andreas Geiger, and Noha Radwan. Regnerf: Regularizing neural radiance fields for view synthesis from sparse inputs. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 6 +[22] Avinash Paliwal, Wei Ye, Jinhui Xiong, Dmytro Kotovenko, Rakesh Ranjan, Vikas Chandra, and Nima Khademi Kalantari. CoherentGS: Sparse novel view synthesis with coherent 3d gaussians. In ECCV, Milano, Italy, 2024. 2 +[23] Barbara Roessle, Jonathan T. Barron, Ben Mildenhall, Pratul P. Srinivasan, and Matthias Nießner. Dense depth priors for neural radiance fields from sparse input views. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12892-12901, 2022. 2 +[24] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 2 +[25] Nagabhushan Somraj, Adithyan Karanayil, and Rajiv Soundararajan. SimpleNeRF: Regularizing sparse input neural radiance fields with simpler solutions. In SIGGRAPH Asia, 2023. 6 +[26] Jiuhn Song, Seonghoon Park, Honggyu An, Seokju Cho, Min-Seop Kwak, Sungjin Cho, and Seungryong Kim. Därf: Boosting radiance fields from sparse input views with monocular depth adaptation. In Advances in Neural Information Processing Systems, pages 68458-68470. Curran Associates, Inc., 2023. 2 + +[27] Thomas Tanay and Matteo Maggioni. Global latent neural rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19723-19733, 2024. 1 +[28] Guangcong Wang, Zhaoxi Chen, Chen Change Loy, and Ziwei Liu. Sparsenerf: Distilling depth ranking for few-shot novel view synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 9065-9076, 2023. 2, 6 +[29] Xulong Wang, Siyan Dong, Youyi Zheng, and Yanchao Yang. InfoNorm: Mutual information shaping of normals for sparse-view reconstruction. In ECCV, Milano, Italy, 2024. 3 +[30] Rundi Wu, Ben Mildenhall, Philipp Henzler, Keunhong Park, Ruiqi Gao, Daniel Watson, Pratul P. Srinivasan, Dor Verbin, Jonathan T. Barron, Ben Poole, and Aleksander Ho?y?ski. Reconfusion: 3d reconstruction with diffusion priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 21551-21561, 2024. 1, 2, 3, 6, 7 +[31] Jamie Wynn and Daniyar Turmukhambetov. Diffusionerf: Regularizing neural radiance fields with denoising diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4180-4189, 2023. 2, 6 +[32] Haolin Xiong, Sairisheek Muttukuru, Rishi Upadhyay, Pradyumna Chari, and Achuta Kadambi. Sparsegs: Realtime $360^{\circ}$ sparse view synthesis using gaussian splatting. Arxiv, 2023. 2 +[33] Qingshan Xu, Xuanyu Yi, Jianyao Xu, Wenbing Tao, Yew Soon Ong, and Hanwang Zhang. Few-shot nerf by adaptive rendering loss regularization. In ECCV, Milano, Italy, 2024. 2 +[34] Wangze Xu, Huachen Gao, Shihe Shen, Rui Peng, Jianbo Jiao, and Ronggang Wang. Mvpgs: Excavating multi-view priors for gaussian splatting from sparse input views. In ECCV, Milano, Italy, 2024. 2, 3 +[35] Chen Yang, Peihao Li, Zanwei Zhou, Shanxin Yuan, Bingbing Liu, Xiaokang Yang, Weichao Qiu, and Wei Shen. NeRFVS: Neural Radiance Fields for Free View Synthesis via Geometry Scaffolds. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 3 +[36] Jiawei Yang, Marco Pavone, and Yue Wang. Freenerf: Improving few-shot neural rendering with free frequency regularization. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2023. 2, 6 +[37] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelnerf: Neural radiance fields from one or few images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4578-4587, 2021. 1 +[38] Jiawei Zhang, Jiahe Li, Xiaohan Yu, Lei Huang, Lin Gu, Jin Zheng, and Xiao Bai. Cor-gs: Sparse-view 3d gaussian splatt-ting via co-regularization. In ECCV, Milano, Italy, 2024. 1, 2, 3, 4, 5, 6, 7, 8 +[39] Zhizhuo Zhou and Shubham Tulsiani. Sparsefusion: Distilling view-conditioned diffusion for 3d reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vi + +sion and Pattern Recognition (CVPR), pages 12588-12597, 2023. 1 +[40] Haidong Zhu, Tianyu Ding, Tianyi Chen, Ilya Zharkov, Ram Nevatia, and Luming Liang. CaesarNeRF: Calibrated semantic representation for few-shot generalizable neural rendering. In ECCV, Milano, Italy, 2024. 1 +[41] Zehao Zhu, Zhiwen Fan, Yifan Jiang, and Zhangyang Wang. Fsgs: Real-time few-shot view synthesis using gaussian splatting. In ECCV, Milano, Italy, 2024. 1, 2, 3, 4, 5, 6, 7, 8 \ No newline at end of file diff --git a/2025/CoMapGS_ Covisibility Map-based Gaussian Splatting for Sparse Novel View Synthesis/images.zip b/2025/CoMapGS_ Covisibility Map-based Gaussian Splatting for Sparse Novel View Synthesis/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..49a6856304ec0ef9075ede24a999ea28d7973099 --- /dev/null +++ b/2025/CoMapGS_ Covisibility Map-based Gaussian Splatting for Sparse Novel View Synthesis/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:74ebcae9b3278c50ed67c16932032f92e1472ab9b1a56ecab5dc3b05510fec43 +size 1005892 diff --git a/2025/CoMapGS_ Covisibility Map-based Gaussian Splatting for Sparse Novel View Synthesis/layout.json b/2025/CoMapGS_ Covisibility Map-based Gaussian Splatting for Sparse Novel View Synthesis/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..ce93b0017b2120a8f154105bcba62fbe1cdcfd94 --- /dev/null +++ b/2025/CoMapGS_ Covisibility Map-based Gaussian Splatting for Sparse Novel View Synthesis/layout.json @@ -0,0 +1,8722 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 136, + 103, + 475, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 103, + 475, + 140 + ], + "spans": [ + { + "bbox": [ + 136, + 103, + 475, + 140 + ], + "type": "text", + "content": "CoMapGS: Covisibility Map-based Gaussian Splating for Sparse Novel View Synthesis" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 192, + 161, + 419, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 192, + 161, + 419, + 190 + ], + "spans": [ + { + "bbox": [ + 192, + 161, + 419, + 190 + ], + "type": "text", + "content": "Youngkyoon Jang and Eduardo Pérez-Pellitero Huawei Noah's Ark Lab, London, UK" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 58, + 203, + 553, + 369 + ], + "blocks": [ + { + "bbox": [ + 58, + 203, + 553, + 369 + ], + "lines": [ + { + "bbox": [ + 58, + 203, + 553, + 369 + ], + "spans": [ + { + "bbox": [ + 58, + 203, + 553, + 369 + ], + "type": "image", + "image_path": "8036dd9eb105e2357f9a64c9bf7788af13ca673d5fb63cc490a72b2ddd5d1aa3.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 375, + 555, + 409 + ], + "lines": [ + { + "bbox": [ + 54, + 375, + 555, + 409 + ], + "spans": [ + { + "bbox": [ + 54, + 375, + 555, + 409 + ], + "type": "text", + "content": "Figure 1. Main idea. Conventional sparse view synthesis methods using 3DGS [38] prioritize frequently captured regions, shown as the brightest areas in the covisibility map, resulting in better reconstruction in these regions but missing details in sparsely captured areas. By using a covisibility map, we guide sparse 3DGS methods to focus on underrepresented regions, enhancing sparse novel view synthesis." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 152, + 424, + 200, + 437 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 424, + 200, + 437 + ], + "spans": [ + { + "bbox": [ + 152, + 424, + 200, + 437 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 451, + 297, + 679 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 451, + 297, + 679 + ], + "spans": [ + { + "bbox": [ + 55, + 451, + 297, + 679 + ], + "type": "text", + "content": "We propose Covisibility Map-based Gaussian Splatting (CoMapGS), designed to recover underrepresented sparse regions in sparse novel view synthesis. CoMapGS addresses both high- and low-uncertainty regions by constructing covisibility maps, enhancing initial point clouds, and applying uncertainty-aware weighted supervision using a proximity classifier. Our contributions are threefold: (1) CoMapGS reframes novel view synthesis by leveraging covisibility maps as a core component to address region-specific uncertainty; (2) Enhanced initial point clouds for both low- and high-uncertainty regions compensate for sparse COLMAP-derived point clouds, improving reconstruction quality and benefiting few-shot 3DGS methods; (3) Adaptive supervision with covisibility-score-based weighting and proximity classification achieves consistent performance gains across scenes with varying sparsity scores derived from covisibility maps. Experimental results demonstrate that CoMapGS outperforms state-of-the-art methods on datasets including Mip-NeRF 360 and LLFF." + }, + { + "bbox": [ + 55, + 451, + 297, + 679 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 315, + 424, + 394, + 437 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 424, + 394, + 437 + ], + "spans": [ + { + "bbox": [ + 315, + 424, + 394, + 437 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 451, + 556, + 716 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 451, + 556, + 716 + ], + "spans": [ + { + "bbox": [ + 313, + 451, + 556, + 716 + ], + "type": "text", + "content": "Recent advances in novel view synthesis have made significant progress in overcoming various challenges, such as generalized model building [27, 40], view synthesis in unbounded scenes [1], fast training [7, 8], and sparse view [38, 41], following the emergence of seminal methodologies such as Neural Radiance Fields (NeRF) [20] and 3D Gaussian Splatting (3DGS) [12]. While dense-view approaches now achieve photo-realistic fidelity, sparse-view synthesis methods often suffer from unrealistic artifacts due to the fundamental shape-radiance ambiguity—i.e., models can overfit to image observations because of implausible geometry, a problem that is exacerbated when training views are limited. Despite the limitations of available training view images, recent studies have proposed various approaches—such as generalized novel view synthesis methods [3, 5, 14, 27, 37, 40], diffusion-based techniques [30, 39], updated densification strategies [38, 41], and the use of pseudo ground truth [30, 38, 41—to improve performance in sparse view synthesis. While recent approaches, particularly those using 3DGS, have significantly improved performance in sparse view synthesis by leveraging rapid point-cloud-based rasterization, efficient gradient-" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 703, + 259, + 715 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 703, + 259, + 715 + ], + "spans": [ + { + "bbox": [ + 66, + 703, + 259, + 715 + ], + "type": "text", + "content": "1 Project page: youngkyoonjang.github.io/projects/comapgs" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "26779" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 96 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 96 + ], + "type": "text", + "content": "descent supervision, and multiview geometry, several open challenges remain:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 99, + 295, + 267 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 55, + 99, + 295, + 159 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 99, + 295, + 159 + ], + "spans": [ + { + "bbox": [ + 55, + 99, + 295, + 159 + ], + "type": "text", + "content": "- Region-wise imbalanced supervision: Conventional novel view synthesis methods prioritize highly covisible regions due to their multiview presence, effectively weighting them more while under-supervising monoview regions that do not have multiview constraints." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 159, + 295, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 159, + 295, + 219 + ], + "spans": [ + { + "bbox": [ + 55, + 159, + 295, + 219 + ], + "type": "text", + "content": "- Sparse point cloud initialization: Conventional sparse point clouds, generated through feature-based keypoint matching (e.g., SIFT [18]) in COLMAP [24], lack the geometric detail that dense view settings offer, especially when the number of training images is limited." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 219, + 295, + 267 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 219, + 295, + 267 + ], + "spans": [ + { + "bbox": [ + 55, + 219, + 295, + 267 + ], + "type": "text", + "content": "- High-uncertainty regions: Methods that rely heavily on multiview geometry often struggle with mono-view regions where pixels do not appear in other views, posing a critical challenge in sparse view synthesis." + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 270, + 295, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 270, + 295, + 545 + ], + "spans": [ + { + "bbox": [ + 55, + 270, + 295, + 545 + ], + "type": "text", + "content": "In response to these challenges, we propose CoMapGS, a Covisibility Map-based Gaussian Splating method that reframes sparse view synthesis by utilizing per-image, pixelwise covisibility counts (i.e., the covisibility map) for adaptive supervision and initial point cloud enhancement, effectively addressing both low- and high-uncertainty regions, as shown in Fig. 2. CoMapGS updates initial point clouds and applies adaptive supervision, using a proximity classifier to regularize plausible geometry in high-uncertainty (i.e., mono-view) regions for enhanced supervision. To effectively manage regions with varying covisibility levels, we construct a covisibility map by accumulating correspondence counts for each pixel from the output of the state-of-the-art dense correspondence prediction model, MASt3R [16]. Additionally, we recover initial point clouds across all regions, including mono-view areas, by learning the correlation between the scale of monocular depth estimation and the point clouds registered using already-optimized COLMAP camera poses, enabling precise rescaling and alignment within the same coordinate system. The updated point clouds are then integrated into CoMapGS to initialize the first set of Gaussians, which are rendered and refined throughout the training process." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "text", + "content": "Our proposed CoMapGS can be applied to most sparse view synthesis methods built on the conventional 3DGS framework. Starting with our updated initial point clouds, we supervise CoMapGS by incorporating an additional proximity loss term into the baseline objective function (e.g., FSGS [41], CoR-GS [38]). This proximity loss is derived from our proposed proximity MLP classifier, which predicts the closeness of each point to the original geometry. Leveraging the covisibility map, we apply adaptive weights to the Gaussians that compose the final proximity term, assigning greater weight to Gaussians projected onto mono-view regions and reduced weight to those in multiview regions. Additionally, we adjust supervision strength for Gaussians outside the view frustum based on scene-" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "text", + "content": "specific可视ability scores, ensuring they contribute effectively to training even if they are not directly represented in rendered images." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 326, + 108, + 548, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 108, + 548, + 120 + ], + "spans": [ + { + "bbox": [ + 326, + 108, + 548, + 120 + ], + "type": "text", + "content": "The main contributions of the CoMapGS are threefold:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 121, + 553, + 323 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 314, + 121, + 553, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 121, + 553, + 168 + ], + "spans": [ + { + "bbox": [ + 314, + 121, + 553, + 168 + ], + "type": "text", + "content": "1. Covisibility Map-based learning: CoMapGS reframes sparse view synthesis by leveraging covisibility maps for adaptive supervision, enabling balanced guidance across different uncertainty levels within the scene." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 169, + 553, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 169, + 553, + 252 + ], + "spans": [ + { + "bbox": [ + 313, + 169, + 553, + 252 + ], + "type": "text", + "content": "2. Enhanced point cloud initialization: By applying learned scaling parameters between monocular depth estimates and triangulated points registered using the given COLMAP camera poses, CoMapGS densely enhances initial point cloud quality, increasing the potential to densify sparse and underrepresented regions, even in mono-view areas." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 253, + 553, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 253, + 553, + 323 + ], + "spans": [ + { + "bbox": [ + 313, + 253, + 553, + 323 + ], + "type": "text", + "content": "3. Weighted supervision using covisibility map: This strategy adaptively reinforces high- and low-uncertainty regions, strengthening supervision based on region-specific covisibility levels and stabilizing training with learned geometric constraints that extend beyond the view frustum." + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 325, + 553, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 325, + 553, + 374 + ], + "spans": [ + { + "bbox": [ + 313, + 325, + 553, + 374 + ], + "type": "text", + "content": "To the best of our knowledge, CoMapGS is the first approach to explicitly address and recover high-uncertainty regions in sparse view synthesis, effectively balancing supervision across regions with varying uncertainty levels." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 314, + 382, + 405, + 395 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 382, + 405, + 395 + ], + "spans": [ + { + "bbox": [ + 314, + 382, + 405, + 395 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 403, + 555, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 403, + 555, + 617 + ], + "spans": [ + { + "bbox": [ + 313, + 403, + 555, + 617 + ], + "type": "text", + "content": "Sparse novel view synthesis. The growing interest in reconstructing spatial photographs from minimal input image sets has driven the development of various methods aimed at generating novel views from just a few shots, utilizing approaches based on NeRF [6, 11, 13, 21, 23, 26, 28, 30, 31, 36] and 3DGS [17, 22, 32, 34, 38, 41]. Due to the inherent challenges of sparse view synthesis, such as the limited number of training views, NeRF-based studies have primarily focused on proposing regularization techniques with depth ranking [6, 15, 23, 26, 28, 31], frequency-based loss [36], semantic consistency with visual encoders [11], between-view smoothness [13], or uncertainty-aware adaptive loss [33] for both seen and unseen views [11, 15, 21, 30]. Some methods have also addressed uncertainty issues [13, 36], particularly for high-uncertainty mono-view or unseen regions. However, these methods either disregard [13] or penalize [36] uncertain areas, rather than fully utilizing the information contained within." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "text", + "content": "While recent 3DGS-based methods continue to rely on regularization techniques, such as depth constraints [17, 22, 32, 34, 41], some have additionally introduced novel densification strategies [38, 41] and explored enhancing initial point clouds [4, 7, 22, 34], capitalizing on the technical characteristics of 3DGS. These methods primarily focus on improving effective multiview regions [9], often biased towards well-represented regions in the training views, rather" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "26780" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 70, + 553, + 246 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 553, + 246 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 553, + 246 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 553, + 246 + ], + "type": "image", + "image_path": "6fc020d41154632b362667b859297b38d6e31a8f398eb8d8cc0b0e616e0ed1b4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 253, + 555, + 301 + ], + "lines": [ + { + "bbox": [ + 54, + 253, + 555, + 301 + ], + "spans": [ + { + "bbox": [ + 54, + 253, + 555, + 301 + ], + "type": "text", + "content": "Figure 2. Overview of the proposed CoMapGS. CoMapGS leverages existing modules, including Structure-from-Motion (SfM), monocular depth, and dense correspondence prediction, to produce preliminary outputs. Covisibility maps are generated and used alongside module outputs to enhance initial point clouds, which initialize Gaussians for CoMapGS. During training, covisibility map-based weighted supervision with the proximity classifier applies varied strengths to Gaussians based on region-specific covisibility." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 319, + 297, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 319, + 297, + 452 + ], + "spans": [ + { + "bbox": [ + 54, + 319, + 297, + 452 + ], + "type": "text", + "content": "than addressing recovery strategies for high-uncertainty mono-view regions. To overcome the inherent limitations of sparse training views, several prior works have explored generating and utilizing pseudo-ground truths for novel viewpoints via mono-depth estimation [41], forward warping with reversed bilinear sampling [34], or diffusion models with CLIP embedding [30]. While these approaches compensate for the lack of training views, they still face unique challenges in recovering unseen or mono-view regions from the original training view images, potentially misleading the learning process in high-uncertainty areas." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 474, + 297, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 474, + 297, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 474, + 297, + 715 + ], + "type": "text", + "content": "Covisibility-aware supervision. Beginning with Dy-Check [9], which emphasizes the importance of covisibility in multiview regions to address fair evaluation in dynamic novel view synthesis, recent studies have explored integrating covisibility information into supervision strategies. NeRFVS [35] uses a view coverage map to enforce stricter regularization in low-coverage (covisible) regions, but it is primarily suited for dense view synthesis with video input, where extensive multiview overlap is expected. Similarly, InfoNorm [29] leverages mutual information among surface normals at highly correlated scene points to regularize geometric modeling. However, InfoNorm captures covisible areas within local subregions in structured scenes (e.g., rooms) rather than using a comprehensive covisibility map, which limits its applicability for addressing both multiview and monoview regions. While these methods demonstrate the value of covisibility in supervision, they lack a framework that addresses the challenges of sparse novel view synthesis for natural scenes, including outdoor environments, through a pixel-wise covisibility map." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 318, + 549, + 332 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 318, + 549, + 332 + ], + "spans": [ + { + "bbox": [ + 313, + 318, + 549, + 332 + ], + "type": "text", + "content": "3. Covisibility Map-based Gaussian Splitting" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 338, + 555, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 338, + 555, + 471 + ], + "spans": [ + { + "bbox": [ + 313, + 338, + 555, + 471 + ], + "type": "text", + "content": "The proposed CoMapGS framework is illustrated in Fig. 2. CoMapGS builds on the latest 3DGS-based sparse novel view synthesis method, CoR-GS [38], with two key improvements: (1) updating initial point clouds specifically for sparse and mono-view regions and (2) applying adaptive supervision across varying covisibility regions using a covisibility map and proximity MLP classifier. We begin by generating covisibility maps using the dense correspondence prediction model MASt3R [16], refining it with morphological operations (e.g., erosion and dilation) to support these strategies effectively." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 476, + 471, + 490 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 476, + 471, + 490 + ], + "spans": [ + { + "bbox": [ + 313, + 476, + 471, + 490 + ], + "type": "text", + "content": "3.1. Covisibility Map Generation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 494, + 556, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 494, + 556, + 616 + ], + "spans": [ + { + "bbox": [ + 313, + 494, + 556, + 616 + ], + "type": "text", + "content": "We generate a covisibility map by identifying dense correspondences between each training view image and all other training view images. Given a set of training view images " + }, + { + "bbox": [ + 313, + 494, + 556, + 616 + ], + "type": "inline_equation", + "content": "T = \\{I_1, I_2, \\ldots, I_n\\}" + }, + { + "bbox": [ + 313, + 494, + 556, + 616 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 494, + 556, + 616 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 494, + 556, + 616 + ], + "type": "text", + "content": " is the total number of training views, we compute a covisibility map " + }, + { + "bbox": [ + 313, + 494, + 556, + 616 + ], + "type": "inline_equation", + "content": "M_i" + }, + { + "bbox": [ + 313, + 494, + 556, + 616 + ], + "type": "text", + "content": " for each view " + }, + { + "bbox": [ + 313, + 494, + 556, + 616 + ], + "type": "inline_equation", + "content": "I_i" + }, + { + "bbox": [ + 313, + 494, + 556, + 616 + ], + "type": "text", + "content": ". For every pair " + }, + { + "bbox": [ + 313, + 494, + 556, + 616 + ], + "type": "inline_equation", + "content": "(I_i, I_{\\{j\\}})" + }, + { + "bbox": [ + 313, + 494, + 556, + 616 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 313, + 494, + 556, + 616 + ], + "type": "inline_equation", + "content": "j \\neq i" + }, + { + "bbox": [ + 313, + 494, + 556, + 616 + ], + "type": "text", + "content": ", dense correspondences " + }, + { + "bbox": [ + 313, + 494, + 556, + 616 + ], + "type": "inline_equation", + "content": "C_i^j" + }, + { + "bbox": [ + 313, + 494, + 556, + 616 + ], + "type": "text", + "content": " are predicted by a dense correspondence prediction function " + }, + { + "bbox": [ + 313, + 494, + 556, + 616 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{cor}}" + }, + { + "bbox": [ + 313, + 494, + 556, + 616 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 494, + 556, + 616 + ], + "type": "inline_equation", + "content": "C_i^j = f_{\\mathrm{cor}}(I_i, I_j)" + }, + { + "bbox": [ + 313, + 494, + 556, + 616 + ], + "type": "text", + "content": ". The covisibility map " + }, + { + "bbox": [ + 313, + 494, + 556, + 616 + ], + "type": "inline_equation", + "content": "M_i" + }, + { + "bbox": [ + 313, + 494, + 556, + 616 + ], + "type": "text", + "content": " is initialized as a zero matrix of size " + }, + { + "bbox": [ + 313, + 494, + 556, + 616 + ], + "type": "inline_equation", + "content": "W \\times H" + }, + { + "bbox": [ + 313, + 494, + 556, + 616 + ], + "type": "text", + "content": " and defined as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 328, + 622, + 555, + 656 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 622, + 555, + 656 + ], + "spans": [ + { + "bbox": [ + 328, + 622, + 555, + 656 + ], + "type": "interline_equation", + "content": "M _ {i} (x, y) = \\sum_ {j = 1, j \\neq i} ^ {n} \\delta_ {(x, y) \\in \\mathcal {P} \\left(C _ {i} ^ {j}\\right)}, \\quad \\forall (x, y) \\in I _ {i}, \\tag {1}", + "image_path": "c1b4dbc58bf60e9422f9c9cb8244be12ad1f7d4cb02f3810bd8f9f9b70f0443f.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 662, + 557, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 662, + 557, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 662, + 557, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 662, + 557, + 715 + ], + "type": "inline_equation", + "content": "\\delta_{(x,y)\\in \\mathcal{P}(C_i^j)}" + }, + { + "bbox": [ + 313, + 662, + 557, + 715 + ], + "type": "text", + "content": " is an indicator variable returning 1 if pixel " + }, + { + "bbox": [ + 313, + 662, + 557, + 715 + ], + "type": "inline_equation", + "content": "(x,y)" + }, + { + "bbox": [ + 313, + 662, + 557, + 715 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 313, + 662, + 557, + 715 + ], + "type": "inline_equation", + "content": "I_{i}" + }, + { + "bbox": [ + 313, + 662, + 557, + 715 + ], + "type": "text", + "content": " has a correspondence match in " + }, + { + "bbox": [ + 313, + 662, + 557, + 715 + ], + "type": "inline_equation", + "content": "I_{j}" + }, + { + "bbox": [ + 313, + 662, + 557, + 715 + ], + "type": "text", + "content": ", and 0 otherwise. Here, " + }, + { + "bbox": [ + 313, + 662, + 557, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{P}(C_i^j)" + }, + { + "bbox": [ + 313, + 662, + 557, + 715 + ], + "type": "text", + "content": " represents the set of pixel coordinates in " + }, + { + "bbox": [ + 313, + 662, + 557, + 715 + ], + "type": "inline_equation", + "content": "I_{i}" + }, + { + "bbox": [ + 313, + 662, + 557, + 715 + ], + "type": "text", + "content": " that have correspondences in " + }, + { + "bbox": [ + 313, + 662, + 557, + 715 + ], + "type": "inline_equation", + "content": "I_{j}" + }, + { + "bbox": [ + 313, + 662, + 557, + 715 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "text", + "content": "26781" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 192 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 192 + ], + "type": "text", + "content": "The resulting covisibility map " + }, + { + "bbox": [ + 55, + 72, + 296, + 192 + ], + "type": "inline_equation", + "content": "M_{i}" + }, + { + "bbox": [ + 55, + 72, + 296, + 192 + ], + "type": "text", + "content": " captures pixel-wise covisibility counts ranging from 0 to " + }, + { + "bbox": [ + 55, + 72, + 296, + 192 + ], + "type": "inline_equation", + "content": "n - 1" + }, + { + "bbox": [ + 55, + 72, + 296, + 192 + ], + "type": "text", + "content": ", indicating the frequency with which a pixel appears across views, as shown in Fig. 1, with values rescaled from 0 (black) to 255 (white) for visualization purposes. Unlike DyCheck [9]'s covisibility mask, which only indicates pixel visibility, our covisibility map provides precise covisibility details (i.e., covisible view count) at the pixel level. This map supports both the initial point cloud update (Sec. 3.2) and the weighted supervision method (Sec. 3.3)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 198, + 224, + 210 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 198, + 224, + 210 + ], + "spans": [ + { + "bbox": [ + 55, + 198, + 224, + 210 + ], + "type": "text", + "content": "3.2. Enhancing Initial Point Clouds" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 215, + 296, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 215, + 296, + 396 + ], + "spans": [ + { + "bbox": [ + 55, + 215, + 296, + 396 + ], + "type": "text", + "content": "In this subsection, we explain how to improve initial point clouds for both low-uncertainty multi-view and high-uncertainty mono-view regions. To enhance initial point clouds in low-uncertainty multi-view regions where sparse SfM points are available, we integrate a triangulated set of 3D points, " + }, + { + "bbox": [ + 55, + 215, + 296, + 396 + ], + "type": "inline_equation", + "content": "P_T" + }, + { + "bbox": [ + 55, + 215, + 296, + 396 + ], + "type": "text", + "content": ", derived from dense correspondences obtained from MASt3R [16], to complement COLMAP's point cloud, " + }, + { + "bbox": [ + 55, + 215, + 296, + 396 + ], + "type": "inline_equation", + "content": "P_C" + }, + { + "bbox": [ + 55, + 215, + 296, + 396 + ], + "type": "text", + "content": ". This approach enhances geometrically plausible structures, especially in regions where COLMAP's keypoint-based matching is sparse. For each view " + }, + { + "bbox": [ + 55, + 215, + 296, + 396 + ], + "type": "inline_equation", + "content": "I_i" + }, + { + "bbox": [ + 55, + 215, + 296, + 396 + ], + "type": "text", + "content": ", we generate " + }, + { + "bbox": [ + 55, + 215, + 296, + 396 + ], + "type": "inline_equation", + "content": "P_T" + }, + { + "bbox": [ + 55, + 215, + 296, + 396 + ], + "type": "text", + "content": " by triangulating dense correspondences " + }, + { + "bbox": [ + 55, + 215, + 296, + 396 + ], + "type": "inline_equation", + "content": "C_i^{\\{j\\}}" + }, + { + "bbox": [ + 55, + 215, + 296, + 396 + ], + "type": "text", + "content": " from view " + }, + { + "bbox": [ + 55, + 215, + 296, + 396 + ], + "type": "inline_equation", + "content": "I_i" + }, + { + "bbox": [ + 55, + 215, + 296, + 396 + ], + "type": "text", + "content": " to other views " + }, + { + "bbox": [ + 55, + 215, + 296, + 396 + ], + "type": "inline_equation", + "content": "I_{\\{j\\}}" + }, + { + "bbox": [ + 55, + 215, + 296, + 396 + ], + "type": "text", + "content": " (e.g., two or three correspondences in a 3-view case). To avoid redundancy, correspondences already reconstructed from other views are excluded from subsequent triangulations." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 396, + 272, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 396, + 272, + 408 + ], + "spans": [ + { + "bbox": [ + 67, + 396, + 272, + 408 + ], + "type": "text", + "content": "Formally, we define the updated point cloud " + }, + { + "bbox": [ + 67, + 396, + 272, + 408 + ], + "type": "inline_equation", + "content": "P_{\\mathrm{u}}" + }, + { + "bbox": [ + 67, + 396, + 272, + 408 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 88, + 413, + 295, + 427 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 413, + 295, + 427 + ], + "spans": [ + { + "bbox": [ + 88, + 413, + 295, + 427 + ], + "type": "interline_equation", + "content": "P _ {\\mathrm {u}} = P _ {C} \\cup \\left\\{p _ {T} \\in P _ {T} \\mid D \\left(p _ {T}, P _ {C}\\right) > \\epsilon \\right\\}, \\tag {2}", + "image_path": "6dec520d6c8cab74d5caa6aef2a8f2ce8e5ea0fdb03d3a77284ca0cd9af76411.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 432, + 296, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 432, + 296, + 528 + ], + "spans": [ + { + "bbox": [ + 55, + 432, + 296, + 528 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 432, + 296, + 528 + ], + "type": "inline_equation", + "content": "D(p_{T}, P_{C})" + }, + { + "bbox": [ + 55, + 432, + 296, + 528 + ], + "type": "text", + "content": " is the distance from a point " + }, + { + "bbox": [ + 55, + 432, + 296, + 528 + ], + "type": "inline_equation", + "content": "p_{T}" + }, + { + "bbox": [ + 55, + 432, + 296, + 528 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 55, + 432, + 296, + 528 + ], + "type": "inline_equation", + "content": "P_{T}" + }, + { + "bbox": [ + 55, + 432, + 296, + 528 + ], + "type": "text", + "content": " to its nearest neighbor in " + }, + { + "bbox": [ + 55, + 432, + 296, + 528 + ], + "type": "inline_equation", + "content": "P_{C}" + }, + { + "bbox": [ + 55, + 432, + 296, + 528 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 432, + 296, + 528 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 55, + 432, + 296, + 528 + ], + "type": "text", + "content": " is a threshold ensuring that only points sufficiently distant from existing points in " + }, + { + "bbox": [ + 55, + 432, + 296, + 528 + ], + "type": "inline_equation", + "content": "P_{C}" + }, + { + "bbox": [ + 55, + 432, + 296, + 528 + ], + "type": "text", + "content": " are added. This selective integration allows the triangulated points to supplement COLMAP's sparse initial point cloud in the same coordinate system, filling gaps in low-uncertainty regions and improving overall density in covisible areas, as shown in Fig. 3." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 529, + 296, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 529, + 296, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 529, + 296, + 696 + ], + "type": "text", + "content": "In high-uncertainty mono-view regions, we use monocular depth estimation [10] to generate a preliminary set of 3D points, " + }, + { + "bbox": [ + 55, + 529, + 296, + 696 + ], + "type": "inline_equation", + "content": "P_{d}" + }, + { + "bbox": [ + 55, + 529, + 296, + 696 + ], + "type": "text", + "content": ", by unprojecting depth estimates using the provided camera intrinsics. This set is defined as " + }, + { + "bbox": [ + 55, + 529, + 296, + 696 + ], + "type": "inline_equation", + "content": "P_{d} = P_{d}^{\\mathrm{low}} \\cup P_{d}^{\\mathrm{high}}" + }, + { + "bbox": [ + 55, + 529, + 296, + 696 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 529, + 296, + 696 + ], + "type": "inline_equation", + "content": "P_{d}^{\\mathrm{low}}" + }, + { + "bbox": [ + 55, + 529, + 296, + 696 + ], + "type": "text", + "content": " comprises points unprojected from multiview regions (where " + }, + { + "bbox": [ + 55, + 529, + 296, + 696 + ], + "type": "inline_equation", + "content": "M_{i} \\geq 1" + }, + { + "bbox": [ + 55, + 529, + 296, + 696 + ], + "type": "text", + "content": "), and " + }, + { + "bbox": [ + 55, + 529, + 296, + 696 + ], + "type": "inline_equation", + "content": "P_{d}^{\\mathrm{high}}" + }, + { + "bbox": [ + 55, + 529, + 296, + 696 + ], + "type": "text", + "content": " comprises points unprojected from mono-view regions (where " + }, + { + "bbox": [ + 55, + 529, + 296, + 696 + ], + "type": "inline_equation", + "content": "M_{i} = 0" + }, + { + "bbox": [ + 55, + 529, + 296, + 696 + ], + "type": "text", + "content": "). To align these randomly scaled unprojected points with triangulated points from multiview geometry, we learn an anisotropic scaling transformation " + }, + { + "bbox": [ + 55, + 529, + 296, + 696 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{scale}}" + }, + { + "bbox": [ + 55, + 529, + 296, + 696 + ], + "type": "text", + "content": ", using covisible points from " + }, + { + "bbox": [ + 55, + 529, + 296, + 696 + ], + "type": "inline_equation", + "content": "P_{u}^{\\mathrm{low}} \\subseteq P_{u}" + }, + { + "bbox": [ + 55, + 529, + 296, + 696 + ], + "type": "text", + "content": " as ground truth and " + }, + { + "bbox": [ + 55, + 529, + 296, + 696 + ], + "type": "inline_equation", + "content": "P_{d}^{\\mathrm{low}}" + }, + { + "bbox": [ + 55, + 529, + 296, + 696 + ], + "type": "text", + "content": " as input. This transformation applies scaling parameters learned through linear regression to match the multiview-derived geometry:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 700, + 295, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 700, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 132, + 700, + 295, + 715 + ], + "type": "interline_equation", + "content": "f _ {\\text {s c a l e}} \\left(P _ {d} ^ {\\text {l o w}}\\right) \\approx P _ {u} ^ {\\text {l o w}}. \\tag {3}", + "image_path": "d417cfdbd8433fa8cbbc1a2dfb7a5ac34e46fd925e80a96428b34cd8d4442e64.jpg" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 317, + 69, + 392, + 132 + ], + "blocks": [ + { + "bbox": [ + 317, + 69, + 392, + 132 + ], + "lines": [ + { + "bbox": [ + 317, + 69, + 392, + 132 + ], + "spans": [ + { + "bbox": [ + 317, + 69, + 392, + 132 + ], + "type": "image", + "image_path": "7ac89f152a45c3aed602face5f0743f28110c5ad04e7e71c0549e6f47f736e12.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 348, + 133, + 359, + 140 + ], + "lines": [ + { + "bbox": [ + 348, + 133, + 359, + 140 + ], + "spans": [ + { + "bbox": [ + 348, + 133, + 359, + 140 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 392, + 69, + 471, + 132 + ], + "blocks": [ + { + "bbox": [ + 392, + 69, + 471, + 132 + ], + "lines": [ + { + "bbox": [ + 392, + 69, + 471, + 132 + ], + "spans": [ + { + "bbox": [ + 392, + 69, + 471, + 132 + ], + "type": "image", + "image_path": "cac3204969951a0f3bce964575f0577b29406bcdd4ae4ff3fbbb61a710f1e9b7.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 427, + 133, + 437, + 140 + ], + "lines": [ + { + "bbox": [ + 427, + 133, + 437, + 140 + ], + "spans": [ + { + "bbox": [ + 427, + 133, + 437, + 140 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 472, + 69, + 551, + 133 + ], + "blocks": [ + { + "bbox": [ + 472, + 69, + 551, + 133 + ], + "lines": [ + { + "bbox": [ + 472, + 69, + 551, + 133 + ], + "spans": [ + { + "bbox": [ + 472, + 69, + 551, + 133 + ], + "type": "image", + "image_path": "0e6f6a6ad6de1689732fb55910d5654658f039daf54df1b211cb8798d412a638.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 506, + 133, + 515, + 140 + ], + "lines": [ + { + "bbox": [ + 506, + 133, + 515, + 140 + ], + "spans": [ + { + "bbox": [ + 506, + 133, + 515, + 140 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 152, + 554, + 208 + ], + "lines": [ + { + "bbox": [ + 313, + 152, + 554, + 208 + ], + "spans": [ + { + "bbox": [ + 313, + 152, + 554, + 208 + ], + "type": "text", + "content": "Figure 3. Examples of enhanced initial PCLs from the Fern scene in the LLFF dataset. (a) Initial SfM points are first (b) updated through triangulation using dense correspondences and COLMAP camera poses, and then (c) fully refined to include mono-view regions (mono-view updates are colored in red for visualization)." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 226, + 554, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 226, + 554, + 251 + ], + "spans": [ + { + "bbox": [ + 313, + 226, + 554, + 251 + ], + "type": "text", + "content": "For consistency, we define " + }, + { + "bbox": [ + 313, + 226, + 554, + 251 + ], + "type": "inline_equation", + "content": "P_{u} = P_{u}^{\\mathrm{low}} \\cup P_{u}^{\\mathrm{high}}" + }, + { + "bbox": [ + 313, + 226, + 554, + 251 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 226, + 554, + 251 + ], + "type": "inline_equation", + "content": "P_{u}^{\\mathrm{low}}" + }, + { + "bbox": [ + 313, + 226, + 554, + 251 + ], + "type": "text", + "content": " is selected based on covisibility:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 361, + 258, + 554, + 273 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 258, + 554, + 273 + ], + "spans": [ + { + "bbox": [ + 361, + 258, + 554, + 273 + ], + "type": "interline_equation", + "content": "P _ {u} ^ {\\text {l o w}} = \\left\\{p \\mid M _ {i} \\left(\\pi \\left(P _ {u}, \\mathbf {H} _ {i}\\right)\\right) \\geq 1 \\right\\}, \\tag {4}", + "image_path": "3207aba6af2dd62efa16b31ffefa96e4b8c46d888c9b66b539ade635c450822f.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 280, + 554, + 328 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 280, + 554, + 328 + ], + "spans": [ + { + "bbox": [ + 313, + 280, + 554, + 328 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 280, + 554, + 328 + ], + "type": "inline_equation", + "content": "\\pi(P_u, \\mathbf{H}_i)" + }, + { + "bbox": [ + 313, + 280, + 554, + 328 + ], + "type": "text", + "content": " denotes the projection of " + }, + { + "bbox": [ + 313, + 280, + 554, + 328 + ], + "type": "inline_equation", + "content": "P_u" + }, + { + "bbox": [ + 313, + 280, + 554, + 328 + ], + "type": "text", + "content": " using the camera transformation matrix " + }, + { + "bbox": [ + 313, + 280, + 554, + 328 + ], + "type": "inline_equation", + "content": "\\mathbf{H}_i" + }, + { + "bbox": [ + 313, + 280, + 554, + 328 + ], + "type": "text", + "content": ", yielding the image coordinates " + }, + { + "bbox": [ + 313, + 280, + 554, + 328 + ], + "type": "inline_equation", + "content": "(x, y)" + }, + { + "bbox": [ + 313, + 280, + 554, + 328 + ], + "type": "text", + "content": " for a given camera viewpoint and ensuring inclusion of only points visible in multiple views." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 329, + 554, + 364 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 329, + 554, + 364 + ], + "spans": [ + { + "bbox": [ + 313, + 329, + 554, + 364 + ], + "type": "text", + "content": "After training, we apply " + }, + { + "bbox": [ + 313, + 329, + 554, + 364 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{scale}}" + }, + { + "bbox": [ + 313, + 329, + 554, + 364 + ], + "type": "text", + "content": " to the randomly scaled unprojected points from depth estimates in high-uncertainty mono-view regions:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 389, + 370, + 553, + 386 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 389, + 370, + 553, + 386 + ], + "spans": [ + { + "bbox": [ + 389, + 370, + 553, + 386 + ], + "type": "interline_equation", + "content": "P _ {s} ^ {\\text {h i g h}} = f _ {\\text {s c a l e}} \\left(P _ {d} ^ {\\text {h i g h}}\\right). \\tag {5}", + "image_path": "2b97d4cb3365ca30a0cafed0dcf76b5f8b50fcd4892b6b0a87a01d1c95cddd01.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 394, + 554, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 394, + 554, + 445 + ], + "spans": [ + { + "bbox": [ + 313, + 394, + 554, + 445 + ], + "type": "text", + "content": "This transformation aligns the randomly scaled points " + }, + { + "bbox": [ + 313, + 394, + 554, + 445 + ], + "type": "inline_equation", + "content": "P_{d}^{\\mathrm{high}}" + }, + { + "bbox": [ + 313, + 394, + 554, + 445 + ], + "type": "text", + "content": " to the scale of COLMAP's triangulated points " + }, + { + "bbox": [ + 313, + 394, + 554, + 445 + ], + "type": "inline_equation", + "content": "P_{s}^{\\mathrm{high}}" + }, + { + "bbox": [ + 313, + 394, + 554, + 445 + ], + "type": "text", + "content": ", ensuring spatial consistency across regions. The final point cloud, " + }, + { + "bbox": [ + 313, + 394, + 554, + 445 + ], + "type": "inline_equation", + "content": "P_{\\mathrm{final}}" + }, + { + "bbox": [ + 313, + 394, + 554, + 445 + ], + "type": "text", + "content": ", is defined as:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 389, + 451, + 553, + 467 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 389, + 451, + 553, + 467 + ], + "spans": [ + { + "bbox": [ + 389, + 451, + 553, + 467 + ], + "type": "interline_equation", + "content": "P _ {\\text {f i n a l}} = P _ {u} ^ {\\text {l o w}} \\cup P _ {s} ^ {\\text {h i g h}}, \\tag {6}", + "image_path": "f2e84471095ff80a6202891b17ce9586c1894fa2c12a8f77617f14c4509c03b9.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 474, + 554, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 474, + 554, + 521 + ], + "spans": [ + { + "bbox": [ + 313, + 474, + 554, + 521 + ], + "type": "text", + "content": "resulting in enhanced initial point cloud density and accuracy across both multiview and mono-view regions, as shown in Fig. 3, thus supporting cohesive scene reconstruction in sparse view synthesis." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 528, + 552, + 541 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 528, + 552, + 541 + ], + "spans": [ + { + "bbox": [ + 313, + 528, + 552, + 541 + ], + "type": "text", + "content": "3.3. Covisibility Map-Based Weighted Supervision" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 545, + 554, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 554, + 666 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 554, + 666 + ], + "type": "text", + "content": "We propose a covisibility map-based supervision strategy that leverages the enhanced initial point cloud, " + }, + { + "bbox": [ + 313, + 545, + 554, + 666 + ], + "type": "inline_equation", + "content": "P_{\\mathrm{final}}" + }, + { + "bbox": [ + 313, + 545, + 554, + 666 + ], + "type": "text", + "content": ", and the covisibility map " + }, + { + "bbox": [ + 313, + 545, + 554, + 666 + ], + "type": "inline_equation", + "content": "M_{i}" + }, + { + "bbox": [ + 313, + 545, + 554, + 666 + ], + "type": "text", + "content": " to adaptively supervise regions based on uncertainty levels represented by covisibility counts, as shown in Fig. 4. This approach improves conventional sparse view Gaussian splatting methods [38, 41] by introducing region-specific weighting across covisibility levels, " + }, + { + "bbox": [ + 313, + 545, + 554, + 666 + ], + "type": "inline_equation", + "content": "M_{i} = \\{0:n - 1\\}" + }, + { + "bbox": [ + 313, + 545, + 554, + 666 + ], + "type": "text", + "content": ", and by supervising Gaussians located outside the training view frustum with a scene-level weight based on the average covisibility score " + }, + { + "bbox": [ + 313, + 545, + 554, + 666 + ], + "type": "inline_equation", + "content": "S = \\mathrm{AVG}(M_{[1:n]})" + }, + { + "bbox": [ + 313, + 545, + 554, + 666 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 666, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 666, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 666, + 554, + 713 + ], + "type": "text", + "content": "Our approach integrates seamlessly with the latest 3DGS-based sparse view synthesis methods [38, 41] by adding a proximity loss term, weighted using inverse proximity scores predicted by a three-layer MLP classifier" + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "26782" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "type": "text", + "content": "trained on " + }, + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "type": "inline_equation", + "content": "P_{\\mathrm{final}}" + }, + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "type": "text", + "content": " (see details in the supplementary materials)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 96, + 296, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 96, + 296, + 180 + ], + "spans": [ + { + "bbox": [ + 55, + 96, + 296, + 180 + ], + "type": "text", + "content": "Proximity loss with adaptive weighting. To implement this approach, we first train an MLP classifier " + }, + { + "bbox": [ + 55, + 96, + 296, + 180 + ], + "type": "inline_equation", + "content": "f_{p}" + }, + { + "bbox": [ + 55, + 96, + 296, + 180 + ], + "type": "text", + "content": " on " + }, + { + "bbox": [ + 55, + 96, + 296, + 180 + ], + "type": "inline_equation", + "content": "P_{\\mathrm{final}}" + }, + { + "bbox": [ + 55, + 96, + 296, + 180 + ], + "type": "text", + "content": ", labeling these curated scene 3D points as positive (representing the original scene geometry) and randomly distributed points, " + }, + { + "bbox": [ + 55, + 96, + 296, + 180 + ], + "type": "inline_equation", + "content": "P_{\\mathrm{random}}" + }, + { + "bbox": [ + 55, + 96, + 296, + 180 + ], + "type": "text", + "content": ", constrained to remain distant from " + }, + { + "bbox": [ + 55, + 96, + 296, + 180 + ], + "type": "inline_equation", + "content": "P_{\\mathrm{final}}" + }, + { + "bbox": [ + 55, + 96, + 296, + 180 + ], + "type": "text", + "content": ", as negative. The classifier " + }, + { + "bbox": [ + 55, + 96, + 296, + 180 + ], + "type": "inline_equation", + "content": "f_{p}: \\mathbb{R}^{3} \\to [0,1]" + }, + { + "bbox": [ + 55, + 96, + 296, + 180 + ], + "type": "text", + "content": " outputs a proximity score " + }, + { + "bbox": [ + 55, + 96, + 296, + 180 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 55, + 96, + 296, + 180 + ], + "type": "text", + "content": " for each 3D point " + }, + { + "bbox": [ + 55, + 96, + 296, + 180 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 55, + 96, + 296, + 180 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 135, + 187, + 295, + 201 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 187, + 295, + 201 + ], + "spans": [ + { + "bbox": [ + 135, + 187, + 295, + 201 + ], + "type": "interline_equation", + "content": "s = f _ {p} (p) \\in [ 0, 1 ], \\tag {7}", + "image_path": "dc2ce34b80547937b2a7da38cb700b5055bb02dd3c98066a100f0a0b102d640f.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 207, + 296, + 256 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 207, + 296, + 256 + ], + "spans": [ + { + "bbox": [ + 55, + 207, + 296, + 256 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 207, + 296, + 256 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 55, + 207, + 296, + 256 + ], + "type": "text", + "content": " approaches 1 for points close to the scene geometry and 0 for distant points. Using batch processing, we compute the proximity scores efficiently for the entire Gaussian set, " + }, + { + "bbox": [ + 55, + 207, + 296, + 256 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 55, + 207, + 296, + 256 + ], + "type": "text", + "content": ", and define the proximity loss " + }, + { + "bbox": [ + 55, + 207, + 296, + 256 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_p" + }, + { + "bbox": [ + 55, + 207, + 296, + 256 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 261, + 295, + 291 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 261, + 295, + 291 + ], + "spans": [ + { + "bbox": [ + 61, + 261, + 295, + 291 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {p} = \\frac {1}{| G |} \\sum_ {g \\in G} (\\chi (g) w _ {\\text {i n}} + (1 - \\chi (g)) w _ {\\text {o u t}}) \\cdot (1 - s), \\tag {8}", + "image_path": "e89749159df50367e38e98778b8edc0477e90a2d3b5bca803a3c84884f122c51.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 297, + 296, + 333 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 297, + 296, + 333 + ], + "spans": [ + { + "bbox": [ + 55, + 297, + 296, + 333 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 297, + 296, + 333 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 55, + 297, + 296, + 333 + ], + "type": "text", + "content": " denotes the 3D position of a Gaussian in the set " + }, + { + "bbox": [ + 55, + 297, + 296, + 333 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 55, + 297, + 296, + 333 + ], + "type": "text", + "content": ", representing the complete Gaussian set. The indicator function " + }, + { + "bbox": [ + 55, + 297, + 296, + 333 + ], + "type": "inline_equation", + "content": "\\chi(g)" + }, + { + "bbox": [ + 55, + 297, + 296, + 333 + ], + "type": "text", + "content": " is defined as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 82, + 339, + 295, + 371 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 339, + 295, + 371 + ], + "spans": [ + { + "bbox": [ + 82, + 339, + 295, + 371 + ], + "type": "interline_equation", + "content": "\\chi (g) = \\left\\{ \\begin{array}{l l} 1, & \\text {i f} \\pi (g, \\mathbf {H} _ {i}) \\in [ 0, \\mathsf {W}) \\times [ 0, \\mathsf {H}) \\\\ 0, & \\text {o t h e r w i s e} \\end{array} , \\right. \\tag {9}", + "image_path": "a3aab3f456e10e473e53ab07be50e5f195dfe4e036d287c177ffca9147a63ffd.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 377, + 295, + 401 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 377, + 295, + 401 + ], + "spans": [ + { + "bbox": [ + 55, + 377, + 295, + 401 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 377, + 295, + 401 + ], + "type": "inline_equation", + "content": "\\pi(g, \\mathbf{H}_i)" + }, + { + "bbox": [ + 55, + 377, + 295, + 401 + ], + "type": "text", + "content": " is the projection of " + }, + { + "bbox": [ + 55, + 377, + 295, + 401 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 55, + 377, + 295, + 401 + ], + "type": "text", + "content": " under camera transformation " + }, + { + "bbox": [ + 55, + 377, + 295, + 401 + ], + "type": "inline_equation", + "content": "\\mathbf{H}_i" + }, + { + "bbox": [ + 55, + 377, + 295, + 401 + ], + "type": "text", + "content": " onto the可视ibility map " + }, + { + "bbox": [ + 55, + 377, + 295, + 401 + ], + "type": "inline_equation", + "content": "M_i" + }, + { + "bbox": [ + 55, + 377, + 295, + 401 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 401, + 296, + 436 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 401, + 296, + 436 + ], + "spans": [ + { + "bbox": [ + 55, + 401, + 296, + 436 + ], + "type": "text", + "content": "Within the camera-view frustum. We use the weight " + }, + { + "bbox": [ + 55, + 401, + 296, + 436 + ], + "type": "inline_equation", + "content": "w_{in}" + }, + { + "bbox": [ + 55, + 401, + 296, + 436 + ], + "type": "text", + "content": " for Gaussians projected onto covisibility mask regions, defined as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 119, + 441, + 295, + 466 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 119, + 441, + 295, + 466 + ], + "spans": [ + { + "bbox": [ + 119, + 441, + 295, + 466 + ], + "type": "interline_equation", + "content": "w _ {\\mathrm {i n}} = \\frac {1}{M _ {i} \\left(\\pi \\left(g , \\mathbf {H} _ {i}\\right)\\right) + 1}, \\tag {10}", + "image_path": "c8a7435fc6d2fda663ea1b09751432068d2ffe990f821a4dfe93a359ba3afbb2.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 472, + 295, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 472, + 295, + 544 + ], + "spans": [ + { + "bbox": [ + 55, + 472, + 295, + 544 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 472, + 295, + 544 + ], + "type": "inline_equation", + "content": "M_{i}(\\pi (g,\\mathbf{H}i))" + }, + { + "bbox": [ + 55, + 472, + 295, + 544 + ], + "type": "text", + "content": " is the covisibility count of " + }, + { + "bbox": [ + 55, + 472, + 295, + 544 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 55, + 472, + 295, + 544 + ], + "type": "text", + "content": " in the covisibility map for the corresponding projection. This weighting ensures that high-covisibility regions, which are frequently supervised by other loss terms, receive weaker proximity-related supervision, while mono-view regions benefit from a more direct application of " + }, + { + "bbox": [ + 55, + 472, + 295, + 544 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_p" + }, + { + "bbox": [ + 55, + 472, + 295, + 544 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 544, + 296, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 544, + 296, + 616 + ], + "spans": [ + { + "bbox": [ + 55, + 544, + 296, + 616 + ], + "type": "text", + "content": "Outside the camera-view frustum. In conventional 3DGS-based methods, only Gaussians within the view frustum are supervised per iteration. However, for scenes with high covisibility scores (" + }, + { + "bbox": [ + 55, + 544, + 296, + 616 + ], + "type": "inline_equation", + "content": "S > 0.7" + }, + { + "bbox": [ + 55, + 544, + 296, + 616 + ], + "type": "text", + "content": "), we extend supervision to Gaussians located outside the view frustum, applying a scene-level weight " + }, + { + "bbox": [ + 55, + 544, + 296, + 616 + ], + "type": "inline_equation", + "content": "w_{out}" + }, + { + "bbox": [ + 55, + 544, + 296, + 616 + ], + "type": "text", + "content": " for these points:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 115, + 621, + 295, + 648 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 621, + 295, + 648 + ], + "spans": [ + { + "bbox": [ + 115, + 621, + 295, + 648 + ], + "type": "interline_equation", + "content": "w _ {o u t} = \\max \\left(0, \\frac {S - 0 . 7}{0 . 3}\\right), \\tag {11}", + "image_path": "a62ff48e302fc16a570f90f5aafd8a585d7a30fcdfe807000d1d7880d4635930.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 653, + 296, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 712 + ], + "type": "text", + "content": "which scales linearly from 1 to 0 as " + }, + { + "bbox": [ + 55, + 653, + 296, + 712 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 55, + 653, + 296, + 712 + ], + "type": "text", + "content": " ranges from 1 to 0.7. This approach improves supervision outside the view frustum in high-covisibility score scenes, ensuring comprehensive guidance while making efficient use of the proximity classifier." + } + ] + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 316, + 71, + 552, + 187 + ], + "blocks": [ + { + "bbox": [ + 316, + 71, + 552, + 187 + ], + "lines": [ + { + "bbox": [ + 316, + 71, + 552, + 187 + ], + "spans": [ + { + "bbox": [ + 316, + 71, + 552, + 187 + ], + "type": "image", + "image_path": "7a91f338337b3a8e1d39c6031da96fbb1027443d5d666fac143882504ccb23d4.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 194, + 555, + 239 + ], + "lines": [ + { + "bbox": [ + 313, + 194, + 555, + 239 + ], + "spans": [ + { + "bbox": [ + 313, + 194, + 555, + 239 + ], + "type": "text", + "content": "Figure 4. Alongside (a) the conventional reconstruction loss " + }, + { + "bbox": [ + 313, + 194, + 555, + 239 + ], + "type": "inline_equation", + "content": "\\mathcal{L}1" + }, + { + "bbox": [ + 313, + 194, + 555, + 239 + ], + "type": "text", + "content": ", we add (b) a weighted proximity loss " + }, + { + "bbox": [ + 313, + 194, + 555, + 239 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_p" + }, + { + "bbox": [ + 313, + 194, + 555, + 239 + ], + "type": "text", + "content": " that applies classifier results with assigned weights based on Gaussian projections onto the可视ability map and their positions relative to the frustum." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 261, + 427, + 274 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 261, + 427, + 274 + ], + "spans": [ + { + "bbox": [ + 313, + 261, + 427, + 274 + ], + "type": "text", + "content": "3.4. Objective Function" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 279, + 555, + 351 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 279, + 555, + 351 + ], + "spans": [ + { + "bbox": [ + 313, + 279, + 555, + 351 + ], + "type": "text", + "content": "Our proposed CoMapGS enhances the objective function by introducing a proximity loss term, effectively complementing the function while preserving the original strategies of state-of-the-art methods, such as Co-pruning and Pseudoview Co-regularization from CoR-GS [38]. The full objective function is expressed as:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 326, + 361, + 554, + 376 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 361, + 554, + 376 + ], + "spans": [ + { + "bbox": [ + 326, + 361, + 554, + 376 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = (1 - \\lambda) \\mathcal {L} 1 \\left(I, I ^ {*}\\right) + \\lambda \\mathcal {L} _ {\\mathrm {D} - \\mathrm {S S I M}} \\left(I, I ^ {*}\\right) + \\mathcal {L} _ {p}, \\tag {12}", + "image_path": "64033c14dc496532734a98de2961f5488fe10dac35d8f7c906d9d38e572adbc9.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 385, + 554, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 385, + 554, + 469 + ], + "spans": [ + { + "bbox": [ + 313, + 385, + 554, + 469 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 385, + 554, + 469 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 313, + 385, + 554, + 469 + ], + "type": "text", + "content": " is the rendered image, supervised against ground truth " + }, + { + "bbox": [ + 313, + 385, + 554, + 469 + ], + "type": "inline_equation", + "content": "I^{*}" + }, + { + "bbox": [ + 313, + 385, + 554, + 469 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 385, + 554, + 469 + ], + "type": "inline_equation", + "content": "\\mathcal{L}1" + }, + { + "bbox": [ + 313, + 385, + 554, + 469 + ], + "type": "text", + "content": " is the L1 reconstruction loss, and " + }, + { + "bbox": [ + 313, + 385, + 554, + 469 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{D - SSIM}}" + }, + { + "bbox": [ + 313, + 385, + 554, + 469 + ], + "type": "text", + "content": " represents the D-SSIM loss term. The balancing parameter " + }, + { + "bbox": [ + 313, + 385, + 554, + 469 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 385, + 554, + 469 + ], + "type": "text", + "content": " controls the weighting between L1 and D-SSIM, in line with standard Gaussian splattering objectives. The proximity loss term " + }, + { + "bbox": [ + 313, + 385, + 554, + 469 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_p" + }, + { + "bbox": [ + 313, + 385, + 554, + 469 + ], + "type": "text", + "content": ", defined in Eq. 8, adds adaptive supervision to improve geometry alignment." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 470, + 554, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 470, + 554, + 518 + ], + "spans": [ + { + "bbox": [ + 313, + 470, + 554, + 518 + ], + "type": "text", + "content": "Our proposed approach extends seamlessly to other recent sparse view synthesis methods, such as FSGS [41], by incorporating " + }, + { + "bbox": [ + 313, + 470, + 554, + 518 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_p" + }, + { + "bbox": [ + 313, + 470, + 554, + 518 + ], + "type": "text", + "content": " into their objective function while preserving their original strategies, as reported in Table 1." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 530, + 440, + 544 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 530, + 440, + 544 + ], + "spans": [ + { + "bbox": [ + 313, + 530, + 440, + 544 + ], + "type": "text", + "content": "4. Experimental Results" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 551, + 364, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 551, + 364, + 563 + ], + "spans": [ + { + "bbox": [ + 313, + 551, + 364, + 563 + ], + "type": "text", + "content": "4.1. Setup" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 569, + 554, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 569, + 554, + 652 + ], + "spans": [ + { + "bbox": [ + 313, + 569, + 554, + 652 + ], + "type": "text", + "content": "Datasets. We conducted experiments on two benchmark datasets: LLFF [19], using 3, 6, and 9 training views with resolutions downsampled to " + }, + { + "bbox": [ + 313, + 569, + 554, + 652 + ], + "type": "inline_equation", + "content": "8 \\times" + }, + { + "bbox": [ + 313, + 569, + 554, + 652 + ], + "type": "text", + "content": ", and Mip-NeRF 360 [2], using 12 and 24 training views per scene with resolutions downsampled to " + }, + { + "bbox": [ + 313, + 569, + 554, + 652 + ], + "type": "inline_equation", + "content": "4 \\times" + }, + { + "bbox": [ + 313, + 569, + 554, + 652 + ], + "type": "text", + "content": ". We adhered to the standard training and testing conventions [38] for datasets commonly used in sparse view synthesis." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 653, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 554, + 713 + ], + "type": "text", + "content": "Comparison Baselines. We compare CoMapGS with existing few-shot novel view synthesis methods, including FSGS [41] and CoR-GS [38]. As our method primarily enhances the performance of 3DGS-based approaches, we reproduced results for these methods. For other methods, we" + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26783" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 65, + 102, + 547, + 317 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "text", + "content": "Table 1. Quantitative results on LLFF with 3, 6, and 9 training views. The best, second-best, and third-best entries are marked in red, orange, and yellow, respectively. " + }, + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "text", + "content": " indicates results reproduced by us using the released code base." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 65, + 102, + 547, + 317 + ], + "lines": [ + { + "bbox": [ + 65, + 102, + 547, + 317 + ], + "spans": [ + { + "bbox": [ + 65, + 102, + 547, + 317 + ], + "type": "table", + "html": "
MethodPSNR ↑SSIM ↑LPIPS ↓
3-view6-view9-view3-view6-view9-view3-view6-view9-view
DietNeRF [11]14.9421.7524.280.3700.7170.8010.4960.2480.183
Mip-NeRF [1]16.1122.9124.880.4010.7560.8260.4600.2130.160
RegNeRF [21]19.0823.1024.860.5870.7600.8200.3360.2060.161
SimpleNeRF [25]19.2423.0523.980.6230.7370.7620.3750.2960.286
FreeNeRF [36]19.6323.7325.130.6120.7790.8270.3080.1950.160
SparseNeRF [28]19.8623.2624.270.7140.7410.7810.2430.2350.228
DiffusioNeRF [31]20.1323.6024.620.6310.7750.8070.3440.2350.216
ReconFusion [30]21.3424.2525.210.7240.8150.8480.2030.1520.134
DNGaussian [17]19.1222.1823.170.5910.7550.7880.2940.1980.180
3DGS [12]19.2223.8025.440.6490.8140.8600.2290.1250.096
FSGS [41]†20.45824.69426.0060.7130.8390.8750.2050.1210.095
CoR-GS [38]†20.47324.77726.4750.7170.8440.8810.1990.1160.086
[41]† + CoMapGS (ours)20.89124.80826.2920.7370.8440.8790.1900.1170.092
[38]† + CoMapGS (ours)21.10525.20426.7310.7470.8540.8860.1820.1080.082
", + "image_path": "af3865340f116d86b37d98feeb210a7dd6bbb618f33e0094fe9b77e8429f00e5.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 58, + 327, + 157, + 402 + ], + "blocks": [ + { + "bbox": [ + 58, + 327, + 157, + 402 + ], + "lines": [ + { + "bbox": [ + 58, + 327, + 157, + 402 + ], + "spans": [ + { + "bbox": [ + 58, + 327, + 157, + 402 + ], + "type": "image", + "image_path": "cdc5b80ddb9038cc63817b9cc6b55aed2a723fc2dc829f9963d54573207839b3.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 59, + 403, + 157, + 476 + ], + "blocks": [ + { + "bbox": [ + 59, + 403, + 157, + 476 + ], + "lines": [ + { + "bbox": [ + 59, + 403, + 157, + 476 + ], + "spans": [ + { + "bbox": [ + 59, + 403, + 157, + 476 + ], + "type": "image", + "image_path": "bb2d96001490401f4cd3b6753cee514f270c6e63b47cc167c83f71bbe56c71a1.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 59, + 477, + 157, + 551 + ], + "blocks": [ + { + "bbox": [ + 59, + 477, + 157, + 551 + ], + "lines": [ + { + "bbox": [ + 59, + 477, + 157, + 551 + ], + "spans": [ + { + "bbox": [ + 59, + 477, + 157, + 551 + ], + "type": "image", + "image_path": "1fd01778b3bda0fd3e2a157c910a5ec072a1c80b53fa32a6c12f730ac44ae21d.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 97, + 553, + 119, + 562 + ], + "lines": [ + { + "bbox": [ + 97, + 553, + 119, + 562 + ], + "spans": [ + { + "bbox": [ + 97, + 553, + 119, + 562 + ], + "type": "text", + "content": "FSGS" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 158, + 327, + 255, + 403 + ], + "blocks": [ + { + "bbox": [ + 158, + 327, + 255, + 403 + ], + "lines": [ + { + "bbox": [ + 158, + 327, + 255, + 403 + ], + "spans": [ + { + "bbox": [ + 158, + 327, + 255, + 403 + ], + "type": "image", + "image_path": "c5391d363b97828f571572ad1bf6b11034cc6075bd3acdf9b09859bff86ce904.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 158, + 403, + 255, + 476 + ], + "blocks": [ + { + "bbox": [ + 158, + 403, + 255, + 476 + ], + "lines": [ + { + "bbox": [ + 158, + 403, + 255, + 476 + ], + "spans": [ + { + "bbox": [ + 158, + 403, + 255, + 476 + ], + "type": "image", + "image_path": "4473f9fd1920f51392b5c01e00b0b306101e5002633053af0dd6a1d0417e7db4.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 158, + 477, + 255, + 551 + ], + "blocks": [ + { + "bbox": [ + 158, + 477, + 255, + 551 + ], + "lines": [ + { + "bbox": [ + 158, + 477, + 255, + 551 + ], + "spans": [ + { + "bbox": [ + 158, + 477, + 255, + 551 + ], + "type": "image", + "image_path": "6e0f8f5891ce569ccecfa80422462dacf1b82806f59c8d96ac9b62219cf971ff.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 192, + 553, + 221, + 562 + ], + "lines": [ + { + "bbox": [ + 192, + 553, + 221, + 562 + ], + "spans": [ + { + "bbox": [ + 192, + 553, + 221, + 562 + ], + "type": "text", + "content": "CoR-GS" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 256, + 327, + 354, + 403 + ], + "blocks": [ + { + "bbox": [ + 256, + 327, + 354, + 403 + ], + "lines": [ + { + "bbox": [ + 256, + 327, + 354, + 403 + ], + "spans": [ + { + "bbox": [ + 256, + 327, + 354, + 403 + ], + "type": "image", + "image_path": "3f9bac9de3ab091468f154e9ed3ab449efdead776ec23e8d4c6803508d45c6fd.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 256, + 403, + 354, + 476 + ], + "blocks": [ + { + "bbox": [ + 256, + 403, + 354, + 476 + ], + "lines": [ + { + "bbox": [ + 256, + 403, + 354, + 476 + ], + "spans": [ + { + "bbox": [ + 256, + 403, + 354, + 476 + ], + "type": "image", + "image_path": "34195533195e7087d72fdb7d3b8205aa224f5434703239bc3c15c672bcfe88d5.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 256, + 477, + 354, + 551 + ], + "blocks": [ + { + "bbox": [ + 256, + 477, + 354, + 551 + ], + "lines": [ + { + "bbox": [ + 256, + 477, + 354, + 551 + ], + "spans": [ + { + "bbox": [ + 256, + 477, + 354, + 551 + ], + "type": "image", + "image_path": "1d53208fa5e2da576cc6539cd1fd4108fd5102a34b86f71152c6a5a0e9799be5.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 296, + 554, + 314, + 562 + ], + "lines": [ + { + "bbox": [ + 296, + 554, + 314, + 562 + ], + "spans": [ + { + "bbox": [ + 296, + 554, + 314, + 562 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 571, + 555, + 605 + ], + "lines": [ + { + "bbox": [ + 55, + 571, + 555, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 571, + 555, + 605 + ], + "type": "text", + "content": "Figure 5. Qualitative comparison on 3 training views from the LLFF dataset. We compare our results with FSGS [41] and CoR-GS [38], demonstrating that our method achieves superior texture synthesis in novel views, particularly in high-uncertainty mono-view regions that appear only once in the training set." + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 355, + 327, + 453, + 403 + ], + "blocks": [ + { + "bbox": [ + 355, + 327, + 453, + 403 + ], + "lines": [ + { + "bbox": [ + 355, + 327, + 453, + 403 + ], + "spans": [ + { + "bbox": [ + 355, + 327, + 453, + 403 + ], + "type": "image", + "image_path": "efbe0a61a970aaff676ad996c7956404d11c4bf77998d5df632ff1f7c4b66992.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 355, + 403, + 452, + 476 + ], + "blocks": [ + { + "bbox": [ + 355, + 403, + 452, + 476 + ], + "lines": [ + { + "bbox": [ + 355, + 403, + 452, + 476 + ], + "spans": [ + { + "bbox": [ + 355, + 403, + 452, + 476 + ], + "type": "image", + "image_path": "bbb2b25365aec124c991a1f3dc5428e2d9b7fa386ee1f516047eb5ee9cc9a8a7.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 355, + 477, + 452, + 551 + ], + "blocks": [ + { + "bbox": [ + 355, + 477, + 452, + 551 + ], + "lines": [ + { + "bbox": [ + 355, + 477, + 452, + 551 + ], + "spans": [ + { + "bbox": [ + 355, + 477, + 452, + 551 + ], + "type": "image", + "image_path": "0e53e929cde0538493dcf585612cb5cf4209e4d7d7845cd26eaafb8ac9551e41.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 381, + 554, + 428, + 562 + ], + "lines": [ + { + "bbox": [ + 381, + 554, + 428, + 562 + ], + "spans": [ + { + "bbox": [ + 381, + 554, + 428, + 562 + ], + "type": "text", + "content": "Ground Truth" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 454, + 327, + 552, + 403 + ], + "blocks": [ + { + "bbox": [ + 454, + 327, + 552, + 403 + ], + "lines": [ + { + "bbox": [ + 454, + 327, + 552, + 403 + ], + "spans": [ + { + "bbox": [ + 454, + 327, + 552, + 403 + ], + "type": "image", + "image_path": "130b119ca75288856f5b6cb312d047e2bbecd3e1c85de00903114c8336554421.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 454, + 403, + 552, + 476 + ], + "blocks": [ + { + "bbox": [ + 454, + 403, + 552, + 476 + ], + "lines": [ + { + "bbox": [ + 454, + 403, + 552, + 476 + ], + "spans": [ + { + "bbox": [ + 454, + 403, + 552, + 476 + ], + "type": "image", + "image_path": "16d3dab5adf8d13e3fb94ee6cca4d753674e7eb29eaeb73a9b2e294308a82183.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 454, + 476, + 552, + 551 + ], + "blocks": [ + { + "bbox": [ + 454, + 476, + 552, + 551 + ], + "lines": [ + { + "bbox": [ + 454, + 476, + 552, + 551 + ], + "spans": [ + { + "bbox": [ + 454, + 476, + 552, + 551 + ], + "type": "image", + "image_path": "f252d99f9e77044a86033fb48eda671de035f1a6441a9b543a08e2cea7e7533d.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 471, + 553, + 533, + 563 + ], + "lines": [ + { + "bbox": [ + 471, + 553, + 533, + 563 + ], + "spans": [ + { + "bbox": [ + 471, + 553, + 533, + 563 + ], + "type": "text", + "content": "Covisibility Mask" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "bbox": [ + 55, + 625, + 295, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 625, + 295, + 649 + ], + "spans": [ + { + "bbox": [ + 55, + 625, + 295, + 649 + ], + "type": "text", + "content": "report the best published scores from prior works, as shown in Tables 1 and 2." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 55, + 651, + 295, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 651, + 295, + 675 + ], + "spans": [ + { + "bbox": [ + 55, + 651, + 295, + 675 + ], + "type": "text", + "content": "Evaluation Metrics. We evaluate our results using PSNR, SSIM, and LPIPS metrics." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": "Implementation. Training was conducted for 10k iterations on the LLFF dataset and 30k iterations on the MipNeRF 360 dataset, following the setup of 3DGS [12]. Con" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 625, + 555, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 625, + 555, + 662 + ], + "spans": [ + { + "bbox": [ + 313, + 625, + 555, + 662 + ], + "type": "text", + "content": "sistent with experimental setups in FSGS [41] and CoR-GS [38], CoMapGS was initialized with point clouds and precomputed camera poses from COLMAP." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 671, + 400, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 671, + 400, + 684 + ], + "spans": [ + { + "bbox": [ + 313, + 671, + 400, + 684 + ], + "type": "text", + "content": "4.2. Comparisons" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": "LLFF. Quantitative and qualitative results on the LLFF dataset are presented in Table 1 and Fig. 5, respectively." + } + ] + } + ], + "index": 28 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "26784" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 113, + 102, + 500, + 188 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "text", + "content": "Table 2. Quantitative results on Mip-NeRF360 with 12 and 24 training views. The best, second-best, and third-best entries are marked in red, orange, and yellow, respectively. " + }, + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "inline_equation", + "content": "{}^{ \\dagger }" + }, + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "text", + "content": " indicates results reproduced by us using the released code base." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 113, + 102, + 500, + 188 + ], + "lines": [ + { + "bbox": [ + 113, + 102, + 500, + 188 + ], + "spans": [ + { + "bbox": [ + 113, + 102, + 500, + 188 + ], + "type": "table", + "html": "
MethodPSNR ↑SSIM ↑LPIPS ↓
12-view24-view12-view24-view12-view24-view
3DGS [12]18.5222.800.5230.7080.4150.276
FSGS [41]18.8023.280.5310.7150.4180.274
CoR-GS [38]†19.16223.3210.5740.7290.4140.271
[38]† + CoMapGS (ours)19.68023.4620.5910.7340.3940.264
", + "image_path": "2ac3ed3e6269f2afab7c9d6e5cdd9ff1fb44528846105a9e835431440a7cad39.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 59, + 196, + 552, + 536 + ], + "blocks": [ + { + "bbox": [ + 59, + 196, + 552, + 536 + ], + "lines": [ + { + "bbox": [ + 59, + 196, + 552, + 536 + ], + "spans": [ + { + "bbox": [ + 59, + 196, + 552, + 536 + ], + "type": "image", + "image_path": "657c6f64c99b25fec1064da2a6118654b62ca309a4d69645f3f1770ab20cd348.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 544, + 555, + 577 + ], + "lines": [ + { + "bbox": [ + 55, + 544, + 555, + 577 + ], + "spans": [ + { + "bbox": [ + 55, + 544, + 555, + 577 + ], + "type": "text", + "content": "Figure 6. Qualitative comparison on 12 training views from the Mip-NeRF 360 dataset. Mip-NeRF 360 highlights diverse covisibility levels, with high-uncertainty regions sparsely captured. CoMapGS effectively synthesizes these regions, while CoR-GS struggles to recover textures." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 598, + 297, + 706 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 598, + 297, + 706 + ], + "spans": [ + { + "bbox": [ + 55, + 598, + 297, + 706 + ], + "type": "text", + "content": "We reproduced evaluation scores for the two baseline methods, FSGS [41] and CoR-GS [38], to highlight the improvements over our reproduced results. The results show consistent improvements, ranking first across various training view counts, with PSNR placing second only to ReconFusion [30] in the 3-view case. This difference is likely due to diffusion-based sparse view synthesis methods, which are optimized for generating visually seamless colors that blend with surrounding regions. However, this approach may not" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 598, + 555, + 682 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 598, + 555, + 682 + ], + "spans": [ + { + "bbox": [ + 313, + 598, + 555, + 682 + ], + "type": "text", + "content": "accurately capture original patterns in mono-view or unseen regions—common in sparse view synthesis datasets, especially with fewer training views. Consequently, while we ranked second in PSNR, our method achieved substantially higher SSIM and lower LPIPS scores under comparable view settings, indicating improved structural fidelity and perceptual quality in the synthesized views." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "content": "Mip-NeRF 360. To validate CoMapGS's applicability to broader scenes that capture large areas with outward-" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26785" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "type": "text", + "content": "facing cameras, we evaluated our method using the MipNeRF 360 dataset. This dataset, which centers objects of interest with outward-facing cameras, captures the motivation of our approach by featuring distinct low-uncertainty, multiview regions alongside high-uncertainty, less-covisible regions within an image, as shown in Fig. 1. In these scenarios, COLMAP struggles to establish stable correspondences, leading to sparse initial point clouds that fail to adequately recover details in high-uncertainty regions, as shown in Fig. 6." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 192, + 295, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 192, + 295, + 335 + ], + "spans": [ + { + "bbox": [ + 55, + 192, + 295, + 335 + ], + "type": "text", + "content": "In contrast, CoMapGS begins with enhanced initial point clouds, preserving the scene's original structure while benefiting from proximity loss supervision, resulting in improved texture representation over conventional methods, as shown in Fig. 6. Quantitative results in Table 2 confirm significant improvements across metrics, particularly LPIPS, compared to recent state-of-the-art methods, which perform worse than their baselines (e.g., 3DGS [12], FSGS [41]). In the 24-view case, while PSNR improvements may appear modest due to unseen regions in test views, the LPIPS gain highlights our method's effectiveness in recovering textures from both low- and high-uncertainty regions in novel view." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 345, + 149, + 357 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 345, + 149, + 357 + ], + "spans": [ + { + "bbox": [ + 55, + 345, + 149, + 357 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 362, + 295, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 362, + 295, + 458 + ], + "spans": [ + { + "bbox": [ + 55, + 362, + 295, + 458 + ], + "type": "text", + "content": "The ablation study in Table 3 highlights that dense reconstruction in low-uncertainty (covisible) regions already provides significant improvement over baseline methods, underscoring the benefit of increased point density in covisible areas. Furthermore, adding points from mono-view regions to the enhanced point cloud yields additional gains, demonstrating the value of incorporating points from underrepresented, high-uncertainty regions in sparse view synthesis." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 460, + 295, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 460, + 295, + 626 + ], + "spans": [ + { + "bbox": [ + 55, + 460, + 295, + 626 + ], + "type": "text", + "content": "We also evaluated the impact of our proximity loss-based weighted supervision both with and without the initial point cloud enhancement strategy. Interestingly, starting from sparse COLMAP point clouds without updates showed limited improvement, likely because 3DGS-based methods tend to densify Gaussians near their original positions. However, when initialized with our enhanced point clouds, the benefits of weighted supervision became evident, delivering significant performance gains in both partial (multiview region) and full (multi- and mono-view region) update scenarios. This demonstrates that enhancing initial point clouds and incorporating proximity loss-based weighted supervision work synergistically to optimize 3DGS-based sparse view synthesis." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 635, + 128, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 635, + 128, + 647 + ], + "spans": [ + { + "bbox": [ + 55, + 635, + 128, + 647 + ], + "type": "text", + "content": "4.4. Discussion" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 653, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 295, + 713 + ], + "type": "text", + "content": "Definition of sparse view synthesis. Although sparse novel view synthesis is often defined by a specific number of training views (e.g., LLFF with 3, 6, or 9 views; Mip-NeRF with 12 or 24 views), we argue that the true challenge lies in reconstructing underrepresented regions with low covisibility" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 323, + 134, + 547, + 245 + ], + "blocks": [ + { + "bbox": [ + 313, + 70, + 553, + 126 + ], + "lines": [ + { + "bbox": [ + 313, + 70, + 553, + 126 + ], + "spans": [ + { + "bbox": [ + 313, + 70, + 553, + 126 + ], + "type": "text", + "content": "Table 3. Ablation study of the proposed CoMapGS with six training views. " + }, + { + "bbox": [ + 313, + 70, + 553, + 126 + ], + "type": "inline_equation", + "content": "\\triangle" + }, + { + "bbox": [ + 313, + 70, + 553, + 126 + ], + "type": "text", + "content": " indicates enhancing initial point clouds only in the low-uncertainty region (where covisible count " + }, + { + "bbox": [ + 313, + 70, + 553, + 126 + ], + "type": "inline_equation", + "content": "\\geq 2" + }, + { + "bbox": [ + 313, + 70, + 553, + 126 + ], + "type": "text", + "content": "), representing a straightforward application of dense point clouds generated by conventional methods such as MAST3R [16]." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 323, + 134, + 547, + 245 + ], + "lines": [ + { + "bbox": [ + 323, + 134, + 547, + 245 + ], + "spans": [ + { + "bbox": [ + 323, + 134, + 547, + 245 + ], + "type": "table", + "html": "
Proposed methodLLFF
Sec. 3.2Sec. 3.3PSNR ↑SSIM ↑LPIPS ↓
CoR-GS [38]†24.7770.8440.116
✓24.7870.8450.116
△24.900.8490.112
△✓25.1530.8540.109
✓25.0760.8520.109
✓✓25.2040.8540.108
", + "image_path": "02c822666b22c4f6cb078096a6930aa92769971ed505ca88dcd07c95e85813de.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 270, + 555, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 270, + 555, + 462 + ], + "spans": [ + { + "bbox": [ + 313, + 270, + 555, + 462 + ], + "type": "text", + "content": "counts rather than the total view count. This problem is inherently region-wise and imbalanced, where certain areas appear in only a few views, leading to sparse representation even when many views are available. For example, as DyCheck [9] categorizes scene types based on effective multiview factors (EMFs), scenarios with high frame counts but low region-specific可视ibility can still be considered sparse view cases. We address this by applying region-wise adaptive supervision based on可视ibility counts, improving results, especially for imbalanced datasets like MipNeRF 360 (12 views). In such cases, central objects are densely captured, while outer areas show low可视ibility due to the 360-degree, outward-facing capture setup. This issue is common in real-world scene capture, where uneven可视ibility across views motivates balanced,可视ibility-aware supervision." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 482, + 391, + 495 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 482, + 391, + 495 + ], + "spans": [ + { + "bbox": [ + 314, + 482, + 391, + 495 + ], + "type": "text", + "content": "5. Conclusions" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 504, + 555, + 704 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 504, + 555, + 704 + ], + "spans": [ + { + "bbox": [ + 313, + 504, + 555, + 704 + ], + "type": "text", + "content": "This paper introduces CoMapGS, a Covisibility Map-based Gaussian Splitting approach designed to enhance sparse view synthesis methods by addressing region-specific uncertainty levels. We generate covisibility maps for a given non-continuous set of distant images, using them to identify low- and high-uncertainty regions and to improve point cloud initialization beyond conventional methods like COLMAP. By introducing covisibility-adaptive weighted supervision, CoMapGS effectively manages varying sparsity across both multiview and mono-view regions, consistently outperforming state-of-the-art 3DGS-based sparse novel view synthesis methods. Experimental results validate CoMapGS's effectiveness across diverse, realistic scenes, including both indoor and outdoor environments and various scales of captured areas. CoMapGS demonstrates superiority in handling complex region-wise covisibility, highlighting its robustness and adaptability in challenging sparse view synthesis tasks." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26786" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "type": "text", + "content": "[1] Jonathan T. Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P. Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In ICCV, pages 5855–5864, 2021. 1, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 136, + 296, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 136, + 296, + 191 + ], + "spans": [ + { + "bbox": [ + 61, + 136, + 296, + 191 + ], + "type": "text", + "content": "[2] Jonathan T. Barron, Ben Mildenhall, Dor Verbin, Pratul P. Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5470-5479, 2022. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 191, + 295, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 191, + 295, + 246 + ], + "spans": [ + { + "bbox": [ + 62, + 191, + 295, + 246 + ], + "type": "text", + "content": "[3] Anpei Chen, Zexiang Xu, Fuqiang Zhao, Xiaoshuai Zhang, Fanbo Xiang, Jingyi Yu, and Hao Su. Mvsnerf: Fast generalizable radiance field reconstruction from multi-view stereo. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 14124-14133, 2021. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 247, + 295, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 247, + 295, + 278 + ], + "spans": [ + { + "bbox": [ + 62, + 247, + 295, + 278 + ], + "type": "text", + "content": "[4] Shan Chen, Jiale Zhou, and Lei Li. Dense point clouds matter: Dust-gs for scene reconstruction from sparse viewpoints, 2024. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 280, + 295, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 280, + 295, + 346 + ], + "spans": [ + { + "bbox": [ + 62, + 280, + 295, + 346 + ], + "type": "text", + "content": "[5] Wenyan Cong, Hanxue Liang, Peihao Wang, Zhiwen Fan, Tianlong Chen, Mukund Varma, Yi Wang, and Zhangyang Wang. Enhancing nerf akin to enhancing llms: Generalizable nerf transformer with mixture-of-view-experts. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 3193-3204, 2023. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 346, + 295, + 401 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 346, + 295, + 401 + ], + "spans": [ + { + "bbox": [ + 62, + 346, + 295, + 401 + ], + "type": "text", + "content": "[6] Kangle Deng, Andrew Liu, Jun-Yan Zhu, and Deva Ramanan. Depth-supervised nerf: Fewer views and faster training for free. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12882-12891, 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 402, + 295, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 402, + 295, + 456 + ], + "spans": [ + { + "bbox": [ + 62, + 402, + 295, + 456 + ], + "type": "text", + "content": "[7] Zhiwen Fan, Wenyan Cong, Kairun Wen, Kevin Wang, Jian Zhang, Xinghao Ding, Danfei Xu, Boris Ivanovic, Marco Pavone, Georgios Pavlakos, Zhangyang Wang, and Yue Wang. Instantsplat: Unbounded sparse-view pose-free gaussian splatting in 40 seconds, 2024. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 458, + 295, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 458, + 295, + 491 + ], + "spans": [ + { + "bbox": [ + 62, + 458, + 295, + 491 + ], + "type": "text", + "content": "[8] Guangchi Fang and Bing Wang. Mini-Splatting: Representing scenes with a constrained number of gaussians. In ECCV, Milano, Italy, 2024. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 491, + 295, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 491, + 295, + 524 + ], + "spans": [ + { + "bbox": [ + 62, + 491, + 295, + 524 + ], + "type": "text", + "content": "[9] Hang Gao, Ruilong Li, Shubham Tulsiani, Bryan Russell, and Angjoo Kanazawa. Monocular dynamic view synthesis: A reality check. In NeurIPS, 2022. 2, 3, 4, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 525, + 295, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 525, + 295, + 590 + ], + "spans": [ + { + "bbox": [ + 57, + 525, + 295, + 590 + ], + "type": "text", + "content": "[10] Mu Hu, Wei Yin, Chi Zhang, Zhipeng Cai, Xiaoxiao Long, Hao Chen, Kaixuan Wang, Gang Yu, Chunhua Shen, and Shaojie Shen. Metric3d v2: A versatile monocular geometric foundation model for zero-shot metric depth and surface normal estimation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(12):10579-10596, 2024. 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 591, + 295, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 591, + 295, + 635 + ], + "spans": [ + { + "bbox": [ + 57, + 591, + 295, + 635 + ], + "type": "text", + "content": "[11] Ajay Jain, Matthew Tancik, and Pieter Abbeel. Putting nerf on a diet: Semantically consistent few-shot view synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 5885-5894, 2021. 2, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 636, + 295, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 636, + 295, + 679 + ], + "spans": [ + { + "bbox": [ + 57, + 636, + 295, + 679 + ], + "type": "text", + "content": "[12] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42 (4), 2023. 1, 6, 7, 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 681, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 681, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 681, + 295, + 713 + ], + "type": "text", + "content": "[13] Mijeong Kim, Seonguk Seo, and Bohyung Han. Infonerf: Ray entropy minimization for few-shot neural volume rendering. In Proceedings of the IEEE/CVF Conference on" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "type": "text", + "content": "Computer Vision and Pattern Recognition (CVPR), pages 12912-12921, 2022. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 95, + 553, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 95, + 553, + 139 + ], + "spans": [ + { + "bbox": [ + 316, + 95, + 553, + 139 + ], + "type": "text", + "content": "[14] Jonás Kulhánek, Erik Derner, Torsten Sattler, and Robert Babuška. Viewformer: Nerf-free neural rendering from few images using transformers. In European Conference on Computer Vision (ECCV), 2022. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 140, + 553, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 140, + 553, + 183 + ], + "spans": [ + { + "bbox": [ + 317, + 140, + 553, + 183 + ], + "type": "text", + "content": "[15] Min-Seop Kwak, Jiuhn Song, and Seungryong Kim. Geconerf: few-shot neural radiance fields via geometric consistency. In Proceedings of the 40th International Conference on Machine Learning. JMLR.org, 2023. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 184, + 553, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 184, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 316, + 184, + 553, + 205 + ], + "type": "text", + "content": "[16] Vincent Leroy, Yohann Cabon, and Jerome Revaud. Grounding image matching in 3d with mast3r, 2024. 2, 3, 4, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 206, + 553, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 206, + 553, + 271 + ], + "spans": [ + { + "bbox": [ + 317, + 206, + 553, + 271 + ], + "type": "text", + "content": "[17] Jiahe Li, Jiawei Zhang, Xiao Bai, Jin Zheng, Xin Ning, Jun Zhou, and Lin Gu. Dngaussian: Optimizing sparse-view 3d gaussian radiance fields with global-local depth normalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20775-20785, 2024. 2, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 272, + 553, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 272, + 553, + 304 + ], + "spans": [ + { + "bbox": [ + 316, + 272, + 553, + 304 + ], + "type": "text", + "content": "[18] David G. Lowe. Distinctive image features from scale-invariant keypoints. Int. J. Comput. Vision, 60(2):91-110, 2004. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 304, + 553, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 304, + 553, + 360 + ], + "spans": [ + { + "bbox": [ + 316, + 304, + 553, + 360 + ], + "type": "text", + "content": "[19] Ben Mildenhall, Pratul P. Srinivasan, Rodrigo Ortiz-Cayon, Nima Khademi Kalantari, Ravi Ramamoorthi, Ren Ng, and Abhishek Kar. Local light field fusion: Practical view synthesis with prescriptive sampling guidelines. ACM Transactions on Graphics (TOG), 2019. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 361, + 553, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 361, + 553, + 404 + ], + "spans": [ + { + "bbox": [ + 316, + 361, + 553, + 404 + ], + "type": "text", + "content": "[20] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 404, + 553, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 404, + 553, + 459 + ], + "spans": [ + { + "bbox": [ + 317, + 404, + 553, + 459 + ], + "type": "text", + "content": "[21] Michael Niemeyer, Jonathan T. Barron, Ben Mildenhall, Mehdi S. M. Sajjadi, Andreas Geiger, and Noha Radwan. Regnerf: Regularizing neural radiance fields for view synthesis from sparse inputs. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 460, + 553, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 460, + 553, + 504 + ], + "spans": [ + { + "bbox": [ + 317, + 460, + 553, + 504 + ], + "type": "text", + "content": "[22] Avinash Paliwal, Wei Ye, Jinhui Xiong, Dmytro Kotovenko, Rakesh Ranjan, Vikas Chandra, and Nima Khademi Kalantari. CoherentGS: Sparse novel view synthesis with coherent 3d gaussians. In ECCV, Milano, Italy, 2024. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 504, + 553, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 504, + 553, + 568 + ], + "spans": [ + { + "bbox": [ + 317, + 504, + 553, + 568 + ], + "type": "text", + "content": "[23] Barbara Roessle, Jonathan T. Barron, Ben Mildenhall, Pratul P. Srinivasan, and Matthias Nießner. Dense depth priors for neural radiance fields from sparse input views. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12892-12901, 2022. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 570, + 553, + 603 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 570, + 553, + 603 + ], + "spans": [ + { + "bbox": [ + 317, + 570, + 553, + 603 + ], + "type": "text", + "content": "[24] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 603, + 553, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 603, + 553, + 646 + ], + "spans": [ + { + "bbox": [ + 316, + 603, + 553, + 646 + ], + "type": "text", + "content": "[25] Nagabhushan Somraj, Adithyan Karanayil, and Rajiv Soundararajan. SimpleNeRF: Regularizing sparse input neural radiance fields with simpler solutions. In SIGGRAPH Asia, 2023. 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 647, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 712 + ], + "type": "text", + "content": "[26] Jiuhn Song, Seonghoon Park, Honggyu An, Seokju Cho, Min-Seop Kwak, Sungjin Cho, and Seungryong Kim. Därf: Boosting radiance fields from sparse input views with monocular depth adaptation. In Advances in Neural Information Processing Systems, pages 68458-68470. Curran Associates, Inc., 2023. 2" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26787" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "type": "text", + "content": "[27] Thomas Tanay and Matteo Maggioni. Global latent neural rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19723-19733, 2024. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 173 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 173 + ], + "type": "text", + "content": "[28] Guangcong Wang, Zhaoxi Chen, Chen Change Loy, and Ziwei Liu. Sparsenerf: Distilling depth ranking for few-shot novel view synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 9065-9076, 2023. 2, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 175, + 294, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 294, + 207 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 294, + 207 + ], + "type": "text", + "content": "[29] Xulong Wang, Siyan Dong, Youyi Zheng, and Yanchao Yang. InfoNorm: Mutual information shaping of normals for sparse-view reconstruction. In ECCV, Milano, Italy, 2024. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 209, + 294, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 209, + 294, + 284 + ], + "spans": [ + { + "bbox": [ + 56, + 209, + 294, + 284 + ], + "type": "text", + "content": "[30] Rundi Wu, Ben Mildenhall, Philipp Henzler, Keunhong Park, Ruiqi Gao, Daniel Watson, Pratul P. Srinivasan, Dor Verbin, Jonathan T. Barron, Ben Poole, and Aleksander Ho?y?ski. Reconfusion: 3d reconstruction with diffusion priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 21551-21561, 2024. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 286, + 294, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 286, + 294, + 341 + ], + "spans": [ + { + "bbox": [ + 56, + 286, + 294, + 341 + ], + "type": "text", + "content": "[31] Jamie Wynn and Daniyar Turmukhambetov. Diffusionerf: Regularizing neural radiance fields with denoising diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4180-4189, 2023. 2, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 342, + 294, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 342, + 294, + 385 + ], + "spans": [ + { + "bbox": [ + 56, + 342, + 294, + 385 + ], + "type": "text", + "content": "[32] Haolin Xiong, Sairisheek Muttukuru, Rishi Upadhyay, Pradyumna Chari, and Achuta Kadambi. Sparsegs: Realtime " + }, + { + "bbox": [ + 56, + 342, + 294, + 385 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 56, + 342, + 294, + 385 + ], + "type": "text", + "content": " sparse view synthesis using gaussian splatting. Arxiv, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 388, + 294, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 388, + 294, + 431 + ], + "spans": [ + { + "bbox": [ + 56, + 388, + 294, + 431 + ], + "type": "text", + "content": "[33] Qingshan Xu, Xuanyu Yi, Jianyao Xu, Wenbing Tao, Yew Soon Ong, and Hanwang Zhang. Few-shot nerf by adaptive rendering loss regularization. In ECCV, Milano, Italy, 2024. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 434, + 294, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 434, + 294, + 476 + ], + "spans": [ + { + "bbox": [ + 56, + 434, + 294, + 476 + ], + "type": "text", + "content": "[34] Wangze Xu, Huachen Gao, Shihe Shen, Rui Peng, Jianbo Jiao, and Ronggang Wang. Mvpgs: Excavating multi-view priors for gaussian splatting from sparse input views. In ECCV, Milano, Italy, 2024. 2, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 478, + 294, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 478, + 294, + 533 + ], + "spans": [ + { + "bbox": [ + 56, + 478, + 294, + 533 + ], + "type": "text", + "content": "[35] Chen Yang, Peihao Li, Zanwei Zhou, Shanxin Yuan, Bingbing Liu, Xiaokang Yang, Weichao Qiu, and Wei Shen. NeRFVS: Neural Radiance Fields for Free View Synthesis via Geometry Scaffolds. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 534, + 294, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 534, + 294, + 578 + ], + "spans": [ + { + "bbox": [ + 56, + 534, + 294, + 578 + ], + "type": "text", + "content": "[36] Jiawei Yang, Marco Pavone, and Yue Wang. Freenerf: Improving few-shot neural rendering with free frequency regularization. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2023. 2, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 579, + 294, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 579, + 294, + 633 + ], + "spans": [ + { + "bbox": [ + 56, + 579, + 294, + 633 + ], + "type": "text", + "content": "[37] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelnerf: Neural radiance fields from one or few images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4578-4587, 2021. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 635, + 294, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 294, + 678 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 294, + 678 + ], + "type": "text", + "content": "[38] Jiawei Zhang, Jiahe Li, Xiaohan Yu, Lei Huang, Lin Gu, Jin Zheng, and Xiao Bai. Cor-gs: Sparse-view 3d gaussian splatt-ting via co-regularization. In ECCV, Milano, Italy, 2024. 1, 2, 3, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 680, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 680, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 680, + 294, + 713 + ], + "type": "text", + "content": "[39] Zhizhuo Zhou and Shubham Tulsiani. Sparsefusion: Distilling view-conditioned diffusion for 3d reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vi" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 183 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "type": "text", + "content": "sion and Pattern Recognition (CVPR), pages 12588-12597, 2023. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "type": "text", + "content": "[40] Haidong Zhu, Tianyu Ding, Tianyi Chen, Ilya Zharkov, Ram Nevatia, and Luming Liang. CaesarNeRF: Calibrated semantic representation for few-shot generalizable neural rendering. In ECCV, Milano, Italy, 2024. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 141, + 553, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 553, + 183 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 553, + 183 + ], + "type": "text", + "content": "[41] Zehao Zhu, Zhiwen Fan, Yifan Jiang, and Zhangyang Wang. Fsgs: Real-time few-shot view synthesis using gaussian splatting. In ECCV, Milano, Italy, 2024. 1, 2, 3, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "26788" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/CoMatcher_ Multi-View Collaborative Feature Matching/b3062684-7794-47f2-b9d1-53d25767d94f_content_list.json b/2025/CoMatcher_ Multi-View Collaborative Feature Matching/b3062684-7794-47f2-b9d1-53d25767d94f_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..84b12a77b014cffd0568c780b9641807d4cc82e5 --- /dev/null +++ b/2025/CoMatcher_ Multi-View Collaborative Feature Matching/b3062684-7794-47f2-b9d1-53d25767d94f_content_list.json @@ -0,0 +1,1490 @@ +[ + { + "type": "text", + "text": "CoMatcher: Multi-View Collaborative Feature Matching", + "text_level": 1, + "bbox": [ + 210, + 130, + 787, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jintao Zhang $^{1}$ , Zimin Xia $^{2*}$ , Mingyue Dong $^{1}$ , Shuhan Shen $^{3}$ , Linwei Yue $^{4}$ , Xianwei Zheng $^{1*}$ $^{1}$ The State Key Lab. LIESMARS, Wuhan University \n $^{2}$ École Polytechnique Fédérale de Lausanne (EPFL) \n $^{3}$ Chinese Academy of Sciences ${ }^{4}$ China University of Geosciences", + "bbox": [ + 135, + 179, + 857, + 253 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 287, + 326, + 303 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "This paper proposes a multi-view collaborative matching strategy for reliable track construction in complex scenarios. We observe that the pairwise matching paradigms applied to image set matching often result in ambiguous estimation when the selected independent pairs exhibit significant occlusions or extreme viewpoint changes. This challenge primarily stems from the inherent uncertainty in interpreting intricate 3D structures based on limited two-view observations, as the 3D-to-2D projection leads to significant information loss. To address this, we introduce CoMatcher, a deep multi-view matcher to (i) leverage complementary context cues from different views to form a holistic 3D scene understanding and (ii) utilize cross-view projection consistency to infer a reliable global solution. Building on CoMatcher, we develop a groupwise framework that fully exploits cross-view relationships for large-scale matching tasks. Extensive experiments on various complex scenarios demonstrate the superiority of our method over the mainstream two-view matching paradigm.", + "bbox": [ + 89, + 319, + 483, + 608 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 637, + 220, + 652 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Research on recovering 3D structures and camera poses from multiple views storing scene information has a long history [20, 23, 34]. A core component of this process is feature matching, which aims to estimate point correspondences across the image set [25, 32, 58]. For computational flexibility, existing methods often decompose the set into pairs of co-visible images [1, 18] and apply a two-view matching approach to each pair independently [31, 43]. The pairwise matches are then merged into comprehensive multi-view tracks [1, 45]. Following this pairwise framework to build systems for localization [22, 42] and mapping [29, 45] has become a entrenched research paradigm.", + "bbox": [ + 89, + 662, + 482, + 844 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent advances in learning-based two-view matchers have greatly improved performance in most cases [15, 30,", + "bbox": [ + 89, + 845, + 482, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e18239df716da0f8358979bd3171e64aed9fd5b1e52c6cccbddd3a248a20e2a0.jpg", + "image_caption": [ + "Figure 1. From two view to multi-view. Unlike pairwise schemes prone to uncertainty, we first partition the image set into co-visible groups and collaboratively estimate group-wide correspondences using a reliable deep multi-view matcher. Benefiting from holistic scene understanding and consistency constraint, our method generates high-quality tracks in complex scenarios." + ], + "image_footnote": [], + "bbox": [ + 519, + 287, + 901, + 501 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "43, 50]. Inspired by Transformer [54], these methods employ deep networks to jointly analyze global context from both views. This enables models to learn to infer underlying 3D scenes by integrating spatial and visual cues, which has been proven crucial for accurate estimations [27, 43].", + "bbox": [ + 511, + 626, + 906, + 702 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite great progress, current two-view matchers still struggle in challenging wide-baseline scenarios, particularly with severe occlusions and repetitive textures [25]. We argue that one potential cause for this difficulty lies in the inherent uncertainty of interpreting complex 3D structures from limited two-view observations—a task that even humans find prone to perceptual ambiguity. Projecting complex 3D geometries onto 2D planes inevitably results in significant loss of scene information. For example, spatially distant regions may appear close in 2D. This makes it highly unreliable to infer the original scene solely from ambiguous two-view context [6, 57]. Moreover, while matchers can learn two-view geometric priors, this alone is also in", + "bbox": [ + 511, + 704, + 908, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author.", + "bbox": [ + 109, + 887, + 235, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "21970", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "adequate—capturing abrupt depth discontinuities requires stronger constraints [4, 27]. Errors in pairwise correspondences compound during merging, significantly impacting downstream tasks, especially in large-scale datasets.", + "bbox": [ + 89, + 90, + 480, + 151 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address these challenges, we propose a shift in focus: instead of optimizing two-view matchers, a more effective strategy is to directly leverage the rich relationships inherent in raw multi-view observations. Building on this, we introduce a unified formulation that collaboratively establishes correspondences between a complementary view group and a target view. This approach fosters holistic scene understanding by integrating contextual information from multiple views, while multi-view geometric priors provide stronger constraints for reliable reasoning. Moreover, matches across views naturally adhere to cross-view projection consistency. This facilitates a mutual verification process, ensuring confident inference of a global solution.", + "bbox": [ + 89, + 152, + 480, + 347 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Inspired by existing two-view methods [43, 50], we develop CoMatcher, a deep-learning architecture for multiview feature matching. Context sharing is achieved through a GNN with multi-view receptive fields. While a larger feature space enriches context, it also introduces noise, complicating learning. To address this, we propose to constrain the search space for each point by leveraging cross-view projection geometry. Furthermore, we redesign the feature correlation layers in cross-attention to enable consistent multiview reasoning. This design exploits the property that correlation scores encode both reasoning outcomes and confidence, enabling progressive layer-by-layer integration of multi-view estimates into a globally consistent solution.", + "bbox": [ + 89, + 348, + 482, + 542 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To scale CoMatcher for large-scale tasks like SfM [36, 45], we develop a groupwise matching pipeline (see Fig. 1). It introduces a novel set partitioning algorithm that uses可视ability information to group images, followed by group-level matching using CoMatcher. By preserving cross-view relationships, our framework achieves comprehensive improvements over pairwise methods in extensive experiments on wide-baseline scenarios. The main contributions are summarized as follows: (i) A novel multi-view matching paradigm, comprising a deep collaborative matcher and a scalable groupwise pipeline, designed to overcome the inherent uncertainty of pairwise approaches. (ii) A multiview representation learning network architecture that enables holistic scene understanding. (iii) A multi-view feature correlation strategy that integrates uncertain individual estimates into a globally consistent solution.", + "bbox": [ + 89, + 544, + 482, + 785 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 89, + 799, + 227, + 814 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Two-view matching is a fundamental problem in computer vision. The most prevalent paradigm involves sparsifying images into keypoints, each represented by a high-dimensional vector encoding its visual context [25, 32]. This transforms matching into a search problem, with match", + "bbox": [ + 89, + 824, + 480, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "likelihood expressed by the inner product of corresponding vectors [31]. Modern methods leverage deep networks for representation learning to extract feature points [10, 13, 14, 39, 53, 59], outperforming traditional hand-crafted encodings [5, 31, 41] and achieving significant advancements on mainstream benchmarks [3, 25, 46]. In parallel, other works aim to address the limitations of heuristic matching strategies by leveraging networks with global modeling capabilities to directly learn the matching process [7, 24, 30, 43]. Such data-driven methods can effectively capture priors of the underlying 3D scene, which are essential for accurate estimation. Unlike keypoint-based approaches, another typical method retains all pixels for dense inference [8, 15, 27, 50, 52]. While this has led to notable progress in weakly textured regions, such methods require greater computational resources [30, 56] and struggle with multi-view inconsistency [21]. In summary, while these methods perform well in most scenarios, they remain constrained in complex wide-baseline settings due to the limited receptive field inherent to two-view systems.", + "bbox": [ + 511, + 90, + 903, + 392 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Multi-view matching focuses on establishing correspondences across a larger-scale image collection. A typical pairwise matching approach decomposes the problem into two-view matching instances and subsequently merges the pairwise results using multi-view projection consistency [1, 42, 44, 45]. During merging, some works explore how consistency assumptions can filter and refine coarse matches [33, 51], while others focus on efficiency, developing better strategies to reduce redundant matching computations [35, 60]. However, as a post-processing step, merging cannot address catastrophic failures that occur during the matching stage. Another line of work treats matching as a point tracking problem [11, 12, 19, 26], using optical flow-like frameworks to reason about the positions of query pixels across multiple images. While this approach is well-suited for video inputs, its effectiveness on unstructured, wide-baseline image collections remains largely unverified. The most relevant method to ours is [40], which performs end-to-end multi-view matching and pose estimation. However, it is limited in scalability, as it can only handle a small image group. In contrast, our framework is highly flexible, with the potential to scale to an arbitrary number of images.", + "bbox": [ + 511, + 393, + 903, + 724 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 511, + 739, + 648, + 756 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "**Formulation.** Given a set of $N_I$ images, $\\mathcal{I} = \\{I_i \\mid i = 1, \\dots, N_I\\}$ , of a scene, the objective of feature matching is to identify the 2D position coordinates $\\mathcal{M}_k = \\{\\mathbf{p}^{I_i}\\}$ in $\\mathcal{I}$ that correspond to identical 3D scene points $\\mathcal{X} = \\{\\mathbf{x}_k \\mid k = 1, \\dots, N_X\\}$ , where $\\mathbf{p}^{I_i} \\in \\mathbb{R}^2$ are the pixel coordinates in $I_i$ , and each $\\mathcal{M}_k$ is called a track.", + "bbox": [ + 511, + 763, + 903, + 854 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Motivation. Constructing complete tracks requires estimating correspondences between each image $I_{t} \\subset \\mathcal{I}$ and its co-visible image set $\\mathcal{C}_t$ . While the pairwise approach of", + "bbox": [ + 511, + 854, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "21971", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/b8d28b5cf08c3d45e508810fe0943624b78ba3655e92c41e14199624d5892134.jpg", + "image_caption": [ + "Figure 2. The CoMatcher architecture can be viewed as an extension of parallel two-view matchers, enhanced with two core components: the multi-view (M-V) feature interaction module (Sec. 3.2) and the multi-view (M-V) feature correlation strategy (Sec. 3.3). The first component employs multi-view cross-attention to enhance the representation of all source and target view features, while leveraging the cross-view projection geometry from $\\mathcal{M}(\\mathcal{G})$ to constrain the point search space. The second component aggregates correlation scores from multiple two-view cross-attention modules based on estimated confidence, guiding the network to achieve multi-view consistent reasoning." + ], + "image_footnote": [], + "bbox": [ + 114, + 88, + 890, + 303 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "fers implementation flexibility for this problem, its reliance on two-view inputs in a single estimation often leads to uncertainty in complex scenes. In practice, images within $\\mathcal{C}_t$ are often not independent—most are captured from adjacent positions but with differing viewing angles. This inspires us to group $\\mathcal{C}_t$ based on image correlations, where each complementary group offers a mapping from multiple planar views to a 3D sub-scene. In the image matching problem, the correspondence between $I_t$ and a group must adhere to certain 2D-to-3D physical constraints, such as cross-view projection consistency. Rather than independently estimating and then integrating them with constraints, it is more efficient to unify this process and learn underlying priors directly from the data. Thus, we formulate CoMatcher to address the 1-to- $N$ matching problem (see Fig. 2).", + "bbox": [ + 88, + 411, + 485, + 638 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Groupwise matching pipeline. Based on the above, we propose a three-step pipeline for constructing tracks: grouping, connecting, and matching. First, we treat the original set of views $\\mathcal{I}$ as a union of smaller image groups $\\{\\mathcal{G}_s\\mid s = 1,\\dots ,N_G\\}$ , each representing a localized scene. This involves a pre-processing step to partition the set. The most critical aspect of this is determining the correlation between two images, i.e., whether they describe the same structure. Building on existing research [2, 38, 48], we employ heuristic techniques to design a grouping algorithm. Next, although CoMatcher can directly learn multi-view geometry, certain fine-grained physical rules are challenging to model. Therefore, we explicitly establish relationships within each group to guide subsequent matching. For each group, we construct tracks using existing frameworks, which provide a set of 3D points along with their multi-view projections and implicitly encode camera poses. These cues", + "bbox": [ + 89, + 643, + 485, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "are incorporated as inputs to the network to steer the inference process. Finally, each group of images is treated as a whole to collaboratively match with other images in $\\mathcal{I}$ using CoMatcher. Additional details on grouping and the overall framework are provided in the supplementary materials.", + "bbox": [ + 511, + 411, + 908, + 489 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. CoMatcher overview", + "text_level": 1, + "bbox": [ + 511, + 497, + 712, + 512 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As shown in Fig. 2, CoMatcher, as a deep sparse matcher, estimates point correspondences between each source view in a group $I_{i} \\in \\mathcal{G}$ and a target view $I_{t}$ . The network takes as input the extracted local features $\\{\\left(\\mathbf{p}_{k}^{I_{i}}, \\mathbf{d}_{k}^{I_{i}}\\right) \\mid k = 1, \\dots, N_{F}\\}$ from the $M$ source views and the target view, along with the precomputed tracks $\\mathcal{M}(\\mathcal{G})$ of $\\mathcal{G}$ . Each local feature consists of keypoint locations $\\mathbf{p}_{k}$ and their corresponding visual descriptors $\\mathbf{d}_{k}$ . Initially, the target view features are broadcast and paired with those of each source view to establish the initial states. These $M$ feature pairs are then refined using a Graph Neural Network (GNN) with alternating self-attention and cross-attention modules, repeated $L$ times. Finally, the enhanced features are fed into matching heads to infer the final correspondences.", + "bbox": [ + 511, + 518, + 908, + 731 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The design of CoMatcher revolves around two core questions: (i) how can multi-view context be leveraged to learn better point representations (Sec. 3.2), and (ii) how can the estimations across multiple views be constrained to satisfy cross-view consistency (Sec. 3.3)? After thoroughly discussing these questions, we introduce the prediction heads (Sec. 3.4) and the loss function (Sec. 3.5).", + "bbox": [ + 511, + 732, + 908, + 838 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Multi-view feature interaction", + "text_level": 1, + "bbox": [ + 511, + 848, + 782, + 862 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Consider an occluded scene. When observing such structures from a frontal view, points near occlusion boundaries", + "bbox": [ + 511, + 869, + 906, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "21972", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "are often contaminated by irrelevant context. This noise significantly compromises the quality of point features, creating substantial challenges for matching. Multi-view learning provides a promising solution to this issue. For these ambiguous point features, it can integrate observations of the same area from multiple other viewpoints to refine their representations. This involves a querying process, which establishes a multi-view receptive field for each point.", + "bbox": [ + 89, + 90, + 480, + 210 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given a set of co-visible source views, CoMatcher leverages a multi-view cross-attention mechanism to aggregate these context (Source Cross). For a point $u$ in source view $I_{i}$ , we attend to all points in the other source views:", + "bbox": [ + 89, + 213, + 480, + 270 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {m} _ {u} ^ {I _ {i} \\leftarrow \\mathcal {W} _ {j}} = \\frac {1}{M - 1} \\sum_ {I _ {j} \\in \\mathcal {G} \\backslash I _ {i}} \\sum_ {v \\in \\mathcal {W} _ {j}} \\operatorname {S o f t m a x} \\left(a _ {u v} ^ {I _ {i} I _ {j}}\\right) \\mathbf {v} _ {v} ^ {I _ {j}}. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 104, + 285, + 482, + 321 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here, $\\mathcal{W}_j$ denotes the point set in the source view $I_{j}$ , while $a_{uv}^{I_iI_j}$ represents the similarity score computed by treating point $u$ in $I_{i}$ as the query and point $v$ in $I_{j}$ as the key. $\\mathbf{v}_v^{I_j}$ is the value of point $v$ . $\\mathbf{m}_u^{I_i}$ represents the message vector from $\\mathcal{W}_j$ to $u$ of the GNN [43]. We uniformly aggregate features from each source view, encouraging the model to integrate multi-view information comprehensively rather than overemphasizing images from similar perspectives.", + "bbox": [ + 89, + 335, + 482, + 462 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "However, for a point near an occlusion boundary in one source view, how does it identify the corresponding points observing the same location in other source views? This implies a matching process. While the network could directly learn this, it would introduce additional complexity, as establishing correspondences between source views is not our primary objective. Moreover, while multi-view setups provide abundant contextual information, much of it is irrelevant, such as regions lacking co-visibility. This redundancy also imposes an additional burden on learning.", + "bbox": [ + 89, + 464, + 482, + 614 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To address this, we propose a geometric constraint mechanism to explicitly guide the attention range of each point. By incorporating relative positional encoding into the attention score computation, we enable the scores to depend on two critical factors: feature correlation and the geometric distance $\\Delta \\mathbf{p}$ between two points:", + "bbox": [ + 89, + 617, + 482, + 708 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\na _ {u v} ^ {I _ {i} I _ {j}} = \\left(\\mathbf {q} _ {u} ^ {I _ {i}}\\right) ^ {\\top} \\mathbf {R} \\left[ \\Delta \\mathbf {p} _ {u v} ^ {I _ {i} I _ {j}} \\right] \\mathbf {k} _ {v} ^ {I _ {j}}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 176, + 723, + 482, + 746 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here, $\\mathbf{R}[\\cdot ]$ is the rotation encoding matrix [49].", + "bbox": [ + 89, + 758, + 406, + 775 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "While the 2D relative distance between two points within a single image is well-defined, across different views, it depends on the projection transformation. We use the precomputed tracks $\\mathcal{M}(\\mathcal{G})$ to find the projection position of $\\mathbf{p}_u^{I_i}$ on $I_{j}$ , which can be denoted as $\\mathbf{p}_w^{I_j}$ . When no corresponding point in $I_{j}$ can be found for $\\mathbf{p}_u^{I_i}$ , we simply assign the relative position to $\\mathbf{0}$ so that the attention score depends solely on the feature similarity. The calculation of the relative po", + "bbox": [ + 89, + 777, + 482, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "sition can be formulated as follows:", + "bbox": [ + 513, + 90, + 751, + 104 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\Delta \\mathbf {p} _ {u v} ^ {I _ {i} I _ {j}} = \\left\\{ \\begin{array}{l l} \\mathbf {p} _ {w} ^ {I _ {j}} - \\mathbf {p} _ {v} ^ {I _ {j}} & (u, w) \\in M (I _ {i}, I _ {j}) \\\\ \\mathbf {0} & (u, \\emptyset) \\in M (I _ {i}, I _ {j}). \\end{array} \\right. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 560, + 114, + 903, + 157 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Additionally, we also interact with all target view features after source cross-attention (Target Cross). For a point $t$ in view $I_{t}$ paired with source view $I_{i}$ , we attend to the same point $t$ across the remaining $M - 1$ target views. This is analogous to operations in some works that aggregate features along the temporal dimension [12, 26]. Unlike the approach for source views, we compute the attention scores solely based on feature similarity. Despite its low computational complexity, this module complements the Source Cross, mutually enhancing each other to help the network implicitly infer globally optimal correspondences.", + "bbox": [ + 511, + 167, + 906, + 333 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Multi-view feature correlation", + "text_level": 1, + "bbox": [ + 513, + 344, + 782, + 358 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For a track in a source view set originating from a 3D point, its corresponding 2D keypoints $\\left\\{\\mathbf{p}_{u}^{I_{i}}, \\mathbf{p}_{v}^{I_{j}}, \\ldots\\right\\}$ should exhibit consistent matches in the target view—either corresponding to the same single point or having no corresponding point at all. This physical constraint is often utilized by existing methods as a post-processing step to filter out erroneous results when merging multiple two-view correspondences. However, as a multi-view matcher, we aim for CoMatcher to inherently satisfy this constraint in its reasoning, thereby enhancing its reliability and confidence.", + "bbox": [ + 511, + 366, + 906, + 523 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Achieving this involves two key processes: (i) identifying points that may lead to errors in the early layers of the network, and (ii) promptly guiding these points toward correct estimation using inference information from other views. To accomplish this, we propose a two-step multiview feature correlation strategy (see Fig. 3).", + "bbox": [ + 511, + 523, + 905, + 614 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "First, to identify ambiguous points in the source views, we utilize a lightweight head at the end of each layer to predict the confidence of each point:", + "bbox": [ + 511, + 614, + 905, + 660 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nc _ {u} ^ {I _ {i}} = \\operatorname {S i g m o i d} \\left(\\operatorname {M L P} \\left(\\mathbf {f} _ {u} ^ {I _ {i}}\\right)\\right) \\in [ 0, 1 ]. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 583, + 671, + 903, + 691 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here, MLP represents a multi-layer perceptron and $\\mathbf{f}_u^{I_i}$ is the feature of $u$ . This confidence score reflects the assessment of each point in its current state: higher values indicate that, at this layer, the point is either converging toward a reliable match or confidently identified as having no match.", + "bbox": [ + 511, + 703, + 905, + 779 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In each layer, we define a distinct hyperparameter threshold $\\theta$ . Points with confidence scores below this threshold are identified as potentially ambiguous and likely to lead to incorrect inference. We progressively increase the value of $\\theta$ in later layers, as the confidence of overall inference generally improves with each successive layer.", + "bbox": [ + 511, + 779, + 905, + 869 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Next, we redesign the parallel two-view cross-attention modules to correct the erroneous estimations of ambiguous", + "bbox": [ + 511, + 869, + 905, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "21973", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/96f4098fe3dbd88256aec1c84d929a9546d018e9a5100a1ccf9d62bdd6564e0f.jpg", + "image_caption": [ + "Figure 3. The multi-view feature correlation strategy consists of two steps: we first identify potentially erroneous points using a confidence estimator, and then correct their estimation by aggregating attention correlation vectors from other views." + ], + "image_footnote": [], + "bbox": [ + 114, + 87, + 472, + 210 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "points. In these modules, each point in the source view computes an attention distribution vector to query [54]:", + "bbox": [ + 89, + 308, + 483, + 339 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\alpha} _ {u} ^ {I _ {i}} = \\underset {x \\in \\mathcal {W} _ {t}} {\\operatorname {S o f t m a x}} \\left(a _ {u x} ^ {I _ {i} I _ {t}}\\right). \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 205, + 348, + 482, + 375 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This vector represents the degree to which the point attends to in view $I_{t}$ , with a peak indicating potential matching regions of interest. For ambiguous points, this similarity is not reliable, often leading to incorrect feature aggregation. Assuming that point $u$ in the source view $I_{i}$ is an ambiguous point, we update its original attention distribution by applying a weighted average to the distributions of its corresponding points across multi-view:", + "bbox": [ + 89, + 383, + 483, + 497 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\alpha} _ {u} ^ {I _ {i} ^ {\\prime}} = c _ {u} ^ {I _ {i}} \\boldsymbol {\\alpha} _ {u} ^ {I _ {i}} + \\left(1 - c _ {u} ^ {I _ {i}}\\right) \\frac {\\sum_ {v \\in \\mathcal {D} _ {m} ^ {u}} c _ {v} ^ {I _ {j}} \\boldsymbol {\\alpha} _ {v} ^ {I _ {j}}}{\\sum_ {v \\in \\mathcal {D} _ {m} ^ {u}} c _ {v} ^ {I _ {j}}} \\quad \\text {s . t .} c _ {u} ^ {I _ {i}} < \\theta . \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 101, + 503, + 482, + 549 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here, $\\mathcal{D}_m^u$ represents other points in the track. We embed this strategy into the model inference process, which significantly enhances the reliability of estimation under challenging conditions despite its simple implementation.", + "bbox": [ + 89, + 556, + 483, + 617 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Correspondence prediction", + "text_level": 1, + "bbox": [ + 89, + 626, + 336, + 643 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As shown in Fig. 2, a lightweight matching head operates in parallel for each pair of network output features to predict correspondences between two views [30, 43]. In each head, we begin by calculating a score matrix from feature correlations between two views: $\\mathbf{S}(u,x) = \\mathrm{L}(\\mathbf{f}_u^{I_i})\\cdot$ $\\mathrm{L}(\\mathbf{f}_x^{I_t})$ , where $\\mathrm{L}$ is the linear transform. Then we apply dual-softmax [30, 50, 53] to derive the matching probabilities of two points: $\\mathbf{S}'(u,x) = \\mathrm{Softmax}(\\mathbf{S}(u,\\cdot))_x\\cdot$ Softmax $(\\mathbf{S}(\\cdot ,x))_u$ . Additionally, we employ the method proposed in [30] to calculate the matching probability for each point: $\\sigma_{i}^{I_{i}} = \\mathrm{Sigmoid}(\\mathrm{L}(\\mathbf{f}_{i}^{I_{i}}))$ . The final assignment matrix is a combination of the two probabilities:", + "bbox": [ + 89, + 648, + 483, + 830 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {P} (u, x) = \\mathbf {S} ^ {\\prime} (u, x) \\sigma_ {u} ^ {I _ {i}} \\sigma_ {x} ^ {I _ {t}}. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 196, + 840, + 482, + 859 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Using mutual nearest neighbors and a threshold, we filter the matching results from the assignment matrix [43, 50].", + "bbox": [ + 89, + 869, + 483, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Loss", + "text_level": 1, + "bbox": [ + 513, + 90, + 586, + 104 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "CoMatcher supervises each source-target image pair individually, with the loss for each pairs composed of two components: correspondence and confidence:", + "bbox": [ + 511, + 112, + 903, + 157 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {t o t a l}} = \\frac {1}{M} \\sum_ {I _ {i} \\in \\mathcal {G}} \\left(\\mathcal {L} _ {\\text {c o r r}} \\left(I _ {i}, I _ {t}\\right) + \\alpha \\mathcal {L} _ {\\text {c o n f}} \\left(I _ {i}\\right)\\right). \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 562, + 167, + 906, + 205 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Correspondence supervision. For each pair $I_{i}$ and $I_{t}$ , we perform two-view transformations using relative poses or homography to compute ground truth labels [30, 43, 50]. We minimize the negative log-likelihood of the assignment matrix, following [30]. More details are provided in the supplementary materials.", + "bbox": [ + 511, + 214, + 906, + 304 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Confidence supervision. To train the confidence estimators at each layer, we define the point confidence as the consistency probability between its correspondence estimated at the current layer and the final estimation [16, 30, 47]. At each layer, predictions are obtained via dual-softmax on current features, without introducing additional matching probabilities or parameters. The ground truth label $y$ indicates whether the two estimations are consistent. This is supervised using a binary cross-entropy loss:", + "bbox": [ + 511, + 305, + 905, + 441 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {c o n f}} \\left(I _ {i}\\right) = \\frac {1}{L - 1} \\sum_ {\\ell} \\sum_ {u \\in \\mathcal {W} _ {i}} \\mathrm {C E} \\left(\\ell c _ {u} ^ {I _ {i}}, \\ell y _ {u} ^ {I _ {i}}\\right), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 562, + 450, + 905, + 488 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\ell \\in \\{1,\\dots ,L - 1\\}$", + "bbox": [ + 513, + 498, + 692, + 513 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 513, + 527, + 645, + 544 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we first introduce the datasets used, followed by our implementation details. Then, our CoMatcher network is compared to previous state-of-the-art baselines for homography and camera pose estimation. Next, we integrate CoMatcher into groupwise matching framework and evaluate it against pairwise method on a large-scale benchmark. Finally, an extensive ablation study is provided.", + "bbox": [ + 511, + 551, + 905, + 657 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Datasets", + "text_level": 1, + "bbox": [ + 513, + 667, + 614, + 681 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The HPatches dataset is used for homography estimation [3]. It consists of 116 strictly planar scenes, each containing 6 images with variations in viewpoint and illumination. For camera pose estimation tasks, we utilize the MegaDepth dataset, an outdoor dataset that exhibits strong occlusions and significant structural changes [28]. We selected two scenes, \"Sacre Coeur\" and \"St. Peter's Square\", from which 1500 co-visible quadruplets were sampled in a way that balances difficulty based on visual overlap [30, 50, 56]. The Image Matching Challenge 2020 benchmark provides a comprehensive evaluation protocol, including datasets that cover multiple challenging outdoor wide-baseline scenes. From the phototourism dataset, we selected 3 validated scenes, each consisting of about 100", + "bbox": [ + 511, + 688, + 906, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "21974", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/e497676a2d20632109343329ef50e38b1f36b268cef778dee4164bac065095f3.jpg", + "image_caption": [ + "Figure 4. Qualitative comparison on MegaDepth. For each quadruple with a target view (top), correspondences (green for correct, red for incorrect) predicted by CoMatcher and LightGlue are shown. Using identical local features as input, CoMatcher achieves significantly more reliable estimations, even at challenging semantic edges with depth discontinuities. This stems from its holistic scene understanding through multi-view cues to effectively address occlusions (Q3), large-scale variations (Q2, 4), and repetitive textures (Q1, 2, 3, 4)." + ], + "image_footnote": [], + "bbox": [ + 112, + 88, + 883, + 520 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "images. Additionally, we conducted an ablation study on a synthetic homography dataset [37] used for training.", + "bbox": [ + 89, + 614, + 482, + 647 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Implementation details", + "text_level": 1, + "bbox": [ + 89, + 652, + 305, + 669 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "CoMatcher is trained in two steps, following [30, 43]. We first pre-train the model on a large-scale synthetic homography dataset [37], leveraging noise-free homography for ground truth. Next, fine-tuning is performed on the MegaDepth dataset [28] using the camera poses recovered by SfM as ground truth. We sample 200 co-visible multiview groups per scene and randomly select one image as the target view. The size of source views is set to $M = 4$ during training. The training process is carried out on two NVIDIA GeForce 4090 GPUs, taking about 6 days in total. More details can be found in the supplementary materials.", + "bbox": [ + 89, + 675, + 482, + 840 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Homography estimation", + "text_level": 1, + "bbox": [ + 89, + 848, + 313, + 864 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "CoMatcher is compared with three types of baselines. First, we include state-of-the-art (SOTA) sparse two-view", + "bbox": [ + 89, + 869, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "matchers: nearest-neighbor with mutual check [31], SuperGlue [43], SGMNet [7], and LightGlue [30], each paired with different feature extractors [10, 31, 53]. Second, we evaluate the multi-view matcher End2End [40]. For reference, we also compare against typical dense matchers LoFTR [50] and PDC-Net [52]. For CoMatcher and End2End, we use a single forward pass to obtain matches across all five image pairs, while others follow a pairwise approach. To ensure a fair comparison, we adopt the setups of [30, 50] for the number of features and image resizing. Homography accuracy is evaluated using both robust (RANSAC [17]) and non-robust (weighted DLT [20]) solvers. For each pair, we compute the mean reprojection error of the four image corners and report the area under the cumulative error curve (AUC) up to 1px, 3px, and 5px.", + "bbox": [ + 511, + 614, + 906, + 843 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Tab. 1 shows that CoMatcher yields more accurate estimates than all sparse two-view matchers, highlighting the advantage of inference in multi-view feature space. How-", + "bbox": [ + 511, + 854, + 906, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "21975", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/eb94d5444d3f00378b91db5c8ba92a699aebc0132eda6259692adfdef7ee3ada.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodAUC - DLT @1px / @3px / @5pxAUC - RANSAC
denseLoFTR38.5 / 66.0 / 71.440.7 / 68.3 / 78.5
PDC-Net36.0 / 65.3 / 73.037.9 / 67.6 / 77.4
sparse 2-viewSIFT+NN+mutual0.0 / 0.0 / 0.035.9 / 65.0 / 75.6
SP+NN+mutual0.0 / 1.9 / 3.434.8 / 64.1 / 74.8
SP+SuperGlue32.2 / 65.1 / 75.737.2 / 68.0 / 78.7
SP+SGMNet31.7 / 64.9 / 76.037.7 / 66.4 / 77.5
SP+LightGlue35.4 / 67.5 / 77.737.2 / 67.8 / 78.1
DISK+NN+mutual1.8 / 5.2 / 7.837.9 / 58.0 / 68.3
DISK+LightGlue34.4 / 64.5 / 74.438.1 / 65.2 / 77.2
multi-viewSP+End2End34.3 / 66.9 / 75.537.0 / 67.2 / 77.5
SP+CoMatcher37.1 / 69.0 / 78.838.4 / 68.9 / 79.0
DISK+CoMatcher36.3 / 66.2 / 75.938.7 / 68.2 / 78.4
", + "bbox": [ + 91, + 88, + 490, + 340 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ever, despite End2End also jointly reasoning over multiple views like ours, its performance is notably inferior. This underscores the superiority of a $N$ -to-1 architecture over $N$ -to- $N$ for multi-view representation learning. When compared to LoFTR, we observe slightly lower performance under low thresholds. We attribute this mainly to the limitations of the feature extractor in keypoint non-repeatability and localization errors [29]. Additionally, DLT achieves accuracy close to RANSAC on most metrics, reflecting the high-quality correspondences from CoMatcher.", + "bbox": [ + 89, + 422, + 483, + 575 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Relative pose estimation", + "text_level": 1, + "bbox": [ + 89, + 588, + 313, + 604 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Next, we evaluate CoMatcher on challenging wide-baseline scenes, using the same baselines as in Sec. 4.3. For each quadruplet, including a selected target view, we still use a single forward pass for multi-view methods. Additionally, we test two newer dense methods, DUS3R [55] and MAS3R [27], following their evaluation setups for image resizing. All other settings remain consistent with [30]. The essential matrix is estimated from the correspondences using both vanilla RANSAC [17] and LO-RANSAC [9], following [30]. The pose error is then computed as the maximum angular error of rotation and translation, derived from the decomposition of the essential matrix. We report the AUC at $5^{\\circ}$ , $10^{\\circ}$ , and $20^{\\circ}$ , and record the average runtime of matching each quadruplet on a single 4090 GPU.", + "bbox": [ + 89, + 612, + 482, + 823 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Tab. 2 shows that CoMatcher, as a sparse method, comprehensively enhances the estimation performance on different local features [10] and [53]. Importantly, compared to other sparse two-view matchers like LightGlue, a key strength of CoMatcher is its reliability, as illustrated in", + "bbox": [ + 89, + 825, + 483, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/3897fd00562f5457995ad0f9e6cadd26d3ba8400d25c4c28f661f494f071a7ff.jpg", + "table_caption": [ + "Table 1. Homography estimation on HPatches. We report the area under the cumulative error curve (AUC) up to values of 1px, 3px and 5px, using DLT and RANSAC [17] as homography solver." + ], + "table_footnote": [], + "table_body": "
MethodAUC RANSACAUC LO-RANSACTime (ms)
5° / 10° / 20°
denseLoFTR58.0 / 73.1 / 84.467.4 / 81.7 / 89.3182
PDCNet54.7 / 73.1 / 83.567.1 / 80.2 / 87.0231
DUSt3R42.4 / 56.7 / 64.258.1 / 70.4 / 79.6264
MASt3R51.5 / 65.3 / 75.663.5 / 76.3 / 85.2317
SuperPointNN+mutual35.3 / 58.3 / 53.751.4 / 67.3 / 75.99
SuperGlue55.8 / 72.8 / 84.165.1 / 77.2 / 89.287
LightGlue56.2 / 72.7 / 83.567.2 / 80.1 / 88.051
End2End55.3 / 71.4 / 81.267.4 / 81.5 / 87.0152
CoMatcher57.2 / 73.9 / 84.868.3 / 82.2 / 89.169
DISKNN+mutual50.9 / 66.7 / 77.764.0 / 79.5 / 87.69
LightGlue53.2 / 69.2 / 80.268.6 / 80.4 / 87.254
CoMatcher54.9 / 71.2 / 82.068.5 / 82.1 / 88.473
", + "bbox": [ + 521, + 88, + 898, + 339 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Relative pose estimation on MegaDepth. We report the AUC at $5^{\\circ}$ , $10^{\\circ}$ , and $20^{\\circ}$ using different robust estimator, and the average runtime of matching each quadruplet.", + "bbox": [ + 511, + 351, + 906, + 393 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Fig. 3. This highlights the advantage of multi-view collaborative reasoning in understanding complex 3D structures, such as occlusions, even with suboptimal input local features in these areas. Additionally, compared to End2End, CoMatcher achieves more accurate relative poses without the need for cumbersome end-to-end training. While some dense methods may achieve better accuracy on certain metrics, their efficiency is generally much lower. By employing a space-for-time strategy to minimize inference steps, Co-Matcher runs significantly faster than most methods.", + "bbox": [ + 511, + 422, + 906, + 575 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Evaluation on the IMC 2020 benchmark", + "text_level": 1, + "bbox": [ + 511, + 588, + 859, + 603 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Next, we evaluate our groupwise framework against pairwise baselines [30, 31, 43] on the Image Matching Challenge (IMC) 2020 benchmark [25]. Given a large-scale unordered image collection, the benchmark requires providing matching results for all image pairs, which are then assessed on two downstream tasks: stereo and multi-view reconstruction. For the stereo task, the accuracy of relative camera poses for each pair is evaluated from correspondences using RANSAC, as in Sec. 4.4. For the multi-view task, all correspondences are fed into COLMAP [45] for SfM, with the final accuracy evaluated based on the estimated multi-view camera poses. We report the AUC at $5^{\\circ}$ and $10^{\\circ}$ across both tasks. Additionally, the average runtime is reported, which is the total matching time for the entire image set divided by the number of pairs.", + "bbox": [ + 511, + 612, + 906, + 838 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Tab. 3 shows that our framework significantly outperforms existing pairwise approaches in both tasks. This advantage primarily stems from our framework's ability to better leverage the relationships within the original image", + "bbox": [ + 511, + 839, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "21976", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/f2a31f32adfe001ee1ae82b4dff0a2fd9ef41cc93c2d5ce8e6a34e7e60b8e0ba.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodStereoMulti-ViewTime (ms)
AUC 5°/10°AUC 5°/10°
two-viewSIFT+NN+mutual31.5 / 44.257.2 / 68.53
SP+NN+mutual28.6 / 40.352.9 / 63.43
SP+SuperGlue36.5 / 50.162.3 / 74.832
SP+LightGlue36.8 / 49.464.6 / 75.417
DISK+NN+mutual35.4 / 47.459.3 / 70.23
DISK+LightGlue42.1 / 55.665.2 / 76.219
m-vSP+CoMatcher39.1 / 52.466.2 / 77.525
DISK+CoMatcher44.9 / 57.367.1 / 78.429
", + "bbox": [ + 99, + 88, + 477, + 266 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "collection, enabling the generation of higher-quality tracks in such large-scale tasks. Additional comparisons on track quality for SfM tasks are provided in the supplementary material. In terms of efficiency, our approach is faster than SuperGlue, with the viewpoint grouping process accounting for only about $4\\%$ of the total matching time.", + "bbox": [ + 89, + 347, + 483, + 439 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.6. Ablation study", + "text_level": 1, + "bbox": [ + 89, + 452, + 241, + 468 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Understanding CoMatcher. The CoMatcher network introduces three key components: multi-view cross-attention modules, a geometric constraint mechanism, and a multiview feature interaction strategy. We validate these design choices by evaluating on two benchmarks: the challenging synthetic homography dataset [37], where precision and recall metrics are employed, and HPatches [3], where the AUC of homography estimation via weighted DLT is measured. For each image, 512 keypoints are extracted.", + "bbox": [ + 89, + 474, + 482, + 611 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Tab. 4 shows the ablation results. Excluding the two cross-attention modules for multi-view interaction reduces the model to a series of parallel two-view matchers [30, 43]. This simplification limits the model's ability to integrate multi-view context, leading to a significant performance decline. Without the geometric constraint mechanism, the model performs even worse than in the previous case. This suggests that direct feature aggregation from multi-view is challenging, as the network requires additional learning to distinguish meaningful context effectively. Without multiview feature correlation, the model loses its ability to guide inference for ambiguous points by leveraging consistency constraints. This highlights the importance of guiding two-view feature correlation in the attention search process, a factor overlooked in previous work.", + "bbox": [ + 89, + 612, + 482, + 838 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Impact of the group size. We evaluated the matching performance on the synthetic homography dataset [37] by varying the size of the source view sets used as network input. As shown in Fig. 5, performance initially improves signif", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5bae3bf10a053c632f3bdc281e2fbc13ae7ba2abcd30da28f2ba6e9de42ced8d.jpg", + "image_caption": [ + "Figure 5. Impact of the group size. Under different numbers of keypoints, the AUC initially increases as the source view set expands, then tends to stabilize or shows a downward trend." + ], + "image_footnote": [], + "bbox": [ + 517, + 95, + 901, + 237 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/d728c9751caacc1315905e599b0ce31cd04c9fa91fc53b5ca2e806d2e1357dd8.jpg", + "table_caption": [ + "Table 3. Evaluation on the IMC 2020 benchmark. We report the pose AUC at $5^{\\circ}$ , $10^{\\circ}$ for two subtasks, stereo and multi-view reconstruction, along with the average matching runtime." + ], + "table_footnote": [], + "table_body": "
CoMatcherSyntheticHpatches-AUC
precisionrecall@1px / @5px
w/o M-V feat. interaction89.796.632.4 / 75.1
w/o attenu. propagation87.794.231.5 / 73.7
w/o M-V feat. correlation90.595.333.1 / 74.9
full92.798.934.7 / 77.1
", + "bbox": [ + 532, + 306, + 888, + 422 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. Ablation study on synthetic homography dataset and HPatches. The \"full\" model is the default model.", + "bbox": [ + 511, + 431, + 906, + 460 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "icantly as more source views are integrated into the collaborative inference process. However, when the number of views exceeds 5, the matching capability begins to decline. We hypothesize that this degradation is attributed to the expanded search space and the excessive redundant information introduced by multi-view attention, which collectively increase the learning complexity of the network.", + "bbox": [ + 511, + 488, + 906, + 595 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 609, + 633, + 625 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We proposed a multi-view collaborative matching strategy to address the challenge of reliable matching uncontrolled image sets in complex scenes. Our method, which consists of a deep collaborative matcher (CoMatcher) and a scalable groupwise pipeline, enables a holistic 3D scene understanding while inherently satisfying the cross-view projection consistency constraint. Extensive experiments have demonstrated that exploiting inter-view connections significantly enhances matching certainty, yielding substantial benefits for downstream tasks such as pose estimation and SfM. We hope this work encourages the research community to expand beyond pairwise matching and further explore the understanding and utilization of multi-view information.", + "bbox": [ + 511, + 633, + 906, + 830 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement. This research is supported by NSFC projects under Grant 42471447, Development Program of China under Grant 2024YFC3811000, and the Fundamental Research Funds for the Central Universities of China under Grant 2042022dx0001.", + "bbox": [ + 511, + 830, + 908, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "21977", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 121, + 187, + 136 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Sameer Agarwal, Noah Snavely, Ian Simon, Steven M Seitz, and Richard Szeliski. Building Rome in a day. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 72-79. IEEE, 2009. 1, 2", + "[2] Relja Arandjelovic, Petr Gronat, Akihiko Torii, Tomas Pajdla, and Josef Sivic. NetVLAD: CNN architecture for weakly supervised place recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5297-5307, 2016. 3", + "[3] Vassileios Balntas, Karel Lenc, Andrea Vedaldi, and Krystian Mikolajczyk. HPatches: A benchmark and evaluation of handcrafted and learned local descriptors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5173-5182, 2017. 2, 5, 8", + "[4] Axel Barroso-Laguna, Sowmya Munukutla, Victor Adrian Prisacariu, and Eric Brachmann. Matching 2D images in 3D: Metric relative pose from metric correspondences. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4852-4863, 2024. 2", + "[5] Herbert Bay, Tinne Tuytelaars, and Luc Van Gool. SURF: Speeded up robust features. In European Conference on Computer Vision, pages 404-417. Springer, 2006. 2", + "[6] Ruojin Cai, Joseph Tung, Qianqian Wang, Hadar Averbuch-Elor, Bharath Hariharan, and Noah Snavely. Doppelgangers: Learning to disambiguate images of similar structures. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 34–44, 2023. 1", + "[7] Hongkai Chen, Zixin Luo, Jiahui Zhang, Lei Zhou, Xuyang Bai, Zeyu Hu, Chiew-Lan Tai, and Long Quan. Learning to match features with seeded graph matching network. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6301-6310, 2021. 2, 6", + "[8] Hongkai Chen, Zixin Luo, Lei Zhou, Yurun Tian, Mingmin Zhen, Tian Fang, David Mckinnon, Yanghai Tsin, and Long Quan. ASpanFormer: Detector-free image matching with adaptive span transformer. In European Conference on Computer Vision, pages 20-36. Springer, 2022. 2", + "[9] Ondrej Chum, Jiří Matas, and Josef Kittler. Locally optimized RANSAC. In Pattern Recognition, pages 236-243. Springer, 2003. 7", + "[10] Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. SuperPoint: Self-supervised interest point detection and description. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 224-236, 2018. 2, 6, 7", + "[11] Carl Doersch, Ankush Gupta, Larisa Markeeeva, Adria Recasens, Lucas Smaira, Yusuf Aytar, Joao Carreira, Andrew Zisserman, and Yi Yang. TAP-Vid: A benchmark for tracking any point in a video. Advances in Neural Information Processing Systems, 35:13610-13626, 2022. 2", + "[12] Carl Doersch, Yi Yang, Mel Vecerik, Dilara Gokay, Ankush Gupta, Yusuf Aytar, Joao Carreira, and Andrew Zisserman. TAPIR: Tracking any point with per-frame initialization and" + ], + "bbox": [ + 93, + 146, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "temporal refinement. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10061-10072, 2023. 2, 4", + "[13] Mihai Dusmanu, Ignacio Rocco, Tomas Pajdla, Marc Pollefeys, Josef Sivic, Akihiko Torii, and Torsten Sattler. D2-Net: A trainable cnn for joint description and detection of local features. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8092–8101, 2019. 2", + "[14] Johan Edstedt, Georg Bokman, Marten Wadenback, and Michael Felsberg. DeDoDe: Detect, don't describe—Describe, don't detect for local feature matching. In International Conference on 3D Vision, pages 148-157. IEEE, 2024. 2", + "[15] Johan Edstedt, Qiyu Sun, Georg Bokman, Marten Wadenback, and Michael Felsberg. RoMa: Robust dense feature matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19790-19800, 2024. 1, 2", + "[16] Maha Elbayad, Jiatao Gu, Edouard Grave, and Michael Auli. Depth-adaptive Transformer. In International Conference on Learning Representations, pages 1-14, 2020. 5", + "[17] Martin A Fischler and Robert C Bolles. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. In Communications of the ACM, 24(6):381-395, 1981. 6, 7", + "[18] Jan-Michael Frahm, Pierre Fite-Georgel, David Gallup, Tim Johnson, Rahul Raguram, Changchang Wu, Yi-Hung Jen, Enrique Dunn, Brian Clipp, Svetlana Lazebnik, et al. Building Rome on a cloudless day. In European Conference on Computer Vision, pages 368-381. Springer, 2010. 1", + "[19] Adam W Harley, Zhaoyuan Fang, and Katerina Fragkiadaki. Particle video revisited: Tracking through occlusions using point trajectories. In European Conference on Computer Vision, pages 59-75. Springer, 2022. 2", + "[20] Richard Hartley and Andrew Zisserman. Multiple view geometry in computer vision. Cambridge university press, 2003. 1, 6", + "[21] Xingyi He, Jiaming Sun, Yifan Wang, Sida Peng, Qixing Huang, Hujun Bao, and Xiaowei Zhou. Detector-free structure-from-motion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21594-21603, 2024. 2", + "[22] Martin Humenberger, Yohann Cabon, Nicolas Guerin, Julien Morat, Vincent Leroy, Jérôme Revaud, Philippe Rerole, Noé Pion, Cesar de Souza, and Gabriela Csurka. Robust image retrieval-based visual localization using kapture. arXiv preprint arXiv:2007.13867, 2020. 1", + "[23] Heinly Jared, Johannes L Schonberger, Enrique Dunn, and Jan-Michael Frahm. Reconstructing the world in six days. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2015. 1", + "[24] Hanwen Jiang, Arjun Karpur, Bingyi Cao, Qixing Huang, and André Araujo. OmniGlue: Generalizable feature matching with foundation model guidance. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19865-19875, 2024. 2" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "21978", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[25] Yuhe Jin, Dmytro Mishkin, Anastasiia Mishchuk, Jiri Matas, Pascal Fua, Kwang Moo Yi, and Eduard Trulls. Image matching across wide baselines: From paper to practice. International Journal of Computer Vision, 129(2):517-547, 2021. 1, 2, 7", + "[26] Nikita Karaev, Ignacio Rocco, Benjamin Graham, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. CoTracker: It is better to track together. In European Conference on Computer Vision, pages 18-35. Springer, 2024. 2, 4", + "[27] Vincent Leroy, Yohann Cabon, and Jérôme Revaud. Grounding image matching in 3D with MAST3R. In European Conference on Computer Vision, pages 71-91. Springer, 2024. 1, 2, 7", + "[28] Zhengqi Li and Noah Snavely. MegaDepth: Learning single-view depth prediction from internet photos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2041-2050, 2018. 5, 6", + "[29] Philipp Lindenberger, Paul-Edouard Sarlin, Viktor Larsson, and Marc Pollefeys. Pixel-perfect structure-from-motion with featuremetric refinement. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5987-5997, 2021. 1, 7", + "[30] Philipp Lindenberger, Paul-Edouard Sarlin, and Marc Pollefeys. LightGlue: Local feature matching at light speed. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17627–17638, 2023. 1, 2, 5, 6, 7, 8", + "[31] David G Lowe. Distinctive image features from scale-invariant keypoints. International Journal of Computer Vision, 60:91-110, 2004. 1, 2, 6, 7", + "[32] Jiayi Ma, Xingyu Jiang, Aoxiang Fan, Junjun Jiang, and Junchi Yan. Image matching from handcrafted to deep features: A survey. International Journal of Computer Vision, 129(1):23-79, 2021. 1, 2", + "[33] Eleonora Maset, Federica Arrigoni, and Andrea Fusiello. Practical and efficient multi-view matching. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4568-4576, 2017. 2", + "[34] Raul Mur-Artal, Jose Maria Martinez Montiel, and Juan D Tardos. ORB-SLAM: A versatile and accurate monocular SLAM system. In IEEE Transactions on Robotics, 31(5): 1147-1163, 2015. 1", + "[35] Carl Olsson and Olof Enqvist. Stable structure-from-motion for unordered image collections. In Image Analysis, pages 524-535. Springer, 2011. 2", + "[36] Linfei Pan, Daniel Barath, Marc Pollefeys, and Johannes L Schonberger. Global structure-from-motion revisited. In European Conference on Computer Vision, 2024. 2", + "[37] Filip Radenović, Ahmet Iscen, Giorgos Tolias, Yannis Avrithis, and Ondrej Chum. Revisiting Oxford and Paris: Large-scale image retrieval benchmarking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5706-5715, 2018. 6, 8", + "[38] Filip Radenovic, Giorgos Tolias, and Ondrej Chum. Finetuning CNN image retrieval with no human annotation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 41(7):1655-1668, 2018. 3" + ], + "bbox": [ + 91, + 90, + 480, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[39] Jerome Revaud, Cesar De Souza, Martin Humenberger, and Philippe Weinzaepfel. R2D2: Reliable and repeatable detector and descriptor. Advances in Neural Information Processing Systems, 32, 2019. 2", + "[40] Barbara Roessle and Matthias Nießner. End2end multi-view feature matching with differentiable pose optimization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 477-487, 2023. 2, 6", + "[41] Ethan Rublee, Vincent Rabaud, Kurt Konolige, and Gary Bradski. ORB: An efficient alternative to SIFT or SURF. In Proceeding of the IEEE/CVF International Conference on Computer Vision, pages 2564-2571. IEEE, 2011. 2", + "[42] Paul-Edouard Sarlin, Cesar Cadena, Roland Siegwart, and Marcin Dymczyk. From coarse to fine: Robust hierarchical localization at large scale. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12716–12725, 2019. 1, 2", + "[43] Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. SuperGlue: Learning feature matching with graph neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4938–4947, 2020. 1, 2, 4, 5, 6, 7, 8", + "[44] Torsten Sattler, Akihiko Torii, Josef Sivic, Marc Pollefeys, Hajime Taira, Masatoshi Okutomi, and Tomas Pajdla. Are large-scale 3d models really necessary for accurate visual localization? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1637-1646, 2017. 2", + "[45] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-Motion revisited. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4104-4113, 2016. 1, 2, 7", + "[46] Johannes L Schonberger, Hans Hardmeier, Torsten Sattler, and Marc Pollefeys. Comparative evaluation of hand-crafted and learned local features. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1482-1491, 2017. 2", + "[47] Tal Schuster, Adam Fisch, Jai Gupta, Mostafa Dehghani, Dara Bahri, Vinh Tran, Yi Tay, and Donald Metzler. Confident adaptive language modeling. Advances in Neural Information Processing Systems, 35:17456-17472, 2022. 5", + "[48] Sivic and Zisserman. Video Google: A text retrieval approach to object matching in videos. In Proceedings in the IEEE/CVF International Conference on Computer Vision, pages 1470-1477. IEEE, 2003. 3", + "[49] Jianlin Su, Murtadha Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu. RoFormer: Enhanced transformer with rotary position embedding. Neurocomputing, 568:127063, 2024. 4", + "[50] Jiaming Sun, Zehong Shen, Yuang Wang, Hujun Bao, and Xiaowei Zhou. LoFTR: Detector-free local feature matching with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8922-8931, 2021. 1, 2, 5, 6", + "[51] Kun Sun, Jinhong Yu, Wenbing Tao, Xin Li, Chang Tang, and Yuhua Qian. A unified feature-spatial cycle consistency fusion framework for robust image matching. Information Fusion, 97:101810, 2023. 2" + ], + "bbox": [ + 516, + 90, + 906, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "21979", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[52] Prune Truong, Martin Danelljan, Luc Van Gool, and Radu Timofte. Learning accurate dense correspondences and when to trust them. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5714-5724, 2021. 2, 6", + "[53] Michal Tyszkiiewicz, Pascal Fua, and Eduard Trulls. DISK: Learning local features with policy gradient. Advances in Neural Information Processing Systems, 33:14254-14265, 2020. 2, 5, 6, 7", + "[54] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in Neural Information Processing Systems, 30, 2017. 1, 5", + "[55] Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. DUSt3R: Geometric 3D vision made easy. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20697-20709, 2024. 7", + "[56] Yifan Wang, Xingyi He, Sida Peng, Dongli Tan, and Xiaowei Zhou. Efficient LoFTR: Semi-dense local feature matching with sparse-like speed. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21666-21675, 2024. 2, 5", + "[57] Yuxi Xiao, Qianqian Wang, Shangzhan Zhang, Nan Xue, Sida Peng, Yujun Shen, and Xiaowei Zhou. SpatialTracker: Tracking any 2D pixels in 3D space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20406-20417, 2024. 1", + "[58] Shibiao Xu, Shunpeng Chen, Rongtao Xu, Changwei Wang, Peng Lu, and Li Guo. Local feature matching using deep learning: A survey. Information Fusion, 107:102344, 2024. 1", + "[59] Kwang Moo Yi, Eduard Trulls, Vincent Lepetit, and Pascal Fua. LIFT: Learned invariant feature transform. In European Conference on Computer Vision, pages 467-483. Springer, 2016. 2", + "[60] Guofeng Zhang, Zilong Dong, Jiaya Jia, Tien-Tsin Wong, and Hujun Bao. Efficient non-consecutive feature tracking for structure-from-Motion. In European Conference on Computer Vision, pages 422-435. Springer, 2010. 2" + ], + "bbox": [ + 91, + 90, + 480, + 656 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "21980", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/CoMatcher_ Multi-View Collaborative Feature Matching/b3062684-7794-47f2-b9d1-53d25767d94f_model.json b/2025/CoMatcher_ Multi-View Collaborative Feature Matching/b3062684-7794-47f2-b9d1-53d25767d94f_model.json new file mode 100644 index 0000000000000000000000000000000000000000..5604ba11c168aca4833edeccdd0cb5a4e8f8dc85 --- /dev/null +++ b/2025/CoMatcher_ Multi-View Collaborative Feature Matching/b3062684-7794-47f2-b9d1-53d25767d94f_model.json @@ -0,0 +1,2125 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.211, + 0.131, + 0.788, + 0.154 + ], + "angle": 0, + "content": "CoMatcher: Multi-View Collaborative Feature Matching" + }, + { + "type": "text", + "bbox": [ + 0.137, + 0.18, + 0.859, + 0.254 + ], + "angle": 0, + "content": "Jintao Zhang\\(^{1}\\), Zimin Xia\\(^{2*}\\), Mingyue Dong\\(^{1}\\), Shuhan Shen\\(^{3}\\), Linwei Yue\\(^{4}\\), Xianwei Zheng\\(^{1*}\\) \n\\(^{1}\\)The State Key Lab. LIESMARS, Wuhan University \n\\(^{2}\\)École Polytechnique Fédérale de Lausanne (EPFL) \n\\(^{3}\\)Chinese Academy of Sciences \\({ }^{4}\\)China University of Geosciences" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.288, + 0.327, + 0.304 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.32, + 0.485, + 0.609 + ], + "angle": 0, + "content": "This paper proposes a multi-view collaborative matching strategy for reliable track construction in complex scenarios. We observe that the pairwise matching paradigms applied to image set matching often result in ambiguous estimation when the selected independent pairs exhibit significant occlusions or extreme viewpoint changes. This challenge primarily stems from the inherent uncertainty in interpreting intricate 3D structures based on limited two-view observations, as the 3D-to-2D projection leads to significant information loss. To address this, we introduce CoMatcher, a deep multi-view matcher to (i) leverage complementary context cues from different views to form a holistic 3D scene understanding and (ii) utilize cross-view projection consistency to infer a reliable global solution. Building on CoMatcher, we develop a groupwise framework that fully exploits cross-view relationships for large-scale matching tasks. Extensive experiments on various complex scenarios demonstrate the superiority of our method over the mainstream two-view matching paradigm." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.638, + 0.222, + 0.654 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.664, + 0.483, + 0.845 + ], + "angle": 0, + "content": "Research on recovering 3D structures and camera poses from multiple views storing scene information has a long history [20, 23, 34]. A core component of this process is feature matching, which aims to estimate point correspondences across the image set [25, 32, 58]. For computational flexibility, existing methods often decompose the set into pairs of co-visible images [1, 18] and apply a two-view matching approach to each pair independently [31, 43]. The pairwise matches are then merged into comprehensive multi-view tracks [1, 45]. Following this pairwise framework to build systems for localization [22, 42] and mapping [29, 45] has become a entrenched research paradigm." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.846, + 0.483, + 0.876 + ], + "angle": 0, + "content": "Recent advances in learning-based two-view matchers have greatly improved performance in most cases [15, 30," + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.288, + 0.902, + 0.502 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.513, + 0.907, + 0.598 + ], + "angle": 0, + "content": "Figure 1. From two view to multi-view. Unlike pairwise schemes prone to uncertainty, we first partition the image set into co-visible groups and collaboratively estimate group-wide correspondences using a reliable deep multi-view matcher. Benefiting from holistic scene understanding and consistency constraint, our method generates high-quality tracks in complex scenarios." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.627, + 0.907, + 0.703 + ], + "angle": 0, + "content": "43, 50]. Inspired by Transformer [54], these methods employ deep networks to jointly analyze global context from both views. This enables models to learn to infer underlying 3D scenes by integrating spatial and visual cues, which has been proven crucial for accurate estimations [27, 43]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Despite great progress, current two-view matchers still struggle in challenging wide-baseline scenarios, particularly with severe occlusions and repetitive textures [25]. We argue that one potential cause for this difficulty lies in the inherent uncertainty of interpreting complex 3D structures from limited two-view observations—a task that even humans find prone to perceptual ambiguity. Projecting complex 3D geometries onto 2D planes inevitably results in significant loss of scene information. For example, spatially distant regions may appear close in 2D. This makes it highly unreliable to infer the original scene solely from ambiguous two-view context [6, 57]. Moreover, while matchers can learn two-view geometric priors, this alone is also in" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.236, + 0.9 + ], + "angle": 0, + "content": "*Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "21970" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.152 + ], + "angle": 0, + "content": "adequate—capturing abrupt depth discontinuities requires stronger constraints [4, 27]. Errors in pairwise correspondences compound during merging, significantly impacting downstream tasks, especially in large-scale datasets." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.153, + 0.482, + 0.348 + ], + "angle": 0, + "content": "To address these challenges, we propose a shift in focus: instead of optimizing two-view matchers, a more effective strategy is to directly leverage the rich relationships inherent in raw multi-view observations. Building on this, we introduce a unified formulation that collaboratively establishes correspondences between a complementary view group and a target view. This approach fosters holistic scene understanding by integrating contextual information from multiple views, while multi-view geometric priors provide stronger constraints for reliable reasoning. Moreover, matches across views naturally adhere to cross-view projection consistency. This facilitates a mutual verification process, ensuring confident inference of a global solution." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.349, + 0.483, + 0.544 + ], + "angle": 0, + "content": "Inspired by existing two-view methods [43, 50], we develop CoMatcher, a deep-learning architecture for multiview feature matching. Context sharing is achieved through a GNN with multi-view receptive fields. While a larger feature space enriches context, it also introduces noise, complicating learning. To address this, we propose to constrain the search space for each point by leveraging cross-view projection geometry. Furthermore, we redesign the feature correlation layers in cross-attention to enable consistent multiview reasoning. This design exploits the property that correlation scores encode both reasoning outcomes and confidence, enabling progressive layer-by-layer integration of multi-view estimates into a globally consistent solution." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.545, + 0.483, + 0.786 + ], + "angle": 0, + "content": "To scale CoMatcher for large-scale tasks like SfM [36, 45], we develop a groupwise matching pipeline (see Fig. 1). It introduces a novel set partitioning algorithm that uses可视ability information to group images, followed by group-level matching using CoMatcher. By preserving cross-view relationships, our framework achieves comprehensive improvements over pairwise methods in extensive experiments on wide-baseline scenarios. The main contributions are summarized as follows: (i) A novel multi-view matching paradigm, comprising a deep collaborative matcher and a scalable groupwise pipeline, designed to overcome the inherent uncertainty of pairwise approaches. (ii) A multiview representation learning network architecture that enables holistic scene understanding. (iii) A multi-view feature correlation strategy that integrates uncertain individual estimates into a globally consistent solution." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.8, + 0.228, + 0.815 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.482, + 0.901 + ], + "angle": 0, + "content": "Two-view matching is a fundamental problem in computer vision. The most prevalent paradigm involves sparsifying images into keypoints, each represented by a high-dimensional vector encoding its visual context [25, 32]. This transforms matching into a search problem, with match" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.393 + ], + "angle": 0, + "content": "likelihood expressed by the inner product of corresponding vectors [31]. Modern methods leverage deep networks for representation learning to extract feature points [10, 13, 14, 39, 53, 59], outperforming traditional hand-crafted encodings [5, 31, 41] and achieving significant advancements on mainstream benchmarks [3, 25, 46]. In parallel, other works aim to address the limitations of heuristic matching strategies by leveraging networks with global modeling capabilities to directly learn the matching process [7, 24, 30, 43]. Such data-driven methods can effectively capture priors of the underlying 3D scene, which are essential for accurate estimation. Unlike keypoint-based approaches, another typical method retains all pixels for dense inference [8, 15, 27, 50, 52]. While this has led to notable progress in weakly textured regions, such methods require greater computational resources [30, 56] and struggle with multi-view inconsistency [21]. In summary, while these methods perform well in most scenarios, they remain constrained in complex wide-baseline settings due to the limited receptive field inherent to two-view systems." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.394, + 0.905, + 0.726 + ], + "angle": 0, + "content": "Multi-view matching focuses on establishing correspondences across a larger-scale image collection. A typical pairwise matching approach decomposes the problem into two-view matching instances and subsequently merges the pairwise results using multi-view projection consistency [1, 42, 44, 45]. During merging, some works explore how consistency assumptions can filter and refine coarse matches [33, 51], while others focus on efficiency, developing better strategies to reduce redundant matching computations [35, 60]. However, as a post-processing step, merging cannot address catastrophic failures that occur during the matching stage. Another line of work treats matching as a point tracking problem [11, 12, 19, 26], using optical flow-like frameworks to reason about the positions of query pixels across multiple images. While this approach is well-suited for video inputs, its effectiveness on unstructured, wide-baseline image collections remains largely unverified. The most relevant method to ours is [40], which performs end-to-end multi-view matching and pose estimation. However, it is limited in scalability, as it can only handle a small image group. In contrast, our framework is highly flexible, with the potential to scale to an arbitrary number of images." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.74, + 0.649, + 0.757 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.765, + 0.905, + 0.855 + ], + "angle": 0, + "content": "**Formulation.** Given a set of \\(N_I\\) images, \\(\\mathcal{I} = \\{I_i \\mid i = 1, \\dots, N_I\\}\\), of a scene, the objective of feature matching is to identify the 2D position coordinates \\(\\mathcal{M}_k = \\{\\mathbf{p}^{I_i}\\}\\) in \\(\\mathcal{I}\\) that correspond to identical 3D scene points \\(\\mathcal{X} = \\{\\mathbf{x}_k \\mid k = 1, \\dots, N_X\\}\\), where \\(\\mathbf{p}^{I_i} \\in \\mathbb{R}^2\\) are the pixel coordinates in \\(I_i\\), and each \\(\\mathcal{M}_k\\) is called a track." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Motivation. Constructing complete tracks requires estimating correspondences between each image \\(I_{t} \\subset \\mathcal{I}\\) and its co-visible image set \\(\\mathcal{C}_t\\). While the pairwise approach of" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "21971" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.116, + 0.089, + 0.891, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.316, + 0.908, + 0.388 + ], + "angle": 0, + "content": "Figure 2. The CoMatcher architecture can be viewed as an extension of parallel two-view matchers, enhanced with two core components: the multi-view (M-V) feature interaction module (Sec. 3.2) and the multi-view (M-V) feature correlation strategy (Sec. 3.3). The first component employs multi-view cross-attention to enhance the representation of all source and target view features, while leveraging the cross-view projection geometry from \\(\\mathcal{M}(\\mathcal{G})\\) to constrain the point search space. The second component aggregates correlation scores from multiple two-view cross-attention modules based on estimated confidence, guiding the network to achieve multi-view consistent reasoning." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.412, + 0.486, + 0.64 + ], + "angle": 0, + "content": "fers implementation flexibility for this problem, its reliance on two-view inputs in a single estimation often leads to uncertainty in complex scenes. In practice, images within \\(\\mathcal{C}_t\\) are often not independent—most are captured from adjacent positions but with differing viewing angles. This inspires us to group \\(\\mathcal{C}_t\\) based on image correlations, where each complementary group offers a mapping from multiple planar views to a 3D sub-scene. In the image matching problem, the correspondence between \\(I_t\\) and a group must adhere to certain 2D-to-3D physical constraints, such as cross-view projection consistency. Rather than independently estimating and then integrating them with constraints, it is more efficient to unify this process and learn underlying priors directly from the data. Thus, we formulate CoMatcher to address the 1-to-\\(N\\) matching problem (see Fig. 2)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.645, + 0.486, + 0.903 + ], + "angle": 0, + "content": "Groupwise matching pipeline. Based on the above, we propose a three-step pipeline for constructing tracks: grouping, connecting, and matching. First, we treat the original set of views \\(\\mathcal{I}\\) as a union of smaller image groups \\(\\{\\mathcal{G}_s\\mid s = 1,\\dots ,N_G\\}\\), each representing a localized scene. This involves a pre-processing step to partition the set. The most critical aspect of this is determining the correlation between two images, i.e., whether they describe the same structure. Building on existing research [2, 38, 48], we employ heuristic techniques to design a grouping algorithm. Next, although CoMatcher can directly learn multi-view geometry, certain fine-grained physical rules are challenging to model. Therefore, we explicitly establish relationships within each group to guide subsequent matching. For each group, we construct tracks using existing frameworks, which provide a set of 3D points along with their multi-view projections and implicitly encode camera poses. These cues" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.412, + 0.909, + 0.49 + ], + "angle": 0, + "content": "are incorporated as inputs to the network to steer the inference process. Finally, each group of images is treated as a whole to collaboratively match with other images in \\(\\mathcal{I}\\) using CoMatcher. Additional details on grouping and the overall framework are provided in the supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.498, + 0.713, + 0.513 + ], + "angle": 0, + "content": "3.1. CoMatcher overview" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.52, + 0.909, + 0.732 + ], + "angle": 0, + "content": "As shown in Fig. 2, CoMatcher, as a deep sparse matcher, estimates point correspondences between each source view in a group \\( I_{i} \\in \\mathcal{G} \\) and a target view \\( I_{t} \\). The network takes as input the extracted local features \\( \\{\\left(\\mathbf{p}_{k}^{I_{i}}, \\mathbf{d}_{k}^{I_{i}}\\right) \\mid k = 1, \\dots, N_{F}\\} \\) from the \\( M \\) source views and the target view, along with the precomputed tracks \\( \\mathcal{M}(\\mathcal{G}) \\) of \\( \\mathcal{G} \\). Each local feature consists of keypoint locations \\( \\mathbf{p}_{k} \\) and their corresponding visual descriptors \\( \\mathbf{d}_{k} \\). Initially, the target view features are broadcast and paired with those of each source view to establish the initial states. These \\( M \\) feature pairs are then refined using a Graph Neural Network (GNN) with alternating self-attention and cross-attention modules, repeated \\( L \\) times. Finally, the enhanced features are fed into matching heads to infer the final correspondences." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.733, + 0.909, + 0.839 + ], + "angle": 0, + "content": "The design of CoMatcher revolves around two core questions: (i) how can multi-view context be leveraged to learn better point representations (Sec. 3.2), and (ii) how can the estimations across multiple views be constrained to satisfy cross-view consistency (Sec. 3.3)? After thoroughly discussing these questions, we introduce the prediction heads (Sec. 3.4) and the loss function (Sec. 3.5)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.849, + 0.783, + 0.863 + ], + "angle": 0, + "content": "3.2. Multi-view feature interaction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.908, + 0.902 + ], + "angle": 0, + "content": "Consider an occluded scene. When observing such structures from a frontal view, points near occlusion boundaries" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "21972" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.212 + ], + "angle": 0, + "content": "are often contaminated by irrelevant context. This noise significantly compromises the quality of point features, creating substantial challenges for matching. Multi-view learning provides a promising solution to this issue. For these ambiguous point features, it can integrate observations of the same area from multiple other viewpoints to refine their representations. This involves a querying process, which establishes a multi-view receptive field for each point." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.214, + 0.482, + 0.271 + ], + "angle": 0, + "content": "Given a set of co-visible source views, CoMatcher leverages a multi-view cross-attention mechanism to aggregate these context (Source Cross). For a point \\( u \\) in source view \\( I_{i} \\), we attend to all points in the other source views:" + }, + { + "type": "equation", + "bbox": [ + 0.105, + 0.286, + 0.483, + 0.322 + ], + "angle": 0, + "content": "\\[\n\\mathbf {m} _ {u} ^ {I _ {i} \\leftarrow \\mathcal {W} _ {j}} = \\frac {1}{M - 1} \\sum_ {I _ {j} \\in \\mathcal {G} \\backslash I _ {i}} \\sum_ {v \\in \\mathcal {W} _ {j}} \\operatorname {S o f t m a x} \\left(a _ {u v} ^ {I _ {i} I _ {j}}\\right) \\mathbf {v} _ {v} ^ {I _ {j}}. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.337, + 0.483, + 0.463 + ], + "angle": 0, + "content": "Here, \\(\\mathcal{W}_j\\) denotes the point set in the source view \\(I_{j}\\), while \\(a_{uv}^{I_iI_j}\\) represents the similarity score computed by treating point \\(u\\) in \\(I_{i}\\) as the query and point \\(v\\) in \\(I_{j}\\) as the key. \\(\\mathbf{v}_v^{I_j}\\) is the value of point \\(v\\). \\(\\mathbf{m}_u^{I_i}\\) represents the message vector from \\(\\mathcal{W}_j\\) to \\(u\\) of the GNN [43]. We uniformly aggregate features from each source view, encouraging the model to integrate multi-view information comprehensively rather than overemphasizing images from similar perspectives." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.465, + 0.483, + 0.616 + ], + "angle": 0, + "content": "However, for a point near an occlusion boundary in one source view, how does it identify the corresponding points observing the same location in other source views? This implies a matching process. While the network could directly learn this, it would introduce additional complexity, as establishing correspondences between source views is not our primary objective. Moreover, while multi-view setups provide abundant contextual information, much of it is irrelevant, such as regions lacking co-visibility. This redundancy also imposes an additional burden on learning." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.618, + 0.483, + 0.709 + ], + "angle": 0, + "content": "To address this, we propose a geometric constraint mechanism to explicitly guide the attention range of each point. By incorporating relative positional encoding into the attention score computation, we enable the scores to depend on two critical factors: feature correlation and the geometric distance \\(\\Delta \\mathbf{p}\\) between two points:" + }, + { + "type": "equation", + "bbox": [ + 0.178, + 0.724, + 0.483, + 0.747 + ], + "angle": 0, + "content": "\\[\na _ {u v} ^ {I _ {i} I _ {j}} = \\left(\\mathbf {q} _ {u} ^ {I _ {i}}\\right) ^ {\\top} \\mathbf {R} \\left[ \\Delta \\mathbf {p} _ {u v} ^ {I _ {i} I _ {j}} \\right] \\mathbf {k} _ {v} ^ {I _ {j}}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.76, + 0.407, + 0.776 + ], + "angle": 0, + "content": "Here, \\(\\mathbf{R}[\\cdot ]\\) is the rotation encoding matrix [49]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.778, + 0.483, + 0.901 + ], + "angle": 0, + "content": "While the 2D relative distance between two points within a single image is well-defined, across different views, it depends on the projection transformation. We use the precomputed tracks \\(\\mathcal{M}(\\mathcal{G})\\) to find the projection position of \\(\\mathbf{p}_u^{I_i}\\) on \\(I_{j}\\), which can be denoted as \\(\\mathbf{p}_w^{I_j}\\). When no corresponding point in \\(I_{j}\\) can be found for \\(\\mathbf{p}_u^{I_i}\\), we simply assign the relative position to \\(\\mathbf{0}\\) so that the attention score depends solely on the feature similarity. The calculation of the relative po" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.092, + 0.752, + 0.106 + ], + "angle": 0, + "content": "sition can be formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.561, + 0.115, + 0.905, + 0.158 + ], + "angle": 0, + "content": "\\[\n\\Delta \\mathbf {p} _ {u v} ^ {I _ {i} I _ {j}} = \\left\\{ \\begin{array}{l l} \\mathbf {p} _ {w} ^ {I _ {j}} - \\mathbf {p} _ {v} ^ {I _ {j}} & (u, w) \\in M (I _ {i}, I _ {j}) \\\\ \\mathbf {0} & (u, \\emptyset) \\in M (I _ {i}, I _ {j}). \\end{array} \\right. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.168, + 0.907, + 0.334 + ], + "angle": 0, + "content": "Additionally, we also interact with all target view features after source cross-attention (Target Cross). For a point \\( t \\) in view \\( I_{t} \\) paired with source view \\( I_{i} \\), we attend to the same point \\( t \\) across the remaining \\( M - 1 \\) target views. This is analogous to operations in some works that aggregate features along the temporal dimension [12, 26]. Unlike the approach for source views, we compute the attention scores solely based on feature similarity. Despite its low computational complexity, this module complements the Source Cross, mutually enhancing each other to help the network implicitly infer globally optimal correspondences." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.345, + 0.783, + 0.359 + ], + "angle": 0, + "content": "3.3. Multi-view feature correlation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.367, + 0.907, + 0.525 + ], + "angle": 0, + "content": "For a track in a source view set originating from a 3D point, its corresponding 2D keypoints \\(\\left\\{\\mathbf{p}_{u}^{I_{i}}, \\mathbf{p}_{v}^{I_{j}}, \\ldots\\right\\}\\) should exhibit consistent matches in the target view—either corresponding to the same single point or having no corresponding point at all. This physical constraint is often utilized by existing methods as a post-processing step to filter out erroneous results when merging multiple two-view correspondences. However, as a multi-view matcher, we aim for CoMatcher to inherently satisfy this constraint in its reasoning, thereby enhancing its reliability and confidence." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.525, + 0.906, + 0.615 + ], + "angle": 0, + "content": "Achieving this involves two key processes: (i) identifying points that may lead to errors in the early layers of the network, and (ii) promptly guiding these points toward correct estimation using inference information from other views. To accomplish this, we propose a two-step multiview feature correlation strategy (see Fig. 3)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.616, + 0.906, + 0.661 + ], + "angle": 0, + "content": "First, to identify ambiguous points in the source views, we utilize a lightweight head at the end of each layer to predict the confidence of each point:" + }, + { + "type": "equation", + "bbox": [ + 0.584, + 0.672, + 0.905, + 0.693 + ], + "angle": 0, + "content": "\\[\nc _ {u} ^ {I _ {i}} = \\operatorname {S i g m o i d} \\left(\\operatorname {M L P} \\left(\\mathbf {f} _ {u} ^ {I _ {i}}\\right)\\right) \\in [ 0, 1 ]. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.704, + 0.906, + 0.78 + ], + "angle": 0, + "content": "Here, MLP represents a multi-layer perceptron and \\(\\mathbf{f}_u^{I_i}\\) is the feature of \\(u\\). This confidence score reflects the assessment of each point in its current state: higher values indicate that, at this layer, the point is either converging toward a reliable match or confidently identified as having no match." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.906, + 0.871 + ], + "angle": 0, + "content": "In each layer, we define a distinct hyperparameter threshold \\(\\theta\\). Points with confidence scores below this threshold are identified as potentially ambiguous and likely to lead to incorrect inference. We progressively increase the value of \\(\\theta\\) in later layers, as the confidence of overall inference generally improves with each successive layer." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Next, we redesign the parallel two-view cross-attention modules to correct the erroneous estimations of ambiguous" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21973" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.116, + 0.088, + 0.473, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.226, + 0.485, + 0.285 + ], + "angle": 0, + "content": "Figure 3. The multi-view feature correlation strategy consists of two steps: we first identify potentially erroneous points using a confidence estimator, and then correct their estimation by aggregating attention correlation vectors from other views." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.309, + 0.484, + 0.34 + ], + "angle": 0, + "content": "points. In these modules, each point in the source view computes an attention distribution vector to query [54]:" + }, + { + "type": "equation", + "bbox": [ + 0.207, + 0.349, + 0.483, + 0.375 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\alpha} _ {u} ^ {I _ {i}} = \\underset {x \\in \\mathcal {W} _ {t}} {\\operatorname {S o f t m a x}} \\left(a _ {u x} ^ {I _ {i} I _ {t}}\\right). \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.385, + 0.484, + 0.498 + ], + "angle": 0, + "content": "This vector represents the degree to which the point attends to in view \\( I_{t} \\), with a peak indicating potential matching regions of interest. For ambiguous points, this similarity is not reliable, often leading to incorrect feature aggregation. Assuming that point \\( u \\) in the source view \\( I_{i} \\) is an ambiguous point, we update its original attention distribution by applying a weighted average to the distributions of its corresponding points across multi-view:" + }, + { + "type": "equation", + "bbox": [ + 0.102, + 0.505, + 0.483, + 0.55 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\alpha} _ {u} ^ {I _ {i} ^ {\\prime}} = c _ {u} ^ {I _ {i}} \\boldsymbol {\\alpha} _ {u} ^ {I _ {i}} + \\left(1 - c _ {u} ^ {I _ {i}}\\right) \\frac {\\sum_ {v \\in \\mathcal {D} _ {m} ^ {u}} c _ {v} ^ {I _ {j}} \\boldsymbol {\\alpha} _ {v} ^ {I _ {j}}}{\\sum_ {v \\in \\mathcal {D} _ {m} ^ {u}} c _ {v} ^ {I _ {j}}} \\quad \\text {s . t .} c _ {u} ^ {I _ {i}} < \\theta . \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.557, + 0.484, + 0.618 + ], + "angle": 0, + "content": "Here, \\( \\mathcal{D}_m^u \\) represents other points in the track. We embed this strategy into the model inference process, which significantly enhances the reliability of estimation under challenging conditions despite its simple implementation." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.627, + 0.338, + 0.644 + ], + "angle": 0, + "content": "3.4. Correspondence prediction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.649, + 0.484, + 0.831 + ], + "angle": 0, + "content": "As shown in Fig. 2, a lightweight matching head operates in parallel for each pair of network output features to predict correspondences between two views [30, 43]. In each head, we begin by calculating a score matrix from feature correlations between two views: \\(\\mathbf{S}(u,x) = \\mathrm{L}(\\mathbf{f}_u^{I_i})\\cdot\\) \\(\\mathrm{L}(\\mathbf{f}_x^{I_t})\\), where \\(\\mathrm{L}\\) is the linear transform. Then we apply dual-softmax [30, 50, 53] to derive the matching probabilities of two points: \\(\\mathbf{S}'(u,x) = \\mathrm{Softmax}(\\mathbf{S}(u,\\cdot))_x\\cdot\\) Softmax \\((\\mathbf{S}(\\cdot ,x))_u\\). Additionally, we employ the method proposed in [30] to calculate the matching probability for each point: \\(\\sigma_{i}^{I_{i}} = \\mathrm{Sigmoid}(\\mathrm{L}(\\mathbf{f}_{i}^{I_{i}}))\\). The final assignment matrix is a combination of the two probabilities:" + }, + { + "type": "equation", + "bbox": [ + 0.197, + 0.841, + 0.483, + 0.861 + ], + "angle": 0, + "content": "\\[\n\\mathbf {P} (u, x) = \\mathbf {S} ^ {\\prime} (u, x) \\sigma_ {u} ^ {I _ {i}} \\sigma_ {x} ^ {I _ {t}}. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Using mutual nearest neighbors and a threshold, we filter the matching results from the assignment matrix [43, 50]." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.091, + 0.587, + 0.105 + ], + "angle": 0, + "content": "3.5. Loss" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.113, + 0.905, + 0.158 + ], + "angle": 0, + "content": "CoMatcher supervises each source-target image pair individually, with the loss for each pairs composed of two components: correspondence and confidence:" + }, + { + "type": "equation", + "bbox": [ + 0.563, + 0.168, + 0.907, + 0.206 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {t o t a l}} = \\frac {1}{M} \\sum_ {I _ {i} \\in \\mathcal {G}} \\left(\\mathcal {L} _ {\\text {c o r r}} \\left(I _ {i}, I _ {t}\\right) + \\alpha \\mathcal {L} _ {\\text {c o n f}} \\left(I _ {i}\\right)\\right). \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.215, + 0.907, + 0.305 + ], + "angle": 0, + "content": "Correspondence supervision. For each pair \\( I_{i} \\) and \\( I_{t} \\), we perform two-view transformations using relative poses or homography to compute ground truth labels [30, 43, 50]. We minimize the negative log-likelihood of the assignment matrix, following [30]. More details are provided in the supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.306, + 0.906, + 0.442 + ], + "angle": 0, + "content": "Confidence supervision. To train the confidence estimators at each layer, we define the point confidence as the consistency probability between its correspondence estimated at the current layer and the final estimation [16, 30, 47]. At each layer, predictions are obtained via dual-softmax on current features, without introducing additional matching probabilities or parameters. The ground truth label \\( y \\) indicates whether the two estimations are consistent. This is supervised using a binary cross-entropy loss:" + }, + { + "type": "equation", + "bbox": [ + 0.563, + 0.452, + 0.906, + 0.489 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {c o n f}} \\left(I _ {i}\\right) = \\frac {1}{L - 1} \\sum_ {\\ell} \\sum_ {u \\in \\mathcal {W} _ {i}} \\mathrm {C E} \\left(\\ell c _ {u} ^ {I _ {i}}, \\ell y _ {u} ^ {I _ {i}}\\right), \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.499, + 0.694, + 0.515 + ], + "angle": 0, + "content": "where \\(\\ell \\in \\{1,\\dots ,L - 1\\}\\)" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.528, + 0.646, + 0.545 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.553, + 0.906, + 0.659 + ], + "angle": 0, + "content": "In this section, we first introduce the datasets used, followed by our implementation details. Then, our CoMatcher network is compared to previous state-of-the-art baselines for homography and camera pose estimation. Next, we integrate CoMatcher into groupwise matching framework and evaluate it against pairwise method on a large-scale benchmark. Finally, an extensive ablation study is provided." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.668, + 0.616, + 0.683 + ], + "angle": 0, + "content": "4.1. Datasets" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.907, + 0.901 + ], + "angle": 0, + "content": "The HPatches dataset is used for homography estimation [3]. It consists of 116 strictly planar scenes, each containing 6 images with variations in viewpoint and illumination. For camera pose estimation tasks, we utilize the MegaDepth dataset, an outdoor dataset that exhibits strong occlusions and significant structural changes [28]. We selected two scenes, \"Sacre Coeur\" and \"St. Peter's Square\", from which 1500 co-visible quadruplets were sampled in a way that balances difficulty based on visual overlap [30, 50, 56]. The Image Matching Challenge 2020 benchmark provides a comprehensive evaluation protocol, including datasets that cover multiple challenging outdoor wide-baseline scenes. From the phototourism dataset, we selected 3 validated scenes, each consisting of about 100" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "21974" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.113, + 0.089, + 0.885, + 0.521 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.534, + 0.907, + 0.591 + ], + "angle": 0, + "content": "Figure 4. Qualitative comparison on MegaDepth. For each quadruple with a target view (top), correspondences (green for correct, red for incorrect) predicted by CoMatcher and LightGlue are shown. Using identical local features as input, CoMatcher achieves significantly more reliable estimations, even at challenging semantic edges with depth discontinuities. This stems from its holistic scene understanding through multi-view cues to effectively address occlusions (Q3), large-scale variations (Q2, 4), and repetitive textures (Q1, 2, 3, 4)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.616, + 0.483, + 0.648 + ], + "angle": 0, + "content": "images. Additionally, we conducted an ablation study on a synthetic homography dataset [37] used for training." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.654, + 0.307, + 0.67 + ], + "angle": 0, + "content": "4.2. Implementation details" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.676, + 0.483, + 0.842 + ], + "angle": 0, + "content": "CoMatcher is trained in two steps, following [30, 43]. We first pre-train the model on a large-scale synthetic homography dataset [37], leveraging noise-free homography for ground truth. Next, fine-tuning is performed on the MegaDepth dataset [28] using the camera poses recovered by SfM as ground truth. We sample 200 co-visible multiview groups per scene and randomly select one image as the target view. The size of source views is set to \\( M = 4 \\) during training. The training process is carried out on two NVIDIA GeForce 4090 GPUs, taking about 6 days in total. More details can be found in the supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.849, + 0.314, + 0.866 + ], + "angle": 0, + "content": "4.3. Homography estimation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.901 + ], + "angle": 0, + "content": "CoMatcher is compared with three types of baselines. First, we include state-of-the-art (SOTA) sparse two-view" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.616, + 0.907, + 0.844 + ], + "angle": 0, + "content": "matchers: nearest-neighbor with mutual check [31], SuperGlue [43], SGMNet [7], and LightGlue [30], each paired with different feature extractors [10, 31, 53]. Second, we evaluate the multi-view matcher End2End [40]. For reference, we also compare against typical dense matchers LoFTR [50] and PDC-Net [52]. For CoMatcher and End2End, we use a single forward pass to obtain matches across all five image pairs, while others follow a pairwise approach. To ensure a fair comparison, we adopt the setups of [30, 50] for the number of features and image resizing. Homography accuracy is evaluated using both robust (RANSAC [17]) and non-robust (weighted DLT [20]) solvers. For each pair, we compute the mean reprojection error of the four image corners and report the area under the cumulative error curve (AUC) up to 1px, 3px, and 5px." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.856, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Tab. 1 shows that CoMatcher yields more accurate estimates than all sparse two-view matchers, highlighting the advantage of inference in multi-view feature space. How-" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21975" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.092, + 0.089, + 0.491, + 0.341 + ], + "angle": 0, + "content": "
MethodAUC - DLT @1px / @3px / @5pxAUC - RANSAC
denseLoFTR38.5 / 66.0 / 71.440.7 / 68.3 / 78.5
PDC-Net36.0 / 65.3 / 73.037.9 / 67.6 / 77.4
sparse 2-viewSIFT+NN+mutual0.0 / 0.0 / 0.035.9 / 65.0 / 75.6
SP+NN+mutual0.0 / 1.9 / 3.434.8 / 64.1 / 74.8
SP+SuperGlue32.2 / 65.1 / 75.737.2 / 68.0 / 78.7
SP+SGMNet31.7 / 64.9 / 76.037.7 / 66.4 / 77.5
SP+LightGlue35.4 / 67.5 / 77.737.2 / 67.8 / 78.1
DISK+NN+mutual1.8 / 5.2 / 7.837.9 / 58.0 / 68.3
DISK+LightGlue34.4 / 64.5 / 74.438.1 / 65.2 / 77.2
multi-viewSP+End2End34.3 / 66.9 / 75.537.0 / 67.2 / 77.5
SP+CoMatcher37.1 / 69.0 / 78.838.4 / 68.9 / 79.0
DISK+CoMatcher36.3 / 66.2 / 75.938.7 / 68.2 / 78.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.352, + 0.483, + 0.394 + ], + "angle": 0, + "content": "Table 1. Homography estimation on HPatches. We report the area under the cumulative error curve (AUC) up to values of 1px, 3px and 5px, using DLT and RANSAC [17] as homography solver." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.424, + 0.484, + 0.576 + ], + "angle": 0, + "content": "ever, despite End2End also jointly reasoning over multiple views like ours, its performance is notably inferior. This underscores the superiority of a \\(N\\)-to-1 architecture over \\(N\\)-to-\\(N\\) for multi-view representation learning. When compared to LoFTR, we observe slightly lower performance under low thresholds. We attribute this mainly to the limitations of the feature extractor in keypoint non-repeatability and localization errors [29]. Additionally, DLT achieves accuracy close to RANSAC on most metrics, reflecting the high-quality correspondences from CoMatcher." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.589, + 0.315, + 0.605 + ], + "angle": 0, + "content": "4.4. Relative pose estimation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.613, + 0.483, + 0.824 + ], + "angle": 0, + "content": "Next, we evaluate CoMatcher on challenging wide-baseline scenes, using the same baselines as in Sec. 4.3. For each quadruplet, including a selected target view, we still use a single forward pass for multi-view methods. Additionally, we test two newer dense methods, DUS3R [55] and MAS3R [27], following their evaluation setups for image resizing. All other settings remain consistent with [30]. The essential matrix is estimated from the correspondences using both vanilla RANSAC [17] and LO-RANSAC [9], following [30]. The pose error is then computed as the maximum angular error of rotation and translation, derived from the decomposition of the essential matrix. We report the AUC at \\(5^{\\circ}\\), \\(10^{\\circ}\\), and \\(20^{\\circ}\\), and record the average runtime of matching each quadruplet on a single 4090 GPU." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Tab. 2 shows that CoMatcher, as a sparse method, comprehensively enhances the estimation performance on different local features [10] and [53]. Importantly, compared to other sparse two-view matchers like LightGlue, a key strength of CoMatcher is its reliability, as illustrated in" + }, + { + "type": "table", + "bbox": [ + 0.522, + 0.089, + 0.9, + 0.34 + ], + "angle": 0, + "content": "
MethodAUC RANSACAUC LO-RANSACTime (ms)
5° / 10° / 20°
denseLoFTR58.0 / 73.1 / 84.467.4 / 81.7 / 89.3182
PDCNet54.7 / 73.1 / 83.567.1 / 80.2 / 87.0231
DUSt3R42.4 / 56.7 / 64.258.1 / 70.4 / 79.6264
MASt3R51.5 / 65.3 / 75.663.5 / 76.3 / 85.2317
SuperPointNN+mutual35.3 / 58.3 / 53.751.4 / 67.3 / 75.99
SuperGlue55.8 / 72.8 / 84.165.1 / 77.2 / 89.287
LightGlue56.2 / 72.7 / 83.567.2 / 80.1 / 88.051
End2End55.3 / 71.4 / 81.267.4 / 81.5 / 87.0152
CoMatcher57.2 / 73.9 / 84.868.3 / 82.2 / 89.169
DISKNN+mutual50.9 / 66.7 / 77.764.0 / 79.5 / 87.69
LightGlue53.2 / 69.2 / 80.268.6 / 80.4 / 87.254
CoMatcher54.9 / 71.2 / 82.068.5 / 82.1 / 88.473
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.352, + 0.907, + 0.394 + ], + "angle": 0, + "content": "Table 2. Relative pose estimation on MegaDepth. We report the AUC at \\(5^{\\circ}\\), \\(10^{\\circ}\\), and \\(20^{\\circ}\\) using different robust estimator, and the average runtime of matching each quadruplet." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.424, + 0.907, + 0.576 + ], + "angle": 0, + "content": "Fig. 3. This highlights the advantage of multi-view collaborative reasoning in understanding complex 3D structures, such as occlusions, even with suboptimal input local features in these areas. Additionally, compared to End2End, CoMatcher achieves more accurate relative poses without the need for cumbersome end-to-end training. While some dense methods may achieve better accuracy on certain metrics, their efficiency is generally much lower. By employing a space-for-time strategy to minimize inference steps, Co-Matcher runs significantly faster than most methods." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.589, + 0.86, + 0.604 + ], + "angle": 0, + "content": "4.5. Evaluation on the IMC 2020 benchmark" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.613, + 0.907, + 0.839 + ], + "angle": 0, + "content": "Next, we evaluate our groupwise framework against pairwise baselines [30, 31, 43] on the Image Matching Challenge (IMC) 2020 benchmark [25]. Given a large-scale unordered image collection, the benchmark requires providing matching results for all image pairs, which are then assessed on two downstream tasks: stereo and multi-view reconstruction. For the stereo task, the accuracy of relative camera poses for each pair is evaluated from correspondences using RANSAC, as in Sec. 4.4. For the multi-view task, all correspondences are fed into COLMAP [45] for SfM, with the final accuracy evaluated based on the estimated multi-view camera poses. We report the AUC at \\(5^{\\circ}\\) and \\(10^{\\circ}\\) across both tasks. Additionally, the average runtime is reported, which is the total matching time for the entire image set divided by the number of pairs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Tab. 3 shows that our framework significantly outperforms existing pairwise approaches in both tasks. This advantage primarily stems from our framework's ability to better leverage the relationships within the original image" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "21976" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.1, + 0.089, + 0.478, + 0.267 + ], + "angle": 0, + "content": "
MethodStereoMulti-ViewTime (ms)
AUC 5°/10°AUC 5°/10°
two-viewSIFT+NN+mutual31.5 / 44.257.2 / 68.53
SP+NN+mutual28.6 / 40.352.9 / 63.43
SP+SuperGlue36.5 / 50.162.3 / 74.832
SP+LightGlue36.8 / 49.464.6 / 75.417
DISK+NN+mutual35.4 / 47.459.3 / 70.23
DISK+LightGlue42.1 / 55.665.2 / 76.219
m-vSP+CoMatcher39.1 / 52.466.2 / 77.525
DISK+CoMatcher44.9 / 57.367.1 / 78.429
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.277, + 0.483, + 0.32 + ], + "angle": 0, + "content": "Table 3. Evaluation on the IMC 2020 benchmark. We report the pose AUC at \\(5^{\\circ}\\), \\(10^{\\circ}\\) for two subtasks, stereo and multi-view reconstruction, along with the average matching runtime." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.348, + 0.484, + 0.44 + ], + "angle": 0, + "content": "collection, enabling the generation of higher-quality tracks in such large-scale tasks. Additional comparisons on track quality for SfM tasks are provided in the supplementary material. In terms of efficiency, our approach is faster than SuperGlue, with the viewpoint grouping process accounting for only about \\(4\\%\\) of the total matching time." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.453, + 0.242, + 0.469 + ], + "angle": 0, + "content": "4.6. Ablation study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.475, + 0.483, + 0.612 + ], + "angle": 0, + "content": "Understanding CoMatcher. The CoMatcher network introduces three key components: multi-view cross-attention modules, a geometric constraint mechanism, and a multiview feature interaction strategy. We validate these design choices by evaluating on two benchmarks: the challenging synthetic homography dataset [37], where precision and recall metrics are employed, and HPatches [3], where the AUC of homography estimation via weighted DLT is measured. For each image, 512 keypoints are extracted." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.613, + 0.483, + 0.839 + ], + "angle": 0, + "content": "Tab. 4 shows the ablation results. Excluding the two cross-attention modules for multi-view interaction reduces the model to a series of parallel two-view matchers [30, 43]. This simplification limits the model's ability to integrate multi-view context, leading to a significant performance decline. Without the geometric constraint mechanism, the model performs even worse than in the previous case. This suggests that direct feature aggregation from multi-view is challenging, as the network requires additional learning to distinguish meaningful context effectively. Without multiview feature correlation, the model loses its ability to guide inference for ambiguous points by leveraging consistency constraints. This highlights the importance of guiding two-view feature correlation in the attention search process, a factor overlooked in previous work." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Impact of the group size. We evaluated the matching performance on the synthetic homography dataset [37] by varying the size of the source view sets used as network input. As shown in Fig. 5, performance initially improves signif" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.097, + 0.903, + 0.238 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.251, + 0.907, + 0.294 + ], + "angle": 0, + "content": "Figure 5. Impact of the group size. Under different numbers of keypoints, the AUC initially increases as the source view set expands, then tends to stabilize or shows a downward trend." + }, + { + "type": "table", + "bbox": [ + 0.534, + 0.307, + 0.889, + 0.423 + ], + "angle": 0, + "content": "
CoMatcherSyntheticHpatches-AUC
precisionrecall@1px / @5px
w/o M-V feat. interaction89.796.632.4 / 75.1
w/o attenu. propagation87.794.231.5 / 73.7
w/o M-V feat. correlation90.595.333.1 / 74.9
full92.798.934.7 / 77.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.433, + 0.907, + 0.461 + ], + "angle": 0, + "content": "Table 4. Ablation study on synthetic homography dataset and HPatches. The \"full\" model is the default model." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.489, + 0.907, + 0.596 + ], + "angle": 0, + "content": "icantly as more source views are integrated into the collaborative inference process. However, when the number of views exceeds 5, the matching capability begins to decline. We hypothesize that this degradation is attributed to the expanded search space and the excessive redundant information introduced by multi-view attention, which collectively increase the learning complexity of the network." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.61, + 0.634, + 0.625 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.635, + 0.907, + 0.831 + ], + "angle": 0, + "content": "We proposed a multi-view collaborative matching strategy to address the challenge of reliable matching uncontrolled image sets in complex scenes. Our method, which consists of a deep collaborative matcher (CoMatcher) and a scalable groupwise pipeline, enables a holistic 3D scene understanding while inherently satisfying the cross-view projection consistency constraint. Extensive experiments have demonstrated that exploiting inter-view connections significantly enhances matching certainty, yielding substantial benefits for downstream tasks such as pose estimation and SfM. We hope this work encourages the research community to expand beyond pairwise matching and further explore the understanding and utilization of multi-view information." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.831, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Acknowledgement. This research is supported by NSFC projects under Grant 42471447, Development Program of China under Grant 2024YFC3811000, and the Fundamental Research Funds for the Central Universities of China under Grant 2042022dx0001." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "21977" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.122, + 0.188, + 0.137 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.147, + 0.483, + 0.203 + ], + "angle": 0, + "content": "[1] Sameer Agarwal, Noah Snavely, Ian Simon, Steven M Seitz, and Richard Szeliski. Building Rome in a day. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 72-79. IEEE, 2009. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.204, + 0.484, + 0.274 + ], + "angle": 0, + "content": "[2] Relja Arandjelovic, Petr Gronat, Akihiko Torii, Tomas Pajdla, and Josef Sivic. NetVLAD: CNN architecture for weakly supervised place recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5297-5307, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.276, + 0.484, + 0.344 + ], + "angle": 0, + "content": "[3] Vassileios Balntas, Karel Lenc, Andrea Vedaldi, and Krystian Mikolajczyk. HPatches: A benchmark and evaluation of handcrafted and learned local descriptors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5173-5182, 2017. 2, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.346, + 0.483, + 0.416 + ], + "angle": 0, + "content": "[4] Axel Barroso-Laguna, Sowmya Munukutla, Victor Adrian Prisacariu, and Eric Brachmann. Matching 2D images in 3D: Metric relative pose from metric correspondences. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4852-4863, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.418, + 0.482, + 0.46 + ], + "angle": 0, + "content": "[5] Herbert Bay, Tinne Tuytelaars, and Luc Van Gool. SURF: Speeded up robust features. In European Conference on Computer Vision, pages 404-417. Springer, 2006. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.461, + 0.482, + 0.53 + ], + "angle": 0, + "content": "[6] Ruojin Cai, Joseph Tung, Qianqian Wang, Hadar Averbuch-Elor, Bharath Hariharan, and Noah Snavely. Doppelgangers: Learning to disambiguate images of similar structures. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 34–44, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.532, + 0.482, + 0.601 + ], + "angle": 0, + "content": "[7] Hongkai Chen, Zixin Luo, Jiahui Zhang, Lei Zhou, Xuyang Bai, Zeyu Hu, Chiew-Lan Tai, and Long Quan. Learning to match features with seeded graph matching network. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6301-6310, 2021. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.603, + 0.482, + 0.673 + ], + "angle": 0, + "content": "[8] Hongkai Chen, Zixin Luo, Lei Zhou, Yurun Tian, Mingmin Zhen, Tian Fang, David Mckinnon, Yanghai Tsin, and Long Quan. ASpanFormer: Detector-free image matching with adaptive span transformer. In European Conference on Computer Vision, pages 20-36. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.674, + 0.482, + 0.715 + ], + "angle": 0, + "content": "[9] Ondrej Chum, Jiří Matas, and Josef Kittler. Locally optimized RANSAC. In Pattern Recognition, pages 236-243. Springer, 2003. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.717, + 0.482, + 0.786 + ], + "angle": 0, + "content": "[10] Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. SuperPoint: Self-supervised interest point detection and description. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 224-236, 2018. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.788, + 0.482, + 0.857 + ], + "angle": 0, + "content": "[11] Carl Doersch, Ankush Gupta, Larisa Markeeeva, Adria Recasens, Lucas Smaira, Yusuf Aytar, Joao Carreira, Andrew Zisserman, and Yi Yang. TAP-Vid: A benchmark for tracking any point in a video. Advances in Neural Information Processing Systems, 35:13610-13626, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.859, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[12] Carl Doersch, Yi Yang, Mel Vecerik, Dilara Gokay, Ankush Gupta, Yusuf Aytar, Joao Carreira, and Andrew Zisserman. TAPIR: Tracking any point with per-frame initialization and" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.147, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "temporal refinement. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10061-10072, 2023. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.136, + 0.905, + 0.218 + ], + "angle": 0, + "content": "[13] Mihai Dusmanu, Ignacio Rocco, Tomas Pajdla, Marc Pollefeys, Josef Sivic, Akihiko Torii, and Torsten Sattler. D2-Net: A trainable cnn for joint description and detection of local features. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8092–8101, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.221, + 0.905, + 0.289 + ], + "angle": 0, + "content": "[14] Johan Edstedt, Georg Bokman, Marten Wadenback, and Michael Felsberg. DeDoDe: Detect, don't describe—Describe, don't detect for local feature matching. In International Conference on 3D Vision, pages 148-157. IEEE, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.291, + 0.905, + 0.359 + ], + "angle": 0, + "content": "[15] Johan Edstedt, Qiyu Sun, Georg Bokman, Marten Wadenback, and Michael Felsberg. RoMa: Robust dense feature matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19790-19800, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.362, + 0.904, + 0.403 + ], + "angle": 0, + "content": "[16] Maha Elbayad, Jiatao Gu, Edouard Grave, and Michael Auli. Depth-adaptive Transformer. In International Conference on Learning Representations, pages 1-14, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.405, + 0.905, + 0.46 + ], + "angle": 0, + "content": "[17] Martin A Fischler and Robert C Bolles. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. In Communications of the ACM, 24(6):381-395, 1981. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.462, + 0.905, + 0.532 + ], + "angle": 0, + "content": "[18] Jan-Michael Frahm, Pierre Fite-Georgel, David Gallup, Tim Johnson, Rahul Raguram, Changchang Wu, Yi-Hung Jen, Enrique Dunn, Brian Clipp, Svetlana Lazebnik, et al. Building Rome on a cloudless day. In European Conference on Computer Vision, pages 368-381. Springer, 2010. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.533, + 0.905, + 0.588 + ], + "angle": 0, + "content": "[19] Adam W Harley, Zhaoyuan Fang, and Katerina Fragkiadaki. Particle video revisited: Tracking through occlusions using point trajectories. In European Conference on Computer Vision, pages 59-75. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.59, + 0.905, + 0.63 + ], + "angle": 0, + "content": "[20] Richard Hartley and Andrew Zisserman. Multiple view geometry in computer vision. Cambridge university press, 2003. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.633, + 0.905, + 0.702 + ], + "angle": 0, + "content": "[21] Xingyi He, Jiaming Sun, Yifan Wang, Sida Peng, Qixing Huang, Hujun Bao, and Xiaowei Zhou. Detector-free structure-from-motion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21594-21603, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.704, + 0.905, + 0.773 + ], + "angle": 0, + "content": "[22] Martin Humenberger, Yohann Cabon, Nicolas Guerin, Julien Morat, Vincent Leroy, Jérôme Revaud, Philippe Rerole, Noé Pion, Cesar de Souza, and Gabriela Csurka. Robust image retrieval-based visual localization using kapture. arXiv preprint arXiv:2007.13867, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.775, + 0.905, + 0.83 + ], + "angle": 0, + "content": "[23] Heinly Jared, Johannes L Schonberger, Enrique Dunn, and Jan-Michael Frahm. Reconstructing the world in six days. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.832, + 0.907, + 0.901 + ], + "angle": 0, + "content": "[24] Hanwen Jiang, Arjun Karpur, Bingyi Cao, Qixing Huang, and André Araujo. OmniGlue: Generalizable feature matching with foundation model guidance. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19865-19875, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.956 + ], + "angle": 0, + "content": "21978" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.161 + ], + "angle": 0, + "content": "[25] Yuhe Jin, Dmytro Mishkin, Anastasiia Mishchuk, Jiri Matas, Pascal Fua, Kwang Moo Yi, and Eduard Trulls. Image matching across wide baselines: From paper to practice. International Journal of Computer Vision, 129(2):517-547, 2021. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.164, + 0.482, + 0.232 + ], + "angle": 0, + "content": "[26] Nikita Karaev, Ignacio Rocco, Benjamin Graham, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. CoTracker: It is better to track together. In European Conference on Computer Vision, pages 18-35. Springer, 2024. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.234, + 0.482, + 0.288 + ], + "angle": 0, + "content": "[27] Vincent Leroy, Yohann Cabon, and Jérôme Revaud. Grounding image matching in 3D with MAST3R. In European Conference on Computer Vision, pages 71-91. Springer, 2024. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.291, + 0.482, + 0.346 + ], + "angle": 0, + "content": "[28] Zhengqi Li and Noah Snavely. MegaDepth: Learning single-view depth prediction from internet photos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2041-2050, 2018. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.348, + 0.482, + 0.417 + ], + "angle": 0, + "content": "[29] Philipp Lindenberger, Paul-Edouard Sarlin, Viktor Larsson, and Marc Pollefeys. Pixel-perfect structure-from-motion with featuremetric refinement. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5987-5997, 2021. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.419, + 0.482, + 0.475 + ], + "angle": 0, + "content": "[30] Philipp Lindenberger, Paul-Edouard Sarlin, and Marc Pollefeys. LightGlue: Local feature matching at light speed. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17627–17638, 2023. 1, 2, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.476, + 0.482, + 0.517 + ], + "angle": 0, + "content": "[31] David G Lowe. Distinctive image features from scale-invariant keypoints. International Journal of Computer Vision, 60:91-110, 2004. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.519, + 0.482, + 0.573 + ], + "angle": 0, + "content": "[32] Jiayi Ma, Xingyu Jiang, Aoxiang Fan, Junjun Jiang, and Junchi Yan. Image matching from handcrafted to deep features: A survey. International Journal of Computer Vision, 129(1):23-79, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.575, + 0.482, + 0.63 + ], + "angle": 0, + "content": "[33] Eleonora Maset, Federica Arrigoni, and Andrea Fusiello. Practical and efficient multi-view matching. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4568-4576, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.632, + 0.482, + 0.687 + ], + "angle": 0, + "content": "[34] Raul Mur-Artal, Jose Maria Martinez Montiel, and Juan D Tardos. ORB-SLAM: A versatile and accurate monocular SLAM system. In IEEE Transactions on Robotics, 31(5): 1147-1163, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.689, + 0.482, + 0.731 + ], + "angle": 0, + "content": "[35] Carl Olsson and Olof Enqvist. Stable structure-from-motion for unordered image collections. In Image Analysis, pages 524-535. Springer, 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.732, + 0.482, + 0.773 + ], + "angle": 0, + "content": "[36] Linfei Pan, Daniel Barath, Marc Pollefeys, and Johannes L Schonberger. Global structure-from-motion revisited. In European Conference on Computer Vision, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.775, + 0.482, + 0.844 + ], + "angle": 0, + "content": "[37] Filip Radenović, Ahmet Iscen, Giorgos Tolias, Yannis Avrithis, and Ondrej Chum. Revisiting Oxford and Paris: Large-scale image retrieval benchmarking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5706-5715, 2018. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[38] Filip Radenovic, Giorgos Tolias, and Ondrej Chum. Finetuning CNN image retrieval with no human annotation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 41(7):1655-1668, 2018. 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[39] Jerome Revaud, Cesar De Souza, Martin Humenberger, and Philippe Weinzaepfel. R2D2: Reliable and repeatable detector and descriptor. Advances in Neural Information Processing Systems, 32, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.149, + 0.905, + 0.204 + ], + "angle": 0, + "content": "[40] Barbara Roessle and Matthias Nießner. End2end multi-view feature matching with differentiable pose optimization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 477-487, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.205, + 0.905, + 0.26 + ], + "angle": 0, + "content": "[41] Ethan Rublee, Vincent Rabaud, Kurt Konolige, and Gary Bradski. ORB: An efficient alternative to SIFT or SURF. In Proceeding of the IEEE/CVF International Conference on Computer Vision, pages 2564-2571. IEEE, 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.261, + 0.905, + 0.328 + ], + "angle": 0, + "content": "[42] Paul-Edouard Sarlin, Cesar Cadena, Roland Siegwart, and Marcin Dymczyk. From coarse to fine: Robust hierarchical localization at large scale. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12716–12725, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.33, + 0.907, + 0.399 + ], + "angle": 0, + "content": "[43] Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. SuperGlue: Learning feature matching with graph neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4938–4947, 2020. 1, 2, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.4, + 0.905, + 0.482 + ], + "angle": 0, + "content": "[44] Torsten Sattler, Akihiko Torii, Josef Sivic, Marc Pollefeys, Hajime Taira, Masatoshi Okutomi, and Tomas Pajdla. Are large-scale 3d models really necessary for accurate visual localization? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1637-1646, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.484, + 0.905, + 0.538 + ], + "angle": 0, + "content": "[45] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-Motion revisited. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4104-4113, 2016. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.539, + 0.905, + 0.607 + ], + "angle": 0, + "content": "[46] Johannes L Schonberger, Hans Hardmeier, Torsten Sattler, and Marc Pollefeys. Comparative evaluation of hand-crafted and learned local features. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1482-1491, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.609, + 0.905, + 0.663 + ], + "angle": 0, + "content": "[47] Tal Schuster, Adam Fisch, Jai Gupta, Mostafa Dehghani, Dara Bahri, Vinh Tran, Yi Tay, and Donald Metzler. Confident adaptive language modeling. Advances in Neural Information Processing Systems, 35:17456-17472, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.665, + 0.905, + 0.719 + ], + "angle": 0, + "content": "[48] Sivic and Zisserman. Video Google: A text retrieval approach to object matching in videos. In Proceedings in the IEEE/CVF International Conference on Computer Vision, pages 1470-1477. IEEE, 2003. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.721, + 0.905, + 0.773 + ], + "angle": 0, + "content": "[49] Jianlin Su, Murtadha Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu. RoFormer: Enhanced transformer with rotary position embedding. Neurocomputing, 568:127063, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.776, + 0.905, + 0.844 + ], + "angle": 0, + "content": "[50] Jiaming Sun, Zehong Shen, Yuang Wang, Hujun Bao, and Xiaowei Zhou. LoFTR: Detector-free local feature matching with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8922-8931, 2021. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[51] Kun Sun, Jinhong Yu, Wenbing Tao, Xin Li, Chang Tang, and Yuhua Qian. A unified feature-spatial cycle consistency fusion framework for robust image matching. Information Fusion, 97:101810, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.907, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "21979" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.161 + ], + "angle": 0, + "content": "[52] Prune Truong, Martin Danelljan, Luc Van Gool, and Radu Timofte. Learning accurate dense correspondences and when to trust them. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5714-5724, 2021. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.163, + 0.482, + 0.218 + ], + "angle": 0, + "content": "[53] Michal Tyszkiiewicz, Pascal Fua, and Eduard Trulls. DISK: Learning local features with policy gradient. Advances in Neural Information Processing Systems, 33:14254-14265, 2020. 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.22, + 0.482, + 0.275 + ], + "angle": 0, + "content": "[54] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in Neural Information Processing Systems, 30, 2017. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.277, + 0.482, + 0.344 + ], + "angle": 0, + "content": "[55] Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. DUSt3R: Geometric 3D vision made easy. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20697-20709, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.346, + 0.482, + 0.415 + ], + "angle": 0, + "content": "[56] Yifan Wang, Xingyi He, Sida Peng, Dongli Tan, and Xiaowei Zhou. Efficient LoFTR: Semi-dense local feature matching with sparse-like speed. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21666-21675, 2024. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.418, + 0.482, + 0.486 + ], + "angle": 0, + "content": "[57] Yuxi Xiao, Qianqian Wang, Shangzhan Zhang, Nan Xue, Sida Peng, Yujun Shen, and Xiaowei Zhou. SpatialTracker: Tracking any 2D pixels in 3D space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20406-20417, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.488, + 0.482, + 0.542 + ], + "angle": 0, + "content": "[58] Shibiao Xu, Shunpeng Chen, Rongtao Xu, Changwei Wang, Peng Lu, and Li Guo. Local feature matching using deep learning: A survey. Information Fusion, 107:102344, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.544, + 0.482, + 0.598 + ], + "angle": 0, + "content": "[59] Kwang Moo Yi, Eduard Trulls, Vincent Lepetit, and Pascal Fua. LIFT: Learned invariant feature transform. In European Conference on Computer Vision, pages 467-483. Springer, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.601, + 0.482, + 0.657 + ], + "angle": 0, + "content": "[60] Guofeng Zhang, Zilong Dong, Jiaya Jia, Tien-Tsin Wong, and Hujun Bao. Efficient non-consecutive feature tracking for structure-from-Motion. In European Conference on Computer Vision, pages 422-435. Springer, 2010. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.657 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "21980" + } + ] +] \ No newline at end of file diff --git a/2025/CoMatcher_ Multi-View Collaborative Feature Matching/b3062684-7794-47f2-b9d1-53d25767d94f_origin.pdf b/2025/CoMatcher_ Multi-View Collaborative Feature Matching/b3062684-7794-47f2-b9d1-53d25767d94f_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..4c94da27a96cc6eedf46253da0ed8e765e757ba8 --- /dev/null +++ b/2025/CoMatcher_ Multi-View Collaborative Feature Matching/b3062684-7794-47f2-b9d1-53d25767d94f_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:85203a4e0ac32b23b4edf059bd31fe51d0bd43223649ff5927b3885bce4b6f64 +size 5882872 diff --git a/2025/CoMatcher_ Multi-View Collaborative Feature Matching/full.md b/2025/CoMatcher_ Multi-View Collaborative Feature Matching/full.md new file mode 100644 index 0000000000000000000000000000000000000000..8508838269591e983c269b7e4d765bd4fc0c40e5 --- /dev/null +++ b/2025/CoMatcher_ Multi-View Collaborative Feature Matching/full.md @@ -0,0 +1,302 @@ +# CoMatcher: Multi-View Collaborative Feature Matching + +Jintao Zhang $^{1}$ , Zimin Xia $^{2*}$ , Mingyue Dong $^{1}$ , Shuhan Shen $^{3}$ , Linwei Yue $^{4}$ , Xianwei Zheng $^{1*}$ $^{1}$ The State Key Lab. LIESMARS, Wuhan University + $^{2}$ École Polytechnique Fédérale de Lausanne (EPFL) + $^{3}$ Chinese Academy of Sciences ${ }^{4}$ China University of Geosciences + +# Abstract + +This paper proposes a multi-view collaborative matching strategy for reliable track construction in complex scenarios. We observe that the pairwise matching paradigms applied to image set matching often result in ambiguous estimation when the selected independent pairs exhibit significant occlusions or extreme viewpoint changes. This challenge primarily stems from the inherent uncertainty in interpreting intricate 3D structures based on limited two-view observations, as the 3D-to-2D projection leads to significant information loss. To address this, we introduce CoMatcher, a deep multi-view matcher to (i) leverage complementary context cues from different views to form a holistic 3D scene understanding and (ii) utilize cross-view projection consistency to infer a reliable global solution. Building on CoMatcher, we develop a groupwise framework that fully exploits cross-view relationships for large-scale matching tasks. Extensive experiments on various complex scenarios demonstrate the superiority of our method over the mainstream two-view matching paradigm. + +# 1. Introduction + +Research on recovering 3D structures and camera poses from multiple views storing scene information has a long history [20, 23, 34]. A core component of this process is feature matching, which aims to estimate point correspondences across the image set [25, 32, 58]. For computational flexibility, existing methods often decompose the set into pairs of co-visible images [1, 18] and apply a two-view matching approach to each pair independently [31, 43]. The pairwise matches are then merged into comprehensive multi-view tracks [1, 45]. Following this pairwise framework to build systems for localization [22, 42] and mapping [29, 45] has become a entrenched research paradigm. + +Recent advances in learning-based two-view matchers have greatly improved performance in most cases [15, 30, + +![](images/e18239df716da0f8358979bd3171e64aed9fd5b1e52c6cccbddd3a248a20e2a0.jpg) +Figure 1. From two view to multi-view. Unlike pairwise schemes prone to uncertainty, we first partition the image set into co-visible groups and collaboratively estimate group-wide correspondences using a reliable deep multi-view matcher. Benefiting from holistic scene understanding and consistency constraint, our method generates high-quality tracks in complex scenarios. + +43, 50]. Inspired by Transformer [54], these methods employ deep networks to jointly analyze global context from both views. This enables models to learn to infer underlying 3D scenes by integrating spatial and visual cues, which has been proven crucial for accurate estimations [27, 43]. + +Despite great progress, current two-view matchers still struggle in challenging wide-baseline scenarios, particularly with severe occlusions and repetitive textures [25]. We argue that one potential cause for this difficulty lies in the inherent uncertainty of interpreting complex 3D structures from limited two-view observations—a task that even humans find prone to perceptual ambiguity. Projecting complex 3D geometries onto 2D planes inevitably results in significant loss of scene information. For example, spatially distant regions may appear close in 2D. This makes it highly unreliable to infer the original scene solely from ambiguous two-view context [6, 57]. Moreover, while matchers can learn two-view geometric priors, this alone is also in + +adequate—capturing abrupt depth discontinuities requires stronger constraints [4, 27]. Errors in pairwise correspondences compound during merging, significantly impacting downstream tasks, especially in large-scale datasets. + +To address these challenges, we propose a shift in focus: instead of optimizing two-view matchers, a more effective strategy is to directly leverage the rich relationships inherent in raw multi-view observations. Building on this, we introduce a unified formulation that collaboratively establishes correspondences between a complementary view group and a target view. This approach fosters holistic scene understanding by integrating contextual information from multiple views, while multi-view geometric priors provide stronger constraints for reliable reasoning. Moreover, matches across views naturally adhere to cross-view projection consistency. This facilitates a mutual verification process, ensuring confident inference of a global solution. + +Inspired by existing two-view methods [43, 50], we develop CoMatcher, a deep-learning architecture for multiview feature matching. Context sharing is achieved through a GNN with multi-view receptive fields. While a larger feature space enriches context, it also introduces noise, complicating learning. To address this, we propose to constrain the search space for each point by leveraging cross-view projection geometry. Furthermore, we redesign the feature correlation layers in cross-attention to enable consistent multiview reasoning. This design exploits the property that correlation scores encode both reasoning outcomes and confidence, enabling progressive layer-by-layer integration of multi-view estimates into a globally consistent solution. + +To scale CoMatcher for large-scale tasks like SfM [36, 45], we develop a groupwise matching pipeline (see Fig. 1). It introduces a novel set partitioning algorithm that uses可视ability information to group images, followed by group-level matching using CoMatcher. By preserving cross-view relationships, our framework achieves comprehensive improvements over pairwise methods in extensive experiments on wide-baseline scenarios. The main contributions are summarized as follows: (i) A novel multi-view matching paradigm, comprising a deep collaborative matcher and a scalable groupwise pipeline, designed to overcome the inherent uncertainty of pairwise approaches. (ii) A multiview representation learning network architecture that enables holistic scene understanding. (iii) A multi-view feature correlation strategy that integrates uncertain individual estimates into a globally consistent solution. + +# 2. Related work + +Two-view matching is a fundamental problem in computer vision. The most prevalent paradigm involves sparsifying images into keypoints, each represented by a high-dimensional vector encoding its visual context [25, 32]. This transforms matching into a search problem, with match + +likelihood expressed by the inner product of corresponding vectors [31]. Modern methods leverage deep networks for representation learning to extract feature points [10, 13, 14, 39, 53, 59], outperforming traditional hand-crafted encodings [5, 31, 41] and achieving significant advancements on mainstream benchmarks [3, 25, 46]. In parallel, other works aim to address the limitations of heuristic matching strategies by leveraging networks with global modeling capabilities to directly learn the matching process [7, 24, 30, 43]. Such data-driven methods can effectively capture priors of the underlying 3D scene, which are essential for accurate estimation. Unlike keypoint-based approaches, another typical method retains all pixels for dense inference [8, 15, 27, 50, 52]. While this has led to notable progress in weakly textured regions, such methods require greater computational resources [30, 56] and struggle with multi-view inconsistency [21]. In summary, while these methods perform well in most scenarios, they remain constrained in complex wide-baseline settings due to the limited receptive field inherent to two-view systems. + +Multi-view matching focuses on establishing correspondences across a larger-scale image collection. A typical pairwise matching approach decomposes the problem into two-view matching instances and subsequently merges the pairwise results using multi-view projection consistency [1, 42, 44, 45]. During merging, some works explore how consistency assumptions can filter and refine coarse matches [33, 51], while others focus on efficiency, developing better strategies to reduce redundant matching computations [35, 60]. However, as a post-processing step, merging cannot address catastrophic failures that occur during the matching stage. Another line of work treats matching as a point tracking problem [11, 12, 19, 26], using optical flow-like frameworks to reason about the positions of query pixels across multiple images. While this approach is well-suited for video inputs, its effectiveness on unstructured, wide-baseline image collections remains largely unverified. The most relevant method to ours is [40], which performs end-to-end multi-view matching and pose estimation. However, it is limited in scalability, as it can only handle a small image group. In contrast, our framework is highly flexible, with the potential to scale to an arbitrary number of images. + +# 3. Methodology + +**Formulation.** Given a set of $N_I$ images, $\mathcal{I} = \{I_i \mid i = 1, \dots, N_I\}$ , of a scene, the objective of feature matching is to identify the 2D position coordinates $\mathcal{M}_k = \{\mathbf{p}^{I_i}\}$ in $\mathcal{I}$ that correspond to identical 3D scene points $\mathcal{X} = \{\mathbf{x}_k \mid k = 1, \dots, N_X\}$ , where $\mathbf{p}^{I_i} \in \mathbb{R}^2$ are the pixel coordinates in $I_i$ , and each $\mathcal{M}_k$ is called a track. + +Motivation. Constructing complete tracks requires estimating correspondences between each image $I_{t} \subset \mathcal{I}$ and its co-visible image set $\mathcal{C}_t$ . While the pairwise approach of + +![](images/b8d28b5cf08c3d45e508810fe0943624b78ba3655e92c41e14199624d5892134.jpg) +Figure 2. The CoMatcher architecture can be viewed as an extension of parallel two-view matchers, enhanced with two core components: the multi-view (M-V) feature interaction module (Sec. 3.2) and the multi-view (M-V) feature correlation strategy (Sec. 3.3). The first component employs multi-view cross-attention to enhance the representation of all source and target view features, while leveraging the cross-view projection geometry from $\mathcal{M}(\mathcal{G})$ to constrain the point search space. The second component aggregates correlation scores from multiple two-view cross-attention modules based on estimated confidence, guiding the network to achieve multi-view consistent reasoning. + +fers implementation flexibility for this problem, its reliance on two-view inputs in a single estimation often leads to uncertainty in complex scenes. In practice, images within $\mathcal{C}_t$ are often not independent—most are captured from adjacent positions but with differing viewing angles. This inspires us to group $\mathcal{C}_t$ based on image correlations, where each complementary group offers a mapping from multiple planar views to a 3D sub-scene. In the image matching problem, the correspondence between $I_t$ and a group must adhere to certain 2D-to-3D physical constraints, such as cross-view projection consistency. Rather than independently estimating and then integrating them with constraints, it is more efficient to unify this process and learn underlying priors directly from the data. Thus, we formulate CoMatcher to address the 1-to- $N$ matching problem (see Fig. 2). + +Groupwise matching pipeline. Based on the above, we propose a three-step pipeline for constructing tracks: grouping, connecting, and matching. First, we treat the original set of views $\mathcal{I}$ as a union of smaller image groups $\{\mathcal{G}_s\mid s = 1,\dots ,N_G\}$ , each representing a localized scene. This involves a pre-processing step to partition the set. The most critical aspect of this is determining the correlation between two images, i.e., whether they describe the same structure. Building on existing research [2, 38, 48], we employ heuristic techniques to design a grouping algorithm. Next, although CoMatcher can directly learn multi-view geometry, certain fine-grained physical rules are challenging to model. Therefore, we explicitly establish relationships within each group to guide subsequent matching. For each group, we construct tracks using existing frameworks, which provide a set of 3D points along with their multi-view projections and implicitly encode camera poses. These cues + +are incorporated as inputs to the network to steer the inference process. Finally, each group of images is treated as a whole to collaboratively match with other images in $\mathcal{I}$ using CoMatcher. Additional details on grouping and the overall framework are provided in the supplementary materials. + +# 3.1. CoMatcher overview + +As shown in Fig. 2, CoMatcher, as a deep sparse matcher, estimates point correspondences between each source view in a group $I_{i} \in \mathcal{G}$ and a target view $I_{t}$ . The network takes as input the extracted local features $\{\left(\mathbf{p}_{k}^{I_{i}}, \mathbf{d}_{k}^{I_{i}}\right) \mid k = 1, \dots, N_{F}\}$ from the $M$ source views and the target view, along with the precomputed tracks $\mathcal{M}(\mathcal{G})$ of $\mathcal{G}$ . Each local feature consists of keypoint locations $\mathbf{p}_{k}$ and their corresponding visual descriptors $\mathbf{d}_{k}$ . Initially, the target view features are broadcast and paired with those of each source view to establish the initial states. These $M$ feature pairs are then refined using a Graph Neural Network (GNN) with alternating self-attention and cross-attention modules, repeated $L$ times. Finally, the enhanced features are fed into matching heads to infer the final correspondences. + +The design of CoMatcher revolves around two core questions: (i) how can multi-view context be leveraged to learn better point representations (Sec. 3.2), and (ii) how can the estimations across multiple views be constrained to satisfy cross-view consistency (Sec. 3.3)? After thoroughly discussing these questions, we introduce the prediction heads (Sec. 3.4) and the loss function (Sec. 3.5). + +# 3.2. Multi-view feature interaction + +Consider an occluded scene. When observing such structures from a frontal view, points near occlusion boundaries + +are often contaminated by irrelevant context. This noise significantly compromises the quality of point features, creating substantial challenges for matching. Multi-view learning provides a promising solution to this issue. For these ambiguous point features, it can integrate observations of the same area from multiple other viewpoints to refine their representations. This involves a querying process, which establishes a multi-view receptive field for each point. + +Given a set of co-visible source views, CoMatcher leverages a multi-view cross-attention mechanism to aggregate these context (Source Cross). For a point $u$ in source view $I_{i}$ , we attend to all points in the other source views: + +$$ +\mathbf {m} _ {u} ^ {I _ {i} \leftarrow \mathcal {W} _ {j}} = \frac {1}{M - 1} \sum_ {I _ {j} \in \mathcal {G} \backslash I _ {i}} \sum_ {v \in \mathcal {W} _ {j}} \operatorname {S o f t m a x} \left(a _ {u v} ^ {I _ {i} I _ {j}}\right) \mathbf {v} _ {v} ^ {I _ {j}}. \tag {1} +$$ + +Here, $\mathcal{W}_j$ denotes the point set in the source view $I_{j}$ , while $a_{uv}^{I_iI_j}$ represents the similarity score computed by treating point $u$ in $I_{i}$ as the query and point $v$ in $I_{j}$ as the key. $\mathbf{v}_v^{I_j}$ is the value of point $v$ . $\mathbf{m}_u^{I_i}$ represents the message vector from $\mathcal{W}_j$ to $u$ of the GNN [43]. We uniformly aggregate features from each source view, encouraging the model to integrate multi-view information comprehensively rather than overemphasizing images from similar perspectives. + +However, for a point near an occlusion boundary in one source view, how does it identify the corresponding points observing the same location in other source views? This implies a matching process. While the network could directly learn this, it would introduce additional complexity, as establishing correspondences between source views is not our primary objective. Moreover, while multi-view setups provide abundant contextual information, much of it is irrelevant, such as regions lacking co-visibility. This redundancy also imposes an additional burden on learning. + +To address this, we propose a geometric constraint mechanism to explicitly guide the attention range of each point. By incorporating relative positional encoding into the attention score computation, we enable the scores to depend on two critical factors: feature correlation and the geometric distance $\Delta \mathbf{p}$ between two points: + +$$ +a _ {u v} ^ {I _ {i} I _ {j}} = \left(\mathbf {q} _ {u} ^ {I _ {i}}\right) ^ {\top} \mathbf {R} \left[ \Delta \mathbf {p} _ {u v} ^ {I _ {i} I _ {j}} \right] \mathbf {k} _ {v} ^ {I _ {j}}. \tag {2} +$$ + +Here, $\mathbf{R}[\cdot ]$ is the rotation encoding matrix [49]. + +While the 2D relative distance between two points within a single image is well-defined, across different views, it depends on the projection transformation. We use the precomputed tracks $\mathcal{M}(\mathcal{G})$ to find the projection position of $\mathbf{p}_u^{I_i}$ on $I_{j}$ , which can be denoted as $\mathbf{p}_w^{I_j}$ . When no corresponding point in $I_{j}$ can be found for $\mathbf{p}_u^{I_i}$ , we simply assign the relative position to $\mathbf{0}$ so that the attention score depends solely on the feature similarity. The calculation of the relative po + +sition can be formulated as follows: + +$$ +\Delta \mathbf {p} _ {u v} ^ {I _ {i} I _ {j}} = \left\{ \begin{array}{l l} \mathbf {p} _ {w} ^ {I _ {j}} - \mathbf {p} _ {v} ^ {I _ {j}} & (u, w) \in M (I _ {i}, I _ {j}) \\ \mathbf {0} & (u, \emptyset) \in M (I _ {i}, I _ {j}). \end{array} \right. \tag {3} +$$ + +Additionally, we also interact with all target view features after source cross-attention (Target Cross). For a point $t$ in view $I_{t}$ paired with source view $I_{i}$ , we attend to the same point $t$ across the remaining $M - 1$ target views. This is analogous to operations in some works that aggregate features along the temporal dimension [12, 26]. Unlike the approach for source views, we compute the attention scores solely based on feature similarity. Despite its low computational complexity, this module complements the Source Cross, mutually enhancing each other to help the network implicitly infer globally optimal correspondences. + +# 3.3. Multi-view feature correlation + +For a track in a source view set originating from a 3D point, its corresponding 2D keypoints $\left\{\mathbf{p}_{u}^{I_{i}}, \mathbf{p}_{v}^{I_{j}}, \ldots\right\}$ should exhibit consistent matches in the target view—either corresponding to the same single point or having no corresponding point at all. This physical constraint is often utilized by existing methods as a post-processing step to filter out erroneous results when merging multiple two-view correspondences. However, as a multi-view matcher, we aim for CoMatcher to inherently satisfy this constraint in its reasoning, thereby enhancing its reliability and confidence. + +Achieving this involves two key processes: (i) identifying points that may lead to errors in the early layers of the network, and (ii) promptly guiding these points toward correct estimation using inference information from other views. To accomplish this, we propose a two-step multiview feature correlation strategy (see Fig. 3). + +First, to identify ambiguous points in the source views, we utilize a lightweight head at the end of each layer to predict the confidence of each point: + +$$ +c _ {u} ^ {I _ {i}} = \operatorname {S i g m o i d} \left(\operatorname {M L P} \left(\mathbf {f} _ {u} ^ {I _ {i}}\right)\right) \in [ 0, 1 ]. \tag {4} +$$ + +Here, MLP represents a multi-layer perceptron and $\mathbf{f}_u^{I_i}$ is the feature of $u$ . This confidence score reflects the assessment of each point in its current state: higher values indicate that, at this layer, the point is either converging toward a reliable match or confidently identified as having no match. + +In each layer, we define a distinct hyperparameter threshold $\theta$ . Points with confidence scores below this threshold are identified as potentially ambiguous and likely to lead to incorrect inference. We progressively increase the value of $\theta$ in later layers, as the confidence of overall inference generally improves with each successive layer. + +Next, we redesign the parallel two-view cross-attention modules to correct the erroneous estimations of ambiguous + +![](images/96f4098fe3dbd88256aec1c84d929a9546d018e9a5100a1ccf9d62bdd6564e0f.jpg) +Figure 3. The multi-view feature correlation strategy consists of two steps: we first identify potentially erroneous points using a confidence estimator, and then correct their estimation by aggregating attention correlation vectors from other views. + +points. In these modules, each point in the source view computes an attention distribution vector to query [54]: + +$$ +\boldsymbol {\alpha} _ {u} ^ {I _ {i}} = \underset {x \in \mathcal {W} _ {t}} {\operatorname {S o f t m a x}} \left(a _ {u x} ^ {I _ {i} I _ {t}}\right). \tag {5} +$$ + +This vector represents the degree to which the point attends to in view $I_{t}$ , with a peak indicating potential matching regions of interest. For ambiguous points, this similarity is not reliable, often leading to incorrect feature aggregation. Assuming that point $u$ in the source view $I_{i}$ is an ambiguous point, we update its original attention distribution by applying a weighted average to the distributions of its corresponding points across multi-view: + +$$ +\boldsymbol {\alpha} _ {u} ^ {I _ {i} ^ {\prime}} = c _ {u} ^ {I _ {i}} \boldsymbol {\alpha} _ {u} ^ {I _ {i}} + \left(1 - c _ {u} ^ {I _ {i}}\right) \frac {\sum_ {v \in \mathcal {D} _ {m} ^ {u}} c _ {v} ^ {I _ {j}} \boldsymbol {\alpha} _ {v} ^ {I _ {j}}}{\sum_ {v \in \mathcal {D} _ {m} ^ {u}} c _ {v} ^ {I _ {j}}} \quad \text {s . t .} c _ {u} ^ {I _ {i}} < \theta . \tag {6} +$$ + +Here, $\mathcal{D}_m^u$ represents other points in the track. We embed this strategy into the model inference process, which significantly enhances the reliability of estimation under challenging conditions despite its simple implementation. + +# 3.4. Correspondence prediction + +As shown in Fig. 2, a lightweight matching head operates in parallel for each pair of network output features to predict correspondences between two views [30, 43]. In each head, we begin by calculating a score matrix from feature correlations between two views: $\mathbf{S}(u,x) = \mathrm{L}(\mathbf{f}_u^{I_i})\cdot$ $\mathrm{L}(\mathbf{f}_x^{I_t})$ , where $\mathrm{L}$ is the linear transform. Then we apply dual-softmax [30, 50, 53] to derive the matching probabilities of two points: $\mathbf{S}'(u,x) = \mathrm{Softmax}(\mathbf{S}(u,\cdot))_x\cdot$ Softmax $(\mathbf{S}(\cdot ,x))_u$ . Additionally, we employ the method proposed in [30] to calculate the matching probability for each point: $\sigma_{i}^{I_{i}} = \mathrm{Sigmoid}(\mathrm{L}(\mathbf{f}_{i}^{I_{i}}))$ . The final assignment matrix is a combination of the two probabilities: + +$$ +\mathbf {P} (u, x) = \mathbf {S} ^ {\prime} (u, x) \sigma_ {u} ^ {I _ {i}} \sigma_ {x} ^ {I _ {t}}. \tag {7} +$$ + +Using mutual nearest neighbors and a threshold, we filter the matching results from the assignment matrix [43, 50]. + +# 3.5. Loss + +CoMatcher supervises each source-target image pair individually, with the loss for each pairs composed of two components: correspondence and confidence: + +$$ +\mathcal {L} _ {\text {t o t a l}} = \frac {1}{M} \sum_ {I _ {i} \in \mathcal {G}} \left(\mathcal {L} _ {\text {c o r r}} \left(I _ {i}, I _ {t}\right) + \alpha \mathcal {L} _ {\text {c o n f}} \left(I _ {i}\right)\right). \tag {8} +$$ + +Correspondence supervision. For each pair $I_{i}$ and $I_{t}$ , we perform two-view transformations using relative poses or homography to compute ground truth labels [30, 43, 50]. We minimize the negative log-likelihood of the assignment matrix, following [30]. More details are provided in the supplementary materials. + +Confidence supervision. To train the confidence estimators at each layer, we define the point confidence as the consistency probability between its correspondence estimated at the current layer and the final estimation [16, 30, 47]. At each layer, predictions are obtained via dual-softmax on current features, without introducing additional matching probabilities or parameters. The ground truth label $y$ indicates whether the two estimations are consistent. This is supervised using a binary cross-entropy loss: + +$$ +\mathcal {L} _ {\text {c o n f}} \left(I _ {i}\right) = \frac {1}{L - 1} \sum_ {\ell} \sum_ {u \in \mathcal {W} _ {i}} \mathrm {C E} \left(\ell c _ {u} ^ {I _ {i}}, \ell y _ {u} ^ {I _ {i}}\right), \tag {9} +$$ + +where $\ell \in \{1,\dots ,L - 1\}$ + +# 4. Experiments + +In this section, we first introduce the datasets used, followed by our implementation details. Then, our CoMatcher network is compared to previous state-of-the-art baselines for homography and camera pose estimation. Next, we integrate CoMatcher into groupwise matching framework and evaluate it against pairwise method on a large-scale benchmark. Finally, an extensive ablation study is provided. + +# 4.1. Datasets + +The HPatches dataset is used for homography estimation [3]. It consists of 116 strictly planar scenes, each containing 6 images with variations in viewpoint and illumination. For camera pose estimation tasks, we utilize the MegaDepth dataset, an outdoor dataset that exhibits strong occlusions and significant structural changes [28]. We selected two scenes, "Sacre Coeur" and "St. Peter's Square", from which 1500 co-visible quadruplets were sampled in a way that balances difficulty based on visual overlap [30, 50, 56]. The Image Matching Challenge 2020 benchmark provides a comprehensive evaluation protocol, including datasets that cover multiple challenging outdoor wide-baseline scenes. From the phototourism dataset, we selected 3 validated scenes, each consisting of about 100 + +![](images/e497676a2d20632109343329ef50e38b1f36b268cef778dee4164bac065095f3.jpg) +Figure 4. Qualitative comparison on MegaDepth. For each quadruple with a target view (top), correspondences (green for correct, red for incorrect) predicted by CoMatcher and LightGlue are shown. Using identical local features as input, CoMatcher achieves significantly more reliable estimations, even at challenging semantic edges with depth discontinuities. This stems from its holistic scene understanding through multi-view cues to effectively address occlusions (Q3), large-scale variations (Q2, 4), and repetitive textures (Q1, 2, 3, 4). + +images. Additionally, we conducted an ablation study on a synthetic homography dataset [37] used for training. + +# 4.2. Implementation details + +CoMatcher is trained in two steps, following [30, 43]. We first pre-train the model on a large-scale synthetic homography dataset [37], leveraging noise-free homography for ground truth. Next, fine-tuning is performed on the MegaDepth dataset [28] using the camera poses recovered by SfM as ground truth. We sample 200 co-visible multiview groups per scene and randomly select one image as the target view. The size of source views is set to $M = 4$ during training. The training process is carried out on two NVIDIA GeForce 4090 GPUs, taking about 6 days in total. More details can be found in the supplementary materials. + +# 4.3. Homography estimation + +CoMatcher is compared with three types of baselines. First, we include state-of-the-art (SOTA) sparse two-view + +matchers: nearest-neighbor with mutual check [31], SuperGlue [43], SGMNet [7], and LightGlue [30], each paired with different feature extractors [10, 31, 53]. Second, we evaluate the multi-view matcher End2End [40]. For reference, we also compare against typical dense matchers LoFTR [50] and PDC-Net [52]. For CoMatcher and End2End, we use a single forward pass to obtain matches across all five image pairs, while others follow a pairwise approach. To ensure a fair comparison, we adopt the setups of [30, 50] for the number of features and image resizing. Homography accuracy is evaluated using both robust (RANSAC [17]) and non-robust (weighted DLT [20]) solvers. For each pair, we compute the mean reprojection error of the four image corners and report the area under the cumulative error curve (AUC) up to 1px, 3px, and 5px. + +Tab. 1 shows that CoMatcher yields more accurate estimates than all sparse two-view matchers, highlighting the advantage of inference in multi-view feature space. How- + +
MethodAUC - DLT @1px / @3px / @5pxAUC - RANSAC
denseLoFTR38.5 / 66.0 / 71.440.7 / 68.3 / 78.5
PDC-Net36.0 / 65.3 / 73.037.9 / 67.6 / 77.4
sparse 2-viewSIFT+NN+mutual0.0 / 0.0 / 0.035.9 / 65.0 / 75.6
SP+NN+mutual0.0 / 1.9 / 3.434.8 / 64.1 / 74.8
SP+SuperGlue32.2 / 65.1 / 75.737.2 / 68.0 / 78.7
SP+SGMNet31.7 / 64.9 / 76.037.7 / 66.4 / 77.5
SP+LightGlue35.4 / 67.5 / 77.737.2 / 67.8 / 78.1
DISK+NN+mutual1.8 / 5.2 / 7.837.9 / 58.0 / 68.3
DISK+LightGlue34.4 / 64.5 / 74.438.1 / 65.2 / 77.2
multi-viewSP+End2End34.3 / 66.9 / 75.537.0 / 67.2 / 77.5
SP+CoMatcher37.1 / 69.0 / 78.838.4 / 68.9 / 79.0
DISK+CoMatcher36.3 / 66.2 / 75.938.7 / 68.2 / 78.4
+ +ever, despite End2End also jointly reasoning over multiple views like ours, its performance is notably inferior. This underscores the superiority of a $N$ -to-1 architecture over $N$ -to- $N$ for multi-view representation learning. When compared to LoFTR, we observe slightly lower performance under low thresholds. We attribute this mainly to the limitations of the feature extractor in keypoint non-repeatability and localization errors [29]. Additionally, DLT achieves accuracy close to RANSAC on most metrics, reflecting the high-quality correspondences from CoMatcher. + +# 4.4. Relative pose estimation + +Next, we evaluate CoMatcher on challenging wide-baseline scenes, using the same baselines as in Sec. 4.3. For each quadruplet, including a selected target view, we still use a single forward pass for multi-view methods. Additionally, we test two newer dense methods, DUS3R [55] and MAS3R [27], following their evaluation setups for image resizing. All other settings remain consistent with [30]. The essential matrix is estimated from the correspondences using both vanilla RANSAC [17] and LO-RANSAC [9], following [30]. The pose error is then computed as the maximum angular error of rotation and translation, derived from the decomposition of the essential matrix. We report the AUC at $5^{\circ}$ , $10^{\circ}$ , and $20^{\circ}$ , and record the average runtime of matching each quadruplet on a single 4090 GPU. + +Tab. 2 shows that CoMatcher, as a sparse method, comprehensively enhances the estimation performance on different local features [10] and [53]. Importantly, compared to other sparse two-view matchers like LightGlue, a key strength of CoMatcher is its reliability, as illustrated in + +Table 1. Homography estimation on HPatches. We report the area under the cumulative error curve (AUC) up to values of 1px, 3px and 5px, using DLT and RANSAC [17] as homography solver. + +
MethodAUC RANSACAUC LO-RANSACTime (ms)
5° / 10° / 20°
denseLoFTR58.0 / 73.1 / 84.467.4 / 81.7 / 89.3182
PDCNet54.7 / 73.1 / 83.567.1 / 80.2 / 87.0231
DUSt3R42.4 / 56.7 / 64.258.1 / 70.4 / 79.6264
MASt3R51.5 / 65.3 / 75.663.5 / 76.3 / 85.2317
SuperPointNN+mutual35.3 / 58.3 / 53.751.4 / 67.3 / 75.99
SuperGlue55.8 / 72.8 / 84.165.1 / 77.2 / 89.287
LightGlue56.2 / 72.7 / 83.567.2 / 80.1 / 88.051
End2End55.3 / 71.4 / 81.267.4 / 81.5 / 87.0152
CoMatcher57.2 / 73.9 / 84.868.3 / 82.2 / 89.169
DISKNN+mutual50.9 / 66.7 / 77.764.0 / 79.5 / 87.69
LightGlue53.2 / 69.2 / 80.268.6 / 80.4 / 87.254
CoMatcher54.9 / 71.2 / 82.068.5 / 82.1 / 88.473
+ +Table 2. Relative pose estimation on MegaDepth. We report the AUC at $5^{\circ}$ , $10^{\circ}$ , and $20^{\circ}$ using different robust estimator, and the average runtime of matching each quadruplet. + +Fig. 3. This highlights the advantage of multi-view collaborative reasoning in understanding complex 3D structures, such as occlusions, even with suboptimal input local features in these areas. Additionally, compared to End2End, CoMatcher achieves more accurate relative poses without the need for cumbersome end-to-end training. While some dense methods may achieve better accuracy on certain metrics, their efficiency is generally much lower. By employing a space-for-time strategy to minimize inference steps, Co-Matcher runs significantly faster than most methods. + +# 4.5. Evaluation on the IMC 2020 benchmark + +Next, we evaluate our groupwise framework against pairwise baselines [30, 31, 43] on the Image Matching Challenge (IMC) 2020 benchmark [25]. Given a large-scale unordered image collection, the benchmark requires providing matching results for all image pairs, which are then assessed on two downstream tasks: stereo and multi-view reconstruction. For the stereo task, the accuracy of relative camera poses for each pair is evaluated from correspondences using RANSAC, as in Sec. 4.4. For the multi-view task, all correspondences are fed into COLMAP [45] for SfM, with the final accuracy evaluated based on the estimated multi-view camera poses. We report the AUC at $5^{\circ}$ and $10^{\circ}$ across both tasks. Additionally, the average runtime is reported, which is the total matching time for the entire image set divided by the number of pairs. + +Tab. 3 shows that our framework significantly outperforms existing pairwise approaches in both tasks. This advantage primarily stems from our framework's ability to better leverage the relationships within the original image + +
MethodStereoMulti-ViewTime (ms)
AUC 5°/10°AUC 5°/10°
two-viewSIFT+NN+mutual31.5 / 44.257.2 / 68.53
SP+NN+mutual28.6 / 40.352.9 / 63.43
SP+SuperGlue36.5 / 50.162.3 / 74.832
SP+LightGlue36.8 / 49.464.6 / 75.417
DISK+NN+mutual35.4 / 47.459.3 / 70.23
DISK+LightGlue42.1 / 55.665.2 / 76.219
m-vSP+CoMatcher39.1 / 52.466.2 / 77.525
DISK+CoMatcher44.9 / 57.367.1 / 78.429
+ +collection, enabling the generation of higher-quality tracks in such large-scale tasks. Additional comparisons on track quality for SfM tasks are provided in the supplementary material. In terms of efficiency, our approach is faster than SuperGlue, with the viewpoint grouping process accounting for only about $4\%$ of the total matching time. + +# 4.6. Ablation study + +Understanding CoMatcher. The CoMatcher network introduces three key components: multi-view cross-attention modules, a geometric constraint mechanism, and a multiview feature interaction strategy. We validate these design choices by evaluating on two benchmarks: the challenging synthetic homography dataset [37], where precision and recall metrics are employed, and HPatches [3], where the AUC of homography estimation via weighted DLT is measured. For each image, 512 keypoints are extracted. + +Tab. 4 shows the ablation results. Excluding the two cross-attention modules for multi-view interaction reduces the model to a series of parallel two-view matchers [30, 43]. This simplification limits the model's ability to integrate multi-view context, leading to a significant performance decline. Without the geometric constraint mechanism, the model performs even worse than in the previous case. This suggests that direct feature aggregation from multi-view is challenging, as the network requires additional learning to distinguish meaningful context effectively. Without multiview feature correlation, the model loses its ability to guide inference for ambiguous points by leveraging consistency constraints. This highlights the importance of guiding two-view feature correlation in the attention search process, a factor overlooked in previous work. + +Impact of the group size. We evaluated the matching performance on the synthetic homography dataset [37] by varying the size of the source view sets used as network input. As shown in Fig. 5, performance initially improves signif + +![](images/5bae3bf10a053c632f3bdc281e2fbc13ae7ba2abcd30da28f2ba6e9de42ced8d.jpg) +Figure 5. Impact of the group size. Under different numbers of keypoints, the AUC initially increases as the source view set expands, then tends to stabilize or shows a downward trend. + +Table 3. Evaluation on the IMC 2020 benchmark. We report the pose AUC at $5^{\circ}$ , $10^{\circ}$ for two subtasks, stereo and multi-view reconstruction, along with the average matching runtime. + +
CoMatcherSyntheticHpatches-AUC
precisionrecall@1px / @5px
w/o M-V feat. interaction89.796.632.4 / 75.1
w/o attenu. propagation87.794.231.5 / 73.7
w/o M-V feat. correlation90.595.333.1 / 74.9
full92.798.934.7 / 77.1
+ +Table 4. Ablation study on synthetic homography dataset and HPatches. The "full" model is the default model. + +icantly as more source views are integrated into the collaborative inference process. However, when the number of views exceeds 5, the matching capability begins to decline. We hypothesize that this degradation is attributed to the expanded search space and the excessive redundant information introduced by multi-view attention, which collectively increase the learning complexity of the network. + +# 5. Conclusion + +We proposed a multi-view collaborative matching strategy to address the challenge of reliable matching uncontrolled image sets in complex scenes. Our method, which consists of a deep collaborative matcher (CoMatcher) and a scalable groupwise pipeline, enables a holistic 3D scene understanding while inherently satisfying the cross-view projection consistency constraint. Extensive experiments have demonstrated that exploiting inter-view connections significantly enhances matching certainty, yielding substantial benefits for downstream tasks such as pose estimation and SfM. We hope this work encourages the research community to expand beyond pairwise matching and further explore the understanding and utilization of multi-view information. + +Acknowledgement. This research is supported by NSFC projects under Grant 42471447, Development Program of China under Grant 2024YFC3811000, and the Fundamental Research Funds for the Central Universities of China under Grant 2042022dx0001. + +# References + +[1] Sameer Agarwal, Noah Snavely, Ian Simon, Steven M Seitz, and Richard Szeliski. Building Rome in a day. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 72-79. IEEE, 2009. 1, 2 +[2] Relja Arandjelovic, Petr Gronat, Akihiko Torii, Tomas Pajdla, and Josef Sivic. NetVLAD: CNN architecture for weakly supervised place recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5297-5307, 2016. 3 +[3] Vassileios Balntas, Karel Lenc, Andrea Vedaldi, and Krystian Mikolajczyk. HPatches: A benchmark and evaluation of handcrafted and learned local descriptors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5173-5182, 2017. 2, 5, 8 +[4] Axel Barroso-Laguna, Sowmya Munukutla, Victor Adrian Prisacariu, and Eric Brachmann. Matching 2D images in 3D: Metric relative pose from metric correspondences. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4852-4863, 2024. 2 +[5] Herbert Bay, Tinne Tuytelaars, and Luc Van Gool. SURF: Speeded up robust features. In European Conference on Computer Vision, pages 404-417. Springer, 2006. 2 +[6] Ruojin Cai, Joseph Tung, Qianqian Wang, Hadar Averbuch-Elor, Bharath Hariharan, and Noah Snavely. Doppelgangers: Learning to disambiguate images of similar structures. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 34–44, 2023. 1 +[7] Hongkai Chen, Zixin Luo, Jiahui Zhang, Lei Zhou, Xuyang Bai, Zeyu Hu, Chiew-Lan Tai, and Long Quan. Learning to match features with seeded graph matching network. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6301-6310, 2021. 2, 6 +[8] Hongkai Chen, Zixin Luo, Lei Zhou, Yurun Tian, Mingmin Zhen, Tian Fang, David Mckinnon, Yanghai Tsin, and Long Quan. ASpanFormer: Detector-free image matching with adaptive span transformer. In European Conference on Computer Vision, pages 20-36. Springer, 2022. 2 +[9] Ondrej Chum, Jiří Matas, and Josef Kittler. Locally optimized RANSAC. In Pattern Recognition, pages 236-243. Springer, 2003. 7 +[10] Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. SuperPoint: Self-supervised interest point detection and description. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 224-236, 2018. 2, 6, 7 +[11] Carl Doersch, Ankush Gupta, Larisa Markeeeva, Adria Recasens, Lucas Smaira, Yusuf Aytar, Joao Carreira, Andrew Zisserman, and Yi Yang. TAP-Vid: A benchmark for tracking any point in a video. Advances in Neural Information Processing Systems, 35:13610-13626, 2022. 2 +[12] Carl Doersch, Yi Yang, Mel Vecerik, Dilara Gokay, Ankush Gupta, Yusuf Aytar, Joao Carreira, and Andrew Zisserman. TAPIR: Tracking any point with per-frame initialization and + +temporal refinement. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10061-10072, 2023. 2, 4 +[13] Mihai Dusmanu, Ignacio Rocco, Tomas Pajdla, Marc Pollefeys, Josef Sivic, Akihiko Torii, and Torsten Sattler. D2-Net: A trainable cnn for joint description and detection of local features. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8092–8101, 2019. 2 +[14] Johan Edstedt, Georg Bokman, Marten Wadenback, and Michael Felsberg. DeDoDe: Detect, don't describe—Describe, don't detect for local feature matching. In International Conference on 3D Vision, pages 148-157. IEEE, 2024. 2 +[15] Johan Edstedt, Qiyu Sun, Georg Bokman, Marten Wadenback, and Michael Felsberg. RoMa: Robust dense feature matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19790-19800, 2024. 1, 2 +[16] Maha Elbayad, Jiatao Gu, Edouard Grave, and Michael Auli. Depth-adaptive Transformer. In International Conference on Learning Representations, pages 1-14, 2020. 5 +[17] Martin A Fischler and Robert C Bolles. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. In Communications of the ACM, 24(6):381-395, 1981. 6, 7 +[18] Jan-Michael Frahm, Pierre Fite-Georgel, David Gallup, Tim Johnson, Rahul Raguram, Changchang Wu, Yi-Hung Jen, Enrique Dunn, Brian Clipp, Svetlana Lazebnik, et al. Building Rome on a cloudless day. In European Conference on Computer Vision, pages 368-381. Springer, 2010. 1 +[19] Adam W Harley, Zhaoyuan Fang, and Katerina Fragkiadaki. Particle video revisited: Tracking through occlusions using point trajectories. In European Conference on Computer Vision, pages 59-75. Springer, 2022. 2 +[20] Richard Hartley and Andrew Zisserman. Multiple view geometry in computer vision. Cambridge university press, 2003. 1, 6 +[21] Xingyi He, Jiaming Sun, Yifan Wang, Sida Peng, Qixing Huang, Hujun Bao, and Xiaowei Zhou. Detector-free structure-from-motion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21594-21603, 2024. 2 +[22] Martin Humenberger, Yohann Cabon, Nicolas Guerin, Julien Morat, Vincent Leroy, Jérôme Revaud, Philippe Rerole, Noé Pion, Cesar de Souza, and Gabriela Csurka. Robust image retrieval-based visual localization using kapture. arXiv preprint arXiv:2007.13867, 2020. 1 +[23] Heinly Jared, Johannes L Schonberger, Enrique Dunn, and Jan-Michael Frahm. Reconstructing the world in six days. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2015. 1 +[24] Hanwen Jiang, Arjun Karpur, Bingyi Cao, Qixing Huang, and André Araujo. OmniGlue: Generalizable feature matching with foundation model guidance. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19865-19875, 2024. 2 + +[25] Yuhe Jin, Dmytro Mishkin, Anastasiia Mishchuk, Jiri Matas, Pascal Fua, Kwang Moo Yi, and Eduard Trulls. Image matching across wide baselines: From paper to practice. International Journal of Computer Vision, 129(2):517-547, 2021. 1, 2, 7 +[26] Nikita Karaev, Ignacio Rocco, Benjamin Graham, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. CoTracker: It is better to track together. In European Conference on Computer Vision, pages 18-35. Springer, 2024. 2, 4 +[27] Vincent Leroy, Yohann Cabon, and Jérôme Revaud. Grounding image matching in 3D with MAST3R. In European Conference on Computer Vision, pages 71-91. Springer, 2024. 1, 2, 7 +[28] Zhengqi Li and Noah Snavely. MegaDepth: Learning single-view depth prediction from internet photos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2041-2050, 2018. 5, 6 +[29] Philipp Lindenberger, Paul-Edouard Sarlin, Viktor Larsson, and Marc Pollefeys. Pixel-perfect structure-from-motion with featuremetric refinement. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5987-5997, 2021. 1, 7 +[30] Philipp Lindenberger, Paul-Edouard Sarlin, and Marc Pollefeys. LightGlue: Local feature matching at light speed. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17627–17638, 2023. 1, 2, 5, 6, 7, 8 +[31] David G Lowe. Distinctive image features from scale-invariant keypoints. International Journal of Computer Vision, 60:91-110, 2004. 1, 2, 6, 7 +[32] Jiayi Ma, Xingyu Jiang, Aoxiang Fan, Junjun Jiang, and Junchi Yan. Image matching from handcrafted to deep features: A survey. International Journal of Computer Vision, 129(1):23-79, 2021. 1, 2 +[33] Eleonora Maset, Federica Arrigoni, and Andrea Fusiello. Practical and efficient multi-view matching. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4568-4576, 2017. 2 +[34] Raul Mur-Artal, Jose Maria Martinez Montiel, and Juan D Tardos. ORB-SLAM: A versatile and accurate monocular SLAM system. In IEEE Transactions on Robotics, 31(5): 1147-1163, 2015. 1 +[35] Carl Olsson and Olof Enqvist. Stable structure-from-motion for unordered image collections. In Image Analysis, pages 524-535. Springer, 2011. 2 +[36] Linfei Pan, Daniel Barath, Marc Pollefeys, and Johannes L Schonberger. Global structure-from-motion revisited. In European Conference on Computer Vision, 2024. 2 +[37] Filip Radenović, Ahmet Iscen, Giorgos Tolias, Yannis Avrithis, and Ondrej Chum. Revisiting Oxford and Paris: Large-scale image retrieval benchmarking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5706-5715, 2018. 6, 8 +[38] Filip Radenovic, Giorgos Tolias, and Ondrej Chum. Finetuning CNN image retrieval with no human annotation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 41(7):1655-1668, 2018. 3 + +[39] Jerome Revaud, Cesar De Souza, Martin Humenberger, and Philippe Weinzaepfel. R2D2: Reliable and repeatable detector and descriptor. Advances in Neural Information Processing Systems, 32, 2019. 2 +[40] Barbara Roessle and Matthias Nießner. End2end multi-view feature matching with differentiable pose optimization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 477-487, 2023. 2, 6 +[41] Ethan Rublee, Vincent Rabaud, Kurt Konolige, and Gary Bradski. ORB: An efficient alternative to SIFT or SURF. In Proceeding of the IEEE/CVF International Conference on Computer Vision, pages 2564-2571. IEEE, 2011. 2 +[42] Paul-Edouard Sarlin, Cesar Cadena, Roland Siegwart, and Marcin Dymczyk. From coarse to fine: Robust hierarchical localization at large scale. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12716–12725, 2019. 1, 2 +[43] Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. SuperGlue: Learning feature matching with graph neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4938–4947, 2020. 1, 2, 4, 5, 6, 7, 8 +[44] Torsten Sattler, Akihiko Torii, Josef Sivic, Marc Pollefeys, Hajime Taira, Masatoshi Okutomi, and Tomas Pajdla. Are large-scale 3d models really necessary for accurate visual localization? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1637-1646, 2017. 2 +[45] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-Motion revisited. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4104-4113, 2016. 1, 2, 7 +[46] Johannes L Schonberger, Hans Hardmeier, Torsten Sattler, and Marc Pollefeys. Comparative evaluation of hand-crafted and learned local features. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1482-1491, 2017. 2 +[47] Tal Schuster, Adam Fisch, Jai Gupta, Mostafa Dehghani, Dara Bahri, Vinh Tran, Yi Tay, and Donald Metzler. Confident adaptive language modeling. Advances in Neural Information Processing Systems, 35:17456-17472, 2022. 5 +[48] Sivic and Zisserman. Video Google: A text retrieval approach to object matching in videos. In Proceedings in the IEEE/CVF International Conference on Computer Vision, pages 1470-1477. IEEE, 2003. 3 +[49] Jianlin Su, Murtadha Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu. RoFormer: Enhanced transformer with rotary position embedding. Neurocomputing, 568:127063, 2024. 4 +[50] Jiaming Sun, Zehong Shen, Yuang Wang, Hujun Bao, and Xiaowei Zhou. LoFTR: Detector-free local feature matching with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8922-8931, 2021. 1, 2, 5, 6 +[51] Kun Sun, Jinhong Yu, Wenbing Tao, Xin Li, Chang Tang, and Yuhua Qian. A unified feature-spatial cycle consistency fusion framework for robust image matching. Information Fusion, 97:101810, 2023. 2 + +[52] Prune Truong, Martin Danelljan, Luc Van Gool, and Radu Timofte. Learning accurate dense correspondences and when to trust them. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5714-5724, 2021. 2, 6 +[53] Michal Tyszkiiewicz, Pascal Fua, and Eduard Trulls. DISK: Learning local features with policy gradient. Advances in Neural Information Processing Systems, 33:14254-14265, 2020. 2, 5, 6, 7 +[54] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in Neural Information Processing Systems, 30, 2017. 1, 5 +[55] Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. DUSt3R: Geometric 3D vision made easy. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20697-20709, 2024. 7 +[56] Yifan Wang, Xingyi He, Sida Peng, Dongli Tan, and Xiaowei Zhou. Efficient LoFTR: Semi-dense local feature matching with sparse-like speed. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21666-21675, 2024. 2, 5 +[57] Yuxi Xiao, Qianqian Wang, Shangzhan Zhang, Nan Xue, Sida Peng, Yujun Shen, and Xiaowei Zhou. SpatialTracker: Tracking any 2D pixels in 3D space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20406-20417, 2024. 1 +[58] Shibiao Xu, Shunpeng Chen, Rongtao Xu, Changwei Wang, Peng Lu, and Li Guo. Local feature matching using deep learning: A survey. Information Fusion, 107:102344, 2024. 1 +[59] Kwang Moo Yi, Eduard Trulls, Vincent Lepetit, and Pascal Fua. LIFT: Learned invariant feature transform. In European Conference on Computer Vision, pages 467-483. Springer, 2016. 2 +[60] Guofeng Zhang, Zilong Dong, Jiaya Jia, Tien-Tsin Wong, and Hujun Bao. Efficient non-consecutive feature tracking for structure-from-Motion. In European Conference on Computer Vision, pages 422-435. Springer, 2010. 2 \ No newline at end of file diff --git a/2025/CoMatcher_ Multi-View Collaborative Feature Matching/images.zip b/2025/CoMatcher_ Multi-View Collaborative Feature Matching/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..7d81e79d7d206f2dce588f8738d84c598bfb25df --- /dev/null +++ b/2025/CoMatcher_ Multi-View Collaborative Feature Matching/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9da213b9ab422559bca9b66bae28953ee4e9d9ac57e6162c582aac41aaeacdc6 +size 700734 diff --git a/2025/CoMatcher_ Multi-View Collaborative Feature Matching/layout.json b/2025/CoMatcher_ Multi-View Collaborative Feature Matching/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..afdef2024e86cf00d23cb37ccf52b9dfcab0f8b9 --- /dev/null +++ b/2025/CoMatcher_ Multi-View Collaborative Feature Matching/layout.json @@ -0,0 +1,8426 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 129, + 103, + 482, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 103, + 482, + 121 + ], + "spans": [ + { + "bbox": [ + 129, + 103, + 482, + 121 + ], + "type": "text", + "content": "CoMatcher: Multi-View Collaborative Feature Matching" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 83, + 142, + 525, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 142, + 525, + 201 + ], + "spans": [ + { + "bbox": [ + 83, + 142, + 525, + 201 + ], + "type": "text", + "content": "Jintao Zhang" + }, + { + "bbox": [ + 83, + 142, + 525, + 201 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 83, + 142, + 525, + 201 + ], + "type": "text", + "content": ", Zimin Xia" + }, + { + "bbox": [ + 83, + 142, + 525, + 201 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 83, + 142, + 525, + 201 + ], + "type": "text", + "content": ", Mingyue Dong" + }, + { + "bbox": [ + 83, + 142, + 525, + 201 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 83, + 142, + 525, + 201 + ], + "type": "text", + "content": ", Shuhan Shen" + }, + { + "bbox": [ + 83, + 142, + 525, + 201 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 83, + 142, + 525, + 201 + ], + "type": "text", + "content": ", Linwei Yue" + }, + { + "bbox": [ + 83, + 142, + 525, + 201 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 83, + 142, + 525, + 201 + ], + "type": "text", + "content": ", Xianwei Zheng" + }, + { + "bbox": [ + 83, + 142, + 525, + 201 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 83, + 142, + 525, + 201 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 83, + 142, + 525, + 201 + ], + "type": "text", + "content": "The State Key Lab. LIESMARS, Wuhan University \n" + }, + { + "bbox": [ + 83, + 142, + 525, + 201 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 83, + 142, + 525, + 201 + ], + "type": "text", + "content": "École Polytechnique Fédérale de Lausanne (EPFL) \n" + }, + { + "bbox": [ + 83, + 142, + 525, + 201 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 83, + 142, + 525, + 201 + ], + "type": "text", + "content": "Chinese Academy of Sciences " + }, + { + "bbox": [ + 83, + 142, + 525, + 201 + ], + "type": "inline_equation", + "content": "{ }^{4}" + }, + { + "bbox": [ + 83, + 142, + 525, + 201 + ], + "type": "text", + "content": "China University of Geosciences" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 151, + 228, + 200, + 240 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 228, + 200, + 240 + ], + "spans": [ + { + "bbox": [ + 151, + 228, + 200, + 240 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 253, + 296, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 253, + 296, + 482 + ], + "spans": [ + { + "bbox": [ + 55, + 253, + 296, + 482 + ], + "type": "text", + "content": "This paper proposes a multi-view collaborative matching strategy for reliable track construction in complex scenarios. We observe that the pairwise matching paradigms applied to image set matching often result in ambiguous estimation when the selected independent pairs exhibit significant occlusions or extreme viewpoint changes. This challenge primarily stems from the inherent uncertainty in interpreting intricate 3D structures based on limited two-view observations, as the 3D-to-2D projection leads to significant information loss. To address this, we introduce CoMatcher, a deep multi-view matcher to (i) leverage complementary context cues from different views to form a holistic 3D scene understanding and (ii) utilize cross-view projection consistency to infer a reliable global solution. Building on CoMatcher, we develop a groupwise framework that fully exploits cross-view relationships for large-scale matching tasks. Extensive experiments on various complex scenarios demonstrate the superiority of our method over the mainstream two-view matching paradigm." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 505, + 135, + 517 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 505, + 135, + 517 + ], + "spans": [ + { + "bbox": [ + 56, + 505, + 135, + 517 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 525, + 295, + 669 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 525, + 295, + 669 + ], + "spans": [ + { + "bbox": [ + 55, + 525, + 295, + 669 + ], + "type": "text", + "content": "Research on recovering 3D structures and camera poses from multiple views storing scene information has a long history [20, 23, 34]. A core component of this process is feature matching, which aims to estimate point correspondences across the image set [25, 32, 58]. For computational flexibility, existing methods often decompose the set into pairs of co-visible images [1, 18] and apply a two-view matching approach to each pair independently [31, 43]. The pairwise matches are then merged into comprehensive multi-view tracks [1, 45]. Following this pairwise framework to build systems for localization [22, 42] and mapping [29, 45] has become a entrenched research paradigm." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 670, + 295, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 670, + 295, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 670, + 295, + 693 + ], + "type": "text", + "content": "Recent advances in learning-based two-view matchers have greatly improved performance in most cases [15, 30," + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 318, + 228, + 552, + 397 + ], + "blocks": [ + { + "bbox": [ + 318, + 228, + 552, + 397 + ], + "lines": [ + { + "bbox": [ + 318, + 228, + 552, + 397 + ], + "spans": [ + { + "bbox": [ + 318, + 228, + 552, + 397 + ], + "type": "image", + "image_path": "e18239df716da0f8358979bd3171e64aed9fd5b1e52c6cccbddd3a248a20e2a0.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 406, + 555, + 473 + ], + "lines": [ + { + "bbox": [ + 313, + 406, + 555, + 473 + ], + "spans": [ + { + "bbox": [ + 313, + 406, + 555, + 473 + ], + "type": "text", + "content": "Figure 1. From two view to multi-view. Unlike pairwise schemes prone to uncertainty, we first partition the image set into co-visible groups and collaboratively estimate group-wide correspondences using a reliable deep multi-view matcher. Benefiting from holistic scene understanding and consistency constraint, our method generates high-quality tracks in complex scenarios." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 496, + 555, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 496, + 555, + 556 + ], + "spans": [ + { + "bbox": [ + 313, + 496, + 555, + 556 + ], + "type": "text", + "content": "43, 50]. Inspired by Transformer [54], these methods employ deep networks to jointly analyze global context from both views. This enables models to learn to infer underlying 3D scenes by integrating spatial and visual cues, which has been proven crucial for accurate estimations [27, 43]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 558, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 556, + 713 + ], + "type": "text", + "content": "Despite great progress, current two-view matchers still struggle in challenging wide-baseline scenarios, particularly with severe occlusions and repetitive textures [25]. We argue that one potential cause for this difficulty lies in the inherent uncertainty of interpreting complex 3D structures from limited two-view observations—a task that even humans find prone to perceptual ambiguity. Projecting complex 3D geometries onto 2D planes inevitably results in significant loss of scene information. For example, spatially distant regions may appear close in 2D. This makes it highly unreliable to infer the original scene solely from ambiguous two-view context [6, 57]. Moreover, while matchers can learn two-view geometric priors, this alone is also in" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "type": "text", + "content": "*Corresponding author." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21970" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "content": "adequate—capturing abrupt depth discontinuities requires stronger constraints [4, 27]. Errors in pairwise correspondences compound during merging, significantly impacting downstream tasks, especially in large-scale datasets." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 121, + 294, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 121, + 294, + 275 + ], + "spans": [ + { + "bbox": [ + 55, + 121, + 294, + 275 + ], + "type": "text", + "content": "To address these challenges, we propose a shift in focus: instead of optimizing two-view matchers, a more effective strategy is to directly leverage the rich relationships inherent in raw multi-view observations. Building on this, we introduce a unified formulation that collaboratively establishes correspondences between a complementary view group and a target view. This approach fosters holistic scene understanding by integrating contextual information from multiple views, while multi-view geometric priors provide stronger constraints for reliable reasoning. Moreover, matches across views naturally adhere to cross-view projection consistency. This facilitates a mutual verification process, ensuring confident inference of a global solution." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 276, + 295, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 276, + 295, + 430 + ], + "spans": [ + { + "bbox": [ + 55, + 276, + 295, + 430 + ], + "type": "text", + "content": "Inspired by existing two-view methods [43, 50], we develop CoMatcher, a deep-learning architecture for multiview feature matching. Context sharing is achieved through a GNN with multi-view receptive fields. While a larger feature space enriches context, it also introduces noise, complicating learning. To address this, we propose to constrain the search space for each point by leveraging cross-view projection geometry. Furthermore, we redesign the feature correlation layers in cross-attention to enable consistent multiview reasoning. This design exploits the property that correlation scores encode both reasoning outcomes and confidence, enabling progressive layer-by-layer integration of multi-view estimates into a globally consistent solution." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 431, + 295, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 431, + 295, + 622 + ], + "spans": [ + { + "bbox": [ + 55, + 431, + 295, + 622 + ], + "type": "text", + "content": "To scale CoMatcher for large-scale tasks like SfM [36, 45], we develop a groupwise matching pipeline (see Fig. 1). It introduces a novel set partitioning algorithm that uses可视ability information to group images, followed by group-level matching using CoMatcher. By preserving cross-view relationships, our framework achieves comprehensive improvements over pairwise methods in extensive experiments on wide-baseline scenarios. The main contributions are summarized as follows: (i) A novel multi-view matching paradigm, comprising a deep collaborative matcher and a scalable groupwise pipeline, designed to overcome the inherent uncertainty of pairwise approaches. (ii) A multiview representation learning network architecture that enables holistic scene understanding. (iii) A multi-view feature correlation strategy that integrates uncertain individual estimates into a globally consistent solution." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 633, + 139, + 645 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 633, + 139, + 645 + ], + "spans": [ + { + "bbox": [ + 55, + 633, + 139, + 645 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 653, + 294, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 294, + 713 + ], + "type": "text", + "content": "Two-view matching is a fundamental problem in computer vision. The most prevalent paradigm involves sparsifying images into keypoints, each represented by a high-dimensional vector encoding its visual context [25, 32]. This transforms matching into a search problem, with match" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 553, + 311 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 311 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 311 + ], + "type": "text", + "content": "likelihood expressed by the inner product of corresponding vectors [31]. Modern methods leverage deep networks for representation learning to extract feature points [10, 13, 14, 39, 53, 59], outperforming traditional hand-crafted encodings [5, 31, 41] and achieving significant advancements on mainstream benchmarks [3, 25, 46]. In parallel, other works aim to address the limitations of heuristic matching strategies by leveraging networks with global modeling capabilities to directly learn the matching process [7, 24, 30, 43]. Such data-driven methods can effectively capture priors of the underlying 3D scene, which are essential for accurate estimation. Unlike keypoint-based approaches, another typical method retains all pixels for dense inference [8, 15, 27, 50, 52]. While this has led to notable progress in weakly textured regions, such methods require greater computational resources [30, 56] and struggle with multi-view inconsistency [21]. In summary, while these methods perform well in most scenarios, they remain constrained in complex wide-baseline settings due to the limited receptive field inherent to two-view systems." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 312, + 553, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 312, + 553, + 574 + ], + "spans": [ + { + "bbox": [ + 313, + 312, + 553, + 574 + ], + "type": "text", + "content": "Multi-view matching focuses on establishing correspondences across a larger-scale image collection. A typical pairwise matching approach decomposes the problem into two-view matching instances and subsequently merges the pairwise results using multi-view projection consistency [1, 42, 44, 45]. During merging, some works explore how consistency assumptions can filter and refine coarse matches [33, 51], while others focus on efficiency, developing better strategies to reduce redundant matching computations [35, 60]. However, as a post-processing step, merging cannot address catastrophic failures that occur during the matching stage. Another line of work treats matching as a point tracking problem [11, 12, 19, 26], using optical flow-like frameworks to reason about the positions of query pixels across multiple images. While this approach is well-suited for video inputs, its effectiveness on unstructured, wide-baseline image collections remains largely unverified. The most relevant method to ours is [40], which performs end-to-end multi-view matching and pose estimation. However, it is limited in scalability, as it can only handle a small image group. In contrast, our framework is highly flexible, with the potential to scale to an arbitrary number of images." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 586, + 397, + 599 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 586, + 397, + 599 + ], + "spans": [ + { + "bbox": [ + 313, + 586, + 397, + 599 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 605, + 553, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 553, + 677 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 553, + 677 + ], + "type": "text", + "content": "**Formulation.** Given a set of " + }, + { + "bbox": [ + 313, + 605, + 553, + 677 + ], + "type": "inline_equation", + "content": "N_I" + }, + { + "bbox": [ + 313, + 605, + 553, + 677 + ], + "type": "text", + "content": " images, " + }, + { + "bbox": [ + 313, + 605, + 553, + 677 + ], + "type": "inline_equation", + "content": "\\mathcal{I} = \\{I_i \\mid i = 1, \\dots, N_I\\}" + }, + { + "bbox": [ + 313, + 605, + 553, + 677 + ], + "type": "text", + "content": ", of a scene, the objective of feature matching is to identify the 2D position coordinates " + }, + { + "bbox": [ + 313, + 605, + 553, + 677 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_k = \\{\\mathbf{p}^{I_i}\\}" + }, + { + "bbox": [ + 313, + 605, + 553, + 677 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 313, + 605, + 553, + 677 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 313, + 605, + 553, + 677 + ], + "type": "text", + "content": " that correspond to identical 3D scene points " + }, + { + "bbox": [ + 313, + 605, + 553, + 677 + ], + "type": "inline_equation", + "content": "\\mathcal{X} = \\{\\mathbf{x}_k \\mid k = 1, \\dots, N_X\\}" + }, + { + "bbox": [ + 313, + 605, + 553, + 677 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 605, + 553, + 677 + ], + "type": "inline_equation", + "content": "\\mathbf{p}^{I_i} \\in \\mathbb{R}^2" + }, + { + "bbox": [ + 313, + 605, + 553, + 677 + ], + "type": "text", + "content": " are the pixel coordinates in " + }, + { + "bbox": [ + 313, + 605, + 553, + 677 + ], + "type": "inline_equation", + "content": "I_i" + }, + { + "bbox": [ + 313, + 605, + 553, + 677 + ], + "type": "text", + "content": ", and each " + }, + { + "bbox": [ + 313, + 605, + 553, + 677 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_k" + }, + { + "bbox": [ + 313, + 605, + 553, + 677 + ], + "type": "text", + "content": " is called a track." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "content": "Motivation. Constructing complete tracks requires estimating correspondences between each image " + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "inline_equation", + "content": "I_{t} \\subset \\mathcal{I}" + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "content": " and its co-visible image set " + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_t" + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "content": ". While the pairwise approach of" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21971" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 70, + 70, + 545, + 240 + ], + "blocks": [ + { + "bbox": [ + 70, + 70, + 545, + 240 + ], + "lines": [ + { + "bbox": [ + 70, + 70, + 545, + 240 + ], + "spans": [ + { + "bbox": [ + 70, + 70, + 545, + 240 + ], + "type": "image", + "image_path": "b8d28b5cf08c3d45e508810fe0943624b78ba3655e92c41e14199624d5892134.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 250, + 555, + 307 + ], + "lines": [ + { + "bbox": [ + 54, + 250, + 555, + 307 + ], + "spans": [ + { + "bbox": [ + 54, + 250, + 555, + 307 + ], + "type": "text", + "content": "Figure 2. The CoMatcher architecture can be viewed as an extension of parallel two-view matchers, enhanced with two core components: the multi-view (M-V) feature interaction module (Sec. 3.2) and the multi-view (M-V) feature correlation strategy (Sec. 3.3). The first component employs multi-view cross-attention to enhance the representation of all source and target view features, while leveraging the cross-view projection geometry from " + }, + { + "bbox": [ + 54, + 250, + 555, + 307 + ], + "type": "inline_equation", + "content": "\\mathcal{M}(\\mathcal{G})" + }, + { + "bbox": [ + 54, + 250, + 555, + 307 + ], + "type": "text", + "content": " to constrain the point search space. The second component aggregates correlation scores from multiple two-view cross-attention modules based on estimated confidence, guiding the network to achieve multi-view consistent reasoning." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 326, + 297, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 326, + 297, + 506 + ], + "spans": [ + { + "bbox": [ + 54, + 326, + 297, + 506 + ], + "type": "text", + "content": "fers implementation flexibility for this problem, its reliance on two-view inputs in a single estimation often leads to uncertainty in complex scenes. In practice, images within " + }, + { + "bbox": [ + 54, + 326, + 297, + 506 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_t" + }, + { + "bbox": [ + 54, + 326, + 297, + 506 + ], + "type": "text", + "content": " are often not independent—most are captured from adjacent positions but with differing viewing angles. This inspires us to group " + }, + { + "bbox": [ + 54, + 326, + 297, + 506 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_t" + }, + { + "bbox": [ + 54, + 326, + 297, + 506 + ], + "type": "text", + "content": " based on image correlations, where each complementary group offers a mapping from multiple planar views to a 3D sub-scene. In the image matching problem, the correspondence between " + }, + { + "bbox": [ + 54, + 326, + 297, + 506 + ], + "type": "inline_equation", + "content": "I_t" + }, + { + "bbox": [ + 54, + 326, + 297, + 506 + ], + "type": "text", + "content": " and a group must adhere to certain 2D-to-3D physical constraints, such as cross-view projection consistency. Rather than independently estimating and then integrating them with constraints, it is more efficient to unify this process and learn underlying priors directly from the data. Thus, we formulate CoMatcher to address the 1-to-" + }, + { + "bbox": [ + 54, + 326, + 297, + 506 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 54, + 326, + 297, + 506 + ], + "type": "text", + "content": " matching problem (see Fig. 2)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 510, + 297, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 297, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 297, + 715 + ], + "type": "text", + "content": "Groupwise matching pipeline. Based on the above, we propose a three-step pipeline for constructing tracks: grouping, connecting, and matching. First, we treat the original set of views " + }, + { + "bbox": [ + 55, + 510, + 297, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 55, + 510, + 297, + 715 + ], + "type": "text", + "content": " as a union of smaller image groups " + }, + { + "bbox": [ + 55, + 510, + 297, + 715 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{G}_s\\mid s = 1,\\dots ,N_G\\}" + }, + { + "bbox": [ + 55, + 510, + 297, + 715 + ], + "type": "text", + "content": ", each representing a localized scene. This involves a pre-processing step to partition the set. The most critical aspect of this is determining the correlation between two images, i.e., whether they describe the same structure. Building on existing research [2, 38, 48], we employ heuristic techniques to design a grouping algorithm. Next, although CoMatcher can directly learn multi-view geometry, certain fine-grained physical rules are challenging to model. Therefore, we explicitly establish relationships within each group to guide subsequent matching. For each group, we construct tracks using existing frameworks, which provide a set of 3D points along with their multi-view projections and implicitly encode camera poses. These cues" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 326, + 556, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 326, + 556, + 388 + ], + "spans": [ + { + "bbox": [ + 313, + 326, + 556, + 388 + ], + "type": "text", + "content": "are incorporated as inputs to the network to steer the inference process. Finally, each group of images is treated as a whole to collaboratively match with other images in " + }, + { + "bbox": [ + 313, + 326, + 556, + 388 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 313, + 326, + 556, + 388 + ], + "type": "text", + "content": " using CoMatcher. Additional details on grouping and the overall framework are provided in the supplementary materials." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 394, + 436, + 406 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 394, + 436, + 406 + ], + "spans": [ + { + "bbox": [ + 313, + 394, + 436, + 406 + ], + "type": "text", + "content": "3.1. CoMatcher overview" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 411, + 556, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 411, + 556, + 579 + ], + "spans": [ + { + "bbox": [ + 313, + 411, + 556, + 579 + ], + "type": "text", + "content": "As shown in Fig. 2, CoMatcher, as a deep sparse matcher, estimates point correspondences between each source view in a group " + }, + { + "bbox": [ + 313, + 411, + 556, + 579 + ], + "type": "inline_equation", + "content": "I_{i} \\in \\mathcal{G}" + }, + { + "bbox": [ + 313, + 411, + 556, + 579 + ], + "type": "text", + "content": " and a target view " + }, + { + "bbox": [ + 313, + 411, + 556, + 579 + ], + "type": "inline_equation", + "content": "I_{t}" + }, + { + "bbox": [ + 313, + 411, + 556, + 579 + ], + "type": "text", + "content": ". The network takes as input the extracted local features " + }, + { + "bbox": [ + 313, + 411, + 556, + 579 + ], + "type": "inline_equation", + "content": "\\{\\left(\\mathbf{p}_{k}^{I_{i}}, \\mathbf{d}_{k}^{I_{i}}\\right) \\mid k = 1, \\dots, N_{F}\\}" + }, + { + "bbox": [ + 313, + 411, + 556, + 579 + ], + "type": "text", + "content": " from the " + }, + { + "bbox": [ + 313, + 411, + 556, + 579 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 411, + 556, + 579 + ], + "type": "text", + "content": " source views and the target view, along with the precomputed tracks " + }, + { + "bbox": [ + 313, + 411, + 556, + 579 + ], + "type": "inline_equation", + "content": "\\mathcal{M}(\\mathcal{G})" + }, + { + "bbox": [ + 313, + 411, + 556, + 579 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 313, + 411, + 556, + 579 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 313, + 411, + 556, + 579 + ], + "type": "text", + "content": ". Each local feature consists of keypoint locations " + }, + { + "bbox": [ + 313, + 411, + 556, + 579 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_{k}" + }, + { + "bbox": [ + 313, + 411, + 556, + 579 + ], + "type": "text", + "content": " and their corresponding visual descriptors " + }, + { + "bbox": [ + 313, + 411, + 556, + 579 + ], + "type": "inline_equation", + "content": "\\mathbf{d}_{k}" + }, + { + "bbox": [ + 313, + 411, + 556, + 579 + ], + "type": "text", + "content": ". Initially, the target view features are broadcast and paired with those of each source view to establish the initial states. These " + }, + { + "bbox": [ + 313, + 411, + 556, + 579 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 411, + 556, + 579 + ], + "type": "text", + "content": " feature pairs are then refined using a Graph Neural Network (GNN) with alternating self-attention and cross-attention modules, repeated " + }, + { + "bbox": [ + 313, + 411, + 556, + 579 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 313, + 411, + 556, + 579 + ], + "type": "text", + "content": " times. Finally, the enhanced features are fed into matching heads to infer the final correspondences." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 580, + 556, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 580, + 556, + 664 + ], + "spans": [ + { + "bbox": [ + 313, + 580, + 556, + 664 + ], + "type": "text", + "content": "The design of CoMatcher revolves around two core questions: (i) how can multi-view context be leveraged to learn better point representations (Sec. 3.2), and (ii) how can the estimations across multiple views be constrained to satisfy cross-view consistency (Sec. 3.3)? After thoroughly discussing these questions, we introduce the prediction heads (Sec. 3.4) and the loss function (Sec. 3.5)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 672, + 479, + 683 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 672, + 479, + 683 + ], + "spans": [ + { + "bbox": [ + 313, + 672, + 479, + 683 + ], + "type": "text", + "content": "3.2. Multi-view feature interaction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": "Consider an occluded scene. When observing such structures from a frontal view, points near occlusion boundaries" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21972" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "content": "are often contaminated by irrelevant context. This noise significantly compromises the quality of point features, creating substantial challenges for matching. Multi-view learning provides a promising solution to this issue. For these ambiguous point features, it can integrate observations of the same area from multiple other viewpoints to refine their representations. This involves a querying process, which establishes a multi-view receptive field for each point." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 169, + 294, + 214 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 169, + 294, + 214 + ], + "spans": [ + { + "bbox": [ + 55, + 169, + 294, + 214 + ], + "type": "text", + "content": "Given a set of co-visible source views, CoMatcher leverages a multi-view cross-attention mechanism to aggregate these context (Source Cross). For a point " + }, + { + "bbox": [ + 55, + 169, + 294, + 214 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 55, + 169, + 294, + 214 + ], + "type": "text", + "content": " in source view " + }, + { + "bbox": [ + 55, + 169, + 294, + 214 + ], + "type": "inline_equation", + "content": "I_{i}" + }, + { + "bbox": [ + 55, + 169, + 294, + 214 + ], + "type": "text", + "content": ", we attend to all points in the other source views:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 64, + 226, + 295, + 255 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 226, + 295, + 255 + ], + "spans": [ + { + "bbox": [ + 64, + 226, + 295, + 255 + ], + "type": "interline_equation", + "content": "\\mathbf {m} _ {u} ^ {I _ {i} \\leftarrow \\mathcal {W} _ {j}} = \\frac {1}{M - 1} \\sum_ {I _ {j} \\in \\mathcal {G} \\backslash I _ {i}} \\sum_ {v \\in \\mathcal {W} _ {j}} \\operatorname {S o f t m a x} \\left(a _ {u v} ^ {I _ {i} I _ {j}}\\right) \\mathbf {v} _ {v} ^ {I _ {j}}. \\tag {1}", + "image_path": "df8b9efc02bc8f44c6fa00c77e46bdd0050c79707578853ee705c9ab0468a816.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 266, + 295, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 266, + 295, + 366 + ], + "spans": [ + { + "bbox": [ + 55, + 266, + 295, + 366 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 55, + 266, + 295, + 366 + ], + "type": "inline_equation", + "content": "\\mathcal{W}_j" + }, + { + "bbox": [ + 55, + 266, + 295, + 366 + ], + "type": "text", + "content": " denotes the point set in the source view " + }, + { + "bbox": [ + 55, + 266, + 295, + 366 + ], + "type": "inline_equation", + "content": "I_{j}" + }, + { + "bbox": [ + 55, + 266, + 295, + 366 + ], + "type": "text", + "content": ", while " + }, + { + "bbox": [ + 55, + 266, + 295, + 366 + ], + "type": "inline_equation", + "content": "a_{uv}^{I_iI_j}" + }, + { + "bbox": [ + 55, + 266, + 295, + 366 + ], + "type": "text", + "content": " represents the similarity score computed by treating point " + }, + { + "bbox": [ + 55, + 266, + 295, + 366 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 55, + 266, + 295, + 366 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 55, + 266, + 295, + 366 + ], + "type": "inline_equation", + "content": "I_{i}" + }, + { + "bbox": [ + 55, + 266, + 295, + 366 + ], + "type": "text", + "content": " as the query and point " + }, + { + "bbox": [ + 55, + 266, + 295, + 366 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 55, + 266, + 295, + 366 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 55, + 266, + 295, + 366 + ], + "type": "inline_equation", + "content": "I_{j}" + }, + { + "bbox": [ + 55, + 266, + 295, + 366 + ], + "type": "text", + "content": " as the key. " + }, + { + "bbox": [ + 55, + 266, + 295, + 366 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_v^{I_j}" + }, + { + "bbox": [ + 55, + 266, + 295, + 366 + ], + "type": "text", + "content": " is the value of point " + }, + { + "bbox": [ + 55, + 266, + 295, + 366 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 55, + 266, + 295, + 366 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 55, + 266, + 295, + 366 + ], + "type": "inline_equation", + "content": "\\mathbf{m}_u^{I_i}" + }, + { + "bbox": [ + 55, + 266, + 295, + 366 + ], + "type": "text", + "content": " represents the message vector from " + }, + { + "bbox": [ + 55, + 266, + 295, + 366 + ], + "type": "inline_equation", + "content": "\\mathcal{W}_j" + }, + { + "bbox": [ + 55, + 266, + 295, + 366 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 266, + 295, + 366 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 55, + 266, + 295, + 366 + ], + "type": "text", + "content": " of the GNN [43]. We uniformly aggregate features from each source view, encouraging the model to integrate multi-view information comprehensively rather than overemphasizing images from similar perspectives." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 368, + 295, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 368, + 295, + 487 + ], + "spans": [ + { + "bbox": [ + 55, + 368, + 295, + 487 + ], + "type": "text", + "content": "However, for a point near an occlusion boundary in one source view, how does it identify the corresponding points observing the same location in other source views? This implies a matching process. While the network could directly learn this, it would introduce additional complexity, as establishing correspondences between source views is not our primary objective. Moreover, while multi-view setups provide abundant contextual information, much of it is irrelevant, such as regions lacking co-visibility. This redundancy also imposes an additional burden on learning." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 489, + 295, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 489, + 295, + 561 + ], + "spans": [ + { + "bbox": [ + 55, + 489, + 295, + 561 + ], + "type": "text", + "content": "To address this, we propose a geometric constraint mechanism to explicitly guide the attention range of each point. By incorporating relative positional encoding into the attention score computation, we enable the scores to depend on two critical factors: feature correlation and the geometric distance " + }, + { + "bbox": [ + 55, + 489, + 295, + 561 + ], + "type": "inline_equation", + "content": "\\Delta \\mathbf{p}" + }, + { + "bbox": [ + 55, + 489, + 295, + 561 + ], + "type": "text", + "content": " between two points:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 108, + 573, + 295, + 591 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 573, + 295, + 591 + ], + "spans": [ + { + "bbox": [ + 108, + 573, + 295, + 591 + ], + "type": "interline_equation", + "content": "a _ {u v} ^ {I _ {i} I _ {j}} = \\left(\\mathbf {q} _ {u} ^ {I _ {i}}\\right) ^ {\\top} \\mathbf {R} \\left[ \\Delta \\mathbf {p} _ {u v} ^ {I _ {i} I _ {j}} \\right] \\mathbf {k} _ {v} ^ {I _ {j}}. \\tag {2}", + "image_path": "682c1bb8cbc4b5c4c02cc3e196fa692b7135e86f1a98c6e310eb0014e4457133.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 601, + 249, + 614 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 601, + 249, + 614 + ], + "spans": [ + { + "bbox": [ + 55, + 601, + 249, + 614 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 55, + 601, + 249, + 614 + ], + "type": "inline_equation", + "content": "\\mathbf{R}[\\cdot ]" + }, + { + "bbox": [ + 55, + 601, + 249, + 614 + ], + "type": "text", + "content": " is the rotation encoding matrix [49]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 616, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 616, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 616, + 295, + 713 + ], + "type": "text", + "content": "While the 2D relative distance between two points within a single image is well-defined, across different views, it depends on the projection transformation. We use the precomputed tracks " + }, + { + "bbox": [ + 55, + 616, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{M}(\\mathcal{G})" + }, + { + "bbox": [ + 55, + 616, + 295, + 713 + ], + "type": "text", + "content": " to find the projection position of " + }, + { + "bbox": [ + 55, + 616, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_u^{I_i}" + }, + { + "bbox": [ + 55, + 616, + 295, + 713 + ], + "type": "text", + "content": " on " + }, + { + "bbox": [ + 55, + 616, + 295, + 713 + ], + "type": "inline_equation", + "content": "I_{j}" + }, + { + "bbox": [ + 55, + 616, + 295, + 713 + ], + "type": "text", + "content": ", which can be denoted as " + }, + { + "bbox": [ + 55, + 616, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_w^{I_j}" + }, + { + "bbox": [ + 55, + 616, + 295, + 713 + ], + "type": "text", + "content": ". When no corresponding point in " + }, + { + "bbox": [ + 55, + 616, + 295, + 713 + ], + "type": "inline_equation", + "content": "I_{j}" + }, + { + "bbox": [ + 55, + 616, + 295, + 713 + ], + "type": "text", + "content": " can be found for " + }, + { + "bbox": [ + 55, + 616, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_u^{I_i}" + }, + { + "bbox": [ + 55, + 616, + 295, + 713 + ], + "type": "text", + "content": ", we simply assign the relative position to " + }, + { + "bbox": [ + 55, + 616, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{0}" + }, + { + "bbox": [ + 55, + 616, + 295, + 713 + ], + "type": "text", + "content": " so that the attention score depends solely on the feature similarity. The calculation of the relative po" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 72, + 460, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 460, + 83 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 460, + 83 + ], + "type": "text", + "content": "sition can be formulated as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 343, + 91, + 553, + 125 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 343, + 91, + 553, + 125 + ], + "spans": [ + { + "bbox": [ + 343, + 91, + 553, + 125 + ], + "type": "interline_equation", + "content": "\\Delta \\mathbf {p} _ {u v} ^ {I _ {i} I _ {j}} = \\left\\{ \\begin{array}{l l} \\mathbf {p} _ {w} ^ {I _ {j}} - \\mathbf {p} _ {v} ^ {I _ {j}} & (u, w) \\in M (I _ {i}, I _ {j}) \\\\ \\mathbf {0} & (u, \\emptyset) \\in M (I _ {i}, I _ {j}). \\end{array} \\right. \\tag {3}", + "image_path": "466deb88e1a80d28a73b836d2e75fca0d241f7e641b3248018934125cf3c35fd.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 133, + 555, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 133, + 555, + 264 + ], + "spans": [ + { + "bbox": [ + 313, + 133, + 555, + 264 + ], + "type": "text", + "content": "Additionally, we also interact with all target view features after source cross-attention (Target Cross). For a point " + }, + { + "bbox": [ + 313, + 133, + 555, + 264 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 133, + 555, + 264 + ], + "type": "text", + "content": " in view " + }, + { + "bbox": [ + 313, + 133, + 555, + 264 + ], + "type": "inline_equation", + "content": "I_{t}" + }, + { + "bbox": [ + 313, + 133, + 555, + 264 + ], + "type": "text", + "content": " paired with source view " + }, + { + "bbox": [ + 313, + 133, + 555, + 264 + ], + "type": "inline_equation", + "content": "I_{i}" + }, + { + "bbox": [ + 313, + 133, + 555, + 264 + ], + "type": "text", + "content": ", we attend to the same point " + }, + { + "bbox": [ + 313, + 133, + 555, + 264 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 133, + 555, + 264 + ], + "type": "text", + "content": " across the remaining " + }, + { + "bbox": [ + 313, + 133, + 555, + 264 + ], + "type": "inline_equation", + "content": "M - 1" + }, + { + "bbox": [ + 313, + 133, + 555, + 264 + ], + "type": "text", + "content": " target views. This is analogous to operations in some works that aggregate features along the temporal dimension [12, 26]. Unlike the approach for source views, we compute the attention scores solely based on feature similarity. Despite its low computational complexity, this module complements the Source Cross, mutually enhancing each other to help the network implicitly infer globally optimal correspondences." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 273, + 479, + 284 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 273, + 479, + 284 + ], + "spans": [ + { + "bbox": [ + 314, + 273, + 479, + 284 + ], + "type": "text", + "content": "3.3. Multi-view feature correlation" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 290, + 555, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 290, + 555, + 415 + ], + "spans": [ + { + "bbox": [ + 313, + 290, + 555, + 415 + ], + "type": "text", + "content": "For a track in a source view set originating from a 3D point, its corresponding 2D keypoints " + }, + { + "bbox": [ + 313, + 290, + 555, + 415 + ], + "type": "inline_equation", + "content": "\\left\\{\\mathbf{p}_{u}^{I_{i}}, \\mathbf{p}_{v}^{I_{j}}, \\ldots\\right\\}" + }, + { + "bbox": [ + 313, + 290, + 555, + 415 + ], + "type": "text", + "content": " should exhibit consistent matches in the target view—either corresponding to the same single point or having no corresponding point at all. This physical constraint is often utilized by existing methods as a post-processing step to filter out erroneous results when merging multiple two-view correspondences. However, as a multi-view matcher, we aim for CoMatcher to inherently satisfy this constraint in its reasoning, thereby enhancing its reliability and confidence." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 415, + 554, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 415, + 554, + 487 + ], + "spans": [ + { + "bbox": [ + 313, + 415, + 554, + 487 + ], + "type": "text", + "content": "Achieving this involves two key processes: (i) identifying points that may lead to errors in the early layers of the network, and (ii) promptly guiding these points toward correct estimation using inference information from other views. To accomplish this, we propose a two-step multiview feature correlation strategy (see Fig. 3)." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 487, + 554, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 487, + 554, + 523 + ], + "spans": [ + { + "bbox": [ + 313, + 487, + 554, + 523 + ], + "type": "text", + "content": "First, to identify ambiguous points in the source views, we utilize a lightweight head at the end of each layer to predict the confidence of each point:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 357, + 532, + 553, + 548 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 357, + 532, + 553, + 548 + ], + "spans": [ + { + "bbox": [ + 357, + 532, + 553, + 548 + ], + "type": "interline_equation", + "content": "c _ {u} ^ {I _ {i}} = \\operatorname {S i g m o i d} \\left(\\operatorname {M L P} \\left(\\mathbf {f} _ {u} ^ {I _ {i}}\\right)\\right) \\in [ 0, 1 ]. \\tag {4}", + "image_path": "cfc1bddd6c90a73f6fbb01fc8449fbd0dcaa7b5ff8b402aded2288e288c6d4b8.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 557, + 554, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 557, + 554, + 617 + ], + "spans": [ + { + "bbox": [ + 313, + 557, + 554, + 617 + ], + "type": "text", + "content": "Here, MLP represents a multi-layer perceptron and " + }, + { + "bbox": [ + 313, + 557, + 554, + 617 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_u^{I_i}" + }, + { + "bbox": [ + 313, + 557, + 554, + 617 + ], + "type": "text", + "content": " is the feature of " + }, + { + "bbox": [ + 313, + 557, + 554, + 617 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 313, + 557, + 554, + 617 + ], + "type": "text", + "content": ". This confidence score reflects the assessment of each point in its current state: higher values indicate that, at this layer, the point is either converging toward a reliable match or confidently identified as having no match." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 617, + 554, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 554, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 554, + 689 + ], + "type": "text", + "content": "In each layer, we define a distinct hyperparameter threshold " + }, + { + "bbox": [ + 313, + 617, + 554, + 689 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 617, + 554, + 689 + ], + "type": "text", + "content": ". Points with confidence scores below this threshold are identified as potentially ambiguous and likely to lead to incorrect inference. We progressively increase the value of " + }, + { + "bbox": [ + 313, + 617, + 554, + 689 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 617, + 554, + 689 + ], + "type": "text", + "content": " in later layers, as the confidence of overall inference generally improves with each successive layer." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "content": "Next, we redesign the parallel two-view cross-attention modules to correct the erroneous estimations of ambiguous" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21973" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 70, + 69, + 289, + 167 + ], + "blocks": [ + { + "bbox": [ + 70, + 69, + 289, + 167 + ], + "lines": [ + { + "bbox": [ + 70, + 69, + 289, + 167 + ], + "spans": [ + { + "bbox": [ + 70, + 69, + 289, + 167 + ], + "type": "image", + "image_path": "96f4098fe3dbd88256aec1c84d929a9546d018e9a5100a1ccf9d62bdd6564e0f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 178, + 296, + 225 + ], + "lines": [ + { + "bbox": [ + 55, + 178, + 296, + 225 + ], + "spans": [ + { + "bbox": [ + 55, + 178, + 296, + 225 + ], + "type": "text", + "content": "Figure 3. The multi-view feature correlation strategy consists of two steps: we first identify potentially erroneous points using a confidence estimator, and then correct their estimation by aggregating attention correlation vectors from other views." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 244, + 296, + 269 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 244, + 296, + 269 + ], + "spans": [ + { + "bbox": [ + 55, + 244, + 296, + 269 + ], + "type": "text", + "content": "points. In these modules, each point in the source view computes an attention distribution vector to query [54]:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 126, + 276, + 295, + 297 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 126, + 276, + 295, + 297 + ], + "spans": [ + { + "bbox": [ + 126, + 276, + 295, + 297 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\alpha} _ {u} ^ {I _ {i}} = \\underset {x \\in \\mathcal {W} _ {t}} {\\operatorname {S o f t m a x}} \\left(a _ {u x} ^ {I _ {i} I _ {t}}\\right). \\tag {5}", + "image_path": "3b20928b022c15a768f53e88a84f6c93145fb5cf863fb11f24efbdf1c2d5aca8.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 304, + 296, + 394 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 304, + 296, + 394 + ], + "spans": [ + { + "bbox": [ + 55, + 304, + 296, + 394 + ], + "type": "text", + "content": "This vector represents the degree to which the point attends to in view " + }, + { + "bbox": [ + 55, + 304, + 296, + 394 + ], + "type": "inline_equation", + "content": "I_{t}" + }, + { + "bbox": [ + 55, + 304, + 296, + 394 + ], + "type": "text", + "content": ", with a peak indicating potential matching regions of interest. For ambiguous points, this similarity is not reliable, often leading to incorrect feature aggregation. Assuming that point " + }, + { + "bbox": [ + 55, + 304, + 296, + 394 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 55, + 304, + 296, + 394 + ], + "type": "text", + "content": " in the source view " + }, + { + "bbox": [ + 55, + 304, + 296, + 394 + ], + "type": "inline_equation", + "content": "I_{i}" + }, + { + "bbox": [ + 55, + 304, + 296, + 394 + ], + "type": "text", + "content": " is an ambiguous point, we update its original attention distribution by applying a weighted average to the distributions of its corresponding points across multi-view:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 399, + 295, + 435 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 399, + 295, + 435 + ], + "spans": [ + { + "bbox": [ + 62, + 399, + 295, + 435 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\alpha} _ {u} ^ {I _ {i} ^ {\\prime}} = c _ {u} ^ {I _ {i}} \\boldsymbol {\\alpha} _ {u} ^ {I _ {i}} + \\left(1 - c _ {u} ^ {I _ {i}}\\right) \\frac {\\sum_ {v \\in \\mathcal {D} _ {m} ^ {u}} c _ {v} ^ {I _ {j}} \\boldsymbol {\\alpha} _ {v} ^ {I _ {j}}}{\\sum_ {v \\in \\mathcal {D} _ {m} ^ {u}} c _ {v} ^ {I _ {j}}} \\quad \\text {s . t .} c _ {u} ^ {I _ {i}} < \\theta . \\tag {6}", + "image_path": "49082eb6109e38f5239a69a23c9b28aaeb695b84baa6e720d52e6f88fc54c40c.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 441, + 296, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 441, + 296, + 489 + ], + "spans": [ + { + "bbox": [ + 55, + 441, + 296, + 489 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 55, + 441, + 296, + 489 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_m^u" + }, + { + "bbox": [ + 55, + 441, + 296, + 489 + ], + "type": "text", + "content": " represents other points in the track. We embed this strategy into the model inference process, which significantly enhances the reliability of estimation under challenging conditions despite its simple implementation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 496, + 206, + 510 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 496, + 206, + 510 + ], + "spans": [ + { + "bbox": [ + 55, + 496, + 206, + 510 + ], + "type": "text", + "content": "3.4. Correspondence prediction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 514, + 296, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 514, + 296, + 658 + ], + "spans": [ + { + "bbox": [ + 55, + 514, + 296, + 658 + ], + "type": "text", + "content": "As shown in Fig. 2, a lightweight matching head operates in parallel for each pair of network output features to predict correspondences between two views [30, 43]. In each head, we begin by calculating a score matrix from feature correlations between two views: " + }, + { + "bbox": [ + 55, + 514, + 296, + 658 + ], + "type": "inline_equation", + "content": "\\mathbf{S}(u,x) = \\mathrm{L}(\\mathbf{f}_u^{I_i})\\cdot" + }, + { + "bbox": [ + 55, + 514, + 296, + 658 + ], + "type": "inline_equation", + "content": "\\mathrm{L}(\\mathbf{f}_x^{I_t})" + }, + { + "bbox": [ + 55, + 514, + 296, + 658 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 514, + 296, + 658 + ], + "type": "inline_equation", + "content": "\\mathrm{L}" + }, + { + "bbox": [ + 55, + 514, + 296, + 658 + ], + "type": "text", + "content": " is the linear transform. Then we apply dual-softmax [30, 50, 53] to derive the matching probabilities of two points: " + }, + { + "bbox": [ + 55, + 514, + 296, + 658 + ], + "type": "inline_equation", + "content": "\\mathbf{S}'(u,x) = \\mathrm{Softmax}(\\mathbf{S}(u,\\cdot))_x\\cdot" + }, + { + "bbox": [ + 55, + 514, + 296, + 658 + ], + "type": "text", + "content": " Softmax " + }, + { + "bbox": [ + 55, + 514, + 296, + 658 + ], + "type": "inline_equation", + "content": "(\\mathbf{S}(\\cdot ,x))_u" + }, + { + "bbox": [ + 55, + 514, + 296, + 658 + ], + "type": "text", + "content": ". Additionally, we employ the method proposed in [30] to calculate the matching probability for each point: " + }, + { + "bbox": [ + 55, + 514, + 296, + 658 + ], + "type": "inline_equation", + "content": "\\sigma_{i}^{I_{i}} = \\mathrm{Sigmoid}(\\mathrm{L}(\\mathbf{f}_{i}^{I_{i}}))" + }, + { + "bbox": [ + 55, + 514, + 296, + 658 + ], + "type": "text", + "content": ". The final assignment matrix is a combination of the two probabilities:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 120, + 666, + 295, + 681 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 120, + 666, + 295, + 681 + ], + "spans": [ + { + "bbox": [ + 120, + 666, + 295, + 681 + ], + "type": "interline_equation", + "content": "\\mathbf {P} (u, x) = \\mathbf {S} ^ {\\prime} (u, x) \\sigma_ {u} ^ {I _ {i}} \\sigma_ {x} ^ {I _ {t}}. \\tag {7}", + "image_path": "d2372ae820942ddb9aead28560cc41e1f7260ba03928a04a1a1ded5bff99ff41.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": "Using mutual nearest neighbors and a threshold, we filter the matching results from the assignment matrix [43, 50]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 72, + 359, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 359, + 83 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 359, + 83 + ], + "type": "text", + "content": "3.5. Loss" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 89, + 553, + 125 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 89, + 553, + 125 + ], + "spans": [ + { + "bbox": [ + 313, + 89, + 553, + 125 + ], + "type": "text", + "content": "CoMatcher supervises each source-target image pair individually, with the loss for each pairs composed of two components: correspondence and confidence:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 344, + 133, + 555, + 163 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 344, + 133, + 555, + 163 + ], + "spans": [ + { + "bbox": [ + 344, + 133, + 555, + 163 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {t o t a l}} = \\frac {1}{M} \\sum_ {I _ {i} \\in \\mathcal {G}} \\left(\\mathcal {L} _ {\\text {c o r r}} \\left(I _ {i}, I _ {t}\\right) + \\alpha \\mathcal {L} _ {\\text {c o n f}} \\left(I _ {i}\\right)\\right). \\tag {8}", + "image_path": "4dd9db0eda95998ed49866ff1b9f6fde0f70b1c118fef8a6504a0c0c623b9d3a.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 170, + 555, + 241 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 170, + 555, + 241 + ], + "spans": [ + { + "bbox": [ + 313, + 170, + 555, + 241 + ], + "type": "text", + "content": "Correspondence supervision. For each pair " + }, + { + "bbox": [ + 313, + 170, + 555, + 241 + ], + "type": "inline_equation", + "content": "I_{i}" + }, + { + "bbox": [ + 313, + 170, + 555, + 241 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 170, + 555, + 241 + ], + "type": "inline_equation", + "content": "I_{t}" + }, + { + "bbox": [ + 313, + 170, + 555, + 241 + ], + "type": "text", + "content": ", we perform two-view transformations using relative poses or homography to compute ground truth labels [30, 43, 50]. We minimize the negative log-likelihood of the assignment matrix, following [30]. More details are provided in the supplementary materials." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 242, + 554, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 242, + 554, + 350 + ], + "spans": [ + { + "bbox": [ + 313, + 242, + 554, + 350 + ], + "type": "text", + "content": "Confidence supervision. To train the confidence estimators at each layer, we define the point confidence as the consistency probability between its correspondence estimated at the current layer and the final estimation [16, 30, 47]. At each layer, predictions are obtained via dual-softmax on current features, without introducing additional matching probabilities or parameters. The ground truth label " + }, + { + "bbox": [ + 313, + 242, + 554, + 350 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 313, + 242, + 554, + 350 + ], + "type": "text", + "content": " indicates whether the two estimations are consistent. This is supervised using a binary cross-entropy loss:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 344, + 357, + 554, + 387 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 344, + 357, + 554, + 387 + ], + "spans": [ + { + "bbox": [ + 344, + 357, + 554, + 387 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {c o n f}} \\left(I _ {i}\\right) = \\frac {1}{L - 1} \\sum_ {\\ell} \\sum_ {u \\in \\mathcal {W} _ {i}} \\mathrm {C E} \\left(\\ell c _ {u} ^ {I _ {i}}, \\ell y _ {u} ^ {I _ {i}}\\right), \\tag {9}", + "image_path": "d172d827151e224b885563340bfedb323cebc8379e2690775c2849198e3a5f4d.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 314, + 395, + 424, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 395, + 424, + 407 + ], + "spans": [ + { + "bbox": [ + 314, + 395, + 424, + 407 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 314, + 395, + 424, + 407 + ], + "type": "inline_equation", + "content": "\\ell \\in \\{1,\\dots ,L - 1\\}" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 314, + 418, + 395, + 431 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 418, + 395, + 431 + ], + "spans": [ + { + "bbox": [ + 314, + 418, + 395, + 431 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 437, + 554, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 554, + 521 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 554, + 521 + ], + "type": "text", + "content": "In this section, we first introduce the datasets used, followed by our implementation details. Then, our CoMatcher network is compared to previous state-of-the-art baselines for homography and camera pose estimation. Next, we integrate CoMatcher into groupwise matching framework and evaluate it against pairwise method on a large-scale benchmark. Finally, an extensive ablation study is provided." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 314, + 529, + 376, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 529, + 376, + 540 + ], + "spans": [ + { + "bbox": [ + 314, + 529, + 376, + 540 + ], + "type": "text", + "content": "4.1. Datasets" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "type": "text", + "content": "The HPatches dataset is used for homography estimation [3]. It consists of 116 strictly planar scenes, each containing 6 images with variations in viewpoint and illumination. For camera pose estimation tasks, we utilize the MegaDepth dataset, an outdoor dataset that exhibits strong occlusions and significant structural changes [28]. We selected two scenes, \"Sacre Coeur\" and \"St. Peter's Square\", from which 1500 co-visible quadruplets were sampled in a way that balances difficulty based on visual overlap [30, 50, 56]. The Image Matching Challenge 2020 benchmark provides a comprehensive evaluation protocol, including datasets that cover multiple challenging outdoor wide-baseline scenes. From the phototourism dataset, we selected 3 validated scenes, each consisting of about 100" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21974" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 69, + 70, + 541, + 412 + ], + "blocks": [ + { + "bbox": [ + 69, + 70, + 541, + 412 + ], + "lines": [ + { + "bbox": [ + 69, + 70, + 541, + 412 + ], + "spans": [ + { + "bbox": [ + 69, + 70, + 541, + 412 + ], + "type": "image", + "image_path": "e497676a2d20632109343329ef50e38b1f36b268cef778dee4164bac065095f3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 422, + 555, + 468 + ], + "lines": [ + { + "bbox": [ + 55, + 422, + 555, + 468 + ], + "spans": [ + { + "bbox": [ + 55, + 422, + 555, + 468 + ], + "type": "text", + "content": "Figure 4. Qualitative comparison on MegaDepth. For each quadruple with a target view (top), correspondences (green for correct, red for incorrect) predicted by CoMatcher and LightGlue are shown. Using identical local features as input, CoMatcher achieves significantly more reliable estimations, even at challenging semantic edges with depth discontinuities. This stems from its holistic scene understanding through multi-view cues to effectively address occlusions (Q3), large-scale variations (Q2, 4), and repetitive textures (Q1, 2, 3, 4)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 487, + 295, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 487, + 295, + 513 + ], + "spans": [ + { + "bbox": [ + 55, + 487, + 295, + 513 + ], + "type": "text", + "content": "images. Additionally, we conducted an ablation study on a synthetic homography dataset [37] used for training." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 517, + 187, + 530 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 517, + 187, + 530 + ], + "spans": [ + { + "bbox": [ + 55, + 517, + 187, + 530 + ], + "type": "text", + "content": "4.2. Implementation details" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 535, + 295, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 535, + 295, + 666 + ], + "spans": [ + { + "bbox": [ + 55, + 535, + 295, + 666 + ], + "type": "text", + "content": "CoMatcher is trained in two steps, following [30, 43]. We first pre-train the model on a large-scale synthetic homography dataset [37], leveraging noise-free homography for ground truth. Next, fine-tuning is performed on the MegaDepth dataset [28] using the camera poses recovered by SfM as ground truth. We sample 200 co-visible multiview groups per scene and randomly select one image as the target view. The size of source views is set to " + }, + { + "bbox": [ + 55, + 535, + 295, + 666 + ], + "type": "inline_equation", + "content": "M = 4" + }, + { + "bbox": [ + 55, + 535, + 295, + 666 + ], + "type": "text", + "content": " during training. The training process is carried out on two NVIDIA GeForce 4090 GPUs, taking about 6 days in total. More details can be found in the supplementary materials." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 672, + 192, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 672, + 192, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 672, + 192, + 685 + ], + "type": "text", + "content": "4.3. Homography estimation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "content": "CoMatcher is compared with three types of baselines. First, we include state-of-the-art (SOTA) sparse two-view" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 487, + 555, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 487, + 555, + 668 + ], + "spans": [ + { + "bbox": [ + 313, + 487, + 555, + 668 + ], + "type": "text", + "content": "matchers: nearest-neighbor with mutual check [31], SuperGlue [43], SGMNet [7], and LightGlue [30], each paired with different feature extractors [10, 31, 53]. Second, we evaluate the multi-view matcher End2End [40]. For reference, we also compare against typical dense matchers LoFTR [50] and PDC-Net [52]. For CoMatcher and End2End, we use a single forward pass to obtain matches across all five image pairs, while others follow a pairwise approach. To ensure a fair comparison, we adopt the setups of [30, 50] for the number of features and image resizing. Homography accuracy is evaluated using both robust (RANSAC [17]) and non-robust (weighted DLT [20]) solvers. For each pair, we compute the mean reprojection error of the four image corners and report the area under the cumulative error curve (AUC) up to 1px, 3px, and 5px." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "content": "Tab. 1 shows that CoMatcher yields more accurate estimates than all sparse two-view matchers, highlighting the advantage of inference in multi-view feature space. How-" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21975" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 70, + 300, + 270 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 300, + 270 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 300, + 270 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 300, + 270 + ], + "type": "table", + "html": "
MethodAUC - DLT @1px / @3px / @5pxAUC - RANSAC
denseLoFTR38.5 / 66.0 / 71.440.7 / 68.3 / 78.5
PDC-Net36.0 / 65.3 / 73.037.9 / 67.6 / 77.4
sparse 2-viewSIFT+NN+mutual0.0 / 0.0 / 0.035.9 / 65.0 / 75.6
SP+NN+mutual0.0 / 1.9 / 3.434.8 / 64.1 / 74.8
SP+SuperGlue32.2 / 65.1 / 75.737.2 / 68.0 / 78.7
SP+SGMNet31.7 / 64.9 / 76.037.7 / 66.4 / 77.5
SP+LightGlue35.4 / 67.5 / 77.737.2 / 67.8 / 78.1
DISK+NN+mutual1.8 / 5.2 / 7.837.9 / 58.0 / 68.3
DISK+LightGlue34.4 / 64.5 / 74.438.1 / 65.2 / 77.2
multi-viewSP+End2End34.3 / 66.9 / 75.537.0 / 67.2 / 77.5
SP+CoMatcher37.1 / 69.0 / 78.838.4 / 68.9 / 79.0
DISK+CoMatcher36.3 / 66.2 / 75.938.7 / 68.2 / 78.4
", + "image_path": "eb94d5444d3f00378b91db5c8ba92a699aebc0132eda6259692adfdef7ee3ada.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 335, + 296, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 335, + 296, + 456 + ], + "spans": [ + { + "bbox": [ + 55, + 335, + 296, + 456 + ], + "type": "text", + "content": "ever, despite End2End also jointly reasoning over multiple views like ours, its performance is notably inferior. This underscores the superiority of a " + }, + { + "bbox": [ + 55, + 335, + 296, + 456 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 335, + 296, + 456 + ], + "type": "text", + "content": "-to-1 architecture over " + }, + { + "bbox": [ + 55, + 335, + 296, + 456 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 335, + 296, + 456 + ], + "type": "text", + "content": "-to-" + }, + { + "bbox": [ + 55, + 335, + 296, + 456 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 335, + 296, + 456 + ], + "type": "text", + "content": " for multi-view representation learning. When compared to LoFTR, we observe slightly lower performance under low thresholds. We attribute this mainly to the limitations of the feature extractor in keypoint non-repeatability and localization errors [29]. Additionally, DLT achieves accuracy close to RANSAC on most metrics, reflecting the high-quality correspondences from CoMatcher." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 466, + 192, + 479 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 466, + 192, + 479 + ], + "spans": [ + { + "bbox": [ + 55, + 466, + 192, + 479 + ], + "type": "text", + "content": "4.4. Relative pose estimation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 485, + 295, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 485, + 295, + 652 + ], + "spans": [ + { + "bbox": [ + 55, + 485, + 295, + 652 + ], + "type": "text", + "content": "Next, we evaluate CoMatcher on challenging wide-baseline scenes, using the same baselines as in Sec. 4.3. For each quadruplet, including a selected target view, we still use a single forward pass for multi-view methods. Additionally, we test two newer dense methods, DUS3R [55] and MAS3R [27], following their evaluation setups for image resizing. All other settings remain consistent with [30]. The essential matrix is estimated from the correspondences using both vanilla RANSAC [17] and LO-RANSAC [9], following [30]. The pose error is then computed as the maximum angular error of rotation and translation, derived from the decomposition of the essential matrix. We report the AUC at " + }, + { + "bbox": [ + 55, + 485, + 295, + 652 + ], + "type": "inline_equation", + "content": "5^{\\circ}" + }, + { + "bbox": [ + 55, + 485, + 295, + 652 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 485, + 295, + 652 + ], + "type": "inline_equation", + "content": "10^{\\circ}" + }, + { + "bbox": [ + 55, + 485, + 295, + 652 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 485, + 295, + 652 + ], + "type": "inline_equation", + "content": "20^{\\circ}" + }, + { + "bbox": [ + 55, + 485, + 295, + 652 + ], + "type": "text", + "content": ", and record the average runtime of matching each quadruplet on a single 4090 GPU." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": "Tab. 2 shows that CoMatcher, as a sparse method, comprehensively enhances the estimation performance on different local features [10] and [53]. Importantly, compared to other sparse two-view matchers like LightGlue, a key strength of CoMatcher is its reliability, as illustrated in" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 319, + 70, + 550, + 269 + ], + "blocks": [ + { + "bbox": [ + 55, + 278, + 295, + 312 + ], + "lines": [ + { + "bbox": [ + 55, + 278, + 295, + 312 + ], + "spans": [ + { + "bbox": [ + 55, + 278, + 295, + 312 + ], + "type": "text", + "content": "Table 1. Homography estimation on HPatches. We report the area under the cumulative error curve (AUC) up to values of 1px, 3px and 5px, using DLT and RANSAC [17] as homography solver." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 319, + 70, + 550, + 269 + ], + "lines": [ + { + "bbox": [ + 319, + 70, + 550, + 269 + ], + "spans": [ + { + "bbox": [ + 319, + 70, + 550, + 269 + ], + "type": "table", + "html": "
MethodAUC RANSACAUC LO-RANSACTime (ms)
5° / 10° / 20°
denseLoFTR58.0 / 73.1 / 84.467.4 / 81.7 / 89.3182
PDCNet54.7 / 73.1 / 83.567.1 / 80.2 / 87.0231
DUSt3R42.4 / 56.7 / 64.258.1 / 70.4 / 79.6264
MASt3R51.5 / 65.3 / 75.663.5 / 76.3 / 85.2317
SuperPointNN+mutual35.3 / 58.3 / 53.751.4 / 67.3 / 75.99
SuperGlue55.8 / 72.8 / 84.165.1 / 77.2 / 89.287
LightGlue56.2 / 72.7 / 83.567.2 / 80.1 / 88.051
End2End55.3 / 71.4 / 81.267.4 / 81.5 / 87.0152
CoMatcher57.2 / 73.9 / 84.868.3 / 82.2 / 89.169
DISKNN+mutual50.9 / 66.7 / 77.764.0 / 79.5 / 87.69
LightGlue53.2 / 69.2 / 80.268.6 / 80.4 / 87.254
CoMatcher54.9 / 71.2 / 82.068.5 / 82.1 / 88.473
", + "image_path": "3897fd00562f5457995ad0f9e6cadd26d3ba8400d25c4c28f661f494f071a7ff.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 278, + 555, + 312 + ], + "lines": [ + { + "bbox": [ + 313, + 278, + 555, + 312 + ], + "spans": [ + { + "bbox": [ + 313, + 278, + 555, + 312 + ], + "type": "text", + "content": "Table 2. Relative pose estimation on MegaDepth. We report the AUC at " + }, + { + "bbox": [ + 313, + 278, + 555, + 312 + ], + "type": "inline_equation", + "content": "5^{\\circ}" + }, + { + "bbox": [ + 313, + 278, + 555, + 312 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 278, + 555, + 312 + ], + "type": "inline_equation", + "content": "10^{\\circ}" + }, + { + "bbox": [ + 313, + 278, + 555, + 312 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 278, + 555, + 312 + ], + "type": "inline_equation", + "content": "20^{\\circ}" + }, + { + "bbox": [ + 313, + 278, + 555, + 312 + ], + "type": "text", + "content": " using different robust estimator, and the average runtime of matching each quadruplet." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 335, + 555, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 335, + 555, + 456 + ], + "spans": [ + { + "bbox": [ + 313, + 335, + 555, + 456 + ], + "type": "text", + "content": "Fig. 3. This highlights the advantage of multi-view collaborative reasoning in understanding complex 3D structures, such as occlusions, even with suboptimal input local features in these areas. Additionally, compared to End2End, CoMatcher achieves more accurate relative poses without the need for cumbersome end-to-end training. While some dense methods may achieve better accuracy on certain metrics, their efficiency is generally much lower. By employing a space-for-time strategy to minimize inference steps, Co-Matcher runs significantly faster than most methods." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 466, + 526, + 478 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 466, + 526, + 478 + ], + "spans": [ + { + "bbox": [ + 313, + 466, + 526, + 478 + ], + "type": "text", + "content": "4.5. Evaluation on the IMC 2020 benchmark" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 485, + 555, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 485, + 555, + 664 + ], + "spans": [ + { + "bbox": [ + 313, + 485, + 555, + 664 + ], + "type": "text", + "content": "Next, we evaluate our groupwise framework against pairwise baselines [30, 31, 43] on the Image Matching Challenge (IMC) 2020 benchmark [25]. Given a large-scale unordered image collection, the benchmark requires providing matching results for all image pairs, which are then assessed on two downstream tasks: stereo and multi-view reconstruction. For the stereo task, the accuracy of relative camera poses for each pair is evaluated from correspondences using RANSAC, as in Sec. 4.4. For the multi-view task, all correspondences are fed into COLMAP [45] for SfM, with the final accuracy evaluated based on the estimated multi-view camera poses. We report the AUC at " + }, + { + "bbox": [ + 313, + 485, + 555, + 664 + ], + "type": "inline_equation", + "content": "5^{\\circ}" + }, + { + "bbox": [ + 313, + 485, + 555, + 664 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 485, + 555, + 664 + ], + "type": "inline_equation", + "content": "10^{\\circ}" + }, + { + "bbox": [ + 313, + 485, + 555, + 664 + ], + "type": "text", + "content": " across both tasks. Additionally, the average runtime is reported, which is the total matching time for the entire image set divided by the number of pairs." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "content": "Tab. 3 shows that our framework significantly outperforms existing pairwise approaches in both tasks. This advantage primarily stems from our framework's ability to better leverage the relationships within the original image" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21976" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 61, + 70, + 292, + 211 + ], + "blocks": [ + { + "bbox": [ + 61, + 70, + 292, + 211 + ], + "lines": [ + { + "bbox": [ + 61, + 70, + 292, + 211 + ], + "spans": [ + { + "bbox": [ + 61, + 70, + 292, + 211 + ], + "type": "table", + "html": "
MethodStereoMulti-ViewTime (ms)
AUC 5°/10°AUC 5°/10°
two-viewSIFT+NN+mutual31.5 / 44.257.2 / 68.53
SP+NN+mutual28.6 / 40.352.9 / 63.43
SP+SuperGlue36.5 / 50.162.3 / 74.832
SP+LightGlue36.8 / 49.464.6 / 75.417
DISK+NN+mutual35.4 / 47.459.3 / 70.23
DISK+LightGlue42.1 / 55.665.2 / 76.219
m-vSP+CoMatcher39.1 / 52.466.2 / 77.525
DISK+CoMatcher44.9 / 57.367.1 / 78.429
", + "image_path": "f2a31f32adfe001ee1ae82b4dff0a2fd9ef41cc93c2d5ce8e6a34e7e60b8e0ba.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 275, + 296, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 275, + 296, + 348 + ], + "spans": [ + { + "bbox": [ + 55, + 275, + 296, + 348 + ], + "type": "text", + "content": "collection, enabling the generation of higher-quality tracks in such large-scale tasks. Additional comparisons on track quality for SfM tasks are provided in the supplementary material. In terms of efficiency, our approach is faster than SuperGlue, with the viewpoint grouping process accounting for only about " + }, + { + "bbox": [ + 55, + 275, + 296, + 348 + ], + "type": "inline_equation", + "content": "4\\%" + }, + { + "bbox": [ + 55, + 275, + 296, + 348 + ], + "type": "text", + "content": " of the total matching time." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 358, + 148, + 371 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 358, + 148, + 371 + ], + "spans": [ + { + "bbox": [ + 55, + 358, + 148, + 371 + ], + "type": "text", + "content": "4.6. Ablation study" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 376, + 295, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 376, + 295, + 484 + ], + "spans": [ + { + "bbox": [ + 55, + 376, + 295, + 484 + ], + "type": "text", + "content": "Understanding CoMatcher. The CoMatcher network introduces three key components: multi-view cross-attention modules, a geometric constraint mechanism, and a multiview feature interaction strategy. We validate these design choices by evaluating on two benchmarks: the challenging synthetic homography dataset [37], where precision and recall metrics are employed, and HPatches [3], where the AUC of homography estimation via weighted DLT is measured. For each image, 512 keypoints are extracted." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 485, + 295, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 485, + 295, + 664 + ], + "spans": [ + { + "bbox": [ + 55, + 485, + 295, + 664 + ], + "type": "text", + "content": "Tab. 4 shows the ablation results. Excluding the two cross-attention modules for multi-view interaction reduces the model to a series of parallel two-view matchers [30, 43]. This simplification limits the model's ability to integrate multi-view context, leading to a significant performance decline. Without the geometric constraint mechanism, the model performs even worse than in the previous case. This suggests that direct feature aggregation from multi-view is challenging, as the network requires additional learning to distinguish meaningful context effectively. Without multiview feature correlation, the model loses its ability to guide inference for ambiguous points by leveraging consistency constraints. This highlights the importance of guiding two-view feature correlation in the attention search process, a factor overlooked in previous work." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "Impact of the group size. We evaluated the matching performance on the synthetic homography dataset [37] by varying the size of the source view sets used as network input. As shown in Fig. 5, performance initially improves signif" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 317, + 76, + 552, + 188 + ], + "blocks": [ + { + "bbox": [ + 317, + 76, + 552, + 188 + ], + "lines": [ + { + "bbox": [ + 317, + 76, + 552, + 188 + ], + "spans": [ + { + "bbox": [ + 317, + 76, + 552, + 188 + ], + "type": "image", + "image_path": "5bae3bf10a053c632f3bdc281e2fbc13ae7ba2abcd30da28f2ba6e9de42ced8d.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 198, + 555, + 232 + ], + "lines": [ + { + "bbox": [ + 313, + 198, + 555, + 232 + ], + "spans": [ + { + "bbox": [ + 313, + 198, + 555, + 232 + ], + "type": "text", + "content": "Figure 5. Impact of the group size. Under different numbers of keypoints, the AUC initially increases as the source view set expands, then tends to stabilize or shows a downward trend." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 326, + 243, + 544, + 335 + ], + "blocks": [ + { + "bbox": [ + 55, + 219, + 295, + 253 + ], + "lines": [ + { + "bbox": [ + 55, + 219, + 295, + 253 + ], + "spans": [ + { + "bbox": [ + 55, + 219, + 295, + 253 + ], + "type": "text", + "content": "Table 3. Evaluation on the IMC 2020 benchmark. We report the pose AUC at " + }, + { + "bbox": [ + 55, + 219, + 295, + 253 + ], + "type": "inline_equation", + "content": "5^{\\circ}" + }, + { + "bbox": [ + 55, + 219, + 295, + 253 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 219, + 295, + 253 + ], + "type": "inline_equation", + "content": "10^{\\circ}" + }, + { + "bbox": [ + 55, + 219, + 295, + 253 + ], + "type": "text", + "content": " for two subtasks, stereo and multi-view reconstruction, along with the average matching runtime." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 326, + 243, + 544, + 335 + ], + "lines": [ + { + "bbox": [ + 326, + 243, + 544, + 335 + ], + "spans": [ + { + "bbox": [ + 326, + 243, + 544, + 335 + ], + "type": "table", + "html": "
CoMatcherSyntheticHpatches-AUC
precisionrecall@1px / @5px
w/o M-V feat. interaction89.796.632.4 / 75.1
w/o attenu. propagation87.794.231.5 / 73.7
w/o M-V feat. correlation90.595.333.1 / 74.9
full92.798.934.7 / 77.1
", + "image_path": "d728c9751caacc1315905e599b0ce31cd04c9fa91fc53b5ca2e806d2e1357dd8.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 342, + 555, + 365 + ], + "lines": [ + { + "bbox": [ + 313, + 342, + 555, + 365 + ], + "spans": [ + { + "bbox": [ + 313, + 342, + 555, + 365 + ], + "type": "text", + "content": "Table 4. Ablation study on synthetic homography dataset and HPatches. The \"full\" model is the default model." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 387, + 555, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 387, + 555, + 472 + ], + "spans": [ + { + "bbox": [ + 313, + 387, + 555, + 472 + ], + "type": "text", + "content": "icantly as more source views are integrated into the collaborative inference process. However, when the number of views exceeds 5, the matching capability begins to decline. We hypothesize that this degradation is attributed to the expanded search space and the excessive redundant information introduced by multi-view attention, which collectively increase the learning complexity of the network." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 483, + 388, + 495 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 483, + 388, + 495 + ], + "spans": [ + { + "bbox": [ + 313, + 483, + 388, + 495 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 502, + 555, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 502, + 555, + 658 + ], + "spans": [ + { + "bbox": [ + 313, + 502, + 555, + 658 + ], + "type": "text", + "content": "We proposed a multi-view collaborative matching strategy to address the challenge of reliable matching uncontrolled image sets in complex scenes. Our method, which consists of a deep collaborative matcher (CoMatcher) and a scalable groupwise pipeline, enables a holistic 3D scene understanding while inherently satisfying the cross-view projection consistency constraint. Extensive experiments have demonstrated that exploiting inter-view connections significantly enhances matching certainty, yielding substantial benefits for downstream tasks such as pose estimation and SfM. We hope this work encourages the research community to expand beyond pairwise matching and further explore the understanding and utilization of multi-view information." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 658, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 658, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 658, + 556, + 713 + ], + "type": "text", + "content": "Acknowledgement. This research is supported by NSFC projects under Grant 42471447, Development Program of China under Grant 2024YFC3811000, and the Fundamental Research Funds for the Central Universities of China under Grant 2042022dx0001." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21977" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 96, + 115, + 108 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 115, + 108 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 115, + 108 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 116, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 116, + 295, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 116, + 295, + 160 + ], + "spans": [ + { + "bbox": [ + 61, + 116, + 295, + 160 + ], + "type": "text", + "content": "[1] Sameer Agarwal, Noah Snavely, Ian Simon, Steven M Seitz, and Richard Szeliski. Building Rome in a day. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 72-79. IEEE, 2009. 1, 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 161, + 296, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 161, + 296, + 217 + ], + "spans": [ + { + "bbox": [ + 61, + 161, + 296, + 217 + ], + "type": "text", + "content": "[2] Relja Arandjelovic, Petr Gronat, Akihiko Torii, Tomas Pajdla, and Josef Sivic. NetVLAD: CNN architecture for weakly supervised place recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5297-5307, 2016. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 218, + 296, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 218, + 296, + 272 + ], + "spans": [ + { + "bbox": [ + 62, + 218, + 296, + 272 + ], + "type": "text", + "content": "[3] Vassileios Balntas, Karel Lenc, Andrea Vedaldi, and Krystian Mikolajczyk. HPatches: A benchmark and evaluation of handcrafted and learned local descriptors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5173-5182, 2017. 2, 5, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 274, + 295, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 274, + 295, + 329 + ], + "spans": [ + { + "bbox": [ + 62, + 274, + 295, + 329 + ], + "type": "text", + "content": "[4] Axel Barroso-Laguna, Sowmya Munukutla, Victor Adrian Prisacariu, and Eric Brachmann. Matching 2D images in 3D: Metric relative pose from metric correspondences. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4852-4863, 2024. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 331, + 294, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 331, + 294, + 364 + ], + "spans": [ + { + "bbox": [ + 62, + 331, + 294, + 364 + ], + "type": "text", + "content": "[5] Herbert Bay, Tinne Tuytelaars, and Luc Van Gool. SURF: Speeded up robust features. In European Conference on Computer Vision, pages 404-417. Springer, 2006. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 365, + 294, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 365, + 294, + 419 + ], + "spans": [ + { + "bbox": [ + 62, + 365, + 294, + 419 + ], + "type": "text", + "content": "[6] Ruojin Cai, Joseph Tung, Qianqian Wang, Hadar Averbuch-Elor, Bharath Hariharan, and Noah Snavely. Doppelgangers: Learning to disambiguate images of similar structures. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 34–44, 2023. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 421, + 294, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 421, + 294, + 475 + ], + "spans": [ + { + "bbox": [ + 62, + 421, + 294, + 475 + ], + "type": "text", + "content": "[7] Hongkai Chen, Zixin Luo, Jiahui Zhang, Lei Zhou, Xuyang Bai, Zeyu Hu, Chiew-Lan Tai, and Long Quan. Learning to match features with seeded graph matching network. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6301-6310, 2021. 2, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 477, + 294, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 477, + 294, + 533 + ], + "spans": [ + { + "bbox": [ + 62, + 477, + 294, + 533 + ], + "type": "text", + "content": "[8] Hongkai Chen, Zixin Luo, Lei Zhou, Yurun Tian, Mingmin Zhen, Tian Fang, David Mckinnon, Yanghai Tsin, and Long Quan. ASpanFormer: Detector-free image matching with adaptive span transformer. In European Conference on Computer Vision, pages 20-36. Springer, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 533, + 294, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 533, + 294, + 566 + ], + "spans": [ + { + "bbox": [ + 62, + 533, + 294, + 566 + ], + "type": "text", + "content": "[9] Ondrej Chum, Jiří Matas, and Josef Kittler. Locally optimized RANSAC. In Pattern Recognition, pages 236-243. Springer, 2003. 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 567, + 294, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 567, + 294, + 622 + ], + "spans": [ + { + "bbox": [ + 57, + 567, + 294, + 622 + ], + "type": "text", + "content": "[10] Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. SuperPoint: Self-supervised interest point detection and description. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 224-236, 2018. 2, 6, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 624, + 294, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 624, + 294, + 678 + ], + "spans": [ + { + "bbox": [ + 57, + 624, + 294, + 678 + ], + "type": "text", + "content": "[11] Carl Doersch, Ankush Gupta, Larisa Markeeeva, Adria Recasens, Lucas Smaira, Yusuf Aytar, Joao Carreira, Andrew Zisserman, and Yi Yang. TAP-Vid: A benchmark for tracking any point in a video. Advances in Neural Information Processing Systems, 35:13610-13626, 2022. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 680, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 680, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 680, + 294, + 713 + ], + "type": "text", + "content": "[12] Carl Doersch, Yi Yang, Mel Vecerik, Dilara Gokay, Ankush Gupta, Yusuf Aytar, Joao Carreira, and Andrew Zisserman. TAPIR: Tracking any point with per-frame initialization and" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 106 + ], + "type": "text", + "content": "temporal refinement. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10061-10072, 2023. 2, 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 107, + 553, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 107, + 553, + 172 + ], + "spans": [ + { + "bbox": [ + 317, + 107, + 553, + 172 + ], + "type": "text", + "content": "[13] Mihai Dusmanu, Ignacio Rocco, Tomas Pajdla, Marc Pollefeys, Josef Sivic, Akihiko Torii, and Torsten Sattler. D2-Net: A trainable cnn for joint description and detection of local features. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8092–8101, 2019. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 175, + 553, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 175, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 317, + 175, + 553, + 228 + ], + "type": "text", + "content": "[14] Johan Edstedt, Georg Bokman, Marten Wadenback, and Michael Felsberg. DeDoDe: Detect, don't describe—Describe, don't detect for local feature matching. In International Conference on 3D Vision, pages 148-157. IEEE, 2024. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 230, + 553, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 230, + 553, + 284 + ], + "spans": [ + { + "bbox": [ + 317, + 230, + 553, + 284 + ], + "type": "text", + "content": "[15] Johan Edstedt, Qiyu Sun, Georg Bokman, Marten Wadenback, and Michael Felsberg. RoMa: Robust dense feature matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19790-19800, 2024. 1, 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 286, + 553, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 286, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 316, + 286, + 553, + 319 + ], + "type": "text", + "content": "[16] Maha Elbayad, Jiatao Gu, Edouard Grave, and Michael Auli. Depth-adaptive Transformer. In International Conference on Learning Representations, pages 1-14, 2020. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 320, + 553, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 320, + 553, + 364 + ], + "spans": [ + { + "bbox": [ + 317, + 320, + 553, + 364 + ], + "type": "text", + "content": "[17] Martin A Fischler and Robert C Bolles. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. In Communications of the ACM, 24(6):381-395, 1981. 6, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 365, + 553, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 365, + 553, + 421 + ], + "spans": [ + { + "bbox": [ + 317, + 365, + 553, + 421 + ], + "type": "text", + "content": "[18] Jan-Michael Frahm, Pierre Fite-Georgel, David Gallup, Tim Johnson, Rahul Raguram, Changchang Wu, Yi-Hung Jen, Enrique Dunn, Brian Clipp, Svetlana Lazebnik, et al. Building Rome on a cloudless day. In European Conference on Computer Vision, pages 368-381. Springer, 2010. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 422, + 553, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 422, + 553, + 465 + ], + "spans": [ + { + "bbox": [ + 317, + 422, + 553, + 465 + ], + "type": "text", + "content": "[19] Adam W Harley, Zhaoyuan Fang, and Katerina Fragkiadaki. Particle video revisited: Tracking through occlusions using point trajectories. In European Conference on Computer Vision, pages 59-75. Springer, 2022. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 467, + 553, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 467, + 553, + 498 + ], + "spans": [ + { + "bbox": [ + 317, + 467, + 553, + 498 + ], + "type": "text", + "content": "[20] Richard Hartley and Andrew Zisserman. Multiple view geometry in computer vision. Cambridge university press, 2003. 1, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 501, + 553, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 501, + 553, + 555 + ], + "spans": [ + { + "bbox": [ + 317, + 501, + 553, + 555 + ], + "type": "text", + "content": "[21] Xingyi He, Jiaming Sun, Yifan Wang, Sida Peng, Qixing Huang, Hujun Bao, and Xiaowei Zhou. Detector-free structure-from-motion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21594-21603, 2024. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 557, + 553, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 557, + 553, + 612 + ], + "spans": [ + { + "bbox": [ + 317, + 557, + 553, + 612 + ], + "type": "text", + "content": "[22] Martin Humenberger, Yohann Cabon, Nicolas Guerin, Julien Morat, Vincent Leroy, Jérôme Revaud, Philippe Rerole, Noé Pion, Cesar de Souza, and Gabriela Csurka. Robust image retrieval-based visual localization using kapture. arXiv preprint arXiv:2007.13867, 2020. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 613, + 553, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 613, + 553, + 657 + ], + "spans": [ + { + "bbox": [ + 317, + 613, + 553, + 657 + ], + "type": "text", + "content": "[23] Heinly Jared, Johannes L Schonberger, Enrique Dunn, and Jan-Michael Frahm. Reconstructing the world in six days. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2015. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 658, + 555, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 658, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 658, + 555, + 713 + ], + "type": "text", + "content": "[24] Hanwen Jiang, Arjun Karpur, Bingyi Cao, Qixing Huang, and André Araujo. OmniGlue: Generalizable feature matching with foundation model guidance. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19865-19875, 2024. 2" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "21978" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "text", + "content": "[25] Yuhe Jin, Dmytro Mishkin, Anastasiia Mishchuk, Jiri Matas, Pascal Fua, Kwang Moo Yi, and Eduard Trulls. Image matching across wide baselines: From paper to practice. International Journal of Computer Vision, 129(2):517-547, 2021. 1, 2, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 294, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 294, + 183 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 294, + 183 + ], + "type": "text", + "content": "[26] Nikita Karaev, Ignacio Rocco, Benjamin Graham, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. CoTracker: It is better to track together. In European Conference on Computer Vision, pages 18-35. Springer, 2024. 2, 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 185, + 294, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 185, + 294, + 228 + ], + "spans": [ + { + "bbox": [ + 56, + 185, + 294, + 228 + ], + "type": "text", + "content": "[27] Vincent Leroy, Yohann Cabon, and Jérôme Revaud. Grounding image matching in 3D with MAST3R. In European Conference on Computer Vision, pages 71-91. Springer, 2024. 1, 2, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 230, + 294, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 230, + 294, + 274 + ], + "spans": [ + { + "bbox": [ + 56, + 230, + 294, + 274 + ], + "type": "text", + "content": "[28] Zhengqi Li and Noah Snavely. MegaDepth: Learning single-view depth prediction from internet photos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2041-2050, 2018. 5, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 275, + 294, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 275, + 294, + 330 + ], + "spans": [ + { + "bbox": [ + 56, + 275, + 294, + 330 + ], + "type": "text", + "content": "[29] Philipp Lindenberger, Paul-Edouard Sarlin, Viktor Larsson, and Marc Pollefeys. Pixel-perfect structure-from-motion with featuremetric refinement. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5987-5997, 2021. 1, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 331, + 294, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 331, + 294, + 376 + ], + "spans": [ + { + "bbox": [ + 56, + 331, + 294, + 376 + ], + "type": "text", + "content": "[30] Philipp Lindenberger, Paul-Edouard Sarlin, and Marc Pollefeys. LightGlue: Local feature matching at light speed. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17627–17638, 2023. 1, 2, 5, 6, 7, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 376, + 294, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 376, + 294, + 409 + ], + "spans": [ + { + "bbox": [ + 56, + 376, + 294, + 409 + ], + "type": "text", + "content": "[31] David G Lowe. Distinctive image features from scale-invariant keypoints. International Journal of Computer Vision, 60:91-110, 2004. 1, 2, 6, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 411, + 294, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 411, + 294, + 453 + ], + "spans": [ + { + "bbox": [ + 56, + 411, + 294, + 453 + ], + "type": "text", + "content": "[32] Jiayi Ma, Xingyu Jiang, Aoxiang Fan, Junjun Jiang, and Junchi Yan. Image matching from handcrafted to deep features: A survey. International Journal of Computer Vision, 129(1):23-79, 2021. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 455, + 294, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 455, + 294, + 498 + ], + "spans": [ + { + "bbox": [ + 56, + 455, + 294, + 498 + ], + "type": "text", + "content": "[33] Eleonora Maset, Federica Arrigoni, and Andrea Fusiello. Practical and efficient multi-view matching. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4568-4576, 2017. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 500, + 294, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 500, + 294, + 544 + ], + "spans": [ + { + "bbox": [ + 56, + 500, + 294, + 544 + ], + "type": "text", + "content": "[34] Raul Mur-Artal, Jose Maria Martinez Montiel, and Juan D Tardos. ORB-SLAM: A versatile and accurate monocular SLAM system. In IEEE Transactions on Robotics, 31(5): 1147-1163, 2015. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 545, + 294, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 545, + 294, + 578 + ], + "spans": [ + { + "bbox": [ + 56, + 545, + 294, + 578 + ], + "type": "text", + "content": "[35] Carl Olsson and Olof Enqvist. Stable structure-from-motion for unordered image collections. In Image Analysis, pages 524-535. Springer, 2011. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 579, + 294, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 579, + 294, + 612 + ], + "spans": [ + { + "bbox": [ + 56, + 579, + 294, + 612 + ], + "type": "text", + "content": "[36] Linfei Pan, Daniel Barath, Marc Pollefeys, and Johannes L Schonberger. Global structure-from-motion revisited. In European Conference on Computer Vision, 2024. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 613, + 294, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 613, + 294, + 668 + ], + "spans": [ + { + "bbox": [ + 56, + 613, + 294, + 668 + ], + "type": "text", + "content": "[37] Filip Radenović, Ahmet Iscen, Giorgos Tolias, Yannis Avrithis, and Ondrej Chum. Revisiting Oxford and Paris: Large-scale image retrieval benchmarking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5706-5715, 2018. 6, 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "type": "text", + "content": "[38] Filip Radenovic, Giorgos Tolias, and Ondrej Chum. Finetuning CNN image retrieval with no human annotation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 41(7):1655-1668, 2018. 3" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 116 + ], + "type": "text", + "content": "[39] Jerome Revaud, Cesar De Souza, Martin Humenberger, and Philippe Weinzaepfel. R2D2: Reliable and repeatable detector and descriptor. Advances in Neural Information Processing Systems, 32, 2019. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 118, + 553, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 553, + 161 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 553, + 161 + ], + "type": "text", + "content": "[40] Barbara Roessle and Matthias Nießner. End2end multi-view feature matching with differentiable pose optimization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 477-487, 2023. 2, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 162, + 553, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 162, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 316, + 162, + 553, + 205 + ], + "type": "text", + "content": "[41] Ethan Rublee, Vincent Rabaud, Kurt Konolige, and Gary Bradski. ORB: An efficient alternative to SIFT or SURF. In Proceeding of the IEEE/CVF International Conference on Computer Vision, pages 2564-2571. IEEE, 2011. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 206, + 553, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 206, + 553, + 259 + ], + "spans": [ + { + "bbox": [ + 316, + 206, + 553, + 259 + ], + "type": "text", + "content": "[42] Paul-Edouard Sarlin, Cesar Cadena, Roland Siegwart, and Marcin Dymczyk. From coarse to fine: Robust hierarchical localization at large scale. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12716–12725, 2019. 1, 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 261, + 555, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 261, + 555, + 316 + ], + "spans": [ + { + "bbox": [ + 316, + 261, + 555, + 316 + ], + "type": "text", + "content": "[43] Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. SuperGlue: Learning feature matching with graph neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4938–4947, 2020. 1, 2, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 316, + 553, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 316, + 553, + 381 + ], + "spans": [ + { + "bbox": [ + 316, + 316, + 553, + 381 + ], + "type": "text", + "content": "[44] Torsten Sattler, Akihiko Torii, Josef Sivic, Marc Pollefeys, Hajime Taira, Masatoshi Okutomi, and Tomas Pajdla. Are large-scale 3d models really necessary for accurate visual localization? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1637-1646, 2017. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 383, + 553, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 383, + 553, + 426 + ], + "spans": [ + { + "bbox": [ + 316, + 383, + 553, + 426 + ], + "type": "text", + "content": "[45] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-Motion revisited. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4104-4113, 2016. 1, 2, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 426, + 553, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 426, + 553, + 480 + ], + "spans": [ + { + "bbox": [ + 316, + 426, + 553, + 480 + ], + "type": "text", + "content": "[46] Johannes L Schonberger, Hans Hardmeier, Torsten Sattler, and Marc Pollefeys. Comparative evaluation of hand-crafted and learned local features. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1482-1491, 2017. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 482, + 553, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 482, + 553, + 525 + ], + "spans": [ + { + "bbox": [ + 316, + 482, + 553, + 525 + ], + "type": "text", + "content": "[47] Tal Schuster, Adam Fisch, Jai Gupta, Mostafa Dehghani, Dara Bahri, Vinh Tran, Yi Tay, and Donald Metzler. Confident adaptive language modeling. Advances in Neural Information Processing Systems, 35:17456-17472, 2022. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 526, + 553, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 526, + 553, + 569 + ], + "spans": [ + { + "bbox": [ + 316, + 526, + 553, + 569 + ], + "type": "text", + "content": "[48] Sivic and Zisserman. Video Google: A text retrieval approach to object matching in videos. In Proceedings in the IEEE/CVF International Conference on Computer Vision, pages 1470-1477. IEEE, 2003. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 571, + 553, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 571, + 553, + 612 + ], + "spans": [ + { + "bbox": [ + 316, + 571, + 553, + 612 + ], + "type": "text", + "content": "[49] Jianlin Su, Murtadha Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu. RoFormer: Enhanced transformer with rotary position embedding. Neurocomputing, 568:127063, 2024. 4" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 614, + 553, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 614, + 553, + 668 + ], + "spans": [ + { + "bbox": [ + 316, + 614, + 553, + 668 + ], + "type": "text", + "content": "[50] Jiaming Sun, Zehong Shen, Yuang Wang, Hujun Bao, and Xiaowei Zhou. LoFTR: Detector-free local feature matching with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8922-8931, 2021. 1, 2, 5, 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "text", + "content": "[51] Kun Sun, Jinhong Yu, Wenbing Tao, Xin Li, Chang Tang, and Yuhua Qian. A unified feature-spatial cycle consistency fusion framework for robust image matching. Information Fusion, 97:101810, 2023. 2" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21979" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 520 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "text", + "content": "[52] Prune Truong, Martin Danelljan, Luc Van Gool, and Radu Timofte. Learning accurate dense correspondences and when to trust them. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5714-5724, 2021. 2, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 294, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 294, + 172 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 294, + 172 + ], + "type": "text", + "content": "[53] Michal Tyszkiiewicz, Pascal Fua, and Eduard Trulls. DISK: Learning local features with policy gradient. Advances in Neural Information Processing Systems, 33:14254-14265, 2020. 2, 5, 6, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 174, + 294, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 174, + 294, + 217 + ], + "spans": [ + { + "bbox": [ + 56, + 174, + 294, + 217 + ], + "type": "text", + "content": "[54] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in Neural Information Processing Systems, 30, 2017. 1, 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 219, + 294, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 219, + 294, + 272 + ], + "spans": [ + { + "bbox": [ + 56, + 219, + 294, + 272 + ], + "type": "text", + "content": "[55] Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. DUSt3R: Geometric 3D vision made easy. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20697-20709, 2024. 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 274, + 294, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 274, + 294, + 328 + ], + "spans": [ + { + "bbox": [ + 56, + 274, + 294, + 328 + ], + "type": "text", + "content": "[56] Yifan Wang, Xingyi He, Sida Peng, Dongli Tan, and Xiaowei Zhou. Efficient LoFTR: Semi-dense local feature matching with sparse-like speed. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21666-21675, 2024. 2, 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 331, + 294, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 331, + 294, + 384 + ], + "spans": [ + { + "bbox": [ + 56, + 331, + 294, + 384 + ], + "type": "text", + "content": "[57] Yuxi Xiao, Qianqian Wang, Shangzhan Zhang, Nan Xue, Sida Peng, Yujun Shen, and Xiaowei Zhou. SpatialTracker: Tracking any 2D pixels in 3D space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20406-20417, 2024. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 386, + 294, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 386, + 294, + 429 + ], + "spans": [ + { + "bbox": [ + 56, + 386, + 294, + 429 + ], + "type": "text", + "content": "[58] Shibiao Xu, Shunpeng Chen, Rongtao Xu, Changwei Wang, Peng Lu, and Li Guo. Local feature matching using deep learning: A survey. Information Fusion, 107:102344, 2024. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 430, + 294, + 473 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 430, + 294, + 473 + ], + "spans": [ + { + "bbox": [ + 56, + 430, + 294, + 473 + ], + "type": "text", + "content": "[59] Kwang Moo Yi, Eduard Trulls, Vincent Lepetit, and Pascal Fua. LIFT: Learned invariant feature transform. In European Conference on Computer Vision, pages 467-483. Springer, 2016. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 475, + 294, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 475, + 294, + 520 + ], + "spans": [ + { + "bbox": [ + 56, + 475, + 294, + 520 + ], + "type": "text", + "content": "[60] Guofeng Zhang, Zilong Dong, Jiaya Jia, Tien-Tsin Wong, and Hujun Bao. Efficient non-consecutive feature tracking for structure-from-Motion. In European Conference on Computer Vision, pages 422-435. Springer, 2010. 2" + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21980" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/CoSDH_ Communication-Efficient Collaborative Perception via Supply-Demand Awareness and Intermediate-Late Hybridization/82cb35c7-62f2-4e6d-b4eb-be441dfd4b29_content_list.json b/2025/CoSDH_ Communication-Efficient Collaborative Perception via Supply-Demand Awareness and Intermediate-Late Hybridization/82cb35c7-62f2-4e6d-b4eb-be441dfd4b29_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..e873a2f241409a39adec8dde53ff4843475a6e20 --- /dev/null +++ b/2025/CoSDH_ Communication-Efficient Collaborative Perception via Supply-Demand Awareness and Intermediate-Late Hybridization/82cb35c7-62f2-4e6d-b4eb-be441dfd4b29_content_list.json @@ -0,0 +1,1383 @@ +[ + { + "type": "text", + "text": "CoSDH: Communication-Efficient Collaborative Perception via Supply-Demand Awareness and Intermediate-Late Hybridization", + "text_level": 1, + "bbox": [ + 94, + 128, + 903, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Junhao Xu $^{1*}$ , Yanan Zhang $^{2*}$ , Zhi Cai $^{1}$ , Di Huang $^{1\\dagger}$", + "bbox": [ + 320, + 200, + 718, + 220 + ], + "page_idx": 0 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "$^{1}$ State Key Laboratory of Complex and Critical Software Environment, Beihang University, Beijing, China", + "$^{2}$ School of Computer Science and Information Engineering, Hefei University of Technology, Hefei, China" + ], + "bbox": [ + 99, + 220, + 941, + 258 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "junhaoxu@buaa.edu.cn, yananzhang@hfut.edu.cn, {caizhi97, dhuang}@buaa.edu.cn", + "bbox": [ + 183, + 261, + 854, + 276 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 310, + 326, + 325 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Multi-agent collaborative perception enhances perceptual capabilities by utilizing information from multiple agents and is considered a fundamental solution to the problem of weak single-vehicle perception in autonomous driving. However, existing collaborative perception methods face a dilemma between communication efficiency and perception accuracy. To address this issue, we propose a novel communication-efficient collaborative perception framework based on supply-demand awareness and intermediate-late hybridization, dubbed as CoSDH. By modeling the supply-demand relationship between agents, the framework refines the selection of collaboration regions, reducing unnecessary communication cost while maintaining accuracy. In addition, we innovatively introduce the intermediate-late hybrid collaboration mode, where late-stage collaboration compensates for the performance degradation in collaborative perception under low communication bandwidth. Extensive experiments on multiple datasets, including both simulated and real-world scenarios, demonstrate that CoSDH achieves state-of-the-art detection accuracy and optimal bandwidth trade-offs, delivering superior detection precision under real communication bandwidths, thus proving its effectiveness and practical applicability. The code will be released at https://github.com/Xu2729/CoSDH.", + "bbox": [ + 88, + 343, + 485, + 720 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 746, + 220, + 762 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Collaborative perception allows multiple agents to exchange complementary perception information. It fundamentally addresses the issues of limited perception range, sensor blind spots, and occlusion from obstacles inherent in single-agent perception, thereby improving both the range and accuracy of perception. Recent studies have demonstrated that collaborative perception can be applied to var", + "bbox": [ + 89, + 767, + 482, + 875 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ious autonomous driving tasks, including 3D object detection [3, 35, 36], semantic segmentation [20, 37], 3D occupancy prediction [31], and trajectory prediction [43], exhibiting enhanced performance and improved robustness to occlusions, thus making it a crucial approach for advancing autonomous driving systems.", + "bbox": [ + 511, + 311, + 906, + 402 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Communication efficiency is a key issue in collaborative perception. An early approach to improving communication efficiency was to use autoencoders to compress the intermediate features that need to be transmitted [36]. Later, communication-efficient methods such as Where2comm [11] were proposed, which select important and sparse foreground regions for collaboration. However, these methods still lack sufficient precision in area selection and require more bandwidth than the real-world constraints1. As shown in Fig. 1a, when the bandwidth is limited to 27 Mbps, Where2comm [11] experiences a decrease of about $1.5\\%$ in average precision (AP@0.7) on the OPV2V [36] dataset with an intersection-over-union (IoU) threshold of 0.7. When the bandwidth is further limited to 5 Mbps and 1 Mbps, AP@0.7 drops by approximately $7.5\\%$ and $14.5\\%$ , respectively, even falling below the accuracy of late fusion methods at the result level. This makes it difficult for existing collaborative perception methods to meet the demands of practical applications.", + "bbox": [ + 511, + 402, + 908, + 689 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we first analyze the problem of selecting collaboration areas and refine this selection based on the supply-demand relationship between agents. As shown in Fig. 1b, for Ego, low foreground confidence areas can be divided into three types: Area 1 (low confidence due to absence of foreground), Area 2 (low confidence due to occlusion), and Area 3 (low confidence due to sparsity of the point cloud). Although all three areas appear as background from the Ego's perspective, Area 1 is well-observed and does not require collaboration, while Areas 2 and 3,", + "bbox": [ + 511, + 690, + 910, + 843 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "To the best of our knowledge, the common V2X communication approach uses IEEE 802.11p-based DSRC (Dedicated Short-Range Communications) technology [1, 24], which provides an information transmission rate of approximately 27 Mbps.", + "bbox": [ + 511, + 849, + 906, + 902 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal Contribution. †Corresponding Author.", + "bbox": [ + 107, + 886, + 352, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "6834", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/06fba8aec52e2e0d60b88c5175542a2224d5edff453a88d56a00d9dd8f457a17.jpg", + "image_caption": [ + "(a) The 3D detection accuracy of Where2comm [11] and CoSDH on OPV2V dataset [36] with different bandwidth limit. Assume the number of collaborative agents is 4 and detection frequency is $10\\mathrm{Hz}$ .", + "Figure 1. Issues of existing communication-efficient collaborative perception methods." + ], + "image_footnote": [], + "bbox": [ + 91, + 88, + 390, + 271 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/cfafa97c36fe5a934c2101d62e17fb153ddc49d8ce0238f4393af0acde4d72be.jpg", + "image_caption": [ + "(b) Point cloud map and classification of areas with low foreground confidence. Although Area 1 has low confidence due to the absence of foreground, it allows for good observation without collaboration. In contrast, Areas 2 and 3 exhibit poor observation and require collaboration, with Area 2 having low confidence due to occlusion and Area 3 due to sparsity of the point cloud." + ], + "image_footnote": [], + "bbox": [ + 403, + 94, + 903, + 271 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "with poorer observation, need collaboration. Additionally, to address the issue of significant accuracy degradation under real bandwidth constraints, we propose that compensating intermediate collaboration results with late fusion is a feasible solution. As shown in Fig. 1a, late fusion achieves much higher accuracy than Where2comm [11] even with a bandwidth of approximately 0.7 Mbps, compared to Where2comm's accuracy with bandwidth less than 5 Mbps. So the use of intermediate-late hybrid collaboration can greatly improve the accuracy lower bound under low-bandwidth conditions.", + "bbox": [ + 88, + 353, + 482, + 518 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Based on these insights, we propose CoSDH, a novel communication-efficient collaborative perception framework for 3D object detection. As shown in Fig. 2, CoSDH consists of three key components: i) Supply-demand-aware information selection, which chooses sparse yet crucial regions for collaboration; ii) Intermediate feature transmission and fusion, which transmits and effectively aggregates information from multiple agents in a communication-efficient manner; iii) Confidence-aware late fusion, which compensates for the intermediate fusion results at a minimal communication cost to improve accuracy. To evaluate CoSDH, we conducted extensive experiments on the simulation datasets OPV2V [36], V2XSim [20], and the real-world dataset DAIR-V2X [42]. The experimental results show that i) In the baseline case, CoSDH uses less bandwidth while achieving satisfactory accuracy, demonstrating a better accuracy-bandwidth trade-off; ii) Under the simulated real-world communication condition with a total bandwidth limit of 27 Mbps, CoSDH experiences less accuracy degradation and outperforms other methods while using less bandwidth.", + "bbox": [ + 91, + 521, + 483, + 837 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our contributions are as follows:", + "bbox": [ + 107, + 839, + 410, + 853 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- We present CoSDH, an innovative communication-efficient collaborative perception framework with better accuracy-bandwidth trade-offs for 3D object detection.", + "bbox": [ + 89, + 854, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a novel supply-demand-aware information selection module, further refining the collaboration area selection to achieve more efficient communication.", + "- We design a novel intermediate-late hybrid collaborative perception paradigm, where confidence-aware late fusion compensates for the intermediate fusion results to maintain high accuracy under low-bandwidth conditions.", + "- We conduct extensive experiments across multiple datasets with bandwidth constraints set closer to real-world conditions, and CoSDH achieves state-of-the-art detection accuracy along with optimal bandwidth trade-offs, demonstrating its feasibility in real-world collaborative perception scenarios." + ], + "bbox": [ + 511, + 354, + 905, + 550 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 511, + 556, + 653, + 571 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. 3D Object Detection", + "text_level": 1, + "bbox": [ + 511, + 575, + 705, + 592 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3D object detection, a key technology in autonomous driving, identifies and localizes objects in 3D scenes using environmental data from onboard sensors and can be categorized into image-based, point-cloud-based, and multimodal-fusion-based methods [26]. Image-based methods can be further classified into monocular [16, 33], stereo [6, 18], and multi-view [13, 44, 45] approaches based on the number of onboard cameras. These methods leverage the rich color and texture information of image, along with its dense data representation, to achieve cost-effective 3D perception for autonomous driving. However, image-based methods are inherently limited by the lack of depth information, which restricts their performance. In contrast, point-cloud-based methods, which adopt voxel-based [7, 39, 48], raw point cloud [27, 28, 46], or point-voxel hybrid [5, 29, 30] representations, can fully leverage the precise 3D geometric information provided by LiDAR, significantly improving perception capabilities. Considering that point-cloud-based methods suffer from sparse characteristic and lack rich semantic information, multimodal-fusion-based meth", + "bbox": [ + 511, + 598, + 906, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "6835", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ods [23, 41, 47] further enhance performance by leveraging the complementary advantages of different modalities.", + "bbox": [ + 89, + 90, + 482, + 121 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "However, these single-agent-based 3D object detection methods are limited by sensor range and susceptible to occlusion, making them unable to detect objects that are further away or completely occluded. This paper focuses on collaborative-perception-based 3D object detection methods, which improve detection performance by supplementing the limitations of single-agent detection with information from other agents.", + "bbox": [ + 89, + 125, + 483, + 247 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Collaborative Perception", + "text_level": 1, + "bbox": [ + 89, + 267, + 318, + 282 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Collaborative perception can be categorized into early collaboration, intermediate collaboration, and late collaboration based on the collaboration timing. Early collaboration [2, 4, 42] shares raw perception data, providing good perception accuracy but with high bandwidth. Late collaboration [20, 36] shares perception results, significantly reducing bandwidth, but leads to a decline in perception accuracy. Intermediate collaboration operates at the feature level and can achieve a better trade-off between accuracy and bandwidth by adjusting the intermediate features transmitted [3, 10, 11, 19, 21, 22, 32, 35-37], which is why it has been widely studied. Some of this research has focused on improving perception accuracy. FCooper [3] and CoFF [10] used manual modeling to fuse multi-agent features. Who2com [22] and When2com [21] perform selective communication and use attention-based fusion. V2VNet [32] and DiscoNet [19] employ communication graph-based fusion methods. However, these methods typically transmit complete, uncompressed intermediate BEV features, leading to enormous bandwidth requirements, which makes them challenging to apply in practice.", + "bbox": [ + 89, + 292, + 483, + 611 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To address the large bandwidth demand in collaborative perception, AttFuse [36] was the first to use autoencoders to compress intermediate features along the channel dimension, which was later adopted by V2X-ViT [35], CoBEVT [37] and others. However, this method leads to significant accuracy degradation at high compression rates. Where2comm [11] reduces bandwidth requirements by selecting sparse but important foreground regions for collaboration while maintaining perception accuracy. However, as bandwidth is further limited, perception accuracy still rapidly degrades, potentially even falling below the accuracy of simple late collaboration methods. To address this issue, this paper proposes a hybrid collaborative method based on both intermediate and late collaboration, which efficiently compensates for the intermediate collaborative results using late collaboration under bandwidth constraints. Furthermore, we also adopt and improve methods based on auto-encoders and information selection to save bandwidth while maintaining accuracy.", + "bbox": [ + 89, + 613, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 513, + 90, + 604, + 104 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Problem Definition", + "text_level": 1, + "bbox": [ + 513, + 109, + 694, + 123 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this paper, we consider the problem of collaborative perception with $N$ agents. Let $X_{i}$ and $y_{i}$ represent the raw observation and the corresponding ground truth supervision of the $i$ -th agent, respectively, and let $P_{j \\rightarrow i}$ be the message sent from agent $j$ to agent $i$ . In the collaborative perception, agent $i$ aggregates its own observations and the messages $\\{P_{j \\rightarrow i}\\}_{j=1}^{N}$ sent from other agents to perform 3D object detection task. Our goal is to maximize the collaborative perception 3D object detection accuracy while ensuring that each agent has a communication budget $B$ :", + "bbox": [ + 511, + 131, + 906, + 282 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\xi_ {\\Phi} (B) = \\arg \\max _ {\\theta} \\sum_ {i = 1} ^ {N} g \\left(\\Phi_ {\\theta} \\left(X _ {i}, \\left\\{P _ {j \\rightarrow i} \\right\\} _ {j = 1} ^ {N}\\right), y _ {i}\\right), (1) \\\\ \\text {s . t .} \\sum_ {j = 1} ^ {N} \\left| P _ {j \\rightarrow i} \\right| \\leq B (2) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 555, + 292, + 906, + 372 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\Phi_{\\theta}$ represents the collaborative perception 3D object detection model, $\\theta$ denotes the model parameters, $|P_{j\\rightarrow i}|$ represents the communication volume of the message sent from agent $j$ to agent $i$ , and $g(\\cdot ,\\cdot)$ denotes the 3D object detection evaluation metric.", + "bbox": [ + 511, + 381, + 906, + 455 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Overall Architecture", + "text_level": 1, + "bbox": [ + 513, + 465, + 709, + 479 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The overall architecture of the proposed $\\mathrm{COSDH}$ is shown in Fig. 2. Each agent first processes its locally observed point cloud $X_{i}$ through a backbone network based on PointPillar [17] and a demand generator to obtain multi-scale BEV features $\\{F_i^{(l)}\\}_{l = 1,2,\\dots,L}$ and a demand mask $D_{i}$ , respectively. Considering the collaboration between Ego agent $i$ and collaborating agent $j$ , agent $j$ generates a supply mask $S_{j}$ from its multi-scale features $\\{F_j^{(l)}\\}_{l = 1,2,\\dots,L}$ via a supply generator, and multiplies it element-wise with the received demand matrix to obtain the supply-demand mask $M_{j\\rightarrow i}$ . Agent $j$ then performs supply-demand-aware information selection by multiplying $\\{F_j^{(l)}\\}_{l = 1,2,\\dots,L}$ with $M_{j\\rightarrow i}$ element-wise to obtain sparse spatial features $\\{Z_j^{(l)}\\}_{l = 1,2,\\dots,L}$ . Subsequently, agent $j$ compresses the features through an autoencoder, sending the non-zero parts of the features along with their corresponding coordinates as the message $P_{j\\rightarrow i}$ to agent $i$ .", + "bbox": [ + 511, + 488, + 906, + 744 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Upon receiving $P_{j \\to i}$ , agent $i$ first decodes the features to restore their dimensions and then fuses them with its local features $\\{F_i^{(l)}\\}_{l = 1,2,\\dots,L}$ across multiple scales to obtain the fused features $\\{F_i^{(l)}\\}_{l = 1,2,\\dots,L}$ , which are then passed to the detection head for intermediate collaborative detection results $\\tilde{y}_i$ . Afterward, we apply confidence-aware late fusion: agent $j$ filters and suppresses its own detection results $y_j$ based on confidence and sends them to agent $i$ for late fusion, yielding the final hybrid collaborative perception detection results $\\hat{y}_i$ .", + "bbox": [ + 511, + 744, + 908, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "6836", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/e61490036051356a3fb3572ee96508d10627968156b304fe0182b559fa40aa0d.jpg", + "image_caption": [ + "Figure 2. The overall architecture of $\\mathrm{COSDH}$ . The Supply-Demand-Aware Information Selection module selects sparse but important information, which is then further compressed by the Message Compression module to achieve efficient communication. Confidence-Aware Late Fusion compensates for the intermediate fusion detection results to improve accuracy." + ], + "image_footnote": [], + "bbox": [ + 102, + 88, + 893, + 300 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Supply-Demand-Aware Information Selection", + "text_level": 1, + "bbox": [ + 89, + 371, + 478, + 386 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Previous methods such as Where2comm [11] and How2comm [40] generally use symmetric supply-demand relationships to select sparse features for collaboration, believing that areas with high foreground confidence in the collaborating agent's view should be provided, and conversely, areas with low foreground confidence in the Ego agent's view need collaboration. However, we believe that areas with low foreground confidence from the Ego's perspective can be divided into areas that are hard to observe and those that can be observed but belong to the background. The latter do not require collaboration. Based on this insight, we propose a novel supply-demand-aware information selection method.", + "bbox": [ + 88, + 396, + 483, + 592 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The demand mask $D_{i}$ indicates where agent $i$ needs information from collaborating agents. Intuitively, the agent requires information from areas that are distant or occluded, which have the common characteristic of having low point cloud density or no point cloud at all. For agent $i$ , we consider using the number of point clouds in each pillar to represent point cloud density, and we map it to the range [0,1], i.e., $A_{i} \\in [0,1]^{H \\times W}$ , where $H$ and $W$ represent the number of Pillars along the height and width dimensions. We then select areas where the point cloud density is below a threshold $\\varepsilon_{a}$ to obtain the demand mask for agent $i$ , $D_{i} = A_{i} < \\varepsilon_{a} \\in \\{0,1\\}^{H \\times W}$ . The demand mask indicates where agent $i$ has poor perception and needs collaborative information from other agents. Filtering information from other agents using the demand mask not only helps save bandwidth but also avoids interference from other agents' information in well-perceived areas.", + "bbox": [ + 89, + 595, + 483, + 851 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For the object detection task, foreground information is more valuable. Providing sparse foreground features can effectively assist other agents in supplementing undetected", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/564f796406eb4f74420469c5a1b582df37d4f9e6f9112793c8facaa7a1529079.jpg", + "image_caption": [ + "Figure 3. Comparison of our multi-scale compression and fusion with other methods. It can achieve thorough fusion with smaller communication volume." + ], + "image_footnote": [], + "bbox": [ + 517, + 369, + 915, + 571 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "and incomplete targets while using less bandwidth. Following previous work [11], we use the spatial confidence map $C_i \\in [0,1]^{H \\times W}$ output by the detection head to select potential foreground areas. We use a supply threshold $\\varepsilon_c$ to obtain the supply mask $S_i^{(l)} = C_i > \\varepsilon_c \\in \\{0,1\\}^{H \\times W}$ . By adjusting the threshold, we can dynamically adjust the bandwidth used for collaborative perception to adapt to varying communication conditions.", + "bbox": [ + 511, + 646, + 906, + 768 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "During collaboration, agent $j$ generates a binary supply-demand selection mask $M_{j\\rightarrow i} = D_i\\odot S_j\\in \\{0,1\\}^{H\\times W}$ based on its supply mask $S_{j}$ and agent $i$ 's demand mask $D_{i}$ and samples it to multiply element-wise with multi-scale BEV features $\\{F_j^{(l)}\\}_{l = 1,2,\\dots,L}$ , obtaining sparse features $\\{Z_{j\\rightarrow i}^{(l)}\\}_{l = 1,2,\\dots,L}$ . During communication, only the non-zero parts and their corresponding coordinates need to be transmitted.", + "bbox": [ + 511, + 771, + 908, + 898 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "6837", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Message Compression and Fusion", + "text_level": 1, + "bbox": [ + 91, + 90, + 385, + 107 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To further reduce communication bandwidth, we also use autoencoders to compress intermediate features along the channel dimension before communication, while making further improvements to existing compression and fusion schemes. As shown in Fig. 3, traditional single-scale fusion schemes [35, 36] only compress and fuse a single-layer BEV feature map before passing it to the detection head, resulting in insufficient fusion. CoAlign [25] proposed a solution that performs layer-wise fusion on multi-scale features, addressing the issue of insufficient fusion. However, it compresses large-scale single-layer features extracted after Pillar Feature Extraction (PFE), which leads to higher bandwidth. Therefore, we propose a multi-scale compression and fusion approach, where separate autoencoders are designed for each scale of features to perform compression, followed by layer-wise fusion. This approach enables more thorough fusion while using less bandwidth.", + "bbox": [ + 89, + 114, + 483, + 372 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For message fusion, we use Max fusion, which has two main advantages: i) it is computationally simple and efficient, with the complexity increasing linearly with the number of agents; ii) max fusion selects the maximum value of features from multiple agents, achieving information complementarity. Considering the collaboration between Ego agent $i$ and collaborating agent $j$ , the process of message compression and fusion can be expressed as:", + "bbox": [ + 89, + 376, + 483, + 498 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nZ _ {j \\rightarrow i} ^ {\\prime (l)} = f _ {e n c o d e} ^ {(l)} \\left(Z _ {j \\rightarrow i} ^ {(l)}\\right) \\in \\mathbb {R} ^ {\\frac {C _ {l}}{c _ {0}} \\times H _ {l} \\times W _ {l}} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 516, + 482, + 542 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nF _ {j \\rightarrow i} ^ {(l)} = f _ {\\text {d e c o d e}} ^ {(l)} \\left(Z _ {j \\rightarrow i} ^ {' (l)}\\right) \\in \\mathbb {R} ^ {C _ {l} \\times H _ {l} \\times W _ {l}} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 544, + 482, + 566 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nF _ {j \\rightarrow i} ^ {\\prime (l)} = f _ {\\text {t r a n s f o r m}} \\left(F _ {j \\rightarrow i} ^ {(l)}, \\xi_ {j \\rightarrow i}\\right) \\in \\mathbb {R} ^ {C _ {l} \\times H _ {l} \\times W _ {l}} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 569, + 482, + 590 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {F _ {i} ^ {(l)}} = \\max \\left(F _ {i} ^ {(l)}, \\left\\{F _ {j \\rightarrow i} ^ {\\prime (l)} \\right\\} _ {j \\neq i}\\right) \\in \\mathbb {R} ^ {C _ {l} \\times H _ {l} \\times W _ {l}} \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 153, + 594, + 480, + 619 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $Z_{j\\rightarrow i}^{(l)}\\in \\mathbb{R}^{C_l\\times H_l\\times W_l}$ represents the sparse features selected based on supply-demand relationships from the previous stage, and $C_l,H_l,W_l$ denote the channel, height, and width dimensions of the $l$ -th layer feature map. $f_{encode}^{(l)},f_{decode}^{(l)}$ represent the encoder and decoder of the $l$ -th layer autoencoder, $c_{0}$ is the compression ratio in the channel dimension, and $f_{transform}$ represents coordinate transformation. During communication, $\\{Z_{j\\rightarrow i}^{\\prime (l)}\\}_{l = 1,2,\\ldots ,L}$ is first converted from float32 to float16, and then only the non-zero parts and corresponding coordinates are transmitted to save bandwidth. Upon receiving the message, agent $i$ decodes the feature dimensions, then aligns the features to its own coordinate system using the coordinate transformation matrix $\\xi_{j\\rightarrow i}$ to obtain $F_{j\\rightarrow i}^{\\prime (l)}$ , and finally fuses its own features with the collaborating agent's features using max fusion to obtain the fused features $F_{i}^{(l)}$", + "bbox": [ + 89, + 641, + 483, + 902 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/80aeadd77d05b25ee99ff4941272c79b32d1f4a1cb80af8bb919803063e6bced.jpg", + "image_caption": [ + "Figure 4. Our confidence-aware late fusion. It filters detection results based on confidence and suppresses suboptimal results from collaborative agents, improving overall detection accuracy." + ], + "image_footnote": [], + "bbox": [ + 517, + 92, + 913, + 272 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Confidence-Aware Late Fusion", + "text_level": 1, + "bbox": [ + 511, + 359, + 785, + 373 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Existing collaborative perception methods focus on the singular intermediate collaboration architecture and neglect the advantages of late collaboration. As shown in Table 1, the experimental results indicate that late fusion can demonstrate acceptable perception accuracy with extremely low bandwidth on the OPV2V [36] and V2XSim [20] datasets. Therefore, late fusion can be used to compensate for the results of intermediate fusion, achieving a better accuracy-bandwidth trade-off.", + "bbox": [ + 511, + 387, + 906, + 523 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Naive late collaboration methods [20, 36] directly merge results from other agents and then apply NMS (Non-Maximum Suppression) to deduplicate and obtain final results. This approach can effectively improve recall for collaborative perception object detection; however, during the merging process, it may introduce some low-confidence false positives, and suboptimal detection results from collaborating agents may override the Ego agent's detection results, lowering precision and consequently decreasing final AP. The lower AP in the late fusion method in Table 1 on the DAIR-V2X [42] dataset is due to this issue.", + "bbox": [ + 511, + 531, + 908, + 698 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To address this problem, we propose a new confidence-aware late fusion method. As shown in Fig. 4, during late fusion, we first filter based on the confidence of the target boxes, discarding target boxes from other agents with confidence lower than $\\varepsilon_{l}$ . Furthermore, considering that if both the Ego agent and other agents detect the same target, the detection result from the Ego agent, which has undergone intermediate fusion, is of higher quality. Therefore, we suppress the detection results from other agents to prevent their lower-quality results from degrading the Ego agent's better detection results. Specifically, before merging the detection results, we multiply the confidence of target boxes from other agents by a coefficient $\\beta \\in (0,1)$ .", + "bbox": [ + 511, + 704, + 908, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "6838", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/60e243a33561c4bfd56b8961edf6821bdd9a1b98bf90aeaa352a6ce69f45a54b.jpg", + "table_caption": [], + "table_footnote": [ + "Table 1. Comparison of detection accuracy and bandwidth of different methods on OPV2V [36], V2XSim [20], and DAIR-V2X [42] datasets. \"BD\" represents the bandwidth required for each collaborative agent, assuming Ego agent collaborates with up to 4 agents and the detection frequency is $10\\mathrm{Hz}$ . \"BD $\\leq$ 6.75 Mbps\" is used to simulate real-world communication limits, assuming a total communication bandwidth of 27 Mbps, with each collaborating agent's bandwidth consumption limited to less than 6.75 Mbps. For intermediate collaboration methods without information selection, we provide a compressed version using autoencoders to meet the bandwidth constraints. For intermediate collaboration methods with information selection, the selection ratio is adjusted to meet the bandwidth constraints." + ], + "table_body": "
SettingMethodOPV2V [36]V2XSim [20]DAIR-V2X [42]
AP@0.5↑AP@0.7↑BD↓AP@0.5↑AP@0.7↑BD↓AP@0.5↑AP@0.7↑BD↓
BasicNo Fusion79.78%67.16%0.0 Mbps70.31%58.55%0.0 Mbps66.51%55.46%0.0 Mbps
Early Fusion95.05%88.98%83.1 Mbps95.68%88.03%55.5 Mbps74.52%59.22%50.2 Mbps
Late Fusion94.70%87.18%0.2 Mbps86.88%78.13%0.1 Mbps67.86%50.47%0.2 Mbps
No LimitWhen2com [21]91.75%81.77%1,320.0 Mbps72.65%62.92%250.0 Mbps64.08%49.14%984.8 Mbps
FCooper [3]90.06%74.03%2,640.0 Mbps72.96%57.61%500.0 Mbps74.58%56.47%1,968.8 Mbps
AttFuse [36]94.31%82.03%2,640.0 Mbps78.06%64.84%500.0 Mbps73.80%56.86%1,968.8 Mbps
V2VNet [32]96.66%92.44%5,280.0 Mbps88.97%85.18%1,000.0 Mbps66.63%47.39%3,937.5 Mbps
DiscoNet [19]90.93%78.90%2,640.0 Mbps77.34%68.77%500.0 Mbps73.58%58.45%1,968.8 Mbps
V2XViT [35]95.87%89.88%2,640.0 Mbps89.01%80.26%500.0 Mbps76.68%57.57%1,920.0 Mbps
Where2comm [11]95.59%91.39%48.7 Mbps88.18%83.66%27.5 Mbps76.13%60.16%172.3 Mbps
CoAlign [25]96.63%92.63%2,640.0 Mbps88.87%85.23%500.0 Mbps78.06%63.09%1,968.8 Mbps
CoSDH96.83%92.99%13.4 Mbps89.23%86.31%1.1 Mbps76.75%63.85%7.1 Mbps
BD ≤ 6.75 MbpsWhen2com [21]79.97%50.88%5.2 Mbps62.02%43.37%4.0 Mbps62.37%44.01%3.8 Mbps
FCooper [3]90.37%72.92%5.2 Mbps76.52%61.83%4.0 Mbps67.71%47.65%3.8 Mbps
AttFuse [36]93.45%80.02%5.2 Mbps84.58%71.36%4.0 Mbps71.06%49.57%3.8 Mbps
V2VNet [32]95.71%87.16%5.2 Mbps85.66%73.72%4.0 Mbps66.49%45.61%3.8 Mbps
DiscoNet [19]90.00%76.72%5.2 Mbps78.04%68.18%4.0 Mbps70.83%53.40%3.8 Mbps
V2XViT [35]95.85%87.13%5.2 Mbps88.94%81.47%4.0 Mbps71.00%52.78%3.8 Mbps
Where2comm [11]94.91%89.86%5.4 Mbps87.51%82.12%4.7 Mbps74.98%59.51%5.3 Mbps
CoAlign [25]94.10%85.99%5.2 Mbps88.01%83.97%4.0 Mbps75.26%60.19%3.8 Mbps
CoSDH96.75%92.92%2.0 Mbps89.23%86.31%1.1 Mbps76.47%63.76%1.4 Mbps
", + "bbox": [ + 94, + 87, + 903, + 345 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 455, + 223, + 470 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Datasets and Experimental Settings", + "text_level": 1, + "bbox": [ + 89, + 478, + 400, + 494 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. We evaluate the proposed CoSDH against other methods on three different collaborative perception datasets (OPV2V [36], V2XSim [20], and DAIR-V2X [42]) for LiDAR-based 3D object detection. The datasets include both simulated and real-world scenarios, and cover two types of collaboration: V2V (Vehicle to Vehicle) and V2I (Vehicle to Infrastructure).", + "bbox": [ + 89, + 502, + 482, + 607 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metrics We use the average precision (AP) with intersection-over-union (IoU) thresholds of 0.5 and 0.7 to evaluate the performance of different methods on 3D object detection. We assume the target detection frequency is $10\\mathrm{Hz}$ and calculate the communication bandwidth based on the average data transmitted by each collaborative agent to the Ego agent, in order to evaluate the communication cost of different methods. Specifically, we consider the bandwidth limitations in real-world collaborative perception scenarios, setting the vehicle's communication rate to 27 Mbps [1, 24]. Considering the typical case where the Ego agent collaborates with up to 4 other agents [36], the bandwidth limit for each collaborative agent is $27/4 = 6.75$ Mbps.", + "bbox": [ + 89, + 611, + 482, + 821 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Our experiments are based on the OpenCOOD [36] framework, using PointPillar [17] as the encoder with a grid size of $(0.4m, 0.4m)$ , and a maximum of 32 points per Pillar. For our method, we set the number of intermediate feature layers to $L = 3$ , the demand threshold", + "bbox": [ + 89, + 825, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "$\\varepsilon_{a} = 4 / 32 = 0.125$ , and the supply threshold $\\varepsilon_{c} = 0.01$ . For the OPV2V [36] and DAIR-V2X [42] datasets, the compression rate is $c_{0} = 16$ , and for the V2XSim [20] dataset, the compression rate is $c_{0} = 8$ due to its smaller perception range and lower inherent bandwidth requirement. The late fusion threshold is set to $\\varepsilon_{l} = 0.3$ , with a suppression coefficient $\\beta = 0.9$ for OPV2V and V2XSim dataset and $\\beta = 0.8$ for DAIR-V2X dataset because of its difficulty. In late fusion, the dense prediction results before NMS are transmitted, as this reduces the computational burden on the collaborating agents. The experiments use the Adam [14] optimizer, with an initial learning rate set between 0.0001 and 0.002 based on the model's testing complexity to ensure proper training. The maximum number of collaborating agents is set to 5. The number of training epochs is set to 40 to ensure model convergence. Other experimental parameters are kept consistent with the OpenCOOD framework. All methods are trained on four NVIDIA GeForce RTX 3090 GPUs.", + "bbox": [ + 511, + 455, + 906, + 743 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2.Quantitative Evaluation", + "text_level": 1, + "bbox": [ + 511, + 756, + 735, + 771 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Benchmark Comparison. Table 1 presents the collaborative 3D object detection accuracy and required bandwidth of the proposed CoSDH compared to previous methods across different datasets. Experimental results show that, with the default uncompressed settings, CoSDH achieves the highest accuracy on the OPV2V [36] and V2XSim [20] datasets while requiring only about 1/100 to 1/1000 of the bandwidth compared to other non-communication-efficient", + "bbox": [ + 511, + 779, + 906, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6839", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/7db46af544b2102d3c9ffe902cb3baa9ea6fb8a5a22940a03b130004298bb20b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 96, + 88, + 364, + 255 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b4969047a0f11c5b7a790bee2ed710f7b5bc4d5af5baed8b60554f1aafa82f0d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 364, + 88, + 630, + 255 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/db0a639c041739ebdbec2fa7a6dc8654654b5df37a6c1f12776c5742af6fa67e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 630, + 88, + 897, + 255 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1432b4733da76085279af19d24efa23ecf1fd4454e9f29ccd4d50ef70a197eab.jpg", + "image_caption": [ + "Figure 5. Comparison of the trade-off between detection accuracy and bandwidth of different methods on OPV2V [36], V2XSim [20] and DAIR-V2X [42] datasets, CoSDH achieves the best accuracy-bandwidth trade-off. The real-world limitation refers to the total bandwidth limit of 27 Mbps, which means that each collaborative agent does not exceed 6.75 Mbps." + ], + "image_footnote": [], + "bbox": [ + 96, + 260, + 364, + 419 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/25960b1269cdf59a40c0f5507e04ed8b6584511eadae182b809599bb39368e1e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 364, + 260, + 630, + 419 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b31a4409efb9422a4d40e23bef779118459e9b3b5ac96a53e320fb7bd610d8fc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 630, + 260, + 898, + 419 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "methods. Although the AP@0.5 of CoSDH on the DAIR-V2X [42] dataset is slightly lower than that of CoAlign [25], it achieves a higher AP@0.7 with only about 1/300 of the bandwidth. Both our CoSDH and Where2comm [11] are communication-efficient methods based on information selection, which enable dynamic accuracy-bandwidth trade-offs by adjusting the selection ratio. The table shows accuracy at a specific bandwidth, and a more detailed comparison of the accuracy-bandwidth curves is provided in the \"Accuracy-Bandwidth Trade-Off Comparison\" part.", + "bbox": [ + 88, + 489, + 482, + 641 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Furthermore, we simulate real-world communication rate limitations. Result shows that $\\mathrm{COSDH}$ achieves improvements in AP@0.7 of $3.06\\% / 2.34\\% / 3.75\\%$ on OPV2V/V2XSim/DAIR-V2X compared to the previous best methods, while using less bandwidth. When comparing the scenarios with and without bandwidth limitations on the OPV2V and DAIR-V2X datasets, we found that $\\mathrm{COSDH}$ achieves less than a $0.3\\%$ decrease in AP under conditions where bandwidth is reduced by $80\\%$ to $85\\%$ , exhibiting less accuracy degradation than Where2comm. Most other methods also show varying degrees of accuracy degradation under bandwidth constraints. Interestingly, under bandwidth limitations, some methods show an improvement in accuracy on certain datasets after using autoencoders to compress intermediate features. For example, FCooper [3] shows improved accuracy on the V2XSim dataset, which may be due to the model's initially poor per", + "bbox": [ + 89, + 643, + 482, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "formance. The compressed version, with the added autoencoder, increases the model's parameter count, thereby enriching its expressive capability.", + "bbox": [ + 511, + 489, + 906, + 536 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Accuracy-Bandwidth Trade-Off Comparison. Fig. 5 shows the accuracy-bandwidth trade-off of the proposed CoSDH and previous advanced methods across different datasets. When the bandwidth grater than 1 Mbps, Where2comm [11] demonstrates a good enough accuracy-bandwidth trade-off, maintaining high detection accuracy as bandwidth decreases. However, as bandwidth further decreases, its accuracy quickly declines, even falling below that of late fusion method. CoSDH can leverage late fusion at low bandwidths to maintain high accuracy, achieving a better performance-bandwidth trade-off. Notably, on the DAIR-V2X dataset, the late fusion method performs worse than the no-fusion case because the naive late fusion method unselectively merges detection results from other agents, introducing a large number of low-quality detections. CoSDH uses a confidence-aware late fusion method to select higher-quality detection results, improving this situation and enhancing perception accuracy. We also observe that the AP@0.7 of CoSDH on the V2XSim dataset initially increases slightly as bandwidth decreases before subsequently declining, indicating that selecting key collaboration areas can help reduce interference from other regions to some extent, thereby slightly improving detection accuracy.", + "bbox": [ + 511, + 553, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "6840", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c67ec94fd94c5dd281f6e942222661da0330f6dfe7c05ae61a60ca082af49f85.jpg", + "image_caption": [ + "CoAlign", + "Bandwidth: 3.85 Mbps" + ], + "image_footnote": [], + "bbox": [ + 94, + 88, + 294, + 179 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e86582efb0df25313fb0e1dc702e3c55eefcce969eb5f89dd88f630f5c436ac9.jpg", + "image_caption": [ + "Where2comm", + "Bandwidth: 5.26 Mbps" + ], + "image_footnote": [], + "bbox": [ + 294, + 88, + 495, + 178 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/26ad5a127e4a1ab99a2ca6d7f0919770d3530f8e83914f19a8855de206c11b9c.jpg", + "image_caption": [ + "CoSDH (w/o late fusion)", + "Figure 6. Visualization of detection results under real-world communication rate limitations on the DAIR-V2X [42] dataset. Green represents ground truth box, and red represents predicted box." + ], + "image_footnote": [], + "bbox": [ + 94, + 189, + 294, + 277 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/df3b39563c59aec3cbec23044cd0307074f1ed1168926dac4f523190071d4d4a.jpg", + "image_caption": [ + "CoSDH", + "Bandwidth: 0.29 Mbps" + ], + "image_footnote": [], + "bbox": [ + 295, + 189, + 493, + 277 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Qualitative Evaluation", + "text_level": 1, + "bbox": [ + 89, + 359, + 302, + 376 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Fig. 6 shows the visualization of detection results for our method compared to other methods under simulated real-world communication rate limitations on the DAIR-V2X dataset. A comparison of the detection results reveals that, while using less bandwidth, our method achieves the same recall rate as CoAlign [25] and Where2comm [11], with fewer false positive predictions and more accurate object localization, demonstrating that our method performs better under low-bandwidth conditions. Comparing the two images at bottom, it can be seen that under low bandwidth conditions, late fusion effectively compensates for the results of intermediate fusion, improving the recall rate of objects.", + "bbox": [ + 89, + 382, + 483, + 564 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Ablation Studies", + "text_level": 1, + "bbox": [ + 89, + 575, + 254, + 589 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 2 presents the results of the ablation study on various modules of the proposed method using the OPV2V [36] dataset. The results show that after using the autoencoder for compression, the bandwidth is reduced by a factor of $c_{0}$ , and there is no significant loss in perception accuracy, even with a slight improvement in AP@0.7. Converting intermediate features to float16 can nearly halve the bandwidth without significant loss, demonstrating the effectiveness of our proposed message compression module. After performing information selection based on the supply mask, perception accuracy slightly decreases, but the bandwidth is reduced by about $95\\%$ , which is a worthwhile trade-off. Further using the demand mask results in no significant decrease in accuracy but reduces bandwidth by about $10\\%$ . This demonstrates that our supply-demand-aware information selection can reduce bandwidth while maintaining accuracy.", + "bbox": [ + 89, + 598, + 482, + 853 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "After adding late fusion, accuracy improves at a relatively small bandwidth cost. However, since this phase uses more bandwidth, the improvement in accuracy is not signif", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/02dd07231ff7a73b3b1f89fea4923a03773404617f641ce62e39802c222074d5.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
CompressionSelectionLate FusionAP@0.5↑AP@0.7↑BD↓
AutoencoderFP16SupplyDemand
96.62%92.62%1,155.00 Mbps
✓96.59%92.99%72.19 Mbps
✓✓96.59%92.99%36.09 Mbps
✓✓✓96.30%92.62%1.99 Mbps
✓✓✓✓96.31%92.60%1.82 Mbps
✓✓✓✓✓96.75%92.92%1.97 Mbps
", + "bbox": [ + 516, + 88, + 911, + 161 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/d9d43b1cb3150781ef732c8c89ecadf87f5ed02235306f654bc8df4a7e2d6eb0.jpg", + "table_caption": [ + "Table 2. Ablation study of the modules in $\\mathrm{C}\\circ \\mathrm{{SDH}}$ on the OPV2V dataset. \"Autoencoder\" refers to the use of autoencoders to compress features along the channel dimension, with a compression ratio of ${c}_{0} = {16}$ . \"FP16\" refers to converting features from float32 to float16 for transmission. \"Supply\" and \"Demand\" refer to using supply and demand masks for information selection, and \"Late Fusion\" refers to applying confidence-aware late fusion." + ], + "table_footnote": [], + "table_body": "
Late Fusionεc0.010.020.030.050.07
AP@0.5↑96.59%96.31%96.19%95.18%93.35%
AP@0.7↑92.95%92.60%92.27%90.67%87.60%
BD↓13.26 Mbps1.82 Mbps0.54 Mbps0.12 Mbps0.05 Mbps
✓AP@0.5↑96.83%96.75%96.72%96.54%96.41%
AP@0.7↑92.99%92.92%92.73%92.23%91.68%
BD↓13.40 Mbps1.97 Mbps0.68 Mbps0.26 Mbps0.19 Mbps
", + "bbox": [ + 516, + 290, + 911, + 359 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3. Ablation study of confidence-aware late fusion under different bandwidths on the OPV2V [36] dataset. The table shows the impact of late fusion on accuracy and bandwidth with different values of $\\varepsilon_{c}$ .", + "bbox": [ + 511, + 369, + 906, + 425 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "cant. Table 3 shows the accuracy improvement due to late collaboration under different bandwidth conditions. It can be observed that late collaboration provides more accuracy improvements under lower bandwidth conditions. For example, when the bandwidth used for intermediate fusion is 0.05 Mbps, late collaboration with a bandwidth cost of 0.14 Mbps leads to a $3\\%$ improvement in AP@0.5 and a $4\\%$ improvement in AP@0.7. This is one of the key reasons why our method achieves significantly higher detection accuracy under lower bandwidth compared to Where2comm [11].", + "bbox": [ + 511, + 452, + 906, + 603 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 622, + 633, + 636 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we propose CoSDH, a novel communication-efficient collaborative perception framework for 3D object detection. By finely modeling the supply-demand relationship between agents, it selects key and sparse regions for collaboration. Additionally, we innovatively incorporate confidence-aware late fusion on top of intermediate collaboration to form an intermediate-late hybrid collaborative perception method. Experiments on multiple datasets show that our method offers a better accuracy-bandwidth trade-off, demonstrating outstanding accuracy under bandwidth constraints close to real-world communication limits, making it highly valuable for practical applications.", + "bbox": [ + 511, + 646, + 906, + 828 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgment", + "text_level": 1, + "bbox": [ + 511, + 845, + 665, + 863 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work is supported by the National Key Research and Development Plan (2024YFB3309302).", + "bbox": [ + 511, + 869, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "6841", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 90, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Fabio Arena and Giovanni Pau. An overview of vehicular communications. Future internet, 11(2):27, 2019. 1, 6", + "[2] Eduardo Arnold, Mehrdad Dianati, Robert de Temple, and Saber Fallah. Cooperative perception for 3d object detection in driving scenarios using infrastructure sensors. IEEE Transactions on Intelligent Transportation Systems, 23(3): 1852-1864, 2020. 3", + "[3] Qi Chen, Xu Ma, Sihai Tang, Jingda Guo, Qing Yang, and Song Fu. F-cooper: Feature based cooperative perception for autonomous vehicle edge computing system using 3d point clouds. In Proceedings of the 4th ACM/IEEE Symposium on Edge Computing, pages 88-100, 2019. 1, 3, 6, 7, 2", + "[4] Qi Chen, Sihai Tang, Qing Yang, and Song Fu. Cooper: Cooperative perception for connected autonomous vehicles based on 3d point clouds. In 2019 IEEE 39th International Conference on Distributed Computing Systems (ICDCS), pages 514-524. IEEE, 2019. 3", + "[5] Yilun Chen, Shu Liu, Xiaoyong Shen, and Jiaya Jia. Fast point r-cnn. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9775-9784, 2019. 2", + "[6] Yilun Chen, Shu Liu, Xiaoyong Shen, and Jiaya Jia. Dsgn: Deep stereo geometry network for 3d object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12536-12545, 2020. 2", + "[7] Jiajun Deng, Shaoshuai Shi, Peiwei Li, Wengang Zhou, Yanyong Zhang, and Houqiang Li. Voxel r-cnn: Towards high performance voxel-based 3d object detection. In Proceedings of the AAAI conference on artificial intelligence, pages 1201-1209, 2021. 2", + "[8] Alexey Dosovitskiy, German Ros, Felipe Codevilla, Antonio Lopez, and Vladlen Koltun. Carla: An open urban driving simulator. In Conference on robot learning, pages 1-16. PMLR, 2017. 1", + "[9] Sebastian Gräfling, Petri Mähönen, and Janne Riihijärvi. Performance evaluation of the IEEE 1609 wave and the IEEE 802.11 p for vehicular communications. In 2010 second international conference on ubiquitous and future networks (ICUFN), pages 344-348. IEEE, 2010. 2", + "[10] Jingda Guo, Dominic Carrillo, Sihai Tang, Qi Chen, Qing Yang, Song Fu, Xi Wang, Nannan Wang, and Paparao Palacharla. Coff: Cooperative spatial feature fusion for 3-d object detection on autonomous vehicles. IEEE Internet of Things Journal, 8(14):11078-11087, 2021. 3", + "[11] Yue Hu, Shaoheng Fang, Zixing Lei, Yiqi Zhong, and Siheng Chen. Where2comm: Communication-efficient collaborative perception via spatial confidence maps. Advances in neural information processing systems, 35:4874-4886, 2022. 1, 2, 3, 4, 6, 7, 8", + "[12] Yue Hu, Juntong Peng, Sifei Liu, Junhao Ge, Si Liu, and Si-heng Chen. Communication-efficient collaborative perception via information filling with codebook. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15481-15490, 2024. 2", + "[13] Zheng Jiang, Jinqing Zhang, Yanan Zhang, Qingjie Liu, Zhenghui Hu, Baohui Wang, and Yunhong Wang. Fsd-bev:" + ], + "bbox": [ + 93, + 109, + 482, + 900 + ], + "page_idx": 8 + }, + { + "type": "ref_text", + "text": "Foreground self-distillation for multi-view 3d object detection. In European Conference on Computer Vision, pages 110-126. Springer, 2024. 2", + "bbox": [ + 544, + 92, + 903, + 133 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Diederik P Kingma. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6", + "[15] Daniel Krajzewicz, Jakob Erdmann, Michael Behrisch, and Laura Bieker. Recent development and applications of sumosimulation of urban mobility. International journal on advances in systems and measurements, 5(3&4), 2012. 1", + "[16] Jason Ku, Alex D Pon, and Steven L Waslander. Monocular 3d object detection leveraging accurate proposals and shape reconstruction. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11867-11876, 2019. 2", + "[17] Alex H Lang, Sourabh Vora, Holger Caesar, Lubing Zhou, Jiong Yang, and Oscar Beijbom. Pointpillars: Fast encoders for object detection from point clouds. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12697-12705, 2019. 3, 6", + "[18] Peiliang Li, Xiaozhi Chen, and Shaojie Shen. Stereo r-cnn based 3d object detection for autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7644-7652, 2019. 2", + "[19] Yiming Li, Shunli Ren, Pengxiang Wu, Siheng Chen, Chen Feng, and Wenjun Zhang. Learning distilled collaboration graph for multi-agent perception. Advances in Neural Information Processing Systems, 34:29541-29552, 2021. 3, 6", + "[20] Yiming Li, Dekun Ma, Ziyan An, Zixun Wang, Yiqi Zhong, Siheng Chen, and Chen Feng. V2x-sim: Multi-agent collaborative perception dataset and benchmark for autonomous driving. IEEE Robotics and Automation Letters, 7(4):10914-10921, 2022. 1, 2, 3, 5, 6, 7", + "[21] Yen-Cheng Liu, Junjiao Tian, Nathaniel Glaser, and Zsolt Kira. When2com: Multi-agent perception via communication graph grouping. In Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition, pages 4106-4115, 2020. 3, 6", + "[22] Yen-Cheng Liu, Junjiao Tian, Chih-Yao Ma, Nathan Glaser, Chia-Wen Kuo, and Zsolt Kira. Who2com: Collaborative perception via learnable handshake communication. In 2020 IEEE International Conference on Robotics and Automation (ICRA), pages 6876-6883. IEEE, 2020. 3", + "[23] Zhijian Liu, Haotian Tang, Alexander Amini, Xinyu Yang, Huizi Mao, Daniela L Rus, and Song Han. Bevfusion: Multi-task multi-sensor fusion with unified bird's-eye view representation. In 2023 IEEE international conference on robotics and automation (ICRA), pages 2774-2781. IEEE, 2023. 3", + "[24] Jose Manuel Lozano Dominguez and Tomas Jesus Mateo Sanguino. Review on v2x, i2x, and p2x communications and their applications: a comprehensive analysis over time. Sensors, 19(12):2756, 2019. 1, 6", + "[25] Yifan Lu, Quanhao Li, Baoan Liu, Mehrdad Dianati, Chen Feng, Siheng Chen, and Yanfeng Wang. Robust collaborative 3d object detection in presence of pose errors. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 4812-4818. IEEE, 2023. 5, 6, 7, 8, 1, 2", + "[26] Jiageng Mao, Shaoshuai Shi, Xiaogang Wang, and Hongsheng Li. 3d object detection for autonomous driving: A" + ], + "bbox": [ + 93, + 109, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "6842", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "comprehensive survey. International Journal of Computer Vision, 131(8):1909-1963, 2023. 2", + "[27] Xuran Pan, Zhuofan Xia, Shiji Song, Li Erran Li, and Gao Huang. 3d object detection with pointformer. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7463-7472, 2021. 2", + "[28] Shaoshuai Shi, Xiaogang Wang, and Hongsheng Li. Pointrcnn: 3d object proposal generation and detection from point cloud. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 770-779, 2019. 2", + "[29] Shaoshuai Shi, Chaoxu Guo, Li Jiang, Zhe Wang, Jianping Shi, Xiaogang Wang, and Hongsheng Li. Pv-rcnn: Pointvoxel feature set abstraction for 3d object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10529–10538, 2020. 2", + "[30] Shaoshuai Shi, Zhe Wang, Jianping Shi, Xiaogang Wang, and Hongsheng Li. From points to parts: 3d object detection from point cloud with part-aware and part-aggregation network. IEEE transactions on pattern analysis and machine intelligence, 43(8):2647-2664, 2020. 2", + "[31] Rui Song, Chenwei Liang, Hu Cao, Zhiran Yan, Walter Zimmer, Markus Gross, Andreas Festag, and Alois Knoll. Collaborative semantic occupancy prediction with hybrid feature fusion in connected automated vehicles. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17996-18006, 2024. 1", + "[32] Tsun-Hsuan Wang, Sivabalan Manivasagam, Ming Liang, Bin Yang, Wenyuan Zeng, and Raquel Urtasun. V2vnet: Vehicle-to-vehicle communication for joint perception and prediction. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part II 16, pages 605–621. Springer, 2020. 3, 6", + "[33] Yan Wang, Wei-Lun Chao, Divyansh Garg, Bharath Hariharan, Mark Campbell, and Kilian Q Weinberger. Pseudolidar from visual depth estimation: Bridging the gap in 3d object detection for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8445-8453, 2019. 2", + "[34] Runsheng Xu, Yi Guo, Xu Han, Xin Xia, Hao Xiang, and Jiaqi Ma. Opencda: an open cooperative driving automation framework integrated with co-simulation. In 2021 IEEE International Intelligent Transportation Systems Conference (ITSC), pages 1155-1162. IEEE, 2021. 1", + "[35] Runsheng Xu, Hao Xiang, Zhengzhong Tu, Xin Xia, Ming-Hsuan Yang, and Jiaqi Ma. V2x-vit: Vehicle-to-everything cooperative perception with vision transformer. In European conference on computer vision, pages 107-124. Springer, 2022. 1, 3, 5, 6, 2", + "[36] Runsheng Xu, Hao Xiang, Xin Xia, Xu Han, Jinlong Li, and Jiaqi Ma. Opv2v: An open benchmark dataset and fusion pipeline for perception with vehicle-to-vehicle communication. In 2022 International Conference on Robotics and Automation (ICRA), pages 2583–2589. IEEE, 2022. 1, 2, 3, 5, 6, 7, 8", + "[37] Runsheng Xu, Zhengzhong Tu, Hao Xiang, Wei Shao, Bolei Zhou, and Jiaqi Ma. Cobevt: Cooperative bird's eye view" + ], + "bbox": [ + 91, + 92, + 483, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "semantic segmentation with sparse transformers. In Conference on Robot Learning, pages 989-1000. PMLR, 2023. 1, 3", + "[38] Runsheng Xu, Xin Xia, Jinlong Li, Hanzhao Li, Shuo Zhang, Zhengzhong Tu, Zonglin Meng, Hao Xiang, Xiaoyu Dong, Rui Song, et al. V2v4real: A real-world large-scale dataset for vehicle-to-vehicle cooperative perception. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13712–13722, 2023. 1, 2", + "[39] Yan Yan, Yuxing Mao, and Bo Li. Second: Sparsely embedded convolutional detection. Sensors, 18(10):3337, 2018. 2", + "[40] Dingkang Yang, Kun Yang, Yuzheng Wang, Jing Liu, Zhi Xu, Rongbin Yin, Peng Zhai, and Lihua Zhang. How2comm: Communication-efficient and collaboration-pragmatic multi-agent perception. Advances in Neural Information Processing Systems, 36, 2024. 4", + "[41] Junbo Yin, Jianbing Shen, Runnan Chen, Wei Li, Ruigang Yang, Pascal Frossard, and Wenguan Wang. Is-fusion: Instance-scene collaborative fusion for multimodal 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14905-14915, 2024. 3", + "[42] Haibao Yu, Yizhen Luo, Mao Shu, Yiyi Huo, Zebang Yang, Yifeng Shi, Zhenglong Guo, Hanyu Li, Xing Hu, Jirui Yuan, et al. Dair-v2x: A large-scale dataset for vehicleinfrastructure cooperative 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21361-21370, 2022. 2, 3, 5, 6, 7, 8, 1", + "[43] Haibao Yu, Wenxian Yang, Hongzhi Ruan, Zhenwei Yang, Yingjuan Tang, Xu Gao, Xin Hao, Yifeng Shi, Yifeng Pan, Ning Sun, et al. V2x-seq: A large-scale sequential dataset for vehicle-infrastructure cooperative perception and forecasting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5486-5495, 2023. 1", + "[44] Jinqing Zhang, Yanan Zhang, Qingjie Liu, and Yunhong Wang. Sa-bev: Generating semantic-aware bird's-eye-view feature for multi-view 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3348-3357, 2023. 2", + "[45] Jinqing Zhang, Yanan Zhang, Yunlong Qi, Zehua Fu, Qingjie Liu, and Yunhong Wang. Geobev: Learning geometric bev representation for multi-view 3d object detection. arXiv preprint arXiv:2409.01816, 2024. 2", + "[46] Yanan Zhang, Di Huang, and Yunhong Wang. Pc-rgnn: Point cloud completion and graph neural network for 3d object detection. In Proceedings of the AAAI conference on artificial intelligence, pages 3430-3437, 2021. 2", + "[47] Yanan Zhang, Jiaxin Chen, and Di Huang. Cat-det: Contrastively augmented transformer for multi-modal 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 908-917, 2022. 3", + "[48] Chao Zhou, Yanan Zhang, Jiaxin Chen, and Di Huang. Octr: Octtree-based transformer for 3d object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5166-5175, 2023. 2" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "6843", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/CoSDH_ Communication-Efficient Collaborative Perception via Supply-Demand Awareness and Intermediate-Late Hybridization/82cb35c7-62f2-4e6d-b4eb-be441dfd4b29_model.json b/2025/CoSDH_ Communication-Efficient Collaborative Perception via Supply-Demand Awareness and Intermediate-Late Hybridization/82cb35c7-62f2-4e6d-b4eb-be441dfd4b29_model.json new file mode 100644 index 0000000000000000000000000000000000000000..e4246e97a92ce7041f2caa4d7a678e4f955b6c56 --- /dev/null +++ b/2025/CoSDH_ Communication-Efficient Collaborative Perception via Supply-Demand Awareness and Intermediate-Late Hybridization/82cb35c7-62f2-4e6d-b4eb-be441dfd4b29_model.json @@ -0,0 +1,2035 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.095, + 0.13, + 0.904, + 0.175 + ], + "angle": 0, + "content": "CoSDH: Communication-Efficient Collaborative Perception via Supply-Demand Awareness and Intermediate-Late Hybridization" + }, + { + "type": "text", + "bbox": [ + 0.321, + 0.202, + 0.72, + 0.221 + ], + "angle": 0, + "content": "Junhao Xu\\(^{1*}\\), Yanan Zhang\\(^{2*}\\), Zhi Cai\\(^{1}\\), Di Huang\\(^{1\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.222, + 0.942, + 0.24 + ], + "angle": 0, + "content": "\\(^{1}\\)State Key Laboratory of Complex and Critical Software Environment, Beihang University, Beijing, China" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.24, + 0.94, + 0.259 + ], + "angle": 0, + "content": "\\(^{2}\\)School of Computer Science and Information Engineering, Hefei University of Technology, Hefei, China" + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.222, + 0.942, + 0.259 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.184, + 0.262, + 0.855, + 0.277 + ], + "angle": 0, + "content": "junhaoxu@buaa.edu.cn, yananzhang@hfut.edu.cn, {caizhi97, dhuang}@buaa.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.311, + 0.327, + 0.327 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.344, + 0.486, + 0.722 + ], + "angle": 0, + "content": "Multi-agent collaborative perception enhances perceptual capabilities by utilizing information from multiple agents and is considered a fundamental solution to the problem of weak single-vehicle perception in autonomous driving. However, existing collaborative perception methods face a dilemma between communication efficiency and perception accuracy. To address this issue, we propose a novel communication-efficient collaborative perception framework based on supply-demand awareness and intermediate-late hybridization, dubbed as CoSDH. By modeling the supply-demand relationship between agents, the framework refines the selection of collaboration regions, reducing unnecessary communication cost while maintaining accuracy. In addition, we innovatively introduce the intermediate-late hybrid collaboration mode, where late-stage collaboration compensates for the performance degradation in collaborative perception under low communication bandwidth. Extensive experiments on multiple datasets, including both simulated and real-world scenarios, demonstrate that CoSDH achieves state-of-the-art detection accuracy and optimal bandwidth trade-offs, delivering superior detection precision under real communication bandwidths, thus proving its effectiveness and practical applicability. The code will be released at https://github.com/Xu2729/CoSDH." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.747, + 0.222, + 0.763 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.768, + 0.483, + 0.875 + ], + "angle": 0, + "content": "Collaborative perception allows multiple agents to exchange complementary perception information. It fundamentally addresses the issues of limited perception range, sensor blind spots, and occlusion from obstacles inherent in single-agent perception, thereby improving both the range and accuracy of perception. Recent studies have demonstrated that collaborative perception can be applied to var" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.313, + 0.907, + 0.403 + ], + "angle": 0, + "content": "ious autonomous driving tasks, including 3D object detection [3, 35, 36], semantic segmentation [20, 37], 3D occupancy prediction [31], and trajectory prediction [43], exhibiting enhanced performance and improved robustness to occlusions, thus making it a crucial approach for advancing autonomous driving systems." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.404, + 0.909, + 0.69 + ], + "angle": 0, + "content": "Communication efficiency is a key issue in collaborative perception. An early approach to improving communication efficiency was to use autoencoders to compress the intermediate features that need to be transmitted [36]. Later, communication-efficient methods such as Where2comm [11] were proposed, which select important and sparse foreground regions for collaboration. However, these methods still lack sufficient precision in area selection and require more bandwidth than the real-world constraints1. As shown in Fig. 1a, when the bandwidth is limited to 27 Mbps, Where2comm [11] experiences a decrease of about \\(1.5\\%\\) in average precision (AP@0.7) on the OPV2V [36] dataset with an intersection-over-union (IoU) threshold of 0.7. When the bandwidth is further limited to 5 Mbps and 1 Mbps, AP@0.7 drops by approximately \\(7.5\\%\\) and \\(14.5\\%\\), respectively, even falling below the accuracy of late fusion methods at the result level. This makes it difficult for existing collaborative perception methods to meet the demands of practical applications." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.691, + 0.911, + 0.844 + ], + "angle": 0, + "content": "In this paper, we first analyze the problem of selecting collaboration areas and refine this selection based on the supply-demand relationship between agents. As shown in Fig. 1b, for Ego, low foreground confidence areas can be divided into three types: Area 1 (low confidence due to absence of foreground), Area 2 (low confidence due to occlusion), and Area 3 (low confidence due to sparsity of the point cloud). Although all three areas appear as background from the Ego's perspective, Area 1 is well-observed and does not require collaboration, while Areas 2 and 3," + }, + { + "type": "page_footnote", + "bbox": [ + 0.513, + 0.851, + 0.908, + 0.903 + ], + "angle": 0, + "content": "To the best of our knowledge, the common V2X communication approach uses IEEE 802.11p-based DSRC (Dedicated Short-Range Communications) technology [1, 24], which provides an information transmission rate of approximately 27 Mbps." + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.887, + 0.353, + 0.901 + ], + "angle": 0, + "content": "*Equal Contribution. †Corresponding Author." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6834" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.089, + 0.391, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.275, + 0.393, + 0.323 + ], + "angle": 0, + "content": "(a) The 3D detection accuracy of Where2comm [11] and CoSDH on OPV2V dataset [36] with different bandwidth limit. Assume the number of collaborative agents is 4 and detection frequency is \\(10\\mathrm{Hz}\\)." + }, + { + "type": "image", + "bbox": [ + 0.405, + 0.095, + 0.905, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.402, + 0.275, + 0.907, + 0.324 + ], + "angle": 0, + "content": "(b) Point cloud map and classification of areas with low foreground confidence. Although Area 1 has low confidence due to the absence of foreground, it allows for good observation without collaboration. In contrast, Areas 2 and 3 exhibit poor observation and require collaboration, with Area 2 having low confidence due to occlusion and Area 3 due to sparsity of the point cloud." + }, + { + "type": "image_caption", + "bbox": [ + 0.24, + 0.329, + 0.755, + 0.344 + ], + "angle": 0, + "content": "Figure 1. Issues of existing communication-efficient collaborative perception methods." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.354, + 0.483, + 0.52 + ], + "angle": 0, + "content": "with poorer observation, need collaboration. Additionally, to address the issue of significant accuracy degradation under real bandwidth constraints, we propose that compensating intermediate collaboration results with late fusion is a feasible solution. As shown in Fig. 1a, late fusion achieves much higher accuracy than Where2comm [11] even with a bandwidth of approximately 0.7 Mbps, compared to Where2comm's accuracy with bandwidth less than 5 Mbps. So the use of intermediate-late hybrid collaboration can greatly improve the accuracy lower bound under low-bandwidth conditions." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.522, + 0.484, + 0.838 + ], + "angle": 0, + "content": "Based on these insights, we propose CoSDH, a novel communication-efficient collaborative perception framework for 3D object detection. As shown in Fig. 2, CoSDH consists of three key components: i) Supply-demand-aware information selection, which chooses sparse yet crucial regions for collaboration; ii) Intermediate feature transmission and fusion, which transmits and effectively aggregates information from multiple agents in a communication-efficient manner; iii) Confidence-aware late fusion, which compensates for the intermediate fusion results at a minimal communication cost to improve accuracy. To evaluate CoSDH, we conducted extensive experiments on the simulation datasets OPV2V [36], V2XSim [20], and the real-world dataset DAIR-V2X [42]. The experimental results show that i) In the baseline case, CoSDH uses less bandwidth while achieving satisfactory accuracy, demonstrating a better accuracy-bandwidth trade-off; ii) Under the simulated real-world communication condition with a total bandwidth limit of 27 Mbps, CoSDH experiences less accuracy degradation and outperforms other methods while using less bandwidth." + }, + { + "type": "text", + "bbox": [ + 0.109, + 0.84, + 0.411, + 0.854 + ], + "angle": 0, + "content": "In summary, our contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.856, + 0.483, + 0.901 + ], + "angle": 0, + "content": "- We present CoSDH, an innovative communication-efficient collaborative perception framework with better accuracy-bandwidth trade-offs for 3D object detection." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.355, + 0.905, + 0.399 + ], + "angle": 0, + "content": "- We propose a novel supply-demand-aware information selection module, further refining the collaboration area selection to achieve more efficient communication." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.4, + 0.905, + 0.46 + ], + "angle": 0, + "content": "- We design a novel intermediate-late hybrid collaborative perception paradigm, where confidence-aware late fusion compensates for the intermediate fusion results to maintain high accuracy under low-bandwidth conditions." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.461, + 0.906, + 0.551 + ], + "angle": 0, + "content": "- We conduct extensive experiments across multiple datasets with bandwidth constraints set closer to real-world conditions, and CoSDH achieves state-of-the-art detection accuracy along with optimal bandwidth trade-offs, demonstrating its feasibility in real-world collaborative perception scenarios." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.355, + 0.906, + 0.551 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.557, + 0.655, + 0.572 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.577, + 0.706, + 0.593 + ], + "angle": 0, + "content": "2.1. 3D Object Detection" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.599, + 0.907, + 0.901 + ], + "angle": 0, + "content": "3D object detection, a key technology in autonomous driving, identifies and localizes objects in 3D scenes using environmental data from onboard sensors and can be categorized into image-based, point-cloud-based, and multimodal-fusion-based methods [26]. Image-based methods can be further classified into monocular [16, 33], stereo [6, 18], and multi-view [13, 44, 45] approaches based on the number of onboard cameras. These methods leverage the rich color and texture information of image, along with its dense data representation, to achieve cost-effective 3D perception for autonomous driving. However, image-based methods are inherently limited by the lack of depth information, which restricts their performance. In contrast, point-cloud-based methods, which adopt voxel-based [7, 39, 48], raw point cloud [27, 28, 46], or point-voxel hybrid [5, 29, 30] representations, can fully leverage the precise 3D geometric information provided by LiDAR, significantly improving perception capabilities. Considering that point-cloud-based methods suffer from sparse characteristic and lack rich semantic information, multimodal-fusion-based meth" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6835" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.122 + ], + "angle": 0, + "content": "ods [23, 41, 47] further enhance performance by leveraging the complementary advantages of different modalities." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.126, + 0.484, + 0.248 + ], + "angle": 0, + "content": "However, these single-agent-based 3D object detection methods are limited by sensor range and susceptible to occlusion, making them unable to detect objects that are further away or completely occluded. This paper focuses on collaborative-perception-based 3D object detection methods, which improve detection performance by supplementing the limitations of single-agent detection with information from other agents." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.268, + 0.32, + 0.284 + ], + "angle": 0, + "content": "2.2. Collaborative Perception" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.293, + 0.485, + 0.612 + ], + "angle": 0, + "content": "Collaborative perception can be categorized into early collaboration, intermediate collaboration, and late collaboration based on the collaboration timing. Early collaboration [2, 4, 42] shares raw perception data, providing good perception accuracy but with high bandwidth. Late collaboration [20, 36] shares perception results, significantly reducing bandwidth, but leads to a decline in perception accuracy. Intermediate collaboration operates at the feature level and can achieve a better trade-off between accuracy and bandwidth by adjusting the intermediate features transmitted [3, 10, 11, 19, 21, 22, 32, 35-37], which is why it has been widely studied. Some of this research has focused on improving perception accuracy. FCooper [3] and CoFF [10] used manual modeling to fuse multi-agent features. Who2com [22] and When2com [21] perform selective communication and use attention-based fusion. V2VNet [32] and DiscoNet [19] employ communication graph-based fusion methods. However, these methods typically transmit complete, uncompressed intermediate BEV features, leading to enormous bandwidth requirements, which makes them challenging to apply in practice." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.614, + 0.484, + 0.902 + ], + "angle": 0, + "content": "To address the large bandwidth demand in collaborative perception, AttFuse [36] was the first to use autoencoders to compress intermediate features along the channel dimension, which was later adopted by V2X-ViT [35], CoBEVT [37] and others. However, this method leads to significant accuracy degradation at high compression rates. Where2comm [11] reduces bandwidth requirements by selecting sparse but important foreground regions for collaboration while maintaining perception accuracy. However, as bandwidth is further limited, perception accuracy still rapidly degrades, potentially even falling below the accuracy of simple late collaboration methods. To address this issue, this paper proposes a hybrid collaborative method based on both intermediate and late collaboration, which efficiently compensates for the intermediate collaborative results using late collaboration under bandwidth constraints. Furthermore, we also adopt and improve methods based on auto-encoders and information selection to save bandwidth while maintaining accuracy." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.091, + 0.605, + 0.105 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.11, + 0.696, + 0.124 + ], + "angle": 0, + "content": "3.1. Problem Definition" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.132, + 0.907, + 0.284 + ], + "angle": 0, + "content": "In this paper, we consider the problem of collaborative perception with \\(N\\) agents. Let \\(X_{i}\\) and \\(y_{i}\\) represent the raw observation and the corresponding ground truth supervision of the \\(i\\)-th agent, respectively, and let \\(P_{j \\rightarrow i}\\) be the message sent from agent \\(j\\) to agent \\(i\\). In the collaborative perception, agent \\(i\\) aggregates its own observations and the messages \\(\\{P_{j \\rightarrow i}\\}_{j=1}^{N}\\) sent from other agents to perform 3D object detection task. Our goal is to maximize the collaborative perception 3D object detection accuracy while ensuring that each agent has a communication budget \\(B\\):" + }, + { + "type": "equation", + "bbox": [ + 0.557, + 0.293, + 0.907, + 0.373 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\xi_ {\\Phi} (B) = \\arg \\max _ {\\theta} \\sum_ {i = 1} ^ {N} g \\left(\\Phi_ {\\theta} \\left(X _ {i}, \\left\\{P _ {j \\rightarrow i} \\right\\} _ {j = 1} ^ {N}\\right), y _ {i}\\right), (1) \\\\ \\text {s . t .} \\sum_ {j = 1} ^ {N} \\left| P _ {j \\rightarrow i} \\right| \\leq B (2) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.382, + 0.907, + 0.457 + ], + "angle": 0, + "content": "where \\(\\Phi_{\\theta}\\) represents the collaborative perception 3D object detection model, \\(\\theta\\) denotes the model parameters, \\(|P_{j\\rightarrow i}|\\) represents the communication volume of the message sent from agent \\(j\\) to agent \\(i\\), and \\(g(\\cdot ,\\cdot)\\) denotes the 3D object detection evaluation metric." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.467, + 0.71, + 0.481 + ], + "angle": 0, + "content": "3.2. Overall Architecture" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.489, + 0.907, + 0.745 + ], + "angle": 0, + "content": "The overall architecture of the proposed \\(\\mathrm{COSDH}\\) is shown in Fig. 2. Each agent first processes its locally observed point cloud \\(X_{i}\\) through a backbone network based on PointPillar [17] and a demand generator to obtain multi-scale BEV features \\(\\{F_i^{(l)}\\}_{l = 1,2,\\dots,L}\\) and a demand mask \\(D_{i}\\), respectively. Considering the collaboration between Ego agent \\(i\\) and collaborating agent \\(j\\), agent \\(j\\) generates a supply mask \\(S_{j}\\) from its multi-scale features \\(\\{F_j^{(l)}\\}_{l = 1,2,\\dots,L}\\) via a supply generator, and multiplies it element-wise with the received demand matrix to obtain the supply-demand mask \\(M_{j\\rightarrow i}\\). Agent \\(j\\) then performs supply-demand-aware information selection by multiplying \\(\\{F_j^{(l)}\\}_{l = 1,2,\\dots,L}\\) with \\(M_{j\\rightarrow i}\\) element-wise to obtain sparse spatial features \\(\\{Z_j^{(l)}\\}_{l = 1,2,\\dots,L}\\). Subsequently, agent \\(j\\) compresses the features through an autoencoder, sending the non-zero parts of the features along with their corresponding coordinates as the message \\(P_{j\\rightarrow i}\\) to agent \\(i\\)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.745, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Upon receiving \\( P_{j \\to i} \\), agent \\( i \\) first decodes the features to restore their dimensions and then fuses them with its local features \\( \\{F_i^{(l)}\\}_{l = 1,2,\\dots,L} \\) across multiple scales to obtain the fused features \\( \\{F_i^{(l)}\\}_{l = 1,2,\\dots,L} \\), which are then passed to the detection head for intermediate collaborative detection results \\( \\tilde{y}_i \\). Afterward, we apply confidence-aware late fusion: agent \\( j \\) filters and suppresses its own detection results \\( y_j \\) based on confidence and sends them to agent \\( i \\) for late fusion, yielding the final hybrid collaborative perception detection results \\( \\hat{y}_i \\)." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6836" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.104, + 0.089, + 0.895, + 0.301 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.312, + 0.908, + 0.357 + ], + "angle": 0, + "content": "Figure 2. The overall architecture of \\(\\mathrm{COSDH}\\). The Supply-Demand-Aware Information Selection module selects sparse but important information, which is then further compressed by the Message Compression module to achieve efficient communication. Confidence-Aware Late Fusion compensates for the intermediate fusion detection results to improve accuracy." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.372, + 0.479, + 0.387 + ], + "angle": 0, + "content": "3.3. Supply-Demand-Aware Information Selection" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.397, + 0.484, + 0.593 + ], + "angle": 0, + "content": "Previous methods such as Where2comm [11] and How2comm [40] generally use symmetric supply-demand relationships to select sparse features for collaboration, believing that areas with high foreground confidence in the collaborating agent's view should be provided, and conversely, areas with low foreground confidence in the Ego agent's view need collaboration. However, we believe that areas with low foreground confidence from the Ego's perspective can be divided into areas that are hard to observe and those that can be observed but belong to the background. The latter do not require collaboration. Based on this insight, we propose a novel supply-demand-aware information selection method." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.596, + 0.484, + 0.852 + ], + "angle": 0, + "content": "The demand mask \\( D_{i} \\) indicates where agent \\( i \\) needs information from collaborating agents. Intuitively, the agent requires information from areas that are distant or occluded, which have the common characteristic of having low point cloud density or no point cloud at all. For agent \\( i \\), we consider using the number of point clouds in each pillar to represent point cloud density, and we map it to the range [0,1], i.e., \\( A_{i} \\in [0,1]^{H \\times W} \\), where \\( H \\) and \\( W \\) represent the number of Pillars along the height and width dimensions. We then select areas where the point cloud density is below a threshold \\( \\varepsilon_{a} \\) to obtain the demand mask for agent \\( i \\), \\( D_{i} = A_{i} < \\varepsilon_{a} \\in \\{0,1\\}^{H \\times W} \\). The demand mask indicates where agent \\( i \\) has poor perception and needs collaborative information from other agents. Filtering information from other agents using the demand mask not only helps save bandwidth but also avoids interference from other agents' information in well-perceived areas." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "For the object detection task, foreground information is more valuable. Providing sparse foreground features can effectively assist other agents in supplementing undetected" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.371, + 0.916, + 0.572 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.583, + 0.908, + 0.625 + ], + "angle": 0, + "content": "Figure 3. Comparison of our multi-scale compression and fusion with other methods. It can achieve thorough fusion with smaller communication volume." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.647, + 0.907, + 0.77 + ], + "angle": 0, + "content": "and incomplete targets while using less bandwidth. Following previous work [11], we use the spatial confidence map \\( C_i \\in [0,1]^{H \\times W} \\) output by the detection head to select potential foreground areas. We use a supply threshold \\( \\varepsilon_c \\) to obtain the supply mask \\( S_i^{(l)} = C_i > \\varepsilon_c \\in \\{0,1\\}^{H \\times W} \\). By adjusting the threshold, we can dynamically adjust the bandwidth used for collaborative perception to adapt to varying communication conditions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.772, + 0.909, + 0.9 + ], + "angle": 0, + "content": "During collaboration, agent \\(j\\) generates a binary supply-demand selection mask \\(M_{j\\rightarrow i} = D_i\\odot S_j\\in \\{0,1\\}^{H\\times W}\\) based on its supply mask \\(S_{j}\\) and agent \\(i\\) 's demand mask \\(D_{i}\\) and samples it to multiply element-wise with multi-scale BEV features \\(\\{F_j^{(l)}\\}_{l = 1,2,\\dots,L}\\), obtaining sparse features \\(\\{Z_{j\\rightarrow i}^{(l)}\\}_{l = 1,2,\\dots,L}\\). During communication, only the non-zero parts and their corresponding coordinates need to be transmitted." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "6837" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.387, + 0.108 + ], + "angle": 0, + "content": "3.4. Message Compression and Fusion" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.116, + 0.484, + 0.373 + ], + "angle": 0, + "content": "To further reduce communication bandwidth, we also use autoencoders to compress intermediate features along the channel dimension before communication, while making further improvements to existing compression and fusion schemes. As shown in Fig. 3, traditional single-scale fusion schemes [35, 36] only compress and fuse a single-layer BEV feature map before passing it to the detection head, resulting in insufficient fusion. CoAlign [25] proposed a solution that performs layer-wise fusion on multi-scale features, addressing the issue of insufficient fusion. However, it compresses large-scale single-layer features extracted after Pillar Feature Extraction (PFE), which leads to higher bandwidth. Therefore, we propose a multi-scale compression and fusion approach, where separate autoencoders are designed for each scale of features to perform compression, followed by layer-wise fusion. This approach enables more thorough fusion while using less bandwidth." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.377, + 0.484, + 0.499 + ], + "angle": 0, + "content": "For message fusion, we use Max fusion, which has two main advantages: i) it is computationally simple and efficient, with the complexity increasing linearly with the number of agents; ii) max fusion selects the maximum value of features from multiple agents, achieving information complementarity. Considering the collaboration between Ego agent \\(i\\) and collaborating agent \\(j\\), the process of message compression and fusion can be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.149, + 0.517, + 0.483, + 0.543 + ], + "angle": 0, + "content": "\\[\nZ _ {j \\rightarrow i} ^ {\\prime (l)} = f _ {e n c o d e} ^ {(l)} \\left(Z _ {j \\rightarrow i} ^ {(l)}\\right) \\in \\mathbb {R} ^ {\\frac {C _ {l}}{c _ {0}} \\times H _ {l} \\times W _ {l}} \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.149, + 0.545, + 0.483, + 0.568 + ], + "angle": 0, + "content": "\\[\nF _ {j \\rightarrow i} ^ {(l)} = f _ {\\text {d e c o d e}} ^ {(l)} \\left(Z _ {j \\rightarrow i} ^ {' (l)}\\right) \\in \\mathbb {R} ^ {C _ {l} \\times H _ {l} \\times W _ {l}} \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.149, + 0.57, + 0.483, + 0.592 + ], + "angle": 0, + "content": "\\[\nF _ {j \\rightarrow i} ^ {\\prime (l)} = f _ {\\text {t r a n s f o r m}} \\left(F _ {j \\rightarrow i} ^ {(l)}, \\xi_ {j \\rightarrow i}\\right) \\in \\mathbb {R} ^ {C _ {l} \\times H _ {l} \\times W _ {l}} \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.154, + 0.595, + 0.482, + 0.62 + ], + "angle": 0, + "content": "\\[\n\\tilde {F _ {i} ^ {(l)}} = \\max \\left(F _ {i} ^ {(l)}, \\left\\{F _ {j \\rightarrow i} ^ {\\prime (l)} \\right\\} _ {j \\neq i}\\right) \\in \\mathbb {R} ^ {C _ {l} \\times H _ {l} \\times W _ {l}} \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.642, + 0.484, + 0.903 + ], + "angle": 0, + "content": "where \\(Z_{j\\rightarrow i}^{(l)}\\in \\mathbb{R}^{C_l\\times H_l\\times W_l}\\) represents the sparse features selected based on supply-demand relationships from the previous stage, and \\(C_l,H_l,W_l\\) denote the channel, height, and width dimensions of the \\(l\\) -th layer feature map. \\(f_{encode}^{(l)},f_{decode}^{(l)}\\) represent the encoder and decoder of the \\(l\\) -th layer autoencoder, \\(c_{0}\\) is the compression ratio in the channel dimension, and \\(f_{transform}\\) represents coordinate transformation. During communication, \\(\\{Z_{j\\rightarrow i}^{\\prime (l)}\\}_{l = 1,2,\\ldots ,L}\\) is first converted from float32 to float16, and then only the non-zero parts and corresponding coordinates are transmitted to save bandwidth. Upon receiving the message, agent \\(i\\) decodes the feature dimensions, then aligns the features to its own coordinate system using the coordinate transformation matrix \\(\\xi_{j\\rightarrow i}\\) to obtain \\(F_{j\\rightarrow i}^{\\prime (l)}\\) , and finally fuses its own features with the collaborating agent's features using max fusion to obtain the fused features \\(F_{i}^{(l)}\\)" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.093, + 0.915, + 0.273 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.286, + 0.907, + 0.329 + ], + "angle": 0, + "content": "Figure 4. Our confidence-aware late fusion. It filters detection results based on confidence and suppresses suboptimal results from collaborative agents, improving overall detection accuracy." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.36, + 0.787, + 0.374 + ], + "angle": 0, + "content": "3.5. Confidence-Aware Late Fusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.388, + 0.907, + 0.525 + ], + "angle": 0, + "content": "Existing collaborative perception methods focus on the singular intermediate collaboration architecture and neglect the advantages of late collaboration. As shown in Table 1, the experimental results indicate that late fusion can demonstrate acceptable perception accuracy with extremely low bandwidth on the OPV2V [36] and V2XSim [20] datasets. Therefore, late fusion can be used to compensate for the results of intermediate fusion, achieving a better accuracy-bandwidth trade-off." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.532, + 0.909, + 0.699 + ], + "angle": 0, + "content": "Naive late collaboration methods [20, 36] directly merge results from other agents and then apply NMS (Non-Maximum Suppression) to deduplicate and obtain final results. This approach can effectively improve recall for collaborative perception object detection; however, during the merging process, it may introduce some low-confidence false positives, and suboptimal detection results from collaborating agents may override the Ego agent's detection results, lowering precision and consequently decreasing final AP. The lower AP in the late fusion method in Table 1 on the DAIR-V2X [42] dataset is due to this issue." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.909, + 0.903 + ], + "angle": 0, + "content": "To address this problem, we propose a new confidence-aware late fusion method. As shown in Fig. 4, during late fusion, we first filter based on the confidence of the target boxes, discarding target boxes from other agents with confidence lower than \\(\\varepsilon_{l}\\). Furthermore, considering that if both the Ego agent and other agents detect the same target, the detection result from the Ego agent, which has undergone intermediate fusion, is of higher quality. Therefore, we suppress the detection results from other agents to prevent their lower-quality results from degrading the Ego agent's better detection results. Specifically, before merging the detection results, we multiply the confidence of target boxes from other agents by a coefficient \\(\\beta \\in (0,1)\\)." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "6838" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.088, + 0.904, + 0.346 + ], + "angle": 0, + "content": "
SettingMethodOPV2V [36]V2XSim [20]DAIR-V2X [42]
AP@0.5↑AP@0.7↑BD↓AP@0.5↑AP@0.7↑BD↓AP@0.5↑AP@0.7↑BD↓
BasicNo Fusion79.78%67.16%0.0 Mbps70.31%58.55%0.0 Mbps66.51%55.46%0.0 Mbps
Early Fusion95.05%88.98%83.1 Mbps95.68%88.03%55.5 Mbps74.52%59.22%50.2 Mbps
Late Fusion94.70%87.18%0.2 Mbps86.88%78.13%0.1 Mbps67.86%50.47%0.2 Mbps
No LimitWhen2com [21]91.75%81.77%1,320.0 Mbps72.65%62.92%250.0 Mbps64.08%49.14%984.8 Mbps
FCooper [3]90.06%74.03%2,640.0 Mbps72.96%57.61%500.0 Mbps74.58%56.47%1,968.8 Mbps
AttFuse [36]94.31%82.03%2,640.0 Mbps78.06%64.84%500.0 Mbps73.80%56.86%1,968.8 Mbps
V2VNet [32]96.66%92.44%5,280.0 Mbps88.97%85.18%1,000.0 Mbps66.63%47.39%3,937.5 Mbps
DiscoNet [19]90.93%78.90%2,640.0 Mbps77.34%68.77%500.0 Mbps73.58%58.45%1,968.8 Mbps
V2XViT [35]95.87%89.88%2,640.0 Mbps89.01%80.26%500.0 Mbps76.68%57.57%1,920.0 Mbps
Where2comm [11]95.59%91.39%48.7 Mbps88.18%83.66%27.5 Mbps76.13%60.16%172.3 Mbps
CoAlign [25]96.63%92.63%2,640.0 Mbps88.87%85.23%500.0 Mbps78.06%63.09%1,968.8 Mbps
CoSDH96.83%92.99%13.4 Mbps89.23%86.31%1.1 Mbps76.75%63.85%7.1 Mbps
BD ≤ 6.75 MbpsWhen2com [21]79.97%50.88%5.2 Mbps62.02%43.37%4.0 Mbps62.37%44.01%3.8 Mbps
FCooper [3]90.37%72.92%5.2 Mbps76.52%61.83%4.0 Mbps67.71%47.65%3.8 Mbps
AttFuse [36]93.45%80.02%5.2 Mbps84.58%71.36%4.0 Mbps71.06%49.57%3.8 Mbps
V2VNet [32]95.71%87.16%5.2 Mbps85.66%73.72%4.0 Mbps66.49%45.61%3.8 Mbps
DiscoNet [19]90.00%76.72%5.2 Mbps78.04%68.18%4.0 Mbps70.83%53.40%3.8 Mbps
V2XViT [35]95.85%87.13%5.2 Mbps88.94%81.47%4.0 Mbps71.00%52.78%3.8 Mbps
Where2comm [11]94.91%89.86%5.4 Mbps87.51%82.12%4.7 Mbps74.98%59.51%5.3 Mbps
CoAlign [25]94.10%85.99%5.2 Mbps88.01%83.97%4.0 Mbps75.26%60.19%3.8 Mbps
CoSDH96.75%92.92%2.0 Mbps89.23%86.31%1.1 Mbps76.47%63.76%1.4 Mbps
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.09, + 0.355, + 0.907, + 0.44 + ], + "angle": 0, + "content": "Table 1. Comparison of detection accuracy and bandwidth of different methods on OPV2V [36], V2XSim [20], and DAIR-V2X [42] datasets. \"BD\" represents the bandwidth required for each collaborative agent, assuming Ego agent collaborates with up to 4 agents and the detection frequency is \\(10\\mathrm{Hz}\\). \"BD \\(\\leq\\) 6.75 Mbps\" is used to simulate real-world communication limits, assuming a total communication bandwidth of 27 Mbps, with each collaborating agent's bandwidth consumption limited to less than 6.75 Mbps. For intermediate collaboration methods without information selection, we provide a compressed version using autoencoders to meet the bandwidth constraints. For intermediate collaboration methods with information selection, the selection ratio is adjusted to meet the bandwidth constraints." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.456, + 0.224, + 0.472 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.479, + 0.401, + 0.495 + ], + "angle": 0, + "content": "4.1. Datasets and Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.503, + 0.483, + 0.608 + ], + "angle": 0, + "content": "Datasets. We evaluate the proposed CoSDH against other methods on three different collaborative perception datasets (OPV2V [36], V2XSim [20], and DAIR-V2X [42]) for LiDAR-based 3D object detection. The datasets include both simulated and real-world scenarios, and cover two types of collaboration: V2V (Vehicle to Vehicle) and V2I (Vehicle to Infrastructure)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.612, + 0.483, + 0.822 + ], + "angle": 0, + "content": "Evaluation Metrics We use the average precision (AP) with intersection-over-union (IoU) thresholds of 0.5 and 0.7 to evaluate the performance of different methods on 3D object detection. We assume the target detection frequency is \\(10\\mathrm{Hz}\\) and calculate the communication bandwidth based on the average data transmitted by each collaborative agent to the Ego agent, in order to evaluate the communication cost of different methods. Specifically, we consider the bandwidth limitations in real-world collaborative perception scenarios, setting the vehicle's communication rate to 27 Mbps [1, 24]. Considering the typical case where the Ego agent collaborates with up to 4 other agents [36], the bandwidth limit for each collaborative agent is \\(27/4 = 6.75\\) Mbps." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Implementation Our experiments are based on the OpenCOOD [36] framework, using PointPillar [17] as the encoder with a grid size of \\((0.4m, 0.4m)\\), and a maximum of 32 points per Pillar. For our method, we set the number of intermediate feature layers to \\(L = 3\\), the demand threshold" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.457, + 0.907, + 0.744 + ], + "angle": 0, + "content": "\\(\\varepsilon_{a} = 4 / 32 = 0.125\\), and the supply threshold \\(\\varepsilon_{c} = 0.01\\). For the OPV2V [36] and DAIR-V2X [42] datasets, the compression rate is \\(c_{0} = 16\\), and for the V2XSim [20] dataset, the compression rate is \\(c_{0} = 8\\) due to its smaller perception range and lower inherent bandwidth requirement. The late fusion threshold is set to \\(\\varepsilon_{l} = 0.3\\), with a suppression coefficient \\(\\beta = 0.9\\) for OPV2V and V2XSim dataset and \\(\\beta = 0.8\\) for DAIR-V2X dataset because of its difficulty. In late fusion, the dense prediction results before NMS are transmitted, as this reduces the computational burden on the collaborating agents. The experiments use the Adam [14] optimizer, with an initial learning rate set between 0.0001 and 0.002 based on the model's testing complexity to ensure proper training. The maximum number of collaborating agents is set to 5. The number of training epochs is set to 40 to ensure model convergence. Other experimental parameters are kept consistent with the OpenCOOD framework. All methods are trained on four NVIDIA GeForce RTX 3090 GPUs." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.757, + 0.736, + 0.772 + ], + "angle": 0, + "content": "4.2.Quantitative Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Benchmark Comparison. Table 1 presents the collaborative 3D object detection accuracy and required bandwidth of the proposed CoSDH compared to previous methods across different datasets. Experimental results show that, with the default uncompressed settings, CoSDH achieves the highest accuracy on the OPV2V [36] and V2XSim [20] datasets while requiring only about 1/100 to 1/1000 of the bandwidth compared to other non-communication-efficient" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "6839" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.089, + 0.365, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.366, + 0.089, + 0.631, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.631, + 0.089, + 0.898, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.261, + 0.365, + 0.42 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.365, + 0.261, + 0.631, + 0.42 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.631, + 0.261, + 0.899, + 0.42 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.431, + 0.908, + 0.473 + ], + "angle": 0, + "content": "Figure 5. Comparison of the trade-off between detection accuracy and bandwidth of different methods on OPV2V [36], V2XSim [20] and DAIR-V2X [42] datasets, CoSDH achieves the best accuracy-bandwidth trade-off. The real-world limitation refers to the total bandwidth limit of 27 Mbps, which means that each collaborative agent does not exceed 6.75 Mbps." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.491, + 0.483, + 0.642 + ], + "angle": 0, + "content": "methods. Although the AP@0.5 of CoSDH on the DAIR-V2X [42] dataset is slightly lower than that of CoAlign [25], it achieves a higher AP@0.7 with only about 1/300 of the bandwidth. Both our CoSDH and Where2comm [11] are communication-efficient methods based on information selection, which enable dynamic accuracy-bandwidth trade-offs by adjusting the selection ratio. The table shows accuracy at a specific bandwidth, and a more detailed comparison of the accuracy-bandwidth curves is provided in the \"Accuracy-Bandwidth Trade-Off Comparison\" part." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.644, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Furthermore, we simulate real-world communication rate limitations. Result shows that \\(\\mathrm{COSDH}\\) achieves improvements in AP@0.7 of \\(3.06\\% / 2.34\\% / 3.75\\%\\) on OPV2V/V2XSim/DAIR-V2X compared to the previous best methods, while using less bandwidth. When comparing the scenarios with and without bandwidth limitations on the OPV2V and DAIR-V2X datasets, we found that \\(\\mathrm{COSDH}\\) achieves less than a \\(0.3\\%\\) decrease in AP under conditions where bandwidth is reduced by \\(80\\%\\) to \\(85\\%\\), exhibiting less accuracy degradation than Where2comm. Most other methods also show varying degrees of accuracy degradation under bandwidth constraints. Interestingly, under bandwidth limitations, some methods show an improvement in accuracy on certain datasets after using autoencoders to compress intermediate features. For example, FCooper [3] shows improved accuracy on the V2XSim dataset, which may be due to the model's initially poor per" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.491, + 0.907, + 0.537 + ], + "angle": 0, + "content": "formance. The compressed version, with the added autoencoder, increases the model's parameter count, thereby enriching its expressive capability." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.554, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Accuracy-Bandwidth Trade-Off Comparison. Fig. 5 shows the accuracy-bandwidth trade-off of the proposed CoSDH and previous advanced methods across different datasets. When the bandwidth grater than 1 Mbps, Where2comm [11] demonstrates a good enough accuracy-bandwidth trade-off, maintaining high detection accuracy as bandwidth decreases. However, as bandwidth further decreases, its accuracy quickly declines, even falling below that of late fusion method. CoSDH can leverage late fusion at low bandwidths to maintain high accuracy, achieving a better performance-bandwidth trade-off. Notably, on the DAIR-V2X dataset, the late fusion method performs worse than the no-fusion case because the naive late fusion method unselectively merges detection results from other agents, introducing a large number of low-quality detections. CoSDH uses a confidence-aware late fusion method to select higher-quality detection results, improving this situation and enhancing perception accuracy. We also observe that the AP@0.7 of CoSDH on the V2XSim dataset initially increases slightly as bandwidth decreases before subsequently declining, indicating that selecting key collaboration areas can help reduce interference from other regions to some extent, thereby slightly improving detection accuracy." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "6840" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.089, + 0.295, + 0.18 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.134, + 0.18, + 0.166, + 0.187 + ], + "angle": 0, + "content": "CoAlign" + }, + { + "type": "image_caption", + "bbox": [ + 0.202, + 0.18, + 0.288, + 0.188 + ], + "angle": 0, + "content": "Bandwidth: 3.85 Mbps" + }, + { + "type": "image", + "bbox": [ + 0.295, + 0.089, + 0.496, + 0.179 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.323, + 0.18, + 0.376, + 0.187 + ], + "angle": 0, + "content": "Where2comm" + }, + { + "type": "image_caption", + "bbox": [ + 0.403, + 0.18, + 0.49, + 0.187 + ], + "angle": 0, + "content": "Bandwidth: 5.26 Mbps" + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.19, + 0.295, + 0.279 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.104, + 0.28, + 0.195, + 0.288 + ], + "angle": 0, + "content": "CoSDH (w/o late fusion)" + }, + { + "type": "image", + "bbox": [ + 0.296, + 0.19, + 0.495, + 0.279 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.336, + 0.28, + 0.365, + 0.287 + ], + "angle": 0, + "content": "CoSDH" + }, + { + "type": "image_caption", + "bbox": [ + 0.403, + 0.28, + 0.489, + 0.288 + ], + "angle": 0, + "content": "Bandwidth: 0.29 Mbps" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.3, + 0.484, + 0.342 + ], + "angle": 0, + "content": "Figure 6. Visualization of detection results under real-world communication rate limitations on the DAIR-V2X [42] dataset. Green represents ground truth box, and red represents predicted box." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.36, + 0.303, + 0.377 + ], + "angle": 0, + "content": "4.3. Qualitative Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.383, + 0.484, + 0.565 + ], + "angle": 0, + "content": "Fig. 6 shows the visualization of detection results for our method compared to other methods under simulated real-world communication rate limitations on the DAIR-V2X dataset. A comparison of the detection results reveals that, while using less bandwidth, our method achieves the same recall rate as CoAlign [25] and Where2comm [11], with fewer false positive predictions and more accurate object localization, demonstrating that our method performs better under low-bandwidth conditions. Comparing the two images at bottom, it can be seen that under low bandwidth conditions, late fusion effectively compensates for the results of intermediate fusion, improving the recall rate of objects." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.576, + 0.255, + 0.59 + ], + "angle": 0, + "content": "4.4. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.599, + 0.483, + 0.854 + ], + "angle": 0, + "content": "Table 2 presents the results of the ablation study on various modules of the proposed method using the OPV2V [36] dataset. The results show that after using the autoencoder for compression, the bandwidth is reduced by a factor of \\( c_{0} \\), and there is no significant loss in perception accuracy, even with a slight improvement in AP@0.7. Converting intermediate features to float16 can nearly halve the bandwidth without significant loss, demonstrating the effectiveness of our proposed message compression module. After performing information selection based on the supply mask, perception accuracy slightly decreases, but the bandwidth is reduced by about \\( 95\\% \\), which is a worthwhile trade-off. Further using the demand mask results in no significant decrease in accuracy but reduces bandwidth by about \\( 10\\% \\). This demonstrates that our supply-demand-aware information selection can reduce bandwidth while maintaining accuracy." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "After adding late fusion, accuracy improves at a relatively small bandwidth cost. However, since this phase uses more bandwidth, the improvement in accuracy is not signif" + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.089, + 0.913, + 0.162 + ], + "angle": 0, + "content": "
CompressionSelectionLate FusionAP@0.5↑AP@0.7↑BD↓
AutoencoderFP16SupplyDemand
96.62%92.62%1,155.00 Mbps
✓96.59%92.99%72.19 Mbps
✓✓96.59%92.99%36.09 Mbps
✓✓✓96.30%92.62%1.99 Mbps
✓✓✓✓96.31%92.60%1.82 Mbps
✓✓✓✓✓96.75%92.92%1.97 Mbps
" + }, + { + "type": "table_caption", + "bbox": [ + 0.512, + 0.173, + 0.907, + 0.271 + ], + "angle": 0, + "content": "Table 2. Ablation study of the modules in \\( \\mathrm{C}\\circ \\mathrm{{SDH}} \\) on the OPV2V dataset. \"Autoencoder\" refers to the use of autoencoders to compress features along the channel dimension, with a compression ratio of \\( {c}_{0} = {16} \\) . \"FP16\" refers to converting features from float32 to float16 for transmission. \"Supply\" and \"Demand\" refer to using supply and demand masks for information selection, and \"Late Fusion\" refers to applying confidence-aware late fusion." + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.291, + 0.913, + 0.361 + ], + "angle": 0, + "content": "
Late Fusionεc0.010.020.030.050.07
AP@0.5↑96.59%96.31%96.19%95.18%93.35%
AP@0.7↑92.95%92.60%92.27%90.67%87.60%
BD↓13.26 Mbps1.82 Mbps0.54 Mbps0.12 Mbps0.05 Mbps
✓AP@0.5↑96.83%96.75%96.72%96.54%96.41%
AP@0.7↑92.99%92.92%92.73%92.23%91.68%
BD↓13.40 Mbps1.97 Mbps0.68 Mbps0.26 Mbps0.19 Mbps
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.371, + 0.907, + 0.426 + ], + "angle": 0, + "content": "Table 3. Ablation study of confidence-aware late fusion under different bandwidths on the OPV2V [36] dataset. The table shows the impact of late fusion on accuracy and bandwidth with different values of \\(\\varepsilon_{c}\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.453, + 0.907, + 0.604 + ], + "angle": 0, + "content": "cant. Table 3 shows the accuracy improvement due to late collaboration under different bandwidth conditions. It can be observed that late collaboration provides more accuracy improvements under lower bandwidth conditions. For example, when the bandwidth used for intermediate fusion is 0.05 Mbps, late collaboration with a bandwidth cost of 0.14 Mbps leads to a \\(3\\%\\) improvement in AP@0.5 and a \\(4\\%\\) improvement in AP@0.7. This is one of the key reasons why our method achieves significantly higher detection accuracy under lower bandwidth compared to Where2comm [11]." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.623, + 0.634, + 0.637 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.647, + 0.907, + 0.829 + ], + "angle": 0, + "content": "In this paper, we propose CoSDH, a novel communication-efficient collaborative perception framework for 3D object detection. By finely modeling the supply-demand relationship between agents, it selects key and sparse regions for collaboration. Additionally, we innovatively incorporate confidence-aware late fusion on top of intermediate collaboration to form an intermediate-late hybrid collaborative perception method. Experiments on multiple datasets show that our method offers a better accuracy-bandwidth trade-off, demonstrating outstanding accuracy under bandwidth constraints close to real-world communication limits, making it highly valuable for practical applications." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.847, + 0.666, + 0.864 + ], + "angle": 0, + "content": "Acknowledgment" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.907, + 0.901 + ], + "angle": 0, + "content": "This work is supported by the National Key Research and Development Plan (2024YFB3309302)." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "6841" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.091, + 0.188, + 0.105 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.111, + 0.483, + 0.137 + ], + "angle": 0, + "content": "[1] Fabio Arena and Giovanni Pau. An overview of vehicular communications. Future internet, 11(2):27, 2019. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.141, + 0.483, + 0.209 + ], + "angle": 0, + "content": "[2] Eduardo Arnold, Mehrdad Dianati, Robert de Temple, and Saber Fallah. Cooperative perception for 3d object detection in driving scenarios using infrastructure sensors. IEEE Transactions on Intelligent Transportation Systems, 23(3): 1852-1864, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.212, + 0.483, + 0.282 + ], + "angle": 0, + "content": "[3] Qi Chen, Xu Ma, Sihai Tang, Jingda Guo, Qing Yang, and Song Fu. F-cooper: Feature based cooperative perception for autonomous vehicle edge computing system using 3d point clouds. In Proceedings of the 4th ACM/IEEE Symposium on Edge Computing, pages 88-100, 2019. 1, 3, 6, 7, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.284, + 0.483, + 0.354 + ], + "angle": 0, + "content": "[4] Qi Chen, Sihai Tang, Qing Yang, and Song Fu. Cooper: Cooperative perception for connected autonomous vehicles based on 3d point clouds. In 2019 IEEE 39th International Conference on Distributed Computing Systems (ICDCS), pages 514-524. IEEE, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.356, + 0.483, + 0.397 + ], + "angle": 0, + "content": "[5] Yilun Chen, Shu Liu, Xiaoyong Shen, and Jiaya Jia. Fast point r-cnn. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9775-9784, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.4, + 0.483, + 0.455 + ], + "angle": 0, + "content": "[6] Yilun Chen, Shu Liu, Xiaoyong Shen, and Jiaya Jia. Dsgn: Deep stereo geometry network for 3d object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12536-12545, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.458, + 0.483, + 0.527 + ], + "angle": 0, + "content": "[7] Jiajun Deng, Shaoshuai Shi, Peiwei Li, Wengang Zhou, Yanyong Zhang, and Houqiang Li. Voxel r-cnn: Towards high performance voxel-based 3d object detection. In Proceedings of the AAAI conference on artificial intelligence, pages 1201-1209, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.529, + 0.483, + 0.583 + ], + "angle": 0, + "content": "[8] Alexey Dosovitskiy, German Ros, Felipe Codevilla, Antonio Lopez, and Vladlen Koltun. Carla: An open urban driving simulator. In Conference on robot learning, pages 1-16. PMLR, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.587, + 0.483, + 0.655 + ], + "angle": 0, + "content": "[9] Sebastian Gräfling, Petri Mähönen, and Janne Riihijärvi. Performance evaluation of the IEEE 1609 wave and the IEEE 802.11 p for vehicular communications. In 2010 second international conference on ubiquitous and future networks (ICUFN), pages 344-348. IEEE, 2010. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.659, + 0.483, + 0.728 + ], + "angle": 0, + "content": "[10] Jingda Guo, Dominic Carrillo, Sihai Tang, Qi Chen, Qing Yang, Song Fu, Xi Wang, Nannan Wang, and Paparao Palacharla. Coff: Cooperative spatial feature fusion for 3-d object detection on autonomous vehicles. IEEE Internet of Things Journal, 8(14):11078-11087, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.73, + 0.483, + 0.798 + ], + "angle": 0, + "content": "[11] Yue Hu, Shaoheng Fang, Zixing Lei, Yiqi Zhong, and Siheng Chen. Where2comm: Communication-efficient collaborative perception via spatial confidence maps. Advances in neural information processing systems, 35:4874-4886, 2022. 1, 2, 3, 4, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.802, + 0.483, + 0.871 + ], + "angle": 0, + "content": "[12] Yue Hu, Juntong Peng, Sifei Liu, Junhao Ge, Si Liu, and Si-heng Chen. Communication-efficient collaborative perception via information filling with codebook. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15481-15490, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[13] Zheng Jiang, Jinqing Zhang, Yanan Zhang, Qingjie Liu, Zhenghui Hu, Baohui Wang, and Yunhong Wang. Fsd-bev:" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.111, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.135 + ], + "angle": 0, + "content": "Foreground self-distillation for multi-view 3d object detection. In European Conference on Computer Vision, pages 110-126. Springer, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.136, + 0.905, + 0.162 + ], + "angle": 0, + "content": "[14] Diederik P Kingma. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.164, + 0.905, + 0.218 + ], + "angle": 0, + "content": "[15] Daniel Krajzewicz, Jakob Erdmann, Michael Behrisch, and Laura Bieker. Recent development and applications of sumosimulation of urban mobility. International journal on advances in systems and measurements, 5(3&4), 2012. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.22, + 0.905, + 0.287 + ], + "angle": 0, + "content": "[16] Jason Ku, Alex D Pon, and Steven L Waslander. Monocular 3d object detection leveraging accurate proposals and shape reconstruction. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11867-11876, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.289, + 0.907, + 0.357 + ], + "angle": 0, + "content": "[17] Alex H Lang, Sourabh Vora, Holger Caesar, Lubing Zhou, Jiong Yang, and Oscar Beijbom. Pointpillars: Fast encoders for object detection from point clouds. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12697-12705, 2019. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.359, + 0.905, + 0.413 + ], + "angle": 0, + "content": "[18] Peiliang Li, Xiaozhi Chen, and Shaojie Shen. Stereo r-cnn based 3d object detection for autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7644-7652, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.415, + 0.905, + 0.469 + ], + "angle": 0, + "content": "[19] Yiming Li, Shunli Ren, Pengxiang Wu, Siheng Chen, Chen Feng, and Wenjun Zhang. Learning distilled collaboration graph for multi-agent perception. Advances in Neural Information Processing Systems, 34:29541-29552, 2021. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.471, + 0.905, + 0.537 + ], + "angle": 0, + "content": "[20] Yiming Li, Dekun Ma, Ziyan An, Zixun Wang, Yiqi Zhong, Siheng Chen, and Chen Feng. V2x-sim: Multi-agent collaborative perception dataset and benchmark for autonomous driving. IEEE Robotics and Automation Letters, 7(4):10914-10921, 2022. 1, 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.539, + 0.905, + 0.607 + ], + "angle": 0, + "content": "[21] Yen-Cheng Liu, Junjiao Tian, Nathaniel Glaser, and Zsolt Kira. When2com: Multi-agent perception via communication graph grouping. In Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition, pages 4106-4115, 2020. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.609, + 0.905, + 0.678 + ], + "angle": 0, + "content": "[22] Yen-Cheng Liu, Junjiao Tian, Chih-Yao Ma, Nathan Glaser, Chia-Wen Kuo, and Zsolt Kira. Who2com: Collaborative perception via learnable handshake communication. In 2020 IEEE International Conference on Robotics and Automation (ICRA), pages 6876-6883. IEEE, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.679, + 0.905, + 0.748 + ], + "angle": 0, + "content": "[23] Zhijian Liu, Haotian Tang, Alexander Amini, Xinyu Yang, Huizi Mao, Daniela L Rus, and Song Han. Bevfusion: Multi-task multi-sensor fusion with unified bird's-eye view representation. In 2023 IEEE international conference on robotics and automation (ICRA), pages 2774-2781. IEEE, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.749, + 0.905, + 0.802 + ], + "angle": 0, + "content": "[24] Jose Manuel Lozano Dominguez and Tomas Jesus Mateo Sanguino. Review on v2x, i2x, and p2x communications and their applications: a comprehensive analysis over time. Sensors, 19(12):2756, 2019. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.804, + 0.905, + 0.873 + ], + "angle": 0, + "content": "[25] Yifan Lu, Quanhao Li, Baoan Liu, Mehrdad Dianati, Chen Feng, Siheng Chen, and Yanfeng Wang. Robust collaborative 3d object detection in presence of pose errors. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 4812-4818. IEEE, 2023. 5, 6, 7, 8, 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[26] Jiageng Mao, Shaoshuai Shi, Xiaogang Wang, and Hongsheng Li. 3d object detection for autonomous driving: A" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.111, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "6842" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.126, + 0.093, + 0.484, + 0.12 + ], + "angle": 0, + "content": "comprehensive survey. International Journal of Computer Vision, 131(8):1909-1963, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.484, + 0.178 + ], + "angle": 0, + "content": "[27] Xuran Pan, Zhuofan Xia, Shiji Song, Li Erran Li, and Gao Huang. 3d object detection with pointformer. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7463-7472, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.18, + 0.482, + 0.246 + ], + "angle": 0, + "content": "[28] Shaoshuai Shi, Xiaogang Wang, and Hongsheng Li. Pointrcnn: 3d object proposal generation and detection from point cloud. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 770-779, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.25, + 0.482, + 0.32 + ], + "angle": 0, + "content": "[29] Shaoshuai Shi, Chaoxu Guo, Li Jiang, Zhe Wang, Jianping Shi, Xiaogang Wang, and Hongsheng Li. Pv-rcnn: Pointvoxel feature set abstraction for 3d object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10529–10538, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.322, + 0.482, + 0.39 + ], + "angle": 0, + "content": "[30] Shaoshuai Shi, Zhe Wang, Jianping Shi, Xiaogang Wang, and Hongsheng Li. From points to parts: 3d object detection from point cloud with part-aware and part-aggregation network. IEEE transactions on pattern analysis and machine intelligence, 43(8):2647-2664, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.392, + 0.482, + 0.476 + ], + "angle": 0, + "content": "[31] Rui Song, Chenwei Liang, Hu Cao, Zhiran Yan, Walter Zimmer, Markus Gross, Andreas Festag, and Alois Knoll. Collaborative semantic occupancy prediction with hybrid feature fusion in connected automated vehicles. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17996-18006, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.478, + 0.482, + 0.56 + ], + "angle": 0, + "content": "[32] Tsun-Hsuan Wang, Sivabalan Manivasagam, Ming Liang, Bin Yang, Wenyuan Zeng, and Raquel Urtasun. V2vnet: Vehicle-to-vehicle communication for joint perception and prediction. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part II 16, pages 605–621. Springer, 2020. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.562, + 0.482, + 0.645 + ], + "angle": 0, + "content": "[33] Yan Wang, Wei-Lun Chao, Divyansh Garg, Bharath Hariharan, Mark Campbell, and Kilian Q Weinberger. Pseudolidar from visual depth estimation: Bridging the gap in 3d object detection for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8445-8453, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.647, + 0.482, + 0.716 + ], + "angle": 0, + "content": "[34] Runsheng Xu, Yi Guo, Xu Han, Xin Xia, Hao Xiang, and Jiaqi Ma. Opencda: an open cooperative driving automation framework integrated with co-simulation. In 2021 IEEE International Intelligent Transportation Systems Conference (ITSC), pages 1155-1162. IEEE, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.718, + 0.482, + 0.786 + ], + "angle": 0, + "content": "[35] Runsheng Xu, Hao Xiang, Zhengzhong Tu, Xin Xia, Ming-Hsuan Yang, and Jiaqi Ma. V2x-vit: Vehicle-to-everything cooperative perception with vision transformer. In European conference on computer vision, pages 107-124. Springer, 2022. 1, 3, 5, 6, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.789, + 0.482, + 0.87 + ], + "angle": 0, + "content": "[36] Runsheng Xu, Hao Xiang, Xin Xia, Xu Han, Jinlong Li, and Jiaqi Ma. Opv2v: An open benchmark dataset and fusion pipeline for perception with vehicle-to-vehicle communication. In 2022 International Conference on Robotics and Automation (ICRA), pages 2583–2589. IEEE, 2022. 1, 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[37] Runsheng Xu, Zhengzhong Tu, Hao Xiang, Wei Shao, Bolei Zhou, and Jiaqi Ma. Cobevt: Cooperative bird's eye view" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.133 + ], + "angle": 0, + "content": "semantic segmentation with sparse transformers. In Conference on Robot Learning, pages 989-1000. PMLR, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.135, + 0.905, + 0.218 + ], + "angle": 0, + "content": "[38] Runsheng Xu, Xin Xia, Jinlong Li, Hanzhao Li, Shuo Zhang, Zhengzhong Tu, Zonglin Meng, Hao Xiang, Xiaoyu Dong, Rui Song, et al. V2v4real: A real-world large-scale dataset for vehicle-to-vehicle cooperative perception. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13712–13722, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.219, + 0.905, + 0.246 + ], + "angle": 0, + "content": "[39] Yan Yan, Yuxing Mao, and Bo Li. Second: Sparsely embedded convolutional detection. Sensors, 18(10):3337, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.247, + 0.905, + 0.315 + ], + "angle": 0, + "content": "[40] Dingkang Yang, Kun Yang, Yuzheng Wang, Jing Liu, Zhi Xu, Rongbin Yin, Peng Zhai, and Lihua Zhang. How2comm: Communication-efficient and collaboration-pragmatic multi-agent perception. Advances in Neural Information Processing Systems, 36, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.317, + 0.905, + 0.398 + ], + "angle": 0, + "content": "[41] Junbo Yin, Jianbing Shen, Runnan Chen, Wei Li, Ruigang Yang, Pascal Frossard, and Wenguan Wang. Is-fusion: Instance-scene collaborative fusion for multimodal 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14905-14915, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.4, + 0.905, + 0.496 + ], + "angle": 0, + "content": "[42] Haibao Yu, Yizhen Luo, Mao Shu, Yiyi Huo, Zebang Yang, Yifeng Shi, Zhenglong Guo, Hanyu Li, Xing Hu, Jirui Yuan, et al. Dair-v2x: A large-scale dataset for vehicleinfrastructure cooperative 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21361-21370, 2022. 2, 3, 5, 6, 7, 8, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.498, + 0.905, + 0.593 + ], + "angle": 0, + "content": "[43] Haibao Yu, Wenxian Yang, Hongzhi Ruan, Zhenwei Yang, Yingjuan Tang, Xu Gao, Xin Hao, Yifeng Shi, Yifeng Pan, Ning Sun, et al. V2x-seq: A large-scale sequential dataset for vehicle-infrastructure cooperative perception and forecasting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5486-5495, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.595, + 0.905, + 0.663 + ], + "angle": 0, + "content": "[44] Jinqing Zhang, Yanan Zhang, Qingjie Liu, and Yunhong Wang. Sa-bev: Generating semantic-aware bird's-eye-view feature for multi-view 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3348-3357, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.665, + 0.905, + 0.719 + ], + "angle": 0, + "content": "[45] Jinqing Zhang, Yanan Zhang, Yunlong Qi, Zehua Fu, Qingjie Liu, and Yunhong Wang. Geobev: Learning geometric bev representation for multi-view 3d object detection. arXiv preprint arXiv:2409.01816, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.721, + 0.905, + 0.775 + ], + "angle": 0, + "content": "[46] Yanan Zhang, Di Huang, and Yunhong Wang. Pc-rgnn: Point cloud completion and graph neural network for 3d object detection. In Proceedings of the AAAI conference on artificial intelligence, pages 3430-3437, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.777, + 0.905, + 0.843 + ], + "angle": 0, + "content": "[47] Yanan Zhang, Jiaxin Chen, and Di Huang. Cat-det: Contrastively augmented transformer for multi-modal 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 908-917, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[48] Chao Zhou, Yanan Zhang, Jiaxin Chen, and Di Huang. Octr: Octtree-based transformer for 3d object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5166-5175, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "6843" + } + ] +] \ No newline at end of file diff --git a/2025/CoSDH_ Communication-Efficient Collaborative Perception via Supply-Demand Awareness and Intermediate-Late Hybridization/82cb35c7-62f2-4e6d-b4eb-be441dfd4b29_origin.pdf b/2025/CoSDH_ Communication-Efficient Collaborative Perception via Supply-Demand Awareness and Intermediate-Late Hybridization/82cb35c7-62f2-4e6d-b4eb-be441dfd4b29_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..4712b228771a8009359cc51c5ae39ca4b82ce6d7 --- /dev/null +++ b/2025/CoSDH_ Communication-Efficient Collaborative Perception via Supply-Demand Awareness and Intermediate-Late Hybridization/82cb35c7-62f2-4e6d-b4eb-be441dfd4b29_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bc4fa67ff436fe0bbb607f66fd59b56fddb9d888bcd5e0e02f662f12ba60a011 +size 2815120 diff --git a/2025/CoSDH_ Communication-Efficient Collaborative Perception via Supply-Demand Awareness and Intermediate-Late Hybridization/full.md b/2025/CoSDH_ Communication-Efficient Collaborative Perception via Supply-Demand Awareness and Intermediate-Late Hybridization/full.md new file mode 100644 index 0000000000000000000000000000000000000000..0f9953ecf0283f32052979c43eab7c03f859f445 --- /dev/null +++ b/2025/CoSDH_ Communication-Efficient Collaborative Perception via Supply-Demand Awareness and Intermediate-Late Hybridization/full.md @@ -0,0 +1,271 @@ +# CoSDH: Communication-Efficient Collaborative Perception via Supply-Demand Awareness and Intermediate-Late Hybridization + +Junhao Xu $^{1*}$ , Yanan Zhang $^{2*}$ , Zhi Cai $^{1}$ , Di Huang $^{1\dagger}$ + +$^{1}$ State Key Laboratory of Complex and Critical Software Environment, Beihang University, Beijing, China +$^{2}$ School of Computer Science and Information Engineering, Hefei University of Technology, Hefei, China + +junhaoxu@buaa.edu.cn, yananzhang@hfut.edu.cn, {caizhi97, dhuang}@buaa.edu.cn + +# Abstract + +Multi-agent collaborative perception enhances perceptual capabilities by utilizing information from multiple agents and is considered a fundamental solution to the problem of weak single-vehicle perception in autonomous driving. However, existing collaborative perception methods face a dilemma between communication efficiency and perception accuracy. To address this issue, we propose a novel communication-efficient collaborative perception framework based on supply-demand awareness and intermediate-late hybridization, dubbed as CoSDH. By modeling the supply-demand relationship between agents, the framework refines the selection of collaboration regions, reducing unnecessary communication cost while maintaining accuracy. In addition, we innovatively introduce the intermediate-late hybrid collaboration mode, where late-stage collaboration compensates for the performance degradation in collaborative perception under low communication bandwidth. Extensive experiments on multiple datasets, including both simulated and real-world scenarios, demonstrate that CoSDH achieves state-of-the-art detection accuracy and optimal bandwidth trade-offs, delivering superior detection precision under real communication bandwidths, thus proving its effectiveness and practical applicability. The code will be released at https://github.com/Xu2729/CoSDH. + +# 1. Introduction + +Collaborative perception allows multiple agents to exchange complementary perception information. It fundamentally addresses the issues of limited perception range, sensor blind spots, and occlusion from obstacles inherent in single-agent perception, thereby improving both the range and accuracy of perception. Recent studies have demonstrated that collaborative perception can be applied to var + +ious autonomous driving tasks, including 3D object detection [3, 35, 36], semantic segmentation [20, 37], 3D occupancy prediction [31], and trajectory prediction [43], exhibiting enhanced performance and improved robustness to occlusions, thus making it a crucial approach for advancing autonomous driving systems. + +Communication efficiency is a key issue in collaborative perception. An early approach to improving communication efficiency was to use autoencoders to compress the intermediate features that need to be transmitted [36]. Later, communication-efficient methods such as Where2comm [11] were proposed, which select important and sparse foreground regions for collaboration. However, these methods still lack sufficient precision in area selection and require more bandwidth than the real-world constraints1. As shown in Fig. 1a, when the bandwidth is limited to 27 Mbps, Where2comm [11] experiences a decrease of about $1.5\%$ in average precision (AP@0.7) on the OPV2V [36] dataset with an intersection-over-union (IoU) threshold of 0.7. When the bandwidth is further limited to 5 Mbps and 1 Mbps, AP@0.7 drops by approximately $7.5\%$ and $14.5\%$ , respectively, even falling below the accuracy of late fusion methods at the result level. This makes it difficult for existing collaborative perception methods to meet the demands of practical applications. + +In this paper, we first analyze the problem of selecting collaboration areas and refine this selection based on the supply-demand relationship between agents. As shown in Fig. 1b, for Ego, low foreground confidence areas can be divided into three types: Area 1 (low confidence due to absence of foreground), Area 2 (low confidence due to occlusion), and Area 3 (low confidence due to sparsity of the point cloud). Although all three areas appear as background from the Ego's perspective, Area 1 is well-observed and does not require collaboration, while Areas 2 and 3, + +![](images/06fba8aec52e2e0d60b88c5175542a2224d5edff453a88d56a00d9dd8f457a17.jpg) +(a) The 3D detection accuracy of Where2comm [11] and CoSDH on OPV2V dataset [36] with different bandwidth limit. Assume the number of collaborative agents is 4 and detection frequency is $10\mathrm{Hz}$ . +Figure 1. Issues of existing communication-efficient collaborative perception methods. + +![](images/cfafa97c36fe5a934c2101d62e17fb153ddc49d8ce0238f4393af0acde4d72be.jpg) +(b) Point cloud map and classification of areas with low foreground confidence. Although Area 1 has low confidence due to the absence of foreground, it allows for good observation without collaboration. In contrast, Areas 2 and 3 exhibit poor observation and require collaboration, with Area 2 having low confidence due to occlusion and Area 3 due to sparsity of the point cloud. + +with poorer observation, need collaboration. Additionally, to address the issue of significant accuracy degradation under real bandwidth constraints, we propose that compensating intermediate collaboration results with late fusion is a feasible solution. As shown in Fig. 1a, late fusion achieves much higher accuracy than Where2comm [11] even with a bandwidth of approximately 0.7 Mbps, compared to Where2comm's accuracy with bandwidth less than 5 Mbps. So the use of intermediate-late hybrid collaboration can greatly improve the accuracy lower bound under low-bandwidth conditions. + +Based on these insights, we propose CoSDH, a novel communication-efficient collaborative perception framework for 3D object detection. As shown in Fig. 2, CoSDH consists of three key components: i) Supply-demand-aware information selection, which chooses sparse yet crucial regions for collaboration; ii) Intermediate feature transmission and fusion, which transmits and effectively aggregates information from multiple agents in a communication-efficient manner; iii) Confidence-aware late fusion, which compensates for the intermediate fusion results at a minimal communication cost to improve accuracy. To evaluate CoSDH, we conducted extensive experiments on the simulation datasets OPV2V [36], V2XSim [20], and the real-world dataset DAIR-V2X [42]. The experimental results show that i) In the baseline case, CoSDH uses less bandwidth while achieving satisfactory accuracy, demonstrating a better accuracy-bandwidth trade-off; ii) Under the simulated real-world communication condition with a total bandwidth limit of 27 Mbps, CoSDH experiences less accuracy degradation and outperforms other methods while using less bandwidth. + +In summary, our contributions are as follows: + +- We present CoSDH, an innovative communication-efficient collaborative perception framework with better accuracy-bandwidth trade-offs for 3D object detection. + +- We propose a novel supply-demand-aware information selection module, further refining the collaboration area selection to achieve more efficient communication. +- We design a novel intermediate-late hybrid collaborative perception paradigm, where confidence-aware late fusion compensates for the intermediate fusion results to maintain high accuracy under low-bandwidth conditions. +- We conduct extensive experiments across multiple datasets with bandwidth constraints set closer to real-world conditions, and CoSDH achieves state-of-the-art detection accuracy along with optimal bandwidth trade-offs, demonstrating its feasibility in real-world collaborative perception scenarios. + +# 2. Related Work + +# 2.1. 3D Object Detection + +3D object detection, a key technology in autonomous driving, identifies and localizes objects in 3D scenes using environmental data from onboard sensors and can be categorized into image-based, point-cloud-based, and multimodal-fusion-based methods [26]. Image-based methods can be further classified into monocular [16, 33], stereo [6, 18], and multi-view [13, 44, 45] approaches based on the number of onboard cameras. These methods leverage the rich color and texture information of image, along with its dense data representation, to achieve cost-effective 3D perception for autonomous driving. However, image-based methods are inherently limited by the lack of depth information, which restricts their performance. In contrast, point-cloud-based methods, which adopt voxel-based [7, 39, 48], raw point cloud [27, 28, 46], or point-voxel hybrid [5, 29, 30] representations, can fully leverage the precise 3D geometric information provided by LiDAR, significantly improving perception capabilities. Considering that point-cloud-based methods suffer from sparse characteristic and lack rich semantic information, multimodal-fusion-based meth + +ods [23, 41, 47] further enhance performance by leveraging the complementary advantages of different modalities. + +However, these single-agent-based 3D object detection methods are limited by sensor range and susceptible to occlusion, making them unable to detect objects that are further away or completely occluded. This paper focuses on collaborative-perception-based 3D object detection methods, which improve detection performance by supplementing the limitations of single-agent detection with information from other agents. + +# 2.2. Collaborative Perception + +Collaborative perception can be categorized into early collaboration, intermediate collaboration, and late collaboration based on the collaboration timing. Early collaboration [2, 4, 42] shares raw perception data, providing good perception accuracy but with high bandwidth. Late collaboration [20, 36] shares perception results, significantly reducing bandwidth, but leads to a decline in perception accuracy. Intermediate collaboration operates at the feature level and can achieve a better trade-off between accuracy and bandwidth by adjusting the intermediate features transmitted [3, 10, 11, 19, 21, 22, 32, 35-37], which is why it has been widely studied. Some of this research has focused on improving perception accuracy. FCooper [3] and CoFF [10] used manual modeling to fuse multi-agent features. Who2com [22] and When2com [21] perform selective communication and use attention-based fusion. V2VNet [32] and DiscoNet [19] employ communication graph-based fusion methods. However, these methods typically transmit complete, uncompressed intermediate BEV features, leading to enormous bandwidth requirements, which makes them challenging to apply in practice. + +To address the large bandwidth demand in collaborative perception, AttFuse [36] was the first to use autoencoders to compress intermediate features along the channel dimension, which was later adopted by V2X-ViT [35], CoBEVT [37] and others. However, this method leads to significant accuracy degradation at high compression rates. Where2comm [11] reduces bandwidth requirements by selecting sparse but important foreground regions for collaboration while maintaining perception accuracy. However, as bandwidth is further limited, perception accuracy still rapidly degrades, potentially even falling below the accuracy of simple late collaboration methods. To address this issue, this paper proposes a hybrid collaborative method based on both intermediate and late collaboration, which efficiently compensates for the intermediate collaborative results using late collaboration under bandwidth constraints. Furthermore, we also adopt and improve methods based on auto-encoders and information selection to save bandwidth while maintaining accuracy. + +# 3. Method + +# 3.1. Problem Definition + +In this paper, we consider the problem of collaborative perception with $N$ agents. Let $X_{i}$ and $y_{i}$ represent the raw observation and the corresponding ground truth supervision of the $i$ -th agent, respectively, and let $P_{j \rightarrow i}$ be the message sent from agent $j$ to agent $i$ . In the collaborative perception, agent $i$ aggregates its own observations and the messages $\{P_{j \rightarrow i}\}_{j=1}^{N}$ sent from other agents to perform 3D object detection task. Our goal is to maximize the collaborative perception 3D object detection accuracy while ensuring that each agent has a communication budget $B$ : + +$$ +\begin{array}{l} \xi_ {\Phi} (B) = \arg \max _ {\theta} \sum_ {i = 1} ^ {N} g \left(\Phi_ {\theta} \left(X _ {i}, \left\{P _ {j \rightarrow i} \right\} _ {j = 1} ^ {N}\right), y _ {i}\right), (1) \\ \text {s . t .} \sum_ {j = 1} ^ {N} \left| P _ {j \rightarrow i} \right| \leq B (2) \\ \end{array} +$$ + +where $\Phi_{\theta}$ represents the collaborative perception 3D object detection model, $\theta$ denotes the model parameters, $|P_{j\rightarrow i}|$ represents the communication volume of the message sent from agent $j$ to agent $i$ , and $g(\cdot ,\cdot)$ denotes the 3D object detection evaluation metric. + +# 3.2. Overall Architecture + +The overall architecture of the proposed $\mathrm{COSDH}$ is shown in Fig. 2. Each agent first processes its locally observed point cloud $X_{i}$ through a backbone network based on PointPillar [17] and a demand generator to obtain multi-scale BEV features $\{F_i^{(l)}\}_{l = 1,2,\dots,L}$ and a demand mask $D_{i}$ , respectively. Considering the collaboration between Ego agent $i$ and collaborating agent $j$ , agent $j$ generates a supply mask $S_{j}$ from its multi-scale features $\{F_j^{(l)}\}_{l = 1,2,\dots,L}$ via a supply generator, and multiplies it element-wise with the received demand matrix to obtain the supply-demand mask $M_{j\rightarrow i}$ . Agent $j$ then performs supply-demand-aware information selection by multiplying $\{F_j^{(l)}\}_{l = 1,2,\dots,L}$ with $M_{j\rightarrow i}$ element-wise to obtain sparse spatial features $\{Z_j^{(l)}\}_{l = 1,2,\dots,L}$ . Subsequently, agent $j$ compresses the features through an autoencoder, sending the non-zero parts of the features along with their corresponding coordinates as the message $P_{j\rightarrow i}$ to agent $i$ . + +Upon receiving $P_{j \to i}$ , agent $i$ first decodes the features to restore their dimensions and then fuses them with its local features $\{F_i^{(l)}\}_{l = 1,2,\dots,L}$ across multiple scales to obtain the fused features $\{F_i^{(l)}\}_{l = 1,2,\dots,L}$ , which are then passed to the detection head for intermediate collaborative detection results $\tilde{y}_i$ . Afterward, we apply confidence-aware late fusion: agent $j$ filters and suppresses its own detection results $y_j$ based on confidence and sends them to agent $i$ for late fusion, yielding the final hybrid collaborative perception detection results $\hat{y}_i$ . + +![](images/e61490036051356a3fb3572ee96508d10627968156b304fe0182b559fa40aa0d.jpg) +Figure 2. The overall architecture of $\mathrm{COSDH}$ . The Supply-Demand-Aware Information Selection module selects sparse but important information, which is then further compressed by the Message Compression module to achieve efficient communication. Confidence-Aware Late Fusion compensates for the intermediate fusion detection results to improve accuracy. + +# 3.3. Supply-Demand-Aware Information Selection + +Previous methods such as Where2comm [11] and How2comm [40] generally use symmetric supply-demand relationships to select sparse features for collaboration, believing that areas with high foreground confidence in the collaborating agent's view should be provided, and conversely, areas with low foreground confidence in the Ego agent's view need collaboration. However, we believe that areas with low foreground confidence from the Ego's perspective can be divided into areas that are hard to observe and those that can be observed but belong to the background. The latter do not require collaboration. Based on this insight, we propose a novel supply-demand-aware information selection method. + +The demand mask $D_{i}$ indicates where agent $i$ needs information from collaborating agents. Intuitively, the agent requires information from areas that are distant or occluded, which have the common characteristic of having low point cloud density or no point cloud at all. For agent $i$ , we consider using the number of point clouds in each pillar to represent point cloud density, and we map it to the range [0,1], i.e., $A_{i} \in [0,1]^{H \times W}$ , where $H$ and $W$ represent the number of Pillars along the height and width dimensions. We then select areas where the point cloud density is below a threshold $\varepsilon_{a}$ to obtain the demand mask for agent $i$ , $D_{i} = A_{i} < \varepsilon_{a} \in \{0,1\}^{H \times W}$ . The demand mask indicates where agent $i$ has poor perception and needs collaborative information from other agents. Filtering information from other agents using the demand mask not only helps save bandwidth but also avoids interference from other agents' information in well-perceived areas. + +For the object detection task, foreground information is more valuable. Providing sparse foreground features can effectively assist other agents in supplementing undetected + +![](images/564f796406eb4f74420469c5a1b582df37d4f9e6f9112793c8facaa7a1529079.jpg) +Figure 3. Comparison of our multi-scale compression and fusion with other methods. It can achieve thorough fusion with smaller communication volume. + +and incomplete targets while using less bandwidth. Following previous work [11], we use the spatial confidence map $C_i \in [0,1]^{H \times W}$ output by the detection head to select potential foreground areas. We use a supply threshold $\varepsilon_c$ to obtain the supply mask $S_i^{(l)} = C_i > \varepsilon_c \in \{0,1\}^{H \times W}$ . By adjusting the threshold, we can dynamically adjust the bandwidth used for collaborative perception to adapt to varying communication conditions. + +During collaboration, agent $j$ generates a binary supply-demand selection mask $M_{j\rightarrow i} = D_i\odot S_j\in \{0,1\}^{H\times W}$ based on its supply mask $S_{j}$ and agent $i$ 's demand mask $D_{i}$ and samples it to multiply element-wise with multi-scale BEV features $\{F_j^{(l)}\}_{l = 1,2,\dots,L}$ , obtaining sparse features $\{Z_{j\rightarrow i}^{(l)}\}_{l = 1,2,\dots,L}$ . During communication, only the non-zero parts and their corresponding coordinates need to be transmitted. + +# 3.4. Message Compression and Fusion + +To further reduce communication bandwidth, we also use autoencoders to compress intermediate features along the channel dimension before communication, while making further improvements to existing compression and fusion schemes. As shown in Fig. 3, traditional single-scale fusion schemes [35, 36] only compress and fuse a single-layer BEV feature map before passing it to the detection head, resulting in insufficient fusion. CoAlign [25] proposed a solution that performs layer-wise fusion on multi-scale features, addressing the issue of insufficient fusion. However, it compresses large-scale single-layer features extracted after Pillar Feature Extraction (PFE), which leads to higher bandwidth. Therefore, we propose a multi-scale compression and fusion approach, where separate autoencoders are designed for each scale of features to perform compression, followed by layer-wise fusion. This approach enables more thorough fusion while using less bandwidth. + +For message fusion, we use Max fusion, which has two main advantages: i) it is computationally simple and efficient, with the complexity increasing linearly with the number of agents; ii) max fusion selects the maximum value of features from multiple agents, achieving information complementarity. Considering the collaboration between Ego agent $i$ and collaborating agent $j$ , the process of message compression and fusion can be expressed as: + +$$ +Z _ {j \rightarrow i} ^ {\prime (l)} = f _ {e n c o d e} ^ {(l)} \left(Z _ {j \rightarrow i} ^ {(l)}\right) \in \mathbb {R} ^ {\frac {C _ {l}}{c _ {0}} \times H _ {l} \times W _ {l}} \tag {3} +$$ + +$$ +F _ {j \rightarrow i} ^ {(l)} = f _ {\text {d e c o d e}} ^ {(l)} \left(Z _ {j \rightarrow i} ^ {' (l)}\right) \in \mathbb {R} ^ {C _ {l} \times H _ {l} \times W _ {l}} \tag {4} +$$ + +$$ +F _ {j \rightarrow i} ^ {\prime (l)} = f _ {\text {t r a n s f o r m}} \left(F _ {j \rightarrow i} ^ {(l)}, \xi_ {j \rightarrow i}\right) \in \mathbb {R} ^ {C _ {l} \times H _ {l} \times W _ {l}} \tag {5} +$$ + +$$ +\tilde {F _ {i} ^ {(l)}} = \max \left(F _ {i} ^ {(l)}, \left\{F _ {j \rightarrow i} ^ {\prime (l)} \right\} _ {j \neq i}\right) \in \mathbb {R} ^ {C _ {l} \times H _ {l} \times W _ {l}} \tag {6} +$$ + +where $Z_{j\rightarrow i}^{(l)}\in \mathbb{R}^{C_l\times H_l\times W_l}$ represents the sparse features selected based on supply-demand relationships from the previous stage, and $C_l,H_l,W_l$ denote the channel, height, and width dimensions of the $l$ -th layer feature map. $f_{encode}^{(l)},f_{decode}^{(l)}$ represent the encoder and decoder of the $l$ -th layer autoencoder, $c_{0}$ is the compression ratio in the channel dimension, and $f_{transform}$ represents coordinate transformation. During communication, $\{Z_{j\rightarrow i}^{\prime (l)}\}_{l = 1,2,\ldots ,L}$ is first converted from float32 to float16, and then only the non-zero parts and corresponding coordinates are transmitted to save bandwidth. Upon receiving the message, agent $i$ decodes the feature dimensions, then aligns the features to its own coordinate system using the coordinate transformation matrix $\xi_{j\rightarrow i}$ to obtain $F_{j\rightarrow i}^{\prime (l)}$ , and finally fuses its own features with the collaborating agent's features using max fusion to obtain the fused features $F_{i}^{(l)}$ + +![](images/80aeadd77d05b25ee99ff4941272c79b32d1f4a1cb80af8bb919803063e6bced.jpg) +Figure 4. Our confidence-aware late fusion. It filters detection results based on confidence and suppresses suboptimal results from collaborative agents, improving overall detection accuracy. + +# 3.5. Confidence-Aware Late Fusion + +Existing collaborative perception methods focus on the singular intermediate collaboration architecture and neglect the advantages of late collaboration. As shown in Table 1, the experimental results indicate that late fusion can demonstrate acceptable perception accuracy with extremely low bandwidth on the OPV2V [36] and V2XSim [20] datasets. Therefore, late fusion can be used to compensate for the results of intermediate fusion, achieving a better accuracy-bandwidth trade-off. + +Naive late collaboration methods [20, 36] directly merge results from other agents and then apply NMS (Non-Maximum Suppression) to deduplicate and obtain final results. This approach can effectively improve recall for collaborative perception object detection; however, during the merging process, it may introduce some low-confidence false positives, and suboptimal detection results from collaborating agents may override the Ego agent's detection results, lowering precision and consequently decreasing final AP. The lower AP in the late fusion method in Table 1 on the DAIR-V2X [42] dataset is due to this issue. + +To address this problem, we propose a new confidence-aware late fusion method. As shown in Fig. 4, during late fusion, we first filter based on the confidence of the target boxes, discarding target boxes from other agents with confidence lower than $\varepsilon_{l}$ . Furthermore, considering that if both the Ego agent and other agents detect the same target, the detection result from the Ego agent, which has undergone intermediate fusion, is of higher quality. Therefore, we suppress the detection results from other agents to prevent their lower-quality results from degrading the Ego agent's better detection results. Specifically, before merging the detection results, we multiply the confidence of target boxes from other agents by a coefficient $\beta \in (0,1)$ . + +
SettingMethodOPV2V [36]V2XSim [20]DAIR-V2X [42]
AP@0.5↑AP@0.7↑BD↓AP@0.5↑AP@0.7↑BD↓AP@0.5↑AP@0.7↑BD↓
BasicNo Fusion79.78%67.16%0.0 Mbps70.31%58.55%0.0 Mbps66.51%55.46%0.0 Mbps
Early Fusion95.05%88.98%83.1 Mbps95.68%88.03%55.5 Mbps74.52%59.22%50.2 Mbps
Late Fusion94.70%87.18%0.2 Mbps86.88%78.13%0.1 Mbps67.86%50.47%0.2 Mbps
No LimitWhen2com [21]91.75%81.77%1,320.0 Mbps72.65%62.92%250.0 Mbps64.08%49.14%984.8 Mbps
FCooper [3]90.06%74.03%2,640.0 Mbps72.96%57.61%500.0 Mbps74.58%56.47%1,968.8 Mbps
AttFuse [36]94.31%82.03%2,640.0 Mbps78.06%64.84%500.0 Mbps73.80%56.86%1,968.8 Mbps
V2VNet [32]96.66%92.44%5,280.0 Mbps88.97%85.18%1,000.0 Mbps66.63%47.39%3,937.5 Mbps
DiscoNet [19]90.93%78.90%2,640.0 Mbps77.34%68.77%500.0 Mbps73.58%58.45%1,968.8 Mbps
V2XViT [35]95.87%89.88%2,640.0 Mbps89.01%80.26%500.0 Mbps76.68%57.57%1,920.0 Mbps
Where2comm [11]95.59%91.39%48.7 Mbps88.18%83.66%27.5 Mbps76.13%60.16%172.3 Mbps
CoAlign [25]96.63%92.63%2,640.0 Mbps88.87%85.23%500.0 Mbps78.06%63.09%1,968.8 Mbps
CoSDH96.83%92.99%13.4 Mbps89.23%86.31%1.1 Mbps76.75%63.85%7.1 Mbps
BD ≤ 6.75 MbpsWhen2com [21]79.97%50.88%5.2 Mbps62.02%43.37%4.0 Mbps62.37%44.01%3.8 Mbps
FCooper [3]90.37%72.92%5.2 Mbps76.52%61.83%4.0 Mbps67.71%47.65%3.8 Mbps
AttFuse [36]93.45%80.02%5.2 Mbps84.58%71.36%4.0 Mbps71.06%49.57%3.8 Mbps
V2VNet [32]95.71%87.16%5.2 Mbps85.66%73.72%4.0 Mbps66.49%45.61%3.8 Mbps
DiscoNet [19]90.00%76.72%5.2 Mbps78.04%68.18%4.0 Mbps70.83%53.40%3.8 Mbps
V2XViT [35]95.85%87.13%5.2 Mbps88.94%81.47%4.0 Mbps71.00%52.78%3.8 Mbps
Where2comm [11]94.91%89.86%5.4 Mbps87.51%82.12%4.7 Mbps74.98%59.51%5.3 Mbps
CoAlign [25]94.10%85.99%5.2 Mbps88.01%83.97%4.0 Mbps75.26%60.19%3.8 Mbps
CoSDH96.75%92.92%2.0 Mbps89.23%86.31%1.1 Mbps76.47%63.76%1.4 Mbps
+ +Table 1. Comparison of detection accuracy and bandwidth of different methods on OPV2V [36], V2XSim [20], and DAIR-V2X [42] datasets. "BD" represents the bandwidth required for each collaborative agent, assuming Ego agent collaborates with up to 4 agents and the detection frequency is $10\mathrm{Hz}$ . "BD $\leq$ 6.75 Mbps" is used to simulate real-world communication limits, assuming a total communication bandwidth of 27 Mbps, with each collaborating agent's bandwidth consumption limited to less than 6.75 Mbps. For intermediate collaboration methods without information selection, we provide a compressed version using autoencoders to meet the bandwidth constraints. For intermediate collaboration methods with information selection, the selection ratio is adjusted to meet the bandwidth constraints. + +# 4. Experiments + +# 4.1. Datasets and Experimental Settings + +Datasets. We evaluate the proposed CoSDH against other methods on three different collaborative perception datasets (OPV2V [36], V2XSim [20], and DAIR-V2X [42]) for LiDAR-based 3D object detection. The datasets include both simulated and real-world scenarios, and cover two types of collaboration: V2V (Vehicle to Vehicle) and V2I (Vehicle to Infrastructure). + +Evaluation Metrics We use the average precision (AP) with intersection-over-union (IoU) thresholds of 0.5 and 0.7 to evaluate the performance of different methods on 3D object detection. We assume the target detection frequency is $10\mathrm{Hz}$ and calculate the communication bandwidth based on the average data transmitted by each collaborative agent to the Ego agent, in order to evaluate the communication cost of different methods. Specifically, we consider the bandwidth limitations in real-world collaborative perception scenarios, setting the vehicle's communication rate to 27 Mbps [1, 24]. Considering the typical case where the Ego agent collaborates with up to 4 other agents [36], the bandwidth limit for each collaborative agent is $27/4 = 6.75$ Mbps. + +Implementation Our experiments are based on the OpenCOOD [36] framework, using PointPillar [17] as the encoder with a grid size of $(0.4m, 0.4m)$ , and a maximum of 32 points per Pillar. For our method, we set the number of intermediate feature layers to $L = 3$ , the demand threshold + +$\varepsilon_{a} = 4 / 32 = 0.125$ , and the supply threshold $\varepsilon_{c} = 0.01$ . For the OPV2V [36] and DAIR-V2X [42] datasets, the compression rate is $c_{0} = 16$ , and for the V2XSim [20] dataset, the compression rate is $c_{0} = 8$ due to its smaller perception range and lower inherent bandwidth requirement. The late fusion threshold is set to $\varepsilon_{l} = 0.3$ , with a suppression coefficient $\beta = 0.9$ for OPV2V and V2XSim dataset and $\beta = 0.8$ for DAIR-V2X dataset because of its difficulty. In late fusion, the dense prediction results before NMS are transmitted, as this reduces the computational burden on the collaborating agents. The experiments use the Adam [14] optimizer, with an initial learning rate set between 0.0001 and 0.002 based on the model's testing complexity to ensure proper training. The maximum number of collaborating agents is set to 5. The number of training epochs is set to 40 to ensure model convergence. Other experimental parameters are kept consistent with the OpenCOOD framework. All methods are trained on four NVIDIA GeForce RTX 3090 GPUs. + +# 4.2.Quantitative Evaluation + +Benchmark Comparison. Table 1 presents the collaborative 3D object detection accuracy and required bandwidth of the proposed CoSDH compared to previous methods across different datasets. Experimental results show that, with the default uncompressed settings, CoSDH achieves the highest accuracy on the OPV2V [36] and V2XSim [20] datasets while requiring only about 1/100 to 1/1000 of the bandwidth compared to other non-communication-efficient + +![](images/7db46af544b2102d3c9ffe902cb3baa9ea6fb8a5a22940a03b130004298bb20b.jpg) + +![](images/b4969047a0f11c5b7a790bee2ed710f7b5bc4d5af5baed8b60554f1aafa82f0d.jpg) + +![](images/db0a639c041739ebdbec2fa7a6dc8654654b5df37a6c1f12776c5742af6fa67e.jpg) + +![](images/1432b4733da76085279af19d24efa23ecf1fd4454e9f29ccd4d50ef70a197eab.jpg) +Figure 5. Comparison of the trade-off between detection accuracy and bandwidth of different methods on OPV2V [36], V2XSim [20] and DAIR-V2X [42] datasets, CoSDH achieves the best accuracy-bandwidth trade-off. The real-world limitation refers to the total bandwidth limit of 27 Mbps, which means that each collaborative agent does not exceed 6.75 Mbps. + +![](images/25960b1269cdf59a40c0f5507e04ed8b6584511eadae182b809599bb39368e1e.jpg) + +![](images/b31a4409efb9422a4d40e23bef779118459e9b3b5ac96a53e320fb7bd610d8fc.jpg) + +methods. Although the AP@0.5 of CoSDH on the DAIR-V2X [42] dataset is slightly lower than that of CoAlign [25], it achieves a higher AP@0.7 with only about 1/300 of the bandwidth. Both our CoSDH and Where2comm [11] are communication-efficient methods based on information selection, which enable dynamic accuracy-bandwidth trade-offs by adjusting the selection ratio. The table shows accuracy at a specific bandwidth, and a more detailed comparison of the accuracy-bandwidth curves is provided in the "Accuracy-Bandwidth Trade-Off Comparison" part. + +Furthermore, we simulate real-world communication rate limitations. Result shows that $\mathrm{COSDH}$ achieves improvements in AP@0.7 of $3.06\% / 2.34\% / 3.75\%$ on OPV2V/V2XSim/DAIR-V2X compared to the previous best methods, while using less bandwidth. When comparing the scenarios with and without bandwidth limitations on the OPV2V and DAIR-V2X datasets, we found that $\mathrm{COSDH}$ achieves less than a $0.3\%$ decrease in AP under conditions where bandwidth is reduced by $80\%$ to $85\%$ , exhibiting less accuracy degradation than Where2comm. Most other methods also show varying degrees of accuracy degradation under bandwidth constraints. Interestingly, under bandwidth limitations, some methods show an improvement in accuracy on certain datasets after using autoencoders to compress intermediate features. For example, FCooper [3] shows improved accuracy on the V2XSim dataset, which may be due to the model's initially poor per + +formance. The compressed version, with the added autoencoder, increases the model's parameter count, thereby enriching its expressive capability. + +Accuracy-Bandwidth Trade-Off Comparison. Fig. 5 shows the accuracy-bandwidth trade-off of the proposed CoSDH and previous advanced methods across different datasets. When the bandwidth grater than 1 Mbps, Where2comm [11] demonstrates a good enough accuracy-bandwidth trade-off, maintaining high detection accuracy as bandwidth decreases. However, as bandwidth further decreases, its accuracy quickly declines, even falling below that of late fusion method. CoSDH can leverage late fusion at low bandwidths to maintain high accuracy, achieving a better performance-bandwidth trade-off. Notably, on the DAIR-V2X dataset, the late fusion method performs worse than the no-fusion case because the naive late fusion method unselectively merges detection results from other agents, introducing a large number of low-quality detections. CoSDH uses a confidence-aware late fusion method to select higher-quality detection results, improving this situation and enhancing perception accuracy. We also observe that the AP@0.7 of CoSDH on the V2XSim dataset initially increases slightly as bandwidth decreases before subsequently declining, indicating that selecting key collaboration areas can help reduce interference from other regions to some extent, thereby slightly improving detection accuracy. + +![](images/c67ec94fd94c5dd281f6e942222661da0330f6dfe7c05ae61a60ca082af49f85.jpg) +CoAlign +Bandwidth: 3.85 Mbps + +![](images/e86582efb0df25313fb0e1dc702e3c55eefcce969eb5f89dd88f630f5c436ac9.jpg) +Where2comm +Bandwidth: 5.26 Mbps + +![](images/26ad5a127e4a1ab99a2ca6d7f0919770d3530f8e83914f19a8855de206c11b9c.jpg) +CoSDH (w/o late fusion) +Figure 6. Visualization of detection results under real-world communication rate limitations on the DAIR-V2X [42] dataset. Green represents ground truth box, and red represents predicted box. + +![](images/df3b39563c59aec3cbec23044cd0307074f1ed1168926dac4f523190071d4d4a.jpg) +CoSDH +Bandwidth: 0.29 Mbps + +# 4.3. Qualitative Evaluation + +Fig. 6 shows the visualization of detection results for our method compared to other methods under simulated real-world communication rate limitations on the DAIR-V2X dataset. A comparison of the detection results reveals that, while using less bandwidth, our method achieves the same recall rate as CoAlign [25] and Where2comm [11], with fewer false positive predictions and more accurate object localization, demonstrating that our method performs better under low-bandwidth conditions. Comparing the two images at bottom, it can be seen that under low bandwidth conditions, late fusion effectively compensates for the results of intermediate fusion, improving the recall rate of objects. + +# 4.4. Ablation Studies + +Table 2 presents the results of the ablation study on various modules of the proposed method using the OPV2V [36] dataset. The results show that after using the autoencoder for compression, the bandwidth is reduced by a factor of $c_{0}$ , and there is no significant loss in perception accuracy, even with a slight improvement in AP@0.7. Converting intermediate features to float16 can nearly halve the bandwidth without significant loss, demonstrating the effectiveness of our proposed message compression module. After performing information selection based on the supply mask, perception accuracy slightly decreases, but the bandwidth is reduced by about $95\%$ , which is a worthwhile trade-off. Further using the demand mask results in no significant decrease in accuracy but reduces bandwidth by about $10\%$ . This demonstrates that our supply-demand-aware information selection can reduce bandwidth while maintaining accuracy. + +After adding late fusion, accuracy improves at a relatively small bandwidth cost. However, since this phase uses more bandwidth, the improvement in accuracy is not signif + +
CompressionSelectionLate FusionAP@0.5↑AP@0.7↑BD↓
AutoencoderFP16SupplyDemand
96.62%92.62%1,155.00 Mbps
✓96.59%92.99%72.19 Mbps
✓✓96.59%92.99%36.09 Mbps
✓✓✓96.30%92.62%1.99 Mbps
✓✓✓✓96.31%92.60%1.82 Mbps
✓✓✓✓✓96.75%92.92%1.97 Mbps
+ +Table 2. Ablation study of the modules in $\mathrm{C}\circ \mathrm{{SDH}}$ on the OPV2V dataset. "Autoencoder" refers to the use of autoencoders to compress features along the channel dimension, with a compression ratio of ${c}_{0} = {16}$ . "FP16" refers to converting features from float32 to float16 for transmission. "Supply" and "Demand" refer to using supply and demand masks for information selection, and "Late Fusion" refers to applying confidence-aware late fusion. + +
Late Fusionεc0.010.020.030.050.07
AP@0.5↑96.59%96.31%96.19%95.18%93.35%
AP@0.7↑92.95%92.60%92.27%90.67%87.60%
BD↓13.26 Mbps1.82 Mbps0.54 Mbps0.12 Mbps0.05 Mbps
✓AP@0.5↑96.83%96.75%96.72%96.54%96.41%
AP@0.7↑92.99%92.92%92.73%92.23%91.68%
BD↓13.40 Mbps1.97 Mbps0.68 Mbps0.26 Mbps0.19 Mbps
+ +Table 3. Ablation study of confidence-aware late fusion under different bandwidths on the OPV2V [36] dataset. The table shows the impact of late fusion on accuracy and bandwidth with different values of $\varepsilon_{c}$ . + +cant. Table 3 shows the accuracy improvement due to late collaboration under different bandwidth conditions. It can be observed that late collaboration provides more accuracy improvements under lower bandwidth conditions. For example, when the bandwidth used for intermediate fusion is 0.05 Mbps, late collaboration with a bandwidth cost of 0.14 Mbps leads to a $3\%$ improvement in AP@0.5 and a $4\%$ improvement in AP@0.7. This is one of the key reasons why our method achieves significantly higher detection accuracy under lower bandwidth compared to Where2comm [11]. + +# 5. Conclusion + +In this paper, we propose CoSDH, a novel communication-efficient collaborative perception framework for 3D object detection. By finely modeling the supply-demand relationship between agents, it selects key and sparse regions for collaboration. Additionally, we innovatively incorporate confidence-aware late fusion on top of intermediate collaboration to form an intermediate-late hybrid collaborative perception method. Experiments on multiple datasets show that our method offers a better accuracy-bandwidth trade-off, demonstrating outstanding accuracy under bandwidth constraints close to real-world communication limits, making it highly valuable for practical applications. + +# Acknowledgment + +This work is supported by the National Key Research and Development Plan (2024YFB3309302). + +# References + +[1] Fabio Arena and Giovanni Pau. An overview of vehicular communications. Future internet, 11(2):27, 2019. 1, 6 +[2] Eduardo Arnold, Mehrdad Dianati, Robert de Temple, and Saber Fallah. Cooperative perception for 3d object detection in driving scenarios using infrastructure sensors. IEEE Transactions on Intelligent Transportation Systems, 23(3): 1852-1864, 2020. 3 +[3] Qi Chen, Xu Ma, Sihai Tang, Jingda Guo, Qing Yang, and Song Fu. F-cooper: Feature based cooperative perception for autonomous vehicle edge computing system using 3d point clouds. In Proceedings of the 4th ACM/IEEE Symposium on Edge Computing, pages 88-100, 2019. 1, 3, 6, 7, 2 +[4] Qi Chen, Sihai Tang, Qing Yang, and Song Fu. Cooper: Cooperative perception for connected autonomous vehicles based on 3d point clouds. In 2019 IEEE 39th International Conference on Distributed Computing Systems (ICDCS), pages 514-524. IEEE, 2019. 3 +[5] Yilun Chen, Shu Liu, Xiaoyong Shen, and Jiaya Jia. Fast point r-cnn. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9775-9784, 2019. 2 +[6] Yilun Chen, Shu Liu, Xiaoyong Shen, and Jiaya Jia. Dsgn: Deep stereo geometry network for 3d object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12536-12545, 2020. 2 +[7] Jiajun Deng, Shaoshuai Shi, Peiwei Li, Wengang Zhou, Yanyong Zhang, and Houqiang Li. Voxel r-cnn: Towards high performance voxel-based 3d object detection. In Proceedings of the AAAI conference on artificial intelligence, pages 1201-1209, 2021. 2 +[8] Alexey Dosovitskiy, German Ros, Felipe Codevilla, Antonio Lopez, and Vladlen Koltun. Carla: An open urban driving simulator. In Conference on robot learning, pages 1-16. PMLR, 2017. 1 +[9] Sebastian Gräfling, Petri Mähönen, and Janne Riihijärvi. Performance evaluation of the IEEE 1609 wave and the IEEE 802.11 p for vehicular communications. In 2010 second international conference on ubiquitous and future networks (ICUFN), pages 344-348. IEEE, 2010. 2 +[10] Jingda Guo, Dominic Carrillo, Sihai Tang, Qi Chen, Qing Yang, Song Fu, Xi Wang, Nannan Wang, and Paparao Palacharla. Coff: Cooperative spatial feature fusion for 3-d object detection on autonomous vehicles. IEEE Internet of Things Journal, 8(14):11078-11087, 2021. 3 +[11] Yue Hu, Shaoheng Fang, Zixing Lei, Yiqi Zhong, and Siheng Chen. Where2comm: Communication-efficient collaborative perception via spatial confidence maps. Advances in neural information processing systems, 35:4874-4886, 2022. 1, 2, 3, 4, 6, 7, 8 +[12] Yue Hu, Juntong Peng, Sifei Liu, Junhao Ge, Si Liu, and Si-heng Chen. Communication-efficient collaborative perception via information filling with codebook. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15481-15490, 2024. 2 +[13] Zheng Jiang, Jinqing Zhang, Yanan Zhang, Qingjie Liu, Zhenghui Hu, Baohui Wang, and Yunhong Wang. Fsd-bev: + +Foreground self-distillation for multi-view 3d object detection. In European Conference on Computer Vision, pages 110-126. Springer, 2024. 2 + +[14] Diederik P Kingma. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6 +[15] Daniel Krajzewicz, Jakob Erdmann, Michael Behrisch, and Laura Bieker. Recent development and applications of sumosimulation of urban mobility. International journal on advances in systems and measurements, 5(3&4), 2012. 1 +[16] Jason Ku, Alex D Pon, and Steven L Waslander. Monocular 3d object detection leveraging accurate proposals and shape reconstruction. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11867-11876, 2019. 2 +[17] Alex H Lang, Sourabh Vora, Holger Caesar, Lubing Zhou, Jiong Yang, and Oscar Beijbom. Pointpillars: Fast encoders for object detection from point clouds. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12697-12705, 2019. 3, 6 +[18] Peiliang Li, Xiaozhi Chen, and Shaojie Shen. Stereo r-cnn based 3d object detection for autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7644-7652, 2019. 2 +[19] Yiming Li, Shunli Ren, Pengxiang Wu, Siheng Chen, Chen Feng, and Wenjun Zhang. Learning distilled collaboration graph for multi-agent perception. Advances in Neural Information Processing Systems, 34:29541-29552, 2021. 3, 6 +[20] Yiming Li, Dekun Ma, Ziyan An, Zixun Wang, Yiqi Zhong, Siheng Chen, and Chen Feng. V2x-sim: Multi-agent collaborative perception dataset and benchmark for autonomous driving. IEEE Robotics and Automation Letters, 7(4):10914-10921, 2022. 1, 2, 3, 5, 6, 7 +[21] Yen-Cheng Liu, Junjiao Tian, Nathaniel Glaser, and Zsolt Kira. When2com: Multi-agent perception via communication graph grouping. In Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition, pages 4106-4115, 2020. 3, 6 +[22] Yen-Cheng Liu, Junjiao Tian, Chih-Yao Ma, Nathan Glaser, Chia-Wen Kuo, and Zsolt Kira. Who2com: Collaborative perception via learnable handshake communication. In 2020 IEEE International Conference on Robotics and Automation (ICRA), pages 6876-6883. IEEE, 2020. 3 +[23] Zhijian Liu, Haotian Tang, Alexander Amini, Xinyu Yang, Huizi Mao, Daniela L Rus, and Song Han. Bevfusion: Multi-task multi-sensor fusion with unified bird's-eye view representation. In 2023 IEEE international conference on robotics and automation (ICRA), pages 2774-2781. IEEE, 2023. 3 +[24] Jose Manuel Lozano Dominguez and Tomas Jesus Mateo Sanguino. Review on v2x, i2x, and p2x communications and their applications: a comprehensive analysis over time. Sensors, 19(12):2756, 2019. 1, 6 +[25] Yifan Lu, Quanhao Li, Baoan Liu, Mehrdad Dianati, Chen Feng, Siheng Chen, and Yanfeng Wang. Robust collaborative 3d object detection in presence of pose errors. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 4812-4818. IEEE, 2023. 5, 6, 7, 8, 1, 2 +[26] Jiageng Mao, Shaoshuai Shi, Xiaogang Wang, and Hongsheng Li. 3d object detection for autonomous driving: A + +comprehensive survey. International Journal of Computer Vision, 131(8):1909-1963, 2023. 2 +[27] Xuran Pan, Zhuofan Xia, Shiji Song, Li Erran Li, and Gao Huang. 3d object detection with pointformer. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7463-7472, 2021. 2 +[28] Shaoshuai Shi, Xiaogang Wang, and Hongsheng Li. Pointrcnn: 3d object proposal generation and detection from point cloud. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 770-779, 2019. 2 +[29] Shaoshuai Shi, Chaoxu Guo, Li Jiang, Zhe Wang, Jianping Shi, Xiaogang Wang, and Hongsheng Li. Pv-rcnn: Pointvoxel feature set abstraction for 3d object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10529–10538, 2020. 2 +[30] Shaoshuai Shi, Zhe Wang, Jianping Shi, Xiaogang Wang, and Hongsheng Li. From points to parts: 3d object detection from point cloud with part-aware and part-aggregation network. IEEE transactions on pattern analysis and machine intelligence, 43(8):2647-2664, 2020. 2 +[31] Rui Song, Chenwei Liang, Hu Cao, Zhiran Yan, Walter Zimmer, Markus Gross, Andreas Festag, and Alois Knoll. Collaborative semantic occupancy prediction with hybrid feature fusion in connected automated vehicles. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17996-18006, 2024. 1 +[32] Tsun-Hsuan Wang, Sivabalan Manivasagam, Ming Liang, Bin Yang, Wenyuan Zeng, and Raquel Urtasun. V2vnet: Vehicle-to-vehicle communication for joint perception and prediction. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part II 16, pages 605–621. Springer, 2020. 3, 6 +[33] Yan Wang, Wei-Lun Chao, Divyansh Garg, Bharath Hariharan, Mark Campbell, and Kilian Q Weinberger. Pseudolidar from visual depth estimation: Bridging the gap in 3d object detection for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8445-8453, 2019. 2 +[34] Runsheng Xu, Yi Guo, Xu Han, Xin Xia, Hao Xiang, and Jiaqi Ma. Opencda: an open cooperative driving automation framework integrated with co-simulation. In 2021 IEEE International Intelligent Transportation Systems Conference (ITSC), pages 1155-1162. IEEE, 2021. 1 +[35] Runsheng Xu, Hao Xiang, Zhengzhong Tu, Xin Xia, Ming-Hsuan Yang, and Jiaqi Ma. V2x-vit: Vehicle-to-everything cooperative perception with vision transformer. In European conference on computer vision, pages 107-124. Springer, 2022. 1, 3, 5, 6, 2 +[36] Runsheng Xu, Hao Xiang, Xin Xia, Xu Han, Jinlong Li, and Jiaqi Ma. Opv2v: An open benchmark dataset and fusion pipeline for perception with vehicle-to-vehicle communication. In 2022 International Conference on Robotics and Automation (ICRA), pages 2583–2589. IEEE, 2022. 1, 2, 3, 5, 6, 7, 8 +[37] Runsheng Xu, Zhengzhong Tu, Hao Xiang, Wei Shao, Bolei Zhou, and Jiaqi Ma. Cobevt: Cooperative bird's eye view + +semantic segmentation with sparse transformers. In Conference on Robot Learning, pages 989-1000. PMLR, 2023. 1, 3 +[38] Runsheng Xu, Xin Xia, Jinlong Li, Hanzhao Li, Shuo Zhang, Zhengzhong Tu, Zonglin Meng, Hao Xiang, Xiaoyu Dong, Rui Song, et al. V2v4real: A real-world large-scale dataset for vehicle-to-vehicle cooperative perception. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13712–13722, 2023. 1, 2 +[39] Yan Yan, Yuxing Mao, and Bo Li. Second: Sparsely embedded convolutional detection. Sensors, 18(10):3337, 2018. 2 +[40] Dingkang Yang, Kun Yang, Yuzheng Wang, Jing Liu, Zhi Xu, Rongbin Yin, Peng Zhai, and Lihua Zhang. How2comm: Communication-efficient and collaboration-pragmatic multi-agent perception. Advances in Neural Information Processing Systems, 36, 2024. 4 +[41] Junbo Yin, Jianbing Shen, Runnan Chen, Wei Li, Ruigang Yang, Pascal Frossard, and Wenguan Wang. Is-fusion: Instance-scene collaborative fusion for multimodal 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14905-14915, 2024. 3 +[42] Haibao Yu, Yizhen Luo, Mao Shu, Yiyi Huo, Zebang Yang, Yifeng Shi, Zhenglong Guo, Hanyu Li, Xing Hu, Jirui Yuan, et al. Dair-v2x: A large-scale dataset for vehicleinfrastructure cooperative 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21361-21370, 2022. 2, 3, 5, 6, 7, 8, 1 +[43] Haibao Yu, Wenxian Yang, Hongzhi Ruan, Zhenwei Yang, Yingjuan Tang, Xu Gao, Xin Hao, Yifeng Shi, Yifeng Pan, Ning Sun, et al. V2x-seq: A large-scale sequential dataset for vehicle-infrastructure cooperative perception and forecasting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5486-5495, 2023. 1 +[44] Jinqing Zhang, Yanan Zhang, Qingjie Liu, and Yunhong Wang. Sa-bev: Generating semantic-aware bird's-eye-view feature for multi-view 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3348-3357, 2023. 2 +[45] Jinqing Zhang, Yanan Zhang, Yunlong Qi, Zehua Fu, Qingjie Liu, and Yunhong Wang. Geobev: Learning geometric bev representation for multi-view 3d object detection. arXiv preprint arXiv:2409.01816, 2024. 2 +[46] Yanan Zhang, Di Huang, and Yunhong Wang. Pc-rgnn: Point cloud completion and graph neural network for 3d object detection. In Proceedings of the AAAI conference on artificial intelligence, pages 3430-3437, 2021. 2 +[47] Yanan Zhang, Jiaxin Chen, and Di Huang. Cat-det: Contrastively augmented transformer for multi-modal 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 908-917, 2022. 3 +[48] Chao Zhou, Yanan Zhang, Jiaxin Chen, and Di Huang. Octr: Octtree-based transformer for 3d object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5166-5175, 2023. 2 \ No newline at end of file diff --git a/2025/CoSDH_ Communication-Efficient Collaborative Perception via Supply-Demand Awareness and Intermediate-Late Hybridization/images.zip b/2025/CoSDH_ Communication-Efficient Collaborative Perception via Supply-Demand Awareness and Intermediate-Late Hybridization/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..2597166bb227e54d5e526ee65b4fb0133fc4b787 --- /dev/null +++ b/2025/CoSDH_ Communication-Efficient Collaborative Perception via Supply-Demand Awareness and Intermediate-Late Hybridization/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:30ef36630662ddd1f8c8f6942c5bcf2f4f2599d075e46edd3f7ff8d675a10cf5 +size 721778 diff --git a/2025/CoSDH_ Communication-Efficient Collaborative Perception via Supply-Demand Awareness and Intermediate-Late Hybridization/layout.json b/2025/CoSDH_ Communication-Efficient Collaborative Perception via Supply-Demand Awareness and Intermediate-Late Hybridization/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..507bd1e966071624a07b8ed14ad9be824a5e451d --- /dev/null +++ b/2025/CoSDH_ Communication-Efficient Collaborative Perception via Supply-Demand Awareness and Intermediate-Late Hybridization/layout.json @@ -0,0 +1,8535 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 58, + 102, + 553, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 102, + 553, + 138 + ], + "spans": [ + { + "bbox": [ + 58, + 102, + 553, + 138 + ], + "type": "text", + "content": "CoSDH: Communication-Efficient Collaborative Perception via Supply-Demand Awareness and Intermediate-Late Hybridization" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 196, + 159, + 440, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 196, + 159, + 440, + 175 + ], + "spans": [ + { + "bbox": [ + 196, + 159, + 440, + 175 + ], + "type": "text", + "content": "Junhao Xu" + }, + { + "bbox": [ + 196, + 159, + 440, + 175 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 196, + 159, + 440, + 175 + ], + "type": "text", + "content": ", Yanan Zhang" + }, + { + "bbox": [ + 196, + 159, + 440, + 175 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 196, + 159, + 440, + 175 + ], + "type": "text", + "content": ", Zhi Cai" + }, + { + "bbox": [ + 196, + 159, + 440, + 175 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 196, + 159, + 440, + 175 + ], + "type": "text", + "content": ", Di Huang" + }, + { + "bbox": [ + 196, + 159, + 440, + 175 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 175, + 576, + 205 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 61, + 175, + 576, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 175, + 576, + 190 + ], + "spans": [ + { + "bbox": [ + 61, + 175, + 576, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 61, + 175, + 576, + 190 + ], + "type": "text", + "content": "State Key Laboratory of Complex and Critical Software Environment, Beihang University, Beijing, China" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 190, + 575, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 190, + 575, + 205 + ], + "spans": [ + { + "bbox": [ + 61, + 190, + 575, + 205 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 61, + 190, + 575, + 205 + ], + "type": "text", + "content": "School of Computer Science and Information Engineering, Hefei University of Technology, Hefei, China" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 112, + 207, + 523, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 112, + 207, + 523, + 219 + ], + "spans": [ + { + "bbox": [ + 112, + 207, + 523, + 219 + ], + "type": "text", + "content": "junhaoxu@buaa.edu.cn, yananzhang@hfut.edu.cn, {caizhi97, dhuang}@buaa.edu.cn" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 151, + 246, + 200, + 258 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 246, + 200, + 258 + ], + "spans": [ + { + "bbox": [ + 151, + 246, + 200, + 258 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 54, + 272, + 297, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 272, + 297, + 571 + ], + "spans": [ + { + "bbox": [ + 54, + 272, + 297, + 571 + ], + "type": "text", + "content": "Multi-agent collaborative perception enhances perceptual capabilities by utilizing information from multiple agents and is considered a fundamental solution to the problem of weak single-vehicle perception in autonomous driving. However, existing collaborative perception methods face a dilemma between communication efficiency and perception accuracy. To address this issue, we propose a novel communication-efficient collaborative perception framework based on supply-demand awareness and intermediate-late hybridization, dubbed as CoSDH. By modeling the supply-demand relationship between agents, the framework refines the selection of collaboration regions, reducing unnecessary communication cost while maintaining accuracy. In addition, we innovatively introduce the intermediate-late hybrid collaboration mode, where late-stage collaboration compensates for the performance degradation in collaborative perception under low communication bandwidth. Extensive experiments on multiple datasets, including both simulated and real-world scenarios, demonstrate that CoSDH achieves state-of-the-art detection accuracy and optimal bandwidth trade-offs, delivering superior detection precision under real communication bandwidths, thus proving its effectiveness and practical applicability. The code will be released at https://github.com/Xu2729/CoSDH." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 591, + 135, + 604 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 591, + 135, + 604 + ], + "spans": [ + { + "bbox": [ + 55, + 591, + 135, + 604 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 608, + 295, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 608, + 295, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 608, + 295, + 693 + ], + "type": "text", + "content": "Collaborative perception allows multiple agents to exchange complementary perception information. It fundamentally addresses the issues of limited perception range, sensor blind spots, and occlusion from obstacles inherent in single-agent perception, thereby improving both the range and accuracy of perception. Recent studies have demonstrated that collaborative perception can be applied to var" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 247, + 555, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 247, + 555, + 319 + ], + "spans": [ + { + "bbox": [ + 313, + 247, + 555, + 319 + ], + "type": "text", + "content": "ious autonomous driving tasks, including 3D object detection [3, 35, 36], semantic segmentation [20, 37], 3D occupancy prediction [31], and trajectory prediction [43], exhibiting enhanced performance and improved robustness to occlusions, thus making it a crucial approach for advancing autonomous driving systems." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 319, + 556, + 546 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 319, + 556, + 546 + ], + "spans": [ + { + "bbox": [ + 313, + 319, + 556, + 546 + ], + "type": "text", + "content": "Communication efficiency is a key issue in collaborative perception. An early approach to improving communication efficiency was to use autoencoders to compress the intermediate features that need to be transmitted [36]. Later, communication-efficient methods such as Where2comm [11] were proposed, which select important and sparse foreground regions for collaboration. However, these methods still lack sufficient precision in area selection and require more bandwidth than the real-world constraints1. As shown in Fig. 1a, when the bandwidth is limited to 27 Mbps, Where2comm [11] experiences a decrease of about " + }, + { + "bbox": [ + 313, + 319, + 556, + 546 + ], + "type": "inline_equation", + "content": "1.5\\%" + }, + { + "bbox": [ + 313, + 319, + 556, + 546 + ], + "type": "text", + "content": " in average precision (AP@0.7) on the OPV2V [36] dataset with an intersection-over-union (IoU) threshold of 0.7. When the bandwidth is further limited to 5 Mbps and 1 Mbps, AP@0.7 drops by approximately " + }, + { + "bbox": [ + 313, + 319, + 556, + 546 + ], + "type": "inline_equation", + "content": "7.5\\%" + }, + { + "bbox": [ + 313, + 319, + 556, + 546 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 319, + 556, + 546 + ], + "type": "inline_equation", + "content": "14.5\\%" + }, + { + "bbox": [ + 313, + 319, + 556, + 546 + ], + "type": "text", + "content": ", respectively, even falling below the accuracy of late fusion methods at the result level. This makes it difficult for existing collaborative perception methods to meet the demands of practical applications." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 547, + 557, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 547, + 557, + 668 + ], + "spans": [ + { + "bbox": [ + 313, + 547, + 557, + 668 + ], + "type": "text", + "content": "In this paper, we first analyze the problem of selecting collaboration areas and refine this selection based on the supply-demand relationship between agents. As shown in Fig. 1b, for Ego, low foreground confidence areas can be divided into three types: Area 1 (low confidence due to absence of foreground), Area 2 (low confidence due to occlusion), and Area 3 (low confidence due to sparsity of the point cloud). Although all three areas appear as background from the Ego's perspective, Area 1 is well-observed and does not require collaboration, while Areas 2 and 3," + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 313, + 673, + 555, + 715 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 673, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 673, + 555, + 715 + ], + "type": "text", + "content": "To the best of our knowledge, the common V2X communication approach uses IEEE 802.11p-based DSRC (Dedicated Short-Range Communications) technology [1, 24], which provides an information transmission rate of approximately 27 Mbps." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 66, + 702, + 216, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 702, + 216, + 713 + ], + "spans": [ + { + "bbox": [ + 66, + 702, + 216, + 713 + ], + "type": "text", + "content": "*Equal Contribution. †Corresponding Author." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6834" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 70, + 239, + 215 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 239, + 215 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 239, + 215 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 239, + 215 + ], + "type": "image", + "image_path": "06fba8aec52e2e0d60b88c5175542a2224d5edff453a88d56a00d9dd8f457a17.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 217, + 240, + 255 + ], + "lines": [ + { + "bbox": [ + 55, + 217, + 240, + 255 + ], + "spans": [ + { + "bbox": [ + 55, + 217, + 240, + 255 + ], + "type": "text", + "content": "(a) The 3D detection accuracy of Where2comm [11] and CoSDH on OPV2V dataset [36] with different bandwidth limit. Assume the number of collaborative agents is 4 and detection frequency is " + }, + { + "bbox": [ + 55, + 217, + 240, + 255 + ], + "type": "inline_equation", + "content": "10\\mathrm{Hz}" + }, + { + "bbox": [ + 55, + 217, + 240, + 255 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 146, + 260, + 462, + 272 + ], + "lines": [ + { + "bbox": [ + 146, + 260, + 462, + 272 + ], + "spans": [ + { + "bbox": [ + 146, + 260, + 462, + 272 + ], + "type": "text", + "content": "Figure 1. Issues of existing communication-efficient collaborative perception methods." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 247, + 75, + 553, + 215 + ], + "blocks": [ + { + "bbox": [ + 247, + 75, + 553, + 215 + ], + "lines": [ + { + "bbox": [ + 247, + 75, + 553, + 215 + ], + "spans": [ + { + "bbox": [ + 247, + 75, + 553, + 215 + ], + "type": "image", + "image_path": "cfafa97c36fe5a934c2101d62e17fb153ddc49d8ce0238f4393af0acde4d72be.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 246, + 217, + 555, + 256 + ], + "lines": [ + { + "bbox": [ + 246, + 217, + 555, + 256 + ], + "spans": [ + { + "bbox": [ + 246, + 217, + 555, + 256 + ], + "type": "text", + "content": "(b) Point cloud map and classification of areas with low foreground confidence. Although Area 1 has low confidence due to the absence of foreground, it allows for good observation without collaboration. In contrast, Areas 2 and 3 exhibit poor observation and require collaboration, with Area 2 having low confidence due to occlusion and Area 3 due to sparsity of the point cloud." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 280, + 295, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 280, + 295, + 411 + ], + "spans": [ + { + "bbox": [ + 54, + 280, + 295, + 411 + ], + "type": "text", + "content": "with poorer observation, need collaboration. Additionally, to address the issue of significant accuracy degradation under real bandwidth constraints, we propose that compensating intermediate collaboration results with late fusion is a feasible solution. As shown in Fig. 1a, late fusion achieves much higher accuracy than Where2comm [11] even with a bandwidth of approximately 0.7 Mbps, compared to Where2comm's accuracy with bandwidth less than 5 Mbps. So the use of intermediate-late hybrid collaboration can greatly improve the accuracy lower bound under low-bandwidth conditions." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 413, + 296, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 413, + 296, + 663 + ], + "spans": [ + { + "bbox": [ + 56, + 413, + 296, + 663 + ], + "type": "text", + "content": "Based on these insights, we propose CoSDH, a novel communication-efficient collaborative perception framework for 3D object detection. As shown in Fig. 2, CoSDH consists of three key components: i) Supply-demand-aware information selection, which chooses sparse yet crucial regions for collaboration; ii) Intermediate feature transmission and fusion, which transmits and effectively aggregates information from multiple agents in a communication-efficient manner; iii) Confidence-aware late fusion, which compensates for the intermediate fusion results at a minimal communication cost to improve accuracy. To evaluate CoSDH, we conducted extensive experiments on the simulation datasets OPV2V [36], V2XSim [20], and the real-world dataset DAIR-V2X [42]. The experimental results show that i) In the baseline case, CoSDH uses less bandwidth while achieving satisfactory accuracy, demonstrating a better accuracy-bandwidth trade-off; ii) Under the simulated real-world communication condition with a total bandwidth limit of 27 Mbps, CoSDH experiences less accuracy degradation and outperforms other methods while using less bandwidth." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 66, + 665, + 251, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 665, + 251, + 676 + ], + "spans": [ + { + "bbox": [ + 66, + 665, + 251, + 676 + ], + "type": "text", + "content": "In summary, our contributions are as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "content": "- We present CoSDH, an innovative communication-efficient collaborative perception framework with better accuracy-bandwidth trade-offs for 3D object detection." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 281, + 554, + 436 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 313, + 281, + 553, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 281, + 553, + 316 + ], + "spans": [ + { + "bbox": [ + 313, + 281, + 553, + 316 + ], + "type": "text", + "content": "- We propose a novel supply-demand-aware information selection module, further refining the collaboration area selection to achieve more efficient communication." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 316, + 553, + 364 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 316, + 553, + 364 + ], + "spans": [ + { + "bbox": [ + 313, + 316, + 553, + 364 + ], + "type": "text", + "content": "- We design a novel intermediate-late hybrid collaborative perception paradigm, where confidence-aware late fusion compensates for the intermediate fusion results to maintain high accuracy under low-bandwidth conditions." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 365, + 554, + 436 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 365, + 554, + 436 + ], + "spans": [ + { + "bbox": [ + 313, + 365, + 554, + 436 + ], + "type": "text", + "content": "- We conduct extensive experiments across multiple datasets with bandwidth constraints set closer to real-world conditions, and CoSDH achieves state-of-the-art detection accuracy along with optimal bandwidth trade-offs, demonstrating its feasibility in real-world collaborative perception scenarios." + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 441, + 400, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 441, + 400, + 453 + ], + "spans": [ + { + "bbox": [ + 313, + 441, + 400, + 453 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 456, + 432, + 469 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 456, + 432, + 469 + ], + "spans": [ + { + "bbox": [ + 313, + 456, + 432, + 469 + ], + "type": "text", + "content": "2.1. 3D Object Detection" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 474, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 474, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 474, + 555, + 713 + ], + "type": "text", + "content": "3D object detection, a key technology in autonomous driving, identifies and localizes objects in 3D scenes using environmental data from onboard sensors and can be categorized into image-based, point-cloud-based, and multimodal-fusion-based methods [26]. Image-based methods can be further classified into monocular [16, 33], stereo [6, 18], and multi-view [13, 44, 45] approaches based on the number of onboard cameras. These methods leverage the rich color and texture information of image, along with its dense data representation, to achieve cost-effective 3D perception for autonomous driving. However, image-based methods are inherently limited by the lack of depth information, which restricts their performance. In contrast, point-cloud-based methods, which adopt voxel-based [7, 39, 48], raw point cloud [27, 28, 46], or point-voxel hybrid [5, 29, 30] representations, can fully leverage the precise 3D geometric information provided by LiDAR, significantly improving perception capabilities. Considering that point-cloud-based methods suffer from sparse characteristic and lack rich semantic information, multimodal-fusion-based meth" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6835" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "type": "text", + "content": "ods [23, 41, 47] further enhance performance by leveraging the complementary advantages of different modalities." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 99, + 296, + 196 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 99, + 296, + 196 + ], + "spans": [ + { + "bbox": [ + 55, + 99, + 296, + 196 + ], + "type": "text", + "content": "However, these single-agent-based 3D object detection methods are limited by sensor range and susceptible to occlusion, making them unable to detect objects that are further away or completely occluded. This paper focuses on collaborative-perception-based 3D object detection methods, which improve detection performance by supplementing the limitations of single-agent detection with information from other agents." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 212, + 195, + 224 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 212, + 195, + 224 + ], + "spans": [ + { + "bbox": [ + 55, + 212, + 195, + 224 + ], + "type": "text", + "content": "2.2. Collaborative Perception" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 232, + 296, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 232, + 296, + 484 + ], + "spans": [ + { + "bbox": [ + 55, + 232, + 296, + 484 + ], + "type": "text", + "content": "Collaborative perception can be categorized into early collaboration, intermediate collaboration, and late collaboration based on the collaboration timing. Early collaboration [2, 4, 42] shares raw perception data, providing good perception accuracy but with high bandwidth. Late collaboration [20, 36] shares perception results, significantly reducing bandwidth, but leads to a decline in perception accuracy. Intermediate collaboration operates at the feature level and can achieve a better trade-off between accuracy and bandwidth by adjusting the intermediate features transmitted [3, 10, 11, 19, 21, 22, 32, 35-37], which is why it has been widely studied. Some of this research has focused on improving perception accuracy. FCooper [3] and CoFF [10] used manual modeling to fuse multi-agent features. Who2com [22] and When2com [21] perform selective communication and use attention-based fusion. V2VNet [32] and DiscoNet [19] employ communication graph-based fusion methods. However, these methods typically transmit complete, uncompressed intermediate BEV features, leading to enormous bandwidth requirements, which makes them challenging to apply in practice." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 486, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 486, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 486, + 296, + 714 + ], + "type": "text", + "content": "To address the large bandwidth demand in collaborative perception, AttFuse [36] was the first to use autoencoders to compress intermediate features along the channel dimension, which was later adopted by V2X-ViT [35], CoBEVT [37] and others. However, this method leads to significant accuracy degradation at high compression rates. Where2comm [11] reduces bandwidth requirements by selecting sparse but important foreground regions for collaboration while maintaining perception accuracy. However, as bandwidth is further limited, perception accuracy still rapidly degrades, potentially even falling below the accuracy of simple late collaboration methods. To address this issue, this paper proposes a hybrid collaborative method based on both intermediate and late collaboration, which efficiently compensates for the intermediate collaborative results using late collaboration under bandwidth constraints. Furthermore, we also adopt and improve methods based on auto-encoders and information selection to save bandwidth while maintaining accuracy." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 314, + 72, + 370, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 370, + 83 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 370, + 83 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 87, + 425, + 98 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 87, + 425, + 98 + ], + "spans": [ + { + "bbox": [ + 314, + 87, + 425, + 98 + ], + "type": "text", + "content": "3.1. Problem Definition" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 104, + 555, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 104, + 555, + 224 + ], + "spans": [ + { + "bbox": [ + 313, + 104, + 555, + 224 + ], + "type": "text", + "content": "In this paper, we consider the problem of collaborative perception with " + }, + { + "bbox": [ + 313, + 104, + 555, + 224 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 104, + 555, + 224 + ], + "type": "text", + "content": " agents. Let " + }, + { + "bbox": [ + 313, + 104, + 555, + 224 + ], + "type": "inline_equation", + "content": "X_{i}" + }, + { + "bbox": [ + 313, + 104, + 555, + 224 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 104, + 555, + 224 + ], + "type": "inline_equation", + "content": "y_{i}" + }, + { + "bbox": [ + 313, + 104, + 555, + 224 + ], + "type": "text", + "content": " represent the raw observation and the corresponding ground truth supervision of the " + }, + { + "bbox": [ + 313, + 104, + 555, + 224 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 104, + 555, + 224 + ], + "type": "text", + "content": "-th agent, respectively, and let " + }, + { + "bbox": [ + 313, + 104, + 555, + 224 + ], + "type": "inline_equation", + "content": "P_{j \\rightarrow i}" + }, + { + "bbox": [ + 313, + 104, + 555, + 224 + ], + "type": "text", + "content": " be the message sent from agent " + }, + { + "bbox": [ + 313, + 104, + 555, + 224 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 104, + 555, + 224 + ], + "type": "text", + "content": " to agent " + }, + { + "bbox": [ + 313, + 104, + 555, + 224 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 104, + 555, + 224 + ], + "type": "text", + "content": ". In the collaborative perception, agent " + }, + { + "bbox": [ + 313, + 104, + 555, + 224 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 104, + 555, + 224 + ], + "type": "text", + "content": " aggregates its own observations and the messages " + }, + { + "bbox": [ + 313, + 104, + 555, + 224 + ], + "type": "inline_equation", + "content": "\\{P_{j \\rightarrow i}\\}_{j=1}^{N}" + }, + { + "bbox": [ + 313, + 104, + 555, + 224 + ], + "type": "text", + "content": " sent from other agents to perform 3D object detection task. Our goal is to maximize the collaborative perception 3D object detection accuracy while ensuring that each agent has a communication budget " + }, + { + "bbox": [ + 313, + 104, + 555, + 224 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 313, + 104, + 555, + 224 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 340, + 232, + 555, + 295 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 340, + 232, + 555, + 295 + ], + "spans": [ + { + "bbox": [ + 340, + 232, + 555, + 295 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\xi_ {\\Phi} (B) = \\arg \\max _ {\\theta} \\sum_ {i = 1} ^ {N} g \\left(\\Phi_ {\\theta} \\left(X _ {i}, \\left\\{P _ {j \\rightarrow i} \\right\\} _ {j = 1} ^ {N}\\right), y _ {i}\\right), (1) \\\\ \\text {s . t .} \\sum_ {j = 1} ^ {N} \\left| P _ {j \\rightarrow i} \\right| \\leq B (2) \\\\ \\end{array}", + "image_path": "01040c0694d23c1c9d9fca478840d20ea05815f68d3b0bb2f167fbd71a1053ee.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 302, + 555, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 302, + 555, + 361 + ], + "spans": [ + { + "bbox": [ + 313, + 302, + 555, + 361 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 302, + 555, + 361 + ], + "type": "inline_equation", + "content": "\\Phi_{\\theta}" + }, + { + "bbox": [ + 313, + 302, + 555, + 361 + ], + "type": "text", + "content": " represents the collaborative perception 3D object detection model, " + }, + { + "bbox": [ + 313, + 302, + 555, + 361 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 302, + 555, + 361 + ], + "type": "text", + "content": " denotes the model parameters, " + }, + { + "bbox": [ + 313, + 302, + 555, + 361 + ], + "type": "inline_equation", + "content": "|P_{j\\rightarrow i}|" + }, + { + "bbox": [ + 313, + 302, + 555, + 361 + ], + "type": "text", + "content": " represents the communication volume of the message sent from agent " + }, + { + "bbox": [ + 313, + 302, + 555, + 361 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 302, + 555, + 361 + ], + "type": "text", + "content": " to agent " + }, + { + "bbox": [ + 313, + 302, + 555, + 361 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 302, + 555, + 361 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 302, + 555, + 361 + ], + "type": "inline_equation", + "content": "g(\\cdot ,\\cdot)" + }, + { + "bbox": [ + 313, + 302, + 555, + 361 + ], + "type": "text", + "content": " denotes the 3D object detection evaluation metric." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 369, + 434, + 380 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 369, + 434, + 380 + ], + "spans": [ + { + "bbox": [ + 314, + 369, + 434, + 380 + ], + "type": "text", + "content": "3.2. Overall Architecture" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 387, + 555, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 387, + 555, + 590 + ], + "spans": [ + { + "bbox": [ + 313, + 387, + 555, + 590 + ], + "type": "text", + "content": "The overall architecture of the proposed " + }, + { + "bbox": [ + 313, + 387, + 555, + 590 + ], + "type": "inline_equation", + "content": "\\mathrm{COSDH}" + }, + { + "bbox": [ + 313, + 387, + 555, + 590 + ], + "type": "text", + "content": " is shown in Fig. 2. Each agent first processes its locally observed point cloud " + }, + { + "bbox": [ + 313, + 387, + 555, + 590 + ], + "type": "inline_equation", + "content": "X_{i}" + }, + { + "bbox": [ + 313, + 387, + 555, + 590 + ], + "type": "text", + "content": " through a backbone network based on PointPillar [17] and a demand generator to obtain multi-scale BEV features " + }, + { + "bbox": [ + 313, + 387, + 555, + 590 + ], + "type": "inline_equation", + "content": "\\{F_i^{(l)}\\}_{l = 1,2,\\dots,L}" + }, + { + "bbox": [ + 313, + 387, + 555, + 590 + ], + "type": "text", + "content": " and a demand mask " + }, + { + "bbox": [ + 313, + 387, + 555, + 590 + ], + "type": "inline_equation", + "content": "D_{i}" + }, + { + "bbox": [ + 313, + 387, + 555, + 590 + ], + "type": "text", + "content": ", respectively. Considering the collaboration between Ego agent " + }, + { + "bbox": [ + 313, + 387, + 555, + 590 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 387, + 555, + 590 + ], + "type": "text", + "content": " and collaborating agent " + }, + { + "bbox": [ + 313, + 387, + 555, + 590 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 387, + 555, + 590 + ], + "type": "text", + "content": ", agent " + }, + { + "bbox": [ + 313, + 387, + 555, + 590 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 387, + 555, + 590 + ], + "type": "text", + "content": " generates a supply mask " + }, + { + "bbox": [ + 313, + 387, + 555, + 590 + ], + "type": "inline_equation", + "content": "S_{j}" + }, + { + "bbox": [ + 313, + 387, + 555, + 590 + ], + "type": "text", + "content": " from its multi-scale features " + }, + { + "bbox": [ + 313, + 387, + 555, + 590 + ], + "type": "inline_equation", + "content": "\\{F_j^{(l)}\\}_{l = 1,2,\\dots,L}" + }, + { + "bbox": [ + 313, + 387, + 555, + 590 + ], + "type": "text", + "content": " via a supply generator, and multiplies it element-wise with the received demand matrix to obtain the supply-demand mask " + }, + { + "bbox": [ + 313, + 387, + 555, + 590 + ], + "type": "inline_equation", + "content": "M_{j\\rightarrow i}" + }, + { + "bbox": [ + 313, + 387, + 555, + 590 + ], + "type": "text", + "content": ". Agent " + }, + { + "bbox": [ + 313, + 387, + 555, + 590 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 387, + 555, + 590 + ], + "type": "text", + "content": " then performs supply-demand-aware information selection by multiplying " + }, + { + "bbox": [ + 313, + 387, + 555, + 590 + ], + "type": "inline_equation", + "content": "\\{F_j^{(l)}\\}_{l = 1,2,\\dots,L}" + }, + { + "bbox": [ + 313, + 387, + 555, + 590 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 313, + 387, + 555, + 590 + ], + "type": "inline_equation", + "content": "M_{j\\rightarrow i}" + }, + { + "bbox": [ + 313, + 387, + 555, + 590 + ], + "type": "text", + "content": " element-wise to obtain sparse spatial features " + }, + { + "bbox": [ + 313, + 387, + 555, + 590 + ], + "type": "inline_equation", + "content": "\\{Z_j^{(l)}\\}_{l = 1,2,\\dots,L}" + }, + { + "bbox": [ + 313, + 387, + 555, + 590 + ], + "type": "text", + "content": ". Subsequently, agent " + }, + { + "bbox": [ + 313, + 387, + 555, + 590 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 387, + 555, + 590 + ], + "type": "text", + "content": " compresses the features through an autoencoder, sending the non-zero parts of the features along with their corresponding coordinates as the message " + }, + { + "bbox": [ + 313, + 387, + 555, + 590 + ], + "type": "inline_equation", + "content": "P_{j\\rightarrow i}" + }, + { + "bbox": [ + 313, + 387, + 555, + 590 + ], + "type": "text", + "content": " to agent " + }, + { + "bbox": [ + 313, + 387, + 555, + 590 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 387, + 555, + 590 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 590, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 590, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 590, + 556, + 715 + ], + "type": "text", + "content": "Upon receiving " + }, + { + "bbox": [ + 313, + 590, + 556, + 715 + ], + "type": "inline_equation", + "content": "P_{j \\to i}" + }, + { + "bbox": [ + 313, + 590, + 556, + 715 + ], + "type": "text", + "content": ", agent " + }, + { + "bbox": [ + 313, + 590, + 556, + 715 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 590, + 556, + 715 + ], + "type": "text", + "content": " first decodes the features to restore their dimensions and then fuses them with its local features " + }, + { + "bbox": [ + 313, + 590, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\{F_i^{(l)}\\}_{l = 1,2,\\dots,L}" + }, + { + "bbox": [ + 313, + 590, + 556, + 715 + ], + "type": "text", + "content": " across multiple scales to obtain the fused features " + }, + { + "bbox": [ + 313, + 590, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\{F_i^{(l)}\\}_{l = 1,2,\\dots,L}" + }, + { + "bbox": [ + 313, + 590, + 556, + 715 + ], + "type": "text", + "content": ", which are then passed to the detection head for intermediate collaborative detection results " + }, + { + "bbox": [ + 313, + 590, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\tilde{y}_i" + }, + { + "bbox": [ + 313, + 590, + 556, + 715 + ], + "type": "text", + "content": ". Afterward, we apply confidence-aware late fusion: agent " + }, + { + "bbox": [ + 313, + 590, + 556, + 715 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 590, + 556, + 715 + ], + "type": "text", + "content": " filters and suppresses its own detection results " + }, + { + "bbox": [ + 313, + 590, + 556, + 715 + ], + "type": "inline_equation", + "content": "y_j" + }, + { + "bbox": [ + 313, + 590, + 556, + 715 + ], + "type": "text", + "content": " based on confidence and sends them to agent " + }, + { + "bbox": [ + 313, + 590, + 556, + 715 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 590, + 556, + 715 + ], + "type": "text", + "content": " for late fusion, yielding the final hybrid collaborative perception detection results " + }, + { + "bbox": [ + 313, + 590, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\hat{y}_i" + }, + { + "bbox": [ + 313, + 590, + 556, + 715 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "6836" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 63, + 70, + 547, + 238 + ], + "blocks": [ + { + "bbox": [ + 63, + 70, + 547, + 238 + ], + "lines": [ + { + "bbox": [ + 63, + 70, + 547, + 238 + ], + "spans": [ + { + "bbox": [ + 63, + 70, + 547, + 238 + ], + "type": "image", + "image_path": "e61490036051356a3fb3572ee96508d10627968156b304fe0182b559fa40aa0d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 247, + 555, + 282 + ], + "lines": [ + { + "bbox": [ + 54, + 247, + 555, + 282 + ], + "spans": [ + { + "bbox": [ + 54, + 247, + 555, + 282 + ], + "type": "text", + "content": "Figure 2. The overall architecture of " + }, + { + "bbox": [ + 54, + 247, + 555, + 282 + ], + "type": "inline_equation", + "content": "\\mathrm{COSDH}" + }, + { + "bbox": [ + 54, + 247, + 555, + 282 + ], + "type": "text", + "content": ". The Supply-Demand-Aware Information Selection module selects sparse but important information, which is then further compressed by the Message Compression module to achieve efficient communication. Confidence-Aware Late Fusion compensates for the intermediate fusion detection results to improve accuracy." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 294, + 293, + 306 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 294, + 293, + 306 + ], + "spans": [ + { + "bbox": [ + 55, + 294, + 293, + 306 + ], + "type": "text", + "content": "3.3. Supply-Demand-Aware Information Selection" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 314, + 296, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 314, + 296, + 469 + ], + "spans": [ + { + "bbox": [ + 54, + 314, + 296, + 469 + ], + "type": "text", + "content": "Previous methods such as Where2comm [11] and How2comm [40] generally use symmetric supply-demand relationships to select sparse features for collaboration, believing that areas with high foreground confidence in the collaborating agent's view should be provided, and conversely, areas with low foreground confidence in the Ego agent's view need collaboration. However, we believe that areas with low foreground confidence from the Ego's perspective can be divided into areas that are hard to observe and those that can be observed but belong to the background. The latter do not require collaboration. Based on this insight, we propose a novel supply-demand-aware information selection method." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 472, + 296, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 472, + 296, + 674 + ], + "spans": [ + { + "bbox": [ + 55, + 472, + 296, + 674 + ], + "type": "text", + "content": "The demand mask " + }, + { + "bbox": [ + 55, + 472, + 296, + 674 + ], + "type": "inline_equation", + "content": "D_{i}" + }, + { + "bbox": [ + 55, + 472, + 296, + 674 + ], + "type": "text", + "content": " indicates where agent " + }, + { + "bbox": [ + 55, + 472, + 296, + 674 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 472, + 296, + 674 + ], + "type": "text", + "content": " needs information from collaborating agents. Intuitively, the agent requires information from areas that are distant or occluded, which have the common characteristic of having low point cloud density or no point cloud at all. For agent " + }, + { + "bbox": [ + 55, + 472, + 296, + 674 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 472, + 296, + 674 + ], + "type": "text", + "content": ", we consider using the number of point clouds in each pillar to represent point cloud density, and we map it to the range [0,1], i.e., " + }, + { + "bbox": [ + 55, + 472, + 296, + 674 + ], + "type": "inline_equation", + "content": "A_{i} \\in [0,1]^{H \\times W}" + }, + { + "bbox": [ + 55, + 472, + 296, + 674 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 472, + 296, + 674 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 55, + 472, + 296, + 674 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 472, + 296, + 674 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 55, + 472, + 296, + 674 + ], + "type": "text", + "content": " represent the number of Pillars along the height and width dimensions. We then select areas where the point cloud density is below a threshold " + }, + { + "bbox": [ + 55, + 472, + 296, + 674 + ], + "type": "inline_equation", + "content": "\\varepsilon_{a}" + }, + { + "bbox": [ + 55, + 472, + 296, + 674 + ], + "type": "text", + "content": " to obtain the demand mask for agent " + }, + { + "bbox": [ + 55, + 472, + 296, + 674 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 472, + 296, + 674 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 472, + 296, + 674 + ], + "type": "inline_equation", + "content": "D_{i} = A_{i} < \\varepsilon_{a} \\in \\{0,1\\}^{H \\times W}" + }, + { + "bbox": [ + 55, + 472, + 296, + 674 + ], + "type": "text", + "content": ". The demand mask indicates where agent " + }, + { + "bbox": [ + 55, + 472, + 296, + 674 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 472, + 296, + 674 + ], + "type": "text", + "content": " has poor perception and needs collaborative information from other agents. Filtering information from other agents using the demand mask not only helps save bandwidth but also avoids interference from other agents' information in well-perceived areas." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "For the object detection task, foreground information is more valuable. Providing sparse foreground features can effectively assist other agents in supplementing undetected" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 317, + 293, + 560, + 453 + ], + "blocks": [ + { + "bbox": [ + 317, + 293, + 560, + 453 + ], + "lines": [ + { + "bbox": [ + 317, + 293, + 560, + 453 + ], + "spans": [ + { + "bbox": [ + 317, + 293, + 560, + 453 + ], + "type": "image", + "image_path": "564f796406eb4f74420469c5a1b582df37d4f9e6f9112793c8facaa7a1529079.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 461, + 555, + 495 + ], + "lines": [ + { + "bbox": [ + 313, + 461, + 555, + 495 + ], + "spans": [ + { + "bbox": [ + 313, + 461, + 555, + 495 + ], + "type": "text", + "content": "Figure 3. Comparison of our multi-scale compression and fusion with other methods. It can achieve thorough fusion with smaller communication volume." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 512, + 555, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 512, + 555, + 609 + ], + "spans": [ + { + "bbox": [ + 313, + 512, + 555, + 609 + ], + "type": "text", + "content": "and incomplete targets while using less bandwidth. Following previous work [11], we use the spatial confidence map " + }, + { + "bbox": [ + 313, + 512, + 555, + 609 + ], + "type": "inline_equation", + "content": "C_i \\in [0,1]^{H \\times W}" + }, + { + "bbox": [ + 313, + 512, + 555, + 609 + ], + "type": "text", + "content": " output by the detection head to select potential foreground areas. We use a supply threshold " + }, + { + "bbox": [ + 313, + 512, + 555, + 609 + ], + "type": "inline_equation", + "content": "\\varepsilon_c" + }, + { + "bbox": [ + 313, + 512, + 555, + 609 + ], + "type": "text", + "content": " to obtain the supply mask " + }, + { + "bbox": [ + 313, + 512, + 555, + 609 + ], + "type": "inline_equation", + "content": "S_i^{(l)} = C_i > \\varepsilon_c \\in \\{0,1\\}^{H \\times W}" + }, + { + "bbox": [ + 313, + 512, + 555, + 609 + ], + "type": "text", + "content": ". By adjusting the threshold, we can dynamically adjust the bandwidth used for collaborative perception to adapt to varying communication conditions." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 611, + 556, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 611, + 556, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 611, + 556, + 712 + ], + "type": "text", + "content": "During collaboration, agent " + }, + { + "bbox": [ + 313, + 611, + 556, + 712 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 611, + 556, + 712 + ], + "type": "text", + "content": " generates a binary supply-demand selection mask " + }, + { + "bbox": [ + 313, + 611, + 556, + 712 + ], + "type": "inline_equation", + "content": "M_{j\\rightarrow i} = D_i\\odot S_j\\in \\{0,1\\}^{H\\times W}" + }, + { + "bbox": [ + 313, + 611, + 556, + 712 + ], + "type": "text", + "content": " based on its supply mask " + }, + { + "bbox": [ + 313, + 611, + 556, + 712 + ], + "type": "inline_equation", + "content": "S_{j}" + }, + { + "bbox": [ + 313, + 611, + 556, + 712 + ], + "type": "text", + "content": " and agent " + }, + { + "bbox": [ + 313, + 611, + 556, + 712 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 611, + 556, + 712 + ], + "type": "text", + "content": " 's demand mask " + }, + { + "bbox": [ + 313, + 611, + 556, + 712 + ], + "type": "inline_equation", + "content": "D_{i}" + }, + { + "bbox": [ + 313, + 611, + 556, + 712 + ], + "type": "text", + "content": " and samples it to multiply element-wise with multi-scale BEV features " + }, + { + "bbox": [ + 313, + 611, + 556, + 712 + ], + "type": "inline_equation", + "content": "\\{F_j^{(l)}\\}_{l = 1,2,\\dots,L}" + }, + { + "bbox": [ + 313, + 611, + 556, + 712 + ], + "type": "text", + "content": ", obtaining sparse features " + }, + { + "bbox": [ + 313, + 611, + 556, + 712 + ], + "type": "inline_equation", + "content": "\\{Z_{j\\rightarrow i}^{(l)}\\}_{l = 1,2,\\dots,L}" + }, + { + "bbox": [ + 313, + 611, + 556, + 712 + ], + "type": "text", + "content": ". During communication, only the non-zero parts and their corresponding coordinates need to be transmitted." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6837" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 236, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 236, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 236, + 85 + ], + "type": "text", + "content": "3.4. Message Compression and Fusion" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 295 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 295 + ], + "type": "text", + "content": "To further reduce communication bandwidth, we also use autoencoders to compress intermediate features along the channel dimension before communication, while making further improvements to existing compression and fusion schemes. As shown in Fig. 3, traditional single-scale fusion schemes [35, 36] only compress and fuse a single-layer BEV feature map before passing it to the detection head, resulting in insufficient fusion. CoAlign [25] proposed a solution that performs layer-wise fusion on multi-scale features, addressing the issue of insufficient fusion. However, it compresses large-scale single-layer features extracted after Pillar Feature Extraction (PFE), which leads to higher bandwidth. Therefore, we propose a multi-scale compression and fusion approach, where separate autoencoders are designed for each scale of features to perform compression, followed by layer-wise fusion. This approach enables more thorough fusion while using less bandwidth." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 298, + 296, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 298, + 296, + 395 + ], + "spans": [ + { + "bbox": [ + 55, + 298, + 296, + 395 + ], + "type": "text", + "content": "For message fusion, we use Max fusion, which has two main advantages: i) it is computationally simple and efficient, with the complexity increasing linearly with the number of agents; ii) max fusion selects the maximum value of features from multiple agents, achieving information complementarity. Considering the collaboration between Ego agent " + }, + { + "bbox": [ + 55, + 298, + 296, + 395 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 298, + 296, + 395 + ], + "type": "text", + "content": " and collaborating agent " + }, + { + "bbox": [ + 55, + 298, + 296, + 395 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 55, + 298, + 296, + 395 + ], + "type": "text", + "content": ", the process of message compression and fusion can be expressed as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 91, + 409, + 295, + 430 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 409, + 295, + 430 + ], + "spans": [ + { + "bbox": [ + 91, + 409, + 295, + 430 + ], + "type": "interline_equation", + "content": "Z _ {j \\rightarrow i} ^ {\\prime (l)} = f _ {e n c o d e} ^ {(l)} \\left(Z _ {j \\rightarrow i} ^ {(l)}\\right) \\in \\mathbb {R} ^ {\\frac {C _ {l}}{c _ {0}} \\times H _ {l} \\times W _ {l}} \\tag {3}", + "image_path": "e24c48d9cbaffba0d19040904047ff5b09fdd12de2fe5d9664aed2dd4252bb1e.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 91, + 431, + 295, + 449 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 431, + 295, + 449 + ], + "spans": [ + { + "bbox": [ + 91, + 431, + 295, + 449 + ], + "type": "interline_equation", + "content": "F _ {j \\rightarrow i} ^ {(l)} = f _ {\\text {d e c o d e}} ^ {(l)} \\left(Z _ {j \\rightarrow i} ^ {' (l)}\\right) \\in \\mathbb {R} ^ {C _ {l} \\times H _ {l} \\times W _ {l}} \\tag {4}", + "image_path": "a218c5521a0d1d763c968b986d71219fbc2e7c855f4b2b31de3d189aefe65ba2.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 91, + 451, + 295, + 468 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 451, + 295, + 468 + ], + "spans": [ + { + "bbox": [ + 91, + 451, + 295, + 468 + ], + "type": "interline_equation", + "content": "F _ {j \\rightarrow i} ^ {\\prime (l)} = f _ {\\text {t r a n s f o r m}} \\left(F _ {j \\rightarrow i} ^ {(l)}, \\xi_ {j \\rightarrow i}\\right) \\in \\mathbb {R} ^ {C _ {l} \\times H _ {l} \\times W _ {l}} \\tag {5}", + "image_path": "b7d638d6b61291bae4dfeceb414fed21160c60399bd2cd6af1e9391920e827ff.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 94, + 471, + 294, + 491 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 471, + 294, + 491 + ], + "spans": [ + { + "bbox": [ + 94, + 471, + 294, + 491 + ], + "type": "interline_equation", + "content": "\\tilde {F _ {i} ^ {(l)}} = \\max \\left(F _ {i} ^ {(l)}, \\left\\{F _ {j \\rightarrow i} ^ {\\prime (l)} \\right\\} _ {j \\neq i}\\right) \\in \\mathbb {R} ^ {C _ {l} \\times H _ {l} \\times W _ {l}} \\tag {6}", + "image_path": "fd27416e8480093a98fe48d7df328f61217f30166d53f2ede3ec59b8afff2e0a.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 508, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 508, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 508, + 296, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 508, + 296, + 715 + ], + "type": "inline_equation", + "content": "Z_{j\\rightarrow i}^{(l)}\\in \\mathbb{R}^{C_l\\times H_l\\times W_l}" + }, + { + "bbox": [ + 55, + 508, + 296, + 715 + ], + "type": "text", + "content": " represents the sparse features selected based on supply-demand relationships from the previous stage, and " + }, + { + "bbox": [ + 55, + 508, + 296, + 715 + ], + "type": "inline_equation", + "content": "C_l,H_l,W_l" + }, + { + "bbox": [ + 55, + 508, + 296, + 715 + ], + "type": "text", + "content": " denote the channel, height, and width dimensions of the " + }, + { + "bbox": [ + 55, + 508, + 296, + 715 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 55, + 508, + 296, + 715 + ], + "type": "text", + "content": " -th layer feature map. " + }, + { + "bbox": [ + 55, + 508, + 296, + 715 + ], + "type": "inline_equation", + "content": "f_{encode}^{(l)},f_{decode}^{(l)}" + }, + { + "bbox": [ + 55, + 508, + 296, + 715 + ], + "type": "text", + "content": " represent the encoder and decoder of the " + }, + { + "bbox": [ + 55, + 508, + 296, + 715 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 55, + 508, + 296, + 715 + ], + "type": "text", + "content": " -th layer autoencoder, " + }, + { + "bbox": [ + 55, + 508, + 296, + 715 + ], + "type": "inline_equation", + "content": "c_{0}" + }, + { + "bbox": [ + 55, + 508, + 296, + 715 + ], + "type": "text", + "content": " is the compression ratio in the channel dimension, and " + }, + { + "bbox": [ + 55, + 508, + 296, + 715 + ], + "type": "inline_equation", + "content": "f_{transform}" + }, + { + "bbox": [ + 55, + 508, + 296, + 715 + ], + "type": "text", + "content": " represents coordinate transformation. During communication, " + }, + { + "bbox": [ + 55, + 508, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\{Z_{j\\rightarrow i}^{\\prime (l)}\\}_{l = 1,2,\\ldots ,L}" + }, + { + "bbox": [ + 55, + 508, + 296, + 715 + ], + "type": "text", + "content": " is first converted from float32 to float16, and then only the non-zero parts and corresponding coordinates are transmitted to save bandwidth. Upon receiving the message, agent " + }, + { + "bbox": [ + 55, + 508, + 296, + 715 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 508, + 296, + 715 + ], + "type": "text", + "content": " decodes the feature dimensions, then aligns the features to its own coordinate system using the coordinate transformation matrix " + }, + { + "bbox": [ + 55, + 508, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\xi_{j\\rightarrow i}" + }, + { + "bbox": [ + 55, + 508, + 296, + 715 + ], + "type": "text", + "content": " to obtain " + }, + { + "bbox": [ + 55, + 508, + 296, + 715 + ], + "type": "inline_equation", + "content": "F_{j\\rightarrow i}^{\\prime (l)}" + }, + { + "bbox": [ + 55, + 508, + 296, + 715 + ], + "type": "text", + "content": " , and finally fuses its own features with the collaborating agent's features using max fusion to obtain the fused features " + }, + { + "bbox": [ + 55, + 508, + 296, + 715 + ], + "type": "inline_equation", + "content": "F_{i}^{(l)}" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 317, + 73, + 559, + 216 + ], + "blocks": [ + { + "bbox": [ + 317, + 73, + 559, + 216 + ], + "lines": [ + { + "bbox": [ + 317, + 73, + 559, + 216 + ], + "spans": [ + { + "bbox": [ + 317, + 73, + 559, + 216 + ], + "type": "image", + "image_path": "80aeadd77d05b25ee99ff4941272c79b32d1f4a1cb80af8bb919803063e6bced.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 226, + 555, + 260 + ], + "lines": [ + { + "bbox": [ + 313, + 226, + 555, + 260 + ], + "spans": [ + { + "bbox": [ + 313, + 226, + 555, + 260 + ], + "type": "text", + "content": "Figure 4. Our confidence-aware late fusion. It filters detection results based on confidence and suppresses suboptimal results from collaborative agents, improving overall detection accuracy." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 285, + 481, + 296 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 285, + 481, + 296 + ], + "spans": [ + { + "bbox": [ + 313, + 285, + 481, + 296 + ], + "type": "text", + "content": "3.5. Confidence-Aware Late Fusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 307, + 555, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 307, + 555, + 415 + ], + "spans": [ + { + "bbox": [ + 313, + 307, + 555, + 415 + ], + "type": "text", + "content": "Existing collaborative perception methods focus on the singular intermediate collaboration architecture and neglect the advantages of late collaboration. As shown in Table 1, the experimental results indicate that late fusion can demonstrate acceptable perception accuracy with extremely low bandwidth on the OPV2V [36] and V2XSim [20] datasets. Therefore, late fusion can be used to compensate for the results of intermediate fusion, achieving a better accuracy-bandwidth trade-off." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 421, + 556, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 421, + 556, + 553 + ], + "spans": [ + { + "bbox": [ + 313, + 421, + 556, + 553 + ], + "type": "text", + "content": "Naive late collaboration methods [20, 36] directly merge results from other agents and then apply NMS (Non-Maximum Suppression) to deduplicate and obtain final results. This approach can effectively improve recall for collaborative perception object detection; however, during the merging process, it may introduce some low-confidence false positives, and suboptimal detection results from collaborating agents may override the Ego agent's detection results, lowering precision and consequently decreasing final AP. The lower AP in the late fusion method in Table 1 on the DAIR-V2X [42] dataset is due to this issue." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "type": "text", + "content": "To address this problem, we propose a new confidence-aware late fusion method. As shown in Fig. 4, during late fusion, we first filter based on the confidence of the target boxes, discarding target boxes from other agents with confidence lower than " + }, + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\varepsilon_{l}" + }, + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "type": "text", + "content": ". Furthermore, considering that if both the Ego agent and other agents detect the same target, the detection result from the Ego agent, which has undergone intermediate fusion, is of higher quality. Therefore, we suppress the detection results from other agents to prevent their lower-quality results from degrading the Ego agent's better detection results. Specifically, before merging the detection results, we multiply the confidence of target boxes from other agents by a coefficient " + }, + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\beta \\in (0,1)" + }, + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6838" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 69, + 553, + 274 + ], + "blocks": [ + { + "bbox": [ + 58, + 69, + 553, + 274 + ], + "lines": [ + { + "bbox": [ + 58, + 69, + 553, + 274 + ], + "spans": [ + { + "bbox": [ + 58, + 69, + 553, + 274 + ], + "type": "table", + "html": "
SettingMethodOPV2V [36]V2XSim [20]DAIR-V2X [42]
AP@0.5↑AP@0.7↑BD↓AP@0.5↑AP@0.7↑BD↓AP@0.5↑AP@0.7↑BD↓
BasicNo Fusion79.78%67.16%0.0 Mbps70.31%58.55%0.0 Mbps66.51%55.46%0.0 Mbps
Early Fusion95.05%88.98%83.1 Mbps95.68%88.03%55.5 Mbps74.52%59.22%50.2 Mbps
Late Fusion94.70%87.18%0.2 Mbps86.88%78.13%0.1 Mbps67.86%50.47%0.2 Mbps
No LimitWhen2com [21]91.75%81.77%1,320.0 Mbps72.65%62.92%250.0 Mbps64.08%49.14%984.8 Mbps
FCooper [3]90.06%74.03%2,640.0 Mbps72.96%57.61%500.0 Mbps74.58%56.47%1,968.8 Mbps
AttFuse [36]94.31%82.03%2,640.0 Mbps78.06%64.84%500.0 Mbps73.80%56.86%1,968.8 Mbps
V2VNet [32]96.66%92.44%5,280.0 Mbps88.97%85.18%1,000.0 Mbps66.63%47.39%3,937.5 Mbps
DiscoNet [19]90.93%78.90%2,640.0 Mbps77.34%68.77%500.0 Mbps73.58%58.45%1,968.8 Mbps
V2XViT [35]95.87%89.88%2,640.0 Mbps89.01%80.26%500.0 Mbps76.68%57.57%1,920.0 Mbps
Where2comm [11]95.59%91.39%48.7 Mbps88.18%83.66%27.5 Mbps76.13%60.16%172.3 Mbps
CoAlign [25]96.63%92.63%2,640.0 Mbps88.87%85.23%500.0 Mbps78.06%63.09%1,968.8 Mbps
CoSDH96.83%92.99%13.4 Mbps89.23%86.31%1.1 Mbps76.75%63.85%7.1 Mbps
BD ≤ 6.75 MbpsWhen2com [21]79.97%50.88%5.2 Mbps62.02%43.37%4.0 Mbps62.37%44.01%3.8 Mbps
FCooper [3]90.37%72.92%5.2 Mbps76.52%61.83%4.0 Mbps67.71%47.65%3.8 Mbps
AttFuse [36]93.45%80.02%5.2 Mbps84.58%71.36%4.0 Mbps71.06%49.57%3.8 Mbps
V2VNet [32]95.71%87.16%5.2 Mbps85.66%73.72%4.0 Mbps66.49%45.61%3.8 Mbps
DiscoNet [19]90.00%76.72%5.2 Mbps78.04%68.18%4.0 Mbps70.83%53.40%3.8 Mbps
V2XViT [35]95.85%87.13%5.2 Mbps88.94%81.47%4.0 Mbps71.00%52.78%3.8 Mbps
Where2comm [11]94.91%89.86%5.4 Mbps87.51%82.12%4.7 Mbps74.98%59.51%5.3 Mbps
CoAlign [25]94.10%85.99%5.2 Mbps88.01%83.97%4.0 Mbps75.26%60.19%3.8 Mbps
CoSDH96.75%92.92%2.0 Mbps89.23%86.31%1.1 Mbps76.47%63.76%1.4 Mbps
", + "image_path": "60e243a33561c4bfd56b8961edf6821bdd9a1b98bf90aeaa352a6ce69f45a54b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 55, + 281, + 555, + 348 + ], + "lines": [ + { + "bbox": [ + 55, + 281, + 555, + 348 + ], + "spans": [ + { + "bbox": [ + 55, + 281, + 555, + 348 + ], + "type": "text", + "content": "Table 1. Comparison of detection accuracy and bandwidth of different methods on OPV2V [36], V2XSim [20], and DAIR-V2X [42] datasets. \"BD\" represents the bandwidth required for each collaborative agent, assuming Ego agent collaborates with up to 4 agents and the detection frequency is " + }, + { + "bbox": [ + 55, + 281, + 555, + 348 + ], + "type": "inline_equation", + "content": "10\\mathrm{Hz}" + }, + { + "bbox": [ + 55, + 281, + 555, + 348 + ], + "type": "text", + "content": ". \"BD " + }, + { + "bbox": [ + 55, + 281, + 555, + 348 + ], + "type": "inline_equation", + "content": "\\leq" + }, + { + "bbox": [ + 55, + 281, + 555, + 348 + ], + "type": "text", + "content": " 6.75 Mbps\" is used to simulate real-world communication limits, assuming a total communication bandwidth of 27 Mbps, with each collaborating agent's bandwidth consumption limited to less than 6.75 Mbps. For intermediate collaboration methods without information selection, we provide a compressed version using autoencoders to meet the bandwidth constraints. For intermediate collaboration methods with information selection, the selection ratio is adjusted to meet the bandwidth constraints." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 361, + 137, + 373 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 361, + 137, + 373 + ], + "spans": [ + { + "bbox": [ + 55, + 361, + 137, + 373 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 379, + 245, + 392 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 379, + 245, + 392 + ], + "spans": [ + { + "bbox": [ + 55, + 379, + 245, + 392 + ], + "type": "text", + "content": "4.1. Datasets and Experimental Settings" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 398, + 295, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 398, + 295, + 481 + ], + "spans": [ + { + "bbox": [ + 55, + 398, + 295, + 481 + ], + "type": "text", + "content": "Datasets. We evaluate the proposed CoSDH against other methods on three different collaborative perception datasets (OPV2V [36], V2XSim [20], and DAIR-V2X [42]) for LiDAR-based 3D object detection. The datasets include both simulated and real-world scenarios, and cover two types of collaboration: V2V (Vehicle to Vehicle) and V2I (Vehicle to Infrastructure)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 484, + 295, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 484, + 295, + 651 + ], + "spans": [ + { + "bbox": [ + 55, + 484, + 295, + 651 + ], + "type": "text", + "content": "Evaluation Metrics We use the average precision (AP) with intersection-over-union (IoU) thresholds of 0.5 and 0.7 to evaluate the performance of different methods on 3D object detection. We assume the target detection frequency is " + }, + { + "bbox": [ + 55, + 484, + 295, + 651 + ], + "type": "inline_equation", + "content": "10\\mathrm{Hz}" + }, + { + "bbox": [ + 55, + 484, + 295, + 651 + ], + "type": "text", + "content": " and calculate the communication bandwidth based on the average data transmitted by each collaborative agent to the Ego agent, in order to evaluate the communication cost of different methods. Specifically, we consider the bandwidth limitations in real-world collaborative perception scenarios, setting the vehicle's communication rate to 27 Mbps [1, 24]. Considering the typical case where the Ego agent collaborates with up to 4 other agents [36], the bandwidth limit for each collaborative agent is " + }, + { + "bbox": [ + 55, + 484, + 295, + 651 + ], + "type": "inline_equation", + "content": "27/4 = 6.75" + }, + { + "bbox": [ + 55, + 484, + 295, + 651 + ], + "type": "text", + "content": " Mbps." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "content": "Implementation Our experiments are based on the OpenCOOD [36] framework, using PointPillar [17] as the encoder with a grid size of " + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "inline_equation", + "content": "(0.4m, 0.4m)" + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "content": ", and a maximum of 32 points per Pillar. For our method, we set the number of intermediate feature layers to " + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "inline_equation", + "content": "L = 3" + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "content": ", the demand threshold" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 361, + 555, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 361, + 555, + 589 + ], + "spans": [ + { + "bbox": [ + 313, + 361, + 555, + 589 + ], + "type": "inline_equation", + "content": "\\varepsilon_{a} = 4 / 32 = 0.125" + }, + { + "bbox": [ + 313, + 361, + 555, + 589 + ], + "type": "text", + "content": ", and the supply threshold " + }, + { + "bbox": [ + 313, + 361, + 555, + 589 + ], + "type": "inline_equation", + "content": "\\varepsilon_{c} = 0.01" + }, + { + "bbox": [ + 313, + 361, + 555, + 589 + ], + "type": "text", + "content": ". For the OPV2V [36] and DAIR-V2X [42] datasets, the compression rate is " + }, + { + "bbox": [ + 313, + 361, + 555, + 589 + ], + "type": "inline_equation", + "content": "c_{0} = 16" + }, + { + "bbox": [ + 313, + 361, + 555, + 589 + ], + "type": "text", + "content": ", and for the V2XSim [20] dataset, the compression rate is " + }, + { + "bbox": [ + 313, + 361, + 555, + 589 + ], + "type": "inline_equation", + "content": "c_{0} = 8" + }, + { + "bbox": [ + 313, + 361, + 555, + 589 + ], + "type": "text", + "content": " due to its smaller perception range and lower inherent bandwidth requirement. The late fusion threshold is set to " + }, + { + "bbox": [ + 313, + 361, + 555, + 589 + ], + "type": "inline_equation", + "content": "\\varepsilon_{l} = 0.3" + }, + { + "bbox": [ + 313, + 361, + 555, + 589 + ], + "type": "text", + "content": ", with a suppression coefficient " + }, + { + "bbox": [ + 313, + 361, + 555, + 589 + ], + "type": "inline_equation", + "content": "\\beta = 0.9" + }, + { + "bbox": [ + 313, + 361, + 555, + 589 + ], + "type": "text", + "content": " for OPV2V and V2XSim dataset and " + }, + { + "bbox": [ + 313, + 361, + 555, + 589 + ], + "type": "inline_equation", + "content": "\\beta = 0.8" + }, + { + "bbox": [ + 313, + 361, + 555, + 589 + ], + "type": "text", + "content": " for DAIR-V2X dataset because of its difficulty. In late fusion, the dense prediction results before NMS are transmitted, as this reduces the computational burden on the collaborating agents. The experiments use the Adam [14] optimizer, with an initial learning rate set between 0.0001 and 0.002 based on the model's testing complexity to ensure proper training. The maximum number of collaborating agents is set to 5. The number of training epochs is set to 40 to ensure model convergence. Other experimental parameters are kept consistent with the OpenCOOD framework. All methods are trained on four NVIDIA GeForce RTX 3090 GPUs." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 599, + 450, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 599, + 450, + 611 + ], + "spans": [ + { + "bbox": [ + 313, + 599, + 450, + 611 + ], + "type": "text", + "content": "4.2.Quantitative Evaluation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "type": "text", + "content": "Benchmark Comparison. Table 1 presents the collaborative 3D object detection accuracy and required bandwidth of the proposed CoSDH compared to previous methods across different datasets. Experimental results show that, with the default uncompressed settings, CoSDH achieves the highest accuracy on the OPV2V [36] and V2XSim [20] datasets while requiring only about 1/100 to 1/1000 of the bandwidth compared to other non-communication-efficient" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6839" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 70, + 223, + 202 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 223, + 202 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 223, + 202 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 223, + 202 + ], + "type": "image", + "image_path": "7db46af544b2102d3c9ffe902cb3baa9ea6fb8a5a22940a03b130004298bb20b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 223, + 70, + 386, + 202 + ], + "blocks": [ + { + "bbox": [ + 223, + 70, + 386, + 202 + ], + "lines": [ + { + "bbox": [ + 223, + 70, + 386, + 202 + ], + "spans": [ + { + "bbox": [ + 223, + 70, + 386, + 202 + ], + "type": "image", + "image_path": "b4969047a0f11c5b7a790bee2ed710f7b5bc4d5af5baed8b60554f1aafa82f0d.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 386, + 70, + 549, + 202 + ], + "blocks": [ + { + "bbox": [ + 386, + 70, + 549, + 202 + ], + "lines": [ + { + "bbox": [ + 386, + 70, + 549, + 202 + ], + "spans": [ + { + "bbox": [ + 386, + 70, + 549, + 202 + ], + "type": "image", + "image_path": "db0a639c041739ebdbec2fa7a6dc8654654b5df37a6c1f12776c5742af6fa67e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 59, + 206, + 223, + 332 + ], + "blocks": [ + { + "bbox": [ + 59, + 206, + 223, + 332 + ], + "lines": [ + { + "bbox": [ + 59, + 206, + 223, + 332 + ], + "spans": [ + { + "bbox": [ + 59, + 206, + 223, + 332 + ], + "type": "image", + "image_path": "1432b4733da76085279af19d24efa23ecf1fd4454e9f29ccd4d50ef70a197eab.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 341, + 555, + 374 + ], + "lines": [ + { + "bbox": [ + 55, + 341, + 555, + 374 + ], + "spans": [ + { + "bbox": [ + 55, + 341, + 555, + 374 + ], + "type": "text", + "content": "Figure 5. Comparison of the trade-off between detection accuracy and bandwidth of different methods on OPV2V [36], V2XSim [20] and DAIR-V2X [42] datasets, CoSDH achieves the best accuracy-bandwidth trade-off. The real-world limitation refers to the total bandwidth limit of 27 Mbps, which means that each collaborative agent does not exceed 6.75 Mbps." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 223, + 206, + 386, + 332 + ], + "blocks": [ + { + "bbox": [ + 223, + 206, + 386, + 332 + ], + "lines": [ + { + "bbox": [ + 223, + 206, + 386, + 332 + ], + "spans": [ + { + "bbox": [ + 223, + 206, + 386, + 332 + ], + "type": "image", + "image_path": "25960b1269cdf59a40c0f5507e04ed8b6584511eadae182b809599bb39368e1e.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 386, + 206, + 550, + 332 + ], + "blocks": [ + { + "bbox": [ + 386, + 206, + 550, + 332 + ], + "lines": [ + { + "bbox": [ + 386, + 206, + 550, + 332 + ], + "spans": [ + { + "bbox": [ + 386, + 206, + 550, + 332 + ], + "type": "image", + "image_path": "b31a4409efb9422a4d40e23bef779118459e9b3b5ac96a53e320fb7bd610d8fc.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 388, + 295, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 388, + 295, + 508 + ], + "spans": [ + { + "bbox": [ + 54, + 388, + 295, + 508 + ], + "type": "text", + "content": "methods. Although the AP@0.5 of CoSDH on the DAIR-V2X [42] dataset is slightly lower than that of CoAlign [25], it achieves a higher AP@0.7 with only about 1/300 of the bandwidth. Both our CoSDH and Where2comm [11] are communication-efficient methods based on information selection, which enable dynamic accuracy-bandwidth trade-offs by adjusting the selection ratio. The table shows accuracy at a specific bandwidth, and a more detailed comparison of the accuracy-bandwidth curves is provided in the \"Accuracy-Bandwidth Trade-Off Comparison\" part." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 510, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 295, + 714 + ], + "type": "text", + "content": "Furthermore, we simulate real-world communication rate limitations. Result shows that " + }, + { + "bbox": [ + 55, + 510, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{COSDH}" + }, + { + "bbox": [ + 55, + 510, + 295, + 714 + ], + "type": "text", + "content": " achieves improvements in AP@0.7 of " + }, + { + "bbox": [ + 55, + 510, + 295, + 714 + ], + "type": "inline_equation", + "content": "3.06\\% / 2.34\\% / 3.75\\%" + }, + { + "bbox": [ + 55, + 510, + 295, + 714 + ], + "type": "text", + "content": " on OPV2V/V2XSim/DAIR-V2X compared to the previous best methods, while using less bandwidth. When comparing the scenarios with and without bandwidth limitations on the OPV2V and DAIR-V2X datasets, we found that " + }, + { + "bbox": [ + 55, + 510, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{COSDH}" + }, + { + "bbox": [ + 55, + 510, + 295, + 714 + ], + "type": "text", + "content": " achieves less than a " + }, + { + "bbox": [ + 55, + 510, + 295, + 714 + ], + "type": "inline_equation", + "content": "0.3\\%" + }, + { + "bbox": [ + 55, + 510, + 295, + 714 + ], + "type": "text", + "content": " decrease in AP under conditions where bandwidth is reduced by " + }, + { + "bbox": [ + 55, + 510, + 295, + 714 + ], + "type": "inline_equation", + "content": "80\\%" + }, + { + "bbox": [ + 55, + 510, + 295, + 714 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 510, + 295, + 714 + ], + "type": "inline_equation", + "content": "85\\%" + }, + { + "bbox": [ + 55, + 510, + 295, + 714 + ], + "type": "text", + "content": ", exhibiting less accuracy degradation than Where2comm. Most other methods also show varying degrees of accuracy degradation under bandwidth constraints. Interestingly, under bandwidth limitations, some methods show an improvement in accuracy on certain datasets after using autoencoders to compress intermediate features. For example, FCooper [3] shows improved accuracy on the V2XSim dataset, which may be due to the model's initially poor per" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 388, + 555, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 388, + 555, + 425 + ], + "spans": [ + { + "bbox": [ + 313, + 388, + 555, + 425 + ], + "type": "text", + "content": "formance. The compressed version, with the added autoencoder, increases the model's parameter count, thereby enriching its expressive capability." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 438, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 438, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 438, + 556, + 715 + ], + "type": "text", + "content": "Accuracy-Bandwidth Trade-Off Comparison. Fig. 5 shows the accuracy-bandwidth trade-off of the proposed CoSDH and previous advanced methods across different datasets. When the bandwidth grater than 1 Mbps, Where2comm [11] demonstrates a good enough accuracy-bandwidth trade-off, maintaining high detection accuracy as bandwidth decreases. However, as bandwidth further decreases, its accuracy quickly declines, even falling below that of late fusion method. CoSDH can leverage late fusion at low bandwidths to maintain high accuracy, achieving a better performance-bandwidth trade-off. Notably, on the DAIR-V2X dataset, the late fusion method performs worse than the no-fusion case because the naive late fusion method unselectively merges detection results from other agents, introducing a large number of low-quality detections. CoSDH uses a confidence-aware late fusion method to select higher-quality detection results, improving this situation and enhancing perception accuracy. We also observe that the AP@0.7 of CoSDH on the V2XSim dataset initially increases slightly as bandwidth decreases before subsequently declining, indicating that selecting key collaboration areas can help reduce interference from other regions to some extent, thereby slightly improving detection accuracy." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6840" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 70, + 180, + 142 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 180, + 142 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 180, + 142 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 180, + 142 + ], + "type": "image", + "image_path": "c67ec94fd94c5dd281f6e942222661da0330f6dfe7c05ae61a60ca082af49f85.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 82, + 142, + 101, + 148 + ], + "lines": [ + { + "bbox": [ + 82, + 142, + 101, + 148 + ], + "spans": [ + { + "bbox": [ + 82, + 142, + 101, + 148 + ], + "type": "text", + "content": "CoAlign" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 123, + 142, + 176, + 148 + ], + "lines": [ + { + "bbox": [ + 123, + 142, + 176, + 148 + ], + "spans": [ + { + "bbox": [ + 123, + 142, + 176, + 148 + ], + "type": "text", + "content": "Bandwidth: 3.85 Mbps" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 180, + 70, + 303, + 141 + ], + "blocks": [ + { + "bbox": [ + 180, + 70, + 303, + 141 + ], + "lines": [ + { + "bbox": [ + 180, + 70, + 303, + 141 + ], + "spans": [ + { + "bbox": [ + 180, + 70, + 303, + 141 + ], + "type": "image", + "image_path": "e86582efb0df25313fb0e1dc702e3c55eefcce969eb5f89dd88f630f5c436ac9.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 197, + 142, + 230, + 148 + ], + "lines": [ + { + "bbox": [ + 197, + 142, + 230, + 148 + ], + "spans": [ + { + "bbox": [ + 197, + 142, + 230, + 148 + ], + "type": "text", + "content": "Where2comm" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 246, + 142, + 299, + 148 + ], + "lines": [ + { + "bbox": [ + 246, + 142, + 299, + 148 + ], + "spans": [ + { + "bbox": [ + 246, + 142, + 299, + 148 + ], + "type": "text", + "content": "Bandwidth: 5.26 Mbps" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 58, + 150, + 180, + 220 + ], + "blocks": [ + { + "bbox": [ + 58, + 150, + 180, + 220 + ], + "lines": [ + { + "bbox": [ + 58, + 150, + 180, + 220 + ], + "spans": [ + { + "bbox": [ + 58, + 150, + 180, + 220 + ], + "type": "image", + "image_path": "26ad5a127e4a1ab99a2ca6d7f0919770d3530f8e83914f19a8855de206c11b9c.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 63, + 221, + 119, + 228 + ], + "lines": [ + { + "bbox": [ + 63, + 221, + 119, + 228 + ], + "spans": [ + { + "bbox": [ + 63, + 221, + 119, + 228 + ], + "type": "text", + "content": "CoSDH (w/o late fusion)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 237, + 296, + 270 + ], + "lines": [ + { + "bbox": [ + 55, + 237, + 296, + 270 + ], + "spans": [ + { + "bbox": [ + 55, + 237, + 296, + 270 + ], + "type": "text", + "content": "Figure 6. Visualization of detection results under real-world communication rate limitations on the DAIR-V2X [42] dataset. Green represents ground truth box, and red represents predicted box." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 181, + 150, + 302, + 220 + ], + "blocks": [ + { + "bbox": [ + 181, + 150, + 302, + 220 + ], + "lines": [ + { + "bbox": [ + 181, + 150, + 302, + 220 + ], + "spans": [ + { + "bbox": [ + 181, + 150, + 302, + 220 + ], + "type": "image", + "image_path": "df3b39563c59aec3cbec23044cd0307074f1ed1168926dac4f523190071d4d4a.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 205, + 221, + 223, + 227 + ], + "lines": [ + { + "bbox": [ + 205, + 221, + 223, + 227 + ], + "spans": [ + { + "bbox": [ + 205, + 221, + 223, + 227 + ], + "type": "text", + "content": "CoSDH" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 246, + 221, + 299, + 228 + ], + "lines": [ + { + "bbox": [ + 246, + 221, + 299, + 228 + ], + "spans": [ + { + "bbox": [ + 246, + 221, + 299, + 228 + ], + "type": "text", + "content": "Bandwidth: 0.29 Mbps" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 285, + 185, + 298 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 285, + 185, + 298 + ], + "spans": [ + { + "bbox": [ + 55, + 285, + 185, + 298 + ], + "type": "text", + "content": "4.3. Qualitative Evaluation" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 303, + 296, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 303, + 296, + 447 + ], + "spans": [ + { + "bbox": [ + 55, + 303, + 296, + 447 + ], + "type": "text", + "content": "Fig. 6 shows the visualization of detection results for our method compared to other methods under simulated real-world communication rate limitations on the DAIR-V2X dataset. A comparison of the detection results reveals that, while using less bandwidth, our method achieves the same recall rate as CoAlign [25] and Where2comm [11], with fewer false positive predictions and more accurate object localization, demonstrating that our method performs better under low-bandwidth conditions. Comparing the two images at bottom, it can be seen that under low bandwidth conditions, late fusion effectively compensates for the results of intermediate fusion, improving the recall rate of objects." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 456, + 156, + 467 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 456, + 156, + 467 + ], + "spans": [ + { + "bbox": [ + 55, + 456, + 156, + 467 + ], + "type": "text", + "content": "4.4. Ablation Studies" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 474, + 295, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 474, + 295, + 676 + ], + "spans": [ + { + "bbox": [ + 55, + 474, + 295, + 676 + ], + "type": "text", + "content": "Table 2 presents the results of the ablation study on various modules of the proposed method using the OPV2V [36] dataset. The results show that after using the autoencoder for compression, the bandwidth is reduced by a factor of " + }, + { + "bbox": [ + 55, + 474, + 295, + 676 + ], + "type": "inline_equation", + "content": "c_{0}" + }, + { + "bbox": [ + 55, + 474, + 295, + 676 + ], + "type": "text", + "content": ", and there is no significant loss in perception accuracy, even with a slight improvement in AP@0.7. Converting intermediate features to float16 can nearly halve the bandwidth without significant loss, demonstrating the effectiveness of our proposed message compression module. After performing information selection based on the supply mask, perception accuracy slightly decreases, but the bandwidth is reduced by about " + }, + { + "bbox": [ + 55, + 474, + 295, + 676 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 55, + 474, + 295, + 676 + ], + "type": "text", + "content": ", which is a worthwhile trade-off. Further using the demand mask results in no significant decrease in accuracy but reduces bandwidth by about " + }, + { + "bbox": [ + 55, + 474, + 295, + 676 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 55, + 474, + 295, + 676 + ], + "type": "text", + "content": ". This demonstrates that our supply-demand-aware information selection can reduce bandwidth while maintaining accuracy." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "After adding late fusion, accuracy improves at a relatively small bandwidth cost. However, since this phase uses more bandwidth, the improvement in accuracy is not signif" + } + ] + } + ], + "index": 16 + }, + { + "type": "table", + "bbox": [ + 316, + 70, + 558, + 128 + ], + "blocks": [ + { + "bbox": [ + 316, + 70, + 558, + 128 + ], + "lines": [ + { + "bbox": [ + 316, + 70, + 558, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 70, + 558, + 128 + ], + "type": "table", + "html": "
CompressionSelectionLate FusionAP@0.5↑AP@0.7↑BD↓
AutoencoderFP16SupplyDemand
96.62%92.62%1,155.00 Mbps
✓96.59%92.99%72.19 Mbps
✓✓96.59%92.99%36.09 Mbps
✓✓✓96.30%92.62%1.99 Mbps
✓✓✓✓96.31%92.60%1.82 Mbps
✓✓✓✓✓96.75%92.92%1.97 Mbps
", + "image_path": "02dd07231ff7a73b3b1f89fea4923a03773404617f641ce62e39802c222074d5.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "table_body" + } + ], + "index": 17 + }, + { + "type": "table", + "bbox": [ + 316, + 230, + 558, + 285 + ], + "blocks": [ + { + "bbox": [ + 313, + 137, + 555, + 214 + ], + "lines": [ + { + "bbox": [ + 313, + 137, + 555, + 214 + ], + "spans": [ + { + "bbox": [ + 313, + 137, + 555, + 214 + ], + "type": "text", + "content": "Table 2. Ablation study of the modules in " + }, + { + "bbox": [ + 313, + 137, + 555, + 214 + ], + "type": "inline_equation", + "content": "\\mathrm{C}\\circ \\mathrm{{SDH}}" + }, + { + "bbox": [ + 313, + 137, + 555, + 214 + ], + "type": "text", + "content": " on the OPV2V dataset. \"Autoencoder\" refers to the use of autoencoders to compress features along the channel dimension, with a compression ratio of " + }, + { + "bbox": [ + 313, + 137, + 555, + 214 + ], + "type": "inline_equation", + "content": "{c}_{0} = {16}" + }, + { + "bbox": [ + 313, + 137, + 555, + 214 + ], + "type": "text", + "content": " . \"FP16\" refers to converting features from float32 to float16 for transmission. \"Supply\" and \"Demand\" refer to using supply and demand masks for information selection, and \"Late Fusion\" refers to applying confidence-aware late fusion." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 316, + 230, + 558, + 285 + ], + "lines": [ + { + "bbox": [ + 316, + 230, + 558, + 285 + ], + "spans": [ + { + "bbox": [ + 316, + 230, + 558, + 285 + ], + "type": "table", + "html": "
Late Fusionεc0.010.020.030.050.07
AP@0.5↑96.59%96.31%96.19%95.18%93.35%
AP@0.7↑92.95%92.60%92.27%90.67%87.60%
BD↓13.26 Mbps1.82 Mbps0.54 Mbps0.12 Mbps0.05 Mbps
✓AP@0.5↑96.83%96.75%96.72%96.54%96.41%
AP@0.7↑92.99%92.92%92.73%92.23%91.68%
BD↓13.40 Mbps1.97 Mbps0.68 Mbps0.26 Mbps0.19 Mbps
", + "image_path": "d9d43b1cb3150781ef732c8c89ecadf87f5ed02235306f654bc8df4a7e2d6eb0.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "table_body" + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 293, + 555, + 337 + ], + "lines": [ + { + "bbox": [ + 313, + 293, + 555, + 337 + ], + "spans": [ + { + "bbox": [ + 313, + 293, + 555, + 337 + ], + "type": "text", + "content": "Table 3. Ablation study of confidence-aware late fusion under different bandwidths on the OPV2V [36] dataset. The table shows the impact of late fusion on accuracy and bandwidth with different values of " + }, + { + "bbox": [ + 313, + 293, + 555, + 337 + ], + "type": "inline_equation", + "content": "\\varepsilon_{c}" + }, + { + "bbox": [ + 313, + 293, + 555, + 337 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 358, + 555, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 358, + 555, + 478 + ], + "spans": [ + { + "bbox": [ + 313, + 358, + 555, + 478 + ], + "type": "text", + "content": "cant. Table 3 shows the accuracy improvement due to late collaboration under different bandwidth conditions. It can be observed that late collaboration provides more accuracy improvements under lower bandwidth conditions. For example, when the bandwidth used for intermediate fusion is 0.05 Mbps, late collaboration with a bandwidth cost of 0.14 Mbps leads to a " + }, + { + "bbox": [ + 313, + 358, + 555, + 478 + ], + "type": "inline_equation", + "content": "3\\%" + }, + { + "bbox": [ + 313, + 358, + 555, + 478 + ], + "type": "text", + "content": " improvement in AP@0.5 and a " + }, + { + "bbox": [ + 313, + 358, + 555, + 478 + ], + "type": "inline_equation", + "content": "4\\%" + }, + { + "bbox": [ + 313, + 358, + 555, + 478 + ], + "type": "text", + "content": " improvement in AP@0.7. This is one of the key reasons why our method achieves significantly higher detection accuracy under lower bandwidth compared to Where2comm [11]." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 493, + 388, + 504 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 493, + 388, + 504 + ], + "spans": [ + { + "bbox": [ + 313, + 493, + 388, + 504 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 512, + 555, + 656 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 512, + 555, + 656 + ], + "spans": [ + { + "bbox": [ + 313, + 512, + 555, + 656 + ], + "type": "text", + "content": "In this paper, we propose CoSDH, a novel communication-efficient collaborative perception framework for 3D object detection. By finely modeling the supply-demand relationship between agents, it selects key and sparse regions for collaboration. Additionally, we innovatively incorporate confidence-aware late fusion on top of intermediate collaboration to form an intermediate-late hybrid collaborative perception method. Experiments on multiple datasets show that our method offers a better accuracy-bandwidth trade-off, demonstrating outstanding accuracy under bandwidth constraints close to real-world communication limits, making it highly valuable for practical applications." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 670, + 407, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 670, + 407, + 684 + ], + "spans": [ + { + "bbox": [ + 313, + 670, + 407, + 684 + ], + "type": "text", + "content": "Acknowledgment" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "content": "This work is supported by the National Key Research and Development Plan (2024YFB3309302)." + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "6841" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 87, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 87, + 295, + 108 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 87, + 295, + 108 + ], + "spans": [ + { + "bbox": [ + 61, + 87, + 295, + 108 + ], + "type": "text", + "content": "[1] Fabio Arena and Giovanni Pau. An overview of vehicular communications. Future internet, 11(2):27, 2019. 1, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 111, + 295, + 165 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 111, + 295, + 165 + ], + "spans": [ + { + "bbox": [ + 61, + 111, + 295, + 165 + ], + "type": "text", + "content": "[2] Eduardo Arnold, Mehrdad Dianati, Robert de Temple, and Saber Fallah. Cooperative perception for 3d object detection in driving scenarios using infrastructure sensors. IEEE Transactions on Intelligent Transportation Systems, 23(3): 1852-1864, 2020. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 167, + 295, + 223 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 167, + 295, + 223 + ], + "spans": [ + { + "bbox": [ + 62, + 167, + 295, + 223 + ], + "type": "text", + "content": "[3] Qi Chen, Xu Ma, Sihai Tang, Jingda Guo, Qing Yang, and Song Fu. F-cooper: Feature based cooperative perception for autonomous vehicle edge computing system using 3d point clouds. In Proceedings of the 4th ACM/IEEE Symposium on Edge Computing, pages 88-100, 2019. 1, 3, 6, 7, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 224, + 295, + 280 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 224, + 295, + 280 + ], + "spans": [ + { + "bbox": [ + 62, + 224, + 295, + 280 + ], + "type": "text", + "content": "[4] Qi Chen, Sihai Tang, Qing Yang, and Song Fu. Cooper: Cooperative perception for connected autonomous vehicles based on 3d point clouds. In 2019 IEEE 39th International Conference on Distributed Computing Systems (ICDCS), pages 514-524. IEEE, 2019. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 281, + 295, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 281, + 295, + 314 + ], + "spans": [ + { + "bbox": [ + 62, + 281, + 295, + 314 + ], + "type": "text", + "content": "[5] Yilun Chen, Shu Liu, Xiaoyong Shen, and Jiaya Jia. Fast point r-cnn. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9775-9784, 2019. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 316, + 295, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 316, + 295, + 360 + ], + "spans": [ + { + "bbox": [ + 62, + 316, + 295, + 360 + ], + "type": "text", + "content": "[6] Yilun Chen, Shu Liu, Xiaoyong Shen, and Jiaya Jia. Dsgn: Deep stereo geometry network for 3d object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12536-12545, 2020. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 362, + 295, + 417 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 362, + 295, + 417 + ], + "spans": [ + { + "bbox": [ + 62, + 362, + 295, + 417 + ], + "type": "text", + "content": "[7] Jiajun Deng, Shaoshuai Shi, Peiwei Li, Wengang Zhou, Yanyong Zhang, and Houqiang Li. Voxel r-cnn: Towards high performance voxel-based 3d object detection. In Proceedings of the AAAI conference on artificial intelligence, pages 1201-1209, 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 418, + 295, + 461 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 418, + 295, + 461 + ], + "spans": [ + { + "bbox": [ + 62, + 418, + 295, + 461 + ], + "type": "text", + "content": "[8] Alexey Dosovitskiy, German Ros, Felipe Codevilla, Antonio Lopez, and Vladlen Koltun. Carla: An open urban driving simulator. In Conference on robot learning, pages 1-16. PMLR, 2017. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 464, + 295, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 464, + 295, + 518 + ], + "spans": [ + { + "bbox": [ + 62, + 464, + 295, + 518 + ], + "type": "text", + "content": "[9] Sebastian Gräfling, Petri Mähönen, and Janne Riihijärvi. Performance evaluation of the IEEE 1609 wave and the IEEE 802.11 p for vehicular communications. In 2010 second international conference on ubiquitous and future networks (ICUFN), pages 344-348. IEEE, 2010. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 521, + 295, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 521, + 295, + 576 + ], + "spans": [ + { + "bbox": [ + 57, + 521, + 295, + 576 + ], + "type": "text", + "content": "[10] Jingda Guo, Dominic Carrillo, Sihai Tang, Qi Chen, Qing Yang, Song Fu, Xi Wang, Nannan Wang, and Paparao Palacharla. Coff: Cooperative spatial feature fusion for 3-d object detection on autonomous vehicles. IEEE Internet of Things Journal, 8(14):11078-11087, 2021. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 578, + 295, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 578, + 295, + 632 + ], + "spans": [ + { + "bbox": [ + 57, + 578, + 295, + 632 + ], + "type": "text", + "content": "[11] Yue Hu, Shaoheng Fang, Zixing Lei, Yiqi Zhong, and Siheng Chen. Where2comm: Communication-efficient collaborative perception via spatial confidence maps. Advances in neural information processing systems, 35:4874-4886, 2022. 1, 2, 3, 4, 6, 7, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 635, + 295, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 635, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 57, + 635, + 295, + 689 + ], + "type": "text", + "content": "[12] Yue Hu, Juntong Peng, Sifei Liu, Junhao Ge, Si Liu, and Si-heng Chen. Communication-efficient collaborative perception via information filling with codebook. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15481-15490, 2024. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "type": "text", + "content": "[13] Zheng Jiang, Jinqing Zhang, Yanan Zhang, Qingjie Liu, Zhenghui Hu, Baohui Wang, and Yunhong Wang. Fsd-bev:" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "text", + "content": "Foreground self-distillation for multi-view 3d object detection. In European Conference on Computer Vision, pages 110-126. Springer, 2024. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 57, + 87, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 317, + 107, + 553, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 107, + 553, + 128 + ], + "spans": [ + { + "bbox": [ + 317, + 107, + 553, + 128 + ], + "type": "text", + "content": "[14] Diederik P Kingma. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 129, + 553, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 129, + 553, + 172 + ], + "spans": [ + { + "bbox": [ + 317, + 129, + 553, + 172 + ], + "type": "text", + "content": "[15] Daniel Krajzewicz, Jakob Erdmann, Michael Behrisch, and Laura Bieker. Recent development and applications of sumosimulation of urban mobility. International journal on advances in systems and measurements, 5(3&4), 2012. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 174, + 553, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 174, + 553, + 227 + ], + "spans": [ + { + "bbox": [ + 317, + 174, + 553, + 227 + ], + "type": "text", + "content": "[16] Jason Ku, Alex D Pon, and Steven L Waslander. Monocular 3d object detection leveraging accurate proposals and shape reconstruction. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11867-11876, 2019. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 228, + 555, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 228, + 555, + 282 + ], + "spans": [ + { + "bbox": [ + 317, + 228, + 555, + 282 + ], + "type": "text", + "content": "[17] Alex H Lang, Sourabh Vora, Holger Caesar, Lubing Zhou, Jiong Yang, and Oscar Beijbom. Pointpillars: Fast encoders for object detection from point clouds. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12697-12705, 2019. 3, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 284, + 553, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 284, + 553, + 327 + ], + "spans": [ + { + "bbox": [ + 316, + 284, + 553, + 327 + ], + "type": "text", + "content": "[18] Peiliang Li, Xiaozhi Chen, and Shaojie Shen. Stereo r-cnn based 3d object detection for autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7644-7652, 2019. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 328, + 553, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 328, + 553, + 371 + ], + "spans": [ + { + "bbox": [ + 316, + 328, + 553, + 371 + ], + "type": "text", + "content": "[19] Yiming Li, Shunli Ren, Pengxiang Wu, Siheng Chen, Chen Feng, and Wenjun Zhang. Learning distilled collaboration graph for multi-agent perception. Advances in Neural Information Processing Systems, 34:29541-29552, 2021. 3, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 373, + 553, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 373, + 553, + 425 + ], + "spans": [ + { + "bbox": [ + 316, + 373, + 553, + 425 + ], + "type": "text", + "content": "[20] Yiming Li, Dekun Ma, Ziyan An, Zixun Wang, Yiqi Zhong, Siheng Chen, and Chen Feng. V2x-sim: Multi-agent collaborative perception dataset and benchmark for autonomous driving. IEEE Robotics and Automation Letters, 7(4):10914-10921, 2022. 1, 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 426, + 553, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 426, + 553, + 480 + ], + "spans": [ + { + "bbox": [ + 316, + 426, + 553, + 480 + ], + "type": "text", + "content": "[21] Yen-Cheng Liu, Junjiao Tian, Nathaniel Glaser, and Zsolt Kira. When2com: Multi-agent perception via communication graph grouping. In Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition, pages 4106-4115, 2020. 3, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 482, + 553, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 482, + 553, + 536 + ], + "spans": [ + { + "bbox": [ + 317, + 482, + 553, + 536 + ], + "type": "text", + "content": "[22] Yen-Cheng Liu, Junjiao Tian, Chih-Yao Ma, Nathan Glaser, Chia-Wen Kuo, and Zsolt Kira. Who2com: Collaborative perception via learnable handshake communication. In 2020 IEEE International Conference on Robotics and Automation (ICRA), pages 6876-6883. IEEE, 2020. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 537, + 553, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 537, + 553, + 592 + ], + "spans": [ + { + "bbox": [ + 317, + 537, + 553, + 592 + ], + "type": "text", + "content": "[23] Zhijian Liu, Haotian Tang, Alexander Amini, Xinyu Yang, Huizi Mao, Daniela L Rus, and Song Han. Bevfusion: Multi-task multi-sensor fusion with unified bird's-eye view representation. In 2023 IEEE international conference on robotics and automation (ICRA), pages 2774-2781. IEEE, 2023. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 593, + 553, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 593, + 553, + 635 + ], + "spans": [ + { + "bbox": [ + 317, + 593, + 553, + 635 + ], + "type": "text", + "content": "[24] Jose Manuel Lozano Dominguez and Tomas Jesus Mateo Sanguino. Review on v2x, i2x, and p2x communications and their applications: a comprehensive analysis over time. Sensors, 19(12):2756, 2019. 1, 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 636, + 553, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 636, + 553, + 691 + ], + "spans": [ + { + "bbox": [ + 317, + 636, + 553, + 691 + ], + "type": "text", + "content": "[25] Yifan Lu, Quanhao Li, Baoan Liu, Mehrdad Dianati, Chen Feng, Siheng Chen, and Yanfeng Wang. Robust collaborative 3d object detection in presence of pose errors. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 4812-4818. IEEE, 2023. 5, 6, 7, 8, 1, 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "text", + "content": "[26] Jiageng Mao, Shaoshuai Shi, Xiaogang Wang, and Hongsheng Li. 3d object detection for autonomous driving: A" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6842" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 77, + 73, + 296, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 73, + 296, + 95 + ], + "spans": [ + { + "bbox": [ + 77, + 73, + 296, + 95 + ], + "type": "text", + "content": "comprehensive survey. International Journal of Computer Vision, 131(8):1909-1963, 2023. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 296, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 296, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 296, + 140 + ], + "type": "text", + "content": "[27] Xuran Pan, Zhuofan Xia, Shiji Song, Li Erran Li, and Gao Huang. 3d object detection with pointformer. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7463-7472, 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 142, + 294, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 142, + 294, + 194 + ], + "spans": [ + { + "bbox": [ + 56, + 142, + 294, + 194 + ], + "type": "text", + "content": "[28] Shaoshuai Shi, Xiaogang Wang, and Hongsheng Li. Pointrcnn: 3d object proposal generation and detection from point cloud. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 770-779, 2019. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 198, + 294, + 253 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 198, + 294, + 253 + ], + "spans": [ + { + "bbox": [ + 56, + 198, + 294, + 253 + ], + "type": "text", + "content": "[29] Shaoshuai Shi, Chaoxu Guo, Li Jiang, Zhe Wang, Jianping Shi, Xiaogang Wang, and Hongsheng Li. Pv-rcnn: Pointvoxel feature set abstraction for 3d object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10529–10538, 2020. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 255, + 294, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 255, + 294, + 308 + ], + "spans": [ + { + "bbox": [ + 56, + 255, + 294, + 308 + ], + "type": "text", + "content": "[30] Shaoshuai Shi, Zhe Wang, Jianping Shi, Xiaogang Wang, and Hongsheng Li. From points to parts: 3d object detection from point cloud with part-aware and part-aggregation network. IEEE transactions on pattern analysis and machine intelligence, 43(8):2647-2664, 2020. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 310, + 294, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 310, + 294, + 376 + ], + "spans": [ + { + "bbox": [ + 56, + 310, + 294, + 376 + ], + "type": "text", + "content": "[31] Rui Song, Chenwei Liang, Hu Cao, Zhiran Yan, Walter Zimmer, Markus Gross, Andreas Festag, and Alois Knoll. Collaborative semantic occupancy prediction with hybrid feature fusion in connected automated vehicles. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17996-18006, 2024. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 378, + 294, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 378, + 294, + 443 + ], + "spans": [ + { + "bbox": [ + 56, + 378, + 294, + 443 + ], + "type": "text", + "content": "[32] Tsun-Hsuan Wang, Sivabalan Manivasagam, Ming Liang, Bin Yang, Wenyuan Zeng, and Raquel Urtasun. V2vnet: Vehicle-to-vehicle communication for joint perception and prediction. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part II 16, pages 605–621. Springer, 2020. 3, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 445, + 294, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 445, + 294, + 510 + ], + "spans": [ + { + "bbox": [ + 56, + 445, + 294, + 510 + ], + "type": "text", + "content": "[33] Yan Wang, Wei-Lun Chao, Divyansh Garg, Bharath Hariharan, Mark Campbell, and Kilian Q Weinberger. Pseudolidar from visual depth estimation: Bridging the gap in 3d object detection for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8445-8453, 2019. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 512, + 294, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 512, + 294, + 567 + ], + "spans": [ + { + "bbox": [ + 56, + 512, + 294, + 567 + ], + "type": "text", + "content": "[34] Runsheng Xu, Yi Guo, Xu Han, Xin Xia, Hao Xiang, and Jiaqi Ma. Opencda: an open cooperative driving automation framework integrated with co-simulation. In 2021 IEEE International Intelligent Transportation Systems Conference (ITSC), pages 1155-1162. IEEE, 2021. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 568, + 294, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 568, + 294, + 622 + ], + "spans": [ + { + "bbox": [ + 56, + 568, + 294, + 622 + ], + "type": "text", + "content": "[35] Runsheng Xu, Hao Xiang, Zhengzhong Tu, Xin Xia, Ming-Hsuan Yang, and Jiaqi Ma. V2x-vit: Vehicle-to-everything cooperative perception with vision transformer. In European conference on computer vision, pages 107-124. Springer, 2022. 1, 3, 5, 6, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 624, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 294, + 689 + ], + "type": "text", + "content": "[36] Runsheng Xu, Hao Xiang, Xin Xia, Xu Han, Jinlong Li, and Jiaqi Ma. Opv2v: An open benchmark dataset and fusion pipeline for perception with vehicle-to-vehicle communication. In 2022 International Conference on Robotics and Automation (ICRA), pages 2583–2589. IEEE, 2022. 1, 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "text", + "content": "[37] Runsheng Xu, Zhengzhong Tu, Hao Xiang, Wei Shao, Bolei Zhou, and Jiaqi Ma. Cobevt: Cooperative bird's eye view" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 105 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 105 + ], + "type": "text", + "content": "semantic segmentation with sparse transformers. In Conference on Robot Learning, pages 989-1000. PMLR, 2023. 1, 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 106, + 553, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 106, + 553, + 172 + ], + "spans": [ + { + "bbox": [ + 316, + 106, + 553, + 172 + ], + "type": "text", + "content": "[38] Runsheng Xu, Xin Xia, Jinlong Li, Hanzhao Li, Shuo Zhang, Zhengzhong Tu, Zonglin Meng, Hao Xiang, Xiaoyu Dong, Rui Song, et al. V2v4real: A real-world large-scale dataset for vehicle-to-vehicle cooperative perception. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13712–13722, 2023. 1, 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 173, + 553, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 173, + 553, + 194 + ], + "spans": [ + { + "bbox": [ + 316, + 173, + 553, + 194 + ], + "type": "text", + "content": "[39] Yan Yan, Yuxing Mao, and Bo Li. Second: Sparsely embedded convolutional detection. Sensors, 18(10):3337, 2018. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 195, + 553, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 195, + 553, + 249 + ], + "spans": [ + { + "bbox": [ + 316, + 195, + 553, + 249 + ], + "type": "text", + "content": "[40] Dingkang Yang, Kun Yang, Yuzheng Wang, Jing Liu, Zhi Xu, Rongbin Yin, Peng Zhai, and Lihua Zhang. How2comm: Communication-efficient and collaboration-pragmatic multi-agent perception. Advances in Neural Information Processing Systems, 36, 2024. 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 251, + 553, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 251, + 553, + 315 + ], + "spans": [ + { + "bbox": [ + 316, + 251, + 553, + 315 + ], + "type": "text", + "content": "[41] Junbo Yin, Jianbing Shen, Runnan Chen, Wei Li, Ruigang Yang, Pascal Frossard, and Wenguan Wang. Is-fusion: Instance-scene collaborative fusion for multimodal 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14905-14915, 2024. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 316, + 553, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 316, + 553, + 392 + ], + "spans": [ + { + "bbox": [ + 316, + 316, + 553, + 392 + ], + "type": "text", + "content": "[42] Haibao Yu, Yizhen Luo, Mao Shu, Yiyi Huo, Zebang Yang, Yifeng Shi, Zhenglong Guo, Hanyu Li, Xing Hu, Jirui Yuan, et al. Dair-v2x: A large-scale dataset for vehicleinfrastructure cooperative 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21361-21370, 2022. 2, 3, 5, 6, 7, 8, 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 394, + 553, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 394, + 553, + 469 + ], + "spans": [ + { + "bbox": [ + 316, + 394, + 553, + 469 + ], + "type": "text", + "content": "[43] Haibao Yu, Wenxian Yang, Hongzhi Ruan, Zhenwei Yang, Yingjuan Tang, Xu Gao, Xin Hao, Yifeng Shi, Yifeng Pan, Ning Sun, et al. V2x-seq: A large-scale sequential dataset for vehicle-infrastructure cooperative perception and forecasting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5486-5495, 2023. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 471, + 553, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 471, + 553, + 525 + ], + "spans": [ + { + "bbox": [ + 316, + 471, + 553, + 525 + ], + "type": "text", + "content": "[44] Jinqing Zhang, Yanan Zhang, Qingjie Liu, and Yunhong Wang. Sa-bev: Generating semantic-aware bird's-eye-view feature for multi-view 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3348-3357, 2023. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 526, + 553, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 526, + 553, + 569 + ], + "spans": [ + { + "bbox": [ + 316, + 526, + 553, + 569 + ], + "type": "text", + "content": "[45] Jinqing Zhang, Yanan Zhang, Yunlong Qi, Zehua Fu, Qingjie Liu, and Yunhong Wang. Geobev: Learning geometric bev representation for multi-view 3d object detection. arXiv preprint arXiv:2409.01816, 2024. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 571, + 553, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 571, + 553, + 613 + ], + "spans": [ + { + "bbox": [ + 316, + 571, + 553, + 613 + ], + "type": "text", + "content": "[46] Yanan Zhang, Di Huang, and Yunhong Wang. Pc-rgnn: Point cloud completion and graph neural network for 3d object detection. In Proceedings of the AAAI conference on artificial intelligence, pages 3430-3437, 2021. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 615, + 553, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 615, + 553, + 667 + ], + "spans": [ + { + "bbox": [ + 316, + 615, + 553, + 667 + ], + "type": "text", + "content": "[47] Yanan Zhang, Jiaxin Chen, and Di Huang. Cat-det: Contrastively augmented transformer for multi-modal 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 908-917, 2022. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "text", + "content": "[48] Chao Zhou, Yanan Zhang, Jiaxin Chen, and Di Huang. Octr: Octtree-based transformer for 3d object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5166-5175, 2023. 2" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6843" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/CoSER_ Towards Consistent Dense Multiview Text-to-Image Generator for 3D Creation/911def72-ec15-4b90-8a72-6946e2b20044_content_list.json b/2025/CoSER_ Towards Consistent Dense Multiview Text-to-Image Generator for 3D Creation/911def72-ec15-4b90-8a72-6946e2b20044_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..0a3f31edd4a63773a7a1738e60704ae2005235d8 --- /dev/null +++ b/2025/CoSER_ Towards Consistent Dense Multiview Text-to-Image Generator for 3D Creation/911def72-ec15-4b90-8a72-6946e2b20044_content_list.json @@ -0,0 +1,1514 @@ +[ + { + "type": "text", + "text": "CoSER: Towards Consistent Dense Multiview Text-to-Image Generator for 3D Creation", + "text_level": 1, + "bbox": [ + 194, + 130, + 805, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Bonan Li $^{1,2*}$ Zicheng Zhang $^{1*}$ Xingyi Yang $^{2}$ Xinchao Wang $^{2\\dagger}$", + "bbox": [ + 228, + 202, + 772, + 222 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ University of Chinese Academy of Sciences $^{2}$ National University of Singapore", + "bbox": [ + 227, + 222, + 771, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "libonacci@ucas.ac.cn zhangzicheng19@mails.ucas.ac.cn xyang@u.nus.edu xinchao@nus.edu.sg", + "bbox": [ + 158, + 239, + 833, + 257 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/708d74411269255e50a089ac20aad4efc83523fe4cbbf99bc2ae0b88c00a6d8f.jpg", + "image_caption": [ + "Figure 1. CoSER aims to generate detailed, diverse and detaild 3D objects from text prompts. All prompts are from T3Bench [13] and the templated is set as \"A DSLR photo of ..., 3d asset\". Better see in color and 2x zoom." + ], + "image_footnote": [], + "bbox": [ + 109, + 262, + 880, + 500 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 577, + 328, + 593 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Generating dense multiview images from text prompts is crucial for creating high-fidelity 3D assets. Nevertheless, existing methods struggle with space-view correspondences, resulting in sparse and low-quality outputs. In this paper, we introduce CoSER, a novel consistent dense Multiview Text-to-Image Generator for Text-to-3D, achieving both efficiency and quality by meticulously learning neighbor-view coherence and further alleviating ambiguity through the swift traversal of all views. For achieving neighbor-view consistency, each viewpoint densely interacts with adjacent viewpoints to perceive the global spatial structure, and aggregates information along motion paths explicitly defined by physical principles to refine details. To further enhance cross-view consistency and alleviate content drift, CoSER rapidly scan all views in spiral bidirectional manner to aware holistic information and then scores each point based on semantic material. Subsequently, we", + "bbox": [ + 88, + 609, + 485, + 867 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "conduct weighted down-sampling along the spatial dimension based on scores, thereby facilitating prominent information fusion across all views with lightweight computation. Technically, the core module is built by integrating the attention mechanism with a selective state space model, exploiting the robust learning capabilities of the former and the low overhead of the latter. Extensive evaluation shows that CoSER is capable of producing dense, high-fidelity, content-consistent multiview images that can be flexibly integrated into various 3D generation models.", + "bbox": [ + 511, + 579, + 908, + 731 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 766, + 645, + 782 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Creating 3D content from text is an important task in computer vision and graphics for its potential applications in advertising, robotics, digital games and the meta-verse. Previous works [3, 26, 33] primarily optimise 3D representations using score distillation sampling [38] of Text-to-Image (T2I) models, ensuring that rendered images from any viewpoint maintain high fidelity while aligning with", + "bbox": [ + 511, + 794, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution", + "bbox": [ + 107, + 875, + 217, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author", + "bbox": [ + 109, + 887, + 233, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "2880", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "the given prompt. Nevertheless, these methods suffer from Janus problems and severe artifacts with asymmetric geometry. As a remedy, multiview text-to-image (MT2I) models [6, 17, 19, 21, 23, 32] have been developed, providing rich 3D priors that help mitigate these problems. Furthermore, using dense views can accelerate the modeling process and enhance the quality of the resulting 3D models.", + "bbox": [ + 89, + 90, + 480, + 196 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To acquire MT2I models, a promising approach is to leverage 3D priors learned from 3D datasets to enhance 2D diffusion models. Seminal works [28, 47] have focused on inflating T2I models with dense cross-view attention to capture 3D information. However, this approach incurs expensive computation and faces challenges when dealing with dense viewpoints. To address these limitations, sparse attention mechanisms [23] have been proposed to model cross-view dependencies, though they can introduce inconsistencies across multiple views. Recent studies [9, 58, 74] have leveraged cross-frame continuity priors from text-to-video (T2V) models to enhance consistency. Nevertheless, azimuthal rotation around an object can cause significant content variations at the same image coordinates, making temporal attention insufficient for capturing intricate details necessary for 3D generation.", + "bbox": [ + 89, + 199, + 482, + 441 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Despite impressive progress in multiview T2I generation, effectively and efficiently modeling the space-view structure within a unified architecture remains an open challenge. In this paper, we advance the field by addressing two fundamental problems: (1) How to design efficient cross-view interactions to ensure dense multiview consistency? Rather than directly optimizing consistency among all views with uniform granularity, we propose a two-step approach: (i) Learning neighbor-view consistency as a prerequisite, as enforcing consistency between adjacent viewpoints naturally enables coherence across all viewpoints. (ii) Ensuring global-view coherence to mitigate discrepancies between distant views and facilitate unified appearance, since subtle inconsistencies are inevitable and deteriorated as the viewing angles increases. (2) Which operator should be applied for the implementation? Unlike previous works that employ attention as the default choice, we leverage hybrid operators, including attention [57] and state space models (SSMs) [12], to harness the benefits of both model families, e.g., the powerful capabilities of attention and the low computational cost of SSMs.", + "bbox": [ + 89, + 444, + 482, + 760 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we address these questions and propose a novel model named CoSER, which enhances consistency in dense MT2I tasks. CoSER utilizes \"factor\" interactions that focus on neighboring views and maintain awareness of the global view. Technically, CoSER initially employs dense attention on adjacent views to capture the rough structure and texture of the current plane, thereby constructing a foundational appearance. Subsequently, we introduce a trajectory-based module to refine local details by accurately tracing", + "bbox": [ + 89, + 763, + 482, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3D trajectories and enhancing point-to-point consistency. Specifically, motivated by the efficiency of calculating rotated locations for centralized objects, we develop a robust method to handle complicate cases. This involves pixel interaction within a window of neighboring frames defined by physical rotation rules, ensuring precise alignment and consistency across views. To further alleviate disparities across arbitrary viewpoints, we utilize state space models (SSMs) equipped with a novel spiral bidirectional scan strategy to aggregate features from the entire set of views. We then score each patch based on textual inputs and perform down-sampling conditioned on these scores to adaptively ignore semantically irrelevant information. This approach allows CoSER to efficiently maintain global context through sparse attention. With these designs, our method establishes robust priors, enabling the creation of exceptionally coherent 3D assets (see Figure 1). In summary, our contributions can be summarized as follows:", + "bbox": [ + 511, + 90, + 903, + 361 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a novel multiview text-to-image framework, CoSER, for 3D content creation which exhibits significantly consistency behavior.", + "- We explore an efficient pipeline to model complex long-span consistency by focusing on neighbors to achieve short-span strict consistency and knowing whole views to obtain cues that correct cumulative errors.", + "- We design an hybrid architecture to implement the interaction by adopting both attention and SSMs as operators, effectively capturing essential dependencies in short sequences while significantly reducing costs for modeling long sequences.", + "- Extensive experiments demonstrate that our CoSER outperforms the state-of-the-art multiview synthesis approaches in both quantitative and qualitative results." + ], + "bbox": [ + 513, + 364, + 903, + 590 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 608, + 653, + 625 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Text-to-3D Creation", + "text_level": 1, + "bbox": [ + 513, + 635, + 702, + 648 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recently, Text-to-3D generative models [4, 8, 15, 18, 35, 38] have made substantial advancements in computer vision [1, 2, 20, 25, 30, 41, 48–51, 64–67, 70, 72], achieving high-quality generation of both 3D objects and complex 3D scenes. Early works focus on directly generating assets using 3D diffusion models based on differentiable representations [7, 36, 61]. However, these approaches often relies on massive 3D data and incurs substantial training costs. Thanks to powerful text-to-image models [44], DreamFusion [38] leverages the robust 2D prior by introducing Score Distillation Sampling, which minimizes the discrepancy between rendered images from the underlying 3D assets and the diffusion prior. This development also sparked a surge of interest in other generative tasks such as texture generation [56, 71], editing [5, 37, 52, 62, 73], and Image-to-3D [39, 53, 59]. Despite this paradigm has bridged the", + "bbox": [ + 511, + 657, + 906, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2881", + "bbox": [ + 482, + 945, + 511, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "gap between imagination and reality, it still faces significant challenges in terms of optimization speed [16, 68], content drift [24, 63] and the Janus problem [31, 46, 69]. In response to inefficient training, Metzer et al. [34] train a NeRF within the latent space of LDM [44] to accelerate the learning process and further introduce the Sketch-Shape method to guide the generation process. To acquire photorealistic appearance, Lin et al. [26] employ a coarse-to-fine strategy, first optimizing a NeRF using a low-resolution diffusion prior, then refining the texture with latent diffusion priors using the DMTet initialized with the coarse NeRF. Chen et al. [3] disentangle the learning of geometry and material, leveraging physics-based rendering techniques to achieve high-fidelity mesh generation. Another line of work lies in generating 3D assets directly through a 3D diffusion model based on differentiable representations. [28] fine-tune image diffusion models to generate multiview image-normal pairs with switch attention and employ NeuS [60] for reconstruction underlying geometries. In this work, we will continue to improve the quality of 3D content by designing the effective multiview text-to-image generator, which provides consistent images for reconstruction method.", + "bbox": [ + 89, + 90, + 483, + 438 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Multiview Image Generation", + "text_level": 1, + "bbox": [ + 89, + 454, + 349, + 469 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Due to the scarcity of 3D data and the formidable training costs, lifting 2D diffusion to multiview generator conditioned on camera pose has become the prevailing paradigm [6, 17, 21, 23, 28, 32, 47, 55]. Tang et al. [55] introduce the novel concept of generating multiview images concurrently with correspondence-aware attention mechanisms, thereby enabling effective cross-view information interaction and application to texturing scene meshes. Meanwhile, Shi et al. [47] and Long et al. [28] simply extends 2D spatial attention to capture 3D structure, yielding remarkable results. Nevertheless, the quadratic complexity along with resolution restricts the number of cameras, thereby compromising the performance of reconstruction algorithms and resulting in inferior results. To reduce computational costs, Li et al. [23] propose an efficient rowwise attention instead of dense interactions, yet struggles to correct local details referencing global information. Latter works [9, 58, 74] capitalized on priors derived from video diffusion model to achieve both dense view generation and consistent outputs. Despite these strides, a fundamental disparity persists between the video continuity and 3D consistency. General videos typically depict objects moving along a fixed plane from the same camera viewpoint, i.e., similar information shifts across different positions on the canvas. In contrast, videos composed of images captured from multiple views often exhibit significant rotation, leading to substantial observable content jitter between frames. Here, we tailor an innovative strategy for cross-view interaction that", + "bbox": [ + 89, + 477, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "optimizes both performance and efficiency, ensuring robust consistency in generating dense-view images for 3D applications.", + "bbox": [ + 511, + 90, + 906, + 136 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 511, + 148, + 604, + 165 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminaries: Latent Diffusion Models", + "text_level": 1, + "bbox": [ + 511, + 174, + 852, + 189 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our work lifts the Latent Diffusion Model (LDM) [44], which maps a noise latent $\\epsilon$ and a text prompt $p$ to an image $x$ . LDM first encodes the image $x_0$ into latent space using a pre-trained encoder $\\mathcal{E}$ , i.e., $z_0 = \\mathcal{E}(x_0)$ . The latent representation $z_0$ are then decoded back into the image $x_0$ using a pre-trained decoder $\\mathcal{D}$ . The model is trained in the latent space in a self-supervised manner, and generates samples by reversing a Markov forward process. Specifically, the forward process injects noise into $z_0$ to obtain noisy inputs $\\{z_t\\}_{t=1}^T$ , where $z_t = \\sqrt{\\alpha^t z_0 + \\sqrt{1 - \\alpha_t}} \\epsilon$ , and $\\epsilon \\sim N(0, I)$ . The generation process is formulated as a sequence of denoising operations $\\epsilon_\\theta(z_t, t, y)$ , predicting a denoised variant of input $z_t$ , where $y = \\phi(p)$ is the condition, and $\\phi$ is a frozen text encoder [42]. The learnable parameter $\\theta$ is typically embedded within structures such as convolutional layers, self-attention mechanisms, and cross-attention mechanisms. The training objective is defined as follows:", + "bbox": [ + 511, + 196, + 906, + 465 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\theta} \\mathbb {E} _ {z _ {0}, \\epsilon , t} | | \\epsilon - \\epsilon_ {\\theta} (z _ {t}, t, y) | | _ {2} ^ {2}. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 607, + 467, + 906, + 489 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The reverse process ensures that the final sample $z_0$ follows the distribution of latents, while aligning the generated image with the given condition.", + "bbox": [ + 511, + 494, + 906, + 540 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Framework of CoSER", + "text_level": 1, + "bbox": [ + 511, + 549, + 720, + 564 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We propose CoSER, a method for generating dense and consistent multiview images based on text prompts. The inputs to CoSER include a set of noisy images $\\{x_{t}^{i}\\}_{i = 1}^{f}$ , a set of camera parameters $\\{c^i\\}_{i = 1}^f$ with a fixed elevation angle, uniformly distributed azimuth angles ranging from $0^{\\circ}$ to $360^{\\circ}$ , and a text prompt $y$ . We fine-tune the lifted Latent Diffusion Model (LDM) on $f$ views using a 3D dataset to synthesize multiview images $\\{x_0^i\\}_{i = 1}^f$ . It is non-trivial to incorporate the LDM architecture originally designed for 2D with 3D awareness. Rather than applying space-view attention or separately fusing spatial and viewpoint dimensions, our framework employs fine-grained interactions for neighboring views and coarse interactions for the global appearance. We retain the LDM structure, including ResBlock and CrossAttention, to maximize parameter reuse and maintain robust generative capabilities. A detailed sketch of CoSER is illustrated in Figure 2 and see more details for each module in Figure 3.", + "bbox": [ + 511, + 571, + 908, + 844 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2.1. Focus on Neighbors with Contextual Attention", + "text_level": 1, + "bbox": [ + 511, + 851, + 883, + 866 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "It is well known that 3D objects exhibit strict cross-view consistency, whereas this does not necessitate considering", + "bbox": [ + 511, + 869, + 906, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "2882", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/cf79c523f5583c58a4947ac6c9a5073d82a9993bd5518a01c0d14dcfdf853384.jpg", + "image_caption": [ + "Figure 2. Illustration of our CoSER. Given images rendered from 12 views at the same elevation, we take a pre-trained text-to-image generation model and fine-tune it by incorporating camera poses and lifting 2D-UNet to generate multi-view images. Specifically, we achieve 3D perception by employing fine-grained learning for neighbors viewpoints and coarse-grained interactions across whole viewpoints. For neighbors, Appearance Awareness (AA) is used to learn the basic appearance and Detail Refinement (DR) is proposed to ensure neighbor consistency in point-level. For whole views, we quickly scan all viewpoints with Rapid Glance (RG) and then eliminate ambiguity with powerful Accumulated Inconsistency Rectification (AIR)." + ], + "image_footnote": [], + "bbox": [ + 91, + 85, + 903, + 244 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the relationships among all views. For instance, the front and back views of an object are usually highly independent. Since adjacent viewpoints share more information and contribute significantly to visual consistency, we propose interacting each view with its context to mitigate the computational expense of space-view attention. Specifically, we use adjacent attention to capture the fundamental appearance and trajectory attention to refine details point-to-point, ensuring consistency between neighboring viewpoints. Contextual attention, in this case, refers to the mechanism of incorporating information from neighboring views to enhance the understanding of the current view.", + "bbox": [ + 88, + 353, + 480, + 532 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Appearance Awareness CoSER begins by integrating content from adjacent viewpoints to outline the basic appearance from the current view, thereby ensuring fundamental neighbor-view consistency. Following the MT2I baseline [47], we encode images $\\{x^i\\}$ to latent feature $\\{z^i\\}$ by the pre-trained encoder $\\mathcal{E}$ ( $t$ is omitted for clarity), followed by mapping camera parameters $\\{c^i\\}$ as high-dimensional condition. Moreover, the original self-attention is replaced with adjacent-attention, which takes the previous view $z^{i-1}$ , current view $z^i$ and next view $z^{i+1}$ as inputs, updating the features of $z^i$ . The formulation of the adjacent-attention is as follows:", + "bbox": [ + 88, + 535, + 482, + 715 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nQ ^ {i} = W _ {Q} z ^ {i}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 102, + 727, + 482, + 752 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nK ^ {i} = W _ {K} \\left[ z ^ {i - 1}, z ^ {i}, z ^ {i + 1} \\right], V ^ {i} = W _ {V} \\left[ z ^ {i - 1}, z ^ {i}, z ^ {i + 1} \\right],\n$$\n", + "text_format": "latex", + "bbox": [ + 102, + 750, + 450, + 763 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $[\\cdot]$ denotes concatenation operation. To inherit the generalizability of the 2D diffusion models, the parameters of self-attention is used to initialize projection matrices $W_{Q}, W_{K}, W_{V}$ which are shared across space and view. Note that adjacent-frame attention, while capable of forming the initial appearance of the object, inevitably disperses attention due to the involvement of numerous interacting patches, making it struggle to focus on details.", + "bbox": [ + 89, + 779, + 482, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Detail Refinement To ensure cross-view consistency in rendered images, it is crucial that the same position of an object corresponds to different pixels in neighboring viewpoints. These co-located pixels across views exhibit stronger correlations than others, playing a key role in maintaining visual consistency in multiview synthesis. For an object centered at the origin, and with some slight abuse of symbols, rigid rotation about the $y$ -axis follows a physical principle:", + "bbox": [ + 511, + 353, + 906, + 474 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nx ^ {\\prime} = \\left(x - \\frac {W}{2}\\right) \\cos \\Delta \\alpha + \\frac {W}{2} - d \\sin \\Delta \\alpha , \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 565, + 483, + 903, + 513 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $(x,y)$ and $(x',y')$ are the pixel coordinates before and after rotation by an angle $\\Delta \\alpha$ at a depth $d$ , and $W$ is the image width. Since the depth $d$ is unknown in the generative task and $\\sin \\Delta \\alpha$ becomes negligible when $\\Delta \\alpha$ is small, we simplify the transformation to:", + "bbox": [ + 511, + 523, + 905, + 599 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nx ^ {\\prime} = \\left(x - \\frac {W}{2}\\right) \\cos \\Delta \\alpha + \\frac {W}{2}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 607, + 608, + 903, + 638 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To make this simplification robust in practical applications, we consider a $3 \\times 3$ window $\\{x' + i, y' + j\\}_{i,j = -1,0,1}$ in neighboring frames as potential results. The feature at pixel $(x,y)$ is then refined through the attention mechanism that interacts with these neighboring pixels, enhancing consistency across views.", + "bbox": [ + 511, + 647, + 905, + 737 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.2. Know the Whole with Spiral Mamba", + "text_level": 1, + "bbox": [ + 511, + 746, + 816, + 760 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To efficiently enhance global consistency, we first quickly scan all viewpoints, then down-sample patches based on textual relevance to filter out irrelevant information. This allows us to utilize powerful attention mechanisms to correct any accumulated inconsistencies.", + "bbox": [ + 511, + 763, + 905, + 839 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Rapid Glance Exchanging information across all viewpoints is crucial for achieving global view consistency. However, dense attention mechanisms [28, 47] can be computationally expensive and memory-intensive, while simple", + "bbox": [ + 511, + 839, + 905, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "2883", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/9d96ae5ad7d78b09a813b424feea8010546057229a895255eaa82fdb19b2d352.jpg", + "image_caption": [ + "Appearance Awareness", + "View i-1" + ], + "image_footnote": [], + "bbox": [ + 94, + 107, + 205, + 191 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/a2b205d773e630c2d53d90ec172321ba039a5c46a27cb0bde5a138f7af44d043.jpg", + "image_caption": [ + "View i" + ], + "image_footnote": [], + "bbox": [ + 230, + 107, + 344, + 191 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/061f00139e0655ec33ca40367068278e9e7501b50c2d6eb9f221ee62b2e6ba0e.jpg", + "image_caption": [ + "View i+1" + ], + "image_footnote": [], + "bbox": [ + 370, + 107, + 480, + 191 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/59c491c3de51fc3fb83723147d06035d4b8c4fdab6029148a30c15839bdb4e88.jpg", + "image_caption": [ + "Detail Refinement", + "View i-1" + ], + "image_footnote": [], + "bbox": [ + 96, + 227, + 197, + 308 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/a68fa2d9f9d5ab7c34b28ead3de6296b21df59e0c88d013b15a96ccd2406d191.jpg", + "image_caption": [ + "View i" + ], + "image_footnote": [], + "bbox": [ + 238, + 228, + 336, + 305 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/7231786a09ce8cd24ff97982134969ef61dcd73da76b10e6f23b63ac4f5a30fc.jpg", + "image_caption": [ + "View $i + 1$" + ], + "image_footnote": [], + "bbox": [ + 375, + 228, + 478, + 308 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/4043873aa485d1af714d00c970500fc1b961acb8881dac2af78ae5a77ee5ed2e.jpg", + "image_caption": [ + "View 1" + ], + "image_footnote": [], + "bbox": [ + 94, + 343, + 205, + 428 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/7e8afe8b0fdc20d36dc84ef32c47d75001d952aff5f41ebfeec549964911df24.jpg", + "image_caption": [ + "Rapid Glance", + "View i" + ], + "image_footnote": [], + "bbox": [ + 238, + 345, + 339, + 422 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/4c171c536dea9b6845f135b1b56825924026d1fd254b4700ec6c49c9fb38233c.jpg", + "image_caption": [ + "View 12" + ], + "image_footnote": [], + "bbox": [ + 370, + 343, + 480, + 428 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/3b79589b275602f7cb200b282006199e24b54951ac0c64644658e4ca1501c048.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 370, + 445, + 480, + 460 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/4ffb6b32815ce9c6540fd786f80e825227046e5706aca771804401ce64cf98a7.jpg", + "image_caption": [ + "Accumulated", + "View 1", + "Figure 3. Visualization of the interaction schemes in CoSER. For illustration, we denote each gird as a token with target token in red and attending token in green." + ], + "image_footnote": [], + "bbox": [ + 94, + 463, + 205, + 549 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/51a47692d1371df9144cfd06043ced87e24c1e308f211f22c40ac3c360fb2391.jpg", + "image_caption": [ + "Inconsistency", + "View i" + ], + "image_footnote": [], + "bbox": [ + 230, + 463, + 344, + 546 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/fea03fc6ce8f2e4446e382adb703622e9bbd969018338dd31cfaf2567c11cd62.jpg", + "image_caption": [ + "View 12" + ], + "image_footnote": [], + "bbox": [ + 369, + 463, + 480, + 549 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "temporal attention [74] fails to effectively learn multi-view consistency. Inspired by Gu et al. [12], we leverage state space models (SSMs) known as linear operators to process all patches of multiview images, offering a more efficient alternative to quadratic-time attention. Specifically, we utilize the advanced selective scan SSMs introduced by Mamba [11] to implement our core SSMs operator. Unlike existing Vision Mamba methods that flatten 2D images into 1D sequences row-wise or column-wise, we propose an effective spiral bidirectional scan strategy. This strategy leverages the prior knowledge that objects in 3D datasets are typically centered. By organizing object-related content within a short sequence window, our approach implicitly mitigates the forgetting problem of SSMs. Additionally, placing background content across distinct viewpoints aids the SSMs in boundary perception.", + "bbox": [ + 89, + 643, + 482, + 883 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Technically, our method organizes spatial tokens start", + "bbox": [ + 109, + 885, + 482, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "ing from the center and expanding in a spiral pattern, then stacks them view by view. Meanwhile, we reverse the scan order within the arrangement of viewpoints to capture robust space-view features. This allows us to perform a rapid glance as follows:", + "bbox": [ + 511, + 90, + 905, + 167 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{z ^ {i} \\right\\} _ {i = 1} ^ {f} = S S M s \\left(S B S c a n \\left(\\left\\{z ^ {i} \\right\\} _ {i = 1} ^ {f}\\right)\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 575, + 175, + 905, + 195 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $SBScan(\\cdot)$ denotes the spiral bidirectional scan. This operator effectively aggregates global information and provides crucial clues for further disambiguation.", + "bbox": [ + 511, + 205, + 905, + 250 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Accumulated Inconsistency Rectification To facilitate the elimination of inconsistencies through attention mechanisms, we implement down-sampling to filter out background and retain features relevant to textual information, thereby substantially reducing the number of patches engaged in interactions. Specifically, we provide a high-level semantic embedding $\\phi(p)$ of the prompt $p$ . A 2-layer MLP $S$ is employed to assess the relevance between each patch and the text, producing a score ranging from 0 to 1 at each position, where a higher score indicates stronger relevance. Notably, we do not explicitly impose supervised constraints on the scoring mechanism. Instead, the model is driven by data to autonomously perceive the regions of interest. Based on the score map $s^i = S([z^i, \\phi(p)])$ as a weight, we perform 2D-pooling along the spatial dimensions on the query, key, and value maps to enable sparse interactions:", + "bbox": [ + 511, + 251, + 906, + 492 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {Q} ^ {i} = P \\left(s ^ {i} \\cdot \\bar {Q} ^ {i}; \\tau\\right),\n$$\n", + "text_format": "latex", + "bbox": [ + 620, + 502, + 754, + 521 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {K} ^ {i} = P \\left(\\left\\{s ^ {j} \\cdot \\bar {K} ^ {i} \\right\\} _ {j = 1} ^ {f}; \\rho\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 620, + 523, + 903, + 544 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {V} ^ {i} = P (\\{s ^ {j} \\cdot \\bar {V} ^ {i} \\} _ {j = 1} ^ {f}; \\rho).\n$$\n", + "text_format": "latex", + "bbox": [ + 622, + 547, + 795, + 566 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here, $P$ denotes the pooling operator, $\\tau$ and $\\rho$ are the strides required for pooling, and the transformations include: $\\bar{Q}^i = W_{\\bar{Q}}z^i$ , $\\bar{K}^i = W_{\\bar{K}}[\\{z^j\\}_{j=1}^f]$ , and $\\bar{V}^j = W_{\\bar{V}}[\\{z^j\\}_{j=1}^f]$ . Notably, in contrast to the intensive down-sampling of $\\hat{V}^i$ and $\\hat{K}^i$ , we maintain a higher resolution for the query feature map $\\hat{Q}^i$ to minimize the loss of content information. Finally, we up-sample the computed feature map using bilinear interpolation to restore its resolution.", + "bbox": [ + 511, + 575, + 905, + 704 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 715, + 643, + 733 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we first provide the implementation details and the dataset. We then conduct both qualitative and quantitative experiments to compare our method with previous state-of-the-art text-to-3D generation methods. Additionally, we perform an ablation study to demonstrate the effectiveness of the proposed modules.", + "bbox": [ + 511, + 742, + 905, + 830 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation We adopt the LDM $^1$ [44], pre-trained on the LAION-5B [45], as our base model. Specifically, we use the AdamW optimizer [29] with a learning rate of 3e-5 and a", + "bbox": [ + 511, + 832, + 905, + 878 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "1https://github.com/CompVis/stable-diffusion", + "bbox": [ + 529, + 887, + 879, + 898 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "2884", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/c925040cdf9122e8ea46408d7de0ca87c5b8f642525e739f37365747278a2596.jpg", + "image_caption": [ + "A bear dressed in medieval armor" + ], + "image_footnote": [], + "bbox": [ + 96, + 94, + 897, + 460 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d82548c0429c60dbbb23905b38ecb44e5be752d9b352cdd385755f41fd1d0b99.jpg", + "image_caption": [ + "Figure 4. Qualitative comparison of VideoMV [74] (Up) and our CoSER (Down).", + "Figure 5. Qualitative comparisons of GaussianDreamer [69], Hash3D [68] and our CoSER." + ], + "image_footnote": [], + "bbox": [ + 96, + 517, + 901, + 868 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "2885", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "batch size of 6 to optimize the weights over approximately 200K iterations on three NVIDIA Tesla A800 GPUs. To preserve diversity and detail, we degrade CoSER to a text-to-image generation model with a $40\\%$ probability and train it with a subset of LAION-5B. During inference, we employ the DDIM sampler with 50 steps and set the scale guidance to 7.5 for generation. Additionally, we set text template as \"A DSLR photo of [prompt], 3d asset\" to acquire more realistic results. Following the Liu et al. [27], we use NeuS[60] for performing reconstruction.", + "bbox": [ + 89, + 90, + 480, + 241 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Dataset We use the public G-Objaverse [40] as the training dataset, utilizing the names of the objects appended with \"3D asset\" as their text labels. For each object, we first normalize the object to be centered within a bounding box ranging from $[-0.5, 0.5]$ . The camera distance is set to 2.0 units. A random elevation is chosen between $[-30^\\circ, 30^\\circ]$ , and starting from the front view, we uniformly sample 12 angles around the object for rendering.", + "bbox": [ + 89, + 243, + 480, + 364 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.1. Comparison with SOTA methods", + "text_level": 1, + "bbox": [ + 89, + 377, + 380, + 393 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Qualitative Comparison We primarily evaluate the performance of the proposed CoSER on multiview text-to-image and compare the results with state-of-the-art methods in Figure 6. For VideoMV [74], 12 images generated from the same camera viewpoints as CoSER are selected for visualization, and all 24 viewpoints were utilized in the reconstruction process to ensure a fair comparison. Qualitatively, our method demonstrates impressive quality improvement in terms of both consistence and texture. Although VideoMV can generate images from a broader range of perspectives, e.g., mask and watch, it significantly compromises cross-view consistency, which poses serious challenges for subsequent 3D reconstruction. Furthermore, Figure 5 presents a comparison between two state-of-the-art text-to-3D methods, including GaussianDreamer [69] and Hash3D [68]. Specifically, the advanced method, Dream-Gaussian [54], is utilized as the foundational model for Hash3D. Thanks to the dense consistent images, CoSER outperforms other methods in terms of fidelity in 3D geometry and visual appearance.", + "bbox": [ + 91, + 400, + 482, + 702 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Quantitative Comparison Following previous work [74], we quantitatively assess the quality of the generated 3D assets using CLIP Score [14]. To evaluate the textual alignment, we followed [10] to compute the recall of the original prompt within the generated caption. We employ 100 text prompts selected from T3Bench [13] to generate 3D models. For each object, we set the elevation as $15^{\\circ}$ and select 120 evenly spaced azimuth angles from $-180^{\\circ}$ to $180^{\\circ}$ , thereby generating 120 rendered images from different viewpoints. During the evaluation, 30 from the 120 rendered images are randomly selected and the CLIP Score is computed as the average semantic similarity between the prompts and these images. We use the CLIP ViT-B [43] to", + "bbox": [ + 89, + 704, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "extract both text and image features. The results, presented in Table 1, demonstrate that our method significantly surpasses other methods, indicating that CoSER is more adept at injecting semantic information into the creation of realistic 3D assets. Given the challenges in automating the measurement of 3D consistency, we resort to user studies to assess which method produces results that are most favored by human evaluators. We conducted three user studies on the results in terms of the Semantic Alignment, Multiview Consistency and Texture Detail. In the first study, participants were asked to select the 3D asset that best aligns with the semantics of the given prompt. In the second study, participants identified the asset with the highest consistency. In the third study, participants were encouraged to choose the object with the finest texture and less distortion artifacts. We receive 100 answers on 30 text-3D pairs in total for each study. As shown in Table 1, over $29\\%$ of our results are selected as the best in both three metrics, which proves a significant advantage in 3D generation.", + "bbox": [ + 511, + 90, + 903, + 378 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Ablation Study", + "text_level": 1, + "bbox": [ + 511, + 392, + 666, + 407 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effect of Module Design We are mainly interested in the 3D-aware module design, given that other unit has been well explored in previous methods. The core of our design is focusing on neighbors in fine-grained manner and knowing whole views with coarse-grained mechanism. To substantiate these claims, we incrementally integrated the proposed modules into four distinct configurations and finetuned the LDM accordingly. As depicted in Figure 6, the use of Appearance Awareness alone proves insufficient for ensuring cross-view consistency. However, the incorporation of Detail Refinement enhances the alignment between adjacent views, thereby improving overall 3D content consistency. It is evident that Rapid Glance offers small reduction in ambiguities; nonetheless, its primary benefit lies in facilitating information exchange across all viewpoints (See Figure 8), thereby establishing a foundation for the effective Accumulated Inconsistency Rectification module. Additionally, we present the parameters and MACs introduced by each module in Table 2. To illustrate that SSM captures multi-view consistency more effectively than temporal attention, we also learn global consistency by temporal attention found that it impaired performance. Based on this observation, we consider the additional 3.31M parameters and 15G MACs computations introduced by RG and AIR to be entirely acceptable.", + "bbox": [ + 511, + 415, + 903, + 792 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Spiral Bidirectional Scan in Rapid Glance To verify the effectiveness of the proposed spiral bidirectional scan strategy, we perform the evaluation by replacing it with vanilla spatial-first bidirectional scan in VideoMamba [22]. As shown in Figure 7, both scanning strategies enhance the consistency. However, spatial-first bidirectional scan inevitably disrupts the continuity of object, which hampers", + "bbox": [ + 511, + 795, + 903, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "2886", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/9e84dae141625baa174541eda61b5f4f07aa41d95aaf358f14da02d5c903c229.jpg", + "table_caption": [ + "Table 1. Quantitative comparison of CoSER and state-of-the-arts based on the Quality, Alignment and User Study." + ], + "table_footnote": [], + "table_body": "
ModelQuality↑Alignment↑User Study
Semantic↑Consistency↑Texture↑
GaussianDreamer [69]30.7833.70.260.270.21
Hash3D [68]31.2630.20.200.190.25
VideoMV [74]32.1934.40.230.250.22
Ours33.0737.70.310.290.32
", + "bbox": [ + 101, + 118, + 620, + 193 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/3c6925f4791c34303aac14eaba23d3007f1aa8997d9c9e0c604a52302b49ec40.jpg", + "table_caption": [ + "Table 2. Model efficiency of each module. TA denotes temporal attention." + ], + "table_footnote": [], + "table_body": "
ModuleParameters (M)MACs (G)
AA867.451036
AA + DR917.041395
AA + DR + RG917.071396
AA + DR + RG + AIR969.951487
AA + DR + TA966.641472
", + "bbox": [ + 630, + 123, + 916, + 188 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f205016f0384fc1af26b56a0c44100d5c1dc7bc3cb1268705c0f8474b2bc46cc.jpg", + "image_caption": [ + "Figure 6. Ablation of proposed moulds. The arrows indicate the changes made to the Appearance Awareness by sequentially plusing Detail Refinement, Rapid Glance, and Accumulated Inconsistency Rectification. Better see in color and $2\\mathrm{x}$ zoom." + ], + "image_footnote": [], + "bbox": [ + 99, + 196, + 895, + 348 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/768572bb6b40cbcfc096d97848991439a0c40871803f71f7adba73e16dbd9697.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 93, + 386, + 478, + 492 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5106cb7081072979b910baca1b63665ead15d1be0a0c841edaa062028b7d9966.jpg", + "image_caption": [ + "Figure 7. Ablation of scanning strategy. Spiral bidirectional scan generates semantic- and view-consistent images.", + "Figure 8. Ablation of score map. Score map aids in capturing critical information, while Rapid Glance contributes to the generation of more accurate score maps." + ], + "image_footnote": [], + "bbox": [ + 96, + 530, + 475, + 652 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "the efficiency of SSM model for feature aggregation. In contrast, our proposed spiral bidirectional scan maximizes the spatial compactness of foreground content while simultaneously enforcing operator awareness of inter-view boundaries, thereby achieving better results. Additionally, we experimented with reversing the spiral scanning order within the image, progressing from the periphery inward towards the center. However, this strategy did not yield further performance improvements, which we attribute to the already efficient information aggregation within an image achieved by the spiral scan.", + "bbox": [ + 88, + 704, + 482, + 869 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Score Map in Accumulated Inconsistency Rectification In Accumulated Inconsistency Rectification, we introduce the", + "bbox": [ + 89, + 869, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "score map to assist in downsampling feature resolution for coarse attention, tailored to each individual view. To understand its impact, we compared the full AIR module generated results with those from two alternative models: AIR without score map and AIR with score map but without Rapid Glance. For clearness, we extract score maps at various resolutions and upsample them to 32x32 for visualization. As shown in Figure 8, the score map assigns higher scores to foreground regions, while background areas receive lower scores, facilitating the retention of critical information during the downsampling process. Additionally, Rapid Glance enhances the exchange of information, contributing to the generation of more precise score maps.", + "bbox": [ + 511, + 398, + 906, + 595 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 602, + 633, + 617 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we introduced CoSER, a novel framework for generating consistent and high-fidelity dense multiview images from text prompts. Our approach leverages a hybrid architecture that encapsulates the strengths of attention mechanisms and state space models (SSMs) to address the crucial challenges of cross-view consistency. The innovative design of contextual interaction includes an adjacent attention for capturing rough structures and textures, a trajectory-based attention for refining local details. A spiral bidirectional Mamba ensures global coherence in a cost-efficient manner. This comprehensive approach allows CoSER to produce dense, high-fidelity, and content-consistent multiview images. Extensive experiments demonstrate that CoSER outperforms state-of-the-art multiview synthesis methods in both quantitative and qualitative evaluations across various datasets. Future work could further explore the hybrid way of neighbor and global interaction to efficiently generate dense and consistent multiview images.", + "bbox": [ + 509, + 628, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "2887", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 91, + 90, + 259, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This paper is supported by the Beijing Natural Science Foundation (1254050), National Natural Science Foundation of China under Grants (U23B2012, 12471308), Fundamental Research Funds for the Central Universities, and by the National Research Foundation, Singapore, under its Medium Sized Center for Advanced Robotics Technology Innovation.", + "bbox": [ + 89, + 114, + 483, + 220 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 234, + 187, + 251 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Haoyu Chen, Jingjing Ren, Jinjin Gu, Hongtao Wu, Xuequan Lu, Haoming Cai, and Lei Zhu. Snow removal in video: A new dataset and a novel method. in 2023 IEEE. In ICCV, pages 13165-13176. 2", + "[2] Haoyu Chen, Jinjin Gu, Yihao Liu, Salma Abdel Magid, Chao Dong, Qiong Wang, Hanspeter Pfister, and Lei Zhu. Masked image training for generalizable deep image denoising. In CVPR, pages 1692-1703, 2023. 2", + "[3] Rui Chen, Yongwei Chen, Ningxin Jiao, and Kui Jia. Fantasia3d: Disentangling geometry and appearance for high-quality text-to-3d content creation. In ICCV, pages 22246–22256, 2023. 1, 3", + "[4] Yongwei Chen, Rui Chen, Jiabao Lei, Yabin Zhang, and Kui Jia. Tango: Text-driven photorealistic and robust 3d stylization via lighting decomposition. In NeurIPS, pages 30923-30936, 2022. 2", + "[5] Yiwen Chen, Zilong Chen, Chi Zhang, Feng Wang, Xiaofeng Yang, Yikai Wang, Zhongang Cai, Lei Yang, Huaping Liu, and Guosheng Lin. Gaussianeditor: Swift and controllable 3d editing with gaussian splatting. In CVPR, pages 21476-21485, 2024. 2", + "[6] Zilong Chen, Yikai Wang, Feng Wang, Zhengyi Wang, and Huaping Liu. V3d: Video diffusion models are effective 3d generators. arXiv preprint arXiv:2403.06738, 2024. 2, 3", + "[7] Yen-Chi Cheng, Hsin-Ying Lee, Sergey Tulyakov, Alexander G Schwing, and Liang-Yan Gui. Sdfusion: Multimodal 3d shape completion, reconstruction, and generation. In CVPR, pages 4456-4465, 2023. 2", + "[8] Lihe Ding, Shaocong Dong, Zhanpeng Huang, Zibin Wang, Yiyuan Zhang, Kaixiong Gong, Dan Xu, and Tianfan Xue. Text-to-3d generation with bidirectional diffusion using both 2d and 3d priors. In CVPR, pages 5115–5124, 2024. 2", + "[9] Ruiqi Gao, Aleksander Holynski, Philipp Henzler, Arthur Brussee, Ricardo Martin-Brualla, Pratul Srinivasan, Jonathan T Barron, and Ben Poole. Cat3d: Create anything in 3d with multi-view diffusion models. arXiv preprint arXiv:2405.10314, 2024. 2, 3", + "[10] Chongjian Ge, Chenfeng Xu, Yuanfeng Ji, Chensheng Peng, Masayoshi Tomizuka, Ping Luo, Mingyu Ding, Varun Jampani, and Wei Zhan. Compgs: Unleashing 2d compositionality for compositional text-to-3d via dynamically optimizing 3d gaussians. arXiv preprint arXiv:2410.20723, 2024. 7", + "[11] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023. 5" + ], + "bbox": [ + 93, + 260, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[12] Albert Gu, Karan Goel, and Christopher Re. Efficiently modeling long sequences with structured state spaces. In ICLR, 2022. 2, 5", + "[13] Yuze He, Yushi Bai, Matthieu Lin, Wang Zhao, Yubin Hu, Jenny Sheng, Ran Yi, Juanzi Li, and Yong-Jin Liu. T3 bench: Benchmarking current progress in text-to-3d generation. arXiv preprint arXiv:2310.02977, 2023. 1, 7", + "[14] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning. In EMNLP, pages 7514-7528, 2021. 7", + "[15] Fangzhou Hong, Jiaxiang Tang, Ziang Cao, Min Shi, Tong Wu, Zhaoxi Chen, Tengfei Wang, Liang Pan, Dahua Lin, and Ziwei Liu. 3dtopia: Large text-to-3d generation model with hybrid diffusion priors. arXiv preprint arXiv:2403.02234, 2024. 2", + "[16] Yukun Huang, Jianan Wang, Yukai Shi, Xianbiao Qi, Zheng-Jun Zha, and Lei Zhang. Dreamtime: An improved optimization strategy for text-to-3d content creation. arXiv preprint arXiv:2306.12422, 2023. 3", + "[17] Zehuan Huang, Hao Wen, Junting Dong, Yaohui Wang, Yangguang Li, Xinyuan Chen, Yan-Pei Cao, Ding Liang, Yu Qiao, Bo Dai, et al. Epidiff: Enhancing multi-view synthesis via localized epipolar-constrained diffusion. In CVPR, pages 9784-9794, 2024. 2, 3", + "[18] Ajay Jain, Ben Mildenhall, Jonathan T Barron, Pieter Abbeel, and Ben Poole. Zero-shot text-guided object generation with dream fields. In CVPR, pages 867-876, 2022. 2", + "[19] Jeong-gi Kwak, Erqun Dong, Yuhe Jin, Hanseok Ko, Shweta Mahajan, and Kwang Moo Yi. Vivid-1-to-3: Novel view synthesis with video diffusion models. In CVPR, pages 6775–6785, 2024. 2", + "[20] Bonan Li, Zicheng Zhang, Xuecheng Nie, Congying Han, Yinhan Hu, and Tiande Guo. Styo: Stylize your face in only one-shot. arXiv preprint arXiv:2303.03231, 2023. 2", + "[21] Jiahao Li, Hao Tan, Kai Zhang, Zexiang Xu, Fujun Luan, Yinghao Xu, Yicong Hong, Kalyan Sunkavalli, Greg Shakhnarovich, and Sai Bi. Instant3d: Fast text-to-3d with sparse-view generation and large reconstruction model. In ICLR, 2023. 2, 3", + "[22] Kunchang Li, Xinhao Li, Yi Wang, Yinan He, Yali Wang, Limin Wang, and Yu Qiao. Videomamba: State space model for efficient video understanding. arXiv preprint arXiv:2403.06977, 2024. 7", + "[23] Peng Li, Yuan Liu, Xiaoxiao Long, Feihu Zhang, Cheng Lin, Mengfei Li, Xingqun Qi, Shanghang Zhang, Wenhan Luo, Ping Tan, et al. Era3d: High-resolution multiview diffusion using efficient row-wise attention. arXiv preprint arXiv:2405.11616, 2024. 2, 3", + "[24] Weiyu Li, Rui Chen, Xuelin Chen, and Ping Tan. Sweetdreamer: Aligning geometric priors in 2d diffusion for consistent text-to-3d. In ICLR, 2024. 3", + "[25] Wenbing Li, Hang Zhou, Junqing Yu, Zikai Song, and Wei Yang. Coupled mamba: Enhanced multimodal fusion with coupled state space model. In NeurIPS, pages 59808-59832, 2025. 2" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "2888", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[26] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. In CVPR, pages 300–309, 2023. 1, 3", + "[27] Yuan Liu, Cheng Lin, Zijiao Zeng, Xiaoxiao Long, Lingjie Liu, Taku Komura, and Wenping Wang. Syncdreamer: Generating multiview-consistent images from a single-view image. In ICLR, 2024. 7", + "[28] Xiaoxiao Long, Yuan-Chen Guo, Cheng Lin, Yuan Liu, Zhiyang Dou, Lingjie Liu, Yuexin Ma, Song-Hai Zhang, Marc Habermann, Christian Theobalt, et al. Wonder3d: Single image to 3d using cross-domain diffusion. In CVPR, pages 9970-9980, 2024. 2, 3, 4", + "[29] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In ICLR, 2019. 5", + "[30] Run Luo, Zikai Song, Lintao Ma, Jinlin Wei, Wei Yang, and Min Yang. Diffusiontrack: Diffusion model for multi-object tracking. In AAAI, pages 3991-3999, 2024. 2", + "[31] Baorui Ma, Haoge Deng, Junsheng Zhou, Yu-Shen Liu, Tiejun Huang, and Xinlong Wang. Geodream: Disentangling 2d and geometric priors for high-fidelity and consistent 3d generation. arXiv preprint arXiv:2311.17971, 2023. 3", + "[32] Luke Melas-Kyriazi, Iro Laina, Christian Rupprecht, Natalia Neverova, Andrea Vedaldi, Oran Gafni, and Filippos Kokkinos. Im-3d: Iterative multiview diffusion and reconstruction for high-quality 3d generation. In ICML, 2024. 2, 3", + "[33] Gal Metzer, Elad Richardson, Or Patashnik, Raja Giryes, and Daniel Cohen-Or. Latent-nerf for shape-guided generation of 3d shapes and textures. In CVPR, pages 12663–12673, 2023. 1", + "[34] Gal Metzer, Elad Richardson, Or Patashnik, Raja Giryes, and Daniel Cohen-Or. Latent-nerf for shape-guided generation of 3d shapes and textures. In CVPR, pages 12663–12673, 2023. 3", + "[35] Nasir Mohammad Khalid, Tianhao Xie, Eugene Belilovsky, and Tiberiu Popa. Clip-mesh: Generating textured meshes from text using pretrained image-text models. In SIGGRAPH Asia, pages 1-8, 2022. 2", + "[36] Alex Nichol, Heewoo Jun, Prafulla Dhariwal, Pamela Mishkin, and Mark Chen. Point-e: A system for generating 3d point clouds from complex prompts. arXiv preprint arXiv:2212.08751, 2022. 2", + "[37] JangHo Park, Gihyun Kwon, and Jong Chul Ye. Ed-nerf: Efficient text-guided editing of 3d scene with latent space nerf. In ICLR, 2024. 2", + "[38] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. In ICLR, 2022. 1, 2", + "[39] Guocheng Qian, Jinjie Mai, Abdullah Hamdi, Jian Ren, Aliaksandr Siarohin, Bing Li, Hsin-Ying Lee, Ivan Skorokhodov, Peter Wonka, Sergey Tulyakov, et al. Magic123: One image to high-quality 3d object generation using both 2d and 3d diffusion priors. arXiv preprint arXiv:2306.17843, 2023. 2", + "[40] Lingteng Qiu, Guanying Chen, Xiaodong Gu, Qi Zuo, Mutian Xu, Yushuang Wu, Weihao Yuan, Zilong Dong, Liefeng" + ], + "bbox": [ + 91, + 90, + 480, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Bo, and Xiaoguang Han. Richdreamer: A generalizable normal-depth diffusion model for detail richness in text-to-3d. In CVPR, pages 9914–9925, 2024. 7", + "[41] Xinmin Qiu, Congying Han, ZiCheng Zhang, Bonan Li, Tiande Guo, and Xuecheng Nie. Diffbfr: Bootstrapping diffusion model for blind face restoration. In ACM MM, pages 7785-7795, 2023. 2", + "[42] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763, 2021. 3", + "[43] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763, 2021. 7", + "[44] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10684-10695, 2022. 2, 3, 5", + "[45] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, Patrick Schramowski, Srivatsa Kundurthy, Katherine Crowson, Ludwig Schmidt, Robert Kaczmarczyk, and Jenia Jitsev. Laion-5b: An open large-scale dataset for training next generation image-text models. ArXiv, abs/2210.08402, 2022. 5", + "[46] Junyoung Seo, Wooseok Jang, Min-Seop Kwak, Hyeonsu Kim, Jaehoon Ko, Junho Kim, Jin-Hwa Kim, Jiyoung Lee, and Seungryong Kim. Let 2d diffusion model know 3d-consistency for robust text-to-3d generation. In ICLR, 2024. 3", + "[47] Yichun Shi, Peng Wang, Jianglong Ye, Long Mai, Kejie Li, and Xiao Yang. Mvdream: Multi-view diffusion for 3d generation. In ICLR, 2023. 2, 3, 4", + "[48] Zikai Song, Junqing Yu, Yi-Ping Phoebe Chen, and Wei Yang. Transformer tracking with cyclic shifting window attention. In CVPR, pages 8791-8800, 2022. 2", + "[49]Zikai Song, Run Luo, Junqing Yu, Yi-Ping Phoebe Chen, and Wei Yang. Compact transformer tracker with correlative masked modeling. In AAAI, pages 2321-2329, 2023.", + "[50] Zikai Song, Ying Tang, Run Luo, Lintao Ma, Junqing Yu, Yi-Ping Phoebe Chen, and Wei Yang. Autogenic language embedding for coherent point tracking. In ACM MM, pages 2021-2030, 2024.", + "[51] Haoze Sun, Wenbo Li, Jianzhuang Liu, Haoyu Chen, Renjing Pei, Xueyi Zou, Youliang Yan, and Yujiu Yang. Coser: Bridging image and language for cognitive super-resolution. In CVPR, pages 25868-25878, 2024. 2", + "[52] Jingxiang Sun, Bo Zhang, Ruizhi Shao, Lizhen Wang, Wen Liu, Zhenda Xie, and Yebin Liu. Dreamcraft3d: Hierarchical 3d generation with bootstrapped diffusion prior. In ICLR, 2024. 2", + "[53] Junshu Tang, Tengfei Wang, Bo Zhang, Ting Zhang, Ran Yi, Lizhuang Ma, and Dong Chen. Make-it-3d: High-fidelity 3d" + ], + "bbox": [ + 516, + 92, + 903, + 901 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "2889", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "creation from a single image with diffusion prior. In ICCV, pages 22819-22829, 2023. 2", + "[54] Jiaxiang Tang, Jiawei Ren, Hang Zhou, Ziwei Liu, and Gang Zeng. Dreamgaussian: Generative gaussian splatting for efficient 3d content creation. In ICLR, 2024. 7", + "[55] Shitao Tang, Fuyang Zhang, Jiacheng Chen, Peng Wang, and Yasutaka Furukawa. Mvdiffusion: Enabling holistic multiview image generation with correspondence-aware diffusion. arXiv, 2023. 3", + "[56] Christina Tsalicoglou, Fabian Manhardt, Alessio Tonioni, Michael Niemeyer, and Federico Tombari. Textmesh: Generation of realistic 3d meshes from text prompts. In 3DV, pages 1554-1563, 2024. 2", + "[57] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 2", + "[58] Vikram Voleti, Chun-Han Yao, Mark Boss, Adam Letts, David Pankratz, Dmitry Tochilkin, Christian Laforte, Robin Rombach, and Varun Jampani. Sv3d: Novel multi-view synthesis and 3d generation from a single image using latent video diffusion. arXiv preprint arXiv:2403.12008, 2024. 2, 3", + "[59] Peng Wang and Yichun Shi. Imagedream: Image-prompt multi-view diffusion for 3d generation. arXiv preprint arXiv:2312.02201, 2023. 2", + "[60] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. In NeurIPS, 2021. 3, 7", + "[61] Tengfei Wang, Bo Zhang, Ting Zhang, Shuyang Gu, Jianmin Bao, Tadas Baltrusaitis, Jingjing Shen, Dong Chen, Fang Wen, Qifeng Chen, et al. Rodin: A generative model for sculpting 3d digital avatars using diffusion. In CVPR, pages 4563-4573, 2023. 2", + "[62] Xinzhou Wang, Yikai Wang, Junliang Ye, Zhengyi Wang, Fuchun Sun, Pengkun Liu, Ling Wang, Kai Sun, Xintong Wang, and Bin He. Animatabledreamer: Text-guided non-rigid 3d model generation and reconstruction with canonical score distillation. arXiv preprint arXiv:2312.03795, 2023. 2", + "[63] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. In NeurIPS, 2024. 3", + "[64] Zeyu Xiao and Zhiwei Xiong. Incorporating degradation estimation in light field spatial super-resolution. Computer Vision and Image Understanding, page 104295, 2025. 2", + "[65] Zeyu Xiao, Yutong Liu, Ruisheng Gao, and Zhiwei Xiong. Cutmib: Boosting light field super-resolution via multi-view image blending. In CVPR, pages 1672-1682, 2023.", + "[66] Zeyu Xiao, Dachun Kai, Yueyi Zhang, Xiaoyan Sun, and Zhiwei Xiong. Asymmetric event-guided video superresolution. In ACM MM, pages 2409-2418, 2024.", + "[67] Xinchao Wang Xingyi Yang. Kolmogorov-arnold transformer. In ICLR, 2025. 2", + "[68] Xingyi Yang and Xinchao Wang. Hash3d: Training-free acceleration for 3d generation. arXiv preprint arXiv:2404.06091, 2024. 3, 6, 7, 8" + ], + "bbox": [ + 91, + 90, + 482, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[69] Taoran Yi, Jiemin Fang, Junjie Wang, Guanjun Wu, Lingxi Xie, Xiaopeng Zhang, Wenyu Liu, Qi Tian, and Xinggang Wang. Gaussian dreamer: Fast generation from text to 3d gaussians by bridging 2d and 3d diffusion models. In CVPR, pages 6796-6807, 2024. 3, 6, 7, 8", + "[70] Weihao Yu and Xinchao Wang. Mambaout: Do we really need mamba for vision? In CVPR, 2025. 2", + "[71] Xianfang Zeng, Xin Chen, Zhongqi Qi, Wen Liu, Zibo Zhao, Zhibin Wang, Bin Fu, Yong Liu, and Gang Yu. Paint3d: Paint anything 3d with lighting-less texture diffusion models. In CVPR, pages 4252-4262, 2024. 2", + "[72] Zicheng Zhang, Bonan Li, Xuecheng Nie, Congying Han, Tiande Guo, and Luoqi Liu. Towards consistent video editing with text-to-image diffusion models. In NeurIPS, pages 58508-58519, 2023. 2", + "[73] Jingyu Zhuang, Chen Wang, Liang Lin, Lingjie Liu, and Guanbin Li. Dreameditor: Text-driven 3d scene editing with neural fields. In SIGGRAPH Asia, pages 1-10, 2023. 2", + "[74] Qi Zuo, Xiaodong Gu, Lingteng Qiu, Yuan Dong, Zhengyi Zhao, Weihao Yuan, Rui Peng, Siyu Zhu, Zilong Dong, Liefeng Bo, et al. Videomv: Consistent multi-view generation based on large video generative model. arXiv preprint arXiv:2403.12010, 2024. 2, 3, 5, 6, 7, 8" + ], + "bbox": [ + 516, + 90, + 903, + 415 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "2890", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/CoSER_ Towards Consistent Dense Multiview Text-to-Image Generator for 3D Creation/911def72-ec15-4b90-8a72-6946e2b20044_model.json b/2025/CoSER_ Towards Consistent Dense Multiview Text-to-Image Generator for 3D Creation/911def72-ec15-4b90-8a72-6946e2b20044_model.json new file mode 100644 index 0000000000000000000000000000000000000000..d9012a4f857671c52d789b079f25b2d25391f73b --- /dev/null +++ b/2025/CoSER_ Towards Consistent Dense Multiview Text-to-Image Generator for 3D Creation/911def72-ec15-4b90-8a72-6946e2b20044_model.json @@ -0,0 +1,2521 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.195, + 0.131, + 0.806, + 0.175 + ], + "angle": 0, + "content": "CoSER: Towards Consistent Dense Multiview Text-to-Image Generator for 3D Creation" + }, + { + "type": "text", + "bbox": [ + 0.229, + 0.203, + 0.773, + 0.223 + ], + "angle": 0, + "content": "Bonan Li\\(^{1,2*}\\) Zicheng Zhang\\(^{1*}\\) Xingyi Yang\\(^{2}\\) Xinchao Wang\\(^{2\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.228, + 0.223, + 0.772, + 0.24 + ], + "angle": 0, + "content": "\\(^{1}\\)University of Chinese Academy of Sciences \\(^{2}\\) National University of Singapore" + }, + { + "type": "text", + "bbox": [ + 0.16, + 0.241, + 0.834, + 0.258 + ], + "angle": 0, + "content": "libonacci@ucas.ac.cn zhangzicheng19@mails.ucas.ac.cn xyang@u.nus.edu xinchao@nus.edu.sg" + }, + { + "type": "image", + "bbox": [ + 0.11, + 0.263, + 0.882, + 0.501 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.516, + 0.908, + 0.545 + ], + "angle": 0, + "content": "Figure 1. CoSER aims to generate detailed, diverse and detaild 3D objects from text prompts. All prompts are from T3Bench [13] and the templated is set as \"A DSLR photo of ..., 3d asset\". Better see in color and 2x zoom." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.578, + 0.329, + 0.594 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.61, + 0.486, + 0.868 + ], + "angle": 0, + "content": "Generating dense multiview images from text prompts is crucial for creating high-fidelity 3D assets. Nevertheless, existing methods struggle with space-view correspondences, resulting in sparse and low-quality outputs. In this paper, we introduce CoSER, a novel consistent dense Multiview Text-to-Image Generator for Text-to-3D, achieving both efficiency and quality by meticulously learning neighbor-view coherence and further alleviating ambiguity through the swift traversal of all views. For achieving neighbor-view consistency, each viewpoint densely interacts with adjacent viewpoints to perceive the global spatial structure, and aggregates information along motion paths explicitly defined by physical principles to refine details. To further enhance cross-view consistency and alleviate content drift, CoSER rapidly scan all views in spiral bidirectional manner to aware holistic information and then scores each point based on semantic material. Subsequently, we" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.58, + 0.909, + 0.732 + ], + "angle": 0, + "content": "conduct weighted down-sampling along the spatial dimension based on scores, thereby facilitating prominent information fusion across all views with lightweight computation. Technically, the core module is built by integrating the attention mechanism with a selective state space model, exploiting the robust learning capabilities of the former and the low overhead of the latter. Extensive evaluation shows that CoSER is capable of producing dense, high-fidelity, content-consistent multiview images that can be flexibly integrated into various 3D generation models." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.767, + 0.646, + 0.783 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Creating 3D content from text is an important task in computer vision and graphics for its potential applications in advertising, robotics, digital games and the meta-verse. Previous works [3, 26, 33] primarily optimise 3D representations using score distillation sampling [38] of Text-to-Image (T2I) models, ensuring that rendered images from any viewpoint maintain high fidelity while aligning with" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.876, + 0.218, + 0.888 + ], + "angle": 0, + "content": "*Equal contribution" + }, + { + "type": "page_footnote", + "bbox": [ + 0.111, + 0.888, + 0.235, + 0.9 + ], + "angle": 0, + "content": "† Corresponding author" + }, + { + "type": "list", + "bbox": [ + 0.109, + 0.876, + 0.235, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "2880" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.198 + ], + "angle": 0, + "content": "the given prompt. Nevertheless, these methods suffer from Janus problems and severe artifacts with asymmetric geometry. As a remedy, multiview text-to-image (MT2I) models [6, 17, 19, 21, 23, 32] have been developed, providing rich 3D priors that help mitigate these problems. Furthermore, using dense views can accelerate the modeling process and enhance the quality of the resulting 3D models." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.2, + 0.483, + 0.442 + ], + "angle": 0, + "content": "To acquire MT2I models, a promising approach is to leverage 3D priors learned from 3D datasets to enhance 2D diffusion models. Seminal works [28, 47] have focused on inflating T2I models with dense cross-view attention to capture 3D information. However, this approach incurs expensive computation and faces challenges when dealing with dense viewpoints. To address these limitations, sparse attention mechanisms [23] have been proposed to model cross-view dependencies, though they can introduce inconsistencies across multiple views. Recent studies [9, 58, 74] have leveraged cross-frame continuity priors from text-to-video (T2V) models to enhance consistency. Nevertheless, azimuthal rotation around an object can cause significant content variations at the same image coordinates, making temporal attention insufficient for capturing intricate details necessary for 3D generation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.445, + 0.483, + 0.761 + ], + "angle": 0, + "content": "Despite impressive progress in multiview T2I generation, effectively and efficiently modeling the space-view structure within a unified architecture remains an open challenge. In this paper, we advance the field by addressing two fundamental problems: (1) How to design efficient cross-view interactions to ensure dense multiview consistency? Rather than directly optimizing consistency among all views with uniform granularity, we propose a two-step approach: (i) Learning neighbor-view consistency as a prerequisite, as enforcing consistency between adjacent viewpoints naturally enables coherence across all viewpoints. (ii) Ensuring global-view coherence to mitigate discrepancies between distant views and facilitate unified appearance, since subtle inconsistencies are inevitable and deteriorated as the viewing angles increases. (2) Which operator should be applied for the implementation? Unlike previous works that employ attention as the default choice, we leverage hybrid operators, including attention [57] and state space models (SSMs) [12], to harness the benefits of both model families, e.g., the powerful capabilities of attention and the low computational cost of SSMs." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.903 + ], + "angle": 0, + "content": "In this paper, we address these questions and propose a novel model named CoSER, which enhances consistency in dense MT2I tasks. CoSER utilizes \"factor\" interactions that focus on neighboring views and maintain awareness of the global view. Technically, CoSER initially employs dense attention on adjacent views to capture the rough structure and texture of the current plane, thereby constructing a foundational appearance. Subsequently, we introduce a trajectory-based module to refine local details by accurately tracing" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.362 + ], + "angle": 0, + "content": "3D trajectories and enhancing point-to-point consistency. Specifically, motivated by the efficiency of calculating rotated locations for centralized objects, we develop a robust method to handle complicate cases. This involves pixel interaction within a window of neighboring frames defined by physical rotation rules, ensuring precise alignment and consistency across views. To further alleviate disparities across arbitrary viewpoints, we utilize state space models (SSMs) equipped with a novel spiral bidirectional scan strategy to aggregate features from the entire set of views. We then score each patch based on textual inputs and perform down-sampling conditioned on these scores to adaptively ignore semantically irrelevant information. This approach allows CoSER to efficiently maintain global context through sparse attention. With these designs, our method establishes robust priors, enabling the creation of exceptionally coherent 3D assets (see Figure 1). In summary, our contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.365, + 0.905, + 0.409 + ], + "angle": 0, + "content": "- We propose a novel multiview text-to-image framework, CoSER, for 3D content creation which exhibits significantly consistency behavior." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.41, + 0.905, + 0.469 + ], + "angle": 0, + "content": "- We explore an efficient pipeline to model complex long-span consistency by focusing on neighbors to achieve short-span strict consistency and knowing whole views to obtain cues that correct cumulative errors." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.471, + 0.905, + 0.546 + ], + "angle": 0, + "content": "- We design an hybrid architecture to implement the interaction by adopting both attention and SSMs as operators, effectively capturing essential dependencies in short sequences while significantly reducing costs for modeling long sequences." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.546, + 0.905, + 0.592 + ], + "angle": 0, + "content": "- Extensive experiments demonstrate that our CoSER outperforms the state-of-the-art multiview synthesis approaches in both quantitative and qualitative results." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.365, + 0.905, + 0.592 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.609, + 0.655, + 0.625 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.636, + 0.704, + 0.65 + ], + "angle": 0, + "content": "2.1. Text-to-3D Creation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.659, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Recently, Text-to-3D generative models [4, 8, 15, 18, 35, 38] have made substantial advancements in computer vision [1, 2, 20, 25, 30, 41, 48–51, 64–67, 70, 72], achieving high-quality generation of both 3D objects and complex 3D scenes. Early works focus on directly generating assets using 3D diffusion models based on differentiable representations [7, 36, 61]. However, these approaches often relies on massive 3D data and incurs substantial training costs. Thanks to powerful text-to-image models [44], DreamFusion [38] leverages the robust 2D prior by introducing Score Distillation Sampling, which minimizes the discrepancy between rendered images from the underlying 3D assets and the diffusion prior. This development also sparked a surge of interest in other generative tasks such as texture generation [56, 71], editing [5, 37, 52, 62, 73], and Image-to-3D [39, 53, 59]. Despite this paradigm has bridged the" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.513, + 0.957 + ], + "angle": 0, + "content": "2881" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.439 + ], + "angle": 0, + "content": "gap between imagination and reality, it still faces significant challenges in terms of optimization speed [16, 68], content drift [24, 63] and the Janus problem [31, 46, 69]. In response to inefficient training, Metzer et al. [34] train a NeRF within the latent space of LDM [44] to accelerate the learning process and further introduce the Sketch-Shape method to guide the generation process. To acquire photorealistic appearance, Lin et al. [26] employ a coarse-to-fine strategy, first optimizing a NeRF using a low-resolution diffusion prior, then refining the texture with latent diffusion priors using the DMTet initialized with the coarse NeRF. Chen et al. [3] disentangle the learning of geometry and material, leveraging physics-based rendering techniques to achieve high-fidelity mesh generation. Another line of work lies in generating 3D assets directly through a 3D diffusion model based on differentiable representations. [28] fine-tune image diffusion models to generate multiview image-normal pairs with switch attention and employ NeuS [60] for reconstruction underlying geometries. In this work, we will continue to improve the quality of 3D content by designing the effective multiview text-to-image generator, which provides consistent images for reconstruction method." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.455, + 0.35, + 0.47 + ], + "angle": 0, + "content": "2.2. Multiview Image Generation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.478, + 0.485, + 0.901 + ], + "angle": 0, + "content": "Due to the scarcity of 3D data and the formidable training costs, lifting 2D diffusion to multiview generator conditioned on camera pose has become the prevailing paradigm [6, 17, 21, 23, 28, 32, 47, 55]. Tang et al. [55] introduce the novel concept of generating multiview images concurrently with correspondence-aware attention mechanisms, thereby enabling effective cross-view information interaction and application to texturing scene meshes. Meanwhile, Shi et al. [47] and Long et al. [28] simply extends 2D spatial attention to capture 3D structure, yielding remarkable results. Nevertheless, the quadratic complexity along with resolution restricts the number of cameras, thereby compromising the performance of reconstruction algorithms and resulting in inferior results. To reduce computational costs, Li et al. [23] propose an efficient rowwise attention instead of dense interactions, yet struggles to correct local details referencing global information. Latter works [9, 58, 74] capitalized on priors derived from video diffusion model to achieve both dense view generation and consistent outputs. Despite these strides, a fundamental disparity persists between the video continuity and 3D consistency. General videos typically depict objects moving along a fixed plane from the same camera viewpoint, i.e., similar information shifts across different positions on the canvas. In contrast, videos composed of images captured from multiple views often exhibit significant rotation, leading to substantial observable content jitter between frames. Here, we tailor an innovative strategy for cross-view interaction that" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.137 + ], + "angle": 0, + "content": "optimizes both performance and efficiency, ensuring robust consistency in generating dense-view images for 3D applications." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.15, + 0.605, + 0.166 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.175, + 0.854, + 0.19 + ], + "angle": 0, + "content": "3.1. Preliminaries: Latent Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.197, + 0.908, + 0.466 + ], + "angle": 0, + "content": "Our work lifts the Latent Diffusion Model (LDM) [44], which maps a noise latent \\(\\epsilon\\) and a text prompt \\(p\\) to an image \\(x\\). LDM first encodes the image \\(x_0\\) into latent space using a pre-trained encoder \\(\\mathcal{E}\\), i.e., \\(z_0 = \\mathcal{E}(x_0)\\). The latent representation \\(z_0\\) are then decoded back into the image \\(x_0\\) using a pre-trained decoder \\(\\mathcal{D}\\). The model is trained in the latent space in a self-supervised manner, and generates samples by reversing a Markov forward process. Specifically, the forward process injects noise into \\(z_0\\) to obtain noisy inputs \\(\\{z_t\\}_{t=1}^T\\), where \\(z_t = \\sqrt{\\alpha^t z_0 + \\sqrt{1 - \\alpha_t}} \\epsilon\\), and \\(\\epsilon \\sim N(0, I)\\). The generation process is formulated as a sequence of denoising operations \\(\\epsilon_\\theta(z_t, t, y)\\), predicting a denoised variant of input \\(z_t\\), where \\(y = \\phi(p)\\) is the condition, and \\(\\phi\\) is a frozen text encoder [42]. The learnable parameter \\(\\theta\\) is typically embedded within structures such as convolutional layers, self-attention mechanisms, and cross-attention mechanisms. The training objective is defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.608, + 0.468, + 0.907, + 0.49 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\theta} \\mathbb {E} _ {z _ {0}, \\epsilon , t} | | \\epsilon - \\epsilon_ {\\theta} (z _ {t}, t, y) | | _ {2} ^ {2}. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.495, + 0.907, + 0.541 + ], + "angle": 0, + "content": "The reverse process ensures that the final sample \\( z_0 \\) follows the distribution of latents, while aligning the generated image with the given condition." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.55, + 0.722, + 0.565 + ], + "angle": 0, + "content": "3.2. Framework of CoSER" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.572, + 0.909, + 0.845 + ], + "angle": 0, + "content": "We propose CoSER, a method for generating dense and consistent multiview images based on text prompts. The inputs to CoSER include a set of noisy images \\(\\{x_{t}^{i}\\}_{i = 1}^{f}\\), a set of camera parameters \\(\\{c^i\\}_{i = 1}^f\\) with a fixed elevation angle, uniformly distributed azimuth angles ranging from \\(0^{\\circ}\\) to \\(360^{\\circ}\\), and a text prompt \\(y\\). We fine-tune the lifted Latent Diffusion Model (LDM) on \\(f\\) views using a 3D dataset to synthesize multiview images \\(\\{x_0^i\\}_{i = 1}^f\\). It is non-trivial to incorporate the LDM architecture originally designed for 2D with 3D awareness. Rather than applying space-view attention or separately fusing spatial and viewpoint dimensions, our framework employs fine-grained interactions for neighboring views and coarse interactions for the global appearance. We retain the LDM structure, including ResBlock and CrossAttention, to maximize parameter reuse and maintain robust generative capabilities. A detailed sketch of CoSER is illustrated in Figure 2 and see more details for each module in Figure 3." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.852, + 0.884, + 0.867 + ], + "angle": 0, + "content": "3.2.1. Focus on Neighbors with Contextual Attention" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.908, + 0.903 + ], + "angle": 0, + "content": "It is well known that 3D objects exhibit strict cross-view consistency, whereas this does not necessitate considering" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2882" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.087, + 0.905, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.254, + 0.908, + 0.339 + ], + "angle": 0, + "content": "Figure 2. Illustration of our CoSER. Given images rendered from 12 views at the same elevation, we take a pre-trained text-to-image generation model and fine-tune it by incorporating camera poses and lifting 2D-UNet to generate multi-view images. Specifically, we achieve 3D perception by employing fine-grained learning for neighbors viewpoints and coarse-grained interactions across whole viewpoints. For neighbors, Appearance Awareness (AA) is used to learn the basic appearance and Detail Refinement (DR) is proposed to ensure neighbor consistency in point-level. For whole views, we quickly scan all viewpoints with Rapid Glance (RG) and then eliminate ambiguity with powerful Accumulated Inconsistency Rectification (AIR)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.354, + 0.482, + 0.534 + ], + "angle": 0, + "content": "the relationships among all views. For instance, the front and back views of an object are usually highly independent. Since adjacent viewpoints share more information and contribute significantly to visual consistency, we propose interacting each view with its context to mitigate the computational expense of space-view attention. Specifically, we use adjacent attention to capture the fundamental appearance and trajectory attention to refine details point-to-point, ensuring consistency between neighboring viewpoints. Contextual attention, in this case, refers to the mechanism of incorporating information from neighboring views to enhance the understanding of the current view." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.536, + 0.483, + 0.716 + ], + "angle": 0, + "content": "Appearance Awareness CoSER begins by integrating content from adjacent viewpoints to outline the basic appearance from the current view, thereby ensuring fundamental neighbor-view consistency. Following the MT2I baseline [47], we encode images \\(\\{x^i\\}\\) to latent feature \\(\\{z^i\\}\\) by the pre-trained encoder \\(\\mathcal{E}\\) (\\(t\\) is omitted for clarity), followed by mapping camera parameters \\(\\{c^i\\}\\) as high-dimensional condition. Moreover, the original self-attention is replaced with adjacent-attention, which takes the previous view \\(z^{i-1}\\), current view \\(z^i\\) and next view \\(z^{i+1}\\) as inputs, updating the features of \\(z^i\\). The formulation of the adjacent-attention is as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.104, + 0.728, + 0.483, + 0.753 + ], + "angle": 0, + "content": "\\[\nQ ^ {i} = W _ {Q} z ^ {i}, \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.104, + 0.75, + 0.452, + 0.765 + ], + "angle": 0, + "content": "\\[\nK ^ {i} = W _ {K} \\left[ z ^ {i - 1}, z ^ {i}, z ^ {i + 1} \\right], V ^ {i} = W _ {V} \\left[ z ^ {i - 1}, z ^ {i}, z ^ {i + 1} \\right],\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.483, + 0.901 + ], + "angle": 0, + "content": "where \\([\\cdot]\\) denotes concatenation operation. To inherit the generalizability of the 2D diffusion models, the parameters of self-attention is used to initialize projection matrices \\(W_{Q}, W_{K}, W_{V}\\) which are shared across space and view. Note that adjacent-frame attention, while capable of forming the initial appearance of the object, inevitably disperses attention due to the involvement of numerous interacting patches, making it struggle to focus on details." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.354, + 0.907, + 0.476 + ], + "angle": 0, + "content": "Detail Refinement To ensure cross-view consistency in rendered images, it is crucial that the same position of an object corresponds to different pixels in neighboring viewpoints. These co-located pixels across views exhibit stronger correlations than others, playing a key role in maintaining visual consistency in multiview synthesis. For an object centered at the origin, and with some slight abuse of symbols, rigid rotation about the \\(y\\)-axis follows a physical principle:" + }, + { + "type": "equation", + "bbox": [ + 0.566, + 0.484, + 0.905, + 0.514 + ], + "angle": 0, + "content": "\\[\nx ^ {\\prime} = \\left(x - \\frac {W}{2}\\right) \\cos \\Delta \\alpha + \\frac {W}{2} - d \\sin \\Delta \\alpha , \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.524, + 0.906, + 0.6 + ], + "angle": 0, + "content": "where \\((x,y)\\) and \\((x',y')\\) are the pixel coordinates before and after rotation by an angle \\(\\Delta \\alpha\\) at a depth \\(d\\), and \\(W\\) is the image width. Since the depth \\(d\\) is unknown in the generative task and \\(\\sin \\Delta \\alpha\\) becomes negligible when \\(\\Delta \\alpha\\) is small, we simplify the transformation to:" + }, + { + "type": "equation", + "bbox": [ + 0.608, + 0.609, + 0.905, + 0.64 + ], + "angle": 0, + "content": "\\[\nx ^ {\\prime} = \\left(x - \\frac {W}{2}\\right) \\cos \\Delta \\alpha + \\frac {W}{2}. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.648, + 0.906, + 0.738 + ], + "angle": 0, + "content": "To make this simplification robust in practical applications, we consider a \\(3 \\times 3\\) window \\(\\{x' + i, y' + j\\}_{i,j = -1,0,1}\\) in neighboring frames as potential results. The feature at pixel \\((x,y)\\) is then refined through the attention mechanism that interacts with these neighboring pixels, enhancing consistency across views." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.747, + 0.818, + 0.761 + ], + "angle": 0, + "content": "3.2.2. Know the Whole with Spiral Mamba" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.906, + 0.84 + ], + "angle": 0, + "content": "To efficiently enhance global consistency, we first quickly scan all viewpoints, then down-sample patches based on textual relevance to filter out irrelevant information. This allows us to utilize powerful attention mechanisms to correct any accumulated inconsistencies." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.906, + 0.902 + ], + "angle": 0, + "content": "Rapid Glance Exchanging information across all viewpoints is crucial for achieving global view consistency. However, dense attention mechanisms [28, 47] can be computationally expensive and memory-intensive, while simple" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2883" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.174, + 0.089, + 0.408, + 0.107 + ], + "angle": 0, + "content": "Appearance Awareness" + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.108, + 0.207, + 0.193 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.113, + 0.195, + 0.183, + 0.206 + ], + "angle": 0, + "content": "View i-1" + }, + { + "type": "image", + "bbox": [ + 0.231, + 0.108, + 0.345, + 0.193 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.266, + 0.195, + 0.32, + 0.206 + ], + "angle": 0, + "content": "View i" + }, + { + "type": "image", + "bbox": [ + 0.371, + 0.108, + 0.482, + 0.193 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.39, + 0.195, + 0.461, + 0.206 + ], + "angle": 0, + "content": "View i+1" + }, + { + "type": "image_caption", + "bbox": [ + 0.196, + 0.209, + 0.395, + 0.226 + ], + "angle": 0, + "content": "Detail Refinement" + }, + { + "type": "image", + "bbox": [ + 0.098, + 0.228, + 0.198, + 0.309 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.113, + 0.312, + 0.183, + 0.324 + ], + "angle": 0, + "content": "View i-1" + }, + { + "type": "image", + "bbox": [ + 0.24, + 0.229, + 0.338, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.266, + 0.312, + 0.32, + 0.324 + ], + "angle": 0, + "content": "View i" + }, + { + "type": "image", + "bbox": [ + 0.377, + 0.229, + 0.48, + 0.309 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.39, + 0.312, + 0.461, + 0.324 + ], + "angle": 0, + "content": "View \\(i + 1\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.223, + 0.327, + 0.364, + 0.345 + ], + "angle": 0, + "content": "Rapid Glance" + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.344, + 0.206, + 0.429 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.123, + 0.431, + 0.177, + 0.442 + ], + "angle": 0, + "content": "View 1" + }, + { + "type": "image", + "bbox": [ + 0.24, + 0.346, + 0.34, + 0.424 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.266, + 0.431, + 0.319, + 0.442 + ], + "angle": 0, + "content": "View i" + }, + { + "type": "image", + "bbox": [ + 0.371, + 0.344, + 0.482, + 0.429 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.394, + 0.431, + 0.458, + 0.442 + ], + "angle": 0, + "content": "View 12" + }, + { + "type": "image_caption", + "bbox": [ + 0.095, + 0.446, + 0.206, + 0.463 + ], + "angle": 0, + "content": "Accumulated" + }, + { + "type": "image_caption", + "bbox": [ + 0.211, + 0.446, + 0.346, + 0.462 + ], + "angle": 0, + "content": "Inconsistency" + }, + { + "type": "image", + "bbox": [ + 0.371, + 0.446, + 0.482, + 0.462 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.464, + 0.206, + 0.55 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.124, + 0.551, + 0.177, + 0.562 + ], + "angle": 0, + "content": "View 1" + }, + { + "type": "image", + "bbox": [ + 0.232, + 0.464, + 0.346, + 0.547 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.266, + 0.55, + 0.32, + 0.562 + ], + "angle": 0, + "content": "View i" + }, + { + "type": "image", + "bbox": [ + 0.37, + 0.464, + 0.482, + 0.55 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.395, + 0.551, + 0.457, + 0.562 + ], + "angle": 0, + "content": "View 12" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.585, + 0.483, + 0.629 + ], + "angle": 0, + "content": "Figure 3. Visualization of the interaction schemes in CoSER. For illustration, we denote each gird as a token with target token in red and attending token in green." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.644, + 0.483, + 0.885 + ], + "angle": 0, + "content": "temporal attention [74] fails to effectively learn multi-view consistency. Inspired by Gu et al. [12], we leverage state space models (SSMs) known as linear operators to process all patches of multiview images, offering a more efficient alternative to quadratic-time attention. Specifically, we utilize the advanced selective scan SSMs introduced by Mamba [11] to implement our core SSMs operator. Unlike existing Vision Mamba methods that flatten 2D images into 1D sequences row-wise or column-wise, we propose an effective spiral bidirectional scan strategy. This strategy leverages the prior knowledge that objects in 3D datasets are typically centered. By organizing object-related content within a short sequence window, our approach implicitly mitigates the forgetting problem of SSMs. Additionally, placing background content across distinct viewpoints aids the SSMs in boundary perception." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.886, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Technically, our method organizes spatial tokens start" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.168 + ], + "angle": 0, + "content": "ing from the center and expanding in a spiral pattern, then stacks them view by view. Meanwhile, we reverse the scan order within the arrangement of viewpoints to capture robust space-view features. This allows us to perform a rapid glance as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.576, + 0.176, + 0.906, + 0.196 + ], + "angle": 0, + "content": "\\[\n\\left\\{z ^ {i} \\right\\} _ {i = 1} ^ {f} = S S M s \\left(S B S c a n \\left(\\left\\{z ^ {i} \\right\\} _ {i = 1} ^ {f}\\right)\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.206, + 0.906, + 0.251 + ], + "angle": 0, + "content": "where \\( SBScan(\\cdot) \\) denotes the spiral bidirectional scan. This operator effectively aggregates global information and provides crucial clues for further disambiguation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.252, + 0.907, + 0.493 + ], + "angle": 0, + "content": "Accumulated Inconsistency Rectification To facilitate the elimination of inconsistencies through attention mechanisms, we implement down-sampling to filter out background and retain features relevant to textual information, thereby substantially reducing the number of patches engaged in interactions. Specifically, we provide a high-level semantic embedding \\(\\phi(p)\\) of the prompt \\(p\\). A 2-layer MLP \\(S\\) is employed to assess the relevance between each patch and the text, producing a score ranging from 0 to 1 at each position, where a higher score indicates stronger relevance. Notably, we do not explicitly impose supervised constraints on the scoring mechanism. Instead, the model is driven by data to autonomously perceive the regions of interest. Based on the score map \\(s^i = S([z^i, \\phi(p)])\\) as a weight, we perform 2D-pooling along the spatial dimensions on the query, key, and value maps to enable sparse interactions:" + }, + { + "type": "equation", + "bbox": [ + 0.621, + 0.503, + 0.756, + 0.522 + ], + "angle": 0, + "content": "\\[\n\\hat {Q} ^ {i} = P \\left(s ^ {i} \\cdot \\bar {Q} ^ {i}; \\tau\\right),\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.621, + 0.525, + 0.905, + 0.545 + ], + "angle": 0, + "content": "\\[\n\\hat {K} ^ {i} = P \\left(\\left\\{s ^ {j} \\cdot \\bar {K} ^ {i} \\right\\} _ {j = 1} ^ {f}; \\rho\\right), \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.624, + 0.548, + 0.797, + 0.568 + ], + "angle": 0, + "content": "\\[\n\\hat {V} ^ {i} = P (\\{s ^ {j} \\cdot \\bar {V} ^ {i} \\} _ {j = 1} ^ {f}; \\rho).\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.577, + 0.906, + 0.705 + ], + "angle": 0, + "content": "Here, \\(P\\) denotes the pooling operator, \\(\\tau\\) and \\(\\rho\\) are the strides required for pooling, and the transformations include: \\(\\bar{Q}^i = W_{\\bar{Q}}z^i\\), \\(\\bar{K}^i = W_{\\bar{K}}[\\{z^j\\}_{j=1}^f]\\), and \\(\\bar{V}^j = W_{\\bar{V}}[\\{z^j\\}_{j=1}^f]\\). Notably, in contrast to the intensive down-sampling of \\(\\hat{V}^i\\) and \\(\\hat{K}^i\\), we maintain a higher resolution for the query feature map \\(\\hat{Q}^i\\) to minimize the loss of content information. Finally, we up-sample the computed feature map using bilinear interpolation to restore its resolution." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.717, + 0.645, + 0.734 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.743, + 0.906, + 0.832 + ], + "angle": 0, + "content": "In this section, we first provide the implementation details and the dataset. We then conduct both qualitative and quantitative experiments to compare our method with previous state-of-the-art text-to-3D generation methods. Additionally, we perform an ablation study to demonstrate the effectiveness of the proposed modules." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.833, + 0.906, + 0.879 + ], + "angle": 0, + "content": "Implementation We adopt the LDM\\(^1\\) [44], pre-trained on the LAION-5B [45], as our base model. Specifically, we use the AdamW optimizer [29] with a learning rate of 3e-5 and a" + }, + { + "type": "page_footnote", + "bbox": [ + 0.531, + 0.888, + 0.88, + 0.9 + ], + "angle": 0, + "content": "1https://github.com/CompVis/stable-diffusion" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2884" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.095, + 0.898, + 0.462 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.369, + 0.463, + 0.624, + 0.475 + ], + "angle": 0, + "content": "A bear dressed in medieval armor" + }, + { + "type": "image_caption", + "bbox": [ + 0.255, + 0.501, + 0.741, + 0.516 + ], + "angle": 0, + "content": "Figure 4. Qualitative comparison of VideoMV [74] (Up) and our CoSER (Down)." + }, + { + "type": "image", + "bbox": [ + 0.098, + 0.518, + 0.903, + 0.869 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.227, + 0.892, + 0.768, + 0.907 + ], + "angle": 0, + "content": "Figure 5. Qualitative comparisons of GaussianDreamer [69], Hash3D [68] and our CoSER." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.515, + 0.957 + ], + "angle": 0, + "content": "2885" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.242 + ], + "angle": 0, + "content": "batch size of 6 to optimize the weights over approximately 200K iterations on three NVIDIA Tesla A800 GPUs. To preserve diversity and detail, we degrade CoSER to a text-to-image generation model with a \\(40\\%\\) probability and train it with a subset of LAION-5B. During inference, we employ the DDIM sampler with 50 steps and set the scale guidance to 7.5 for generation. Additionally, we set text template as \"A DSLR photo of [prompt], 3d asset\" to acquire more realistic results. Following the Liu et al. [27], we use NeuS[60] for performing reconstruction." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.244, + 0.482, + 0.365 + ], + "angle": 0, + "content": "Dataset We use the public G-Objaverse [40] as the training dataset, utilizing the names of the objects appended with \"3D asset\" as their text labels. For each object, we first normalize the object to be centered within a bounding box ranging from \\([-0.5, 0.5]\\). The camera distance is set to 2.0 units. A random elevation is chosen between \\([-30^\\circ, 30^\\circ]\\), and starting from the front view, we uniformly sample 12 angles around the object for rendering." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.378, + 0.382, + 0.394 + ], + "angle": 0, + "content": "4.1. Comparison with SOTA methods" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.401, + 0.483, + 0.703 + ], + "angle": 0, + "content": "Qualitative Comparison We primarily evaluate the performance of the proposed CoSER on multiview text-to-image and compare the results with state-of-the-art methods in Figure 6. For VideoMV [74], 12 images generated from the same camera viewpoints as CoSER are selected for visualization, and all 24 viewpoints were utilized in the reconstruction process to ensure a fair comparison. Qualitatively, our method demonstrates impressive quality improvement in terms of both consistence and texture. Although VideoMV can generate images from a broader range of perspectives, e.g., mask and watch, it significantly compromises cross-view consistency, which poses serious challenges for subsequent 3D reconstruction. Furthermore, Figure 5 presents a comparison between two state-of-the-art text-to-3D methods, including GaussianDreamer [69] and Hash3D [68]. Specifically, the advanced method, Dream-Gaussian [54], is utilized as the foundational model for Hash3D. Thanks to the dense consistent images, CoSER outperforms other methods in terms of fidelity in 3D geometry and visual appearance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Quantitative Comparison Following previous work [74], we quantitatively assess the quality of the generated 3D assets using CLIP Score [14]. To evaluate the textual alignment, we followed [10] to compute the recall of the original prompt within the generated caption. We employ 100 text prompts selected from T3Bench [13] to generate 3D models. For each object, we set the elevation as \\(15^{\\circ}\\) and select 120 evenly spaced azimuth angles from \\(-180^{\\circ}\\) to \\(180^{\\circ}\\), thereby generating 120 rendered images from different viewpoints. During the evaluation, 30 from the 120 rendered images are randomly selected and the CLIP Score is computed as the average semantic similarity between the prompts and these images. We use the CLIP ViT-B [43] to" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.38 + ], + "angle": 0, + "content": "extract both text and image features. The results, presented in Table 1, demonstrate that our method significantly surpasses other methods, indicating that CoSER is more adept at injecting semantic information into the creation of realistic 3D assets. Given the challenges in automating the measurement of 3D consistency, we resort to user studies to assess which method produces results that are most favored by human evaluators. We conducted three user studies on the results in terms of the Semantic Alignment, Multiview Consistency and Texture Detail. In the first study, participants were asked to select the 3D asset that best aligns with the semantics of the given prompt. In the second study, participants identified the asset with the highest consistency. In the third study, participants were encouraged to choose the object with the finest texture and less distortion artifacts. We receive 100 answers on 30 text-3D pairs in total for each study. As shown in Table 1, over \\(29\\%\\) of our results are selected as the best in both three metrics, which proves a significant advantage in 3D generation." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.393, + 0.667, + 0.409 + ], + "angle": 0, + "content": "4.2. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.416, + 0.905, + 0.793 + ], + "angle": 0, + "content": "Effect of Module Design We are mainly interested in the 3D-aware module design, given that other unit has been well explored in previous methods. The core of our design is focusing on neighbors in fine-grained manner and knowing whole views with coarse-grained mechanism. To substantiate these claims, we incrementally integrated the proposed modules into four distinct configurations and finetuned the LDM accordingly. As depicted in Figure 6, the use of Appearance Awareness alone proves insufficient for ensuring cross-view consistency. However, the incorporation of Detail Refinement enhances the alignment between adjacent views, thereby improving overall 3D content consistency. It is evident that Rapid Glance offers small reduction in ambiguities; nonetheless, its primary benefit lies in facilitating information exchange across all viewpoints (See Figure 8), thereby establishing a foundation for the effective Accumulated Inconsistency Rectification module. Additionally, we present the parameters and MACs introduced by each module in Table 2. To illustrate that SSM captures multi-view consistency more effectively than temporal attention, we also learn global consistency by temporal attention found that it impaired performance. Based on this observation, we consider the additional 3.31M parameters and 15G MACs computations introduced by RG and AIR to be entirely acceptable." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Spiral Bidirectional Scan in Rapid Glance To verify the effectiveness of the proposed spiral bidirectional scan strategy, we perform the evaluation by replacing it with vanilla spatial-first bidirectional scan in VideoMamba [22]. As shown in Figure 7, both scanning strategies enhance the consistency. However, spatial-first bidirectional scan inevitably disrupts the continuity of object, which hampers" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "2886" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.101, + 0.09, + 0.592, + 0.117 + ], + "angle": 0, + "content": "Table 1. Quantitative comparison of CoSER and state-of-the-arts based on the Quality, Alignment and User Study." + }, + { + "type": "table", + "bbox": [ + 0.102, + 0.119, + 0.622, + 0.194 + ], + "angle": 0, + "content": "
ModelQuality↑Alignment↑User Study
Semantic↑Consistency↑Texture↑
GaussianDreamer [69]30.7833.70.260.270.21
Hash3D [68]31.2630.20.200.190.25
VideoMV [74]32.1934.40.230.250.22
Ours33.0737.70.310.290.32
" + }, + { + "type": "table_caption", + "bbox": [ + 0.632, + 0.095, + 0.894, + 0.122 + ], + "angle": 0, + "content": "Table 2. Model efficiency of each module. TA denotes temporal attention." + }, + { + "type": "table", + "bbox": [ + 0.631, + 0.124, + 0.918, + 0.189 + ], + "angle": 0, + "content": "
ModuleParameters (M)MACs (G)
AA867.451036
AA + DR917.041395
AA + DR + RG917.071396
AA + DR + RG + AIR969.951487
AA + DR + TA966.641472
" + }, + { + "type": "image", + "bbox": [ + 0.101, + 0.198, + 0.896, + 0.349 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.356, + 0.905, + 0.384 + ], + "angle": 0, + "content": "Figure 6. Ablation of proposed moulds. The arrows indicate the changes made to the Appearance Awareness by sequentially plusing Detail Refinement, Rapid Glance, and Accumulated Inconsistency Rectification. Better see in color and \\(2\\mathrm{x}\\) zoom." + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.387, + 0.48, + 0.493 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.502, + 0.483, + 0.529 + ], + "angle": 0, + "content": "Figure 7. Ablation of scanning strategy. Spiral bidirectional scan generates semantic- and view-consistent images." + }, + { + "type": "image", + "bbox": [ + 0.098, + 0.531, + 0.477, + 0.653 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.664, + 0.483, + 0.706 + ], + "angle": 0, + "content": "Figure 8. Ablation of score map. Score map aids in capturing critical information, while Rapid Glance contributes to the generation of more accurate score maps." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.705, + 0.483, + 0.87 + ], + "angle": 0, + "content": "the efficiency of SSM model for feature aggregation. In contrast, our proposed spiral bidirectional scan maximizes the spatial compactness of foreground content while simultaneously enforcing operator awareness of inter-view boundaries, thereby achieving better results. Additionally, we experimented with reversing the spiral scanning order within the image, progressing from the periphery inward towards the center. However, this strategy did not yield further performance improvements, which we attribute to the already efficient information aggregation within an image achieved by the spiral scan." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Score Map in Accumulated Inconsistency Rectification In Accumulated Inconsistency Rectification, we introduce the" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.4, + 0.907, + 0.596 + ], + "angle": 0, + "content": "score map to assist in downsampling feature resolution for coarse attention, tailored to each individual view. To understand its impact, we compared the full AIR module generated results with those from two alternative models: AIR without score map and AIR with score map but without Rapid Glance. For clearness, we extract score maps at various resolutions and upsample them to 32x32 for visualization. As shown in Figure 8, the score map assigns higher scores to foreground regions, while background areas receive lower scores, facilitating the retention of critical information during the downsampling process. Additionally, Rapid Glance enhances the exchange of information, contributing to the generation of more precise score maps." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.603, + 0.634, + 0.618 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.629, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In this paper, we introduced CoSER, a novel framework for generating consistent and high-fidelity dense multiview images from text prompts. Our approach leverages a hybrid architecture that encapsulates the strengths of attention mechanisms and state space models (SSMs) to address the crucial challenges of cross-view consistency. The innovative design of contextual interaction includes an adjacent attention for capturing rough structures and textures, a trajectory-based attention for refining local details. A spiral bidirectional Mamba ensures global coherence in a cost-efficient manner. This comprehensive approach allows CoSER to produce dense, high-fidelity, and content-consistent multiview images. Extensive experiments demonstrate that CoSER outperforms state-of-the-art multiview synthesis methods in both quantitative and qualitative evaluations across various datasets. Future work could further explore the hybrid way of neighbor and global interaction to efficiently generate dense and consistent multiview images." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2887" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.26, + 0.108 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.484, + 0.221 + ], + "angle": 0, + "content": "This paper is supported by the Beijing Natural Science Foundation (1254050), National Natural Science Foundation of China under Grants (U23B2012, 12471308), Fundamental Research Funds for the Central Universities, and by the National Research Foundation, Singapore, under its Medium Sized Center for Advanced Robotics Technology Innovation." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.236, + 0.188, + 0.252 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.261, + 0.484, + 0.317 + ], + "angle": 0, + "content": "[1] Haoyu Chen, Jingjing Ren, Jinjin Gu, Hongtao Wu, Xuequan Lu, Haoming Cai, and Lei Zhu. Snow removal in video: A new dataset and a novel method. in 2023 IEEE. In ICCV, pages 13165-13176. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.319, + 0.484, + 0.374 + ], + "angle": 0, + "content": "[2] Haoyu Chen, Jinjin Gu, Yihao Liu, Salma Abdel Magid, Chao Dong, Qiong Wang, Hanspeter Pfister, and Lei Zhu. Masked image training for generalizable deep image denoising. In CVPR, pages 1692-1703, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.376, + 0.484, + 0.431 + ], + "angle": 0, + "content": "[3] Rui Chen, Yongwei Chen, Ningxin Jiao, and Kui Jia. Fantasia3d: Disentangling geometry and appearance for high-quality text-to-3d content creation. In ICCV, pages 22246–22256, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.433, + 0.484, + 0.488 + ], + "angle": 0, + "content": "[4] Yongwei Chen, Rui Chen, Jiabao Lei, Yabin Zhang, and Kui Jia. Tango: Text-driven photorealistic and robust 3d stylization via lighting decomposition. In NeurIPS, pages 30923-30936, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.49, + 0.484, + 0.558 + ], + "angle": 0, + "content": "[5] Yiwen Chen, Zilong Chen, Chi Zhang, Feng Wang, Xiaofeng Yang, Yikai Wang, Zhongang Cai, Lei Yang, Huaping Liu, and Guosheng Lin. Gaussianeditor: Swift and controllable 3d editing with gaussian splatting. In CVPR, pages 21476-21485, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.561, + 0.484, + 0.603 + ], + "angle": 0, + "content": "[6] Zilong Chen, Yikai Wang, Feng Wang, Zhengyi Wang, and Huaping Liu. V3d: Video diffusion models are effective 3d generators. arXiv preprint arXiv:2403.06738, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.604, + 0.484, + 0.659 + ], + "angle": 0, + "content": "[7] Yen-Chi Cheng, Hsin-Ying Lee, Sergey Tulyakov, Alexander G Schwing, and Liang-Yan Gui. Sdfusion: Multimodal 3d shape completion, reconstruction, and generation. In CVPR, pages 4456-4465, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.661, + 0.484, + 0.716 + ], + "angle": 0, + "content": "[8] Lihe Ding, Shaocong Dong, Zhanpeng Huang, Zibin Wang, Yiyuan Zhang, Kaixiong Gong, Dan Xu, and Tianfan Xue. Text-to-3d generation with bidirectional diffusion using both 2d and 3d priors. In CVPR, pages 5115–5124, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.718, + 0.484, + 0.786 + ], + "angle": 0, + "content": "[9] Ruiqi Gao, Aleksander Holynski, Philipp Henzler, Arthur Brussee, Ricardo Martin-Brualla, Pratul Srinivasan, Jonathan T Barron, and Ben Poole. Cat3d: Create anything in 3d with multi-view diffusion models. arXiv preprint arXiv:2405.10314, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.789, + 0.484, + 0.858 + ], + "angle": 0, + "content": "[10] Chongjian Ge, Chenfeng Xu, Yuanfeng Ji, Chensheng Peng, Masayoshi Tomizuka, Ping Luo, Mingyu Ding, Varun Jampani, and Wei Zhan. Compgs: Unleashing 2d compositionality for compositional text-to-3d via dynamically optimizing 3d gaussians. arXiv preprint arXiv:2410.20723, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.86, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[11] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023. 5" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.261, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "[12] Albert Gu, Karan Goel, and Christopher Re. Efficiently modeling long sequences with structured state spaces. In ICLR, 2022. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.136, + 0.905, + 0.192 + ], + "angle": 0, + "content": "[13] Yuze He, Yushi Bai, Matthieu Lin, Wang Zhao, Yubin Hu, Jenny Sheng, Ran Yi, Juanzi Li, and Yong-Jin Liu. T3 bench: Benchmarking current progress in text-to-3d generation. arXiv preprint arXiv:2310.02977, 2023. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.193, + 0.905, + 0.247 + ], + "angle": 0, + "content": "[14] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning. In EMNLP, pages 7514-7528, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.25, + 0.905, + 0.317 + ], + "angle": 0, + "content": "[15] Fangzhou Hong, Jiaxiang Tang, Ziang Cao, Min Shi, Tong Wu, Zhaoxi Chen, Tengfei Wang, Liang Pan, Dahua Lin, and Ziwei Liu. 3dtopia: Large text-to-3d generation model with hybrid diffusion priors. arXiv preprint arXiv:2403.02234, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.321, + 0.905, + 0.375 + ], + "angle": 0, + "content": "[16] Yukun Huang, Jianan Wang, Yukai Shi, Xianbiao Qi, Zheng-Jun Zha, and Lei Zhang. Dreamtime: An improved optimization strategy for text-to-3d content creation. arXiv preprint arXiv:2306.12422, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.378, + 0.905, + 0.446 + ], + "angle": 0, + "content": "[17] Zehuan Huang, Hao Wen, Junting Dong, Yaohui Wang, Yangguang Li, Xinyuan Chen, Yan-Pei Cao, Ding Liang, Yu Qiao, Bo Dai, et al. Epidiff: Enhancing multi-view synthesis via localized epipolar-constrained diffusion. In CVPR, pages 9784-9794, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.448, + 0.905, + 0.502 + ], + "angle": 0, + "content": "[18] Ajay Jain, Ben Mildenhall, Jonathan T Barron, Pieter Abbeel, and Ben Poole. Zero-shot text-guided object generation with dream fields. In CVPR, pages 867-876, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.504, + 0.905, + 0.56 + ], + "angle": 0, + "content": "[19] Jeong-gi Kwak, Erqun Dong, Yuhe Jin, Hanseok Ko, Shweta Mahajan, and Kwang Moo Yi. Vivid-1-to-3: Novel view synthesis with video diffusion models. In CVPR, pages 6775–6785, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.561, + 0.905, + 0.603 + ], + "angle": 0, + "content": "[20] Bonan Li, Zicheng Zhang, Xuecheng Nie, Congying Han, Yinhan Hu, and Tiande Guo. Styo: Stylize your face in only one-shot. arXiv preprint arXiv:2303.03231, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.604, + 0.905, + 0.673 + ], + "angle": 0, + "content": "[21] Jiahao Li, Hao Tan, Kai Zhang, Zexiang Xu, Fujun Luan, Yinghao Xu, Yicong Hong, Kalyan Sunkavalli, Greg Shakhnarovich, and Sai Bi. Instant3d: Fast text-to-3d with sparse-view generation and large reconstruction model. In ICLR, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.676, + 0.905, + 0.731 + ], + "angle": 0, + "content": "[22] Kunchang Li, Xinhao Li, Yi Wang, Yinan He, Yali Wang, Limin Wang, and Yu Qiao. Videomamba: State space model for efficient video understanding. arXiv preprint arXiv:2403.06977, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.733, + 0.905, + 0.801 + ], + "angle": 0, + "content": "[23] Peng Li, Yuan Liu, Xiaoxiao Long, Feihu Zhang, Cheng Lin, Mengfei Li, Xingqun Qi, Shanghang Zhang, Wenhan Luo, Ping Tan, et al. Era3d: High-resolution multiview diffusion using efficient row-wise attention. arXiv preprint arXiv:2405.11616, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.803, + 0.905, + 0.845 + ], + "angle": 0, + "content": "[24] Weiyu Li, Rui Chen, Xuelin Chen, and Ping Tan. Sweetdreamer: Aligning geometric priors in 2d diffusion for consistent text-to-3d. In ICLR, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[25] Wenbing Li, Hang Zhou, Junqing Yu, Zikai Song, and Wei Yang. Coupled mamba: Enhanced multimodal fusion with coupled state space model. In NeurIPS, pages 59808-59832, 2025. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "2888" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.161 + ], + "angle": 0, + "content": "[26] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. In CVPR, pages 300–309, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.163, + 0.482, + 0.219 + ], + "angle": 0, + "content": "[27] Yuan Liu, Cheng Lin, Zijiao Zeng, Xiaoxiao Long, Lingjie Liu, Taku Komura, and Wenping Wang. Syncdreamer: Generating multiview-consistent images from a single-view image. In ICLR, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.221, + 0.482, + 0.289 + ], + "angle": 0, + "content": "[28] Xiaoxiao Long, Yuan-Chen Guo, Cheng Lin, Yuan Liu, Zhiyang Dou, Lingjie Liu, Yuexin Ma, Song-Hai Zhang, Marc Habermann, Christian Theobalt, et al. Wonder3d: Single image to 3d using cross-domain diffusion. In CVPR, pages 9970-9980, 2024. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.291, + 0.482, + 0.319 + ], + "angle": 0, + "content": "[29] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In ICLR, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.321, + 0.482, + 0.361 + ], + "angle": 0, + "content": "[30] Run Luo, Zikai Song, Lintao Ma, Jinlin Wei, Wei Yang, and Min Yang. Diffusiontrack: Diffusion model for multi-object tracking. In AAAI, pages 3991-3999, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.362, + 0.482, + 0.418 + ], + "angle": 0, + "content": "[31] Baorui Ma, Haoge Deng, Junsheng Zhou, Yu-Shen Liu, Tiejun Huang, and Xinlong Wang. Geodream: Disentangling 2d and geometric priors for high-fidelity and consistent 3d generation. arXiv preprint arXiv:2311.17971, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.42, + 0.482, + 0.475 + ], + "angle": 0, + "content": "[32] Luke Melas-Kyriazi, Iro Laina, Christian Rupprecht, Natalia Neverova, Andrea Vedaldi, Oran Gafni, and Filippos Kokkinos. Im-3d: Iterative multiview diffusion and reconstruction for high-quality 3d generation. In ICML, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.477, + 0.482, + 0.531 + ], + "angle": 0, + "content": "[33] Gal Metzer, Elad Richardson, Or Patashnik, Raja Giryes, and Daniel Cohen-Or. Latent-nerf for shape-guided generation of 3d shapes and textures. In CVPR, pages 12663–12673, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.533, + 0.482, + 0.586 + ], + "angle": 0, + "content": "[34] Gal Metzer, Elad Richardson, Or Patashnik, Raja Giryes, and Daniel Cohen-Or. Latent-nerf for shape-guided generation of 3d shapes and textures. In CVPR, pages 12663–12673, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.59, + 0.482, + 0.645 + ], + "angle": 0, + "content": "[35] Nasir Mohammad Khalid, Tianhao Xie, Eugene Belilovsky, and Tiberiu Popa. Clip-mesh: Generating textured meshes from text using pretrained image-text models. In SIGGRAPH Asia, pages 1-8, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.647, + 0.482, + 0.701 + ], + "angle": 0, + "content": "[36] Alex Nichol, Heewoo Jun, Prafulla Dhariwal, Pamela Mishkin, and Mark Chen. Point-e: A system for generating 3d point clouds from complex prompts. arXiv preprint arXiv:2212.08751, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.703, + 0.482, + 0.744 + ], + "angle": 0, + "content": "[37] JangHo Park, Gihyun Kwon, and Jong Chul Ye. Ed-nerf: Efficient text-guided editing of 3d scene with latent space nerf. In ICLR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.746, + 0.482, + 0.786 + ], + "angle": 0, + "content": "[38] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. In ICLR, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.789, + 0.482, + 0.87 + ], + "angle": 0, + "content": "[39] Guocheng Qian, Jinjie Mai, Abdullah Hamdi, Jian Ren, Aliaksandr Siarohin, Bing Li, Hsin-Ying Lee, Ivan Skorokhodov, Peter Wonka, Sergey Tulyakov, et al. Magic123: One image to high-quality 3d object generation using both 2d and 3d diffusion priors. arXiv preprint arXiv:2306.17843, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.482, + 0.902 + ], + "angle": 0, + "content": "[40] Lingteng Qiu, Guanying Chen, Xiaodong Gu, Qi Zuo, Mutian Xu, Yushuang Wu, Weihao Yuan, Zilong Dong, Liefeng" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "Bo, and Xiaoguang Han. Richdreamer: A generalizable normal-depth diffusion model for detail richness in text-to-3d. In CVPR, pages 9914–9925, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[41] Xinmin Qiu, Congying Han, ZiCheng Zhang, Bonan Li, Tiande Guo, and Xuecheng Nie. Diffbfr: Bootstrapping diffusion model for blind face restoration. In ACM MM, pages 7785-7795, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.193, + 0.905, + 0.262 + ], + "angle": 0, + "content": "[42] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.264, + 0.905, + 0.332 + ], + "angle": 0, + "content": "[43] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.334, + 0.905, + 0.389 + ], + "angle": 0, + "content": "[44] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10684-10695, 2022. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.391, + 0.905, + 0.5 + ], + "angle": 0, + "content": "[45] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, Patrick Schramowski, Srivatsa Kundurthy, Katherine Crowson, Ludwig Schmidt, Robert Kaczmarczyk, and Jenia Jitsev. Laion-5b: An open large-scale dataset for training next generation image-text models. ArXiv, abs/2210.08402, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.503, + 0.905, + 0.571 + ], + "angle": 0, + "content": "[46] Junyoung Seo, Wooseok Jang, Min-Seop Kwak, Hyeonsu Kim, Jaehoon Ko, Junho Kim, Jin-Hwa Kim, Jiyoung Lee, and Seungryong Kim. Let 2d diffusion model know 3d-consistency for robust text-to-3d generation. In ICLR, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.574, + 0.905, + 0.614 + ], + "angle": 0, + "content": "[47] Yichun Shi, Peng Wang, Jianglong Ye, Long Mai, Kejie Li, and Xiao Yang. Mvdream: Multi-view diffusion for 3d generation. In ICLR, 2023. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.617, + 0.905, + 0.657 + ], + "angle": 0, + "content": "[48] Zikai Song, Junqing Yu, Yi-Ping Phoebe Chen, and Wei Yang. Transformer tracking with cyclic shifting window attention. In CVPR, pages 8791-8800, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.66, + 0.905, + 0.701 + ], + "angle": 0, + "content": "[49]Zikai Song, Run Luo, Junqing Yu, Yi-Ping Phoebe Chen, and Wei Yang. Compact transformer tracker with correlative masked modeling. In AAAI, pages 2321-2329, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.703, + 0.905, + 0.757 + ], + "angle": 0, + "content": "[50] Zikai Song, Ying Tang, Run Luo, Lintao Ma, Junqing Yu, Yi-Ping Phoebe Chen, and Wei Yang. Autogenic language embedding for coherent point tracking. In ACM MM, pages 2021-2030, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.76, + 0.905, + 0.815 + ], + "angle": 0, + "content": "[51] Haoze Sun, Wenbo Li, Jianzhuang Liu, Haoyu Chen, Renjing Pei, Xueyi Zou, Youliang Yan, and Yujiu Yang. Coser: Bridging image and language for cognitive super-resolution. In CVPR, pages 25868-25878, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.817, + 0.905, + 0.87 + ], + "angle": 0, + "content": "[52] Jingxiang Sun, Bo Zhang, Ruizhi Shao, Lizhen Wang, Wen Liu, Zhenda Xie, and Yebin Liu. Dreamcraft3d: Hierarchical 3d generation with bootstrapped diffusion prior. In ICLR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.902 + ], + "angle": 0, + "content": "[53] Junshu Tang, Tengfei Wang, Bo Zhang, Ting Zhang, Ran Yi, Lizhuang Ma, and Dong Chen. Make-it-3d: High-fidelity 3d" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "2889" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.12 + ], + "angle": 0, + "content": "creation from a single image with diffusion prior. In ICCV, pages 22819-22829, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.162 + ], + "angle": 0, + "content": "[54] Jiaxiang Tang, Jiawei Ren, Hang Zhou, Ziwei Liu, and Gang Zeng. Dreamgaussian: Generative gaussian splatting for efficient 3d content creation. In ICLR, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.165, + 0.482, + 0.219 + ], + "angle": 0, + "content": "[55] Shitao Tang, Fuyang Zhang, Jiacheng Chen, Peng Wang, and Yasutaka Furukawa. Mvdiffusion: Enabling holistic multiview image generation with correspondence-aware diffusion. arXiv, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.221, + 0.482, + 0.276 + ], + "angle": 0, + "content": "[56] Christina Tsalicoglou, Fabian Manhardt, Alessio Tonioni, Michael Niemeyer, and Federico Tombari. Textmesh: Generation of realistic 3d meshes from text prompts. In 3DV, pages 1554-1563, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.278, + 0.482, + 0.319 + ], + "angle": 0, + "content": "[57] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.321, + 0.482, + 0.401 + ], + "angle": 0, + "content": "[58] Vikram Voleti, Chun-Han Yao, Mark Boss, Adam Letts, David Pankratz, Dmitry Tochilkin, Christian Laforte, Robin Rombach, and Varun Jampani. Sv3d: Novel multi-view synthesis and 3d generation from a single image using latent video diffusion. arXiv preprint arXiv:2403.12008, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.405, + 0.482, + 0.445 + ], + "angle": 0, + "content": "[59] Peng Wang and Yichun Shi. Imagedream: Image-prompt multi-view diffusion for 3d generation. arXiv preprint arXiv:2312.02201, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.448, + 0.482, + 0.502 + ], + "angle": 0, + "content": "[60] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. In NeurIPS, 2021. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.504, + 0.482, + 0.572 + ], + "angle": 0, + "content": "[61] Tengfei Wang, Bo Zhang, Ting Zhang, Shuyang Gu, Jianmin Bao, Tadas Baltrusaitis, Jingjing Shen, Dong Chen, Fang Wen, Qifeng Chen, et al. Rodin: A generative model for sculpting 3d digital avatars using diffusion. In CVPR, pages 4563-4573, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.575, + 0.482, + 0.644 + ], + "angle": 0, + "content": "[62] Xinzhou Wang, Yikai Wang, Junliang Ye, Zhengyi Wang, Fuchun Sun, Pengkun Liu, Ling Wang, Kai Sun, Xintong Wang, and Bin He. Animatabledreamer: Text-guided non-rigid 3d model generation and reconstruction with canonical score distillation. arXiv preprint arXiv:2312.03795, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.645, + 0.482, + 0.7 + ], + "angle": 0, + "content": "[63] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. In NeurIPS, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.702, + 0.482, + 0.743 + ], + "angle": 0, + "content": "[64] Zeyu Xiao and Zhiwei Xiong. Incorporating degradation estimation in light field spatial super-resolution. Computer Vision and Image Understanding, page 104295, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.745, + 0.482, + 0.786 + ], + "angle": 0, + "content": "[65] Zeyu Xiao, Yutong Liu, Ruisheng Gao, and Zhiwei Xiong. Cutmib: Boosting light field super-resolution via multi-view image blending. In CVPR, pages 1672-1682, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.788, + 0.482, + 0.829 + ], + "angle": 0, + "content": "[66] Zeyu Xiao, Dachun Kai, Yueyi Zhang, Xiaoyan Sun, and Zhiwei Xiong. Asymmetric event-guided video superresolution. In ACM MM, pages 2409-2418, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.831, + 0.482, + 0.857 + ], + "angle": 0, + "content": "[67] Xinchao Wang Xingyi Yang. Kolmogorov-arnold transformer. In ICLR, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.859, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[68] Xingyi Yang and Xinchao Wang. Hash3d: Training-free acceleration for 3d generation. arXiv preprint arXiv:2404.06091, 2024. 3, 6, 7, 8" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.162 + ], + "angle": 0, + "content": "[69] Taoran Yi, Jiemin Fang, Junjie Wang, Guanjun Wu, Lingxi Xie, Xiaopeng Zhang, Wenyu Liu, Qi Tian, and Xinggang Wang. Gaussian dreamer: Fast generation from text to 3d gaussians by bridging 2d and 3d diffusion models. In CVPR, pages 6796-6807, 2024. 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.164, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[70] Weihao Yu and Xinchao Wang. Mambaout: Do we really need mamba for vision? In CVPR, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.192, + 0.905, + 0.247 + ], + "angle": 0, + "content": "[71] Xianfang Zeng, Xin Chen, Zhongqi Qi, Wen Liu, Zibo Zhao, Zhibin Wang, Bin Fu, Yong Liu, and Gang Yu. Paint3d: Paint anything 3d with lighting-less texture diffusion models. In CVPR, pages 4252-4262, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.249, + 0.905, + 0.303 + ], + "angle": 0, + "content": "[72] Zicheng Zhang, Bonan Li, Xuecheng Nie, Congying Han, Tiande Guo, and Luoqi Liu. Towards consistent video editing with text-to-image diffusion models. In NeurIPS, pages 58508-58519, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.305, + 0.905, + 0.347 + ], + "angle": 0, + "content": "[73] Jingyu Zhuang, Chen Wang, Liang Lin, Lingjie Liu, and Guanbin Li. Dreameditor: Text-driven 3d scene editing with neural fields. In SIGGRAPH Asia, pages 1-10, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.349, + 0.905, + 0.416 + ], + "angle": 0, + "content": "[74] Qi Zuo, Xiaodong Gu, Lingteng Qiu, Yuan Dong, Zhengyi Zhao, Weihao Yuan, Rui Peng, Siyu Zhu, Zilong Dong, Liefeng Bo, et al. Videomv: Consistent multi-view generation based on large video generative model. arXiv preprint arXiv:2403.12010, 2024. 2, 3, 5, 6, 7, 8" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.416 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "2890" + } + ] +] \ No newline at end of file diff --git a/2025/CoSER_ Towards Consistent Dense Multiview Text-to-Image Generator for 3D Creation/911def72-ec15-4b90-8a72-6946e2b20044_origin.pdf b/2025/CoSER_ Towards Consistent Dense Multiview Text-to-Image Generator for 3D Creation/911def72-ec15-4b90-8a72-6946e2b20044_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..25726de1f5e37992d02d4dc35c3ba1d2c6b3642c --- /dev/null +++ b/2025/CoSER_ Towards Consistent Dense Multiview Text-to-Image Generator for 3D Creation/911def72-ec15-4b90-8a72-6946e2b20044_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9e97b4bd76c7b4c293b40962154517e54efa76416c78f0bd5da93120a77be5da +size 7930615 diff --git a/2025/CoSER_ Towards Consistent Dense Multiview Text-to-Image Generator for 3D Creation/full.md b/2025/CoSER_ Towards Consistent Dense Multiview Text-to-Image Generator for 3D Creation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..a77563c161f5286d74ec22889e5d908b7feebd70 --- /dev/null +++ b/2025/CoSER_ Towards Consistent Dense Multiview Text-to-Image Generator for 3D Creation/full.md @@ -0,0 +1,324 @@ +# CoSER: Towards Consistent Dense Multiview Text-to-Image Generator for 3D Creation + +Bonan Li $^{1,2*}$ Zicheng Zhang $^{1*}$ Xingyi Yang $^{2}$ Xinchao Wang $^{2\dagger}$ + +$^{1}$ University of Chinese Academy of Sciences $^{2}$ National University of Singapore + +libonacci@ucas.ac.cn zhangzicheng19@mails.ucas.ac.cn xyang@u.nus.edu xinchao@nus.edu.sg + +![](images/708d74411269255e50a089ac20aad4efc83523fe4cbbf99bc2ae0b88c00a6d8f.jpg) +Figure 1. CoSER aims to generate detailed, diverse and detaild 3D objects from text prompts. All prompts are from T3Bench [13] and the templated is set as "A DSLR photo of ..., 3d asset". Better see in color and 2x zoom. + +# Abstract + +Generating dense multiview images from text prompts is crucial for creating high-fidelity 3D assets. Nevertheless, existing methods struggle with space-view correspondences, resulting in sparse and low-quality outputs. In this paper, we introduce CoSER, a novel consistent dense Multiview Text-to-Image Generator for Text-to-3D, achieving both efficiency and quality by meticulously learning neighbor-view coherence and further alleviating ambiguity through the swift traversal of all views. For achieving neighbor-view consistency, each viewpoint densely interacts with adjacent viewpoints to perceive the global spatial structure, and aggregates information along motion paths explicitly defined by physical principles to refine details. To further enhance cross-view consistency and alleviate content drift, CoSER rapidly scan all views in spiral bidirectional manner to aware holistic information and then scores each point based on semantic material. Subsequently, we + +conduct weighted down-sampling along the spatial dimension based on scores, thereby facilitating prominent information fusion across all views with lightweight computation. Technically, the core module is built by integrating the attention mechanism with a selective state space model, exploiting the robust learning capabilities of the former and the low overhead of the latter. Extensive evaluation shows that CoSER is capable of producing dense, high-fidelity, content-consistent multiview images that can be flexibly integrated into various 3D generation models. + +# 1. Introduction + +Creating 3D content from text is an important task in computer vision and graphics for its potential applications in advertising, robotics, digital games and the meta-verse. Previous works [3, 26, 33] primarily optimise 3D representations using score distillation sampling [38] of Text-to-Image (T2I) models, ensuring that rendered images from any viewpoint maintain high fidelity while aligning with + +the given prompt. Nevertheless, these methods suffer from Janus problems and severe artifacts with asymmetric geometry. As a remedy, multiview text-to-image (MT2I) models [6, 17, 19, 21, 23, 32] have been developed, providing rich 3D priors that help mitigate these problems. Furthermore, using dense views can accelerate the modeling process and enhance the quality of the resulting 3D models. + +To acquire MT2I models, a promising approach is to leverage 3D priors learned from 3D datasets to enhance 2D diffusion models. Seminal works [28, 47] have focused on inflating T2I models with dense cross-view attention to capture 3D information. However, this approach incurs expensive computation and faces challenges when dealing with dense viewpoints. To address these limitations, sparse attention mechanisms [23] have been proposed to model cross-view dependencies, though they can introduce inconsistencies across multiple views. Recent studies [9, 58, 74] have leveraged cross-frame continuity priors from text-to-video (T2V) models to enhance consistency. Nevertheless, azimuthal rotation around an object can cause significant content variations at the same image coordinates, making temporal attention insufficient for capturing intricate details necessary for 3D generation. + +Despite impressive progress in multiview T2I generation, effectively and efficiently modeling the space-view structure within a unified architecture remains an open challenge. In this paper, we advance the field by addressing two fundamental problems: (1) How to design efficient cross-view interactions to ensure dense multiview consistency? Rather than directly optimizing consistency among all views with uniform granularity, we propose a two-step approach: (i) Learning neighbor-view consistency as a prerequisite, as enforcing consistency between adjacent viewpoints naturally enables coherence across all viewpoints. (ii) Ensuring global-view coherence to mitigate discrepancies between distant views and facilitate unified appearance, since subtle inconsistencies are inevitable and deteriorated as the viewing angles increases. (2) Which operator should be applied for the implementation? Unlike previous works that employ attention as the default choice, we leverage hybrid operators, including attention [57] and state space models (SSMs) [12], to harness the benefits of both model families, e.g., the powerful capabilities of attention and the low computational cost of SSMs. + +In this paper, we address these questions and propose a novel model named CoSER, which enhances consistency in dense MT2I tasks. CoSER utilizes "factor" interactions that focus on neighboring views and maintain awareness of the global view. Technically, CoSER initially employs dense attention on adjacent views to capture the rough structure and texture of the current plane, thereby constructing a foundational appearance. Subsequently, we introduce a trajectory-based module to refine local details by accurately tracing + +3D trajectories and enhancing point-to-point consistency. Specifically, motivated by the efficiency of calculating rotated locations for centralized objects, we develop a robust method to handle complicate cases. This involves pixel interaction within a window of neighboring frames defined by physical rotation rules, ensuring precise alignment and consistency across views. To further alleviate disparities across arbitrary viewpoints, we utilize state space models (SSMs) equipped with a novel spiral bidirectional scan strategy to aggregate features from the entire set of views. We then score each patch based on textual inputs and perform down-sampling conditioned on these scores to adaptively ignore semantically irrelevant information. This approach allows CoSER to efficiently maintain global context through sparse attention. With these designs, our method establishes robust priors, enabling the creation of exceptionally coherent 3D assets (see Figure 1). In summary, our contributions can be summarized as follows: + +- We propose a novel multiview text-to-image framework, CoSER, for 3D content creation which exhibits significantly consistency behavior. +- We explore an efficient pipeline to model complex long-span consistency by focusing on neighbors to achieve short-span strict consistency and knowing whole views to obtain cues that correct cumulative errors. +- We design an hybrid architecture to implement the interaction by adopting both attention and SSMs as operators, effectively capturing essential dependencies in short sequences while significantly reducing costs for modeling long sequences. +- Extensive experiments demonstrate that our CoSER outperforms the state-of-the-art multiview synthesis approaches in both quantitative and qualitative results. + +# 2. Related Work + +# 2.1. Text-to-3D Creation + +Recently, Text-to-3D generative models [4, 8, 15, 18, 35, 38] have made substantial advancements in computer vision [1, 2, 20, 25, 30, 41, 48–51, 64–67, 70, 72], achieving high-quality generation of both 3D objects and complex 3D scenes. Early works focus on directly generating assets using 3D diffusion models based on differentiable representations [7, 36, 61]. However, these approaches often relies on massive 3D data and incurs substantial training costs. Thanks to powerful text-to-image models [44], DreamFusion [38] leverages the robust 2D prior by introducing Score Distillation Sampling, which minimizes the discrepancy between rendered images from the underlying 3D assets and the diffusion prior. This development also sparked a surge of interest in other generative tasks such as texture generation [56, 71], editing [5, 37, 52, 62, 73], and Image-to-3D [39, 53, 59]. Despite this paradigm has bridged the + +gap between imagination and reality, it still faces significant challenges in terms of optimization speed [16, 68], content drift [24, 63] and the Janus problem [31, 46, 69]. In response to inefficient training, Metzer et al. [34] train a NeRF within the latent space of LDM [44] to accelerate the learning process and further introduce the Sketch-Shape method to guide the generation process. To acquire photorealistic appearance, Lin et al. [26] employ a coarse-to-fine strategy, first optimizing a NeRF using a low-resolution diffusion prior, then refining the texture with latent diffusion priors using the DMTet initialized with the coarse NeRF. Chen et al. [3] disentangle the learning of geometry and material, leveraging physics-based rendering techniques to achieve high-fidelity mesh generation. Another line of work lies in generating 3D assets directly through a 3D diffusion model based on differentiable representations. [28] fine-tune image diffusion models to generate multiview image-normal pairs with switch attention and employ NeuS [60] for reconstruction underlying geometries. In this work, we will continue to improve the quality of 3D content by designing the effective multiview text-to-image generator, which provides consistent images for reconstruction method. + +# 2.2. Multiview Image Generation + +Due to the scarcity of 3D data and the formidable training costs, lifting 2D diffusion to multiview generator conditioned on camera pose has become the prevailing paradigm [6, 17, 21, 23, 28, 32, 47, 55]. Tang et al. [55] introduce the novel concept of generating multiview images concurrently with correspondence-aware attention mechanisms, thereby enabling effective cross-view information interaction and application to texturing scene meshes. Meanwhile, Shi et al. [47] and Long et al. [28] simply extends 2D spatial attention to capture 3D structure, yielding remarkable results. Nevertheless, the quadratic complexity along with resolution restricts the number of cameras, thereby compromising the performance of reconstruction algorithms and resulting in inferior results. To reduce computational costs, Li et al. [23] propose an efficient rowwise attention instead of dense interactions, yet struggles to correct local details referencing global information. Latter works [9, 58, 74] capitalized on priors derived from video diffusion model to achieve both dense view generation and consistent outputs. Despite these strides, a fundamental disparity persists between the video continuity and 3D consistency. General videos typically depict objects moving along a fixed plane from the same camera viewpoint, i.e., similar information shifts across different positions on the canvas. In contrast, videos composed of images captured from multiple views often exhibit significant rotation, leading to substantial observable content jitter between frames. Here, we tailor an innovative strategy for cross-view interaction that + +optimizes both performance and efficiency, ensuring robust consistency in generating dense-view images for 3D applications. + +# 3. Method + +# 3.1. Preliminaries: Latent Diffusion Models + +Our work lifts the Latent Diffusion Model (LDM) [44], which maps a noise latent $\epsilon$ and a text prompt $p$ to an image $x$ . LDM first encodes the image $x_0$ into latent space using a pre-trained encoder $\mathcal{E}$ , i.e., $z_0 = \mathcal{E}(x_0)$ . The latent representation $z_0$ are then decoded back into the image $x_0$ using a pre-trained decoder $\mathcal{D}$ . The model is trained in the latent space in a self-supervised manner, and generates samples by reversing a Markov forward process. Specifically, the forward process injects noise into $z_0$ to obtain noisy inputs $\{z_t\}_{t=1}^T$ , where $z_t = \sqrt{\alpha^t z_0 + \sqrt{1 - \alpha_t}} \epsilon$ , and $\epsilon \sim N(0, I)$ . The generation process is formulated as a sequence of denoising operations $\epsilon_\theta(z_t, t, y)$ , predicting a denoised variant of input $z_t$ , where $y = \phi(p)$ is the condition, and $\phi$ is a frozen text encoder [42]. The learnable parameter $\theta$ is typically embedded within structures such as convolutional layers, self-attention mechanisms, and cross-attention mechanisms. The training objective is defined as follows: + +$$ +\min _ {\theta} \mathbb {E} _ {z _ {0}, \epsilon , t} | | \epsilon - \epsilon_ {\theta} (z _ {t}, t, y) | | _ {2} ^ {2}. \tag {1} +$$ + +The reverse process ensures that the final sample $z_0$ follows the distribution of latents, while aligning the generated image with the given condition. + +# 3.2. Framework of CoSER + +We propose CoSER, a method for generating dense and consistent multiview images based on text prompts. The inputs to CoSER include a set of noisy images $\{x_{t}^{i}\}_{i = 1}^{f}$ , a set of camera parameters $\{c^i\}_{i = 1}^f$ with a fixed elevation angle, uniformly distributed azimuth angles ranging from $0^{\circ}$ to $360^{\circ}$ , and a text prompt $y$ . We fine-tune the lifted Latent Diffusion Model (LDM) on $f$ views using a 3D dataset to synthesize multiview images $\{x_0^i\}_{i = 1}^f$ . It is non-trivial to incorporate the LDM architecture originally designed for 2D with 3D awareness. Rather than applying space-view attention or separately fusing spatial and viewpoint dimensions, our framework employs fine-grained interactions for neighboring views and coarse interactions for the global appearance. We retain the LDM structure, including ResBlock and CrossAttention, to maximize parameter reuse and maintain robust generative capabilities. A detailed sketch of CoSER is illustrated in Figure 2 and see more details for each module in Figure 3. + +# 3.2.1. Focus on Neighbors with Contextual Attention + +It is well known that 3D objects exhibit strict cross-view consistency, whereas this does not necessitate considering + +![](images/cf79c523f5583c58a4947ac6c9a5073d82a9993bd5518a01c0d14dcfdf853384.jpg) +Figure 2. Illustration of our CoSER. Given images rendered from 12 views at the same elevation, we take a pre-trained text-to-image generation model and fine-tune it by incorporating camera poses and lifting 2D-UNet to generate multi-view images. Specifically, we achieve 3D perception by employing fine-grained learning for neighbors viewpoints and coarse-grained interactions across whole viewpoints. For neighbors, Appearance Awareness (AA) is used to learn the basic appearance and Detail Refinement (DR) is proposed to ensure neighbor consistency in point-level. For whole views, we quickly scan all viewpoints with Rapid Glance (RG) and then eliminate ambiguity with powerful Accumulated Inconsistency Rectification (AIR). + +the relationships among all views. For instance, the front and back views of an object are usually highly independent. Since adjacent viewpoints share more information and contribute significantly to visual consistency, we propose interacting each view with its context to mitigate the computational expense of space-view attention. Specifically, we use adjacent attention to capture the fundamental appearance and trajectory attention to refine details point-to-point, ensuring consistency between neighboring viewpoints. Contextual attention, in this case, refers to the mechanism of incorporating information from neighboring views to enhance the understanding of the current view. + +Appearance Awareness CoSER begins by integrating content from adjacent viewpoints to outline the basic appearance from the current view, thereby ensuring fundamental neighbor-view consistency. Following the MT2I baseline [47], we encode images $\{x^i\}$ to latent feature $\{z^i\}$ by the pre-trained encoder $\mathcal{E}$ ( $t$ is omitted for clarity), followed by mapping camera parameters $\{c^i\}$ as high-dimensional condition. Moreover, the original self-attention is replaced with adjacent-attention, which takes the previous view $z^{i-1}$ , current view $z^i$ and next view $z^{i+1}$ as inputs, updating the features of $z^i$ . The formulation of the adjacent-attention is as follows: + +$$ +Q ^ {i} = W _ {Q} z ^ {i}, \tag {2} +$$ + +$$ +K ^ {i} = W _ {K} \left[ z ^ {i - 1}, z ^ {i}, z ^ {i + 1} \right], V ^ {i} = W _ {V} \left[ z ^ {i - 1}, z ^ {i}, z ^ {i + 1} \right], +$$ + +where $[\cdot]$ denotes concatenation operation. To inherit the generalizability of the 2D diffusion models, the parameters of self-attention is used to initialize projection matrices $W_{Q}, W_{K}, W_{V}$ which are shared across space and view. Note that adjacent-frame attention, while capable of forming the initial appearance of the object, inevitably disperses attention due to the involvement of numerous interacting patches, making it struggle to focus on details. + +Detail Refinement To ensure cross-view consistency in rendered images, it is crucial that the same position of an object corresponds to different pixels in neighboring viewpoints. These co-located pixels across views exhibit stronger correlations than others, playing a key role in maintaining visual consistency in multiview synthesis. For an object centered at the origin, and with some slight abuse of symbols, rigid rotation about the $y$ -axis follows a physical principle: + +$$ +x ^ {\prime} = \left(x - \frac {W}{2}\right) \cos \Delta \alpha + \frac {W}{2} - d \sin \Delta \alpha , \tag {3} +$$ + +where $(x,y)$ and $(x',y')$ are the pixel coordinates before and after rotation by an angle $\Delta \alpha$ at a depth $d$ , and $W$ is the image width. Since the depth $d$ is unknown in the generative task and $\sin \Delta \alpha$ becomes negligible when $\Delta \alpha$ is small, we simplify the transformation to: + +$$ +x ^ {\prime} = \left(x - \frac {W}{2}\right) \cos \Delta \alpha + \frac {W}{2}. \tag {4} +$$ + +To make this simplification robust in practical applications, we consider a $3 \times 3$ window $\{x' + i, y' + j\}_{i,j = -1,0,1}$ in neighboring frames as potential results. The feature at pixel $(x,y)$ is then refined through the attention mechanism that interacts with these neighboring pixels, enhancing consistency across views. + +# 3.2.2. Know the Whole with Spiral Mamba + +To efficiently enhance global consistency, we first quickly scan all viewpoints, then down-sample patches based on textual relevance to filter out irrelevant information. This allows us to utilize powerful attention mechanisms to correct any accumulated inconsistencies. + +Rapid Glance Exchanging information across all viewpoints is crucial for achieving global view consistency. However, dense attention mechanisms [28, 47] can be computationally expensive and memory-intensive, while simple + +![](images/9d96ae5ad7d78b09a813b424feea8010546057229a895255eaa82fdb19b2d352.jpg) +Appearance Awareness +View i-1 + +![](images/a2b205d773e630c2d53d90ec172321ba039a5c46a27cb0bde5a138f7af44d043.jpg) +View i + +![](images/061f00139e0655ec33ca40367068278e9e7501b50c2d6eb9f221ee62b2e6ba0e.jpg) +View i+1 + +![](images/59c491c3de51fc3fb83723147d06035d4b8c4fdab6029148a30c15839bdb4e88.jpg) +Detail Refinement +View i-1 + +![](images/a68fa2d9f9d5ab7c34b28ead3de6296b21df59e0c88d013b15a96ccd2406d191.jpg) +View i + +![](images/7231786a09ce8cd24ff97982134969ef61dcd73da76b10e6f23b63ac4f5a30fc.jpg) +View $i + 1$ + +![](images/4043873aa485d1af714d00c970500fc1b961acb8881dac2af78ae5a77ee5ed2e.jpg) +View 1 + +![](images/7e8afe8b0fdc20d36dc84ef32c47d75001d952aff5f41ebfeec549964911df24.jpg) +Rapid Glance +View i + +![](images/4c171c536dea9b6845f135b1b56825924026d1fd254b4700ec6c49c9fb38233c.jpg) +View 12 + +![](images/3b79589b275602f7cb200b282006199e24b54951ac0c64644658e4ca1501c048.jpg) + +![](images/4ffb6b32815ce9c6540fd786f80e825227046e5706aca771804401ce64cf98a7.jpg) +Accumulated +View 1 +Figure 3. Visualization of the interaction schemes in CoSER. For illustration, we denote each gird as a token with target token in red and attending token in green. + +![](images/51a47692d1371df9144cfd06043ced87e24c1e308f211f22c40ac3c360fb2391.jpg) +Inconsistency +View i + +![](images/fea03fc6ce8f2e4446e382adb703622e9bbd969018338dd31cfaf2567c11cd62.jpg) +View 12 + +temporal attention [74] fails to effectively learn multi-view consistency. Inspired by Gu et al. [12], we leverage state space models (SSMs) known as linear operators to process all patches of multiview images, offering a more efficient alternative to quadratic-time attention. Specifically, we utilize the advanced selective scan SSMs introduced by Mamba [11] to implement our core SSMs operator. Unlike existing Vision Mamba methods that flatten 2D images into 1D sequences row-wise or column-wise, we propose an effective spiral bidirectional scan strategy. This strategy leverages the prior knowledge that objects in 3D datasets are typically centered. By organizing object-related content within a short sequence window, our approach implicitly mitigates the forgetting problem of SSMs. Additionally, placing background content across distinct viewpoints aids the SSMs in boundary perception. + +Technically, our method organizes spatial tokens start + +ing from the center and expanding in a spiral pattern, then stacks them view by view. Meanwhile, we reverse the scan order within the arrangement of viewpoints to capture robust space-view features. This allows us to perform a rapid glance as follows: + +$$ +\left\{z ^ {i} \right\} _ {i = 1} ^ {f} = S S M s \left(S B S c a n \left(\left\{z ^ {i} \right\} _ {i = 1} ^ {f}\right)\right), \tag {5} +$$ + +where $SBScan(\cdot)$ denotes the spiral bidirectional scan. This operator effectively aggregates global information and provides crucial clues for further disambiguation. + +Accumulated Inconsistency Rectification To facilitate the elimination of inconsistencies through attention mechanisms, we implement down-sampling to filter out background and retain features relevant to textual information, thereby substantially reducing the number of patches engaged in interactions. Specifically, we provide a high-level semantic embedding $\phi(p)$ of the prompt $p$ . A 2-layer MLP $S$ is employed to assess the relevance between each patch and the text, producing a score ranging from 0 to 1 at each position, where a higher score indicates stronger relevance. Notably, we do not explicitly impose supervised constraints on the scoring mechanism. Instead, the model is driven by data to autonomously perceive the regions of interest. Based on the score map $s^i = S([z^i, \phi(p)])$ as a weight, we perform 2D-pooling along the spatial dimensions on the query, key, and value maps to enable sparse interactions: + +$$ +\hat {Q} ^ {i} = P \left(s ^ {i} \cdot \bar {Q} ^ {i}; \tau\right), +$$ + +$$ +\hat {K} ^ {i} = P \left(\left\{s ^ {j} \cdot \bar {K} ^ {i} \right\} _ {j = 1} ^ {f}; \rho\right), \tag {6} +$$ + +$$ +\hat {V} ^ {i} = P (\{s ^ {j} \cdot \bar {V} ^ {i} \} _ {j = 1} ^ {f}; \rho). +$$ + +Here, $P$ denotes the pooling operator, $\tau$ and $\rho$ are the strides required for pooling, and the transformations include: $\bar{Q}^i = W_{\bar{Q}}z^i$ , $\bar{K}^i = W_{\bar{K}}[\{z^j\}_{j=1}^f]$ , and $\bar{V}^j = W_{\bar{V}}[\{z^j\}_{j=1}^f]$ . Notably, in contrast to the intensive down-sampling of $\hat{V}^i$ and $\hat{K}^i$ , we maintain a higher resolution for the query feature map $\hat{Q}^i$ to minimize the loss of content information. Finally, we up-sample the computed feature map using bilinear interpolation to restore its resolution. + +# 4. Experiments + +In this section, we first provide the implementation details and the dataset. We then conduct both qualitative and quantitative experiments to compare our method with previous state-of-the-art text-to-3D generation methods. Additionally, we perform an ablation study to demonstrate the effectiveness of the proposed modules. + +Implementation We adopt the LDM $^1$ [44], pre-trained on the LAION-5B [45], as our base model. Specifically, we use the AdamW optimizer [29] with a learning rate of 3e-5 and a + +![](images/c925040cdf9122e8ea46408d7de0ca87c5b8f642525e739f37365747278a2596.jpg) +A bear dressed in medieval armor + +![](images/d82548c0429c60dbbb23905b38ecb44e5be752d9b352cdd385755f41fd1d0b99.jpg) +Figure 4. Qualitative comparison of VideoMV [74] (Up) and our CoSER (Down). +Figure 5. Qualitative comparisons of GaussianDreamer [69], Hash3D [68] and our CoSER. + +batch size of 6 to optimize the weights over approximately 200K iterations on three NVIDIA Tesla A800 GPUs. To preserve diversity and detail, we degrade CoSER to a text-to-image generation model with a $40\%$ probability and train it with a subset of LAION-5B. During inference, we employ the DDIM sampler with 50 steps and set the scale guidance to 7.5 for generation. Additionally, we set text template as "A DSLR photo of [prompt], 3d asset" to acquire more realistic results. Following the Liu et al. [27], we use NeuS[60] for performing reconstruction. + +Dataset We use the public G-Objaverse [40] as the training dataset, utilizing the names of the objects appended with "3D asset" as their text labels. For each object, we first normalize the object to be centered within a bounding box ranging from $[-0.5, 0.5]$ . The camera distance is set to 2.0 units. A random elevation is chosen between $[-30^\circ, 30^\circ]$ , and starting from the front view, we uniformly sample 12 angles around the object for rendering. + +# 4.1. Comparison with SOTA methods + +Qualitative Comparison We primarily evaluate the performance of the proposed CoSER on multiview text-to-image and compare the results with state-of-the-art methods in Figure 6. For VideoMV [74], 12 images generated from the same camera viewpoints as CoSER are selected for visualization, and all 24 viewpoints were utilized in the reconstruction process to ensure a fair comparison. Qualitatively, our method demonstrates impressive quality improvement in terms of both consistence and texture. Although VideoMV can generate images from a broader range of perspectives, e.g., mask and watch, it significantly compromises cross-view consistency, which poses serious challenges for subsequent 3D reconstruction. Furthermore, Figure 5 presents a comparison between two state-of-the-art text-to-3D methods, including GaussianDreamer [69] and Hash3D [68]. Specifically, the advanced method, Dream-Gaussian [54], is utilized as the foundational model for Hash3D. Thanks to the dense consistent images, CoSER outperforms other methods in terms of fidelity in 3D geometry and visual appearance. + +Quantitative Comparison Following previous work [74], we quantitatively assess the quality of the generated 3D assets using CLIP Score [14]. To evaluate the textual alignment, we followed [10] to compute the recall of the original prompt within the generated caption. We employ 100 text prompts selected from T3Bench [13] to generate 3D models. For each object, we set the elevation as $15^{\circ}$ and select 120 evenly spaced azimuth angles from $-180^{\circ}$ to $180^{\circ}$ , thereby generating 120 rendered images from different viewpoints. During the evaluation, 30 from the 120 rendered images are randomly selected and the CLIP Score is computed as the average semantic similarity between the prompts and these images. We use the CLIP ViT-B [43] to + +extract both text and image features. The results, presented in Table 1, demonstrate that our method significantly surpasses other methods, indicating that CoSER is more adept at injecting semantic information into the creation of realistic 3D assets. Given the challenges in automating the measurement of 3D consistency, we resort to user studies to assess which method produces results that are most favored by human evaluators. We conducted three user studies on the results in terms of the Semantic Alignment, Multiview Consistency and Texture Detail. In the first study, participants were asked to select the 3D asset that best aligns with the semantics of the given prompt. In the second study, participants identified the asset with the highest consistency. In the third study, participants were encouraged to choose the object with the finest texture and less distortion artifacts. We receive 100 answers on 30 text-3D pairs in total for each study. As shown in Table 1, over $29\%$ of our results are selected as the best in both three metrics, which proves a significant advantage in 3D generation. + +# 4.2. Ablation Study + +Effect of Module Design We are mainly interested in the 3D-aware module design, given that other unit has been well explored in previous methods. The core of our design is focusing on neighbors in fine-grained manner and knowing whole views with coarse-grained mechanism. To substantiate these claims, we incrementally integrated the proposed modules into four distinct configurations and finetuned the LDM accordingly. As depicted in Figure 6, the use of Appearance Awareness alone proves insufficient for ensuring cross-view consistency. However, the incorporation of Detail Refinement enhances the alignment between adjacent views, thereby improving overall 3D content consistency. It is evident that Rapid Glance offers small reduction in ambiguities; nonetheless, its primary benefit lies in facilitating information exchange across all viewpoints (See Figure 8), thereby establishing a foundation for the effective Accumulated Inconsistency Rectification module. Additionally, we present the parameters and MACs introduced by each module in Table 2. To illustrate that SSM captures multi-view consistency more effectively than temporal attention, we also learn global consistency by temporal attention found that it impaired performance. Based on this observation, we consider the additional 3.31M parameters and 15G MACs computations introduced by RG and AIR to be entirely acceptable. + +Spiral Bidirectional Scan in Rapid Glance To verify the effectiveness of the proposed spiral bidirectional scan strategy, we perform the evaluation by replacing it with vanilla spatial-first bidirectional scan in VideoMamba [22]. As shown in Figure 7, both scanning strategies enhance the consistency. However, spatial-first bidirectional scan inevitably disrupts the continuity of object, which hampers + +Table 1. Quantitative comparison of CoSER and state-of-the-arts based on the Quality, Alignment and User Study. + +
ModelQuality↑Alignment↑User Study
Semantic↑Consistency↑Texture↑
GaussianDreamer [69]30.7833.70.260.270.21
Hash3D [68]31.2630.20.200.190.25
VideoMV [74]32.1934.40.230.250.22
Ours33.0737.70.310.290.32
+ +Table 2. Model efficiency of each module. TA denotes temporal attention. + +
ModuleParameters (M)MACs (G)
AA867.451036
AA + DR917.041395
AA + DR + RG917.071396
AA + DR + RG + AIR969.951487
AA + DR + TA966.641472
+ +![](images/f205016f0384fc1af26b56a0c44100d5c1dc7bc3cb1268705c0f8474b2bc46cc.jpg) +Figure 6. Ablation of proposed moulds. The arrows indicate the changes made to the Appearance Awareness by sequentially plusing Detail Refinement, Rapid Glance, and Accumulated Inconsistency Rectification. Better see in color and $2\mathrm{x}$ zoom. + +![](images/768572bb6b40cbcfc096d97848991439a0c40871803f71f7adba73e16dbd9697.jpg) + +![](images/5106cb7081072979b910baca1b63665ead15d1be0a0c841edaa062028b7d9966.jpg) +Figure 7. Ablation of scanning strategy. Spiral bidirectional scan generates semantic- and view-consistent images. +Figure 8. Ablation of score map. Score map aids in capturing critical information, while Rapid Glance contributes to the generation of more accurate score maps. + +the efficiency of SSM model for feature aggregation. In contrast, our proposed spiral bidirectional scan maximizes the spatial compactness of foreground content while simultaneously enforcing operator awareness of inter-view boundaries, thereby achieving better results. Additionally, we experimented with reversing the spiral scanning order within the image, progressing from the periphery inward towards the center. However, this strategy did not yield further performance improvements, which we attribute to the already efficient information aggregation within an image achieved by the spiral scan. + +Score Map in Accumulated Inconsistency Rectification In Accumulated Inconsistency Rectification, we introduce the + +score map to assist in downsampling feature resolution for coarse attention, tailored to each individual view. To understand its impact, we compared the full AIR module generated results with those from two alternative models: AIR without score map and AIR with score map but without Rapid Glance. For clearness, we extract score maps at various resolutions and upsample them to 32x32 for visualization. As shown in Figure 8, the score map assigns higher scores to foreground regions, while background areas receive lower scores, facilitating the retention of critical information during the downsampling process. Additionally, Rapid Glance enhances the exchange of information, contributing to the generation of more precise score maps. + +# 5. Conclusion + +In this paper, we introduced CoSER, a novel framework for generating consistent and high-fidelity dense multiview images from text prompts. Our approach leverages a hybrid architecture that encapsulates the strengths of attention mechanisms and state space models (SSMs) to address the crucial challenges of cross-view consistency. The innovative design of contextual interaction includes an adjacent attention for capturing rough structures and textures, a trajectory-based attention for refining local details. A spiral bidirectional Mamba ensures global coherence in a cost-efficient manner. This comprehensive approach allows CoSER to produce dense, high-fidelity, and content-consistent multiview images. Extensive experiments demonstrate that CoSER outperforms state-of-the-art multiview synthesis methods in both quantitative and qualitative evaluations across various datasets. Future work could further explore the hybrid way of neighbor and global interaction to efficiently generate dense and consistent multiview images. + +# Acknowledgements + +This paper is supported by the Beijing Natural Science Foundation (1254050), National Natural Science Foundation of China under Grants (U23B2012, 12471308), Fundamental Research Funds for the Central Universities, and by the National Research Foundation, Singapore, under its Medium Sized Center for Advanced Robotics Technology Innovation. + +# References + +[1] Haoyu Chen, Jingjing Ren, Jinjin Gu, Hongtao Wu, Xuequan Lu, Haoming Cai, and Lei Zhu. Snow removal in video: A new dataset and a novel method. in 2023 IEEE. In ICCV, pages 13165-13176. 2 +[2] Haoyu Chen, Jinjin Gu, Yihao Liu, Salma Abdel Magid, Chao Dong, Qiong Wang, Hanspeter Pfister, and Lei Zhu. Masked image training for generalizable deep image denoising. In CVPR, pages 1692-1703, 2023. 2 +[3] Rui Chen, Yongwei Chen, Ningxin Jiao, and Kui Jia. Fantasia3d: Disentangling geometry and appearance for high-quality text-to-3d content creation. In ICCV, pages 22246–22256, 2023. 1, 3 +[4] Yongwei Chen, Rui Chen, Jiabao Lei, Yabin Zhang, and Kui Jia. Tango: Text-driven photorealistic and robust 3d stylization via lighting decomposition. In NeurIPS, pages 30923-30936, 2022. 2 +[5] Yiwen Chen, Zilong Chen, Chi Zhang, Feng Wang, Xiaofeng Yang, Yikai Wang, Zhongang Cai, Lei Yang, Huaping Liu, and Guosheng Lin. Gaussianeditor: Swift and controllable 3d editing with gaussian splatting. In CVPR, pages 21476-21485, 2024. 2 +[6] Zilong Chen, Yikai Wang, Feng Wang, Zhengyi Wang, and Huaping Liu. V3d: Video diffusion models are effective 3d generators. arXiv preprint arXiv:2403.06738, 2024. 2, 3 +[7] Yen-Chi Cheng, Hsin-Ying Lee, Sergey Tulyakov, Alexander G Schwing, and Liang-Yan Gui. Sdfusion: Multimodal 3d shape completion, reconstruction, and generation. In CVPR, pages 4456-4465, 2023. 2 +[8] Lihe Ding, Shaocong Dong, Zhanpeng Huang, Zibin Wang, Yiyuan Zhang, Kaixiong Gong, Dan Xu, and Tianfan Xue. Text-to-3d generation with bidirectional diffusion using both 2d and 3d priors. In CVPR, pages 5115–5124, 2024. 2 +[9] Ruiqi Gao, Aleksander Holynski, Philipp Henzler, Arthur Brussee, Ricardo Martin-Brualla, Pratul Srinivasan, Jonathan T Barron, and Ben Poole. Cat3d: Create anything in 3d with multi-view diffusion models. arXiv preprint arXiv:2405.10314, 2024. 2, 3 +[10] Chongjian Ge, Chenfeng Xu, Yuanfeng Ji, Chensheng Peng, Masayoshi Tomizuka, Ping Luo, Mingyu Ding, Varun Jampani, and Wei Zhan. Compgs: Unleashing 2d compositionality for compositional text-to-3d via dynamically optimizing 3d gaussians. arXiv preprint arXiv:2410.20723, 2024. 7 +[11] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023. 5 + +[12] Albert Gu, Karan Goel, and Christopher Re. Efficiently modeling long sequences with structured state spaces. In ICLR, 2022. 2, 5 +[13] Yuze He, Yushi Bai, Matthieu Lin, Wang Zhao, Yubin Hu, Jenny Sheng, Ran Yi, Juanzi Li, and Yong-Jin Liu. T3 bench: Benchmarking current progress in text-to-3d generation. arXiv preprint arXiv:2310.02977, 2023. 1, 7 +[14] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning. In EMNLP, pages 7514-7528, 2021. 7 +[15] Fangzhou Hong, Jiaxiang Tang, Ziang Cao, Min Shi, Tong Wu, Zhaoxi Chen, Tengfei Wang, Liang Pan, Dahua Lin, and Ziwei Liu. 3dtopia: Large text-to-3d generation model with hybrid diffusion priors. arXiv preprint arXiv:2403.02234, 2024. 2 +[16] Yukun Huang, Jianan Wang, Yukai Shi, Xianbiao Qi, Zheng-Jun Zha, and Lei Zhang. Dreamtime: An improved optimization strategy for text-to-3d content creation. arXiv preprint arXiv:2306.12422, 2023. 3 +[17] Zehuan Huang, Hao Wen, Junting Dong, Yaohui Wang, Yangguang Li, Xinyuan Chen, Yan-Pei Cao, Ding Liang, Yu Qiao, Bo Dai, et al. Epidiff: Enhancing multi-view synthesis via localized epipolar-constrained diffusion. In CVPR, pages 9784-9794, 2024. 2, 3 +[18] Ajay Jain, Ben Mildenhall, Jonathan T Barron, Pieter Abbeel, and Ben Poole. Zero-shot text-guided object generation with dream fields. In CVPR, pages 867-876, 2022. 2 +[19] Jeong-gi Kwak, Erqun Dong, Yuhe Jin, Hanseok Ko, Shweta Mahajan, and Kwang Moo Yi. Vivid-1-to-3: Novel view synthesis with video diffusion models. In CVPR, pages 6775–6785, 2024. 2 +[20] Bonan Li, Zicheng Zhang, Xuecheng Nie, Congying Han, Yinhan Hu, and Tiande Guo. Styo: Stylize your face in only one-shot. arXiv preprint arXiv:2303.03231, 2023. 2 +[21] Jiahao Li, Hao Tan, Kai Zhang, Zexiang Xu, Fujun Luan, Yinghao Xu, Yicong Hong, Kalyan Sunkavalli, Greg Shakhnarovich, and Sai Bi. Instant3d: Fast text-to-3d with sparse-view generation and large reconstruction model. In ICLR, 2023. 2, 3 +[22] Kunchang Li, Xinhao Li, Yi Wang, Yinan He, Yali Wang, Limin Wang, and Yu Qiao. Videomamba: State space model for efficient video understanding. arXiv preprint arXiv:2403.06977, 2024. 7 +[23] Peng Li, Yuan Liu, Xiaoxiao Long, Feihu Zhang, Cheng Lin, Mengfei Li, Xingqun Qi, Shanghang Zhang, Wenhan Luo, Ping Tan, et al. Era3d: High-resolution multiview diffusion using efficient row-wise attention. arXiv preprint arXiv:2405.11616, 2024. 2, 3 +[24] Weiyu Li, Rui Chen, Xuelin Chen, and Ping Tan. Sweetdreamer: Aligning geometric priors in 2d diffusion for consistent text-to-3d. In ICLR, 2024. 3 +[25] Wenbing Li, Hang Zhou, Junqing Yu, Zikai Song, and Wei Yang. Coupled mamba: Enhanced multimodal fusion with coupled state space model. In NeurIPS, pages 59808-59832, 2025. 2 + +[26] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. In CVPR, pages 300–309, 2023. 1, 3 +[27] Yuan Liu, Cheng Lin, Zijiao Zeng, Xiaoxiao Long, Lingjie Liu, Taku Komura, and Wenping Wang. Syncdreamer: Generating multiview-consistent images from a single-view image. In ICLR, 2024. 7 +[28] Xiaoxiao Long, Yuan-Chen Guo, Cheng Lin, Yuan Liu, Zhiyang Dou, Lingjie Liu, Yuexin Ma, Song-Hai Zhang, Marc Habermann, Christian Theobalt, et al. Wonder3d: Single image to 3d using cross-domain diffusion. In CVPR, pages 9970-9980, 2024. 2, 3, 4 +[29] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In ICLR, 2019. 5 +[30] Run Luo, Zikai Song, Lintao Ma, Jinlin Wei, Wei Yang, and Min Yang. Diffusiontrack: Diffusion model for multi-object tracking. In AAAI, pages 3991-3999, 2024. 2 +[31] Baorui Ma, Haoge Deng, Junsheng Zhou, Yu-Shen Liu, Tiejun Huang, and Xinlong Wang. Geodream: Disentangling 2d and geometric priors for high-fidelity and consistent 3d generation. arXiv preprint arXiv:2311.17971, 2023. 3 +[32] Luke Melas-Kyriazi, Iro Laina, Christian Rupprecht, Natalia Neverova, Andrea Vedaldi, Oran Gafni, and Filippos Kokkinos. Im-3d: Iterative multiview diffusion and reconstruction for high-quality 3d generation. In ICML, 2024. 2, 3 +[33] Gal Metzer, Elad Richardson, Or Patashnik, Raja Giryes, and Daniel Cohen-Or. Latent-nerf for shape-guided generation of 3d shapes and textures. In CVPR, pages 12663–12673, 2023. 1 +[34] Gal Metzer, Elad Richardson, Or Patashnik, Raja Giryes, and Daniel Cohen-Or. Latent-nerf for shape-guided generation of 3d shapes and textures. In CVPR, pages 12663–12673, 2023. 3 +[35] Nasir Mohammad Khalid, Tianhao Xie, Eugene Belilovsky, and Tiberiu Popa. Clip-mesh: Generating textured meshes from text using pretrained image-text models. In SIGGRAPH Asia, pages 1-8, 2022. 2 +[36] Alex Nichol, Heewoo Jun, Prafulla Dhariwal, Pamela Mishkin, and Mark Chen. Point-e: A system for generating 3d point clouds from complex prompts. arXiv preprint arXiv:2212.08751, 2022. 2 +[37] JangHo Park, Gihyun Kwon, and Jong Chul Ye. Ed-nerf: Efficient text-guided editing of 3d scene with latent space nerf. In ICLR, 2024. 2 +[38] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. In ICLR, 2022. 1, 2 +[39] Guocheng Qian, Jinjie Mai, Abdullah Hamdi, Jian Ren, Aliaksandr Siarohin, Bing Li, Hsin-Ying Lee, Ivan Skorokhodov, Peter Wonka, Sergey Tulyakov, et al. Magic123: One image to high-quality 3d object generation using both 2d and 3d diffusion priors. arXiv preprint arXiv:2306.17843, 2023. 2 +[40] Lingteng Qiu, Guanying Chen, Xiaodong Gu, Qi Zuo, Mutian Xu, Yushuang Wu, Weihao Yuan, Zilong Dong, Liefeng + +Bo, and Xiaoguang Han. Richdreamer: A generalizable normal-depth diffusion model for detail richness in text-to-3d. In CVPR, pages 9914–9925, 2024. 7 +[41] Xinmin Qiu, Congying Han, ZiCheng Zhang, Bonan Li, Tiande Guo, and Xuecheng Nie. Diffbfr: Bootstrapping diffusion model for blind face restoration. In ACM MM, pages 7785-7795, 2023. 2 +[42] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763, 2021. 3 +[43] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763, 2021. 7 +[44] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10684-10695, 2022. 2, 3, 5 +[45] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, Patrick Schramowski, Srivatsa Kundurthy, Katherine Crowson, Ludwig Schmidt, Robert Kaczmarczyk, and Jenia Jitsev. Laion-5b: An open large-scale dataset for training next generation image-text models. ArXiv, abs/2210.08402, 2022. 5 +[46] Junyoung Seo, Wooseok Jang, Min-Seop Kwak, Hyeonsu Kim, Jaehoon Ko, Junho Kim, Jin-Hwa Kim, Jiyoung Lee, and Seungryong Kim. Let 2d diffusion model know 3d-consistency for robust text-to-3d generation. In ICLR, 2024. 3 +[47] Yichun Shi, Peng Wang, Jianglong Ye, Long Mai, Kejie Li, and Xiao Yang. Mvdream: Multi-view diffusion for 3d generation. In ICLR, 2023. 2, 3, 4 +[48] Zikai Song, Junqing Yu, Yi-Ping Phoebe Chen, and Wei Yang. Transformer tracking with cyclic shifting window attention. In CVPR, pages 8791-8800, 2022. 2 +[49]Zikai Song, Run Luo, Junqing Yu, Yi-Ping Phoebe Chen, and Wei Yang. Compact transformer tracker with correlative masked modeling. In AAAI, pages 2321-2329, 2023. +[50] Zikai Song, Ying Tang, Run Luo, Lintao Ma, Junqing Yu, Yi-Ping Phoebe Chen, and Wei Yang. Autogenic language embedding for coherent point tracking. In ACM MM, pages 2021-2030, 2024. +[51] Haoze Sun, Wenbo Li, Jianzhuang Liu, Haoyu Chen, Renjing Pei, Xueyi Zou, Youliang Yan, and Yujiu Yang. Coser: Bridging image and language for cognitive super-resolution. In CVPR, pages 25868-25878, 2024. 2 +[52] Jingxiang Sun, Bo Zhang, Ruizhi Shao, Lizhen Wang, Wen Liu, Zhenda Xie, and Yebin Liu. Dreamcraft3d: Hierarchical 3d generation with bootstrapped diffusion prior. In ICLR, 2024. 2 +[53] Junshu Tang, Tengfei Wang, Bo Zhang, Ting Zhang, Ran Yi, Lizhuang Ma, and Dong Chen. Make-it-3d: High-fidelity 3d + +creation from a single image with diffusion prior. In ICCV, pages 22819-22829, 2023. 2 +[54] Jiaxiang Tang, Jiawei Ren, Hang Zhou, Ziwei Liu, and Gang Zeng. Dreamgaussian: Generative gaussian splatting for efficient 3d content creation. In ICLR, 2024. 7 +[55] Shitao Tang, Fuyang Zhang, Jiacheng Chen, Peng Wang, and Yasutaka Furukawa. Mvdiffusion: Enabling holistic multiview image generation with correspondence-aware diffusion. arXiv, 2023. 3 +[56] Christina Tsalicoglou, Fabian Manhardt, Alessio Tonioni, Michael Niemeyer, and Federico Tombari. Textmesh: Generation of realistic 3d meshes from text prompts. In 3DV, pages 1554-1563, 2024. 2 +[57] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 2 +[58] Vikram Voleti, Chun-Han Yao, Mark Boss, Adam Letts, David Pankratz, Dmitry Tochilkin, Christian Laforte, Robin Rombach, and Varun Jampani. Sv3d: Novel multi-view synthesis and 3d generation from a single image using latent video diffusion. arXiv preprint arXiv:2403.12008, 2024. 2, 3 +[59] Peng Wang and Yichun Shi. Imagedream: Image-prompt multi-view diffusion for 3d generation. arXiv preprint arXiv:2312.02201, 2023. 2 +[60] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. In NeurIPS, 2021. 3, 7 +[61] Tengfei Wang, Bo Zhang, Ting Zhang, Shuyang Gu, Jianmin Bao, Tadas Baltrusaitis, Jingjing Shen, Dong Chen, Fang Wen, Qifeng Chen, et al. Rodin: A generative model for sculpting 3d digital avatars using diffusion. In CVPR, pages 4563-4573, 2023. 2 +[62] Xinzhou Wang, Yikai Wang, Junliang Ye, Zhengyi Wang, Fuchun Sun, Pengkun Liu, Ling Wang, Kai Sun, Xintong Wang, and Bin He. Animatabledreamer: Text-guided non-rigid 3d model generation and reconstruction with canonical score distillation. arXiv preprint arXiv:2312.03795, 2023. 2 +[63] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. In NeurIPS, 2024. 3 +[64] Zeyu Xiao and Zhiwei Xiong. Incorporating degradation estimation in light field spatial super-resolution. Computer Vision and Image Understanding, page 104295, 2025. 2 +[65] Zeyu Xiao, Yutong Liu, Ruisheng Gao, and Zhiwei Xiong. Cutmib: Boosting light field super-resolution via multi-view image blending. In CVPR, pages 1672-1682, 2023. +[66] Zeyu Xiao, Dachun Kai, Yueyi Zhang, Xiaoyan Sun, and Zhiwei Xiong. Asymmetric event-guided video superresolution. In ACM MM, pages 2409-2418, 2024. +[67] Xinchao Wang Xingyi Yang. Kolmogorov-arnold transformer. In ICLR, 2025. 2 +[68] Xingyi Yang and Xinchao Wang. Hash3d: Training-free acceleration for 3d generation. arXiv preprint arXiv:2404.06091, 2024. 3, 6, 7, 8 + +[69] Taoran Yi, Jiemin Fang, Junjie Wang, Guanjun Wu, Lingxi Xie, Xiaopeng Zhang, Wenyu Liu, Qi Tian, and Xinggang Wang. Gaussian dreamer: Fast generation from text to 3d gaussians by bridging 2d and 3d diffusion models. In CVPR, pages 6796-6807, 2024. 3, 6, 7, 8 +[70] Weihao Yu and Xinchao Wang. Mambaout: Do we really need mamba for vision? In CVPR, 2025. 2 +[71] Xianfang Zeng, Xin Chen, Zhongqi Qi, Wen Liu, Zibo Zhao, Zhibin Wang, Bin Fu, Yong Liu, and Gang Yu. Paint3d: Paint anything 3d with lighting-less texture diffusion models. In CVPR, pages 4252-4262, 2024. 2 +[72] Zicheng Zhang, Bonan Li, Xuecheng Nie, Congying Han, Tiande Guo, and Luoqi Liu. Towards consistent video editing with text-to-image diffusion models. In NeurIPS, pages 58508-58519, 2023. 2 +[73] Jingyu Zhuang, Chen Wang, Liang Lin, Lingjie Liu, and Guanbin Li. Dreameditor: Text-driven 3d scene editing with neural fields. In SIGGRAPH Asia, pages 1-10, 2023. 2 +[74] Qi Zuo, Xiaodong Gu, Lingteng Qiu, Yuan Dong, Zhengyi Zhao, Weihao Yuan, Rui Peng, Siyu Zhu, Zilong Dong, Liefeng Bo, et al. Videomv: Consistent multi-view generation based on large video generative model. arXiv preprint arXiv:2403.12010, 2024. 2, 3, 5, 6, 7, 8 \ No newline at end of file diff --git a/2025/CoSER_ Towards Consistent Dense Multiview Text-to-Image Generator for 3D Creation/images.zip b/2025/CoSER_ Towards Consistent Dense Multiview Text-to-Image Generator for 3D Creation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..2bf4a3fa2145088f36c97c044003ed35832fd698 --- /dev/null +++ b/2025/CoSER_ Towards Consistent Dense Multiview Text-to-Image Generator for 3D Creation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d15070828b8b98831693105ce0d4e4b7fb05dc9e800d35dea7f2b1142c009fa5 +size 682887 diff --git a/2025/CoSER_ Towards Consistent Dense Multiview Text-to-Image Generator for 3D Creation/layout.json b/2025/CoSER_ Towards Consistent Dense Multiview Text-to-Image Generator for 3D Creation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..39b2dc7943e2cfd6b7f798f1ddafb96a95390588 --- /dev/null +++ b/2025/CoSER_ Towards Consistent Dense Multiview Text-to-Image Generator for 3D Creation/layout.json @@ -0,0 +1,9277 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 119, + 103, + 493, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 119, + 103, + 493, + 138 + ], + "spans": [ + { + "bbox": [ + 119, + 103, + 493, + 138 + ], + "type": "text", + "content": "CoSER: Towards Consistent Dense Multiview Text-to-Image Generator for 3D Creation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 140, + 160, + 473, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 160, + 473, + 176 + ], + "spans": [ + { + "bbox": [ + 140, + 160, + 473, + 176 + ], + "type": "text", + "content": "Bonan Li" + }, + { + "bbox": [ + 140, + 160, + 473, + 176 + ], + "type": "inline_equation", + "content": "^{1,2*}" + }, + { + "bbox": [ + 140, + 160, + 473, + 176 + ], + "type": "text", + "content": " Zicheng Zhang" + }, + { + "bbox": [ + 140, + 160, + 473, + 176 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 140, + 160, + 473, + 176 + ], + "type": "text", + "content": " Xingyi Yang" + }, + { + "bbox": [ + 140, + 160, + 473, + 176 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 140, + 160, + 473, + 176 + ], + "type": "text", + "content": " Xinchao Wang" + }, + { + "bbox": [ + 140, + 160, + 473, + 176 + ], + "type": "inline_equation", + "content": "^{2\\dagger}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 139, + 176, + 472, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 176, + 472, + 190 + ], + "spans": [ + { + "bbox": [ + 139, + 176, + 472, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 139, + 176, + 472, + 190 + ], + "type": "text", + "content": "University of Chinese Academy of Sciences " + }, + { + "bbox": [ + 139, + 176, + 472, + 190 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 139, + 176, + 472, + 190 + ], + "type": "text", + "content": " National University of Singapore" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 97, + 190, + 510, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 190, + 510, + 204 + ], + "spans": [ + { + "bbox": [ + 97, + 190, + 510, + 204 + ], + "type": "text", + "content": "libonacci@ucas.ac.cn zhangzicheng19@mails.ucas.ac.cn xyang@u.nus.edu xinchao@nus.edu.sg" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 67, + 208, + 539, + 396 + ], + "blocks": [ + { + "bbox": [ + 67, + 208, + 539, + 396 + ], + "lines": [ + { + "bbox": [ + 67, + 208, + 539, + 396 + ], + "spans": [ + { + "bbox": [ + 67, + 208, + 539, + 396 + ], + "type": "image", + "image_path": "708d74411269255e50a089ac20aad4efc83523fe4cbbf99bc2ae0b88c00a6d8f.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 408, + 555, + 431 + ], + "lines": [ + { + "bbox": [ + 54, + 408, + 555, + 431 + ], + "spans": [ + { + "bbox": [ + 54, + 408, + 555, + 431 + ], + "type": "text", + "content": "Figure 1. CoSER aims to generate detailed, diverse and detaild 3D objects from text prompts. All prompts are from T3Bench [13] and the templated is set as \"A DSLR photo of ..., 3d asset\". Better see in color and 2x zoom." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 151, + 457, + 201, + 470 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 457, + 201, + 470 + ], + "spans": [ + { + "bbox": [ + 151, + 457, + 201, + 470 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 54, + 483, + 297, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 483, + 297, + 687 + ], + "spans": [ + { + "bbox": [ + 54, + 483, + 297, + 687 + ], + "type": "text", + "content": "Generating dense multiview images from text prompts is crucial for creating high-fidelity 3D assets. Nevertheless, existing methods struggle with space-view correspondences, resulting in sparse and low-quality outputs. In this paper, we introduce CoSER, a novel consistent dense Multiview Text-to-Image Generator for Text-to-3D, achieving both efficiency and quality by meticulously learning neighbor-view coherence and further alleviating ambiguity through the swift traversal of all views. For achieving neighbor-view consistency, each viewpoint densely interacts with adjacent viewpoints to perceive the global spatial structure, and aggregates information along motion paths explicitly defined by physical principles to refine details. To further enhance cross-view consistency and alleviate content drift, CoSER rapidly scan all views in spiral bidirectional manner to aware holistic information and then scores each point based on semantic material. Subsequently, we" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 459, + 556, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 459, + 556, + 579 + ], + "spans": [ + { + "bbox": [ + 313, + 459, + 556, + 579 + ], + "type": "text", + "content": "conduct weighted down-sampling along the spatial dimension based on scores, thereby facilitating prominent information fusion across all views with lightweight computation. Technically, the core module is built by integrating the attention mechanism with a selective state space model, exploiting the robust learning capabilities of the former and the low overhead of the latter. Extensive evaluation shows that CoSER is capable of producing dense, high-fidelity, content-consistent multiview images that can be flexibly integrated into various 3D generation models." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 607, + 395, + 620 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 607, + 395, + 620 + ], + "spans": [ + { + "bbox": [ + 314, + 607, + 395, + 620 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "content": "Creating 3D content from text is an important task in computer vision and graphics for its potential applications in advertising, robotics, digital games and the meta-verse. Previous works [3, 26, 33] primarily optimise 3D representations using score distillation sampling [38] of Text-to-Image (T2I) models, ensuring that rendered images from any viewpoint maintain high fidelity while aligning with" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 693, + 133, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 693, + 133, + 703 + ], + "spans": [ + { + "bbox": [ + 66, + 693, + 133, + 703 + ], + "type": "text", + "content": "*Equal contribution" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 67, + 703, + 143, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 143, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 143, + 712 + ], + "type": "text", + "content": "† Corresponding author" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "2880" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": "the given prompt. Nevertheless, these methods suffer from Janus problems and severe artifacts with asymmetric geometry. As a remedy, multiview text-to-image (MT2I) models [6, 17, 19, 21, 23, 32] have been developed, providing rich 3D priors that help mitigate these problems. Furthermore, using dense views can accelerate the modeling process and enhance the quality of the resulting 3D models." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 158, + 295, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 158, + 295, + 350 + ], + "spans": [ + { + "bbox": [ + 55, + 158, + 295, + 350 + ], + "type": "text", + "content": "To acquire MT2I models, a promising approach is to leverage 3D priors learned from 3D datasets to enhance 2D diffusion models. Seminal works [28, 47] have focused on inflating T2I models with dense cross-view attention to capture 3D information. However, this approach incurs expensive computation and faces challenges when dealing with dense viewpoints. To address these limitations, sparse attention mechanisms [23] have been proposed to model cross-view dependencies, though they can introduce inconsistencies across multiple views. Recent studies [9, 58, 74] have leveraged cross-frame continuity priors from text-to-video (T2V) models to enhance consistency. Nevertheless, azimuthal rotation around an object can cause significant content variations at the same image coordinates, making temporal attention insufficient for capturing intricate details necessary for 3D generation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 352, + 295, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 352, + 295, + 602 + ], + "spans": [ + { + "bbox": [ + 55, + 352, + 295, + 602 + ], + "type": "text", + "content": "Despite impressive progress in multiview T2I generation, effectively and efficiently modeling the space-view structure within a unified architecture remains an open challenge. In this paper, we advance the field by addressing two fundamental problems: (1) How to design efficient cross-view interactions to ensure dense multiview consistency? Rather than directly optimizing consistency among all views with uniform granularity, we propose a two-step approach: (i) Learning neighbor-view consistency as a prerequisite, as enforcing consistency between adjacent viewpoints naturally enables coherence across all viewpoints. (ii) Ensuring global-view coherence to mitigate discrepancies between distant views and facilitate unified appearance, since subtle inconsistencies are inevitable and deteriorated as the viewing angles increases. (2) Which operator should be applied for the implementation? Unlike previous works that employ attention as the default choice, we leverage hybrid operators, including attention [57] and state space models (SSMs) [12], to harness the benefits of both model families, e.g., the powerful capabilities of attention and the low computational cost of SSMs." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 605, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 715 + ], + "type": "text", + "content": "In this paper, we address these questions and propose a novel model named CoSER, which enhances consistency in dense MT2I tasks. CoSER utilizes \"factor\" interactions that focus on neighboring views and maintain awareness of the global view. Technically, CoSER initially employs dense attention on adjacent views to capture the rough structure and texture of the current plane, thereby constructing a foundational appearance. Subsequently, we introduce a trajectory-based module to refine local details by accurately tracing" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 553, + 286 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 286 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 286 + ], + "type": "text", + "content": "3D trajectories and enhancing point-to-point consistency. Specifically, motivated by the efficiency of calculating rotated locations for centralized objects, we develop a robust method to handle complicate cases. This involves pixel interaction within a window of neighboring frames defined by physical rotation rules, ensuring precise alignment and consistency across views. To further alleviate disparities across arbitrary viewpoints, we utilize state space models (SSMs) equipped with a novel spiral bidirectional scan strategy to aggregate features from the entire set of views. We then score each patch based on textual inputs and perform down-sampling conditioned on these scores to adaptively ignore semantically irrelevant information. This approach allows CoSER to efficiently maintain global context through sparse attention. With these designs, our method establishes robust priors, enabling the creation of exceptionally coherent 3D assets (see Figure 1). In summary, our contributions can be summarized as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 314, + 289, + 553, + 468 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 314, + 289, + 553, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 289, + 553, + 323 + ], + "spans": [ + { + "bbox": [ + 314, + 289, + 553, + 323 + ], + "type": "text", + "content": "- We propose a novel multiview text-to-image framework, CoSER, for 3D content creation which exhibits significantly consistency behavior." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 324, + 553, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 324, + 553, + 371 + ], + "spans": [ + { + "bbox": [ + 314, + 324, + 553, + 371 + ], + "type": "text", + "content": "- We explore an efficient pipeline to model complex long-span consistency by focusing on neighbors to achieve short-span strict consistency and knowing whole views to obtain cues that correct cumulative errors." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 373, + 553, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 373, + 553, + 432 + ], + "spans": [ + { + "bbox": [ + 314, + 373, + 553, + 432 + ], + "type": "text", + "content": "- We design an hybrid architecture to implement the interaction by adopting both attention and SSMs as operators, effectively capturing essential dependencies in short sequences while significantly reducing costs for modeling long sequences." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 432, + 553, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 432, + 553, + 468 + ], + "spans": [ + { + "bbox": [ + 314, + 432, + 553, + 468 + ], + "type": "text", + "content": "- Extensive experiments demonstrate that our CoSER outperforms the state-of-the-art multiview synthesis approaches in both quantitative and qualitative results." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 482, + 400, + 495 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 482, + 400, + 495 + ], + "spans": [ + { + "bbox": [ + 314, + 482, + 400, + 495 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 503, + 430, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 503, + 430, + 514 + ], + "spans": [ + { + "bbox": [ + 314, + 503, + 430, + 514 + ], + "type": "text", + "content": "2.1. Text-to-3D Creation" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "type": "text", + "content": "Recently, Text-to-3D generative models [4, 8, 15, 18, 35, 38] have made substantial advancements in computer vision [1, 2, 20, 25, 30, 41, 48–51, 64–67, 70, 72], achieving high-quality generation of both 3D objects and complex 3D scenes. Early works focus on directly generating assets using 3D diffusion models based on differentiable representations [7, 36, 61]. However, these approaches often relies on massive 3D data and incurs substantial training costs. Thanks to powerful text-to-image models [44], DreamFusion [38] leverages the robust 2D prior by introducing Score Distillation Sampling, which minimizes the discrepancy between rendered images from the underlying 3D assets and the diffusion prior. This development also sparked a surge of interest in other generative tasks such as texture generation [56, 71], editing [5, 37, 52, 62, 73], and Image-to-3D [39, 53, 59]. Despite this paradigm has bridged the" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 313, + 757 + ], + "type": "text", + "content": "2881" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 347 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 347 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 347 + ], + "type": "text", + "content": "gap between imagination and reality, it still faces significant challenges in terms of optimization speed [16, 68], content drift [24, 63] and the Janus problem [31, 46, 69]. In response to inefficient training, Metzer et al. [34] train a NeRF within the latent space of LDM [44] to accelerate the learning process and further introduce the Sketch-Shape method to guide the generation process. To acquire photorealistic appearance, Lin et al. [26] employ a coarse-to-fine strategy, first optimizing a NeRF using a low-resolution diffusion prior, then refining the texture with latent diffusion priors using the DMTet initialized with the coarse NeRF. Chen et al. [3] disentangle the learning of geometry and material, leveraging physics-based rendering techniques to achieve high-fidelity mesh generation. Another line of work lies in generating 3D assets directly through a 3D diffusion model based on differentiable representations. [28] fine-tune image diffusion models to generate multiview image-normal pairs with switch attention and employ NeuS [60] for reconstruction underlying geometries. In this work, we will continue to improve the quality of 3D content by designing the effective multiview text-to-image generator, which provides consistent images for reconstruction method." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 360, + 214, + 372 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 360, + 214, + 372 + ], + "spans": [ + { + "bbox": [ + 55, + 360, + 214, + 372 + ], + "type": "text", + "content": "2.2. Multiview Image Generation" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 378, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 378, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 378, + 296, + 713 + ], + "type": "text", + "content": "Due to the scarcity of 3D data and the formidable training costs, lifting 2D diffusion to multiview generator conditioned on camera pose has become the prevailing paradigm [6, 17, 21, 23, 28, 32, 47, 55]. Tang et al. [55] introduce the novel concept of generating multiview images concurrently with correspondence-aware attention mechanisms, thereby enabling effective cross-view information interaction and application to texturing scene meshes. Meanwhile, Shi et al. [47] and Long et al. [28] simply extends 2D spatial attention to capture 3D structure, yielding remarkable results. Nevertheless, the quadratic complexity along with resolution restricts the number of cameras, thereby compromising the performance of reconstruction algorithms and resulting in inferior results. To reduce computational costs, Li et al. [23] propose an efficient rowwise attention instead of dense interactions, yet struggles to correct local details referencing global information. Latter works [9, 58, 74] capitalized on priors derived from video diffusion model to achieve both dense view generation and consistent outputs. Despite these strides, a fundamental disparity persists between the video continuity and 3D consistency. General videos typically depict objects moving along a fixed plane from the same camera viewpoint, i.e., similar information shifts across different positions on the canvas. In contrast, videos composed of images captured from multiple views often exhibit significant rotation, leading to substantial observable content jitter between frames. Here, we tailor an innovative strategy for cross-view interaction that" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 313, + 72, + 555, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 108 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 108 + ], + "type": "text", + "content": "optimizes both performance and efficiency, ensuring robust consistency in generating dense-view images for 3D applications." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 118, + 370, + 131 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 118, + 370, + 131 + ], + "spans": [ + { + "bbox": [ + 313, + 118, + 370, + 131 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 138, + 522, + 150 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 138, + 522, + 150 + ], + "spans": [ + { + "bbox": [ + 313, + 138, + 522, + 150 + ], + "type": "text", + "content": "3.1. Preliminaries: Latent Diffusion Models" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 156, + 555, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 156, + 555, + 369 + ], + "spans": [ + { + "bbox": [ + 313, + 156, + 555, + 369 + ], + "type": "text", + "content": "Our work lifts the Latent Diffusion Model (LDM) [44], which maps a noise latent " + }, + { + "bbox": [ + 313, + 156, + 555, + 369 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 313, + 156, + 555, + 369 + ], + "type": "text", + "content": " and a text prompt " + }, + { + "bbox": [ + 313, + 156, + 555, + 369 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 313, + 156, + 555, + 369 + ], + "type": "text", + "content": " to an image " + }, + { + "bbox": [ + 313, + 156, + 555, + 369 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 156, + 555, + 369 + ], + "type": "text", + "content": ". LDM first encodes the image " + }, + { + "bbox": [ + 313, + 156, + 555, + 369 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 313, + 156, + 555, + 369 + ], + "type": "text", + "content": " into latent space using a pre-trained encoder " + }, + { + "bbox": [ + 313, + 156, + 555, + 369 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 313, + 156, + 555, + 369 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 313, + 156, + 555, + 369 + ], + "type": "inline_equation", + "content": "z_0 = \\mathcal{E}(x_0)" + }, + { + "bbox": [ + 313, + 156, + 555, + 369 + ], + "type": "text", + "content": ". The latent representation " + }, + { + "bbox": [ + 313, + 156, + 555, + 369 + ], + "type": "inline_equation", + "content": "z_0" + }, + { + "bbox": [ + 313, + 156, + 555, + 369 + ], + "type": "text", + "content": " are then decoded back into the image " + }, + { + "bbox": [ + 313, + 156, + 555, + 369 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 313, + 156, + 555, + 369 + ], + "type": "text", + "content": " using a pre-trained decoder " + }, + { + "bbox": [ + 313, + 156, + 555, + 369 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 313, + 156, + 555, + 369 + ], + "type": "text", + "content": ". The model is trained in the latent space in a self-supervised manner, and generates samples by reversing a Markov forward process. Specifically, the forward process injects noise into " + }, + { + "bbox": [ + 313, + 156, + 555, + 369 + ], + "type": "inline_equation", + "content": "z_0" + }, + { + "bbox": [ + 313, + 156, + 555, + 369 + ], + "type": "text", + "content": " to obtain noisy inputs " + }, + { + "bbox": [ + 313, + 156, + 555, + 369 + ], + "type": "inline_equation", + "content": "\\{z_t\\}_{t=1}^T" + }, + { + "bbox": [ + 313, + 156, + 555, + 369 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 156, + 555, + 369 + ], + "type": "inline_equation", + "content": "z_t = \\sqrt{\\alpha^t z_0 + \\sqrt{1 - \\alpha_t}} \\epsilon" + }, + { + "bbox": [ + 313, + 156, + 555, + 369 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 156, + 555, + 369 + ], + "type": "inline_equation", + "content": "\\epsilon \\sim N(0, I)" + }, + { + "bbox": [ + 313, + 156, + 555, + 369 + ], + "type": "text", + "content": ". The generation process is formulated as a sequence of denoising operations " + }, + { + "bbox": [ + 313, + 156, + 555, + 369 + ], + "type": "inline_equation", + "content": "\\epsilon_\\theta(z_t, t, y)" + }, + { + "bbox": [ + 313, + 156, + 555, + 369 + ], + "type": "text", + "content": ", predicting a denoised variant of input " + }, + { + "bbox": [ + 313, + 156, + 555, + 369 + ], + "type": "inline_equation", + "content": "z_t" + }, + { + "bbox": [ + 313, + 156, + 555, + 369 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 156, + 555, + 369 + ], + "type": "inline_equation", + "content": "y = \\phi(p)" + }, + { + "bbox": [ + 313, + 156, + 555, + 369 + ], + "type": "text", + "content": " is the condition, and " + }, + { + "bbox": [ + 313, + 156, + 555, + 369 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 313, + 156, + 555, + 369 + ], + "type": "text", + "content": " is a frozen text encoder [42]. The learnable parameter " + }, + { + "bbox": [ + 313, + 156, + 555, + 369 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 156, + 555, + 369 + ], + "type": "text", + "content": " is typically embedded within structures such as convolutional layers, self-attention mechanisms, and cross-attention mechanisms. The training objective is defined as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 372, + 370, + 555, + 388 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 370, + 555, + 388 + ], + "spans": [ + { + "bbox": [ + 372, + 370, + 555, + 388 + ], + "type": "interline_equation", + "content": "\\min _ {\\theta} \\mathbb {E} _ {z _ {0}, \\epsilon , t} | | \\epsilon - \\epsilon_ {\\theta} (z _ {t}, t, y) | | _ {2} ^ {2}. \\tag {1}", + "image_path": "e485cec378efbe82ae88c5a81ab5db6c70a767825a07f51f7d3f6bffb1663e14.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 392, + 555, + 428 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 392, + 555, + 428 + ], + "spans": [ + { + "bbox": [ + 313, + 392, + 555, + 428 + ], + "type": "text", + "content": "The reverse process ensures that the final sample " + }, + { + "bbox": [ + 313, + 392, + 555, + 428 + ], + "type": "inline_equation", + "content": "z_0" + }, + { + "bbox": [ + 313, + 392, + 555, + 428 + ], + "type": "text", + "content": " follows the distribution of latents, while aligning the generated image with the given condition." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 435, + 441, + 447 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 435, + 441, + 447 + ], + "spans": [ + { + "bbox": [ + 313, + 435, + 441, + 447 + ], + "type": "text", + "content": "3.2. Framework of CoSER" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 453, + 556, + 669 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 453, + 556, + 669 + ], + "spans": [ + { + "bbox": [ + 313, + 453, + 556, + 669 + ], + "type": "text", + "content": "We propose CoSER, a method for generating dense and consistent multiview images based on text prompts. The inputs to CoSER include a set of noisy images " + }, + { + "bbox": [ + 313, + 453, + 556, + 669 + ], + "type": "inline_equation", + "content": "\\{x_{t}^{i}\\}_{i = 1}^{f}" + }, + { + "bbox": [ + 313, + 453, + 556, + 669 + ], + "type": "text", + "content": ", a set of camera parameters " + }, + { + "bbox": [ + 313, + 453, + 556, + 669 + ], + "type": "inline_equation", + "content": "\\{c^i\\}_{i = 1}^f" + }, + { + "bbox": [ + 313, + 453, + 556, + 669 + ], + "type": "text", + "content": " with a fixed elevation angle, uniformly distributed azimuth angles ranging from " + }, + { + "bbox": [ + 313, + 453, + 556, + 669 + ], + "type": "inline_equation", + "content": "0^{\\circ}" + }, + { + "bbox": [ + 313, + 453, + 556, + 669 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 453, + 556, + 669 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 313, + 453, + 556, + 669 + ], + "type": "text", + "content": ", and a text prompt " + }, + { + "bbox": [ + 313, + 453, + 556, + 669 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 313, + 453, + 556, + 669 + ], + "type": "text", + "content": ". We fine-tune the lifted Latent Diffusion Model (LDM) on " + }, + { + "bbox": [ + 313, + 453, + 556, + 669 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 313, + 453, + 556, + 669 + ], + "type": "text", + "content": " views using a 3D dataset to synthesize multiview images " + }, + { + "bbox": [ + 313, + 453, + 556, + 669 + ], + "type": "inline_equation", + "content": "\\{x_0^i\\}_{i = 1}^f" + }, + { + "bbox": [ + 313, + 453, + 556, + 669 + ], + "type": "text", + "content": ". It is non-trivial to incorporate the LDM architecture originally designed for 2D with 3D awareness. Rather than applying space-view attention or separately fusing spatial and viewpoint dimensions, our framework employs fine-grained interactions for neighboring views and coarse interactions for the global appearance. We retain the LDM structure, including ResBlock and CrossAttention, to maximize parameter reuse and maintain robust generative capabilities. A detailed sketch of CoSER is illustrated in Figure 2 and see more details for each module in Figure 3." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 674, + 541, + 686 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 674, + 541, + 686 + ], + "spans": [ + { + "bbox": [ + 313, + 674, + 541, + 686 + ], + "type": "text", + "content": "3.2.1. Focus on Neighbors with Contextual Attention" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "content": "It is well known that 3D objects exhibit strict cross-view consistency, whereas this does not necessitate considering" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2882" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 68, + 553, + 194 + ], + "blocks": [ + { + "bbox": [ + 56, + 68, + 553, + 194 + ], + "lines": [ + { + "bbox": [ + 56, + 68, + 553, + 194 + ], + "spans": [ + { + "bbox": [ + 56, + 68, + 553, + 194 + ], + "type": "image", + "image_path": "cf79c523f5583c58a4947ac6c9a5073d82a9993bd5518a01c0d14dcfdf853384.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 201, + 555, + 268 + ], + "lines": [ + { + "bbox": [ + 54, + 201, + 555, + 268 + ], + "spans": [ + { + "bbox": [ + 54, + 201, + 555, + 268 + ], + "type": "text", + "content": "Figure 2. Illustration of our CoSER. Given images rendered from 12 views at the same elevation, we take a pre-trained text-to-image generation model and fine-tune it by incorporating camera poses and lifting 2D-UNet to generate multi-view images. Specifically, we achieve 3D perception by employing fine-grained learning for neighbors viewpoints and coarse-grained interactions across whole viewpoints. For neighbors, Appearance Awareness (AA) is used to learn the basic appearance and Detail Refinement (DR) is proposed to ensure neighbor consistency in point-level. For whole views, we quickly scan all viewpoints with Rapid Glance (RG) and then eliminate ambiguity with powerful Accumulated Inconsistency Rectification (AIR)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 280, + 294, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 280, + 294, + 422 + ], + "spans": [ + { + "bbox": [ + 54, + 280, + 294, + 422 + ], + "type": "text", + "content": "the relationships among all views. For instance, the front and back views of an object are usually highly independent. Since adjacent viewpoints share more information and contribute significantly to visual consistency, we propose interacting each view with its context to mitigate the computational expense of space-view attention. Specifically, we use adjacent attention to capture the fundamental appearance and trajectory attention to refine details point-to-point, ensuring consistency between neighboring viewpoints. Contextual attention, in this case, refers to the mechanism of incorporating information from neighboring views to enhance the understanding of the current view." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 424, + 295, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 424, + 295, + 567 + ], + "spans": [ + { + "bbox": [ + 54, + 424, + 295, + 567 + ], + "type": "text", + "content": "Appearance Awareness CoSER begins by integrating content from adjacent viewpoints to outline the basic appearance from the current view, thereby ensuring fundamental neighbor-view consistency. Following the MT2I baseline [47], we encode images " + }, + { + "bbox": [ + 54, + 424, + 295, + 567 + ], + "type": "inline_equation", + "content": "\\{x^i\\}" + }, + { + "bbox": [ + 54, + 424, + 295, + 567 + ], + "type": "text", + "content": " to latent feature " + }, + { + "bbox": [ + 54, + 424, + 295, + 567 + ], + "type": "inline_equation", + "content": "\\{z^i\\}" + }, + { + "bbox": [ + 54, + 424, + 295, + 567 + ], + "type": "text", + "content": " by the pre-trained encoder " + }, + { + "bbox": [ + 54, + 424, + 295, + 567 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 54, + 424, + 295, + 567 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 54, + 424, + 295, + 567 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 54, + 424, + 295, + 567 + ], + "type": "text", + "content": " is omitted for clarity), followed by mapping camera parameters " + }, + { + "bbox": [ + 54, + 424, + 295, + 567 + ], + "type": "inline_equation", + "content": "\\{c^i\\}" + }, + { + "bbox": [ + 54, + 424, + 295, + 567 + ], + "type": "text", + "content": " as high-dimensional condition. Moreover, the original self-attention is replaced with adjacent-attention, which takes the previous view " + }, + { + "bbox": [ + 54, + 424, + 295, + 567 + ], + "type": "inline_equation", + "content": "z^{i-1}" + }, + { + "bbox": [ + 54, + 424, + 295, + 567 + ], + "type": "text", + "content": ", current view " + }, + { + "bbox": [ + 54, + 424, + 295, + 567 + ], + "type": "inline_equation", + "content": "z^i" + }, + { + "bbox": [ + 54, + 424, + 295, + 567 + ], + "type": "text", + "content": " and next view " + }, + { + "bbox": [ + 54, + 424, + 295, + 567 + ], + "type": "inline_equation", + "content": "z^{i+1}" + }, + { + "bbox": [ + 54, + 424, + 295, + 567 + ], + "type": "text", + "content": " as inputs, updating the features of " + }, + { + "bbox": [ + 54, + 424, + 295, + 567 + ], + "type": "inline_equation", + "content": "z^i" + }, + { + "bbox": [ + 54, + 424, + 295, + 567 + ], + "type": "text", + "content": ". The formulation of the adjacent-attention is as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 63, + 576, + 295, + 596 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 576, + 295, + 596 + ], + "spans": [ + { + "bbox": [ + 63, + 576, + 295, + 596 + ], + "type": "interline_equation", + "content": "Q ^ {i} = W _ {Q} z ^ {i}, \\tag {2}", + "image_path": "b96e3695682c4077bf89a4bfcb14fa699aaa9c930ffbe27d39c82991825382cf.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 63, + 594, + 276, + 605 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 594, + 276, + 605 + ], + "spans": [ + { + "bbox": [ + 63, + 594, + 276, + 605 + ], + "type": "interline_equation", + "content": "K ^ {i} = W _ {K} \\left[ z ^ {i - 1}, z ^ {i}, z ^ {i + 1} \\right], V ^ {i} = W _ {V} \\left[ z ^ {i - 1}, z ^ {i}, z ^ {i + 1} \\right],", + "image_path": "6a307ca9ee29b440fe89c2c8c8fed5a707f15909b8a3905922ac18880d735b4c.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "inline_equation", + "content": "[\\cdot]" + }, + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "text", + "content": " denotes concatenation operation. To inherit the generalizability of the 2D diffusion models, the parameters of self-attention is used to initialize projection matrices " + }, + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "inline_equation", + "content": "W_{Q}, W_{K}, W_{V}" + }, + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "text", + "content": " which are shared across space and view. Note that adjacent-frame attention, while capable of forming the initial appearance of the object, inevitably disperses attention due to the involvement of numerous interacting patches, making it struggle to focus on details." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 280, + 555, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 280, + 555, + 376 + ], + "spans": [ + { + "bbox": [ + 313, + 280, + 555, + 376 + ], + "type": "text", + "content": "Detail Refinement To ensure cross-view consistency in rendered images, it is crucial that the same position of an object corresponds to different pixels in neighboring viewpoints. These co-located pixels across views exhibit stronger correlations than others, playing a key role in maintaining visual consistency in multiview synthesis. For an object centered at the origin, and with some slight abuse of symbols, rigid rotation about the " + }, + { + "bbox": [ + 313, + 280, + 555, + 376 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 313, + 280, + 555, + 376 + ], + "type": "text", + "content": "-axis follows a physical principle:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 346, + 383, + 553, + 407 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 346, + 383, + 553, + 407 + ], + "spans": [ + { + "bbox": [ + 346, + 383, + 553, + 407 + ], + "type": "interline_equation", + "content": "x ^ {\\prime} = \\left(x - \\frac {W}{2}\\right) \\cos \\Delta \\alpha + \\frac {W}{2} - d \\sin \\Delta \\alpha , \\tag {3}", + "image_path": "7d2f89418642ebba06983254ed8a1b4b504fbaf1889f76f3fd28b6c55bedcd49.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 415, + 554, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 415, + 554, + 475 + ], + "spans": [ + { + "bbox": [ + 313, + 415, + 554, + 475 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 415, + 554, + 475 + ], + "type": "inline_equation", + "content": "(x,y)" + }, + { + "bbox": [ + 313, + 415, + 554, + 475 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 415, + 554, + 475 + ], + "type": "inline_equation", + "content": "(x',y')" + }, + { + "bbox": [ + 313, + 415, + 554, + 475 + ], + "type": "text", + "content": " are the pixel coordinates before and after rotation by an angle " + }, + { + "bbox": [ + 313, + 415, + 554, + 475 + ], + "type": "inline_equation", + "content": "\\Delta \\alpha" + }, + { + "bbox": [ + 313, + 415, + 554, + 475 + ], + "type": "text", + "content": " at a depth " + }, + { + "bbox": [ + 313, + 415, + 554, + 475 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 313, + 415, + 554, + 475 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 415, + 554, + 475 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 313, + 415, + 554, + 475 + ], + "type": "text", + "content": " is the image width. Since the depth " + }, + { + "bbox": [ + 313, + 415, + 554, + 475 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 313, + 415, + 554, + 475 + ], + "type": "text", + "content": " is unknown in the generative task and " + }, + { + "bbox": [ + 313, + 415, + 554, + 475 + ], + "type": "inline_equation", + "content": "\\sin \\Delta \\alpha" + }, + { + "bbox": [ + 313, + 415, + 554, + 475 + ], + "type": "text", + "content": " becomes negligible when " + }, + { + "bbox": [ + 313, + 415, + 554, + 475 + ], + "type": "inline_equation", + "content": "\\Delta \\alpha" + }, + { + "bbox": [ + 313, + 415, + 554, + 475 + ], + "type": "text", + "content": " is small, we simplify the transformation to:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 372, + 482, + 553, + 506 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 482, + 553, + 506 + ], + "spans": [ + { + "bbox": [ + 372, + 482, + 553, + 506 + ], + "type": "interline_equation", + "content": "x ^ {\\prime} = \\left(x - \\frac {W}{2}\\right) \\cos \\Delta \\alpha + \\frac {W}{2}. \\tag {4}", + "image_path": "4f61711f911487b265786e62ebe71269e85c2e8f607162aaf1e623c83c306c07.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 513, + 554, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 513, + 554, + 584 + ], + "spans": [ + { + "bbox": [ + 313, + 513, + 554, + 584 + ], + "type": "text", + "content": "To make this simplification robust in practical applications, we consider a " + }, + { + "bbox": [ + 313, + 513, + 554, + 584 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 313, + 513, + 554, + 584 + ], + "type": "text", + "content": " window " + }, + { + "bbox": [ + 313, + 513, + 554, + 584 + ], + "type": "inline_equation", + "content": "\\{x' + i, y' + j\\}_{i,j = -1,0,1}" + }, + { + "bbox": [ + 313, + 513, + 554, + 584 + ], + "type": "text", + "content": " in neighboring frames as potential results. The feature at pixel " + }, + { + "bbox": [ + 313, + 513, + 554, + 584 + ], + "type": "inline_equation", + "content": "(x,y)" + }, + { + "bbox": [ + 313, + 513, + 554, + 584 + ], + "type": "text", + "content": " is then refined through the attention mechanism that interacts with these neighboring pixels, enhancing consistency across views." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 591, + 500, + 602 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 591, + 500, + 602 + ], + "spans": [ + { + "bbox": [ + 313, + 591, + 500, + 602 + ], + "type": "text", + "content": "3.2.2. Know the Whole with Spiral Mamba" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 605, + 554, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 554, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 554, + 665 + ], + "type": "text", + "content": "To efficiently enhance global consistency, we first quickly scan all viewpoints, then down-sample patches based on textual relevance to filter out irrelevant information. This allows us to utilize powerful attention mechanisms to correct any accumulated inconsistencies." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "type": "text", + "content": "Rapid Glance Exchanging information across all viewpoints is crucial for achieving global view consistency. However, dense attention mechanisms [28, 47] can be computationally expensive and memory-intensive, while simple" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2883" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 85, + 126, + 152 + ], + "blocks": [ + { + "bbox": [ + 106, + 70, + 249, + 84 + ], + "lines": [ + { + "bbox": [ + 106, + 70, + 249, + 84 + ], + "spans": [ + { + "bbox": [ + 106, + 70, + 249, + 84 + ], + "type": "text", + "content": "Appearance Awareness" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 58, + 85, + 126, + 152 + ], + "lines": [ + { + "bbox": [ + 58, + 85, + 126, + 152 + ], + "spans": [ + { + "bbox": [ + 58, + 85, + 126, + 152 + ], + "type": "image", + "image_path": "9d96ae5ad7d78b09a813b424feea8010546057229a895255eaa82fdb19b2d352.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 69, + 154, + 111, + 163 + ], + "lines": [ + { + "bbox": [ + 69, + 154, + 111, + 163 + ], + "spans": [ + { + "bbox": [ + 69, + 154, + 111, + 163 + ], + "type": "text", + "content": "View i-1" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 141, + 85, + 211, + 152 + ], + "blocks": [ + { + "bbox": [ + 141, + 85, + 211, + 152 + ], + "lines": [ + { + "bbox": [ + 141, + 85, + 211, + 152 + ], + "spans": [ + { + "bbox": [ + 141, + 85, + 211, + 152 + ], + "type": "image", + "image_path": "a2b205d773e630c2d53d90ec172321ba039a5c46a27cb0bde5a138f7af44d043.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 162, + 154, + 195, + 163 + ], + "lines": [ + { + "bbox": [ + 162, + 154, + 195, + 163 + ], + "spans": [ + { + "bbox": [ + 162, + 154, + 195, + 163 + ], + "type": "text", + "content": "View i" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 227, + 85, + 294, + 152 + ], + "blocks": [ + { + "bbox": [ + 227, + 85, + 294, + 152 + ], + "lines": [ + { + "bbox": [ + 227, + 85, + 294, + 152 + ], + "spans": [ + { + "bbox": [ + 227, + 85, + 294, + 152 + ], + "type": "image", + "image_path": "061f00139e0655ec33ca40367068278e9e7501b50c2d6eb9f221ee62b2e6ba0e.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 238, + 154, + 282, + 163 + ], + "lines": [ + { + "bbox": [ + 238, + 154, + 282, + 163 + ], + "spans": [ + { + "bbox": [ + 238, + 154, + 282, + 163 + ], + "type": "text", + "content": "View i+1" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 59, + 180, + 121, + 244 + ], + "blocks": [ + { + "bbox": [ + 119, + 165, + 241, + 178 + ], + "lines": [ + { + "bbox": [ + 119, + 165, + 241, + 178 + ], + "spans": [ + { + "bbox": [ + 119, + 165, + 241, + 178 + ], + "type": "text", + "content": "Detail Refinement" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 59, + 180, + 121, + 244 + ], + "lines": [ + { + "bbox": [ + 59, + 180, + 121, + 244 + ], + "spans": [ + { + "bbox": [ + 59, + 180, + 121, + 244 + ], + "type": "image", + "image_path": "59c491c3de51fc3fb83723147d06035d4b8c4fdab6029148a30c15839bdb4e88.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 69, + 247, + 111, + 256 + ], + "lines": [ + { + "bbox": [ + 69, + 247, + 111, + 256 + ], + "spans": [ + { + "bbox": [ + 69, + 247, + 111, + 256 + ], + "type": "text", + "content": "View i-1" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 146, + 181, + 206, + 242 + ], + "blocks": [ + { + "bbox": [ + 146, + 181, + 206, + 242 + ], + "lines": [ + { + "bbox": [ + 146, + 181, + 206, + 242 + ], + "spans": [ + { + "bbox": [ + 146, + 181, + 206, + 242 + ], + "type": "image", + "image_path": "a68fa2d9f9d5ab7c34b28ead3de6296b21df59e0c88d013b15a96ccd2406d191.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 162, + 247, + 195, + 256 + ], + "lines": [ + { + "bbox": [ + 162, + 247, + 195, + 256 + ], + "spans": [ + { + "bbox": [ + 162, + 247, + 195, + 256 + ], + "type": "text", + "content": "View i" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 230, + 181, + 293, + 244 + ], + "blocks": [ + { + "bbox": [ + 230, + 181, + 293, + 244 + ], + "lines": [ + { + "bbox": [ + 230, + 181, + 293, + 244 + ], + "spans": [ + { + "bbox": [ + 230, + 181, + 293, + 244 + ], + "type": "image", + "image_path": "7231786a09ce8cd24ff97982134969ef61dcd73da76b10e6f23b63ac4f5a30fc.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 238, + 247, + 282, + 256 + ], + "lines": [ + { + "bbox": [ + 238, + 247, + 282, + 256 + ], + "spans": [ + { + "bbox": [ + 238, + 247, + 282, + 256 + ], + "type": "text", + "content": "View " + }, + { + "bbox": [ + 238, + 247, + 282, + 256 + ], + "type": "inline_equation", + "content": "i + 1" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 58, + 272, + 126, + 339 + ], + "blocks": [ + { + "bbox": [ + 58, + 272, + 126, + 339 + ], + "lines": [ + { + "bbox": [ + 58, + 272, + 126, + 339 + ], + "spans": [ + { + "bbox": [ + 58, + 272, + 126, + 339 + ], + "type": "image", + "image_path": "4043873aa485d1af714d00c970500fc1b961acb8881dac2af78ae5a77ee5ed2e.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 75, + 341, + 108, + 350 + ], + "lines": [ + { + "bbox": [ + 75, + 341, + 108, + 350 + ], + "spans": [ + { + "bbox": [ + 75, + 341, + 108, + 350 + ], + "type": "text", + "content": "View 1" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 146, + 274, + 208, + 335 + ], + "blocks": [ + { + "bbox": [ + 136, + 258, + 222, + 273 + ], + "lines": [ + { + "bbox": [ + 136, + 258, + 222, + 273 + ], + "spans": [ + { + "bbox": [ + 136, + 258, + 222, + 273 + ], + "type": "text", + "content": "Rapid Glance" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 146, + 274, + 208, + 335 + ], + "lines": [ + { + "bbox": [ + 146, + 274, + 208, + 335 + ], + "spans": [ + { + "bbox": [ + 146, + 274, + 208, + 335 + ], + "type": "image", + "image_path": "7e8afe8b0fdc20d36dc84ef32c47d75001d952aff5f41ebfeec549964911df24.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 162, + 341, + 195, + 350 + ], + "lines": [ + { + "bbox": [ + 162, + 341, + 195, + 350 + ], + "spans": [ + { + "bbox": [ + 162, + 341, + 195, + 350 + ], + "type": "text", + "content": "View i" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 227, + 272, + 294, + 339 + ], + "blocks": [ + { + "bbox": [ + 227, + 272, + 294, + 339 + ], + "lines": [ + { + "bbox": [ + 227, + 272, + 294, + 339 + ], + "spans": [ + { + "bbox": [ + 227, + 272, + 294, + 339 + ], + "type": "image", + "image_path": "4c171c536dea9b6845f135b1b56825924026d1fd254b4700ec6c49c9fb38233c.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 241, + 341, + 280, + 350 + ], + "lines": [ + { + "bbox": [ + 241, + 341, + 280, + 350 + ], + "spans": [ + { + "bbox": [ + 241, + 341, + 280, + 350 + ], + "type": "text", + "content": "View 12" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 227, + 353, + 294, + 365 + ], + "blocks": [ + { + "bbox": [ + 227, + 353, + 294, + 365 + ], + "lines": [ + { + "bbox": [ + 227, + 353, + 294, + 365 + ], + "spans": [ + { + "bbox": [ + 227, + 353, + 294, + 365 + ], + "type": "image", + "image_path": "3b79589b275602f7cb200b282006199e24b54951ac0c64644658e4ca1501c048.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 58, + 367, + 126, + 435 + ], + "blocks": [ + { + "bbox": [ + 58, + 353, + 126, + 366 + ], + "lines": [ + { + "bbox": [ + 58, + 353, + 126, + 366 + ], + "spans": [ + { + "bbox": [ + 58, + 353, + 126, + 366 + ], + "type": "text", + "content": "Accumulated" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 58, + 367, + 126, + 435 + ], + "lines": [ + { + "bbox": [ + 58, + 367, + 126, + 435 + ], + "spans": [ + { + "bbox": [ + 58, + 367, + 126, + 435 + ], + "type": "image", + "image_path": "4ffb6b32815ce9c6540fd786f80e825227046e5706aca771804401ce64cf98a7.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 75, + 436, + 108, + 445 + ], + "lines": [ + { + "bbox": [ + 75, + 436, + 108, + 445 + ], + "spans": [ + { + "bbox": [ + 75, + 436, + 108, + 445 + ], + "type": "text", + "content": "View 1" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 463, + 295, + 498 + ], + "lines": [ + { + "bbox": [ + 55, + 463, + 295, + 498 + ], + "spans": [ + { + "bbox": [ + 55, + 463, + 295, + 498 + ], + "type": "text", + "content": "Figure 3. Visualization of the interaction schemes in CoSER. For illustration, we denote each gird as a token with target token in red and attending token in green." + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 141, + 367, + 211, + 433 + ], + "blocks": [ + { + "bbox": [ + 129, + 353, + 211, + 365 + ], + "lines": [ + { + "bbox": [ + 129, + 353, + 211, + 365 + ], + "spans": [ + { + "bbox": [ + 129, + 353, + 211, + 365 + ], + "type": "text", + "content": "Inconsistency" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 141, + 367, + 211, + 433 + ], + "lines": [ + { + "bbox": [ + 141, + 367, + 211, + 433 + ], + "spans": [ + { + "bbox": [ + 141, + 367, + 211, + 433 + ], + "type": "image", + "image_path": "51a47692d1371df9144cfd06043ced87e24c1e308f211f22c40ac3c360fb2391.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 162, + 435, + 195, + 445 + ], + "lines": [ + { + "bbox": [ + 162, + 435, + 195, + 445 + ], + "spans": [ + { + "bbox": [ + 162, + 435, + 195, + 445 + ], + "type": "text", + "content": "View i" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 226, + 367, + 294, + 435 + ], + "blocks": [ + { + "bbox": [ + 226, + 367, + 294, + 435 + ], + "lines": [ + { + "bbox": [ + 226, + 367, + 294, + 435 + ], + "spans": [ + { + "bbox": [ + 226, + 367, + 294, + 435 + ], + "type": "image", + "image_path": "fea03fc6ce8f2e4446e382adb703622e9bbd969018338dd31cfaf2567c11cd62.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 241, + 436, + 279, + 445 + ], + "lines": [ + { + "bbox": [ + 241, + 436, + 279, + 445 + ], + "spans": [ + { + "bbox": [ + 241, + 436, + 279, + 445 + ], + "type": "text", + "content": "View 12" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 28 + }, + { + "bbox": [ + 55, + 510, + 295, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 295, + 700 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 295, + 700 + ], + "type": "text", + "content": "temporal attention [74] fails to effectively learn multi-view consistency. Inspired by Gu et al. [12], we leverage state space models (SSMs) known as linear operators to process all patches of multiview images, offering a more efficient alternative to quadratic-time attention. Specifically, we utilize the advanced selective scan SSMs introduced by Mamba [11] to implement our core SSMs operator. Unlike existing Vision Mamba methods that flatten 2D images into 1D sequences row-wise or column-wise, we propose an effective spiral bidirectional scan strategy. This strategy leverages the prior knowledge that objects in 3D datasets are typically centered. By organizing object-related content within a short sequence window, our approach implicitly mitigates the forgetting problem of SSMs. Additionally, placing background content across distinct viewpoints aids the SSMs in boundary perception." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "type": "text", + "content": "Technically, our method organizes spatial tokens start" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "text", + "content": "ing from the center and expanding in a spiral pattern, then stacks them view by view. Meanwhile, we reverse the scan order within the arrangement of viewpoints to capture robust space-view features. This allows us to perform a rapid glance as follows:" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 352, + 139, + 554, + 155 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 352, + 139, + 554, + 155 + ], + "spans": [ + { + "bbox": [ + 352, + 139, + 554, + 155 + ], + "type": "interline_equation", + "content": "\\left\\{z ^ {i} \\right\\} _ {i = 1} ^ {f} = S S M s \\left(S B S c a n \\left(\\left\\{z ^ {i} \\right\\} _ {i = 1} ^ {f}\\right)\\right), \\tag {5}", + "image_path": "7695f57ab1882e5739d8e6ac2c689de76cdc57f3fce891d8b3c68c2884e1b672.jpg" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 313, + 163, + 554, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 163, + 554, + 198 + ], + "spans": [ + { + "bbox": [ + 313, + 163, + 554, + 198 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 163, + 554, + 198 + ], + "type": "inline_equation", + "content": "SBScan(\\cdot)" + }, + { + "bbox": [ + 313, + 163, + 554, + 198 + ], + "type": "text", + "content": " denotes the spiral bidirectional scan. This operator effectively aggregates global information and provides crucial clues for further disambiguation." + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 313, + 199, + 555, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 199, + 555, + 390 + ], + "spans": [ + { + "bbox": [ + 313, + 199, + 555, + 390 + ], + "type": "text", + "content": "Accumulated Inconsistency Rectification To facilitate the elimination of inconsistencies through attention mechanisms, we implement down-sampling to filter out background and retain features relevant to textual information, thereby substantially reducing the number of patches engaged in interactions. Specifically, we provide a high-level semantic embedding " + }, + { + "bbox": [ + 313, + 199, + 555, + 390 + ], + "type": "inline_equation", + "content": "\\phi(p)" + }, + { + "bbox": [ + 313, + 199, + 555, + 390 + ], + "type": "text", + "content": " of the prompt " + }, + { + "bbox": [ + 313, + 199, + 555, + 390 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 313, + 199, + 555, + 390 + ], + "type": "text", + "content": ". A 2-layer MLP " + }, + { + "bbox": [ + 313, + 199, + 555, + 390 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 313, + 199, + 555, + 390 + ], + "type": "text", + "content": " is employed to assess the relevance between each patch and the text, producing a score ranging from 0 to 1 at each position, where a higher score indicates stronger relevance. Notably, we do not explicitly impose supervised constraints on the scoring mechanism. Instead, the model is driven by data to autonomously perceive the regions of interest. Based on the score map " + }, + { + "bbox": [ + 313, + 199, + 555, + 390 + ], + "type": "inline_equation", + "content": "s^i = S([z^i, \\phi(p)])" + }, + { + "bbox": [ + 313, + 199, + 555, + 390 + ], + "type": "text", + "content": " as a weight, we perform 2D-pooling along the spatial dimensions on the query, key, and value maps to enable sparse interactions:" + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 380, + 398, + 462, + 413 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 380, + 398, + 462, + 413 + ], + "spans": [ + { + "bbox": [ + 380, + 398, + 462, + 413 + ], + "type": "interline_equation", + "content": "\\hat {Q} ^ {i} = P \\left(s ^ {i} \\cdot \\bar {Q} ^ {i}; \\tau\\right),", + "image_path": "172711472cfc2c7ece0d7c97e1e1d10f9bb99881b03c2982a6b095f457e2103e.jpg" + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 380, + 415, + 553, + 431 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 380, + 415, + 553, + 431 + ], + "spans": [ + { + "bbox": [ + 380, + 415, + 553, + 431 + ], + "type": "interline_equation", + "content": "\\hat {K} ^ {i} = P \\left(\\left\\{s ^ {j} \\cdot \\bar {K} ^ {i} \\right\\} _ {j = 1} ^ {f}; \\rho\\right), \\tag {6}", + "image_path": "6f968ffed4a91da1590cf0a76d04a2870eb4f85562151e0495c0adac37538030.jpg" + } + ] + } + ], + "index": 38 + }, + { + "bbox": [ + 381, + 434, + 487, + 449 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 381, + 434, + 487, + 449 + ], + "spans": [ + { + "bbox": [ + 381, + 434, + 487, + 449 + ], + "type": "interline_equation", + "content": "\\hat {V} ^ {i} = P (\\{s ^ {j} \\cdot \\bar {V} ^ {i} \\} _ {j = 1} ^ {f}; \\rho).", + "image_path": "a0e55a8222cee486a6ae062699b04ee130fbc3e50dddbe85694bf9efa465e130.jpg" + } + ] + } + ], + "index": 39 + }, + { + "bbox": [ + 313, + 456, + 554, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 456, + 554, + 558 + ], + "spans": [ + { + "bbox": [ + 313, + 456, + 554, + 558 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 313, + 456, + 554, + 558 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 313, + 456, + 554, + 558 + ], + "type": "text", + "content": " denotes the pooling operator, " + }, + { + "bbox": [ + 313, + 456, + 554, + 558 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 456, + 554, + 558 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 456, + 554, + 558 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 313, + 456, + 554, + 558 + ], + "type": "text", + "content": " are the strides required for pooling, and the transformations include: " + }, + { + "bbox": [ + 313, + 456, + 554, + 558 + ], + "type": "inline_equation", + "content": "\\bar{Q}^i = W_{\\bar{Q}}z^i" + }, + { + "bbox": [ + 313, + 456, + 554, + 558 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 456, + 554, + 558 + ], + "type": "inline_equation", + "content": "\\bar{K}^i = W_{\\bar{K}}[\\{z^j\\}_{j=1}^f]" + }, + { + "bbox": [ + 313, + 456, + 554, + 558 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 456, + 554, + 558 + ], + "type": "inline_equation", + "content": "\\bar{V}^j = W_{\\bar{V}}[\\{z^j\\}_{j=1}^f]" + }, + { + "bbox": [ + 313, + 456, + 554, + 558 + ], + "type": "text", + "content": ". Notably, in contrast to the intensive down-sampling of " + }, + { + "bbox": [ + 313, + 456, + 554, + 558 + ], + "type": "inline_equation", + "content": "\\hat{V}^i" + }, + { + "bbox": [ + 313, + 456, + 554, + 558 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 456, + 554, + 558 + ], + "type": "inline_equation", + "content": "\\hat{K}^i" + }, + { + "bbox": [ + 313, + 456, + 554, + 558 + ], + "type": "text", + "content": ", we maintain a higher resolution for the query feature map " + }, + { + "bbox": [ + 313, + 456, + 554, + 558 + ], + "type": "inline_equation", + "content": "\\hat{Q}^i" + }, + { + "bbox": [ + 313, + 456, + 554, + 558 + ], + "type": "text", + "content": " to minimize the loss of content information. Finally, we up-sample the computed feature map using bilinear interpolation to restore its resolution." + } + ] + } + ], + "index": 40 + }, + { + "bbox": [ + 313, + 567, + 394, + 581 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 567, + 394, + 581 + ], + "spans": [ + { + "bbox": [ + 313, + 567, + 394, + 581 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 41 + }, + { + "bbox": [ + 313, + 588, + 554, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 588, + 554, + 658 + ], + "spans": [ + { + "bbox": [ + 313, + 588, + 554, + 658 + ], + "type": "text", + "content": "In this section, we first provide the implementation details and the dataset. We then conduct both qualitative and quantitative experiments to compare our method with previous state-of-the-art text-to-3D generation methods. Additionally, we perform an ablation study to demonstrate the effectiveness of the proposed modules." + } + ] + } + ], + "index": 42 + }, + { + "bbox": [ + 313, + 659, + 554, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 659, + 554, + 696 + ], + "spans": [ + { + "bbox": [ + 313, + 659, + 554, + 696 + ], + "type": "text", + "content": "Implementation We adopt the LDM" + }, + { + "bbox": [ + 313, + 659, + 554, + 696 + ], + "type": "inline_equation", + "content": "^1" + }, + { + "bbox": [ + 313, + 659, + 554, + 696 + ], + "type": "text", + "content": " [44], pre-trained on the LAION-5B [45], as our base model. Specifically, we use the AdamW optimizer [29] with a learning rate of 3e-5 and a" + } + ] + } + ], + "index": 43 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 324, + 703, + 538, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 703, + 538, + 712 + ], + "spans": [ + { + "bbox": [ + 324, + 703, + 538, + 712 + ], + "type": "text", + "content": "1https://github.com/CompVis/stable-diffusion" + } + ] + } + ], + "index": 44 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2884" + } + ] + } + ], + "index": 45 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 75, + 549, + 365 + ], + "blocks": [ + { + "bbox": [ + 59, + 75, + 549, + 365 + ], + "lines": [ + { + "bbox": [ + 59, + 75, + 549, + 365 + ], + "spans": [ + { + "bbox": [ + 59, + 75, + 549, + 365 + ], + "type": "image", + "image_path": "c925040cdf9122e8ea46408d7de0ca87c5b8f642525e739f37365747278a2596.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 225, + 366, + 381, + 376 + ], + "lines": [ + { + "bbox": [ + 225, + 366, + 381, + 376 + ], + "spans": [ + { + "bbox": [ + 225, + 366, + 381, + 376 + ], + "type": "text", + "content": "A bear dressed in medieval armor" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 59, + 410, + 552, + 688 + ], + "blocks": [ + { + "bbox": [ + 156, + 396, + 453, + 408 + ], + "lines": [ + { + "bbox": [ + 156, + 396, + 453, + 408 + ], + "spans": [ + { + "bbox": [ + 156, + 396, + 453, + 408 + ], + "type": "text", + "content": "Figure 4. Qualitative comparison of VideoMV [74] (Up) and our CoSER (Down)." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 59, + 410, + 552, + 688 + ], + "lines": [ + { + "bbox": [ + 59, + 410, + 552, + 688 + ], + "spans": [ + { + "bbox": [ + 59, + 410, + 552, + 688 + ], + "type": "image", + "image_path": "d82548c0429c60dbbb23905b38ecb44e5be752d9b352cdd385755f41fd1d0b99.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 138, + 706, + 470, + 718 + ], + "lines": [ + { + "bbox": [ + 138, + 706, + 470, + 718 + ], + "spans": [ + { + "bbox": [ + 138, + 706, + 470, + 718 + ], + "type": "text", + "content": "Figure 5. Qualitative comparisons of GaussianDreamer [69], Hash3D [68] and our CoSER." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2885" + } + ] + } + ], + "index": 5 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "type": "text", + "content": "batch size of 6 to optimize the weights over approximately 200K iterations on three NVIDIA Tesla A800 GPUs. To preserve diversity and detail, we degrade CoSER to a text-to-image generation model with a " + }, + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "type": "text", + "content": " probability and train it with a subset of LAION-5B. During inference, we employ the DDIM sampler with 50 steps and set the scale guidance to 7.5 for generation. Additionally, we set text template as \"A DSLR photo of [prompt], 3d asset\" to acquire more realistic results. Following the Liu et al. [27], we use NeuS[60] for performing reconstruction." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 193, + 294, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 193, + 294, + 289 + ], + "spans": [ + { + "bbox": [ + 55, + 193, + 294, + 289 + ], + "type": "text", + "content": "Dataset We use the public G-Objaverse [40] as the training dataset, utilizing the names of the objects appended with \"3D asset\" as their text labels. For each object, we first normalize the object to be centered within a bounding box ranging from " + }, + { + "bbox": [ + 55, + 193, + 294, + 289 + ], + "type": "inline_equation", + "content": "[-0.5, 0.5]" + }, + { + "bbox": [ + 55, + 193, + 294, + 289 + ], + "type": "text", + "content": ". The camera distance is set to 2.0 units. A random elevation is chosen between " + }, + { + "bbox": [ + 55, + 193, + 294, + 289 + ], + "type": "inline_equation", + "content": "[-30^\\circ, 30^\\circ]" + }, + { + "bbox": [ + 55, + 193, + 294, + 289 + ], + "type": "text", + "content": ", and starting from the front view, we uniformly sample 12 angles around the object for rendering." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 299, + 233, + 312 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 299, + 233, + 312 + ], + "spans": [ + { + "bbox": [ + 55, + 299, + 233, + 312 + ], + "type": "text", + "content": "4.1. Comparison with SOTA methods" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 317, + 295, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 317, + 295, + 556 + ], + "spans": [ + { + "bbox": [ + 56, + 317, + 295, + 556 + ], + "type": "text", + "content": "Qualitative Comparison We primarily evaluate the performance of the proposed CoSER on multiview text-to-image and compare the results with state-of-the-art methods in Figure 6. For VideoMV [74], 12 images generated from the same camera viewpoints as CoSER are selected for visualization, and all 24 viewpoints were utilized in the reconstruction process to ensure a fair comparison. Qualitatively, our method demonstrates impressive quality improvement in terms of both consistence and texture. Although VideoMV can generate images from a broader range of perspectives, e.g., mask and watch, it significantly compromises cross-view consistency, which poses serious challenges for subsequent 3D reconstruction. Furthermore, Figure 5 presents a comparison between two state-of-the-art text-to-3D methods, including GaussianDreamer [69] and Hash3D [68]. Specifically, the advanced method, Dream-Gaussian [54], is utilized as the foundational model for Hash3D. Thanks to the dense consistent images, CoSER outperforms other methods in terms of fidelity in 3D geometry and visual appearance." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "type": "text", + "content": "Quantitative Comparison Following previous work [74], we quantitatively assess the quality of the generated 3D assets using CLIP Score [14]. To evaluate the textual alignment, we followed [10] to compute the recall of the original prompt within the generated caption. We employ 100 text prompts selected from T3Bench [13] to generate 3D models. For each object, we set the elevation as " + }, + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "type": "inline_equation", + "content": "15^{\\circ}" + }, + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "type": "text", + "content": " and select 120 evenly spaced azimuth angles from " + }, + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "type": "inline_equation", + "content": "-180^{\\circ}" + }, + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "type": "inline_equation", + "content": "180^{\\circ}" + }, + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "type": "text", + "content": ", thereby generating 120 rendered images from different viewpoints. During the evaluation, 30 from the 120 rendered images are randomly selected and the CLIP Score is computed as the average semantic similarity between the prompts and these images. We use the CLIP ViT-B [43] to" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 553, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 300 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 300 + ], + "type": "text", + "content": "extract both text and image features. The results, presented in Table 1, demonstrate that our method significantly surpasses other methods, indicating that CoSER is more adept at injecting semantic information into the creation of realistic 3D assets. Given the challenges in automating the measurement of 3D consistency, we resort to user studies to assess which method produces results that are most favored by human evaluators. We conducted three user studies on the results in terms of the Semantic Alignment, Multiview Consistency and Texture Detail. In the first study, participants were asked to select the 3D asset that best aligns with the semantics of the given prompt. In the second study, participants identified the asset with the highest consistency. In the third study, participants were encouraged to choose the object with the finest texture and less distortion artifacts. We receive 100 answers on 30 text-3D pairs in total for each study. As shown in Table 1, over " + }, + { + "bbox": [ + 313, + 72, + 553, + 300 + ], + "type": "inline_equation", + "content": "29\\%" + }, + { + "bbox": [ + 313, + 72, + 553, + 300 + ], + "type": "text", + "content": " of our results are selected as the best in both three metrics, which proves a significant advantage in 3D generation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 311, + 408, + 323 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 311, + 408, + 323 + ], + "spans": [ + { + "bbox": [ + 313, + 311, + 408, + 323 + ], + "type": "text", + "content": "4.2. Ablation Study" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 329, + 553, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 329, + 553, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 329, + 553, + 628 + ], + "type": "text", + "content": "Effect of Module Design We are mainly interested in the 3D-aware module design, given that other unit has been well explored in previous methods. The core of our design is focusing on neighbors in fine-grained manner and knowing whole views with coarse-grained mechanism. To substantiate these claims, we incrementally integrated the proposed modules into four distinct configurations and finetuned the LDM accordingly. As depicted in Figure 6, the use of Appearance Awareness alone proves insufficient for ensuring cross-view consistency. However, the incorporation of Detail Refinement enhances the alignment between adjacent views, thereby improving overall 3D content consistency. It is evident that Rapid Glance offers small reduction in ambiguities; nonetheless, its primary benefit lies in facilitating information exchange across all viewpoints (See Figure 8), thereby establishing a foundation for the effective Accumulated Inconsistency Rectification module. Additionally, we present the parameters and MACs introduced by each module in Table 2. To illustrate that SSM captures multi-view consistency more effectively than temporal attention, we also learn global consistency by temporal attention found that it impaired performance. Based on this observation, we consider the additional 3.31M parameters and 15G MACs computations introduced by RG and AIR to be entirely acceptable." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 630, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 553, + 713 + ], + "type": "text", + "content": "Spiral Bidirectional Scan in Rapid Glance To verify the effectiveness of the proposed spiral bidirectional scan strategy, we perform the evaluation by replacing it with vanilla spatial-first bidirectional scan in VideoMamba [22]. As shown in Figure 7, both scanning strategies enhance the consistency. However, spatial-first bidirectional scan inevitably disrupts the continuity of object, which hampers" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "2886" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 62, + 94, + 380, + 153 + ], + "blocks": [ + { + "bbox": [ + 61, + 71, + 362, + 92 + ], + "lines": [ + { + "bbox": [ + 61, + 71, + 362, + 92 + ], + "spans": [ + { + "bbox": [ + 61, + 71, + 362, + 92 + ], + "type": "text", + "content": "Table 1. Quantitative comparison of CoSER and state-of-the-arts based on the Quality, Alignment and User Study." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 62, + 94, + 380, + 153 + ], + "lines": [ + { + "bbox": [ + 62, + 94, + 380, + 153 + ], + "spans": [ + { + "bbox": [ + 62, + 94, + 380, + 153 + ], + "type": "table", + "html": "
ModelQuality↑Alignment↑User Study
Semantic↑Consistency↑Texture↑
GaussianDreamer [69]30.7833.70.260.270.21
Hash3D [68]31.2630.20.200.190.25
VideoMV [74]32.1934.40.230.250.22
Ours33.0737.70.310.290.32
", + "image_path": "9e84dae141625baa174541eda61b5f4f07aa41d95aaf358f14da02d5c903c229.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 386, + 98, + 561, + 149 + ], + "blocks": [ + { + "bbox": [ + 386, + 75, + 547, + 96 + ], + "lines": [ + { + "bbox": [ + 386, + 75, + 547, + 96 + ], + "spans": [ + { + "bbox": [ + 386, + 75, + 547, + 96 + ], + "type": "text", + "content": "Table 2. Model efficiency of each module. TA denotes temporal attention." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 386, + 98, + 561, + 149 + ], + "lines": [ + { + "bbox": [ + 386, + 98, + 561, + 149 + ], + "spans": [ + { + "bbox": [ + 386, + 98, + 561, + 149 + ], + "type": "table", + "html": "
ModuleParameters (M)MACs (G)
AA867.451036
AA + DR917.041395
AA + DR + RG917.071396
AA + DR + RG + AIR969.951487
AA + DR + TA966.641472
", + "image_path": "3c6925f4791c34303aac14eaba23d3007f1aa8997d9c9e0c604a52302b49ec40.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 61, + 156, + 548, + 276 + ], + "blocks": [ + { + "bbox": [ + 61, + 156, + 548, + 276 + ], + "lines": [ + { + "bbox": [ + 61, + 156, + 548, + 276 + ], + "spans": [ + { + "bbox": [ + 61, + 156, + 548, + 276 + ], + "type": "image", + "image_path": "f205016f0384fc1af26b56a0c44100d5c1dc7bc3cb1268705c0f8474b2bc46cc.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 281, + 553, + 304 + ], + "lines": [ + { + "bbox": [ + 55, + 281, + 553, + 304 + ], + "spans": [ + { + "bbox": [ + 55, + 281, + 553, + 304 + ], + "type": "text", + "content": "Figure 6. Ablation of proposed moulds. The arrows indicate the changes made to the Appearance Awareness by sequentially plusing Detail Refinement, Rapid Glance, and Accumulated Inconsistency Rectification. Better see in color and " + }, + { + "bbox": [ + 55, + 281, + 553, + 304 + ], + "type": "inline_equation", + "content": "2\\mathrm{x}" + }, + { + "bbox": [ + 55, + 281, + 553, + 304 + ], + "type": "text", + "content": " zoom." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 57, + 306, + 293, + 390 + ], + "blocks": [ + { + "bbox": [ + 57, + 306, + 293, + 390 + ], + "lines": [ + { + "bbox": [ + 57, + 306, + 293, + 390 + ], + "spans": [ + { + "bbox": [ + 57, + 306, + 293, + 390 + ], + "type": "image", + "image_path": "768572bb6b40cbcfc096d97848991439a0c40871803f71f7adba73e16dbd9697.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 59, + 420, + 291, + 517 + ], + "blocks": [ + { + "bbox": [ + 55, + 397, + 295, + 418 + ], + "lines": [ + { + "bbox": [ + 55, + 397, + 295, + 418 + ], + "spans": [ + { + "bbox": [ + 55, + 397, + 295, + 418 + ], + "type": "text", + "content": "Figure 7. Ablation of scanning strategy. Spiral bidirectional scan generates semantic- and view-consistent images." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 59, + 420, + 291, + 517 + ], + "lines": [ + { + "bbox": [ + 59, + 420, + 291, + 517 + ], + "spans": [ + { + "bbox": [ + 59, + 420, + 291, + 517 + ], + "type": "image", + "image_path": "5106cb7081072979b910baca1b63665ead15d1be0a0c841edaa062028b7d9966.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 525, + 295, + 559 + ], + "lines": [ + { + "bbox": [ + 55, + 525, + 295, + 559 + ], + "spans": [ + { + "bbox": [ + 55, + 525, + 295, + 559 + ], + "type": "text", + "content": "Figure 8. Ablation of score map. Score map aids in capturing critical information, while Rapid Glance contributes to the generation of more accurate score maps." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 54, + 558, + 295, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 558, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 54, + 558, + 295, + 689 + ], + "type": "text", + "content": "the efficiency of SSM model for feature aggregation. In contrast, our proposed spiral bidirectional scan maximizes the spatial compactness of foreground content while simultaneously enforcing operator awareness of inter-view boundaries, thereby achieving better results. Additionally, we experimented with reversing the spiral scanning order within the image, progressing from the periphery inward towards the center. However, this strategy did not yield further performance improvements, which we attribute to the already efficient information aggregation within an image achieved by the spiral scan." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": "Score Map in Accumulated Inconsistency Rectification In Accumulated Inconsistency Rectification, we introduce the" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 316, + 555, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 316, + 555, + 472 + ], + "spans": [ + { + "bbox": [ + 313, + 316, + 555, + 472 + ], + "type": "text", + "content": "score map to assist in downsampling feature resolution for coarse attention, tailored to each individual view. To understand its impact, we compared the full AIR module generated results with those from two alternative models: AIR without score map and AIR with score map but without Rapid Glance. For clearness, we extract score maps at various resolutions and upsample them to 32x32 for visualization. As shown in Figure 8, the score map assigns higher scores to foreground regions, while background areas receive lower scores, facilitating the retention of critical information during the downsampling process. Additionally, Rapid Glance enhances the exchange of information, contributing to the generation of more precise score maps." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 477, + 388, + 489 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 477, + 388, + 489 + ], + "spans": [ + { + "bbox": [ + 313, + 477, + 388, + 489 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 312, + 498, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 498, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 312, + 498, + 555, + 713 + ], + "type": "text", + "content": "In this paper, we introduced CoSER, a novel framework for generating consistent and high-fidelity dense multiview images from text prompts. Our approach leverages a hybrid architecture that encapsulates the strengths of attention mechanisms and state space models (SSMs) to address the crucial challenges of cross-view consistency. The innovative design of contextual interaction includes an adjacent attention for capturing rough structures and textures, a trajectory-based attention for refining local details. A spiral bidirectional Mamba ensures global coherence in a cost-efficient manner. This comprehensive approach allows CoSER to produce dense, high-fidelity, and content-consistent multiview images. Extensive experiments demonstrate that CoSER outperforms state-of-the-art multiview synthesis methods in both quantitative and qualitative evaluations across various datasets. Future work could further explore the hybrid way of neighbor and global interaction to efficiently generate dense and consistent multiview images." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2887" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 159, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 159, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 159, + 85 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 175 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 175 + ], + "type": "text", + "content": "This paper is supported by the Beijing Natural Science Foundation (1254050), National Natural Science Foundation of China under Grants (U23B2012, 12471308), Fundamental Research Funds for the Central Universities, and by the National Research Foundation, Singapore, under its Medium Sized Center for Advanced Robotics Technology Innovation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 186, + 115, + 199 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 186, + 115, + 199 + ], + "spans": [ + { + "bbox": [ + 56, + 186, + 115, + 199 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 206, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 206, + 296, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 206, + 296, + 251 + ], + "spans": [ + { + "bbox": [ + 61, + 206, + 296, + 251 + ], + "type": "text", + "content": "[1] Haoyu Chen, Jingjing Ren, Jinjin Gu, Hongtao Wu, Xuequan Lu, Haoming Cai, and Lei Zhu. Snow removal in video: A new dataset and a novel method. in 2023 IEEE. In ICCV, pages 13165-13176. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 252, + 296, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 252, + 296, + 296 + ], + "spans": [ + { + "bbox": [ + 61, + 252, + 296, + 296 + ], + "type": "text", + "content": "[2] Haoyu Chen, Jinjin Gu, Yihao Liu, Salma Abdel Magid, Chao Dong, Qiong Wang, Hanspeter Pfister, and Lei Zhu. Masked image training for generalizable deep image denoising. In CVPR, pages 1692-1703, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 297, + 296, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 297, + 296, + 341 + ], + "spans": [ + { + "bbox": [ + 62, + 297, + 296, + 341 + ], + "type": "text", + "content": "[3] Rui Chen, Yongwei Chen, Ningxin Jiao, and Kui Jia. Fantasia3d: Disentangling geometry and appearance for high-quality text-to-3d content creation. In ICCV, pages 22246–22256, 2023. 1, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 342, + 296, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 342, + 296, + 386 + ], + "spans": [ + { + "bbox": [ + 62, + 342, + 296, + 386 + ], + "type": "text", + "content": "[4] Yongwei Chen, Rui Chen, Jiabao Lei, Yabin Zhang, and Kui Jia. Tango: Text-driven photorealistic and robust 3d stylization via lighting decomposition. In NeurIPS, pages 30923-30936, 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 388, + 296, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 388, + 296, + 441 + ], + "spans": [ + { + "bbox": [ + 62, + 388, + 296, + 441 + ], + "type": "text", + "content": "[5] Yiwen Chen, Zilong Chen, Chi Zhang, Feng Wang, Xiaofeng Yang, Yikai Wang, Zhongang Cai, Lei Yang, Huaping Liu, and Guosheng Lin. Gaussianeditor: Swift and controllable 3d editing with gaussian splatting. In CVPR, pages 21476-21485, 2024. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 444, + 296, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 444, + 296, + 477 + ], + "spans": [ + { + "bbox": [ + 62, + 444, + 296, + 477 + ], + "type": "text", + "content": "[6] Zilong Chen, Yikai Wang, Feng Wang, Zhengyi Wang, and Huaping Liu. V3d: Video diffusion models are effective 3d generators. arXiv preprint arXiv:2403.06738, 2024. 2, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 478, + 296, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 478, + 296, + 521 + ], + "spans": [ + { + "bbox": [ + 62, + 478, + 296, + 521 + ], + "type": "text", + "content": "[7] Yen-Chi Cheng, Hsin-Ying Lee, Sergey Tulyakov, Alexander G Schwing, and Liang-Yan Gui. Sdfusion: Multimodal 3d shape completion, reconstruction, and generation. In CVPR, pages 4456-4465, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 523, + 296, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 523, + 296, + 567 + ], + "spans": [ + { + "bbox": [ + 62, + 523, + 296, + 567 + ], + "type": "text", + "content": "[8] Lihe Ding, Shaocong Dong, Zhanpeng Huang, Zibin Wang, Yiyuan Zhang, Kaixiong Gong, Dan Xu, and Tianfan Xue. Text-to-3d generation with bidirectional diffusion using both 2d and 3d priors. In CVPR, pages 5115–5124, 2024. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 568, + 296, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 568, + 296, + 622 + ], + "spans": [ + { + "bbox": [ + 62, + 568, + 296, + 622 + ], + "type": "text", + "content": "[9] Ruiqi Gao, Aleksander Holynski, Philipp Henzler, Arthur Brussee, Ricardo Martin-Brualla, Pratul Srinivasan, Jonathan T Barron, and Ben Poole. Cat3d: Create anything in 3d with multi-view diffusion models. arXiv preprint arXiv:2405.10314, 2024. 2, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 624, + 296, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 624, + 296, + 679 + ], + "spans": [ + { + "bbox": [ + 57, + 624, + 296, + 679 + ], + "type": "text", + "content": "[10] Chongjian Ge, Chenfeng Xu, Yuanfeng Ji, Chensheng Peng, Masayoshi Tomizuka, Ping Luo, Mingyu Ding, Varun Jampani, and Wei Zhan. Compgs: Unleashing 2d compositionality for compositional text-to-3d via dynamically optimizing 3d gaussians. arXiv preprint arXiv:2410.20723, 2024. 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 681, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 681, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 681, + 296, + 713 + ], + "type": "text", + "content": "[11] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023. 5" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "text", + "content": "[12] Albert Gu, Karan Goel, and Christopher Re. Efficiently modeling long sequences with structured state spaces. In ICLR, 2022. 2, 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 107, + 553, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 107, + 553, + 152 + ], + "spans": [ + { + "bbox": [ + 317, + 107, + 553, + 152 + ], + "type": "text", + "content": "[13] Yuze He, Yushi Bai, Matthieu Lin, Wang Zhao, Yubin Hu, Jenny Sheng, Ran Yi, Juanzi Li, and Yong-Jin Liu. T3 bench: Benchmarking current progress in text-to-3d generation. arXiv preprint arXiv:2310.02977, 2023. 1, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 152, + 553, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 553, + 195 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 553, + 195 + ], + "type": "text", + "content": "[14] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning. In EMNLP, pages 7514-7528, 2021. 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 198, + 553, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 198, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 316, + 198, + 553, + 251 + ], + "type": "text", + "content": "[15] Fangzhou Hong, Jiaxiang Tang, Ziang Cao, Min Shi, Tong Wu, Zhaoxi Chen, Tengfei Wang, Liang Pan, Dahua Lin, and Ziwei Liu. 3dtopia: Large text-to-3d generation model with hybrid diffusion priors. arXiv preprint arXiv:2403.02234, 2024. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 254, + 553, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 254, + 553, + 297 + ], + "spans": [ + { + "bbox": [ + 316, + 254, + 553, + 297 + ], + "type": "text", + "content": "[16] Yukun Huang, Jianan Wang, Yukai Shi, Xianbiao Qi, Zheng-Jun Zha, and Lei Zhang. Dreamtime: An improved optimization strategy for text-to-3d content creation. arXiv preprint arXiv:2306.12422, 2023. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 299, + 553, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 299, + 553, + 353 + ], + "spans": [ + { + "bbox": [ + 316, + 299, + 553, + 353 + ], + "type": "text", + "content": "[17] Zehuan Huang, Hao Wen, Junting Dong, Yaohui Wang, Yangguang Li, Xinyuan Chen, Yan-Pei Cao, Ding Liang, Yu Qiao, Bo Dai, et al. Epidiff: Enhancing multi-view synthesis via localized epipolar-constrained diffusion. In CVPR, pages 9784-9794, 2024. 2, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 354, + 553, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 354, + 553, + 397 + ], + "spans": [ + { + "bbox": [ + 316, + 354, + 553, + 397 + ], + "type": "text", + "content": "[18] Ajay Jain, Ben Mildenhall, Jonathan T Barron, Pieter Abbeel, and Ben Poole. Zero-shot text-guided object generation with dream fields. In CVPR, pages 867-876, 2022. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 399, + 553, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 399, + 553, + 443 + ], + "spans": [ + { + "bbox": [ + 316, + 399, + 553, + 443 + ], + "type": "text", + "content": "[19] Jeong-gi Kwak, Erqun Dong, Yuhe Jin, Hanseok Ko, Shweta Mahajan, and Kwang Moo Yi. Vivid-1-to-3: Novel view synthesis with video diffusion models. In CVPR, pages 6775–6785, 2024. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 444, + 553, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 444, + 553, + 477 + ], + "spans": [ + { + "bbox": [ + 316, + 444, + 553, + 477 + ], + "type": "text", + "content": "[20] Bonan Li, Zicheng Zhang, Xuecheng Nie, Congying Han, Yinhan Hu, and Tiande Guo. Styo: Stylize your face in only one-shot. arXiv preprint arXiv:2303.03231, 2023. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 478, + 553, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 478, + 553, + 533 + ], + "spans": [ + { + "bbox": [ + 316, + 478, + 553, + 533 + ], + "type": "text", + "content": "[21] Jiahao Li, Hao Tan, Kai Zhang, Zexiang Xu, Fujun Luan, Yinghao Xu, Yicong Hong, Kalyan Sunkavalli, Greg Shakhnarovich, and Sai Bi. Instant3d: Fast text-to-3d with sparse-view generation and large reconstruction model. In ICLR, 2023. 2, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 535, + 553, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 535, + 553, + 578 + ], + "spans": [ + { + "bbox": [ + 316, + 535, + 553, + 578 + ], + "type": "text", + "content": "[22] Kunchang Li, Xinhao Li, Yi Wang, Yinan He, Yali Wang, Limin Wang, and Yu Qiao. Videomamba: State space model for efficient video understanding. arXiv preprint arXiv:2403.06977, 2024. 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 580, + 553, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 580, + 553, + 634 + ], + "spans": [ + { + "bbox": [ + 316, + 580, + 553, + 634 + ], + "type": "text", + "content": "[23] Peng Li, Yuan Liu, Xiaoxiao Long, Feihu Zhang, Cheng Lin, Mengfei Li, Xingqun Qi, Shanghang Zhang, Wenhan Luo, Ping Tan, et al. Era3d: High-resolution multiview diffusion using efficient row-wise attention. arXiv preprint arXiv:2405.11616, 2024. 2, 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 635, + 553, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 553, + 669 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 553, + 669 + ], + "type": "text", + "content": "[24] Weiyu Li, Rui Chen, Xuelin Chen, and Ping Tan. Sweetdreamer: Aligning geometric priors in 2d diffusion for consistent text-to-3d. In ICLR, 2024. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "text", + "content": "[25] Wenbing Li, Hang Zhou, Junqing Yu, Zikai Song, and Wei Yang. Coupled mamba: Enhanced multimodal fusion with coupled state space model. In NeurIPS, pages 59808-59832, 2025. 2" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "2888" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 714 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "text", + "content": "[26] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. In CVPR, pages 300–309, 2023. 1, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 294, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 294, + 173 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 294, + 173 + ], + "type": "text", + "content": "[27] Yuan Liu, Cheng Lin, Zijiao Zeng, Xiaoxiao Long, Lingjie Liu, Taku Komura, and Wenping Wang. Syncdreamer: Generating multiview-consistent images from a single-view image. In ICLR, 2024. 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 175, + 294, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 294, + 228 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 294, + 228 + ], + "type": "text", + "content": "[28] Xiaoxiao Long, Yuan-Chen Guo, Cheng Lin, Yuan Liu, Zhiyang Dou, Lingjie Liu, Yuexin Ma, Song-Hai Zhang, Marc Habermann, Christian Theobalt, et al. Wonder3d: Single image to 3d using cross-domain diffusion. In CVPR, pages 9970-9980, 2024. 2, 3, 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 230, + 294, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 230, + 294, + 252 + ], + "spans": [ + { + "bbox": [ + 57, + 230, + 294, + 252 + ], + "type": "text", + "content": "[29] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In ICLR, 2019. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 57, + 254, + 294, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 254, + 294, + 285 + ], + "spans": [ + { + "bbox": [ + 57, + 254, + 294, + 285 + ], + "type": "text", + "content": "[30] Run Luo, Zikai Song, Lintao Ma, Jinlin Wei, Wei Yang, and Min Yang. Diffusiontrack: Diffusion model for multi-object tracking. In AAAI, pages 3991-3999, 2024. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 57, + 286, + 294, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 286, + 294, + 331 + ], + "spans": [ + { + "bbox": [ + 57, + 286, + 294, + 331 + ], + "type": "text", + "content": "[31] Baorui Ma, Haoge Deng, Junsheng Zhou, Yu-Shen Liu, Tiejun Huang, and Xinlong Wang. Geodream: Disentangling 2d and geometric priors for high-fidelity and consistent 3d generation. arXiv preprint arXiv:2311.17971, 2023. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 332, + 294, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 332, + 294, + 376 + ], + "spans": [ + { + "bbox": [ + 56, + 332, + 294, + 376 + ], + "type": "text", + "content": "[32] Luke Melas-Kyriazi, Iro Laina, Christian Rupprecht, Natalia Neverova, Andrea Vedaldi, Oran Gafni, and Filippos Kokkinos. Im-3d: Iterative multiview diffusion and reconstruction for high-quality 3d generation. In ICML, 2024. 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 57, + 377, + 294, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 377, + 294, + 420 + ], + "spans": [ + { + "bbox": [ + 57, + 377, + 294, + 420 + ], + "type": "text", + "content": "[33] Gal Metzer, Elad Richardson, Or Patashnik, Raja Giryes, and Daniel Cohen-Or. Latent-nerf for shape-guided generation of 3d shapes and textures. In CVPR, pages 12663–12673, 2023. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 422, + 294, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 422, + 294, + 464 + ], + "spans": [ + { + "bbox": [ + 56, + 422, + 294, + 464 + ], + "type": "text", + "content": "[34] Gal Metzer, Elad Richardson, Or Patashnik, Raja Giryes, and Daniel Cohen-Or. Latent-nerf for shape-guided generation of 3d shapes and textures. In CVPR, pages 12663–12673, 2023. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 57, + 467, + 294, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 467, + 294, + 510 + ], + "spans": [ + { + "bbox": [ + 57, + 467, + 294, + 510 + ], + "type": "text", + "content": "[35] Nasir Mohammad Khalid, Tianhao Xie, Eugene Belilovsky, and Tiberiu Popa. Clip-mesh: Generating textured meshes from text using pretrained image-text models. In SIGGRAPH Asia, pages 1-8, 2022. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 512, + 294, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 512, + 294, + 555 + ], + "spans": [ + { + "bbox": [ + 56, + 512, + 294, + 555 + ], + "type": "text", + "content": "[36] Alex Nichol, Heewoo Jun, Prafulla Dhariwal, Pamela Mishkin, and Mark Chen. Point-e: A system for generating 3d point clouds from complex prompts. arXiv preprint arXiv:2212.08751, 2022. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 556, + 294, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 556, + 294, + 589 + ], + "spans": [ + { + "bbox": [ + 56, + 556, + 294, + 589 + ], + "type": "text", + "content": "[37] JangHo Park, Gihyun Kwon, and Jong Chul Ye. Ed-nerf: Efficient text-guided editing of 3d scene with latent space nerf. In ICLR, 2024. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 590, + 294, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 590, + 294, + 622 + ], + "spans": [ + { + "bbox": [ + 56, + 590, + 294, + 622 + ], + "type": "text", + "content": "[38] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. In ICLR, 2022. 1, 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 624, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 294, + 689 + ], + "type": "text", + "content": "[39] Guocheng Qian, Jinjie Mai, Abdullah Hamdi, Jian Ren, Aliaksandr Siarohin, Bing Li, Hsin-Ying Lee, Ivan Skorokhodov, Peter Wonka, Sergey Tulyakov, et al. Magic123: One image to high-quality 3d object generation using both 2d and 3d diffusion priors. arXiv preprint arXiv:2306.17843, 2023. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 691, + 294, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 294, + 714 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 294, + 714 + ], + "type": "text", + "content": "[40] Lingteng Qiu, Guanying Chen, Xiaodong Gu, Qi Zuo, Mutian Xu, Yushuang Wu, Weihao Yuan, Zilong Dong, Liefeng" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 714 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "text", + "content": "Bo, and Xiaoguang Han. Richdreamer: A generalizable normal-depth diffusion model for detail richness in text-to-3d. In CVPR, pages 9914–9925, 2024. 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 107, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 150 + ], + "type": "text", + "content": "[41] Xinmin Qiu, Congying Han, ZiCheng Zhang, Bonan Li, Tiande Guo, and Xuecheng Nie. Diffbfr: Bootstrapping diffusion model for blind face restoration. In ACM MM, pages 7785-7795, 2023. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 152, + 553, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 553, + 207 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 553, + 207 + ], + "type": "text", + "content": "[42] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763, 2021. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 209, + 553, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 209, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 316, + 209, + 553, + 262 + ], + "type": "text", + "content": "[43] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763, 2021. 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 264, + 553, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 264, + 553, + 308 + ], + "spans": [ + { + "bbox": [ + 316, + 264, + 553, + 308 + ], + "type": "text", + "content": "[44] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10684-10695, 2022. 2, 3, 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 309, + 553, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 309, + 553, + 396 + ], + "spans": [ + { + "bbox": [ + 316, + 309, + 553, + 396 + ], + "type": "text", + "content": "[45] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, Patrick Schramowski, Srivatsa Kundurthy, Katherine Crowson, Ludwig Schmidt, Robert Kaczmarczyk, and Jenia Jitsev. Laion-5b: An open large-scale dataset for training next generation image-text models. ArXiv, abs/2210.08402, 2022. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 398, + 553, + 452 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 398, + 553, + 452 + ], + "spans": [ + { + "bbox": [ + 316, + 398, + 553, + 452 + ], + "type": "text", + "content": "[46] Junyoung Seo, Wooseok Jang, Min-Seop Kwak, Hyeonsu Kim, Jaehoon Ko, Junho Kim, Jin-Hwa Kim, Jiyoung Lee, and Seungryong Kim. Let 2d diffusion model know 3d-consistency for robust text-to-3d generation. In ICLR, 2024. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 454, + 553, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 454, + 553, + 486 + ], + "spans": [ + { + "bbox": [ + 316, + 454, + 553, + 486 + ], + "type": "text", + "content": "[47] Yichun Shi, Peng Wang, Jianglong Ye, Long Mai, Kejie Li, and Xiao Yang. Mvdream: Multi-view diffusion for 3d generation. In ICLR, 2023. 2, 3, 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 488, + 553, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 488, + 553, + 520 + ], + "spans": [ + { + "bbox": [ + 316, + 488, + 553, + 520 + ], + "type": "text", + "content": "[48] Zikai Song, Junqing Yu, Yi-Ping Phoebe Chen, and Wei Yang. Transformer tracking with cyclic shifting window attention. In CVPR, pages 8791-8800, 2022. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 522, + 553, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 522, + 553, + 555 + ], + "spans": [ + { + "bbox": [ + 316, + 522, + 553, + 555 + ], + "type": "text", + "content": "[49]Zikai Song, Run Luo, Junqing Yu, Yi-Ping Phoebe Chen, and Wei Yang. Compact transformer tracker with correlative masked modeling. In AAAI, pages 2321-2329, 2023." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 556, + 553, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 556, + 553, + 599 + ], + "spans": [ + { + "bbox": [ + 316, + 556, + 553, + 599 + ], + "type": "text", + "content": "[50] Zikai Song, Ying Tang, Run Luo, Lintao Ma, Junqing Yu, Yi-Ping Phoebe Chen, and Wei Yang. Autogenic language embedding for coherent point tracking. In ACM MM, pages 2021-2030, 2024." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 601, + 553, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 601, + 553, + 645 + ], + "spans": [ + { + "bbox": [ + 316, + 601, + 553, + 645 + ], + "type": "text", + "content": "[51] Haoze Sun, Wenbo Li, Jianzhuang Liu, Haoyu Chen, Renjing Pei, Xueyi Zou, Youliang Yan, and Yujiu Yang. Coser: Bridging image and language for cognitive super-resolution. In CVPR, pages 25868-25878, 2024. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "type": "text", + "content": "[52] Jingxiang Sun, Bo Zhang, Ruizhi Shao, Lizhen Wang, Wen Liu, Zhenda Xie, and Yebin Liu. Dreamcraft3d: Hierarchical 3d generation with bootstrapped diffusion prior. In ICLR, 2024. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 691, + 553, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 714 + ], + "type": "text", + "content": "[53] Junshu Tang, Tengfei Wang, Bo Zhang, Ting Zhang, Ran Yi, Lizhuang Ma, and Dong Chen. Make-it-3d: High-fidelity 3d" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2889" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "text", + "content": "creation from a single image with diffusion prior. In ICCV, pages 22819-22829, 2023. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 128 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 128 + ], + "type": "text", + "content": "[54] Jiaxiang Tang, Jiawei Ren, Hang Zhou, Ziwei Liu, and Gang Zeng. Dreamgaussian: Generative gaussian splatting for efficient 3d content creation. In ICLR, 2024. 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 130, + 294, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 130, + 294, + 173 + ], + "spans": [ + { + "bbox": [ + 56, + 130, + 294, + 173 + ], + "type": "text", + "content": "[55] Shitao Tang, Fuyang Zhang, Jiacheng Chen, Peng Wang, and Yasutaka Furukawa. Mvdiffusion: Enabling holistic multiview image generation with correspondence-aware diffusion. arXiv, 2023. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 175, + 294, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 294, + 218 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 294, + 218 + ], + "type": "text", + "content": "[56] Christina Tsalicoglou, Fabian Manhardt, Alessio Tonioni, Michael Niemeyer, and Federico Tombari. Textmesh: Generation of realistic 3d meshes from text prompts. In 3DV, pages 1554-1563, 2024. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 220, + 294, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 220, + 294, + 252 + ], + "spans": [ + { + "bbox": [ + 56, + 220, + 294, + 252 + ], + "type": "text", + "content": "[57] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 254, + 294, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 254, + 294, + 317 + ], + "spans": [ + { + "bbox": [ + 56, + 254, + 294, + 317 + ], + "type": "text", + "content": "[58] Vikram Voleti, Chun-Han Yao, Mark Boss, Adam Letts, David Pankratz, Dmitry Tochilkin, Christian Laforte, Robin Rombach, and Varun Jampani. Sv3d: Novel multi-view synthesis and 3d generation from a single image using latent video diffusion. arXiv preprint arXiv:2403.12008, 2024. 2, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 320, + 294, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 320, + 294, + 352 + ], + "spans": [ + { + "bbox": [ + 56, + 320, + 294, + 352 + ], + "type": "text", + "content": "[59] Peng Wang and Yichun Shi. Imagedream: Image-prompt multi-view diffusion for 3d generation. arXiv preprint arXiv:2312.02201, 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 354, + 294, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 354, + 294, + 397 + ], + "spans": [ + { + "bbox": [ + 56, + 354, + 294, + 397 + ], + "type": "text", + "content": "[60] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. In NeurIPS, 2021. 3, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 399, + 294, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 399, + 294, + 453 + ], + "spans": [ + { + "bbox": [ + 56, + 399, + 294, + 453 + ], + "type": "text", + "content": "[61] Tengfei Wang, Bo Zhang, Ting Zhang, Shuyang Gu, Jianmin Bao, Tadas Baltrusaitis, Jingjing Shen, Dong Chen, Fang Wen, Qifeng Chen, et al. Rodin: A generative model for sculpting 3d digital avatars using diffusion. In CVPR, pages 4563-4573, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 455, + 294, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 455, + 294, + 510 + ], + "spans": [ + { + "bbox": [ + 56, + 455, + 294, + 510 + ], + "type": "text", + "content": "[62] Xinzhou Wang, Yikai Wang, Junliang Ye, Zhengyi Wang, Fuchun Sun, Pengkun Liu, Ling Wang, Kai Sun, Xintong Wang, and Bin He. Animatabledreamer: Text-guided non-rigid 3d model generation and reconstruction with canonical score distillation. arXiv preprint arXiv:2312.03795, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 510, + 294, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 510, + 294, + 554 + ], + "spans": [ + { + "bbox": [ + 56, + 510, + 294, + 554 + ], + "type": "text", + "content": "[63] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. In NeurIPS, 2024. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 555, + 294, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 555, + 294, + 588 + ], + "spans": [ + { + "bbox": [ + 56, + 555, + 294, + 588 + ], + "type": "text", + "content": "[64] Zeyu Xiao and Zhiwei Xiong. Incorporating degradation estimation in light field spatial super-resolution. Computer Vision and Image Understanding, page 104295, 2025. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 590, + 294, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 590, + 294, + 622 + ], + "spans": [ + { + "bbox": [ + 56, + 590, + 294, + 622 + ], + "type": "text", + "content": "[65] Zeyu Xiao, Yutong Liu, Ruisheng Gao, and Zhiwei Xiong. Cutmib: Boosting light field super-resolution via multi-view image blending. In CVPR, pages 1672-1682, 2023." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 624, + 294, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 294, + 656 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 294, + 656 + ], + "type": "text", + "content": "[66] Zeyu Xiao, Dachun Kai, Yueyi Zhang, Xiaoyan Sun, and Zhiwei Xiong. Asymmetric event-guided video superresolution. In ACM MM, pages 2409-2418, 2024." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 658, + 294, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 294, + 678 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 294, + 678 + ], + "type": "text", + "content": "[67] Xinchao Wang Xingyi Yang. Kolmogorov-arnold transformer. In ICLR, 2025. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 56, + 680, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 680, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 680, + 294, + 712 + ], + "type": "text", + "content": "[68] Xingyi Yang and Xinchao Wang. Hash3d: Training-free acceleration for 3d generation. arXiv preprint arXiv:2404.06091, 2024. 3, 6, 7, 8" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 329 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 128 + ], + "type": "text", + "content": "[69] Taoran Yi, Jiemin Fang, Junjie Wang, Guanjun Wu, Lingxi Xie, Xiaopeng Zhang, Wenyu Liu, Qi Tian, and Xinggang Wang. Gaussian dreamer: Fast generation from text to 3d gaussians by bridging 2d and 3d diffusion models. In CVPR, pages 6796-6807, 2024. 3, 6, 7, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 129, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 553, + 150 + ], + "type": "text", + "content": "[70] Weihao Yu and Xinchao Wang. Mambaout: Do we really need mamba for vision? In CVPR, 2025. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 152, + 553, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 553, + 195 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 553, + 195 + ], + "type": "text", + "content": "[71] Xianfang Zeng, Xin Chen, Zhongqi Qi, Wen Liu, Zibo Zhao, Zhibin Wang, Bin Fu, Yong Liu, and Gang Yu. Paint3d: Paint anything 3d with lighting-less texture diffusion models. In CVPR, pages 4252-4262, 2024. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 197, + 553, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 197, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 316, + 197, + 553, + 239 + ], + "type": "text", + "content": "[72] Zicheng Zhang, Bonan Li, Xuecheng Nie, Congying Han, Tiande Guo, and Luoqi Liu. Towards consistent video editing with text-to-image diffusion models. In NeurIPS, pages 58508-58519, 2023. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 241, + 553, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 241, + 553, + 274 + ], + "spans": [ + { + "bbox": [ + 316, + 241, + 553, + 274 + ], + "type": "text", + "content": "[73] Jingyu Zhuang, Chen Wang, Liang Lin, Lingjie Liu, and Guanbin Li. Dreameditor: Text-driven 3d scene editing with neural fields. In SIGGRAPH Asia, pages 1-10, 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 276, + 553, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 276, + 553, + 329 + ], + "spans": [ + { + "bbox": [ + 316, + 276, + 553, + 329 + ], + "type": "text", + "content": "[74] Qi Zuo, Xiaodong Gu, Lingteng Qiu, Yuan Dong, Zhengyi Zhao, Weihao Yuan, Rui Peng, Siyu Zhu, Zilong Dong, Liefeng Bo, et al. Videomv: Consistent multi-view generation based on large video generative model. arXiv preprint arXiv:2403.12010, 2024. 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "2890" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/CoSpace_ Benchmarking Continuous Space Perception Ability for Vision-Language Models/45f50ce2-59a4-414e-86e0-abbb59b51daa_content_list.json b/2025/CoSpace_ Benchmarking Continuous Space Perception Ability for Vision-Language Models/45f50ce2-59a4-414e-86e0-abbb59b51daa_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..ee2b7d337885885739a6d732d13f435e00a6477f --- /dev/null +++ b/2025/CoSpace_ Benchmarking Continuous Space Perception Ability for Vision-Language Models/45f50ce2-59a4-414e-86e0-abbb59b51daa_content_list.json @@ -0,0 +1,2512 @@ +[ + { + "type": "text", + "text": "CoSpace: Benchmarking Continuous Space Perception Ability for Vision-Language Models", + "text_level": 1, + "bbox": [ + 184, + 130, + 816, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yiqi Zhu $^{1*}$ , Ziyue Wang $^{1*}$ , Can Zhang $^{3}$ , Peng Li $^{2,4\\dagger}$ , Yang Liu $^{1,2,4,5\\dagger}$", + "bbox": [ + 236, + 202, + 769, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Dept. of Comp. Sci. & Tech., Institute for AI, Tsinghua University, Beijing, China", + "bbox": [ + 176, + 220, + 833, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Institute for AI Industry Research (AIR), Tsinghua University, Beijing, China", + "bbox": [ + 194, + 239, + 813, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{3}$ School of Computer and Communication Engineering, University of Science and Technology Beijing", + "bbox": [ + 101, + 256, + 906, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{4}$ Shanghai Artificial Intelligence Laboratory, Shanghai, China", + "bbox": [ + 259, + 275, + 748, + 291 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{5}$ Jiangsu Collaborative Innovation Center for Language Competence, Jiangsu, China", + "bbox": [ + 169, + 292, + 834, + 310 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e7ef0ad57d49aabffbf8bb5e3d9aaee49e8be2cfeddc8d8e0c6db3eb5d4b33f9.jpg", + "image_caption": [ + "The way models perceive real-world scenarios" + ], + "image_footnote": [], + "bbox": [ + 98, + 318, + 514, + 459 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a67683c6cd9e489633cf9362ca9a8745fec18ceaadaee4d0016724819c9cef44.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 521, + 319, + 901, + 415 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/36e08331fc60a1a87e70a404552c46e77d1fb66fa7ac796d4fc66d460223305f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 517, + 417, + 544, + 438 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Each of the four images contain one car respectively. However 1 2 feature the same car, and 3 4 feature another identical car. Thus, there are only two cars in the scene.", + "bbox": [ + 547, + 419, + 839, + 448 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b5f684a0dc7fb59074ce74c40d6e4c7c784d80a08f7b552dcd1e0ff195ada2e0.jpg", + "image_caption": [ + "Figure 1. Illustration of continuous space perception: The left side depicts the construction of an image sequence representing a continuous space. The right side presents an example task related to the continuous space, showing how adjacent images are connected. Note that to answer the question correctly, one must recognize that the cars in images 1 and 2 are the same, as well as those in images 3 and 4." + ], + "image_footnote": [], + "bbox": [ + 517, + 450, + 544, + 470 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The four images respectively contains one car. Four cars all together are visible in the scene.", + "bbox": [ + 547, + 450, + 844, + 469 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 529, + 326, + 544 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Vision-Language Models (VLMs) have recently witnessed significant progress in visual comprehension. As the permitting length of image context grows, VLMs can now comprehend a broader range of views and spaces. Current benchmarks provide insightful analysis of VLMs in tasks involving complex visual instructions following, multi-image understanding and spatial reasoning. However, they usually focus on spatially irrelevant images or discrete images captured from varied viewpoints. The compositional characteristic of images captured from a static viewpoint remains underestimated. We term this characteristic as Continuous Space Perception. When observing a scene from a static viewpoint while shifting orientations, it produces a series of spatially continuous images, enabling the reconstruction of the entire space. In this paper, we present CoSpace, a multi-image visual understanding benchmark designed to assess the Continuous Space perception ability for VLMs. CoSpace contains 2,918 images and 1,626 question-answer pairs, covering seven types of tasks. We conduct evaluation across 19 proprietary and open-source VLMs. Results reveal that there exist pitfalls on the continuous space perception ability for most of the evaluated models, including pro", + "bbox": [ + 88, + 568, + 485, + 902 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "prietary ones. Interestingly, we find that the main discrepancy between open-source and proprietary models lies not in accuracy but in the consistency of responses. We believe that enhancing the ability of continuous space perception is essential for VLMs to perform effectively in real-world tasks and encourage further research to advance this capability.", + "bbox": [ + 511, + 530, + 906, + 621 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 630, + 645, + 643 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "With the advancements in deep learning and Large Language Models (LLMs), Vision-Language Models (VLMs) have demonstrated exceptional performance across a range of downstream tasks [4, 17, 22, 30, 47, 60]. Driven by this, VLMs are now being used in diverse applications, such as planning [5, 56], navigation [8, 63] and embodied tasks [28, 42], and are further applied in real-world scenarios [3, 61]. Real-world scenario, featuring the complicated environment, cast significant challenges on VLMs, especially the ability to capture sophisticated spatial information from surrounding spaces.", + "bbox": [ + 509, + 657, + 906, + 823 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite previous attempts to advance spatial understanding in VLMs [4, 38], there remain areas requiring further exploration. Various strategies have been explored to empower models with the ability to identify and ground objects based on spatial relationship [4, 16]. Shiri et al. [47]", + "bbox": [ + 509, + 825, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "29569", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "table", + "img_path": "images/b9e9b35e9445a6335db31052747050423a76d9b664d966251c12788654e790b0.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
BenchmarksEvaluation TargetFeaturesNum.ImgEvaluationInstances
Spatial ReasoningMulti-image PerceptionStatic Viewpoint
MMRel [38]Object-relationship○××15k15k
Spatial-MM [47]Object-relationship and spatial-CoT✓××2.3k400
SpatialEval [50]Spatial Reasoning✓××4.6k4.6k
GSR-Bench [41]Grounded spatial reasoning✓××4.9k4.9k
MileBench [9]Multimodal Long Context×✓×97.9k6.4k
Demon-Core [26]Demonstrative instruction following×✓×62.8k18.2k
Mementos-Val [53]Temporal sequential-image understanding×✓×-699
MIBench [33]Fine-grained multi-image understanding×✓×125k13k
BLINK [14]General visual perception○○×7.3k3.8k
SEED-Bench [23]Visual comprehension○○×1.9k24k
Video-MME [13]Visual comprehension○✓×0.9k2.7k
MMIU [37]Visual comprehension○✓×77.7k11.7k
SparklesEval [15]Multi-image Chat○✓×550150
CoSpace (Ours)continuous space perception✓✓✓2,9181,626
", + "bbox": [ + 187, + 88, + 807, + 287 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Table 1. Comparison with selected benchmarks that address multi-image evaluation and/or spatial evaluation. $\\checkmark$ denotes the main focus of the corresponding feature, $\\circ$ indicates that the feature exists but is not the primary focus, while $X$ implies a lack of the feature.", + "bbox": [ + 89, + 290, + 906, + 318 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "and Wang et al. [50] assess spatial understanding ability for VLMs with the recognition of basic tasks in this domain. However, much of these work center on single-image inputs, without further extending to multi-image scenarios. Works on video understanding [52, 60] present the concept of spatial temporal understanding, but the spatial information of videos involve sequential frames that inherently linking spatial understanding with additional temporal dynamics. Given these attempts in spatial understanding, we pose the question: Are the current efforts adequate for advancing the research of spatial understanding?", + "bbox": [ + 88, + 327, + 480, + 492 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We claim that a gap still exists between the concept proposed in previous work and more advanced spatial understanding. For humans, exposing to an unfamiliar environment requires determining the position, identifying surrounding objects, and making decisions based on spatial context. Such process typically involves integrating various information about the surrounding space. We refer to it as continuous space perception, which is an essential ability in real-world towards effectively completion of daily tasks.", + "bbox": [ + 88, + 493, + 482, + 628 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Figure 1 shows an example of continuous space perception: Conceptually, humans perceive continuous visual space as a series of \"images\", each representing a segment of the space with overlaps between adjacent views. Mimicking this perception, Figure 1 illustrates the continuous visual space using four images. To answer questions accurately about these images requires a comprehensive understanding of all four images and reconstructing the entire space they represent. For example, to accurately count the number of cars in the space, it is essential to recognize that the cars appearing in the first and second images are the same, as are the cars in the third and fourth images.", + "bbox": [ + 88, + 628, + 482, + 809 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The continuous space perception ability is equally important for VLMs as for human. Many real-world scenario tasks, such as planning and navigation, require models to effectively integrate the information in the surroundings to take the optimal actions. During this process, the reconstruction of the space heavily relies on continuous space", + "bbox": [ + 88, + 810, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "perception. However, as shown in Table 1, this ability is largely overlooked in existing benchmarks for VLMs.", + "bbox": [ + 511, + 327, + 903, + 357 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To this end, we present CoSpace, a visual understanding benchmark designed to evaluate the Continuous Space perception ability for VLMs. CoSpace includes four categories and seven tasks, emphasizing the following core abilities: (1) Can VLMs recognize direction and position when presented with spatially continuous images captured from a static viewpoint? (2) Can models identify the correspondence between adjacent images in the same scene? (3) Can models make optimal navigation decisions based on the perception of the space? We posit the importance of these three core abilities, with the first one forms the basic understanding of space, the second one is essential for integrating the information across different images and the last one surpasses mere literal understanding and makes a step towards more practical applications.", + "bbox": [ + 509, + 357, + 906, + 584 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We evaluated 19 widely used open-source and proprietary VLMs on our proposed CoSpace benchmark. Results show that though some of the open-source models achieve approaching or even higher accuracy compared with state-of-the-art proprietary models on certain tasks, proprietary models exhibit remarkably higher consistency in their responses across multi-dimensional evaluation metrics. Additionally, we found that with the alteration of input context and image sequence, the tasks that humans find more challenging do not always pose greater difficulty for the models.", + "bbox": [ + 509, + 585, + 905, + 734 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 511, + 761, + 653, + 777 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Multi-image Perception Ability of VLMs", + "text_level": 1, + "bbox": [ + 511, + 787, + 861, + 803 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Large Vision-Language Models (VLMs), such as LLaVA models [31, 32] and Sphinx [30], have shown impressive performances on visual reasoning and grounding tasks that focus on only one image at a time. However, it is more natural to allow multiple images and arrange them in interleave format in one input. Along with the growing capability of large vision models, there are also many other applications. For example, the use of large vision models has been demonstrated for large datasets, such as the US Census dataset [33].", + "bbox": [ + 509, + 809, + 905, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "29570", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ities of single-image models, multi-image perception and understanding abilities are gaining increasing attentions, including research topics such as multimodal in-context learning [10, 22, 62], interleaved visual instruction following [18, 25, 26], and temporal visual understanding [49], and practical applications such as navigation and embodied QA [63]. To achieve multi-image perception, some methods focus on constructing interleaved visual instruction tuning data [17, 18, 22, 62], some improve the model structure and training strategies for better cross-image understanding [26, 52, 55, 58], and others address the image processing strategy for adaptable visual feature encoding concerning multi-image scenarios [21, 32, 59].", + "bbox": [ + 89, + 90, + 480, + 287 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "With the enhancements of the multi-image perception ability, VLMs are able to obtain visual information covering a wider range of views from multiple images at one time, resulting in more profound observations and understandings of the surrounding environment. In our paper, we take this as basis, and dedicate to investigating the spatial perception ability of VLMs.", + "bbox": [ + 89, + 301, + 480, + 407 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Spatial Comprehension Evaluation of VLMs", + "text_level": 1, + "bbox": [ + 89, + 455, + 467, + 472 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For the past years, the evaluation of spatial comprehension for VLMs has always been a commonly engaged topic, and it often refers to the spatial relation recognition among objects, such as MMRel, GSR-Bench and SpatialMM shown in Table 1. Recently, multi-image related evaluation for VLMs are widely explored, including interleaved visual instruction following [26, 33], temporal sequential-image comprehension [53], and long context understanding [9]. In multi-image scenarios, such as video comprehension [13, 23] and multi-image dialogue [15], the spatial reasoning ability is also one of the evaluated aspects. However, these benchmarks primarily focus on general multi-image comprehension where most of input images come from difference scenes. The compositional features across multiple images to integrate richer spatial information of a wider scene and more profound spatial understanding are seldom emphasized. Although Blink [14], ActiView [54], and video-based benchmarks [13, 27] require spatial understanding across multiple images, they either focus on different views of the same object, or discrete views from dynamic viewpoint. These views are not continuous regarding a static viewpoint, and the resulting images are not necessarily bounded regarding the same scene. In contrast to existing works, we address the continuous spatial perception from a static viewpoint, that requires models to connect all the perceived views together to form an intact understanding of the whole scene.", + "bbox": [ + 89, + 493, + 482, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. CoSpace", + "text_level": 1, + "bbox": [ + 513, + 90, + 612, + 107 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Continuous Space Perception Ability", + "text_level": 1, + "bbox": [ + 511, + 114, + 831, + 132 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this paper, we define the term, continuous visual space, as the perceived views from a static viewpoint towards different orientations that can naturally compose a complete scene or environment. As is shown in Figure 1, if one would like to correctly answer the question, it is needed to first capture the visual information of all the cars, and consequently, the ability of continuous space perception refers to capturing and processing information from these images of continuous visual spaces.", + "bbox": [ + 511, + 138, + 903, + 273 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To obtain an overall observation of environment, human perceive the whole surroundings from the same viewpoint towards different directions, and the percept continuous visual spaces can be reconstructed as a panorama in the brain. Different from human that directly views the space continuously, models typically take discrete images as input. Therefore, the way to present continuous visual spaces remains flexible. In our CoSpace, we present a complete continuous visual space as follows: we provide several images shot at the same viewpoint in the same scene, but towards different directions, e.g. respectively towards front, right, back and left. These images together form a complete space without omitting any visual information of the surrounding environment. Naturally, such strategy could produce overlapped spaces between adjacent images, which tightly connects images describing adjacent spaces, and can consequently help models better identify the relationships between images from continuous visual space. In this way, we posit that such strategy can effectively preserve continuous spatial information of an extensive space as well as be suitable to understanding for VLMs.", + "bbox": [ + 511, + 275, + 903, + 592 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Task Design", + "text_level": 1, + "bbox": [ + 513, + 604, + 643, + 619 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We comprehensively evaluate the continuous space perception ability of VLMs regarding four aspects. Two of them are fundamental ability of spatial perception, detailed as direction perception and recognition (Direction), and rotation-centric perception (Rotation). The other two are further practical tasks, detailed as counting-based perception (Counting), and embodied-based planning (Planning). These four categories are subdivided into seven tasks for elaborate investigation as shown in Figure 2.", + "bbox": [ + 511, + 627, + 903, + 762 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We employ two multiple choice tasks, both with four options, for the Direction category as follows:", + "bbox": [ + 511, + 763, + 903, + 792 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "- Direction Recognition (DIR-Rec): In real world, identifying direction is inevitable when one is placed in a new environment. Similarly, the DIR-Rec task requires models to recognize the direction of targets by answering questions such as \"where is the building located?\" Meanwhile, we provide detailed task instructions, including direction references such as \"the first image is facing north,", + "bbox": [ + 513, + 795, + 903, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "29571", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Direction", + "text_level": 1, + "bbox": [ + 401, + 93, + 452, + 103 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$\\clubsuit$ Direction Recognition (DIR-Rec)", + "bbox": [ + 163, + 108, + 344, + 119 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/114f03be34d4468d3a10b547709cd68894dcdf6af81a77d3e849293e92ab478e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 94, + 119, + 418, + 160 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Question: Which direction is the alleyway going in the images?", + "bbox": [ + 94, + 160, + 385, + 170 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Options: A. South to North", + "bbox": [ + 94, + 170, + 230, + 179 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "C. Northeast to Southwest", + "bbox": [ + 145, + 180, + 266, + 188 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "B. East to West", + "bbox": [ + 277, + 170, + 351, + 180 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "D. Northwest to Southeast", + "bbox": [ + 282, + 180, + 400, + 188 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Rationale: The $2^{nd}$ and $4^{th}$ images indicate a direction of East-West, the $1^{st}$ and $3^{rd}$ images indicate the width of the alleyway.", + "bbox": [ + 96, + 190, + 405, + 212 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Directional Object Perception (DIR-Obj)", + "bbox": [ + 486, + 108, + 710, + 119 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/22af0283074d7c8bc2bc1b8252d1b9d164edb5bb734d732fdf9aa9071376d10f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 434, + 119, + 756, + 160 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Question: What is visible in the southeast direction?", + "bbox": [ + 439, + 160, + 679, + 170 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Options: A. A pedestrian bridge", + "bbox": [ + 439, + 170, + 594, + 179 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "C. A \"no parking\" sign", + "bbox": [ + 491, + 180, + 596, + 189 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Rationale: Southeast refers to the left part of the $3^{rd}$ image and the right part of the $4^{th}$ image, only a pedestrian bridge presents.", + "bbox": [ + 434, + 190, + 753, + 212 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Rotation", + "text_level": 1, + "bbox": [ + 403, + 220, + 450, + 231 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Rotation-Angle (ROT-Ang)", + "bbox": [ + 178, + 234, + 330, + 246 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/1554f2b582b6c91996922e0d27a13e6890690504b76f6731e929484cc98020fc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 94, + 246, + 419, + 287 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Question: What is the angle between the adjacent image?", + "bbox": [ + 96, + 291, + 359, + 301 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Options: A 30 B.60", + "bbox": [ + 96, + 301, + 251, + 314 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Rationale: Observing from the overlapping areas of the tall building, and the displacement of the short buildings. 30 degree is more feasible.", + "bbox": [ + 96, + 315, + 413, + 335 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Rotation-Difference (ROT-Dif)", + "bbox": [ + 514, + 234, + 684, + 244 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/d76124066281e59112b3b9b77dc8421710c641d62c0b920bff0db435f17796f1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 437, + 244, + 756, + 303 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Question: Which image does not belong to this image sequence?", + "bbox": [ + 439, + 304, + 736, + 314 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Rationale: Observing from the range of the overlaps regarding the road., image 4 present the most rotation degree.", + "bbox": [ + 434, + 315, + 720, + 337 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Counting", + "text_level": 1, + "bbox": [ + 403, + 344, + 450, + 354 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "- Counting (CNT)", + "bbox": [ + 380, + 358, + 470, + 369 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/7e2c0df7eaaf10792f257ec4dfb6d65b5fcd55f1a7c57ebd699ce0e31dc10731.jpg", + "image_caption": [ + "Figure 2. Examples of CoSpace, along with prompt templates for evaluation. We also provide extra visual guidance and textual rationales for readers (invisible to models during evaluation) for easier understanding." + ], + "image_footnote": [], + "bbox": [ + 96, + 371, + 189, + 410 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/9f29c1ad98575cb90fc026b7a4b90c2ed3e8c633420162a9c15cbc9af4fd3a67.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 191, + 375, + 305, + 411 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/9b6656572a95637c9279c6210c83597fc61f408ad64447af2c03086f967d072a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 307, + 375, + 419, + 419 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/5d8997331dfde03496b702d1ffa3509f97b3ba5c6be9677062be56ec6035ff9c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 421, + 375, + 570, + 419 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Rationale:", + "bbox": [ + 580, + 368, + 643, + 381 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "One red truck spotted in the $3^{rd}$ image, and it is spotted again in the $4^{th}$ image.", + "bbox": [ + 578, + 383, + 746, + 402 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "There is indeed 1 red truck in the scene.", + "bbox": [ + 580, + 404, + 750, + 412 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Question: How many red trucks are visible in the scene?", + "bbox": [ + 96, + 421, + 356, + 431 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Answer: 1", + "bbox": [ + 460, + 422, + 511, + 431 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Note that $\\mathbb{O}$ denotes repeated cars", + "bbox": [ + 614, + 421, + 756, + 431 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Planning", + "text_level": 1, + "bbox": [ + 403, + 439, + 449, + 450 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/689c430bf92b78135be9e9764e13e542626ef1dde256d5e0e0831d2fa33799b3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 94, + 452, + 419, + 502 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Question: Where is the television regarding your position?", + "bbox": [ + 98, + 503, + 369, + 513 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Options: A. Front-left", + "bbox": [ + 99, + 513, + 210, + 523 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "C. Front-right", + "bbox": [ + 151, + 525, + 217, + 534 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "B. Back-left", + "bbox": [ + 277, + 513, + 334, + 523 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "D. Back-right", + "bbox": [ + 277, + 523, + 343, + 535 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Rationale: TV is placed in the right part of $2^{nd}$ image. Right part of the right view indicate back-right direction.", + "bbox": [ + 96, + 535, + 403, + 555 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/7af29df7f5feca7d9de3d7bb07ca96bbb2cd6c49585cbe7ac18293aa3e50afff.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 434, + 454, + 756, + 494 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "- Planning-Decision (PLA-Dec)", + "bbox": [ + 517, + 494, + 673, + 503 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Question: What should you do to reach the fireplace?", + "bbox": [ + 439, + 503, + 684, + 513 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Options: A. Turn right and go ahead", + "bbox": [ + 439, + 513, + 614, + 523 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "C. Turn left and go ahead", + "bbox": [ + 490, + 525, + 607, + 534 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "B. Turn back and go ahead", + "bbox": [ + 625, + 513, + 751, + 523 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "D. Go ahead and turn back", + "bbox": [ + 627, + 525, + 750, + 534 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Rationale:Fireplace locates in the far back, requiring both a turn action and a moving action, need to first turn backwards and then go ahead", + "bbox": [ + 434, + 535, + 743, + 555 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/808e7415e17194c8358cfca03e7e3d73fae6482126560c3aa864213c6cac7e9f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 764, + 93, + 782, + 108 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Task instructions", + "text_level": 1, + "bbox": [ + 795, + 97, + 887, + 107 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "General prompt template:", + "bbox": [ + 764, + 111, + 893, + 119 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "", + "bbox": [ + 764, + 121, + 838, + 128 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "", + "bbox": [ + 764, + 130, + 890, + 138 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "", + "bbox": [ + 764, + 140, + 852, + 148 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": " ", + "bbox": [ + 764, + 148, + 856, + 157 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "You should generate", + "bbox": [ + 764, + 157, + 857, + 167 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "answer from A, B, C or D.", + "bbox": [ + 764, + 169, + 890, + 176 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Your answer:", + "bbox": [ + 766, + 176, + 826, + 184 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "", + "bbox": [ + 764, + 191, + 885, + 200 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "for Direction and Counting", + "bbox": [ + 764, + 202, + 880, + 210 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "categories:", + "bbox": [ + 764, + 210, + 810, + 220 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the first image is facing north,", + "bbox": [ + 764, + 220, + 888, + 229 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the second facing east, the third", + "bbox": [ + 764, + 229, + 895, + 238 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "facing south and the fourth", + "bbox": [ + 764, + 239, + 875, + 247 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "facing west.", + "bbox": [ + 764, + 248, + 815, + 256 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "", + "bbox": [ + 764, + 266, + 885, + 275 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "for Rotation category:", + "bbox": [ + 764, + 276, + 857, + 282 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "images are arranged clockwise .", + "bbox": [ + 764, + 284, + 895, + 292 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "for example, after tak", + "bbox": [ + 764, + 294, + 852, + 303 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "first image, the photographer", + "bbox": [ + 764, + 304, + 887, + 311 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "turns a certain degree clockwise", + "bbox": [ + 764, + 313, + 895, + 321 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "for capturin", + "bbox": [ + 764, + 321, + 815, + 330 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "", + "bbox": [ + 764, + 340, + 885, + 349 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "for Planning category:", + "bbox": [ + 764, + 351, + 857, + 358 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "you are facing the direction in the first image, with the second image showing the scene to your right, the third image showing the scene at your back, and the fourth image showing the scene to your left.", + "bbox": [ + 764, + 359, + 893, + 412 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/983e9298d2aa988e14b08eb46fb0686ef916722a18575f904eaa277612d43c8d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 764, + 422, + 782, + 436 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Notes for readers", + "text_level": 1, + "bbox": [ + 790, + 426, + 890, + 436 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "- The directional symbols:", + "bbox": [ + 764, + 443, + 885, + 452 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/d409c9bb24490783d6a278550e46775252122cb6e2e2850dc17a1ef23b5001f0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 774, + 454, + 879, + 474 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "are invisible to models. They", + "bbox": [ + 761, + 476, + 895, + 484 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "are used for easier under", + "bbox": [ + 761, + 486, + 883, + 494 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "standing of image orders.", + "bbox": [ + 761, + 496, + 880, + 503 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "- The Rationale is also", + "bbox": [ + 761, + 512, + 866, + 521 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "invisible to models, as they", + "bbox": [ + 761, + 522, + 888, + 531 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "intend for explaining the", + "bbox": [ + 761, + 532, + 879, + 541 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "results for readers.", + "bbox": [ + 761, + 541, + 857, + 549 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the second image is facing east\". We standardize 8 directional options, including four cardinal directions, North, East, South and West, and four intercardinal directions, Northeast, Southeast, Southwest, and Northwest.", + "bbox": [ + 102, + 604, + 482, + 662 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "- Directional Object Perception (DIR-Obj): Some real-world tasks demand abilities far beyond merely identifying the direction of a specified target. It is also a critical ability to distinguish content regarding different directions and to align objects with correct directions. Given a specified direction, the DIR-Obj task requires models to be aware of the appearing objects, and also distinguish those absent from that direction. To achieve this, models need to first reconstruct the entire space from continuous views, and then correctly identify the objects dedicate to a given direction. Specifically, in this task, questions typically focus on the intercardinal directions such as \"What is visible in the southeast direction?\", which necessitates a more comprehensive understanding of the entire space.", + "bbox": [ + 89, + 664, + 482, + 876 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Observing the surroundings from a static viewpoint natu", + "bbox": [ + 109, + 885, + 482, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "rally involves rotating the perspectives. Therefore, we summaries Rotation as another fundamental category of continuous space perception, with two tasks designed as follows:", + "bbox": [ + 511, + 604, + 903, + 650 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "- Rotation-Angle (ROT-Ang): For an observer, when rotating the perspective for more spatial information, there could be overlaps between adjacent observations. Specifically, human can tell the approximate turning angle between consecutive images observed from a static viewpoint by noticing identical furnishings appearing across images. In this task, we investigate if models exhibit similar ability as human does, which requires fine-grained continuous spatial understanding. We apply unified question \"What is the turning angle between the adjacent image?\" and provide two candidate options for model. This task requires models to carefully exam the overlapping and different areas between adjacent images and the shared features in spatially continuous image sequence.", + "bbox": [ + 513, + 659, + 905, + 869 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "- Rotation-Difference (ROT-Dif): Following the ROT-Ang task, we also investigate the model ability to iden", + "bbox": [ + 513, + 869, + 903, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "29572", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "tify distinguished rotation angles. For the ROT-Dif task, models are given a sequence of five images, four of which share the same turning angle and the other is exceptional. Compared to ROT-Ang, this task emphasizes more on the global understanding of images regarding continuous visual space. The unified question of this task is \"Which image does not belong to this image sequence?\" It is an open-ended question without options, where models are required to answer with index of the exceptional image.", + "bbox": [ + 102, + 90, + 482, + 226 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "On the basis of the above two fundamental categories, we further challenge models with two practical categories. The Counting category addresses the occurrence of objects in the entire scene. The Planning category comprises two spatial-related tasks widely evaluated in embodied AI tasks. For the Counting category, the task is described as:", + "bbox": [ + 89, + 227, + 482, + 318 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "- Counting (CNT): Counting is a widely adopted task in existing visual benchmarks that asks models to recognize objects and deduce the time of occurrence of target objects. Generally, existing counting tasks only require models to deal with a single image or multiple spatially discrete images, while in our benchmark, images from continuous visual space where identical objects can occur in multiple images are focused on. This raise a challenge for models to not only recognize targets and count for their occurrences, but also be aware of the existence of the same object across different images. To achieve this, models should locate the overlapping area of adjacent images and align the same object appearing different images. The CNT task is an open-ended task, where models should respond with the total count of the target.", + "bbox": [ + 89, + 321, + 482, + 547 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The ability of continuous space perception is also significant in embodied scenarios. Intuitively, regarding the location of an embodied agent as the static viewpoint, it should have a profound understanding of the overall surroundings to effectively solve questions, and to efficiently plan for the following action sequence. Embodied question-answering (EQA) [36, 45] and navigation [19, 20] are typical tasks that require understand continuous visual space. Following these, we split the Planning category into two tasks addressing the relative locations of targets regarding the VLM-based agent as follows:", + "bbox": [ + 89, + 551, + 482, + 717 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Planning-Question Answering (PLA-QA): Following the implementation of EQA, we develop the PLA-QA task, requiring models to identify the location of a certain object given a continuous embodied space. In this task, instructions like \"Where is the television regarding your position?\" are provided for the models, and we formulate four options, containing candidate directions relative to the agent, for each question.", + "- Planning-Decision (PLA-Dec): This task further investigates the understanding of continuous visual space by asking models to select the proper route to reach the target object. For disambiguation, we standardize the action" + ], + "bbox": [ + 91, + 719, + 482, + 901 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/dc56adf0e9888a06d753626b2fd0a10512cdb11acb5eaa2037d09011a3b445f1.jpg", + "image_caption": [ + "Figure 3. Distribution of categories and tasks in our CoSpace." + ], + "image_footnote": [], + "bbox": [ + 568, + 90, + 866, + 258 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "space as turning (turning to other directions without displacement) and go ahead. The PLA-Dec task especially focus on the order of actions. As is shown in the case in Figure 2, the options \"B. Turn back and go ahead\" and \"D. Go ahead and turn back\" represent two totally different actions and end up in different positions.", + "bbox": [ + 524, + 294, + 903, + 383 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Data Collection and Statistics", + "text_level": 1, + "bbox": [ + 513, + 402, + 774, + 416 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The construction of CoSpace follows the collect-filter-annotate pipeline. We first adopt automatic collection approaches to gather images, and then manually select images that satisfy detailed requirements of our designed tasks and generate questions for these images. Finally human annotators provide answers to each questions. The process is detailed in the following of this section.", + "bbox": [ + 511, + 426, + 903, + 532 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Firstly, to automatically collect sets of images captured in the same scene at static viewpoint but towards different directions, we employ two primary sources, Baidu Map Panorama API that allows us to acquire data towards any desired directions in the real world scene, and HM3D [43], a dataset of 1,000 building-scale 3D reconstructions for embodied tasks. For HM3D, by using Habitat-Sim [40, 46, 48], we instruct an agent wandering in the habitats, we are able to collect images of any views through the attached camera. Secondly, we manually filtered the auto-collected images according to our task requirements, ensuring that the selected images all contain sufficient visual information. With the filtered images, we employed GPT-4o to generate questions via few-shot prompting for the required tasks. Finally, human annotators, an average of two annotators per question, provided answers to these generated questions. This process allowed us to discard vague or unanswerable questions, ensuring that only high-quality question-answer pairs were included in our proposed CoSpace.", + "bbox": [ + 511, + 535, + 905, + 823 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our CoSpace comprises a total of 1,626 question-answer pairs and 2,918 images. Among the images, 2,302 are accessed from Baidu Map API and 616 are derived from HM3D dataset. The distribution of instances over categories and tasks is shown in Figure 3.", + "bbox": [ + 511, + 825, + 903, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "29573", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/343cc711b0a759b7eb2c52e37b67cf118a900b09043d24d005b54c29bef49d24.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelsDIR-RecDIR-ObjCNTROT-AngROT-DifPLA-QAPLA-DecAverage
ACCqACCpACCqACCpACCqACCpACCACCACCACC
Proprietary Models
Claude-3.7-sonnet [2]44.4029.2045.6035.6045.0038.0064.3393.5054.7369.3451.97
Gemini-1.5-pro [44]37.6015.6040.6031.6038.2524.0059.6782.0051.6462.9144.39
GPT-4o [39]40.4022.8046.0032.0040.0023.5058.3350.5053.0554.4642.10
>70B Open-source Models
InternVL2_5-78B [6]32.2024.4054.4047.6051.2543.0050.0077.0032.3942.7245.50
Qwen2-VL-72B [51]31.0023.2053.6044.8044.7537.0050.0062.0042.7259.1544.82
InternVL2-76B [7]33.6013.2046.4039.2050.0043.5050.0023.0025.3530.0535.43
LLaVA-OneVision-72B [24]19.2010.8044.2034.4024.2519.0050.0026.0032.8630.9929.17
<13B Open-source Models
MiniCPM-V 2.6 [57]32.8021.2040.4031.6038.5031.5050.0056.0041.3127.7037.10
Qwen2-VL-7B [51]26.4016.4039.2031.6045.0036.0050.0051.5034.2726.7635.71
Mantis-8B [18]30.6024.4031.2028.0041.7536.5050.0039.5033.3327.7034.30
InternVL2-8B [7]29.0012.8037.2030.4038.5031.0050.3347.5032.3927.7033.68
VILA1.5-8B [29]34.0030.4028.2024.0042.2538.0050.0018.5022.0742.7233.01
Idefics3-8B [21]34.0021.2038.6028.0032.7525.0048.0028.0026.7625.3530.77
LLaVA-OneVision-7B [24]18.8012.0037.6033.6037.7533.0050.0021.5029.5826.2930.01
Phi-3.5-vision [1]20.4012.0033.2028.4036.7534.0050.0022.5034.7426.7629.88
Brote-IM-XXL [55]33.0013.2031.0024.8030.7529.5050.0018.0026.7620.1927.72
LongVA-7B [60]23.0016.0031.6028.0028.0022.0048.6721.5019.2530.9926.90
Mono-InternVL-2B [35]30.0027.6032.0027.6014.008.5050.0018.0025.3524.8825.79
mPLUG-Owl3-7B [58]22.806.8029.2015.6022.259.0046.3316.5027.7021.1321.73
", + "bbox": [ + 151, + 88, + 844, + 358 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Main results on our benchmark. The proprietary models are accessed via APIs, and the open-source models are accessed from Huggingface checkpoints.", + "bbox": [ + 89, + 359, + 906, + 390 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 404, + 225, + 422 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As discussed above, CoSpace features reconstructing the surrounding space via multiple images, with each image representing a different space segments. Different from previous multi-image benchmark involving spatial reasoning tasks, the input order of images is crucial in our tasks. To enable that models can associate each image with a specific direction to answer questions in our benchmark, we provide direction references along with images following the prompt templates in Figure 2 through out the evaluation.", + "bbox": [ + 89, + 430, + 483, + 566 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Evaluation Metrics", + "text_level": 1, + "bbox": [ + 89, + 575, + 274, + 590 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For the Direction and the Counting categories, for fair evaluation, we adopt both pair-wise accuracy $ACC_{p}$ and query-wise accuracy $ACC_{q}$ as introduced by Fu et al. [12] and Luo et al. [34] instead of simple accuracy. For each question, we task models twice with different prompts. $ACC_{p}$ only counts when models correctly response to both prompts, while $ACC_{q}$ is the averaged score of the two responses. For these different prompts regarding the same question, we vary only the order of images by preserving the clockwise order but changing the beginning view, so that the answer to the question remains unchanged. For example, if original image sequence begins with the northern image, the modified sequence could start from the east, south or west.", + "bbox": [ + 89, + 598, + 482, + 794 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For the Rotation category, some of the instances does not cover the entire space, and the such a modification of input image sequence will inevitable disrupt the comprehension of the continuous space. Similarly, for the Planning category, considering that the original input sequence features the front view as the first image and this manner inherently matches the thinking pattern when models take", + "bbox": [ + 89, + 795, + 483, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "actions. Therefore, for these two categories, we apply only a single context for each instances during evaluation, and report accuracy, $ACC$ , as final scores.", + "bbox": [ + 511, + 406, + 906, + 450 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Models", + "text_level": 1, + "bbox": [ + 511, + 463, + 607, + 478 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We accessed three proprietary models, including GPT-4o [39], Gemini-1.5-pro [44] and Claude-3.7-sonnet [2], along with 16 widely used open-source VLMs accepting multiple images simultaneously. Among the open-source models, four have more than 70B parameters, while the others range in size from 2B to 13B. We also enable the evaluation for powerful single-image models, such as LLaVA-1.6 [31] and LLaMA-3.2 [11], for further investigation. For fair comparison, we design a single-image pipeline for some our designed tasks and conduct evaluations on both single-image and multi-image models in Section 5.3.", + "bbox": [ + 511, + 487, + 906, + 654 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Results", + "text_level": 1, + "bbox": [ + 511, + 666, + 607, + 681 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The overall results on our benchmark is reported in Table 2. We derive three main findings from this table: 1. The performance of proprietary models and $>70B$ open-source models largely varies. As expected, most proprietary models and $>70B$ open-source models marginally outperform the evaluated open-source models. However, there are significant dependencies among these models, where Claude-3.5-sonnet achieves an average score of $49.37\\%$ , InternVL2_5-78B achieves $45.50\\%$ , while LLaVA-OneVision-72B only presents an average score of $29.17\\%$ , even lower than more than half of the $<13B$ open-source models. 2. Proprietary models do not equally perform the best across all the tasks. Proprietary models fail to beat open-source models on tasks including DIR-Obj and CNT,", + "bbox": [ + 511, + 689, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "29574", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/e05a594ce3ebddec012d9ba10d5cdf0dcd54e6abf2990ef64766cbc2d9042de5.jpg", + "table_caption": [], + "table_footnote": [ + "Table 3. Results for disrupted image orders. $ACC_{q}$ is reported for Direction and Counting tasks and $ACC$ is reported for the Planning task. * indicates the performance is higher than that in Table 2." + ], + "table_body": "
ModelsDIR-RecDIR-ObjCNTPLA-QAPLA-Dec
Proprietary Models
Claude-3.7-sonnet36.3631.1737.2044.4044.50
Gemini-1.5-pro37.2037.2027.5038.0336.15
GPT-4o37.2041.2039.0045.0739.91
>70B Open-source Models
InternVL2_5-78B29.6052.8044.0034.27*36.15
Qwen2-VL-72B30.4047.6046.00*37.0933.33
InternVL2-76B25.2038.8052.00*25.3525.35
LLaVA-OneVision-72B20.40*43.6022.5036.62*32.39*
<13B Open-source Models
MiniCPM-V 2.622.8040.0037.0039.4426.76
Qwen2-VL-7B19.6035.2040.5034.2723.00
Mantis-8B34.80*30.8042.50*31.9229.11*
InternVL2-8B31.20*36.4042.50*22.5426.76
VILA1.5-8B33.6029.20*42.0019.7239.91
Idefics3-8B37.60*37.2032.5023.9424.88
LLaVA-OneVision-7B17.2038.00*35.0030.05*27.70*
Phi-3.5-vision26.00*34.40*36.5031.4624.41
Brote-IM-XXL34.40*29.6029.5029.58*17.84
LongVA-7B16.0031.6028.50*17.8428.64
Mono-InternVL-2B18.8033.60*15.50*25.3522.07
mPLUG-Owl3-7B26.40*29.2029.50*29.11*32.39*
", + "bbox": [ + 91, + 88, + 486, + 345 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $>70\\mathrm{B}$ open-source models exhibit stronger performance. 3. The ROT-Ang task is quite challenging for open-source models. It is noticeable that for the ROT-Ang task, all open-source models achieve similar results near $50\\%$ , which are almost equal to random selection. Such result is caused by the fixed predictions over all different context and images input. We will elaborately investigate such phenomena in Section 5.4.", + "bbox": [ + 88, + 407, + 482, + 526 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We conduct manual evaluation on all instances to assess the reasonability and difficulty of CoSpace and results are provided in Appendix ??. Results show that human performance excels the best models across all tasks. Human annotators achieved an average $ACC_{q}$ of $86.25\\%$ , which is significantly higher than that of Claude-3.7-sonnet (59.56%).", + "bbox": [ + 89, + 529, + 483, + 621 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5. Analysis", + "text_level": 1, + "bbox": [ + 89, + 637, + 187, + 654 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.1. Disrupted Image Order", + "text_level": 1, + "bbox": [ + 89, + 662, + 308, + 680 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Considering the spatial relationship between input images of our CoSpace, the order of input images can make significant difference on the understanding of the space, especially when the order being thoroughly disrupted. Based on this speculation, we randomly shuffled the input sequence, and provided the directional orientation only of the first image for the model, with directions of the rest images to be inferred by the model itself. We conducted evaluation under such setting on three tasks: Direction, Counting and Planning, and the results are illustrated in Table 3.", + "bbox": [ + 89, + 686, + 482, + 837 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Surprisingly, the performance of $< 13\\mathrm{B}$ models under disrupted image order was not always lower than the basic setting. In fact, half of the $< 13\\mathrm{B}$ open-source models achieved a higher score in the DIR-Rec task under disrupted set", + "bbox": [ + 89, + 839, + 483, + 900 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8db564116b8fb674fff9f811684485b15026bd1df307597ab2fb592aeadc7458.jpg", + "image_caption": [ + "Figure 4. Variance in $ACC_{q}$ of the average on the Direction and Counting categories for eight selected models. Model abbreviations are used for simplicity. Accuracy in the figure represents the average of $ACC_{q}$ over two categories." + ], + "image_footnote": [], + "bbox": [ + 506, + 89, + 911, + 244 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/37023e20b2dbacaf1783a933c188a52f71c750cc31504d1653620e3a01ff812f.jpg", + "table_caption": [], + "table_footnote": [ + "Table 4. Results of conditional accuracy on the Planning category for eight selected models. Model names are abbreviated. Open-source models all refer to $<13\\mathrm{B}$ ones." + ], + "table_body": "
Img OrderClaude GeminiQwen2InternVL2OneVisionPhi-3.5LongVA
Sequential79.5873.6469.9134.2534.7833.3321.6236.59
Disrupted60.7156.7956.2524.6639.5835.9426.8731.58
", + "bbox": [ + 519, + 310, + 897, + 356 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ting. Additionally, for other tasks, an average of three to five open-source models performed slightly better with disrupted image order. Nevertheless, similar phenomenon was not observed for proprietary models and most of the $>70\\mathrm{B}$ open-source models, except LLaVA-OneVision-72B. Also, the evaluated proprietary models similarly underperformed across all tasks under disrupted setting, which aligns more closely with our expectations. For $>70\\mathrm{B}$ open-source models, three of four outperformed on only one task against basic setting, with LLaVA-OneVision-72B producing a completely opposite result. Considering the such phenomenon, we conclude that the results for $<13\\mathrm{B}$ open-source models are partly due to the instability of predictions. Moreover, we acknowledge that the impact of disrupting the image order may not be as significant for models as it is for human, exhibiting the discrepancy between models and human when perceiving continuous visual space via multiple images.", + "bbox": [ + 511, + 415, + 906, + 674 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Performance Consistency", + "text_level": 1, + "bbox": [ + 511, + 681, + 746, + 698 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We mentioned in Section 4.1 that for the Direction and Counting categories, both $ACC_{q}$ and $ACC_{p}$ are adopted as evaluation metrics. We present the gap between $ACC_{q}$ and $ACC_{p}$ , i.e. variance in accuracy, in Figure 4. A smaller variance indicates that a model responds more consistently to input changes, reflecting better performance consistency. As shown, all models exhibit relatively large variance, with proprietary models (Claude-3.7-sonnet, Gemini-1.5-pro, GPT-4o) displaying even higher variances than open-source models, suggesting room for improvement. Such result highlights the importance of introducing $ACC_{p}$ .", + "bbox": [ + 511, + 703, + 906, + 869 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To further analyze the performance consistency of models, we introduce the metric conditional accuracy in the", + "bbox": [ + 511, + 869, + 905, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "29575", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/e3d96dbb1ce1c5ecd33edcb1e1979dd2cae30225eb577c363a85e93dc0354e70.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelsDIR-RecDIR-ObjCNTPLA-QAPLA-DecAVG
Models with only one image in the input
LLaVA-v1.627.6032.0021.0021.6025.8225.60
LLama-3.2-11B29.2028.8028.5017.8416.9024.25
Models with multi-image perception ability
InternVL2-8B36.40*34.8027.0033.80*31.92*32.78
MiniCPM-V 2.623.2027.2026.0031.9232.39*28.14
Phi-3.5-vision18.4036.00*28.0031.4623.0027.37
LLaVA-OneVision26.80*32.8016.0022.0724.4124.42
", + "bbox": [ + 98, + 89, + 473, + 198 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. Results for single image pipeline. We report $ACC_{q}$ for Direction and Counting tasks and $ACC$ for Planning tasks. * indicates the performance is higher than that in Table 2.", + "bbox": [ + 89, + 200, + 482, + 242 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Planning category. Conditional accuracy measures how often models successfully solve the PLA-Dec question given that they answer the correspondent PLA-QA question correctly, as in PLA-QA task and PLA-Dec task, two respective questions regarding the same set of images typically feature the same object or direction as target. Results in Tabel 4 reveal that open-source models exhibit a greater gap in conditional accuracy against proprietary models than on the Direction and Counting categories. While proprietary models make more consistent decisions based on correct detection, the gap remains non-negligible.", + "bbox": [ + 88, + 258, + 482, + 425 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Above all, we argue that maintaining performance consistency is rather important for the further improvement of models, as it is closely related to robustness and rather essential for more complicated tasks.", + "bbox": [ + 89, + 425, + 483, + 484 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.3. Single Image Pipeline", + "text_level": 1, + "bbox": [ + 89, + 498, + 294, + 513 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We presented in Section 3.1 that the understanding of continuous visual space inherently relies on multiple images as input to represent the whole space. Also, recent work [54] explored to construct a pipeline that can utilize caption and memory to transform the multi-image task into single image format. Similarly, we developed a single image pipeline for the Direction, Counting and Planning categories in our benchmark. For the Rotation category, as the single image input will inevitably compromise the uniform understanding of continuous visual space and the relationship between images can hardly be replaced by mere captions, we consider it not suitable for the single image pipeline. For evaluation, we selected LLaVA-v1.6 [31] and Llama-3.2-Vision [11] as the representatives of single image models. We also selected some multi-image models and conducted the single image evaluation.", + "bbox": [ + 91, + 521, + 482, + 762 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Results of the single image pipeline are reported in Table 5. Overall, the multi-image models underperformed under this setting compared with the multi-image pipeline. However, in the Direction-Rec and Planning-Dec task, two of four models achieve a higher accuracy, which is far beyond our expectations. We conclude that with the effective summarization of captions, models can still have an understanding of visual clues and sometimes the complete of all images can even add to the performance. Also,", + "bbox": [ + 89, + 763, + 482, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "three of the four multi-image models outperform the single-image models on the average accuracy, indicating that models with inherent multi-image capabilities boast an overall better understanding of the visual clues and information extracted from the effectively multi-image scene. Based on this observation, the importance of empowering models with multi-image capabilities needs again to be emphasized.", + "bbox": [ + 511, + 90, + 903, + 196 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.4. Investigation on Angle-related Reasoning", + "text_level": 1, + "bbox": [ + 511, + 210, + 864, + 226 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 2 shows that all the assessed open-source models achieve near-half ACCs on the Rotation-Angle (ROT-Ang) task. After checking the responses of different models, we found out that these models tend to produce the same answer across all the queries, e.g. \"A\" for all the 300 questions in this task. Therefore, due to the even distribution of options, these models achieve the similar near-half ACC.", + "bbox": [ + 511, + 233, + 903, + 339 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To discover the reason behind such phenomenon, we asked models to generate the rationale when answering the questions in the ROT-Ang task and observed some of these models can follow the instruction to generate rationales. For example, InternVL2-8B no longer always selected the same option and consequently correctly answered some question that it originally failed. However, we also noticed that it tended to identify the whole surrounding covered by the images, thus perceiving the sum of turning angles as 360 degrees. Therefore, it frequently output 90 degrees as the final answer, even if 90 degrees not included in the options. In contrast, proprietary models like Claude-3.7-sonnet could firstly calculate the sum of covered angles across all images based on the two give options and then decided which answer better matched the scene in given images.", + "bbox": [ + 511, + 340, + 905, + 566 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Though all the assessed open-source models fail the Rotation-Angle task, we still need to underline the importance of this task and the related capability to correctly identify the rotation in the space. Such capability lays the foundation for more advanced tasks in practical use and has already been boasted by some of the proprietary models.", + "bbox": [ + 511, + 569, + 905, + 659 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 676, + 633, + 693 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Recognizing the gap between previous work and more advanced spatial understanding, we propose the concept of continuous space perception and present a novel benchmark CoSpace that comprises four categories and seven tasks. CoSpace covers both fundamental spatial comprehending abilities and more practical and commonly-used capabilities to comprehensively assess continuous space perception ability for VLMs. We evaluated three widely used proprietary models and sixteen most advanced open-source models and conducted detailed analysis for noticeable results in the experiments. Results reveal a gap between the assessed open-source and proprietary models, especially for performance consistency and in rotation-based perception.", + "bbox": [ + 511, + 704, + 905, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "29576", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 91, + 90, + 250, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work is supported by the National Key R&D Program of China (2022ZD0160502), the National Natural Science Foundation of China (No. 62276152), and funding from Wuxi Research Institute of Applied Technologies, Tsinghua University under Grant 20242001120. We acknowledge Yan Liu, Jidong Chen, Jianyang Liu, Shuangyu Li and Shu Chi for their effort during the process of data annotation and human evaluation.", + "bbox": [ + 89, + 113, + 483, + 226 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 251, + 187, + 266 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Phi-3 technical report: A highly capable language model locally on your phone, 2024. 6", + "[2] Anthropic. The Claude 3 Model Family: Opus, Sonnet, Haiku. 2024. 6", + "[3] Xu Cao, Tong Zhou, Yunsheng Ma, Wenqian Ye, Can Cui, Kun Tang, Zhipeng Cao, Kaizhao Liang, Ziran Wang, James M Rehg, et al. MAPLM: a real-world large-scale vision-language benchmark for map and traffic scene understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21819-21830, 2024. 1", + "[4] Boyuan Chen, Zhuo Xu, Sean Kirmani, Brain Ichter, Dorsa Sadigh, Leonidas Guibas, and Fei Xia. SpatialVLM: endowing vision-language models with spatial reasoning capabilities. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14455-14465, 2024. 1", + "[5] Liang Chen, Yichi Zhang, Shuhuai Ren, Haozhe Zhao, Zefan Cai, Yuchi Wang, Peiyi Wang, Tianyu Liu, and Baobao Chang. Towards end-to-end embodied decision making via multi-modal large language model: Explorations with gpt4- vision and beyond. arXiv preprint arXiv:2310.02071, 2023. 1", + "[6] Zhe Chen, Weiyun Wang, Yue Cao, Yangzhou Liu, Zhangwei Gao, Erfei Cui, Jinguo Zhu, Shenglong Ye, Hao Tian, Zhaoyang Liu, et al. Expanding performance boundaries of open-source multimodal models with model, data, and test-time scaling. arXiv preprint arXiv:2412.05271, 2024. 6", + "[7] Zhe Chen, Weiyun Wang, Hao Tian, Shenglong Ye, Zhangwei Gao, Erfei Cui, Wenwen Tong, Kongzhi Hu, Jiapeng Luo, Zheng Ma, et al. How far are we to GPT-4V? closing the gap to commercial multimodal models with open-source suites. ArXiv preprint, abs/2404.16821, 2024. 6", + "[8] Hao-Tien Lewis Chiang, Zhuo Xu, Zipeng Fu, Mithun George Jacob, Tingnan Zhang, Tsang-Wei Edward Lee, Wenhao Yu, Connor Schenck, David Rendleman, Dhruv Shah, et al. Mobility VLA: multimodal instruction navigation with long-context vlms and topological graphs. arXiv preprint arXiv:2407.07775, 2024. 1", + "[9] Song Dingjie, Shunian Chen, Guiming Hardy Chen, Fei Yu, Xiang Wan, and Benyou Wang. Milebench: Benchmarking MLLMs in long context. In First Conference on Language Modeling, 2024. 2, 3", + "[10] Sivan Doveh, Shaked Perek, M Jehanzeb Mirza, Wei Lin, Amit Alfassy, Assaf Arbelle, Shimon Ullman, and Leonid" + ], + "bbox": [ + 93, + 276, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Karlinsky. Towards multimodal in-context learning for vision & language models. arXiv preprint arXiv:2403.12736, 2024. 3", + "[11] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, and et al. The llama 3 herd of models, 2024. 6, 8", + "[12] Chaoyou Fu, Peixian Chen, Yunhang Shen, Yulei Qin, Mengdan Zhang, Xu Lin, Jinrui Yang, Xiawu Zheng, Ke Li, Xing Sun, et al. Mme: A comprehensive evaluation benchmark for multimodal large language models. arXiv preprint arXiv:2306.13394, 2023. 6", + "[13] Chaoyou Fu, Yuhan Dai, Yondong Luo, Lei Li, Shuhuai Ren, Renrui Zhang, Zihan Wang, Chenyu Zhou, Yunhang Shen, Mengdan Zhang, et al. Video-MME: the first-ever comprehensive evaluation benchmark of multi-modal llms in video analysis. ArXiv preprint, abs/2405.21075, 2024. 2, 3", + "[14] Xingyu Fu, Yushi Hu, Bangzheng Li, Yu Feng, Haoyu Wang, Xudong Lin, Dan Roth, Noah A. Smith, Wei-Chiu Ma, and Ranjay Krishna. Blink: Multimodal large language models can see butA not perceive. In Computer Vision - ECCV 2024, pages 148-166, Cham, 2025. Springer Nature Switzerland. 2, 3", + "[15] Yupan Huang, Zaiqiao Meng, Fangyu Liu, Yixuan Su, Nigel Collier, and Yutong Lu. Sparkles: Unlocking chats across multiple images for multimodal instruction-following models. In ICLR 2024 Workshop on Navigating and Addressing Data Problems for Foundation Models, 2024. 2, 3", + "[16] Baoxiong Jia, Yixin Chen, Huangyue Yu, Yan Wang, Xuesong Niu, Tengyu Liu, Qing Li, and Siyuan Huang. Sceneverse: Scaling 3d vision-language learning for grounded scene understanding. arXiv preprint arXiv:2401.09340, 2024. 1", + "[17] Mengzhao Jia, Wenhao Yu, Kaixin Ma, Tianqing Fang, Zhihan Zhang, Siru Ouyang, Hongming Zhang, Meng Jiang, and Dong Yu. Leopard: A Vision Language Model For Text-Rich Multi-Image Tasks, 2024. 1, 3", + "[18] Dongfu Jiang, Xuan He, Huaye Zeng, Cong Wei, Max Ku, Qian Liu, and Wenhu Chen. Mantis: interleaved multi-image instruction tuning. ArXiv preprint, abs/2405.01483, 2024. 3, 6", + "[19] Mukul Khanna*, Ram Ramrakhya*, Gunjan Chhablani, Sriram Yenamandra, Theophile Gervet, Matthew Chang, Zsolt Kira, Devendra Singh Chaplot, Dhruv Batra, and Roozbeh Mottaghi. Goat-bench: A benchmark for multi-modal life-long navigation. In CVPR, 2024. 5", + "[20] Taewoong Kim, Cheolhong Min, Byeonghwi Kim, Jinyeon Kim, Wonje Jeung, and Jonghyun Choi. Realfred: An embodied instruction following benchmark in photo-realistic environment. In ECCV, 2024. 5", + "[21] Hugo Laurençon, Andres Marafioti, Victor Sanh, and Léo Tronchon. Building and better understanding vision-language models: insights and future directions, 2024. 3, 6", + "[22] Bo Li, Yuanhan Zhang, Liangyu Chen, Jinghao Wang, Jingkang Yang, and Ziwei Liu. Otter: A Multi-Modal Model with In-Context Instruction Tuning, 2023. 1, 3" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "29577", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[23] Bohao Li, Yuying Ge, Yixiao Ge, Guangzhi Wang, Rui Wang, Ruimao Zhang, and Ying Shan. SEED-Bench: Benchmarking multimodal large language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13299-13308, 2024. 2, 3", + "[24] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. LLaVA-OneVision: Easy visual task transfer. ArXiv preprint, abs/2408.03326, 2024. 6", + "[25] Feng Li, Renrui Zhang, Hao Zhang, Yuanhan Zhang, Bo Li, Wei Li, Zejun Ma, and Chunyuan Li. Llava-last-interleave: Tackling multi-image, video, and 3d in large multimodal models, 2024. 3", + "[26] Juncheng Li, Kaihang Pan, Zhiqi Ge, Minghe Gao, Wei Ji, Wenqiao Zhang, Tat-Seng Chua, Siliang Tang, Hanwang Zhang, and Yueting Zhuang. Fine-tuning multimodal LLMs to follow zero-shot demonstrative instructions. In The Twelfth International Conference on Learning Representations, 2024. 2, 3", + "[27] Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Yi Liu, Zun Wang, Jilan Xu, Guo Chen, Ping Luo, et al. Mvbench: A comprehensive multi-modal video understanding benchmark. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22195-22206, 2024. 3", + "[28] Xiaoqi Li, Mingxu Zhang, Yiran Geng, Haoran Geng, Yuxing Long, Yan Shen, Renrui Zhang, Jiaming Liu, and Hao Dong. ManipLLM: Embodied multimodal large language model for object-centric robotic manipulation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18061-18070, 2024. 1", + "[29] Ji Lin, Hongxu Yin, Wei Ping, Pavlo Molchanov, Mohammad Shoeybi, and Song Han. Vila: On pre-training for visual language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26689-26699, 2024. 6", + "[30] Ziyi Lin, Chris Liu, Renrui Zhang, Peng Gao, Longtian Qiu, Han Xiao, Han Qiu, Chen Lin, Wenqi Shao, Keqin Chen, et al. Sphinx: The joint mixing of weights, tasks, and visual embeddings for multi-modal large language models. arXiv preprint arXiv:2311.07575, 2023. 1, 2", + "[31] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning, 2023. 2, 6, 8", + "[32] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. LLaVA-NeXT: Improved reasoning, OCR, and world knowledge, 2024. 2, 3", + "[33] Haowei Liu, Xi Zhang, Haiyang Xu, Yaya Shi, Chaoya Jiang, Ming Yan, Ji Zhang, Fei Huang, Chunfeng Yuan, Bing Li, and Weiming Hu. MIBench: Evaluating multimodal large language models over multiple images, 2024. 2, 3", + "[34] Fuwen Luo, Chi Chen, Zihao Wan, Zhaolu Kang, Qidong Yan, Yingjie Li, Xiaolong Wang, Siyu Wang, Ziyue Wang, Xiaoyue Mi, Peng Li, Ning Ma, Maosong Sun, and Yang Liu. Codis: Benchmarking context-dependent visual comprehension for multimodal large language models. arXiv preprint arXiv:2402.13607, 2024. 6", + "[35] Gen Luo, Xue Yang, Wenhan Dou, Zhaokai Wang, Jifeng Dai, Yu Qiao, and Xizhou Zhu. Mono-internvl: Pushing the" + ], + "bbox": [ + 91, + 90, + 480, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "boundaries of monolithic multimodal large language models with endogenous visual pre-training, 2024. 6", + "[36] Arjun Majumdar, Anurag Ajay, Xiaohan Zhang, Pranav Putta, Sriram Yenamandra, Mikael Henaff, Sneha Silwal, Paul Mcvay, Oleksandr Maksymets, Sergio Arnaud, et al. OpenEQA: embodied question answering in the era of foundation models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16488-16498, 2024. 5", + "[37] Fanqing Meng, Jin Wang, Chuanhao Li, Quanfeng Lu, Hao Tian, Jiaqi Liao, Xizhou Zhu, Jifeng Dai, Yu Qiao, Ping Luo, Kaipeng Zhang, and Wenqi Shao. MMIU: Multimodal multi-image understanding for evaluating large vision-language models, 2024. 2", + "[38] Jiahao Nie, Gongjie Zhang, Wenbin An, Yap-Peng Tan, Alex C Kot, and Shijian Lu. Mmrel: A relation understanding dataset and benchmark in the mllm era. arXiv preprint arXiv:2406.09121, 2024. 1, 2", + "[39] OpenAI. Hello GPT-4o, 2024. Accessed on June 16, 2024. 6", + "[40] Xavier Puig, Eric Undersander, Andrew Szot, Mikael Dallaire Cote, Tsung-Yen Yang, Ruslan Partsey, Ruta Desai, Alexander Clegg, Michal Hlavac, So Yeon Min, Vladimir Vondruš, Theophile Gervet, Vincent-Pierre Berges, John M Turner, Oleksandr Maksymets, Zsolt Kira, Mrinal Kalakrishnan, Jitendra Malik, Devendra Singh Chaplot, Unnat Jain, Dhruv Batra, Akshara Rai, and Roozbeh Mottaghi. Habitat 3.0: A co-habitat for humans, avatars, and robots. In Proceedings of the Twelfth International Conference on Learning Representations (ICLR 2024), 2024. 5", + "[41] Navid Rajabi and Jana Kosecka. Gsr-bench: A benchmark for grounded spatial reasoning evaluation via multimodal llms. arXiv preprint arXiv:2406.13246, 2024. 2", + "[42] Abhinav Rajvanshi, Karan Sikka, Xiao Lin, Bhoram Lee, Han pang Chiu, and Alvaro Velasquez. SayNav: Grounding large language models for dynamic planning to navigation in new environments. In 34th International Conference on Automated Planning and Scheduling, 2024. 1", + "[43] Santhosh Kumar Ramakrishnan, Aaron Gokaslan, Erik Wijmans, Oleksandr Maksymets, Alexander Clegg, John M Turner, Eric Undersander, Wojciech Galuba, Andrew Westbury, Angel X Chang, Manolis Savva, Yili Zhao, and Dhruv Batra. Habitat-matterport 3d dataset (HM3d): 1000 large-scale 3d environments for embodied AI. In Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 2), 2021. 5", + "[44] Machel Reid, Nikolay Savinov, Denis Teplyashin, Dmitry Lepikhin, Timothy Lillicrap, Jean-baptiste Alayrac, Radu Soricut, Angeliki Lazaridou, Orhan First, Julian Schrittwieser, et al. Gemini 1.5: unlocking multimodal understanding across millions of tokens of context. ArXiv preprint, abs/2403.05530, 2024. 6", + "[45] Allen Z. Ren, Jaden Clark, Anushri Dixit, Masha Itkina, Anirudha Majumdar, and Dorsa Sadigh. Explore until confident: Efficient exploration for embodied question answering, 2024. 5", + "[46] Manolis Savva, Abhishek Kadian, Oleksandr Maksymets, Yili Zhao, Erik Wijmans, Bhavana Jain, Julian Straub, Jia" + ], + "bbox": [ + 516, + 90, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "29578", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Liu, Vladlen Koltun, Jitendra Malik, Devi Parikh, and Dhruv Batra. Habitat: A Platform for Embodied AI Research. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2019. 5", + "[47] Fatemeh Shiri, Xiao-Yu Guo, Mona Far, Xin Yu, Reza Haf, and Yuan-Fang Li. An empirical analysis on spatial reasoning capabilities of large multimodal models. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 21440-21455, 2024. 1, 2", + "[48] Andrew Szot, Alex Clegg, Eric Undersander, Erik Wijmans, Yili Zhao, John Turner, Noah Maestre, Mustafa Mukadam, Devendra Chaplot, Oleksandr Maksymets, Aaron Gokaslan, Vladimir Vondrus, Sameer Dharur, Franziska Meier, Wojciech Galuba, Angel Chang, Zsolt Kira, Vladlen Koltun, Jitendra Malik, Manolis Savva, and Dhruv Batra. Habitat 2.0: Training home assistants to rearrange their habitat. In Advances in Neural Information Processing Systems (NeurIPS), 2021. 5", + "[49] Yunlong Tang, Jing Bi, Siting Xu, Luchuan Song, Susan Liang, Teng Wang, Daoan Zhang, Jie An, Jingyang Lin, Rongyi Zhu, et al. Video understanding with large language models: A survey. arXiv preprint arXiv:2312.17432, 2023. 3", + "[50] Jiayu Wang, Yifei Ming, Zhenmei Shi, Vibhav Vineet, Xin Wang, Yixuan Li, and Neel Joshi. Is a picture worth a thousand words? delving into spatial reasoning for vision language models. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 2", + "[51] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Yang Fan, Kai Dang, Mengfei Du, Xuancheng Ren, Rui Men, Dayiheng Liu, Chang Zhou, Jingren Zhou, and Junyang Lin. Qwen2-VL: Enhancing vision-language model's perception of the world at any resolution, 2024. 6", + "[52] Xidong Wang, Dingjie Song, Shunian Chen, Chen Zhang, and Benyou Wang. LongLLaVA: Scaling Multi-modal LLMs to 1000 Images Efficiently via a Hybrid Architecture, 2024. 2, 3", + "[53] Xiyao Wang, Yuhang Zhou, Xiaoyu Liu, Hongjin Lu, Yuancheng Xu, Feihong He, Jaehong Yoon, Taixi Lu, Fuxiao Liu, Gedas Bertasius, Mohit Bansal, Huaxiu Yao, and Furong Huang. Mementos: A comprehensive benchmark for multimodal large language model reasoning over image sequences. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 416-442, Bangkok, Thailand, 2024. Association for Computational Linguistics. 2, 3", + "[54] Ziyue Wang, Chi Chen, Fuwen Luo, Yurui Dong, Yuanchi Zhang, Yuzhuang Xu, Xiaolong Wang, Peng Li, and Yang Liu. ActiView: Evaluating active perception ability for multimodal large language models. arXiv preprint arXiv:2410.04659, 2024. 3, 8", + "[55] Ziyue Wang, Chi Chen, Yiqi Zhu, Fuwen Luo, Peng Li, Ming Yan, Ji Zhang, Fei Huang, Maosong Sun, and Yang Liu. Browse and Concentrate: comprehending multimodal content via prior-llm context fusion. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 11229–11245," + ], + "bbox": [ + 91, + 90, + 483, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Bangkok, Thailand, 2024. Association for Computational Linguistics. 3, 6", + "[56] Zhutian Yang, Caelan Garrett, Dieter Fox, Tomás Lozano-Pérez, and Leslie Pack Kaelbling. Guiding long-horizon task and motion planning with vision language models, 2024. 1", + "[57] Yuan Yao, Tianyu Yu, Ao Zhang, Chongyi Wang, Junbo Cui, Hongji Zhu, Tianchi Cai, Haoyu Li, Weilin Zhao, Zhihui He, et al. MiniCPM-V: A GPT-4V level MLLM on your phone. ArXiv preprint, abs/2408.01800, 2024. 6", + "[58] Jiabo Ye, Haiyang Xu, Haowei Liu, Anwen Hu, Ming Yan, Qi Qian, Ji Zhang, Fei Huang, and Jingren Zhou. mPLUG-Owl3: towards long image-sequence understanding in multi-modal large language models. ArXiv preprint, abs/2408.04840, 2024. 3, 6", + "[59] Haotian Zhang, Mingfei Gao, Zhe Gan, Philipp Duffer, Nina Wenzel, Forrest Huang, Dhruti Shah, Xianzhi Du, Bowen Zhang, Yanghao Li, Sam Dodge, Keen You, Zhen Yang, Aleksei Timofeev, Mingze Xu, Hong-You Chen, JeanPhilippe Fauconnier, Zhengfeng Lai, Haoxuan You, Zirui Wang, Afshin Dehghan, Peter Grasch, and Yinfei Yang. MM1.5: Methods, analysis & insights from multimodal llm fine-tuning, 2024. 3", + "[60] Peiyuan Zhang, Kaichen Zhang, Bo Li, Guangtao Zeng, Jingkang Yang, Yuanhan Zhang, Ziyue Wang, Haoran Tan, Chunyuan Li, and Ziwei Liu. Long context transfer from language to vision. arXiv preprint arXiv:2406.16852, 2024. 1, 2, 6", + "[61] Yi-Fan Zhang, Huanyu Zhang, Haochen Tian, Chaoyou Fu, Shuangqing Zhang, Junfei Wu, Feng Li, Kun Wang, Qing-song Wen, Zhang Zhang, Liang Wang, Rong Jin, and Tieniu Tan. Mme-realworld: Could your multimodal llm challenge high-resolution real-world scenarios that are difficult for humans?, 2024. 1", + "[62] Haozhe Zhao, Zefan Cai, Shuzheng Si, Xiaojian Ma, Kaikai An, Liang Chen, Zixuan Liu, Sheng Wang, Wenjuan Han, and Baobao Chang. MMICL: Empowering vision-language model with multi-modal in-context learning. In The Twelfth International Conference on Learning Representations, 2024. 3", + "[63] Gengze Zhou, Yicong Hong, Zun Wang, Xin Eric Wang, and Qi Wu. NavGPT-2: Unleashing navigational reasoning capability for large vision-language models. In European Conference on Computer Vision, pages 260–278. Springer, 2025. 1, 3" + ], + "bbox": [ + 516, + 92, + 903, + 710 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "29579", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/CoSpace_ Benchmarking Continuous Space Perception Ability for Vision-Language Models/45f50ce2-59a4-414e-86e0-abbb59b51daa_model.json b/2025/CoSpace_ Benchmarking Continuous Space Perception Ability for Vision-Language Models/45f50ce2-59a4-414e-86e0-abbb59b51daa_model.json new file mode 100644 index 0000000000000000000000000000000000000000..ee0839a78ef472d925bb638d4ec03f775d223320 --- /dev/null +++ b/2025/CoSpace_ Benchmarking Continuous Space Perception Ability for Vision-Language Models/45f50ce2-59a4-414e-86e0-abbb59b51daa_model.json @@ -0,0 +1,3192 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.185, + 0.131, + 0.817, + 0.177 + ], + "angle": 0, + "content": "CoSpace: Benchmarking Continuous Space Perception Ability for Vision-Language Models" + }, + { + "type": "text", + "bbox": [ + 0.238, + 0.203, + 0.77, + 0.222 + ], + "angle": 0, + "content": "Yiqi Zhu\\(^{1*}\\), Ziyue Wang\\(^{1*}\\), Can Zhang\\(^{3}\\), Peng Li\\(^{2,4\\dagger}\\), Yang Liu\\(^{1,2,4,5\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.177, + 0.222, + 0.834, + 0.24 + ], + "angle": 0, + "content": "\\(^{1}\\)Dept. of Comp. Sci. & Tech., Institute for AI, Tsinghua University, Beijing, China" + }, + { + "type": "text", + "bbox": [ + 0.195, + 0.24, + 0.814, + 0.257 + ], + "angle": 0, + "content": "\\(^{2}\\)Institute for AI Industry Research (AIR), Tsinghua University, Beijing, China" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.257, + 0.907, + 0.275 + ], + "angle": 0, + "content": "\\(^{3}\\)School of Computer and Communication Engineering, University of Science and Technology Beijing" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.276, + 0.749, + 0.292 + ], + "angle": 0, + "content": "\\(^{4}\\)Shanghai Artificial Intelligence Laboratory, Shanghai, China" + }, + { + "type": "text", + "bbox": [ + 0.171, + 0.293, + 0.836, + 0.311 + ], + "angle": 0, + "content": "\\(^{5}\\)Jiangsu Collaborative Innovation Center for Language Competence, Jiangsu, China" + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.319, + 0.516, + 0.46 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.208, + 0.463, + 0.402, + 0.473 + ], + "angle": 0, + "content": "The way models perceive real-world scenarios" + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.32, + 0.902, + 0.416 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.418, + 0.545, + 0.439 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.548, + 0.42, + 0.84, + 0.449 + ], + "angle": 0, + "content": "Each of the four images contain one car respectively. However 1 2 feature the same car, and 3 4 feature another identical car. Thus, there are only two cars in the scene." + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.451, + 0.545, + 0.471 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.549, + 0.452, + 0.846, + 0.47 + ], + "angle": 0, + "content": "The four images respectively contains one car. Four cars all together are visible in the scene." + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.479, + 0.907, + 0.521 + ], + "angle": 0, + "content": "Figure 1. Illustration of continuous space perception: The left side depicts the construction of an image sequence representing a continuous space. The right side presents an example task related to the continuous space, showing how adjacent images are connected. Note that to answer the question correctly, one must recognize that the cars in images 1 and 2 are the same, as well as those in images 3 and 4." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.53, + 0.327, + 0.545 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.569, + 0.486, + 0.903 + ], + "angle": 0, + "content": "Vision-Language Models (VLMs) have recently witnessed significant progress in visual comprehension. As the permitting length of image context grows, VLMs can now comprehend a broader range of views and spaces. Current benchmarks provide insightful analysis of VLMs in tasks involving complex visual instructions following, multi-image understanding and spatial reasoning. However, they usually focus on spatially irrelevant images or discrete images captured from varied viewpoints. The compositional characteristic of images captured from a static viewpoint remains underestimated. We term this characteristic as Continuous Space Perception. When observing a scene from a static viewpoint while shifting orientations, it produces a series of spatially continuous images, enabling the reconstruction of the entire space. In this paper, we present CoSpace, a multi-image visual understanding benchmark designed to assess the Continuous Space perception ability for VLMs. CoSpace contains 2,918 images and 1,626 question-answer pairs, covering seven types of tasks. We conduct evaluation across 19 proprietary and open-source VLMs. Results reveal that there exist pitfalls on the continuous space perception ability for most of the evaluated models, including pro" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.531, + 0.907, + 0.622 + ], + "angle": 0, + "content": "prietary ones. Interestingly, we find that the main discrepancy between open-source and proprietary models lies not in accuracy but in the consistency of responses. We believe that enhancing the ability of continuous space perception is essential for VLMs to perform effectively in real-world tasks and encourage further research to advance this capability." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.631, + 0.646, + 0.645 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.658, + 0.907, + 0.824 + ], + "angle": 0, + "content": "With the advancements in deep learning and Large Language Models (LLMs), Vision-Language Models (VLMs) have demonstrated exceptional performance across a range of downstream tasks [4, 17, 22, 30, 47, 60]. Driven by this, VLMs are now being used in diverse applications, such as planning [5, 56], navigation [8, 63] and embodied tasks [28, 42], and are further applied in real-world scenarios [3, 61]. Real-world scenario, featuring the complicated environment, cast significant challenges on VLMs, especially the ability to capture sophisticated spatial information from surrounding spaces." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.826, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Despite previous attempts to advance spatial understanding in VLMs [4, 38], there remain areas requiring further exploration. Various strategies have been explored to empower models with the ability to identify and ground objects based on spatial relationship [4, 16]. Shiri et al. [47]" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "29569" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.188, + 0.089, + 0.808, + 0.288 + ], + "angle": 0, + "content": "
BenchmarksEvaluation TargetFeaturesNum.ImgEvaluationInstances
Spatial ReasoningMulti-image PerceptionStatic Viewpoint
MMRel [38]Object-relationship○××15k15k
Spatial-MM [47]Object-relationship and spatial-CoT✓××2.3k400
SpatialEval [50]Spatial Reasoning✓××4.6k4.6k
GSR-Bench [41]Grounded spatial reasoning✓××4.9k4.9k
MileBench [9]Multimodal Long Context×✓×97.9k6.4k
Demon-Core [26]Demonstrative instruction following×✓×62.8k18.2k
Mementos-Val [53]Temporal sequential-image understanding×✓×-699
MIBench [33]Fine-grained multi-image understanding×✓×125k13k
BLINK [14]General visual perception○○×7.3k3.8k
SEED-Bench [23]Visual comprehension○○×1.9k24k
Video-MME [13]Visual comprehension○✓×0.9k2.7k
MMIU [37]Visual comprehension○✓×77.7k11.7k
SparklesEval [15]Multi-image Chat○✓×550150
CoSpace (Ours)continuous space perception✓✓✓2,9181,626
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.291, + 0.907, + 0.319 + ], + "angle": 0, + "content": "Table 1. Comparison with selected benchmarks that address multi-image evaluation and/or spatial evaluation. \\( \\checkmark \\) denotes the main focus of the corresponding feature, \\( \\circ \\) indicates that the feature exists but is not the primary focus, while \\( X \\) implies a lack of the feature." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.328, + 0.482, + 0.493 + ], + "angle": 0, + "content": "and Wang et al. [50] assess spatial understanding ability for VLMs with the recognition of basic tasks in this domain. However, much of these work center on single-image inputs, without further extending to multi-image scenarios. Works on video understanding [52, 60] present the concept of spatial temporal understanding, but the spatial information of videos involve sequential frames that inherently linking spatial understanding with additional temporal dynamics. Given these attempts in spatial understanding, we pose the question: Are the current efforts adequate for advancing the research of spatial understanding?" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.494, + 0.483, + 0.63 + ], + "angle": 0, + "content": "We claim that a gap still exists between the concept proposed in previous work and more advanced spatial understanding. For humans, exposing to an unfamiliar environment requires determining the position, identifying surrounding objects, and making decisions based on spatial context. Such process typically involves integrating various information about the surrounding space. We refer to it as continuous space perception, which is an essential ability in real-world towards effectively completion of daily tasks." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.63, + 0.483, + 0.81 + ], + "angle": 0, + "content": "Figure 1 shows an example of continuous space perception: Conceptually, humans perceive continuous visual space as a series of \"images\", each representing a segment of the space with overlaps between adjacent views. Mimicking this perception, Figure 1 illustrates the continuous visual space using four images. To answer questions accurately about these images requires a comprehensive understanding of all four images and reconstructing the entire space they represent. For example, to accurately count the number of cars in the space, it is essential to recognize that the cars appearing in the first and second images are the same, as are the cars in the third and fourth images." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.811, + 0.483, + 0.901 + ], + "angle": 0, + "content": "The continuous space perception ability is equally important for VLMs as for human. Many real-world scenario tasks, such as planning and navigation, require models to effectively integrate the information in the surroundings to take the optimal actions. During this process, the reconstruction of the space heavily relies on continuous space" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.328, + 0.905, + 0.358 + ], + "angle": 0, + "content": "perception. However, as shown in Table 1, this ability is largely overlooked in existing benchmarks for VLMs." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.358, + 0.907, + 0.585 + ], + "angle": 0, + "content": "To this end, we present CoSpace, a visual understanding benchmark designed to evaluate the Continuous Space perception ability for VLMs. CoSpace includes four categories and seven tasks, emphasizing the following core abilities: (1) Can VLMs recognize direction and position when presented with spatially continuous images captured from a static viewpoint? (2) Can models identify the correspondence between adjacent images in the same scene? (3) Can models make optimal navigation decisions based on the perception of the space? We posit the importance of these three core abilities, with the first one forms the basic understanding of space, the second one is essential for integrating the information across different images and the last one surpasses mere literal understanding and makes a step towards more practical applications." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.586, + 0.906, + 0.736 + ], + "angle": 0, + "content": "We evaluated 19 widely used open-source and proprietary VLMs on our proposed CoSpace benchmark. Results show that though some of the open-source models achieve approaching or even higher accuracy compared with state-of-the-art proprietary models on certain tasks, proprietary models exhibit remarkably higher consistency in their responses across multi-dimensional evaluation metrics. Additionally, we found that with the alteration of input context and image sequence, the tasks that humans find more challenging do not always pose greater difficulty for the models." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.762, + 0.655, + 0.779 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.788, + 0.862, + 0.804 + ], + "angle": 0, + "content": "2.1. Multi-image Perception Ability of VLMs" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.81, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Large Vision-Language Models (VLMs), such as LLaVA models [31, 32] and Sphinx [30], have shown impressive performances on visual reasoning and grounding tasks that focus on only one image at a time. However, it is more natural to allow multiple images and arrange them in interleave format in one input. Along with the growing capability of large vision models, there are also many other applications. For example, the use of large vision models has been demonstrated for large datasets, such as the US Census dataset [33]." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "29570" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.289 + ], + "angle": 0, + "content": "ities of single-image models, multi-image perception and understanding abilities are gaining increasing attentions, including research topics such as multimodal in-context learning [10, 22, 62], interleaved visual instruction following [18, 25, 26], and temporal visual understanding [49], and practical applications such as navigation and embodied QA [63]. To achieve multi-image perception, some methods focus on constructing interleaved visual instruction tuning data [17, 18, 22, 62], some improve the model structure and training strategies for better cross-image understanding [26, 52, 55, 58], and others address the image processing strategy for adaptable visual feature encoding concerning multi-image scenarios [21, 32, 59]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.302, + 0.482, + 0.409 + ], + "angle": 0, + "content": "With the enhancements of the multi-image perception ability, VLMs are able to obtain visual information covering a wider range of views from multiple images at one time, resulting in more profound observations and understandings of the surrounding environment. In our paper, we take this as basis, and dedicate to investigating the spatial perception ability of VLMs." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.457, + 0.468, + 0.473 + ], + "angle": 0, + "content": "2.2. Spatial Comprehension Evaluation of VLMs" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.494, + 0.483, + 0.901 + ], + "angle": 0, + "content": "For the past years, the evaluation of spatial comprehension for VLMs has always been a commonly engaged topic, and it often refers to the spatial relation recognition among objects, such as MMRel, GSR-Bench and SpatialMM shown in Table 1. Recently, multi-image related evaluation for VLMs are widely explored, including interleaved visual instruction following [26, 33], temporal sequential-image comprehension [53], and long context understanding [9]. In multi-image scenarios, such as video comprehension [13, 23] and multi-image dialogue [15], the spatial reasoning ability is also one of the evaluated aspects. However, these benchmarks primarily focus on general multi-image comprehension where most of input images come from difference scenes. The compositional features across multiple images to integrate richer spatial information of a wider scene and more profound spatial understanding are seldom emphasized. Although Blink [14], ActiView [54], and video-based benchmarks [13, 27] require spatial understanding across multiple images, they either focus on different views of the same object, or discrete views from dynamic viewpoint. These views are not continuous regarding a static viewpoint, and the resulting images are not necessarily bounded regarding the same scene. In contrast to existing works, we address the continuous spatial perception from a static viewpoint, that requires models to connect all the perceived views together to form an intact understanding of the whole scene." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.091, + 0.613, + 0.108 + ], + "angle": 0, + "content": "3. CoSpace" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.116, + 0.833, + 0.133 + ], + "angle": 0, + "content": "3.1. Continuous Space Perception Ability" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.139, + 0.905, + 0.274 + ], + "angle": 0, + "content": "In this paper, we define the term, continuous visual space, as the perceived views from a static viewpoint towards different orientations that can naturally compose a complete scene or environment. As is shown in Figure 1, if one would like to correctly answer the question, it is needed to first capture the visual information of all the cars, and consequently, the ability of continuous space perception refers to capturing and processing information from these images of continuous visual spaces." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.276, + 0.905, + 0.593 + ], + "angle": 0, + "content": "To obtain an overall observation of environment, human perceive the whole surroundings from the same viewpoint towards different directions, and the percept continuous visual spaces can be reconstructed as a panorama in the brain. Different from human that directly views the space continuously, models typically take discrete images as input. Therefore, the way to present continuous visual spaces remains flexible. In our CoSpace, we present a complete continuous visual space as follows: we provide several images shot at the same viewpoint in the same scene, but towards different directions, e.g. respectively towards front, right, back and left. These images together form a complete space without omitting any visual information of the surrounding environment. Naturally, such strategy could produce overlapped spaces between adjacent images, which tightly connects images describing adjacent spaces, and can consequently help models better identify the relationships between images from continuous visual space. In this way, we posit that such strategy can effectively preserve continuous spatial information of an extensive space as well as be suitable to understanding for VLMs." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.605, + 0.644, + 0.621 + ], + "angle": 0, + "content": "3.2. Task Design" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.628, + 0.905, + 0.763 + ], + "angle": 0, + "content": "We comprehensively evaluate the continuous space perception ability of VLMs regarding four aspects. Two of them are fundamental ability of spatial perception, detailed as direction perception and recognition (Direction), and rotation-centric perception (Rotation). The other two are further practical tasks, detailed as counting-based perception (Counting), and embodied-based planning (Planning). These four categories are subdivided into seven tasks for elaborate investigation as shown in Figure 2." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.765, + 0.905, + 0.794 + ], + "angle": 0, + "content": "We employ two multiple choice tasks, both with four options, for the Direction category as follows:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.796, + 0.905, + 0.901 + ], + "angle": 0, + "content": "- Direction Recognition (DIR-Rec): In real world, identifying direction is inevitable when one is placed in a new environment. Similarly, the DIR-Rec task requires models to recognize the direction of targets by answering questions such as \"where is the building located?\" Meanwhile, we provide detailed task instructions, including direction references such as \"the first image is facing north," + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "29571" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.402, + 0.094, + 0.453, + 0.104 + ], + "angle": 0, + "content": "Direction" + }, + { + "type": "text", + "bbox": [ + 0.165, + 0.109, + 0.345, + 0.12 + ], + "angle": 0, + "content": "\\(\\clubsuit\\) Direction Recognition (DIR-Rec)" + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.121, + 0.419, + 0.161 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.095, + 0.161, + 0.387, + 0.171 + ], + "angle": 0, + "content": "Question: Which direction is the alleyway going in the images?" + }, + { + "type": "text", + "bbox": [ + 0.095, + 0.171, + 0.231, + 0.18 + ], + "angle": 0, + "content": "Options: A. South to North" + }, + { + "type": "text", + "bbox": [ + 0.147, + 0.181, + 0.267, + 0.189 + ], + "angle": 0, + "content": "C. Northeast to Southwest" + }, + { + "type": "text", + "bbox": [ + 0.278, + 0.171, + 0.352, + 0.181 + ], + "angle": 0, + "content": "B. East to West" + }, + { + "type": "text", + "bbox": [ + 0.283, + 0.181, + 0.401, + 0.189 + ], + "angle": 0, + "content": "D. Northwest to Southeast" + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.191, + 0.406, + 0.213 + ], + "angle": 0, + "content": "Rationale: The \\(2^{nd}\\) and \\(4^{th}\\) images indicate a direction of East-West, the \\(1^{st}\\) and \\(3^{rd}\\) images indicate the width of the alleyway." + }, + { + "type": "text", + "bbox": [ + 0.487, + 0.109, + 0.712, + 0.12 + ], + "angle": 0, + "content": "Directional Object Perception (DIR-Obj)" + }, + { + "type": "image", + "bbox": [ + 0.435, + 0.121, + 0.757, + 0.161 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.44, + 0.161, + 0.681, + 0.171 + ], + "angle": 0, + "content": "Question: What is visible in the southeast direction?" + }, + { + "type": "text", + "bbox": [ + 0.44, + 0.171, + 0.595, + 0.18 + ], + "angle": 0, + "content": "Options: A. A pedestrian bridge" + }, + { + "type": "text", + "bbox": [ + 0.492, + 0.181, + 0.597, + 0.19 + ], + "angle": 0, + "content": "C. A \"no parking\" sign" + }, + { + "type": "text", + "bbox": [ + 0.435, + 0.191, + 0.754, + 0.213 + ], + "angle": 0, + "content": "Rationale: Southeast refers to the left part of the \\(3^{rd}\\) image and the right part of the \\(4^{th}\\) image, only a pedestrian bridge presents." + }, + { + "type": "title", + "bbox": [ + 0.404, + 0.221, + 0.451, + 0.232 + ], + "angle": 0, + "content": "Rotation" + }, + { + "type": "text", + "bbox": [ + 0.179, + 0.235, + 0.331, + 0.247 + ], + "angle": 0, + "content": "Rotation-Angle (ROT-Ang)" + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.247, + 0.42, + 0.289 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.292, + 0.36, + 0.302 + ], + "angle": 0, + "content": "Question: What is the angle between the adjacent image?" + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.303, + 0.253, + 0.315 + ], + "angle": 0, + "content": "Options: A 30 B.60" + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.316, + 0.415, + 0.337 + ], + "angle": 0, + "content": "Rationale: Observing from the overlapping areas of the tall building, and the displacement of the short buildings. 30 degree is more feasible." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.235, + 0.685, + 0.245 + ], + "angle": 0, + "content": "Rotation-Difference (ROT-Dif)" + }, + { + "type": "image", + "bbox": [ + 0.439, + 0.245, + 0.757, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.44, + 0.305, + 0.737, + 0.315 + ], + "angle": 0, + "content": "Question: Which image does not belong to this image sequence?" + }, + { + "type": "text", + "bbox": [ + 0.435, + 0.316, + 0.721, + 0.338 + ], + "angle": 0, + "content": "Rationale: Observing from the range of the overlaps regarding the road., image 4 present the most rotation degree." + }, + { + "type": "title", + "bbox": [ + 0.404, + 0.345, + 0.451, + 0.356 + ], + "angle": 0, + "content": "Counting" + }, + { + "type": "text", + "bbox": [ + 0.382, + 0.359, + 0.472, + 0.37 + ], + "angle": 0, + "content": "- Counting (CNT)" + }, + { + "type": "image", + "bbox": [ + 0.098, + 0.372, + 0.19, + 0.411 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.192, + 0.376, + 0.307, + 0.412 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.308, + 0.376, + 0.421, + 0.42 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.423, + 0.375, + 0.571, + 0.42 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.581, + 0.369, + 0.645, + 0.382 + ], + "angle": 0, + "content": "Rationale:" + }, + { + "type": "text", + "bbox": [ + 0.58, + 0.385, + 0.747, + 0.404 + ], + "angle": 0, + "content": "One red truck spotted in the \\(3^{rd}\\) image, and it is spotted again in the \\(4^{th}\\) image." + }, + { + "type": "text", + "bbox": [ + 0.581, + 0.405, + 0.75, + 0.413 + ], + "angle": 0, + "content": "There is indeed 1 red truck in the scene." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.422, + 0.357, + 0.433 + ], + "angle": 0, + "content": "Question: How many red trucks are visible in the scene?" + }, + { + "type": "text", + "bbox": [ + 0.462, + 0.423, + 0.512, + 0.432 + ], + "angle": 0, + "content": "Answer: 1" + }, + { + "type": "text", + "bbox": [ + 0.615, + 0.422, + 0.757, + 0.432 + ], + "angle": 0, + "content": "Note that \\(\\mathbb{O}\\) denotes repeated cars" + }, + { + "type": "title", + "bbox": [ + 0.405, + 0.44, + 0.45, + 0.451 + ], + "angle": 0, + "content": "Planning" + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.453, + 0.421, + 0.503 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.099, + 0.505, + 0.37, + 0.515 + ], + "angle": 0, + "content": "Question: Where is the television regarding your position?" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.515, + 0.212, + 0.525 + ], + "angle": 0, + "content": "Options: A. Front-left" + }, + { + "type": "text", + "bbox": [ + 0.152, + 0.526, + 0.218, + 0.535 + ], + "angle": 0, + "content": "C. Front-right" + }, + { + "type": "text", + "bbox": [ + 0.278, + 0.515, + 0.336, + 0.525 + ], + "angle": 0, + "content": "B. Back-left" + }, + { + "type": "text", + "bbox": [ + 0.278, + 0.525, + 0.344, + 0.536 + ], + "angle": 0, + "content": "D. Back-right" + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.536, + 0.404, + 0.556 + ], + "angle": 0, + "content": "Rationale: TV is placed in the right part of \\(2^{nd}\\) image. Right part of the right view indicate back-right direction." + }, + { + "type": "image", + "bbox": [ + 0.435, + 0.455, + 0.758, + 0.496 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.496, + 0.674, + 0.505 + ], + "angle": 0, + "content": "- Planning-Decision (PLA-Dec)" + }, + { + "type": "text", + "bbox": [ + 0.441, + 0.505, + 0.686, + 0.515 + ], + "angle": 0, + "content": "Question: What should you do to reach the fireplace?" + }, + { + "type": "text", + "bbox": [ + 0.441, + 0.515, + 0.615, + 0.525 + ], + "angle": 0, + "content": "Options: A. Turn right and go ahead" + }, + { + "type": "text", + "bbox": [ + 0.491, + 0.526, + 0.608, + 0.535 + ], + "angle": 0, + "content": "C. Turn left and go ahead" + }, + { + "type": "text", + "bbox": [ + 0.627, + 0.515, + 0.752, + 0.525 + ], + "angle": 0, + "content": "B. Turn back and go ahead" + }, + { + "type": "text", + "bbox": [ + 0.629, + 0.526, + 0.751, + 0.535 + ], + "angle": 0, + "content": "D. Go ahead and turn back" + }, + { + "type": "text", + "bbox": [ + 0.435, + 0.536, + 0.745, + 0.556 + ], + "angle": 0, + "content": "Rationale:Fireplace locates in the far back, requiring both a turn action and a moving action, need to first turn backwards and then go ahead" + }, + { + "type": "image", + "bbox": [ + 0.765, + 0.094, + 0.783, + 0.109 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.796, + 0.098, + 0.888, + 0.108 + ], + "angle": 0, + "content": "Task instructions" + }, + { + "type": "text", + "bbox": [ + 0.766, + 0.112, + 0.895, + 0.121 + ], + "angle": 0, + "content": "General prompt template:" + }, + { + "type": "text", + "bbox": [ + 0.766, + 0.122, + 0.839, + 0.13 + ], + "angle": 0, + "content": "" + }, + { + "type": "text", + "bbox": [ + 0.766, + 0.131, + 0.892, + 0.14 + ], + "angle": 0, + "content": "" + }, + { + "type": "text", + "bbox": [ + 0.766, + 0.141, + 0.854, + 0.149 + ], + "angle": 0, + "content": "" + }, + { + "type": "text", + "bbox": [ + 0.766, + 0.15, + 0.857, + 0.158 + ], + "angle": 0, + "content": " " + }, + { + "type": "text", + "bbox": [ + 0.766, + 0.159, + 0.858, + 0.168 + ], + "angle": 0, + "content": "You should generate" + }, + { + "type": "text", + "bbox": [ + 0.766, + 0.17, + 0.891, + 0.178 + ], + "angle": 0, + "content": "answer from A, B, C or D." + }, + { + "type": "text", + "bbox": [ + 0.767, + 0.178, + 0.828, + 0.185 + ], + "angle": 0, + "content": "Your answer:" + }, + { + "type": "text", + "bbox": [ + 0.766, + 0.193, + 0.886, + 0.202 + ], + "angle": 0, + "content": "" + }, + { + "type": "text", + "bbox": [ + 0.766, + 0.203, + 0.882, + 0.211 + ], + "angle": 0, + "content": "for Direction and Counting" + }, + { + "type": "text", + "bbox": [ + 0.766, + 0.212, + 0.811, + 0.221 + ], + "angle": 0, + "content": "categories:" + }, + { + "type": "text", + "bbox": [ + 0.766, + 0.222, + 0.889, + 0.23 + ], + "angle": 0, + "content": "the first image is facing north," + }, + { + "type": "text", + "bbox": [ + 0.766, + 0.231, + 0.896, + 0.239 + ], + "angle": 0, + "content": "the second facing east, the third" + }, + { + "type": "text", + "bbox": [ + 0.766, + 0.24, + 0.876, + 0.248 + ], + "angle": 0, + "content": "facing south and the fourth" + }, + { + "type": "text", + "bbox": [ + 0.766, + 0.249, + 0.816, + 0.257 + ], + "angle": 0, + "content": "facing west." + }, + { + "type": "text", + "bbox": [ + 0.766, + 0.267, + 0.887, + 0.276 + ], + "angle": 0, + "content": "" + }, + { + "type": "text", + "bbox": [ + 0.766, + 0.277, + 0.859, + 0.284 + ], + "angle": 0, + "content": "for Rotation category:" + }, + { + "type": "text", + "bbox": [ + 0.766, + 0.285, + 0.896, + 0.294 + ], + "angle": 0, + "content": "images are arranged clockwise ." + }, + { + "type": "text", + "bbox": [ + 0.766, + 0.295, + 0.854, + 0.304 + ], + "angle": 0, + "content": "for example, after tak" + }, + { + "type": "text", + "bbox": [ + 0.766, + 0.305, + 0.888, + 0.313 + ], + "angle": 0, + "content": "first image, the photographer" + }, + { + "type": "text", + "bbox": [ + 0.766, + 0.314, + 0.897, + 0.322 + ], + "angle": 0, + "content": "turns a certain degree clockwise" + }, + { + "type": "text", + "bbox": [ + 0.766, + 0.323, + 0.816, + 0.331 + ], + "angle": 0, + "content": "for capturin" + }, + { + "type": "text", + "bbox": [ + 0.766, + 0.342, + 0.887, + 0.351 + ], + "angle": 0, + "content": "" + }, + { + "type": "text", + "bbox": [ + 0.766, + 0.352, + 0.859, + 0.359 + ], + "angle": 0, + "content": "for Planning category:" + }, + { + "type": "text", + "bbox": [ + 0.766, + 0.36, + 0.895, + 0.413 + ], + "angle": 0, + "content": "you are facing the direction in the first image, with the second image showing the scene to your right, the third image showing the scene at your back, and the fourth image showing the scene to your left." + }, + { + "type": "image", + "bbox": [ + 0.766, + 0.424, + 0.783, + 0.438 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.792, + 0.427, + 0.892, + 0.438 + ], + "angle": 0, + "content": "Notes for readers" + }, + { + "type": "text", + "bbox": [ + 0.766, + 0.444, + 0.886, + 0.453 + ], + "angle": 0, + "content": "- The directional symbols:" + }, + { + "type": "image", + "bbox": [ + 0.775, + 0.455, + 0.88, + 0.475 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.763, + 0.477, + 0.897, + 0.486 + ], + "angle": 0, + "content": "are invisible to models. They" + }, + { + "type": "text", + "bbox": [ + 0.763, + 0.487, + 0.884, + 0.496 + ], + "angle": 0, + "content": "are used for easier under" + }, + { + "type": "text", + "bbox": [ + 0.763, + 0.497, + 0.882, + 0.505 + ], + "angle": 0, + "content": "standing of image orders." + }, + { + "type": "text", + "bbox": [ + 0.763, + 0.513, + 0.867, + 0.522 + ], + "angle": 0, + "content": "- The Rationale is also" + }, + { + "type": "text", + "bbox": [ + 0.763, + 0.523, + 0.89, + 0.532 + ], + "angle": 0, + "content": "invisible to models, as they" + }, + { + "type": "text", + "bbox": [ + 0.763, + 0.533, + 0.88, + 0.542 + ], + "angle": 0, + "content": "intend for explaining the" + }, + { + "type": "text", + "bbox": [ + 0.763, + 0.542, + 0.858, + 0.55 + ], + "angle": 0, + "content": "results for readers." + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.563, + 0.905, + 0.591 + ], + "angle": 0, + "content": "Figure 2. Examples of CoSpace, along with prompt templates for evaluation. We also provide extra visual guidance and textual rationales for readers (invisible to models during evaluation) for easier understanding." + }, + { + "type": "text", + "bbox": [ + 0.104, + 0.605, + 0.483, + 0.664 + ], + "angle": 0, + "content": "the second image is facing east\". We standardize 8 directional options, including four cardinal directions, North, East, South and West, and four intercardinal directions, Northeast, Southeast, Southwest, and Northwest." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.665, + 0.483, + 0.877 + ], + "angle": 0, + "content": "- Directional Object Perception (DIR-Obj): Some real-world tasks demand abilities far beyond merely identifying the direction of a specified target. It is also a critical ability to distinguish content regarding different directions and to align objects with correct directions. Given a specified direction, the DIR-Obj task requires models to be aware of the appearing objects, and also distinguish those absent from that direction. To achieve this, models need to first reconstruct the entire space from continuous views, and then correctly identify the objects dedicate to a given direction. Specifically, in this task, questions typically focus on the intercardinal directions such as \"What is visible in the southeast direction?\", which necessitates a more comprehensive understanding of the entire space." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.886, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Observing the surroundings from a static viewpoint natu" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.605, + 0.905, + 0.651 + ], + "angle": 0, + "content": "rally involves rotating the perspectives. Therefore, we summaries Rotation as another fundamental category of continuous space perception, with two tasks designed as follows:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.66, + 0.906, + 0.87 + ], + "angle": 0, + "content": "- Rotation-Angle (ROT-Ang): For an observer, when rotating the perspective for more spatial information, there could be overlaps between adjacent observations. Specifically, human can tell the approximate turning angle between consecutive images observed from a static viewpoint by noticing identical furnishings appearing across images. In this task, we investigate if models exhibit similar ability as human does, which requires fine-grained continuous spatial understanding. We apply unified question \"What is the turning angle between the adjacent image?\" and provide two candidate options for model. This task requires models to carefully exam the overlapping and different areas between adjacent images and the shared features in spatially continuous image sequence." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.871, + 0.905, + 0.901 + ], + "angle": 0, + "content": "- Rotation-Difference (ROT-Dif): Following the ROT-Ang task, we also investigate the model ability to iden" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "29572" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.104, + 0.092, + 0.483, + 0.227 + ], + "angle": 0, + "content": "tify distinguished rotation angles. For the ROT-Dif task, models are given a sequence of five images, four of which share the same turning angle and the other is exceptional. Compared to ROT-Ang, this task emphasizes more on the global understanding of images regarding continuous visual space. The unified question of this task is \"Which image does not belong to this image sequence?\" It is an open-ended question without options, where models are required to answer with index of the exceptional image." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.228, + 0.483, + 0.319 + ], + "angle": 0, + "content": "On the basis of the above two fundamental categories, we further challenge models with two practical categories. The Counting category addresses the occurrence of objects in the entire scene. The Planning category comprises two spatial-related tasks widely evaluated in embodied AI tasks. For the Counting category, the task is described as:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.323, + 0.483, + 0.549 + ], + "angle": 0, + "content": "- Counting (CNT): Counting is a widely adopted task in existing visual benchmarks that asks models to recognize objects and deduce the time of occurrence of target objects. Generally, existing counting tasks only require models to deal with a single image or multiple spatially discrete images, while in our benchmark, images from continuous visual space where identical objects can occur in multiple images are focused on. This raise a challenge for models to not only recognize targets and count for their occurrences, but also be aware of the existence of the same object across different images. To achieve this, models should locate the overlapping area of adjacent images and align the same object appearing different images. The CNT task is an open-ended task, where models should respond with the total count of the target." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.553, + 0.483, + 0.718 + ], + "angle": 0, + "content": "The ability of continuous space perception is also significant in embodied scenarios. Intuitively, regarding the location of an embodied agent as the static viewpoint, it should have a profound understanding of the overall surroundings to effectively solve questions, and to efficiently plan for the following action sequence. Embodied question-answering (EQA) [36, 45] and navigation [19, 20] are typical tasks that require understand continuous visual space. Following these, we split the Planning category into two tasks addressing the relative locations of targets regarding the VLM-based agent as follows:" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.72, + 0.483, + 0.84 + ], + "angle": 0, + "content": "- Planning-Question Answering (PLA-QA): Following the implementation of EQA, we develop the PLA-QA task, requiring models to identify the location of a certain object given a continuous embodied space. In this task, instructions like \"Where is the television regarding your position?\" are provided for the models, and we formulate four options, containing candidate directions relative to the agent, for each question." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.84, + 0.483, + 0.902 + ], + "angle": 0, + "content": "- Planning-Decision (PLA-Dec): This task further investigates the understanding of continuous visual space by asking models to select the proper route to reach the target object. For disambiguation, we standardize the action" + }, + { + "type": "list", + "bbox": [ + 0.092, + 0.72, + 0.483, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.57, + 0.091, + 0.867, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.526, + 0.263, + 0.892, + 0.277 + ], + "angle": 0, + "content": "Figure 3. Distribution of categories and tasks in our CoSpace." + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.295, + 0.905, + 0.385 + ], + "angle": 0, + "content": "space as turning (turning to other directions without displacement) and go ahead. The PLA-Dec task especially focus on the order of actions. As is shown in the case in Figure 2, the options \"B. Turn back and go ahead\" and \"D. Go ahead and turn back\" represent two totally different actions and end up in different positions." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.403, + 0.776, + 0.417 + ], + "angle": 0, + "content": "3.3. Data Collection and Statistics" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.427, + 0.905, + 0.533 + ], + "angle": 0, + "content": "The construction of CoSpace follows the collect-filter-annotate pipeline. We first adopt automatic collection approaches to gather images, and then manually select images that satisfy detailed requirements of our designed tasks and generate questions for these images. Finally human annotators provide answers to each questions. The process is detailed in the following of this section." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.536, + 0.906, + 0.824 + ], + "angle": 0, + "content": "Firstly, to automatically collect sets of images captured in the same scene at static viewpoint but towards different directions, we employ two primary sources, Baidu Map Panorama API that allows us to acquire data towards any desired directions in the real world scene, and HM3D [43], a dataset of 1,000 building-scale 3D reconstructions for embodied tasks. For HM3D, by using Habitat-Sim [40, 46, 48], we instruct an agent wandering in the habitats, we are able to collect images of any views through the attached camera. Secondly, we manually filtered the auto-collected images according to our task requirements, ensuring that the selected images all contain sufficient visual information. With the filtered images, we employed GPT-4o to generate questions via few-shot prompting for the required tasks. Finally, human annotators, an average of two annotators per question, provided answers to these generated questions. This process allowed us to discard vague or unanswerable questions, ensuring that only high-quality question-answer pairs were included in our proposed CoSpace." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Our CoSpace comprises a total of 1,626 question-answer pairs and 2,918 images. Among the images, 2,302 are accessed from Baidu Map API and 616 are derived from HM3D dataset. The distribution of instances over categories and tasks is shown in Figure 3." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "29573" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.153, + 0.089, + 0.845, + 0.359 + ], + "angle": 0, + "content": "
ModelsDIR-RecDIR-ObjCNTROT-AngROT-DifPLA-QAPLA-DecAverage
ACCqACCpACCqACCpACCqACCpACCACCACCACC
Proprietary Models
Claude-3.7-sonnet [2]44.4029.2045.6035.6045.0038.0064.3393.5054.7369.3451.97
Gemini-1.5-pro [44]37.6015.6040.6031.6038.2524.0059.6782.0051.6462.9144.39
GPT-4o [39]40.4022.8046.0032.0040.0023.5058.3350.5053.0554.4642.10
>70B Open-source Models
InternVL2_5-78B [6]32.2024.4054.4047.6051.2543.0050.0077.0032.3942.7245.50
Qwen2-VL-72B [51]31.0023.2053.6044.8044.7537.0050.0062.0042.7259.1544.82
InternVL2-76B [7]33.6013.2046.4039.2050.0043.5050.0023.0025.3530.0535.43
LLaVA-OneVision-72B [24]19.2010.8044.2034.4024.2519.0050.0026.0032.8630.9929.17
<13B Open-source Models
MiniCPM-V 2.6 [57]32.8021.2040.4031.6038.5031.5050.0056.0041.3127.7037.10
Qwen2-VL-7B [51]26.4016.4039.2031.6045.0036.0050.0051.5034.2726.7635.71
Mantis-8B [18]30.6024.4031.2028.0041.7536.5050.0039.5033.3327.7034.30
InternVL2-8B [7]29.0012.8037.2030.4038.5031.0050.3347.5032.3927.7033.68
VILA1.5-8B [29]34.0030.4028.2024.0042.2538.0050.0018.5022.0742.7233.01
Idefics3-8B [21]34.0021.2038.6028.0032.7525.0048.0028.0026.7625.3530.77
LLaVA-OneVision-7B [24]18.8012.0037.6033.6037.7533.0050.0021.5029.5826.2930.01
Phi-3.5-vision [1]20.4012.0033.2028.4036.7534.0050.0022.5034.7426.7629.88
Brote-IM-XXL [55]33.0013.2031.0024.8030.7529.5050.0018.0026.7620.1927.72
LongVA-7B [60]23.0016.0031.6028.0028.0022.0048.6721.5019.2530.9926.90
Mono-InternVL-2B [35]30.0027.6032.0027.6014.008.5050.0018.0025.3524.8825.79
mPLUG-Owl3-7B [58]22.806.8029.2015.6022.259.0046.3316.5027.7021.1321.73
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.36, + 0.908, + 0.391 + ], + "angle": 0, + "content": "Table 2. Main results on our benchmark. The proprietary models are accessed via APIs, and the open-source models are accessed from Huggingface checkpoints." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.405, + 0.226, + 0.423 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.431, + 0.484, + 0.567 + ], + "angle": 0, + "content": "As discussed above, CoSpace features reconstructing the surrounding space via multiple images, with each image representing a different space segments. Different from previous multi-image benchmark involving spatial reasoning tasks, the input order of images is crucial in our tasks. To enable that models can associate each image with a specific direction to answer questions in our benchmark, we provide direction references along with images following the prompt templates in Figure 2 through out the evaluation." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.577, + 0.276, + 0.592 + ], + "angle": 0, + "content": "4.1. Evaluation Metrics" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.599, + 0.483, + 0.795 + ], + "angle": 0, + "content": "For the Direction and the Counting categories, for fair evaluation, we adopt both pair-wise accuracy \\( ACC_{p} \\) and query-wise accuracy \\( ACC_{q} \\) as introduced by Fu et al. [12] and Luo et al. [34] instead of simple accuracy. For each question, we task models twice with different prompts. \\( ACC_{p} \\) only counts when models correctly response to both prompts, while \\( ACC_{q} \\) is the averaged score of the two responses. For these different prompts regarding the same question, we vary only the order of images by preserving the clockwise order but changing the beginning view, so that the answer to the question remains unchanged. For example, if original image sequence begins with the northern image, the modified sequence could start from the east, south or west." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.484, + 0.903 + ], + "angle": 0, + "content": "For the Rotation category, some of the instances does not cover the entire space, and the such a modification of input image sequence will inevitable disrupt the comprehension of the continuous space. Similarly, for the Planning category, considering that the original input sequence features the front view as the first image and this manner inherently matches the thinking pattern when models take" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.407, + 0.907, + 0.452 + ], + "angle": 0, + "content": "actions. Therefore, for these two categories, we apply only a single context for each instances during evaluation, and report accuracy, \\(ACC\\), as final scores." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.464, + 0.608, + 0.479 + ], + "angle": 0, + "content": "4.2. Models" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.488, + 0.907, + 0.655 + ], + "angle": 0, + "content": "We accessed three proprietary models, including GPT-4o [39], Gemini-1.5-pro [44] and Claude-3.7-sonnet [2], along with 16 widely used open-source VLMs accepting multiple images simultaneously. Among the open-source models, four have more than 70B parameters, while the others range in size from 2B to 13B. We also enable the evaluation for powerful single-image models, such as LLaVA-1.6 [31] and LLaMA-3.2 [11], for further investigation. For fair comparison, we design a single-image pipeline for some our designed tasks and conduct evaluations on both single-image and multi-image models in Section 5.3." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.667, + 0.608, + 0.682 + ], + "angle": 0, + "content": "4.3. Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.69, + 0.909, + 0.903 + ], + "angle": 0, + "content": "The overall results on our benchmark is reported in Table 2. We derive three main findings from this table: 1. The performance of proprietary models and \\(>70B\\) open-source models largely varies. As expected, most proprietary models and \\(>70B\\) open-source models marginally outperform the evaluated open-source models. However, there are significant dependencies among these models, where Claude-3.5-sonnet achieves an average score of \\(49.37\\%\\), InternVL2_5-78B achieves \\(45.50\\%\\), while LLaVA-OneVision-72B only presents an average score of \\(29.17\\%\\), even lower than more than half of the \\(<13B\\) open-source models. 2. Proprietary models do not equally perform the best across all the tasks. Proprietary models fail to beat open-source models on tasks including DIR-Obj and CNT," + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "29574" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.093, + 0.089, + 0.487, + 0.346 + ], + "angle": 0, + "content": "
ModelsDIR-RecDIR-ObjCNTPLA-QAPLA-Dec
Proprietary Models
Claude-3.7-sonnet36.3631.1737.2044.4044.50
Gemini-1.5-pro37.2037.2027.5038.0336.15
GPT-4o37.2041.2039.0045.0739.91
>70B Open-source Models
InternVL2_5-78B29.6052.8044.0034.27*36.15
Qwen2-VL-72B30.4047.6046.00*37.0933.33
InternVL2-76B25.2038.8052.00*25.3525.35
LLaVA-OneVision-72B20.40*43.6022.5036.62*32.39*
<13B Open-source Models
MiniCPM-V 2.622.8040.0037.0039.4426.76
Qwen2-VL-7B19.6035.2040.5034.2723.00
Mantis-8B34.80*30.8042.50*31.9229.11*
InternVL2-8B31.20*36.4042.50*22.5426.76
VILA1.5-8B33.6029.20*42.0019.7239.91
Idefics3-8B37.60*37.2032.5023.9424.88
LLaVA-OneVision-7B17.2038.00*35.0030.05*27.70*
Phi-3.5-vision26.00*34.40*36.5031.4624.41
Brote-IM-XXL34.40*29.6029.5029.58*17.84
LongVA-7B16.0031.6028.50*17.8428.64
Mono-InternVL-2B18.8033.60*15.50*25.3522.07
mPLUG-Owl3-7B26.40*29.2029.50*29.11*32.39*
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.09, + 0.35, + 0.483, + 0.392 + ], + "angle": 0, + "content": "Table 3. Results for disrupted image orders. \\( ACC_{q} \\) is reported for Direction and Counting tasks and \\( ACC \\) is reported for the Planning task. * indicates the performance is higher than that in Table 2." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.408, + 0.483, + 0.527 + ], + "angle": 0, + "content": "where \\(>70\\mathrm{B}\\) open-source models exhibit stronger performance. 3. The ROT-Ang task is quite challenging for open-source models. It is noticeable that for the ROT-Ang task, all open-source models achieve similar results near \\(50\\%\\), which are almost equal to random selection. Such result is caused by the fixed predictions over all different context and images input. We will elaborately investigate such phenomena in Section 5.4." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.53, + 0.484, + 0.622 + ], + "angle": 0, + "content": "We conduct manual evaluation on all instances to assess the reasonability and difficulty of CoSpace and results are provided in Appendix ??. Results show that human performance excels the best models across all tasks. Human annotators achieved an average \\( ACC_{q} \\) of \\( 86.25\\% \\), which is significantly higher than that of Claude-3.7-sonnet (59.56%)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.638, + 0.188, + 0.655 + ], + "angle": 0, + "content": "5. Analysis" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.664, + 0.31, + 0.681 + ], + "angle": 0, + "content": "5.1. Disrupted Image Order" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.688, + 0.483, + 0.838 + ], + "angle": 0, + "content": "Considering the spatial relationship between input images of our CoSpace, the order of input images can make significant difference on the understanding of the space, especially when the order being thoroughly disrupted. Based on this speculation, we randomly shuffled the input sequence, and provided the directional orientation only of the first image for the model, with directions of the rest images to be inferred by the model itself. We conducted evaluation under such setting on three tasks: Direction, Counting and Planning, and the results are illustrated in Table 3." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Surprisingly, the performance of \\(< 13\\mathrm{B}\\) models under disrupted image order was not always lower than the basic setting. In fact, half of the \\(< 13\\mathrm{B}\\) open-source models achieved a higher score in the DIR-Rec task under disrupted set" + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.09, + 0.912, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.248, + 0.907, + 0.304 + ], + "angle": 0, + "content": "Figure 4. Variance in \\( ACC_{q} \\) of the average on the Direction and Counting categories for eight selected models. Model abbreviations are used for simplicity. Accuracy in the figure represents the average of \\( ACC_{q} \\) over two categories." + }, + { + "type": "table", + "bbox": [ + 0.52, + 0.311, + 0.898, + 0.357 + ], + "angle": 0, + "content": "
Img OrderClaude GeminiQwen2InternVL2OneVisionPhi-3.5LongVA
Sequential79.5873.6469.9134.2534.7833.3321.6236.59
Disrupted60.7156.7956.2524.6639.5835.9426.8731.58
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.513, + 0.36, + 0.907, + 0.402 + ], + "angle": 0, + "content": "Table 4. Results of conditional accuracy on the Planning category for eight selected models. Model names are abbreviated. Open-source models all refer to \\(<13\\mathrm{B}\\) ones." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.416, + 0.907, + 0.675 + ], + "angle": 0, + "content": "ting. Additionally, for other tasks, an average of three to five open-source models performed slightly better with disrupted image order. Nevertheless, similar phenomenon was not observed for proprietary models and most of the \\(>70\\mathrm{B}\\) open-source models, except LLaVA-OneVision-72B. Also, the evaluated proprietary models similarly underperformed across all tasks under disrupted setting, which aligns more closely with our expectations. For \\(>70\\mathrm{B}\\) open-source models, three of four outperformed on only one task against basic setting, with LLaVA-OneVision-72B producing a completely opposite result. Considering the such phenomenon, we conclude that the results for \\(<13\\mathrm{B}\\) open-source models are partly due to the instability of predictions. Moreover, we acknowledge that the impact of disrupting the image order may not be as significant for models as it is for human, exhibiting the discrepancy between models and human when perceiving continuous visual space via multiple images." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.683, + 0.747, + 0.699 + ], + "angle": 0, + "content": "5.2. Performance Consistency" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.704, + 0.907, + 0.871 + ], + "angle": 0, + "content": "We mentioned in Section 4.1 that for the Direction and Counting categories, both \\( ACC_{q} \\) and \\( ACC_{p} \\) are adopted as evaluation metrics. We present the gap between \\( ACC_{q} \\) and \\( ACC_{p} \\), i.e. variance in accuracy, in Figure 4. A smaller variance indicates that a model responds more consistently to input changes, reflecting better performance consistency. As shown, all models exhibit relatively large variance, with proprietary models (Claude-3.7-sonnet, Gemini-1.5-pro, GPT-4o) displaying even higher variances than open-source models, suggesting room for improvement. Such result highlights the importance of introducing \\( ACC_{p} \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.906, + 0.901 + ], + "angle": 0, + "content": "To further analyze the performance consistency of models, we introduce the metric conditional accuracy in the" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "29575" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.099, + 0.09, + 0.475, + 0.199 + ], + "angle": 0, + "content": "
ModelsDIR-RecDIR-ObjCNTPLA-QAPLA-DecAVG
Models with only one image in the input
LLaVA-v1.627.6032.0021.0021.6025.8225.60
LLama-3.2-11B29.2028.8028.5017.8416.9024.25
Models with multi-image perception ability
InternVL2-8B36.40*34.8027.0033.80*31.92*32.78
MiniCPM-V 2.623.2027.2026.0031.9232.39*28.14
Phi-3.5-vision18.4036.00*28.0031.4623.0027.37
LLaVA-OneVision26.80*32.8016.0022.0724.4124.42
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.201, + 0.483, + 0.243 + ], + "angle": 0, + "content": "Table 5. Results for single image pipeline. We report \\( ACC_{q} \\) for Direction and Counting tasks and \\( ACC \\) for Planning tasks. * indicates the performance is higher than that in Table 2." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.259, + 0.483, + 0.426 + ], + "angle": 0, + "content": "Planning category. Conditional accuracy measures how often models successfully solve the PLA-Dec question given that they answer the correspondent PLA-QA question correctly, as in PLA-QA task and PLA-Dec task, two respective questions regarding the same set of images typically feature the same object or direction as target. Results in Tabel 4 reveal that open-source models exhibit a greater gap in conditional accuracy against proprietary models than on the Direction and Counting categories. While proprietary models make more consistent decisions based on correct detection, the gap remains non-negligible." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.426, + 0.484, + 0.486 + ], + "angle": 0, + "content": "Above all, we argue that maintaining performance consistency is rather important for the further improvement of models, as it is closely related to robustness and rather essential for more complicated tasks." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.499, + 0.295, + 0.515 + ], + "angle": 0, + "content": "5.3. Single Image Pipeline" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.522, + 0.483, + 0.763 + ], + "angle": 0, + "content": "We presented in Section 3.1 that the understanding of continuous visual space inherently relies on multiple images as input to represent the whole space. Also, recent work [54] explored to construct a pipeline that can utilize caption and memory to transform the multi-image task into single image format. Similarly, we developed a single image pipeline for the Direction, Counting and Planning categories in our benchmark. For the Rotation category, as the single image input will inevitably compromise the uniform understanding of continuous visual space and the relationship between images can hardly be replaced by mere captions, we consider it not suitable for the single image pipeline. For evaluation, we selected LLaVA-v1.6 [31] and Llama-3.2-Vision [11] as the representatives of single image models. We also selected some multi-image models and conducted the single image evaluation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Results of the single image pipeline are reported in Table 5. Overall, the multi-image models underperformed under this setting compared with the multi-image pipeline. However, in the Direction-Rec and Planning-Dec task, two of four models achieve a higher accuracy, which is far beyond our expectations. We conclude that with the effective summarization of captions, models can still have an understanding of visual clues and sometimes the complete of all images can even add to the performance. Also," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.198 + ], + "angle": 0, + "content": "three of the four multi-image models outperform the single-image models on the average accuracy, indicating that models with inherent multi-image capabilities boast an overall better understanding of the visual clues and information extracted from the effectively multi-image scene. Based on this observation, the importance of empowering models with multi-image capabilities needs again to be emphasized." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.211, + 0.866, + 0.227 + ], + "angle": 0, + "content": "5.4. Investigation on Angle-related Reasoning" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.234, + 0.905, + 0.34 + ], + "angle": 0, + "content": "Table 2 shows that all the assessed open-source models achieve near-half ACCs on the Rotation-Angle (ROT-Ang) task. After checking the responses of different models, we found out that these models tend to produce the same answer across all the queries, e.g. \"A\" for all the 300 questions in this task. Therefore, due to the even distribution of options, these models achieve the similar near-half ACC." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.342, + 0.906, + 0.568 + ], + "angle": 0, + "content": "To discover the reason behind such phenomenon, we asked models to generate the rationale when answering the questions in the ROT-Ang task and observed some of these models can follow the instruction to generate rationales. For example, InternVL2-8B no longer always selected the same option and consequently correctly answered some question that it originally failed. However, we also noticed that it tended to identify the whole surrounding covered by the images, thus perceiving the sum of turning angles as 360 degrees. Therefore, it frequently output 90 degrees as the final answer, even if 90 degrees not included in the options. In contrast, proprietary models like Claude-3.7-sonnet could firstly calculate the sum of covered angles across all images based on the two give options and then decided which answer better matched the scene in given images." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.57, + 0.906, + 0.66 + ], + "angle": 0, + "content": "Though all the assessed open-source models fail the Rotation-Angle task, we still need to underline the importance of this task and the related capability to correctly identify the rotation in the space. Such capability lays the foundation for more advanced tasks in practical use and has already been boasted by some of the proprietary models." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.678, + 0.634, + 0.694 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Recognizing the gap between previous work and more advanced spatial understanding, we propose the concept of continuous space perception and present a novel benchmark CoSpace that comprises four categories and seven tasks. CoSpace covers both fundamental spatial comprehending abilities and more practical and commonly-used capabilities to comprehensively assess continuous space perception ability for VLMs. We evaluated three widely used proprietary models and sixteen most advanced open-source models and conducted detailed analysis for noticeable results in the experiments. Results reveal a gap between the assessed open-source and proprietary models, especially for performance consistency and in rotation-based perception." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "29576" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.251, + 0.108 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.114, + 0.484, + 0.227 + ], + "angle": 0, + "content": "This work is supported by the National Key R&D Program of China (2022ZD0160502), the National Natural Science Foundation of China (No. 62276152), and funding from Wuxi Research Institute of Applied Technologies, Tsinghua University under Grant 20242001120. We acknowledge Yan Liu, Jidong Chen, Jianyang Liu, Shuangyu Li and Shu Chi for their effort during the process of data annotation and human evaluation." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.252, + 0.188, + 0.267 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.277, + 0.483, + 0.304 + ], + "angle": 0, + "content": "[1] Phi-3 technical report: A highly capable language model locally on your phone, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.306, + 0.483, + 0.331 + ], + "angle": 0, + "content": "[2] Anthropic. The Claude 3 Model Family: Opus, Sonnet, Haiku. 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.333, + 0.484, + 0.428 + ], + "angle": 0, + "content": "[3] Xu Cao, Tong Zhou, Yunsheng Ma, Wenqian Ye, Can Cui, Kun Tang, Zhipeng Cao, Kaizhao Liang, Ziran Wang, James M Rehg, et al. MAPLM: a real-world large-scale vision-language benchmark for map and traffic scene understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21819-21830, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.43, + 0.483, + 0.512 + ], + "angle": 0, + "content": "[4] Boyuan Chen, Zhuo Xu, Sean Kirmani, Brain Ichter, Dorsa Sadigh, Leonidas Guibas, and Fei Xia. SpatialVLM: endowing vision-language models with spatial reasoning capabilities. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14455-14465, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.513, + 0.483, + 0.594 + ], + "angle": 0, + "content": "[5] Liang Chen, Yichi Zhang, Shuhuai Ren, Haozhe Zhao, Zefan Cai, Yuchi Wang, Peiyi Wang, Tianyu Liu, and Baobao Chang. Towards end-to-end embodied decision making via multi-modal large language model: Explorations with gpt4- vision and beyond. arXiv preprint arXiv:2310.02071, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.597, + 0.483, + 0.664 + ], + "angle": 0, + "content": "[6] Zhe Chen, Weiyun Wang, Yue Cao, Yangzhou Liu, Zhangwei Gao, Erfei Cui, Jinguo Zhu, Shenglong Ye, Hao Tian, Zhaoyang Liu, et al. Expanding performance boundaries of open-source multimodal models with model, data, and test-time scaling. arXiv preprint arXiv:2412.05271, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.666, + 0.483, + 0.734 + ], + "angle": 0, + "content": "[7] Zhe Chen, Weiyun Wang, Hao Tian, Shenglong Ye, Zhangwei Gao, Erfei Cui, Wenwen Tong, Kongzhi Hu, Jiapeng Luo, Zheng Ma, et al. How far are we to GPT-4V? closing the gap to commercial multimodal models with open-source suites. ArXiv preprint, abs/2404.16821, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.735, + 0.483, + 0.817 + ], + "angle": 0, + "content": "[8] Hao-Tien Lewis Chiang, Zhuo Xu, Zipeng Fu, Mithun George Jacob, Tingnan Zhang, Tsang-Wei Edward Lee, Wenhao Yu, Connor Schenck, David Rendleman, Dhruv Shah, et al. Mobility VLA: multimodal instruction navigation with long-context vlms and topological graphs. arXiv preprint arXiv:2407.07775, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.818, + 0.483, + 0.873 + ], + "angle": 0, + "content": "[9] Song Dingjie, Shunian Chen, Guiming Hardy Chen, Fei Yu, Xiang Wan, and Benyou Wang. Milebench: Benchmarking MLLMs in long context. In First Conference on Language Modeling, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.874, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[10] Sivan Doveh, Shaked Perek, M Jehanzeb Mirza, Wei Lin, Amit Alfassy, Assaf Arbelle, Shimon Ullman, and Leonid" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.277, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.133 + ], + "angle": 0, + "content": "Karlinsky. Towards multimodal in-context learning for vision & language models. arXiv preprint arXiv:2403.12736, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[11] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, and et al. The llama 3 herd of models, 2024. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.193, + 0.905, + 0.261 + ], + "angle": 0, + "content": "[12] Chaoyou Fu, Peixian Chen, Yunhang Shen, Yulei Qin, Mengdan Zhang, Xu Lin, Jinrui Yang, Xiawu Zheng, Ke Li, Xing Sun, et al. Mme: A comprehensive evaluation benchmark for multimodal large language models. arXiv preprint arXiv:2306.13394, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.264, + 0.905, + 0.332 + ], + "angle": 0, + "content": "[13] Chaoyou Fu, Yuhan Dai, Yondong Luo, Lei Li, Shuhuai Ren, Renrui Zhang, Zihan Wang, Chenyu Zhou, Yunhang Shen, Mengdan Zhang, et al. Video-MME: the first-ever comprehensive evaluation benchmark of multi-modal llms in video analysis. ArXiv preprint, abs/2405.21075, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.334, + 0.905, + 0.417 + ], + "angle": 0, + "content": "[14] Xingyu Fu, Yushi Hu, Bangzheng Li, Yu Feng, Haoyu Wang, Xudong Lin, Dan Roth, Noah A. Smith, Wei-Chiu Ma, and Ranjay Krishna. Blink: Multimodal large language models can see butA not perceive. In Computer Vision - ECCV 2024, pages 148-166, Cham, 2025. Springer Nature Switzerland. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.42, + 0.905, + 0.488 + ], + "angle": 0, + "content": "[15] Yupan Huang, Zaiqiao Meng, Fangyu Liu, Yixuan Su, Nigel Collier, and Yutong Lu. Sparkles: Unlocking chats across multiple images for multimodal instruction-following models. In ICLR 2024 Workshop on Navigating and Addressing Data Problems for Foundation Models, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.49, + 0.905, + 0.558 + ], + "angle": 0, + "content": "[16] Baoxiong Jia, Yixin Chen, Huangyue Yu, Yan Wang, Xuesong Niu, Tengyu Liu, Qing Li, and Siyuan Huang. Sceneverse: Scaling 3d vision-language learning for grounded scene understanding. arXiv preprint arXiv:2401.09340, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.561, + 0.905, + 0.615 + ], + "angle": 0, + "content": "[17] Mengzhao Jia, Wenhao Yu, Kaixin Ma, Tianqing Fang, Zhihan Zhang, Siru Ouyang, Hongming Zhang, Meng Jiang, and Dong Yu. Leopard: A Vision Language Model For Text-Rich Multi-Image Tasks, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.618, + 0.905, + 0.671 + ], + "angle": 0, + "content": "[18] Dongfu Jiang, Xuan He, Huaye Zeng, Cong Wei, Max Ku, Qian Liu, and Wenhu Chen. Mantis: interleaved multi-image instruction tuning. ArXiv preprint, abs/2405.01483, 2024. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.675, + 0.905, + 0.743 + ], + "angle": 0, + "content": "[19] Mukul Khanna*, Ram Ramrakhya*, Gunjan Chhablani, Sriram Yenamandra, Theophile Gervet, Matthew Chang, Zsolt Kira, Devendra Singh Chaplot, Dhruv Batra, and Roozbeh Mottaghi. Goat-bench: A benchmark for multi-modal life-long navigation. In CVPR, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.746, + 0.905, + 0.8 + ], + "angle": 0, + "content": "[20] Taewoong Kim, Cheolhong Min, Byeonghwi Kim, Jinyeon Kim, Wonje Jeung, and Jonghyun Choi. Realfred: An embodied instruction following benchmark in photo-realistic environment. In ECCV, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.802, + 0.905, + 0.856 + ], + "angle": 0, + "content": "[21] Hugo Laurençon, Andres Marafioti, Victor Sanh, and Léo Tronchon. Building and better understanding vision-language models: insights and future directions, 2024. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.859, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[22] Bo Li, Yuanhan Zhang, Liangyu Chen, Jinghao Wang, Jingkang Yang, and Ziwei Liu. Otter: A Multi-Modal Model with In-Context Instruction Tuning, 2023. 1, 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "29577" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.161 + ], + "angle": 0, + "content": "[23] Bohao Li, Yuying Ge, Yixiao Ge, Guangzhi Wang, Rui Wang, Ruimao Zhang, and Ying Shan. SEED-Bench: Benchmarking multimodal large language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13299-13308, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.162, + 0.482, + 0.218 + ], + "angle": 0, + "content": "[24] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. LLaVA-OneVision: Easy visual task transfer. ArXiv preprint, abs/2408.03326, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.218, + 0.482, + 0.272 + ], + "angle": 0, + "content": "[25] Feng Li, Renrui Zhang, Hao Zhang, Yuanhan Zhang, Bo Li, Wei Li, Zejun Ma, and Chunyuan Li. Llava-last-interleave: Tackling multi-image, video, and 3d in large multimodal models, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.273, + 0.482, + 0.356 + ], + "angle": 0, + "content": "[26] Juncheng Li, Kaihang Pan, Zhiqi Ge, Minghe Gao, Wei Ji, Wenqiao Zhang, Tat-Seng Chua, Siliang Tang, Hanwang Zhang, and Yueting Zhuang. Fine-tuning multimodal LLMs to follow zero-shot demonstrative instructions. In The Twelfth International Conference on Learning Representations, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.357, + 0.482, + 0.439 + ], + "angle": 0, + "content": "[27] Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Yi Liu, Zun Wang, Jilan Xu, Guo Chen, Ping Luo, et al. Mvbench: A comprehensive multi-modal video understanding benchmark. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22195-22206, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.441, + 0.482, + 0.523 + ], + "angle": 0, + "content": "[28] Xiaoqi Li, Mingxu Zhang, Yiran Geng, Haoran Geng, Yuxing Long, Yan Shen, Renrui Zhang, Jiaming Liu, and Hao Dong. ManipLLM: Embodied multimodal large language model for object-centric robotic manipulation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18061-18070, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.524, + 0.482, + 0.593 + ], + "angle": 0, + "content": "[29] Ji Lin, Hongxu Yin, Wei Ping, Pavlo Molchanov, Mohammad Shoeybi, and Song Han. Vila: On pre-training for visual language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26689-26699, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.594, + 0.482, + 0.663 + ], + "angle": 0, + "content": "[30] Ziyi Lin, Chris Liu, Renrui Zhang, Peng Gao, Longtian Qiu, Han Xiao, Han Qiu, Chen Lin, Wenqi Shao, Keqin Chen, et al. Sphinx: The joint mixing of weights, tasks, and visual embeddings for multi-modal large language models. arXiv preprint arXiv:2311.07575, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.664, + 0.482, + 0.691 + ], + "angle": 0, + "content": "[31] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning, 2023. 2, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.692, + 0.482, + 0.734 + ], + "angle": 0, + "content": "[32] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. LLaVA-NeXT: Improved reasoning, OCR, and world knowledge, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.735, + 0.482, + 0.789 + ], + "angle": 0, + "content": "[33] Haowei Liu, Xi Zhang, Haiyang Xu, Yaya Shi, Chaoya Jiang, Ming Yan, Ji Zhang, Fei Huang, Chunfeng Yuan, Bing Li, and Weiming Hu. MIBench: Evaluating multimodal large language models over multiple images, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.79, + 0.482, + 0.872 + ], + "angle": 0, + "content": "[34] Fuwen Luo, Chi Chen, Zihao Wan, Zhaolu Kang, Qidong Yan, Yingjie Li, Xiaolong Wang, Siyu Wang, Ziyue Wang, Xiaoyue Mi, Peng Li, Ning Ma, Maosong Sun, and Yang Liu. Codis: Benchmarking context-dependent visual comprehension for multimodal large language models. arXiv preprint arXiv:2402.13607, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[35] Gen Luo, Xue Yang, Wenhan Dou, Zhaokai Wang, Jifeng Dai, Yu Qiao, and Xizhou Zhu. Mono-internvl: Pushing the" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.092, + 0.905, + 0.121 + ], + "angle": 0, + "content": "boundaries of monolithic multimodal large language models with endogenous visual pre-training, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.217 + ], + "angle": 0, + "content": "[36] Arjun Majumdar, Anurag Ajay, Xiaohan Zhang, Pranav Putta, Sriram Yenamandra, Mikael Henaff, Sneha Silwal, Paul Mcvay, Oleksandr Maksymets, Sergio Arnaud, et al. OpenEQA: embodied question answering in the era of foundation models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16488-16498, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.218, + 0.905, + 0.288 + ], + "angle": 0, + "content": "[37] Fanqing Meng, Jin Wang, Chuanhao Li, Quanfeng Lu, Hao Tian, Jiaqi Liao, Xizhou Zhu, Jifeng Dai, Yu Qiao, Ping Luo, Kaipeng Zhang, and Wenqi Shao. MMIU: Multimodal multi-image understanding for evaluating large vision-language models, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.289, + 0.905, + 0.343 + ], + "angle": 0, + "content": "[38] Jiahao Nie, Gongjie Zhang, Wenbin An, Yap-Peng Tan, Alex C Kot, and Shijian Lu. Mmrel: A relation understanding dataset and benchmark in the mllm era. arXiv preprint arXiv:2406.09121, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.344, + 0.905, + 0.371 + ], + "angle": 0, + "content": "[39] OpenAI. Hello GPT-4o, 2024. Accessed on June 16, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.372, + 0.905, + 0.51 + ], + "angle": 0, + "content": "[40] Xavier Puig, Eric Undersander, Andrew Szot, Mikael Dallaire Cote, Tsung-Yen Yang, Ruslan Partsey, Ruta Desai, Alexander Clegg, Michal Hlavac, So Yeon Min, Vladimir Vondruš, Theophile Gervet, Vincent-Pierre Berges, John M Turner, Oleksandr Maksymets, Zsolt Kira, Mrinal Kalakrishnan, Jitendra Malik, Devendra Singh Chaplot, Unnat Jain, Dhruv Batra, Akshara Rai, and Roozbeh Mottaghi. Habitat 3.0: A co-habitat for humans, avatars, and robots. In Proceedings of the Twelfth International Conference on Learning Representations (ICLR 2024), 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.511, + 0.905, + 0.552 + ], + "angle": 0, + "content": "[41] Navid Rajabi and Jana Kosecka. Gsr-bench: A benchmark for grounded spatial reasoning evaluation via multimodal llms. arXiv preprint arXiv:2406.13246, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.553, + 0.905, + 0.622 + ], + "angle": 0, + "content": "[42] Abhinav Rajvanshi, Karan Sikka, Xiao Lin, Bhoram Lee, Han pang Chiu, and Alvaro Velasquez. SayNav: Grounding large language models for dynamic planning to navigation in new environments. In 34th International Conference on Automated Planning and Scheduling, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.622, + 0.905, + 0.733 + ], + "angle": 0, + "content": "[43] Santhosh Kumar Ramakrishnan, Aaron Gokaslan, Erik Wijmans, Oleksandr Maksymets, Alexander Clegg, John M Turner, Eric Undersander, Wojciech Galuba, Andrew Westbury, Angel X Chang, Manolis Savva, Yili Zhao, and Dhruv Batra. Habitat-matterport 3d dataset (HM3d): 1000 large-scale 3d environments for embodied AI. In Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 2), 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.734, + 0.905, + 0.815 + ], + "angle": 0, + "content": "[44] Machel Reid, Nikolay Savinov, Denis Teplyashin, Dmitry Lepikhin, Timothy Lillicrap, Jean-baptiste Alayrac, Radu Soricut, Angeliki Lazaridou, Orhan First, Julian Schrittwieser, et al. Gemini 1.5: unlocking multimodal understanding across millions of tokens of context. ArXiv preprint, abs/2403.05530, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.817, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[45] Allen Z. Ren, Jaden Clark, Anushri Dixit, Masha Itkina, Anirudha Majumdar, and Dorsa Sadigh. Explore until confident: Efficient exploration for embodied question answering, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.872, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[46] Manolis Savva, Abhishek Kadian, Oleksandr Maksymets, Yili Zhao, Erik Wijmans, Bhavana Jain, Julian Straub, Jia" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "29578" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.484, + 0.147 + ], + "angle": 0, + "content": "Liu, Vladlen Koltun, Jitendra Malik, Devi Parikh, and Dhruv Batra. Habitat: A Platform for Embodied AI Research. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.484, + 0.219 + ], + "angle": 0, + "content": "[47] Fatemeh Shiri, Xiao-Yu Guo, Mona Far, Xin Yu, Reza Haf, and Yuan-Fang Li. An empirical analysis on spatial reasoning capabilities of large multimodal models. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 21440-21455, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.219, + 0.483, + 0.342 + ], + "angle": 0, + "content": "[48] Andrew Szot, Alex Clegg, Eric Undersander, Erik Wijmans, Yili Zhao, John Turner, Noah Maestre, Mustafa Mukadam, Devendra Chaplot, Oleksandr Maksymets, Aaron Gokaslan, Vladimir Vondrus, Sameer Dharur, Franziska Meier, Wojciech Galuba, Angel Chang, Zsolt Kira, Vladlen Koltun, Jitendra Malik, Manolis Savva, and Dhruv Batra. Habitat 2.0: Training home assistants to rearrange their habitat. In Advances in Neural Information Processing Systems (NeurIPS), 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.344, + 0.483, + 0.411 + ], + "angle": 0, + "content": "[49] Yunlong Tang, Jing Bi, Siting Xu, Luchuan Song, Susan Liang, Teng Wang, Daoan Zhang, Jie An, Jingyang Lin, Rongyi Zhu, et al. Video understanding with large language models: A survey. arXiv preprint arXiv:2312.17432, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.413, + 0.483, + 0.482 + ], + "angle": 0, + "content": "[50] Jiayu Wang, Yifei Ming, Zhenmei Shi, Vibhav Vineet, Xin Wang, Yixuan Li, and Neel Joshi. Is a picture worth a thousand words? delving into spatial reasoning for vision language models. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.484, + 0.483, + 0.567 + ], + "angle": 0, + "content": "[51] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Yang Fan, Kai Dang, Mengfei Du, Xuancheng Ren, Rui Men, Dayiheng Liu, Chang Zhou, Jingren Zhou, and Junyang Lin. Qwen2-VL: Enhancing vision-language model's perception of the world at any resolution, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.568, + 0.483, + 0.621 + ], + "angle": 0, + "content": "[52] Xidong Wang, Dingjie Song, Shunian Chen, Chen Zhang, and Benyou Wang. LongLLaVA: Scaling Multi-modal LLMs to 1000 Images Efficiently via a Hybrid Architecture, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.623, + 0.483, + 0.747 + ], + "angle": 0, + "content": "[53] Xiyao Wang, Yuhang Zhou, Xiaoyu Liu, Hongjin Lu, Yuancheng Xu, Feihong He, Jaehong Yoon, Taixi Lu, Fuxiao Liu, Gedas Bertasius, Mohit Bansal, Huaxiu Yao, and Furong Huang. Mementos: A comprehensive benchmark for multimodal large language model reasoning over image sequences. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 416-442, Bangkok, Thailand, 2024. Association for Computational Linguistics. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.748, + 0.483, + 0.816 + ], + "angle": 0, + "content": "[54] Ziyue Wang, Chi Chen, Fuwen Luo, Yurui Dong, Yuanchi Zhang, Yuzhuang Xu, Xiaolong Wang, Peng Li, and Yang Liu. ActiView: Evaluating active perception ability for multimodal large language models. arXiv preprint arXiv:2410.04659, 2024. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.818, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[55] Ziyue Wang, Chi Chen, Yiqi Zhu, Fuwen Luo, Peng Li, Ming Yan, Ji Zhang, Fei Huang, Maosong Sun, and Yang Liu. Browse and Concentrate: comprehending multimodal content via prior-llm context fusion. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 11229–11245," + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "Bangkok, Thailand, 2024. Association for Computational Linguistics. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.163 + ], + "angle": 0, + "content": "[56] Zhutian Yang, Caelan Garrett, Dieter Fox, Tomás Lozano-Pérez, and Leslie Pack Kaelbling. Guiding long-horizon task and motion planning with vision language models, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.165, + 0.905, + 0.22 + ], + "angle": 0, + "content": "[57] Yuan Yao, Tianyu Yu, Ao Zhang, Chongyi Wang, Junbo Cui, Hongji Zhu, Tianchi Cai, Haoyu Li, Weilin Zhao, Zhihui He, et al. MiniCPM-V: A GPT-4V level MLLM on your phone. ArXiv preprint, abs/2408.01800, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.222, + 0.905, + 0.289 + ], + "angle": 0, + "content": "[58] Jiabo Ye, Haiyang Xu, Haowei Liu, Anwen Hu, Ming Yan, Qi Qian, Ji Zhang, Fei Huang, and Jingren Zhou. mPLUG-Owl3: towards long image-sequence understanding in multi-modal large language models. ArXiv preprint, abs/2408.04840, 2024. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.291, + 0.905, + 0.402 + ], + "angle": 0, + "content": "[59] Haotian Zhang, Mingfei Gao, Zhe Gan, Philipp Duffer, Nina Wenzel, Forrest Huang, Dhruti Shah, Xianzhi Du, Bowen Zhang, Yanghao Li, Sam Dodge, Keen You, Zhen Yang, Aleksei Timofeev, Mingze Xu, Hong-You Chen, JeanPhilippe Fauconnier, Zhengfeng Lai, Haoxuan You, Zirui Wang, Afshin Dehghan, Peter Grasch, and Yinfei Yang. MM1.5: Methods, analysis & insights from multimodal llm fine-tuning, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.404, + 0.905, + 0.471 + ], + "angle": 0, + "content": "[60] Peiyuan Zhang, Kaichen Zhang, Bo Li, Guangtao Zeng, Jingkang Yang, Yuanhan Zhang, Ziyue Wang, Haoran Tan, Chunyuan Li, and Ziwei Liu. Long context transfer from language to vision. arXiv preprint arXiv:2406.16852, 2024. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.474, + 0.905, + 0.556 + ], + "angle": 0, + "content": "[61] Yi-Fan Zhang, Huanyu Zhang, Haochen Tian, Chaoyou Fu, Shuangqing Zhang, Junfei Wu, Feng Li, Kun Wang, Qing-song Wen, Zhang Zhang, Liang Wang, Rong Jin, and Tieniu Tan. Mme-realworld: Could your multimodal llm challenge high-resolution real-world scenarios that are difficult for humans?, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.558, + 0.905, + 0.64 + ], + "angle": 0, + "content": "[62] Haozhe Zhao, Zefan Cai, Shuzheng Si, Xiaojian Ma, Kaikai An, Liang Chen, Zixuan Liu, Sheng Wang, Wenjuan Han, and Baobao Chang. MMICL: Empowering vision-language model with multi-modal in-context learning. In The Twelfth International Conference on Learning Representations, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.642, + 0.905, + 0.711 + ], + "angle": 0, + "content": "[63] Gengze Zhou, Yicong Hong, Zun Wang, Xin Eric Wang, and Qi Wu. NavGPT-2: Unleashing navigational reasoning capability for large vision-language models. In European Conference on Computer Vision, pages 260–278. Springer, 2025. 1, 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.711 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "29579" + } + ] +] \ No newline at end of file diff --git a/2025/CoSpace_ Benchmarking Continuous Space Perception Ability for Vision-Language Models/45f50ce2-59a4-414e-86e0-abbb59b51daa_origin.pdf b/2025/CoSpace_ Benchmarking Continuous Space Perception Ability for Vision-Language Models/45f50ce2-59a4-414e-86e0-abbb59b51daa_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..eca3cffeb2a9f7603c34bd93406645b5af189e5b --- /dev/null +++ b/2025/CoSpace_ Benchmarking Continuous Space Perception Ability for Vision-Language Models/45f50ce2-59a4-414e-86e0-abbb59b51daa_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:322485f8b3b1598f6ed8a28caddd6544ac08adeb0c9fc74b524340f1dbf1f0c1 +size 5350931 diff --git a/2025/CoSpace_ Benchmarking Continuous Space Perception Ability for Vision-Language Models/full.md b/2025/CoSpace_ Benchmarking Continuous Space Perception Ability for Vision-Language Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..6644dcac8c89cbef5d129345bbb80a114e0a37ee --- /dev/null +++ b/2025/CoSpace_ Benchmarking Continuous Space Perception Ability for Vision-Language Models/full.md @@ -0,0 +1,468 @@ +# CoSpace: Benchmarking Continuous Space Perception Ability for Vision-Language Models + +Yiqi Zhu $^{1*}$ , Ziyue Wang $^{1*}$ , Can Zhang $^{3}$ , Peng Li $^{2,4\dagger}$ , Yang Liu $^{1,2,4,5\dagger}$ + +$^{1}$ Dept. of Comp. Sci. & Tech., Institute for AI, Tsinghua University, Beijing, China + +$^{2}$ Institute for AI Industry Research (AIR), Tsinghua University, Beijing, China + +$^{3}$ School of Computer and Communication Engineering, University of Science and Technology Beijing + +$^{4}$ Shanghai Artificial Intelligence Laboratory, Shanghai, China + +$^{5}$ Jiangsu Collaborative Innovation Center for Language Competence, Jiangsu, China + +![](images/e7ef0ad57d49aabffbf8bb5e3d9aaee49e8be2cfeddc8d8e0c6db3eb5d4b33f9.jpg) +The way models perceive real-world scenarios + +![](images/a67683c6cd9e489633cf9362ca9a8745fec18ceaadaee4d0016724819c9cef44.jpg) + +![](images/36e08331fc60a1a87e70a404552c46e77d1fb66fa7ac796d4fc66d460223305f.jpg) + +Each of the four images contain one car respectively. However 1 2 feature the same car, and 3 4 feature another identical car. Thus, there are only two cars in the scene. + +![](images/b5f684a0dc7fb59074ce74c40d6e4c7c784d80a08f7b552dcd1e0ff195ada2e0.jpg) +Figure 1. Illustration of continuous space perception: The left side depicts the construction of an image sequence representing a continuous space. The right side presents an example task related to the continuous space, showing how adjacent images are connected. Note that to answer the question correctly, one must recognize that the cars in images 1 and 2 are the same, as well as those in images 3 and 4. + +The four images respectively contains one car. Four cars all together are visible in the scene. + +# Abstract + +Vision-Language Models (VLMs) have recently witnessed significant progress in visual comprehension. As the permitting length of image context grows, VLMs can now comprehend a broader range of views and spaces. Current benchmarks provide insightful analysis of VLMs in tasks involving complex visual instructions following, multi-image understanding and spatial reasoning. However, they usually focus on spatially irrelevant images or discrete images captured from varied viewpoints. The compositional characteristic of images captured from a static viewpoint remains underestimated. We term this characteristic as Continuous Space Perception. When observing a scene from a static viewpoint while shifting orientations, it produces a series of spatially continuous images, enabling the reconstruction of the entire space. In this paper, we present CoSpace, a multi-image visual understanding benchmark designed to assess the Continuous Space perception ability for VLMs. CoSpace contains 2,918 images and 1,626 question-answer pairs, covering seven types of tasks. We conduct evaluation across 19 proprietary and open-source VLMs. Results reveal that there exist pitfalls on the continuous space perception ability for most of the evaluated models, including pro + +prietary ones. Interestingly, we find that the main discrepancy between open-source and proprietary models lies not in accuracy but in the consistency of responses. We believe that enhancing the ability of continuous space perception is essential for VLMs to perform effectively in real-world tasks and encourage further research to advance this capability. + +# 1. Introduction + +With the advancements in deep learning and Large Language Models (LLMs), Vision-Language Models (VLMs) have demonstrated exceptional performance across a range of downstream tasks [4, 17, 22, 30, 47, 60]. Driven by this, VLMs are now being used in diverse applications, such as planning [5, 56], navigation [8, 63] and embodied tasks [28, 42], and are further applied in real-world scenarios [3, 61]. Real-world scenario, featuring the complicated environment, cast significant challenges on VLMs, especially the ability to capture sophisticated spatial information from surrounding spaces. + +Despite previous attempts to advance spatial understanding in VLMs [4, 38], there remain areas requiring further exploration. Various strategies have been explored to empower models with the ability to identify and ground objects based on spatial relationship [4, 16]. Shiri et al. [47] + +
BenchmarksEvaluation TargetFeaturesNum.ImgEvaluationInstances
Spatial ReasoningMulti-image PerceptionStatic Viewpoint
MMRel [38]Object-relationship○××15k15k
Spatial-MM [47]Object-relationship and spatial-CoT✓××2.3k400
SpatialEval [50]Spatial Reasoning✓××4.6k4.6k
GSR-Bench [41]Grounded spatial reasoning✓××4.9k4.9k
MileBench [9]Multimodal Long Context×✓×97.9k6.4k
Demon-Core [26]Demonstrative instruction following×✓×62.8k18.2k
Mementos-Val [53]Temporal sequential-image understanding×✓×-699
MIBench [33]Fine-grained multi-image understanding×✓×125k13k
BLINK [14]General visual perception○○×7.3k3.8k
SEED-Bench [23]Visual comprehension○○×1.9k24k
Video-MME [13]Visual comprehension○✓×0.9k2.7k
MMIU [37]Visual comprehension○✓×77.7k11.7k
SparklesEval [15]Multi-image Chat○✓×550150
CoSpace (Ours)continuous space perception✓✓✓2,9181,626
+ +Table 1. Comparison with selected benchmarks that address multi-image evaluation and/or spatial evaluation. $\checkmark$ denotes the main focus of the corresponding feature, $\circ$ indicates that the feature exists but is not the primary focus, while $X$ implies a lack of the feature. + +and Wang et al. [50] assess spatial understanding ability for VLMs with the recognition of basic tasks in this domain. However, much of these work center on single-image inputs, without further extending to multi-image scenarios. Works on video understanding [52, 60] present the concept of spatial temporal understanding, but the spatial information of videos involve sequential frames that inherently linking spatial understanding with additional temporal dynamics. Given these attempts in spatial understanding, we pose the question: Are the current efforts adequate for advancing the research of spatial understanding? + +We claim that a gap still exists between the concept proposed in previous work and more advanced spatial understanding. For humans, exposing to an unfamiliar environment requires determining the position, identifying surrounding objects, and making decisions based on spatial context. Such process typically involves integrating various information about the surrounding space. We refer to it as continuous space perception, which is an essential ability in real-world towards effectively completion of daily tasks. + +Figure 1 shows an example of continuous space perception: Conceptually, humans perceive continuous visual space as a series of "images", each representing a segment of the space with overlaps between adjacent views. Mimicking this perception, Figure 1 illustrates the continuous visual space using four images. To answer questions accurately about these images requires a comprehensive understanding of all four images and reconstructing the entire space they represent. For example, to accurately count the number of cars in the space, it is essential to recognize that the cars appearing in the first and second images are the same, as are the cars in the third and fourth images. + +The continuous space perception ability is equally important for VLMs as for human. Many real-world scenario tasks, such as planning and navigation, require models to effectively integrate the information in the surroundings to take the optimal actions. During this process, the reconstruction of the space heavily relies on continuous space + +perception. However, as shown in Table 1, this ability is largely overlooked in existing benchmarks for VLMs. + +To this end, we present CoSpace, a visual understanding benchmark designed to evaluate the Continuous Space perception ability for VLMs. CoSpace includes four categories and seven tasks, emphasizing the following core abilities: (1) Can VLMs recognize direction and position when presented with spatially continuous images captured from a static viewpoint? (2) Can models identify the correspondence between adjacent images in the same scene? (3) Can models make optimal navigation decisions based on the perception of the space? We posit the importance of these three core abilities, with the first one forms the basic understanding of space, the second one is essential for integrating the information across different images and the last one surpasses mere literal understanding and makes a step towards more practical applications. + +We evaluated 19 widely used open-source and proprietary VLMs on our proposed CoSpace benchmark. Results show that though some of the open-source models achieve approaching or even higher accuracy compared with state-of-the-art proprietary models on certain tasks, proprietary models exhibit remarkably higher consistency in their responses across multi-dimensional evaluation metrics. Additionally, we found that with the alteration of input context and image sequence, the tasks that humans find more challenging do not always pose greater difficulty for the models. + +# 2. Related Work + +# 2.1. Multi-image Perception Ability of VLMs + +Large Vision-Language Models (VLMs), such as LLaVA models [31, 32] and Sphinx [30], have shown impressive performances on visual reasoning and grounding tasks that focus on only one image at a time. However, it is more natural to allow multiple images and arrange them in interleave format in one input. Along with the growing capability of large vision models, there are also many other applications. For example, the use of large vision models has been demonstrated for large datasets, such as the US Census dataset [33]. + +ities of single-image models, multi-image perception and understanding abilities are gaining increasing attentions, including research topics such as multimodal in-context learning [10, 22, 62], interleaved visual instruction following [18, 25, 26], and temporal visual understanding [49], and practical applications such as navigation and embodied QA [63]. To achieve multi-image perception, some methods focus on constructing interleaved visual instruction tuning data [17, 18, 22, 62], some improve the model structure and training strategies for better cross-image understanding [26, 52, 55, 58], and others address the image processing strategy for adaptable visual feature encoding concerning multi-image scenarios [21, 32, 59]. + +With the enhancements of the multi-image perception ability, VLMs are able to obtain visual information covering a wider range of views from multiple images at one time, resulting in more profound observations and understandings of the surrounding environment. In our paper, we take this as basis, and dedicate to investigating the spatial perception ability of VLMs. + +# 2.2. Spatial Comprehension Evaluation of VLMs + +For the past years, the evaluation of spatial comprehension for VLMs has always been a commonly engaged topic, and it often refers to the spatial relation recognition among objects, such as MMRel, GSR-Bench and SpatialMM shown in Table 1. Recently, multi-image related evaluation for VLMs are widely explored, including interleaved visual instruction following [26, 33], temporal sequential-image comprehension [53], and long context understanding [9]. In multi-image scenarios, such as video comprehension [13, 23] and multi-image dialogue [15], the spatial reasoning ability is also one of the evaluated aspects. However, these benchmarks primarily focus on general multi-image comprehension where most of input images come from difference scenes. The compositional features across multiple images to integrate richer spatial information of a wider scene and more profound spatial understanding are seldom emphasized. Although Blink [14], ActiView [54], and video-based benchmarks [13, 27] require spatial understanding across multiple images, they either focus on different views of the same object, or discrete views from dynamic viewpoint. These views are not continuous regarding a static viewpoint, and the resulting images are not necessarily bounded regarding the same scene. In contrast to existing works, we address the continuous spatial perception from a static viewpoint, that requires models to connect all the perceived views together to form an intact understanding of the whole scene. + +# 3. CoSpace + +# 3.1. Continuous Space Perception Ability + +In this paper, we define the term, continuous visual space, as the perceived views from a static viewpoint towards different orientations that can naturally compose a complete scene or environment. As is shown in Figure 1, if one would like to correctly answer the question, it is needed to first capture the visual information of all the cars, and consequently, the ability of continuous space perception refers to capturing and processing information from these images of continuous visual spaces. + +To obtain an overall observation of environment, human perceive the whole surroundings from the same viewpoint towards different directions, and the percept continuous visual spaces can be reconstructed as a panorama in the brain. Different from human that directly views the space continuously, models typically take discrete images as input. Therefore, the way to present continuous visual spaces remains flexible. In our CoSpace, we present a complete continuous visual space as follows: we provide several images shot at the same viewpoint in the same scene, but towards different directions, e.g. respectively towards front, right, back and left. These images together form a complete space without omitting any visual information of the surrounding environment. Naturally, such strategy could produce overlapped spaces between adjacent images, which tightly connects images describing adjacent spaces, and can consequently help models better identify the relationships between images from continuous visual space. In this way, we posit that such strategy can effectively preserve continuous spatial information of an extensive space as well as be suitable to understanding for VLMs. + +# 3.2. Task Design + +We comprehensively evaluate the continuous space perception ability of VLMs regarding four aspects. Two of them are fundamental ability of spatial perception, detailed as direction perception and recognition (Direction), and rotation-centric perception (Rotation). The other two are further practical tasks, detailed as counting-based perception (Counting), and embodied-based planning (Planning). These four categories are subdivided into seven tasks for elaborate investigation as shown in Figure 2. + +We employ two multiple choice tasks, both with four options, for the Direction category as follows: + +- Direction Recognition (DIR-Rec): In real world, identifying direction is inevitable when one is placed in a new environment. Similarly, the DIR-Rec task requires models to recognize the direction of targets by answering questions such as "where is the building located?" Meanwhile, we provide detailed task instructions, including direction references such as "the first image is facing north, + +# Direction + +$\clubsuit$ Direction Recognition (DIR-Rec) + +![](images/114f03be34d4468d3a10b547709cd68894dcdf6af81a77d3e849293e92ab478e.jpg) + +Question: Which direction is the alleyway going in the images? + +Options: A. South to North + +C. Northeast to Southwest + +B. East to West + +D. Northwest to Southeast + +Rationale: The $2^{nd}$ and $4^{th}$ images indicate a direction of East-West, the $1^{st}$ and $3^{rd}$ images indicate the width of the alleyway. + +Directional Object Perception (DIR-Obj) + +![](images/22af0283074d7c8bc2bc1b8252d1b9d164edb5bb734d732fdf9aa9071376d10f.jpg) + +Question: What is visible in the southeast direction? + +Options: A. A pedestrian bridge + +C. A "no parking" sign + +Rationale: Southeast refers to the left part of the $3^{rd}$ image and the right part of the $4^{th}$ image, only a pedestrian bridge presents. + +# Rotation + +Rotation-Angle (ROT-Ang) + +![](images/1554f2b582b6c91996922e0d27a13e6890690504b76f6731e929484cc98020fc.jpg) + +Question: What is the angle between the adjacent image? + +Options: A 30 B.60 + +Rationale: Observing from the overlapping areas of the tall building, and the displacement of the short buildings. 30 degree is more feasible. + +Rotation-Difference (ROT-Dif) + +![](images/d76124066281e59112b3b9b77dc8421710c641d62c0b920bff0db435f17796f1.jpg) + +Question: Which image does not belong to this image sequence? + +Rationale: Observing from the range of the overlaps regarding the road., image 4 present the most rotation degree. + +# Counting + +- Counting (CNT) + +![](images/7e2c0df7eaaf10792f257ec4dfb6d65b5fcd55f1a7c57ebd699ce0e31dc10731.jpg) +Figure 2. Examples of CoSpace, along with prompt templates for evaluation. We also provide extra visual guidance and textual rationales for readers (invisible to models during evaluation) for easier understanding. + +![](images/9f29c1ad98575cb90fc026b7a4b90c2ed3e8c633420162a9c15cbc9af4fd3a67.jpg) + +![](images/9b6656572a95637c9279c6210c83597fc61f408ad64447af2c03086f967d072a.jpg) + +![](images/5d8997331dfde03496b702d1ffa3509f97b3ba5c6be9677062be56ec6035ff9c.jpg) + +Rationale: + +One red truck spotted in the $3^{rd}$ image, and it is spotted again in the $4^{th}$ image. + +There is indeed 1 red truck in the scene. + +Question: How many red trucks are visible in the scene? + +Answer: 1 + +Note that $\mathbb{O}$ denotes repeated cars + +# Planning + +![](images/689c430bf92b78135be9e9764e13e542626ef1dde256d5e0e0831d2fa33799b3.jpg) + +Question: Where is the television regarding your position? + +Options: A. Front-left + +C. Front-right + +B. Back-left + +D. Back-right + +Rationale: TV is placed in the right part of $2^{nd}$ image. Right part of the right view indicate back-right direction. + +![](images/7af29df7f5feca7d9de3d7bb07ca96bbb2cd6c49585cbe7ac18293aa3e50afff.jpg) + +- Planning-Decision (PLA-Dec) + +Question: What should you do to reach the fireplace? + +Options: A. Turn right and go ahead + +C. Turn left and go ahead + +B. Turn back and go ahead + +D. Go ahead and turn back + +Rationale:Fireplace locates in the far back, requiring both a turn action and a moving action, need to first turn backwards and then go ahead + +![](images/808e7415e17194c8358cfca03e7e3d73fae6482126560c3aa864213c6cac7e9f.jpg) + +# Task instructions + +General prompt template: + + + + + + + + + +You should generate + +answer from A, B, C or D. + +Your answer: + + + +for Direction and Counting + +categories: + +the first image is facing north, + +the second facing east, the third + +facing south and the fourth + +facing west. + + + +for Rotation category: + +images are arranged clockwise . + +for example, after tak + +first image, the photographer + +turns a certain degree clockwise + +for capturin + + + +for Planning category: + +you are facing the direction in the first image, with the second image showing the scene to your right, the third image showing the scene at your back, and the fourth image showing the scene to your left. + +![](images/983e9298d2aa988e14b08eb46fb0686ef916722a18575f904eaa277612d43c8d.jpg) + +# Notes for readers + +- The directional symbols: + +![](images/d409c9bb24490783d6a278550e46775252122cb6e2e2850dc17a1ef23b5001f0.jpg) + +are invisible to models. They + +are used for easier under + +standing of image orders. + +- The Rationale is also + +invisible to models, as they + +intend for explaining the + +results for readers. + +the second image is facing east". We standardize 8 directional options, including four cardinal directions, North, East, South and West, and four intercardinal directions, Northeast, Southeast, Southwest, and Northwest. + +- Directional Object Perception (DIR-Obj): Some real-world tasks demand abilities far beyond merely identifying the direction of a specified target. It is also a critical ability to distinguish content regarding different directions and to align objects with correct directions. Given a specified direction, the DIR-Obj task requires models to be aware of the appearing objects, and also distinguish those absent from that direction. To achieve this, models need to first reconstruct the entire space from continuous views, and then correctly identify the objects dedicate to a given direction. Specifically, in this task, questions typically focus on the intercardinal directions such as "What is visible in the southeast direction?", which necessitates a more comprehensive understanding of the entire space. + +Observing the surroundings from a static viewpoint natu + +rally involves rotating the perspectives. Therefore, we summaries Rotation as another fundamental category of continuous space perception, with two tasks designed as follows: + +- Rotation-Angle (ROT-Ang): For an observer, when rotating the perspective for more spatial information, there could be overlaps between adjacent observations. Specifically, human can tell the approximate turning angle between consecutive images observed from a static viewpoint by noticing identical furnishings appearing across images. In this task, we investigate if models exhibit similar ability as human does, which requires fine-grained continuous spatial understanding. We apply unified question "What is the turning angle between the adjacent image?" and provide two candidate options for model. This task requires models to carefully exam the overlapping and different areas between adjacent images and the shared features in spatially continuous image sequence. + +- Rotation-Difference (ROT-Dif): Following the ROT-Ang task, we also investigate the model ability to iden + +tify distinguished rotation angles. For the ROT-Dif task, models are given a sequence of five images, four of which share the same turning angle and the other is exceptional. Compared to ROT-Ang, this task emphasizes more on the global understanding of images regarding continuous visual space. The unified question of this task is "Which image does not belong to this image sequence?" It is an open-ended question without options, where models are required to answer with index of the exceptional image. + +On the basis of the above two fundamental categories, we further challenge models with two practical categories. The Counting category addresses the occurrence of objects in the entire scene. The Planning category comprises two spatial-related tasks widely evaluated in embodied AI tasks. For the Counting category, the task is described as: + +- Counting (CNT): Counting is a widely adopted task in existing visual benchmarks that asks models to recognize objects and deduce the time of occurrence of target objects. Generally, existing counting tasks only require models to deal with a single image or multiple spatially discrete images, while in our benchmark, images from continuous visual space where identical objects can occur in multiple images are focused on. This raise a challenge for models to not only recognize targets and count for their occurrences, but also be aware of the existence of the same object across different images. To achieve this, models should locate the overlapping area of adjacent images and align the same object appearing different images. The CNT task is an open-ended task, where models should respond with the total count of the target. + +The ability of continuous space perception is also significant in embodied scenarios. Intuitively, regarding the location of an embodied agent as the static viewpoint, it should have a profound understanding of the overall surroundings to effectively solve questions, and to efficiently plan for the following action sequence. Embodied question-answering (EQA) [36, 45] and navigation [19, 20] are typical tasks that require understand continuous visual space. Following these, we split the Planning category into two tasks addressing the relative locations of targets regarding the VLM-based agent as follows: + +- Planning-Question Answering (PLA-QA): Following the implementation of EQA, we develop the PLA-QA task, requiring models to identify the location of a certain object given a continuous embodied space. In this task, instructions like "Where is the television regarding your position?" are provided for the models, and we formulate four options, containing candidate directions relative to the agent, for each question. +- Planning-Decision (PLA-Dec): This task further investigates the understanding of continuous visual space by asking models to select the proper route to reach the target object. For disambiguation, we standardize the action + +![](images/dc56adf0e9888a06d753626b2fd0a10512cdb11acb5eaa2037d09011a3b445f1.jpg) +Figure 3. Distribution of categories and tasks in our CoSpace. + +space as turning (turning to other directions without displacement) and go ahead. The PLA-Dec task especially focus on the order of actions. As is shown in the case in Figure 2, the options "B. Turn back and go ahead" and "D. Go ahead and turn back" represent two totally different actions and end up in different positions. + +# 3.3. Data Collection and Statistics + +The construction of CoSpace follows the collect-filter-annotate pipeline. We first adopt automatic collection approaches to gather images, and then manually select images that satisfy detailed requirements of our designed tasks and generate questions for these images. Finally human annotators provide answers to each questions. The process is detailed in the following of this section. + +Firstly, to automatically collect sets of images captured in the same scene at static viewpoint but towards different directions, we employ two primary sources, Baidu Map Panorama API that allows us to acquire data towards any desired directions in the real world scene, and HM3D [43], a dataset of 1,000 building-scale 3D reconstructions for embodied tasks. For HM3D, by using Habitat-Sim [40, 46, 48], we instruct an agent wandering in the habitats, we are able to collect images of any views through the attached camera. Secondly, we manually filtered the auto-collected images according to our task requirements, ensuring that the selected images all contain sufficient visual information. With the filtered images, we employed GPT-4o to generate questions via few-shot prompting for the required tasks. Finally, human annotators, an average of two annotators per question, provided answers to these generated questions. This process allowed us to discard vague or unanswerable questions, ensuring that only high-quality question-answer pairs were included in our proposed CoSpace. + +Our CoSpace comprises a total of 1,626 question-answer pairs and 2,918 images. Among the images, 2,302 are accessed from Baidu Map API and 616 are derived from HM3D dataset. The distribution of instances over categories and tasks is shown in Figure 3. + +
ModelsDIR-RecDIR-ObjCNTROT-AngROT-DifPLA-QAPLA-DecAverage
ACCqACCpACCqACCpACCqACCpACCACCACCACC
Proprietary Models
Claude-3.7-sonnet [2]44.4029.2045.6035.6045.0038.0064.3393.5054.7369.3451.97
Gemini-1.5-pro [44]37.6015.6040.6031.6038.2524.0059.6782.0051.6462.9144.39
GPT-4o [39]40.4022.8046.0032.0040.0023.5058.3350.5053.0554.4642.10
>70B Open-source Models
InternVL2_5-78B [6]32.2024.4054.4047.6051.2543.0050.0077.0032.3942.7245.50
Qwen2-VL-72B [51]31.0023.2053.6044.8044.7537.0050.0062.0042.7259.1544.82
InternVL2-76B [7]33.6013.2046.4039.2050.0043.5050.0023.0025.3530.0535.43
LLaVA-OneVision-72B [24]19.2010.8044.2034.4024.2519.0050.0026.0032.8630.9929.17
<13B Open-source Models
MiniCPM-V 2.6 [57]32.8021.2040.4031.6038.5031.5050.0056.0041.3127.7037.10
Qwen2-VL-7B [51]26.4016.4039.2031.6045.0036.0050.0051.5034.2726.7635.71
Mantis-8B [18]30.6024.4031.2028.0041.7536.5050.0039.5033.3327.7034.30
InternVL2-8B [7]29.0012.8037.2030.4038.5031.0050.3347.5032.3927.7033.68
VILA1.5-8B [29]34.0030.4028.2024.0042.2538.0050.0018.5022.0742.7233.01
Idefics3-8B [21]34.0021.2038.6028.0032.7525.0048.0028.0026.7625.3530.77
LLaVA-OneVision-7B [24]18.8012.0037.6033.6037.7533.0050.0021.5029.5826.2930.01
Phi-3.5-vision [1]20.4012.0033.2028.4036.7534.0050.0022.5034.7426.7629.88
Brote-IM-XXL [55]33.0013.2031.0024.8030.7529.5050.0018.0026.7620.1927.72
LongVA-7B [60]23.0016.0031.6028.0028.0022.0048.6721.5019.2530.9926.90
Mono-InternVL-2B [35]30.0027.6032.0027.6014.008.5050.0018.0025.3524.8825.79
mPLUG-Owl3-7B [58]22.806.8029.2015.6022.259.0046.3316.5027.7021.1321.73
+ +Table 2. Main results on our benchmark. The proprietary models are accessed via APIs, and the open-source models are accessed from Huggingface checkpoints. + +# 4. Experiments + +As discussed above, CoSpace features reconstructing the surrounding space via multiple images, with each image representing a different space segments. Different from previous multi-image benchmark involving spatial reasoning tasks, the input order of images is crucial in our tasks. To enable that models can associate each image with a specific direction to answer questions in our benchmark, we provide direction references along with images following the prompt templates in Figure 2 through out the evaluation. + +# 4.1. Evaluation Metrics + +For the Direction and the Counting categories, for fair evaluation, we adopt both pair-wise accuracy $ACC_{p}$ and query-wise accuracy $ACC_{q}$ as introduced by Fu et al. [12] and Luo et al. [34] instead of simple accuracy. For each question, we task models twice with different prompts. $ACC_{p}$ only counts when models correctly response to both prompts, while $ACC_{q}$ is the averaged score of the two responses. For these different prompts regarding the same question, we vary only the order of images by preserving the clockwise order but changing the beginning view, so that the answer to the question remains unchanged. For example, if original image sequence begins with the northern image, the modified sequence could start from the east, south or west. + +For the Rotation category, some of the instances does not cover the entire space, and the such a modification of input image sequence will inevitable disrupt the comprehension of the continuous space. Similarly, for the Planning category, considering that the original input sequence features the front view as the first image and this manner inherently matches the thinking pattern when models take + +actions. Therefore, for these two categories, we apply only a single context for each instances during evaluation, and report accuracy, $ACC$ , as final scores. + +# 4.2. Models + +We accessed three proprietary models, including GPT-4o [39], Gemini-1.5-pro [44] and Claude-3.7-sonnet [2], along with 16 widely used open-source VLMs accepting multiple images simultaneously. Among the open-source models, four have more than 70B parameters, while the others range in size from 2B to 13B. We also enable the evaluation for powerful single-image models, such as LLaVA-1.6 [31] and LLaMA-3.2 [11], for further investigation. For fair comparison, we design a single-image pipeline for some our designed tasks and conduct evaluations on both single-image and multi-image models in Section 5.3. + +# 4.3. Results + +The overall results on our benchmark is reported in Table 2. We derive three main findings from this table: 1. The performance of proprietary models and $>70B$ open-source models largely varies. As expected, most proprietary models and $>70B$ open-source models marginally outperform the evaluated open-source models. However, there are significant dependencies among these models, where Claude-3.5-sonnet achieves an average score of $49.37\%$ , InternVL2_5-78B achieves $45.50\%$ , while LLaVA-OneVision-72B only presents an average score of $29.17\%$ , even lower than more than half of the $<13B$ open-source models. 2. Proprietary models do not equally perform the best across all the tasks. Proprietary models fail to beat open-source models on tasks including DIR-Obj and CNT, + +
ModelsDIR-RecDIR-ObjCNTPLA-QAPLA-Dec
Proprietary Models
Claude-3.7-sonnet36.3631.1737.2044.4044.50
Gemini-1.5-pro37.2037.2027.5038.0336.15
GPT-4o37.2041.2039.0045.0739.91
>70B Open-source Models
InternVL2_5-78B29.6052.8044.0034.27*36.15
Qwen2-VL-72B30.4047.6046.00*37.0933.33
InternVL2-76B25.2038.8052.00*25.3525.35
LLaVA-OneVision-72B20.40*43.6022.5036.62*32.39*
<13B Open-source Models
MiniCPM-V 2.622.8040.0037.0039.4426.76
Qwen2-VL-7B19.6035.2040.5034.2723.00
Mantis-8B34.80*30.8042.50*31.9229.11*
InternVL2-8B31.20*36.4042.50*22.5426.76
VILA1.5-8B33.6029.20*42.0019.7239.91
Idefics3-8B37.60*37.2032.5023.9424.88
LLaVA-OneVision-7B17.2038.00*35.0030.05*27.70*
Phi-3.5-vision26.00*34.40*36.5031.4624.41
Brote-IM-XXL34.40*29.6029.5029.58*17.84
LongVA-7B16.0031.6028.50*17.8428.64
Mono-InternVL-2B18.8033.60*15.50*25.3522.07
mPLUG-Owl3-7B26.40*29.2029.50*29.11*32.39*
+ +Table 3. Results for disrupted image orders. $ACC_{q}$ is reported for Direction and Counting tasks and $ACC$ is reported for the Planning task. * indicates the performance is higher than that in Table 2. + +where $>70\mathrm{B}$ open-source models exhibit stronger performance. 3. The ROT-Ang task is quite challenging for open-source models. It is noticeable that for the ROT-Ang task, all open-source models achieve similar results near $50\%$ , which are almost equal to random selection. Such result is caused by the fixed predictions over all different context and images input. We will elaborately investigate such phenomena in Section 5.4. + +We conduct manual evaluation on all instances to assess the reasonability and difficulty of CoSpace and results are provided in Appendix ??. Results show that human performance excels the best models across all tasks. Human annotators achieved an average $ACC_{q}$ of $86.25\%$ , which is significantly higher than that of Claude-3.7-sonnet (59.56%). + +# 5. Analysis + +# 5.1. Disrupted Image Order + +Considering the spatial relationship between input images of our CoSpace, the order of input images can make significant difference on the understanding of the space, especially when the order being thoroughly disrupted. Based on this speculation, we randomly shuffled the input sequence, and provided the directional orientation only of the first image for the model, with directions of the rest images to be inferred by the model itself. We conducted evaluation under such setting on three tasks: Direction, Counting and Planning, and the results are illustrated in Table 3. + +Surprisingly, the performance of $< 13\mathrm{B}$ models under disrupted image order was not always lower than the basic setting. In fact, half of the $< 13\mathrm{B}$ open-source models achieved a higher score in the DIR-Rec task under disrupted set + +![](images/8db564116b8fb674fff9f811684485b15026bd1df307597ab2fb592aeadc7458.jpg) +Figure 4. Variance in $ACC_{q}$ of the average on the Direction and Counting categories for eight selected models. Model abbreviations are used for simplicity. Accuracy in the figure represents the average of $ACC_{q}$ over two categories. + +
Img OrderClaude GeminiQwen2InternVL2OneVisionPhi-3.5LongVA
Sequential79.5873.6469.9134.2534.7833.3321.6236.59
Disrupted60.7156.7956.2524.6639.5835.9426.8731.58
+ +Table 4. Results of conditional accuracy on the Planning category for eight selected models. Model names are abbreviated. Open-source models all refer to $<13\mathrm{B}$ ones. + +ting. Additionally, for other tasks, an average of three to five open-source models performed slightly better with disrupted image order. Nevertheless, similar phenomenon was not observed for proprietary models and most of the $>70\mathrm{B}$ open-source models, except LLaVA-OneVision-72B. Also, the evaluated proprietary models similarly underperformed across all tasks under disrupted setting, which aligns more closely with our expectations. For $>70\mathrm{B}$ open-source models, three of four outperformed on only one task against basic setting, with LLaVA-OneVision-72B producing a completely opposite result. Considering the such phenomenon, we conclude that the results for $<13\mathrm{B}$ open-source models are partly due to the instability of predictions. Moreover, we acknowledge that the impact of disrupting the image order may not be as significant for models as it is for human, exhibiting the discrepancy between models and human when perceiving continuous visual space via multiple images. + +# 5.2. Performance Consistency + +We mentioned in Section 4.1 that for the Direction and Counting categories, both $ACC_{q}$ and $ACC_{p}$ are adopted as evaluation metrics. We present the gap between $ACC_{q}$ and $ACC_{p}$ , i.e. variance in accuracy, in Figure 4. A smaller variance indicates that a model responds more consistently to input changes, reflecting better performance consistency. As shown, all models exhibit relatively large variance, with proprietary models (Claude-3.7-sonnet, Gemini-1.5-pro, GPT-4o) displaying even higher variances than open-source models, suggesting room for improvement. Such result highlights the importance of introducing $ACC_{p}$ . + +To further analyze the performance consistency of models, we introduce the metric conditional accuracy in the + +
ModelsDIR-RecDIR-ObjCNTPLA-QAPLA-DecAVG
Models with only one image in the input
LLaVA-v1.627.6032.0021.0021.6025.8225.60
LLama-3.2-11B29.2028.8028.5017.8416.9024.25
Models with multi-image perception ability
InternVL2-8B36.40*34.8027.0033.80*31.92*32.78
MiniCPM-V 2.623.2027.2026.0031.9232.39*28.14
Phi-3.5-vision18.4036.00*28.0031.4623.0027.37
LLaVA-OneVision26.80*32.8016.0022.0724.4124.42
+ +Table 5. Results for single image pipeline. We report $ACC_{q}$ for Direction and Counting tasks and $ACC$ for Planning tasks. * indicates the performance is higher than that in Table 2. + +Planning category. Conditional accuracy measures how often models successfully solve the PLA-Dec question given that they answer the correspondent PLA-QA question correctly, as in PLA-QA task and PLA-Dec task, two respective questions regarding the same set of images typically feature the same object or direction as target. Results in Tabel 4 reveal that open-source models exhibit a greater gap in conditional accuracy against proprietary models than on the Direction and Counting categories. While proprietary models make more consistent decisions based on correct detection, the gap remains non-negligible. + +Above all, we argue that maintaining performance consistency is rather important for the further improvement of models, as it is closely related to robustness and rather essential for more complicated tasks. + +# 5.3. Single Image Pipeline + +We presented in Section 3.1 that the understanding of continuous visual space inherently relies on multiple images as input to represent the whole space. Also, recent work [54] explored to construct a pipeline that can utilize caption and memory to transform the multi-image task into single image format. Similarly, we developed a single image pipeline for the Direction, Counting and Planning categories in our benchmark. For the Rotation category, as the single image input will inevitably compromise the uniform understanding of continuous visual space and the relationship between images can hardly be replaced by mere captions, we consider it not suitable for the single image pipeline. For evaluation, we selected LLaVA-v1.6 [31] and Llama-3.2-Vision [11] as the representatives of single image models. We also selected some multi-image models and conducted the single image evaluation. + +Results of the single image pipeline are reported in Table 5. Overall, the multi-image models underperformed under this setting compared with the multi-image pipeline. However, in the Direction-Rec and Planning-Dec task, two of four models achieve a higher accuracy, which is far beyond our expectations. We conclude that with the effective summarization of captions, models can still have an understanding of visual clues and sometimes the complete of all images can even add to the performance. Also, + +three of the four multi-image models outperform the single-image models on the average accuracy, indicating that models with inherent multi-image capabilities boast an overall better understanding of the visual clues and information extracted from the effectively multi-image scene. Based on this observation, the importance of empowering models with multi-image capabilities needs again to be emphasized. + +# 5.4. Investigation on Angle-related Reasoning + +Table 2 shows that all the assessed open-source models achieve near-half ACCs on the Rotation-Angle (ROT-Ang) task. After checking the responses of different models, we found out that these models tend to produce the same answer across all the queries, e.g. "A" for all the 300 questions in this task. Therefore, due to the even distribution of options, these models achieve the similar near-half ACC. + +To discover the reason behind such phenomenon, we asked models to generate the rationale when answering the questions in the ROT-Ang task and observed some of these models can follow the instruction to generate rationales. For example, InternVL2-8B no longer always selected the same option and consequently correctly answered some question that it originally failed. However, we also noticed that it tended to identify the whole surrounding covered by the images, thus perceiving the sum of turning angles as 360 degrees. Therefore, it frequently output 90 degrees as the final answer, even if 90 degrees not included in the options. In contrast, proprietary models like Claude-3.7-sonnet could firstly calculate the sum of covered angles across all images based on the two give options and then decided which answer better matched the scene in given images. + +Though all the assessed open-source models fail the Rotation-Angle task, we still need to underline the importance of this task and the related capability to correctly identify the rotation in the space. Such capability lays the foundation for more advanced tasks in practical use and has already been boasted by some of the proprietary models. + +# 6. Conclusion + +Recognizing the gap between previous work and more advanced spatial understanding, we propose the concept of continuous space perception and present a novel benchmark CoSpace that comprises four categories and seven tasks. CoSpace covers both fundamental spatial comprehending abilities and more practical and commonly-used capabilities to comprehensively assess continuous space perception ability for VLMs. We evaluated three widely used proprietary models and sixteen most advanced open-source models and conducted detailed analysis for noticeable results in the experiments. Results reveal a gap between the assessed open-source and proprietary models, especially for performance consistency and in rotation-based perception. + +# Acknowledgments + +This work is supported by the National Key R&D Program of China (2022ZD0160502), the National Natural Science Foundation of China (No. 62276152), and funding from Wuxi Research Institute of Applied Technologies, Tsinghua University under Grant 20242001120. We acknowledge Yan Liu, Jidong Chen, Jianyang Liu, Shuangyu Li and Shu Chi for their effort during the process of data annotation and human evaluation. + +# References + +[1] Phi-3 technical report: A highly capable language model locally on your phone, 2024. 6 +[2] Anthropic. The Claude 3 Model Family: Opus, Sonnet, Haiku. 2024. 6 +[3] Xu Cao, Tong Zhou, Yunsheng Ma, Wenqian Ye, Can Cui, Kun Tang, Zhipeng Cao, Kaizhao Liang, Ziran Wang, James M Rehg, et al. MAPLM: a real-world large-scale vision-language benchmark for map and traffic scene understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21819-21830, 2024. 1 +[4] Boyuan Chen, Zhuo Xu, Sean Kirmani, Brain Ichter, Dorsa Sadigh, Leonidas Guibas, and Fei Xia. SpatialVLM: endowing vision-language models with spatial reasoning capabilities. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14455-14465, 2024. 1 +[5] Liang Chen, Yichi Zhang, Shuhuai Ren, Haozhe Zhao, Zefan Cai, Yuchi Wang, Peiyi Wang, Tianyu Liu, and Baobao Chang. Towards end-to-end embodied decision making via multi-modal large language model: Explorations with gpt4- vision and beyond. arXiv preprint arXiv:2310.02071, 2023. 1 +[6] Zhe Chen, Weiyun Wang, Yue Cao, Yangzhou Liu, Zhangwei Gao, Erfei Cui, Jinguo Zhu, Shenglong Ye, Hao Tian, Zhaoyang Liu, et al. Expanding performance boundaries of open-source multimodal models with model, data, and test-time scaling. arXiv preprint arXiv:2412.05271, 2024. 6 +[7] Zhe Chen, Weiyun Wang, Hao Tian, Shenglong Ye, Zhangwei Gao, Erfei Cui, Wenwen Tong, Kongzhi Hu, Jiapeng Luo, Zheng Ma, et al. How far are we to GPT-4V? closing the gap to commercial multimodal models with open-source suites. ArXiv preprint, abs/2404.16821, 2024. 6 +[8] Hao-Tien Lewis Chiang, Zhuo Xu, Zipeng Fu, Mithun George Jacob, Tingnan Zhang, Tsang-Wei Edward Lee, Wenhao Yu, Connor Schenck, David Rendleman, Dhruv Shah, et al. Mobility VLA: multimodal instruction navigation with long-context vlms and topological graphs. arXiv preprint arXiv:2407.07775, 2024. 1 +[9] Song Dingjie, Shunian Chen, Guiming Hardy Chen, Fei Yu, Xiang Wan, and Benyou Wang. Milebench: Benchmarking MLLMs in long context. In First Conference on Language Modeling, 2024. 2, 3 +[10] Sivan Doveh, Shaked Perek, M Jehanzeb Mirza, Wei Lin, Amit Alfassy, Assaf Arbelle, Shimon Ullman, and Leonid + +Karlinsky. Towards multimodal in-context learning for vision & language models. arXiv preprint arXiv:2403.12736, 2024. 3 +[11] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, and et al. The llama 3 herd of models, 2024. 6, 8 +[12] Chaoyou Fu, Peixian Chen, Yunhang Shen, Yulei Qin, Mengdan Zhang, Xu Lin, Jinrui Yang, Xiawu Zheng, Ke Li, Xing Sun, et al. Mme: A comprehensive evaluation benchmark for multimodal large language models. arXiv preprint arXiv:2306.13394, 2023. 6 +[13] Chaoyou Fu, Yuhan Dai, Yondong Luo, Lei Li, Shuhuai Ren, Renrui Zhang, Zihan Wang, Chenyu Zhou, Yunhang Shen, Mengdan Zhang, et al. Video-MME: the first-ever comprehensive evaluation benchmark of multi-modal llms in video analysis. ArXiv preprint, abs/2405.21075, 2024. 2, 3 +[14] Xingyu Fu, Yushi Hu, Bangzheng Li, Yu Feng, Haoyu Wang, Xudong Lin, Dan Roth, Noah A. Smith, Wei-Chiu Ma, and Ranjay Krishna. Blink: Multimodal large language models can see butA not perceive. In Computer Vision - ECCV 2024, pages 148-166, Cham, 2025. Springer Nature Switzerland. 2, 3 +[15] Yupan Huang, Zaiqiao Meng, Fangyu Liu, Yixuan Su, Nigel Collier, and Yutong Lu. Sparkles: Unlocking chats across multiple images for multimodal instruction-following models. In ICLR 2024 Workshop on Navigating and Addressing Data Problems for Foundation Models, 2024. 2, 3 +[16] Baoxiong Jia, Yixin Chen, Huangyue Yu, Yan Wang, Xuesong Niu, Tengyu Liu, Qing Li, and Siyuan Huang. Sceneverse: Scaling 3d vision-language learning for grounded scene understanding. arXiv preprint arXiv:2401.09340, 2024. 1 +[17] Mengzhao Jia, Wenhao Yu, Kaixin Ma, Tianqing Fang, Zhihan Zhang, Siru Ouyang, Hongming Zhang, Meng Jiang, and Dong Yu. Leopard: A Vision Language Model For Text-Rich Multi-Image Tasks, 2024. 1, 3 +[18] Dongfu Jiang, Xuan He, Huaye Zeng, Cong Wei, Max Ku, Qian Liu, and Wenhu Chen. Mantis: interleaved multi-image instruction tuning. ArXiv preprint, abs/2405.01483, 2024. 3, 6 +[19] Mukul Khanna*, Ram Ramrakhya*, Gunjan Chhablani, Sriram Yenamandra, Theophile Gervet, Matthew Chang, Zsolt Kira, Devendra Singh Chaplot, Dhruv Batra, and Roozbeh Mottaghi. Goat-bench: A benchmark for multi-modal life-long navigation. In CVPR, 2024. 5 +[20] Taewoong Kim, Cheolhong Min, Byeonghwi Kim, Jinyeon Kim, Wonje Jeung, and Jonghyun Choi. Realfred: An embodied instruction following benchmark in photo-realistic environment. In ECCV, 2024. 5 +[21] Hugo Laurençon, Andres Marafioti, Victor Sanh, and Léo Tronchon. Building and better understanding vision-language models: insights and future directions, 2024. 3, 6 +[22] Bo Li, Yuanhan Zhang, Liangyu Chen, Jinghao Wang, Jingkang Yang, and Ziwei Liu. Otter: A Multi-Modal Model with In-Context Instruction Tuning, 2023. 1, 3 + +[23] Bohao Li, Yuying Ge, Yixiao Ge, Guangzhi Wang, Rui Wang, Ruimao Zhang, and Ying Shan. SEED-Bench: Benchmarking multimodal large language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13299-13308, 2024. 2, 3 +[24] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. LLaVA-OneVision: Easy visual task transfer. ArXiv preprint, abs/2408.03326, 2024. 6 +[25] Feng Li, Renrui Zhang, Hao Zhang, Yuanhan Zhang, Bo Li, Wei Li, Zejun Ma, and Chunyuan Li. Llava-last-interleave: Tackling multi-image, video, and 3d in large multimodal models, 2024. 3 +[26] Juncheng Li, Kaihang Pan, Zhiqi Ge, Minghe Gao, Wei Ji, Wenqiao Zhang, Tat-Seng Chua, Siliang Tang, Hanwang Zhang, and Yueting Zhuang. Fine-tuning multimodal LLMs to follow zero-shot demonstrative instructions. In The Twelfth International Conference on Learning Representations, 2024. 2, 3 +[27] Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Yi Liu, Zun Wang, Jilan Xu, Guo Chen, Ping Luo, et al. Mvbench: A comprehensive multi-modal video understanding benchmark. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22195-22206, 2024. 3 +[28] Xiaoqi Li, Mingxu Zhang, Yiran Geng, Haoran Geng, Yuxing Long, Yan Shen, Renrui Zhang, Jiaming Liu, and Hao Dong. ManipLLM: Embodied multimodal large language model for object-centric robotic manipulation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18061-18070, 2024. 1 +[29] Ji Lin, Hongxu Yin, Wei Ping, Pavlo Molchanov, Mohammad Shoeybi, and Song Han. Vila: On pre-training for visual language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26689-26699, 2024. 6 +[30] Ziyi Lin, Chris Liu, Renrui Zhang, Peng Gao, Longtian Qiu, Han Xiao, Han Qiu, Chen Lin, Wenqi Shao, Keqin Chen, et al. Sphinx: The joint mixing of weights, tasks, and visual embeddings for multi-modal large language models. arXiv preprint arXiv:2311.07575, 2023. 1, 2 +[31] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning, 2023. 2, 6, 8 +[32] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. LLaVA-NeXT: Improved reasoning, OCR, and world knowledge, 2024. 2, 3 +[33] Haowei Liu, Xi Zhang, Haiyang Xu, Yaya Shi, Chaoya Jiang, Ming Yan, Ji Zhang, Fei Huang, Chunfeng Yuan, Bing Li, and Weiming Hu. MIBench: Evaluating multimodal large language models over multiple images, 2024. 2, 3 +[34] Fuwen Luo, Chi Chen, Zihao Wan, Zhaolu Kang, Qidong Yan, Yingjie Li, Xiaolong Wang, Siyu Wang, Ziyue Wang, Xiaoyue Mi, Peng Li, Ning Ma, Maosong Sun, and Yang Liu. Codis: Benchmarking context-dependent visual comprehension for multimodal large language models. arXiv preprint arXiv:2402.13607, 2024. 6 +[35] Gen Luo, Xue Yang, Wenhan Dou, Zhaokai Wang, Jifeng Dai, Yu Qiao, and Xizhou Zhu. Mono-internvl: Pushing the + +boundaries of monolithic multimodal large language models with endogenous visual pre-training, 2024. 6 +[36] Arjun Majumdar, Anurag Ajay, Xiaohan Zhang, Pranav Putta, Sriram Yenamandra, Mikael Henaff, Sneha Silwal, Paul Mcvay, Oleksandr Maksymets, Sergio Arnaud, et al. OpenEQA: embodied question answering in the era of foundation models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16488-16498, 2024. 5 +[37] Fanqing Meng, Jin Wang, Chuanhao Li, Quanfeng Lu, Hao Tian, Jiaqi Liao, Xizhou Zhu, Jifeng Dai, Yu Qiao, Ping Luo, Kaipeng Zhang, and Wenqi Shao. MMIU: Multimodal multi-image understanding for evaluating large vision-language models, 2024. 2 +[38] Jiahao Nie, Gongjie Zhang, Wenbin An, Yap-Peng Tan, Alex C Kot, and Shijian Lu. Mmrel: A relation understanding dataset and benchmark in the mllm era. arXiv preprint arXiv:2406.09121, 2024. 1, 2 +[39] OpenAI. Hello GPT-4o, 2024. Accessed on June 16, 2024. 6 +[40] Xavier Puig, Eric Undersander, Andrew Szot, Mikael Dallaire Cote, Tsung-Yen Yang, Ruslan Partsey, Ruta Desai, Alexander Clegg, Michal Hlavac, So Yeon Min, Vladimir Vondruš, Theophile Gervet, Vincent-Pierre Berges, John M Turner, Oleksandr Maksymets, Zsolt Kira, Mrinal Kalakrishnan, Jitendra Malik, Devendra Singh Chaplot, Unnat Jain, Dhruv Batra, Akshara Rai, and Roozbeh Mottaghi. Habitat 3.0: A co-habitat for humans, avatars, and robots. In Proceedings of the Twelfth International Conference on Learning Representations (ICLR 2024), 2024. 5 +[41] Navid Rajabi and Jana Kosecka. Gsr-bench: A benchmark for grounded spatial reasoning evaluation via multimodal llms. arXiv preprint arXiv:2406.13246, 2024. 2 +[42] Abhinav Rajvanshi, Karan Sikka, Xiao Lin, Bhoram Lee, Han pang Chiu, and Alvaro Velasquez. SayNav: Grounding large language models for dynamic planning to navigation in new environments. In 34th International Conference on Automated Planning and Scheduling, 2024. 1 +[43] Santhosh Kumar Ramakrishnan, Aaron Gokaslan, Erik Wijmans, Oleksandr Maksymets, Alexander Clegg, John M Turner, Eric Undersander, Wojciech Galuba, Andrew Westbury, Angel X Chang, Manolis Savva, Yili Zhao, and Dhruv Batra. Habitat-matterport 3d dataset (HM3d): 1000 large-scale 3d environments for embodied AI. In Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 2), 2021. 5 +[44] Machel Reid, Nikolay Savinov, Denis Teplyashin, Dmitry Lepikhin, Timothy Lillicrap, Jean-baptiste Alayrac, Radu Soricut, Angeliki Lazaridou, Orhan First, Julian Schrittwieser, et al. Gemini 1.5: unlocking multimodal understanding across millions of tokens of context. ArXiv preprint, abs/2403.05530, 2024. 6 +[45] Allen Z. Ren, Jaden Clark, Anushri Dixit, Masha Itkina, Anirudha Majumdar, and Dorsa Sadigh. Explore until confident: Efficient exploration for embodied question answering, 2024. 5 +[46] Manolis Savva, Abhishek Kadian, Oleksandr Maksymets, Yili Zhao, Erik Wijmans, Bhavana Jain, Julian Straub, Jia + +Liu, Vladlen Koltun, Jitendra Malik, Devi Parikh, and Dhruv Batra. Habitat: A Platform for Embodied AI Research. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2019. 5 +[47] Fatemeh Shiri, Xiao-Yu Guo, Mona Far, Xin Yu, Reza Haf, and Yuan-Fang Li. An empirical analysis on spatial reasoning capabilities of large multimodal models. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 21440-21455, 2024. 1, 2 +[48] Andrew Szot, Alex Clegg, Eric Undersander, Erik Wijmans, Yili Zhao, John Turner, Noah Maestre, Mustafa Mukadam, Devendra Chaplot, Oleksandr Maksymets, Aaron Gokaslan, Vladimir Vondrus, Sameer Dharur, Franziska Meier, Wojciech Galuba, Angel Chang, Zsolt Kira, Vladlen Koltun, Jitendra Malik, Manolis Savva, and Dhruv Batra. Habitat 2.0: Training home assistants to rearrange their habitat. In Advances in Neural Information Processing Systems (NeurIPS), 2021. 5 +[49] Yunlong Tang, Jing Bi, Siting Xu, Luchuan Song, Susan Liang, Teng Wang, Daoan Zhang, Jie An, Jingyang Lin, Rongyi Zhu, et al. Video understanding with large language models: A survey. arXiv preprint arXiv:2312.17432, 2023. 3 +[50] Jiayu Wang, Yifei Ming, Zhenmei Shi, Vibhav Vineet, Xin Wang, Yixuan Li, and Neel Joshi. Is a picture worth a thousand words? delving into spatial reasoning for vision language models. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 2 +[51] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Yang Fan, Kai Dang, Mengfei Du, Xuancheng Ren, Rui Men, Dayiheng Liu, Chang Zhou, Jingren Zhou, and Junyang Lin. Qwen2-VL: Enhancing vision-language model's perception of the world at any resolution, 2024. 6 +[52] Xidong Wang, Dingjie Song, Shunian Chen, Chen Zhang, and Benyou Wang. LongLLaVA: Scaling Multi-modal LLMs to 1000 Images Efficiently via a Hybrid Architecture, 2024. 2, 3 +[53] Xiyao Wang, Yuhang Zhou, Xiaoyu Liu, Hongjin Lu, Yuancheng Xu, Feihong He, Jaehong Yoon, Taixi Lu, Fuxiao Liu, Gedas Bertasius, Mohit Bansal, Huaxiu Yao, and Furong Huang. Mementos: A comprehensive benchmark for multimodal large language model reasoning over image sequences. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 416-442, Bangkok, Thailand, 2024. Association for Computational Linguistics. 2, 3 +[54] Ziyue Wang, Chi Chen, Fuwen Luo, Yurui Dong, Yuanchi Zhang, Yuzhuang Xu, Xiaolong Wang, Peng Li, and Yang Liu. ActiView: Evaluating active perception ability for multimodal large language models. arXiv preprint arXiv:2410.04659, 2024. 3, 8 +[55] Ziyue Wang, Chi Chen, Yiqi Zhu, Fuwen Luo, Peng Li, Ming Yan, Ji Zhang, Fei Huang, Maosong Sun, and Yang Liu. Browse and Concentrate: comprehending multimodal content via prior-llm context fusion. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 11229–11245, + +Bangkok, Thailand, 2024. Association for Computational Linguistics. 3, 6 +[56] Zhutian Yang, Caelan Garrett, Dieter Fox, Tomás Lozano-Pérez, and Leslie Pack Kaelbling. Guiding long-horizon task and motion planning with vision language models, 2024. 1 +[57] Yuan Yao, Tianyu Yu, Ao Zhang, Chongyi Wang, Junbo Cui, Hongji Zhu, Tianchi Cai, Haoyu Li, Weilin Zhao, Zhihui He, et al. MiniCPM-V: A GPT-4V level MLLM on your phone. ArXiv preprint, abs/2408.01800, 2024. 6 +[58] Jiabo Ye, Haiyang Xu, Haowei Liu, Anwen Hu, Ming Yan, Qi Qian, Ji Zhang, Fei Huang, and Jingren Zhou. mPLUG-Owl3: towards long image-sequence understanding in multi-modal large language models. ArXiv preprint, abs/2408.04840, 2024. 3, 6 +[59] Haotian Zhang, Mingfei Gao, Zhe Gan, Philipp Duffer, Nina Wenzel, Forrest Huang, Dhruti Shah, Xianzhi Du, Bowen Zhang, Yanghao Li, Sam Dodge, Keen You, Zhen Yang, Aleksei Timofeev, Mingze Xu, Hong-You Chen, JeanPhilippe Fauconnier, Zhengfeng Lai, Haoxuan You, Zirui Wang, Afshin Dehghan, Peter Grasch, and Yinfei Yang. MM1.5: Methods, analysis & insights from multimodal llm fine-tuning, 2024. 3 +[60] Peiyuan Zhang, Kaichen Zhang, Bo Li, Guangtao Zeng, Jingkang Yang, Yuanhan Zhang, Ziyue Wang, Haoran Tan, Chunyuan Li, and Ziwei Liu. Long context transfer from language to vision. arXiv preprint arXiv:2406.16852, 2024. 1, 2, 6 +[61] Yi-Fan Zhang, Huanyu Zhang, Haochen Tian, Chaoyou Fu, Shuangqing Zhang, Junfei Wu, Feng Li, Kun Wang, Qing-song Wen, Zhang Zhang, Liang Wang, Rong Jin, and Tieniu Tan. Mme-realworld: Could your multimodal llm challenge high-resolution real-world scenarios that are difficult for humans?, 2024. 1 +[62] Haozhe Zhao, Zefan Cai, Shuzheng Si, Xiaojian Ma, Kaikai An, Liang Chen, Zixuan Liu, Sheng Wang, Wenjuan Han, and Baobao Chang. MMICL: Empowering vision-language model with multi-modal in-context learning. In The Twelfth International Conference on Learning Representations, 2024. 3 +[63] Gengze Zhou, Yicong Hong, Zun Wang, Xin Eric Wang, and Qi Wu. NavGPT-2: Unleashing navigational reasoning capability for large vision-language models. In European Conference on Computer Vision, pages 260–278. Springer, 2025. 1, 3 \ No newline at end of file diff --git a/2025/CoSpace_ Benchmarking Continuous Space Perception Ability for Vision-Language Models/images.zip b/2025/CoSpace_ Benchmarking Continuous Space Perception Ability for Vision-Language Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..1715fce169376d93796f3900caab615935d23a61 --- /dev/null +++ b/2025/CoSpace_ Benchmarking Continuous Space Perception Ability for Vision-Language Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d3717538983de4da5ff347a0edbc537b380cb7f3b5ae3a95683447e9f6ca36aa +size 534803 diff --git a/2025/CoSpace_ Benchmarking Continuous Space Perception Ability for Vision-Language Models/layout.json b/2025/CoSpace_ Benchmarking Continuous Space Perception Ability for Vision-Language Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..772dee79f2a24ce904a36a0a7355d4e706d88409 --- /dev/null +++ b/2025/CoSpace_ Benchmarking Continuous Space Perception Ability for Vision-Language Models/layout.json @@ -0,0 +1,10841 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 113, + 103, + 500, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 103, + 500, + 140 + ], + "spans": [ + { + "bbox": [ + 113, + 103, + 500, + 140 + ], + "type": "text", + "content": "CoSpace: Benchmarking Continuous Space Perception Ability for Vision-Language Models" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 145, + 160, + 471, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 160, + 471, + 175 + ], + "spans": [ + { + "bbox": [ + 145, + 160, + 471, + 175 + ], + "type": "text", + "content": "Yiqi Zhu" + }, + { + "bbox": [ + 145, + 160, + 471, + 175 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 145, + 160, + 471, + 175 + ], + "type": "text", + "content": ", Ziyue Wang" + }, + { + "bbox": [ + 145, + 160, + 471, + 175 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 145, + 160, + 471, + 175 + ], + "type": "text", + "content": ", Can Zhang" + }, + { + "bbox": [ + 145, + 160, + 471, + 175 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 145, + 160, + 471, + 175 + ], + "type": "text", + "content": ", Peng Li" + }, + { + "bbox": [ + 145, + 160, + 471, + 175 + ], + "type": "inline_equation", + "content": "^{2,4\\dagger}" + }, + { + "bbox": [ + 145, + 160, + 471, + 175 + ], + "type": "text", + "content": ", Yang Liu" + }, + { + "bbox": [ + 145, + 160, + 471, + 175 + ], + "type": "inline_equation", + "content": "^{1,2,4,5\\dagger}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 108, + 175, + 510, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 175, + 510, + 190 + ], + "spans": [ + { + "bbox": [ + 108, + 175, + 510, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 108, + 175, + 510, + 190 + ], + "type": "text", + "content": "Dept. of Comp. Sci. & Tech., Institute for AI, Tsinghua University, Beijing, China" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 119, + 190, + 498, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 119, + 190, + 498, + 203 + ], + "spans": [ + { + "bbox": [ + 119, + 190, + 498, + 203 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 119, + 190, + 498, + 203 + ], + "type": "text", + "content": "Institute for AI Industry Research (AIR), Tsinghua University, Beijing, China" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 203, + 555, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 203, + 555, + 217 + ], + "spans": [ + { + "bbox": [ + 62, + 203, + 555, + 217 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 62, + 203, + 555, + 217 + ], + "type": "text", + "content": "School of Computer and Communication Engineering, University of Science and Technology Beijing" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 159, + 218, + 458, + 231 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 218, + 458, + 231 + ], + "spans": [ + { + "bbox": [ + 159, + 218, + 458, + 231 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 159, + 218, + 458, + 231 + ], + "type": "text", + "content": "Shanghai Artificial Intelligence Laboratory, Shanghai, China" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 104, + 232, + 511, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 232, + 511, + 246 + ], + "spans": [ + { + "bbox": [ + 104, + 232, + 511, + 246 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 104, + 232, + 511, + 246 + ], + "type": "text", + "content": "Jiangsu Collaborative Innovation Center for Language Competence, Jiangsu, China" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 60, + 252, + 315, + 364 + ], + "blocks": [ + { + "bbox": [ + 60, + 252, + 315, + 364 + ], + "lines": [ + { + "bbox": [ + 60, + 252, + 315, + 364 + ], + "spans": [ + { + "bbox": [ + 60, + 252, + 315, + 364 + ], + "type": "image", + "image_path": "e7ef0ad57d49aabffbf8bb5e3d9aaee49e8be2cfeddc8d8e0c6db3eb5d4b33f9.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 127, + 366, + 246, + 374 + ], + "lines": [ + { + "bbox": [ + 127, + 366, + 246, + 374 + ], + "spans": [ + { + "bbox": [ + 127, + 366, + 246, + 374 + ], + "type": "text", + "content": "The way models perceive real-world scenarios" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 319, + 253, + 552, + 329 + ], + "blocks": [ + { + "bbox": [ + 319, + 253, + 552, + 329 + ], + "lines": [ + { + "bbox": [ + 319, + 253, + 552, + 329 + ], + "spans": [ + { + "bbox": [ + 319, + 253, + 552, + 329 + ], + "type": "image", + "image_path": "a67683c6cd9e489633cf9362ca9a8745fec18ceaadaee4d0016724819c9cef44.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 317, + 331, + 333, + 347 + ], + "blocks": [ + { + "bbox": [ + 317, + 331, + 333, + 347 + ], + "lines": [ + { + "bbox": [ + 317, + 331, + 333, + 347 + ], + "spans": [ + { + "bbox": [ + 317, + 331, + 333, + 347 + ], + "type": "image", + "image_path": "36e08331fc60a1a87e70a404552c46e77d1fb66fa7ac796d4fc66d460223305f.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 335, + 332, + 514, + 355 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 332, + 514, + 355 + ], + "spans": [ + { + "bbox": [ + 335, + 332, + 514, + 355 + ], + "type": "text", + "content": "Each of the four images contain one car respectively. However 1 2 feature the same car, and 3 4 feature another identical car. Thus, there are only two cars in the scene." + } + ] + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 317, + 357, + 333, + 373 + ], + "blocks": [ + { + "bbox": [ + 317, + 357, + 333, + 373 + ], + "lines": [ + { + "bbox": [ + 317, + 357, + 333, + 373 + ], + "spans": [ + { + "bbox": [ + 317, + 357, + 333, + 373 + ], + "type": "image", + "image_path": "b5f684a0dc7fb59074ce74c40d6e4c7c784d80a08f7b552dcd1e0ff195ada2e0.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 379, + 555, + 412 + ], + "lines": [ + { + "bbox": [ + 54, + 379, + 555, + 412 + ], + "spans": [ + { + "bbox": [ + 54, + 379, + 555, + 412 + ], + "type": "text", + "content": "Figure 1. Illustration of continuous space perception: The left side depicts the construction of an image sequence representing a continuous space. The right side presents an example task related to the continuous space, showing how adjacent images are connected. Note that to answer the question correctly, one must recognize that the cars in images 1 and 2 are the same, as well as those in images 3 and 4." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 335, + 357, + 517, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 357, + 517, + 372 + ], + "spans": [ + { + "bbox": [ + 335, + 357, + 517, + 372 + ], + "type": "text", + "content": "The four images respectively contains one car. Four cars all together are visible in the scene." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 151, + 419, + 200, + 431 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 419, + 200, + 431 + ], + "spans": [ + { + "bbox": [ + 151, + 419, + 200, + 431 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 54, + 450, + 297, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 450, + 297, + 715 + ], + "spans": [ + { + "bbox": [ + 54, + 450, + 297, + 715 + ], + "type": "text", + "content": "Vision-Language Models (VLMs) have recently witnessed significant progress in visual comprehension. As the permitting length of image context grows, VLMs can now comprehend a broader range of views and spaces. Current benchmarks provide insightful analysis of VLMs in tasks involving complex visual instructions following, multi-image understanding and spatial reasoning. However, they usually focus on spatially irrelevant images or discrete images captured from varied viewpoints. The compositional characteristic of images captured from a static viewpoint remains underestimated. We term this characteristic as Continuous Space Perception. When observing a scene from a static viewpoint while shifting orientations, it produces a series of spatially continuous images, enabling the reconstruction of the entire space. In this paper, we present CoSpace, a multi-image visual understanding benchmark designed to assess the Continuous Space perception ability for VLMs. CoSpace contains 2,918 images and 1,626 question-answer pairs, covering seven types of tasks. We conduct evaluation across 19 proprietary and open-source VLMs. Results reveal that there exist pitfalls on the continuous space perception ability for most of the evaluated models, including pro" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 420, + 555, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 420, + 555, + 492 + ], + "spans": [ + { + "bbox": [ + 313, + 420, + 555, + 492 + ], + "type": "text", + "content": "prietary ones. Interestingly, we find that the main discrepancy between open-source and proprietary models lies not in accuracy but in the consistency of responses. We believe that enhancing the ability of continuous space perception is essential for VLMs to perform effectively in real-world tasks and encourage further research to advance this capability." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 314, + 499, + 395, + 510 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 499, + 395, + 510 + ], + "spans": [ + { + "bbox": [ + 314, + 499, + 395, + 510 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 312, + 521, + 555, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 521, + 555, + 652 + ], + "spans": [ + { + "bbox": [ + 312, + 521, + 555, + 652 + ], + "type": "text", + "content": "With the advancements in deep learning and Large Language Models (LLMs), Vision-Language Models (VLMs) have demonstrated exceptional performance across a range of downstream tasks [4, 17, 22, 30, 47, 60]. Driven by this, VLMs are now being used in diverse applications, such as planning [5, 56], navigation [8, 63] and embodied tasks [28, 42], and are further applied in real-world scenarios [3, 61]. Real-world scenario, featuring the complicated environment, cast significant challenges on VLMs, especially the ability to capture sophisticated spatial information from surrounding spaces." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 312, + 654, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 654, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 312, + 654, + 556, + 715 + ], + "type": "text", + "content": "Despite previous attempts to advance spatial understanding in VLMs [4, 38], there remain areas requiring further exploration. Various strategies have been explored to empower models with the ability to identify and ground objects based on spatial relationship [4, 16]. Shiri et al. [47]" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "29569" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 115, + 70, + 494, + 228 + ], + "blocks": [ + { + "bbox": [ + 115, + 70, + 494, + 228 + ], + "lines": [ + { + "bbox": [ + 115, + 70, + 494, + 228 + ], + "spans": [ + { + "bbox": [ + 115, + 70, + 494, + 228 + ], + "type": "table", + "html": "
BenchmarksEvaluation TargetFeaturesNum.ImgEvaluationInstances
Spatial ReasoningMulti-image PerceptionStatic Viewpoint
MMRel [38]Object-relationship○××15k15k
Spatial-MM [47]Object-relationship and spatial-CoT✓××2.3k400
SpatialEval [50]Spatial Reasoning✓××4.6k4.6k
GSR-Bench [41]Grounded spatial reasoning✓××4.9k4.9k
MileBench [9]Multimodal Long Context×✓×97.9k6.4k
Demon-Core [26]Demonstrative instruction following×✓×62.8k18.2k
Mementos-Val [53]Temporal sequential-image understanding×✓×-699
MIBench [33]Fine-grained multi-image understanding×✓×125k13k
BLINK [14]General visual perception○○×7.3k3.8k
SEED-Bench [23]Visual comprehension○○×1.9k24k
Video-MME [13]Visual comprehension○✓×0.9k2.7k
MMIU [37]Visual comprehension○✓×77.7k11.7k
SparklesEval [15]Multi-image Chat○✓×550150
CoSpace (Ours)continuous space perception✓✓✓2,9181,626
", + "image_path": "b9e9b35e9445a6335db31052747050423a76d9b664d966251c12788654e790b0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 230, + 555, + 252 + ], + "lines": [ + { + "bbox": [ + 55, + 230, + 555, + 252 + ], + "spans": [ + { + "bbox": [ + 55, + 230, + 555, + 252 + ], + "type": "text", + "content": "Table 1. Comparison with selected benchmarks that address multi-image evaluation and/or spatial evaluation. " + }, + { + "bbox": [ + 55, + 230, + 555, + 252 + ], + "type": "inline_equation", + "content": "\\checkmark" + }, + { + "bbox": [ + 55, + 230, + 555, + 252 + ], + "type": "text", + "content": " denotes the main focus of the corresponding feature, " + }, + { + "bbox": [ + 55, + 230, + 555, + 252 + ], + "type": "inline_equation", + "content": "\\circ" + }, + { + "bbox": [ + 55, + 230, + 555, + 252 + ], + "type": "text", + "content": " indicates that the feature exists but is not the primary focus, while " + }, + { + "bbox": [ + 55, + 230, + 555, + 252 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 55, + 230, + 555, + 252 + ], + "type": "text", + "content": " implies a lack of the feature." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 259, + 294, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 259, + 294, + 390 + ], + "spans": [ + { + "bbox": [ + 54, + 259, + 294, + 390 + ], + "type": "text", + "content": "and Wang et al. [50] assess spatial understanding ability for VLMs with the recognition of basic tasks in this domain. However, much of these work center on single-image inputs, without further extending to multi-image scenarios. Works on video understanding [52, 60] present the concept of spatial temporal understanding, but the spatial information of videos involve sequential frames that inherently linking spatial understanding with additional temporal dynamics. Given these attempts in spatial understanding, we pose the question: Are the current efforts adequate for advancing the research of spatial understanding?" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 391, + 295, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 391, + 295, + 498 + ], + "spans": [ + { + "bbox": [ + 54, + 391, + 295, + 498 + ], + "type": "text", + "content": "We claim that a gap still exists between the concept proposed in previous work and more advanced spatial understanding. For humans, exposing to an unfamiliar environment requires determining the position, identifying surrounding objects, and making decisions based on spatial context. Such process typically involves integrating various information about the surrounding space. We refer to it as continuous space perception, which is an essential ability in real-world towards effectively completion of daily tasks." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 498, + 295, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 498, + 295, + 641 + ], + "spans": [ + { + "bbox": [ + 54, + 498, + 295, + 641 + ], + "type": "text", + "content": "Figure 1 shows an example of continuous space perception: Conceptually, humans perceive continuous visual space as a series of \"images\", each representing a segment of the space with overlaps between adjacent views. Mimicking this perception, Figure 1 illustrates the continuous visual space using four images. To answer questions accurately about these images requires a comprehensive understanding of all four images and reconstructing the entire space they represent. For example, to accurately count the number of cars in the space, it is essential to recognize that the cars appearing in the first and second images are the same, as are the cars in the third and fourth images." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 642, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 642, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 54, + 642, + 295, + 713 + ], + "type": "text", + "content": "The continuous space perception ability is equally important for VLMs as for human. Many real-world scenario tasks, such as planning and navigation, require models to effectively integrate the information in the surroundings to take the optimal actions. During this process, the reconstruction of the space heavily relies on continuous space" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 259, + 553, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 259, + 553, + 283 + ], + "spans": [ + { + "bbox": [ + 313, + 259, + 553, + 283 + ], + "type": "text", + "content": "perception. However, as shown in Table 1, this ability is largely overlooked in existing benchmarks for VLMs." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 312, + 283, + 555, + 463 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 283, + 555, + 463 + ], + "spans": [ + { + "bbox": [ + 312, + 283, + 555, + 463 + ], + "type": "text", + "content": "To this end, we present CoSpace, a visual understanding benchmark designed to evaluate the Continuous Space perception ability for VLMs. CoSpace includes four categories and seven tasks, emphasizing the following core abilities: (1) Can VLMs recognize direction and position when presented with spatially continuous images captured from a static viewpoint? (2) Can models identify the correspondence between adjacent images in the same scene? (3) Can models make optimal navigation decisions based on the perception of the space? We posit the importance of these three core abilities, with the first one forms the basic understanding of space, the second one is essential for integrating the information across different images and the last one surpasses mere literal understanding and makes a step towards more practical applications." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 312, + 464, + 554, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 464, + 554, + 582 + ], + "spans": [ + { + "bbox": [ + 312, + 464, + 554, + 582 + ], + "type": "text", + "content": "We evaluated 19 widely used open-source and proprietary VLMs on our proposed CoSpace benchmark. Results show that though some of the open-source models achieve approaching or even higher accuracy compared with state-of-the-art proprietary models on certain tasks, proprietary models exhibit remarkably higher consistency in their responses across multi-dimensional evaluation metrics. Additionally, we found that with the alteration of input context and image sequence, the tasks that humans find more challenging do not always pose greater difficulty for the models." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 603, + 400, + 616 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 603, + 400, + 616 + ], + "spans": [ + { + "bbox": [ + 313, + 603, + 400, + 616 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 624, + 527, + 636 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 624, + 527, + 636 + ], + "spans": [ + { + "bbox": [ + 313, + 624, + 527, + 636 + ], + "type": "text", + "content": "2.1. Multi-image Perception Ability of VLMs" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 312, + 641, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 641, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 312, + 641, + 554, + 713 + ], + "type": "text", + "content": "Large Vision-Language Models (VLMs), such as LLaVA models [31, 32] and Sphinx [30], have shown impressive performances on visual reasoning and grounding tasks that focus on only one image at a time. However, it is more natural to allow multiple images and arrange them in interleave format in one input. Along with the growing capability of large vision models, there are also many other applications. For example, the use of large vision models has been demonstrated for large datasets, such as the US Census dataset [33]." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "29570" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 228 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 228 + ], + "type": "text", + "content": "ities of single-image models, multi-image perception and understanding abilities are gaining increasing attentions, including research topics such as multimodal in-context learning [10, 22, 62], interleaved visual instruction following [18, 25, 26], and temporal visual understanding [49], and practical applications such as navigation and embodied QA [63]. To achieve multi-image perception, some methods focus on constructing interleaved visual instruction tuning data [17, 18, 22, 62], some improve the model structure and training strategies for better cross-image understanding [26, 52, 55, 58], and others address the image processing strategy for adaptable visual feature encoding concerning multi-image scenarios [21, 32, 59]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 239, + 294, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 239, + 294, + 323 + ], + "spans": [ + { + "bbox": [ + 55, + 239, + 294, + 323 + ], + "type": "text", + "content": "With the enhancements of the multi-image perception ability, VLMs are able to obtain visual information covering a wider range of views from multiple images at one time, resulting in more profound observations and understandings of the surrounding environment. In our paper, we take this as basis, and dedicate to investigating the spatial perception ability of VLMs." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 361, + 286, + 374 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 361, + 286, + 374 + ], + "spans": [ + { + "bbox": [ + 55, + 361, + 286, + 374 + ], + "type": "text", + "content": "2.2. Spatial Comprehension Evaluation of VLMs" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 391, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 391, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 391, + 295, + 713 + ], + "type": "text", + "content": "For the past years, the evaluation of spatial comprehension for VLMs has always been a commonly engaged topic, and it often refers to the spatial relation recognition among objects, such as MMRel, GSR-Bench and SpatialMM shown in Table 1. Recently, multi-image related evaluation for VLMs are widely explored, including interleaved visual instruction following [26, 33], temporal sequential-image comprehension [53], and long context understanding [9]. In multi-image scenarios, such as video comprehension [13, 23] and multi-image dialogue [15], the spatial reasoning ability is also one of the evaluated aspects. However, these benchmarks primarily focus on general multi-image comprehension where most of input images come from difference scenes. The compositional features across multiple images to integrate richer spatial information of a wider scene and more profound spatial understanding are seldom emphasized. Although Blink [14], ActiView [54], and video-based benchmarks [13, 27] require spatial understanding across multiple images, they either focus on different views of the same object, or discrete views from dynamic viewpoint. These views are not continuous regarding a static viewpoint, and the resulting images are not necessarily bounded regarding the same scene. In contrast to existing works, we address the continuous spatial perception from a static viewpoint, that requires models to connect all the perceived views together to form an intact understanding of the whole scene." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 314, + 72, + 375, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 375, + 85 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 375, + 85 + ], + "type": "text", + "content": "3. CoSpace" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 91, + 509, + 105 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 91, + 509, + 105 + ], + "spans": [ + { + "bbox": [ + 313, + 91, + 509, + 105 + ], + "type": "text", + "content": "3.1. Continuous Space Perception Ability" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 110, + 553, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 110, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 313, + 110, + 553, + 217 + ], + "type": "text", + "content": "In this paper, we define the term, continuous visual space, as the perceived views from a static viewpoint towards different orientations that can naturally compose a complete scene or environment. As is shown in Figure 1, if one would like to correctly answer the question, it is needed to first capture the visual information of all the cars, and consequently, the ability of continuous space perception refers to capturing and processing information from these images of continuous visual spaces." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 218, + 553, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 218, + 553, + 469 + ], + "spans": [ + { + "bbox": [ + 313, + 218, + 553, + 469 + ], + "type": "text", + "content": "To obtain an overall observation of environment, human perceive the whole surroundings from the same viewpoint towards different directions, and the percept continuous visual spaces can be reconstructed as a panorama in the brain. Different from human that directly views the space continuously, models typically take discrete images as input. Therefore, the way to present continuous visual spaces remains flexible. In our CoSpace, we present a complete continuous visual space as follows: we provide several images shot at the same viewpoint in the same scene, but towards different directions, e.g. respectively towards front, right, back and left. These images together form a complete space without omitting any visual information of the surrounding environment. Naturally, such strategy could produce overlapped spaces between adjacent images, which tightly connects images describing adjacent spaces, and can consequently help models better identify the relationships between images from continuous visual space. In this way, we posit that such strategy can effectively preserve continuous spatial information of an extensive space as well as be suitable to understanding for VLMs." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 479, + 394, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 479, + 394, + 491 + ], + "spans": [ + { + "bbox": [ + 314, + 479, + 394, + 491 + ], + "type": "text", + "content": "3.2. Task Design" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 497, + 553, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 497, + 553, + 604 + ], + "spans": [ + { + "bbox": [ + 313, + 497, + 553, + 604 + ], + "type": "text", + "content": "We comprehensively evaluate the continuous space perception ability of VLMs regarding four aspects. Two of them are fundamental ability of spatial perception, detailed as direction perception and recognition (Direction), and rotation-centric perception (Rotation). The other two are further practical tasks, detailed as counting-based perception (Counting), and embodied-based planning (Planning). These four categories are subdivided into seven tasks for elaborate investigation as shown in Figure 2." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 605, + 553, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 553, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 553, + 628 + ], + "type": "text", + "content": "We employ two multiple choice tasks, both with four options, for the Direction category as follows:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 630, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 630, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 314, + 630, + 553, + 713 + ], + "type": "text", + "content": "- Direction Recognition (DIR-Rec): In real world, identifying direction is inevitable when one is placed in a new environment. Similarly, the DIR-Rec task requires models to recognize the direction of targets by answering questions such as \"where is the building located?\" Meanwhile, we provide detailed task instructions, including direction references such as \"the first image is facing north," + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "29571" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 246, + 74, + 277, + 82 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 246, + 74, + 277, + 82 + ], + "spans": [ + { + "bbox": [ + 246, + 74, + 277, + 82 + ], + "type": "text", + "content": "Direction" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 100, + 86, + 211, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 86, + 211, + 95 + ], + "spans": [ + { + "bbox": [ + 100, + 86, + 211, + 95 + ], + "type": "inline_equation", + "content": "\\clubsuit" + }, + { + "bbox": [ + 100, + 86, + 211, + 95 + ], + "type": "text", + "content": " Direction Recognition (DIR-Rec)" + } + ] + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 58, + 95, + 256, + 127 + ], + "blocks": [ + { + "bbox": [ + 58, + 95, + 256, + 127 + ], + "lines": [ + { + "bbox": [ + 58, + 95, + 256, + 127 + ], + "spans": [ + { + "bbox": [ + 58, + 95, + 256, + 127 + ], + "type": "image", + "image_path": "114f03be34d4468d3a10b547709cd68894dcdf6af81a77d3e849293e92ab478e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 58, + 127, + 236, + 135 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 127, + 236, + 135 + ], + "spans": [ + { + "bbox": [ + 58, + 127, + 236, + 135 + ], + "type": "text", + "content": "Question: Which direction is the alleyway going in the images?" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 58, + 135, + 141, + 142 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 135, + 141, + 142 + ], + "spans": [ + { + "bbox": [ + 58, + 135, + 141, + 142 + ], + "type": "text", + "content": "Options: A. South to North" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 89, + 143, + 163, + 149 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 143, + 163, + 149 + ], + "spans": [ + { + "bbox": [ + 89, + 143, + 163, + 149 + ], + "type": "text", + "content": "C. Northeast to Southwest" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 170, + 135, + 215, + 143 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 170, + 135, + 215, + 143 + ], + "spans": [ + { + "bbox": [ + 170, + 135, + 215, + 143 + ], + "type": "text", + "content": "B. East to West" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 173, + 143, + 245, + 149 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 173, + 143, + 245, + 149 + ], + "spans": [ + { + "bbox": [ + 173, + 143, + 245, + 149 + ], + "type": "text", + "content": "D. Northwest to Southeast" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 59, + 151, + 248, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 151, + 248, + 168 + ], + "spans": [ + { + "bbox": [ + 59, + 151, + 248, + 168 + ], + "type": "text", + "content": "Rationale: The " + }, + { + "bbox": [ + 59, + 151, + 248, + 168 + ], + "type": "inline_equation", + "content": "2^{nd}" + }, + { + "bbox": [ + 59, + 151, + 248, + 168 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 59, + 151, + 248, + 168 + ], + "type": "inline_equation", + "content": "4^{th}" + }, + { + "bbox": [ + 59, + 151, + 248, + 168 + ], + "type": "text", + "content": " images indicate a direction of East-West, the " + }, + { + "bbox": [ + 59, + 151, + 248, + 168 + ], + "type": "inline_equation", + "content": "1^{st}" + }, + { + "bbox": [ + 59, + 151, + 248, + 168 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 59, + 151, + 248, + 168 + ], + "type": "inline_equation", + "content": "3^{rd}" + }, + { + "bbox": [ + 59, + 151, + 248, + 168 + ], + "type": "text", + "content": " images indicate the width of the alleyway." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 298, + 86, + 435, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 86, + 435, + 95 + ], + "spans": [ + { + "bbox": [ + 298, + 86, + 435, + 95 + ], + "type": "text", + "content": "Directional Object Perception (DIR-Obj)" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 266, + 95, + 463, + 127 + ], + "blocks": [ + { + "bbox": [ + 266, + 95, + 463, + 127 + ], + "lines": [ + { + "bbox": [ + 266, + 95, + 463, + 127 + ], + "spans": [ + { + "bbox": [ + 266, + 95, + 463, + 127 + ], + "type": "image", + "image_path": "22af0283074d7c8bc2bc1b8252d1b9d164edb5bb734d732fdf9aa9071376d10f.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 269, + 127, + 416, + 135 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 269, + 127, + 416, + 135 + ], + "spans": [ + { + "bbox": [ + 269, + 127, + 416, + 135 + ], + "type": "text", + "content": "Question: What is visible in the southeast direction?" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 269, + 135, + 364, + 142 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 269, + 135, + 364, + 142 + ], + "spans": [ + { + "bbox": [ + 269, + 135, + 364, + 142 + ], + "type": "text", + "content": "Options: A. A pedestrian bridge" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 301, + 143, + 365, + 150 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 301, + 143, + 365, + 150 + ], + "spans": [ + { + "bbox": [ + 301, + 143, + 365, + 150 + ], + "type": "text", + "content": "C. A \"no parking\" sign" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 266, + 151, + 461, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 266, + 151, + 461, + 168 + ], + "spans": [ + { + "bbox": [ + 266, + 151, + 461, + 168 + ], + "type": "text", + "content": "Rationale: Southeast refers to the left part of the " + }, + { + "bbox": [ + 266, + 151, + 461, + 168 + ], + "type": "inline_equation", + "content": "3^{rd}" + }, + { + "bbox": [ + 266, + 151, + 461, + 168 + ], + "type": "text", + "content": " image and the right part of the " + }, + { + "bbox": [ + 266, + 151, + 461, + 168 + ], + "type": "inline_equation", + "content": "4^{th}" + }, + { + "bbox": [ + 266, + 151, + 461, + 168 + ], + "type": "text", + "content": " image, only a pedestrian bridge presents." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 247, + 175, + 276, + 183 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 247, + 175, + 276, + 183 + ], + "spans": [ + { + "bbox": [ + 247, + 175, + 276, + 183 + ], + "type": "text", + "content": "Rotation" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 109, + 186, + 202, + 195 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 109, + 186, + 202, + 195 + ], + "spans": [ + { + "bbox": [ + 109, + 186, + 202, + 195 + ], + "type": "text", + "content": "Rotation-Angle (ROT-Ang)" + } + ] + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 58, + 195, + 257, + 228 + ], + "blocks": [ + { + "bbox": [ + 58, + 195, + 257, + 228 + ], + "lines": [ + { + "bbox": [ + 58, + 195, + 257, + 228 + ], + "spans": [ + { + "bbox": [ + 58, + 195, + 257, + 228 + ], + "type": "image", + "image_path": "1554f2b582b6c91996922e0d27a13e6890690504b76f6731e929484cc98020fc.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "bbox": [ + 59, + 231, + 220, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 231, + 220, + 239 + ], + "spans": [ + { + "bbox": [ + 59, + 231, + 220, + 239 + ], + "type": "text", + "content": "Question: What is the angle between the adjacent image?" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 59, + 239, + 154, + 249 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 239, + 154, + 249 + ], + "spans": [ + { + "bbox": [ + 59, + 239, + 154, + 249 + ], + "type": "text", + "content": "Options: A 30 B.60" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 59, + 250, + 253, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 250, + 253, + 266 + ], + "spans": [ + { + "bbox": [ + 59, + 250, + 253, + 266 + ], + "type": "text", + "content": "Rationale: Observing from the overlapping areas of the tall building, and the displacement of the short buildings. 30 degree is more feasible." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 315, + 186, + 419, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 186, + 419, + 194 + ], + "spans": [ + { + "bbox": [ + 315, + 186, + 419, + 194 + ], + "type": "text", + "content": "Rotation-Difference (ROT-Dif)" + } + ] + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 268, + 194, + 463, + 240 + ], + "blocks": [ + { + "bbox": [ + 268, + 194, + 463, + 240 + ], + "lines": [ + { + "bbox": [ + 268, + 194, + 463, + 240 + ], + "spans": [ + { + "bbox": [ + 268, + 194, + 463, + 240 + ], + "type": "image", + "image_path": "d76124066281e59112b3b9b77dc8421710c641d62c0b920bff0db435f17796f1.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "bbox": [ + 269, + 241, + 451, + 249 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 269, + 241, + 451, + 249 + ], + "spans": [ + { + "bbox": [ + 269, + 241, + 451, + 249 + ], + "type": "text", + "content": "Question: Which image does not belong to this image sequence?" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 266, + 250, + 441, + 267 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 266, + 250, + 441, + 267 + ], + "spans": [ + { + "bbox": [ + 266, + 250, + 441, + 267 + ], + "type": "text", + "content": "Rationale: Observing from the range of the overlaps regarding the road., image 4 present the most rotation degree." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 247, + 273, + 276, + 281 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 247, + 273, + 276, + 281 + ], + "spans": [ + { + "bbox": [ + 247, + 273, + 276, + 281 + ], + "type": "text", + "content": "Counting" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 233, + 284, + 288, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 233, + 284, + 288, + 293 + ], + "spans": [ + { + "bbox": [ + 233, + 284, + 288, + 293 + ], + "type": "text", + "content": "- Counting (CNT)" + } + ] + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 59, + 294, + 116, + 325 + ], + "blocks": [ + { + "bbox": [ + 59, + 294, + 116, + 325 + ], + "lines": [ + { + "bbox": [ + 59, + 294, + 116, + 325 + ], + "spans": [ + { + "bbox": [ + 59, + 294, + 116, + 325 + ], + "type": "image", + "image_path": "7e2c0df7eaaf10792f257ec4dfb6d65b5fcd55f1a7c57ebd699ce0e31dc10731.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 445, + 553, + 468 + ], + "lines": [ + { + "bbox": [ + 55, + 445, + 553, + 468 + ], + "spans": [ + { + "bbox": [ + 55, + 445, + 553, + 468 + ], + "type": "text", + "content": "Figure 2. Examples of CoSpace, along with prompt templates for evaluation. We also provide extra visual guidance and textual rationales for readers (invisible to models during evaluation) for easier understanding." + } + ] + } + ], + "index": 91, + "angle": 0, + "type": "image_caption" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 117, + 297, + 187, + 326 + ], + "blocks": [ + { + "bbox": [ + 117, + 297, + 187, + 326 + ], + "lines": [ + { + "bbox": [ + 117, + 297, + 187, + 326 + ], + "spans": [ + { + "bbox": [ + 117, + 297, + 187, + 326 + ], + "type": "image", + "image_path": "9f29c1ad98575cb90fc026b7a4b90c2ed3e8c633420162a9c15cbc9af4fd3a67.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 188, + 297, + 257, + 332 + ], + "blocks": [ + { + "bbox": [ + 188, + 297, + 257, + 332 + ], + "lines": [ + { + "bbox": [ + 188, + 297, + 257, + 332 + ], + "spans": [ + { + "bbox": [ + 188, + 297, + 257, + 332 + ], + "type": "image", + "image_path": "9b6656572a95637c9279c6210c83597fc61f408ad64447af2c03086f967d072a.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 258, + 297, + 349, + 332 + ], + "blocks": [ + { + "bbox": [ + 258, + 297, + 349, + 332 + ], + "lines": [ + { + "bbox": [ + 258, + 297, + 349, + 332 + ], + "spans": [ + { + "bbox": [ + 258, + 297, + 349, + 332 + ], + "type": "image", + "image_path": "5d8997331dfde03496b702d1ffa3509f97b3ba5c6be9677062be56ec6035ff9c.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + } + ], + "index": 30 + }, + { + "bbox": [ + 355, + 292, + 394, + 302 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 355, + 292, + 394, + 302 + ], + "spans": [ + { + "bbox": [ + 355, + 292, + 394, + 302 + ], + "type": "text", + "content": "Rationale:" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 354, + 304, + 457, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 304, + 457, + 319 + ], + "spans": [ + { + "bbox": [ + 354, + 304, + 457, + 319 + ], + "type": "text", + "content": "One red truck spotted in the " + }, + { + "bbox": [ + 354, + 304, + 457, + 319 + ], + "type": "inline_equation", + "content": "3^{rd}" + }, + { + "bbox": [ + 354, + 304, + 457, + 319 + ], + "type": "text", + "content": " image, and it is spotted again in the " + }, + { + "bbox": [ + 354, + 304, + 457, + 319 + ], + "type": "inline_equation", + "content": "4^{th}" + }, + { + "bbox": [ + 354, + 304, + 457, + 319 + ], + "type": "text", + "content": " image." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 355, + 320, + 459, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 355, + 320, + 459, + 327 + ], + "spans": [ + { + "bbox": [ + 355, + 320, + 459, + 327 + ], + "type": "text", + "content": "There is indeed 1 red truck in the scene." + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 59, + 334, + 218, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 334, + 218, + 342 + ], + "spans": [ + { + "bbox": [ + 59, + 334, + 218, + 342 + ], + "type": "text", + "content": "Question: How many red trucks are visible in the scene?" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 282, + 335, + 313, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 282, + 335, + 313, + 342 + ], + "spans": [ + { + "bbox": [ + 282, + 335, + 313, + 342 + ], + "type": "text", + "content": "Answer: 1" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 376, + 334, + 463, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 376, + 334, + 463, + 342 + ], + "spans": [ + { + "bbox": [ + 376, + 334, + 463, + 342 + ], + "type": "text", + "content": "Note that " + }, + { + "bbox": [ + 376, + 334, + 463, + 342 + ], + "type": "inline_equation", + "content": "\\mathbb{O}" + }, + { + "bbox": [ + 376, + 334, + 463, + 342 + ], + "type": "text", + "content": " denotes repeated cars" + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 247, + 348, + 275, + 357 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 247, + 348, + 275, + 357 + ], + "spans": [ + { + "bbox": [ + 247, + 348, + 275, + 357 + ], + "type": "text", + "content": "Planning" + } + ] + } + ], + "index": 37 + }, + { + "type": "image", + "bbox": [ + 58, + 358, + 257, + 398 + ], + "blocks": [ + { + "bbox": [ + 58, + 358, + 257, + 398 + ], + "lines": [ + { + "bbox": [ + 58, + 358, + 257, + 398 + ], + "spans": [ + { + "bbox": [ + 58, + 358, + 257, + 398 + ], + "type": "image", + "image_path": "689c430bf92b78135be9e9764e13e542626ef1dde256d5e0e0831d2fa33799b3.jpg" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_body" + } + ], + "index": 38 + }, + { + "bbox": [ + 60, + 399, + 226, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 399, + 226, + 407 + ], + "spans": [ + { + "bbox": [ + 60, + 399, + 226, + 407 + ], + "type": "text", + "content": "Question: Where is the television regarding your position?" + } + ] + } + ], + "index": 39 + }, + { + "bbox": [ + 61, + 407, + 129, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 407, + 129, + 415 + ], + "spans": [ + { + "bbox": [ + 61, + 407, + 129, + 415 + ], + "type": "text", + "content": "Options: A. Front-left" + } + ] + } + ], + "index": 40 + }, + { + "bbox": [ + 93, + 416, + 133, + 423 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 416, + 133, + 423 + ], + "spans": [ + { + "bbox": [ + 93, + 416, + 133, + 423 + ], + "type": "text", + "content": "C. Front-right" + } + ] + } + ], + "index": 41 + }, + { + "bbox": [ + 170, + 407, + 205, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 170, + 407, + 205, + 415 + ], + "spans": [ + { + "bbox": [ + 170, + 407, + 205, + 415 + ], + "type": "text", + "content": "B. Back-left" + } + ] + } + ], + "index": 42 + }, + { + "bbox": [ + 170, + 415, + 210, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 170, + 415, + 210, + 424 + ], + "spans": [ + { + "bbox": [ + 170, + 415, + 210, + 424 + ], + "type": "text", + "content": "D. Back-right" + } + ] + } + ], + "index": 43 + }, + { + "bbox": [ + 59, + 424, + 247, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 424, + 247, + 440 + ], + "spans": [ + { + "bbox": [ + 59, + 424, + 247, + 440 + ], + "type": "text", + "content": "Rationale: TV is placed in the right part of " + }, + { + "bbox": [ + 59, + 424, + 247, + 440 + ], + "type": "inline_equation", + "content": "2^{nd}" + }, + { + "bbox": [ + 59, + 424, + 247, + 440 + ], + "type": "text", + "content": " image. Right part of the right view indicate back-right direction." + } + ] + } + ], + "index": 44 + }, + { + "type": "image", + "bbox": [ + 266, + 360, + 463, + 392 + ], + "blocks": [ + { + "bbox": [ + 266, + 360, + 463, + 392 + ], + "lines": [ + { + "bbox": [ + 266, + 360, + 463, + 392 + ], + "spans": [ + { + "bbox": [ + 266, + 360, + 463, + 392 + ], + "type": "image", + "image_path": "7af29df7f5feca7d9de3d7bb07ca96bbb2cd6c49585cbe7ac18293aa3e50afff.jpg" + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "image_body" + } + ], + "index": 45 + }, + { + "bbox": [ + 317, + 392, + 412, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 392, + 412, + 399 + ], + "spans": [ + { + "bbox": [ + 317, + 392, + 412, + 399 + ], + "type": "text", + "content": "- Planning-Decision (PLA-Dec)" + } + ] + } + ], + "index": 46 + }, + { + "bbox": [ + 269, + 399, + 419, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 269, + 399, + 419, + 407 + ], + "spans": [ + { + "bbox": [ + 269, + 399, + 419, + 407 + ], + "type": "text", + "content": "Question: What should you do to reach the fireplace?" + } + ] + } + ], + "index": 47 + }, + { + "bbox": [ + 269, + 407, + 376, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 269, + 407, + 376, + 415 + ], + "spans": [ + { + "bbox": [ + 269, + 407, + 376, + 415 + ], + "type": "text", + "content": "Options: A. Turn right and go ahead" + } + ] + } + ], + "index": 48 + }, + { + "bbox": [ + 300, + 416, + 372, + 423 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 300, + 416, + 372, + 423 + ], + "spans": [ + { + "bbox": [ + 300, + 416, + 372, + 423 + ], + "type": "text", + "content": "C. Turn left and go ahead" + } + ] + } + ], + "index": 49 + }, + { + "bbox": [ + 383, + 407, + 460, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 383, + 407, + 460, + 415 + ], + "spans": [ + { + "bbox": [ + 383, + 407, + 460, + 415 + ], + "type": "text", + "content": "B. Turn back and go ahead" + } + ] + } + ], + "index": 50 + }, + { + "bbox": [ + 384, + 416, + 459, + 423 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 384, + 416, + 459, + 423 + ], + "spans": [ + { + "bbox": [ + 384, + 416, + 459, + 423 + ], + "type": "text", + "content": "D. Go ahead and turn back" + } + ] + } + ], + "index": 51 + }, + { + "bbox": [ + 266, + 424, + 455, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 266, + 424, + 455, + 440 + ], + "spans": [ + { + "bbox": [ + 266, + 424, + 455, + 440 + ], + "type": "text", + "content": "Rationale:Fireplace locates in the far back, requiring both a turn action and a moving action, need to first turn backwards and then go ahead" + } + ] + } + ], + "index": 52 + }, + { + "type": "image", + "bbox": [ + 468, + 74, + 479, + 86 + ], + "blocks": [ + { + "bbox": [ + 468, + 74, + 479, + 86 + ], + "lines": [ + { + "bbox": [ + 468, + 74, + 479, + 86 + ], + "spans": [ + { + "bbox": [ + 468, + 74, + 479, + 86 + ], + "type": "image", + "image_path": "808e7415e17194c8358cfca03e7e3d73fae6482126560c3aa864213c6cac7e9f.jpg" + } + ] + } + ], + "index": 53, + "angle": 0, + "type": "image_body" + } + ], + "index": 53 + }, + { + "bbox": [ + 487, + 77, + 543, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 487, + 77, + 543, + 85 + ], + "spans": [ + { + "bbox": [ + 487, + 77, + 543, + 85 + ], + "type": "text", + "content": "Task instructions" + } + ] + } + ], + "index": 54 + }, + { + "bbox": [ + 468, + 88, + 547, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 468, + 88, + 547, + 95 + ], + "spans": [ + { + "bbox": [ + 468, + 88, + 547, + 95 + ], + "type": "text", + "content": "General prompt template:" + } + ] + } + ], + "index": 55 + }, + { + "bbox": [ + 468, + 96, + 513, + 102 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 468, + 96, + 513, + 102 + ], + "spans": [ + { + "bbox": [ + 468, + 96, + 513, + 102 + ], + "type": "text", + "content": "" + } + ] + } + ], + "index": 56 + }, + { + "bbox": [ + 468, + 103, + 545, + 110 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 468, + 103, + 545, + 110 + ], + "spans": [ + { + "bbox": [ + 468, + 103, + 545, + 110 + ], + "type": "text", + "content": "" + } + ] + } + ], + "index": 57 + }, + { + "bbox": [ + 468, + 111, + 522, + 118 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 468, + 111, + 522, + 118 + ], + "spans": [ + { + "bbox": [ + 468, + 111, + 522, + 118 + ], + "type": "text", + "content": "" + } + ] + } + ], + "index": 58 + }, + { + "bbox": [ + 468, + 118, + 524, + 125 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 468, + 118, + 524, + 125 + ], + "spans": [ + { + "bbox": [ + 468, + 118, + 524, + 125 + ], + "type": "text", + "content": " " + } + ] + } + ], + "index": 59 + }, + { + "bbox": [ + 468, + 125, + 525, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 468, + 125, + 525, + 133 + ], + "spans": [ + { + "bbox": [ + 468, + 125, + 525, + 133 + ], + "type": "text", + "content": "You should generate" + } + ] + } + ], + "index": 60 + }, + { + "bbox": [ + 468, + 134, + 545, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 468, + 134, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 468, + 134, + 545, + 140 + ], + "type": "text", + "content": "answer from A, B, C or D." + } + ] + } + ], + "index": 61 + }, + { + "bbox": [ + 469, + 140, + 506, + 146 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 469, + 140, + 506, + 146 + ], + "spans": [ + { + "bbox": [ + 469, + 140, + 506, + 146 + ], + "type": "text", + "content": "Your answer:" + } + ] + } + ], + "index": 62 + }, + { + "bbox": [ + 468, + 152, + 542, + 159 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 468, + 152, + 542, + 159 + ], + "spans": [ + { + "bbox": [ + 468, + 152, + 542, + 159 + ], + "type": "text", + "content": "" + } + ] + } + ], + "index": 63 + }, + { + "bbox": [ + 468, + 160, + 539, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 468, + 160, + 539, + 167 + ], + "spans": [ + { + "bbox": [ + 468, + 160, + 539, + 167 + ], + "type": "text", + "content": "for Direction and Counting" + } + ] + } + ], + "index": 64 + }, + { + "bbox": [ + 468, + 167, + 496, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 468, + 167, + 496, + 175 + ], + "spans": [ + { + "bbox": [ + 468, + 167, + 496, + 175 + ], + "type": "text", + "content": "categories:" + } + ] + } + ], + "index": 65 + }, + { + "bbox": [ + 468, + 175, + 544, + 182 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 468, + 175, + 544, + 182 + ], + "spans": [ + { + "bbox": [ + 468, + 175, + 544, + 182 + ], + "type": "text", + "content": "the first image is facing north," + } + ] + } + ], + "index": 66 + }, + { + "bbox": [ + 468, + 182, + 548, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 468, + 182, + 548, + 189 + ], + "spans": [ + { + "bbox": [ + 468, + 182, + 548, + 189 + ], + "type": "text", + "content": "the second facing east, the third" + } + ] + } + ], + "index": 67 + }, + { + "bbox": [ + 468, + 190, + 536, + 196 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 468, + 190, + 536, + 196 + ], + "spans": [ + { + "bbox": [ + 468, + 190, + 536, + 196 + ], + "type": "text", + "content": "facing south and the fourth" + } + ] + } + ], + "index": 68 + }, + { + "bbox": [ + 468, + 197, + 499, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 468, + 197, + 499, + 203 + ], + "spans": [ + { + "bbox": [ + 468, + 197, + 499, + 203 + ], + "type": "text", + "content": "facing west." + } + ] + } + ], + "index": 69 + }, + { + "bbox": [ + 468, + 211, + 542, + 218 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 468, + 211, + 542, + 218 + ], + "spans": [ + { + "bbox": [ + 468, + 211, + 542, + 218 + ], + "type": "text", + "content": "" + } + ] + } + ], + "index": 70 + }, + { + "bbox": [ + 468, + 219, + 525, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 468, + 219, + 525, + 224 + ], + "spans": [ + { + "bbox": [ + 468, + 219, + 525, + 224 + ], + "type": "text", + "content": "for Rotation category:" + } + ] + } + ], + "index": 71 + }, + { + "bbox": [ + 468, + 225, + 548, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 468, + 225, + 548, + 232 + ], + "spans": [ + { + "bbox": [ + 468, + 225, + 548, + 232 + ], + "type": "text", + "content": "images are arranged clockwise ." + } + ] + } + ], + "index": 72 + }, + { + "bbox": [ + 468, + 233, + 522, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 468, + 233, + 522, + 240 + ], + "spans": [ + { + "bbox": [ + 468, + 233, + 522, + 240 + ], + "type": "text", + "content": "for example, after tak" + } + ] + } + ], + "index": 73 + }, + { + "bbox": [ + 468, + 241, + 543, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 468, + 241, + 543, + 247 + ], + "spans": [ + { + "bbox": [ + 468, + 241, + 543, + 247 + ], + "type": "text", + "content": "first image, the photographer" + } + ] + } + ], + "index": 74 + }, + { + "bbox": [ + 468, + 248, + 548, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 468, + 248, + 548, + 255 + ], + "spans": [ + { + "bbox": [ + 468, + 248, + 548, + 255 + ], + "type": "text", + "content": "turns a certain degree clockwise" + } + ] + } + ], + "index": 75 + }, + { + "bbox": [ + 468, + 255, + 499, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 468, + 255, + 499, + 262 + ], + "spans": [ + { + "bbox": [ + 468, + 255, + 499, + 262 + ], + "type": "text", + "content": "for capturin" + } + ] + } + ], + "index": 76 + }, + { + "bbox": [ + 468, + 270, + 542, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 468, + 270, + 542, + 277 + ], + "spans": [ + { + "bbox": [ + 468, + 270, + 542, + 277 + ], + "type": "text", + "content": "" + } + ] + } + ], + "index": 77 + }, + { + "bbox": [ + 468, + 278, + 525, + 284 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 468, + 278, + 525, + 284 + ], + "spans": [ + { + "bbox": [ + 468, + 278, + 525, + 284 + ], + "type": "text", + "content": "for Planning category:" + } + ] + } + ], + "index": 78 + }, + { + "bbox": [ + 468, + 285, + 547, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 468, + 285, + 547, + 327 + ], + "spans": [ + { + "bbox": [ + 468, + 285, + 547, + 327 + ], + "type": "text", + "content": "you are facing the direction in the first image, with the second image showing the scene to your right, the third image showing the scene at your back, and the fourth image showing the scene to your left." + } + ] + } + ], + "index": 79 + }, + { + "type": "image", + "bbox": [ + 468, + 335, + 479, + 346 + ], + "blocks": [ + { + "bbox": [ + 468, + 335, + 479, + 346 + ], + "lines": [ + { + "bbox": [ + 468, + 335, + 479, + 346 + ], + "spans": [ + { + "bbox": [ + 468, + 335, + 479, + 346 + ], + "type": "image", + "image_path": "983e9298d2aa988e14b08eb46fb0686ef916722a18575f904eaa277612d43c8d.jpg" + } + ] + } + ], + "index": 80, + "angle": 0, + "type": "image_body" + } + ], + "index": 80 + }, + { + "bbox": [ + 484, + 338, + 545, + 346 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 484, + 338, + 545, + 346 + ], + "spans": [ + { + "bbox": [ + 484, + 338, + 545, + 346 + ], + "type": "text", + "content": "Notes for readers" + } + ] + } + ], + "index": 81 + }, + { + "bbox": [ + 468, + 351, + 542, + 358 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 468, + 351, + 542, + 358 + ], + "spans": [ + { + "bbox": [ + 468, + 351, + 542, + 358 + ], + "type": "text", + "content": "- The directional symbols:" + } + ] + } + ], + "index": 82 + }, + { + "type": "image", + "bbox": [ + 474, + 360, + 538, + 376 + ], + "blocks": [ + { + "bbox": [ + 474, + 360, + 538, + 376 + ], + "lines": [ + { + "bbox": [ + 474, + 360, + 538, + 376 + ], + "spans": [ + { + "bbox": [ + 474, + 360, + 538, + 376 + ], + "type": "image", + "image_path": "d409c9bb24490783d6a278550e46775252122cb6e2e2850dc17a1ef23b5001f0.jpg" + } + ] + } + ], + "index": 83, + "angle": 0, + "type": "image_body" + } + ], + "index": 83 + }, + { + "bbox": [ + 466, + 377, + 548, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 466, + 377, + 548, + 384 + ], + "spans": [ + { + "bbox": [ + 466, + 377, + 548, + 384 + ], + "type": "text", + "content": "are invisible to models. They" + } + ] + } + ], + "index": 84 + }, + { + "bbox": [ + 466, + 385, + 541, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 466, + 385, + 541, + 392 + ], + "spans": [ + { + "bbox": [ + 466, + 385, + 541, + 392 + ], + "type": "text", + "content": "are used for easier under" + } + ] + } + ], + "index": 85 + }, + { + "bbox": [ + 466, + 393, + 539, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 466, + 393, + 539, + 399 + ], + "spans": [ + { + "bbox": [ + 466, + 393, + 539, + 399 + ], + "type": "text", + "content": "standing of image orders." + } + ] + } + ], + "index": 86 + }, + { + "bbox": [ + 466, + 406, + 530, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 466, + 406, + 530, + 413 + ], + "spans": [ + { + "bbox": [ + 466, + 406, + 530, + 413 + ], + "type": "text", + "content": "- The Rationale is also" + } + ] + } + ], + "index": 87 + }, + { + "bbox": [ + 466, + 414, + 544, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 466, + 414, + 544, + 421 + ], + "spans": [ + { + "bbox": [ + 466, + 414, + 544, + 421 + ], + "type": "text", + "content": "invisible to models, as they" + } + ] + } + ], + "index": 88 + }, + { + "bbox": [ + 466, + 422, + 538, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 466, + 422, + 538, + 429 + ], + "spans": [ + { + "bbox": [ + 466, + 422, + 538, + 429 + ], + "type": "text", + "content": "intend for explaining the" + } + ] + } + ], + "index": 89 + }, + { + "bbox": [ + 466, + 429, + 525, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 466, + 429, + 525, + 435 + ], + "spans": [ + { + "bbox": [ + 466, + 429, + 525, + 435 + ], + "type": "text", + "content": "results for readers." + } + ] + } + ], + "index": 90 + }, + { + "bbox": [ + 63, + 479, + 295, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 479, + 295, + 525 + ], + "spans": [ + { + "bbox": [ + 63, + 479, + 295, + 525 + ], + "type": "text", + "content": "the second image is facing east\". We standardize 8 directional options, including four cardinal directions, North, East, South and West, and four intercardinal directions, Northeast, Southeast, Southwest, and Northwest." + } + ] + } + ], + "index": 92 + }, + { + "bbox": [ + 55, + 526, + 295, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 526, + 295, + 694 + ], + "spans": [ + { + "bbox": [ + 55, + 526, + 295, + 694 + ], + "type": "text", + "content": "- Directional Object Perception (DIR-Obj): Some real-world tasks demand abilities far beyond merely identifying the direction of a specified target. It is also a critical ability to distinguish content regarding different directions and to align objects with correct directions. Given a specified direction, the DIR-Obj task requires models to be aware of the appearing objects, and also distinguish those absent from that direction. To achieve this, models need to first reconstruct the entire space from continuous views, and then correctly identify the objects dedicate to a given direction. Specifically, in this task, questions typically focus on the intercardinal directions such as \"What is visible in the southeast direction?\", which necessitates a more comprehensive understanding of the entire space." + } + ] + } + ], + "index": 93 + }, + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "type": "text", + "content": "Observing the surroundings from a static viewpoint natu" + } + ] + } + ], + "index": 94 + }, + { + "bbox": [ + 313, + 479, + 553, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 479, + 553, + 515 + ], + "spans": [ + { + "bbox": [ + 313, + 479, + 553, + 515 + ], + "type": "text", + "content": "rally involves rotating the perspectives. Therefore, we summaries Rotation as another fundamental category of continuous space perception, with two tasks designed as follows:" + } + ] + } + ], + "index": 95 + }, + { + "bbox": [ + 314, + 522, + 554, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 522, + 554, + 689 + ], + "spans": [ + { + "bbox": [ + 314, + 522, + 554, + 689 + ], + "type": "text", + "content": "- Rotation-Angle (ROT-Ang): For an observer, when rotating the perspective for more spatial information, there could be overlaps between adjacent observations. Specifically, human can tell the approximate turning angle between consecutive images observed from a static viewpoint by noticing identical furnishings appearing across images. In this task, we investigate if models exhibit similar ability as human does, which requires fine-grained continuous spatial understanding. We apply unified question \"What is the turning angle between the adjacent image?\" and provide two candidate options for model. This task requires models to carefully exam the overlapping and different areas between adjacent images and the shared features in spatially continuous image sequence." + } + ] + } + ], + "index": 96 + }, + { + "bbox": [ + 314, + 689, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 689, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 314, + 689, + 553, + 713 + ], + "type": "text", + "content": "- Rotation-Difference (ROT-Dif): Following the ROT-Ang task, we also investigate the model ability to iden" + } + ] + } + ], + "index": 97 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "29572" + } + ] + } + ], + "index": 98 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 63, + 72, + 295, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 72, + 295, + 179 + ], + "spans": [ + { + "bbox": [ + 63, + 72, + 295, + 179 + ], + "type": "text", + "content": "tify distinguished rotation angles. For the ROT-Dif task, models are given a sequence of five images, four of which share the same turning angle and the other is exceptional. Compared to ROT-Ang, this task emphasizes more on the global understanding of images regarding continuous visual space. The unified question of this task is \"Which image does not belong to this image sequence?\" It is an open-ended question without options, where models are required to answer with index of the exceptional image." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 180, + 295, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 180, + 295, + 252 + ], + "spans": [ + { + "bbox": [ + 55, + 180, + 295, + 252 + ], + "type": "text", + "content": "On the basis of the above two fundamental categories, we further challenge models with two practical categories. The Counting category addresses the occurrence of objects in the entire scene. The Planning category comprises two spatial-related tasks widely evaluated in embodied AI tasks. For the Counting category, the task is described as:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 255, + 295, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 255, + 295, + 434 + ], + "spans": [ + { + "bbox": [ + 55, + 255, + 295, + 434 + ], + "type": "text", + "content": "- Counting (CNT): Counting is a widely adopted task in existing visual benchmarks that asks models to recognize objects and deduce the time of occurrence of target objects. Generally, existing counting tasks only require models to deal with a single image or multiple spatially discrete images, while in our benchmark, images from continuous visual space where identical objects can occur in multiple images are focused on. This raise a challenge for models to not only recognize targets and count for their occurrences, but also be aware of the existence of the same object across different images. To achieve this, models should locate the overlapping area of adjacent images and align the same object appearing different images. The CNT task is an open-ended task, where models should respond with the total count of the target." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 437, + 295, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 437, + 295, + 568 + ], + "spans": [ + { + "bbox": [ + 55, + 437, + 295, + 568 + ], + "type": "text", + "content": "The ability of continuous space perception is also significant in embodied scenarios. Intuitively, regarding the location of an embodied agent as the static viewpoint, it should have a profound understanding of the overall surroundings to effectively solve questions, and to efficiently plan for the following action sequence. Embodied question-answering (EQA) [36, 45] and navigation [19, 20] are typical tasks that require understand continuous visual space. Following these, we split the Planning category into two tasks addressing the relative locations of targets regarding the VLM-based agent as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 570, + 295, + 714 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 56, + 570, + 295, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 570, + 295, + 665 + ], + "spans": [ + { + "bbox": [ + 56, + 570, + 295, + 665 + ], + "type": "text", + "content": "- Planning-Question Answering (PLA-QA): Following the implementation of EQA, we develop the PLA-QA task, requiring models to identify the location of a certain object given a continuous embodied space. In this task, instructions like \"Where is the television regarding your position?\" are provided for the models, and we formulate four options, containing candidate directions relative to the agent, for each question." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 665, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 665, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 665, + 295, + 714 + ], + "type": "text", + "content": "- Planning-Decision (PLA-Dec): This task further investigates the understanding of continuous visual space by asking models to select the proper route to reach the target object. For disambiguation, we standardize the action" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "type": "image", + "bbox": [ + 348, + 72, + 530, + 205 + ], + "blocks": [ + { + "bbox": [ + 348, + 72, + 530, + 205 + ], + "lines": [ + { + "bbox": [ + 348, + 72, + 530, + 205 + ], + "spans": [ + { + "bbox": [ + 348, + 72, + 530, + 205 + ], + "type": "image", + "image_path": "dc56adf0e9888a06d753626b2fd0a10512cdb11acb5eaa2037d09011a3b445f1.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 321, + 208, + 545, + 219 + ], + "lines": [ + { + "bbox": [ + 321, + 208, + 545, + 219 + ], + "spans": [ + { + "bbox": [ + 321, + 208, + 545, + 219 + ], + "type": "text", + "content": "Figure 3. Distribution of categories and tasks in our CoSpace." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 321, + 233, + 553, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 233, + 553, + 304 + ], + "spans": [ + { + "bbox": [ + 321, + 233, + 553, + 304 + ], + "type": "text", + "content": "space as turning (turning to other directions without displacement) and go ahead. The PLA-Dec task especially focus on the order of actions. As is shown in the case in Figure 2, the options \"B. Turn back and go ahead\" and \"D. Go ahead and turn back\" represent two totally different actions and end up in different positions." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 319, + 474, + 330 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 319, + 474, + 330 + ], + "spans": [ + { + "bbox": [ + 314, + 319, + 474, + 330 + ], + "type": "text", + "content": "3.3. Data Collection and Statistics" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 338, + 553, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 338, + 553, + 422 + ], + "spans": [ + { + "bbox": [ + 313, + 338, + 553, + 422 + ], + "type": "text", + "content": "The construction of CoSpace follows the collect-filter-annotate pipeline. We first adopt automatic collection approaches to gather images, and then manually select images that satisfy detailed requirements of our designed tasks and generate questions for these images. Finally human annotators provide answers to each questions. The process is detailed in the following of this section." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 424, + 554, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 424, + 554, + 652 + ], + "spans": [ + { + "bbox": [ + 313, + 424, + 554, + 652 + ], + "type": "text", + "content": "Firstly, to automatically collect sets of images captured in the same scene at static viewpoint but towards different directions, we employ two primary sources, Baidu Map Panorama API that allows us to acquire data towards any desired directions in the real world scene, and HM3D [43], a dataset of 1,000 building-scale 3D reconstructions for embodied tasks. For HM3D, by using Habitat-Sim [40, 46, 48], we instruct an agent wandering in the habitats, we are able to collect images of any views through the attached camera. Secondly, we manually filtered the auto-collected images according to our task requirements, ensuring that the selected images all contain sufficient visual information. With the filtered images, we employed GPT-4o to generate questions via few-shot prompting for the required tasks. Finally, human annotators, an average of two annotators per question, provided answers to these generated questions. This process allowed us to discard vague or unanswerable questions, ensuring that only high-quality question-answer pairs were included in our proposed CoSpace." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "text", + "content": "Our CoSpace comprises a total of 1,626 question-answer pairs and 2,918 images. Among the images, 2,302 are accessed from Baidu Map API and 616 are derived from HM3D dataset. The distribution of instances over categories and tasks is shown in Figure 3." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "29573" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 93, + 70, + 517, + 284 + ], + "blocks": [ + { + "bbox": [ + 93, + 70, + 517, + 284 + ], + "lines": [ + { + "bbox": [ + 93, + 70, + 517, + 284 + ], + "spans": [ + { + "bbox": [ + 93, + 70, + 517, + 284 + ], + "type": "table", + "html": "
ModelsDIR-RecDIR-ObjCNTROT-AngROT-DifPLA-QAPLA-DecAverage
ACCqACCpACCqACCpACCqACCpACCACCACCACC
Proprietary Models
Claude-3.7-sonnet [2]44.4029.2045.6035.6045.0038.0064.3393.5054.7369.3451.97
Gemini-1.5-pro [44]37.6015.6040.6031.6038.2524.0059.6782.0051.6462.9144.39
GPT-4o [39]40.4022.8046.0032.0040.0023.5058.3350.5053.0554.4642.10
>70B Open-source Models
InternVL2_5-78B [6]32.2024.4054.4047.6051.2543.0050.0077.0032.3942.7245.50
Qwen2-VL-72B [51]31.0023.2053.6044.8044.7537.0050.0062.0042.7259.1544.82
InternVL2-76B [7]33.6013.2046.4039.2050.0043.5050.0023.0025.3530.0535.43
LLaVA-OneVision-72B [24]19.2010.8044.2034.4024.2519.0050.0026.0032.8630.9929.17
<13B Open-source Models
MiniCPM-V 2.6 [57]32.8021.2040.4031.6038.5031.5050.0056.0041.3127.7037.10
Qwen2-VL-7B [51]26.4016.4039.2031.6045.0036.0050.0051.5034.2726.7635.71
Mantis-8B [18]30.6024.4031.2028.0041.7536.5050.0039.5033.3327.7034.30
InternVL2-8B [7]29.0012.8037.2030.4038.5031.0050.3347.5032.3927.7033.68
VILA1.5-8B [29]34.0030.4028.2024.0042.2538.0050.0018.5022.0742.7233.01
Idefics3-8B [21]34.0021.2038.6028.0032.7525.0048.0028.0026.7625.3530.77
LLaVA-OneVision-7B [24]18.8012.0037.6033.6037.7533.0050.0021.5029.5826.2930.01
Phi-3.5-vision [1]20.4012.0033.2028.4036.7534.0050.0022.5034.7426.7629.88
Brote-IM-XXL [55]33.0013.2031.0024.8030.7529.5050.0018.0026.7620.1927.72
LongVA-7B [60]23.0016.0031.6028.0028.0022.0048.6721.5019.2530.9926.90
Mono-InternVL-2B [35]30.0027.6032.0027.6014.008.5050.0018.0025.3524.8825.79
mPLUG-Owl3-7B [58]22.806.8029.2015.6022.259.0046.3316.5027.7021.1321.73
", + "image_path": "343cc711b0a759b7eb2c52e37b67cf118a900b09043d24d005b54c29bef49d24.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 285, + 555, + 309 + ], + "lines": [ + { + "bbox": [ + 55, + 285, + 555, + 309 + ], + "spans": [ + { + "bbox": [ + 55, + 285, + 555, + 309 + ], + "type": "text", + "content": "Table 2. Main results on our benchmark. The proprietary models are accessed via APIs, and the open-source models are accessed from Huggingface checkpoints." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 320, + 138, + 335 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 320, + 138, + 335 + ], + "spans": [ + { + "bbox": [ + 55, + 320, + 138, + 335 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 341, + 296, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 341, + 296, + 449 + ], + "spans": [ + { + "bbox": [ + 55, + 341, + 296, + 449 + ], + "type": "text", + "content": "As discussed above, CoSpace features reconstructing the surrounding space via multiple images, with each image representing a different space segments. Different from previous multi-image benchmark involving spatial reasoning tasks, the input order of images is crucial in our tasks. To enable that models can associate each image with a specific direction to answer questions in our benchmark, we provide direction references along with images following the prompt templates in Figure 2 through out the evaluation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 456, + 168, + 468 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 456, + 168, + 468 + ], + "spans": [ + { + "bbox": [ + 55, + 456, + 168, + 468 + ], + "type": "text", + "content": "4.1. Evaluation Metrics" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 474, + 295, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 474, + 295, + 629 + ], + "spans": [ + { + "bbox": [ + 55, + 474, + 295, + 629 + ], + "type": "text", + "content": "For the Direction and the Counting categories, for fair evaluation, we adopt both pair-wise accuracy " + }, + { + "bbox": [ + 55, + 474, + 295, + 629 + ], + "type": "inline_equation", + "content": "ACC_{p}" + }, + { + "bbox": [ + 55, + 474, + 295, + 629 + ], + "type": "text", + "content": " and query-wise accuracy " + }, + { + "bbox": [ + 55, + 474, + 295, + 629 + ], + "type": "inline_equation", + "content": "ACC_{q}" + }, + { + "bbox": [ + 55, + 474, + 295, + 629 + ], + "type": "text", + "content": " as introduced by Fu et al. [12] and Luo et al. [34] instead of simple accuracy. For each question, we task models twice with different prompts. " + }, + { + "bbox": [ + 55, + 474, + 295, + 629 + ], + "type": "inline_equation", + "content": "ACC_{p}" + }, + { + "bbox": [ + 55, + 474, + 295, + 629 + ], + "type": "text", + "content": " only counts when models correctly response to both prompts, while " + }, + { + "bbox": [ + 55, + 474, + 295, + 629 + ], + "type": "inline_equation", + "content": "ACC_{q}" + }, + { + "bbox": [ + 55, + 474, + 295, + 629 + ], + "type": "text", + "content": " is the averaged score of the two responses. For these different prompts regarding the same question, we vary only the order of images by preserving the clockwise order but changing the beginning view, so that the answer to the question remains unchanged. For example, if original image sequence begins with the northern image, the modified sequence could start from the east, south or west." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 630, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 296, + 715 + ], + "type": "text", + "content": "For the Rotation category, some of the instances does not cover the entire space, and the such a modification of input image sequence will inevitable disrupt the comprehension of the continuous space. Similarly, for the Planning category, considering that the original input sequence features the front view as the first image and this manner inherently matches the thinking pattern when models take" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 322, + 555, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 322, + 555, + 357 + ], + "spans": [ + { + "bbox": [ + 313, + 322, + 555, + 357 + ], + "type": "text", + "content": "actions. Therefore, for these two categories, we apply only a single context for each instances during evaluation, and report accuracy, " + }, + { + "bbox": [ + 313, + 322, + 555, + 357 + ], + "type": "inline_equation", + "content": "ACC" + }, + { + "bbox": [ + 313, + 322, + 555, + 357 + ], + "type": "text", + "content": ", as final scores." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 367, + 372, + 379 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 367, + 372, + 379 + ], + "spans": [ + { + "bbox": [ + 313, + 367, + 372, + 379 + ], + "type": "text", + "content": "4.2. Models" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 386, + 555, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 386, + 555, + 518 + ], + "spans": [ + { + "bbox": [ + 313, + 386, + 555, + 518 + ], + "type": "text", + "content": "We accessed three proprietary models, including GPT-4o [39], Gemini-1.5-pro [44] and Claude-3.7-sonnet [2], along with 16 widely used open-source VLMs accepting multiple images simultaneously. Among the open-source models, four have more than 70B parameters, while the others range in size from 2B to 13B. We also enable the evaluation for powerful single-image models, such as LLaVA-1.6 [31] and LLaMA-3.2 [11], for further investigation. For fair comparison, we design a single-image pipeline for some our designed tasks and conduct evaluations on both single-image and multi-image models in Section 5.3." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 528, + 372, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 528, + 372, + 540 + ], + "spans": [ + { + "bbox": [ + 313, + 528, + 372, + 540 + ], + "type": "text", + "content": "4.3. Results" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 546, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 546, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 546, + 556, + 715 + ], + "type": "text", + "content": "The overall results on our benchmark is reported in Table 2. We derive three main findings from this table: 1. The performance of proprietary models and " + }, + { + "bbox": [ + 313, + 546, + 556, + 715 + ], + "type": "inline_equation", + "content": ">70B" + }, + { + "bbox": [ + 313, + 546, + 556, + 715 + ], + "type": "text", + "content": " open-source models largely varies. As expected, most proprietary models and " + }, + { + "bbox": [ + 313, + 546, + 556, + 715 + ], + "type": "inline_equation", + "content": ">70B" + }, + { + "bbox": [ + 313, + 546, + 556, + 715 + ], + "type": "text", + "content": " open-source models marginally outperform the evaluated open-source models. However, there are significant dependencies among these models, where Claude-3.5-sonnet achieves an average score of " + }, + { + "bbox": [ + 313, + 546, + 556, + 715 + ], + "type": "inline_equation", + "content": "49.37\\%" + }, + { + "bbox": [ + 313, + 546, + 556, + 715 + ], + "type": "text", + "content": ", InternVL2_5-78B achieves " + }, + { + "bbox": [ + 313, + 546, + 556, + 715 + ], + "type": "inline_equation", + "content": "45.50\\%" + }, + { + "bbox": [ + 313, + 546, + 556, + 715 + ], + "type": "text", + "content": ", while LLaVA-OneVision-72B only presents an average score of " + }, + { + "bbox": [ + 313, + 546, + 556, + 715 + ], + "type": "inline_equation", + "content": "29.17\\%" + }, + { + "bbox": [ + 313, + 546, + 556, + 715 + ], + "type": "text", + "content": ", even lower than more than half of the " + }, + { + "bbox": [ + 313, + 546, + 556, + 715 + ], + "type": "inline_equation", + "content": "<13B" + }, + { + "bbox": [ + 313, + 546, + 556, + 715 + ], + "type": "text", + "content": " open-source models. 2. Proprietary models do not equally perform the best across all the tasks. Proprietary models fail to beat open-source models on tasks including DIR-Obj and CNT," + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "29574" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 70, + 298, + 274 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 298, + 274 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 298, + 274 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 298, + 274 + ], + "type": "table", + "html": "
ModelsDIR-RecDIR-ObjCNTPLA-QAPLA-Dec
Proprietary Models
Claude-3.7-sonnet36.3631.1737.2044.4044.50
Gemini-1.5-pro37.2037.2027.5038.0336.15
GPT-4o37.2041.2039.0045.0739.91
>70B Open-source Models
InternVL2_5-78B29.6052.8044.0034.27*36.15
Qwen2-VL-72B30.4047.6046.00*37.0933.33
InternVL2-76B25.2038.8052.00*25.3525.35
LLaVA-OneVision-72B20.40*43.6022.5036.62*32.39*
<13B Open-source Models
MiniCPM-V 2.622.8040.0037.0039.4426.76
Qwen2-VL-7B19.6035.2040.5034.2723.00
Mantis-8B34.80*30.8042.50*31.9229.11*
InternVL2-8B31.20*36.4042.50*22.5426.76
VILA1.5-8B33.6029.20*42.0019.7239.91
Idefics3-8B37.60*37.2032.5023.9424.88
LLaVA-OneVision-7B17.2038.00*35.0030.05*27.70*
Phi-3.5-vision26.00*34.40*36.5031.4624.41
Brote-IM-XXL34.40*29.6029.5029.58*17.84
LongVA-7B16.0031.6028.50*17.8428.64
Mono-InternVL-2B18.8033.60*15.50*25.3522.07
mPLUG-Owl3-7B26.40*29.2029.50*29.11*32.39*
", + "image_path": "e05a594ce3ebddec012d9ba10d5cdf0dcd54e6abf2990ef64766cbc2d9042de5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 55, + 277, + 295, + 310 + ], + "lines": [ + { + "bbox": [ + 55, + 277, + 295, + 310 + ], + "spans": [ + { + "bbox": [ + 55, + 277, + 295, + 310 + ], + "type": "text", + "content": "Table 3. Results for disrupted image orders. " + }, + { + "bbox": [ + 55, + 277, + 295, + 310 + ], + "type": "inline_equation", + "content": "ACC_{q}" + }, + { + "bbox": [ + 55, + 277, + 295, + 310 + ], + "type": "text", + "content": " is reported for Direction and Counting tasks and " + }, + { + "bbox": [ + 55, + 277, + 295, + 310 + ], + "type": "inline_equation", + "content": "ACC" + }, + { + "bbox": [ + 55, + 277, + 295, + 310 + ], + "type": "text", + "content": " is reported for the Planning task. * indicates the performance is higher than that in Table 2." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 323, + 295, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 323, + 295, + 417 + ], + "spans": [ + { + "bbox": [ + 54, + 323, + 295, + 417 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 54, + 323, + 295, + 417 + ], + "type": "inline_equation", + "content": ">70\\mathrm{B}" + }, + { + "bbox": [ + 54, + 323, + 295, + 417 + ], + "type": "text", + "content": " open-source models exhibit stronger performance. 3. The ROT-Ang task is quite challenging for open-source models. It is noticeable that for the ROT-Ang task, all open-source models achieve similar results near " + }, + { + "bbox": [ + 54, + 323, + 295, + 417 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 54, + 323, + 295, + 417 + ], + "type": "text", + "content": ", which are almost equal to random selection. Such result is caused by the fixed predictions over all different context and images input. We will elaborately investigate such phenomena in Section 5.4." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 419, + 296, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 419, + 296, + 492 + ], + "spans": [ + { + "bbox": [ + 55, + 419, + 296, + 492 + ], + "type": "text", + "content": "We conduct manual evaluation on all instances to assess the reasonability and difficulty of CoSpace and results are provided in Appendix ??. Results show that human performance excels the best models across all tasks. Human annotators achieved an average " + }, + { + "bbox": [ + 55, + 419, + 296, + 492 + ], + "type": "inline_equation", + "content": "ACC_{q}" + }, + { + "bbox": [ + 55, + 419, + 296, + 492 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 55, + 419, + 296, + 492 + ], + "type": "inline_equation", + "content": "86.25\\%" + }, + { + "bbox": [ + 55, + 419, + 296, + 492 + ], + "type": "text", + "content": ", which is significantly higher than that of Claude-3.7-sonnet (59.56%)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 505, + 115, + 518 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 505, + 115, + 518 + ], + "spans": [ + { + "bbox": [ + 55, + 505, + 115, + 518 + ], + "type": "text", + "content": "5. Analysis" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 525, + 189, + 539 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 525, + 189, + 539 + ], + "spans": [ + { + "bbox": [ + 55, + 525, + 189, + 539 + ], + "type": "text", + "content": "5.1. Disrupted Image Order" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 544, + 295, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 544, + 295, + 663 + ], + "spans": [ + { + "bbox": [ + 55, + 544, + 295, + 663 + ], + "type": "text", + "content": "Considering the spatial relationship between input images of our CoSpace, the order of input images can make significant difference on the understanding of the space, especially when the order being thoroughly disrupted. Based on this speculation, we randomly shuffled the input sequence, and provided the directional orientation only of the first image for the model, with directions of the rest images to be inferred by the model itself. We conducted evaluation under such setting on three tasks: Direction, Counting and Planning, and the results are illustrated in Table 3." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "content": "Surprisingly, the performance of " + }, + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "inline_equation", + "content": "< 13\\mathrm{B}" + }, + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "content": " models under disrupted image order was not always lower than the basic setting. In fact, half of the " + }, + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "inline_equation", + "content": "< 13\\mathrm{B}" + }, + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "content": " open-source models achieved a higher score in the DIR-Rec task under disrupted set" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 310, + 71, + 558, + 194 + ], + "blocks": [ + { + "bbox": [ + 310, + 71, + 558, + 194 + ], + "lines": [ + { + "bbox": [ + 310, + 71, + 558, + 194 + ], + "spans": [ + { + "bbox": [ + 310, + 71, + 558, + 194 + ], + "type": "image", + "image_path": "8db564116b8fb674fff9f811684485b15026bd1df307597ab2fb592aeadc7458.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 196, + 555, + 240 + ], + "lines": [ + { + "bbox": [ + 313, + 196, + 555, + 240 + ], + "spans": [ + { + "bbox": [ + 313, + 196, + 555, + 240 + ], + "type": "text", + "content": "Figure 4. Variance in " + }, + { + "bbox": [ + 313, + 196, + 555, + 240 + ], + "type": "inline_equation", + "content": "ACC_{q}" + }, + { + "bbox": [ + 313, + 196, + 555, + 240 + ], + "type": "text", + "content": " of the average on the Direction and Counting categories for eight selected models. Model abbreviations are used for simplicity. Accuracy in the figure represents the average of " + }, + { + "bbox": [ + 313, + 196, + 555, + 240 + ], + "type": "inline_equation", + "content": "ACC_{q}" + }, + { + "bbox": [ + 313, + 196, + 555, + 240 + ], + "type": "text", + "content": " over two categories." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 318, + 246, + 549, + 282 + ], + "blocks": [ + { + "bbox": [ + 318, + 246, + 549, + 282 + ], + "lines": [ + { + "bbox": [ + 318, + 246, + 549, + 282 + ], + "spans": [ + { + "bbox": [ + 318, + 246, + 549, + 282 + ], + "type": "table", + "html": "
Img OrderClaude GeminiQwen2InternVL2OneVisionPhi-3.5LongVA
Sequential79.5873.6469.9134.2534.7833.3321.6236.59
Disrupted60.7156.7956.2524.6639.5835.9426.8731.58
", + "image_path": "37023e20b2dbacaf1783a933c188a52f71c750cc31504d1653620e3a01ff812f.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 313, + 285, + 555, + 318 + ], + "lines": [ + { + "bbox": [ + 313, + 285, + 555, + 318 + ], + "spans": [ + { + "bbox": [ + 313, + 285, + 555, + 318 + ], + "type": "text", + "content": "Table 4. Results of conditional accuracy on the Planning category for eight selected models. Model names are abbreviated. Open-source models all refer to " + }, + { + "bbox": [ + 313, + 285, + 555, + 318 + ], + "type": "inline_equation", + "content": "<13\\mathrm{B}" + }, + { + "bbox": [ + 313, + 285, + 555, + 318 + ], + "type": "text", + "content": " ones." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 329, + 555, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 329, + 555, + 534 + ], + "spans": [ + { + "bbox": [ + 313, + 329, + 555, + 534 + ], + "type": "text", + "content": "ting. Additionally, for other tasks, an average of three to five open-source models performed slightly better with disrupted image order. Nevertheless, similar phenomenon was not observed for proprietary models and most of the " + }, + { + "bbox": [ + 313, + 329, + 555, + 534 + ], + "type": "inline_equation", + "content": ">70\\mathrm{B}" + }, + { + "bbox": [ + 313, + 329, + 555, + 534 + ], + "type": "text", + "content": " open-source models, except LLaVA-OneVision-72B. Also, the evaluated proprietary models similarly underperformed across all tasks under disrupted setting, which aligns more closely with our expectations. For " + }, + { + "bbox": [ + 313, + 329, + 555, + 534 + ], + "type": "inline_equation", + "content": ">70\\mathrm{B}" + }, + { + "bbox": [ + 313, + 329, + 555, + 534 + ], + "type": "text", + "content": " open-source models, three of four outperformed on only one task against basic setting, with LLaVA-OneVision-72B producing a completely opposite result. Considering the such phenomenon, we conclude that the results for " + }, + { + "bbox": [ + 313, + 329, + 555, + 534 + ], + "type": "inline_equation", + "content": "<13\\mathrm{B}" + }, + { + "bbox": [ + 313, + 329, + 555, + 534 + ], + "type": "text", + "content": " open-source models are partly due to the instability of predictions. Moreover, we acknowledge that the impact of disrupting the image order may not be as significant for models as it is for human, exhibiting the discrepancy between models and human when perceiving continuous visual space via multiple images." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 540, + 457, + 553 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 540, + 457, + 553 + ], + "spans": [ + { + "bbox": [ + 313, + 540, + 457, + 553 + ], + "type": "text", + "content": "5.2. Performance Consistency" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 557, + 555, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 557, + 555, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 557, + 555, + 689 + ], + "type": "text", + "content": "We mentioned in Section 4.1 that for the Direction and Counting categories, both " + }, + { + "bbox": [ + 313, + 557, + 555, + 689 + ], + "type": "inline_equation", + "content": "ACC_{q}" + }, + { + "bbox": [ + 313, + 557, + 555, + 689 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 557, + 555, + 689 + ], + "type": "inline_equation", + "content": "ACC_{p}" + }, + { + "bbox": [ + 313, + 557, + 555, + 689 + ], + "type": "text", + "content": " are adopted as evaluation metrics. We present the gap between " + }, + { + "bbox": [ + 313, + 557, + 555, + 689 + ], + "type": "inline_equation", + "content": "ACC_{q}" + }, + { + "bbox": [ + 313, + 557, + 555, + 689 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 557, + 555, + 689 + ], + "type": "inline_equation", + "content": "ACC_{p}" + }, + { + "bbox": [ + 313, + 557, + 555, + 689 + ], + "type": "text", + "content": ", i.e. variance in accuracy, in Figure 4. A smaller variance indicates that a model responds more consistently to input changes, reflecting better performance consistency. As shown, all models exhibit relatively large variance, with proprietary models (Claude-3.7-sonnet, Gemini-1.5-pro, GPT-4o) displaying even higher variances than open-source models, suggesting room for improvement. Such result highlights the importance of introducing " + }, + { + "bbox": [ + 313, + 557, + 555, + 689 + ], + "type": "inline_equation", + "content": "ACC_{p}" + }, + { + "bbox": [ + 313, + 557, + 555, + 689 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "content": "To further analyze the performance consistency of models, we introduce the metric conditional accuracy in the" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "29575" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 60, + 71, + 290, + 157 + ], + "blocks": [ + { + "bbox": [ + 60, + 71, + 290, + 157 + ], + "lines": [ + { + "bbox": [ + 60, + 71, + 290, + 157 + ], + "spans": [ + { + "bbox": [ + 60, + 71, + 290, + 157 + ], + "type": "table", + "html": "
ModelsDIR-RecDIR-ObjCNTPLA-QAPLA-DecAVG
Models with only one image in the input
LLaVA-v1.627.6032.0021.0021.6025.8225.60
LLama-3.2-11B29.2028.8028.5017.8416.9024.25
Models with multi-image perception ability
InternVL2-8B36.40*34.8027.0033.80*31.92*32.78
MiniCPM-V 2.623.2027.2026.0031.9232.39*28.14
Phi-3.5-vision18.4036.00*28.0031.4623.0027.37
LLaVA-OneVision26.80*32.8016.0022.0724.4124.42
", + "image_path": "e3d96dbb1ce1c5ecd33edcb1e1979dd2cae30225eb577c363a85e93dc0354e70.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 159, + 295, + 192 + ], + "lines": [ + { + "bbox": [ + 55, + 159, + 295, + 192 + ], + "spans": [ + { + "bbox": [ + 55, + 159, + 295, + 192 + ], + "type": "text", + "content": "Table 5. Results for single image pipeline. We report " + }, + { + "bbox": [ + 55, + 159, + 295, + 192 + ], + "type": "inline_equation", + "content": "ACC_{q}" + }, + { + "bbox": [ + 55, + 159, + 295, + 192 + ], + "type": "text", + "content": " for Direction and Counting tasks and " + }, + { + "bbox": [ + 55, + 159, + 295, + 192 + ], + "type": "inline_equation", + "content": "ACC" + }, + { + "bbox": [ + 55, + 159, + 295, + 192 + ], + "type": "text", + "content": " for Planning tasks. * indicates the performance is higher than that in Table 2." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 205, + 295, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 205, + 295, + 337 + ], + "spans": [ + { + "bbox": [ + 54, + 205, + 295, + 337 + ], + "type": "text", + "content": "Planning category. Conditional accuracy measures how often models successfully solve the PLA-Dec question given that they answer the correspondent PLA-QA question correctly, as in PLA-QA task and PLA-Dec task, two respective questions regarding the same set of images typically feature the same object or direction as target. Results in Tabel 4 reveal that open-source models exhibit a greater gap in conditional accuracy against proprietary models than on the Direction and Counting categories. While proprietary models make more consistent decisions based on correct detection, the gap remains non-negligible." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 337, + 296, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 337, + 296, + 384 + ], + "spans": [ + { + "bbox": [ + 55, + 337, + 296, + 384 + ], + "type": "text", + "content": "Above all, we argue that maintaining performance consistency is rather important for the further improvement of models, as it is closely related to robustness and rather essential for more complicated tasks." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 395, + 180, + 407 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 395, + 180, + 407 + ], + "spans": [ + { + "bbox": [ + 55, + 395, + 180, + 407 + ], + "type": "text", + "content": "5.3. Single Image Pipeline" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 413, + 295, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 413, + 295, + 604 + ], + "spans": [ + { + "bbox": [ + 56, + 413, + 295, + 604 + ], + "type": "text", + "content": "We presented in Section 3.1 that the understanding of continuous visual space inherently relies on multiple images as input to represent the whole space. Also, recent work [54] explored to construct a pipeline that can utilize caption and memory to transform the multi-image task into single image format. Similarly, we developed a single image pipeline for the Direction, Counting and Planning categories in our benchmark. For the Rotation category, as the single image input will inevitably compromise the uniform understanding of continuous visual space and the relationship between images can hardly be replaced by mere captions, we consider it not suitable for the single image pipeline. For evaluation, we selected LLaVA-v1.6 [31] and Llama-3.2-Vision [11] as the representatives of single image models. We also selected some multi-image models and conducted the single image evaluation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "content": "Results of the single image pipeline are reported in Table 5. Overall, the multi-image models underperformed under this setting compared with the multi-image pipeline. However, in the Direction-Rec and Planning-Dec task, two of four models achieve a higher accuracy, which is far beyond our expectations. We conclude that with the effective summarization of captions, models can still have an understanding of visual clues and sometimes the complete of all images can even add to the performance. Also," + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "content": "three of the four multi-image models outperform the single-image models on the average accuracy, indicating that models with inherent multi-image capabilities boast an overall better understanding of the visual clues and information extracted from the effectively multi-image scene. Based on this observation, the importance of empowering models with multi-image capabilities needs again to be emphasized." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 167, + 529, + 179 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 167, + 529, + 179 + ], + "spans": [ + { + "bbox": [ + 313, + 167, + 529, + 179 + ], + "type": "text", + "content": "5.4. Investigation on Angle-related Reasoning" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 185, + 553, + 269 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 185, + 553, + 269 + ], + "spans": [ + { + "bbox": [ + 313, + 185, + 553, + 269 + ], + "type": "text", + "content": "Table 2 shows that all the assessed open-source models achieve near-half ACCs on the Rotation-Angle (ROT-Ang) task. After checking the responses of different models, we found out that these models tend to produce the same answer across all the queries, e.g. \"A\" for all the 300 questions in this task. Therefore, due to the even distribution of options, these models achieve the similar near-half ACC." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 270, + 554, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 270, + 554, + 449 + ], + "spans": [ + { + "bbox": [ + 313, + 270, + 554, + 449 + ], + "type": "text", + "content": "To discover the reason behind such phenomenon, we asked models to generate the rationale when answering the questions in the ROT-Ang task and observed some of these models can follow the instruction to generate rationales. For example, InternVL2-8B no longer always selected the same option and consequently correctly answered some question that it originally failed. However, we also noticed that it tended to identify the whole surrounding covered by the images, thus perceiving the sum of turning angles as 360 degrees. Therefore, it frequently output 90 degrees as the final answer, even if 90 degrees not included in the options. In contrast, proprietary models like Claude-3.7-sonnet could firstly calculate the sum of covered angles across all images based on the two give options and then decided which answer better matched the scene in given images." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 451, + 554, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 451, + 554, + 522 + ], + "spans": [ + { + "bbox": [ + 313, + 451, + 554, + 522 + ], + "type": "text", + "content": "Though all the assessed open-source models fail the Rotation-Angle task, we still need to underline the importance of this task and the related capability to correctly identify the rotation in the space. Such capability lays the foundation for more advanced tasks in practical use and has already been boasted by some of the proprietary models." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 536, + 388, + 549 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 536, + 388, + 549 + ], + "spans": [ + { + "bbox": [ + 313, + 536, + 388, + 549 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 558, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 554, + 713 + ], + "type": "text", + "content": "Recognizing the gap between previous work and more advanced spatial understanding, we propose the concept of continuous space perception and present a novel benchmark CoSpace that comprises four categories and seven tasks. CoSpace covers both fundamental spatial comprehending abilities and more practical and commonly-used capabilities to comprehensively assess continuous space perception ability for VLMs. We evaluated three widely used proprietary models and sixteen most advanced open-source models and conducted detailed analysis for noticeable results in the experiments. Results reveal a gap between the assessed open-source and proprietary models, especially for performance consistency and in rotation-based perception." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "29576" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 90, + 296, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 90, + 296, + 179 + ], + "spans": [ + { + "bbox": [ + 55, + 90, + 296, + 179 + ], + "type": "text", + "content": "This work is supported by the National Key R&D Program of China (2022ZD0160502), the National Natural Science Foundation of China (No. 62276152), and funding from Wuxi Research Institute of Applied Technologies, Tsinghua University under Grant 20242001120. We acknowledge Yan Liu, Jidong Chen, Jianyang Liu, Shuangyu Li and Shu Chi for their effort during the process of data annotation and human evaluation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 199, + 115, + 211 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 199, + 115, + 211 + ], + "spans": [ + { + "bbox": [ + 56, + 199, + 115, + 211 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 219, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 219, + 295, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 219, + 295, + 240 + ], + "spans": [ + { + "bbox": [ + 61, + 219, + 295, + 240 + ], + "type": "text", + "content": "[1] Phi-3 technical report: A highly capable language model locally on your phone, 2024. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 242, + 295, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 242, + 295, + 262 + ], + "spans": [ + { + "bbox": [ + 62, + 242, + 295, + 262 + ], + "type": "text", + "content": "[2] Anthropic. The Claude 3 Model Family: Opus, Sonnet, Haiku. 2024. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 263, + 296, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 263, + 296, + 338 + ], + "spans": [ + { + "bbox": [ + 62, + 263, + 296, + 338 + ], + "type": "text", + "content": "[3] Xu Cao, Tong Zhou, Yunsheng Ma, Wenqian Ye, Can Cui, Kun Tang, Zhipeng Cao, Kaizhao Liang, Ziran Wang, James M Rehg, et al. MAPLM: a real-world large-scale vision-language benchmark for map and traffic scene understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21819-21830, 2024. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 340, + 295, + 405 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 340, + 295, + 405 + ], + "spans": [ + { + "bbox": [ + 62, + 340, + 295, + 405 + ], + "type": "text", + "content": "[4] Boyuan Chen, Zhuo Xu, Sean Kirmani, Brain Ichter, Dorsa Sadigh, Leonidas Guibas, and Fei Xia. SpatialVLM: endowing vision-language models with spatial reasoning capabilities. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14455-14465, 2024. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 406, + 295, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 406, + 295, + 470 + ], + "spans": [ + { + "bbox": [ + 62, + 406, + 295, + 470 + ], + "type": "text", + "content": "[5] Liang Chen, Yichi Zhang, Shuhuai Ren, Haozhe Zhao, Zefan Cai, Yuchi Wang, Peiyi Wang, Tianyu Liu, and Baobao Chang. Towards end-to-end embodied decision making via multi-modal large language model: Explorations with gpt4- vision and beyond. arXiv preprint arXiv:2310.02071, 2023. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 472, + 295, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 472, + 295, + 525 + ], + "spans": [ + { + "bbox": [ + 62, + 472, + 295, + 525 + ], + "type": "text", + "content": "[6] Zhe Chen, Weiyun Wang, Yue Cao, Yangzhou Liu, Zhangwei Gao, Erfei Cui, Jinguo Zhu, Shenglong Ye, Hao Tian, Zhaoyang Liu, et al. Expanding performance boundaries of open-source multimodal models with model, data, and test-time scaling. arXiv preprint arXiv:2412.05271, 2024. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 527, + 295, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 527, + 295, + 581 + ], + "spans": [ + { + "bbox": [ + 62, + 527, + 295, + 581 + ], + "type": "text", + "content": "[7] Zhe Chen, Weiyun Wang, Hao Tian, Shenglong Ye, Zhangwei Gao, Erfei Cui, Wenwen Tong, Kongzhi Hu, Jiapeng Luo, Zheng Ma, et al. How far are we to GPT-4V? closing the gap to commercial multimodal models with open-source suites. ArXiv preprint, abs/2404.16821, 2024. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 582, + 295, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 582, + 295, + 647 + ], + "spans": [ + { + "bbox": [ + 62, + 582, + 295, + 647 + ], + "type": "text", + "content": "[8] Hao-Tien Lewis Chiang, Zhuo Xu, Zipeng Fu, Mithun George Jacob, Tingnan Zhang, Tsang-Wei Edward Lee, Wenhao Yu, Connor Schenck, David Rendleman, Dhruv Shah, et al. Mobility VLA: multimodal instruction navigation with long-context vlms and topological graphs. arXiv preprint arXiv:2407.07775, 2024. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 647, + 295, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 647, + 295, + 691 + ], + "spans": [ + { + "bbox": [ + 62, + 647, + 295, + 691 + ], + "type": "text", + "content": "[9] Song Dingjie, Shunian Chen, Guiming Hardy Chen, Fei Yu, Xiang Wan, and Benyou Wang. Milebench: Benchmarking MLLMs in long context. In First Conference on Language Modeling, 2024. 2, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 692, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 692, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 692, + 295, + 713 + ], + "type": "text", + "content": "[10] Sivan Doveh, Shaked Perek, M Jehanzeb Mirza, Wei Lin, Amit Alfassy, Assaf Arbelle, Shimon Ullman, and Leonid" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 105 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 105 + ], + "type": "text", + "content": "Karlinsky. Towards multimodal in-context learning for vision & language models. arXiv preprint arXiv:2403.12736, 2024. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 107, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 150 + ], + "type": "text", + "content": "[11] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, and et al. The llama 3 herd of models, 2024. 6, 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 152, + 553, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 553, + 206 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 553, + 206 + ], + "type": "text", + "content": "[12] Chaoyou Fu, Peixian Chen, Yunhang Shen, Yulei Qin, Mengdan Zhang, Xu Lin, Jinrui Yang, Xiawu Zheng, Ke Li, Xing Sun, et al. Mme: A comprehensive evaluation benchmark for multimodal large language models. arXiv preprint arXiv:2306.13394, 2023. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 209, + 553, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 209, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 317, + 209, + 553, + 262 + ], + "type": "text", + "content": "[13] Chaoyou Fu, Yuhan Dai, Yondong Luo, Lei Li, Shuhuai Ren, Renrui Zhang, Zihan Wang, Chenyu Zhou, Yunhang Shen, Mengdan Zhang, et al. Video-MME: the first-ever comprehensive evaluation benchmark of multi-modal llms in video analysis. ArXiv preprint, abs/2405.21075, 2024. 2, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 264, + 553, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 264, + 553, + 330 + ], + "spans": [ + { + "bbox": [ + 316, + 264, + 553, + 330 + ], + "type": "text", + "content": "[14] Xingyu Fu, Yushi Hu, Bangzheng Li, Yu Feng, Haoyu Wang, Xudong Lin, Dan Roth, Noah A. Smith, Wei-Chiu Ma, and Ranjay Krishna. Blink: Multimodal large language models can see butA not perceive. In Computer Vision - ECCV 2024, pages 148-166, Cham, 2025. Springer Nature Switzerland. 2, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 332, + 553, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 332, + 553, + 386 + ], + "spans": [ + { + "bbox": [ + 316, + 332, + 553, + 386 + ], + "type": "text", + "content": "[15] Yupan Huang, Zaiqiao Meng, Fangyu Liu, Yixuan Su, Nigel Collier, and Yutong Lu. Sparkles: Unlocking chats across multiple images for multimodal instruction-following models. In ICLR 2024 Workshop on Navigating and Addressing Data Problems for Foundation Models, 2024. 2, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 388, + 553, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 388, + 553, + 441 + ], + "spans": [ + { + "bbox": [ + 317, + 388, + 553, + 441 + ], + "type": "text", + "content": "[16] Baoxiong Jia, Yixin Chen, Huangyue Yu, Yan Wang, Xuesong Niu, Tengyu Liu, Qing Li, and Siyuan Huang. Sceneverse: Scaling 3d vision-language learning for grounded scene understanding. arXiv preprint arXiv:2401.09340, 2024. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 444, + 553, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 444, + 553, + 487 + ], + "spans": [ + { + "bbox": [ + 317, + 444, + 553, + 487 + ], + "type": "text", + "content": "[17] Mengzhao Jia, Wenhao Yu, Kaixin Ma, Tianqing Fang, Zhihan Zhang, Siru Ouyang, Hongming Zhang, Meng Jiang, and Dong Yu. Leopard: A Vision Language Model For Text-Rich Multi-Image Tasks, 2024. 1, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 489, + 553, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 489, + 553, + 531 + ], + "spans": [ + { + "bbox": [ + 317, + 489, + 553, + 531 + ], + "type": "text", + "content": "[18] Dongfu Jiang, Xuan He, Huaye Zeng, Cong Wei, Max Ku, Qian Liu, and Wenhu Chen. Mantis: interleaved multi-image instruction tuning. ArXiv preprint, abs/2405.01483, 2024. 3, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 534, + 553, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 534, + 553, + 588 + ], + "spans": [ + { + "bbox": [ + 317, + 534, + 553, + 588 + ], + "type": "text", + "content": "[19] Mukul Khanna*, Ram Ramrakhya*, Gunjan Chhablani, Sriram Yenamandra, Theophile Gervet, Matthew Chang, Zsolt Kira, Devendra Singh Chaplot, Dhruv Batra, and Roozbeh Mottaghi. Goat-bench: A benchmark for multi-modal life-long navigation. In CVPR, 2024. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 590, + 553, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 590, + 553, + 633 + ], + "spans": [ + { + "bbox": [ + 317, + 590, + 553, + 633 + ], + "type": "text", + "content": "[20] Taewoong Kim, Cheolhong Min, Byeonghwi Kim, Jinyeon Kim, Wonje Jeung, and Jonghyun Choi. Realfred: An embodied instruction following benchmark in photo-realistic environment. In ECCV, 2024. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 635, + 553, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 553, + 677 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 553, + 677 + ], + "type": "text", + "content": "[21] Hugo Laurençon, Andres Marafioti, Victor Sanh, and Léo Tronchon. Building and better understanding vision-language models: insights and future directions, 2024. 3, 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 680, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 680, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 680, + 553, + 713 + ], + "type": "text", + "content": "[22] Bo Li, Yuanhan Zhang, Liangyu Chen, Jinghao Wang, Jingkang Yang, and Ziwei Liu. Otter: A Multi-Modal Model with In-Context Instruction Tuning, 2023. 1, 3" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "29577" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "text", + "content": "[23] Bohao Li, Yuying Ge, Yixiao Ge, Guangzhi Wang, Rui Wang, Ruimao Zhang, and Ying Shan. SEED-Bench: Benchmarking multimodal large language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13299-13308, 2024. 2, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 128, + 294, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 128, + 294, + 172 + ], + "spans": [ + { + "bbox": [ + 56, + 128, + 294, + 172 + ], + "type": "text", + "content": "[24] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. LLaVA-OneVision: Easy visual task transfer. ArXiv preprint, abs/2408.03326, 2024. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 172, + 294, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 172, + 294, + 215 + ], + "spans": [ + { + "bbox": [ + 56, + 172, + 294, + 215 + ], + "type": "text", + "content": "[25] Feng Li, Renrui Zhang, Hao Zhang, Yuanhan Zhang, Bo Li, Wei Li, Zejun Ma, and Chunyuan Li. Llava-last-interleave: Tackling multi-image, video, and 3d in large multimodal models, 2024. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 216, + 294, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 216, + 294, + 281 + ], + "spans": [ + { + "bbox": [ + 56, + 216, + 294, + 281 + ], + "type": "text", + "content": "[26] Juncheng Li, Kaihang Pan, Zhiqi Ge, Minghe Gao, Wei Ji, Wenqiao Zhang, Tat-Seng Chua, Siliang Tang, Hanwang Zhang, and Yueting Zhuang. Fine-tuning multimodal LLMs to follow zero-shot demonstrative instructions. In The Twelfth International Conference on Learning Representations, 2024. 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 282, + 294, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 282, + 294, + 347 + ], + "spans": [ + { + "bbox": [ + 56, + 282, + 294, + 347 + ], + "type": "text", + "content": "[27] Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Yi Liu, Zun Wang, Jilan Xu, Guo Chen, Ping Luo, et al. Mvbench: A comprehensive multi-modal video understanding benchmark. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22195-22206, 2024. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 349, + 294, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 349, + 294, + 414 + ], + "spans": [ + { + "bbox": [ + 56, + 349, + 294, + 414 + ], + "type": "text", + "content": "[28] Xiaoqi Li, Mingxu Zhang, Yiran Geng, Haoran Geng, Yuxing Long, Yan Shen, Renrui Zhang, Jiaming Liu, and Hao Dong. ManipLLM: Embodied multimodal large language model for object-centric robotic manipulation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18061-18070, 2024. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 415, + 294, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 415, + 294, + 469 + ], + "spans": [ + { + "bbox": [ + 56, + 415, + 294, + 469 + ], + "type": "text", + "content": "[29] Ji Lin, Hongxu Yin, Wei Ping, Pavlo Molchanov, Mohammad Shoeybi, and Song Han. Vila: On pre-training for visual language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26689-26699, 2024. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 470, + 294, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 470, + 294, + 525 + ], + "spans": [ + { + "bbox": [ + 56, + 470, + 294, + 525 + ], + "type": "text", + "content": "[30] Ziyi Lin, Chris Liu, Renrui Zhang, Peng Gao, Longtian Qiu, Han Xiao, Han Qiu, Chen Lin, Wenqi Shao, Keqin Chen, et al. Sphinx: The joint mixing of weights, tasks, and visual embeddings for multi-modal large language models. arXiv preprint arXiv:2311.07575, 2023. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 525, + 294, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 525, + 294, + 547 + ], + "spans": [ + { + "bbox": [ + 56, + 525, + 294, + 547 + ], + "type": "text", + "content": "[31] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning, 2023. 2, 6, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 548, + 294, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 548, + 294, + 581 + ], + "spans": [ + { + "bbox": [ + 56, + 548, + 294, + 581 + ], + "type": "text", + "content": "[32] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. LLaVA-NeXT: Improved reasoning, OCR, and world knowledge, 2024. 2, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 582, + 294, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 582, + 294, + 624 + ], + "spans": [ + { + "bbox": [ + 56, + 582, + 294, + 624 + ], + "type": "text", + "content": "[33] Haowei Liu, Xi Zhang, Haiyang Xu, Yaya Shi, Chaoya Jiang, Ming Yan, Ji Zhang, Fei Huang, Chunfeng Yuan, Bing Li, and Weiming Hu. MIBench: Evaluating multimodal large language models over multiple images, 2024. 2, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 625, + 294, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 625, + 294, + 690 + ], + "spans": [ + { + "bbox": [ + 56, + 625, + 294, + 690 + ], + "type": "text", + "content": "[34] Fuwen Luo, Chi Chen, Zihao Wan, Zhaolu Kang, Qidong Yan, Yingjie Li, Xiaolong Wang, Siyu Wang, Ziyue Wang, Xiaoyue Mi, Peng Li, Ning Ma, Maosong Sun, and Yang Liu. Codis: Benchmarking context-dependent visual comprehension for multimodal large language models. arXiv preprint arXiv:2402.13607, 2024. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "text", + "content": "[35] Gen Luo, Xue Yang, Wenhan Dou, Zhaokai Wang, Jifeng Dai, Yu Qiao, and Xizhou Zhu. Mono-internvl: Pushing the" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 333, + 72, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 72, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 72, + 553, + 95 + ], + "type": "text", + "content": "boundaries of monolithic multimodal large language models with endogenous visual pre-training, 2024. 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 96, + 553, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 171 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 171 + ], + "type": "text", + "content": "[36] Arjun Majumdar, Anurag Ajay, Xiaohan Zhang, Pranav Putta, Sriram Yenamandra, Mikael Henaff, Sneha Silwal, Paul Mcvay, Oleksandr Maksymets, Sergio Arnaud, et al. OpenEQA: embodied question answering in the era of foundation models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16488-16498, 2024. 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 172, + 553, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 172, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 316, + 172, + 553, + 228 + ], + "type": "text", + "content": "[37] Fanqing Meng, Jin Wang, Chuanhao Li, Quanfeng Lu, Hao Tian, Jiaqi Liao, Xizhou Zhu, Jifeng Dai, Yu Qiao, Ping Luo, Kaipeng Zhang, and Wenqi Shao. MMIU: Multimodal multi-image understanding for evaluating large vision-language models, 2024. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 228, + 553, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 228, + 553, + 271 + ], + "spans": [ + { + "bbox": [ + 316, + 228, + 553, + 271 + ], + "type": "text", + "content": "[38] Jiahao Nie, Gongjie Zhang, Wenbin An, Yap-Peng Tan, Alex C Kot, and Shijian Lu. Mmrel: A relation understanding dataset and benchmark in the mllm era. arXiv preprint arXiv:2406.09121, 2024. 1, 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 272, + 553, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 272, + 553, + 293 + ], + "spans": [ + { + "bbox": [ + 316, + 272, + 553, + 293 + ], + "type": "text", + "content": "[39] OpenAI. Hello GPT-4o, 2024. Accessed on June 16, 2024. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 294, + 553, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 294, + 553, + 403 + ], + "spans": [ + { + "bbox": [ + 316, + 294, + 553, + 403 + ], + "type": "text", + "content": "[40] Xavier Puig, Eric Undersander, Andrew Szot, Mikael Dallaire Cote, Tsung-Yen Yang, Ruslan Partsey, Ruta Desai, Alexander Clegg, Michal Hlavac, So Yeon Min, Vladimir Vondruš, Theophile Gervet, Vincent-Pierre Berges, John M Turner, Oleksandr Maksymets, Zsolt Kira, Mrinal Kalakrishnan, Jitendra Malik, Devendra Singh Chaplot, Unnat Jain, Dhruv Batra, Akshara Rai, and Roozbeh Mottaghi. Habitat 3.0: A co-habitat for humans, avatars, and robots. In Proceedings of the Twelfth International Conference on Learning Representations (ICLR 2024), 2024. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 404, + 553, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 404, + 553, + 437 + ], + "spans": [ + { + "bbox": [ + 316, + 404, + 553, + 437 + ], + "type": "text", + "content": "[41] Navid Rajabi and Jana Kosecka. Gsr-bench: A benchmark for grounded spatial reasoning evaluation via multimodal llms. arXiv preprint arXiv:2406.13246, 2024. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 437, + 553, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 437, + 553, + 492 + ], + "spans": [ + { + "bbox": [ + 316, + 437, + 553, + 492 + ], + "type": "text", + "content": "[42] Abhinav Rajvanshi, Karan Sikka, Xiao Lin, Bhoram Lee, Han pang Chiu, and Alvaro Velasquez. SayNav: Grounding large language models for dynamic planning to navigation in new environments. In 34th International Conference on Automated Planning and Scheduling, 2024. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 492, + 553, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 492, + 553, + 580 + ], + "spans": [ + { + "bbox": [ + 316, + 492, + 553, + 580 + ], + "type": "text", + "content": "[43] Santhosh Kumar Ramakrishnan, Aaron Gokaslan, Erik Wijmans, Oleksandr Maksymets, Alexander Clegg, John M Turner, Eric Undersander, Wojciech Galuba, Andrew Westbury, Angel X Chang, Manolis Savva, Yili Zhao, and Dhruv Batra. Habitat-matterport 3d dataset (HM3d): 1000 large-scale 3d environments for embodied AI. In Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 2), 2021. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 581, + 553, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 581, + 553, + 645 + ], + "spans": [ + { + "bbox": [ + 316, + 581, + 553, + 645 + ], + "type": "text", + "content": "[44] Machel Reid, Nikolay Savinov, Denis Teplyashin, Dmitry Lepikhin, Timothy Lillicrap, Jean-baptiste Alayrac, Radu Soricut, Angeliki Lazaridou, Orhan First, Julian Schrittwieser, et al. Gemini 1.5: unlocking multimodal understanding across millions of tokens of context. ArXiv preprint, abs/2403.05530, 2024. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "type": "text", + "content": "[45] Allen Z. Ren, Jaden Clark, Anushri Dixit, Masha Itkina, Anirudha Majumdar, and Dorsa Sadigh. Explore until confident: Efficient exploration for embodied question answering, 2024. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 690, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 690, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 690, + 553, + 713 + ], + "type": "text", + "content": "[46] Manolis Savva, Abhishek Kadian, Oleksandr Maksymets, Yili Zhao, Erik Wijmans, Bhavana Jain, Julian Straub, Jia" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "29578" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 76, + 72, + 296, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 296, + 116 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 296, + 116 + ], + "type": "text", + "content": "Liu, Vladlen Koltun, Jitendra Malik, Devi Parikh, and Dhruv Batra. Habitat: A Platform for Embodied AI Research. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2019. 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 296, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 296, + 173 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 296, + 173 + ], + "type": "text", + "content": "[47] Fatemeh Shiri, Xiao-Yu Guo, Mona Far, Xin Yu, Reza Haf, and Yuan-Fang Li. An empirical analysis on spatial reasoning capabilities of large multimodal models. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 21440-21455, 2024. 1, 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 173, + 295, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 173, + 295, + 270 + ], + "spans": [ + { + "bbox": [ + 56, + 173, + 295, + 270 + ], + "type": "text", + "content": "[48] Andrew Szot, Alex Clegg, Eric Undersander, Erik Wijmans, Yili Zhao, John Turner, Noah Maestre, Mustafa Mukadam, Devendra Chaplot, Oleksandr Maksymets, Aaron Gokaslan, Vladimir Vondrus, Sameer Dharur, Franziska Meier, Wojciech Galuba, Angel Chang, Zsolt Kira, Vladlen Koltun, Jitendra Malik, Manolis Savva, and Dhruv Batra. Habitat 2.0: Training home assistants to rearrange their habitat. In Advances in Neural Information Processing Systems (NeurIPS), 2021. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 272, + 295, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 272, + 295, + 325 + ], + "spans": [ + { + "bbox": [ + 56, + 272, + 295, + 325 + ], + "type": "text", + "content": "[49] Yunlong Tang, Jing Bi, Siting Xu, Luchuan Song, Susan Liang, Teng Wang, Daoan Zhang, Jie An, Jingyang Lin, Rongyi Zhu, et al. Video understanding with large language models: A survey. arXiv preprint arXiv:2312.17432, 2023. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 327, + 295, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 327, + 295, + 381 + ], + "spans": [ + { + "bbox": [ + 56, + 327, + 295, + 381 + ], + "type": "text", + "content": "[50] Jiayu Wang, Yifei Ming, Zhenmei Shi, Vibhav Vineet, Xin Wang, Yixuan Li, and Neel Joshi. Is a picture worth a thousand words? delving into spatial reasoning for vision language models. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 383, + 295, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 383, + 295, + 449 + ], + "spans": [ + { + "bbox": [ + 56, + 383, + 295, + 449 + ], + "type": "text", + "content": "[51] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Yang Fan, Kai Dang, Mengfei Du, Xuancheng Ren, Rui Men, Dayiheng Liu, Chang Zhou, Jingren Zhou, and Junyang Lin. Qwen2-VL: Enhancing vision-language model's perception of the world at any resolution, 2024. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 449, + 295, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 449, + 295, + 491 + ], + "spans": [ + { + "bbox": [ + 56, + 449, + 295, + 491 + ], + "type": "text", + "content": "[52] Xidong Wang, Dingjie Song, Shunian Chen, Chen Zhang, and Benyou Wang. LongLLaVA: Scaling Multi-modal LLMs to 1000 Images Efficiently via a Hybrid Architecture, 2024. 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 493, + 295, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 493, + 295, + 591 + ], + "spans": [ + { + "bbox": [ + 56, + 493, + 295, + 591 + ], + "type": "text", + "content": "[53] Xiyao Wang, Yuhang Zhou, Xiaoyu Liu, Hongjin Lu, Yuancheng Xu, Feihong He, Jaehong Yoon, Taixi Lu, Fuxiao Liu, Gedas Bertasius, Mohit Bansal, Huaxiu Yao, and Furong Huang. Mementos: A comprehensive benchmark for multimodal large language model reasoning over image sequences. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 416-442, Bangkok, Thailand, 2024. Association for Computational Linguistics. 2, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 592, + 295, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 592, + 295, + 646 + ], + "spans": [ + { + "bbox": [ + 56, + 592, + 295, + 646 + ], + "type": "text", + "content": "[54] Ziyue Wang, Chi Chen, Fuwen Luo, Yurui Dong, Yuanchi Zhang, Yuzhuang Xu, Xiaolong Wang, Peng Li, and Yang Liu. ActiView: Evaluating active perception ability for multimodal large language models. arXiv preprint arXiv:2410.04659, 2024. 3, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 647, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 295, + 713 + ], + "type": "text", + "content": "[55] Ziyue Wang, Chi Chen, Yiqi Zhu, Fuwen Luo, Peng Li, Ming Yan, Ji Zhang, Fei Huang, Maosong Sun, and Yang Liu. Browse and Concentrate: comprehending multimodal content via prior-llm context fusion. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 11229–11245," + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 563 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "Bangkok, Thailand, 2024. Association for Computational Linguistics. 3, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 316, + 96, + 553, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 129 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 129 + ], + "type": "text", + "content": "[56] Zhutian Yang, Caelan Garrett, Dieter Fox, Tomás Lozano-Pérez, and Leslie Pack Kaelbling. Guiding long-horizon task and motion planning with vision language models, 2024. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 316, + 130, + 553, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 130, + 553, + 174 + ], + "spans": [ + { + "bbox": [ + 316, + 130, + 553, + 174 + ], + "type": "text", + "content": "[57] Yuan Yao, Tianyu Yu, Ao Zhang, Chongyi Wang, Junbo Cui, Hongji Zhu, Tianchi Cai, Haoyu Li, Weilin Zhao, Zhihui He, et al. MiniCPM-V: A GPT-4V level MLLM on your phone. ArXiv preprint, abs/2408.01800, 2024. 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 175, + 553, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 553, + 228 + ], + "type": "text", + "content": "[58] Jiabo Ye, Haiyang Xu, Haowei Liu, Anwen Hu, Ming Yan, Qi Qian, Ji Zhang, Fei Huang, and Jingren Zhou. mPLUG-Owl3: towards long image-sequence understanding in multi-modal large language models. ArXiv preprint, abs/2408.04840, 2024. 3, 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 230, + 553, + 318 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 230, + 553, + 318 + ], + "spans": [ + { + "bbox": [ + 316, + 230, + 553, + 318 + ], + "type": "text", + "content": "[59] Haotian Zhang, Mingfei Gao, Zhe Gan, Philipp Duffer, Nina Wenzel, Forrest Huang, Dhruti Shah, Xianzhi Du, Bowen Zhang, Yanghao Li, Sam Dodge, Keen You, Zhen Yang, Aleksei Timofeev, Mingze Xu, Hong-You Chen, JeanPhilippe Fauconnier, Zhengfeng Lai, Haoxuan You, Zirui Wang, Afshin Dehghan, Peter Grasch, and Yinfei Yang. MM1.5: Methods, analysis & insights from multimodal llm fine-tuning, 2024. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 319, + 553, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 319, + 553, + 373 + ], + "spans": [ + { + "bbox": [ + 316, + 319, + 553, + 373 + ], + "type": "text", + "content": "[60] Peiyuan Zhang, Kaichen Zhang, Bo Li, Guangtao Zeng, Jingkang Yang, Yuanhan Zhang, Ziyue Wang, Haoran Tan, Chunyuan Li, and Ziwei Liu. Long context transfer from language to vision. arXiv preprint arXiv:2406.16852, 2024. 1, 2, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 375, + 553, + 440 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 375, + 553, + 440 + ], + "spans": [ + { + "bbox": [ + 316, + 375, + 553, + 440 + ], + "type": "text", + "content": "[61] Yi-Fan Zhang, Huanyu Zhang, Haochen Tian, Chaoyou Fu, Shuangqing Zhang, Junfei Wu, Feng Li, Kun Wang, Qing-song Wen, Zhang Zhang, Liang Wang, Rong Jin, and Tieniu Tan. Mme-realworld: Could your multimodal llm challenge high-resolution real-world scenarios that are difficult for humans?, 2024. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 441, + 553, + 506 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 441, + 553, + 506 + ], + "spans": [ + { + "bbox": [ + 316, + 441, + 553, + 506 + ], + "type": "text", + "content": "[62] Haozhe Zhao, Zefan Cai, Shuzheng Si, Xiaojian Ma, Kaikai An, Liang Chen, Zixuan Liu, Sheng Wang, Wenjuan Han, and Baobao Chang. MMICL: Empowering vision-language model with multi-modal in-context learning. In The Twelfth International Conference on Learning Representations, 2024. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 508, + 553, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 508, + 553, + 563 + ], + "spans": [ + { + "bbox": [ + 316, + 508, + 553, + 563 + ], + "type": "text", + "content": "[63] Gengze Zhou, Yicong Hong, Zun Wang, Xin Eric Wang, and Qi Wu. NavGPT-2: Unleashing navigational reasoning capability for large vision-language models. In European Conference on Computer Vision, pages 260–278. Springer, 2025. 1, 3" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "29579" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/CoT-VLA_ Visual Chain-of-Thought Reasoning for Vision-Language-Action Models/4f2ea357-dbc3-435d-b0c3-c4ec69dfb496_content_list.json b/2025/CoT-VLA_ Visual Chain-of-Thought Reasoning for Vision-Language-Action Models/4f2ea357-dbc3-435d-b0c3-c4ec69dfb496_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..c51597cb341bae8438fab600a893ecdc49b35eeb --- /dev/null +++ b/2025/CoT-VLA_ Visual Chain-of-Thought Reasoning for Vision-Language-Action Models/4f2ea357-dbc3-435d-b0c3-c4ec69dfb496_content_list.json @@ -0,0 +1,1343 @@ +[ + { + "type": "text", + "text": "CoT-VLA: Visual Chain-of-Thought Reasoning for Vision-Language-Action Models", + "text_level": 1, + "bbox": [ + 241, + 130, + 756, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Qingqing Zhao $^{1,2,*}$ Yao Lu $^{1}$ Moo Jin Kim $^{2}$ Zipeng Fu $^{2}$ Zhuoyang Zhang $^{3}$ Yecheng Wu $^{1,3}$ Zhaoshuo Li $^{1}$ Qianli Ma $^{1}$ Song Han $^{1,3}$ Chelsea Finn $^{2}$ Ankur Handa $^{1}$ Tsung-Yi Lin $^{1\\dagger}$ Gordon Wetzstein $^{2\\dagger}$ Ming-Yu Liu $^{1\\dagger}$ Donglai Xiang $^{1\\dagger}$ NVIDIA ${}^{2}$ Stanford University ${}^{3}$ MIT", + "bbox": [ + 99, + 200, + 883, + 282 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 316, + 326, + 333 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Vision-language-action models (VLAs) have shown potential in leveraging pretrained vision-language models and diverse robot demonstrations for learning generalizable sensorimotor control. While this paradigm effectively utilizes large-scale data from both robotic and non-robotic sources, current VLAs primarily focus on direct input-output mappings, lacking the intermediate reasoning steps crucial for complex manipulation tasks. As a result, existing VLAs lack temporal planning or reasoning capabilities. In this paper, we introduce a method that incorporates explicit visual chain-of-thought (CoT) reasoning into vision-language-action models (VLAs) by predicting future image frames autoregressively as visual goals before generating a short action sequence to achieve these goals. We introduce CoT-VLA, a state-of-the-art 7B VLA that can understand and generate visual and action tokens. Our experimental results demonstrate that CoT-VLA achieves strong performance, outperforming the state-of-the-art VLA model by $17\\%$ in real-world manipulation tasks and $6\\%$ in simulation benchmarks. Videos are available at: https://cot-vla.github.io/.", + "bbox": [ + 89, + 349, + 485, + 652 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 684, + 220, + 699 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent advances in robot learning have demonstrated impressive progress in training policies that can act across diverse tasks and environments [1, 5, 12, 14, 18, 29, 36, 44, 45, 48, 54, 59, 63, 66, 70, 76, 78]. One promising direction is vision-language-action (VLA) models, which leverage the rich understanding capabilities of pretrained vision-language models (VLMs) to map natural language instructions and visual observations to robot actions [12, 29, 48]. By training VLMs on robot demonstrations, VLAs inherit their ability to understand diverse scenes, objects, and natural language in-", + "bbox": [ + 89, + 710, + 485, + 861 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/bd662ec31266da1396db03846cc4a2f7312b32a1a370c3c820dd1e662ac80f6a.jpg", + "image_caption": [ + "Figure 1. Comparison between vanilla VLA and CoT-VLA frameworks. Prior VLA models (top) directly predict robot actions from task inputs without explicit reasoning steps and only use action-annotated robot demonstration data for training. Unlike vanilla VLAs, CoT-VLA (bottom) can also leverage action-less datasets like EPIC-KITCHEN-100 [27] to enhance subgoal image generation ability, unlocking the potential of using abundant unlabeled video data to improve VLA's visual reasoning capability. CoT-VLA first generates a subgoal image as an intermediate reasoning step, and then generate a short action sequence to achieve the subgoal. We outline the robot arm for better visualization." + ], + "image_footnote": [], + "bbox": [ + 519, + 316, + 901, + 664 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "structions, leading to better generalization capabilities when fine-tuned for downstream testing scenarios. While these approaches have shown impressive results, they typically map", + "bbox": [ + 511, + 854, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Work done during an internship at NVIDIA.", + "bbox": [ + 107, + 875, + 351, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "Equal Advising", + "bbox": [ + 109, + 887, + 200, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "1702", + "bbox": [ + 483, + 945, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "directly from observations to actions without explicit intermediate reasoning steps that could improve interpretability and, potentially, performance.", + "bbox": [ + 89, + 90, + 485, + 136 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In the language domain, chain-of-thought (CoT) prompting has emerged as a powerful technique for improving the reasoning capabilities of large language models (LLMs) by encouraging step-by-step thinking [62, 75]. Applying these concepts to robotics presents exciting opportunities for grounding reasoning in text, visual observations, and physical actions. Recent works have made progress in this direction by incorporating intermediate reasoning steps like language descriptions, keypoints, or bounding boxes [15, 44, 45, 63]. These intermediate representations capture abstracted states of scenes, objects, and tasks and often require additional pre-processing pipelines. In our work, we explore subgoal images as an intermediate reasoning step before action generation. These images capture the state of the model's reasoning process and are naturally available within robot demonstration datasets. While prior work has explored subgoal generation and goal-conditioned imitation learning [2, 11, 46, 55], to the best of our knowledge, our approach is the first to integrate these concepts with VLAs as intermediate chain-of-thought reasoning steps.", + "bbox": [ + 91, + 138, + 485, + 441 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We propose visual chain-of-thought reasoning for VLAs, a new method that uses subgoal image generation as a form of chain-of-thought reasoning for robotic tasks. Rather than directly predicting actions, our method first generates a subgoal image that represents the robot's planned state in pixel space, and then conditions its action on both the current observation and the generated subgoal image. This approach allows the model to \"think visually\" about how to accomplish a task before acting. By using the subgoal image as intermediate reasoning step, we leverage information that already exists in robot manipulation data with minimal preprocessing required. Furthermore, since subgoal image generation does not require action annotations, this unlocks the potential of using abundant video data for improved visual reasoning and understanding.", + "bbox": [ + 89, + 444, + 485, + 671 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We build our CoT-VLA system that leverages visual chain-of-thought reasoning upon recent advances in unified multimodal foundation models that can understand and generate text and images [39, 58, 61, 67, 69]. We train our base model [67] on both the Open X-Embodiment dataset [48] and action-less video datasets [20, 27], and then fine-tune the model on task demonstrations collected on downstream robot setups used for deployment and evaluation. We design a hybrid attention mechanism for CoT-VLA: we use causal attention with next-token prediction for text and image generation, and leverage full attention to predict all action dimensions at once. Additionally, inspired by recent advances in robot learning [10, 17, 77], we predict sequences of actions (action chunking) rather than a single action at each timestep. We demonstrate that both action chunking and the hybrid", + "bbox": [ + 89, + 674, + 485, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "attention mechanism improve the model's performance.", + "bbox": [ + 511, + 90, + 880, + 106 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Through extensive experiments in both simulation benchmarks [37] and real-world experiments[48, 60], we demonstrate that our visual chain-of-thought reasoning helps improve policy performance compared to prior VLA approaches. Our key contributions include:", + "bbox": [ + 511, + 106, + 906, + 181 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce a method of visual chain-of-thought reasoning through subgoal image generation as an intermediate reasoning step for robotic control.", + "- We introduce a system CoT-VLA that incorporates visual chain-of-thought reasoning, and a hybrid attention mechanism that combines causal attention for pixel and text generation and full attention for action prediction.", + "- We conduct comprehensive evaluations in both simulation and the real world, demonstrating that visual chain-of-thought reasoning improves VLA performance, and our system achieves state-of-the-art performance across multiple robot platforms and tasks." + ], + "bbox": [ + 513, + 181, + 906, + 362 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 375, + 653, + 388 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Chain-of-Thought (CoT) Reasoning CoT reasoning has gained prominence in natural language processing, particularly for enabling models to perform complex, multi-step reasoning tasks by breaking down problem-solving into sequential, explainable steps. Early work on CoT reasoning [62] has demonstrated the effectiveness of prompting large language models to generate intermediate reasoning steps before arriving at a final answer. Extending this paradigm to the visual domain, researchers have explored multimodal chain-of-thought methods, where visual information is processed iteratively in a stepwise fashion to reason about future outcomes or states, including generating bounding boxes [53], intermediate image infillments using Stable Diffusion [50] or standard Python packages [24], or generating CLIP embeddings [22]. Recently, CoT reasoning has been explored in embodied applications. It can generate textual plans for multi-stage execution [44, 45], point trajectories [63], label bounding boxes of objects and gripper positions as additional observations [44], generate future image trajectories for open-loop following [35, 47], and generate fine-grained reward guidance for reinforcement learning [76]. In this work, we introduce Visual-CoT reasoning for robotic manipulation, where predicted subgoal images serve as intermediate reasoning steps for closed-loop action generation. This approach leverages demonstration videos as natural intermediate reasoning states without requiring additional annotations.", + "bbox": [ + 511, + 398, + 908, + 806 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Vision-Language-Action Models Large pretrained vision-language models (VLMs) [9, 28, 38] have emerged as a powerful tool for robot learning, and recent works have explored various approaches to integrate them into robot systems. Several works utilize VLMs as intermediate", + "bbox": [ + 511, + 825, + 908, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "1703", + "bbox": [ + 483, + 945, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/50788cdfa18198255180faf8692616ccb541f830221d05133caaf4dace9548cd.jpg", + "image_caption": [ + "Figure 2. Overview of CoT-VLA framework. We build our model on VILA-U [67], a generative multimodal model pretrained on interleaved text-image data. The base model then trains on robot demonstrations [48] and action-less videos [20, 27]. During deployment, given a visual observation and a text instruction, the model performs visual chain-of-thought reasoning by generating a subgoal image (upper blue) with causal attention. It then generates a short action sequence with full attention $(\\mathbf{a}_1 \\cdots \\mathbf{a}_n)$ for robot execution. The system operates in a closed-loop control manner by capturing new observations after executing predicted action sequences." + ], + "image_footnote": [], + "bbox": [ + 94, + 95, + 906, + 287 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "components for perception and control, leveraging their strong semantic understanding and reasoning capabilities to decompose complex tasks [21, 26, 34, 56], detect objects [19, 25], or generate dense rewards [13, 40, 74] or goals [2, 11, 15, 46, 47, 55, 71, 80]. Some approaches incorporate VLMs [4, 49, 64] into end-to-end trainable policies by using them as pretrained backbone for better visuo-language representation [12, 29, 41, 48, 59]. Most relevant to our work are recent approaches that fine-tune pretrained VLMs on robot demonstration data for direct action prediction [12, 29, 48]. These VLAs demonstrate improved generalization to novel objects, environments, and natural language instructions through pretraining on internet-scale vision-language datasets, providing a promising direction for transferring visual and language knowledge to robotic control tasks. However, most existing VLAs do not leverage the step-by-step reasoning capabilities demonstrated in large language models, which have been shown to significantly improve performance across various tasks [62]. In the past, researchers have used chain-of-thought reasoning on language instructions or intermediate keypoints/bounding boxes for robotics [6, 13, 45, 63]. We introduce visual chain-of-thought reasoning to the VLA frameworks, using subgoal images as intermediate reasoning steps before action generation.", + "bbox": [ + 91, + 395, + 485, + 771 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. CoT-VLA", + "text_level": 1, + "bbox": [ + 89, + 784, + 199, + 800 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we present our visual chain-of-thought reasoning framework for VLAs. We begin with the formulation of our method (3.1), followed by a detailed description of the system architecture (3.2). We then explain our training procedures (3.3) and outline the deployment strategy for downstream tasks (3.3).", + "bbox": [ + 89, + 810, + 485, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Visual Chain-of-Thought Reasoning", + "text_level": 1, + "bbox": [ + 511, + 393, + 828, + 411 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We consider two types of training data for VLA pretraining: robot demonstrations dataset $D_r$ and action-less videos dataset $D_v$ . Robot demonstrations are represented as $D_r = \\{(l, \\mathbf{a}_{1\\dots T}, \\mathbf{s}_{1\\dots T})\\}$ , where $l$ denotes the natural language instruction, $\\mathbf{a}_{1\\dots T} = \\{\\mathbf{a}_1, \\dots, \\mathbf{a}_T\\}$ denotes the sequence of robot actions, and $\\mathbf{s}_{1\\dots T} = \\{\\mathbf{s}_1, \\dots, \\mathbf{s}_T\\}$ denotes the visual observations as a sequence of images. Action-less videos $D_v = \\{(l, \\mathbf{s}_{1\\dots T})\\}$ consist of language descriptions and images without action annotations.", + "bbox": [ + 511, + 417, + 908, + 551 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "VLA: Vanilla VLA approaches fine-tune a pretrained VLM, $P_{\\theta}$ , on $D_{r}$ , learning to predict actions $\\hat{\\mathbf{a}}_{t+1}$ directly from the current observation $\\mathbf{s}_{t}$ and language instruction $l$ (Figure 1, top):", + "bbox": [ + 511, + 570, + 908, + 631 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {a}} _ {t} \\sim P _ {\\theta} (\\mathbf {a} _ {t} | \\mathbf {s} _ {t}, l) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 650, + 638, + 906, + 655 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "CoT-VLA: Our key insight is to incorporate explicit visual reasoning before action generation. As illustrated in Figure 2, our approach operates in two sequential phases:", + "bbox": [ + 511, + 670, + 906, + 717 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {s}} _ {t + n} \\sim P _ {\\theta} \\left(\\mathbf {s} _ {t + n} \\mid \\mathbf {s} _ {t}, l\\right) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 673, + 723, + 906, + 739 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\{\\hat {\\mathbf {a}} _ {t}, \\dots , \\hat {\\mathbf {a}} _ {t + m} \\} \\sim P _ {\\theta} (\\{\\mathbf {a} _ {t}, \\dots , \\mathbf {a} _ {t + m} | \\mathbf {s} _ {t}, l, \\mathbf {s} _ {t + n}) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 557, + 741, + 906, + 758 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where we first predict a subgoal image $\\hat{\\mathbf{s}}_{t + n}$ , $n$ frames ahead, as an intermediate visual reasoning step (Equation 2). Then we generate a sequence of $m$ actions to achieve this subgoal state (Equation 3). This enables the model to \"think visually\" first by explicitly reasoning about desired future states before predicting the actions. The visual reasoning step, Equation (2), is trained on both robot demonstrations $D_r$ and actionless videos $D_v$ , and the action generation step, Equation (3), is trained on robot demonstrations $D_r$ only.", + "bbox": [ + 511, + 763, + 910, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "1704", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/5db19d53593396530375d73b8f75a3bbac011a98bf56c372acb801dadf706ffa.jpg", + "image_caption": [ + "Figure 3. Hybrid attention mechanism in CoT-VLA. We use causal attention for image or text generation and full attention for action generation. $[x]$ , $[\\theta]$ and $[g]$ are special tokens for parallel decoding of actions." + ], + "image_footnote": [], + "bbox": [ + 147, + 87, + 433, + 287 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. The Base Vision-Language Model", + "text_level": 1, + "bbox": [ + 89, + 382, + 387, + 398 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To enable the visual reasoning capabilities described in Equation (2), we build upon VILA-U [67], an unified multimodal foundation model capable of both understanding and generating image and text tokens.", + "bbox": [ + 89, + 407, + 485, + 467 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "VILA-U unifies video, image, and language understanding through an autoregressive next-token prediction framework. At its core is a unified vision tower that encodes visual inputs as discrete tokens aligned with textual information. This enables autoregressive image and video generation while significantly enhancing the understanding capabilities of VLMs that leverage discrete visual features. VILA-U utilizes residual quantization [32] to improve the representational capacity of discrete visual features - incorporating a depth transformer, as introduced in RQ-VAE [32], to gradually predict the residual tokens. The extracted visual features are then passed through a projector before being processed by the LLM backbone. The base model is trained on multimodal pairs including [image, text], [text, image], [video, text], and [text, video]. We use the VILA-U model trained on $256 \\times 256$ resolution images, where each image is encoded into $16 \\times 16 \\times 4$ tokens with a residual depth of 4 [32]. For detailed information about VILA-U training and architecture, we refer readers to [67].", + "bbox": [ + 89, + 469, + 483, + 757 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Training Procedures", + "text_level": 1, + "bbox": [ + 89, + 770, + 284, + 786 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We pretrain the base 7B VILA-U model on a combination of robot demonstrations $D_r$ and action-less videos $D_v$ . During training, we optimize three components, the LLM backbone, projector, and depth transformer, while keeping the vision tower fixed. Our training objective has two key components: subgoal image generation with causal attention (2) and action generation with full attention (3).", + "bbox": [ + 89, + 794, + 485, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Visual Tokens Prediction For subgoal image generation, each training sequence is of form $(l, \\mathbf{s}_t, \\mathbf{s}_{t + n})$ . We follow the training objective used in [67]. At each visual position $j$ , the depth transformer, $P_{\\delta}$ , autoregressively predicts $D$ residual tokens $(k_{j1}, \\dots, k_{jD})$ based on the LLM-generated code embedding $h_j$ . The training objective for visual tokens is then formulated as:", + "bbox": [ + 511, + 90, + 906, + 196 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {v i s u a l}} = - \\sum_ {j} \\sum_ {d = 1} ^ {D} \\log P _ {\\delta} \\left(k _ {j d} \\mid k _ {j, < d}\\right) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 584, + 203, + 906, + 247 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $j$ indexes the positions containing visual tokens. For a more detailed explanation of this loss function, we refer readers to [67], Section 3.2.", + "bbox": [ + 511, + 253, + 906, + 300 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Action Tokens Prediction For action prediction, each training sequence takes the form $(l, \\mathbf{s}_t, \\mathbf{s}_{t+n}, \\mathbf{a}_t, \\dots, \\mathbf{a}_{t+m})$ . Each action $\\mathbf{a}_i$ is represented by 7 tokens, with each action dimension independently discretized. Following [29], we map each continuous action dimension into 256 discrete bins, with bin widths determined by uniformly dividing the interval between the 1st and 99th percentiles of the training data's action distribution. We repurpose the 256 least frequently used tokens in the text tokenizer's vocabulary as action bin tokens. Unlike prior works [12, 29, 48], we employ full attention for processing and predicting action tokens, enabling all action tokens to interact with each other. This attention mechanism is illustrated in Figure 3. During training, we minimize the cross-entropy loss for action predictions:", + "bbox": [ + 511, + 316, + 908, + 529 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {a c t i o n}} = - \\sum_ {i = 1} ^ {m} \\log P _ {\\theta} \\left(\\mathbf {a} _ {t} \\dots \\mathbf {a} _ {t + m} \\mid l, s _ {t}, s _ {t + n}\\right) \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 560, + 537, + 906, + 577 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given a batch of input sequences, The overall training objective combines the action and visual losses:", + "bbox": [ + 511, + 585, + 906, + 614 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {\\text {a c t i o n}} + \\mathcal {L} _ {\\text {v i s u a l}} \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 640, + 626, + 906, + 642 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Pretraining Phase We pretrain CoT-VLA on both robot demonstrations $D_r$ and action-less videos $D_v$ as described in Section 3.1. For robot demonstrations, we curate a subset of the Open X-Embodiment dataset [48] (OpenX). Following the pre-processing pipeline established in OpenVLA [29], we select and process datasets with third-person camera views and single-arm end-effector control (7-DoF). For action-less videos $D_v$ , we incorporate the EPIC-KITCHENS [27] and Something-Something V2 [20] datasets. All images are processed at $256 \\times 256$ resolution. For visual reasoning, we use subgoal images at future timestep $n$ uniformly sampled from a dataset-specific range $[n_l, n_u]$ , where $n_l$ and $n_u$ define the lower and upper bounds of the prediction horizon. We use an action chunk size of 10. For complete dataset specifications and training hyperparameters, please refer to the supplementary material.", + "bbox": [ + 511, + 657, + 908, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "1705", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/3e9a7d9e9b5b15c18745f19d1f037c682e73622b9600f865f7c14ad21897aa5d.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Average (↑)Spatial (↑)Object (↑)Goal (↑)Long (↑)
Diffusion Policy72.4 ± 0.7%78.3 ± 1.1%92.5 ± 0.7%68.3 ± 1.2%50.5 ± 1.3%
Octo fine-tuned75.1 ± 0.6%78.9 ± 1.0%85.7 ± 0.9%84.6 ± 0.9%51.1 ± 1.3%
OpenVLA fine-tuned76.5 ± 0.6%84.7 ± 0.9%88.4 ± 0.8%79.2 ± 1.0%53.7 ± 1.3%
CoT-VLA-7B (ours)81.13 ± 0.6 %87.5 ± 1.4%91.6 ± 0.5%87.6 ± 0.6%69.0 ± 0.8%
", + "bbox": [ + 147, + 88, + 849, + 181 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1. LIBERO benchmark experimental results. For each task suite (Spatial, Object, Goal, Long), we report the average success rate and standard error across 3 seeds with 500 episodes each. CoT-VLA achieves the best or competitive performance across all LIBERO benchmarks suites compared to baseline approaches. The bolded entries correspond to highest success rates while underlined entries correspond to second-highest.", + "bbox": [ + 88, + 190, + 906, + 247 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Adaptation Phase for Downstream Closed-Loop Deployment For adaptation to downstream tasks, we fine-tune our pretrained model using task-specific robot demonstration data $D_{r}$ collected on the target robot setups. During this phase, we optimize the LLM backbone, projector, and depth transformer while keeping the vision tower frozen, maintaining the same training setup as the pretraining stage. The resulting model can execute new manipulation tasks based on natural language commands $l$ . Algorithm 1 describes our robot control procedure at test time.", + "bbox": [ + 89, + 261, + 485, + 412 + ], + "page_idx": 4 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1 CoT-VLA test-time closed-loop control" + ], + "code_body": "Require: CoT-VLA Model $P_{\\theta}$ , initial state $s_0^{\\mathrm{obs}}$ , language instruction $l$ \n0: $t \\gets 0$ \n0: while True do \n0: sample $\\hat{\\mathbf{s}}_{t+n} \\sim P_{\\theta}(\\mathbf{s}_{t+n} \\mid l, \\mathbf{s}_t^{\\mathrm{obs}})$ \n0: sample $[\\hat{\\mathbf{a}}_t, \\dots, \\hat{\\mathbf{a}}_{t+m}] \\sim P_{\\theta}(\\mathbf{a}_t, \\dots, \\mathbf{a}_{t+m} \\mid l, \\mathbf{s}_t^{\\mathrm{obs}}, \\mathbf{s}_{t+n})$ \n0: for $j = 0$ to $m$ do \n0: execute $\\hat{\\mathbf{a}}_{t+j}$ \n0: end for \n0: $t \\gets t + m + 1$ \n0: $\\mathbf{s}_t^{\\mathrm{obs}} \\gets$ robot observation", + "bbox": [ + 89, + 455, + 483, + 633 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 666, + 223, + 683 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We evaluate the effectiveness of our approach and our system through a set of experiments spanning both simulation benchmarks and real-world robot manipulation tasks. Our experiments aim to addresses following questions:", + "bbox": [ + 89, + 696, + 485, + 758 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- How does our system perform compared to state-of-the-art baselines across multiple benchmarks and embodiments? (Section 4.2)", + "- What is the impact of pretraining, visual chain-of-thought reasoning and hybrid attention on task performance? (Section 4.3)", + "- To what extent does improved generalization in visual reasoning enhance the action prediction capabilities? (Section 4.4)" + ], + "bbox": [ + 89, + 762, + 483, + 896 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 511, + 260, + 702, + 277 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We conduct evaluations across three complementary settings: the LIBERO benchmark [37] for evaluation in simulation environments, the Bridge-V2 platform [60] with its dataset of 45k robot demonstrations, and the Franka-Tabletop setup with a stationary, table-mounted Franka Emika Panda robot with limited 10 to 150 robot demonstrations for each testing scenario.", + "bbox": [ + 511, + 282, + 906, + 388 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "LIBERO Simulation Benchmark We perform evaluation on LIBERO [37], a simulation benchmark comprising four distinct task suites: LIBERO-Spatial, LIBERO-Object, LIBERO-Goal, and LIBERO-Long. Each suite contains 10 diverse tasks with 50 human-teleoperated demonstrations per task, aiming to evaluate the robot's comprehension of spatial relationships, object interactions, and task-specific objectives. We follow the same preprocessed pipeline as in [29]: (1) removing pause intervals from trajectories, (2) standardizing image resolution to $256 \\times 256$ pixels, and (3) applying a 180-degree rotation to all images.", + "bbox": [ + 511, + 407, + 908, + 575 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Bridge-V2 Real-Robot Experiments We use a 6-DoF WidowX robotic arm, following the experimental setup from Bridge-V2 [60]. Our training data has $45\\mathrm{k}$ language-annotated trajectories from the Bridge-V2 dataset, encompassing diverse manipulation tasks. While the dataset was incorporated into the pretraining phase alongside OpenX, we performed additional task-specific fine-tuning exclusively on Bridge-V2 until achieving a training action prediction accuracy threshold of $95\\%$ . Following [29], we evaluate on four tasks designed in [29] to evaluate visual robustness (varying distractors), motion generalization (novel object positions), semantic generalization (unseen language instruction), and language grounding (instruction following).", + "bbox": [ + 511, + 594, + 908, + 791 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Franka-Tabletop Real-Robot Experiments We use a stationary, table-mounted Franka Emika Panda 7-DoF robot arm denoted as Franka-Tabletop. The setup is not seen during the pretraining stage and is designed to assess our model's adaptation capability to novel real-world environments with small amounts of robot demonstrations. We", + "bbox": [ + 511, + 809, + 908, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "1706", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/26969250864415a73f4cc4a6fc6cd168154b53ece3d178986990c9cda66a259c.jpg", + "image_caption": [ + "Figure 4. Franka-Tabletop comparisons. Evaluation across six distinct manipulation tasks, with separate models trained per task. Left: Representative initial states for each task setup. Right: Task-specific success rates and cross-task averages for our method and baselines. CoT-VLA achieves best average performance and demonstrates strong capabilities in both single-instruction and multi-instruction scenarios." + ], + "image_footnote": [], + "bbox": [ + 96, + 93, + 911, + 237 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "perform evaluations across 6 tasks: 3 narrow domain single-instruction tasks for and 3 diverse multi-instruction tasks outlined in Figure 4 and introduced in [29]. For each task, the dataset contains between 10 and 150 demonstrations.", + "bbox": [ + 89, + 306, + 485, + 367 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines We evaluate our approach against four state-of-the-art baselines. Diffusion Policy [10], a state-of-the-art imitation learning algorithm, is trained from scratch for each test scenario in LIBERO and Franka-Tabletop. The implementation incorporates action chunking and proprioception while conditioning on DistilBERT [52] language embeddings. OpenVLA [29] is an open-source VLA model that fine-tunes pretrained vision-language models on the OpenX dataset; and Octo [59] is a generalist model pretrained on OpenX without VLM initialization. For both OpenVLA and Octo, we use their published checkpoints for Bridge-V2 evaluations and fine-tune them for our LIBERO and Franka-Tabletop experiments. SUSIE [2], a two-stage approach, combines instruction-guided image editing for goal generation with a goal-conditioned policy for action generation. We evaluate SUSIE using their published checkpoint on Bridge-V2.", + "bbox": [ + 89, + 385, + 483, + 641 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Evaluations Results", + "text_level": 1, + "bbox": [ + 89, + 650, + 277, + 666 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "LIBERO We present quantitative results in Table 1, where each method is evaluated over 500 trials per task suite, with 3 random seeds. Success rates are reported with means and standard error. Qualitative examples of our method's reasoning and execution trajectories are illustrated in Figure 5. Results demonstrate that CoT-VLA effectively adapts to tasks in the LIBERO simulation environment, achieving best or competitive performance compared to baseline approaches. By analyzing rollout videos of failure cases, we found that baseline methods occasionally overfit to visual cues while disregarding language instructions. Specifically, when initial states appear visually similar across different tasks (e.g., in LIBERO-Spatial), baseline methods execute a different task compared to the commanded task in some episodes. CoT-VLA exhibits better instruction following", + "bbox": [ + 89, + 674, + 485, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ability by first reasoning visually about the desired actions via language-grounded subgoal generation, and then predicting the relevant actions for achieving the goal.", + "bbox": [ + 511, + 306, + 906, + 354 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Bridge-V2 We evaluate CoT-VLA and baselines on the Bridge-V2 benchmark across four generalization categories identified in [29]: visual generalization (\"put eggplant into pot\" with cluttered environments), motion generalization (\"put carrot on plate\" with height variations), semantic generalization (\"take purple grapes out of pot\"), and language grounding (\"put eggplant or red bottle into pot\"). We report the quantitative results in Table 2, where each task is tested with 10 trials. SUSIE [2] generates visually higher-quality goal images through its diffusion prior (see Section 5 for a detailed discussion on our limitations) but achieves lower success rates on tasks involving novel objects or requiring complex language grounding. Compared to OpenVLA [29], CoT-VLA shows slightly lower success rates in visual and language generalization tasks due to grasping failures from action chunking (see Section 5) rather than errors in visual reasoning. However, CoT-VLA demonstrates competitive performance across all four generalization categories overall, achieving comparable or better results to baseline approaches.", + "bbox": [ + 511, + 377, + 908, + 679 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Franka-Tabletop We present quantitative results in Table 4 and example execution trajectories in Figure 5. In this experiment, models are fine-tuned on a relatively small set of demonstrations. While Diffusion Policy achieves top performance on single-instruction tasks (e.g., \"put corn in bowl\"), its performance degrades on multi-instruction tasks involving diverse objects and complex language instructions. Models pretrained on the OpenX dataset - Octo, OpenVLA, and CoT-VLA - demonstrate better adaptation and performance on multi-instruction tasks where language grounding is critical. Overall, CoT-VLA achieves the highest average performance compared to baseline approaches, showing improvements in both single and multi-instruction scenarios.", + "bbox": [ + 511, + 704, + 910, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "1707", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ddecb96de6c4fe2fe9935672eab756d2ca0a6be1e313cef3f5f9502ad7b74dcd.jpg", + "image_caption": [ + "Figure 5. Task execution examples for LIBERO, Bridge-V2, and Franka-Tabletop using CoT-VLA. For each task: Left: text instruction $(l)$ and initial state $(s_0^{\\mathrm{obs}})$ . Middle: generated intermediate goal states $(\\hat{s}_t)$ demonstrating visual chain-of-thought reasoning, where each goal image is conditioned on both the instruction and the most recent observation. Right: final state $(s_T^{\\mathrm{obs}})$ upon task completion. Complete execution trajectories are available in the supplementary video." + ], + "image_footnote": [], + "bbox": [ + 89, + 85, + 880, + 502 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 597, + 243, + 614 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Visual CoT, Hybrid Attention, and Action Chunking We conduct comprehensive ablation studies on two LIBERO benchmark suites: LIBERO-Spatial and LIBERO-Goal. We evaluate four model variants: VLA - a baseline implementation following the standard VLA framework [29], with the same VILA-U backbone but without chain-of-thought reasoning and action chunking; + action chunking - extending the vanilla VLA to predict action sequences of length $m$ ; + hybrid attention - further adding full attention mechanisms for action sequence prediction, as illustrated in Figure 3; and + CoT (ours): our complete approach with hybrid attention mechanism and visual chain-of-thought reasoning.", + "bbox": [ + 88, + 625, + 485, + 806 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As shown in Figure 6, both benchmark suites demonstrate that action sequence prediction consistently outperforms single-action prediction. The addition of hybrid attention mechanisms further improves performance. Our CoT-VLA achieves the best results validating the effectiveness of visual chain-of-thought reasoning for VLA tasks.", + "bbox": [ + 89, + 810, + 483, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/97606d9ff20fa34d7f742170b707e4341531fc2321f2afaf2e87d27bea623221.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
CategorySUSIEOctoOpenVLACoT-VLA
Visual30%35%75%65%
Motion10%10%45%60%
Semantic20%0%40%50%
Language40%40%75%70%
", + "bbox": [ + 519, + 595, + 901, + 688 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Bridge-V2 Comparison. Success rates across four generalization categories, with 10 trials per category and partial credit scoring following [29]. Visual: \"put eggplant into pot\" with cluttered environments; Motion: \"put carrot on plate\" with height variations; Semantic: \"take purple grapes out of pot\"; Language: \"put eggplant or red bottle into pot\".", + "bbox": [ + 511, + 698, + 908, + 782 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "**Pretraining** Our training pipeline has two stages, pretraining VILA-U on the OpenX dataset augmented with actionless video data (Section 3.3), and task-specific post-training on robot demonstration data. To assess the importance of our pretraining stage, we conduct ablation studies on the Franka-Tabletop setup. We report the quantitative results", + "bbox": [ + 511, + 810, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "1708", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6ea0ad3b3ccb98fdac323ac664b2e908e72db861c0add6f21cf035b150c7659f.jpg", + "image_caption": [ + "a)", + "VLA + action chunking" + ], + "image_footnote": [], + "bbox": [ + 96, + 111, + 483, + 223 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8b7eb700dc94010f650f5243af400c3b1badf4c194800db729df8fcc393cb9bc.jpg", + "image_caption": [ + "b)", + "Figure 6. Ablation studies of CoT-VLA components. a) Results on LIBERO-Spatial and LIBERO-Goal benchmarks demonstrate the effectiveness of three components: action chunking, hybrid attention, and visual chain-of-thought reasoning. b) Pretraining ablation experiments on Franka-Tabletop show performance improvements from the OpenX and action-less video pretraining process." + ], + "image_footnote": [], + "bbox": [ + 102, + 247, + 478, + 297 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "in Figure 6. Our results show that CoT-VLA with our pretraining stage achieves a $46.7\\%$ relative improvement, from $53.7\\%$ to $78.8\\%$ , compared to directly fine-tuning the base VILA-U model on Franka-Tabletop demonstrations, demonstrating better downstream task adaptation.", + "bbox": [ + 89, + 412, + 483, + 488 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Better Visual Reasoning Helps", + "text_level": 1, + "bbox": [ + 89, + 500, + 359, + 515 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Unlike prior VLAs that only use robot demonstration data $D_{\\mathrm{r}}$ during training, CoT-VLA also leverages action-less video data $D_{\\mathrm{v}}$ for pretraining through its intermediate visual chain-of-thought reasoning steps. This enables learning of both dynamics and instruction following from captioned videos alone, which are substantially more abundant than robot demonstrations. To investigate how visual reasoning capabilities transfer to robot performance, we conduct an ablation study on the Franka-Tabletop setup using novel, long-horizon tasks that combine two unseen subtasks. We design two tasks – (1) “move the green scallion to the apple-covered book” and (2) “move the green cauliflower to the bear-covered book” – that are challenging for our model’s out-of-distribution generalization. For each task, we collect one demonstration trajectory to obtain ground-truth goal images. We evaluate each task across 5 trials under two conditions: (1) CoT-VLA using its generated goal images and (2) CoT-VLA using ground-truth goal images from the collected demonstrations. As shown in Table 3, using ground-truth goal images improves the absolute success rate by $40\\%$ for both tasks. This performance boost suggests that advances in visual reasoning and goal image generation could directly translate to better robotic task performance. While our method still struggles with out-of-distribution subgoal generation, recent advances in large-scale video and image", + "bbox": [ + 89, + 523, + 483, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "models show promising directions for improving visual reasoning capabilities with scaling.", + "bbox": [ + 511, + 90, + 906, + 121 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/9c04b07ec2b2be0db2c9b1d8532eff46a04cfef1d5cd1ca53de7ccf25d25a5d7.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Sub-task 1Sub-task 2
Generated Goal Images20%0%
Ground-truth Goal Images60%40%
", + "bbox": [ + 521, + 131, + 898, + 193 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3. Better visual reasoning helps. Success rates comparing CoT-VLA using generated versus ground-truth goal images on out-of-distribution tasks. Results demonstrate that improved visual reasoning (simulated by ground-truth goals) leads to better task performance, suggesting that advances in goal generation can translate to improved action execution.", + "bbox": [ + 511, + 204, + 906, + 287 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion, Limitations and Future Work", + "text_level": 1, + "bbox": [ + 511, + 301, + 888, + 316 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we introduce CoT-VLA, bridging vision-language-action models with chain-of-thought reasoning by introducing intermediate visual goals as explicit reasoning steps. Rather than using abstract representations like bounding boxes or keypoints, we propose using subgoal images sampled from videos as an interpretable and effective intermediate representation. We build our system upon VILA-U, demonstrating strong performance across diverse robotic manipulation tasks.", + "bbox": [ + 511, + 325, + 906, + 462 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "While our approach demonstrates effectiveness, there are certain limitations. First, generating intermediate image tokens during inference introduces significant computational overhead compared to direct action generation approaches. Our method requires generating 256 image tokens before action tokens, leading to a $7 \\times$ slowdown on average with an action chunk size of 10. While action chunking and parallel decoding improve inference speed, image generation remains the primary bottleneck. Recent advancement in fast image generation or fast LLM inference techniques could potentially improve the throughput of the model [7, 31, 33, 57, 73] and be integrated into our system. Second, our autoregressive image generation produces lower visual quality compared to state-of-the-art diffusion-based models. Recent advances in unified multimodal models [61, 65, 69, 79] suggest promising directions for improvements. Additionally, while effective, our action chunking approach can introduce discontinuous actions between chunks and lacks high-frequency feedback during execution. These limitations could be addressed through temporal smoothing techniques and per-step prediction approaches similar to those proposed in [10]. Finally, while CoT-VLA leverages action-less video data during pretraining, current computational constraints limit its ability to achieve visual-reasoning generalization for entirely new tasks. Looking forward, we believe recent advances in video/image generation and world models [15, 23, 30, 68, 72] present promising opportunities to enhance generalization capabilities through improved visual reasoning and predictive modeling.", + "bbox": [ + 511, + 463, + 908, + 901 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "1709", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Homanga Bharadhwaj, Debidatta Dwibedi, Abhinav Gupta, Shubham Tulsiani, Carl Doersch, Ted Xiao, Dhruv Shah, Fei Xia, Dorsa Sadigh, and Sean Kirmani. Gen2act: Human video generation in novel scenarios enables generalizable robot manipulation. arXiv preprint arXiv:2409.16283, 2024. 1", + "[2] Kevin Black, Mitsuhiko Nakamoto, Pranav Atreya, Homer Walke, Chelsea Finn, Aviral Kumar, and Sergey Levine. Zero-shot robotic manipulation with pretrained image-editing diffusion models. arXiv preprint arXiv:2310.10639, 2023. 2, 3, 6", + "[3] Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Joseph Dabis, Chelsea Finn, Keerthana Gopalakrishnan, Karol Hausman, Alex Herzog, Jasmine Hsu, et al. Rt-1: Robotics transformer for real-world control at scale. arXiv preprint arXiv:2212.06817, 2022. 1", + "[4] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 3", + "[5] Chi-Lam Cheang, Guangzeng Chen, Ya Jing, Tao Kong, Hang Li, Yifeng Li, Yuxiao Liu, Hongtao Wu, Jiafeng Xu, Yichu Yang, et al. Gr-2: A generative video-language-action model with web-scale knowledge for robot manipulation. arXiv preprint arXiv:2410.06158, 2024. 1", + "[6] Boyuan Chen, Zhuo Xu, Sean Kirmani, Brain Ichter, Dorsa Sadigh, Leonidas Guibas, and Fei Xia. Spatialvlm: Endowing vision-language models with spatial reasoning capabilities. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14455-14465, 2024. 3", + "[7] Junyu Chen, Han Cai, Junsong Chen, Enze Xie, Shang Yang, Haotian Tang, Muyang Li, Yao Lu, and Song Han. Deep compression autoencoder for efficient high-resolution diffusion models. arXiv preprint arXiv:2410.10733, 2024. 8", + "[8] Lawrence Yunliang Chen, Simeon Adebola, and Ken Goldberg. Berkeley UR5 demonstration dataset. https:// sites.google.com/view/berkeley-ur5/home. 1", + "[9] Xi Chen, Xiao Wang, Lucas Beyer, Alexander Kolesnikov, Jialin Wu, Paul Voigtlaender, Basil Mustafa, Sebastian Goodman, Ibrahim Alabdulmohsin, Piotr Padlewski, et al. Pali-3 vision language models: Smaller, faster, stronger. arXiv preprint arXiv:2310.09199, 2023. 2", + "[10] Cheng Chi, Zhenjia Xu, Siyuan Feng, Eric Cousineau, Yilun Du, Benjamin Burchfiel, Russ Tedrake, and Shuran Song. Diffusion policy: Visuomotor policy learning via action diffusion. The International Journal of Robotics Research, page 02783649241273668, 2023. 2, 6, 8", + "[11] Yiming Ding, Carlos Florensa, Pieter Abbeel, and Mariano Pielipp. Goal-conditioned imitation learning. Advances in neural information processing systems, 32, 2019. 2, 3", + "[12] Danny Driess, Fei Xia, Mehdi SM Sajjadi, Corey Lynch, Aakanksha Chowdhery, Brian Ichter, Ayzaan Wahid, Jonathan Thompson, Quan Vuong, Tianhe Yu, et al. Palm-e: An" + ], + "bbox": [ + 93, + 114, + 485, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "embodied multimodal language model. arXiv preprint arXiv:2303.03378, 2023.1,3,4", + "[13] Yuqing Du, Ksenia Konyushkova, Misha Denil, Akhil Raju, Jessica Landon, Felix Hill, Nando de Freitas, and Serkan Cabi. Vision-language models as success detectors. arXiv preprint arXiv:2303.07280, 2023. 3", + "[14] Yilun Du, Mengjiao Yang, Pete Florence, Fei Xia, Ayzaan Wahid, Brian Ichter, Pierre Sermanet, Tianhe Yu, Pieter Abbeel, Joshua B Tenenbaum, et al. Video language planning. arXiv preprint arXiv:2310.10625, 2023. 1", + "[15] Yilun Du, Sherry Yang, Bo Dai, Hanjun Dai, Ofir Nachum, Josh Tenenbaum, Dale Schuurmans, and Pieter Abbeel. Learning universal policies via text-guided video generation. Advances in Neural Information Processing Systems, 36, 2024. 2, 3, 8", + "[16] Frederik Ebert, Yanlai Yang, Karl Schmeckpeper, Bernadette Bucher, Georgios Georgakis, Kostas Daniilidis, Chelsea Finn, and Sergey Levine. Bridge data: Boosting generalization of robotic skills with cross-domain datasets. arXiv preprint arXiv:2109.13396, 2021. 1", + "[17] Zipeng Fu, Qingqing Zhao, Qi Wu, Gordon Wetzstein, and Chelsea Finn. Humanplus: Humanoid shadowing and imitation from humans. arXiv preprint arXiv:2406.10454, 2024. 2", + "[18] Zipeng Fu, Tony Z. Zhao, and Chelsea Finn. Mobile aloha: Learning bimanual mobile manipulation with low-cost whole-body teleoperation. In Conference on Robot Learning (CoRL), 2024. 1", + "[19] Samir Yitzhak Gadre, Mitchell Wortsman, Gabriel Ilharco, Ludwig Schmidt, and Shuran Song. Cows on pasture: Baselines and benchmarks for language-driven zero-shot object navigation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23171-23181, 2023. 3", + "[20] Raghav Goyal, Samira Ebrahimi Kahou, Vincent Michalski, Joanna Materzynska, Susanne Westphal, Heuna Kim, Valentin Haenel, Ingo Fruend, Peter Yianilos, Moritz Mueller-Freitag, et al. The\" something something\" video database for learning and evaluating visual common sense. In Proceedings of the IEEE international conference on computer vision, pages 5842-5850, 2017. 2, 3, 4, 1", + "[21] Huy Ha, Pete Florence, and Shuran Song. Scaling up and distilling down: Language-guided robot skill acquisition. In Conference on Robot Learning, pages 3766-3777. PMLR, 2023. 3", + "[22] William Harvey and Frank Wood. Visual chain-of-thought diffusion models. arXiv preprint arXiv:2303.16187, 2023. 2", + "[23] Anthony Hu, Lloyd Russell, Hudson Yeo, Zak Murez, George Fedoseev, Alex Kendall, Jamie Shotton, and Gianluca Corrado. Gaia-1: A generative world model for autonomous driving. arXiv preprint arXiv:2309.17080, 2023. 8", + "[24] Yushi Hu, Weijia Shi, Xingyu Fu, Dan Roth, Mari Ostendorf, Luke Zettlemoyer, Noah A Smith, and Ranjay Krishna. Visual sketchpad: Sketching as a visual chain of thought for multimodal language models. arXiv preprint arXiv:2406.09403, 2024. 2", + "[25] Chenguang Huang, Oier Mees, Andy Zeng, and Wolfram Burgard. Visual language maps for robot navigation. In 2023" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "1710", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "IEEE International Conference on Robotics and Automation (ICRA), pages 10608-10615. IEEE, 2023. 3", + "[26] Wenlong Huang, Chen Wang, Yunzhu Li, Ruohan Zhang, and Li Fei-Fei. Rekep: Spatio-temporal reasoning of relational keypoint constraints for robotic manipulation. arXiv preprint arXiv:2409.01652, 2024. 3", + "[27] Georgios Kapidis, Ronald Poppe, Elsbeth Van Dam, Lucas Noldus, and Remco Veltkamp. Egocentric hand track and object-based human action recognition. In 2019 IEEE SmartWorld, Ubiquitous Intelligence & Computing, Advanced & Trusted Computing, Scalable Computing & Communications, Cloud & Big Data Computing, Internet of People and Smart City Innovation (SmartWorld/SCALCOM/UIC/ATC/CBDCom/IOP/SCI), pages 922-929. IEEE, 2019. 1, 2, 3, 4", + "[28] Siddharth Karamcheti, Suraj Nair, Ashwin Balakrishna, Percy Liang, Thomas Kollar, and Dorsa Sadigh. Prismatic vlms: Investigating the design space of visually-conditioned language models. arXiv preprint arXiv:2402.07865, 2024. 2", + "[29] Moo Jin Kim, Karl Pertsch, Siddharth Karamcheti, Ted Xiao, Ashwin Balakrishna, Suraj Nair, Rafael Rafailov, Ethan Foster, Grace Lam, Pannag Sanketi, et al. Openvla: An open-source vision-language-action model. arXiv preprint arXiv:2406.09246, 2024. 1, 3, 4, 5, 6, 7", + "[30] Dan Kondratyuk, Lijun Yu, Xiuye Gu, José Lezama, Jonathan Huang, Grant Schindler, Rachel Hornung, Vighnesh Birodkar, Jimmy Yan, Ming-Chang Chiu, et al. Videopoet: A large language model for zero-shot video generation. arXiv preprint arXiv:2312.14125, 2023. 8", + "[31] Woosuk Kwon, Zhuohan Li, Siyuan Zhuang, Ying Sheng, Lianmin Zheng, Cody Hao Yu, Joseph E. Gonzalez, Hao Zhang, and Ion Stoica. Efficient memory management for large language model serving with pagedattention. In Proceedings of the ACM SIGOPS 29th Symposium on Operating Systems Principles, 2023. 8", + "[32] Doyup Lee, Chiheon Kim, Saehoon Kim, Minsu Cho, and Wook-Shin Han. Autoregressive image generation using residual quantization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11523-11532, 2022. 4", + "[33] Yaniv Leviathan, Matan Kalman, and Yossi Matias. Fast inference from transformers via speculative decoding. In International Conference on Machine Learning, pages 19274-19286. PMLR, 2023. 8", + "[34] Boyi Li, Yue Wang, Jiageng Mao, Boris Ivanovic, Sushant Veer, Karen Leung, and Marco Pavone. Driving everywhere with large language model policy adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14948-14957, 2024. 3", + "[35] Junbang Liang, Ruoshi Liu, Ege Ozguroglu, Sruthi Sudhakar, Achal Dave, Pavel Tokmakov, Shuran Song, and Carl Vondrick. Dreamitate: Real-world visuomotor policy learning via video generation. arXiv preprint arXiv:2406.16862, 2024. 2", + "[36] Fanqi Lin, Yingdong Hu, Pingyue Sheng, Chuan Wen, Jiacheng You, and Yang Gao. Data scaling laws in imitation learning for robotic manipulation. arXiv preprint arXiv:2410.18647, 2024. 1" + ], + "bbox": [ + 91, + 90, + 482, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[37] Bo Liu, Yifeng Zhu, Chongkai Gao, Yihao Feng, Qiang Liu, Yuke Zhu, and Peter Stone. Libero: Benchmarking knowledge transfer for lifelong robot learning. arXiv preprint arXiv:2306.03310, 2023. 2, 5, 1", + "[38] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 2", + "[39] Jiasen Lu, Christopher Clark, Sangho Lee, Zichen Zhang, Savya Khosla, Ryan Marten, Derek Hoiem, and Aniruddha Kembhavi. Unified-io 2: Scaling autoregressive multimodal models with vision language audio and action. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26439-26455, 2024. 2", + "[40] Yecheng Jason Ma, William Liang, Guanzhi Wang, De-An Huang, Osbert Bastani, Dinesh Jayaraman, Yuke Zhu, Linxi Fan, and Anima Anandkumar. Eureka: Human-level reward design via coding large language models. arXiv preprint arXiv:Arxiv-2310.12931, 2023.3", + "[41] Arjun Majumdar, Karmesh Yadav, Sergio Arnaud, Jason Ma, Claire Chen, Sneha Silwal, Aryan Jain, Vincent-Pierre Berges, Tingfan Wu, Jay Vakil, et al. Where are we in the search for an artificial visual cortex for embodied intelligence? Advances in Neural Information Processing Systems, 36:655-677, 2023. 3", + "[42] Ajay Mandlekar, Yuke Zhu, Animesh Garg, Jonathan Booher, Max Spero, Albert Tung, Julian Gao, John Emmons, Anchit Gupta, Emre Orbay, et al. Roboturk: A crowdsourcing platform for robotic skill learning through imitation. In Conference on Robot Learning, pages 879-893. PMLR, 2018. 1", + "[43] Oier Mees, Jessica Borja-Diaz, and Wolfram Burgard. Grounding language with visual affordances over unstructured data. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 11576-11582. IEEE, 2023. 1", + "[44] Zawalski Michal, Chen William, Pertsch Karl, Mees Oier, Finn Chelsea, and Levine Sergey. Robotic control via embodied chain-of-thought reasoning. arXiv preprint arXiv:2407.08693, 2024. 1, 2", + "[45] Yao Mu, Qinglong Zhang, Mengkang Hu, Wenhai Wang, Mingyu Ding, Jun Jin, Bin Wang, Jifeng Dai, Yu Qiao, and Ping Luo. Embodiedgpt: Vision-language pre-training via embodied chain of thought. Advances in Neural Information Processing Systems, 36, 2024. 1, 2, 3", + "[46] Ashvin V Nair, Vitchyr Pong, Murtaza Dalal, Shikhar Bahl, Steven Lin, and Sergey Levine. Visual reinforcement learning with imagined goals. Advances in neural information processing systems, 31, 2018. 2, 3", + "[47] Fei Ni, Jianye Hao, Shiguang Wu, Longxin Kou, Jiashun Liu, Yan Zheng, Bin Wang, and Yuzheng Zhuang. Generate subgoal images before act: Unlocking the chain-of-thought reasoning in diffusion model for robot manipulation with multimodal prompts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 2, 3", + "[48] Abby O'Neill, Abdul Rehman, Abhinav Gupta, Abhiram Maddukuri, Abhishek Gupta, Abhishek Padalkar, Abraham" + ], + "bbox": [ + 516, + 90, + 905, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "1711", + "bbox": [ + 483, + 945, + 513, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Lee, Acorn Pooley, Agrim Gupta, Ajay Mandlekar, et al. Open x-embodiment: Robotic learning datasets and rt-x models. arXiv preprint arXiv:2310.08864, 2023. 1, 2, 3, 4", + "[49] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 3", + "[50] Daniel Rose, Vaishnavi Himakunthala, Andy Ouyang, Ryan He, Alex Mei, Yujie Lu, Michael Saxon, Chinmay Sonar, Diba Mirza, and William Yang Wang. Visual chain of thought: bridging logical gaps with multimodal infillings. arXiv preprint arXiv:2305.02317, 2023. 2", + "[51] Erick Rosete-Beas, Oier Mees, Gabriel Kalweit, Joschka Boedecker, and Wolfram Burgard. Latent plans for task-agnostic offline reinforcement learning. In Conference on Robot Learning, pages 1838–1849. PMLR, 2023. 1", + "[52] V Sanh. Distilbert, a distilled version of bert: smaller, faster, cheaper and lighter. arXiv preprint arXiv:1910.01108, 2019. 6", + "[53] Hao Shao, Shengju Qian, Han Xiao, Guanglu Song, Zhuofan Zong, Letian Wang, Yu Liu, and Hongsheng Li. Visual cot: Unleashing chain-of-thought reasoning in multi-modal language models. arXiv preprint arXiv:2403.16999, 2024. 2", + "[54] Mohit Shridhar, Lucas Manuelli, and Dieter Fox. CIport: What and where pathways for robotic manipulation. In Conference on robot learning, pages 894-906. PMLR, 2022. 1", + "[55] Mohit Shridhar, Yat Long Lo, and Stephen James. Generative image as action models. arXiv preprint arXiv:2407.07875, 2024. 2, 3", + "[56] Ishika Singh, Valts Blukis, Arsalan Mousavian, Ankit Goyal, Danfei Xu, Jonathan Tremblay, Dieter Fox, Jesse Thomason, and Animesh Garg. Progprompt: Generating situated robot task plans using large language models. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 11523-11530. IEEE, 2023. 3", + "[57] Yang Song, Prafulla Dhariwal, Mark Chen, and Ilya Sutskever. Consistency models. arXiv preprint arXiv:2303.01469, 2023.8", + "[58] Chameleon Team. Chameleon: Mixed-modal early-fusion foundation models. arXiv preprint arXiv:2405.09818, 2024. 2", + "[59] Octo Model Team, Dibya Ghosh, Homer Walke, Karl Pertsch, Kevin Black, Oier Mees, Sudeep Dasari, Joel Hejna, Tobias Kreiman, Charles Xu, et al. Octo: An open-source generalist robot policy. arXiv preprint arXiv:2405.12213, 2024. 1, 3, 6", + "[60] Homer Rich Walke, Kevin Black, Tony Z Zhao, Quan Vuong, Chongyi Zheng, Philippe Hansen-Estruch, Andre Wang He, Vivek Myers, Moo Jin Kim, Max Du, et al. Bridgedata v2: A dataset for robot learning at scale. In Conference on Robot Learning, pages 1723-1736. PMLR, 2023. 2, 5, 1", + "[61] Xinlong Wang, Xiaosong Zhang, Zhengxiong Luo, Quan Sun, Yufeng Cui, Jinsheng Wang, Fan Zhang, Yueze Wang, Zhen Li, Qiying Yu, et al. Emu3: Next-token prediction is all you need. arXiv preprint arXiv:2409.18869, 2024. 2, 8" + ], + "bbox": [ + 91, + 90, + 485, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[62] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 2022. 2, 3", + "[63] Chuan Wen, Xingyu Lin, John So, Kai Chen, Qi Dou, Yang Gao, and Pieter Abbeel. Any-point trajectory modeling for policy learning. arXiv preprint arXiv:2401.00025, 2023. 1, 2, 3", + "[64] Junjie Wen, Yichen Zhu, Jinming Li, Minjie Zhu, Kun Wu, Zhiyuan Xu, Ran Cheng, Chaomin Shen, Yaxin Peng, Feifei Feng, et al. Tinyvla: Towards fast, data-efficient vision-language-action models for robotic manipulation. arXiv preprint arXiv:2409.12514, 2024.3", + "[65] Chengyue Wu, Xiaokang Chen, Zhiyu Wu, Yiyang Ma, Xingchao Liu, Zizheng Pan, Wen Liu, Zhenda Xie, Xingkai Yu, Chong Ruan, et al. Janus: Decoupling visual encoding for unified multimodal understanding and generation. arXiv preprint arXiv:2410.13848, 2024. 8", + "[66] Hongtao Wu, Ya Jing, Chilam Cheang, Guangzeng Chen, Jiafeng Xu, Xinghang Li, Minghuan Liu, Hang Li, and Tao Kong. Unleashing large-scale video generative pretraining for visual robot manipulation. arXiv preprint arXiv:2312.13139, 2023. 1", + "[67] Yecheng Wu, Zhuoyang Zhang, Junyu Chen, Haotian Tang, Dacheng Li, Yunhao Fang, Ligeng Zhu, Enze Xie, Hongxu Yin, Li Yi, et al. Vila-u: a unified foundation model integrating visual understanding and generation. arXiv preprint arXiv:2409.04429, 2024. 2, 3, 4", + "[68] Jiannan Xiang, Guangyi Liu, Yi Gu, Qiyue Gao, Yuting Ning, Yuheng Zha, Zeyu Feng, Tianhua Tao, Shibo Hao, Yemin Shi, et al. Pandora: Towards general world model with natural language actions and video states. arXiv preprint arXiv:2406.09455, 2024. 8", + "[69] Jinheng Xie, Weijia Mao, Zechen Bai, David Junhao Zhang, Weihao Wang, Kevin Qinghong Lin, Yuchao Gu, Zhijie Chen, Zhenheng Yang, and Mike Zheng Shou. Show-o: One single transformer to unify multimodal understanding and generation. arXiv preprint arXiv:2408.12528, 2024. 2, 8", + "[70] Jonathan Yang, Catherine Glossop, Arjun Bhorkar, Dhruv Shah, Quan Vuong, Chelsea Finn, Dorsa Sadigh, and Sergey Levine. Pushing the limits of cross-embodiment learning for manipulation and navigation. arXiv preprint arXiv:2402.19432, 2024. 1", + "[71] Mengjiao Yang, Yilun Du, Kamyar Ghasemipour, Jonathan Thompson, Dale Schuurmans, and Pieter Abbeel. Learning interactive real-world simulators. arXiv preprint arXiv:2310.06114, 2023. 3", + "[72] Sherry Yang, Jacob Walker, Jack Parker-Holder, Yilun Du, Jake Bruce, Andre Barreto, Pieter Abbeel, and Dale Schuurmans. Video as the new language for real-world decision making. arXiv preprint arXiv:2402.17139, 2024. 8", + "[73] Qihang Yu, Mark Weber, Xueqing Deng, Xiaohui Shen, Daniel Cremers, and Liang-Chieh Chen. An image is worth 32 tokens for reconstruction and generation. arxiv: 2406.07550, 2024.8", + "[74] Wenhao Yu, Nimrod Gileadi, Chuyuan Fu, Sean Kirmani, Kuang-Huei Lee, Montse Gonzalez Arenas, Hao-Tien Lewis" + ], + "bbox": [ + 516, + 90, + 906, + 900 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "1712", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Chiang, Tom Erez, Leonard Hasenclever, Jan Humplik, et al. Language to rewards for robotic skill synthesis. arXiv preprint arXiv:2306.08647, 2023. 3", + "[75] Eric Zelikman, Yuhuai Wu, Jesse Mu, and Noah Goodman. Star: Bootstrapping reasoning with reasoning. Advances in Neural Information Processing Systems, 35:15476-15488, 2022. 2", + "[76] Kaifeng Zhang, Zhao-Heng Yin, Weirui Ye, and Yang Gao. Learning manipulation skills through robot chain-of-thought with sparse failure guidance. arXiv preprint arXiv:2405.13573, 2024. 1, 2", + "[77] Tony Z Zhao, Vikash Kumar, Sergey Levine, and Chelsea Finn. Learning fine-grained bimanual manipulation with low-cost hardware. arXiv preprint arXiv:2304.13705, 2023. 2", + "[78] Haoyu Zhen, Xiaowen Qiu, Peihao Chen, Jincheng Yang, Xin Yan, Yilun Du, Yining Hong, and Chuang Gan. 3d-vla: A 3d vision-language-action generative world model. arXiv preprint arXiv:2403.09631, 2024. 1", + "[79] Chunting Zhou, Lili Yu, Arun Babu, Kushal Tirumala, Michihiro Yasunaga, Leonid Shamis, Jacob Kahn, Xuezhe Ma, Luke Zettlemoyer, and Omer Levy. Transfusion: Predict the next token and diffuse images with one multi-modal model. 2024. 8", + "[80] Enshen Zhou, Yiran Qin, Zhenfei Yin, Yuzhou Huang, Ruimao Zhang, Lu Sheng, Yu Qiao, and Jing Shao. Mine-dreamer: Learning to follow instructions via chain-of-imagination for simulated-world control. arXiv preprint arXiv:2403.12037, 2024. 3", + "[81] Gaoyue Zhou, Victoria Dean, Mohan Kumar Srirama, Aravind Rajeswaran, Jyothish Pari, Kyle Hatch, Aryan Jain, Tianhe Yu, Pieter Abbeel, Lerrel Pinto, et al. Train offline, test online: A real robot learning benchmark. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 9197-9203. IEEE, 2023. 1", + "[82] Xinghao Zhu, Ran Tian, Chenfeng Xu, Mingxiao Huo, Wei Zhan, Masayoshi Tomizuka, and Mingyu Ding. Fanuc manipulation: A dataset for learning-based manipulation with fanuc mate 200id robot. https://sites.google.com/berkeley.edu/fanuc-manipulation, 2023.1", + "[83] Yifeng Zhu, Abhishek Joshi, Peter Stone, and Yuke Zhu. Viola: Imitation learning for vision-based manipulation with object proposal priors. In Conference on Robot Learning, pages 1199-1210. PMLR, 2023. 1" + ], + "bbox": [ + 91, + 90, + 483, + 698 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "1713", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2025/CoT-VLA_ Visual Chain-of-Thought Reasoning for Vision-Language-Action Models/4f2ea357-dbc3-435d-b0c3-c4ec69dfb496_model.json b/2025/CoT-VLA_ Visual Chain-of-Thought Reasoning for Vision-Language-Action Models/4f2ea357-dbc3-435d-b0c3-c4ec69dfb496_model.json new file mode 100644 index 0000000000000000000000000000000000000000..d263b07b96aadce588e94b0b4fa261f307d96248 --- /dev/null +++ b/2025/CoT-VLA_ Visual Chain-of-Thought Reasoning for Vision-Language-Action Models/4f2ea357-dbc3-435d-b0c3-c4ec69dfb496_model.json @@ -0,0 +1,2336 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.243, + 0.131, + 0.758, + 0.177 + ], + "angle": 0, + "content": "CoT-VLA: Visual Chain-of-Thought Reasoning for Vision-Language-Action Models" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.202, + 0.885, + 0.284 + ], + "angle": 0, + "content": "Qingqing Zhao\\(^{1,2,*}\\) Yao Lu\\(^{1}\\) Moo Jin Kim\\(^{2}\\) Zipeng Fu\\(^{2}\\) Zhuoyang Zhang\\(^{3}\\) Yecheng Wu\\(^{1,3}\\) Zhaoshuo Li\\(^{1}\\) Qianli Ma\\(^{1}\\) Song Han\\(^{1,3}\\) Chelsea Finn\\(^{2}\\) Ankur Handa\\(^{1}\\) Tsung-Yi Lin\\(^{1\\dagger}\\) Gordon Wetzstein\\(^{2\\dagger}\\) Ming-Yu Liu\\(^{1\\dagger}\\) Donglai Xiang\\(^{1\\dagger}\\) NVIDIA \\({}^{2}\\)Stanford University \\({}^{3}\\)MIT" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.318, + 0.327, + 0.334 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.351, + 0.486, + 0.653 + ], + "angle": 0, + "content": "Vision-language-action models (VLAs) have shown potential in leveraging pretrained vision-language models and diverse robot demonstrations for learning generalizable sensorimotor control. While this paradigm effectively utilizes large-scale data from both robotic and non-robotic sources, current VLAs primarily focus on direct input-output mappings, lacking the intermediate reasoning steps crucial for complex manipulation tasks. As a result, existing VLAs lack temporal planning or reasoning capabilities. In this paper, we introduce a method that incorporates explicit visual chain-of-thought (CoT) reasoning into vision-language-action models (VLAs) by predicting future image frames autoregressively as visual goals before generating a short action sequence to achieve these goals. We introduce CoT-VLA, a state-of-the-art 7B VLA that can understand and generate visual and action tokens. Our experimental results demonstrate that CoT-VLA achieves strong performance, outperforming the state-of-the-art VLA model by \\(17\\%\\) in real-world manipulation tasks and \\(6\\%\\) in simulation benchmarks. Videos are available at: https://cot-vla.github.io/." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.685, + 0.222, + 0.7 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.711, + 0.486, + 0.862 + ], + "angle": 0, + "content": "Recent advances in robot learning have demonstrated impressive progress in training policies that can act across diverse tasks and environments [1, 5, 12, 14, 18, 29, 36, 44, 45, 48, 54, 59, 63, 66, 70, 76, 78]. One promising direction is vision-language-action (VLA) models, which leverage the rich understanding capabilities of pretrained vision-language models (VLMs) to map natural language instructions and visual observations to robot actions [12, 29, 48]. By training VLMs on robot demonstrations, VLAs inherit their ability to understand diverse scenes, objects, and natural language in-" + }, + { + "type": "image", + "bbox": [ + 0.521, + 0.318, + 0.902, + 0.665 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.674, + 0.909, + 0.828 + ], + "angle": 0, + "content": "Figure 1. Comparison between vanilla VLA and CoT-VLA frameworks. Prior VLA models (top) directly predict robot actions from task inputs without explicit reasoning steps and only use action-annotated robot demonstration data for training. Unlike vanilla VLAs, CoT-VLA (bottom) can also leverage action-less datasets like EPIC-KITCHEN-100 [27] to enhance subgoal image generation ability, unlocking the potential of using abundant unlabeled video data to improve VLA's visual reasoning capability. CoT-VLA first generates a subgoal image as an intermediate reasoning step, and then generate a short action sequence to achieve the subgoal. We outline the robot arm for better visualization." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.909, + 0.903 + ], + "angle": 0, + "content": "structions, leading to better generalization capabilities when fine-tuned for downstream testing scenarios. While these approaches have shown impressive results, they typically map" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.875, + 0.352, + 0.888 + ], + "angle": 0, + "content": "*Work done during an internship at NVIDIA." + }, + { + "type": "page_footnote", + "bbox": [ + 0.111, + 0.888, + 0.202, + 0.901 + ], + "angle": 0, + "content": "Equal Advising" + }, + { + "type": "list", + "bbox": [ + 0.109, + 0.875, + 0.352, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1702" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.137 + ], + "angle": 0, + "content": "directly from observations to actions without explicit intermediate reasoning steps that could improve interpretability and, potentially, performance." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.14, + 0.486, + 0.442 + ], + "angle": 0, + "content": "In the language domain, chain-of-thought (CoT) prompting has emerged as a powerful technique for improving the reasoning capabilities of large language models (LLMs) by encouraging step-by-step thinking [62, 75]. Applying these concepts to robotics presents exciting opportunities for grounding reasoning in text, visual observations, and physical actions. Recent works have made progress in this direction by incorporating intermediate reasoning steps like language descriptions, keypoints, or bounding boxes [15, 44, 45, 63]. These intermediate representations capture abstracted states of scenes, objects, and tasks and often require additional pre-processing pipelines. In our work, we explore subgoal images as an intermediate reasoning step before action generation. These images capture the state of the model's reasoning process and are naturally available within robot demonstration datasets. While prior work has explored subgoal generation and goal-conditioned imitation learning [2, 11, 46, 55], to the best of our knowledge, our approach is the first to integrate these concepts with VLAs as intermediate chain-of-thought reasoning steps." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.445, + 0.486, + 0.672 + ], + "angle": 0, + "content": "We propose visual chain-of-thought reasoning for VLAs, a new method that uses subgoal image generation as a form of chain-of-thought reasoning for robotic tasks. Rather than directly predicting actions, our method first generates a subgoal image that represents the robot's planned state in pixel space, and then conditions its action on both the current observation and the generated subgoal image. This approach allows the model to \"think visually\" about how to accomplish a task before acting. By using the subgoal image as intermediate reasoning step, we leverage information that already exists in robot manipulation data with minimal preprocessing required. Furthermore, since subgoal image generation does not require action annotations, this unlocks the potential of using abundant video data for improved visual reasoning and understanding." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.675, + 0.486, + 0.901 + ], + "angle": 0, + "content": "We build our CoT-VLA system that leverages visual chain-of-thought reasoning upon recent advances in unified multimodal foundation models that can understand and generate text and images [39, 58, 61, 67, 69]. We train our base model [67] on both the Open X-Embodiment dataset [48] and action-less video datasets [20, 27], and then fine-tune the model on task demonstrations collected on downstream robot setups used for deployment and evaluation. We design a hybrid attention mechanism for CoT-VLA: we use causal attention with next-token prediction for text and image generation, and leverage full attention to predict all action dimensions at once. Additionally, inspired by recent advances in robot learning [10, 17, 77], we predict sequences of actions (action chunking) rather than a single action at each timestep. We demonstrate that both action chunking and the hybrid" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.882, + 0.107 + ], + "angle": 0, + "content": "attention mechanism improve the model's performance." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.107, + 0.908, + 0.182 + ], + "angle": 0, + "content": "Through extensive experiments in both simulation benchmarks [37] and real-world experiments[48, 60], we demonstrate that our visual chain-of-thought reasoning helps improve policy performance compared to prior VLA approaches. Our key contributions include:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.183, + 0.907, + 0.227 + ], + "angle": 0, + "content": "- We introduce a method of visual chain-of-thought reasoning through subgoal image generation as an intermediate reasoning step for robotic control." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.228, + 0.907, + 0.288 + ], + "angle": 0, + "content": "- We introduce a system CoT-VLA that incorporates visual chain-of-thought reasoning, and a hybrid attention mechanism that combines causal attention for pixel and text generation and full attention for action prediction." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.289, + 0.907, + 0.363 + ], + "angle": 0, + "content": "- We conduct comprehensive evaluations in both simulation and the real world, demonstrating that visual chain-of-thought reasoning improves VLA performance, and our system achieves state-of-the-art performance across multiple robot platforms and tasks." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.183, + 0.907, + 0.363 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.375, + 0.655, + 0.39 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.4, + 0.909, + 0.807 + ], + "angle": 0, + "content": "Chain-of-Thought (CoT) Reasoning CoT reasoning has gained prominence in natural language processing, particularly for enabling models to perform complex, multi-step reasoning tasks by breaking down problem-solving into sequential, explainable steps. Early work on CoT reasoning [62] has demonstrated the effectiveness of prompting large language models to generate intermediate reasoning steps before arriving at a final answer. Extending this paradigm to the visual domain, researchers have explored multimodal chain-of-thought methods, where visual information is processed iteratively in a stepwise fashion to reason about future outcomes or states, including generating bounding boxes [53], intermediate image infillments using Stable Diffusion [50] or standard Python packages [24], or generating CLIP embeddings [22]. Recently, CoT reasoning has been explored in embodied applications. It can generate textual plans for multi-stage execution [44, 45], point trajectories [63], label bounding boxes of objects and gripper positions as additional observations [44], generate future image trajectories for open-loop following [35, 47], and generate fine-grained reward guidance for reinforcement learning [76]. In this work, we introduce Visual-CoT reasoning for robotic manipulation, where predicted subgoal images serve as intermediate reasoning steps for closed-loop action generation. This approach leverages demonstration videos as natural intermediate reasoning states without requiring additional annotations." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Vision-Language-Action Models Large pretrained vision-language models (VLMs) [9, 28, 38] have emerged as a powerful tool for robot learning, and recent works have explored various approaches to integrate them into robot systems. Several works utilize VLMs as intermediate" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1703" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.096, + 0.907, + 0.288 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.299, + 0.908, + 0.371 + ], + "angle": 0, + "content": "Figure 2. Overview of CoT-VLA framework. We build our model on VILA-U [67], a generative multimodal model pretrained on interleaved text-image data. The base model then trains on robot demonstrations [48] and action-less videos [20, 27]. During deployment, given a visual observation and a text instruction, the model performs visual chain-of-thought reasoning by generating a subgoal image (upper blue) with causal attention. It then generates a short action sequence with full attention \\((\\mathbf{a}_1 \\cdots \\mathbf{a}_n)\\) for robot execution. The system operates in a closed-loop control manner by capturing new observations after executing predicted action sequences." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.396, + 0.486, + 0.772 + ], + "angle": 0, + "content": "components for perception and control, leveraging their strong semantic understanding and reasoning capabilities to decompose complex tasks [21, 26, 34, 56], detect objects [19, 25], or generate dense rewards [13, 40, 74] or goals [2, 11, 15, 46, 47, 55, 71, 80]. Some approaches incorporate VLMs [4, 49, 64] into end-to-end trainable policies by using them as pretrained backbone for better visuo-language representation [12, 29, 41, 48, 59]. Most relevant to our work are recent approaches that fine-tune pretrained VLMs on robot demonstration data for direct action prediction [12, 29, 48]. These VLAs demonstrate improved generalization to novel objects, environments, and natural language instructions through pretraining on internet-scale vision-language datasets, providing a promising direction for transferring visual and language knowledge to robotic control tasks. However, most existing VLAs do not leverage the step-by-step reasoning capabilities demonstrated in large language models, which have been shown to significantly improve performance across various tasks [62]. In the past, researchers have used chain-of-thought reasoning on language instructions or intermediate keypoints/bounding boxes for robotics [6, 13, 45, 63]. We introduce visual chain-of-thought reasoning to the VLA frameworks, using subgoal images as intermediate reasoning steps before action generation." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.785, + 0.2, + 0.801 + ], + "angle": 0, + "content": "3. CoT-VLA" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.486, + 0.903 + ], + "angle": 0, + "content": "In this section, we present our visual chain-of-thought reasoning framework for VLAs. We begin with the formulation of our method (3.1), followed by a detailed description of the system architecture (3.2). We then explain our training procedures (3.3) and outline the deployment strategy for downstream tasks (3.3)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.395, + 0.829, + 0.412 + ], + "angle": 0, + "content": "3.1. Visual Chain-of-Thought Reasoning" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.418, + 0.909, + 0.553 + ], + "angle": 0, + "content": "We consider two types of training data for VLA pretraining: robot demonstrations dataset \\(D_r\\) and action-less videos dataset \\(D_v\\). Robot demonstrations are represented as \\(D_r = \\{(l, \\mathbf{a}_{1\\dots T}, \\mathbf{s}_{1\\dots T})\\}\\), where \\(l\\) denotes the natural language instruction, \\(\\mathbf{a}_{1\\dots T} = \\{\\mathbf{a}_1, \\dots, \\mathbf{a}_T\\}\\) denotes the sequence of robot actions, and \\(\\mathbf{s}_{1\\dots T} = \\{\\mathbf{s}_1, \\dots, \\mathbf{s}_T\\}\\) denotes the visual observations as a sequence of images. Action-less videos \\(D_v = \\{(l, \\mathbf{s}_{1\\dots T})\\}\\) consist of language descriptions and images without action annotations." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.571, + 0.909, + 0.632 + ], + "angle": 0, + "content": "VLA: Vanilla VLA approaches fine-tune a pretrained VLM, \\(P_{\\theta}\\), on \\(D_{r}\\), learning to predict actions \\(\\hat{\\mathbf{a}}_{t+1}\\) directly from the current observation \\(\\mathbf{s}_{t}\\) and language instruction \\(l\\) (Figure 1, top):" + }, + { + "type": "equation", + "bbox": [ + 0.651, + 0.639, + 0.907, + 0.656 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {a}} _ {t} \\sim P _ {\\theta} (\\mathbf {a} _ {t} | \\mathbf {s} _ {t}, l) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.671, + 0.908, + 0.718 + ], + "angle": 0, + "content": "CoT-VLA: Our key insight is to incorporate explicit visual reasoning before action generation. As illustrated in Figure 2, our approach operates in two sequential phases:" + }, + { + "type": "equation", + "bbox": [ + 0.674, + 0.724, + 0.907, + 0.74 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {s}} _ {t + n} \\sim P _ {\\theta} \\left(\\mathbf {s} _ {t + n} \\mid \\mathbf {s} _ {t}, l\\right) \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.558, + 0.742, + 0.907, + 0.759 + ], + "angle": 0, + "content": "\\[\n\\{\\hat {\\mathbf {a}} _ {t}, \\dots , \\hat {\\mathbf {a}} _ {t + m} \\} \\sim P _ {\\theta} (\\{\\mathbf {a} _ {t}, \\dots , \\mathbf {a} _ {t + m} | \\mathbf {s} _ {t}, l, \\mathbf {s} _ {t + n}) \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.911, + 0.903 + ], + "angle": 0, + "content": "where we first predict a subgoal image \\(\\hat{\\mathbf{s}}_{t + n}\\), \\(n\\) frames ahead, as an intermediate visual reasoning step (Equation 2). Then we generate a sequence of \\(m\\) actions to achieve this subgoal state (Equation 3). This enables the model to \"think visually\" first by explicitly reasoning about desired future states before predicting the actions. The visual reasoning step, Equation (2), is trained on both robot demonstrations \\(D_r\\) and actionless videos \\(D_v\\), and the action generation step, Equation (3), is trained on robot demonstrations \\(D_r\\) only." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1704" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.148, + 0.088, + 0.434, + 0.288 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.298, + 0.486, + 0.356 + ], + "angle": 0, + "content": "Figure 3. Hybrid attention mechanism in CoT-VLA. We use causal attention for image or text generation and full attention for action generation. \\([x]\\), \\([\\theta]\\) and \\([g]\\) are special tokens for parallel decoding of actions." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.383, + 0.388, + 0.4 + ], + "angle": 0, + "content": "3.2. The Base Vision-Language Model" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.408, + 0.486, + 0.468 + ], + "angle": 0, + "content": "To enable the visual reasoning capabilities described in Equation (2), we build upon VILA-U [67], an unified multimodal foundation model capable of both understanding and generating image and text tokens." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.47, + 0.485, + 0.758 + ], + "angle": 0, + "content": "VILA-U unifies video, image, and language understanding through an autoregressive next-token prediction framework. At its core is a unified vision tower that encodes visual inputs as discrete tokens aligned with textual information. This enables autoregressive image and video generation while significantly enhancing the understanding capabilities of VLMs that leverage discrete visual features. VILA-U utilizes residual quantization [32] to improve the representational capacity of discrete visual features - incorporating a depth transformer, as introduced in RQ-VAE [32], to gradually predict the residual tokens. The extracted visual features are then passed through a projector before being processed by the LLM backbone. The base model is trained on multimodal pairs including [image, text], [text, image], [video, text], and [text, video]. We use the VILA-U model trained on \\(256 \\times 256\\) resolution images, where each image is encoded into \\(16 \\times 16 \\times 4\\) tokens with a residual depth of 4 [32]. For detailed information about VILA-U training and architecture, we refer readers to [67]." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.771, + 0.285, + 0.787 + ], + "angle": 0, + "content": "3.3. Training Procedures" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.486, + 0.903 + ], + "angle": 0, + "content": "We pretrain the base 7B VILA-U model on a combination of robot demonstrations \\( D_r \\) and action-less videos \\( D_v \\). During training, we optimize three components, the LLM backbone, projector, and depth transformer, while keeping the vision tower fixed. Our training objective has two key components: subgoal image generation with causal attention (2) and action generation with full attention (3)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.908, + 0.198 + ], + "angle": 0, + "content": "Visual Tokens Prediction For subgoal image generation, each training sequence is of form \\((l, \\mathbf{s}_t, \\mathbf{s}_{t + n})\\). We follow the training objective used in [67]. At each visual position \\(j\\), the depth transformer, \\(P_{\\delta}\\), autoregressively predicts \\(D\\) residual tokens \\((k_{j1}, \\dots, k_{jD})\\) based on the LLM-generated code embedding \\(h_j\\). The training objective for visual tokens is then formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.585, + 0.204, + 0.908, + 0.248 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {v i s u a l}} = - \\sum_ {j} \\sum_ {d = 1} ^ {D} \\log P _ {\\delta} \\left(k _ {j d} \\mid k _ {j, < d}\\right) \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.255, + 0.908, + 0.301 + ], + "angle": 0, + "content": "where \\( j \\) indexes the positions containing visual tokens. For a more detailed explanation of this loss function, we refer readers to [67], Section 3.2." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.318, + 0.909, + 0.53 + ], + "angle": 0, + "content": "Action Tokens Prediction For action prediction, each training sequence takes the form \\((l, \\mathbf{s}_t, \\mathbf{s}_{t+n}, \\mathbf{a}_t, \\dots, \\mathbf{a}_{t+m})\\). Each action \\(\\mathbf{a}_i\\) is represented by 7 tokens, with each action dimension independently discretized. Following [29], we map each continuous action dimension into 256 discrete bins, with bin widths determined by uniformly dividing the interval between the 1st and 99th percentiles of the training data's action distribution. We repurpose the 256 least frequently used tokens in the text tokenizer's vocabulary as action bin tokens. Unlike prior works [12, 29, 48], we employ full attention for processing and predicting action tokens, enabling all action tokens to interact with each other. This attention mechanism is illustrated in Figure 3. During training, we minimize the cross-entropy loss for action predictions:" + }, + { + "type": "equation", + "bbox": [ + 0.561, + 0.538, + 0.907, + 0.578 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {a c t i o n}} = - \\sum_ {i = 1} ^ {m} \\log P _ {\\theta} \\left(\\mathbf {a} _ {t} \\dots \\mathbf {a} _ {t + m} \\mid l, s _ {t}, s _ {t + n}\\right) \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.586, + 0.908, + 0.616 + ], + "angle": 0, + "content": "Given a batch of input sequences, The overall training objective combines the action and visual losses:" + }, + { + "type": "equation", + "bbox": [ + 0.641, + 0.627, + 0.907, + 0.643 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {\\text {a c t i o n}} + \\mathcal {L} _ {\\text {v i s u a l}} \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.659, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Pretraining Phase We pretrain CoT-VLA on both robot demonstrations \\( D_r \\) and action-less videos \\( D_v \\) as described in Section 3.1. For robot demonstrations, we curate a subset of the Open X-Embodiment dataset [48] (OpenX). Following the pre-processing pipeline established in OpenVLA [29], we select and process datasets with third-person camera views and single-arm end-effector control (7-DoF). For action-less videos \\( D_v \\), we incorporate the EPIC-KITCHENS [27] and Something-Something V2 [20] datasets. All images are processed at \\( 256 \\times 256 \\) resolution. For visual reasoning, we use subgoal images at future timestep \\( n \\) uniformly sampled from a dataset-specific range \\( [n_l, n_u] \\), where \\( n_l \\) and \\( n_u \\) define the lower and upper bounds of the prediction horizon. We use an action chunk size of 10. For complete dataset specifications and training hyperparameters, please refer to the supplementary material." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1705" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.148, + 0.089, + 0.851, + 0.182 + ], + "angle": 0, + "content": "
Average (↑)Spatial (↑)Object (↑)Goal (↑)Long (↑)
Diffusion Policy72.4 ± 0.7%78.3 ± 1.1%92.5 ± 0.7%68.3 ± 1.2%50.5 ± 1.3%
Octo fine-tuned75.1 ± 0.6%78.9 ± 1.0%85.7 ± 0.9%84.6 ± 0.9%51.1 ± 1.3%
OpenVLA fine-tuned76.5 ± 0.6%84.7 ± 0.9%88.4 ± 0.8%79.2 ± 1.0%53.7 ± 1.3%
CoT-VLA-7B (ours)81.13 ± 0.6 %87.5 ± 1.4%91.6 ± 0.5%87.6 ± 0.6%69.0 ± 0.8%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.089, + 0.191, + 0.908, + 0.248 + ], + "angle": 0, + "content": "Table 1. LIBERO benchmark experimental results. For each task suite (Spatial, Object, Goal, Long), we report the average success rate and standard error across 3 seeds with 500 episodes each. CoT-VLA achieves the best or competitive performance across all LIBERO benchmarks suites compared to baseline approaches. The bolded entries correspond to highest success rates while underlined entries correspond to second-highest." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.262, + 0.486, + 0.414 + ], + "angle": 0, + "content": "Adaptation Phase for Downstream Closed-Loop Deployment For adaptation to downstream tasks, we fine-tune our pretrained model using task-specific robot demonstration data \\(D_{r}\\) collected on the target robot setups. During this phase, we optimize the LLM backbone, projector, and depth transformer while keeping the vision tower frozen, maintaining the same training setup as the pretraining stage. The resulting model can execute new manipulation tasks based on natural language commands \\(l\\). Algorithm 1 describes our robot control procedure at test time." + }, + { + "type": "code_caption", + "bbox": [ + 0.093, + 0.439, + 0.442, + 0.455 + ], + "angle": 0, + "content": "Algorithm 1 CoT-VLA test-time closed-loop control" + }, + { + "type": "algorithm", + "bbox": [ + 0.09, + 0.457, + 0.484, + 0.634 + ], + "angle": 0, + "content": "Require: CoT-VLA Model \\(P_{\\theta}\\), initial state \\(s_0^{\\mathrm{obs}}\\), language instruction \\(l\\) \n0: \\(t \\gets 0\\) \n0: while True do \n0: sample \\(\\hat{\\mathbf{s}}_{t+n} \\sim P_{\\theta}(\\mathbf{s}_{t+n} \\mid l, \\mathbf{s}_t^{\\mathrm{obs}})\\) \n0: sample \\([\\hat{\\mathbf{a}}_t, \\dots, \\hat{\\mathbf{a}}_{t+m}] \\sim P_{\\theta}(\\mathbf{a}_t, \\dots, \\mathbf{a}_{t+m} \\mid l, \\mathbf{s}_t^{\\mathrm{obs}}, \\mathbf{s}_{t+n})\\) \n0: for \\(j = 0\\) to \\(m\\) do \n0: execute \\(\\hat{\\mathbf{a}}_{t+j}\\) \n0: end for \n0: \\(t \\gets t + m + 1\\) \n0: \\(\\mathbf{s}_t^{\\mathrm{obs}} \\gets\\) robot observation" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.667, + 0.224, + 0.684 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.697, + 0.486, + 0.759 + ], + "angle": 0, + "content": "We evaluate the effectiveness of our approach and our system through a set of experiments spanning both simulation benchmarks and real-world robot manipulation tasks. Our experiments aim to addresses following questions:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.763, + 0.483, + 0.806 + ], + "angle": 0, + "content": "- How does our system perform compared to state-of-the-art baselines across multiple benchmarks and embodiments? (Section 4.2)" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.808, + 0.484, + 0.851 + ], + "angle": 0, + "content": "- What is the impact of pretraining, visual chain-of-thought reasoning and hybrid attention on task performance? (Section 4.3)" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.853, + 0.484, + 0.897 + ], + "angle": 0, + "content": "- To what extent does improved generalization in visual reasoning enhance the action prediction capabilities? (Section 4.4)" + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.763, + 0.484, + 0.897 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.261, + 0.704, + 0.278 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.284, + 0.907, + 0.39 + ], + "angle": 0, + "content": "We conduct evaluations across three complementary settings: the LIBERO benchmark [37] for evaluation in simulation environments, the Bridge-V2 platform [60] with its dataset of 45k robot demonstrations, and the Franka-Tabletop setup with a stationary, table-mounted Franka Emika Panda robot with limited 10 to 150 robot demonstrations for each testing scenario." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.409, + 0.909, + 0.577 + ], + "angle": 0, + "content": "LIBERO Simulation Benchmark We perform evaluation on LIBERO [37], a simulation benchmark comprising four distinct task suites: LIBERO-Spatial, LIBERO-Object, LIBERO-Goal, and LIBERO-Long. Each suite contains 10 diverse tasks with 50 human-teleoperated demonstrations per task, aiming to evaluate the robot's comprehension of spatial relationships, object interactions, and task-specific objectives. We follow the same preprocessed pipeline as in [29]: (1) removing pause intervals from trajectories, (2) standardizing image resolution to \\(256 \\times 256\\) pixels, and (3) applying a 180-degree rotation to all images." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.595, + 0.909, + 0.792 + ], + "angle": 0, + "content": "Bridge-V2 Real-Robot Experiments We use a 6-DoF WidowX robotic arm, following the experimental setup from Bridge-V2 [60]. Our training data has \\(45\\mathrm{k}\\) language-annotated trajectories from the Bridge-V2 dataset, encompassing diverse manipulation tasks. While the dataset was incorporated into the pretraining phase alongside OpenX, we performed additional task-specific fine-tuning exclusively on Bridge-V2 until achieving a training action prediction accuracy threshold of \\(95\\%\\). Following [29], we evaluate on four tasks designed in [29] to evaluate visual robustness (varying distractors), motion generalization (novel object positions), semantic generalization (unseen language instruction), and language grounding (instruction following)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Franka-Tabletop Real-Robot Experiments We use a stationary, table-mounted Franka Emika Panda 7-DoF robot arm denoted as Franka-Tabletop. The setup is not seen during the pretraining stage and is designed to assess our model's adaptation capability to novel real-world environments with small amounts of robot demonstrations. We" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "1706" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.098, + 0.094, + 0.912, + 0.238 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.251, + 0.908, + 0.294 + ], + "angle": 0, + "content": "Figure 4. Franka-Tabletop comparisons. Evaluation across six distinct manipulation tasks, with separate models trained per task. Left: Representative initial states for each task setup. Right: Task-specific success rates and cross-task averages for our method and baselines. CoT-VLA achieves best average performance and demonstrates strong capabilities in both single-instruction and multi-instruction scenarios." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.307, + 0.486, + 0.368 + ], + "angle": 0, + "content": "perform evaluations across 6 tasks: 3 narrow domain single-instruction tasks for and 3 diverse multi-instruction tasks outlined in Figure 4 and introduced in [29]. For each task, the dataset contains between 10 and 150 demonstrations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.386, + 0.485, + 0.642 + ], + "angle": 0, + "content": "Baselines We evaluate our approach against four state-of-the-art baselines. Diffusion Policy [10], a state-of-the-art imitation learning algorithm, is trained from scratch for each test scenario in LIBERO and Franka-Tabletop. The implementation incorporates action chunking and proprioception while conditioning on DistilBERT [52] language embeddings. OpenVLA [29] is an open-source VLA model that fine-tunes pretrained vision-language models on the OpenX dataset; and Octo [59] is a generalist model pretrained on OpenX without VLM initialization. For both OpenVLA and Octo, we use their published checkpoints for Bridge-V2 evaluations and fine-tune them for our LIBERO and Franka-Tabletop experiments. SUSIE [2], a two-stage approach, combines instruction-guided image editing for goal generation with a goal-conditioned policy for action generation. We evaluate SUSIE using their published checkpoint on Bridge-V2." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.651, + 0.279, + 0.667 + ], + "angle": 0, + "content": "4.2. Evaluations Results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.675, + 0.486, + 0.901 + ], + "angle": 0, + "content": "LIBERO We present quantitative results in Table 1, where each method is evaluated over 500 trials per task suite, with 3 random seeds. Success rates are reported with means and standard error. Qualitative examples of our method's reasoning and execution trajectories are illustrated in Figure 5. Results demonstrate that CoT-VLA effectively adapts to tasks in the LIBERO simulation environment, achieving best or competitive performance compared to baseline approaches. By analyzing rollout videos of failure cases, we found that baseline methods occasionally overfit to visual cues while disregarding language instructions. Specifically, when initial states appear visually similar across different tasks (e.g., in LIBERO-Spatial), baseline methods execute a different task compared to the commanded task in some episodes. CoT-VLA exhibits better instruction following" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.307, + 0.908, + 0.355 + ], + "angle": 0, + "content": "ability by first reasoning visually about the desired actions via language-grounded subgoal generation, and then predicting the relevant actions for achieving the goal." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.378, + 0.909, + 0.68 + ], + "angle": 0, + "content": "Bridge-V2 We evaluate CoT-VLA and baselines on the Bridge-V2 benchmark across four generalization categories identified in [29]: visual generalization (\"put eggplant into pot\" with cluttered environments), motion generalization (\"put carrot on plate\" with height variations), semantic generalization (\"take purple grapes out of pot\"), and language grounding (\"put eggplant or red bottle into pot\"). We report the quantitative results in Table 2, where each task is tested with 10 trials. SUSIE [2] generates visually higher-quality goal images through its diffusion prior (see Section 5 for a detailed discussion on our limitations) but achieves lower success rates on tasks involving novel objects or requiring complex language grounding. Compared to OpenVLA [29], CoT-VLA shows slightly lower success rates in visual and language generalization tasks due to grasping failures from action chunking (see Section 5) rather than errors in visual reasoning. However, CoT-VLA demonstrates competitive performance across all four generalization categories overall, achieving comparable or better results to baseline approaches." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.911, + 0.901 + ], + "angle": 0, + "content": "Franka-Tabletop We present quantitative results in Table 4 and example execution trajectories in Figure 5. In this experiment, models are fine-tuned on a relatively small set of demonstrations. While Diffusion Policy achieves top performance on single-instruction tasks (e.g., \"put corn in bowl\"), its performance degrades on multi-instruction tasks involving diverse objects and complex language instructions. Models pretrained on the OpenX dataset - Octo, OpenVLA, and CoT-VLA - demonstrate better adaptation and performance on multi-instruction tasks where language grounding is critical. Overall, CoT-VLA achieves the highest average performance compared to baseline approaches, showing improvements in both single and multi-instruction scenarios." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "1707" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.091, + 0.087, + 0.881, + 0.503 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.517, + 0.907, + 0.573 + ], + "angle": 0, + "content": "Figure 5. Task execution examples for LIBERO, Bridge-V2, and Franka-Tabletop using CoT-VLA. For each task: Left: text instruction \\((l)\\) and initial state \\((s_0^{\\mathrm{obs}})\\). Middle: generated intermediate goal states \\((\\hat{s}_t)\\) demonstrating visual chain-of-thought reasoning, where each goal image is conditioned on both the instruction and the most recent observation. Right: final state \\((s_T^{\\mathrm{obs}})\\) upon task completion. Complete execution trajectories are available in the supplementary video." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.598, + 0.245, + 0.615 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.625, + 0.486, + 0.807 + ], + "angle": 0, + "content": "Visual CoT, Hybrid Attention, and Action Chunking We conduct comprehensive ablation studies on two LIBERO benchmark suites: LIBERO-Spatial and LIBERO-Goal. We evaluate four model variants: VLA - a baseline implementation following the standard VLA framework [29], with the same VILA-U backbone but without chain-of-thought reasoning and action chunking; + action chunking - extending the vanilla VLA to predict action sequences of length \\( m \\); + hybrid attention - further adding full attention mechanisms for action sequence prediction, as illustrated in Figure 3; and + CoT (ours): our complete approach with hybrid attention mechanism and visual chain-of-thought reasoning." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.485, + 0.902 + ], + "angle": 0, + "content": "As shown in Figure 6, both benchmark suites demonstrate that action sequence prediction consistently outperforms single-action prediction. The addition of hybrid attention mechanisms further improves performance. Our CoT-VLA achieves the best results validating the effectiveness of visual chain-of-thought reasoning for VLA tasks." + }, + { + "type": "table", + "bbox": [ + 0.521, + 0.596, + 0.902, + 0.689 + ], + "angle": 0, + "content": "
CategorySUSIEOctoOpenVLACoT-VLA
Visual30%35%75%65%
Motion10%10%45%60%
Semantic20%0%40%50%
Language40%40%75%70%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.512, + 0.699, + 0.909, + 0.783 + ], + "angle": 0, + "content": "Table 2. Bridge-V2 Comparison. Success rates across four generalization categories, with 10 trials per category and partial credit scoring following [29]. Visual: \"put eggplant into pot\" with cluttered environments; Motion: \"put carrot on plate\" with height variations; Semantic: \"take purple grapes out of pot\"; Language: \"put eggplant or red bottle into pot\"." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.909, + 0.903 + ], + "angle": 0, + "content": "**Pretraining** Our training pipeline has two stages, pretraining VILA-U on the OpenX dataset augmented with actionless video data (Section 3.3), and task-specific post-training on robot demonstration data. To assess the importance of our pretraining stage, we conduct ablation studies on the Franka-Tabletop setup. We report the quantitative results" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "1708" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.104, + 0.098, + 0.12, + 0.11 + ], + "angle": 0, + "content": "a)" + }, + { + "type": "image_caption", + "bbox": [ + 0.212, + 0.087, + 0.401, + 0.101 + ], + "angle": 0, + "content": "VLA + action chunking" + }, + { + "type": "image", + "bbox": [ + 0.098, + 0.112, + 0.484, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.104, + 0.235, + 0.12, + 0.248 + ], + "angle": 0, + "content": "b)" + }, + { + "type": "image", + "bbox": [ + 0.104, + 0.248, + 0.48, + 0.299 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.313, + 0.484, + 0.398 + ], + "angle": 0, + "content": "Figure 6. Ablation studies of CoT-VLA components. a) Results on LIBERO-Spatial and LIBERO-Goal benchmarks demonstrate the effectiveness of three components: action chunking, hybrid attention, and visual chain-of-thought reasoning. b) Pretraining ablation experiments on Franka-Tabletop show performance improvements from the OpenX and action-less video pretraining process." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.413, + 0.484, + 0.489 + ], + "angle": 0, + "content": "in Figure 6. Our results show that CoT-VLA with our pretraining stage achieves a \\(46.7\\%\\) relative improvement, from \\(53.7\\%\\) to \\(78.8\\%\\), compared to directly fine-tuning the base VILA-U model on Franka-Tabletop demonstrations, demonstrating better downstream task adaptation." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.5, + 0.361, + 0.516 + ], + "angle": 0, + "content": "4.4. Better Visual Reasoning Helps" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.524, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Unlike prior VLAs that only use robot demonstration data \\( D_{\\mathrm{r}} \\) during training, CoT-VLA also leverages action-less video data \\( D_{\\mathrm{v}} \\) for pretraining through its intermediate visual chain-of-thought reasoning steps. This enables learning of both dynamics and instruction following from captioned videos alone, which are substantially more abundant than robot demonstrations. To investigate how visual reasoning capabilities transfer to robot performance, we conduct an ablation study on the Franka-Tabletop setup using novel, long-horizon tasks that combine two unseen subtasks. We design two tasks – (1) “move the green scallion to the apple-covered book” and (2) “move the green cauliflower to the bear-covered book” – that are challenging for our model’s out-of-distribution generalization. For each task, we collect one demonstration trajectory to obtain ground-truth goal images. We evaluate each task across 5 trials under two conditions: (1) CoT-VLA using its generated goal images and (2) CoT-VLA using ground-truth goal images from the collected demonstrations. As shown in Table 3, using ground-truth goal images improves the absolute success rate by \\( 40\\% \\) for both tasks. This performance boost suggests that advances in visual reasoning and goal image generation could directly translate to better robotic task performance. While our method still struggles with out-of-distribution subgoal generation, recent advances in large-scale video and image" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.908, + 0.122 + ], + "angle": 0, + "content": "models show promising directions for improving visual reasoning capabilities with scaling." + }, + { + "type": "table", + "bbox": [ + 0.522, + 0.132, + 0.899, + 0.194 + ], + "angle": 0, + "content": "
Sub-task 1Sub-task 2
Generated Goal Images20%0%
Ground-truth Goal Images60%40%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.512, + 0.205, + 0.907, + 0.289 + ], + "angle": 0, + "content": "Table 3. Better visual reasoning helps. Success rates comparing CoT-VLA using generated versus ground-truth goal images on out-of-distribution tasks. Results demonstrate that improved visual reasoning (simulated by ground-truth goals) leads to better task performance, suggesting that advances in goal generation can translate to improved action execution." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.302, + 0.89, + 0.318 + ], + "angle": 0, + "content": "5. Conclusion, Limitations and Future Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.327, + 0.907, + 0.463 + ], + "angle": 0, + "content": "In this work, we introduce CoT-VLA, bridging vision-language-action models with chain-of-thought reasoning by introducing intermediate visual goals as explicit reasoning steps. Rather than using abstract representations like bounding boxes or keypoints, we propose using subgoal images sampled from videos as an interpretable and effective intermediate representation. We build our system upon VILA-U, demonstrating strong performance across diverse robotic manipulation tasks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.464, + 0.909, + 0.902 + ], + "angle": 0, + "content": "While our approach demonstrates effectiveness, there are certain limitations. First, generating intermediate image tokens during inference introduces significant computational overhead compared to direct action generation approaches. Our method requires generating 256 image tokens before action tokens, leading to a \\(7 \\times\\) slowdown on average with an action chunk size of 10. While action chunking and parallel decoding improve inference speed, image generation remains the primary bottleneck. Recent advancement in fast image generation or fast LLM inference techniques could potentially improve the throughput of the model [7, 31, 33, 57, 73] and be integrated into our system. Second, our autoregressive image generation produces lower visual quality compared to state-of-the-art diffusion-based models. Recent advances in unified multimodal models [61, 65, 69, 79] suggest promising directions for improvements. Additionally, while effective, our action chunking approach can introduce discontinuous actions between chunks and lacks high-frequency feedback during execution. These limitations could be addressed through temporal smoothing techniques and per-step prediction approaches similar to those proposed in [10]. Finally, while CoT-VLA leverages action-less video data during pretraining, current computational constraints limit its ability to achieve visual-reasoning generalization for entirely new tasks. Looking forward, we believe recent advances in video/image generation and world models [15, 23, 30, 68, 72] present promising opportunities to enhance generalization capabilities through improved visual reasoning and predictive modeling." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1709" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.486, + 0.197 + ], + "angle": 0, + "content": "[1] Homanga Bharadhwaj, Debidatta Dwibedi, Abhinav Gupta, Shubham Tulsiani, Carl Doersch, Ted Xiao, Dhruv Shah, Fei Xia, Dorsa Sadigh, and Sean Kirmani. Gen2act: Human video generation in novel scenarios enables generalizable robot manipulation. arXiv preprint arXiv:2409.16283, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.2, + 0.485, + 0.267 + ], + "angle": 0, + "content": "[2] Kevin Black, Mitsuhiko Nakamoto, Pranav Atreya, Homer Walke, Chelsea Finn, Aviral Kumar, and Sergey Levine. Zero-shot robotic manipulation with pretrained image-editing diffusion models. arXiv preprint arXiv:2310.10639, 2023. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.271, + 0.484, + 0.338 + ], + "angle": 0, + "content": "[3] Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Joseph Dabis, Chelsea Finn, Keerthana Gopalakrishnan, Karol Hausman, Alex Herzog, Jasmine Hsu, et al. Rt-1: Robotics transformer for real-world control at scale. arXiv preprint arXiv:2212.06817, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.341, + 0.484, + 0.409 + ], + "angle": 0, + "content": "[4] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.411, + 0.484, + 0.479 + ], + "angle": 0, + "content": "[5] Chi-Lam Cheang, Guangzeng Chen, Ya Jing, Tao Kong, Hang Li, Yifeng Li, Yuxiao Liu, Hongtao Wu, Jiafeng Xu, Yichu Yang, et al. Gr-2: A generative video-language-action model with web-scale knowledge for robot manipulation. arXiv preprint arXiv:2410.06158, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.481, + 0.484, + 0.562 + ], + "angle": 0, + "content": "[6] Boyuan Chen, Zhuo Xu, Sean Kirmani, Brain Ichter, Dorsa Sadigh, Leonidas Guibas, and Fei Xia. Spatialvlm: Endowing vision-language models with spatial reasoning capabilities. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14455-14465, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.565, + 0.484, + 0.619 + ], + "angle": 0, + "content": "[7] Junyu Chen, Han Cai, Junsong Chen, Enze Xie, Shang Yang, Haotian Tang, Muyang Li, Yao Lu, and Song Han. Deep compression autoencoder for efficient high-resolution diffusion models. arXiv preprint arXiv:2410.10733, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.621, + 0.485, + 0.675 + ], + "angle": 0, + "content": "[8] Lawrence Yunliang Chen, Simeon Adebola, and Ken Goldberg. Berkeley UR5 demonstration dataset. https:// sites.google.com/view/berkeley-ur5/home. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.677, + 0.484, + 0.746 + ], + "angle": 0, + "content": "[9] Xi Chen, Xiao Wang, Lucas Beyer, Alexander Kolesnikov, Jialin Wu, Paul Voigtlaender, Basil Mustafa, Sebastian Goodman, Ibrahim Alabdulmohsin, Piotr Padlewski, et al. Pali-3 vision language models: Smaller, faster, stronger. arXiv preprint arXiv:2310.09199, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.747, + 0.484, + 0.815 + ], + "angle": 0, + "content": "[10] Cheng Chi, Zhenjia Xu, Siyuan Feng, Eric Cousineau, Yilun Du, Benjamin Burchfiel, Russ Tedrake, and Shuran Song. Diffusion policy: Visuomotor policy learning via action diffusion. The International Journal of Robotics Research, page 02783649241273668, 2023. 2, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.818, + 0.482, + 0.859 + ], + "angle": 0, + "content": "[11] Yiming Ding, Carlos Florensa, Pieter Abbeel, and Mariano Pielipp. Goal-conditioned imitation learning. Advances in neural information processing systems, 32, 2019. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.86, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[12] Danny Driess, Fei Xia, Mehdi SM Sajjadi, Corey Lynch, Aakanksha Chowdhery, Brian Ichter, Ayzaan Wahid, Jonathan Thompson, Quan Vuong, Tianhe Yu, et al. Palm-e: An" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.486, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.907, + 0.12 + ], + "angle": 0, + "content": "embodied multimodal language model. arXiv preprint arXiv:2303.03378, 2023.1,3,4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.908, + 0.176 + ], + "angle": 0, + "content": "[13] Yuqing Du, Ksenia Konyushkova, Misha Denil, Akhil Raju, Jessica Landon, Felix Hill, Nando de Freitas, and Serkan Cabi. Vision-language models as success detectors. arXiv preprint arXiv:2303.07280, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.178, + 0.907, + 0.232 + ], + "angle": 0, + "content": "[14] Yilun Du, Mengjiao Yang, Pete Florence, Fei Xia, Ayzaan Wahid, Brian Ichter, Pierre Sermanet, Tianhe Yu, Pieter Abbeel, Joshua B Tenenbaum, et al. Video language planning. arXiv preprint arXiv:2310.10625, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.234, + 0.908, + 0.302 + ], + "angle": 0, + "content": "[15] Yilun Du, Sherry Yang, Bo Dai, Hanjun Dai, Ofir Nachum, Josh Tenenbaum, Dale Schuurmans, and Pieter Abbeel. Learning universal policies via text-guided video generation. Advances in Neural Information Processing Systems, 36, 2024. 2, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.304, + 0.907, + 0.37 + ], + "angle": 0, + "content": "[16] Frederik Ebert, Yanlai Yang, Karl Schmeckpeper, Bernadette Bucher, Georgios Georgakis, Kostas Daniilidis, Chelsea Finn, and Sergey Levine. Bridge data: Boosting generalization of robotic skills with cross-domain datasets. arXiv preprint arXiv:2109.13396, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.373, + 0.908, + 0.425 + ], + "angle": 0, + "content": "[17] Zipeng Fu, Qingqing Zhao, Qi Wu, Gordon Wetzstein, and Chelsea Finn. Humanplus: Humanoid shadowing and imitation from humans. arXiv preprint arXiv:2406.10454, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.428, + 0.908, + 0.481 + ], + "angle": 0, + "content": "[18] Zipeng Fu, Tony Z. Zhao, and Chelsea Finn. Mobile aloha: Learning bimanual mobile manipulation with low-cost whole-body teleoperation. In Conference on Robot Learning (CoRL), 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.484, + 0.908, + 0.565 + ], + "angle": 0, + "content": "[19] Samir Yitzhak Gadre, Mitchell Wortsman, Gabriel Ilharco, Ludwig Schmidt, and Shuran Song. Cows on pasture: Baselines and benchmarks for language-driven zero-shot object navigation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23171-23181, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.568, + 0.908, + 0.663 + ], + "angle": 0, + "content": "[20] Raghav Goyal, Samira Ebrahimi Kahou, Vincent Michalski, Joanna Materzynska, Susanne Westphal, Heuna Kim, Valentin Haenel, Ingo Fruend, Peter Yianilos, Moritz Mueller-Freitag, et al. The\" something something\" video database for learning and evaluating visual common sense. In Proceedings of the IEEE international conference on computer vision, pages 5842-5850, 2017. 2, 3, 4, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.665, + 0.907, + 0.718 + ], + "angle": 0, + "content": "[21] Huy Ha, Pete Florence, and Shuran Song. Scaling up and distilling down: Language-guided robot skill acquisition. In Conference on Robot Learning, pages 3766-3777. PMLR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.721, + 0.907, + 0.747 + ], + "angle": 0, + "content": "[22] William Harvey and Frank Wood. Visual chain-of-thought diffusion models. arXiv preprint arXiv:2303.16187, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.749, + 0.907, + 0.803 + ], + "angle": 0, + "content": "[23] Anthony Hu, Lloyd Russell, Hudson Yeo, Zak Murez, George Fedoseev, Alex Kendall, Jamie Shotton, and Gianluca Corrado. Gaia-1: A generative world model for autonomous driving. arXiv preprint arXiv:2309.17080, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.805, + 0.908, + 0.871 + ], + "angle": 0, + "content": "[24] Yushi Hu, Weijia Shi, Xingyu Fu, Dan Roth, Mari Ostendorf, Luke Zettlemoyer, Noah A Smith, and Ranjay Krishna. Visual sketchpad: Sketching as a visual chain of thought for multimodal language models. arXiv preprint arXiv:2406.09403, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.907, + 0.901 + ], + "angle": 0, + "content": "[25] Chenguang Huang, Oier Mees, Andy Zeng, and Wolfram Burgard. Visual language maps for robot navigation. In 2023" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "1710" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.12 + ], + "angle": 0, + "content": "IEEE International Conference on Robotics and Automation (ICRA), pages 10608-10615. IEEE, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.178 + ], + "angle": 0, + "content": "[26] Wenlong Huang, Chen Wang, Yunzhu Li, Ruohan Zhang, and Li Fei-Fei. Rekep: Spatio-temporal reasoning of relational keypoint constraints for robotic manipulation. arXiv preprint arXiv:2409.01652, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.18, + 0.483, + 0.303 + ], + "angle": 0, + "content": "[27] Georgios Kapidis, Ronald Poppe, Elsbeth Van Dam, Lucas Noldus, and Remco Veltkamp. Egocentric hand track and object-based human action recognition. In 2019 IEEE SmartWorld, Ubiquitous Intelligence & Computing, Advanced & Trusted Computing, Scalable Computing & Communications, Cloud & Big Data Computing, Internet of People and Smart City Innovation (SmartWorld/SCALCOM/UIC/ATC/CBDCom/IOP/SCI), pages 922-929. IEEE, 2019. 1, 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.305, + 0.483, + 0.36 + ], + "angle": 0, + "content": "[28] Siddharth Karamcheti, Suraj Nair, Ashwin Balakrishna, Percy Liang, Thomas Kollar, and Dorsa Sadigh. Prismatic vlms: Investigating the design space of visually-conditioned language models. arXiv preprint arXiv:2402.07865, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.362, + 0.483, + 0.431 + ], + "angle": 0, + "content": "[29] Moo Jin Kim, Karl Pertsch, Siddharth Karamcheti, Ted Xiao, Ashwin Balakrishna, Suraj Nair, Rafael Rafailov, Ethan Foster, Grace Lam, Pannag Sanketi, et al. Openvla: An open-source vision-language-action model. arXiv preprint arXiv:2406.09246, 2024. 1, 3, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.433, + 0.483, + 0.502 + ], + "angle": 0, + "content": "[30] Dan Kondratyuk, Lijun Yu, Xiuye Gu, José Lezama, Jonathan Huang, Grant Schindler, Rachel Hornung, Vighnesh Birodkar, Jimmy Yan, Ming-Chang Chiu, et al. Videopoet: A large language model for zero-shot video generation. arXiv preprint arXiv:2312.14125, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.504, + 0.483, + 0.588 + ], + "angle": 0, + "content": "[31] Woosuk Kwon, Zhuohan Li, Siyuan Zhuang, Ying Sheng, Lianmin Zheng, Cody Hao Yu, Joseph E. Gonzalez, Hao Zhang, and Ion Stoica. Efficient memory management for large language model serving with pagedattention. In Proceedings of the ACM SIGOPS 29th Symposium on Operating Systems Principles, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.589, + 0.483, + 0.657 + ], + "angle": 0, + "content": "[32] Doyup Lee, Chiheon Kim, Saehoon Kim, Minsu Cho, and Wook-Shin Han. Autoregressive image generation using residual quantization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11523-11532, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.66, + 0.483, + 0.715 + ], + "angle": 0, + "content": "[33] Yaniv Leviathan, Matan Kalman, and Yossi Matias. Fast inference from transformers via speculative decoding. In International Conference on Machine Learning, pages 19274-19286. PMLR, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.717, + 0.483, + 0.787 + ], + "angle": 0, + "content": "[34] Boyi Li, Yue Wang, Jiageng Mao, Boris Ivanovic, Sushant Veer, Karen Leung, and Marco Pavone. Driving everywhere with large language model policy adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14948-14957, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.788, + 0.483, + 0.844 + ], + "angle": 0, + "content": "[35] Junbang Liang, Ruoshi Liu, Ege Ozguroglu, Sruthi Sudhakar, Achal Dave, Pavel Tokmakov, Shuran Song, and Carl Vondrick. Dreamitate: Real-world visuomotor policy learning via video generation. arXiv preprint arXiv:2406.16862, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.845, + 0.483, + 0.9 + ], + "angle": 0, + "content": "[36] Fanqi Lin, Yingdong Hu, Pingyue Sheng, Chuan Wen, Jiacheng You, and Yang Gao. Data scaling laws in imitation learning for robotic manipulation. arXiv preprint arXiv:2410.18647, 2024. 1" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.906, + 0.147 + ], + "angle": 0, + "content": "[37] Bo Liu, Yifeng Zhu, Chongkai Gao, Yihao Feng, Qiang Liu, Yuke Zhu, and Peter Stone. Libero: Benchmarking knowledge transfer for lifelong robot learning. arXiv preprint arXiv:2306.03310, 2023. 2, 5, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.15, + 0.906, + 0.191 + ], + "angle": 0, + "content": "[38] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.193, + 0.906, + 0.276 + ], + "angle": 0, + "content": "[39] Jiasen Lu, Christopher Clark, Sangho Lee, Zichen Zhang, Savya Khosla, Ryan Marten, Derek Hoiem, and Aniruddha Kembhavi. Unified-io 2: Scaling autoregressive multimodal models with vision language audio and action. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26439-26455, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.278, + 0.906, + 0.346 + ], + "angle": 0, + "content": "[40] Yecheng Jason Ma, William Liang, Guanzhi Wang, De-An Huang, Osbert Bastani, Dinesh Jayaraman, Yuke Zhu, Linxi Fan, and Anima Anandkumar. Eureka: Human-level reward design via coding large language models. arXiv preprint arXiv:Arxiv-2310.12931, 2023.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.349, + 0.906, + 0.43 + ], + "angle": 0, + "content": "[41] Arjun Majumdar, Karmesh Yadav, Sergio Arnaud, Jason Ma, Claire Chen, Sneha Silwal, Aryan Jain, Vincent-Pierre Berges, Tingfan Wu, Jay Vakil, et al. Where are we in the search for an artificial visual cortex for embodied intelligence? Advances in Neural Information Processing Systems, 36:655-677, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.433, + 0.906, + 0.516 + ], + "angle": 0, + "content": "[42] Ajay Mandlekar, Yuke Zhu, Animesh Garg, Jonathan Booher, Max Spero, Albert Tung, Julian Gao, John Emmons, Anchit Gupta, Emre Orbay, et al. Roboturk: A crowdsourcing platform for robotic skill learning through imitation. In Conference on Robot Learning, pages 879-893. PMLR, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.519, + 0.906, + 0.586 + ], + "angle": 0, + "content": "[43] Oier Mees, Jessica Borja-Diaz, and Wolfram Burgard. Grounding language with visual affordances over unstructured data. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 11576-11582. IEEE, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.589, + 0.906, + 0.644 + ], + "angle": 0, + "content": "[44] Zawalski Michal, Chen William, Pertsch Karl, Mees Oier, Finn Chelsea, and Levine Sergey. Robotic control via embodied chain-of-thought reasoning. arXiv preprint arXiv:2407.08693, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.647, + 0.906, + 0.715 + ], + "angle": 0, + "content": "[45] Yao Mu, Qinglong Zhang, Mengkang Hu, Wenhai Wang, Mingyu Ding, Jun Jin, Bin Wang, Jifeng Dai, Yu Qiao, and Ping Luo. Embodiedgpt: Vision-language pre-training via embodied chain of thought. Advances in Neural Information Processing Systems, 36, 2024. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.717, + 0.906, + 0.773 + ], + "angle": 0, + "content": "[46] Ashvin V Nair, Vitchyr Pong, Murtaza Dalal, Shikhar Bahl, Steven Lin, and Sergey Levine. Visual reinforcement learning with imagined goals. Advances in neural information processing systems, 31, 2018. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.775, + 0.906, + 0.87 + ], + "angle": 0, + "content": "[47] Fei Ni, Jianye Hao, Shiguang Wu, Longxin Kou, Jiashun Liu, Yan Zheng, Bin Wang, and Yuzheng Zhuang. Generate subgoal images before act: Unlocking the chain-of-thought reasoning in diffusion model for robot manipulation with multimodal prompts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.906, + 0.9 + ], + "angle": 0, + "content": "[48] Abby O'Neill, Abdul Rehman, Abhinav Gupta, Abhiram Maddukuri, Abhishek Gupta, Abhishek Padalkar, Abraham" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.906, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.946, + 0.514, + 0.957 + ], + "angle": 0, + "content": "1711" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.486, + 0.134 + ], + "angle": 0, + "content": "Lee, Acorn Pooley, Agrim Gupta, Ajay Mandlekar, et al. Open x-embodiment: Robotic learning datasets and rt-x models. arXiv preprint arXiv:2310.08864, 2023. 1, 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.137, + 0.484, + 0.219 + ], + "angle": 0, + "content": "[49] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.222, + 0.484, + 0.292 + ], + "angle": 0, + "content": "[50] Daniel Rose, Vaishnavi Himakunthala, Andy Ouyang, Ryan He, Alex Mei, Yujie Lu, Michael Saxon, Chinmay Sonar, Diba Mirza, and William Yang Wang. Visual chain of thought: bridging logical gaps with multimodal infillings. arXiv preprint arXiv:2305.02317, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.294, + 0.484, + 0.35 + ], + "angle": 0, + "content": "[51] Erick Rosete-Beas, Oier Mees, Gabriel Kalweit, Joschka Boedecker, and Wolfram Burgard. Latent plans for task-agnostic offline reinforcement learning. In Conference on Robot Learning, pages 1838–1849. PMLR, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.352, + 0.484, + 0.393 + ], + "angle": 0, + "content": "[52] V Sanh. Distilbert, a distilled version of bert: smaller, faster, cheaper and lighter. arXiv preprint arXiv:1910.01108, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.396, + 0.484, + 0.452 + ], + "angle": 0, + "content": "[53] Hao Shao, Shengju Qian, Han Xiao, Guanglu Song, Zhuofan Zong, Letian Wang, Yu Liu, and Hongsheng Li. Visual cot: Unleashing chain-of-thought reasoning in multi-modal language models. arXiv preprint arXiv:2403.16999, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.453, + 0.484, + 0.496 + ], + "angle": 0, + "content": "[54] Mohit Shridhar, Lucas Manuelli, and Dieter Fox. CIport: What and where pathways for robotic manipulation. In Conference on robot learning, pages 894-906. PMLR, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.498, + 0.484, + 0.539 + ], + "angle": 0, + "content": "[55] Mohit Shridhar, Yat Long Lo, and Stephen James. Generative image as action models. arXiv preprint arXiv:2407.07875, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.542, + 0.484, + 0.625 + ], + "angle": 0, + "content": "[56] Ishika Singh, Valts Blukis, Arsalan Mousavian, Ankit Goyal, Danfei Xu, Jonathan Tremblay, Dieter Fox, Jesse Thomason, and Animesh Garg. Progprompt: Generating situated robot task plans using large language models. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 11523-11530. IEEE, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.628, + 0.484, + 0.668 + ], + "angle": 0, + "content": "[57] Yang Song, Prafulla Dhariwal, Mark Chen, and Ilya Sutskever. Consistency models. arXiv preprint arXiv:2303.01469, 2023.8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.672, + 0.484, + 0.711 + ], + "angle": 0, + "content": "[58] Chameleon Team. Chameleon: Mixed-modal early-fusion foundation models. arXiv preprint arXiv:2405.09818, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.715, + 0.484, + 0.771 + ], + "angle": 0, + "content": "[59] Octo Model Team, Dibya Ghosh, Homer Walke, Karl Pertsch, Kevin Black, Oier Mees, Sudeep Dasari, Joel Hejna, Tobias Kreiman, Charles Xu, et al. Octo: An open-source generalist robot policy. arXiv preprint arXiv:2405.12213, 2024. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.774, + 0.484, + 0.844 + ], + "angle": 0, + "content": "[60] Homer Rich Walke, Kevin Black, Tony Z Zhao, Quan Vuong, Chongyi Zheng, Philippe Hansen-Estruch, Andre Wang He, Vivek Myers, Moo Jin Kim, Max Du, et al. Bridgedata v2: A dataset for robot learning at scale. In Conference on Robot Learning, pages 1723-1736. PMLR, 2023. 2, 5, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[61] Xinlong Wang, Xiaosong Zhang, Zhengxiong Luo, Quan Sun, Yufeng Cui, Jinsheng Wang, Fan Zhang, Yueze Wang, Zhen Li, Qiying Yu, et al. Emu3: Next-token prediction is all you need. arXiv preprint arXiv:2409.18869, 2024. 2, 8" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.908, + 0.16 + ], + "angle": 0, + "content": "[62] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.163, + 0.908, + 0.216 + ], + "angle": 0, + "content": "[63] Chuan Wen, Xingyu Lin, John So, Kai Chen, Qi Dou, Yang Gao, and Pieter Abbeel. Any-point trajectory modeling for policy learning. arXiv preprint arXiv:2401.00025, 2023. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.219, + 0.908, + 0.287 + ], + "angle": 0, + "content": "[64] Junjie Wen, Yichen Zhu, Jinming Li, Minjie Zhu, Kun Wu, Zhiyuan Xu, Ran Cheng, Chaomin Shen, Yaxin Peng, Feifei Feng, et al. Tinyvla: Towards fast, data-efficient vision-language-action models for robotic manipulation. arXiv preprint arXiv:2409.12514, 2024.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.288, + 0.908, + 0.357 + ], + "angle": 0, + "content": "[65] Chengyue Wu, Xiaokang Chen, Zhiyu Wu, Yiyang Ma, Xingchao Liu, Zizheng Pan, Wen Liu, Zhenda Xie, Xingkai Yu, Chong Ruan, et al. Janus: Decoupling visual encoding for unified multimodal understanding and generation. arXiv preprint arXiv:2410.13848, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.358, + 0.908, + 0.426 + ], + "angle": 0, + "content": "[66] Hongtao Wu, Ya Jing, Chilam Cheang, Guangzeng Chen, Jiafeng Xu, Xinghang Li, Minghuan Liu, Hang Li, and Tao Kong. Unleashing large-scale video generative pretraining for visual robot manipulation. arXiv preprint arXiv:2312.13139, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.428, + 0.908, + 0.495 + ], + "angle": 0, + "content": "[67] Yecheng Wu, Zhuoyang Zhang, Junyu Chen, Haotian Tang, Dacheng Li, Yunhao Fang, Ligeng Zhu, Enze Xie, Hongxu Yin, Li Yi, et al. Vila-u: a unified foundation model integrating visual understanding and generation. arXiv preprint arXiv:2409.04429, 2024. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.497, + 0.908, + 0.564 + ], + "angle": 0, + "content": "[68] Jiannan Xiang, Guangyi Liu, Yi Gu, Qiyue Gao, Yuting Ning, Yuheng Zha, Zeyu Feng, Tianhua Tao, Shibo Hao, Yemin Shi, et al. Pandora: Towards general world model with natural language actions and video states. arXiv preprint arXiv:2406.09455, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.567, + 0.908, + 0.635 + ], + "angle": 0, + "content": "[69] Jinheng Xie, Weijia Mao, Zechen Bai, David Junhao Zhang, Weihao Wang, Kevin Qinghong Lin, Yuchao Gu, Zhijie Chen, Zhenheng Yang, and Mike Zheng Shou. Show-o: One single transformer to unify multimodal understanding and generation. arXiv preprint arXiv:2408.12528, 2024. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.637, + 0.908, + 0.704 + ], + "angle": 0, + "content": "[70] Jonathan Yang, Catherine Glossop, Arjun Bhorkar, Dhruv Shah, Quan Vuong, Chelsea Finn, Dorsa Sadigh, and Sergey Levine. Pushing the limits of cross-embodiment learning for manipulation and navigation. arXiv preprint arXiv:2402.19432, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.706, + 0.908, + 0.761 + ], + "angle": 0, + "content": "[71] Mengjiao Yang, Yilun Du, Kamyar Ghasemipour, Jonathan Thompson, Dale Schuurmans, and Pieter Abbeel. Learning interactive real-world simulators. arXiv preprint arXiv:2310.06114, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.762, + 0.908, + 0.817 + ], + "angle": 0, + "content": "[72] Sherry Yang, Jacob Walker, Jack Parker-Holder, Yilun Du, Jake Bruce, Andre Barreto, Pieter Abbeel, and Dale Schuurmans. Video as the new language for real-world decision making. arXiv preprint arXiv:2402.17139, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.818, + 0.908, + 0.871 + ], + "angle": 0, + "content": "[73] Qihang Yu, Mark Weber, Xueqing Deng, Xiaohui Shen, Daniel Cremers, and Liang-Chieh Chen. An image is worth 32 tokens for reconstruction and generation. arxiv: 2406.07550, 2024.8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.908, + 0.901 + ], + "angle": 0, + "content": "[74] Wenhao Yu, Nimrod Gileadi, Chuyuan Fu, Sean Kirmani, Kuang-Huei Lee, Montse Gonzalez Arenas, Hao-Tien Lewis" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.908, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1712" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.484, + 0.134 + ], + "angle": 0, + "content": "Chiang, Tom Erez, Leonard Hasenclever, Jan Humplik, et al. Language to rewards for robotic skill synthesis. arXiv preprint arXiv:2306.08647, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.484, + 0.19 + ], + "angle": 0, + "content": "[75] Eric Zelikman, Yuhuai Wu, Jesse Mu, and Noah Goodman. Star: Bootstrapping reasoning with reasoning. Advances in Neural Information Processing Systems, 35:15476-15488, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.192, + 0.484, + 0.247 + ], + "angle": 0, + "content": "[76] Kaifeng Zhang, Zhao-Heng Yin, Weirui Ye, and Yang Gao. Learning manipulation skills through robot chain-of-thought with sparse failure guidance. arXiv preprint arXiv:2405.13573, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.249, + 0.484, + 0.29 + ], + "angle": 0, + "content": "[77] Tony Z Zhao, Vikash Kumar, Sergey Levine, and Chelsea Finn. Learning fine-grained bimanual manipulation with low-cost hardware. arXiv preprint arXiv:2304.13705, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.292, + 0.484, + 0.347 + ], + "angle": 0, + "content": "[78] Haoyu Zhen, Xiaowen Qiu, Peihao Chen, Jincheng Yang, Xin Yan, Yilun Du, Yining Hong, and Chuang Gan. 3d-vla: A 3d vision-language-action generative world model. arXiv preprint arXiv:2403.09631, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.349, + 0.484, + 0.416 + ], + "angle": 0, + "content": "[79] Chunting Zhou, Lili Yu, Arun Babu, Kushal Tirumala, Michihiro Yasunaga, Leonid Shamis, Jacob Kahn, Xuezhe Ma, Luke Zettlemoyer, and Omer Levy. Transfusion: Predict the next token and diffuse images with one multi-modal model. 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.418, + 0.484, + 0.487 + ], + "angle": 0, + "content": "[80] Enshen Zhou, Yiran Qin, Zhenfei Yin, Yuzhou Huang, Ruimao Zhang, Lu Sheng, Yu Qiao, and Jing Shao. Mine-dreamer: Learning to follow instructions via chain-of-imagination for simulated-world control. arXiv preprint arXiv:2403.12037, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.489, + 0.484, + 0.572 + ], + "angle": 0, + "content": "[81] Gaoyue Zhou, Victoria Dean, Mohan Kumar Srirama, Aravind Rajeswaran, Jyothish Pari, Kyle Hatch, Aryan Jain, Tianhe Yu, Pieter Abbeel, Lerrel Pinto, et al. Train offline, test online: A real robot learning benchmark. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 9197-9203. IEEE, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.574, + 0.484, + 0.643 + ], + "angle": 0, + "content": "[82] Xinghao Zhu, Ran Tian, Chenfeng Xu, Mingxiao Huo, Wei Zhan, Masayoshi Tomizuka, and Mingyu Ding. Fanuc manipulation: A dataset for learning-based manipulation with fanuc mate 200id robot. https://sites.google.com/berkeley.edu/fanuc-manipulation, 2023.1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.645, + 0.484, + 0.699 + ], + "angle": 0, + "content": "[83] Yifeng Zhu, Abhishek Joshi, Peter Stone, and Yuke Zhu. Viola: Imitation learning for vision-based manipulation with object proposal priors. In Conference on Robot Learning, pages 1199-1210. PMLR, 2023. 1" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.699 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1713" + } + ] +] \ No newline at end of file diff --git a/2025/CoT-VLA_ Visual Chain-of-Thought Reasoning for Vision-Language-Action Models/4f2ea357-dbc3-435d-b0c3-c4ec69dfb496_origin.pdf b/2025/CoT-VLA_ Visual Chain-of-Thought Reasoning for Vision-Language-Action Models/4f2ea357-dbc3-435d-b0c3-c4ec69dfb496_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..84acd956e2571510c879ac96d9879b981c4922f8 --- /dev/null +++ b/2025/CoT-VLA_ Visual Chain-of-Thought Reasoning for Vision-Language-Action Models/4f2ea357-dbc3-435d-b0c3-c4ec69dfb496_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e37837292fda758c667af6596abbaa0eecd41b4ea07483f43ad71a4ec5a9e623 +size 6873362 diff --git a/2025/CoT-VLA_ Visual Chain-of-Thought Reasoning for Vision-Language-Action Models/full.md b/2025/CoT-VLA_ Visual Chain-of-Thought Reasoning for Vision-Language-Action Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..e5e6eae6c73162d92fe7af8507505e6f87bd8905 --- /dev/null +++ b/2025/CoT-VLA_ Visual Chain-of-Thought Reasoning for Vision-Language-Action Models/full.md @@ -0,0 +1,294 @@ +# CoT-VLA: Visual Chain-of-Thought Reasoning for Vision-Language-Action Models + +Qingqing Zhao $^{1,2,*}$ Yao Lu $^{1}$ Moo Jin Kim $^{2}$ Zipeng Fu $^{2}$ Zhuoyang Zhang $^{3}$ Yecheng Wu $^{1,3}$ Zhaoshuo Li $^{1}$ Qianli Ma $^{1}$ Song Han $^{1,3}$ Chelsea Finn $^{2}$ Ankur Handa $^{1}$ Tsung-Yi Lin $^{1\dagger}$ Gordon Wetzstein $^{2\dagger}$ Ming-Yu Liu $^{1\dagger}$ Donglai Xiang $^{1\dagger}$ NVIDIA ${}^{2}$ Stanford University ${}^{3}$ MIT + +# Abstract + +Vision-language-action models (VLAs) have shown potential in leveraging pretrained vision-language models and diverse robot demonstrations for learning generalizable sensorimotor control. While this paradigm effectively utilizes large-scale data from both robotic and non-robotic sources, current VLAs primarily focus on direct input-output mappings, lacking the intermediate reasoning steps crucial for complex manipulation tasks. As a result, existing VLAs lack temporal planning or reasoning capabilities. In this paper, we introduce a method that incorporates explicit visual chain-of-thought (CoT) reasoning into vision-language-action models (VLAs) by predicting future image frames autoregressively as visual goals before generating a short action sequence to achieve these goals. We introduce CoT-VLA, a state-of-the-art 7B VLA that can understand and generate visual and action tokens. Our experimental results demonstrate that CoT-VLA achieves strong performance, outperforming the state-of-the-art VLA model by $17\%$ in real-world manipulation tasks and $6\%$ in simulation benchmarks. Videos are available at: https://cot-vla.github.io/. + +# 1. Introduction + +Recent advances in robot learning have demonstrated impressive progress in training policies that can act across diverse tasks and environments [1, 5, 12, 14, 18, 29, 36, 44, 45, 48, 54, 59, 63, 66, 70, 76, 78]. One promising direction is vision-language-action (VLA) models, which leverage the rich understanding capabilities of pretrained vision-language models (VLMs) to map natural language instructions and visual observations to robot actions [12, 29, 48]. By training VLMs on robot demonstrations, VLAs inherit their ability to understand diverse scenes, objects, and natural language in- + +![](images/bd662ec31266da1396db03846cc4a2f7312b32a1a370c3c820dd1e662ac80f6a.jpg) +Figure 1. Comparison between vanilla VLA and CoT-VLA frameworks. Prior VLA models (top) directly predict robot actions from task inputs without explicit reasoning steps and only use action-annotated robot demonstration data for training. Unlike vanilla VLAs, CoT-VLA (bottom) can also leverage action-less datasets like EPIC-KITCHEN-100 [27] to enhance subgoal image generation ability, unlocking the potential of using abundant unlabeled video data to improve VLA's visual reasoning capability. CoT-VLA first generates a subgoal image as an intermediate reasoning step, and then generate a short action sequence to achieve the subgoal. We outline the robot arm for better visualization. + +structions, leading to better generalization capabilities when fine-tuned for downstream testing scenarios. While these approaches have shown impressive results, they typically map + +directly from observations to actions without explicit intermediate reasoning steps that could improve interpretability and, potentially, performance. + +In the language domain, chain-of-thought (CoT) prompting has emerged as a powerful technique for improving the reasoning capabilities of large language models (LLMs) by encouraging step-by-step thinking [62, 75]. Applying these concepts to robotics presents exciting opportunities for grounding reasoning in text, visual observations, and physical actions. Recent works have made progress in this direction by incorporating intermediate reasoning steps like language descriptions, keypoints, or bounding boxes [15, 44, 45, 63]. These intermediate representations capture abstracted states of scenes, objects, and tasks and often require additional pre-processing pipelines. In our work, we explore subgoal images as an intermediate reasoning step before action generation. These images capture the state of the model's reasoning process and are naturally available within robot demonstration datasets. While prior work has explored subgoal generation and goal-conditioned imitation learning [2, 11, 46, 55], to the best of our knowledge, our approach is the first to integrate these concepts with VLAs as intermediate chain-of-thought reasoning steps. + +We propose visual chain-of-thought reasoning for VLAs, a new method that uses subgoal image generation as a form of chain-of-thought reasoning for robotic tasks. Rather than directly predicting actions, our method first generates a subgoal image that represents the robot's planned state in pixel space, and then conditions its action on both the current observation and the generated subgoal image. This approach allows the model to "think visually" about how to accomplish a task before acting. By using the subgoal image as intermediate reasoning step, we leverage information that already exists in robot manipulation data with minimal preprocessing required. Furthermore, since subgoal image generation does not require action annotations, this unlocks the potential of using abundant video data for improved visual reasoning and understanding. + +We build our CoT-VLA system that leverages visual chain-of-thought reasoning upon recent advances in unified multimodal foundation models that can understand and generate text and images [39, 58, 61, 67, 69]. We train our base model [67] on both the Open X-Embodiment dataset [48] and action-less video datasets [20, 27], and then fine-tune the model on task demonstrations collected on downstream robot setups used for deployment and evaluation. We design a hybrid attention mechanism for CoT-VLA: we use causal attention with next-token prediction for text and image generation, and leverage full attention to predict all action dimensions at once. Additionally, inspired by recent advances in robot learning [10, 17, 77], we predict sequences of actions (action chunking) rather than a single action at each timestep. We demonstrate that both action chunking and the hybrid + +attention mechanism improve the model's performance. + +Through extensive experiments in both simulation benchmarks [37] and real-world experiments[48, 60], we demonstrate that our visual chain-of-thought reasoning helps improve policy performance compared to prior VLA approaches. Our key contributions include: + +- We introduce a method of visual chain-of-thought reasoning through subgoal image generation as an intermediate reasoning step for robotic control. +- We introduce a system CoT-VLA that incorporates visual chain-of-thought reasoning, and a hybrid attention mechanism that combines causal attention for pixel and text generation and full attention for action prediction. +- We conduct comprehensive evaluations in both simulation and the real world, demonstrating that visual chain-of-thought reasoning improves VLA performance, and our system achieves state-of-the-art performance across multiple robot platforms and tasks. + +# 2. Related Work + +Chain-of-Thought (CoT) Reasoning CoT reasoning has gained prominence in natural language processing, particularly for enabling models to perform complex, multi-step reasoning tasks by breaking down problem-solving into sequential, explainable steps. Early work on CoT reasoning [62] has demonstrated the effectiveness of prompting large language models to generate intermediate reasoning steps before arriving at a final answer. Extending this paradigm to the visual domain, researchers have explored multimodal chain-of-thought methods, where visual information is processed iteratively in a stepwise fashion to reason about future outcomes or states, including generating bounding boxes [53], intermediate image infillments using Stable Diffusion [50] or standard Python packages [24], or generating CLIP embeddings [22]. Recently, CoT reasoning has been explored in embodied applications. It can generate textual plans for multi-stage execution [44, 45], point trajectories [63], label bounding boxes of objects and gripper positions as additional observations [44], generate future image trajectories for open-loop following [35, 47], and generate fine-grained reward guidance for reinforcement learning [76]. In this work, we introduce Visual-CoT reasoning for robotic manipulation, where predicted subgoal images serve as intermediate reasoning steps for closed-loop action generation. This approach leverages demonstration videos as natural intermediate reasoning states without requiring additional annotations. + +Vision-Language-Action Models Large pretrained vision-language models (VLMs) [9, 28, 38] have emerged as a powerful tool for robot learning, and recent works have explored various approaches to integrate them into robot systems. Several works utilize VLMs as intermediate + +![](images/50788cdfa18198255180faf8692616ccb541f830221d05133caaf4dace9548cd.jpg) +Figure 2. Overview of CoT-VLA framework. We build our model on VILA-U [67], a generative multimodal model pretrained on interleaved text-image data. The base model then trains on robot demonstrations [48] and action-less videos [20, 27]. During deployment, given a visual observation and a text instruction, the model performs visual chain-of-thought reasoning by generating a subgoal image (upper blue) with causal attention. It then generates a short action sequence with full attention $(\mathbf{a}_1 \cdots \mathbf{a}_n)$ for robot execution. The system operates in a closed-loop control manner by capturing new observations after executing predicted action sequences. + +components for perception and control, leveraging their strong semantic understanding and reasoning capabilities to decompose complex tasks [21, 26, 34, 56], detect objects [19, 25], or generate dense rewards [13, 40, 74] or goals [2, 11, 15, 46, 47, 55, 71, 80]. Some approaches incorporate VLMs [4, 49, 64] into end-to-end trainable policies by using them as pretrained backbone for better visuo-language representation [12, 29, 41, 48, 59]. Most relevant to our work are recent approaches that fine-tune pretrained VLMs on robot demonstration data for direct action prediction [12, 29, 48]. These VLAs demonstrate improved generalization to novel objects, environments, and natural language instructions through pretraining on internet-scale vision-language datasets, providing a promising direction for transferring visual and language knowledge to robotic control tasks. However, most existing VLAs do not leverage the step-by-step reasoning capabilities demonstrated in large language models, which have been shown to significantly improve performance across various tasks [62]. In the past, researchers have used chain-of-thought reasoning on language instructions or intermediate keypoints/bounding boxes for robotics [6, 13, 45, 63]. We introduce visual chain-of-thought reasoning to the VLA frameworks, using subgoal images as intermediate reasoning steps before action generation. + +# 3. CoT-VLA + +In this section, we present our visual chain-of-thought reasoning framework for VLAs. We begin with the formulation of our method (3.1), followed by a detailed description of the system architecture (3.2). We then explain our training procedures (3.3) and outline the deployment strategy for downstream tasks (3.3). + +# 3.1. Visual Chain-of-Thought Reasoning + +We consider two types of training data for VLA pretraining: robot demonstrations dataset $D_r$ and action-less videos dataset $D_v$ . Robot demonstrations are represented as $D_r = \{(l, \mathbf{a}_{1\dots T}, \mathbf{s}_{1\dots T})\}$ , where $l$ denotes the natural language instruction, $\mathbf{a}_{1\dots T} = \{\mathbf{a}_1, \dots, \mathbf{a}_T\}$ denotes the sequence of robot actions, and $\mathbf{s}_{1\dots T} = \{\mathbf{s}_1, \dots, \mathbf{s}_T\}$ denotes the visual observations as a sequence of images. Action-less videos $D_v = \{(l, \mathbf{s}_{1\dots T})\}$ consist of language descriptions and images without action annotations. + +VLA: Vanilla VLA approaches fine-tune a pretrained VLM, $P_{\theta}$ , on $D_{r}$ , learning to predict actions $\hat{\mathbf{a}}_{t+1}$ directly from the current observation $\mathbf{s}_{t}$ and language instruction $l$ (Figure 1, top): + +$$ +\hat {\mathbf {a}} _ {t} \sim P _ {\theta} (\mathbf {a} _ {t} | \mathbf {s} _ {t}, l) \tag {1} +$$ + +CoT-VLA: Our key insight is to incorporate explicit visual reasoning before action generation. As illustrated in Figure 2, our approach operates in two sequential phases: + +$$ +\hat {\mathbf {s}} _ {t + n} \sim P _ {\theta} \left(\mathbf {s} _ {t + n} \mid \mathbf {s} _ {t}, l\right) \tag {2} +$$ + +$$ +\{\hat {\mathbf {a}} _ {t}, \dots , \hat {\mathbf {a}} _ {t + m} \} \sim P _ {\theta} (\{\mathbf {a} _ {t}, \dots , \mathbf {a} _ {t + m} | \mathbf {s} _ {t}, l, \mathbf {s} _ {t + n}) \tag {3} +$$ + +where we first predict a subgoal image $\hat{\mathbf{s}}_{t + n}$ , $n$ frames ahead, as an intermediate visual reasoning step (Equation 2). Then we generate a sequence of $m$ actions to achieve this subgoal state (Equation 3). This enables the model to "think visually" first by explicitly reasoning about desired future states before predicting the actions. The visual reasoning step, Equation (2), is trained on both robot demonstrations $D_r$ and actionless videos $D_v$ , and the action generation step, Equation (3), is trained on robot demonstrations $D_r$ only. + +![](images/5db19d53593396530375d73b8f75a3bbac011a98bf56c372acb801dadf706ffa.jpg) +Figure 3. Hybrid attention mechanism in CoT-VLA. We use causal attention for image or text generation and full attention for action generation. $[x]$ , $[\theta]$ and $[g]$ are special tokens for parallel decoding of actions. + +# 3.2. The Base Vision-Language Model + +To enable the visual reasoning capabilities described in Equation (2), we build upon VILA-U [67], an unified multimodal foundation model capable of both understanding and generating image and text tokens. + +VILA-U unifies video, image, and language understanding through an autoregressive next-token prediction framework. At its core is a unified vision tower that encodes visual inputs as discrete tokens aligned with textual information. This enables autoregressive image and video generation while significantly enhancing the understanding capabilities of VLMs that leverage discrete visual features. VILA-U utilizes residual quantization [32] to improve the representational capacity of discrete visual features - incorporating a depth transformer, as introduced in RQ-VAE [32], to gradually predict the residual tokens. The extracted visual features are then passed through a projector before being processed by the LLM backbone. The base model is trained on multimodal pairs including [image, text], [text, image], [video, text], and [text, video]. We use the VILA-U model trained on $256 \times 256$ resolution images, where each image is encoded into $16 \times 16 \times 4$ tokens with a residual depth of 4 [32]. For detailed information about VILA-U training and architecture, we refer readers to [67]. + +# 3.3. Training Procedures + +We pretrain the base 7B VILA-U model on a combination of robot demonstrations $D_r$ and action-less videos $D_v$ . During training, we optimize three components, the LLM backbone, projector, and depth transformer, while keeping the vision tower fixed. Our training objective has two key components: subgoal image generation with causal attention (2) and action generation with full attention (3). + +Visual Tokens Prediction For subgoal image generation, each training sequence is of form $(l, \mathbf{s}_t, \mathbf{s}_{t + n})$ . We follow the training objective used in [67]. At each visual position $j$ , the depth transformer, $P_{\delta}$ , autoregressively predicts $D$ residual tokens $(k_{j1}, \dots, k_{jD})$ based on the LLM-generated code embedding $h_j$ . The training objective for visual tokens is then formulated as: + +$$ +\mathcal {L} _ {\text {v i s u a l}} = - \sum_ {j} \sum_ {d = 1} ^ {D} \log P _ {\delta} \left(k _ {j d} \mid k _ {j, < d}\right) \tag {4} +$$ + +where $j$ indexes the positions containing visual tokens. For a more detailed explanation of this loss function, we refer readers to [67], Section 3.2. + +Action Tokens Prediction For action prediction, each training sequence takes the form $(l, \mathbf{s}_t, \mathbf{s}_{t+n}, \mathbf{a}_t, \dots, \mathbf{a}_{t+m})$ . Each action $\mathbf{a}_i$ is represented by 7 tokens, with each action dimension independently discretized. Following [29], we map each continuous action dimension into 256 discrete bins, with bin widths determined by uniformly dividing the interval between the 1st and 99th percentiles of the training data's action distribution. We repurpose the 256 least frequently used tokens in the text tokenizer's vocabulary as action bin tokens. Unlike prior works [12, 29, 48], we employ full attention for processing and predicting action tokens, enabling all action tokens to interact with each other. This attention mechanism is illustrated in Figure 3. During training, we minimize the cross-entropy loss for action predictions: + +$$ +\mathcal {L} _ {\text {a c t i o n}} = - \sum_ {i = 1} ^ {m} \log P _ {\theta} \left(\mathbf {a} _ {t} \dots \mathbf {a} _ {t + m} \mid l, s _ {t}, s _ {t + n}\right) \tag {5} +$$ + +Given a batch of input sequences, The overall training objective combines the action and visual losses: + +$$ +\mathcal {L} = \mathcal {L} _ {\text {a c t i o n}} + \mathcal {L} _ {\text {v i s u a l}} \tag {6} +$$ + +Pretraining Phase We pretrain CoT-VLA on both robot demonstrations $D_r$ and action-less videos $D_v$ as described in Section 3.1. For robot demonstrations, we curate a subset of the Open X-Embodiment dataset [48] (OpenX). Following the pre-processing pipeline established in OpenVLA [29], we select and process datasets with third-person camera views and single-arm end-effector control (7-DoF). For action-less videos $D_v$ , we incorporate the EPIC-KITCHENS [27] and Something-Something V2 [20] datasets. All images are processed at $256 \times 256$ resolution. For visual reasoning, we use subgoal images at future timestep $n$ uniformly sampled from a dataset-specific range $[n_l, n_u]$ , where $n_l$ and $n_u$ define the lower and upper bounds of the prediction horizon. We use an action chunk size of 10. For complete dataset specifications and training hyperparameters, please refer to the supplementary material. + +
Average (↑)Spatial (↑)Object (↑)Goal (↑)Long (↑)
Diffusion Policy72.4 ± 0.7%78.3 ± 1.1%92.5 ± 0.7%68.3 ± 1.2%50.5 ± 1.3%
Octo fine-tuned75.1 ± 0.6%78.9 ± 1.0%85.7 ± 0.9%84.6 ± 0.9%51.1 ± 1.3%
OpenVLA fine-tuned76.5 ± 0.6%84.7 ± 0.9%88.4 ± 0.8%79.2 ± 1.0%53.7 ± 1.3%
CoT-VLA-7B (ours)81.13 ± 0.6 %87.5 ± 1.4%91.6 ± 0.5%87.6 ± 0.6%69.0 ± 0.8%
+ +Table 1. LIBERO benchmark experimental results. For each task suite (Spatial, Object, Goal, Long), we report the average success rate and standard error across 3 seeds with 500 episodes each. CoT-VLA achieves the best or competitive performance across all LIBERO benchmarks suites compared to baseline approaches. The bolded entries correspond to highest success rates while underlined entries correspond to second-highest. + +Adaptation Phase for Downstream Closed-Loop Deployment For adaptation to downstream tasks, we fine-tune our pretrained model using task-specific robot demonstration data $D_{r}$ collected on the target robot setups. During this phase, we optimize the LLM backbone, projector, and depth transformer while keeping the vision tower frozen, maintaining the same training setup as the pretraining stage. The resulting model can execute new manipulation tasks based on natural language commands $l$ . Algorithm 1 describes our robot control procedure at test time. + +Algorithm 1 CoT-VLA test-time closed-loop control +Require: CoT-VLA Model $P_{\theta}$ , initial state $s_0^{\mathrm{obs}}$ , language instruction $l$ +0: $t \gets 0$ +0: while True do +0: sample $\hat{\mathbf{s}}_{t+n} \sim P_{\theta}(\mathbf{s}_{t+n} \mid l, \mathbf{s}_t^{\mathrm{obs}})$ +0: sample $[\hat{\mathbf{a}}_t, \dots, \hat{\mathbf{a}}_{t+m}] \sim P_{\theta}(\mathbf{a}_t, \dots, \mathbf{a}_{t+m} \mid l, \mathbf{s}_t^{\mathrm{obs}}, \mathbf{s}_{t+n})$ +0: for $j = 0$ to $m$ do +0: execute $\hat{\mathbf{a}}_{t+j}$ +0: end for +0: $t \gets t + m + 1$ +0: $\mathbf{s}_t^{\mathrm{obs}} \gets$ robot observation + +# 4. Experiments + +We evaluate the effectiveness of our approach and our system through a set of experiments spanning both simulation benchmarks and real-world robot manipulation tasks. Our experiments aim to addresses following questions: + +- How does our system perform compared to state-of-the-art baselines across multiple benchmarks and embodiments? (Section 4.2) +- What is the impact of pretraining, visual chain-of-thought reasoning and hybrid attention on task performance? (Section 4.3) +- To what extent does improved generalization in visual reasoning enhance the action prediction capabilities? (Section 4.4) + +# 4.1. Experimental Setup + +We conduct evaluations across three complementary settings: the LIBERO benchmark [37] for evaluation in simulation environments, the Bridge-V2 platform [60] with its dataset of 45k robot demonstrations, and the Franka-Tabletop setup with a stationary, table-mounted Franka Emika Panda robot with limited 10 to 150 robot demonstrations for each testing scenario. + +LIBERO Simulation Benchmark We perform evaluation on LIBERO [37], a simulation benchmark comprising four distinct task suites: LIBERO-Spatial, LIBERO-Object, LIBERO-Goal, and LIBERO-Long. Each suite contains 10 diverse tasks with 50 human-teleoperated demonstrations per task, aiming to evaluate the robot's comprehension of spatial relationships, object interactions, and task-specific objectives. We follow the same preprocessed pipeline as in [29]: (1) removing pause intervals from trajectories, (2) standardizing image resolution to $256 \times 256$ pixels, and (3) applying a 180-degree rotation to all images. + +Bridge-V2 Real-Robot Experiments We use a 6-DoF WidowX robotic arm, following the experimental setup from Bridge-V2 [60]. Our training data has $45\mathrm{k}$ language-annotated trajectories from the Bridge-V2 dataset, encompassing diverse manipulation tasks. While the dataset was incorporated into the pretraining phase alongside OpenX, we performed additional task-specific fine-tuning exclusively on Bridge-V2 until achieving a training action prediction accuracy threshold of $95\%$ . Following [29], we evaluate on four tasks designed in [29] to evaluate visual robustness (varying distractors), motion generalization (novel object positions), semantic generalization (unseen language instruction), and language grounding (instruction following). + +Franka-Tabletop Real-Robot Experiments We use a stationary, table-mounted Franka Emika Panda 7-DoF robot arm denoted as Franka-Tabletop. The setup is not seen during the pretraining stage and is designed to assess our model's adaptation capability to novel real-world environments with small amounts of robot demonstrations. We + +![](images/26969250864415a73f4cc4a6fc6cd168154b53ece3d178986990c9cda66a259c.jpg) +Figure 4. Franka-Tabletop comparisons. Evaluation across six distinct manipulation tasks, with separate models trained per task. Left: Representative initial states for each task setup. Right: Task-specific success rates and cross-task averages for our method and baselines. CoT-VLA achieves best average performance and demonstrates strong capabilities in both single-instruction and multi-instruction scenarios. + +perform evaluations across 6 tasks: 3 narrow domain single-instruction tasks for and 3 diverse multi-instruction tasks outlined in Figure 4 and introduced in [29]. For each task, the dataset contains between 10 and 150 demonstrations. + +Baselines We evaluate our approach against four state-of-the-art baselines. Diffusion Policy [10], a state-of-the-art imitation learning algorithm, is trained from scratch for each test scenario in LIBERO and Franka-Tabletop. The implementation incorporates action chunking and proprioception while conditioning on DistilBERT [52] language embeddings. OpenVLA [29] is an open-source VLA model that fine-tunes pretrained vision-language models on the OpenX dataset; and Octo [59] is a generalist model pretrained on OpenX without VLM initialization. For both OpenVLA and Octo, we use their published checkpoints for Bridge-V2 evaluations and fine-tune them for our LIBERO and Franka-Tabletop experiments. SUSIE [2], a two-stage approach, combines instruction-guided image editing for goal generation with a goal-conditioned policy for action generation. We evaluate SUSIE using their published checkpoint on Bridge-V2. + +# 4.2. Evaluations Results + +LIBERO We present quantitative results in Table 1, where each method is evaluated over 500 trials per task suite, with 3 random seeds. Success rates are reported with means and standard error. Qualitative examples of our method's reasoning and execution trajectories are illustrated in Figure 5. Results demonstrate that CoT-VLA effectively adapts to tasks in the LIBERO simulation environment, achieving best or competitive performance compared to baseline approaches. By analyzing rollout videos of failure cases, we found that baseline methods occasionally overfit to visual cues while disregarding language instructions. Specifically, when initial states appear visually similar across different tasks (e.g., in LIBERO-Spatial), baseline methods execute a different task compared to the commanded task in some episodes. CoT-VLA exhibits better instruction following + +ability by first reasoning visually about the desired actions via language-grounded subgoal generation, and then predicting the relevant actions for achieving the goal. + +Bridge-V2 We evaluate CoT-VLA and baselines on the Bridge-V2 benchmark across four generalization categories identified in [29]: visual generalization ("put eggplant into pot" with cluttered environments), motion generalization ("put carrot on plate" with height variations), semantic generalization ("take purple grapes out of pot"), and language grounding ("put eggplant or red bottle into pot"). We report the quantitative results in Table 2, where each task is tested with 10 trials. SUSIE [2] generates visually higher-quality goal images through its diffusion prior (see Section 5 for a detailed discussion on our limitations) but achieves lower success rates on tasks involving novel objects or requiring complex language grounding. Compared to OpenVLA [29], CoT-VLA shows slightly lower success rates in visual and language generalization tasks due to grasping failures from action chunking (see Section 5) rather than errors in visual reasoning. However, CoT-VLA demonstrates competitive performance across all four generalization categories overall, achieving comparable or better results to baseline approaches. + +Franka-Tabletop We present quantitative results in Table 4 and example execution trajectories in Figure 5. In this experiment, models are fine-tuned on a relatively small set of demonstrations. While Diffusion Policy achieves top performance on single-instruction tasks (e.g., "put corn in bowl"), its performance degrades on multi-instruction tasks involving diverse objects and complex language instructions. Models pretrained on the OpenX dataset - Octo, OpenVLA, and CoT-VLA - demonstrate better adaptation and performance on multi-instruction tasks where language grounding is critical. Overall, CoT-VLA achieves the highest average performance compared to baseline approaches, showing improvements in both single and multi-instruction scenarios. + +![](images/ddecb96de6c4fe2fe9935672eab756d2ca0a6be1e313cef3f5f9502ad7b74dcd.jpg) +Figure 5. Task execution examples for LIBERO, Bridge-V2, and Franka-Tabletop using CoT-VLA. For each task: Left: text instruction $(l)$ and initial state $(s_0^{\mathrm{obs}})$ . Middle: generated intermediate goal states $(\hat{s}_t)$ demonstrating visual chain-of-thought reasoning, where each goal image is conditioned on both the instruction and the most recent observation. Right: final state $(s_T^{\mathrm{obs}})$ upon task completion. Complete execution trajectories are available in the supplementary video. + +# 4.3. Ablation Study + +Visual CoT, Hybrid Attention, and Action Chunking We conduct comprehensive ablation studies on two LIBERO benchmark suites: LIBERO-Spatial and LIBERO-Goal. We evaluate four model variants: VLA - a baseline implementation following the standard VLA framework [29], with the same VILA-U backbone but without chain-of-thought reasoning and action chunking; + action chunking - extending the vanilla VLA to predict action sequences of length $m$ ; + hybrid attention - further adding full attention mechanisms for action sequence prediction, as illustrated in Figure 3; and + CoT (ours): our complete approach with hybrid attention mechanism and visual chain-of-thought reasoning. + +As shown in Figure 6, both benchmark suites demonstrate that action sequence prediction consistently outperforms single-action prediction. The addition of hybrid attention mechanisms further improves performance. Our CoT-VLA achieves the best results validating the effectiveness of visual chain-of-thought reasoning for VLA tasks. + +
CategorySUSIEOctoOpenVLACoT-VLA
Visual30%35%75%65%
Motion10%10%45%60%
Semantic20%0%40%50%
Language40%40%75%70%
+ +Table 2. Bridge-V2 Comparison. Success rates across four generalization categories, with 10 trials per category and partial credit scoring following [29]. Visual: "put eggplant into pot" with cluttered environments; Motion: "put carrot on plate" with height variations; Semantic: "take purple grapes out of pot"; Language: "put eggplant or red bottle into pot". + +**Pretraining** Our training pipeline has two stages, pretraining VILA-U on the OpenX dataset augmented with actionless video data (Section 3.3), and task-specific post-training on robot demonstration data. To assess the importance of our pretraining stage, we conduct ablation studies on the Franka-Tabletop setup. We report the quantitative results + +![](images/6ea0ad3b3ccb98fdac323ac664b2e908e72db861c0add6f21cf035b150c7659f.jpg) +a) +VLA + action chunking + +![](images/8b7eb700dc94010f650f5243af400c3b1badf4c194800db729df8fcc393cb9bc.jpg) +b) +Figure 6. Ablation studies of CoT-VLA components. a) Results on LIBERO-Spatial and LIBERO-Goal benchmarks demonstrate the effectiveness of three components: action chunking, hybrid attention, and visual chain-of-thought reasoning. b) Pretraining ablation experiments on Franka-Tabletop show performance improvements from the OpenX and action-less video pretraining process. + +in Figure 6. Our results show that CoT-VLA with our pretraining stage achieves a $46.7\%$ relative improvement, from $53.7\%$ to $78.8\%$ , compared to directly fine-tuning the base VILA-U model on Franka-Tabletop demonstrations, demonstrating better downstream task adaptation. + +# 4.4. Better Visual Reasoning Helps + +Unlike prior VLAs that only use robot demonstration data $D_{\mathrm{r}}$ during training, CoT-VLA also leverages action-less video data $D_{\mathrm{v}}$ for pretraining through its intermediate visual chain-of-thought reasoning steps. This enables learning of both dynamics and instruction following from captioned videos alone, which are substantially more abundant than robot demonstrations. To investigate how visual reasoning capabilities transfer to robot performance, we conduct an ablation study on the Franka-Tabletop setup using novel, long-horizon tasks that combine two unseen subtasks. We design two tasks – (1) “move the green scallion to the apple-covered book” and (2) “move the green cauliflower to the bear-covered book” – that are challenging for our model’s out-of-distribution generalization. For each task, we collect one demonstration trajectory to obtain ground-truth goal images. We evaluate each task across 5 trials under two conditions: (1) CoT-VLA using its generated goal images and (2) CoT-VLA using ground-truth goal images from the collected demonstrations. As shown in Table 3, using ground-truth goal images improves the absolute success rate by $40\%$ for both tasks. This performance boost suggests that advances in visual reasoning and goal image generation could directly translate to better robotic task performance. While our method still struggles with out-of-distribution subgoal generation, recent advances in large-scale video and image + +models show promising directions for improving visual reasoning capabilities with scaling. + +
Sub-task 1Sub-task 2
Generated Goal Images20%0%
Ground-truth Goal Images60%40%
+ +Table 3. Better visual reasoning helps. Success rates comparing CoT-VLA using generated versus ground-truth goal images on out-of-distribution tasks. Results demonstrate that improved visual reasoning (simulated by ground-truth goals) leads to better task performance, suggesting that advances in goal generation can translate to improved action execution. + +# 5. Conclusion, Limitations and Future Work + +In this work, we introduce CoT-VLA, bridging vision-language-action models with chain-of-thought reasoning by introducing intermediate visual goals as explicit reasoning steps. Rather than using abstract representations like bounding boxes or keypoints, we propose using subgoal images sampled from videos as an interpretable and effective intermediate representation. We build our system upon VILA-U, demonstrating strong performance across diverse robotic manipulation tasks. + +While our approach demonstrates effectiveness, there are certain limitations. First, generating intermediate image tokens during inference introduces significant computational overhead compared to direct action generation approaches. Our method requires generating 256 image tokens before action tokens, leading to a $7 \times$ slowdown on average with an action chunk size of 10. While action chunking and parallel decoding improve inference speed, image generation remains the primary bottleneck. Recent advancement in fast image generation or fast LLM inference techniques could potentially improve the throughput of the model [7, 31, 33, 57, 73] and be integrated into our system. Second, our autoregressive image generation produces lower visual quality compared to state-of-the-art diffusion-based models. Recent advances in unified multimodal models [61, 65, 69, 79] suggest promising directions for improvements. Additionally, while effective, our action chunking approach can introduce discontinuous actions between chunks and lacks high-frequency feedback during execution. These limitations could be addressed through temporal smoothing techniques and per-step prediction approaches similar to those proposed in [10]. Finally, while CoT-VLA leverages action-less video data during pretraining, current computational constraints limit its ability to achieve visual-reasoning generalization for entirely new tasks. Looking forward, we believe recent advances in video/image generation and world models [15, 23, 30, 68, 72] present promising opportunities to enhance generalization capabilities through improved visual reasoning and predictive modeling. + +# References + +[1] Homanga Bharadhwaj, Debidatta Dwibedi, Abhinav Gupta, Shubham Tulsiani, Carl Doersch, Ted Xiao, Dhruv Shah, Fei Xia, Dorsa Sadigh, and Sean Kirmani. Gen2act: Human video generation in novel scenarios enables generalizable robot manipulation. arXiv preprint arXiv:2409.16283, 2024. 1 +[2] Kevin Black, Mitsuhiko Nakamoto, Pranav Atreya, Homer Walke, Chelsea Finn, Aviral Kumar, and Sergey Levine. Zero-shot robotic manipulation with pretrained image-editing diffusion models. arXiv preprint arXiv:2310.10639, 2023. 2, 3, 6 +[3] Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Joseph Dabis, Chelsea Finn, Keerthana Gopalakrishnan, Karol Hausman, Alex Herzog, Jasmine Hsu, et al. Rt-1: Robotics transformer for real-world control at scale. arXiv preprint arXiv:2212.06817, 2022. 1 +[4] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 3 +[5] Chi-Lam Cheang, Guangzeng Chen, Ya Jing, Tao Kong, Hang Li, Yifeng Li, Yuxiao Liu, Hongtao Wu, Jiafeng Xu, Yichu Yang, et al. Gr-2: A generative video-language-action model with web-scale knowledge for robot manipulation. arXiv preprint arXiv:2410.06158, 2024. 1 +[6] Boyuan Chen, Zhuo Xu, Sean Kirmani, Brain Ichter, Dorsa Sadigh, Leonidas Guibas, and Fei Xia. Spatialvlm: Endowing vision-language models with spatial reasoning capabilities. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14455-14465, 2024. 3 +[7] Junyu Chen, Han Cai, Junsong Chen, Enze Xie, Shang Yang, Haotian Tang, Muyang Li, Yao Lu, and Song Han. Deep compression autoencoder for efficient high-resolution diffusion models. arXiv preprint arXiv:2410.10733, 2024. 8 +[8] Lawrence Yunliang Chen, Simeon Adebola, and Ken Goldberg. Berkeley UR5 demonstration dataset. https:// sites.google.com/view/berkeley-ur5/home. 1 +[9] Xi Chen, Xiao Wang, Lucas Beyer, Alexander Kolesnikov, Jialin Wu, Paul Voigtlaender, Basil Mustafa, Sebastian Goodman, Ibrahim Alabdulmohsin, Piotr Padlewski, et al. Pali-3 vision language models: Smaller, faster, stronger. arXiv preprint arXiv:2310.09199, 2023. 2 +[10] Cheng Chi, Zhenjia Xu, Siyuan Feng, Eric Cousineau, Yilun Du, Benjamin Burchfiel, Russ Tedrake, and Shuran Song. Diffusion policy: Visuomotor policy learning via action diffusion. The International Journal of Robotics Research, page 02783649241273668, 2023. 2, 6, 8 +[11] Yiming Ding, Carlos Florensa, Pieter Abbeel, and Mariano Pielipp. Goal-conditioned imitation learning. Advances in neural information processing systems, 32, 2019. 2, 3 +[12] Danny Driess, Fei Xia, Mehdi SM Sajjadi, Corey Lynch, Aakanksha Chowdhery, Brian Ichter, Ayzaan Wahid, Jonathan Thompson, Quan Vuong, Tianhe Yu, et al. Palm-e: An + +embodied multimodal language model. arXiv preprint arXiv:2303.03378, 2023.1,3,4 +[13] Yuqing Du, Ksenia Konyushkova, Misha Denil, Akhil Raju, Jessica Landon, Felix Hill, Nando de Freitas, and Serkan Cabi. Vision-language models as success detectors. arXiv preprint arXiv:2303.07280, 2023. 3 +[14] Yilun Du, Mengjiao Yang, Pete Florence, Fei Xia, Ayzaan Wahid, Brian Ichter, Pierre Sermanet, Tianhe Yu, Pieter Abbeel, Joshua B Tenenbaum, et al. Video language planning. arXiv preprint arXiv:2310.10625, 2023. 1 +[15] Yilun Du, Sherry Yang, Bo Dai, Hanjun Dai, Ofir Nachum, Josh Tenenbaum, Dale Schuurmans, and Pieter Abbeel. Learning universal policies via text-guided video generation. Advances in Neural Information Processing Systems, 36, 2024. 2, 3, 8 +[16] Frederik Ebert, Yanlai Yang, Karl Schmeckpeper, Bernadette Bucher, Georgios Georgakis, Kostas Daniilidis, Chelsea Finn, and Sergey Levine. Bridge data: Boosting generalization of robotic skills with cross-domain datasets. arXiv preprint arXiv:2109.13396, 2021. 1 +[17] Zipeng Fu, Qingqing Zhao, Qi Wu, Gordon Wetzstein, and Chelsea Finn. Humanplus: Humanoid shadowing and imitation from humans. arXiv preprint arXiv:2406.10454, 2024. 2 +[18] Zipeng Fu, Tony Z. Zhao, and Chelsea Finn. Mobile aloha: Learning bimanual mobile manipulation with low-cost whole-body teleoperation. In Conference on Robot Learning (CoRL), 2024. 1 +[19] Samir Yitzhak Gadre, Mitchell Wortsman, Gabriel Ilharco, Ludwig Schmidt, and Shuran Song. Cows on pasture: Baselines and benchmarks for language-driven zero-shot object navigation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23171-23181, 2023. 3 +[20] Raghav Goyal, Samira Ebrahimi Kahou, Vincent Michalski, Joanna Materzynska, Susanne Westphal, Heuna Kim, Valentin Haenel, Ingo Fruend, Peter Yianilos, Moritz Mueller-Freitag, et al. The" something something" video database for learning and evaluating visual common sense. In Proceedings of the IEEE international conference on computer vision, pages 5842-5850, 2017. 2, 3, 4, 1 +[21] Huy Ha, Pete Florence, and Shuran Song. Scaling up and distilling down: Language-guided robot skill acquisition. In Conference on Robot Learning, pages 3766-3777. PMLR, 2023. 3 +[22] William Harvey and Frank Wood. Visual chain-of-thought diffusion models. arXiv preprint arXiv:2303.16187, 2023. 2 +[23] Anthony Hu, Lloyd Russell, Hudson Yeo, Zak Murez, George Fedoseev, Alex Kendall, Jamie Shotton, and Gianluca Corrado. Gaia-1: A generative world model for autonomous driving. arXiv preprint arXiv:2309.17080, 2023. 8 +[24] Yushi Hu, Weijia Shi, Xingyu Fu, Dan Roth, Mari Ostendorf, Luke Zettlemoyer, Noah A Smith, and Ranjay Krishna. Visual sketchpad: Sketching as a visual chain of thought for multimodal language models. arXiv preprint arXiv:2406.09403, 2024. 2 +[25] Chenguang Huang, Oier Mees, Andy Zeng, and Wolfram Burgard. Visual language maps for robot navigation. In 2023 + +IEEE International Conference on Robotics and Automation (ICRA), pages 10608-10615. IEEE, 2023. 3 +[26] Wenlong Huang, Chen Wang, Yunzhu Li, Ruohan Zhang, and Li Fei-Fei. Rekep: Spatio-temporal reasoning of relational keypoint constraints for robotic manipulation. arXiv preprint arXiv:2409.01652, 2024. 3 +[27] Georgios Kapidis, Ronald Poppe, Elsbeth Van Dam, Lucas Noldus, and Remco Veltkamp. Egocentric hand track and object-based human action recognition. In 2019 IEEE SmartWorld, Ubiquitous Intelligence & Computing, Advanced & Trusted Computing, Scalable Computing & Communications, Cloud & Big Data Computing, Internet of People and Smart City Innovation (SmartWorld/SCALCOM/UIC/ATC/CBDCom/IOP/SCI), pages 922-929. IEEE, 2019. 1, 2, 3, 4 +[28] Siddharth Karamcheti, Suraj Nair, Ashwin Balakrishna, Percy Liang, Thomas Kollar, and Dorsa Sadigh. Prismatic vlms: Investigating the design space of visually-conditioned language models. arXiv preprint arXiv:2402.07865, 2024. 2 +[29] Moo Jin Kim, Karl Pertsch, Siddharth Karamcheti, Ted Xiao, Ashwin Balakrishna, Suraj Nair, Rafael Rafailov, Ethan Foster, Grace Lam, Pannag Sanketi, et al. Openvla: An open-source vision-language-action model. arXiv preprint arXiv:2406.09246, 2024. 1, 3, 4, 5, 6, 7 +[30] Dan Kondratyuk, Lijun Yu, Xiuye Gu, José Lezama, Jonathan Huang, Grant Schindler, Rachel Hornung, Vighnesh Birodkar, Jimmy Yan, Ming-Chang Chiu, et al. Videopoet: A large language model for zero-shot video generation. arXiv preprint arXiv:2312.14125, 2023. 8 +[31] Woosuk Kwon, Zhuohan Li, Siyuan Zhuang, Ying Sheng, Lianmin Zheng, Cody Hao Yu, Joseph E. Gonzalez, Hao Zhang, and Ion Stoica. Efficient memory management for large language model serving with pagedattention. In Proceedings of the ACM SIGOPS 29th Symposium on Operating Systems Principles, 2023. 8 +[32] Doyup Lee, Chiheon Kim, Saehoon Kim, Minsu Cho, and Wook-Shin Han. Autoregressive image generation using residual quantization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11523-11532, 2022. 4 +[33] Yaniv Leviathan, Matan Kalman, and Yossi Matias. Fast inference from transformers via speculative decoding. In International Conference on Machine Learning, pages 19274-19286. PMLR, 2023. 8 +[34] Boyi Li, Yue Wang, Jiageng Mao, Boris Ivanovic, Sushant Veer, Karen Leung, and Marco Pavone. Driving everywhere with large language model policy adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14948-14957, 2024. 3 +[35] Junbang Liang, Ruoshi Liu, Ege Ozguroglu, Sruthi Sudhakar, Achal Dave, Pavel Tokmakov, Shuran Song, and Carl Vondrick. Dreamitate: Real-world visuomotor policy learning via video generation. arXiv preprint arXiv:2406.16862, 2024. 2 +[36] Fanqi Lin, Yingdong Hu, Pingyue Sheng, Chuan Wen, Jiacheng You, and Yang Gao. Data scaling laws in imitation learning for robotic manipulation. arXiv preprint arXiv:2410.18647, 2024. 1 + +[37] Bo Liu, Yifeng Zhu, Chongkai Gao, Yihao Feng, Qiang Liu, Yuke Zhu, and Peter Stone. Libero: Benchmarking knowledge transfer for lifelong robot learning. arXiv preprint arXiv:2306.03310, 2023. 2, 5, 1 +[38] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 2 +[39] Jiasen Lu, Christopher Clark, Sangho Lee, Zichen Zhang, Savya Khosla, Ryan Marten, Derek Hoiem, and Aniruddha Kembhavi. Unified-io 2: Scaling autoregressive multimodal models with vision language audio and action. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26439-26455, 2024. 2 +[40] Yecheng Jason Ma, William Liang, Guanzhi Wang, De-An Huang, Osbert Bastani, Dinesh Jayaraman, Yuke Zhu, Linxi Fan, and Anima Anandkumar. Eureka: Human-level reward design via coding large language models. arXiv preprint arXiv:Arxiv-2310.12931, 2023.3 +[41] Arjun Majumdar, Karmesh Yadav, Sergio Arnaud, Jason Ma, Claire Chen, Sneha Silwal, Aryan Jain, Vincent-Pierre Berges, Tingfan Wu, Jay Vakil, et al. Where are we in the search for an artificial visual cortex for embodied intelligence? Advances in Neural Information Processing Systems, 36:655-677, 2023. 3 +[42] Ajay Mandlekar, Yuke Zhu, Animesh Garg, Jonathan Booher, Max Spero, Albert Tung, Julian Gao, John Emmons, Anchit Gupta, Emre Orbay, et al. Roboturk: A crowdsourcing platform for robotic skill learning through imitation. In Conference on Robot Learning, pages 879-893. PMLR, 2018. 1 +[43] Oier Mees, Jessica Borja-Diaz, and Wolfram Burgard. Grounding language with visual affordances over unstructured data. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 11576-11582. IEEE, 2023. 1 +[44] Zawalski Michal, Chen William, Pertsch Karl, Mees Oier, Finn Chelsea, and Levine Sergey. Robotic control via embodied chain-of-thought reasoning. arXiv preprint arXiv:2407.08693, 2024. 1, 2 +[45] Yao Mu, Qinglong Zhang, Mengkang Hu, Wenhai Wang, Mingyu Ding, Jun Jin, Bin Wang, Jifeng Dai, Yu Qiao, and Ping Luo. Embodiedgpt: Vision-language pre-training via embodied chain of thought. Advances in Neural Information Processing Systems, 36, 2024. 1, 2, 3 +[46] Ashvin V Nair, Vitchyr Pong, Murtaza Dalal, Shikhar Bahl, Steven Lin, and Sergey Levine. Visual reinforcement learning with imagined goals. Advances in neural information processing systems, 31, 2018. 2, 3 +[47] Fei Ni, Jianye Hao, Shiguang Wu, Longxin Kou, Jiashun Liu, Yan Zheng, Bin Wang, and Yuzheng Zhuang. Generate subgoal images before act: Unlocking the chain-of-thought reasoning in diffusion model for robot manipulation with multimodal prompts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 2, 3 +[48] Abby O'Neill, Abdul Rehman, Abhinav Gupta, Abhiram Maddukuri, Abhishek Gupta, Abhishek Padalkar, Abraham + +Lee, Acorn Pooley, Agrim Gupta, Ajay Mandlekar, et al. Open x-embodiment: Robotic learning datasets and rt-x models. arXiv preprint arXiv:2310.08864, 2023. 1, 2, 3, 4 +[49] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 3 +[50] Daniel Rose, Vaishnavi Himakunthala, Andy Ouyang, Ryan He, Alex Mei, Yujie Lu, Michael Saxon, Chinmay Sonar, Diba Mirza, and William Yang Wang. Visual chain of thought: bridging logical gaps with multimodal infillings. arXiv preprint arXiv:2305.02317, 2023. 2 +[51] Erick Rosete-Beas, Oier Mees, Gabriel Kalweit, Joschka Boedecker, and Wolfram Burgard. Latent plans for task-agnostic offline reinforcement learning. In Conference on Robot Learning, pages 1838–1849. PMLR, 2023. 1 +[52] V Sanh. Distilbert, a distilled version of bert: smaller, faster, cheaper and lighter. arXiv preprint arXiv:1910.01108, 2019. 6 +[53] Hao Shao, Shengju Qian, Han Xiao, Guanglu Song, Zhuofan Zong, Letian Wang, Yu Liu, and Hongsheng Li. Visual cot: Unleashing chain-of-thought reasoning in multi-modal language models. arXiv preprint arXiv:2403.16999, 2024. 2 +[54] Mohit Shridhar, Lucas Manuelli, and Dieter Fox. CIport: What and where pathways for robotic manipulation. In Conference on robot learning, pages 894-906. PMLR, 2022. 1 +[55] Mohit Shridhar, Yat Long Lo, and Stephen James. Generative image as action models. arXiv preprint arXiv:2407.07875, 2024. 2, 3 +[56] Ishika Singh, Valts Blukis, Arsalan Mousavian, Ankit Goyal, Danfei Xu, Jonathan Tremblay, Dieter Fox, Jesse Thomason, and Animesh Garg. Progprompt: Generating situated robot task plans using large language models. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 11523-11530. IEEE, 2023. 3 +[57] Yang Song, Prafulla Dhariwal, Mark Chen, and Ilya Sutskever. Consistency models. arXiv preprint arXiv:2303.01469, 2023.8 +[58] Chameleon Team. Chameleon: Mixed-modal early-fusion foundation models. arXiv preprint arXiv:2405.09818, 2024. 2 +[59] Octo Model Team, Dibya Ghosh, Homer Walke, Karl Pertsch, Kevin Black, Oier Mees, Sudeep Dasari, Joel Hejna, Tobias Kreiman, Charles Xu, et al. Octo: An open-source generalist robot policy. arXiv preprint arXiv:2405.12213, 2024. 1, 3, 6 +[60] Homer Rich Walke, Kevin Black, Tony Z Zhao, Quan Vuong, Chongyi Zheng, Philippe Hansen-Estruch, Andre Wang He, Vivek Myers, Moo Jin Kim, Max Du, et al. Bridgedata v2: A dataset for robot learning at scale. In Conference on Robot Learning, pages 1723-1736. PMLR, 2023. 2, 5, 1 +[61] Xinlong Wang, Xiaosong Zhang, Zhengxiong Luo, Quan Sun, Yufeng Cui, Jinsheng Wang, Fan Zhang, Yueze Wang, Zhen Li, Qiying Yu, et al. Emu3: Next-token prediction is all you need. arXiv preprint arXiv:2409.18869, 2024. 2, 8 + +[62] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 2022. 2, 3 +[63] Chuan Wen, Xingyu Lin, John So, Kai Chen, Qi Dou, Yang Gao, and Pieter Abbeel. Any-point trajectory modeling for policy learning. arXiv preprint arXiv:2401.00025, 2023. 1, 2, 3 +[64] Junjie Wen, Yichen Zhu, Jinming Li, Minjie Zhu, Kun Wu, Zhiyuan Xu, Ran Cheng, Chaomin Shen, Yaxin Peng, Feifei Feng, et al. Tinyvla: Towards fast, data-efficient vision-language-action models for robotic manipulation. arXiv preprint arXiv:2409.12514, 2024.3 +[65] Chengyue Wu, Xiaokang Chen, Zhiyu Wu, Yiyang Ma, Xingchao Liu, Zizheng Pan, Wen Liu, Zhenda Xie, Xingkai Yu, Chong Ruan, et al. Janus: Decoupling visual encoding for unified multimodal understanding and generation. arXiv preprint arXiv:2410.13848, 2024. 8 +[66] Hongtao Wu, Ya Jing, Chilam Cheang, Guangzeng Chen, Jiafeng Xu, Xinghang Li, Minghuan Liu, Hang Li, and Tao Kong. Unleashing large-scale video generative pretraining for visual robot manipulation. arXiv preprint arXiv:2312.13139, 2023. 1 +[67] Yecheng Wu, Zhuoyang Zhang, Junyu Chen, Haotian Tang, Dacheng Li, Yunhao Fang, Ligeng Zhu, Enze Xie, Hongxu Yin, Li Yi, et al. Vila-u: a unified foundation model integrating visual understanding and generation. arXiv preprint arXiv:2409.04429, 2024. 2, 3, 4 +[68] Jiannan Xiang, Guangyi Liu, Yi Gu, Qiyue Gao, Yuting Ning, Yuheng Zha, Zeyu Feng, Tianhua Tao, Shibo Hao, Yemin Shi, et al. Pandora: Towards general world model with natural language actions and video states. arXiv preprint arXiv:2406.09455, 2024. 8 +[69] Jinheng Xie, Weijia Mao, Zechen Bai, David Junhao Zhang, Weihao Wang, Kevin Qinghong Lin, Yuchao Gu, Zhijie Chen, Zhenheng Yang, and Mike Zheng Shou. Show-o: One single transformer to unify multimodal understanding and generation. arXiv preprint arXiv:2408.12528, 2024. 2, 8 +[70] Jonathan Yang, Catherine Glossop, Arjun Bhorkar, Dhruv Shah, Quan Vuong, Chelsea Finn, Dorsa Sadigh, and Sergey Levine. Pushing the limits of cross-embodiment learning for manipulation and navigation. arXiv preprint arXiv:2402.19432, 2024. 1 +[71] Mengjiao Yang, Yilun Du, Kamyar Ghasemipour, Jonathan Thompson, Dale Schuurmans, and Pieter Abbeel. Learning interactive real-world simulators. arXiv preprint arXiv:2310.06114, 2023. 3 +[72] Sherry Yang, Jacob Walker, Jack Parker-Holder, Yilun Du, Jake Bruce, Andre Barreto, Pieter Abbeel, and Dale Schuurmans. Video as the new language for real-world decision making. arXiv preprint arXiv:2402.17139, 2024. 8 +[73] Qihang Yu, Mark Weber, Xueqing Deng, Xiaohui Shen, Daniel Cremers, and Liang-Chieh Chen. An image is worth 32 tokens for reconstruction and generation. arxiv: 2406.07550, 2024.8 +[74] Wenhao Yu, Nimrod Gileadi, Chuyuan Fu, Sean Kirmani, Kuang-Huei Lee, Montse Gonzalez Arenas, Hao-Tien Lewis + +Chiang, Tom Erez, Leonard Hasenclever, Jan Humplik, et al. Language to rewards for robotic skill synthesis. arXiv preprint arXiv:2306.08647, 2023. 3 +[75] Eric Zelikman, Yuhuai Wu, Jesse Mu, and Noah Goodman. Star: Bootstrapping reasoning with reasoning. Advances in Neural Information Processing Systems, 35:15476-15488, 2022. 2 +[76] Kaifeng Zhang, Zhao-Heng Yin, Weirui Ye, and Yang Gao. Learning manipulation skills through robot chain-of-thought with sparse failure guidance. arXiv preprint arXiv:2405.13573, 2024. 1, 2 +[77] Tony Z Zhao, Vikash Kumar, Sergey Levine, and Chelsea Finn. Learning fine-grained bimanual manipulation with low-cost hardware. arXiv preprint arXiv:2304.13705, 2023. 2 +[78] Haoyu Zhen, Xiaowen Qiu, Peihao Chen, Jincheng Yang, Xin Yan, Yilun Du, Yining Hong, and Chuang Gan. 3d-vla: A 3d vision-language-action generative world model. arXiv preprint arXiv:2403.09631, 2024. 1 +[79] Chunting Zhou, Lili Yu, Arun Babu, Kushal Tirumala, Michihiro Yasunaga, Leonid Shamis, Jacob Kahn, Xuezhe Ma, Luke Zettlemoyer, and Omer Levy. Transfusion: Predict the next token and diffuse images with one multi-modal model. 2024. 8 +[80] Enshen Zhou, Yiran Qin, Zhenfei Yin, Yuzhou Huang, Ruimao Zhang, Lu Sheng, Yu Qiao, and Jing Shao. Mine-dreamer: Learning to follow instructions via chain-of-imagination for simulated-world control. arXiv preprint arXiv:2403.12037, 2024. 3 +[81] Gaoyue Zhou, Victoria Dean, Mohan Kumar Srirama, Aravind Rajeswaran, Jyothish Pari, Kyle Hatch, Aryan Jain, Tianhe Yu, Pieter Abbeel, Lerrel Pinto, et al. Train offline, test online: A real robot learning benchmark. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 9197-9203. IEEE, 2023. 1 +[82] Xinghao Zhu, Ran Tian, Chenfeng Xu, Mingxiao Huo, Wei Zhan, Masayoshi Tomizuka, and Mingyu Ding. Fanuc manipulation: A dataset for learning-based manipulation with fanuc mate 200id robot. https://sites.google.com/berkeley.edu/fanuc-manipulation, 2023.1 +[83] Yifeng Zhu, Abhishek Joshi, Peter Stone, and Yuke Zhu. Viola: Imitation learning for vision-based manipulation with object proposal priors. In Conference on Robot Learning, pages 1199-1210. PMLR, 2023. 1 \ No newline at end of file diff --git a/2025/CoT-VLA_ Visual Chain-of-Thought Reasoning for Vision-Language-Action Models/images.zip b/2025/CoT-VLA_ Visual Chain-of-Thought Reasoning for Vision-Language-Action Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..fa9b8b9c9ece47185e9c37da50fe896387a2e3b9 --- /dev/null +++ b/2025/CoT-VLA_ Visual Chain-of-Thought Reasoning for Vision-Language-Action Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:db6cc00a0ef7f2735dc8f739add89e4935d7fb6ec3477a9047e9362b9bc6a6f8 +size 482664 diff --git a/2025/CoT-VLA_ Visual Chain-of-Thought Reasoning for Vision-Language-Action Models/layout.json b/2025/CoT-VLA_ Visual Chain-of-Thought Reasoning for Vision-Language-Action Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..82aac9b37d774536c1465acee0b8b0e65abc734e --- /dev/null +++ b/2025/CoT-VLA_ Visual Chain-of-Thought Reasoning for Vision-Language-Action Models/layout.json @@ -0,0 +1,8806 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 148, + 103, + 463, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 103, + 463, + 140 + ], + "spans": [ + { + "bbox": [ + 148, + 103, + 463, + 140 + ], + "type": "text", + "content": "CoT-VLA: Visual Chain-of-Thought Reasoning for Vision-Language-Action Models" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 159, + 541, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 159, + 541, + 224 + ], + "spans": [ + { + "bbox": [ + 61, + 159, + 541, + 224 + ], + "type": "text", + "content": "Qingqing Zhao" + }, + { + "bbox": [ + 61, + 159, + 541, + 224 + ], + "type": "inline_equation", + "content": "^{1,2,*}" + }, + { + "bbox": [ + 61, + 159, + 541, + 224 + ], + "type": "text", + "content": " Yao Lu" + }, + { + "bbox": [ + 61, + 159, + 541, + 224 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 61, + 159, + 541, + 224 + ], + "type": "text", + "content": " Moo Jin Kim" + }, + { + "bbox": [ + 61, + 159, + 541, + 224 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 61, + 159, + 541, + 224 + ], + "type": "text", + "content": " Zipeng Fu" + }, + { + "bbox": [ + 61, + 159, + 541, + 224 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 61, + 159, + 541, + 224 + ], + "type": "text", + "content": " Zhuoyang Zhang" + }, + { + "bbox": [ + 61, + 159, + 541, + 224 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 61, + 159, + 541, + 224 + ], + "type": "text", + "content": " Yecheng Wu" + }, + { + "bbox": [ + 61, + 159, + 541, + 224 + ], + "type": "inline_equation", + "content": "^{1,3}" + }, + { + "bbox": [ + 61, + 159, + 541, + 224 + ], + "type": "text", + "content": " Zhaoshuo Li" + }, + { + "bbox": [ + 61, + 159, + 541, + 224 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 61, + 159, + 541, + 224 + ], + "type": "text", + "content": " Qianli Ma" + }, + { + "bbox": [ + 61, + 159, + 541, + 224 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 61, + 159, + 541, + 224 + ], + "type": "text", + "content": " Song Han" + }, + { + "bbox": [ + 61, + 159, + 541, + 224 + ], + "type": "inline_equation", + "content": "^{1,3}" + }, + { + "bbox": [ + 61, + 159, + 541, + 224 + ], + "type": "text", + "content": " Chelsea Finn" + }, + { + "bbox": [ + 61, + 159, + 541, + 224 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 61, + 159, + 541, + 224 + ], + "type": "text", + "content": " Ankur Handa" + }, + { + "bbox": [ + 61, + 159, + 541, + 224 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 61, + 159, + 541, + 224 + ], + "type": "text", + "content": " Tsung-Yi Lin" + }, + { + "bbox": [ + 61, + 159, + 541, + 224 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 61, + 159, + 541, + 224 + ], + "type": "text", + "content": " Gordon Wetzstein" + }, + { + "bbox": [ + 61, + 159, + 541, + 224 + ], + "type": "inline_equation", + "content": "^{2\\dagger}" + }, + { + "bbox": [ + 61, + 159, + 541, + 224 + ], + "type": "text", + "content": " Ming-Yu Liu" + }, + { + "bbox": [ + 61, + 159, + 541, + 224 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 61, + 159, + 541, + 224 + ], + "type": "text", + "content": " Donglai Xiang" + }, + { + "bbox": [ + 61, + 159, + 541, + 224 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 61, + 159, + 541, + 224 + ], + "type": "text", + "content": " NVIDIA " + }, + { + "bbox": [ + 61, + 159, + 541, + 224 + ], + "type": "inline_equation", + "content": "{}^{2}" + }, + { + "bbox": [ + 61, + 159, + 541, + 224 + ], + "type": "text", + "content": "Stanford University " + }, + { + "bbox": [ + 61, + 159, + 541, + 224 + ], + "type": "inline_equation", + "content": "{}^{3}" + }, + { + "bbox": [ + 61, + 159, + 541, + 224 + ], + "type": "text", + "content": "MIT" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 151, + 251, + 200, + 264 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 251, + 200, + 264 + ], + "spans": [ + { + "bbox": [ + 151, + 251, + 200, + 264 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 277, + 297, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 277, + 297, + 517 + ], + "spans": [ + { + "bbox": [ + 55, + 277, + 297, + 517 + ], + "type": "text", + "content": "Vision-language-action models (VLAs) have shown potential in leveraging pretrained vision-language models and diverse robot demonstrations for learning generalizable sensorimotor control. While this paradigm effectively utilizes large-scale data from both robotic and non-robotic sources, current VLAs primarily focus on direct input-output mappings, lacking the intermediate reasoning steps crucial for complex manipulation tasks. As a result, existing VLAs lack temporal planning or reasoning capabilities. In this paper, we introduce a method that incorporates explicit visual chain-of-thought (CoT) reasoning into vision-language-action models (VLAs) by predicting future image frames autoregressively as visual goals before generating a short action sequence to achieve these goals. We introduce CoT-VLA, a state-of-the-art 7B VLA that can understand and generate visual and action tokens. Our experimental results demonstrate that CoT-VLA achieves strong performance, outperforming the state-of-the-art VLA model by " + }, + { + "bbox": [ + 55, + 277, + 297, + 517 + ], + "type": "inline_equation", + "content": "17\\%" + }, + { + "bbox": [ + 55, + 277, + 297, + 517 + ], + "type": "text", + "content": " in real-world manipulation tasks and " + }, + { + "bbox": [ + 55, + 277, + 297, + 517 + ], + "type": "inline_equation", + "content": "6\\%" + }, + { + "bbox": [ + 55, + 277, + 297, + 517 + ], + "type": "text", + "content": " in simulation benchmarks. Videos are available at: https://cot-vla.github.io/." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 542, + 135, + 554 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 542, + 135, + 554 + ], + "spans": [ + { + "bbox": [ + 56, + 542, + 135, + 554 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 563, + 297, + 682 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 563, + 297, + 682 + ], + "spans": [ + { + "bbox": [ + 55, + 563, + 297, + 682 + ], + "type": "text", + "content": "Recent advances in robot learning have demonstrated impressive progress in training policies that can act across diverse tasks and environments [1, 5, 12, 14, 18, 29, 36, 44, 45, 48, 54, 59, 63, 66, 70, 76, 78]. One promising direction is vision-language-action (VLA) models, which leverage the rich understanding capabilities of pretrained vision-language models (VLMs) to map natural language instructions and visual observations to robot actions [12, 29, 48]. By training VLMs on robot demonstrations, VLAs inherit their ability to understand diverse scenes, objects, and natural language in-" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 318, + 251, + 552, + 526 + ], + "blocks": [ + { + "bbox": [ + 318, + 251, + 552, + 526 + ], + "lines": [ + { + "bbox": [ + 318, + 251, + 552, + 526 + ], + "spans": [ + { + "bbox": [ + 318, + 251, + 552, + 526 + ], + "type": "image", + "image_path": "bd662ec31266da1396db03846cc4a2f7312b32a1a370c3c820dd1e662ac80f6a.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 533, + 556, + 655 + ], + "lines": [ + { + "bbox": [ + 313, + 533, + 556, + 655 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 556, + 655 + ], + "type": "text", + "content": "Figure 1. Comparison between vanilla VLA and CoT-VLA frameworks. Prior VLA models (top) directly predict robot actions from task inputs without explicit reasoning steps and only use action-annotated robot demonstration data for training. Unlike vanilla VLAs, CoT-VLA (bottom) can also leverage action-less datasets like EPIC-KITCHEN-100 [27] to enhance subgoal image generation ability, unlocking the potential of using abundant unlabeled video data to improve VLA's visual reasoning capability. CoT-VLA first generates a subgoal image as an intermediate reasoning step, and then generate a short action sequence to achieve the subgoal. We outline the robot arm for better visualization." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "text", + "content": "structions, leading to better generalization capabilities when fine-tuned for downstream testing scenarios. While these approaches have shown impressive results, they typically map" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 693, + 215, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 693, + 215, + 703 + ], + "spans": [ + { + "bbox": [ + 66, + 693, + 215, + 703 + ], + "type": "text", + "content": "*Work done during an internship at NVIDIA." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 67, + 703, + 123, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 123, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 123, + 713 + ], + "type": "text", + "content": "Equal Advising" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "text", + "content": "1702" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 108 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 108 + ], + "type": "text", + "content": "directly from observations to actions without explicit intermediate reasoning steps that could improve interpretability and, potentially, performance." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 110, + 297, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 110, + 297, + 350 + ], + "spans": [ + { + "bbox": [ + 56, + 110, + 297, + 350 + ], + "type": "text", + "content": "In the language domain, chain-of-thought (CoT) prompting has emerged as a powerful technique for improving the reasoning capabilities of large language models (LLMs) by encouraging step-by-step thinking [62, 75]. Applying these concepts to robotics presents exciting opportunities for grounding reasoning in text, visual observations, and physical actions. Recent works have made progress in this direction by incorporating intermediate reasoning steps like language descriptions, keypoints, or bounding boxes [15, 44, 45, 63]. These intermediate representations capture abstracted states of scenes, objects, and tasks and often require additional pre-processing pipelines. In our work, we explore subgoal images as an intermediate reasoning step before action generation. These images capture the state of the model's reasoning process and are naturally available within robot demonstration datasets. While prior work has explored subgoal generation and goal-conditioned imitation learning [2, 11, 46, 55], to the best of our knowledge, our approach is the first to integrate these concepts with VLAs as intermediate chain-of-thought reasoning steps." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 352, + 297, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 352, + 297, + 532 + ], + "spans": [ + { + "bbox": [ + 55, + 352, + 297, + 532 + ], + "type": "text", + "content": "We propose visual chain-of-thought reasoning for VLAs, a new method that uses subgoal image generation as a form of chain-of-thought reasoning for robotic tasks. Rather than directly predicting actions, our method first generates a subgoal image that represents the robot's planned state in pixel space, and then conditions its action on both the current observation and the generated subgoal image. This approach allows the model to \"think visually\" about how to accomplish a task before acting. By using the subgoal image as intermediate reasoning step, we leverage information that already exists in robot manipulation data with minimal preprocessing required. Furthermore, since subgoal image generation does not require action annotations, this unlocks the potential of using abundant video data for improved visual reasoning and understanding." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 534, + 297, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 297, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 297, + 713 + ], + "type": "text", + "content": "We build our CoT-VLA system that leverages visual chain-of-thought reasoning upon recent advances in unified multimodal foundation models that can understand and generate text and images [39, 58, 61, 67, 69]. We train our base model [67] on both the Open X-Embodiment dataset [48] and action-less video datasets [20, 27], and then fine-tune the model on task demonstrations collected on downstream robot setups used for deployment and evaluation. We design a hybrid attention mechanism for CoT-VLA: we use causal attention with next-token prediction for text and image generation, and leverage full attention to predict all action dimensions at once. Additionally, inspired by recent advances in robot learning [10, 17, 77], we predict sequences of actions (action chunking) rather than a single action at each timestep. We demonstrate that both action chunking and the hybrid" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 539, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 539, + 84 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 539, + 84 + ], + "type": "text", + "content": "attention mechanism improve the model's performance." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 84, + 555, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 84, + 555, + 144 + ], + "spans": [ + { + "bbox": [ + 313, + 84, + 555, + 144 + ], + "type": "text", + "content": "Through extensive experiments in both simulation benchmarks [37] and real-world experiments[48, 60], we demonstrate that our visual chain-of-thought reasoning helps improve policy performance compared to prior VLA approaches. Our key contributions include:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 144, + 555, + 287 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 314, + 144, + 555, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 144, + 555, + 179 + ], + "spans": [ + { + "bbox": [ + 314, + 144, + 555, + 179 + ], + "type": "text", + "content": "- We introduce a method of visual chain-of-thought reasoning through subgoal image generation as an intermediate reasoning step for robotic control." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 180, + 555, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 180, + 555, + 228 + ], + "spans": [ + { + "bbox": [ + 314, + 180, + 555, + 228 + ], + "type": "text", + "content": "- We introduce a system CoT-VLA that incorporates visual chain-of-thought reasoning, and a hybrid attention mechanism that combines causal attention for pixel and text generation and full attention for action prediction." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 228, + 555, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 228, + 555, + 287 + ], + "spans": [ + { + "bbox": [ + 314, + 228, + 555, + 287 + ], + "type": "text", + "content": "- We conduct comprehensive evaluations in both simulation and the real world, demonstrating that visual chain-of-thought reasoning improves VLA performance, and our system achieves state-of-the-art performance across multiple robot platforms and tasks." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 297, + 400, + 308 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 297, + 400, + 308 + ], + "spans": [ + { + "bbox": [ + 314, + 297, + 400, + 308 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 316, + 556, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 316, + 556, + 639 + ], + "spans": [ + { + "bbox": [ + 313, + 316, + 556, + 639 + ], + "type": "text", + "content": "Chain-of-Thought (CoT) Reasoning CoT reasoning has gained prominence in natural language processing, particularly for enabling models to perform complex, multi-step reasoning tasks by breaking down problem-solving into sequential, explainable steps. Early work on CoT reasoning [62] has demonstrated the effectiveness of prompting large language models to generate intermediate reasoning steps before arriving at a final answer. Extending this paradigm to the visual domain, researchers have explored multimodal chain-of-thought methods, where visual information is processed iteratively in a stepwise fashion to reason about future outcomes or states, including generating bounding boxes [53], intermediate image infillments using Stable Diffusion [50] or standard Python packages [24], or generating CLIP embeddings [22]. Recently, CoT reasoning has been explored in embodied applications. It can generate textual plans for multi-stage execution [44, 45], point trajectories [63], label bounding boxes of objects and gripper positions as additional observations [44], generate future image trajectories for open-loop following [35, 47], and generate fine-grained reward guidance for reinforcement learning [76]. In this work, we introduce Visual-CoT reasoning for robotic manipulation, where predicted subgoal images serve as intermediate reasoning steps for closed-loop action generation. This approach leverages demonstration videos as natural intermediate reasoning states without requiring additional annotations." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "text", + "content": "Vision-Language-Action Models Large pretrained vision-language models (VLMs) [9, 28, 38] have emerged as a powerful tool for robot learning, and recent works have explored various approaches to integrate them into robot systems. Several works utilize VLMs as intermediate" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "text", + "content": "1703" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 76, + 555, + 228 + ], + "blocks": [ + { + "bbox": [ + 58, + 76, + 555, + 228 + ], + "lines": [ + { + "bbox": [ + 58, + 76, + 555, + 228 + ], + "spans": [ + { + "bbox": [ + 58, + 76, + 555, + 228 + ], + "type": "image", + "image_path": "50788cdfa18198255180faf8692616ccb541f830221d05133caaf4dace9548cd.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 236, + 555, + 293 + ], + "lines": [ + { + "bbox": [ + 54, + 236, + 555, + 293 + ], + "spans": [ + { + "bbox": [ + 54, + 236, + 555, + 293 + ], + "type": "text", + "content": "Figure 2. Overview of CoT-VLA framework. We build our model on VILA-U [67], a generative multimodal model pretrained on interleaved text-image data. The base model then trains on robot demonstrations [48] and action-less videos [20, 27]. During deployment, given a visual observation and a text instruction, the model performs visual chain-of-thought reasoning by generating a subgoal image (upper blue) with causal attention. It then generates a short action sequence with full attention " + }, + { + "bbox": [ + 54, + 236, + 555, + 293 + ], + "type": "inline_equation", + "content": "(\\mathbf{a}_1 \\cdots \\mathbf{a}_n)" + }, + { + "bbox": [ + 54, + 236, + 555, + 293 + ], + "type": "text", + "content": " for robot execution. The system operates in a closed-loop control manner by capturing new observations after executing predicted action sequences." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 313, + 297, + 611 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 313, + 297, + 611 + ], + "spans": [ + { + "bbox": [ + 56, + 313, + 297, + 611 + ], + "type": "text", + "content": "components for perception and control, leveraging their strong semantic understanding and reasoning capabilities to decompose complex tasks [21, 26, 34, 56], detect objects [19, 25], or generate dense rewards [13, 40, 74] or goals [2, 11, 15, 46, 47, 55, 71, 80]. Some approaches incorporate VLMs [4, 49, 64] into end-to-end trainable policies by using them as pretrained backbone for better visuo-language representation [12, 29, 41, 48, 59]. Most relevant to our work are recent approaches that fine-tune pretrained VLMs on robot demonstration data for direct action prediction [12, 29, 48]. These VLAs demonstrate improved generalization to novel objects, environments, and natural language instructions through pretraining on internet-scale vision-language datasets, providing a promising direction for transferring visual and language knowledge to robotic control tasks. However, most existing VLAs do not leverage the step-by-step reasoning capabilities demonstrated in large language models, which have been shown to significantly improve performance across various tasks [62]. In the past, researchers have used chain-of-thought reasoning on language instructions or intermediate keypoints/bounding boxes for robotics [6, 13, 45, 63]. We introduce visual chain-of-thought reasoning to the VLA frameworks, using subgoal images as intermediate reasoning steps before action generation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 621, + 122, + 634 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 621, + 122, + 634 + ], + "spans": [ + { + "bbox": [ + 55, + 621, + 122, + 634 + ], + "type": "text", + "content": "3. CoT-VLA" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 642, + 297, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 297, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 297, + 715 + ], + "type": "text", + "content": "In this section, we present our visual chain-of-thought reasoning framework for VLAs. We begin with the formulation of our method (3.1), followed by a detailed description of the system architecture (3.2). We then explain our training procedures (3.3) and outline the deployment strategy for downstream tasks (3.3)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 312, + 507, + 326 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 312, + 507, + 326 + ], + "spans": [ + { + "bbox": [ + 313, + 312, + 507, + 326 + ], + "type": "text", + "content": "3.1. Visual Chain-of-Thought Reasoning" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 331, + 556, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 331, + 556, + 437 + ], + "spans": [ + { + "bbox": [ + 313, + 331, + 556, + 437 + ], + "type": "text", + "content": "We consider two types of training data for VLA pretraining: robot demonstrations dataset " + }, + { + "bbox": [ + 313, + 331, + 556, + 437 + ], + "type": "inline_equation", + "content": "D_r" + }, + { + "bbox": [ + 313, + 331, + 556, + 437 + ], + "type": "text", + "content": " and action-less videos dataset " + }, + { + "bbox": [ + 313, + 331, + 556, + 437 + ], + "type": "inline_equation", + "content": "D_v" + }, + { + "bbox": [ + 313, + 331, + 556, + 437 + ], + "type": "text", + "content": ". Robot demonstrations are represented as " + }, + { + "bbox": [ + 313, + 331, + 556, + 437 + ], + "type": "inline_equation", + "content": "D_r = \\{(l, \\mathbf{a}_{1\\dots T}, \\mathbf{s}_{1\\dots T})\\}" + }, + { + "bbox": [ + 313, + 331, + 556, + 437 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 331, + 556, + 437 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 313, + 331, + 556, + 437 + ], + "type": "text", + "content": " denotes the natural language instruction, " + }, + { + "bbox": [ + 313, + 331, + 556, + 437 + ], + "type": "inline_equation", + "content": "\\mathbf{a}_{1\\dots T} = \\{\\mathbf{a}_1, \\dots, \\mathbf{a}_T\\}" + }, + { + "bbox": [ + 313, + 331, + 556, + 437 + ], + "type": "text", + "content": " denotes the sequence of robot actions, and " + }, + { + "bbox": [ + 313, + 331, + 556, + 437 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_{1\\dots T} = \\{\\mathbf{s}_1, \\dots, \\mathbf{s}_T\\}" + }, + { + "bbox": [ + 313, + 331, + 556, + 437 + ], + "type": "text", + "content": " denotes the visual observations as a sequence of images. Action-less videos " + }, + { + "bbox": [ + 313, + 331, + 556, + 437 + ], + "type": "inline_equation", + "content": "D_v = \\{(l, \\mathbf{s}_{1\\dots T})\\}" + }, + { + "bbox": [ + 313, + 331, + 556, + 437 + ], + "type": "text", + "content": " consist of language descriptions and images without action annotations." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 452, + 556, + 500 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 452, + 556, + 500 + ], + "spans": [ + { + "bbox": [ + 313, + 452, + 556, + 500 + ], + "type": "text", + "content": "VLA: Vanilla VLA approaches fine-tune a pretrained VLM, " + }, + { + "bbox": [ + 313, + 452, + 556, + 500 + ], + "type": "inline_equation", + "content": "P_{\\theta}" + }, + { + "bbox": [ + 313, + 452, + 556, + 500 + ], + "type": "text", + "content": ", on " + }, + { + "bbox": [ + 313, + 452, + 556, + 500 + ], + "type": "inline_equation", + "content": "D_{r}" + }, + { + "bbox": [ + 313, + 452, + 556, + 500 + ], + "type": "text", + "content": ", learning to predict actions " + }, + { + "bbox": [ + 313, + 452, + 556, + 500 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{a}}_{t+1}" + }, + { + "bbox": [ + 313, + 452, + 556, + 500 + ], + "type": "text", + "content": " directly from the current observation " + }, + { + "bbox": [ + 313, + 452, + 556, + 500 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_{t}" + }, + { + "bbox": [ + 313, + 452, + 556, + 500 + ], + "type": "text", + "content": " and language instruction " + }, + { + "bbox": [ + 313, + 452, + 556, + 500 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 313, + 452, + 556, + 500 + ], + "type": "text", + "content": " (Figure 1, top):" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 398, + 506, + 555, + 519 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 398, + 506, + 555, + 519 + ], + "spans": [ + { + "bbox": [ + 398, + 506, + 555, + 519 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {a}} _ {t} \\sim P _ {\\theta} (\\mathbf {a} _ {t} | \\mathbf {s} _ {t}, l) \\tag {1}", + "image_path": "e8f52abd084ad55abe27e81362730d39d9d281fb95e430c20f7bee01e00046bc.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 531, + 555, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 531, + 555, + 568 + ], + "spans": [ + { + "bbox": [ + 313, + 531, + 555, + 568 + ], + "type": "text", + "content": "CoT-VLA: Our key insight is to incorporate explicit visual reasoning before action generation. As illustrated in Figure 2, our approach operates in two sequential phases:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 412, + 573, + 555, + 586 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 412, + 573, + 555, + 586 + ], + "spans": [ + { + "bbox": [ + 412, + 573, + 555, + 586 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {s}} _ {t + n} \\sim P _ {\\theta} \\left(\\mathbf {s} _ {t + n} \\mid \\mathbf {s} _ {t}, l\\right) \\tag {2}", + "image_path": "6b1b3d64012ef552e6bfce890d4c913570a6f509c8a5f1fdb881df3a37399f14.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 341, + 587, + 555, + 601 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 341, + 587, + 555, + 601 + ], + "spans": [ + { + "bbox": [ + 341, + 587, + 555, + 601 + ], + "type": "interline_equation", + "content": "\\{\\hat {\\mathbf {a}} _ {t}, \\dots , \\hat {\\mathbf {a}} _ {t + m} \\} \\sim P _ {\\theta} (\\{\\mathbf {a} _ {t}, \\dots , \\mathbf {a} _ {t + m} | \\mathbf {s} _ {t}, l, \\mathbf {s} _ {t + n}) \\tag {3}", + "image_path": "ff457129206a2196f6c1ddb8549b44749d1311269848d8e8bbd22f043a7fceef.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 605, + 557, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 557, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 557, + 715 + ], + "type": "text", + "content": "where we first predict a subgoal image " + }, + { + "bbox": [ + 313, + 605, + 557, + 715 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{s}}_{t + n}" + }, + { + "bbox": [ + 313, + 605, + 557, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 605, + 557, + 715 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 605, + 557, + 715 + ], + "type": "text", + "content": " frames ahead, as an intermediate visual reasoning step (Equation 2). Then we generate a sequence of " + }, + { + "bbox": [ + 313, + 605, + 557, + 715 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 605, + 557, + 715 + ], + "type": "text", + "content": " actions to achieve this subgoal state (Equation 3). This enables the model to \"think visually\" first by explicitly reasoning about desired future states before predicting the actions. The visual reasoning step, Equation (2), is trained on both robot demonstrations " + }, + { + "bbox": [ + 313, + 605, + 557, + 715 + ], + "type": "inline_equation", + "content": "D_r" + }, + { + "bbox": [ + 313, + 605, + 557, + 715 + ], + "type": "text", + "content": " and actionless videos " + }, + { + "bbox": [ + 313, + 605, + 557, + 715 + ], + "type": "inline_equation", + "content": "D_v" + }, + { + "bbox": [ + 313, + 605, + 557, + 715 + ], + "type": "text", + "content": ", and the action generation step, Equation (3), is trained on robot demonstrations " + }, + { + "bbox": [ + 313, + 605, + 557, + 715 + ], + "type": "inline_equation", + "content": "D_r" + }, + { + "bbox": [ + 313, + 605, + 557, + 715 + ], + "type": "text", + "content": " only." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1704" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 90, + 69, + 265, + 228 + ], + "blocks": [ + { + "bbox": [ + 90, + 69, + 265, + 228 + ], + "lines": [ + { + "bbox": [ + 90, + 69, + 265, + 228 + ], + "spans": [ + { + "bbox": [ + 90, + 69, + 265, + 228 + ], + "type": "image", + "image_path": "5db19d53593396530375d73b8f75a3bbac011a98bf56c372acb801dadf706ffa.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 236, + 297, + 281 + ], + "lines": [ + { + "bbox": [ + 55, + 236, + 297, + 281 + ], + "spans": [ + { + "bbox": [ + 55, + 236, + 297, + 281 + ], + "type": "text", + "content": "Figure 3. Hybrid attention mechanism in CoT-VLA. We use causal attention for image or text generation and full attention for action generation. " + }, + { + "bbox": [ + 55, + 236, + 297, + 281 + ], + "type": "inline_equation", + "content": "[x]" + }, + { + "bbox": [ + 55, + 236, + 297, + 281 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 236, + 297, + 281 + ], + "type": "inline_equation", + "content": "[\\theta]" + }, + { + "bbox": [ + 55, + 236, + 297, + 281 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 236, + 297, + 281 + ], + "type": "inline_equation", + "content": "[g]" + }, + { + "bbox": [ + 55, + 236, + 297, + 281 + ], + "type": "text", + "content": " are special tokens for parallel decoding of actions." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 303, + 237, + 316 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 303, + 237, + 316 + ], + "spans": [ + { + "bbox": [ + 55, + 303, + 237, + 316 + ], + "type": "text", + "content": "3.2. The Base Vision-Language Model" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 323, + 297, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 323, + 297, + 370 + ], + "spans": [ + { + "bbox": [ + 55, + 323, + 297, + 370 + ], + "type": "text", + "content": "To enable the visual reasoning capabilities described in Equation (2), we build upon VILA-U [67], an unified multimodal foundation model capable of both understanding and generating image and text tokens." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 372, + 296, + 600 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 372, + 296, + 600 + ], + "spans": [ + { + "bbox": [ + 55, + 372, + 296, + 600 + ], + "type": "text", + "content": "VILA-U unifies video, image, and language understanding through an autoregressive next-token prediction framework. At its core is a unified vision tower that encodes visual inputs as discrete tokens aligned with textual information. This enables autoregressive image and video generation while significantly enhancing the understanding capabilities of VLMs that leverage discrete visual features. VILA-U utilizes residual quantization [32] to improve the representational capacity of discrete visual features - incorporating a depth transformer, as introduced in RQ-VAE [32], to gradually predict the residual tokens. The extracted visual features are then passed through a projector before being processed by the LLM backbone. The base model is trained on multimodal pairs including [image, text], [text, image], [video, text], and [text, video]. We use the VILA-U model trained on " + }, + { + "bbox": [ + 55, + 372, + 296, + 600 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 55, + 372, + 296, + 600 + ], + "type": "text", + "content": " resolution images, where each image is encoded into " + }, + { + "bbox": [ + 55, + 372, + 296, + 600 + ], + "type": "inline_equation", + "content": "16 \\times 16 \\times 4" + }, + { + "bbox": [ + 55, + 372, + 296, + 600 + ], + "type": "text", + "content": " tokens with a residual depth of 4 [32]. For detailed information about VILA-U training and architecture, we refer readers to [67]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 610, + 174, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 610, + 174, + 623 + ], + "spans": [ + { + "bbox": [ + 55, + 610, + 174, + 623 + ], + "type": "text", + "content": "3.3. Training Procedures" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 629, + 297, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 297, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 297, + 715 + ], + "type": "text", + "content": "We pretrain the base 7B VILA-U model on a combination of robot demonstrations " + }, + { + "bbox": [ + 55, + 629, + 297, + 715 + ], + "type": "inline_equation", + "content": "D_r" + }, + { + "bbox": [ + 55, + 629, + 297, + 715 + ], + "type": "text", + "content": " and action-less videos " + }, + { + "bbox": [ + 55, + 629, + 297, + 715 + ], + "type": "inline_equation", + "content": "D_v" + }, + { + "bbox": [ + 55, + 629, + 297, + 715 + ], + "type": "text", + "content": ". During training, we optimize three components, the LLM backbone, projector, and depth transformer, while keeping the vision tower fixed. Our training objective has two key components: subgoal image generation with causal attention (2) and action generation with full attention (3)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "type": "text", + "content": "Visual Tokens Prediction For subgoal image generation, each training sequence is of form " + }, + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "type": "inline_equation", + "content": "(l, \\mathbf{s}_t, \\mathbf{s}_{t + n})" + }, + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "type": "text", + "content": ". We follow the training objective used in [67]. At each visual position " + }, + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "type": "text", + "content": ", the depth transformer, " + }, + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "type": "inline_equation", + "content": "P_{\\delta}" + }, + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "type": "text", + "content": ", autoregressively predicts " + }, + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "type": "text", + "content": " residual tokens " + }, + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "type": "inline_equation", + "content": "(k_{j1}, \\dots, k_{jD})" + }, + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "type": "text", + "content": " based on the LLM-generated code embedding " + }, + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "type": "inline_equation", + "content": "h_j" + }, + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "type": "text", + "content": ". The training objective for visual tokens is then formulated as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 358, + 161, + 555, + 196 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 161, + 555, + 196 + ], + "spans": [ + { + "bbox": [ + 358, + 161, + 555, + 196 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {v i s u a l}} = - \\sum_ {j} \\sum_ {d = 1} ^ {D} \\log P _ {\\delta} \\left(k _ {j d} \\mid k _ {j, < d}\\right) \\tag {4}", + "image_path": "ca3f5c08b137315915b06fa41b50414411cba5d9c9b2434f742f3f2c15dfff9a.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 201, + 555, + 238 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 201, + 555, + 238 + ], + "spans": [ + { + "bbox": [ + 313, + 201, + 555, + 238 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 201, + 555, + 238 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 201, + 555, + 238 + ], + "type": "text", + "content": " indexes the positions containing visual tokens. For a more detailed explanation of this loss function, we refer readers to [67], Section 3.2." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 251, + 556, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 251, + 556, + 419 + ], + "spans": [ + { + "bbox": [ + 313, + 251, + 556, + 419 + ], + "type": "text", + "content": "Action Tokens Prediction For action prediction, each training sequence takes the form " + }, + { + "bbox": [ + 313, + 251, + 556, + 419 + ], + "type": "inline_equation", + "content": "(l, \\mathbf{s}_t, \\mathbf{s}_{t+n}, \\mathbf{a}_t, \\dots, \\mathbf{a}_{t+m})" + }, + { + "bbox": [ + 313, + 251, + 556, + 419 + ], + "type": "text", + "content": ". Each action " + }, + { + "bbox": [ + 313, + 251, + 556, + 419 + ], + "type": "inline_equation", + "content": "\\mathbf{a}_i" + }, + { + "bbox": [ + 313, + 251, + 556, + 419 + ], + "type": "text", + "content": " is represented by 7 tokens, with each action dimension independently discretized. Following [29], we map each continuous action dimension into 256 discrete bins, with bin widths determined by uniformly dividing the interval between the 1st and 99th percentiles of the training data's action distribution. We repurpose the 256 least frequently used tokens in the text tokenizer's vocabulary as action bin tokens. Unlike prior works [12, 29, 48], we employ full attention for processing and predicting action tokens, enabling all action tokens to interact with each other. This attention mechanism is illustrated in Figure 3. During training, we minimize the cross-entropy loss for action predictions:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 343, + 426, + 555, + 457 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 343, + 426, + 555, + 457 + ], + "spans": [ + { + "bbox": [ + 343, + 426, + 555, + 457 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {a c t i o n}} = - \\sum_ {i = 1} ^ {m} \\log P _ {\\theta} \\left(\\mathbf {a} _ {t} \\dots \\mathbf {a} _ {t + m} \\mid l, s _ {t}, s _ {t + n}\\right) \\tag {5}", + "image_path": "78e26ad10032c3b9006c8b55a41e016e5d5c96fdc6af71abbda40ed80bc4ee0e.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 464, + 555, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 464, + 555, + 487 + ], + "spans": [ + { + "bbox": [ + 313, + 464, + 555, + 487 + ], + "type": "text", + "content": "Given a batch of input sequences, The overall training objective combines the action and visual losses:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 392, + 496, + 555, + 509 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 392, + 496, + 555, + 509 + ], + "spans": [ + { + "bbox": [ + 392, + 496, + 555, + 509 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {\\text {a c t i o n}} + \\mathcal {L} _ {\\text {v i s u a l}} \\tag {6}", + "image_path": "6fa2e0835abda6a02693a381abb8c2176737575800e6e908ef79032ddc0ed285.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 521, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 556, + 715 + ], + "type": "text", + "content": "Pretraining Phase We pretrain CoT-VLA on both robot demonstrations " + }, + { + "bbox": [ + 313, + 521, + 556, + 715 + ], + "type": "inline_equation", + "content": "D_r" + }, + { + "bbox": [ + 313, + 521, + 556, + 715 + ], + "type": "text", + "content": " and action-less videos " + }, + { + "bbox": [ + 313, + 521, + 556, + 715 + ], + "type": "inline_equation", + "content": "D_v" + }, + { + "bbox": [ + 313, + 521, + 556, + 715 + ], + "type": "text", + "content": " as described in Section 3.1. For robot demonstrations, we curate a subset of the Open X-Embodiment dataset [48] (OpenX). Following the pre-processing pipeline established in OpenVLA [29], we select and process datasets with third-person camera views and single-arm end-effector control (7-DoF). For action-less videos " + }, + { + "bbox": [ + 313, + 521, + 556, + 715 + ], + "type": "inline_equation", + "content": "D_v" + }, + { + "bbox": [ + 313, + 521, + 556, + 715 + ], + "type": "text", + "content": ", we incorporate the EPIC-KITCHENS [27] and Something-Something V2 [20] datasets. All images are processed at " + }, + { + "bbox": [ + 313, + 521, + 556, + 715 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 313, + 521, + 556, + 715 + ], + "type": "text", + "content": " resolution. For visual reasoning, we use subgoal images at future timestep " + }, + { + "bbox": [ + 313, + 521, + 556, + 715 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 521, + 556, + 715 + ], + "type": "text", + "content": " uniformly sampled from a dataset-specific range " + }, + { + "bbox": [ + 313, + 521, + 556, + 715 + ], + "type": "inline_equation", + "content": "[n_l, n_u]" + }, + { + "bbox": [ + 313, + 521, + 556, + 715 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 521, + 556, + 715 + ], + "type": "inline_equation", + "content": "n_l" + }, + { + "bbox": [ + 313, + 521, + 556, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 521, + 556, + 715 + ], + "type": "inline_equation", + "content": "n_u" + }, + { + "bbox": [ + 313, + 521, + 556, + 715 + ], + "type": "text", + "content": " define the lower and upper bounds of the prediction horizon. We use an action chunk size of 10. For complete dataset specifications and training hyperparameters, please refer to the supplementary material." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1705" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 90, + 70, + 520, + 144 + ], + "blocks": [ + { + "bbox": [ + 90, + 70, + 520, + 144 + ], + "lines": [ + { + "bbox": [ + 90, + 70, + 520, + 144 + ], + "spans": [ + { + "bbox": [ + 90, + 70, + 520, + 144 + ], + "type": "table", + "html": "
Average (↑)Spatial (↑)Object (↑)Goal (↑)Long (↑)
Diffusion Policy72.4 ± 0.7%78.3 ± 1.1%92.5 ± 0.7%68.3 ± 1.2%50.5 ± 1.3%
Octo fine-tuned75.1 ± 0.6%78.9 ± 1.0%85.7 ± 0.9%84.6 ± 0.9%51.1 ± 1.3%
OpenVLA fine-tuned76.5 ± 0.6%84.7 ± 0.9%88.4 ± 0.8%79.2 ± 1.0%53.7 ± 1.3%
CoT-VLA-7B (ours)81.13 ± 0.6 %87.5 ± 1.4%91.6 ± 0.5%87.6 ± 0.6%69.0 ± 0.8%
", + "image_path": "3e9a7d9e9b5b15c18745f19d1f037c682e73622b9600f865f7c14ad21897aa5d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 151, + 555, + 196 + ], + "lines": [ + { + "bbox": [ + 54, + 151, + 555, + 196 + ], + "spans": [ + { + "bbox": [ + 54, + 151, + 555, + 196 + ], + "type": "text", + "content": "Table 1. LIBERO benchmark experimental results. For each task suite (Spatial, Object, Goal, Long), we report the average success rate and standard error across 3 seeds with 500 episodes each. CoT-VLA achieves the best or competitive performance across all LIBERO benchmarks suites compared to baseline approaches. The bolded entries correspond to highest success rates while underlined entries correspond to second-highest." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 207, + 297, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 207, + 297, + 327 + ], + "spans": [ + { + "bbox": [ + 55, + 207, + 297, + 327 + ], + "type": "text", + "content": "Adaptation Phase for Downstream Closed-Loop Deployment For adaptation to downstream tasks, we fine-tune our pretrained model using task-specific robot demonstration data " + }, + { + "bbox": [ + 55, + 207, + 297, + 327 + ], + "type": "inline_equation", + "content": "D_{r}" + }, + { + "bbox": [ + 55, + 207, + 297, + 327 + ], + "type": "text", + "content": " collected on the target robot setups. During this phase, we optimize the LLM backbone, projector, and depth transformer while keeping the vision tower frozen, maintaining the same training setup as the pretraining stage. The resulting model can execute new manipulation tasks based on natural language commands " + }, + { + "bbox": [ + 55, + 207, + 297, + 327 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 55, + 207, + 297, + 327 + ], + "type": "text", + "content": ". Algorithm 1 describes our robot control procedure at test time." + } + ] + } + ], + "index": 2 + }, + { + "type": "code", + "bbox": [ + 55, + 361, + 296, + 502 + ], + "blocks": [ + { + "bbox": [ + 56, + 347, + 270, + 360 + ], + "lines": [ + { + "bbox": [ + 56, + 347, + 270, + 360 + ], + "spans": [ + { + "bbox": [ + 56, + 347, + 270, + 360 + ], + "type": "text", + "content": "Algorithm 1 CoT-VLA test-time closed-loop control" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 55, + 361, + 296, + 502 + ], + "lines": [ + { + "bbox": [ + 55, + 361, + 296, + 502 + ], + "spans": [ + { + "bbox": [ + 55, + 361, + 296, + 502 + ], + "type": "text", + "content": "Require: CoT-VLA Model " + }, + { + "bbox": [ + 55, + 361, + 296, + 502 + ], + "type": "inline_equation", + "content": "P_{\\theta}" + }, + { + "bbox": [ + 55, + 361, + 296, + 502 + ], + "type": "text", + "content": ", initial state " + }, + { + "bbox": [ + 55, + 361, + 296, + 502 + ], + "type": "inline_equation", + "content": "s_0^{\\mathrm{obs}}" + }, + { + "bbox": [ + 55, + 361, + 296, + 502 + ], + "type": "text", + "content": ", language instruction " + }, + { + "bbox": [ + 55, + 361, + 296, + 502 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 55, + 361, + 296, + 502 + ], + "type": "text", + "content": " \n0: " + }, + { + "bbox": [ + 55, + 361, + 296, + 502 + ], + "type": "inline_equation", + "content": "t \\gets 0" + }, + { + "bbox": [ + 55, + 361, + 296, + 502 + ], + "type": "text", + "content": " \n0: while True do \n0: sample " + }, + { + "bbox": [ + 55, + 361, + 296, + 502 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{s}}_{t+n} \\sim P_{\\theta}(\\mathbf{s}_{t+n} \\mid l, \\mathbf{s}_t^{\\mathrm{obs}})" + }, + { + "bbox": [ + 55, + 361, + 296, + 502 + ], + "type": "text", + "content": " \n0: sample " + }, + { + "bbox": [ + 55, + 361, + 296, + 502 + ], + "type": "inline_equation", + "content": "[\\hat{\\mathbf{a}}_t, \\dots, \\hat{\\mathbf{a}}_{t+m}] \\sim P_{\\theta}(\\mathbf{a}_t, \\dots, \\mathbf{a}_{t+m} \\mid l, \\mathbf{s}_t^{\\mathrm{obs}}, \\mathbf{s}_{t+n})" + }, + { + "bbox": [ + 55, + 361, + 296, + 502 + ], + "type": "text", + "content": " \n0: for " + }, + { + "bbox": [ + 55, + 361, + 296, + 502 + ], + "type": "inline_equation", + "content": "j = 0" + }, + { + "bbox": [ + 55, + 361, + 296, + 502 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 361, + 296, + 502 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 55, + 361, + 296, + 502 + ], + "type": "text", + "content": " do \n0: execute " + }, + { + "bbox": [ + 55, + 361, + 296, + 502 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{a}}_{t+j}" + }, + { + "bbox": [ + 55, + 361, + 296, + 502 + ], + "type": "text", + "content": " \n0: end for \n0: " + }, + { + "bbox": [ + 55, + 361, + 296, + 502 + ], + "type": "inline_equation", + "content": "t \\gets t + m + 1" + }, + { + "bbox": [ + 55, + 361, + 296, + 502 + ], + "type": "text", + "content": " \n0: " + }, + { + "bbox": [ + 55, + 361, + 296, + 502 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_t^{\\mathrm{obs}} \\gets" + }, + { + "bbox": [ + 55, + 361, + 296, + 502 + ], + "type": "text", + "content": " robot observation" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "code_body" + } + ], + "index": 4, + "sub_type": "algorithm" + }, + { + "bbox": [ + 55, + 528, + 137, + 541 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 528, + 137, + 541 + ], + "spans": [ + { + "bbox": [ + 55, + 528, + 137, + 541 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 552, + 297, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 552, + 297, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 552, + 297, + 601 + ], + "type": "text", + "content": "We evaluate the effectiveness of our approach and our system through a set of experiments spanning both simulation benchmarks and real-world robot manipulation tasks. Our experiments aim to addresses following questions:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 604, + 296, + 710 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 55, + 604, + 295, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 604, + 295, + 638 + ], + "spans": [ + { + "bbox": [ + 55, + 604, + 295, + 638 + ], + "type": "text", + "content": "- How does our system perform compared to state-of-the-art baselines across multiple benchmarks and embodiments? (Section 4.2)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 639, + 296, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 639, + 296, + 673 + ], + "spans": [ + { + "bbox": [ + 55, + 639, + 296, + 673 + ], + "type": "text", + "content": "- What is the impact of pretraining, visual chain-of-thought reasoning and hybrid attention on task performance? (Section 4.3)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 675, + 296, + 710 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 675, + 296, + 710 + ], + "spans": [ + { + "bbox": [ + 55, + 675, + 296, + 710 + ], + "type": "text", + "content": "- To what extent does improved generalization in visual reasoning enhance the action prediction capabilities? (Section 4.4)" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 206, + 430, + 220 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 206, + 430, + 220 + ], + "spans": [ + { + "bbox": [ + 313, + 206, + 430, + 220 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 224, + 555, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 224, + 555, + 308 + ], + "spans": [ + { + "bbox": [ + 313, + 224, + 555, + 308 + ], + "type": "text", + "content": "We conduct evaluations across three complementary settings: the LIBERO benchmark [37] for evaluation in simulation environments, the Bridge-V2 platform [60] with its dataset of 45k robot demonstrations, and the Franka-Tabletop setup with a stationary, table-mounted Franka Emika Panda robot with limited 10 to 150 robot demonstrations for each testing scenario." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 323, + 556, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 323, + 556, + 456 + ], + "spans": [ + { + "bbox": [ + 313, + 323, + 556, + 456 + ], + "type": "text", + "content": "LIBERO Simulation Benchmark We perform evaluation on LIBERO [37], a simulation benchmark comprising four distinct task suites: LIBERO-Spatial, LIBERO-Object, LIBERO-Goal, and LIBERO-Long. Each suite contains 10 diverse tasks with 50 human-teleoperated demonstrations per task, aiming to evaluate the robot's comprehension of spatial relationships, object interactions, and task-specific objectives. We follow the same preprocessed pipeline as in [29]: (1) removing pause intervals from trajectories, (2) standardizing image resolution to " + }, + { + "bbox": [ + 313, + 323, + 556, + 456 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 313, + 323, + 556, + 456 + ], + "type": "text", + "content": " pixels, and (3) applying a 180-degree rotation to all images." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 471, + 556, + 627 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 471, + 556, + 627 + ], + "spans": [ + { + "bbox": [ + 313, + 471, + 556, + 627 + ], + "type": "text", + "content": "Bridge-V2 Real-Robot Experiments We use a 6-DoF WidowX robotic arm, following the experimental setup from Bridge-V2 [60]. Our training data has " + }, + { + "bbox": [ + 313, + 471, + 556, + 627 + ], + "type": "inline_equation", + "content": "45\\mathrm{k}" + }, + { + "bbox": [ + 313, + 471, + 556, + 627 + ], + "type": "text", + "content": " language-annotated trajectories from the Bridge-V2 dataset, encompassing diverse manipulation tasks. While the dataset was incorporated into the pretraining phase alongside OpenX, we performed additional task-specific fine-tuning exclusively on Bridge-V2 until achieving a training action prediction accuracy threshold of " + }, + { + "bbox": [ + 313, + 471, + 556, + 627 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 313, + 471, + 556, + 627 + ], + "type": "text", + "content": ". Following [29], we evaluate on four tasks designed in [29] to evaluate visual robustness (varying distractors), motion generalization (novel object positions), semantic generalization (unseen language instruction), and language grounding (instruction following)." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 641, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 556, + 713 + ], + "type": "text", + "content": "Franka-Tabletop Real-Robot Experiments We use a stationary, table-mounted Franka Emika Panda 7-DoF robot arm denoted as Franka-Tabletop. The setup is not seen during the pretraining stage and is designed to assess our model's adaptation capability to novel real-world environments with small amounts of robot demonstrations. We" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1706" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 74, + 558, + 188 + ], + "blocks": [ + { + "bbox": [ + 59, + 74, + 558, + 188 + ], + "lines": [ + { + "bbox": [ + 59, + 74, + 558, + 188 + ], + "spans": [ + { + "bbox": [ + 59, + 74, + 558, + 188 + ], + "type": "image", + "image_path": "26969250864415a73f4cc4a6fc6cd168154b53ece3d178986990c9cda66a259c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 198, + 555, + 232 + ], + "lines": [ + { + "bbox": [ + 55, + 198, + 555, + 232 + ], + "spans": [ + { + "bbox": [ + 55, + 198, + 555, + 232 + ], + "type": "text", + "content": "Figure 4. Franka-Tabletop comparisons. Evaluation across six distinct manipulation tasks, with separate models trained per task. Left: Representative initial states for each task setup. Right: Task-specific success rates and cross-task averages for our method and baselines. CoT-VLA achieves best average performance and demonstrates strong capabilities in both single-instruction and multi-instruction scenarios." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 243, + 297, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 243, + 297, + 291 + ], + "spans": [ + { + "bbox": [ + 55, + 243, + 297, + 291 + ], + "type": "text", + "content": "perform evaluations across 6 tasks: 3 narrow domain single-instruction tasks for and 3 diverse multi-instruction tasks outlined in Figure 4 and introduced in [29]. For each task, the dataset contains between 10 and 150 demonstrations." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 305, + 296, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 305, + 296, + 508 + ], + "spans": [ + { + "bbox": [ + 55, + 305, + 296, + 508 + ], + "type": "text", + "content": "Baselines We evaluate our approach against four state-of-the-art baselines. Diffusion Policy [10], a state-of-the-art imitation learning algorithm, is trained from scratch for each test scenario in LIBERO and Franka-Tabletop. The implementation incorporates action chunking and proprioception while conditioning on DistilBERT [52] language embeddings. OpenVLA [29] is an open-source VLA model that fine-tunes pretrained vision-language models on the OpenX dataset; and Octo [59] is a generalist model pretrained on OpenX without VLM initialization. For both OpenVLA and Octo, we use their published checkpoints for Bridge-V2 evaluations and fine-tune them for our LIBERO and Franka-Tabletop experiments. SUSIE [2], a two-stage approach, combines instruction-guided image editing for goal generation with a goal-conditioned policy for action generation. We evaluate SUSIE using their published checkpoint on Bridge-V2." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 515, + 170, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 515, + 170, + 528 + ], + "spans": [ + { + "bbox": [ + 55, + 515, + 170, + 528 + ], + "type": "text", + "content": "4.2. Evaluations Results" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 534, + 297, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 297, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 297, + 713 + ], + "type": "text", + "content": "LIBERO We present quantitative results in Table 1, where each method is evaluated over 500 trials per task suite, with 3 random seeds. Success rates are reported with means and standard error. Qualitative examples of our method's reasoning and execution trajectories are illustrated in Figure 5. Results demonstrate that CoT-VLA effectively adapts to tasks in the LIBERO simulation environment, achieving best or competitive performance compared to baseline approaches. By analyzing rollout videos of failure cases, we found that baseline methods occasionally overfit to visual cues while disregarding language instructions. Specifically, when initial states appear visually similar across different tasks (e.g., in LIBERO-Spatial), baseline methods execute a different task compared to the commanded task in some episodes. CoT-VLA exhibits better instruction following" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 243, + 555, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 243, + 555, + 281 + ], + "spans": [ + { + "bbox": [ + 313, + 243, + 555, + 281 + ], + "type": "text", + "content": "ability by first reasoning visually about the desired actions via language-grounded subgoal generation, and then predicting the relevant actions for achieving the goal." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 299, + 556, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 299, + 556, + 538 + ], + "spans": [ + { + "bbox": [ + 313, + 299, + 556, + 538 + ], + "type": "text", + "content": "Bridge-V2 We evaluate CoT-VLA and baselines on the Bridge-V2 benchmark across four generalization categories identified in [29]: visual generalization (\"put eggplant into pot\" with cluttered environments), motion generalization (\"put carrot on plate\" with height variations), semantic generalization (\"take purple grapes out of pot\"), and language grounding (\"put eggplant or red bottle into pot\"). We report the quantitative results in Table 2, where each task is tested with 10 trials. SUSIE [2] generates visually higher-quality goal images through its diffusion prior (see Section 5 for a detailed discussion on our limitations) but achieves lower success rates on tasks involving novel objects or requiring complex language grounding. Compared to OpenVLA [29], CoT-VLA shows slightly lower success rates in visual and language generalization tasks due to grasping failures from action chunking (see Section 5) rather than errors in visual reasoning. However, CoT-VLA demonstrates competitive performance across all four generalization categories overall, achieving comparable or better results to baseline approaches." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 558, + 557, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 557, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 557, + 713 + ], + "type": "text", + "content": "Franka-Tabletop We present quantitative results in Table 4 and example execution trajectories in Figure 5. In this experiment, models are fine-tuned on a relatively small set of demonstrations. While Diffusion Policy achieves top performance on single-instruction tasks (e.g., \"put corn in bowl\"), its performance degrades on multi-instruction tasks involving diverse objects and complex language instructions. Models pretrained on the OpenX dataset - Octo, OpenVLA, and CoT-VLA - demonstrate better adaptation and performance on multi-instruction tasks where language grounding is critical. Overall, CoT-VLA achieves the highest average performance compared to baseline approaches, showing improvements in both single and multi-instruction scenarios." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1707" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 55, + 68, + 539, + 398 + ], + "blocks": [ + { + "bbox": [ + 55, + 68, + 539, + 398 + ], + "lines": [ + { + "bbox": [ + 55, + 68, + 539, + 398 + ], + "spans": [ + { + "bbox": [ + 55, + 68, + 539, + 398 + ], + "type": "image", + "image_path": "ddecb96de6c4fe2fe9935672eab756d2ca0a6be1e313cef3f5f9502ad7b74dcd.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 409, + 555, + 453 + ], + "lines": [ + { + "bbox": [ + 55, + 409, + 555, + 453 + ], + "spans": [ + { + "bbox": [ + 55, + 409, + 555, + 453 + ], + "type": "text", + "content": "Figure 5. Task execution examples for LIBERO, Bridge-V2, and Franka-Tabletop using CoT-VLA. For each task: Left: text instruction " + }, + { + "bbox": [ + 55, + 409, + 555, + 453 + ], + "type": "inline_equation", + "content": "(l)" + }, + { + "bbox": [ + 55, + 409, + 555, + 453 + ], + "type": "text", + "content": " and initial state " + }, + { + "bbox": [ + 55, + 409, + 555, + 453 + ], + "type": "inline_equation", + "content": "(s_0^{\\mathrm{obs}})" + }, + { + "bbox": [ + 55, + 409, + 555, + 453 + ], + "type": "text", + "content": ". Middle: generated intermediate goal states " + }, + { + "bbox": [ + 55, + 409, + 555, + 453 + ], + "type": "inline_equation", + "content": "(\\hat{s}_t)" + }, + { + "bbox": [ + 55, + 409, + 555, + 453 + ], + "type": "text", + "content": " demonstrating visual chain-of-thought reasoning, where each goal image is conditioned on both the instruction and the most recent observation. Right: final state " + }, + { + "bbox": [ + 55, + 409, + 555, + 453 + ], + "type": "inline_equation", + "content": "(s_T^{\\mathrm{obs}})" + }, + { + "bbox": [ + 55, + 409, + 555, + 453 + ], + "type": "text", + "content": " upon task completion. Complete execution trajectories are available in the supplementary video." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 473, + 149, + 487 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 473, + 149, + 487 + ], + "spans": [ + { + "bbox": [ + 55, + 473, + 149, + 487 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 495, + 297, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 495, + 297, + 639 + ], + "spans": [ + { + "bbox": [ + 54, + 495, + 297, + 639 + ], + "type": "text", + "content": "Visual CoT, Hybrid Attention, and Action Chunking We conduct comprehensive ablation studies on two LIBERO benchmark suites: LIBERO-Spatial and LIBERO-Goal. We evaluate four model variants: VLA - a baseline implementation following the standard VLA framework [29], with the same VILA-U backbone but without chain-of-thought reasoning and action chunking; + action chunking - extending the vanilla VLA to predict action sequences of length " + }, + { + "bbox": [ + 54, + 495, + 297, + 639 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 54, + 495, + 297, + 639 + ], + "type": "text", + "content": "; + hybrid attention - further adding full attention mechanisms for action sequence prediction, as illustrated in Figure 3; and + CoT (ours): our complete approach with hybrid attention mechanism and visual chain-of-thought reasoning." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": "As shown in Figure 6, both benchmark suites demonstrate that action sequence prediction consistently outperforms single-action prediction. The addition of hybrid attention mechanisms further improves performance. Our CoT-VLA achieves the best results validating the effectiveness of visual chain-of-thought reasoning for VLA tasks." + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 318, + 472, + 552, + 545 + ], + "blocks": [ + { + "bbox": [ + 318, + 472, + 552, + 545 + ], + "lines": [ + { + "bbox": [ + 318, + 472, + 552, + 545 + ], + "spans": [ + { + "bbox": [ + 318, + 472, + 552, + 545 + ], + "type": "table", + "html": "
CategorySUSIEOctoOpenVLACoT-VLA
Visual30%35%75%65%
Motion10%10%45%60%
Semantic20%0%40%50%
Language40%40%75%70%
", + "image_path": "97606d9ff20fa34d7f742170b707e4341531fc2321f2afaf2e87d27bea623221.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 553, + 556, + 620 + ], + "lines": [ + { + "bbox": [ + 313, + 553, + 556, + 620 + ], + "spans": [ + { + "bbox": [ + 313, + 553, + 556, + 620 + ], + "type": "text", + "content": "Table 2. Bridge-V2 Comparison. Success rates across four generalization categories, with 10 trials per category and partial credit scoring following [29]. Visual: \"put eggplant into pot\" with cluttered environments; Motion: \"put carrot on plate\" with height variations; Semantic: \"take purple grapes out of pot\"; Language: \"put eggplant or red bottle into pot\"." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "text", + "content": "**Pretraining** Our training pipeline has two stages, pretraining VILA-U on the OpenX dataset augmented with actionless video data (Section 3.3), and task-specific post-training on robot demonstration data. To assess the importance of our pretraining stage, we conduct ablation studies on the Franka-Tabletop setup. We report the quantitative results" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1708" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 88, + 296, + 177 + ], + "blocks": [ + { + "bbox": [ + 63, + 77, + 73, + 87 + ], + "lines": [ + { + "bbox": [ + 63, + 77, + 73, + 87 + ], + "spans": [ + { + "bbox": [ + 63, + 77, + 73, + 87 + ], + "type": "text", + "content": "a)" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 129, + 68, + 245, + 79 + ], + "lines": [ + { + "bbox": [ + 129, + 68, + 245, + 79 + ], + "spans": [ + { + "bbox": [ + 129, + 68, + 245, + 79 + ], + "type": "text", + "content": "VLA + action chunking" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 59, + 88, + 296, + 177 + ], + "lines": [ + { + "bbox": [ + 59, + 88, + 296, + 177 + ], + "spans": [ + { + "bbox": [ + 59, + 88, + 296, + 177 + ], + "type": "image", + "image_path": "6ea0ad3b3ccb98fdac323ac664b2e908e72db861c0add6f21cf035b150c7659f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 63, + 196, + 293, + 236 + ], + "blocks": [ + { + "bbox": [ + 63, + 186, + 73, + 196 + ], + "lines": [ + { + "bbox": [ + 63, + 186, + 73, + 196 + ], + "spans": [ + { + "bbox": [ + 63, + 186, + 73, + 196 + ], + "type": "text", + "content": "b)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 63, + 196, + 293, + 236 + ], + "lines": [ + { + "bbox": [ + 63, + 196, + 293, + 236 + ], + "spans": [ + { + "bbox": [ + 63, + 196, + 293, + 236 + ], + "type": "image", + "image_path": "8b7eb700dc94010f650f5243af400c3b1badf4c194800db729df8fcc393cb9bc.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 247, + 296, + 315 + ], + "lines": [ + { + "bbox": [ + 55, + 247, + 296, + 315 + ], + "spans": [ + { + "bbox": [ + 55, + 247, + 296, + 315 + ], + "type": "text", + "content": "Figure 6. Ablation studies of CoT-VLA components. a) Results on LIBERO-Spatial and LIBERO-Goal benchmarks demonstrate the effectiveness of three components: action chunking, hybrid attention, and visual chain-of-thought reasoning. b) Pretraining ablation experiments on Franka-Tabletop show performance improvements from the OpenX and action-less video pretraining process." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 327, + 296, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 327, + 296, + 387 + ], + "spans": [ + { + "bbox": [ + 55, + 327, + 296, + 387 + ], + "type": "text", + "content": "in Figure 6. Our results show that CoT-VLA with our pretraining stage achieves a " + }, + { + "bbox": [ + 55, + 327, + 296, + 387 + ], + "type": "inline_equation", + "content": "46.7\\%" + }, + { + "bbox": [ + 55, + 327, + 296, + 387 + ], + "type": "text", + "content": " relative improvement, from " + }, + { + "bbox": [ + 55, + 327, + 296, + 387 + ], + "type": "inline_equation", + "content": "53.7\\%" + }, + { + "bbox": [ + 55, + 327, + 296, + 387 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 327, + 296, + 387 + ], + "type": "inline_equation", + "content": "78.8\\%" + }, + { + "bbox": [ + 55, + 327, + 296, + 387 + ], + "type": "text", + "content": ", compared to directly fine-tuning the base VILA-U model on Franka-Tabletop demonstrations, demonstrating better downstream task adaptation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 396, + 220, + 408 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 396, + 220, + 408 + ], + "spans": [ + { + "bbox": [ + 55, + 396, + 220, + 408 + ], + "type": "text", + "content": "4.4. Better Visual Reasoning Helps" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 415, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 415, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 415, + 296, + 714 + ], + "type": "text", + "content": "Unlike prior VLAs that only use robot demonstration data " + }, + { + "bbox": [ + 55, + 415, + 296, + 714 + ], + "type": "inline_equation", + "content": "D_{\\mathrm{r}}" + }, + { + "bbox": [ + 55, + 415, + 296, + 714 + ], + "type": "text", + "content": " during training, CoT-VLA also leverages action-less video data " + }, + { + "bbox": [ + 55, + 415, + 296, + 714 + ], + "type": "inline_equation", + "content": "D_{\\mathrm{v}}" + }, + { + "bbox": [ + 55, + 415, + 296, + 714 + ], + "type": "text", + "content": " for pretraining through its intermediate visual chain-of-thought reasoning steps. This enables learning of both dynamics and instruction following from captioned videos alone, which are substantially more abundant than robot demonstrations. To investigate how visual reasoning capabilities transfer to robot performance, we conduct an ablation study on the Franka-Tabletop setup using novel, long-horizon tasks that combine two unseen subtasks. We design two tasks – (1) “move the green scallion to the apple-covered book” and (2) “move the green cauliflower to the bear-covered book” – that are challenging for our model’s out-of-distribution generalization. For each task, we collect one demonstration trajectory to obtain ground-truth goal images. We evaluate each task across 5 trials under two conditions: (1) CoT-VLA using its generated goal images and (2) CoT-VLA using ground-truth goal images from the collected demonstrations. As shown in Table 3, using ground-truth goal images improves the absolute success rate by " + }, + { + "bbox": [ + 55, + 415, + 296, + 714 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 55, + 415, + 296, + 714 + ], + "type": "text", + "content": " for both tasks. This performance boost suggests that advances in visual reasoning and goal image generation could directly translate to better robotic task performance. While our method still struggles with out-of-distribution subgoal generation, recent advances in large-scale video and image" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "type": "text", + "content": "models show promising directions for improving visual reasoning capabilities with scaling." + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 319, + 104, + 550, + 153 + ], + "blocks": [ + { + "bbox": [ + 319, + 104, + 550, + 153 + ], + "lines": [ + { + "bbox": [ + 319, + 104, + 550, + 153 + ], + "spans": [ + { + "bbox": [ + 319, + 104, + 550, + 153 + ], + "type": "table", + "html": "
Sub-task 1Sub-task 2
Generated Goal Images20%0%
Ground-truth Goal Images60%40%
", + "image_path": "9c04b07ec2b2be0db2c9b1d8532eff46a04cfef1d5cd1ca53de7ccf25d25a5d7.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 162, + 555, + 228 + ], + "lines": [ + { + "bbox": [ + 313, + 162, + 555, + 228 + ], + "spans": [ + { + "bbox": [ + 313, + 162, + 555, + 228 + ], + "type": "text", + "content": "Table 3. Better visual reasoning helps. Success rates comparing CoT-VLA using generated versus ground-truth goal images on out-of-distribution tasks. Results demonstrate that improved visual reasoning (simulated by ground-truth goals) leads to better task performance, suggesting that advances in goal generation can translate to improved action execution." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 239, + 544, + 251 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 239, + 544, + 251 + ], + "spans": [ + { + "bbox": [ + 313, + 239, + 544, + 251 + ], + "type": "text", + "content": "5. Conclusion, Limitations and Future Work" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 258, + 555, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 258, + 555, + 366 + ], + "spans": [ + { + "bbox": [ + 313, + 258, + 555, + 366 + ], + "type": "text", + "content": "In this work, we introduce CoT-VLA, bridging vision-language-action models with chain-of-thought reasoning by introducing intermediate visual goals as explicit reasoning steps. Rather than using abstract representations like bounding boxes or keypoints, we propose using subgoal images sampled from videos as an interpretable and effective intermediate representation. We build our system upon VILA-U, demonstrating strong performance across diverse robotic manipulation tasks." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 367, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 367, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 367, + 556, + 714 + ], + "type": "text", + "content": "While our approach demonstrates effectiveness, there are certain limitations. First, generating intermediate image tokens during inference introduces significant computational overhead compared to direct action generation approaches. Our method requires generating 256 image tokens before action tokens, leading to a " + }, + { + "bbox": [ + 313, + 367, + 556, + 714 + ], + "type": "inline_equation", + "content": "7 \\times" + }, + { + "bbox": [ + 313, + 367, + 556, + 714 + ], + "type": "text", + "content": " slowdown on average with an action chunk size of 10. While action chunking and parallel decoding improve inference speed, image generation remains the primary bottleneck. Recent advancement in fast image generation or fast LLM inference techniques could potentially improve the throughput of the model [7, 31, 33, 57, 73] and be integrated into our system. Second, our autoregressive image generation produces lower visual quality compared to state-of-the-art diffusion-based models. Recent advances in unified multimodal models [61, 65, 69, 79] suggest promising directions for improvements. Additionally, while effective, our action chunking approach can introduce discontinuous actions between chunks and lacks high-frequency feedback during execution. These limitations could be addressed through temporal smoothing techniques and per-step prediction approaches similar to those proposed in [10]. Finally, while CoT-VLA leverages action-less video data during pretraining, current computational constraints limit its ability to achieve visual-reasoning generalization for entirely new tasks. Looking forward, we believe recent advances in video/image generation and world models [15, 23, 30, 68, 72] present promising opportunities to enhance generalization capabilities through improved visual reasoning and predictive modeling." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1709" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 297, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 91, + 297, + 156 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 297, + 156 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 297, + 156 + ], + "type": "text", + "content": "[1] Homanga Bharadhwaj, Debidatta Dwibedi, Abhinav Gupta, Shubham Tulsiani, Carl Doersch, Ted Xiao, Dhruv Shah, Fei Xia, Dorsa Sadigh, and Sean Kirmani. Gen2act: Human video generation in novel scenarios enables generalizable robot manipulation. arXiv preprint arXiv:2409.16283, 2024. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 158, + 296, + 211 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 158, + 296, + 211 + ], + "spans": [ + { + "bbox": [ + 61, + 158, + 296, + 211 + ], + "type": "text", + "content": "[2] Kevin Black, Mitsuhiko Nakamoto, Pranav Atreya, Homer Walke, Chelsea Finn, Aviral Kumar, and Sergey Levine. Zero-shot robotic manipulation with pretrained image-editing diffusion models. arXiv preprint arXiv:2310.10639, 2023. 2, 3, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 214, + 296, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 214, + 296, + 267 + ], + "spans": [ + { + "bbox": [ + 62, + 214, + 296, + 267 + ], + "type": "text", + "content": "[3] Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Joseph Dabis, Chelsea Finn, Keerthana Gopalakrishnan, Karol Hausman, Alex Herzog, Jasmine Hsu, et al. Rt-1: Robotics transformer for real-world control at scale. arXiv preprint arXiv:2212.06817, 2022. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 270, + 296, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 270, + 296, + 323 + ], + "spans": [ + { + "bbox": [ + 62, + 270, + 296, + 323 + ], + "type": "text", + "content": "[4] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 325, + 296, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 325, + 296, + 379 + ], + "spans": [ + { + "bbox": [ + 62, + 325, + 296, + 379 + ], + "type": "text", + "content": "[5] Chi-Lam Cheang, Guangzeng Chen, Ya Jing, Tao Kong, Hang Li, Yifeng Li, Yuxiao Liu, Hongtao Wu, Jiafeng Xu, Yichu Yang, et al. Gr-2: A generative video-language-action model with web-scale knowledge for robot manipulation. arXiv preprint arXiv:2410.06158, 2024. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 380, + 296, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 380, + 296, + 445 + ], + "spans": [ + { + "bbox": [ + 62, + 380, + 296, + 445 + ], + "type": "text", + "content": "[6] Boyuan Chen, Zhuo Xu, Sean Kirmani, Brain Ichter, Dorsa Sadigh, Leonidas Guibas, and Fei Xia. Spatialvlm: Endowing vision-language models with spatial reasoning capabilities. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14455-14465, 2024. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 447, + 296, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 447, + 296, + 490 + ], + "spans": [ + { + "bbox": [ + 62, + 447, + 296, + 490 + ], + "type": "text", + "content": "[7] Junyu Chen, Han Cai, Junsong Chen, Enze Xie, Shang Yang, Haotian Tang, Muyang Li, Yao Lu, and Song Han. Deep compression autoencoder for efficient high-resolution diffusion models. arXiv preprint arXiv:2410.10733, 2024. 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 491, + 296, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 491, + 296, + 534 + ], + "spans": [ + { + "bbox": [ + 62, + 491, + 296, + 534 + ], + "type": "text", + "content": "[8] Lawrence Yunliang Chen, Simeon Adebola, and Ken Goldberg. Berkeley UR5 demonstration dataset. https:// sites.google.com/view/berkeley-ur5/home. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 536, + 296, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 536, + 296, + 590 + ], + "spans": [ + { + "bbox": [ + 62, + 536, + 296, + 590 + ], + "type": "text", + "content": "[9] Xi Chen, Xiao Wang, Lucas Beyer, Alexander Kolesnikov, Jialin Wu, Paul Voigtlaender, Basil Mustafa, Sebastian Goodman, Ibrahim Alabdulmohsin, Piotr Padlewski, et al. Pali-3 vision language models: Smaller, faster, stronger. arXiv preprint arXiv:2310.09199, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 591, + 296, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 591, + 296, + 645 + ], + "spans": [ + { + "bbox": [ + 57, + 591, + 296, + 645 + ], + "type": "text", + "content": "[10] Cheng Chi, Zhenjia Xu, Siyuan Feng, Eric Cousineau, Yilun Du, Benjamin Burchfiel, Russ Tedrake, and Shuran Song. Diffusion policy: Visuomotor policy learning via action diffusion. The International Journal of Robotics Research, page 02783649241273668, 2023. 2, 6, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 647, + 294, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 294, + 680 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 294, + 680 + ], + "type": "text", + "content": "[11] Yiming Ding, Carlos Florensa, Pieter Abbeel, and Mariano Pielipp. Goal-conditioned imitation learning. Advances in neural information processing systems, 32, 2019. 2, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 681, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 681, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 681, + 296, + 713 + ], + "type": "text", + "content": "[12] Danny Driess, Fei Xia, Mehdi SM Sajjadi, Corey Lynch, Aakanksha Chowdhery, Brian Ichter, Ayzaan Wahid, Jonathan Thompson, Quan Vuong, Tianhe Yu, et al. Palm-e: An" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 333, + 73, + 555, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 555, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 555, + 95 + ], + "type": "text", + "content": "embodied multimodal language model. arXiv preprint arXiv:2303.03378, 2023.1,3,4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 96, + 555, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 555, + 139 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 555, + 139 + ], + "type": "text", + "content": "[13] Yuqing Du, Ksenia Konyushkova, Misha Denil, Akhil Raju, Jessica Landon, Felix Hill, Nando de Freitas, and Serkan Cabi. Vision-language models as success detectors. arXiv preprint arXiv:2303.07280, 2023. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 140, + 555, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 140, + 555, + 183 + ], + "spans": [ + { + "bbox": [ + 317, + 140, + 555, + 183 + ], + "type": "text", + "content": "[14] Yilun Du, Mengjiao Yang, Pete Florence, Fei Xia, Ayzaan Wahid, Brian Ichter, Pierre Sermanet, Tianhe Yu, Pieter Abbeel, Joshua B Tenenbaum, et al. Video language planning. arXiv preprint arXiv:2310.10625, 2023. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 185, + 555, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 185, + 555, + 239 + ], + "spans": [ + { + "bbox": [ + 317, + 185, + 555, + 239 + ], + "type": "text", + "content": "[15] Yilun Du, Sherry Yang, Bo Dai, Hanjun Dai, Ofir Nachum, Josh Tenenbaum, Dale Schuurmans, and Pieter Abbeel. Learning universal policies via text-guided video generation. Advances in Neural Information Processing Systems, 36, 2024. 2, 3, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 240, + 555, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 240, + 555, + 293 + ], + "spans": [ + { + "bbox": [ + 316, + 240, + 555, + 293 + ], + "type": "text", + "content": "[16] Frederik Ebert, Yanlai Yang, Karl Schmeckpeper, Bernadette Bucher, Georgios Georgakis, Kostas Daniilidis, Chelsea Finn, and Sergey Levine. Bridge data: Boosting generalization of robotic skills with cross-domain datasets. arXiv preprint arXiv:2109.13396, 2021. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 295, + 555, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 295, + 555, + 336 + ], + "spans": [ + { + "bbox": [ + 316, + 295, + 555, + 336 + ], + "type": "text", + "content": "[17] Zipeng Fu, Qingqing Zhao, Qi Wu, Gordon Wetzstein, and Chelsea Finn. Humanplus: Humanoid shadowing and imitation from humans. arXiv preprint arXiv:2406.10454, 2024. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 338, + 555, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 338, + 555, + 380 + ], + "spans": [ + { + "bbox": [ + 317, + 338, + 555, + 380 + ], + "type": "text", + "content": "[18] Zipeng Fu, Tony Z. Zhao, and Chelsea Finn. Mobile aloha: Learning bimanual mobile manipulation with low-cost whole-body teleoperation. In Conference on Robot Learning (CoRL), 2024. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 383, + 555, + 447 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 383, + 555, + 447 + ], + "spans": [ + { + "bbox": [ + 317, + 383, + 555, + 447 + ], + "type": "text", + "content": "[19] Samir Yitzhak Gadre, Mitchell Wortsman, Gabriel Ilharco, Ludwig Schmidt, and Shuran Song. Cows on pasture: Baselines and benchmarks for language-driven zero-shot object navigation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23171-23181, 2023. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 449, + 555, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 449, + 555, + 525 + ], + "spans": [ + { + "bbox": [ + 317, + 449, + 555, + 525 + ], + "type": "text", + "content": "[20] Raghav Goyal, Samira Ebrahimi Kahou, Vincent Michalski, Joanna Materzynska, Susanne Westphal, Heuna Kim, Valentin Haenel, Ingo Fruend, Peter Yianilos, Moritz Mueller-Freitag, et al. The\" something something\" video database for learning and evaluating visual common sense. In Proceedings of the IEEE international conference on computer vision, pages 5842-5850, 2017. 2, 3, 4, 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 526, + 555, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 526, + 555, + 568 + ], + "spans": [ + { + "bbox": [ + 317, + 526, + 555, + 568 + ], + "type": "text", + "content": "[21] Huy Ha, Pete Florence, and Shuran Song. Scaling up and distilling down: Language-guided robot skill acquisition. In Conference on Robot Learning, pages 3766-3777. PMLR, 2023. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 571, + 555, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 571, + 555, + 591 + ], + "spans": [ + { + "bbox": [ + 317, + 571, + 555, + 591 + ], + "type": "text", + "content": "[22] William Harvey and Frank Wood. Visual chain-of-thought diffusion models. arXiv preprint arXiv:2303.16187, 2023. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 593, + 555, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 593, + 555, + 635 + ], + "spans": [ + { + "bbox": [ + 317, + 593, + 555, + 635 + ], + "type": "text", + "content": "[23] Anthony Hu, Lloyd Russell, Hudson Yeo, Zak Murez, George Fedoseev, Alex Kendall, Jamie Shotton, and Gianluca Corrado. Gaia-1: A generative world model for autonomous driving. arXiv preprint arXiv:2309.17080, 2023. 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 637, + 555, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 637, + 555, + 689 + ], + "spans": [ + { + "bbox": [ + 317, + 637, + 555, + 689 + ], + "type": "text", + "content": "[24] Yushi Hu, Weijia Shi, Xingyu Fu, Dan Roth, Mari Ostendorf, Luke Zettlemoyer, Noah A Smith, and Ranjay Krishna. Visual sketchpad: Sketching as a visual chain of thought for multimodal language models. arXiv preprint arXiv:2406.09403, 2024. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 692, + 555, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 555, + 713 + ], + "type": "text", + "content": "[25] Chenguang Huang, Oier Mees, Andy Zeng, and Wolfram Burgard. Visual language maps for robot navigation. In 2023" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1710" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "text", + "content": "IEEE International Conference on Robotics and Automation (ICRA), pages 10608-10615. IEEE, 2023. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 140 + ], + "type": "text", + "content": "[26] Wenlong Huang, Chen Wang, Yunzhu Li, Ruohan Zhang, and Li Fei-Fei. Rekep: Spatio-temporal reasoning of relational keypoint constraints for robotic manipulation. arXiv preprint arXiv:2409.01652, 2024. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 142, + 295, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 142, + 295, + 239 + ], + "spans": [ + { + "bbox": [ + 56, + 142, + 295, + 239 + ], + "type": "text", + "content": "[27] Georgios Kapidis, Ronald Poppe, Elsbeth Van Dam, Lucas Noldus, and Remco Veltkamp. Egocentric hand track and object-based human action recognition. In 2019 IEEE SmartWorld, Ubiquitous Intelligence & Computing, Advanced & Trusted Computing, Scalable Computing & Communications, Cloud & Big Data Computing, Internet of People and Smart City Innovation (SmartWorld/SCALCOM/UIC/ATC/CBDCom/IOP/SCI), pages 922-929. IEEE, 2019. 1, 2, 3, 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 241, + 295, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 241, + 295, + 285 + ], + "spans": [ + { + "bbox": [ + 56, + 241, + 295, + 285 + ], + "type": "text", + "content": "[28] Siddharth Karamcheti, Suraj Nair, Ashwin Balakrishna, Percy Liang, Thomas Kollar, and Dorsa Sadigh. Prismatic vlms: Investigating the design space of visually-conditioned language models. arXiv preprint arXiv:2402.07865, 2024. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 286, + 295, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 286, + 295, + 341 + ], + "spans": [ + { + "bbox": [ + 56, + 286, + 295, + 341 + ], + "type": "text", + "content": "[29] Moo Jin Kim, Karl Pertsch, Siddharth Karamcheti, Ted Xiao, Ashwin Balakrishna, Suraj Nair, Rafael Rafailov, Ethan Foster, Grace Lam, Pannag Sanketi, et al. Openvla: An open-source vision-language-action model. arXiv preprint arXiv:2406.09246, 2024. 1, 3, 4, 5, 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 342, + 295, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 342, + 295, + 397 + ], + "spans": [ + { + "bbox": [ + 56, + 342, + 295, + 397 + ], + "type": "text", + "content": "[30] Dan Kondratyuk, Lijun Yu, Xiuye Gu, José Lezama, Jonathan Huang, Grant Schindler, Rachel Hornung, Vighnesh Birodkar, Jimmy Yan, Ming-Chang Chiu, et al. Videopoet: A large language model for zero-shot video generation. arXiv preprint arXiv:2312.14125, 2023. 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 399, + 295, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 399, + 295, + 465 + ], + "spans": [ + { + "bbox": [ + 56, + 399, + 295, + 465 + ], + "type": "text", + "content": "[31] Woosuk Kwon, Zhuohan Li, Siyuan Zhuang, Ying Sheng, Lianmin Zheng, Cody Hao Yu, Joseph E. Gonzalez, Hao Zhang, and Ion Stoica. Efficient memory management for large language model serving with pagedattention. In Proceedings of the ACM SIGOPS 29th Symposium on Operating Systems Principles, 2023. 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 466, + 295, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 466, + 295, + 520 + ], + "spans": [ + { + "bbox": [ + 56, + 466, + 295, + 520 + ], + "type": "text", + "content": "[32] Doyup Lee, Chiheon Kim, Saehoon Kim, Minsu Cho, and Wook-Shin Han. Autoregressive image generation using residual quantization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11523-11532, 2022. 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 522, + 295, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 522, + 295, + 566 + ], + "spans": [ + { + "bbox": [ + 56, + 522, + 295, + 566 + ], + "type": "text", + "content": "[33] Yaniv Leviathan, Matan Kalman, and Yossi Matias. Fast inference from transformers via speculative decoding. In International Conference on Machine Learning, pages 19274-19286. PMLR, 2023. 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 567, + 295, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 567, + 295, + 623 + ], + "spans": [ + { + "bbox": [ + 56, + 567, + 295, + 623 + ], + "type": "text", + "content": "[34] Boyi Li, Yue Wang, Jiageng Mao, Boris Ivanovic, Sushant Veer, Karen Leung, and Marco Pavone. Driving everywhere with large language model policy adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14948-14957, 2024. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 624, + 295, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 295, + 668 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 295, + 668 + ], + "type": "text", + "content": "[35] Junbang Liang, Ruoshi Liu, Ege Ozguroglu, Sruthi Sudhakar, Achal Dave, Pavel Tokmakov, Shuran Song, and Carl Vondrick. Dreamitate: Real-world visuomotor policy learning via video generation. arXiv preprint arXiv:2406.16862, 2024. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 669, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 669, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 669, + 295, + 712 + ], + "type": "text", + "content": "[36] Fanqi Lin, Yingdong Hu, Pingyue Sheng, Chuan Wen, Jiacheng You, and Yang Gao. Data scaling laws in imitation learning for robotic manipulation. arXiv preprint arXiv:2410.18647, 2024. 1" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 554, + 712 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 316, + 72, + 554, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 554, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 554, + 116 + ], + "type": "text", + "content": "[37] Bo Liu, Yifeng Zhu, Chongkai Gao, Yihao Feng, Qiang Liu, Yuke Zhu, and Peter Stone. Libero: Benchmarking knowledge transfer for lifelong robot learning. arXiv preprint arXiv:2306.03310, 2023. 2, 5, 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 118, + 554, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 554, + 151 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 554, + 151 + ], + "type": "text", + "content": "[38] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 152, + 554, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 554, + 218 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 554, + 218 + ], + "type": "text", + "content": "[39] Jiasen Lu, Christopher Clark, Sangho Lee, Zichen Zhang, Savya Khosla, Ryan Marten, Derek Hoiem, and Aniruddha Kembhavi. Unified-io 2: Scaling autoregressive multimodal models with vision language audio and action. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26439-26455, 2024. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 220, + 554, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 220, + 554, + 274 + ], + "spans": [ + { + "bbox": [ + 316, + 220, + 554, + 274 + ], + "type": "text", + "content": "[40] Yecheng Jason Ma, William Liang, Guanzhi Wang, De-An Huang, Osbert Bastani, Dinesh Jayaraman, Yuke Zhu, Linxi Fan, and Anima Anandkumar. Eureka: Human-level reward design via coding large language models. arXiv preprint arXiv:Arxiv-2310.12931, 2023.3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 276, + 554, + 340 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 276, + 554, + 340 + ], + "spans": [ + { + "bbox": [ + 316, + 276, + 554, + 340 + ], + "type": "text", + "content": "[41] Arjun Majumdar, Karmesh Yadav, Sergio Arnaud, Jason Ma, Claire Chen, Sneha Silwal, Aryan Jain, Vincent-Pierre Berges, Tingfan Wu, Jay Vakil, et al. Where are we in the search for an artificial visual cortex for embodied intelligence? Advances in Neural Information Processing Systems, 36:655-677, 2023. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 342, + 554, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 342, + 554, + 408 + ], + "spans": [ + { + "bbox": [ + 316, + 342, + 554, + 408 + ], + "type": "text", + "content": "[42] Ajay Mandlekar, Yuke Zhu, Animesh Garg, Jonathan Booher, Max Spero, Albert Tung, Julian Gao, John Emmons, Anchit Gupta, Emre Orbay, et al. Roboturk: A crowdsourcing platform for robotic skill learning through imitation. In Conference on Robot Learning, pages 879-893. PMLR, 2018. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 411, + 554, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 411, + 554, + 464 + ], + "spans": [ + { + "bbox": [ + 316, + 411, + 554, + 464 + ], + "type": "text", + "content": "[43] Oier Mees, Jessica Borja-Diaz, and Wolfram Burgard. Grounding language with visual affordances over unstructured data. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 11576-11582. IEEE, 2023. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 466, + 554, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 466, + 554, + 510 + ], + "spans": [ + { + "bbox": [ + 316, + 466, + 554, + 510 + ], + "type": "text", + "content": "[44] Zawalski Michal, Chen William, Pertsch Karl, Mees Oier, Finn Chelsea, and Levine Sergey. Robotic control via embodied chain-of-thought reasoning. arXiv preprint arXiv:2407.08693, 2024. 1, 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 512, + 554, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 512, + 554, + 566 + ], + "spans": [ + { + "bbox": [ + 316, + 512, + 554, + 566 + ], + "type": "text", + "content": "[45] Yao Mu, Qinglong Zhang, Mengkang Hu, Wenhai Wang, Mingyu Ding, Jun Jin, Bin Wang, Jifeng Dai, Yu Qiao, and Ping Luo. Embodiedgpt: Vision-language pre-training via embodied chain of thought. Advances in Neural Information Processing Systems, 36, 2024. 1, 2, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 567, + 554, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 567, + 554, + 612 + ], + "spans": [ + { + "bbox": [ + 316, + 567, + 554, + 612 + ], + "type": "text", + "content": "[46] Ashvin V Nair, Vitchyr Pong, Murtaza Dalal, Shikhar Bahl, Steven Lin, and Sergey Levine. Visual reinforcement learning with imagined goals. Advances in neural information processing systems, 31, 2018. 2, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 613, + 554, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 613, + 554, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 613, + 554, + 689 + ], + "type": "text", + "content": "[47] Fei Ni, Jianye Hao, Shiguang Wu, Longxin Kou, Jiashun Liu, Yan Zheng, Bin Wang, and Yuzheng Zhuang. Generate subgoal images before act: Unlocking the chain-of-thought reasoning in diffusion model for robot manipulation with multimodal prompts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 2, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 691, + 554, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 554, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 554, + 712 + ], + "type": "text", + "content": "[48] Abby O'Neill, Abdul Rehman, Abhinav Gupta, Abhiram Maddukuri, Abhishek Gupta, Abhishek Padalkar, Abraham" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 749, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 749, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 749, + 314, + 757 + ], + "type": "text", + "content": "1711" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 76, + 72, + 297, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 297, + 106 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 297, + 106 + ], + "type": "text", + "content": "Lee, Acorn Pooley, Agrim Gupta, Ajay Mandlekar, et al. Open x-embodiment: Robotic learning datasets and rt-x models. arXiv preprint arXiv:2310.08864, 2023. 1, 2, 3, 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 108, + 296, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 108, + 296, + 173 + ], + "spans": [ + { + "bbox": [ + 56, + 108, + 296, + 173 + ], + "type": "text", + "content": "[49] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 175, + 296, + 231 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 296, + 231 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 296, + 231 + ], + "type": "text", + "content": "[50] Daniel Rose, Vaishnavi Himakunthala, Andy Ouyang, Ryan He, Alex Mei, Yujie Lu, Michael Saxon, Chinmay Sonar, Diba Mirza, and William Yang Wang. Visual chain of thought: bridging logical gaps with multimodal infillings. arXiv preprint arXiv:2305.02317, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 232, + 296, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 232, + 296, + 277 + ], + "spans": [ + { + "bbox": [ + 56, + 232, + 296, + 277 + ], + "type": "text", + "content": "[51] Erick Rosete-Beas, Oier Mees, Gabriel Kalweit, Joschka Boedecker, and Wolfram Burgard. Latent plans for task-agnostic offline reinforcement learning. In Conference on Robot Learning, pages 1838–1849. PMLR, 2023. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 278, + 296, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 278, + 296, + 311 + ], + "spans": [ + { + "bbox": [ + 56, + 278, + 296, + 311 + ], + "type": "text", + "content": "[52] V Sanh. Distilbert, a distilled version of bert: smaller, faster, cheaper and lighter. arXiv preprint arXiv:1910.01108, 2019. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 313, + 296, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 313, + 296, + 357 + ], + "spans": [ + { + "bbox": [ + 56, + 313, + 296, + 357 + ], + "type": "text", + "content": "[53] Hao Shao, Shengju Qian, Han Xiao, Guanglu Song, Zhuofan Zong, Letian Wang, Yu Liu, and Hongsheng Li. Visual cot: Unleashing chain-of-thought reasoning in multi-modal language models. arXiv preprint arXiv:2403.16999, 2024. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 358, + 296, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 358, + 296, + 392 + ], + "spans": [ + { + "bbox": [ + 56, + 358, + 296, + 392 + ], + "type": "text", + "content": "[54] Mohit Shridhar, Lucas Manuelli, and Dieter Fox. CIport: What and where pathways for robotic manipulation. In Conference on robot learning, pages 894-906. PMLR, 2022. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 394, + 296, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 394, + 296, + 426 + ], + "spans": [ + { + "bbox": [ + 56, + 394, + 296, + 426 + ], + "type": "text", + "content": "[55] Mohit Shridhar, Yat Long Lo, and Stephen James. Generative image as action models. arXiv preprint arXiv:2407.07875, 2024. 2, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 429, + 296, + 495 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 429, + 296, + 495 + ], + "spans": [ + { + "bbox": [ + 56, + 429, + 296, + 495 + ], + "type": "text", + "content": "[56] Ishika Singh, Valts Blukis, Arsalan Mousavian, Ankit Goyal, Danfei Xu, Jonathan Tremblay, Dieter Fox, Jesse Thomason, and Animesh Garg. Progprompt: Generating situated robot task plans using large language models. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 11523-11530. IEEE, 2023. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 497, + 296, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 497, + 296, + 529 + ], + "spans": [ + { + "bbox": [ + 56, + 497, + 296, + 529 + ], + "type": "text", + "content": "[57] Yang Song, Prafulla Dhariwal, Mark Chen, and Ilya Sutskever. Consistency models. arXiv preprint arXiv:2303.01469, 2023.8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 532, + 296, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 532, + 296, + 563 + ], + "spans": [ + { + "bbox": [ + 56, + 532, + 296, + 563 + ], + "type": "text", + "content": "[58] Chameleon Team. Chameleon: Mixed-modal early-fusion foundation models. arXiv preprint arXiv:2405.09818, 2024. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 566, + 296, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 566, + 296, + 610 + ], + "spans": [ + { + "bbox": [ + 56, + 566, + 296, + 610 + ], + "type": "text", + "content": "[59] Octo Model Team, Dibya Ghosh, Homer Walke, Karl Pertsch, Kevin Black, Oier Mees, Sudeep Dasari, Joel Hejna, Tobias Kreiman, Charles Xu, et al. Octo: An open-source generalist robot policy. arXiv preprint arXiv:2405.12213, 2024. 1, 3, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 613, + 296, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 613, + 296, + 668 + ], + "spans": [ + { + "bbox": [ + 56, + 613, + 296, + 668 + ], + "type": "text", + "content": "[60] Homer Rich Walke, Kevin Black, Tony Z Zhao, Quan Vuong, Chongyi Zheng, Philippe Hansen-Estruch, Andre Wang He, Vivek Myers, Moo Jin Kim, Max Du, et al. Bridgedata v2: A dataset for robot learning at scale. In Conference on Robot Learning, pages 1723-1736. PMLR, 2023. 2, 5, 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 670, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 296, + 713 + ], + "type": "text", + "content": "[61] Xinlong Wang, Xiaosong Zhang, Zhengxiong Luo, Quan Sun, Yufeng Cui, Jinsheng Wang, Fan Zhang, Yueze Wang, Zhen Li, Qiying Yu, et al. Emu3: Next-token prediction is all you need. arXiv preprint arXiv:2409.18869, 2024. 2, 8" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 316, + 72, + 555, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 555, + 126 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 555, + 126 + ], + "type": "text", + "content": "[62] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 2022. 2, 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 129, + 555, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 555, + 171 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 555, + 171 + ], + "type": "text", + "content": "[63] Chuan Wen, Xingyu Lin, John So, Kai Chen, Qi Dou, Yang Gao, and Pieter Abbeel. Any-point trajectory modeling for policy learning. arXiv preprint arXiv:2401.00025, 2023. 1, 2, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 173, + 555, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 173, + 555, + 227 + ], + "spans": [ + { + "bbox": [ + 316, + 173, + 555, + 227 + ], + "type": "text", + "content": "[64] Junjie Wen, Yichen Zhu, Jinming Li, Minjie Zhu, Kun Wu, Zhiyuan Xu, Ran Cheng, Chaomin Shen, Yaxin Peng, Feifei Feng, et al. Tinyvla: Towards fast, data-efficient vision-language-action models for robotic manipulation. arXiv preprint arXiv:2409.12514, 2024.3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 228, + 555, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 228, + 555, + 282 + ], + "spans": [ + { + "bbox": [ + 316, + 228, + 555, + 282 + ], + "type": "text", + "content": "[65] Chengyue Wu, Xiaokang Chen, Zhiyu Wu, Yiyang Ma, Xingchao Liu, Zizheng Pan, Wen Liu, Zhenda Xie, Xingkai Yu, Chong Ruan, et al. Janus: Decoupling visual encoding for unified multimodal understanding and generation. arXiv preprint arXiv:2410.13848, 2024. 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 283, + 555, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 283, + 555, + 337 + ], + "spans": [ + { + "bbox": [ + 316, + 283, + 555, + 337 + ], + "type": "text", + "content": "[66] Hongtao Wu, Ya Jing, Chilam Cheang, Guangzeng Chen, Jiafeng Xu, Xinghang Li, Minghuan Liu, Hang Li, and Tao Kong. Unleashing large-scale video generative pretraining for visual robot manipulation. arXiv preprint arXiv:2312.13139, 2023. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 338, + 555, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 338, + 555, + 392 + ], + "spans": [ + { + "bbox": [ + 316, + 338, + 555, + 392 + ], + "type": "text", + "content": "[67] Yecheng Wu, Zhuoyang Zhang, Junyu Chen, Haotian Tang, Dacheng Li, Yunhao Fang, Ligeng Zhu, Enze Xie, Hongxu Yin, Li Yi, et al. Vila-u: a unified foundation model integrating visual understanding and generation. arXiv preprint arXiv:2409.04429, 2024. 2, 3, 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 393, + 555, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 393, + 555, + 446 + ], + "spans": [ + { + "bbox": [ + 316, + 393, + 555, + 446 + ], + "type": "text", + "content": "[68] Jiannan Xiang, Guangyi Liu, Yi Gu, Qiyue Gao, Yuting Ning, Yuheng Zha, Zeyu Feng, Tianhua Tao, Shibo Hao, Yemin Shi, et al. Pandora: Towards general world model with natural language actions and video states. arXiv preprint arXiv:2406.09455, 2024. 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 449, + 555, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 449, + 555, + 502 + ], + "spans": [ + { + "bbox": [ + 316, + 449, + 555, + 502 + ], + "type": "text", + "content": "[69] Jinheng Xie, Weijia Mao, Zechen Bai, David Junhao Zhang, Weihao Wang, Kevin Qinghong Lin, Yuchao Gu, Zhijie Chen, Zhenheng Yang, and Mike Zheng Shou. Show-o: One single transformer to unify multimodal understanding and generation. arXiv preprint arXiv:2408.12528, 2024. 2, 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 504, + 555, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 504, + 555, + 557 + ], + "spans": [ + { + "bbox": [ + 316, + 504, + 555, + 557 + ], + "type": "text", + "content": "[70] Jonathan Yang, Catherine Glossop, Arjun Bhorkar, Dhruv Shah, Quan Vuong, Chelsea Finn, Dorsa Sadigh, and Sergey Levine. Pushing the limits of cross-embodiment learning for manipulation and navigation. arXiv preprint arXiv:2402.19432, 2024. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 559, + 555, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 559, + 555, + 602 + ], + "spans": [ + { + "bbox": [ + 316, + 559, + 555, + 602 + ], + "type": "text", + "content": "[71] Mengjiao Yang, Yilun Du, Kamyar Ghasemipour, Jonathan Thompson, Dale Schuurmans, and Pieter Abbeel. Learning interactive real-world simulators. arXiv preprint arXiv:2310.06114, 2023. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 603, + 555, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 603, + 555, + 647 + ], + "spans": [ + { + "bbox": [ + 316, + 603, + 555, + 647 + ], + "type": "text", + "content": "[72] Sherry Yang, Jacob Walker, Jack Parker-Holder, Yilun Du, Jake Bruce, Andre Barreto, Pieter Abbeel, and Dale Schuurmans. Video as the new language for real-world decision making. arXiv preprint arXiv:2402.17139, 2024. 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 647, + 555, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 555, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 555, + 689 + ], + "type": "text", + "content": "[73] Qihang Yu, Mark Weber, Xueqing Deng, Xiaohui Shen, Daniel Cremers, and Liang-Chieh Chen. An image is worth 32 tokens for reconstruction and generation. arxiv: 2406.07550, 2024.8" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 691, + 555, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 555, + 713 + ], + "type": "text", + "content": "[74] Wenhao Yu, Nimrod Gileadi, Chuyuan Fu, Sean Kirmani, Kuang-Huei Lee, Montse Gonzalez Arenas, Hao-Tien Lewis" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1712" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 553 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 76, + 72, + 296, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 296, + 106 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 296, + 106 + ], + "type": "text", + "content": "Chiang, Tom Erez, Leonard Hasenclever, Jan Humplik, et al. Language to rewards for robotic skill synthesis. arXiv preprint arXiv:2306.08647, 2023. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 296, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 296, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 296, + 150 + ], + "type": "text", + "content": "[75] Eric Zelikman, Yuhuai Wu, Jesse Mu, and Noah Goodman. Star: Bootstrapping reasoning with reasoning. Advances in Neural Information Processing Systems, 35:15476-15488, 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 296, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 296, + 195 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 296, + 195 + ], + "type": "text", + "content": "[76] Kaifeng Zhang, Zhao-Heng Yin, Weirui Ye, and Yang Gao. Learning manipulation skills through robot chain-of-thought with sparse failure guidance. arXiv preprint arXiv:2405.13573, 2024. 1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 197, + 296, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 197, + 296, + 229 + ], + "spans": [ + { + "bbox": [ + 56, + 197, + 296, + 229 + ], + "type": "text", + "content": "[77] Tony Z Zhao, Vikash Kumar, Sergey Levine, and Chelsea Finn. Learning fine-grained bimanual manipulation with low-cost hardware. arXiv preprint arXiv:2304.13705, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 231, + 296, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 231, + 296, + 274 + ], + "spans": [ + { + "bbox": [ + 56, + 231, + 296, + 274 + ], + "type": "text", + "content": "[78] Haoyu Zhen, Xiaowen Qiu, Peihao Chen, Jincheng Yang, Xin Yan, Yilun Du, Yining Hong, and Chuang Gan. 3d-vla: A 3d vision-language-action generative world model. arXiv preprint arXiv:2403.09631, 2024. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 276, + 296, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 276, + 296, + 329 + ], + "spans": [ + { + "bbox": [ + 56, + 276, + 296, + 329 + ], + "type": "text", + "content": "[79] Chunting Zhou, Lili Yu, Arun Babu, Kushal Tirumala, Michihiro Yasunaga, Leonid Shamis, Jacob Kahn, Xuezhe Ma, Luke Zettlemoyer, and Omer Levy. Transfusion: Predict the next token and diffuse images with one multi-modal model. 2024. 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 331, + 296, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 331, + 296, + 385 + ], + "spans": [ + { + "bbox": [ + 56, + 331, + 296, + 385 + ], + "type": "text", + "content": "[80] Enshen Zhou, Yiran Qin, Zhenfei Yin, Yuzhou Huang, Ruimao Zhang, Lu Sheng, Yu Qiao, and Jing Shao. Mine-dreamer: Learning to follow instructions via chain-of-imagination for simulated-world control. arXiv preprint arXiv:2403.12037, 2024. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 387, + 296, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 387, + 296, + 453 + ], + "spans": [ + { + "bbox": [ + 56, + 387, + 296, + 453 + ], + "type": "text", + "content": "[81] Gaoyue Zhou, Victoria Dean, Mohan Kumar Srirama, Aravind Rajeswaran, Jyothish Pari, Kyle Hatch, Aryan Jain, Tianhe Yu, Pieter Abbeel, Lerrel Pinto, et al. Train offline, test online: A real robot learning benchmark. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 9197-9203. IEEE, 2023. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 454, + 296, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 454, + 296, + 509 + ], + "spans": [ + { + "bbox": [ + 56, + 454, + 296, + 509 + ], + "type": "text", + "content": "[82] Xinghao Zhu, Ran Tian, Chenfeng Xu, Mingxiao Huo, Wei Zhan, Masayoshi Tomizuka, and Mingyu Ding. Fanuc manipulation: A dataset for learning-based manipulation with fanuc mate 200id robot. https://sites.google.com/berkeley.edu/fanuc-manipulation, 2023.1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 510, + 296, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 510, + 296, + 553 + ], + "spans": [ + { + "bbox": [ + 56, + 510, + 296, + 553 + ], + "type": "text", + "content": "[83] Yifeng Zhu, Abhishek Joshi, Peter Stone, and Yuke Zhu. Viola: Imitation learning for vision-based manipulation with object proposal priors. In Conference on Robot Learning, pages 1199-1210. PMLR, 2023. 1" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1713" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Coarse Correspondences Boost Spatial-Temporal Reasoning in Multimodal Language Model/53ff4ccd-60e6-479a-bd87-54a994f71bf8_content_list.json b/2025/Coarse Correspondences Boost Spatial-Temporal Reasoning in Multimodal Language Model/53ff4ccd-60e6-479a-bd87-54a994f71bf8_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..3f1b8d384ea0b56fd3e81a1036ce86b7bff23dd5 --- /dev/null +++ b/2025/Coarse Correspondences Boost Spatial-Temporal Reasoning in Multimodal Language Model/53ff4ccd-60e6-479a-bd87-54a994f71bf8_content_list.json @@ -0,0 +1,1647 @@ +[ + { + "type": "text", + "text": "COARSE CORRESPONDENCES Boost Spatial-Temporal Reasoning in Multimodal Language Model", + "text_level": 1, + "bbox": [ + 166, + 130, + 833, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Benlin Liu1*, Yuhao Dong2,3, Yiqin Wang2*, Zixian Ma1, Yansong Tang2, Luming Tang5, Yongming Rao3, Wei-Chiu Ma5,6, Ranjay Krishna1,5 \n1University of Washington, 2Tsinghua University, 3Tencent, 4Allen Institute for AI, 5Cornell University \ncoarse-correspondence.github.io", + "bbox": [ + 99, + 202, + 906, + 292 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 325, + 326, + 343 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Multimodal language models (MLLMs) are increasingly being applied in real-world environments, necessitating their ability to interpret 3D spaces and comprehend temporal dynamics. Current methods often rely on specialized architectural designs or task-specific fine-tuning to achieve this. We introduce COARSE CORRESPONDENCES, a simple lightweight method that enhances MLLMs' spatial-temporal reasoning with 2D images as input, without modifying the architecture or requiring task-specific fine-tuning. Our method uses a lightweight tracking model to identify primary object correspondences between frames in a video or across different image viewpoints, and then conveys this information to MLLMs through visual prompting. We demonstrate that this simple training-free approach brings substantial gains to GPT4-V/O consistently on four benchmarks that require spatial-temporal reasoning, including $+20.5\\%$ improvement on ScanQA, $+9.7\\%$ on OpenEQA's episodic memory subset, $+6.0\\%$ on the long-form video benchmark EgoSchema, and $+11\\%$ on the R2R navigation benchmark. Additionally, we show that COARSE CORRESPONDENCES can also enhance open-source MLLMs' spatial reasoning (by $+6.9\\%$ on ScanQA) when applied in both training and inference and that the improvement can generalize to unseen datasets such as SQA3D $(+3.1\\%)$ . Taken together, we show that COARSE CORRESPONDENCES effectively and efficiently boosts models' performance on downstream tasks requiring spatial-temporal reasoning.", + "bbox": [ + 88, + 358, + 485, + 767 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 787, + 220, + 804 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Intelligence is multi-faceted. While multi-modal language models [28] have shown remarkable linguistic, logical and even mathematical intelligence, many remain doubtful about their visual and spatial intelligence. Despite their excel", + "bbox": [ + 89, + 814, + 485, + 876 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "lent performance on visual-linguistic tasks, many recent works [24, 37] demonstrate that state-of-the-art MLLMs still struggle at 3D and long video understanding benchmarks, performing only marginally better than blind text-only baselines. These results suggest that spatial-temporal reasoning is a major bottleneck on MLLMs' path to general visual intelligence.", + "bbox": [ + 511, + 328, + 906, + 434 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To enhance MLLMs' 3D understanding, researchers have mainly explored three approaches: providing MLLMs with 3D data as input [45], designing specialized architectures for 3D tasks [13], or employing supervised fine-tuning with 3D data [6]. Similarly, to boost MLLMs' temporal understanding, prior works have proposed new model architectures designed for long video understanding [4, 30], or adopted Socratic-based methods [15, 50] (i.e., converting each frame of a video into text using a caption model, and then using text-only LLMs to summarize).", + "bbox": [ + 511, + 436, + 908, + 588 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In contrast to prior works that separately enhance spatial and temporal reasoning, we propose COARSE CORRESPON-DENCES, a simple yet effective training-free visual prompting method to jointly boost spatial-temporal reasoning in MLLMs. COARSE CORRESPONDENCES uses a tracking model to extract object-level correspondences across multiple images, and then represent the most salient correspondence relationships on the images through visual prompting. We refer to our method as Coarse because of the following:", + "bbox": [ + 511, + 592, + 910, + 728 + ], + "page_idx": 0 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. We only visually prompt for instance-level correspondences and not point-level correspondences.", + "2. The instance-level correspondences are extracted using off-the-shelf tracking models.", + "3. We only visualize a handful of prominent corresponding instances." + ], + "bbox": [ + 514, + 729, + 906, + 819 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite the obtained correspondence information not being perfectly precise, our method can still significantly boost MLLMs' spatial-temporal reasoning with only 2D image inputs and without any specialized architectural design or task-specific fine-tuning.", + "bbox": [ + 511, + 825, + 906, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution.", + "bbox": [ + 109, + 887, + 218, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "3783", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/377e4dd11bf3046000ae07f806f03b2809b2434a209d339a8b154f4e392f34db.jpg", + "image_caption": [ + "(a) track objects in high frame rate videos" + ], + "image_footnote": [], + "bbox": [ + 99, + 108, + 524, + 186 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/080e16c9a9846667bbfa8e4ac89e364d21d052fb5e7396f2ba56d5d9fcb3c009.jpg", + "image_caption": [ + "(b) construct coarse correspondence on sparsified views" + ], + "image_footnote": [], + "bbox": [ + 102, + 238, + 524, + 329 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/5c32a801503c311eb192b487492c30b884979cec9dd567ae39d349c15132ed28.jpg", + "image_caption": [ + "(c) help MLLMs understand 3D space-time with prompted images", + "Figure 1. We combined light-weight video tracking models and multimodal LLMs to achieve a better understanding of 3D spacetime. (a) We use a tracking model at a high frame rate to obtain instance segmentation masks for each frame. (b) Then, we sequentially sparsify input frames, select prominent coarse correspondences, and visualize the constructed coarse correspondences on the images. (c) Finally, we enable MLLMs to better understand 3D spacetime from the prompted images." + ], + "image_footnote": [], + "bbox": [ + 560, + 90, + 893, + 309 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We have demonstrated substantial performance gains of COARSE CORRESPONDENCES through extensive experiments with both open-source and closed-source models across 6 benchmarks on spatial-temporal reasoning. For closed-source models, we apply COARSE CORRESPONDENCES on GPT4-V/O during inference and achieve compelling gains. First, on spatial reasoning, we show that our method significantly surpasses state-of-the-art models by $20.5\\%$ and $9.7\\%$ on ScanQA [2] and OpenEQA [24] respectively. Second, for long video understanding, our method leads to a $6\\%$ gain in performance on the EgoSchema benchmark [25]. Notably, our method uses much fewer input images and, in a zero-shot manner, outperforms many fine-tuned models that use far more images. For example, on EgoSchema, COARSE CORRESPONDENCES surpasses state-of-the-art results with just 8 uniformly sampled frames from a 3-minute video, greatly reducing the computational costs of MLLMs compared to existing methods. In addition to 3D and video QA tasks, we further demonstrate that our method enhances models' performance on embodied tasks such as navigation [16], which require strong spatial-temporal reasoning, by $11\\%$ in success rate on R2R. These results suggest that COARSE CORRESPONDENCES boosts MLLMs' spatial-temporal reasoning both effectively and efficiently. Last but not least, we experiment with open-source MLLMs [22] by applying COARSE CORRESPONDENCES in both instruction tuning and inference; again, our method shows performance gains against the baseline (by $6.9\\%$ on ScanQA), and the improvement even generalizes to unseen datasets such as SQA3D $(+3.1\\%)$ . These results suggest that COARSE COR", + "bbox": [ + 91, + 446, + 485, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "RESPONDENCES works well universally with any model - both closed-source and open-source - that can take in multiple images and understand visual markers.", + "bbox": [ + 511, + 446, + 906, + 491 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Overall, we want to highlight with this work that, despite its simplicity and being underestimated for semantic tasks in deep learning, visual correspondence can still bring significant utility to spatial-temporal reasoning in MLLMs, just as it has long contributed to 3D reconstruction [35]. We hope our work demonstrates the potential of leveraging general-purpose MLLMs to better understand our physical world.", + "bbox": [ + 511, + 492, + 908, + 599 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Method", + "text_level": 1, + "bbox": [ + 511, + 613, + 604, + 628 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We introduce CoARSE CORRESPONDERENCES, a visual prompting method that allows MLLMs to reason about 3D space and time.", + "bbox": [ + 511, + 638, + 906, + 684 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Problem formulation. Given a question $\\mathcal{Q}$ and a sequence or set of observations in an environment $[I_1, \\ldots, I_n]$ , our aim is to design a visual prompt $\\mathcal{P}(\\ldots)$ that modifies the input image set. These image inputs don't have to be a video. They can also represent a set of images of a specific scene from multiple viewpoints. We evaluate the prompt by measuring its utility in prompting an MLLM $\\mathcal{M}$ :", + "bbox": [ + 511, + 685, + 908, + 791 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} [ I _ {1} ^ {\\prime}, \\dots , I _ {n} ^ {\\prime} ] = P ([ I _ {1}, \\dots , I _ {n} ]) \\\\ \\hat {\\mathcal {A}} = \\mathcal {M} (([ I _ {1} ^ {\\prime}, \\dots , I _ {n} ^ {\\prime} ]), \\mathcal {Q}) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 604, + 801, + 810, + 842 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We compare generated answer $\\hat{\\mathcal{A}}$ with the ground truth $\\mathcal{A}$ .", + "bbox": [ + 511, + 854, + 897, + 869 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In our framework, the MLLMs can be any general-purpose model without requiring any special architecture", + "bbox": [ + 511, + 869, + 906, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "3784", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "or training for spatial-temporal reasoning. Our aim is to develop a prompting strategy that allows models to improve such capabilities without any training (Figure 1).", + "bbox": [ + 89, + 90, + 483, + 137 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Coarse Correspondence", + "text_level": 1, + "bbox": [ + 89, + 143, + 279, + 159 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our prompting method, COARSE CORRESPONDENCES, contains four steps: (1) tracking correspondences, (2) sparsify frames, (3) selecting, and (4) visualizing coarse correspondences.", + "bbox": [ + 89, + 166, + 485, + 224 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "(1) Tracking correspondences. Given $n$ input images, $[I_1, \\ldots, I_n]$ , we first use an off-the-shelf video object tracking model, such as Tracking Anything [47]. This model extracts class-agnostic instance segmentation masks $(M_1, \\ldots, M_n)$ for each image. Each $M_i$ is a $H \\times W$ dimensional matrix where $H$ and $W$ are the height and width of the input image $I_i$ . Each pixel location in $M_i$ contains an instance ID, indicating which instance the pixel at that position belongs to within the image sequence.", + "bbox": [ + 89, + 226, + 483, + 362 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "(2) Sparsify frames. Since most MLLMs contain a large number of parameters, directly using them to process long image sequences is very computationally intensive. Additionally, proprietary MLLMs like GPT-4O can also incur significant costs if the number of image tokens that need to be processed increases. Reducing the number of input images might lose vital information necessary for MLLMs.", + "bbox": [ + 89, + 362, + 483, + 467 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "COARSE CORRESPONDENCES strikes a balance in this tradeoff by extracting meaningful video object tracks (a relatively cheaper operation) from high-frame-rate image sequences, and then samples a few image inputs along with the tracks, to retain—and even improve—performance while reducing the MLLM's computation cost. From this extracted video object tracks, we perform temporal downsampling, retaining only $m << n$ uniformly sampled images and their corresponding masks, denoted as $[I_{s_1},\\dots,I_{s_m}]$ and $[M_{s_1},\\dots,M_{s_m}]$ , where $s_i \\in \\{1,\\dots,n\\}$ . This downsampling reduces the number of images we feed into $\\mathcal{M}$ .", + "bbox": [ + 89, + 468, + 483, + 633 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "(3) Selecting coarse correspondences. Prompting an MLLM with all the detected correspondences results in information overload. In fact, our ablations (discussed in Sec 5) find that adding all the correspondences reduces the MLLM's performance. Therefore, we select a subset of prominent instances to retain. We select the prominent instances of the top-K objects that co-occur in the most number of frames. We first calculate the occurrence frequency and area sum of each unique instance ID in the retained $m$ masks using the following equation:", + "bbox": [ + 89, + 633, + 485, + 785 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {F} r e q (\\mathrm {I D}) = \\sum_ {i = s _ {1}} ^ {s _ {m}} \\mathbf {1} _ {\\{\\mathrm {I D} \\in M _ {i} \\}}, \\mathcal {A r e a} (\\mathrm {I D}) = \\sum_ {i = s _ {1}} ^ {s _ {m}} \\sum_ {p \\in M _ {i}} \\mathbf {1} _ {\\{\\mathrm {I D} = p \\}}.\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 792, + 504, + 834 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Then, we first sort all instance IDs in descending order based on $\\mathcal{Freq}(\\mathrm{ID})$ . If there are ties, we further sort based on $\\mathcal{Area}(\\mathrm{ID})$ . Finally, we retain the top $k$ instance IDs as tracklets, denoted as $[T_1, \\dots, T_k]$ , to visualize for MLLMs.", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "(4) Visualizing coarse correspondences. For each set of obtained correspondence relationships, we visualize the correspondences directly in the image as a marker. Specifically, for each identified primary instance ID $T_{i}$ , if it exists in the mask $M_{s_j}$ of a retained image $I_{s_j}$ , we overlay a mark with a fixed size and shape labeled with $T_{i}$ at the position $(\\bar{x}_{ij},\\bar{y}_{ij})$ on $I_{s_j}$ to produce $I_{s_j}'$ . The specific placement position can be easily obtained by:", + "bbox": [ + 511, + 90, + 906, + 212 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\left(\\bar {x} _ {i j}, \\bar {y} _ {i j}\\right) = \\frac {\\sum_ {(x , y)} (x , y) \\cdot \\mathbf {1} _ {\\{M _ {s _ {j}} (x , y) = T _ {i} \\}}}{\\sum_ {(x , y)} \\mathbf {1} _ {\\{M _ {s _ {j}} (x , y) = T _ {i} \\}}}\n$$\n", + "text_format": "latex", + "bbox": [ + 565, + 220, + 849, + 263 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Naturally, we can overlay not just the markers but also the segmentation outlines or even the segmentation masks associated with each retained prominent instance. We explore these ablations later. In the end, we obtain the prompted image sequence $[I_1',\\dots,I_m']$ , which is then used as the input to MLLMs.", + "bbox": [ + 511, + 271, + 906, + 361 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Prompting Proprietary Models", + "text_level": 1, + "bbox": [ + 511, + 375, + 797, + 393 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We first evaluated the utility of our COARSE CORRESPON-DENCES on multiple tasks using proprietary models, including understanding 3D space (ScanQA [2] and OpenEQA [24] in §3.1) as well as temporal events (EgoSchema [25] in §3.2). Building on the improvements our method brings to spatial-temporal reasoning, we further demonstrate that our method also delivers significant gains in navigation tasks (VLN-CE [16]). Across all these benchmarks, we augment proprietary MLLMs (e.g., GPT-4V and GPT-4O) with COARSE CORRESPON-DENCES and evaluate its zero-shot performance. We show that COARSE CORRESPON-DENCES significantly improves the base GPT models and can substantially surpass many current state-of-the-art methods that have undergone specialized fine-tuning, even while using much fewer images as input. All experiments were conducted using A100 80G GPUs.", + "bbox": [ + 511, + 401, + 908, + 642 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Spatial understanding", + "text_level": 1, + "bbox": [ + 511, + 652, + 720, + 667 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "**Benchmarks.** The validation set of ScanQA dataset contains 4675 questions about 71 scenes. The questions require basic recognition, 3D localization, and 3D embodied capabilities [9]. It contains two ground-truth answers per question for evaluation with models that produce free-form answers. OpenEQA Dataset is an open-vocabulary dataset benchmarking spatial environment understanding and embodied reasoning. We evaluate on OpenEQA's EM-EQA data split, which contains over 1600 high-quality human-generated questions. The subset tests the episodic memory of an agent moving through a 3D environment over time.", + "bbox": [ + 511, + 674, + 908, + 839 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Baselines. For ScanQA, we evaluate COARSE CORRESPON-DENCES by augmenting both GPT-4{V,O}, Gemini and Claude models. Besides, we also consider 3D specialized models [2, 13, 49] fine-tuned on ScanQA. For OpenEQA,", + "bbox": [ + 511, + 839, + 908, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3785", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/ba9d5ab8a0bed2386b4d5a405926f822e1a413ae807c564a1a2d896e0aa3849c.jpg", + "table_caption": [], + "table_footnote": [ + "Table 1. Comparison on ScanQA validation set. We conduct experiments on the ScanQA validation set to demonstrate the effectiveness of COARSE CORRESPONDENCES with different MLLMs. Our method enables both proprietary models and open-source models to surpass all 3D-specific models." + ], + "table_body": "
ModelFrameBLEU-1BLEU-2METEORROUGE-LCIDEr
3D-Specific Models
ScanQA [2]-26.916.611.53055.4
ScanRefer+MCAN [49]-30.220.413.133.364.9
3D-LLM [13]-39.325.214.535.769.4
Open-source Multi-modal Models
LLaVA(Fine-tuned)6434.722.013.831.167.3
LLaVA+Coarse Correspondences6438.624.715.438.374.2
Proprietary Multi-modal Models
GPT-4V828.613.413.533.459.6
GPT-4V+Coarse Correspondences839.725.517.440.879.2
GPT-4O430.519.814.836.172.2
GPT-4O+Coarse Correspondences435.425.518.042.687.0
", + "bbox": [ + 94, + 87, + 906, + 338 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "we compare against language-only models to account for language bias (LLaMA2 [40]), commonly used general-purpose multimodal LLMs (GPT-4 [29], Claude3 [1], Gemini-Pro [38], GPT-4V with 15 and 50 frames.", + "bbox": [ + 89, + 391, + 485, + 450 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Metrics. For ScanQA, following prior works, we adopt BLEU [31], METEOR [5], ROUGE-L [20], and CIDEr [41] as our evaluation metrics. For OpenEQA, we follow their evaluation approach by using GPT-4 to compare the generated answers with the ground-truth answers and assign a score. We report the average score across all questions.", + "bbox": [ + 89, + 455, + 483, + 546 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Results. For ScanQA, as shown in Table 1, compared to raw input, COARSE CORRESPONDENCES consistently improves the overall performance of different proprietary models. For instance, on the strongest model, GPT-40, COARSE CORRESPONDENCES brings improvements of 5.7 BLEU-2, 3.2 METEOR, 6.5 ROUGE-L, and 15 CIDEr points. Compared to methods that are specifically designed for 3D understanding tasks, fine-tuned with specialized 3D SFT data, or even those that use 3D point clouds instead of 2D images as input, we observe that a general-purpose MLLM can still outperform them, especially when enhanced with COARSE CORRESPONDENCES. Moreover, we found that this can be achieved using far fewer images as input.", + "bbox": [ + 88, + 549, + 485, + 747 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We also demonstrated the same conclusion on OpenEQA, as indicated in Table 2. By applying COARSE CORRESPON-DENCES, we significantly improved the performance of both GPT-4v and GPT-4o, achieving better results with fewer input images. These findings suggest that general-purpose MLLMs are indeed capable of understanding 3D space, and COARSE CORRESPON-DENCES can significantly enhance their spatial understanding while simultaneously reducing the number of views needed, which could lower the inference cost and make MLLMs more useful for embodied tasks.", + "bbox": [ + 88, + 750, + 485, + 900 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/02d75a269fd06992a56cb26575913935846d3a345a5a37ec84417a37646a5d13.jpg", + "table_caption": [], + "table_footnote": [ + "Table 2. Comparisons on EM-EQA setting of OpenEQA. Our method further enhances the embodied ability of MLLMs and exceeds previous methods by a large margin." + ], + "table_body": "
ModelsFrameAccuracy
LLaMA2 [40]028.3
GPT-4 [29]033.5
Claude3 [1]2036.3
Gemini-Pro [38]1544.9
GPT-4V [27]1554.6
GPT-4V [27]5055.3
HumanFull86.8
GPT-4V844.8
GPT-4V+CC858.5
GPT-4O449.4
GPT-4O+CC459.1
", + "bbox": [ + 526, + 388, + 890, + 595 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Temporal understanding", + "text_level": 1, + "bbox": [ + 511, + 659, + 740, + 676 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "**Benchmarks.** We evaluated the improvements of our method for long video understanding using the EgoSchema dataset. Each video in EgoSchema is 3 minutes long, with a corresponding question that includes five multiple-choice options. These questions are designed to ensure that answering them requires viewing a sufficient number of frames from the video. Due to budget constraints, we limited our evaluation to 500 questions from the validation set.", + "bbox": [ + 511, + 685, + 908, + 806 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Baselines. The baseline methods we compared against include newly designed and trained model architectures specifically for long video understanding, such as LongViviT [30] and MC-ViT-L [4]. On the other hand, we also compared methods that rely on text-only foundation models (e.g., GPT-4), i.e., Socratic-based approaches [15, 50], which first use an", + "bbox": [ + 511, + 810, + 910, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "3786", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/77aaadb0cb7fbb63c08d18703860b8998c944fc9ccb09d0f1a73969e38990284.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelsFrameAccuracy
LongViviT [30]25656.8
MC-ViT-L [4]128+62.6
LLoVi [50]18058.3
VideoAgent [42]8.460.2
MVU [33]1660.3
VideoAgent [10]-62.8
LangRepo [15]-66.2
GPT-4V864.2
GPT-4V+CC867.4
GPT-4O867.2
GPT-4O+CC873.2
", + "bbox": [ + 102, + 88, + 467, + 295 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 3. Comparisons on EgoSchema validation set. COARSE CORRESPONDENCES improves existing MLLMs and surpasses previous fine-tuned models in a zero-shot manner.", + "bbox": [ + 89, + 297, + 482, + 339 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "off-the-shelf image captioning model [51] to convert video frames into captions, and then prompt GPT-4 to answer questions based on those captions. Additionally, we compared agent-based methods [10, 42], which involve using GPT-4 alongside an image captioning model in an agent framework to iteratively perform a series of multi-step reasoning operations to understand long videos. In contrast to these approaches, our method is entirely based on an end-to-end general MLLM architecture, exploring how to further enhance its ability to understand long videos.", + "bbox": [ + 88, + 353, + 482, + 503 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Results. COARSE CORRESPONDENCES demonstrates state-of-the-art performance, significantly outperforming existing approaches in a zero-shot manner (Table 3). Compared to the original GPT-4o model, our method improves its performance by $6\\%$ . Notably, our method uses far fewer frames than other approaches, yet achieves higher results compared to methods that use many more frames. It is also worth highlighting that even the original GPT-4o, when limited to just 8 frames, already serves as a very strong baseline. This indicates the potential of a sufficiently powerful general-purpose MLLM in long video understanding. We provide a detailed breakdown of COARSE CORRESPONDENCES's performance across different question types in the supplementary material.", + "bbox": [ + 89, + 506, + 485, + 717 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Navigation", + "text_level": 1, + "bbox": [ + 89, + 727, + 210, + 742 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Building on the improvements in 3DQA and VideoQA, we hope that our method can also prove effective in embodied tasks such as navigation. Navigation requires an agent to understand 3D space, such as being able to determine the spatial relationship between objects in the instruction and itself, while also performing temporal reasoning to assess the progress toward completing the instruction. We consider conducting experiments on the VLN-CE benchmark [16], which is a continuous simulation environment for low-level action execution in indoor scenes.", + "bbox": [ + 89, + 750, + 482, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Setup. We adopt the val-unseen split from R2R [16] for evaluation. Unlike the previous QA tasks, where all images could be processed at once, in navigation tasks, each image is processed in an online fashion. Specifically, we feed in one image at each iteration of the conversation. Given the significant variation in viewpoints during navigation, we use SAMv2 [34], the state-of-the-art model for long-range tracking, to label each new input image based on episodic history. Then, using the prompted images, we induce the MLLM to output one of four actions at each step: FORWARD (distance), TURN-LEFT (rotation angle), TURN-RIGHT (rotation angle), and STOP. We follow NavGPT [52] to craft input prompts. Considering the high computational cost of navigation tasks, we selected 100 samples from the val-unseen split. Our primary goal is to demonstrate that our method can enhance GPT models' capabilities in zero-shot navigation tasks, which remains a significant challenge for various types of models.", + "bbox": [ + 511, + 90, + 906, + 362 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Metrics. We follow the standard VLN evaluation metrics to evaluate the navigation performance, including success rate (SR), oracle success rate (OS), success weighted by path length (SPL), trajectory length (TL), and navigation error from goal (NE). Note that an episode is considered successful if the agent calls the STOP action within $3\\mathrm{m}$ of the goal in the VLN-CE.", + "bbox": [ + 511, + 363, + 908, + 468 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Results. As shown in Table 4, our method achieved improvements across all metrics. While GPT-4o performs impressively on many QA tasks, its zero-shot performance on navigation tasks is relatively low. This may partly be due to the lack of specialized training on action data, making it less accurate in outputs such as determining how many meters to move forward. However, our method reveals another dimension of the problem: MLLMs' understanding of the 3D spacetime in which they operate can be further enhanced. This is evidenced by the significant improvements in navigation when using COARSE CORRESPONDENCES. We believe that our approach holds great potential for embodied tasks, which can be explored in future research.", + "bbox": [ + 511, + 469, + 910, + 664 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Prompting Open Models", + "text_level": 1, + "bbox": [ + 511, + 679, + 743, + 696 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We further validate the effectiveness of our COARSE CORRESPONDENCES on open-source models. Our primary goal is to demonstrate that our method not only enhances proprietary models but also yields general improvements across various MLLMs. However, directly prompting open-source models presents two challenges: first, these models lack the ability to interpret prompting marks as effectively as proprietary models; second, they currently cannot handle multiple image inputs well. Thus, we first fine-tune the open-source models to understand prompting marks and manage multiple image inputs, enabling COARSE CORRESPONDENCES to enhance these models' environmental understanding capabilities at the inference stage. In this work, we primarily focus", + "bbox": [ + 511, + 703, + 908, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "3787", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/d600c6ff85301e1e8b6697f6962e931f2a485914d4e618f629fa573567421a43.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsSuccess Rate ↑Oracle Success Rate ↑Success weighted by Path Length ↑Trajectory Length ↑Navigation Error ↓
GPT-4O12.0018.0010.377.318.49
GPT-4O+CC23.0029.0021.038.127.37
", + "bbox": [ + 94, + 88, + 903, + 170 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/9d3c9b7f6685d2d910b423f11c7163af91b63308a6e16858dfc40230bab5ad3f.jpg", + "table_caption": [ + "Table 4. Comparison on Navigation task. COARSE CORRESPONDENCES improve GPT-4o's performance on R2R dataset for different evaluation metrics." + ], + "table_footnote": [], + "table_body": "
MethodAcc
LLaVA(w/o CC during finetune)36.0
LLaVA+CC(train only)37.17
LLaVA+CC(train & test)39.13
", + "bbox": [ + 91, + 214, + 480, + 292 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 5. Comparisons on SQA3D dataset. We tested the effectiveness of COARSE CORRESPONDENCES on the out-of-domain SQA3D dataset. With the same training data, we found that using COARSE CORRESPONDENCES not only improves performance during inference but also enhances zero-shot evaluation when applied solely during the training phase.", + "bbox": [ + 89, + 292, + 485, + 378 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "on proving that COARSE CORRESPONDENCES improves open-source models' understanding of 3D space.", + "bbox": [ + 89, + 395, + 482, + 425 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We start with the LLaVA model [22] and fine-tune it using a dataset comprising approximately 1.2 million samples of image and video data. Notably, the ScanQA dataset is the only 3D-related dataset in this collection. COARSE CORRESPONDENCES is applied exclusively to the ScanQA data, while other data maintain their original format. Fine-tuning on such a dataset enables the open-source LLaVA model to better interpret overlayed marks on images and process multiple image inputs.", + "bbox": [ + 88, + 426, + 485, + 563 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In-domain Evaluation. We first evaluate whether COARSE CORRESPONDENCES can improve open-source models on the ScanQA validation set. As shown in Table 1, our method significantly enhances the model's 3D spatial understanding compared to the baseline fine-tuned on the same data without using COARSE CORRESPONDENCES, even outperforming previous VLMs [13] specifically designed for 3D tasks, which incorporate specialized architectural designs and are fine-tuned on substantially larger amounts of 3D-related data. This in-domain evaluation demonstrates that COARSE CORRESPONDENCES is also effective for open MLLMs.", + "bbox": [ + 89, + 565, + 485, + 744 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Out-domain Evaluation. To further demonstrate the generalizability of our method, we conduct experiments on the SQA3D dataset to evaluate whether COARSE CORRESPONDENCES can also provide improvements in a zero-shot setting. As shown in Table 5, on this previously unseen dataset, COARSE CORRESPONDENCES still outperforms the baseline simply fine-tuned on the same data, proving that our method can enhance open-source models' understanding of 3D space on out-of-domain datasets.", + "bbox": [ + 89, + 747, + 483, + 883 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Notably, even without using CoARSE CORRESPON-", + "bbox": [ + 109, + 885, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "DENCES during inference, simply applying it during training already yields improvements. This highlights that our method is effective not only as a prompting technique for inference but also holds potential as a data augmentation method during training, a prospect worth exploring further in future work.", + "bbox": [ + 511, + 218, + 906, + 308 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Analysis", + "text_level": 1, + "bbox": [ + 511, + 325, + 609, + 343 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Here, we explore the various design decisions in our method. Although the study is conducted only on ScanQA, applying the design choices derived from ScanQA to all other benchmarks also yields significant performance improvements. This strongly demonstrates the generalizability of these design choices. Additionally, we further analyze the impact of our method on mitigating the camera motion bias present in MLLMs.", + "bbox": [ + 511, + 352, + 908, + 473 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "How does COARSE CORRESPONDENCES differ from other visual prompting methods? Our proposed method calculates and highlights correspondences between images, aiming to elicit 3D and temporal understanding. Other visual prompting methods (namely Set-of-Mark [48], 3DAxiesPrompts [21], and Chain-of-thought [43]) can also be viewed as alternative prompting methods. Given that the ground-truth answers in existing benchmarks are relatively brief, we selected a scene from ScanQA and manually designed a new question. We qualitatively compare COARSE CORRESPONDENCES against other prompting methods on this new question, as shown in Figure 2.", + "bbox": [ + 511, + 474, + 908, + 656 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The orange part of Figure 2 shows our Coarse Correspondence labels are recognized by GPT-4V. The output answer provides evidence that our coarse correspondence helps GPT-4V develop a mental 3D model of the scene. Set-of-Marks provides no spatial corresponding information and therefore is unhelpful. The Axis labels in 3DAxies can be easily misrecognized by GPT-4V, leading to misleading spatial information. Though Chain-of-Thought helps identify objects, it fails to resolve the \"spatial perspective-taking\" issue.", + "bbox": [ + 511, + 657, + 908, + 794 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Why use coarse instead of dense correspondences? Instead of filtering and retaining only a handful of coarse correspondences, one ablation we considered is the possibility of using all dense correspondence. Unfortunately, we find that excessively overlaying too many instance marks can degrade performance (Figure 3) as they occlude the visual content in the images.", + "bbox": [ + 511, + 795, + 910, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "3788", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "System: You are an AI with the ability to analyze a series of images, each representing a different perspective of a single scene. [Prompt-about-Marks]. Your task is to construct a 3D understanding based on these images.", + "bbox": [ + 119, + 97, + 872, + 119 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "User: You are sitting on the sofa and the electric fan is on your left. Describe the location of the room door from your perspective: A. to the front left of you; B. to the front right of you; C. to the back left of you; D. to the back right of you.", + "bbox": [ + 119, + 125, + 875, + 148 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "+Coarse Correspondence (Ours)", + "text_level": 1, + "bbox": [ + 129, + 157, + 352, + 172 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0fe6c28e518bdc27bd7fb5f4d1280aa2ffeb6fbf693cd70bd02e580b692921df.jpg", + "image_caption": [ + "View 1" + ], + "image_footnote": [], + "bbox": [ + 124, + 176, + 243, + 246 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2ff9112e172f2835a79129ddc506b432ca7aebcb262c1bd791444c887217654b.jpg", + "image_caption": [ + "View 2" + ], + "image_footnote": [], + "bbox": [ + 253, + 176, + 372, + 246 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "GPT-4V: The first image ... a door marked with number 7, a fan marked with number 39, and items on shelves marked with numbers 11 and 37. In the second image, ... the fan again at number 39, and a sofa marked with number 3(√).", + "bbox": [ + 122, + 265, + 369, + 323 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We can infer that the door and the fan are on the same side of the room(✓). Since the fan appears to the left in the first image and is also on the left when sitting on the sofa as per the second image and your description, the door must be to the front of the sofa, as it is not visible in the second image from sofa's perspective(✓). By 'sitting on the sofa' as instructed, and with the electric fan to the left, I can determine the door is A. to the front left of you (✓).", + "bbox": [ + 122, + 321, + 370, + 433 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "+Set-of-Mark", + "text_level": 1, + "bbox": [ + 395, + 157, + 490, + 172 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8c7e9682a85506eb5ffaf9bd9bde061a44e7719f499a7dc6f7caf310a7ef0c17.jpg", + "image_caption": [ + "View 1" + ], + "image_footnote": [], + "bbox": [ + 385, + 176, + 506, + 244 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/bf90788f5d01e2b03543d31adaae137c152a0b3b8f0f2efb3a29f6a5eb2267b2.jpg", + "image_caption": [ + "View 2" + ], + "image_footnote": [], + "bbox": [ + 509, + 175, + 629, + 244 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "GPT-4V: The fan is seen on the right side from the sofa's perspective, its leftward when sitting means the orientation is flipped(√). Therefore, the door that is visible on the right in the first image would be D. behind and to the right when seated on the sofa(X).", + "bbox": [ + 633, + 176, + 862, + 239 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "+3DAxesPrompts", + "text_level": 1, + "bbox": [ + 395, + 270, + 519, + 282 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/79719678b177040e7303dce4bcbabb3c35fd5cb7b4f41a8efd9c81bd658259a4.jpg", + "image_caption": [ + "View 1" + ], + "image_footnote": [], + "bbox": [ + 385, + 284, + 506, + 354 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8b39bec3be4d458965620e2b801246b81dadc9a99a50c8f4e86746f53bbbc7a0.jpg", + "image_caption": [ + "View 2", + "Figure 2. Comparison of different prompting method. Our proposed Coarse Correspondences successfully guided GPT-40 to understand 3D spatial relationships and generate the right answer. Other existing prompting method including image-based Set-of-Marks, 3DAxies and text-based Chain-of-Thought failed to answer correctly." + ], + "image_footnote": [], + "bbox": [ + 509, + 284, + 629, + 354 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "GPT-4V: The first image shows ... and a coordinate system with X-axis (horizontally rightwards) $(\\mathbf{X})$ and Z-axis (vertically upwards). The second image shows a sofa on the left with a coordinate system's Y-axis (horizontally extending away from the viewer) $(\\mathbf{X})$ and Z-axis as before. When sitting on the sofa with the fan to your left, the door would be B. diagonally opposite to you on the right side $(\\mathbf{X})$ .", + "bbox": [ + 633, + 282, + 877, + 362 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "+Chain-of-Thought", + "text_level": 1, + "bbox": [ + 395, + 375, + 529, + 387 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "System: Your response should be as follows. \n1)Identification: Identify all labeled objects. \n2)Analysis: Analyze the objects and their positions construct a 3D model.", + "bbox": [ + 390, + 390, + 629, + 424 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3)Synthesis: Combine the info above, form a comprehensive understanding.", + "bbox": [ + 390, + 424, + 629, + 440 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4)Response: Provide a concise answer to the question based on your comprehensive 3D modelling.", + "bbox": [ + 390, + 440, + 629, + 457 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "GPT-4V:", + "text_level": 1, + "bbox": [ + 638, + 378, + 684, + 388 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Identification: Door, shelves, sofa,... $(\\checkmark)$ Analysis: in image 1 the door is on the right with the shelves adjacent to the door ... In image 2 the sofa is ... $(\\checkmark)$", + "bbox": [ + 637, + 388, + 874, + 424 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Synthesis: When sitting on the sofa, the shelves are on the right(X). The door is next to the shelves.", + "bbox": [ + 635, + 424, + 874, + 448 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Response: B. (X).", + "bbox": [ + 637, + 448, + 725, + 455 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "How large should the marks be? We inject the correspondences into MLLMs by overlaying the marks into images. We empirically find an optimal mark size (where 'px' represents the mark's diameter in pixels) in Figure 3. Marks that are too small tend to be ignored while those that are too large occlude visual content.", + "bbox": [ + 89, + 525, + 485, + 616 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "What shape should the marks be? We further studied the appearance of the marks. In addition to red circles with white text, we experimented with adding segmentation outlines and segmentation masks. As shown in Figure 3, using segmentation outlines enhances object grounding. However, using segmentation marks occludes visual content and reduces performance.", + "bbox": [ + 89, + 622, + 485, + 728 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Improvement of Camera Motion Invariance with CoARSE CORRESPONDENCES. A crucial aspect of strong spatial reasoning is the model's ability to maintain an understanding of spatial relationships invariant to camera motion; in other words, whether the image sequence is captured from left-to-right or right-to-left, the model's comprehension of static space should remain consistent. However, current benchmarks lack annotations regarding whether the 3D space scan was conducted from left-to-right or right-to-left, making it challenging to analyze the effect of camera motion on MLLMs' spatial understanding. To address this,", + "bbox": [ + 89, + 734, + 485, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "we curated a small diagnostic benchmark specifically for this analysis. This benchmark includes 10 scenes captured from left-to-right, with five questions per scene focusing on left-right spatial relationships between objects—fundamental aspects of spatial understanding. We ensured that all questions required information across multiple frames and could not be answered from a single frame alone. Ideally, regardless of whether the image sequence is input in the original or reversed order, the output should remain consistent. We leave additional details on data curation in the supplementary material.", + "bbox": [ + 511, + 525, + 908, + 690 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To mitigate randomness, each question was tested 20 times, yielding a total of 1000 trials, and we report the average accuracy in Table 6. We then reversed the sequence order to simulate a right-to-left capture and repeated 1000 trials with this reversed input, reporting the average accuracy for this reverse sequence. Finally, we report the harmonic mean of the average accuracies for the forward and reverse inputs.", + "bbox": [ + 511, + 698, + 908, + 819 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Without CoARSE CORRESPONDENCES, the accuracy in understanding spatial relationships in right-to-left sequences was significantly lower than in left-to-right sequences. With CoARSE CORRESPONDENCES, not only did MLLMs achieve a substantial increase in spatial un", + "bbox": [ + 511, + 825, + 910, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "3789", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/bca43439fed192cbc8fcb8351b17493dc4145d1ea329596027de5730b2e276f0.jpg", + "image_caption": [ + "Figure 3. Ablations on different design choices of COARSE CORRESPONDENCES. We studied the impact of the number, size, and type of marks on performance. All experiments were conducted on ScanQA using GPT-4V." + ], + "image_footnote": [], + "bbox": [ + 98, + 92, + 377, + 292 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d9e791dfd0896750629d08fd18032293b53343eaf54d936d0f28d9141647c815.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 377, + 93, + 637, + 292 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/42eac979b8b802b8830f4100c9584b88c550e7117c6b8c35f934381261188779.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 638, + 95, + 898, + 292 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/07aab8d4a4ca3b2ec6e4d6fc397a998df064f7f1898448aa0a393bef5e145c99.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelsFrameForwardReverseHarmonic Mean
GPT-4O458.050.453.9
GPT-4O+CC471.271.271.2
", + "bbox": [ + 94, + 339, + 478, + 393 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 6. Effect of COARSE CORRESPONDENCES on Camera Motion Bias. COARSE CORRESPONDENCES shows strong capability of enhancing 3D spatial understanding of MLLMs. It can also ease the camera motion bias of current MLLMs.", + "bbox": [ + 89, + 398, + 483, + 455 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "derstanding accuracy with the original sequence $(+13.2\\%)$ , but more importantly, reversing the input order no longer drastically impacted the model's comprehension of static spatial relationships. Specifically, COARSE CORRESPON-DENCES improved the harmonic mean of forward and reverse sequence accuracy by $17.3\\%$ . This demonstrates that our method enables MLLMs to perform spatial reasoning invariant to camera motion, thus achieving more robust spatial understanding capabilities.", + "bbox": [ + 88, + 460, + 483, + 597 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Related work", + "text_level": 1, + "bbox": [ + 89, + 616, + 228, + 632 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Multimodal language models Multimodal LLMs [3, 23] integrate vision encoders [32] into large LLMs [8, 40], enabling direct reasoning over visual input. Proprietary models like GPT-4 [27], Gemini [38], and Claude [1], along with open-source models such as the LLaVA series [23] and BLIP series [18], have made significant strides in 2D vision-language tasks like image captioning [7] and visual question answering (VQA) [11, 14]. However, for applications in real-world scenarios like autonomous driving [39] and robotics [46], MLLMs still require stronger spatial-temporal reasoning capabilities. Recent work [19] has enhanced MLLMs' capacity to process multiple image frames and understand temporal dynamics, achieving notable improvements in video dense captioning [17] and videoQA [12, 44]. Nonetheless, MLLMs still face challenges in understanding complex 3D spatial relations [24] and long-term temporal dynamics [25] present in image sequences. Many", + "bbox": [ + 89, + 643, + 485, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "approaches have introduced specialized architectures [13] or fine-tuning methods targeting spatial and temporal reasoning individually. Our goal, however, is to jointly enhance spatial-temporal reasoning within the existing general-purpose MLLM framework. By leveraging off-the-shelf tracking models to extract instance-level correspondences, we achieve this in a training-free manner.", + "bbox": [ + 511, + 342, + 906, + 446 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Visual prompting. Effective prompting has been widely proven to improve LLMs across multiple domains. Methods, such as chain-of-thought prompting [43], force the model to reason before answering a question. In the context of visual prompting, we can enhance MLLMs' grounding abilities [36, 48], their ability to use numerical expressions to describe 3D object relationships [21], and their capability to generate outputs for robotics control [26], all using single-image inputs. Unlike these approaches, COARSE CORRESPONDENCES is designed to enhance MLLMs' spatial-temporal reasoning capabilities for interpreting complex spacetime structures represented in image sequences.", + "bbox": [ + 511, + 449, + 908, + 630 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Conclusion.", + "text_level": 1, + "bbox": [ + 511, + 647, + 635, + 662 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We propose a framework called COARSE CORRESPON-DENCES prompting. By using off-the-shelf video tracking models to obtain class-agnostic, instance-level correspondences and conveying this information to MLLMs through visual prompting, we discovered that this simple method, using only 2D images as input—without any specialized architectural design or task-specific SFT—can effectively enhance MLLMs' spatial-temporal reasoning. This improvement extends to embodied tasks like navigation. Our method not only works on proprietary models but also generalizes to open-source models, and it performs well on both in-domain and out-of-domain datasets. Moreover, it enhances not just inference but also training. Further analysis shows that our method helps MLLMs become more robust to camera motion bias.", + "bbox": [ + 511, + 674, + 908, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "3790", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 91, + 90, + 250, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We thank Xiaojuan Wang, Ruitao Zhang, Yuntian Deng for helpful discussions, feedback and collecting data. This project is partially funded by Amazon Science. This paper was partially supported by a gift from Ai2 and DARPA Transfer from Imprecise and Abstract Models to Autonomous Technologies (TIAMAT) program No. HR00112490422. Its contents are solely the responsibility of the authors and do not necessarily represent the official views of DARPA.", + "bbox": [ + 89, + 113, + 485, + 239 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 253, + 187, + 268 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] AI Anthropic. The claude 3 model family: Opus, sonnet, haiku. Claude-3 Model Card, 2024. 4, 8", + "[2] Daichi Azuma, Taiki Miyanishi, Shuhei Kurita, and Motoaki Kawanabe. Scanqa: 3d question answering for spatial scene understanding. In proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 19129-19139, 2022. 2, 3, 4", + "[3] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond, 2023. 8", + "[4] Ivana Balažević, Yuge Shi, Pinelopi Papalampidi, Rahma Chaabouni, Skanda Koppula, and Olivier J. Henaff. Memory consolidation enables long-context video understanding, 2024. 1, 4, 5", + "[5] Satanjeev Banerjee and Alon Lavie. Meteor: An automatic metric for mt evaluation with improved correlation with human judgments. In Proceedings of the acl workshop on intrinsic and extrinsic evaluation measures for machine translation and/or summarization, pages 65-72, 2005. 4", + "[6] Boyuan Chen, Zhuo Xu, Sean Kirmani, Brian Ichter, Danny Driess, Pete Florence, Dorsa Sadigh, Leonidas Guibas, and Fei Xia. Spatialvlm: Endowing vision-language models with spatial reasoning capabilities. arXiv preprint arXiv:2401.12168, 2024. 1", + "[7] Xinlei Chen, Hao Fang, Tsung-Yi Lin, Ramakrishna Vedantam, Saurabh Gupta, Piotr Dollar, and C. Lawrence Zitnick. Microsoft coco captions: Data collection and evaluation server, 2015. 8", + "[8] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhanghao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E. Gonzalez, Ion Stoica, and Eric P. Xing. Vicuna: An open-source chatbot impressing gpt-4 with $90\\%$ chatgpt quality, 2023. 8", + "[9] Jiafei Duan, Samson Yu, Hui Li Tan, Hongyuan Zhu, and Cheston Tan. A survey of embodied ai: From simulators to research tasks. IEEE Transactions on Emerging Topics in Computational Intelligence, 6(2):230-244, 2022. 3", + "[10] Yue Fan, Xiaojian Ma, Rujie Wu, Yuntao Du, Jiaqi Li, Zhi Gao, and Qing Li. Videoagent: A memory-augmented multimodal agent for video understanding. arXiv preprint arXiv:2403.11481, 2024.5", + "[11] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Batra, and Devi Parikh. Making the v in vqa matter: Elevating" + ], + "bbox": [ + 93, + 277, + 483, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "the role of image understanding in visual question answering, 2017. 8", + "[12] Madeleine Grunde-McLaughlin, Ranjay Krishna, and Maneesh Agrawala. Agqa: A benchmark for compositional spatio-temporal reasoning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11287-11297, 2021. 8", + "[13] Yining Hong, Haoyu Zhen, Peihao Chen, Shuhong Zheng, Yilun Du, Zhenfang Chen, and Chuang Gan. 3d-llm: Injecting the 3d world into large language models. Advances in Neural Information Processing Systems, 36, 2024. 1, 3, 4, 6, 8", + "[14] Drew A. Hudson and Christopher D. Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering, 2019. 8", + "[15] Kumara Kahatapitiya, Kanchana Ranasinghe, Jongwoo Park, and Michael S. Ryoo. Language repository for long video understanding, 2024. 1, 4, 5", + "[16] Jacob Krantz, Erik Wijmans, Arjun Majumdar, Dhruv Batra, and Stefan Lee. Beyond the nav-graph: Vision-and-language navigation in continuous environments. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXVIII 16, pages 104-120. Springer, 2020. 2, 3, 5", + "[17] Ranjay Krishna, Kenji Hata, Frederic Ren, Li Fei-Fei, and Juan Carlos Niebles. Dense-captioning events in videos, 2017. 8", + "[18] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip2: Bootstrapping language-image pre-training with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597, 2023. 8", + "[19] Bin Lin, Bin Zhu, Yang Ye, Munan Ning, Peng Jin, and Li Yuan. Video-llava: Learning united visual representation by alignment before projection. arXiv preprint arXiv:2311.10122, 2023. 8", + "[20] Chin-Yew Lin. Rouge: A package for automatic evaluation of summaries. In Text summarization branches out, pages 74-81, 2004. 4", + "[21] Dingning Liu, Xiaomeng Dong, Renrui Zhang, Xu Luo, Peng Gao, Xiaoshui Huang, Yongshun Gong, and Zhihui Wang. 3daxiesprompts: Unleashing the 3d spatial task capabilities of gpt-4v. arXiv preprint arXiv:2312.09738, 2023. 6, 8", + "[22] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. NeurIPS, 36, 2024. 2, 6", + "[23] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 8", + "[24] Arjun Majumdar, Anurag Ajay, Xiaohan Zhang, Pranav Putta, Sriram Yenamandra, Mikael Henaff, Sneha Silwal, Paul Mcvay, Oleksandr Maksymets, Sergio Arnaud, Karmesh Yadav, Qiyang Li, Ben Newman, Mohit Sharma, Vincent Berges, Shiqi Zhang, Pulkit Agrawal, Yonatan Bisk, Dhruv Batra, Mrinal Kalakrishnan, Franziska Meier, Chris Paxton, Sasha Sax, and Aravind Rajeswaran. Openeqa: Embodied question answering in the era of foundation models. In Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 1, 2, 3, 8" + ], + "bbox": [ + 514, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "3791", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[25] Karttikeya Mangalam, Raiymbek Akshulakov, and Jitendra Malik. Egoschema: A diagnostic benchmark for very long-form video language understanding, 2023. 2, 3, 8", + "[26] Soroush Nasiriany, Fei Xia, Wenhao Yu, Ted Xiao, Jacky Liang, Ishita Dasgupta, Annie Xie, Danny Driess, Ayzaan Wahid, Zhuo Xu, et al. Pivot: Iterative visual prompting elicits actionable knowledge for vlms. arXiv preprint arXiv:2402.07872, 2024. 8", + "[27] OpenAI. Gpt-4v(ision) system card. OpenAI Blog, 2023. 4, 8", + "[28] OpenAI. Hello gpt-4o. https://openai.com/index/hello-gpt-4o/, 2024. Accessed: 2024-05-22.1", + "[29] OpenAI, Josh Achiam, and et al. Gpt-4 technical report, 2024. 4", + "[30] Pinelopi Papalampidi, Skanda Koppula, Shreya Pathak, Justin Chiu, Joe Heyward, Viorica Patraucean, Jiajun Shen, Antoine Miech, Andrew Zisserman, and Aida Nematzdeh. A simple recipe for contrastively pre-training video-first encoders beyond 16 frames, 2023. 1, 4, 5", + "[31] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th annual meeting of the Association for Computational Linguistics, pages 311-318, 2002. 4", + "[32] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 8", + "[33] Kanchana Ranasinghe, Xiang Li, Kumara Kahatapitiya, and Michael S. Ryoo. Understanding long videos in one multimodal language model pass, 2024. 5", + "[34] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, et al. Sam 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714, 2024. 5", + "[35] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4104-4113, 2016. 2", + "[36] Aleksandar Shtedritski, Christian Rupprecht, and Andrea Vedaldi. What does clip know about a red circle? visual prompt engineering for vlms. arXiv preprint arXiv:2304.06712, 2023. 8", + "[37] Simranjit Singh, Georgios Pavlakos, and Dimitrios Stamoulis. Evaluating zero-shot gpt-4v performance on 3d visual question answering benchmarks. arXiv preprint arXiv:2405.18831, 2024. 1", + "[38] Gemini Team, Rohan Anil, and et al. Gemini: A family of highly capable multimodal models, 2024. 4, 8", + "[39] Xiaoyu Tian, Junru Gu, Bailin Li, Yicheng Liu, Chenxu Hu, Yang Wang, Kun Zhan, Peng Jia, Xianpeng Lang, and Hang Zhao. Drivevlm: The convergence of autonomous driving and large vision-language models, 2024. 8", + "[40] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama:" + ], + "bbox": [ + 91, + 92, + 483, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023.4,8", + "[41] Ramakrishna Vedantam, C Lawrence Zitnick, and Devi Parikh. Cider: Consensus-based image description evaluation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4566-4575, 2015. 4", + "[42] Xiaohan Wang, Yuhui Zhang, Orr Zohar, and Serena Yeung-Levy. Videoagent: Long-form video understanding with large language model as agent, 2024. 5", + "[43] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Brian Ichter, Fei Xia, Ed Chi, Quoc Le, and Denny Zhou. Chain-of-thought prompting elicits reasoning in large language models, 2023. 6, 8", + "[44] Junbin Xiao, Xindi Shang, Angela Yao, and Tat-Seng Chua. Next-qa:next phase of question-answering to explaining temporal actions, 2021. 8", + "[45] Runsen Xu, Xiaolong Wang, Tai Wang, Yilun Chen, Jiangmiao Pang, and Dahua Lin. Pointllm: Empowering large language models to understand point clouds. In European Conference on Computer Vision, pages 131-147. Springer, 2025. 1", + "[46] Jingkang Yang, Yuhao Dong, Shuai Liu, Bo Li, Ziyue Wang, Chencheng Jiang, Haoran Tan, Jiamu Kang, Yuanhan Zhang, Kaiyang Zhou, and Ziwei Liu. Octopus: Embodied vision-language programmer from environmental feedback, 2023. 8", + "[47] Jinyu Yang, Mingqi Gao, Zhe Li, Shang Gao, Fangjing Wang, and Feng Zheng. Track anything: Segment anything meets videos. arXiv preprint arXiv:2304.11968, 2023. 3", + "[48] Jianwei Yang, Hao Zhang, Feng Li, Xueyan Zou, Chunyuan Li, and Jianfeng Gao. Set-of-mark prompting unleashes extraordinary visual grounding in gpt-4v. arXiv preprint arXiv:2310.11441, 2023. 6, 8", + "[49] Zhou Yu, Jun Yu, Yuhao Cui, Dacheng Tao, and Qi Tian. Deep modular co-attention networks for visual question answering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6281–6290, 2019. 3, 4", + "[50] Ce Zhang, Taixi Lu, Md Mohaiminul Islam, Ziyang Wang, Shoubin Yu, Mohit Bansal, and Gedas Bertasius. A simple llm framework for long-range video question-answering, 2024. 1, 4, 5", + "[51] Yue Zhao, Ishan Misra, Philipp Krahenbuhl, and Rohit Girdhar. Learning video representations from large language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6586-6597, 2023. 5", + "[52] Gengze Zhou, Yicong Hong, and Qi Wu. Navgpt: Explicit reasoning in vision-and-language navigation with large language models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 7641-7649, 2024. 5" + ], + "bbox": [ + 516, + 92, + 906, + 799 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "3792", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Coarse Correspondences Boost Spatial-Temporal Reasoning in Multimodal Language Model/53ff4ccd-60e6-479a-bd87-54a994f71bf8_model.json b/2025/Coarse Correspondences Boost Spatial-Temporal Reasoning in Multimodal Language Model/53ff4ccd-60e6-479a-bd87-54a994f71bf8_model.json new file mode 100644 index 0000000000000000000000000000000000000000..6a73d771e95f233b749252dd75ce2cd4de6bec52 --- /dev/null +++ b/2025/Coarse Correspondences Boost Spatial-Temporal Reasoning in Multimodal Language Model/53ff4ccd-60e6-479a-bd87-54a994f71bf8_model.json @@ -0,0 +1,2299 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.167, + 0.131, + 0.834, + 0.177 + ], + "angle": 0, + "content": "COARSE CORRESPONDENCES Boost Spatial-Temporal Reasoning in Multimodal Language Model" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.203, + 0.908, + 0.294 + ], + "angle": 0, + "content": "Benlin Liu1*, Yuhao Dong2,3, Yiqin Wang2*, Zixian Ma1, Yansong Tang2, Luming Tang5, Yongming Rao3, Wei-Chiu Ma5,6, Ranjay Krishna1,5 \n1University of Washington, 2Tsinghua University, 3Tencent, 4Allen Institute for AI, 5Cornell University \ncoarse-correspondence.github.io" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.327, + 0.328, + 0.344 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.359, + 0.486, + 0.768 + ], + "angle": 0, + "content": "Multimodal language models (MLLMs) are increasingly being applied in real-world environments, necessitating their ability to interpret 3D spaces and comprehend temporal dynamics. Current methods often rely on specialized architectural designs or task-specific fine-tuning to achieve this. We introduce COARSE CORRESPONDENCES, a simple lightweight method that enhances MLLMs' spatial-temporal reasoning with 2D images as input, without modifying the architecture or requiring task-specific fine-tuning. Our method uses a lightweight tracking model to identify primary object correspondences between frames in a video or across different image viewpoints, and then conveys this information to MLLMs through visual prompting. We demonstrate that this simple training-free approach brings substantial gains to GPT4-V/O consistently on four benchmarks that require spatial-temporal reasoning, including \\(+20.5\\%\\) improvement on ScanQA, \\(+9.7\\%\\) on OpenEQA's episodic memory subset, \\(+6.0\\%\\) on the long-form video benchmark EgoSchema, and \\(+11\\%\\) on the R2R navigation benchmark. Additionally, we show that COARSE CORRESPONDENCES can also enhance open-source MLLMs' spatial reasoning (by \\(+6.9\\%\\) on ScanQA) when applied in both training and inference and that the improvement can generalize to unseen datasets such as SQA3D \\((+3.1\\%)\\). Taken together, we show that COARSE CORRESPONDENCES effectively and efficiently boosts models' performance on downstream tasks requiring spatial-temporal reasoning." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.789, + 0.222, + 0.805 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.815, + 0.486, + 0.877 + ], + "angle": 0, + "content": "Intelligence is multi-faceted. While multi-modal language models [28] have shown remarkable linguistic, logical and even mathematical intelligence, many remain doubtful about their visual and spatial intelligence. Despite their excel" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.329, + 0.908, + 0.435 + ], + "angle": 0, + "content": "lent performance on visual-linguistic tasks, many recent works [24, 37] demonstrate that state-of-the-art MLLMs still struggle at 3D and long video understanding benchmarks, performing only marginally better than blind text-only baselines. These results suggest that spatial-temporal reasoning is a major bottleneck on MLLMs' path to general visual intelligence." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.438, + 0.909, + 0.589 + ], + "angle": 0, + "content": "To enhance MLLMs' 3D understanding, researchers have mainly explored three approaches: providing MLLMs with 3D data as input [45], designing specialized architectures for 3D tasks [13], or employing supervised fine-tuning with 3D data [6]. Similarly, to boost MLLMs' temporal understanding, prior works have proposed new model architectures designed for long video understanding [4, 30], or adopted Socratic-based methods [15, 50] (i.e., converting each frame of a video into text using a caption model, and then using text-only LLMs to summarize)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.593, + 0.911, + 0.729 + ], + "angle": 0, + "content": "In contrast to prior works that separately enhance spatial and temporal reasoning, we propose COARSE CORRESPON-DENCES, a simple yet effective training-free visual prompting method to jointly boost spatial-temporal reasoning in MLLMs. COARSE CORRESPONDENCES uses a tracking model to extract object-level correspondences across multiple images, and then represent the most salient correspondence relationships on the images through visual prompting. We refer to our method as Coarse because of the following:" + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.731, + 0.908, + 0.76 + ], + "angle": 0, + "content": "1. We only visually prompt for instance-level correspondences and not point-level correspondences." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.762, + 0.905, + 0.791 + ], + "angle": 0, + "content": "2. The instance-level correspondences are extracted using off-the-shelf tracking models." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.792, + 0.905, + 0.82 + ], + "angle": 0, + "content": "3. We only visualize a handful of prominent corresponding instances." + }, + { + "type": "list", + "bbox": [ + 0.515, + 0.731, + 0.908, + 0.82 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.908, + 0.902 + ], + "angle": 0, + "content": "Despite the obtained correspondence information not being perfectly precise, our method can still significantly boost MLLMs' spatial-temporal reasoning with only 2D image inputs and without any specialized architectural design or task-specific fine-tuning." + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.22, + 0.9 + ], + "angle": 0, + "content": "*Equal contribution." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3783" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.101, + 0.109, + 0.526, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.176, + 0.201, + 0.461, + 0.217 + ], + "angle": 0, + "content": "(a) track objects in high frame rate videos" + }, + { + "type": "image", + "bbox": [ + 0.104, + 0.239, + 0.526, + 0.33 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.122, + 0.342, + 0.516, + 0.357 + ], + "angle": 0, + "content": "(b) construct coarse correspondence on sparsified views" + }, + { + "type": "image", + "bbox": [ + 0.562, + 0.092, + 0.895, + 0.31 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.58, + 0.326, + 0.875, + 0.356 + ], + "angle": 0, + "content": "(c) help MLLMs understand 3D space-time with prompted images" + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.377, + 0.908, + 0.435 + ], + "angle": 0, + "content": "Figure 1. We combined light-weight video tracking models and multimodal LLMs to achieve a better understanding of 3D spacetime. (a) We use a tracking model at a high frame rate to obtain instance segmentation masks for each frame. (b) Then, we sequentially sparsify input frames, select prominent coarse correspondences, and visualize the constructed coarse correspondences on the images. (c) Finally, we enable MLLMs to better understand 3D spacetime from the prompted images." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.448, + 0.486, + 0.901 + ], + "angle": 0, + "content": "We have demonstrated substantial performance gains of COARSE CORRESPONDENCES through extensive experiments with both open-source and closed-source models across 6 benchmarks on spatial-temporal reasoning. For closed-source models, we apply COARSE CORRESPONDENCES on GPT4-V/O during inference and achieve compelling gains. First, on spatial reasoning, we show that our method significantly surpasses state-of-the-art models by \\(20.5\\%\\) and \\(9.7\\%\\) on ScanQA [2] and OpenEQA [24] respectively. Second, for long video understanding, our method leads to a \\(6\\%\\) gain in performance on the EgoSchema benchmark [25]. Notably, our method uses much fewer input images and, in a zero-shot manner, outperforms many fine-tuned models that use far more images. For example, on EgoSchema, COARSE CORRESPONDENCES surpasses state-of-the-art results with just 8 uniformly sampled frames from a 3-minute video, greatly reducing the computational costs of MLLMs compared to existing methods. In addition to 3D and video QA tasks, we further demonstrate that our method enhances models' performance on embodied tasks such as navigation [16], which require strong spatial-temporal reasoning, by \\(11\\%\\) in success rate on R2R. These results suggest that COARSE CORRESPONDENCES boosts MLLMs' spatial-temporal reasoning both effectively and efficiently. Last but not least, we experiment with open-source MLLMs [22] by applying COARSE CORRESPONDENCES in both instruction tuning and inference; again, our method shows performance gains against the baseline (by \\(6.9\\%\\) on ScanQA), and the improvement even generalizes to unseen datasets such as SQA3D \\((+3.1\\%)\\). These results suggest that COARSE COR" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.448, + 0.908, + 0.492 + ], + "angle": 0, + "content": "RESPONDENCES works well universally with any model - both closed-source and open-source - that can take in multiple images and understand visual markers." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.493, + 0.909, + 0.6 + ], + "angle": 0, + "content": "Overall, we want to highlight with this work that, despite its simplicity and being underestimated for semantic tasks in deep learning, visual correspondence can still bring significant utility to spatial-temporal reasoning in MLLMs, just as it has long contributed to 3D reconstruction [35]. We hope our work demonstrates the potential of leveraging general-purpose MLLMs to better understand our physical world." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.614, + 0.605, + 0.63 + ], + "angle": 0, + "content": "2. Method" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.64, + 0.907, + 0.685 + ], + "angle": 0, + "content": "We introduce CoARSE CORRESPONDERENCES, a visual prompting method that allows MLLMs to reason about 3D space and time." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.686, + 0.909, + 0.792 + ], + "angle": 0, + "content": "Problem formulation. Given a question \\(\\mathcal{Q}\\) and a sequence or set of observations in an environment \\([I_1, \\ldots, I_n]\\), our aim is to design a visual prompt \\(\\mathcal{P}(\\ldots)\\) that modifies the input image set. These image inputs don't have to be a video. They can also represent a set of images of a specific scene from multiple viewpoints. We evaluate the prompt by measuring its utility in prompting an MLLM \\(\\mathcal{M}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.606, + 0.802, + 0.811, + 0.843 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} [ I _ {1} ^ {\\prime}, \\dots , I _ {n} ^ {\\prime} ] = P ([ I _ {1}, \\dots , I _ {n} ]) \\\\ \\hat {\\mathcal {A}} = \\mathcal {M} (([ I _ {1} ^ {\\prime}, \\dots , I _ {n} ^ {\\prime} ]), \\mathcal {Q}) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.855, + 0.898, + 0.87 + ], + "angle": 0, + "content": "We compare generated answer \\(\\hat{\\mathcal{A}}\\) with the ground truth \\(\\mathcal{A}\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.908, + 0.902 + ], + "angle": 0, + "content": "In our framework, the MLLMs can be any general-purpose model without requiring any special architecture" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3784" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.138 + ], + "angle": 0, + "content": "or training for spatial-temporal reasoning. Our aim is to develop a prompting strategy that allows models to improve such capabilities without any training (Figure 1)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.145, + 0.281, + 0.16 + ], + "angle": 0, + "content": "Coarse Correspondence" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.167, + 0.486, + 0.226 + ], + "angle": 0, + "content": "Our prompting method, COARSE CORRESPONDENCES, contains four steps: (1) tracking correspondences, (2) sparsify frames, (3) selecting, and (4) visualizing coarse correspondences." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.227, + 0.485, + 0.363 + ], + "angle": 0, + "content": "(1) Tracking correspondences. Given \\( n \\) input images, \\( [I_1, \\ldots, I_n] \\), we first use an off-the-shelf video object tracking model, such as Tracking Anything [47]. This model extracts class-agnostic instance segmentation masks \\( (M_1, \\ldots, M_n) \\) for each image. Each \\( M_i \\) is a \\( H \\times W \\) dimensional matrix where \\( H \\) and \\( W \\) are the height and width of the input image \\( I_i \\). Each pixel location in \\( M_i \\) contains an instance ID, indicating which instance the pixel at that position belongs to within the image sequence." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.363, + 0.484, + 0.468 + ], + "angle": 0, + "content": "(2) Sparsify frames. Since most MLLMs contain a large number of parameters, directly using them to process long image sequences is very computationally intensive. Additionally, proprietary MLLMs like GPT-4O can also incur significant costs if the number of image tokens that need to be processed increases. Reducing the number of input images might lose vital information necessary for MLLMs." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.469, + 0.485, + 0.635 + ], + "angle": 0, + "content": "COARSE CORRESPONDENCES strikes a balance in this tradeoff by extracting meaningful video object tracks (a relatively cheaper operation) from high-frame-rate image sequences, and then samples a few image inputs along with the tracks, to retain—and even improve—performance while reducing the MLLM's computation cost. From this extracted video object tracks, we perform temporal downsampling, retaining only \\( m << n \\) uniformly sampled images and their corresponding masks, denoted as \\( [I_{s_1},\\dots,I_{s_m}] \\) and \\( [M_{s_1},\\dots,M_{s_m}] \\), where \\( s_i \\in \\{1,\\dots,n\\} \\). This downsampling reduces the number of images we feed into \\( \\mathcal{M} \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.635, + 0.486, + 0.786 + ], + "angle": 0, + "content": "(3) Selecting coarse correspondences. Prompting an MLLM with all the detected correspondences results in information overload. In fact, our ablations (discussed in Sec 5) find that adding all the correspondences reduces the MLLM's performance. Therefore, we select a subset of prominent instances to retain. We select the prominent instances of the top-K objects that co-occur in the most number of frames. We first calculate the occurrence frequency and area sum of each unique instance ID in the retained \\( m \\) masks using the following equation:" + }, + { + "type": "equation", + "bbox": [ + 0.09, + 0.793, + 0.506, + 0.835 + ], + "angle": 0, + "content": "\\[\n\\mathcal {F} r e q (\\mathrm {I D}) = \\sum_ {i = s _ {1}} ^ {s _ {m}} \\mathbf {1} _ {\\{\\mathrm {I D} \\in M _ {i} \\}}, \\mathcal {A r e a} (\\mathrm {I D}) = \\sum_ {i = s _ {1}} ^ {s _ {m}} \\sum_ {p \\in M _ {i}} \\mathbf {1} _ {\\{\\mathrm {I D} = p \\}}.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.485, + 0.902 + ], + "angle": 0, + "content": "Then, we first sort all instance IDs in descending order based on \\( \\mathcal{Freq}(\\mathrm{ID}) \\). If there are ties, we further sort based on \\( \\mathcal{Area}(\\mathrm{ID}) \\). Finally, we retain the top \\( k \\) instance IDs as tracklets, denoted as \\( [T_1, \\dots, T_k] \\), to visualize for MLLMs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.908, + 0.213 + ], + "angle": 0, + "content": "(4) Visualizing coarse correspondences. For each set of obtained correspondence relationships, we visualize the correspondences directly in the image as a marker. Specifically, for each identified primary instance ID \\( T_{i} \\), if it exists in the mask \\( M_{s_j} \\) of a retained image \\( I_{s_j} \\), we overlay a mark with a fixed size and shape labeled with \\( T_{i} \\) at the position \\( (\\bar{x}_{ij},\\bar{y}_{ij}) \\) on \\( I_{s_j} \\) to produce \\( I_{s_j}' \\). The specific placement position can be easily obtained by:" + }, + { + "type": "equation", + "bbox": [ + 0.566, + 0.222, + 0.851, + 0.264 + ], + "angle": 0, + "content": "\\[\n\\left(\\bar {x} _ {i j}, \\bar {y} _ {i j}\\right) = \\frac {\\sum_ {(x , y)} (x , y) \\cdot \\mathbf {1} _ {\\{M _ {s _ {j}} (x , y) = T _ {i} \\}}}{\\sum_ {(x , y)} \\mathbf {1} _ {\\{M _ {s _ {j}} (x , y) = T _ {i} \\}}}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.272, + 0.908, + 0.362 + ], + "angle": 0, + "content": "Naturally, we can overlay not just the markers but also the segmentation outlines or even the segmentation masks associated with each retained prominent instance. We explore these ablations later. In the end, we obtain the prompted image sequence \\([I_1',\\dots,I_m']\\), which is then used as the input to MLLMs." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.376, + 0.798, + 0.394 + ], + "angle": 0, + "content": "3. Prompting Proprietary Models" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.402, + 0.909, + 0.643 + ], + "angle": 0, + "content": "We first evaluated the utility of our COARSE CORRESPON-DENCES on multiple tasks using proprietary models, including understanding 3D space (ScanQA [2] and OpenEQA [24] in §3.1) as well as temporal events (EgoSchema [25] in §3.2). Building on the improvements our method brings to spatial-temporal reasoning, we further demonstrate that our method also delivers significant gains in navigation tasks (VLN-CE [16]). Across all these benchmarks, we augment proprietary MLLMs (e.g., GPT-4V and GPT-4O) with COARSE CORRESPON-DENCES and evaluate its zero-shot performance. We show that COARSE CORRESPON-DENCES significantly improves the base GPT models and can substantially surpass many current state-of-the-art methods that have undergone specialized fine-tuning, even while using much fewer images as input. All experiments were conducted using A100 80G GPUs." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.653, + 0.722, + 0.669 + ], + "angle": 0, + "content": "3.1. Spatial understanding" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.909, + 0.84 + ], + "angle": 0, + "content": "**Benchmarks.** The validation set of ScanQA dataset contains 4675 questions about 71 scenes. The questions require basic recognition, 3D localization, and 3D embodied capabilities [9]. It contains two ground-truth answers per question for evaluation with models that produce free-form answers. OpenEQA Dataset is an open-vocabulary dataset benchmarking spatial environment understanding and embodied reasoning. We evaluate on OpenEQA's EM-EQA data split, which contains over 1600 high-quality human-generated questions. The subset tests the episodic memory of an agent moving through a 3D environment over time." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Baselines. For ScanQA, we evaluate COARSE CORRESPON-DENCES by augmenting both GPT-4{V,O}, Gemini and Claude models. Besides, we also consider 3D specialized models [2, 13, 49] fine-tuned on ScanQA. For OpenEQA," + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3785" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.088, + 0.907, + 0.339 + ], + "angle": 0, + "content": "
ModelFrameBLEU-1BLEU-2METEORROUGE-LCIDEr
3D-Specific Models
ScanQA [2]-26.916.611.53055.4
ScanRefer+MCAN [49]-30.220.413.133.364.9
3D-LLM [13]-39.325.214.535.769.4
Open-source Multi-modal Models
LLaVA(Fine-tuned)6434.722.013.831.167.3
LLaVA+Coarse Correspondences6438.624.715.438.374.2
Proprietary Multi-modal Models
GPT-4V828.613.413.533.459.6
GPT-4V+Coarse Correspondences839.725.517.440.879.2
GPT-4O430.519.814.836.172.2
GPT-4O+Coarse Correspondences435.425.518.042.687.0
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.09, + 0.342, + 0.907, + 0.384 + ], + "angle": 0, + "content": "Table 1. Comparison on ScanQA validation set. We conduct experiments on the ScanQA validation set to demonstrate the effectiveness of COARSE CORRESPONDENCES with different MLLMs. Our method enables both proprietary models and open-source models to surpass all 3D-specific models." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.392, + 0.486, + 0.452 + ], + "angle": 0, + "content": "we compare against language-only models to account for language bias (LLaMA2 [40]), commonly used general-purpose multimodal LLMs (GPT-4 [29], Claude3 [1], Gemini-Pro [38], GPT-4V with 15 and 50 frames." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.456, + 0.484, + 0.547 + ], + "angle": 0, + "content": "Metrics. For ScanQA, following prior works, we adopt BLEU [31], METEOR [5], ROUGE-L [20], and CIDEr [41] as our evaluation metrics. For OpenEQA, we follow their evaluation approach by using GPT-4 to compare the generated answers with the ground-truth answers and assign a score. We report the average score across all questions." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.55, + 0.486, + 0.748 + ], + "angle": 0, + "content": "Results. For ScanQA, as shown in Table 1, compared to raw input, COARSE CORRESPONDENCES consistently improves the overall performance of different proprietary models. For instance, on the strongest model, GPT-40, COARSE CORRESPONDENCES brings improvements of 5.7 BLEU-2, 3.2 METEOR, 6.5 ROUGE-L, and 15 CIDEr points. Compared to methods that are specifically designed for 3D understanding tasks, fine-tuned with specialized 3D SFT data, or even those that use 3D point clouds instead of 2D images as input, we observe that a general-purpose MLLM can still outperform them, especially when enhanced with COARSE CORRESPONDENCES. Moreover, we found that this can be achieved using far fewer images as input." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.75, + 0.486, + 0.901 + ], + "angle": 0, + "content": "We also demonstrated the same conclusion on OpenEQA, as indicated in Table 2. By applying COARSE CORRESPON-DENCES, we significantly improved the performance of both GPT-4v and GPT-4o, achieving better results with fewer input images. These findings suggest that general-purpose MLLMs are indeed capable of understanding 3D space, and COARSE CORRESPON-DENCES can significantly enhance their spatial understanding while simultaneously reducing the number of views needed, which could lower the inference cost and make MLLMs more useful for embodied tasks." + }, + { + "type": "table", + "bbox": [ + 0.527, + 0.389, + 0.891, + 0.596 + ], + "angle": 0, + "content": "
ModelsFrameAccuracy
LLaMA2 [40]028.3
GPT-4 [29]033.5
Claude3 [1]2036.3
Gemini-Pro [38]1544.9
GPT-4V [27]1554.6
GPT-4V [27]5055.3
HumanFull86.8
GPT-4V844.8
GPT-4V+CC858.5
GPT-4O449.4
GPT-4O+CC459.1
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.513, + 0.598, + 0.907, + 0.641 + ], + "angle": 0, + "content": "Table 2. Comparisons on EM-EQA setting of OpenEQA. Our method further enhances the embodied ability of MLLMs and exceeds previous methods by a large margin." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.66, + 0.741, + 0.677 + ], + "angle": 0, + "content": "3.2. Temporal understanding" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.686, + 0.909, + 0.807 + ], + "angle": 0, + "content": "**Benchmarks.** We evaluated the improvements of our method for long video understanding using the EgoSchema dataset. Each video in EgoSchema is 3 minutes long, with a corresponding question that includes five multiple-choice options. These questions are designed to ensure that answering them requires viewing a sufficient number of frames from the video. Due to budget constraints, we limited our evaluation to 500 questions from the validation set." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.911, + 0.902 + ], + "angle": 0, + "content": "Baselines. The baseline methods we compared against include newly designed and trained model architectures specifically for long video understanding, such as LongViviT [30] and MC-ViT-L [4]. On the other hand, we also compared methods that rely on text-only foundation models (e.g., GPT-4), i.e., Socratic-based approaches [15, 50], which first use an" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3786" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.104, + 0.089, + 0.468, + 0.296 + ], + "angle": 0, + "content": "
ModelsFrameAccuracy
LongViviT [30]25656.8
MC-ViT-L [4]128+62.6
LLoVi [50]18058.3
VideoAgent [42]8.460.2
MVU [33]1660.3
VideoAgent [10]-62.8
LangRepo [15]-66.2
GPT-4V864.2
GPT-4V+CC867.4
GPT-4O867.2
GPT-4O+CC873.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.298, + 0.483, + 0.34 + ], + "angle": 0, + "content": "Table 3. Comparisons on EgoSchema validation set. COARSE CORRESPONDENCES improves existing MLLMs and surpasses previous fine-tuned models in a zero-shot manner." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.354, + 0.483, + 0.505 + ], + "angle": 0, + "content": "off-the-shelf image captioning model [51] to convert video frames into captions, and then prompt GPT-4 to answer questions based on those captions. Additionally, we compared agent-based methods [10, 42], which involve using GPT-4 alongside an image captioning model in an agent framework to iteratively perform a series of multi-step reasoning operations to understand long videos. In contrast to these approaches, our method is entirely based on an end-to-end general MLLM architecture, exploring how to further enhance its ability to understand long videos." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.507, + 0.486, + 0.718 + ], + "angle": 0, + "content": "Results. COARSE CORRESPONDENCES demonstrates state-of-the-art performance, significantly outperforming existing approaches in a zero-shot manner (Table 3). Compared to the original GPT-4o model, our method improves its performance by \\(6\\%\\). Notably, our method uses far fewer frames than other approaches, yet achieves higher results compared to methods that use many more frames. It is also worth highlighting that even the original GPT-4o, when limited to just 8 frames, already serves as a very strong baseline. This indicates the potential of a sufficiently powerful general-purpose MLLM in long video understanding. We provide a detailed breakdown of COARSE CORRESPONDENCES's performance across different question types in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.728, + 0.212, + 0.743 + ], + "angle": 0, + "content": "3.3. Navigation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Building on the improvements in 3DQA and VideoQA, we hope that our method can also prove effective in embodied tasks such as navigation. Navigation requires an agent to understand 3D space, such as being able to determine the spatial relationship between objects in the instruction and itself, while also performing temporal reasoning to assess the progress toward completing the instruction. We consider conducting experiments on the VLN-CE benchmark [16], which is a continuous simulation environment for low-level action execution in indoor scenes." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.908, + 0.363 + ], + "angle": 0, + "content": "Setup. We adopt the val-unseen split from R2R [16] for evaluation. Unlike the previous QA tasks, where all images could be processed at once, in navigation tasks, each image is processed in an online fashion. Specifically, we feed in one image at each iteration of the conversation. Given the significant variation in viewpoints during navigation, we use SAMv2 [34], the state-of-the-art model for long-range tracking, to label each new input image based on episodic history. Then, using the prompted images, we induce the MLLM to output one of four actions at each step: FORWARD (distance), TURN-LEFT (rotation angle), TURN-RIGHT (rotation angle), and STOP. We follow NavGPT [52] to craft input prompts. Considering the high computational cost of navigation tasks, we selected 100 samples from the val-unseen split. Our primary goal is to demonstrate that our method can enhance GPT models' capabilities in zero-shot navigation tasks, which remains a significant challenge for various types of models." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.364, + 0.909, + 0.469 + ], + "angle": 0, + "content": "Metrics. We follow the standard VLN evaluation metrics to evaluate the navigation performance, including success rate (SR), oracle success rate (OS), success weighted by path length (SPL), trajectory length (TL), and navigation error from goal (NE). Note that an episode is considered successful if the agent calls the STOP action within \\(3\\mathrm{m}\\) of the goal in the VLN-CE." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.47, + 0.911, + 0.665 + ], + "angle": 0, + "content": "Results. As shown in Table 4, our method achieved improvements across all metrics. While GPT-4o performs impressively on many QA tasks, its zero-shot performance on navigation tasks is relatively low. This may partly be due to the lack of specialized training on action data, making it less accurate in outputs such as determining how many meters to move forward. However, our method reveals another dimension of the problem: MLLMs' understanding of the 3D spacetime in which they operate can be further enhanced. This is evidenced by the significant improvements in navigation when using COARSE CORRESPONDENCES. We believe that our approach holds great potential for embodied tasks, which can be explored in future research." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.68, + 0.745, + 0.697 + ], + "angle": 0, + "content": "4. Prompting Open Models" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.704, + 0.909, + 0.901 + ], + "angle": 0, + "content": "We further validate the effectiveness of our COARSE CORRESPONDENCES on open-source models. Our primary goal is to demonstrate that our method not only enhances proprietary models but also yields general improvements across various MLLMs. However, directly prompting open-source models presents two challenges: first, these models lack the ability to interpret prompting marks as effectively as proprietary models; second, they currently cannot handle multiple image inputs well. Thus, we first fine-tune the open-source models to understand prompting marks and manage multiple image inputs, enabling COARSE CORRESPONDENCES to enhance these models' environmental understanding capabilities at the inference stage. In this work, we primarily focus" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3787" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.089, + 0.904, + 0.171 + ], + "angle": 0, + "content": "
MethodsSuccess Rate ↑Oracle Success Rate ↑Success weighted by Path Length ↑Trajectory Length ↑Navigation Error ↓
GPT-4O12.0018.0010.377.318.49
GPT-4O+CC23.0029.0021.038.127.37
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.172, + 0.907, + 0.201 + ], + "angle": 0, + "content": "Table 4. Comparison on Navigation task. COARSE CORRESPONDENCES improve GPT-4o's performance on R2R dataset for different evaluation metrics." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.215, + 0.481, + 0.294 + ], + "angle": 0, + "content": "
MethodAcc
LLaVA(w/o CC during finetune)36.0
LLaVA+CC(train only)37.17
LLaVA+CC(train & test)39.13
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.294, + 0.486, + 0.38 + ], + "angle": 0, + "content": "Table 5. Comparisons on SQA3D dataset. We tested the effectiveness of COARSE CORRESPONDENCES on the out-of-domain SQA3D dataset. With the same training data, we found that using COARSE CORRESPONDENCES not only improves performance during inference but also enhances zero-shot evaluation when applied solely during the training phase." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.396, + 0.483, + 0.426 + ], + "angle": 0, + "content": "on proving that COARSE CORRESPONDENCES improves open-source models' understanding of 3D space." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.427, + 0.486, + 0.564 + ], + "angle": 0, + "content": "We start with the LLaVA model [22] and fine-tune it using a dataset comprising approximately 1.2 million samples of image and video data. Notably, the ScanQA dataset is the only 3D-related dataset in this collection. COARSE CORRESPONDENCES is applied exclusively to the ScanQA data, while other data maintain their original format. Fine-tuning on such a dataset enables the open-source LLaVA model to better interpret overlayed marks on images and process multiple image inputs." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.566, + 0.486, + 0.746 + ], + "angle": 0, + "content": "In-domain Evaluation. We first evaluate whether COARSE CORRESPONDENCES can improve open-source models on the ScanQA validation set. As shown in Table 1, our method significantly enhances the model's 3D spatial understanding compared to the baseline fine-tuned on the same data without using COARSE CORRESPONDENCES, even outperforming previous VLMs [13] specifically designed for 3D tasks, which incorporate specialized architectural designs and are fine-tuned on substantially larger amounts of 3D-related data. This in-domain evaluation demonstrates that COARSE CORRESPONDENCES is also effective for open MLLMs." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.748, + 0.485, + 0.884 + ], + "angle": 0, + "content": "Out-domain Evaluation. To further demonstrate the generalizability of our method, we conduct experiments on the SQA3D dataset to evaluate whether COARSE CORRESPONDENCES can also provide improvements in a zero-shot setting. As shown in Table 5, on this previously unseen dataset, COARSE CORRESPONDENCES still outperforms the baseline simply fine-tuned on the same data, proving that our method can enhance open-source models' understanding of 3D space on out-of-domain datasets." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.886, + 0.485, + 0.901 + ], + "angle": 0, + "content": "Notably, even without using CoARSE CORRESPON-" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.219, + 0.908, + 0.309 + ], + "angle": 0, + "content": "DENCES during inference, simply applying it during training already yields improvements. This highlights that our method is effective not only as a prompting technique for inference but also holds potential as a data augmentation method during training, a prospect worth exploring further in future work." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.327, + 0.611, + 0.344 + ], + "angle": 0, + "content": "5. Analysis" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.353, + 0.909, + 0.474 + ], + "angle": 0, + "content": "Here, we explore the various design decisions in our method. Although the study is conducted only on ScanQA, applying the design choices derived from ScanQA to all other benchmarks also yields significant performance improvements. This strongly demonstrates the generalizability of these design choices. Additionally, we further analyze the impact of our method on mitigating the camera motion bias present in MLLMs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.476, + 0.909, + 0.657 + ], + "angle": 0, + "content": "How does COARSE CORRESPONDENCES differ from other visual prompting methods? Our proposed method calculates and highlights correspondences between images, aiming to elicit 3D and temporal understanding. Other visual prompting methods (namely Set-of-Mark [48], 3DAxiesPrompts [21], and Chain-of-thought [43]) can also be viewed as alternative prompting methods. Given that the ground-truth answers in existing benchmarks are relatively brief, we selected a scene from ScanQA and manually designed a new question. We qualitatively compare COARSE CORRESPONDENCES against other prompting methods on this new question, as shown in Figure 2." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.658, + 0.909, + 0.795 + ], + "angle": 0, + "content": "The orange part of Figure 2 shows our Coarse Correspondence labels are recognized by GPT-4V. The output answer provides evidence that our coarse correspondence helps GPT-4V develop a mental 3D model of the scene. Set-of-Marks provides no spatial corresponding information and therefore is unhelpful. The Axis labels in 3DAxies can be easily misrecognized by GPT-4V, leading to misleading spatial information. Though Chain-of-Thought helps identify objects, it fails to resolve the \"spatial perspective-taking\" issue." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.911, + 0.902 + ], + "angle": 0, + "content": "Why use coarse instead of dense correspondences? Instead of filtering and retaining only a handful of coarse correspondences, one ablation we considered is the possibility of using all dense correspondence. Unfortunately, we find that excessively overlaying too many instance marks can degrade performance (Figure 3) as they occlude the visual content in the images." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3788" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.12, + 0.098, + 0.873, + 0.121 + ], + "angle": 0, + "content": "System: You are an AI with the ability to analyze a series of images, each representing a different perspective of a single scene. [Prompt-about-Marks]. Your task is to construct a 3D understanding based on these images." + }, + { + "type": "text", + "bbox": [ + 0.12, + 0.126, + 0.877, + 0.149 + ], + "angle": 0, + "content": "User: You are sitting on the sofa and the electric fan is on your left. Describe the location of the room door from your perspective: A. to the front left of you; B. to the front right of you; C. to the back left of you; D. to the back right of you." + }, + { + "type": "title", + "bbox": [ + 0.13, + 0.159, + 0.354, + 0.173 + ], + "angle": 0, + "content": "+Coarse Correspondence (Ours)" + }, + { + "type": "image", + "bbox": [ + 0.125, + 0.177, + 0.245, + 0.247 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.166, + 0.249, + 0.203, + 0.258 + ], + "angle": 0, + "content": "View 1" + }, + { + "type": "image", + "bbox": [ + 0.254, + 0.178, + 0.373, + 0.247 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.295, + 0.249, + 0.332, + 0.258 + ], + "angle": 0, + "content": "View 2" + }, + { + "type": "text", + "bbox": [ + 0.123, + 0.266, + 0.37, + 0.324 + ], + "angle": 0, + "content": "GPT-4V: The first image ... a door marked with number 7, a fan marked with number 39, and items on shelves marked with numbers 11 and 37. In the second image, ... the fan again at number 39, and a sofa marked with number 3(√)." + }, + { + "type": "text", + "bbox": [ + 0.123, + 0.323, + 0.371, + 0.434 + ], + "angle": 0, + "content": "We can infer that the door and the fan are on the same side of the room(✓). Since the fan appears to the left in the first image and is also on the left when sitting on the sofa as per the second image and your description, the door must be to the front of the sofa, as it is not visible in the second image from sofa's perspective(✓). By 'sitting on the sofa' as instructed, and with the electric fan to the left, I can determine the door is A. to the front left of you (✓)." + }, + { + "type": "title", + "bbox": [ + 0.397, + 0.159, + 0.491, + 0.173 + ], + "angle": 0, + "content": "+Set-of-Mark" + }, + { + "type": "image", + "bbox": [ + 0.386, + 0.177, + 0.507, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.427, + 0.248, + 0.463, + 0.257 + ], + "angle": 0, + "content": "View 1" + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.176, + 0.63, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.556, + 0.248, + 0.592, + 0.257 + ], + "angle": 0, + "content": "View 2" + }, + { + "type": "text", + "bbox": [ + 0.635, + 0.178, + 0.863, + 0.24 + ], + "angle": 0, + "content": "GPT-4V: The fan is seen on the right side from the sofa's perspective, its leftward when sitting means the orientation is flipped(√). Therefore, the door that is visible on the right in the first image would be D. behind and to the right when seated on the sofa(X)." + }, + { + "type": "title", + "bbox": [ + 0.397, + 0.271, + 0.521, + 0.284 + ], + "angle": 0, + "content": "+3DAxesPrompts" + }, + { + "type": "image", + "bbox": [ + 0.386, + 0.285, + 0.507, + 0.355 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.427, + 0.356, + 0.463, + 0.365 + ], + "angle": 0, + "content": "View 1" + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.285, + 0.63, + 0.355 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.556, + 0.356, + 0.592, + 0.365 + ], + "angle": 0, + "content": "View 2" + }, + { + "type": "text", + "bbox": [ + 0.635, + 0.284, + 0.879, + 0.363 + ], + "angle": 0, + "content": "GPT-4V: The first image shows ... and a coordinate system with X-axis (horizontally rightwards) \\((\\mathbf{X})\\) and Z-axis (vertically upwards). The second image shows a sofa on the left with a coordinate system's Y-axis (horizontally extending away from the viewer) \\((\\mathbf{X})\\) and Z-axis as before. When sitting on the sofa with the fan to your left, the door would be B. diagonally opposite to you on the right side \\((\\mathbf{X})\\)." + }, + { + "type": "title", + "bbox": [ + 0.397, + 0.375, + 0.531, + 0.388 + ], + "angle": 0, + "content": "+Chain-of-Thought" + }, + { + "type": "text", + "bbox": [ + 0.391, + 0.391, + 0.63, + 0.425 + ], + "angle": 0, + "content": "System: Your response should be as follows. \n1)Identification: Identify all labeled objects. \n2)Analysis: Analyze the objects and their positions construct a 3D model." + }, + { + "type": "text", + "bbox": [ + 0.391, + 0.425, + 0.63, + 0.441 + ], + "angle": 0, + "content": "3)Synthesis: Combine the info above, form a comprehensive understanding." + }, + { + "type": "text", + "bbox": [ + 0.391, + 0.441, + 0.63, + 0.458 + ], + "angle": 0, + "content": "4)Response: Provide a concise answer to the question based on your comprehensive 3D modelling." + }, + { + "type": "title", + "bbox": [ + 0.639, + 0.379, + 0.685, + 0.389 + ], + "angle": 0, + "content": "GPT-4V:" + }, + { + "type": "text", + "bbox": [ + 0.638, + 0.39, + 0.875, + 0.425 + ], + "angle": 0, + "content": "Identification: Door, shelves, sofa,... \\((\\checkmark)\\) Analysis: in image 1 the door is on the right with the shelves adjacent to the door ... In image 2 the sofa is ... \\((\\checkmark)\\)" + }, + { + "type": "text", + "bbox": [ + 0.637, + 0.425, + 0.875, + 0.449 + ], + "angle": 0, + "content": "Synthesis: When sitting on the sofa, the shelves are on the right(X). The door is next to the shelves." + }, + { + "type": "text", + "bbox": [ + 0.638, + 0.449, + 0.726, + 0.457 + ], + "angle": 0, + "content": "Response: B. (X)." + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.479, + 0.907, + 0.521 + ], + "angle": 0, + "content": "Figure 2. Comparison of different prompting method. Our proposed Coarse Correspondences successfully guided GPT-40 to understand 3D spatial relationships and generate the right answer. Other existing prompting method including image-based Set-of-Marks, 3DAxies and text-based Chain-of-Thought failed to answer correctly." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.526, + 0.486, + 0.617 + ], + "angle": 0, + "content": "How large should the marks be? We inject the correspondences into MLLMs by overlaying the marks into images. We empirically find an optimal mark size (where 'px' represents the mark's diameter in pixels) in Figure 3. Marks that are too small tend to be ignored while those that are too large occlude visual content." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.623, + 0.486, + 0.729 + ], + "angle": 0, + "content": "What shape should the marks be? We further studied the appearance of the marks. In addition to red circles with white text, we experimented with adding segmentation outlines and segmentation masks. As shown in Figure 3, using segmentation outlines enhances object grounding. However, using segmentation marks occludes visual content and reduces performance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.486, + 0.902 + ], + "angle": 0, + "content": "Improvement of Camera Motion Invariance with CoARSE CORRESPONDENCES. A crucial aspect of strong spatial reasoning is the model's ability to maintain an understanding of spatial relationships invariant to camera motion; in other words, whether the image sequence is captured from left-to-right or right-to-left, the model's comprehension of static space should remain consistent. However, current benchmarks lack annotations regarding whether the 3D space scan was conducted from left-to-right or right-to-left, making it challenging to analyze the effect of camera motion on MLLMs' spatial understanding. To address this," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.526, + 0.909, + 0.691 + ], + "angle": 0, + "content": "we curated a small diagnostic benchmark specifically for this analysis. This benchmark includes 10 scenes captured from left-to-right, with five questions per scene focusing on left-right spatial relationships between objects—fundamental aspects of spatial understanding. We ensured that all questions required information across multiple frames and could not be answered from a single frame alone. Ideally, regardless of whether the image sequence is input in the original or reversed order, the output should remain consistent. We leave additional details on data curation in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.699, + 0.909, + 0.82 + ], + "angle": 0, + "content": "To mitigate randomness, each question was tested 20 times, yielding a total of 1000 trials, and we report the average accuracy in Table 6. We then reversed the sequence order to simulate a right-to-left capture and repeated 1000 trials with this reversed input, reporting the average accuracy for this reverse sequence. Finally, we report the harmonic mean of the average accuracies for the forward and reverse inputs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.911, + 0.903 + ], + "angle": 0, + "content": "Without CoARSE CORRESPONDENCES, the accuracy in understanding spatial relationships in right-to-left sequences was significantly lower than in left-to-right sequences. With CoARSE CORRESPONDENCES, not only did MLLMs achieve a substantial increase in spatial un" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3789" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.093, + 0.379, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.379, + 0.094, + 0.638, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.64, + 0.096, + 0.9, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.3, + 0.908, + 0.33 + ], + "angle": 0, + "content": "Figure 3. Ablations on different design choices of COARSE CORRESPONDENCES. We studied the impact of the number, size, and type of marks on performance. All experiments were conducted on ScanQA using GPT-4V." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.34, + 0.48, + 0.394 + ], + "angle": 0, + "content": "
ModelsFrameForwardReverseHarmonic Mean
GPT-4O458.050.453.9
GPT-4O+CC471.271.271.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.399, + 0.484, + 0.456 + ], + "angle": 0, + "content": "Table 6. Effect of COARSE CORRESPONDENCES on Camera Motion Bias. COARSE CORRESPONDENCES shows strong capability of enhancing 3D spatial understanding of MLLMs. It can also ease the camera motion bias of current MLLMs." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.461, + 0.484, + 0.598 + ], + "angle": 0, + "content": "derstanding accuracy with the original sequence \\((+13.2\\%)\\), but more importantly, reversing the input order no longer drastically impacted the model's comprehension of static spatial relationships. Specifically, COARSE CORRESPON-DENCES improved the harmonic mean of forward and reverse sequence accuracy by \\(17.3\\%\\). This demonstrates that our method enables MLLMs to perform spatial reasoning invariant to camera motion, thus achieving more robust spatial understanding capabilities." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.617, + 0.229, + 0.633 + ], + "angle": 0, + "content": "6. Related work" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.644, + 0.486, + 0.903 + ], + "angle": 0, + "content": "Multimodal language models Multimodal LLMs [3, 23] integrate vision encoders [32] into large LLMs [8, 40], enabling direct reasoning over visual input. Proprietary models like GPT-4 [27], Gemini [38], and Claude [1], along with open-source models such as the LLaVA series [23] and BLIP series [18], have made significant strides in 2D vision-language tasks like image captioning [7] and visual question answering (VQA) [11, 14]. However, for applications in real-world scenarios like autonomous driving [39] and robotics [46], MLLMs still require stronger spatial-temporal reasoning capabilities. Recent work [19] has enhanced MLLMs' capacity to process multiple image frames and understand temporal dynamics, achieving notable improvements in video dense captioning [17] and videoQA [12, 44]. Nonetheless, MLLMs still face challenges in understanding complex 3D spatial relations [24] and long-term temporal dynamics [25] present in image sequences. Many" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.343, + 0.908, + 0.448 + ], + "angle": 0, + "content": "approaches have introduced specialized architectures [13] or fine-tuning methods targeting spatial and temporal reasoning individually. Our goal, however, is to jointly enhance spatial-temporal reasoning within the existing general-purpose MLLM framework. By leveraging off-the-shelf tracking models to extract instance-level correspondences, we achieve this in a training-free manner." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.45, + 0.909, + 0.631 + ], + "angle": 0, + "content": "Visual prompting. Effective prompting has been widely proven to improve LLMs across multiple domains. Methods, such as chain-of-thought prompting [43], force the model to reason before answering a question. In the context of visual prompting, we can enhance MLLMs' grounding abilities [36, 48], their ability to use numerical expressions to describe 3D object relationships [21], and their capability to generate outputs for robotics control [26], all using single-image inputs. Unlike these approaches, COARSE CORRESPONDENCES is designed to enhance MLLMs' spatial-temporal reasoning capabilities for interpreting complex spacetime structures represented in image sequences." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.648, + 0.637, + 0.664 + ], + "angle": 0, + "content": "7. Conclusion." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.909, + 0.9 + ], + "angle": 0, + "content": "We propose a framework called COARSE CORRESPON-DENCES prompting. By using off-the-shelf video tracking models to obtain class-agnostic, instance-level correspondences and conveying this information to MLLMs through visual prompting, we discovered that this simple method, using only 2D images as input—without any specialized architectural design or task-specific SFT—can effectively enhance MLLMs' spatial-temporal reasoning. This improvement extends to embodied tasks like navigation. Our method not only works on proprietary models but also generalizes to open-source models, and it performs well on both in-domain and out-of-domain datasets. Moreover, it enhances not just inference but also training. Further analysis shows that our method helps MLLMs become more robust to camera motion bias." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3790" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.251, + 0.108 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.114, + 0.486, + 0.24 + ], + "angle": 0, + "content": "We thank Xiaojuan Wang, Ruitao Zhang, Yuntian Deng for helpful discussions, feedback and collecting data. This project is partially funded by Amazon Science. This paper was partially supported by a gift from Ai2 and DARPA Transfer from Imprecise and Abstract Models to Autonomous Technologies (TIAMAT) program No. HR00112490422. Its contents are solely the responsibility of the authors and do not necessarily represent the official views of DARPA." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.254, + 0.188, + 0.269 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.279, + 0.484, + 0.307 + ], + "angle": 0, + "content": "[1] AI Anthropic. The claude 3 model family: Opus, sonnet, haiku. Claude-3 Model Card, 2024. 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.308, + 0.484, + 0.376 + ], + "angle": 0, + "content": "[2] Daichi Azuma, Taiki Miyanishi, Shuhei Kurita, and Motoaki Kawanabe. Scanqa: 3d question answering for spatial scene understanding. In proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 19129-19139, 2022. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.379, + 0.485, + 0.434 + ], + "angle": 0, + "content": "[3] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.436, + 0.484, + 0.49 + ], + "angle": 0, + "content": "[4] Ivana Balažević, Yuge Shi, Pinelopi Papalampidi, Rahma Chaabouni, Skanda Koppula, and Olivier J. Henaff. Memory consolidation enables long-context video understanding, 2024. 1, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.492, + 0.484, + 0.561 + ], + "angle": 0, + "content": "[5] Satanjeev Banerjee and Alon Lavie. Meteor: An automatic metric for mt evaluation with improved correlation with human judgments. In Proceedings of the acl workshop on intrinsic and extrinsic evaluation measures for machine translation and/or summarization, pages 65-72, 2005. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.563, + 0.484, + 0.631 + ], + "angle": 0, + "content": "[6] Boyuan Chen, Zhuo Xu, Sean Kirmani, Brian Ichter, Danny Driess, Pete Florence, Dorsa Sadigh, Leonidas Guibas, and Fei Xia. Spatialvlm: Endowing vision-language models with spatial reasoning capabilities. arXiv preprint arXiv:2401.12168, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.633, + 0.484, + 0.688 + ], + "angle": 0, + "content": "[7] Xinlei Chen, Hao Fang, Tsung-Yi Lin, Ramakrishna Vedantam, Saurabh Gupta, Piotr Dollar, and C. Lawrence Zitnick. Microsoft coco captions: Data collection and evaluation server, 2015. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.689, + 0.484, + 0.759 + ], + "angle": 0, + "content": "[8] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhanghao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E. Gonzalez, Ion Stoica, and Eric P. Xing. Vicuna: An open-source chatbot impressing gpt-4 with \\(90\\%\\) chatgpt quality, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.76, + 0.484, + 0.815 + ], + "angle": 0, + "content": "[9] Jiafei Duan, Samson Yu, Hui Li Tan, Hongyuan Zhu, and Cheston Tan. A survey of embodied ai: From simulators to research tasks. IEEE Transactions on Emerging Topics in Computational Intelligence, 6(2):230-244, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.817, + 0.484, + 0.871 + ], + "angle": 0, + "content": "[10] Yue Fan, Xiaojian Ma, Rujie Wu, Yuntao Du, Jiaqi Li, Zhi Gao, and Qing Li. Videoagent: A memory-augmented multimodal agent for video understanding. arXiv preprint arXiv:2403.11481, 2024.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.484, + 0.902 + ], + "angle": 0, + "content": "[11] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Batra, and Devi Parikh. Making the v in vqa matter: Elevating" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.279, + 0.485, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.908, + 0.12 + ], + "angle": 0, + "content": "the role of image understanding in visual question answering, 2017. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.123, + 0.908, + 0.192 + ], + "angle": 0, + "content": "[12] Madeleine Grunde-McLaughlin, Ranjay Krishna, and Maneesh Agrawala. Agqa: A benchmark for compositional spatio-temporal reasoning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11287-11297, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.194, + 0.907, + 0.25 + ], + "angle": 0, + "content": "[13] Yining Hong, Haoyu Zhen, Peihao Chen, Shuhong Zheng, Yilun Du, Zhenfang Chen, and Chuang Gan. 3d-llm: Injecting the 3d world into large language models. Advances in Neural Information Processing Systems, 36, 2024. 1, 3, 4, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.252, + 0.906, + 0.294 + ], + "angle": 0, + "content": "[14] Drew A. Hudson and Christopher D. Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering, 2019. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.297, + 0.907, + 0.338 + ], + "angle": 0, + "content": "[15] Kumara Kahatapitiya, Kanchana Ranasinghe, Jongwoo Park, and Michael S. Ryoo. Language repository for long video understanding, 2024. 1, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.341, + 0.908, + 0.424 + ], + "angle": 0, + "content": "[16] Jacob Krantz, Erik Wijmans, Arjun Majumdar, Dhruv Batra, and Stefan Lee. Beyond the nav-graph: Vision-and-language navigation in continuous environments. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXVIII 16, pages 104-120. Springer, 2020. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.426, + 0.908, + 0.466 + ], + "angle": 0, + "content": "[17] Ranjay Krishna, Kenji Hata, Frederic Ren, Li Fei-Fei, and Juan Carlos Niebles. Dense-captioning events in videos, 2017. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.47, + 0.907, + 0.525 + ], + "angle": 0, + "content": "[18] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip2: Bootstrapping language-image pre-training with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.528, + 0.908, + 0.582 + ], + "angle": 0, + "content": "[19] Bin Lin, Bin Zhu, Yang Ye, Munan Ning, Peng Jin, and Li Yuan. Video-llava: Learning united visual representation by alignment before projection. arXiv preprint arXiv:2311.10122, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.586, + 0.906, + 0.627 + ], + "angle": 0, + "content": "[20] Chin-Yew Lin. Rouge: A package for automatic evaluation of summaries. In Text summarization branches out, pages 74-81, 2004. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.63, + 0.907, + 0.686 + ], + "angle": 0, + "content": "[21] Dingning Liu, Xiaomeng Dong, Renrui Zhang, Xu Luo, Peng Gao, Xiaoshui Huang, Yongshun Gong, and Zhihui Wang. 3daxiesprompts: Unleashing the 3d spatial task capabilities of gpt-4v. arXiv preprint arXiv:2312.09738, 2023. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.688, + 0.907, + 0.716 + ], + "angle": 0, + "content": "[22] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. NeurIPS, 36, 2024. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.719, + 0.908, + 0.761 + ], + "angle": 0, + "content": "[23] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.763, + 0.908, + 0.901 + ], + "angle": 0, + "content": "[24] Arjun Majumdar, Anurag Ajay, Xiaohan Zhang, Pranav Putta, Sriram Yenamandra, Mikael Henaff, Sneha Silwal, Paul Mcvay, Oleksandr Maksymets, Sergio Arnaud, Karmesh Yadav, Qiyang Li, Ben Newman, Mohit Sharma, Vincent Berges, Shiqi Zhang, Pulkit Agrawal, Yonatan Bisk, Dhruv Batra, Mrinal Kalakrishnan, Franziska Meier, Chris Paxton, Sasha Sax, and Aravind Rajeswaran. Openeqa: Embodied question answering in the era of foundation models. In Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 1, 2, 3, 8" + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.093, + 0.908, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "3791" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.484, + 0.134 + ], + "angle": 0, + "content": "[25] Karttikeya Mangalam, Raiymbek Akshulakov, and Jitendra Malik. Egoschema: A diagnostic benchmark for very long-form video language understanding, 2023. 2, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.134, + 0.484, + 0.203 + ], + "angle": 0, + "content": "[26] Soroush Nasiriany, Fei Xia, Wenhao Yu, Ted Xiao, Jacky Liang, Ishita Dasgupta, Annie Xie, Danny Driess, Ayzaan Wahid, Zhuo Xu, et al. Pivot: Iterative visual prompting elicits actionable knowledge for vlms. arXiv preprint arXiv:2402.07872, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.204, + 0.482, + 0.219 + ], + "angle": 0, + "content": "[27] OpenAI. Gpt-4v(ision) system card. OpenAI Blog, 2023. 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.219, + 0.482, + 0.246 + ], + "angle": 0, + "content": "[28] OpenAI. Hello gpt-4o. https://openai.com/index/hello-gpt-4o/, 2024. Accessed: 2024-05-22.1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.246, + 0.482, + 0.273 + ], + "angle": 0, + "content": "[29] OpenAI, Josh Achiam, and et al. Gpt-4 technical report, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.274, + 0.482, + 0.343 + ], + "angle": 0, + "content": "[30] Pinelopi Papalampidi, Skanda Koppula, Shreya Pathak, Justin Chiu, Joe Heyward, Viorica Patraucean, Jiajun Shen, Antoine Miech, Andrew Zisserman, and Aida Nematzdeh. A simple recipe for contrastively pre-training video-first encoders beyond 16 frames, 2023. 1, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.343, + 0.482, + 0.412 + ], + "angle": 0, + "content": "[31] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th annual meeting of the Association for Computational Linguistics, pages 311-318, 2002. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.413, + 0.482, + 0.495 + ], + "angle": 0, + "content": "[32] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.497, + 0.482, + 0.538 + ], + "angle": 0, + "content": "[33] Kanchana Ranasinghe, Xiang Li, Kumara Kahatapitiya, and Michael S. Ryoo. Understanding long videos in one multimodal language model pass, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.539, + 0.482, + 0.607 + ], + "angle": 0, + "content": "[34] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, et al. Sam 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.608, + 0.482, + 0.663 + ], + "angle": 0, + "content": "[35] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4104-4113, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.664, + 0.482, + 0.718 + ], + "angle": 0, + "content": "[36] Aleksandar Shtedritski, Christian Rupprecht, and Andrea Vedaldi. What does clip know about a red circle? visual prompt engineering for vlms. arXiv preprint arXiv:2304.06712, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.719, + 0.482, + 0.774 + ], + "angle": 0, + "content": "[37] Simranjit Singh, Georgios Pavlakos, and Dimitrios Stamoulis. Evaluating zero-shot gpt-4v performance on 3d visual question answering benchmarks. arXiv preprint arXiv:2405.18831, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.775, + 0.482, + 0.803 + ], + "angle": 0, + "content": "[38] Gemini Team, Rohan Anil, and et al. Gemini: A family of highly capable multimodal models, 2024. 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.804, + 0.482, + 0.859 + ], + "angle": 0, + "content": "[39] Xiaoyu Tian, Junru Gu, Bailin Li, Yicheng Liu, Chenxu Hu, Yang Wang, Kun Zhan, Peng Jia, Xianpeng Lang, and Hang Zhao. Drivevlm: The convergence of autonomous driving and large vision-language models, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.859, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[40] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama:" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.121 + ], + "angle": 0, + "content": "Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023.4,8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.907, + 0.178 + ], + "angle": 0, + "content": "[41] Ramakrishna Vedantam, C Lawrence Zitnick, and Devi Parikh. Cider: Consensus-based image description evaluation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4566-4575, 2015. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.907, + 0.22 + ], + "angle": 0, + "content": "[42] Xiaohan Wang, Yuhui Zhang, Orr Zohar, and Serena Yeung-Levy. Videoagent: Long-form video understanding with large language model as agent, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.221, + 0.907, + 0.276 + ], + "angle": 0, + "content": "[43] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Brian Ichter, Fei Xia, Ed Chi, Quoc Le, and Denny Zhou. Chain-of-thought prompting elicits reasoning in large language models, 2023. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.277, + 0.907, + 0.319 + ], + "angle": 0, + "content": "[44] Junbin Xiao, Xindi Shang, Angela Yao, and Tat-Seng Chua. Next-qa:next phase of question-answering to explaining temporal actions, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.32, + 0.907, + 0.388 + ], + "angle": 0, + "content": "[45] Runsen Xu, Xiaolong Wang, Tai Wang, Yilun Chen, Jiangmiao Pang, and Dahua Lin. Pointllm: Empowering large language models to understand point clouds. In European Conference on Computer Vision, pages 131-147. Springer, 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.39, + 0.907, + 0.459 + ], + "angle": 0, + "content": "[46] Jingkang Yang, Yuhao Dong, Shuai Liu, Bo Li, Ziyue Wang, Chencheng Jiang, Haoran Tan, Jiamu Kang, Yuanhan Zhang, Kaiyang Zhou, and Ziwei Liu. Octopus: Embodied vision-language programmer from environmental feedback, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.461, + 0.906, + 0.503 + ], + "angle": 0, + "content": "[47] Jinyu Yang, Mingqi Gao, Zhe Li, Shang Gao, Fangjing Wang, and Feng Zheng. Track anything: Segment anything meets videos. arXiv preprint arXiv:2304.11968, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.504, + 0.906, + 0.559 + ], + "angle": 0, + "content": "[48] Jianwei Yang, Hao Zhang, Feng Li, Xueyan Zou, Chunyuan Li, and Jianfeng Gao. Set-of-mark prompting unleashes extraordinary visual grounding in gpt-4v. arXiv preprint arXiv:2310.11441, 2023. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.56, + 0.907, + 0.616 + ], + "angle": 0, + "content": "[49] Zhou Yu, Jun Yu, Yuhao Cui, Dacheng Tao, and Qi Tian. Deep modular co-attention networks for visual question answering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6281–6290, 2019. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.617, + 0.907, + 0.671 + ], + "angle": 0, + "content": "[50] Ce Zhang, Taixi Lu, Md Mohaiminul Islam, Ziyang Wang, Shoubin Yu, Mohit Bansal, and Gedas Bertasius. A simple llm framework for long-range video question-answering, 2024. 1, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.673, + 0.907, + 0.741 + ], + "angle": 0, + "content": "[51] Yue Zhao, Ishan Misra, Philipp Krahenbuhl, and Rohit Girdhar. Learning video representations from large language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6586-6597, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.743, + 0.907, + 0.8 + ], + "angle": 0, + "content": "[52] Gengze Zhou, Yicong Hong, and Qi Wu. Navgpt: Explicit reasoning in vision-and-language navigation with large language models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 7641-7649, 2024. 5" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.8 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3792" + } + ] +] \ No newline at end of file diff --git a/2025/Coarse Correspondences Boost Spatial-Temporal Reasoning in Multimodal Language Model/53ff4ccd-60e6-479a-bd87-54a994f71bf8_origin.pdf b/2025/Coarse Correspondences Boost Spatial-Temporal Reasoning in Multimodal Language Model/53ff4ccd-60e6-479a-bd87-54a994f71bf8_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..cbf2785c37a406730957e7ad98078a085dbfce6c --- /dev/null +++ b/2025/Coarse Correspondences Boost Spatial-Temporal Reasoning in Multimodal Language Model/53ff4ccd-60e6-479a-bd87-54a994f71bf8_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:71e4262fd1380631a677db443ef40cb2caec0554643b1dab9337c6f1a91f2850 +size 4715618 diff --git a/2025/Coarse Correspondences Boost Spatial-Temporal Reasoning in Multimodal Language Model/full.md b/2025/Coarse Correspondences Boost Spatial-Temporal Reasoning in Multimodal Language Model/full.md new file mode 100644 index 0000000000000000000000000000000000000000..f744af0a9f5f629ef0daf6a4ff0dea6e6d49edf4 --- /dev/null +++ b/2025/Coarse Correspondences Boost Spatial-Temporal Reasoning in Multimodal Language Model/full.md @@ -0,0 +1,323 @@ +# COARSE CORRESPONDENCES Boost Spatial-Temporal Reasoning in Multimodal Language Model + +Benlin Liu1*, Yuhao Dong2,3, Yiqin Wang2*, Zixian Ma1, Yansong Tang2, Luming Tang5, Yongming Rao3, Wei-Chiu Ma5,6, Ranjay Krishna1,5 +1University of Washington, 2Tsinghua University, 3Tencent, 4Allen Institute for AI, 5Cornell University +coarse-correspondence.github.io + +# Abstract + +Multimodal language models (MLLMs) are increasingly being applied in real-world environments, necessitating their ability to interpret 3D spaces and comprehend temporal dynamics. Current methods often rely on specialized architectural designs or task-specific fine-tuning to achieve this. We introduce COARSE CORRESPONDENCES, a simple lightweight method that enhances MLLMs' spatial-temporal reasoning with 2D images as input, without modifying the architecture or requiring task-specific fine-tuning. Our method uses a lightweight tracking model to identify primary object correspondences between frames in a video or across different image viewpoints, and then conveys this information to MLLMs through visual prompting. We demonstrate that this simple training-free approach brings substantial gains to GPT4-V/O consistently on four benchmarks that require spatial-temporal reasoning, including $+20.5\%$ improvement on ScanQA, $+9.7\%$ on OpenEQA's episodic memory subset, $+6.0\%$ on the long-form video benchmark EgoSchema, and $+11\%$ on the R2R navigation benchmark. Additionally, we show that COARSE CORRESPONDENCES can also enhance open-source MLLMs' spatial reasoning (by $+6.9\%$ on ScanQA) when applied in both training and inference and that the improvement can generalize to unseen datasets such as SQA3D $(+3.1\%)$ . Taken together, we show that COARSE CORRESPONDENCES effectively and efficiently boosts models' performance on downstream tasks requiring spatial-temporal reasoning. + +# 1. Introduction + +Intelligence is multi-faceted. While multi-modal language models [28] have shown remarkable linguistic, logical and even mathematical intelligence, many remain doubtful about their visual and spatial intelligence. Despite their excel + +lent performance on visual-linguistic tasks, many recent works [24, 37] demonstrate that state-of-the-art MLLMs still struggle at 3D and long video understanding benchmarks, performing only marginally better than blind text-only baselines. These results suggest that spatial-temporal reasoning is a major bottleneck on MLLMs' path to general visual intelligence. + +To enhance MLLMs' 3D understanding, researchers have mainly explored three approaches: providing MLLMs with 3D data as input [45], designing specialized architectures for 3D tasks [13], or employing supervised fine-tuning with 3D data [6]. Similarly, to boost MLLMs' temporal understanding, prior works have proposed new model architectures designed for long video understanding [4, 30], or adopted Socratic-based methods [15, 50] (i.e., converting each frame of a video into text using a caption model, and then using text-only LLMs to summarize). + +In contrast to prior works that separately enhance spatial and temporal reasoning, we propose COARSE CORRESPON-DENCES, a simple yet effective training-free visual prompting method to jointly boost spatial-temporal reasoning in MLLMs. COARSE CORRESPONDENCES uses a tracking model to extract object-level correspondences across multiple images, and then represent the most salient correspondence relationships on the images through visual prompting. We refer to our method as Coarse because of the following: + +1. We only visually prompt for instance-level correspondences and not point-level correspondences. +2. The instance-level correspondences are extracted using off-the-shelf tracking models. +3. We only visualize a handful of prominent corresponding instances. + +Despite the obtained correspondence information not being perfectly precise, our method can still significantly boost MLLMs' spatial-temporal reasoning with only 2D image inputs and without any specialized architectural design or task-specific fine-tuning. + +![](images/377e4dd11bf3046000ae07f806f03b2809b2434a209d339a8b154f4e392f34db.jpg) +(a) track objects in high frame rate videos + +![](images/080e16c9a9846667bbfa8e4ac89e364d21d052fb5e7396f2ba56d5d9fcb3c009.jpg) +(b) construct coarse correspondence on sparsified views + +![](images/5c32a801503c311eb192b487492c30b884979cec9dd567ae39d349c15132ed28.jpg) +(c) help MLLMs understand 3D space-time with prompted images +Figure 1. We combined light-weight video tracking models and multimodal LLMs to achieve a better understanding of 3D spacetime. (a) We use a tracking model at a high frame rate to obtain instance segmentation masks for each frame. (b) Then, we sequentially sparsify input frames, select prominent coarse correspondences, and visualize the constructed coarse correspondences on the images. (c) Finally, we enable MLLMs to better understand 3D spacetime from the prompted images. + +We have demonstrated substantial performance gains of COARSE CORRESPONDENCES through extensive experiments with both open-source and closed-source models across 6 benchmarks on spatial-temporal reasoning. For closed-source models, we apply COARSE CORRESPONDENCES on GPT4-V/O during inference and achieve compelling gains. First, on spatial reasoning, we show that our method significantly surpasses state-of-the-art models by $20.5\%$ and $9.7\%$ on ScanQA [2] and OpenEQA [24] respectively. Second, for long video understanding, our method leads to a $6\%$ gain in performance on the EgoSchema benchmark [25]. Notably, our method uses much fewer input images and, in a zero-shot manner, outperforms many fine-tuned models that use far more images. For example, on EgoSchema, COARSE CORRESPONDENCES surpasses state-of-the-art results with just 8 uniformly sampled frames from a 3-minute video, greatly reducing the computational costs of MLLMs compared to existing methods. In addition to 3D and video QA tasks, we further demonstrate that our method enhances models' performance on embodied tasks such as navigation [16], which require strong spatial-temporal reasoning, by $11\%$ in success rate on R2R. These results suggest that COARSE CORRESPONDENCES boosts MLLMs' spatial-temporal reasoning both effectively and efficiently. Last but not least, we experiment with open-source MLLMs [22] by applying COARSE CORRESPONDENCES in both instruction tuning and inference; again, our method shows performance gains against the baseline (by $6.9\%$ on ScanQA), and the improvement even generalizes to unseen datasets such as SQA3D $(+3.1\%)$ . These results suggest that COARSE COR + +RESPONDENCES works well universally with any model - both closed-source and open-source - that can take in multiple images and understand visual markers. + +Overall, we want to highlight with this work that, despite its simplicity and being underestimated for semantic tasks in deep learning, visual correspondence can still bring significant utility to spatial-temporal reasoning in MLLMs, just as it has long contributed to 3D reconstruction [35]. We hope our work demonstrates the potential of leveraging general-purpose MLLMs to better understand our physical world. + +# 2. Method + +We introduce CoARSE CORRESPONDERENCES, a visual prompting method that allows MLLMs to reason about 3D space and time. + +Problem formulation. Given a question $\mathcal{Q}$ and a sequence or set of observations in an environment $[I_1, \ldots, I_n]$ , our aim is to design a visual prompt $\mathcal{P}(\ldots)$ that modifies the input image set. These image inputs don't have to be a video. They can also represent a set of images of a specific scene from multiple viewpoints. We evaluate the prompt by measuring its utility in prompting an MLLM $\mathcal{M}$ : + +$$ +\begin{array}{l} [ I _ {1} ^ {\prime}, \dots , I _ {n} ^ {\prime} ] = P ([ I _ {1}, \dots , I _ {n} ]) \\ \hat {\mathcal {A}} = \mathcal {M} (([ I _ {1} ^ {\prime}, \dots , I _ {n} ^ {\prime} ]), \mathcal {Q}) \\ \end{array} +$$ + +We compare generated answer $\hat{\mathcal{A}}$ with the ground truth $\mathcal{A}$ . + +In our framework, the MLLMs can be any general-purpose model without requiring any special architecture + +or training for spatial-temporal reasoning. Our aim is to develop a prompting strategy that allows models to improve such capabilities without any training (Figure 1). + +# Coarse Correspondence + +Our prompting method, COARSE CORRESPONDENCES, contains four steps: (1) tracking correspondences, (2) sparsify frames, (3) selecting, and (4) visualizing coarse correspondences. + +(1) Tracking correspondences. Given $n$ input images, $[I_1, \ldots, I_n]$ , we first use an off-the-shelf video object tracking model, such as Tracking Anything [47]. This model extracts class-agnostic instance segmentation masks $(M_1, \ldots, M_n)$ for each image. Each $M_i$ is a $H \times W$ dimensional matrix where $H$ and $W$ are the height and width of the input image $I_i$ . Each pixel location in $M_i$ contains an instance ID, indicating which instance the pixel at that position belongs to within the image sequence. + +(2) Sparsify frames. Since most MLLMs contain a large number of parameters, directly using them to process long image sequences is very computationally intensive. Additionally, proprietary MLLMs like GPT-4O can also incur significant costs if the number of image tokens that need to be processed increases. Reducing the number of input images might lose vital information necessary for MLLMs. + +COARSE CORRESPONDENCES strikes a balance in this tradeoff by extracting meaningful video object tracks (a relatively cheaper operation) from high-frame-rate image sequences, and then samples a few image inputs along with the tracks, to retain—and even improve—performance while reducing the MLLM's computation cost. From this extracted video object tracks, we perform temporal downsampling, retaining only $m << n$ uniformly sampled images and their corresponding masks, denoted as $[I_{s_1},\dots,I_{s_m}]$ and $[M_{s_1},\dots,M_{s_m}]$ , where $s_i \in \{1,\dots,n\}$ . This downsampling reduces the number of images we feed into $\mathcal{M}$ . + +(3) Selecting coarse correspondences. Prompting an MLLM with all the detected correspondences results in information overload. In fact, our ablations (discussed in Sec 5) find that adding all the correspondences reduces the MLLM's performance. Therefore, we select a subset of prominent instances to retain. We select the prominent instances of the top-K objects that co-occur in the most number of frames. We first calculate the occurrence frequency and area sum of each unique instance ID in the retained $m$ masks using the following equation: + +$$ +\mathcal {F} r e q (\mathrm {I D}) = \sum_ {i = s _ {1}} ^ {s _ {m}} \mathbf {1} _ {\{\mathrm {I D} \in M _ {i} \}}, \mathcal {A r e a} (\mathrm {I D}) = \sum_ {i = s _ {1}} ^ {s _ {m}} \sum_ {p \in M _ {i}} \mathbf {1} _ {\{\mathrm {I D} = p \}}. +$$ + +Then, we first sort all instance IDs in descending order based on $\mathcal{Freq}(\mathrm{ID})$ . If there are ties, we further sort based on $\mathcal{Area}(\mathrm{ID})$ . Finally, we retain the top $k$ instance IDs as tracklets, denoted as $[T_1, \dots, T_k]$ , to visualize for MLLMs. + +(4) Visualizing coarse correspondences. For each set of obtained correspondence relationships, we visualize the correspondences directly in the image as a marker. Specifically, for each identified primary instance ID $T_{i}$ , if it exists in the mask $M_{s_j}$ of a retained image $I_{s_j}$ , we overlay a mark with a fixed size and shape labeled with $T_{i}$ at the position $(\bar{x}_{ij},\bar{y}_{ij})$ on $I_{s_j}$ to produce $I_{s_j}'$ . The specific placement position can be easily obtained by: + +$$ +\left(\bar {x} _ {i j}, \bar {y} _ {i j}\right) = \frac {\sum_ {(x , y)} (x , y) \cdot \mathbf {1} _ {\{M _ {s _ {j}} (x , y) = T _ {i} \}}}{\sum_ {(x , y)} \mathbf {1} _ {\{M _ {s _ {j}} (x , y) = T _ {i} \}}} +$$ + +Naturally, we can overlay not just the markers but also the segmentation outlines or even the segmentation masks associated with each retained prominent instance. We explore these ablations later. In the end, we obtain the prompted image sequence $[I_1',\dots,I_m']$ , which is then used as the input to MLLMs. + +# 3. Prompting Proprietary Models + +We first evaluated the utility of our COARSE CORRESPON-DENCES on multiple tasks using proprietary models, including understanding 3D space (ScanQA [2] and OpenEQA [24] in §3.1) as well as temporal events (EgoSchema [25] in §3.2). Building on the improvements our method brings to spatial-temporal reasoning, we further demonstrate that our method also delivers significant gains in navigation tasks (VLN-CE [16]). Across all these benchmarks, we augment proprietary MLLMs (e.g., GPT-4V and GPT-4O) with COARSE CORRESPON-DENCES and evaluate its zero-shot performance. We show that COARSE CORRESPON-DENCES significantly improves the base GPT models and can substantially surpass many current state-of-the-art methods that have undergone specialized fine-tuning, even while using much fewer images as input. All experiments were conducted using A100 80G GPUs. + +# 3.1. Spatial understanding + +**Benchmarks.** The validation set of ScanQA dataset contains 4675 questions about 71 scenes. The questions require basic recognition, 3D localization, and 3D embodied capabilities [9]. It contains two ground-truth answers per question for evaluation with models that produce free-form answers. OpenEQA Dataset is an open-vocabulary dataset benchmarking spatial environment understanding and embodied reasoning. We evaluate on OpenEQA's EM-EQA data split, which contains over 1600 high-quality human-generated questions. The subset tests the episodic memory of an agent moving through a 3D environment over time. + +Baselines. For ScanQA, we evaluate COARSE CORRESPON-DENCES by augmenting both GPT-4{V,O}, Gemini and Claude models. Besides, we also consider 3D specialized models [2, 13, 49] fine-tuned on ScanQA. For OpenEQA, + +
ModelFrameBLEU-1BLEU-2METEORROUGE-LCIDEr
3D-Specific Models
ScanQA [2]-26.916.611.53055.4
ScanRefer+MCAN [49]-30.220.413.133.364.9
3D-LLM [13]-39.325.214.535.769.4
Open-source Multi-modal Models
LLaVA(Fine-tuned)6434.722.013.831.167.3
LLaVA+Coarse Correspondences6438.624.715.438.374.2
Proprietary Multi-modal Models
GPT-4V828.613.413.533.459.6
GPT-4V+Coarse Correspondences839.725.517.440.879.2
GPT-4O430.519.814.836.172.2
GPT-4O+Coarse Correspondences435.425.518.042.687.0
+ +Table 1. Comparison on ScanQA validation set. We conduct experiments on the ScanQA validation set to demonstrate the effectiveness of COARSE CORRESPONDENCES with different MLLMs. Our method enables both proprietary models and open-source models to surpass all 3D-specific models. + +we compare against language-only models to account for language bias (LLaMA2 [40]), commonly used general-purpose multimodal LLMs (GPT-4 [29], Claude3 [1], Gemini-Pro [38], GPT-4V with 15 and 50 frames. + +Metrics. For ScanQA, following prior works, we adopt BLEU [31], METEOR [5], ROUGE-L [20], and CIDEr [41] as our evaluation metrics. For OpenEQA, we follow their evaluation approach by using GPT-4 to compare the generated answers with the ground-truth answers and assign a score. We report the average score across all questions. + +Results. For ScanQA, as shown in Table 1, compared to raw input, COARSE CORRESPONDENCES consistently improves the overall performance of different proprietary models. For instance, on the strongest model, GPT-40, COARSE CORRESPONDENCES brings improvements of 5.7 BLEU-2, 3.2 METEOR, 6.5 ROUGE-L, and 15 CIDEr points. Compared to methods that are specifically designed for 3D understanding tasks, fine-tuned with specialized 3D SFT data, or even those that use 3D point clouds instead of 2D images as input, we observe that a general-purpose MLLM can still outperform them, especially when enhanced with COARSE CORRESPONDENCES. Moreover, we found that this can be achieved using far fewer images as input. + +We also demonstrated the same conclusion on OpenEQA, as indicated in Table 2. By applying COARSE CORRESPON-DENCES, we significantly improved the performance of both GPT-4v and GPT-4o, achieving better results with fewer input images. These findings suggest that general-purpose MLLMs are indeed capable of understanding 3D space, and COARSE CORRESPON-DENCES can significantly enhance their spatial understanding while simultaneously reducing the number of views needed, which could lower the inference cost and make MLLMs more useful for embodied tasks. + +
ModelsFrameAccuracy
LLaMA2 [40]028.3
GPT-4 [29]033.5
Claude3 [1]2036.3
Gemini-Pro [38]1544.9
GPT-4V [27]1554.6
GPT-4V [27]5055.3
HumanFull86.8
GPT-4V844.8
GPT-4V+CC858.5
GPT-4O449.4
GPT-4O+CC459.1
+ +Table 2. Comparisons on EM-EQA setting of OpenEQA. Our method further enhances the embodied ability of MLLMs and exceeds previous methods by a large margin. + +# 3.2. Temporal understanding + +**Benchmarks.** We evaluated the improvements of our method for long video understanding using the EgoSchema dataset. Each video in EgoSchema is 3 minutes long, with a corresponding question that includes five multiple-choice options. These questions are designed to ensure that answering them requires viewing a sufficient number of frames from the video. Due to budget constraints, we limited our evaluation to 500 questions from the validation set. + +Baselines. The baseline methods we compared against include newly designed and trained model architectures specifically for long video understanding, such as LongViviT [30] and MC-ViT-L [4]. On the other hand, we also compared methods that rely on text-only foundation models (e.g., GPT-4), i.e., Socratic-based approaches [15, 50], which first use an + +
ModelsFrameAccuracy
LongViviT [30]25656.8
MC-ViT-L [4]128+62.6
LLoVi [50]18058.3
VideoAgent [42]8.460.2
MVU [33]1660.3
VideoAgent [10]-62.8
LangRepo [15]-66.2
GPT-4V864.2
GPT-4V+CC867.4
GPT-4O867.2
GPT-4O+CC873.2
+ +Table 3. Comparisons on EgoSchema validation set. COARSE CORRESPONDENCES improves existing MLLMs and surpasses previous fine-tuned models in a zero-shot manner. + +off-the-shelf image captioning model [51] to convert video frames into captions, and then prompt GPT-4 to answer questions based on those captions. Additionally, we compared agent-based methods [10, 42], which involve using GPT-4 alongside an image captioning model in an agent framework to iteratively perform a series of multi-step reasoning operations to understand long videos. In contrast to these approaches, our method is entirely based on an end-to-end general MLLM architecture, exploring how to further enhance its ability to understand long videos. + +Results. COARSE CORRESPONDENCES demonstrates state-of-the-art performance, significantly outperforming existing approaches in a zero-shot manner (Table 3). Compared to the original GPT-4o model, our method improves its performance by $6\%$ . Notably, our method uses far fewer frames than other approaches, yet achieves higher results compared to methods that use many more frames. It is also worth highlighting that even the original GPT-4o, when limited to just 8 frames, already serves as a very strong baseline. This indicates the potential of a sufficiently powerful general-purpose MLLM in long video understanding. We provide a detailed breakdown of COARSE CORRESPONDENCES's performance across different question types in the supplementary material. + +# 3.3. Navigation + +Building on the improvements in 3DQA and VideoQA, we hope that our method can also prove effective in embodied tasks such as navigation. Navigation requires an agent to understand 3D space, such as being able to determine the spatial relationship between objects in the instruction and itself, while also performing temporal reasoning to assess the progress toward completing the instruction. We consider conducting experiments on the VLN-CE benchmark [16], which is a continuous simulation environment for low-level action execution in indoor scenes. + +Setup. We adopt the val-unseen split from R2R [16] for evaluation. Unlike the previous QA tasks, where all images could be processed at once, in navigation tasks, each image is processed in an online fashion. Specifically, we feed in one image at each iteration of the conversation. Given the significant variation in viewpoints during navigation, we use SAMv2 [34], the state-of-the-art model for long-range tracking, to label each new input image based on episodic history. Then, using the prompted images, we induce the MLLM to output one of four actions at each step: FORWARD (distance), TURN-LEFT (rotation angle), TURN-RIGHT (rotation angle), and STOP. We follow NavGPT [52] to craft input prompts. Considering the high computational cost of navigation tasks, we selected 100 samples from the val-unseen split. Our primary goal is to demonstrate that our method can enhance GPT models' capabilities in zero-shot navigation tasks, which remains a significant challenge for various types of models. + +Metrics. We follow the standard VLN evaluation metrics to evaluate the navigation performance, including success rate (SR), oracle success rate (OS), success weighted by path length (SPL), trajectory length (TL), and navigation error from goal (NE). Note that an episode is considered successful if the agent calls the STOP action within $3\mathrm{m}$ of the goal in the VLN-CE. + +Results. As shown in Table 4, our method achieved improvements across all metrics. While GPT-4o performs impressively on many QA tasks, its zero-shot performance on navigation tasks is relatively low. This may partly be due to the lack of specialized training on action data, making it less accurate in outputs such as determining how many meters to move forward. However, our method reveals another dimension of the problem: MLLMs' understanding of the 3D spacetime in which they operate can be further enhanced. This is evidenced by the significant improvements in navigation when using COARSE CORRESPONDENCES. We believe that our approach holds great potential for embodied tasks, which can be explored in future research. + +# 4. Prompting Open Models + +We further validate the effectiveness of our COARSE CORRESPONDENCES on open-source models. Our primary goal is to demonstrate that our method not only enhances proprietary models but also yields general improvements across various MLLMs. However, directly prompting open-source models presents two challenges: first, these models lack the ability to interpret prompting marks as effectively as proprietary models; second, they currently cannot handle multiple image inputs well. Thus, we first fine-tune the open-source models to understand prompting marks and manage multiple image inputs, enabling COARSE CORRESPONDENCES to enhance these models' environmental understanding capabilities at the inference stage. In this work, we primarily focus + +
MethodsSuccess Rate ↑Oracle Success Rate ↑Success weighted by Path Length ↑Trajectory Length ↑Navigation Error ↓
GPT-4O12.0018.0010.377.318.49
GPT-4O+CC23.0029.0021.038.127.37
+ +Table 4. Comparison on Navigation task. COARSE CORRESPONDENCES improve GPT-4o's performance on R2R dataset for different evaluation metrics. + +
MethodAcc
LLaVA(w/o CC during finetune)36.0
LLaVA+CC(train only)37.17
LLaVA+CC(train & test)39.13
+ +Table 5. Comparisons on SQA3D dataset. We tested the effectiveness of COARSE CORRESPONDENCES on the out-of-domain SQA3D dataset. With the same training data, we found that using COARSE CORRESPONDENCES not only improves performance during inference but also enhances zero-shot evaluation when applied solely during the training phase. + +on proving that COARSE CORRESPONDENCES improves open-source models' understanding of 3D space. + +We start with the LLaVA model [22] and fine-tune it using a dataset comprising approximately 1.2 million samples of image and video data. Notably, the ScanQA dataset is the only 3D-related dataset in this collection. COARSE CORRESPONDENCES is applied exclusively to the ScanQA data, while other data maintain their original format. Fine-tuning on such a dataset enables the open-source LLaVA model to better interpret overlayed marks on images and process multiple image inputs. + +In-domain Evaluation. We first evaluate whether COARSE CORRESPONDENCES can improve open-source models on the ScanQA validation set. As shown in Table 1, our method significantly enhances the model's 3D spatial understanding compared to the baseline fine-tuned on the same data without using COARSE CORRESPONDENCES, even outperforming previous VLMs [13] specifically designed for 3D tasks, which incorporate specialized architectural designs and are fine-tuned on substantially larger amounts of 3D-related data. This in-domain evaluation demonstrates that COARSE CORRESPONDENCES is also effective for open MLLMs. + +Out-domain Evaluation. To further demonstrate the generalizability of our method, we conduct experiments on the SQA3D dataset to evaluate whether COARSE CORRESPONDENCES can also provide improvements in a zero-shot setting. As shown in Table 5, on this previously unseen dataset, COARSE CORRESPONDENCES still outperforms the baseline simply fine-tuned on the same data, proving that our method can enhance open-source models' understanding of 3D space on out-of-domain datasets. + +Notably, even without using CoARSE CORRESPON- + +DENCES during inference, simply applying it during training already yields improvements. This highlights that our method is effective not only as a prompting technique for inference but also holds potential as a data augmentation method during training, a prospect worth exploring further in future work. + +# 5. Analysis + +Here, we explore the various design decisions in our method. Although the study is conducted only on ScanQA, applying the design choices derived from ScanQA to all other benchmarks also yields significant performance improvements. This strongly demonstrates the generalizability of these design choices. Additionally, we further analyze the impact of our method on mitigating the camera motion bias present in MLLMs. + +How does COARSE CORRESPONDENCES differ from other visual prompting methods? Our proposed method calculates and highlights correspondences between images, aiming to elicit 3D and temporal understanding. Other visual prompting methods (namely Set-of-Mark [48], 3DAxiesPrompts [21], and Chain-of-thought [43]) can also be viewed as alternative prompting methods. Given that the ground-truth answers in existing benchmarks are relatively brief, we selected a scene from ScanQA and manually designed a new question. We qualitatively compare COARSE CORRESPONDENCES against other prompting methods on this new question, as shown in Figure 2. + +The orange part of Figure 2 shows our Coarse Correspondence labels are recognized by GPT-4V. The output answer provides evidence that our coarse correspondence helps GPT-4V develop a mental 3D model of the scene. Set-of-Marks provides no spatial corresponding information and therefore is unhelpful. The Axis labels in 3DAxies can be easily misrecognized by GPT-4V, leading to misleading spatial information. Though Chain-of-Thought helps identify objects, it fails to resolve the "spatial perspective-taking" issue. + +Why use coarse instead of dense correspondences? Instead of filtering and retaining only a handful of coarse correspondences, one ablation we considered is the possibility of using all dense correspondence. Unfortunately, we find that excessively overlaying too many instance marks can degrade performance (Figure 3) as they occlude the visual content in the images. + +System: You are an AI with the ability to analyze a series of images, each representing a different perspective of a single scene. [Prompt-about-Marks]. Your task is to construct a 3D understanding based on these images. + +User: You are sitting on the sofa and the electric fan is on your left. Describe the location of the room door from your perspective: A. to the front left of you; B. to the front right of you; C. to the back left of you; D. to the back right of you. + +# +Coarse Correspondence (Ours) + +![](images/0fe6c28e518bdc27bd7fb5f4d1280aa2ffeb6fbf693cd70bd02e580b692921df.jpg) +View 1 + +![](images/2ff9112e172f2835a79129ddc506b432ca7aebcb262c1bd791444c887217654b.jpg) +View 2 + +GPT-4V: The first image ... a door marked with number 7, a fan marked with number 39, and items on shelves marked with numbers 11 and 37. In the second image, ... the fan again at number 39, and a sofa marked with number 3(√). + +We can infer that the door and the fan are on the same side of the room(✓). Since the fan appears to the left in the first image and is also on the left when sitting on the sofa as per the second image and your description, the door must be to the front of the sofa, as it is not visible in the second image from sofa's perspective(✓). By 'sitting on the sofa' as instructed, and with the electric fan to the left, I can determine the door is A. to the front left of you (✓). + +# +Set-of-Mark + +![](images/8c7e9682a85506eb5ffaf9bd9bde061a44e7719f499a7dc6f7caf310a7ef0c17.jpg) +View 1 + +![](images/bf90788f5d01e2b03543d31adaae137c152a0b3b8f0f2efb3a29f6a5eb2267b2.jpg) +View 2 + +GPT-4V: The fan is seen on the right side from the sofa's perspective, its leftward when sitting means the orientation is flipped(√). Therefore, the door that is visible on the right in the first image would be D. behind and to the right when seated on the sofa(X). + +# +3DAxesPrompts + +![](images/79719678b177040e7303dce4bcbabb3c35fd5cb7b4f41a8efd9c81bd658259a4.jpg) +View 1 + +![](images/8b39bec3be4d458965620e2b801246b81dadc9a99a50c8f4e86746f53bbbc7a0.jpg) +View 2 +Figure 2. Comparison of different prompting method. Our proposed Coarse Correspondences successfully guided GPT-40 to understand 3D spatial relationships and generate the right answer. Other existing prompting method including image-based Set-of-Marks, 3DAxies and text-based Chain-of-Thought failed to answer correctly. + +GPT-4V: The first image shows ... and a coordinate system with X-axis (horizontally rightwards) $(\mathbf{X})$ and Z-axis (vertically upwards). The second image shows a sofa on the left with a coordinate system's Y-axis (horizontally extending away from the viewer) $(\mathbf{X})$ and Z-axis as before. When sitting on the sofa with the fan to your left, the door would be B. diagonally opposite to you on the right side $(\mathbf{X})$ . + +# +Chain-of-Thought + +System: Your response should be as follows. +1)Identification: Identify all labeled objects. +2)Analysis: Analyze the objects and their positions construct a 3D model. + +3)Synthesis: Combine the info above, form a comprehensive understanding. + +4)Response: Provide a concise answer to the question based on your comprehensive 3D modelling. + +# GPT-4V: + +Identification: Door, shelves, sofa,... $(\checkmark)$ Analysis: in image 1 the door is on the right with the shelves adjacent to the door ... In image 2 the sofa is ... $(\checkmark)$ + +Synthesis: When sitting on the sofa, the shelves are on the right(X). The door is next to the shelves. + +Response: B. (X). + +How large should the marks be? We inject the correspondences into MLLMs by overlaying the marks into images. We empirically find an optimal mark size (where 'px' represents the mark's diameter in pixels) in Figure 3. Marks that are too small tend to be ignored while those that are too large occlude visual content. + +What shape should the marks be? We further studied the appearance of the marks. In addition to red circles with white text, we experimented with adding segmentation outlines and segmentation masks. As shown in Figure 3, using segmentation outlines enhances object grounding. However, using segmentation marks occludes visual content and reduces performance. + +Improvement of Camera Motion Invariance with CoARSE CORRESPONDENCES. A crucial aspect of strong spatial reasoning is the model's ability to maintain an understanding of spatial relationships invariant to camera motion; in other words, whether the image sequence is captured from left-to-right or right-to-left, the model's comprehension of static space should remain consistent. However, current benchmarks lack annotations regarding whether the 3D space scan was conducted from left-to-right or right-to-left, making it challenging to analyze the effect of camera motion on MLLMs' spatial understanding. To address this, + +we curated a small diagnostic benchmark specifically for this analysis. This benchmark includes 10 scenes captured from left-to-right, with five questions per scene focusing on left-right spatial relationships between objects—fundamental aspects of spatial understanding. We ensured that all questions required information across multiple frames and could not be answered from a single frame alone. Ideally, regardless of whether the image sequence is input in the original or reversed order, the output should remain consistent. We leave additional details on data curation in the supplementary material. + +To mitigate randomness, each question was tested 20 times, yielding a total of 1000 trials, and we report the average accuracy in Table 6. We then reversed the sequence order to simulate a right-to-left capture and repeated 1000 trials with this reversed input, reporting the average accuracy for this reverse sequence. Finally, we report the harmonic mean of the average accuracies for the forward and reverse inputs. + +Without CoARSE CORRESPONDENCES, the accuracy in understanding spatial relationships in right-to-left sequences was significantly lower than in left-to-right sequences. With CoARSE CORRESPONDENCES, not only did MLLMs achieve a substantial increase in spatial un + +![](images/bca43439fed192cbc8fcb8351b17493dc4145d1ea329596027de5730b2e276f0.jpg) +Figure 3. Ablations on different design choices of COARSE CORRESPONDENCES. We studied the impact of the number, size, and type of marks on performance. All experiments were conducted on ScanQA using GPT-4V. + +![](images/d9e791dfd0896750629d08fd18032293b53343eaf54d936d0f28d9141647c815.jpg) + +![](images/42eac979b8b802b8830f4100c9584b88c550e7117c6b8c35f934381261188779.jpg) + +
ModelsFrameForwardReverseHarmonic Mean
GPT-4O458.050.453.9
GPT-4O+CC471.271.271.2
+ +Table 6. Effect of COARSE CORRESPONDENCES on Camera Motion Bias. COARSE CORRESPONDENCES shows strong capability of enhancing 3D spatial understanding of MLLMs. It can also ease the camera motion bias of current MLLMs. + +derstanding accuracy with the original sequence $(+13.2\%)$ , but more importantly, reversing the input order no longer drastically impacted the model's comprehension of static spatial relationships. Specifically, COARSE CORRESPON-DENCES improved the harmonic mean of forward and reverse sequence accuracy by $17.3\%$ . This demonstrates that our method enables MLLMs to perform spatial reasoning invariant to camera motion, thus achieving more robust spatial understanding capabilities. + +# 6. Related work + +Multimodal language models Multimodal LLMs [3, 23] integrate vision encoders [32] into large LLMs [8, 40], enabling direct reasoning over visual input. Proprietary models like GPT-4 [27], Gemini [38], and Claude [1], along with open-source models such as the LLaVA series [23] and BLIP series [18], have made significant strides in 2D vision-language tasks like image captioning [7] and visual question answering (VQA) [11, 14]. However, for applications in real-world scenarios like autonomous driving [39] and robotics [46], MLLMs still require stronger spatial-temporal reasoning capabilities. Recent work [19] has enhanced MLLMs' capacity to process multiple image frames and understand temporal dynamics, achieving notable improvements in video dense captioning [17] and videoQA [12, 44]. Nonetheless, MLLMs still face challenges in understanding complex 3D spatial relations [24] and long-term temporal dynamics [25] present in image sequences. Many + +approaches have introduced specialized architectures [13] or fine-tuning methods targeting spatial and temporal reasoning individually. Our goal, however, is to jointly enhance spatial-temporal reasoning within the existing general-purpose MLLM framework. By leveraging off-the-shelf tracking models to extract instance-level correspondences, we achieve this in a training-free manner. + +Visual prompting. Effective prompting has been widely proven to improve LLMs across multiple domains. Methods, such as chain-of-thought prompting [43], force the model to reason before answering a question. In the context of visual prompting, we can enhance MLLMs' grounding abilities [36, 48], their ability to use numerical expressions to describe 3D object relationships [21], and their capability to generate outputs for robotics control [26], all using single-image inputs. Unlike these approaches, COARSE CORRESPONDENCES is designed to enhance MLLMs' spatial-temporal reasoning capabilities for interpreting complex spacetime structures represented in image sequences. + +# 7. Conclusion. + +We propose a framework called COARSE CORRESPON-DENCES prompting. By using off-the-shelf video tracking models to obtain class-agnostic, instance-level correspondences and conveying this information to MLLMs through visual prompting, we discovered that this simple method, using only 2D images as input—without any specialized architectural design or task-specific SFT—can effectively enhance MLLMs' spatial-temporal reasoning. This improvement extends to embodied tasks like navigation. Our method not only works on proprietary models but also generalizes to open-source models, and it performs well on both in-domain and out-of-domain datasets. Moreover, it enhances not just inference but also training. Further analysis shows that our method helps MLLMs become more robust to camera motion bias. + +# Acknowledgement + +We thank Xiaojuan Wang, Ruitao Zhang, Yuntian Deng for helpful discussions, feedback and collecting data. This project is partially funded by Amazon Science. This paper was partially supported by a gift from Ai2 and DARPA Transfer from Imprecise and Abstract Models to Autonomous Technologies (TIAMAT) program No. HR00112490422. Its contents are solely the responsibility of the authors and do not necessarily represent the official views of DARPA. + +# References + +[1] AI Anthropic. The claude 3 model family: Opus, sonnet, haiku. Claude-3 Model Card, 2024. 4, 8 +[2] Daichi Azuma, Taiki Miyanishi, Shuhei Kurita, and Motoaki Kawanabe. Scanqa: 3d question answering for spatial scene understanding. In proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 19129-19139, 2022. 2, 3, 4 +[3] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond, 2023. 8 +[4] Ivana Balažević, Yuge Shi, Pinelopi Papalampidi, Rahma Chaabouni, Skanda Koppula, and Olivier J. Henaff. Memory consolidation enables long-context video understanding, 2024. 1, 4, 5 +[5] Satanjeev Banerjee and Alon Lavie. Meteor: An automatic metric for mt evaluation with improved correlation with human judgments. In Proceedings of the acl workshop on intrinsic and extrinsic evaluation measures for machine translation and/or summarization, pages 65-72, 2005. 4 +[6] Boyuan Chen, Zhuo Xu, Sean Kirmani, Brian Ichter, Danny Driess, Pete Florence, Dorsa Sadigh, Leonidas Guibas, and Fei Xia. Spatialvlm: Endowing vision-language models with spatial reasoning capabilities. arXiv preprint arXiv:2401.12168, 2024. 1 +[7] Xinlei Chen, Hao Fang, Tsung-Yi Lin, Ramakrishna Vedantam, Saurabh Gupta, Piotr Dollar, and C. Lawrence Zitnick. Microsoft coco captions: Data collection and evaluation server, 2015. 8 +[8] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhanghao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E. Gonzalez, Ion Stoica, and Eric P. Xing. Vicuna: An open-source chatbot impressing gpt-4 with $90\%$ chatgpt quality, 2023. 8 +[9] Jiafei Duan, Samson Yu, Hui Li Tan, Hongyuan Zhu, and Cheston Tan. A survey of embodied ai: From simulators to research tasks. IEEE Transactions on Emerging Topics in Computational Intelligence, 6(2):230-244, 2022. 3 +[10] Yue Fan, Xiaojian Ma, Rujie Wu, Yuntao Du, Jiaqi Li, Zhi Gao, and Qing Li. Videoagent: A memory-augmented multimodal agent for video understanding. arXiv preprint arXiv:2403.11481, 2024.5 +[11] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Batra, and Devi Parikh. Making the v in vqa matter: Elevating + +the role of image understanding in visual question answering, 2017. 8 +[12] Madeleine Grunde-McLaughlin, Ranjay Krishna, and Maneesh Agrawala. Agqa: A benchmark for compositional spatio-temporal reasoning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11287-11297, 2021. 8 +[13] Yining Hong, Haoyu Zhen, Peihao Chen, Shuhong Zheng, Yilun Du, Zhenfang Chen, and Chuang Gan. 3d-llm: Injecting the 3d world into large language models. Advances in Neural Information Processing Systems, 36, 2024. 1, 3, 4, 6, 8 +[14] Drew A. Hudson and Christopher D. Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering, 2019. 8 +[15] Kumara Kahatapitiya, Kanchana Ranasinghe, Jongwoo Park, and Michael S. Ryoo. Language repository for long video understanding, 2024. 1, 4, 5 +[16] Jacob Krantz, Erik Wijmans, Arjun Majumdar, Dhruv Batra, and Stefan Lee. Beyond the nav-graph: Vision-and-language navigation in continuous environments. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXVIII 16, pages 104-120. Springer, 2020. 2, 3, 5 +[17] Ranjay Krishna, Kenji Hata, Frederic Ren, Li Fei-Fei, and Juan Carlos Niebles. Dense-captioning events in videos, 2017. 8 +[18] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip2: Bootstrapping language-image pre-training with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597, 2023. 8 +[19] Bin Lin, Bin Zhu, Yang Ye, Munan Ning, Peng Jin, and Li Yuan. Video-llava: Learning united visual representation by alignment before projection. arXiv preprint arXiv:2311.10122, 2023. 8 +[20] Chin-Yew Lin. Rouge: A package for automatic evaluation of summaries. In Text summarization branches out, pages 74-81, 2004. 4 +[21] Dingning Liu, Xiaomeng Dong, Renrui Zhang, Xu Luo, Peng Gao, Xiaoshui Huang, Yongshun Gong, and Zhihui Wang. 3daxiesprompts: Unleashing the 3d spatial task capabilities of gpt-4v. arXiv preprint arXiv:2312.09738, 2023. 6, 8 +[22] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. NeurIPS, 36, 2024. 2, 6 +[23] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 8 +[24] Arjun Majumdar, Anurag Ajay, Xiaohan Zhang, Pranav Putta, Sriram Yenamandra, Mikael Henaff, Sneha Silwal, Paul Mcvay, Oleksandr Maksymets, Sergio Arnaud, Karmesh Yadav, Qiyang Li, Ben Newman, Mohit Sharma, Vincent Berges, Shiqi Zhang, Pulkit Agrawal, Yonatan Bisk, Dhruv Batra, Mrinal Kalakrishnan, Franziska Meier, Chris Paxton, Sasha Sax, and Aravind Rajeswaran. Openeqa: Embodied question answering in the era of foundation models. In Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 1, 2, 3, 8 + +[25] Karttikeya Mangalam, Raiymbek Akshulakov, and Jitendra Malik. Egoschema: A diagnostic benchmark for very long-form video language understanding, 2023. 2, 3, 8 +[26] Soroush Nasiriany, Fei Xia, Wenhao Yu, Ted Xiao, Jacky Liang, Ishita Dasgupta, Annie Xie, Danny Driess, Ayzaan Wahid, Zhuo Xu, et al. Pivot: Iterative visual prompting elicits actionable knowledge for vlms. arXiv preprint arXiv:2402.07872, 2024. 8 +[27] OpenAI. Gpt-4v(ision) system card. OpenAI Blog, 2023. 4, 8 +[28] OpenAI. Hello gpt-4o. https://openai.com/index/hello-gpt-4o/, 2024. Accessed: 2024-05-22.1 +[29] OpenAI, Josh Achiam, and et al. Gpt-4 technical report, 2024. 4 +[30] Pinelopi Papalampidi, Skanda Koppula, Shreya Pathak, Justin Chiu, Joe Heyward, Viorica Patraucean, Jiajun Shen, Antoine Miech, Andrew Zisserman, and Aida Nematzdeh. A simple recipe for contrastively pre-training video-first encoders beyond 16 frames, 2023. 1, 4, 5 +[31] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th annual meeting of the Association for Computational Linguistics, pages 311-318, 2002. 4 +[32] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 8 +[33] Kanchana Ranasinghe, Xiang Li, Kumara Kahatapitiya, and Michael S. Ryoo. Understanding long videos in one multimodal language model pass, 2024. 5 +[34] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, et al. Sam 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714, 2024. 5 +[35] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4104-4113, 2016. 2 +[36] Aleksandar Shtedritski, Christian Rupprecht, and Andrea Vedaldi. What does clip know about a red circle? visual prompt engineering for vlms. arXiv preprint arXiv:2304.06712, 2023. 8 +[37] Simranjit Singh, Georgios Pavlakos, and Dimitrios Stamoulis. Evaluating zero-shot gpt-4v performance on 3d visual question answering benchmarks. arXiv preprint arXiv:2405.18831, 2024. 1 +[38] Gemini Team, Rohan Anil, and et al. Gemini: A family of highly capable multimodal models, 2024. 4, 8 +[39] Xiaoyu Tian, Junru Gu, Bailin Li, Yicheng Liu, Chenxu Hu, Yang Wang, Kun Zhan, Peng Jia, Xianpeng Lang, and Hang Zhao. Drivevlm: The convergence of autonomous driving and large vision-language models, 2024. 8 +[40] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: + +Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023.4,8 +[41] Ramakrishna Vedantam, C Lawrence Zitnick, and Devi Parikh. Cider: Consensus-based image description evaluation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4566-4575, 2015. 4 +[42] Xiaohan Wang, Yuhui Zhang, Orr Zohar, and Serena Yeung-Levy. Videoagent: Long-form video understanding with large language model as agent, 2024. 5 +[43] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Brian Ichter, Fei Xia, Ed Chi, Quoc Le, and Denny Zhou. Chain-of-thought prompting elicits reasoning in large language models, 2023. 6, 8 +[44] Junbin Xiao, Xindi Shang, Angela Yao, and Tat-Seng Chua. Next-qa:next phase of question-answering to explaining temporal actions, 2021. 8 +[45] Runsen Xu, Xiaolong Wang, Tai Wang, Yilun Chen, Jiangmiao Pang, and Dahua Lin. Pointllm: Empowering large language models to understand point clouds. In European Conference on Computer Vision, pages 131-147. Springer, 2025. 1 +[46] Jingkang Yang, Yuhao Dong, Shuai Liu, Bo Li, Ziyue Wang, Chencheng Jiang, Haoran Tan, Jiamu Kang, Yuanhan Zhang, Kaiyang Zhou, and Ziwei Liu. Octopus: Embodied vision-language programmer from environmental feedback, 2023. 8 +[47] Jinyu Yang, Mingqi Gao, Zhe Li, Shang Gao, Fangjing Wang, and Feng Zheng. Track anything: Segment anything meets videos. arXiv preprint arXiv:2304.11968, 2023. 3 +[48] Jianwei Yang, Hao Zhang, Feng Li, Xueyan Zou, Chunyuan Li, and Jianfeng Gao. Set-of-mark prompting unleashes extraordinary visual grounding in gpt-4v. arXiv preprint arXiv:2310.11441, 2023. 6, 8 +[49] Zhou Yu, Jun Yu, Yuhao Cui, Dacheng Tao, and Qi Tian. Deep modular co-attention networks for visual question answering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6281–6290, 2019. 3, 4 +[50] Ce Zhang, Taixi Lu, Md Mohaiminul Islam, Ziyang Wang, Shoubin Yu, Mohit Bansal, and Gedas Bertasius. A simple llm framework for long-range video question-answering, 2024. 1, 4, 5 +[51] Yue Zhao, Ishan Misra, Philipp Krahenbuhl, and Rohit Girdhar. Learning video representations from large language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6586-6597, 2023. 5 +[52] Gengze Zhou, Yicong Hong, and Qi Wu. Navgpt: Explicit reasoning in vision-and-language navigation with large language models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 7641-7649, 2024. 5 \ No newline at end of file diff --git a/2025/Coarse Correspondences Boost Spatial-Temporal Reasoning in Multimodal Language Model/images.zip b/2025/Coarse Correspondences Boost Spatial-Temporal Reasoning in Multimodal Language Model/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..8520ef7fea401149533c739beb31cfe128d9aba8 --- /dev/null +++ b/2025/Coarse Correspondences Boost Spatial-Temporal Reasoning in Multimodal Language Model/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:194e8af3b5bdae9c502d015f5ece5b2efc0ea0af8d9e5bf0d2c95e604c404649 +size 343705 diff --git a/2025/Coarse Correspondences Boost Spatial-Temporal Reasoning in Multimodal Language Model/layout.json b/2025/Coarse Correspondences Boost Spatial-Temporal Reasoning in Multimodal Language Model/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..94837a921116b7011e764a3d8975ad54b4201d9b --- /dev/null +++ b/2025/Coarse Correspondences Boost Spatial-Temporal Reasoning in Multimodal Language Model/layout.json @@ -0,0 +1,8167 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 102, + 103, + 510, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 103, + 510, + 140 + ], + "spans": [ + { + "bbox": [ + 102, + 103, + 510, + 140 + ], + "type": "text", + "content": "COARSE CORRESPONDENCES Boost Spatial-Temporal Reasoning in Multimodal Language Model" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 160, + 555, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 160, + 555, + 232 + ], + "spans": [ + { + "bbox": [ + 61, + 160, + 555, + 232 + ], + "type": "text", + "content": "Benlin Liu1*, Yuhao Dong2,3, Yiqin Wang2*, Zixian Ma1, Yansong Tang2, Luming Tang5, Yongming Rao3, Wei-Chiu Ma5,6, Ranjay Krishna1,5 \n1University of Washington, 2Tsinghua University, 3Tencent, 4Allen Institute for AI, 5Cornell University \ncoarse-correspondence.github.io" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 151, + 258, + 200, + 272 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 258, + 200, + 272 + ], + "spans": [ + { + "bbox": [ + 151, + 258, + 200, + 272 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 284, + 297, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 284, + 297, + 608 + ], + "spans": [ + { + "bbox": [ + 54, + 284, + 297, + 608 + ], + "type": "text", + "content": "Multimodal language models (MLLMs) are increasingly being applied in real-world environments, necessitating their ability to interpret 3D spaces and comprehend temporal dynamics. Current methods often rely on specialized architectural designs or task-specific fine-tuning to achieve this. We introduce COARSE CORRESPONDENCES, a simple lightweight method that enhances MLLMs' spatial-temporal reasoning with 2D images as input, without modifying the architecture or requiring task-specific fine-tuning. Our method uses a lightweight tracking model to identify primary object correspondences between frames in a video or across different image viewpoints, and then conveys this information to MLLMs through visual prompting. We demonstrate that this simple training-free approach brings substantial gains to GPT4-V/O consistently on four benchmarks that require spatial-temporal reasoning, including " + }, + { + "bbox": [ + 54, + 284, + 297, + 608 + ], + "type": "inline_equation", + "content": "+20.5\\%" + }, + { + "bbox": [ + 54, + 284, + 297, + 608 + ], + "type": "text", + "content": " improvement on ScanQA, " + }, + { + "bbox": [ + 54, + 284, + 297, + 608 + ], + "type": "inline_equation", + "content": "+9.7\\%" + }, + { + "bbox": [ + 54, + 284, + 297, + 608 + ], + "type": "text", + "content": " on OpenEQA's episodic memory subset, " + }, + { + "bbox": [ + 54, + 284, + 297, + 608 + ], + "type": "inline_equation", + "content": "+6.0\\%" + }, + { + "bbox": [ + 54, + 284, + 297, + 608 + ], + "type": "text", + "content": " on the long-form video benchmark EgoSchema, and " + }, + { + "bbox": [ + 54, + 284, + 297, + 608 + ], + "type": "inline_equation", + "content": "+11\\%" + }, + { + "bbox": [ + 54, + 284, + 297, + 608 + ], + "type": "text", + "content": " on the R2R navigation benchmark. Additionally, we show that COARSE CORRESPONDENCES can also enhance open-source MLLMs' spatial reasoning (by " + }, + { + "bbox": [ + 54, + 284, + 297, + 608 + ], + "type": "inline_equation", + "content": "+6.9\\%" + }, + { + "bbox": [ + 54, + 284, + 297, + 608 + ], + "type": "text", + "content": " on ScanQA) when applied in both training and inference and that the improvement can generalize to unseen datasets such as SQA3D " + }, + { + "bbox": [ + 54, + 284, + 297, + 608 + ], + "type": "inline_equation", + "content": "(+3.1\\%)" + }, + { + "bbox": [ + 54, + 284, + 297, + 608 + ], + "type": "text", + "content": ". Taken together, we show that COARSE CORRESPONDENCES effectively and efficiently boosts models' performance on downstream tasks requiring spatial-temporal reasoning." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 624, + 135, + 637 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 624, + 135, + 637 + ], + "spans": [ + { + "bbox": [ + 55, + 624, + 135, + 637 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 645, + 297, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 645, + 297, + 694 + ], + "spans": [ + { + "bbox": [ + 55, + 645, + 297, + 694 + ], + "type": "text", + "content": "Intelligence is multi-faceted. While multi-modal language models [28] have shown remarkable linguistic, logical and even mathematical intelligence, many remain doubtful about their visual and spatial intelligence. Despite their excel" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 260, + 555, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 260, + 555, + 344 + ], + "spans": [ + { + "bbox": [ + 313, + 260, + 555, + 344 + ], + "type": "text", + "content": "lent performance on visual-linguistic tasks, many recent works [24, 37] demonstrate that state-of-the-art MLLMs still struggle at 3D and long video understanding benchmarks, performing only marginally better than blind text-only baselines. These results suggest that spatial-temporal reasoning is a major bottleneck on MLLMs' path to general visual intelligence." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 346, + 556, + 466 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 346, + 556, + 466 + ], + "spans": [ + { + "bbox": [ + 313, + 346, + 556, + 466 + ], + "type": "text", + "content": "To enhance MLLMs' 3D understanding, researchers have mainly explored three approaches: providing MLLMs with 3D data as input [45], designing specialized architectures for 3D tasks [13], or employing supervised fine-tuning with 3D data [6]. Similarly, to boost MLLMs' temporal understanding, prior works have proposed new model architectures designed for long video understanding [4, 30], or adopted Socratic-based methods [15, 50] (i.e., converting each frame of a video into text using a caption model, and then using text-only LLMs to summarize)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 469, + 557, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 469, + 557, + 577 + ], + "spans": [ + { + "bbox": [ + 313, + 469, + 557, + 577 + ], + "type": "text", + "content": "In contrast to prior works that separately enhance spatial and temporal reasoning, we propose COARSE CORRESPON-DENCES, a simple yet effective training-free visual prompting method to jointly boost spatial-temporal reasoning in MLLMs. COARSE CORRESPONDENCES uses a tracking model to extract object-level correspondences across multiple images, and then represent the most salient correspondence relationships on the images through visual prompting. We refer to our method as Coarse because of the following:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 315, + 578, + 555, + 649 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 315, + 578, + 555, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 578, + 555, + 601 + ], + "spans": [ + { + "bbox": [ + 315, + 578, + 555, + 601 + ], + "type": "text", + "content": "1. We only visually prompt for instance-level correspondences and not point-level correspondences." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 315, + 603, + 553, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 603, + 553, + 626 + ], + "spans": [ + { + "bbox": [ + 315, + 603, + 553, + 626 + ], + "type": "text", + "content": "2. The instance-level correspondences are extracted using off-the-shelf tracking models." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 315, + 627, + 553, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 627, + 553, + 649 + ], + "spans": [ + { + "bbox": [ + 315, + 627, + 553, + 649 + ], + "type": "text", + "content": "3. We only visualize a handful of prominent corresponding instances." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 654, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 555, + 714 + ], + "type": "text", + "content": "Despite the obtained correspondence information not being perfectly precise, our method can still significantly boost MLLMs' spatial-temporal reasoning with only 2D image inputs and without any specialized architectural design or task-specific fine-tuning." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 703, + 134, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 134, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 134, + 712 + ], + "type": "text", + "content": "*Equal contribution." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3783" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 86, + 321, + 148 + ], + "blocks": [ + { + "bbox": [ + 61, + 86, + 321, + 148 + ], + "lines": [ + { + "bbox": [ + 61, + 86, + 321, + 148 + ], + "spans": [ + { + "bbox": [ + 61, + 86, + 321, + 148 + ], + "type": "image", + "image_path": "377e4dd11bf3046000ae07f806f03b2809b2434a209d339a8b154f4e392f34db.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 107, + 159, + 282, + 171 + ], + "lines": [ + { + "bbox": [ + 107, + 159, + 282, + 171 + ], + "spans": [ + { + "bbox": [ + 107, + 159, + 282, + 171 + ], + "type": "text", + "content": "(a) track objects in high frame rate videos" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 63, + 189, + 321, + 261 + ], + "blocks": [ + { + "bbox": [ + 63, + 189, + 321, + 261 + ], + "lines": [ + { + "bbox": [ + 63, + 189, + 321, + 261 + ], + "spans": [ + { + "bbox": [ + 63, + 189, + 321, + 261 + ], + "type": "image", + "image_path": "080e16c9a9846667bbfa8e4ac89e364d21d052fb5e7396f2ba56d5d9fcb3c009.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 74, + 270, + 315, + 282 + ], + "lines": [ + { + "bbox": [ + 74, + 270, + 315, + 282 + ], + "spans": [ + { + "bbox": [ + 74, + 270, + 315, + 282 + ], + "type": "text", + "content": "(b) construct coarse correspondence on sparsified views" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 343, + 72, + 547, + 245 + ], + "blocks": [ + { + "bbox": [ + 343, + 72, + 547, + 245 + ], + "lines": [ + { + "bbox": [ + 343, + 72, + 547, + 245 + ], + "spans": [ + { + "bbox": [ + 343, + 72, + 547, + 245 + ], + "type": "image", + "image_path": "5c32a801503c311eb192b487492c30b884979cec9dd567ae39d349c15132ed28.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 354, + 258, + 535, + 281 + ], + "lines": [ + { + "bbox": [ + 354, + 258, + 535, + 281 + ], + "spans": [ + { + "bbox": [ + 354, + 258, + 535, + 281 + ], + "type": "text", + "content": "(c) help MLLMs understand 3D space-time with prompted images" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 54, + 298, + 555, + 344 + ], + "lines": [ + { + "bbox": [ + 54, + 298, + 555, + 344 + ], + "spans": [ + { + "bbox": [ + 54, + 298, + 555, + 344 + ], + "type": "text", + "content": "Figure 1. We combined light-weight video tracking models and multimodal LLMs to achieve a better understanding of 3D spacetime. (a) We use a tracking model at a high frame rate to obtain instance segmentation masks for each frame. (b) Then, we sequentially sparsify input frames, select prominent coarse correspondences, and visualize the constructed coarse correspondences on the images. (c) Finally, we enable MLLMs to better understand 3D spacetime from the prompted images." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 354, + 297, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 354, + 297, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 354, + 297, + 713 + ], + "type": "text", + "content": "We have demonstrated substantial performance gains of COARSE CORRESPONDENCES through extensive experiments with both open-source and closed-source models across 6 benchmarks on spatial-temporal reasoning. For closed-source models, we apply COARSE CORRESPONDENCES on GPT4-V/O during inference and achieve compelling gains. First, on spatial reasoning, we show that our method significantly surpasses state-of-the-art models by " + }, + { + "bbox": [ + 56, + 354, + 297, + 713 + ], + "type": "inline_equation", + "content": "20.5\\%" + }, + { + "bbox": [ + 56, + 354, + 297, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 56, + 354, + 297, + 713 + ], + "type": "inline_equation", + "content": "9.7\\%" + }, + { + "bbox": [ + 56, + 354, + 297, + 713 + ], + "type": "text", + "content": " on ScanQA [2] and OpenEQA [24] respectively. Second, for long video understanding, our method leads to a " + }, + { + "bbox": [ + 56, + 354, + 297, + 713 + ], + "type": "inline_equation", + "content": "6\\%" + }, + { + "bbox": [ + 56, + 354, + 297, + 713 + ], + "type": "text", + "content": " gain in performance on the EgoSchema benchmark [25]. Notably, our method uses much fewer input images and, in a zero-shot manner, outperforms many fine-tuned models that use far more images. For example, on EgoSchema, COARSE CORRESPONDENCES surpasses state-of-the-art results with just 8 uniformly sampled frames from a 3-minute video, greatly reducing the computational costs of MLLMs compared to existing methods. In addition to 3D and video QA tasks, we further demonstrate that our method enhances models' performance on embodied tasks such as navigation [16], which require strong spatial-temporal reasoning, by " + }, + { + "bbox": [ + 56, + 354, + 297, + 713 + ], + "type": "inline_equation", + "content": "11\\%" + }, + { + "bbox": [ + 56, + 354, + 297, + 713 + ], + "type": "text", + "content": " in success rate on R2R. These results suggest that COARSE CORRESPONDENCES boosts MLLMs' spatial-temporal reasoning both effectively and efficiently. Last but not least, we experiment with open-source MLLMs [22] by applying COARSE CORRESPONDENCES in both instruction tuning and inference; again, our method shows performance gains against the baseline (by " + }, + { + "bbox": [ + 56, + 354, + 297, + 713 + ], + "type": "inline_equation", + "content": "6.9\\%" + }, + { + "bbox": [ + 56, + 354, + 297, + 713 + ], + "type": "text", + "content": " on ScanQA), and the improvement even generalizes to unseen datasets such as SQA3D " + }, + { + "bbox": [ + 56, + 354, + 297, + 713 + ], + "type": "inline_equation", + "content": "(+3.1\\%)" + }, + { + "bbox": [ + 56, + 354, + 297, + 713 + ], + "type": "text", + "content": ". These results suggest that COARSE COR" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 354, + 555, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 354, + 555, + 389 + ], + "spans": [ + { + "bbox": [ + 313, + 354, + 555, + 389 + ], + "type": "text", + "content": "RESPONDENCES works well universally with any model - both closed-source and open-source - that can take in multiple images and understand visual markers." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 390, + 556, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 390, + 556, + 475 + ], + "spans": [ + { + "bbox": [ + 313, + 390, + 556, + 475 + ], + "type": "text", + "content": "Overall, we want to highlight with this work that, despite its simplicity and being underestimated for semantic tasks in deep learning, visual correspondence can still bring significant utility to spatial-temporal reasoning in MLLMs, just as it has long contributed to 3D reconstruction [35]. We hope our work demonstrates the potential of leveraging general-purpose MLLMs to better understand our physical world." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 486, + 370, + 498 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 486, + 370, + 498 + ], + "spans": [ + { + "bbox": [ + 313, + 486, + 370, + 498 + ], + "type": "text", + "content": "2. Method" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 506, + 555, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 506, + 555, + 542 + ], + "spans": [ + { + "bbox": [ + 313, + 506, + 555, + 542 + ], + "type": "text", + "content": "We introduce CoARSE CORRESPONDERENCES, a visual prompting method that allows MLLMs to reason about 3D space and time." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 543, + 556, + 627 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 543, + 556, + 627 + ], + "spans": [ + { + "bbox": [ + 313, + 543, + 556, + 627 + ], + "type": "text", + "content": "Problem formulation. Given a question " + }, + { + "bbox": [ + 313, + 543, + 556, + 627 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}" + }, + { + "bbox": [ + 313, + 543, + 556, + 627 + ], + "type": "text", + "content": " and a sequence or set of observations in an environment " + }, + { + "bbox": [ + 313, + 543, + 556, + 627 + ], + "type": "inline_equation", + "content": "[I_1, \\ldots, I_n]" + }, + { + "bbox": [ + 313, + 543, + 556, + 627 + ], + "type": "text", + "content": ", our aim is to design a visual prompt " + }, + { + "bbox": [ + 313, + 543, + 556, + 627 + ], + "type": "inline_equation", + "content": "\\mathcal{P}(\\ldots)" + }, + { + "bbox": [ + 313, + 543, + 556, + 627 + ], + "type": "text", + "content": " that modifies the input image set. These image inputs don't have to be a video. They can also represent a set of images of a specific scene from multiple viewpoints. We evaluate the prompt by measuring its utility in prompting an MLLM " + }, + { + "bbox": [ + 313, + 543, + 556, + 627 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 313, + 543, + 556, + 627 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 370, + 635, + 496, + 667 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 635, + 496, + 667 + ], + "spans": [ + { + "bbox": [ + 370, + 635, + 496, + 667 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} [ I _ {1} ^ {\\prime}, \\dots , I _ {n} ^ {\\prime} ] = P ([ I _ {1}, \\dots , I _ {n} ]) \\\\ \\hat {\\mathcal {A}} = \\mathcal {M} (([ I _ {1} ^ {\\prime}, \\dots , I _ {n} ^ {\\prime} ]), \\mathcal {Q}) \\\\ \\end{array}", + "image_path": "8ac67807bc020410c5d9bef51ec3a3aec0439e12d8837e50acc8eb5f3b21f383.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 677, + 549, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 549, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 549, + 689 + ], + "type": "text", + "content": "We compare generated answer " + }, + { + "bbox": [ + 313, + 677, + 549, + 689 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{A}}" + }, + { + "bbox": [ + 313, + 677, + 549, + 689 + ], + "type": "text", + "content": " with the ground truth " + }, + { + "bbox": [ + 313, + 677, + 549, + 689 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 313, + 677, + 549, + 689 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": "In our framework, the MLLMs can be any general-purpose model without requiring any special architecture" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3784" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 109 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 109 + ], + "type": "text", + "content": "or training for spatial-temporal reasoning. Our aim is to develop a prompting strategy that allows models to improve such capabilities without any training (Figure 1)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 114, + 171, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 114, + 171, + 126 + ], + "spans": [ + { + "bbox": [ + 55, + 114, + 171, + 126 + ], + "type": "text", + "content": "Coarse Correspondence" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 132, + 297, + 178 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 132, + 297, + 178 + ], + "spans": [ + { + "bbox": [ + 55, + 132, + 297, + 178 + ], + "type": "text", + "content": "Our prompting method, COARSE CORRESPONDENCES, contains four steps: (1) tracking correspondences, (2) sparsify frames, (3) selecting, and (4) visualizing coarse correspondences." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 179, + 296, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 179, + 296, + 287 + ], + "spans": [ + { + "bbox": [ + 55, + 179, + 296, + 287 + ], + "type": "text", + "content": "(1) Tracking correspondences. Given " + }, + { + "bbox": [ + 55, + 179, + 296, + 287 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 179, + 296, + 287 + ], + "type": "text", + "content": " input images, " + }, + { + "bbox": [ + 55, + 179, + 296, + 287 + ], + "type": "inline_equation", + "content": "[I_1, \\ldots, I_n]" + }, + { + "bbox": [ + 55, + 179, + 296, + 287 + ], + "type": "text", + "content": ", we first use an off-the-shelf video object tracking model, such as Tracking Anything [47]. This model extracts class-agnostic instance segmentation masks " + }, + { + "bbox": [ + 55, + 179, + 296, + 287 + ], + "type": "inline_equation", + "content": "(M_1, \\ldots, M_n)" + }, + { + "bbox": [ + 55, + 179, + 296, + 287 + ], + "type": "text", + "content": " for each image. Each " + }, + { + "bbox": [ + 55, + 179, + 296, + 287 + ], + "type": "inline_equation", + "content": "M_i" + }, + { + "bbox": [ + 55, + 179, + 296, + 287 + ], + "type": "text", + "content": " is a " + }, + { + "bbox": [ + 55, + 179, + 296, + 287 + ], + "type": "inline_equation", + "content": "H \\times W" + }, + { + "bbox": [ + 55, + 179, + 296, + 287 + ], + "type": "text", + "content": " dimensional matrix where " + }, + { + "bbox": [ + 55, + 179, + 296, + 287 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 55, + 179, + 296, + 287 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 179, + 296, + 287 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 55, + 179, + 296, + 287 + ], + "type": "text", + "content": " are the height and width of the input image " + }, + { + "bbox": [ + 55, + 179, + 296, + 287 + ], + "type": "inline_equation", + "content": "I_i" + }, + { + "bbox": [ + 55, + 179, + 296, + 287 + ], + "type": "text", + "content": ". Each pixel location in " + }, + { + "bbox": [ + 55, + 179, + 296, + 287 + ], + "type": "inline_equation", + "content": "M_i" + }, + { + "bbox": [ + 55, + 179, + 296, + 287 + ], + "type": "text", + "content": " contains an instance ID, indicating which instance the pixel at that position belongs to within the image sequence." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 287, + 296, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 287, + 296, + 370 + ], + "spans": [ + { + "bbox": [ + 55, + 287, + 296, + 370 + ], + "type": "text", + "content": "(2) Sparsify frames. Since most MLLMs contain a large number of parameters, directly using them to process long image sequences is very computationally intensive. Additionally, proprietary MLLMs like GPT-4O can also incur significant costs if the number of image tokens that need to be processed increases. Reducing the number of input images might lose vital information necessary for MLLMs." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 371, + 296, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 371, + 296, + 502 + ], + "spans": [ + { + "bbox": [ + 55, + 371, + 296, + 502 + ], + "type": "text", + "content": "COARSE CORRESPONDENCES strikes a balance in this tradeoff by extracting meaningful video object tracks (a relatively cheaper operation) from high-frame-rate image sequences, and then samples a few image inputs along with the tracks, to retain—and even improve—performance while reducing the MLLM's computation cost. From this extracted video object tracks, we perform temporal downsampling, retaining only " + }, + { + "bbox": [ + 55, + 371, + 296, + 502 + ], + "type": "inline_equation", + "content": "m << n" + }, + { + "bbox": [ + 55, + 371, + 296, + 502 + ], + "type": "text", + "content": " uniformly sampled images and their corresponding masks, denoted as " + }, + { + "bbox": [ + 55, + 371, + 296, + 502 + ], + "type": "inline_equation", + "content": "[I_{s_1},\\dots,I_{s_m}]" + }, + { + "bbox": [ + 55, + 371, + 296, + 502 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 371, + 296, + 502 + ], + "type": "inline_equation", + "content": "[M_{s_1},\\dots,M_{s_m}]" + }, + { + "bbox": [ + 55, + 371, + 296, + 502 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 371, + 296, + 502 + ], + "type": "inline_equation", + "content": "s_i \\in \\{1,\\dots,n\\}" + }, + { + "bbox": [ + 55, + 371, + 296, + 502 + ], + "type": "text", + "content": ". This downsampling reduces the number of images we feed into " + }, + { + "bbox": [ + 55, + 371, + 296, + 502 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 55, + 371, + 296, + 502 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 502, + 297, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 502, + 297, + 622 + ], + "spans": [ + { + "bbox": [ + 55, + 502, + 297, + 622 + ], + "type": "text", + "content": "(3) Selecting coarse correspondences. Prompting an MLLM with all the detected correspondences results in information overload. In fact, our ablations (discussed in Sec 5) find that adding all the correspondences reduces the MLLM's performance. Therefore, we select a subset of prominent instances to retain. We select the prominent instances of the top-K objects that co-occur in the most number of frames. We first calculate the occurrence frequency and area sum of each unique instance ID in the retained " + }, + { + "bbox": [ + 55, + 502, + 297, + 622 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 55, + 502, + 297, + 622 + ], + "type": "text", + "content": " masks using the following equation:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 628, + 309, + 661 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 628, + 309, + 661 + ], + "spans": [ + { + "bbox": [ + 55, + 628, + 309, + 661 + ], + "type": "interline_equation", + "content": "\\mathcal {F} r e q (\\mathrm {I D}) = \\sum_ {i = s _ {1}} ^ {s _ {m}} \\mathbf {1} _ {\\{\\mathrm {I D} \\in M _ {i} \\}}, \\mathcal {A r e a} (\\mathrm {I D}) = \\sum_ {i = s _ {1}} ^ {s _ {m}} \\sum_ {p \\in M _ {i}} \\mathbf {1} _ {\\{\\mathrm {I D} = p \\}}.", + "image_path": "bc9aa7015072c5cdf5b204acc3f4982828f2340c8448c9c27c47528984cee70b.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "Then, we first sort all instance IDs in descending order based on " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{Freq}(\\mathrm{ID})" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": ". If there are ties, we further sort based on " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{Area}(\\mathrm{ID})" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": ". Finally, we retain the top " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": " instance IDs as tracklets, denoted as " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "[T_1, \\dots, T_k]" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": ", to visualize for MLLMs." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "type": "text", + "content": "(4) Visualizing coarse correspondences. For each set of obtained correspondence relationships, we visualize the correspondences directly in the image as a marker. Specifically, for each identified primary instance ID " + }, + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "type": "inline_equation", + "content": "T_{i}" + }, + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "type": "text", + "content": ", if it exists in the mask " + }, + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "type": "inline_equation", + "content": "M_{s_j}" + }, + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "type": "text", + "content": " of a retained image " + }, + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "type": "inline_equation", + "content": "I_{s_j}" + }, + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "type": "text", + "content": ", we overlay a mark with a fixed size and shape labeled with " + }, + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "type": "inline_equation", + "content": "T_{i}" + }, + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "type": "text", + "content": " at the position " + }, + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "type": "inline_equation", + "content": "(\\bar{x}_{ij},\\bar{y}_{ij})" + }, + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "type": "text", + "content": " on " + }, + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "type": "inline_equation", + "content": "I_{s_j}" + }, + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "type": "text", + "content": " to produce " + }, + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "type": "inline_equation", + "content": "I_{s_j}'" + }, + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "type": "text", + "content": ". The specific placement position can be easily obtained by:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 346, + 175, + 520, + 209 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 346, + 175, + 520, + 209 + ], + "spans": [ + { + "bbox": [ + 346, + 175, + 520, + 209 + ], + "type": "interline_equation", + "content": "\\left(\\bar {x} _ {i j}, \\bar {y} _ {i j}\\right) = \\frac {\\sum_ {(x , y)} (x , y) \\cdot \\mathbf {1} _ {\\{M _ {s _ {j}} (x , y) = T _ {i} \\}}}{\\sum_ {(x , y)} \\mathbf {1} _ {\\{M _ {s _ {j}} (x , y) = T _ {i} \\}}}", + "image_path": "e267c7049dfec12e5b9220c5bc824b80f4b034f9b6007731f1abcff7335acc91.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 215, + 555, + 286 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 215, + 555, + 286 + ], + "spans": [ + { + "bbox": [ + 313, + 215, + 555, + 286 + ], + "type": "text", + "content": "Naturally, we can overlay not just the markers but also the segmentation outlines or even the segmentation masks associated with each retained prominent instance. We explore these ablations later. In the end, we obtain the prompted image sequence " + }, + { + "bbox": [ + 313, + 215, + 555, + 286 + ], + "type": "inline_equation", + "content": "[I_1',\\dots,I_m']" + }, + { + "bbox": [ + 313, + 215, + 555, + 286 + ], + "type": "text", + "content": ", which is then used as the input to MLLMs." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 297, + 488, + 312 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 297, + 488, + 312 + ], + "spans": [ + { + "bbox": [ + 313, + 297, + 488, + 312 + ], + "type": "text", + "content": "3. Prompting Proprietary Models" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 318, + 556, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 318, + 556, + 509 + ], + "spans": [ + { + "bbox": [ + 313, + 318, + 556, + 509 + ], + "type": "text", + "content": "We first evaluated the utility of our COARSE CORRESPON-DENCES on multiple tasks using proprietary models, including understanding 3D space (ScanQA [2] and OpenEQA [24] in §3.1) as well as temporal events (EgoSchema [25] in §3.2). Building on the improvements our method brings to spatial-temporal reasoning, we further demonstrate that our method also delivers significant gains in navigation tasks (VLN-CE [16]). Across all these benchmarks, we augment proprietary MLLMs (e.g., GPT-4V and GPT-4O) with COARSE CORRESPON-DENCES and evaluate its zero-shot performance. We show that COARSE CORRESPON-DENCES significantly improves the base GPT models and can substantially surpass many current state-of-the-art methods that have undergone specialized fine-tuning, even while using much fewer images as input. All experiments were conducted using A100 80G GPUs." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 517, + 441, + 529 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 517, + 441, + 529 + ], + "spans": [ + { + "bbox": [ + 313, + 517, + 441, + 529 + ], + "type": "text", + "content": "3.1. Spatial understanding" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 534, + 556, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 556, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 556, + 665 + ], + "type": "text", + "content": "**Benchmarks.** The validation set of ScanQA dataset contains 4675 questions about 71 scenes. The questions require basic recognition, 3D localization, and 3D embodied capabilities [9]. It contains two ground-truth answers per question for evaluation with models that produce free-form answers. OpenEQA Dataset is an open-vocabulary dataset benchmarking spatial environment understanding and embodied reasoning. We evaluate on OpenEQA's EM-EQA data split, which contains over 1600 high-quality human-generated questions. The subset tests the episodic memory of an agent moving through a 3D environment over time." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 665, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 556, + 714 + ], + "type": "text", + "content": "Baselines. For ScanQA, we evaluate COARSE CORRESPON-DENCES by augmenting both GPT-4{V,O}, Gemini and Claude models. Besides, we also consider 3D specialized models [2, 13, 49] fine-tuned on ScanQA. For OpenEQA," + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3785" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 69, + 555, + 268 + ], + "blocks": [ + { + "bbox": [ + 58, + 69, + 555, + 268 + ], + "lines": [ + { + "bbox": [ + 58, + 69, + 555, + 268 + ], + "spans": [ + { + "bbox": [ + 58, + 69, + 555, + 268 + ], + "type": "table", + "html": "
ModelFrameBLEU-1BLEU-2METEORROUGE-LCIDEr
3D-Specific Models
ScanQA [2]-26.916.611.53055.4
ScanRefer+MCAN [49]-30.220.413.133.364.9
3D-LLM [13]-39.325.214.535.769.4
Open-source Multi-modal Models
LLaVA(Fine-tuned)6434.722.013.831.167.3
LLaVA+Coarse Correspondences6438.624.715.438.374.2
Proprietary Multi-modal Models
GPT-4V828.613.413.533.459.6
GPT-4V+Coarse Correspondences839.725.517.440.879.2
GPT-4O430.519.814.836.172.2
GPT-4O+Coarse Correspondences435.425.518.042.687.0
", + "image_path": "ba9d5ab8a0bed2386b4d5a405926f822e1a413ae807c564a1a2d896e0aa3849c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 55, + 270, + 555, + 304 + ], + "lines": [ + { + "bbox": [ + 55, + 270, + 555, + 304 + ], + "spans": [ + { + "bbox": [ + 55, + 270, + 555, + 304 + ], + "type": "text", + "content": "Table 1. Comparison on ScanQA validation set. We conduct experiments on the ScanQA validation set to demonstrate the effectiveness of COARSE CORRESPONDENCES with different MLLMs. Our method enables both proprietary models and open-source models to surpass all 3D-specific models." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 310, + 297, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 310, + 297, + 357 + ], + "spans": [ + { + "bbox": [ + 55, + 310, + 297, + 357 + ], + "type": "text", + "content": "we compare against language-only models to account for language bias (LLaMA2 [40]), commonly used general-purpose multimodal LLMs (GPT-4 [29], Claude3 [1], Gemini-Pro [38], GPT-4V with 15 and 50 frames." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 361, + 296, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 361, + 296, + 433 + ], + "spans": [ + { + "bbox": [ + 55, + 361, + 296, + 433 + ], + "type": "text", + "content": "Metrics. For ScanQA, following prior works, we adopt BLEU [31], METEOR [5], ROUGE-L [20], and CIDEr [41] as our evaluation metrics. For OpenEQA, we follow their evaluation approach by using GPT-4 to compare the generated answers with the ground-truth answers and assign a score. We report the average score across all questions." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 435, + 297, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 435, + 297, + 592 + ], + "spans": [ + { + "bbox": [ + 54, + 435, + 297, + 592 + ], + "type": "text", + "content": "Results. For ScanQA, as shown in Table 1, compared to raw input, COARSE CORRESPONDENCES consistently improves the overall performance of different proprietary models. For instance, on the strongest model, GPT-40, COARSE CORRESPONDENCES brings improvements of 5.7 BLEU-2, 3.2 METEOR, 6.5 ROUGE-L, and 15 CIDEr points. Compared to methods that are specifically designed for 3D understanding tasks, fine-tuned with specialized 3D SFT data, or even those that use 3D point clouds instead of 2D images as input, we observe that a general-purpose MLLM can still outperform them, especially when enhanced with COARSE CORRESPONDENCES. Moreover, we found that this can be achieved using far fewer images as input." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 594, + 297, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 594, + 297, + 713 + ], + "spans": [ + { + "bbox": [ + 54, + 594, + 297, + 713 + ], + "type": "text", + "content": "We also demonstrated the same conclusion on OpenEQA, as indicated in Table 2. By applying COARSE CORRESPON-DENCES, we significantly improved the performance of both GPT-4v and GPT-4o, achieving better results with fewer input images. These findings suggest that general-purpose MLLMs are indeed capable of understanding 3D space, and COARSE CORRESPON-DENCES can significantly enhance their spatial understanding while simultaneously reducing the number of views needed, which could lower the inference cost and make MLLMs more useful for embodied tasks." + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 322, + 308, + 545, + 472 + ], + "blocks": [ + { + "bbox": [ + 322, + 308, + 545, + 472 + ], + "lines": [ + { + "bbox": [ + 322, + 308, + 545, + 472 + ], + "spans": [ + { + "bbox": [ + 322, + 308, + 545, + 472 + ], + "type": "table", + "html": "
ModelsFrameAccuracy
LLaMA2 [40]028.3
GPT-4 [29]033.5
Claude3 [1]2036.3
Gemini-Pro [38]1544.9
GPT-4V [27]1554.6
GPT-4V [27]5055.3
HumanFull86.8
GPT-4V844.8
GPT-4V+CC858.5
GPT-4O449.4
GPT-4O+CC459.1
", + "image_path": "02d75a269fd06992a56cb26575913935846d3a345a5a37ec84417a37646a5d13.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 313, + 473, + 555, + 507 + ], + "lines": [ + { + "bbox": [ + 313, + 473, + 555, + 507 + ], + "spans": [ + { + "bbox": [ + 313, + 473, + 555, + 507 + ], + "type": "text", + "content": "Table 2. Comparisons on EM-EQA setting of OpenEQA. Our method further enhances the embodied ability of MLLMs and exceeds previous methods by a large margin." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 522, + 453, + 536 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 522, + 453, + 536 + ], + "spans": [ + { + "bbox": [ + 313, + 522, + 453, + 536 + ], + "type": "text", + "content": "3.2. Temporal understanding" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 543, + 556, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 543, + 556, + 639 + ], + "spans": [ + { + "bbox": [ + 313, + 543, + 556, + 639 + ], + "type": "text", + "content": "**Benchmarks.** We evaluated the improvements of our method for long video understanding using the EgoSchema dataset. Each video in EgoSchema is 3 minutes long, with a corresponding question that includes five multiple-choice options. These questions are designed to ensure that answering them requires viewing a sufficient number of frames from the video. Due to budget constraints, we limited our evaluation to 500 questions from the validation set." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 642, + 557, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 557, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 557, + 714 + ], + "type": "text", + "content": "Baselines. The baseline methods we compared against include newly designed and trained model architectures specifically for long video understanding, such as LongViviT [30] and MC-ViT-L [4]. On the other hand, we also compared methods that rely on text-only foundation models (e.g., GPT-4), i.e., Socratic-based approaches [15, 50], which first use an" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "3786" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 63, + 70, + 286, + 234 + ], + "blocks": [ + { + "bbox": [ + 63, + 70, + 286, + 234 + ], + "lines": [ + { + "bbox": [ + 63, + 70, + 286, + 234 + ], + "spans": [ + { + "bbox": [ + 63, + 70, + 286, + 234 + ], + "type": "table", + "html": "
ModelsFrameAccuracy
LongViviT [30]25656.8
MC-ViT-L [4]128+62.6
LLoVi [50]18058.3
VideoAgent [42]8.460.2
MVU [33]1660.3
VideoAgent [10]-62.8
LangRepo [15]-66.2
GPT-4V864.2
GPT-4V+CC867.4
GPT-4O867.2
GPT-4O+CC873.2
", + "image_path": "77aaadb0cb7fbb63c08d18703860b8998c944fc9ccb09d0f1a73969e38990284.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 236, + 295, + 269 + ], + "lines": [ + { + "bbox": [ + 55, + 236, + 295, + 269 + ], + "spans": [ + { + "bbox": [ + 55, + 236, + 295, + 269 + ], + "type": "text", + "content": "Table 3. Comparisons on EgoSchema validation set. COARSE CORRESPONDENCES improves existing MLLMs and surpasses previous fine-tuned models in a zero-shot manner." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 280, + 295, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 280, + 295, + 399 + ], + "spans": [ + { + "bbox": [ + 54, + 280, + 295, + 399 + ], + "type": "text", + "content": "off-the-shelf image captioning model [51] to convert video frames into captions, and then prompt GPT-4 to answer questions based on those captions. Additionally, we compared agent-based methods [10, 42], which involve using GPT-4 alongside an image captioning model in an agent framework to iteratively perform a series of multi-step reasoning operations to understand long videos. In contrast to these approaches, our method is entirely based on an end-to-end general MLLM architecture, exploring how to further enhance its ability to understand long videos." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 401, + 297, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 401, + 297, + 568 + ], + "spans": [ + { + "bbox": [ + 55, + 401, + 297, + 568 + ], + "type": "text", + "content": "Results. COARSE CORRESPONDENCES demonstrates state-of-the-art performance, significantly outperforming existing approaches in a zero-shot manner (Table 3). Compared to the original GPT-4o model, our method improves its performance by " + }, + { + "bbox": [ + 55, + 401, + 297, + 568 + ], + "type": "inline_equation", + "content": "6\\%" + }, + { + "bbox": [ + 55, + 401, + 297, + 568 + ], + "type": "text", + "content": ". Notably, our method uses far fewer frames than other approaches, yet achieves higher results compared to methods that use many more frames. It is also worth highlighting that even the original GPT-4o, when limited to just 8 frames, already serves as a very strong baseline. This indicates the potential of a sufficiently powerful general-purpose MLLM in long video understanding. We provide a detailed breakdown of COARSE CORRESPONDENCES's performance across different question types in the supplementary material." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 576, + 129, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 576, + 129, + 588 + ], + "spans": [ + { + "bbox": [ + 55, + 576, + 129, + 588 + ], + "type": "text", + "content": "3.3. Navigation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "text", + "content": "Building on the improvements in 3DQA and VideoQA, we hope that our method can also prove effective in embodied tasks such as navigation. Navigation requires an agent to understand 3D space, such as being able to determine the spatial relationship between objects in the instruction and itself, while also performing temporal reasoning to assess the progress toward completing the instruction. We consider conducting experiments on the VLN-CE benchmark [16], which is a continuous simulation environment for low-level action execution in indoor scenes." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 555, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 287 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 287 + ], + "type": "text", + "content": "Setup. We adopt the val-unseen split from R2R [16] for evaluation. Unlike the previous QA tasks, where all images could be processed at once, in navigation tasks, each image is processed in an online fashion. Specifically, we feed in one image at each iteration of the conversation. Given the significant variation in viewpoints during navigation, we use SAMv2 [34], the state-of-the-art model for long-range tracking, to label each new input image based on episodic history. Then, using the prompted images, we induce the MLLM to output one of four actions at each step: FORWARD (distance), TURN-LEFT (rotation angle), TURN-RIGHT (rotation angle), and STOP. We follow NavGPT [52] to craft input prompts. Considering the high computational cost of navigation tasks, we selected 100 samples from the val-unseen split. Our primary goal is to demonstrate that our method can enhance GPT models' capabilities in zero-shot navigation tasks, which remains a significant challenge for various types of models." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 288, + 556, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 288, + 556, + 371 + ], + "spans": [ + { + "bbox": [ + 313, + 288, + 556, + 371 + ], + "type": "text", + "content": "Metrics. We follow the standard VLN evaluation metrics to evaluate the navigation performance, including success rate (SR), oracle success rate (OS), success weighted by path length (SPL), trajectory length (TL), and navigation error from goal (NE). Note that an episode is considered successful if the agent calls the STOP action within " + }, + { + "bbox": [ + 313, + 288, + 556, + 371 + ], + "type": "inline_equation", + "content": "3\\mathrm{m}" + }, + { + "bbox": [ + 313, + 288, + 556, + 371 + ], + "type": "text", + "content": " of the goal in the VLN-CE." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 372, + 557, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 372, + 557, + 526 + ], + "spans": [ + { + "bbox": [ + 313, + 372, + 557, + 526 + ], + "type": "text", + "content": "Results. As shown in Table 4, our method achieved improvements across all metrics. While GPT-4o performs impressively on many QA tasks, its zero-shot performance on navigation tasks is relatively low. This may partly be due to the lack of specialized training on action data, making it less accurate in outputs such as determining how many meters to move forward. However, our method reveals another dimension of the problem: MLLMs' understanding of the 3D spacetime in which they operate can be further enhanced. This is evidenced by the significant improvements in navigation when using COARSE CORRESPONDENCES. We believe that our approach holds great potential for embodied tasks, which can be explored in future research." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 538, + 455, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 538, + 455, + 552 + ], + "spans": [ + { + "bbox": [ + 313, + 538, + 455, + 552 + ], + "type": "text", + "content": "4. Prompting Open Models" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 557, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 557, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 557, + 556, + 713 + ], + "type": "text", + "content": "We further validate the effectiveness of our COARSE CORRESPONDENCES on open-source models. Our primary goal is to demonstrate that our method not only enhances proprietary models but also yields general improvements across various MLLMs. However, directly prompting open-source models presents two challenges: first, these models lack the ability to interpret prompting marks as effectively as proprietary models; second, they currently cannot handle multiple image inputs well. Thus, we first fine-tune the open-source models to understand prompting marks and manage multiple image inputs, enabling COARSE CORRESPONDENCES to enhance these models' environmental understanding capabilities at the inference stage. In this work, we primarily focus" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3787" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 70, + 553, + 135 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 553, + 135 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 553, + 135 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 553, + 135 + ], + "type": "table", + "html": "
MethodsSuccess Rate ↑Oracle Success Rate ↑Success weighted by Path Length ↑Trajectory Length ↑Navigation Error ↓
GPT-4O12.0018.0010.377.318.49
GPT-4O+CC23.0029.0021.038.127.37
", + "image_path": "d600c6ff85301e1e8b6697f6962e931f2a485914d4e618f629fa573567421a43.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 56, + 170, + 294, + 232 + ], + "blocks": [ + { + "bbox": [ + 55, + 136, + 555, + 159 + ], + "lines": [ + { + "bbox": [ + 55, + 136, + 555, + 159 + ], + "spans": [ + { + "bbox": [ + 55, + 136, + 555, + 159 + ], + "type": "text", + "content": "Table 4. Comparison on Navigation task. COARSE CORRESPONDENCES improve GPT-4o's performance on R2R dataset for different evaluation metrics." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 170, + 294, + 232 + ], + "lines": [ + { + "bbox": [ + 56, + 170, + 294, + 232 + ], + "spans": [ + { + "bbox": [ + 56, + 170, + 294, + 232 + ], + "type": "table", + "html": "
MethodAcc
LLaVA(w/o CC during finetune)36.0
LLaVA+CC(train only)37.17
LLaVA+CC(train & test)39.13
", + "image_path": "9d3c9b7f6685d2d910b423f11c7163af91b63308a6e16858dfc40230bab5ad3f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 232, + 297, + 300 + ], + "lines": [ + { + "bbox": [ + 55, + 232, + 297, + 300 + ], + "spans": [ + { + "bbox": [ + 55, + 232, + 297, + 300 + ], + "type": "text", + "content": "Table 5. Comparisons on SQA3D dataset. We tested the effectiveness of COARSE CORRESPONDENCES on the out-of-domain SQA3D dataset. With the same training data, we found that using COARSE CORRESPONDENCES not only improves performance during inference but also enhances zero-shot evaluation when applied solely during the training phase." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 313, + 295, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 313, + 295, + 337 + ], + "spans": [ + { + "bbox": [ + 55, + 313, + 295, + 337 + ], + "type": "text", + "content": "on proving that COARSE CORRESPONDENCES improves open-source models' understanding of 3D space." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 338, + 297, + 446 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 338, + 297, + 446 + ], + "spans": [ + { + "bbox": [ + 54, + 338, + 297, + 446 + ], + "type": "text", + "content": "We start with the LLaVA model [22] and fine-tune it using a dataset comprising approximately 1.2 million samples of image and video data. Notably, the ScanQA dataset is the only 3D-related dataset in this collection. COARSE CORRESPONDENCES is applied exclusively to the ScanQA data, while other data maintain their original format. Fine-tuning on such a dataset enables the open-source LLaVA model to better interpret overlayed marks on images and process multiple image inputs." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 448, + 297, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 448, + 297, + 590 + ], + "spans": [ + { + "bbox": [ + 55, + 448, + 297, + 590 + ], + "type": "text", + "content": "In-domain Evaluation. We first evaluate whether COARSE CORRESPONDENCES can improve open-source models on the ScanQA validation set. As shown in Table 1, our method significantly enhances the model's 3D spatial understanding compared to the baseline fine-tuned on the same data without using COARSE CORRESPONDENCES, even outperforming previous VLMs [13] specifically designed for 3D tasks, which incorporate specialized architectural designs and are fine-tuned on substantially larger amounts of 3D-related data. This in-domain evaluation demonstrates that COARSE CORRESPONDENCES is also effective for open MLLMs." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 592, + 296, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 592, + 296, + 700 + ], + "spans": [ + { + "bbox": [ + 55, + 592, + 296, + 700 + ], + "type": "text", + "content": "Out-domain Evaluation. To further demonstrate the generalizability of our method, we conduct experiments on the SQA3D dataset to evaluate whether COARSE CORRESPONDENCES can also provide improvements in a zero-shot setting. As shown in Table 5, on this previously unseen dataset, COARSE CORRESPONDENCES still outperforms the baseline simply fine-tuned on the same data, proving that our method can enhance open-source models' understanding of 3D space on out-of-domain datasets." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 67, + 701, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 701, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 701, + 296, + 713 + ], + "type": "text", + "content": "Notably, even without using CoARSE CORRESPON-" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 173, + 555, + 244 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 173, + 555, + 244 + ], + "spans": [ + { + "bbox": [ + 313, + 173, + 555, + 244 + ], + "type": "text", + "content": "DENCES during inference, simply applying it during training already yields improvements. This highlights that our method is effective not only as a prompting technique for inference but also holds potential as a data augmentation method during training, a prospect worth exploring further in future work." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 258, + 373, + 272 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 258, + 373, + 272 + ], + "spans": [ + { + "bbox": [ + 313, + 258, + 373, + 272 + ], + "type": "text", + "content": "5. Analysis" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 279, + 556, + 375 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 279, + 556, + 375 + ], + "spans": [ + { + "bbox": [ + 313, + 279, + 556, + 375 + ], + "type": "text", + "content": "Here, we explore the various design decisions in our method. Although the study is conducted only on ScanQA, applying the design choices derived from ScanQA to all other benchmarks also yields significant performance improvements. This strongly demonstrates the generalizability of these design choices. Additionally, we further analyze the impact of our method on mitigating the camera motion bias present in MLLMs." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 376, + 556, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 376, + 556, + 520 + ], + "spans": [ + { + "bbox": [ + 313, + 376, + 556, + 520 + ], + "type": "text", + "content": "How does COARSE CORRESPONDENCES differ from other visual prompting methods? Our proposed method calculates and highlights correspondences between images, aiming to elicit 3D and temporal understanding. Other visual prompting methods (namely Set-of-Mark [48], 3DAxiesPrompts [21], and Chain-of-thought [43]) can also be viewed as alternative prompting methods. Given that the ground-truth answers in existing benchmarks are relatively brief, we selected a scene from ScanQA and manually designed a new question. We qualitatively compare COARSE CORRESPONDENCES against other prompting methods on this new question, as shown in Figure 2." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 521, + 556, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 556, + 629 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 556, + 629 + ], + "type": "text", + "content": "The orange part of Figure 2 shows our Coarse Correspondence labels are recognized by GPT-4V. The output answer provides evidence that our coarse correspondence helps GPT-4V develop a mental 3D model of the scene. Set-of-Marks provides no spatial corresponding information and therefore is unhelpful. The Axis labels in 3DAxies can be easily misrecognized by GPT-4V, leading to misleading spatial information. Though Chain-of-Thought helps identify objects, it fails to resolve the \"spatial perspective-taking\" issue." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 630, + 557, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 557, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 557, + 714 + ], + "type": "text", + "content": "Why use coarse instead of dense correspondences? Instead of filtering and retaining only a handful of coarse correspondences, one ablation we considered is the possibility of using all dense correspondence. Unfortunately, we find that excessively overlaying too many instance marks can degrade performance (Figure 3) as they occlude the visual content in the images." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3788" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 73, + 77, + 534, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 77, + 534, + 95 + ], + "spans": [ + { + "bbox": [ + 73, + 77, + 534, + 95 + ], + "type": "text", + "content": "System: You are an AI with the ability to analyze a series of images, each representing a different perspective of a single scene. [Prompt-about-Marks]. Your task is to construct a 3D understanding based on these images." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 73, + 99, + 536, + 118 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 99, + 536, + 118 + ], + "spans": [ + { + "bbox": [ + 73, + 99, + 536, + 118 + ], + "type": "text", + "content": "User: You are sitting on the sofa and the electric fan is on your left. Describe the location of the room door from your perspective: A. to the front left of you; B. to the front right of you; C. to the back left of you; D. to the back right of you." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 79, + 125, + 216, + 137 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 125, + 216, + 137 + ], + "spans": [ + { + "bbox": [ + 79, + 125, + 216, + 137 + ], + "type": "text", + "content": "+Coarse Correspondence (Ours)" + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 76, + 140, + 149, + 195 + ], + "blocks": [ + { + "bbox": [ + 76, + 140, + 149, + 195 + ], + "lines": [ + { + "bbox": [ + 76, + 140, + 149, + 195 + ], + "spans": [ + { + "bbox": [ + 76, + 140, + 149, + 195 + ], + "type": "image", + "image_path": "0fe6c28e518bdc27bd7fb5f4d1280aa2ffeb6fbf693cd70bd02e580b692921df.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 101, + 197, + 124, + 204 + ], + "lines": [ + { + "bbox": [ + 101, + 197, + 124, + 204 + ], + "spans": [ + { + "bbox": [ + 101, + 197, + 124, + 204 + ], + "type": "text", + "content": "View 1" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 155, + 140, + 228, + 195 + ], + "blocks": [ + { + "bbox": [ + 155, + 140, + 228, + 195 + ], + "lines": [ + { + "bbox": [ + 155, + 140, + 228, + 195 + ], + "spans": [ + { + "bbox": [ + 155, + 140, + 228, + 195 + ], + "type": "image", + "image_path": "2ff9112e172f2835a79129ddc506b432ca7aebcb262c1bd791444c887217654b.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 180, + 197, + 203, + 204 + ], + "lines": [ + { + "bbox": [ + 180, + 197, + 203, + 204 + ], + "spans": [ + { + "bbox": [ + 180, + 197, + 203, + 204 + ], + "type": "text", + "content": "View 2" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 75, + 210, + 226, + 256 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 210, + 226, + 256 + ], + "spans": [ + { + "bbox": [ + 75, + 210, + 226, + 256 + ], + "type": "text", + "content": "GPT-4V: The first image ... a door marked with number 7, a fan marked with number 39, and items on shelves marked with numbers 11 and 37. In the second image, ... the fan again at number 39, and a sofa marked with number 3(√)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 75, + 255, + 227, + 343 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 255, + 227, + 343 + ], + "spans": [ + { + "bbox": [ + 75, + 255, + 227, + 343 + ], + "type": "text", + "content": "We can infer that the door and the fan are on the same side of the room(✓). Since the fan appears to the left in the first image and is also on the left when sitting on the sofa as per the second image and your description, the door must be to the front of the sofa, as it is not visible in the second image from sofa's perspective(✓). By 'sitting on the sofa' as instructed, and with the electric fan to the left, I can determine the door is A. to the front left of you (✓)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 242, + 125, + 300, + 137 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 242, + 125, + 300, + 137 + ], + "spans": [ + { + "bbox": [ + 242, + 125, + 300, + 137 + ], + "type": "text", + "content": "+Set-of-Mark" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 236, + 140, + 310, + 194 + ], + "blocks": [ + { + "bbox": [ + 236, + 140, + 310, + 194 + ], + "lines": [ + { + "bbox": [ + 236, + 140, + 310, + 194 + ], + "spans": [ + { + "bbox": [ + 236, + 140, + 310, + 194 + ], + "type": "image", + "image_path": "8c7e9682a85506eb5ffaf9bd9bde061a44e7719f499a7dc6f7caf310a7ef0c17.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 261, + 196, + 283, + 203 + ], + "lines": [ + { + "bbox": [ + 261, + 196, + 283, + 203 + ], + "spans": [ + { + "bbox": [ + 261, + 196, + 283, + 203 + ], + "type": "text", + "content": "View 1" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 312, + 139, + 385, + 194 + ], + "blocks": [ + { + "bbox": [ + 312, + 139, + 385, + 194 + ], + "lines": [ + { + "bbox": [ + 312, + 139, + 385, + 194 + ], + "spans": [ + { + "bbox": [ + 312, + 139, + 385, + 194 + ], + "type": "image", + "image_path": "bf90788f5d01e2b03543d31adaae137c152a0b3b8f0f2efb3a29f6a5eb2267b2.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 340, + 196, + 362, + 203 + ], + "lines": [ + { + "bbox": [ + 340, + 196, + 362, + 203 + ], + "spans": [ + { + "bbox": [ + 340, + 196, + 362, + 203 + ], + "type": "text", + "content": "View 2" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 388, + 140, + 528, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 388, + 140, + 528, + 190 + ], + "spans": [ + { + "bbox": [ + 388, + 140, + 528, + 190 + ], + "type": "text", + "content": "GPT-4V: The fan is seen on the right side from the sofa's perspective, its leftward when sitting means the orientation is flipped(√). Therefore, the door that is visible on the right in the first image would be D. behind and to the right when seated on the sofa(X)." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 242, + 214, + 318, + 224 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 242, + 214, + 318, + 224 + ], + "spans": [ + { + "bbox": [ + 242, + 214, + 318, + 224 + ], + "type": "text", + "content": "+3DAxesPrompts" + } + ] + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 236, + 225, + 310, + 281 + ], + "blocks": [ + { + "bbox": [ + 236, + 225, + 310, + 281 + ], + "lines": [ + { + "bbox": [ + 236, + 225, + 310, + 281 + ], + "spans": [ + { + "bbox": [ + 236, + 225, + 310, + 281 + ], + "type": "image", + "image_path": "79719678b177040e7303dce4bcbabb3c35fd5cb7b4f41a8efd9c81bd658259a4.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 261, + 281, + 283, + 289 + ], + "lines": [ + { + "bbox": [ + 261, + 281, + 283, + 289 + ], + "spans": [ + { + "bbox": [ + 261, + 281, + 283, + 289 + ], + "type": "text", + "content": "View 1" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 312, + 225, + 385, + 281 + ], + "blocks": [ + { + "bbox": [ + 312, + 225, + 385, + 281 + ], + "lines": [ + { + "bbox": [ + 312, + 225, + 385, + 281 + ], + "spans": [ + { + "bbox": [ + 312, + 225, + 385, + 281 + ], + "type": "image", + "image_path": "8b39bec3be4d458965620e2b801246b81dadc9a99a50c8f4e86746f53bbbc7a0.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 340, + 281, + 362, + 289 + ], + "lines": [ + { + "bbox": [ + 340, + 281, + 362, + 289 + ], + "spans": [ + { + "bbox": [ + 340, + 281, + 362, + 289 + ], + "type": "text", + "content": "View 2" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 379, + 555, + 412 + ], + "lines": [ + { + "bbox": [ + 55, + 379, + 555, + 412 + ], + "spans": [ + { + "bbox": [ + 55, + 379, + 555, + 412 + ], + "type": "text", + "content": "Figure 2. Comparison of different prompting method. Our proposed Coarse Correspondences successfully guided GPT-40 to understand 3D spatial relationships and generate the right answer. Other existing prompting method including image-based Set-of-Marks, 3DAxies and text-based Chain-of-Thought failed to answer correctly." + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "bbox": [ + 388, + 224, + 537, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 388, + 224, + 537, + 287 + ], + "spans": [ + { + "bbox": [ + 388, + 224, + 537, + 287 + ], + "type": "text", + "content": "GPT-4V: The first image shows ... and a coordinate system with X-axis (horizontally rightwards) " + }, + { + "bbox": [ + 388, + 224, + 537, + 287 + ], + "type": "inline_equation", + "content": "(\\mathbf{X})" + }, + { + "bbox": [ + 388, + 224, + 537, + 287 + ], + "type": "text", + "content": " and Z-axis (vertically upwards). The second image shows a sofa on the left with a coordinate system's Y-axis (horizontally extending away from the viewer) " + }, + { + "bbox": [ + 388, + 224, + 537, + 287 + ], + "type": "inline_equation", + "content": "(\\mathbf{X})" + }, + { + "bbox": [ + 388, + 224, + 537, + 287 + ], + "type": "text", + "content": " and Z-axis as before. When sitting on the sofa with the fan to your left, the door would be B. diagonally opposite to you on the right side " + }, + { + "bbox": [ + 388, + 224, + 537, + 287 + ], + "type": "inline_equation", + "content": "(\\mathbf{X})" + }, + { + "bbox": [ + 388, + 224, + 537, + 287 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 242, + 297, + 324, + 307 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 242, + 297, + 324, + 307 + ], + "spans": [ + { + "bbox": [ + 242, + 297, + 324, + 307 + ], + "type": "text", + "content": "+Chain-of-Thought" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 239, + 309, + 385, + 336 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 239, + 309, + 385, + 336 + ], + "spans": [ + { + "bbox": [ + 239, + 309, + 385, + 336 + ], + "type": "text", + "content": "System: Your response should be as follows. \n1)Identification: Identify all labeled objects. \n2)Analysis: Analyze the objects and their positions construct a 3D model." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 239, + 336, + 385, + 349 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 239, + 336, + 385, + 349 + ], + "spans": [ + { + "bbox": [ + 239, + 336, + 385, + 349 + ], + "type": "text", + "content": "3)Synthesis: Combine the info above, form a comprehensive understanding." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 239, + 349, + 385, + 362 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 239, + 349, + 385, + 362 + ], + "spans": [ + { + "bbox": [ + 239, + 349, + 385, + 362 + ], + "type": "text", + "content": "4)Response: Provide a concise answer to the question based on your comprehensive 3D modelling." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 391, + 300, + 419, + 308 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 391, + 300, + 419, + 308 + ], + "spans": [ + { + "bbox": [ + 391, + 300, + 419, + 308 + ], + "type": "text", + "content": "GPT-4V:" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 390, + 308, + 535, + 336 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 390, + 308, + 535, + 336 + ], + "spans": [ + { + "bbox": [ + 390, + 308, + 535, + 336 + ], + "type": "text", + "content": "Identification: Door, shelves, sofa,... " + }, + { + "bbox": [ + 390, + 308, + 535, + 336 + ], + "type": "inline_equation", + "content": "(\\checkmark)" + }, + { + "bbox": [ + 390, + 308, + 535, + 336 + ], + "type": "text", + "content": " Analysis: in image 1 the door is on the right with the shelves adjacent to the door ... In image 2 the sofa is ... " + }, + { + "bbox": [ + 390, + 308, + 535, + 336 + ], + "type": "inline_equation", + "content": "(\\checkmark)" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 389, + 336, + 535, + 355 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 389, + 336, + 535, + 355 + ], + "spans": [ + { + "bbox": [ + 389, + 336, + 535, + 355 + ], + "type": "text", + "content": "Synthesis: When sitting on the sofa, the shelves are on the right(X). The door is next to the shelves." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 390, + 355, + 444, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 390, + 355, + 444, + 361 + ], + "spans": [ + { + "bbox": [ + 390, + 355, + 444, + 361 + ], + "type": "text", + "content": "Response: B. (X)." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 55, + 416, + 297, + 488 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 416, + 297, + 488 + ], + "spans": [ + { + "bbox": [ + 55, + 416, + 297, + 488 + ], + "type": "text", + "content": "How large should the marks be? We inject the correspondences into MLLMs by overlaying the marks into images. We empirically find an optimal mark size (where 'px' represents the mark's diameter in pixels) in Figure 3. Marks that are too small tend to be ignored while those that are too large occlude visual content." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 55, + 493, + 297, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 493, + 297, + 577 + ], + "spans": [ + { + "bbox": [ + 55, + 493, + 297, + 577 + ], + "type": "text", + "content": "What shape should the marks be? We further studied the appearance of the marks. In addition to red circles with white text, we experimented with adding segmentation outlines and segmentation masks. As shown in Figure 3, using segmentation outlines enhances object grounding. However, using segmentation marks occludes visual content and reduces performance." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 55, + 582, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 297, + 714 + ], + "type": "text", + "content": "Improvement of Camera Motion Invariance with CoARSE CORRESPONDENCES. A crucial aspect of strong spatial reasoning is the model's ability to maintain an understanding of spatial relationships invariant to camera motion; in other words, whether the image sequence is captured from left-to-right or right-to-left, the model's comprehension of static space should remain consistent. However, current benchmarks lack annotations regarding whether the 3D space scan was conducted from left-to-right or right-to-left, making it challenging to analyze the effect of camera motion on MLLMs' spatial understanding. To address this," + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 313, + 416, + 556, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 416, + 556, + 547 + ], + "spans": [ + { + "bbox": [ + 313, + 416, + 556, + 547 + ], + "type": "text", + "content": "we curated a small diagnostic benchmark specifically for this analysis. This benchmark includes 10 scenes captured from left-to-right, with five questions per scene focusing on left-right spatial relationships between objects—fundamental aspects of spatial understanding. We ensured that all questions required information across multiple frames and could not be answered from a single frame alone. Ideally, regardless of whether the image sequence is input in the original or reversed order, the output should remain consistent. We leave additional details on data curation in the supplementary material." + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 313, + 553, + 556, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 553, + 556, + 649 + ], + "spans": [ + { + "bbox": [ + 313, + 553, + 556, + 649 + ], + "type": "text", + "content": "To mitigate randomness, each question was tested 20 times, yielding a total of 1000 trials, and we report the average accuracy in Table 6. We then reversed the sequence order to simulate a right-to-left capture and repeated 1000 trials with this reversed input, reporting the average accuracy for this reverse sequence. Finally, we report the harmonic mean of the average accuracies for the forward and reverse inputs." + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 313, + 654, + 557, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 557, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 557, + 715 + ], + "type": "text", + "content": "Without CoARSE CORRESPONDENCES, the accuracy in understanding spatial relationships in right-to-left sequences was significantly lower than in left-to-right sequences. With CoARSE CORRESPONDENCES, not only did MLLMs achieve a substantial increase in spatial un" + } + ] + } + ], + "index": 35 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3789" + } + ] + } + ], + "index": 36 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 73, + 231, + 232 + ], + "blocks": [ + { + "bbox": [ + 60, + 73, + 231, + 232 + ], + "lines": [ + { + "bbox": [ + 60, + 73, + 231, + 232 + ], + "spans": [ + { + "bbox": [ + 60, + 73, + 231, + 232 + ], + "type": "image", + "image_path": "bca43439fed192cbc8fcb8351b17493dc4145d1ea329596027de5730b2e276f0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 237, + 555, + 261 + ], + "lines": [ + { + "bbox": [ + 55, + 237, + 555, + 261 + ], + "spans": [ + { + "bbox": [ + 55, + 237, + 555, + 261 + ], + "type": "text", + "content": "Figure 3. Ablations on different design choices of COARSE CORRESPONDENCES. We studied the impact of the number, size, and type of marks on performance. All experiments were conducted on ScanQA using GPT-4V." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 231, + 74, + 390, + 232 + ], + "blocks": [ + { + "bbox": [ + 231, + 74, + 390, + 232 + ], + "lines": [ + { + "bbox": [ + 231, + 74, + 390, + 232 + ], + "spans": [ + { + "bbox": [ + 231, + 74, + 390, + 232 + ], + "type": "image", + "image_path": "d9e791dfd0896750629d08fd18032293b53343eaf54d936d0f28d9141647c815.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 391, + 76, + 550, + 232 + ], + "blocks": [ + { + "bbox": [ + 391, + 76, + 550, + 232 + ], + "lines": [ + { + "bbox": [ + 391, + 76, + 550, + 232 + ], + "spans": [ + { + "bbox": [ + 391, + 76, + 550, + 232 + ], + "type": "image", + "image_path": "42eac979b8b802b8830f4100c9584b88c550e7117c6b8c35f934381261188779.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 58, + 269, + 293, + 312 + ], + "blocks": [ + { + "bbox": [ + 58, + 269, + 293, + 312 + ], + "lines": [ + { + "bbox": [ + 58, + 269, + 293, + 312 + ], + "spans": [ + { + "bbox": [ + 58, + 269, + 293, + 312 + ], + "type": "table", + "html": "
ModelsFrameForwardReverseHarmonic Mean
GPT-4O458.050.453.9
GPT-4O+CC471.271.271.2
", + "image_path": "07aab8d4a4ca3b2ec6e4d6fc397a998df064f7f1898448aa0a393bef5e145c99.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 316, + 296, + 361 + ], + "lines": [ + { + "bbox": [ + 55, + 316, + 296, + 361 + ], + "spans": [ + { + "bbox": [ + 55, + 316, + 296, + 361 + ], + "type": "text", + "content": "Table 6. Effect of COARSE CORRESPONDENCES on Camera Motion Bias. COARSE CORRESPONDENCES shows strong capability of enhancing 3D spatial understanding of MLLMs. It can also ease the camera motion bias of current MLLMs." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 365, + 296, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 365, + 296, + 473 + ], + "spans": [ + { + "bbox": [ + 54, + 365, + 296, + 473 + ], + "type": "text", + "content": "derstanding accuracy with the original sequence " + }, + { + "bbox": [ + 54, + 365, + 296, + 473 + ], + "type": "inline_equation", + "content": "(+13.2\\%)" + }, + { + "bbox": [ + 54, + 365, + 296, + 473 + ], + "type": "text", + "content": ", but more importantly, reversing the input order no longer drastically impacted the model's comprehension of static spatial relationships. Specifically, COARSE CORRESPON-DENCES improved the harmonic mean of forward and reverse sequence accuracy by " + }, + { + "bbox": [ + 54, + 365, + 296, + 473 + ], + "type": "inline_equation", + "content": "17.3\\%" + }, + { + "bbox": [ + 54, + 365, + 296, + 473 + ], + "type": "text", + "content": ". This demonstrates that our method enables MLLMs to perform spatial reasoning invariant to camera motion, thus achieving more robust spatial understanding capabilities." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 488, + 140, + 501 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 488, + 140, + 501 + ], + "spans": [ + { + "bbox": [ + 55, + 488, + 140, + 501 + ], + "type": "text", + "content": "6. Related work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 510, + 297, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 297, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 297, + 715 + ], + "type": "text", + "content": "Multimodal language models Multimodal LLMs [3, 23] integrate vision encoders [32] into large LLMs [8, 40], enabling direct reasoning over visual input. Proprietary models like GPT-4 [27], Gemini [38], and Claude [1], along with open-source models such as the LLaVA series [23] and BLIP series [18], have made significant strides in 2D vision-language tasks like image captioning [7] and visual question answering (VQA) [11, 14]. However, for applications in real-world scenarios like autonomous driving [39] and robotics [46], MLLMs still require stronger spatial-temporal reasoning capabilities. Recent work [19] has enhanced MLLMs' capacity to process multiple image frames and understand temporal dynamics, achieving notable improvements in video dense captioning [17] and videoQA [12, 44]. Nonetheless, MLLMs still face challenges in understanding complex 3D spatial relations [24] and long-term temporal dynamics [25] present in image sequences. Many" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 271, + 555, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 271, + 555, + 354 + ], + "spans": [ + { + "bbox": [ + 313, + 271, + 555, + 354 + ], + "type": "text", + "content": "approaches have introduced specialized architectures [13] or fine-tuning methods targeting spatial and temporal reasoning individually. Our goal, however, is to jointly enhance spatial-temporal reasoning within the existing general-purpose MLLM framework. By leveraging off-the-shelf tracking models to extract instance-level correspondences, we achieve this in a training-free manner." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 356, + 556, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 356, + 556, + 499 + ], + "spans": [ + { + "bbox": [ + 313, + 356, + 556, + 499 + ], + "type": "text", + "content": "Visual prompting. Effective prompting has been widely proven to improve LLMs across multiple domains. Methods, such as chain-of-thought prompting [43], force the model to reason before answering a question. In the context of visual prompting, we can enhance MLLMs' grounding abilities [36, 48], their ability to use numerical expressions to describe 3D object relationships [21], and their capability to generate outputs for robotics control [26], all using single-image inputs. Unlike these approaches, COARSE CORRESPONDENCES is designed to enhance MLLMs' spatial-temporal reasoning capabilities for interpreting complex spacetime structures represented in image sequences." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 513, + 389, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 513, + 389, + 525 + ], + "spans": [ + { + "bbox": [ + 313, + 513, + 389, + 525 + ], + "type": "text", + "content": "7. Conclusion." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 534, + 556, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 556, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 556, + 712 + ], + "type": "text", + "content": "We propose a framework called COARSE CORRESPON-DENCES prompting. By using off-the-shelf video tracking models to obtain class-agnostic, instance-level correspondences and conveying this information to MLLMs through visual prompting, we discovered that this simple method, using only 2D images as input—without any specialized architectural design or task-specific SFT—can effectively enhance MLLMs' spatial-temporal reasoning. This improvement extends to embodied tasks like navigation. Our method not only works on proprietary models but also generalizes to open-source models, and it performs well on both in-domain and out-of-domain datasets. Moreover, it enhances not just inference but also training. Further analysis shows that our method helps MLLMs become more robust to camera motion bias." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3790" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 90, + 297, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 90, + 297, + 190 + ], + "spans": [ + { + "bbox": [ + 55, + 90, + 297, + 190 + ], + "type": "text", + "content": "We thank Xiaojuan Wang, Ruitao Zhang, Yuntian Deng for helpful discussions, feedback and collecting data. This project is partially funded by Amazon Science. This paper was partially supported by a gift from Ai2 and DARPA Transfer from Imprecise and Abstract Models to Autonomous Technologies (TIAMAT) program No. HR00112490422. Its contents are solely the responsibility of the authors and do not necessarily represent the official views of DARPA." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 201, + 115, + 213 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 201, + 115, + 213 + ], + "spans": [ + { + "bbox": [ + 56, + 201, + 115, + 213 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 220, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 220, + 296, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 220, + 296, + 243 + ], + "spans": [ + { + "bbox": [ + 61, + 220, + 296, + 243 + ], + "type": "text", + "content": "[1] AI Anthropic. The claude 3 model family: Opus, sonnet, haiku. Claude-3 Model Card, 2024. 4, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 243, + 296, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 243, + 296, + 297 + ], + "spans": [ + { + "bbox": [ + 61, + 243, + 296, + 297 + ], + "type": "text", + "content": "[2] Daichi Azuma, Taiki Miyanishi, Shuhei Kurita, and Motoaki Kawanabe. Scanqa: 3d question answering for spatial scene understanding. In proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 19129-19139, 2022. 2, 3, 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 300, + 296, + 343 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 300, + 296, + 343 + ], + "spans": [ + { + "bbox": [ + 62, + 300, + 296, + 343 + ], + "type": "text", + "content": "[3] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond, 2023. 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 345, + 296, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 345, + 296, + 388 + ], + "spans": [ + { + "bbox": [ + 62, + 345, + 296, + 388 + ], + "type": "text", + "content": "[4] Ivana Balažević, Yuge Shi, Pinelopi Papalampidi, Rahma Chaabouni, Skanda Koppula, and Olivier J. Henaff. Memory consolidation enables long-context video understanding, 2024. 1, 4, 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 389, + 296, + 444 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 389, + 296, + 444 + ], + "spans": [ + { + "bbox": [ + 62, + 389, + 296, + 444 + ], + "type": "text", + "content": "[5] Satanjeev Banerjee and Alon Lavie. Meteor: An automatic metric for mt evaluation with improved correlation with human judgments. In Proceedings of the acl workshop on intrinsic and extrinsic evaluation measures for machine translation and/or summarization, pages 65-72, 2005. 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 445, + 296, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 445, + 296, + 499 + ], + "spans": [ + { + "bbox": [ + 62, + 445, + 296, + 499 + ], + "type": "text", + "content": "[6] Boyuan Chen, Zhuo Xu, Sean Kirmani, Brian Ichter, Danny Driess, Pete Florence, Dorsa Sadigh, Leonidas Guibas, and Fei Xia. Spatialvlm: Endowing vision-language models with spatial reasoning capabilities. arXiv preprint arXiv:2401.12168, 2024. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 501, + 296, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 501, + 296, + 544 + ], + "spans": [ + { + "bbox": [ + 62, + 501, + 296, + 544 + ], + "type": "text", + "content": "[7] Xinlei Chen, Hao Fang, Tsung-Yi Lin, Ramakrishna Vedantam, Saurabh Gupta, Piotr Dollar, and C. Lawrence Zitnick. Microsoft coco captions: Data collection and evaluation server, 2015. 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 545, + 296, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 545, + 296, + 601 + ], + "spans": [ + { + "bbox": [ + 62, + 545, + 296, + 601 + ], + "type": "text", + "content": "[8] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhanghao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E. Gonzalez, Ion Stoica, and Eric P. Xing. Vicuna: An open-source chatbot impressing gpt-4 with " + }, + { + "bbox": [ + 62, + 545, + 296, + 601 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 62, + 545, + 296, + 601 + ], + "type": "text", + "content": " chatgpt quality, 2023. 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 601, + 296, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 601, + 296, + 645 + ], + "spans": [ + { + "bbox": [ + 62, + 601, + 296, + 645 + ], + "type": "text", + "content": "[9] Jiafei Duan, Samson Yu, Hui Li Tan, Hongyuan Zhu, and Cheston Tan. A survey of embodied ai: From simulators to research tasks. IEEE Transactions on Emerging Topics in Computational Intelligence, 6(2):230-244, 2022. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 647, + 296, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 296, + 689 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 296, + 689 + ], + "type": "text", + "content": "[10] Yue Fan, Xiaojian Ma, Rujie Wu, Yuntao Du, Jiaqi Li, Zhi Gao, and Qing Li. Videoagent: A memory-augmented multimodal agent for video understanding. arXiv preprint arXiv:2403.11481, 2024.5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 691, + 296, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 296, + 714 + ], + "type": "text", + "content": "[11] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Batra, and Devi Parikh. Making the v in vqa matter: Elevating" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 333, + 73, + 555, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 555, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 555, + 95 + ], + "type": "text", + "content": "the role of image understanding in visual question answering, 2017. 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 97, + 555, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 97, + 555, + 152 + ], + "spans": [ + { + "bbox": [ + 316, + 97, + 555, + 152 + ], + "type": "text", + "content": "[12] Madeleine Grunde-McLaughlin, Ranjay Krishna, and Maneesh Agrawala. Agqa: A benchmark for compositional spatio-temporal reasoning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11287-11297, 2021. 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 153, + 555, + 198 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 153, + 555, + 198 + ], + "spans": [ + { + "bbox": [ + 316, + 153, + 555, + 198 + ], + "type": "text", + "content": "[13] Yining Hong, Haoyu Zhen, Peihao Chen, Shuhong Zheng, Yilun Du, Zhenfang Chen, and Chuang Gan. 3d-llm: Injecting the 3d world into large language models. Advances in Neural Information Processing Systems, 36, 2024. 1, 3, 4, 6, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 199, + 554, + 232 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 199, + 554, + 232 + ], + "spans": [ + { + "bbox": [ + 316, + 199, + 554, + 232 + ], + "type": "text", + "content": "[14] Drew A. Hudson and Christopher D. Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering, 2019. 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 235, + 555, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 235, + 555, + 267 + ], + "spans": [ + { + "bbox": [ + 316, + 235, + 555, + 267 + ], + "type": "text", + "content": "[15] Kumara Kahatapitiya, Kanchana Ranasinghe, Jongwoo Park, and Michael S. Ryoo. Language repository for long video understanding, 2024. 1, 4, 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 270, + 555, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 270, + 555, + 335 + ], + "spans": [ + { + "bbox": [ + 316, + 270, + 555, + 335 + ], + "type": "text", + "content": "[16] Jacob Krantz, Erik Wijmans, Arjun Majumdar, Dhruv Batra, and Stefan Lee. Beyond the nav-graph: Vision-and-language navigation in continuous environments. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXVIII 16, pages 104-120. Springer, 2020. 2, 3, 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 337, + 555, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 337, + 555, + 369 + ], + "spans": [ + { + "bbox": [ + 316, + 337, + 555, + 369 + ], + "type": "text", + "content": "[17] Ranjay Krishna, Kenji Hata, Frederic Ren, Li Fei-Fei, and Juan Carlos Niebles. Dense-captioning events in videos, 2017. 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 372, + 555, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 372, + 555, + 415 + ], + "spans": [ + { + "bbox": [ + 316, + 372, + 555, + 415 + ], + "type": "text", + "content": "[18] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip2: Bootstrapping language-image pre-training with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597, 2023. 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 418, + 555, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 418, + 555, + 460 + ], + "spans": [ + { + "bbox": [ + 316, + 418, + 555, + 460 + ], + "type": "text", + "content": "[19] Bin Lin, Bin Zhu, Yang Ye, Munan Ning, Peng Jin, and Li Yuan. Video-llava: Learning united visual representation by alignment before projection. arXiv preprint arXiv:2311.10122, 2023. 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 464, + 554, + 496 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 464, + 554, + 496 + ], + "spans": [ + { + "bbox": [ + 316, + 464, + 554, + 496 + ], + "type": "text", + "content": "[20] Chin-Yew Lin. Rouge: A package for automatic evaluation of summaries. In Text summarization branches out, pages 74-81, 2004. 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 498, + 555, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 498, + 555, + 543 + ], + "spans": [ + { + "bbox": [ + 316, + 498, + 555, + 543 + ], + "type": "text", + "content": "[21] Dingning Liu, Xiaomeng Dong, Renrui Zhang, Xu Luo, Peng Gao, Xiaoshui Huang, Yongshun Gong, and Zhihui Wang. 3daxiesprompts: Unleashing the 3d spatial task capabilities of gpt-4v. arXiv preprint arXiv:2312.09738, 2023. 6, 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 315, + 544, + 555, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 544, + 555, + 567 + ], + "spans": [ + { + "bbox": [ + 315, + 544, + 555, + 567 + ], + "type": "text", + "content": "[22] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. NeurIPS, 36, 2024. 2, 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 569, + 555, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 569, + 555, + 602 + ], + "spans": [ + { + "bbox": [ + 316, + 569, + 555, + 602 + ], + "type": "text", + "content": "[23] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 8" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 604, + 555, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 604, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 604, + 555, + 713 + ], + "type": "text", + "content": "[24] Arjun Majumdar, Anurag Ajay, Xiaohan Zhang, Pranav Putta, Sriram Yenamandra, Mikael Henaff, Sneha Silwal, Paul Mcvay, Oleksandr Maksymets, Sergio Arnaud, Karmesh Yadav, Qiyang Li, Ben Newman, Mohit Sharma, Vincent Berges, Shiqi Zhang, Pulkit Agrawal, Yonatan Bisk, Dhruv Batra, Mrinal Kalakrishnan, Franziska Meier, Chris Paxton, Sasha Sax, and Aravind Rajeswaran. Openeqa: Embodied question answering in the era of foundation models. In Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 1, 2, 3, 8" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "3791" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 56, + 73, + 296, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 296, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 296, + 106 + ], + "type": "text", + "content": "[25] Karttikeya Mangalam, Raiymbek Akshulakov, and Jitendra Malik. Egoschema: A diagnostic benchmark for very long-form video language understanding, 2023. 2, 3, 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 106, + 296, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 106, + 296, + 160 + ], + "spans": [ + { + "bbox": [ + 56, + 106, + 296, + 160 + ], + "type": "text", + "content": "[26] Soroush Nasiriany, Fei Xia, Wenhao Yu, Ted Xiao, Jacky Liang, Ishita Dasgupta, Annie Xie, Danny Driess, Ayzaan Wahid, Zhuo Xu, et al. Pivot: Iterative visual prompting elicits actionable knowledge for vlms. arXiv preprint arXiv:2402.07872, 2024. 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 161, + 294, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 161, + 294, + 173 + ], + "spans": [ + { + "bbox": [ + 56, + 161, + 294, + 173 + ], + "type": "text", + "content": "[27] OpenAI. Gpt-4v(ision) system card. OpenAI Blog, 2023. 4, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 173, + 294, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 173, + 294, + 194 + ], + "spans": [ + { + "bbox": [ + 56, + 173, + 294, + 194 + ], + "type": "text", + "content": "[28] OpenAI. Hello gpt-4o. https://openai.com/index/hello-gpt-4o/, 2024. Accessed: 2024-05-22.1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 194, + 294, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 194, + 294, + 216 + ], + "spans": [ + { + "bbox": [ + 56, + 194, + 294, + 216 + ], + "type": "text", + "content": "[29] OpenAI, Josh Achiam, and et al. Gpt-4 technical report, 2024. 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 217, + 294, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 217, + 294, + 271 + ], + "spans": [ + { + "bbox": [ + 56, + 217, + 294, + 271 + ], + "type": "text", + "content": "[30] Pinelopi Papalampidi, Skanda Koppula, Shreya Pathak, Justin Chiu, Joe Heyward, Viorica Patraucean, Jiajun Shen, Antoine Miech, Andrew Zisserman, and Aida Nematzdeh. A simple recipe for contrastively pre-training video-first encoders beyond 16 frames, 2023. 1, 4, 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 271, + 294, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 271, + 294, + 326 + ], + "spans": [ + { + "bbox": [ + 56, + 271, + 294, + 326 + ], + "type": "text", + "content": "[31] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th annual meeting of the Association for Computational Linguistics, pages 311-318, 2002. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 327, + 294, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 327, + 294, + 392 + ], + "spans": [ + { + "bbox": [ + 56, + 327, + 294, + 392 + ], + "type": "text", + "content": "[32] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 393, + 294, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 393, + 294, + 426 + ], + "spans": [ + { + "bbox": [ + 56, + 393, + 294, + 426 + ], + "type": "text", + "content": "[33] Kanchana Ranasinghe, Xiang Li, Kumara Kahatapitiya, and Michael S. Ryoo. Understanding long videos in one multimodal language model pass, 2024. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 426, + 294, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 426, + 294, + 480 + ], + "spans": [ + { + "bbox": [ + 56, + 426, + 294, + 480 + ], + "type": "text", + "content": "[34] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, et al. Sam 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714, 2024. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 481, + 294, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 481, + 294, + 525 + ], + "spans": [ + { + "bbox": [ + 56, + 481, + 294, + 525 + ], + "type": "text", + "content": "[35] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4104-4113, 2016. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 525, + 294, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 525, + 294, + 568 + ], + "spans": [ + { + "bbox": [ + 56, + 525, + 294, + 568 + ], + "type": "text", + "content": "[36] Aleksandar Shtedritski, Christian Rupprecht, and Andrea Vedaldi. What does clip know about a red circle? visual prompt engineering for vlms. arXiv preprint arXiv:2304.06712, 2023. 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 569, + 294, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 569, + 294, + 613 + ], + "spans": [ + { + "bbox": [ + 56, + 569, + 294, + 613 + ], + "type": "text", + "content": "[37] Simranjit Singh, Georgios Pavlakos, and Dimitrios Stamoulis. Evaluating zero-shot gpt-4v performance on 3d visual question answering benchmarks. arXiv preprint arXiv:2405.18831, 2024. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 613, + 294, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 613, + 294, + 635 + ], + "spans": [ + { + "bbox": [ + 56, + 613, + 294, + 635 + ], + "type": "text", + "content": "[38] Gemini Team, Rohan Anil, and et al. Gemini: A family of highly capable multimodal models, 2024. 4, 8" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 636, + 294, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 636, + 294, + 680 + ], + "spans": [ + { + "bbox": [ + 56, + 636, + 294, + 680 + ], + "type": "text", + "content": "[39] Xiaoyu Tian, Junru Gu, Bailin Li, Yicheng Liu, Chenxu Hu, Yang Wang, Kun Zhan, Peng Jia, Xianpeng Lang, and Hang Zhao. Drivevlm: The convergence of autonomous driving and large vision-language models, 2024. 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 56, + 680, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 680, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 680, + 294, + 713 + ], + "type": "text", + "content": "[40] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama:" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 633 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023.4,8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 96, + 555, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 555, + 140 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 555, + 140 + ], + "type": "text", + "content": "[41] Ramakrishna Vedantam, C Lawrence Zitnick, and Devi Parikh. Cider: Consensus-based image description evaluation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4566-4575, 2015. 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 141, + 555, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 555, + 174 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 555, + 174 + ], + "type": "text", + "content": "[42] Xiaohan Wang, Yuhui Zhang, Orr Zohar, and Serena Yeung-Levy. Videoagent: Long-form video understanding with large language model as agent, 2024. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 175, + 555, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 555, + 218 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 555, + 218 + ], + "type": "text", + "content": "[43] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Brian Ichter, Fei Xia, Ed Chi, Quoc Le, and Denny Zhou. Chain-of-thought prompting elicits reasoning in large language models, 2023. 6, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 219, + 555, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 219, + 555, + 252 + ], + "spans": [ + { + "bbox": [ + 316, + 219, + 555, + 252 + ], + "type": "text", + "content": "[44] Junbin Xiao, Xindi Shang, Angela Yao, and Tat-Seng Chua. Next-qa:next phase of question-answering to explaining temporal actions, 2021. 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 253, + 555, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 253, + 555, + 307 + ], + "spans": [ + { + "bbox": [ + 316, + 253, + 555, + 307 + ], + "type": "text", + "content": "[45] Runsen Xu, Xiaolong Wang, Tai Wang, Yilun Chen, Jiangmiao Pang, and Dahua Lin. Pointllm: Empowering large language models to understand point clouds. In European Conference on Computer Vision, pages 131-147. Springer, 2025. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 308, + 555, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 308, + 555, + 363 + ], + "spans": [ + { + "bbox": [ + 316, + 308, + 555, + 363 + ], + "type": "text", + "content": "[46] Jingkang Yang, Yuhao Dong, Shuai Liu, Bo Li, Ziyue Wang, Chencheng Jiang, Haoran Tan, Jiamu Kang, Yuanhan Zhang, Kaiyang Zhou, and Ziwei Liu. Octopus: Embodied vision-language programmer from environmental feedback, 2023. 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 365, + 554, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 365, + 554, + 398 + ], + "spans": [ + { + "bbox": [ + 316, + 365, + 554, + 398 + ], + "type": "text", + "content": "[47] Jinyu Yang, Mingqi Gao, Zhe Li, Shang Gao, Fangjing Wang, and Feng Zheng. Track anything: Segment anything meets videos. arXiv preprint arXiv:2304.11968, 2023. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 399, + 554, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 399, + 554, + 442 + ], + "spans": [ + { + "bbox": [ + 316, + 399, + 554, + 442 + ], + "type": "text", + "content": "[48] Jianwei Yang, Hao Zhang, Feng Li, Xueyan Zou, Chunyuan Li, and Jianfeng Gao. Set-of-mark prompting unleashes extraordinary visual grounding in gpt-4v. arXiv preprint arXiv:2310.11441, 2023. 6, 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 443, + 555, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 443, + 555, + 487 + ], + "spans": [ + { + "bbox": [ + 316, + 443, + 555, + 487 + ], + "type": "text", + "content": "[49] Zhou Yu, Jun Yu, Yuhao Cui, Dacheng Tao, and Qi Tian. Deep modular co-attention networks for visual question answering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6281–6290, 2019. 3, 4" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 488, + 555, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 488, + 555, + 531 + ], + "spans": [ + { + "bbox": [ + 316, + 488, + 555, + 531 + ], + "type": "text", + "content": "[50] Ce Zhang, Taixi Lu, Md Mohaiminul Islam, Ziyang Wang, Shoubin Yu, Mohit Bansal, and Gedas Bertasius. A simple llm framework for long-range video question-answering, 2024. 1, 4, 5" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 533, + 555, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 533, + 555, + 586 + ], + "spans": [ + { + "bbox": [ + 316, + 533, + 555, + 586 + ], + "type": "text", + "content": "[51] Yue Zhao, Ishan Misra, Philipp Krahenbuhl, and Rohit Girdhar. Learning video representations from large language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6586-6597, 2023. 5" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 588, + 555, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 588, + 555, + 633 + ], + "spans": [ + { + "bbox": [ + 316, + 588, + 555, + 633 + ], + "type": "text", + "content": "[52] Gengze Zhou, Yicong Hong, and Qi Wu. Navgpt: Explicit reasoning in vision-and-language navigation with large language models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 7641-7649, 2024. 5" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3792" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/CocoER_ Aligning Multi-Level Feature by Competition and Coordination for Emotion Recognition/91abd7f6-c62e-4947-ae7d-c03321b50ab1_content_list.json b/2025/CocoER_ Aligning Multi-Level Feature by Competition and Coordination for Emotion Recognition/91abd7f6-c62e-4947-ae7d-c03321b50ab1_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..956f3f92b6dbea3b1fe9d6e0988b6cb43840c234 --- /dev/null +++ b/2025/CocoER_ Aligning Multi-Level Feature by Competition and Coordination for Emotion Recognition/91abd7f6-c62e-4947-ae7d-c03321b50ab1_content_list.json @@ -0,0 +1,1697 @@ +[ + { + "type": "text", + "text": "CocoER: Aligning Multi-Level Feature by Competition and Coordination for Emotion Recognition", + "text_level": 1, + "bbox": [ + 109, + 128, + 885, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xuli Shen $^{1,2}$ Hua Cai $^{2}$ Weilin Shen $^{2}$ Qing Xu $^{2}$ Dingding Yu $^{2}$ Weifeng Ge $^{1\\star}$ Xiangyang Xue $^{1\\star}$ Fudan University $^{2}$ UniDT", + "bbox": [ + 99, + 200, + 895, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{x1shen20,wfge,xyxue}@fudan.edu.cn， {hua.cai,weilin.shen,qing.xu,dingding.yu}@unidt.com", + "bbox": [ + 106, + 241, + 880, + 258 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 291, + 326, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "With the explosion of human-machine interaction, emotion recognition has reignited attention. Previous works focus on improving visual feature fusion and reasoning from multiple image levels. Although it is non-trivial to deduce a person's emotion by integrating multi-level feature (head, body and context), the emotion recognition results of each level is usually different from one another, which creates inconsistency in the prevailing feature alignment method and decrease recognition performance. In this work, we propose a multi-level image feature refinement method for emotion recognition (CocoER) to mitigate the impact caused by conflicting results from multi-level recognition. First, we leverage cross-level attention to improve visual feature consistency between hierarchically cropped head, body and context windows. Then, vocabulary informed alignment is incorporated into the recognition framework to produce pseudo label and guide hierarchical visual feature refinement. To effectively fuse multi-level feature, we elaborate on a competition process of eliminating irrelevant image level predictions and a coordination process to enhance the feature across all levels. Extensive experiments are executed on two popular datasets, and our method achieves state-of-the-art performance with multi-level interpretation results. Code is available at: https://github.com/bisno/CocoER.", + "bbox": [ + 89, + 323, + 485, + 686 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 715, + 220, + 729 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Visual emotion recognition plays a vital role in human computer interaction [27], surveillance [11], and robotics [5]. It aims to perceive the emotion state of human, and machine can modify responses and behaviors by considering the sentiment attitudes. Recently, impressive progress in recognizing emotions is achieved with expression from additional modality, such as head, body, background, depth, and so on [24]. Besides, emotion representation can be enhanced in view of discrete categories [24], points in a continuous", + "bbox": [ + 89, + 739, + 483, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/858bcb9d3548b96e064815fc3c03f48ebe0cad3e87f19d9a30aa07ab24d8ca60.jpg", + "image_caption": [ + "Figure 1. To avoid conflicts arising from inconsistent multi-level predictions, competition and coordination are applied in refining multi-level visual representation for emotion recognition, guided by pseudo label from vocabulary-informed alignment." + ], + "image_footnote": [], + "bbox": [ + 529, + 290, + 890, + 520 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "space [13, 42] and label distributions [41].", + "bbox": [ + 511, + 608, + 794, + 623 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "As emotion is naturally coupled in human and environments, the complex background in image still hinders the research of emotion recognition. Pioneering researches focus on perceiving the emotion states of the specific individuals in the given image with hierarchical visual feature [24, 37, 40] on context-aware emotion recognition datasets [13, 15]. [14] trains different backbones to combine the information of the target body and the entire image. Subsequently, a context-aware emotion recognition architecture enhances the classified precision to a high level representation [24]. In addition to image feature, knowledge graph is incorporated to reason the emotion states in [36]. To align feature across different modalities, [37] aims to address distribution gap between heterogeneous modalities.", + "bbox": [ + 509, + 626, + 906, + 838 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite previous progress in decoupling irrelevant contextual information in the inference stage [38, 40], there still lacks of multi-level image feature alignment method to precisely recognize human emotions. Each level has dif", + "bbox": [ + 511, + 839, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Corresponding authors.", + "bbox": [ + 114, + 887, + 246, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "29591", + "bbox": [ + 478, + 944, + 517, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ferent and potentially misleading representation and predictions, which decreases accuracy of emotion recognition. As shown in Fig. 1, multi-level emotion output from the head (fear and doubt), and context (sad) confuse final emotion prediction. We hypothesize that conflicting results from each level interfere with emotional decision-making process. To better align multi-level feature, researchers often draws upon prior knowledge, such as knowledge graph, to select relevant object. Although knowledge graph [36] is applied to enhance interpretability, extracting object relevance does not directly address the challenge of conflicted multi-level predictions and there is a lack of a guidance that can combine head, body, and background feature.", + "bbox": [ + 89, + 90, + 480, + 287 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we propose competition and coordination mechanism to effectively perform multi-level fusion of head, body, and context emotion representation. Specifically, we apply cross-level attention mechanism with linear mapping on extracted feature maps to extract correlation from multi-level windows and obtain multi-level embeddings and recognition results. Furthermore, we leverage prototype learning to enhance the recognition result from external image-text aligned model, leading to pseudo label as guidance for multi-level feature refinement. Next, based on the pseudo label, we introduce competition process as an emotional feature selection method that filters out inconsistent multi-level feature. We leverage coordination process to update the parameters of the linear mapping for all image levels in inference-stage. At last, we re-send the multi-level feature to the updated linear mapping and achieve emotional consistency for final emotion recognition. In summary, the contributions of our paper are:", + "bbox": [ + 91, + 289, + 482, + 559 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We enhance multi-level image correlation via hierarchical feature extraction module, and produce multi-level recognition. We leverage prediction from vocabulary-informed module as pseudo label to identify inconsistent levels.", + "- We introduce workspace module that incorporates competition and coordination process to rectify multi-level feature through updated model in inference-stage.", + "- Extensive experiments have been conducted to compare with baseline methods and large vision-language models (VLM), and the outcomes demonstrate superior performance and interpretability of our proposed method." + ], + "bbox": [ + 89, + 561, + 482, + 728 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 744, + 232, + 760 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Visual Emotion Recognition", + "text_level": 1, + "bbox": [ + 89, + 771, + 341, + 787 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Visual emotion recognition is a comprehensive task in computer vision area, which is deeply associated with cognitive intelligence, and has attracted increasing attentions from researchers. Extensive works utilize facial expression to deduce the emotion state of specific person [7, 23, 25, 32, 35, 38]. Considering that recognizing a person's emotions can be subjective and has raised inherent ambiguity among sim", + "bbox": [ + 89, + 794, + 480, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ilar emotion categories, some researches focus on the latent distribution on the candidate emotion label sets [32, 41] and incorporates models from cognitive computing into the recognition algorithms [7, 21, 24, 26, 41]. Regarding network structures, [3, 18, 39] present transformer-based architectures to capture contextual embeddings. [10, 34, 37] utilize multi-head attention to fuse multiple modalities. In the aspect of inference stage, [38, 40] use causal relationships to eliminate information in the background that is inconsistent with target emotions. In contrast, we focuses on emotion recognition via multi-level visual representation that aligned by vocabulary-informed pseudo label.", + "bbox": [ + 511, + 90, + 903, + 272 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Multi-modal Emotion Recognition", + "text_level": 1, + "bbox": [ + 511, + 281, + 815, + 297 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "[22, 23, 26, 28, 40] enhance multi-modal fusion performance, and [12, 14, 24, 30] explore context-based visual expressions. To fully take advantages of various context information, multi-modal visual feature are incorporated into model inputs, such as face, pose and depth [24, 40]. Different modality information requires different sensors or off-the-shelf extractor, resulting in increasing model complexity and a narrowed application range. On the other hand, [14, 30] focus on addressing this problem with only global and body information, and are still able to achieve competitive performance. [36] introduces a hierarchical context-based emotion recognition method that uses scene graphs to infer emotions from images.", + "bbox": [ + 511, + 301, + 903, + 498 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In the field of large vision language model [17, 20], despite competitive performance in many multi-modal tasks, existing models do not prioritize emotion recognition. The modality alignment methods for large vision language models mostly focus on tokenized pair in the whole image, neglecting the multi-level image feature that misleads the image-text alignment. In this paper, we target this problem by eliminating irrelevant level of feature and further propose a novel multi-level feature alignment framework to enhance the performance of emotion recognition.", + "bbox": [ + 511, + 500, + 903, + 650 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Methods", + "text_level": 1, + "bbox": [ + 511, + 662, + 611, + 678 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this section, we first introduce the hierarchical feature extraction module on the cropped multi-level image, where each image level has its own emotion recognition outcome. Next, we elaborate on the vocabulary-informed alignment module, which serves as guidance that provides pseudo label for eliminating inconsistent multi-level predictions. Finally, based on such pseudo label, we apply the workspace module to enhance multi-level feature fusion. The proposed framework is depicted in Fig. 2.", + "bbox": [ + 511, + 688, + 903, + 825 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3.1. Hierarchical Feature Extraction", + "text_level": 1, + "bbox": [ + 511, + 833, + 795, + 848 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We construct hierarchical emotion recognition by dividing the input image into three windows: head, body, and context. Specifically, given an input image $\\mathbf{I} \\in \\mathbb{R}^{H \\times W \\times 3}$ , we", + "bbox": [ + 511, + 854, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "29592", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/6bb02efdd310c837e29ea24d7a49f7d5f308150aab87a3ccb8d701ffbc66a8b9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 158, + 89, + 519, + 292 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/90889120229087d16b4b638f4e5a4e755c842f7e65f248648edf292a5ebde9be.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 522, + 90, + 843, + 292 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/ce603e5c4c4e0963bfcec4b26747f4d3ac12ed83ab52f29cb8223ace58c7fa91.jpg", + "image_caption": [ + "Figure 2. Illustration of the proposed framework." + ], + "image_footnote": [], + "bbox": [ + 156, + 294, + 834, + 496 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "leverage object detection model [4] to obtain head, body, and context windows using detected bounding boxes. All three cropped windows are resized to $(H,W)$ , and are then fed into trainable backbones $\\mathcal{R}$ to extract multi-level feature maps $\\mathbf{u}_m$ with a size of $(h,w)$ and $c$ channels:", + "bbox": [ + 89, + 537, + 483, + 613 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {u} _ {m} = \\mathcal {R} _ {m} (\\mathbf {I} _ {m}), m \\in \\{\\mathbf {c}, \\mathbf {b}, \\mathbf {h} \\}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 179, + 621, + 480, + 638 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbf{c},\\mathbf{b},\\mathbf{h}$ are context, body, and head windows, respectively. Each $\\mathbf{u}_m\\in \\mathbb{R}^{h\\times w\\times c}$ is divided and reshaped into $N$ non-overlapping patches $\\mathbf{u}_m^p\\in \\mathbb{R}^{p^2\\cdot c}$ and then mapped into $d_h$ dimensions with linear projection $\\mathbf{E}_m\\in \\mathbb{R}^{(p^2\\cdot c)\\times d_h}$ . Thus each level $m$ consists of a sequence of 2D embeddings $\\bar{\\mathbf{u}}_m = [\\mathbf{u}_m^{p,0}\\mathbf{E}_m;\\mathbf{u}_m^{p,1}\\mathbf{E}_m;\\dots ;\\mathbf{u}_m^{p,N - 1}\\mathbf{E}_m]\\in \\mathbb{R}^{N\\times d_h}$ , where $N = hw / p^2$ is patch number [6]. $\\bar{\\mathbf{u}}_m$ is sent to cross-level attention $\\mathcal{T}$ for enhancing multi-level correlation via connecting local and context feature. For instance, the head-context attention $\\mathcal{T}_{\\mathbf{h},\\mathbf{c}}$ is formalized with linear projections $\\{W_{\\mathbf{h}}^{0},W_{\\mathbf{h}}^{1},W_{\\mathbf{c}}\\} \\in \\mathbb{R}^{d_h\\times d_h}$ with feature maps $\\{\\bar{\\mathbf{u}}_{\\mathbf{h}},\\bar{\\mathbf{u}}_{\\mathbf{c}}\\}$ :", + "bbox": [ + 89, + 645, + 483, + 830 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {Q} = \\bar {\\mathbf {u}} _ {\\mathbf {h}} W _ {\\mathbf {h}} ^ {0}, \\mathbf {K} = \\bar {\\mathbf {u}} _ {\\mathbf {c}} W _ {\\mathbf {c}}, \\mathbf {V} = \\bar {\\mathbf {u}} _ {\\mathbf {h}} W _ {\\mathbf {h}} ^ {1} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 155, + 835, + 482, + 854 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\gamma = \\operatorname {s o f t m a x} \\left(\\mathbf {Q} \\mathbf {K} ^ {T}\\right) / \\sqrt {d _ {h}} \\quad \\in \\mathbb {R} ^ {N \\times N} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 156, + 857, + 482, + 876 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {\\mathbf {z}} _ {\\mathbf {h}} = \\boldsymbol {\\gamma} \\mathbf {V} \\in \\mathbb {R} ^ {N \\times d _ {h}}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 151, + 878, + 482, + 896 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Then $\\bar{\\mathbf{z}}_{\\mathbf{h}}$ is added with $\\bar{\\mathbf{u}}_{\\mathbf{h}}$ and sent to typical fully-connected layers with skip-connection:", + "bbox": [ + 511, + 537, + 906, + 568 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\dot {\\mathbf {z}} _ {\\mathbf {h}} = \\bar {\\mathbf {z}} _ {\\mathbf {h}} + \\bar {\\mathbf {u}} _ {\\mathbf {h}} \\in \\mathbb {R} ^ {N \\times d _ {h}} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 594, + 579, + 903, + 597 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {z} _ {\\mathbf {h}} = \\operatorname {F F N} _ {0} (\\operatorname {L N} (\\dot {\\mathbf {z}} _ {\\mathbf {h}})) + \\dot {\\mathbf {z}} _ {\\mathbf {h}} \\in \\mathbb {R} ^ {d}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 594, + 601, + 903, + 619 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where LN is LayerNorm, $\\mathrm{FFN}_0$ refers to the feed-forward network and $\\mathbf{z_h} \\in \\mathbb{R}^d$ is the head-level representation. Thus we bridge hierarchical information to enhance head-level representation $\\mathbf{z_h}$ via head-context cross attention $\\mathcal{T}_{\\mathbf{h},\\mathbf{c}}$ . Similarly we have $\\mathbf{z_b}$ via body-context cross attention $\\mathcal{T}_{\\mathbf{b},\\mathbf{c}}$ . Therefore we leverage attention mechanism to enrich local feature representation $\\{\\bar{\\mathbf{u}}_{\\mathbf{h}},\\bar{\\mathbf{u}}_{\\mathbf{c}}\\}$ . Note that we conduct self-attention $\\mathcal{T}_{\\mathbf{c},\\mathbf{c}}$ and obtain context feature $\\mathbf{z_c}$ . Next, we pass each level representation $\\mathbf{z}_m$ through a linear mapping $g_{\\theta_m}$ , followed by an operation $\\mathcal{F}$ to produce $k$ emotion category corresponding to multi-class task or multi-label task:", + "bbox": [ + 511, + 631, + 906, + 797 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\ny _ {m} = \\mathcal {F} \\left(g _ {\\boldsymbol {\\theta} _ {m}} \\left(\\mathbf {z} _ {m}\\right)\\right), m \\in \\{\\mathbf {c}, \\mathbf {b}, \\mathbf {h} \\}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 810, + 903, + 828 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $y_{m}$ represents the prediction of each level $m$ . The parameter of linear mapping $g_{\\theta_m}$ will be updated in inference-stage, and each level representation $\\mathbf{z}_m$ will be refined for enhancing emotion recognition.", + "bbox": [ + 511, + 840, + 906, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "29593", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Vocabulary-Informed Alignment", + "text_level": 1, + "bbox": [ + 89, + 90, + 379, + 107 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In order to semantically guide feature refinement from hierarchical feature extraction, we propose a vocabulary-informed alignment method that leverages pre-trained image and text encoder to gain pseudo label and calibrate image representation. Specifically, given same input image $\\mathbf{I} \\in \\mathbb{R}^{H \\times W \\times 3}$ , we improve CLIP [29] prediction to guide image feature alignment from text space. As shown in Fig. 2, we first encode the image through an CLIP image encoder $\\mathbf{Enc}_{\\mathrm{img}}$ . Then, we provide a text prompt for $k$ classes of emotion recognition task. The $k$ classes are injected and expanded into $k$ sentences of \"This photo shows {} emotion\", denoted as $\\mathbf{S}_k$ , and are sent to CLIP text encoder $\\mathbf{Enc}_{\\mathrm{text}}$ . Thus, we have", + "bbox": [ + 89, + 112, + 483, + 308 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {u} _ {\\mathrm {i m g}} = \\mathbf {E n c} _ {\\mathrm {i m g}} (\\mathbf {I}) \\in \\mathbb {R} ^ {d} \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 189, + 316, + 480, + 333 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {U} _ {\\text {t e x t}} = \\mathbf {E n c} _ {\\text {t e x t}} \\left(\\mathbf {S} _ {k}\\right) \\in \\mathbb {R} ^ {d \\times k}. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 189, + 335, + 480, + 354 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "One can directly use dot-product similarity to obtain recognition result of CLIP. However, merely applying CLIP prediction is insufficient because it is trained on open-world image-text pairs, which leads to difficulties in differentiating precise expressions such as \"surprise\" and \"fear\". To improve the recognition performance in vocabulary-informed module, we apply prototype learning [8] that projects image representation into text representation space and pushes closer to emotion prototype. In detail, after extracting image feature, we expand $\\mathbf{u}_{\\mathrm{img}}$ into $\\hat{\\mathbf{U}}_{\\mathrm{img}}$ which has the same shape of $\\mathbf{U}_{\\mathrm{text}}$ and aim to minimize the distance $\\mathcal{D}(\\hat{\\mathbf{U}}_{\\mathrm{img}},\\mathbf{U}_{\\mathrm{text}})$ :", + "bbox": [ + 89, + 362, + 483, + 545 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {D} \\left(\\hat {\\mathbf {U}} _ {\\mathrm {i m g}}, \\mathbf {U} _ {\\text {t e x t}}\\right) = \\left\\| \\omega^ {T} \\hat {\\mathbf {U}} _ {\\mathrm {i m g}} - \\mathbf {U} _ {\\text {t e x t}} \\right\\| _ {2} ^ {2}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 156, + 551, + 480, + 571 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\hat{\\mathbf{U}}_{\\mathrm{img}} \\in \\mathbb{R}^{d \\times k}$ , and $\\omega \\in \\mathbb{R}^{d \\times d}$ is a linear projection which is trained by minimizing the above distance of image feature and semantic emotion prototypes. Upon obtaining the trained $\\omega$ , we feed the emotion-aligned image feature $\\omega^T \\hat{\\mathbf{U}}_{\\mathrm{img}}$ to $\\mathrm{FFN}_1$ . Similar to (7) in the previous module that each hierarchical window has its own recognition result, the vocabulary-informed module also provides emotion recognition result:", + "bbox": [ + 89, + 579, + 483, + 700 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ny _ {\\mathbf {t}} = \\mathcal {F} \\left(\\operatorname {F F N} _ {1} \\left(\\omega^ {T} \\hat {\\mathbf {U}} _ {\\mathrm {i m g}}\\right)\\right). \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 197, + 708, + 480, + 727 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$y_{\\mathbf{t}}$ serves as pseudo label for further feature refinement in workspace. We use the embeddings of last hidden layer of $\\mathrm{FFN}_1$ as the vocabulary-informed feature $\\mathbf{z}_{\\mathbf{t}}$ , which has same feature dimension of image level feature $\\mathbf{z} \\in \\mathbb{R}^d$ . Note that we freeze parameters of pre-trained $\\omega$ , as well as $\\mathbf{Enc}_{\\mathrm{img}}$ and $\\mathbf{Enc}_{\\mathrm{text}}$ during the training stage of our framework, leaving $\\mathrm{FFN}_1$ trainable.", + "bbox": [ + 89, + 734, + 483, + 840 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Workspace", + "text_level": 1, + "bbox": [ + 89, + 848, + 212, + 864 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Different from the previous feature refinement methods, such as the entity-context feature alignment [36] or emotion", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/0027eb2ef9440995765100063b19b441b345f207b99a5519d7af155cf2992193.jpg", + "image_caption": [ + "Figure 3. Detailed illustration of competition and coordination process, incorporated with inference-stage parameter update." + ], + "image_footnote": [], + "bbox": [ + 526, + 88, + 893, + 277 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "feature disentanglement [16, 37], we introduce a workspace module that can adaptively refine all levels of emotion feature based on the vocabulary-informed classification results, as illustrated in the bottom part of Fig. 2. The name \"workspace\" is inspired by the Workspace Theory [1, 2], which describes how multi-level predictions compete and coordinate for final recognition.", + "bbox": [ + 511, + 328, + 905, + 433 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Formally, given the multi-level emotion feature set $\\mathbf{Z} = \\{\\mathbf{z}_{\\mathrm{c}},\\mathbf{z}_{\\mathrm{b}},\\mathbf{z}_{\\mathrm{h}}\\}$ , we aim to gradually eliminate inconsistent level of feature in $\\mathbf{Z}$ according to the pseudo label $y_{t}$ (11) until there is only one element in $\\mathbf{Z}$ . We refer to this process as multi-level feature competition. Subsequently, we leverage the selection process to update parameters of $g_{\\theta_m}$ and re-send each $\\mathbf{z}_m$ to $g_{\\theta_m}$ to obtain the refined feature $\\hat{\\mathbf{z}}_m$ . We call this multi-level feature coordination. Technical details are provided in the next two sections.", + "bbox": [ + 511, + 434, + 906, + 569 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3.1. Competition Process", + "text_level": 1, + "bbox": [ + 511, + 575, + 702, + 590 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The competition process is designed as an iteration and contains two steps during each iteration: 1) Determine the descent direction for updating parameters of $g_{\\theta_m}$ ; 2) Eliminate inconsistent element in feature set $\\mathbf{Z}$ . The iteration ends when there is only one element in $\\mathbf{Z}$ , shown in Fig. 3.", + "bbox": [ + 511, + 594, + 905, + 669 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Step 1. We leverage the value of loss function $\\mathcal{L}_{\\mathrm{hier}}(\\mathcal{F}(g_{\\theta}(\\mathbf{z})),y_{\\mathbf{t}})$ as a measure of inconsistency between multi-level recognition and pseudo label. Then, in order to target the most inconsistent level, we take the maximum of the loss function groups for the results from $\\mathcal{F}(g_{\\theta_m}(\\mathbf{z}_m))$ against pseudo label $y_{\\mathbf{t}}$ :", + "bbox": [ + 511, + 670, + 906, + 760 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\max _ {\\mathbf {z} _ {m} \\in \\mathbf {Z}} \\left\\{\\mathcal {L} _ {\\text {h i e r}} \\left(g _ {\\boldsymbol {\\theta} _ {m}} \\left(\\mathbf {z} _ {m}\\right), y _ {\\mathbf {t}}\\right) \\right\\}. \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 616, + 767, + 903, + 789 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For the convenience of notation, let $\\mathbf{f}_{\\mathbf{z}_m}(\\pmb{\\theta}_m) = \\{\\mathcal{L}_{\\mathrm{hier}}(g_{\\pmb{\\theta}_m}(\\mathbf{z}_m),y_t),\\mathbf{z}_m\\in \\mathbf{Z}\\}$ . During each iteration in the competition process, we aim to calculate the gradient of $\\max_{\\mathbf{z}_m\\in \\mathbf{Z}}\\mathbf{f}_{\\mathbf{z}_m}(\\pmb{\\theta}_m)$ so as to update parameters of $g_{\\pmb{\\theta}_m}$ . Thus, our goal transforms into a min-max problem:", + "bbox": [ + 511, + 797, + 905, + 873 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\boldsymbol {\\theta} _ {m}} \\max _ {\\mathbf {z} _ {m} \\in \\mathbf {Z}} \\mathbf {f} _ {\\mathbf {z} _ {m}} \\left(\\boldsymbol {\\theta} _ {m}\\right). \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 640, + 880, + 903, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "29594", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "However, there exists a non-trivial challenge in finding the descent direction when we implement $\\max(\\cdot)$ on the groups of loss functions because the above function is nondifferentiable. Inspired by [33], we bring differentiability to $\\max_{\\mathbf{z}_m \\in \\mathbf{Z}} \\mathbf{f}_{\\mathbf{z}_m}(\\theta_m)$ and leverage a regularization term to stabilize the descent direction. Please refer to technical details in Appendix (Sec.1). Then, the min-max problem is reduced to the following quadratic programming (QP) problem with a Lagrange multiplier $\\lambda$ :", + "bbox": [ + 89, + 90, + 483, + 227 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\min _ {\\boldsymbol {\\lambda}} \\left(\\frac {1}{2} \\boldsymbol {\\lambda} ^ {T} \\mathbf {G} \\mathbf {G} ^ {T} \\boldsymbol {\\lambda} - \\boldsymbol {\\lambda} ^ {T} \\mathbf {f} _ {\\mathbf {z} _ {m}}\\right) \\tag {14} \\\\ \\text {s . t .} \\quad \\sum_ {\\mathbf {z} _ {m} \\in \\mathbf {Z}} \\lambda_ {\\mathbf {z} _ {m}} = 1, \\lambda_ {\\mathbf {z} _ {m}} \\geq 0 \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 181, + 239, + 483, + 303 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{G} = \\nabla_{\\boldsymbol{\\theta}_m^l}\\mathbf{f}_{\\mathbf{z}_m} \\in \\mathbb{R}^{q \\times (d \\cdot k)}$ contains gradient of the prediction layer $\\boldsymbol{\\theta}_m^l$ with respect to the loss of $\\mathbf{f}_{\\mathbf{z}_m}$ and $q = |\\mathbf{Z}|$ refers to the number of elements left in $\\mathbf{Z}$ . After solving for $\\boldsymbol{\\lambda}$ in the QP problem, we have $\\delta = -\\mathbf{G}^T\\boldsymbol{\\lambda} \\in \\mathbb{R}^{d \\cdot k}$ as the approximate solution of (13). Here, $\\boldsymbol{\\delta}$ serves as the shared parameter updating path for linear mapping $g_{\\boldsymbol{\\theta}_m}$ of all three levels, which will be used for feature refinement in coordination process.", + "bbox": [ + 89, + 316, + 483, + 439 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Step 2. Let $n = \\arg \\max_{\\mathbf{z}_m \\in \\mathbf{Z}} \\left\\{ \\mathcal{L}_{\\mathrm{hier}}(g_{\\theta_m}(\\mathbf{z}_m), y_t) \\right\\}$ be the index of most inconsistent feature in $\\mathbf{Z}$ . We iteratively eliminate the corresponding $n$ of feature $\\mathbf{z}$ in $\\mathbf{Z}$ , denoted as $\\mathbf{Z}_{-n}$ . This elimination is illustrated in the top part of Fig. 3.", + "bbox": [ + 89, + 441, + 483, + 502 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3.2. Coordination Process", + "text_level": 1, + "bbox": [ + 89, + 511, + 287, + 523 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The goal of coordination process is to rectify the multi-level feature $\\mathbf{z}$ via updated linear mapping $g_{\\theta}$ . Specifically, after iteratively obtaining the respective descent direction $\\delta$ during the competition process, we accumulate the overall descent direction, shown in the bottom part of Fig. 3. For example, if the context-level feature $\\mathbf{z}_{\\mathbf{c}}$ and head-level feature $\\mathbf{z_h}$ is gradually eliminated through competition, we have $\\delta_{\\mathbf{c}}$ and $\\delta_{\\mathbf{h}}$ based on (14), which construct the shared parameter updating path for the prediction layer $\\theta_m^l$ as follows:", + "bbox": [ + 89, + 530, + 483, + 667 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\delta^ {l} = \\delta_ {\\mathrm {c}} + \\delta_ {\\mathrm {h}}. \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 238, + 679, + 480, + 696 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Then, for each image level $m$ we update entire parameter $\\theta_{m}$ via chain rule and back-propagation, denoted as $\\longleftarrow$ :", + "bbox": [ + 89, + 709, + 483, + 739 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\theta} _ {\\mathbf {c}} \\leftarrow \\theta_ {\\mathbf {c}} ^ {l} + \\delta^ {l}; \\tilde {\\theta} _ {\\mathbf {b}} \\leftarrow \\theta_ {\\mathbf {b}} ^ {l} + \\delta^ {l}; \\tilde {\\theta} _ {\\mathbf {h}} \\leftarrow \\theta_ {\\mathbf {h}} ^ {l} + \\delta^ {l}. \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 119, + 751, + 480, + 770 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Next, we re-send $\\mathbf{z}_m$ to its corresponding linear mapping $g_{\\tilde{\\theta}}$ and acquire refined feature from last hidden layer and construct multi-level concatenated representation $[\\hat{\\mathbf{z}}_{\\mathrm{c}};\\hat{\\mathbf{z}}_{\\mathrm{b}};\\hat{\\mathbf{z}}_{\\mathrm{h}}]$ :", + "bbox": [ + 89, + 782, + 483, + 829 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {z}} _ {\\mathrm {c}} = g _ {\\tilde {\\theta} _ {\\mathrm {c}}} \\left(\\mathbf {z} _ {\\mathrm {c}}\\right); \\hat {\\mathbf {z}} _ {\\mathrm {b}} = g _ {\\tilde {\\theta} _ {\\mathrm {b}}} \\left(\\mathbf {z} _ {\\mathrm {b}}\\right); \\hat {\\mathbf {z}} _ {\\mathrm {h}} = g _ {\\tilde {\\theta} _ {\\mathrm {h}}} \\left(\\mathbf {z} _ {\\mathrm {h}}\\right). \\tag {17}\n$$\n", + "text_format": "latex", + "bbox": [ + 143, + 845, + 480, + 864 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "With such inference-stage parameter updating, we calibrate multi-level emotion representation, ensuring that the final", + "bbox": [ + 89, + 869, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "emotion recognition results are less affected by the inconsistent emotion recognition outcomes from hierarchical feature. Please refer to algorithm in Appendix that describes competition and coordination process within workspace.", + "bbox": [ + 511, + 90, + 905, + 151 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Training Objectives", + "text_level": 1, + "bbox": [ + 511, + 160, + 702, + 176 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We concatenate the refined multi-level representation with vocabulary-informed feature $\\mathbf{z_t}$ in $\\mathbf{z_G} = [\\hat{\\mathbf{z}}_{\\mathrm{c}};\\hat{\\mathbf{z}}_{\\mathrm{b}};\\hat{\\mathbf{z}}_{\\mathrm{h}};\\mathbf{z}_{\\mathrm{t}}]$ . Then $\\mathbf{z_G}$ is passed to feed-forward network, and output of the proposed method is:", + "bbox": [ + 511, + 181, + 906, + 242 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\ny _ {\\mathbf {G}} = \\mathcal {F} \\left(\\operatorname {F F N} _ {2} \\left(\\mathbf {z} _ {\\mathbf {G}}\\right)\\right). \\tag {18}\n$$\n", + "text_format": "latex", + "bbox": [ + 635, + 253, + 903, + 270 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We collectively form loss functions from HFE, V-I and workspace module as $\\mathcal{L}_{\\mathbf{z}}(y_{\\mathbf{z}},y)$ with respect to ground-truth $y$ . The combined loss function $\\mathcal{L}_{\\mathrm{emo}}$ is defined as follows:", + "bbox": [ + 511, + 280, + 905, + 325 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {e m o}} = \\mathcal {L} _ {\\mathbf {c}} + \\mathcal {L} _ {\\mathbf {b}} + \\mathcal {L} _ {\\mathbf {h}} + \\mathcal {L} _ {\\mathbf {t}} + \\mathcal {L} _ {\\mathbf {G}}. \\tag {19}\n$$\n", + "text_format": "latex", + "bbox": [ + 593, + 337, + 903, + 353 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Next, we add norm of $\\delta^l$ [43] to regularize the shared updating path, aiming to avoid over-dependence on the vocabulary-informed pseudo label. Finally, all parameters of our proposed framework are trained based on:", + "bbox": [ + 511, + 363, + 905, + 424 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {\\mathrm {e m o}} + \\alpha \\| \\boldsymbol {\\delta} ^ {l} \\| _ {2}, \\tag {20}\n$$\n", + "text_format": "latex", + "bbox": [ + 638, + 434, + 903, + 452 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\alpha$ is hyperparameter, empirically set to 0.1.", + "bbox": [ + 511, + 462, + 843, + 478 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 491, + 645, + 507 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Setup", + "text_level": 1, + "bbox": [ + 511, + 515, + 594, + 531 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. EMOTIC [12] comprises 23,571 images featuring 34,320 individuals, predominantly captured in unconstrained environments. Each image encompasses multiple emotions, presenting a more intricate challenge of multi-label emotion recognition across 26 discrete emotion categories. CAER-S [15] is derived by extracting static frames from video clips in the CAER dataset [15], containing over 70,000 images. These images are annotated with an expanded set of 7 emotion categories: anger, disgust, fear, happiness, sadness, surprise, and neutral. Both datasets are divided into training (80%) and testing (20%) subsets. Implementation Details. Please refer to Appendix.", + "bbox": [ + 511, + 537, + 906, + 719 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Comparison with State-of-the-Arts", + "text_level": 1, + "bbox": [ + 511, + 727, + 816, + 743 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "EMOTIC test Split. We first compare our method with several multimodal or multi-object based emotion recognition competitors on EMOTIC test split. As shown in Tab. 1, we report the average precision (AP) over all categories as well as the average results (mAP). From the table, we can observe that: (1) Our method achieves state-of-the-art performance with $37.32\\%$ on mAP. Compared with [9] which utilizes Faster R-CNN to detect all possible objects in images for reasoning, we surpass it by a margin of $2.16\\%$ using only three hierarchical cropped windows. Moreover, [24]", + "bbox": [ + 511, + 750, + 906, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "29595", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/04a3462b63a13dfa5cc5435ac56004aea03aef50a1ef460993eb65dfa9f56081.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
CategoriesLi et al. [19] Hoang et al. [9] Mittal et al. [24]† Wu et al. [36]BLIP2-6.7b [17] LLaVa-1.6-7b [20] GPT-4o [27]
2021202120202024202220232024
Affection37.9344.4845.2332.2524.2534.3222.53
Anger13.7330.7115.4610.3708.3004.9810.32
Annoyance20.8726.4721.9212.8408.9912.3709.98
Anticipation61.0859.8972.1256.8391.3591.8754.16
Aversion09.6112.4317.8107.7006.4506.4603.51
Confidence80.0879.2468.6577.5159.3762.0568.09
Disapproval21.5424.5419.8212.4707.5610.3808.64
Disconnection28.3234.2443.1226.4125.8531.0517.03
Disquietment22.5718.7324.2316.7513.5314.4517.17
Doubt/Confusion33.5025.4235.1218.6218.5819.0631.20
Embarrassment04.1604.2614.3702.8005.4505.4402.17
Engagement88.1288.7191.1286.8395.1496.1080.77
Esteem20.5017.9923.6215.5023.1923.2718.30
Excitement80.1174.2183.2670.7364.5170.3669.59
Fatigue17.5122.6216.2312.6407.6811.0309.53
Fear15.5613.9223.6505.4807.8408.4912.79
Happiness76.0183.0274.7169.4570.7775.1453.65
Pain14.5616.6813.2110.6405.5205.8107.45
Peace26.7628.9134.2724.6320.3221.2628.91
Pleasure55.6455.4765.5344.3343.3043.4942.86
Sadness30.8042.8723.4121.2107.0618.7210.70
Sensitivity09.5915.8908.3205.8904.7704.7503.96
Suffering30.7046.2326.3921.4805.2206.1709.74
Surprise17.9216.2717.3707.9415.4216.1417.66
Sympathy15.2615.3734.2813.2421.4821.6209.10
Yearning10.1110.0414.2909.1213.0512.7907.93
mAP32.4135.1635.4826.6825.9627.9923.80
", + "bbox": [ + 94, + 88, + 903, + 512 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/fa441faf561ae2b29f4f136c87747defc0123a42728dbe558e456fbb62003f74.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 96, + 566, + 326, + 700 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/d63616a10c6b7c497bf675e7a05d0342f625edfe259fc558b8748042d8e47b59.jpg", + "table_caption": [ + "Table 1. Performance comparisons on EMOTIC test split. The 1st/2nd best results are highlighted in red/blue. †' denotes the method that contains additional visual modalities, such as depth, which is not used in our method." + ], + "table_footnote": [], + "table_body": "
GTPredictionHeadBodyContextV-I (Pseudo label)
AffectionAffectionDisconnectionAffectionEngagementAffection
AnticipationConfidenceFatigueConfidenceExcitementConfidence
ConfidenceExcitementPleasureHappinessExcitement
EngagementHappinessPeaceHappiness
ExcitementPeacePleasure
HappinessPleasure
PeaceLhier=2.35Lhier=0.61Lhier=1.33
Pleasure
Exclusion Sequence: (1) Head→(2) Context→(3) Body
", + "bbox": [ + 328, + 566, + 898, + 700 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/7de892624f59c881d029e55340def0523c5da2e1b00bf7544a8c4d575e036438.jpg", + "image_caption": [ + "Figure 4. Case study of randomly sampled results on EMOTIC. The V-I column shows pseudo label in vocabulary-informed module. We display the elimination of inconsistent recognition results in orange, which are not shown in either the V-I column or the prediction column." + ], + "image_footnote": [], + "bbox": [ + 96, + 708, + 326, + 840 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/3f8deb009b6cb91653a1787612cb5f76db36b6ed70b95b597a141a00ce137780.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
GTPredictionHeadBodyContextV-I (Pseudo label)
AnticipationAnticipationAnticipationAffectionEngagementAnticipation
ConfidenceConfidenceConfidenceSensitivityEmbarrassmentEngagement
EngagementEngagementExcitementSufferingPleasureEsteem
ExcitementExcitementHappinessHappiness
PeacePleasure
Lhier=1.42Lhier=2.36Lhier=1.03
Exclusion Sequence: (1) Body→(2) Context→(3) Head
", + "bbox": [ + 328, + 708, + 898, + 840 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "29596", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/003ef8adafd7ccda900b086d91734471a3e1626ea0fb95bfb229a83f72800d8c.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsC-F1 (%)O-F1 (%)
Kosti et al. [14] (2019)29.3354.45
Li et al. [19] (2021)15.1048.07
Wu et al. [36] (2024)34.2454.78
Ours36.7757.23
", + "bbox": [ + 99, + 88, + 475, + 186 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/346dc86a4cffd6893c55a6ed37da30fa12bdb64c1d4333012b7d728542de61d3.jpg", + "table_caption": [ + "Table 2. Label-based F1 score (C-F1) and example-based F1 score (O-F1) on EMOTIC test split." + ], + "table_footnote": [], + "table_body": "
MethodsAccuracy (%)
CLIP-ViT-L [29] (2021)21.57
BLIP2-6.7b [17] (2022)14.21
LLaVa-1.6-7b [20] (2023)28.59
GPT-4o [27] (2024)29.33
EfficientFace [45] (2021)85.87
MA-Net [44] (2021)88.42
Li et al. [19] (2021)84.42
Wu et al. [36] (2024)90.83
Ours91.86
", + "bbox": [ + 106, + 227, + 468, + 407 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "additionally leverages the depth modality to model sociodynamic influence, but the mAP is lower than ours. What's more, there are 18 top-two categories (illustrated in red/blue color) of our methods, whereas [24] has 12 top-two categories. (2) For commonly used categories (anger, fear, happiness, sadness, surprise), [24] reports an averaged accuracy of $30.70\\%$ . In contrast, our method achieves a higher mean result of $37.19\\%$ . (3) In Tab. 2, we demonstrate that our method outperforms the state-of-the-art method [36] in both C-F1 and O-F1 scores. (4) Our approach also outperforms VLM methods (BLIP2, LLaVa, GPT-4o). This can be attributed to the lack of alignment of multi-level features in VLM for emotion recognition, suggesting that the alignment of tokenized image and text still produce confused results. Please refer to Appendix (Sec.10.1) for more detailed analysis of multi-level prediction confusion in VLM.", + "bbox": [ + 89, + 446, + 482, + 686 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "CAER-S test Split. Our method is further evaluated on the extensive CAER-S dataset. Unlike EMOTIC, CAER-S encompasses a more intricate background. Thus the results of VLM-related methods are significantly inferior compared to those of other emotion recognition baselines. The accuracy results are presented in Tab. 3 and compared with prominent and recently-published works, demonstrating the exceptional performance of our proposed method. Notably, our approach shows improvement compared to [19, 36], which involves the extraction of hierarchical feature from body and background regions, as well as the exploration of interdependencies across emotion labels using word embeddings or graph embeddings. This comparison highlights the efficacy of our proposed hierarchical feature fusion method.", + "bbox": [ + 89, + 688, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/bc60af2a744247e7743ac10a29459d2eb24cdb17611f3b2de9af8567d91db209.jpg", + "table_caption": [ + "Table 3. Performance comparisons on CAER-S test split." + ], + "table_footnote": [], + "table_body": "
DatasetHFE TypeV-I TypeWorkspacemAP
EMOTICResNet-50Prototype✓36.71
Cross-Level Attn.CLIP✓36.12
Cross-Level Attn.PrototypeX35.73
Cross-Level Attn.Prototype✓37.32
DatasetHFE TypeV-I TypeWorkspaceAcc.
CAER-SResNet-50Prototype✓90.16
Cross-Level Attn.CLIP✓88.28
Cross-Level Attn.PrototypeX87.60
Cross-Level Attn.Prototype✓91.86
", + "bbox": [ + 516, + 89, + 903, + 268 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. Ablation Study on EMOTIC and CAER-S datasets.", + "bbox": [ + 527, + 279, + 883, + 292 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 511, + 304, + 666, + 320 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We conduct comprehensive ablation experiments by different module combinations in our method:", + "bbox": [ + 511, + 325, + 903, + 354 + ], + "page_idx": 6 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- For hierarchical feature extraction (HFE) module, we compare the proposed method with vanilla ResNet-50.", + "- In vocabulary-informed (V-I) module, we substitute pseudo label from prototype-aligned projections (11) with CLIP predictions.", + "- In terms of the workspace, we compare with feature concatenation $\\mathbf{z}_{\\mathbf{G}}$ without feature refinement, denoted as $X$ . First, in Tab. 4, we observe that performance decreases significantly on both datasets when the workspace is degraded. This indicates that the workspace module, along with the competition and coordination mechanism, plays a vital role in coordinating multi-level image representation, thus benefiting the emotion recognition task. Secondly, we show that prototype learning is more important in the complex-environment dataset CAER-S, as it exhibits a greater performance decrease than EMOTIC. Finally, by replacing cross-level attention with ResNet-50, we observe lower performance in both datasets, indicating that modeling multi-level correlation facilitates effective emotion inference." + ], + "bbox": [ + 511, + 357, + 903, + 643 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Qualitative Analysis", + "text_level": 1, + "bbox": [ + 511, + 652, + 705, + 667 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Case Study. In Fig. 4 we demonstrate the effectiveness of our proposed method via two randomly-chosen results of EMOTIC dataset. The left two columns contain and ground truth (GT) label and our prediction results. The prediction successfully matches GT, regardless of different and inconsistent recognition results of head, body and context window. Our approach further delineates the elimination in step two (see Fig 3) of competition process based on value of $\\mathcal{L}_{\\mathrm{hier}}$ of maximum loss. It illustrates that with the guidance of pseudo label from V-I module, the inconsistent recognition results are gradually eliminated. We use orange color to high-light the labels in multi-level results that are not shown in V-I and prediction column. For example, from the top row of Fig. 4, the head level, which is the most inconsistent level of recognition results, is initially filtered out, and", + "bbox": [ + 509, + 674, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "29597", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/87de515feaea1f2141585b08b14d76b581601b0aabfb75e90f349e08af33e456.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 148, + 89, + 318, + 234 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a77b652b353d3a71efa4a7f75669096296a80baa972a737b4e60913bf1bd7933.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 323, + 90, + 496, + 237 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/476deb5786f21a2fc3960a5ec981bdf487ed7a5e2ec8c2c645862c668a7fb24b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 90, + 671, + 237 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/43d77da0600a0cfe4246b8921a7a59d3d0f6668f77746586f1cd1e9af4d68d36.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 676, + 90, + 849, + 237 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/260951db3fa588ad13282164965402d4e345823e2f4b3b3716ba704548cf1073.jpg", + "image_caption": [ + "Figure 5. t-SNE and saliency maps visualization of feature refinement on CAER-S. The red star refers to a random selected sample feature that has been clustered in a more compact group in t-SNE visualization, and displaying more semantic regions on saliency maps." + ], + "image_footnote": [], + "bbox": [ + 148, + 231, + 320, + 359 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/65f8498f799a40f870622ec680b5325819540b4a4824000f062915bc0c97b5e3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 323, + 238, + 496, + 378 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4656457d35e05ba8aa9d14aa754cfdc808a7589b7ec9533aa9f7f5ca7fba36d2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 238, + 671, + 378 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e775eed61d8e7b74624bbb0e896e3b955f3f622643310421090cb0b4c7e3d70e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 676, + 238, + 849, + 377 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "then prediction from the context level is subsequently eliminated. This supports our hypothesis that if the inconsistent multi-level emotion recognition is excluded, the final prediction performance is enhanced.", + "bbox": [ + 88, + 430, + 482, + 491 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "How is multi-level feature refined? To understand the feature refinement in workspace, we employ t-SNE on CAERS to visualize multi-level feature, with 250 randomly sampled testing images per class. As shown in Fig. 5, features $\\mathbf{z}_m$ from different level $m$ are refined after inference-stage parameter update of $\\theta$ , displaying more clustered groups. Consequently, concatenated feature $[\\hat{\\mathbf{z}}_{\\mathbf{c}}; \\hat{\\mathbf{z}}_{\\mathbf{b}}; \\hat{\\mathbf{z}}_{\\mathbf{h}}]$ encompasses compact clusters of in-group representation. We also provide a clustered sample point (shown in red star in Fig. 5) by Grad-CAM [31], a wildly used saliency maps visualization technique of highlighting the semantic pixels. We can observe that with the parameter update of $\\theta$ , body-level and head-level saliency maps align with hand gestures and facial expressions, which contain more semantic information for emotion recognition.", + "bbox": [ + 88, + 491, + 482, + 717 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Is the proposed framework potential to enhance VLM? Since the performance of VLM is worse than expected, we aim to explore whether we can use our proposed framework to improve emotion recognition of VLM. First, based on step two of the competition process, we can obtain the first excluded image level, which we denote as the most inconsistent level $\\kappa$ . Next, using this level $\\kappa$ , we modify the input of the VLM in two ways:", + "bbox": [ + 88, + 718, + 482, + 838 + ], + "page_idx": 7 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Prompt $\\kappa$ : we add the prompt with: \"Please avoid potentially inconsistent recognition results from $\\{\\kappa\\}$ .\"", + "- Multi-Level $^{\\kappa}$ : we use multi-level image inputs but eliminate the level $\\kappa$ ." + ], + "bbox": [ + 89, + 839, + 482, + 898 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/41ceb0d718ca5d0db0cea3b6ad9822e666d76dcca235506b951ffc0fbe3d19af.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodBLIP2 [17]LLaVa [20]GPT-4o [27]
PromptK26.77+0.8128.56+0.5725.11+1.31
Multi-LevelK27.14+1.1829.03+1.0426.64+2.84
", + "bbox": [ + 516, + 428, + 903, + 489 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. We report the mAP of enhanced VLM results on EMOTIC dataset. $\\kappa$ refers to modified input via our proposed framework.", + "bbox": [ + 511, + 500, + 906, + 529 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As shown in Tab. 5, both of them successfully enhance emotion recognition capabilities of VLM, especially for GPT-40. Thus, we validate the inconsistency of multi-level inputs also exists in VLM and we improve it using our framework. Please see Appendix (Sec.10.2) for detailed cases.", + "bbox": [ + 511, + 542, + 906, + 619 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 633, + 633, + 648 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we propose a competition and coordination mechanism to avoid hierarchical visual feature inconsistency in emotion recognition. To guide the alignment of visual features, a vocabulary-informed module is applied to produce labels that refine hierarchical visual representation. We elaborate on the competition process to eliminate irrelevant image-level representations and the coordination process to enhance representations across all image levels. Extensive experiments are conducted on two popular benchmarks, the EMOTIC and CAER-S datasets. Our method not only achieves state-of-the-art performance with better interpretability but also improves large vision-language models.", + "bbox": [ + 511, + 657, + 906, + 840 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments. This work was supported by Shanghai Science and Technology Committee under Grant No. 24511103900, 24511103200, and National Key RD Program of China under grant No. 2022YFC3601405.", + "bbox": [ + 511, + 840, + 903, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "29598", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Bernard J Baars. A cognitive theory of consciousness. Cambridge University Press, 1993. 4", + "[2] Lenore Blum and Manuel Blum. A theory of consciousness from a theoretical computer science perspective: Insights from the conscious Turing machine. Proceedings of the National Academy of Sciences, 119(21):e2115934119, 2022. 4", + "[3] Jean-Benoit Delbrouck, Noé Tits, Mathilde Brousmiche, and Stéphane Dupont. A transformer-based joint-encoding for emotion recognition and sentiment analysis. arXiv preprint arXiv:2006.15955, 2020. 2", + "[4] Jiankang Deng, Jia Guo, Xue Niannan, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In CVPR, 2019. 3", + "[5] Jiayuan Dong, Alex Santiago-Anaya, and Myounghoon Jeon. Facial expressions increase emotion recognition clarity and improve warmth and attractiveness on a humanoid robot without adding the uncanny valley. Proceedings of the Human Factors and Ergonomics Society Annual Meeting, 67 (1):933-939, 2023. 1", + "[6] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 3", + "[7] Bernd Dudzik, Joost Broekens, Mark Neerincx, and Hayley Hung. Exploring personal memories and video content as context for facial behavior in predictions of video-induced emotions. In Proceedings of the 2020 International Conference on Multimodal Interaction, pages 153-162, 2020. 2", + "[8] Yanwei Fu, Xiaomei Wang, Hanze Dong, Yu-Gang Jiang, Meng Wang, Xiangyang Xue, and Leonid Sigal. Vocabulary-informed zero-shot and open-set learning. IEEE Transactions on Pattern Analysis and Machine Intelligence, 42(12): 3136-3152, 2020. 4", + "[9] Manh-Hung Hoang, Soo-Hyung Kim, Hyung-Jeong Yang, and Guee-Sang Lee. Context-aware emotion recognition based on visual relationship detection. IEEE Access, 9: 90465-90474, 2021. 5, 6", + "[10] Jian Huang, Jianhua Tao, Bin Liu, Zheng Lian, and Mingyue Niu. Multimodal transformer fusion for continuous emotion recognition. In ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 3507-3511. IEEE, 2020. 2", + "[11] Oleg Kalyta. Information technology of facial emotion recognition for visual safety surveillance. 2022. 1", + "[12] Ronak Kosti, Jose M Alvarez, Adria Recasens, and Agata Lapedriza. Emotic: Emotions in context dataset. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 61-69, 2017. 2, 5", + "[13] Ronak Kosti, Jose M Alvarez, Adria Recasens, and Agata Lapedriza. Emotion recognition in context. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1667-1675, 2017. 1", + "[14] Ronak Kosti, Jose M Alvarez, Adria Recasens, and Agata Lapedriza. Context based emotion recognition using emotic" + ], + "bbox": [ + 93, + 114, + 482, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "dataset. IEEE transactions on pattern analysis and machine intelligence, 42(11):2755-2766, 2019. 1, 2, 7", + "[15] Jiyoung Lee, Seungryong Kim, Sunok Kim, Jungin Park, and Kwanghoon Sohn. Context-aware emotion recognition networks. In 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 10142-10151, 2019. 1, 5", + "[16] Li et al. Decoupled multimodal distilling for emotion recognition. In CVPR, 2023. 4", + "[17] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, 2022. 2, 6, 7, 8", + "[18] Qingbiao Li, Chunhua Wu, Zhe Wang, and Kangfeng Zheng. Hierarchical transformer network for utterance-level emotion recognition. Applied Sciences, 10(13):4447, 2020. 2", + "[19] Weixin Li, Xuan Dong, and Yunhong Wang. Human emotion recognition with relational region-level analysis. IEEE Transactions on Affective Computing, 2021. 6, 7", + "[20] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, 2023. 2, 6, 7, 8", + "[21] Xin Liu, Henglin Shi, Haoyu Chen, Zitong Yu, Xiaobai Li, and Guoying Zhao. imigue: An identity-free video dataset for micro-gesture understanding and emotion analysis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10631-10642, 2021. 2", + "[22] Fengmao Lv, Xiang Chen, Yanyong Huang, Lixin Duan, and Guosheng Lin. Progressive modality reinforcement for human multimodal emotion recognition from unaligned multimodal sequences. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2554-2562, 2021. 2", + "[23] Trisha Mittal, Uttaran Bhattacharya, Rohan Chandra, Aniket Bera, and Dinesh Manocha. M3er: Multiplicative multimodal emotion recognition using facial, textual, and speech cues. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1359-1367, 2020. 2", + "[24] Trisha Mittal, Pooja Guhan, Uttaran Bhattacharya, Rohan Chandra, Aniket Bera, and Dinesh Manocha. Emoticon: Context-aware multimodal emotion recognition using frege's principle. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14234-14243, 2020. 1, 2, 5, 6, 7", + "[25] Trisha Mittal, Aniket Bera, and Dinesh Manocha. Multimodal and context-aware emotion perception model with multiplicative fusion. IEEE MultiMedia, 28(2):67-75, 2021. 2", + "[26] Trisha Mittal, Puneet Mathur, Aniket Bera, and Dinesh Manocha. Affect2mm: Affective analysis of multimedia content using emotion causality. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5661-5671, 2021. 2", + "[27] OpenAI. GPT-4o API, 2024. Available at: https://openai.com/index/hello-gpt-4o, last accessed on 2024-11-01. 1, 6, 7, 8", + "[28] Kunyu Peng, Alina Roitberg, David Schneider, Marios Koulakis, Kailun Yang, and Rainer Stiefelhagen. Affect-dml: Context-aware one-shot recognition of human affect" + ], + "bbox": [ + 516, + 93, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "29599", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "using deep metric learning. In 2021 16th IEEE International Conference on Automatic Face and Gesture Recognition (FG 2021), pages 1-8, 2021. 2", + "[29] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision, 2021. 4, 7", + "[30] Shulan Ruan, Kun Zhang, Yijun Wang, Hanqing Tao, Weidong He, Guangyi Lv, and Enhong Chen. Context-aware generation-based net for multi-label visual emotion recognition. In 2020 IEEE International Conference on Multimedia and Expo (ICME), pages 1-6. IEEE Computer Society, 2020. 2", + "[31] Ramprasaath R. Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In 2017 IEEE International Conference on Computer Vision (ICCV), pages 618-626, 2017. 8", + "[32] Jiahui She, Yibo Hu, Hailin Shi, Jun Wang, Qiu Shen, and Tao Mei. Dive into ambiguity: latent distribution mining and pairwise uncertainty estimation for facial expression recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6248-6257, 2021. 2", + "[33] Xuli Shen, Xiaomei Wang, Qing Xu, Weifeng Ge, and Xi-angyang Xue. Towards scalable and fast distributionally robust optimization for data-driven deep learning. In 2022 IEEE International Conference on Data Mining (ICDM), pages 448–457, 2022. 5", + "[34] Shamane Siriwardhana, Tharindu Kaluarachchi, Mark Billinghurst, and Suranga Nanayakkara. Multimodal emotion recognition with transformer-based self supervised feature fusion. IEEE Access, 8:176274-176285, 2020. 2", + "[35] Zili Wang, Lingjie Lao, Xiaoya Zhang, Yong Li, Tong Zhang, and Zhen Cui. Context-dependent emotion recognition. Journal of Visual Communication and Image Representation, 89:103679, 2022. 2", + "[36] Shichao Wu, Lei Zhou, Zhengxi Hu, and Jingtai Liu. Hierarchical context-based emotion recognition with scene graphs. IEEE Transactions on Neural Networks and Learning Systems, 35(3):3725-3739, 2024. 1, 2, 4, 6, 7", + "[37] Dingkang Yang, Shuai Huang, Haopeng Kuang, Yangtao Du, and Lihua Zhang. Disentangled representation learning for multimodal emotion recognition. In Proceedings of the 30th ACM International Conference on Multimedia, pages 1642-1651, New York, NY, USA, 2022. Association for Computing Machinery. 1, 2, 4", + "[38] Dingkang Yang, Zhaoyu Chen, Yuzheng Wang, Shunli Wang, Mingcheng Li, Siao Liu, Xiao Zhao, Shuai Huang, Zhiyan Dong, Peng Zhai, and Lihua Zhang. Context de-confounded emotion recognition. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2023. 1, 2", + "[39] Dingkang Yang, Mingcheng Li, Dongling Xiao, Yang Liu, Kun Yang, Zhaoyu Chen, Yuzheng Wang, Peng Zhai, Ke Li," + ], + "bbox": [ + 91, + 90, + 483, + 902 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "and Lihua Zhang. Towards multimodal sentiment analysis debiasing via bias purification, 2024. 2", + "[40] Dingkang Yang, Kun Yang, Mingcheng Li, Shunli Wang, Shuaibing Wang, and Lihua Zhang. Robust emotion recognition in context debiasing. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2024. 1, 2", + "[41] Jingyuan Yang, Jie Li, Leida Li, Xiumei Wang, and Xinbo Gao. A circular-structured representation for visual emotion distribution learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4237–4246, 2021. 1, 2", + "[42] Minghui Zhang, Yumeng Liang, and Huadong Ma. Context-aware affective graph reasoning for emotion recognition. In 2019 IEEE International Conference on Multimedia and Expo (ICME), pages 151-156. IEEE, 2019. 1", + "[43] Yang Zhao, Hao Zhang, and Xiuyuan Hu. Penalizing gradient norm for efficiently improving generalization in deep learning. In Proceedings of the 39th International Conference on Machine Learning, pages 26982-26992. PMLR, 2022. 5", + "[44] Zengqun Zhao, Qingshan Liu, and Shanmin Wang. Learning deep global multi-scale and local attention features for facial expression recognition in the wild. IEEE Transactions on Image Processing, 30:6544-6556, 2021. 7", + "[45] Zengqun Zhao, Qingshan Liu, and Feng Zhou. Robust lightweight facial expression recognition network with label distribution training. In Proceedings of the AAAI conference on artificial intelligence, pages 3510-3519, 2021. 7" + ], + "bbox": [ + 516, + 90, + 903, + 502 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "29600", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/CocoER_ Aligning Multi-Level Feature by Competition and Coordination for Emotion Recognition/91abd7f6-c62e-4947-ae7d-c03321b50ab1_model.json b/2025/CocoER_ Aligning Multi-Level Feature by Competition and Coordination for Emotion Recognition/91abd7f6-c62e-4947-ae7d-c03321b50ab1_model.json new file mode 100644 index 0000000000000000000000000000000000000000..7cdef68abfee9210504bee34195f6bbf92c73ef6 --- /dev/null +++ b/2025/CocoER_ Aligning Multi-Level Feature by Competition and Coordination for Emotion Recognition/91abd7f6-c62e-4947-ae7d-c03321b50ab1_model.json @@ -0,0 +1,2244 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.111, + 0.13, + 0.887, + 0.177 + ], + "angle": 0, + "content": "CocoER: Aligning Multi-Level Feature by Competition and Coordination for Emotion Recognition" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.202, + 0.897, + 0.24 + ], + "angle": 0, + "content": "Xuli Shen\\(^{1,2}\\) Hua Cai\\(^{2}\\) Weilin Shen\\(^{2}\\) Qing Xu\\(^{2}\\) Dingding Yu\\(^{2}\\) Weifeng Ge\\(^{1\\star}\\) Xiangyang Xue\\(^{1\\star}\\) Fudan University \\(^{2}\\)UniDT" + }, + { + "type": "text", + "bbox": [ + 0.107, + 0.242, + 0.882, + 0.259 + ], + "angle": 0, + "content": "{x1shen20,wfge,xyxue}@fudan.edu.cn， {hua.cai,weilin.shen,qing.xu,dingding.yu}@unidt.com" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.292, + 0.327, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.324, + 0.486, + 0.687 + ], + "angle": 0, + "content": "With the explosion of human-machine interaction, emotion recognition has reignited attention. Previous works focus on improving visual feature fusion and reasoning from multiple image levels. Although it is non-trivial to deduce a person's emotion by integrating multi-level feature (head, body and context), the emotion recognition results of each level is usually different from one another, which creates inconsistency in the prevailing feature alignment method and decrease recognition performance. In this work, we propose a multi-level image feature refinement method for emotion recognition (CocoER) to mitigate the impact caused by conflicting results from multi-level recognition. First, we leverage cross-level attention to improve visual feature consistency between hierarchically cropped head, body and context windows. Then, vocabulary informed alignment is incorporated into the recognition framework to produce pseudo label and guide hierarchical visual feature refinement. To effectively fuse multi-level feature, we elaborate on a competition process of eliminating irrelevant image level predictions and a coordination process to enhance the feature across all levels. Extensive experiments are executed on two popular datasets, and our method achieves state-of-the-art performance with multi-level interpretation results. Code is available at: https://github.com/bisno/CocoER." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.716, + 0.222, + 0.731 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.741, + 0.484, + 0.879 + ], + "angle": 0, + "content": "Visual emotion recognition plays a vital role in human computer interaction [27], surveillance [11], and robotics [5]. It aims to perceive the emotion state of human, and machine can modify responses and behaviors by considering the sentiment attitudes. Recently, impressive progress in recognizing emotions is achieved with expression from additional modality, such as head, body, background, depth, and so on [24]. Besides, emotion representation can be enhanced in view of discrete categories [24], points in a continuous" + }, + { + "type": "image", + "bbox": [ + 0.53, + 0.291, + 0.891, + 0.521 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.534, + 0.908, + 0.591 + ], + "angle": 0, + "content": "Figure 1. To avoid conflicts arising from inconsistent multi-level predictions, competition and coordination are applied in refining multi-level visual representation for emotion recognition, guided by pseudo label from vocabulary-informed alignment." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.609, + 0.795, + 0.624 + ], + "angle": 0, + "content": "space [13, 42] and label distributions [41]." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.627, + 0.907, + 0.839 + ], + "angle": 0, + "content": "As emotion is naturally coupled in human and environments, the complex background in image still hinders the research of emotion recognition. Pioneering researches focus on perceiving the emotion states of the specific individuals in the given image with hierarchical visual feature [24, 37, 40] on context-aware emotion recognition datasets [13, 15]. [14] trains different backbones to combine the information of the target body and the entire image. Subsequently, a context-aware emotion recognition architecture enhances the classified precision to a high level representation [24]. In addition to image feature, knowledge graph is incorporated to reason the emotion states in [36]. To align feature across different modalities, [37] aims to address distribution gap between heterogeneous modalities." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Despite previous progress in decoupling irrelevant contextual information in the inference stage [38, 40], there still lacks of multi-level image feature alignment method to precisely recognize human emotions. Each level has dif" + }, + { + "type": "page_footnote", + "bbox": [ + 0.115, + 0.888, + 0.248, + 0.901 + ], + "angle": 0, + "content": "* Corresponding authors." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.958 + ], + "angle": 0, + "content": "29591" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.288 + ], + "angle": 0, + "content": "ferent and potentially misleading representation and predictions, which decreases accuracy of emotion recognition. As shown in Fig. 1, multi-level emotion output from the head (fear and doubt), and context (sad) confuse final emotion prediction. We hypothesize that conflicting results from each level interfere with emotional decision-making process. To better align multi-level feature, researchers often draws upon prior knowledge, such as knowledge graph, to select relevant object. Although knowledge graph [36] is applied to enhance interpretability, extracting object relevance does not directly address the challenge of conflicted multi-level predictions and there is a lack of a guidance that can combine head, body, and background feature." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.29, + 0.483, + 0.56 + ], + "angle": 0, + "content": "In this work, we propose competition and coordination mechanism to effectively perform multi-level fusion of head, body, and context emotion representation. Specifically, we apply cross-level attention mechanism with linear mapping on extracted feature maps to extract correlation from multi-level windows and obtain multi-level embeddings and recognition results. Furthermore, we leverage prototype learning to enhance the recognition result from external image-text aligned model, leading to pseudo label as guidance for multi-level feature refinement. Next, based on the pseudo label, we introduce competition process as an emotional feature selection method that filters out inconsistent multi-level feature. We leverage coordination process to update the parameters of the linear mapping for all image levels in inference-stage. At last, we re-send the multi-level feature to the updated linear mapping and achieve emotional consistency for final emotion recognition. In summary, the contributions of our paper are:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.563, + 0.483, + 0.623 + ], + "angle": 0, + "content": "- We enhance multi-level image correlation via hierarchical feature extraction module, and produce multi-level recognition. We leverage prediction from vocabulary-informed module as pseudo label to identify inconsistent levels." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.623, + 0.483, + 0.667 + ], + "angle": 0, + "content": "- We introduce workspace module that incorporates competition and coordination process to rectify multi-level feature through updated model in inference-stage." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.668, + 0.483, + 0.729 + ], + "angle": 0, + "content": "- Extensive experiments have been conducted to compare with baseline methods and large vision-language models (VLM), and the outcomes demonstrate superior performance and interpretability of our proposed method." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.563, + 0.483, + 0.729 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.746, + 0.233, + 0.761 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.772, + 0.342, + 0.788 + ], + "angle": 0, + "content": "2.1. Visual Emotion Recognition" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.482, + 0.901 + ], + "angle": 0, + "content": "Visual emotion recognition is a comprehensive task in computer vision area, which is deeply associated with cognitive intelligence, and has attracted increasing attentions from researchers. Extensive works utilize facial expression to deduce the emotion state of specific person [7, 23, 25, 32, 35, 38]. Considering that recognizing a person's emotions can be subjective and has raised inherent ambiguity among sim" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.273 + ], + "angle": 0, + "content": "ilar emotion categories, some researches focus on the latent distribution on the candidate emotion label sets [32, 41] and incorporates models from cognitive computing into the recognition algorithms [7, 21, 24, 26, 41]. Regarding network structures, [3, 18, 39] present transformer-based architectures to capture contextual embeddings. [10, 34, 37] utilize multi-head attention to fuse multiple modalities. In the aspect of inference stage, [38, 40] use causal relationships to eliminate information in the background that is inconsistent with target emotions. In contrast, we focuses on emotion recognition via multi-level visual representation that aligned by vocabulary-informed pseudo label." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.282, + 0.816, + 0.298 + ], + "angle": 0, + "content": "2.2. Multi-modal Emotion Recognition" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.303, + 0.905, + 0.499 + ], + "angle": 0, + "content": "[22, 23, 26, 28, 40] enhance multi-modal fusion performance, and [12, 14, 24, 30] explore context-based visual expressions. To fully take advantages of various context information, multi-modal visual feature are incorporated into model inputs, such as face, pose and depth [24, 40]. Different modality information requires different sensors or off-the-shelf extractor, resulting in increasing model complexity and a narrowed application range. On the other hand, [14, 30] focus on addressing this problem with only global and body information, and are still able to achieve competitive performance. [36] introduces a hierarchical context-based emotion recognition method that uses scene graphs to infer emotions from images." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.5, + 0.905, + 0.651 + ], + "angle": 0, + "content": "In the field of large vision language model [17, 20], despite competitive performance in many multi-modal tasks, existing models do not prioritize emotion recognition. The modality alignment methods for large vision language models mostly focus on tokenized pair in the whole image, neglecting the multi-level image feature that misleads the image-text alignment. In this paper, we target this problem by eliminating irrelevant level of feature and further propose a novel multi-level feature alignment framework to enhance the performance of emotion recognition." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.664, + 0.612, + 0.679 + ], + "angle": 0, + "content": "3. Methods" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.905, + 0.826 + ], + "angle": 0, + "content": "In this section, we first introduce the hierarchical feature extraction module on the cropped multi-level image, where each image level has its own emotion recognition outcome. Next, we elaborate on the vocabulary-informed alignment module, which serves as guidance that provides pseudo label for eliminating inconsistent multi-level predictions. Finally, based on such pseudo label, we apply the workspace module to enhance multi-level feature fusion. The proposed framework is depicted in Fig. 2." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.834, + 0.797, + 0.849 + ], + "angle": 0, + "content": "3.1. Hierarchical Feature Extraction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.905, + 0.901 + ], + "angle": 0, + "content": "We construct hierarchical emotion recognition by dividing the input image into three windows: head, body, and context. Specifically, given an input image \\(\\mathbf{I} \\in \\mathbb{R}^{H \\times W \\times 3}\\), we" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "29592" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.159, + 0.09, + 0.52, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.524, + 0.091, + 0.844, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.158, + 0.295, + 0.836, + 0.497 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.349, + 0.511, + 0.645, + 0.526 + ], + "angle": 0, + "content": "Figure 2. Illustration of the proposed framework." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.539, + 0.484, + 0.614 + ], + "angle": 0, + "content": "leverage object detection model [4] to obtain head, body, and context windows using detected bounding boxes. All three cropped windows are resized to \\((H,W)\\), and are then fed into trainable backbones \\(\\mathcal{R}\\) to extract multi-level feature maps \\(\\mathbf{u}_m\\) with a size of \\((h,w)\\) and \\(c\\) channels:" + }, + { + "type": "equation", + "bbox": [ + 0.18, + 0.622, + 0.482, + 0.639 + ], + "angle": 0, + "content": "\\[\n\\mathbf {u} _ {m} = \\mathcal {R} _ {m} (\\mathbf {I} _ {m}), m \\in \\{\\mathbf {c}, \\mathbf {b}, \\mathbf {h} \\}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.646, + 0.484, + 0.831 + ], + "angle": 0, + "content": "where \\(\\mathbf{c},\\mathbf{b},\\mathbf{h}\\) are context, body, and head windows, respectively. Each \\(\\mathbf{u}_m\\in \\mathbb{R}^{h\\times w\\times c}\\) is divided and reshaped into \\(N\\) non-overlapping patches \\(\\mathbf{u}_m^p\\in \\mathbb{R}^{p^2\\cdot c}\\) and then mapped into \\(d_h\\) dimensions with linear projection \\(\\mathbf{E}_m\\in \\mathbb{R}^{(p^2\\cdot c)\\times d_h}\\). Thus each level \\(m\\) consists of a sequence of 2D embeddings \\(\\bar{\\mathbf{u}}_m = [\\mathbf{u}_m^{p,0}\\mathbf{E}_m;\\mathbf{u}_m^{p,1}\\mathbf{E}_m;\\dots ;\\mathbf{u}_m^{p,N - 1}\\mathbf{E}_m]\\in \\mathbb{R}^{N\\times d_h}\\), where \\(N = hw / p^2\\) is patch number [6]. \\(\\bar{\\mathbf{u}}_m\\) is sent to cross-level attention \\(\\mathcal{T}\\) for enhancing multi-level correlation via connecting local and context feature. For instance, the head-context attention \\(\\mathcal{T}_{\\mathbf{h},\\mathbf{c}}\\) is formalized with linear projections \\(\\{W_{\\mathbf{h}}^{0},W_{\\mathbf{h}}^{1},W_{\\mathbf{c}}\\} \\in \\mathbb{R}^{d_h\\times d_h}\\) with feature maps \\(\\{\\bar{\\mathbf{u}}_{\\mathbf{h}},\\bar{\\mathbf{u}}_{\\mathbf{c}}\\}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.156, + 0.837, + 0.483, + 0.855 + ], + "angle": 0, + "content": "\\[\n\\mathbf {Q} = \\bar {\\mathbf {u}} _ {\\mathbf {h}} W _ {\\mathbf {h}} ^ {0}, \\mathbf {K} = \\bar {\\mathbf {u}} _ {\\mathbf {c}} W _ {\\mathbf {c}}, \\mathbf {V} = \\bar {\\mathbf {u}} _ {\\mathbf {h}} W _ {\\mathbf {h}} ^ {1} \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.158, + 0.858, + 0.483, + 0.877 + ], + "angle": 0, + "content": "\\[\n\\gamma = \\operatorname {s o f t m a x} \\left(\\mathbf {Q} \\mathbf {K} ^ {T}\\right) / \\sqrt {d _ {h}} \\quad \\in \\mathbb {R} ^ {N \\times N} \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.153, + 0.879, + 0.483, + 0.897 + ], + "angle": 0, + "content": "\\[\n\\bar {\\mathbf {z}} _ {\\mathbf {h}} = \\boldsymbol {\\gamma} \\mathbf {V} \\in \\mathbb {R} ^ {N \\times d _ {h}}. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.539, + 0.907, + 0.569 + ], + "angle": 0, + "content": "Then \\(\\bar{\\mathbf{z}}_{\\mathbf{h}}\\) is added with \\(\\bar{\\mathbf{u}}_{\\mathbf{h}}\\) and sent to typical fully-connected layers with skip-connection:" + }, + { + "type": "equation", + "bbox": [ + 0.596, + 0.58, + 0.905, + 0.598 + ], + "angle": 0, + "content": "\\[\n\\dot {\\mathbf {z}} _ {\\mathbf {h}} = \\bar {\\mathbf {z}} _ {\\mathbf {h}} + \\bar {\\mathbf {u}} _ {\\mathbf {h}} \\in \\mathbb {R} ^ {N \\times d _ {h}} \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.596, + 0.602, + 0.905, + 0.62 + ], + "angle": 0, + "content": "\\[\n\\mathbf {z} _ {\\mathbf {h}} = \\operatorname {F F N} _ {0} (\\operatorname {L N} (\\dot {\\mathbf {z}} _ {\\mathbf {h}})) + \\dot {\\mathbf {z}} _ {\\mathbf {h}} \\in \\mathbb {R} ^ {d}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.632, + 0.907, + 0.799 + ], + "angle": 0, + "content": "where LN is LayerNorm, \\(\\mathrm{FFN}_0\\) refers to the feed-forward network and \\(\\mathbf{z_h} \\in \\mathbb{R}^d\\) is the head-level representation. Thus we bridge hierarchical information to enhance head-level representation \\(\\mathbf{z_h}\\) via head-context cross attention \\(\\mathcal{T}_{\\mathbf{h},\\mathbf{c}}\\). Similarly we have \\(\\mathbf{z_b}\\) via body-context cross attention \\(\\mathcal{T}_{\\mathbf{b},\\mathbf{c}}\\). Therefore we leverage attention mechanism to enrich local feature representation \\(\\{\\bar{\\mathbf{u}}_{\\mathbf{h}},\\bar{\\mathbf{u}}_{\\mathbf{c}}\\}\\). Note that we conduct self-attention \\(\\mathcal{T}_{\\mathbf{c},\\mathbf{c}}\\) and obtain context feature \\(\\mathbf{z_c}\\). Next, we pass each level representation \\(\\mathbf{z}_m\\) through a linear mapping \\(g_{\\theta_m}\\), followed by an operation \\(\\mathcal{F}\\) to produce \\(k\\) emotion category corresponding to multi-class task or multi-label task:" + }, + { + "type": "equation", + "bbox": [ + 0.59, + 0.811, + 0.905, + 0.829 + ], + "angle": 0, + "content": "\\[\ny _ {m} = \\mathcal {F} \\left(g _ {\\boldsymbol {\\theta} _ {m}} \\left(\\mathbf {z} _ {m}\\right)\\right), m \\in \\{\\mathbf {c}, \\mathbf {b}, \\mathbf {h} \\}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.841, + 0.907, + 0.902 + ], + "angle": 0, + "content": "where \\( y_{m} \\) represents the prediction of each level \\( m \\). The parameter of linear mapping \\( g_{\\theta_m} \\) will be updated in inference-stage, and each level representation \\( \\mathbf{z}_m \\) will be refined for enhancing emotion recognition." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "29593" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.38, + 0.108 + ], + "angle": 0, + "content": "3.2. Vocabulary-Informed Alignment" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.113, + 0.484, + 0.309 + ], + "angle": 0, + "content": "In order to semantically guide feature refinement from hierarchical feature extraction, we propose a vocabulary-informed alignment method that leverages pre-trained image and text encoder to gain pseudo label and calibrate image representation. Specifically, given same input image \\(\\mathbf{I} \\in \\mathbb{R}^{H \\times W \\times 3}\\), we improve CLIP [29] prediction to guide image feature alignment from text space. As shown in Fig. 2, we first encode the image through an CLIP image encoder \\(\\mathbf{Enc}_{\\mathrm{img}}\\). Then, we provide a text prompt for \\(k\\) classes of emotion recognition task. The \\(k\\) classes are injected and expanded into \\(k\\) sentences of \"This photo shows {} emotion\", denoted as \\(\\mathbf{S}_k\\), and are sent to CLIP text encoder \\(\\mathbf{Enc}_{\\mathrm{text}}\\). Thus, we have" + }, + { + "type": "equation", + "bbox": [ + 0.191, + 0.317, + 0.482, + 0.334 + ], + "angle": 0, + "content": "\\[\n\\mathbf {u} _ {\\mathrm {i m g}} = \\mathbf {E n c} _ {\\mathrm {i m g}} (\\mathbf {I}) \\in \\mathbb {R} ^ {d} \\tag {8}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.19, + 0.337, + 0.482, + 0.356 + ], + "angle": 0, + "content": "\\[\n\\mathbf {U} _ {\\text {t e x t}} = \\mathbf {E n c} _ {\\text {t e x t}} \\left(\\mathbf {S} _ {k}\\right) \\in \\mathbb {R} ^ {d \\times k}. \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.363, + 0.484, + 0.546 + ], + "angle": 0, + "content": "One can directly use dot-product similarity to obtain recognition result of CLIP. However, merely applying CLIP prediction is insufficient because it is trained on open-world image-text pairs, which leads to difficulties in differentiating precise expressions such as \"surprise\" and \"fear\". To improve the recognition performance in vocabulary-informed module, we apply prototype learning [8] that projects image representation into text representation space and pushes closer to emotion prototype. In detail, after extracting image feature, we expand \\(\\mathbf{u}_{\\mathrm{img}}\\) into \\(\\hat{\\mathbf{U}}_{\\mathrm{img}}\\) which has the same shape of \\(\\mathbf{U}_{\\mathrm{text}}\\) and aim to minimize the distance \\(\\mathcal{D}(\\hat{\\mathbf{U}}_{\\mathrm{img}},\\mathbf{U}_{\\mathrm{text}})\\):" + }, + { + "type": "equation", + "bbox": [ + 0.157, + 0.553, + 0.482, + 0.573 + ], + "angle": 0, + "content": "\\[\n\\mathcal {D} \\left(\\hat {\\mathbf {U}} _ {\\mathrm {i m g}}, \\mathbf {U} _ {\\text {t e x t}}\\right) = \\left\\| \\omega^ {T} \\hat {\\mathbf {U}} _ {\\mathrm {i m g}} - \\mathbf {U} _ {\\text {t e x t}} \\right\\| _ {2} ^ {2}, \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.58, + 0.484, + 0.701 + ], + "angle": 0, + "content": "where \\(\\hat{\\mathbf{U}}_{\\mathrm{img}} \\in \\mathbb{R}^{d \\times k}\\), and \\(\\omega \\in \\mathbb{R}^{d \\times d}\\) is a linear projection which is trained by minimizing the above distance of image feature and semantic emotion prototypes. Upon obtaining the trained \\(\\omega\\), we feed the emotion-aligned image feature \\(\\omega^T \\hat{\\mathbf{U}}_{\\mathrm{img}}\\) to \\(\\mathrm{FFN}_1\\). Similar to (7) in the previous module that each hierarchical window has its own recognition result, the vocabulary-informed module also provides emotion recognition result:" + }, + { + "type": "equation", + "bbox": [ + 0.199, + 0.709, + 0.482, + 0.728 + ], + "angle": 0, + "content": "\\[\ny _ {\\mathbf {t}} = \\mathcal {F} \\left(\\operatorname {F F N} _ {1} \\left(\\omega^ {T} \\hat {\\mathbf {U}} _ {\\mathrm {i m g}}\\right)\\right). \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.484, + 0.841 + ], + "angle": 0, + "content": "\\(y_{\\mathbf{t}}\\) serves as pseudo label for further feature refinement in workspace. We use the embeddings of last hidden layer of \\(\\mathrm{FFN}_1\\) as the vocabulary-informed feature \\(\\mathbf{z}_{\\mathbf{t}}\\), which has same feature dimension of image level feature \\(\\mathbf{z} \\in \\mathbb{R}^d\\). Note that we freeze parameters of pre-trained \\(\\omega\\), as well as \\(\\mathbf{Enc}_{\\mathrm{img}}\\) and \\(\\mathbf{Enc}_{\\mathrm{text}}\\) during the training stage of our framework, leaving \\(\\mathrm{FFN}_1\\) trainable." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.849, + 0.214, + 0.865 + ], + "angle": 0, + "content": "3.3. Workspace" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Different from the previous feature refinement methods, such as the entity-context feature alignment [36] or emotion" + }, + { + "type": "image", + "bbox": [ + 0.527, + 0.089, + 0.895, + 0.278 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.291, + 0.906, + 0.32 + ], + "angle": 0, + "content": "Figure 3. Detailed illustration of competition and coordination process, incorporated with inference-stage parameter update." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.329, + 0.906, + 0.434 + ], + "angle": 0, + "content": "feature disentanglement [16, 37], we introduce a workspace module that can adaptively refine all levels of emotion feature based on the vocabulary-informed classification results, as illustrated in the bottom part of Fig. 2. The name \"workspace\" is inspired by the Workspace Theory [1, 2], which describes how multi-level predictions compete and coordinate for final recognition." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.435, + 0.907, + 0.57 + ], + "angle": 0, + "content": "Formally, given the multi-level emotion feature set \\(\\mathbf{Z} = \\{\\mathbf{z}_{\\mathrm{c}},\\mathbf{z}_{\\mathrm{b}},\\mathbf{z}_{\\mathrm{h}}\\}\\), we aim to gradually eliminate inconsistent level of feature in \\(\\mathbf{Z}\\) according to the pseudo label \\(y_{t}\\) (11) until there is only one element in \\(\\mathbf{Z}\\). We refer to this process as multi-level feature competition. Subsequently, we leverage the selection process to update parameters of \\(g_{\\theta_m}\\) and re-send each \\(\\mathbf{z}_m\\) to \\(g_{\\theta_m}\\) to obtain the refined feature \\(\\hat{\\mathbf{z}}_m\\). We call this multi-level feature coordination. Technical details are provided in the next two sections." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.576, + 0.704, + 0.591 + ], + "angle": 0, + "content": "3.3.1. Competition Process" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.595, + 0.906, + 0.67 + ], + "angle": 0, + "content": "The competition process is designed as an iteration and contains two steps during each iteration: 1) Determine the descent direction for updating parameters of \\( g_{\\theta_m} \\); 2) Eliminate inconsistent element in feature set \\( \\mathbf{Z} \\). The iteration ends when there is only one element in \\( \\mathbf{Z} \\), shown in Fig. 3." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.671, + 0.907, + 0.761 + ], + "angle": 0, + "content": "Step 1. We leverage the value of loss function \\(\\mathcal{L}_{\\mathrm{hier}}(\\mathcal{F}(g_{\\theta}(\\mathbf{z})),y_{\\mathbf{t}})\\) as a measure of inconsistency between multi-level recognition and pseudo label. Then, in order to target the most inconsistent level, we take the maximum of the loss function groups for the results from \\(\\mathcal{F}(g_{\\theta_m}(\\mathbf{z}_m))\\) against pseudo label \\(y_{\\mathbf{t}}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.617, + 0.768, + 0.905, + 0.79 + ], + "angle": 0, + "content": "\\[\n\\max _ {\\mathbf {z} _ {m} \\in \\mathbf {Z}} \\left\\{\\mathcal {L} _ {\\text {h i e r}} \\left(g _ {\\boldsymbol {\\theta} _ {m}} \\left(\\mathbf {z} _ {m}\\right), y _ {\\mathbf {t}}\\right) \\right\\}. \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.798, + 0.906, + 0.874 + ], + "angle": 0, + "content": "For the convenience of notation, let \\(\\mathbf{f}_{\\mathbf{z}_m}(\\pmb{\\theta}_m) = \\{\\mathcal{L}_{\\mathrm{hier}}(g_{\\pmb{\\theta}_m}(\\mathbf{z}_m),y_t),\\mathbf{z}_m\\in \\mathbf{Z}\\}\\). During each iteration in the competition process, we aim to calculate the gradient of \\(\\max_{\\mathbf{z}_m\\in \\mathbf{Z}}\\mathbf{f}_{\\mathbf{z}_m}(\\pmb{\\theta}_m)\\) so as to update parameters of \\(g_{\\pmb{\\theta}_m}\\). Thus, our goal transforms into a min-max problem:" + }, + { + "type": "equation", + "bbox": [ + 0.642, + 0.881, + 0.905, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\boldsymbol {\\theta} _ {m}} \\max _ {\\mathbf {z} _ {m} \\in \\mathbf {Z}} \\mathbf {f} _ {\\mathbf {z} _ {m}} \\left(\\boldsymbol {\\theta} _ {m}\\right). \\tag {13}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "29594" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.228 + ], + "angle": 0, + "content": "However, there exists a non-trivial challenge in finding the descent direction when we implement \\(\\max(\\cdot)\\) on the groups of loss functions because the above function is nondifferentiable. Inspired by [33], we bring differentiability to \\(\\max_{\\mathbf{z}_m \\in \\mathbf{Z}} \\mathbf{f}_{\\mathbf{z}_m}(\\theta_m)\\) and leverage a regularization term to stabilize the descent direction. Please refer to technical details in Appendix (Sec.1). Then, the min-max problem is reduced to the following quadratic programming (QP) problem with a Lagrange multiplier \\(\\lambda\\):" + }, + { + "type": "equation", + "bbox": [ + 0.182, + 0.24, + 0.484, + 0.304 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\min _ {\\boldsymbol {\\lambda}} \\left(\\frac {1}{2} \\boldsymbol {\\lambda} ^ {T} \\mathbf {G} \\mathbf {G} ^ {T} \\boldsymbol {\\lambda} - \\boldsymbol {\\lambda} ^ {T} \\mathbf {f} _ {\\mathbf {z} _ {m}}\\right) \\tag {14} \\\\ \\text {s . t .} \\quad \\sum_ {\\mathbf {z} _ {m} \\in \\mathbf {Z}} \\lambda_ {\\mathbf {z} _ {m}} = 1, \\lambda_ {\\mathbf {z} _ {m}} \\geq 0 \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.318, + 0.484, + 0.44 + ], + "angle": 0, + "content": "where \\(\\mathbf{G} = \\nabla_{\\boldsymbol{\\theta}_m^l}\\mathbf{f}_{\\mathbf{z}_m} \\in \\mathbb{R}^{q \\times (d \\cdot k)}\\) contains gradient of the prediction layer \\(\\boldsymbol{\\theta}_m^l\\) with respect to the loss of \\(\\mathbf{f}_{\\mathbf{z}_m}\\) and \\(q = |\\mathbf{Z}|\\) refers to the number of elements left in \\(\\mathbf{Z}\\). After solving for \\(\\boldsymbol{\\lambda}\\) in the QP problem, we have \\(\\delta = -\\mathbf{G}^T\\boldsymbol{\\lambda} \\in \\mathbb{R}^{d \\cdot k}\\) as the approximate solution of (13). Here, \\(\\boldsymbol{\\delta}\\) serves as the shared parameter updating path for linear mapping \\(g_{\\boldsymbol{\\theta}_m}\\) of all three levels, which will be used for feature refinement in coordination process." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.442, + 0.484, + 0.503 + ], + "angle": 0, + "content": "Step 2. Let \\( n = \\arg \\max_{\\mathbf{z}_m \\in \\mathbf{Z}} \\left\\{ \\mathcal{L}_{\\mathrm{hier}}(g_{\\theta_m}(\\mathbf{z}_m), y_t) \\right\\} \\) be the index of most inconsistent feature in \\( \\mathbf{Z} \\). We iteratively eliminate the corresponding \\( n \\) of feature \\( \\mathbf{z} \\) in \\( \\mathbf{Z} \\), denoted as \\( \\mathbf{Z}_{-n} \\). This elimination is illustrated in the top part of Fig. 3." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.512, + 0.288, + 0.525 + ], + "angle": 0, + "content": "3.3.2. Coordination Process" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.531, + 0.484, + 0.669 + ], + "angle": 0, + "content": "The goal of coordination process is to rectify the multi-level feature \\(\\mathbf{z}\\) via updated linear mapping \\(g_{\\theta}\\). Specifically, after iteratively obtaining the respective descent direction \\(\\delta\\) during the competition process, we accumulate the overall descent direction, shown in the bottom part of Fig. 3. For example, if the context-level feature \\(\\mathbf{z}_{\\mathbf{c}}\\) and head-level feature \\(\\mathbf{z_h}\\) is gradually eliminated through competition, we have \\(\\delta_{\\mathbf{c}}\\) and \\(\\delta_{\\mathbf{h}}\\) based on (14), which construct the shared parameter updating path for the prediction layer \\(\\theta_m^l\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.239, + 0.68, + 0.482, + 0.697 + ], + "angle": 0, + "content": "\\[\n\\delta^ {l} = \\delta_ {\\mathrm {c}} + \\delta_ {\\mathrm {h}}. \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.71, + 0.484, + 0.741 + ], + "angle": 0, + "content": "Then, for each image level \\(m\\) we update entire parameter \\(\\theta_{m}\\) via chain rule and back-propagation, denoted as \\(\\longleftarrow\\):" + }, + { + "type": "equation", + "bbox": [ + 0.12, + 0.752, + 0.482, + 0.771 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\theta} _ {\\mathbf {c}} \\leftarrow \\theta_ {\\mathbf {c}} ^ {l} + \\delta^ {l}; \\tilde {\\theta} _ {\\mathbf {b}} \\leftarrow \\theta_ {\\mathbf {b}} ^ {l} + \\delta^ {l}; \\tilde {\\theta} _ {\\mathbf {h}} \\leftarrow \\theta_ {\\mathbf {h}} ^ {l} + \\delta^ {l}. \\tag {16}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.784, + 0.484, + 0.83 + ], + "angle": 0, + "content": "Next, we re-send \\(\\mathbf{z}_m\\) to its corresponding linear mapping \\(g_{\\tilde{\\theta}}\\) and acquire refined feature from last hidden layer and construct multi-level concatenated representation \\([\\hat{\\mathbf{z}}_{\\mathrm{c}};\\hat{\\mathbf{z}}_{\\mathrm{b}};\\hat{\\mathbf{z}}_{\\mathrm{h}}]\\):" + }, + { + "type": "equation", + "bbox": [ + 0.145, + 0.846, + 0.482, + 0.865 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {z}} _ {\\mathrm {c}} = g _ {\\tilde {\\theta} _ {\\mathrm {c}}} \\left(\\mathbf {z} _ {\\mathrm {c}}\\right); \\hat {\\mathbf {z}} _ {\\mathrm {b}} = g _ {\\tilde {\\theta} _ {\\mathrm {b}}} \\left(\\mathbf {z} _ {\\mathrm {b}}\\right); \\hat {\\mathbf {z}} _ {\\mathrm {h}} = g _ {\\tilde {\\theta} _ {\\mathrm {h}}} \\left(\\mathbf {z} _ {\\mathrm {h}}\\right). \\tag {17}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.901 + ], + "angle": 0, + "content": "With such inference-stage parameter updating, we calibrate multi-level emotion representation, ensuring that the final" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.152 + ], + "angle": 0, + "content": "emotion recognition results are less affected by the inconsistent emotion recognition outcomes from hierarchical feature. Please refer to algorithm in Appendix that describes competition and coordination process within workspace." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.161, + 0.704, + 0.178 + ], + "angle": 0, + "content": "3.4. Training Objectives" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.183, + 0.907, + 0.243 + ], + "angle": 0, + "content": "We concatenate the refined multi-level representation with vocabulary-informed feature \\(\\mathbf{z_t}\\) in \\(\\mathbf{z_G} = [\\hat{\\mathbf{z}}_{\\mathrm{c}};\\hat{\\mathbf{z}}_{\\mathrm{b}};\\hat{\\mathbf{z}}_{\\mathrm{h}};\\mathbf{z}_{\\mathrm{t}}]\\). Then \\(\\mathbf{z_G}\\) is passed to feed-forward network, and output of the proposed method is:" + }, + { + "type": "equation", + "bbox": [ + 0.637, + 0.255, + 0.905, + 0.271 + ], + "angle": 0, + "content": "\\[\ny _ {\\mathbf {G}} = \\mathcal {F} \\left(\\operatorname {F F N} _ {2} \\left(\\mathbf {z} _ {\\mathbf {G}}\\right)\\right). \\tag {18}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.281, + 0.906, + 0.327 + ], + "angle": 0, + "content": "We collectively form loss functions from HFE, V-I and workspace module as \\(\\mathcal{L}_{\\mathbf{z}}(y_{\\mathbf{z}},y)\\) with respect to ground-truth \\(y\\). The combined loss function \\(\\mathcal{L}_{\\mathrm{emo}}\\) is defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.594, + 0.338, + 0.905, + 0.354 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {e m o}} = \\mathcal {L} _ {\\mathbf {c}} + \\mathcal {L} _ {\\mathbf {b}} + \\mathcal {L} _ {\\mathbf {h}} + \\mathcal {L} _ {\\mathbf {t}} + \\mathcal {L} _ {\\mathbf {G}}. \\tag {19}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.364, + 0.906, + 0.425 + ], + "angle": 0, + "content": "Next, we add norm of \\(\\delta^l\\) [43] to regularize the shared updating path, aiming to avoid over-dependence on the vocabulary-informed pseudo label. Finally, all parameters of our proposed framework are trained based on:" + }, + { + "type": "equation", + "bbox": [ + 0.639, + 0.435, + 0.905, + 0.453 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {\\mathrm {e m o}} + \\alpha \\| \\boldsymbol {\\delta} ^ {l} \\| _ {2}, \\tag {20}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.463, + 0.844, + 0.479 + ], + "angle": 0, + "content": "where \\(\\alpha\\) is hyperparameter, empirically set to 0.1." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.492, + 0.646, + 0.508 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.516, + 0.596, + 0.532 + ], + "angle": 0, + "content": "4.1. Setup" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.538, + 0.907, + 0.72 + ], + "angle": 0, + "content": "Datasets. EMOTIC [12] comprises 23,571 images featuring 34,320 individuals, predominantly captured in unconstrained environments. Each image encompasses multiple emotions, presenting a more intricate challenge of multi-label emotion recognition across 26 discrete emotion categories. CAER-S [15] is derived by extracting static frames from video clips in the CAER dataset [15], containing over 70,000 images. These images are annotated with an expanded set of 7 emotion categories: anger, disgust, fear, happiness, sadness, surprise, and neutral. Both datasets are divided into training (80%) and testing (20%) subsets. Implementation Details. Please refer to Appendix." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.728, + 0.818, + 0.744 + ], + "angle": 0, + "content": "4.2. Comparison with State-of-the-Arts" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.907, + 0.902 + ], + "angle": 0, + "content": "EMOTIC test Split. We first compare our method with several multimodal or multi-object based emotion recognition competitors on EMOTIC test split. As shown in Tab. 1, we report the average precision (AP) over all categories as well as the average results (mAP). From the table, we can observe that: (1) Our method achieves state-of-the-art performance with \\(37.32\\%\\) on mAP. Compared with [9] which utilizes Faster R-CNN to detect all possible objects in images for reasoning, we surpass it by a margin of \\(2.16\\%\\) using only three hierarchical cropped windows. Moreover, [24]" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "29595" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.089, + 0.904, + 0.513 + ], + "angle": 0, + "content": "
CategoriesLi et al. [19] Hoang et al. [9] Mittal et al. [24]† Wu et al. [36]BLIP2-6.7b [17] LLaVa-1.6-7b [20] GPT-4o [27]
2021202120202024202220232024
Affection37.9344.4845.2332.2524.2534.3222.53
Anger13.7330.7115.4610.3708.3004.9810.32
Annoyance20.8726.4721.9212.8408.9912.3709.98
Anticipation61.0859.8972.1256.8391.3591.8754.16
Aversion09.6112.4317.8107.7006.4506.4603.51
Confidence80.0879.2468.6577.5159.3762.0568.09
Disapproval21.5424.5419.8212.4707.5610.3808.64
Disconnection28.3234.2443.1226.4125.8531.0517.03
Disquietment22.5718.7324.2316.7513.5314.4517.17
Doubt/Confusion33.5025.4235.1218.6218.5819.0631.20
Embarrassment04.1604.2614.3702.8005.4505.4402.17
Engagement88.1288.7191.1286.8395.1496.1080.77
Esteem20.5017.9923.6215.5023.1923.2718.30
Excitement80.1174.2183.2670.7364.5170.3669.59
Fatigue17.5122.6216.2312.6407.6811.0309.53
Fear15.5613.9223.6505.4807.8408.4912.79
Happiness76.0183.0274.7169.4570.7775.1453.65
Pain14.5616.6813.2110.6405.5205.8107.45
Peace26.7628.9134.2724.6320.3221.2628.91
Pleasure55.6455.4765.5344.3343.3043.4942.86
Sadness30.8042.8723.4121.2107.0618.7210.70
Sensitivity09.5915.8908.3205.8904.7704.7503.96
Suffering30.7046.2326.3921.4805.2206.1709.74
Surprise17.9216.2717.3707.9415.4216.1417.66
Sympathy15.2615.3734.2813.2421.4821.6209.10
Yearning10.1110.0414.2909.1213.0512.7907.93
mAP32.4135.1635.4826.6825.9627.9923.80
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.524, + 0.907, + 0.553 + ], + "angle": 0, + "content": "Table 1. Performance comparisons on EMOTIC test split. The 1st/2nd best results are highlighted in red/blue. †' denotes the method that contains additional visual modalities, such as depth, which is not used in our method." + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.568, + 0.327, + 0.701 + ], + "angle": 0, + "content": null + }, + { + "type": "table", + "bbox": [ + 0.33, + 0.568, + 0.9, + 0.701 + ], + "angle": 0, + "content": "
GTPredictionHeadBodyContextV-I (Pseudo label)
AffectionAffectionDisconnectionAffectionEngagementAffection
AnticipationConfidenceFatigueConfidenceExcitementConfidence
ConfidenceExcitementPleasureHappinessExcitement
EngagementHappinessPeaceHappiness
ExcitementPeacePleasure
HappinessPleasure
PeaceLhier=2.35Lhier=0.61Lhier=1.33
Pleasure
Exclusion Sequence: (1) Head→(2) Context→(3) Body
" + }, + { + "type": "image", + "bbox": [ + 0.098, + 0.709, + 0.327, + 0.842 + ], + "angle": 0, + "content": null + }, + { + "type": "table", + "bbox": [ + 0.33, + 0.709, + 0.9, + 0.842 + ], + "angle": 0, + "content": "
GTPredictionHeadBodyContextV-I (Pseudo label)
AnticipationAnticipationAnticipationAffectionEngagementAnticipation
ConfidenceConfidenceConfidenceSensitivityEmbarrassmentEngagement
EngagementEngagementExcitementSufferingPleasureEsteem
ExcitementExcitementHappinessHappiness
PeacePleasure
Lhier=1.42Lhier=2.36Lhier=1.03
Exclusion Sequence: (1) Body→(2) Context→(3) Head
" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.853, + 0.907, + 0.882 + ], + "angle": 0, + "content": "Figure 4. Case study of randomly sampled results on EMOTIC. The V-I column shows pseudo label in vocabulary-informed module. We display the elimination of inconsistent recognition results in orange, which are not shown in either the V-I column or the prediction column." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.946, + 0.521, + 0.957 + ], + "angle": 0, + "content": "29596" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.1, + 0.089, + 0.477, + 0.188 + ], + "angle": 0, + "content": "
MethodsC-F1 (%)O-F1 (%)
Kosti et al. [14] (2019)29.3354.45
Li et al. [19] (2021)15.1048.07
Wu et al. [36] (2024)34.2454.78
Ours36.7757.23
" + }, + { + "type": "table_caption", + "bbox": [ + 0.092, + 0.198, + 0.483, + 0.226 + ], + "angle": 0, + "content": "Table 2. Label-based F1 score (C-F1) and example-based F1 score (O-F1) on EMOTIC test split." + }, + { + "type": "table", + "bbox": [ + 0.107, + 0.228, + 0.47, + 0.408 + ], + "angle": 0, + "content": "
MethodsAccuracy (%)
CLIP-ViT-L [29] (2021)21.57
BLIP2-6.7b [17] (2022)14.21
LLaVa-1.6-7b [20] (2023)28.59
GPT-4o [27] (2024)29.33
EfficientFace [45] (2021)85.87
MA-Net [44] (2021)88.42
Li et al. [19] (2021)84.42
Wu et al. [36] (2024)90.83
Ours91.86
" + }, + { + "type": "table_caption", + "bbox": [ + 0.115, + 0.419, + 0.455, + 0.433 + ], + "angle": 0, + "content": "Table 3. Performance comparisons on CAER-S test split." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.447, + 0.483, + 0.688 + ], + "angle": 0, + "content": "additionally leverages the depth modality to model sociodynamic influence, but the mAP is lower than ours. What's more, there are 18 top-two categories (illustrated in red/blue color) of our methods, whereas [24] has 12 top-two categories. (2) For commonly used categories (anger, fear, happiness, sadness, surprise), [24] reports an averaged accuracy of \\(30.70\\%\\). In contrast, our method achieves a higher mean result of \\(37.19\\%\\). (3) In Tab. 2, we demonstrate that our method outperforms the state-of-the-art method [36] in both C-F1 and O-F1 scores. (4) Our approach also outperforms VLM methods (BLIP2, LLaVa, GPT-4o). This can be attributed to the lack of alignment of multi-level features in VLM for emotion recognition, suggesting that the alignment of tokenized image and text still produce confused results. Please refer to Appendix (Sec.10.1) for more detailed analysis of multi-level prediction confusion in VLM." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.483, + 0.901 + ], + "angle": 0, + "content": "CAER-S test Split. Our method is further evaluated on the extensive CAER-S dataset. Unlike EMOTIC, CAER-S encompasses a more intricate background. Thus the results of VLM-related methods are significantly inferior compared to those of other emotion recognition baselines. The accuracy results are presented in Tab. 3 and compared with prominent and recently-published works, demonstrating the exceptional performance of our proposed method. Notably, our approach shows improvement compared to [19, 36], which involves the extraction of hierarchical feature from body and background regions, as well as the exploration of interdependencies across emotion labels using word embeddings or graph embeddings. This comparison highlights the efficacy of our proposed hierarchical feature fusion method." + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.09, + 0.905, + 0.269 + ], + "angle": 0, + "content": "
DatasetHFE TypeV-I TypeWorkspacemAP
EMOTICResNet-50Prototype✓36.71
Cross-Level Attn.CLIP✓36.12
Cross-Level Attn.PrototypeX35.73
Cross-Level Attn.Prototype✓37.32
DatasetHFE TypeV-I TypeWorkspaceAcc.
CAER-SResNet-50Prototype✓90.16
Cross-Level Attn.CLIP✓88.28
Cross-Level Attn.PrototypeX87.60
Cross-Level Attn.Prototype✓91.86
" + }, + { + "type": "table_caption", + "bbox": [ + 0.528, + 0.28, + 0.885, + 0.294 + ], + "angle": 0, + "content": "Table 4. Ablation Study on EMOTIC and CAER-S datasets." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.305, + 0.667, + 0.321 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.327, + 0.905, + 0.356 + ], + "angle": 0, + "content": "We conduct comprehensive ablation experiments by different module combinations in our method:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.358, + 0.905, + 0.387 + ], + "angle": 0, + "content": "- For hierarchical feature extraction (HFE) module, we compare the proposed method with vanilla ResNet-50." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.388, + 0.905, + 0.432 + ], + "angle": 0, + "content": "- In vocabulary-informed (V-I) module, we substitute pseudo label from prototype-aligned projections (11) with CLIP predictions." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.433, + 0.905, + 0.644 + ], + "angle": 0, + "content": "- In terms of the workspace, we compare with feature concatenation \\(\\mathbf{z}_{\\mathbf{G}}\\) without feature refinement, denoted as \\(X\\). First, in Tab. 4, we observe that performance decreases significantly on both datasets when the workspace is degraded. This indicates that the workspace module, along with the competition and coordination mechanism, plays a vital role in coordinating multi-level image representation, thus benefiting the emotion recognition task. Secondly, we show that prototype learning is more important in the complex-environment dataset CAER-S, as it exhibits a greater performance decrease than EMOTIC. Finally, by replacing cross-level attention with ResNet-50, we observe lower performance in both datasets, indicating that modeling multi-level correlation facilitates effective emotion inference." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.358, + 0.905, + 0.644 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.653, + 0.706, + 0.669 + ], + "angle": 0, + "content": "4.4. Qualitative Analysis" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.675, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Case Study. In Fig. 4 we demonstrate the effectiveness of our proposed method via two randomly-chosen results of EMOTIC dataset. The left two columns contain and ground truth (GT) label and our prediction results. The prediction successfully matches GT, regardless of different and inconsistent recognition results of head, body and context window. Our approach further delineates the elimination in step two (see Fig 3) of competition process based on value of \\(\\mathcal{L}_{\\mathrm{hier}}\\) of maximum loss. It illustrates that with the guidance of pseudo label from V-I module, the inconsistent recognition results are gradually eliminated. We use orange color to high-light the labels in multi-level results that are not shown in V-I and prediction column. For example, from the top row of Fig. 4, the head level, which is the most inconsistent level of recognition results, is initially filtered out, and" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "29597" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.149, + 0.09, + 0.32, + 0.235 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.324, + 0.091, + 0.498, + 0.238 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.091, + 0.673, + 0.238 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.678, + 0.091, + 0.85, + 0.238 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.149, + 0.232, + 0.321, + 0.361 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.324, + 0.239, + 0.498, + 0.38 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.239, + 0.673, + 0.38 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.678, + 0.239, + 0.851, + 0.378 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.39, + 0.908, + 0.42 + ], + "angle": 0, + "content": "Figure 5. t-SNE and saliency maps visualization of feature refinement on CAER-S. The red star refers to a random selected sample feature that has been clustered in a more compact group in t-SNE visualization, and displaying more semantic regions on saliency maps." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.431, + 0.483, + 0.492 + ], + "angle": 0, + "content": "then prediction from the context level is subsequently eliminated. This supports our hypothesis that if the inconsistent multi-level emotion recognition is excluded, the final prediction performance is enhanced." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.492, + 0.483, + 0.718 + ], + "angle": 0, + "content": "How is multi-level feature refined? To understand the feature refinement in workspace, we employ t-SNE on CAERS to visualize multi-level feature, with 250 randomly sampled testing images per class. As shown in Fig. 5, features \\(\\mathbf{z}_m\\) from different level \\(m\\) are refined after inference-stage parameter update of \\(\\theta\\), displaying more clustered groups. Consequently, concatenated feature \\([\\hat{\\mathbf{z}}_{\\mathbf{c}}; \\hat{\\mathbf{z}}_{\\mathbf{b}}; \\hat{\\mathbf{z}}_{\\mathbf{h}}]\\) encompasses compact clusters of in-group representation. We also provide a clustered sample point (shown in red star in Fig. 5) by Grad-CAM [31], a wildly used saliency maps visualization technique of highlighting the semantic pixels. We can observe that with the parameter update of \\(\\theta\\), body-level and head-level saliency maps align with hand gestures and facial expressions, which contain more semantic information for emotion recognition." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.719, + 0.483, + 0.839 + ], + "angle": 0, + "content": "Is the proposed framework potential to enhance VLM? Since the performance of VLM is worse than expected, we aim to explore whether we can use our proposed framework to improve emotion recognition of VLM. First, based on step two of the competition process, we can obtain the first excluded image level, which we denote as the most inconsistent level \\(\\kappa\\). Next, using this level \\(\\kappa\\), we modify the input of the VLM in two ways:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.84, + 0.483, + 0.871 + ], + "angle": 0, + "content": "- Prompt\\(\\kappa\\): we add the prompt with: \"Please avoid potentially inconsistent recognition results from \\(\\{\\kappa\\}\\).\"" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.483, + 0.9 + ], + "angle": 0, + "content": "- Multi-Level\\(^{\\kappa}\\): we use multi-level image inputs but eliminate the level \\(\\kappa\\)." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.84, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.429, + 0.905, + 0.491 + ], + "angle": 0, + "content": "
MethodBLIP2 [17]LLaVa [20]GPT-4o [27]
PromptK26.77+0.8128.56+0.5725.11+1.31
Multi-LevelK27.14+1.1829.03+1.0426.64+2.84
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.501, + 0.907, + 0.53 + ], + "angle": 0, + "content": "Table 5. We report the mAP of enhanced VLM results on EMOTIC dataset. \\(\\kappa\\) refers to modified input via our proposed framework." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.543, + 0.907, + 0.62 + ], + "angle": 0, + "content": "As shown in Tab. 5, both of them successfully enhance emotion recognition capabilities of VLM, especially for GPT-40. Thus, we validate the inconsistency of multi-level inputs also exists in VLM and we improve it using our framework. Please see Appendix (Sec.10.2) for detailed cases." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.634, + 0.634, + 0.649 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.659, + 0.907, + 0.841 + ], + "angle": 0, + "content": "In this paper, we propose a competition and coordination mechanism to avoid hierarchical visual feature inconsistency in emotion recognition. To guide the alignment of visual features, a vocabulary-informed module is applied to produce labels that refine hierarchical visual representation. We elaborate on the competition process to eliminate irrelevant image-level representations and the coordination process to enhance representations across all image levels. Extensive experiments are conducted on two popular benchmarks, the EMOTIC and CAER-S datasets. Our method not only achieves state-of-the-art performance with better interpretability but also improves large vision-language models." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.841, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Acknowledgments. This work was supported by Shanghai Science and Technology Committee under Grant No. 24511103900, 24511103200, and National Key RD Program of China under grant No. 2022YFC3601405." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "29598" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.116, + 0.482, + 0.142 + ], + "angle": 0, + "content": "[1] Bernard J Baars. A cognitive theory of consciousness. Cambridge University Press, 1993. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.145, + 0.483, + 0.199 + ], + "angle": 0, + "content": "[2] Lenore Blum and Manuel Blum. A theory of consciousness from a theoretical computer science perspective: Insights from the conscious Turing machine. Proceedings of the National Academy of Sciences, 119(21):e2115934119, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.201, + 0.483, + 0.255 + ], + "angle": 0, + "content": "[3] Jean-Benoit Delbrouck, Noé Tits, Mathilde Brousmiche, and Stéphane Dupont. A transformer-based joint-encoding for emotion recognition and sentiment analysis. arXiv preprint arXiv:2006.15955, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.257, + 0.482, + 0.297 + ], + "angle": 0, + "content": "[4] Jiankang Deng, Jia Guo, Xue Niannan, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In CVPR, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.299, + 0.482, + 0.381 + ], + "angle": 0, + "content": "[5] Jiayuan Dong, Alex Santiago-Anaya, and Myounghoon Jeon. Facial expressions increase emotion recognition clarity and improve warmth and attractiveness on a humanoid robot without adding the uncanny valley. Proceedings of the Human Factors and Ergonomics Society Annual Meeting, 67 (1):933-939, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.383, + 0.482, + 0.464 + ], + "angle": 0, + "content": "[6] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.467, + 0.482, + 0.535 + ], + "angle": 0, + "content": "[7] Bernd Dudzik, Joost Broekens, Mark Neerincx, and Hayley Hung. Exploring personal memories and video content as context for facial behavior in predictions of video-induced emotions. In Proceedings of the 2020 International Conference on Multimodal Interaction, pages 153-162, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.537, + 0.482, + 0.605 + ], + "angle": 0, + "content": "[8] Yanwei Fu, Xiaomei Wang, Hanze Dong, Yu-Gang Jiang, Meng Wang, Xiangyang Xue, and Leonid Sigal. Vocabulary-informed zero-shot and open-set learning. IEEE Transactions on Pattern Analysis and Machine Intelligence, 42(12): 3136-3152, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.607, + 0.482, + 0.661 + ], + "angle": 0, + "content": "[9] Manh-Hung Hoang, Soo-Hyung Kim, Hyung-Jeong Yang, and Guee-Sang Lee. Context-aware emotion recognition based on visual relationship detection. IEEE Access, 9: 90465-90474, 2021. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.663, + 0.482, + 0.732 + ], + "angle": 0, + "content": "[10] Jian Huang, Jianhua Tao, Bin Liu, Zheng Lian, and Mingyue Niu. Multimodal transformer fusion for continuous emotion recognition. In ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 3507-3511. IEEE, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.734, + 0.482, + 0.76 + ], + "angle": 0, + "content": "[11] Oleg Kalyta. Information technology of facial emotion recognition for visual safety surveillance. 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.762, + 0.483, + 0.816 + ], + "angle": 0, + "content": "[12] Ronak Kosti, Jose M Alvarez, Adria Recasens, and Agata Lapedriza. Emotic: Emotions in context dataset. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 61-69, 2017. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.818, + 0.483, + 0.872 + ], + "angle": 0, + "content": "[13] Ronak Kosti, Jose M Alvarez, Adria Recasens, and Agata Lapedriza. Emotion recognition in context. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1667-1675, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.874, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[14] Ronak Kosti, Jose M Alvarez, Adria Recasens, and Agata Lapedriza. Context based emotion recognition using emotic" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.116, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.094, + 0.905, + 0.12 + ], + "angle": 0, + "content": "dataset. IEEE transactions on pattern analysis and machine intelligence, 42(11):2755-2766, 2019. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.123, + 0.905, + 0.178 + ], + "angle": 0, + "content": "[15] Jiyoung Lee, Seungryong Kim, Sunok Kim, Jungin Park, and Kwanghoon Sohn. Context-aware emotion recognition networks. In 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 10142-10151, 2019. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.18, + 0.905, + 0.206 + ], + "angle": 0, + "content": "[16] Li et al. Decoupled multimodal distilling for emotion recognition. In CVPR, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.208, + 0.905, + 0.262 + ], + "angle": 0, + "content": "[17] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, 2022. 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.265, + 0.905, + 0.306 + ], + "angle": 0, + "content": "[18] Qingbiao Li, Chunhua Wu, Zhe Wang, and Kangfeng Zheng. Hierarchical transformer network for utterance-level emotion recognition. Applied Sciences, 10(13):4447, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.308, + 0.905, + 0.349 + ], + "angle": 0, + "content": "[19] Weixin Li, Xuan Dong, and Yunhong Wang. Human emotion recognition with relational region-level analysis. IEEE Transactions on Affective Computing, 2021. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.351, + 0.904, + 0.377 + ], + "angle": 0, + "content": "[20] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, 2023. 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.38, + 0.905, + 0.449 + ], + "angle": 0, + "content": "[21] Xin Liu, Henglin Shi, Haoyu Chen, Zitong Yu, Xiaobai Li, and Guoying Zhao. imigue: An identity-free video dataset for micro-gesture understanding and emotion analysis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10631-10642, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.451, + 0.905, + 0.532 + ], + "angle": 0, + "content": "[22] Fengmao Lv, Xiang Chen, Yanyong Huang, Lixin Duan, and Guosheng Lin. Progressive modality reinforcement for human multimodal emotion recognition from unaligned multimodal sequences. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2554-2562, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.535, + 0.905, + 0.604 + ], + "angle": 0, + "content": "[23] Trisha Mittal, Uttaran Bhattacharya, Rohan Chandra, Aniket Bera, and Dinesh Manocha. M3er: Multiplicative multimodal emotion recognition using facial, textual, and speech cues. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1359-1367, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.606, + 0.905, + 0.687 + ], + "angle": 0, + "content": "[24] Trisha Mittal, Pooja Guhan, Uttaran Bhattacharya, Rohan Chandra, Aniket Bera, and Dinesh Manocha. Emoticon: Context-aware multimodal emotion recognition using frege's principle. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14234-14243, 2020. 1, 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.69, + 0.905, + 0.743 + ], + "angle": 0, + "content": "[25] Trisha Mittal, Aniket Bera, and Dinesh Manocha. Multimodal and context-aware emotion perception model with multiplicative fusion. IEEE MultiMedia, 28(2):67-75, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.746, + 0.907, + 0.815 + ], + "angle": 0, + "content": "[26] Trisha Mittal, Puneet Mathur, Aniket Bera, and Dinesh Manocha. Affect2mm: Affective analysis of multimedia content using emotion causality. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5661-5671, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.817, + 0.905, + 0.857 + ], + "angle": 0, + "content": "[27] OpenAI. GPT-4o API, 2024. Available at: https://openai.com/index/hello-gpt-4o, last accessed on 2024-11-01. 1, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[28] Kunyu Peng, Alina Roitberg, David Schneider, Marios Koulakis, Kailun Yang, and Rainer Stiefelhagen. Affect-dml: Context-aware one-shot recognition of human affect" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.094, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "29599" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.484, + 0.134 + ], + "angle": 0, + "content": "using deep metric learning. In 2021 16th IEEE International Conference on Automatic Face and Gesture Recognition (FG 2021), pages 1-8, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.484, + 0.207 + ], + "angle": 0, + "content": "[29] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision, 2021. 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.208, + 0.484, + 0.288 + ], + "angle": 0, + "content": "[30] Shulan Ruan, Kun Zhang, Yijun Wang, Hanqing Tao, Weidong He, Guangyi Lv, and Enhong Chen. Context-aware generation-based net for multi-label visual emotion recognition. In 2020 IEEE International Conference on Multimedia and Expo (ICME), pages 1-6. IEEE Computer Society, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.291, + 0.484, + 0.375 + ], + "angle": 0, + "content": "[31] Ramprasaath R. Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In 2017 IEEE International Conference on Computer Vision (ICCV), pages 618-626, 2017. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.377, + 0.484, + 0.459 + ], + "angle": 0, + "content": "[32] Jiahui She, Yibo Hu, Hailin Shi, Jun Wang, Qiu Shen, and Tao Mei. Dive into ambiguity: latent distribution mining and pairwise uncertainty estimation for facial expression recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6248-6257, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.461, + 0.484, + 0.532 + ], + "angle": 0, + "content": "[33] Xuli Shen, Xiaomei Wang, Qing Xu, Weifeng Ge, and Xi-angyang Xue. Towards scalable and fast distributionally robust optimization for data-driven deep learning. In 2022 IEEE International Conference on Data Mining (ICDM), pages 448–457, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.533, + 0.484, + 0.588 + ], + "angle": 0, + "content": "[34] Shamane Siriwardhana, Tharindu Kaluarachchi, Mark Billinghurst, and Suranga Nanayakkara. Multimodal emotion recognition with transformer-based self supervised feature fusion. IEEE Access, 8:176274-176285, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.59, + 0.484, + 0.645 + ], + "angle": 0, + "content": "[35] Zili Wang, Lingjie Lao, Xiaoya Zhang, Yong Li, Tong Zhang, and Zhen Cui. Context-dependent emotion recognition. Journal of Visual Communication and Image Representation, 89:103679, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.647, + 0.484, + 0.702 + ], + "angle": 0, + "content": "[36] Shichao Wu, Lei Zhou, Zhengxi Hu, and Jingtai Liu. Hierarchical context-based emotion recognition with scene graphs. IEEE Transactions on Neural Networks and Learning Systems, 35(3):3725-3739, 2024. 1, 2, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.704, + 0.484, + 0.788 + ], + "angle": 0, + "content": "[37] Dingkang Yang, Shuai Huang, Haopeng Kuang, Yangtao Du, and Lihua Zhang. Disentangled representation learning for multimodal emotion recognition. In Proceedings of the 30th ACM International Conference on Multimedia, pages 1642-1651, New York, NY, USA, 2022. Association for Computing Machinery. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.789, + 0.484, + 0.871 + ], + "angle": 0, + "content": "[38] Dingkang Yang, Zhaoyu Chen, Yuzheng Wang, Shunli Wang, Mingcheng Li, Siao Liu, Xiao Zhao, Shuai Huang, Zhiyan Dong, Peng Zhai, and Lihua Zhang. Context de-confounded emotion recognition. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.484, + 0.903 + ], + "angle": 0, + "content": "[39] Dingkang Yang, Mingcheng Li, Dongling Xiao, Yang Liu, Kun Yang, Zhaoyu Chen, Yuzheng Wang, Peng Zhai, Ke Li," + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.092, + 0.905, + 0.121 + ], + "angle": 0, + "content": "and Lihua Zhang. Towards multimodal sentiment analysis debiasing via bias purification, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[40] Dingkang Yang, Kun Yang, Mingcheng Li, Shunli Wang, Shuaibing Wang, and Lihua Zhang. Robust emotion recognition in context debiasing. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.192, + 0.905, + 0.261 + ], + "angle": 0, + "content": "[41] Jingyuan Yang, Jie Li, Leida Li, Xiumei Wang, and Xinbo Gao. A circular-structured representation for visual emotion distribution learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4237–4246, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.263, + 0.905, + 0.319 + ], + "angle": 0, + "content": "[42] Minghui Zhang, Yumeng Liang, and Huadong Ma. Context-aware affective graph reasoning for emotion recognition. In 2019 IEEE International Conference on Multimedia and Expo (ICME), pages 151-156. IEEE, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.32, + 0.905, + 0.387 + ], + "angle": 0, + "content": "[43] Yang Zhao, Hao Zhang, and Xiuyuan Hu. Penalizing gradient norm for efficiently improving generalization in deep learning. In Proceedings of the 39th International Conference on Machine Learning, pages 26982-26992. PMLR, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.39, + 0.905, + 0.445 + ], + "angle": 0, + "content": "[44] Zengqun Zhao, Qingshan Liu, and Shanmin Wang. Learning deep global multi-scale and local attention features for facial expression recognition in the wild. IEEE Transactions on Image Processing, 30:6544-6556, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.447, + 0.905, + 0.503 + ], + "angle": 0, + "content": "[45] Zengqun Zhao, Qingshan Liu, and Feng Zhou. Robust lightweight facial expression recognition network with label distribution training. In Proceedings of the AAAI conference on artificial intelligence, pages 3510-3519, 2021. 7" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.503 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "29600" + } + ] +] \ No newline at end of file diff --git a/2025/CocoER_ Aligning Multi-Level Feature by Competition and Coordination for Emotion Recognition/91abd7f6-c62e-4947-ae7d-c03321b50ab1_origin.pdf b/2025/CocoER_ Aligning Multi-Level Feature by Competition and Coordination for Emotion Recognition/91abd7f6-c62e-4947-ae7d-c03321b50ab1_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..1f6e375d8562163e657b66409664851d68df6d04 --- /dev/null +++ b/2025/CocoER_ Aligning Multi-Level Feature by Competition and Coordination for Emotion Recognition/91abd7f6-c62e-4947-ae7d-c03321b50ab1_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c32f08c9981cfa514871b397dfa321a1e47093d4a777fb9624d1b84ece8db5d4 +size 6541693 diff --git a/2025/CocoER_ Aligning Multi-Level Feature by Competition and Coordination for Emotion Recognition/full.md b/2025/CocoER_ Aligning Multi-Level Feature by Competition and Coordination for Emotion Recognition/full.md new file mode 100644 index 0000000000000000000000000000000000000000..3324d65a2dafaad47c7fdf2a3b588d6d8e4090e3 --- /dev/null +++ b/2025/CocoER_ Aligning Multi-Level Feature by Competition and Coordination for Emotion Recognition/full.md @@ -0,0 +1,349 @@ +# CocoER: Aligning Multi-Level Feature by Competition and Coordination for Emotion Recognition + +Xuli Shen $^{1,2}$ Hua Cai $^{2}$ Weilin Shen $^{2}$ Qing Xu $^{2}$ Dingding Yu $^{2}$ Weifeng Ge $^{1\star}$ Xiangyang Xue $^{1\star}$ Fudan University $^{2}$ UniDT + +{x1shen20,wfge,xyxue}@fudan.edu.cn， {hua.cai,weilin.shen,qing.xu,dingding.yu}@unidt.com + +# Abstract + +With the explosion of human-machine interaction, emotion recognition has reignited attention. Previous works focus on improving visual feature fusion and reasoning from multiple image levels. Although it is non-trivial to deduce a person's emotion by integrating multi-level feature (head, body and context), the emotion recognition results of each level is usually different from one another, which creates inconsistency in the prevailing feature alignment method and decrease recognition performance. In this work, we propose a multi-level image feature refinement method for emotion recognition (CocoER) to mitigate the impact caused by conflicting results from multi-level recognition. First, we leverage cross-level attention to improve visual feature consistency between hierarchically cropped head, body and context windows. Then, vocabulary informed alignment is incorporated into the recognition framework to produce pseudo label and guide hierarchical visual feature refinement. To effectively fuse multi-level feature, we elaborate on a competition process of eliminating irrelevant image level predictions and a coordination process to enhance the feature across all levels. Extensive experiments are executed on two popular datasets, and our method achieves state-of-the-art performance with multi-level interpretation results. Code is available at: https://github.com/bisno/CocoER. + +# 1. Introduction + +Visual emotion recognition plays a vital role in human computer interaction [27], surveillance [11], and robotics [5]. It aims to perceive the emotion state of human, and machine can modify responses and behaviors by considering the sentiment attitudes. Recently, impressive progress in recognizing emotions is achieved with expression from additional modality, such as head, body, background, depth, and so on [24]. Besides, emotion representation can be enhanced in view of discrete categories [24], points in a continuous + +![](images/858bcb9d3548b96e064815fc3c03f48ebe0cad3e87f19d9a30aa07ab24d8ca60.jpg) +Figure 1. To avoid conflicts arising from inconsistent multi-level predictions, competition and coordination are applied in refining multi-level visual representation for emotion recognition, guided by pseudo label from vocabulary-informed alignment. + +space [13, 42] and label distributions [41]. + +As emotion is naturally coupled in human and environments, the complex background in image still hinders the research of emotion recognition. Pioneering researches focus on perceiving the emotion states of the specific individuals in the given image with hierarchical visual feature [24, 37, 40] on context-aware emotion recognition datasets [13, 15]. [14] trains different backbones to combine the information of the target body and the entire image. Subsequently, a context-aware emotion recognition architecture enhances the classified precision to a high level representation [24]. In addition to image feature, knowledge graph is incorporated to reason the emotion states in [36]. To align feature across different modalities, [37] aims to address distribution gap between heterogeneous modalities. + +Despite previous progress in decoupling irrelevant contextual information in the inference stage [38, 40], there still lacks of multi-level image feature alignment method to precisely recognize human emotions. Each level has dif + +ferent and potentially misleading representation and predictions, which decreases accuracy of emotion recognition. As shown in Fig. 1, multi-level emotion output from the head (fear and doubt), and context (sad) confuse final emotion prediction. We hypothesize that conflicting results from each level interfere with emotional decision-making process. To better align multi-level feature, researchers often draws upon prior knowledge, such as knowledge graph, to select relevant object. Although knowledge graph [36] is applied to enhance interpretability, extracting object relevance does not directly address the challenge of conflicted multi-level predictions and there is a lack of a guidance that can combine head, body, and background feature. + +In this work, we propose competition and coordination mechanism to effectively perform multi-level fusion of head, body, and context emotion representation. Specifically, we apply cross-level attention mechanism with linear mapping on extracted feature maps to extract correlation from multi-level windows and obtain multi-level embeddings and recognition results. Furthermore, we leverage prototype learning to enhance the recognition result from external image-text aligned model, leading to pseudo label as guidance for multi-level feature refinement. Next, based on the pseudo label, we introduce competition process as an emotional feature selection method that filters out inconsistent multi-level feature. We leverage coordination process to update the parameters of the linear mapping for all image levels in inference-stage. At last, we re-send the multi-level feature to the updated linear mapping and achieve emotional consistency for final emotion recognition. In summary, the contributions of our paper are: + +- We enhance multi-level image correlation via hierarchical feature extraction module, and produce multi-level recognition. We leverage prediction from vocabulary-informed module as pseudo label to identify inconsistent levels. +- We introduce workspace module that incorporates competition and coordination process to rectify multi-level feature through updated model in inference-stage. +- Extensive experiments have been conducted to compare with baseline methods and large vision-language models (VLM), and the outcomes demonstrate superior performance and interpretability of our proposed method. + +# 2. Related Work + +# 2.1. Visual Emotion Recognition + +Visual emotion recognition is a comprehensive task in computer vision area, which is deeply associated with cognitive intelligence, and has attracted increasing attentions from researchers. Extensive works utilize facial expression to deduce the emotion state of specific person [7, 23, 25, 32, 35, 38]. Considering that recognizing a person's emotions can be subjective and has raised inherent ambiguity among sim + +ilar emotion categories, some researches focus on the latent distribution on the candidate emotion label sets [32, 41] and incorporates models from cognitive computing into the recognition algorithms [7, 21, 24, 26, 41]. Regarding network structures, [3, 18, 39] present transformer-based architectures to capture contextual embeddings. [10, 34, 37] utilize multi-head attention to fuse multiple modalities. In the aspect of inference stage, [38, 40] use causal relationships to eliminate information in the background that is inconsistent with target emotions. In contrast, we focuses on emotion recognition via multi-level visual representation that aligned by vocabulary-informed pseudo label. + +# 2.2. Multi-modal Emotion Recognition + +[22, 23, 26, 28, 40] enhance multi-modal fusion performance, and [12, 14, 24, 30] explore context-based visual expressions. To fully take advantages of various context information, multi-modal visual feature are incorporated into model inputs, such as face, pose and depth [24, 40]. Different modality information requires different sensors or off-the-shelf extractor, resulting in increasing model complexity and a narrowed application range. On the other hand, [14, 30] focus on addressing this problem with only global and body information, and are still able to achieve competitive performance. [36] introduces a hierarchical context-based emotion recognition method that uses scene graphs to infer emotions from images. + +In the field of large vision language model [17, 20], despite competitive performance in many multi-modal tasks, existing models do not prioritize emotion recognition. The modality alignment methods for large vision language models mostly focus on tokenized pair in the whole image, neglecting the multi-level image feature that misleads the image-text alignment. In this paper, we target this problem by eliminating irrelevant level of feature and further propose a novel multi-level feature alignment framework to enhance the performance of emotion recognition. + +# 3. Methods + +In this section, we first introduce the hierarchical feature extraction module on the cropped multi-level image, where each image level has its own emotion recognition outcome. Next, we elaborate on the vocabulary-informed alignment module, which serves as guidance that provides pseudo label for eliminating inconsistent multi-level predictions. Finally, based on such pseudo label, we apply the workspace module to enhance multi-level feature fusion. The proposed framework is depicted in Fig. 2. + +# 3.1. Hierarchical Feature Extraction + +We construct hierarchical emotion recognition by dividing the input image into three windows: head, body, and context. Specifically, given an input image $\mathbf{I} \in \mathbb{R}^{H \times W \times 3}$ , we + +![](images/6bb02efdd310c837e29ea24d7a49f7d5f308150aab87a3ccb8d701ffbc66a8b9.jpg) + +![](images/90889120229087d16b4b638f4e5a4e755c842f7e65f248648edf292a5ebde9be.jpg) + +![](images/ce603e5c4c4e0963bfcec4b26747f4d3ac12ed83ab52f29cb8223ace58c7fa91.jpg) +Figure 2. Illustration of the proposed framework. + +leverage object detection model [4] to obtain head, body, and context windows using detected bounding boxes. All three cropped windows are resized to $(H,W)$ , and are then fed into trainable backbones $\mathcal{R}$ to extract multi-level feature maps $\mathbf{u}_m$ with a size of $(h,w)$ and $c$ channels: + +$$ +\mathbf {u} _ {m} = \mathcal {R} _ {m} (\mathbf {I} _ {m}), m \in \{\mathbf {c}, \mathbf {b}, \mathbf {h} \}, \tag {1} +$$ + +where $\mathbf{c},\mathbf{b},\mathbf{h}$ are context, body, and head windows, respectively. Each $\mathbf{u}_m\in \mathbb{R}^{h\times w\times c}$ is divided and reshaped into $N$ non-overlapping patches $\mathbf{u}_m^p\in \mathbb{R}^{p^2\cdot c}$ and then mapped into $d_h$ dimensions with linear projection $\mathbf{E}_m\in \mathbb{R}^{(p^2\cdot c)\times d_h}$ . Thus each level $m$ consists of a sequence of 2D embeddings $\bar{\mathbf{u}}_m = [\mathbf{u}_m^{p,0}\mathbf{E}_m;\mathbf{u}_m^{p,1}\mathbf{E}_m;\dots ;\mathbf{u}_m^{p,N - 1}\mathbf{E}_m]\in \mathbb{R}^{N\times d_h}$ , where $N = hw / p^2$ is patch number [6]. $\bar{\mathbf{u}}_m$ is sent to cross-level attention $\mathcal{T}$ for enhancing multi-level correlation via connecting local and context feature. For instance, the head-context attention $\mathcal{T}_{\mathbf{h},\mathbf{c}}$ is formalized with linear projections $\{W_{\mathbf{h}}^{0},W_{\mathbf{h}}^{1},W_{\mathbf{c}}\} \in \mathbb{R}^{d_h\times d_h}$ with feature maps $\{\bar{\mathbf{u}}_{\mathbf{h}},\bar{\mathbf{u}}_{\mathbf{c}}\}$ : + +$$ +\mathbf {Q} = \bar {\mathbf {u}} _ {\mathbf {h}} W _ {\mathbf {h}} ^ {0}, \mathbf {K} = \bar {\mathbf {u}} _ {\mathbf {c}} W _ {\mathbf {c}}, \mathbf {V} = \bar {\mathbf {u}} _ {\mathbf {h}} W _ {\mathbf {h}} ^ {1} \tag {2} +$$ + +$$ +\gamma = \operatorname {s o f t m a x} \left(\mathbf {Q} \mathbf {K} ^ {T}\right) / \sqrt {d _ {h}} \quad \in \mathbb {R} ^ {N \times N} \tag {3} +$$ + +$$ +\bar {\mathbf {z}} _ {\mathbf {h}} = \boldsymbol {\gamma} \mathbf {V} \in \mathbb {R} ^ {N \times d _ {h}}. \tag {4} +$$ + +Then $\bar{\mathbf{z}}_{\mathbf{h}}$ is added with $\bar{\mathbf{u}}_{\mathbf{h}}$ and sent to typical fully-connected layers with skip-connection: + +$$ +\dot {\mathbf {z}} _ {\mathbf {h}} = \bar {\mathbf {z}} _ {\mathbf {h}} + \bar {\mathbf {u}} _ {\mathbf {h}} \in \mathbb {R} ^ {N \times d _ {h}} \tag {5} +$$ + +$$ +\mathbf {z} _ {\mathbf {h}} = \operatorname {F F N} _ {0} (\operatorname {L N} (\dot {\mathbf {z}} _ {\mathbf {h}})) + \dot {\mathbf {z}} _ {\mathbf {h}} \in \mathbb {R} ^ {d}, \tag {6} +$$ + +where LN is LayerNorm, $\mathrm{FFN}_0$ refers to the feed-forward network and $\mathbf{z_h} \in \mathbb{R}^d$ is the head-level representation. Thus we bridge hierarchical information to enhance head-level representation $\mathbf{z_h}$ via head-context cross attention $\mathcal{T}_{\mathbf{h},\mathbf{c}}$ . Similarly we have $\mathbf{z_b}$ via body-context cross attention $\mathcal{T}_{\mathbf{b},\mathbf{c}}$ . Therefore we leverage attention mechanism to enrich local feature representation $\{\bar{\mathbf{u}}_{\mathbf{h}},\bar{\mathbf{u}}_{\mathbf{c}}\}$ . Note that we conduct self-attention $\mathcal{T}_{\mathbf{c},\mathbf{c}}$ and obtain context feature $\mathbf{z_c}$ . Next, we pass each level representation $\mathbf{z}_m$ through a linear mapping $g_{\theta_m}$ , followed by an operation $\mathcal{F}$ to produce $k$ emotion category corresponding to multi-class task or multi-label task: + +$$ +y _ {m} = \mathcal {F} \left(g _ {\boldsymbol {\theta} _ {m}} \left(\mathbf {z} _ {m}\right)\right), m \in \{\mathbf {c}, \mathbf {b}, \mathbf {h} \}, \tag {7} +$$ + +where $y_{m}$ represents the prediction of each level $m$ . The parameter of linear mapping $g_{\theta_m}$ will be updated in inference-stage, and each level representation $\mathbf{z}_m$ will be refined for enhancing emotion recognition. + +# 3.2. Vocabulary-Informed Alignment + +In order to semantically guide feature refinement from hierarchical feature extraction, we propose a vocabulary-informed alignment method that leverages pre-trained image and text encoder to gain pseudo label and calibrate image representation. Specifically, given same input image $\mathbf{I} \in \mathbb{R}^{H \times W \times 3}$ , we improve CLIP [29] prediction to guide image feature alignment from text space. As shown in Fig. 2, we first encode the image through an CLIP image encoder $\mathbf{Enc}_{\mathrm{img}}$ . Then, we provide a text prompt for $k$ classes of emotion recognition task. The $k$ classes are injected and expanded into $k$ sentences of "This photo shows {} emotion", denoted as $\mathbf{S}_k$ , and are sent to CLIP text encoder $\mathbf{Enc}_{\mathrm{text}}$ . Thus, we have + +$$ +\mathbf {u} _ {\mathrm {i m g}} = \mathbf {E n c} _ {\mathrm {i m g}} (\mathbf {I}) \in \mathbb {R} ^ {d} \tag {8} +$$ + +$$ +\mathbf {U} _ {\text {t e x t}} = \mathbf {E n c} _ {\text {t e x t}} \left(\mathbf {S} _ {k}\right) \in \mathbb {R} ^ {d \times k}. \tag {9} +$$ + +One can directly use dot-product similarity to obtain recognition result of CLIP. However, merely applying CLIP prediction is insufficient because it is trained on open-world image-text pairs, which leads to difficulties in differentiating precise expressions such as "surprise" and "fear". To improve the recognition performance in vocabulary-informed module, we apply prototype learning [8] that projects image representation into text representation space and pushes closer to emotion prototype. In detail, after extracting image feature, we expand $\mathbf{u}_{\mathrm{img}}$ into $\hat{\mathbf{U}}_{\mathrm{img}}$ which has the same shape of $\mathbf{U}_{\mathrm{text}}$ and aim to minimize the distance $\mathcal{D}(\hat{\mathbf{U}}_{\mathrm{img}},\mathbf{U}_{\mathrm{text}})$ : + +$$ +\mathcal {D} \left(\hat {\mathbf {U}} _ {\mathrm {i m g}}, \mathbf {U} _ {\text {t e x t}}\right) = \left\| \omega^ {T} \hat {\mathbf {U}} _ {\mathrm {i m g}} - \mathbf {U} _ {\text {t e x t}} \right\| _ {2} ^ {2}, \tag {10} +$$ + +where $\hat{\mathbf{U}}_{\mathrm{img}} \in \mathbb{R}^{d \times k}$ , and $\omega \in \mathbb{R}^{d \times d}$ is a linear projection which is trained by minimizing the above distance of image feature and semantic emotion prototypes. Upon obtaining the trained $\omega$ , we feed the emotion-aligned image feature $\omega^T \hat{\mathbf{U}}_{\mathrm{img}}$ to $\mathrm{FFN}_1$ . Similar to (7) in the previous module that each hierarchical window has its own recognition result, the vocabulary-informed module also provides emotion recognition result: + +$$ +y _ {\mathbf {t}} = \mathcal {F} \left(\operatorname {F F N} _ {1} \left(\omega^ {T} \hat {\mathbf {U}} _ {\mathrm {i m g}}\right)\right). \tag {11} +$$ + +$y_{\mathbf{t}}$ serves as pseudo label for further feature refinement in workspace. We use the embeddings of last hidden layer of $\mathrm{FFN}_1$ as the vocabulary-informed feature $\mathbf{z}_{\mathbf{t}}$ , which has same feature dimension of image level feature $\mathbf{z} \in \mathbb{R}^d$ . Note that we freeze parameters of pre-trained $\omega$ , as well as $\mathbf{Enc}_{\mathrm{img}}$ and $\mathbf{Enc}_{\mathrm{text}}$ during the training stage of our framework, leaving $\mathrm{FFN}_1$ trainable. + +# 3.3. Workspace + +Different from the previous feature refinement methods, such as the entity-context feature alignment [36] or emotion + +![](images/0027eb2ef9440995765100063b19b441b345f207b99a5519d7af155cf2992193.jpg) +Figure 3. Detailed illustration of competition and coordination process, incorporated with inference-stage parameter update. + +feature disentanglement [16, 37], we introduce a workspace module that can adaptively refine all levels of emotion feature based on the vocabulary-informed classification results, as illustrated in the bottom part of Fig. 2. The name "workspace" is inspired by the Workspace Theory [1, 2], which describes how multi-level predictions compete and coordinate for final recognition. + +Formally, given the multi-level emotion feature set $\mathbf{Z} = \{\mathbf{z}_{\mathrm{c}},\mathbf{z}_{\mathrm{b}},\mathbf{z}_{\mathrm{h}}\}$ , we aim to gradually eliminate inconsistent level of feature in $\mathbf{Z}$ according to the pseudo label $y_{t}$ (11) until there is only one element in $\mathbf{Z}$ . We refer to this process as multi-level feature competition. Subsequently, we leverage the selection process to update parameters of $g_{\theta_m}$ and re-send each $\mathbf{z}_m$ to $g_{\theta_m}$ to obtain the refined feature $\hat{\mathbf{z}}_m$ . We call this multi-level feature coordination. Technical details are provided in the next two sections. + +# 3.3.1. Competition Process + +The competition process is designed as an iteration and contains two steps during each iteration: 1) Determine the descent direction for updating parameters of $g_{\theta_m}$ ; 2) Eliminate inconsistent element in feature set $\mathbf{Z}$ . The iteration ends when there is only one element in $\mathbf{Z}$ , shown in Fig. 3. + +Step 1. We leverage the value of loss function $\mathcal{L}_{\mathrm{hier}}(\mathcal{F}(g_{\theta}(\mathbf{z})),y_{\mathbf{t}})$ as a measure of inconsistency between multi-level recognition and pseudo label. Then, in order to target the most inconsistent level, we take the maximum of the loss function groups for the results from $\mathcal{F}(g_{\theta_m}(\mathbf{z}_m))$ against pseudo label $y_{\mathbf{t}}$ : + +$$ +\max _ {\mathbf {z} _ {m} \in \mathbf {Z}} \left\{\mathcal {L} _ {\text {h i e r}} \left(g _ {\boldsymbol {\theta} _ {m}} \left(\mathbf {z} _ {m}\right), y _ {\mathbf {t}}\right) \right\}. \tag {12} +$$ + +For the convenience of notation, let $\mathbf{f}_{\mathbf{z}_m}(\pmb{\theta}_m) = \{\mathcal{L}_{\mathrm{hier}}(g_{\pmb{\theta}_m}(\mathbf{z}_m),y_t),\mathbf{z}_m\in \mathbf{Z}\}$ . During each iteration in the competition process, we aim to calculate the gradient of $\max_{\mathbf{z}_m\in \mathbf{Z}}\mathbf{f}_{\mathbf{z}_m}(\pmb{\theta}_m)$ so as to update parameters of $g_{\pmb{\theta}_m}$ . Thus, our goal transforms into a min-max problem: + +$$ +\min _ {\boldsymbol {\theta} _ {m}} \max _ {\mathbf {z} _ {m} \in \mathbf {Z}} \mathbf {f} _ {\mathbf {z} _ {m}} \left(\boldsymbol {\theta} _ {m}\right). \tag {13} +$$ + +However, there exists a non-trivial challenge in finding the descent direction when we implement $\max(\cdot)$ on the groups of loss functions because the above function is nondifferentiable. Inspired by [33], we bring differentiability to $\max_{\mathbf{z}_m \in \mathbf{Z}} \mathbf{f}_{\mathbf{z}_m}(\theta_m)$ and leverage a regularization term to stabilize the descent direction. Please refer to technical details in Appendix (Sec.1). Then, the min-max problem is reduced to the following quadratic programming (QP) problem with a Lagrange multiplier $\lambda$ : + +$$ +\begin{array}{l} \min _ {\boldsymbol {\lambda}} \left(\frac {1}{2} \boldsymbol {\lambda} ^ {T} \mathbf {G} \mathbf {G} ^ {T} \boldsymbol {\lambda} - \boldsymbol {\lambda} ^ {T} \mathbf {f} _ {\mathbf {z} _ {m}}\right) \tag {14} \\ \text {s . t .} \quad \sum_ {\mathbf {z} _ {m} \in \mathbf {Z}} \lambda_ {\mathbf {z} _ {m}} = 1, \lambda_ {\mathbf {z} _ {m}} \geq 0 \\ \end{array} +$$ + +where $\mathbf{G} = \nabla_{\boldsymbol{\theta}_m^l}\mathbf{f}_{\mathbf{z}_m} \in \mathbb{R}^{q \times (d \cdot k)}$ contains gradient of the prediction layer $\boldsymbol{\theta}_m^l$ with respect to the loss of $\mathbf{f}_{\mathbf{z}_m}$ and $q = |\mathbf{Z}|$ refers to the number of elements left in $\mathbf{Z}$ . After solving for $\boldsymbol{\lambda}$ in the QP problem, we have $\delta = -\mathbf{G}^T\boldsymbol{\lambda} \in \mathbb{R}^{d \cdot k}$ as the approximate solution of (13). Here, $\boldsymbol{\delta}$ serves as the shared parameter updating path for linear mapping $g_{\boldsymbol{\theta}_m}$ of all three levels, which will be used for feature refinement in coordination process. + +Step 2. Let $n = \arg \max_{\mathbf{z}_m \in \mathbf{Z}} \left\{ \mathcal{L}_{\mathrm{hier}}(g_{\theta_m}(\mathbf{z}_m), y_t) \right\}$ be the index of most inconsistent feature in $\mathbf{Z}$ . We iteratively eliminate the corresponding $n$ of feature $\mathbf{z}$ in $\mathbf{Z}$ , denoted as $\mathbf{Z}_{-n}$ . This elimination is illustrated in the top part of Fig. 3. + +# 3.3.2. Coordination Process + +The goal of coordination process is to rectify the multi-level feature $\mathbf{z}$ via updated linear mapping $g_{\theta}$ . Specifically, after iteratively obtaining the respective descent direction $\delta$ during the competition process, we accumulate the overall descent direction, shown in the bottom part of Fig. 3. For example, if the context-level feature $\mathbf{z}_{\mathbf{c}}$ and head-level feature $\mathbf{z_h}$ is gradually eliminated through competition, we have $\delta_{\mathbf{c}}$ and $\delta_{\mathbf{h}}$ based on (14), which construct the shared parameter updating path for the prediction layer $\theta_m^l$ as follows: + +$$ +\delta^ {l} = \delta_ {\mathrm {c}} + \delta_ {\mathrm {h}}. \tag {15} +$$ + +Then, for each image level $m$ we update entire parameter $\theta_{m}$ via chain rule and back-propagation, denoted as $\longleftarrow$ : + +$$ +\tilde {\theta} _ {\mathbf {c}} \leftarrow \theta_ {\mathbf {c}} ^ {l} + \delta^ {l}; \tilde {\theta} _ {\mathbf {b}} \leftarrow \theta_ {\mathbf {b}} ^ {l} + \delta^ {l}; \tilde {\theta} _ {\mathbf {h}} \leftarrow \theta_ {\mathbf {h}} ^ {l} + \delta^ {l}. \tag {16} +$$ + +Next, we re-send $\mathbf{z}_m$ to its corresponding linear mapping $g_{\tilde{\theta}}$ and acquire refined feature from last hidden layer and construct multi-level concatenated representation $[\hat{\mathbf{z}}_{\mathrm{c}};\hat{\mathbf{z}}_{\mathrm{b}};\hat{\mathbf{z}}_{\mathrm{h}}]$ : + +$$ +\hat {\mathbf {z}} _ {\mathrm {c}} = g _ {\tilde {\theta} _ {\mathrm {c}}} \left(\mathbf {z} _ {\mathrm {c}}\right); \hat {\mathbf {z}} _ {\mathrm {b}} = g _ {\tilde {\theta} _ {\mathrm {b}}} \left(\mathbf {z} _ {\mathrm {b}}\right); \hat {\mathbf {z}} _ {\mathrm {h}} = g _ {\tilde {\theta} _ {\mathrm {h}}} \left(\mathbf {z} _ {\mathrm {h}}\right). \tag {17} +$$ + +With such inference-stage parameter updating, we calibrate multi-level emotion representation, ensuring that the final + +emotion recognition results are less affected by the inconsistent emotion recognition outcomes from hierarchical feature. Please refer to algorithm in Appendix that describes competition and coordination process within workspace. + +# 3.4. Training Objectives + +We concatenate the refined multi-level representation with vocabulary-informed feature $\mathbf{z_t}$ in $\mathbf{z_G} = [\hat{\mathbf{z}}_{\mathrm{c}};\hat{\mathbf{z}}_{\mathrm{b}};\hat{\mathbf{z}}_{\mathrm{h}};\mathbf{z}_{\mathrm{t}}]$ . Then $\mathbf{z_G}$ is passed to feed-forward network, and output of the proposed method is: + +$$ +y _ {\mathbf {G}} = \mathcal {F} \left(\operatorname {F F N} _ {2} \left(\mathbf {z} _ {\mathbf {G}}\right)\right). \tag {18} +$$ + +We collectively form loss functions from HFE, V-I and workspace module as $\mathcal{L}_{\mathbf{z}}(y_{\mathbf{z}},y)$ with respect to ground-truth $y$ . The combined loss function $\mathcal{L}_{\mathrm{emo}}$ is defined as follows: + +$$ +\mathcal {L} _ {\mathrm {e m o}} = \mathcal {L} _ {\mathbf {c}} + \mathcal {L} _ {\mathbf {b}} + \mathcal {L} _ {\mathbf {h}} + \mathcal {L} _ {\mathbf {t}} + \mathcal {L} _ {\mathbf {G}}. \tag {19} +$$ + +Next, we add norm of $\delta^l$ [43] to regularize the shared updating path, aiming to avoid over-dependence on the vocabulary-informed pseudo label. Finally, all parameters of our proposed framework are trained based on: + +$$ +\mathcal {L} = \mathcal {L} _ {\mathrm {e m o}} + \alpha \| \boldsymbol {\delta} ^ {l} \| _ {2}, \tag {20} +$$ + +where $\alpha$ is hyperparameter, empirically set to 0.1. + +# 4. Experiments + +# 4.1. Setup + +Datasets. EMOTIC [12] comprises 23,571 images featuring 34,320 individuals, predominantly captured in unconstrained environments. Each image encompasses multiple emotions, presenting a more intricate challenge of multi-label emotion recognition across 26 discrete emotion categories. CAER-S [15] is derived by extracting static frames from video clips in the CAER dataset [15], containing over 70,000 images. These images are annotated with an expanded set of 7 emotion categories: anger, disgust, fear, happiness, sadness, surprise, and neutral. Both datasets are divided into training (80%) and testing (20%) subsets. Implementation Details. Please refer to Appendix. + +# 4.2. Comparison with State-of-the-Arts + +EMOTIC test Split. We first compare our method with several multimodal or multi-object based emotion recognition competitors on EMOTIC test split. As shown in Tab. 1, we report the average precision (AP) over all categories as well as the average results (mAP). From the table, we can observe that: (1) Our method achieves state-of-the-art performance with $37.32\%$ on mAP. Compared with [9] which utilizes Faster R-CNN to detect all possible objects in images for reasoning, we surpass it by a margin of $2.16\%$ using only three hierarchical cropped windows. Moreover, [24] + +
CategoriesLi et al. [19] Hoang et al. [9] Mittal et al. [24]† Wu et al. [36]BLIP2-6.7b [17] LLaVa-1.6-7b [20] GPT-4o [27]
2021202120202024202220232024
Affection37.9344.4845.2332.2524.2534.3222.53
Anger13.7330.7115.4610.3708.3004.9810.32
Annoyance20.8726.4721.9212.8408.9912.3709.98
Anticipation61.0859.8972.1256.8391.3591.8754.16
Aversion09.6112.4317.8107.7006.4506.4603.51
Confidence80.0879.2468.6577.5159.3762.0568.09
Disapproval21.5424.5419.8212.4707.5610.3808.64
Disconnection28.3234.2443.1226.4125.8531.0517.03
Disquietment22.5718.7324.2316.7513.5314.4517.17
Doubt/Confusion33.5025.4235.1218.6218.5819.0631.20
Embarrassment04.1604.2614.3702.8005.4505.4402.17
Engagement88.1288.7191.1286.8395.1496.1080.77
Esteem20.5017.9923.6215.5023.1923.2718.30
Excitement80.1174.2183.2670.7364.5170.3669.59
Fatigue17.5122.6216.2312.6407.6811.0309.53
Fear15.5613.9223.6505.4807.8408.4912.79
Happiness76.0183.0274.7169.4570.7775.1453.65
Pain14.5616.6813.2110.6405.5205.8107.45
Peace26.7628.9134.2724.6320.3221.2628.91
Pleasure55.6455.4765.5344.3343.3043.4942.86
Sadness30.8042.8723.4121.2107.0618.7210.70
Sensitivity09.5915.8908.3205.8904.7704.7503.96
Suffering30.7046.2326.3921.4805.2206.1709.74
Surprise17.9216.2717.3707.9415.4216.1417.66
Sympathy15.2615.3734.2813.2421.4821.6209.10
Yearning10.1110.0414.2909.1213.0512.7907.93
mAP32.4135.1635.4826.6825.9627.9923.80
+ +![](images/fa441faf561ae2b29f4f136c87747defc0123a42728dbe558e456fbb62003f74.jpg) + +Table 1. Performance comparisons on EMOTIC test split. The 1st/2nd best results are highlighted in red/blue. †' denotes the method that contains additional visual modalities, such as depth, which is not used in our method. + +
GTPredictionHeadBodyContextV-I (Pseudo label)
AffectionAffectionDisconnectionAffectionEngagementAffection
AnticipationConfidenceFatigueConfidenceExcitementConfidence
ConfidenceExcitementPleasureHappinessExcitement
EngagementHappinessPeaceHappiness
ExcitementPeacePleasure
HappinessPleasure
PeaceLhier=2.35Lhier=0.61Lhier=1.33
Pleasure
Exclusion Sequence: (1) Head→(2) Context→(3) Body
+ +![](images/7de892624f59c881d029e55340def0523c5da2e1b00bf7544a8c4d575e036438.jpg) +Figure 4. Case study of randomly sampled results on EMOTIC. The V-I column shows pseudo label in vocabulary-informed module. We display the elimination of inconsistent recognition results in orange, which are not shown in either the V-I column or the prediction column. + +
GTPredictionHeadBodyContextV-I (Pseudo label)
AnticipationAnticipationAnticipationAffectionEngagementAnticipation
ConfidenceConfidenceConfidenceSensitivityEmbarrassmentEngagement
EngagementEngagementExcitementSufferingPleasureEsteem
ExcitementExcitementHappinessHappiness
PeacePleasure
Lhier=1.42Lhier=2.36Lhier=1.03
Exclusion Sequence: (1) Body→(2) Context→(3) Head
+ +
MethodsC-F1 (%)O-F1 (%)
Kosti et al. [14] (2019)29.3354.45
Li et al. [19] (2021)15.1048.07
Wu et al. [36] (2024)34.2454.78
Ours36.7757.23
+ +Table 2. Label-based F1 score (C-F1) and example-based F1 score (O-F1) on EMOTIC test split. + +
MethodsAccuracy (%)
CLIP-ViT-L [29] (2021)21.57
BLIP2-6.7b [17] (2022)14.21
LLaVa-1.6-7b [20] (2023)28.59
GPT-4o [27] (2024)29.33
EfficientFace [45] (2021)85.87
MA-Net [44] (2021)88.42
Li et al. [19] (2021)84.42
Wu et al. [36] (2024)90.83
Ours91.86
+ +additionally leverages the depth modality to model sociodynamic influence, but the mAP is lower than ours. What's more, there are 18 top-two categories (illustrated in red/blue color) of our methods, whereas [24] has 12 top-two categories. (2) For commonly used categories (anger, fear, happiness, sadness, surprise), [24] reports an averaged accuracy of $30.70\%$ . In contrast, our method achieves a higher mean result of $37.19\%$ . (3) In Tab. 2, we demonstrate that our method outperforms the state-of-the-art method [36] in both C-F1 and O-F1 scores. (4) Our approach also outperforms VLM methods (BLIP2, LLaVa, GPT-4o). This can be attributed to the lack of alignment of multi-level features in VLM for emotion recognition, suggesting that the alignment of tokenized image and text still produce confused results. Please refer to Appendix (Sec.10.1) for more detailed analysis of multi-level prediction confusion in VLM. + +CAER-S test Split. Our method is further evaluated on the extensive CAER-S dataset. Unlike EMOTIC, CAER-S encompasses a more intricate background. Thus the results of VLM-related methods are significantly inferior compared to those of other emotion recognition baselines. The accuracy results are presented in Tab. 3 and compared with prominent and recently-published works, demonstrating the exceptional performance of our proposed method. Notably, our approach shows improvement compared to [19, 36], which involves the extraction of hierarchical feature from body and background regions, as well as the exploration of interdependencies across emotion labels using word embeddings or graph embeddings. This comparison highlights the efficacy of our proposed hierarchical feature fusion method. + +Table 3. Performance comparisons on CAER-S test split. + +
DatasetHFE TypeV-I TypeWorkspacemAP
EMOTICResNet-50Prototype✓36.71
Cross-Level Attn.CLIP✓36.12
Cross-Level Attn.PrototypeX35.73
Cross-Level Attn.Prototype✓37.32
DatasetHFE TypeV-I TypeWorkspaceAcc.
CAER-SResNet-50Prototype✓90.16
Cross-Level Attn.CLIP✓88.28
Cross-Level Attn.PrototypeX87.60
Cross-Level Attn.Prototype✓91.86
+ +Table 4. Ablation Study on EMOTIC and CAER-S datasets. + +# 4.3. Ablation Study + +We conduct comprehensive ablation experiments by different module combinations in our method: + +- For hierarchical feature extraction (HFE) module, we compare the proposed method with vanilla ResNet-50. +- In vocabulary-informed (V-I) module, we substitute pseudo label from prototype-aligned projections (11) with CLIP predictions. +- In terms of the workspace, we compare with feature concatenation $\mathbf{z}_{\mathbf{G}}$ without feature refinement, denoted as $X$ . First, in Tab. 4, we observe that performance decreases significantly on both datasets when the workspace is degraded. This indicates that the workspace module, along with the competition and coordination mechanism, plays a vital role in coordinating multi-level image representation, thus benefiting the emotion recognition task. Secondly, we show that prototype learning is more important in the complex-environment dataset CAER-S, as it exhibits a greater performance decrease than EMOTIC. Finally, by replacing cross-level attention with ResNet-50, we observe lower performance in both datasets, indicating that modeling multi-level correlation facilitates effective emotion inference. + +# 4.4. Qualitative Analysis + +Case Study. In Fig. 4 we demonstrate the effectiveness of our proposed method via two randomly-chosen results of EMOTIC dataset. The left two columns contain and ground truth (GT) label and our prediction results. The prediction successfully matches GT, regardless of different and inconsistent recognition results of head, body and context window. Our approach further delineates the elimination in step two (see Fig 3) of competition process based on value of $\mathcal{L}_{\mathrm{hier}}$ of maximum loss. It illustrates that with the guidance of pseudo label from V-I module, the inconsistent recognition results are gradually eliminated. We use orange color to high-light the labels in multi-level results that are not shown in V-I and prediction column. For example, from the top row of Fig. 4, the head level, which is the most inconsistent level of recognition results, is initially filtered out, and + +![](images/87de515feaea1f2141585b08b14d76b581601b0aabfb75e90f349e08af33e456.jpg) + +![](images/a77b652b353d3a71efa4a7f75669096296a80baa972a737b4e60913bf1bd7933.jpg) + +![](images/476deb5786f21a2fc3960a5ec981bdf487ed7a5e2ec8c2c645862c668a7fb24b.jpg) + +![](images/43d77da0600a0cfe4246b8921a7a59d3d0f6668f77746586f1cd1e9af4d68d36.jpg) + +![](images/260951db3fa588ad13282164965402d4e345823e2f4b3b3716ba704548cf1073.jpg) +Figure 5. t-SNE and saliency maps visualization of feature refinement on CAER-S. The red star refers to a random selected sample feature that has been clustered in a more compact group in t-SNE visualization, and displaying more semantic regions on saliency maps. + +![](images/65f8498f799a40f870622ec680b5325819540b4a4824000f062915bc0c97b5e3.jpg) + +![](images/4656457d35e05ba8aa9d14aa754cfdc808a7589b7ec9533aa9f7f5ca7fba36d2.jpg) + +![](images/e775eed61d8e7b74624bbb0e896e3b955f3f622643310421090cb0b4c7e3d70e.jpg) + +then prediction from the context level is subsequently eliminated. This supports our hypothesis that if the inconsistent multi-level emotion recognition is excluded, the final prediction performance is enhanced. + +How is multi-level feature refined? To understand the feature refinement in workspace, we employ t-SNE on CAERS to visualize multi-level feature, with 250 randomly sampled testing images per class. As shown in Fig. 5, features $\mathbf{z}_m$ from different level $m$ are refined after inference-stage parameter update of $\theta$ , displaying more clustered groups. Consequently, concatenated feature $[\hat{\mathbf{z}}_{\mathbf{c}}; \hat{\mathbf{z}}_{\mathbf{b}}; \hat{\mathbf{z}}_{\mathbf{h}}]$ encompasses compact clusters of in-group representation. We also provide a clustered sample point (shown in red star in Fig. 5) by Grad-CAM [31], a wildly used saliency maps visualization technique of highlighting the semantic pixels. We can observe that with the parameter update of $\theta$ , body-level and head-level saliency maps align with hand gestures and facial expressions, which contain more semantic information for emotion recognition. + +Is the proposed framework potential to enhance VLM? Since the performance of VLM is worse than expected, we aim to explore whether we can use our proposed framework to improve emotion recognition of VLM. First, based on step two of the competition process, we can obtain the first excluded image level, which we denote as the most inconsistent level $\kappa$ . Next, using this level $\kappa$ , we modify the input of the VLM in two ways: + +- Prompt $\kappa$ : we add the prompt with: "Please avoid potentially inconsistent recognition results from $\{\kappa\}$ ." +- Multi-Level $^{\kappa}$ : we use multi-level image inputs but eliminate the level $\kappa$ . + +
MethodBLIP2 [17]LLaVa [20]GPT-4o [27]
PromptK26.77+0.8128.56+0.5725.11+1.31
Multi-LevelK27.14+1.1829.03+1.0426.64+2.84
+ +Table 5. We report the mAP of enhanced VLM results on EMOTIC dataset. $\kappa$ refers to modified input via our proposed framework. + +As shown in Tab. 5, both of them successfully enhance emotion recognition capabilities of VLM, especially for GPT-40. Thus, we validate the inconsistency of multi-level inputs also exists in VLM and we improve it using our framework. Please see Appendix (Sec.10.2) for detailed cases. + +# 5. Conclusion + +In this paper, we propose a competition and coordination mechanism to avoid hierarchical visual feature inconsistency in emotion recognition. To guide the alignment of visual features, a vocabulary-informed module is applied to produce labels that refine hierarchical visual representation. We elaborate on the competition process to eliminate irrelevant image-level representations and the coordination process to enhance representations across all image levels. Extensive experiments are conducted on two popular benchmarks, the EMOTIC and CAER-S datasets. Our method not only achieves state-of-the-art performance with better interpretability but also improves large vision-language models. + +Acknowledgments. This work was supported by Shanghai Science and Technology Committee under Grant No. 24511103900, 24511103200, and National Key RD Program of China under grant No. 2022YFC3601405. + +# References + +[1] Bernard J Baars. A cognitive theory of consciousness. Cambridge University Press, 1993. 4 +[2] Lenore Blum and Manuel Blum. A theory of consciousness from a theoretical computer science perspective: Insights from the conscious Turing machine. Proceedings of the National Academy of Sciences, 119(21):e2115934119, 2022. 4 +[3] Jean-Benoit Delbrouck, Noé Tits, Mathilde Brousmiche, and Stéphane Dupont. A transformer-based joint-encoding for emotion recognition and sentiment analysis. arXiv preprint arXiv:2006.15955, 2020. 2 +[4] Jiankang Deng, Jia Guo, Xue Niannan, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In CVPR, 2019. 3 +[5] Jiayuan Dong, Alex Santiago-Anaya, and Myounghoon Jeon. Facial expressions increase emotion recognition clarity and improve warmth and attractiveness on a humanoid robot without adding the uncanny valley. Proceedings of the Human Factors and Ergonomics Society Annual Meeting, 67 (1):933-939, 2023. 1 +[6] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 3 +[7] Bernd Dudzik, Joost Broekens, Mark Neerincx, and Hayley Hung. Exploring personal memories and video content as context for facial behavior in predictions of video-induced emotions. In Proceedings of the 2020 International Conference on Multimodal Interaction, pages 153-162, 2020. 2 +[8] Yanwei Fu, Xiaomei Wang, Hanze Dong, Yu-Gang Jiang, Meng Wang, Xiangyang Xue, and Leonid Sigal. Vocabulary-informed zero-shot and open-set learning. IEEE Transactions on Pattern Analysis and Machine Intelligence, 42(12): 3136-3152, 2020. 4 +[9] Manh-Hung Hoang, Soo-Hyung Kim, Hyung-Jeong Yang, and Guee-Sang Lee. Context-aware emotion recognition based on visual relationship detection. IEEE Access, 9: 90465-90474, 2021. 5, 6 +[10] Jian Huang, Jianhua Tao, Bin Liu, Zheng Lian, and Mingyue Niu. Multimodal transformer fusion for continuous emotion recognition. In ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 3507-3511. IEEE, 2020. 2 +[11] Oleg Kalyta. Information technology of facial emotion recognition for visual safety surveillance. 2022. 1 +[12] Ronak Kosti, Jose M Alvarez, Adria Recasens, and Agata Lapedriza. Emotic: Emotions in context dataset. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 61-69, 2017. 2, 5 +[13] Ronak Kosti, Jose M Alvarez, Adria Recasens, and Agata Lapedriza. Emotion recognition in context. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1667-1675, 2017. 1 +[14] Ronak Kosti, Jose M Alvarez, Adria Recasens, and Agata Lapedriza. Context based emotion recognition using emotic + +dataset. IEEE transactions on pattern analysis and machine intelligence, 42(11):2755-2766, 2019. 1, 2, 7 +[15] Jiyoung Lee, Seungryong Kim, Sunok Kim, Jungin Park, and Kwanghoon Sohn. Context-aware emotion recognition networks. In 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 10142-10151, 2019. 1, 5 +[16] Li et al. Decoupled multimodal distilling for emotion recognition. In CVPR, 2023. 4 +[17] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, 2022. 2, 6, 7, 8 +[18] Qingbiao Li, Chunhua Wu, Zhe Wang, and Kangfeng Zheng. Hierarchical transformer network for utterance-level emotion recognition. Applied Sciences, 10(13):4447, 2020. 2 +[19] Weixin Li, Xuan Dong, and Yunhong Wang. Human emotion recognition with relational region-level analysis. IEEE Transactions on Affective Computing, 2021. 6, 7 +[20] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, 2023. 2, 6, 7, 8 +[21] Xin Liu, Henglin Shi, Haoyu Chen, Zitong Yu, Xiaobai Li, and Guoying Zhao. imigue: An identity-free video dataset for micro-gesture understanding and emotion analysis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10631-10642, 2021. 2 +[22] Fengmao Lv, Xiang Chen, Yanyong Huang, Lixin Duan, and Guosheng Lin. Progressive modality reinforcement for human multimodal emotion recognition from unaligned multimodal sequences. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2554-2562, 2021. 2 +[23] Trisha Mittal, Uttaran Bhattacharya, Rohan Chandra, Aniket Bera, and Dinesh Manocha. M3er: Multiplicative multimodal emotion recognition using facial, textual, and speech cues. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1359-1367, 2020. 2 +[24] Trisha Mittal, Pooja Guhan, Uttaran Bhattacharya, Rohan Chandra, Aniket Bera, and Dinesh Manocha. Emoticon: Context-aware multimodal emotion recognition using frege's principle. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14234-14243, 2020. 1, 2, 5, 6, 7 +[25] Trisha Mittal, Aniket Bera, and Dinesh Manocha. Multimodal and context-aware emotion perception model with multiplicative fusion. IEEE MultiMedia, 28(2):67-75, 2021. 2 +[26] Trisha Mittal, Puneet Mathur, Aniket Bera, and Dinesh Manocha. Affect2mm: Affective analysis of multimedia content using emotion causality. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5661-5671, 2021. 2 +[27] OpenAI. GPT-4o API, 2024. Available at: https://openai.com/index/hello-gpt-4o, last accessed on 2024-11-01. 1, 6, 7, 8 +[28] Kunyu Peng, Alina Roitberg, David Schneider, Marios Koulakis, Kailun Yang, and Rainer Stiefelhagen. Affect-dml: Context-aware one-shot recognition of human affect + +using deep metric learning. In 2021 16th IEEE International Conference on Automatic Face and Gesture Recognition (FG 2021), pages 1-8, 2021. 2 +[29] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision, 2021. 4, 7 +[30] Shulan Ruan, Kun Zhang, Yijun Wang, Hanqing Tao, Weidong He, Guangyi Lv, and Enhong Chen. Context-aware generation-based net for multi-label visual emotion recognition. In 2020 IEEE International Conference on Multimedia and Expo (ICME), pages 1-6. IEEE Computer Society, 2020. 2 +[31] Ramprasaath R. Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In 2017 IEEE International Conference on Computer Vision (ICCV), pages 618-626, 2017. 8 +[32] Jiahui She, Yibo Hu, Hailin Shi, Jun Wang, Qiu Shen, and Tao Mei. Dive into ambiguity: latent distribution mining and pairwise uncertainty estimation for facial expression recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6248-6257, 2021. 2 +[33] Xuli Shen, Xiaomei Wang, Qing Xu, Weifeng Ge, and Xi-angyang Xue. Towards scalable and fast distributionally robust optimization for data-driven deep learning. In 2022 IEEE International Conference on Data Mining (ICDM), pages 448–457, 2022. 5 +[34] Shamane Siriwardhana, Tharindu Kaluarachchi, Mark Billinghurst, and Suranga Nanayakkara. Multimodal emotion recognition with transformer-based self supervised feature fusion. IEEE Access, 8:176274-176285, 2020. 2 +[35] Zili Wang, Lingjie Lao, Xiaoya Zhang, Yong Li, Tong Zhang, and Zhen Cui. Context-dependent emotion recognition. Journal of Visual Communication and Image Representation, 89:103679, 2022. 2 +[36] Shichao Wu, Lei Zhou, Zhengxi Hu, and Jingtai Liu. Hierarchical context-based emotion recognition with scene graphs. IEEE Transactions on Neural Networks and Learning Systems, 35(3):3725-3739, 2024. 1, 2, 4, 6, 7 +[37] Dingkang Yang, Shuai Huang, Haopeng Kuang, Yangtao Du, and Lihua Zhang. Disentangled representation learning for multimodal emotion recognition. In Proceedings of the 30th ACM International Conference on Multimedia, pages 1642-1651, New York, NY, USA, 2022. Association for Computing Machinery. 1, 2, 4 +[38] Dingkang Yang, Zhaoyu Chen, Yuzheng Wang, Shunli Wang, Mingcheng Li, Siao Liu, Xiao Zhao, Shuai Huang, Zhiyan Dong, Peng Zhai, and Lihua Zhang. Context de-confounded emotion recognition. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2023. 1, 2 +[39] Dingkang Yang, Mingcheng Li, Dongling Xiao, Yang Liu, Kun Yang, Zhaoyu Chen, Yuzheng Wang, Peng Zhai, Ke Li, + +and Lihua Zhang. Towards multimodal sentiment analysis debiasing via bias purification, 2024. 2 +[40] Dingkang Yang, Kun Yang, Mingcheng Li, Shunli Wang, Shuaibing Wang, and Lihua Zhang. Robust emotion recognition in context debiasing. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2024. 1, 2 +[41] Jingyuan Yang, Jie Li, Leida Li, Xiumei Wang, and Xinbo Gao. A circular-structured representation for visual emotion distribution learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4237–4246, 2021. 1, 2 +[42] Minghui Zhang, Yumeng Liang, and Huadong Ma. Context-aware affective graph reasoning for emotion recognition. In 2019 IEEE International Conference on Multimedia and Expo (ICME), pages 151-156. IEEE, 2019. 1 +[43] Yang Zhao, Hao Zhang, and Xiuyuan Hu. Penalizing gradient norm for efficiently improving generalization in deep learning. In Proceedings of the 39th International Conference on Machine Learning, pages 26982-26992. PMLR, 2022. 5 +[44] Zengqun Zhao, Qingshan Liu, and Shanmin Wang. Learning deep global multi-scale and local attention features for facial expression recognition in the wild. IEEE Transactions on Image Processing, 30:6544-6556, 2021. 7 +[45] Zengqun Zhao, Qingshan Liu, and Feng Zhou. Robust lightweight facial expression recognition network with label distribution training. In Proceedings of the AAAI conference on artificial intelligence, pages 3510-3519, 2021. 7 \ No newline at end of file diff --git a/2025/CocoER_ Aligning Multi-Level Feature by Competition and Coordination for Emotion Recognition/images.zip b/2025/CocoER_ Aligning Multi-Level Feature by Competition and Coordination for Emotion Recognition/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..d9c15fa8b3bb1a18c1fc84bad77851baf296f733 --- /dev/null +++ b/2025/CocoER_ Aligning Multi-Level Feature by Competition and Coordination for Emotion Recognition/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:00d075924023b6c00086a8336fa4400e03bff31c891fe4a714e9920e1a113d5b +size 874849 diff --git a/2025/CocoER_ Aligning Multi-Level Feature by Competition and Coordination for Emotion Recognition/layout.json b/2025/CocoER_ Aligning Multi-Level Feature by Competition and Coordination for Emotion Recognition/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..4adfaf8c8b84398b072e13e2b5a0bdf0d51f0913 --- /dev/null +++ b/2025/CocoER_ Aligning Multi-Level Feature by Competition and Coordination for Emotion Recognition/layout.json @@ -0,0 +1,9870 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 67, + 102, + 542, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 102, + 542, + 140 + ], + "spans": [ + { + "bbox": [ + 67, + 102, + 542, + 140 + ], + "type": "text", + "content": "CocoER: Aligning Multi-Level Feature by Competition and Coordination for Emotion Recognition" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 159, + 548, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 159, + 548, + 190 + ], + "spans": [ + { + "bbox": [ + 61, + 159, + 548, + 190 + ], + "type": "text", + "content": "Xuli Shen" + }, + { + "bbox": [ + 61, + 159, + 548, + 190 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 61, + 159, + 548, + 190 + ], + "type": "text", + "content": " Hua Cai" + }, + { + "bbox": [ + 61, + 159, + 548, + 190 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 61, + 159, + 548, + 190 + ], + "type": "text", + "content": " Weilin Shen" + }, + { + "bbox": [ + 61, + 159, + 548, + 190 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 61, + 159, + 548, + 190 + ], + "type": "text", + "content": " Qing Xu" + }, + { + "bbox": [ + 61, + 159, + 548, + 190 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 61, + 159, + 548, + 190 + ], + "type": "text", + "content": " Dingding Yu" + }, + { + "bbox": [ + 61, + 159, + 548, + 190 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 61, + 159, + 548, + 190 + ], + "type": "text", + "content": " Weifeng Ge" + }, + { + "bbox": [ + 61, + 159, + 548, + 190 + ], + "type": "inline_equation", + "content": "^{1\\star}" + }, + { + "bbox": [ + 61, + 159, + 548, + 190 + ], + "type": "text", + "content": " Xiangyang Xue" + }, + { + "bbox": [ + 61, + 159, + 548, + 190 + ], + "type": "inline_equation", + "content": "^{1\\star}" + }, + { + "bbox": [ + 61, + 159, + 548, + 190 + ], + "type": "text", + "content": " Fudan University " + }, + { + "bbox": [ + 61, + 159, + 548, + 190 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 61, + 159, + 548, + 190 + ], + "type": "text", + "content": "UniDT" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 65, + 191, + 539, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 191, + 539, + 205 + ], + "spans": [ + { + "bbox": [ + 65, + 191, + 539, + 205 + ], + "type": "text", + "content": "{x1shen20,wfge,xyxue}@fudan.edu.cn， {hua.cai,weilin.shen,qing.xu,dingding.yu}@unidt.com" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "spans": [ + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 256, + 297, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 256, + 297, + 544 + ], + "spans": [ + { + "bbox": [ + 55, + 256, + 297, + 544 + ], + "type": "text", + "content": "With the explosion of human-machine interaction, emotion recognition has reignited attention. Previous works focus on improving visual feature fusion and reasoning from multiple image levels. Although it is non-trivial to deduce a person's emotion by integrating multi-level feature (head, body and context), the emotion recognition results of each level is usually different from one another, which creates inconsistency in the prevailing feature alignment method and decrease recognition performance. In this work, we propose a multi-level image feature refinement method for emotion recognition (CocoER) to mitigate the impact caused by conflicting results from multi-level recognition. First, we leverage cross-level attention to improve visual feature consistency between hierarchically cropped head, body and context windows. Then, vocabulary informed alignment is incorporated into the recognition framework to produce pseudo label and guide hierarchical visual feature refinement. To effectively fuse multi-level feature, we elaborate on a competition process of eliminating irrelevant image level predictions and a coordination process to enhance the feature across all levels. Extensive experiments are executed on two popular datasets, and our method achieves state-of-the-art performance with multi-level interpretation results. Code is available at: https://github.com/bisno/CocoER." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 567, + 135, + 578 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 567, + 135, + 578 + ], + "spans": [ + { + "bbox": [ + 55, + 567, + 135, + 578 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 586, + 296, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 586, + 296, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 586, + 296, + 696 + ], + "type": "text", + "content": "Visual emotion recognition plays a vital role in human computer interaction [27], surveillance [11], and robotics [5]. It aims to perceive the emotion state of human, and machine can modify responses and behaviors by considering the sentiment attitudes. Recently, impressive progress in recognizing emotions is achieved with expression from additional modality, such as head, body, background, depth, and so on [24]. Besides, emotion representation can be enhanced in view of discrete categories [24], points in a continuous" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 324, + 230, + 545, + 412 + ], + "blocks": [ + { + "bbox": [ + 324, + 230, + 545, + 412 + ], + "lines": [ + { + "bbox": [ + 324, + 230, + 545, + 412 + ], + "spans": [ + { + "bbox": [ + 324, + 230, + 545, + 412 + ], + "type": "image", + "image_path": "858bcb9d3548b96e064815fc3c03f48ebe0cad3e87f19d9a30aa07ab24d8ca60.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 422, + 555, + 468 + ], + "lines": [ + { + "bbox": [ + 313, + 422, + 555, + 468 + ], + "spans": [ + { + "bbox": [ + 313, + 422, + 555, + 468 + ], + "type": "text", + "content": "Figure 1. To avoid conflicts arising from inconsistent multi-level predictions, competition and coordination are applied in refining multi-level visual representation for emotion recognition, guided by pseudo label from vocabulary-informed alignment." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 482, + 486, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 482, + 486, + 494 + ], + "spans": [ + { + "bbox": [ + 313, + 482, + 486, + 494 + ], + "type": "text", + "content": "space [13, 42] and label distributions [41]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 312, + 496, + 555, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 496, + 555, + 664 + ], + "spans": [ + { + "bbox": [ + 312, + 496, + 555, + 664 + ], + "type": "text", + "content": "As emotion is naturally coupled in human and environments, the complex background in image still hinders the research of emotion recognition. Pioneering researches focus on perceiving the emotion states of the specific individuals in the given image with hierarchical visual feature [24, 37, 40] on context-aware emotion recognition datasets [13, 15]. [14] trains different backbones to combine the information of the target body and the entire image. Subsequently, a context-aware emotion recognition architecture enhances the classified precision to a high level representation [24]. In addition to image feature, knowledge graph is incorporated to reason the emotion states in [36]. To align feature across different modalities, [37] aims to address distribution gap between heterogeneous modalities." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "content": "Despite previous progress in decoupling irrelevant contextual information in the inference stage [38, 40], there still lacks of multi-level image feature alignment method to precisely recognize human emotions. Each level has dif" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 70, + 703, + 151, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 703, + 151, + 713 + ], + "spans": [ + { + "bbox": [ + 70, + 703, + 151, + 713 + ], + "type": "text", + "content": "* Corresponding authors." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "text", + "content": "29591" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 228 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 228 + ], + "type": "text", + "content": "ferent and potentially misleading representation and predictions, which decreases accuracy of emotion recognition. As shown in Fig. 1, multi-level emotion output from the head (fear and doubt), and context (sad) confuse final emotion prediction. We hypothesize that conflicting results from each level interfere with emotional decision-making process. To better align multi-level feature, researchers often draws upon prior knowledge, such as knowledge graph, to select relevant object. Although knowledge graph [36] is applied to enhance interpretability, extracting object relevance does not directly address the challenge of conflicted multi-level predictions and there is a lack of a guidance that can combine head, body, and background feature." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 229, + 295, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 229, + 295, + 443 + ], + "spans": [ + { + "bbox": [ + 56, + 229, + 295, + 443 + ], + "type": "text", + "content": "In this work, we propose competition and coordination mechanism to effectively perform multi-level fusion of head, body, and context emotion representation. Specifically, we apply cross-level attention mechanism with linear mapping on extracted feature maps to extract correlation from multi-level windows and obtain multi-level embeddings and recognition results. Furthermore, we leverage prototype learning to enhance the recognition result from external image-text aligned model, leading to pseudo label as guidance for multi-level feature refinement. Next, based on the pseudo label, we introduce competition process as an emotional feature selection method that filters out inconsistent multi-level feature. We leverage coordination process to update the parameters of the linear mapping for all image levels in inference-stage. At last, we re-send the multi-level feature to the updated linear mapping and achieve emotional consistency for final emotion recognition. In summary, the contributions of our paper are:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 445, + 295, + 577 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 55, + 445, + 295, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 445, + 295, + 493 + ], + "spans": [ + { + "bbox": [ + 55, + 445, + 295, + 493 + ], + "type": "text", + "content": "- We enhance multi-level image correlation via hierarchical feature extraction module, and produce multi-level recognition. We leverage prediction from vocabulary-informed module as pseudo label to identify inconsistent levels." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 493, + 295, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 493, + 295, + 528 + ], + "spans": [ + { + "bbox": [ + 56, + 493, + 295, + 528 + ], + "type": "text", + "content": "- We introduce workspace module that incorporates competition and coordination process to rectify multi-level feature through updated model in inference-stage." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 529, + 295, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 529, + 295, + 577 + ], + "spans": [ + { + "bbox": [ + 56, + 529, + 295, + 577 + ], + "type": "text", + "content": "- Extensive experiments have been conducted to compare with baseline methods and large vision-language models (VLM), and the outcomes demonstrate superior performance and interpretability of our proposed method." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 590, + 142, + 602 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 590, + 142, + 602 + ], + "spans": [ + { + "bbox": [ + 55, + 590, + 142, + 602 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 611, + 209, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 611, + 209, + 624 + ], + "spans": [ + { + "bbox": [ + 55, + 611, + 209, + 624 + ], + "type": "text", + "content": "2.1. Visual Emotion Recognition" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 629, + 294, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 294, + 713 + ], + "type": "text", + "content": "Visual emotion recognition is a comprehensive task in computer vision area, which is deeply associated with cognitive intelligence, and has attracted increasing attentions from researchers. Extensive works utilize facial expression to deduce the emotion state of specific person [7, 23, 25, 32, 35, 38]. Considering that recognizing a person's emotions can be subjective and has raised inherent ambiguity among sim" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 553, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 216 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 216 + ], + "type": "text", + "content": "ilar emotion categories, some researches focus on the latent distribution on the candidate emotion label sets [32, 41] and incorporates models from cognitive computing into the recognition algorithms [7, 21, 24, 26, 41]. Regarding network structures, [3, 18, 39] present transformer-based architectures to capture contextual embeddings. [10, 34, 37] utilize multi-head attention to fuse multiple modalities. In the aspect of inference stage, [38, 40] use causal relationships to eliminate information in the background that is inconsistent with target emotions. In contrast, we focuses on emotion recognition via multi-level visual representation that aligned by vocabulary-informed pseudo label." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 223, + 499, + 236 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 223, + 499, + 236 + ], + "spans": [ + { + "bbox": [ + 313, + 223, + 499, + 236 + ], + "type": "text", + "content": "2.2. Multi-modal Emotion Recognition" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 239, + 553, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 239, + 553, + 395 + ], + "spans": [ + { + "bbox": [ + 313, + 239, + 553, + 395 + ], + "type": "text", + "content": "[22, 23, 26, 28, 40] enhance multi-modal fusion performance, and [12, 14, 24, 30] explore context-based visual expressions. To fully take advantages of various context information, multi-modal visual feature are incorporated into model inputs, such as face, pose and depth [24, 40]. Different modality information requires different sensors or off-the-shelf extractor, resulting in increasing model complexity and a narrowed application range. On the other hand, [14, 30] focus on addressing this problem with only global and body information, and are still able to achieve competitive performance. [36] introduces a hierarchical context-based emotion recognition method that uses scene graphs to infer emotions from images." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 396, + 553, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 396, + 553, + 515 + ], + "spans": [ + { + "bbox": [ + 313, + 396, + 553, + 515 + ], + "type": "text", + "content": "In the field of large vision language model [17, 20], despite competitive performance in many multi-modal tasks, existing models do not prioritize emotion recognition. The modality alignment methods for large vision language models mostly focus on tokenized pair in the whole image, neglecting the multi-level image feature that misleads the image-text alignment. In this paper, we target this problem by eliminating irrelevant level of feature and further propose a novel multi-level feature alignment framework to enhance the performance of emotion recognition." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 525, + 374, + 537 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 525, + 374, + 537 + ], + "spans": [ + { + "bbox": [ + 313, + 525, + 374, + 537 + ], + "type": "text", + "content": "3. Methods" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 545, + 553, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 553, + 654 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 553, + 654 + ], + "type": "text", + "content": "In this section, we first introduce the hierarchical feature extraction module on the cropped multi-level image, where each image level has its own emotion recognition outcome. Next, we elaborate on the vocabulary-informed alignment module, which serves as guidance that provides pseudo label for eliminating inconsistent multi-level predictions. Finally, based on such pseudo label, we apply the workspace module to enhance multi-level feature fusion. The proposed framework is depicted in Fig. 2." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 660, + 487, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 660, + 487, + 672 + ], + "spans": [ + { + "bbox": [ + 313, + 660, + 487, + 672 + ], + "type": "text", + "content": "3.1. Hierarchical Feature Extraction" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "content": "We construct hierarchical emotion recognition by dividing the input image into three windows: head, body, and context. Specifically, given an input image " + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{I} \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "content": ", we" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "29592" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 97, + 71, + 318, + 232 + ], + "blocks": [ + { + "bbox": [ + 97, + 71, + 318, + 232 + ], + "lines": [ + { + "bbox": [ + 97, + 71, + 318, + 232 + ], + "spans": [ + { + "bbox": [ + 97, + 71, + 318, + 232 + ], + "type": "image", + "image_path": "6bb02efdd310c837e29ea24d7a49f7d5f308150aab87a3ccb8d701ffbc66a8b9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 320, + 72, + 516, + 232 + ], + "blocks": [ + { + "bbox": [ + 320, + 72, + 516, + 232 + ], + "lines": [ + { + "bbox": [ + 320, + 72, + 516, + 232 + ], + "spans": [ + { + "bbox": [ + 320, + 72, + 516, + 232 + ], + "type": "image", + "image_path": "90889120229087d16b4b638f4e5a4e755c842f7e65f248648edf292a5ebde9be.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 96, + 233, + 511, + 393 + ], + "blocks": [ + { + "bbox": [ + 96, + 233, + 511, + 393 + ], + "lines": [ + { + "bbox": [ + 96, + 233, + 511, + 393 + ], + "spans": [ + { + "bbox": [ + 96, + 233, + 511, + 393 + ], + "type": "image", + "image_path": "ce603e5c4c4e0963bfcec4b26747f4d3ac12ed83ab52f29cb8223ace58c7fa91.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 213, + 404, + 394, + 416 + ], + "lines": [ + { + "bbox": [ + 213, + 404, + 394, + 416 + ], + "spans": [ + { + "bbox": [ + 213, + 404, + 394, + 416 + ], + "type": "text", + "content": "Figure 2. Illustration of the proposed framework." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 426, + 296, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 426, + 296, + 486 + ], + "spans": [ + { + "bbox": [ + 55, + 426, + 296, + 486 + ], + "type": "text", + "content": "leverage object detection model [4] to obtain head, body, and context windows using detected bounding boxes. All three cropped windows are resized to " + }, + { + "bbox": [ + 55, + 426, + 296, + 486 + ], + "type": "inline_equation", + "content": "(H,W)" + }, + { + "bbox": [ + 55, + 426, + 296, + 486 + ], + "type": "text", + "content": ", and are then fed into trainable backbones " + }, + { + "bbox": [ + 55, + 426, + 296, + 486 + ], + "type": "inline_equation", + "content": "\\mathcal{R}" + }, + { + "bbox": [ + 55, + 426, + 296, + 486 + ], + "type": "text", + "content": " to extract multi-level feature maps " + }, + { + "bbox": [ + 55, + 426, + 296, + 486 + ], + "type": "inline_equation", + "content": "\\mathbf{u}_m" + }, + { + "bbox": [ + 55, + 426, + 296, + 486 + ], + "type": "text", + "content": " with a size of " + }, + { + "bbox": [ + 55, + 426, + 296, + 486 + ], + "type": "inline_equation", + "content": "(h,w)" + }, + { + "bbox": [ + 55, + 426, + 296, + 486 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 426, + 296, + 486 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 426, + 296, + 486 + ], + "type": "text", + "content": " channels:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 110, + 492, + 294, + 506 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 492, + 294, + 506 + ], + "spans": [ + { + "bbox": [ + 110, + 492, + 294, + 506 + ], + "type": "interline_equation", + "content": "\\mathbf {u} _ {m} = \\mathcal {R} _ {m} (\\mathbf {I} _ {m}), m \\in \\{\\mathbf {c}, \\mathbf {b}, \\mathbf {h} \\}, \\tag {1}", + "image_path": "eb81a9c4a31e5fe200626302ca4a0db6714adc0e3171a9e3bfcc103a9ca39b10.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 511, + 296, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 511, + 296, + 658 + ], + "spans": [ + { + "bbox": [ + 55, + 511, + 296, + 658 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 511, + 296, + 658 + ], + "type": "inline_equation", + "content": "\\mathbf{c},\\mathbf{b},\\mathbf{h}" + }, + { + "bbox": [ + 55, + 511, + 296, + 658 + ], + "type": "text", + "content": " are context, body, and head windows, respectively. Each " + }, + { + "bbox": [ + 55, + 511, + 296, + 658 + ], + "type": "inline_equation", + "content": "\\mathbf{u}_m\\in \\mathbb{R}^{h\\times w\\times c}" + }, + { + "bbox": [ + 55, + 511, + 296, + 658 + ], + "type": "text", + "content": " is divided and reshaped into " + }, + { + "bbox": [ + 55, + 511, + 296, + 658 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 511, + 296, + 658 + ], + "type": "text", + "content": " non-overlapping patches " + }, + { + "bbox": [ + 55, + 511, + 296, + 658 + ], + "type": "inline_equation", + "content": "\\mathbf{u}_m^p\\in \\mathbb{R}^{p^2\\cdot c}" + }, + { + "bbox": [ + 55, + 511, + 296, + 658 + ], + "type": "text", + "content": " and then mapped into " + }, + { + "bbox": [ + 55, + 511, + 296, + 658 + ], + "type": "inline_equation", + "content": "d_h" + }, + { + "bbox": [ + 55, + 511, + 296, + 658 + ], + "type": "text", + "content": " dimensions with linear projection " + }, + { + "bbox": [ + 55, + 511, + 296, + 658 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_m\\in \\mathbb{R}^{(p^2\\cdot c)\\times d_h}" + }, + { + "bbox": [ + 55, + 511, + 296, + 658 + ], + "type": "text", + "content": ". Thus each level " + }, + { + "bbox": [ + 55, + 511, + 296, + 658 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 55, + 511, + 296, + 658 + ], + "type": "text", + "content": " consists of a sequence of 2D embeddings " + }, + { + "bbox": [ + 55, + 511, + 296, + 658 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{u}}_m = [\\mathbf{u}_m^{p,0}\\mathbf{E}_m;\\mathbf{u}_m^{p,1}\\mathbf{E}_m;\\dots ;\\mathbf{u}_m^{p,N - 1}\\mathbf{E}_m]\\in \\mathbb{R}^{N\\times d_h}" + }, + { + "bbox": [ + 55, + 511, + 296, + 658 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 511, + 296, + 658 + ], + "type": "inline_equation", + "content": "N = hw / p^2" + }, + { + "bbox": [ + 55, + 511, + 296, + 658 + ], + "type": "text", + "content": " is patch number [6]. " + }, + { + "bbox": [ + 55, + 511, + 296, + 658 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{u}}_m" + }, + { + "bbox": [ + 55, + 511, + 296, + 658 + ], + "type": "text", + "content": " is sent to cross-level attention " + }, + { + "bbox": [ + 55, + 511, + 296, + 658 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 55, + 511, + 296, + 658 + ], + "type": "text", + "content": " for enhancing multi-level correlation via connecting local and context feature. For instance, the head-context attention " + }, + { + "bbox": [ + 55, + 511, + 296, + 658 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{\\mathbf{h},\\mathbf{c}}" + }, + { + "bbox": [ + 55, + 511, + 296, + 658 + ], + "type": "text", + "content": " is formalized with linear projections " + }, + { + "bbox": [ + 55, + 511, + 296, + 658 + ], + "type": "inline_equation", + "content": "\\{W_{\\mathbf{h}}^{0},W_{\\mathbf{h}}^{1},W_{\\mathbf{c}}\\} \\in \\mathbb{R}^{d_h\\times d_h}" + }, + { + "bbox": [ + 55, + 511, + 296, + 658 + ], + "type": "text", + "content": " with feature maps " + }, + { + "bbox": [ + 55, + 511, + 296, + 658 + ], + "type": "inline_equation", + "content": "\\{\\bar{\\mathbf{u}}_{\\mathbf{h}},\\bar{\\mathbf{u}}_{\\mathbf{c}}\\}" + }, + { + "bbox": [ + 55, + 511, + 296, + 658 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 95, + 662, + 295, + 677 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 95, + 662, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 95, + 662, + 295, + 677 + ], + "type": "interline_equation", + "content": "\\mathbf {Q} = \\bar {\\mathbf {u}} _ {\\mathbf {h}} W _ {\\mathbf {h}} ^ {0}, \\mathbf {K} = \\bar {\\mathbf {u}} _ {\\mathbf {c}} W _ {\\mathbf {c}}, \\mathbf {V} = \\bar {\\mathbf {u}} _ {\\mathbf {h}} W _ {\\mathbf {h}} ^ {1} \\tag {2}", + "image_path": "33ff75ec5a01810bfe5d8f2f5bf5aa7d7154b307623e2a5052d809d333ae99f4.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 96, + 679, + 295, + 694 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 679, + 295, + 694 + ], + "spans": [ + { + "bbox": [ + 96, + 679, + 295, + 694 + ], + "type": "interline_equation", + "content": "\\gamma = \\operatorname {s o f t m a x} \\left(\\mathbf {Q} \\mathbf {K} ^ {T}\\right) / \\sqrt {d _ {h}} \\quad \\in \\mathbb {R} ^ {N \\times N} \\tag {3}", + "image_path": "ab12a54c282667956e22fcd6f01fdb92268a2128eb0afddaf4fb52871a53bae2.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 93, + 696, + 295, + 710 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 696, + 295, + 710 + ], + "spans": [ + { + "bbox": [ + 93, + 696, + 295, + 710 + ], + "type": "interline_equation", + "content": "\\bar {\\mathbf {z}} _ {\\mathbf {h}} = \\boldsymbol {\\gamma} \\mathbf {V} \\in \\mathbb {R} ^ {N \\times d _ {h}}. \\tag {4}", + "image_path": "e251649dc5943c0af383323762ba09029736e07b0ae1248492c0d2ae0a22d11b.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 426, + 555, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 555, + 450 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 555, + 450 + ], + "type": "text", + "content": "Then " + }, + { + "bbox": [ + 313, + 426, + 555, + 450 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{z}}_{\\mathbf{h}}" + }, + { + "bbox": [ + 313, + 426, + 555, + 450 + ], + "type": "text", + "content": " is added with " + }, + { + "bbox": [ + 313, + 426, + 555, + 450 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{u}}_{\\mathbf{h}}" + }, + { + "bbox": [ + 313, + 426, + 555, + 450 + ], + "type": "text", + "content": " and sent to typical fully-connected layers with skip-connection:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 364, + 459, + 553, + 473 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 364, + 459, + 553, + 473 + ], + "spans": [ + { + "bbox": [ + 364, + 459, + 553, + 473 + ], + "type": "interline_equation", + "content": "\\dot {\\mathbf {z}} _ {\\mathbf {h}} = \\bar {\\mathbf {z}} _ {\\mathbf {h}} + \\bar {\\mathbf {u}} _ {\\mathbf {h}} \\in \\mathbb {R} ^ {N \\times d _ {h}} \\tag {5}", + "image_path": "35d596af037d938288f9c08d9374801ca2dd130c8300e4e6164d362a029036ac.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 364, + 476, + 553, + 491 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 364, + 476, + 553, + 491 + ], + "spans": [ + { + "bbox": [ + 364, + 476, + 553, + 491 + ], + "type": "interline_equation", + "content": "\\mathbf {z} _ {\\mathbf {h}} = \\operatorname {F F N} _ {0} (\\operatorname {L N} (\\dot {\\mathbf {z}} _ {\\mathbf {h}})) + \\dot {\\mathbf {z}} _ {\\mathbf {h}} \\in \\mathbb {R} ^ {d}, \\tag {6}", + "image_path": "41a650f0f08bf1f149354a76e3fe222660c080ce269745350e1d322061abdd4b.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 500, + 555, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 500, + 555, + 632 + ], + "spans": [ + { + "bbox": [ + 313, + 500, + 555, + 632 + ], + "type": "text", + "content": "where LN is LayerNorm, " + }, + { + "bbox": [ + 313, + 500, + 555, + 632 + ], + "type": "inline_equation", + "content": "\\mathrm{FFN}_0" + }, + { + "bbox": [ + 313, + 500, + 555, + 632 + ], + "type": "text", + "content": " refers to the feed-forward network and " + }, + { + "bbox": [ + 313, + 500, + 555, + 632 + ], + "type": "inline_equation", + "content": "\\mathbf{z_h} \\in \\mathbb{R}^d" + }, + { + "bbox": [ + 313, + 500, + 555, + 632 + ], + "type": "text", + "content": " is the head-level representation. Thus we bridge hierarchical information to enhance head-level representation " + }, + { + "bbox": [ + 313, + 500, + 555, + 632 + ], + "type": "inline_equation", + "content": "\\mathbf{z_h}" + }, + { + "bbox": [ + 313, + 500, + 555, + 632 + ], + "type": "text", + "content": " via head-context cross attention " + }, + { + "bbox": [ + 313, + 500, + 555, + 632 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{\\mathbf{h},\\mathbf{c}}" + }, + { + "bbox": [ + 313, + 500, + 555, + 632 + ], + "type": "text", + "content": ". Similarly we have " + }, + { + "bbox": [ + 313, + 500, + 555, + 632 + ], + "type": "inline_equation", + "content": "\\mathbf{z_b}" + }, + { + "bbox": [ + 313, + 500, + 555, + 632 + ], + "type": "text", + "content": " via body-context cross attention " + }, + { + "bbox": [ + 313, + 500, + 555, + 632 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{\\mathbf{b},\\mathbf{c}}" + }, + { + "bbox": [ + 313, + 500, + 555, + 632 + ], + "type": "text", + "content": ". Therefore we leverage attention mechanism to enrich local feature representation " + }, + { + "bbox": [ + 313, + 500, + 555, + 632 + ], + "type": "inline_equation", + "content": "\\{\\bar{\\mathbf{u}}_{\\mathbf{h}},\\bar{\\mathbf{u}}_{\\mathbf{c}}\\}" + }, + { + "bbox": [ + 313, + 500, + 555, + 632 + ], + "type": "text", + "content": ". Note that we conduct self-attention " + }, + { + "bbox": [ + 313, + 500, + 555, + 632 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{\\mathbf{c},\\mathbf{c}}" + }, + { + "bbox": [ + 313, + 500, + 555, + 632 + ], + "type": "text", + "content": " and obtain context feature " + }, + { + "bbox": [ + 313, + 500, + 555, + 632 + ], + "type": "inline_equation", + "content": "\\mathbf{z_c}" + }, + { + "bbox": [ + 313, + 500, + 555, + 632 + ], + "type": "text", + "content": ". Next, we pass each level representation " + }, + { + "bbox": [ + 313, + 500, + 555, + 632 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_m" + }, + { + "bbox": [ + 313, + 500, + 555, + 632 + ], + "type": "text", + "content": " through a linear mapping " + }, + { + "bbox": [ + 313, + 500, + 555, + 632 + ], + "type": "inline_equation", + "content": "g_{\\theta_m}" + }, + { + "bbox": [ + 313, + 500, + 555, + 632 + ], + "type": "text", + "content": ", followed by an operation " + }, + { + "bbox": [ + 313, + 500, + 555, + 632 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 313, + 500, + 555, + 632 + ], + "type": "text", + "content": " to produce " + }, + { + "bbox": [ + 313, + 500, + 555, + 632 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 500, + 555, + 632 + ], + "type": "text", + "content": " emotion category corresponding to multi-class task or multi-label task:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 361, + 642, + 553, + 656 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 642, + 553, + 656 + ], + "spans": [ + { + "bbox": [ + 361, + 642, + 553, + 656 + ], + "type": "interline_equation", + "content": "y _ {m} = \\mathcal {F} \\left(g _ {\\boldsymbol {\\theta} _ {m}} \\left(\\mathbf {z} _ {m}\\right)\\right), m \\in \\{\\mathbf {c}, \\mathbf {b}, \\mathbf {h} \\}, \\tag {7}", + "image_path": "5c817287d9911c6f1d7af4a30816f8a9558c3694dd79e73eb6e3d21c60c5ed82.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 666, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 666, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 666, + 555, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 666, + 555, + 714 + ], + "type": "inline_equation", + "content": "y_{m}" + }, + { + "bbox": [ + 313, + 666, + 555, + 714 + ], + "type": "text", + "content": " represents the prediction of each level " + }, + { + "bbox": [ + 313, + 666, + 555, + 714 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 666, + 555, + 714 + ], + "type": "text", + "content": ". The parameter of linear mapping " + }, + { + "bbox": [ + 313, + 666, + 555, + 714 + ], + "type": "inline_equation", + "content": "g_{\\theta_m}" + }, + { + "bbox": [ + 313, + 666, + 555, + 714 + ], + "type": "text", + "content": " will be updated in inference-stage, and each level representation " + }, + { + "bbox": [ + 313, + 666, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_m" + }, + { + "bbox": [ + 313, + 666, + 555, + 714 + ], + "type": "text", + "content": " will be refined for enhancing emotion recognition." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "29593" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 232, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 232, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 232, + 85 + ], + "type": "text", + "content": "3.2. Vocabulary-Informed Alignment" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 89, + 296, + 244 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 89, + 296, + 244 + ], + "spans": [ + { + "bbox": [ + 55, + 89, + 296, + 244 + ], + "type": "text", + "content": "In order to semantically guide feature refinement from hierarchical feature extraction, we propose a vocabulary-informed alignment method that leverages pre-trained image and text encoder to gain pseudo label and calibrate image representation. Specifically, given same input image " + }, + { + "bbox": [ + 55, + 89, + 296, + 244 + ], + "type": "inline_equation", + "content": "\\mathbf{I} \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 55, + 89, + 296, + 244 + ], + "type": "text", + "content": ", we improve CLIP [29] prediction to guide image feature alignment from text space. As shown in Fig. 2, we first encode the image through an CLIP image encoder " + }, + { + "bbox": [ + 55, + 89, + 296, + 244 + ], + "type": "inline_equation", + "content": "\\mathbf{Enc}_{\\mathrm{img}}" + }, + { + "bbox": [ + 55, + 89, + 296, + 244 + ], + "type": "text", + "content": ". Then, we provide a text prompt for " + }, + { + "bbox": [ + 55, + 89, + 296, + 244 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 89, + 296, + 244 + ], + "type": "text", + "content": " classes of emotion recognition task. The " + }, + { + "bbox": [ + 55, + 89, + 296, + 244 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 89, + 296, + 244 + ], + "type": "text", + "content": " classes are injected and expanded into " + }, + { + "bbox": [ + 55, + 89, + 296, + 244 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 89, + 296, + 244 + ], + "type": "text", + "content": " sentences of \"This photo shows {} emotion\", denoted as " + }, + { + "bbox": [ + 55, + 89, + 296, + 244 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_k" + }, + { + "bbox": [ + 55, + 89, + 296, + 244 + ], + "type": "text", + "content": ", and are sent to CLIP text encoder " + }, + { + "bbox": [ + 55, + 89, + 296, + 244 + ], + "type": "inline_equation", + "content": "\\mathbf{Enc}_{\\mathrm{text}}" + }, + { + "bbox": [ + 55, + 89, + 296, + 244 + ], + "type": "text", + "content": ". Thus, we have" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 116, + 251, + 294, + 264 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 251, + 294, + 264 + ], + "spans": [ + { + "bbox": [ + 116, + 251, + 294, + 264 + ], + "type": "interline_equation", + "content": "\\mathbf {u} _ {\\mathrm {i m g}} = \\mathbf {E n c} _ {\\mathrm {i m g}} (\\mathbf {I}) \\in \\mathbb {R} ^ {d} \\tag {8}", + "image_path": "975039342b3072cd3d3d8ea4d54f476ef767144661e4ca4fd66a629389861b20.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 116, + 266, + 294, + 281 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 266, + 294, + 281 + ], + "spans": [ + { + "bbox": [ + 116, + 266, + 294, + 281 + ], + "type": "interline_equation", + "content": "\\mathbf {U} _ {\\text {t e x t}} = \\mathbf {E n c} _ {\\text {t e x t}} \\left(\\mathbf {S} _ {k}\\right) \\in \\mathbb {R} ^ {d \\times k}. \\tag {9}", + "image_path": "3f25d6aa2f385d5b12fa6f2d7319d9a136c5e062f3748a3780849db6681ea1e9.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 287, + 296, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 287, + 296, + 432 + ], + "spans": [ + { + "bbox": [ + 55, + 287, + 296, + 432 + ], + "type": "text", + "content": "One can directly use dot-product similarity to obtain recognition result of CLIP. However, merely applying CLIP prediction is insufficient because it is trained on open-world image-text pairs, which leads to difficulties in differentiating precise expressions such as \"surprise\" and \"fear\". To improve the recognition performance in vocabulary-informed module, we apply prototype learning [8] that projects image representation into text representation space and pushes closer to emotion prototype. In detail, after extracting image feature, we expand " + }, + { + "bbox": [ + 55, + 287, + 296, + 432 + ], + "type": "inline_equation", + "content": "\\mathbf{u}_{\\mathrm{img}}" + }, + { + "bbox": [ + 55, + 287, + 296, + 432 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 55, + 287, + 296, + 432 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{U}}_{\\mathrm{img}}" + }, + { + "bbox": [ + 55, + 287, + 296, + 432 + ], + "type": "text", + "content": " which has the same shape of " + }, + { + "bbox": [ + 55, + 287, + 296, + 432 + ], + "type": "inline_equation", + "content": "\\mathbf{U}_{\\mathrm{text}}" + }, + { + "bbox": [ + 55, + 287, + 296, + 432 + ], + "type": "text", + "content": " and aim to minimize the distance " + }, + { + "bbox": [ + 55, + 287, + 296, + 432 + ], + "type": "inline_equation", + "content": "\\mathcal{D}(\\hat{\\mathbf{U}}_{\\mathrm{img}},\\mathbf{U}_{\\mathrm{text}})" + }, + { + "bbox": [ + 55, + 287, + 296, + 432 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 96, + 437, + 294, + 453 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 437, + 294, + 453 + ], + "spans": [ + { + "bbox": [ + 96, + 437, + 294, + 453 + ], + "type": "interline_equation", + "content": "\\mathcal {D} \\left(\\hat {\\mathbf {U}} _ {\\mathrm {i m g}}, \\mathbf {U} _ {\\text {t e x t}}\\right) = \\left\\| \\omega^ {T} \\hat {\\mathbf {U}} _ {\\mathrm {i m g}} - \\mathbf {U} _ {\\text {t e x t}} \\right\\| _ {2} ^ {2}, \\tag {10}", + "image_path": "ffcc1613599370aeac8a3c270cc49ee97a6fa53848a26cef7b8f476df3d1569f.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 459, + 296, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 459, + 296, + 555 + ], + "spans": [ + { + "bbox": [ + 55, + 459, + 296, + 555 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 459, + 296, + 555 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{U}}_{\\mathrm{img}} \\in \\mathbb{R}^{d \\times k}" + }, + { + "bbox": [ + 55, + 459, + 296, + 555 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 459, + 296, + 555 + ], + "type": "inline_equation", + "content": "\\omega \\in \\mathbb{R}^{d \\times d}" + }, + { + "bbox": [ + 55, + 459, + 296, + 555 + ], + "type": "text", + "content": " is a linear projection which is trained by minimizing the above distance of image feature and semantic emotion prototypes. Upon obtaining the trained " + }, + { + "bbox": [ + 55, + 459, + 296, + 555 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 55, + 459, + 296, + 555 + ], + "type": "text", + "content": ", we feed the emotion-aligned image feature " + }, + { + "bbox": [ + 55, + 459, + 296, + 555 + ], + "type": "inline_equation", + "content": "\\omega^T \\hat{\\mathbf{U}}_{\\mathrm{img}}" + }, + { + "bbox": [ + 55, + 459, + 296, + 555 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 459, + 296, + 555 + ], + "type": "inline_equation", + "content": "\\mathrm{FFN}_1" + }, + { + "bbox": [ + 55, + 459, + 296, + 555 + ], + "type": "text", + "content": ". Similar to (7) in the previous module that each hierarchical window has its own recognition result, the vocabulary-informed module also provides emotion recognition result:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 121, + 561, + 294, + 576 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 561, + 294, + 576 + ], + "spans": [ + { + "bbox": [ + 121, + 561, + 294, + 576 + ], + "type": "interline_equation", + "content": "y _ {\\mathbf {t}} = \\mathcal {F} \\left(\\operatorname {F F N} _ {1} \\left(\\omega^ {T} \\hat {\\mathbf {U}} _ {\\mathrm {i m g}}\\right)\\right). \\tag {11}", + "image_path": "aa641582c23e0e9caed26d5bc874544e9e93dc3ca1a11c37b527bea4a59110f7.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 582, + 296, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 296, + 666 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 296, + 666 + ], + "type": "inline_equation", + "content": "y_{\\mathbf{t}}" + }, + { + "bbox": [ + 55, + 582, + 296, + 666 + ], + "type": "text", + "content": " serves as pseudo label for further feature refinement in workspace. We use the embeddings of last hidden layer of " + }, + { + "bbox": [ + 55, + 582, + 296, + 666 + ], + "type": "inline_equation", + "content": "\\mathrm{FFN}_1" + }, + { + "bbox": [ + 55, + 582, + 296, + 666 + ], + "type": "text", + "content": " as the vocabulary-informed feature " + }, + { + "bbox": [ + 55, + 582, + 296, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{\\mathbf{t}}" + }, + { + "bbox": [ + 55, + 582, + 296, + 666 + ], + "type": "text", + "content": ", which has same feature dimension of image level feature " + }, + { + "bbox": [ + 55, + 582, + 296, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{z} \\in \\mathbb{R}^d" + }, + { + "bbox": [ + 55, + 582, + 296, + 666 + ], + "type": "text", + "content": ". Note that we freeze parameters of pre-trained " + }, + { + "bbox": [ + 55, + 582, + 296, + 666 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 55, + 582, + 296, + 666 + ], + "type": "text", + "content": ", as well as " + }, + { + "bbox": [ + 55, + 582, + 296, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{Enc}_{\\mathrm{img}}" + }, + { + "bbox": [ + 55, + 582, + 296, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 582, + 296, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{Enc}_{\\mathrm{text}}" + }, + { + "bbox": [ + 55, + 582, + 296, + 666 + ], + "type": "text", + "content": " during the training stage of our framework, leaving " + }, + { + "bbox": [ + 55, + 582, + 296, + 666 + ], + "type": "inline_equation", + "content": "\\mathrm{FFN}_1" + }, + { + "bbox": [ + 55, + 582, + 296, + 666 + ], + "type": "text", + "content": " trainable." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 672, + 130, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 672, + 130, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 672, + 130, + 685 + ], + "type": "text", + "content": "3.3. Workspace" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "Different from the previous feature refinement methods, such as the entity-context feature alignment [36] or emotion" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 322, + 70, + 547, + 220 + ], + "blocks": [ + { + "bbox": [ + 322, + 70, + 547, + 220 + ], + "lines": [ + { + "bbox": [ + 322, + 70, + 547, + 220 + ], + "spans": [ + { + "bbox": [ + 322, + 70, + 547, + 220 + ], + "type": "image", + "image_path": "0027eb2ef9440995765100063b19b441b345f207b99a5519d7af155cf2992193.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 230, + 554, + 253 + ], + "lines": [ + { + "bbox": [ + 313, + 230, + 554, + 253 + ], + "spans": [ + { + "bbox": [ + 313, + 230, + 554, + 253 + ], + "type": "text", + "content": "Figure 3. Detailed illustration of competition and coordination process, incorporated with inference-stage parameter update." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 260, + 554, + 343 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 260, + 554, + 343 + ], + "spans": [ + { + "bbox": [ + 313, + 260, + 554, + 343 + ], + "type": "text", + "content": "feature disentanglement [16, 37], we introduce a workspace module that can adaptively refine all levels of emotion feature based on the vocabulary-informed classification results, as illustrated in the bottom part of Fig. 2. The name \"workspace\" is inspired by the Workspace Theory [1, 2], which describes how multi-level predictions compete and coordinate for final recognition." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 344, + 555, + 451 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 344, + 555, + 451 + ], + "spans": [ + { + "bbox": [ + 313, + 344, + 555, + 451 + ], + "type": "text", + "content": "Formally, given the multi-level emotion feature set " + }, + { + "bbox": [ + 313, + 344, + 555, + 451 + ], + "type": "inline_equation", + "content": "\\mathbf{Z} = \\{\\mathbf{z}_{\\mathrm{c}},\\mathbf{z}_{\\mathrm{b}},\\mathbf{z}_{\\mathrm{h}}\\}" + }, + { + "bbox": [ + 313, + 344, + 555, + 451 + ], + "type": "text", + "content": ", we aim to gradually eliminate inconsistent level of feature in " + }, + { + "bbox": [ + 313, + 344, + 555, + 451 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}" + }, + { + "bbox": [ + 313, + 344, + 555, + 451 + ], + "type": "text", + "content": " according to the pseudo label " + }, + { + "bbox": [ + 313, + 344, + 555, + 451 + ], + "type": "inline_equation", + "content": "y_{t}" + }, + { + "bbox": [ + 313, + 344, + 555, + 451 + ], + "type": "text", + "content": " (11) until there is only one element in " + }, + { + "bbox": [ + 313, + 344, + 555, + 451 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}" + }, + { + "bbox": [ + 313, + 344, + 555, + 451 + ], + "type": "text", + "content": ". We refer to this process as multi-level feature competition. Subsequently, we leverage the selection process to update parameters of " + }, + { + "bbox": [ + 313, + 344, + 555, + 451 + ], + "type": "inline_equation", + "content": "g_{\\theta_m}" + }, + { + "bbox": [ + 313, + 344, + 555, + 451 + ], + "type": "text", + "content": " and re-send each " + }, + { + "bbox": [ + 313, + 344, + 555, + 451 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_m" + }, + { + "bbox": [ + 313, + 344, + 555, + 451 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 344, + 555, + 451 + ], + "type": "inline_equation", + "content": "g_{\\theta_m}" + }, + { + "bbox": [ + 313, + 344, + 555, + 451 + ], + "type": "text", + "content": " to obtain the refined feature " + }, + { + "bbox": [ + 313, + 344, + 555, + 451 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{z}}_m" + }, + { + "bbox": [ + 313, + 344, + 555, + 451 + ], + "type": "text", + "content": ". We call this multi-level feature coordination. Technical details are provided in the next two sections." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 456, + 430, + 468 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 456, + 430, + 468 + ], + "spans": [ + { + "bbox": [ + 313, + 456, + 430, + 468 + ], + "type": "text", + "content": "3.3.1. Competition Process" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 471, + 554, + 530 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 471, + 554, + 530 + ], + "spans": [ + { + "bbox": [ + 313, + 471, + 554, + 530 + ], + "type": "text", + "content": "The competition process is designed as an iteration and contains two steps during each iteration: 1) Determine the descent direction for updating parameters of " + }, + { + "bbox": [ + 313, + 471, + 554, + 530 + ], + "type": "inline_equation", + "content": "g_{\\theta_m}" + }, + { + "bbox": [ + 313, + 471, + 554, + 530 + ], + "type": "text", + "content": "; 2) Eliminate inconsistent element in feature set " + }, + { + "bbox": [ + 313, + 471, + 554, + 530 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}" + }, + { + "bbox": [ + 313, + 471, + 554, + 530 + ], + "type": "text", + "content": ". The iteration ends when there is only one element in " + }, + { + "bbox": [ + 313, + 471, + 554, + 530 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}" + }, + { + "bbox": [ + 313, + 471, + 554, + 530 + ], + "type": "text", + "content": ", shown in Fig. 3." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 531, + 555, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 531, + 555, + 602 + ], + "spans": [ + { + "bbox": [ + 313, + 531, + 555, + 602 + ], + "type": "text", + "content": "Step 1. We leverage the value of loss function " + }, + { + "bbox": [ + 313, + 531, + 555, + 602 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{hier}}(\\mathcal{F}(g_{\\theta}(\\mathbf{z})),y_{\\mathbf{t}})" + }, + { + "bbox": [ + 313, + 531, + 555, + 602 + ], + "type": "text", + "content": " as a measure of inconsistency between multi-level recognition and pseudo label. Then, in order to target the most inconsistent level, we take the maximum of the loss function groups for the results from " + }, + { + "bbox": [ + 313, + 531, + 555, + 602 + ], + "type": "inline_equation", + "content": "\\mathcal{F}(g_{\\theta_m}(\\mathbf{z}_m))" + }, + { + "bbox": [ + 313, + 531, + 555, + 602 + ], + "type": "text", + "content": " against pseudo label " + }, + { + "bbox": [ + 313, + 531, + 555, + 602 + ], + "type": "inline_equation", + "content": "y_{\\mathbf{t}}" + }, + { + "bbox": [ + 313, + 531, + 555, + 602 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 377, + 608, + 553, + 625 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 377, + 608, + 553, + 625 + ], + "spans": [ + { + "bbox": [ + 377, + 608, + 553, + 625 + ], + "type": "interline_equation", + "content": "\\max _ {\\mathbf {z} _ {m} \\in \\mathbf {Z}} \\left\\{\\mathcal {L} _ {\\text {h i e r}} \\left(g _ {\\boldsymbol {\\theta} _ {m}} \\left(\\mathbf {z} _ {m}\\right), y _ {\\mathbf {t}}\\right) \\right\\}. \\tag {12}", + "image_path": "93c828eb9da606c75ce7af26176a8e68a25b3baa16786c9e1027ffbce0514cba.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 632, + 554, + 692 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 632, + 554, + 692 + ], + "spans": [ + { + "bbox": [ + 313, + 632, + 554, + 692 + ], + "type": "text", + "content": "For the convenience of notation, let " + }, + { + "bbox": [ + 313, + 632, + 554, + 692 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{\\mathbf{z}_m}(\\pmb{\\theta}_m) = \\{\\mathcal{L}_{\\mathrm{hier}}(g_{\\pmb{\\theta}_m}(\\mathbf{z}_m),y_t),\\mathbf{z}_m\\in \\mathbf{Z}\\}" + }, + { + "bbox": [ + 313, + 632, + 554, + 692 + ], + "type": "text", + "content": ". During each iteration in the competition process, we aim to calculate the gradient of " + }, + { + "bbox": [ + 313, + 632, + 554, + 692 + ], + "type": "inline_equation", + "content": "\\max_{\\mathbf{z}_m\\in \\mathbf{Z}}\\mathbf{f}_{\\mathbf{z}_m}(\\pmb{\\theta}_m)" + }, + { + "bbox": [ + 313, + 632, + 554, + 692 + ], + "type": "text", + "content": " so as to update parameters of " + }, + { + "bbox": [ + 313, + 632, + 554, + 692 + ], + "type": "inline_equation", + "content": "g_{\\pmb{\\theta}_m}" + }, + { + "bbox": [ + 313, + 632, + 554, + 692 + ], + "type": "text", + "content": ". Thus, our goal transforms into a min-max problem:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 392, + 697, + 553, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 392, + 697, + 553, + 715 + ], + "spans": [ + { + "bbox": [ + 392, + 697, + 553, + 715 + ], + "type": "interline_equation", + "content": "\\min _ {\\boldsymbol {\\theta} _ {m}} \\max _ {\\mathbf {z} _ {m} \\in \\mathbf {Z}} \\mathbf {f} _ {\\mathbf {z} _ {m}} \\left(\\boldsymbol {\\theta} _ {m}\\right). \\tag {13}", + "image_path": "e57d28bee7c1af4b0094279c7f937c578827bb2333574277930487492e8802d1.jpg" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "29594" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "content": "However, there exists a non-trivial challenge in finding the descent direction when we implement " + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "inline_equation", + "content": "\\max(\\cdot)" + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "content": " on the groups of loss functions because the above function is nondifferentiable. Inspired by [33], we bring differentiability to " + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "inline_equation", + "content": "\\max_{\\mathbf{z}_m \\in \\mathbf{Z}} \\mathbf{f}_{\\mathbf{z}_m}(\\theta_m)" + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "content": " and leverage a regularization term to stabilize the descent direction. Please refer to technical details in Appendix (Sec.1). Then, the min-max problem is reduced to the following quadratic programming (QP) problem with a Lagrange multiplier " + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 111, + 190, + 296, + 240 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 190, + 296, + 240 + ], + "spans": [ + { + "bbox": [ + 111, + 190, + 296, + 240 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\min _ {\\boldsymbol {\\lambda}} \\left(\\frac {1}{2} \\boldsymbol {\\lambda} ^ {T} \\mathbf {G} \\mathbf {G} ^ {T} \\boldsymbol {\\lambda} - \\boldsymbol {\\lambda} ^ {T} \\mathbf {f} _ {\\mathbf {z} _ {m}}\\right) \\tag {14} \\\\ \\text {s . t .} \\quad \\sum_ {\\mathbf {z} _ {m} \\in \\mathbf {Z}} \\lambda_ {\\mathbf {z} _ {m}} = 1, \\lambda_ {\\mathbf {z} _ {m}} \\geq 0 \\\\ \\end{array}", + "image_path": "49a775baa79ed96513041b4c6a0395322d8bd3cb8963cd53501d577665cc9dd6.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 251, + 296, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 251, + 296, + 348 + ], + "spans": [ + { + "bbox": [ + 55, + 251, + 296, + 348 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 251, + 296, + 348 + ], + "type": "inline_equation", + "content": "\\mathbf{G} = \\nabla_{\\boldsymbol{\\theta}_m^l}\\mathbf{f}_{\\mathbf{z}_m} \\in \\mathbb{R}^{q \\times (d \\cdot k)}" + }, + { + "bbox": [ + 55, + 251, + 296, + 348 + ], + "type": "text", + "content": " contains gradient of the prediction layer " + }, + { + "bbox": [ + 55, + 251, + 296, + 348 + ], + "type": "inline_equation", + "content": "\\boldsymbol{\\theta}_m^l" + }, + { + "bbox": [ + 55, + 251, + 296, + 348 + ], + "type": "text", + "content": " with respect to the loss of " + }, + { + "bbox": [ + 55, + 251, + 296, + 348 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{\\mathbf{z}_m}" + }, + { + "bbox": [ + 55, + 251, + 296, + 348 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 251, + 296, + 348 + ], + "type": "inline_equation", + "content": "q = |\\mathbf{Z}|" + }, + { + "bbox": [ + 55, + 251, + 296, + 348 + ], + "type": "text", + "content": " refers to the number of elements left in " + }, + { + "bbox": [ + 55, + 251, + 296, + 348 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}" + }, + { + "bbox": [ + 55, + 251, + 296, + 348 + ], + "type": "text", + "content": ". After solving for " + }, + { + "bbox": [ + 55, + 251, + 296, + 348 + ], + "type": "inline_equation", + "content": "\\boldsymbol{\\lambda}" + }, + { + "bbox": [ + 55, + 251, + 296, + 348 + ], + "type": "text", + "content": " in the QP problem, we have " + }, + { + "bbox": [ + 55, + 251, + 296, + 348 + ], + "type": "inline_equation", + "content": "\\delta = -\\mathbf{G}^T\\boldsymbol{\\lambda} \\in \\mathbb{R}^{d \\cdot k}" + }, + { + "bbox": [ + 55, + 251, + 296, + 348 + ], + "type": "text", + "content": " as the approximate solution of (13). Here, " + }, + { + "bbox": [ + 55, + 251, + 296, + 348 + ], + "type": "inline_equation", + "content": "\\boldsymbol{\\delta}" + }, + { + "bbox": [ + 55, + 251, + 296, + 348 + ], + "type": "text", + "content": " serves as the shared parameter updating path for linear mapping " + }, + { + "bbox": [ + 55, + 251, + 296, + 348 + ], + "type": "inline_equation", + "content": "g_{\\boldsymbol{\\theta}_m}" + }, + { + "bbox": [ + 55, + 251, + 296, + 348 + ], + "type": "text", + "content": " of all three levels, which will be used for feature refinement in coordination process." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 350, + 296, + 398 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 350, + 296, + 398 + ], + "spans": [ + { + "bbox": [ + 55, + 350, + 296, + 398 + ], + "type": "text", + "content": "Step 2. Let " + }, + { + "bbox": [ + 55, + 350, + 296, + 398 + ], + "type": "inline_equation", + "content": "n = \\arg \\max_{\\mathbf{z}_m \\in \\mathbf{Z}} \\left\\{ \\mathcal{L}_{\\mathrm{hier}}(g_{\\theta_m}(\\mathbf{z}_m), y_t) \\right\\}" + }, + { + "bbox": [ + 55, + 350, + 296, + 398 + ], + "type": "text", + "content": " be the index of most inconsistent feature in " + }, + { + "bbox": [ + 55, + 350, + 296, + 398 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}" + }, + { + "bbox": [ + 55, + 350, + 296, + 398 + ], + "type": "text", + "content": ". We iteratively eliminate the corresponding " + }, + { + "bbox": [ + 55, + 350, + 296, + 398 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 350, + 296, + 398 + ], + "type": "text", + "content": " of feature " + }, + { + "bbox": [ + 55, + 350, + 296, + 398 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 55, + 350, + 296, + 398 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 55, + 350, + 296, + 398 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}" + }, + { + "bbox": [ + 55, + 350, + 296, + 398 + ], + "type": "text", + "content": ", denoted as " + }, + { + "bbox": [ + 55, + 350, + 296, + 398 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}_{-n}" + }, + { + "bbox": [ + 55, + 350, + 296, + 398 + ], + "type": "text", + "content": ". This elimination is illustrated in the top part of Fig. 3." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 405, + 176, + 415 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 405, + 176, + 415 + ], + "spans": [ + { + "bbox": [ + 55, + 405, + 176, + 415 + ], + "type": "text", + "content": "3.3.2. Coordination Process" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 420, + 296, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 420, + 296, + 529 + ], + "spans": [ + { + "bbox": [ + 55, + 420, + 296, + 529 + ], + "type": "text", + "content": "The goal of coordination process is to rectify the multi-level feature " + }, + { + "bbox": [ + 55, + 420, + 296, + 529 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 55, + 420, + 296, + 529 + ], + "type": "text", + "content": " via updated linear mapping " + }, + { + "bbox": [ + 55, + 420, + 296, + 529 + ], + "type": "inline_equation", + "content": "g_{\\theta}" + }, + { + "bbox": [ + 55, + 420, + 296, + 529 + ], + "type": "text", + "content": ". Specifically, after iteratively obtaining the respective descent direction " + }, + { + "bbox": [ + 55, + 420, + 296, + 529 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 55, + 420, + 296, + 529 + ], + "type": "text", + "content": " during the competition process, we accumulate the overall descent direction, shown in the bottom part of Fig. 3. For example, if the context-level feature " + }, + { + "bbox": [ + 55, + 420, + 296, + 529 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{\\mathbf{c}}" + }, + { + "bbox": [ + 55, + 420, + 296, + 529 + ], + "type": "text", + "content": " and head-level feature " + }, + { + "bbox": [ + 55, + 420, + 296, + 529 + ], + "type": "inline_equation", + "content": "\\mathbf{z_h}" + }, + { + "bbox": [ + 55, + 420, + 296, + 529 + ], + "type": "text", + "content": " is gradually eliminated through competition, we have " + }, + { + "bbox": [ + 55, + 420, + 296, + 529 + ], + "type": "inline_equation", + "content": "\\delta_{\\mathbf{c}}" + }, + { + "bbox": [ + 55, + 420, + 296, + 529 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 420, + 296, + 529 + ], + "type": "inline_equation", + "content": "\\delta_{\\mathbf{h}}" + }, + { + "bbox": [ + 55, + 420, + 296, + 529 + ], + "type": "text", + "content": " based on (14), which construct the shared parameter updating path for the prediction layer " + }, + { + "bbox": [ + 55, + 420, + 296, + 529 + ], + "type": "inline_equation", + "content": "\\theta_m^l" + }, + { + "bbox": [ + 55, + 420, + 296, + 529 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 146, + 538, + 294, + 552 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 538, + 294, + 552 + ], + "spans": [ + { + "bbox": [ + 146, + 538, + 294, + 552 + ], + "type": "interline_equation", + "content": "\\delta^ {l} = \\delta_ {\\mathrm {c}} + \\delta_ {\\mathrm {h}}. \\tag {15}", + "image_path": "885f230d4311ae1e2f976ba7e9992e92731be5eeb37243b55fa3ddbe814183c8.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 562, + 296, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 562, + 296, + 586 + ], + "spans": [ + { + "bbox": [ + 55, + 562, + 296, + 586 + ], + "type": "text", + "content": "Then, for each image level " + }, + { + "bbox": [ + 55, + 562, + 296, + 586 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 55, + 562, + 296, + 586 + ], + "type": "text", + "content": " we update entire parameter " + }, + { + "bbox": [ + 55, + 562, + 296, + 586 + ], + "type": "inline_equation", + "content": "\\theta_{m}" + }, + { + "bbox": [ + 55, + 562, + 296, + 586 + ], + "type": "text", + "content": " via chain rule and back-propagation, denoted as " + }, + { + "bbox": [ + 55, + 562, + 296, + 586 + ], + "type": "inline_equation", + "content": "\\longleftarrow" + }, + { + "bbox": [ + 55, + 562, + 296, + 586 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 73, + 595, + 294, + 610 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 595, + 294, + 610 + ], + "spans": [ + { + "bbox": [ + 73, + 595, + 294, + 610 + ], + "type": "interline_equation", + "content": "\\tilde {\\theta} _ {\\mathbf {c}} \\leftarrow \\theta_ {\\mathbf {c}} ^ {l} + \\delta^ {l}; \\tilde {\\theta} _ {\\mathbf {b}} \\leftarrow \\theta_ {\\mathbf {b}} ^ {l} + \\delta^ {l}; \\tilde {\\theta} _ {\\mathbf {h}} \\leftarrow \\theta_ {\\mathbf {h}} ^ {l} + \\delta^ {l}. \\tag {16}", + "image_path": "2bd750be0e3e24b4c1d57bffd957b99afb805c4f840648e9b464cb7c7c9d0d27.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 620, + 296, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 620, + 296, + 657 + ], + "spans": [ + { + "bbox": [ + 55, + 620, + 296, + 657 + ], + "type": "text", + "content": "Next, we re-send " + }, + { + "bbox": [ + 55, + 620, + 296, + 657 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_m" + }, + { + "bbox": [ + 55, + 620, + 296, + 657 + ], + "type": "text", + "content": " to its corresponding linear mapping " + }, + { + "bbox": [ + 55, + 620, + 296, + 657 + ], + "type": "inline_equation", + "content": "g_{\\tilde{\\theta}}" + }, + { + "bbox": [ + 55, + 620, + 296, + 657 + ], + "type": "text", + "content": " and acquire refined feature from last hidden layer and construct multi-level concatenated representation " + }, + { + "bbox": [ + 55, + 620, + 296, + 657 + ], + "type": "inline_equation", + "content": "[\\hat{\\mathbf{z}}_{\\mathrm{c}};\\hat{\\mathbf{z}}_{\\mathrm{b}};\\hat{\\mathbf{z}}_{\\mathrm{h}}]" + }, + { + "bbox": [ + 55, + 620, + 296, + 657 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 88, + 670, + 294, + 685 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 670, + 294, + 685 + ], + "spans": [ + { + "bbox": [ + 88, + 670, + 294, + 685 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {z}} _ {\\mathrm {c}} = g _ {\\tilde {\\theta} _ {\\mathrm {c}}} \\left(\\mathbf {z} _ {\\mathrm {c}}\\right); \\hat {\\mathbf {z}} _ {\\mathrm {b}} = g _ {\\tilde {\\theta} _ {\\mathrm {b}}} \\left(\\mathbf {z} _ {\\mathrm {b}}\\right); \\hat {\\mathbf {z}} _ {\\mathrm {h}} = g _ {\\tilde {\\theta} _ {\\mathrm {h}}} \\left(\\mathbf {z} _ {\\mathrm {h}}\\right). \\tag {17}", + "image_path": "592e0661b12461d283433f8bdbe94d09379e932e2dcff7ed3c5de7c87107d5f7.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": "With such inference-stage parameter updating, we calibrate multi-level emotion representation, ensuring that the final" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "text", + "content": "emotion recognition results are less affected by the inconsistent emotion recognition outcomes from hierarchical feature. Please refer to algorithm in Appendix that describes competition and coordination process within workspace." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 127, + 430, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 127, + 430, + 140 + ], + "spans": [ + { + "bbox": [ + 313, + 127, + 430, + 140 + ], + "type": "text", + "content": "3.4. Training Objectives" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 144, + 555, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 144, + 555, + 192 + ], + "spans": [ + { + "bbox": [ + 313, + 144, + 555, + 192 + ], + "type": "text", + "content": "We concatenate the refined multi-level representation with vocabulary-informed feature " + }, + { + "bbox": [ + 313, + 144, + 555, + 192 + ], + "type": "inline_equation", + "content": "\\mathbf{z_t}" + }, + { + "bbox": [ + 313, + 144, + 555, + 192 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 313, + 144, + 555, + 192 + ], + "type": "inline_equation", + "content": "\\mathbf{z_G} = [\\hat{\\mathbf{z}}_{\\mathrm{c}};\\hat{\\mathbf{z}}_{\\mathrm{b}};\\hat{\\mathbf{z}}_{\\mathrm{h}};\\mathbf{z}_{\\mathrm{t}}]" + }, + { + "bbox": [ + 313, + 144, + 555, + 192 + ], + "type": "text", + "content": ". Then " + }, + { + "bbox": [ + 313, + 144, + 555, + 192 + ], + "type": "inline_equation", + "content": "\\mathbf{z_G}" + }, + { + "bbox": [ + 313, + 144, + 555, + 192 + ], + "type": "text", + "content": " is passed to feed-forward network, and output of the proposed method is:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 389, + 201, + 553, + 214 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 389, + 201, + 553, + 214 + ], + "spans": [ + { + "bbox": [ + 389, + 201, + 553, + 214 + ], + "type": "interline_equation", + "content": "y _ {\\mathbf {G}} = \\mathcal {F} \\left(\\operatorname {F F N} _ {2} \\left(\\mathbf {z} _ {\\mathbf {G}}\\right)\\right). \\tag {18}", + "image_path": "0d6542fc3b6af3be2dacd0c1528d10d4ea58a85fb06ae7f439f2ca0b047c8b63.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 222, + 554, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 222, + 554, + 258 + ], + "spans": [ + { + "bbox": [ + 313, + 222, + 554, + 258 + ], + "type": "text", + "content": "We collectively form loss functions from HFE, V-I and workspace module as " + }, + { + "bbox": [ + 313, + 222, + 554, + 258 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathbf{z}}(y_{\\mathbf{z}},y)" + }, + { + "bbox": [ + 313, + 222, + 554, + 258 + ], + "type": "text", + "content": " with respect to ground-truth " + }, + { + "bbox": [ + 313, + 222, + 554, + 258 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 313, + 222, + 554, + 258 + ], + "type": "text", + "content": ". The combined loss function " + }, + { + "bbox": [ + 313, + 222, + 554, + 258 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{emo}}" + }, + { + "bbox": [ + 313, + 222, + 554, + 258 + ], + "type": "text", + "content": " is defined as follows:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 363, + 267, + 553, + 280 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 363, + 267, + 553, + 280 + ], + "spans": [ + { + "bbox": [ + 363, + 267, + 553, + 280 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {e m o}} = \\mathcal {L} _ {\\mathbf {c}} + \\mathcal {L} _ {\\mathbf {b}} + \\mathcal {L} _ {\\mathbf {h}} + \\mathcal {L} _ {\\mathbf {t}} + \\mathcal {L} _ {\\mathbf {G}}. \\tag {19}", + "image_path": "0533b33e16b8be7240cc8205987ee42b30514054247979bbe054e59a3c2e944e.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 288, + 554, + 336 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 288, + 554, + 336 + ], + "spans": [ + { + "bbox": [ + 313, + 288, + 554, + 336 + ], + "type": "text", + "content": "Next, we add norm of " + }, + { + "bbox": [ + 313, + 288, + 554, + 336 + ], + "type": "inline_equation", + "content": "\\delta^l" + }, + { + "bbox": [ + 313, + 288, + 554, + 336 + ], + "type": "text", + "content": " [43] to regularize the shared updating path, aiming to avoid over-dependence on the vocabulary-informed pseudo label. Finally, all parameters of our proposed framework are trained based on:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 391, + 344, + 553, + 358 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 391, + 344, + 553, + 358 + ], + "spans": [ + { + "bbox": [ + 391, + 344, + 553, + 358 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {\\mathrm {e m o}} + \\alpha \\| \\boldsymbol {\\delta} ^ {l} \\| _ {2}, \\tag {20}", + "image_path": "77a81388ecb7f2e2101b0b9c3eeb64c3c4782c419b7f9414ca8aeed8430e342f.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 366, + 516, + 379 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 366, + 516, + 379 + ], + "spans": [ + { + "bbox": [ + 313, + 366, + 516, + 379 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 366, + 516, + 379 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 366, + 516, + 379 + ], + "type": "text", + "content": " is hyperparameter, empirically set to 0.1." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 389, + 395, + 402 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 389, + 395, + 402 + ], + "spans": [ + { + "bbox": [ + 313, + 389, + 395, + 402 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 408, + 364, + 421 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 408, + 364, + 421 + ], + "spans": [ + { + "bbox": [ + 313, + 408, + 364, + 421 + ], + "type": "text", + "content": "4.1. Setup" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 426, + 555, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 555, + 570 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 555, + 570 + ], + "type": "text", + "content": "Datasets. EMOTIC [12] comprises 23,571 images featuring 34,320 individuals, predominantly captured in unconstrained environments. Each image encompasses multiple emotions, presenting a more intricate challenge of multi-label emotion recognition across 26 discrete emotion categories. CAER-S [15] is derived by extracting static frames from video clips in the CAER dataset [15], containing over 70,000 images. These images are annotated with an expanded set of 7 emotion categories: anger, disgust, fear, happiness, sadness, surprise, and neutral. Both datasets are divided into training (80%) and testing (20%) subsets. Implementation Details. Please refer to Appendix." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 576, + 500, + 589 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 576, + 500, + 589 + ], + "spans": [ + { + "bbox": [ + 313, + 576, + 500, + 589 + ], + "type": "text", + "content": "4.2. Comparison with State-of-the-Arts" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "content": "EMOTIC test Split. We first compare our method with several multimodal or multi-object based emotion recognition competitors on EMOTIC test split. As shown in Tab. 1, we report the average precision (AP) over all categories as well as the average results (mAP). From the table, we can observe that: (1) Our method achieves state-of-the-art performance with " + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "inline_equation", + "content": "37.32\\%" + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "content": " on mAP. Compared with [9] which utilizes Faster R-CNN to detect all possible objects in images for reasoning, we surpass it by a margin of " + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "inline_equation", + "content": "2.16\\%" + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "content": " using only three hierarchical cropped windows. Moreover, [24]" + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "29595" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 70, + 553, + 406 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 553, + 406 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 553, + 406 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 553, + 406 + ], + "type": "table", + "html": "
CategoriesLi et al. [19] Hoang et al. [9] Mittal et al. [24]† Wu et al. [36]BLIP2-6.7b [17] LLaVa-1.6-7b [20] GPT-4o [27]
2021202120202024202220232024
Affection37.9344.4845.2332.2524.2534.3222.53
Anger13.7330.7115.4610.3708.3004.9810.32
Annoyance20.8726.4721.9212.8408.9912.3709.98
Anticipation61.0859.8972.1256.8391.3591.8754.16
Aversion09.6112.4317.8107.7006.4506.4603.51
Confidence80.0879.2468.6577.5159.3762.0568.09
Disapproval21.5424.5419.8212.4707.5610.3808.64
Disconnection28.3234.2443.1226.4125.8531.0517.03
Disquietment22.5718.7324.2316.7513.5314.4517.17
Doubt/Confusion33.5025.4235.1218.6218.5819.0631.20
Embarrassment04.1604.2614.3702.8005.4505.4402.17
Engagement88.1288.7191.1286.8395.1496.1080.77
Esteem20.5017.9923.6215.5023.1923.2718.30
Excitement80.1174.2183.2670.7364.5170.3669.59
Fatigue17.5122.6216.2312.6407.6811.0309.53
Fear15.5613.9223.6505.4807.8408.4912.79
Happiness76.0183.0274.7169.4570.7775.1453.65
Pain14.5616.6813.2110.6405.5205.8107.45
Peace26.7628.9134.2724.6320.3221.2628.91
Pleasure55.6455.4765.5344.3343.3043.4942.86
Sadness30.8042.8723.4121.2107.0618.7210.70
Sensitivity09.5915.8908.3205.8904.7704.7503.96
Suffering30.7046.2326.3921.4805.2206.1709.74
Surprise17.9216.2717.3707.9415.4216.1417.66
Sympathy15.2615.3734.2813.2421.4821.6209.10
Yearning10.1110.0414.2909.1213.0512.7907.93
mAP32.4135.1635.4826.6825.9627.9923.80
", + "image_path": "04a3462b63a13dfa5cc5435ac56004aea03aef50a1ef460993eb65dfa9f56081.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 59, + 449, + 200, + 555 + ], + "blocks": [ + { + "bbox": [ + 59, + 449, + 200, + 555 + ], + "lines": [ + { + "bbox": [ + 59, + 449, + 200, + 555 + ], + "spans": [ + { + "bbox": [ + 59, + 449, + 200, + 555 + ], + "type": "image", + "image_path": "fa441faf561ae2b29f4f136c87747defc0123a42728dbe558e456fbb62003f74.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 201, + 449, + 550, + 555 + ], + "blocks": [ + { + "bbox": [ + 55, + 415, + 555, + 437 + ], + "lines": [ + { + "bbox": [ + 55, + 415, + 555, + 437 + ], + "spans": [ + { + "bbox": [ + 55, + 415, + 555, + 437 + ], + "type": "text", + "content": "Table 1. Performance comparisons on EMOTIC test split. The 1st/2nd best results are highlighted in red/blue. †' denotes the method that contains additional visual modalities, such as depth, which is not used in our method." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 201, + 449, + 550, + 555 + ], + "lines": [ + { + "bbox": [ + 201, + 449, + 550, + 555 + ], + "spans": [ + { + "bbox": [ + 201, + 449, + 550, + 555 + ], + "type": "table", + "html": "
GTPredictionHeadBodyContextV-I (Pseudo label)
AffectionAffectionDisconnectionAffectionEngagementAffection
AnticipationConfidenceFatigueConfidenceExcitementConfidence
ConfidenceExcitementPleasureHappinessExcitement
EngagementHappinessPeaceHappiness
ExcitementPeacePleasure
HappinessPleasure
PeaceLhier=2.35Lhier=0.61Lhier=1.33
Pleasure
Exclusion Sequence: (1) Head→(2) Context→(3) Body
", + "image_path": "d63616a10c6b7c497bf675e7a05d0342f625edfe259fc558b8748042d8e47b59.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 59, + 561, + 200, + 666 + ], + "blocks": [ + { + "bbox": [ + 59, + 561, + 200, + 666 + ], + "lines": [ + { + "bbox": [ + 59, + 561, + 200, + 666 + ], + "spans": [ + { + "bbox": [ + 59, + 561, + 200, + 666 + ], + "type": "image", + "image_path": "7de892624f59c881d029e55340def0523c5da2e1b00bf7544a8c4d575e036438.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 675, + 555, + 698 + ], + "lines": [ + { + "bbox": [ + 55, + 675, + 555, + 698 + ], + "spans": [ + { + "bbox": [ + 55, + 675, + 555, + 698 + ], + "type": "text", + "content": "Figure 4. Case study of randomly sampled results on EMOTIC. The V-I column shows pseudo label in vocabulary-informed module. We display the elimination of inconsistent recognition results in orange, which are not shown in either the V-I column or the prediction column." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 201, + 561, + 550, + 666 + ], + "blocks": [ + { + "bbox": [ + 201, + 561, + 550, + 666 + ], + "lines": [ + { + "bbox": [ + 201, + 561, + 550, + 666 + ], + "spans": [ + { + "bbox": [ + 201, + 561, + 550, + 666 + ], + "type": "table", + "html": "
GTPredictionHeadBodyContextV-I (Pseudo label)
AnticipationAnticipationAnticipationAffectionEngagementAnticipation
ConfidenceConfidenceConfidenceSensitivityEmbarrassmentEngagement
EngagementEngagementExcitementSufferingPleasureEsteem
ExcitementExcitementHappinessHappiness
PeacePleasure
Lhier=1.42Lhier=2.36Lhier=1.03
Exclusion Sequence: (1) Body→(2) Context→(3) Head
", + "image_path": "3f8deb009b6cb91653a1787612cb5f76db36b6ed70b95b597a141a00ce137780.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "29596" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 61, + 70, + 291, + 148 + ], + "blocks": [ + { + "bbox": [ + 61, + 70, + 291, + 148 + ], + "lines": [ + { + "bbox": [ + 61, + 70, + 291, + 148 + ], + "spans": [ + { + "bbox": [ + 61, + 70, + 291, + 148 + ], + "type": "table", + "html": "
MethodsC-F1 (%)O-F1 (%)
Kosti et al. [14] (2019)29.3354.45
Li et al. [19] (2021)15.1048.07
Wu et al. [36] (2024)34.2454.78
Ours36.7757.23
", + "image_path": "003ef8adafd7ccda900b086d91734471a3e1626ea0fb95bfb229a83f72800d8c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 65, + 180, + 287, + 323 + ], + "blocks": [ + { + "bbox": [ + 56, + 156, + 295, + 178 + ], + "lines": [ + { + "bbox": [ + 56, + 156, + 295, + 178 + ], + "spans": [ + { + "bbox": [ + 56, + 156, + 295, + 178 + ], + "type": "text", + "content": "Table 2. Label-based F1 score (C-F1) and example-based F1 score (O-F1) on EMOTIC test split." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 65, + 180, + 287, + 323 + ], + "lines": [ + { + "bbox": [ + 65, + 180, + 287, + 323 + ], + "spans": [ + { + "bbox": [ + 65, + 180, + 287, + 323 + ], + "type": "table", + "html": "
MethodsAccuracy (%)
CLIP-ViT-L [29] (2021)21.57
BLIP2-6.7b [17] (2022)14.21
LLaVa-1.6-7b [20] (2023)28.59
GPT-4o [27] (2024)29.33
EfficientFace [45] (2021)85.87
MA-Net [44] (2021)88.42
Li et al. [19] (2021)84.42
Wu et al. [36] (2024)90.83
Ours91.86
", + "image_path": "346dc86a4cffd6893c55a6ed37da30fa12bdb64c1d4333012b7d728542de61d3.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 354, + 295, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 354, + 295, + 544 + ], + "spans": [ + { + "bbox": [ + 55, + 354, + 295, + 544 + ], + "type": "text", + "content": "additionally leverages the depth modality to model sociodynamic influence, but the mAP is lower than ours. What's more, there are 18 top-two categories (illustrated in red/blue color) of our methods, whereas [24] has 12 top-two categories. (2) For commonly used categories (anger, fear, happiness, sadness, surprise), [24] reports an averaged accuracy of " + }, + { + "bbox": [ + 55, + 354, + 295, + 544 + ], + "type": "inline_equation", + "content": "30.70\\%" + }, + { + "bbox": [ + 55, + 354, + 295, + 544 + ], + "type": "text", + "content": ". In contrast, our method achieves a higher mean result of " + }, + { + "bbox": [ + 55, + 354, + 295, + 544 + ], + "type": "inline_equation", + "content": "37.19\\%" + }, + { + "bbox": [ + 55, + 354, + 295, + 544 + ], + "type": "text", + "content": ". (3) In Tab. 2, we demonstrate that our method outperforms the state-of-the-art method [36] in both C-F1 and O-F1 scores. (4) Our approach also outperforms VLM methods (BLIP2, LLaVa, GPT-4o). This can be attributed to the lack of alignment of multi-level features in VLM for emotion recognition, suggesting that the alignment of tokenized image and text still produce confused results. Please refer to Appendix (Sec.10.1) for more detailed analysis of multi-level prediction confusion in VLM." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "text", + "content": "CAER-S test Split. Our method is further evaluated on the extensive CAER-S dataset. Unlike EMOTIC, CAER-S encompasses a more intricate background. Thus the results of VLM-related methods are significantly inferior compared to those of other emotion recognition baselines. The accuracy results are presented in Tab. 3 and compared with prominent and recently-published works, demonstrating the exceptional performance of our proposed method. Notably, our approach shows improvement compared to [19, 36], which involves the extraction of hierarchical feature from body and background regions, as well as the exploration of interdependencies across emotion labels using word embeddings or graph embeddings. This comparison highlights the efficacy of our proposed hierarchical feature fusion method." + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 316, + 71, + 553, + 213 + ], + "blocks": [ + { + "bbox": [ + 70, + 331, + 278, + 342 + ], + "lines": [ + { + "bbox": [ + 70, + 331, + 278, + 342 + ], + "spans": [ + { + "bbox": [ + 70, + 331, + 278, + 342 + ], + "type": "text", + "content": "Table 3. Performance comparisons on CAER-S test split." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 316, + 71, + 553, + 213 + ], + "lines": [ + { + "bbox": [ + 316, + 71, + 553, + 213 + ], + "spans": [ + { + "bbox": [ + 316, + 71, + 553, + 213 + ], + "type": "table", + "html": "
DatasetHFE TypeV-I TypeWorkspacemAP
EMOTICResNet-50Prototype✓36.71
Cross-Level Attn.CLIP✓36.12
Cross-Level Attn.PrototypeX35.73
Cross-Level Attn.Prototype✓37.32
DatasetHFE TypeV-I TypeWorkspaceAcc.
CAER-SResNet-50Prototype✓90.16
Cross-Level Attn.CLIP✓88.28
Cross-Level Attn.PrototypeX87.60
Cross-Level Attn.Prototype✓91.86
", + "image_path": "bc60af2a744247e7743ac10a29459d2eb24cdb17611f3b2de9af8567d91db209.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 323, + 221, + 541, + 232 + ], + "lines": [ + { + "bbox": [ + 323, + 221, + 541, + 232 + ], + "spans": [ + { + "bbox": [ + 323, + 221, + 541, + 232 + ], + "type": "text", + "content": "Table 4. Ablation Study on EMOTIC and CAER-S datasets." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 241, + 408, + 254 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 241, + 408, + 254 + ], + "spans": [ + { + "bbox": [ + 313, + 241, + 408, + 254 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 258, + 553, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 258, + 553, + 281 + ], + "spans": [ + { + "bbox": [ + 313, + 258, + 553, + 281 + ], + "type": "text", + "content": "We conduct comprehensive ablation experiments by different module combinations in our method:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 283, + 553, + 510 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 314, + 283, + 553, + 306 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 283, + 553, + 306 + ], + "spans": [ + { + "bbox": [ + 314, + 283, + 553, + 306 + ], + "type": "text", + "content": "- For hierarchical feature extraction (HFE) module, we compare the proposed method with vanilla ResNet-50." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 307, + 553, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 307, + 553, + 342 + ], + "spans": [ + { + "bbox": [ + 314, + 307, + 553, + 342 + ], + "type": "text", + "content": "- In vocabulary-informed (V-I) module, we substitute pseudo label from prototype-aligned projections (11) with CLIP predictions." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 342, + 553, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 342, + 553, + 510 + ], + "spans": [ + { + "bbox": [ + 313, + 342, + 553, + 510 + ], + "type": "text", + "content": "- In terms of the workspace, we compare with feature concatenation " + }, + { + "bbox": [ + 313, + 342, + 553, + 510 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{\\mathbf{G}}" + }, + { + "bbox": [ + 313, + 342, + 553, + 510 + ], + "type": "text", + "content": " without feature refinement, denoted as " + }, + { + "bbox": [ + 313, + 342, + 553, + 510 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 313, + 342, + 553, + 510 + ], + "type": "text", + "content": ". First, in Tab. 4, we observe that performance decreases significantly on both datasets when the workspace is degraded. This indicates that the workspace module, along with the competition and coordination mechanism, plays a vital role in coordinating multi-level image representation, thus benefiting the emotion recognition task. Secondly, we show that prototype learning is more important in the complex-environment dataset CAER-S, as it exhibits a greater performance decrease than EMOTIC. Finally, by replacing cross-level attention with ResNet-50, we observe lower performance in both datasets, indicating that modeling multi-level correlation facilitates effective emotion inference." + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 517, + 432, + 529 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 517, + 432, + 529 + ], + "spans": [ + { + "bbox": [ + 313, + 517, + 432, + 529 + ], + "type": "text", + "content": "4.4. Qualitative Analysis" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 312, + 534, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 534, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 312, + 534, + 555, + 713 + ], + "type": "text", + "content": "Case Study. In Fig. 4 we demonstrate the effectiveness of our proposed method via two randomly-chosen results of EMOTIC dataset. The left two columns contain and ground truth (GT) label and our prediction results. The prediction successfully matches GT, regardless of different and inconsistent recognition results of head, body and context window. Our approach further delineates the elimination in step two (see Fig 3) of competition process based on value of " + }, + { + "bbox": [ + 312, + 534, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{hier}}" + }, + { + "bbox": [ + 312, + 534, + 555, + 713 + ], + "type": "text", + "content": " of maximum loss. It illustrates that with the guidance of pseudo label from V-I module, the inconsistent recognition results are gradually eliminated. We use orange color to high-light the labels in multi-level results that are not shown in V-I and prediction column. For example, from the top row of Fig. 4, the head level, which is the most inconsistent level of recognition results, is initially filtered out, and" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "29597" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 91, + 71, + 195, + 186 + ], + "blocks": [ + { + "bbox": [ + 91, + 71, + 195, + 186 + ], + "lines": [ + { + "bbox": [ + 91, + 71, + 195, + 186 + ], + "spans": [ + { + "bbox": [ + 91, + 71, + 195, + 186 + ], + "type": "image", + "image_path": "87de515feaea1f2141585b08b14d76b581601b0aabfb75e90f349e08af33e456.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 198, + 72, + 304, + 188 + ], + "blocks": [ + { + "bbox": [ + 198, + 72, + 304, + 188 + ], + "lines": [ + { + "bbox": [ + 198, + 72, + 304, + 188 + ], + "spans": [ + { + "bbox": [ + 198, + 72, + 304, + 188 + ], + "type": "image", + "image_path": "a77b652b353d3a71efa4a7f75669096296a80baa972a737b4e60913bf1bd7933.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 307, + 72, + 411, + 188 + ], + "blocks": [ + { + "bbox": [ + 307, + 72, + 411, + 188 + ], + "lines": [ + { + "bbox": [ + 307, + 72, + 411, + 188 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 411, + 188 + ], + "type": "image", + "image_path": "476deb5786f21a2fc3960a5ec981bdf487ed7a5e2ec8c2c645862c668a7fb24b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 414, + 72, + 520, + 188 + ], + "blocks": [ + { + "bbox": [ + 414, + 72, + 520, + 188 + ], + "lines": [ + { + "bbox": [ + 414, + 72, + 520, + 188 + ], + "spans": [ + { + "bbox": [ + 414, + 72, + 520, + 188 + ], + "type": "image", + "image_path": "43d77da0600a0cfe4246b8921a7a59d3d0f6668f77746586f1cd1e9af4d68d36.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 91, + 183, + 196, + 285 + ], + "blocks": [ + { + "bbox": [ + 91, + 183, + 196, + 285 + ], + "lines": [ + { + "bbox": [ + 91, + 183, + 196, + 285 + ], + "spans": [ + { + "bbox": [ + 91, + 183, + 196, + 285 + ], + "type": "image", + "image_path": "260951db3fa588ad13282164965402d4e345823e2f4b3b3716ba704548cf1073.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 308, + 555, + 332 + ], + "lines": [ + { + "bbox": [ + 55, + 308, + 555, + 332 + ], + "spans": [ + { + "bbox": [ + 55, + 308, + 555, + 332 + ], + "type": "text", + "content": "Figure 5. t-SNE and saliency maps visualization of feature refinement on CAER-S. The red star refers to a random selected sample feature that has been clustered in a more compact group in t-SNE visualization, and displaying more semantic regions on saliency maps." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 198, + 189, + 304, + 300 + ], + "blocks": [ + { + "bbox": [ + 198, + 189, + 304, + 300 + ], + "lines": [ + { + "bbox": [ + 198, + 189, + 304, + 300 + ], + "spans": [ + { + "bbox": [ + 198, + 189, + 304, + 300 + ], + "type": "image", + "image_path": "65f8498f799a40f870622ec680b5325819540b4a4824000f062915bc0c97b5e3.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 307, + 189, + 411, + 300 + ], + "blocks": [ + { + "bbox": [ + 307, + 189, + 411, + 300 + ], + "lines": [ + { + "bbox": [ + 307, + 189, + 411, + 300 + ], + "spans": [ + { + "bbox": [ + 307, + 189, + 411, + 300 + ], + "type": "image", + "image_path": "4656457d35e05ba8aa9d14aa754cfdc808a7589b7ec9533aa9f7f5ca7fba36d2.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 414, + 189, + 520, + 299 + ], + "blocks": [ + { + "bbox": [ + 414, + 189, + 520, + 299 + ], + "lines": [ + { + "bbox": [ + 414, + 189, + 520, + 299 + ], + "spans": [ + { + "bbox": [ + 414, + 189, + 520, + 299 + ], + "type": "image", + "image_path": "e775eed61d8e7b74624bbb0e896e3b955f3f622643310421090cb0b4c7e3d70e.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 341, + 295, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 341, + 295, + 389 + ], + "spans": [ + { + "bbox": [ + 54, + 341, + 295, + 389 + ], + "type": "text", + "content": "then prediction from the context level is subsequently eliminated. This supports our hypothesis that if the inconsistent multi-level emotion recognition is excluded, the final prediction performance is enhanced." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 54, + 389, + 295, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 389, + 295, + 568 + ], + "spans": [ + { + "bbox": [ + 54, + 389, + 295, + 568 + ], + "type": "text", + "content": "How is multi-level feature refined? To understand the feature refinement in workspace, we employ t-SNE on CAERS to visualize multi-level feature, with 250 randomly sampled testing images per class. As shown in Fig. 5, features " + }, + { + "bbox": [ + 54, + 389, + 295, + 568 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_m" + }, + { + "bbox": [ + 54, + 389, + 295, + 568 + ], + "type": "text", + "content": " from different level " + }, + { + "bbox": [ + 54, + 389, + 295, + 568 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 54, + 389, + 295, + 568 + ], + "type": "text", + "content": " are refined after inference-stage parameter update of " + }, + { + "bbox": [ + 54, + 389, + 295, + 568 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 54, + 389, + 295, + 568 + ], + "type": "text", + "content": ", displaying more clustered groups. Consequently, concatenated feature " + }, + { + "bbox": [ + 54, + 389, + 295, + 568 + ], + "type": "inline_equation", + "content": "[\\hat{\\mathbf{z}}_{\\mathbf{c}}; \\hat{\\mathbf{z}}_{\\mathbf{b}}; \\hat{\\mathbf{z}}_{\\mathbf{h}}]" + }, + { + "bbox": [ + 54, + 389, + 295, + 568 + ], + "type": "text", + "content": " encompasses compact clusters of in-group representation. We also provide a clustered sample point (shown in red star in Fig. 5) by Grad-CAM [31], a wildly used saliency maps visualization technique of highlighting the semantic pixels. We can observe that with the parameter update of " + }, + { + "bbox": [ + 54, + 389, + 295, + 568 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 54, + 389, + 295, + 568 + ], + "type": "text", + "content": ", body-level and head-level saliency maps align with hand gestures and facial expressions, which contain more semantic information for emotion recognition." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 54, + 569, + 295, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 569, + 295, + 664 + ], + "spans": [ + { + "bbox": [ + 54, + 569, + 295, + 664 + ], + "type": "text", + "content": "Is the proposed framework potential to enhance VLM? Since the performance of VLM is worse than expected, we aim to explore whether we can use our proposed framework to improve emotion recognition of VLM. First, based on step two of the competition process, we can obtain the first excluded image level, which we denote as the most inconsistent level " + }, + { + "bbox": [ + 54, + 569, + 295, + 664 + ], + "type": "inline_equation", + "content": "\\kappa" + }, + { + "bbox": [ + 54, + 569, + 295, + 664 + ], + "type": "text", + "content": ". Next, using this level " + }, + { + "bbox": [ + 54, + 569, + 295, + 664 + ], + "type": "inline_equation", + "content": "\\kappa" + }, + { + "bbox": [ + 54, + 569, + 295, + 664 + ], + "type": "text", + "content": ", we modify the input of the VLM in two ways:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 665, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 55, + 665, + 295, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 295, + 689 + ], + "type": "text", + "content": "- Prompt" + }, + { + "bbox": [ + 55, + 665, + 295, + 689 + ], + "type": "inline_equation", + "content": "\\kappa" + }, + { + "bbox": [ + 55, + 665, + 295, + 689 + ], + "type": "text", + "content": ": we add the prompt with: \"Please avoid potentially inconsistent recognition results from " + }, + { + "bbox": [ + 55, + 665, + 295, + 689 + ], + "type": "inline_equation", + "content": "\\{\\kappa\\}" + }, + { + "bbox": [ + 55, + 665, + 295, + 689 + ], + "type": "text", + "content": ".\"" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 689, + 295, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 712 + ], + "type": "text", + "content": "- Multi-Level" + }, + { + "bbox": [ + 55, + 689, + 295, + 712 + ], + "type": "inline_equation", + "content": "^{\\kappa}" + }, + { + "bbox": [ + 55, + 689, + 295, + 712 + ], + "type": "text", + "content": ": we use multi-level image inputs but eliminate the level " + }, + { + "bbox": [ + 55, + 689, + 295, + 712 + ], + "type": "inline_equation", + "content": "\\kappa" + }, + { + "bbox": [ + 55, + 689, + 295, + 712 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "text" + }, + { + "type": "table", + "bbox": [ + 316, + 339, + 553, + 388 + ], + "blocks": [ + { + "bbox": [ + 316, + 339, + 553, + 388 + ], + "lines": [ + { + "bbox": [ + 316, + 339, + 553, + 388 + ], + "spans": [ + { + "bbox": [ + 316, + 339, + 553, + 388 + ], + "type": "table", + "html": "
MethodBLIP2 [17]LLaVa [20]GPT-4o [27]
PromptK26.77+0.8128.56+0.5725.11+1.31
Multi-LevelK27.14+1.1829.03+1.0426.64+2.84
", + "image_path": "41ceb0d718ca5d0db0cea3b6ad9822e666d76dcca235506b951ffc0fbe3d19af.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_body" + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 396, + 555, + 419 + ], + "lines": [ + { + "bbox": [ + 313, + 396, + 555, + 419 + ], + "spans": [ + { + "bbox": [ + 313, + 396, + 555, + 419 + ], + "type": "text", + "content": "Table 5. We report the mAP of enhanced VLM results on EMOTIC dataset. " + }, + { + "bbox": [ + 313, + 396, + 555, + 419 + ], + "type": "inline_equation", + "content": "\\kappa" + }, + { + "bbox": [ + 313, + 396, + 555, + 419 + ], + "type": "text", + "content": " refers to modified input via our proposed framework." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 430, + 555, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 430, + 555, + 491 + ], + "spans": [ + { + "bbox": [ + 313, + 430, + 555, + 491 + ], + "type": "text", + "content": "As shown in Tab. 5, both of them successfully enhance emotion recognition capabilities of VLM, especially for GPT-40. Thus, we validate the inconsistency of multi-level inputs also exists in VLM and we improve it using our framework. Please see Appendix (Sec.10.2) for detailed cases." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 502, + 388, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 502, + 388, + 514 + ], + "spans": [ + { + "bbox": [ + 313, + 502, + 388, + 514 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 521, + 555, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 555, + 666 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 555, + 666 + ], + "type": "text", + "content": "In this paper, we propose a competition and coordination mechanism to avoid hierarchical visual feature inconsistency in emotion recognition. To guide the alignment of visual features, a vocabulary-informed module is applied to produce labels that refine hierarchical visual representation. We elaborate on the competition process to eliminate irrelevant image-level representations and the coordination process to enhance representations across all image levels. Extensive experiments are conducted on two popular benchmarks, the EMOTIC and CAER-S datasets. Our method not only achieves state-of-the-art performance with better interpretability but also improves large vision-language models." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 666, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 666, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 666, + 553, + 713 + ], + "type": "text", + "content": "Acknowledgments. This work was supported by Shanghai Science and Technology Committee under Grant No. 24511103900, 24511103200, and National Key RD Program of China under grant No. 2022YFC3601405." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "29598" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 91, + 294, + 112 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 294, + 112 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 294, + 112 + ], + "type": "text", + "content": "[1] Bernard J Baars. A cognitive theory of consciousness. Cambridge University Press, 1993. 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 114, + 295, + 157 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 114, + 295, + 157 + ], + "spans": [ + { + "bbox": [ + 61, + 114, + 295, + 157 + ], + "type": "text", + "content": "[2] Lenore Blum and Manuel Blum. A theory of consciousness from a theoretical computer science perspective: Insights from the conscious Turing machine. Proceedings of the National Academy of Sciences, 119(21):e2115934119, 2022. 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 159, + 295, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 159, + 295, + 201 + ], + "spans": [ + { + "bbox": [ + 62, + 159, + 295, + 201 + ], + "type": "text", + "content": "[3] Jean-Benoit Delbrouck, Noé Tits, Mathilde Brousmiche, and Stéphane Dupont. A transformer-based joint-encoding for emotion recognition and sentiment analysis. arXiv preprint arXiv:2006.15955, 2020. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 203, + 294, + 235 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 203, + 294, + 235 + ], + "spans": [ + { + "bbox": [ + 62, + 203, + 294, + 235 + ], + "type": "text", + "content": "[4] Jiankang Deng, Jia Guo, Xue Niannan, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In CVPR, 2019. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 236, + 294, + 301 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 236, + 294, + 301 + ], + "spans": [ + { + "bbox": [ + 62, + 236, + 294, + 301 + ], + "type": "text", + "content": "[5] Jiayuan Dong, Alex Santiago-Anaya, and Myounghoon Jeon. Facial expressions increase emotion recognition clarity and improve warmth and attractiveness on a humanoid robot without adding the uncanny valley. Proceedings of the Human Factors and Ergonomics Society Annual Meeting, 67 (1):933-939, 2023. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 303, + 294, + 367 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 303, + 294, + 367 + ], + "spans": [ + { + "bbox": [ + 62, + 303, + 294, + 367 + ], + "type": "text", + "content": "[6] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 369, + 294, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 369, + 294, + 423 + ], + "spans": [ + { + "bbox": [ + 62, + 369, + 294, + 423 + ], + "type": "text", + "content": "[7] Bernd Dudzik, Joost Broekens, Mark Neerincx, and Hayley Hung. Exploring personal memories and video content as context for facial behavior in predictions of video-induced emotions. In Proceedings of the 2020 International Conference on Multimodal Interaction, pages 153-162, 2020. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 425, + 294, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 425, + 294, + 479 + ], + "spans": [ + { + "bbox": [ + 62, + 425, + 294, + 479 + ], + "type": "text", + "content": "[8] Yanwei Fu, Xiaomei Wang, Hanze Dong, Yu-Gang Jiang, Meng Wang, Xiangyang Xue, and Leonid Sigal. Vocabulary-informed zero-shot and open-set learning. IEEE Transactions on Pattern Analysis and Machine Intelligence, 42(12): 3136-3152, 2020. 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 480, + 294, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 480, + 294, + 523 + ], + "spans": [ + { + "bbox": [ + 62, + 480, + 294, + 523 + ], + "type": "text", + "content": "[9] Manh-Hung Hoang, Soo-Hyung Kim, Hyung-Jeong Yang, and Guee-Sang Lee. Context-aware emotion recognition based on visual relationship detection. IEEE Access, 9: 90465-90474, 2021. 5, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 525, + 294, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 525, + 294, + 579 + ], + "spans": [ + { + "bbox": [ + 57, + 525, + 294, + 579 + ], + "type": "text", + "content": "[10] Jian Huang, Jianhua Tao, Bin Liu, Zheng Lian, and Mingyue Niu. Multimodal transformer fusion for continuous emotion recognition. In ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 3507-3511. IEEE, 2020. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 581, + 294, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 581, + 294, + 601 + ], + "spans": [ + { + "bbox": [ + 57, + 581, + 294, + 601 + ], + "type": "text", + "content": "[11] Oleg Kalyta. Information technology of facial emotion recognition for visual safety surveillance. 2022. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 603, + 295, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 603, + 295, + 646 + ], + "spans": [ + { + "bbox": [ + 57, + 603, + 295, + 646 + ], + "type": "text", + "content": "[12] Ronak Kosti, Jose M Alvarez, Adria Recasens, and Agata Lapedriza. Emotic: Emotions in context dataset. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 61-69, 2017. 2, 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 647, + 295, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 295, + 690 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 295, + 690 + ], + "type": "text", + "content": "[13] Ronak Kosti, Jose M Alvarez, Adria Recasens, and Agata Lapedriza. Emotion recognition in context. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1667-1675, 2017. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 692, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 692, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 692, + 295, + 713 + ], + "type": "text", + "content": "[14] Ronak Kosti, Jose M Alvarez, Adria Recasens, and Agata Lapedriza. Context based emotion recognition using emotic" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 74, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 333, + 74, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 74, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 74, + 553, + 95 + ], + "type": "text", + "content": "dataset. IEEE transactions on pattern analysis and machine intelligence, 42(11):2755-2766, 2019. 1, 2, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 97, + 553, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 97, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 316, + 97, + 553, + 140 + ], + "type": "text", + "content": "[15] Jiyoung Lee, Seungryong Kim, Sunok Kim, Jungin Park, and Kwanghoon Sohn. Context-aware emotion recognition networks. In 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 10142-10151, 2019. 1, 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 142, + 553, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 142, + 553, + 163 + ], + "spans": [ + { + "bbox": [ + 316, + 142, + 553, + 163 + ], + "type": "text", + "content": "[16] Li et al. Decoupled multimodal distilling for emotion recognition. In CVPR, 2023. 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 164, + 553, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 164, + 553, + 207 + ], + "spans": [ + { + "bbox": [ + 317, + 164, + 553, + 207 + ], + "type": "text", + "content": "[17] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, 2022. 2, 6, 7, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 209, + 553, + 242 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 209, + 553, + 242 + ], + "spans": [ + { + "bbox": [ + 317, + 209, + 553, + 242 + ], + "type": "text", + "content": "[18] Qingbiao Li, Chunhua Wu, Zhe Wang, and Kangfeng Zheng. Hierarchical transformer network for utterance-level emotion recognition. Applied Sciences, 10(13):4447, 2020. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 243, + 553, + 276 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 243, + 553, + 276 + ], + "spans": [ + { + "bbox": [ + 317, + 243, + 553, + 276 + ], + "type": "text", + "content": "[19] Weixin Li, Xuan Dong, and Yunhong Wang. Human emotion recognition with relational region-level analysis. IEEE Transactions on Affective Computing, 2021. 6, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 277, + 553, + 298 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 277, + 553, + 298 + ], + "spans": [ + { + "bbox": [ + 316, + 277, + 553, + 298 + ], + "type": "text", + "content": "[20] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, 2023. 2, 6, 7, 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 300, + 553, + 355 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 300, + 553, + 355 + ], + "spans": [ + { + "bbox": [ + 317, + 300, + 553, + 355 + ], + "type": "text", + "content": "[21] Xin Liu, Henglin Shi, Haoyu Chen, Zitong Yu, Xiaobai Li, and Guoying Zhao. imigue: An identity-free video dataset for micro-gesture understanding and emotion analysis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10631-10642, 2021. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 357, + 553, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 357, + 553, + 421 + ], + "spans": [ + { + "bbox": [ + 317, + 357, + 553, + 421 + ], + "type": "text", + "content": "[22] Fengmao Lv, Xiang Chen, Yanyong Huang, Lixin Duan, and Guosheng Lin. Progressive modality reinforcement for human multimodal emotion recognition from unaligned multimodal sequences. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2554-2562, 2021. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 423, + 553, + 478 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 423, + 553, + 478 + ], + "spans": [ + { + "bbox": [ + 317, + 423, + 553, + 478 + ], + "type": "text", + "content": "[23] Trisha Mittal, Uttaran Bhattacharya, Rohan Chandra, Aniket Bera, and Dinesh Manocha. M3er: Multiplicative multimodal emotion recognition using facial, textual, and speech cues. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1359-1367, 2020. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 479, + 553, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 479, + 553, + 544 + ], + "spans": [ + { + "bbox": [ + 317, + 479, + 553, + 544 + ], + "type": "text", + "content": "[24] Trisha Mittal, Pooja Guhan, Uttaran Bhattacharya, Rohan Chandra, Aniket Bera, and Dinesh Manocha. Emoticon: Context-aware multimodal emotion recognition using frege's principle. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14234-14243, 2020. 1, 2, 5, 6, 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 546, + 553, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 546, + 553, + 588 + ], + "spans": [ + { + "bbox": [ + 317, + 546, + 553, + 588 + ], + "type": "text", + "content": "[25] Trisha Mittal, Aniket Bera, and Dinesh Manocha. Multimodal and context-aware emotion perception model with multiplicative fusion. IEEE MultiMedia, 28(2):67-75, 2021. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 590, + 555, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 590, + 555, + 645 + ], + "spans": [ + { + "bbox": [ + 317, + 590, + 555, + 645 + ], + "type": "text", + "content": "[26] Trisha Mittal, Puneet Mathur, Aniket Bera, and Dinesh Manocha. Affect2mm: Affective analysis of multimedia content using emotion causality. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5661-5671, 2021. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 647, + 553, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 678 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 678 + ], + "type": "text", + "content": "[27] OpenAI. GPT-4o API, 2024. Available at: https://openai.com/index/hello-gpt-4o, last accessed on 2024-11-01. 1, 6, 7, 8" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "text", + "content": "[28] Kunyu Peng, Alina Roitberg, David Schneider, Marios Koulakis, Kailun Yang, and Rainer Stiefelhagen. Affect-dml: Context-aware one-shot recognition of human affect" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "29599" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 715 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 76, + 72, + 296, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 296, + 106 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 296, + 106 + ], + "type": "text", + "content": "using deep metric learning. In 2021 16th IEEE International Conference on Automatic Face and Gesture Recognition (FG 2021), pages 1-8, 2021. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 296, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 296, + 163 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 296, + 163 + ], + "type": "text", + "content": "[29] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision, 2021. 4, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 164, + 296, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 164, + 296, + 228 + ], + "spans": [ + { + "bbox": [ + 56, + 164, + 296, + 228 + ], + "type": "text", + "content": "[30] Shulan Ruan, Kun Zhang, Yijun Wang, Hanqing Tao, Weidong He, Guangyi Lv, and Enhong Chen. Context-aware generation-based net for multi-label visual emotion recognition. In 2020 IEEE International Conference on Multimedia and Expo (ICME), pages 1-6. IEEE Computer Society, 2020. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 230, + 296, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 230, + 296, + 297 + ], + "spans": [ + { + "bbox": [ + 56, + 230, + 296, + 297 + ], + "type": "text", + "content": "[31] Ramprasaath R. Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In 2017 IEEE International Conference on Computer Vision (ICCV), pages 618-626, 2017. 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 298, + 296, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 298, + 296, + 363 + ], + "spans": [ + { + "bbox": [ + 56, + 298, + 296, + 363 + ], + "type": "text", + "content": "[32] Jiahui She, Yibo Hu, Hailin Shi, Jun Wang, Qiu Shen, and Tao Mei. Dive into ambiguity: latent distribution mining and pairwise uncertainty estimation for facial expression recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6248-6257, 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 365, + 296, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 365, + 296, + 421 + ], + "spans": [ + { + "bbox": [ + 56, + 365, + 296, + 421 + ], + "type": "text", + "content": "[33] Xuli Shen, Xiaomei Wang, Qing Xu, Weifeng Ge, and Xi-angyang Xue. Towards scalable and fast distributionally robust optimization for data-driven deep learning. In 2022 IEEE International Conference on Data Mining (ICDM), pages 448–457, 2022. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 422, + 296, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 422, + 296, + 465 + ], + "spans": [ + { + "bbox": [ + 56, + 422, + 296, + 465 + ], + "type": "text", + "content": "[34] Shamane Siriwardhana, Tharindu Kaluarachchi, Mark Billinghurst, and Suranga Nanayakkara. Multimodal emotion recognition with transformer-based self supervised feature fusion. IEEE Access, 8:176274-176285, 2020. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 467, + 296, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 467, + 296, + 510 + ], + "spans": [ + { + "bbox": [ + 56, + 467, + 296, + 510 + ], + "type": "text", + "content": "[35] Zili Wang, Lingjie Lao, Xiaoya Zhang, Yong Li, Tong Zhang, and Zhen Cui. Context-dependent emotion recognition. Journal of Visual Communication and Image Representation, 89:103679, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 512, + 296, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 512, + 296, + 555 + ], + "spans": [ + { + "bbox": [ + 56, + 512, + 296, + 555 + ], + "type": "text", + "content": "[36] Shichao Wu, Lei Zhou, Zhengxi Hu, and Jingtai Liu. Hierarchical context-based emotion recognition with scene graphs. IEEE Transactions on Neural Networks and Learning Systems, 35(3):3725-3739, 2024. 1, 2, 4, 6, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 557, + 296, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 557, + 296, + 624 + ], + "spans": [ + { + "bbox": [ + 56, + 557, + 296, + 624 + ], + "type": "text", + "content": "[37] Dingkang Yang, Shuai Huang, Haopeng Kuang, Yangtao Du, and Lihua Zhang. Disentangled representation learning for multimodal emotion recognition. In Proceedings of the 30th ACM International Conference on Multimedia, pages 1642-1651, New York, NY, USA, 2022. Association for Computing Machinery. 1, 2, 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 624, + 296, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 296, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 296, + 689 + ], + "type": "text", + "content": "[38] Dingkang Yang, Zhaoyu Chen, Yuzheng Wang, Shunli Wang, Mingcheng Li, Siao Liu, Xiao Zhao, Shuai Huang, Zhiyan Dong, Peng Zhai, and Lihua Zhang. Context de-confounded emotion recognition. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2023. 1, 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 691, + 296, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 296, + 715 + ], + "type": "text", + "content": "[39] Dingkang Yang, Mingcheng Li, Dongling Xiao, Yang Liu, Kun Yang, Zhaoyu Chen, Yuzheng Wang, Peng Zhai, Ke Li," + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 398 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 333, + 72, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 72, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 72, + 553, + 95 + ], + "type": "text", + "content": "and Lihua Zhang. Towards multimodal sentiment analysis debiasing via bias purification, 2024. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 96, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 150 + ], + "type": "text", + "content": "[40] Dingkang Yang, Kun Yang, Mingcheng Li, Shunli Wang, Shuaibing Wang, and Lihua Zhang. Robust emotion recognition in context debiasing. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2024. 1, 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 152, + 553, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 553, + 206 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 553, + 206 + ], + "type": "text", + "content": "[41] Jingyuan Yang, Jie Li, Leida Li, Xiumei Wang, and Xinbo Gao. A circular-structured representation for visual emotion distribution learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4237–4246, 2021. 1, 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 208, + 553, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 208, + 553, + 252 + ], + "spans": [ + { + "bbox": [ + 316, + 208, + 553, + 252 + ], + "type": "text", + "content": "[42] Minghui Zhang, Yumeng Liang, and Huadong Ma. Context-aware affective graph reasoning for emotion recognition. In 2019 IEEE International Conference on Multimedia and Expo (ICME), pages 151-156. IEEE, 2019. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 253, + 553, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 253, + 553, + 306 + ], + "spans": [ + { + "bbox": [ + 316, + 253, + 553, + 306 + ], + "type": "text", + "content": "[43] Yang Zhao, Hao Zhang, and Xiuyuan Hu. Penalizing gradient norm for efficiently improving generalization in deep learning. In Proceedings of the 39th International Conference on Machine Learning, pages 26982-26992. PMLR, 2022. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 308, + 553, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 308, + 553, + 352 + ], + "spans": [ + { + "bbox": [ + 316, + 308, + 553, + 352 + ], + "type": "text", + "content": "[44] Zengqun Zhao, Qingshan Liu, and Shanmin Wang. Learning deep global multi-scale and local attention features for facial expression recognition in the wild. IEEE Transactions on Image Processing, 30:6544-6556, 2021. 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 354, + 553, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 354, + 553, + 398 + ], + "spans": [ + { + "bbox": [ + 316, + 354, + 553, + 398 + ], + "type": "text", + "content": "[45] Zengqun Zhao, Qingshan Liu, and Feng Zhou. Robust lightweight facial expression recognition network with label distribution training. In Proceedings of the AAAI conference on artificial intelligence, pages 3510-3519, 2021. 7" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "29600" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Code-as-Monitor_ Constraint-aware Visual Programming for Reactive and Proactive Robotic Failure Detection/3086dac5-625c-4ed2-b32d-12ee6b15f252_content_list.json b/2025/Code-as-Monitor_ Constraint-aware Visual Programming for Reactive and Proactive Robotic Failure Detection/3086dac5-625c-4ed2-b32d-12ee6b15f252_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..33d3665a340543ad46cf50be5cba63a56d3f349a --- /dev/null +++ b/2025/Code-as-Monitor_ Constraint-aware Visual Programming for Reactive and Proactive Robotic Failure Detection/3086dac5-625c-4ed2-b32d-12ee6b15f252_content_list.json @@ -0,0 +1,1606 @@ +[ + { + "type": "text", + "text": "Code-as-Monitor: Constraint-aware Visual Programming for Reactive and Proactive Robotic Failure Detection", + "text_level": 1, + "bbox": [ + 189, + 130, + 810, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Enshen Zhou $^{1,3*}$ , Qi Su $^{2,3,4*}$ , Cheng Chi $^{3*†}$ , Zhizheng Zhang $^{4}$ , Zhongyuan Wang $^{3}$ , Tiejun Huang $^{2,3}$ , Lu Sheng $^{1†}$ , He Wang $^{2,3,4†}$ , $^{1}$ Beihang University $^{2}$ Peking University $^{3}$ Beijing Academy of Artificial Intelligence $^{4}$ Galbot", + "bbox": [ + 156, + 202, + 844, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "zhouenshen@buaa.edu.cn chicheng15@mails.ucas.ac.cn lsheng@buaa.edu.cn hewang@pku.edu.cn", + "bbox": [ + 132, + 260, + 864, + 273 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/02514d55c0840685d9698546835ab2425a6bcdb26c727621e2fc7aca1f41445f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 102, + 287, + 135, + 311 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Task: Move the pan with the lobster to the stove, and be careful not to let the lobster drop out.", + "bbox": [ + 140, + 292, + 756, + 309 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ad2a9b3059d16dfce8fb97ac8eaa7705617ecc383e20ea57c9c28f978527b444.jpg", + "image_caption": [ + "(a) Reactive Failure Detection" + ], + "image_footnote": [], + "bbox": [ + 94, + 311, + 437, + 496 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/afb1dd77eb2d946721e72019f19549490e43e8d7270770e895d49b66c2d4fb41.jpg", + "image_caption": [ + "(b) Proactive Failure Detection" + ], + "image_footnote": [], + "bbox": [ + 437, + 311, + 781, + 496 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/813e5a2c28f246840e12ff1c4a1c1c84d785cda759cce1ba9f8fc0af6e4bc184.jpg", + "image_caption": [ + "(c) Real-world Test", + "Figure 1. For the task \"Move the pan with lobster to the stove without losing the lobster\", (a) reactive failure detection identifies failures after they occur, and (b) proactive failure detection prevents foreseeable failures. In (a), at $t_4^R$ , the robot detects the failure after the lobster unpredictably jumps out due to the heat. In (b), pan tilting is detected at $t_3^P$ and corrected it at $t_3^{P'}$ , requiring real-time precision. We formulate both tasks as spatio-temporal constraint satisfaction problems, leveraging our proposed constraint elements for precise, real-time checking. For example, in (a), a large relative distance between pan and lobster indicates failure; in (b), a large angle between the pan and the horizontal plane needs correction. (c) shows that our method combined with an open-loop policy forms a closed-loop system, enabling proactive (e.g., detecting moving glass during grasping) and reactive (e.g., removing toy after grasping) failure detection in cluttered scenes." + ], + "image_footnote": [], + "bbox": [ + 782, + 290, + 903, + 497 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 628, + 326, + 643 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Automatic detection and prevention of open-set failures are crucial in closed-loop robotic systems. Recent studies often struggle to simultaneously identify unexpected failures reactively after they occur and prevent foreseeable ones proactively. To this end, we propose Code-as-Monitor (CaM), a novel paradigm leveraging the vision-language model (VLM) for both open-set reactive and proactive failure detection. The core of our method is to formulate both tasks as a unified set of spatio-temporal constraint satisfaction problems and use VLM-generated code to evaluate them for real-time monitoring. To enhance the accuracy and efficiency of monitoring, we further introduce constraint elements that abstract constraint-related entities or their parts into compact geometric ele", + "bbox": [ + 86, + 661, + 485, + 875 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ments. This approach offers greater generality, simplifies tracking, and facilitates constraint-aware visual programming by leveraging these elements as visual prompts. Experiments show that CaM achieves a $28.7\\%$ higher success rate and reduces execution time by $31.8\\%$ under severe disturbances compared to baselines across three simulators and a real-world setting. Moreover, CaM can be integrated with open-loop control policies to form closed-loop systems, enabling long-horizon tasks in cluttered scenes with dynamic environments. See the project page at https://zhoues.github.io/Code-as-Monitor.", + "bbox": [ + 511, + 628, + 908, + 796 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 813, + 645, + 829 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "As expectations grow for robots to handle long-horizon tasks within intricate environments, failures are unavoidable. Therefore, automatically detecting and preventing those failures plays a vital role in ensuring the tasks can", + "bbox": [ + 511, + 839, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Equal contribution † Corresponding author", + "bbox": [ + 114, + 886, + 362, + 901 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "6919", + "bbox": [ + 482, + 944, + 514, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "eventually be solved, especially for closed-loop robotic systems. There are two modes of failure detection [30], reactive and proactive. As depicted in Fig. 1, reactive failure detection aims to identify failures after they occur (e.g., recognizing that lobster lands on the table, indicating a delivery failure). In contrast, proactive failure detection aims to prevent foreseeable failures (e.g., recognizing that a tilted pan could cause the lobster to fall out, leading to a delivery failure). Both detection modes are even more challenging in open-set scenarios, where the failures are not predefined.", + "bbox": [ + 89, + 90, + 480, + 242 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "With the help of large language models (LLMs) [14, 58] and vision-language models (VLMs) [1, 36], recent studies can achieve open-set reactive failure detection [10, 12, 13, 15, 20, 37, 55, 56, 66, 74] as a special case of visual question answering (VQA) tasks. However, these methods often bear compromised execution speeds and coarse-grained detection accuracy, due to the high computational costs and inadequate 3D spatio-temporal perception capability of recent LLMs/VLMs. Moreover, open-set proactive failure detection, which has been less explored in the literature, presents even severer challenges as it is required to foresee potential causes of failure and monitor them in real-time with high precision to anticipate and prevent imminent failure. Simply adapting LLMs/VLMs cannot meet such expectations.", + "bbox": [ + 89, + 244, + 480, + 455 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we aim to develop an open-set failure detection framework that achieves reactive and proactive detection simultaneously, not only benefiting from the generalization power of VLMs but also enjoying high precision in monitoring failure characteristics with real-time efficiency. We address this by formulating both tasks as a unified set of spatio-temporal constraint satisfaction problems, which can be precisely translated by VLMs into executable programs. Such visual programs can efficiently verify whether entities (e.g., robots, objects) or their parts in the captured environment maintain or achieve required states during or after execution (i.e., satisfying constraints), so as to immediately prevent or detect failures. To the best of our knowledge, this is the first attempt to integrate both detection modes within a single framework. We name this constraint-aware visual programming framework as Code-as-Monitor $(CaM)$ .", + "bbox": [ + 89, + 459, + 482, + 700 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To be specific, the proposed spatio-temporal constraint satisfaction scheme abstracts the constraint-related entity or part segments from the observed images into compact geometric elements (e.g., points, lines, and surfaces), as shown in Fig. 1. It simplifies the monitoring of constraint satisfaction by tracking and evaluating the spatio-temporal combinational dynamics of these elements, eliminating the most irrelevant geometric and visual details of the raw entities/parts. The constraint element detection and tracking are grounded by our trained constraint-aware segmentation and off-the-shelf tracking models, ensuring speed, accuracy, and certain open-set adaptation capabilities. The evaluation protocol is in the form of VLM-generated code, i.e., monitor", + "bbox": [ + 89, + 704, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "code, which is visually prompted by the starting frames of a sub-goal and its associated constraint elements, in addition to the textual constraints that must be fulfilled. Once generated, this monitor code could detect reactive or proactive failures just by being executed according to the tracked constraint elements, without needing to call the VLMs again. Therefore, this minimalist scheme is generalizable to open-set failure detection for unseen entities and scenes (enabled by the potential diversity of the structured associations of constraint elements) as well as common skills (powered by the rich prior knowledge offered by VLMs), maintaining sufficient detection accuracy and real-time execution speed.", + "bbox": [ + 511, + 90, + 903, + 271 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We conduct extensive experiments in three simulators (i.e., CLIPort [54], Omnigibson [31], and RL-Bench [22]) and one real-world setting, spanning diverse manipulation tasks (e.g., pick & place, articulated objects, tool-use), robot platforms (e.g., UR5, Fetch, Franka) and end-effectors (e.g., suction cup, gripper, dexterous hand). The results show that CaM is generalizable and achieves both reactive and proactive failure detection in real-time, resulting in $28.7\\%$ higher success rates and reduced execution time by $31.8\\%$ under severe disturbances compared to the baselines. Moreover, in Sec. 4.3, CaM can be integrated with the existing open-loop control policy to form a closed-loop system, enabling long-horizon tasks in cluttered scenes with environment dynamics and human disturbances. Our contributions are summarized as follows:", + "bbox": [ + 511, + 272, + 903, + 496 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce Code-as-Monitor (CaM), a novel paradigm that leverages VLMs for both reactive and proactive failure detection via constraint-aware visual programming.", + "- We propose the constraint elements to enhance the accuracy and efficiency of constraint satisfaction monitoring.", + "- Extensive experiments show that CaM is generalizable and achieves more real-time and precise failure detection than baselines across simulators and real-world settings." + ], + "bbox": [ + 513, + 500, + 903, + 618 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 513, + 633, + 651, + 648 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Robotic Failure Detection. Recent advances in LLMs [4, 14, 58, 59] and VLMs [1, 3, 18, 33, 36, 44, 45, 52, 71, 76] greatly improve open-set reactive failure detection for robotics [7, 9, 23, 24, 43, 46, 57, 72] Current LLM-based methods either convert visual inputs into text [38, 53, 64], potentially losing visual details, or rely on ground-truth feedback [20, 47, 55, 56], which is impractical in real-world scenarios. Recent studies employ VLMs as failure detectors, offering binary success indicators [12, 37, 66, 74] or textual explanations [10, 13] through visual question answering (VQA), such as DoReMi [15]. However, they often bear compromised execution speeds and coarse-grained detection accuracy. Meanwhile, open-set proactive failure detection is rarely explored, as previous methods [2, 11] are confined to predefined failures. This detection mode requires foreseeing potential failure causes and monitoring", + "bbox": [ + 511, + 657, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "6920", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/ddded73361f0a4434fdf1b72370125485c3d7d83cf3d628cb57583acb5b99e7d.jpg", + "image_caption": [ + "Figure 2. Overview of Code-as-Monitor. Given task instructions and prior information, the Constraint Generator derives the next subgoal and corresponding textual constraints based on multi-view observations. The Painter maps these constraints onto images as constraint elements. The Monitor generates monitor code from these images and tracks them for real-time monitoring. If any constraint is violated, it outputs the reason for failure and triggers re-planning. This framework unifies reactive and proactive failure detection via constraints, more generally abstracts relevant entities/parts through constraint elements, and ensures precise and real-time monitoring via code evaluation." + ], + "image_footnote": [], + "bbox": [ + 91, + 89, + 906, + 368 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "them in real-time with high precision. Herein we formulate both reactive and proactive failure detection as constraint satisfaction problems and use VLM-generated code to evaluate them, meeting the expectations of both modes above.", + "bbox": [ + 89, + 450, + 482, + 511 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Visual Prompting. Visual prompts enhance the visual reasoning abilities of VLMs and are categorized into mask-based, point-based, and element-based methods. Mask-based approaches like SoM [63] apply numbered segmentation masks on images without considering instructions, whereas instruction-guided methods like IVM [73] and API [65] generate masks to block irrelevant regions. Point-based prompting encodes functionalities via points, including affordances [29, 35, 67], motion [35, 41], and constraints like ReKep [21]. However, ReKep [21] extracts semantic keypoints using DINOv2 [42] and struggles to capture precise keypoints that accurately represent desired constraints. Element-based methods like CoPa [19], represent robotic control using basic elements (e.g., points, lines) via pre-trained models like SAM [27] and GPT-4V [1] simply. In contrast, this work explores constraint satisfaction monitoring and introduces constraint elements that more precisely represent relevant entities/parts. These elements are tracked and evaluated in real-time to simplify monitoring.", + "bbox": [ + 89, + 516, + 482, + 804 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Visual Programming. Visual programming requires strong visual concept understanding and reasoning. Prior works show strong generalization (e.g., zero-shot) across various tasks, such as image editing [8, 16], 3D visual grounding [68], and robotics control [34, 60], by integrating LLMs with visual modules but lose fine-grained details. Recent", + "bbox": [ + 89, + 810, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "studies [40, 61] explore visual programming using VLMs, but these methods take raw RGB images as inputs with predefined primitive libraries. In contrast, this work leverages constraint elements for visual programming and uses arithmetic operations in the code to encode their spatio-temporal combinational dynamics, presenting greater challenges.", + "bbox": [ + 511, + 450, + 906, + 542 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 513, + 556, + 604, + 571 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We first give an overview of the proposed Code-as-Monitor (CaM) (Sec. 3.1). Then, we elaborate on the constraint element in Constraint Painter, especially constraint-aware segmentation (Sec. 3.2). Finally, we present Constraint Monitor for real-time detection (Sec. 3.3).", + "bbox": [ + 511, + 584, + 905, + 659 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Overview", + "text_level": 1, + "bbox": [ + 511, + 671, + 624, + 686 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The proposed CaM comprises three key modules: the Constraint Generator, Painter, and Monitor. We focus on long-horizon manipulation task instructions $\\mathcal{L}_{\\mathrm{global}}$ (e.g., \"Move the pan with the bread to the stove, and be careful not to let the bread fall out\"), using RGB-D observations $\\mathcal{O}$ from two camera views (front and top). As shown in Fig. 2, the RGB images $\\mathcal{O}$ , along with instructions $\\mathcal{L}_{\\mathrm{global}}$ , previous subgoal $l_{\\mathrm{pre}}$ , and failure feedback from the Constraint Monitor $f_{\\mathrm{pre}}$ (e.g., subgoal success or failure reason), are fed into the Constraint Generator $\\mathcal{F}_{\\mathrm{VLM}}$ (i.e., GPT-4o [1]) to generate the next subgoal $l_{\\mathrm{next}}$ and associated textual constraints $\\mathcal{C}$ . This process can be formulated as follows:", + "bbox": [ + 511, + 696, + 906, + 876 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nl _ {\\text {n e x t}}, \\mathcal {C} _ {d}, \\mathcal {C} _ {u} = \\mathcal {F} _ {\\mathrm {V L M}} \\left(\\mathcal {O}, \\mathcal {L} _ {\\text {g l o b a l}}, l _ {\\text {p r e}}, f _ {\\text {p r e}}\\right) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 573, + 886, + 906, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "6921", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/886a1ff14757f6786cfc75f4e25d63c20b4fca4f0f7a5e4cf0e114601e8474ef.jpg", + "image_caption": [ + "Figure 3. Constraint Element Pipeline. Given a constraint, our model ConSeg generates instance-level and part-level masks across multiple views, which are projected into 3D space. Through a series of heuristics, the desired elements are produced. Once all elements are obtained, they are annotated onto the original multi-view images. Here we display the annotation result of one element." + ], + "image_footnote": [], + "bbox": [ + 96, + 89, + 903, + 224 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{C}_d = \\{c_d^0, c_d^1, \\ldots, c_d^n\\}$ denotes the constraints that must be maintained during subgoal execution (e.g., pan handle must be grasped, bread must remain in the pan, pan should remain horizontal during transfer), and $\\mathcal{C}_u = \\{c_u^0, c_u^1, \\ldots, c_u^k\\}$ refers to the constraints that must be met upon subgoal completion (e.g., pan should be directly above the stove). We successfully unify reactive and proactive failure detection as these task-specific, situation-aware constraint satisfaction problems.", + "bbox": [ + 89, + 282, + 483, + 417 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In Painter, for each textual constraint $c$ from $\\mathcal{C}_d$ or $\\mathcal{C}_u$ , we generate corresponding constraint elements $e$ (detailed in Sec. 3.2) from observations $\\mathcal{O}$ . These elements, which are composed of 3D points, abstract the constraint-related entities or their parts to represent the desired textual constraint more easily (e.g., the constant distance between green points on bread and pan determines if bread remains in pan, as shown in Fig. 2.) These generated elements are then aggregated into the final set $\\mathcal{E} = \\{e^0, e^1, \\ldots, e^{n + k}\\}$ , and numerically annotated across all views to produce the final visual prompted images $\\mathcal{O}_{\\mathcal{E}}$ .", + "bbox": [ + 88, + 420, + 483, + 585 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In Monitor, we provide GPT-4o [1] with the next subgoal $l_{\\mathrm{next}}$ , textual constraints $\\mathcal{C}$ , and annotated observations $\\mathcal{O}_{\\mathcal{E}}$ for constraint-aware visual programming to generate the evaluation protocol, i.e., monitor code. This code inputs the elements' 3D positions, calculates arithmetic operations within it, and returns a boolean to indicate potential or actual failure and a string to describe its reason. During subgoal execution, Monitor tracks the elements and evaluates the spatio-temporal combinational dynamics of these elements. If the code returns False, the policy execution halts immediately, and the accompanying string is used as feedback $f_{\\mathrm{pre}}$ for re-planning. Otherwise, the subgoal is considered completed. In either case, the cycle is repeated.", + "bbox": [ + 89, + 588, + 483, + 785 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Constraint Element", + "text_level": 1, + "bbox": [ + 89, + 800, + 277, + 814 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To simplify the monitoring of constraint satisfaction, we introduce constraint elements by abstracting constraint-related entities/parts into compact geometric elements (e.g., points, lines) to get rid of the most irrelevant geometric and visual details, making them easier to track and generalize.", + "bbox": [ + 89, + 825, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Pipeline. The constraint element generation pipeline is shown in Fig. 3. Our trained multi-granularity constraint-aware model ConSeg performs two inference steps for each $c$ and each RGB image $o$ from the set of views $O$ . First, instance-level masks $\\mathcal{M}_i$ are generated to capture constraint-related entities. Then, fine-grained part-level masks $\\mathcal{M}_p$ and corresponding element type descriptions $l_{\\mathrm{e}}$ are produced, as shown in Fig. 4. Using corresponding depth data, we project $\\mathcal{M}_p$ from all views into 3D space, fusing them into a point cloud. However, directly tracking and evaluating the spatio-temporal combinational dynamics of these constraint-related entities/parts is challenging. Therefore, we convert these entities into proposed constraint elements. We first apply voxelization to the point cloud with voxel size determined by element type $l_{\\mathrm{e}}$ (e.g., the surface needs at least 3 points, we divide the occupied space into $2\\times 2$ voxels.). Then we cluster one representative point within each voxel and filter them to a specified number, also determined by $l_{\\mathrm{e}}$ to extract the final desired 3D points. We connect points within each instance-level mask $\\mathcal{M}_i$ to form the constraint element $e$ associated with constraint $c$ . Notably, points modeled as end-effectors (e.g., the points of the fingertips and hand's center represent a dexterous hand) can be directly obtained from forward kinematics, bypassing the process above. Additionally, we perform parallel inference of ConSeg across all views to expedite the acquisition of the final set $\\mathcal{E}$ and their annotations onto the corresponding views $\\mathcal{O}_{\\mathcal{E}}$ . Moreover, this minimalist approach, i.e., constraint elements, emphasizes the most relevant entities/parts, enabling generalization to unseen scenes and entities, which is critical for open-set failure detection. For more details, please check Supp. B.3.", + "bbox": [ + 511, + 282, + 906, + 767 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Constraint-aware Segmentation. Since the textual constraint does not explicitly specify relevant entities/parts, we require a model capable of logical reasoning and certain open-set adaptation, enabling precise constraint-aware instance and part-level segmentation. As depicted in Fig. 4, we propose ConSeg, which builds upon LISA [28] comprising a VLM $\\mathcal{F}_{\\mathrm{VLM}}$ (i.e., LLaVA [36]) with the vision encoder $\\mathcal{F}_{\\mathrm{enc}}$ and decoder $\\mathcal{F}_{\\mathrm{dec}}$ from SAM [27]. The textual", + "bbox": [ + 511, + 780, + 908, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "6922", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/ca68e9310cb053ab41342b51c7c6ab82de32f443ec5dbcc4378e2116ec9e5ba0.jpg", + "image_caption": [ + "Response $(l_{\\mathrm{e}})$ : It needs a surface $<\\mathrm{SEG}>.$", + "Figure 4. ConSeg architecture. Here we display the part-level segmentation, which will output the desired element type and mask." + ], + "image_footnote": [], + "bbox": [ + 93, + 87, + 480, + 233 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "constraint $c$ and the image $o$ are input into the VLM $\\mathcal{F}_{\\mathrm{VLM}}$ to generate an embedding for the token. The VLM outputs a textual description $l_{\\mathrm{e}}$ of the desired element type additionally for part-level segmentation. The decoder $\\mathcal{F}_{\\mathrm{dec}}$ then generates the segmentation mask $\\mathcal{M}$ by leveraging visual features from the vision encoder $\\mathcal{F}_{\\mathrm{enc}}$ based on the current observation $o$ , and features derived from the final-layer embedding $h_{<\\mathrm{SEG}>}$ of the token, transformed via an MLP ( $\\mathcal{F}_{\\mathrm{MLP}}$ ). The process can be formulated as follows:", + "bbox": [ + 89, + 289, + 483, + 425 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nl _ {\\mathrm {e}}, h _ {< \\mathrm {S E G} >} = \\mathcal {F} _ {\\mathrm {V L M}} (c, o), \\mathcal {M} = \\mathcal {F} _ {\\mathrm {d e c}} (\\mathcal {F} _ {\\mathrm {M L P}} \\left(h _ {< \\mathrm {S E G} >}\\right), \\mathcal {F} _ {\\mathrm {e n c}} (o)) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 96, + 435, + 482, + 450 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The entire pipeline, including additional LoRA [17] parameters, is optimized end-to-end while keeping the parameters of the VLM $\\mathcal{F}_{\\mathrm{VLM}}$ and vision encoder $\\mathcal{F}_{\\mathrm{enc}}$ frozen.", + "bbox": [ + 89, + 455, + 482, + 502 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Dataset and Collection Pipeline. Our dataset utilizes images from BridgeData V2 [62]. While BridgeData V2 provides trajectory-level instructions, we require frame-level annotations for per-frame subgoals and constraints. To address this, we sample pick-and-place data, using external references (e.g., gripper open/close states) to generate 10,181 trajectories with 219,356 images. We employ GPT-4o to decompose trajectory instructions into subgoals, constraints, and object-part associations to generate groundtruth annotations. Instance-level and part-level segmentations are performed using Grounded SAM [49] and Semantic SAM [32], respectively. These outputs are integrated and refined through manual inspection to produce the multi-granularity dataset, which is combined with LISA's training data to fine-tune our model. More details are in Supp. B.", + "bbox": [ + 89, + 505, + 483, + 733 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Real-time Monitor Module", + "text_level": 1, + "bbox": [ + 89, + 741, + 333, + 756 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After annotating constraint elements with unique colors and numerical labels across all views, these images serve as visual prompts to generate monitor code via GPT-4o [1] with constraints $\\mathcal{C}$ and the next subgoal $l_{\\mathrm{next}}$ . The generated code accepts the elements' 3D positions as input and performs arithmetic operations (e.g., NumPy-supported calculations) to evaluate spatio-temporal constraints at various stages (i.e., during execution or upon completion), as shown in Fig. 2 code block. It returns a boolean flag indicating", + "bbox": [ + 89, + 763, + 482, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "potential or actual failure if the computed result exceeds a predefined tolerance (e.g., accepting deviations of the pan's surface normal from the z-axis within $15^{\\circ}$ ), along with a string explaining the cause. By tracking the elements using CoTracker [25], we achieve real-time detection via VLMs without frequent calls, reducing the computational cost.", + "bbox": [ + 511, + 90, + 903, + 181 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Notably, including both current positions and historical trajectories of the elements as input allows us to monitor tasks needing high precision, such as the requirement of $2\\mathrm{cm}$ movement or $180^{\\circ}$ rotation, which previous reactive detection methods could not address through direct VLM queries using raw RGB images. Moreover, the potential diversity in the structured associations of constraint elements within the code (e.g., calculating the angle between the normal vector of the pan's surface element and the z-axis determines if the pan is level, as shown in Fig. 2 code block) and rich prior knowledge offered by VLMs enable generalization to certain unseen tasks requiring common skills.", + "bbox": [ + 511, + 181, + 906, + 364 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 378, + 645, + 395 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our experiments aim to address the following questions: (1) Can CaM achieve open-set reactive and proactive failure detection across diverse robots, end-effectors, and objects, both in simulator (Sec. 4.2) and on real robots (Sec. 4.3)? (2) Can ConSeg infer multi-granularity constraint-aware masks for unseen scenes and objects (Sec. 4.4)? (3) Which design choices greatly enhance the performance (Sec. 4.5)?", + "bbox": [ + 511, + 404, + 906, + 510 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 511, + 521, + 702, + 537 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Environment Settings. We evaluate CaM across three simulators (i.e., CLIPort [54], Omnigibson [31], and RL-Bench [22]) and a real-world setting. In CLIPort, we employ a UR5 arm equipped with a suction cup for pick-and-place and a spatula for pushing, controlled by pre-trained CLIPort policy [54]. Omnigibson features a Fetch robot with a gripper, controlled by ReKep [21]. RLBench uses a Franka arm with a gripper, controlled by ARP [70]. We collect a small dataset to fine-tune ConSeg for each simulator. We use a UR5 arm with a Leap Hand [51] for real-world experiments, controlled by an open-loop policy named DexGraspNet 2.0 [69]. More details are provided in Supp. C.", + "bbox": [ + 511, + 544, + 905, + 727 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Task Settings. We introduce disturbances across all environments to induce failures, classified by the affected constraints (e.g., points, lines, surfaces). We evaluate task success rate, execution time, and additional token usage in Omnigibson. Notably, we don't assess failure judgment success rate because our code operates continuously in real-time, rendering this metric inapplicable. Detailed evaluation settings are provided in each section.", + "bbox": [ + 511, + 729, + 905, + 851 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Baselines. We compare DoReMi (DRM) [15] as the baseline for all settings, which uses VLMs to detect intermediate failures during execution via frequent VQA-based queries.", + "bbox": [ + 511, + 854, + 903, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "6923", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In CLIPort, we include Inner Monologue [20] baseline, which detects failures only upon each subgoal completion.", + "bbox": [ + 89, + 90, + 480, + 122 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Main Results in Simulator", + "text_level": 1, + "bbox": [ + 89, + 128, + 330, + 143 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2.1 Results in CLIPort", + "text_level": 1, + "bbox": [ + 89, + 152, + 277, + 166 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We evaluate two tasks in CLIPort: (1) Stack in Order: The robot must stack blocks in a specified order, including two point-level disturbances: (a) with per-step probability $p$ , the suction cup may release a block, causing it to drop; (b) placement positions are perturbed by uniform noise in $[0, q]$ cm, potentially leading to tower collapse. Success is defined as correctly stacking the blocks within 70s. (2) Sweep Half the Blocks: To address the pre-trained policy's tendency to sweep all blocks, the robot must determine when to halt, sweeping half of the blocks ( $\\pm 10\\%$ ) into a specified colored area within 30s. Success is achieved by meeting these criteria. Tab. 1 shows the mean results over 5 different seeds, each with 12 episodes. Check Supp. D.1 for details.", + "bbox": [ + 89, + 176, + 483, + 372 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Code better monitors 3D space relations. As shown in the Tab. 1, Under the most severe disturbances $(p = 0.3, q = 3)$ in \"Stack in Order\", CaM achieves a $17.5\\%$ higher success rate than DoReMi. Frequent VLM queries lead to increased incorrect failure judgments due to a limited 3D spatial understanding from single images, compared to code-based evaluation of block positions. For example, after placing the green block on the red and preparing to pick up the blue, DoReMi mistakenly thinks that the green block is no longer atop the red, causing it to re-execute the previous subgoal.", + "bbox": [ + 89, + 376, + 483, + 526 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Code with elements achieve reactive and proactive failure detection. As shown in Tab. 1, under severe disturbances in \"Stack in Order\", CaM reduces execution time by $38.7\\%$ and $14.4\\%$ compared to Inner Monologue [20], which only detects failures upon subgoal completion, and DoReMi [15], which suffers from misjudgments due to repeated VLM queries, respectively. In contrast, CaM unifies both reactive and proactive failure detection with high precision in real-time, especially in preventing foreseeable failures. For example, if the green block is placed on the red block with heavy placement position disturbance, CaM anticipates that further stacking the blue may lead to collapse and then stabilizes the green block before proceeding.", + "bbox": [ + 89, + 530, + 483, + 727 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Code with elements leads to more accurate counting. As shown in the Tab. 1, in \"Sweep Half the Blocks\", CaM achieves average success rates $4.5 \\times$ higher than DoReMi [15]. Calculating the block points in the designated surface region provides more accurate results than directly using the VLM to count, enabling a more precise halting of the policy to complete the task.", + "bbox": [ + 89, + 729, + 483, + 837 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2.2 Results in Omnigibson", + "text_level": 1, + "bbox": [ + 89, + 847, + 303, + 862 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We conduct experiments on three tasks in Omnigibson, each involving a distinct type of constraint-element disturbance:", + "bbox": [ + 89, + 869, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/d443fc67a5fc90ec15ed640490a77859089cbb39df46db8c5a51cafc8b8f0f77.jpg", + "table_caption": [ + "Table 1. Performance in CLIPort. We report the success rate and execution time, compared to CLIPort (CP) [54], with Inner Monologue (IM) [20] and DoReMi (DRM) [15]." + ], + "table_footnote": [], + "table_body": "
Tasks with disturbanceSuccess Rate(%) ↑Execution Time(s) ↓
CP+IM+DRM+Ours+IM+DRM+Ours
Stack in order with drop pp=0.0100.0100.0100.0100.013.413.413.4
p=0.1556.6781.6783.3395.0034.826.0021.0
p=0.321.6775.0076.6788.3342.834.2025.4
Stack in order with noise qq=190.0090.0096.6798.3324.824.624.2
q=241.6771.6775.0083.3339.437.029.2
q=315.0040.0040.0063.3358.254.236.8
Sweep Half the Blocks0.0018.3316.6775.0022.016.616.4
", + "bbox": [ + 519, + 133, + 895, + 232 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "(1) Slot Pen: Insert a pen into a holder, facing point-level disturbances wherein (a) pen is moved during grasping; (b) pen is dropped during transport; (c) holder is moved during insertion. (2) Stow Book: Place a book to a bookshelf vertically, with line-level disturbances where (a) book is randomly rotated during grasping; (b) end-effector joint is randomly actuated to alter the book pose; (c) book is reoriented horizontally after placement. (3) Pour Tea: Pours from a teapot into a teacup, encountering surface-level disturbances wherein (a) teapot is tilted forward/backward during movement; (b) end-effector joint is actuated to induce a lateral tilt of the teapot during movement; (c) teapot is returned to a horizontal position during pouring. Tab. 2 reports the results across three tasks, each including one no-disturbance trial and three specific-disturbance trials, with 10 runs for each setting. More details are provided in Supp. D.2.", + "bbox": [ + 511, + 241, + 906, + 484 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Code with elements detects richer failures. In Tab. 2, only CaM can detect failures caused by surface-level disturbances in \"Pour Tea\" compared to DoReMi [15]. The reason is that changes in the teapot's pitch and roll angles are hard to detect by querying VLM via VQA using one current image. The misjudgment leads to a $0\\%$ success rate of DoReMi in this task. Notably, DoReMi's success rate is sometimes lower than that of ReKep alone due to VLM's limited spatio-temporal understanding of single images.", + "bbox": [ + 511, + 488, + 908, + 625 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Code with elements detects failure with lower computational cost. As shown in Tab. 2, we achieve a $34.8\\%$ reduction in execution time and a $52.2\\%$ decrease in token count compared to DoReMi [15]. This improvement stems from proactive failure detection, which prevents more severe failures ahead in real-time for timely re-planning while generating code only once per subgoal. For example, in \"Pour Tea\", CaM detects failure by monitoring the teapot's tilt angle in real-time, avoiding frequent VLM checks.", + "bbox": [ + 511, + 628, + 908, + 765 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2.3 Results in RLBench", + "text_level": 1, + "bbox": [ + 511, + 779, + 705, + 792 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We further evaluate CaM on RLbench and demonstrate its superior generalization. Details can be found in Supp. D.3.", + "bbox": [ + 511, + 804, + 903, + 835 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Main Results in Real World", + "text_level": 1, + "bbox": [ + 511, + 845, + 763, + 861 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We conduct real-world evaluations on two tasks: (1) Simple Pick & Place: The robot has 70s to pick up objects and", + "bbox": [ + 511, + 869, + 906, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6924", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/b65cecf2cf416741b69ef18bf2d4a36ce958dced030908d853646d2f46173150.jpg", + "table_caption": [ + "Table 2. Performance in Omnigibson. We report the success rate (SR), execution time, and token usage, compared to DoReMi (DRM) [15]." + ], + "table_footnote": [], + "table_body": "
MethodSlot Pen (Point-level Disturbances)Stow Book (Line-level Disturbances)Pour Tea (Surface-level Disturbances)
SR with Disturbance(%) ↑Time (s)↓Token (k)↓SR with Disturbance(%) ↑Time (s)↓Token (k)↓SR with Disturbance(%) ↑Time (s)↓Token (k)↓
NoneDist.(a)Dist.(b)Dist.(c)NoneDist.(a)Dist.(b)Dist.(c)NoneDist.(a)Dist.(b)Dist.(c)
ReKep [21]30201010--40303020--20202010--
+DRM40102020177.8454.5450402040127.1738.670000--
+Ours60504040101.8525.827060706093.0818.6750404030174.5544.19
", + "bbox": [ + 91, + 106, + 903, + 172 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/660a001cf7222d7caab60e84243d3fb50d8e79091238225a2bfc7a625d82ef67.jpg", + "image_caption": [ + "Reasoning Pick &Place Task: Grasp the animals according to their distances to the fruits, from nearest to farthest." + ], + "image_footnote": [], + "bbox": [ + 96, + 196, + 196, + 277 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/30b32e9acf971d7764d35087851fed9ed881248853e32b716bcbcde1517ab5b1.jpg", + "image_caption": [ + "Horse is moved, target changed" + ], + "image_footnote": [], + "bbox": [ + 197, + 196, + 297, + 277 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d36186f6b5c1dbf886a26c79c2abd776300562c00a411ea8e9fd516ebaf0568a.jpg", + "image_caption": [ + "Figure 5. Example of Real-world Evaluation. The red bounding box shows the current grasp target, which may shift due to environmental changes. CaM monitors and adapts to these changes in real-time, resulting in a closed-loop system with an open-loop policy." + ], + "image_footnote": [], + "bbox": [ + 297, + 196, + 398, + 277 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a4097cc7255874f1477e174cb053a7d4c83bdb3dfcd789a82b0557e19377088b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 398, + 196, + 498, + 277 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d1fb0406b88c25c6813c4d294041583f54667d02849a04c591f1a43ba9a73e5f.jpg", + "image_caption": [ + "Pear is moved, target changed" + ], + "image_footnote": [], + "bbox": [ + 498, + 196, + 598, + 277 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/fa256f96746d862e29ef9c4138ce178c58e1a22d6c4e45d2c0950b354aa0defd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 599, + 196, + 699, + 277 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c76d3c8fea3c5410202a403736f70d6009c391076cdb5eb7b23c8b4bd590b806.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 699, + 196, + 800, + 277 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/928f0c7cf2182f87c7deebdbee23ac940cb16c7e1dd3d312dadae4ec1986a1b0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 800, + 196, + 901, + 277 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/7669d1805c87750ba7e38ffd591dfb1ebc2ec5fd04eecfb9b35bcaf1b2957fc0.jpg", + "table_caption": [ + "Table 3. Performance of Single Pick & Place. We report the success rate and execution time. DGN donates DexGraspNet 2.0 [69]." + ], + "table_footnote": [], + "table_body": "
Tasks with disturbanceObject typesSuccess Rate(%) ↑Execution Time(s) ↓
DGN+DRM+Ours+DRM+Ours
Pick & PlaceDeformable0.0083.3396.6761.846.3
with theTransparent0.0066.6793.3372.648.1
objects beingSmall Rigid0.0080.096.6765.745.4
movedLarge Geometric0.0086.6796.6768.945.3
Pick & PlaceDeformable0.0076.6793.3368.762.5
with theTransparent0.0060.0090.0077.762.7
objects beingSmall Rigid0.0063.3393.3369.860.5
removedLarge Geometric0.0076.6796.6772.360.3
", + "bbox": [ + 91, + 356, + 478, + 465 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/50b0ccaca0a87798ed5ec621f376ea15ee4670ced9e862f960d88b454e6f4e04.jpg", + "table_caption": [ + "Table 4. Performance of Reasoning Pick & Place in cluttered scene. We report the success rate. The robot is controlled by an open-loop policy named DexGraspNet 2.0 (DGN) [69]. w/ CE with $\\checkmark$ indicates using constraint elements; otherwise, constraint-aware entities or parts are used for tracking and code computation." + ], + "table_footnote": [], + "table_body": "
Tasksw/ CESuccess Rate(%) ↑
DGN+DRM+Ours
Clear all objects on table except for animals✓0.0010.0060.00
X0.0010.0020.00
Grasp the animals according to their distances to fruits, from nearest to farthest✓0.000.0090.00
", + "bbox": [ + 96, + 545, + 470, + 614 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "place them at specified locations, facing two disturbances: (a) moving the object during grasping, and (b) removing the object from hand during movement. We test four object types, selecting 3 examples per type and conducting 10 trials for each (see Tab. 3). (2) Reasoning Pick & Place: The robot executes long-horizon tasks, involving ambiguous terms (e.g., \"fruit\"), under the same disturbances. We evaluate 2 long-horizon tasks in cluttered scenes, performing 10 trials each (see Tab. 4). Check Supp. D.4 for details.", + "bbox": [ + 88, + 622, + 482, + 758 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Elements generally abstract constraints and relevant entities. As shown in Tab. 3, using a different end-effector (i.e., Leap Hand [51]) in Simple Pick & Place, CaM also achieves success rates surpassing DoReMi [15] by $20.4\\%$ when handling different kinds of objects (e.g., deformable). We find that abstracting constraint-related entities/parts removes the irrelevant visual details, enabling generalization to different kinds of entities in unseen scenes (as discussed in Sec. 3.2), leading to easier tracking and code evaluation.", + "bbox": [ + 89, + 763, + 483, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/8e367f57950329b95acdab592a44a39605e835fdb4990b1d71bf4d6550bb3131.jpg", + "table_caption": [ + "Table 5. Performance of reasoning and constraint-aware segmentation. FMC denotes the foundation model combination baseline." + ], + "table_footnote": [], + "table_body": "
MethodReasonSegConstraintSeg
Instance-levelInstance-levelPart-level
gIoUcIoUgIoUcIoUgIoUcIoU
PixelLM [50]56.061.444.443.224.122.6
LISA-13B [28]56.665.142.138.923.424.3
FMC51.253.349.349.640.839.3
ConSeg-13B55.763.962.168.760.265.3
", + "bbox": [ + 529, + 356, + 883, + 434 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/895bd2b94ff4532079a96a21cfb34999501520c9a74ae9bdf9ddf2e8b25000a9.jpg", + "table_caption": [ + "Table 6. Ablation studies in Omnigibson's \"Stow Book\", assessing the impact of Multi-View (MV), Constraint-aware Segmentation (CS) for elements, and Connect Points (CP) for element formation." + ], + "table_footnote": [], + "table_body": "
MVCSCPSuccess Rate with Disturbance(%) ↑
NoneDist.(a)Dist.(b)Dist.(c)Avg
X✓✓40.040.030.050.040.0
✓X✓50.040.040.040.042.5
✓✓X60.050.060.050.055.0
✓✓✓70.060.070.060.065.0
", + "bbox": [ + 527, + 487, + 885, + 554 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Reactive and Proactive failure detection combined with an open-loop policy forms a close-loop system. Tab. 4 shows that only $CaM$ successfully handles long-horizon tasks in cluttered scenes. These tasks are challenging as the robot is controlled by an open-loop policy that cannot handle environment dynamics and human disturbances in a closed-loop manner. By incorporating both reactive and proactive failure detection with the open-loop policy (see Fig. 5), the robot can dynamically adjust its target in real-time. For example, when a human moves the horse or pear during the task, the robot adapts by grasping the animal closest to the fruit, forming a closed-loop system.", + "bbox": [ + 511, + 561, + 906, + 743 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Main Results of Segmentation", + "text_level": 1, + "bbox": [ + 511, + 755, + 779, + 771 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Tab. 5 presents segmentation results comparing our ConSeg, with SOTA models and a Foundation Model Combination (FMC) baseline. FMC integrates GPT-4o for reasoning over tasks and constraints to identify relevant instances and parts, along with Grounded SAM [49] and Semantic SAM [32] for instance and part-level segmentation, respectively, similar to our data collection pipeline. We evaluate performance on the ReasonSeg [28] benchmark and our", + "bbox": [ + 511, + 780, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "6925", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/dafd99b9a9b9a1c2b9729ad2ca841eddfb7e36588fc6fc73c4eb0b05070bf488.jpg", + "image_caption": [ + "Figure 6. Visual comparison between our ConSeg and LISA [28] at instance and part level. The red masks are the segmentation results." + ], + "image_footnote": [], + "bbox": [ + 94, + 89, + 903, + 311 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/62004fe47f27a58d3ee66f85584ac2bc5b4ec40934198e067fded82a97219a2d.jpg", + "table_caption": [ + "Table 7. Ablation study on training data. SemanticSeg includes ADE20K [75], COCO-Stuff [5], PACO-LVIS [48] and PASCALPart [6]. ReferSeg includes refCLEF, refCOCO, refCOCO+ [26] and refCOCOg [39]. VQA indicates LLaVAInstruct-150k [36]." + ], + "table_footnote": [], + "table_body": "
SemanticSegReferSegVQAReasonSegConstraintSegIns-level gIoUcloUPart-level gIoUcloU
✓✓✓✓✘✘42.138.923.424.3
✓✓✓✓✓✘60.765.940.445.6
✓✓✓✓✘✓51.550.656.561.7
✓✓✓✓✓✓62.168.760.265.3
", + "bbox": [ + 94, + 398, + 473, + 465 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "proposed Constraint-Aware Segmentation (ConstraintSeg) benchmark. Our benchmark evaluates performance using gIoU and cIoU, following ReasonSeg's evaluation setting.", + "bbox": [ + 89, + 474, + 482, + 520 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ConSeg performs both reasoning and multi-granularity constraint-aware segmentation. As shown in Tab. 5, ConSeg performs comparably to LISA and PixelLM on the ReasonSeg but significantly surpasses them on ConstraintSeg, achieving nearly a $40\\%$ improvement at the part level. Visual comparisons in Fig. 6 further demonstrate ConSeg's strong generalization to unseen objects, scenes, and tasks.", + "bbox": [ + 89, + 523, + 483, + 630 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 641, + 243, + 657 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We conduct ablation studies and present analyses below.", + "bbox": [ + 89, + 665, + 464, + 680 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Multi views are critical for visual programming. As shown in Tab. 6, using only front-view images with constraint elements reduces the average success rate from $65\\%$ to $40\\%$ . This is primarily due to (1) occlusion leading to suboptimal element generation and (2) errors in visual programming, where dimensional reduction causes surfaces to be misinterpreted as lines or lines as points.", + "bbox": [ + 89, + 684, + 482, + 790 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Constraint-aware segmentation enhances element. As shown in Tab. 6, using DINOv2 [42] to generate semantic points to form elements, rather than through constraint-aware segmentation, reduces the success rate from $65\\%$ to $42.5\\%$ , because these constructed elements fail to represent the desired constraints accurately. For example, capturing a book's vertical orientation requires two precise points on its", + "bbox": [ + 89, + 795, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "edge, which DINOv2 can not provide.", + "bbox": [ + 511, + 340, + 767, + 356 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Forming elements improves code generation. As shown in Tab. 6, using unconnected 3D points as final elements reduces the success rate from $65\\%$ to $55\\%$ . Pre-formed elements contain more prior constraint information, serving as visual cues better encoded in code, whereas standalone points require recombination during visual programming.", + "bbox": [ + 511, + 363, + 905, + 455 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Elements simplify constraints computation. As shown in Tab. 4, the success rate decreases from $60\\%$ to $20\\%$ when entities or parts are not transformed into proposed elements in real-world evaluation. We find two key issues: (1) inaccurate 3D position and slow pose tracking of entities or parts; and (2) difficulties in performing arithmetic operations on their 3D positions or poses in code, hindering the encoding of their spatio-temporal constraints.", + "bbox": [ + 511, + 460, + 906, + 583 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Both instance-level and part-level data improve performance. We conduct ablation studies to assess the impact of our proposed dataset on the ConstraintSeg benchmark. The results are shown in Tab. 7. The results indicate that both the instance-level and part-level subsets contribute to performance improvements and enhance each other's effects.", + "bbox": [ + 511, + 589, + 906, + 679 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 705, + 633, + 720 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we present a novel paradigm termed Code-as-Monitor, leveraging the VLMs for both open-set reactive and proactive failure detection. In detail, We formulate both detection modes as spatio-temporal constraint satisfaction problems and use VLM-generated code to evaluate them for real-time monitoring. We further propose constraint elements, which abstract constraints-related entities or their parts into compact geometric elements, to improve the precision and efficiency of monitoring. Extensive experiments demonstrate the superiority of the proposed approach and highlight its potential to advance closed-loop robot systems.", + "bbox": [ + 511, + 734, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "6926", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 91, + 90, + 250, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported by the National Natural Science Foundation of China (62132001), the National Science and Technology Major Project (2022ZD0116314), and the Fundamental Research Funds for the Central Universities.", + "bbox": [ + 89, + 114, + 483, + 176 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 190, + 187, + 207 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv, 2023. 2, 3, 4, 5", + "[2] Aneseh Alvanpour, Sumit Kumar Das, Christopher Kevin Robinson, Olfa Nasraoui, and Dan Popa. Robot failure mode prediction with explainable machine learning. CASE, 2020. 2", + "[3] Jingkun An, Yinghao Zhu, Zongjian Li, Enshen Zhou, Haoran Feng, Xijie Huang, Bohua Chen, Yemin Shi, and Chengwei Pan. Agfpsync: Leveraging ai-generated feedback for preference optimization in text-to-image generation. arXiv preprint arXiv:2403.13352, 2024. 2", + "[4] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. *NeurIPS*, 2020. 2", + "[5] Holger Caesar, Jasper Uijlings, and Vittorio Ferrari. Cocostuff: Thing and stuff classes in context. CVPR, 2018. 8", + "[6] Xianjie Chen, Roozbeh Mottaghi, Xiaobai Liu, Sanja Fidler, Raquel Urtasun, and Alan Yuille. Detect what you can: Detecting and representing objects using holistic models and body parts. CVPR, 2014. 8", + "[7] Zeren Chen, Zhelun Shi, Xiaoya Lu, Lehan He, Sucheng Qian, Zhenfei Yin, Wanli Ouyang, Jing Shao, Yu Qiao, Cewu Lu, et al. Rh20t-p: A primitive-level robotic dataset towards composable generalization agents. arXiv preprint arXiv:2403.19622, 2024. 2", + "[8] Jaemin Cho, Abhay Zala, and Mohit Bansal. Visual programming for step-by-step text-to-image generation and evaluation. NeurIPS, 2024. 3", + "[9] Wenbo Cui, Chengyang Zhao, Songlin Wei, Jiazhao Zhang, Haoran Geng, Yaran Chen, and He Wang. Gapartmanip: A large-scale part-centric dataset for material-agnostic articulated object manipulation. arXiv preprint arXiv:2411.18276, 2024. 2", + "[10] Yinpei Dai, Jayjun Lee, Nima Fazeli, and Joyce Chai. Racer: Rich language-guided failure recovery policies for imitation learning. arXiv, 2024. 2", + "[11] Maximilian Diehl and Karinne Ramirez-Amaro. A causal-based approach to explain, predict and prevent failures in robotic tasks. RAS, 2023. 2", + "[12] Yuqing Du, Ksenia Konyushkova, Misha Denil, Akhil Raju, Jessica Landon, Felix Hill, Nando de Freitas, and Serkan Cabi. Vision-language models as success detectors. arXiv, 2023. 2", + "[13] Jiafei Duan, Wilbert Pumacay, Nishanth Kumar, Yi Ru Wang, Shulin Tian, Wentao Yuan, Ranjay Krishna, Dieter" + ], + "bbox": [ + 93, + 215, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Fox, Ajay Mandlekar, and Yijie Guo. Aha: A vision-language-model for detecting and reasoning over failures in robotic manipulation. arXiv, 2024. 2", + "[14] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv, 2024. 2", + "[15] Yanjiang Guo, Yen-Jen Wang, Lihan Zha, Zheyuan Jiang, and Jianyu Chen. Doremi: Grounding language model by detecting and recovering from plan-execution misalignment. arXiv, 2023. 2, 5, 6, 7", + "[16] Tanmay Gupta and Aniruddha Kembhavi. Visual programming: Compositional visual reasoning without training. CVPR, 2023. 3", + "[17] Edward J Hu, Phillip Wallis, Zeyuan Allen-Zhu, Yanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. *ICLR*, 2021. 5", + "[18] Xuhao Hu, Dongrui Liu, Hao Li, Xuanjing Huang, and Jing Shao. Vlsbench: Unveiling visual leakage in multimodal safety. arXiv preprint arXiv:2411.19939, 2024. 2", + "[19] Haoxu Huang, Fanqi Lin, Yingdong Hu, Shengjie Wang, and Yang Gao. Copa: General robotic manipulation through spatial constraints of parts with foundation models. arXiv, 2024. 3", + "[20] Wenlong Huang, Fei Xia, Ted Xiao, Harris Chan, Jacky Liang, Pete Florence, Andy Zeng, Jonathan Tompson, Igor Mordatch, Yevgen Chebotar, et al. Inner monologue: Embodied reasoning through planning with language models. arXiv, 2022. 2, 6", + "[21] Wenlong Huang, Chen Wang, Yunzhu Li, Ruohan Zhang, and Li Fei-Fei. Rekep: Spatio-temporal reasoning of relational keypoint constraints for robotic manipulation. arXiv, 2024. 3, 5, 7", + "[22] Stephen James, Zicong Ma, David Rovick Arrojo, and Andrew J. Davison. Rlbench: The robot learning benchmark & learning environment. RAL, 2020. 2, 5", + "[23] Yuheng Ji, Huajie Tan, Jiayu Shi, Xiaoshuai Hao, Yuan Zhang, Hengyuan Zhang, Pengwei Wang, Mengdi Zhao, Yao Mu, Pengju An, et al. Robobrain: A unified brain model for robotic manipulation from abstract to concrete. arXiv preprint arXiv:2502.21257, 2025. 2", + "[24] Yueru Jia, Jiaming Liu, Sixiang Chen, Chenyang Gu, Zhilue Wang, Longzan Luo, Lily Lee, Pengwei Wang, Zhongyuan Wang, Renrui Zhang, et al. Lift3d foundation policy: Lifting 2d large-scale pretrained models for robust 3d robotic manipulation. arXiv preprint arXiv:2411.18623, 2024. 2", + "[25] Nikita Karaev, Ignacio Rocco, Benjamin Graham, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. Co-tracker: It is better to track together. ECCV, 2024. 5", + "[26] Sahar Kazemzadeh, Vicente Ordonez, Mark Matten, and Tamara Berg. Referitag: Referring to objects in photographs of natural scenes. EMNLP, 2014. 8", + "[27] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. ICCV, 2023. 3, 4" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "6927", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[28] Xin Lai, Zhuotao Tian, Yukang Chen, Yanwei Li, Yuhui Yuan, Shu Liu, and Jiaya Jia. Lisa: Reasoning segmentation via large language model. CVPR, 2024. 4, 7, 8", + "[29] Olivia Y Lee, Annie Xie, Kuan Fang, Karl Pertsch, and Chelsea Finn. Affordance-guided reinforcement learning via visual prompting. arXiv, 2024. 3", + "[30] Gregory LeMasurier, Alvika Gautam, Zhao Han, Jacob W Crandall, and Holly A Yanco. Reactive or proactive? how robots should explain failures. HRI, 2024. 2", + "[31] Chengshu Li, Ruohan Zhang, Josiah Wong, Cem Gokmen, Sanjana Srivastava, Roberto Martin-Martin, Chen Wang, Gabrael Levine, Michael Lingelbach, Jiankai Sun, Mona Anvari, Minjune Hwang, Manasi Sharma, Arman Aydin, Dhruva Bansal, Samuel Hunter, Kyu-Young Kim, Alan Lou, Caleb R Matthews, Ivan Villa-Renteria, Jerry Huayang Tang, Claire Tang, Fei Xia, Silvio Savarese, Hyowon Gweon, Karen Liu, Jiajun Wu, and Li Fei-Fei. Behavior-1k: A benchmark for embodied ai with 1,000 everyday activities and realistic simulation. CoRL, 2022. 2, 5", + "[32] Feng Li, Hao Zhang, Peize Sun, Xueyan Zou, Shilong Liu, Jianwei Yang, Chunyuan Li, Lei Zhang, and Jianfeng Gao. Semantic-sam: Segment and recognize anything at any granularity. arXiv, 2023. 5, 7", + "[33] Lijun Li, Zhelun Shi, Xuhao Hu, Bowen Dong, Yiran Qin, Xihui Liu, Lu Sheng, and Jing Shao. T2isafety: Benchmark for assessing fairness, toxicity, and privacy in image generation. arXiv preprint arXiv:2501.12612, 2025. 2", + "[34] Jacky Liang, Wenlong Huang, Fei Xia, Peng Xu, Karol Hausman, Brian Ichter, Pete Florence, and Andy Zeng. Code as policies: Language model programs for embodied control. ICRA, 2023. 3", + "[35] Fangchen Liu, Kuan Fang, Pieter Abbeel, and Sergey Levine. Moka: Open-vocabulary robotic manipulation through mark-based visual prompting. arXiv, 2024. 3", + "[36] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. NeurIPS, 2024. 2, 4, 8", + "[37] Jiaming Liu, Chenxuan Li, Guanqun Wang, Lily Lee, Kaichen Zhou, Sixiang Chen, Chuyan Xiong, Jiaxin Ge, Renrui Zhang, and Shanghang Zhang. Self-corrected multimodal large language model for end-to-end robot manipulation. arXiv, 2024. 2", + "[38] Zeyi Liu, Arpit Bahety, and Shuran Song. Reflect: Summarizing robot experiences for failure explanation and correction. CoRL, 2023. 2", + "[39] Junhua Mao, Jonathan Huang, Alexander Toshev, Oana Camburu, Alan L Yuille, and Kevin Murphy. Generation and comprehension of unambiguous object descriptions. CVPR, 2016. 8", + "[40] Yao Mu, Junting Chen, Qinglong Zhang, Shoufa Chen, Qiaojun Yu, Chongjian Ge, Runjian Chen, Zhixuan Liang, Mengkang Hu, Chaofan Tao, et al. Robocodex: Multimodal code generation for robotic behavior synthesis. arXiv, 2024. 3", + "[41] Soroush Nasiriany, Fei Xia, Wenhao Yu, Ted Xiao, Jacky Liang, Ishita Dasgupta, Annie Xie, Danny Driess, Ayzaan Wahid, Zhuo Xu, et al. Pivot: Iterative visual prompting elicits actionable knowledge for vlms. arXiv, 2024. 3" + ], + "bbox": [ + 91, + 90, + 482, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[42] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv, 2023. 3, 8", + "[43] Zekun Qi, Wenyao Zhang, Yufei Ding, Runpei Dong, Xinqiang Yu, Jingwen Li, Lingyun Xu, Baoyu Li, Xialin He, Guofan Fan, et al. Sofar: Language-grounded orientation bridges spatial reasoning and object manipulation. arXiv preprint arXiv:2502.13143, 2025. 2", + "[44] Yiran Qin, Zhelun Shi, Jiwen Yu, Xijun Wang, Enshen Zhou, Lijun Li, Zhenfei Yin, Xihui Liu, Lu Sheng, Jing Shao, et al. Worldsimbench: Towards video generation models as world simulators. arXiv preprint arXiv:2410.18072, 2024. 2", + "[45] Yiran Qin, Enshen Zhou, Qichang Liu, Zhenfei Yin, Lu Sheng, Ruimao Zhang, Yu Qiao, and Jing Shao. Mp5: A multi-modal open-ended embodied system in mycraf via active perception. In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16307-16316. IEEE, 2024. 2", + "[46] Yiran Qin, Ao Sun, Yuze Hong, Benyou Wang, and Ruimao Zhang. Navigatediff: Visual predictors are zero-shot navigation assistants. arXiv preprint arXiv:2502.13894, 2025. 2", + "[47] Shreyas Sundara Raman, Vanya Cohen, Eric Rosen, Ifrah Idrees, David Paulius, and Stefanie Tellex. Planning with large language models via corrective re-prompting. NeurIPS, 2022. 2", + "[48] Vignesh Ramanathan, Anmol Kalia, Vladan Petrovic, Yi Wen, Baixue Zheng, Baishan Guo, Rui Wang, Aaron Marquez, Rama Kovvuri, Abhishek Kadian, et al. Paco: Parts and attributes of common objects. CVPR, 2023. 8", + "[49] Tianhe Ren, Shilong Liu, Ailing Zeng, Jing Lin, Kunchang Li, He Cao, Jiayu Chen, Xinyu Huang, Yukang Chen, Feng Yan, Zhaoyang Zeng, Hao Zhang, Feng Li, Jie Yang, Hongyang Li, Qing Jiang, and Lei Zhang. Grounded sam: Assembling open-world models for diverse visual tasks. arXiv, 2024. 5, 7", + "[50] Zhongwei Ren, Zhicheng Huang, Yunchao Wei, Yao Zhao, Dongmei Fu, Jiashi Feng, and Xiaojie Jin. Pixel reasoning with large multimodal model. CVPR, 2024. 7", + "[51] Kenneth Shaw, Ananye Agarwal, and Deepak Pathak. Leap hand: Low-cost, efficient, and anthropomorphic hand for robot learning. arXiv, 2023. 5, 7", + "[52] Zhelun Shi, Zhipin Wang, Hongxing Fan, Zaibin Zhang, Lijun Li, Yongting Zhang, Zhenfei Yin, Lu Sheng, Yu Qiao, and Jing Shao. Assessment of multimodal large language models in alignment with human values. arXiv preprint arXiv:2403.17830, 2024. 2", + "[53] Noah Shinn, Beck Labash, and Ashwin Gopinath. Reflection: an autonomous agent with dynamic memory and self-reflection. arXiv, 2023. 2", + "[54] Mohit Shridhar, Lucas Manuelli, and Dieter Fox. CIport: What and where pathways for robotic manipulation. CoRL, 2021. 2, 5, 6", + "[55] Tom Silver, Soham Dan, Kavitha Srinivas, Joshua B Tenenbaum, Leslie Kaelbling, and Michael Katz. Generalized planning in pddl domains with pretrained large language models. AAAI, 2024. 2" + ], + "bbox": [ + 516, + 90, + 903, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "6928", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[56] Marta Skreta, Naruki Yoshikawa, Sebastian Arellano-Rubach, Zhi Ji, Lasse Bjørn Kristensen, Kourosh Darvish, Alán Aspuru-Guzik, Florian Shkurti, and Animesh Garg. Errors are useful prompts: Instruction guided task programming with verifier-assisted iterative prompting. arXiv, 2023. 2", + "[57] Yingbo Tang, Shuaike Zhang, Xiaoshuai Hao, Pengwei Wang, Jianlong Wu, Zhongyuan Wang, and Shanghang Zhang. Affordgrasp: In-context affordance reasoning for open-vocabulary task-oriented grasping in clutter. arXiv preprint arXiv:2503.00778, 2025. 2", + "[58] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv, 2023. 2", + "[59] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv, 2023. 2", + "[60] Sai H Vemprala, Rogerio Bonatti, Arthur Tucker, and Ashish Kapoor. Chatgpt for robotics: Design principles and model abilities. IEEE Access, 2024. 3", + "[61] David Venuto, Sami Nur Islam, Martin Klissarov, Doina Precup, Sherry Yang, and Ankit Anand. Code as reward: Empowering reinforcement learning with vlms. arXiv, 2024. 3", + "[62] Homer Rich Walke, Kevin Black, Tony Z Zhao, Quan Vuong, Chongyi Zheng, Philippe Hansen-Estruch, Andre Wang He, Vivek Myers, Moo Jin Kim, Max Du, et al. Bridgedata v2: A dataset for robot learning at scale. CoRL, 2023. 5", + "[63] Jianwei Yang, Hao Zhang, Feng Li, Xueyan Zou, Chunyuan Li, and Jianfeng Gao. Set-of-mark prompting unleashes extraordinary visual grounding in gpt-4v. arXiv, 2023. 3", + "[64] Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, and Yuan Cao. React: Synergizing reasoning and acting in language models. arXiv, 2022. 2", + "[65] Runpeng Yu, Weihao Yu, and Xinchao Wang. Attention prompting on image for large vision-language models. arXiv, 2024. 3", + "[66] Zhouliang Yu, Jie Fu, Yao Mu, Chenguang Wang, Lin Shao, and Yaodong Yang. Multireact: Multimodal tools augmented reasoning-acting traces for embodied agent planning. *ICLR*, 2024. 2", + "[67] Wentao Yuan, Jiafei Duan, Valts Blukis, Wilbert Pumacay, Ranjay Krishna, Adithyavairavan Murali, Arsalan Mousavian, and Dieter Fox. Robopoint: A vision-language model for spatial affordance prediction for robotics. arXiv, 2024. 3", + "[68] Zhihao Yuan, Jinke Ren, Chun-Mei Feng, Hengshuang Zhao, Shuguang Cui, and Zhen Li. Visual programming for zero-shot open-vocabulary 3d visual grounding. CVPR, 2024. 3", + "[69] Jialiang Zhang, Haoran Liu, Danshi Li, XinQiang Yu, Hao ran Geng, Yufei Ding, Jiayi Chen, and He Wang. Dexgrasp-net 2.0: Learning generative dexterous grasping in largescale synthetic cluttered scenes. CoRL, 2024. 5, 7" + ], + "bbox": [ + 91, + 92, + 480, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[70] Xinyu Zhang, Yuhan Liu, Haonan Chang, Liam Schramm, and Abdeslam Boularias. Autoregressive action sequence learning for robotic manipulation. arXiv, 2024. 5", + "[71] Yongting Zhang, Lu Chen, Guodong Zheng, Yifeng Gao, Rui Zheng, Jinlan Fu, Zhenfei Yin, Senjie Jin, Yu Qiao, Xuanjing Huang, et al. Spa-vl: A comprehensive safety preference alignment dataset for vision language model. arXiv preprint arXiv:2406.12030, 2024. 2", + "[72] Zaibin Zhang, Shiyu Tang, Yuanhang Zhang, Talas Fu, Yifan Wang, Yang Liu, Dong Wang, Jing Shao, Lijun Wang, and Huchuan Lu. Ad-h: Autonomous driving with hierarchical agents. arXiv preprint arXiv:2406.03474, 2024. 2", + "[73] Jinliang Zheng, Jianxiong Li, Sijie Cheng, Yinan Zheng, Ji-aming Li, Jihao Liu, Yu Liu, Jingjing Liu, and Xianyuan Zhan. Instruction-guided visual masking. arXiv, 2024. 3", + "[74] Zhi Zheng, Qian Feng, Hang Li, Alois Knoll, and Jianxiang Feng. Evaluating uncertainty-based failure detection for closed-loop llm planners. arXiv, 2024. 2", + "[75] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. CVPR, 2017. 8", + "[76] Enshen Zhou, Yiran Qin, Zhenfei Yin, Yuzhou Huang, Ruimao Zhang, Lu Sheng, Yu Qiao, and Jing Shao. Minedreamer: Learning to follow instructions via chain-of-imagination for simulated-world control. arXiv preprint arXiv:2403.12037, 2024. 2" + ], + "bbox": [ + 516, + 92, + 903, + 458 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "6929", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Code-as-Monitor_ Constraint-aware Visual Programming for Reactive and Proactive Robotic Failure Detection/3086dac5-625c-4ed2-b32d-12ee6b15f252_model.json b/2025/Code-as-Monitor_ Constraint-aware Visual Programming for Reactive and Proactive Robotic Failure Detection/3086dac5-625c-4ed2-b32d-12ee6b15f252_model.json new file mode 100644 index 0000000000000000000000000000000000000000..c6306325a1b7dca0b1c2cddfa31a3a3aa98e3bfd --- /dev/null +++ b/2025/Code-as-Monitor_ Constraint-aware Visual Programming for Reactive and Proactive Robotic Failure Detection/3086dac5-625c-4ed2-b32d-12ee6b15f252_model.json @@ -0,0 +1,2521 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.19, + 0.131, + 0.812, + 0.175 + ], + "angle": 0, + "content": "Code-as-Monitor: Constraint-aware Visual Programming for Reactive and Proactive Robotic Failure Detection" + }, + { + "type": "text", + "bbox": [ + 0.158, + 0.203, + 0.845, + 0.257 + ], + "angle": 0, + "content": "Enshen Zhou\\(^{1,3*}\\), Qi Su\\(^{2,3,4*}\\), Cheng Chi\\(^{3*†}\\), Zhizheng Zhang\\(^{4}\\), Zhongyuan Wang\\(^{3}\\), Tiejun Huang\\(^{2,3}\\), Lu Sheng\\(^{1†}\\), He Wang\\(^{2,3,4†}\\), \\(^{1}\\)Beihang University \\(^{2}\\)Peking University \\(^{3}\\)Beijing Academy of Artificial Intelligence \\(^{4}\\)Galbot" + }, + { + "type": "text", + "bbox": [ + 0.133, + 0.261, + 0.865, + 0.275 + ], + "angle": 0, + "content": "zhouenshen@buaa.edu.cn chicheng15@mails.ucas.ac.cn lsheng@buaa.edu.cn hewang@pku.edu.cn" + }, + { + "type": "image", + "bbox": [ + 0.104, + 0.289, + 0.136, + 0.313 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.141, + 0.294, + 0.757, + 0.31 + ], + "angle": 0, + "content": "Task: Move the pan with the lobster to the stove, and be careful not to let the lobster drop out." + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.313, + 0.439, + 0.497 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.17, + 0.498, + 0.364, + 0.512 + ], + "angle": 0, + "content": "(a) Reactive Failure Detection" + }, + { + "type": "image", + "bbox": [ + 0.439, + 0.313, + 0.782, + 0.497 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.51, + 0.498, + 0.71, + 0.51 + ], + "angle": 0, + "content": "(b) Proactive Failure Detection" + }, + { + "type": "image", + "bbox": [ + 0.784, + 0.291, + 0.905, + 0.498 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.78, + 0.499, + 0.903, + 0.512 + ], + "angle": 0, + "content": "(c) Real-world Test" + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.517, + 0.907, + 0.618 + ], + "angle": 0, + "content": "Figure 1. For the task \"Move the pan with lobster to the stove without losing the lobster\", (a) reactive failure detection identifies failures after they occur, and (b) proactive failure detection prevents foreseeable failures. In (a), at \\( t_4^R \\), the robot detects the failure after the lobster unpredictably jumps out due to the heat. In (b), pan tilting is detected at \\( t_3^P \\) and corrected it at \\( t_3^{P'} \\), requiring real-time precision. We formulate both tasks as spatio-temporal constraint satisfaction problems, leveraging our proposed constraint elements for precise, real-time checking. For example, in (a), a large relative distance between pan and lobster indicates failure; in (b), a large angle between the pan and the horizontal plane needs correction. (c) shows that our method combined with an open-loop policy forms a closed-loop system, enabling proactive (e.g., detecting moving glass during grasping) and reactive (e.g., removing toy after grasping) failure detection in cluttered scenes." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.629, + 0.327, + 0.645 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.662, + 0.486, + 0.875 + ], + "angle": 0, + "content": "Automatic detection and prevention of open-set failures are crucial in closed-loop robotic systems. Recent studies often struggle to simultaneously identify unexpected failures reactively after they occur and prevent foreseeable ones proactively. To this end, we propose Code-as-Monitor (CaM), a novel paradigm leveraging the vision-language model (VLM) for both open-set reactive and proactive failure detection. The core of our method is to formulate both tasks as a unified set of spatio-temporal constraint satisfaction problems and use VLM-generated code to evaluate them for real-time monitoring. To enhance the accuracy and efficiency of monitoring, we further introduce constraint elements that abstract constraint-related entities or their parts into compact geometric ele" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.63, + 0.909, + 0.797 + ], + "angle": 0, + "content": "ments. This approach offers greater generality, simplifies tracking, and facilitates constraint-aware visual programming by leveraging these elements as visual prompts. Experiments show that CaM achieves a \\(28.7\\%\\) higher success rate and reduces execution time by \\(31.8\\%\\) under severe disturbances compared to baselines across three simulators and a real-world setting. Moreover, CaM can be integrated with open-loop control policies to form closed-loop systems, enabling long-horizon tasks in cluttered scenes with dynamic environments. See the project page at https://zhoues.github.io/Code-as-Monitor." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.814, + 0.646, + 0.83 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.909, + 0.903 + ], + "angle": 0, + "content": "As expectations grow for robots to handle long-horizon tasks within intricate environments, failures are unavoidable. Therefore, automatically detecting and preventing those failures plays a vital role in ensuring the tasks can" + }, + { + "type": "page_footnote", + "bbox": [ + 0.115, + 0.887, + 0.364, + 0.902 + ], + "angle": 0, + "content": "* Equal contribution † Corresponding author" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.958 + ], + "angle": 0, + "content": "6919" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.243 + ], + "angle": 0, + "content": "eventually be solved, especially for closed-loop robotic systems. There are two modes of failure detection [30], reactive and proactive. As depicted in Fig. 1, reactive failure detection aims to identify failures after they occur (e.g., recognizing that lobster lands on the table, indicating a delivery failure). In contrast, proactive failure detection aims to prevent foreseeable failures (e.g., recognizing that a tilted pan could cause the lobster to fall out, leading to a delivery failure). Both detection modes are even more challenging in open-set scenarios, where the failures are not predefined." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.246, + 0.482, + 0.457 + ], + "angle": 0, + "content": "With the help of large language models (LLMs) [14, 58] and vision-language models (VLMs) [1, 36], recent studies can achieve open-set reactive failure detection [10, 12, 13, 15, 20, 37, 55, 56, 66, 74] as a special case of visual question answering (VQA) tasks. However, these methods often bear compromised execution speeds and coarse-grained detection accuracy, due to the high computational costs and inadequate 3D spatio-temporal perception capability of recent LLMs/VLMs. Moreover, open-set proactive failure detection, which has been less explored in the literature, presents even severer challenges as it is required to foresee potential causes of failure and monitor them in real-time with high precision to anticipate and prevent imminent failure. Simply adapting LLMs/VLMs cannot meet such expectations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.46, + 0.483, + 0.702 + ], + "angle": 0, + "content": "In this work, we aim to develop an open-set failure detection framework that achieves reactive and proactive detection simultaneously, not only benefiting from the generalization power of VLMs but also enjoying high precision in monitoring failure characteristics with real-time efficiency. We address this by formulating both tasks as a unified set of spatio-temporal constraint satisfaction problems, which can be precisely translated by VLMs into executable programs. Such visual programs can efficiently verify whether entities (e.g., robots, objects) or their parts in the captured environment maintain or achieve required states during or after execution (i.e., satisfying constraints), so as to immediately prevent or detect failures. To the best of our knowledge, this is the first attempt to integrate both detection modes within a single framework. We name this constraint-aware visual programming framework as Code-as-Monitor \\((CaM)\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.483, + 0.901 + ], + "angle": 0, + "content": "To be specific, the proposed spatio-temporal constraint satisfaction scheme abstracts the constraint-related entity or part segments from the observed images into compact geometric elements (e.g., points, lines, and surfaces), as shown in Fig. 1. It simplifies the monitoring of constraint satisfaction by tracking and evaluating the spatio-temporal combinational dynamics of these elements, eliminating the most irrelevant geometric and visual details of the raw entities/parts. The constraint element detection and tracking are grounded by our trained constraint-aware segmentation and off-the-shelf tracking models, ensuring speed, accuracy, and certain open-set adaptation capabilities. The evaluation protocol is in the form of VLM-generated code, i.e., monitor" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.272 + ], + "angle": 0, + "content": "code, which is visually prompted by the starting frames of a sub-goal and its associated constraint elements, in addition to the textual constraints that must be fulfilled. Once generated, this monitor code could detect reactive or proactive failures just by being executed according to the tracked constraint elements, without needing to call the VLMs again. Therefore, this minimalist scheme is generalizable to open-set failure detection for unseen entities and scenes (enabled by the potential diversity of the structured associations of constraint elements) as well as common skills (powered by the rich prior knowledge offered by VLMs), maintaining sufficient detection accuracy and real-time execution speed." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.273, + 0.905, + 0.497 + ], + "angle": 0, + "content": "We conduct extensive experiments in three simulators (i.e., CLIPort [54], Omnigibson [31], and RL-Bench [22]) and one real-world setting, spanning diverse manipulation tasks (e.g., pick & place, articulated objects, tool-use), robot platforms (e.g., UR5, Fetch, Franka) and end-effectors (e.g., suction cup, gripper, dexterous hand). The results show that CaM is generalizable and achieves both reactive and proactive failure detection in real-time, resulting in \\(28.7\\%\\) higher success rates and reduced execution time by \\(31.8\\%\\) under severe disturbances compared to the baselines. Moreover, in Sec. 4.3, CaM can be integrated with the existing open-loop control policy to form a closed-loop system, enabling long-horizon tasks in cluttered scenes with environment dynamics and human disturbances. Our contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.5, + 0.904, + 0.543 + ], + "angle": 0, + "content": "- We introduce Code-as-Monitor (CaM), a novel paradigm that leverages VLMs for both reactive and proactive failure detection via constraint-aware visual programming." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.545, + 0.904, + 0.574 + ], + "angle": 0, + "content": "- We propose the constraint elements to enhance the accuracy and efficiency of constraint satisfaction monitoring." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.576, + 0.904, + 0.619 + ], + "angle": 0, + "content": "- Extensive experiments show that CaM is generalizable and achieves more real-time and precise failure detection than baselines across simulators and real-world settings." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.5, + 0.904, + 0.619 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.634, + 0.652, + 0.649 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.659, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Robotic Failure Detection. Recent advances in LLMs [4, 14, 58, 59] and VLMs [1, 3, 18, 33, 36, 44, 45, 52, 71, 76] greatly improve open-set reactive failure detection for robotics [7, 9, 23, 24, 43, 46, 57, 72] Current LLM-based methods either convert visual inputs into text [38, 53, 64], potentially losing visual details, or rely on ground-truth feedback [20, 47, 55, 56], which is impractical in real-world scenarios. Recent studies employ VLMs as failure detectors, offering binary success indicators [12, 37, 66, 74] or textual explanations [10, 13] through visual question answering (VQA), such as DoReMi [15]. However, they often bear compromised execution speeds and coarse-grained detection accuracy. Meanwhile, open-set proactive failure detection is rarely explored, as previous methods [2, 11] are confined to predefined failures. This detection mode requires foreseeing potential failure causes and monitoring" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "6920" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.093, + 0.09, + 0.907, + 0.369 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.373, + 0.907, + 0.444 + ], + "angle": 0, + "content": "Figure 2. Overview of Code-as-Monitor. Given task instructions and prior information, the Constraint Generator derives the next subgoal and corresponding textual constraints based on multi-view observations. The Painter maps these constraints onto images as constraint elements. The Monitor generates monitor code from these images and tracks them for real-time monitoring. If any constraint is violated, it outputs the reason for failure and triggers re-planning. This framework unifies reactive and proactive failure detection via constraints, more generally abstracts relevant entities/parts through constraint elements, and ensures precise and real-time monitoring via code evaluation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.452, + 0.483, + 0.512 + ], + "angle": 0, + "content": "them in real-time with high precision. Herein we formulate both reactive and proactive failure detection as constraint satisfaction problems and use VLM-generated code to evaluate them, meeting the expectations of both modes above." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.517, + 0.483, + 0.805 + ], + "angle": 0, + "content": "Visual Prompting. Visual prompts enhance the visual reasoning abilities of VLMs and are categorized into mask-based, point-based, and element-based methods. Mask-based approaches like SoM [63] apply numbered segmentation masks on images without considering instructions, whereas instruction-guided methods like IVM [73] and API [65] generate masks to block irrelevant regions. Point-based prompting encodes functionalities via points, including affordances [29, 35, 67], motion [35, 41], and constraints like ReKep [21]. However, ReKep [21] extracts semantic keypoints using DINOv2 [42] and struggles to capture precise keypoints that accurately represent desired constraints. Element-based methods like CoPa [19], represent robotic control using basic elements (e.g., points, lines) via pre-trained models like SAM [27] and GPT-4V [1] simply. In contrast, this work explores constraint satisfaction monitoring and introduces constraint elements that more precisely represent relevant entities/parts. These elements are tracked and evaluated in real-time to simplify monitoring." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Visual Programming. Visual programming requires strong visual concept understanding and reasoning. Prior works show strong generalization (e.g., zero-shot) across various tasks, such as image editing [8, 16], 3D visual grounding [68], and robotics control [34, 60], by integrating LLMs with visual modules but lose fine-grained details. Recent" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.452, + 0.907, + 0.543 + ], + "angle": 0, + "content": "studies [40, 61] explore visual programming using VLMs, but these methods take raw RGB images as inputs with predefined primitive libraries. In contrast, this work leverages constraint elements for visual programming and uses arithmetic operations in the code to encode their spatio-temporal combinational dynamics, presenting greater challenges." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.558, + 0.606, + 0.573 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.585, + 0.906, + 0.66 + ], + "angle": 0, + "content": "We first give an overview of the proposed Code-as-Monitor (CaM) (Sec. 3.1). Then, we elaborate on the constraint element in Constraint Painter, especially constraint-aware segmentation (Sec. 3.2). Finally, we present Constraint Monitor for real-time detection (Sec. 3.3)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.672, + 0.625, + 0.687 + ], + "angle": 0, + "content": "3.1. Overview" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.697, + 0.907, + 0.877 + ], + "angle": 0, + "content": "The proposed CaM comprises three key modules: the Constraint Generator, Painter, and Monitor. We focus on long-horizon manipulation task instructions \\(\\mathcal{L}_{\\mathrm{global}}\\) (e.g., \"Move the pan with the bread to the stove, and be careful not to let the bread fall out\"), using RGB-D observations \\(\\mathcal{O}\\) from two camera views (front and top). As shown in Fig. 2, the RGB images \\(\\mathcal{O}\\), along with instructions \\(\\mathcal{L}_{\\mathrm{global}}\\), previous subgoal \\(l_{\\mathrm{pre}}\\), and failure feedback from the Constraint Monitor \\(f_{\\mathrm{pre}}\\) (e.g., subgoal success or failure reason), are fed into the Constraint Generator \\(\\mathcal{F}_{\\mathrm{VLM}}\\) (i.e., GPT-4o [1]) to generate the next subgoal \\(l_{\\mathrm{next}}\\) and associated textual constraints \\(\\mathcal{C}\\). This process can be formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.574, + 0.887, + 0.907, + 0.902 + ], + "angle": 0, + "content": "\\[\nl _ {\\text {n e x t}}, \\mathcal {C} _ {d}, \\mathcal {C} _ {u} = \\mathcal {F} _ {\\mathrm {V L M}} \\left(\\mathcal {O}, \\mathcal {L} _ {\\text {g l o b a l}}, l _ {\\text {p r e}}, f _ {\\text {p r e}}\\right) \\tag {1}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "6921" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.09, + 0.904, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.229, + 0.908, + 0.272 + ], + "angle": 0, + "content": "Figure 3. Constraint Element Pipeline. Given a constraint, our model ConSeg generates instance-level and part-level masks across multiple views, which are projected into 3D space. Through a series of heuristics, the desired elements are produced. Once all elements are obtained, they are annotated onto the original multi-view images. Here we display the annotation result of one element." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.283, + 0.484, + 0.419 + ], + "angle": 0, + "content": "where \\(\\mathcal{C}_d = \\{c_d^0, c_d^1, \\ldots, c_d^n\\}\\) denotes the constraints that must be maintained during subgoal execution (e.g., pan handle must be grasped, bread must remain in the pan, pan should remain horizontal during transfer), and \\(\\mathcal{C}_u = \\{c_u^0, c_u^1, \\ldots, c_u^k\\}\\) refers to the constraints that must be met upon subgoal completion (e.g., pan should be directly above the stove). We successfully unify reactive and proactive failure detection as these task-specific, situation-aware constraint satisfaction problems." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.421, + 0.484, + 0.587 + ], + "angle": 0, + "content": "In Painter, for each textual constraint \\( c \\) from \\( \\mathcal{C}_d \\) or \\( \\mathcal{C}_u \\), we generate corresponding constraint elements \\( e \\) (detailed in Sec. 3.2) from observations \\( \\mathcal{O} \\). These elements, which are composed of 3D points, abstract the constraint-related entities or their parts to represent the desired textual constraint more easily (e.g., the constant distance between green points on bread and pan determines if bread remains in pan, as shown in Fig. 2.) These generated elements are then aggregated into the final set \\( \\mathcal{E} = \\{e^0, e^1, \\ldots, e^{n + k}\\} \\), and numerically annotated across all views to produce the final visual prompted images \\( \\mathcal{O}_{\\mathcal{E}} \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.589, + 0.484, + 0.786 + ], + "angle": 0, + "content": "In Monitor, we provide GPT-4o [1] with the next subgoal \\( l_{\\mathrm{next}} \\), textual constraints \\( \\mathcal{C} \\), and annotated observations \\( \\mathcal{O}_{\\mathcal{E}} \\) for constraint-aware visual programming to generate the evaluation protocol, i.e., monitor code. This code inputs the elements' 3D positions, calculates arithmetic operations within it, and returns a boolean to indicate potential or actual failure and a string to describe its reason. During subgoal execution, Monitor tracks the elements and evaluates the spatio-temporal combinational dynamics of these elements. If the code returns False, the policy execution halts immediately, and the accompanying string is used as feedback \\( f_{\\mathrm{pre}} \\) for re-planning. Otherwise, the subgoal is considered completed. In either case, the cycle is repeated." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.801, + 0.279, + 0.815 + ], + "angle": 0, + "content": "3.2. Constraint Element" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.484, + 0.901 + ], + "angle": 0, + "content": "To simplify the monitoring of constraint satisfaction, we introduce constraint elements by abstracting constraint-related entities/parts into compact geometric elements (e.g., points, lines) to get rid of the most irrelevant geometric and visual details, making them easier to track and generalize." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.283, + 0.908, + 0.768 + ], + "angle": 0, + "content": "Pipeline. The constraint element generation pipeline is shown in Fig. 3. Our trained multi-granularity constraint-aware model ConSeg performs two inference steps for each \\(c\\) and each RGB image \\(o\\) from the set of views \\(O\\). First, instance-level masks \\(\\mathcal{M}_i\\) are generated to capture constraint-related entities. Then, fine-grained part-level masks \\(\\mathcal{M}_p\\) and corresponding element type descriptions \\(l_{\\mathrm{e}}\\) are produced, as shown in Fig. 4. Using corresponding depth data, we project \\(\\mathcal{M}_p\\) from all views into 3D space, fusing them into a point cloud. However, directly tracking and evaluating the spatio-temporal combinational dynamics of these constraint-related entities/parts is challenging. Therefore, we convert these entities into proposed constraint elements. We first apply voxelization to the point cloud with voxel size determined by element type \\(l_{\\mathrm{e}}\\) (e.g., the surface needs at least 3 points, we divide the occupied space into \\(2\\times 2\\) voxels.). Then we cluster one representative point within each voxel and filter them to a specified number, also determined by \\(l_{\\mathrm{e}}\\) to extract the final desired 3D points. We connect points within each instance-level mask \\(\\mathcal{M}_i\\) to form the constraint element \\(e\\) associated with constraint \\(c\\). Notably, points modeled as end-effectors (e.g., the points of the fingertips and hand's center represent a dexterous hand) can be directly obtained from forward kinematics, bypassing the process above. Additionally, we perform parallel inference of ConSeg across all views to expedite the acquisition of the final set \\(\\mathcal{E}\\) and their annotations onto the corresponding views \\(\\mathcal{O}_{\\mathcal{E}}\\). Moreover, this minimalist approach, i.e., constraint elements, emphasizes the most relevant entities/parts, enabling generalization to unseen scenes and entities, which is critical for open-set failure detection. For more details, please check Supp. B.3." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.781, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Constraint-aware Segmentation. Since the textual constraint does not explicitly specify relevant entities/parts, we require a model capable of logical reasoning and certain open-set adaptation, enabling precise constraint-aware instance and part-level segmentation. As depicted in Fig. 4, we propose ConSeg, which builds upon LISA [28] comprising a VLM \\(\\mathcal{F}_{\\mathrm{VLM}}\\) (i.e., LLaVA [36]) with the vision encoder \\(\\mathcal{F}_{\\mathrm{enc}}\\) and decoder \\(\\mathcal{F}_{\\mathrm{dec}}\\) from SAM [27]. The textual" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6922" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.088, + 0.482, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.095, + 0.236, + 0.303, + 0.248 + ], + "angle": 0, + "content": "Response \\((l_{\\mathrm{e}})\\) : It needs a surface \\(<\\mathrm{SEG}>.\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.091, + 0.254, + 0.483, + 0.281 + ], + "angle": 0, + "content": "Figure 4. ConSeg architecture. Here we display the part-level segmentation, which will output the desired element type and mask." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.29, + 0.484, + 0.426 + ], + "angle": 0, + "content": "constraint \\(c\\) and the image \\(o\\) are input into the VLM \\(\\mathcal{F}_{\\mathrm{VLM}}\\) to generate an embedding for the token. The VLM outputs a textual description \\(l_{\\mathrm{e}}\\) of the desired element type additionally for part-level segmentation. The decoder \\(\\mathcal{F}_{\\mathrm{dec}}\\) then generates the segmentation mask \\(\\mathcal{M}\\) by leveraging visual features from the vision encoder \\(\\mathcal{F}_{\\mathrm{enc}}\\) based on the current observation \\(o\\), and features derived from the final-layer embedding \\(h_{<\\mathrm{SEG}>}\\) of the token, transformed via an MLP (\\(\\mathcal{F}_{\\mathrm{MLP}}\\)). The process can be formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.097, + 0.436, + 0.483, + 0.451 + ], + "angle": 0, + "content": "\\[\nl _ {\\mathrm {e}}, h _ {< \\mathrm {S E G} >} = \\mathcal {F} _ {\\mathrm {V L M}} (c, o), \\mathcal {M} = \\mathcal {F} _ {\\mathrm {d e c}} (\\mathcal {F} _ {\\mathrm {M L P}} \\left(h _ {< \\mathrm {S E G} >}\\right), \\mathcal {F} _ {\\mathrm {e n c}} (o)) \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.457, + 0.483, + 0.503 + ], + "angle": 0, + "content": "The entire pipeline, including additional LoRA [17] parameters, is optimized end-to-end while keeping the parameters of the VLM \\(\\mathcal{F}_{\\mathrm{VLM}}\\) and vision encoder \\(\\mathcal{F}_{\\mathrm{enc}}\\) frozen." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.506, + 0.484, + 0.734 + ], + "angle": 0, + "content": "Dataset and Collection Pipeline. Our dataset utilizes images from BridgeData V2 [62]. While BridgeData V2 provides trajectory-level instructions, we require frame-level annotations for per-frame subgoals and constraints. To address this, we sample pick-and-place data, using external references (e.g., gripper open/close states) to generate 10,181 trajectories with 219,356 images. We employ GPT-4o to decompose trajectory instructions into subgoals, constraints, and object-part associations to generate groundtruth annotations. Instance-level and part-level segmentations are performed using Grounded SAM [49] and Semantic SAM [32], respectively. These outputs are integrated and refined through manual inspection to produce the multi-granularity dataset, which is combined with LISA's training data to fine-tune our model. More details are in Supp. B." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.742, + 0.334, + 0.757 + ], + "angle": 0, + "content": "3.3. Real-time Monitor Module" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.902 + ], + "angle": 0, + "content": "After annotating constraint elements with unique colors and numerical labels across all views, these images serve as visual prompts to generate monitor code via GPT-4o [1] with constraints \\(\\mathcal{C}\\) and the next subgoal \\(l_{\\mathrm{next}}\\). The generated code accepts the elements' 3D positions as input and performs arithmetic operations (e.g., NumPy-supported calculations) to evaluate spatio-temporal constraints at various stages (i.e., during execution or upon completion), as shown in Fig. 2 code block. It returns a boolean flag indicating" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.182 + ], + "angle": 0, + "content": "potential or actual failure if the computed result exceeds a predefined tolerance (e.g., accepting deviations of the pan's surface normal from the z-axis within \\(15^{\\circ}\\)), along with a string explaining the cause. By tracking the elements using CoTracker [25], we achieve real-time detection via VLMs without frequent calls, reducing the computational cost." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.183, + 0.907, + 0.365 + ], + "angle": 0, + "content": "Notably, including both current positions and historical trajectories of the elements as input allows us to monitor tasks needing high precision, such as the requirement of \\(2\\mathrm{cm}\\) movement or \\(180^{\\circ}\\) rotation, which previous reactive detection methods could not address through direct VLM queries using raw RGB images. Moreover, the potential diversity in the structured associations of constraint elements within the code (e.g., calculating the angle between the normal vector of the pan's surface element and the z-axis determines if the pan is level, as shown in Fig. 2 code block) and rich prior knowledge offered by VLMs enable generalization to certain unseen tasks requiring common skills." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.379, + 0.646, + 0.396 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.405, + 0.907, + 0.511 + ], + "angle": 0, + "content": "Our experiments aim to address the following questions: (1) Can CaM achieve open-set reactive and proactive failure detection across diverse robots, end-effectors, and objects, both in simulator (Sec. 4.2) and on real robots (Sec. 4.3)? (2) Can ConSeg infer multi-granularity constraint-aware masks for unseen scenes and objects (Sec. 4.4)? (3) Which design choices greatly enhance the performance (Sec. 4.5)?" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.522, + 0.704, + 0.538 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.545, + 0.906, + 0.728 + ], + "angle": 0, + "content": "Environment Settings. We evaluate CaM across three simulators (i.e., CLIPort [54], Omnigibson [31], and RL-Bench [22]) and a real-world setting. In CLIPort, we employ a UR5 arm equipped with a suction cup for pick-and-place and a spatula for pushing, controlled by pre-trained CLIPort policy [54]. Omnigibson features a Fetch robot with a gripper, controlled by ReKep [21]. RLBench uses a Franka arm with a gripper, controlled by ARP [70]. We collect a small dataset to fine-tune ConSeg for each simulator. We use a UR5 arm with a Leap Hand [51] for real-world experiments, controlled by an open-loop policy named DexGraspNet 2.0 [69]. More details are provided in Supp. C." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.731, + 0.906, + 0.852 + ], + "angle": 0, + "content": "Task Settings. We introduce disturbances across all environments to induce failures, classified by the affected constraints (e.g., points, lines, surfaces). We evaluate task success rate, execution time, and additional token usage in Omnigibson. Notably, we don't assess failure judgment success rate because our code operates continuously in real-time, rendering this metric inapplicable. Detailed evaluation settings are provided in each section." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Baselines. We compare DoReMi (DRM) [15] as the baseline for all settings, which uses VLMs to detect intermediate failures during execution via frequent VQA-based queries." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6923" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.123 + ], + "angle": 0, + "content": "In CLIPort, we include Inner Monologue [20] baseline, which detects failures only upon each subgoal completion." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.13, + 0.331, + 0.145 + ], + "angle": 0, + "content": "4.2. Main Results in Simulator" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.153, + 0.279, + 0.167 + ], + "angle": 0, + "content": "4.2.1 Results in CLIPort" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.177, + 0.484, + 0.373 + ], + "angle": 0, + "content": "We evaluate two tasks in CLIPort: (1) Stack in Order: The robot must stack blocks in a specified order, including two point-level disturbances: (a) with per-step probability \\( p \\), the suction cup may release a block, causing it to drop; (b) placement positions are perturbed by uniform noise in \\([0, q]\\) cm, potentially leading to tower collapse. Success is defined as correctly stacking the blocks within 70s. (2) Sweep Half the Blocks: To address the pre-trained policy's tendency to sweep all blocks, the robot must determine when to halt, sweeping half of the blocks (\\(\\pm 10\\%\\)) into a specified colored area within 30s. Success is achieved by meeting these criteria. Tab. 1 shows the mean results over 5 different seeds, each with 12 episodes. Check Supp. D.1 for details." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.377, + 0.484, + 0.527 + ], + "angle": 0, + "content": "Code better monitors 3D space relations. As shown in the Tab. 1, Under the most severe disturbances \\((p = 0.3, q = 3)\\) in \"Stack in Order\", CaM achieves a \\(17.5\\%\\) higher success rate than DoReMi. Frequent VLM queries lead to increased incorrect failure judgments due to a limited 3D spatial understanding from single images, compared to code-based evaluation of block positions. For example, after placing the green block on the red and preparing to pick up the blue, DoReMi mistakenly thinks that the green block is no longer atop the red, causing it to re-execute the previous subgoal." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.531, + 0.484, + 0.728 + ], + "angle": 0, + "content": "Code with elements achieve reactive and proactive failure detection. As shown in Tab. 1, under severe disturbances in \"Stack in Order\", CaM reduces execution time by \\(38.7\\%\\) and \\(14.4\\%\\) compared to Inner Monologue [20], which only detects failures upon subgoal completion, and DoReMi [15], which suffers from misjudgments due to repeated VLM queries, respectively. In contrast, CaM unifies both reactive and proactive failure detection with high precision in real-time, especially in preventing foreseeable failures. For example, if the green block is placed on the red block with heavy placement position disturbance, CaM anticipates that further stacking the blue may lead to collapse and then stabilizes the green block before proceeding." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.731, + 0.484, + 0.838 + ], + "angle": 0, + "content": "Code with elements leads to more accurate counting. As shown in the Tab. 1, in \"Sweep Half the Blocks\", CaM achieves average success rates \\(4.5 \\times\\) higher than DoReMi [15]. Calculating the block points in the designated surface region provides more accurate results than directly using the VLM to count, enabling a more precise halting of the policy to complete the task." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.848, + 0.304, + 0.863 + ], + "angle": 0, + "content": "4.2.2 Results in Omnigibson" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.901 + ], + "angle": 0, + "content": "We conduct experiments on three tasks in Omnigibson, each involving a distinct type of constraint-element disturbance:" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.09, + 0.905, + 0.131 + ], + "angle": 0, + "content": "Table 1. Performance in CLIPort. We report the success rate and execution time, compared to CLIPort (CP) [54], with Inner Monologue (IM) [20] and DoReMi (DRM) [15]." + }, + { + "type": "table", + "bbox": [ + 0.52, + 0.135, + 0.897, + 0.233 + ], + "angle": 0, + "content": "
Tasks with disturbanceSuccess Rate(%) ↑Execution Time(s) ↓
CP+IM+DRM+Ours+IM+DRM+Ours
Stack in order with drop pp=0.0100.0100.0100.0100.013.413.413.4
p=0.1556.6781.6783.3395.0034.826.0021.0
p=0.321.6775.0076.6788.3342.834.2025.4
Stack in order with noise qq=190.0090.0096.6798.3324.824.624.2
q=241.6771.6775.0083.3339.437.029.2
q=315.0040.0040.0063.3358.254.236.8
Sweep Half the Blocks0.0018.3316.6775.0022.016.616.4
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.242, + 0.907, + 0.485 + ], + "angle": 0, + "content": "(1) Slot Pen: Insert a pen into a holder, facing point-level disturbances wherein (a) pen is moved during grasping; (b) pen is dropped during transport; (c) holder is moved during insertion. (2) Stow Book: Place a book to a bookshelf vertically, with line-level disturbances where (a) book is randomly rotated during grasping; (b) end-effector joint is randomly actuated to alter the book pose; (c) book is reoriented horizontally after placement. (3) Pour Tea: Pours from a teapot into a teacup, encountering surface-level disturbances wherein (a) teapot is tilted forward/backward during movement; (b) end-effector joint is actuated to induce a lateral tilt of the teapot during movement; (c) teapot is returned to a horizontal position during pouring. Tab. 2 reports the results across three tasks, each including one no-disturbance trial and three specific-disturbance trials, with 10 runs for each setting. More details are provided in Supp. D.2." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.489, + 0.909, + 0.625 + ], + "angle": 0, + "content": "Code with elements detects richer failures. In Tab. 2, only CaM can detect failures caused by surface-level disturbances in \"Pour Tea\" compared to DoReMi [15]. The reason is that changes in the teapot's pitch and roll angles are hard to detect by querying VLM via VQA using one current image. The misjudgment leads to a \\(0\\%\\) success rate of DoReMi in this task. Notably, DoReMi's success rate is sometimes lower than that of ReKep alone due to VLM's limited spatio-temporal understanding of single images." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.629, + 0.909, + 0.766 + ], + "angle": 0, + "content": "Code with elements detects failure with lower computational cost. As shown in Tab. 2, we achieve a \\(34.8\\%\\) reduction in execution time and a \\(52.2\\%\\) decrease in token count compared to DoReMi [15]. This improvement stems from proactive failure detection, which prevents more severe failures ahead in real-time for timely re-planning while generating code only once per subgoal. For example, in \"Pour Tea\", CaM detects failure by monitoring the teapot's tilt angle in real-time, avoiding frequent VLM checks." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.78, + 0.707, + 0.794 + ], + "angle": 0, + "content": "4.2.3 Results in RLBench" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.805, + 0.905, + 0.836 + ], + "angle": 0, + "content": "We further evaluate CaM on RLbench and demonstrate its superior generalization. Details can be found in Supp. D.3." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.847, + 0.764, + 0.862 + ], + "angle": 0, + "content": "4.3. Main Results in Real World" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.908, + 0.903 + ], + "angle": 0, + "content": "We conduct real-world evaluations on two tasks: (1) Simple Pick & Place: The robot has 70s to pick up objects and" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6924" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.09, + 0.904, + 0.104 + ], + "angle": 0, + "content": "Table 2. Performance in Omnigibson. We report the success rate (SR), execution time, and token usage, compared to DoReMi (DRM) [15]." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.107, + 0.905, + 0.174 + ], + "angle": 0, + "content": "
MethodSlot Pen (Point-level Disturbances)Stow Book (Line-level Disturbances)Pour Tea (Surface-level Disturbances)
SR with Disturbance(%) ↑Time (s)↓Token (k)↓SR with Disturbance(%) ↑Time (s)↓Token (k)↓SR with Disturbance(%) ↑Time (s)↓Token (k)↓
NoneDist.(a)Dist.(b)Dist.(c)NoneDist.(a)Dist.(b)Dist.(c)NoneDist.(a)Dist.(b)Dist.(c)
ReKep [21]30201010--40303020--20202010--
+DRM40102020177.8454.5450402040127.1738.670000--
+Ours60504040101.8525.827060706093.0818.6750404030174.5544.19
" + }, + { + "type": "image_caption", + "bbox": [ + 0.21, + 0.182, + 0.786, + 0.194 + ], + "angle": 0, + "content": "Reasoning Pick &Place Task: Grasp the animals according to their distances to the fruits, from nearest to farthest." + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.197, + 0.197, + 0.279 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.198, + 0.197, + 0.298, + 0.279 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.298, + 0.197, + 0.399, + 0.279 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.399, + 0.197, + 0.499, + 0.279 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.499, + 0.197, + 0.599, + 0.279 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.6, + 0.197, + 0.7, + 0.279 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.7, + 0.197, + 0.801, + 0.279 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.801, + 0.197, + 0.902, + 0.279 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.28, + 0.28, + 0.415, + 0.29 + ], + "angle": 0, + "content": "Horse is moved, target changed" + }, + { + "type": "image_caption", + "bbox": [ + 0.588, + 0.281, + 0.718, + 0.29 + ], + "angle": 0, + "content": "Pear is moved, target changed" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.292, + 0.905, + 0.32 + ], + "angle": 0, + "content": "Figure 5. Example of Real-world Evaluation. The red bounding box shows the current grasp target, which may shift due to environmental changes. CaM monitors and adapts to these changes in real-time, resulting in a closed-loop system with an open-loop policy." + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.327, + 0.482, + 0.354 + ], + "angle": 0, + "content": "Table 3. Performance of Single Pick & Place. We report the success rate and execution time. DGN donates DexGraspNet 2.0 [69]." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.357, + 0.479, + 0.466 + ], + "angle": 0, + "content": "
Tasks with disturbanceObject typesSuccess Rate(%) ↑Execution Time(s) ↓
DGN+DRM+Ours+DRM+Ours
Pick & PlaceDeformable0.0083.3396.6761.846.3
with theTransparent0.0066.6793.3372.648.1
objects beingSmall Rigid0.0080.096.6765.745.4
movedLarge Geometric0.0086.6796.6768.945.3
Pick & PlaceDeformable0.0076.6793.3368.762.5
with theTransparent0.0060.0090.0077.762.7
objects beingSmall Rigid0.0063.3393.3369.860.5
removedLarge Geometric0.0076.6796.6772.360.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.474, + 0.483, + 0.543 + ], + "angle": 0, + "content": "Table 4. Performance of Reasoning Pick & Place in cluttered scene. We report the success rate. The robot is controlled by an open-loop policy named DexGraspNet 2.0 (DGN) [69]. w/ CE with \\(\\checkmark\\) indicates using constraint elements; otherwise, constraint-aware entities or parts are used for tracking and code computation." + }, + { + "type": "table", + "bbox": [ + 0.097, + 0.546, + 0.472, + 0.615 + ], + "angle": 0, + "content": "
Tasksw/ CESuccess Rate(%) ↑
DGN+DRM+Ours
Clear all objects on table except for animals✓0.0010.0060.00
X0.0010.0020.00
Grasp the animals according to their distances to fruits, from nearest to farthest✓0.000.0090.00
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.623, + 0.483, + 0.76 + ], + "angle": 0, + "content": "place them at specified locations, facing two disturbances: (a) moving the object during grasping, and (b) removing the object from hand during movement. We test four object types, selecting 3 examples per type and conducting 10 trials for each (see Tab. 3). (2) Reasoning Pick & Place: The robot executes long-horizon tasks, involving ambiguous terms (e.g., \"fruit\"), under the same disturbances. We evaluate 2 long-horizon tasks in cluttered scenes, performing 10 trials each (see Tab. 4). Check Supp. D.4 for details." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Elements generally abstract constraints and relevant entities. As shown in Tab. 3, using a different end-effector (i.e., Leap Hand [51]) in Simple Pick & Place, CaM also achieves success rates surpassing DoReMi [15] by \\(20.4\\%\\) when handling different kinds of objects (e.g., deformable). We find that abstracting constraint-related entities/parts removes the irrelevant visual details, enabling generalization to different kinds of entities in unseen scenes (as discussed in Sec. 3.2), leading to easier tracking and code evaluation." + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.327, + 0.905, + 0.353 + ], + "angle": 0, + "content": "Table 5. Performance of reasoning and constraint-aware segmentation. FMC denotes the foundation model combination baseline." + }, + { + "type": "table", + "bbox": [ + 0.531, + 0.357, + 0.885, + 0.435 + ], + "angle": 0, + "content": "
MethodReasonSegConstraintSeg
Instance-levelInstance-levelPart-level
gIoUcIoUgIoUcIoUgIoUcIoU
PixelLM [50]56.061.444.443.224.122.6
LISA-13B [28]56.665.142.138.923.424.3
FMC51.253.349.349.640.839.3
ConSeg-13B55.763.962.168.760.265.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.443, + 0.905, + 0.484 + ], + "angle": 0, + "content": "Table 6. Ablation studies in Omnigibson's \"Stow Book\", assessing the impact of Multi-View (MV), Constraint-aware Segmentation (CS) for elements, and Connect Points (CP) for element formation." + }, + { + "type": "table", + "bbox": [ + 0.529, + 0.488, + 0.886, + 0.555 + ], + "angle": 0, + "content": "
MVCSCPSuccess Rate with Disturbance(%) ↑
NoneDist.(a)Dist.(b)Dist.(c)Avg
X✓✓40.040.030.050.040.0
✓X✓50.040.040.040.042.5
✓✓X60.050.060.050.055.0
✓✓✓70.060.070.060.065.0
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.562, + 0.907, + 0.744 + ], + "angle": 0, + "content": "Reactive and Proactive failure detection combined with an open-loop policy forms a close-loop system. Tab. 4 shows that only \\( CaM \\) successfully handles long-horizon tasks in cluttered scenes. These tasks are challenging as the robot is controlled by an open-loop policy that cannot handle environment dynamics and human disturbances in a closed-loop manner. By incorporating both reactive and proactive failure detection with the open-loop policy (see Fig. 5), the robot can dynamically adjust its target in real-time. For example, when a human moves the horse or pear during the task, the robot adapts by grasping the animal closest to the fruit, forming a closed-loop system." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.756, + 0.78, + 0.772 + ], + "angle": 0, + "content": "4.4. Main Results of Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.781, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Tab. 5 presents segmentation results comparing our ConSeg, with SOTA models and a Foundation Model Combination (FMC) baseline. FMC integrates GPT-4o for reasoning over tasks and constraints to identify relevant instances and parts, along with Grounded SAM [49] and Semantic SAM [32] for instance and part-level segmentation, respectively, similar to our data collection pipeline. We evaluate performance on the ReasonSeg [28] benchmark and our" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6925" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.09, + 0.904, + 0.313 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.097, + 0.319, + 0.898, + 0.333 + ], + "angle": 0, + "content": "Figure 6. Visual comparison between our ConSeg and LISA [28] at instance and part level. The red masks are the segmentation results." + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.34, + 0.483, + 0.396 + ], + "angle": 0, + "content": "Table 7. Ablation study on training data. SemanticSeg includes ADE20K [75], COCO-Stuff [5], PACO-LVIS [48] and PASCALPart [6]. ReferSeg includes refCLEF, refCOCO, refCOCO+ [26] and refCOCOg [39]. VQA indicates LLaVAInstruct-150k [36]." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.399, + 0.475, + 0.466 + ], + "angle": 0, + "content": "
SemanticSegReferSegVQAReasonSegConstraintSegIns-level gIoUcloUPart-level gIoUcloU
✓✓✓✓✘✘42.138.923.424.3
✓✓✓✓✓✘60.765.940.445.6
✓✓✓✓✘✓51.550.656.561.7
✓✓✓✓✓✓62.168.760.265.3
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.475, + 0.483, + 0.521 + ], + "angle": 0, + "content": "proposed Constraint-Aware Segmentation (ConstraintSeg) benchmark. Our benchmark evaluates performance using gIoU and cIoU, following ReasonSeg's evaluation setting." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.525, + 0.484, + 0.631 + ], + "angle": 0, + "content": "ConSeg performs both reasoning and multi-granularity constraint-aware segmentation. As shown in Tab. 5, ConSeg performs comparably to LISA and PixelLM on the ReasonSeg but significantly surpasses them on ConstraintSeg, achieving nearly a \\(40\\%\\) improvement at the part level. Visual comparisons in Fig. 6 further demonstrate ConSeg's strong generalization to unseen objects, scenes, and tasks." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.642, + 0.245, + 0.658 + ], + "angle": 0, + "content": "4.5. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.666, + 0.465, + 0.681 + ], + "angle": 0, + "content": "We conduct ablation studies and present analyses below." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.685, + 0.483, + 0.791 + ], + "angle": 0, + "content": "Multi views are critical for visual programming. As shown in Tab. 6, using only front-view images with constraint elements reduces the average success rate from \\(65\\%\\) to \\(40\\%\\). This is primarily due to (1) occlusion leading to suboptimal element generation and (2) errors in visual programming, where dimensional reduction causes surfaces to be misinterpreted as lines or lines as points." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Constraint-aware segmentation enhances element. As shown in Tab. 6, using DINOv2 [42] to generate semantic points to form elements, rather than through constraint-aware segmentation, reduces the success rate from \\(65\\%\\) to \\(42.5\\%\\), because these constructed elements fail to represent the desired constraints accurately. For example, capturing a book's vertical orientation requires two precise points on its" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.342, + 0.768, + 0.357 + ], + "angle": 0, + "content": "edge, which DINOv2 can not provide." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.364, + 0.906, + 0.456 + ], + "angle": 0, + "content": "Forming elements improves code generation. As shown in Tab. 6, using unconnected 3D points as final elements reduces the success rate from \\(65\\%\\) to \\(55\\%\\). Pre-formed elements contain more prior constraint information, serving as visual cues better encoded in code, whereas standalone points require recombination during visual programming." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.462, + 0.907, + 0.584 + ], + "angle": 0, + "content": "Elements simplify constraints computation. As shown in Tab. 4, the success rate decreases from \\(60\\%\\) to \\(20\\%\\) when entities or parts are not transformed into proposed elements in real-world evaluation. We find two key issues: (1) inaccurate 3D position and slow pose tracking of entities or parts; and (2) difficulties in performing arithmetic operations on their 3D positions or poses in code, hindering the encoding of their spatio-temporal constraints." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.59, + 0.907, + 0.68 + ], + "angle": 0, + "content": "Both instance-level and part-level data improve performance. We conduct ablation studies to assess the impact of our proposed dataset on the ConstraintSeg benchmark. The results are shown in Tab. 7. The results indicate that both the instance-level and part-level subsets contribute to performance improvements and enhance each other's effects." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.706, + 0.634, + 0.721 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In this paper, we present a novel paradigm termed Code-as-Monitor, leveraging the VLMs for both open-set reactive and proactive failure detection. In detail, We formulate both detection modes as spatio-temporal constraint satisfaction problems and use VLM-generated code to evaluate them for real-time monitoring. We further propose constraint elements, which abstract constraints-related entities or their parts into compact geometric elements, to improve the precision and efficiency of monitoring. Extensive experiments demonstrate the superiority of the proposed approach and highlight its potential to advance closed-loop robot systems." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6926" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.251, + 0.108 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.116, + 0.484, + 0.177 + ], + "angle": 0, + "content": "This work was supported by the National Natural Science Foundation of China (62132001), the National Science and Technology Major Project (2022ZD0116314), and the Fundamental Research Funds for the Central Universities." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.191, + 0.188, + 0.208 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.216, + 0.484, + 0.272 + ], + "angle": 0, + "content": "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv, 2023. 2, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.274, + 0.484, + 0.327 + ], + "angle": 0, + "content": "[2] Aneseh Alvanpour, Sumit Kumar Das, Christopher Kevin Robinson, Olfa Nasraoui, and Dan Popa. Robot failure mode prediction with explainable machine learning. CASE, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.33, + 0.484, + 0.4 + ], + "angle": 0, + "content": "[3] Jingkun An, Yinghao Zhu, Zongjian Li, Enshen Zhou, Haoran Feng, Xijie Huang, Bohua Chen, Yemin Shi, and Chengwei Pan. Agfpsync: Leveraging ai-generated feedback for preference optimization in text-to-image generation. arXiv preprint arXiv:2403.13352, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.402, + 0.483, + 0.457 + ], + "angle": 0, + "content": "[4] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. *NeurIPS*, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.459, + 0.483, + 0.486 + ], + "angle": 0, + "content": "[5] Holger Caesar, Jasper Uijlings, and Vittorio Ferrari. Cocostuff: Thing and stuff classes in context. CVPR, 2018. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.488, + 0.484, + 0.543 + ], + "angle": 0, + "content": "[6] Xianjie Chen, Roozbeh Mottaghi, Xiaobai Liu, Sanja Fidler, Raquel Urtasun, and Alan Yuille. Detect what you can: Detecting and representing objects using holistic models and body parts. CVPR, 2014. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.545, + 0.484, + 0.613 + ], + "angle": 0, + "content": "[7] Zeren Chen, Zhelun Shi, Xiaoya Lu, Lehan He, Sucheng Qian, Zhenfei Yin, Wanli Ouyang, Jing Shao, Yu Qiao, Cewu Lu, et al. Rh20t-p: A primitive-level robotic dataset towards composable generalization agents. arXiv preprint arXiv:2403.19622, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.616, + 0.484, + 0.657 + ], + "angle": 0, + "content": "[8] Jaemin Cho, Abhay Zala, and Mohit Bansal. Visual programming for step-by-step text-to-image generation and evaluation. NeurIPS, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.66, + 0.484, + 0.727 + ], + "angle": 0, + "content": "[9] Wenbo Cui, Chengyang Zhao, Songlin Wei, Jiazhao Zhang, Haoran Geng, Yaran Chen, and He Wang. Gapartmanip: A large-scale part-centric dataset for material-agnostic articulated object manipulation. arXiv preprint arXiv:2411.18276, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.73, + 0.483, + 0.771 + ], + "angle": 0, + "content": "[10] Yinpei Dai, Jayjun Lee, Nima Fazeli, and Joyce Chai. Racer: Rich language-guided failure recovery policies for imitation learning. arXiv, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.774, + 0.483, + 0.814 + ], + "angle": 0, + "content": "[11] Maximilian Diehl and Karinne Ramirez-Amaro. A causal-based approach to explain, predict and prevent failures in robotic tasks. RAS, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.817, + 0.483, + 0.87 + ], + "angle": 0, + "content": "[12] Yuqing Du, Ksenia Konyushkova, Misha Denil, Akhil Raju, Jessica Landon, Felix Hill, Nando de Freitas, and Serkan Cabi. Vision-language models as success detectors. arXiv, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[13] Jiafei Duan, Wilbert Pumacay, Nishanth Kumar, Yi Ru Wang, Shulin Tian, Wentao Yuan, Ranjay Krishna, Dieter" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.216, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.135 + ], + "angle": 0, + "content": "Fox, Ajay Mandlekar, and Yijie Guo. Aha: A vision-language-model for detecting and reasoning over failures in robotic manipulation. arXiv, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.137, + 0.905, + 0.191 + ], + "angle": 0, + "content": "[14] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.194, + 0.905, + 0.249 + ], + "angle": 0, + "content": "[15] Yanjiang Guo, Yen-Jen Wang, Lihan Zha, Zheyuan Jiang, and Jianyu Chen. Doremi: Grounding language model by detecting and recovering from plan-execution misalignment. arXiv, 2023. 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.252, + 0.904, + 0.293 + ], + "angle": 0, + "content": "[16] Tanmay Gupta and Aniruddha Kembhavi. Visual programming: Compositional visual reasoning without training. CVPR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.296, + 0.905, + 0.338 + ], + "angle": 0, + "content": "[17] Edward J Hu, Phillip Wallis, Zeyuan Allen-Zhu, Yanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. *ICLR*, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.34, + 0.905, + 0.382 + ], + "angle": 0, + "content": "[18] Xuhao Hu, Dongrui Liu, Hao Li, Xuanjing Huang, and Jing Shao. Vlsbench: Unveiling visual leakage in multimodal safety. arXiv preprint arXiv:2411.19939, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.384, + 0.905, + 0.437 + ], + "angle": 0, + "content": "[19] Haoxu Huang, Fanqi Lin, Yingdong Hu, Shengjie Wang, and Yang Gao. Copa: General robotic manipulation through spatial constraints of parts with foundation models. arXiv, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.442, + 0.905, + 0.51 + ], + "angle": 0, + "content": "[20] Wenlong Huang, Fei Xia, Ted Xiao, Harris Chan, Jacky Liang, Pete Florence, Andy Zeng, Jonathan Tompson, Igor Mordatch, Yevgen Chebotar, et al. Inner monologue: Embodied reasoning through planning with language models. arXiv, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.513, + 0.905, + 0.568 + ], + "angle": 0, + "content": "[21] Wenlong Huang, Chen Wang, Yunzhu Li, Ruohan Zhang, and Li Fei-Fei. Rekep: Spatio-temporal reasoning of relational keypoint constraints for robotic manipulation. arXiv, 2024. 3, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.571, + 0.905, + 0.613 + ], + "angle": 0, + "content": "[22] Stephen James, Zicong Ma, David Rovick Arrojo, and Andrew J. Davison. Rlbench: The robot learning benchmark & learning environment. RAL, 2020. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.615, + 0.905, + 0.684 + ], + "angle": 0, + "content": "[23] Yuheng Ji, Huajie Tan, Jiayu Shi, Xiaoshuai Hao, Yuan Zhang, Hengyuan Zhang, Pengwei Wang, Mengdi Zhao, Yao Mu, Pengju An, et al. Robobrain: A unified brain model for robotic manipulation from abstract to concrete. arXiv preprint arXiv:2502.21257, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.687, + 0.905, + 0.756 + ], + "angle": 0, + "content": "[24] Yueru Jia, Jiaming Liu, Sixiang Chen, Chenyang Gu, Zhilue Wang, Longzan Luo, Lily Lee, Pengwei Wang, Zhongyuan Wang, Renrui Zhang, et al. Lift3d foundation policy: Lifting 2d large-scale pretrained models for robust 3d robotic manipulation. arXiv preprint arXiv:2411.18623, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.758, + 0.905, + 0.8 + ], + "angle": 0, + "content": "[25] Nikita Karaev, Ignacio Rocco, Benjamin Graham, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. Co-tracker: It is better to track together. ECCV, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.802, + 0.905, + 0.843 + ], + "angle": 0, + "content": "[26] Sahar Kazemzadeh, Vicente Ordonez, Mark Matten, and Tamara Berg. Referitag: Referring to objects in photographs of natural scenes. EMNLP, 2014. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[27] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. ICCV, 2023. 3, 4" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6927" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.134 + ], + "angle": 0, + "content": "[28] Xin Lai, Zhuotao Tian, Yukang Chen, Yanwei Li, Yuhui Yuan, Shu Liu, and Jiaya Jia. Lisa: Reasoning segmentation via large language model. CVPR, 2024. 4, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.483, + 0.177 + ], + "angle": 0, + "content": "[29] Olivia Y Lee, Annie Xie, Kuan Fang, Karl Pertsch, and Chelsea Finn. Affordance-guided reinforcement learning via visual prompting. arXiv, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.179, + 0.482, + 0.22 + ], + "angle": 0, + "content": "[30] Gregory LeMasurier, Alvika Gautam, Zhao Han, Jacob W Crandall, and Holly A Yanco. Reactive or proactive? how robots should explain failures. HRI, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.222, + 0.482, + 0.358 + ], + "angle": 0, + "content": "[31] Chengshu Li, Ruohan Zhang, Josiah Wong, Cem Gokmen, Sanjana Srivastava, Roberto Martin-Martin, Chen Wang, Gabrael Levine, Michael Lingelbach, Jiankai Sun, Mona Anvari, Minjune Hwang, Manasi Sharma, Arman Aydin, Dhruva Bansal, Samuel Hunter, Kyu-Young Kim, Alan Lou, Caleb R Matthews, Ivan Villa-Renteria, Jerry Huayang Tang, Claire Tang, Fei Xia, Silvio Savarese, Hyowon Gweon, Karen Liu, Jiajun Wu, and Li Fei-Fei. Behavior-1k: A benchmark for embodied ai with 1,000 everyday activities and realistic simulation. CoRL, 2022. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.362, + 0.482, + 0.416 + ], + "angle": 0, + "content": "[32] Feng Li, Hao Zhang, Peize Sun, Xueyan Zou, Shilong Liu, Jianwei Yang, Chunyuan Li, Lei Zhang, and Jianfeng Gao. Semantic-sam: Segment and recognize anything at any granularity. arXiv, 2023. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.419, + 0.482, + 0.473 + ], + "angle": 0, + "content": "[33] Lijun Li, Zhelun Shi, Xuhao Hu, Bowen Dong, Yiran Qin, Xihui Liu, Lu Sheng, and Jing Shao. T2isafety: Benchmark for assessing fairness, toxicity, and privacy in image generation. arXiv preprint arXiv:2501.12612, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.476, + 0.482, + 0.53 + ], + "angle": 0, + "content": "[34] Jacky Liang, Wenlong Huang, Fei Xia, Peng Xu, Karol Hausman, Brian Ichter, Pete Florence, and Andy Zeng. Code as policies: Language model programs for embodied control. ICRA, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.532, + 0.482, + 0.574 + ], + "angle": 0, + "content": "[35] Fangchen Liu, Kuan Fang, Pieter Abbeel, and Sergey Levine. Moka: Open-vocabulary robotic manipulation through mark-based visual prompting. arXiv, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.576, + 0.482, + 0.602 + ], + "angle": 0, + "content": "[36] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. NeurIPS, 2024. 2, 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.605, + 0.482, + 0.673 + ], + "angle": 0, + "content": "[37] Jiaming Liu, Chenxuan Li, Guanqun Wang, Lily Lee, Kaichen Zhou, Sixiang Chen, Chuyan Xiong, Jiaxin Ge, Renrui Zhang, and Shanghang Zhang. Self-corrected multimodal large language model for end-to-end robot manipulation. arXiv, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.676, + 0.482, + 0.716 + ], + "angle": 0, + "content": "[38] Zeyi Liu, Arpit Bahety, and Shuran Song. Reflect: Summarizing robot experiences for failure explanation and correction. CoRL, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.719, + 0.482, + 0.771 + ], + "angle": 0, + "content": "[39] Junhua Mao, Jonathan Huang, Alexander Toshev, Oana Camburu, Alan L Yuille, and Kevin Murphy. Generation and comprehension of unambiguous object descriptions. CVPR, 2016. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.775, + 0.482, + 0.842 + ], + "angle": 0, + "content": "[40] Yao Mu, Junting Chen, Qinglong Zhang, Shoufa Chen, Qiaojun Yu, Chongjian Ge, Runjian Chen, Zhixuan Liang, Mengkang Hu, Chaofan Tao, et al. Robocodex: Multimodal code generation for robotic behavior synthesis. arXiv, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[41] Soroush Nasiriany, Fei Xia, Wenhao Yu, Ted Xiao, Jacky Liang, Ishita Dasgupta, Annie Xie, Danny Driess, Ayzaan Wahid, Zhuo Xu, et al. Pivot: Iterative visual prompting elicits actionable knowledge for vlms. arXiv, 2024. 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.16 + ], + "angle": 0, + "content": "[42] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv, 2023. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.163, + 0.905, + 0.232 + ], + "angle": 0, + "content": "[43] Zekun Qi, Wenyao Zhang, Yufei Ding, Runpei Dong, Xinqiang Yu, Jingwen Li, Lingyun Xu, Baoyu Li, Xialin He, Guofan Fan, et al. Sofar: Language-grounded orientation bridges spatial reasoning and object manipulation. arXiv preprint arXiv:2502.13143, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.234, + 0.905, + 0.287 + ], + "angle": 0, + "content": "[44] Yiran Qin, Zhelun Shi, Jiwen Yu, Xijun Wang, Enshen Zhou, Lijun Li, Zhenfei Yin, Xihui Liu, Lu Sheng, Jing Shao, et al. Worldsimbench: Towards video generation models as world simulators. arXiv preprint arXiv:2410.18072, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.289, + 0.905, + 0.37 + ], + "angle": 0, + "content": "[45] Yiran Qin, Enshen Zhou, Qichang Liu, Zhenfei Yin, Lu Sheng, Ruimao Zhang, Yu Qiao, and Jing Shao. Mp5: A multi-modal open-ended embodied system in mycraf via active perception. In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16307-16316. IEEE, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.372, + 0.905, + 0.412 + ], + "angle": 0, + "content": "[46] Yiran Qin, Ao Sun, Yuze Hong, Benyou Wang, and Ruimao Zhang. Navigatediff: Visual predictors are zero-shot navigation assistants. arXiv preprint arXiv:2502.13894, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.414, + 0.905, + 0.467 + ], + "angle": 0, + "content": "[47] Shreyas Sundara Raman, Vanya Cohen, Eric Rosen, Ifrah Idrees, David Paulius, and Stefanie Tellex. Planning with large language models via corrective re-prompting. NeurIPS, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.469, + 0.905, + 0.524 + ], + "angle": 0, + "content": "[48] Vignesh Ramanathan, Anmol Kalia, Vladan Petrovic, Yi Wen, Baixue Zheng, Baishan Guo, Rui Wang, Aaron Marquez, Rama Kovvuri, Abhishek Kadian, et al. Paco: Parts and attributes of common objects. CVPR, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.526, + 0.905, + 0.607 + ], + "angle": 0, + "content": "[49] Tianhe Ren, Shilong Liu, Ailing Zeng, Jing Lin, Kunchang Li, He Cao, Jiayu Chen, Xinyu Huang, Yukang Chen, Feng Yan, Zhaoyang Zeng, Hao Zhang, Feng Li, Jie Yang, Hongyang Li, Qing Jiang, and Lei Zhang. Grounded sam: Assembling open-world models for diverse visual tasks. arXiv, 2024. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.609, + 0.905, + 0.649 + ], + "angle": 0, + "content": "[50] Zhongwei Ren, Zhicheng Huang, Yunchao Wei, Yao Zhao, Dongmei Fu, Jiashi Feng, and Xiaojie Jin. Pixel reasoning with large multimodal model. CVPR, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.651, + 0.905, + 0.691 + ], + "angle": 0, + "content": "[51] Kenneth Shaw, Ananye Agarwal, and Deepak Pathak. Leap hand: Low-cost, efficient, and anthropomorphic hand for robot learning. arXiv, 2023. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.693, + 0.905, + 0.76 + ], + "angle": 0, + "content": "[52] Zhelun Shi, Zhipin Wang, Hongxing Fan, Zaibin Zhang, Lijun Li, Yongting Zhang, Zhenfei Yin, Lu Sheng, Yu Qiao, and Jing Shao. Assessment of multimodal large language models in alignment with human values. arXiv preprint arXiv:2403.17830, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.763, + 0.905, + 0.801 + ], + "angle": 0, + "content": "[53] Noah Shinn, Beck Labash, and Ashwin Gopinath. Reflection: an autonomous agent with dynamic memory and self-reflection. arXiv, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.804, + 0.905, + 0.844 + ], + "angle": 0, + "content": "[54] Mohit Shridhar, Lucas Manuelli, and Dieter Fox. CIport: What and where pathways for robotic manipulation. CoRL, 2021. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[55] Tom Silver, Soham Dan, Kavitha Srinivas, Joshua B Tenenbaum, Leslie Kaelbling, and Michael Katz. Generalized planning in pddl domains with pretrained large language models. AAAI, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "6928" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.174 + ], + "angle": 0, + "content": "[56] Marta Skreta, Naruki Yoshikawa, Sebastian Arellano-Rubach, Zhi Ji, Lasse Bjørn Kristensen, Kourosh Darvish, Alán Aspuru-Guzik, Florian Shkurti, and Animesh Garg. Errors are useful prompts: Instruction guided task programming with verifier-assisted iterative prompting. arXiv, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.178, + 0.482, + 0.247 + ], + "angle": 0, + "content": "[57] Yingbo Tang, Shuaike Zhang, Xiaoshuai Hao, Pengwei Wang, Jianlong Wu, Zhongyuan Wang, and Shanghang Zhang. Affordgrasp: In-context affordance reasoning for open-vocabulary task-oriented grasping in clutter. arXiv preprint arXiv:2503.00778, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.249, + 0.482, + 0.316 + ], + "angle": 0, + "content": "[58] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.319, + 0.482, + 0.387 + ], + "angle": 0, + "content": "[59] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.39, + 0.482, + 0.431 + ], + "angle": 0, + "content": "[60] Sai H Vemprala, Rogerio Bonatti, Arthur Tucker, and Ashish Kapoor. Chatgpt for robotics: Design principles and model abilities. IEEE Access, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.433, + 0.482, + 0.474 + ], + "angle": 0, + "content": "[61] David Venuto, Sami Nur Islam, Martin Klissarov, Doina Precup, Sherry Yang, and Ankit Anand. Code as reward: Empowering reinforcement learning with vlms. arXiv, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.476, + 0.482, + 0.544 + ], + "angle": 0, + "content": "[62] Homer Rich Walke, Kevin Black, Tony Z Zhao, Quan Vuong, Chongyi Zheng, Philippe Hansen-Estruch, Andre Wang He, Vivek Myers, Moo Jin Kim, Max Du, et al. Bridgedata v2: A dataset for robot learning at scale. CoRL, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.547, + 0.482, + 0.588 + ], + "angle": 0, + "content": "[63] Jianwei Yang, Hao Zhang, Feng Li, Xueyan Zou, Chunyuan Li, and Jianfeng Gao. Set-of-mark prompting unleashes extraordinary visual grounding in gpt-4v. arXiv, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.59, + 0.482, + 0.631 + ], + "angle": 0, + "content": "[64] Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, and Yuan Cao. React: Synergizing reasoning and acting in language models. arXiv, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.633, + 0.482, + 0.672 + ], + "angle": 0, + "content": "[65] Runpeng Yu, Weihao Yu, and Xinchao Wang. Attention prompting on image for large vision-language models. arXiv, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.676, + 0.482, + 0.729 + ], + "angle": 0, + "content": "[66] Zhouliang Yu, Jie Fu, Yao Mu, Chenguang Wang, Lin Shao, and Yaodong Yang. Multireact: Multimodal tools augmented reasoning-acting traces for embodied agent planning. *ICLR*, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.733, + 0.482, + 0.787 + ], + "angle": 0, + "content": "[67] Wentao Yuan, Jiafei Duan, Valts Blukis, Wilbert Pumacay, Ranjay Krishna, Adithyavairavan Murali, Arsalan Mousavian, and Dieter Fox. Robopoint: A vision-language model for spatial affordance prediction for robotics. arXiv, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.789, + 0.482, + 0.842 + ], + "angle": 0, + "content": "[68] Zhihao Yuan, Jinke Ren, Chun-Mei Feng, Hengshuang Zhao, Shuguang Cui, and Zhen Li. Visual programming for zero-shot open-vocabulary 3d visual grounding. CVPR, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[69] Jialiang Zhang, Haoran Liu, Danshi Li, XinQiang Yu, Hao ran Geng, Yufei Ding, Jiayi Chen, and He Wang. Dexgrasp-net 2.0: Learning generative dexterous grasping in largescale synthetic cluttered scenes. CoRL, 2024. 5, 7" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "[70] Xinyu Zhang, Yuhan Liu, Haonan Chang, Liam Schramm, and Abdeslam Boularias. Autoregressive action sequence learning for robotic manipulation. arXiv, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.137, + 0.905, + 0.204 + ], + "angle": 0, + "content": "[71] Yongting Zhang, Lu Chen, Guodong Zheng, Yifeng Gao, Rui Zheng, Jinlan Fu, Zhenfei Yin, Senjie Jin, Yu Qiao, Xuanjing Huang, et al. Spa-vl: A comprehensive safety preference alignment dataset for vision language model. arXiv preprint arXiv:2406.12030, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.207, + 0.905, + 0.261 + ], + "angle": 0, + "content": "[72] Zaibin Zhang, Shiyu Tang, Yuanhang Zhang, Talas Fu, Yifan Wang, Yang Liu, Dong Wang, Jing Shao, Lijun Wang, and Huchuan Lu. Ad-h: Autonomous driving with hierarchical agents. arXiv preprint arXiv:2406.03474, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.264, + 0.905, + 0.304 + ], + "angle": 0, + "content": "[73] Jinliang Zheng, Jianxiong Li, Sijie Cheng, Yinan Zheng, Ji-aming Li, Jihao Liu, Yu Liu, Jingjing Liu, and Xianyuan Zhan. Instruction-guided visual masking. arXiv, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.307, + 0.905, + 0.347 + ], + "angle": 0, + "content": "[74] Zhi Zheng, Qian Feng, Hang Li, Alois Knoll, and Jianxiang Feng. Evaluating uncertainty-based failure detection for closed-loop llm planners. arXiv, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.349, + 0.905, + 0.388 + ], + "angle": 0, + "content": "[75] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. CVPR, 2017. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.391, + 0.905, + 0.459 + ], + "angle": 0, + "content": "[76] Enshen Zhou, Yiran Qin, Zhenfei Yin, Yuzhou Huang, Ruimao Zhang, Lu Sheng, Yu Qiao, and Jing Shao. Minedreamer: Learning to follow instructions via chain-of-imagination for simulated-world control. arXiv preprint arXiv:2403.12037, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.459 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "6929" + } + ] +] \ No newline at end of file diff --git a/2025/Code-as-Monitor_ Constraint-aware Visual Programming for Reactive and Proactive Robotic Failure Detection/3086dac5-625c-4ed2-b32d-12ee6b15f252_origin.pdf b/2025/Code-as-Monitor_ Constraint-aware Visual Programming for Reactive and Proactive Robotic Failure Detection/3086dac5-625c-4ed2-b32d-12ee6b15f252_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..531862a313e10fb4f512f1d79b74d7f501707242 --- /dev/null +++ b/2025/Code-as-Monitor_ Constraint-aware Visual Programming for Reactive and Proactive Robotic Failure Detection/3086dac5-625c-4ed2-b32d-12ee6b15f252_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b40cd537288e5356a8798d6bb33d0cdc65ee5003dce978e4f0722847986f5532 +size 1944996 diff --git a/2025/Code-as-Monitor_ Constraint-aware Visual Programming for Reactive and Proactive Robotic Failure Detection/full.md b/2025/Code-as-Monitor_ Constraint-aware Visual Programming for Reactive and Proactive Robotic Failure Detection/full.md new file mode 100644 index 0000000000000000000000000000000000000000..296d4927baa759e2466e4bda2d2f29949dca0dbd --- /dev/null +++ b/2025/Code-as-Monitor_ Constraint-aware Visual Programming for Reactive and Proactive Robotic Failure Detection/full.md @@ -0,0 +1,329 @@ +# Code-as-Monitor: Constraint-aware Visual Programming for Reactive and Proactive Robotic Failure Detection + +Enshen Zhou $^{1,3*}$ , Qi Su $^{2,3,4*}$ , Cheng Chi $^{3*†}$ , Zhizheng Zhang $^{4}$ , Zhongyuan Wang $^{3}$ , Tiejun Huang $^{2,3}$ , Lu Sheng $^{1†}$ , He Wang $^{2,3,4†}$ , $^{1}$ Beihang University $^{2}$ Peking University $^{3}$ Beijing Academy of Artificial Intelligence $^{4}$ Galbot + +zhouenshen@buaa.edu.cn chicheng15@mails.ucas.ac.cn lsheng@buaa.edu.cn hewang@pku.edu.cn + +![](images/02514d55c0840685d9698546835ab2425a6bcdb26c727621e2fc7aca1f41445f.jpg) + +Task: Move the pan with the lobster to the stove, and be careful not to let the lobster drop out. + +![](images/ad2a9b3059d16dfce8fb97ac8eaa7705617ecc383e20ea57c9c28f978527b444.jpg) +(a) Reactive Failure Detection + +![](images/afb1dd77eb2d946721e72019f19549490e43e8d7270770e895d49b66c2d4fb41.jpg) +(b) Proactive Failure Detection + +![](images/813e5a2c28f246840e12ff1c4a1c1c84d785cda759cce1ba9f8fc0af6e4bc184.jpg) +(c) Real-world Test +Figure 1. For the task "Move the pan with lobster to the stove without losing the lobster", (a) reactive failure detection identifies failures after they occur, and (b) proactive failure detection prevents foreseeable failures. In (a), at $t_4^R$ , the robot detects the failure after the lobster unpredictably jumps out due to the heat. In (b), pan tilting is detected at $t_3^P$ and corrected it at $t_3^{P'}$ , requiring real-time precision. We formulate both tasks as spatio-temporal constraint satisfaction problems, leveraging our proposed constraint elements for precise, real-time checking. For example, in (a), a large relative distance between pan and lobster indicates failure; in (b), a large angle between the pan and the horizontal plane needs correction. (c) shows that our method combined with an open-loop policy forms a closed-loop system, enabling proactive (e.g., detecting moving glass during grasping) and reactive (e.g., removing toy after grasping) failure detection in cluttered scenes. + +# Abstract + +Automatic detection and prevention of open-set failures are crucial in closed-loop robotic systems. Recent studies often struggle to simultaneously identify unexpected failures reactively after they occur and prevent foreseeable ones proactively. To this end, we propose Code-as-Monitor (CaM), a novel paradigm leveraging the vision-language model (VLM) for both open-set reactive and proactive failure detection. The core of our method is to formulate both tasks as a unified set of spatio-temporal constraint satisfaction problems and use VLM-generated code to evaluate them for real-time monitoring. To enhance the accuracy and efficiency of monitoring, we further introduce constraint elements that abstract constraint-related entities or their parts into compact geometric ele + +ments. This approach offers greater generality, simplifies tracking, and facilitates constraint-aware visual programming by leveraging these elements as visual prompts. Experiments show that CaM achieves a $28.7\%$ higher success rate and reduces execution time by $31.8\%$ under severe disturbances compared to baselines across three simulators and a real-world setting. Moreover, CaM can be integrated with open-loop control policies to form closed-loop systems, enabling long-horizon tasks in cluttered scenes with dynamic environments. See the project page at https://zhoues.github.io/Code-as-Monitor. + +# 1. Introduction + +As expectations grow for robots to handle long-horizon tasks within intricate environments, failures are unavoidable. Therefore, automatically detecting and preventing those failures plays a vital role in ensuring the tasks can + +eventually be solved, especially for closed-loop robotic systems. There are two modes of failure detection [30], reactive and proactive. As depicted in Fig. 1, reactive failure detection aims to identify failures after they occur (e.g., recognizing that lobster lands on the table, indicating a delivery failure). In contrast, proactive failure detection aims to prevent foreseeable failures (e.g., recognizing that a tilted pan could cause the lobster to fall out, leading to a delivery failure). Both detection modes are even more challenging in open-set scenarios, where the failures are not predefined. + +With the help of large language models (LLMs) [14, 58] and vision-language models (VLMs) [1, 36], recent studies can achieve open-set reactive failure detection [10, 12, 13, 15, 20, 37, 55, 56, 66, 74] as a special case of visual question answering (VQA) tasks. However, these methods often bear compromised execution speeds and coarse-grained detection accuracy, due to the high computational costs and inadequate 3D spatio-temporal perception capability of recent LLMs/VLMs. Moreover, open-set proactive failure detection, which has been less explored in the literature, presents even severer challenges as it is required to foresee potential causes of failure and monitor them in real-time with high precision to anticipate and prevent imminent failure. Simply adapting LLMs/VLMs cannot meet such expectations. + +In this work, we aim to develop an open-set failure detection framework that achieves reactive and proactive detection simultaneously, not only benefiting from the generalization power of VLMs but also enjoying high precision in monitoring failure characteristics with real-time efficiency. We address this by formulating both tasks as a unified set of spatio-temporal constraint satisfaction problems, which can be precisely translated by VLMs into executable programs. Such visual programs can efficiently verify whether entities (e.g., robots, objects) or their parts in the captured environment maintain or achieve required states during or after execution (i.e., satisfying constraints), so as to immediately prevent or detect failures. To the best of our knowledge, this is the first attempt to integrate both detection modes within a single framework. We name this constraint-aware visual programming framework as Code-as-Monitor $(CaM)$ . + +To be specific, the proposed spatio-temporal constraint satisfaction scheme abstracts the constraint-related entity or part segments from the observed images into compact geometric elements (e.g., points, lines, and surfaces), as shown in Fig. 1. It simplifies the monitoring of constraint satisfaction by tracking and evaluating the spatio-temporal combinational dynamics of these elements, eliminating the most irrelevant geometric and visual details of the raw entities/parts. The constraint element detection and tracking are grounded by our trained constraint-aware segmentation and off-the-shelf tracking models, ensuring speed, accuracy, and certain open-set adaptation capabilities. The evaluation protocol is in the form of VLM-generated code, i.e., monitor + +code, which is visually prompted by the starting frames of a sub-goal and its associated constraint elements, in addition to the textual constraints that must be fulfilled. Once generated, this monitor code could detect reactive or proactive failures just by being executed according to the tracked constraint elements, without needing to call the VLMs again. Therefore, this minimalist scheme is generalizable to open-set failure detection for unseen entities and scenes (enabled by the potential diversity of the structured associations of constraint elements) as well as common skills (powered by the rich prior knowledge offered by VLMs), maintaining sufficient detection accuracy and real-time execution speed. + +We conduct extensive experiments in three simulators (i.e., CLIPort [54], Omnigibson [31], and RL-Bench [22]) and one real-world setting, spanning diverse manipulation tasks (e.g., pick & place, articulated objects, tool-use), robot platforms (e.g., UR5, Fetch, Franka) and end-effectors (e.g., suction cup, gripper, dexterous hand). The results show that CaM is generalizable and achieves both reactive and proactive failure detection in real-time, resulting in $28.7\%$ higher success rates and reduced execution time by $31.8\%$ under severe disturbances compared to the baselines. Moreover, in Sec. 4.3, CaM can be integrated with the existing open-loop control policy to form a closed-loop system, enabling long-horizon tasks in cluttered scenes with environment dynamics and human disturbances. Our contributions are summarized as follows: + +- We introduce Code-as-Monitor (CaM), a novel paradigm that leverages VLMs for both reactive and proactive failure detection via constraint-aware visual programming. +- We propose the constraint elements to enhance the accuracy and efficiency of constraint satisfaction monitoring. +- Extensive experiments show that CaM is generalizable and achieves more real-time and precise failure detection than baselines across simulators and real-world settings. + +# 2. Related work + +Robotic Failure Detection. Recent advances in LLMs [4, 14, 58, 59] and VLMs [1, 3, 18, 33, 36, 44, 45, 52, 71, 76] greatly improve open-set reactive failure detection for robotics [7, 9, 23, 24, 43, 46, 57, 72] Current LLM-based methods either convert visual inputs into text [38, 53, 64], potentially losing visual details, or rely on ground-truth feedback [20, 47, 55, 56], which is impractical in real-world scenarios. Recent studies employ VLMs as failure detectors, offering binary success indicators [12, 37, 66, 74] or textual explanations [10, 13] through visual question answering (VQA), such as DoReMi [15]. However, they often bear compromised execution speeds and coarse-grained detection accuracy. Meanwhile, open-set proactive failure detection is rarely explored, as previous methods [2, 11] are confined to predefined failures. This detection mode requires foreseeing potential failure causes and monitoring + +![](images/ddded73361f0a4434fdf1b72370125485c3d7d83cf3d628cb57583acb5b99e7d.jpg) +Figure 2. Overview of Code-as-Monitor. Given task instructions and prior information, the Constraint Generator derives the next subgoal and corresponding textual constraints based on multi-view observations. The Painter maps these constraints onto images as constraint elements. The Monitor generates monitor code from these images and tracks them for real-time monitoring. If any constraint is violated, it outputs the reason for failure and triggers re-planning. This framework unifies reactive and proactive failure detection via constraints, more generally abstracts relevant entities/parts through constraint elements, and ensures precise and real-time monitoring via code evaluation. + +them in real-time with high precision. Herein we formulate both reactive and proactive failure detection as constraint satisfaction problems and use VLM-generated code to evaluate them, meeting the expectations of both modes above. + +Visual Prompting. Visual prompts enhance the visual reasoning abilities of VLMs and are categorized into mask-based, point-based, and element-based methods. Mask-based approaches like SoM [63] apply numbered segmentation masks on images without considering instructions, whereas instruction-guided methods like IVM [73] and API [65] generate masks to block irrelevant regions. Point-based prompting encodes functionalities via points, including affordances [29, 35, 67], motion [35, 41], and constraints like ReKep [21]. However, ReKep [21] extracts semantic keypoints using DINOv2 [42] and struggles to capture precise keypoints that accurately represent desired constraints. Element-based methods like CoPa [19], represent robotic control using basic elements (e.g., points, lines) via pre-trained models like SAM [27] and GPT-4V [1] simply. In contrast, this work explores constraint satisfaction monitoring and introduces constraint elements that more precisely represent relevant entities/parts. These elements are tracked and evaluated in real-time to simplify monitoring. + +Visual Programming. Visual programming requires strong visual concept understanding and reasoning. Prior works show strong generalization (e.g., zero-shot) across various tasks, such as image editing [8, 16], 3D visual grounding [68], and robotics control [34, 60], by integrating LLMs with visual modules but lose fine-grained details. Recent + +studies [40, 61] explore visual programming using VLMs, but these methods take raw RGB images as inputs with predefined primitive libraries. In contrast, this work leverages constraint elements for visual programming and uses arithmetic operations in the code to encode their spatio-temporal combinational dynamics, presenting greater challenges. + +# 3. Method + +We first give an overview of the proposed Code-as-Monitor (CaM) (Sec. 3.1). Then, we elaborate on the constraint element in Constraint Painter, especially constraint-aware segmentation (Sec. 3.2). Finally, we present Constraint Monitor for real-time detection (Sec. 3.3). + +# 3.1. Overview + +The proposed CaM comprises three key modules: the Constraint Generator, Painter, and Monitor. We focus on long-horizon manipulation task instructions $\mathcal{L}_{\mathrm{global}}$ (e.g., "Move the pan with the bread to the stove, and be careful not to let the bread fall out"), using RGB-D observations $\mathcal{O}$ from two camera views (front and top). As shown in Fig. 2, the RGB images $\mathcal{O}$ , along with instructions $\mathcal{L}_{\mathrm{global}}$ , previous subgoal $l_{\mathrm{pre}}$ , and failure feedback from the Constraint Monitor $f_{\mathrm{pre}}$ (e.g., subgoal success or failure reason), are fed into the Constraint Generator $\mathcal{F}_{\mathrm{VLM}}$ (i.e., GPT-4o [1]) to generate the next subgoal $l_{\mathrm{next}}$ and associated textual constraints $\mathcal{C}$ . This process can be formulated as follows: + +$$ +l _ {\text {n e x t}}, \mathcal {C} _ {d}, \mathcal {C} _ {u} = \mathcal {F} _ {\mathrm {V L M}} \left(\mathcal {O}, \mathcal {L} _ {\text {g l o b a l}}, l _ {\text {p r e}}, f _ {\text {p r e}}\right) \tag {1} +$$ + +![](images/886a1ff14757f6786cfc75f4e25d63c20b4fca4f0f7a5e4cf0e114601e8474ef.jpg) +Figure 3. Constraint Element Pipeline. Given a constraint, our model ConSeg generates instance-level and part-level masks across multiple views, which are projected into 3D space. Through a series of heuristics, the desired elements are produced. Once all elements are obtained, they are annotated onto the original multi-view images. Here we display the annotation result of one element. + +where $\mathcal{C}_d = \{c_d^0, c_d^1, \ldots, c_d^n\}$ denotes the constraints that must be maintained during subgoal execution (e.g., pan handle must be grasped, bread must remain in the pan, pan should remain horizontal during transfer), and $\mathcal{C}_u = \{c_u^0, c_u^1, \ldots, c_u^k\}$ refers to the constraints that must be met upon subgoal completion (e.g., pan should be directly above the stove). We successfully unify reactive and proactive failure detection as these task-specific, situation-aware constraint satisfaction problems. + +In Painter, for each textual constraint $c$ from $\mathcal{C}_d$ or $\mathcal{C}_u$ , we generate corresponding constraint elements $e$ (detailed in Sec. 3.2) from observations $\mathcal{O}$ . These elements, which are composed of 3D points, abstract the constraint-related entities or their parts to represent the desired textual constraint more easily (e.g., the constant distance between green points on bread and pan determines if bread remains in pan, as shown in Fig. 2.) These generated elements are then aggregated into the final set $\mathcal{E} = \{e^0, e^1, \ldots, e^{n + k}\}$ , and numerically annotated across all views to produce the final visual prompted images $\mathcal{O}_{\mathcal{E}}$ . + +In Monitor, we provide GPT-4o [1] with the next subgoal $l_{\mathrm{next}}$ , textual constraints $\mathcal{C}$ , and annotated observations $\mathcal{O}_{\mathcal{E}}$ for constraint-aware visual programming to generate the evaluation protocol, i.e., monitor code. This code inputs the elements' 3D positions, calculates arithmetic operations within it, and returns a boolean to indicate potential or actual failure and a string to describe its reason. During subgoal execution, Monitor tracks the elements and evaluates the spatio-temporal combinational dynamics of these elements. If the code returns False, the policy execution halts immediately, and the accompanying string is used as feedback $f_{\mathrm{pre}}$ for re-planning. Otherwise, the subgoal is considered completed. In either case, the cycle is repeated. + +# 3.2. Constraint Element + +To simplify the monitoring of constraint satisfaction, we introduce constraint elements by abstracting constraint-related entities/parts into compact geometric elements (e.g., points, lines) to get rid of the most irrelevant geometric and visual details, making them easier to track and generalize. + +Pipeline. The constraint element generation pipeline is shown in Fig. 3. Our trained multi-granularity constraint-aware model ConSeg performs two inference steps for each $c$ and each RGB image $o$ from the set of views $O$ . First, instance-level masks $\mathcal{M}_i$ are generated to capture constraint-related entities. Then, fine-grained part-level masks $\mathcal{M}_p$ and corresponding element type descriptions $l_{\mathrm{e}}$ are produced, as shown in Fig. 4. Using corresponding depth data, we project $\mathcal{M}_p$ from all views into 3D space, fusing them into a point cloud. However, directly tracking and evaluating the spatio-temporal combinational dynamics of these constraint-related entities/parts is challenging. Therefore, we convert these entities into proposed constraint elements. We first apply voxelization to the point cloud with voxel size determined by element type $l_{\mathrm{e}}$ (e.g., the surface needs at least 3 points, we divide the occupied space into $2\times 2$ voxels.). Then we cluster one representative point within each voxel and filter them to a specified number, also determined by $l_{\mathrm{e}}$ to extract the final desired 3D points. We connect points within each instance-level mask $\mathcal{M}_i$ to form the constraint element $e$ associated with constraint $c$ . Notably, points modeled as end-effectors (e.g., the points of the fingertips and hand's center represent a dexterous hand) can be directly obtained from forward kinematics, bypassing the process above. Additionally, we perform parallel inference of ConSeg across all views to expedite the acquisition of the final set $\mathcal{E}$ and their annotations onto the corresponding views $\mathcal{O}_{\mathcal{E}}$ . Moreover, this minimalist approach, i.e., constraint elements, emphasizes the most relevant entities/parts, enabling generalization to unseen scenes and entities, which is critical for open-set failure detection. For more details, please check Supp. B.3. + +Constraint-aware Segmentation. Since the textual constraint does not explicitly specify relevant entities/parts, we require a model capable of logical reasoning and certain open-set adaptation, enabling precise constraint-aware instance and part-level segmentation. As depicted in Fig. 4, we propose ConSeg, which builds upon LISA [28] comprising a VLM $\mathcal{F}_{\mathrm{VLM}}$ (i.e., LLaVA [36]) with the vision encoder $\mathcal{F}_{\mathrm{enc}}$ and decoder $\mathcal{F}_{\mathrm{dec}}$ from SAM [27]. The textual + +![](images/ca68e9310cb053ab41342b51c7c6ab82de32f443ec5dbcc4378e2116ec9e5ba0.jpg) +Response $(l_{\mathrm{e}})$ : It needs a surface $<\mathrm{SEG}>.$ +Figure 4. ConSeg architecture. Here we display the part-level segmentation, which will output the desired element type and mask. + +constraint $c$ and the image $o$ are input into the VLM $\mathcal{F}_{\mathrm{VLM}}$ to generate an embedding for the token. The VLM outputs a textual description $l_{\mathrm{e}}$ of the desired element type additionally for part-level segmentation. The decoder $\mathcal{F}_{\mathrm{dec}}$ then generates the segmentation mask $\mathcal{M}$ by leveraging visual features from the vision encoder $\mathcal{F}_{\mathrm{enc}}$ based on the current observation $o$ , and features derived from the final-layer embedding $h_{<\mathrm{SEG}>}$ of the token, transformed via an MLP ( $\mathcal{F}_{\mathrm{MLP}}$ ). The process can be formulated as follows: + +$$ +l _ {\mathrm {e}}, h _ {< \mathrm {S E G} >} = \mathcal {F} _ {\mathrm {V L M}} (c, o), \mathcal {M} = \mathcal {F} _ {\mathrm {d e c}} (\mathcal {F} _ {\mathrm {M L P}} \left(h _ {< \mathrm {S E G} >}\right), \mathcal {F} _ {\mathrm {e n c}} (o)) \tag {2} +$$ + +The entire pipeline, including additional LoRA [17] parameters, is optimized end-to-end while keeping the parameters of the VLM $\mathcal{F}_{\mathrm{VLM}}$ and vision encoder $\mathcal{F}_{\mathrm{enc}}$ frozen. + +Dataset and Collection Pipeline. Our dataset utilizes images from BridgeData V2 [62]. While BridgeData V2 provides trajectory-level instructions, we require frame-level annotations for per-frame subgoals and constraints. To address this, we sample pick-and-place data, using external references (e.g., gripper open/close states) to generate 10,181 trajectories with 219,356 images. We employ GPT-4o to decompose trajectory instructions into subgoals, constraints, and object-part associations to generate groundtruth annotations. Instance-level and part-level segmentations are performed using Grounded SAM [49] and Semantic SAM [32], respectively. These outputs are integrated and refined through manual inspection to produce the multi-granularity dataset, which is combined with LISA's training data to fine-tune our model. More details are in Supp. B. + +# 3.3. Real-time Monitor Module + +After annotating constraint elements with unique colors and numerical labels across all views, these images serve as visual prompts to generate monitor code via GPT-4o [1] with constraints $\mathcal{C}$ and the next subgoal $l_{\mathrm{next}}$ . The generated code accepts the elements' 3D positions as input and performs arithmetic operations (e.g., NumPy-supported calculations) to evaluate spatio-temporal constraints at various stages (i.e., during execution or upon completion), as shown in Fig. 2 code block. It returns a boolean flag indicating + +potential or actual failure if the computed result exceeds a predefined tolerance (e.g., accepting deviations of the pan's surface normal from the z-axis within $15^{\circ}$ ), along with a string explaining the cause. By tracking the elements using CoTracker [25], we achieve real-time detection via VLMs without frequent calls, reducing the computational cost. + +Notably, including both current positions and historical trajectories of the elements as input allows us to monitor tasks needing high precision, such as the requirement of $2\mathrm{cm}$ movement or $180^{\circ}$ rotation, which previous reactive detection methods could not address through direct VLM queries using raw RGB images. Moreover, the potential diversity in the structured associations of constraint elements within the code (e.g., calculating the angle between the normal vector of the pan's surface element and the z-axis determines if the pan is level, as shown in Fig. 2 code block) and rich prior knowledge offered by VLMs enable generalization to certain unseen tasks requiring common skills. + +# 4. Experiments + +Our experiments aim to address the following questions: (1) Can CaM achieve open-set reactive and proactive failure detection across diverse robots, end-effectors, and objects, both in simulator (Sec. 4.2) and on real robots (Sec. 4.3)? (2) Can ConSeg infer multi-granularity constraint-aware masks for unseen scenes and objects (Sec. 4.4)? (3) Which design choices greatly enhance the performance (Sec. 4.5)? + +# 4.1. Experimental Setup + +Environment Settings. We evaluate CaM across three simulators (i.e., CLIPort [54], Omnigibson [31], and RL-Bench [22]) and a real-world setting. In CLIPort, we employ a UR5 arm equipped with a suction cup for pick-and-place and a spatula for pushing, controlled by pre-trained CLIPort policy [54]. Omnigibson features a Fetch robot with a gripper, controlled by ReKep [21]. RLBench uses a Franka arm with a gripper, controlled by ARP [70]. We collect a small dataset to fine-tune ConSeg for each simulator. We use a UR5 arm with a Leap Hand [51] for real-world experiments, controlled by an open-loop policy named DexGraspNet 2.0 [69]. More details are provided in Supp. C. + +Task Settings. We introduce disturbances across all environments to induce failures, classified by the affected constraints (e.g., points, lines, surfaces). We evaluate task success rate, execution time, and additional token usage in Omnigibson. Notably, we don't assess failure judgment success rate because our code operates continuously in real-time, rendering this metric inapplicable. Detailed evaluation settings are provided in each section. + +Baselines. We compare DoReMi (DRM) [15] as the baseline for all settings, which uses VLMs to detect intermediate failures during execution via frequent VQA-based queries. + +In CLIPort, we include Inner Monologue [20] baseline, which detects failures only upon each subgoal completion. + +# 4.2. Main Results in Simulator + +# 4.2.1 Results in CLIPort + +We evaluate two tasks in CLIPort: (1) Stack in Order: The robot must stack blocks in a specified order, including two point-level disturbances: (a) with per-step probability $p$ , the suction cup may release a block, causing it to drop; (b) placement positions are perturbed by uniform noise in $[0, q]$ cm, potentially leading to tower collapse. Success is defined as correctly stacking the blocks within 70s. (2) Sweep Half the Blocks: To address the pre-trained policy's tendency to sweep all blocks, the robot must determine when to halt, sweeping half of the blocks ( $\pm 10\%$ ) into a specified colored area within 30s. Success is achieved by meeting these criteria. Tab. 1 shows the mean results over 5 different seeds, each with 12 episodes. Check Supp. D.1 for details. + +Code better monitors 3D space relations. As shown in the Tab. 1, Under the most severe disturbances $(p = 0.3, q = 3)$ in "Stack in Order", CaM achieves a $17.5\%$ higher success rate than DoReMi. Frequent VLM queries lead to increased incorrect failure judgments due to a limited 3D spatial understanding from single images, compared to code-based evaluation of block positions. For example, after placing the green block on the red and preparing to pick up the blue, DoReMi mistakenly thinks that the green block is no longer atop the red, causing it to re-execute the previous subgoal. + +Code with elements achieve reactive and proactive failure detection. As shown in Tab. 1, under severe disturbances in "Stack in Order", CaM reduces execution time by $38.7\%$ and $14.4\%$ compared to Inner Monologue [20], which only detects failures upon subgoal completion, and DoReMi [15], which suffers from misjudgments due to repeated VLM queries, respectively. In contrast, CaM unifies both reactive and proactive failure detection with high precision in real-time, especially in preventing foreseeable failures. For example, if the green block is placed on the red block with heavy placement position disturbance, CaM anticipates that further stacking the blue may lead to collapse and then stabilizes the green block before proceeding. + +Code with elements leads to more accurate counting. As shown in the Tab. 1, in "Sweep Half the Blocks", CaM achieves average success rates $4.5 \times$ higher than DoReMi [15]. Calculating the block points in the designated surface region provides more accurate results than directly using the VLM to count, enabling a more precise halting of the policy to complete the task. + +# 4.2.2 Results in Omnigibson + +We conduct experiments on three tasks in Omnigibson, each involving a distinct type of constraint-element disturbance: + +Table 1. Performance in CLIPort. We report the success rate and execution time, compared to CLIPort (CP) [54], with Inner Monologue (IM) [20] and DoReMi (DRM) [15]. + +
Tasks with disturbanceSuccess Rate(%) ↑Execution Time(s) ↓
CP+IM+DRM+Ours+IM+DRM+Ours
Stack in order with drop pp=0.0100.0100.0100.0100.013.413.413.4
p=0.1556.6781.6783.3395.0034.826.0021.0
p=0.321.6775.0076.6788.3342.834.2025.4
Stack in order with noise qq=190.0090.0096.6798.3324.824.624.2
q=241.6771.6775.0083.3339.437.029.2
q=315.0040.0040.0063.3358.254.236.8
Sweep Half the Blocks0.0018.3316.6775.0022.016.616.4
+ +(1) Slot Pen: Insert a pen into a holder, facing point-level disturbances wherein (a) pen is moved during grasping; (b) pen is dropped during transport; (c) holder is moved during insertion. (2) Stow Book: Place a book to a bookshelf vertically, with line-level disturbances where (a) book is randomly rotated during grasping; (b) end-effector joint is randomly actuated to alter the book pose; (c) book is reoriented horizontally after placement. (3) Pour Tea: Pours from a teapot into a teacup, encountering surface-level disturbances wherein (a) teapot is tilted forward/backward during movement; (b) end-effector joint is actuated to induce a lateral tilt of the teapot during movement; (c) teapot is returned to a horizontal position during pouring. Tab. 2 reports the results across three tasks, each including one no-disturbance trial and three specific-disturbance trials, with 10 runs for each setting. More details are provided in Supp. D.2. + +Code with elements detects richer failures. In Tab. 2, only CaM can detect failures caused by surface-level disturbances in "Pour Tea" compared to DoReMi [15]. The reason is that changes in the teapot's pitch and roll angles are hard to detect by querying VLM via VQA using one current image. The misjudgment leads to a $0\%$ success rate of DoReMi in this task. Notably, DoReMi's success rate is sometimes lower than that of ReKep alone due to VLM's limited spatio-temporal understanding of single images. + +Code with elements detects failure with lower computational cost. As shown in Tab. 2, we achieve a $34.8\%$ reduction in execution time and a $52.2\%$ decrease in token count compared to DoReMi [15]. This improvement stems from proactive failure detection, which prevents more severe failures ahead in real-time for timely re-planning while generating code only once per subgoal. For example, in "Pour Tea", CaM detects failure by monitoring the teapot's tilt angle in real-time, avoiding frequent VLM checks. + +# 4.2.3 Results in RLBench + +We further evaluate CaM on RLbench and demonstrate its superior generalization. Details can be found in Supp. D.3. + +# 4.3. Main Results in Real World + +We conduct real-world evaluations on two tasks: (1) Simple Pick & Place: The robot has 70s to pick up objects and + +Table 2. Performance in Omnigibson. We report the success rate (SR), execution time, and token usage, compared to DoReMi (DRM) [15]. + +
MethodSlot Pen (Point-level Disturbances)Stow Book (Line-level Disturbances)Pour Tea (Surface-level Disturbances)
SR with Disturbance(%) ↑Time (s)↓Token (k)↓SR with Disturbance(%) ↑Time (s)↓Token (k)↓SR with Disturbance(%) ↑Time (s)↓Token (k)↓
NoneDist.(a)Dist.(b)Dist.(c)NoneDist.(a)Dist.(b)Dist.(c)NoneDist.(a)Dist.(b)Dist.(c)
ReKep [21]30201010--40303020--20202010--
+DRM40102020177.8454.5450402040127.1738.670000--
+Ours60504040101.8525.827060706093.0818.6750404030174.5544.19
+ +![](images/660a001cf7222d7caab60e84243d3fb50d8e79091238225a2bfc7a625d82ef67.jpg) +Reasoning Pick &Place Task: Grasp the animals according to their distances to the fruits, from nearest to farthest. + +![](images/30b32e9acf971d7764d35087851fed9ed881248853e32b716bcbcde1517ab5b1.jpg) +Horse is moved, target changed + +![](images/d36186f6b5c1dbf886a26c79c2abd776300562c00a411ea8e9fd516ebaf0568a.jpg) +Figure 5. Example of Real-world Evaluation. The red bounding box shows the current grasp target, which may shift due to environmental changes. CaM monitors and adapts to these changes in real-time, resulting in a closed-loop system with an open-loop policy. + +![](images/a4097cc7255874f1477e174cb053a7d4c83bdb3dfcd789a82b0557e19377088b.jpg) + +![](images/d1fb0406b88c25c6813c4d294041583f54667d02849a04c591f1a43ba9a73e5f.jpg) +Pear is moved, target changed + +![](images/fa256f96746d862e29ef9c4138ce178c58e1a22d6c4e45d2c0950b354aa0defd.jpg) + +![](images/c76d3c8fea3c5410202a403736f70d6009c391076cdb5eb7b23c8b4bd590b806.jpg) + +![](images/928f0c7cf2182f87c7deebdbee23ac940cb16c7e1dd3d312dadae4ec1986a1b0.jpg) + +Table 3. Performance of Single Pick & Place. We report the success rate and execution time. DGN donates DexGraspNet 2.0 [69]. + +
Tasks with disturbanceObject typesSuccess Rate(%) ↑Execution Time(s) ↓
DGN+DRM+Ours+DRM+Ours
Pick & PlaceDeformable0.0083.3396.6761.846.3
with theTransparent0.0066.6793.3372.648.1
objects beingSmall Rigid0.0080.096.6765.745.4
movedLarge Geometric0.0086.6796.6768.945.3
Pick & PlaceDeformable0.0076.6793.3368.762.5
with theTransparent0.0060.0090.0077.762.7
objects beingSmall Rigid0.0063.3393.3369.860.5
removedLarge Geometric0.0076.6796.6772.360.3
+ +Table 4. Performance of Reasoning Pick & Place in cluttered scene. We report the success rate. The robot is controlled by an open-loop policy named DexGraspNet 2.0 (DGN) [69]. w/ CE with $\checkmark$ indicates using constraint elements; otherwise, constraint-aware entities or parts are used for tracking and code computation. + +
Tasksw/ CESuccess Rate(%) ↑
DGN+DRM+Ours
Clear all objects on table except for animals✓0.0010.0060.00
X0.0010.0020.00
Grasp the animals according to their distances to fruits, from nearest to farthest✓0.000.0090.00
+ +place them at specified locations, facing two disturbances: (a) moving the object during grasping, and (b) removing the object from hand during movement. We test four object types, selecting 3 examples per type and conducting 10 trials for each (see Tab. 3). (2) Reasoning Pick & Place: The robot executes long-horizon tasks, involving ambiguous terms (e.g., "fruit"), under the same disturbances. We evaluate 2 long-horizon tasks in cluttered scenes, performing 10 trials each (see Tab. 4). Check Supp. D.4 for details. + +Elements generally abstract constraints and relevant entities. As shown in Tab. 3, using a different end-effector (i.e., Leap Hand [51]) in Simple Pick & Place, CaM also achieves success rates surpassing DoReMi [15] by $20.4\%$ when handling different kinds of objects (e.g., deformable). We find that abstracting constraint-related entities/parts removes the irrelevant visual details, enabling generalization to different kinds of entities in unseen scenes (as discussed in Sec. 3.2), leading to easier tracking and code evaluation. + +Table 5. Performance of reasoning and constraint-aware segmentation. FMC denotes the foundation model combination baseline. + +
MethodReasonSegConstraintSeg
Instance-levelInstance-levelPart-level
gIoUcIoUgIoUcIoUgIoUcIoU
PixelLM [50]56.061.444.443.224.122.6
LISA-13B [28]56.665.142.138.923.424.3
FMC51.253.349.349.640.839.3
ConSeg-13B55.763.962.168.760.265.3
+ +Table 6. Ablation studies in Omnigibson's "Stow Book", assessing the impact of Multi-View (MV), Constraint-aware Segmentation (CS) for elements, and Connect Points (CP) for element formation. + +
MVCSCPSuccess Rate with Disturbance(%) ↑
NoneDist.(a)Dist.(b)Dist.(c)Avg
X✓✓40.040.030.050.040.0
✓X✓50.040.040.040.042.5
✓✓X60.050.060.050.055.0
✓✓✓70.060.070.060.065.0
+ +Reactive and Proactive failure detection combined with an open-loop policy forms a close-loop system. Tab. 4 shows that only $CaM$ successfully handles long-horizon tasks in cluttered scenes. These tasks are challenging as the robot is controlled by an open-loop policy that cannot handle environment dynamics and human disturbances in a closed-loop manner. By incorporating both reactive and proactive failure detection with the open-loop policy (see Fig. 5), the robot can dynamically adjust its target in real-time. For example, when a human moves the horse or pear during the task, the robot adapts by grasping the animal closest to the fruit, forming a closed-loop system. + +# 4.4. Main Results of Segmentation + +Tab. 5 presents segmentation results comparing our ConSeg, with SOTA models and a Foundation Model Combination (FMC) baseline. FMC integrates GPT-4o for reasoning over tasks and constraints to identify relevant instances and parts, along with Grounded SAM [49] and Semantic SAM [32] for instance and part-level segmentation, respectively, similar to our data collection pipeline. We evaluate performance on the ReasonSeg [28] benchmark and our + +![](images/dafd99b9a9b9a1c2b9729ad2ca841eddfb7e36588fc6fc73c4eb0b05070bf488.jpg) +Figure 6. Visual comparison between our ConSeg and LISA [28] at instance and part level. The red masks are the segmentation results. + +Table 7. Ablation study on training data. SemanticSeg includes ADE20K [75], COCO-Stuff [5], PACO-LVIS [48] and PASCALPart [6]. ReferSeg includes refCLEF, refCOCO, refCOCO+ [26] and refCOCOg [39]. VQA indicates LLaVAInstruct-150k [36]. + +
SemanticSegReferSegVQAReasonSegConstraintSegIns-level gIoUcloUPart-level gIoUcloU
✓✓✓✓✘✘42.138.923.424.3
✓✓✓✓✓✘60.765.940.445.6
✓✓✓✓✘✓51.550.656.561.7
✓✓✓✓✓✓62.168.760.265.3
+ +proposed Constraint-Aware Segmentation (ConstraintSeg) benchmark. Our benchmark evaluates performance using gIoU and cIoU, following ReasonSeg's evaluation setting. + +ConSeg performs both reasoning and multi-granularity constraint-aware segmentation. As shown in Tab. 5, ConSeg performs comparably to LISA and PixelLM on the ReasonSeg but significantly surpasses them on ConstraintSeg, achieving nearly a $40\%$ improvement at the part level. Visual comparisons in Fig. 6 further demonstrate ConSeg's strong generalization to unseen objects, scenes, and tasks. + +# 4.5. Ablation Study + +We conduct ablation studies and present analyses below. + +Multi views are critical for visual programming. As shown in Tab. 6, using only front-view images with constraint elements reduces the average success rate from $65\%$ to $40\%$ . This is primarily due to (1) occlusion leading to suboptimal element generation and (2) errors in visual programming, where dimensional reduction causes surfaces to be misinterpreted as lines or lines as points. + +Constraint-aware segmentation enhances element. As shown in Tab. 6, using DINOv2 [42] to generate semantic points to form elements, rather than through constraint-aware segmentation, reduces the success rate from $65\%$ to $42.5\%$ , because these constructed elements fail to represent the desired constraints accurately. For example, capturing a book's vertical orientation requires two precise points on its + +edge, which DINOv2 can not provide. + +Forming elements improves code generation. As shown in Tab. 6, using unconnected 3D points as final elements reduces the success rate from $65\%$ to $55\%$ . Pre-formed elements contain more prior constraint information, serving as visual cues better encoded in code, whereas standalone points require recombination during visual programming. + +Elements simplify constraints computation. As shown in Tab. 4, the success rate decreases from $60\%$ to $20\%$ when entities or parts are not transformed into proposed elements in real-world evaluation. We find two key issues: (1) inaccurate 3D position and slow pose tracking of entities or parts; and (2) difficulties in performing arithmetic operations on their 3D positions or poses in code, hindering the encoding of their spatio-temporal constraints. + +Both instance-level and part-level data improve performance. We conduct ablation studies to assess the impact of our proposed dataset on the ConstraintSeg benchmark. The results are shown in Tab. 7. The results indicate that both the instance-level and part-level subsets contribute to performance improvements and enhance each other's effects. + +# 5. Conclusion + +In this paper, we present a novel paradigm termed Code-as-Monitor, leveraging the VLMs for both open-set reactive and proactive failure detection. In detail, We formulate both detection modes as spatio-temporal constraint satisfaction problems and use VLM-generated code to evaluate them for real-time monitoring. We further propose constraint elements, which abstract constraints-related entities or their parts into compact geometric elements, to improve the precision and efficiency of monitoring. Extensive experiments demonstrate the superiority of the proposed approach and highlight its potential to advance closed-loop robot systems. + +# Acknowledgement + +This work was supported by the National Natural Science Foundation of China (62132001), the National Science and Technology Major Project (2022ZD0116314), and the Fundamental Research Funds for the Central Universities. + +# References + +[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv, 2023. 2, 3, 4, 5 +[2] Aneseh Alvanpour, Sumit Kumar Das, Christopher Kevin Robinson, Olfa Nasraoui, and Dan Popa. Robot failure mode prediction with explainable machine learning. CASE, 2020. 2 +[3] Jingkun An, Yinghao Zhu, Zongjian Li, Enshen Zhou, Haoran Feng, Xijie Huang, Bohua Chen, Yemin Shi, and Chengwei Pan. Agfpsync: Leveraging ai-generated feedback for preference optimization in text-to-image generation. arXiv preprint arXiv:2403.13352, 2024. 2 +[4] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. *NeurIPS*, 2020. 2 +[5] Holger Caesar, Jasper Uijlings, and Vittorio Ferrari. Cocostuff: Thing and stuff classes in context. CVPR, 2018. 8 +[6] Xianjie Chen, Roozbeh Mottaghi, Xiaobai Liu, Sanja Fidler, Raquel Urtasun, and Alan Yuille. Detect what you can: Detecting and representing objects using holistic models and body parts. CVPR, 2014. 8 +[7] Zeren Chen, Zhelun Shi, Xiaoya Lu, Lehan He, Sucheng Qian, Zhenfei Yin, Wanli Ouyang, Jing Shao, Yu Qiao, Cewu Lu, et al. Rh20t-p: A primitive-level robotic dataset towards composable generalization agents. arXiv preprint arXiv:2403.19622, 2024. 2 +[8] Jaemin Cho, Abhay Zala, and Mohit Bansal. Visual programming for step-by-step text-to-image generation and evaluation. NeurIPS, 2024. 3 +[9] Wenbo Cui, Chengyang Zhao, Songlin Wei, Jiazhao Zhang, Haoran Geng, Yaran Chen, and He Wang. Gapartmanip: A large-scale part-centric dataset for material-agnostic articulated object manipulation. arXiv preprint arXiv:2411.18276, 2024. 2 +[10] Yinpei Dai, Jayjun Lee, Nima Fazeli, and Joyce Chai. Racer: Rich language-guided failure recovery policies for imitation learning. arXiv, 2024. 2 +[11] Maximilian Diehl and Karinne Ramirez-Amaro. A causal-based approach to explain, predict and prevent failures in robotic tasks. RAS, 2023. 2 +[12] Yuqing Du, Ksenia Konyushkova, Misha Denil, Akhil Raju, Jessica Landon, Felix Hill, Nando de Freitas, and Serkan Cabi. Vision-language models as success detectors. arXiv, 2023. 2 +[13] Jiafei Duan, Wilbert Pumacay, Nishanth Kumar, Yi Ru Wang, Shulin Tian, Wentao Yuan, Ranjay Krishna, Dieter + +Fox, Ajay Mandlekar, and Yijie Guo. Aha: A vision-language-model for detecting and reasoning over failures in robotic manipulation. arXiv, 2024. 2 +[14] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv, 2024. 2 +[15] Yanjiang Guo, Yen-Jen Wang, Lihan Zha, Zheyuan Jiang, and Jianyu Chen. Doremi: Grounding language model by detecting and recovering from plan-execution misalignment. arXiv, 2023. 2, 5, 6, 7 +[16] Tanmay Gupta and Aniruddha Kembhavi. Visual programming: Compositional visual reasoning without training. CVPR, 2023. 3 +[17] Edward J Hu, Phillip Wallis, Zeyuan Allen-Zhu, Yanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. *ICLR*, 2021. 5 +[18] Xuhao Hu, Dongrui Liu, Hao Li, Xuanjing Huang, and Jing Shao. Vlsbench: Unveiling visual leakage in multimodal safety. arXiv preprint arXiv:2411.19939, 2024. 2 +[19] Haoxu Huang, Fanqi Lin, Yingdong Hu, Shengjie Wang, and Yang Gao. Copa: General robotic manipulation through spatial constraints of parts with foundation models. arXiv, 2024. 3 +[20] Wenlong Huang, Fei Xia, Ted Xiao, Harris Chan, Jacky Liang, Pete Florence, Andy Zeng, Jonathan Tompson, Igor Mordatch, Yevgen Chebotar, et al. Inner monologue: Embodied reasoning through planning with language models. arXiv, 2022. 2, 6 +[21] Wenlong Huang, Chen Wang, Yunzhu Li, Ruohan Zhang, and Li Fei-Fei. Rekep: Spatio-temporal reasoning of relational keypoint constraints for robotic manipulation. arXiv, 2024. 3, 5, 7 +[22] Stephen James, Zicong Ma, David Rovick Arrojo, and Andrew J. Davison. Rlbench: The robot learning benchmark & learning environment. RAL, 2020. 2, 5 +[23] Yuheng Ji, Huajie Tan, Jiayu Shi, Xiaoshuai Hao, Yuan Zhang, Hengyuan Zhang, Pengwei Wang, Mengdi Zhao, Yao Mu, Pengju An, et al. Robobrain: A unified brain model for robotic manipulation from abstract to concrete. arXiv preprint arXiv:2502.21257, 2025. 2 +[24] Yueru Jia, Jiaming Liu, Sixiang Chen, Chenyang Gu, Zhilue Wang, Longzan Luo, Lily Lee, Pengwei Wang, Zhongyuan Wang, Renrui Zhang, et al. Lift3d foundation policy: Lifting 2d large-scale pretrained models for robust 3d robotic manipulation. arXiv preprint arXiv:2411.18623, 2024. 2 +[25] Nikita Karaev, Ignacio Rocco, Benjamin Graham, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. Co-tracker: It is better to track together. ECCV, 2024. 5 +[26] Sahar Kazemzadeh, Vicente Ordonez, Mark Matten, and Tamara Berg. Referitag: Referring to objects in photographs of natural scenes. EMNLP, 2014. 8 +[27] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. ICCV, 2023. 3, 4 + +[28] Xin Lai, Zhuotao Tian, Yukang Chen, Yanwei Li, Yuhui Yuan, Shu Liu, and Jiaya Jia. Lisa: Reasoning segmentation via large language model. CVPR, 2024. 4, 7, 8 +[29] Olivia Y Lee, Annie Xie, Kuan Fang, Karl Pertsch, and Chelsea Finn. Affordance-guided reinforcement learning via visual prompting. arXiv, 2024. 3 +[30] Gregory LeMasurier, Alvika Gautam, Zhao Han, Jacob W Crandall, and Holly A Yanco. Reactive or proactive? how robots should explain failures. HRI, 2024. 2 +[31] Chengshu Li, Ruohan Zhang, Josiah Wong, Cem Gokmen, Sanjana Srivastava, Roberto Martin-Martin, Chen Wang, Gabrael Levine, Michael Lingelbach, Jiankai Sun, Mona Anvari, Minjune Hwang, Manasi Sharma, Arman Aydin, Dhruva Bansal, Samuel Hunter, Kyu-Young Kim, Alan Lou, Caleb R Matthews, Ivan Villa-Renteria, Jerry Huayang Tang, Claire Tang, Fei Xia, Silvio Savarese, Hyowon Gweon, Karen Liu, Jiajun Wu, and Li Fei-Fei. Behavior-1k: A benchmark for embodied ai with 1,000 everyday activities and realistic simulation. CoRL, 2022. 2, 5 +[32] Feng Li, Hao Zhang, Peize Sun, Xueyan Zou, Shilong Liu, Jianwei Yang, Chunyuan Li, Lei Zhang, and Jianfeng Gao. Semantic-sam: Segment and recognize anything at any granularity. arXiv, 2023. 5, 7 +[33] Lijun Li, Zhelun Shi, Xuhao Hu, Bowen Dong, Yiran Qin, Xihui Liu, Lu Sheng, and Jing Shao. T2isafety: Benchmark for assessing fairness, toxicity, and privacy in image generation. arXiv preprint arXiv:2501.12612, 2025. 2 +[34] Jacky Liang, Wenlong Huang, Fei Xia, Peng Xu, Karol Hausman, Brian Ichter, Pete Florence, and Andy Zeng. Code as policies: Language model programs for embodied control. ICRA, 2023. 3 +[35] Fangchen Liu, Kuan Fang, Pieter Abbeel, and Sergey Levine. Moka: Open-vocabulary robotic manipulation through mark-based visual prompting. arXiv, 2024. 3 +[36] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. NeurIPS, 2024. 2, 4, 8 +[37] Jiaming Liu, Chenxuan Li, Guanqun Wang, Lily Lee, Kaichen Zhou, Sixiang Chen, Chuyan Xiong, Jiaxin Ge, Renrui Zhang, and Shanghang Zhang. Self-corrected multimodal large language model for end-to-end robot manipulation. arXiv, 2024. 2 +[38] Zeyi Liu, Arpit Bahety, and Shuran Song. Reflect: Summarizing robot experiences for failure explanation and correction. CoRL, 2023. 2 +[39] Junhua Mao, Jonathan Huang, Alexander Toshev, Oana Camburu, Alan L Yuille, and Kevin Murphy. Generation and comprehension of unambiguous object descriptions. CVPR, 2016. 8 +[40] Yao Mu, Junting Chen, Qinglong Zhang, Shoufa Chen, Qiaojun Yu, Chongjian Ge, Runjian Chen, Zhixuan Liang, Mengkang Hu, Chaofan Tao, et al. Robocodex: Multimodal code generation for robotic behavior synthesis. arXiv, 2024. 3 +[41] Soroush Nasiriany, Fei Xia, Wenhao Yu, Ted Xiao, Jacky Liang, Ishita Dasgupta, Annie Xie, Danny Driess, Ayzaan Wahid, Zhuo Xu, et al. Pivot: Iterative visual prompting elicits actionable knowledge for vlms. arXiv, 2024. 3 + +[42] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv, 2023. 3, 8 +[43] Zekun Qi, Wenyao Zhang, Yufei Ding, Runpei Dong, Xinqiang Yu, Jingwen Li, Lingyun Xu, Baoyu Li, Xialin He, Guofan Fan, et al. Sofar: Language-grounded orientation bridges spatial reasoning and object manipulation. arXiv preprint arXiv:2502.13143, 2025. 2 +[44] Yiran Qin, Zhelun Shi, Jiwen Yu, Xijun Wang, Enshen Zhou, Lijun Li, Zhenfei Yin, Xihui Liu, Lu Sheng, Jing Shao, et al. Worldsimbench: Towards video generation models as world simulators. arXiv preprint arXiv:2410.18072, 2024. 2 +[45] Yiran Qin, Enshen Zhou, Qichang Liu, Zhenfei Yin, Lu Sheng, Ruimao Zhang, Yu Qiao, and Jing Shao. Mp5: A multi-modal open-ended embodied system in mycraf via active perception. In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16307-16316. IEEE, 2024. 2 +[46] Yiran Qin, Ao Sun, Yuze Hong, Benyou Wang, and Ruimao Zhang. Navigatediff: Visual predictors are zero-shot navigation assistants. arXiv preprint arXiv:2502.13894, 2025. 2 +[47] Shreyas Sundara Raman, Vanya Cohen, Eric Rosen, Ifrah Idrees, David Paulius, and Stefanie Tellex. Planning with large language models via corrective re-prompting. NeurIPS, 2022. 2 +[48] Vignesh Ramanathan, Anmol Kalia, Vladan Petrovic, Yi Wen, Baixue Zheng, Baishan Guo, Rui Wang, Aaron Marquez, Rama Kovvuri, Abhishek Kadian, et al. Paco: Parts and attributes of common objects. CVPR, 2023. 8 +[49] Tianhe Ren, Shilong Liu, Ailing Zeng, Jing Lin, Kunchang Li, He Cao, Jiayu Chen, Xinyu Huang, Yukang Chen, Feng Yan, Zhaoyang Zeng, Hao Zhang, Feng Li, Jie Yang, Hongyang Li, Qing Jiang, and Lei Zhang. Grounded sam: Assembling open-world models for diverse visual tasks. arXiv, 2024. 5, 7 +[50] Zhongwei Ren, Zhicheng Huang, Yunchao Wei, Yao Zhao, Dongmei Fu, Jiashi Feng, and Xiaojie Jin. Pixel reasoning with large multimodal model. CVPR, 2024. 7 +[51] Kenneth Shaw, Ananye Agarwal, and Deepak Pathak. Leap hand: Low-cost, efficient, and anthropomorphic hand for robot learning. arXiv, 2023. 5, 7 +[52] Zhelun Shi, Zhipin Wang, Hongxing Fan, Zaibin Zhang, Lijun Li, Yongting Zhang, Zhenfei Yin, Lu Sheng, Yu Qiao, and Jing Shao. Assessment of multimodal large language models in alignment with human values. arXiv preprint arXiv:2403.17830, 2024. 2 +[53] Noah Shinn, Beck Labash, and Ashwin Gopinath. Reflection: an autonomous agent with dynamic memory and self-reflection. arXiv, 2023. 2 +[54] Mohit Shridhar, Lucas Manuelli, and Dieter Fox. CIport: What and where pathways for robotic manipulation. CoRL, 2021. 2, 5, 6 +[55] Tom Silver, Soham Dan, Kavitha Srinivas, Joshua B Tenenbaum, Leslie Kaelbling, and Michael Katz. Generalized planning in pddl domains with pretrained large language models. AAAI, 2024. 2 + +[56] Marta Skreta, Naruki Yoshikawa, Sebastian Arellano-Rubach, Zhi Ji, Lasse Bjørn Kristensen, Kourosh Darvish, Alán Aspuru-Guzik, Florian Shkurti, and Animesh Garg. Errors are useful prompts: Instruction guided task programming with verifier-assisted iterative prompting. arXiv, 2023. 2 +[57] Yingbo Tang, Shuaike Zhang, Xiaoshuai Hao, Pengwei Wang, Jianlong Wu, Zhongyuan Wang, and Shanghang Zhang. Affordgrasp: In-context affordance reasoning for open-vocabulary task-oriented grasping in clutter. arXiv preprint arXiv:2503.00778, 2025. 2 +[58] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv, 2023. 2 +[59] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv, 2023. 2 +[60] Sai H Vemprala, Rogerio Bonatti, Arthur Tucker, and Ashish Kapoor. Chatgpt for robotics: Design principles and model abilities. IEEE Access, 2024. 3 +[61] David Venuto, Sami Nur Islam, Martin Klissarov, Doina Precup, Sherry Yang, and Ankit Anand. Code as reward: Empowering reinforcement learning with vlms. arXiv, 2024. 3 +[62] Homer Rich Walke, Kevin Black, Tony Z Zhao, Quan Vuong, Chongyi Zheng, Philippe Hansen-Estruch, Andre Wang He, Vivek Myers, Moo Jin Kim, Max Du, et al. Bridgedata v2: A dataset for robot learning at scale. CoRL, 2023. 5 +[63] Jianwei Yang, Hao Zhang, Feng Li, Xueyan Zou, Chunyuan Li, and Jianfeng Gao. Set-of-mark prompting unleashes extraordinary visual grounding in gpt-4v. arXiv, 2023. 3 +[64] Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, and Yuan Cao. React: Synergizing reasoning and acting in language models. arXiv, 2022. 2 +[65] Runpeng Yu, Weihao Yu, and Xinchao Wang. Attention prompting on image for large vision-language models. arXiv, 2024. 3 +[66] Zhouliang Yu, Jie Fu, Yao Mu, Chenguang Wang, Lin Shao, and Yaodong Yang. Multireact: Multimodal tools augmented reasoning-acting traces for embodied agent planning. *ICLR*, 2024. 2 +[67] Wentao Yuan, Jiafei Duan, Valts Blukis, Wilbert Pumacay, Ranjay Krishna, Adithyavairavan Murali, Arsalan Mousavian, and Dieter Fox. Robopoint: A vision-language model for spatial affordance prediction for robotics. arXiv, 2024. 3 +[68] Zhihao Yuan, Jinke Ren, Chun-Mei Feng, Hengshuang Zhao, Shuguang Cui, and Zhen Li. Visual programming for zero-shot open-vocabulary 3d visual grounding. CVPR, 2024. 3 +[69] Jialiang Zhang, Haoran Liu, Danshi Li, XinQiang Yu, Hao ran Geng, Yufei Ding, Jiayi Chen, and He Wang. Dexgrasp-net 2.0: Learning generative dexterous grasping in largescale synthetic cluttered scenes. CoRL, 2024. 5, 7 + +[70] Xinyu Zhang, Yuhan Liu, Haonan Chang, Liam Schramm, and Abdeslam Boularias. Autoregressive action sequence learning for robotic manipulation. arXiv, 2024. 5 +[71] Yongting Zhang, Lu Chen, Guodong Zheng, Yifeng Gao, Rui Zheng, Jinlan Fu, Zhenfei Yin, Senjie Jin, Yu Qiao, Xuanjing Huang, et al. Spa-vl: A comprehensive safety preference alignment dataset for vision language model. arXiv preprint arXiv:2406.12030, 2024. 2 +[72] Zaibin Zhang, Shiyu Tang, Yuanhang Zhang, Talas Fu, Yifan Wang, Yang Liu, Dong Wang, Jing Shao, Lijun Wang, and Huchuan Lu. Ad-h: Autonomous driving with hierarchical agents. arXiv preprint arXiv:2406.03474, 2024. 2 +[73] Jinliang Zheng, Jianxiong Li, Sijie Cheng, Yinan Zheng, Ji-aming Li, Jihao Liu, Yu Liu, Jingjing Liu, and Xianyuan Zhan. Instruction-guided visual masking. arXiv, 2024. 3 +[74] Zhi Zheng, Qian Feng, Hang Li, Alois Knoll, and Jianxiang Feng. Evaluating uncertainty-based failure detection for closed-loop llm planners. arXiv, 2024. 2 +[75] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. CVPR, 2017. 8 +[76] Enshen Zhou, Yiran Qin, Zhenfei Yin, Yuzhou Huang, Ruimao Zhang, Lu Sheng, Yu Qiao, and Jing Shao. Minedreamer: Learning to follow instructions via chain-of-imagination for simulated-world control. arXiv preprint arXiv:2403.12037, 2024. 2 \ No newline at end of file diff --git a/2025/Code-as-Monitor_ Constraint-aware Visual Programming for Reactive and Proactive Robotic Failure Detection/images.zip b/2025/Code-as-Monitor_ Constraint-aware Visual Programming for Reactive and Proactive Robotic Failure Detection/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..dbd762ccafb7666a5cb6ef9dfa801fdcc00bcd17 --- /dev/null +++ b/2025/Code-as-Monitor_ Constraint-aware Visual Programming for Reactive and Proactive Robotic Failure Detection/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:eac617f48c2849324a78d362b0f1bd6d46142c5ff99b1dbe7d068122d13b687d +size 686027 diff --git a/2025/Code-as-Monitor_ Constraint-aware Visual Programming for Reactive and Proactive Robotic Failure Detection/layout.json b/2025/Code-as-Monitor_ Constraint-aware Visual Programming for Reactive and Proactive Robotic Failure Detection/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..e6dffb493694f4ec3ede89294c3392545b702c00 --- /dev/null +++ b/2025/Code-as-Monitor_ Constraint-aware Visual Programming for Reactive and Proactive Robotic Failure Detection/layout.json @@ -0,0 +1,9800 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 116, + 103, + 496, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 103, + 496, + 138 + ], + "spans": [ + { + "bbox": [ + 116, + 103, + 496, + 138 + ], + "type": "text", + "content": "Code-as-Monitor: Constraint-aware Visual Programming for Reactive and Proactive Robotic Failure Detection" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 96, + 160, + 517, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 160, + 517, + 203 + ], + "spans": [ + { + "bbox": [ + 96, + 160, + 517, + 203 + ], + "type": "text", + "content": "Enshen Zhou" + }, + { + "bbox": [ + 96, + 160, + 517, + 203 + ], + "type": "inline_equation", + "content": "^{1,3*}" + }, + { + "bbox": [ + 96, + 160, + 517, + 203 + ], + "type": "text", + "content": ", Qi Su" + }, + { + "bbox": [ + 96, + 160, + 517, + 203 + ], + "type": "inline_equation", + "content": "^{2,3,4*}" + }, + { + "bbox": [ + 96, + 160, + 517, + 203 + ], + "type": "text", + "content": ", Cheng Chi" + }, + { + "bbox": [ + 96, + 160, + 517, + 203 + ], + "type": "inline_equation", + "content": "^{3*†}" + }, + { + "bbox": [ + 96, + 160, + 517, + 203 + ], + "type": "text", + "content": ", Zhizheng Zhang" + }, + { + "bbox": [ + 96, + 160, + 517, + 203 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 96, + 160, + 517, + 203 + ], + "type": "text", + "content": ", Zhongyuan Wang" + }, + { + "bbox": [ + 96, + 160, + 517, + 203 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 96, + 160, + 517, + 203 + ], + "type": "text", + "content": ", Tiejun Huang" + }, + { + "bbox": [ + 96, + 160, + 517, + 203 + ], + "type": "inline_equation", + "content": "^{2,3}" + }, + { + "bbox": [ + 96, + 160, + 517, + 203 + ], + "type": "text", + "content": ", Lu Sheng" + }, + { + "bbox": [ + 96, + 160, + 517, + 203 + ], + "type": "inline_equation", + "content": "^{1†}" + }, + { + "bbox": [ + 96, + 160, + 517, + 203 + ], + "type": "text", + "content": ", He Wang" + }, + { + "bbox": [ + 96, + 160, + 517, + 203 + ], + "type": "inline_equation", + "content": "^{2,3,4†}" + }, + { + "bbox": [ + 96, + 160, + 517, + 203 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 96, + 160, + 517, + 203 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 96, + 160, + 517, + 203 + ], + "type": "text", + "content": "Beihang University " + }, + { + "bbox": [ + 96, + 160, + 517, + 203 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 96, + 160, + 517, + 203 + ], + "type": "text", + "content": "Peking University " + }, + { + "bbox": [ + 96, + 160, + 517, + 203 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 96, + 160, + 517, + 203 + ], + "type": "text", + "content": "Beijing Academy of Artificial Intelligence " + }, + { + "bbox": [ + 96, + 160, + 517, + 203 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 96, + 160, + 517, + 203 + ], + "type": "text", + "content": "Galbot" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 81, + 206, + 529, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 206, + 529, + 217 + ], + "spans": [ + { + "bbox": [ + 81, + 206, + 529, + 217 + ], + "type": "text", + "content": "zhouenshen@buaa.edu.cn chicheng15@mails.ucas.ac.cn lsheng@buaa.edu.cn hewang@pku.edu.cn" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 63, + 228, + 83, + 247 + ], + "blocks": [ + { + "bbox": [ + 63, + 228, + 83, + 247 + ], + "lines": [ + { + "bbox": [ + 63, + 228, + 83, + 247 + ], + "spans": [ + { + "bbox": [ + 63, + 228, + 83, + 247 + ], + "type": "image", + "image_path": "02514d55c0840685d9698546835ab2425a6bcdb26c727621e2fc7aca1f41445f.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 86, + 232, + 463, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 232, + 463, + 245 + ], + "spans": [ + { + "bbox": [ + 86, + 232, + 463, + 245 + ], + "type": "text", + "content": "Task: Move the pan with the lobster to the stove, and be careful not to let the lobster drop out." + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 58, + 247, + 268, + 393 + ], + "blocks": [ + { + "bbox": [ + 58, + 247, + 268, + 393 + ], + "lines": [ + { + "bbox": [ + 58, + 247, + 268, + 393 + ], + "spans": [ + { + "bbox": [ + 58, + 247, + 268, + 393 + ], + "type": "image", + "image_path": "ad2a9b3059d16dfce8fb97ac8eaa7705617ecc383e20ea57c9c28f978527b444.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 104, + 394, + 222, + 405 + ], + "lines": [ + { + "bbox": [ + 104, + 394, + 222, + 405 + ], + "spans": [ + { + "bbox": [ + 104, + 394, + 222, + 405 + ], + "type": "text", + "content": "(a) Reactive Failure Detection" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 268, + 247, + 478, + 393 + ], + "blocks": [ + { + "bbox": [ + 268, + 247, + 478, + 393 + ], + "lines": [ + { + "bbox": [ + 268, + 247, + 478, + 393 + ], + "spans": [ + { + "bbox": [ + 268, + 247, + 478, + 393 + ], + "type": "image", + "image_path": "afb1dd77eb2d946721e72019f19549490e43e8d7270770e895d49b66c2d4fb41.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 312, + 394, + 434, + 403 + ], + "lines": [ + { + "bbox": [ + 312, + 394, + 434, + 403 + ], + "spans": [ + { + "bbox": [ + 312, + 394, + 434, + 403 + ], + "type": "text", + "content": "(b) Proactive Failure Detection" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 479, + 230, + 553, + 394 + ], + "blocks": [ + { + "bbox": [ + 479, + 230, + 553, + 394 + ], + "lines": [ + { + "bbox": [ + 479, + 230, + 553, + 394 + ], + "spans": [ + { + "bbox": [ + 479, + 230, + 553, + 394 + ], + "type": "image", + "image_path": "813e5a2c28f246840e12ff1c4a1c1c84d785cda759cce1ba9f8fc0af6e4bc184.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 477, + 395, + 552, + 405 + ], + "lines": [ + { + "bbox": [ + 477, + 395, + 552, + 405 + ], + "spans": [ + { + "bbox": [ + 477, + 395, + 552, + 405 + ], + "type": "text", + "content": "(c) Real-world Test" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 54, + 409, + 555, + 489 + ], + "lines": [ + { + "bbox": [ + 54, + 409, + 555, + 489 + ], + "spans": [ + { + "bbox": [ + 54, + 409, + 555, + 489 + ], + "type": "text", + "content": "Figure 1. For the task \"Move the pan with lobster to the stove without losing the lobster\", (a) reactive failure detection identifies failures after they occur, and (b) proactive failure detection prevents foreseeable failures. In (a), at " + }, + { + "bbox": [ + 54, + 409, + 555, + 489 + ], + "type": "inline_equation", + "content": "t_4^R" + }, + { + "bbox": [ + 54, + 409, + 555, + 489 + ], + "type": "text", + "content": ", the robot detects the failure after the lobster unpredictably jumps out due to the heat. In (b), pan tilting is detected at " + }, + { + "bbox": [ + 54, + 409, + 555, + 489 + ], + "type": "inline_equation", + "content": "t_3^P" + }, + { + "bbox": [ + 54, + 409, + 555, + 489 + ], + "type": "text", + "content": " and corrected it at " + }, + { + "bbox": [ + 54, + 409, + 555, + 489 + ], + "type": "inline_equation", + "content": "t_3^{P'}" + }, + { + "bbox": [ + 54, + 409, + 555, + 489 + ], + "type": "text", + "content": ", requiring real-time precision. We formulate both tasks as spatio-temporal constraint satisfaction problems, leveraging our proposed constraint elements for precise, real-time checking. For example, in (a), a large relative distance between pan and lobster indicates failure; in (b), a large angle between the pan and the horizontal plane needs correction. (c) shows that our method combined with an open-loop policy forms a closed-loop system, enabling proactive (e.g., detecting moving glass during grasping) and reactive (e.g., removing toy after grasping) failure detection in cluttered scenes." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 151, + 498, + 200, + 510 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 498, + 200, + 510 + ], + "spans": [ + { + "bbox": [ + 151, + 498, + 200, + 510 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 53, + 524, + 297, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 524, + 297, + 693 + ], + "spans": [ + { + "bbox": [ + 53, + 524, + 297, + 693 + ], + "type": "text", + "content": "Automatic detection and prevention of open-set failures are crucial in closed-loop robotic systems. Recent studies often struggle to simultaneously identify unexpected failures reactively after they occur and prevent foreseeable ones proactively. To this end, we propose Code-as-Monitor (CaM), a novel paradigm leveraging the vision-language model (VLM) for both open-set reactive and proactive failure detection. The core of our method is to formulate both tasks as a unified set of spatio-temporal constraint satisfaction problems and use VLM-generated code to evaluate them for real-time monitoring. To enhance the accuracy and efficiency of monitoring, we further introduce constraint elements that abstract constraint-related entities or their parts into compact geometric ele" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 498, + 556, + 631 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 556, + 631 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 556, + 631 + ], + "type": "text", + "content": "ments. This approach offers greater generality, simplifies tracking, and facilitates constraint-aware visual programming by leveraging these elements as visual prompts. Experiments show that CaM achieves a " + }, + { + "bbox": [ + 313, + 498, + 556, + 631 + ], + "type": "inline_equation", + "content": "28.7\\%" + }, + { + "bbox": [ + 313, + 498, + 556, + 631 + ], + "type": "text", + "content": " higher success rate and reduces execution time by " + }, + { + "bbox": [ + 313, + 498, + 556, + 631 + ], + "type": "inline_equation", + "content": "31.8\\%" + }, + { + "bbox": [ + 313, + 498, + 556, + 631 + ], + "type": "text", + "content": " under severe disturbances compared to baselines across three simulators and a real-world setting. Moreover, CaM can be integrated with open-loop control policies to form closed-loop systems, enabling long-horizon tasks in cluttered scenes with dynamic environments. See the project page at https://zhoues.github.io/Code-as-Monitor." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 314, + 644, + 395, + 657 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 644, + 395, + 657 + ], + "spans": [ + { + "bbox": [ + 314, + 644, + 395, + 657 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "content": "As expectations grow for robots to handle long-horizon tasks within intricate environments, failures are unavoidable. Therefore, automatically detecting and preventing those failures plays a vital role in ensuring the tasks can" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 70, + 702, + 222, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 702, + 222, + 714 + ], + "spans": [ + { + "bbox": [ + 70, + 702, + 222, + 714 + ], + "type": "text", + "content": "* Equal contribution † Corresponding author" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "type": "text", + "content": "6919" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 192 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 192 + ], + "type": "text", + "content": "eventually be solved, especially for closed-loop robotic systems. There are two modes of failure detection [30], reactive and proactive. As depicted in Fig. 1, reactive failure detection aims to identify failures after they occur (e.g., recognizing that lobster lands on the table, indicating a delivery failure). In contrast, proactive failure detection aims to prevent foreseeable failures (e.g., recognizing that a tilted pan could cause the lobster to fall out, leading to a delivery failure). Both detection modes are even more challenging in open-set scenarios, where the failures are not predefined." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 194, + 294, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 194, + 294, + 361 + ], + "spans": [ + { + "bbox": [ + 55, + 194, + 294, + 361 + ], + "type": "text", + "content": "With the help of large language models (LLMs) [14, 58] and vision-language models (VLMs) [1, 36], recent studies can achieve open-set reactive failure detection [10, 12, 13, 15, 20, 37, 55, 56, 66, 74] as a special case of visual question answering (VQA) tasks. However, these methods often bear compromised execution speeds and coarse-grained detection accuracy, due to the high computational costs and inadequate 3D spatio-temporal perception capability of recent LLMs/VLMs. Moreover, open-set proactive failure detection, which has been less explored in the literature, presents even severer challenges as it is required to foresee potential causes of failure and monitor them in real-time with high precision to anticipate and prevent imminent failure. Simply adapting LLMs/VLMs cannot meet such expectations." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 364, + 295, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 364, + 295, + 555 + ], + "spans": [ + { + "bbox": [ + 55, + 364, + 295, + 555 + ], + "type": "text", + "content": "In this work, we aim to develop an open-set failure detection framework that achieves reactive and proactive detection simultaneously, not only benefiting from the generalization power of VLMs but also enjoying high precision in monitoring failure characteristics with real-time efficiency. We address this by formulating both tasks as a unified set of spatio-temporal constraint satisfaction problems, which can be precisely translated by VLMs into executable programs. Such visual programs can efficiently verify whether entities (e.g., robots, objects) or their parts in the captured environment maintain or achieve required states during or after execution (i.e., satisfying constraints), so as to immediately prevent or detect failures. To the best of our knowledge, this is the first attempt to integrate both detection modes within a single framework. We name this constraint-aware visual programming framework as Code-as-Monitor " + }, + { + "bbox": [ + 55, + 364, + 295, + 555 + ], + "type": "inline_equation", + "content": "(CaM)" + }, + { + "bbox": [ + 55, + 364, + 295, + 555 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "type": "text", + "content": "To be specific, the proposed spatio-temporal constraint satisfaction scheme abstracts the constraint-related entity or part segments from the observed images into compact geometric elements (e.g., points, lines, and surfaces), as shown in Fig. 1. It simplifies the monitoring of constraint satisfaction by tracking and evaluating the spatio-temporal combinational dynamics of these elements, eliminating the most irrelevant geometric and visual details of the raw entities/parts. The constraint element detection and tracking are grounded by our trained constraint-aware segmentation and off-the-shelf tracking models, ensuring speed, accuracy, and certain open-set adaptation capabilities. The evaluation protocol is in the form of VLM-generated code, i.e., monitor" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 553, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 215 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 215 + ], + "type": "text", + "content": "code, which is visually prompted by the starting frames of a sub-goal and its associated constraint elements, in addition to the textual constraints that must be fulfilled. Once generated, this monitor code could detect reactive or proactive failures just by being executed according to the tracked constraint elements, without needing to call the VLMs again. Therefore, this minimalist scheme is generalizable to open-set failure detection for unseen entities and scenes (enabled by the potential diversity of the structured associations of constraint elements) as well as common skills (powered by the rich prior knowledge offered by VLMs), maintaining sufficient detection accuracy and real-time execution speed." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 216, + 553, + 393 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 216, + 553, + 393 + ], + "spans": [ + { + "bbox": [ + 313, + 216, + 553, + 393 + ], + "type": "text", + "content": "We conduct extensive experiments in three simulators (i.e., CLIPort [54], Omnigibson [31], and RL-Bench [22]) and one real-world setting, spanning diverse manipulation tasks (e.g., pick & place, articulated objects, tool-use), robot platforms (e.g., UR5, Fetch, Franka) and end-effectors (e.g., suction cup, gripper, dexterous hand). The results show that CaM is generalizable and achieves both reactive and proactive failure detection in real-time, resulting in " + }, + { + "bbox": [ + 313, + 216, + 553, + 393 + ], + "type": "inline_equation", + "content": "28.7\\%" + }, + { + "bbox": [ + 313, + 216, + 553, + 393 + ], + "type": "text", + "content": " higher success rates and reduced execution time by " + }, + { + "bbox": [ + 313, + 216, + 553, + 393 + ], + "type": "inline_equation", + "content": "31.8\\%" + }, + { + "bbox": [ + 313, + 216, + 553, + 393 + ], + "type": "text", + "content": " under severe disturbances compared to the baselines. Moreover, in Sec. 4.3, CaM can be integrated with the existing open-loop control policy to form a closed-loop system, enabling long-horizon tasks in cluttered scenes with environment dynamics and human disturbances. Our contributions are summarized as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 396, + 553, + 490 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 314, + 396, + 553, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 396, + 553, + 430 + ], + "spans": [ + { + "bbox": [ + 314, + 396, + 553, + 430 + ], + "type": "text", + "content": "- We introduce Code-as-Monitor (CaM), a novel paradigm that leverages VLMs for both reactive and proactive failure detection via constraint-aware visual programming." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 431, + 553, + 454 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 431, + 553, + 454 + ], + "spans": [ + { + "bbox": [ + 314, + 431, + 553, + 454 + ], + "type": "text", + "content": "- We propose the constraint elements to enhance the accuracy and efficiency of constraint satisfaction monitoring." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 456, + 553, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 456, + 553, + 490 + ], + "spans": [ + { + "bbox": [ + 314, + 456, + 553, + 490 + ], + "type": "text", + "content": "- Extensive experiments show that CaM is generalizable and achieves more real-time and precise failure detection than baselines across simulators and real-world settings." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 502, + 399, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 502, + 399, + 514 + ], + "spans": [ + { + "bbox": [ + 314, + 502, + 399, + 514 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 521, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 553, + 713 + ], + "type": "text", + "content": "Robotic Failure Detection. Recent advances in LLMs [4, 14, 58, 59] and VLMs [1, 3, 18, 33, 36, 44, 45, 52, 71, 76] greatly improve open-set reactive failure detection for robotics [7, 9, 23, 24, 43, 46, 57, 72] Current LLM-based methods either convert visual inputs into text [38, 53, 64], potentially losing visual details, or rely on ground-truth feedback [20, 47, 55, 56], which is impractical in real-world scenarios. Recent studies employ VLMs as failure detectors, offering binary success indicators [12, 37, 66, 74] or textual explanations [10, 13] through visual question answering (VQA), such as DoReMi [15]. However, they often bear compromised execution speeds and coarse-grained detection accuracy. Meanwhile, open-set proactive failure detection is rarely explored, as previous methods [2, 11] are confined to predefined failures. This detection mode requires foreseeing potential failure causes and monitoring" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "6920" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 71, + 555, + 292 + ], + "blocks": [ + { + "bbox": [ + 56, + 71, + 555, + 292 + ], + "lines": [ + { + "bbox": [ + 56, + 71, + 555, + 292 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 555, + 292 + ], + "type": "image", + "image_path": "ddded73361f0a4434fdf1b72370125485c3d7d83cf3d628cb57583acb5b99e7d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 295, + 555, + 351 + ], + "lines": [ + { + "bbox": [ + 55, + 295, + 555, + 351 + ], + "spans": [ + { + "bbox": [ + 55, + 295, + 555, + 351 + ], + "type": "text", + "content": "Figure 2. Overview of Code-as-Monitor. Given task instructions and prior information, the Constraint Generator derives the next subgoal and corresponding textual constraints based on multi-view observations. The Painter maps these constraints onto images as constraint elements. The Monitor generates monitor code from these images and tracks them for real-time monitoring. If any constraint is violated, it outputs the reason for failure and triggers re-planning. This framework unifies reactive and proactive failure detection via constraints, more generally abstracts relevant entities/parts through constraint elements, and ensures precise and real-time monitoring via code evaluation." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 357, + 295, + 405 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 357, + 295, + 405 + ], + "spans": [ + { + "bbox": [ + 55, + 357, + 295, + 405 + ], + "type": "text", + "content": "them in real-time with high precision. Herein we formulate both reactive and proactive failure detection as constraint satisfaction problems and use VLM-generated code to evaluate them, meeting the expectations of both modes above." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 409, + 295, + 637 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 409, + 295, + 637 + ], + "spans": [ + { + "bbox": [ + 55, + 409, + 295, + 637 + ], + "type": "text", + "content": "Visual Prompting. Visual prompts enhance the visual reasoning abilities of VLMs and are categorized into mask-based, point-based, and element-based methods. Mask-based approaches like SoM [63] apply numbered segmentation masks on images without considering instructions, whereas instruction-guided methods like IVM [73] and API [65] generate masks to block irrelevant regions. Point-based prompting encodes functionalities via points, including affordances [29, 35, 67], motion [35, 41], and constraints like ReKep [21]. However, ReKep [21] extracts semantic keypoints using DINOv2 [42] and struggles to capture precise keypoints that accurately represent desired constraints. Element-based methods like CoPa [19], represent robotic control using basic elements (e.g., points, lines) via pre-trained models like SAM [27] and GPT-4V [1] simply. In contrast, this work explores constraint satisfaction monitoring and introduces constraint elements that more precisely represent relevant entities/parts. These elements are tracked and evaluated in real-time to simplify monitoring." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": "Visual Programming. Visual programming requires strong visual concept understanding and reasoning. Prior works show strong generalization (e.g., zero-shot) across various tasks, such as image editing [8, 16], 3D visual grounding [68], and robotics control [34, 60], by integrating LLMs with visual modules but lose fine-grained details. Recent" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 357, + 555, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 357, + 555, + 430 + ], + "spans": [ + { + "bbox": [ + 313, + 357, + 555, + 430 + ], + "type": "text", + "content": "studies [40, 61] explore visual programming using VLMs, but these methods take raw RGB images as inputs with predefined primitive libraries. In contrast, this work leverages constraint elements for visual programming and uses arithmetic operations in the code to encode their spatio-temporal combinational dynamics, presenting greater challenges." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 441, + 370, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 441, + 370, + 453 + ], + "spans": [ + { + "bbox": [ + 314, + 441, + 370, + 453 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 463, + 554, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 463, + 554, + 522 + ], + "spans": [ + { + "bbox": [ + 313, + 463, + 554, + 522 + ], + "type": "text", + "content": "We first give an overview of the proposed Code-as-Monitor (CaM) (Sec. 3.1). Then, we elaborate on the constraint element in Constraint Painter, especially constraint-aware segmentation (Sec. 3.2). Finally, we present Constraint Monitor for real-time detection (Sec. 3.3)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 532, + 382, + 544 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 532, + 382, + 544 + ], + "spans": [ + { + "bbox": [ + 313, + 532, + 382, + 544 + ], + "type": "text", + "content": "3.1. Overview" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 552, + 555, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 552, + 555, + 694 + ], + "spans": [ + { + "bbox": [ + 313, + 552, + 555, + 694 + ], + "type": "text", + "content": "The proposed CaM comprises three key modules: the Constraint Generator, Painter, and Monitor. We focus on long-horizon manipulation task instructions " + }, + { + "bbox": [ + 313, + 552, + 555, + 694 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{global}}" + }, + { + "bbox": [ + 313, + 552, + 555, + 694 + ], + "type": "text", + "content": " (e.g., \"Move the pan with the bread to the stove, and be careful not to let the bread fall out\"), using RGB-D observations " + }, + { + "bbox": [ + 313, + 552, + 555, + 694 + ], + "type": "inline_equation", + "content": "\\mathcal{O}" + }, + { + "bbox": [ + 313, + 552, + 555, + 694 + ], + "type": "text", + "content": " from two camera views (front and top). As shown in Fig. 2, the RGB images " + }, + { + "bbox": [ + 313, + 552, + 555, + 694 + ], + "type": "inline_equation", + "content": "\\mathcal{O}" + }, + { + "bbox": [ + 313, + 552, + 555, + 694 + ], + "type": "text", + "content": ", along with instructions " + }, + { + "bbox": [ + 313, + 552, + 555, + 694 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{global}}" + }, + { + "bbox": [ + 313, + 552, + 555, + 694 + ], + "type": "text", + "content": ", previous subgoal " + }, + { + "bbox": [ + 313, + 552, + 555, + 694 + ], + "type": "inline_equation", + "content": "l_{\\mathrm{pre}}" + }, + { + "bbox": [ + 313, + 552, + 555, + 694 + ], + "type": "text", + "content": ", and failure feedback from the Constraint Monitor " + }, + { + "bbox": [ + 313, + 552, + 555, + 694 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{pre}}" + }, + { + "bbox": [ + 313, + 552, + 555, + 694 + ], + "type": "text", + "content": " (e.g., subgoal success or failure reason), are fed into the Constraint Generator " + }, + { + "bbox": [ + 313, + 552, + 555, + 694 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{VLM}}" + }, + { + "bbox": [ + 313, + 552, + 555, + 694 + ], + "type": "text", + "content": " (i.e., GPT-4o [1]) to generate the next subgoal " + }, + { + "bbox": [ + 313, + 552, + 555, + 694 + ], + "type": "inline_equation", + "content": "l_{\\mathrm{next}}" + }, + { + "bbox": [ + 313, + 552, + 555, + 694 + ], + "type": "text", + "content": " and associated textual constraints " + }, + { + "bbox": [ + 313, + 552, + 555, + 694 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 313, + 552, + 555, + 694 + ], + "type": "text", + "content": ". This process can be formulated as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 351, + 702, + 555, + 714 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 702, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 351, + 702, + 555, + 714 + ], + "type": "interline_equation", + "content": "l _ {\\text {n e x t}}, \\mathcal {C} _ {d}, \\mathcal {C} _ {u} = \\mathcal {F} _ {\\mathrm {V L M}} \\left(\\mathcal {O}, \\mathcal {L} _ {\\text {g l o b a l}}, l _ {\\text {p r e}}, f _ {\\text {p r e}}\\right) \\tag {1}", + "image_path": "944addc9bce70affce1f7f93e9f432b9800b0a63af71be7bac708e376cb222fa.jpg" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "6921" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 71, + 553, + 178 + ], + "blocks": [ + { + "bbox": [ + 59, + 71, + 553, + 178 + ], + "lines": [ + { + "bbox": [ + 59, + 71, + 553, + 178 + ], + "spans": [ + { + "bbox": [ + 59, + 71, + 553, + 178 + ], + "type": "image", + "image_path": "886a1ff14757f6786cfc75f4e25d63c20b4fca4f0f7a5e4cf0e114601e8474ef.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 181, + 555, + 215 + ], + "lines": [ + { + "bbox": [ + 55, + 181, + 555, + 215 + ], + "spans": [ + { + "bbox": [ + 55, + 181, + 555, + 215 + ], + "type": "text", + "content": "Figure 3. Constraint Element Pipeline. Given a constraint, our model ConSeg generates instance-level and part-level masks across multiple views, which are projected into 3D space. Through a series of heuristics, the desired elements are produced. Once all elements are obtained, they are annotated onto the original multi-view images. Here we display the annotation result of one element." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 224, + 296, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 224, + 296, + 331 + ], + "spans": [ + { + "bbox": [ + 55, + 224, + 296, + 331 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 224, + 296, + 331 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_d = \\{c_d^0, c_d^1, \\ldots, c_d^n\\}" + }, + { + "bbox": [ + 55, + 224, + 296, + 331 + ], + "type": "text", + "content": " denotes the constraints that must be maintained during subgoal execution (e.g., pan handle must be grasped, bread must remain in the pan, pan should remain horizontal during transfer), and " + }, + { + "bbox": [ + 55, + 224, + 296, + 331 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_u = \\{c_u^0, c_u^1, \\ldots, c_u^k\\}" + }, + { + "bbox": [ + 55, + 224, + 296, + 331 + ], + "type": "text", + "content": " refers to the constraints that must be met upon subgoal completion (e.g., pan should be directly above the stove). We successfully unify reactive and proactive failure detection as these task-specific, situation-aware constraint satisfaction problems." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 333, + 296, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 333, + 296, + 464 + ], + "spans": [ + { + "bbox": [ + 54, + 333, + 296, + 464 + ], + "type": "text", + "content": "In Painter, for each textual constraint " + }, + { + "bbox": [ + 54, + 333, + 296, + 464 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 54, + 333, + 296, + 464 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 54, + 333, + 296, + 464 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_d" + }, + { + "bbox": [ + 54, + 333, + 296, + 464 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 54, + 333, + 296, + 464 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_u" + }, + { + "bbox": [ + 54, + 333, + 296, + 464 + ], + "type": "text", + "content": ", we generate corresponding constraint elements " + }, + { + "bbox": [ + 54, + 333, + 296, + 464 + ], + "type": "inline_equation", + "content": "e" + }, + { + "bbox": [ + 54, + 333, + 296, + 464 + ], + "type": "text", + "content": " (detailed in Sec. 3.2) from observations " + }, + { + "bbox": [ + 54, + 333, + 296, + 464 + ], + "type": "inline_equation", + "content": "\\mathcal{O}" + }, + { + "bbox": [ + 54, + 333, + 296, + 464 + ], + "type": "text", + "content": ". These elements, which are composed of 3D points, abstract the constraint-related entities or their parts to represent the desired textual constraint more easily (e.g., the constant distance between green points on bread and pan determines if bread remains in pan, as shown in Fig. 2.) These generated elements are then aggregated into the final set " + }, + { + "bbox": [ + 54, + 333, + 296, + 464 + ], + "type": "inline_equation", + "content": "\\mathcal{E} = \\{e^0, e^1, \\ldots, e^{n + k}\\}" + }, + { + "bbox": [ + 54, + 333, + 296, + 464 + ], + "type": "text", + "content": ", and numerically annotated across all views to produce the final visual prompted images " + }, + { + "bbox": [ + 54, + 333, + 296, + 464 + ], + "type": "inline_equation", + "content": "\\mathcal{O}_{\\mathcal{E}}" + }, + { + "bbox": [ + 54, + 333, + 296, + 464 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 466, + 296, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 466, + 296, + 622 + ], + "spans": [ + { + "bbox": [ + 55, + 466, + 296, + 622 + ], + "type": "text", + "content": "In Monitor, we provide GPT-4o [1] with the next subgoal " + }, + { + "bbox": [ + 55, + 466, + 296, + 622 + ], + "type": "inline_equation", + "content": "l_{\\mathrm{next}}" + }, + { + "bbox": [ + 55, + 466, + 296, + 622 + ], + "type": "text", + "content": ", textual constraints " + }, + { + "bbox": [ + 55, + 466, + 296, + 622 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 55, + 466, + 296, + 622 + ], + "type": "text", + "content": ", and annotated observations " + }, + { + "bbox": [ + 55, + 466, + 296, + 622 + ], + "type": "inline_equation", + "content": "\\mathcal{O}_{\\mathcal{E}}" + }, + { + "bbox": [ + 55, + 466, + 296, + 622 + ], + "type": "text", + "content": " for constraint-aware visual programming to generate the evaluation protocol, i.e., monitor code. This code inputs the elements' 3D positions, calculates arithmetic operations within it, and returns a boolean to indicate potential or actual failure and a string to describe its reason. During subgoal execution, Monitor tracks the elements and evaluates the spatio-temporal combinational dynamics of these elements. If the code returns False, the policy execution halts immediately, and the accompanying string is used as feedback " + }, + { + "bbox": [ + 55, + 466, + 296, + 622 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{pre}}" + }, + { + "bbox": [ + 55, + 466, + 296, + 622 + ], + "type": "text", + "content": " for re-planning. Otherwise, the subgoal is considered completed. In either case, the cycle is repeated." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 634, + 170, + 645 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 634, + 170, + 645 + ], + "spans": [ + { + "bbox": [ + 55, + 634, + 170, + 645 + ], + "type": "text", + "content": "3.2. Constraint Element" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "content": "To simplify the monitoring of constraint satisfaction, we introduce constraint elements by abstracting constraint-related entities/parts into compact geometric elements (e.g., points, lines) to get rid of the most irrelevant geometric and visual details, making them easier to track and generalize." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 224, + 555, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 224, + 555, + 608 + ], + "spans": [ + { + "bbox": [ + 313, + 224, + 555, + 608 + ], + "type": "text", + "content": "Pipeline. The constraint element generation pipeline is shown in Fig. 3. Our trained multi-granularity constraint-aware model ConSeg performs two inference steps for each " + }, + { + "bbox": [ + 313, + 224, + 555, + 608 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 224, + 555, + 608 + ], + "type": "text", + "content": " and each RGB image " + }, + { + "bbox": [ + 313, + 224, + 555, + 608 + ], + "type": "inline_equation", + "content": "o" + }, + { + "bbox": [ + 313, + 224, + 555, + 608 + ], + "type": "text", + "content": " from the set of views " + }, + { + "bbox": [ + 313, + 224, + 555, + 608 + ], + "type": "inline_equation", + "content": "O" + }, + { + "bbox": [ + 313, + 224, + 555, + 608 + ], + "type": "text", + "content": ". First, instance-level masks " + }, + { + "bbox": [ + 313, + 224, + 555, + 608 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_i" + }, + { + "bbox": [ + 313, + 224, + 555, + 608 + ], + "type": "text", + "content": " are generated to capture constraint-related entities. Then, fine-grained part-level masks " + }, + { + "bbox": [ + 313, + 224, + 555, + 608 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_p" + }, + { + "bbox": [ + 313, + 224, + 555, + 608 + ], + "type": "text", + "content": " and corresponding element type descriptions " + }, + { + "bbox": [ + 313, + 224, + 555, + 608 + ], + "type": "inline_equation", + "content": "l_{\\mathrm{e}}" + }, + { + "bbox": [ + 313, + 224, + 555, + 608 + ], + "type": "text", + "content": " are produced, as shown in Fig. 4. Using corresponding depth data, we project " + }, + { + "bbox": [ + 313, + 224, + 555, + 608 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_p" + }, + { + "bbox": [ + 313, + 224, + 555, + 608 + ], + "type": "text", + "content": " from all views into 3D space, fusing them into a point cloud. However, directly tracking and evaluating the spatio-temporal combinational dynamics of these constraint-related entities/parts is challenging. Therefore, we convert these entities into proposed constraint elements. We first apply voxelization to the point cloud with voxel size determined by element type " + }, + { + "bbox": [ + 313, + 224, + 555, + 608 + ], + "type": "inline_equation", + "content": "l_{\\mathrm{e}}" + }, + { + "bbox": [ + 313, + 224, + 555, + 608 + ], + "type": "text", + "content": " (e.g., the surface needs at least 3 points, we divide the occupied space into " + }, + { + "bbox": [ + 313, + 224, + 555, + 608 + ], + "type": "inline_equation", + "content": "2\\times 2" + }, + { + "bbox": [ + 313, + 224, + 555, + 608 + ], + "type": "text", + "content": " voxels.). Then we cluster one representative point within each voxel and filter them to a specified number, also determined by " + }, + { + "bbox": [ + 313, + 224, + 555, + 608 + ], + "type": "inline_equation", + "content": "l_{\\mathrm{e}}" + }, + { + "bbox": [ + 313, + 224, + 555, + 608 + ], + "type": "text", + "content": " to extract the final desired 3D points. We connect points within each instance-level mask " + }, + { + "bbox": [ + 313, + 224, + 555, + 608 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_i" + }, + { + "bbox": [ + 313, + 224, + 555, + 608 + ], + "type": "text", + "content": " to form the constraint element " + }, + { + "bbox": [ + 313, + 224, + 555, + 608 + ], + "type": "inline_equation", + "content": "e" + }, + { + "bbox": [ + 313, + 224, + 555, + 608 + ], + "type": "text", + "content": " associated with constraint " + }, + { + "bbox": [ + 313, + 224, + 555, + 608 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 224, + 555, + 608 + ], + "type": "text", + "content": ". Notably, points modeled as end-effectors (e.g., the points of the fingertips and hand's center represent a dexterous hand) can be directly obtained from forward kinematics, bypassing the process above. Additionally, we perform parallel inference of ConSeg across all views to expedite the acquisition of the final set " + }, + { + "bbox": [ + 313, + 224, + 555, + 608 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 313, + 224, + 555, + 608 + ], + "type": "text", + "content": " and their annotations onto the corresponding views " + }, + { + "bbox": [ + 313, + 224, + 555, + 608 + ], + "type": "inline_equation", + "content": "\\mathcal{O}_{\\mathcal{E}}" + }, + { + "bbox": [ + 313, + 224, + 555, + 608 + ], + "type": "text", + "content": ". Moreover, this minimalist approach, i.e., constraint elements, emphasizes the most relevant entities/parts, enabling generalization to unseen scenes and entities, which is critical for open-set failure detection. For more details, please check Supp. B.3." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 618, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 618, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 618, + 556, + 713 + ], + "type": "text", + "content": "Constraint-aware Segmentation. Since the textual constraint does not explicitly specify relevant entities/parts, we require a model capable of logical reasoning and certain open-set adaptation, enabling precise constraint-aware instance and part-level segmentation. As depicted in Fig. 4, we propose ConSeg, which builds upon LISA [28] comprising a VLM " + }, + { + "bbox": [ + 313, + 618, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{VLM}}" + }, + { + "bbox": [ + 313, + 618, + 556, + 713 + ], + "type": "text", + "content": " (i.e., LLaVA [36]) with the vision encoder " + }, + { + "bbox": [ + 313, + 618, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{enc}}" + }, + { + "bbox": [ + 313, + 618, + 556, + 713 + ], + "type": "text", + "content": " and decoder " + }, + { + "bbox": [ + 313, + 618, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{dec}}" + }, + { + "bbox": [ + 313, + 618, + 556, + 713 + ], + "type": "text", + "content": " from SAM [27]. The textual" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6922" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 69, + 294, + 185 + ], + "blocks": [ + { + "bbox": [ + 57, + 69, + 294, + 185 + ], + "lines": [ + { + "bbox": [ + 57, + 69, + 294, + 185 + ], + "spans": [ + { + "bbox": [ + 57, + 69, + 294, + 185 + ], + "type": "image", + "image_path": "ca68e9310cb053ab41342b51c7c6ab82de32f443ec5dbcc4378e2116ec9e5ba0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 58, + 186, + 185, + 196 + ], + "lines": [ + { + "bbox": [ + 58, + 186, + 185, + 196 + ], + "spans": [ + { + "bbox": [ + 58, + 186, + 185, + 196 + ], + "type": "text", + "content": "Response " + }, + { + "bbox": [ + 58, + 186, + 185, + 196 + ], + "type": "inline_equation", + "content": "(l_{\\mathrm{e}})" + }, + { + "bbox": [ + 58, + 186, + 185, + 196 + ], + "type": "text", + "content": " : It needs a surface " + }, + { + "bbox": [ + 58, + 186, + 185, + 196 + ], + "type": "inline_equation", + "content": "<\\mathrm{SEG}>." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 201, + 295, + 222 + ], + "lines": [ + { + "bbox": [ + 55, + 201, + 295, + 222 + ], + "spans": [ + { + "bbox": [ + 55, + 201, + 295, + 222 + ], + "type": "text", + "content": "Figure 4. ConSeg architecture. Here we display the part-level segmentation, which will output the desired element type and mask." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 229, + 296, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 229, + 296, + 337 + ], + "spans": [ + { + "bbox": [ + 55, + 229, + 296, + 337 + ], + "type": "text", + "content": "constraint " + }, + { + "bbox": [ + 55, + 229, + 296, + 337 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 229, + 296, + 337 + ], + "type": "text", + "content": " and the image " + }, + { + "bbox": [ + 55, + 229, + 296, + 337 + ], + "type": "inline_equation", + "content": "o" + }, + { + "bbox": [ + 55, + 229, + 296, + 337 + ], + "type": "text", + "content": " are input into the VLM " + }, + { + "bbox": [ + 55, + 229, + 296, + 337 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{VLM}}" + }, + { + "bbox": [ + 55, + 229, + 296, + 337 + ], + "type": "text", + "content": " to generate an embedding for the token. The VLM outputs a textual description " + }, + { + "bbox": [ + 55, + 229, + 296, + 337 + ], + "type": "inline_equation", + "content": "l_{\\mathrm{e}}" + }, + { + "bbox": [ + 55, + 229, + 296, + 337 + ], + "type": "text", + "content": " of the desired element type additionally for part-level segmentation. The decoder " + }, + { + "bbox": [ + 55, + 229, + 296, + 337 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{dec}}" + }, + { + "bbox": [ + 55, + 229, + 296, + 337 + ], + "type": "text", + "content": " then generates the segmentation mask " + }, + { + "bbox": [ + 55, + 229, + 296, + 337 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 55, + 229, + 296, + 337 + ], + "type": "text", + "content": " by leveraging visual features from the vision encoder " + }, + { + "bbox": [ + 55, + 229, + 296, + 337 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{enc}}" + }, + { + "bbox": [ + 55, + 229, + 296, + 337 + ], + "type": "text", + "content": " based on the current observation " + }, + { + "bbox": [ + 55, + 229, + 296, + 337 + ], + "type": "inline_equation", + "content": "o" + }, + { + "bbox": [ + 55, + 229, + 296, + 337 + ], + "type": "text", + "content": ", and features derived from the final-layer embedding " + }, + { + "bbox": [ + 55, + 229, + 296, + 337 + ], + "type": "inline_equation", + "content": "h_{<\\mathrm{SEG}>}" + }, + { + "bbox": [ + 55, + 229, + 296, + 337 + ], + "type": "text", + "content": " of the token, transformed via an MLP (" + }, + { + "bbox": [ + 55, + 229, + 296, + 337 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{MLP}}" + }, + { + "bbox": [ + 55, + 229, + 296, + 337 + ], + "type": "text", + "content": "). The process can be formulated as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 59, + 345, + 295, + 357 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 345, + 295, + 357 + ], + "spans": [ + { + "bbox": [ + 59, + 345, + 295, + 357 + ], + "type": "interline_equation", + "content": "l _ {\\mathrm {e}}, h _ {< \\mathrm {S E G} >} = \\mathcal {F} _ {\\mathrm {V L M}} (c, o), \\mathcal {M} = \\mathcal {F} _ {\\mathrm {d e c}} (\\mathcal {F} _ {\\mathrm {M L P}} \\left(h _ {< \\mathrm {S E G} >}\\right), \\mathcal {F} _ {\\mathrm {e n c}} (o)) \\tag {2}", + "image_path": "7757d6740e197a6a9d0c183652486fd7317fc7112ec39de595a4066edc259442.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 361, + 295, + 398 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 361, + 295, + 398 + ], + "spans": [ + { + "bbox": [ + 55, + 361, + 295, + 398 + ], + "type": "text", + "content": "The entire pipeline, including additional LoRA [17] parameters, is optimized end-to-end while keeping the parameters of the VLM " + }, + { + "bbox": [ + 55, + 361, + 295, + 398 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{VLM}}" + }, + { + "bbox": [ + 55, + 361, + 295, + 398 + ], + "type": "text", + "content": " and vision encoder " + }, + { + "bbox": [ + 55, + 361, + 295, + 398 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{enc}}" + }, + { + "bbox": [ + 55, + 361, + 295, + 398 + ], + "type": "text", + "content": " frozen." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 400, + 296, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 400, + 296, + 581 + ], + "spans": [ + { + "bbox": [ + 55, + 400, + 296, + 581 + ], + "type": "text", + "content": "Dataset and Collection Pipeline. Our dataset utilizes images from BridgeData V2 [62]. While BridgeData V2 provides trajectory-level instructions, we require frame-level annotations for per-frame subgoals and constraints. To address this, we sample pick-and-place data, using external references (e.g., gripper open/close states) to generate 10,181 trajectories with 219,356 images. We employ GPT-4o to decompose trajectory instructions into subgoals, constraints, and object-part associations to generate groundtruth annotations. Instance-level and part-level segmentations are performed using Grounded SAM [49] and Semantic SAM [32], respectively. These outputs are integrated and refined through manual inspection to produce the multi-granularity dataset, which is combined with LISA's training data to fine-tune our model. More details are in Supp. B." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 587, + 204, + 599 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 587, + 204, + 599 + ], + "spans": [ + { + "bbox": [ + 55, + 587, + 204, + 599 + ], + "type": "text", + "content": "3.3. Real-time Monitor Module" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "content": "After annotating constraint elements with unique colors and numerical labels across all views, these images serve as visual prompts to generate monitor code via GPT-4o [1] with constraints " + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "content": " and the next subgoal " + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "inline_equation", + "content": "l_{\\mathrm{next}}" + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "content": ". The generated code accepts the elements' 3D positions as input and performs arithmetic operations (e.g., NumPy-supported calculations) to evaluate spatio-temporal constraints at various stages (i.e., during execution or upon completion), as shown in Fig. 2 code block. It returns a boolean flag indicating" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": "potential or actual failure if the computed result exceeds a predefined tolerance (e.g., accepting deviations of the pan's surface normal from the z-axis within " + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "inline_equation", + "content": "15^{\\circ}" + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": "), along with a string explaining the cause. By tracking the elements using CoTracker [25], we achieve real-time detection via VLMs without frequent calls, reducing the computational cost." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 144, + 555, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 144, + 555, + 289 + ], + "spans": [ + { + "bbox": [ + 313, + 144, + 555, + 289 + ], + "type": "text", + "content": "Notably, including both current positions and historical trajectories of the elements as input allows us to monitor tasks needing high precision, such as the requirement of " + }, + { + "bbox": [ + 313, + 144, + 555, + 289 + ], + "type": "inline_equation", + "content": "2\\mathrm{cm}" + }, + { + "bbox": [ + 313, + 144, + 555, + 289 + ], + "type": "text", + "content": " movement or " + }, + { + "bbox": [ + 313, + 144, + 555, + 289 + ], + "type": "inline_equation", + "content": "180^{\\circ}" + }, + { + "bbox": [ + 313, + 144, + 555, + 289 + ], + "type": "text", + "content": " rotation, which previous reactive detection methods could not address through direct VLM queries using raw RGB images. Moreover, the potential diversity in the structured associations of constraint elements within the code (e.g., calculating the angle between the normal vector of the pan's surface element and the z-axis determines if the pan is level, as shown in Fig. 2 code block) and rich prior knowledge offered by VLMs enable generalization to certain unseen tasks requiring common skills." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 300, + 395, + 313 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 300, + 395, + 313 + ], + "spans": [ + { + "bbox": [ + 313, + 300, + 395, + 313 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 320, + 555, + 404 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 320, + 555, + 404 + ], + "spans": [ + { + "bbox": [ + 313, + 320, + 555, + 404 + ], + "type": "text", + "content": "Our experiments aim to address the following questions: (1) Can CaM achieve open-set reactive and proactive failure detection across diverse robots, end-effectors, and objects, both in simulator (Sec. 4.2) and on real robots (Sec. 4.3)? (2) Can ConSeg infer multi-granularity constraint-aware masks for unseen scenes and objects (Sec. 4.4)? (3) Which design choices greatly enhance the performance (Sec. 4.5)?" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 413, + 430, + 426 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 413, + 430, + 426 + ], + "spans": [ + { + "bbox": [ + 313, + 413, + 430, + 426 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 431, + 554, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 431, + 554, + 576 + ], + "spans": [ + { + "bbox": [ + 313, + 431, + 554, + 576 + ], + "type": "text", + "content": "Environment Settings. We evaluate CaM across three simulators (i.e., CLIPort [54], Omnigibson [31], and RL-Bench [22]) and a real-world setting. In CLIPort, we employ a UR5 arm equipped with a suction cup for pick-and-place and a spatula for pushing, controlled by pre-trained CLIPort policy [54]. Omnigibson features a Fetch robot with a gripper, controlled by ReKep [21]. RLBench uses a Franka arm with a gripper, controlled by ARP [70]. We collect a small dataset to fine-tune ConSeg for each simulator. We use a UR5 arm with a Leap Hand [51] for real-world experiments, controlled by an open-loop policy named DexGraspNet 2.0 [69]. More details are provided in Supp. C." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 578, + 554, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 578, + 554, + 674 + ], + "spans": [ + { + "bbox": [ + 313, + 578, + 554, + 674 + ], + "type": "text", + "content": "Task Settings. We introduce disturbances across all environments to induce failures, classified by the affected constraints (e.g., points, lines, surfaces). We evaluate task success rate, execution time, and additional token usage in Omnigibson. Notably, we don't assess failure judgment success rate because our code operates continuously in real-time, rendering this metric inapplicable. Detailed evaluation settings are provided in each section." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "content": "Baselines. We compare DoReMi (DRM) [15] as the baseline for all settings, which uses VLMs to detect intermediate failures during execution via frequent VQA-based queries." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6923" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 97 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 97 + ], + "type": "text", + "content": "In CLIPort, we include Inner Monologue [20] baseline, which detects failures only upon each subgoal completion." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 102, + 202, + 114 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 102, + 202, + 114 + ], + "spans": [ + { + "bbox": [ + 55, + 102, + 202, + 114 + ], + "type": "text", + "content": "4.2. Main Results in Simulator" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 121, + 170, + 132 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 121, + 170, + 132 + ], + "spans": [ + { + "bbox": [ + 55, + 121, + 170, + 132 + ], + "type": "text", + "content": "4.2.1 Results in CLIPort" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 140, + 296, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 140, + 296, + 295 + ], + "spans": [ + { + "bbox": [ + 55, + 140, + 296, + 295 + ], + "type": "text", + "content": "We evaluate two tasks in CLIPort: (1) Stack in Order: The robot must stack blocks in a specified order, including two point-level disturbances: (a) with per-step probability " + }, + { + "bbox": [ + 55, + 140, + 296, + 295 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 55, + 140, + 296, + 295 + ], + "type": "text", + "content": ", the suction cup may release a block, causing it to drop; (b) placement positions are perturbed by uniform noise in " + }, + { + "bbox": [ + 55, + 140, + 296, + 295 + ], + "type": "inline_equation", + "content": "[0, q]" + }, + { + "bbox": [ + 55, + 140, + 296, + 295 + ], + "type": "text", + "content": " cm, potentially leading to tower collapse. Success is defined as correctly stacking the blocks within 70s. (2) Sweep Half the Blocks: To address the pre-trained policy's tendency to sweep all blocks, the robot must determine when to halt, sweeping half of the blocks (" + }, + { + "bbox": [ + 55, + 140, + 296, + 295 + ], + "type": "inline_equation", + "content": "\\pm 10\\%" + }, + { + "bbox": [ + 55, + 140, + 296, + 295 + ], + "type": "text", + "content": ") into a specified colored area within 30s. Success is achieved by meeting these criteria. Tab. 1 shows the mean results over 5 different seeds, each with 12 episodes. Check Supp. D.1 for details." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 298, + 296, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 298, + 296, + 417 + ], + "spans": [ + { + "bbox": [ + 55, + 298, + 296, + 417 + ], + "type": "text", + "content": "Code better monitors 3D space relations. As shown in the Tab. 1, Under the most severe disturbances " + }, + { + "bbox": [ + 55, + 298, + 296, + 417 + ], + "type": "inline_equation", + "content": "(p = 0.3, q = 3)" + }, + { + "bbox": [ + 55, + 298, + 296, + 417 + ], + "type": "text", + "content": " in \"Stack in Order\", CaM achieves a " + }, + { + "bbox": [ + 55, + 298, + 296, + 417 + ], + "type": "inline_equation", + "content": "17.5\\%" + }, + { + "bbox": [ + 55, + 298, + 296, + 417 + ], + "type": "text", + "content": " higher success rate than DoReMi. Frequent VLM queries lead to increased incorrect failure judgments due to a limited 3D spatial understanding from single images, compared to code-based evaluation of block positions. For example, after placing the green block on the red and preparing to pick up the blue, DoReMi mistakenly thinks that the green block is no longer atop the red, causing it to re-execute the previous subgoal." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 420, + 296, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 420, + 296, + 576 + ], + "spans": [ + { + "bbox": [ + 55, + 420, + 296, + 576 + ], + "type": "text", + "content": "Code with elements achieve reactive and proactive failure detection. As shown in Tab. 1, under severe disturbances in \"Stack in Order\", CaM reduces execution time by " + }, + { + "bbox": [ + 55, + 420, + 296, + 576 + ], + "type": "inline_equation", + "content": "38.7\\%" + }, + { + "bbox": [ + 55, + 420, + 296, + 576 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 420, + 296, + 576 + ], + "type": "inline_equation", + "content": "14.4\\%" + }, + { + "bbox": [ + 55, + 420, + 296, + 576 + ], + "type": "text", + "content": " compared to Inner Monologue [20], which only detects failures upon subgoal completion, and DoReMi [15], which suffers from misjudgments due to repeated VLM queries, respectively. In contrast, CaM unifies both reactive and proactive failure detection with high precision in real-time, especially in preventing foreseeable failures. For example, if the green block is placed on the red block with heavy placement position disturbance, CaM anticipates that further stacking the blue may lead to collapse and then stabilizes the green block before proceeding." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 578, + 296, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 578, + 296, + 663 + ], + "spans": [ + { + "bbox": [ + 55, + 578, + 296, + 663 + ], + "type": "text", + "content": "Code with elements leads to more accurate counting. As shown in the Tab. 1, in \"Sweep Half the Blocks\", CaM achieves average success rates " + }, + { + "bbox": [ + 55, + 578, + 296, + 663 + ], + "type": "inline_equation", + "content": "4.5 \\times" + }, + { + "bbox": [ + 55, + 578, + 296, + 663 + ], + "type": "text", + "content": " higher than DoReMi [15]. Calculating the block points in the designated surface region provides more accurate results than directly using the VLM to count, enabling a more precise halting of the policy to complete the task." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 671, + 186, + 683 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 671, + 186, + 683 + ], + "spans": [ + { + "bbox": [ + 55, + 671, + 186, + 683 + ], + "type": "text", + "content": "4.2.2 Results in Omnigibson" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": "We conduct experiments on three tasks in Omnigibson, each involving a distinct type of constraint-element disturbance:" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 318, + 106, + 548, + 184 + ], + "blocks": [ + { + "bbox": [ + 313, + 71, + 553, + 103 + ], + "lines": [ + { + "bbox": [ + 313, + 71, + 553, + 103 + ], + "spans": [ + { + "bbox": [ + 313, + 71, + 553, + 103 + ], + "type": "text", + "content": "Table 1. Performance in CLIPort. We report the success rate and execution time, compared to CLIPort (CP) [54], with Inner Monologue (IM) [20] and DoReMi (DRM) [15]." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 318, + 106, + 548, + 184 + ], + "lines": [ + { + "bbox": [ + 318, + 106, + 548, + 184 + ], + "spans": [ + { + "bbox": [ + 318, + 106, + 548, + 184 + ], + "type": "table", + "html": "
Tasks with disturbanceSuccess Rate(%) ↑Execution Time(s) ↓
CP+IM+DRM+Ours+IM+DRM+Ours
Stack in order with drop pp=0.0100.0100.0100.0100.013.413.413.4
p=0.1556.6781.6783.3395.0034.826.0021.0
p=0.321.6775.0076.6788.3342.834.2025.4
Stack in order with noise qq=190.0090.0096.6798.3324.824.624.2
q=241.6771.6775.0083.3339.437.029.2
q=315.0040.0040.0063.3358.254.236.8
Sweep Half the Blocks0.0018.3316.6775.0022.016.616.4
", + "image_path": "d443fc67a5fc90ec15ed640490a77859089cbb39df46db8c5a51cafc8b8f0f77.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 191, + 555, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 191, + 555, + 384 + ], + "spans": [ + { + "bbox": [ + 313, + 191, + 555, + 384 + ], + "type": "text", + "content": "(1) Slot Pen: Insert a pen into a holder, facing point-level disturbances wherein (a) pen is moved during grasping; (b) pen is dropped during transport; (c) holder is moved during insertion. (2) Stow Book: Place a book to a bookshelf vertically, with line-level disturbances where (a) book is randomly rotated during grasping; (b) end-effector joint is randomly actuated to alter the book pose; (c) book is reoriented horizontally after placement. (3) Pour Tea: Pours from a teapot into a teacup, encountering surface-level disturbances wherein (a) teapot is tilted forward/backward during movement; (b) end-effector joint is actuated to induce a lateral tilt of the teapot during movement; (c) teapot is returned to a horizontal position during pouring. Tab. 2 reports the results across three tasks, each including one no-disturbance trial and three specific-disturbance trials, with 10 runs for each setting. More details are provided in Supp. D.2." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 387, + 556, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 387, + 556, + 495 + ], + "spans": [ + { + "bbox": [ + 313, + 387, + 556, + 495 + ], + "type": "text", + "content": "Code with elements detects richer failures. In Tab. 2, only CaM can detect failures caused by surface-level disturbances in \"Pour Tea\" compared to DoReMi [15]. The reason is that changes in the teapot's pitch and roll angles are hard to detect by querying VLM via VQA using one current image. The misjudgment leads to a " + }, + { + "bbox": [ + 313, + 387, + 556, + 495 + ], + "type": "inline_equation", + "content": "0\\%" + }, + { + "bbox": [ + 313, + 387, + 556, + 495 + ], + "type": "text", + "content": " success rate of DoReMi in this task. Notably, DoReMi's success rate is sometimes lower than that of ReKep alone due to VLM's limited spatio-temporal understanding of single images." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 498, + 556, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 556, + 606 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 556, + 606 + ], + "type": "text", + "content": "Code with elements detects failure with lower computational cost. As shown in Tab. 2, we achieve a " + }, + { + "bbox": [ + 313, + 498, + 556, + 606 + ], + "type": "inline_equation", + "content": "34.8\\%" + }, + { + "bbox": [ + 313, + 498, + 556, + 606 + ], + "type": "text", + "content": " reduction in execution time and a " + }, + { + "bbox": [ + 313, + 498, + 556, + 606 + ], + "type": "inline_equation", + "content": "52.2\\%" + }, + { + "bbox": [ + 313, + 498, + 556, + 606 + ], + "type": "text", + "content": " decrease in token count compared to DoReMi [15]. This improvement stems from proactive failure detection, which prevents more severe failures ahead in real-time for timely re-planning while generating code only once per subgoal. For example, in \"Pour Tea\", CaM detects failure by monitoring the teapot's tilt angle in real-time, avoiding frequent VLM checks." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 617, + 432, + 628 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 432, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 432, + 628 + ], + "type": "text", + "content": "4.2.3 Results in RLBench" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 637, + 553, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 637, + 553, + 662 + ], + "spans": [ + { + "bbox": [ + 313, + 637, + 553, + 662 + ], + "type": "text", + "content": "We further evaluate CaM on RLbench and demonstrate its superior generalization. Details can be found in Supp. D.3." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 670, + 467, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 670, + 467, + 682 + ], + "spans": [ + { + "bbox": [ + 313, + 670, + 467, + 682 + ], + "type": "text", + "content": "4.3. Main Results in Real World" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "content": "We conduct real-world evaluations on two tasks: (1) Simple Pick & Place: The robot has 70s to pick up objects and" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6924" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 84, + 553, + 137 + ], + "blocks": [ + { + "bbox": [ + 55, + 71, + 553, + 82 + ], + "lines": [ + { + "bbox": [ + 55, + 71, + 553, + 82 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 553, + 82 + ], + "type": "text", + "content": "Table 2. Performance in Omnigibson. We report the success rate (SR), execution time, and token usage, compared to DoReMi (DRM) [15]." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 84, + 553, + 137 + ], + "lines": [ + { + "bbox": [ + 56, + 84, + 553, + 137 + ], + "spans": [ + { + "bbox": [ + 56, + 84, + 553, + 137 + ], + "type": "table", + "html": "
MethodSlot Pen (Point-level Disturbances)Stow Book (Line-level Disturbances)Pour Tea (Surface-level Disturbances)
SR with Disturbance(%) ↑Time (s)↓Token (k)↓SR with Disturbance(%) ↑Time (s)↓Token (k)↓SR with Disturbance(%) ↑Time (s)↓Token (k)↓
NoneDist.(a)Dist.(b)Dist.(c)NoneDist.(a)Dist.(b)Dist.(c)NoneDist.(a)Dist.(b)Dist.(c)
ReKep [21]30201010--40303020--20202010--
+DRM40102020177.8454.5450402040127.1738.670000--
+Ours60504040101.8525.827060706093.0818.6750404030174.5544.19
", + "image_path": "b65cecf2cf416741b69ef18bf2d4a36ce958dced030908d853646d2f46173150.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 59, + 156, + 120, + 220 + ], + "blocks": [ + { + "bbox": [ + 128, + 144, + 481, + 153 + ], + "lines": [ + { + "bbox": [ + 128, + 144, + 481, + 153 + ], + "spans": [ + { + "bbox": [ + 128, + 144, + 481, + 153 + ], + "type": "text", + "content": "Reasoning Pick &Place Task: Grasp the animals according to their distances to the fruits, from nearest to farthest." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 59, + 156, + 120, + 220 + ], + "lines": [ + { + "bbox": [ + 59, + 156, + 120, + 220 + ], + "spans": [ + { + "bbox": [ + 59, + 156, + 120, + 220 + ], + "type": "image", + "image_path": "660a001cf7222d7caab60e84243d3fb50d8e79091238225a2bfc7a625d82ef67.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 121, + 156, + 182, + 220 + ], + "blocks": [ + { + "bbox": [ + 121, + 156, + 182, + 220 + ], + "lines": [ + { + "bbox": [ + 121, + 156, + 182, + 220 + ], + "spans": [ + { + "bbox": [ + 121, + 156, + 182, + 220 + ], + "type": "image", + "image_path": "30b32e9acf971d7764d35087851fed9ed881248853e32b716bcbcde1517ab5b1.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 171, + 221, + 253, + 229 + ], + "lines": [ + { + "bbox": [ + 171, + 221, + 253, + 229 + ], + "spans": [ + { + "bbox": [ + 171, + 221, + 253, + 229 + ], + "type": "text", + "content": "Horse is moved, target changed" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 182, + 156, + 244, + 220 + ], + "blocks": [ + { + "bbox": [ + 182, + 156, + 244, + 220 + ], + "lines": [ + { + "bbox": [ + 182, + 156, + 244, + 220 + ], + "spans": [ + { + "bbox": [ + 182, + 156, + 244, + 220 + ], + "type": "image", + "image_path": "d36186f6b5c1dbf886a26c79c2abd776300562c00a411ea8e9fd516ebaf0568a.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 231, + 553, + 253 + ], + "lines": [ + { + "bbox": [ + 55, + 231, + 553, + 253 + ], + "spans": [ + { + "bbox": [ + 55, + 231, + 553, + 253 + ], + "type": "text", + "content": "Figure 5. Example of Real-world Evaluation. The red bounding box shows the current grasp target, which may shift due to environmental changes. CaM monitors and adapts to these changes in real-time, resulting in a closed-loop system with an open-loop policy." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 244, + 156, + 305, + 220 + ], + "blocks": [ + { + "bbox": [ + 244, + 156, + 305, + 220 + ], + "lines": [ + { + "bbox": [ + 244, + 156, + 305, + 220 + ], + "spans": [ + { + "bbox": [ + 244, + 156, + 305, + 220 + ], + "type": "image", + "image_path": "a4097cc7255874f1477e174cb053a7d4c83bdb3dfcd789a82b0557e19377088b.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 305, + 156, + 366, + 220 + ], + "blocks": [ + { + "bbox": [ + 305, + 156, + 366, + 220 + ], + "lines": [ + { + "bbox": [ + 305, + 156, + 366, + 220 + ], + "spans": [ + { + "bbox": [ + 305, + 156, + 366, + 220 + ], + "type": "image", + "image_path": "d1fb0406b88c25c6813c4d294041583f54667d02849a04c591f1a43ba9a73e5f.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 359, + 222, + 439, + 229 + ], + "lines": [ + { + "bbox": [ + 359, + 222, + 439, + 229 + ], + "spans": [ + { + "bbox": [ + 359, + 222, + 439, + 229 + ], + "type": "text", + "content": "Pear is moved, target changed" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 367, + 156, + 428, + 220 + ], + "blocks": [ + { + "bbox": [ + 367, + 156, + 428, + 220 + ], + "lines": [ + { + "bbox": [ + 367, + 156, + 428, + 220 + ], + "spans": [ + { + "bbox": [ + 367, + 156, + 428, + 220 + ], + "type": "image", + "image_path": "fa256f96746d862e29ef9c4138ce178c58e1a22d6c4e45d2c0950b354aa0defd.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 428, + 156, + 490, + 220 + ], + "blocks": [ + { + "bbox": [ + 428, + 156, + 490, + 220 + ], + "lines": [ + { + "bbox": [ + 428, + 156, + 490, + 220 + ], + "spans": [ + { + "bbox": [ + 428, + 156, + 490, + 220 + ], + "type": "image", + "image_path": "c76d3c8fea3c5410202a403736f70d6009c391076cdb5eb7b23c8b4bd590b806.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 490, + 156, + 552, + 220 + ], + "blocks": [ + { + "bbox": [ + 490, + 156, + 552, + 220 + ], + "lines": [ + { + "bbox": [ + 490, + 156, + 552, + 220 + ], + "spans": [ + { + "bbox": [ + 490, + 156, + 552, + 220 + ], + "type": "image", + "image_path": "928f0c7cf2182f87c7deebdbee23ac940cb16c7e1dd3d312dadae4ec1986a1b0.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 56, + 282, + 293, + 369 + ], + "blocks": [ + { + "bbox": [ + 55, + 258, + 294, + 280 + ], + "lines": [ + { + "bbox": [ + 55, + 258, + 294, + 280 + ], + "spans": [ + { + "bbox": [ + 55, + 258, + 294, + 280 + ], + "type": "text", + "content": "Table 3. Performance of Single Pick & Place. We report the success rate and execution time. DGN donates DexGraspNet 2.0 [69]." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 282, + 293, + 369 + ], + "lines": [ + { + "bbox": [ + 56, + 282, + 293, + 369 + ], + "spans": [ + { + "bbox": [ + 56, + 282, + 293, + 369 + ], + "type": "table", + "html": "
Tasks with disturbanceObject typesSuccess Rate(%) ↑Execution Time(s) ↓
DGN+DRM+Ours+DRM+Ours
Pick & PlaceDeformable0.0083.3396.6761.846.3
with theTransparent0.0066.6793.3372.648.1
objects beingSmall Rigid0.0080.096.6765.745.4
movedLarge Geometric0.0086.6796.6768.945.3
Pick & PlaceDeformable0.0076.6793.3368.762.5
with theTransparent0.0060.0090.0077.762.7
objects beingSmall Rigid0.0063.3393.3369.860.5
removedLarge Geometric0.0076.6796.6772.360.3
", + "image_path": "7669d1805c87750ba7e38ffd591dfb1ebc2ec5fd04eecfb9b35bcaf1b2957fc0.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_body" + } + ], + "index": 15 + }, + { + "type": "table", + "bbox": [ + 59, + 432, + 288, + 487 + ], + "blocks": [ + { + "bbox": [ + 55, + 375, + 295, + 430 + ], + "lines": [ + { + "bbox": [ + 55, + 375, + 295, + 430 + ], + "spans": [ + { + "bbox": [ + 55, + 375, + 295, + 430 + ], + "type": "text", + "content": "Table 4. Performance of Reasoning Pick & Place in cluttered scene. We report the success rate. The robot is controlled by an open-loop policy named DexGraspNet 2.0 (DGN) [69]. w/ CE with " + }, + { + "bbox": [ + 55, + 375, + 295, + 430 + ], + "type": "inline_equation", + "content": "\\checkmark" + }, + { + "bbox": [ + 55, + 375, + 295, + 430 + ], + "type": "text", + "content": " indicates using constraint elements; otherwise, constraint-aware entities or parts are used for tracking and code computation." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 59, + 432, + 288, + 487 + ], + "lines": [ + { + "bbox": [ + 59, + 432, + 288, + 487 + ], + "spans": [ + { + "bbox": [ + 59, + 432, + 288, + 487 + ], + "type": "table", + "html": "
Tasksw/ CESuccess Rate(%) ↑
DGN+DRM+Ours
Clear all objects on table except for animals✓0.0010.0060.00
X0.0010.0020.00
Grasp the animals according to their distances to fruits, from nearest to farthest✓0.000.0090.00
", + "image_path": "50b0ccaca0a87798ed5ec621f376ea15ee4670ced9e862f960d88b454e6f4e04.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "table_body" + } + ], + "index": 17 + }, + { + "bbox": [ + 54, + 493, + 295, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 493, + 295, + 601 + ], + "spans": [ + { + "bbox": [ + 54, + 493, + 295, + 601 + ], + "type": "text", + "content": "place them at specified locations, facing two disturbances: (a) moving the object during grasping, and (b) removing the object from hand during movement. We test four object types, selecting 3 examples per type and conducting 10 trials for each (see Tab. 3). (2) Reasoning Pick & Place: The robot executes long-horizon tasks, involving ambiguous terms (e.g., \"fruit\"), under the same disturbances. We evaluate 2 long-horizon tasks in cluttered scenes, performing 10 trials each (see Tab. 4). Check Supp. D.4 for details." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "content": "Elements generally abstract constraints and relevant entities. As shown in Tab. 3, using a different end-effector (i.e., Leap Hand [51]) in Simple Pick & Place, CaM also achieves success rates surpassing DoReMi [15] by " + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "inline_equation", + "content": "20.4\\%" + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "content": " when handling different kinds of objects (e.g., deformable). We find that abstracting constraint-related entities/parts removes the irrelevant visual details, enabling generalization to different kinds of entities in unseen scenes (as discussed in Sec. 3.2), leading to easier tracking and code evaluation." + } + ] + } + ], + "index": 19 + }, + { + "type": "table", + "bbox": [ + 324, + 282, + 541, + 344 + ], + "blocks": [ + { + "bbox": [ + 313, + 258, + 553, + 279 + ], + "lines": [ + { + "bbox": [ + 313, + 258, + 553, + 279 + ], + "spans": [ + { + "bbox": [ + 313, + 258, + 553, + 279 + ], + "type": "text", + "content": "Table 5. Performance of reasoning and constraint-aware segmentation. FMC denotes the foundation model combination baseline." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 324, + 282, + 541, + 344 + ], + "lines": [ + { + "bbox": [ + 324, + 282, + 541, + 344 + ], + "spans": [ + { + "bbox": [ + 324, + 282, + 541, + 344 + ], + "type": "table", + "html": "
MethodReasonSegConstraintSeg
Instance-levelInstance-levelPart-level
gIoUcIoUgIoUcIoUgIoUcIoU
PixelLM [50]56.061.444.443.224.122.6
LISA-13B [28]56.665.142.138.923.424.3
FMC51.253.349.349.640.839.3
ConSeg-13B55.763.962.168.760.265.3
", + "image_path": "8e367f57950329b95acdab592a44a39605e835fdb4990b1d71bf4d6550bb3131.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "table_body" + } + ], + "index": 21 + }, + { + "type": "table", + "bbox": [ + 323, + 386, + 542, + 439 + ], + "blocks": [ + { + "bbox": [ + 313, + 350, + 553, + 383 + ], + "lines": [ + { + "bbox": [ + 313, + 350, + 553, + 383 + ], + "spans": [ + { + "bbox": [ + 313, + 350, + 553, + 383 + ], + "type": "text", + "content": "Table 6. Ablation studies in Omnigibson's \"Stow Book\", assessing the impact of Multi-View (MV), Constraint-aware Segmentation (CS) for elements, and Connect Points (CP) for element formation." + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 323, + 386, + 542, + 439 + ], + "lines": [ + { + "bbox": [ + 323, + 386, + 542, + 439 + ], + "spans": [ + { + "bbox": [ + 323, + 386, + 542, + 439 + ], + "type": "table", + "html": "
MVCSCPSuccess Rate with Disturbance(%) ↑
NoneDist.(a)Dist.(b)Dist.(c)Avg
X✓✓40.040.030.050.040.0
✓X✓50.040.040.040.042.5
✓✓X60.050.060.050.055.0
✓✓✓70.060.070.060.065.0
", + "image_path": "895bd2b94ff4532079a96a21cfb34999501520c9a74ae9bdf9ddf2e8b25000a9.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "table_body" + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 445, + 555, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 445, + 555, + 589 + ], + "spans": [ + { + "bbox": [ + 313, + 445, + 555, + 589 + ], + "type": "text", + "content": "Reactive and Proactive failure detection combined with an open-loop policy forms a close-loop system. Tab. 4 shows that only " + }, + { + "bbox": [ + 313, + 445, + 555, + 589 + ], + "type": "inline_equation", + "content": "CaM" + }, + { + "bbox": [ + 313, + 445, + 555, + 589 + ], + "type": "text", + "content": " successfully handles long-horizon tasks in cluttered scenes. These tasks are challenging as the robot is controlled by an open-loop policy that cannot handle environment dynamics and human disturbances in a closed-loop manner. By incorporating both reactive and proactive failure detection with the open-loop policy (see Fig. 5), the robot can dynamically adjust its target in real-time. For example, when a human moves the horse or pear during the task, the robot adapts by grasping the animal closest to the fruit, forming a closed-loop system." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 598, + 477, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 598, + 477, + 611 + ], + "spans": [ + { + "bbox": [ + 313, + 598, + 477, + 611 + ], + "type": "text", + "content": "4.4. Main Results of Segmentation" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 618, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 618, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 618, + 555, + 713 + ], + "type": "text", + "content": "Tab. 5 presents segmentation results comparing our ConSeg, with SOTA models and a Foundation Model Combination (FMC) baseline. FMC integrates GPT-4o for reasoning over tasks and constraints to identify relevant instances and parts, along with Grounded SAM [49] and Semantic SAM [32] for instance and part-level segmentation, respectively, similar to our data collection pipeline. We evaluate performance on the ReasonSeg [28] benchmark and our" + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "6925" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 71, + 553, + 247 + ], + "blocks": [ + { + "bbox": [ + 58, + 71, + 553, + 247 + ], + "lines": [ + { + "bbox": [ + 58, + 71, + 553, + 247 + ], + "spans": [ + { + "bbox": [ + 58, + 71, + 553, + 247 + ], + "type": "image", + "image_path": "dafd99b9a9b9a1c2b9729ad2ca841eddfb7e36588fc6fc73c4eb0b05070bf488.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 59, + 252, + 549, + 263 + ], + "lines": [ + { + "bbox": [ + 59, + 252, + 549, + 263 + ], + "spans": [ + { + "bbox": [ + 59, + 252, + 549, + 263 + ], + "type": "text", + "content": "Figure 6. Visual comparison between our ConSeg and LISA [28] at instance and part level. The red masks are the segmentation results." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 58, + 316, + 290, + 369 + ], + "blocks": [ + { + "bbox": [ + 55, + 269, + 295, + 313 + ], + "lines": [ + { + "bbox": [ + 55, + 269, + 295, + 313 + ], + "spans": [ + { + "bbox": [ + 55, + 269, + 295, + 313 + ], + "type": "text", + "content": "Table 7. Ablation study on training data. SemanticSeg includes ADE20K [75], COCO-Stuff [5], PACO-LVIS [48] and PASCALPart [6]. ReferSeg includes refCLEF, refCOCO, refCOCO+ [26] and refCOCOg [39]. VQA indicates LLaVAInstruct-150k [36]." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 316, + 290, + 369 + ], + "lines": [ + { + "bbox": [ + 58, + 316, + 290, + 369 + ], + "spans": [ + { + "bbox": [ + 58, + 316, + 290, + 369 + ], + "type": "table", + "html": "
SemanticSegReferSegVQAReasonSegConstraintSegIns-level gIoUcloUPart-level gIoUcloU
✓✓✓✓✘✘42.138.923.424.3
✓✓✓✓✓✘60.765.940.445.6
✓✓✓✓✘✓51.550.656.561.7
✓✓✓✓✓✓62.168.760.265.3
", + "image_path": "62004fe47f27a58d3ee66f85584ac2bc5b4ec40934198e067fded82a97219a2d.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 376, + 295, + 412 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 376, + 295, + 412 + ], + "spans": [ + { + "bbox": [ + 55, + 376, + 295, + 412 + ], + "type": "text", + "content": "proposed Constraint-Aware Segmentation (ConstraintSeg) benchmark. Our benchmark evaluates performance using gIoU and cIoU, following ReasonSeg's evaluation setting." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 415, + 296, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 415, + 296, + 499 + ], + "spans": [ + { + "bbox": [ + 55, + 415, + 296, + 499 + ], + "type": "text", + "content": "ConSeg performs both reasoning and multi-granularity constraint-aware segmentation. As shown in Tab. 5, ConSeg performs comparably to LISA and PixelLM on the ReasonSeg but significantly surpasses them on ConstraintSeg, achieving nearly a " + }, + { + "bbox": [ + 55, + 415, + 296, + 499 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 55, + 415, + 296, + 499 + ], + "type": "text", + "content": " improvement at the part level. Visual comparisons in Fig. 6 further demonstrate ConSeg's strong generalization to unseen objects, scenes, and tasks." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 508, + 149, + 521 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 508, + 149, + 521 + ], + "spans": [ + { + "bbox": [ + 55, + 508, + 149, + 521 + ], + "type": "text", + "content": "4.5. Ablation Study" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 527, + 284, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 527, + 284, + 539 + ], + "spans": [ + { + "bbox": [ + 55, + 527, + 284, + 539 + ], + "type": "text", + "content": "We conduct ablation studies and present analyses below." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 542, + 295, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 542, + 295, + 626 + ], + "spans": [ + { + "bbox": [ + 55, + 542, + 295, + 626 + ], + "type": "text", + "content": "Multi views are critical for visual programming. As shown in Tab. 6, using only front-view images with constraint elements reduces the average success rate from " + }, + { + "bbox": [ + 55, + 542, + 295, + 626 + ], + "type": "inline_equation", + "content": "65\\%" + }, + { + "bbox": [ + 55, + 542, + 295, + 626 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 542, + 295, + 626 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 55, + 542, + 295, + 626 + ], + "type": "text", + "content": ". This is primarily due to (1) occlusion leading to suboptimal element generation and (2) errors in visual programming, where dimensional reduction causes surfaces to be misinterpreted as lines or lines as points." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "text", + "content": "Constraint-aware segmentation enhances element. As shown in Tab. 6, using DINOv2 [42] to generate semantic points to form elements, rather than through constraint-aware segmentation, reduces the success rate from " + }, + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "inline_equation", + "content": "65\\%" + }, + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "inline_equation", + "content": "42.5\\%" + }, + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "text", + "content": ", because these constructed elements fail to represent the desired constraints accurately. For example, capturing a book's vertical orientation requires two precise points on its" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 270, + 470, + 282 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 270, + 470, + 282 + ], + "spans": [ + { + "bbox": [ + 313, + 270, + 470, + 282 + ], + "type": "text", + "content": "edge, which DINOv2 can not provide." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 288, + 554, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 288, + 554, + 361 + ], + "spans": [ + { + "bbox": [ + 313, + 288, + 554, + 361 + ], + "type": "text", + "content": "Forming elements improves code generation. As shown in Tab. 6, using unconnected 3D points as final elements reduces the success rate from " + }, + { + "bbox": [ + 313, + 288, + 554, + 361 + ], + "type": "inline_equation", + "content": "65\\%" + }, + { + "bbox": [ + 313, + 288, + 554, + 361 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 288, + 554, + 361 + ], + "type": "inline_equation", + "content": "55\\%" + }, + { + "bbox": [ + 313, + 288, + 554, + 361 + ], + "type": "text", + "content": ". Pre-formed elements contain more prior constraint information, serving as visual cues better encoded in code, whereas standalone points require recombination during visual programming." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 365, + 555, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 365, + 555, + 462 + ], + "spans": [ + { + "bbox": [ + 313, + 365, + 555, + 462 + ], + "type": "text", + "content": "Elements simplify constraints computation. As shown in Tab. 4, the success rate decreases from " + }, + { + "bbox": [ + 313, + 365, + 555, + 462 + ], + "type": "inline_equation", + "content": "60\\%" + }, + { + "bbox": [ + 313, + 365, + 555, + 462 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 365, + 555, + 462 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 313, + 365, + 555, + 462 + ], + "type": "text", + "content": " when entities or parts are not transformed into proposed elements in real-world evaluation. We find two key issues: (1) inaccurate 3D position and slow pose tracking of entities or parts; and (2) difficulties in performing arithmetic operations on their 3D positions or poses in code, hindering the encoding of their spatio-temporal constraints." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 467, + 555, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 467, + 555, + 538 + ], + "spans": [ + { + "bbox": [ + 313, + 467, + 555, + 538 + ], + "type": "text", + "content": "Both instance-level and part-level data improve performance. We conduct ablation studies to assess the impact of our proposed dataset on the ConstraintSeg benchmark. The results are shown in Tab. 7. The results indicate that both the instance-level and part-level subsets contribute to performance improvements and enhance each other's effects." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 559, + 388, + 571 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 559, + 388, + 571 + ], + "spans": [ + { + "bbox": [ + 313, + 559, + 388, + 571 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "content": "In this paper, we present a novel paradigm termed Code-as-Monitor, leveraging the VLMs for both open-set reactive and proactive failure detection. In detail, We formulate both detection modes as spatio-temporal constraint satisfaction problems and use VLM-generated code to evaluate them for real-time monitoring. We further propose constraint elements, which abstract constraints-related entities or their parts into compact geometric elements, to improve the precision and efficiency of monitoring. Extensive experiments demonstrate the superiority of the proposed approach and highlight its potential to advance closed-loop robot systems." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "6926" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 140 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 140 + ], + "type": "text", + "content": "This work was supported by the National Natural Science Foundation of China (62132001), the National Science and Technology Major Project (2022ZD0116314), and the Fundamental Research Funds for the Central Universities." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 151, + 115, + 164 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 151, + 115, + 164 + ], + "spans": [ + { + "bbox": [ + 56, + 151, + 115, + 164 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 171, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 61, + 171, + 296, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 171, + 296, + 215 + ], + "spans": [ + { + "bbox": [ + 61, + 171, + 296, + 215 + ], + "type": "text", + "content": "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv, 2023. 2, 3, 4, 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 217, + 296, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 217, + 296, + 258 + ], + "spans": [ + { + "bbox": [ + 61, + 217, + 296, + 258 + ], + "type": "text", + "content": "[2] Aneseh Alvanpour, Sumit Kumar Das, Christopher Kevin Robinson, Olfa Nasraoui, and Dan Popa. Robot failure mode prediction with explainable machine learning. CASE, 2020. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 261, + 296, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 261, + 296, + 316 + ], + "spans": [ + { + "bbox": [ + 61, + 261, + 296, + 316 + ], + "type": "text", + "content": "[3] Jingkun An, Yinghao Zhu, Zongjian Li, Enshen Zhou, Haoran Feng, Xijie Huang, Bohua Chen, Yemin Shi, and Chengwei Pan. Agfpsync: Leveraging ai-generated feedback for preference optimization in text-to-image generation. arXiv preprint arXiv:2403.13352, 2024. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 61, + 318, + 295, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 318, + 295, + 361 + ], + "spans": [ + { + "bbox": [ + 61, + 318, + 295, + 361 + ], + "type": "text", + "content": "[4] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. *NeurIPS*, 2020. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 363, + 295, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 363, + 295, + 384 + ], + "spans": [ + { + "bbox": [ + 62, + 363, + 295, + 384 + ], + "type": "text", + "content": "[5] Holger Caesar, Jasper Uijlings, and Vittorio Ferrari. Cocostuff: Thing and stuff classes in context. CVPR, 2018. 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 386, + 296, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 386, + 296, + 430 + ], + "spans": [ + { + "bbox": [ + 62, + 386, + 296, + 430 + ], + "type": "text", + "content": "[6] Xianjie Chen, Roozbeh Mottaghi, Xiaobai Liu, Sanja Fidler, Raquel Urtasun, and Alan Yuille. Detect what you can: Detecting and representing objects using holistic models and body parts. CVPR, 2014. 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 431, + 296, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 431, + 296, + 485 + ], + "spans": [ + { + "bbox": [ + 62, + 431, + 296, + 485 + ], + "type": "text", + "content": "[7] Zeren Chen, Zhelun Shi, Xiaoya Lu, Lehan He, Sucheng Qian, Zhenfei Yin, Wanli Ouyang, Jing Shao, Yu Qiao, Cewu Lu, et al. Rh20t-p: A primitive-level robotic dataset towards composable generalization agents. arXiv preprint arXiv:2403.19622, 2024. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 487, + 296, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 487, + 296, + 520 + ], + "spans": [ + { + "bbox": [ + 62, + 487, + 296, + 520 + ], + "type": "text", + "content": "[8] Jaemin Cho, Abhay Zala, and Mohit Bansal. Visual programming for step-by-step text-to-image generation and evaluation. NeurIPS, 2024. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 522, + 296, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 522, + 296, + 575 + ], + "spans": [ + { + "bbox": [ + 62, + 522, + 296, + 575 + ], + "type": "text", + "content": "[9] Wenbo Cui, Chengyang Zhao, Songlin Wei, Jiazhao Zhang, Haoran Geng, Yaran Chen, and He Wang. Gapartmanip: A large-scale part-centric dataset for material-agnostic articulated object manipulation. arXiv preprint arXiv:2411.18276, 2024. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 578, + 295, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 578, + 295, + 610 + ], + "spans": [ + { + "bbox": [ + 57, + 578, + 295, + 610 + ], + "type": "text", + "content": "[10] Yinpei Dai, Jayjun Lee, Nima Fazeli, and Joyce Chai. Racer: Rich language-guided failure recovery policies for imitation learning. arXiv, 2024. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 613, + 295, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 613, + 295, + 644 + ], + "spans": [ + { + "bbox": [ + 57, + 613, + 295, + 644 + ], + "type": "text", + "content": "[11] Maximilian Diehl and Karinne Ramirez-Amaro. A causal-based approach to explain, predict and prevent failures in robotic tasks. RAS, 2023. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 647, + 295, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 295, + 689 + ], + "type": "text", + "content": "[12] Yuqing Du, Ksenia Konyushkova, Misha Denil, Akhil Raju, Jessica Landon, Felix Hill, Nando de Freitas, and Serkan Cabi. Vision-language models as success detectors. arXiv, 2023. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "type": "text", + "content": "[13] Jiafei Duan, Wilbert Pumacay, Nishanth Kumar, Yi Ru Wang, Shulin Tian, Wentao Yuan, Ranjay Krishna, Dieter" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "text", + "content": "Fox, Ajay Mandlekar, and Yijie Guo. Aha: A vision-language-model for detecting and reasoning over failures in robotic manipulation. arXiv, 2024. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 108, + 553, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 108, + 553, + 151 + ], + "spans": [ + { + "bbox": [ + 316, + 108, + 553, + 151 + ], + "type": "text", + "content": "[14] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv, 2024. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 153, + 553, + 197 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 153, + 553, + 197 + ], + "spans": [ + { + "bbox": [ + 316, + 153, + 553, + 197 + ], + "type": "text", + "content": "[15] Yanjiang Guo, Yen-Jen Wang, Lihan Zha, Zheyuan Jiang, and Jianyu Chen. Doremi: Grounding language model by detecting and recovering from plan-execution misalignment. arXiv, 2023. 2, 5, 6, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 199, + 553, + 232 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 199, + 553, + 232 + ], + "spans": [ + { + "bbox": [ + 316, + 199, + 553, + 232 + ], + "type": "text", + "content": "[16] Tanmay Gupta and Aniruddha Kembhavi. Visual programming: Compositional visual reasoning without training. CVPR, 2023. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 234, + 553, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 234, + 553, + 267 + ], + "spans": [ + { + "bbox": [ + 316, + 234, + 553, + 267 + ], + "type": "text", + "content": "[17] Edward J Hu, Phillip Wallis, Zeyuan Allen-Zhu, Yanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. *ICLR*, 2021. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 269, + 553, + 302 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 269, + 553, + 302 + ], + "spans": [ + { + "bbox": [ + 316, + 269, + 553, + 302 + ], + "type": "text", + "content": "[18] Xuhao Hu, Dongrui Liu, Hao Li, Xuanjing Huang, and Jing Shao. Vlsbench: Unveiling visual leakage in multimodal safety. arXiv preprint arXiv:2411.19939, 2024. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 304, + 553, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 304, + 553, + 346 + ], + "spans": [ + { + "bbox": [ + 316, + 304, + 553, + 346 + ], + "type": "text", + "content": "[19] Haoxu Huang, Fanqi Lin, Yingdong Hu, Shengjie Wang, and Yang Gao. Copa: General robotic manipulation through spatial constraints of parts with foundation models. arXiv, 2024. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 350, + 553, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 350, + 553, + 403 + ], + "spans": [ + { + "bbox": [ + 316, + 350, + 553, + 403 + ], + "type": "text", + "content": "[20] Wenlong Huang, Fei Xia, Ted Xiao, Harris Chan, Jacky Liang, Pete Florence, Andy Zeng, Jonathan Tompson, Igor Mordatch, Yevgen Chebotar, et al. Inner monologue: Embodied reasoning through planning with language models. arXiv, 2022. 2, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 406, + 553, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 406, + 553, + 449 + ], + "spans": [ + { + "bbox": [ + 316, + 406, + 553, + 449 + ], + "type": "text", + "content": "[21] Wenlong Huang, Chen Wang, Yunzhu Li, Ruohan Zhang, and Li Fei-Fei. Rekep: Spatio-temporal reasoning of relational keypoint constraints for robotic manipulation. arXiv, 2024. 3, 5, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 452, + 553, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 452, + 553, + 485 + ], + "spans": [ + { + "bbox": [ + 316, + 452, + 553, + 485 + ], + "type": "text", + "content": "[22] Stephen James, Zicong Ma, David Rovick Arrojo, and Andrew J. Davison. Rlbench: The robot learning benchmark & learning environment. RAL, 2020. 2, 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 487, + 553, + 541 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 487, + 553, + 541 + ], + "spans": [ + { + "bbox": [ + 316, + 487, + 553, + 541 + ], + "type": "text", + "content": "[23] Yuheng Ji, Huajie Tan, Jiayu Shi, Xiaoshuai Hao, Yuan Zhang, Hengyuan Zhang, Pengwei Wang, Mengdi Zhao, Yao Mu, Pengju An, et al. Robobrain: A unified brain model for robotic manipulation from abstract to concrete. arXiv preprint arXiv:2502.21257, 2025. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 544, + 553, + 598 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 544, + 553, + 598 + ], + "spans": [ + { + "bbox": [ + 316, + 544, + 553, + 598 + ], + "type": "text", + "content": "[24] Yueru Jia, Jiaming Liu, Sixiang Chen, Chenyang Gu, Zhilue Wang, Longzan Luo, Lily Lee, Pengwei Wang, Zhongyuan Wang, Renrui Zhang, et al. Lift3d foundation policy: Lifting 2d large-scale pretrained models for robust 3d robotic manipulation. arXiv preprint arXiv:2411.18623, 2024. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 600, + 553, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 600, + 553, + 633 + ], + "spans": [ + { + "bbox": [ + 316, + 600, + 553, + 633 + ], + "type": "text", + "content": "[25] Nikita Karaev, Ignacio Rocco, Benjamin Graham, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. Co-tracker: It is better to track together. ECCV, 2024. 5" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 635, + 553, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 553, + 667 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 553, + 667 + ], + "type": "text", + "content": "[26] Sahar Kazemzadeh, Vicente Ordonez, Mark Matten, and Tamara Berg. Referitag: Referring to objects in photographs of natural scenes. EMNLP, 2014. 8" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "text", + "content": "[27] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. ICCV, 2023. 3, 4" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6927" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "type": "text", + "content": "[28] Xin Lai, Zhuotao Tian, Yukang Chen, Yanwei Li, Yuhui Yuan, Shu Liu, and Jiaya Jia. Lisa: Reasoning segmentation via large language model. CVPR, 2024. 4, 7, 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 295, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 295, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 295, + 140 + ], + "type": "text", + "content": "[29] Olivia Y Lee, Annie Xie, Kuan Fang, Karl Pertsch, and Chelsea Finn. Affordance-guided reinforcement learning via visual prompting. arXiv, 2024. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 141, + 294, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 141, + 294, + 174 + ], + "spans": [ + { + "bbox": [ + 56, + 141, + 294, + 174 + ], + "type": "text", + "content": "[30] Gregory LeMasurier, Alvika Gautam, Zhao Han, Jacob W Crandall, and Holly A Yanco. Reactive or proactive? how robots should explain failures. HRI, 2024. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 175, + 294, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 294, + 283 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 294, + 283 + ], + "type": "text", + "content": "[31] Chengshu Li, Ruohan Zhang, Josiah Wong, Cem Gokmen, Sanjana Srivastava, Roberto Martin-Martin, Chen Wang, Gabrael Levine, Michael Lingelbach, Jiankai Sun, Mona Anvari, Minjune Hwang, Manasi Sharma, Arman Aydin, Dhruva Bansal, Samuel Hunter, Kyu-Young Kim, Alan Lou, Caleb R Matthews, Ivan Villa-Renteria, Jerry Huayang Tang, Claire Tang, Fei Xia, Silvio Savarese, Hyowon Gweon, Karen Liu, Jiajun Wu, and Li Fei-Fei. Behavior-1k: A benchmark for embodied ai with 1,000 everyday activities and realistic simulation. CoRL, 2022. 2, 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 286, + 294, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 286, + 294, + 329 + ], + "spans": [ + { + "bbox": [ + 56, + 286, + 294, + 329 + ], + "type": "text", + "content": "[32] Feng Li, Hao Zhang, Peize Sun, Xueyan Zou, Shilong Liu, Jianwei Yang, Chunyuan Li, Lei Zhang, and Jianfeng Gao. Semantic-sam: Segment and recognize anything at any granularity. arXiv, 2023. 5, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 331, + 294, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 331, + 294, + 374 + ], + "spans": [ + { + "bbox": [ + 56, + 331, + 294, + 374 + ], + "type": "text", + "content": "[33] Lijun Li, Zhelun Shi, Xuhao Hu, Bowen Dong, Yiran Qin, Xihui Liu, Lu Sheng, and Jing Shao. T2isafety: Benchmark for assessing fairness, toxicity, and privacy in image generation. arXiv preprint arXiv:2501.12612, 2025. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 376, + 294, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 376, + 294, + 419 + ], + "spans": [ + { + "bbox": [ + 56, + 376, + 294, + 419 + ], + "type": "text", + "content": "[34] Jacky Liang, Wenlong Huang, Fei Xia, Peng Xu, Karol Hausman, Brian Ichter, Pete Florence, and Andy Zeng. Code as policies: Language model programs for embodied control. ICRA, 2023. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 421, + 294, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 421, + 294, + 454 + ], + "spans": [ + { + "bbox": [ + 56, + 421, + 294, + 454 + ], + "type": "text", + "content": "[35] Fangchen Liu, Kuan Fang, Pieter Abbeel, and Sergey Levine. Moka: Open-vocabulary robotic manipulation through mark-based visual prompting. arXiv, 2024. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 456, + 294, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 456, + 294, + 476 + ], + "spans": [ + { + "bbox": [ + 56, + 456, + 294, + 476 + ], + "type": "text", + "content": "[36] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. NeurIPS, 2024. 2, 4, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 479, + 294, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 479, + 294, + 533 + ], + "spans": [ + { + "bbox": [ + 56, + 479, + 294, + 533 + ], + "type": "text", + "content": "[37] Jiaming Liu, Chenxuan Li, Guanqun Wang, Lily Lee, Kaichen Zhou, Sixiang Chen, Chuyan Xiong, Jiaxin Ge, Renrui Zhang, and Shanghang Zhang. Self-corrected multimodal large language model for end-to-end robot manipulation. arXiv, 2024. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 535, + 294, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 535, + 294, + 567 + ], + "spans": [ + { + "bbox": [ + 56, + 535, + 294, + 567 + ], + "type": "text", + "content": "[38] Zeyi Liu, Arpit Bahety, and Shuran Song. Reflect: Summarizing robot experiences for failure explanation and correction. CoRL, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 569, + 294, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 569, + 294, + 610 + ], + "spans": [ + { + "bbox": [ + 56, + 569, + 294, + 610 + ], + "type": "text", + "content": "[39] Junhua Mao, Jonathan Huang, Alexander Toshev, Oana Camburu, Alan L Yuille, and Kevin Murphy. Generation and comprehension of unambiguous object descriptions. CVPR, 2016. 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 613, + 294, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 613, + 294, + 666 + ], + "spans": [ + { + "bbox": [ + 56, + 613, + 294, + 666 + ], + "type": "text", + "content": "[40] Yao Mu, Junting Chen, Qinglong Zhang, Shoufa Chen, Qiaojun Yu, Chongjian Ge, Runjian Chen, Zhixuan Liang, Mengkang Hu, Chaofan Tao, et al. Robocodex: Multimodal code generation for robotic behavior synthesis. arXiv, 2024. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "type": "text", + "content": "[41] Soroush Nasiriany, Fei Xia, Wenhao Yu, Ted Xiao, Jacky Liang, Ishita Dasgupta, Annie Xie, Danny Driess, Ayzaan Wahid, Zhuo Xu, et al. Pivot: Iterative visual prompting elicits actionable knowledge for vlms. arXiv, 2024. 3" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 126 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 126 + ], + "type": "text", + "content": "[42] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv, 2023. 3, 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 129, + 553, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 553, + 183 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 553, + 183 + ], + "type": "text", + "content": "[43] Zekun Qi, Wenyao Zhang, Yufei Ding, Runpei Dong, Xinqiang Yu, Jingwen Li, Lingyun Xu, Baoyu Li, Xialin He, Guofan Fan, et al. Sofar: Language-grounded orientation bridges spatial reasoning and object manipulation. arXiv preprint arXiv:2502.13143, 2025. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 185, + 553, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 185, + 553, + 227 + ], + "spans": [ + { + "bbox": [ + 316, + 185, + 553, + 227 + ], + "type": "text", + "content": "[44] Yiran Qin, Zhelun Shi, Jiwen Yu, Xijun Wang, Enshen Zhou, Lijun Li, Zhenfei Yin, Xihui Liu, Lu Sheng, Jing Shao, et al. Worldsimbench: Towards video generation models as world simulators. arXiv preprint arXiv:2410.18072, 2024. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 228, + 553, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 228, + 553, + 293 + ], + "spans": [ + { + "bbox": [ + 316, + 228, + 553, + 293 + ], + "type": "text", + "content": "[45] Yiran Qin, Enshen Zhou, Qichang Liu, Zhenfei Yin, Lu Sheng, Ruimao Zhang, Yu Qiao, and Jing Shao. Mp5: A multi-modal open-ended embodied system in mycraf via active perception. In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16307-16316. IEEE, 2024. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 294, + 553, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 294, + 553, + 326 + ], + "spans": [ + { + "bbox": [ + 316, + 294, + 553, + 326 + ], + "type": "text", + "content": "[46] Yiran Qin, Ao Sun, Yuze Hong, Benyou Wang, and Ruimao Zhang. Navigatediff: Visual predictors are zero-shot navigation assistants. arXiv preprint arXiv:2502.13894, 2025. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 327, + 553, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 327, + 553, + 369 + ], + "spans": [ + { + "bbox": [ + 316, + 327, + 553, + 369 + ], + "type": "text", + "content": "[47] Shreyas Sundara Raman, Vanya Cohen, Eric Rosen, Ifrah Idrees, David Paulius, and Stefanie Tellex. Planning with large language models via corrective re-prompting. NeurIPS, 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 371, + 553, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 371, + 553, + 415 + ], + "spans": [ + { + "bbox": [ + 316, + 371, + 553, + 415 + ], + "type": "text", + "content": "[48] Vignesh Ramanathan, Anmol Kalia, Vladan Petrovic, Yi Wen, Baixue Zheng, Baishan Guo, Rui Wang, Aaron Marquez, Rama Kovvuri, Abhishek Kadian, et al. Paco: Parts and attributes of common objects. CVPR, 2023. 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 416, + 553, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 416, + 553, + 480 + ], + "spans": [ + { + "bbox": [ + 316, + 416, + 553, + 480 + ], + "type": "text", + "content": "[49] Tianhe Ren, Shilong Liu, Ailing Zeng, Jing Lin, Kunchang Li, He Cao, Jiayu Chen, Xinyu Huang, Yukang Chen, Feng Yan, Zhaoyang Zeng, Hao Zhang, Feng Li, Jie Yang, Hongyang Li, Qing Jiang, and Lei Zhang. Grounded sam: Assembling open-world models for diverse visual tasks. arXiv, 2024. 5, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 482, + 553, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 482, + 553, + 514 + ], + "spans": [ + { + "bbox": [ + 316, + 482, + 553, + 514 + ], + "type": "text", + "content": "[50] Zhongwei Ren, Zhicheng Huang, Yunchao Wei, Yao Zhao, Dongmei Fu, Jiashi Feng, and Xiaojie Jin. Pixel reasoning with large multimodal model. CVPR, 2024. 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 515, + 553, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 515, + 553, + 547 + ], + "spans": [ + { + "bbox": [ + 316, + 515, + 553, + 547 + ], + "type": "text", + "content": "[51] Kenneth Shaw, Ananye Agarwal, and Deepak Pathak. Leap hand: Low-cost, efficient, and anthropomorphic hand for robot learning. arXiv, 2023. 5, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 548, + 553, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 548, + 553, + 601 + ], + "spans": [ + { + "bbox": [ + 316, + 548, + 553, + 601 + ], + "type": "text", + "content": "[52] Zhelun Shi, Zhipin Wang, Hongxing Fan, Zaibin Zhang, Lijun Li, Yongting Zhang, Zhenfei Yin, Lu Sheng, Yu Qiao, and Jing Shao. Assessment of multimodal large language models in alignment with human values. arXiv preprint arXiv:2403.17830, 2024. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 604, + 553, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 604, + 553, + 634 + ], + "spans": [ + { + "bbox": [ + 316, + 604, + 553, + 634 + ], + "type": "text", + "content": "[53] Noah Shinn, Beck Labash, and Ashwin Gopinath. Reflection: an autonomous agent with dynamic memory and self-reflection. arXiv, 2023. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 636, + 553, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 636, + 553, + 668 + ], + "spans": [ + { + "bbox": [ + 316, + 636, + 553, + 668 + ], + "type": "text", + "content": "[54] Mohit Shridhar, Lucas Manuelli, and Dieter Fox. CIport: What and where pathways for robotic manipulation. CoRL, 2021. 2, 5, 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "text", + "content": "[55] Tom Silver, Soham Dan, Kavitha Srinivas, Joshua B Tenenbaum, Leslie Kaelbling, and Michael Katz. Generalized planning in pddl domains with pretrained large language models. AAAI, 2024. 2" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "6928" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 137 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 137 + ], + "type": "text", + "content": "[56] Marta Skreta, Naruki Yoshikawa, Sebastian Arellano-Rubach, Zhi Ji, Lasse Bjørn Kristensen, Kourosh Darvish, Alán Aspuru-Guzik, Florian Shkurti, and Animesh Garg. Errors are useful prompts: Instruction guided task programming with verifier-assisted iterative prompting. arXiv, 2023. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 140, + 294, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 140, + 294, + 195 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 294, + 195 + ], + "type": "text", + "content": "[57] Yingbo Tang, Shuaike Zhang, Xiaoshuai Hao, Pengwei Wang, Jianlong Wu, Zhongyuan Wang, and Shanghang Zhang. Affordgrasp: In-context affordance reasoning for open-vocabulary task-oriented grasping in clutter. arXiv preprint arXiv:2503.00778, 2025. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 197, + 294, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 197, + 294, + 250 + ], + "spans": [ + { + "bbox": [ + 56, + 197, + 294, + 250 + ], + "type": "text", + "content": "[58] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 252, + 294, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 252, + 294, + 306 + ], + "spans": [ + { + "bbox": [ + 56, + 252, + 294, + 306 + ], + "type": "text", + "content": "[59] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 308, + 294, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 308, + 294, + 341 + ], + "spans": [ + { + "bbox": [ + 56, + 308, + 294, + 341 + ], + "type": "text", + "content": "[60] Sai H Vemprala, Rogerio Bonatti, Arthur Tucker, and Ashish Kapoor. Chatgpt for robotics: Design principles and model abilities. IEEE Access, 2024. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 342, + 294, + 375 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 342, + 294, + 375 + ], + "spans": [ + { + "bbox": [ + 56, + 342, + 294, + 375 + ], + "type": "text", + "content": "[61] David Venuto, Sami Nur Islam, Martin Klissarov, Doina Precup, Sherry Yang, and Ankit Anand. Code as reward: Empowering reinforcement learning with vlms. arXiv, 2024. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 376, + 294, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 376, + 294, + 430 + ], + "spans": [ + { + "bbox": [ + 56, + 376, + 294, + 430 + ], + "type": "text", + "content": "[62] Homer Rich Walke, Kevin Black, Tony Z Zhao, Quan Vuong, Chongyi Zheng, Philippe Hansen-Estruch, Andre Wang He, Vivek Myers, Moo Jin Kim, Max Du, et al. Bridgedata v2: A dataset for robot learning at scale. CoRL, 2023. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 433, + 294, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 433, + 294, + 465 + ], + "spans": [ + { + "bbox": [ + 56, + 433, + 294, + 465 + ], + "type": "text", + "content": "[63] Jianwei Yang, Hao Zhang, Feng Li, Xueyan Zou, Chunyuan Li, and Jianfeng Gao. Set-of-mark prompting unleashes extraordinary visual grounding in gpt-4v. arXiv, 2023. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 467, + 294, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 467, + 294, + 499 + ], + "spans": [ + { + "bbox": [ + 56, + 467, + 294, + 499 + ], + "type": "text", + "content": "[64] Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, and Yuan Cao. React: Synergizing reasoning and acting in language models. arXiv, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 501, + 294, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 501, + 294, + 532 + ], + "spans": [ + { + "bbox": [ + 56, + 501, + 294, + 532 + ], + "type": "text", + "content": "[65] Runpeng Yu, Weihao Yu, and Xinchao Wang. Attention prompting on image for large vision-language models. arXiv, 2024. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 535, + 294, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 535, + 294, + 577 + ], + "spans": [ + { + "bbox": [ + 56, + 535, + 294, + 577 + ], + "type": "text", + "content": "[66] Zhouliang Yu, Jie Fu, Yao Mu, Chenguang Wang, Lin Shao, and Yaodong Yang. Multireact: Multimodal tools augmented reasoning-acting traces for embodied agent planning. *ICLR*, 2024. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 580, + 294, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 580, + 294, + 623 + ], + "spans": [ + { + "bbox": [ + 56, + 580, + 294, + 623 + ], + "type": "text", + "content": "[67] Wentao Yuan, Jiafei Duan, Valts Blukis, Wilbert Pumacay, Ranjay Krishna, Adithyavairavan Murali, Arsalan Mousavian, and Dieter Fox. Robopoint: A vision-language model for spatial affordance prediction for robotics. arXiv, 2024. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 624, + 294, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 294, + 666 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 294, + 666 + ], + "type": "text", + "content": "[68] Zhihao Yuan, Jinke Ren, Chun-Mei Feng, Hengshuang Zhao, Shuguang Cui, and Zhen Li. Visual programming for zero-shot open-vocabulary 3d visual grounding. CVPR, 2024. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 670, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 294, + 713 + ], + "type": "text", + "content": "[69] Jialiang Zhang, Haoran Liu, Danshi Li, XinQiang Yu, Hao ran Geng, Yufei Ding, Jiayi Chen, and He Wang. Dexgrasp-net 2.0: Learning generative dexterous grasping in largescale synthetic cluttered scenes. CoRL, 2024. 5, 7" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 363 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "text", + "content": "[70] Xinyu Zhang, Yuhan Liu, Haonan Chang, Liam Schramm, and Abdeslam Boularias. Autoregressive action sequence learning for robotic manipulation. arXiv, 2024. 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 108, + 553, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 108, + 553, + 161 + ], + "spans": [ + { + "bbox": [ + 316, + 108, + 553, + 161 + ], + "type": "text", + "content": "[71] Yongting Zhang, Lu Chen, Guodong Zheng, Yifeng Gao, Rui Zheng, Jinlan Fu, Zhenfei Yin, Senjie Jin, Yu Qiao, Xuanjing Huang, et al. Spa-vl: A comprehensive safety preference alignment dataset for vision language model. arXiv preprint arXiv:2406.12030, 2024. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 163, + 553, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 553, + 206 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 553, + 206 + ], + "type": "text", + "content": "[72] Zaibin Zhang, Shiyu Tang, Yuanhang Zhang, Talas Fu, Yifan Wang, Yang Liu, Dong Wang, Jing Shao, Lijun Wang, and Huchuan Lu. Ad-h: Autonomous driving with hierarchical agents. arXiv preprint arXiv:2406.03474, 2024. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 209, + 553, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 209, + 553, + 240 + ], + "spans": [ + { + "bbox": [ + 316, + 209, + 553, + 240 + ], + "type": "text", + "content": "[73] Jinliang Zheng, Jianxiong Li, Sijie Cheng, Yinan Zheng, Ji-aming Li, Jihao Liu, Yu Liu, Jingjing Liu, and Xianyuan Zhan. Instruction-guided visual masking. arXiv, 2024. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 243, + 553, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 243, + 553, + 274 + ], + "spans": [ + { + "bbox": [ + 316, + 243, + 553, + 274 + ], + "type": "text", + "content": "[74] Zhi Zheng, Qian Feng, Hang Li, Alois Knoll, and Jianxiang Feng. Evaluating uncertainty-based failure detection for closed-loop llm planners. arXiv, 2024. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 276, + 553, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 276, + 553, + 307 + ], + "spans": [ + { + "bbox": [ + 316, + 276, + 553, + 307 + ], + "type": "text", + "content": "[75] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. CVPR, 2017. 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 309, + 553, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 309, + 553, + 363 + ], + "spans": [ + { + "bbox": [ + 316, + 309, + 553, + 363 + ], + "type": "text", + "content": "[76] Enshen Zhou, Yiran Qin, Zhenfei Yin, Yuzhou Huang, Ruimao Zhang, Lu Sheng, Yu Qiao, and Jing Shao. Minedreamer: Learning to follow instructions via chain-of-imagination for simulated-world control. arXiv preprint arXiv:2403.12037, 2024. 2" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "6929" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Coeff-Tuning_ A Graph Filter Subspace View for Tuning Attention-Based Large Models/e7a23930-dc20-4352-a4dc-e2a2616c610f_content_list.json b/2025/Coeff-Tuning_ A Graph Filter Subspace View for Tuning Attention-Based Large Models/e7a23930-dc20-4352-a4dc-e2a2616c610f_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..66c53d0519632b6da0de1d802a3efc55532c7177 --- /dev/null +++ b/2025/Coeff-Tuning_ A Graph Filter Subspace View for Tuning Attention-Based Large Models/e7a23930-dc20-4352-a4dc-e2a2616c610f_content_list.json @@ -0,0 +1,1761 @@ +[ + { + "type": "text", + "text": "Coeff-Tuning: A Graph Filter Subspace View for Tuning Attention-Based Large Models", + "text_level": 1, + "bbox": [ + 96, + 128, + 898, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zichen Miao†, Wei Chen†, Qiang Qiu Purdue University, IN, USA", + "bbox": [ + 346, + 202, + 647, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{miaoz, chen2732, qqiu}@purdue.edu", + "bbox": [ + 346, + 241, + 647, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 291, + 326, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Transformer-based large pre-trained models have shown remarkable generalization ability, and various parameter-efficient fine-tuning (PEFT) methods have been proposed to customize these models on downstream tasks with minimal computational and memory budgets. Previous PEFT methods are primarily designed from a tensor-decomposition perspective that tries to effectively tune the linear transformation by finding the smallest subset of parameters to train. Our study adopts an orthogonal view by representing the attention operation as a graph convolution and formulating the multi-head attention maps as a convolutional filter subspace, with each attention map as a subspace element. In this paper, we propose to tune the large pre-trained transformers by learning a small set of combination coefficients that construct a more expressive filter subspace from the original multi-head attention maps. We show analytically and experimentally that the tuned filter subspace can effectively expand the feature space of the multi-head attention and further enhance the capacity of transformers. We further stabilize the fine-tuning with a residual parameterization of the tunable subspace coefficients, and enhance the generalization with a regularization design by directly applying dropout on the tunable coefficient during training. The tunable coefficients take a tiny number of parameters and can be combined with previous PEFT methods in a plug-and-play manner. Extensive experiments show that our approach achieves superior performances than PEFT baselines with neglectable additional parameters. $^{1}$", + "bbox": [ + 86, + 323, + 485, + 748 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 777, + 223, + 792 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Large foundational models have demonstrated remarkable performance and generalization ability across a wide range of natural language processing and computer vision", + "bbox": [ + 89, + 804, + 483, + 849 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0815525a3a84eb0cfff84eee395e64934a19e0f9a5d2a1c968962f9036d3cc75.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 519, + 290, + 717, + 415 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/49bb2896c59a987c7204b513e0614d167910713c6d31bab73f622a30fbfb3e67.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 723, + 296, + 895, + 412 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/08444f97283a41285cc1e6a3fac51216f3a3927e61125f45277baec4bdd4d0c4.jpg", + "image_caption": [ + "A [V] plushie panda on top of a purple rug in a forest." + ], + "image_footnote": [], + "bbox": [ + 537, + 435, + 687, + 551 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/74646b37219c208b016d61193f640aa8da780874b6c2829704fbeaa9300ff873.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 707, + 435, + 856, + 550 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/498db363057d9dbe54d77e5f5fcfbf7156bc5ec867e58301e3bb0e9f6348f0fa.jpg", + "image_caption": [ + "A [V] plushie panda on in the snow.", + "LoRA", + "Figure 1. (Up) Under the graph convolution view of the attention, the proposed Coeff-Tuning enhances the expressiveness by combining the attention maps with a tiny tunable subspace coefficient $\\alpha$ . It can be seamlessly integrated with other PEFT methods, e.g., LoRA [21], by introducing negligible additional parameters. (Bottom) Combining Coeff-Tuning with LoRA achieves better results in the concept customization task with SDXL [51]." + ], + "image_footnote": [], + "bbox": [ + 540, + 566, + 689, + 684 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/236494ecb81262f72717a042ff784bd414d1ceb29f4bd63f97f0ec10e063104f.jpg", + "image_caption": [ + "LoRA + Coeff-Tuning" + ], + "image_footnote": [], + "bbox": [ + 707, + 566, + 856, + 683 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tasks [11, 28, 55, 68]. The success of these models can be attributed to their ability to learn rich, transferable representations by leveraging the power of the transformer architecture [68] and the extensive knowledge captured during", + "bbox": [ + 511, + 839, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$\\dagger$ Equal contribution.", + "bbox": [ + 107, + 862, + 218, + 875 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "1Code is available at: https://github.com/ZichenMiao/Coeff_Tuning", + "bbox": [ + 91, + 875, + 480, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "20146", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "pre-training on large-scale datasets [1, 58]. However, fine-tuning these pre-trained transformers on downstream tasks with full-parameter updates can be challenging due to the high computational and memory costs.", + "bbox": [ + 89, + 90, + 480, + 151 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address this, various parameter-efficient fine-tuning (PEFT) methods have been developed to adapt pre-trained models on downstream tasks while minimizing the number of trainable parameters [21, 38, 53, 79]. A representative family of these approaches focuses on tuning the linear transformation layers in transformers, where they apply a lightweight parameterization, e.g., low-rank parameterization [21, 38], to adapt the weight matrices. However, none of these methods are designed from a holistic view of the core operation in transformers, the multi-head attention layer [68], which plays a key role in the expressiveness of transformers [10, 59].", + "bbox": [ + 89, + 154, + 482, + 333 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we propose a new PEFT method tailored to enhance the multi-head attention's expressiveness by tuning a minimal set of parameters. We first formulate the multi-head attention as a graph convolution with a filter subspace, with each graph filter as the attention map of each attention head. We then tune the filter subspace by simply learning a small set of subspace coefficients that linearly combine the pre-trained attention maps. In this way, we adapt for each attention head a new task-adaptive attention map that better captures the token-wise correlation.", + "bbox": [ + 89, + 335, + 482, + 487 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We attribute the effectiveness of our approach to its ability to expand the feature space of the original multi-head attention layer. By applying the softmax function to the query-key correlation matrices, all elements in a row of the original attention map are non-negative and have a sum of one [68]. This limits the expressiveness of the multi-head attention as the output token embeddings lie within the convex hull of the value embeddings. Our method constructs a new filter subspace with non-constrained coefficients as shown in Figure 1, allowing negative elements learned in the combined attention map to capture more diverse token-wise relationships. This also enables the output tokens to obtain representation outside the convex hull of the value embeddings, which further enhances the capacity of the multi-head attention. We provide analysis and an illustrative toy example to better demonstrate this point.", + "bbox": [ + 89, + 489, + 482, + 731 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We show that our approach can be seamlessly combined with previous PEFT methods in a plug-and-play manner. As the filter subspace in each attention layer is spanned by a tiny number of attention maps (usually $10\\sim 30$ ), we only introduce neglectable additional tunable parameters and training costs over the base PEFT methods.", + "bbox": [ + 89, + 732, + 482, + 823 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We validate our approach on a large variety of fine-tuning tasks and datasets, along with pre-trained transformers of different scales and modalities. Specifically, we first show the effectiveness of our method with ViT [11] on the few-shot image classification fine-tuning. We also", + "bbox": [ + 89, + 824, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "show our method can achieve superior performance with vision-language transformers on image-text multi-modal fine-tuning [8]. Moreover, we also show our method can better customize the text-to-image diffusion models on personalized generation [56].", + "bbox": [ + 511, + 90, + 903, + 166 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We summarize our contributions as follows,", + "bbox": [ + 532, + 167, + 823, + 180 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We provide a filter subspace view of the multi-head attention and propose to fine-tune a new filter subspace by learning a small set of subspace coefficients.", + "- We demonstrate that our method can expand the feature space of the original multi-head attention and enhance its expressiveness on downstream tasks.", + "- We validate our method on extensive fine-tuning tasks and various transformer architectures." + ], + "bbox": [ + 511, + 181, + 903, + 301 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 316, + 653, + 333 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Parameter-Efficient Fine-Tuning. Large attention-based transformers [11, 54, 55, 68] gain a significant amount of generalizable knowledge by pre-training on large-scale datasets [57, 58]. A common way to utilize those general skills is to fine-tune large transformers on domain-specific tasks and data. Recent advancements in parameter-efficient fine-tuning (PEFT) methods aim to reduce the computational and memory costs of adapting large models by fine-tuning only a subset of parameters. LoRA [21] and its variants [4, 24, 29, 36, 78, 82] fine-tune lower-rank matrices at each layer to represent weight updates. DoRA [38] enhances LoRA through weight decomposition into magnitude and direction, thus improving training stability and learning capacity without additional inference costs. The adapter [20] approach inserts small modules between layers and reduces parameters by only tuning these adapters [2, 24, 34, 50, 79]. Visual prompt tuning (VPT) [23, 60] has introduced a limited number of learnable parameters for optimization while keeping the backbone frozen. SSF [35] proposes scaling and shifting features of each layer in a pre-trained large model. OFT [53] tackles fine-tuning in text-to-image diffusion models by preserving the hyperspherical energy of neuron relationships. However, these methods primarily focus on efficient tuning of the parameter weight matrices, lacking a holistic view of the attention layer. In this paper, we take an orthogonal approach by tuning a tiny combination coefficient given a novel graph filter subspace view of the multi-head attention.", + "bbox": [ + 511, + 343, + 903, + 781 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The Expressiveness of Attention. Transformer-based models have achieved success across vision and language tasks [8, 11, 14, 22, 31, 37] due to the effectiveness of the attention mechanism [68]. Despite their success, various studies have examined the limitations in the expressiveness of transformers [10, 59, 71, 80]. Shi et al. [59] and Zhou et al. [84] identify the attention collapse in transformers as", + "bbox": [ + 511, + 794, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "20147", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/80e2fe75c334f99f28bd010b2c8adf95a63b5aee8e96e79f60c2303965e95bf8.jpg", + "image_caption": [ + "Figure 2. Illustration of the proposed Coeff-Tuning. We first view each token $\\mathbf{X}[i]$ as a node $V_{i}$ in an all-connected graph $\\mathcal{G} = (V,E)$ . The attention operation then turns into a graph convolution with each attention map $\\mathbf{A}^h$ as a graph convolutional filter $\\mathbf{F}^h (\\mathbf{X})$ . In this way, the attention maps in the multi-head attention layer form a filter subspace $\\mathcal{F}(\\mathbf{X})$ , and each element in the filter is constrained in $(0,1)$ after softmax. The proposed Coeff-Tuning then tunes a subspace coefficient $\\alpha \\in \\mathbb{R}^{H\\times H}$ with a tiny number of parameters, that expands the range of attention scores by linearly combining the original filter subspace and enhances the expressiveness of the multi-head attention." + ], + "image_footnote": [], + "bbox": [ + 112, + 89, + 880, + 333 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "a result of self-attention's convergence to a low-rank space, which leads to unstable training as attention scores concentrate [10, 45]. To counteract this, techniques such as hierarchical fusion [59] and stabilizing attention entropy [80] have been proposed to prevent performance stagnation in vision transformers. DeepViT [84] proposes to re-generate diverse attention maps to avoid attention collapse in training vision transformers from scratch. In contrast, our work studies the expressiveness of the attention layer in the context of the large model fine-tuning. We also provide a novel interpretation of multi-head attention maps as a graph filter subspace, and propose the residual parameterization and regularization to stabilize the tuning process.", + "bbox": [ + 88, + 429, + 485, + 626 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 645, + 183, + 660 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. A Filter Subspace View on Attention", + "text_level": 1, + "bbox": [ + 89, + 671, + 408, + 688 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Attention. The attention layer transforms the input sequence $\\mathbf{X} \\in \\mathbb{R}^{N \\times C_i}$ to the output sequence $\\mathbf{O} \\in \\mathbb{R}^{N \\times C_o}$ , where $N$ denotes the sequence length, $C_i$ and $C_o$ are the dimension of input and output features. The attention layer projects the input as $\\mathbf{Q} = \\mathbf{X}\\mathbf{W}_q$ , $\\mathbf{K} = \\mathbf{X}\\mathbf{W}_k$ , $\\mathbf{V} = \\mathbf{X}\\mathbf{W}_v$ using the corresponding projection matrices $\\mathbf{W}_q, \\mathbf{W}_k, \\mathbf{W}_v \\in \\mathbb{R}^{C_i \\times C_o}$ , and calculates the attention map,", + "bbox": [ + 89, + 696, + 483, + 803 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathbf {A} = \\operatorname {s o f t m a x} (\\mathbf {Q K} ^ {\\intercal}) \\tag {1} \\\\ = \\operatorname {s o f t m a x} \\left(\\mathbf {X} \\mathbf {W} _ {q} \\mathbf {W} _ {k} ^ {\\intercal} \\mathbf {X} ^ {\\intercal}\\right). \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 184, + 816, + 482, + 854 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The attention map $\\mathbf{A} \\in \\mathbb{R}^{N \\times N}$ captures the token-wise relationship by doing inner-product in a space transformed by", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "$\\mathbf{W}_q, \\mathbf{W}_k$ . Then the output of the attention layer is,", + "bbox": [ + 513, + 428, + 852, + 445 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathbf {O} = \\operatorname {a t t n} \\left(\\mathbf {W} _ {q}, \\mathbf {W} _ {k}, \\mathbf {W} _ {v}; \\mathbf {X}\\right) \\\\ = \\operatorname {s o f t m a x} \\left(\\mathbf {X} \\mathbf {W} _ {q} \\mathbf {W} _ {k} ^ {\\intercal} \\mathbf {X} ^ {\\intercal}\\right) \\mathbf {X} \\mathbf {W} _ {v} \\tag {2} \\\\ = \\mathbf {A X W} _ {v}. \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 454, + 903, + 507 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Multi-head Attention. Multi-head attention extends this by allowing multiple attention mechanisms, i.e., heads, to work in parallel, with each head independently learning attention patterns. For each attention head, we have,", + "bbox": [ + 511, + 523, + 906, + 585 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {O} ^ {h} = \\operatorname {a t t n} \\left(\\mathbf {W} _ {q} ^ {h}, \\mathbf {W} _ {k} ^ {h}, \\mathbf {W} _ {v} ^ {h}; \\mathbf {X}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 602, + 593, + 906, + 613 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $h = 1,\\dots ,H$ and $H$ is the number of heads. The project matrices for each head are $\\mathbf{W}_q^h,\\mathbf{W}_k^h,\\mathbf{W}_v^h\\in$ $\\mathbb{R}^{C_i\\times \\frac{C_o}{H}}$ . The multi-head attention represents as,", + "bbox": [ + 511, + 619, + 906, + 671 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathbf {Z} = \\operatorname {C o n c a t} \\left[ \\mathbf {O} ^ {1}, \\dots , \\mathbf {O} ^ {H} \\right] \\mathbf {W} _ {o} \\\\ = \\underset {H} {\\operatorname {C o n c a t}} \\left[ \\mathbf {A} ^ {1} \\mathbf {X} \\mathbf {W} _ {v} ^ {1}, \\dots , \\mathbf {A} ^ {H} \\mathbf {X} \\mathbf {W} _ {v} ^ {H} \\right] \\mathbf {W} _ {o} \\tag {4} \\\\ = \\sum_ {h = 1} ^ {H} \\mathbf {A} ^ {h} \\mathbf {X} \\mathbf {W} _ {v o} ^ {h}, \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 555, + 676, + 903, + 762 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbf{W}_o\\in \\mathbb{R}^{C_o\\times C_o}$ and $\\mathbf{W}_{vo}^h = \\mathbf{W}_v^h (\\mathbf{W}_o^h)^\\intercal$ $\\mathbf{W}_o^h\\in$ $\\mathbb{R}^{C_o\\times \\frac{C_o}{H}}$ $\\mathbf{W}_{vo}^h\\in \\mathbb{R}^{C_i\\times C_o}$", + "bbox": [ + 511, + 770, + 906, + 805 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "A Filter Subspace View on Multi-head Attention. We now present our filter subspace view on the multi-head attention. By viewing each token embedding $\\mathbf{X}[n] \\in \\mathbb{R}^{C_i}$ as a node $V_{n}$ , which are connected with all the other nodes with edges $E[m, n] = 1$ , $\\forall m \\in [N]$ , we obtain a graph representation of all the tokens, $\\mathcal{G} = (V, E)$ .", + "bbox": [ + 511, + 809, + 908, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "20148", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We then reformulate the attention operation as the message passing as in Graph Convolution Nerworks [27]. Specifically, inspired by [6], we write the single-head attention as a graph convolution with the filter $\\mathbf{F}^h (\\mathbf{X})$", + "bbox": [ + 89, + 90, + 483, + 154 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {O} ^ {h} [ n,: ] = \\sum_ {\\substack {n ^ {\\prime} \\in [ N ] \\\\ c ^ {\\prime} \\in [ C _ {i} ]}} \\mathbf {F} ^ {h} (\\mathbf {X}) [ n, n ^ {\\prime} ] \\mathbf {X} [ n ^ {\\prime}, c ^ {\\prime} ] \\mathbf {W} ^ {h} [ c ^ {\\prime},,: ], \\tag{5}\n$$\n", + "text_format": "latex", + "bbox": [ + 107, + 161, + 482, + 208 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{W}^h = \\mathbf{W}_{vo}^h\\in \\mathbb{R}^{C_i\\times C_0}$ denotes the weight for linear projection on each node, and the $n$ -th row of the filter, i.e., $\\mathbf{F}^h (\\mathbf{X})[n,:]$ , represents the local graph filter around the $n$ -th node (token). Comparing with (1) and (2), we have,", + "bbox": [ + 89, + 217, + 483, + 280 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {F} ^ {h} (\\mathbf {X}) = \\operatorname {s o f t m a x} \\left(\\mathbf {X} \\mathbf {W} _ {q} \\mathbf {W} _ {k} ^ {\\intercal} \\mathbf {X} ^ {\\intercal}\\right). \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 168, + 289, + 482, + 309 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The multi-head attention individually learns $H$ filters $\\{\\mathbf{F}^1 (X),\\dots,\\mathbf{F}^H (X)\\}$ to aggregate complementary information across nodes. These $H$ filters naturally form a graph convolutional filter subspace $\\mathcal{F}(\\mathbf{X})$ as,", + "bbox": [ + 89, + 316, + 483, + 378 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {F} (\\mathbf {X}) = \\left\\{\\mathbf {F} ^ {h} (\\mathbf {X}) \\right\\} _ {h = 1} ^ {H} \\subseteq (0, 1) ^ {N \\times N}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 158, + 386, + 482, + 407 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "and we denote each $\\mathbf{F}^h (\\mathbf{X})$ as a filter atom. The multi-head attention can then be represented in the way of decomposed graph convolution [6] as,", + "bbox": [ + 89, + 414, + 483, + 462 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {O} [ n,: ] = \\sum_ {h = 1} ^ {H} \\sum_ {\\substack {n ^ {\\prime} \\in [ N ] \\\\ c ^ {\\prime} \\in [ C _ {i} ]}} \\mathbf {F} ^ {h} (\\mathbf {X}) [ n, n ^ {\\prime} ] \\mathbf {X} [ n ^ {\\prime}, c ^ {\\prime} ] \\mathbf {W} ^ {h} [ c ^ {\\prime},: ]. \\tag{8}\n$$\n", + "text_format": "latex", + "bbox": [ + 101, + 470, + 482, + 527 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "It can be seen from (8) that the parameters $\\mathbf{F}^h (\\mathbf{X}),\\mathbf{W}^h$ plays two key yet independent roles in the graph convolution. The $\\mathbf{W}^h$ updates the knowledge for each node by projecting its representation into a new space, while the graph filter $\\mathbf{F}^h (\\mathbf{X})$ guides the information flow across nodes. Intuitively, with a more expressive graph convolutional filter subspace $\\mathcal{F}(\\mathbf{X})$ , each node can aggregate more informative features from other nodes.", + "bbox": [ + 89, + 537, + 483, + 657 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Subspace Coefficient Tuning", + "text_level": 1, + "bbox": [ + 89, + 666, + 346, + 683 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given the re-formulation of multi-head attention maps as a graph convolutional filter subspace, we present an efficient fine-tuning method specifically designed for transformers to enhance attention's expressiveness. Our motivation is that after the softmax function, each graph filter $\\mathbf{F}^h (\\mathbf{X})$ has values in the range (0, 1) with row-sum as one. The normalized attention maps restrict the expressiveness in the way that each output embedding lies in the convex hull of the value embeddings.", + "bbox": [ + 89, + 688, + 482, + 824 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To expand the feature space of the output embeddings, we propose Coeff-Tuning, to tune the filter subspace on the target dataset with a small set of combination coefficients, i.e., subspace coefficient $\\alpha \\in \\mathbb{R}^{H\\times H}$ . As shown in Figure 2, the tunable coefficient $\\alpha$ constructs a more expressive", + "bbox": [ + 89, + 825, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "filter subspace $\\hat{\\mathcal{F}}_{\\alpha}(\\mathbf{X})$ by linearly combining the pre-trained multi-head attention maps $\\mathbf{F}(\\mathbf{X})$ . Each element $\\hat{\\mathbf{F}}^h (\\mathbf{X})$ in the learned filter subspace can be written as,", + "bbox": [ + 511, + 90, + 906, + 138 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {F}} ^ {h} (\\mathbf {X}) = \\boldsymbol {\\alpha} [ h, : ] ^ {\\intercal} \\mathcal {F} (\\mathbf {X}) = \\sum_ {i = 1} ^ {H} \\boldsymbol {\\alpha} [ h, i ] \\mathbf {F} ^ {i} (\\mathbf {X}). \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 555, + 148, + 906, + 191 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The multi-head attention then represents as,", + "bbox": [ + 513, + 202, + 803, + 217 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {O} = \\sum_ {h = 1} ^ {H} \\hat {\\mathbf {F}} ^ {h} (\\mathbf {X}) \\mathbf {X} \\mathbf {W} ^ {h} = \\sum_ {h = 1} ^ {H} \\sum_ {i = 1} ^ {H} \\alpha [ h, i ] \\mathbf {F} ^ {i} (\\mathbf {X}) \\mathbf {X} \\mathbf {W} ^ {h}. \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 228, + 905, + 287 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Note that, mathematically, the subspace coefficient $\\alpha$ is unconstrained, allowing it to combine attention maps with a wider range of values from the original filter subspace $\\mathbf{F}(\\mathbf{X})$ . With attention scores exceeding the range $(0,1)$ , we effectively expand the space for $\\mathbf{O}$ by allowing a more flexible combination of value matrices.", + "bbox": [ + 511, + 297, + 906, + 388 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Residual Parameterization. To facilitate the tuning and preserve the most knowledge in the pre-trained transformer, we propose a residual parameterization of the subspace coefficient,", + "bbox": [ + 511, + 402, + 906, + 463 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\alpha} ^ {\\prime} = \\boldsymbol {\\alpha} + \\mathbf {I},\n$$\n", + "text_format": "latex", + "bbox": [ + 611, + 474, + 697, + 489 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {F}} ^ {h} (\\mathbf {X}) = \\mathbf {F} ^ {h} (\\mathbf {X}) + \\sum_ {i = 1} ^ {H} \\boldsymbol {\\alpha} [ h, i ] \\mathbf {F} ^ {i} (\\mathbf {X}). \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 580, + 494, + 905, + 536 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this way, we maintain the original filter subspace with the identity matrix, while parameterizing the subspace coefficient $\\alpha'$ to learn the residual combination of attention maps on fine-tuning tasks.", + "bbox": [ + 511, + 547, + 906, + 608 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Reguralization on Subspace Coefficient. The proposed Coeff-Tuning substantially enhances the expressiveness of multi-head attention with tunable coefficient $\\alpha$ . However, it may also lead to overfitting and preventing channel projection parameters $\\mathbf{W}_v, \\mathbf{W}_o$ from learning generalizable features. To alleviate this, we propose a simple regularization design that performs dropout [63] directly on the tuned subspace coefficient $\\alpha$ during training,", + "bbox": [ + 511, + 621, + 906, + 743 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\alpha} ^ {\\prime} = \\operatorname {D r o p o u t} (\\boldsymbol {\\alpha}; p) + \\mathbf {I}. \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 622, + 755, + 906, + 772 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this way, we randomly remove an element from $\\alpha$ with a probability of $p$ to avoid overfitting.", + "bbox": [ + 511, + 784, + 905, + 815 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Analysis", + "text_level": 1, + "bbox": [ + 513, + 829, + 609, + 845 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we analyze that adjusting the filter subspace using the proposed subspace coefficient $\\alpha$ can effectively enhance the expressiveness of the multi-head attention.", + "bbox": [ + 511, + 854, + 905, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "20149", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/4d56caa0f51e396ee56aad728933bff47314989d7280470ba025ee4dfe966f6e.jpg", + "image_caption": [ + "Figure 3. The effectiveness of tuning subspace coefficient $\\alpha$ . The goal is to convert the input graph $\\mathbf{X}$ to the target graph $\\mathbf{O}$ using a single attention layer, $\\mathbf{O} = \\sum_{h=1}^{H} \\mathbf{F}^h(\\mathbf{X}) \\mathbf{X} \\mathbf{W}_v^h$ , where $\\mathbf{F}(\\mathbf{X}) = \\text{softmax}(\\mathbf{X} \\mathbf{W}_q \\mathbf{W}_k^\\top \\mathbf{X}^\\top)$ . We have the following observations: (i) When only tuning $\\mathbf{W}_q$ and $\\mathbf{W}_k$ , i.e., updating $\\mathbf{F}(\\mathbf{X})$ , the positions of output nodes are constrained within the range of the input nodes $\\mathbf{X}$ , since the values of $\\mathbf{F}(\\mathbf{X})$ are in the range of $(0,1)$ . (ii) In addition to tuning $\\mathbf{F}(\\mathbf{X})$ , updating $\\mathbf{W}_v$ projects the input as $\\mathbf{X} \\mathbf{W}_v$ , but the output graph is still not aligned with the target graph. (iii) In addition to tuning $\\mathbf{F}(\\mathbf{X})$ , adapting $\\alpha$ successfully transforms the input graph into the target graph. Compared with (i), $\\alpha$ expand the positions of output nodes $\\sum_{i=1}^{H} \\alpha[;, i] \\mathbf{F}^i(\\mathbf{X}) \\mathbf{X}$ beyond the constraints of the input nodes, enhancing the expressiveness of the attention effectively." + ], + "image_footnote": [], + "bbox": [ + 106, + 88, + 869, + 223 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For the filter $\\mathbf{F}(\\mathbf{X})$ in the filter subspace $\\mathcal{F}$ since $\\mathbf{F}(\\mathbf{X}) = \\mathrm{softmax}(\\mathbf{X}\\mathbf{W}_q\\mathbf{W}_k^\\top \\mathbf{X}^\\top)$ , their elements $\\mathbf{F}(\\mathbf{X})[n,n^{\\prime}]$ satisfy the following properties,", + "bbox": [ + 89, + 348, + 482, + 393 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n0 < \\mathbf {F} (\\mathbf {X}) [ n, n ^ {\\prime} ] < 1, \\sum_ {n ^ {\\prime} \\in [ N ]} \\mathbf {F} (\\mathbf {X}) [ n, n ^ {\\prime} ] = 1. \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 114, + 398, + 480, + 433 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Definition 4.1. Given $N$ nodes $\\{\\mathbf{x}_i\\}_{i=1}^N$ , the set $\\mathcal{S}(\\mathbf{x}) = \\{\\sum_{i=1}^{N} \\lambda_i \\mathbf{x}_i \\mid \\lambda_i > 0, \\sum_i \\lambda_i = 1\\}$ is a bounded convex set.", + "bbox": [ + 89, + 439, + 483, + 487 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "According to the Definition 4.1, when the projected nodes $\\mathbf{X}\\mathbf{W}^h\\in \\mathbb{R}^{N\\times C_o}$ of each head $h$ are given, the output $\\mathbf{O}^h [n,:] = \\sum_{n'\\in [N]}\\mathbf{F}^h (\\mathbf{X})[n,n']\\mathbf{X}[n']\\mathbf{W}^h$ lies within $S(\\mathbf{X}\\mathbf{W}^h)$ , which means the outputs $\\mathbf{O}^h [n,:]$ are bounded. This property also applies to the multi-head attention mechanism.", + "bbox": [ + 89, + 493, + 483, + 585 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Proposition 4.2. If $S^h$ is a bounded convex set, $\\forall h = 1, \\dots, H$ , the set $S = S^1 + \\dots + S^H = \\{ \\mathbf{x}^1 + \\dots + \\mathbf{x}^H \\mid \\forall \\mathbf{x}^h \\in S^h \\}$ is also a bounded convex set.", + "bbox": [ + 89, + 590, + 483, + 638 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "According to Proposition 4.2, when the projected nodes $\\mathbf{X}\\mathbf{W}^h$ are given, the outputs of the multi-head attention $\\mathbf{O}[n,:] = \\sum_{h=1}^{H} \\sum_{n' \\in [N]} \\mathbf{F}^h(\\mathbf{X})[n,n']\\mathbf{X}[n']\\mathbf{W}^h$ are also bounded.", + "bbox": [ + 89, + 642, + 483, + 702 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Proposition 4.3. After applying the subspace coefficients $\\alpha$ , the outputs of the reconstructed filter subspace $\\hat{\\mathcal{F}}_{\\alpha}$ form a set $\\hat{S} = \\{\\sum_{h=1}^{H} \\sum_{i=1}^{H} \\boldsymbol{\\alpha}[h,i]\\mathbf{F}^i(\\mathbf{X})[n,n']\\mathbf{X}\\mathbf{W}^h \\mid \\mathbf{F}^i(\\mathbf{X})[n,n'] > 0, \\sum_{j} \\mathbf{F}^i(\\mathbf{X})[n,n'] = 1, \\boldsymbol{\\alpha}[h,i] \\in \\mathbb{R}\\}$ , which satisfies $S \\subseteq \\hat{S}$ .", + "bbox": [ + 88, + 708, + 482, + 787 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "According to Proposition 4.3, when applying filter subspace coefficients $\\alpha$ , we can reconstruct the filter subspace $\\mathcal{F}$ and expand the set $\\mathcal{S}$ . Specifically, given $\\mathbf{X}\\mathbf{W}^h$ , the output $\\mathbf{O}[n,:] = \\sum_{h=1}^{H} \\sum_{i=1}^{H} \\alpha[h,i]\\mathbf{F}^i(\\mathbf{X})[n,n']\\mathbf{X}\\mathbf{W}^h$ can be in the set $\\hat{\\mathcal{S}} \\setminus \\mathcal{S}$ . Our method obtains an expanded set, enhancing the expressiveness of the model. We provide the support of our analyses in Appendix A.", + "bbox": [ + 89, + 792, + 483, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Validation Experiment", + "text_level": 1, + "bbox": [ + 511, + 347, + 723, + 364 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We assess the effectiveness of the subspace coefficients using a single attention layer. Given an input graph $\\mathbf{X}$ , the output of the attention layer is calculated as $\\mathbf{O} = \\mathbf{F}(\\mathbf{X})\\mathbf{X}\\mathbf{W}_v$ , where $\\mathbf{F}(\\mathbf{X}) = \\mathrm{softmax}(\\mathbf{X}\\mathbf{W}_q\\mathbf{W}_k^\\intercal\\mathbf{X}^\\intercal)$ .", + "bbox": [ + 511, + 369, + 906, + 430 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our goal is to optimize $\\mathbf{W}_q$ , $\\mathbf{W}_k$ , $\\mathbf{W}_v$ to transform the input graph $\\mathbf{X}$ into the target graph $\\mathbf{O}$ as shown in Figure 3. The input graph contains 8 nodes $\\mathbf{X} = [[-1,1],[0,1],[1,1],[1,0],[1,-1],[0,-1],[-1,-1],[-1,0]]$ positioned in a range of $[-1,1]$ and the target graph $\\mathbf{O} = [[-1,1],[0,2],[1,1],[2,0],[1,-1],[0,-2],[-1,-1],[-2,0]]$ , with node coordinates in the range of $[-2,2]$ .", + "bbox": [ + 511, + 430, + 921, + 536 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As shown in Figure 3, by only tuning the $\\mathbf{W}_q$ and $\\mathbf{W}_k$ , i.e., $\\mathbf{F}(\\mathbf{X}) = \\mathrm{softmax}(\\mathbf{X}\\mathbf{W}_q\\mathbf{W}_k^\\top \\mathbf{X}^\\top)$ , the output is constrained inside the range of the input, which is aligned with Proposition 4.1. When we tune $\\mathbf{W}_q$ , $\\mathbf{W}_k$ , and $\\mathbf{W}_v$ , the output exceeds the input range but still cannot fully match the target graph. However, when we tune the subspace $\\mathbf{F}(\\mathbf{X})$ and subspace coefficient $\\alpha$ , the attention layer can successfully transform the input graph into the target graph as in Figure 3. This observation is aligned with Proposition 4.3, and indicates that the proposed Coeff-Tuning can effectively expand the feature space of the multi-head attention layer.", + "bbox": [ + 511, + 536, + 908, + 703 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Complexity Analysis", + "text_level": 1, + "bbox": [ + 511, + 709, + 709, + 724 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We now formally analyze the complexity of the proposed Coeff-Tuning, and compare it with other baseline PEFT methods. The number of parameters required for our method is $H \\times H$ , where $H$ is the number of heads at each attention layer. Considering $\\mathbf{W}_q, \\mathbf{W}_k, \\mathbf{W}_v \\in \\mathbb{R}^{C_i \\times C_o}$ and $\\mathbf{W}_o \\in \\mathbb{R}^{C_o \\times C_o}$ of the attention layer, our method introduces relative $\\frac{H^2}{3C_iC_o + C_o^2}$ parameters. In modern large transformer models, we typically have $C_i = 2048$ and $C_o = 1280$ . With $H = 10$ , our method requires approximately 0.001% additional parameters. Compared with other PEFT methods such as LoRA, our method introduces", + "bbox": [ + 511, + 731, + 908, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "20150", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/708fe0328f1842289aad51ed3885a124efa0037b3b92bf5b56dfbea1fe3fbb18.jpg", + "table_caption": [ + "Table 1. Fine-tuning Results on VTAB-1k with ViT" + ], + "table_footnote": [], + "table_body": "
NaturalSpecializedStructured
C100Cal101DTDF102PetsSVHNS397CameE.SATRes45ReinoCle_ctmCle_disDMLabKITTIdSlocationdS oriS_aziS_eleMeanParam.GTL_OPs
Full fine-tuning68.987.764.397.286.987.438.879.795.784.273.956.358.641.765.557.546.725.729.165.5785.845.15
Adapter74.186.163.297.787.034.650.876.388.073.170.545.737.431.253.230.325.413.822.155.820.23416.44
BitFit72.887.059.297.585.359.951.478.791.672.969.861.555.632.455.966.640.015.725.162.050.10515.08
VPT77.786.962.697.587.374.551.278.292.075.672.950.558.640.567.168.736.120.234.164.850.0616.68
LoRA67.191.469.498.890.485.354.084.995.384.473.682.969.249.881.781.848.332.844.272.910.30515.74
SPT-Adapter72.993.272.599.391.484.655.285.396.083.975.882.268.049.380.082.451.931.741.273.520.26616.44
Linear Probing63.485.063.297.086.336.651.078.587.568.674.034.330.633.255.412.520.09.619.252.94015.05
Coeff. α Only (Ours)69.590.874.599.190.983.051.183.695.384.675.857.459.146.578.675.051.425.634.269.780.00215.06
FACT-TT≤1671.389.670.798.991.087.854.685.295.583.475.782.069.049.880.079.248.434.241.473.040.04015.14
w/ Coeff-Tuning (Ours)72.189.771.899.191.288.954.885.395.784.275.383.068.851.081.380.049.335.042.773.640.04215.15
SNELL-873.792.772.499.291.489.255.484.996.186.475.284.068.553.581.082.749.933.939.274.170.26815.47
w/ Coeff-Tuning (Ours)74.292.973.199.291.690.355.685.296.286.375.884.168.853.781.384.052.135.039.974.700.27015.48
", + "bbox": [ + 93, + 108, + 906, + 289 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "relative $\\frac{H^2}{3rC_i + 5rC_o}$ parameters, where $r$ is the LoRA rank. Using the same values for $C_i$ , $C_o$ , $H$ , and setting $r = 4$ , our approach requires less than $0.2\\%$ additional parameters compared to LoRA and its variants. This also indicates that our Coeff-Tuning can maintain almost the same quantity of parameters when combined with other PEFT methods.", + "bbox": [ + 89, + 313, + 483, + 407 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 89, + 421, + 225, + 439 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we validate the proposed Coeff-Tuning with various tasks and datasets, including tuning the vision transformer [11] on few-shot image classification, tuning text-to-image diffusion model [51] for concept customization, and tuning vision-language BART [8] on image-to-text understanding. Combining with other weight-based PEFT methods, the proposed Coeff-Tuning consistently achieves superior results with negligible parameters added.", + "bbox": [ + 89, + 446, + 483, + 569 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Few-shot Image Classification", + "text_level": 1, + "bbox": [ + 89, + 580, + 356, + 595 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "we first validate the proposed method by fine-tuning Vision Transformers on a series of downstream datasets for few-shot image classification. Our method achieves superior results while maintaining a high parameter efficiency compared with baseline methods.", + "bbox": [ + 89, + 602, + 483, + 679 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Experimental Settings. We utilize the VTAB-1k [81] dataset which comprises 19 diverse computer vision tasks across three domains: standard natural imagery, specialized capture, and synthetic/structured environments. The benchmark evaluates model performance on various task", + "bbox": [ + 89, + 686, + 483, + 763 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/e7580b354df23c7052215a6949f33da330de507f399d2e8b1cd4b49a3180fcfa.jpg", + "table_caption": [ + "Table 2. Fine-tuning Results on the FGVC with ViT." + ], + "table_footnote": [], + "table_body": "
CUB-200NABirdsFlowersS-DogsS-CarsMean
Full fine-tuning87.382.798.889.484.588.5
Linear Probing85.375.997.986.251.379.3
Adapter87.384.398.488.868.485.5
BitFit88.484.298.891.279.488.4
VPT86.778.898.490.768.784.6
LoRA84.979.098.188.179.886.0
SPT-LoRA88.683.499.591.487.390.1
SNELL-889.686.899.392.189.991.5
w/ Coeff-Tuning (Ours)89.887.199.592.290.791.9
", + "bbox": [ + 83, + 781, + 475, + 898 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "objectives. With only 1,000 training samples per task, it presents a significant few-shot learning challenge. We also benchmark our method on five fine-grained datasets in few-shot settings, including CUB-200-2011 [69], NABirds [67], Oxford Flowers [44], Stanford Dogs [25] and Stanford Cars [13]. A detailed description is provided in Appendix B.2.", + "bbox": [ + 511, + 314, + 906, + 421 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As for the vision transformer, we follow VPT [23], SSF [35], and select ViT-B/16 [11] pre-trained on ImageNet-21K [9] as the base model for fine-tuning. We benchmark the proposed Coeff-Tuning with a series of PEFT methods for few-shot learning including SSF [35], VPT [23], Adapter [20], LoRA [21], and BitFit [79].", + "bbox": [ + 511, + 422, + 906, + 513 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We present two settings of Coeff-Tuning in this task. We first explore the representation ability of tuning subspace coefficient $\\alpha$ only, which tunes the transformer bottleneck with only tuning $H^2$ (144 for ViT [11]) parameters for each transformer block. We also show the performance of combining Coeff-Tuning and SSF [35] to demonstrate that our method can plug and play with other PEFT methods. We set the coefficient dropout rates as $p = 0$ and $p = 0.1$ for these two settings respectively, and provide other hyperparameters in Appendix B.2.", + "bbox": [ + 511, + 515, + 908, + 667 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Results. As shown in Table 1, the proposed Coeff-Tuning shows superior performances in both settings. Compared with the linear probing baseline which achieves a mean accuracy of 61.69, Coeff-Tuning obtains a significantly higher result of 73.49 with tuning only $(12 * 12) * 12 = 1$ , 728 additional subspace coefficients $\\alpha$ in ViT [11] (ViT [11] has 12 attention layers, each attention has 12 heads). Thus, we achieve a large performance increase in few-shot image classification $(11.8^{\\uparrow})$ with negligible additional parameter costs. Note that this result has already surpassed almost all baselines. We further achieve the best performance by combining Coeff-Tuning with SSF while maintaining almost the same parameter cost.", + "bbox": [ + 511, + 688, + 908, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "20151", + "bbox": [ + 478, + 944, + 517, + 957 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/2941f246d25d534a5c8707518a81d42e9c918146ed65822cd22a66b4a3805b00.jpg", + "image_caption": [ + "Figure 4. Results on Concept Customization. Compared to the baseline method (middle row), Coeff-Tuning (bottom row) leads to higher concept fidelity and text alignment of the generated images. For instance, our method accurately generates images of a “ $\\langle V \\rangle$ plushie sloth in the snow,” while the baseline fails to align with the text prompt. Additionally, our method generates a “ $\\langle V \\rangle$ wooden pot” with higher fidelity to the input image." + ], + "image_footnote": [], + "bbox": [ + 133, + 89, + 859, + 388 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/69606d837a52a1f2b97bdb5aad104845b812ee214cca357fdbe320bed1e46b85.jpg", + "table_caption": [ + "Table 3. Concept Customization Results with SDXL." + ], + "table_footnote": [], + "table_body": "
Method# Params (%)Text Align.Concept Align.Diversity
DoRA0.03710.330.7310.9
Ours + DoRA0.03740.340.7511.37
LoRA0.03060.330.7410.08
Ours + LoRA0.03090.340.7510.45
", + "bbox": [ + 94, + 486, + 483, + 563 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Personalized Text-to-Image Generation", + "text_level": 1, + "bbox": [ + 89, + 575, + 428, + 593 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We now show the effectiveness of Coeff-Tuning with the concept customization with Stable Diffusion (SDXL) [51, 55]. We validate the effectiveness of Coeff-Tuning by combining it with baseline PEFT methods, LoRA [21] and DoRA [38], and compare the performance with baselines without tuning $\\alpha$ . Specifically, we fine-tune SDXL [51] to associate the special token with a concept image from Dreambooth [56] and CustomConcept101 [30] datasets. We choose 10 different concepts, and fine-tune the model on $4 \\sim 9$ images. We utilize 25 different text prompts following [53] and generate 4 images for each prompt. We evaluate the fidelity, diversity, and text-to-image alignment of the generated images. Other training details are described in Appendix B.3.", + "bbox": [ + 89, + 599, + 483, + 811 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Evaluation metrics. We assess the concept alignment with the average cosine similarity between CLIP embeddings [54] of the generated images and the concept images. The diversity of generated images is measured by the Vendi score [12] calculated with the DINOv2 embeddings [46].", + "bbox": [ + 89, + 824, + 483, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The text alignment of the generated images is measured via average cosine similarity in the CLIP feature space [54].", + "bbox": [ + 511, + 469, + 906, + 500 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results. The quantitative results are shown in Table 3. By combining with LoRA [21] or DoRA [38], the proposed Coeff-Tuning generates images with greater diversity and better alignment with the both text prompt and input concept images. The qualitative results of our method compared with LoRA [21] are shown in Figure 4. The generated images with Coeff-Tuning have higher concept fidelity, preserving more characteristics from the input images.", + "bbox": [ + 511, + 506, + 906, + 627 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As an example, consider the concept “ $\\langle V \\rangle$ cat” as shown in the 4th column in Figure 4, the image generated using our method retains the cat's color from the input image, while the one generated by the baseline method alter the color. Additionally, the backgrounds in images generated with our method show more detailed features compared to those from the baseline method. Coeff-Tuning achieves the above improvements by introducing a limited number of additional parameters. We provide more generated images in Appendix B.3.", + "bbox": [ + 511, + 627, + 908, + 779 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Image-Text Understanding", + "text_level": 1, + "bbox": [ + 511, + 787, + 756, + 804 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As we have shown the effectiveness of Coeff-Tuning in fine-tuning models on both few-shot image classification and personalized image generation, we now validate it on multitask image-text understanding, which is a much larger-scale multi-modal fine-tuning task that requires a higher model capacity. By combining with baseline PEFT meth", + "bbox": [ + 511, + 810, + 906, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "20152", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/40a839bea6f4f9035dde5b929ad5dbb8b30ca1b9489e8cb430374fea293d4eed.jpg", + "table_caption": [ + "Table 4. Multi-Model Fine-tuning Results with VL-Bart." + ], + "table_footnote": [], + "table_body": "
Method# Params (%)VQAV2GQANVLR2COCO CapAvg.
Full Model10066.956.773.7112.077.3
LoRA5.9465.253.671.9115.376.5
Ours + LoRA5.9466.254.173.0115.277.1
DoRA5.9665.854.773.1115.977.4
Ours + DoRA5.9666.454.973.0116.477.7
", + "bbox": [ + 253, + 108, + 736, + 215 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ods, our method again achieves superior performance with negligible parameter increase.", + "bbox": [ + 89, + 232, + 482, + 263 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Experimental Settings. We evaluate our method on VL-BART [8] multi-modal transformer, which utilizes CLIP-ResNet101 [54] for vision encoding and sends the visual features to BART-Base [31] for image-text joint modeling. Following [65], we evaluate our method with the multitask fine-tuning of VL-Bart [8] on four large-scale image-text understanding tasks simultaneously: VQAv2 [14] and GQA [22] for visual question answering, NLVR2 [64] for visual reasoning, and MSCOCO [5] for image captioning. Following the plug-and-play practice, we combine the proposed Coeff-Tuning with LoRA [21] and DoRA [38] respectively, and set the subspace coefficient dropout rate to $p = 0.2$ . Detailed descriptions of other settings and hyperparameters are provided in Appendix B.4.", + "bbox": [ + 89, + 290, + 483, + 503 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Results. As shown in Table 4, our method surpasses baseline LoRA [21] and DoRA [38] respectively in almost all image-text datasets consistently, and achieves higher average scores. The Coeff-Tuning combined with LoRA [21] obtains comparable results with full-model fine-tuning, and it achieves better results than full-model tuning when it is combined with DoRA [38]. Notably, these improvements are achieved while maintaining the same parameter budgets as we only tune a hundred parameters for each attention.", + "bbox": [ + 89, + 523, + 483, + 660 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.4. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 674, + 243, + 690 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this section, we examine the effectiveness of each component of our design with ablation experiments.", + "bbox": [ + 89, + 698, + 482, + 729 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. Ablation studies on (a) residual parameterization and (b) coefficient dropout showing their effects on model performance.", + "bbox": [ + 89, + 755, + 482, + 784 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/64ca2f7436c2f55894d7e53c545b86d4dae545183619cb03d813627467ce15d4.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ParametrizationAvg.
α w/ rand. init.44.28
α w/ identity init.61.57
I + α, zero init.73.49
", + "bbox": [ + 112, + 808, + 294, + 878 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/e7fc21c7e75b61732e582127804b9055368fa67c24e2848449710fedc543b72b.jpg", + "table_caption": [ + "(a)" + ], + "table_footnote": [], + "table_body": "
Dropout pAvg. Score
0.075.1
0.177.3
0.277.7
0.472.3
", + "bbox": [ + 308, + 795, + 478, + 880 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "(b)", + "bbox": [ + 377, + 881, + 395, + 892 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Residual Parameterization. We first demonstrate the importance of residual parameterization in Eq. 11 in Coeff-Tuning. We compare our residual parameterization with the direct parameterization where we directly tune $\\alpha$ without adding it to the identity matrix, and initialize it with either a random matrix or an identity matrix. We adopt the few-shot tuning task as in Sec. 5.1 and only tune $\\alpha$ . Results in Table 5a show that models with residual parameterization achieve superior performance than direct parameterization with two different initializations. This validates our design that maintaining the original filter subspace while learning the residual combination helps preserve pre-trained knowledge and leads to better fine-tuning performance.", + "bbox": [ + 511, + 232, + 906, + 429 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Subspace Coefficient Dropout. The dropout regularization on the subspace coefficient as in Eq. 12 plays a crucial role in preventing overfitting. We adopt the experiment setting in multi-modal fine-tuning as in Sec. 5.3. We experiment with different dropout rates $p \\in \\{0, 0.1, 0.2, 0.4\\}$ during training. The results in Table 5b demonstrate that a moderate dropout rate ( $p = 0.2$ ) yields the best performance while either a lower or higher dropout rate leads to sub-optimal results. This suggests that applying dropout directly on the tuned subspace coefficient $\\alpha$ can effectively prevent overfitting and lead to better generalization.", + "bbox": [ + 511, + 440, + 908, + 608 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 619, + 633, + 636 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we presented Coeff-Tuning, a new parameter-efficient fine-tuning approach that enhances multi-head attention's expressiveness by learning a set of subspace coefficients. By reformulating attention as a graph convolution with a filter subspace, we demonstrated both theoretically and empirically that our method expands the feature space of multi-head attention beyond the original convex hull constraints. The proposed Coeff-Tuning can be seamlessly integrated with existing PEFT methods while introducing negligible additional parameters. Through extensive experiments across diverse tasks and modalities, we consistently demonstrated superior performance compared to competitive PEFT baselines. Our method achieved significant improvements in various metrics while maintaining high parameter efficiency, suggesting a promising direction for enhancing the adaptability of large pre-trained transformers.", + "bbox": [ + 511, + 645, + 908, + 888 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "20153", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Tom B Brown. Language models are few-shot learners. arXiv preprint arXiv:2005.14165, 2020. 2", + "[2] Shoufa Chen, Chongjian Ge, Zhan Tong, Jiangliu Wang, Yibing Song, Jue Wang, and Ping Luo. Adaptformer: Adapting vision transformers for scalable visual recognition. Advances in Neural Information Processing Systems, 2022. 2", + "[3] Wei Chen, Zichen Miao, and Qiang Qiu. Inner product-based neural network similarity. Advances in Neural Information Processing Systems, 36, 2024. 1", + "[4] Wei Chen, Zichen Miao, and Qiang Qiu. Large convolutional model tuning via filter subspace. arXiv preprint arXiv:2403.00269, 2024.2, 1", + "[5] Xinlei Chen, Hao Fang, Tsung-Yi Lin, Ramakrishna Vedantam, Saurabh Gupta, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco captions: Data collection and evaluation server. arXiv preprint arXiv:1504.00325, 2015. 8", + "[6] Xiuyuan Cheng, Zichen Miao, and Qiang Qiu. Graph convolution with low-rank learnable local filters. arXiv preprint arXiv:2008.01818, 2020. 4", + "[7] Xiuyuan Cheng, Zichen Miao, and Qiang Qiu. Graph convolution with low-rank learnable local filters. In International Conference on Learning Representations, 2021. 1", + "[8] Jaemin Cho, Jie Lei, Hao Tan, and Mohit Bansal. Unifying vision-and-language tasks via text generation. In International Conference on Machine Learning, pages 1931-1942. PMLR, 2021. 2, 6, 8", + "[9] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 6, 1", + "[10] Yihe Dong, Jean-Baptiste Cordonnier, and Andreas Loukas. Attention is not all you need: Pure attention loses rank doubly exponentially with depth. In International Conference on Machine Learning, pages 2793-2803. PMLR, 2021. 2, 3", + "[11] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. ICLR, 2021. 1, 2, 6", + "[12] Dan Friedman and Adji Bousso Dieng. The vendi score: A diversity evaluation metric for machine learning. arXiv preprint arXiv:2210.02410, 2022. 7", + "[13] Timnit Gebru, Jonathan Krause, Yilun Wang, Duyun Chen, Jia Deng, and Li Fei-Fei. Fine-grained car detection for visual census estimation. In Proceedings of the AAAI Conference on Artificial Intelligence, 2017. 6", + "[14] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Bartra, and Devi Parikh. Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6904-6913, 2017. 2, 8", + "[15] Ligong Han, Yinxiao Li, Han Zhang, Peyman Milanfar, Dimitris Metaxas, and Feng Yang. Svdiff: Compact parameter space for diffusion fine-tuning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7323-7334, 2023. 1", + "[16] Liu He and Daniel Aliaga. Globalmapper: Arbitrary-shaped urban layout generation. In Proceedings of the IEEE/CVF" + ], + "bbox": [ + 93, + 114, + 482, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "International Conference on Computer Vision, pages 454-464, 2023. 1", + "[17] Liu He and Daniel Aliaga. Coho: Context-sensitive city-scale hierarchical urban layout generation. In European Conference on Computer Vision, pages 1-18. Springer, 2024.", + "[18] Liu He, Yijuan Lu, John Corring, Dinei Florencio, and Cha Zhang. Diffusion-based document layout generation. In International Conference on Document Analysis and Recognition, pages 361-378. Springer, 2023.", + "[19] Liu He, Yizhi Song, Hejun Huang, Daniel Aliaga, and Xin Zhou. Kubrick: Multimodal agent collaborations for synthetic video generation. arXiv preprint arXiv:2408.10453, 2024. 1", + "[20] Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin De Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. Parameter-efficient transfer learning for nlp. In International conference on machine learning, pages 2790-2799. PMLR, 2019. 2, 6", + "[21] Edward J Hu, Phillip Wallis, Zeyuan Allen-Zhu, Yuzhhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. In International Conference on Learning Representations, 2021. 1, 2, 6, 7, 8", + "[22] Drew A Hudson and Christopher D Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6700-6709, 2019. 2, 8", + "[23] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In European Conference on Computer Vision, pages 709-727. Springer, 2022. 2, 6, 1", + "[24] Rabeeh Karimi Mahabadi, James Henderson, and Sebastian Ruder. Compacter: Efficient low-rank hypercomplex adapter layers. Advances in Neural Information Processing Systems, 2021. 2, 1", + "[25] Aditya Khosla, Nityananda Jayadevaprakash, Bangpeng Yao, and Fei-Fei Li. Novel dataset for fine-grained image categorization: Stanford dogs. In Proc. CVPR workshop on fine-grained visual categorization (FGVC), 2011. 6", + "[26] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In International Conference on Learning Representations, 2015. 1", + "[27] Thomas N Kipf and Max Welling. Semi-supervised classification with graph convolutional networks. arXiv preprint arXiv:1609.02907, 2016. 4", + "[28] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 1", + "[29] Dawid J Kopiczko, Tijmen Blankevoort, and Yuki M Asano. Vera: Vector-based random matrix adaptation. *ICLR*, 2024. 2", + "[30] Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization of text-to-image diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 7, 1" + ], + "bbox": [ + 517, + 93, + 903, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "20154", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[31] M Lewis. Bart: Denoising sequence-to-sequence pretraining for natural language generation, translation, and comprehension. arXiv preprint arXiv:1910.13461, 2019. 2, 8", + "[32] Junbo Li, Zichen Miao, Qiang Qiu, and Ruqi Zhang. Training bayesian neural networks with sparse subspace variational inference. arXiv preprint arXiv:2402.11025, 2024. 1", + "[33] Tianchun Li, Chengxiang Wu, Pengyi Shi, and Xiaoqian Wang. Cumulative difference learning vae for time-series with temporally correlated inflow-outflow. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 13619-13627, 2024. 1", + "[34] Xiang Lisa Li and Percy Liang. Prefix-tuning: Optimizing continuous prompts for generation. In Proceedings of the Association for Computational Linguistics, 2021. 2", + "[35] Dongze Lian, Daquan Zhou, Jiashi Feng, and Xinchao Wang. Scaling & shifting your features: A new baseline for efficient model tuning. Advances in Neural Information Processing Systems, 35:109-123, 2022. 2, 6, 1", + "[36] Haokun Liu, Derek Tam, Mohammed Muqeeth, Jay Mohta, Tenghao Huang, Mohit Bansal, and Colin A Raffel. Few-shot parameter-efficient fine-tuning is better and cheaper than in-context learning. Advances in Neural Information Processing Systems, 35:1950-1965, 2022. 2", + "[37] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 2", + "[38] Shih-Yang Liu, Chien-Yi Wang, Hongxu Yin, Pavlo Molchanov, Yu-Chiang Frank Wang, Kwang-Ting Cheng, and Min-Hung Chen. Dora: Weight-decomposed low-rank adaptation. arXiv preprint arXiv:2402.09353, 2024. 2, 7, 8, 1", + "[39] Tianci Liu, Ruirui Li, Yunzhe Qi, Hui Liu, Xianfeng Tang, Tianqi Zheng, Qingyu Yin, Monica Xiao Cheng, Jun Huan, Haoyu Wang, et al. Unlocking efficient, scalable, and continual knowledge editing with basis-level representation finetuning. arXiv preprint arXiv:2503.00306, 2025. 1", + "[40] Zichen Miao, Ze Wang, Wei Chen, and Qiang Qiu. Continual learning with filter atom swapping. In International Conference on Learning Representations, 2021. 1", + "[41] Zichen Miao, Ze Wang, Xiuyuan Cheng, and Qiang Qiu. Spatiotemporal joint filter decomposition in 3d convolutional neural networks. Advances in Neural Information Processing Systems, 34:3376-3388, 2021. 1", + "[42] Zichen Miao, Jiang Wang, Ze Wang, Zhengyuan Yang, Lijuan Wang, Qiang Qiu, and Zicheng Liu. Training diffusion models towards diverse image generation with reinforcement learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10844-10853, 2024. 1", + "[43] Zichen Miao, Zhengyuan Yang, Kevin Lin, Ze Wang, Zicheng Liu, Lijuan Wang, and Qiang Qiu. Tuning timestep-distilled diffusion model using pairwise sample optimization. arXiv preprint arXiv:2410.03190, 2024. 1", + "[44] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In 2008 Sixth Indian conference on computer vision, graphics & image processing, pages 722-729. IEEE, 2008. 6" + ], + "bbox": [ + 91, + 92, + 480, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[45] Lorenzo Noci, Sotiris Anagnostidis, Luca Biggio, Antonio Orvieto, Sidak Pal Singh, and Aurelien Lucchi. Signal propagation in transformers: Theoretical perspectives and the role of rank collapse. Advances in Neural Information Processing Systems, 35:27198-27211, 2022. 3", + "[46] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 7", + "[47] Gaurav Patel and Qiang Qiu. Learning to unlearn while retaining: Combating gradient conflicts in machine unlearning. arXiv preprint arXiv:2503.06339, 2025. 1", + "[48] Gaurav Patel, Konda Reddy Mopuri, and Qiang Qiu. Learning to retain while acquiring: Combating distribution-shift in adversarial data-free knowledge distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7786-7794, 2023.", + "[49] Gaurav Patel, Christopher Sandino, Behrooz Mahasseni, Ellen L Zippi, Erdrin Azemi, Ali Moin, and Juri Minxha. Efficient source-free time-series adaptation via parameter subspace disentanglement. arXiv preprint arXiv:2410.02147, 2024. 1", + "[50] Jonas Pfeiffer, Aishwarya Kamath, Andreas Rücklé, Kyunghyun Cho, and Iryna Gurevych. Adapterfusion: Nondestructive task composition for transfer learning. arXiv preprint arXiv:2005.00247, 2020. 2", + "[51] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. *ICLR*, 2024. 1, 6, 7", + "[52] Qiang Qiu, Xiuyuan Cheng, Guillermo Sapiro, et al. Dcfnet: Deep neural network with decomposed convolutional filters. In International Conference on Machine Learning, pages 4198-4207. PMLR, 2018. 1", + "[53] Zeju Qiu, Weiyang Liu, Haiwen Feng, Yuxuan Xue, Yao Feng, Zhen Liu, Dan Zhang, Adrian Weller, and Bernhard Scholkopf. Controlling text-to-image diffusion by orthogonal finetuning. Advances in Neural Information Processing Systems, 36:79320-79362, 2023. 2, 7", + "[54] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, 2021. 2, 7, 8", + "[55] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10684-10695, 2022. 1, 2, 7", + "[56] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 22500-22510, 2023. 2, 7, 1", + "[57] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy," + ], + "bbox": [ + 516, + 92, + 903, + 902 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "20155", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Aditya Khosla, Michael Bernstein, Alexander C. Berg, and Li Fei-Fei. ImageNet Large Scale Visual Recognition Challenge. International Journal of Computer Vision (IJCV), 2015. 2", + "[58] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278-25294, 2022. 2", + "[59] Han Shi, Jiahui Gao, Hang Xu, Xiaodan Liang, Zhenguo Li, Lingpeng Kong, Stephen Lee, and James T Kwok. Revisiting over-smoothing in bert from the perspective of graph. arXiv preprint arXiv:2202.08625, 2022. 2, 3", + "[60] Kihyuk Sohn, Huiwen Chang, José Lezama, Luisa Polania, Han Zhang, Yuan Hao, Irfan Essa, and Lu Jiang. Visual prompt tuning for generative transfer learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 2", + "[61] Yizhi Song, Liu He, Zhifei Zhang, Soo Ye Kim, He Zhang, Wei Xiong, Zhe Lin, Brian Price, Scott Cohen, Jianming Zhang, et al. Refine-by-align: Reference-guided artifacts refinement through semantic alignment. arXiv preprint arXiv:2412.00306, 2024. 1", + "[62] Yizhi Song, Zhifei Zhang, Zhe Lin, Scott Cohen, Brian Price, Jianming Zhang, Soo Ye Kim, He Zhang, Wei Xiong, and Daniel Aliaga. Imprint: Generative object compositing by learning identity-preserving representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8048-8058, 2024. 1", + "[63] Nitish Srivastava, Geoffrey Hinton, Alex Krizhevsky, Ilya Sutskever, and Ruslan Salakhutdinov. Dropout: a simple way to prevent neural networks from overfitting. The journal of machine learning research, 15(1):1929-1958, 2014. 4", + "[64] Alane Suhr, Stephanie Zhou, Ally Zhang, Iris Zhang, Hua-jun Bai, and Yoav Artzi. A corpus for reasoning about natural language grounded in photographs. arXiv preprint arXiv:1811.00491, 2018. 8", + "[65] Yi-Lin Sung, Jaemin Cho, and Mohit Bansal. Vl-adapter: Parameter-efficient transfer learning for vision-and-language tasks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5227-5237, 2022. 8", + "[66] Gemma Canet Tarres, Zhe Lin, Zhifei Zhang, Jianming Zhang, Yizhi Song, Dan Ruta, Andrew Gilbert, John Colomosse, and Soo Ye Kim. Thinking outside the bbox: Unconstrained generative object compositing. arXiv preprint arXiv:2409.04559, 2024. 1", + "[67] Grant Van Horn, Steve Branson, Ryan Farrell, Scott Haber, Jessie Barry, Panos Ipeirotis, Pietro Perona, and Serge Belongie. Building a bird recognition app and large scale dataset with citizen scientists: The fine print in fine-grained dataset collection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 595-604, 2015. 6", + "[68] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 1, 2" + ], + "bbox": [ + 91, + 90, + 483, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[69] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The caltech-ucsd birds-200-2011 dataset. 2011. 6", + "[70] Haoyu Wang, Tianci Liu, Ruirui Li, Monica Cheng, Tuo Zhao, and Jing Gao. Roselora: Row and column-wise sparse low-rank adaptation of pre-trained language model for knowledge editing and fine-tuning. arXiv preprint arXiv:2406.10777, 2024. 1", + "[71] Peihao Wang, Wenqing Zheng, Tianlong Chen, and Zhangyang Wang. Anti-oversmoothing in deep vision transformers via the fourier domain analysis: From theory to practice. arXiv preprint arXiv:2203.05962, 2022. 2", + "[72] Ze Wang, Xiuyuan Cheng, Guillermo Sapiro, and Qiang Qiu. Stochastic conditional generative networks with basis decomposition. arXiv preprint arXiv:1909.11286, 2019. 1", + "[73] Ze Wang, Seunghyun Hwang, Zichen Miao, and Qiang Qiu. Image generation using continuous filter atoms. Advances in Neural Information Processing Systems, 34:17826-17838, 2021.", + "[74] Ze Wang, Zichen Miao, Jun Hu, and Qiang Qiu. Adaptive convolutions with per-pixel dynamic filter atom. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12302-12311, 2021. 1", + "[75] Ze Wang, Zichen Miao, Xiantong Zhen, and Qiang Qiu. Learning to learn dense gaussian processes for few-shot learning. Advances in Neural Information Processing Systems, 34:13230-13241, 2021. 1", + "[76] Enze Xie, Lewei Yao, Han Shi, Zhili Liu, Daquan Zhou, Zhaoqiang Liu, Jiawei Li, and Zhenguo Li. Difffit: Unlocking transferability of large diffusion models via simple parameter-efficient fine-tuning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4230-4239, 2023. 1", + "[77] Zhexiao Xiong, Wei Xiong, Jing Shi, He Zhang, Yizhi Song, and Nathan Jacobs. Groundingbooth: Grounding text-to-image customization. arXiv preprint arXiv:2409.08520, 2024.1", + "[78] SHIH-YING YEH, Yu-Guan Hsieh, Zhidong Gao, Bernard BW Yang, Gigeong Oh, and Yanmin Gong. Navigating text-to-image customization: From lycoris fine-tuning to model evaluation. In International Conference on Learning Representations, 2023. 2", + "[79] Elad Ben Zaken, Yoav Goldberg, and Shauli Ravfogel. Bitfit: Simple parameter-efficient fine-tuning for transformer-based masked language-models. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics, 2022. 2, 6", + "[80] Shuangfei Zhai, Tatiana Likhomanenko, Etai Littwin, Dan Busbridge, Jason Ramapuram, Yizhe Zhang, Jiatao Gu, and Joshua M Susskind. Stabilizing transformer training by preventing attention entropy collapse. In International Conference on Machine Learning, pages 40770-40803. PMLR, 2023. 2, 3", + "[81] Xiaohua Zhai, Joan Puigcerver, Alexander Kolesnikov, Pierre Ruyssen, Carlos Riquelme, Mario Lucic, Josip Djolonga, Andre Susano Pinto, Maxim Neumann, Alexey Dosovitskiy, et al. A large-scale study of representation learning" + ], + "bbox": [ + 516, + 90, + 903, + 900 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "20156", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "with the visual task adaptation benchmark. arXiv preprint arXiv:1910.04867, 2019. 6, 1", + "[82] Qingru Zhang, Minshuo Chen, Alexander Bukharin, Nikos Karampatziakis, Pengcheng He, Yu Cheng, Weizhu Chen, and Tuo Zhao. Adalora: Adaptive budget allocation for parameter-efficient fine-tuning. *ICLR*, 2022. 2", + "[83] Yibo Zhong, Haoxiang Jiang, Lincan Li, Ryumei Nakada, Tianci Liu, Linjun Zhang, Huaxiu Yao, and Haoyu Wang. Neat: Nonlinear parameter-efficient adaptation of pretrained models. arXiv preprint arXiv:2410.01870, 2024. 1", + "[84] Daquan Zhou, Bingyi Kang, Xiaojie Jin, Linjie Yang, Xiaochen Lian, Zihang Jiang, Qibin Hou, and Jiashi Feng. Deepvit: Towards deeper vision transformer. arXiv preprint arXiv:2103.11886, 2021. 2, 3" + ], + "bbox": [ + 91, + 90, + 483, + 289 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "20157", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2025/Coeff-Tuning_ A Graph Filter Subspace View for Tuning Attention-Based Large Models/e7a23930-dc20-4352-a4dc-e2a2616c610f_model.json b/2025/Coeff-Tuning_ A Graph Filter Subspace View for Tuning Attention-Based Large Models/e7a23930-dc20-4352-a4dc-e2a2616c610f_model.json new file mode 100644 index 0000000000000000000000000000000000000000..896b266afdfa6959dc760ed0491e5bd4306d1a52 --- /dev/null +++ b/2025/Coeff-Tuning_ A Graph Filter Subspace View for Tuning Attention-Based Large Models/e7a23930-dc20-4352-a4dc-e2a2616c610f_model.json @@ -0,0 +1,2732 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.097, + 0.13, + 0.9, + 0.175 + ], + "angle": 0, + "content": "Coeff-Tuning: A Graph Filter Subspace View for Tuning Attention-Based Large Models" + }, + { + "type": "text", + "bbox": [ + 0.348, + 0.203, + 0.648, + 0.24 + ], + "angle": 0, + "content": "Zichen Miao†, Wei Chen†, Qiang Qiu Purdue University, IN, USA" + }, + { + "type": "text", + "bbox": [ + 0.348, + 0.242, + 0.648, + 0.258 + ], + "angle": 0, + "content": "{miaoz, chen2732, qqiu}@purdue.edu" + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.292, + 0.327, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.324, + 0.486, + 0.749 + ], + "angle": 0, + "content": "Transformer-based large pre-trained models have shown remarkable generalization ability, and various parameter-efficient fine-tuning (PEFT) methods have been proposed to customize these models on downstream tasks with minimal computational and memory budgets. Previous PEFT methods are primarily designed from a tensor-decomposition perspective that tries to effectively tune the linear transformation by finding the smallest subset of parameters to train. Our study adopts an orthogonal view by representing the attention operation as a graph convolution and formulating the multi-head attention maps as a convolutional filter subspace, with each attention map as a subspace element. In this paper, we propose to tune the large pre-trained transformers by learning a small set of combination coefficients that construct a more expressive filter subspace from the original multi-head attention maps. We show analytically and experimentally that the tuned filter subspace can effectively expand the feature space of the multi-head attention and further enhance the capacity of transformers. We further stabilize the fine-tuning with a residual parameterization of the tunable subspace coefficients, and enhance the generalization with a regularization design by directly applying dropout on the tunable coefficient during training. The tunable coefficients take a tiny number of parameters and can be combined with previous PEFT methods in a plug-and-play manner. Extensive experiments show that our approach achieves superior performances than PEFT baselines with neglectable additional parameters.\\(^{1}\\)" + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.778, + 0.224, + 0.794 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.805, + 0.484, + 0.851 + ], + "angle": 0, + "content": "Large foundational models have demonstrated remarkable performance and generalization ability across a wide range of natural language processing and computer vision" + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.291, + 0.718, + 0.416 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.725, + 0.297, + 0.896, + 0.414 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.55, + 0.422, + 0.824, + 0.435 + ], + "angle": 0, + "content": "A [V] plushie panda on top of a purple rug in a forest." + }, + { + "type": "image", + "bbox": [ + 0.539, + 0.436, + 0.689, + 0.552 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.708, + 0.436, + 0.857, + 0.551 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.614, + 0.554, + 0.797, + 0.566 + ], + "angle": 0, + "content": "A [V] plushie panda on in the snow." + }, + { + "type": "image", + "bbox": [ + 0.541, + 0.568, + 0.691, + 0.685 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.595, + 0.689, + 0.635, + 0.7 + ], + "angle": 0, + "content": "LoRA" + }, + { + "type": "image", + "bbox": [ + 0.708, + 0.568, + 0.857, + 0.684 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.714, + 0.689, + 0.845, + 0.702 + ], + "angle": 0, + "content": "LoRA + Coeff-Tuning" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.72, + 0.907, + 0.817 + ], + "angle": 0, + "content": "Figure 1. (Up) Under the graph convolution view of the attention, the proposed Coeff-Tuning enhances the expressiveness by combining the attention maps with a tiny tunable subspace coefficient \\(\\alpha\\). It can be seamlessly integrated with other PEFT methods, e.g., LoRA [21], by introducing negligible additional parameters. (Bottom) Combining Coeff-Tuning with LoRA achieves better results in the concept customization task with SDXL [51]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.909, + 0.903 + ], + "angle": 0, + "content": "tasks [11, 28, 55, 68]. The success of these models can be attributed to their ability to learn rich, transferable representations by leveraging the power of the transformer architecture [68] and the extensive knowledge captured during" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.863, + 0.22, + 0.876 + ], + "angle": 0, + "content": "\\(\\dagger\\) Equal contribution." + }, + { + "type": "page_footnote", + "bbox": [ + 0.093, + 0.876, + 0.482, + 0.901 + ], + "angle": 0, + "content": "1Code is available at: https://github.com/ZichenMiao/Coeff_Tuning" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.863, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "20146" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.152 + ], + "angle": 0, + "content": "pre-training on large-scale datasets [1, 58]. However, fine-tuning these pre-trained transformers on downstream tasks with full-parameter updates can be challenging due to the high computational and memory costs." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.155, + 0.483, + 0.334 + ], + "angle": 0, + "content": "To address this, various parameter-efficient fine-tuning (PEFT) methods have been developed to adapt pre-trained models on downstream tasks while minimizing the number of trainable parameters [21, 38, 53, 79]. A representative family of these approaches focuses on tuning the linear transformation layers in transformers, where they apply a lightweight parameterization, e.g., low-rank parameterization [21, 38], to adapt the weight matrices. However, none of these methods are designed from a holistic view of the core operation in transformers, the multi-head attention layer [68], which plays a key role in the expressiveness of transformers [10, 59]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.337, + 0.483, + 0.488 + ], + "angle": 0, + "content": "In this paper, we propose a new PEFT method tailored to enhance the multi-head attention's expressiveness by tuning a minimal set of parameters. We first formulate the multi-head attention as a graph convolution with a filter subspace, with each graph filter as the attention map of each attention head. We then tune the filter subspace by simply learning a small set of subspace coefficients that linearly combine the pre-trained attention maps. In this way, we adapt for each attention head a new task-adaptive attention map that better captures the token-wise correlation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.49, + 0.483, + 0.732 + ], + "angle": 0, + "content": "We attribute the effectiveness of our approach to its ability to expand the feature space of the original multi-head attention layer. By applying the softmax function to the query-key correlation matrices, all elements in a row of the original attention map are non-negative and have a sum of one [68]. This limits the expressiveness of the multi-head attention as the output token embeddings lie within the convex hull of the value embeddings. Our method constructs a new filter subspace with non-constrained coefficients as shown in Figure 1, allowing negative elements learned in the combined attention map to capture more diverse token-wise relationships. This also enables the output tokens to obtain representation outside the convex hull of the value embeddings, which further enhances the capacity of the multi-head attention. We provide analysis and an illustrative toy example to better demonstrate this point." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.733, + 0.483, + 0.824 + ], + "angle": 0, + "content": "We show that our approach can be seamlessly combined with previous PEFT methods in a plug-and-play manner. As the filter subspace in each attention layer is spanned by a tiny number of attention maps (usually \\(10\\sim 30\\)), we only introduce neglectable additional tunable parameters and training costs over the base PEFT methods." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.483, + 0.901 + ], + "angle": 0, + "content": "We validate our approach on a large variety of fine-tuning tasks and datasets, along with pre-trained transformers of different scales and modalities. Specifically, we first show the effectiveness of our method with ViT [11] on the few-shot image classification fine-tuning. We also" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.167 + ], + "angle": 0, + "content": "show our method can achieve superior performance with vision-language transformers on image-text multi-modal fine-tuning [8]. Moreover, we also show our method can better customize the text-to-image diffusion models on personalized generation [56]." + }, + { + "type": "text", + "bbox": [ + 0.534, + 0.168, + 0.825, + 0.181 + ], + "angle": 0, + "content": "We summarize our contributions as follows," + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.183, + 0.905, + 0.228 + ], + "angle": 0, + "content": "- We provide a filter subspace view of the multi-head attention and propose to fine-tune a new filter subspace by learning a small set of subspace coefficients." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.229, + 0.905, + 0.273 + ], + "angle": 0, + "content": "- We demonstrate that our method can expand the feature space of the original multi-head attention and enhance its expressiveness on downstream tasks." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.274, + 0.905, + 0.303 + ], + "angle": 0, + "content": "- We validate our method on extensive fine-tuning tasks and various transformer architectures." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.183, + 0.905, + 0.303 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.318, + 0.655, + 0.334 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.344, + 0.905, + 0.782 + ], + "angle": 0, + "content": "Parameter-Efficient Fine-Tuning. Large attention-based transformers [11, 54, 55, 68] gain a significant amount of generalizable knowledge by pre-training on large-scale datasets [57, 58]. A common way to utilize those general skills is to fine-tune large transformers on domain-specific tasks and data. Recent advancements in parameter-efficient fine-tuning (PEFT) methods aim to reduce the computational and memory costs of adapting large models by fine-tuning only a subset of parameters. LoRA [21] and its variants [4, 24, 29, 36, 78, 82] fine-tune lower-rank matrices at each layer to represent weight updates. DoRA [38] enhances LoRA through weight decomposition into magnitude and direction, thus improving training stability and learning capacity without additional inference costs. The adapter [20] approach inserts small modules between layers and reduces parameters by only tuning these adapters [2, 24, 34, 50, 79]. Visual prompt tuning (VPT) [23, 60] has introduced a limited number of learnable parameters for optimization while keeping the backbone frozen. SSF [35] proposes scaling and shifting features of each layer in a pre-trained large model. OFT [53] tackles fine-tuning in text-to-image diffusion models by preserving the hyperspherical energy of neuron relationships. However, these methods primarily focus on efficient tuning of the parameter weight matrices, lacking a holistic view of the attention layer. In this paper, we take an orthogonal approach by tuning a tiny combination coefficient given a novel graph filter subspace view of the multi-head attention." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.795, + 0.905, + 0.901 + ], + "angle": 0, + "content": "The Expressiveness of Attention. Transformer-based models have achieved success across vision and language tasks [8, 11, 14, 22, 31, 37] due to the effectiveness of the attention mechanism [68]. Despite their success, various studies have examined the limitations in the expressiveness of transformers [10, 59, 71, 80]. Shi et al. [59] and Zhou et al. [84] identify the attention collapse in transformers as" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20147" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.114, + 0.09, + 0.882, + 0.334 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.345, + 0.907, + 0.417 + ], + "angle": 0, + "content": "Figure 2. Illustration of the proposed Coeff-Tuning. We first view each token \\(\\mathbf{X}[i]\\) as a node \\(V_{i}\\) in an all-connected graph \\(\\mathcal{G} = (V,E)\\). The attention operation then turns into a graph convolution with each attention map \\(\\mathbf{A}^h\\) as a graph convolutional filter \\(\\mathbf{F}^h (\\mathbf{X})\\). In this way, the attention maps in the multi-head attention layer form a filter subspace \\(\\mathcal{F}(\\mathbf{X})\\), and each element in the filter is constrained in \\((0,1)\\) after softmax. The proposed Coeff-Tuning then tunes a subspace coefficient \\(\\alpha \\in \\mathbb{R}^{H\\times H}\\) with a tiny number of parameters, that expands the range of attention scores by linearly combining the original filter subspace and enhances the expressiveness of the multi-head attention." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.43, + 0.486, + 0.627 + ], + "angle": 0, + "content": "a result of self-attention's convergence to a low-rank space, which leads to unstable training as attention scores concentrate [10, 45]. To counteract this, techniques such as hierarchical fusion [59] and stabilizing attention entropy [80] have been proposed to prevent performance stagnation in vision transformers. DeepViT [84] proposes to re-generate diverse attention maps to avoid attention collapse in training vision transformers from scratch. In contrast, our work studies the expressiveness of the attention layer in the context of the large model fine-tuning. We also provide a novel interpretation of multi-head attention maps as a graph filter subspace, and propose the residual parameterization and regularization to stabilize the tuning process." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.646, + 0.184, + 0.661 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.672, + 0.409, + 0.689 + ], + "angle": 0, + "content": "3.1. A Filter Subspace View on Attention" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.697, + 0.485, + 0.804 + ], + "angle": 0, + "content": "Attention. The attention layer transforms the input sequence \\(\\mathbf{X} \\in \\mathbb{R}^{N \\times C_i}\\) to the output sequence \\(\\mathbf{O} \\in \\mathbb{R}^{N \\times C_o}\\), where \\(N\\) denotes the sequence length, \\(C_i\\) and \\(C_o\\) are the dimension of input and output features. The attention layer projects the input as \\(\\mathbf{Q} = \\mathbf{X}\\mathbf{W}_q\\), \\(\\mathbf{K} = \\mathbf{X}\\mathbf{W}_k\\), \\(\\mathbf{V} = \\mathbf{X}\\mathbf{W}_v\\) using the corresponding projection matrices \\(\\mathbf{W}_q, \\mathbf{W}_k, \\mathbf{W}_v \\in \\mathbb{R}^{C_i \\times C_o}\\), and calculates the attention map," + }, + { + "type": "equation", + "bbox": [ + 0.186, + 0.818, + 0.483, + 0.855 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathbf {A} = \\operatorname {s o f t m a x} (\\mathbf {Q K} ^ {\\intercal}) \\tag {1} \\\\ = \\operatorname {s o f t m a x} \\left(\\mathbf {X} \\mathbf {W} _ {q} \\mathbf {W} _ {k} ^ {\\intercal} \\mathbf {X} ^ {\\intercal}\\right). \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.87, + 0.484, + 0.902 + ], + "angle": 0, + "content": "The attention map \\(\\mathbf{A} \\in \\mathbb{R}^{N \\times N}\\) captures the token-wise relationship by doing inner-product in a space transformed by" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.429, + 0.853, + 0.446 + ], + "angle": 0, + "content": "\\(\\mathbf{W}_q, \\mathbf{W}_k\\). Then the output of the attention layer is," + }, + { + "type": "equation", + "bbox": [ + 0.59, + 0.455, + 0.905, + 0.508 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathbf {O} = \\operatorname {a t t n} \\left(\\mathbf {W} _ {q}, \\mathbf {W} _ {k}, \\mathbf {W} _ {v}; \\mathbf {X}\\right) \\\\ = \\operatorname {s o f t m a x} \\left(\\mathbf {X} \\mathbf {W} _ {q} \\mathbf {W} _ {k} ^ {\\intercal} \\mathbf {X} ^ {\\intercal}\\right) \\mathbf {X} \\mathbf {W} _ {v} \\tag {2} \\\\ = \\mathbf {A X W} _ {v}. \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.525, + 0.907, + 0.587 + ], + "angle": 0, + "content": "Multi-head Attention. Multi-head attention extends this by allowing multiple attention mechanisms, i.e., heads, to work in parallel, with each head independently learning attention patterns. For each attention head, we have," + }, + { + "type": "equation", + "bbox": [ + 0.604, + 0.594, + 0.907, + 0.614 + ], + "angle": 0, + "content": "\\[\n\\mathbf {O} ^ {h} = \\operatorname {a t t n} \\left(\\mathbf {W} _ {q} ^ {h}, \\mathbf {W} _ {k} ^ {h}, \\mathbf {W} _ {v} ^ {h}; \\mathbf {X}\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.621, + 0.907, + 0.672 + ], + "angle": 0, + "content": "where \\(h = 1,\\dots ,H\\) and \\(H\\) is the number of heads. The project matrices for each head are \\(\\mathbf{W}_q^h,\\mathbf{W}_k^h,\\mathbf{W}_v^h\\in\\) \\(\\mathbb{R}^{C_i\\times \\frac{C_o}{H}}\\) . The multi-head attention represents as," + }, + { + "type": "equation", + "bbox": [ + 0.557, + 0.678, + 0.905, + 0.763 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathbf {Z} = \\operatorname {C o n c a t} \\left[ \\mathbf {O} ^ {1}, \\dots , \\mathbf {O} ^ {H} \\right] \\mathbf {W} _ {o} \\\\ = \\underset {H} {\\operatorname {C o n c a t}} \\left[ \\mathbf {A} ^ {1} \\mathbf {X} \\mathbf {W} _ {v} ^ {1}, \\dots , \\mathbf {A} ^ {H} \\mathbf {X} \\mathbf {W} _ {v} ^ {H} \\right] \\mathbf {W} _ {o} \\tag {4} \\\\ = \\sum_ {h = 1} ^ {H} \\mathbf {A} ^ {h} \\mathbf {X} \\mathbf {W} _ {v o} ^ {h}, \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.771, + 0.907, + 0.806 + ], + "angle": 0, + "content": "where \\(\\mathbf{W}_o\\in \\mathbb{R}^{C_o\\times C_o}\\) and \\(\\mathbf{W}_{vo}^h = \\mathbf{W}_v^h (\\mathbf{W}_o^h)^\\intercal\\) \\(\\mathbf{W}_o^h\\in\\) \\(\\mathbb{R}^{C_o\\times \\frac{C_o}{H}}\\) \\(\\mathbf{W}_{vo}^h\\in \\mathbb{R}^{C_i\\times C_o}\\)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.909, + 0.902 + ], + "angle": 0, + "content": "A Filter Subspace View on Multi-head Attention. We now present our filter subspace view on the multi-head attention. By viewing each token embedding \\(\\mathbf{X}[n] \\in \\mathbb{R}^{C_i}\\) as a node \\(V_{n}\\), which are connected with all the other nodes with edges \\(E[m, n] = 1\\), \\(\\forall m \\in [N]\\), we obtain a graph representation of all the tokens, \\(\\mathcal{G} = (V, E)\\)." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20148" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.155 + ], + "angle": 0, + "content": "We then reformulate the attention operation as the message passing as in Graph Convolution Nerworks [27]. Specifically, inspired by [6], we write the single-head attention as a graph convolution with the filter \\(\\mathbf{F}^h (\\mathbf{X})\\)" + }, + { + "type": "equation", + "bbox": [ + 0.108, + 0.162, + 0.483, + 0.209 + ], + "angle": 0, + "content": "\\[\n\\mathbf {O} ^ {h} [ n,: ] = \\sum_ {\\substack {n ^ {\\prime} \\in [ N ] \\\\ c ^ {\\prime} \\in [ C _ {i} ]}} \\mathbf {F} ^ {h} (\\mathbf {X}) [ n, n ^ {\\prime} ] \\mathbf {X} [ n ^ {\\prime}, c ^ {\\prime} ] \\mathbf {W} ^ {h} [ c ^ {\\prime},,: ], \\tag{5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.218, + 0.485, + 0.281 + ], + "angle": 0, + "content": "where \\(\\mathbf{W}^h = \\mathbf{W}_{vo}^h\\in \\mathbb{R}^{C_i\\times C_0}\\) denotes the weight for linear projection on each node, and the \\(n\\) -th row of the filter, i.e., \\(\\mathbf{F}^h (\\mathbf{X})[n,:]\\), represents the local graph filter around the \\(n\\) -th node (token). Comparing with (1) and (2), we have," + }, + { + "type": "equation", + "bbox": [ + 0.169, + 0.29, + 0.483, + 0.31 + ], + "angle": 0, + "content": "\\[\n\\mathbf {F} ^ {h} (\\mathbf {X}) = \\operatorname {s o f t m a x} \\left(\\mathbf {X} \\mathbf {W} _ {q} \\mathbf {W} _ {k} ^ {\\intercal} \\mathbf {X} ^ {\\intercal}\\right). \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.318, + 0.485, + 0.38 + ], + "angle": 0, + "content": "The multi-head attention individually learns \\(H\\) filters \\(\\{\\mathbf{F}^1 (X),\\dots,\\mathbf{F}^H (X)\\}\\) to aggregate complementary information across nodes. These \\(H\\) filters naturally form a graph convolutional filter subspace \\(\\mathcal{F}(\\mathbf{X})\\) as," + }, + { + "type": "equation", + "bbox": [ + 0.16, + 0.387, + 0.483, + 0.408 + ], + "angle": 0, + "content": "\\[\n\\mathcal {F} (\\mathbf {X}) = \\left\\{\\mathbf {F} ^ {h} (\\mathbf {X}) \\right\\} _ {h = 1} ^ {H} \\subseteq (0, 1) ^ {N \\times N}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.415, + 0.484, + 0.463 + ], + "angle": 0, + "content": "and we denote each \\(\\mathbf{F}^h (\\mathbf{X})\\) as a filter atom. The multi-head attention can then be represented in the way of decomposed graph convolution [6] as," + }, + { + "type": "equation", + "bbox": [ + 0.102, + 0.472, + 0.483, + 0.528 + ], + "angle": 0, + "content": "\\[\n\\mathbf {O} [ n,: ] = \\sum_ {h = 1} ^ {H} \\sum_ {\\substack {n ^ {\\prime} \\in [ N ] \\\\ c ^ {\\prime} \\in [ C _ {i} ]}} \\mathbf {F} ^ {h} (\\mathbf {X}) [ n, n ^ {\\prime} ] \\mathbf {X} [ n ^ {\\prime}, c ^ {\\prime} ] \\mathbf {W} ^ {h} [ c ^ {\\prime},: ]. \\tag{8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.538, + 0.484, + 0.659 + ], + "angle": 0, + "content": "It can be seen from (8) that the parameters \\(\\mathbf{F}^h (\\mathbf{X}),\\mathbf{W}^h\\) plays two key yet independent roles in the graph convolution. The \\(\\mathbf{W}^h\\) updates the knowledge for each node by projecting its representation into a new space, while the graph filter \\(\\mathbf{F}^h (\\mathbf{X})\\) guides the information flow across nodes. Intuitively, with a more expressive graph convolutional filter subspace \\(\\mathcal{F}(\\mathbf{X})\\), each node can aggregate more informative features from other nodes." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.667, + 0.347, + 0.684 + ], + "angle": 0, + "content": "3.2. Subspace Coefficient Tuning" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.483, + 0.825 + ], + "angle": 0, + "content": "Given the re-formulation of multi-head attention maps as a graph convolutional filter subspace, we present an efficient fine-tuning method specifically designed for transformers to enhance attention's expressiveness. Our motivation is that after the softmax function, each graph filter \\(\\mathbf{F}^h (\\mathbf{X})\\) has values in the range (0, 1) with row-sum as one. The normalized attention maps restrict the expressiveness in the way that each output embedding lies in the convex hull of the value embeddings." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.484, + 0.902 + ], + "angle": 0, + "content": "To expand the feature space of the output embeddings, we propose Coeff-Tuning, to tune the filter subspace on the target dataset with a small set of combination coefficients, i.e., subspace coefficient \\(\\alpha \\in \\mathbb{R}^{H\\times H}\\). As shown in Figure 2, the tunable coefficient \\(\\alpha\\) constructs a more expressive" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.091, + 0.907, + 0.139 + ], + "angle": 0, + "content": "filter subspace \\(\\hat{\\mathcal{F}}_{\\alpha}(\\mathbf{X})\\) by linearly combining the pre-trained multi-head attention maps \\(\\mathbf{F}(\\mathbf{X})\\). Each element \\(\\hat{\\mathbf{F}}^h (\\mathbf{X})\\) in the learned filter subspace can be written as," + }, + { + "type": "equation", + "bbox": [ + 0.557, + 0.15, + 0.907, + 0.192 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {F}} ^ {h} (\\mathbf {X}) = \\boldsymbol {\\alpha} [ h, : ] ^ {\\intercal} \\mathcal {F} (\\mathbf {X}) = \\sum_ {i = 1} ^ {H} \\boldsymbol {\\alpha} [ h, i ] \\mathbf {F} ^ {i} (\\mathbf {X}). \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.203, + 0.805, + 0.218 + ], + "angle": 0, + "content": "The multi-head attention then represents as," + }, + { + "type": "equation", + "bbox": [ + 0.525, + 0.229, + 0.906, + 0.288 + ], + "angle": 0, + "content": "\\[\n\\mathbf {O} = \\sum_ {h = 1} ^ {H} \\hat {\\mathbf {F}} ^ {h} (\\mathbf {X}) \\mathbf {X} \\mathbf {W} ^ {h} = \\sum_ {h = 1} ^ {H} \\sum_ {i = 1} ^ {H} \\alpha [ h, i ] \\mathbf {F} ^ {i} (\\mathbf {X}) \\mathbf {X} \\mathbf {W} ^ {h}. \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.299, + 0.907, + 0.39 + ], + "angle": 0, + "content": "Note that, mathematically, the subspace coefficient \\(\\alpha\\) is unconstrained, allowing it to combine attention maps with a wider range of values from the original filter subspace \\(\\mathbf{F}(\\mathbf{X})\\). With attention scores exceeding the range \\((0,1)\\), we effectively expand the space for \\(\\mathbf{O}\\) by allowing a more flexible combination of value matrices." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.403, + 0.907, + 0.464 + ], + "angle": 0, + "content": "Residual Parameterization. To facilitate the tuning and preserve the most knowledge in the pre-trained transformer, we propose a residual parameterization of the subspace coefficient," + }, + { + "type": "equation", + "bbox": [ + 0.612, + 0.475, + 0.699, + 0.491 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\alpha} ^ {\\prime} = \\boldsymbol {\\alpha} + \\mathbf {I},\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.581, + 0.495, + 0.906, + 0.537 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {F}} ^ {h} (\\mathbf {X}) = \\mathbf {F} ^ {h} (\\mathbf {X}) + \\sum_ {i = 1} ^ {H} \\boldsymbol {\\alpha} [ h, i ] \\mathbf {F} ^ {i} (\\mathbf {X}). \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.548, + 0.907, + 0.609 + ], + "angle": 0, + "content": "In this way, we maintain the original filter subspace with the identity matrix, while parameterizing the subspace coefficient \\(\\alpha'\\) to learn the residual combination of attention maps on fine-tuning tasks." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.622, + 0.907, + 0.744 + ], + "angle": 0, + "content": "Reguralization on Subspace Coefficient. The proposed Coeff-Tuning substantially enhances the expressiveness of multi-head attention with tunable coefficient \\(\\alpha\\). However, it may also lead to overfitting and preventing channel projection parameters \\(\\mathbf{W}_v, \\mathbf{W}_o\\) from learning generalizable features. To alleviate this, we propose a simple regularization design that performs dropout [63] directly on the tuned subspace coefficient \\(\\alpha\\) during training," + }, + { + "type": "equation", + "bbox": [ + 0.624, + 0.756, + 0.907, + 0.773 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\alpha} ^ {\\prime} = \\operatorname {D r o p o u t} (\\boldsymbol {\\alpha}; p) + \\mathbf {I}. \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.785, + 0.906, + 0.816 + ], + "angle": 0, + "content": "In this way, we randomly remove an element from \\(\\alpha\\) with a probability of \\(p\\) to avoid overfitting." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.83, + 0.611, + 0.847 + ], + "angle": 0, + "content": "4. Analysis" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.856, + 0.906, + 0.901 + ], + "angle": 0, + "content": "In this section, we analyze that adjusting the filter subspace using the proposed subspace coefficient \\(\\alpha\\) can effectively enhance the expressiveness of the multi-head attention." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "20149" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.107, + 0.089, + 0.87, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.236, + 0.907, + 0.336 + ], + "angle": 0, + "content": "Figure 3. The effectiveness of tuning subspace coefficient \\(\\alpha\\). The goal is to convert the input graph \\(\\mathbf{X}\\) to the target graph \\(\\mathbf{O}\\) using a single attention layer, \\(\\mathbf{O} = \\sum_{h=1}^{H} \\mathbf{F}^h(\\mathbf{X}) \\mathbf{X} \\mathbf{W}_v^h\\), where \\(\\mathbf{F}(\\mathbf{X}) = \\text{softmax}(\\mathbf{X} \\mathbf{W}_q \\mathbf{W}_k^\\top \\mathbf{X}^\\top)\\). We have the following observations: (i) When only tuning \\(\\mathbf{W}_q\\) and \\(\\mathbf{W}_k\\), i.e., updating \\(\\mathbf{F}(\\mathbf{X})\\), the positions of output nodes are constrained within the range of the input nodes \\(\\mathbf{X}\\), since the values of \\(\\mathbf{F}(\\mathbf{X})\\) are in the range of \\((0,1)\\). (ii) In addition to tuning \\(\\mathbf{F}(\\mathbf{X})\\), updating \\(\\mathbf{W}_v\\) projects the input as \\(\\mathbf{X} \\mathbf{W}_v\\), but the output graph is still not aligned with the target graph. (iii) In addition to tuning \\(\\mathbf{F}(\\mathbf{X})\\), adapting \\(\\alpha\\) successfully transforms the input graph into the target graph. Compared with (i), \\(\\alpha\\) expand the positions of output nodes \\(\\sum_{i=1}^{H} \\alpha[;, i] \\mathbf{F}^i(\\mathbf{X}) \\mathbf{X}\\) beyond the constraints of the input nodes, enhancing the expressiveness of the attention effectively." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.349, + 0.483, + 0.395 + ], + "angle": 0, + "content": "For the filter \\(\\mathbf{F}(\\mathbf{X})\\) in the filter subspace \\(\\mathcal{F}\\) since \\(\\mathbf{F}(\\mathbf{X}) = \\mathrm{softmax}(\\mathbf{X}\\mathbf{W}_q\\mathbf{W}_k^\\top \\mathbf{X}^\\top)\\), their elements \\(\\mathbf{F}(\\mathbf{X})[n,n^{\\prime}]\\) satisfy the following properties," + }, + { + "type": "equation", + "bbox": [ + 0.116, + 0.4, + 0.482, + 0.434 + ], + "angle": 0, + "content": "\\[\n0 < \\mathbf {F} (\\mathbf {X}) [ n, n ^ {\\prime} ] < 1, \\sum_ {n ^ {\\prime} \\in [ N ]} \\mathbf {F} (\\mathbf {X}) [ n, n ^ {\\prime} ] = 1. \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.44, + 0.485, + 0.488 + ], + "angle": 0, + "content": "Definition 4.1. Given \\(N\\) nodes \\(\\{\\mathbf{x}_i\\}_{i=1}^N\\), the set \\(\\mathcal{S}(\\mathbf{x}) = \\{\\sum_{i=1}^{N} \\lambda_i \\mathbf{x}_i \\mid \\lambda_i > 0, \\sum_i \\lambda_i = 1\\}\\) is a bounded convex set." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.494, + 0.484, + 0.587 + ], + "angle": 0, + "content": "According to the Definition 4.1, when the projected nodes \\(\\mathbf{X}\\mathbf{W}^h\\in \\mathbb{R}^{N\\times C_o}\\) of each head \\(h\\) are given, the output \\(\\mathbf{O}^h [n,:] = \\sum_{n'\\in [N]}\\mathbf{F}^h (\\mathbf{X})[n,n']\\mathbf{X}[n']\\mathbf{W}^h\\) lies within \\(S(\\mathbf{X}\\mathbf{W}^h)\\), which means the outputs \\(\\mathbf{O}^h [n,:]\\) are bounded. This property also applies to the multi-head attention mechanism." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.592, + 0.484, + 0.639 + ], + "angle": 0, + "content": "Proposition 4.2. If \\( S^h \\) is a bounded convex set, \\( \\forall h = 1, \\dots, H \\), the set \\( S = S^1 + \\dots + S^H = \\{ \\mathbf{x}^1 + \\dots + \\mathbf{x}^H \\mid \\forall \\mathbf{x}^h \\in S^h \\} \\) is also a bounded convex set." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.643, + 0.484, + 0.703 + ], + "angle": 0, + "content": "According to Proposition 4.2, when the projected nodes \\(\\mathbf{X}\\mathbf{W}^h\\) are given, the outputs of the multi-head attention \\(\\mathbf{O}[n,:] = \\sum_{h=1}^{H} \\sum_{n' \\in [N]} \\mathbf{F}^h(\\mathbf{X})[n,n']\\mathbf{X}[n']\\mathbf{W}^h\\) are also bounded." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.709, + 0.483, + 0.789 + ], + "angle": 0, + "content": "Proposition 4.3. After applying the subspace coefficients \\(\\alpha\\), the outputs of the reconstructed filter subspace \\(\\hat{\\mathcal{F}}_{\\alpha}\\) form a set \\(\\hat{S} = \\{\\sum_{h=1}^{H} \\sum_{i=1}^{H} \\boldsymbol{\\alpha}[h,i]\\mathbf{F}^i(\\mathbf{X})[n,n']\\mathbf{X}\\mathbf{W}^h \\mid \\mathbf{F}^i(\\mathbf{X})[n,n'] > 0, \\sum_{j} \\mathbf{F}^i(\\mathbf{X})[n,n'] = 1, \\boldsymbol{\\alpha}[h,i] \\in \\mathbb{R}\\}\\), which satisfies \\(S \\subseteq \\hat{S}\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.794, + 0.484, + 0.903 + ], + "angle": 0, + "content": "According to Proposition 4.3, when applying filter subspace coefficients \\(\\alpha\\), we can reconstruct the filter subspace \\(\\mathcal{F}\\) and expand the set \\(\\mathcal{S}\\). Specifically, given \\(\\mathbf{X}\\mathbf{W}^h\\), the output \\(\\mathbf{O}[n,:] = \\sum_{h=1}^{H} \\sum_{i=1}^{H} \\alpha[h,i]\\mathbf{F}^i(\\mathbf{X})[n,n']\\mathbf{X}\\mathbf{W}^h\\) can be in the set \\(\\hat{\\mathcal{S}} \\setminus \\mathcal{S}\\). Our method obtains an expanded set, enhancing the expressiveness of the model. We provide the support of our analyses in Appendix A." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.348, + 0.725, + 0.365 + ], + "angle": 0, + "content": "4.1. Validation Experiment" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.37, + 0.907, + 0.431 + ], + "angle": 0, + "content": "We assess the effectiveness of the subspace coefficients using a single attention layer. Given an input graph \\(\\mathbf{X}\\), the output of the attention layer is calculated as \\(\\mathbf{O} = \\mathbf{F}(\\mathbf{X})\\mathbf{X}\\mathbf{W}_v\\), where \\(\\mathbf{F}(\\mathbf{X}) = \\mathrm{softmax}(\\mathbf{X}\\mathbf{W}_q\\mathbf{W}_k^\\intercal\\mathbf{X}^\\intercal)\\)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.431, + 0.923, + 0.537 + ], + "angle": 0, + "content": "Our goal is to optimize \\(\\mathbf{W}_q\\), \\(\\mathbf{W}_k\\), \\(\\mathbf{W}_v\\) to transform the input graph \\(\\mathbf{X}\\) into the target graph \\(\\mathbf{O}\\) as shown in Figure 3. The input graph contains 8 nodes \\(\\mathbf{X} = [[-1,1],[0,1],[1,1],[1,0],[1,-1],[0,-1],[-1,-1],[-1,0]]\\) positioned in a range of \\([-1,1]\\) and the target graph \\(\\mathbf{O} = [[-1,1],[0,2],[1,1],[2,0],[1,-1],[0,-2],[-1,-1],[-2,0]]\\), with node coordinates in the range of \\([-2,2]\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.537, + 0.909, + 0.704 + ], + "angle": 0, + "content": "As shown in Figure 3, by only tuning the \\(\\mathbf{W}_q\\) and \\(\\mathbf{W}_k\\), i.e., \\(\\mathbf{F}(\\mathbf{X}) = \\mathrm{softmax}(\\mathbf{X}\\mathbf{W}_q\\mathbf{W}_k^\\top \\mathbf{X}^\\top)\\), the output is constrained inside the range of the input, which is aligned with Proposition 4.1. When we tune \\(\\mathbf{W}_q\\), \\(\\mathbf{W}_k\\), and \\(\\mathbf{W}_v\\), the output exceeds the input range but still cannot fully match the target graph. However, when we tune the subspace \\(\\mathbf{F}(\\mathbf{X})\\) and subspace coefficient \\(\\alpha\\), the attention layer can successfully transform the input graph into the target graph as in Figure 3. This observation is aligned with Proposition 4.3, and indicates that the proposed Coeff-Tuning can effectively expand the feature space of the multi-head attention layer." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.71, + 0.71, + 0.726 + ], + "angle": 0, + "content": "4.2. Complexity Analysis" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.732, + 0.909, + 0.901 + ], + "angle": 0, + "content": "We now formally analyze the complexity of the proposed Coeff-Tuning, and compare it with other baseline PEFT methods. The number of parameters required for our method is \\( H \\times H \\), where \\( H \\) is the number of heads at each attention layer. Considering \\( \\mathbf{W}_q, \\mathbf{W}_k, \\mathbf{W}_v \\in \\mathbb{R}^{C_i \\times C_o} \\) and \\( \\mathbf{W}_o \\in \\mathbb{R}^{C_o \\times C_o} \\) of the attention layer, our method introduces relative \\( \\frac{H^2}{3C_iC_o + C_o^2} \\) parameters. In modern large transformer models, we typically have \\( C_i = 2048 \\) and \\( C_o = 1280 \\). With \\( H = 10 \\), our method requires approximately 0.001% additional parameters. Compared with other PEFT methods such as LoRA, our method introduces" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20150" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.345, + 0.089, + 0.654, + 0.104 + ], + "angle": 0, + "content": "Table 1. Fine-tuning Results on VTAB-1k with ViT" + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.109, + 0.907, + 0.29 + ], + "angle": 0, + "content": "
NaturalSpecializedStructured
C100Cal101DTDF102PetsSVHNS397CameE.SATRes45ReinoCle_ctmCle_disDMLabKITTIdSlocationdS oriS_aziS_eleMeanParam.GTL_OPs
Full fine-tuning68.987.764.397.286.987.438.879.795.784.273.956.358.641.765.557.546.725.729.165.5785.845.15
Adapter74.186.163.297.787.034.650.876.388.073.170.545.737.431.253.230.325.413.822.155.820.23416.44
BitFit72.887.059.297.585.359.951.478.791.672.969.861.555.632.455.966.640.015.725.162.050.10515.08
VPT77.786.962.697.587.374.551.278.292.075.672.950.558.640.567.168.736.120.234.164.850.0616.68
LoRA67.191.469.498.890.485.354.084.995.384.473.682.969.249.881.781.848.332.844.272.910.30515.74
SPT-Adapter72.993.272.599.391.484.655.285.396.083.975.882.268.049.380.082.451.931.741.273.520.26616.44
Linear Probing63.485.063.297.086.336.651.078.587.568.674.034.330.633.255.412.520.09.619.252.94015.05
Coeff. α Only (Ours)69.590.874.599.190.983.051.183.695.384.675.857.459.146.578.675.051.425.634.269.780.00215.06
FACT-TT≤1671.389.670.798.991.087.854.685.295.583.475.782.069.049.880.079.248.434.241.473.040.04015.14
w/ Coeff-Tuning (Ours)72.189.771.899.191.288.954.885.395.784.275.383.068.851.081.380.049.335.042.773.640.04215.15
SNELL-873.792.772.499.291.489.255.484.996.186.475.284.068.553.581.082.749.933.939.274.170.26815.47
w/ Coeff-Tuning (Ours)74.292.973.199.291.690.355.685.296.286.375.884.168.853.781.384.052.135.039.974.700.27015.48
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.314, + 0.484, + 0.409 + ], + "angle": 0, + "content": "relative \\(\\frac{H^2}{3rC_i + 5rC_o}\\) parameters, where \\(r\\) is the LoRA rank. Using the same values for \\(C_i\\), \\(C_o\\), \\(H\\), and setting \\(r = 4\\), our approach requires less than \\(0.2\\%\\) additional parameters compared to LoRA and its variants. This also indicates that our Coeff-Tuning can maintain almost the same quantity of parameters when combined with other PEFT methods." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.422, + 0.226, + 0.44 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.448, + 0.484, + 0.57 + ], + "angle": 0, + "content": "In this section, we validate the proposed Coeff-Tuning with various tasks and datasets, including tuning the vision transformer [11] on few-shot image classification, tuning text-to-image diffusion model [51] for concept customization, and tuning vision-language BART [8] on image-to-text understanding. Combining with other weight-based PEFT methods, the proposed Coeff-Tuning consistently achieves superior results with negligible parameters added." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.581, + 0.357, + 0.597 + ], + "angle": 0, + "content": "5.1. Few-shot Image Classification" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.603, + 0.484, + 0.68 + ], + "angle": 0, + "content": "we first validate the proposed method by fine-tuning Vision Transformers on a series of downstream datasets for few-shot image classification. Our method achieves superior results while maintaining a high parameter efficiency compared with baseline methods." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.688, + 0.484, + 0.765 + ], + "angle": 0, + "content": "Experimental Settings. We utilize the VTAB-1k [81] dataset which comprises 19 diverse computer vision tasks across three domains: standard natural imagery, specialized capture, and synthetic/structured environments. The benchmark evaluates model performance on various task" + }, + { + "type": "table_caption", + "bbox": [ + 0.145, + 0.77, + 0.429, + 0.782 + ], + "angle": 0, + "content": "Table 2. Fine-tuning Results on the FGVC with ViT." + }, + { + "type": "table", + "bbox": [ + 0.084, + 0.782, + 0.477, + 0.899 + ], + "angle": 0, + "content": "
CUB-200NABirdsFlowersS-DogsS-CarsMean
Full fine-tuning87.382.798.889.484.588.5
Linear Probing85.375.997.986.251.379.3
Adapter87.384.398.488.868.485.5
BitFit88.484.298.891.279.488.4
VPT86.778.898.490.768.784.6
LoRA84.979.098.188.179.886.0
SPT-LoRA88.683.499.591.487.390.1
SNELL-889.686.899.392.189.991.5
w/ Coeff-Tuning (Ours)89.887.199.592.290.791.9
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.315, + 0.907, + 0.422 + ], + "angle": 0, + "content": "objectives. With only 1,000 training samples per task, it presents a significant few-shot learning challenge. We also benchmark our method on five fine-grained datasets in few-shot settings, including CUB-200-2011 [69], NABirds [67], Oxford Flowers [44], Stanford Dogs [25] and Stanford Cars [13]. A detailed description is provided in Appendix B.2." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.424, + 0.907, + 0.515 + ], + "angle": 0, + "content": "As for the vision transformer, we follow VPT [23], SSF [35], and select ViT-B/16 [11] pre-trained on ImageNet-21K [9] as the base model for fine-tuning. We benchmark the proposed Coeff-Tuning with a series of PEFT methods for few-shot learning including SSF [35], VPT [23], Adapter [20], LoRA [21], and BitFit [79]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.516, + 0.909, + 0.669 + ], + "angle": 0, + "content": "We present two settings of Coeff-Tuning in this task. We first explore the representation ability of tuning subspace coefficient \\(\\alpha\\) only, which tunes the transformer bottleneck with only tuning \\(H^2\\) (144 for ViT [11]) parameters for each transformer block. We also show the performance of combining Coeff-Tuning and SSF [35] to demonstrate that our method can plug and play with other PEFT methods. We set the coefficient dropout rates as \\(p = 0\\) and \\(p = 0.1\\) for these two settings respectively, and provide other hyperparameters in Appendix B.2." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Results. As shown in Table 1, the proposed Coeff-Tuning shows superior performances in both settings. Compared with the linear probing baseline which achieves a mean accuracy of 61.69, Coeff-Tuning obtains a significantly higher result of 73.49 with tuning only \\((12 * 12) * 12 = 1\\), 728 additional subspace coefficients \\(\\alpha\\) in ViT [11] (ViT [11] has 12 attention layers, each attention has 12 heads). Thus, we achieve a large performance increase in few-shot image classification \\((11.8^{\\uparrow})\\) with negligible additional parameter costs. Note that this result has already surpassed almost all baselines. We further achieve the best performance by combining Coeff-Tuning with SSF while maintaining almost the same parameter cost." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.958 + ], + "angle": 0, + "content": "20151" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.135, + 0.09, + 0.861, + 0.389 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.396, + 0.908, + 0.455 + ], + "angle": 0, + "content": "Figure 4. Results on Concept Customization. Compared to the baseline method (middle row), Coeff-Tuning (bottom row) leads to higher concept fidelity and text alignment of the generated images. For instance, our method accurately generates images of a “\\(\\langle V \\rangle\\) plushie sloth in the snow,” while the baseline fails to align with the text prompt. Additionally, our method generates a “\\(\\langle V \\rangle\\) wooden pot” with higher fidelity to the input image." + }, + { + "type": "table_caption", + "bbox": [ + 0.129, + 0.468, + 0.444, + 0.482 + ], + "angle": 0, + "content": "Table 3. Concept Customization Results with SDXL." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.487, + 0.484, + 0.564 + ], + "angle": 0, + "content": "
Method# Params (%)Text Align.Concept Align.Diversity
DoRA0.03710.330.7310.9
Ours + DoRA0.03740.340.7511.37
LoRA0.03060.330.7410.08
Ours + LoRA0.03090.340.7510.45
" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.577, + 0.429, + 0.594 + ], + "angle": 0, + "content": "5.2. Personalized Text-to-Image Generation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.6, + 0.484, + 0.812 + ], + "angle": 0, + "content": "We now show the effectiveness of Coeff-Tuning with the concept customization with Stable Diffusion (SDXL) [51, 55]. We validate the effectiveness of Coeff-Tuning by combining it with baseline PEFT methods, LoRA [21] and DoRA [38], and compare the performance with baselines without tuning \\(\\alpha\\). Specifically, we fine-tune SDXL [51] to associate the special token with a concept image from Dreambooth [56] and CustomConcept101 [30] datasets. We choose 10 different concepts, and fine-tune the model on \\(4 \\sim 9\\) images. We utilize 25 different text prompts following [53] and generate 4 images for each prompt. We evaluate the fidelity, diversity, and text-to-image alignment of the generated images. Other training details are described in Appendix B.3." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Evaluation metrics. We assess the concept alignment with the average cosine similarity between CLIP embeddings [54] of the generated images and the concept images. The diversity of generated images is measured by the Vendi score [12] calculated with the DINOv2 embeddings [46]." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.47, + 0.907, + 0.501 + ], + "angle": 0, + "content": "The text alignment of the generated images is measured via average cosine similarity in the CLIP feature space [54]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.507, + 0.907, + 0.628 + ], + "angle": 0, + "content": "Results. The quantitative results are shown in Table 3. By combining with LoRA [21] or DoRA [38], the proposed Coeff-Tuning generates images with greater diversity and better alignment with the both text prompt and input concept images. The qualitative results of our method compared with LoRA [21] are shown in Figure 4. The generated images with Coeff-Tuning have higher concept fidelity, preserving more characteristics from the input images." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.628, + 0.909, + 0.78 + ], + "angle": 0, + "content": "As an example, consider the concept “\\(\\langle V \\rangle\\) cat” as shown in the 4th column in Figure 4, the image generated using our method retains the cat's color from the input image, while the one generated by the baseline method alter the color. Additionally, the backgrounds in images generated with our method show more detailed features compared to those from the baseline method. Coeff-Tuning achieves the above improvements by introducing a limited number of additional parameters. We provide more generated images in Appendix B.3." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.789, + 0.757, + 0.805 + ], + "angle": 0, + "content": "5.3. Image-Text Understanding" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.907, + 0.903 + ], + "angle": 0, + "content": "As we have shown the effectiveness of Coeff-Tuning in fine-tuning models on both few-shot image classification and personalized image generation, we now validate it on multitask image-text understanding, which is a much larger-scale multi-modal fine-tuning task that requires a higher model capacity. By combining with baseline PEFT meth" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20152" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.33, + 0.089, + 0.668, + 0.104 + ], + "angle": 0, + "content": "Table 4. Multi-Model Fine-tuning Results with VL-Bart." + }, + { + "type": "table", + "bbox": [ + 0.254, + 0.109, + 0.737, + 0.216 + ], + "angle": 0, + "content": "
Method# Params (%)VQAV2GQANVLR2COCO CapAvg.
Full Model10066.956.773.7112.077.3
LoRA5.9465.253.671.9115.376.5
Ours + LoRA5.9466.254.173.0115.277.1
DoRA5.9665.854.773.1115.977.4
Ours + DoRA5.9666.454.973.0116.477.7
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.233, + 0.483, + 0.264 + ], + "angle": 0, + "content": "ods, our method again achieves superior performance with negligible parameter increase." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.291, + 0.484, + 0.504 + ], + "angle": 0, + "content": "Experimental Settings. We evaluate our method on VL-BART [8] multi-modal transformer, which utilizes CLIP-ResNet101 [54] for vision encoding and sends the visual features to BART-Base [31] for image-text joint modeling. Following [65], we evaluate our method with the multitask fine-tuning of VL-Bart [8] on four large-scale image-text understanding tasks simultaneously: VQAv2 [14] and GQA [22] for visual question answering, NLVR2 [64] for visual reasoning, and MSCOCO [5] for image captioning. Following the plug-and-play practice, we combine the proposed Coeff-Tuning with LoRA [21] and DoRA [38] respectively, and set the subspace coefficient dropout rate to \\( p = 0.2 \\). Detailed descriptions of other settings and hyperparameters are provided in Appendix B.4." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.525, + 0.484, + 0.661 + ], + "angle": 0, + "content": "Results. As shown in Table 4, our method surpasses baseline LoRA [21] and DoRA [38] respectively in almost all image-text datasets consistently, and achieves higher average scores. The Coeff-Tuning combined with LoRA [21] obtains comparable results with full-model fine-tuning, and it achieves better results than full-model tuning when it is combined with DoRA [38]. Notably, these improvements are achieved while maintaining the same parameter budgets as we only tune a hundred parameters for each attention." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.675, + 0.245, + 0.691 + ], + "angle": 0, + "content": "5.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.699, + 0.483, + 0.73 + ], + "angle": 0, + "content": "In this section, we examine the effectiveness of each component of our design with ablation experiments." + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.756, + 0.483, + 0.785 + ], + "angle": 0, + "content": "Table 5. Ablation studies on (a) residual parameterization and (b) coefficient dropout showing their effects on model performance." + }, + { + "type": "table", + "bbox": [ + 0.113, + 0.809, + 0.295, + 0.88 + ], + "angle": 0, + "content": "
ParametrizationAvg.
α w/ rand. init.44.28
α w/ identity init.61.57
I + α, zero init.73.49
" + }, + { + "type": "table_caption", + "bbox": [ + 0.194, + 0.882, + 0.212, + 0.893 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "table", + "bbox": [ + 0.309, + 0.796, + 0.48, + 0.881 + ], + "angle": 0, + "content": "
Dropout pAvg. Score
0.075.1
0.177.3
0.277.7
0.472.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.379, + 0.882, + 0.397, + 0.893 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.233, + 0.907, + 0.43 + ], + "angle": 0, + "content": "Residual Parameterization. We first demonstrate the importance of residual parameterization in Eq. 11 in Coeff-Tuning. We compare our residual parameterization with the direct parameterization where we directly tune \\(\\alpha\\) without adding it to the identity matrix, and initialize it with either a random matrix or an identity matrix. We adopt the few-shot tuning task as in Sec. 5.1 and only tune \\(\\alpha\\). Results in Table 5a show that models with residual parameterization achieve superior performance than direct parameterization with two different initializations. This validates our design that maintaining the original filter subspace while learning the residual combination helps preserve pre-trained knowledge and leads to better fine-tuning performance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.441, + 0.909, + 0.609 + ], + "angle": 0, + "content": "Subspace Coefficient Dropout. The dropout regularization on the subspace coefficient as in Eq. 12 plays a crucial role in preventing overfitting. We adopt the experiment setting in multi-modal fine-tuning as in Sec. 5.3. We experiment with different dropout rates \\( p \\in \\{0, 0.1, 0.2, 0.4\\} \\) during training. The results in Table 5b demonstrate that a moderate dropout rate (\\( p = 0.2 \\)) yields the best performance while either a lower or higher dropout rate leads to sub-optimal results. This suggests that applying dropout directly on the tuned subspace coefficient \\( \\alpha \\) can effectively prevent overfitting and lead to better generalization." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.621, + 0.634, + 0.637 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.646, + 0.909, + 0.889 + ], + "angle": 0, + "content": "In this paper, we presented Coeff-Tuning, a new parameter-efficient fine-tuning approach that enhances multi-head attention's expressiveness by learning a set of subspace coefficients. By reformulating attention as a graph convolution with a filter subspace, we demonstrated both theoretically and empirically that our method expands the feature space of multi-head attention beyond the original convex hull constraints. The proposed Coeff-Tuning can be seamlessly integrated with existing PEFT methods while introducing negligible additional parameters. Through extensive experiments across diverse tasks and modalities, we consistently demonstrated superior performance compared to competitive PEFT baselines. Our method achieved significant improvements in various metrics while maintaining high parameter efficiency, suggesting a promising direction for enhancing the adaptability of large pre-trained transformers." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20153" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.116, + 0.482, + 0.143 + ], + "angle": 0, + "content": "[1] Tom B Brown. Language models are few-shot learners. arXiv preprint arXiv:2005.14165, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.145, + 0.483, + 0.2 + ], + "angle": 0, + "content": "[2] Shoufa Chen, Chongjian Ge, Zhan Tong, Jiangliu Wang, Yibing Song, Jue Wang, and Ping Luo. Adaptformer: Adapting vision transformers for scalable visual recognition. Advances in Neural Information Processing Systems, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.202, + 0.483, + 0.243 + ], + "angle": 0, + "content": "[3] Wei Chen, Zichen Miao, and Qiang Qiu. Inner product-based neural network similarity. Advances in Neural Information Processing Systems, 36, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.245, + 0.483, + 0.285 + ], + "angle": 0, + "content": "[4] Wei Chen, Zichen Miao, and Qiang Qiu. Large convolutional model tuning via filter subspace. arXiv preprint arXiv:2403.00269, 2024.2, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.288, + 0.483, + 0.343 + ], + "angle": 0, + "content": "[5] Xinlei Chen, Hao Fang, Tsung-Yi Lin, Ramakrishna Vedantam, Saurabh Gupta, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco captions: Data collection and evaluation server. arXiv preprint arXiv:1504.00325, 2015. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.345, + 0.483, + 0.385 + ], + "angle": 0, + "content": "[6] Xiuyuan Cheng, Zichen Miao, and Qiang Qiu. Graph convolution with low-rank learnable local filters. arXiv preprint arXiv:2008.01818, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.388, + 0.483, + 0.43 + ], + "angle": 0, + "content": "[7] Xiuyuan Cheng, Zichen Miao, and Qiang Qiu. Graph convolution with low-rank learnable local filters. In International Conference on Learning Representations, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.432, + 0.483, + 0.487 + ], + "angle": 0, + "content": "[8] Jaemin Cho, Jie Lei, Hao Tan, and Mohit Bansal. Unifying vision-and-language tasks via text generation. In International Conference on Machine Learning, pages 1931-1942. PMLR, 2021. 2, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.489, + 0.483, + 0.544 + ], + "angle": 0, + "content": "[9] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 6, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.546, + 0.483, + 0.601 + ], + "angle": 0, + "content": "[10] Yihe Dong, Jean-Baptiste Cordonnier, and Andreas Loukas. Attention is not all you need: Pure attention loses rank doubly exponentially with depth. In International Conference on Machine Learning, pages 2793-2803. PMLR, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.603, + 0.483, + 0.63 + ], + "angle": 0, + "content": "[11] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. ICLR, 2021. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.632, + 0.483, + 0.673 + ], + "angle": 0, + "content": "[12] Dan Friedman and Adji Bousso Dieng. The vendi score: A diversity evaluation metric for machine learning. arXiv preprint arXiv:2210.02410, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.675, + 0.483, + 0.731 + ], + "angle": 0, + "content": "[13] Timnit Gebru, Jonathan Krause, Yilun Wang, Duyun Chen, Jia Deng, and Li Fei-Fei. Fine-grained car detection for visual census estimation. In Proceedings of the AAAI Conference on Artificial Intelligence, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.732, + 0.483, + 0.802 + ], + "angle": 0, + "content": "[14] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Bartra, and Devi Parikh. Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6904-6913, 2017. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.803, + 0.483, + 0.871 + ], + "angle": 0, + "content": "[15] Ligong Han, Yinxiao Li, Han Zhang, Peyman Milanfar, Dimitris Metaxas, and Feng Yang. Svdiff: Compact parameter space for diffusion fine-tuning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7323-7334, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[16] Liu He and Daniel Aliaga. Globalmapper: Arbitrary-shaped urban layout generation. In Proceedings of the IEEE/CVF" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.116, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.548, + 0.094, + 0.905, + 0.12 + ], + "angle": 0, + "content": "International Conference on Computer Vision, pages 454-464, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.122, + 0.905, + 0.163 + ], + "angle": 0, + "content": "[17] Liu He and Daniel Aliaga. Coho: Context-sensitive city-scale hierarchical urban layout generation. In European Conference on Computer Vision, pages 1-18. Springer, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.164, + 0.905, + 0.218 + ], + "angle": 0, + "content": "[18] Liu He, Yijuan Lu, John Corring, Dinei Florencio, and Cha Zhang. Diffusion-based document layout generation. In International Conference on Document Analysis and Recognition, pages 361-378. Springer, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.219, + 0.905, + 0.273 + ], + "angle": 0, + "content": "[19] Liu He, Yizhi Song, Hejun Huang, Daniel Aliaga, and Xin Zhou. Kubrick: Multimodal agent collaborations for synthetic video generation. arXiv preprint arXiv:2408.10453, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.275, + 0.905, + 0.343 + ], + "angle": 0, + "content": "[20] Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin De Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. Parameter-efficient transfer learning for nlp. In International conference on machine learning, pages 2790-2799. PMLR, 2019. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.345, + 0.905, + 0.399 + ], + "angle": 0, + "content": "[21] Edward J Hu, Phillip Wallis, Zeyuan Allen-Zhu, Yuzhhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. In International Conference on Learning Representations, 2021. 1, 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.401, + 0.905, + 0.468 + ], + "angle": 0, + "content": "[22] Drew A Hudson and Christopher D Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6700-6709, 2019. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.47, + 0.905, + 0.525 + ], + "angle": 0, + "content": "[23] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In European Conference on Computer Vision, pages 709-727. Springer, 2022. 2, 6, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.526, + 0.905, + 0.579 + ], + "angle": 0, + "content": "[24] Rabeeh Karimi Mahabadi, James Henderson, and Sebastian Ruder. Compacter: Efficient low-rank hypercomplex adapter layers. Advances in Neural Information Processing Systems, 2021. 2, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.581, + 0.905, + 0.636 + ], + "angle": 0, + "content": "[25] Aditya Khosla, Nityananda Jayadevaprakash, Bangpeng Yao, and Fei-Fei Li. Novel dataset for fine-grained image categorization: Stanford dogs. In Proc. CVPR workshop on fine-grained visual categorization (FGVC), 2011. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.637, + 0.905, + 0.678 + ], + "angle": 0, + "content": "[26] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In International Conference on Learning Representations, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.679, + 0.905, + 0.719 + ], + "angle": 0, + "content": "[27] Thomas N Kipf and Max Welling. Semi-supervised classification with graph convolutional networks. arXiv preprint arXiv:1609.02907, 2016. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.721, + 0.905, + 0.789 + ], + "angle": 0, + "content": "[28] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.79, + 0.905, + 0.83 + ], + "angle": 0, + "content": "[29] Dawid J Kopiczko, Tijmen Blankevoort, and Yuki M Asano. Vera: Vector-based random matrix adaptation. *ICLR*, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.832, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[30] Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization of text-to-image diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 7, 1" + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.094, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "20154" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.146 + ], + "angle": 0, + "content": "[31] M Lewis. Bart: Denoising sequence-to-sequence pretraining for natural language generation, translation, and comprehension. arXiv preprint arXiv:1910.13461, 2019. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.189 + ], + "angle": 0, + "content": "[32] Junbo Li, Zichen Miao, Qiang Qiu, and Ruqi Zhang. Training bayesian neural networks with sparse subspace variational inference. arXiv preprint arXiv:2402.11025, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.191, + 0.482, + 0.258 + ], + "angle": 0, + "content": "[33] Tianchun Li, Chengxiang Wu, Pengyi Shi, and Xiaoqian Wang. Cumulative difference learning vae for time-series with temporally correlated inflow-outflow. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 13619-13627, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.261, + 0.482, + 0.302 + ], + "angle": 0, + "content": "[34] Xiang Lisa Li and Percy Liang. Prefix-tuning: Optimizing continuous prompts for generation. In Proceedings of the Association for Computational Linguistics, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.303, + 0.482, + 0.357 + ], + "angle": 0, + "content": "[35] Dongze Lian, Daquan Zhou, Jiashi Feng, and Xinchao Wang. Scaling & shifting your features: A new baseline for efficient model tuning. Advances in Neural Information Processing Systems, 35:109-123, 2022. 2, 6, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.359, + 0.482, + 0.426 + ], + "angle": 0, + "content": "[36] Haokun Liu, Derek Tam, Mohammed Muqeeth, Jay Mohta, Tenghao Huang, Mohit Bansal, and Colin A Raffel. Few-shot parameter-efficient fine-tuning is better and cheaper than in-context learning. Advances in Neural Information Processing Systems, 35:1950-1965, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.428, + 0.482, + 0.469 + ], + "angle": 0, + "content": "[37] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.47, + 0.482, + 0.536 + ], + "angle": 0, + "content": "[38] Shih-Yang Liu, Chien-Yi Wang, Hongxu Yin, Pavlo Molchanov, Yu-Chiang Frank Wang, Kwang-Ting Cheng, and Min-Hung Chen. Dora: Weight-decomposed low-rank adaptation. arXiv preprint arXiv:2402.09353, 2024. 2, 7, 8, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.539, + 0.482, + 0.607 + ], + "angle": 0, + "content": "[39] Tianci Liu, Ruirui Li, Yunzhe Qi, Hui Liu, Xianfeng Tang, Tianqi Zheng, Qingyu Yin, Monica Xiao Cheng, Jun Huan, Haoyu Wang, et al. Unlocking efficient, scalable, and continual knowledge editing with basis-level representation finetuning. arXiv preprint arXiv:2503.00306, 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.609, + 0.482, + 0.65 + ], + "angle": 0, + "content": "[40] Zichen Miao, Ze Wang, Wei Chen, and Qiang Qiu. Continual learning with filter atom swapping. In International Conference on Learning Representations, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.651, + 0.482, + 0.705 + ], + "angle": 0, + "content": "[41] Zichen Miao, Ze Wang, Xiuyuan Cheng, and Qiang Qiu. Spatiotemporal joint filter decomposition in 3d convolutional neural networks. Advances in Neural Information Processing Systems, 34:3376-3388, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.707, + 0.482, + 0.788 + ], + "angle": 0, + "content": "[42] Zichen Miao, Jiang Wang, Ze Wang, Zhengyuan Yang, Lijuan Wang, Qiang Qiu, and Zicheng Liu. Training diffusion models towards diverse image generation with reinforcement learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10844-10853, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.79, + 0.482, + 0.845 + ], + "angle": 0, + "content": "[43] Zichen Miao, Zhengyuan Yang, Kevin Lin, Ze Wang, Zicheng Liu, Lijuan Wang, and Qiang Qiu. Tuning timestep-distilled diffusion model using pairwise sample optimization. arXiv preprint arXiv:2410.03190, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[44] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In 2008 Sixth Indian conference on computer vision, graphics & image processing, pages 722-729. IEEE, 2008. 6" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.161 + ], + "angle": 0, + "content": "[45] Lorenzo Noci, Sotiris Anagnostidis, Luca Biggio, Antonio Orvieto, Sidak Pal Singh, and Aurelien Lucchi. Signal propagation in transformers: Theoretical perspectives and the role of rank collapse. Advances in Neural Information Processing Systems, 35:27198-27211, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.163, + 0.905, + 0.232 + ], + "angle": 0, + "content": "[46] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.233, + 0.905, + 0.273 + ], + "angle": 0, + "content": "[47] Gaurav Patel and Qiang Qiu. Learning to unlearn while retaining: Combating gradient conflicts in machine unlearning. arXiv preprint arXiv:2503.06339, 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.274, + 0.905, + 0.343 + ], + "angle": 0, + "content": "[48] Gaurav Patel, Konda Reddy Mopuri, and Qiang Qiu. Learning to retain while acquiring: Combating distribution-shift in adversarial data-free knowledge distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7786-7794, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.344, + 0.905, + 0.411 + ], + "angle": 0, + "content": "[49] Gaurav Patel, Christopher Sandino, Behrooz Mahasseni, Ellen L Zippi, Erdrin Azemi, Ali Moin, and Juri Minxha. Efficient source-free time-series adaptation via parameter subspace disentanglement. arXiv preprint arXiv:2410.02147, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.413, + 0.905, + 0.469 + ], + "angle": 0, + "content": "[50] Jonas Pfeiffer, Aishwarya Kamath, Andreas Rücklé, Kyunghyun Cho, and Iryna Gurevych. Adapterfusion: Nondestructive task composition for transfer learning. arXiv preprint arXiv:2005.00247, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.47, + 0.905, + 0.525 + ], + "angle": 0, + "content": "[51] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. *ICLR*, 2024. 1, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.526, + 0.905, + 0.579 + ], + "angle": 0, + "content": "[52] Qiang Qiu, Xiuyuan Cheng, Guillermo Sapiro, et al. Dcfnet: Deep neural network with decomposed convolutional filters. In International Conference on Machine Learning, pages 4198-4207. PMLR, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.581, + 0.905, + 0.649 + ], + "angle": 0, + "content": "[53] Zeju Qiu, Weiyang Liu, Haiwen Feng, Yuxuan Xue, Yao Feng, Zhen Liu, Dan Zhang, Adrian Weller, and Bernhard Scholkopf. Controlling text-to-image diffusion by orthogonal finetuning. Advances in Neural Information Processing Systems, 36:79320-79362, 2023. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.65, + 0.905, + 0.733 + ], + "angle": 0, + "content": "[54] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, 2021. 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.735, + 0.905, + 0.788 + ], + "angle": 0, + "content": "[55] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10684-10695, 2022. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.79, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[56] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 22500-22510, 2023. 2, 7, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.903 + ], + "angle": 0, + "content": "[57] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy," + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20155" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.484, + 0.147 + ], + "angle": 0, + "content": "Aditya Khosla, Michael Bernstein, Alexander C. Berg, and Li Fei-Fei. ImageNet Large Scale Visual Recognition Challenge. International Journal of Computer Vision (IJCV), 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.483, + 0.234 + ], + "angle": 0, + "content": "[58] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278-25294, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.235, + 0.483, + 0.29 + ], + "angle": 0, + "content": "[59] Han Shi, Jiahui Gao, Hang Xu, Xiaodan Liang, Zhenguo Li, Lingpeng Kong, Stephen Lee, and James T Kwok. Revisiting over-smoothing in bert from the perspective of graph. arXiv preprint arXiv:2202.08625, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.291, + 0.483, + 0.361 + ], + "angle": 0, + "content": "[60] Kihyuk Sohn, Huiwen Chang, José Lezama, Luisa Polania, Han Zhang, Yuan Hao, Irfan Essa, and Lu Jiang. Visual prompt tuning for generative transfer learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.362, + 0.483, + 0.432 + ], + "angle": 0, + "content": "[61] Yizhi Song, Liu He, Zhifei Zhang, Soo Ye Kim, He Zhang, Wei Xiong, Zhe Lin, Brian Price, Scott Cohen, Jianming Zhang, et al. Refine-by-align: Reference-guided artifacts refinement through semantic alignment. arXiv preprint arXiv:2412.00306, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.434, + 0.483, + 0.517 + ], + "angle": 0, + "content": "[62] Yizhi Song, Zhifei Zhang, Zhe Lin, Scott Cohen, Brian Price, Jianming Zhang, Soo Ye Kim, He Zhang, Wei Xiong, and Daniel Aliaga. Imprint: Generative object compositing by learning identity-preserving representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8048-8058, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.518, + 0.483, + 0.573 + ], + "angle": 0, + "content": "[63] Nitish Srivastava, Geoffrey Hinton, Alex Krizhevsky, Ilya Sutskever, and Ruslan Salakhutdinov. Dropout: a simple way to prevent neural networks from overfitting. The journal of machine learning research, 15(1):1929-1958, 2014. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.575, + 0.483, + 0.631 + ], + "angle": 0, + "content": "[64] Alane Suhr, Stephanie Zhou, Ally Zhang, Iris Zhang, Hua-jun Bai, and Yoav Artzi. A corpus for reasoning about natural language grounded in photographs. arXiv preprint arXiv:1811.00491, 2018. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.632, + 0.483, + 0.7 + ], + "angle": 0, + "content": "[65] Yi-Lin Sung, Jaemin Cho, and Mohit Bansal. Vl-adapter: Parameter-efficient transfer learning for vision-and-language tasks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5227-5237, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.703, + 0.483, + 0.771 + ], + "angle": 0, + "content": "[66] Gemma Canet Tarres, Zhe Lin, Zhifei Zhang, Jianming Zhang, Yizhi Song, Dan Ruta, Andrew Gilbert, John Colomosse, and Soo Ye Kim. Thinking outside the bbox: Unconstrained generative object compositing. arXiv preprint arXiv:2409.04559, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.774, + 0.483, + 0.87 + ], + "angle": 0, + "content": "[67] Grant Van Horn, Steve Branson, Ryan Farrell, Scott Haber, Jessie Barry, Panos Ipeirotis, Pietro Perona, and Serge Belongie. Building a bird recognition app and large scale dataset with citizen scientists: The fine print in fine-grained dataset collection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 595-604, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.872, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[68] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.134 + ], + "angle": 0, + "content": "[69] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The caltech-ucsd birds-200-2011 dataset. 2011. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.204 + ], + "angle": 0, + "content": "[70] Haoyu Wang, Tianci Liu, Ruirui Li, Monica Cheng, Tuo Zhao, and Jing Gao. Roselora: Row and column-wise sparse low-rank adaptation of pre-trained language model for knowledge editing and fine-tuning. arXiv preprint arXiv:2406.10777, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.206, + 0.905, + 0.262 + ], + "angle": 0, + "content": "[71] Peihao Wang, Wenqing Zheng, Tianlong Chen, and Zhangyang Wang. Anti-oversmoothing in deep vision transformers via the fourier domain analysis: From theory to practice. arXiv preprint arXiv:2203.05962, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.264, + 0.905, + 0.305 + ], + "angle": 0, + "content": "[72] Ze Wang, Xiuyuan Cheng, Guillermo Sapiro, and Qiang Qiu. Stochastic conditional generative networks with basis decomposition. arXiv preprint arXiv:1909.11286, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.307, + 0.905, + 0.361 + ], + "angle": 0, + "content": "[73] Ze Wang, Seunghyun Hwang, Zichen Miao, and Qiang Qiu. Image generation using continuous filter atoms. Advances in Neural Information Processing Systems, 34:17826-17838, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.363, + 0.905, + 0.42 + ], + "angle": 0, + "content": "[74] Ze Wang, Zichen Miao, Jun Hu, and Qiang Qiu. Adaptive convolutions with per-pixel dynamic filter atom. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12302-12311, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.421, + 0.905, + 0.475 + ], + "angle": 0, + "content": "[75] Ze Wang, Zichen Miao, Xiantong Zhen, and Qiang Qiu. Learning to learn dense gaussian processes for few-shot learning. Advances in Neural Information Processing Systems, 34:13230-13241, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.477, + 0.905, + 0.561 + ], + "angle": 0, + "content": "[76] Enze Xie, Lewei Yao, Han Shi, Zhili Liu, Daquan Zhou, Zhaoqiang Liu, Jiawei Li, and Zhenguo Li. Difffit: Unlocking transferability of large diffusion models via simple parameter-efficient fine-tuning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4230-4239, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.562, + 0.905, + 0.617 + ], + "angle": 0, + "content": "[77] Zhexiao Xiong, Wei Xiong, Jing Shi, He Zhang, Yizhi Song, and Nathan Jacobs. Groundingbooth: Grounding text-to-image customization. arXiv preprint arXiv:2409.08520, 2024.1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.619, + 0.905, + 0.689 + ], + "angle": 0, + "content": "[78] SHIH-YING YEH, Yu-Guan Hsieh, Zhidong Gao, Bernard BW Yang, Gigeong Oh, and Yanmin Gong. Navigating text-to-image customization: From lycoris fine-tuning to model evaluation. In International Conference on Learning Representations, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.69, + 0.905, + 0.758 + ], + "angle": 0, + "content": "[79] Elad Ben Zaken, Yoav Goldberg, and Shauli Ravfogel. Bitfit: Simple parameter-efficient fine-tuning for transformer-based masked language-models. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.76, + 0.905, + 0.843 + ], + "angle": 0, + "content": "[80] Shuangfei Zhai, Tatiana Likhomanenko, Etai Littwin, Dan Busbridge, Jason Ramapuram, Yizhe Zhang, Jiatao Gu, and Joshua M Susskind. Stabilizing transformer training by preventing attention entropy collapse. In International Conference on Machine Learning, pages 40770-40803. PMLR, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.845, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[81] Xiaohua Zhai, Joan Puigcerver, Alexander Kolesnikov, Pierre Ruyssen, Carlos Riquelme, Mario Lucic, Josip Djolonga, Andre Susano Pinto, Maxim Neumann, Alexey Dosovitskiy, et al. A large-scale study of representation learning" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "20156" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.124, + 0.092, + 0.484, + 0.12 + ], + "angle": 0, + "content": "with the visual task adaptation benchmark. arXiv preprint arXiv:1910.04867, 2019. 6, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.484, + 0.177 + ], + "angle": 0, + "content": "[82] Qingru Zhang, Minshuo Chen, Alexander Bukharin, Nikos Karampatziakis, Pengcheng He, Yu Cheng, Weizhu Chen, and Tuo Zhao. Adalora: Adaptive budget allocation for parameter-efficient fine-tuning. *ICLR*, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.179, + 0.482, + 0.234 + ], + "angle": 0, + "content": "[83] Yibo Zhong, Haoxiang Jiang, Lincan Li, Ryumei Nakada, Tianci Liu, Linjun Zhang, Huaxiu Yao, and Haoyu Wang. Neat: Nonlinear parameter-efficient adaptation of pretrained models. arXiv preprint arXiv:2410.01870, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.235, + 0.484, + 0.29 + ], + "angle": 0, + "content": "[84] Daquan Zhou, Bingyi Kang, Xiaojie Jin, Linjie Yang, Xiaochen Lian, Zihang Jiang, Qibin Hou, and Jiashi Feng. Deepvit: Towards deeper vision transformer. arXiv preprint arXiv:2103.11886, 2021. 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.29 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20157" + } + ] +] \ No newline at end of file diff --git a/2025/Coeff-Tuning_ A Graph Filter Subspace View for Tuning Attention-Based Large Models/e7a23930-dc20-4352-a4dc-e2a2616c610f_origin.pdf b/2025/Coeff-Tuning_ A Graph Filter Subspace View for Tuning Attention-Based Large Models/e7a23930-dc20-4352-a4dc-e2a2616c610f_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..47541d2141d80147c98cec646dec2a5bf74f28bd --- /dev/null +++ b/2025/Coeff-Tuning_ A Graph Filter Subspace View for Tuning Attention-Based Large Models/e7a23930-dc20-4352-a4dc-e2a2616c610f_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:855cffdffecd817a86f01cee6c988951d183f787e088950bd28da230a571e50d +size 1492152 diff --git a/2025/Coeff-Tuning_ A Graph Filter Subspace View for Tuning Attention-Based Large Models/full.md b/2025/Coeff-Tuning_ A Graph Filter Subspace View for Tuning Attention-Based Large Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..76cb2394ebac5c4e37be3f37a0cf1df682830daa --- /dev/null +++ b/2025/Coeff-Tuning_ A Graph Filter Subspace View for Tuning Attention-Based Large Models/full.md @@ -0,0 +1,379 @@ +# Coeff-Tuning: A Graph Filter Subspace View for Tuning Attention-Based Large Models + +Zichen Miao†, Wei Chen†, Qiang Qiu Purdue University, IN, USA + +{miaoz, chen2732, qqiu}@purdue.edu + +# Abstract + +Transformer-based large pre-trained models have shown remarkable generalization ability, and various parameter-efficient fine-tuning (PEFT) methods have been proposed to customize these models on downstream tasks with minimal computational and memory budgets. Previous PEFT methods are primarily designed from a tensor-decomposition perspective that tries to effectively tune the linear transformation by finding the smallest subset of parameters to train. Our study adopts an orthogonal view by representing the attention operation as a graph convolution and formulating the multi-head attention maps as a convolutional filter subspace, with each attention map as a subspace element. In this paper, we propose to tune the large pre-trained transformers by learning a small set of combination coefficients that construct a more expressive filter subspace from the original multi-head attention maps. We show analytically and experimentally that the tuned filter subspace can effectively expand the feature space of the multi-head attention and further enhance the capacity of transformers. We further stabilize the fine-tuning with a residual parameterization of the tunable subspace coefficients, and enhance the generalization with a regularization design by directly applying dropout on the tunable coefficient during training. The tunable coefficients take a tiny number of parameters and can be combined with previous PEFT methods in a plug-and-play manner. Extensive experiments show that our approach achieves superior performances than PEFT baselines with neglectable additional parameters. $^{1}$ + +# 1. Introduction + +Large foundational models have demonstrated remarkable performance and generalization ability across a wide range of natural language processing and computer vision + +![](images/0815525a3a84eb0cfff84eee395e64934a19e0f9a5d2a1c968962f9036d3cc75.jpg) + +![](images/49bb2896c59a987c7204b513e0614d167910713c6d31bab73f622a30fbfb3e67.jpg) + +![](images/08444f97283a41285cc1e6a3fac51216f3a3927e61125f45277baec4bdd4d0c4.jpg) +A [V] plushie panda on top of a purple rug in a forest. + +![](images/74646b37219c208b016d61193f640aa8da780874b6c2829704fbeaa9300ff873.jpg) + +![](images/498db363057d9dbe54d77e5f5fcfbf7156bc5ec867e58301e3bb0e9f6348f0fa.jpg) +A [V] plushie panda on in the snow. +LoRA +Figure 1. (Up) Under the graph convolution view of the attention, the proposed Coeff-Tuning enhances the expressiveness by combining the attention maps with a tiny tunable subspace coefficient $\alpha$ . It can be seamlessly integrated with other PEFT methods, e.g., LoRA [21], by introducing negligible additional parameters. (Bottom) Combining Coeff-Tuning with LoRA achieves better results in the concept customization task with SDXL [51]. + +![](images/236494ecb81262f72717a042ff784bd414d1ceb29f4bd63f97f0ec10e063104f.jpg) +LoRA + Coeff-Tuning + +tasks [11, 28, 55, 68]. The success of these models can be attributed to their ability to learn rich, transferable representations by leveraging the power of the transformer architecture [68] and the extensive knowledge captured during + +pre-training on large-scale datasets [1, 58]. However, fine-tuning these pre-trained transformers on downstream tasks with full-parameter updates can be challenging due to the high computational and memory costs. + +To address this, various parameter-efficient fine-tuning (PEFT) methods have been developed to adapt pre-trained models on downstream tasks while minimizing the number of trainable parameters [21, 38, 53, 79]. A representative family of these approaches focuses on tuning the linear transformation layers in transformers, where they apply a lightweight parameterization, e.g., low-rank parameterization [21, 38], to adapt the weight matrices. However, none of these methods are designed from a holistic view of the core operation in transformers, the multi-head attention layer [68], which plays a key role in the expressiveness of transformers [10, 59]. + +In this paper, we propose a new PEFT method tailored to enhance the multi-head attention's expressiveness by tuning a minimal set of parameters. We first formulate the multi-head attention as a graph convolution with a filter subspace, with each graph filter as the attention map of each attention head. We then tune the filter subspace by simply learning a small set of subspace coefficients that linearly combine the pre-trained attention maps. In this way, we adapt for each attention head a new task-adaptive attention map that better captures the token-wise correlation. + +We attribute the effectiveness of our approach to its ability to expand the feature space of the original multi-head attention layer. By applying the softmax function to the query-key correlation matrices, all elements in a row of the original attention map are non-negative and have a sum of one [68]. This limits the expressiveness of the multi-head attention as the output token embeddings lie within the convex hull of the value embeddings. Our method constructs a new filter subspace with non-constrained coefficients as shown in Figure 1, allowing negative elements learned in the combined attention map to capture more diverse token-wise relationships. This also enables the output tokens to obtain representation outside the convex hull of the value embeddings, which further enhances the capacity of the multi-head attention. We provide analysis and an illustrative toy example to better demonstrate this point. + +We show that our approach can be seamlessly combined with previous PEFT methods in a plug-and-play manner. As the filter subspace in each attention layer is spanned by a tiny number of attention maps (usually $10\sim 30$ ), we only introduce neglectable additional tunable parameters and training costs over the base PEFT methods. + +We validate our approach on a large variety of fine-tuning tasks and datasets, along with pre-trained transformers of different scales and modalities. Specifically, we first show the effectiveness of our method with ViT [11] on the few-shot image classification fine-tuning. We also + +show our method can achieve superior performance with vision-language transformers on image-text multi-modal fine-tuning [8]. Moreover, we also show our method can better customize the text-to-image diffusion models on personalized generation [56]. + +We summarize our contributions as follows, + +- We provide a filter subspace view of the multi-head attention and propose to fine-tune a new filter subspace by learning a small set of subspace coefficients. +- We demonstrate that our method can expand the feature space of the original multi-head attention and enhance its expressiveness on downstream tasks. +- We validate our method on extensive fine-tuning tasks and various transformer architectures. + +# 2. Related Work + +Parameter-Efficient Fine-Tuning. Large attention-based transformers [11, 54, 55, 68] gain a significant amount of generalizable knowledge by pre-training on large-scale datasets [57, 58]. A common way to utilize those general skills is to fine-tune large transformers on domain-specific tasks and data. Recent advancements in parameter-efficient fine-tuning (PEFT) methods aim to reduce the computational and memory costs of adapting large models by fine-tuning only a subset of parameters. LoRA [21] and its variants [4, 24, 29, 36, 78, 82] fine-tune lower-rank matrices at each layer to represent weight updates. DoRA [38] enhances LoRA through weight decomposition into magnitude and direction, thus improving training stability and learning capacity without additional inference costs. The adapter [20] approach inserts small modules between layers and reduces parameters by only tuning these adapters [2, 24, 34, 50, 79]. Visual prompt tuning (VPT) [23, 60] has introduced a limited number of learnable parameters for optimization while keeping the backbone frozen. SSF [35] proposes scaling and shifting features of each layer in a pre-trained large model. OFT [53] tackles fine-tuning in text-to-image diffusion models by preserving the hyperspherical energy of neuron relationships. However, these methods primarily focus on efficient tuning of the parameter weight matrices, lacking a holistic view of the attention layer. In this paper, we take an orthogonal approach by tuning a tiny combination coefficient given a novel graph filter subspace view of the multi-head attention. + +The Expressiveness of Attention. Transformer-based models have achieved success across vision and language tasks [8, 11, 14, 22, 31, 37] due to the effectiveness of the attention mechanism [68]. Despite their success, various studies have examined the limitations in the expressiveness of transformers [10, 59, 71, 80]. Shi et al. [59] and Zhou et al. [84] identify the attention collapse in transformers as + +![](images/80e2fe75c334f99f28bd010b2c8adf95a63b5aee8e96e79f60c2303965e95bf8.jpg) +Figure 2. Illustration of the proposed Coeff-Tuning. We first view each token $\mathbf{X}[i]$ as a node $V_{i}$ in an all-connected graph $\mathcal{G} = (V,E)$ . The attention operation then turns into a graph convolution with each attention map $\mathbf{A}^h$ as a graph convolutional filter $\mathbf{F}^h (\mathbf{X})$ . In this way, the attention maps in the multi-head attention layer form a filter subspace $\mathcal{F}(\mathbf{X})$ , and each element in the filter is constrained in $(0,1)$ after softmax. The proposed Coeff-Tuning then tunes a subspace coefficient $\alpha \in \mathbb{R}^{H\times H}$ with a tiny number of parameters, that expands the range of attention scores by linearly combining the original filter subspace and enhances the expressiveness of the multi-head attention. + +a result of self-attention's convergence to a low-rank space, which leads to unstable training as attention scores concentrate [10, 45]. To counteract this, techniques such as hierarchical fusion [59] and stabilizing attention entropy [80] have been proposed to prevent performance stagnation in vision transformers. DeepViT [84] proposes to re-generate diverse attention maps to avoid attention collapse in training vision transformers from scratch. In contrast, our work studies the expressiveness of the attention layer in the context of the large model fine-tuning. We also provide a novel interpretation of multi-head attention maps as a graph filter subspace, and propose the residual parameterization and regularization to stabilize the tuning process. + +# 3. Method + +# 3.1. A Filter Subspace View on Attention + +Attention. The attention layer transforms the input sequence $\mathbf{X} \in \mathbb{R}^{N \times C_i}$ to the output sequence $\mathbf{O} \in \mathbb{R}^{N \times C_o}$ , where $N$ denotes the sequence length, $C_i$ and $C_o$ are the dimension of input and output features. The attention layer projects the input as $\mathbf{Q} = \mathbf{X}\mathbf{W}_q$ , $\mathbf{K} = \mathbf{X}\mathbf{W}_k$ , $\mathbf{V} = \mathbf{X}\mathbf{W}_v$ using the corresponding projection matrices $\mathbf{W}_q, \mathbf{W}_k, \mathbf{W}_v \in \mathbb{R}^{C_i \times C_o}$ , and calculates the attention map, + +$$ +\begin{array}{l} \mathbf {A} = \operatorname {s o f t m a x} (\mathbf {Q K} ^ {\intercal}) \tag {1} \\ = \operatorname {s o f t m a x} \left(\mathbf {X} \mathbf {W} _ {q} \mathbf {W} _ {k} ^ {\intercal} \mathbf {X} ^ {\intercal}\right). \\ \end{array} +$$ + +The attention map $\mathbf{A} \in \mathbb{R}^{N \times N}$ captures the token-wise relationship by doing inner-product in a space transformed by + +$\mathbf{W}_q, \mathbf{W}_k$ . Then the output of the attention layer is, + +$$ +\begin{array}{l} \mathbf {O} = \operatorname {a t t n} \left(\mathbf {W} _ {q}, \mathbf {W} _ {k}, \mathbf {W} _ {v}; \mathbf {X}\right) \\ = \operatorname {s o f t m a x} \left(\mathbf {X} \mathbf {W} _ {q} \mathbf {W} _ {k} ^ {\intercal} \mathbf {X} ^ {\intercal}\right) \mathbf {X} \mathbf {W} _ {v} \tag {2} \\ = \mathbf {A X W} _ {v}. \\ \end{array} +$$ + +Multi-head Attention. Multi-head attention extends this by allowing multiple attention mechanisms, i.e., heads, to work in parallel, with each head independently learning attention patterns. For each attention head, we have, + +$$ +\mathbf {O} ^ {h} = \operatorname {a t t n} \left(\mathbf {W} _ {q} ^ {h}, \mathbf {W} _ {k} ^ {h}, \mathbf {W} _ {v} ^ {h}; \mathbf {X}\right), \tag {3} +$$ + +where $h = 1,\dots ,H$ and $H$ is the number of heads. The project matrices for each head are $\mathbf{W}_q^h,\mathbf{W}_k^h,\mathbf{W}_v^h\in$ $\mathbb{R}^{C_i\times \frac{C_o}{H}}$ . The multi-head attention represents as, + +$$ +\begin{array}{l} \mathbf {Z} = \operatorname {C o n c a t} \left[ \mathbf {O} ^ {1}, \dots , \mathbf {O} ^ {H} \right] \mathbf {W} _ {o} \\ = \underset {H} {\operatorname {C o n c a t}} \left[ \mathbf {A} ^ {1} \mathbf {X} \mathbf {W} _ {v} ^ {1}, \dots , \mathbf {A} ^ {H} \mathbf {X} \mathbf {W} _ {v} ^ {H} \right] \mathbf {W} _ {o} \tag {4} \\ = \sum_ {h = 1} ^ {H} \mathbf {A} ^ {h} \mathbf {X} \mathbf {W} _ {v o} ^ {h}, \\ \end{array} +$$ + +where $\mathbf{W}_o\in \mathbb{R}^{C_o\times C_o}$ and $\mathbf{W}_{vo}^h = \mathbf{W}_v^h (\mathbf{W}_o^h)^\intercal$ $\mathbf{W}_o^h\in$ $\mathbb{R}^{C_o\times \frac{C_o}{H}}$ $\mathbf{W}_{vo}^h\in \mathbb{R}^{C_i\times C_o}$ + +A Filter Subspace View on Multi-head Attention. We now present our filter subspace view on the multi-head attention. By viewing each token embedding $\mathbf{X}[n] \in \mathbb{R}^{C_i}$ as a node $V_{n}$ , which are connected with all the other nodes with edges $E[m, n] = 1$ , $\forall m \in [N]$ , we obtain a graph representation of all the tokens, $\mathcal{G} = (V, E)$ . + +We then reformulate the attention operation as the message passing as in Graph Convolution Nerworks [27]. Specifically, inspired by [6], we write the single-head attention as a graph convolution with the filter $\mathbf{F}^h (\mathbf{X})$ + +$$ +\mathbf {O} ^ {h} [ n,: ] = \sum_ {\substack {n ^ {\prime} \in [ N ] \\ c ^ {\prime} \in [ C _ {i} ]}} \mathbf {F} ^ {h} (\mathbf {X}) [ n, n ^ {\prime} ] \mathbf {X} [ n ^ {\prime}, c ^ {\prime} ] \mathbf {W} ^ {h} [ c ^ {\prime},,: ], \tag{5} +$$ + +where $\mathbf{W}^h = \mathbf{W}_{vo}^h\in \mathbb{R}^{C_i\times C_0}$ denotes the weight for linear projection on each node, and the $n$ -th row of the filter, i.e., $\mathbf{F}^h (\mathbf{X})[n,:]$ , represents the local graph filter around the $n$ -th node (token). Comparing with (1) and (2), we have, + +$$ +\mathbf {F} ^ {h} (\mathbf {X}) = \operatorname {s o f t m a x} \left(\mathbf {X} \mathbf {W} _ {q} \mathbf {W} _ {k} ^ {\intercal} \mathbf {X} ^ {\intercal}\right). \tag {6} +$$ + +The multi-head attention individually learns $H$ filters $\{\mathbf{F}^1 (X),\dots,\mathbf{F}^H (X)\}$ to aggregate complementary information across nodes. These $H$ filters naturally form a graph convolutional filter subspace $\mathcal{F}(\mathbf{X})$ as, + +$$ +\mathcal {F} (\mathbf {X}) = \left\{\mathbf {F} ^ {h} (\mathbf {X}) \right\} _ {h = 1} ^ {H} \subseteq (0, 1) ^ {N \times N}, \tag {7} +$$ + +and we denote each $\mathbf{F}^h (\mathbf{X})$ as a filter atom. The multi-head attention can then be represented in the way of decomposed graph convolution [6] as, + +$$ +\mathbf {O} [ n,: ] = \sum_ {h = 1} ^ {H} \sum_ {\substack {n ^ {\prime} \in [ N ] \\ c ^ {\prime} \in [ C _ {i} ]}} \mathbf {F} ^ {h} (\mathbf {X}) [ n, n ^ {\prime} ] \mathbf {X} [ n ^ {\prime}, c ^ {\prime} ] \mathbf {W} ^ {h} [ c ^ {\prime},: ]. \tag{8} +$$ + +It can be seen from (8) that the parameters $\mathbf{F}^h (\mathbf{X}),\mathbf{W}^h$ plays two key yet independent roles in the graph convolution. The $\mathbf{W}^h$ updates the knowledge for each node by projecting its representation into a new space, while the graph filter $\mathbf{F}^h (\mathbf{X})$ guides the information flow across nodes. Intuitively, with a more expressive graph convolutional filter subspace $\mathcal{F}(\mathbf{X})$ , each node can aggregate more informative features from other nodes. + +# 3.2. Subspace Coefficient Tuning + +Given the re-formulation of multi-head attention maps as a graph convolutional filter subspace, we present an efficient fine-tuning method specifically designed for transformers to enhance attention's expressiveness. Our motivation is that after the softmax function, each graph filter $\mathbf{F}^h (\mathbf{X})$ has values in the range (0, 1) with row-sum as one. The normalized attention maps restrict the expressiveness in the way that each output embedding lies in the convex hull of the value embeddings. + +To expand the feature space of the output embeddings, we propose Coeff-Tuning, to tune the filter subspace on the target dataset with a small set of combination coefficients, i.e., subspace coefficient $\alpha \in \mathbb{R}^{H\times H}$ . As shown in Figure 2, the tunable coefficient $\alpha$ constructs a more expressive + +filter subspace $\hat{\mathcal{F}}_{\alpha}(\mathbf{X})$ by linearly combining the pre-trained multi-head attention maps $\mathbf{F}(\mathbf{X})$ . Each element $\hat{\mathbf{F}}^h (\mathbf{X})$ in the learned filter subspace can be written as, + +$$ +\hat {\mathbf {F}} ^ {h} (\mathbf {X}) = \boldsymbol {\alpha} [ h, : ] ^ {\intercal} \mathcal {F} (\mathbf {X}) = \sum_ {i = 1} ^ {H} \boldsymbol {\alpha} [ h, i ] \mathbf {F} ^ {i} (\mathbf {X}). \tag {9} +$$ + +The multi-head attention then represents as, + +$$ +\mathbf {O} = \sum_ {h = 1} ^ {H} \hat {\mathbf {F}} ^ {h} (\mathbf {X}) \mathbf {X} \mathbf {W} ^ {h} = \sum_ {h = 1} ^ {H} \sum_ {i = 1} ^ {H} \alpha [ h, i ] \mathbf {F} ^ {i} (\mathbf {X}) \mathbf {X} \mathbf {W} ^ {h}. \tag {10} +$$ + +Note that, mathematically, the subspace coefficient $\alpha$ is unconstrained, allowing it to combine attention maps with a wider range of values from the original filter subspace $\mathbf{F}(\mathbf{X})$ . With attention scores exceeding the range $(0,1)$ , we effectively expand the space for $\mathbf{O}$ by allowing a more flexible combination of value matrices. + +Residual Parameterization. To facilitate the tuning and preserve the most knowledge in the pre-trained transformer, we propose a residual parameterization of the subspace coefficient, + +$$ +\boldsymbol {\alpha} ^ {\prime} = \boldsymbol {\alpha} + \mathbf {I}, +$$ + +$$ +\hat {\mathbf {F}} ^ {h} (\mathbf {X}) = \mathbf {F} ^ {h} (\mathbf {X}) + \sum_ {i = 1} ^ {H} \boldsymbol {\alpha} [ h, i ] \mathbf {F} ^ {i} (\mathbf {X}). \tag {11} +$$ + +In this way, we maintain the original filter subspace with the identity matrix, while parameterizing the subspace coefficient $\alpha'$ to learn the residual combination of attention maps on fine-tuning tasks. + +Reguralization on Subspace Coefficient. The proposed Coeff-Tuning substantially enhances the expressiveness of multi-head attention with tunable coefficient $\alpha$ . However, it may also lead to overfitting and preventing channel projection parameters $\mathbf{W}_v, \mathbf{W}_o$ from learning generalizable features. To alleviate this, we propose a simple regularization design that performs dropout [63] directly on the tuned subspace coefficient $\alpha$ during training, + +$$ +\boldsymbol {\alpha} ^ {\prime} = \operatorname {D r o p o u t} (\boldsymbol {\alpha}; p) + \mathbf {I}. \tag {12} +$$ + +In this way, we randomly remove an element from $\alpha$ with a probability of $p$ to avoid overfitting. + +# 4. Analysis + +In this section, we analyze that adjusting the filter subspace using the proposed subspace coefficient $\alpha$ can effectively enhance the expressiveness of the multi-head attention. + +![](images/4d56caa0f51e396ee56aad728933bff47314989d7280470ba025ee4dfe966f6e.jpg) +Figure 3. The effectiveness of tuning subspace coefficient $\alpha$ . The goal is to convert the input graph $\mathbf{X}$ to the target graph $\mathbf{O}$ using a single attention layer, $\mathbf{O} = \sum_{h=1}^{H} \mathbf{F}^h(\mathbf{X}) \mathbf{X} \mathbf{W}_v^h$ , where $\mathbf{F}(\mathbf{X}) = \text{softmax}(\mathbf{X} \mathbf{W}_q \mathbf{W}_k^\top \mathbf{X}^\top)$ . We have the following observations: (i) When only tuning $\mathbf{W}_q$ and $\mathbf{W}_k$ , i.e., updating $\mathbf{F}(\mathbf{X})$ , the positions of output nodes are constrained within the range of the input nodes $\mathbf{X}$ , since the values of $\mathbf{F}(\mathbf{X})$ are in the range of $(0,1)$ . (ii) In addition to tuning $\mathbf{F}(\mathbf{X})$ , updating $\mathbf{W}_v$ projects the input as $\mathbf{X} \mathbf{W}_v$ , but the output graph is still not aligned with the target graph. (iii) In addition to tuning $\mathbf{F}(\mathbf{X})$ , adapting $\alpha$ successfully transforms the input graph into the target graph. Compared with (i), $\alpha$ expand the positions of output nodes $\sum_{i=1}^{H} \alpha[;, i] \mathbf{F}^i(\mathbf{X}) \mathbf{X}$ beyond the constraints of the input nodes, enhancing the expressiveness of the attention effectively. + +For the filter $\mathbf{F}(\mathbf{X})$ in the filter subspace $\mathcal{F}$ since $\mathbf{F}(\mathbf{X}) = \mathrm{softmax}(\mathbf{X}\mathbf{W}_q\mathbf{W}_k^\top \mathbf{X}^\top)$ , their elements $\mathbf{F}(\mathbf{X})[n,n^{\prime}]$ satisfy the following properties, + +$$ +0 < \mathbf {F} (\mathbf {X}) [ n, n ^ {\prime} ] < 1, \sum_ {n ^ {\prime} \in [ N ]} \mathbf {F} (\mathbf {X}) [ n, n ^ {\prime} ] = 1. \tag {13} +$$ + +Definition 4.1. Given $N$ nodes $\{\mathbf{x}_i\}_{i=1}^N$ , the set $\mathcal{S}(\mathbf{x}) = \{\sum_{i=1}^{N} \lambda_i \mathbf{x}_i \mid \lambda_i > 0, \sum_i \lambda_i = 1\}$ is a bounded convex set. + +According to the Definition 4.1, when the projected nodes $\mathbf{X}\mathbf{W}^h\in \mathbb{R}^{N\times C_o}$ of each head $h$ are given, the output $\mathbf{O}^h [n,:] = \sum_{n'\in [N]}\mathbf{F}^h (\mathbf{X})[n,n']\mathbf{X}[n']\mathbf{W}^h$ lies within $S(\mathbf{X}\mathbf{W}^h)$ , which means the outputs $\mathbf{O}^h [n,:]$ are bounded. This property also applies to the multi-head attention mechanism. + +Proposition 4.2. If $S^h$ is a bounded convex set, $\forall h = 1, \dots, H$ , the set $S = S^1 + \dots + S^H = \{ \mathbf{x}^1 + \dots + \mathbf{x}^H \mid \forall \mathbf{x}^h \in S^h \}$ is also a bounded convex set. + +According to Proposition 4.2, when the projected nodes $\mathbf{X}\mathbf{W}^h$ are given, the outputs of the multi-head attention $\mathbf{O}[n,:] = \sum_{h=1}^{H} \sum_{n' \in [N]} \mathbf{F}^h(\mathbf{X})[n,n']\mathbf{X}[n']\mathbf{W}^h$ are also bounded. + +Proposition 4.3. After applying the subspace coefficients $\alpha$ , the outputs of the reconstructed filter subspace $\hat{\mathcal{F}}_{\alpha}$ form a set $\hat{S} = \{\sum_{h=1}^{H} \sum_{i=1}^{H} \boldsymbol{\alpha}[h,i]\mathbf{F}^i(\mathbf{X})[n,n']\mathbf{X}\mathbf{W}^h \mid \mathbf{F}^i(\mathbf{X})[n,n'] > 0, \sum_{j} \mathbf{F}^i(\mathbf{X})[n,n'] = 1, \boldsymbol{\alpha}[h,i] \in \mathbb{R}\}$ , which satisfies $S \subseteq \hat{S}$ . + +According to Proposition 4.3, when applying filter subspace coefficients $\alpha$ , we can reconstruct the filter subspace $\mathcal{F}$ and expand the set $\mathcal{S}$ . Specifically, given $\mathbf{X}\mathbf{W}^h$ , the output $\mathbf{O}[n,:] = \sum_{h=1}^{H} \sum_{i=1}^{H} \alpha[h,i]\mathbf{F}^i(\mathbf{X})[n,n']\mathbf{X}\mathbf{W}^h$ can be in the set $\hat{\mathcal{S}} \setminus \mathcal{S}$ . Our method obtains an expanded set, enhancing the expressiveness of the model. We provide the support of our analyses in Appendix A. + +# 4.1. Validation Experiment + +We assess the effectiveness of the subspace coefficients using a single attention layer. Given an input graph $\mathbf{X}$ , the output of the attention layer is calculated as $\mathbf{O} = \mathbf{F}(\mathbf{X})\mathbf{X}\mathbf{W}_v$ , where $\mathbf{F}(\mathbf{X}) = \mathrm{softmax}(\mathbf{X}\mathbf{W}_q\mathbf{W}_k^\intercal\mathbf{X}^\intercal)$ . + +Our goal is to optimize $\mathbf{W}_q$ , $\mathbf{W}_k$ , $\mathbf{W}_v$ to transform the input graph $\mathbf{X}$ into the target graph $\mathbf{O}$ as shown in Figure 3. The input graph contains 8 nodes $\mathbf{X} = [[-1,1],[0,1],[1,1],[1,0],[1,-1],[0,-1],[-1,-1],[-1,0]]$ positioned in a range of $[-1,1]$ and the target graph $\mathbf{O} = [[-1,1],[0,2],[1,1],[2,0],[1,-1],[0,-2],[-1,-1],[-2,0]]$ , with node coordinates in the range of $[-2,2]$ . + +As shown in Figure 3, by only tuning the $\mathbf{W}_q$ and $\mathbf{W}_k$ , i.e., $\mathbf{F}(\mathbf{X}) = \mathrm{softmax}(\mathbf{X}\mathbf{W}_q\mathbf{W}_k^\top \mathbf{X}^\top)$ , the output is constrained inside the range of the input, which is aligned with Proposition 4.1. When we tune $\mathbf{W}_q$ , $\mathbf{W}_k$ , and $\mathbf{W}_v$ , the output exceeds the input range but still cannot fully match the target graph. However, when we tune the subspace $\mathbf{F}(\mathbf{X})$ and subspace coefficient $\alpha$ , the attention layer can successfully transform the input graph into the target graph as in Figure 3. This observation is aligned with Proposition 4.3, and indicates that the proposed Coeff-Tuning can effectively expand the feature space of the multi-head attention layer. + +# 4.2. Complexity Analysis + +We now formally analyze the complexity of the proposed Coeff-Tuning, and compare it with other baseline PEFT methods. The number of parameters required for our method is $H \times H$ , where $H$ is the number of heads at each attention layer. Considering $\mathbf{W}_q, \mathbf{W}_k, \mathbf{W}_v \in \mathbb{R}^{C_i \times C_o}$ and $\mathbf{W}_o \in \mathbb{R}^{C_o \times C_o}$ of the attention layer, our method introduces relative $\frac{H^2}{3C_iC_o + C_o^2}$ parameters. In modern large transformer models, we typically have $C_i = 2048$ and $C_o = 1280$ . With $H = 10$ , our method requires approximately 0.001% additional parameters. Compared with other PEFT methods such as LoRA, our method introduces + +Table 1. Fine-tuning Results on VTAB-1k with ViT + +
NaturalSpecializedStructured
C100Cal101DTDF102PetsSVHNS397CameE.SATRes45ReinoCle_ctmCle_disDMLabKITTIdSlocationdS oriS_aziS_eleMeanParam.GTL_OPs
Full fine-tuning68.987.764.397.286.987.438.879.795.784.273.956.358.641.765.557.546.725.729.165.5785.845.15
Adapter74.186.163.297.787.034.650.876.388.073.170.545.737.431.253.230.325.413.822.155.820.23416.44
BitFit72.887.059.297.585.359.951.478.791.672.969.861.555.632.455.966.640.015.725.162.050.10515.08
VPT77.786.962.697.587.374.551.278.292.075.672.950.558.640.567.168.736.120.234.164.850.0616.68
LoRA67.191.469.498.890.485.354.084.995.384.473.682.969.249.881.781.848.332.844.272.910.30515.74
SPT-Adapter72.993.272.599.391.484.655.285.396.083.975.882.268.049.380.082.451.931.741.273.520.26616.44
Linear Probing63.485.063.297.086.336.651.078.587.568.674.034.330.633.255.412.520.09.619.252.94015.05
Coeff. α Only (Ours)69.590.874.599.190.983.051.183.695.384.675.857.459.146.578.675.051.425.634.269.780.00215.06
FACT-TT≤1671.389.670.798.991.087.854.685.295.583.475.782.069.049.880.079.248.434.241.473.040.04015.14
w/ Coeff-Tuning (Ours)72.189.771.899.191.288.954.885.395.784.275.383.068.851.081.380.049.335.042.773.640.04215.15
SNELL-873.792.772.499.291.489.255.484.996.186.475.284.068.553.581.082.749.933.939.274.170.26815.47
w/ Coeff-Tuning (Ours)74.292.973.199.291.690.355.685.296.286.375.884.168.853.781.384.052.135.039.974.700.27015.48
+ +relative $\frac{H^2}{3rC_i + 5rC_o}$ parameters, where $r$ is the LoRA rank. Using the same values for $C_i$ , $C_o$ , $H$ , and setting $r = 4$ , our approach requires less than $0.2\%$ additional parameters compared to LoRA and its variants. This also indicates that our Coeff-Tuning can maintain almost the same quantity of parameters when combined with other PEFT methods. + +# 5. Experiments + +In this section, we validate the proposed Coeff-Tuning with various tasks and datasets, including tuning the vision transformer [11] on few-shot image classification, tuning text-to-image diffusion model [51] for concept customization, and tuning vision-language BART [8] on image-to-text understanding. Combining with other weight-based PEFT methods, the proposed Coeff-Tuning consistently achieves superior results with negligible parameters added. + +# 5.1. Few-shot Image Classification + +we first validate the proposed method by fine-tuning Vision Transformers on a series of downstream datasets for few-shot image classification. Our method achieves superior results while maintaining a high parameter efficiency compared with baseline methods. + +Experimental Settings. We utilize the VTAB-1k [81] dataset which comprises 19 diverse computer vision tasks across three domains: standard natural imagery, specialized capture, and synthetic/structured environments. The benchmark evaluates model performance on various task + +Table 2. Fine-tuning Results on the FGVC with ViT. + +
CUB-200NABirdsFlowersS-DogsS-CarsMean
Full fine-tuning87.382.798.889.484.588.5
Linear Probing85.375.997.986.251.379.3
Adapter87.384.398.488.868.485.5
BitFit88.484.298.891.279.488.4
VPT86.778.898.490.768.784.6
LoRA84.979.098.188.179.886.0
SPT-LoRA88.683.499.591.487.390.1
SNELL-889.686.899.392.189.991.5
w/ Coeff-Tuning (Ours)89.887.199.592.290.791.9
+ +objectives. With only 1,000 training samples per task, it presents a significant few-shot learning challenge. We also benchmark our method on five fine-grained datasets in few-shot settings, including CUB-200-2011 [69], NABirds [67], Oxford Flowers [44], Stanford Dogs [25] and Stanford Cars [13]. A detailed description is provided in Appendix B.2. + +As for the vision transformer, we follow VPT [23], SSF [35], and select ViT-B/16 [11] pre-trained on ImageNet-21K [9] as the base model for fine-tuning. We benchmark the proposed Coeff-Tuning with a series of PEFT methods for few-shot learning including SSF [35], VPT [23], Adapter [20], LoRA [21], and BitFit [79]. + +We present two settings of Coeff-Tuning in this task. We first explore the representation ability of tuning subspace coefficient $\alpha$ only, which tunes the transformer bottleneck with only tuning $H^2$ (144 for ViT [11]) parameters for each transformer block. We also show the performance of combining Coeff-Tuning and SSF [35] to demonstrate that our method can plug and play with other PEFT methods. We set the coefficient dropout rates as $p = 0$ and $p = 0.1$ for these two settings respectively, and provide other hyperparameters in Appendix B.2. + +Results. As shown in Table 1, the proposed Coeff-Tuning shows superior performances in both settings. Compared with the linear probing baseline which achieves a mean accuracy of 61.69, Coeff-Tuning obtains a significantly higher result of 73.49 with tuning only $(12 * 12) * 12 = 1$ , 728 additional subspace coefficients $\alpha$ in ViT [11] (ViT [11] has 12 attention layers, each attention has 12 heads). Thus, we achieve a large performance increase in few-shot image classification $(11.8^{\uparrow})$ with negligible additional parameter costs. Note that this result has already surpassed almost all baselines. We further achieve the best performance by combining Coeff-Tuning with SSF while maintaining almost the same parameter cost. + +![](images/2941f246d25d534a5c8707518a81d42e9c918146ed65822cd22a66b4a3805b00.jpg) +Figure 4. Results on Concept Customization. Compared to the baseline method (middle row), Coeff-Tuning (bottom row) leads to higher concept fidelity and text alignment of the generated images. For instance, our method accurately generates images of a “ $\langle V \rangle$ plushie sloth in the snow,” while the baseline fails to align with the text prompt. Additionally, our method generates a “ $\langle V \rangle$ wooden pot” with higher fidelity to the input image. + +Table 3. Concept Customization Results with SDXL. + +
Method# Params (%)Text Align.Concept Align.Diversity
DoRA0.03710.330.7310.9
Ours + DoRA0.03740.340.7511.37
LoRA0.03060.330.7410.08
Ours + LoRA0.03090.340.7510.45
+ +# 5.2. Personalized Text-to-Image Generation + +We now show the effectiveness of Coeff-Tuning with the concept customization with Stable Diffusion (SDXL) [51, 55]. We validate the effectiveness of Coeff-Tuning by combining it with baseline PEFT methods, LoRA [21] and DoRA [38], and compare the performance with baselines without tuning $\alpha$ . Specifically, we fine-tune SDXL [51] to associate the special token with a concept image from Dreambooth [56] and CustomConcept101 [30] datasets. We choose 10 different concepts, and fine-tune the model on $4 \sim 9$ images. We utilize 25 different text prompts following [53] and generate 4 images for each prompt. We evaluate the fidelity, diversity, and text-to-image alignment of the generated images. Other training details are described in Appendix B.3. + +Evaluation metrics. We assess the concept alignment with the average cosine similarity between CLIP embeddings [54] of the generated images and the concept images. The diversity of generated images is measured by the Vendi score [12] calculated with the DINOv2 embeddings [46]. + +The text alignment of the generated images is measured via average cosine similarity in the CLIP feature space [54]. + +Results. The quantitative results are shown in Table 3. By combining with LoRA [21] or DoRA [38], the proposed Coeff-Tuning generates images with greater diversity and better alignment with the both text prompt and input concept images. The qualitative results of our method compared with LoRA [21] are shown in Figure 4. The generated images with Coeff-Tuning have higher concept fidelity, preserving more characteristics from the input images. + +As an example, consider the concept “ $\langle V \rangle$ cat” as shown in the 4th column in Figure 4, the image generated using our method retains the cat's color from the input image, while the one generated by the baseline method alter the color. Additionally, the backgrounds in images generated with our method show more detailed features compared to those from the baseline method. Coeff-Tuning achieves the above improvements by introducing a limited number of additional parameters. We provide more generated images in Appendix B.3. + +# 5.3. Image-Text Understanding + +As we have shown the effectiveness of Coeff-Tuning in fine-tuning models on both few-shot image classification and personalized image generation, we now validate it on multitask image-text understanding, which is a much larger-scale multi-modal fine-tuning task that requires a higher model capacity. By combining with baseline PEFT meth + +Table 4. Multi-Model Fine-tuning Results with VL-Bart. + +
Method# Params (%)VQAV2GQANVLR2COCO CapAvg.
Full Model10066.956.773.7112.077.3
LoRA5.9465.253.671.9115.376.5
Ours + LoRA5.9466.254.173.0115.277.1
DoRA5.9665.854.773.1115.977.4
Ours + DoRA5.9666.454.973.0116.477.7
+ +ods, our method again achieves superior performance with negligible parameter increase. + +Experimental Settings. We evaluate our method on VL-BART [8] multi-modal transformer, which utilizes CLIP-ResNet101 [54] for vision encoding and sends the visual features to BART-Base [31] for image-text joint modeling. Following [65], we evaluate our method with the multitask fine-tuning of VL-Bart [8] on four large-scale image-text understanding tasks simultaneously: VQAv2 [14] and GQA [22] for visual question answering, NLVR2 [64] for visual reasoning, and MSCOCO [5] for image captioning. Following the plug-and-play practice, we combine the proposed Coeff-Tuning with LoRA [21] and DoRA [38] respectively, and set the subspace coefficient dropout rate to $p = 0.2$ . Detailed descriptions of other settings and hyperparameters are provided in Appendix B.4. + +Results. As shown in Table 4, our method surpasses baseline LoRA [21] and DoRA [38] respectively in almost all image-text datasets consistently, and achieves higher average scores. The Coeff-Tuning combined with LoRA [21] obtains comparable results with full-model fine-tuning, and it achieves better results than full-model tuning when it is combined with DoRA [38]. Notably, these improvements are achieved while maintaining the same parameter budgets as we only tune a hundred parameters for each attention. + +# 5.4. Ablation Study + +In this section, we examine the effectiveness of each component of our design with ablation experiments. + +Table 5. Ablation studies on (a) residual parameterization and (b) coefficient dropout showing their effects on model performance. + +
ParametrizationAvg.
α w/ rand. init.44.28
α w/ identity init.61.57
I + α, zero init.73.49
+ +(a) + +
Dropout pAvg. Score
0.075.1
0.177.3
0.277.7
0.472.3
+ +(b) + +Residual Parameterization. We first demonstrate the importance of residual parameterization in Eq. 11 in Coeff-Tuning. We compare our residual parameterization with the direct parameterization where we directly tune $\alpha$ without adding it to the identity matrix, and initialize it with either a random matrix or an identity matrix. We adopt the few-shot tuning task as in Sec. 5.1 and only tune $\alpha$ . Results in Table 5a show that models with residual parameterization achieve superior performance than direct parameterization with two different initializations. This validates our design that maintaining the original filter subspace while learning the residual combination helps preserve pre-trained knowledge and leads to better fine-tuning performance. + +Subspace Coefficient Dropout. The dropout regularization on the subspace coefficient as in Eq. 12 plays a crucial role in preventing overfitting. We adopt the experiment setting in multi-modal fine-tuning as in Sec. 5.3. We experiment with different dropout rates $p \in \{0, 0.1, 0.2, 0.4\}$ during training. The results in Table 5b demonstrate that a moderate dropout rate ( $p = 0.2$ ) yields the best performance while either a lower or higher dropout rate leads to sub-optimal results. This suggests that applying dropout directly on the tuned subspace coefficient $\alpha$ can effectively prevent overfitting and lead to better generalization. + +# 6. Conclusion + +In this paper, we presented Coeff-Tuning, a new parameter-efficient fine-tuning approach that enhances multi-head attention's expressiveness by learning a set of subspace coefficients. By reformulating attention as a graph convolution with a filter subspace, we demonstrated both theoretically and empirically that our method expands the feature space of multi-head attention beyond the original convex hull constraints. The proposed Coeff-Tuning can be seamlessly integrated with existing PEFT methods while introducing negligible additional parameters. Through extensive experiments across diverse tasks and modalities, we consistently demonstrated superior performance compared to competitive PEFT baselines. Our method achieved significant improvements in various metrics while maintaining high parameter efficiency, suggesting a promising direction for enhancing the adaptability of large pre-trained transformers. + +# References + +[1] Tom B Brown. Language models are few-shot learners. arXiv preprint arXiv:2005.14165, 2020. 2 +[2] Shoufa Chen, Chongjian Ge, Zhan Tong, Jiangliu Wang, Yibing Song, Jue Wang, and Ping Luo. Adaptformer: Adapting vision transformers for scalable visual recognition. Advances in Neural Information Processing Systems, 2022. 2 +[3] Wei Chen, Zichen Miao, and Qiang Qiu. Inner product-based neural network similarity. Advances in Neural Information Processing Systems, 36, 2024. 1 +[4] Wei Chen, Zichen Miao, and Qiang Qiu. Large convolutional model tuning via filter subspace. arXiv preprint arXiv:2403.00269, 2024.2, 1 +[5] Xinlei Chen, Hao Fang, Tsung-Yi Lin, Ramakrishna Vedantam, Saurabh Gupta, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco captions: Data collection and evaluation server. arXiv preprint arXiv:1504.00325, 2015. 8 +[6] Xiuyuan Cheng, Zichen Miao, and Qiang Qiu. Graph convolution with low-rank learnable local filters. arXiv preprint arXiv:2008.01818, 2020. 4 +[7] Xiuyuan Cheng, Zichen Miao, and Qiang Qiu. Graph convolution with low-rank learnable local filters. In International Conference on Learning Representations, 2021. 1 +[8] Jaemin Cho, Jie Lei, Hao Tan, and Mohit Bansal. Unifying vision-and-language tasks via text generation. In International Conference on Machine Learning, pages 1931-1942. PMLR, 2021. 2, 6, 8 +[9] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 6, 1 +[10] Yihe Dong, Jean-Baptiste Cordonnier, and Andreas Loukas. Attention is not all you need: Pure attention loses rank doubly exponentially with depth. In International Conference on Machine Learning, pages 2793-2803. PMLR, 2021. 2, 3 +[11] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. ICLR, 2021. 1, 2, 6 +[12] Dan Friedman and Adji Bousso Dieng. The vendi score: A diversity evaluation metric for machine learning. arXiv preprint arXiv:2210.02410, 2022. 7 +[13] Timnit Gebru, Jonathan Krause, Yilun Wang, Duyun Chen, Jia Deng, and Li Fei-Fei. Fine-grained car detection for visual census estimation. In Proceedings of the AAAI Conference on Artificial Intelligence, 2017. 6 +[14] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Bartra, and Devi Parikh. Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6904-6913, 2017. 2, 8 +[15] Ligong Han, Yinxiao Li, Han Zhang, Peyman Milanfar, Dimitris Metaxas, and Feng Yang. Svdiff: Compact parameter space for diffusion fine-tuning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7323-7334, 2023. 1 +[16] Liu He and Daniel Aliaga. Globalmapper: Arbitrary-shaped urban layout generation. In Proceedings of the IEEE/CVF + +International Conference on Computer Vision, pages 454-464, 2023. 1 +[17] Liu He and Daniel Aliaga. Coho: Context-sensitive city-scale hierarchical urban layout generation. In European Conference on Computer Vision, pages 1-18. Springer, 2024. +[18] Liu He, Yijuan Lu, John Corring, Dinei Florencio, and Cha Zhang. Diffusion-based document layout generation. In International Conference on Document Analysis and Recognition, pages 361-378. Springer, 2023. +[19] Liu He, Yizhi Song, Hejun Huang, Daniel Aliaga, and Xin Zhou. Kubrick: Multimodal agent collaborations for synthetic video generation. arXiv preprint arXiv:2408.10453, 2024. 1 +[20] Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin De Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. Parameter-efficient transfer learning for nlp. In International conference on machine learning, pages 2790-2799. PMLR, 2019. 2, 6 +[21] Edward J Hu, Phillip Wallis, Zeyuan Allen-Zhu, Yuzhhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. In International Conference on Learning Representations, 2021. 1, 2, 6, 7, 8 +[22] Drew A Hudson and Christopher D Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6700-6709, 2019. 2, 8 +[23] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In European Conference on Computer Vision, pages 709-727. Springer, 2022. 2, 6, 1 +[24] Rabeeh Karimi Mahabadi, James Henderson, and Sebastian Ruder. Compacter: Efficient low-rank hypercomplex adapter layers. Advances in Neural Information Processing Systems, 2021. 2, 1 +[25] Aditya Khosla, Nityananda Jayadevaprakash, Bangpeng Yao, and Fei-Fei Li. Novel dataset for fine-grained image categorization: Stanford dogs. In Proc. CVPR workshop on fine-grained visual categorization (FGVC), 2011. 6 +[26] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In International Conference on Learning Representations, 2015. 1 +[27] Thomas N Kipf and Max Welling. Semi-supervised classification with graph convolutional networks. arXiv preprint arXiv:1609.02907, 2016. 4 +[28] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 1 +[29] Dawid J Kopiczko, Tijmen Blankevoort, and Yuki M Asano. Vera: Vector-based random matrix adaptation. *ICLR*, 2024. 2 +[30] Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization of text-to-image diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 7, 1 + +[31] M Lewis. Bart: Denoising sequence-to-sequence pretraining for natural language generation, translation, and comprehension. arXiv preprint arXiv:1910.13461, 2019. 2, 8 +[32] Junbo Li, Zichen Miao, Qiang Qiu, and Ruqi Zhang. Training bayesian neural networks with sparse subspace variational inference. arXiv preprint arXiv:2402.11025, 2024. 1 +[33] Tianchun Li, Chengxiang Wu, Pengyi Shi, and Xiaoqian Wang. Cumulative difference learning vae for time-series with temporally correlated inflow-outflow. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 13619-13627, 2024. 1 +[34] Xiang Lisa Li and Percy Liang. Prefix-tuning: Optimizing continuous prompts for generation. In Proceedings of the Association for Computational Linguistics, 2021. 2 +[35] Dongze Lian, Daquan Zhou, Jiashi Feng, and Xinchao Wang. Scaling & shifting your features: A new baseline for efficient model tuning. Advances in Neural Information Processing Systems, 35:109-123, 2022. 2, 6, 1 +[36] Haokun Liu, Derek Tam, Mohammed Muqeeth, Jay Mohta, Tenghao Huang, Mohit Bansal, and Colin A Raffel. Few-shot parameter-efficient fine-tuning is better and cheaper than in-context learning. Advances in Neural Information Processing Systems, 35:1950-1965, 2022. 2 +[37] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 2 +[38] Shih-Yang Liu, Chien-Yi Wang, Hongxu Yin, Pavlo Molchanov, Yu-Chiang Frank Wang, Kwang-Ting Cheng, and Min-Hung Chen. Dora: Weight-decomposed low-rank adaptation. arXiv preprint arXiv:2402.09353, 2024. 2, 7, 8, 1 +[39] Tianci Liu, Ruirui Li, Yunzhe Qi, Hui Liu, Xianfeng Tang, Tianqi Zheng, Qingyu Yin, Monica Xiao Cheng, Jun Huan, Haoyu Wang, et al. Unlocking efficient, scalable, and continual knowledge editing with basis-level representation finetuning. arXiv preprint arXiv:2503.00306, 2025. 1 +[40] Zichen Miao, Ze Wang, Wei Chen, and Qiang Qiu. Continual learning with filter atom swapping. In International Conference on Learning Representations, 2021. 1 +[41] Zichen Miao, Ze Wang, Xiuyuan Cheng, and Qiang Qiu. Spatiotemporal joint filter decomposition in 3d convolutional neural networks. Advances in Neural Information Processing Systems, 34:3376-3388, 2021. 1 +[42] Zichen Miao, Jiang Wang, Ze Wang, Zhengyuan Yang, Lijuan Wang, Qiang Qiu, and Zicheng Liu. Training diffusion models towards diverse image generation with reinforcement learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10844-10853, 2024. 1 +[43] Zichen Miao, Zhengyuan Yang, Kevin Lin, Ze Wang, Zicheng Liu, Lijuan Wang, and Qiang Qiu. Tuning timestep-distilled diffusion model using pairwise sample optimization. arXiv preprint arXiv:2410.03190, 2024. 1 +[44] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In 2008 Sixth Indian conference on computer vision, graphics & image processing, pages 722-729. IEEE, 2008. 6 + +[45] Lorenzo Noci, Sotiris Anagnostidis, Luca Biggio, Antonio Orvieto, Sidak Pal Singh, and Aurelien Lucchi. Signal propagation in transformers: Theoretical perspectives and the role of rank collapse. Advances in Neural Information Processing Systems, 35:27198-27211, 2022. 3 +[46] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 7 +[47] Gaurav Patel and Qiang Qiu. Learning to unlearn while retaining: Combating gradient conflicts in machine unlearning. arXiv preprint arXiv:2503.06339, 2025. 1 +[48] Gaurav Patel, Konda Reddy Mopuri, and Qiang Qiu. Learning to retain while acquiring: Combating distribution-shift in adversarial data-free knowledge distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7786-7794, 2023. +[49] Gaurav Patel, Christopher Sandino, Behrooz Mahasseni, Ellen L Zippi, Erdrin Azemi, Ali Moin, and Juri Minxha. Efficient source-free time-series adaptation via parameter subspace disentanglement. arXiv preprint arXiv:2410.02147, 2024. 1 +[50] Jonas Pfeiffer, Aishwarya Kamath, Andreas Rücklé, Kyunghyun Cho, and Iryna Gurevych. Adapterfusion: Nondestructive task composition for transfer learning. arXiv preprint arXiv:2005.00247, 2020. 2 +[51] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. *ICLR*, 2024. 1, 6, 7 +[52] Qiang Qiu, Xiuyuan Cheng, Guillermo Sapiro, et al. Dcfnet: Deep neural network with decomposed convolutional filters. In International Conference on Machine Learning, pages 4198-4207. PMLR, 2018. 1 +[53] Zeju Qiu, Weiyang Liu, Haiwen Feng, Yuxuan Xue, Yao Feng, Zhen Liu, Dan Zhang, Adrian Weller, and Bernhard Scholkopf. Controlling text-to-image diffusion by orthogonal finetuning. Advances in Neural Information Processing Systems, 36:79320-79362, 2023. 2, 7 +[54] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, 2021. 2, 7, 8 +[55] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10684-10695, 2022. 1, 2, 7 +[56] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 22500-22510, 2023. 2, 7, 1 +[57] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, + +Aditya Khosla, Michael Bernstein, Alexander C. Berg, and Li Fei-Fei. ImageNet Large Scale Visual Recognition Challenge. International Journal of Computer Vision (IJCV), 2015. 2 +[58] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278-25294, 2022. 2 +[59] Han Shi, Jiahui Gao, Hang Xu, Xiaodan Liang, Zhenguo Li, Lingpeng Kong, Stephen Lee, and James T Kwok. Revisiting over-smoothing in bert from the perspective of graph. arXiv preprint arXiv:2202.08625, 2022. 2, 3 +[60] Kihyuk Sohn, Huiwen Chang, José Lezama, Luisa Polania, Han Zhang, Yuan Hao, Irfan Essa, and Lu Jiang. Visual prompt tuning for generative transfer learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 2 +[61] Yizhi Song, Liu He, Zhifei Zhang, Soo Ye Kim, He Zhang, Wei Xiong, Zhe Lin, Brian Price, Scott Cohen, Jianming Zhang, et al. Refine-by-align: Reference-guided artifacts refinement through semantic alignment. arXiv preprint arXiv:2412.00306, 2024. 1 +[62] Yizhi Song, Zhifei Zhang, Zhe Lin, Scott Cohen, Brian Price, Jianming Zhang, Soo Ye Kim, He Zhang, Wei Xiong, and Daniel Aliaga. Imprint: Generative object compositing by learning identity-preserving representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8048-8058, 2024. 1 +[63] Nitish Srivastava, Geoffrey Hinton, Alex Krizhevsky, Ilya Sutskever, and Ruslan Salakhutdinov. Dropout: a simple way to prevent neural networks from overfitting. The journal of machine learning research, 15(1):1929-1958, 2014. 4 +[64] Alane Suhr, Stephanie Zhou, Ally Zhang, Iris Zhang, Hua-jun Bai, and Yoav Artzi. A corpus for reasoning about natural language grounded in photographs. arXiv preprint arXiv:1811.00491, 2018. 8 +[65] Yi-Lin Sung, Jaemin Cho, and Mohit Bansal. Vl-adapter: Parameter-efficient transfer learning for vision-and-language tasks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5227-5237, 2022. 8 +[66] Gemma Canet Tarres, Zhe Lin, Zhifei Zhang, Jianming Zhang, Yizhi Song, Dan Ruta, Andrew Gilbert, John Colomosse, and Soo Ye Kim. Thinking outside the bbox: Unconstrained generative object compositing. arXiv preprint arXiv:2409.04559, 2024. 1 +[67] Grant Van Horn, Steve Branson, Ryan Farrell, Scott Haber, Jessie Barry, Panos Ipeirotis, Pietro Perona, and Serge Belongie. Building a bird recognition app and large scale dataset with citizen scientists: The fine print in fine-grained dataset collection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 595-604, 2015. 6 +[68] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 1, 2 + +[69] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The caltech-ucsd birds-200-2011 dataset. 2011. 6 +[70] Haoyu Wang, Tianci Liu, Ruirui Li, Monica Cheng, Tuo Zhao, and Jing Gao. Roselora: Row and column-wise sparse low-rank adaptation of pre-trained language model for knowledge editing and fine-tuning. arXiv preprint arXiv:2406.10777, 2024. 1 +[71] Peihao Wang, Wenqing Zheng, Tianlong Chen, and Zhangyang Wang. Anti-oversmoothing in deep vision transformers via the fourier domain analysis: From theory to practice. arXiv preprint arXiv:2203.05962, 2022. 2 +[72] Ze Wang, Xiuyuan Cheng, Guillermo Sapiro, and Qiang Qiu. Stochastic conditional generative networks with basis decomposition. arXiv preprint arXiv:1909.11286, 2019. 1 +[73] Ze Wang, Seunghyun Hwang, Zichen Miao, and Qiang Qiu. Image generation using continuous filter atoms. Advances in Neural Information Processing Systems, 34:17826-17838, 2021. +[74] Ze Wang, Zichen Miao, Jun Hu, and Qiang Qiu. Adaptive convolutions with per-pixel dynamic filter atom. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12302-12311, 2021. 1 +[75] Ze Wang, Zichen Miao, Xiantong Zhen, and Qiang Qiu. Learning to learn dense gaussian processes for few-shot learning. Advances in Neural Information Processing Systems, 34:13230-13241, 2021. 1 +[76] Enze Xie, Lewei Yao, Han Shi, Zhili Liu, Daquan Zhou, Zhaoqiang Liu, Jiawei Li, and Zhenguo Li. Difffit: Unlocking transferability of large diffusion models via simple parameter-efficient fine-tuning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4230-4239, 2023. 1 +[77] Zhexiao Xiong, Wei Xiong, Jing Shi, He Zhang, Yizhi Song, and Nathan Jacobs. Groundingbooth: Grounding text-to-image customization. arXiv preprint arXiv:2409.08520, 2024.1 +[78] SHIH-YING YEH, Yu-Guan Hsieh, Zhidong Gao, Bernard BW Yang, Gigeong Oh, and Yanmin Gong. Navigating text-to-image customization: From lycoris fine-tuning to model evaluation. In International Conference on Learning Representations, 2023. 2 +[79] Elad Ben Zaken, Yoav Goldberg, and Shauli Ravfogel. Bitfit: Simple parameter-efficient fine-tuning for transformer-based masked language-models. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics, 2022. 2, 6 +[80] Shuangfei Zhai, Tatiana Likhomanenko, Etai Littwin, Dan Busbridge, Jason Ramapuram, Yizhe Zhang, Jiatao Gu, and Joshua M Susskind. Stabilizing transformer training by preventing attention entropy collapse. In International Conference on Machine Learning, pages 40770-40803. PMLR, 2023. 2, 3 +[81] Xiaohua Zhai, Joan Puigcerver, Alexander Kolesnikov, Pierre Ruyssen, Carlos Riquelme, Mario Lucic, Josip Djolonga, Andre Susano Pinto, Maxim Neumann, Alexey Dosovitskiy, et al. A large-scale study of representation learning + +with the visual task adaptation benchmark. arXiv preprint arXiv:1910.04867, 2019. 6, 1 +[82] Qingru Zhang, Minshuo Chen, Alexander Bukharin, Nikos Karampatziakis, Pengcheng He, Yu Cheng, Weizhu Chen, and Tuo Zhao. Adalora: Adaptive budget allocation for parameter-efficient fine-tuning. *ICLR*, 2022. 2 +[83] Yibo Zhong, Haoxiang Jiang, Lincan Li, Ryumei Nakada, Tianci Liu, Linjun Zhang, Huaxiu Yao, and Haoyu Wang. Neat: Nonlinear parameter-efficient adaptation of pretrained models. arXiv preprint arXiv:2410.01870, 2024. 1 +[84] Daquan Zhou, Bingyi Kang, Xiaojie Jin, Linjie Yang, Xiaochen Lian, Zihang Jiang, Qibin Hou, and Jiashi Feng. Deepvit: Towards deeper vision transformer. arXiv preprint arXiv:2103.11886, 2021. 2, 3 \ No newline at end of file diff --git a/2025/Coeff-Tuning_ A Graph Filter Subspace View for Tuning Attention-Based Large Models/images.zip b/2025/Coeff-Tuning_ A Graph Filter Subspace View for Tuning Attention-Based Large Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..dfc6ff74d52e02e3d884a8373e347b929f02377a --- /dev/null +++ b/2025/Coeff-Tuning_ A Graph Filter Subspace View for Tuning Attention-Based Large Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:da73e3da6cbb3eaed7031d9c276a8fe198e80aac00840b67915bfd2de06900a7 +size 610880 diff --git a/2025/Coeff-Tuning_ A Graph Filter Subspace View for Tuning Attention-Based Large Models/layout.json b/2025/Coeff-Tuning_ A Graph Filter Subspace View for Tuning Attention-Based Large Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..1a629d1b63598bee642d6aecebc2aa6710330f46 --- /dev/null +++ b/2025/Coeff-Tuning_ A Graph Filter Subspace View for Tuning Attention-Based Large Models/layout.json @@ -0,0 +1,11501 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 59, + 102, + 550, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 102, + 550, + 138 + ], + "spans": [ + { + "bbox": [ + 59, + 102, + 550, + 138 + ], + "type": "text", + "content": "Coeff-Tuning: A Graph Filter Subspace View for Tuning Attention-Based Large Models" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 212, + 160, + 396, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 212, + 160, + 396, + 190 + ], + "spans": [ + { + "bbox": [ + 212, + 160, + 396, + 190 + ], + "type": "text", + "content": "Zichen Miao†, Wei Chen†, Qiang Qiu Purdue University, IN, USA" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 212, + 191, + 396, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 212, + 191, + 396, + 204 + ], + "spans": [ + { + "bbox": [ + 212, + 191, + 396, + 204 + ], + "type": "text", + "content": "{miaoz, chen2732, qqiu}@purdue.edu" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 152, + 231, + 200, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 231, + 200, + 243 + ], + "spans": [ + { + "bbox": [ + 152, + 231, + 200, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 256, + 297, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 256, + 297, + 593 + ], + "spans": [ + { + "bbox": [ + 53, + 256, + 297, + 593 + ], + "type": "text", + "content": "Transformer-based large pre-trained models have shown remarkable generalization ability, and various parameter-efficient fine-tuning (PEFT) methods have been proposed to customize these models on downstream tasks with minimal computational and memory budgets. Previous PEFT methods are primarily designed from a tensor-decomposition perspective that tries to effectively tune the linear transformation by finding the smallest subset of parameters to train. Our study adopts an orthogonal view by representing the attention operation as a graph convolution and formulating the multi-head attention maps as a convolutional filter subspace, with each attention map as a subspace element. In this paper, we propose to tune the large pre-trained transformers by learning a small set of combination coefficients that construct a more expressive filter subspace from the original multi-head attention maps. We show analytically and experimentally that the tuned filter subspace can effectively expand the feature space of the multi-head attention and further enhance the capacity of transformers. We further stabilize the fine-tuning with a residual parameterization of the tunable subspace coefficients, and enhance the generalization with a regularization design by directly applying dropout on the tunable coefficient during training. The tunable coefficients take a tiny number of parameters and can be combined with previous PEFT methods in a plug-and-play manner. Extensive experiments show that our approach achieves superior performances than PEFT baselines with neglectable additional parameters." + }, + { + "bbox": [ + 53, + 256, + 297, + 593 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 616, + 137, + 628 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 616, + 137, + 628 + ], + "spans": [ + { + "bbox": [ + 56, + 616, + 137, + 628 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 637, + 296, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 637, + 296, + 673 + ], + "spans": [ + { + "bbox": [ + 55, + 637, + 296, + 673 + ], + "type": "text", + "content": "Large foundational models have demonstrated remarkable performance and generalization ability across a wide range of natural language processing and computer vision" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 318, + 230, + 439, + 329 + ], + "blocks": [ + { + "bbox": [ + 318, + 230, + 439, + 329 + ], + "lines": [ + { + "bbox": [ + 318, + 230, + 439, + 329 + ], + "spans": [ + { + "bbox": [ + 318, + 230, + 439, + 329 + ], + "type": "image", + "image_path": "0815525a3a84eb0cfff84eee395e64934a19e0f9a5d2a1c968962f9036d3cc75.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 443, + 235, + 548, + 327 + ], + "blocks": [ + { + "bbox": [ + 443, + 235, + 548, + 327 + ], + "lines": [ + { + "bbox": [ + 443, + 235, + 548, + 327 + ], + "spans": [ + { + "bbox": [ + 443, + 235, + 548, + 327 + ], + "type": "image", + "image_path": "49bb2896c59a987c7204b513e0614d167910713c6d31bab73f622a30fbfb3e67.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 329, + 345, + 421, + 437 + ], + "blocks": [ + { + "bbox": [ + 336, + 334, + 504, + 344 + ], + "lines": [ + { + "bbox": [ + 336, + 334, + 504, + 344 + ], + "spans": [ + { + "bbox": [ + 336, + 334, + 504, + 344 + ], + "type": "text", + "content": "A [V] plushie panda on top of a purple rug in a forest." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 329, + 345, + 421, + 437 + ], + "lines": [ + { + "bbox": [ + 329, + 345, + 421, + 437 + ], + "spans": [ + { + "bbox": [ + 329, + 345, + 421, + 437 + ], + "type": "image", + "image_path": "08444f97283a41285cc1e6a3fac51216f3a3927e61125f45277baec4bdd4d0c4.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 433, + 345, + 524, + 436 + ], + "blocks": [ + { + "bbox": [ + 433, + 345, + 524, + 436 + ], + "lines": [ + { + "bbox": [ + 433, + 345, + 524, + 436 + ], + "spans": [ + { + "bbox": [ + 433, + 345, + 524, + 436 + ], + "type": "image", + "image_path": "74646b37219c208b016d61193f640aa8da780874b6c2829704fbeaa9300ff873.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 331, + 449, + 422, + 542 + ], + "blocks": [ + { + "bbox": [ + 375, + 438, + 487, + 448 + ], + "lines": [ + { + "bbox": [ + 375, + 438, + 487, + 448 + ], + "spans": [ + { + "bbox": [ + 375, + 438, + 487, + 448 + ], + "type": "text", + "content": "A [V] plushie panda on in the snow." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 331, + 449, + 422, + 542 + ], + "lines": [ + { + "bbox": [ + 331, + 449, + 422, + 542 + ], + "spans": [ + { + "bbox": [ + 331, + 449, + 422, + 542 + ], + "type": "image", + "image_path": "498db363057d9dbe54d77e5f5fcfbf7156bc5ec867e58301e3bb0e9f6348f0fa.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 364, + 545, + 388, + 554 + ], + "lines": [ + { + "bbox": [ + 364, + 545, + 388, + 554 + ], + "spans": [ + { + "bbox": [ + 364, + 545, + 388, + 554 + ], + "type": "text", + "content": "LoRA" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 570, + 555, + 647 + ], + "lines": [ + { + "bbox": [ + 313, + 570, + 555, + 647 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 555, + 647 + ], + "type": "text", + "content": "Figure 1. (Up) Under the graph convolution view of the attention, the proposed Coeff-Tuning enhances the expressiveness by combining the attention maps with a tiny tunable subspace coefficient " + }, + { + "bbox": [ + 313, + 570, + 555, + 647 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 570, + 555, + 647 + ], + "type": "text", + "content": ". It can be seamlessly integrated with other PEFT methods, e.g., LoRA [21], by introducing negligible additional parameters. (Bottom) Combining Coeff-Tuning with LoRA achieves better results in the concept customization task with SDXL [51]." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 433, + 449, + 524, + 541 + ], + "blocks": [ + { + "bbox": [ + 433, + 449, + 524, + 541 + ], + "lines": [ + { + "bbox": [ + 433, + 449, + 524, + 541 + ], + "spans": [ + { + "bbox": [ + 433, + 449, + 524, + 541 + ], + "type": "image", + "image_path": "236494ecb81262f72717a042ff784bd414d1ceb29f4bd63f97f0ec10e063104f.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 436, + 545, + 517, + 555 + ], + "lines": [ + { + "bbox": [ + 436, + 545, + 517, + 555 + ], + "spans": [ + { + "bbox": [ + 436, + 545, + 517, + 555 + ], + "type": "text", + "content": "LoRA + Coeff-Tuning" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "content": "tasks [11, 28, 55, 68]. The success of these models can be attributed to their ability to learn rich, transferable representations by leveraging the power of the transformer architecture [68] and the extensive knowledge captured during" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 683, + 134, + 693 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 683, + 134, + 693 + ], + "spans": [ + { + "bbox": [ + 66, + 683, + 134, + 693 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 66, + 683, + 134, + 693 + ], + "type": "text", + "content": " Equal contribution." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 56, + 693, + 294, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 693, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 693, + 294, + 713 + ], + "type": "text", + "content": "1Code is available at: https://github.com/ZichenMiao/Coeff_Tuning" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "20146" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "content": "pre-training on large-scale datasets [1, 58]. However, fine-tuning these pre-trained transformers on downstream tasks with full-parameter updates can be challenging due to the high computational and memory costs." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 122, + 295, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 122, + 295, + 264 + ], + "spans": [ + { + "bbox": [ + 55, + 122, + 295, + 264 + ], + "type": "text", + "content": "To address this, various parameter-efficient fine-tuning (PEFT) methods have been developed to adapt pre-trained models on downstream tasks while minimizing the number of trainable parameters [21, 38, 53, 79]. A representative family of these approaches focuses on tuning the linear transformation layers in transformers, where they apply a lightweight parameterization, e.g., low-rank parameterization [21, 38], to adapt the weight matrices. However, none of these methods are designed from a holistic view of the core operation in transformers, the multi-head attention layer [68], which plays a key role in the expressiveness of transformers [10, 59]." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 266, + 295, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 266, + 295, + 386 + ], + "spans": [ + { + "bbox": [ + 55, + 266, + 295, + 386 + ], + "type": "text", + "content": "In this paper, we propose a new PEFT method tailored to enhance the multi-head attention's expressiveness by tuning a minimal set of parameters. We first formulate the multi-head attention as a graph convolution with a filter subspace, with each graph filter as the attention map of each attention head. We then tune the filter subspace by simply learning a small set of subspace coefficients that linearly combine the pre-trained attention maps. In this way, we adapt for each attention head a new task-adaptive attention map that better captures the token-wise correlation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 388, + 295, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 388, + 295, + 579 + ], + "spans": [ + { + "bbox": [ + 55, + 388, + 295, + 579 + ], + "type": "text", + "content": "We attribute the effectiveness of our approach to its ability to expand the feature space of the original multi-head attention layer. By applying the softmax function to the query-key correlation matrices, all elements in a row of the original attention map are non-negative and have a sum of one [68]. This limits the expressiveness of the multi-head attention as the output token embeddings lie within the convex hull of the value embeddings. Our method constructs a new filter subspace with non-constrained coefficients as shown in Figure 1, allowing negative elements learned in the combined attention map to capture more diverse token-wise relationships. This also enables the output tokens to obtain representation outside the convex hull of the value embeddings, which further enhances the capacity of the multi-head attention. We provide analysis and an illustrative toy example to better demonstrate this point." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 580, + 295, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 580, + 295, + 652 + ], + "spans": [ + { + "bbox": [ + 55, + 580, + 295, + 652 + ], + "type": "text", + "content": "We show that our approach can be seamlessly combined with previous PEFT methods in a plug-and-play manner. As the filter subspace in each attention layer is spanned by a tiny number of attention maps (usually " + }, + { + "bbox": [ + 55, + 580, + 295, + 652 + ], + "type": "inline_equation", + "content": "10\\sim 30" + }, + { + "bbox": [ + 55, + 580, + 295, + 652 + ], + "type": "text", + "content": "), we only introduce neglectable additional tunable parameters and training costs over the base PEFT methods." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 653, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 295, + 713 + ], + "type": "text", + "content": "We validate our approach on a large variety of fine-tuning tasks and datasets, along with pre-trained transformers of different scales and modalities. Specifically, we first show the effectiveness of our method with ViT [11] on the few-shot image classification fine-tuning. We also" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 553, + 132 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 132 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 132 + ], + "type": "text", + "content": "show our method can achieve superior performance with vision-language transformers on image-text multi-modal fine-tuning [8]. Moreover, we also show our method can better customize the text-to-image diffusion models on personalized generation [56]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 326, + 133, + 504, + 143 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 133, + 504, + 143 + ], + "spans": [ + { + "bbox": [ + 326, + 133, + 504, + 143 + ], + "type": "text", + "content": "We summarize our contributions as follows," + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 144, + 553, + 239 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 314, + 144, + 553, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 144, + 553, + 180 + ], + "spans": [ + { + "bbox": [ + 314, + 144, + 553, + 180 + ], + "type": "text", + "content": "- We provide a filter subspace view of the multi-head attention and propose to fine-tune a new filter subspace by learning a small set of subspace coefficients." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 181, + 553, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 181, + 553, + 216 + ], + "spans": [ + { + "bbox": [ + 313, + 181, + 553, + 216 + ], + "type": "text", + "content": "- We demonstrate that our method can expand the feature space of the original multi-head attention and enhance its expressiveness on downstream tasks." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 217, + 553, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 217, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 313, + 217, + 553, + 239 + ], + "type": "text", + "content": "- We validate our method on extensive fine-tuning tasks and various transformer architectures." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 251, + 400, + 264 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 251, + 400, + 264 + ], + "spans": [ + { + "bbox": [ + 314, + 251, + 400, + 264 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 272, + 553, + 619 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 272, + 553, + 619 + ], + "spans": [ + { + "bbox": [ + 313, + 272, + 553, + 619 + ], + "type": "text", + "content": "Parameter-Efficient Fine-Tuning. Large attention-based transformers [11, 54, 55, 68] gain a significant amount of generalizable knowledge by pre-training on large-scale datasets [57, 58]. A common way to utilize those general skills is to fine-tune large transformers on domain-specific tasks and data. Recent advancements in parameter-efficient fine-tuning (PEFT) methods aim to reduce the computational and memory costs of adapting large models by fine-tuning only a subset of parameters. LoRA [21] and its variants [4, 24, 29, 36, 78, 82] fine-tune lower-rank matrices at each layer to represent weight updates. DoRA [38] enhances LoRA through weight decomposition into magnitude and direction, thus improving training stability and learning capacity without additional inference costs. The adapter [20] approach inserts small modules between layers and reduces parameters by only tuning these adapters [2, 24, 34, 50, 79]. Visual prompt tuning (VPT) [23, 60] has introduced a limited number of learnable parameters for optimization while keeping the backbone frozen. SSF [35] proposes scaling and shifting features of each layer in a pre-trained large model. OFT [53] tackles fine-tuning in text-to-image diffusion models by preserving the hyperspherical energy of neuron relationships. However, these methods primarily focus on efficient tuning of the parameter weight matrices, lacking a holistic view of the attention layer. In this paper, we take an orthogonal approach by tuning a tiny combination coefficient given a novel graph filter subspace view of the multi-head attention." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 629, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 553, + 713 + ], + "type": "text", + "content": "The Expressiveness of Attention. Transformer-based models have achieved success across vision and language tasks [8, 11, 14, 22, 31, 37] due to the effectiveness of the attention mechanism [68]. Despite their success, various studies have examined the limitations in the expressiveness of transformers [10, 59, 71, 80]. Shi et al. [59] and Zhou et al. [84] identify the attention collapse in transformers as" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20147" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 69, + 71, + 539, + 264 + ], + "blocks": [ + { + "bbox": [ + 69, + 71, + 539, + 264 + ], + "lines": [ + { + "bbox": [ + 69, + 71, + 539, + 264 + ], + "spans": [ + { + "bbox": [ + 69, + 71, + 539, + 264 + ], + "type": "image", + "image_path": "80e2fe75c334f99f28bd010b2c8adf95a63b5aee8e96e79f60c2303965e95bf8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 273, + 555, + 330 + ], + "lines": [ + { + "bbox": [ + 54, + 273, + 555, + 330 + ], + "spans": [ + { + "bbox": [ + 54, + 273, + 555, + 330 + ], + "type": "text", + "content": "Figure 2. Illustration of the proposed Coeff-Tuning. We first view each token " + }, + { + "bbox": [ + 54, + 273, + 555, + 330 + ], + "type": "inline_equation", + "content": "\\mathbf{X}[i]" + }, + { + "bbox": [ + 54, + 273, + 555, + 330 + ], + "type": "text", + "content": " as a node " + }, + { + "bbox": [ + 54, + 273, + 555, + 330 + ], + "type": "inline_equation", + "content": "V_{i}" + }, + { + "bbox": [ + 54, + 273, + 555, + 330 + ], + "type": "text", + "content": " in an all-connected graph " + }, + { + "bbox": [ + 54, + 273, + 555, + 330 + ], + "type": "inline_equation", + "content": "\\mathcal{G} = (V,E)" + }, + { + "bbox": [ + 54, + 273, + 555, + 330 + ], + "type": "text", + "content": ". The attention operation then turns into a graph convolution with each attention map " + }, + { + "bbox": [ + 54, + 273, + 555, + 330 + ], + "type": "inline_equation", + "content": "\\mathbf{A}^h" + }, + { + "bbox": [ + 54, + 273, + 555, + 330 + ], + "type": "text", + "content": " as a graph convolutional filter " + }, + { + "bbox": [ + 54, + 273, + 555, + 330 + ], + "type": "inline_equation", + "content": "\\mathbf{F}^h (\\mathbf{X})" + }, + { + "bbox": [ + 54, + 273, + 555, + 330 + ], + "type": "text", + "content": ". In this way, the attention maps in the multi-head attention layer form a filter subspace " + }, + { + "bbox": [ + 54, + 273, + 555, + 330 + ], + "type": "inline_equation", + "content": "\\mathcal{F}(\\mathbf{X})" + }, + { + "bbox": [ + 54, + 273, + 555, + 330 + ], + "type": "text", + "content": ", and each element in the filter is constrained in " + }, + { + "bbox": [ + 54, + 273, + 555, + 330 + ], + "type": "inline_equation", + "content": "(0,1)" + }, + { + "bbox": [ + 54, + 273, + 555, + 330 + ], + "type": "text", + "content": " after softmax. The proposed Coeff-Tuning then tunes a subspace coefficient " + }, + { + "bbox": [ + 54, + 273, + 555, + 330 + ], + "type": "inline_equation", + "content": "\\alpha \\in \\mathbb{R}^{H\\times H}" + }, + { + "bbox": [ + 54, + 273, + 555, + 330 + ], + "type": "text", + "content": " with a tiny number of parameters, that expands the range of attention scores by linearly combining the original filter subspace and enhances the expressiveness of the multi-head attention." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 340, + 297, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 340, + 297, + 496 + ], + "spans": [ + { + "bbox": [ + 54, + 340, + 297, + 496 + ], + "type": "text", + "content": "a result of self-attention's convergence to a low-rank space, which leads to unstable training as attention scores concentrate [10, 45]. To counteract this, techniques such as hierarchical fusion [59] and stabilizing attention entropy [80] have been proposed to prevent performance stagnation in vision transformers. DeepViT [84] proposes to re-generate diverse attention maps to avoid attention collapse in training vision transformers from scratch. In contrast, our work studies the expressiveness of the attention layer in the context of the large model fine-tuning. We also provide a novel interpretation of multi-head attention maps as a graph filter subspace, and propose the residual parameterization and regularization to stabilize the tuning process." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 511, + 112, + 523 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 511, + 112, + 523 + ], + "spans": [ + { + "bbox": [ + 55, + 511, + 112, + 523 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 532, + 250, + 545 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 532, + 250, + 545 + ], + "spans": [ + { + "bbox": [ + 55, + 532, + 250, + 545 + ], + "type": "text", + "content": "3.1. A Filter Subspace View on Attention" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 552, + 296, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 552, + 296, + 636 + ], + "spans": [ + { + "bbox": [ + 55, + 552, + 296, + 636 + ], + "type": "text", + "content": "Attention. The attention layer transforms the input sequence " + }, + { + "bbox": [ + 55, + 552, + 296, + 636 + ], + "type": "inline_equation", + "content": "\\mathbf{X} \\in \\mathbb{R}^{N \\times C_i}" + }, + { + "bbox": [ + 55, + 552, + 296, + 636 + ], + "type": "text", + "content": " to the output sequence " + }, + { + "bbox": [ + 55, + 552, + 296, + 636 + ], + "type": "inline_equation", + "content": "\\mathbf{O} \\in \\mathbb{R}^{N \\times C_o}" + }, + { + "bbox": [ + 55, + 552, + 296, + 636 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 552, + 296, + 636 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 552, + 296, + 636 + ], + "type": "text", + "content": " denotes the sequence length, " + }, + { + "bbox": [ + 55, + 552, + 296, + 636 + ], + "type": "inline_equation", + "content": "C_i" + }, + { + "bbox": [ + 55, + 552, + 296, + 636 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 552, + 296, + 636 + ], + "type": "inline_equation", + "content": "C_o" + }, + { + "bbox": [ + 55, + 552, + 296, + 636 + ], + "type": "text", + "content": " are the dimension of input and output features. The attention layer projects the input as " + }, + { + "bbox": [ + 55, + 552, + 296, + 636 + ], + "type": "inline_equation", + "content": "\\mathbf{Q} = \\mathbf{X}\\mathbf{W}_q" + }, + { + "bbox": [ + 55, + 552, + 296, + 636 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 552, + 296, + 636 + ], + "type": "inline_equation", + "content": "\\mathbf{K} = \\mathbf{X}\\mathbf{W}_k" + }, + { + "bbox": [ + 55, + 552, + 296, + 636 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 552, + 296, + 636 + ], + "type": "inline_equation", + "content": "\\mathbf{V} = \\mathbf{X}\\mathbf{W}_v" + }, + { + "bbox": [ + 55, + 552, + 296, + 636 + ], + "type": "text", + "content": " using the corresponding projection matrices " + }, + { + "bbox": [ + 55, + 552, + 296, + 636 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_q, \\mathbf{W}_k, \\mathbf{W}_v \\in \\mathbb{R}^{C_i \\times C_o}" + }, + { + "bbox": [ + 55, + 552, + 296, + 636 + ], + "type": "text", + "content": ", and calculates the attention map," + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 113, + 647, + 295, + 677 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 647, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 113, + 647, + 295, + 677 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathbf {A} = \\operatorname {s o f t m a x} (\\mathbf {Q K} ^ {\\intercal}) \\tag {1} \\\\ = \\operatorname {s o f t m a x} \\left(\\mathbf {X} \\mathbf {W} _ {q} \\mathbf {W} _ {k} ^ {\\intercal} \\mathbf {X} ^ {\\intercal}\\right). \\\\ \\end{array}", + "image_path": "9cbf7d85e882bec5b9236e5c08cab9a65266379b0776829f9d8a6f0c90dfcc3b.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "The attention map " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{A} \\in \\mathbb{R}^{N \\times N}" + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": " captures the token-wise relationship by doing inner-product in a space transformed by" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 339, + 522, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 339, + 522, + 353 + ], + "spans": [ + { + "bbox": [ + 314, + 339, + 522, + 353 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_q, \\mathbf{W}_k" + }, + { + "bbox": [ + 314, + 339, + 522, + 353 + ], + "type": "text", + "content": ". Then the output of the attention layer is," + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 361, + 360, + 553, + 402 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 360, + 553, + 402 + ], + "spans": [ + { + "bbox": [ + 361, + 360, + 553, + 402 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathbf {O} = \\operatorname {a t t n} \\left(\\mathbf {W} _ {q}, \\mathbf {W} _ {k}, \\mathbf {W} _ {v}; \\mathbf {X}\\right) \\\\ = \\operatorname {s o f t m a x} \\left(\\mathbf {X} \\mathbf {W} _ {q} \\mathbf {W} _ {k} ^ {\\intercal} \\mathbf {X} ^ {\\intercal}\\right) \\mathbf {X} \\mathbf {W} _ {v} \\tag {2} \\\\ = \\mathbf {A X W} _ {v}. \\\\ \\end{array}", + "image_path": "1afcccc446e792a6347f1206c1f040920042f0f966e6c4ec7ea9217e6f8c9984.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 415, + 555, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 415, + 555, + 464 + ], + "spans": [ + { + "bbox": [ + 313, + 415, + 555, + 464 + ], + "type": "text", + "content": "Multi-head Attention. Multi-head attention extends this by allowing multiple attention mechanisms, i.e., heads, to work in parallel, with each head independently learning attention patterns. For each attention head, we have," + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 369, + 470, + 555, + 486 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 369, + 470, + 555, + 486 + ], + "spans": [ + { + "bbox": [ + 369, + 470, + 555, + 486 + ], + "type": "interline_equation", + "content": "\\mathbf {O} ^ {h} = \\operatorname {a t t n} \\left(\\mathbf {W} _ {q} ^ {h}, \\mathbf {W} _ {k} ^ {h}, \\mathbf {W} _ {v} ^ {h}; \\mathbf {X}\\right), \\tag {3}", + "image_path": "79b6cd846588d416805e697ffd5ad0658339bafeefe7df5627000aa2dda36907.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 491, + 555, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 491, + 555, + 532 + ], + "spans": [ + { + "bbox": [ + 313, + 491, + 555, + 532 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 491, + 555, + 532 + ], + "type": "inline_equation", + "content": "h = 1,\\dots ,H" + }, + { + "bbox": [ + 313, + 491, + 555, + 532 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 491, + 555, + 532 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 313, + 491, + 555, + 532 + ], + "type": "text", + "content": " is the number of heads. The project matrices for each head are " + }, + { + "bbox": [ + 313, + 491, + 555, + 532 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_q^h,\\mathbf{W}_k^h,\\mathbf{W}_v^h\\in" + }, + { + "bbox": [ + 313, + 491, + 555, + 532 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{C_i\\times \\frac{C_o}{H}}" + }, + { + "bbox": [ + 313, + 491, + 555, + 532 + ], + "type": "text", + "content": " . The multi-head attention represents as," + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 340, + 536, + 553, + 604 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 340, + 536, + 553, + 604 + ], + "spans": [ + { + "bbox": [ + 340, + 536, + 553, + 604 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathbf {Z} = \\operatorname {C o n c a t} \\left[ \\mathbf {O} ^ {1}, \\dots , \\mathbf {O} ^ {H} \\right] \\mathbf {W} _ {o} \\\\ = \\underset {H} {\\operatorname {C o n c a t}} \\left[ \\mathbf {A} ^ {1} \\mathbf {X} \\mathbf {W} _ {v} ^ {1}, \\dots , \\mathbf {A} ^ {H} \\mathbf {X} \\mathbf {W} _ {v} ^ {H} \\right] \\mathbf {W} _ {o} \\tag {4} \\\\ = \\sum_ {h = 1} ^ {H} \\mathbf {A} ^ {h} \\mathbf {X} \\mathbf {W} _ {v o} ^ {h}, \\\\ \\end{array}", + "image_path": "de3104be6db297d30e4115254bc24593b40a0e22a5e2c320fec8252d4d925192.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 610, + 555, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 610, + 555, + 638 + ], + "spans": [ + { + "bbox": [ + 313, + 610, + 555, + 638 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 610, + 555, + 638 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_o\\in \\mathbb{R}^{C_o\\times C_o}" + }, + { + "bbox": [ + 313, + 610, + 555, + 638 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 610, + 555, + 638 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_{vo}^h = \\mathbf{W}_v^h (\\mathbf{W}_o^h)^\\intercal" + }, + { + "bbox": [ + 313, + 610, + 555, + 638 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_o^h\\in" + }, + { + "bbox": [ + 313, + 610, + 555, + 638 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{C_o\\times \\frac{C_o}{H}}" + }, + { + "bbox": [ + 313, + 610, + 555, + 638 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_{vo}^h\\in \\mathbb{R}^{C_i\\times C_o}" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 641, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 556, + 714 + ], + "type": "text", + "content": "A Filter Subspace View on Multi-head Attention. We now present our filter subspace view on the multi-head attention. By viewing each token embedding " + }, + { + "bbox": [ + 313, + 641, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{X}[n] \\in \\mathbb{R}^{C_i}" + }, + { + "bbox": [ + 313, + 641, + 556, + 714 + ], + "type": "text", + "content": " as a node " + }, + { + "bbox": [ + 313, + 641, + 556, + 714 + ], + "type": "inline_equation", + "content": "V_{n}" + }, + { + "bbox": [ + 313, + 641, + 556, + 714 + ], + "type": "text", + "content": ", which are connected with all the other nodes with edges " + }, + { + "bbox": [ + 313, + 641, + 556, + 714 + ], + "type": "inline_equation", + "content": "E[m, n] = 1" + }, + { + "bbox": [ + 313, + 641, + 556, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 641, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\forall m \\in [N]" + }, + { + "bbox": [ + 313, + 641, + 556, + 714 + ], + "type": "text", + "content": ", we obtain a graph representation of all the tokens, " + }, + { + "bbox": [ + 313, + 641, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{G} = (V, E)" + }, + { + "bbox": [ + 313, + 641, + 556, + 714 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20148" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 122 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 122 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 122 + ], + "type": "text", + "content": "We then reformulate the attention operation as the message passing as in Graph Convolution Nerworks [27]. Specifically, inspired by [6], we write the single-head attention as a graph convolution with the filter " + }, + { + "bbox": [ + 55, + 72, + 296, + 122 + ], + "type": "inline_equation", + "content": "\\mathbf{F}^h (\\mathbf{X})" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 66, + 128, + 295, + 165 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 128, + 295, + 165 + ], + "spans": [ + { + "bbox": [ + 66, + 128, + 295, + 165 + ], + "type": "interline_equation", + "content": "\\mathbf {O} ^ {h} [ n,: ] = \\sum_ {\\substack {n ^ {\\prime} \\in [ N ] \\\\ c ^ {\\prime} \\in [ C _ {i} ]}} \\mathbf {F} ^ {h} (\\mathbf {X}) [ n, n ^ {\\prime} ] \\mathbf {X} [ n ^ {\\prime}, c ^ {\\prime} ] \\mathbf {W} ^ {h} [ c ^ {\\prime},,: ], \\tag{5}", + "image_path": "447d971678a5132a7e32cfa514a4dbde054063800bc542b9127bc9e7c44f879b.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 172, + 296, + 222 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 172, + 296, + 222 + ], + "spans": [ + { + "bbox": [ + 55, + 172, + 296, + 222 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 172, + 296, + 222 + ], + "type": "inline_equation", + "content": "\\mathbf{W}^h = \\mathbf{W}_{vo}^h\\in \\mathbb{R}^{C_i\\times C_0}" + }, + { + "bbox": [ + 55, + 172, + 296, + 222 + ], + "type": "text", + "content": " denotes the weight for linear projection on each node, and the " + }, + { + "bbox": [ + 55, + 172, + 296, + 222 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 172, + 296, + 222 + ], + "type": "text", + "content": " -th row of the filter, i.e., " + }, + { + "bbox": [ + 55, + 172, + 296, + 222 + ], + "type": "inline_equation", + "content": "\\mathbf{F}^h (\\mathbf{X})[n,:]" + }, + { + "bbox": [ + 55, + 172, + 296, + 222 + ], + "type": "text", + "content": ", represents the local graph filter around the " + }, + { + "bbox": [ + 55, + 172, + 296, + 222 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 172, + 296, + 222 + ], + "type": "text", + "content": " -th node (token). Comparing with (1) and (2), we have," + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 103, + 229, + 295, + 245 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 103, + 229, + 295, + 245 + ], + "spans": [ + { + "bbox": [ + 103, + 229, + 295, + 245 + ], + "type": "interline_equation", + "content": "\\mathbf {F} ^ {h} (\\mathbf {X}) = \\operatorname {s o f t m a x} \\left(\\mathbf {X} \\mathbf {W} _ {q} \\mathbf {W} _ {k} ^ {\\intercal} \\mathbf {X} ^ {\\intercal}\\right). \\tag {6}", + "image_path": "84d09a6c642cedd0d22131e4a1ec0e81683a96c7dfc3e3b7bd9c7d0386793a34.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 251, + 296, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 251, + 296, + 300 + ], + "spans": [ + { + "bbox": [ + 55, + 251, + 296, + 300 + ], + "type": "text", + "content": "The multi-head attention individually learns " + }, + { + "bbox": [ + 55, + 251, + 296, + 300 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 55, + 251, + 296, + 300 + ], + "type": "text", + "content": " filters " + }, + { + "bbox": [ + 55, + 251, + 296, + 300 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{F}^1 (X),\\dots,\\mathbf{F}^H (X)\\}" + }, + { + "bbox": [ + 55, + 251, + 296, + 300 + ], + "type": "text", + "content": " to aggregate complementary information across nodes. These " + }, + { + "bbox": [ + 55, + 251, + 296, + 300 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 55, + 251, + 296, + 300 + ], + "type": "text", + "content": " filters naturally form a graph convolutional filter subspace " + }, + { + "bbox": [ + 55, + 251, + 296, + 300 + ], + "type": "inline_equation", + "content": "\\mathcal{F}(\\mathbf{X})" + }, + { + "bbox": [ + 55, + 251, + 296, + 300 + ], + "type": "text", + "content": " as," + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 97, + 306, + 295, + 323 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 306, + 295, + 323 + ], + "spans": [ + { + "bbox": [ + 97, + 306, + 295, + 323 + ], + "type": "interline_equation", + "content": "\\mathcal {F} (\\mathbf {X}) = \\left\\{\\mathbf {F} ^ {h} (\\mathbf {X}) \\right\\} _ {h = 1} ^ {H} \\subseteq (0, 1) ^ {N \\times N}, \\tag {7}", + "image_path": "df1044edd1d1e62bd2a3a26bed3e9a1926080cb153348dc243887fcdc7f74784.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 328, + 296, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 328, + 296, + 366 + ], + "spans": [ + { + "bbox": [ + 55, + 328, + 296, + 366 + ], + "type": "text", + "content": "and we denote each " + }, + { + "bbox": [ + 55, + 328, + 296, + 366 + ], + "type": "inline_equation", + "content": "\\mathbf{F}^h (\\mathbf{X})" + }, + { + "bbox": [ + 55, + 328, + 296, + 366 + ], + "type": "text", + "content": " as a filter atom. The multi-head attention can then be represented in the way of decomposed graph convolution [6] as," + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 373, + 295, + 418 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 373, + 295, + 418 + ], + "spans": [ + { + "bbox": [ + 62, + 373, + 295, + 418 + ], + "type": "interline_equation", + "content": "\\mathbf {O} [ n,: ] = \\sum_ {h = 1} ^ {H} \\sum_ {\\substack {n ^ {\\prime} \\in [ N ] \\\\ c ^ {\\prime} \\in [ C _ {i} ]}} \\mathbf {F} ^ {h} (\\mathbf {X}) [ n, n ^ {\\prime} ] \\mathbf {X} [ n ^ {\\prime}, c ^ {\\prime} ] \\mathbf {W} ^ {h} [ c ^ {\\prime},: ]. \\tag{8}", + "image_path": "5e31eb3b0ef011cbec7a0076c53e547352019b68f7d64c776d9ef96d4d70ab95.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 426, + 296, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 426, + 296, + 521 + ], + "spans": [ + { + "bbox": [ + 55, + 426, + 296, + 521 + ], + "type": "text", + "content": "It can be seen from (8) that the parameters " + }, + { + "bbox": [ + 55, + 426, + 296, + 521 + ], + "type": "inline_equation", + "content": "\\mathbf{F}^h (\\mathbf{X}),\\mathbf{W}^h" + }, + { + "bbox": [ + 55, + 426, + 296, + 521 + ], + "type": "text", + "content": " plays two key yet independent roles in the graph convolution. The " + }, + { + "bbox": [ + 55, + 426, + 296, + 521 + ], + "type": "inline_equation", + "content": "\\mathbf{W}^h" + }, + { + "bbox": [ + 55, + 426, + 296, + 521 + ], + "type": "text", + "content": " updates the knowledge for each node by projecting its representation into a new space, while the graph filter " + }, + { + "bbox": [ + 55, + 426, + 296, + 521 + ], + "type": "inline_equation", + "content": "\\mathbf{F}^h (\\mathbf{X})" + }, + { + "bbox": [ + 55, + 426, + 296, + 521 + ], + "type": "text", + "content": " guides the information flow across nodes. Intuitively, with a more expressive graph convolutional filter subspace " + }, + { + "bbox": [ + 55, + 426, + 296, + 521 + ], + "type": "inline_equation", + "content": "\\mathcal{F}(\\mathbf{X})" + }, + { + "bbox": [ + 55, + 426, + 296, + 521 + ], + "type": "text", + "content": ", each node can aggregate more informative features from other nodes." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 528, + 212, + 541 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 528, + 212, + 541 + ], + "spans": [ + { + "bbox": [ + 55, + 528, + 212, + 541 + ], + "type": "text", + "content": "3.2. Subspace Coefficient Tuning" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 545, + 295, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 295, + 653 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 295, + 653 + ], + "type": "text", + "content": "Given the re-formulation of multi-head attention maps as a graph convolutional filter subspace, we present an efficient fine-tuning method specifically designed for transformers to enhance attention's expressiveness. Our motivation is that after the softmax function, each graph filter " + }, + { + "bbox": [ + 55, + 545, + 295, + 653 + ], + "type": "inline_equation", + "content": "\\mathbf{F}^h (\\mathbf{X})" + }, + { + "bbox": [ + 55, + 545, + 295, + 653 + ], + "type": "text", + "content": " has values in the range (0, 1) with row-sum as one. The normalized attention maps restrict the expressiveness in the way that each output embedding lies in the convex hull of the value embeddings." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": "To expand the feature space of the output embeddings, we propose Coeff-Tuning, to tune the filter subspace on the target dataset with a small set of combination coefficients, i.e., subspace coefficient " + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\alpha \\in \\mathbb{R}^{H\\times H}" + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": ". As shown in Figure 2, the tunable coefficient " + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": " constructs a more expressive" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 72, + 555, + 110 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 110 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 110 + ], + "type": "text", + "content": "filter subspace " + }, + { + "bbox": [ + 313, + 72, + 555, + 110 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{F}}_{\\alpha}(\\mathbf{X})" + }, + { + "bbox": [ + 313, + 72, + 555, + 110 + ], + "type": "text", + "content": " by linearly combining the pre-trained multi-head attention maps " + }, + { + "bbox": [ + 313, + 72, + 555, + 110 + ], + "type": "inline_equation", + "content": "\\mathbf{F}(\\mathbf{X})" + }, + { + "bbox": [ + 313, + 72, + 555, + 110 + ], + "type": "text", + "content": ". Each element " + }, + { + "bbox": [ + 313, + 72, + 555, + 110 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{F}}^h (\\mathbf{X})" + }, + { + "bbox": [ + 313, + 72, + 555, + 110 + ], + "type": "text", + "content": " in the learned filter subspace can be written as," + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 340, + 118, + 555, + 152 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 340, + 118, + 555, + 152 + ], + "spans": [ + { + "bbox": [ + 340, + 118, + 555, + 152 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {F}} ^ {h} (\\mathbf {X}) = \\boldsymbol {\\alpha} [ h, : ] ^ {\\intercal} \\mathcal {F} (\\mathbf {X}) = \\sum_ {i = 1} ^ {H} \\boldsymbol {\\alpha} [ h, i ] \\mathbf {F} ^ {i} (\\mathbf {X}). \\tag {9}", + "image_path": "0dce20d72e26d5a051fc364fcbde156071616cd1352471dc6ada4a9e1099ffa1.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 314, + 160, + 492, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 160, + 492, + 172 + ], + "spans": [ + { + "bbox": [ + 314, + 160, + 492, + 172 + ], + "type": "text", + "content": "The multi-head attention then represents as," + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 321, + 181, + 554, + 228 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 181, + 554, + 228 + ], + "spans": [ + { + "bbox": [ + 321, + 181, + 554, + 228 + ], + "type": "interline_equation", + "content": "\\mathbf {O} = \\sum_ {h = 1} ^ {H} \\hat {\\mathbf {F}} ^ {h} (\\mathbf {X}) \\mathbf {X} \\mathbf {W} ^ {h} = \\sum_ {h = 1} ^ {H} \\sum_ {i = 1} ^ {H} \\alpha [ h, i ] \\mathbf {F} ^ {i} (\\mathbf {X}) \\mathbf {X} \\mathbf {W} ^ {h}. \\tag {10}", + "image_path": "7fa4e9d7ff9550846958259e01976753e0cf3931eb9bb36d893b9aab9e3ded4f.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 236, + 555, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 236, + 555, + 308 + ], + "spans": [ + { + "bbox": [ + 313, + 236, + 555, + 308 + ], + "type": "text", + "content": "Note that, mathematically, the subspace coefficient " + }, + { + "bbox": [ + 313, + 236, + 555, + 308 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 236, + 555, + 308 + ], + "type": "text", + "content": " is unconstrained, allowing it to combine attention maps with a wider range of values from the original filter subspace " + }, + { + "bbox": [ + 313, + 236, + 555, + 308 + ], + "type": "inline_equation", + "content": "\\mathbf{F}(\\mathbf{X})" + }, + { + "bbox": [ + 313, + 236, + 555, + 308 + ], + "type": "text", + "content": ". With attention scores exceeding the range " + }, + { + "bbox": [ + 313, + 236, + 555, + 308 + ], + "type": "inline_equation", + "content": "(0,1)" + }, + { + "bbox": [ + 313, + 236, + 555, + 308 + ], + "type": "text", + "content": ", we effectively expand the space for " + }, + { + "bbox": [ + 313, + 236, + 555, + 308 + ], + "type": "inline_equation", + "content": "\\mathbf{O}" + }, + { + "bbox": [ + 313, + 236, + 555, + 308 + ], + "type": "text", + "content": " by allowing a more flexible combination of value matrices." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 319, + 555, + 367 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 319, + 555, + 367 + ], + "spans": [ + { + "bbox": [ + 313, + 319, + 555, + 367 + ], + "type": "text", + "content": "Residual Parameterization. To facilitate the tuning and preserve the most knowledge in the pre-trained transformer, we propose a residual parameterization of the subspace coefficient," + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 374, + 376, + 427, + 388 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 374, + 376, + 427, + 388 + ], + "spans": [ + { + "bbox": [ + 374, + 376, + 427, + 388 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\alpha} ^ {\\prime} = \\boldsymbol {\\alpha} + \\mathbf {I},", + "image_path": "0b67a8e96cbec4149c8997599d3c2af1875e21bde382a864e91528310825510a.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 355, + 392, + 554, + 425 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 355, + 392, + 554, + 425 + ], + "spans": [ + { + "bbox": [ + 355, + 392, + 554, + 425 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {F}} ^ {h} (\\mathbf {X}) = \\mathbf {F} ^ {h} (\\mathbf {X}) + \\sum_ {i = 1} ^ {H} \\boldsymbol {\\alpha} [ h, i ] \\mathbf {F} ^ {i} (\\mathbf {X}). \\tag {11}", + "image_path": "0d66c330d5b9e3eafffee8f1b93d2a9d67c7e664a929f0d58b9e4622edc9fb3c.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 434, + 555, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 434, + 555, + 482 + ], + "spans": [ + { + "bbox": [ + 313, + 434, + 555, + 482 + ], + "type": "text", + "content": "In this way, we maintain the original filter subspace with the identity matrix, while parameterizing the subspace coefficient " + }, + { + "bbox": [ + 313, + 434, + 555, + 482 + ], + "type": "inline_equation", + "content": "\\alpha'" + }, + { + "bbox": [ + 313, + 434, + 555, + 482 + ], + "type": "text", + "content": " to learn the residual combination of attention maps on fine-tuning tasks." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 492, + 555, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 492, + 555, + 589 + ], + "spans": [ + { + "bbox": [ + 313, + 492, + 555, + 589 + ], + "type": "text", + "content": "Reguralization on Subspace Coefficient. The proposed Coeff-Tuning substantially enhances the expressiveness of multi-head attention with tunable coefficient " + }, + { + "bbox": [ + 313, + 492, + 555, + 589 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 492, + 555, + 589 + ], + "type": "text", + "content": ". However, it may also lead to overfitting and preventing channel projection parameters " + }, + { + "bbox": [ + 313, + 492, + 555, + 589 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_v, \\mathbf{W}_o" + }, + { + "bbox": [ + 313, + 492, + 555, + 589 + ], + "type": "text", + "content": " from learning generalizable features. To alleviate this, we propose a simple regularization design that performs dropout [63] directly on the tuned subspace coefficient " + }, + { + "bbox": [ + 313, + 492, + 555, + 589 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 492, + 555, + 589 + ], + "type": "text", + "content": " during training," + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 381, + 598, + 555, + 612 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 381, + 598, + 555, + 612 + ], + "spans": [ + { + "bbox": [ + 381, + 598, + 555, + 612 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\alpha} ^ {\\prime} = \\operatorname {D r o p o u t} (\\boldsymbol {\\alpha}; p) + \\mathbf {I}. \\tag {12}", + "image_path": "0738e6154d28fe0bf960f7bde81cb9b57e229f2e66f95cc8fcc8a7ac6bfb6068.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 621, + 554, + 646 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 621, + 554, + 646 + ], + "spans": [ + { + "bbox": [ + 313, + 621, + 554, + 646 + ], + "type": "text", + "content": "In this way, we randomly remove an element from " + }, + { + "bbox": [ + 313, + 621, + 554, + 646 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 621, + 554, + 646 + ], + "type": "text", + "content": " with a probability of " + }, + { + "bbox": [ + 313, + 621, + 554, + 646 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 313, + 621, + 554, + 646 + ], + "type": "text", + "content": " to avoid overfitting." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 314, + 657, + 373, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 657, + 373, + 670 + ], + "spans": [ + { + "bbox": [ + 314, + 657, + 373, + 670 + ], + "type": "text", + "content": "4. Analysis" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "content": "In this section, we analyze that adjusting the filter subspace using the proposed subspace coefficient " + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "content": " can effectively enhance the expressiveness of the multi-head attention." + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20149" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 65, + 70, + 532, + 177 + ], + "blocks": [ + { + "bbox": [ + 65, + 70, + 532, + 177 + ], + "lines": [ + { + "bbox": [ + 65, + 70, + 532, + 177 + ], + "spans": [ + { + "bbox": [ + 65, + 70, + 532, + 177 + ], + "type": "image", + "image_path": "4d56caa0f51e396ee56aad728933bff47314989d7280470ba025ee4dfe966f6e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 186, + 555, + 266 + ], + "lines": [ + { + "bbox": [ + 54, + 186, + 555, + 266 + ], + "spans": [ + { + "bbox": [ + 54, + 186, + 555, + 266 + ], + "type": "text", + "content": "Figure 3. The effectiveness of tuning subspace coefficient " + }, + { + "bbox": [ + 54, + 186, + 555, + 266 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 54, + 186, + 555, + 266 + ], + "type": "text", + "content": ". The goal is to convert the input graph " + }, + { + "bbox": [ + 54, + 186, + 555, + 266 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 54, + 186, + 555, + 266 + ], + "type": "text", + "content": " to the target graph " + }, + { + "bbox": [ + 54, + 186, + 555, + 266 + ], + "type": "inline_equation", + "content": "\\mathbf{O}" + }, + { + "bbox": [ + 54, + 186, + 555, + 266 + ], + "type": "text", + "content": " using a single attention layer, " + }, + { + "bbox": [ + 54, + 186, + 555, + 266 + ], + "type": "inline_equation", + "content": "\\mathbf{O} = \\sum_{h=1}^{H} \\mathbf{F}^h(\\mathbf{X}) \\mathbf{X} \\mathbf{W}_v^h" + }, + { + "bbox": [ + 54, + 186, + 555, + 266 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 54, + 186, + 555, + 266 + ], + "type": "inline_equation", + "content": "\\mathbf{F}(\\mathbf{X}) = \\text{softmax}(\\mathbf{X} \\mathbf{W}_q \\mathbf{W}_k^\\top \\mathbf{X}^\\top)" + }, + { + "bbox": [ + 54, + 186, + 555, + 266 + ], + "type": "text", + "content": ". We have the following observations: (i) When only tuning " + }, + { + "bbox": [ + 54, + 186, + 555, + 266 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_q" + }, + { + "bbox": [ + 54, + 186, + 555, + 266 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 186, + 555, + 266 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_k" + }, + { + "bbox": [ + 54, + 186, + 555, + 266 + ], + "type": "text", + "content": ", i.e., updating " + }, + { + "bbox": [ + 54, + 186, + 555, + 266 + ], + "type": "inline_equation", + "content": "\\mathbf{F}(\\mathbf{X})" + }, + { + "bbox": [ + 54, + 186, + 555, + 266 + ], + "type": "text", + "content": ", the positions of output nodes are constrained within the range of the input nodes " + }, + { + "bbox": [ + 54, + 186, + 555, + 266 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 54, + 186, + 555, + 266 + ], + "type": "text", + "content": ", since the values of " + }, + { + "bbox": [ + 54, + 186, + 555, + 266 + ], + "type": "inline_equation", + "content": "\\mathbf{F}(\\mathbf{X})" + }, + { + "bbox": [ + 54, + 186, + 555, + 266 + ], + "type": "text", + "content": " are in the range of " + }, + { + "bbox": [ + 54, + 186, + 555, + 266 + ], + "type": "inline_equation", + "content": "(0,1)" + }, + { + "bbox": [ + 54, + 186, + 555, + 266 + ], + "type": "text", + "content": ". (ii) In addition to tuning " + }, + { + "bbox": [ + 54, + 186, + 555, + 266 + ], + "type": "inline_equation", + "content": "\\mathbf{F}(\\mathbf{X})" + }, + { + "bbox": [ + 54, + 186, + 555, + 266 + ], + "type": "text", + "content": ", updating " + }, + { + "bbox": [ + 54, + 186, + 555, + 266 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_v" + }, + { + "bbox": [ + 54, + 186, + 555, + 266 + ], + "type": "text", + "content": " projects the input as " + }, + { + "bbox": [ + 54, + 186, + 555, + 266 + ], + "type": "inline_equation", + "content": "\\mathbf{X} \\mathbf{W}_v" + }, + { + "bbox": [ + 54, + 186, + 555, + 266 + ], + "type": "text", + "content": ", but the output graph is still not aligned with the target graph. (iii) In addition to tuning " + }, + { + "bbox": [ + 54, + 186, + 555, + 266 + ], + "type": "inline_equation", + "content": "\\mathbf{F}(\\mathbf{X})" + }, + { + "bbox": [ + 54, + 186, + 555, + 266 + ], + "type": "text", + "content": ", adapting " + }, + { + "bbox": [ + 54, + 186, + 555, + 266 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 54, + 186, + 555, + 266 + ], + "type": "text", + "content": " successfully transforms the input graph into the target graph. Compared with (i), " + }, + { + "bbox": [ + 54, + 186, + 555, + 266 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 54, + 186, + 555, + 266 + ], + "type": "text", + "content": " expand the positions of output nodes " + }, + { + "bbox": [ + 54, + 186, + 555, + 266 + ], + "type": "inline_equation", + "content": "\\sum_{i=1}^{H} \\alpha[;, i] \\mathbf{F}^i(\\mathbf{X}) \\mathbf{X}" + }, + { + "bbox": [ + 54, + 186, + 555, + 266 + ], + "type": "text", + "content": " beyond the constraints of the input nodes, enhancing the expressiveness of the attention effectively." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 276, + 295, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 276, + 295, + 312 + ], + "spans": [ + { + "bbox": [ + 55, + 276, + 295, + 312 + ], + "type": "text", + "content": "For the filter " + }, + { + "bbox": [ + 55, + 276, + 295, + 312 + ], + "type": "inline_equation", + "content": "\\mathbf{F}(\\mathbf{X})" + }, + { + "bbox": [ + 55, + 276, + 295, + 312 + ], + "type": "text", + "content": " in the filter subspace " + }, + { + "bbox": [ + 55, + 276, + 295, + 312 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 55, + 276, + 295, + 312 + ], + "type": "text", + "content": " since " + }, + { + "bbox": [ + 55, + 276, + 295, + 312 + ], + "type": "inline_equation", + "content": "\\mathbf{F}(\\mathbf{X}) = \\mathrm{softmax}(\\mathbf{X}\\mathbf{W}_q\\mathbf{W}_k^\\top \\mathbf{X}^\\top)" + }, + { + "bbox": [ + 55, + 276, + 295, + 312 + ], + "type": "text", + "content": ", their elements " + }, + { + "bbox": [ + 55, + 276, + 295, + 312 + ], + "type": "inline_equation", + "content": "\\mathbf{F}(\\mathbf{X})[n,n^{\\prime}]" + }, + { + "bbox": [ + 55, + 276, + 295, + 312 + ], + "type": "text", + "content": " satisfy the following properties," + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 70, + 316, + 294, + 343 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 316, + 294, + 343 + ], + "spans": [ + { + "bbox": [ + 70, + 316, + 294, + 343 + ], + "type": "interline_equation", + "content": "0 < \\mathbf {F} (\\mathbf {X}) [ n, n ^ {\\prime} ] < 1, \\sum_ {n ^ {\\prime} \\in [ N ]} \\mathbf {F} (\\mathbf {X}) [ n, n ^ {\\prime} ] = 1. \\tag {13}", + "image_path": "1322ed510511e624c18cb98ea7511a2844bfb7f5856983f848c4b80deb163c2a.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 348, + 296, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 348, + 296, + 386 + ], + "spans": [ + { + "bbox": [ + 55, + 348, + 296, + 386 + ], + "type": "text", + "content": "Definition 4.1. Given " + }, + { + "bbox": [ + 55, + 348, + 296, + 386 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 348, + 296, + 386 + ], + "type": "text", + "content": " nodes " + }, + { + "bbox": [ + 55, + 348, + 296, + 386 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{x}_i\\}_{i=1}^N" + }, + { + "bbox": [ + 55, + 348, + 296, + 386 + ], + "type": "text", + "content": ", the set " + }, + { + "bbox": [ + 55, + 348, + 296, + 386 + ], + "type": "inline_equation", + "content": "\\mathcal{S}(\\mathbf{x}) = \\{\\sum_{i=1}^{N} \\lambda_i \\mathbf{x}_i \\mid \\lambda_i > 0, \\sum_i \\lambda_i = 1\\}" + }, + { + "bbox": [ + 55, + 348, + 296, + 386 + ], + "type": "text", + "content": " is a bounded convex set." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 391, + 296, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 391, + 296, + 464 + ], + "spans": [ + { + "bbox": [ + 55, + 391, + 296, + 464 + ], + "type": "text", + "content": "According to the Definition 4.1, when the projected nodes " + }, + { + "bbox": [ + 55, + 391, + 296, + 464 + ], + "type": "inline_equation", + "content": "\\mathbf{X}\\mathbf{W}^h\\in \\mathbb{R}^{N\\times C_o}" + }, + { + "bbox": [ + 55, + 391, + 296, + 464 + ], + "type": "text", + "content": " of each head " + }, + { + "bbox": [ + 55, + 391, + 296, + 464 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 55, + 391, + 296, + 464 + ], + "type": "text", + "content": " are given, the output " + }, + { + "bbox": [ + 55, + 391, + 296, + 464 + ], + "type": "inline_equation", + "content": "\\mathbf{O}^h [n,:] = \\sum_{n'\\in [N]}\\mathbf{F}^h (\\mathbf{X})[n,n']\\mathbf{X}[n']\\mathbf{W}^h" + }, + { + "bbox": [ + 55, + 391, + 296, + 464 + ], + "type": "text", + "content": " lies within " + }, + { + "bbox": [ + 55, + 391, + 296, + 464 + ], + "type": "inline_equation", + "content": "S(\\mathbf{X}\\mathbf{W}^h)" + }, + { + "bbox": [ + 55, + 391, + 296, + 464 + ], + "type": "text", + "content": ", which means the outputs " + }, + { + "bbox": [ + 55, + 391, + 296, + 464 + ], + "type": "inline_equation", + "content": "\\mathbf{O}^h [n,:]" + }, + { + "bbox": [ + 55, + 391, + 296, + 464 + ], + "type": "text", + "content": " are bounded. This property also applies to the multi-head attention mechanism." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 468, + 296, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 468, + 296, + 506 + ], + "spans": [ + { + "bbox": [ + 55, + 468, + 296, + 506 + ], + "type": "text", + "content": "Proposition 4.2. If " + }, + { + "bbox": [ + 55, + 468, + 296, + 506 + ], + "type": "inline_equation", + "content": "S^h" + }, + { + "bbox": [ + 55, + 468, + 296, + 506 + ], + "type": "text", + "content": " is a bounded convex set, " + }, + { + "bbox": [ + 55, + 468, + 296, + 506 + ], + "type": "inline_equation", + "content": "\\forall h = 1, \\dots, H" + }, + { + "bbox": [ + 55, + 468, + 296, + 506 + ], + "type": "text", + "content": ", the set " + }, + { + "bbox": [ + 55, + 468, + 296, + 506 + ], + "type": "inline_equation", + "content": "S = S^1 + \\dots + S^H = \\{ \\mathbf{x}^1 + \\dots + \\mathbf{x}^H \\mid \\forall \\mathbf{x}^h \\in S^h \\}" + }, + { + "bbox": [ + 55, + 468, + 296, + 506 + ], + "type": "text", + "content": " is also a bounded convex set." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 509, + 296, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 509, + 296, + 556 + ], + "spans": [ + { + "bbox": [ + 55, + 509, + 296, + 556 + ], + "type": "text", + "content": "According to Proposition 4.2, when the projected nodes " + }, + { + "bbox": [ + 55, + 509, + 296, + 556 + ], + "type": "inline_equation", + "content": "\\mathbf{X}\\mathbf{W}^h" + }, + { + "bbox": [ + 55, + 509, + 296, + 556 + ], + "type": "text", + "content": " are given, the outputs of the multi-head attention " + }, + { + "bbox": [ + 55, + 509, + 296, + 556 + ], + "type": "inline_equation", + "content": "\\mathbf{O}[n,:] = \\sum_{h=1}^{H} \\sum_{n' \\in [N]} \\mathbf{F}^h(\\mathbf{X})[n,n']\\mathbf{X}[n']\\mathbf{W}^h" + }, + { + "bbox": [ + 55, + 509, + 296, + 556 + ], + "type": "text", + "content": " are also bounded." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 561, + 295, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 561, + 295, + 624 + ], + "spans": [ + { + "bbox": [ + 54, + 561, + 295, + 624 + ], + "type": "text", + "content": "Proposition 4.3. After applying the subspace coefficients " + }, + { + "bbox": [ + 54, + 561, + 295, + 624 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 54, + 561, + 295, + 624 + ], + "type": "text", + "content": ", the outputs of the reconstructed filter subspace " + }, + { + "bbox": [ + 54, + 561, + 295, + 624 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{F}}_{\\alpha}" + }, + { + "bbox": [ + 54, + 561, + 295, + 624 + ], + "type": "text", + "content": " form a set " + }, + { + "bbox": [ + 54, + 561, + 295, + 624 + ], + "type": "inline_equation", + "content": "\\hat{S} = \\{\\sum_{h=1}^{H} \\sum_{i=1}^{H} \\boldsymbol{\\alpha}[h,i]\\mathbf{F}^i(\\mathbf{X})[n,n']\\mathbf{X}\\mathbf{W}^h \\mid \\mathbf{F}^i(\\mathbf{X})[n,n'] > 0, \\sum_{j} \\mathbf{F}^i(\\mathbf{X})[n,n'] = 1, \\boldsymbol{\\alpha}[h,i] \\in \\mathbb{R}\\}" + }, + { + "bbox": [ + 54, + 561, + 295, + 624 + ], + "type": "text", + "content": ", which satisfies " + }, + { + "bbox": [ + 54, + 561, + 295, + 624 + ], + "type": "inline_equation", + "content": "S \\subseteq \\hat{S}" + }, + { + "bbox": [ + 54, + 561, + 295, + 624 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 628, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 628, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 628, + 296, + 715 + ], + "type": "text", + "content": "According to Proposition 4.3, when applying filter subspace coefficients " + }, + { + "bbox": [ + 55, + 628, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 628, + 296, + 715 + ], + "type": "text", + "content": ", we can reconstruct the filter subspace " + }, + { + "bbox": [ + 55, + 628, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 55, + 628, + 296, + 715 + ], + "type": "text", + "content": " and expand the set " + }, + { + "bbox": [ + 55, + 628, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{S}" + }, + { + "bbox": [ + 55, + 628, + 296, + 715 + ], + "type": "text", + "content": ". Specifically, given " + }, + { + "bbox": [ + 55, + 628, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{X}\\mathbf{W}^h" + }, + { + "bbox": [ + 55, + 628, + 296, + 715 + ], + "type": "text", + "content": ", the output " + }, + { + "bbox": [ + 55, + 628, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{O}[n,:] = \\sum_{h=1}^{H} \\sum_{i=1}^{H} \\alpha[h,i]\\mathbf{F}^i(\\mathbf{X})[n,n']\\mathbf{X}\\mathbf{W}^h" + }, + { + "bbox": [ + 55, + 628, + 296, + 715 + ], + "type": "text", + "content": " can be in the set " + }, + { + "bbox": [ + 55, + 628, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{S}} \\setminus \\mathcal{S}" + }, + { + "bbox": [ + 55, + 628, + 296, + 715 + ], + "type": "text", + "content": ". Our method obtains an expanded set, enhancing the expressiveness of the model. We provide the support of our analyses in Appendix A." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 275, + 443, + 289 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 275, + 443, + 289 + ], + "spans": [ + { + "bbox": [ + 313, + 275, + 443, + 289 + ], + "type": "text", + "content": "4.1. Validation Experiment" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 293, + 555, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 293, + 555, + 341 + ], + "spans": [ + { + "bbox": [ + 313, + 293, + 555, + 341 + ], + "type": "text", + "content": "We assess the effectiveness of the subspace coefficients using a single attention layer. Given an input graph " + }, + { + "bbox": [ + 313, + 293, + 555, + 341 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 313, + 293, + 555, + 341 + ], + "type": "text", + "content": ", the output of the attention layer is calculated as " + }, + { + "bbox": [ + 313, + 293, + 555, + 341 + ], + "type": "inline_equation", + "content": "\\mathbf{O} = \\mathbf{F}(\\mathbf{X})\\mathbf{X}\\mathbf{W}_v" + }, + { + "bbox": [ + 313, + 293, + 555, + 341 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 293, + 555, + 341 + ], + "type": "inline_equation", + "content": "\\mathbf{F}(\\mathbf{X}) = \\mathrm{softmax}(\\mathbf{X}\\mathbf{W}_q\\mathbf{W}_k^\\intercal\\mathbf{X}^\\intercal)" + }, + { + "bbox": [ + 313, + 293, + 555, + 341 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 341, + 564, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 341, + 564, + 425 + ], + "spans": [ + { + "bbox": [ + 313, + 341, + 564, + 425 + ], + "type": "text", + "content": "Our goal is to optimize " + }, + { + "bbox": [ + 313, + 341, + 564, + 425 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_q" + }, + { + "bbox": [ + 313, + 341, + 564, + 425 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 341, + 564, + 425 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_k" + }, + { + "bbox": [ + 313, + 341, + 564, + 425 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 341, + 564, + 425 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_v" + }, + { + "bbox": [ + 313, + 341, + 564, + 425 + ], + "type": "text", + "content": " to transform the input graph " + }, + { + "bbox": [ + 313, + 341, + 564, + 425 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 313, + 341, + 564, + 425 + ], + "type": "text", + "content": " into the target graph " + }, + { + "bbox": [ + 313, + 341, + 564, + 425 + ], + "type": "inline_equation", + "content": "\\mathbf{O}" + }, + { + "bbox": [ + 313, + 341, + 564, + 425 + ], + "type": "text", + "content": " as shown in Figure 3. The input graph contains 8 nodes " + }, + { + "bbox": [ + 313, + 341, + 564, + 425 + ], + "type": "inline_equation", + "content": "\\mathbf{X} = [[-1,1],[0,1],[1,1],[1,0],[1,-1],[0,-1],[-1,-1],[-1,0]]" + }, + { + "bbox": [ + 313, + 341, + 564, + 425 + ], + "type": "text", + "content": " positioned in a range of " + }, + { + "bbox": [ + 313, + 341, + 564, + 425 + ], + "type": "inline_equation", + "content": "[-1,1]" + }, + { + "bbox": [ + 313, + 341, + 564, + 425 + ], + "type": "text", + "content": " and the target graph " + }, + { + "bbox": [ + 313, + 341, + 564, + 425 + ], + "type": "inline_equation", + "content": "\\mathbf{O} = [[-1,1],[0,2],[1,1],[2,0],[1,-1],[0,-2],[-1,-1],[-2,0]]" + }, + { + "bbox": [ + 313, + 341, + 564, + 425 + ], + "type": "text", + "content": ", with node coordinates in the range of " + }, + { + "bbox": [ + 313, + 341, + 564, + 425 + ], + "type": "inline_equation", + "content": "[-2,2]" + }, + { + "bbox": [ + 313, + 341, + 564, + 425 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 425, + 556, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 425, + 556, + 557 + ], + "spans": [ + { + "bbox": [ + 313, + 425, + 556, + 557 + ], + "type": "text", + "content": "As shown in Figure 3, by only tuning the " + }, + { + "bbox": [ + 313, + 425, + 556, + 557 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_q" + }, + { + "bbox": [ + 313, + 425, + 556, + 557 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 425, + 556, + 557 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_k" + }, + { + "bbox": [ + 313, + 425, + 556, + 557 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 313, + 425, + 556, + 557 + ], + "type": "inline_equation", + "content": "\\mathbf{F}(\\mathbf{X}) = \\mathrm{softmax}(\\mathbf{X}\\mathbf{W}_q\\mathbf{W}_k^\\top \\mathbf{X}^\\top)" + }, + { + "bbox": [ + 313, + 425, + 556, + 557 + ], + "type": "text", + "content": ", the output is constrained inside the range of the input, which is aligned with Proposition 4.1. When we tune " + }, + { + "bbox": [ + 313, + 425, + 556, + 557 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_q" + }, + { + "bbox": [ + 313, + 425, + 556, + 557 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 425, + 556, + 557 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_k" + }, + { + "bbox": [ + 313, + 425, + 556, + 557 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 425, + 556, + 557 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_v" + }, + { + "bbox": [ + 313, + 425, + 556, + 557 + ], + "type": "text", + "content": ", the output exceeds the input range but still cannot fully match the target graph. However, when we tune the subspace " + }, + { + "bbox": [ + 313, + 425, + 556, + 557 + ], + "type": "inline_equation", + "content": "\\mathbf{F}(\\mathbf{X})" + }, + { + "bbox": [ + 313, + 425, + 556, + 557 + ], + "type": "text", + "content": " and subspace coefficient " + }, + { + "bbox": [ + 313, + 425, + 556, + 557 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 425, + 556, + 557 + ], + "type": "text", + "content": ", the attention layer can successfully transform the input graph into the target graph as in Figure 3. This observation is aligned with Proposition 4.3, and indicates that the proposed Coeff-Tuning can effectively expand the feature space of the multi-head attention layer." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 562, + 434, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 562, + 434, + 574 + ], + "spans": [ + { + "bbox": [ + 313, + 562, + 434, + 574 + ], + "type": "text", + "content": "4.2. Complexity Analysis" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 579, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 579, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 579, + 556, + 713 + ], + "type": "text", + "content": "We now formally analyze the complexity of the proposed Coeff-Tuning, and compare it with other baseline PEFT methods. The number of parameters required for our method is " + }, + { + "bbox": [ + 313, + 579, + 556, + 713 + ], + "type": "inline_equation", + "content": "H \\times H" + }, + { + "bbox": [ + 313, + 579, + 556, + 713 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 579, + 556, + 713 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 313, + 579, + 556, + 713 + ], + "type": "text", + "content": " is the number of heads at each attention layer. Considering " + }, + { + "bbox": [ + 313, + 579, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_q, \\mathbf{W}_k, \\mathbf{W}_v \\in \\mathbb{R}^{C_i \\times C_o}" + }, + { + "bbox": [ + 313, + 579, + 556, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 579, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_o \\in \\mathbb{R}^{C_o \\times C_o}" + }, + { + "bbox": [ + 313, + 579, + 556, + 713 + ], + "type": "text", + "content": " of the attention layer, our method introduces relative " + }, + { + "bbox": [ + 313, + 579, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\frac{H^2}{3C_iC_o + C_o^2}" + }, + { + "bbox": [ + 313, + 579, + 556, + 713 + ], + "type": "text", + "content": " parameters. In modern large transformer models, we typically have " + }, + { + "bbox": [ + 313, + 579, + 556, + 713 + ], + "type": "inline_equation", + "content": "C_i = 2048" + }, + { + "bbox": [ + 313, + 579, + 556, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 579, + 556, + 713 + ], + "type": "inline_equation", + "content": "C_o = 1280" + }, + { + "bbox": [ + 313, + 579, + 556, + 713 + ], + "type": "text", + "content": ". With " + }, + { + "bbox": [ + 313, + 579, + 556, + 713 + ], + "type": "inline_equation", + "content": "H = 10" + }, + { + "bbox": [ + 313, + 579, + 556, + 713 + ], + "type": "text", + "content": ", our method requires approximately 0.001% additional parameters. Compared with other PEFT methods such as LoRA, our method introduces" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20150" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 57, + 86, + 555, + 229 + ], + "blocks": [ + { + "bbox": [ + 211, + 70, + 400, + 82 + ], + "lines": [ + { + "bbox": [ + 211, + 70, + 400, + 82 + ], + "spans": [ + { + "bbox": [ + 211, + 70, + 400, + 82 + ], + "type": "text", + "content": "Table 1. Fine-tuning Results on VTAB-1k with ViT" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 86, + 555, + 229 + ], + "lines": [ + { + "bbox": [ + 57, + 86, + 555, + 229 + ], + "spans": [ + { + "bbox": [ + 57, + 86, + 555, + 229 + ], + "type": "table", + "html": "
NaturalSpecializedStructured
C100Cal101DTDF102PetsSVHNS397CameE.SATRes45ReinoCle_ctmCle_disDMLabKITTIdSlocationdS oriS_aziS_eleMeanParam.GTL_OPs
Full fine-tuning68.987.764.397.286.987.438.879.795.784.273.956.358.641.765.557.546.725.729.165.5785.845.15
Adapter74.186.163.297.787.034.650.876.388.073.170.545.737.431.253.230.325.413.822.155.820.23416.44
BitFit72.887.059.297.585.359.951.478.791.672.969.861.555.632.455.966.640.015.725.162.050.10515.08
VPT77.786.962.697.587.374.551.278.292.075.672.950.558.640.567.168.736.120.234.164.850.0616.68
LoRA67.191.469.498.890.485.354.084.995.384.473.682.969.249.881.781.848.332.844.272.910.30515.74
SPT-Adapter72.993.272.599.391.484.655.285.396.083.975.882.268.049.380.082.451.931.741.273.520.26616.44
Linear Probing63.485.063.297.086.336.651.078.587.568.674.034.330.633.255.412.520.09.619.252.94015.05
Coeff. α Only (Ours)69.590.874.599.190.983.051.183.695.384.675.857.459.146.578.675.051.425.634.269.780.00215.06
FACT-TT≤1671.389.670.798.991.087.854.685.295.583.475.782.069.049.880.079.248.434.241.473.040.04015.14
w/ Coeff-Tuning (Ours)72.189.771.899.191.288.954.885.395.784.275.383.068.851.081.380.049.335.042.773.640.04215.15
SNELL-873.792.772.499.291.489.255.484.996.186.475.284.068.553.581.082.749.933.939.274.170.26815.47
w/ Coeff-Tuning (Ours)74.292.973.199.291.690.355.685.296.286.375.884.168.853.781.384.052.135.039.974.700.27015.48
", + "image_path": "708fe0328f1842289aad51ed3885a124efa0037b3b92bf5b56dfbea1fe3fbb18.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 248, + 296, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 248, + 296, + 323 + ], + "spans": [ + { + "bbox": [ + 55, + 248, + 296, + 323 + ], + "type": "text", + "content": "relative " + }, + { + "bbox": [ + 55, + 248, + 296, + 323 + ], + "type": "inline_equation", + "content": "\\frac{H^2}{3rC_i + 5rC_o}" + }, + { + "bbox": [ + 55, + 248, + 296, + 323 + ], + "type": "text", + "content": " parameters, where " + }, + { + "bbox": [ + 55, + 248, + 296, + 323 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 55, + 248, + 296, + 323 + ], + "type": "text", + "content": " is the LoRA rank. Using the same values for " + }, + { + "bbox": [ + 55, + 248, + 296, + 323 + ], + "type": "inline_equation", + "content": "C_i" + }, + { + "bbox": [ + 55, + 248, + 296, + 323 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 248, + 296, + 323 + ], + "type": "inline_equation", + "content": "C_o" + }, + { + "bbox": [ + 55, + 248, + 296, + 323 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 248, + 296, + 323 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 55, + 248, + 296, + 323 + ], + "type": "text", + "content": ", and setting " + }, + { + "bbox": [ + 55, + 248, + 296, + 323 + ], + "type": "inline_equation", + "content": "r = 4" + }, + { + "bbox": [ + 55, + 248, + 296, + 323 + ], + "type": "text", + "content": ", our approach requires less than " + }, + { + "bbox": [ + 55, + 248, + 296, + 323 + ], + "type": "inline_equation", + "content": "0.2\\%" + }, + { + "bbox": [ + 55, + 248, + 296, + 323 + ], + "type": "text", + "content": " additional parameters compared to LoRA and its variants. This also indicates that our Coeff-Tuning can maintain almost the same quantity of parameters when combined with other PEFT methods." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 334, + 138, + 348 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 334, + 138, + 348 + ], + "spans": [ + { + "bbox": [ + 55, + 334, + 138, + 348 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 354, + 296, + 451 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 354, + 296, + 451 + ], + "spans": [ + { + "bbox": [ + 55, + 354, + 296, + 451 + ], + "type": "text", + "content": "In this section, we validate the proposed Coeff-Tuning with various tasks and datasets, including tuning the vision transformer [11] on few-shot image classification, tuning text-to-image diffusion model [51] for concept customization, and tuning vision-language BART [8] on image-to-text understanding. Combining with other weight-based PEFT methods, the proposed Coeff-Tuning consistently achieves superior results with negligible parameters added." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 460, + 218, + 472 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 460, + 218, + 472 + ], + "spans": [ + { + "bbox": [ + 55, + 460, + 218, + 472 + ], + "type": "text", + "content": "5.1. Few-shot Image Classification" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 477, + 296, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 477, + 296, + 538 + ], + "spans": [ + { + "bbox": [ + 55, + 477, + 296, + 538 + ], + "type": "text", + "content": "we first validate the proposed method by fine-tuning Vision Transformers on a series of downstream datasets for few-shot image classification. Our method achieves superior results while maintaining a high parameter efficiency compared with baseline methods." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 544, + 296, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 544, + 296, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 544, + 296, + 605 + ], + "type": "text", + "content": "Experimental Settings. We utilize the VTAB-1k [81] dataset which comprises 19 diverse computer vision tasks across three domains: standard natural imagery, specialized capture, and synthetic/structured environments. The benchmark evaluates model performance on various task" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 51, + 619, + 291, + 712 + ], + "blocks": [ + { + "bbox": [ + 88, + 609, + 262, + 619 + ], + "lines": [ + { + "bbox": [ + 88, + 609, + 262, + 619 + ], + "spans": [ + { + "bbox": [ + 88, + 609, + 262, + 619 + ], + "type": "text", + "content": "Table 2. Fine-tuning Results on the FGVC with ViT." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 51, + 619, + 291, + 712 + ], + "lines": [ + { + "bbox": [ + 51, + 619, + 291, + 712 + ], + "spans": [ + { + "bbox": [ + 51, + 619, + 291, + 712 + ], + "type": "table", + "html": "
CUB-200NABirdsFlowersS-DogsS-CarsMean
Full fine-tuning87.382.798.889.484.588.5
Linear Probing85.375.997.986.251.379.3
Adapter87.384.398.488.868.485.5
BitFit88.484.298.891.279.488.4
VPT86.778.898.490.768.784.6
LoRA84.979.098.188.179.886.0
SPT-LoRA88.683.499.591.487.390.1
SNELL-889.686.899.392.189.991.5
w/ Coeff-Tuning (Ours)89.887.199.592.290.791.9
", + "image_path": "e7580b354df23c7052215a6949f33da330de507f399d2e8b1cd4b49a3180fcfa.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 249, + 555, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 249, + 555, + 334 + ], + "spans": [ + { + "bbox": [ + 313, + 249, + 555, + 334 + ], + "type": "text", + "content": "objectives. With only 1,000 training samples per task, it presents a significant few-shot learning challenge. We also benchmark our method on five fine-grained datasets in few-shot settings, including CUB-200-2011 [69], NABirds [67], Oxford Flowers [44], Stanford Dogs [25] and Stanford Cars [13]. A detailed description is provided in Appendix B.2." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 335, + 555, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 335, + 555, + 407 + ], + "spans": [ + { + "bbox": [ + 313, + 335, + 555, + 407 + ], + "type": "text", + "content": "As for the vision transformer, we follow VPT [23], SSF [35], and select ViT-B/16 [11] pre-trained on ImageNet-21K [9] as the base model for fine-tuning. We benchmark the proposed Coeff-Tuning with a series of PEFT methods for few-shot learning including SSF [35], VPT [23], Adapter [20], LoRA [21], and BitFit [79]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 408, + 556, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 408, + 556, + 529 + ], + "spans": [ + { + "bbox": [ + 313, + 408, + 556, + 529 + ], + "type": "text", + "content": "We present two settings of Coeff-Tuning in this task. We first explore the representation ability of tuning subspace coefficient " + }, + { + "bbox": [ + 313, + 408, + 556, + 529 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 408, + 556, + 529 + ], + "type": "text", + "content": " only, which tunes the transformer bottleneck with only tuning " + }, + { + "bbox": [ + 313, + 408, + 556, + 529 + ], + "type": "inline_equation", + "content": "H^2" + }, + { + "bbox": [ + 313, + 408, + 556, + 529 + ], + "type": "text", + "content": " (144 for ViT [11]) parameters for each transformer block. We also show the performance of combining Coeff-Tuning and SSF [35] to demonstrate that our method can plug and play with other PEFT methods. We set the coefficient dropout rates as " + }, + { + "bbox": [ + 313, + 408, + 556, + 529 + ], + "type": "inline_equation", + "content": "p = 0" + }, + { + "bbox": [ + 313, + 408, + 556, + 529 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 408, + 556, + 529 + ], + "type": "inline_equation", + "content": "p = 0.1" + }, + { + "bbox": [ + 313, + 408, + 556, + 529 + ], + "type": "text", + "content": " for these two settings respectively, and provide other hyperparameters in Appendix B.2." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 545, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 556, + 713 + ], + "type": "text", + "content": "Results. As shown in Table 1, the proposed Coeff-Tuning shows superior performances in both settings. Compared with the linear probing baseline which achieves a mean accuracy of 61.69, Coeff-Tuning obtains a significantly higher result of 73.49 with tuning only " + }, + { + "bbox": [ + 313, + 545, + 556, + 713 + ], + "type": "inline_equation", + "content": "(12 * 12) * 12 = 1" + }, + { + "bbox": [ + 313, + 545, + 556, + 713 + ], + "type": "text", + "content": ", 728 additional subspace coefficients " + }, + { + "bbox": [ + 313, + 545, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 545, + 556, + 713 + ], + "type": "text", + "content": " in ViT [11] (ViT [11] has 12 attention layers, each attention has 12 heads). Thus, we achieve a large performance increase in few-shot image classification " + }, + { + "bbox": [ + 313, + 545, + 556, + 713 + ], + "type": "inline_equation", + "content": "(11.8^{\\uparrow})" + }, + { + "bbox": [ + 313, + 545, + 556, + 713 + ], + "type": "text", + "content": " with negligible additional parameter costs. Note that this result has already surpassed almost all baselines. We further achieve the best performance by combining Coeff-Tuning with SSF while maintaining almost the same parameter cost." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "text", + "content": "20151" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 82, + 71, + 526, + 308 + ], + "blocks": [ + { + "bbox": [ + 82, + 71, + 526, + 308 + ], + "lines": [ + { + "bbox": [ + 82, + 71, + 526, + 308 + ], + "spans": [ + { + "bbox": [ + 82, + 71, + 526, + 308 + ], + "type": "image", + "image_path": "2941f246d25d534a5c8707518a81d42e9c918146ed65822cd22a66b4a3805b00.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 313, + 555, + 360 + ], + "lines": [ + { + "bbox": [ + 55, + 313, + 555, + 360 + ], + "spans": [ + { + "bbox": [ + 55, + 313, + 555, + 360 + ], + "type": "text", + "content": "Figure 4. Results on Concept Customization. Compared to the baseline method (middle row), Coeff-Tuning (bottom row) leads to higher concept fidelity and text alignment of the generated images. For instance, our method accurately generates images of a “" + }, + { + "bbox": [ + 55, + 313, + 555, + 360 + ], + "type": "inline_equation", + "content": "\\langle V \\rangle" + }, + { + "bbox": [ + 55, + 313, + 555, + 360 + ], + "type": "text", + "content": " plushie sloth in the snow,” while the baseline fails to align with the text prompt. Additionally, our method generates a “" + }, + { + "bbox": [ + 55, + 313, + 555, + 360 + ], + "type": "inline_equation", + "content": "\\langle V \\rangle" + }, + { + "bbox": [ + 55, + 313, + 555, + 360 + ], + "type": "text", + "content": " wooden pot” with higher fidelity to the input image." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 58, + 385, + 296, + 446 + ], + "blocks": [ + { + "bbox": [ + 78, + 370, + 271, + 381 + ], + "lines": [ + { + "bbox": [ + 78, + 370, + 271, + 381 + ], + "spans": [ + { + "bbox": [ + 78, + 370, + 271, + 381 + ], + "type": "text", + "content": "Table 3. Concept Customization Results with SDXL." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 385, + 296, + 446 + ], + "lines": [ + { + "bbox": [ + 58, + 385, + 296, + 446 + ], + "spans": [ + { + "bbox": [ + 58, + 385, + 296, + 446 + ], + "type": "table", + "html": "
Method# Params (%)Text Align.Concept Align.Diversity
DoRA0.03710.330.7310.9
Ours + DoRA0.03740.340.7511.37
LoRA0.03060.330.7410.08
Ours + LoRA0.03090.340.7510.45
", + "image_path": "69606d837a52a1f2b97bdb5aad104845b812ee214cca357fdbe320bed1e46b85.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 456, + 262, + 470 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 456, + 262, + 470 + ], + "spans": [ + { + "bbox": [ + 55, + 456, + 262, + 470 + ], + "type": "text", + "content": "5.2. Personalized Text-to-Image Generation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 475, + 296, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 475, + 296, + 643 + ], + "spans": [ + { + "bbox": [ + 55, + 475, + 296, + 643 + ], + "type": "text", + "content": "We now show the effectiveness of Coeff-Tuning with the concept customization with Stable Diffusion (SDXL) [51, 55]. We validate the effectiveness of Coeff-Tuning by combining it with baseline PEFT methods, LoRA [21] and DoRA [38], and compare the performance with baselines without tuning " + }, + { + "bbox": [ + 55, + 475, + 296, + 643 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 475, + 296, + 643 + ], + "type": "text", + "content": ". Specifically, we fine-tune SDXL [51] to associate the special token with a concept image from Dreambooth [56] and CustomConcept101 [30] datasets. We choose 10 different concepts, and fine-tune the model on " + }, + { + "bbox": [ + 55, + 475, + 296, + 643 + ], + "type": "inline_equation", + "content": "4 \\sim 9" + }, + { + "bbox": [ + 55, + 475, + 296, + 643 + ], + "type": "text", + "content": " images. We utilize 25 different text prompts following [53] and generate 4 images for each prompt. We evaluate the fidelity, diversity, and text-to-image alignment of the generated images. Other training details are described in Appendix B.3." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "content": "Evaluation metrics. We assess the concept alignment with the average cosine similarity between CLIP embeddings [54] of the generated images and the concept images. The diversity of generated images is measured by the Vendi score [12] calculated with the DINOv2 embeddings [46]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 372, + 555, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 372, + 555, + 396 + ], + "spans": [ + { + "bbox": [ + 313, + 372, + 555, + 396 + ], + "type": "text", + "content": "The text alignment of the generated images is measured via average cosine similarity in the CLIP feature space [54]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 401, + 555, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 401, + 555, + 497 + ], + "spans": [ + { + "bbox": [ + 313, + 401, + 555, + 497 + ], + "type": "text", + "content": "Results. The quantitative results are shown in Table 3. By combining with LoRA [21] or DoRA [38], the proposed Coeff-Tuning generates images with greater diversity and better alignment with the both text prompt and input concept images. The qualitative results of our method compared with LoRA [21] are shown in Figure 4. The generated images with Coeff-Tuning have higher concept fidelity, preserving more characteristics from the input images." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 497, + 556, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 497, + 556, + 617 + ], + "spans": [ + { + "bbox": [ + 313, + 497, + 556, + 617 + ], + "type": "text", + "content": "As an example, consider the concept “" + }, + { + "bbox": [ + 313, + 497, + 556, + 617 + ], + "type": "inline_equation", + "content": "\\langle V \\rangle" + }, + { + "bbox": [ + 313, + 497, + 556, + 617 + ], + "type": "text", + "content": " cat” as shown in the 4th column in Figure 4, the image generated using our method retains the cat's color from the input image, while the one generated by the baseline method alter the color. Additionally, the backgrounds in images generated with our method show more detailed features compared to those from the baseline method. Coeff-Tuning achieves the above improvements by introducing a limited number of additional parameters. We provide more generated images in Appendix B.3." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 624, + 463, + 637 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 624, + 463, + 637 + ], + "spans": [ + { + "bbox": [ + 313, + 624, + 463, + 637 + ], + "type": "text", + "content": "5.3. Image-Text Understanding" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 642, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 555, + 715 + ], + "type": "text", + "content": "As we have shown the effectiveness of Coeff-Tuning in fine-tuning models on both few-shot image classification and personalized image generation, we now validate it on multitask image-text understanding, which is a much larger-scale multi-modal fine-tuning task that requires a higher model capacity. By combining with baseline PEFT meth" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20152" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 155, + 86, + 451, + 171 + ], + "blocks": [ + { + "bbox": [ + 201, + 70, + 408, + 82 + ], + "lines": [ + { + "bbox": [ + 201, + 70, + 408, + 82 + ], + "spans": [ + { + "bbox": [ + 201, + 70, + 408, + 82 + ], + "type": "text", + "content": "Table 4. Multi-Model Fine-tuning Results with VL-Bart." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 155, + 86, + 451, + 171 + ], + "lines": [ + { + "bbox": [ + 155, + 86, + 451, + 171 + ], + "spans": [ + { + "bbox": [ + 155, + 86, + 451, + 171 + ], + "type": "table", + "html": "
Method# Params (%)VQAV2GQANVLR2COCO CapAvg.
Full Model10066.956.773.7112.077.3
LoRA5.9465.253.671.9115.376.5
Ours + LoRA5.9466.254.173.0115.277.1
DoRA5.9665.854.773.1115.977.4
Ours + DoRA5.9666.454.973.0116.477.7
", + "image_path": "40a839bea6f4f9035dde5b929ad5dbb8b30ca1b9489e8cb430374fea293d4eed.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 184, + 295, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 184, + 295, + 209 + ], + "spans": [ + { + "bbox": [ + 55, + 184, + 295, + 209 + ], + "type": "text", + "content": "ods, our method again achieves superior performance with negligible parameter increase." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 230, + 296, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 230, + 296, + 399 + ], + "spans": [ + { + "bbox": [ + 55, + 230, + 296, + 399 + ], + "type": "text", + "content": "Experimental Settings. We evaluate our method on VL-BART [8] multi-modal transformer, which utilizes CLIP-ResNet101 [54] for vision encoding and sends the visual features to BART-Base [31] for image-text joint modeling. Following [65], we evaluate our method with the multitask fine-tuning of VL-Bart [8] on four large-scale image-text understanding tasks simultaneously: VQAv2 [14] and GQA [22] for visual question answering, NLVR2 [64] for visual reasoning, and MSCOCO [5] for image captioning. Following the plug-and-play practice, we combine the proposed Coeff-Tuning with LoRA [21] and DoRA [38] respectively, and set the subspace coefficient dropout rate to " + }, + { + "bbox": [ + 55, + 230, + 296, + 399 + ], + "type": "inline_equation", + "content": "p = 0.2" + }, + { + "bbox": [ + 55, + 230, + 296, + 399 + ], + "type": "text", + "content": ". Detailed descriptions of other settings and hyperparameters are provided in Appendix B.4." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 415, + 296, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 415, + 296, + 523 + ], + "spans": [ + { + "bbox": [ + 55, + 415, + 296, + 523 + ], + "type": "text", + "content": "Results. As shown in Table 4, our method surpasses baseline LoRA [21] and DoRA [38] respectively in almost all image-text datasets consistently, and achieves higher average scores. The Coeff-Tuning combined with LoRA [21] obtains comparable results with full-model fine-tuning, and it achieves better results than full-model tuning when it is combined with DoRA [38]. Notably, these improvements are achieved while maintaining the same parameter budgets as we only tune a hundred parameters for each attention." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 534, + 149, + 547 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 149, + 547 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 149, + 547 + ], + "type": "text", + "content": "5.4. Ablation Study" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 553, + 295, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 553, + 295, + 578 + ], + "spans": [ + { + "bbox": [ + 55, + 553, + 295, + 578 + ], + "type": "text", + "content": "In this section, we examine the effectiveness of each component of our design with ablation experiments." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 598, + 295, + 621 + ], + "lines": [ + { + "bbox": [ + 55, + 598, + 295, + 621 + ], + "spans": [ + { + "bbox": [ + 55, + 598, + 295, + 621 + ], + "type": "text", + "content": "Table 5. Ablation studies on (a) residual parameterization and (b) coefficient dropout showing their effects on model performance." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "type": "table", + "bbox": [ + 69, + 640, + 180, + 696 + ], + "blocks": [ + { + "bbox": [ + 69, + 640, + 180, + 696 + ], + "lines": [ + { + "bbox": [ + 69, + 640, + 180, + 696 + ], + "spans": [ + { + "bbox": [ + 69, + 640, + 180, + 696 + ], + "type": "table", + "html": "
ParametrizationAvg.
α w/ rand. init.44.28
α w/ identity init.61.57
I + α, zero init.73.49
", + "image_path": "64ca2f7436c2f55894d7e53c545b86d4dae545183619cb03d813627467ce15d4.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 189, + 630, + 293, + 697 + ], + "blocks": [ + { + "bbox": [ + 118, + 698, + 129, + 707 + ], + "lines": [ + { + "bbox": [ + 118, + 698, + 129, + 707 + ], + "spans": [ + { + "bbox": [ + 118, + 698, + 129, + 707 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 189, + 630, + 293, + 697 + ], + "lines": [ + { + "bbox": [ + 189, + 630, + 293, + 697 + ], + "spans": [ + { + "bbox": [ + 189, + 630, + 293, + 697 + ], + "type": "table", + "html": "
Dropout pAvg. Score
0.075.1
0.177.3
0.277.7
0.472.3
", + "image_path": "e7fc21c7e75b61732e582127804b9055368fa67c24e2848449710fedc543b72b.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 231, + 698, + 242, + 707 + ], + "lines": [ + { + "bbox": [ + 231, + 698, + 242, + 707 + ], + "spans": [ + { + "bbox": [ + 231, + 698, + 242, + 707 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 184, + 555, + 340 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 184, + 555, + 340 + ], + "spans": [ + { + "bbox": [ + 313, + 184, + 555, + 340 + ], + "type": "text", + "content": "Residual Parameterization. We first demonstrate the importance of residual parameterization in Eq. 11 in Coeff-Tuning. We compare our residual parameterization with the direct parameterization where we directly tune " + }, + { + "bbox": [ + 313, + 184, + 555, + 340 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 184, + 555, + 340 + ], + "type": "text", + "content": " without adding it to the identity matrix, and initialize it with either a random matrix or an identity matrix. We adopt the few-shot tuning task as in Sec. 5.1 and only tune " + }, + { + "bbox": [ + 313, + 184, + 555, + 340 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 184, + 555, + 340 + ], + "type": "text", + "content": ". Results in Table 5a show that models with residual parameterization achieve superior performance than direct parameterization with two different initializations. This validates our design that maintaining the original filter subspace while learning the residual combination helps preserve pre-trained knowledge and leads to better fine-tuning performance." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 349, + 556, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 349, + 556, + 482 + ], + "spans": [ + { + "bbox": [ + 313, + 349, + 556, + 482 + ], + "type": "text", + "content": "Subspace Coefficient Dropout. The dropout regularization on the subspace coefficient as in Eq. 12 plays a crucial role in preventing overfitting. We adopt the experiment setting in multi-modal fine-tuning as in Sec. 5.3. We experiment with different dropout rates " + }, + { + "bbox": [ + 313, + 349, + 556, + 482 + ], + "type": "inline_equation", + "content": "p \\in \\{0, 0.1, 0.2, 0.4\\}" + }, + { + "bbox": [ + 313, + 349, + 556, + 482 + ], + "type": "text", + "content": " during training. The results in Table 5b demonstrate that a moderate dropout rate (" + }, + { + "bbox": [ + 313, + 349, + 556, + 482 + ], + "type": "inline_equation", + "content": "p = 0.2" + }, + { + "bbox": [ + 313, + 349, + 556, + 482 + ], + "type": "text", + "content": ") yields the best performance while either a lower or higher dropout rate leads to sub-optimal results. This suggests that applying dropout directly on the tuned subspace coefficient " + }, + { + "bbox": [ + 313, + 349, + 556, + 482 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 349, + 556, + 482 + ], + "type": "text", + "content": " can effectively prevent overfitting and lead to better generalization." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 491, + 388, + 504 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 491, + 388, + 504 + ], + "spans": [ + { + "bbox": [ + 313, + 491, + 388, + 504 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 511, + 556, + 704 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 511, + 556, + 704 + ], + "spans": [ + { + "bbox": [ + 313, + 511, + 556, + 704 + ], + "type": "text", + "content": "In this paper, we presented Coeff-Tuning, a new parameter-efficient fine-tuning approach that enhances multi-head attention's expressiveness by learning a set of subspace coefficients. By reformulating attention as a graph convolution with a filter subspace, we demonstrated both theoretically and empirically that our method expands the feature space of multi-head attention beyond the original convex hull constraints. The proposed Coeff-Tuning can be seamlessly integrated with existing PEFT methods while introducing negligible additional parameters. Through extensive experiments across diverse tasks and modalities, we consistently demonstrated superior performance compared to competitive PEFT baselines. Our method achieved significant improvements in various metrics while maintaining high parameter efficiency, suggesting a promising direction for enhancing the adaptability of large pre-trained transformers." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20153" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 61, + 91, + 294, + 113 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 294, + 113 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 294, + 113 + ], + "type": "text", + "content": "[1] Tom B Brown. Language models are few-shot learners. arXiv preprint arXiv:2005.14165, 2020. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 114, + 295, + 158 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 114, + 295, + 158 + ], + "spans": [ + { + "bbox": [ + 61, + 114, + 295, + 158 + ], + "type": "text", + "content": "[2] Shoufa Chen, Chongjian Ge, Zhan Tong, Jiangliu Wang, Yibing Song, Jue Wang, and Ping Luo. Adaptformer: Adapting vision transformers for scalable visual recognition. Advances in Neural Information Processing Systems, 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 159, + 295, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 159, + 295, + 192 + ], + "spans": [ + { + "bbox": [ + 62, + 159, + 295, + 192 + ], + "type": "text", + "content": "[3] Wei Chen, Zichen Miao, and Qiang Qiu. Inner product-based neural network similarity. Advances in Neural Information Processing Systems, 36, 2024. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 194, + 295, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 194, + 295, + 225 + ], + "spans": [ + { + "bbox": [ + 62, + 194, + 295, + 225 + ], + "type": "text", + "content": "[4] Wei Chen, Zichen Miao, and Qiang Qiu. Large convolutional model tuning via filter subspace. arXiv preprint arXiv:2403.00269, 2024.2, 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 228, + 295, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 228, + 295, + 271 + ], + "spans": [ + { + "bbox": [ + 62, + 228, + 295, + 271 + ], + "type": "text", + "content": "[5] Xinlei Chen, Hao Fang, Tsung-Yi Lin, Ramakrishna Vedantam, Saurabh Gupta, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco captions: Data collection and evaluation server. arXiv preprint arXiv:1504.00325, 2015. 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 273, + 295, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 273, + 295, + 304 + ], + "spans": [ + { + "bbox": [ + 62, + 273, + 295, + 304 + ], + "type": "text", + "content": "[6] Xiuyuan Cheng, Zichen Miao, and Qiang Qiu. Graph convolution with low-rank learnable local filters. arXiv preprint arXiv:2008.01818, 2020. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 307, + 295, + 340 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 307, + 295, + 340 + ], + "spans": [ + { + "bbox": [ + 62, + 307, + 295, + 340 + ], + "type": "text", + "content": "[7] Xiuyuan Cheng, Zichen Miao, and Qiang Qiu. Graph convolution with low-rank learnable local filters. In International Conference on Learning Representations, 2021. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 342, + 295, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 342, + 295, + 385 + ], + "spans": [ + { + "bbox": [ + 62, + 342, + 295, + 385 + ], + "type": "text", + "content": "[8] Jaemin Cho, Jie Lei, Hao Tan, and Mohit Bansal. Unifying vision-and-language tasks via text generation. In International Conference on Machine Learning, pages 1931-1942. PMLR, 2021. 2, 6, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 387, + 295, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 387, + 295, + 430 + ], + "spans": [ + { + "bbox": [ + 62, + 387, + 295, + 430 + ], + "type": "text", + "content": "[9] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 6, 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 432, + 295, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 432, + 295, + 475 + ], + "spans": [ + { + "bbox": [ + 57, + 432, + 295, + 475 + ], + "type": "text", + "content": "[10] Yihe Dong, Jean-Baptiste Cordonnier, and Andreas Loukas. Attention is not all you need: Pure attention loses rank doubly exponentially with depth. In International Conference on Machine Learning, pages 2793-2803. PMLR, 2021. 2, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 477, + 295, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 477, + 295, + 498 + ], + "spans": [ + { + "bbox": [ + 57, + 477, + 295, + 498 + ], + "type": "text", + "content": "[11] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. ICLR, 2021. 1, 2, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 500, + 295, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 500, + 295, + 533 + ], + "spans": [ + { + "bbox": [ + 57, + 500, + 295, + 533 + ], + "type": "text", + "content": "[12] Dan Friedman and Adji Bousso Dieng. The vendi score: A diversity evaluation metric for machine learning. arXiv preprint arXiv:2210.02410, 2022. 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 534, + 295, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 534, + 295, + 578 + ], + "spans": [ + { + "bbox": [ + 57, + 534, + 295, + 578 + ], + "type": "text", + "content": "[13] Timnit Gebru, Jonathan Krause, Yilun Wang, Duyun Chen, Jia Deng, and Li Fei-Fei. Fine-grained car detection for visual census estimation. In Proceedings of the AAAI Conference on Artificial Intelligence, 2017. 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 579, + 295, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 579, + 295, + 635 + ], + "spans": [ + { + "bbox": [ + 57, + 579, + 295, + 635 + ], + "type": "text", + "content": "[14] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Bartra, and Devi Parikh. Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6904-6913, 2017. 2, 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 635, + 295, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 635, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 57, + 635, + 295, + 689 + ], + "type": "text", + "content": "[15] Ligong Han, Yinxiao Li, Han Zhang, Peyman Milanfar, Dimitris Metaxas, and Feng Yang. Svdiff: Compact parameter space for diffusion fine-tuning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7323-7334, 2023. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "type": "text", + "content": "[16] Liu He and Daniel Aliaga. Globalmapper: Arbitrary-shaped urban layout generation. In Proceedings of the IEEE/CVF" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 317, + 74, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 335, + 74, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 74, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 74, + 553, + 95 + ], + "type": "text", + "content": "International Conference on Computer Vision, pages 454-464, 2023. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 96, + 553, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 96, + 553, + 129 + ], + "spans": [ + { + "bbox": [ + 317, + 96, + 553, + 129 + ], + "type": "text", + "content": "[17] Liu He and Daniel Aliaga. Coho: Context-sensitive city-scale hierarchical urban layout generation. In European Conference on Computer Vision, pages 1-18. Springer, 2024." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 129, + 553, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 129, + 553, + 172 + ], + "spans": [ + { + "bbox": [ + 317, + 129, + 553, + 172 + ], + "type": "text", + "content": "[18] Liu He, Yijuan Lu, John Corring, Dinei Florencio, and Cha Zhang. Diffusion-based document layout generation. In International Conference on Document Analysis and Recognition, pages 361-378. Springer, 2023." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 173, + 553, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 173, + 553, + 216 + ], + "spans": [ + { + "bbox": [ + 317, + 173, + 553, + 216 + ], + "type": "text", + "content": "[19] Liu He, Yizhi Song, Hejun Huang, Daniel Aliaga, and Xin Zhou. Kubrick: Multimodal agent collaborations for synthetic video generation. arXiv preprint arXiv:2408.10453, 2024. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 217, + 553, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 217, + 553, + 271 + ], + "spans": [ + { + "bbox": [ + 317, + 217, + 553, + 271 + ], + "type": "text", + "content": "[20] Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin De Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. Parameter-efficient transfer learning for nlp. In International conference on machine learning, pages 2790-2799. PMLR, 2019. 2, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 273, + 553, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 273, + 553, + 316 + ], + "spans": [ + { + "bbox": [ + 317, + 273, + 553, + 316 + ], + "type": "text", + "content": "[21] Edward J Hu, Phillip Wallis, Zeyuan Allen-Zhu, Yuzhhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. In International Conference on Learning Representations, 2021. 1, 2, 6, 7, 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 317, + 553, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 317, + 553, + 370 + ], + "spans": [ + { + "bbox": [ + 317, + 317, + 553, + 370 + ], + "type": "text", + "content": "[22] Drew A Hudson and Christopher D Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6700-6709, 2019. 2, 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 372, + 553, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 372, + 553, + 415 + ], + "spans": [ + { + "bbox": [ + 317, + 372, + 553, + 415 + ], + "type": "text", + "content": "[23] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In European Conference on Computer Vision, pages 709-727. Springer, 2022. 2, 6, 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 416, + 553, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 416, + 553, + 458 + ], + "spans": [ + { + "bbox": [ + 317, + 416, + 553, + 458 + ], + "type": "text", + "content": "[24] Rabeeh Karimi Mahabadi, James Henderson, and Sebastian Ruder. Compacter: Efficient low-rank hypercomplex adapter layers. Advances in Neural Information Processing Systems, 2021. 2, 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 460, + 553, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 460, + 553, + 503 + ], + "spans": [ + { + "bbox": [ + 317, + 460, + 553, + 503 + ], + "type": "text", + "content": "[25] Aditya Khosla, Nityananda Jayadevaprakash, Bangpeng Yao, and Fei-Fei Li. Novel dataset for fine-grained image categorization: Stanford dogs. In Proc. CVPR workshop on fine-grained visual categorization (FGVC), 2011. 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 504, + 553, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 504, + 553, + 536 + ], + "spans": [ + { + "bbox": [ + 317, + 504, + 553, + 536 + ], + "type": "text", + "content": "[26] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In International Conference on Learning Representations, 2015. 1" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 317, + 537, + 553, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 537, + 553, + 569 + ], + "spans": [ + { + "bbox": [ + 317, + 537, + 553, + 569 + ], + "type": "text", + "content": "[27] Thomas N Kipf and Max Welling. Semi-supervised classification with graph convolutional networks. arXiv preprint arXiv:1609.02907, 2016. 4" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 317, + 571, + 553, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 571, + 553, + 624 + ], + "spans": [ + { + "bbox": [ + 317, + 571, + 553, + 624 + ], + "type": "text", + "content": "[28] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 1" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 317, + 625, + 553, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 625, + 553, + 657 + ], + "spans": [ + { + "bbox": [ + 317, + 625, + 553, + 657 + ], + "type": "text", + "content": "[29] Dawid J Kopiczko, Tijmen Blankevoort, and Yuki M Asano. Vera: Vector-based random matrix adaptation. *ICLR*, 2024. 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 317, + 658, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 658, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 317, + 658, + 553, + 712 + ], + "type": "text", + "content": "[30] Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization of text-to-image diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 7, 1" + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "20154" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 115 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 115 + ], + "type": "text", + "content": "[31] M Lewis. Bart: Denoising sequence-to-sequence pretraining for natural language generation, translation, and comprehension. arXiv preprint arXiv:1910.13461, 2019. 2, 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 149 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 149 + ], + "type": "text", + "content": "[32] Junbo Li, Zichen Miao, Qiang Qiu, and Ruqi Zhang. Training bayesian neural networks with sparse subspace variational inference. arXiv preprint arXiv:2402.11025, 2024. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 151, + 294, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 151, + 294, + 204 + ], + "spans": [ + { + "bbox": [ + 56, + 151, + 294, + 204 + ], + "type": "text", + "content": "[33] Tianchun Li, Chengxiang Wu, Pengyi Shi, and Xiaoqian Wang. Cumulative difference learning vae for time-series with temporally correlated inflow-outflow. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 13619-13627, 2024. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 206, + 294, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 206, + 294, + 239 + ], + "spans": [ + { + "bbox": [ + 56, + 206, + 294, + 239 + ], + "type": "text", + "content": "[34] Xiang Lisa Li and Percy Liang. Prefix-tuning: Optimizing continuous prompts for generation. In Proceedings of the Association for Computational Linguistics, 2021. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 239, + 294, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 239, + 294, + 282 + ], + "spans": [ + { + "bbox": [ + 56, + 239, + 294, + 282 + ], + "type": "text", + "content": "[35] Dongze Lian, Daquan Zhou, Jiashi Feng, and Xinchao Wang. Scaling & shifting your features: A new baseline for efficient model tuning. Advances in Neural Information Processing Systems, 35:109-123, 2022. 2, 6, 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 284, + 294, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 284, + 294, + 337 + ], + "spans": [ + { + "bbox": [ + 56, + 284, + 294, + 337 + ], + "type": "text", + "content": "[36] Haokun Liu, Derek Tam, Mohammed Muqeeth, Jay Mohta, Tenghao Huang, Mohit Bansal, and Colin A Raffel. Few-shot parameter-efficient fine-tuning is better and cheaper than in-context learning. Advances in Neural Information Processing Systems, 35:1950-1965, 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 338, + 294, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 338, + 294, + 371 + ], + "spans": [ + { + "bbox": [ + 56, + 338, + 294, + 371 + ], + "type": "text", + "content": "[37] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 372, + 294, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 372, + 294, + 424 + ], + "spans": [ + { + "bbox": [ + 56, + 372, + 294, + 424 + ], + "type": "text", + "content": "[38] Shih-Yang Liu, Chien-Yi Wang, Hongxu Yin, Pavlo Molchanov, Yu-Chiang Frank Wang, Kwang-Ting Cheng, and Min-Hung Chen. Dora: Weight-decomposed low-rank adaptation. arXiv preprint arXiv:2402.09353, 2024. 2, 7, 8, 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 426, + 294, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 426, + 294, + 480 + ], + "spans": [ + { + "bbox": [ + 56, + 426, + 294, + 480 + ], + "type": "text", + "content": "[39] Tianci Liu, Ruirui Li, Yunzhe Qi, Hui Liu, Xianfeng Tang, Tianqi Zheng, Qingyu Yin, Monica Xiao Cheng, Jun Huan, Haoyu Wang, et al. Unlocking efficient, scalable, and continual knowledge editing with basis-level representation finetuning. arXiv preprint arXiv:2503.00306, 2025. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 482, + 294, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 482, + 294, + 514 + ], + "spans": [ + { + "bbox": [ + 56, + 482, + 294, + 514 + ], + "type": "text", + "content": "[40] Zichen Miao, Ze Wang, Wei Chen, and Qiang Qiu. Continual learning with filter atom swapping. In International Conference on Learning Representations, 2021. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 515, + 294, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 515, + 294, + 558 + ], + "spans": [ + { + "bbox": [ + 56, + 515, + 294, + 558 + ], + "type": "text", + "content": "[41] Zichen Miao, Ze Wang, Xiuyuan Cheng, and Qiang Qiu. Spatiotemporal joint filter decomposition in 3d convolutional neural networks. Advances in Neural Information Processing Systems, 34:3376-3388, 2021. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 559, + 294, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 559, + 294, + 624 + ], + "spans": [ + { + "bbox": [ + 56, + 559, + 294, + 624 + ], + "type": "text", + "content": "[42] Zichen Miao, Jiang Wang, Ze Wang, Zhengyuan Yang, Lijuan Wang, Qiang Qiu, and Zicheng Liu. Training diffusion models towards diverse image generation with reinforcement learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10844-10853, 2024. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 625, + 294, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 625, + 294, + 669 + ], + "spans": [ + { + "bbox": [ + 56, + 625, + 294, + 669 + ], + "type": "text", + "content": "[43] Zichen Miao, Zhengyuan Yang, Kevin Lin, Ze Wang, Zicheng Liu, Lijuan Wang, and Qiang Qiu. Tuning timestep-distilled diffusion model using pairwise sample optimization. arXiv preprint arXiv:2410.03190, 2024. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 670, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 294, + 713 + ], + "type": "text", + "content": "[44] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In 2008 Sixth Indian conference on computer vision, graphics & image processing, pages 722-729. IEEE, 2008. 6" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 715 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "type": "text", + "content": "[45] Lorenzo Noci, Sotiris Anagnostidis, Luca Biggio, Antonio Orvieto, Sidak Pal Singh, and Aurelien Lucchi. Signal propagation in transformers: Theoretical perspectives and the role of rank collapse. Advances in Neural Information Processing Systems, 35:27198-27211, 2022. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 129, + 553, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 553, + 183 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 553, + 183 + ], + "type": "text", + "content": "[46] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 184, + 553, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 184, + 553, + 216 + ], + "spans": [ + { + "bbox": [ + 316, + 184, + 553, + 216 + ], + "type": "text", + "content": "[47] Gaurav Patel and Qiang Qiu. Learning to unlearn while retaining: Combating gradient conflicts in machine unlearning. arXiv preprint arXiv:2503.06339, 2025. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 217, + 553, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 217, + 553, + 271 + ], + "spans": [ + { + "bbox": [ + 316, + 217, + 553, + 271 + ], + "type": "text", + "content": "[48] Gaurav Patel, Konda Reddy Mopuri, and Qiang Qiu. Learning to retain while acquiring: Combating distribution-shift in adversarial data-free knowledge distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7786-7794, 2023." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 272, + 553, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 272, + 553, + 325 + ], + "spans": [ + { + "bbox": [ + 316, + 272, + 553, + 325 + ], + "type": "text", + "content": "[49] Gaurav Patel, Christopher Sandino, Behrooz Mahasseni, Ellen L Zippi, Erdrin Azemi, Ali Moin, and Juri Minxha. Efficient source-free time-series adaptation via parameter subspace disentanglement. arXiv preprint arXiv:2410.02147, 2024. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 327, + 553, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 327, + 553, + 371 + ], + "spans": [ + { + "bbox": [ + 316, + 327, + 553, + 371 + ], + "type": "text", + "content": "[50] Jonas Pfeiffer, Aishwarya Kamath, Andreas Rücklé, Kyunghyun Cho, and Iryna Gurevych. Adapterfusion: Nondestructive task composition for transfer learning. arXiv preprint arXiv:2005.00247, 2020. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 372, + 553, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 372, + 553, + 415 + ], + "spans": [ + { + "bbox": [ + 316, + 372, + 553, + 415 + ], + "type": "text", + "content": "[51] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. *ICLR*, 2024. 1, 6, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 416, + 553, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 416, + 553, + 458 + ], + "spans": [ + { + "bbox": [ + 316, + 416, + 553, + 458 + ], + "type": "text", + "content": "[52] Qiang Qiu, Xiuyuan Cheng, Guillermo Sapiro, et al. Dcfnet: Deep neural network with decomposed convolutional filters. In International Conference on Machine Learning, pages 4198-4207. PMLR, 2018. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 460, + 553, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 460, + 553, + 514 + ], + "spans": [ + { + "bbox": [ + 316, + 460, + 553, + 514 + ], + "type": "text", + "content": "[53] Zeju Qiu, Weiyang Liu, Haiwen Feng, Yuxuan Xue, Yao Feng, Zhen Liu, Dan Zhang, Adrian Weller, and Bernhard Scholkopf. Controlling text-to-image diffusion by orthogonal finetuning. Advances in Neural Information Processing Systems, 36:79320-79362, 2023. 2, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 514, + 553, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 514, + 553, + 580 + ], + "spans": [ + { + "bbox": [ + 316, + 514, + 553, + 580 + ], + "type": "text", + "content": "[54] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, 2021. 2, 7, 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 582, + 553, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 582, + 553, + 624 + ], + "spans": [ + { + "bbox": [ + 316, + 582, + 553, + 624 + ], + "type": "text", + "content": "[55] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10684-10695, 2022. 1, 2, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 625, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 625, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 625, + 553, + 689 + ], + "type": "text", + "content": "[56] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 22500-22510, 2023. 2, 7, 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 691, + 553, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 715 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 715 + ], + "type": "text", + "content": "[57] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy," + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20155" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 76, + 72, + 296, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 296, + 116 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 296, + 116 + ], + "type": "text", + "content": "Aditya Khosla, Michael Bernstein, Alexander C. Berg, and Li Fei-Fei. ImageNet Large Scale Visual Recognition Challenge. International Journal of Computer Vision (IJCV), 2015. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 295, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 295, + 185 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 295, + 185 + ], + "type": "text", + "content": "[58] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278-25294, 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 186, + 295, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 186, + 295, + 229 + ], + "spans": [ + { + "bbox": [ + 56, + 186, + 295, + 229 + ], + "type": "text", + "content": "[59] Han Shi, Jiahui Gao, Hang Xu, Xiaodan Liang, Zhenguo Li, Lingpeng Kong, Stephen Lee, and James T Kwok. Revisiting over-smoothing in bert from the perspective of graph. arXiv preprint arXiv:2202.08625, 2022. 2, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 230, + 295, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 230, + 295, + 285 + ], + "spans": [ + { + "bbox": [ + 56, + 230, + 295, + 285 + ], + "type": "text", + "content": "[60] Kihyuk Sohn, Huiwen Chang, José Lezama, Luisa Polania, Han Zhang, Yuan Hao, Irfan Essa, and Lu Jiang. Visual prompt tuning for generative transfer learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 286, + 295, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 286, + 295, + 342 + ], + "spans": [ + { + "bbox": [ + 56, + 286, + 295, + 342 + ], + "type": "text", + "content": "[61] Yizhi Song, Liu He, Zhifei Zhang, Soo Ye Kim, He Zhang, Wei Xiong, Zhe Lin, Brian Price, Scott Cohen, Jianming Zhang, et al. Refine-by-align: Reference-guided artifacts refinement through semantic alignment. arXiv preprint arXiv:2412.00306, 2024. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 343, + 295, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 343, + 295, + 409 + ], + "spans": [ + { + "bbox": [ + 56, + 343, + 295, + 409 + ], + "type": "text", + "content": "[62] Yizhi Song, Zhifei Zhang, Zhe Lin, Scott Cohen, Brian Price, Jianming Zhang, Soo Ye Kim, He Zhang, Wei Xiong, and Daniel Aliaga. Imprint: Generative object compositing by learning identity-preserving representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8048-8058, 2024. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 410, + 295, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 410, + 295, + 453 + ], + "spans": [ + { + "bbox": [ + 56, + 410, + 295, + 453 + ], + "type": "text", + "content": "[63] Nitish Srivastava, Geoffrey Hinton, Alex Krizhevsky, Ilya Sutskever, and Ruslan Salakhutdinov. Dropout: a simple way to prevent neural networks from overfitting. The journal of machine learning research, 15(1):1929-1958, 2014. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 455, + 295, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 455, + 295, + 499 + ], + "spans": [ + { + "bbox": [ + 56, + 455, + 295, + 499 + ], + "type": "text", + "content": "[64] Alane Suhr, Stephanie Zhou, Ally Zhang, Iris Zhang, Hua-jun Bai, and Yoav Artzi. A corpus for reasoning about natural language grounded in photographs. arXiv preprint arXiv:1811.00491, 2018. 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 500, + 295, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 500, + 295, + 554 + ], + "spans": [ + { + "bbox": [ + 56, + 500, + 295, + 554 + ], + "type": "text", + "content": "[65] Yi-Lin Sung, Jaemin Cho, and Mohit Bansal. Vl-adapter: Parameter-efficient transfer learning for vision-and-language tasks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5227-5237, 2022. 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 556, + 295, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 556, + 295, + 610 + ], + "spans": [ + { + "bbox": [ + 56, + 556, + 295, + 610 + ], + "type": "text", + "content": "[66] Gemma Canet Tarres, Zhe Lin, Zhifei Zhang, Jianming Zhang, Yizhi Song, Dan Ruta, Andrew Gilbert, John Colomosse, and Soo Ye Kim. Thinking outside the bbox: Unconstrained generative object compositing. arXiv preprint arXiv:2409.04559, 2024. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 613, + 295, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 613, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 613, + 295, + 689 + ], + "type": "text", + "content": "[67] Grant Van Horn, Steve Branson, Ryan Farrell, Scott Haber, Jessie Barry, Panos Ipeirotis, Pietro Perona, and Serge Belongie. Building a bird recognition app and large scale dataset with citizen scientists: The fine print in fine-grained dataset collection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 595-604, 2015. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 690, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 690, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 690, + 295, + 713 + ], + "type": "text", + "content": "[68] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 1, 2" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 106 + ], + "type": "text", + "content": "[69] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The caltech-ucsd birds-200-2011 dataset. 2011. 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 107, + 553, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 161 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 161 + ], + "type": "text", + "content": "[70] Haoyu Wang, Tianci Liu, Ruirui Li, Monica Cheng, Tuo Zhao, and Jing Gao. Roselora: Row and column-wise sparse low-rank adaptation of pre-trained language model for knowledge editing and fine-tuning. arXiv preprint arXiv:2406.10777, 2024. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 163, + 553, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 553, + 207 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 553, + 207 + ], + "type": "text", + "content": "[71] Peihao Wang, Wenqing Zheng, Tianlong Chen, and Zhangyang Wang. Anti-oversmoothing in deep vision transformers via the fourier domain analysis: From theory to practice. arXiv preprint arXiv:2203.05962, 2022. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 209, + 553, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 209, + 553, + 241 + ], + "spans": [ + { + "bbox": [ + 316, + 209, + 553, + 241 + ], + "type": "text", + "content": "[72] Ze Wang, Xiuyuan Cheng, Guillermo Sapiro, and Qiang Qiu. Stochastic conditional generative networks with basis decomposition. arXiv preprint arXiv:1909.11286, 2019. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 243, + 553, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 243, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 316, + 243, + 553, + 285 + ], + "type": "text", + "content": "[73] Ze Wang, Seunghyun Hwang, Zichen Miao, and Qiang Qiu. Image generation using continuous filter atoms. Advances in Neural Information Processing Systems, 34:17826-17838, 2021." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 287, + 553, + 332 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 287, + 553, + 332 + ], + "spans": [ + { + "bbox": [ + 316, + 287, + 553, + 332 + ], + "type": "text", + "content": "[74] Ze Wang, Zichen Miao, Jun Hu, and Qiang Qiu. Adaptive convolutions with per-pixel dynamic filter atom. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12302-12311, 2021. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 333, + 553, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 333, + 553, + 376 + ], + "spans": [ + { + "bbox": [ + 316, + 333, + 553, + 376 + ], + "type": "text", + "content": "[75] Ze Wang, Zichen Miao, Xiantong Zhen, and Qiang Qiu. Learning to learn dense gaussian processes for few-shot learning. Advances in Neural Information Processing Systems, 34:13230-13241, 2021. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 377, + 553, + 444 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 377, + 553, + 444 + ], + "spans": [ + { + "bbox": [ + 316, + 377, + 553, + 444 + ], + "type": "text", + "content": "[76] Enze Xie, Lewei Yao, Han Shi, Zhili Liu, Daquan Zhou, Zhaoqiang Liu, Jiawei Li, and Zhenguo Li. Difffit: Unlocking transferability of large diffusion models via simple parameter-efficient fine-tuning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4230-4239, 2023. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 445, + 553, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 445, + 553, + 488 + ], + "spans": [ + { + "bbox": [ + 316, + 445, + 553, + 488 + ], + "type": "text", + "content": "[77] Zhexiao Xiong, Wei Xiong, Jing Shi, He Zhang, Yizhi Song, and Nathan Jacobs. Groundingbooth: Grounding text-to-image customization. arXiv preprint arXiv:2409.08520, 2024.1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 490, + 553, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 490, + 553, + 545 + ], + "spans": [ + { + "bbox": [ + 316, + 490, + 553, + 545 + ], + "type": "text", + "content": "[78] SHIH-YING YEH, Yu-Guan Hsieh, Zhidong Gao, Bernard BW Yang, Gigeong Oh, and Yanmin Gong. Navigating text-to-image customization: From lycoris fine-tuning to model evaluation. In International Conference on Learning Representations, 2023. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 546, + 553, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 546, + 553, + 600 + ], + "spans": [ + { + "bbox": [ + 316, + 546, + 553, + 600 + ], + "type": "text", + "content": "[79] Elad Ben Zaken, Yoav Goldberg, and Shauli Ravfogel. Bitfit: Simple parameter-efficient fine-tuning for transformer-based masked language-models. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics, 2022. 2, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 601, + 553, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 601, + 553, + 667 + ], + "spans": [ + { + "bbox": [ + 316, + 601, + 553, + 667 + ], + "type": "text", + "content": "[80] Shuangfei Zhai, Tatiana Likhomanenko, Etai Littwin, Dan Busbridge, Jason Ramapuram, Yizhe Zhang, Jiatao Gu, and Joshua M Susskind. Stabilizing transformer training by preventing attention entropy collapse. In International Conference on Machine Learning, pages 40770-40803. PMLR, 2023. 2, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 669, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 669, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 669, + 553, + 713 + ], + "type": "text", + "content": "[81] Xiaohua Zhai, Joan Puigcerver, Alexander Kolesnikov, Pierre Ruyssen, Carlos Riquelme, Mario Lucic, Josip Djolonga, Andre Susano Pinto, Maxim Neumann, Alexey Dosovitskiy, et al. A large-scale study of representation learning" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "20156" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 229 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 75, + 72, + 296, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 72, + 296, + 95 + ], + "spans": [ + { + "bbox": [ + 75, + 72, + 296, + 95 + ], + "type": "text", + "content": "with the visual task adaptation benchmark. arXiv preprint arXiv:1910.04867, 2019. 6, 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 296, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 296, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 296, + 140 + ], + "type": "text", + "content": "[82] Qingru Zhang, Minshuo Chen, Alexander Bukharin, Nikos Karampatziakis, Pengcheng He, Yu Cheng, Weizhu Chen, and Tuo Zhao. Adalora: Adaptive budget allocation for parameter-efficient fine-tuning. *ICLR*, 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 141, + 294, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 141, + 294, + 185 + ], + "spans": [ + { + "bbox": [ + 56, + 141, + 294, + 185 + ], + "type": "text", + "content": "[83] Yibo Zhong, Haoxiang Jiang, Lincan Li, Ryumei Nakada, Tianci Liu, Linjun Zhang, Huaxiu Yao, and Haoyu Wang. Neat: Nonlinear parameter-efficient adaptation of pretrained models. arXiv preprint arXiv:2410.01870, 2024. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 186, + 296, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 186, + 296, + 229 + ], + "spans": [ + { + "bbox": [ + 56, + 186, + 296, + 229 + ], + "type": "text", + "content": "[84] Daquan Zhou, Bingyi Kang, Xiaojie Jin, Linjie Yang, Xiaochen Lian, Zihang Jiang, Qibin Hou, and Jiashi Feng. Deepvit: Towards deeper vision transformer. arXiv preprint arXiv:2103.11886, 2021. 2, 3" + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20157" + } + ] + } + ], + "index": 5 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Coherent 3D Portrait Video Reconstruction via Triplane Fusion/2088b6b2-40e1-4341-a0c5-c5303c3acd65_content_list.json b/2025/Coherent 3D Portrait Video Reconstruction via Triplane Fusion/2088b6b2-40e1-4341-a0c5-c5303c3acd65_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..41d6fbd234d7887ca27470da5f80ed21a05bf904 --- /dev/null +++ b/2025/Coherent 3D Portrait Video Reconstruction via Triplane Fusion/2088b6b2-40e1-4341-a0c5-c5303c3acd65_content_list.json @@ -0,0 +1,1585 @@ +[ + { + "type": "text", + "text": "Coherent 3D Portrait Video Reconstruction via Triplane Fusion", + "text_level": 1, + "bbox": [ + 178, + 128, + 818, + 151 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Shengze Wang $^{1*}$ Xueting Li $^{2}$ Chao Liu $^{2}$ Matthew Chan $^{2}$ Michael Stengel $^{2}$ Henry Fuchs $^{1}$ Shalini De Mello $^{2\\dagger}$ Koki Nagano $^{2\\dagger}$ $^{1}$ UNC Chapel Hill ${}^{2}$ NVIDIA $\\dagger$ Equal Contribution \nhttps://research.nvidia.com/labs/amri/projects/coherent3d/", + "bbox": [ + 116, + 178, + 879, + 251 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 284, + 326, + 300 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent breakthroughs in single-image 3D portrait reconstruction have enabled telepresence systems to stream 3D portrait videos from a single camera in real-time, democratizing telepresence. However, per-frame 3D reconstruction exhibits temporal inconsistency and forgets the user's appearance. On the other hand, self-reenactment methods can render coherent 3D portraits by driving a 3D avatar built from a single reference image but fail to faithfully preserve the user's per-frame appearance (e.g., instantaneous facial expressions and lighting). As a result, neither of these two frameworks is an ideal solution for democratized 3D telepresence. In this work, we address this dilemma and propose a novel solution that maintains both coherent identity and dynamic per-frame appearance to enable the best possible realism. To this end, we propose a new fusion-based method that takes the best of both worlds by fusing a canonical 3D prior from a reference view with dynamic appearance from per-frame input views, producing temporally stable 3D videos with faithful reconstruction of the user's per-frame appearance. Trained only using synthetic data produced by an expression-conditioned 3D GAN, our encoder-based method achieves both state-of-the-art 3D reconstruction and temporal consistency on in-studio and in-the-wild datasets.", + "bbox": [ + 88, + 316, + 485, + 667 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 680, + 220, + 696 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Telepresence aims at bringing distant people face-to-face and stands out as a particularly compelling application of computer vision and graphics. Over the last decades, various successful telepresence systems [19, 21, 21, 25, 31, 33, 37, 43] have been developed. However, most employ bulky multi-view 3D scanners or depth sensors to ensure high-quality volumetric per-frame reconstruction. Unlike these classical 3D/4D reconstruction methods, recent AI-based feed-forward 3D lifting techniques [2, 52] can lift a single RGB image from an off-the-shelf webcam into a neural radiance field (NeRF) representation encoded into a set of triplanes in real time, paving the way towards making 3D", + "bbox": [ + 89, + 699, + 483, + 875 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "telepresence accessible to anyone [48].", + "bbox": [ + 513, + 285, + 772, + 301 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/af1c171076518b0c923ef4c3c10de2463aa514630b3845c04053af66948df8fa.jpg", + "image_caption": [ + "Figure 1. Comparisons. Given a single reference image and a single-view video frame, our method reconstructs the authentic dynamic appearance of the user (e.g., facial expressions and lighting) while producing a temporally coherent 3D video. A previous single-view 3D lifting method (LP3D) that reconstructs the avatar from the video frame on a per-frame basis suffers from distortions and temporal inconsistency. A portrait reenactment method (GPAvatar) drives the identity in the reference image using the video frame, but fails to capture accurate facial expressions (e.g., smile) and per-frame appearance (e.g., lighting). The output should be compared to the appearance of the input video frame (green box)." + ], + "image_footnote": [], + "bbox": [ + 516, + 313, + 906, + 595 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Currently, there are two major paradigms in democratized 3D telepresence solutions from a single-view video: (1) single-view per-frame 3D lifting methods and (2) 3D portrait reenactment, which drives an identity from a reference image using another driving frame, but none of them is an ideal solution. For (1), single-frame-based lifting techniques such as LP3D [52], have the advantage of faithfully preserving the instantaneous dynamic conditions present in an input video, e.g. lighting, expressions, and posture, all", + "bbox": [ + 511, + 768, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Shengze Wang was an intern at NVIDIA during the project", + "bbox": [ + 107, + 886, + 426, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "10712", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "of which are crucial to an authentic telepresence experience. However, single-image reconstruction methods operate independently on each frame and thus have fundamental limitations in maintaining temporal consistency. This difficulty stems from the inherent ill-posed nature of single-image-based reconstruction. In order to render novel views that are significantly far from the input view, the system cannot rely on information present in the input view and hence must hallucinate generate plausible content, which cannot be guaranteed to be consistent across multiple temporal frames. For example, LP3D's reconstruction changes significantly depending on the user's head pose in the input frame (see second row in Fig. 1 and third column in Fig. 2). In comparison, 3D self-reenactment methods create an avatar model from one or multiple reference images and use a separate driving video to drive the facial expressions and poses of the avatar. [6, 27, 51, 59]. While these approaches allow for temporally consistent results, they often cannot faithfully reconstruct the input video's dynamic conditions such as changes in lighting. Moreover, reenactment methods often struggle to generate accurate expressions because their expression detection and control are not precise enough (see the first row in Fig. 1).", + "bbox": [ + 89, + 90, + 483, + 426 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we address this dilemma of democratized 3D telepresence approaches and propose a novel solution to the problem of simultaneously maintaining temporal stability while preserving real-time dynamics of input videos for single-camera telepresence applications. Our proposed solution is a fusion-based approach that leverages the stability and accuracy of a canonical 3D prior, and also captures the diverse deviations from the prior in newer video frames (see the third row in Fig. 1).", + "bbox": [ + 89, + 428, + 483, + 559 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our model first uses LP3D [52] to construct a canonical triplane prior from a (near) frontal image of the user, which can be casually captured or extracted from a video. During video reconstruction, our model lifts each input frame into a raw triplane, which is then fused with the canonical triplane (see Fig. 3). For images with oblique yaw head poses, artifacts and identity-related distortions may be present in its lifted triplane (see \"LP3D\" in Fig. 1 and 2). Hence we propose an undistorter module, which learns to undistort the raw instantaneous triplane to more closely match the identity of the correctly structured canonical triplane. We then propose a fuser module that combines the undistorted triplane and the canonical triplane in order to preserve identity consistency and recover occluded regions while reconstructing the dynamic lighting, nuanced expressions, and posture information in the input frame.", + "bbox": [ + 89, + 560, + 483, + 792 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We summarize our contributions as follows:", + "bbox": [ + 109, + 794, + 401, + 808 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- We contribute a novel triplane fusion method that combines the dynamic information from per-frame triplanes with a canonical triplane extracted from a reference image. Trained only using a synthetic multi-view video", + "bbox": [ + 89, + 809, + 482, + 868 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "dataset, our feed-forward approach generates 3D portrait videos that demonstrate both temporal consistency and faithful reconstruction of the dynamic appearance of the user (e.g. lighting and expression), whereas prior solutions can only achieve one of the two properties.", + "bbox": [ + 524, + 90, + 906, + 165 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a new framework to evaluate single-view 3D portrait reconstruction methods using multi-view data and gain insight into the method's reconstruction quality and robustness.", + "- We present evaluations on both in-studio and in-the-wild data and demonstrate that our method achieves state-of-the-art performance in terms of temporal consistency and reconstruction accuracy." + ], + "bbox": [ + 513, + 165, + 906, + 281 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/10853aafdfdf15407b71a6bf588e4b8fab5223564dfef49a2ada7f3d83b9a07b.jpg", + "image_caption": [ + "Figure 2. View-Dependent Distortion. \"NV\" refers to \"Novel View Rendering\". Top: inputs to our model and LP3D. Second & Third Rows: LP3D's reconstruction varies greatly under challenging viewpoints, showing a predictable pattern of artifacts including abnormally strong activations on the side being captured (red circle), as well as geometric distortion along the view direction of the camera due to depth ambiguity. We refer to this phenomenon as \"View-Dependent Distortion\". Fourth row: Our method removes such artifacts and achieves better coherence." + ], + "image_footnote": [], + "bbox": [ + 519, + 297, + 900, + 523 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 667, + 653, + 684 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2D portrait reenactment. Given a single or a few reference portrait images and a driving video, recent talking-head generators can reenact 2D portraits by transferring the facial expressions and poses from the driving video onto facial portraits [9, 10, 17, 46, 53, 54, 60, 62-64, 66]. However, being 2D, they cannot be rendered from novel viewpoints, which is crucial for 3D telepresence.", + "bbox": [ + 511, + 694, + 905, + 796 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3D-aware portrait generation and reenactment. Some recent works use deformable volumetric implicit radiance fields [34, 38, 39] or Gaussian splatting [22] combined by 3DMMs to reconstruct a photorealistic and animatable volumetric head avatar [1, 4, 14, 41, 45, 57, 67]. However, they require extensive data captures from videos or multiview cameras and person-specific training. Others use", + "bbox": [ + 511, + 797, + 906, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "1We find that the 3D lifting from a near frontal reference view is reliable, hence use this as a canonical 3D prior. See Fig. 2 first column.", + "bbox": [ + 89, + 875, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "10713", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "large-scale video datasets and learn a disentangled triplane 3D [5] for 3D facial reenactment in a feedforward fashion [6, 27, 28, 32, 51, 59, 61]. They construct a canonical 3D head from a reference image (often a neutral frame), and use facial expressions and head poses extracted from a separate driving video to animate it. As such, fine-grained facial expressions may not be captured due to errors in disentanglement. Most importantly, these reenactment methods fail to preserve the dynamic appearance of users (e.g., person-specific wrinkles or lighting) across time.", + "bbox": [ + 89, + 90, + 480, + 237 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3D GAN inversion. By combining GANs [15] and neural volume rendering [34], recent breakthroughs in 3D-aware GANs [5, 8, 16, 36, 44, 47, 55, 56, 58, 65, 68] can learn to generate photorealistic 3D heads from in-the-wild 2D images. Notably, EG3D proposes an efficient and compact triplane representation [5] to generate 3D heads, and Next3D [50] extends EG3D to create 3D portrait videos controlled by 3DMM facial expression and pose parameters. Besides generating 3D heads, these models can also be used for single-view 3D reconstruction using GAN inversion [13, 24, 30, 49], manipulation of the 3D head [18, 49, 69], and 3D personalization [3, 40]. However, inverting a few seconds of video can take many minutes or hours, and the inversion quality is often unsatisfactory due to inaccurate identity and expression preservation, as evaluated by GPAvatar [6]. Therefore, recent works [2, 52] propose encoder-based solutions to lift each video frame into a triplane. However, when each frame is independently lifted into 3D, the resulting 3D head video suffers in terms of temporal consistency - a key requirement of 3D telepresence systems. To enhance single-frame 2D-to-3D encoders for human heads, e.g., LP3D [52], we propose a triplane-fusion-based method, which improves their temporal stability while preserving temporal dynamics across time.", + "bbox": [ + 91, + 238, + 486, + 590 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 603, + 181, + 618 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Background: 3D Portrait from a Single Image", + "text_level": 1, + "bbox": [ + 89, + 627, + 480, + 643 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "LP3D [52] performs photorealistic 3D portrait reconstruction by using a feedforward encoder to lift an RGB image into a triplane $\\mathbf{T} \\in \\mathbb{R}^{3 \\times 32 \\times 256 \\times 256}$ , which can be volume rendered to an RGB image from any viewpoint. LP3D can run in real-time and has been developed into a complete realtime telepresence system [48]. We lift 2D faces into triplanes with a slightly modified LP3D trained on larger face crops containing shoulders and a camera estimator to recover the input image's camera parameters $M \\in \\mathbb{R}^{25}$ . It performs better than the original version (Table 3).", + "bbox": [ + 89, + 648, + 482, + 796 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Definitions", + "text_level": 1, + "bbox": [ + 89, + 805, + 210, + 820 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We call a current video frame an \"input frame\" into our system. Additionally, we encode a near-frontal \"reference image\" of the subject into a \"canonical triplane\", which is used to stabilize the 3D video generation process. Lastly, we refer to the viewpoint of the camera in the input frame", + "bbox": [ + 89, + 825, + 482, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "as \"input viewpoint\".", + "bbox": [ + 513, + 90, + 663, + 106 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.3. Overview", + "text_level": 1, + "bbox": [ + 513, + 114, + 624, + 130 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "An overview of our method is illustrated in Fig. 3. To improve temporal consistency and reconstruct occluded parts of the face, we leverage an additional near-frontal reference image, which can be a selfie image or a video frame automatically obtained from the same video using an off-the-shelf head pose estimator like DECA [11]. We lift this reference image into the canonical triplane using pre-trained LP3D. When processing the video, we first lift each input frame into a raw triplane. Then, the Triplane Undistorter (Sec. 3.5) removes view-dependent distortions and recovers the identity using the canonical triplane as a reference. This produces an undistorted triplane. Finally, to recover regions that are occluded in the input frame and further improve stability, our Triplane Fuser (Sec. 3.6) combines the undistorted triplane with the canonical triplane to generate the final coherent triplane.", + "bbox": [ + 511, + 136, + 906, + 372 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.4. Generating Synthetic Dynamic Multiview Data", + "text_level": 1, + "bbox": [ + 511, + 380, + 906, + 396 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We synthesize ground truth multiview training images using Next3D [50]. We first extract 2D landmarks and FLAME [26] coefficient labels for the FFHQ [20] dataset using DECA [12]. Then, during training, we sample a pair of random FLAME coefficients from our FFHQ labels and input them to Next3D alongside a single random identity code $z$ . We notice that FFHQ contains mostly less extreme expressions. Thus, we randomly multiply the expression codes by a small scaling factor to exaggerate the expressions during training. Notice that the Next3D generator is much more expressive than FLAME due to its use of triplanes and GAN training even though it is conditioned on FLAME. Next3D then generates a pair of triplanes for $t = 0$ and $t = 1$ of the same person with 2 expressions. We render the $t = 0$ triplane into a near-frontal reference image $I_{ref}$ (Fig. 3 orange box) and the $t = 1$ triplane into 3 images: the input frame rendered from a random viewpoint (green box), the ground truth image at a different viewpoint (red box), and a frontal-view image (red box). Only the latter two are used for supervision. Additionally, to simulate different lighting conditions, we also apply different color augmentations (brightness, contrast, saturation, and hue) to images at $t = 0$ and $t = 1$ .", + "bbox": [ + 511, + 402, + 908, + 738 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Shoulder augmentation. It is important that 3D portraits capture dynamic shoulder movements to convey body language and achieve eye contact in telepresence. However, Next3D does not provide control over shoulder posture. Thus, we simulate shoulder movement in the rendered images without modifying the Next3D triplanes by warping camera rays during volume rendering. Please see the supplement for more detail.", + "bbox": [ + 511, + 739, + 906, + 854 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Pseudo-groundtruth triplanes. Due to our shoulder augmentation, the Next3D triplanes and their 2D renderings become inconsistent. Thus, Next3D's triplanes cannot be used", + "bbox": [ + 511, + 856, + 906, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "10714", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/d9070d5ebef1551b6b6a61f0bebca5b7fdd80545b2baa92dc0c5b294d45db6ea.jpg", + "image_caption": [ + "Figure 3. Method Overview. Given a near-frontal reference image and an input frame, we first reconstruct a canonical triplane and a raw triplane using LP3D [52] (Sec. 3.1). Next, we combine them through a Triplane Fusion module (blue box) that ensures temporal consistency while preserving dynamic information (e.g., lighting and shoulder pose) (Sec. 3.5 and 3.6). Our model is trained with only synthetic video data generated by a 3D GAN [50], with carefully designed augmentations to preserve shoulder motion and lighting (Sec. 3.4)." + ], + "image_footnote": [], + "bbox": [ + 106, + 95, + 883, + 425 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "as direct supervisory signals. To mitigate this, we leverage the fact that LP3D generates reasonably accurate triplanes from frontal view images. We use a frozen LP3D to predict pseudo-groundtruth triplanes $T_{\\text{triplaneGT}}$ from the frontal novel view $I_{\\text{frontal}}$ for $t = 1$ (Fig. 3 bottom right).", + "bbox": [ + 88, + 496, + 486, + 571 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.5. Removing Distortion and Preserving Identity", + "text_level": 1, + "bbox": [ + 89, + 582, + 472, + 599 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "LP3D's reconstruction quality strongly depends on the input frame's viewpoint. When subjects are captured from the side, LP3D tends to produce errors in identity and exhibits stretching distortions (e.g., in the \"Input Frame 1\" and \"Input Frame 2\" columns in Fig. 2), primarily due to the inherent ambiguity of single-image reconstruction. In contrast, near-frontal views offer more complete identity cues and suffer less from occlusion (e.g. the \"Reference Image\" column in Fig. 2), leading to more reliable reconstruction. To reduce the single-image ambiguity, we adopt a near-frontal reconstruction as the canonical triplane $T_{\\text{cano}}$ . Although incorporating more reference views could potentially improve accuracy, we find that using a single near-frontal view not only simplifies the model but also sufficiently enhances temporal coherence.", + "bbox": [ + 88, + 604, + 482, + 824 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To correct the distortion in an input raw triplane $T_{raw}$ using the canonical triplane $T_{cano}$ as reference, we devise a Triplane Undistorter $U$ (Fig. 3):", + "bbox": [ + 89, + 825, + 482, + 871 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nU \\left(T _ {\\text {r a w}}, T _ {\\text {c a n o}}\\right) = T _ {\\text {u n d i s t}} \\in \\mathbb {R} ^ {3 \\times 3 2 \\times 2 5 6 \\times 2 5 6}. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 130, + 883, + 482, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Since the distortion in each triplane is a 2D warping artifact, it can be reversed through a 2D undistortion warping of each plane. To this end, our Undistorter $U$ adopts the SPyNet[42] architecture to predict a 2D correction warping $T_{corr} \\in \\mathbb{R}^{3 \\times 2 \\times 256 \\times 256}$ for the three planes. Then, the Undistorter corrects the raw triplane $T_{raw}$ by warping it based on the predicted 2D offsets $(\\Delta u, \\Delta v)$ in $T_{corr}$ as:", + "bbox": [ + 511, + 496, + 906, + 599 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nT _ {c o r r} = S P y N e t \\left(T _ {\\text {r a w}}, T _ {\\text {c a n o}}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 598, + 604, + 906, + 619 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nT _ {u n d i s t} = \\operatorname {W a r p} \\left(T _ {\\text {r a w}}, T _ {\\text {c o r r}}\\right). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 602, + 625, + 903, + 641 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "It is important to note the difference between optical flow and undistortion. An ideal optical flow model would warp $T_{\\text{raw}}$ towards $T_{\\text{cano}}$ in order to make the two identical. However, this alignment leads to significant artifacts and destroys the reconstruction (Table. 3 and Fig. 5). This is because the purpose of undistortion is not to make the two triplanes identical, but to correct $T_{\\text{raw}}$ such that it has the same identity as the canonical triplane $T_{\\text{cano}}$ , while preserving dynamic information such as expressions and lighting. Therefore, instead of using the canonical triplane $T_{\\text{cano}}$ as the warping target, our Undistorter $U$ merely uses $T_{\\text{cano}}$ as the identity conditioning to correct the raw triplane $T_{\\text{raw}}$ towards the ground truth triplane $T_{\\text{triplaneGT}}$ (Fig. 3, lower-right), which is not available during test time. The correction warping is supervised by the consistency between the undistorted triplane $T_{\\text{undist}}$ and pseudo-groundtruth tri", + "bbox": [ + 511, + 646, + 908, + 882 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "10715", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "plane $T_{\\text{triplane}GT}$ (Sec. 3.4) via a triplane loss:", + "bbox": [ + 89, + 90, + 403, + 107 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {u n d i s t} = L _ {1} \\left(T _ {u n d i s t}, T _ {t r i p l a n e G T}\\right). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 163, + 116, + 480, + 133 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.6. Reconstructing Occluded Regions through Triplane Fusion", + "text_level": 1, + "bbox": [ + 89, + 138, + 482, + 169 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As the user moves around in the video, different parts of their head become occluded. To recover occluded areas in the input frame and further stabilize the subject's identity across the video, our Fuser $F$ enhances the reconstruction by utilizing the canonical triplane $T_{\\text{cano}}$ , in which the currently occluded areas are often visible. Therefore, it is important for the Fuser $F$ to identify and recover the occluded regions while preserving information from the visible regions in $T_{\\text{undist}}$ . To accomplish this, we thus use a 5-layer ConvNet-based visibility estimator $V$ to predict a visibility triplane as $V(T_{\\text{raw}}) = T_{\\text{raw}}^{\\text{vis}} \\in \\mathbb{R}^{3 \\times 1 \\times 128 \\times 128}$ , i.e., one visibility map for each plane. $T_{\\text{raw}}^{\\text{vis}}$ is undistorted alongside $T_{\\text{undist}}$ to produce $T_{\\text{undist}}^{\\text{vis}}$ . We also predict a visibility triplane for the canonical triplane as $T_{\\text{cano}}^{\\text{vis}} = V(T_{\\text{cano}})$ . Finally, the Fuser $F$ produces the fused triplane $T_{\\text{fused}}$ by combining information from the undistorted input tranplane $T_{\\text{undist}}$ , its visibility triplane $T_{\\text{undist}}^{\\text{vis}}$ , the canonical triplane $T_{\\text{cano}}$ , and its visibility triplane $T_{\\text{cano}}^{\\text{vis}}$ .", + "bbox": [ + 89, + 174, + 483, + 441 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nT _ {f u s e d} = F \\left(T _ {u n d i s t}, T _ {u n d i s t} ^ {\\text {v i s}}, T _ {c a n o}, T _ {c a n o} ^ {\\text {v i s}}\\right). \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 137, + 449, + 482, + 467 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this way, Fuser $F$ merges new information in the input frame with occluded regions visible in the reference image. To train the visibility predictor $V$ , we calculate the visibility loss $L_{vis}$ as the $L_{1}$ distance between the predicted visibility triplanes versus the ground truth:", + "bbox": [ + 89, + 476, + 483, + 551 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {v i s} = L _ {1} (T _ {r a w} ^ {v i s}, T _ {r a w} ^ {v i s G T}) + L _ {1} (T _ {c a n o} ^ {v i s}, T _ {c a n o} ^ {v i s G T}). \\quad (6)\n$$\n", + "text_format": "latex", + "bbox": [ + 112, + 558, + 482, + 578 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The ground truth visibility triplanes contain 1 for pixels that are visible, and 0 otherwise. The generation of the ground truth visibility triplanes $T_{raw}^{visGT}$ and $T_{cano}^{visGT}$ is discussed in the supplementary.", + "bbox": [ + 89, + 585, + 483, + 645 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To supervise the Fuser $F$ , we calculate the fusion loss $L_{fusion}$ as the $L_{1}$ loss between the fused triplane $T_{fused}$ and the pseudo-groundtruth triplane $T_{frontalGT}$ . To highlight the currently occluded region during training, we also upweight the occluded region using an occlusion mask $T_{occMask} \\in \\mathbb{R}^{3 \\times 1 \\times 256 \\times 256}$ as:", + "bbox": [ + 89, + 645, + 483, + 734 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nT _ {d i f f} = \\left| T _ {f u s e d} - T _ {t r i p l a n e G T} \\right|, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 228, + 743, + 480, + 760 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {f u s i o n} = \\operatorname {M e a n} \\left(T _ {d i f f}\\right) + \\frac {T _ {d i f f} \\cdot T _ {v i s G T}}{\\left| T _ {v i s G T} \\right|} + \\frac {T _ {d i f f} \\cdot T _ {o c c M a s k}}{\\left| T _ {o c c M a s k} \\right|}. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 102, + 762, + 482, + 789 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Please refer to the supplementary for the calculation of the occlusion mask. We use the Recurrent Video Restoration Transformer (RVRT) [29] as the backbone of our Fuser $F$ because of its memory efficiency. We find that the final summation skip connection in RVRT prevents effective learning. This is because the original RVRT was designed to correct local blurriness and noise in a corrupted RGB video,", + "bbox": [ + 89, + 797, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "whereas our triplane videos exhibit structural distortion on a much larger scale and the summation skip connection thus limits the model's ability to correct the general structure. We thus replace the summation with a small 5-layer ConvNet.", + "bbox": [ + 511, + 90, + 903, + 162 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Lastly, note that both the Undistorter $U$ and the Fuser $F$ consist of 3 separate copies, one for each of the 3 planes, because we find that processing all three planes jointly leads to collapse to 2D (please see supplementary for visualization and analysis).", + "bbox": [ + 511, + 164, + 905, + 238 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.7. Training Losses", + "text_level": 1, + "bbox": [ + 511, + 244, + 671, + 260 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our loss function is the summation of four loss terms that provide two types of supervision: (a) direct triplane space guidance used to supervise the undistortion process in the Undistorter $U$ , the visibility prediction process, and the fusion process in the Fuser $F$ ; and (b) image space guidance for overall learning of high-quality image synthesis:", + "bbox": [ + 511, + 266, + 905, + 354 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} L = w _ {u n d i s t} L _ {u n d i s t} + w _ {f u s i o n} L _ {f u s i o n} \\\\ + w _ {v i s} L _ {v i s} + w _ {r e n d e r} L _ {r e n d e r}, \\tag {9} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 354, + 903, + 388 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $w_{undist}$ , $w_{viz}$ , $w_{fusion}$ , and $w_{render}$ are scalar weights for the different loss terms. $L_{render}$ is calculated as the perceptual loss $L_{LPIPS}$ between the ground truth novel view $I_{GT}$ and the rendered novel view $I_{render}$ as:", + "bbox": [ + 511, + 393, + 905, + 453 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\text {r e n d e r}} = L _ {\\text {L P I P S}} \\left(I _ {G T}, I _ {\\text {r e n d e r}}\\right). \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 591, + 458, + 903, + 474 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Evaluation", + "text_level": 1, + "bbox": [ + 511, + 479, + 630, + 496 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Multi-view evaluation of single-view reconstruction. Due to the lack of 3D ground truth for in-the-wild images, prior methods are often evaluated on the input image reconstruction task, which compares the rendered image against the input image using metrics like PSNR, whereas the novel view synthesis task often relies on visual assessments. However, this practice can lead to ambiguities and inaccurate conclusions. For example, if a 3D reconstruction is only rendered and evaluated from the input viewpoint, then that reconstruction can overfit to the input view to achieve high scores even if it is inaccurate when rendered from other viewpoints (as proven quantitatively in Fig. 11 of the supplementary). Moreover, single-view reconstruction methods can be heavily affected by the choice of input viewpoints (Fig. 2). Therefore, both the input viewpoint and the novel viewpoint are crucial variables to consider when concluding the performance of a method. We thus propose a new evaluation framework that evaluates a model across every input-novel viewpoint combination. In this way, a method can only achieve high numerical performance when it consistently generates high-quality reconstructions regardless of the choice of input or novel viewpoint.", + "bbox": [ + 511, + 505, + 905, + 843 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Evaluation Dataset", + "text_level": 1, + "bbox": [ + 511, + 849, + 696, + 864 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We quantitatively evaluate various methods on the NeRSemble [23] dataset, which is a multiview portrait", + "bbox": [ + 511, + 871, + 905, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "10716", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "video dataset that allows us to evaluate the methods using different input-novel viewpoint combinations. It is recorded with 16 calibrated time-synchronized cameras in a controlled studio environment. The images are captured at $7.1\\mathrm{MP}$ resolution and 73 frames per second. There are 10 sequences in the test set, each capturing a different individual performing different expressions. One of the 10 test sequences involves severe facial occlusion from hair that causes most of the methods' face trackers to fail for significant portions of the recording for many of the viewpoints. We thus leave out that sequence because the results would not be a reliable assessment of quality. We also use 8 roughly evenly separated cameras out of all 16 cameras during evaluation.", + "bbox": [ + 89, + 90, + 480, + 296 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Metrics", + "text_level": 1, + "bbox": [ + 89, + 305, + 187, + 319 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Synthesis Quality. Given $N$ views in the dataset, we evaluate a method's average performance across every input-novel viewpoint combination. More specifically, for each frame, each of the $N$ cameras is used as the input viewpoint, producing $N$ reconstructions in total. Then, each of the $N$ reconstructions is rendered and evaluated on the $N$ viewpoints, resulting in an $N \\times N$ score matrix. As mentioned before, we use $N = 8$ of the camera views from the NeRSemnble[23] dataset as input and novel views. Thus, for a test sequence of $T$ frames, we can calculate a spatial-temporal score matrices $\\mathbf{S}^{T \\times 8 \\times 8}$ for each of the metrics (as visualized in Fig. 11 in the supplementary):", + "bbox": [ + 89, + 325, + 482, + 503 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {S} _ {t, i, j} = \\operatorname {M e t r i c} \\left(\\mathbf {I} _ {\\text {r e n d e r}} ^ {t, i, j}, \\mathbf {I} _ {G T} ^ {t, j}\\right), 1 \\leq i, j \\leq N, 1 \\leq t \\leq T, \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 120, + 506, + 480, + 523 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\ns = \\operatorname {M e a n} \\left(\\left\\{\\mathbf {S} _ {t, i, j} \\right\\}\\right), \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 282, + 525, + 480, + 541 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $Metric(\\cdot)$ can be LPIPS, PSNR, etc.. $\\mathbf{I}_{\\text{render}}^{t,i,j}$ is the image rendered using camera $i$ as the input frame and camera $j$ as the output rendering view at frame $t$ . $\\mathbf{I}_{GT}^{t,j}$ is the ground truth frame captured by camera $j$ at frame $t$ . The Synthesis Quality $s$ is thus the average over all entries in S. For a dataset with multiple sequences, such as NeRSample, the final Synthesis Quality is the average across all sequences.", + "bbox": [ + 89, + 553, + 482, + 671 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Novel View Synthesis (NVS) Quality. Novel View Synthesis Quality $s_{NV}$ is the average over all entries where the input view $i$ is different from the novel view $j$ , essentially removing the diagonal entries from the score matrix:", + "bbox": [ + 89, + 671, + 482, + 731 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\ns _ {N V} = \\operatorname {M e a n} \\left(\\left\\{\\mathbf {S} _ {t, i, j} \\mid i \\neq j, 1 \\leq t \\leq T \\right\\}\\right). \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 143, + 741, + 480, + 758 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Identity Accuracy (ID) We measure the ArcFace [7] cosine distance between the ground truth image $I_{GT}$ and rendered image $I_{render}$ both from the frontal camera.", + "bbox": [ + 89, + 767, + 482, + 813 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Expression Accuracy (Expr) We use NVIDIA Maxine AR SDK [35] to measure the $L_{2}$ distance between expression coefficients $e_{GT}$ of the ground truth image and $e_{render}$ of the rendered image both from the frontal camera.", + "bbox": [ + 89, + 811, + 482, + 869 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Dynamic Appearance While we believe that it is important to measure the accuracy of dynamic appearance such", + "bbox": [ + 89, + 871, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "as lighting and shoulder poses, there is no existing multiview in-the-wild portrait video dataset to support such evaluation. We thus qualitatively evaluate various methods on challenging in-the-wild portrait videos. Please see the supplementary material for image examples and video results.", + "bbox": [ + 511, + 90, + 903, + 165 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Comparisons", + "text_level": 1, + "bbox": [ + 511, + 174, + 653, + 191 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines. We compare recent methods from 3 categories: (i) Reconstruction: We evaluate LP3D [52] using the above protocol. LP3D generates $N$ 3D portraits using each of the $N$ viewpoints as input, and the $N$ 3D portraits are then evaluated on the $N$ ground truth viewpoints. (ii) Reenactment: We evaluate Li et al. [28] and GPAvatar [6] in the self-reenactment setting. Li et al. [28] reconstruct 3D portraits as triplanes from reference images without test-time optimization, and they drive the 3D portraits via dynamic frontal renderings of 3DMMs. The authors of Li et al. [28] evaluated their approach using the input views as the only novel view, instead of all $N$ views. GPAvatar[6] reconstructs 3D portraits by leveraging multiple source images and driving them through a FLAME [26] mesh model. We use the first frame of the frontal camera in each NeRSample test sequence as the reference image to generate the 3D portrait, and drive it using videos from each of the $N$ viewpoints. We evaluate GPAvatar using the same evaluation protocol as our method and LP3D. (iii) Inversion: We evaluate VIVE3D [13], which is a state-of-the-art 3D GAN inversion method for videos, and it can also perform semantic video editing. We first provide VIVE3D 3 video frames from the input viewpoint, which it requires for personalization. Then, we invert the input video frames into 3D portraits and render the 3D portraits from all $N$ viewpoints. Each of the above method uses a different cropping of the face. We standardize the evaluation by re-cropping all methods to our cropping protocol, which is the largest of all. Please see the supplement for an additional table, where we evaluate the methods using different cropings around the face and arrive at conclusions consistent with Table 1.", + "bbox": [ + 511, + 196, + 906, + 651 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Quantitative results. We evaluate various methods using different input-novel viewpoint combinations, providing a robust multi-view assessment of a model's performance. Table. 1 shows that our model achieves state-of-the-art performance across all metrics versus recent works. Notably, while LP3D overfits to the input viewpoint (as shown by the difference of $0.806\\mathrm{dB}$ PSNR between Synthesis Quality and NVS Quality), our method generalizes better to novel viewpoints (as shown by the difference of $0.330\\mathrm{dB}$ ). Moreover, our method best preserves subject identity and expression (Table. 1, Fig. 2, and 4). On the other hand, the reenactment methods struggle to capture authentic expressions because of the use of morphable face models, which have limited expressiveness and accuracy. Moreover, they cannot reconstruct dynamic appearance (e.g. the stuck-out tongue in the second and third rows of Fig. 4) because they solely rely on information present in the reference image(s) and", + "bbox": [ + 511, + 651, + 908, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "10717", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4e299ad7c03841300890bb99d16bffa0a5a809729a23a80a534e72d20a65bd4f.jpg", + "image_caption": [ + "Groundtruth", + "Ours", + "LP3D", + "GPAvatar", + "Input / Driving Frame", + "Reference / Source", + "Figure 4. Visual comparisons with baseline methods. Our method strikes a balance between coherent reconstruction and faithfully preserving dynamic conditions like expressions. LP3D (third column) exhibits inconsistency in identities, hairstyles, and artifacts (red circles). GPAvatar (fourth column) fails to capture challenging expressions (first row), new information not present in the reference image, (the stuck-out tongue in second and third rows), and the identity of the person (last row)." + ], + "image_footnote": [], + "bbox": [ + 94, + 104, + 903, + 521 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "do not incorporate new per-frame information. On the other hand, our method achieves faithful dynamic appearance and coherent reconstruction at the same time.", + "bbox": [ + 89, + 602, + 482, + 646 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/895e615cb663b340f45519e636c291ac699f3bb4a68f77f17c31fc0cca5e8cdf.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodTypeExpr↓ID↓SynthesisLIPS↓LVS Quality
PSNR↑LPIPS↓PSNR↑LPIPS↓
Li et al. [28]reenact0.2660.24118.5730.25518.2020.262
GPAvatar[6]reenact0.2040.20721.9490.23321.9490.2334
VIVE3D[13]invert0.2900.39518.5770.25918.1450.271
LP3D[52]recon0.1680.21522.3310.22321.5250.237
Oursrecon0.1580.18722.7700.21922.4400.224
", + "bbox": [ + 91, + 672, + 482, + 795 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 1. Comparisons on Nersemble [23]. Our evaluation protocol (Sec. 4.2) utilizes multiview groundtruth for thorough analysis. Our method achieves state-of-the-art performance across all metrics, maintains temporal coherence, and preserves accurate expressions.", + "bbox": [ + 89, + 805, + 482, + 873 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Ablations", + "text_level": 1, + "bbox": [ + 513, + 601, + 624, + 614 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We evaluate the effectiveness of our proposed Triplane Undistorter and Fuser modules on NeRSemble. In addition to the Synthesis Quality in terms of PSNR and LPIPS, we develop two new metrics to separately measure the robustness to variations in (1) rendering viewpoints and (2) input viewpoints:", + "bbox": [ + 511, + 622, + 906, + 710 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "(1) Novel View Variation (NVV). We evaluate how much a method's reconstruction quality varies across different novel views. We quantify this as the standard deviation of performance across the $N$ novel views when using the same input view, i.e. the average standard deviation of each horizontal row of the score matrix $S$ :", + "bbox": [ + 511, + 710, + 906, + 799 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {N V V} = \\frac {1}{T N} \\sum_ {t = 1} ^ {T} \\sum_ {i = 1} ^ {N} \\operatorname {s t d d e v} \\left(\\left\\{S _ {t, i, j} \\mid 1 \\leq j \\leq N \\right\\}\\right). \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 806, + 906, + 848 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "(2) Input View Variation (IVV). We measure how much a method's reconstruction quality varies when using the $N$ different input viewpoints (Sec. 4.2 second column from the", + "bbox": [ + 511, + 856, + 906, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "10718", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/6549d0c6c73901cfe3faf3b718e3c7973a877153f8079d815f9bee5dc27252c2.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodU FSynthesis QualityIVV↓NVV↓
PSNR↑LPIPS↓(PSNR)(LPIPS)(PSNR)(LPIPS)
LP3D[52]XX22.3310.2231.0250.0152.2000.053
Ours✓X22.1960.2210.9070.0091.6990.038
X✓22.2650.2230.5590.0061.3150.029
✓✓22.7700.2190.2450.0051.3830.037
", + "bbox": [ + 91, + 88, + 486, + 195 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 2. Undistorter and Fuser Ablations. The Undistorter-only (row 2) and Fuser-only (row 3) variants both lead to improvements in viewpoint robustness metrics (IVV and NVV), but not in Synthesis Quality. The best performance is achieved by using both the Undistorter and the Fuser at the same time.", + "bbox": [ + 89, + 204, + 482, + 272 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "right). We quantify this variation as the average standard deviation of performance on the same novel view when using different input views, i.e. the average standard deviation of each vertical column of the score matrix $S$ :", + "bbox": [ + 89, + 287, + 483, + 347 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {I V V} = \\frac {1}{T N} \\sum_ {t = 1} ^ {T} \\sum_ {j = 1} ^ {N} \\operatorname {s t d d e v} \\left(\\left\\{S _ {t, i, j} \\mid 1 \\leq i \\leq N \\right\\}\\right). \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 102, + 351, + 482, + 393 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "$NVV$ and $IVV$ are measured for both PSNR and LPIPS, and they are better if lower.", + "bbox": [ + 89, + 401, + 482, + 430 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As shown in Table. 2, the Undistorter module consistently improves the NVV and IVV metrics versus LP3D, indicating better robustness to different input viewpoints and more consistent rendering quality across novel rendered views. However, when only the Undistorter is added (Tab. 2 second row) the PSNR is slightly reduced. This is likely because this model does not leverage the reference image to recover occluded areas while also overfitting to the input view. Similarly, the Fuser-only variant (Tab. 2 third row) achieves better NVV and IVV scores, but also scores lower in terms of PSNR. A likely cause is that, without the Undistorter, the Fuser needs to merge highly misaligned triplanes, where the person looks drastically different in $T_{raw}$ and $T_{cano}$ . This misalignment possibly induces more blurriness and alignment artifacts that lower the PSNR. The best performance is achieved by our full model, which includes both the Undistorter and the Fuser and thus achieves better PSNR, LPIPS, IVV and NVV scores. This means that the Undistorter and the Fuser complement each other: The Undistorter corrects the distortion in the raw triplane and thus reduces the challenges in fusing misaligned triplanes, and the Fuser recovers the occluded areas in the raw triplane.", + "bbox": [ + 89, + 431, + 482, + 767 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Shoulder Augmentation. without shoulder augmentation, the model fails to correctly reconstruct varying shoulder poses, resulting in worse performance (Fig. 5 and Table. 3).", + "bbox": [ + 89, + 768, + 482, + 813 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Optical Flow. As mentioned in Sec. 3.5, replacing the Undistorter with optical flow worsens the reconstruction (Table. 3). As shown by Fig. 5, this variant fails to correctly reconstruct the shoulder and the mouth accurately.", + "bbox": [ + 89, + 813, + 482, + 869 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "GAN Training. We retrained our model with an additional adversarial loss incorporated into Eqn. 10. As expected, the", + "bbox": [ + 89, + 869, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0bf2bdf2ae8f88065d6249dc2bbb6387132e237314eb57e37ea0f8f069b5b960.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 519, + 90, + 901, + 198 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/69e9b4d60aacd9bb22256783be28169cc859b9e1ba35bbb35d853a074bd8d22c.jpg", + "table_caption": [ + "Figure 5. Visual ablation. Our method with optical flow and w/o shoulder augmentation on two different input frames (top and bottom rows)." + ], + "table_footnote": [], + "table_body": "
MethodLPIPS↓PSNR↑IVV (PSNR) ↓ID↓Expr↓
LP3D (orig.)0.33418.7212.1300.2470.451
LP3D (ours)0.22322.3311.0250.1680.215
w optical flow0.22722.0851.1750.1780.335
w/o shoulder aug.0.21822.3420.8290.1530.244
Ours0.21922.7700.2450.1580.187
Ours + GAN0.20722.0740.3140.1520.198
", + "bbox": [ + 514, + 250, + 903, + 371 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3. Other Ablations. We compare our implementation of LP3D to the original one [52] (row 1&2); the effect of replacing our Undistorter with optical flow warping (row 4); the effect of removing shoulder augmentation during training (row 4); and the effect of adding an adversarial GAN loss (row 6) alongside Eqn. 10.", + "bbox": [ + 511, + 375, + 905, + 444 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "GAN training (Table. 3) improves the sharpness of images as seen from the LPIPS error reduction from 0.219 to 0.207. Additional results are included in the supplementary.", + "bbox": [ + 511, + 460, + 903, + 506 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Qualitative results. In-the-wild experiments show that our method achieves better temporal consistency than LP3D [52] and more accurately captures dynamic information like expressions and lighting changes than GPAvatar [6]. We refer the readers to our supplementary materials for visual examples of these results.", + "bbox": [ + 511, + 507, + 905, + 595 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Discussion", + "text_level": 1, + "bbox": [ + 511, + 609, + 627, + 626 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Conclusion. Recognizing the individual limitations of per-frame single-view reconstruction and 3D reenactment methods, we presented the first single-view 3D lifting method to reconstruct a 3D photorealistic avatar with faithful dynamic appearance as well as temporal consistency, marrying the best of both worlds. We believe our method paves the way forward for creating a high-quality telepresence system accessible to consumers.", + "bbox": [ + 511, + 636, + 905, + 753 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations and future work. We use a single reference image, but incorporating multiple ones with different expressions and head poses could lead to further improvements. With our method, fusing an extreme side view with a very different expression to the reference view may result in blurry reconstruction due to ambiguity in triplane alignment. Finally, due to the additional components, our current run-time performance is slower than real time, which could be improved in future work.", + "bbox": [ + 511, + 763, + 905, + 886 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "10719", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] ShahRukh Athar, Zexiang Xu, Kalyan Sunkavalli, Eli Shechtman, and Zhixin Shu. Rignerf: Fully controllable neural 3d portraits. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2", + "[2] Ananta R. Bhattarai, Matthias Nießner, and Artem Sevastopolsky. Triplanenet: An encoder for eg3d inversion. 2024. 1, 3", + "[3] Marcel C. Buehler, Kripasindhu Sarkar, Tanmay Shah, Gengyan Li, Daoye Wang, Leonhard Helminger, Sergio Orts-Escolano, Dmitry Lagun, Otmar Hilliges, Thabo Beeler, and Abhimitra Meka. Preface: A data-driven volumetric prior for few-shot ultra high-resolution face synthesis. In IEEE International Conference on Computer Vision (ICCV), 2023. 3", + "[4] Chen Cao, Tomas Simon, Jin Kyu Kim, Gabe Schwartz, Michael Zollhoefer, Shun-Suke Saito, Stephen Lombardi, Shih-En Wei, Danielle Belko, Shouou-I Yu, Yaser Sheikh, and Jason Saragih. Authentic volumetric avatars from a phone scan. ACM Transactions on Graphics (SIGGRAPH), 2022. 2", + "[5] Eric R. Chan, Connor Z. Lin, Matthew A. Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas Guibas, Jonathan Tremblay, Sameh Khamis, Tero Karras, and Gordon Wetzstein. Efficient geometry-aware 3D generative adversarial networks. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3", + "[6] Xuangeng Chu, Yu Li, Ailing Zeng, Tianyu Yang, Lijian Lin, Yunfei Liu, and Tatsuya Harada. Gpavatar: Generalizable and precise head avatar from image(s). In International Conference on Learning Representations (ICLR), 2024. 2, 3, 6, 7, 8", + "[7] Jiankang Deng, Jia Guo, Xue Niannan, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 6", + "[8] Yu Deng, Jiaolong Yang, Jianfeng Xiang, and Xin Tong. Gram: Generative radiance manifolds for 3d-aware image generation. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3", + "[9] Michail Christos Doukas, Stefanos Zafeiriou, and Viktoriia Sharmanska. Headgan: One-shot neural head synthesis and editing. In IEEE International Conference on Computer Vision (ICCV), 2021. 2", + "[10] Nikita Drobyshev, Jenya Chelishev, Taras Khakhulin, Aleksei Ivakhnenko, Victor Lempitsky, and Egor Zakharov. Megaportraits: One-shot megapixel neural head avatars. arXiv preprint arXiv:2207.07621, 2022. 2", + "[11] Yao Feng, Haiwen Feng, Michael J. Black, and Timo Bolkart. Learning an animatable detailed 3D face model from in-the-wild images. 2021. 3", + "[12] Yao Feng, Haiwen Feng, Michael J Black, and Timo Bolkart. Learning an animatable detailed 3d face model from in-the-wild images. ACM Transactions on Graphics (ToG), 40(4): 1-13, 2021. 3", + "[13] Anna Fruhstück, Nikolaos Sarafianos, Yuanlu Xu, Peter Wonka, and Tony Tung. VIVE3D: Viewpoint-independent video editing using 3D-Aware GANs. In IEEE Conference" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "on Computer Vision and Pattern Recognition (CVPR), 2023. 3, 6, 7", + "[14] Guy Gafni, Justus Thies, Michael Zollhöfer, and Matthias Nießner. Dynamic neural radiance fields for monocular 4d facial avatar reconstruction. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2", + "[15] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In Advances in Neural Information Processing Systems (NeurIPS), 2014. 3", + "[16] Jiatao Gu, Lingjie Liu, Peng Wang, and Christian Theobalt. StyleNeRF: A style-based 3D-aware generator for high-resolution image synthesis. arXiv preprint arXiv:2110.08985, 2021. 3", + "[17] Fa-Ting Hong, Longhao Zhang, Li Shen, and Dan Xu. Depth-aware generative adversarial network for talking head video generation. 2022. 2", + "[18] Yang Hong, Bo Peng, Haiyao Xiao, Ligang Liu, and Juyong Zhang. Headnerf: A real-time nerf-based parametric head model. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3", + "[19] Andrew Jones, Magnus Lang, Graham Fyffe, Xueming Yu, Jay Busch, Ian McDowall, Mark Bolas, and Paul Debevec. Achieving eye contact in a one-to-many 3d video teleconferencing system. ACM Transactions on Graphics (SIGGRAPH), 2009. 1", + "[20] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 3", + "[21] Peter Kauff and Oliver Schreer. An immersive 3d videoconferencing system using shared virtual team user environments. In Proceedings of the 4th International Conference on Collaborative Virtual Environments, 2002. 1", + "[22] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics (SIGGRAPH), 2023. 2", + "[23] Tobias Kirschstein, Shenhan Qian, Simon Giebenhain, Tim Walter, and Matthias Nießner. Nersemble: Multi-view radiance field reconstruction of human heads. ACM Trans. Graph., 2023. 5, 6, 7", + "[24] Jaehoon Ko, Kyusun Cho, Daewon Choi, Kwangrok Ryoo, and Seungryong Kim. 3d gan inversion with pose optimization. In IEEE Winter Conference on Applications of Computer Vision (WACV), 2023. 3", + "[25] Jason Lawrence, Dan B Goldman, Supreeth Achar, Gregory Major Blascovich, Joseph G. Desloge, Tommy Fortes, Eric M. Gomez, Sascha Haberling, Hugues Hoppe, Andy Huibers, Claude Knaus, Brian Kuschak, Ricardo Martin-Brualla, Harris Nover, Andrew Ian Russell, Steven M. Seitz, and Kevin Tong. Project starline: A high-fidelity telepresence system. ACM Transactions on Graphics (SIGGRAPH ASIA), 2021. 1", + "[26] Tianye Li, Timo Bolkart, Michael. J. Black, Hao Li, and Javier Romero. Learning a model of facial shape and expression from 4D scans. 2017. 3, 6", + "[27] Weichuang Li, Longhao Zhang, Dong Wang, Bin Zhao, Zhi-gang Wang, Mulin Chen, Bang Zhang, Zhongjian Wang," + ], + "bbox": [ + 517, + 93, + 903, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "10720", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Liefeng Bo, and Xuelong Li. One-shot high-fidelity talking-head synthesis with deformable neural radiance field. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2, 3", + "[28] Xueting Li, Shalini De Mello, Sifei Liu, Koki Nagano, Umar Iqbal, and Jan Kautz. Generalizable one-shot neural head avatar. 2023. 3, 6, 7", + "[29] Jingyun Liang, Yuchen Fan, Xiaoyu Xiang, Rakesh Ranjan, Eddy Ilg, Simon Green, Jiezhang Cao, Kai Zhang, Radu Timofte, and Luc Van Gool. Recurrent video restoration transformer with guided deformable attention. 2022. 5", + "[30] C.Z. Lin, D.B. Lindell, E.R. Chan, and G. Wetzstein. 3d gan inversion for controllable portrait image animation. In ECCV Workshop on Learning to Generate 3D Shapes and Scenes, 2022. 3", + "[31] S. Ma, T. Simon, J. Saragih, D. Wang, Y. Li, F. La Torre, and Y. Sheikh. Pixel codec avatars. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 1", + "[32] Zhiyuan Ma, Xiangyu Zhu, Guojun Qi, Zhen Lei, and Lei Zhang. Otavatar: One-shot talking face avatar with controllable tri-plane rendering. 2023. 3", + "[33] Andrew Maimone, Jonathan Bidwell, Kun Peng, and Henry Fuchs. Enhanced personal autostereoscopic telepresence system using commodity depth cameras. Computers & Graphics, 2012. 1", + "[34] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing scenes as neural radiance fields for view synthesis. In European Conference on Computer Vision (ECCV), 2020. 2, 3", + "[35] NVIDIA. Nvidia maxine ar sdk. https://github.com/NVIDIA/MAXINE-AR-SDK, 2024.6", + "[36] Roy Or-El, Xuan Luo, Mengyi Shan, Eli Shechtman, Jeong Joon Park, and Ira Kemelmacher-Shlizerman. StyleSDF: High-Resolution 3D-Consistent Image and Geometry Generation. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3", + "[37] Sergio Orts-Escalano, Christoph Rhemann, Sean Fanello, Wayne Chang, Adarsh Kowdle, Yury Degtyarev, David Kim, Philip L. Davidson, Sameh Khamis, Mingsong Dou, Vladimir Tankovich, Charles Loop, Qin Cai, Philip A. Chou, Sarah Mennicken, Julien Valentin, Vivek Pradeep, Shenlong Wang, Sing Bing Kang, Pushmeet Kohli, Yuliya Lutchyn, Cem Keskin, and Shahram Izadi. Holoportation: Virtual 3d teleportation in real-time. In UIST 2016, 2016. 1", + "[38] Keunhong Park, Utkarsh Sinha, Jonathan T. Barron, Sofien Bouaziz, Dan B Goldman, Steven M. Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. 2021. 2", + "[39] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-NeRF: Neural Radiance Fields for Dynamic Scenes. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2", + "[40] Luchao Qi, Jiaye Wu, Annie N. Wang, Shengze Wang, and Roni Sengupta. My3dgen: A scalable personalized 3d generative model, 2023. 3", + "[41] Shenhan Qian, Tobias Kirschstein, Liam Schoneveld, Davide Davoli, Simon Giebenhain, and Matthias Nießner. Gaussianavatars: Photorealistic head avatars with rigged 3d gaussians. 2024. 2" + ], + "bbox": [ + 91, + 90, + 482, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[42] Anurag Ranjan and Michael J Black. Optical flow estimation using a spatial pyramid network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4161-4170, 2017. 4", + "[43] Ramesh Raskar, Greg Welch, Matt Cutts, Adam Lake, Lev Stesin, and Henry Fuchs. The office of the future: a unified approach to image-based modeling and spatially immersive displays. In ACM Transactions on Graphics (SIGGRAPH), 1998. 1", + "[44] Daniel Rebain, Mark Matthews, Kwang Moo Yi, Dmitry Lagan, and Andrea Tagliasacchi. Lolnerf: Learn from one look. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3", + "[45] Shunsuke Saito, Gabriel Schwartz, Tomas Simon, Junxuan Li, and Giljoo Nam. Relightable gaussian codec avatars. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2", + "[46] Aliaksandr Siarohin, Stéphane Lathuilière, Sergey Tulyakov, Elisa Ricci, and Nicu Sebe. First order motion model for image animation. In Advances in Neural Information Processing Systems (NeurIPS), 2019. 2", + "[47] Ivan Skorokhodov, Sergey Tulyakov, Yiqun Wang, and Peter Wonka. Epigraph: Rethinking training of 3d gans. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 3", + "[48] Michael Stengel, Koki Nagano, Chao Liu, Matthew Chan, Alex Trevithick, Shalini De Mello, Jonghyun Kim, and David Luebke. AI-Mediated 3D Video Conferencing. In ACM SIGGRAPH 2023 Emerging Technologies, 2023. 1, 3", + "[49] Jingxiang Sun, Xuan Wang, Yichun Shi, Lizhen Wang, Jue Wang, and Yebin Liu. Ide-3d: Interactive disentangled editing for high-resolution 3d-aware portrait synthesis. ACM Transactions on Graphics (SIGGRAPH ASIA), 2022. 3", + "[50] Jingxiang Sun, Xuan Wang, Lizhen Wang, Xiaoyu Li, Yong Zhang, Hongwen Zhang, and Yebin Liu. Next3d: Generative neural texture rasterization for 3d-aware head avatars. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 3, 4", + "[51] Phong Tran, Egor Zakharov, Long-Nhat Ho, Anh Tuan Tran, Liwen Hu, and Hao Li. Voodoo 3d: Volumetric portrait disentanglement for one-shot 3d head reenactment. 2024. 2, 3", + "[52] Alex Trevithick, Matthew Chan, Michael Stengel, Eric R. Chan, Chao Liu, Zhiding Yu, Sameh Khamis, Manmohan Chandraker, Ravi Ramamoorthi, and Koki Nagano. Realtime radiance fields for single-image portrait view synthesis. In ACM Transactions on Graphics (SIGGRAPH), 2023. 1, 2, 3, 4, 6, 7, 8", + "[53] Ting-Chun Wang, Arun Mallya, and Ming-Yu Liu. One-shot free-view neural talking-head synthesis for video conferencing. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2", + "[54] Yaohui Wang, Di Yang, Francois Bremond, and Antitza Dantcheva. Latent image animator: Learning to animate images via latent space navigation. In International Conference on Learning Representations (ICLR), 2022. 2", + "[55] Jianfeng Xiang, Jiaolong Yang, Yu Deng, and Xin Tong. Gram-hd: 3d-consistent image generation at high resolution with generative radiance manifolds. arXiv preprint arXiv:2206.07255, 2022. 3" + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10721", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[56] Yinghao Xu, Sida Peng, Ceyuan Yang, Yujun Shen, and Bolei Zhou. 3d-aware image synthesis via learning structural and textural representations. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3", + "[57] Yuelang Xu, Benwang Chen, Zhe Li, Hongwen Zhang, Lizhen Wang, Zerong Zheng, and Yebin Liu. Gaussian head avatar: Ultra high-fidelity head avatar via dynamic gaussians. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2", + "[58] Zhongcong Xu, Jianfeng Zhang, Junhao Liew, Wenqing Zhang, Song Bai, Jiashi Feng, and Mike Zheng Shou. Pv3d: A 3d generative model for portrait video generation. In The Tenth International Conference on Learning Representations, 2023. 3", + "[59] Zhenhui Ye, Tianyun Zhong, Yi Ren, Jiaqi Yang, Weichuang Li, Jiangwei Huang, Ziyue Jiang, Jinzheng He, Rongjie Huang, Jinglin Liu, Chen Zhang, Xiang Yin, Zejun Ma, and Zhou Zhao. Real3d-portrait: One-shot realistic 3d talking portrait synthesis. In International Conference on Learning Representations (ICLR), 2024. 2, 3", + "[60] Fei Yin, Yong Zhang, Xiaodong Cun, Mingdeng Cao, Yanbo Fan, Xuan Wang, Qingyan Bai, Baoyuan Wu, Jue Wang, and Yujiu Yang. Styleheat: One-shot high-resolution editable talking face generation via pre-trained stylegan. 2022. 2", + "[61] Wangbo Yu, Yanbo Fan, Yong Zhang, Xuan Wang, Fei Yin, Yunpeng Bai, Yan-Pei Cao, Ying Shan, Yang Wu, Zhongqian Sun, and Baoyuan Wu. Nofa: Nerf-based one-shot facial avatar reconstruction. In ACM SIGGRAPH 2023 Conference Proceedings, 2023. 3", + "[62] Egor Zakharov, Aliaksandra Shysheya, Egor Burkov, and Victor Lempitsky. Few-shot adversarial learning of realistic neural talking head models. In IEEE International Conference on Computer Vision (ICCV), 2019. 2", + "[63] Egor Zakharov, Aleksei Ivakhnenko, Aliaksandra Shysheya, and Victor Lempitsky. Fast bi-layer neural synthesis of one-shot realistic head avatars. In European Conference on Computer Vision (ECCV), 2020.", + "[64] Bowen Zhang, Chenyang Qi, Pan Zhang, Bo Zhang, Hsiang-Tao Wu, Dong Chen, Qifeng Chen, Yong Wang, and Fang Wen. Metaportrait: Identity-preserving talking head generation with fast personalized adaptation, 2023. 2", + "[65] Xuanmeng Zhang, Zhedong Zheng, Daiheng Gao, Bang Zhang, Pan Pan, and Yi Yang. Multi-view consistent generative adversarial networks for 3d-aware image synthesis. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3", + "[66] Jian Zhao and Hui Zhang. Thin-plate spline motion model for image animation. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 3657-3666, 2022. 2", + "[67] Yufeng Zheng, Victoria Fernández Abrevaya, Marcel C. Bühler, Xu Chen, Michael J. Black, and Otmar Hilliges. I M Avatar: Implicit morphable head avatars from videos. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2", + "[68] Peng Zhou, Lingxi Xie, Bingbing Ni, and Qi Tian. CIPS-3D: A 3D-Aware Generator of GANs Based on Conditionally-Independent Pixel Synthesis. arXiv preprint arXiv:2110.09788, 2021. 3" + ], + "bbox": [ + 91, + 90, + 480, + 898 + ], + "page_idx": 10 + }, + { + "type": "ref_text", + "text": "[69] Yiyu Zhuang, Hao Zhu, Xusen Sun, and Xun Cao. Mofanerf: Morphable facial neural radiance field. In European Conference on Computer Vision (ECCV), 2022. 3", + "bbox": [ + 516, + 90, + 903, + 133 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "10722", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Coherent 3D Portrait Video Reconstruction via Triplane Fusion/2088b6b2-40e1-4341-a0c5-c5303c3acd65_model.json b/2025/Coherent 3D Portrait Video Reconstruction via Triplane Fusion/2088b6b2-40e1-4341-a0c5-c5303c3acd65_model.json new file mode 100644 index 0000000000000000000000000000000000000000..4dcdcf1b3adf4d5c618fd364bb9e4ad9acdc8f4f --- /dev/null +++ b/2025/Coherent 3D Portrait Video Reconstruction via Triplane Fusion/2088b6b2-40e1-4341-a0c5-c5303c3acd65_model.json @@ -0,0 +1,2367 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.179, + 0.13, + 0.82, + 0.152 + ], + "angle": 0, + "content": "Coherent 3D Portrait Video Reconstruction via Triplane Fusion" + }, + { + "type": "text", + "bbox": [ + 0.117, + 0.179, + 0.88, + 0.252 + ], + "angle": 0, + "content": "Shengze Wang\\(^{1*}\\) Xueting Li\\(^{2}\\) Chao Liu\\(^{2}\\) Matthew Chan\\(^{2}\\) Michael Stengel\\(^{2}\\) Henry Fuchs\\(^{1}\\) Shalini De Mello\\(^{2\\dagger}\\) Koki Nagano\\(^{2\\dagger}\\) \n\\(^{1}\\) UNC Chapel Hill \\({}^{2}\\) NVIDIA \\(\\dagger\\) Equal Contribution \nhttps://research.nvidia.com/labs/amri/projects/coherent3d/" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.285, + 0.327, + 0.301 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.317, + 0.486, + 0.669 + ], + "angle": 0, + "content": "Recent breakthroughs in single-image 3D portrait reconstruction have enabled telepresence systems to stream 3D portrait videos from a single camera in real-time, democratizing telepresence. However, per-frame 3D reconstruction exhibits temporal inconsistency and forgets the user's appearance. On the other hand, self-reenactment methods can render coherent 3D portraits by driving a 3D avatar built from a single reference image but fail to faithfully preserve the user's per-frame appearance (e.g., instantaneous facial expressions and lighting). As a result, neither of these two frameworks is an ideal solution for democratized 3D telepresence. In this work, we address this dilemma and propose a novel solution that maintains both coherent identity and dynamic per-frame appearance to enable the best possible realism. To this end, we propose a new fusion-based method that takes the best of both worlds by fusing a canonical 3D prior from a reference view with dynamic appearance from per-frame input views, producing temporally stable 3D videos with faithful reconstruction of the user's per-frame appearance. Trained only using synthetic data produced by an expression-conditioned 3D GAN, our encoder-based method achieves both state-of-the-art 3D reconstruction and temporal consistency on in-studio and in-the-wild datasets." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.681, + 0.222, + 0.697 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.7, + 0.485, + 0.876 + ], + "angle": 0, + "content": "Telepresence aims at bringing distant people face-to-face and stands out as a particularly compelling application of computer vision and graphics. Over the last decades, various successful telepresence systems [19, 21, 21, 25, 31, 33, 37, 43] have been developed. However, most employ bulky multi-view 3D scanners or depth sensors to ensure high-quality volumetric per-frame reconstruction. Unlike these classical 3D/4D reconstruction methods, recent AI-based feed-forward 3D lifting techniques [2, 52] can lift a single RGB image from an off-the-shelf webcam into a neural radiance field (NeRF) representation encoded into a set of triplanes in real time, paving the way towards making 3D" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.286, + 0.774, + 0.302 + ], + "angle": 0, + "content": "telepresence accessible to anyone [48]." + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.314, + 0.907, + 0.596 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.605, + 0.907, + 0.756 + ], + "angle": 0, + "content": "Figure 1. Comparisons. Given a single reference image and a single-view video frame, our method reconstructs the authentic dynamic appearance of the user (e.g., facial expressions and lighting) while producing a temporally coherent 3D video. A previous single-view 3D lifting method (LP3D) that reconstructs the avatar from the video frame on a per-frame basis suffers from distortions and temporal inconsistency. A portrait reenactment method (GPAvatar) drives the identity in the reference image using the video frame, but fails to capture accurate facial expressions (e.g., smile) and per-frame appearance (e.g., lighting). The output should be compared to the appearance of the input video frame (green box)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.769, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Currently, there are two major paradigms in democratized 3D telepresence solutions from a single-view video: (1) single-view per-frame 3D lifting methods and (2) 3D portrait reenactment, which drives an identity from a reference image using another driving frame, but none of them is an ideal solution. For (1), single-frame-based lifting techniques such as LP3D [52], have the advantage of faithfully preserving the instantaneous dynamic conditions present in an input video, e.g. lighting, expressions, and posture, all" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.887, + 0.428, + 0.901 + ], + "angle": 0, + "content": "*Shengze Wang was an intern at NVIDIA during the project" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10712" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.428 + ], + "angle": 0, + "content": "of which are crucial to an authentic telepresence experience. However, single-image reconstruction methods operate independently on each frame and thus have fundamental limitations in maintaining temporal consistency. This difficulty stems from the inherent ill-posed nature of single-image-based reconstruction. In order to render novel views that are significantly far from the input view, the system cannot rely on information present in the input view and hence must hallucinate generate plausible content, which cannot be guaranteed to be consistent across multiple temporal frames. For example, LP3D's reconstruction changes significantly depending on the user's head pose in the input frame (see second row in Fig. 1 and third column in Fig. 2). In comparison, 3D self-reenactment methods create an avatar model from one or multiple reference images and use a separate driving video to drive the facial expressions and poses of the avatar. [6, 27, 51, 59]. While these approaches allow for temporally consistent results, they often cannot faithfully reconstruct the input video's dynamic conditions such as changes in lighting. Moreover, reenactment methods often struggle to generate accurate expressions because their expression detection and control are not precise enough (see the first row in Fig. 1)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.429, + 0.484, + 0.56 + ], + "angle": 0, + "content": "In this work, we address this dilemma of democratized 3D telepresence approaches and propose a novel solution to the problem of simultaneously maintaining temporal stability while preserving real-time dynamics of input videos for single-camera telepresence applications. Our proposed solution is a fusion-based approach that leverages the stability and accuracy of a canonical 3D prior, and also captures the diverse deviations from the prior in newer video frames (see the third row in Fig. 1)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.561, + 0.484, + 0.794 + ], + "angle": 0, + "content": "Our model first uses LP3D [52] to construct a canonical triplane prior from a (near) frontal image of the user, which can be casually captured or extracted from a video. During video reconstruction, our model lifts each input frame into a raw triplane, which is then fused with the canonical triplane (see Fig. 3). For images with oblique yaw head poses, artifacts and identity-related distortions may be present in its lifted triplane (see \"LP3D\" in Fig. 1 and 2). Hence we propose an undistorter module, which learns to undistort the raw instantaneous triplane to more closely match the identity of the correctly structured canonical triplane. We then propose a fuser module that combines the undistorted triplane and the canonical triplane in order to preserve identity consistency and recover occluded regions while reconstructing the dynamic lighting, nuanced expressions, and posture information in the input frame." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.795, + 0.403, + 0.809 + ], + "angle": 0, + "content": "We summarize our contributions as follows:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.81, + 0.483, + 0.869 + ], + "angle": 0, + "content": "- We contribute a novel triplane fusion method that combines the dynamic information from per-frame triplanes with a canonical triplane extracted from a reference image. Trained only using a synthetic multi-view video" + }, + { + "type": "page_footnote", + "bbox": [ + 0.09, + 0.875, + 0.483, + 0.901 + ], + "angle": 0, + "content": "1We find that the 3D lifting from a near frontal reference view is reliable, hence use this as a canonical 3D prior. See Fig. 2 first column." + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.092, + 0.907, + 0.166 + ], + "angle": 0, + "content": "dataset, our feed-forward approach generates 3D portrait videos that demonstrate both temporal consistency and faithful reconstruction of the dynamic appearance of the user (e.g. lighting and expression), whereas prior solutions can only achieve one of the two properties." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.166, + 0.907, + 0.223 + ], + "angle": 0, + "content": "- We propose a new framework to evaluate single-view 3D portrait reconstruction methods using multi-view data and gain insight into the method's reconstruction quality and robustness." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.224, + 0.906, + 0.282 + ], + "angle": 0, + "content": "- We present evaluations on both in-studio and in-the-wild data and demonstrate that our method achieves state-of-the-art performance in terms of temporal consistency and reconstruction accuracy." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.166, + 0.907, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.298, + 0.901, + 0.525 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.537, + 0.907, + 0.658 + ], + "angle": 0, + "content": "Figure 2. View-Dependent Distortion. \"NV\" refers to \"Novel View Rendering\". Top: inputs to our model and LP3D. Second & Third Rows: LP3D's reconstruction varies greatly under challenging viewpoints, showing a predictable pattern of artifacts including abnormally strong activations on the side being captured (red circle), as well as geometric distortion along the view direction of the camera due to depth ambiguity. We refer to this phenomenon as \"View-Dependent Distortion\". Fourth row: Our method removes such artifacts and achieves better coherence." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.669, + 0.655, + 0.685 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.695, + 0.906, + 0.797 + ], + "angle": 0, + "content": "2D portrait reenactment. Given a single or a few reference portrait images and a driving video, recent talking-head generators can reenact 2D portraits by transferring the facial expressions and poses from the driving video onto facial portraits [9, 10, 17, 46, 53, 54, 60, 62-64, 66]. However, being 2D, they cannot be rendered from novel viewpoints, which is crucial for 3D telepresence." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.799, + 0.907, + 0.901 + ], + "angle": 0, + "content": "3D-aware portrait generation and reenactment. Some recent works use deformable volumetric implicit radiance fields [34, 38, 39] or Gaussian splatting [22] combined by 3DMMs to reconstruct a photorealistic and animatable volumetric head avatar [1, 4, 14, 41, 45, 57, 67]. However, they require extensive data captures from videos or multiview cameras and person-specific training. Others use" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10713" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.238 + ], + "angle": 0, + "content": "large-scale video datasets and learn a disentangled triplane 3D [5] for 3D facial reenactment in a feedforward fashion [6, 27, 28, 32, 51, 59, 61]. They construct a canonical 3D head from a reference image (often a neutral frame), and use facial expressions and head poses extracted from a separate driving video to animate it. As such, fine-grained facial expressions may not be captured due to errors in disentanglement. Most importantly, these reenactment methods fail to preserve the dynamic appearance of users (e.g., person-specific wrinkles or lighting) across time." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.239, + 0.487, + 0.591 + ], + "angle": 0, + "content": "3D GAN inversion. By combining GANs [15] and neural volume rendering [34], recent breakthroughs in 3D-aware GANs [5, 8, 16, 36, 44, 47, 55, 56, 58, 65, 68] can learn to generate photorealistic 3D heads from in-the-wild 2D images. Notably, EG3D proposes an efficient and compact triplane representation [5] to generate 3D heads, and Next3D [50] extends EG3D to create 3D portrait videos controlled by 3DMM facial expression and pose parameters. Besides generating 3D heads, these models can also be used for single-view 3D reconstruction using GAN inversion [13, 24, 30, 49], manipulation of the 3D head [18, 49, 69], and 3D personalization [3, 40]. However, inverting a few seconds of video can take many minutes or hours, and the inversion quality is often unsatisfactory due to inaccurate identity and expression preservation, as evaluated by GPAvatar [6]. Therefore, recent works [2, 52] propose encoder-based solutions to lift each video frame into a triplane. However, when each frame is independently lifted into 3D, the resulting 3D head video suffers in terms of temporal consistency - a key requirement of 3D telepresence systems. To enhance single-frame 2D-to-3D encoders for human heads, e.g., LP3D [52], we propose a triplane-fusion-based method, which improves their temporal stability while preserving temporal dynamics across time." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.604, + 0.182, + 0.619 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.628, + 0.482, + 0.645 + ], + "angle": 0, + "content": "3.1. Background: 3D Portrait from a Single Image" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.649, + 0.483, + 0.797 + ], + "angle": 0, + "content": "LP3D [52] performs photorealistic 3D portrait reconstruction by using a feedforward encoder to lift an RGB image into a triplane \\(\\mathbf{T} \\in \\mathbb{R}^{3 \\times 32 \\times 256 \\times 256}\\), which can be volume rendered to an RGB image from any viewpoint. LP3D can run in real-time and has been developed into a complete realtime telepresence system [48]. We lift 2D faces into triplanes with a slightly modified LP3D trained on larger face crops containing shoulders and a camera estimator to recover the input image's camera parameters \\(M \\in \\mathbb{R}^{25}\\). It performs better than the original version (Table 3)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.806, + 0.212, + 0.821 + ], + "angle": 0, + "content": "3.2. Definitions" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.827, + 0.483, + 0.901 + ], + "angle": 0, + "content": "We call a current video frame an \"input frame\" into our system. Additionally, we encode a near-frontal \"reference image\" of the subject into a \"canonical triplane\", which is used to stabilize the 3D video generation process. Lastly, we refer to the viewpoint of the camera in the input frame" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.092, + 0.664, + 0.107 + ], + "angle": 0, + "content": "as \"input viewpoint\"." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.116, + 0.625, + 0.131 + ], + "angle": 0, + "content": "3.3. Overview" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.137, + 0.907, + 0.373 + ], + "angle": 0, + "content": "An overview of our method is illustrated in Fig. 3. To improve temporal consistency and reconstruct occluded parts of the face, we leverage an additional near-frontal reference image, which can be a selfie image or a video frame automatically obtained from the same video using an off-the-shelf head pose estimator like DECA [11]. We lift this reference image into the canonical triplane using pre-trained LP3D. When processing the video, we first lift each input frame into a raw triplane. Then, the Triplane Undistorter (Sec. 3.5) removes view-dependent distortions and recovers the identity using the canonical triplane as a reference. This produces an undistorted triplane. Finally, to recover regions that are occluded in the input frame and further improve stability, our Triplane Fuser (Sec. 3.6) combines the undistorted triplane with the canonical triplane to generate the final coherent triplane." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.381, + 0.907, + 0.397 + ], + "angle": 0, + "content": "3.4. Generating Synthetic Dynamic Multiview Data" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.403, + 0.909, + 0.739 + ], + "angle": 0, + "content": "We synthesize ground truth multiview training images using Next3D [50]. We first extract 2D landmarks and FLAME [26] coefficient labels for the FFHQ [20] dataset using DECA [12]. Then, during training, we sample a pair of random FLAME coefficients from our FFHQ labels and input them to Next3D alongside a single random identity code \\( z \\). We notice that FFHQ contains mostly less extreme expressions. Thus, we randomly multiply the expression codes by a small scaling factor to exaggerate the expressions during training. Notice that the Next3D generator is much more expressive than FLAME due to its use of triplanes and GAN training even though it is conditioned on FLAME. Next3D then generates a pair of triplanes for \\( t = 0 \\) and \\( t = 1 \\) of the same person with 2 expressions. We render the \\( t = 0 \\) triplane into a near-frontal reference image \\( I_{ref} \\) (Fig. 3 orange box) and the \\( t = 1 \\) triplane into 3 images: the input frame rendered from a random viewpoint (green box), the ground truth image at a different viewpoint (red box), and a frontal-view image (red box). Only the latter two are used for supervision. Additionally, to simulate different lighting conditions, we also apply different color augmentations (brightness, contrast, saturation, and hue) to images at \\( t = 0 \\) and \\( t = 1 \\)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.74, + 0.907, + 0.856 + ], + "angle": 0, + "content": "Shoulder augmentation. It is important that 3D portraits capture dynamic shoulder movements to convey body language and achieve eye contact in telepresence. However, Next3D does not provide control over shoulder posture. Thus, we simulate shoulder movement in the rendered images without modifying the Next3D triplanes by warping camera rays during volume rendering. Please see the supplement for more detail." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.857, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Pseudo-groundtruth triplanes. Due to our shoulder augmentation, the Next3D triplanes and their 2D renderings become inconsistent. Thus, Next3D's triplanes cannot be used" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "10714" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.107, + 0.096, + 0.885, + 0.426 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.434, + 0.908, + 0.492 + ], + "angle": 0, + "content": "Figure 3. Method Overview. Given a near-frontal reference image and an input frame, we first reconstruct a canonical triplane and a raw triplane using LP3D [52] (Sec. 3.1). Next, we combine them through a Triplane Fusion module (blue box) that ensures temporal consistency while preserving dynamic information (e.g., lighting and shoulder pose) (Sec. 3.5 and 3.6). Our model is trained with only synthetic video data generated by a 3D GAN [50], with carefully designed augmentations to preserve shoulder motion and lighting (Sec. 3.4)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.497, + 0.487, + 0.572 + ], + "angle": 0, + "content": "as direct supervisory signals. To mitigate this, we leverage the fact that LP3D generates reasonably accurate triplanes from frontal view images. We use a frozen LP3D to predict pseudo-groundtruth triplanes \\( T_{\\text{triplaneGT}} \\) from the frontal novel view \\( I_{\\text{frontal}} \\) for \\( t = 1 \\) (Fig. 3 bottom right)." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.583, + 0.473, + 0.601 + ], + "angle": 0, + "content": "3.5. Removing Distortion and Preserving Identity" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.606, + 0.483, + 0.825 + ], + "angle": 0, + "content": "LP3D's reconstruction quality strongly depends on the input frame's viewpoint. When subjects are captured from the side, LP3D tends to produce errors in identity and exhibits stretching distortions (e.g., in the \"Input Frame 1\" and \"Input Frame 2\" columns in Fig. 2), primarily due to the inherent ambiguity of single-image reconstruction. In contrast, near-frontal views offer more complete identity cues and suffer less from occlusion (e.g. the \"Reference Image\" column in Fig. 2), leading to more reliable reconstruction. To reduce the single-image ambiguity, we adopt a near-frontal reconstruction as the canonical triplane \\( T_{\\text{cano}} \\). Although incorporating more reference views could potentially improve accuracy, we find that using a single near-frontal view not only simplifies the model but also sufficiently enhances temporal coherence." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.827, + 0.483, + 0.872 + ], + "angle": 0, + "content": "To correct the distortion in an input raw triplane \\( T_{raw} \\) using the canonical triplane \\( T_{cano} \\) as reference, we devise a Triplane Undistorter \\( U \\) (Fig. 3):" + }, + { + "type": "equation", + "bbox": [ + 0.131, + 0.884, + 0.483, + 0.903 + ], + "angle": 0, + "content": "\\[\nU \\left(T _ {\\text {r a w}}, T _ {\\text {c a n o}}\\right) = T _ {\\text {u n d i s t}} \\in \\mathbb {R} ^ {3 \\times 3 2 \\times 2 5 6 \\times 2 5 6}. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.497, + 0.907, + 0.6 + ], + "angle": 0, + "content": "Since the distortion in each triplane is a 2D warping artifact, it can be reversed through a 2D undistortion warping of each plane. To this end, our Undistorter \\(U\\) adopts the SPyNet[42] architecture to predict a 2D correction warping \\(T_{corr} \\in \\mathbb{R}^{3 \\times 2 \\times 256 \\times 256}\\) for the three planes. Then, the Undistorter corrects the raw triplane \\(T_{raw}\\) by warping it based on the predicted 2D offsets \\((\\Delta u, \\Delta v)\\) in \\(T_{corr}\\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.599, + 0.605, + 0.907, + 0.621 + ], + "angle": 0, + "content": "\\[\nT _ {c o r r} = S P y N e t \\left(T _ {\\text {r a w}}, T _ {\\text {c a n o}}\\right), \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.604, + 0.626, + 0.905, + 0.642 + ], + "angle": 0, + "content": "\\[\nT _ {u n d i s t} = \\operatorname {W a r p} \\left(T _ {\\text {r a w}}, T _ {\\text {c o r r}}\\right). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.647, + 0.909, + 0.883 + ], + "angle": 0, + "content": "It is important to note the difference between optical flow and undistortion. An ideal optical flow model would warp \\( T_{\\text{raw}} \\) towards \\( T_{\\text{cano}} \\) in order to make the two identical. However, this alignment leads to significant artifacts and destroys the reconstruction (Table. 3 and Fig. 5). This is because the purpose of undistortion is not to make the two triplanes identical, but to correct \\( T_{\\text{raw}} \\) such that it has the same identity as the canonical triplane \\( T_{\\text{cano}} \\), while preserving dynamic information such as expressions and lighting. Therefore, instead of using the canonical triplane \\( T_{\\text{cano}} \\) as the warping target, our Undistorter \\( U \\) merely uses \\( T_{\\text{cano}} \\) as the identity conditioning to correct the raw triplane \\( T_{\\text{raw}} \\) towards the ground truth triplane \\( T_{\\text{triplaneGT}} \\) (Fig. 3, lower-right), which is not available during test time. The correction warping is supervised by the consistency between the undistorted triplane \\( T_{\\text{undist}} \\) and pseudo-groundtruth tri" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10715" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.405, + 0.108 + ], + "angle": 0, + "content": "plane \\(T_{\\text{triplane}GT}\\) (Sec. 3.4) via a triplane loss:" + }, + { + "type": "equation", + "bbox": [ + 0.164, + 0.117, + 0.482, + 0.134 + ], + "angle": 0, + "content": "\\[\nL _ {u n d i s t} = L _ {1} \\left(T _ {u n d i s t}, T _ {t r i p l a n e G T}\\right). \\tag {4}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.139, + 0.483, + 0.17 + ], + "angle": 0, + "content": "3.6. Reconstructing Occluded Regions through Triplane Fusion" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.175, + 0.484, + 0.442 + ], + "angle": 0, + "content": "As the user moves around in the video, different parts of their head become occluded. To recover occluded areas in the input frame and further stabilize the subject's identity across the video, our Fuser \\(F\\) enhances the reconstruction by utilizing the canonical triplane \\(T_{\\text{cano}}\\), in which the currently occluded areas are often visible. Therefore, it is important for the Fuser \\(F\\) to identify and recover the occluded regions while preserving information from the visible regions in \\(T_{\\text{undist}}\\). To accomplish this, we thus use a 5-layer ConvNet-based visibility estimator \\(V\\) to predict a visibility triplane as \\(V(T_{\\text{raw}}) = T_{\\text{raw}}^{\\text{vis}} \\in \\mathbb{R}^{3 \\times 1 \\times 128 \\times 128}\\), i.e., one visibility map for each plane. \\(T_{\\text{raw}}^{\\text{vis}}\\) is undistorted alongside \\(T_{\\text{undist}}\\) to produce \\(T_{\\text{undist}}^{\\text{vis}}\\). We also predict a visibility triplane for the canonical triplane as \\(T_{\\text{cano}}^{\\text{vis}} = V(T_{\\text{cano}})\\). Finally, the Fuser \\(F\\) produces the fused triplane \\(T_{\\text{fused}}\\) by combining information from the undistorted input tranplane \\(T_{\\text{undist}}\\), its visibility triplane \\(T_{\\text{undist}}^{\\text{vis}}\\), the canonical triplane \\(T_{\\text{cano}}\\), and its visibility triplane \\(T_{\\text{cano}}^{\\text{vis}}\\)." + }, + { + "type": "equation", + "bbox": [ + 0.138, + 0.45, + 0.483, + 0.468 + ], + "angle": 0, + "content": "\\[\nT _ {f u s e d} = F \\left(T _ {u n d i s t}, T _ {u n d i s t} ^ {\\text {v i s}}, T _ {c a n o}, T _ {c a n o} ^ {\\text {v i s}}\\right). \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.477, + 0.484, + 0.552 + ], + "angle": 0, + "content": "In this way, Fuser \\(F\\) merges new information in the input frame with occluded regions visible in the reference image. To train the visibility predictor \\(V\\), we calculate the visibility loss \\(L_{vis}\\) as the \\(L_{1}\\) distance between the predicted visibility triplanes versus the ground truth:" + }, + { + "type": "equation", + "bbox": [ + 0.114, + 0.559, + 0.483, + 0.579 + ], + "angle": 0, + "content": "\\[\nL _ {v i s} = L _ {1} (T _ {r a w} ^ {v i s}, T _ {r a w} ^ {v i s G T}) + L _ {1} (T _ {c a n o} ^ {v i s}, T _ {c a n o} ^ {v i s G T}). \\quad (6)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.587, + 0.484, + 0.646 + ], + "angle": 0, + "content": "The ground truth visibility triplanes contain 1 for pixels that are visible, and 0 otherwise. The generation of the ground truth visibility triplanes \\( T_{raw}^{visGT} \\) and \\( T_{cano}^{visGT} \\) is discussed in the supplementary." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.646, + 0.484, + 0.735 + ], + "angle": 0, + "content": "To supervise the Fuser \\(F\\), we calculate the fusion loss \\(L_{fusion}\\) as the \\(L_{1}\\) loss between the fused triplane \\(T_{fused}\\) and the pseudo-groundtruth triplane \\(T_{frontalGT}\\). To highlight the currently occluded region during training, we also upweight the occluded region using an occlusion mask \\(T_{occMask} \\in \\mathbb{R}^{3 \\times 1 \\times 256 \\times 256}\\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.23, + 0.744, + 0.482, + 0.761 + ], + "angle": 0, + "content": "\\[\nT _ {d i f f} = \\left| T _ {f u s e d} - T _ {t r i p l a n e G T} \\right|, \\tag {7}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.104, + 0.763, + 0.483, + 0.79 + ], + "angle": 0, + "content": "\\[\nL _ {f u s i o n} = \\operatorname {M e a n} \\left(T _ {d i f f}\\right) + \\frac {T _ {d i f f} \\cdot T _ {v i s G T}}{\\left| T _ {v i s G T} \\right|} + \\frac {T _ {d i f f} \\cdot T _ {o c c M a s k}}{\\left| T _ {o c c M a s k} \\right|}. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.798, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Please refer to the supplementary for the calculation of the occlusion mask. We use the Recurrent Video Restoration Transformer (RVRT) [29] as the backbone of our Fuser \\(F\\) because of its memory efficiency. We find that the final summation skip connection in RVRT prevents effective learning. This is because the original RVRT was designed to correct local blurriness and noise in a corrupted RGB video," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.164 + ], + "angle": 0, + "content": "whereas our triplane videos exhibit structural distortion on a much larger scale and the summation skip connection thus limits the model's ability to correct the general structure. We thus replace the summation with a small 5-layer ConvNet." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.165, + 0.906, + 0.239 + ], + "angle": 0, + "content": "Lastly, note that both the Undistorter \\( U \\) and the Fuser \\( F \\) consist of 3 separate copies, one for each of the 3 planes, because we find that processing all three planes jointly leads to collapse to 2D (please see supplementary for visualization and analysis)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.245, + 0.673, + 0.261 + ], + "angle": 0, + "content": "3.7. Training Losses" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.267, + 0.906, + 0.355 + ], + "angle": 0, + "content": "Our loss function is the summation of four loss terms that provide two types of supervision: (a) direct triplane space guidance used to supervise the undistortion process in the Undistorter \\(U\\), the visibility prediction process, and the fusion process in the Fuser \\(F\\); and (b) image space guidance for overall learning of high-quality image synthesis:" + }, + { + "type": "equation", + "bbox": [ + 0.577, + 0.356, + 0.905, + 0.389 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} L = w _ {u n d i s t} L _ {u n d i s t} + w _ {f u s i o n} L _ {f u s i o n} \\\\ + w _ {v i s} L _ {v i s} + w _ {r e n d e r} L _ {r e n d e r}, \\tag {9} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.395, + 0.906, + 0.454 + ], + "angle": 0, + "content": "where \\( w_{undist} \\), \\( w_{viz} \\), \\( w_{fusion} \\), and \\( w_{render} \\) are scalar weights for the different loss terms. \\( L_{render} \\) is calculated as the perceptual loss \\( L_{LPIPS} \\) between the ground truth novel view \\( I_{GT} \\) and the rendered novel view \\( I_{render} \\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.592, + 0.459, + 0.905, + 0.475 + ], + "angle": 0, + "content": "\\[\nL _ {\\text {r e n d e r}} = L _ {\\text {L P I P S}} \\left(I _ {G T}, I _ {\\text {r e n d e r}}\\right). \\tag {10}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.481, + 0.631, + 0.497 + ], + "angle": 0, + "content": "4. Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.506, + 0.906, + 0.844 + ], + "angle": 0, + "content": "Multi-view evaluation of single-view reconstruction. Due to the lack of 3D ground truth for in-the-wild images, prior methods are often evaluated on the input image reconstruction task, which compares the rendered image against the input image using metrics like PSNR, whereas the novel view synthesis task often relies on visual assessments. However, this practice can lead to ambiguities and inaccurate conclusions. For example, if a 3D reconstruction is only rendered and evaluated from the input viewpoint, then that reconstruction can overfit to the input view to achieve high scores even if it is inaccurate when rendered from other viewpoints (as proven quantitatively in Fig. 11 of the supplementary). Moreover, single-view reconstruction methods can be heavily affected by the choice of input viewpoints (Fig. 2). Therefore, both the input viewpoint and the novel viewpoint are crucial variables to consider when concluding the performance of a method. We thus propose a new evaluation framework that evaluates a model across every input-novel viewpoint combination. In this way, a method can only achieve high numerical performance when it consistently generates high-quality reconstructions regardless of the choice of input or novel viewpoint." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.85, + 0.697, + 0.865 + ], + "angle": 0, + "content": "4.1. Evaluation Dataset" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.872, + 0.906, + 0.901 + ], + "angle": 0, + "content": "We quantitatively evaluate various methods on the NeRSemble [23] dataset, which is a multiview portrait" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10716" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.297 + ], + "angle": 0, + "content": "video dataset that allows us to evaluate the methods using different input-novel viewpoint combinations. It is recorded with 16 calibrated time-synchronized cameras in a controlled studio environment. The images are captured at \\(7.1\\mathrm{MP}\\) resolution and 73 frames per second. There are 10 sequences in the test set, each capturing a different individual performing different expressions. One of the 10 test sequences involves severe facial occlusion from hair that causes most of the methods' face trackers to fail for significant portions of the recording for many of the viewpoints. We thus leave out that sequence because the results would not be a reliable assessment of quality. We also use 8 roughly evenly separated cameras out of all 16 cameras during evaluation." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.306, + 0.188, + 0.32 + ], + "angle": 0, + "content": "4.2. Metrics" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.327, + 0.483, + 0.504 + ], + "angle": 0, + "content": "Synthesis Quality. Given \\(N\\) views in the dataset, we evaluate a method's average performance across every input-novel viewpoint combination. More specifically, for each frame, each of the \\(N\\) cameras is used as the input viewpoint, producing \\(N\\) reconstructions in total. Then, each of the \\(N\\) reconstructions is rendered and evaluated on the \\(N\\) viewpoints, resulting in an \\(N \\times N\\) score matrix. As mentioned before, we use \\(N = 8\\) of the camera views from the NeRSemnble[23] dataset as input and novel views. Thus, for a test sequence of \\(T\\) frames, we can calculate a spatial-temporal score matrices \\(\\mathbf{S}^{T \\times 8 \\times 8}\\) for each of the metrics (as visualized in Fig. 11 in the supplementary):" + }, + { + "type": "equation", + "bbox": [ + 0.121, + 0.507, + 0.482, + 0.524 + ], + "angle": 0, + "content": "\\[\n\\mathbf {S} _ {t, i, j} = \\operatorname {M e t r i c} \\left(\\mathbf {I} _ {\\text {r e n d e r}} ^ {t, i, j}, \\mathbf {I} _ {G T} ^ {t, j}\\right), 1 \\leq i, j \\leq N, 1 \\leq t \\leq T, \\tag {11}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.284, + 0.526, + 0.482, + 0.542 + ], + "angle": 0, + "content": "\\[\ns = \\operatorname {M e a n} \\left(\\left\\{\\mathbf {S} _ {t, i, j} \\right\\}\\right), \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.554, + 0.483, + 0.672 + ], + "angle": 0, + "content": "where \\( Metric(\\cdot) \\) can be LPIPS, PSNR, etc.. \\( \\mathbf{I}_{\\text{render}}^{t,i,j} \\) is the image rendered using camera \\( i \\) as the input frame and camera \\( j \\) as the output rendering view at frame \\( t \\). \\( \\mathbf{I}_{GT}^{t,j} \\) is the ground truth frame captured by camera \\( j \\) at frame \\( t \\). The Synthesis Quality \\( s \\) is thus the average over all entries in S. For a dataset with multiple sequences, such as NeRSample, the final Synthesis Quality is the average across all sequences." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.672, + 0.483, + 0.732 + ], + "angle": 0, + "content": "Novel View Synthesis (NVS) Quality. Novel View Synthesis Quality \\( s_{NV} \\) is the average over all entries where the input view \\( i \\) is different from the novel view \\( j \\), essentially removing the diagonal entries from the score matrix:" + }, + { + "type": "equation", + "bbox": [ + 0.145, + 0.742, + 0.482, + 0.759 + ], + "angle": 0, + "content": "\\[\ns _ {N V} = \\operatorname {M e a n} \\left(\\left\\{\\mathbf {S} _ {t, i, j} \\mid i \\neq j, 1 \\leq t \\leq T \\right\\}\\right). \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.768, + 0.483, + 0.814 + ], + "angle": 0, + "content": "Identity Accuracy (ID) We measure the ArcFace [7] cosine distance between the ground truth image \\( I_{GT} \\) and rendered image \\( I_{render} \\) both from the frontal camera." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.813, + 0.483, + 0.871 + ], + "angle": 0, + "content": "Expression Accuracy (Expr) We use NVIDIA Maxine AR SDK [35] to measure the \\( L_{2} \\) distance between expression coefficients \\( e_{GT} \\) of the ground truth image and \\( e_{render} \\) of the rendered image both from the frontal camera." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.872, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Dynamic Appearance While we believe that it is important to measure the accuracy of dynamic appearance such" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.166 + ], + "angle": 0, + "content": "as lighting and shoulder poses, there is no existing multiview in-the-wild portrait video dataset to support such evaluation. We thus qualitatively evaluate various methods on challenging in-the-wild portrait videos. Please see the supplementary material for image examples and video results." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.175, + 0.655, + 0.192 + ], + "angle": 0, + "content": "4.3. Comparisons" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.197, + 0.907, + 0.652 + ], + "angle": 0, + "content": "Baselines. We compare recent methods from 3 categories: (i) Reconstruction: We evaluate LP3D [52] using the above protocol. LP3D generates \\(N\\) 3D portraits using each of the \\(N\\) viewpoints as input, and the \\(N\\) 3D portraits are then evaluated on the \\(N\\) ground truth viewpoints. (ii) Reenactment: We evaluate Li et al. [28] and GPAvatar [6] in the self-reenactment setting. Li et al. [28] reconstruct 3D portraits as triplanes from reference images without test-time optimization, and they drive the 3D portraits via dynamic frontal renderings of 3DMMs. The authors of Li et al. [28] evaluated their approach using the input views as the only novel view, instead of all \\(N\\) views. GPAvatar[6] reconstructs 3D portraits by leveraging multiple source images and driving them through a FLAME [26] mesh model. We use the first frame of the frontal camera in each NeRSample test sequence as the reference image to generate the 3D portrait, and drive it using videos from each of the \\(N\\) viewpoints. We evaluate GPAvatar using the same evaluation protocol as our method and LP3D. (iii) Inversion: We evaluate VIVE3D [13], which is a state-of-the-art 3D GAN inversion method for videos, and it can also perform semantic video editing. We first provide VIVE3D 3 video frames from the input viewpoint, which it requires for personalization. Then, we invert the input video frames into 3D portraits and render the 3D portraits from all \\(N\\) viewpoints. Each of the above method uses a different cropping of the face. We standardize the evaluation by re-cropping all methods to our cropping protocol, which is the largest of all. Please see the supplement for an additional table, where we evaluate the methods using different cropings around the face and arrive at conclusions consistent with Table 1." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.652, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Quantitative results. We evaluate various methods using different input-novel viewpoint combinations, providing a robust multi-view assessment of a model's performance. Table. 1 shows that our model achieves state-of-the-art performance across all metrics versus recent works. Notably, while LP3D overfits to the input viewpoint (as shown by the difference of \\(0.806\\mathrm{dB}\\) PSNR between Synthesis Quality and NVS Quality), our method generalizes better to novel viewpoints (as shown by the difference of \\(0.330\\mathrm{dB}\\)). Moreover, our method best preserves subject identity and expression (Table. 1, Fig. 2, and 4). On the other hand, the reenactment methods struggle to capture authentic expressions because of the use of morphable face models, which have limited expressiveness and accuracy. Moreover, they cannot reconstruct dynamic appearance (e.g. the stuck-out tongue in the second and third rows of Fig. 4) because they solely rely on information present in the reference image(s) and" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10717" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.125, + 0.092, + 0.2, + 0.104 + ], + "angle": 0, + "content": "Groundtruth" + }, + { + "type": "image_caption", + "bbox": [ + 0.28, + 0.092, + 0.312, + 0.103 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.412, + 0.092, + 0.45, + 0.103 + ], + "angle": 0, + "content": "LP3D" + }, + { + "type": "image_caption", + "bbox": [ + 0.537, + 0.092, + 0.597, + 0.103 + ], + "angle": 0, + "content": "GPAvatar" + }, + { + "type": "image_caption", + "bbox": [ + 0.634, + 0.092, + 0.766, + 0.105 + ], + "angle": 0, + "content": "Input / Driving Frame" + }, + { + "type": "image_caption", + "bbox": [ + 0.778, + 0.092, + 0.892, + 0.104 + ], + "angle": 0, + "content": "Reference / Source" + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.106, + 0.904, + 0.522 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.534, + 0.907, + 0.59 + ], + "angle": 0, + "content": "Figure 4. Visual comparisons with baseline methods. Our method strikes a balance between coherent reconstruction and faithfully preserving dynamic conditions like expressions. LP3D (third column) exhibits inconsistency in identities, hairstyles, and artifacts (red circles). GPAvatar (fourth column) fails to capture challenging expressions (first row), new information not present in the reference image, (the stuck-out tongue in second and third rows), and the identity of the person (last row)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.603, + 0.483, + 0.647 + ], + "angle": 0, + "content": "do not incorporate new per-frame information. On the other hand, our method achieves faithful dynamic appearance and coherent reconstruction at the same time." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.673, + 0.483, + 0.796 + ], + "angle": 0, + "content": "
MethodTypeExpr↓ID↓SynthesisLIPS↓LVS Quality
PSNR↑LPIPS↓PSNR↑LPIPS↓
Li et al. [28]reenact0.2660.24118.5730.25518.2020.262
GPAvatar[6]reenact0.2040.20721.9490.23321.9490.2334
VIVE3D[13]invert0.2900.39518.5770.25918.1450.271
LP3D[52]recon0.1680.21522.3310.22321.5250.237
Oursrecon0.1580.18722.7700.21922.4400.224
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.806, + 0.483, + 0.874 + ], + "angle": 0, + "content": "Table 1. Comparisons on Nersemble [23]. Our evaluation protocol (Sec. 4.2) utilizes multiview groundtruth for thorough analysis. Our method achieves state-of-the-art performance across all metrics, maintains temporal coherence, and preserves accurate expressions." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.602, + 0.625, + 0.616 + ], + "angle": 0, + "content": "4.4. Ablations" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.623, + 0.907, + 0.711 + ], + "angle": 0, + "content": "We evaluate the effectiveness of our proposed Triplane Undistorter and Fuser modules on NeRSemble. In addition to the Synthesis Quality in terms of PSNR and LPIPS, we develop two new metrics to separately measure the robustness to variations in (1) rendering viewpoints and (2) input viewpoints:" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.711, + 0.907, + 0.8 + ], + "angle": 0, + "content": "(1) Novel View Variation (NVV). We evaluate how much a method's reconstruction quality varies across different novel views. We quantify this as the standard deviation of performance across the \\(N\\) novel views when using the same input view, i.e. the average standard deviation of each horizontal row of the score matrix \\(S\\):" + }, + { + "type": "equation", + "bbox": [ + 0.52, + 0.807, + 0.907, + 0.849 + ], + "angle": 0, + "content": "\\[\n\\mathrm {N V V} = \\frac {1}{T N} \\sum_ {t = 1} ^ {T} \\sum_ {i = 1} ^ {N} \\operatorname {s t d d e v} \\left(\\left\\{S _ {t, i, j} \\mid 1 \\leq j \\leq N \\right\\}\\right). \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.857, + 0.907, + 0.902 + ], + "angle": 0, + "content": "(2) Input View Variation (IVV). We measure how much a method's reconstruction quality varies when using the \\(N\\) different input viewpoints (Sec. 4.2 second column from the" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10718" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.093, + 0.089, + 0.487, + 0.196 + ], + "angle": 0, + "content": "
MethodU FSynthesis QualityIVV↓NVV↓
PSNR↑LPIPS↓(PSNR)(LPIPS)(PSNR)(LPIPS)
LP3D[52]XX22.3310.2231.0250.0152.2000.053
Ours✓X22.1960.2210.9070.0091.6990.038
X✓22.2650.2230.5590.0061.3150.029
✓✓22.7700.2190.2450.0051.3830.037
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.205, + 0.483, + 0.273 + ], + "angle": 0, + "content": "Table 2. Undistorter and Fuser Ablations. The Undistorter-only (row 2) and Fuser-only (row 3) variants both lead to improvements in viewpoint robustness metrics (IVV and NVV), but not in Synthesis Quality. The best performance is achieved by using both the Undistorter and the Fuser at the same time." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.288, + 0.484, + 0.348 + ], + "angle": 0, + "content": "right). We quantify this variation as the average standard deviation of performance on the same novel view when using different input views, i.e. the average standard deviation of each vertical column of the score matrix \\( S \\):" + }, + { + "type": "equation", + "bbox": [ + 0.103, + 0.352, + 0.483, + 0.395 + ], + "angle": 0, + "content": "\\[\n\\mathrm {I V V} = \\frac {1}{T N} \\sum_ {t = 1} ^ {T} \\sum_ {j = 1} ^ {N} \\operatorname {s t d d e v} \\left(\\left\\{S _ {t, i, j} \\mid 1 \\leq i \\leq N \\right\\}\\right). \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.402, + 0.483, + 0.431 + ], + "angle": 0, + "content": "\\(NVV\\) and \\(IVV\\) are measured for both PSNR and LPIPS, and they are better if lower." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.432, + 0.483, + 0.768 + ], + "angle": 0, + "content": "As shown in Table. 2, the Undistorter module consistently improves the NVV and IVV metrics versus LP3D, indicating better robustness to different input viewpoints and more consistent rendering quality across novel rendered views. However, when only the Undistorter is added (Tab. 2 second row) the PSNR is slightly reduced. This is likely because this model does not leverage the reference image to recover occluded areas while also overfitting to the input view. Similarly, the Fuser-only variant (Tab. 2 third row) achieves better NVV and IVV scores, but also scores lower in terms of PSNR. A likely cause is that, without the Undistorter, the Fuser needs to merge highly misaligned triplanes, where the person looks drastically different in \\( T_{raw} \\) and \\( T_{cano} \\). This misalignment possibly induces more blurriness and alignment artifacts that lower the PSNR. The best performance is achieved by our full model, which includes both the Undistorter and the Fuser and thus achieves better PSNR, LPIPS, IVV and NVV scores. This means that the Undistorter and the Fuser complement each other: The Undistorter corrects the distortion in the raw triplane and thus reduces the challenges in fusing misaligned triplanes, and the Fuser recovers the occluded areas in the raw triplane." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.769, + 0.483, + 0.814 + ], + "angle": 0, + "content": "Shoulder Augmentation. without shoulder augmentation, the model fails to correctly reconstruct varying shoulder poses, resulting in worse performance (Fig. 5 and Table. 3)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.814, + 0.483, + 0.871 + ], + "angle": 0, + "content": "Optical Flow. As mentioned in Sec. 3.5, replacing the Undistorter with optical flow worsens the reconstruction (Table. 3). As shown by Fig. 5, this variant fails to correctly reconstruct the shoulder and the mouth accurately." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.901 + ], + "angle": 0, + "content": "GAN Training. We retrained our model with an additional adversarial loss incorporated into Eqn. 10. As expected, the" + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.091, + 0.903, + 0.199 + ], + "angle": 0, + "content": null + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.21, + 0.906, + 0.25 + ], + "angle": 0, + "content": "Figure 5. Visual ablation. Our method with optical flow and w/o shoulder augmentation on two different input frames (top and bottom rows)." + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.25, + 0.905, + 0.372 + ], + "angle": 0, + "content": "
MethodLPIPS↓PSNR↑IVV (PSNR) ↓ID↓Expr↓
LP3D (orig.)0.33418.7212.1300.2470.451
LP3D (ours)0.22322.3311.0250.1680.215
w optical flow0.22722.0851.1750.1780.335
w/o shoulder aug.0.21822.3420.8290.1530.244
Ours0.21922.7700.2450.1580.187
Ours + GAN0.20722.0740.3140.1520.198
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.376, + 0.906, + 0.445 + ], + "angle": 0, + "content": "Table 3. Other Ablations. We compare our implementation of LP3D to the original one [52] (row 1&2); the effect of replacing our Undistorter with optical flow warping (row 4); the effect of removing shoulder augmentation during training (row 4); and the effect of adding an adversarial GAN loss (row 6) alongside Eqn. 10." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.462, + 0.905, + 0.507 + ], + "angle": 0, + "content": "GAN training (Table. 3) improves the sharpness of images as seen from the LPIPS error reduction from 0.219 to 0.207. Additional results are included in the supplementary." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.508, + 0.906, + 0.596 + ], + "angle": 0, + "content": "Qualitative results. In-the-wild experiments show that our method achieves better temporal consistency than LP3D [52] and more accurately captures dynamic information like expressions and lighting changes than GPAvatar [6]. We refer the readers to our supplementary materials for visual examples of these results." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.611, + 0.628, + 0.627 + ], + "angle": 0, + "content": "5. Discussion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.637, + 0.906, + 0.754 + ], + "angle": 0, + "content": "Conclusion. Recognizing the individual limitations of per-frame single-view reconstruction and 3D reenactment methods, we presented the first single-view 3D lifting method to reconstruct a 3D photorealistic avatar with faithful dynamic appearance as well as temporal consistency, marrying the best of both worlds. We believe our method paves the way forward for creating a high-quality telepresence system accessible to consumers." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.906, + 0.887 + ], + "angle": 0, + "content": "Limitations and future work. We use a single reference image, but incorporating multiple ones with different expressions and head poses could lead to further improvements. With our method, fusing an extreme side view with a very different expression to the reference view may result in blurry reconstruction due to ambiguity in triplane alignment. Finally, due to the additional components, our current run-time performance is slower than real time, which could be improved in future work." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10719" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.168 + ], + "angle": 0, + "content": "[1] ShahRukh Athar, Zexiang Xu, Kalyan Sunkavalli, Eli Shechtman, and Zhixin Shu. Rignerf: Fully controllable neural 3d portraits. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.171, + 0.484, + 0.21 + ], + "angle": 0, + "content": "[2] Ananta R. Bhattarai, Matthias Nießner, and Artem Sevastopolsky. Triplanenet: An encoder for eg3d inversion. 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.213, + 0.484, + 0.306 + ], + "angle": 0, + "content": "[3] Marcel C. Buehler, Kripasindhu Sarkar, Tanmay Shah, Gengyan Li, Daoye Wang, Leonhard Helminger, Sergio Orts-Escolano, Dmitry Lagun, Otmar Hilliges, Thabo Beeler, and Abhimitra Meka. Preface: A data-driven volumetric prior for few-shot ultra high-resolution face synthesis. In IEEE International Conference on Computer Vision (ICCV), 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.309, + 0.484, + 0.387 + ], + "angle": 0, + "content": "[4] Chen Cao, Tomas Simon, Jin Kyu Kim, Gabe Schwartz, Michael Zollhoefer, Shun-Suke Saito, Stephen Lombardi, Shih-En Wei, Danielle Belko, Shouou-I Yu, Yaser Sheikh, and Jason Saragih. Authentic volumetric avatars from a phone scan. ACM Transactions on Graphics (SIGGRAPH), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.391, + 0.484, + 0.471 + ], + "angle": 0, + "content": "[5] Eric R. Chan, Connor Z. Lin, Matthew A. Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas Guibas, Jonathan Tremblay, Sameh Khamis, Tero Karras, and Gordon Wetzstein. Efficient geometry-aware 3D generative adversarial networks. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.474, + 0.484, + 0.539 + ], + "angle": 0, + "content": "[6] Xuangeng Chu, Yu Li, Ailing Zeng, Tianyu Yang, Lijian Lin, Yunfei Liu, and Tatsuya Harada. Gpavatar: Generalizable and precise head avatar from image(s). In International Conference on Learning Representations (ICLR), 2024. 2, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.542, + 0.484, + 0.595 + ], + "angle": 0, + "content": "[7] Jiankang Deng, Jia Guo, Xue Niannan, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.598, + 0.484, + 0.651 + ], + "angle": 0, + "content": "[8] Yu Deng, Jiaolong Yang, Jianfeng Xiang, and Xin Tong. Gram: Generative radiance manifolds for 3d-aware image generation. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.653, + 0.484, + 0.706 + ], + "angle": 0, + "content": "[9] Michail Christos Doukas, Stefanos Zafeiriou, and Viktoriia Sharmanska. Headgan: One-shot neural head synthesis and editing. In IEEE International Conference on Computer Vision (ICCV), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.708, + 0.483, + 0.762 + ], + "angle": 0, + "content": "[10] Nikita Drobyshev, Jenya Chelishev, Taras Khakhulin, Aleksei Ivakhnenko, Victor Lempitsky, and Egor Zakharov. Megaportraits: One-shot megapixel neural head avatars. arXiv preprint arXiv:2207.07621, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.764, + 0.483, + 0.804 + ], + "angle": 0, + "content": "[11] Yao Feng, Haiwen Feng, Michael J. Black, and Timo Bolkart. Learning an animatable detailed 3D face model from in-the-wild images. 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.806, + 0.483, + 0.858 + ], + "angle": 0, + "content": "[12] Yao Feng, Haiwen Feng, Michael J Black, and Timo Bolkart. Learning an animatable detailed 3d face model from in-the-wild images. ACM Transactions on Graphics (ToG), 40(4): 1-13, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.861, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[13] Anna Fruhstück, Nikolaos Sarafianos, Yuanlu Xu, Peter Wonka, and Tony Tung. VIVE3D: Viewpoint-independent video editing using 3D-Aware GANs. In IEEE Conference" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.548, + 0.094, + 0.905, + 0.119 + ], + "angle": 0, + "content": "on Computer Vision and Pattern Recognition (CVPR), 2023. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.122, + 0.905, + 0.175 + ], + "angle": 0, + "content": "[14] Guy Gafni, Justus Thies, Michael Zollhöfer, and Matthias Nießner. Dynamic neural radiance fields for monocular 4d facial avatar reconstruction. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.177, + 0.905, + 0.229 + ], + "angle": 0, + "content": "[15] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In Advances in Neural Information Processing Systems (NeurIPS), 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.231, + 0.905, + 0.284 + ], + "angle": 0, + "content": "[16] Jiatao Gu, Lingjie Liu, Peng Wang, and Christian Theobalt. StyleNeRF: A style-based 3D-aware generator for high-resolution image synthesis. arXiv preprint arXiv:2110.08985, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.286, + 0.905, + 0.326 + ], + "angle": 0, + "content": "[17] Fa-Ting Hong, Longhao Zhang, Li Shen, and Dan Xu. Depth-aware generative adversarial network for talking head video generation. 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.328, + 0.905, + 0.38 + ], + "angle": 0, + "content": "[18] Yang Hong, Bo Peng, Haiyao Xiao, Ligang Liu, and Juyong Zhang. Headnerf: A real-time nerf-based parametric head model. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.382, + 0.905, + 0.449 + ], + "angle": 0, + "content": "[19] Andrew Jones, Magnus Lang, Graham Fyffe, Xueming Yu, Jay Busch, Ian McDowall, Mark Bolas, and Paul Debevec. Achieving eye contact in a one-to-many 3d video teleconferencing system. ACM Transactions on Graphics (SIGGRAPH), 2009. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.451, + 0.905, + 0.503 + ], + "angle": 0, + "content": "[20] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.505, + 0.905, + 0.558 + ], + "angle": 0, + "content": "[21] Peter Kauff and Oliver Schreer. An immersive 3d videoconferencing system using shared virtual team user environments. In Proceedings of the 4th International Conference on Collaborative Virtual Environments, 2002. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.56, + 0.905, + 0.613 + ], + "angle": 0, + "content": "[22] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics (SIGGRAPH), 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.615, + 0.905, + 0.668 + ], + "angle": 0, + "content": "[23] Tobias Kirschstein, Shenhan Qian, Simon Giebenhain, Tim Walter, and Matthias Nießner. Nersemble: Multi-view radiance field reconstruction of human heads. ACM Trans. Graph., 2023. 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.67, + 0.905, + 0.723 + ], + "angle": 0, + "content": "[24] Jaehoon Ko, Kyusun Cho, Daewon Choi, Kwangrok Ryoo, and Seungryong Kim. 3d gan inversion with pose optimization. In IEEE Winter Conference on Applications of Computer Vision (WACV), 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.725, + 0.905, + 0.831 + ], + "angle": 0, + "content": "[25] Jason Lawrence, Dan B Goldman, Supreeth Achar, Gregory Major Blascovich, Joseph G. Desloge, Tommy Fortes, Eric M. Gomez, Sascha Haberling, Hugues Hoppe, Andy Huibers, Claude Knaus, Brian Kuschak, Ricardo Martin-Brualla, Harris Nover, Andrew Ian Russell, Steven M. Seitz, and Kevin Tong. Project starline: A high-fidelity telepresence system. ACM Transactions on Graphics (SIGGRAPH ASIA), 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.833, + 0.905, + 0.872 + ], + "angle": 0, + "content": "[26] Tianye Li, Timo Bolkart, Michael. J. Black, Hao Li, and Javier Romero. Learning a model of facial shape and expression from 4D scans. 2017. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.874, + 0.905, + 0.902 + ], + "angle": 0, + "content": "[27] Weichuang Li, Longhao Zhang, Dong Wang, Bin Zhao, Zhi-gang Wang, Mulin Chen, Bang Zhang, Zhongjian Wang," + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.094, + 0.905, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10720" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.146 + ], + "angle": 0, + "content": "Liefeng Bo, and Xuelong Li. One-shot high-fidelity talking-head synthesis with deformable neural radiance field. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.147, + 0.483, + 0.186 + ], + "angle": 0, + "content": "[28] Xueting Li, Shalini De Mello, Sifei Liu, Koki Nagano, Umar Iqbal, and Jan Kautz. Generalizable one-shot neural head avatar. 2023. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.187, + 0.482, + 0.242 + ], + "angle": 0, + "content": "[29] Jingyun Liang, Yuchen Fan, Xiaoyu Xiang, Rakesh Ranjan, Eddy Ilg, Simon Green, Jiezhang Cao, Kai Zhang, Radu Timofte, and Luc Van Gool. Recurrent video restoration transformer with guided deformable attention. 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.242, + 0.483, + 0.294 + ], + "angle": 0, + "content": "[30] C.Z. Lin, D.B. Lindell, E.R. Chan, and G. Wetzstein. 3d gan inversion for controllable portrait image animation. In ECCV Workshop on Learning to Generate 3D Shapes and Scenes, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.295, + 0.482, + 0.336 + ], + "angle": 0, + "content": "[31] S. Ma, T. Simon, J. Saragih, D. Wang, Y. Li, F. La Torre, and Y. Sheikh. Pixel codec avatars. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.336, + 0.482, + 0.376 + ], + "angle": 0, + "content": "[32] Zhiyuan Ma, Xiangyu Zhu, Guojun Qi, Zhen Lei, and Lei Zhang. Otavatar: One-shot talking face avatar with controllable tri-plane rendering. 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.377, + 0.482, + 0.429 + ], + "angle": 0, + "content": "[33] Andrew Maimone, Jonathan Bidwell, Kun Peng, and Henry Fuchs. Enhanced personal autostereoscopic telepresence system using commodity depth cameras. Computers & Graphics, 2012. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.43, + 0.482, + 0.496 + ], + "angle": 0, + "content": "[34] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing scenes as neural radiance fields for view synthesis. In European Conference on Computer Vision (ECCV), 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.497, + 0.479, + 0.523 + ], + "angle": 0, + "content": "[35] NVIDIA. Nvidia maxine ar sdk. https://github.com/NVIDIA/MAXINE-AR-SDK, 2024.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.525, + 0.482, + 0.59 + ], + "angle": 0, + "content": "[36] Roy Or-El, Xuan Luo, Mengyi Shan, Eli Shechtman, Jeong Joon Park, and Ira Kemelmacher-Shlizerman. StyleSDF: High-Resolution 3D-Consistent Image and Geometry Generation. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.591, + 0.482, + 0.698 + ], + "angle": 0, + "content": "[37] Sergio Orts-Escalano, Christoph Rhemann, Sean Fanello, Wayne Chang, Adarsh Kowdle, Yury Degtyarev, David Kim, Philip L. Davidson, Sameh Khamis, Mingsong Dou, Vladimir Tankovich, Charles Loop, Qin Cai, Philip A. Chou, Sarah Mennicken, Julien Valentin, Vivek Pradeep, Shenlong Wang, Sing Bing Kang, Pushmeet Kohli, Yuliya Lutchyn, Cem Keskin, and Shahram Izadi. Holoportation: Virtual 3d teleportation in real-time. In UIST 2016, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.699, + 0.482, + 0.751 + ], + "angle": 0, + "content": "[38] Keunhong Park, Utkarsh Sinha, Jonathan T. Barron, Sofien Bouaziz, Dan B Goldman, Steven M. Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.753, + 0.482, + 0.806 + ], + "angle": 0, + "content": "[39] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-NeRF: Neural Radiance Fields for Dynamic Scenes. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.807, + 0.482, + 0.846 + ], + "angle": 0, + "content": "[40] Luchao Qi, Jiaye Wu, Annie N. Wang, Shengze Wang, and Roni Sengupta. My3dgen: A scalable personalized 3d generative model, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.847, + 0.482, + 0.899 + ], + "angle": 0, + "content": "[41] Shenhan Qian, Tobias Kirschstein, Liam Schoneveld, Davide Davoli, Simon Giebenhain, and Matthias Nießner. Gaussianavatars: Photorealistic head avatars with rigged 3d gaussians. 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.146 + ], + "angle": 0, + "content": "[42] Anurag Ranjan and Michael J Black. Optical flow estimation using a spatial pyramid network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4161-4170, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.147, + 0.905, + 0.213 + ], + "angle": 0, + "content": "[43] Ramesh Raskar, Greg Welch, Matt Cutts, Adam Lake, Lev Stesin, and Henry Fuchs. The office of the future: a unified approach to image-based modeling and spatially immersive displays. In ACM Transactions on Graphics (SIGGRAPH), 1998. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.215, + 0.905, + 0.268 + ], + "angle": 0, + "content": "[44] Daniel Rebain, Mark Matthews, Kwang Moo Yi, Dmitry Lagan, and Andrea Tagliasacchi. Lolnerf: Learn from one look. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.269, + 0.905, + 0.322 + ], + "angle": 0, + "content": "[45] Shunsuke Saito, Gabriel Schwartz, Tomas Simon, Junxuan Li, and Giljoo Nam. Relightable gaussian codec avatars. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.322, + 0.905, + 0.376 + ], + "angle": 0, + "content": "[46] Aliaksandr Siarohin, Stéphane Lathuilière, Sergey Tulyakov, Elisa Ricci, and Nicu Sebe. First order motion model for image animation. In Advances in Neural Information Processing Systems (NeurIPS), 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.376, + 0.905, + 0.428 + ], + "angle": 0, + "content": "[47] Ivan Skorokhodov, Sergey Tulyakov, Yiqun Wang, and Peter Wonka. Epigraph: Rethinking training of 3d gans. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.429, + 0.905, + 0.483 + ], + "angle": 0, + "content": "[48] Michael Stengel, Koki Nagano, Chao Liu, Matthew Chan, Alex Trevithick, Shalini De Mello, Jonghyun Kim, and David Luebke. AI-Mediated 3D Video Conferencing. In ACM SIGGRAPH 2023 Emerging Technologies, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.484, + 0.905, + 0.537 + ], + "angle": 0, + "content": "[49] Jingxiang Sun, Xuan Wang, Yichun Shi, Lizhen Wang, Jue Wang, and Yebin Liu. Ide-3d: Interactive disentangled editing for high-resolution 3d-aware portrait synthesis. ACM Transactions on Graphics (SIGGRAPH ASIA), 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.537, + 0.905, + 0.604 + ], + "angle": 0, + "content": "[50] Jingxiang Sun, Xuan Wang, Lizhen Wang, Xiaoyu Li, Yong Zhang, Hongwen Zhang, and Yebin Liu. Next3d: Generative neural texture rasterization for 3d-aware head avatars. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.605, + 0.905, + 0.657 + ], + "angle": 0, + "content": "[51] Phong Tran, Egor Zakharov, Long-Nhat Ho, Anh Tuan Tran, Liwen Hu, and Hao Li. Voodoo 3d: Volumetric portrait disentanglement for one-shot 3d head reenactment. 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.659, + 0.905, + 0.738 + ], + "angle": 0, + "content": "[52] Alex Trevithick, Matthew Chan, Michael Stengel, Eric R. Chan, Chao Liu, Zhiding Yu, Sameh Khamis, Manmohan Chandraker, Ravi Ramamoorthi, and Koki Nagano. Realtime radiance fields for single-image portrait view synthesis. In ACM Transactions on Graphics (SIGGRAPH), 2023. 1, 2, 3, 4, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.739, + 0.905, + 0.793 + ], + "angle": 0, + "content": "[53] Ting-Chun Wang, Arun Mallya, and Ming-Yu Liu. One-shot free-view neural talking-head synthesis for video conferencing. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.794, + 0.905, + 0.846 + ], + "angle": 0, + "content": "[54] Yaohui Wang, Di Yang, Francois Bremond, and Antitza Dantcheva. Latent image animator: Learning to animate images via latent space navigation. In International Conference on Learning Representations (ICLR), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.847, + 0.905, + 0.899 + ], + "angle": 0, + "content": "[55] Jianfeng Xiang, Jiaolong Yang, Yu Deng, and Xin Tong. Gram-hd: 3d-consistent image generation at high resolution with generative radiance manifolds. arXiv preprint arXiv:2206.07255, 2022. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "10721" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.146 + ], + "angle": 0, + "content": "[56] Yinghao Xu, Sida Peng, Ceyuan Yang, Yujun Shen, and Bolei Zhou. 3d-aware image synthesis via learning structural and textural representations. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.148, + 0.482, + 0.215 + ], + "angle": 0, + "content": "[57] Yuelang Xu, Benwang Chen, Zhe Li, Hongwen Zhang, Lizhen Wang, Zerong Zheng, and Yebin Liu. Gaussian head avatar: Ultra high-fidelity head avatar via dynamic gaussians. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.217, + 0.482, + 0.284 + ], + "angle": 0, + "content": "[58] Zhongcong Xu, Jianfeng Zhang, Junhao Liew, Wenqing Zhang, Song Bai, Jiashi Feng, and Mike Zheng Shou. Pv3d: A 3d generative model for portrait video generation. In The Tenth International Conference on Learning Representations, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.285, + 0.482, + 0.365 + ], + "angle": 0, + "content": "[59] Zhenhui Ye, Tianyun Zhong, Yi Ren, Jiaqi Yang, Weichuang Li, Jiangwei Huang, Ziyue Jiang, Jinzheng He, Rongjie Huang, Jinglin Liu, Chen Zhang, Xiang Yin, Zejun Ma, and Zhou Zhao. Real3d-portrait: One-shot realistic 3d talking portrait synthesis. In International Conference on Learning Representations (ICLR), 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.367, + 0.482, + 0.42 + ], + "angle": 0, + "content": "[60] Fei Yin, Yong Zhang, Xiaodong Cun, Mingdeng Cao, Yanbo Fan, Xuan Wang, Qingyan Bai, Baoyuan Wu, Jue Wang, and Yujiu Yang. Styleheat: One-shot high-resolution editable talking face generation via pre-trained stylegan. 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.422, + 0.482, + 0.489 + ], + "angle": 0, + "content": "[61] Wangbo Yu, Yanbo Fan, Yong Zhang, Xuan Wang, Fei Yin, Yunpeng Bai, Yan-Pei Cao, Ying Shan, Yang Wu, Zhongqian Sun, and Baoyuan Wu. Nofa: Nerf-based one-shot facial avatar reconstruction. In ACM SIGGRAPH 2023 Conference Proceedings, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.49, + 0.482, + 0.544 + ], + "angle": 0, + "content": "[62] Egor Zakharov, Aliaksandra Shysheya, Egor Burkov, and Victor Lempitsky. Few-shot adversarial learning of realistic neural talking head models. In IEEE International Conference on Computer Vision (ICCV), 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.545, + 0.482, + 0.599 + ], + "angle": 0, + "content": "[63] Egor Zakharov, Aleksei Ivakhnenko, Aliaksandra Shysheya, and Victor Lempitsky. Fast bi-layer neural synthesis of one-shot realistic head avatars. In European Conference on Computer Vision (ECCV), 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.6, + 0.482, + 0.654 + ], + "angle": 0, + "content": "[64] Bowen Zhang, Chenyang Qi, Pan Zhang, Bo Zhang, Hsiang-Tao Wu, Dong Chen, Qifeng Chen, Yong Wang, and Fang Wen. Metaportrait: Identity-preserving talking head generation with fast personalized adaptation, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.655, + 0.482, + 0.722 + ], + "angle": 0, + "content": "[65] Xuanmeng Zhang, Zhedong Zheng, Daiheng Gao, Bang Zhang, Pan Pan, and Yi Yang. Multi-view consistent generative adversarial networks for 3d-aware image synthesis. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.723, + 0.482, + 0.776 + ], + "angle": 0, + "content": "[66] Jian Zhao and Hui Zhang. Thin-plate spline motion model for image animation. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 3657-3666, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.778, + 0.482, + 0.845 + ], + "angle": 0, + "content": "[67] Yufeng Zheng, Victoria Fernández Abrevaya, Marcel C. Bühler, Xu Chen, Michael J. Black, and Otmar Hilliges. I M Avatar: Implicit morphable head avatars from videos. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.847, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[68] Peng Zhou, Lingxi Xie, Bingbing Ni, and Qi Tian. CIPS-3D: A 3D-Aware Generator of GANs Based on Conditionally-Independent Pixel Synthesis. arXiv preprint arXiv:2110.09788, 2021. 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.134 + ], + "angle": 0, + "content": "[69] Yiyu Zhuang, Hao Zhu, Xusen Sun, and Xun Cao. Mofanerf: Morphable facial neural radiance field. In European Conference on Computer Vision (ECCV), 2022. 3" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10722" + } + ] +] \ No newline at end of file diff --git a/2025/Coherent 3D Portrait Video Reconstruction via Triplane Fusion/2088b6b2-40e1-4341-a0c5-c5303c3acd65_origin.pdf b/2025/Coherent 3D Portrait Video Reconstruction via Triplane Fusion/2088b6b2-40e1-4341-a0c5-c5303c3acd65_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..777c678114926ca49f041d7a6f29dd6caa72b22d --- /dev/null +++ b/2025/Coherent 3D Portrait Video Reconstruction via Triplane Fusion/2088b6b2-40e1-4341-a0c5-c5303c3acd65_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cdedf2426b27493154afc38bb436bc2ed5104ee8d1de4ef40ea5edf8fc7ed38b +size 8065609 diff --git a/2025/Coherent 3D Portrait Video Reconstruction via Triplane Fusion/full.md b/2025/Coherent 3D Portrait Video Reconstruction via Triplane Fusion/full.md new file mode 100644 index 0000000000000000000000000000000000000000..906582ca34a70c1dad1effd16530832e00098129 --- /dev/null +++ b/2025/Coherent 3D Portrait Video Reconstruction via Triplane Fusion/full.md @@ -0,0 +1,338 @@ +# Coherent 3D Portrait Video Reconstruction via Triplane Fusion + +Shengze Wang $^{1*}$ Xueting Li $^{2}$ Chao Liu $^{2}$ Matthew Chan $^{2}$ Michael Stengel $^{2}$ Henry Fuchs $^{1}$ Shalini De Mello $^{2\dagger}$ Koki Nagano $^{2\dagger}$ $^{1}$ UNC Chapel Hill ${}^{2}$ NVIDIA $\dagger$ Equal Contribution +https://research.nvidia.com/labs/amri/projects/coherent3d/ + +# Abstract + +Recent breakthroughs in single-image 3D portrait reconstruction have enabled telepresence systems to stream 3D portrait videos from a single camera in real-time, democratizing telepresence. However, per-frame 3D reconstruction exhibits temporal inconsistency and forgets the user's appearance. On the other hand, self-reenactment methods can render coherent 3D portraits by driving a 3D avatar built from a single reference image but fail to faithfully preserve the user's per-frame appearance (e.g., instantaneous facial expressions and lighting). As a result, neither of these two frameworks is an ideal solution for democratized 3D telepresence. In this work, we address this dilemma and propose a novel solution that maintains both coherent identity and dynamic per-frame appearance to enable the best possible realism. To this end, we propose a new fusion-based method that takes the best of both worlds by fusing a canonical 3D prior from a reference view with dynamic appearance from per-frame input views, producing temporally stable 3D videos with faithful reconstruction of the user's per-frame appearance. Trained only using synthetic data produced by an expression-conditioned 3D GAN, our encoder-based method achieves both state-of-the-art 3D reconstruction and temporal consistency on in-studio and in-the-wild datasets. + +# 1. Introduction + +Telepresence aims at bringing distant people face-to-face and stands out as a particularly compelling application of computer vision and graphics. Over the last decades, various successful telepresence systems [19, 21, 21, 25, 31, 33, 37, 43] have been developed. However, most employ bulky multi-view 3D scanners or depth sensors to ensure high-quality volumetric per-frame reconstruction. Unlike these classical 3D/4D reconstruction methods, recent AI-based feed-forward 3D lifting techniques [2, 52] can lift a single RGB image from an off-the-shelf webcam into a neural radiance field (NeRF) representation encoded into a set of triplanes in real time, paving the way towards making 3D + +telepresence accessible to anyone [48]. + +![](images/af1c171076518b0c923ef4c3c10de2463aa514630b3845c04053af66948df8fa.jpg) +Figure 1. Comparisons. Given a single reference image and a single-view video frame, our method reconstructs the authentic dynamic appearance of the user (e.g., facial expressions and lighting) while producing a temporally coherent 3D video. A previous single-view 3D lifting method (LP3D) that reconstructs the avatar from the video frame on a per-frame basis suffers from distortions and temporal inconsistency. A portrait reenactment method (GPAvatar) drives the identity in the reference image using the video frame, but fails to capture accurate facial expressions (e.g., smile) and per-frame appearance (e.g., lighting). The output should be compared to the appearance of the input video frame (green box). + +Currently, there are two major paradigms in democratized 3D telepresence solutions from a single-view video: (1) single-view per-frame 3D lifting methods and (2) 3D portrait reenactment, which drives an identity from a reference image using another driving frame, but none of them is an ideal solution. For (1), single-frame-based lifting techniques such as LP3D [52], have the advantage of faithfully preserving the instantaneous dynamic conditions present in an input video, e.g. lighting, expressions, and posture, all + +of which are crucial to an authentic telepresence experience. However, single-image reconstruction methods operate independently on each frame and thus have fundamental limitations in maintaining temporal consistency. This difficulty stems from the inherent ill-posed nature of single-image-based reconstruction. In order to render novel views that are significantly far from the input view, the system cannot rely on information present in the input view and hence must hallucinate generate plausible content, which cannot be guaranteed to be consistent across multiple temporal frames. For example, LP3D's reconstruction changes significantly depending on the user's head pose in the input frame (see second row in Fig. 1 and third column in Fig. 2). In comparison, 3D self-reenactment methods create an avatar model from one or multiple reference images and use a separate driving video to drive the facial expressions and poses of the avatar. [6, 27, 51, 59]. While these approaches allow for temporally consistent results, they often cannot faithfully reconstruct the input video's dynamic conditions such as changes in lighting. Moreover, reenactment methods often struggle to generate accurate expressions because their expression detection and control are not precise enough (see the first row in Fig. 1). + +In this work, we address this dilemma of democratized 3D telepresence approaches and propose a novel solution to the problem of simultaneously maintaining temporal stability while preserving real-time dynamics of input videos for single-camera telepresence applications. Our proposed solution is a fusion-based approach that leverages the stability and accuracy of a canonical 3D prior, and also captures the diverse deviations from the prior in newer video frames (see the third row in Fig. 1). + +Our model first uses LP3D [52] to construct a canonical triplane prior from a (near) frontal image of the user, which can be casually captured or extracted from a video. During video reconstruction, our model lifts each input frame into a raw triplane, which is then fused with the canonical triplane (see Fig. 3). For images with oblique yaw head poses, artifacts and identity-related distortions may be present in its lifted triplane (see "LP3D" in Fig. 1 and 2). Hence we propose an undistorter module, which learns to undistort the raw instantaneous triplane to more closely match the identity of the correctly structured canonical triplane. We then propose a fuser module that combines the undistorted triplane and the canonical triplane in order to preserve identity consistency and recover occluded regions while reconstructing the dynamic lighting, nuanced expressions, and posture information in the input frame. + +We summarize our contributions as follows: + +- We contribute a novel triplane fusion method that combines the dynamic information from per-frame triplanes with a canonical triplane extracted from a reference image. Trained only using a synthetic multi-view video + +dataset, our feed-forward approach generates 3D portrait videos that demonstrate both temporal consistency and faithful reconstruction of the dynamic appearance of the user (e.g. lighting and expression), whereas prior solutions can only achieve one of the two properties. + +- We propose a new framework to evaluate single-view 3D portrait reconstruction methods using multi-view data and gain insight into the method's reconstruction quality and robustness. +- We present evaluations on both in-studio and in-the-wild data and demonstrate that our method achieves state-of-the-art performance in terms of temporal consistency and reconstruction accuracy. + +![](images/10853aafdfdf15407b71a6bf588e4b8fab5223564dfef49a2ada7f3d83b9a07b.jpg) +Figure 2. View-Dependent Distortion. "NV" refers to "Novel View Rendering". Top: inputs to our model and LP3D. Second & Third Rows: LP3D's reconstruction varies greatly under challenging viewpoints, showing a predictable pattern of artifacts including abnormally strong activations on the side being captured (red circle), as well as geometric distortion along the view direction of the camera due to depth ambiguity. We refer to this phenomenon as "View-Dependent Distortion". Fourth row: Our method removes such artifacts and achieves better coherence. + +# 2. Related Work + +2D portrait reenactment. Given a single or a few reference portrait images and a driving video, recent talking-head generators can reenact 2D portraits by transferring the facial expressions and poses from the driving video onto facial portraits [9, 10, 17, 46, 53, 54, 60, 62-64, 66]. However, being 2D, they cannot be rendered from novel viewpoints, which is crucial for 3D telepresence. + +3D-aware portrait generation and reenactment. Some recent works use deformable volumetric implicit radiance fields [34, 38, 39] or Gaussian splatting [22] combined by 3DMMs to reconstruct a photorealistic and animatable volumetric head avatar [1, 4, 14, 41, 45, 57, 67]. However, they require extensive data captures from videos or multiview cameras and person-specific training. Others use + +large-scale video datasets and learn a disentangled triplane 3D [5] for 3D facial reenactment in a feedforward fashion [6, 27, 28, 32, 51, 59, 61]. They construct a canonical 3D head from a reference image (often a neutral frame), and use facial expressions and head poses extracted from a separate driving video to animate it. As such, fine-grained facial expressions may not be captured due to errors in disentanglement. Most importantly, these reenactment methods fail to preserve the dynamic appearance of users (e.g., person-specific wrinkles or lighting) across time. + +3D GAN inversion. By combining GANs [15] and neural volume rendering [34], recent breakthroughs in 3D-aware GANs [5, 8, 16, 36, 44, 47, 55, 56, 58, 65, 68] can learn to generate photorealistic 3D heads from in-the-wild 2D images. Notably, EG3D proposes an efficient and compact triplane representation [5] to generate 3D heads, and Next3D [50] extends EG3D to create 3D portrait videos controlled by 3DMM facial expression and pose parameters. Besides generating 3D heads, these models can also be used for single-view 3D reconstruction using GAN inversion [13, 24, 30, 49], manipulation of the 3D head [18, 49, 69], and 3D personalization [3, 40]. However, inverting a few seconds of video can take many minutes or hours, and the inversion quality is often unsatisfactory due to inaccurate identity and expression preservation, as evaluated by GPAvatar [6]. Therefore, recent works [2, 52] propose encoder-based solutions to lift each video frame into a triplane. However, when each frame is independently lifted into 3D, the resulting 3D head video suffers in terms of temporal consistency - a key requirement of 3D telepresence systems. To enhance single-frame 2D-to-3D encoders for human heads, e.g., LP3D [52], we propose a triplane-fusion-based method, which improves their temporal stability while preserving temporal dynamics across time. + +# 3. Method + +# 3.1. Background: 3D Portrait from a Single Image + +LP3D [52] performs photorealistic 3D portrait reconstruction by using a feedforward encoder to lift an RGB image into a triplane $\mathbf{T} \in \mathbb{R}^{3 \times 32 \times 256 \times 256}$ , which can be volume rendered to an RGB image from any viewpoint. LP3D can run in real-time and has been developed into a complete realtime telepresence system [48]. We lift 2D faces into triplanes with a slightly modified LP3D trained on larger face crops containing shoulders and a camera estimator to recover the input image's camera parameters $M \in \mathbb{R}^{25}$ . It performs better than the original version (Table 3). + +# 3.2. Definitions + +We call a current video frame an "input frame" into our system. Additionally, we encode a near-frontal "reference image" of the subject into a "canonical triplane", which is used to stabilize the 3D video generation process. Lastly, we refer to the viewpoint of the camera in the input frame + +as "input viewpoint". + +# 3.3. Overview + +An overview of our method is illustrated in Fig. 3. To improve temporal consistency and reconstruct occluded parts of the face, we leverage an additional near-frontal reference image, which can be a selfie image or a video frame automatically obtained from the same video using an off-the-shelf head pose estimator like DECA [11]. We lift this reference image into the canonical triplane using pre-trained LP3D. When processing the video, we first lift each input frame into a raw triplane. Then, the Triplane Undistorter (Sec. 3.5) removes view-dependent distortions and recovers the identity using the canonical triplane as a reference. This produces an undistorted triplane. Finally, to recover regions that are occluded in the input frame and further improve stability, our Triplane Fuser (Sec. 3.6) combines the undistorted triplane with the canonical triplane to generate the final coherent triplane. + +# 3.4. Generating Synthetic Dynamic Multiview Data + +We synthesize ground truth multiview training images using Next3D [50]. We first extract 2D landmarks and FLAME [26] coefficient labels for the FFHQ [20] dataset using DECA [12]. Then, during training, we sample a pair of random FLAME coefficients from our FFHQ labels and input them to Next3D alongside a single random identity code $z$ . We notice that FFHQ contains mostly less extreme expressions. Thus, we randomly multiply the expression codes by a small scaling factor to exaggerate the expressions during training. Notice that the Next3D generator is much more expressive than FLAME due to its use of triplanes and GAN training even though it is conditioned on FLAME. Next3D then generates a pair of triplanes for $t = 0$ and $t = 1$ of the same person with 2 expressions. We render the $t = 0$ triplane into a near-frontal reference image $I_{ref}$ (Fig. 3 orange box) and the $t = 1$ triplane into 3 images: the input frame rendered from a random viewpoint (green box), the ground truth image at a different viewpoint (red box), and a frontal-view image (red box). Only the latter two are used for supervision. Additionally, to simulate different lighting conditions, we also apply different color augmentations (brightness, contrast, saturation, and hue) to images at $t = 0$ and $t = 1$ . + +Shoulder augmentation. It is important that 3D portraits capture dynamic shoulder movements to convey body language and achieve eye contact in telepresence. However, Next3D does not provide control over shoulder posture. Thus, we simulate shoulder movement in the rendered images without modifying the Next3D triplanes by warping camera rays during volume rendering. Please see the supplement for more detail. + +Pseudo-groundtruth triplanes. Due to our shoulder augmentation, the Next3D triplanes and their 2D renderings become inconsistent. Thus, Next3D's triplanes cannot be used + +![](images/d9070d5ebef1551b6b6a61f0bebca5b7fdd80545b2baa92dc0c5b294d45db6ea.jpg) +Figure 3. Method Overview. Given a near-frontal reference image and an input frame, we first reconstruct a canonical triplane and a raw triplane using LP3D [52] (Sec. 3.1). Next, we combine them through a Triplane Fusion module (blue box) that ensures temporal consistency while preserving dynamic information (e.g., lighting and shoulder pose) (Sec. 3.5 and 3.6). Our model is trained with only synthetic video data generated by a 3D GAN [50], with carefully designed augmentations to preserve shoulder motion and lighting (Sec. 3.4). + +as direct supervisory signals. To mitigate this, we leverage the fact that LP3D generates reasonably accurate triplanes from frontal view images. We use a frozen LP3D to predict pseudo-groundtruth triplanes $T_{\text{triplaneGT}}$ from the frontal novel view $I_{\text{frontal}}$ for $t = 1$ (Fig. 3 bottom right). + +# 3.5. Removing Distortion and Preserving Identity + +LP3D's reconstruction quality strongly depends on the input frame's viewpoint. When subjects are captured from the side, LP3D tends to produce errors in identity and exhibits stretching distortions (e.g., in the "Input Frame 1" and "Input Frame 2" columns in Fig. 2), primarily due to the inherent ambiguity of single-image reconstruction. In contrast, near-frontal views offer more complete identity cues and suffer less from occlusion (e.g. the "Reference Image" column in Fig. 2), leading to more reliable reconstruction. To reduce the single-image ambiguity, we adopt a near-frontal reconstruction as the canonical triplane $T_{\text{cano}}$ . Although incorporating more reference views could potentially improve accuracy, we find that using a single near-frontal view not only simplifies the model but also sufficiently enhances temporal coherence. + +To correct the distortion in an input raw triplane $T_{raw}$ using the canonical triplane $T_{cano}$ as reference, we devise a Triplane Undistorter $U$ (Fig. 3): + +$$ +U \left(T _ {\text {r a w}}, T _ {\text {c a n o}}\right) = T _ {\text {u n d i s t}} \in \mathbb {R} ^ {3 \times 3 2 \times 2 5 6 \times 2 5 6}. \tag {1} +$$ + +Since the distortion in each triplane is a 2D warping artifact, it can be reversed through a 2D undistortion warping of each plane. To this end, our Undistorter $U$ adopts the SPyNet[42] architecture to predict a 2D correction warping $T_{corr} \in \mathbb{R}^{3 \times 2 \times 256 \times 256}$ for the three planes. Then, the Undistorter corrects the raw triplane $T_{raw}$ by warping it based on the predicted 2D offsets $(\Delta u, \Delta v)$ in $T_{corr}$ as: + +$$ +T _ {c o r r} = S P y N e t \left(T _ {\text {r a w}}, T _ {\text {c a n o}}\right), \tag {2} +$$ + +$$ +T _ {u n d i s t} = \operatorname {W a r p} \left(T _ {\text {r a w}}, T _ {\text {c o r r}}\right). \tag {3} +$$ + +It is important to note the difference between optical flow and undistortion. An ideal optical flow model would warp $T_{\text{raw}}$ towards $T_{\text{cano}}$ in order to make the two identical. However, this alignment leads to significant artifacts and destroys the reconstruction (Table. 3 and Fig. 5). This is because the purpose of undistortion is not to make the two triplanes identical, but to correct $T_{\text{raw}}$ such that it has the same identity as the canonical triplane $T_{\text{cano}}$ , while preserving dynamic information such as expressions and lighting. Therefore, instead of using the canonical triplane $T_{\text{cano}}$ as the warping target, our Undistorter $U$ merely uses $T_{\text{cano}}$ as the identity conditioning to correct the raw triplane $T_{\text{raw}}$ towards the ground truth triplane $T_{\text{triplaneGT}}$ (Fig. 3, lower-right), which is not available during test time. The correction warping is supervised by the consistency between the undistorted triplane $T_{\text{undist}}$ and pseudo-groundtruth tri + +plane $T_{\text{triplane}GT}$ (Sec. 3.4) via a triplane loss: + +$$ +L _ {u n d i s t} = L _ {1} \left(T _ {u n d i s t}, T _ {t r i p l a n e G T}\right). \tag {4} +$$ + +# 3.6. Reconstructing Occluded Regions through Triplane Fusion + +As the user moves around in the video, different parts of their head become occluded. To recover occluded areas in the input frame and further stabilize the subject's identity across the video, our Fuser $F$ enhances the reconstruction by utilizing the canonical triplane $T_{\text{cano}}$ , in which the currently occluded areas are often visible. Therefore, it is important for the Fuser $F$ to identify and recover the occluded regions while preserving information from the visible regions in $T_{\text{undist}}$ . To accomplish this, we thus use a 5-layer ConvNet-based visibility estimator $V$ to predict a visibility triplane as $V(T_{\text{raw}}) = T_{\text{raw}}^{\text{vis}} \in \mathbb{R}^{3 \times 1 \times 128 \times 128}$ , i.e., one visibility map for each plane. $T_{\text{raw}}^{\text{vis}}$ is undistorted alongside $T_{\text{undist}}$ to produce $T_{\text{undist}}^{\text{vis}}$ . We also predict a visibility triplane for the canonical triplane as $T_{\text{cano}}^{\text{vis}} = V(T_{\text{cano}})$ . Finally, the Fuser $F$ produces the fused triplane $T_{\text{fused}}$ by combining information from the undistorted input tranplane $T_{\text{undist}}$ , its visibility triplane $T_{\text{undist}}^{\text{vis}}$ , the canonical triplane $T_{\text{cano}}$ , and its visibility triplane $T_{\text{cano}}^{\text{vis}}$ . + +$$ +T _ {f u s e d} = F \left(T _ {u n d i s t}, T _ {u n d i s t} ^ {\text {v i s}}, T _ {c a n o}, T _ {c a n o} ^ {\text {v i s}}\right). \tag {5} +$$ + +In this way, Fuser $F$ merges new information in the input frame with occluded regions visible in the reference image. To train the visibility predictor $V$ , we calculate the visibility loss $L_{vis}$ as the $L_{1}$ distance between the predicted visibility triplanes versus the ground truth: + +$$ +L _ {v i s} = L _ {1} (T _ {r a w} ^ {v i s}, T _ {r a w} ^ {v i s G T}) + L _ {1} (T _ {c a n o} ^ {v i s}, T _ {c a n o} ^ {v i s G T}). \quad (6) +$$ + +The ground truth visibility triplanes contain 1 for pixels that are visible, and 0 otherwise. The generation of the ground truth visibility triplanes $T_{raw}^{visGT}$ and $T_{cano}^{visGT}$ is discussed in the supplementary. + +To supervise the Fuser $F$ , we calculate the fusion loss $L_{fusion}$ as the $L_{1}$ loss between the fused triplane $T_{fused}$ and the pseudo-groundtruth triplane $T_{frontalGT}$ . To highlight the currently occluded region during training, we also upweight the occluded region using an occlusion mask $T_{occMask} \in \mathbb{R}^{3 \times 1 \times 256 \times 256}$ as: + +$$ +T _ {d i f f} = \left| T _ {f u s e d} - T _ {t r i p l a n e G T} \right|, \tag {7} +$$ + +$$ +L _ {f u s i o n} = \operatorname {M e a n} \left(T _ {d i f f}\right) + \frac {T _ {d i f f} \cdot T _ {v i s G T}}{\left| T _ {v i s G T} \right|} + \frac {T _ {d i f f} \cdot T _ {o c c M a s k}}{\left| T _ {o c c M a s k} \right|}. \tag {8} +$$ + +Please refer to the supplementary for the calculation of the occlusion mask. We use the Recurrent Video Restoration Transformer (RVRT) [29] as the backbone of our Fuser $F$ because of its memory efficiency. We find that the final summation skip connection in RVRT prevents effective learning. This is because the original RVRT was designed to correct local blurriness and noise in a corrupted RGB video, + +whereas our triplane videos exhibit structural distortion on a much larger scale and the summation skip connection thus limits the model's ability to correct the general structure. We thus replace the summation with a small 5-layer ConvNet. + +Lastly, note that both the Undistorter $U$ and the Fuser $F$ consist of 3 separate copies, one for each of the 3 planes, because we find that processing all three planes jointly leads to collapse to 2D (please see supplementary for visualization and analysis). + +# 3.7. Training Losses + +Our loss function is the summation of four loss terms that provide two types of supervision: (a) direct triplane space guidance used to supervise the undistortion process in the Undistorter $U$ , the visibility prediction process, and the fusion process in the Fuser $F$ ; and (b) image space guidance for overall learning of high-quality image synthesis: + +$$ +\begin{array}{l} L = w _ {u n d i s t} L _ {u n d i s t} + w _ {f u s i o n} L _ {f u s i o n} \\ + w _ {v i s} L _ {v i s} + w _ {r e n d e r} L _ {r e n d e r}, \tag {9} \\ \end{array} +$$ + +where $w_{undist}$ , $w_{viz}$ , $w_{fusion}$ , and $w_{render}$ are scalar weights for the different loss terms. $L_{render}$ is calculated as the perceptual loss $L_{LPIPS}$ between the ground truth novel view $I_{GT}$ and the rendered novel view $I_{render}$ as: + +$$ +L _ {\text {r e n d e r}} = L _ {\text {L P I P S}} \left(I _ {G T}, I _ {\text {r e n d e r}}\right). \tag {10} +$$ + +# 4. Evaluation + +Multi-view evaluation of single-view reconstruction. Due to the lack of 3D ground truth for in-the-wild images, prior methods are often evaluated on the input image reconstruction task, which compares the rendered image against the input image using metrics like PSNR, whereas the novel view synthesis task often relies on visual assessments. However, this practice can lead to ambiguities and inaccurate conclusions. For example, if a 3D reconstruction is only rendered and evaluated from the input viewpoint, then that reconstruction can overfit to the input view to achieve high scores even if it is inaccurate when rendered from other viewpoints (as proven quantitatively in Fig. 11 of the supplementary). Moreover, single-view reconstruction methods can be heavily affected by the choice of input viewpoints (Fig. 2). Therefore, both the input viewpoint and the novel viewpoint are crucial variables to consider when concluding the performance of a method. We thus propose a new evaluation framework that evaluates a model across every input-novel viewpoint combination. In this way, a method can only achieve high numerical performance when it consistently generates high-quality reconstructions regardless of the choice of input or novel viewpoint. + +# 4.1. Evaluation Dataset + +We quantitatively evaluate various methods on the NeRSemble [23] dataset, which is a multiview portrait + +video dataset that allows us to evaluate the methods using different input-novel viewpoint combinations. It is recorded with 16 calibrated time-synchronized cameras in a controlled studio environment. The images are captured at $7.1\mathrm{MP}$ resolution and 73 frames per second. There are 10 sequences in the test set, each capturing a different individual performing different expressions. One of the 10 test sequences involves severe facial occlusion from hair that causes most of the methods' face trackers to fail for significant portions of the recording for many of the viewpoints. We thus leave out that sequence because the results would not be a reliable assessment of quality. We also use 8 roughly evenly separated cameras out of all 16 cameras during evaluation. + +# 4.2. Metrics + +Synthesis Quality. Given $N$ views in the dataset, we evaluate a method's average performance across every input-novel viewpoint combination. More specifically, for each frame, each of the $N$ cameras is used as the input viewpoint, producing $N$ reconstructions in total. Then, each of the $N$ reconstructions is rendered and evaluated on the $N$ viewpoints, resulting in an $N \times N$ score matrix. As mentioned before, we use $N = 8$ of the camera views from the NeRSemnble[23] dataset as input and novel views. Thus, for a test sequence of $T$ frames, we can calculate a spatial-temporal score matrices $\mathbf{S}^{T \times 8 \times 8}$ for each of the metrics (as visualized in Fig. 11 in the supplementary): + +$$ +\mathbf {S} _ {t, i, j} = \operatorname {M e t r i c} \left(\mathbf {I} _ {\text {r e n d e r}} ^ {t, i, j}, \mathbf {I} _ {G T} ^ {t, j}\right), 1 \leq i, j \leq N, 1 \leq t \leq T, \tag {11} +$$ + +$$ +s = \operatorname {M e a n} \left(\left\{\mathbf {S} _ {t, i, j} \right\}\right), \tag {12} +$$ + +where $Metric(\cdot)$ can be LPIPS, PSNR, etc.. $\mathbf{I}_{\text{render}}^{t,i,j}$ is the image rendered using camera $i$ as the input frame and camera $j$ as the output rendering view at frame $t$ . $\mathbf{I}_{GT}^{t,j}$ is the ground truth frame captured by camera $j$ at frame $t$ . The Synthesis Quality $s$ is thus the average over all entries in S. For a dataset with multiple sequences, such as NeRSample, the final Synthesis Quality is the average across all sequences. + +Novel View Synthesis (NVS) Quality. Novel View Synthesis Quality $s_{NV}$ is the average over all entries where the input view $i$ is different from the novel view $j$ , essentially removing the diagonal entries from the score matrix: + +$$ +s _ {N V} = \operatorname {M e a n} \left(\left\{\mathbf {S} _ {t, i, j} \mid i \neq j, 1 \leq t \leq T \right\}\right). \tag {13} +$$ + +Identity Accuracy (ID) We measure the ArcFace [7] cosine distance between the ground truth image $I_{GT}$ and rendered image $I_{render}$ both from the frontal camera. + +Expression Accuracy (Expr) We use NVIDIA Maxine AR SDK [35] to measure the $L_{2}$ distance between expression coefficients $e_{GT}$ of the ground truth image and $e_{render}$ of the rendered image both from the frontal camera. + +Dynamic Appearance While we believe that it is important to measure the accuracy of dynamic appearance such + +as lighting and shoulder poses, there is no existing multiview in-the-wild portrait video dataset to support such evaluation. We thus qualitatively evaluate various methods on challenging in-the-wild portrait videos. Please see the supplementary material for image examples and video results. + +# 4.3. Comparisons + +Baselines. We compare recent methods from 3 categories: (i) Reconstruction: We evaluate LP3D [52] using the above protocol. LP3D generates $N$ 3D portraits using each of the $N$ viewpoints as input, and the $N$ 3D portraits are then evaluated on the $N$ ground truth viewpoints. (ii) Reenactment: We evaluate Li et al. [28] and GPAvatar [6] in the self-reenactment setting. Li et al. [28] reconstruct 3D portraits as triplanes from reference images without test-time optimization, and they drive the 3D portraits via dynamic frontal renderings of 3DMMs. The authors of Li et al. [28] evaluated their approach using the input views as the only novel view, instead of all $N$ views. GPAvatar[6] reconstructs 3D portraits by leveraging multiple source images and driving them through a FLAME [26] mesh model. We use the first frame of the frontal camera in each NeRSample test sequence as the reference image to generate the 3D portrait, and drive it using videos from each of the $N$ viewpoints. We evaluate GPAvatar using the same evaluation protocol as our method and LP3D. (iii) Inversion: We evaluate VIVE3D [13], which is a state-of-the-art 3D GAN inversion method for videos, and it can also perform semantic video editing. We first provide VIVE3D 3 video frames from the input viewpoint, which it requires for personalization. Then, we invert the input video frames into 3D portraits and render the 3D portraits from all $N$ viewpoints. Each of the above method uses a different cropping of the face. We standardize the evaluation by re-cropping all methods to our cropping protocol, which is the largest of all. Please see the supplement for an additional table, where we evaluate the methods using different cropings around the face and arrive at conclusions consistent with Table 1. + +Quantitative results. We evaluate various methods using different input-novel viewpoint combinations, providing a robust multi-view assessment of a model's performance. Table. 1 shows that our model achieves state-of-the-art performance across all metrics versus recent works. Notably, while LP3D overfits to the input viewpoint (as shown by the difference of $0.806\mathrm{dB}$ PSNR between Synthesis Quality and NVS Quality), our method generalizes better to novel viewpoints (as shown by the difference of $0.330\mathrm{dB}$ ). Moreover, our method best preserves subject identity and expression (Table. 1, Fig. 2, and 4). On the other hand, the reenactment methods struggle to capture authentic expressions because of the use of morphable face models, which have limited expressiveness and accuracy. Moreover, they cannot reconstruct dynamic appearance (e.g. the stuck-out tongue in the second and third rows of Fig. 4) because they solely rely on information present in the reference image(s) and + +![](images/4e299ad7c03841300890bb99d16bffa0a5a809729a23a80a534e72d20a65bd4f.jpg) +Groundtruth +Ours +LP3D +GPAvatar +Input / Driving Frame +Reference / Source +Figure 4. Visual comparisons with baseline methods. Our method strikes a balance between coherent reconstruction and faithfully preserving dynamic conditions like expressions. LP3D (third column) exhibits inconsistency in identities, hairstyles, and artifacts (red circles). GPAvatar (fourth column) fails to capture challenging expressions (first row), new information not present in the reference image, (the stuck-out tongue in second and third rows), and the identity of the person (last row). + +do not incorporate new per-frame information. On the other hand, our method achieves faithful dynamic appearance and coherent reconstruction at the same time. + +
MethodTypeExpr↓ID↓SynthesisLIPS↓LVS Quality
PSNR↑LPIPS↓PSNR↑LPIPS↓
Li et al. [28]reenact0.2660.24118.5730.25518.2020.262
GPAvatar[6]reenact0.2040.20721.9490.23321.9490.2334
VIVE3D[13]invert0.2900.39518.5770.25918.1450.271
LP3D[52]recon0.1680.21522.3310.22321.5250.237
Oursrecon0.1580.18722.7700.21922.4400.224
+ +Table 1. Comparisons on Nersemble [23]. Our evaluation protocol (Sec. 4.2) utilizes multiview groundtruth for thorough analysis. Our method achieves state-of-the-art performance across all metrics, maintains temporal coherence, and preserves accurate expressions. + +# 4.4. Ablations + +We evaluate the effectiveness of our proposed Triplane Undistorter and Fuser modules on NeRSemble. In addition to the Synthesis Quality in terms of PSNR and LPIPS, we develop two new metrics to separately measure the robustness to variations in (1) rendering viewpoints and (2) input viewpoints: + +(1) Novel View Variation (NVV). We evaluate how much a method's reconstruction quality varies across different novel views. We quantify this as the standard deviation of performance across the $N$ novel views when using the same input view, i.e. the average standard deviation of each horizontal row of the score matrix $S$ : + +$$ +\mathrm {N V V} = \frac {1}{T N} \sum_ {t = 1} ^ {T} \sum_ {i = 1} ^ {N} \operatorname {s t d d e v} \left(\left\{S _ {t, i, j} \mid 1 \leq j \leq N \right\}\right). \tag {14} +$$ + +(2) Input View Variation (IVV). We measure how much a method's reconstruction quality varies when using the $N$ different input viewpoints (Sec. 4.2 second column from the + +
MethodU FSynthesis QualityIVV↓NVV↓
PSNR↑LPIPS↓(PSNR)(LPIPS)(PSNR)(LPIPS)
LP3D[52]XX22.3310.2231.0250.0152.2000.053
Ours✓X22.1960.2210.9070.0091.6990.038
X✓22.2650.2230.5590.0061.3150.029
✓✓22.7700.2190.2450.0051.3830.037
+ +Table 2. Undistorter and Fuser Ablations. The Undistorter-only (row 2) and Fuser-only (row 3) variants both lead to improvements in viewpoint robustness metrics (IVV and NVV), but not in Synthesis Quality. The best performance is achieved by using both the Undistorter and the Fuser at the same time. + +right). We quantify this variation as the average standard deviation of performance on the same novel view when using different input views, i.e. the average standard deviation of each vertical column of the score matrix $S$ : + +$$ +\mathrm {I V V} = \frac {1}{T N} \sum_ {t = 1} ^ {T} \sum_ {j = 1} ^ {N} \operatorname {s t d d e v} \left(\left\{S _ {t, i, j} \mid 1 \leq i \leq N \right\}\right). \tag {15} +$$ + +$NVV$ and $IVV$ are measured for both PSNR and LPIPS, and they are better if lower. + +As shown in Table. 2, the Undistorter module consistently improves the NVV and IVV metrics versus LP3D, indicating better robustness to different input viewpoints and more consistent rendering quality across novel rendered views. However, when only the Undistorter is added (Tab. 2 second row) the PSNR is slightly reduced. This is likely because this model does not leverage the reference image to recover occluded areas while also overfitting to the input view. Similarly, the Fuser-only variant (Tab. 2 third row) achieves better NVV and IVV scores, but also scores lower in terms of PSNR. A likely cause is that, without the Undistorter, the Fuser needs to merge highly misaligned triplanes, where the person looks drastically different in $T_{raw}$ and $T_{cano}$ . This misalignment possibly induces more blurriness and alignment artifacts that lower the PSNR. The best performance is achieved by our full model, which includes both the Undistorter and the Fuser and thus achieves better PSNR, LPIPS, IVV and NVV scores. This means that the Undistorter and the Fuser complement each other: The Undistorter corrects the distortion in the raw triplane and thus reduces the challenges in fusing misaligned triplanes, and the Fuser recovers the occluded areas in the raw triplane. + +Shoulder Augmentation. without shoulder augmentation, the model fails to correctly reconstruct varying shoulder poses, resulting in worse performance (Fig. 5 and Table. 3). + +Optical Flow. As mentioned in Sec. 3.5, replacing the Undistorter with optical flow worsens the reconstruction (Table. 3). As shown by Fig. 5, this variant fails to correctly reconstruct the shoulder and the mouth accurately. + +GAN Training. We retrained our model with an additional adversarial loss incorporated into Eqn. 10. As expected, the + +![](images/0bf2bdf2ae8f88065d6249dc2bbb6387132e237314eb57e37ea0f8f069b5b960.jpg) + +Figure 5. Visual ablation. Our method with optical flow and w/o shoulder augmentation on two different input frames (top and bottom rows). + +
MethodLPIPS↓PSNR↑IVV (PSNR) ↓ID↓Expr↓
LP3D (orig.)0.33418.7212.1300.2470.451
LP3D (ours)0.22322.3311.0250.1680.215
w optical flow0.22722.0851.1750.1780.335
w/o shoulder aug.0.21822.3420.8290.1530.244
Ours0.21922.7700.2450.1580.187
Ours + GAN0.20722.0740.3140.1520.198
+ +Table 3. Other Ablations. We compare our implementation of LP3D to the original one [52] (row 1&2); the effect of replacing our Undistorter with optical flow warping (row 4); the effect of removing shoulder augmentation during training (row 4); and the effect of adding an adversarial GAN loss (row 6) alongside Eqn. 10. + +GAN training (Table. 3) improves the sharpness of images as seen from the LPIPS error reduction from 0.219 to 0.207. Additional results are included in the supplementary. + +Qualitative results. In-the-wild experiments show that our method achieves better temporal consistency than LP3D [52] and more accurately captures dynamic information like expressions and lighting changes than GPAvatar [6]. We refer the readers to our supplementary materials for visual examples of these results. + +# 5. Discussion + +Conclusion. Recognizing the individual limitations of per-frame single-view reconstruction and 3D reenactment methods, we presented the first single-view 3D lifting method to reconstruct a 3D photorealistic avatar with faithful dynamic appearance as well as temporal consistency, marrying the best of both worlds. We believe our method paves the way forward for creating a high-quality telepresence system accessible to consumers. + +Limitations and future work. We use a single reference image, but incorporating multiple ones with different expressions and head poses could lead to further improvements. With our method, fusing an extreme side view with a very different expression to the reference view may result in blurry reconstruction due to ambiguity in triplane alignment. Finally, due to the additional components, our current run-time performance is slower than real time, which could be improved in future work. + +# References + +[1] ShahRukh Athar, Zexiang Xu, Kalyan Sunkavalli, Eli Shechtman, and Zhixin Shu. Rignerf: Fully controllable neural 3d portraits. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2 +[2] Ananta R. Bhattarai, Matthias Nießner, and Artem Sevastopolsky. Triplanenet: An encoder for eg3d inversion. 2024. 1, 3 +[3] Marcel C. Buehler, Kripasindhu Sarkar, Tanmay Shah, Gengyan Li, Daoye Wang, Leonhard Helminger, Sergio Orts-Escolano, Dmitry Lagun, Otmar Hilliges, Thabo Beeler, and Abhimitra Meka. Preface: A data-driven volumetric prior for few-shot ultra high-resolution face synthesis. In IEEE International Conference on Computer Vision (ICCV), 2023. 3 +[4] Chen Cao, Tomas Simon, Jin Kyu Kim, Gabe Schwartz, Michael Zollhoefer, Shun-Suke Saito, Stephen Lombardi, Shih-En Wei, Danielle Belko, Shouou-I Yu, Yaser Sheikh, and Jason Saragih. Authentic volumetric avatars from a phone scan. ACM Transactions on Graphics (SIGGRAPH), 2022. 2 +[5] Eric R. Chan, Connor Z. Lin, Matthew A. Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas Guibas, Jonathan Tremblay, Sameh Khamis, Tero Karras, and Gordon Wetzstein. Efficient geometry-aware 3D generative adversarial networks. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3 +[6] Xuangeng Chu, Yu Li, Ailing Zeng, Tianyu Yang, Lijian Lin, Yunfei Liu, and Tatsuya Harada. Gpavatar: Generalizable and precise head avatar from image(s). In International Conference on Learning Representations (ICLR), 2024. 2, 3, 6, 7, 8 +[7] Jiankang Deng, Jia Guo, Xue Niannan, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 6 +[8] Yu Deng, Jiaolong Yang, Jianfeng Xiang, and Xin Tong. Gram: Generative radiance manifolds for 3d-aware image generation. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3 +[9] Michail Christos Doukas, Stefanos Zafeiriou, and Viktoriia Sharmanska. Headgan: One-shot neural head synthesis and editing. In IEEE International Conference on Computer Vision (ICCV), 2021. 2 +[10] Nikita Drobyshev, Jenya Chelishev, Taras Khakhulin, Aleksei Ivakhnenko, Victor Lempitsky, and Egor Zakharov. Megaportraits: One-shot megapixel neural head avatars. arXiv preprint arXiv:2207.07621, 2022. 2 +[11] Yao Feng, Haiwen Feng, Michael J. Black, and Timo Bolkart. Learning an animatable detailed 3D face model from in-the-wild images. 2021. 3 +[12] Yao Feng, Haiwen Feng, Michael J Black, and Timo Bolkart. Learning an animatable detailed 3d face model from in-the-wild images. ACM Transactions on Graphics (ToG), 40(4): 1-13, 2021. 3 +[13] Anna Fruhstück, Nikolaos Sarafianos, Yuanlu Xu, Peter Wonka, and Tony Tung. VIVE3D: Viewpoint-independent video editing using 3D-Aware GANs. In IEEE Conference + +on Computer Vision and Pattern Recognition (CVPR), 2023. 3, 6, 7 +[14] Guy Gafni, Justus Thies, Michael Zollhöfer, and Matthias Nießner. Dynamic neural radiance fields for monocular 4d facial avatar reconstruction. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2 +[15] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In Advances in Neural Information Processing Systems (NeurIPS), 2014. 3 +[16] Jiatao Gu, Lingjie Liu, Peng Wang, and Christian Theobalt. StyleNeRF: A style-based 3D-aware generator for high-resolution image synthesis. arXiv preprint arXiv:2110.08985, 2021. 3 +[17] Fa-Ting Hong, Longhao Zhang, Li Shen, and Dan Xu. Depth-aware generative adversarial network for talking head video generation. 2022. 2 +[18] Yang Hong, Bo Peng, Haiyao Xiao, Ligang Liu, and Juyong Zhang. Headnerf: A real-time nerf-based parametric head model. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3 +[19] Andrew Jones, Magnus Lang, Graham Fyffe, Xueming Yu, Jay Busch, Ian McDowall, Mark Bolas, and Paul Debevec. Achieving eye contact in a one-to-many 3d video teleconferencing system. ACM Transactions on Graphics (SIGGRAPH), 2009. 1 +[20] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 3 +[21] Peter Kauff and Oliver Schreer. An immersive 3d videoconferencing system using shared virtual team user environments. In Proceedings of the 4th International Conference on Collaborative Virtual Environments, 2002. 1 +[22] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics (SIGGRAPH), 2023. 2 +[23] Tobias Kirschstein, Shenhan Qian, Simon Giebenhain, Tim Walter, and Matthias Nießner. Nersemble: Multi-view radiance field reconstruction of human heads. ACM Trans. Graph., 2023. 5, 6, 7 +[24] Jaehoon Ko, Kyusun Cho, Daewon Choi, Kwangrok Ryoo, and Seungryong Kim. 3d gan inversion with pose optimization. In IEEE Winter Conference on Applications of Computer Vision (WACV), 2023. 3 +[25] Jason Lawrence, Dan B Goldman, Supreeth Achar, Gregory Major Blascovich, Joseph G. Desloge, Tommy Fortes, Eric M. Gomez, Sascha Haberling, Hugues Hoppe, Andy Huibers, Claude Knaus, Brian Kuschak, Ricardo Martin-Brualla, Harris Nover, Andrew Ian Russell, Steven M. Seitz, and Kevin Tong. Project starline: A high-fidelity telepresence system. ACM Transactions on Graphics (SIGGRAPH ASIA), 2021. 1 +[26] Tianye Li, Timo Bolkart, Michael. J. Black, Hao Li, and Javier Romero. Learning a model of facial shape and expression from 4D scans. 2017. 3, 6 +[27] Weichuang Li, Longhao Zhang, Dong Wang, Bin Zhao, Zhi-gang Wang, Mulin Chen, Bang Zhang, Zhongjian Wang, + +Liefeng Bo, and Xuelong Li. One-shot high-fidelity talking-head synthesis with deformable neural radiance field. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2, 3 +[28] Xueting Li, Shalini De Mello, Sifei Liu, Koki Nagano, Umar Iqbal, and Jan Kautz. Generalizable one-shot neural head avatar. 2023. 3, 6, 7 +[29] Jingyun Liang, Yuchen Fan, Xiaoyu Xiang, Rakesh Ranjan, Eddy Ilg, Simon Green, Jiezhang Cao, Kai Zhang, Radu Timofte, and Luc Van Gool. Recurrent video restoration transformer with guided deformable attention. 2022. 5 +[30] C.Z. Lin, D.B. Lindell, E.R. Chan, and G. Wetzstein. 3d gan inversion for controllable portrait image animation. In ECCV Workshop on Learning to Generate 3D Shapes and Scenes, 2022. 3 +[31] S. Ma, T. Simon, J. Saragih, D. Wang, Y. Li, F. La Torre, and Y. Sheikh. Pixel codec avatars. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 1 +[32] Zhiyuan Ma, Xiangyu Zhu, Guojun Qi, Zhen Lei, and Lei Zhang. Otavatar: One-shot talking face avatar with controllable tri-plane rendering. 2023. 3 +[33] Andrew Maimone, Jonathan Bidwell, Kun Peng, and Henry Fuchs. Enhanced personal autostereoscopic telepresence system using commodity depth cameras. Computers & Graphics, 2012. 1 +[34] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing scenes as neural radiance fields for view synthesis. In European Conference on Computer Vision (ECCV), 2020. 2, 3 +[35] NVIDIA. Nvidia maxine ar sdk. https://github.com/NVIDIA/MAXINE-AR-SDK, 2024.6 +[36] Roy Or-El, Xuan Luo, Mengyi Shan, Eli Shechtman, Jeong Joon Park, and Ira Kemelmacher-Shlizerman. StyleSDF: High-Resolution 3D-Consistent Image and Geometry Generation. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3 +[37] Sergio Orts-Escalano, Christoph Rhemann, Sean Fanello, Wayne Chang, Adarsh Kowdle, Yury Degtyarev, David Kim, Philip L. Davidson, Sameh Khamis, Mingsong Dou, Vladimir Tankovich, Charles Loop, Qin Cai, Philip A. Chou, Sarah Mennicken, Julien Valentin, Vivek Pradeep, Shenlong Wang, Sing Bing Kang, Pushmeet Kohli, Yuliya Lutchyn, Cem Keskin, and Shahram Izadi. Holoportation: Virtual 3d teleportation in real-time. In UIST 2016, 2016. 1 +[38] Keunhong Park, Utkarsh Sinha, Jonathan T. Barron, Sofien Bouaziz, Dan B Goldman, Steven M. Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. 2021. 2 +[39] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-NeRF: Neural Radiance Fields for Dynamic Scenes. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2 +[40] Luchao Qi, Jiaye Wu, Annie N. Wang, Shengze Wang, and Roni Sengupta. My3dgen: A scalable personalized 3d generative model, 2023. 3 +[41] Shenhan Qian, Tobias Kirschstein, Liam Schoneveld, Davide Davoli, Simon Giebenhain, and Matthias Nießner. Gaussianavatars: Photorealistic head avatars with rigged 3d gaussians. 2024. 2 + +[42] Anurag Ranjan and Michael J Black. Optical flow estimation using a spatial pyramid network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4161-4170, 2017. 4 +[43] Ramesh Raskar, Greg Welch, Matt Cutts, Adam Lake, Lev Stesin, and Henry Fuchs. The office of the future: a unified approach to image-based modeling and spatially immersive displays. In ACM Transactions on Graphics (SIGGRAPH), 1998. 1 +[44] Daniel Rebain, Mark Matthews, Kwang Moo Yi, Dmitry Lagan, and Andrea Tagliasacchi. Lolnerf: Learn from one look. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3 +[45] Shunsuke Saito, Gabriel Schwartz, Tomas Simon, Junxuan Li, and Giljoo Nam. Relightable gaussian codec avatars. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2 +[46] Aliaksandr Siarohin, Stéphane Lathuilière, Sergey Tulyakov, Elisa Ricci, and Nicu Sebe. First order motion model for image animation. In Advances in Neural Information Processing Systems (NeurIPS), 2019. 2 +[47] Ivan Skorokhodov, Sergey Tulyakov, Yiqun Wang, and Peter Wonka. Epigraph: Rethinking training of 3d gans. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 3 +[48] Michael Stengel, Koki Nagano, Chao Liu, Matthew Chan, Alex Trevithick, Shalini De Mello, Jonghyun Kim, and David Luebke. AI-Mediated 3D Video Conferencing. In ACM SIGGRAPH 2023 Emerging Technologies, 2023. 1, 3 +[49] Jingxiang Sun, Xuan Wang, Yichun Shi, Lizhen Wang, Jue Wang, and Yebin Liu. Ide-3d: Interactive disentangled editing for high-resolution 3d-aware portrait synthesis. ACM Transactions on Graphics (SIGGRAPH ASIA), 2022. 3 +[50] Jingxiang Sun, Xuan Wang, Lizhen Wang, Xiaoyu Li, Yong Zhang, Hongwen Zhang, and Yebin Liu. Next3d: Generative neural texture rasterization for 3d-aware head avatars. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 3, 4 +[51] Phong Tran, Egor Zakharov, Long-Nhat Ho, Anh Tuan Tran, Liwen Hu, and Hao Li. Voodoo 3d: Volumetric portrait disentanglement for one-shot 3d head reenactment. 2024. 2, 3 +[52] Alex Trevithick, Matthew Chan, Michael Stengel, Eric R. Chan, Chao Liu, Zhiding Yu, Sameh Khamis, Manmohan Chandraker, Ravi Ramamoorthi, and Koki Nagano. Realtime radiance fields for single-image portrait view synthesis. In ACM Transactions on Graphics (SIGGRAPH), 2023. 1, 2, 3, 4, 6, 7, 8 +[53] Ting-Chun Wang, Arun Mallya, and Ming-Yu Liu. One-shot free-view neural talking-head synthesis for video conferencing. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2 +[54] Yaohui Wang, Di Yang, Francois Bremond, and Antitza Dantcheva. Latent image animator: Learning to animate images via latent space navigation. In International Conference on Learning Representations (ICLR), 2022. 2 +[55] Jianfeng Xiang, Jiaolong Yang, Yu Deng, and Xin Tong. Gram-hd: 3d-consistent image generation at high resolution with generative radiance manifolds. arXiv preprint arXiv:2206.07255, 2022. 3 + +[56] Yinghao Xu, Sida Peng, Ceyuan Yang, Yujun Shen, and Bolei Zhou. 3d-aware image synthesis via learning structural and textural representations. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3 +[57] Yuelang Xu, Benwang Chen, Zhe Li, Hongwen Zhang, Lizhen Wang, Zerong Zheng, and Yebin Liu. Gaussian head avatar: Ultra high-fidelity head avatar via dynamic gaussians. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2 +[58] Zhongcong Xu, Jianfeng Zhang, Junhao Liew, Wenqing Zhang, Song Bai, Jiashi Feng, and Mike Zheng Shou. Pv3d: A 3d generative model for portrait video generation. In The Tenth International Conference on Learning Representations, 2023. 3 +[59] Zhenhui Ye, Tianyun Zhong, Yi Ren, Jiaqi Yang, Weichuang Li, Jiangwei Huang, Ziyue Jiang, Jinzheng He, Rongjie Huang, Jinglin Liu, Chen Zhang, Xiang Yin, Zejun Ma, and Zhou Zhao. Real3d-portrait: One-shot realistic 3d talking portrait synthesis. In International Conference on Learning Representations (ICLR), 2024. 2, 3 +[60] Fei Yin, Yong Zhang, Xiaodong Cun, Mingdeng Cao, Yanbo Fan, Xuan Wang, Qingyan Bai, Baoyuan Wu, Jue Wang, and Yujiu Yang. Styleheat: One-shot high-resolution editable talking face generation via pre-trained stylegan. 2022. 2 +[61] Wangbo Yu, Yanbo Fan, Yong Zhang, Xuan Wang, Fei Yin, Yunpeng Bai, Yan-Pei Cao, Ying Shan, Yang Wu, Zhongqian Sun, and Baoyuan Wu. Nofa: Nerf-based one-shot facial avatar reconstruction. In ACM SIGGRAPH 2023 Conference Proceedings, 2023. 3 +[62] Egor Zakharov, Aliaksandra Shysheya, Egor Burkov, and Victor Lempitsky. Few-shot adversarial learning of realistic neural talking head models. In IEEE International Conference on Computer Vision (ICCV), 2019. 2 +[63] Egor Zakharov, Aleksei Ivakhnenko, Aliaksandra Shysheya, and Victor Lempitsky. Fast bi-layer neural synthesis of one-shot realistic head avatars. In European Conference on Computer Vision (ECCV), 2020. +[64] Bowen Zhang, Chenyang Qi, Pan Zhang, Bo Zhang, Hsiang-Tao Wu, Dong Chen, Qifeng Chen, Yong Wang, and Fang Wen. Metaportrait: Identity-preserving talking head generation with fast personalized adaptation, 2023. 2 +[65] Xuanmeng Zhang, Zhedong Zheng, Daiheng Gao, Bang Zhang, Pan Pan, and Yi Yang. Multi-view consistent generative adversarial networks for 3d-aware image synthesis. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3 +[66] Jian Zhao and Hui Zhang. Thin-plate spline motion model for image animation. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 3657-3666, 2022. 2 +[67] Yufeng Zheng, Victoria Fernández Abrevaya, Marcel C. Bühler, Xu Chen, Michael J. Black, and Otmar Hilliges. I M Avatar: Implicit morphable head avatars from videos. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2 +[68] Peng Zhou, Lingxi Xie, Bingbing Ni, and Qi Tian. CIPS-3D: A 3D-Aware Generator of GANs Based on Conditionally-Independent Pixel Synthesis. arXiv preprint arXiv:2110.09788, 2021. 3 + +[69] Yiyu Zhuang, Hao Zhu, Xusen Sun, and Xun Cao. Mofanerf: Morphable facial neural radiance field. In European Conference on Computer Vision (ECCV), 2022. 3 \ No newline at end of file diff --git a/2025/Coherent 3D Portrait Video Reconstruction via Triplane Fusion/images.zip b/2025/Coherent 3D Portrait Video Reconstruction via Triplane Fusion/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..0b69c648e9691b76984a534e9ffcf332a450efb2 --- /dev/null +++ b/2025/Coherent 3D Portrait Video Reconstruction via Triplane Fusion/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:52c7e81b8d1f3adea40fe4ff3722c50728e819baae18667cc84808f358f1cd19 +size 474951 diff --git a/2025/Coherent 3D Portrait Video Reconstruction via Triplane Fusion/layout.json b/2025/Coherent 3D Portrait Video Reconstruction via Triplane Fusion/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..f28bf37740b537d2c1da5028ee6bcaf20d41c580 --- /dev/null +++ b/2025/Coherent 3D Portrait Video Reconstruction via Triplane Fusion/layout.json @@ -0,0 +1,9716 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 109, + 102, + 501, + 120 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 109, + 102, + 501, + 120 + ], + "spans": [ + { + "bbox": [ + 109, + 102, + 501, + 120 + ], + "type": "text", + "content": "Coherent 3D Portrait Video Reconstruction via Triplane Fusion" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 71, + 141, + 538, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 141, + 538, + 199 + ], + "spans": [ + { + "bbox": [ + 71, + 141, + 538, + 199 + ], + "type": "text", + "content": "Shengze Wang" + }, + { + "bbox": [ + 71, + 141, + 538, + 199 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 71, + 141, + 538, + 199 + ], + "type": "text", + "content": " Xueting Li" + }, + { + "bbox": [ + 71, + 141, + 538, + 199 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 71, + 141, + 538, + 199 + ], + "type": "text", + "content": " Chao Liu" + }, + { + "bbox": [ + 71, + 141, + 538, + 199 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 71, + 141, + 538, + 199 + ], + "type": "text", + "content": " Matthew Chan" + }, + { + "bbox": [ + 71, + 141, + 538, + 199 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 71, + 141, + 538, + 199 + ], + "type": "text", + "content": " Michael Stengel" + }, + { + "bbox": [ + 71, + 141, + 538, + 199 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 71, + 141, + 538, + 199 + ], + "type": "text", + "content": " Henry Fuchs" + }, + { + "bbox": [ + 71, + 141, + 538, + 199 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 71, + 141, + 538, + 199 + ], + "type": "text", + "content": " Shalini De Mello" + }, + { + "bbox": [ + 71, + 141, + 538, + 199 + ], + "type": "inline_equation", + "content": "^{2\\dagger}" + }, + { + "bbox": [ + 71, + 141, + 538, + 199 + ], + "type": "text", + "content": " Koki Nagano" + }, + { + "bbox": [ + 71, + 141, + 538, + 199 + ], + "type": "inline_equation", + "content": "^{2\\dagger}" + }, + { + "bbox": [ + 71, + 141, + 538, + 199 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 71, + 141, + 538, + 199 + ], + "type": "text", + "content": " UNC Chapel Hill " + }, + { + "bbox": [ + 71, + 141, + 538, + 199 + ], + "type": "inline_equation", + "content": "{}^{2}" + }, + { + "bbox": [ + 71, + 141, + 538, + 199 + ], + "type": "text", + "content": " NVIDIA " + }, + { + "bbox": [ + 71, + 141, + 538, + 199 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 71, + 141, + 538, + 199 + ], + "type": "text", + "content": " Equal Contribution \nhttps://research.nvidia.com/labs/amri/projects/coherent3d/" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 151, + 225, + 200, + 238 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 225, + 200, + 238 + ], + "spans": [ + { + "bbox": [ + 151, + 225, + 200, + 238 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 251, + 297, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 251, + 297, + 529 + ], + "spans": [ + { + "bbox": [ + 54, + 251, + 297, + 529 + ], + "type": "text", + "content": "Recent breakthroughs in single-image 3D portrait reconstruction have enabled telepresence systems to stream 3D portrait videos from a single camera in real-time, democratizing telepresence. However, per-frame 3D reconstruction exhibits temporal inconsistency and forgets the user's appearance. On the other hand, self-reenactment methods can render coherent 3D portraits by driving a 3D avatar built from a single reference image but fail to faithfully preserve the user's per-frame appearance (e.g., instantaneous facial expressions and lighting). As a result, neither of these two frameworks is an ideal solution for democratized 3D telepresence. In this work, we address this dilemma and propose a novel solution that maintains both coherent identity and dynamic per-frame appearance to enable the best possible realism. To this end, we propose a new fusion-based method that takes the best of both worlds by fusing a canonical 3D prior from a reference view with dynamic appearance from per-frame input views, producing temporally stable 3D videos with faithful reconstruction of the user's per-frame appearance. Trained only using synthetic data produced by an expression-conditioned 3D GAN, our encoder-based method achieves both state-of-the-art 3D reconstruction and temporal consistency on in-studio and in-the-wild datasets." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 539, + 135, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 539, + 135, + 552 + ], + "spans": [ + { + "bbox": [ + 55, + 539, + 135, + 552 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 554, + 296, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 554, + 296, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 554, + 296, + 693 + ], + "type": "text", + "content": "Telepresence aims at bringing distant people face-to-face and stands out as a particularly compelling application of computer vision and graphics. Over the last decades, various successful telepresence systems [19, 21, 21, 25, 31, 33, 37, 43] have been developed. However, most employ bulky multi-view 3D scanners or depth sensors to ensure high-quality volumetric per-frame reconstruction. Unlike these classical 3D/4D reconstruction methods, recent AI-based feed-forward 3D lifting techniques [2, 52] can lift a single RGB image from an off-the-shelf webcam into a neural radiance field (NeRF) representation encoded into a set of triplanes in real time, paving the way towards making 3D" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 226, + 473, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 226, + 473, + 239 + ], + "spans": [ + { + "bbox": [ + 314, + 226, + 473, + 239 + ], + "type": "text", + "content": "telepresence accessible to anyone [48]." + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 316, + 248, + 555, + 472 + ], + "blocks": [ + { + "bbox": [ + 316, + 248, + 555, + 472 + ], + "lines": [ + { + "bbox": [ + 316, + 248, + 555, + 472 + ], + "spans": [ + { + "bbox": [ + 316, + 248, + 555, + 472 + ], + "type": "image", + "image_path": "af1c171076518b0c923ef4c3c10de2463aa514630b3845c04053af66948df8fa.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 479, + 555, + 598 + ], + "lines": [ + { + "bbox": [ + 313, + 479, + 555, + 598 + ], + "spans": [ + { + "bbox": [ + 313, + 479, + 555, + 598 + ], + "type": "text", + "content": "Figure 1. Comparisons. Given a single reference image and a single-view video frame, our method reconstructs the authentic dynamic appearance of the user (e.g., facial expressions and lighting) while producing a temporally coherent 3D video. A previous single-view 3D lifting method (LP3D) that reconstructs the avatar from the video frame on a per-frame basis suffers from distortions and temporal inconsistency. A portrait reenactment method (GPAvatar) drives the identity in the reference image using the video frame, but fails to capture accurate facial expressions (e.g., smile) and per-frame appearance (e.g., lighting). The output should be compared to the appearance of the input video frame (green box)." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 609, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 609, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 609, + 556, + 715 + ], + "type": "text", + "content": "Currently, there are two major paradigms in democratized 3D telepresence solutions from a single-view video: (1) single-view per-frame 3D lifting methods and (2) 3D portrait reenactment, which drives an identity from a reference image using another driving frame, but none of them is an ideal solution. For (1), single-frame-based lifting techniques such as LP3D [52], have the advantage of faithfully preserving the instantaneous dynamic conditions present in an input video, e.g. lighting, expressions, and posture, all" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 702, + 261, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 702, + 261, + 713 + ], + "spans": [ + { + "bbox": [ + 66, + 702, + 261, + 713 + ], + "type": "text", + "content": "*Shengze Wang was an intern at NVIDIA during the project" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "10712" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 338 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 338 + ], + "type": "text", + "content": "of which are crucial to an authentic telepresence experience. However, single-image reconstruction methods operate independently on each frame and thus have fundamental limitations in maintaining temporal consistency. This difficulty stems from the inherent ill-posed nature of single-image-based reconstruction. In order to render novel views that are significantly far from the input view, the system cannot rely on information present in the input view and hence must hallucinate generate plausible content, which cannot be guaranteed to be consistent across multiple temporal frames. For example, LP3D's reconstruction changes significantly depending on the user's head pose in the input frame (see second row in Fig. 1 and third column in Fig. 2). In comparison, 3D self-reenactment methods create an avatar model from one or multiple reference images and use a separate driving video to drive the facial expressions and poses of the avatar. [6, 27, 51, 59]. While these approaches allow for temporally consistent results, they often cannot faithfully reconstruct the input video's dynamic conditions such as changes in lighting. Moreover, reenactment methods often struggle to generate accurate expressions because their expression detection and control are not precise enough (see the first row in Fig. 1)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 339, + 296, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 339, + 296, + 443 + ], + "spans": [ + { + "bbox": [ + 55, + 339, + 296, + 443 + ], + "type": "text", + "content": "In this work, we address this dilemma of democratized 3D telepresence approaches and propose a novel solution to the problem of simultaneously maintaining temporal stability while preserving real-time dynamics of input videos for single-camera telepresence applications. Our proposed solution is a fusion-based approach that leverages the stability and accuracy of a canonical 3D prior, and also captures the diverse deviations from the prior in newer video frames (see the third row in Fig. 1)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 444, + 296, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 444, + 296, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 444, + 296, + 628 + ], + "type": "text", + "content": "Our model first uses LP3D [52] to construct a canonical triplane prior from a (near) frontal image of the user, which can be casually captured or extracted from a video. During video reconstruction, our model lifts each input frame into a raw triplane, which is then fused with the canonical triplane (see Fig. 3). For images with oblique yaw head poses, artifacts and identity-related distortions may be present in its lifted triplane (see \"LP3D\" in Fig. 1 and 2). Hence we propose an undistorter module, which learns to undistort the raw instantaneous triplane to more closely match the identity of the correctly structured canonical triplane. We then propose a fuser module that combines the undistorted triplane and the canonical triplane in order to preserve identity consistency and recover occluded regions while reconstructing the dynamic lighting, nuanced expressions, and posture information in the input frame." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 629, + 246, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 629, + 246, + 640 + ], + "spans": [ + { + "bbox": [ + 67, + 629, + 246, + 640 + ], + "type": "text", + "content": "We summarize our contributions as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 641, + 295, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 295, + 688 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 295, + 688 + ], + "type": "text", + "content": "- We contribute a novel triplane fusion method that combines the dynamic information from per-frame triplanes with a canonical triplane extracted from a reference image. Trained only using a synthetic multi-view video" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 321, + 72, + 555, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 72, + 555, + 131 + ], + "spans": [ + { + "bbox": [ + 321, + 72, + 555, + 131 + ], + "type": "text", + "content": "dataset, our feed-forward approach generates 3D portrait videos that demonstrate both temporal consistency and faithful reconstruction of the dynamic appearance of the user (e.g. lighting and expression), whereas prior solutions can only achieve one of the two properties." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 131, + 555, + 223 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 314, + 131, + 555, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 131, + 555, + 176 + ], + "spans": [ + { + "bbox": [ + 314, + 131, + 555, + 176 + ], + "type": "text", + "content": "- We propose a new framework to evaluate single-view 3D portrait reconstruction methods using multi-view data and gain insight into the method's reconstruction quality and robustness." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 177, + 554, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 177, + 554, + 223 + ], + "spans": [ + { + "bbox": [ + 314, + 177, + 554, + 223 + ], + "type": "text", + "content": "- We present evaluations on both in-studio and in-the-wild data and demonstrate that our method achieves state-of-the-art performance in terms of temporal consistency and reconstruction accuracy." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "type": "image", + "bbox": [ + 318, + 236, + 551, + 415 + ], + "blocks": [ + { + "bbox": [ + 318, + 236, + 551, + 415 + ], + "lines": [ + { + "bbox": [ + 318, + 236, + 551, + 415 + ], + "spans": [ + { + "bbox": [ + 318, + 236, + 551, + 415 + ], + "type": "image", + "image_path": "10853aafdfdf15407b71a6bf588e4b8fab5223564dfef49a2ada7f3d83b9a07b.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 425, + 555, + 521 + ], + "lines": [ + { + "bbox": [ + 313, + 425, + 555, + 521 + ], + "spans": [ + { + "bbox": [ + 313, + 425, + 555, + 521 + ], + "type": "text", + "content": "Figure 2. View-Dependent Distortion. \"NV\" refers to \"Novel View Rendering\". Top: inputs to our model and LP3D. Second & Third Rows: LP3D's reconstruction varies greatly under challenging viewpoints, showing a predictable pattern of artifacts including abnormally strong activations on the side being captured (red circle), as well as geometric distortion along the view direction of the camera due to depth ambiguity. We refer to this phenomenon as \"View-Dependent Distortion\". Fourth row: Our method removes such artifacts and achieves better coherence." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 529, + 400, + 542 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 529, + 400, + 542 + ], + "spans": [ + { + "bbox": [ + 314, + 529, + 400, + 542 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 550, + 554, + 631 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 550, + 554, + 631 + ], + "spans": [ + { + "bbox": [ + 313, + 550, + 554, + 631 + ], + "type": "text", + "content": "2D portrait reenactment. Given a single or a few reference portrait images and a driving video, recent talking-head generators can reenact 2D portraits by transferring the facial expressions and poses from the driving video onto facial portraits [9, 10, 17, 46, 53, 54, 60, 62-64, 66]. However, being 2D, they cannot be rendered from novel viewpoints, which is crucial for 3D telepresence." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 632, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 632, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 632, + 555, + 713 + ], + "type": "text", + "content": "3D-aware portrait generation and reenactment. Some recent works use deformable volumetric implicit radiance fields [34, 38, 39] or Gaussian splatting [22] combined by 3DMMs to reconstruct a photorealistic and animatable volumetric head avatar [1, 4, 14, 41, 45, 57, 67]. However, they require extensive data captures from videos or multiview cameras and person-specific training. Others use" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "type": "text", + "content": "1We find that the 3D lifting from a near frontal reference view is reliable, hence use this as a canonical 3D prior. See Fig. 2 first column." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10713" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 188 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 188 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 188 + ], + "type": "text", + "content": "large-scale video datasets and learn a disentangled triplane 3D [5] for 3D facial reenactment in a feedforward fashion [6, 27, 28, 32, 51, 59, 61]. They construct a canonical 3D head from a reference image (often a neutral frame), and use facial expressions and head poses extracted from a separate driving video to animate it. As such, fine-grained facial expressions may not be captured due to errors in disentanglement. Most importantly, these reenactment methods fail to preserve the dynamic appearance of users (e.g., person-specific wrinkles or lighting) across time." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 189, + 298, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 189, + 298, + 468 + ], + "spans": [ + { + "bbox": [ + 56, + 189, + 298, + 468 + ], + "type": "text", + "content": "3D GAN inversion. By combining GANs [15] and neural volume rendering [34], recent breakthroughs in 3D-aware GANs [5, 8, 16, 36, 44, 47, 55, 56, 58, 65, 68] can learn to generate photorealistic 3D heads from in-the-wild 2D images. Notably, EG3D proposes an efficient and compact triplane representation [5] to generate 3D heads, and Next3D [50] extends EG3D to create 3D portrait videos controlled by 3DMM facial expression and pose parameters. Besides generating 3D heads, these models can also be used for single-view 3D reconstruction using GAN inversion [13, 24, 30, 49], manipulation of the 3D head [18, 49, 69], and 3D personalization [3, 40]. However, inverting a few seconds of video can take many minutes or hours, and the inversion quality is often unsatisfactory due to inaccurate identity and expression preservation, as evaluated by GPAvatar [6]. Therefore, recent works [2, 52] propose encoder-based solutions to lift each video frame into a triplane. However, when each frame is independently lifted into 3D, the resulting 3D head video suffers in terms of temporal consistency - a key requirement of 3D telepresence systems. To enhance single-frame 2D-to-3D encoders for human heads, e.g., LP3D [52], we propose a triplane-fusion-based method, which improves their temporal stability while preserving temporal dynamics across time." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 478, + 111, + 490 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 478, + 111, + 490 + ], + "spans": [ + { + "bbox": [ + 55, + 478, + 111, + 490 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 497, + 294, + 510 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 497, + 294, + 510 + ], + "spans": [ + { + "bbox": [ + 55, + 497, + 294, + 510 + ], + "type": "text", + "content": "3.1. Background: 3D Portrait from a Single Image" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 514, + 295, + 631 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 514, + 295, + 631 + ], + "spans": [ + { + "bbox": [ + 55, + 514, + 295, + 631 + ], + "type": "text", + "content": "LP3D [52] performs photorealistic 3D portrait reconstruction by using a feedforward encoder to lift an RGB image into a triplane " + }, + { + "bbox": [ + 55, + 514, + 295, + 631 + ], + "type": "inline_equation", + "content": "\\mathbf{T} \\in \\mathbb{R}^{3 \\times 32 \\times 256 \\times 256}" + }, + { + "bbox": [ + 55, + 514, + 295, + 631 + ], + "type": "text", + "content": ", which can be volume rendered to an RGB image from any viewpoint. LP3D can run in real-time and has been developed into a complete realtime telepresence system [48]. We lift 2D faces into triplanes with a slightly modified LP3D trained on larger face crops containing shoulders and a camera estimator to recover the input image's camera parameters " + }, + { + "bbox": [ + 55, + 514, + 295, + 631 + ], + "type": "inline_equation", + "content": "M \\in \\mathbb{R}^{25}" + }, + { + "bbox": [ + 55, + 514, + 295, + 631 + ], + "type": "text", + "content": ". It performs better than the original version (Table 3)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 638, + 129, + 650 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 638, + 129, + 650 + ], + "spans": [ + { + "bbox": [ + 55, + 638, + 129, + 650 + ], + "type": "text", + "content": "3.2. Definitions" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 654, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 295, + 713 + ], + "type": "text", + "content": "We call a current video frame an \"input frame\" into our system. Additionally, we encode a near-frontal \"reference image\" of the subject into a \"canonical triplane\", which is used to stabilize the 3D video generation process. Lastly, we refer to the viewpoint of the camera in the input frame" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 72, + 406, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 406, + 84 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 406, + 84 + ], + "type": "text", + "content": "as \"input viewpoint\"." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 91, + 382, + 103 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 91, + 382, + 103 + ], + "spans": [ + { + "bbox": [ + 314, + 91, + 382, + 103 + ], + "type": "text", + "content": "3.3. Overview" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 108, + 555, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 108, + 555, + 295 + ], + "spans": [ + { + "bbox": [ + 313, + 108, + 555, + 295 + ], + "type": "text", + "content": "An overview of our method is illustrated in Fig. 3. To improve temporal consistency and reconstruct occluded parts of the face, we leverage an additional near-frontal reference image, which can be a selfie image or a video frame automatically obtained from the same video using an off-the-shelf head pose estimator like DECA [11]. We lift this reference image into the canonical triplane using pre-trained LP3D. When processing the video, we first lift each input frame into a raw triplane. Then, the Triplane Undistorter (Sec. 3.5) removes view-dependent distortions and recovers the identity using the canonical triplane as a reference. This produces an undistorted triplane. Finally, to recover regions that are occluded in the input frame and further improve stability, our Triplane Fuser (Sec. 3.6) combines the undistorted triplane with the canonical triplane to generate the final coherent triplane." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 301, + 555, + 314 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 301, + 555, + 314 + ], + "spans": [ + { + "bbox": [ + 313, + 301, + 555, + 314 + ], + "type": "text", + "content": "3.4. Generating Synthetic Dynamic Multiview Data" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 319, + 556, + 585 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 319, + 556, + 585 + ], + "spans": [ + { + "bbox": [ + 313, + 319, + 556, + 585 + ], + "type": "text", + "content": "We synthesize ground truth multiview training images using Next3D [50]. We first extract 2D landmarks and FLAME [26] coefficient labels for the FFHQ [20] dataset using DECA [12]. Then, during training, we sample a pair of random FLAME coefficients from our FFHQ labels and input them to Next3D alongside a single random identity code " + }, + { + "bbox": [ + 313, + 319, + 556, + 585 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 313, + 319, + 556, + 585 + ], + "type": "text", + "content": ". We notice that FFHQ contains mostly less extreme expressions. Thus, we randomly multiply the expression codes by a small scaling factor to exaggerate the expressions during training. Notice that the Next3D generator is much more expressive than FLAME due to its use of triplanes and GAN training even though it is conditioned on FLAME. Next3D then generates a pair of triplanes for " + }, + { + "bbox": [ + 313, + 319, + 556, + 585 + ], + "type": "inline_equation", + "content": "t = 0" + }, + { + "bbox": [ + 313, + 319, + 556, + 585 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 319, + 556, + 585 + ], + "type": "inline_equation", + "content": "t = 1" + }, + { + "bbox": [ + 313, + 319, + 556, + 585 + ], + "type": "text", + "content": " of the same person with 2 expressions. We render the " + }, + { + "bbox": [ + 313, + 319, + 556, + 585 + ], + "type": "inline_equation", + "content": "t = 0" + }, + { + "bbox": [ + 313, + 319, + 556, + 585 + ], + "type": "text", + "content": " triplane into a near-frontal reference image " + }, + { + "bbox": [ + 313, + 319, + 556, + 585 + ], + "type": "inline_equation", + "content": "I_{ref}" + }, + { + "bbox": [ + 313, + 319, + 556, + 585 + ], + "type": "text", + "content": " (Fig. 3 orange box) and the " + }, + { + "bbox": [ + 313, + 319, + 556, + 585 + ], + "type": "inline_equation", + "content": "t = 1" + }, + { + "bbox": [ + 313, + 319, + 556, + 585 + ], + "type": "text", + "content": " triplane into 3 images: the input frame rendered from a random viewpoint (green box), the ground truth image at a different viewpoint (red box), and a frontal-view image (red box). Only the latter two are used for supervision. Additionally, to simulate different lighting conditions, we also apply different color augmentations (brightness, contrast, saturation, and hue) to images at " + }, + { + "bbox": [ + 313, + 319, + 556, + 585 + ], + "type": "inline_equation", + "content": "t = 0" + }, + { + "bbox": [ + 313, + 319, + 556, + 585 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 319, + 556, + 585 + ], + "type": "inline_equation", + "content": "t = 1" + }, + { + "bbox": [ + 313, + 319, + 556, + 585 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 586, + 555, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 586, + 555, + 677 + ], + "spans": [ + { + "bbox": [ + 313, + 586, + 555, + 677 + ], + "type": "text", + "content": "Shoulder augmentation. It is important that 3D portraits capture dynamic shoulder movements to convey body language and achieve eye contact in telepresence. However, Next3D does not provide control over shoulder posture. Thus, we simulate shoulder movement in the rendered images without modifying the Next3D triplanes by warping camera rays during volume rendering. Please see the supplement for more detail." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 678, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 678, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 678, + 555, + 713 + ], + "type": "text", + "content": "Pseudo-groundtruth triplanes. Due to our shoulder augmentation, the Next3D triplanes and their 2D renderings become inconsistent. Thus, Next3D's triplanes cannot be used" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "10714" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 65, + 76, + 541, + 337 + ], + "blocks": [ + { + "bbox": [ + 65, + 76, + 541, + 337 + ], + "lines": [ + { + "bbox": [ + 65, + 76, + 541, + 337 + ], + "spans": [ + { + "bbox": [ + 65, + 76, + 541, + 337 + ], + "type": "image", + "image_path": "d9070d5ebef1551b6b6a61f0bebca5b7fdd80545b2baa92dc0c5b294d45db6ea.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 343, + 555, + 389 + ], + "lines": [ + { + "bbox": [ + 54, + 343, + 555, + 389 + ], + "spans": [ + { + "bbox": [ + 54, + 343, + 555, + 389 + ], + "type": "text", + "content": "Figure 3. Method Overview. Given a near-frontal reference image and an input frame, we first reconstruct a canonical triplane and a raw triplane using LP3D [52] (Sec. 3.1). Next, we combine them through a Triplane Fusion module (blue box) that ensures temporal consistency while preserving dynamic information (e.g., lighting and shoulder pose) (Sec. 3.5 and 3.6). Our model is trained with only synthetic video data generated by a 3D GAN [50], with carefully designed augmentations to preserve shoulder motion and lighting (Sec. 3.4)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 393, + 298, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 393, + 298, + 453 + ], + "spans": [ + { + "bbox": [ + 54, + 393, + 298, + 453 + ], + "type": "text", + "content": "as direct supervisory signals. To mitigate this, we leverage the fact that LP3D generates reasonably accurate triplanes from frontal view images. We use a frozen LP3D to predict pseudo-groundtruth triplanes " + }, + { + "bbox": [ + 54, + 393, + 298, + 453 + ], + "type": "inline_equation", + "content": "T_{\\text{triplaneGT}}" + }, + { + "bbox": [ + 54, + 393, + 298, + 453 + ], + "type": "text", + "content": " from the frontal novel view " + }, + { + "bbox": [ + 54, + 393, + 298, + 453 + ], + "type": "inline_equation", + "content": "I_{\\text{frontal}}" + }, + { + "bbox": [ + 54, + 393, + 298, + 453 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 54, + 393, + 298, + 453 + ], + "type": "inline_equation", + "content": "t = 1" + }, + { + "bbox": [ + 54, + 393, + 298, + 453 + ], + "type": "text", + "content": " (Fig. 3 bottom right)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 461, + 289, + 475 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 461, + 289, + 475 + ], + "spans": [ + { + "bbox": [ + 55, + 461, + 289, + 475 + ], + "type": "text", + "content": "3.5. Removing Distortion and Preserving Identity" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 479, + 295, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 479, + 295, + 653 + ], + "spans": [ + { + "bbox": [ + 54, + 479, + 295, + 653 + ], + "type": "text", + "content": "LP3D's reconstruction quality strongly depends on the input frame's viewpoint. When subjects are captured from the side, LP3D tends to produce errors in identity and exhibits stretching distortions (e.g., in the \"Input Frame 1\" and \"Input Frame 2\" columns in Fig. 2), primarily due to the inherent ambiguity of single-image reconstruction. In contrast, near-frontal views offer more complete identity cues and suffer less from occlusion (e.g. the \"Reference Image\" column in Fig. 2), leading to more reliable reconstruction. To reduce the single-image ambiguity, we adopt a near-frontal reconstruction as the canonical triplane " + }, + { + "bbox": [ + 54, + 479, + 295, + 653 + ], + "type": "inline_equation", + "content": "T_{\\text{cano}}" + }, + { + "bbox": [ + 54, + 479, + 295, + 653 + ], + "type": "text", + "content": ". Although incorporating more reference views could potentially improve accuracy, we find that using a single near-frontal view not only simplifies the model but also sufficiently enhances temporal coherence." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 654, + 295, + 690 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 295, + 690 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 295, + 690 + ], + "type": "text", + "content": "To correct the distortion in an input raw triplane " + }, + { + "bbox": [ + 55, + 654, + 295, + 690 + ], + "type": "inline_equation", + "content": "T_{raw}" + }, + { + "bbox": [ + 55, + 654, + 295, + 690 + ], + "type": "text", + "content": " using the canonical triplane " + }, + { + "bbox": [ + 55, + 654, + 295, + 690 + ], + "type": "inline_equation", + "content": "T_{cano}" + }, + { + "bbox": [ + 55, + 654, + 295, + 690 + ], + "type": "text", + "content": " as reference, we devise a Triplane Undistorter " + }, + { + "bbox": [ + 55, + 654, + 295, + 690 + ], + "type": "inline_equation", + "content": "U" + }, + { + "bbox": [ + 55, + 654, + 295, + 690 + ], + "type": "text", + "content": " (Fig. 3):" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 80, + 700, + 295, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 700, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 80, + 700, + 295, + 715 + ], + "type": "interline_equation", + "content": "U \\left(T _ {\\text {r a w}}, T _ {\\text {c a n o}}\\right) = T _ {\\text {u n d i s t}} \\in \\mathbb {R} ^ {3 \\times 3 2 \\times 2 5 6 \\times 2 5 6}. \\tag {1}", + "image_path": "0206004795c9d7bc8fa81e3dcd04732ab15371212c52fefd47ab9240137a34ba.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 393, + 555, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 393, + 555, + 475 + ], + "spans": [ + { + "bbox": [ + 313, + 393, + 555, + 475 + ], + "type": "text", + "content": "Since the distortion in each triplane is a 2D warping artifact, it can be reversed through a 2D undistortion warping of each plane. To this end, our Undistorter " + }, + { + "bbox": [ + 313, + 393, + 555, + 475 + ], + "type": "inline_equation", + "content": "U" + }, + { + "bbox": [ + 313, + 393, + 555, + 475 + ], + "type": "text", + "content": " adopts the SPyNet[42] architecture to predict a 2D correction warping " + }, + { + "bbox": [ + 313, + 393, + 555, + 475 + ], + "type": "inline_equation", + "content": "T_{corr} \\in \\mathbb{R}^{3 \\times 2 \\times 256 \\times 256}" + }, + { + "bbox": [ + 313, + 393, + 555, + 475 + ], + "type": "text", + "content": " for the three planes. Then, the Undistorter corrects the raw triplane " + }, + { + "bbox": [ + 313, + 393, + 555, + 475 + ], + "type": "inline_equation", + "content": "T_{raw}" + }, + { + "bbox": [ + 313, + 393, + 555, + 475 + ], + "type": "text", + "content": " by warping it based on the predicted 2D offsets " + }, + { + "bbox": [ + 313, + 393, + 555, + 475 + ], + "type": "inline_equation", + "content": "(\\Delta u, \\Delta v)" + }, + { + "bbox": [ + 313, + 393, + 555, + 475 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 313, + 393, + 555, + 475 + ], + "type": "inline_equation", + "content": "T_{corr}" + }, + { + "bbox": [ + 313, + 393, + 555, + 475 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 366, + 479, + 555, + 491 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 366, + 479, + 555, + 491 + ], + "spans": [ + { + "bbox": [ + 366, + 479, + 555, + 491 + ], + "type": "interline_equation", + "content": "T _ {c o r r} = S P y N e t \\left(T _ {\\text {r a w}}, T _ {\\text {c a n o}}\\right), \\tag {2}", + "image_path": "6a8f9d480a4af0807c34962616f8780072cd7a5928c9ea4207316f62fde5a5a9.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 369, + 495, + 553, + 508 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 369, + 495, + 553, + 508 + ], + "spans": [ + { + "bbox": [ + 369, + 495, + 553, + 508 + ], + "type": "interline_equation", + "content": "T _ {u n d i s t} = \\operatorname {W a r p} \\left(T _ {\\text {r a w}}, T _ {\\text {c o r r}}\\right). \\tag {3}", + "image_path": "0f32f947ab38c464b9a237de7ef738e7a114f25dd3b74808e587c583d725c94f.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 512, + 556, + 699 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 512, + 556, + 699 + ], + "spans": [ + { + "bbox": [ + 313, + 512, + 556, + 699 + ], + "type": "text", + "content": "It is important to note the difference between optical flow and undistortion. An ideal optical flow model would warp " + }, + { + "bbox": [ + 313, + 512, + 556, + 699 + ], + "type": "inline_equation", + "content": "T_{\\text{raw}}" + }, + { + "bbox": [ + 313, + 512, + 556, + 699 + ], + "type": "text", + "content": " towards " + }, + { + "bbox": [ + 313, + 512, + 556, + 699 + ], + "type": "inline_equation", + "content": "T_{\\text{cano}}" + }, + { + "bbox": [ + 313, + 512, + 556, + 699 + ], + "type": "text", + "content": " in order to make the two identical. However, this alignment leads to significant artifacts and destroys the reconstruction (Table. 3 and Fig. 5). This is because the purpose of undistortion is not to make the two triplanes identical, but to correct " + }, + { + "bbox": [ + 313, + 512, + 556, + 699 + ], + "type": "inline_equation", + "content": "T_{\\text{raw}}" + }, + { + "bbox": [ + 313, + 512, + 556, + 699 + ], + "type": "text", + "content": " such that it has the same identity as the canonical triplane " + }, + { + "bbox": [ + 313, + 512, + 556, + 699 + ], + "type": "inline_equation", + "content": "T_{\\text{cano}}" + }, + { + "bbox": [ + 313, + 512, + 556, + 699 + ], + "type": "text", + "content": ", while preserving dynamic information such as expressions and lighting. Therefore, instead of using the canonical triplane " + }, + { + "bbox": [ + 313, + 512, + 556, + 699 + ], + "type": "inline_equation", + "content": "T_{\\text{cano}}" + }, + { + "bbox": [ + 313, + 512, + 556, + 699 + ], + "type": "text", + "content": " as the warping target, our Undistorter " + }, + { + "bbox": [ + 313, + 512, + 556, + 699 + ], + "type": "inline_equation", + "content": "U" + }, + { + "bbox": [ + 313, + 512, + 556, + 699 + ], + "type": "text", + "content": " merely uses " + }, + { + "bbox": [ + 313, + 512, + 556, + 699 + ], + "type": "inline_equation", + "content": "T_{\\text{cano}}" + }, + { + "bbox": [ + 313, + 512, + 556, + 699 + ], + "type": "text", + "content": " as the identity conditioning to correct the raw triplane " + }, + { + "bbox": [ + 313, + 512, + 556, + 699 + ], + "type": "inline_equation", + "content": "T_{\\text{raw}}" + }, + { + "bbox": [ + 313, + 512, + 556, + 699 + ], + "type": "text", + "content": " towards the ground truth triplane " + }, + { + "bbox": [ + 313, + 512, + 556, + 699 + ], + "type": "inline_equation", + "content": "T_{\\text{triplaneGT}}" + }, + { + "bbox": [ + 313, + 512, + 556, + 699 + ], + "type": "text", + "content": " (Fig. 3, lower-right), which is not available during test time. The correction warping is supervised by the consistency between the undistorted triplane " + }, + { + "bbox": [ + 313, + 512, + 556, + 699 + ], + "type": "inline_equation", + "content": "T_{\\text{undist}}" + }, + { + "bbox": [ + 313, + 512, + 556, + 699 + ], + "type": "text", + "content": " and pseudo-groundtruth tri" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "10715" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 247, + 85 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 247, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 247, + 85 + ], + "type": "text", + "content": "plane " + }, + { + "bbox": [ + 55, + 72, + 247, + 85 + ], + "type": "inline_equation", + "content": "T_{\\text{triplane}GT}" + }, + { + "bbox": [ + 55, + 72, + 247, + 85 + ], + "type": "text", + "content": " (Sec. 3.4) via a triplane loss:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 100, + 92, + 294, + 106 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 92, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 100, + 92, + 294, + 106 + ], + "type": "interline_equation", + "content": "L _ {u n d i s t} = L _ {1} \\left(T _ {u n d i s t}, T _ {t r i p l a n e G T}\\right). \\tag {4}", + "image_path": "9f616c8d9f7164deb6e427f3d4940e40e903d774f0c6818ab8dd8fccb0dd260d.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 110, + 295, + 134 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 110, + 295, + 134 + ], + "spans": [ + { + "bbox": [ + 55, + 110, + 295, + 134 + ], + "type": "text", + "content": "3.6. Reconstructing Occluded Regions through Triplane Fusion" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 138, + 296, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 138, + 296, + 350 + ], + "spans": [ + { + "bbox": [ + 55, + 138, + 296, + 350 + ], + "type": "text", + "content": "As the user moves around in the video, different parts of their head become occluded. To recover occluded areas in the input frame and further stabilize the subject's identity across the video, our Fuser " + }, + { + "bbox": [ + 55, + 138, + 296, + 350 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 55, + 138, + 296, + 350 + ], + "type": "text", + "content": " enhances the reconstruction by utilizing the canonical triplane " + }, + { + "bbox": [ + 55, + 138, + 296, + 350 + ], + "type": "inline_equation", + "content": "T_{\\text{cano}}" + }, + { + "bbox": [ + 55, + 138, + 296, + 350 + ], + "type": "text", + "content": ", in which the currently occluded areas are often visible. Therefore, it is important for the Fuser " + }, + { + "bbox": [ + 55, + 138, + 296, + 350 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 55, + 138, + 296, + 350 + ], + "type": "text", + "content": " to identify and recover the occluded regions while preserving information from the visible regions in " + }, + { + "bbox": [ + 55, + 138, + 296, + 350 + ], + "type": "inline_equation", + "content": "T_{\\text{undist}}" + }, + { + "bbox": [ + 55, + 138, + 296, + 350 + ], + "type": "text", + "content": ". To accomplish this, we thus use a 5-layer ConvNet-based visibility estimator " + }, + { + "bbox": [ + 55, + 138, + 296, + 350 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 55, + 138, + 296, + 350 + ], + "type": "text", + "content": " to predict a visibility triplane as " + }, + { + "bbox": [ + 55, + 138, + 296, + 350 + ], + "type": "inline_equation", + "content": "V(T_{\\text{raw}}) = T_{\\text{raw}}^{\\text{vis}} \\in \\mathbb{R}^{3 \\times 1 \\times 128 \\times 128}" + }, + { + "bbox": [ + 55, + 138, + 296, + 350 + ], + "type": "text", + "content": ", i.e., one visibility map for each plane. " + }, + { + "bbox": [ + 55, + 138, + 296, + 350 + ], + "type": "inline_equation", + "content": "T_{\\text{raw}}^{\\text{vis}}" + }, + { + "bbox": [ + 55, + 138, + 296, + 350 + ], + "type": "text", + "content": " is undistorted alongside " + }, + { + "bbox": [ + 55, + 138, + 296, + 350 + ], + "type": "inline_equation", + "content": "T_{\\text{undist}}" + }, + { + "bbox": [ + 55, + 138, + 296, + 350 + ], + "type": "text", + "content": " to produce " + }, + { + "bbox": [ + 55, + 138, + 296, + 350 + ], + "type": "inline_equation", + "content": "T_{\\text{undist}}^{\\text{vis}}" + }, + { + "bbox": [ + 55, + 138, + 296, + 350 + ], + "type": "text", + "content": ". We also predict a visibility triplane for the canonical triplane as " + }, + { + "bbox": [ + 55, + 138, + 296, + 350 + ], + "type": "inline_equation", + "content": "T_{\\text{cano}}^{\\text{vis}} = V(T_{\\text{cano}})" + }, + { + "bbox": [ + 55, + 138, + 296, + 350 + ], + "type": "text", + "content": ". Finally, the Fuser " + }, + { + "bbox": [ + 55, + 138, + 296, + 350 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 55, + 138, + 296, + 350 + ], + "type": "text", + "content": " produces the fused triplane " + }, + { + "bbox": [ + 55, + 138, + 296, + 350 + ], + "type": "inline_equation", + "content": "T_{\\text{fused}}" + }, + { + "bbox": [ + 55, + 138, + 296, + 350 + ], + "type": "text", + "content": " by combining information from the undistorted input tranplane " + }, + { + "bbox": [ + 55, + 138, + 296, + 350 + ], + "type": "inline_equation", + "content": "T_{\\text{undist}}" + }, + { + "bbox": [ + 55, + 138, + 296, + 350 + ], + "type": "text", + "content": ", its visibility triplane " + }, + { + "bbox": [ + 55, + 138, + 296, + 350 + ], + "type": "inline_equation", + "content": "T_{\\text{undist}}^{\\text{vis}}" + }, + { + "bbox": [ + 55, + 138, + 296, + 350 + ], + "type": "text", + "content": ", the canonical triplane " + }, + { + "bbox": [ + 55, + 138, + 296, + 350 + ], + "type": "inline_equation", + "content": "T_{\\text{cano}}" + }, + { + "bbox": [ + 55, + 138, + 296, + 350 + ], + "type": "text", + "content": ", and its visibility triplane " + }, + { + "bbox": [ + 55, + 138, + 296, + 350 + ], + "type": "inline_equation", + "content": "T_{\\text{cano}}^{\\text{vis}}" + }, + { + "bbox": [ + 55, + 138, + 296, + 350 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 84, + 356, + 295, + 370 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 84, + 356, + 295, + 370 + ], + "spans": [ + { + "bbox": [ + 84, + 356, + 295, + 370 + ], + "type": "interline_equation", + "content": "T _ {f u s e d} = F \\left(T _ {u n d i s t}, T _ {u n d i s t} ^ {\\text {v i s}}, T _ {c a n o}, T _ {c a n o} ^ {\\text {v i s}}\\right). \\tag {5}", + "image_path": "2349845232613d604ef262a22ae1f829352802a741c0c460257a5cae261d5088.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 377, + 296, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 377, + 296, + 437 + ], + "spans": [ + { + "bbox": [ + 55, + 377, + 296, + 437 + ], + "type": "text", + "content": "In this way, Fuser " + }, + { + "bbox": [ + 55, + 377, + 296, + 437 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 55, + 377, + 296, + 437 + ], + "type": "text", + "content": " merges new information in the input frame with occluded regions visible in the reference image. To train the visibility predictor " + }, + { + "bbox": [ + 55, + 377, + 296, + 437 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 55, + 377, + 296, + 437 + ], + "type": "text", + "content": ", we calculate the visibility loss " + }, + { + "bbox": [ + 55, + 377, + 296, + 437 + ], + "type": "inline_equation", + "content": "L_{vis}" + }, + { + "bbox": [ + 55, + 377, + 296, + 437 + ], + "type": "text", + "content": " as the " + }, + { + "bbox": [ + 55, + 377, + 296, + 437 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 55, + 377, + 296, + 437 + ], + "type": "text", + "content": " distance between the predicted visibility triplanes versus the ground truth:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 69, + 442, + 295, + 458 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 442, + 295, + 458 + ], + "spans": [ + { + "bbox": [ + 69, + 442, + 295, + 458 + ], + "type": "interline_equation", + "content": "L _ {v i s} = L _ {1} (T _ {r a w} ^ {v i s}, T _ {r a w} ^ {v i s G T}) + L _ {1} (T _ {c a n o} ^ {v i s}, T _ {c a n o} ^ {v i s G T}). \\quad (6)", + "image_path": "15d57f63fc62b906019309aa8920006a7d04a4d7bece50e900945ea9f9bf5a0a.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 464, + 296, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 464, + 296, + 511 + ], + "spans": [ + { + "bbox": [ + 55, + 464, + 296, + 511 + ], + "type": "text", + "content": "The ground truth visibility triplanes contain 1 for pixels that are visible, and 0 otherwise. The generation of the ground truth visibility triplanes " + }, + { + "bbox": [ + 55, + 464, + 296, + 511 + ], + "type": "inline_equation", + "content": "T_{raw}^{visGT}" + }, + { + "bbox": [ + 55, + 464, + 296, + 511 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 464, + 296, + 511 + ], + "type": "inline_equation", + "content": "T_{cano}^{visGT}" + }, + { + "bbox": [ + 55, + 464, + 296, + 511 + ], + "type": "text", + "content": " is discussed in the supplementary." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 511, + 296, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 511, + 296, + 582 + ], + "spans": [ + { + "bbox": [ + 55, + 511, + 296, + 582 + ], + "type": "text", + "content": "To supervise the Fuser " + }, + { + "bbox": [ + 55, + 511, + 296, + 582 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 55, + 511, + 296, + 582 + ], + "type": "text", + "content": ", we calculate the fusion loss " + }, + { + "bbox": [ + 55, + 511, + 296, + 582 + ], + "type": "inline_equation", + "content": "L_{fusion}" + }, + { + "bbox": [ + 55, + 511, + 296, + 582 + ], + "type": "text", + "content": " as the " + }, + { + "bbox": [ + 55, + 511, + 296, + 582 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 55, + 511, + 296, + 582 + ], + "type": "text", + "content": " loss between the fused triplane " + }, + { + "bbox": [ + 55, + 511, + 296, + 582 + ], + "type": "inline_equation", + "content": "T_{fused}" + }, + { + "bbox": [ + 55, + 511, + 296, + 582 + ], + "type": "text", + "content": " and the pseudo-groundtruth triplane " + }, + { + "bbox": [ + 55, + 511, + 296, + 582 + ], + "type": "inline_equation", + "content": "T_{frontalGT}" + }, + { + "bbox": [ + 55, + 511, + 296, + 582 + ], + "type": "text", + "content": ". To highlight the currently occluded region during training, we also upweight the occluded region using an occlusion mask " + }, + { + "bbox": [ + 55, + 511, + 296, + 582 + ], + "type": "inline_equation", + "content": "T_{occMask} \\in \\mathbb{R}^{3 \\times 1 \\times 256 \\times 256}" + }, + { + "bbox": [ + 55, + 511, + 296, + 582 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 140, + 589, + 294, + 602 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 589, + 294, + 602 + ], + "spans": [ + { + "bbox": [ + 140, + 589, + 294, + 602 + ], + "type": "interline_equation", + "content": "T _ {d i f f} = \\left| T _ {f u s e d} - T _ {t r i p l a n e G T} \\right|, \\tag {7}", + "image_path": "f524240af8227c8b483bf7436fbca56447e2cf26c85f0504ed139df57fe65776.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 63, + 604, + 295, + 625 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 604, + 295, + 625 + ], + "spans": [ + { + "bbox": [ + 63, + 604, + 295, + 625 + ], + "type": "interline_equation", + "content": "L _ {f u s i o n} = \\operatorname {M e a n} \\left(T _ {d i f f}\\right) + \\frac {T _ {d i f f} \\cdot T _ {v i s G T}}{\\left| T _ {v i s G T} \\right|} + \\frac {T _ {d i f f} \\cdot T _ {o c c M a s k}}{\\left| T _ {o c c M a s k} \\right|}. \\tag {8}", + "image_path": "1e780ac2823066e78d518e7307b436f42f0a22becf71724d698d4ac1496f6cce.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 632, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 632, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 632, + 296, + 713 + ], + "type": "text", + "content": "Please refer to the supplementary for the calculation of the occlusion mask. We use the Recurrent Video Restoration Transformer (RVRT) [29] as the backbone of our Fuser " + }, + { + "bbox": [ + 55, + 632, + 296, + 713 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 55, + 632, + 296, + 713 + ], + "type": "text", + "content": " because of its memory efficiency. We find that the final summation skip connection in RVRT prevents effective learning. This is because the original RVRT was designed to correct local blurriness and noise in a corrupted RGB video," + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 72, + 553, + 129 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 129 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 129 + ], + "type": "text", + "content": "whereas our triplane videos exhibit structural distortion on a much larger scale and the summation skip connection thus limits the model's ability to correct the general structure. We thus replace the summation with a small 5-layer ConvNet." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 130, + 554, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 130, + 554, + 189 + ], + "spans": [ + { + "bbox": [ + 313, + 130, + 554, + 189 + ], + "type": "text", + "content": "Lastly, note that both the Undistorter " + }, + { + "bbox": [ + 313, + 130, + 554, + 189 + ], + "type": "inline_equation", + "content": "U" + }, + { + "bbox": [ + 313, + 130, + 554, + 189 + ], + "type": "text", + "content": " and the Fuser " + }, + { + "bbox": [ + 313, + 130, + 554, + 189 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 313, + 130, + 554, + 189 + ], + "type": "text", + "content": " consist of 3 separate copies, one for each of the 3 planes, because we find that processing all three planes jointly leads to collapse to 2D (please see supplementary for visualization and analysis)." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 194, + 411, + 206 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 194, + 411, + 206 + ], + "spans": [ + { + "bbox": [ + 313, + 194, + 411, + 206 + ], + "type": "text", + "content": "3.7. Training Losses" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 211, + 554, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 211, + 554, + 281 + ], + "spans": [ + { + "bbox": [ + 313, + 211, + 554, + 281 + ], + "type": "text", + "content": "Our loss function is the summation of four loss terms that provide two types of supervision: (a) direct triplane space guidance used to supervise the undistortion process in the Undistorter " + }, + { + "bbox": [ + 313, + 211, + 554, + 281 + ], + "type": "inline_equation", + "content": "U" + }, + { + "bbox": [ + 313, + 211, + 554, + 281 + ], + "type": "text", + "content": ", the visibility prediction process, and the fusion process in the Fuser " + }, + { + "bbox": [ + 313, + 211, + 554, + 281 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 313, + 211, + 554, + 281 + ], + "type": "text", + "content": "; and (b) image space guidance for overall learning of high-quality image synthesis:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 353, + 281, + 553, + 308 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 281, + 553, + 308 + ], + "spans": [ + { + "bbox": [ + 353, + 281, + 553, + 308 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} L = w _ {u n d i s t} L _ {u n d i s t} + w _ {f u s i o n} L _ {f u s i o n} \\\\ + w _ {v i s} L _ {v i s} + w _ {r e n d e r} L _ {r e n d e r}, \\tag {9} \\\\ \\end{array}", + "image_path": "e03c6b3847e82ff247e404a75d1bb276645f6dc7417c2a591733cd3cda17229d.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 312, + 554, + 359 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 312, + 554, + 359 + ], + "spans": [ + { + "bbox": [ + 313, + 312, + 554, + 359 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 312, + 554, + 359 + ], + "type": "inline_equation", + "content": "w_{undist}" + }, + { + "bbox": [ + 313, + 312, + 554, + 359 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 312, + 554, + 359 + ], + "type": "inline_equation", + "content": "w_{viz}" + }, + { + "bbox": [ + 313, + 312, + 554, + 359 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 312, + 554, + 359 + ], + "type": "inline_equation", + "content": "w_{fusion}" + }, + { + "bbox": [ + 313, + 312, + 554, + 359 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 312, + 554, + 359 + ], + "type": "inline_equation", + "content": "w_{render}" + }, + { + "bbox": [ + 313, + 312, + 554, + 359 + ], + "type": "text", + "content": " are scalar weights for the different loss terms. " + }, + { + "bbox": [ + 313, + 312, + 554, + 359 + ], + "type": "inline_equation", + "content": "L_{render}" + }, + { + "bbox": [ + 313, + 312, + 554, + 359 + ], + "type": "text", + "content": " is calculated as the perceptual loss " + }, + { + "bbox": [ + 313, + 312, + 554, + 359 + ], + "type": "inline_equation", + "content": "L_{LPIPS}" + }, + { + "bbox": [ + 313, + 312, + 554, + 359 + ], + "type": "text", + "content": " between the ground truth novel view " + }, + { + "bbox": [ + 313, + 312, + 554, + 359 + ], + "type": "inline_equation", + "content": "I_{GT}" + }, + { + "bbox": [ + 313, + 312, + 554, + 359 + ], + "type": "text", + "content": " and the rendered novel view " + }, + { + "bbox": [ + 313, + 312, + 554, + 359 + ], + "type": "inline_equation", + "content": "I_{render}" + }, + { + "bbox": [ + 313, + 312, + 554, + 359 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 362, + 363, + 553, + 376 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 362, + 363, + 553, + 376 + ], + "spans": [ + { + "bbox": [ + 362, + 363, + 553, + 376 + ], + "type": "interline_equation", + "content": "L _ {\\text {r e n d e r}} = L _ {\\text {L P I P S}} \\left(I _ {G T}, I _ {\\text {r e n d e r}}\\right). \\tag {10}", + "image_path": "780d1bee618769cc45fad345497f08d33b7e3fc19f395f09b60596d5e8e73b28.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 380, + 386, + 393 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 380, + 386, + 393 + ], + "spans": [ + { + "bbox": [ + 313, + 380, + 386, + 393 + ], + "type": "text", + "content": "4. Evaluation" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 400, + 554, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 400, + 554, + 668 + ], + "spans": [ + { + "bbox": [ + 313, + 400, + 554, + 668 + ], + "type": "text", + "content": "Multi-view evaluation of single-view reconstruction. Due to the lack of 3D ground truth for in-the-wild images, prior methods are often evaluated on the input image reconstruction task, which compares the rendered image against the input image using metrics like PSNR, whereas the novel view synthesis task often relies on visual assessments. However, this practice can lead to ambiguities and inaccurate conclusions. For example, if a 3D reconstruction is only rendered and evaluated from the input viewpoint, then that reconstruction can overfit to the input view to achieve high scores even if it is inaccurate when rendered from other viewpoints (as proven quantitatively in Fig. 11 of the supplementary). Moreover, single-view reconstruction methods can be heavily affected by the choice of input viewpoints (Fig. 2). Therefore, both the input viewpoint and the novel viewpoint are crucial variables to consider when concluding the performance of a method. We thus propose a new evaluation framework that evaluates a model across every input-novel viewpoint combination. In this way, a method can only achieve high numerical performance when it consistently generates high-quality reconstructions regardless of the choice of input or novel viewpoint." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 673, + 426, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 673, + 426, + 685 + ], + "spans": [ + { + "bbox": [ + 313, + 673, + 426, + 685 + ], + "type": "text", + "content": "4.1. Evaluation Dataset" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 690, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 690, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 690, + 554, + 713 + ], + "type": "text", + "content": "We quantitatively evaluate various methods on the NeRSemble [23] dataset, which is a multiview portrait" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "10716" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 235 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 235 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 235 + ], + "type": "text", + "content": "video dataset that allows us to evaluate the methods using different input-novel viewpoint combinations. It is recorded with 16 calibrated time-synchronized cameras in a controlled studio environment. The images are captured at " + }, + { + "bbox": [ + 55, + 72, + 294, + 235 + ], + "type": "inline_equation", + "content": "7.1\\mathrm{MP}" + }, + { + "bbox": [ + 55, + 72, + 294, + 235 + ], + "type": "text", + "content": " resolution and 73 frames per second. There are 10 sequences in the test set, each capturing a different individual performing different expressions. One of the 10 test sequences involves severe facial occlusion from hair that causes most of the methods' face trackers to fail for significant portions of the recording for many of the viewpoints. We thus leave out that sequence because the results would not be a reliable assessment of quality. We also use 8 roughly evenly separated cameras out of all 16 cameras during evaluation." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 242, + 115, + 253 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 242, + 115, + 253 + ], + "spans": [ + { + "bbox": [ + 55, + 242, + 115, + 253 + ], + "type": "text", + "content": "4.2. Metrics" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 258, + 295, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 258, + 295, + 399 + ], + "spans": [ + { + "bbox": [ + 55, + 258, + 295, + 399 + ], + "type": "text", + "content": "Synthesis Quality. Given " + }, + { + "bbox": [ + 55, + 258, + 295, + 399 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 258, + 295, + 399 + ], + "type": "text", + "content": " views in the dataset, we evaluate a method's average performance across every input-novel viewpoint combination. More specifically, for each frame, each of the " + }, + { + "bbox": [ + 55, + 258, + 295, + 399 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 258, + 295, + 399 + ], + "type": "text", + "content": " cameras is used as the input viewpoint, producing " + }, + { + "bbox": [ + 55, + 258, + 295, + 399 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 258, + 295, + 399 + ], + "type": "text", + "content": " reconstructions in total. Then, each of the " + }, + { + "bbox": [ + 55, + 258, + 295, + 399 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 258, + 295, + 399 + ], + "type": "text", + "content": " reconstructions is rendered and evaluated on the " + }, + { + "bbox": [ + 55, + 258, + 295, + 399 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 258, + 295, + 399 + ], + "type": "text", + "content": " viewpoints, resulting in an " + }, + { + "bbox": [ + 55, + 258, + 295, + 399 + ], + "type": "inline_equation", + "content": "N \\times N" + }, + { + "bbox": [ + 55, + 258, + 295, + 399 + ], + "type": "text", + "content": " score matrix. As mentioned before, we use " + }, + { + "bbox": [ + 55, + 258, + 295, + 399 + ], + "type": "inline_equation", + "content": "N = 8" + }, + { + "bbox": [ + 55, + 258, + 295, + 399 + ], + "type": "text", + "content": " of the camera views from the NeRSemnble[23] dataset as input and novel views. Thus, for a test sequence of " + }, + { + "bbox": [ + 55, + 258, + 295, + 399 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 258, + 295, + 399 + ], + "type": "text", + "content": " frames, we can calculate a spatial-temporal score matrices " + }, + { + "bbox": [ + 55, + 258, + 295, + 399 + ], + "type": "inline_equation", + "content": "\\mathbf{S}^{T \\times 8 \\times 8}" + }, + { + "bbox": [ + 55, + 258, + 295, + 399 + ], + "type": "text", + "content": " for each of the metrics (as visualized in Fig. 11 in the supplementary):" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 74, + 401, + 294, + 415 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 401, + 294, + 415 + ], + "spans": [ + { + "bbox": [ + 74, + 401, + 294, + 415 + ], + "type": "interline_equation", + "content": "\\mathbf {S} _ {t, i, j} = \\operatorname {M e t r i c} \\left(\\mathbf {I} _ {\\text {r e n d e r}} ^ {t, i, j}, \\mathbf {I} _ {G T} ^ {t, j}\\right), 1 \\leq i, j \\leq N, 1 \\leq t \\leq T, \\tag {11}", + "image_path": "543012f37eac75c3984c48cb920b420c8723acc9f31c66b1edd0133d4d0201b9.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 173, + 416, + 294, + 429 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 173, + 416, + 294, + 429 + ], + "spans": [ + { + "bbox": [ + 173, + 416, + 294, + 429 + ], + "type": "interline_equation", + "content": "s = \\operatorname {M e a n} \\left(\\left\\{\\mathbf {S} _ {t, i, j} \\right\\}\\right), \\tag {12}", + "image_path": "74f1c5461c193c58f7654140075595309b03cb059ee56a3ae03c4fcfbb4cc5f9.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 438, + 295, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 438, + 295, + 532 + ], + "spans": [ + { + "bbox": [ + 55, + 438, + 295, + 532 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 438, + 295, + 532 + ], + "type": "inline_equation", + "content": "Metric(\\cdot)" + }, + { + "bbox": [ + 55, + 438, + 295, + 532 + ], + "type": "text", + "content": " can be LPIPS, PSNR, etc.. " + }, + { + "bbox": [ + 55, + 438, + 295, + 532 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{\\text{render}}^{t,i,j}" + }, + { + "bbox": [ + 55, + 438, + 295, + 532 + ], + "type": "text", + "content": " is the image rendered using camera " + }, + { + "bbox": [ + 55, + 438, + 295, + 532 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 438, + 295, + 532 + ], + "type": "text", + "content": " as the input frame and camera " + }, + { + "bbox": [ + 55, + 438, + 295, + 532 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 55, + 438, + 295, + 532 + ], + "type": "text", + "content": " as the output rendering view at frame " + }, + { + "bbox": [ + 55, + 438, + 295, + 532 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 438, + 295, + 532 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 55, + 438, + 295, + 532 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{GT}^{t,j}" + }, + { + "bbox": [ + 55, + 438, + 295, + 532 + ], + "type": "text", + "content": " is the ground truth frame captured by camera " + }, + { + "bbox": [ + 55, + 438, + 295, + 532 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 55, + 438, + 295, + 532 + ], + "type": "text", + "content": " at frame " + }, + { + "bbox": [ + 55, + 438, + 295, + 532 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 438, + 295, + 532 + ], + "type": "text", + "content": ". The Synthesis Quality " + }, + { + "bbox": [ + 55, + 438, + 295, + 532 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 55, + 438, + 295, + 532 + ], + "type": "text", + "content": " is thus the average over all entries in S. For a dataset with multiple sequences, such as NeRSample, the final Synthesis Quality is the average across all sequences." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 532, + 295, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 532, + 295, + 579 + ], + "spans": [ + { + "bbox": [ + 55, + 532, + 295, + 579 + ], + "type": "text", + "content": "Novel View Synthesis (NVS) Quality. Novel View Synthesis Quality " + }, + { + "bbox": [ + 55, + 532, + 295, + 579 + ], + "type": "inline_equation", + "content": "s_{NV}" + }, + { + "bbox": [ + 55, + 532, + 295, + 579 + ], + "type": "text", + "content": " is the average over all entries where the input view " + }, + { + "bbox": [ + 55, + 532, + 295, + 579 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 532, + 295, + 579 + ], + "type": "text", + "content": " is different from the novel view " + }, + { + "bbox": [ + 55, + 532, + 295, + 579 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 55, + 532, + 295, + 579 + ], + "type": "text", + "content": ", essentially removing the diagonal entries from the score matrix:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 88, + 587, + 294, + 601 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 587, + 294, + 601 + ], + "spans": [ + { + "bbox": [ + 88, + 587, + 294, + 601 + ], + "type": "interline_equation", + "content": "s _ {N V} = \\operatorname {M e a n} \\left(\\left\\{\\mathbf {S} _ {t, i, j} \\mid i \\neq j, 1 \\leq t \\leq T \\right\\}\\right). \\tag {13}", + "image_path": "809bace667d98f400cc432bd969f3f0650200828cfcf84e3d262f356b5bf2149.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 608, + 295, + 644 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 608, + 295, + 644 + ], + "spans": [ + { + "bbox": [ + 55, + 608, + 295, + 644 + ], + "type": "text", + "content": "Identity Accuracy (ID) We measure the ArcFace [7] cosine distance between the ground truth image " + }, + { + "bbox": [ + 55, + 608, + 295, + 644 + ], + "type": "inline_equation", + "content": "I_{GT}" + }, + { + "bbox": [ + 55, + 608, + 295, + 644 + ], + "type": "text", + "content": " and rendered image " + }, + { + "bbox": [ + 55, + 608, + 295, + 644 + ], + "type": "inline_equation", + "content": "I_{render}" + }, + { + "bbox": [ + 55, + 608, + 295, + 644 + ], + "type": "text", + "content": " both from the frontal camera." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 643, + 295, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 643, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 643, + 295, + 689 + ], + "type": "text", + "content": "Expression Accuracy (Expr) We use NVIDIA Maxine AR SDK [35] to measure the " + }, + { + "bbox": [ + 55, + 643, + 295, + 689 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 55, + 643, + 295, + 689 + ], + "type": "text", + "content": " distance between expression coefficients " + }, + { + "bbox": [ + 55, + 643, + 295, + 689 + ], + "type": "inline_equation", + "content": "e_{GT}" + }, + { + "bbox": [ + 55, + 643, + 295, + 689 + ], + "type": "text", + "content": " of the ground truth image and " + }, + { + "bbox": [ + 55, + 643, + 295, + 689 + ], + "type": "inline_equation", + "content": "e_{render}" + }, + { + "bbox": [ + 55, + 643, + 295, + 689 + ], + "type": "text", + "content": " of the rendered image both from the frontal camera." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 690, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 690, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 690, + 295, + 713 + ], + "type": "text", + "content": "Dynamic Appearance While we believe that it is important to measure the accuracy of dynamic appearance such" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 72, + 553, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 131 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 131 + ], + "type": "text", + "content": "as lighting and shoulder poses, there is no existing multiview in-the-wild portrait video dataset to support such evaluation. We thus qualitatively evaluate various methods on challenging in-the-wild portrait videos. Please see the supplementary material for image examples and video results." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 138, + 400, + 152 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 138, + 400, + 152 + ], + "spans": [ + { + "bbox": [ + 313, + 138, + 400, + 152 + ], + "type": "text", + "content": "4.3. Comparisons" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 156, + 555, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 156, + 555, + 516 + ], + "spans": [ + { + "bbox": [ + 313, + 156, + 555, + 516 + ], + "type": "text", + "content": "Baselines. We compare recent methods from 3 categories: (i) Reconstruction: We evaluate LP3D [52] using the above protocol. LP3D generates " + }, + { + "bbox": [ + 313, + 156, + 555, + 516 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 156, + 555, + 516 + ], + "type": "text", + "content": " 3D portraits using each of the " + }, + { + "bbox": [ + 313, + 156, + 555, + 516 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 156, + 555, + 516 + ], + "type": "text", + "content": " viewpoints as input, and the " + }, + { + "bbox": [ + 313, + 156, + 555, + 516 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 156, + 555, + 516 + ], + "type": "text", + "content": " 3D portraits are then evaluated on the " + }, + { + "bbox": [ + 313, + 156, + 555, + 516 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 156, + 555, + 516 + ], + "type": "text", + "content": " ground truth viewpoints. (ii) Reenactment: We evaluate Li et al. [28] and GPAvatar [6] in the self-reenactment setting. Li et al. [28] reconstruct 3D portraits as triplanes from reference images without test-time optimization, and they drive the 3D portraits via dynamic frontal renderings of 3DMMs. The authors of Li et al. [28] evaluated their approach using the input views as the only novel view, instead of all " + }, + { + "bbox": [ + 313, + 156, + 555, + 516 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 156, + 555, + 516 + ], + "type": "text", + "content": " views. GPAvatar[6] reconstructs 3D portraits by leveraging multiple source images and driving them through a FLAME [26] mesh model. We use the first frame of the frontal camera in each NeRSample test sequence as the reference image to generate the 3D portrait, and drive it using videos from each of the " + }, + { + "bbox": [ + 313, + 156, + 555, + 516 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 156, + 555, + 516 + ], + "type": "text", + "content": " viewpoints. We evaluate GPAvatar using the same evaluation protocol as our method and LP3D. (iii) Inversion: We evaluate VIVE3D [13], which is a state-of-the-art 3D GAN inversion method for videos, and it can also perform semantic video editing. We first provide VIVE3D 3 video frames from the input viewpoint, which it requires for personalization. Then, we invert the input video frames into 3D portraits and render the 3D portraits from all " + }, + { + "bbox": [ + 313, + 156, + 555, + 516 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 156, + 555, + 516 + ], + "type": "text", + "content": " viewpoints. Each of the above method uses a different cropping of the face. We standardize the evaluation by re-cropping all methods to our cropping protocol, which is the largest of all. Please see the supplement for an additional table, where we evaluate the methods using different cropings around the face and arrive at conclusions consistent with Table 1." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 516, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 516, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 516, + 556, + 714 + ], + "type": "text", + "content": "Quantitative results. We evaluate various methods using different input-novel viewpoint combinations, providing a robust multi-view assessment of a model's performance. Table. 1 shows that our model achieves state-of-the-art performance across all metrics versus recent works. Notably, while LP3D overfits to the input viewpoint (as shown by the difference of " + }, + { + "bbox": [ + 313, + 516, + 556, + 714 + ], + "type": "inline_equation", + "content": "0.806\\mathrm{dB}" + }, + { + "bbox": [ + 313, + 516, + 556, + 714 + ], + "type": "text", + "content": " PSNR between Synthesis Quality and NVS Quality), our method generalizes better to novel viewpoints (as shown by the difference of " + }, + { + "bbox": [ + 313, + 516, + 556, + 714 + ], + "type": "inline_equation", + "content": "0.330\\mathrm{dB}" + }, + { + "bbox": [ + 313, + 516, + 556, + 714 + ], + "type": "text", + "content": "). Moreover, our method best preserves subject identity and expression (Table. 1, Fig. 2, and 4). On the other hand, the reenactment methods struggle to capture authentic expressions because of the use of morphable face models, which have limited expressiveness and accuracy. Moreover, they cannot reconstruct dynamic appearance (e.g. the stuck-out tongue in the second and third rows of Fig. 4) because they solely rely on information present in the reference image(s) and" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "10717" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 83, + 553, + 413 + ], + "blocks": [ + { + "bbox": [ + 76, + 72, + 122, + 82 + ], + "lines": [ + { + "bbox": [ + 76, + 72, + 122, + 82 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 122, + 82 + ], + "type": "text", + "content": "Groundtruth" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 171, + 72, + 190, + 81 + ], + "lines": [ + { + "bbox": [ + 171, + 72, + 190, + 81 + ], + "spans": [ + { + "bbox": [ + 171, + 72, + 190, + 81 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 252, + 72, + 275, + 81 + ], + "lines": [ + { + "bbox": [ + 252, + 72, + 275, + 81 + ], + "spans": [ + { + "bbox": [ + 252, + 72, + 275, + 81 + ], + "type": "text", + "content": "LP3D" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 328, + 72, + 365, + 81 + ], + "lines": [ + { + "bbox": [ + 328, + 72, + 365, + 81 + ], + "spans": [ + { + "bbox": [ + 328, + 72, + 365, + 81 + ], + "type": "text", + "content": "GPAvatar" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 388, + 72, + 468, + 83 + ], + "lines": [ + { + "bbox": [ + 388, + 72, + 468, + 83 + ], + "spans": [ + { + "bbox": [ + 388, + 72, + 468, + 83 + ], + "type": "text", + "content": "Input / Driving Frame" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 476, + 72, + 545, + 82 + ], + "lines": [ + { + "bbox": [ + 476, + 72, + 545, + 82 + ], + "spans": [ + { + "bbox": [ + 476, + 72, + 545, + 82 + ], + "type": "text", + "content": "Reference / Source" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 58, + 83, + 553, + 413 + ], + "lines": [ + { + "bbox": [ + 58, + 83, + 553, + 413 + ], + "spans": [ + { + "bbox": [ + 58, + 83, + 553, + 413 + ], + "type": "image", + "image_path": "4e299ad7c03841300890bb99d16bffa0a5a809729a23a80a534e72d20a65bd4f.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 422, + 555, + 467 + ], + "lines": [ + { + "bbox": [ + 55, + 422, + 555, + 467 + ], + "spans": [ + { + "bbox": [ + 55, + 422, + 555, + 467 + ], + "type": "text", + "content": "Figure 4. Visual comparisons with baseline methods. Our method strikes a balance between coherent reconstruction and faithfully preserving dynamic conditions like expressions. LP3D (third column) exhibits inconsistency in identities, hairstyles, and artifacts (red circles). GPAvatar (fourth column) fails to capture challenging expressions (first row), new information not present in the reference image, (the stuck-out tongue in second and third rows), and the identity of the person (last row)." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 477, + 295, + 512 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 477, + 295, + 512 + ], + "spans": [ + { + "bbox": [ + 55, + 477, + 295, + 512 + ], + "type": "text", + "content": "do not incorporate new per-frame information. On the other hand, our method achieves faithful dynamic appearance and coherent reconstruction at the same time." + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 56, + 533, + 295, + 630 + ], + "blocks": [ + { + "bbox": [ + 56, + 533, + 295, + 630 + ], + "lines": [ + { + "bbox": [ + 56, + 533, + 295, + 630 + ], + "spans": [ + { + "bbox": [ + 56, + 533, + 295, + 630 + ], + "type": "table", + "html": "
MethodTypeExpr↓ID↓SynthesisLIPS↓LVS Quality
PSNR↑LPIPS↓PSNR↑LPIPS↓
Li et al. [28]reenact0.2660.24118.5730.25518.2020.262
GPAvatar[6]reenact0.2040.20721.9490.23321.9490.2334
VIVE3D[13]invert0.2900.39518.5770.25918.1450.271
LP3D[52]recon0.1680.21522.3310.22321.5250.237
Oursrecon0.1580.18722.7700.21922.4400.224
", + "image_path": "895e615cb663b340f45519e636c291ac699f3bb4a68f77f17c31fc0cca5e8cdf.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 638, + 295, + 692 + ], + "lines": [ + { + "bbox": [ + 55, + 638, + 295, + 692 + ], + "spans": [ + { + "bbox": [ + 55, + 638, + 295, + 692 + ], + "type": "text", + "content": "Table 1. Comparisons on Nersemble [23]. Our evaluation protocol (Sec. 4.2) utilizes multiview groundtruth for thorough analysis. Our method achieves state-of-the-art performance across all metrics, maintains temporal coherence, and preserves accurate expressions." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 314, + 476, + 382, + 487 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 476, + 382, + 487 + ], + "spans": [ + { + "bbox": [ + 314, + 476, + 382, + 487 + ], + "type": "text", + "content": "4.4. Ablations" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 493, + 555, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 493, + 555, + 563 + ], + "spans": [ + { + "bbox": [ + 313, + 493, + 555, + 563 + ], + "type": "text", + "content": "We evaluate the effectiveness of our proposed Triplane Undistorter and Fuser modules on NeRSemble. In addition to the Synthesis Quality in terms of PSNR and LPIPS, we develop two new metrics to separately measure the robustness to variations in (1) rendering viewpoints and (2) input viewpoints:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 563, + 555, + 633 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 563, + 555, + 633 + ], + "spans": [ + { + "bbox": [ + 313, + 563, + 555, + 633 + ], + "type": "text", + "content": "(1) Novel View Variation (NVV). We evaluate how much a method's reconstruction quality varies across different novel views. We quantify this as the standard deviation of performance across the " + }, + { + "bbox": [ + 313, + 563, + 555, + 633 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 563, + 555, + 633 + ], + "type": "text", + "content": " novel views when using the same input view, i.e. the average standard deviation of each horizontal row of the score matrix " + }, + { + "bbox": [ + 313, + 563, + 555, + 633 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 313, + 563, + 555, + 633 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 318, + 639, + 555, + 672 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 639, + 555, + 672 + ], + "spans": [ + { + "bbox": [ + 318, + 639, + 555, + 672 + ], + "type": "interline_equation", + "content": "\\mathrm {N V V} = \\frac {1}{T N} \\sum_ {t = 1} ^ {T} \\sum_ {i = 1} ^ {N} \\operatorname {s t d d e v} \\left(\\left\\{S _ {t, i, j} \\mid 1 \\leq j \\leq N \\right\\}\\right). \\tag {14}", + "image_path": "70423882ab2aadde1125e1f0fafb7480a2cfaaf0fd6dd69cefa243847beaefe3.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 678, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 678, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 678, + 555, + 714 + ], + "type": "text", + "content": "(2) Input View Variation (IVV). We measure how much a method's reconstruction quality varies when using the " + }, + { + "bbox": [ + 313, + 678, + 555, + 714 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 678, + 555, + 714 + ], + "type": "text", + "content": " different input viewpoints (Sec. 4.2 second column from the" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10718" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 70, + 298, + 155 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 298, + 155 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 298, + 155 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 298, + 155 + ], + "type": "table", + "html": "
MethodU FSynthesis QualityIVV↓NVV↓
PSNR↑LPIPS↓(PSNR)(LPIPS)(PSNR)(LPIPS)
LP3D[52]XX22.3310.2231.0250.0152.2000.053
Ours✓X22.1960.2210.9070.0091.6990.038
X✓22.2650.2230.5590.0061.3150.029
✓✓22.7700.2190.2450.0051.3830.037
", + "image_path": "6549d0c6c73901cfe3faf3b718e3c7973a877153f8079d815f9bee5dc27252c2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 162, + 295, + 216 + ], + "lines": [ + { + "bbox": [ + 55, + 162, + 295, + 216 + ], + "spans": [ + { + "bbox": [ + 55, + 162, + 295, + 216 + ], + "type": "text", + "content": "Table 2. Undistorter and Fuser Ablations. The Undistorter-only (row 2) and Fuser-only (row 3) variants both lead to improvements in viewpoint robustness metrics (IVV and NVV), but not in Synthesis Quality. The best performance is achieved by using both the Undistorter and the Fuser at the same time." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 228, + 296, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 228, + 296, + 275 + ], + "spans": [ + { + "bbox": [ + 55, + 228, + 296, + 275 + ], + "type": "text", + "content": "right). We quantify this variation as the average standard deviation of performance on the same novel view when using different input views, i.e. the average standard deviation of each vertical column of the score matrix " + }, + { + "bbox": [ + 55, + 228, + 296, + 275 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 55, + 228, + 296, + 275 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 63, + 278, + 295, + 312 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 278, + 295, + 312 + ], + "spans": [ + { + "bbox": [ + 63, + 278, + 295, + 312 + ], + "type": "interline_equation", + "content": "\\mathrm {I V V} = \\frac {1}{T N} \\sum_ {t = 1} ^ {T} \\sum_ {j = 1} ^ {N} \\operatorname {s t d d e v} \\left(\\left\\{S _ {t, i, j} \\mid 1 \\leq i \\leq N \\right\\}\\right). \\tag {15}", + "image_path": "ab28c478764c7aeafdb4338bc23ec9cb4f3cc088927240cd87aeab8558c3e21e.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 318, + 295, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 318, + 295, + 341 + ], + "spans": [ + { + "bbox": [ + 55, + 318, + 295, + 341 + ], + "type": "inline_equation", + "content": "NVV" + }, + { + "bbox": [ + 55, + 318, + 295, + 341 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 318, + 295, + 341 + ], + "type": "inline_equation", + "content": "IVV" + }, + { + "bbox": [ + 55, + 318, + 295, + 341 + ], + "type": "text", + "content": " are measured for both PSNR and LPIPS, and they are better if lower." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 342, + 295, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 342, + 295, + 608 + ], + "spans": [ + { + "bbox": [ + 55, + 342, + 295, + 608 + ], + "type": "text", + "content": "As shown in Table. 2, the Undistorter module consistently improves the NVV and IVV metrics versus LP3D, indicating better robustness to different input viewpoints and more consistent rendering quality across novel rendered views. However, when only the Undistorter is added (Tab. 2 second row) the PSNR is slightly reduced. This is likely because this model does not leverage the reference image to recover occluded areas while also overfitting to the input view. Similarly, the Fuser-only variant (Tab. 2 third row) achieves better NVV and IVV scores, but also scores lower in terms of PSNR. A likely cause is that, without the Undistorter, the Fuser needs to merge highly misaligned triplanes, where the person looks drastically different in " + }, + { + "bbox": [ + 55, + 342, + 295, + 608 + ], + "type": "inline_equation", + "content": "T_{raw}" + }, + { + "bbox": [ + 55, + 342, + 295, + 608 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 342, + 295, + 608 + ], + "type": "inline_equation", + "content": "T_{cano}" + }, + { + "bbox": [ + 55, + 342, + 295, + 608 + ], + "type": "text", + "content": ". This misalignment possibly induces more blurriness and alignment artifacts that lower the PSNR. The best performance is achieved by our full model, which includes both the Undistorter and the Fuser and thus achieves better PSNR, LPIPS, IVV and NVV scores. This means that the Undistorter and the Fuser complement each other: The Undistorter corrects the distortion in the raw triplane and thus reduces the challenges in fusing misaligned triplanes, and the Fuser recovers the occluded areas in the raw triplane." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 609, + 295, + 644 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 609, + 295, + 644 + ], + "spans": [ + { + "bbox": [ + 55, + 609, + 295, + 644 + ], + "type": "text", + "content": "Shoulder Augmentation. without shoulder augmentation, the model fails to correctly reconstruct varying shoulder poses, resulting in worse performance (Fig. 5 and Table. 3)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 644, + 295, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 644, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 644, + 295, + 689 + ], + "type": "text", + "content": "Optical Flow. As mentioned in Sec. 3.5, replacing the Undistorter with optical flow worsens the reconstruction (Table. 3). As shown by Fig. 5, this variant fails to correctly reconstruct the shoulder and the mouth accurately." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "content": "GAN Training. We retrained our model with an additional adversarial loss incorporated into Eqn. 10. As expected, the" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 318, + 72, + 552, + 157 + ], + "blocks": [ + { + "bbox": [ + 318, + 72, + 552, + 157 + ], + "lines": [ + { + "bbox": [ + 318, + 72, + 552, + 157 + ], + "spans": [ + { + "bbox": [ + 318, + 72, + 552, + 157 + ], + "type": "image", + "image_path": "0bf2bdf2ae8f88065d6249dc2bbb6387132e237314eb57e37ea0f8f069b5b960.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 315, + 198, + 553, + 294 + ], + "blocks": [ + { + "bbox": [ + 313, + 166, + 554, + 198 + ], + "lines": [ + { + "bbox": [ + 313, + 166, + 554, + 198 + ], + "spans": [ + { + "bbox": [ + 313, + 166, + 554, + 198 + ], + "type": "text", + "content": "Figure 5. Visual ablation. Our method with optical flow and w/o shoulder augmentation on two different input frames (top and bottom rows)." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 198, + 553, + 294 + ], + "lines": [ + { + "bbox": [ + 315, + 198, + 553, + 294 + ], + "spans": [ + { + "bbox": [ + 315, + 198, + 553, + 294 + ], + "type": "table", + "html": "
MethodLPIPS↓PSNR↑IVV (PSNR) ↓ID↓Expr↓
LP3D (orig.)0.33418.7212.1300.2470.451
LP3D (ours)0.22322.3311.0250.1680.215
w optical flow0.22722.0851.1750.1780.335
w/o shoulder aug.0.21822.3420.8290.1530.244
Ours0.21922.7700.2450.1580.187
Ours + GAN0.20722.0740.3140.1520.198
", + "image_path": "69e9b4d60aacd9bb22256783be28169cc859b9e1ba35bbb35d853a074bd8d22c.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 297, + 554, + 352 + ], + "lines": [ + { + "bbox": [ + 313, + 297, + 554, + 352 + ], + "spans": [ + { + "bbox": [ + 313, + 297, + 554, + 352 + ], + "type": "text", + "content": "Table 3. Other Ablations. We compare our implementation of LP3D to the original one [52] (row 1&2); the effect of replacing our Undistorter with optical flow warping (row 4); the effect of removing shoulder augmentation during training (row 4); and the effect of adding an adversarial GAN loss (row 6) alongside Eqn. 10." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 365, + 553, + 401 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 365, + 553, + 401 + ], + "spans": [ + { + "bbox": [ + 313, + 365, + 553, + 401 + ], + "type": "text", + "content": "GAN training (Table. 3) improves the sharpness of images as seen from the LPIPS error reduction from 0.219 to 0.207. Additional results are included in the supplementary." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 402, + 554, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 402, + 554, + 472 + ], + "spans": [ + { + "bbox": [ + 313, + 402, + 554, + 472 + ], + "type": "text", + "content": "Qualitative results. In-the-wild experiments show that our method achieves better temporal consistency than LP3D [52] and more accurately captures dynamic information like expressions and lighting changes than GPAvatar [6]. We refer the readers to our supplementary materials for visual examples of these results." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 483, + 384, + 496 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 483, + 384, + 496 + ], + "spans": [ + { + "bbox": [ + 313, + 483, + 384, + 496 + ], + "type": "text", + "content": "5. Discussion" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 504, + 554, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 504, + 554, + 597 + ], + "spans": [ + { + "bbox": [ + 313, + 504, + 554, + 597 + ], + "type": "text", + "content": "Conclusion. Recognizing the individual limitations of per-frame single-view reconstruction and 3D reenactment methods, we presented the first single-view 3D lifting method to reconstruct a 3D photorealistic avatar with faithful dynamic appearance as well as temporal consistency, marrying the best of both worlds. We believe our method paves the way forward for creating a high-quality telepresence system accessible to consumers." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 605, + 554, + 702 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 554, + 702 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 554, + 702 + ], + "type": "text", + "content": "Limitations and future work. We use a single reference image, but incorporating multiple ones with different expressions and head poses could lead to further improvements. With our method, fusing an extreme side view with a very different expression to the reference view may result in blurry reconstruction due to ambiguity in triplane alignment. Finally, due to the additional components, our current run-time performance is slower than real time, which could be improved in future work." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10719" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 133 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 133 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 133 + ], + "type": "text", + "content": "[1] ShahRukh Athar, Zexiang Xu, Kalyan Sunkavalli, Eli Shechtman, and Zhixin Shu. Rignerf: Fully controllable neural 3d portraits. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 135, + 296, + 166 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 135, + 296, + 166 + ], + "spans": [ + { + "bbox": [ + 61, + 135, + 296, + 166 + ], + "type": "text", + "content": "[2] Ananta R. Bhattarai, Matthias Nießner, and Artem Sevastopolsky. Triplanenet: An encoder for eg3d inversion. 2024. 1, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 168, + 296, + 242 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 168, + 296, + 242 + ], + "spans": [ + { + "bbox": [ + 62, + 168, + 296, + 242 + ], + "type": "text", + "content": "[3] Marcel C. Buehler, Kripasindhu Sarkar, Tanmay Shah, Gengyan Li, Daoye Wang, Leonhard Helminger, Sergio Orts-Escolano, Dmitry Lagun, Otmar Hilliges, Thabo Beeler, and Abhimitra Meka. Preface: A data-driven volumetric prior for few-shot ultra high-resolution face synthesis. In IEEE International Conference on Computer Vision (ICCV), 2023. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 244, + 296, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 244, + 296, + 306 + ], + "spans": [ + { + "bbox": [ + 62, + 244, + 296, + 306 + ], + "type": "text", + "content": "[4] Chen Cao, Tomas Simon, Jin Kyu Kim, Gabe Schwartz, Michael Zollhoefer, Shun-Suke Saito, Stephen Lombardi, Shih-En Wei, Danielle Belko, Shouou-I Yu, Yaser Sheikh, and Jason Saragih. Authentic volumetric avatars from a phone scan. ACM Transactions on Graphics (SIGGRAPH), 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 309, + 296, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 309, + 296, + 373 + ], + "spans": [ + { + "bbox": [ + 62, + 309, + 296, + 373 + ], + "type": "text", + "content": "[5] Eric R. Chan, Connor Z. Lin, Matthew A. Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas Guibas, Jonathan Tremblay, Sameh Khamis, Tero Karras, and Gordon Wetzstein. Efficient geometry-aware 3D generative adversarial networks. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 375, + 296, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 375, + 296, + 426 + ], + "spans": [ + { + "bbox": [ + 62, + 375, + 296, + 426 + ], + "type": "text", + "content": "[6] Xuangeng Chu, Yu Li, Ailing Zeng, Tianyu Yang, Lijian Lin, Yunfei Liu, and Tatsuya Harada. Gpavatar: Generalizable and precise head avatar from image(s). In International Conference on Learning Representations (ICLR), 2024. 2, 3, 6, 7, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 429, + 296, + 471 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 429, + 296, + 471 + ], + "spans": [ + { + "bbox": [ + 62, + 429, + 296, + 471 + ], + "type": "text", + "content": "[7] Jiankang Deng, Jia Guo, Xue Niannan, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 473, + 296, + 515 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 473, + 296, + 515 + ], + "spans": [ + { + "bbox": [ + 62, + 473, + 296, + 515 + ], + "type": "text", + "content": "[8] Yu Deng, Jiaolong Yang, Jianfeng Xiang, and Xin Tong. Gram: Generative radiance manifolds for 3d-aware image generation. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 517, + 296, + 559 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 517, + 296, + 559 + ], + "spans": [ + { + "bbox": [ + 62, + 517, + 296, + 559 + ], + "type": "text", + "content": "[9] Michail Christos Doukas, Stefanos Zafeiriou, and Viktoriia Sharmanska. Headgan: One-shot neural head synthesis and editing. In IEEE International Conference on Computer Vision (ICCV), 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 560, + 295, + 603 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 560, + 295, + 603 + ], + "spans": [ + { + "bbox": [ + 57, + 560, + 295, + 603 + ], + "type": "text", + "content": "[10] Nikita Drobyshev, Jenya Chelishev, Taras Khakhulin, Aleksei Ivakhnenko, Victor Lempitsky, and Egor Zakharov. Megaportraits: One-shot megapixel neural head avatars. arXiv preprint arXiv:2207.07621, 2022. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 605, + 295, + 636 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 605, + 295, + 636 + ], + "spans": [ + { + "bbox": [ + 57, + 605, + 295, + 636 + ], + "type": "text", + "content": "[11] Yao Feng, Haiwen Feng, Michael J. Black, and Timo Bolkart. Learning an animatable detailed 3D face model from in-the-wild images. 2021. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 638, + 295, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 638, + 295, + 679 + ], + "spans": [ + { + "bbox": [ + 57, + 638, + 295, + 679 + ], + "type": "text", + "content": "[12] Yao Feng, Haiwen Feng, Michael J Black, and Timo Bolkart. Learning an animatable detailed 3d face model from in-the-wild images. ACM Transactions on Graphics (ToG), 40(4): 1-13, 2021. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 681, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 681, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 681, + 295, + 713 + ], + "type": "text", + "content": "[13] Anna Fruhstück, Nikolaos Sarafianos, Yuanlu Xu, Peter Wonka, and Tony Tung. VIVE3D: Viewpoint-independent video editing using 3D-Aware GANs. In IEEE Conference" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 317, + 74, + 553, + 714 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 335, + 74, + 553, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 74, + 553, + 94 + ], + "spans": [ + { + "bbox": [ + 335, + 74, + 553, + 94 + ], + "type": "text", + "content": "on Computer Vision and Pattern Recognition (CVPR), 2023. 3, 6, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 96, + 553, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 96, + 553, + 138 + ], + "spans": [ + { + "bbox": [ + 317, + 96, + 553, + 138 + ], + "type": "text", + "content": "[14] Guy Gafni, Justus Thies, Michael Zollhöfer, and Matthias Nießner. Dynamic neural radiance fields for monocular 4d facial avatar reconstruction. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 140, + 553, + 181 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 140, + 553, + 181 + ], + "spans": [ + { + "bbox": [ + 317, + 140, + 553, + 181 + ], + "type": "text", + "content": "[15] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In Advances in Neural Information Processing Systems (NeurIPS), 2014. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 182, + 553, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 182, + 553, + 224 + ], + "spans": [ + { + "bbox": [ + 317, + 182, + 553, + 224 + ], + "type": "text", + "content": "[16] Jiatao Gu, Lingjie Liu, Peng Wang, and Christian Theobalt. StyleNeRF: A style-based 3D-aware generator for high-resolution image synthesis. arXiv preprint arXiv:2110.08985, 2021. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 226, + 553, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 226, + 553, + 258 + ], + "spans": [ + { + "bbox": [ + 317, + 226, + 553, + 258 + ], + "type": "text", + "content": "[17] Fa-Ting Hong, Longhao Zhang, Li Shen, and Dan Xu. Depth-aware generative adversarial network for talking head video generation. 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 259, + 553, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 259, + 553, + 300 + ], + "spans": [ + { + "bbox": [ + 317, + 259, + 553, + 300 + ], + "type": "text", + "content": "[18] Yang Hong, Bo Peng, Haiyao Xiao, Ligang Liu, and Juyong Zhang. Headnerf: A real-time nerf-based parametric head model. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 302, + 553, + 355 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 302, + 553, + 355 + ], + "spans": [ + { + "bbox": [ + 317, + 302, + 553, + 355 + ], + "type": "text", + "content": "[19] Andrew Jones, Magnus Lang, Graham Fyffe, Xueming Yu, Jay Busch, Ian McDowall, Mark Bolas, and Paul Debevec. Achieving eye contact in a one-to-many 3d video teleconferencing system. ACM Transactions on Graphics (SIGGRAPH), 2009. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 357, + 553, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 357, + 553, + 398 + ], + "spans": [ + { + "bbox": [ + 317, + 357, + 553, + 398 + ], + "type": "text", + "content": "[20] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 399, + 553, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 399, + 553, + 441 + ], + "spans": [ + { + "bbox": [ + 317, + 399, + 553, + 441 + ], + "type": "text", + "content": "[21] Peter Kauff and Oliver Schreer. An immersive 3d videoconferencing system using shared virtual team user environments. In Proceedings of the 4th International Conference on Collaborative Virtual Environments, 2002. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 443, + 553, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 443, + 553, + 485 + ], + "spans": [ + { + "bbox": [ + 317, + 443, + 553, + 485 + ], + "type": "text", + "content": "[22] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics (SIGGRAPH), 2023. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 487, + 553, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 487, + 553, + 529 + ], + "spans": [ + { + "bbox": [ + 317, + 487, + 553, + 529 + ], + "type": "text", + "content": "[23] Tobias Kirschstein, Shenhan Qian, Simon Giebenhain, Tim Walter, and Matthias Nießner. Nersemble: Multi-view radiance field reconstruction of human heads. ACM Trans. Graph., 2023. 5, 6, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 530, + 553, + 572 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 530, + 553, + 572 + ], + "spans": [ + { + "bbox": [ + 317, + 530, + 553, + 572 + ], + "type": "text", + "content": "[24] Jaehoon Ko, Kyusun Cho, Daewon Choi, Kwangrok Ryoo, and Seungryong Kim. 3d gan inversion with pose optimization. In IEEE Winter Conference on Applications of Computer Vision (WACV), 2023. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 574, + 553, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 574, + 553, + 658 + ], + "spans": [ + { + "bbox": [ + 317, + 574, + 553, + 658 + ], + "type": "text", + "content": "[25] Jason Lawrence, Dan B Goldman, Supreeth Achar, Gregory Major Blascovich, Joseph G. Desloge, Tommy Fortes, Eric M. Gomez, Sascha Haberling, Hugues Hoppe, Andy Huibers, Claude Knaus, Brian Kuschak, Ricardo Martin-Brualla, Harris Nover, Andrew Ian Russell, Steven M. Seitz, and Kevin Tong. Project starline: A high-fidelity telepresence system. ACM Transactions on Graphics (SIGGRAPH ASIA), 2021. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 659, + 553, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 659, + 553, + 690 + ], + "spans": [ + { + "bbox": [ + 317, + 659, + 553, + 690 + ], + "type": "text", + "content": "[26] Tianye Li, Timo Bolkart, Michael. J. Black, Hao Li, and Javier Romero. Learning a model of facial shape and expression from 4D scans. 2017. 3, 6" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 317, + 692, + 553, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 692, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 317, + 692, + 553, + 714 + ], + "type": "text", + "content": "[27] Weichuang Li, Longhao Zhang, Dong Wang, Bin Zhao, Zhi-gang Wang, Mulin Chen, Bang Zhang, Zhongjian Wang," + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "10720" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 115 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 115 + ], + "type": "text", + "content": "Liefeng Bo, and Xuelong Li. One-shot high-fidelity talking-head synthesis with deformable neural radiance field. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 116, + 295, + 147 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 116, + 295, + 147 + ], + "spans": [ + { + "bbox": [ + 56, + 116, + 295, + 147 + ], + "type": "text", + "content": "[28] Xueting Li, Shalini De Mello, Sifei Liu, Koki Nagano, Umar Iqbal, and Jan Kautz. Generalizable one-shot neural head avatar. 2023. 3, 6, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 148, + 294, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 148, + 294, + 191 + ], + "spans": [ + { + "bbox": [ + 56, + 148, + 294, + 191 + ], + "type": "text", + "content": "[29] Jingyun Liang, Yuchen Fan, Xiaoyu Xiang, Rakesh Ranjan, Eddy Ilg, Simon Green, Jiezhang Cao, Kai Zhang, Radu Timofte, and Luc Van Gool. Recurrent video restoration transformer with guided deformable attention. 2022. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 191, + 295, + 232 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 191, + 295, + 232 + ], + "spans": [ + { + "bbox": [ + 56, + 191, + 295, + 232 + ], + "type": "text", + "content": "[30] C.Z. Lin, D.B. Lindell, E.R. Chan, and G. Wetzstein. 3d gan inversion for controllable portrait image animation. In ECCV Workshop on Learning to Generate 3D Shapes and Scenes, 2022. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 233, + 294, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 233, + 294, + 266 + ], + "spans": [ + { + "bbox": [ + 56, + 233, + 294, + 266 + ], + "type": "text", + "content": "[31] S. Ma, T. Simon, J. Saragih, D. Wang, Y. Li, F. La Torre, and Y. Sheikh. Pixel codec avatars. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 266, + 294, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 266, + 294, + 297 + ], + "spans": [ + { + "bbox": [ + 56, + 266, + 294, + 297 + ], + "type": "text", + "content": "[32] Zhiyuan Ma, Xiangyu Zhu, Guojun Qi, Zhen Lei, and Lei Zhang. Otavatar: One-shot talking face avatar with controllable tri-plane rendering. 2023. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 298, + 294, + 339 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 298, + 294, + 339 + ], + "spans": [ + { + "bbox": [ + 56, + 298, + 294, + 339 + ], + "type": "text", + "content": "[33] Andrew Maimone, Jonathan Bidwell, Kun Peng, and Henry Fuchs. Enhanced personal autostereoscopic telepresence system using commodity depth cameras. Computers & Graphics, 2012. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 340, + 294, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 340, + 294, + 392 + ], + "spans": [ + { + "bbox": [ + 56, + 340, + 294, + 392 + ], + "type": "text", + "content": "[34] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing scenes as neural radiance fields for view synthesis. In European Conference on Computer Vision (ECCV), 2020. 2, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 393, + 293, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 393, + 293, + 414 + ], + "spans": [ + { + "bbox": [ + 56, + 393, + 293, + 414 + ], + "type": "text", + "content": "[35] NVIDIA. Nvidia maxine ar sdk. https://github.com/NVIDIA/MAXINE-AR-SDK, 2024.6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 415, + 294, + 467 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 415, + 294, + 467 + ], + "spans": [ + { + "bbox": [ + 56, + 415, + 294, + 467 + ], + "type": "text", + "content": "[36] Roy Or-El, Xuan Luo, Mengyi Shan, Eli Shechtman, Jeong Joon Park, and Ira Kemelmacher-Shlizerman. StyleSDF: High-Resolution 3D-Consistent Image and Geometry Generation. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 468, + 294, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 468, + 294, + 552 + ], + "spans": [ + { + "bbox": [ + 56, + 468, + 294, + 552 + ], + "type": "text", + "content": "[37] Sergio Orts-Escalano, Christoph Rhemann, Sean Fanello, Wayne Chang, Adarsh Kowdle, Yury Degtyarev, David Kim, Philip L. Davidson, Sameh Khamis, Mingsong Dou, Vladimir Tankovich, Charles Loop, Qin Cai, Philip A. Chou, Sarah Mennicken, Julien Valentin, Vivek Pradeep, Shenlong Wang, Sing Bing Kang, Pushmeet Kohli, Yuliya Lutchyn, Cem Keskin, and Shahram Izadi. Holoportation: Virtual 3d teleportation in real-time. In UIST 2016, 2016. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 553, + 294, + 594 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 553, + 294, + 594 + ], + "spans": [ + { + "bbox": [ + 56, + 553, + 294, + 594 + ], + "type": "text", + "content": "[38] Keunhong Park, Utkarsh Sinha, Jonathan T. Barron, Sofien Bouaziz, Dan B Goldman, Steven M. Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. 2021. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 596, + 294, + 638 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 596, + 294, + 638 + ], + "spans": [ + { + "bbox": [ + 56, + 596, + 294, + 638 + ], + "type": "text", + "content": "[39] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-NeRF: Neural Radiance Fields for Dynamic Scenes. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 639, + 294, + 670 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 639, + 294, + 670 + ], + "spans": [ + { + "bbox": [ + 56, + 639, + 294, + 670 + ], + "type": "text", + "content": "[40] Luchao Qi, Jiaye Wu, Annie N. Wang, Shengze Wang, and Roni Sengupta. My3dgen: A scalable personalized 3d generative model, 2023. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "type": "text", + "content": "[41] Shenhan Qian, Tobias Kirschstein, Liam Schoneveld, Davide Davoli, Simon Giebenhain, and Matthias Nießner. Gaussianavatars: Photorealistic head avatars with rigged 3d gaussians. 2024. 2" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 115 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 115 + ], + "type": "text", + "content": "[42] Anurag Ranjan and Michael J Black. Optical flow estimation using a spatial pyramid network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4161-4170, 2017. 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 116, + 553, + 168 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 116, + 553, + 168 + ], + "spans": [ + { + "bbox": [ + 316, + 116, + 553, + 168 + ], + "type": "text", + "content": "[43] Ramesh Raskar, Greg Welch, Matt Cutts, Adam Lake, Lev Stesin, and Henry Fuchs. The office of the future: a unified approach to image-based modeling and spatially immersive displays. In ACM Transactions on Graphics (SIGGRAPH), 1998. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 170, + 553, + 212 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 170, + 553, + 212 + ], + "spans": [ + { + "bbox": [ + 316, + 170, + 553, + 212 + ], + "type": "text", + "content": "[44] Daniel Rebain, Mark Matthews, Kwang Moo Yi, Dmitry Lagan, and Andrea Tagliasacchi. Lolnerf: Learn from one look. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 213, + 553, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 213, + 553, + 255 + ], + "spans": [ + { + "bbox": [ + 316, + 213, + 553, + 255 + ], + "type": "text", + "content": "[45] Shunsuke Saito, Gabriel Schwartz, Tomas Simon, Junxuan Li, and Giljoo Nam. Relightable gaussian codec avatars. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 255, + 553, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 255, + 553, + 297 + ], + "spans": [ + { + "bbox": [ + 316, + 255, + 553, + 297 + ], + "type": "text", + "content": "[46] Aliaksandr Siarohin, Stéphane Lathuilière, Sergey Tulyakov, Elisa Ricci, and Nicu Sebe. First order motion model for image animation. In Advances in Neural Information Processing Systems (NeurIPS), 2019. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 297, + 553, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 297, + 553, + 338 + ], + "spans": [ + { + "bbox": [ + 316, + 297, + 553, + 338 + ], + "type": "text", + "content": "[47] Ivan Skorokhodov, Sergey Tulyakov, Yiqun Wang, and Peter Wonka. Epigraph: Rethinking training of 3d gans. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 339, + 553, + 382 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 339, + 553, + 382 + ], + "spans": [ + { + "bbox": [ + 316, + 339, + 553, + 382 + ], + "type": "text", + "content": "[48] Michael Stengel, Koki Nagano, Chao Liu, Matthew Chan, Alex Trevithick, Shalini De Mello, Jonghyun Kim, and David Luebke. AI-Mediated 3D Video Conferencing. In ACM SIGGRAPH 2023 Emerging Technologies, 2023. 1, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 383, + 553, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 383, + 553, + 425 + ], + "spans": [ + { + "bbox": [ + 316, + 383, + 553, + 425 + ], + "type": "text", + "content": "[49] Jingxiang Sun, Xuan Wang, Yichun Shi, Lizhen Wang, Jue Wang, and Yebin Liu. Ide-3d: Interactive disentangled editing for high-resolution 3d-aware portrait synthesis. ACM Transactions on Graphics (SIGGRAPH ASIA), 2022. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 425, + 553, + 478 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 425, + 553, + 478 + ], + "spans": [ + { + "bbox": [ + 316, + 425, + 553, + 478 + ], + "type": "text", + "content": "[50] Jingxiang Sun, Xuan Wang, Lizhen Wang, Xiaoyu Li, Yong Zhang, Hongwen Zhang, and Yebin Liu. Next3d: Generative neural texture rasterization for 3d-aware head avatars. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 3, 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 479, + 553, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 479, + 553, + 520 + ], + "spans": [ + { + "bbox": [ + 316, + 479, + 553, + 520 + ], + "type": "text", + "content": "[51] Phong Tran, Egor Zakharov, Long-Nhat Ho, Anh Tuan Tran, Liwen Hu, and Hao Li. Voodoo 3d: Volumetric portrait disentanglement for one-shot 3d head reenactment. 2024. 2, 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 521, + 553, + 584 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 521, + 553, + 584 + ], + "spans": [ + { + "bbox": [ + 316, + 521, + 553, + 584 + ], + "type": "text", + "content": "[52] Alex Trevithick, Matthew Chan, Michael Stengel, Eric R. Chan, Chao Liu, Zhiding Yu, Sameh Khamis, Manmohan Chandraker, Ravi Ramamoorthi, and Koki Nagano. Realtime radiance fields for single-image portrait view synthesis. In ACM Transactions on Graphics (SIGGRAPH), 2023. 1, 2, 3, 4, 6, 7, 8" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 585, + 553, + 628 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 585, + 553, + 628 + ], + "spans": [ + { + "bbox": [ + 316, + 585, + 553, + 628 + ], + "type": "text", + "content": "[53] Ting-Chun Wang, Arun Mallya, and Ming-Yu Liu. One-shot free-view neural talking-head synthesis for video conferencing. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 628, + 553, + 670 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 628, + 553, + 670 + ], + "spans": [ + { + "bbox": [ + 316, + 628, + 553, + 670 + ], + "type": "text", + "content": "[54] Yaohui Wang, Di Yang, Francois Bremond, and Antitza Dantcheva. Latent image animator: Learning to animate images via latent space navigation. In International Conference on Learning Representations (ICLR), 2022. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "text", + "content": "[55] Jianfeng Xiang, Jiaolong Yang, Yu Deng, and Xin Tong. Gram-hd: 3d-consistent image generation at high resolution with generative radiance manifolds. arXiv preprint arXiv:2206.07255, 2022. 3" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "10721" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 115 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 115 + ], + "type": "text", + "content": "[56] Yinghao Xu, Sida Peng, Ceyuan Yang, Yujun Shen, and Bolei Zhou. 3d-aware image synthesis via learning structural and textural representations. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 117, + 294, + 170 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 117, + 294, + 170 + ], + "spans": [ + { + "bbox": [ + 56, + 117, + 294, + 170 + ], + "type": "text", + "content": "[57] Yuelang Xu, Benwang Chen, Zhe Li, Hongwen Zhang, Lizhen Wang, Zerong Zheng, and Yebin Liu. Gaussian head avatar: Ultra high-fidelity head avatar via dynamic gaussians. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 171, + 294, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 171, + 294, + 224 + ], + "spans": [ + { + "bbox": [ + 56, + 171, + 294, + 224 + ], + "type": "text", + "content": "[58] Zhongcong Xu, Jianfeng Zhang, Junhao Liew, Wenqing Zhang, Song Bai, Jiashi Feng, and Mike Zheng Shou. Pv3d: A 3d generative model for portrait video generation. In The Tenth International Conference on Learning Representations, 2023. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 225, + 294, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 225, + 294, + 289 + ], + "spans": [ + { + "bbox": [ + 56, + 225, + 294, + 289 + ], + "type": "text", + "content": "[59] Zhenhui Ye, Tianyun Zhong, Yi Ren, Jiaqi Yang, Weichuang Li, Jiangwei Huang, Ziyue Jiang, Jinzheng He, Rongjie Huang, Jinglin Liu, Chen Zhang, Xiang Yin, Zejun Ma, and Zhou Zhao. Real3d-portrait: One-shot realistic 3d talking portrait synthesis. In International Conference on Learning Representations (ICLR), 2024. 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 290, + 294, + 332 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 290, + 294, + 332 + ], + "spans": [ + { + "bbox": [ + 56, + 290, + 294, + 332 + ], + "type": "text", + "content": "[60] Fei Yin, Yong Zhang, Xiaodong Cun, Mingdeng Cao, Yanbo Fan, Xuan Wang, Qingyan Bai, Baoyuan Wu, Jue Wang, and Yujiu Yang. Styleheat: One-shot high-resolution editable talking face generation via pre-trained stylegan. 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 334, + 294, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 334, + 294, + 387 + ], + "spans": [ + { + "bbox": [ + 56, + 334, + 294, + 387 + ], + "type": "text", + "content": "[61] Wangbo Yu, Yanbo Fan, Yong Zhang, Xuan Wang, Fei Yin, Yunpeng Bai, Yan-Pei Cao, Ying Shan, Yang Wu, Zhongqian Sun, and Baoyuan Wu. Nofa: Nerf-based one-shot facial avatar reconstruction. In ACM SIGGRAPH 2023 Conference Proceedings, 2023. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 388, + 294, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 388, + 294, + 430 + ], + "spans": [ + { + "bbox": [ + 56, + 388, + 294, + 430 + ], + "type": "text", + "content": "[62] Egor Zakharov, Aliaksandra Shysheya, Egor Burkov, and Victor Lempitsky. Few-shot adversarial learning of realistic neural talking head models. In IEEE International Conference on Computer Vision (ICCV), 2019. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 431, + 294, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 431, + 294, + 474 + ], + "spans": [ + { + "bbox": [ + 56, + 431, + 294, + 474 + ], + "type": "text", + "content": "[63] Egor Zakharov, Aleksei Ivakhnenko, Aliaksandra Shysheya, and Victor Lempitsky. Fast bi-layer neural synthesis of one-shot realistic head avatars. In European Conference on Computer Vision (ECCV), 2020." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 475, + 294, + 517 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 475, + 294, + 517 + ], + "spans": [ + { + "bbox": [ + 56, + 475, + 294, + 517 + ], + "type": "text", + "content": "[64] Bowen Zhang, Chenyang Qi, Pan Zhang, Bo Zhang, Hsiang-Tao Wu, Dong Chen, Qifeng Chen, Yong Wang, and Fang Wen. Metaportrait: Identity-preserving talking head generation with fast personalized adaptation, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 518, + 294, + 571 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 518, + 294, + 571 + ], + "spans": [ + { + "bbox": [ + 56, + 518, + 294, + 571 + ], + "type": "text", + "content": "[65] Xuanmeng Zhang, Zhedong Zheng, Daiheng Gao, Bang Zhang, Pan Pan, and Yi Yang. Multi-view consistent generative adversarial networks for 3d-aware image synthesis. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 572, + 294, + 614 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 572, + 294, + 614 + ], + "spans": [ + { + "bbox": [ + 56, + 572, + 294, + 614 + ], + "type": "text", + "content": "[66] Jian Zhao and Hui Zhang. Thin-plate spline motion model for image animation. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 3657-3666, 2022. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 616, + 294, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 616, + 294, + 669 + ], + "spans": [ + { + "bbox": [ + 56, + 616, + 294, + 669 + ], + "type": "text", + "content": "[67] Yufeng Zheng, Victoria Fernández Abrevaya, Marcel C. Bühler, Xu Chen, Michael J. Black, and Otmar Hilliges. I M Avatar: Implicit morphable head avatars from videos. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "type": "text", + "content": "[68] Peng Zhou, Lingxi Xie, Bingbing Ni, and Qi Tian. CIPS-3D: A 3D-Aware Generator of GANs Based on Conditionally-Independent Pixel Synthesis. arXiv preprint arXiv:2110.09788, 2021. 3" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 106 + ], + "type": "text", + "content": "[69] Yiyu Zhuang, Hao Zhu, Xusen Sun, and Xun Cao. Mofanerf: Morphable facial neural radiance field. In European Conference on Computer Vision (ECCV), 2022. 3" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10722" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/ColabSfM_ Collaborative Structure-from-Motion by Point Cloud Registration/52ce366e-9afd-441c-9c7e-e1d8e8bedeb7_content_list.json b/2025/ColabSfM_ Collaborative Structure-from-Motion by Point Cloud Registration/52ce366e-9afd-441c-9c7e-e1d8e8bedeb7_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..fe75cfea8f015a51b702cd043d9a12c654263fec --- /dev/null +++ b/2025/ColabSfM_ Collaborative Structure-from-Motion by Point Cloud Registration/52ce366e-9afd-441c-9c7e-e1d8e8bedeb7_content_list.json @@ -0,0 +1,1284 @@ +[ + { + "type": "text", + "text": "ColabSfM: Collaborative Structure-from-Motion by Point Cloud Registration", + "text_level": 1, + "bbox": [ + 107, + 130, + 890, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Johan Edstedt\\*1 André Mateus2 Alberto Jaenal2", + "bbox": [ + 292, + 185, + 702, + 204 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Linköping University $^{2}$ Ericsson Research, Sweden", + "bbox": [ + 315, + 212, + 676, + 229 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/63d8556a7d3a7ee377336bc1d92af36db15e3b370e54ae4c18f88e5a2d2a60ec.jpg", + "image_caption": [ + "Figure 1. Our proposed registration paradigm for collaborative SfM reconstructions (ColabSfM). Given two input SfM reconstructions $\\mathcal{P},\\mathcal{Q}$ of the same scene, the task is to estimate the relative similarity transform $(s,R,t)$ between them. Our first contribution is to address this as a point cloud registration problem, using only 3D SfM tracks. For this, we do not rely on the visual descriptors, but on the 3D coordinates of the points $\\mathbf{P},\\mathbf{Q}$ , their normals $\\mathbf{N},\\mathbf{M}$ and, optionally, but not necessarily, features $\\mathbf{X},\\mathbf{Y}$ . To make point cloud registration methods perform well on this task, we propose as our second contribution a scalable pipeline to construct synthetic training datasets for SfM registration. Finally, we propose an improved version of RoITr [64] as registration method $f_{\\theta}$ ." + ], + "image_footnote": [], + "bbox": [ + 109, + 261, + 888, + 500 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 603, + 328, + 619 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Structure-from-Motion (SfM) is the task of estimating 3D structure and camera poses from images. We define Collaborative SfM (ColabSfM) as sharing distributed SfM reconstructions. Sharing maps requires estimating a joint reference frame, which is typically referred to as registration. However, there is a lack of scalable methods and training datasets for registering SfM reconstructions. In this paper, we tackle this challenge by proposing the scalable task of point cloud registration for SfM reconstructions. We find that current registration methods cannot register SfM point clouds when trained on existing datasets. To this end, we propose a SfM registration dataset generation pipeline, leveraging partial reconstructions from synthetically generated camera trajectories for each scene. Finally, we propose a simple but impactful neural refiner on top of the SotA registration method RoITr that yields significant improve", + "bbox": [ + 88, + 630, + 485, + 875 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ments, which we call RefineRoITr. Our extensive experimental evaluation shows that our proposed pipeline and model enables ColabSfM. Code is available at https://github.com/EricssonResearch/ColabSfM", + "bbox": [ + 511, + 604, + 908, + 666 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 691, + 645, + 708 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "For robots and extended reality (XR) devices to become ubiquitous in our lives, they must be able to accurately build and localize in maps of the environment. Most available solutions rely on visual data and onboard computing, which can limit the accuracy in the context of small devices. Thus, in the last years there has been a push from industry to move computation to the cloud (e.g. Google's Visual Positioning System [40], Microsoft's Azure Spatial Anchors [27], and Niantic's Lightship [47]). However, the heterogeneous nature of these systems prevents devices of different vendors from interoperating, i.e., sharing their maps and localizing in others' maps. If we are to make visual localization in-", + "bbox": [ + 509, + 719, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Work done during an internship at Ericsson Research.", + "bbox": [ + 107, + 886, + 400, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "6573", + "bbox": [ + 482, + 944, + 514, + 957 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/70d2831a980b6a8bb1a51e38bb366823f1fde1599554af37c6432bab8b1c5471.jpg", + "image_caption": [ + "Figure 2. Qualitative comparison. We compare our approach to the previous point cloud registration method OverlapPredator [24] on the St. Peter's Basilica test scene. Without training on our proposed SfM registration dataset (column 1), previous methods are unable to produce sufficiently good matches (top row) and accurate relative pose estimation results (bottom row). In contrast, our proposed model RefineRoITr, trained on the proposed dataset, is able to find better matches and hence register the scenes well. The source and target point clouds are depicted in yellow and blue, respectively." + ], + "image_footnote": [], + "bbox": [ + 106, + 85, + 890, + 406 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "teroperable, we need ways to merge such maps. Structurefrom-Motion (SfM) [44] is the most used method to process images and produce 3D maps (scene reconstructions consisting of a set of 3D points and the poses of the images expressed in a common coordinate system). To enable localization, a visual descriptor is stored for each map point, and for localization, 2D-3D matches are found by matching the descriptors of the query image with those of the map. Typically, SfM maps are merged using these visual descriptors, by finding 3D-3D matches [22, 33], or inserting new sub-maps based on 2D-3D correspondences [50].", + "bbox": [ + 88, + 518, + 480, + 685 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, these methods have limitations. First, different SfM pipelines tend to use different feature extractors, which yield incompatible descriptors. This is partially addressed by the cross-descriptor approach proposed in [17], but the assumption that the descriptors are computed from the same keypoints is unrealistic. Additionally, exposing visual descriptors may allow feature inversion attacks [12, 15, 38, 58], which can recover private elements from the images. The second limitation concerns scalability, since storing visual descriptors leads to a map size increase between two and three orders of magnitude [67]; and matching using a traditional pipeline [35] may suffer in performance; which can be enhanced by using learned methods [31, 43] at the cost of excessive runtimes. Visual", + "bbox": [ + 89, + 688, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "localization without descriptors was tackled in [55, 66, 67], demonstrating competitive performance, but since they rely solely on 2D-3D geometry, which suffers from perspective distortions, their accuracy is not as high as their descriptor-based counterparts. Furthermore, they still rely on image retrieval, which again raises privacy, scalability, and interoperability concerns. Nevertheless, their results show that geometry alone could be sufficient for matching. Based on that observation, we pose the following question", + "bbox": [ + 511, + 518, + 906, + 656 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Can we merge SfM maps using only 3D geometry?", + "text_level": 1, + "bbox": [ + 547, + 665, + 883, + 680 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To this end, we propose to merge SfM 3D maps through 3D point cloud registration methods [1, 65], which rely solely on geometric information, see Fig. 1. We call our proposed paradigm Collaborative SfM (ColabSfM). In contrast to visual localization, which aims to localize a single image to a map, our proposed task involves registering a map to a map, and is image retrieval-free.", + "bbox": [ + 511, + 688, + 906, + 794 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "State of the art methods for 3D point cloud registration rely on learned features [24, 39, 64], therefore it is necessary to provide suitable datasets to train such networks. There is a wide range of datasets available [34] for the point cloud registration task. However, those mostly consist of 3D sensor data, e.g., RGB-D [65] and/or LiDAR scans [20], which differ from SfM point clouds. To the best of our knowl", + "bbox": [ + 511, + 795, + 908, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "6574", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "edge, there are no large datasets for SfM point cloud registration. SfM datasets [30] provide a 3D point cloud for each scene, but since each scene is disconnected, no matching is possible between different scenes. To tackle this issue, we propose a pipeline to generate synthetic datasets for SfM point cloud registration, which generates multiple synthetic trajectories for each scene. The synthetic trajectories provide a trade-off between random internet images and sequential images (e.g., Cambridge Landmarks Dataset [28]). We apply our pipeline and generate training and benchmarks datasets from MegaDepth [30] and a benchmark dataset from Quad6k [11]. Our pipeline is presented in detail in Sec. 3.", + "bbox": [ + 89, + 90, + 480, + 286 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "SfM reference frames may be arbitrarily aligned relative to each other. As such, using a registration model invariant to rotation and translation SE(3), and robust to scale Sim(3), is important. To this end, we build on top of the recently proposed SotA rotation invariant model RoITr [64]. While this method already produces good results when trained on our dataset, we find that we can further improve results on ColabSfM by introducing a neural refinement stage to the model. We introduce our baseline model and our improved version, which we call RefineRoITr, in Sec. 4.", + "bbox": [ + 89, + 287, + 480, + 438 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To summarize, our contributions are:", + "bbox": [ + 109, + 439, + 356, + 452 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. A simple but powerful pipeline to synthetically generate SfM registration datasets, detailed in Sec. 3.", + "2. An improved model for SfM Registration, detailed in Sec. 4.", + "3. The task of SfM point cloud registration, with extensive experiments in Sec. 5 demonstrating its potential." + ], + "bbox": [ + 89, + 453, + 482, + 544 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 558, + 232, + 573 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Point cloud registration is typically defined as estimating the rigid transform that aligns two 3D point clouds to a joint reference frame. The classic approach is the Iterative Closest Point (ICP) algorithm [3, 6], which estimates the pose by alternating between computing point correspondences using the current pose and estimating the pose using those correspondences. Several improvements have been proposed over the years to handle outliers and to attempt to find a global minimum [29, 60]. More recently, point cloud registration has been addressed with neural networks, such as DCP [56] and its iterative extension [57], with similar methods presented in [19, 62]. Another class of iterative methods [1, 25, 59], extract global features from each point cloud, and exploits them to regress the pose.", + "bbox": [ + 89, + 583, + 482, + 794 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "An alternative approach, which will be our main focus, consists of finding and extracting correspondences between the two point clouds, being those points [37, 45], planes and lines [36], and finding the pose using robust estimators such as RANSAC [18]. Several point correspondence methods have been proposed, both handcrafted [16, 41] and learned [7, 14, 21]. In [24] an overlap-attention block is", + "bbox": [ + 89, + 795, + 482, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "proposed to improve matching of point clouds with low overlap. Overlap-aware methods are also presented in [26, 39, 63], which focus on improving the matching accuracy and thus avoid the use of RANSAC to filter outliers. The previous methods are not robust to large rotations, which are common in SfM. This issue has been tackled by proposing methods that are rotation invariant [4, 13, 64]. In this paper we will mainly consider point pair features (PPFs) [16] to ensure invariance to global rigid transformations, which have recently shown SotA performance [64].", + "bbox": [ + 511, + 90, + 903, + 243 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Localization without descriptors has recently received renewed attention as a way to reduce computation, and preserve privacy. BPnPNet [5] jointly estimates the pose and the correspondences between 2D keypoints on a query image and 3D points from a scene point cloud. However, it is susceptible to outlier correspondences. This issue was tackled by Zhou et al. [67], who use only 2D keypoints and the 3D map point cloud to localize a query image to a 3D map. They do this by projecting the map point cloud into retrieved map images and estimating 2D-2D correspondences with a graph neural network. A similar method DGC-GNN [55] proposes to exploit geometry and color information to guide point matching by computing a geometric embedding of Euclidean and angular relations. These methods have the advantage of not requiring storage of descriptors. However, they still require image retrieval to get sufficiently close map images to the query. This is problematic if the map is sparse in the region of the query, and introduces additional complexity. Furthermore, the projection from 3D into 2D loses important geometrical information, which we retain by directly registering the point clouds.", + "bbox": [ + 511, + 246, + 903, + 564 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Collaborative mapping is a general term, typically referring to fusion of partial 3D reconstructions. Cohen et al. [8] use the Manhattan world assumption [10] to derive geometric constraints on partial reconstructions to enable model fusion. Cohen et al. [9] fuse indoor and outdoor reconstructions by first leveraging images to obtain semantic masks with window positions. They then match the windows between the two models to solve for the pose with RANSAC [18]. This approach requires images to be stored alongside the point clouds which brings additional space constraints [67]. Strecha et al. [48] and Untzelmann et al. [54] merge multiple building reconstructions to a joint map by combining GPS coordinates and known building outlines to construct an alignment optimization objective. This has the downside of restricting the reconstructions to buildings that have previously been mapped and requires geo-tags. Dusmanu et al. [17] investigated collaborative mapping and localization by translating descriptors to a joint embedding space. They consider a closed set of descriptors and learn a common embedding space for them. Collaborative mapping is achieved through embedding the descriptors into the joint space and running feature match-", + "bbox": [ + 511, + 568, + 906, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "6575", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ing on the joint space. This requires access to all features for all images, which is expensive, and the required compute scales as standard SfM with the number of images. In contrast to these methods, we make far fewer assumptions on the data available. Namely, we do not store any descriptors (neither local nor global), do not make assumptions on the structure of the scenes, and do not use any topological maps or data. Thus, we constrain ourselves to using only the geometry. Closest to our work Liu et al. [32] align line reconstructions and simulate partial reconstructions by adding noise and dropout of lines to a single reconstruction. In contrast we consider point-based reconstructions and propose a more realistic way to generate partial reconstructions through synthetic camera trajectories and retriangulation.", + "bbox": [ + 89, + 90, + 483, + 303 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3. Pipeline for SfM Registration Dataset Generation", + "text_level": 1, + "bbox": [ + 89, + 314, + 483, + 347 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We find that current SotA point cloud registration models trained on 3DMatch [65], the most commonly used training set for 3D registration, are not able to register SfM reconstruction out of the box (cf. Tabs. 1, 2 and 4). Intuitively this is due to the large distribution shift between registration of RGB-D and/or LiDAR scans, and the much larger scenes covered by SfM reconstructions. However, it is not obvious how to generate a diverse dataset with multiple aligned reconstructions. Manually collecting such datasets is prohibitively expensive, and publicly available datasets typically contain a single reconstruction per scene.", + "bbox": [ + 89, + 357, + 483, + 523 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To tackle this issue, we propose a pipeline to generate a large synthetic dataset for registration between SfM models from random image collection datasets, e.g., MegaDepth [30]. Our key insight is that we can leverage these models and for each scene to construct partial reconstructions by sampling subsets of images/c cameras. An overview of our approach is presented in Fig. 3. In the next sections we describe our approach in more detail.", + "bbox": [ + 89, + 523, + 483, + 643 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Creating Partial Reconstructions", + "text_level": 1, + "bbox": [ + 89, + 652, + 380, + 667 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We use hloc [42] with COLMAP [44] to create partial reconstructions from scenes. We use SIFT [35] and SOSNet [51] retriangulations for training and evaluation. We consider two approaches to create partial reconstructions.", + "bbox": [ + 89, + 674, + 483, + 734 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Random point sampling: In the simplest approach, we sample a set of 3D points from the scene for each partial reconstruction, and add the images in which they are visible until reaching approximately 200 images. For each scene we create 10 such partial reconstructions. We found that while this approach works well when registering reconstructions from collections of random images, it does not generalize as well to reconstructions from a single camera trajectory, which are common in real-life scenarios, e.g., video. We believe that this is due to the point clouds exhibiting more significant variance when the scene is cap-", + "bbox": [ + 89, + 734, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "tured from different viewpoints. For example, keypoints are likely to be denser for nearby cameras, and natural occlusions will depend on viewpoints. As random image based datasets are not naturally divisible into videos, it is not obvious how to bridge this gap. We next describe our approach addressing this.", + "bbox": [ + 511, + 90, + 903, + 181 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Partial Trajectories: Our approach to bridge the random image $\\rightarrow$ video gap is to generate synthetic trajectories from reconstructions of random images. Concretely, we randomly pick a starting image, and then sequentially (without replacement) pick the nearest neighbour of the image using a distance metric (which we refer to as dist). In practice we found that, for each trajectory, using a randomly weighted combination of geodesic rotation distance and Euclidean distance on the translation produced satisfactory results. Each trajectory has a size between $n_{\\mathrm{low}} = 75$ and $n_{\\mathrm{high}} = 300$ . The algorithm is presented in Algorithm 1. We show a qualitative example of three generated synthetic trajectories in Fig. 4. Running this on all scenes gave us between 5 to 20 synthetic trajectories per scene. Training is conducted on an equally weighted combination of the random point sampling approach and the synthetic trajectories.", + "bbox": [ + 511, + 181, + 906, + 425 + ], + "page_idx": 3 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1 Generation of synthetic trajectory from a set of posed images" + ], + "code_body": "Precondition: $\\mathbb{T}$ set of posed images, $\\mathbb{I}$ remaining index set, distance function dist \n1: function GENERATETRAJECTORY(T, I) \n2: $\\tau \\gets \\emptyset$ \n3: i $\\leftarrow$ Random sample from I \n4: n $\\leftarrow$ Random sample from $\\{n_{\\mathrm{low}},\\dots ,n_{\\mathrm{high}}\\}$ \n5: $\\mathbb{I}\\gets \\mathbb{I}\\backslash i$ \n6: $\\tau \\gets \\tau \\cup i$ \n7: for $k\\gets 1$ to $n$ do \n8: $i^{\\prime}\\gets \\arg \\min_{i^{\\prime}\\in \\mathbb{I}}\\mathrm{dist}(\\mathbb{T}_{i},\\mathbb{T}_{i^{\\prime}})$ \n9: $\\tau \\gets \\tau \\cup i^{\\prime}$ \n10: $\\mathbb{I}\\gets \\mathbb{I}\\backslash i^{\\prime}$ \n11: $i\\gets i^{\\prime}$ \n12: end for \n13: return $\\tau ,\\mathbb{I}$ \n14: end function", + "bbox": [ + 514, + 484, + 905, + 742 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Pipeline and Dataset", + "text_level": 1, + "bbox": [ + 513, + 772, + 709, + 787 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Reconstruction Pipeline: We are interested in having good ground truth correspondences between the partial reconstructions. As the SfM pipeline accuracy usually degrades with fewer available cameras, we use fixed camera poses from the joint reconstruction using all images and simply run triangulation from feature matching. This also has the benefit of being significantly less computationally", + "bbox": [ + 511, + 794, + 906, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "6576", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/8c6176577430faea20741dd3474d55eae1e91e84627eb5f3d7ddd10f6a8938ce.jpg", + "image_caption": [ + "Figure 3. Overview of our pipeline. For each scene, from a random image SfM dataset, e.g., MegaDepth [30], we retriangulate partial reconstructions using partial trajectories from the scene. Since these trajectories are in the global SfM reference frame, the relative transformation is simply the identity mapping." + ], + "image_footnote": [], + "bbox": [ + 173, + 87, + 823, + 280 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/5e26daff9cf077a7f0594b642cfd5aa9131102f539201cfac5a45660b73d5bd7.jpg", + "image_caption": [ + "Figure 4. Example of synthetic trajectories in our proposed dataset. We start with a large scale scene, consisting of hundreds of cameras (left). From this set of cameras we run Algorithm 1 until the remaining set of cameras is small. Plotting the trajectories shows that plausible camera motion is achieved by this procedure (right). Sampling in this way bridges the gap between random image collections and video-based trajectories." + ], + "image_footnote": [], + "bbox": [ + 94, + 353, + 472, + 478 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "demanding than running the full reconstructions.", + "bbox": [ + 89, + 614, + 413, + 630 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Dataset Construction: We run our pipeline on all 196 MegaDepth scenes where 10 of them are reserved for the test benchmark. We use either SIFT [35] or SOSNet [51] features for the retriangulation, however any local features could be used. We compute the overlap between all pairs of partial reconstructions. To train models we enforce the overlap to be larger than $30\\%$ . In total, after enforcing sufficient overlap, we are left with about 22000 pairs ( $\\sim$ 20000 for training and $\\sim$ 2000 for testing), which constitutes our dataset. Details on how point cloud overlap is computed are presented in the supplementary material.", + "bbox": [ + 89, + 631, + 483, + 797 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. RoITr and RefineRoITr for ColabSfM", + "text_level": 1, + "bbox": [ + 89, + 813, + 434, + 830 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we discuss our baseline and our proposed improvements for ColabSfM. An overview of the proposed registration model $f_{\\theta}$ is presented in Figure 5. For consistency we follow the notation of Yu et al. [64].", + "bbox": [ + 89, + 839, + 483, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Preliminaries", + "text_level": 1, + "bbox": [ + 511, + 356, + 653, + 371 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We are interested in registration methods that are robust to the challenges posed by SfM reconstructions. In particular, one of the main challenges is that SfM reconstructions are defined up-to-scale, and the rotation and origin of the reference frame is arbitrary1. Hence, we would like the registration model to be approximately invariant to the action of the similarity group, i.e., $\\mathrm{Sim}(3) = \\mathrm{SE}(3) + \\mathrm{scale}$ . In practice we will achieve this by using an SE(3) invariant model, and learn approximate scale invariance. We next describe our baseline, and then our proposed improvements.", + "bbox": [ + 511, + 378, + 906, + 531 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Baseline Model", + "text_level": 1, + "bbox": [ + 511, + 542, + 668, + 556 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As a strong baseline, we base our approach on the SotA SE(3) invariant registration method RoITr [64]. We next give an overview of our baseline. An more thorough description of the architecture is given in the supplementary and by Yu et al. [64].", + "bbox": [ + 511, + 565, + 906, + 640 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Rotation Invariance through PPFs: RoITr [64] is based on point pair features (PPFs) [16], which computes a 4D feature for point pairs by the angles between their relative position and their normals. Mathematically, the PPF is defined as $\\mathrm{PPF}(p_1,p_2,n_1,n_2) =$ $[\\| d\\| ,\\angle (n_1,d),\\angle (n_2,d),\\angle (n_1,n_2)]$ , where $d = p_{2} - p_{1}$ It can easily be shown that this construction is invariant to rotations,", + "bbox": [ + 511, + 641, + 908, + 762 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\operatorname {P P F} \\left(R p _ {1}, R p _ {2}, R n _ {1}, R n _ {2}\\right) = [ \\| R d \\|, \\angle \\left(R n _ {1}, R d\\right), \\\\ \\angle (R n _ {2}, R d), \\angle (R n _ {1}, R n _ {2}) ] = [ \\| d \\|, \\angle (n _ {1}, d), \\angle (n _ {2}, d), \\\\ \\left. \\angle \\left(n _ {1}, n _ {2}\\right) \\right] = \\operatorname {P P F} \\left(p _ {1}, p _ {2}, n _ {1}, n _ {2}\\right), \\tag {1} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 527, + 779, + 903, + 834 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "since norms and angles are invariant to rotations.", + "bbox": [ + 511, + 847, + 836, + 861 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "1Although in practice, approximate scale can be obtained by, e.g., standardization, objects with known size, or additional sensors.", + "bbox": [ + 511, + 875, + 906, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "6577", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/624943ce69130c89c89e4623403e4a9cae83d8a5ef7844f1d3cfc722152ba5d4.jpg", + "image_caption": [ + "Figure 5. Overview of our proposed model RefineRoITr. As input we take two point clouds $\\mathcal{P} = (\\mathbf{P},\\mathbf{N},\\mathbf{X})$ $\\mathcal{Q} = (\\mathbf{Q},\\mathbf{M},\\mathbf{Y})$ consisting of 3D points $\\mathbf{P} / \\mathbf{Q}\\in \\mathbb{R}^{n / m\\times 3}$ , and normals $\\mathbf{N} / \\mathbf{M}\\in \\mathbb{R}^{n / m\\times 3}$ . In this work the features $\\mathbf{X} / \\mathbf{Y}$ are always assumed to be constant. We ensure rotation invariance by Point Pair Feature (PPF) encoding. The PPFs are fed together with the features into an encoder $e_{\\theta}$ , producing $N^{\\prime}$ coarse superpoints $\\mathcal{P}'/\\mathcal{Q}'$ . These are passed through a global Transformer $g_{\\theta}$ , from which the top- $k$ coarse correspondences $\\mathcal{C}'\\in \\mathbb{R}^{k\\times 2}$ are extracted. A decoder $d_{\\theta}$ takes the coarse features from $g_{\\theta}$ and the finer features from $e_{\\theta}$ and produces fine features $\\hat{\\mathbf{X}},\\hat{\\mathbf{Y}}$ . Using the coarse correspondences we extract neighbourhoods $\\hat{\\mathbf{G}}^X,\\hat{\\mathbf{G}}^Y\\in \\mathbb{R}^{k\\times 64\\times c}$ , which we feed into our proposed refinement Transformer. The refined features from the Transformer are used to construct a cost matrix, where the Sinkhorn algorithm [43, 54, 64] is used to solve the optimal transport (OT) problem, producing our final matches $\\mathcal{C}$ . The network is trained using $\\mathcal{L}_s + \\mathcal{L}_p$ ." + ], + "image_footnote": [], + "bbox": [ + 130, + 97, + 867, + 244 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Normalization Procedure: SfM reconstructions do not in general come with metric scale. This poses several challenges for registration methods, that typically assume both point clouds to share the same scale. We tackle this by normalization. Specifically, for Sim(3) training we normalize the point clouds $\\mathbf{P}$ , $\\mathbf{Q}$ with $\\sigma_{\\mathbf{P}} / \\sqrt{2}$ , $\\sigma_{\\mathbf{Q}} / \\sqrt{2}$ , where $\\sigma$ is the largest singular value of the centered point cloud. For SE(3) training, we simply use $\\sigma_{\\mathbf{P}}$ to normalize both $\\mathbf{P}$ and $\\mathbf{Q}$ . We find that this simple approach leads to models robust to scale variation, as we demonstrate in Tab. 1.", + "bbox": [ + 88, + 390, + 480, + 541 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Normal Estimation: We follow RoITr [64] and use Open3D [68] with a neighborhood size of 33 for estimating normals, we found that this works well in practice. In contrast to the RGB-D scans of 3DMatch, our dataset consists of point clouds from multiple sensors. This means that the simple approach of aligning the normal orientation towards the origin of the coordinate system, as done in RoITr, will not be consistent in general. To remedy this, we use the fact that all cameras observing a 3D track will yield consistent orientation estimates. In practice, we align the orientation towards the camera center of a random image observing the track, which we found yields consistent orientation.", + "bbox": [ + 88, + 545, + 482, + 726 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Overview of Architecture: RoITr consists of an encoder $e_{\\theta}$ , a global Transformer $g_{\\theta}$ , and a decoder $d_{\\theta}$ . $e_{\\theta}$ and $d_{\\theta}$ take as input sets of PPF encoded local neighbourhoods (typically around 8/16 points) around each point/superpoint. In contrast, $g_{\\theta}$ performs global attention on the coarse point cloud $\\hat{\\mathcal{P}}$ . $e_{\\theta}, d_{\\theta}, g_{\\theta}$ all use the Transformer architectures as described by Yu et al. [64]. Note that RoITr does not include the refining Transformer $r_{\\theta}$ , which will be detailed in our RefineRoITr.", + "bbox": [ + 88, + 729, + 482, + 864 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Losses: RoITr uses a superpoint matching loss $\\mathcal{L}_s$ and a point matching $\\mathcal{L}_p$ loss for training. We follow RoITr", + "bbox": [ + 88, + 869, + 482, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "and use their overlap-aware [39] circle loss [49] as $\\mathcal{L}_s$ . The point matching loss $\\mathcal{L}_p$ is the Negative Log Likelihood (NLL) of the GT correspondences after running the Sinkhorn algorithm [43] on the fine feature similarities $\\tilde{S}$ .", + "bbox": [ + 511, + 390, + 903, + 450 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. RefineRoITr (Ours)", + "text_level": 1, + "bbox": [ + 511, + 460, + 702, + 474 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Refinement Transformer $r_{\\theta}$ : The refinement in RoITr is performed by optimizing the Sinkhorn [54] distance between shallow features in a local neighbourhood around the coarse matches, denoted $\\hat{\\mathbf{G}}^X, \\hat{\\mathbf{G}}^Y$ . While this provides satisfactory results on 3DMatch, we found that we could improve results by incorporating a joint local Transformer $r_{\\theta}$ . This Transformer takes $\\hat{\\mathbf{G}}^X, \\hat{\\mathbf{G}}^Y$ as inputs and outputs enriched neighbourhoods $\\hat{\\mathbf{G}}^X, \\hat{\\mathbf{G}}^Y$ . We found that these cross-point cloud enriched fine features from $r_{\\theta}$ yielded improved results.", + "bbox": [ + 511, + 482, + 905, + 632 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Architectural Details of $r_{\\theta}$ : $r_{\\theta}$ takes in local neighbourhoods of shape $m \\times 64 \\times c$ , where $m$ is the number of neighbourhoods, 64 is the number of neighbours (identical to RoITr), and $c$ is the feature dimensionality. We use layers of interleaved self-attention and cross-attention as in LightGlue [31]. Note however, that in contrast to LightGlue that performs global attention, our refinement Transformer performs attention locally, which is significantly cheaper. We use an input/output dimensionality of $c_{\\mathrm{in}} = c_{\\mathrm{out}} = c = 64$ , and 4 self+cross attention layers. Each self/cross attention layer consists of single head attention, using dim $c$ , which is stacked with the input into a 1 hidden layer MLP (input dim $2c$ , hidden dim $2c$ , output dim $c$ ) using the GELU [23] activation function, with layer normalization [2] applied before GELU. Each self/cross block is residual, i.e., the output of the block is the output of the MLP plus the input.", + "bbox": [ + 511, + 633, + 906, + 875 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6578", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/d32433905fba16354b8e2537aafce4887ca5a4fe98694837d0a270a7181d5491.jpg", + "table_caption": [ + "Table 1. Results of SfM registration on MegaDepth. We evaluate 2 different scenarios. Unknown relative orientation SO(3) + unknown relative position = SE(3), SE(3) + unknown relative scale = Sim(3). The top portion contains methods only trained on the 3DMatch (3DM) dataset, the middle portion methods trained only on our proposed dataset (Mega), while the lower portion contains methods trained on the former and fine-tuned on the latter." + ], + "table_footnote": [], + "table_body": "
MethodSE(3)Sim(3)
IRFMRRRIRFMRRR
OverlapPredator [24] (3DM)6.135.510.03.621.32.1
GeoTransformer [39] (3DM)2.114.63.31.38.60.4
PareNet [61] (3DM)7.338.65.24.727.70.8
RoITr [64] (3DM)3.012.60.01.67.00.8
RefineRoITr (3DM)10.032.813.96.222.83.5
RefineRoITr (Mega)51.096.570.244.692.842.7
OverlapPredator [24] (3DM + Mega)21.374.935.411.456.210.7
RoITr [64] (3DM + Mega)44.690.960.038.486.138.7
RefineRoITr (3DM + Mega)48.795.167.744.693.944.3
", + "bbox": [ + 91, + 196, + 480, + 335 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 89, + 378, + 222, + 397 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we evaluate the RefineRoITr model presented in Sec. 4. Three different model instances were trained: 1) pre-trained RefineRoITr on 3DMatch [65]; 2) the model in 1) fine-tuned on the dataset in Sec. 3 and; 3) RefineRoITr trained from scratch on our dataset. For comparison, we used RoITr [64] and OverlapPredator [24], both with pre-trained weights from 3DMatch, and fine-tuned weights trained on our dataset. Additionally, we tested GeoTransformer [39] also pre-trained on 3DMatch [65]. As we found that the performance, of fine-tuning 3DMatch trained models, on our dataset was similar for our proposed model, we did not train other models in our dataset from scratch.", + "bbox": [ + 88, + 412, + 482, + 595 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Metrics: We use the standard registration metric Inlier Ratio (IR). This measures the percentage of estimated correspondences lying within a threshold distance $\\tau_{\\mathrm{IR}} = 0.1$ from each other. We additionally measure the Feature Matching Recall (FMR), which is defined as the percentage of point cloud pairs whose IR is above a threshold $\\tau_{\\mathrm{FMR}} = 5\\%$ . Finally, we use Registration Recall (RR), which consists of the percentage of pairs for which the rotation error and translation error are below $5^{\\circ}$ and $0.05$ , respectively. We solve for the pose using the RANSAC [18] implementation from Open3D [68] on the matches retrieved by the neural networks. Details on error computation and RANSAC can be found in the Supplemental material.", + "bbox": [ + 89, + 604, + 482, + 801 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Computational cost: On a desktop machine equipped with an i9-13900K @ 3.00GHz CPU and an NVIDIA 4090 GPU, the median runtime for RoITr is 99.94 ms and for RefineRoITris 102.84, which shows that the computational burden added by the modifications described in Sec. 4.3 is around $3\\%$ .", + "bbox": [ + 89, + 809, + 482, + 898 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.1. MegaDepth SfM Registration Benchmark", + "text_level": 1, + "bbox": [ + 511, + 90, + 870, + 107 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We select 10 scenes from MegaDepth [30] dataset, which consists of 196 large scale 3D scenes reconstructed from internet images for our benchmark. The test scenes are, in order, [0008, 0015, 0019, 0021, 0022, 0024, 0025, 0032, 0063, 1589]. The average size of the point clouds is $15.3\\mathrm{K}$ points, using either SIFT or SOSNet features for the triangulation. To train and evaluate the network presented in Sec. 4 on our dataset, we define two different scenarios SE(3) and Sim(3). We define a random rotation by sampling three Euler angles $\\alpha, \\beta, \\gamma \\in [0,2\\pi]$ .", + "bbox": [ + 509, + 113, + 903, + 265 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The results of the RefineRoITr method and the baselines are presented in Tab. 1. As expected, the methods which are not fine-tuned in our data perform the worst in both SE(3) and Sim(3) data. This shows the lack of generalization of current registration methods on SfM point clouds. When considering FMR, the methods trained from scratch or fine-tuned on our data obtain high scores ( $>90\\%$ ) on SE(3) data with slightly lower scores on Sim(3). When looking at the three variations of RefineRoITr, we see that while the version trained on 3DMatch outperforms vanilla RoITr, the major boost in performance is due to fine-tuning the model on our dataset. The best overall version is the one trained from scratch on our dataset, even though the improvement is not as significant. Qualitative results are presented in Fig. 2 and the Supplementary material.", + "bbox": [ + 511, + 266, + 906, + 493 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Standard Benchmark: Cambridge Landmarks", + "text_level": 1, + "bbox": [ + 511, + 507, + 903, + 523 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The Cambridge Landmarks [28] dataset consists of 5 large-scale 3D scenes recorded with a smartphone by a pedestrian in Cambridge, each of which consisting of a train and a test trajectories with camera poses. We retriangulate the train/test sequences independently, which results in two reconstructions for each scene. In this case, we evaluate the IR and number of matches. Here we only test the SE(3) case. Rotation and translation errors and qualitative results for this dataset are reported in the supplementary material. The results are presented in Tab. 2. Similar to the evaluation results in the previous section, the models trained only on 3DMatch [65] perform poorly, while the ones fine-tuned and or trained from scratch on our dataset present the highest IR and number of matches, with small differences in favor of the version trained from scratch.", + "bbox": [ + 511, + 530, + 906, + 757 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Noisy Indoor Reconstructions: 7-Scenes", + "text_level": 1, + "bbox": [ + 511, + 771, + 856, + 787 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The 7-Scenes [46] dataset is an indoor RGB-D dataset. This dataset is challenging for SfM, since textured regions lead to significantly more noisy point clouds, and less geometric features. We present results in Tab. 3. Encouragingly, we find that models trained using our pipeline perform well also on these indoor scenes, showing the generalizability of our approach.", + "bbox": [ + 511, + 794, + 903, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "6579", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/b869fec888bc2ba0f26af053082366cbd8a9173a70d33b669767992ce130c00f.jpg", + "table_caption": [ + "Table 2. Results of SfM registration on Cambridge Landmarks [28]. We evaluate unknown relative orientation $\\mathrm{SO}(3) +$ unknown relative position $= \\mathrm{{SE}}\\left( 3\\right)$ . The top portion contains methods only trained on the 3DMatch (3DM) dataset, the middle portion methods trained only on our proposed dataset (Mega), while the lower portion contains methods trained on the former and fine-tuned on the latter." + ], + "table_footnote": [], + "table_body": "
MethodGreat CourtKings CollegeOld HospitalShop FacadeSt Mary's Church
IRMatchesIRMatchesIRMatchesIRMatchesIRMatches
OverlapPredator [24] (3DM)1.43560.53720.03632.53540.2382
GeoTransformer [39] (3DM)0.03820.17850.04380.113680.0431
PureNet [61] (3DM)4.320001.62000020001.220006.72000
RoiTr [64] (3DM)0.02850.04470.03210.39591.5407
RefineRoiTr (3DM)0.24122.24080.05440.97372.8567
RefineRoiTr (Mega)70.9453857.6397731.5221441.6132581.85167
RoiTr [64] (3DM + Mega)52.1142639.6137721.960328.070064.52481
RefineRoiTr (3DM + Mega)69.4455154.0324628.0159139.1134277.85168
", + "bbox": [ + 210, + 141, + 782, + 271 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/1432871e0f8a88c437e493c8814857d0d950cc8d539fa4ef98ade6d799f47c39.jpg", + "table_caption": [ + "Table 3. Results of SfM registration on 7-Scenes. We evaluate unknown relative orientation $\\mathrm{SO}(3) +$ unknown relative position $= \\mathrm{{SE}}\\left( 3\\right)$ . The top portion contains methods only trained on the 3DMatch (3DM) dataset, the middle portion methods trained only on our proposed dataset (Mega), while the lower portion contains methods trained on the former and fine-tuned on the latter." + ], + "table_footnote": [], + "table_body": "
MethodChessFireHeadsOfficePumpkinRed KitchenStairs
IRMatchesIRMatchesIRMatchesIRMatchesIRMatchesIRMatchesIRMatches
RoITr [64] (3DM)0.03070.010040.016990.013220.03270.011630.0730
RefineRoITr (3DM)0.08420.022040.011040.010730.010160.017210.01043
RefineRoITr (Mega)23.4671738.8904350.6541968.0946114.9289327.980506.23230
RoITr [64] (3DM+Mega)20.4209830.5416536.4381250.047527.9127118.524389.11883
RefineRoITr (3DM+Mega)23.6563233.3833950.5599662.788878.9226826.866166.42863
", + "bbox": [ + 150, + 335, + 844, + 431 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/f4e78ae21f432629b3d687b8901527829f02b2f93ef38342a4239b77194c1f57.jpg", + "table_caption": [ + "Table 4. Results of SfM registration on Quad6k [11]. We evaluate unknown relative orientation $\\mathrm{SO}(3) +$ unknown relative position $=$ SE(3). The top portion contains methods only trained on the 3DMatch (3DM) dataset, the middle portion methods trained only on our proposed dataset (Mega), while the lower portion contains methods trained on the former and fine-tuned on the latter." + ], + "table_footnote": [], + "table_body": "
MethodIRFMRMatchesRR
OverlapPredator [24] (3DM)1.610.21210.0
GeoTransformer [39] (3DM)0.60.019360.0
PareNet [61] (3DM)1.610.420000.0
RoITr [64] (3DM)0.40.05740.0
RefineRoITr (3DM)1.76.16172.0
RefineRoITr (Mega)17.267.487822.4
RoITr [64] (3DM + Mega)14.263.349224.5
RefineRoITr (3DM + Mega)18.165.371516.3
", + "bbox": [ + 138, + 547, + 431, + 659 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.4. Low Overlap: Quad6k", + "text_level": 1, + "bbox": [ + 89, + 691, + 302, + 709 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The Quad6k dataset [11] consists of about 6500 images from the Arts Quad at Cornell University. To create partial reconstructions for testing, we first perform a full scene retriangulation using the refined poses provided by [52] and then follow the procedure in Sec. 3.1, generating a total of 24 trajectories, and 49 pairs. The results are presented in Tab. 4. Similar to the previous datasets, the performance of the methods has a substantial boost after training on the dataset presented in Sec. 3. GeoTransformer [39] trained only on 3DMatch [65] presents the most matches on average, but only a small fraction (0.6%) are inliers, and its FMR and RR are both zero. Our Quad6k benchmark is", + "bbox": [ + 89, + 719, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "more challenging and thus the performance is lower than in Tabs. 1 and 2, due to a low overlap of $16.5\\%$ , in comparison with the $61.2\\%$ of MegaDepth. Qualitative results for this dataset are shown in the supplementary material.", + "bbox": [ + 511, + 454, + 906, + 516 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 529, + 633, + 544 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We have proposed Collaborative Structure-from-Motion by Point Cloud Registration, ColabSfM, a challenging new task to enable scalable collaborative mapping and localization. By extensive experiments we showed that with our proposed dataset generation pipeline and SE(3) invariant model, this task is solvable, and showed that our model, while trained only on MegaDepth, is able to generalize to several challenging scenarios, such as noisy indoor reconstructions, and low-overlap registration. We believe our initial results will encourage further research on collaborative 3D reconstruction. Limitations: a) Our proposed model RefineRoITr is E(3) invariant like RoITr, and therefore can struggle with estimating correct matches in symmetric scenes. b) Our models are trained and evaluated on retriangulated partial reconstructions. In practice, issues such as drift occur, making global alignment more difficult, however we find that our model still performs well in this setting in Table 7.", + "bbox": [ + 511, + 554, + 906, + 825 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 511, + 840, + 669, + 857 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work was supported by the strategic research environment ELLIIT, funded by the Swedish government.", + "bbox": [ + 511, + 864, + 905, + 896 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "6580", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Yasuhiro Aoki, Hunter Goforth, Rangaprasad Arun Srivatsan, and Simon Lucey. Pointnetlk: Robust & efficient point cloud registration using pointnet. In IEEE Conf. Comput. Vis. Pattern Recog., pages 7163-7172, 2019. 2, 3", + "[2] Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. arXiv preprint arXiv:1607.06450, 2016. 6", + "[3] Paul J. Besl and Neil D. McKay. A method for registration of 3-d shapes. IEEE Trans. Pattern Anal. Mach. Intell., 14 (2):239-256, 1992. 3", + "[4] Georg Bokman and Fredrik Kahl. A case for using rotation invariant features in state of the art feature matchers. In IEEE Conf. Comput. Vis. Pattern Recog., pages 5110-5119, 2022. 3", + "[5] Dylan Campbell, Liu Liu, and Stephen Gould. Solving the blind perspective-n-point problem end-to-end with robust differentiable geometric optimization. In Eur. Conf. Comput. Vis., pages 244-261, 2020. 3", + "[6] Yang Chen and Gerard Medioni. Object modeling by registration of multiple range images. In IEEE Int'l Conf. Robotics and Auto., pages 2724-2729, 1991. 3", + "[7] Christopher Choy, Jaesik Park, and Vladlen Koltun. Fully convolutional geometric features. In Int. Conf. Comput. Vis., pages 8958-8966, 2019. 3", + "[8] Andrea Cohen, Torsten Sattler, and Marc Pollefeys. Merging the unmatched: Stitching visually disconnected sfm models. In Int. Conf. Comput. Vis., pages 2129-2137, 2015. 3", + "[9] Andrea Cohen, Johannes L Schonberger, Pablo Speciale, Torsten Sattler, Jan-Michael Frahm, and Marc Pollefeys. Indoor-outdoor 3d reconstruction alignment. In Eur. Conf. Comput. Vis., pages 285-300, 2016. 3", + "[10] James Coughlan and Alan L Yuille. The manhattan world assumption: Regularities in scene statistics which enable bayesian inference. Adv. Neural Inform. Process. Syst., 13, 2000. 3", + "[11] David Crandall, Andrew Owens, Noah Snavely, and Dan Huttenlocher. Discrete-continuous optimization for large-scale structure from motion. In IEEE Conf. Comput. Vis. Pattern Recog., pages 3001–3008, 2011. 3, 8", + "[12] Deeksha Dangwal, Vincent T Lee, Hyo Jin Kim, Tianwei Shen, Meghan Cowan, Rajvi Shah, Caroline Trippel, Brandon Reagen, Timothy Sherwood, Vasileios Balntas, et al. Mitigating reverse engineering attacks on local feature descriptors. In Brit. Mach. Vis. Conf., 2021. 2", + "[13] Haowen Deng, Tolga Birdal, and Slobodan Ilic. Ppf-foldnet: Unsupervised learning of rotation invariant 3d local descriptors. In Eur. Conf. Comput. Vis., pages 602-618, 2018. 3", + "[14] Haowen Deng, Tolga Birdal, and Slobodan Ilic. Ppfnet: Global context aware local features for robust 3d point matching. In IEEE Conf. Comput. Vis. Pattern Recog., pages 195-205, 2018. 3", + "[15] Alexey Dosovitskiy and Thomas Brox. Inverting visual representations with convolutional networks. In IEEE Conf. Comput. Vis. Pattern Recog., pages 4829-4837, 2016. 2", + "[16] Bertram Drost, Markus Ulrich, Nassir Navab, and Slobodan Ilic. Model globally, match locally: Efficient and robust" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "3d object recognition. In IEEE Conf. Comput. Vis. Pattern Recog., pages 998-1005, 2010. 3, 5", + "[17] Mihai Dusmanu, Ondrej Miksik, Johannes L Schonberger, and Marc Pollefeys. Cross-descriptor visual localization and mapping. In Int. Conf. Comput. Vis., pages 6058-6067, 2021. 2, 3", + "[18] Martin A Fischler and Robert C Bolles. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM, 24(6):381-395, 1981. 3, 7, 1", + "[19] Kexue Fu, Shaolei Liu, Xiaoyuan Luo, and Manning Wang. Robust point cloud registration framework based on deep graph matching. In IEEE Conf. Comput. Vis. Pattern Recog., pages 8893-8902, 2021. 3", + "[20] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In IEEE Conf. Comput. Vis. Pattern Recog., pages 3354-3361, 2012. 2", + "[21] Zan Gojcic, Caifa Zhou, Jan D Wegner, and Andreas Wieser. The perfect match: 3d point cloud matching with smoothed densities. In IEEE Conf. Comput. Vis. Pattern Recog., pages 5545-5554, 2019. 3", + "[22] Michal Havlena, Akihiko Torii, and Tomáš Pajdla. Efficient structure from motion by graph optimization. In Eur. Conf. Comput. Vis., pages 100-113, 2010. 2", + "[23] Dan Hendrycks and Kevin Gimpel. Gaussian error linear units (gelus). arXiv preprint arXiv:1606.08415, 2016. 6", + "[24] Shengyu Huang, Zan Gojcic, Mikhail Usvyatsov, Andreas Wieser, and Konrad Schindler. Predator: Registration of 3d point clouds with low overlap. In IEEE Conf. Comput. Vis. Pattern Recog., pages 4267-4276, 2021. 2, 3, 7, 8, 1, 4, 5", + "[25] Xiaoshui Huang, Guofeng Mei, and Jian Zhang. Feature-metric registration: A fast semi-supervised approach for robust point cloud registration without correspondences. In IEEE Conf. Comput. Vis. Pattern Recog., pages 11366-11374, 2020. 3", + "[26] Shengze Jin, Iro Armeni, Marc Pollefeys, and Daniel Barath. Multiway point cloud mosaicking with diffusion and global optimization. In IEEE Conf. Comput. Vis. Pattern Recog., pages 20838-20849, 2024. 3", + "[27] Neena Kamath. Announcing azure spatial anchors for collaborative, cross-platform mixed reality apps, 2019. URL https://azure.microsoft.com/en-us/blog/announcing-azure-spatial-anchors-for-collaborative-cross-platform-mixed-reality-apps/.1", + "[28] Alex Kendall, Matthew Grimes, and Roberto Cipolla. Posenet: A convolutional network for real-time 6-dof camera relocalization. In Int. Conf. Comput. Vis., pages 2938-2946, 2015. 3, 7, 8, 2, 6", + "[29] Hongdong Li and Richard Hartley. The 3d-3d registration problem revisited. In Int. Conf. Comput. Vis., pages 1-8, 2007. 3", + "[30] Zhengqi Li and Noah Snavely. Megadepth: Learning single-view depth prediction from internet photos. In IEEE Conf. Comput. Vis. Pattern Recog., pages 2041-2050, 2018. 3, 4, 5, 7" + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "6581", + "bbox": [ + 482, + 945, + 513, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[31] Philipp Lindenberger, Paul-Edouard Sarlin, and Marc Pollefeys. LightGlue: Local Feature Matching at Light Speed. In Int. Conf. Comput. Vis., 2023. 2, 6, 1", + "[32] Liu Liu, Hongdong Li, Haodong Yao, and Ruyi Zha. Pluckernet: Learn to register 3d line reconstructions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1842-1852, 2021. 4", + "[33] Alex Locher, Michal Havlena, and Luc Van Gool. Progressive structure from motion. In Eur. Conf. Comput. Vis., pages 20-35, 2018. 2", + "[34] Alexandre Lopes, Roberto Souza, and Helio Pedrini. A survey on rgb-d datasets. Comput. Vis. Img. Underst., 222: 103489, 2022. 2", + "[35] David G Lowe. Distinctive image features from scale-invariant keypoints. Int. J. Comput. Vis., 60:91-110, 2004. 2, 4, 5", + "[36] André Mateus, Siddhant Ranade, Srikumar Ramalingam, and Pedro Miraldo. Fast and accurate 3d registration from line intersection constraints. Int. J. Comput. Vis., pages 1-26, 2023. 3", + "[37] Pedro Miraldo, Surojit Saha, and Srikumar Ramalingam. Minimal solvers for mini-loop closures in 3d multi-scan alignment. In IEEE Conf. Comput. Vis. Pattern Recog., 2019. 3", + "[38] Francesco Pittaluga, Sanjeev J Koppal, Sing Bing Kang, and Sudipta N Sinha. Revealing scenes by inverting structure from motion reconstructions. In IEEE Conf. Comput. Vis. Pattern Recog., pages 145-154, 2019. 2", + "[39] Zheng Qin, Hao Yu, Changjian Wang, Yulan Guo, Yuxing Peng, and Kai Xu. Geometric transformer for fast and robust point cloud registration. In IEEE Conf. Comput. Vis. Pattern Recog., pages 11143-11152, 2022. 2, 3, 6, 7, 8", + "[40] Tilman Reinhardt. Using global localization to improve navigation, 2019. URL https://blog.research.google/2019/02/using-global-localization-to-improve.html.1", + "[41] Radu Bogdan Rusu, Nico Blodow, and Michael Beetz. Fast point feature histograms (fpfh) for 3d registration. In IEEE Int'l Conf. Robotics and Auto., pages 3212-3217, 2009. 3", + "[42] Paul-Edouard Sarlin, Cesar Cadena, Roland Siegwart, and Marcin Dymczyk. From coarse to fine: Robust hierarchical localization at large scale. In IEEE Conf. Comput. Vis. Pattern Recog., pages 12716–12725, 2019. 4", + "[43] Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. SuperGlue: Learning feature matching with graph neural networks. In IEEE Conf. Comput. Vis. Pattern Recog., 2020. 2, 6", + "[44] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In IEEE Conf. Comput. Vis. Pattern Recog., 2016. 2, 4", + "[45] Peter H. Schonemann. A generalized solution of the orthogonal procrustes problem. Psychometrika, 31(1):1-10, 1966. 3", + "[46] Jamie Shotton, Ben Glocker, Christopher Zach, Shahram Izadi, Antonio Criminisi, and Andrew Fitzgibbon. Scene coordinate regression forests for camera relocalization in rgb-d images. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2930-2937, 2013. 7" + ], + "bbox": [ + 91, + 90, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[47] Tory Smith. Introducing lightship vps, 2022. URL https://lightship.dev/blog/introducing-lightship-vps/.1", + "[48] Christoph Strecha, Timo Pylvänäinen, and Pascal Fua. Dynamic and scalable large scale image reconstruction. In IEEE Conf. Comput. Vis. Pattern Recog., pages 406-413, 2010. 3", + "[49] Yifan Sun, Changmao Cheng, Yuhan Zhang, Chi Zhang, Liang Zheng, Zhongdao Wang, and Yichen Wei. Circle loss: A unified perspective of pair similarity optimization. In IEEE Conf. Comput. Vis. Pattern Recog., pages 6398-6407, 2020. 6", + "[50] Chris Sweeney, Victor Fragoso, Tobias Höllerer, and Matthew Turk. Large scale sfm with the distributed camera model. In Int'l Conf. 3D Vision, pages 230-238, 2016. 2", + "[51] Yurun Tian, Xin Yu, Bin Fan, Fuchao Wu, Huub Heijnen, and Vassileios Balntas. Sosnet: Second order similarity regularization for local descriptor learning. In IEEE Conf. Comput. Vis. Pattern Recog., pages 11016-11025, 2019. 4, 5", + "[52] Haithem Turki, Deva Ramanan, and Mahadev Satyanarayanan. Mega-nerf: Scalable construction of large-scale nerfs for virtual fly-throughs. In IEEE Conf. Comput. Vis. Pattern Recog., pages 12922–12931, 2022. 8", + "[53] Michal Tyszkiewicz, Pascal Fua, and Eduard Trulls. Disk: Learning local features with policy gradient. Adv. Neural Inform. Process. Syst., 33:14254-14265, 2020. 1", + "[54] Ole Untzelmann, Torsten Sattler, Sven Middelberg, and Leif Kobbelt. A scalable collaborative online system for city reconstruction. In Proceedings of the IEEE International Conference on Computer Vision Workshops, pages 644-651, 2013. 3, 6", + "[55] Shuzhe Wang, Juho Kannala, and Daniel Barath. Dgc-gnn: Leveraging geometry and color cues for visual descriptor-free 2d-3d matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20881–20891, 2024. 2, 3", + "[56] Yue Wang and Justin M Solomon. Deep closest point: Learning representations for point cloud registration. In Int. Conf. Comput. Vis., pages 3523-3532, 2019. 3", + "[57] Yue Wang and Justin M Solomon. Prnet: Self-supervised learning for partial-to-partial registration. Adv. Neural Inform. Process. Syst., 32, 2019. 3", + "[58] Philippe Weinzaepfel, Hervé Jégou, and Patrick Pérez. Reconstructing an image from its local descriptors. In IEEE Conf. Comput. Vis. Pattern Recog., pages 337-344, 2011. 2", + "[59] Hao Xu, Shuaicheng Liu, Guangfu Wang, Guanghui Liu, and Bing Zeng. Omnet: Learning overlapping mask for partial-to-partial point cloud registration. In Int. Conf. Comput. Vis., pages 3132-3141, 2021. 3", + "[60] Jiaolong Yang, Hongdong Li, and Yunde Jia. Go-icp: Solving 3d registration efficiently and globally optimally. In Int. Conf. Comput. Vis., pages 1457–1464, 2013. 3", + "[61] Runzhao Yao, Shaoyi Du, Wenting Cui, Canhui Tang, and Chengwu Yang. Pare-net: Position-aware rotation-equivariant networks for robust point cloud registration. In ECCV, 2024. 7, 8, 2" + ], + "bbox": [ + 516, + 90, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "6582", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[62] Zi Jian Yew and Gim Hee Lee. RPM-net: Robust point matching using learned features. In IEEE Conf. Comput. Vis. Pattern Recog., pages 11824-11833, 2020. 3", + "[63] Zi Jian Yew and Gim Hee Lee. Regtr: End-to-end point cloud correspondences with transformers. In IEEE Conf. Comput. Vis. Pattern Recog., pages 6677-6686, 2022. 3", + "[64] Hao Yu, Zheng Qin, Ji Hou, Mahdi Saleh, Dongsheng Li, Benjamin Busam, and Slobodan Ilic. Rotation-invariant transformer for point cloud matching. In IEEE Conf. Comput. Vis. Pattern Recog., pages 5384-5393, 2023. 1, 2, 3, 5, 6, 7, 8, 4", + "[65] Andy Zeng, Shuran Song, Matthias Nießner, Matthew Fisher, Jianxiong Xiao, and Thomas Funkhouser. 3dmatch: Learning local geometric descriptors from rgb-d reconstructions. In IEEE Conf. Comput. Vis. Pattern Recog., pages 1802-1811, 2017. 2, 4, 7, 8, 1", + "[66] Yejun Zhang, Shuzhe Wang, and Juho Kannala. A2-gnn: Angle-annular gnn for visual descriptor-free camera relocalization. In International Conference on 3D Vision (3DV), 2025. arXiv preprint arXiv:2502.20036. 2", + "[67] Qunjie Zhou, Sérgio Agostinho, Aljosa Osep, and Laura Leal-Taixe. Is geometry enough for matching in visual localization? In Eur. Conf. Comput. Vis., pages 407-425, 2022. 2, 3", + "[68] Qian-Yi Zhou, Jaesik Park, and Vladlen Koltun. Open3D: A modern library for 3D data processing. arXiv:1801.09847, 2018. 6, 7, 1" + ], + "bbox": [ + 91, + 92, + 482, + 472 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "6583", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/ColabSfM_ Collaborative Structure-from-Motion by Point Cloud Registration/52ce366e-9afd-441c-9c7e-e1d8e8bedeb7_model.json b/2025/ColabSfM_ Collaborative Structure-from-Motion by Point Cloud Registration/52ce366e-9afd-441c-9c7e-e1d8e8bedeb7_model.json new file mode 100644 index 0000000000000000000000000000000000000000..78f476d24ac3a58998d341af2999d5c15636f8ad --- /dev/null +++ b/2025/ColabSfM_ Collaborative Structure-from-Motion by Point Cloud Registration/52ce366e-9afd-441c-9c7e-e1d8e8bedeb7_model.json @@ -0,0 +1,2059 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.108, + 0.131, + 0.892, + 0.154 + ], + "angle": 0, + "content": "ColabSfM: Collaborative Structure-from-Motion by Point Cloud Registration" + }, + { + "type": "text", + "bbox": [ + 0.294, + 0.186, + 0.704, + 0.205 + ], + "angle": 0, + "content": "Johan Edstedt\\*1 André Mateus2 Alberto Jaenal2" + }, + { + "type": "text", + "bbox": [ + 0.316, + 0.213, + 0.678, + 0.23 + ], + "angle": 0, + "content": "\\(^{1}\\)Linköping University \\(^{2}\\)Ericsson Research, Sweden" + }, + { + "type": "image", + "bbox": [ + 0.11, + 0.262, + 0.889, + 0.5 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.512, + 0.908, + 0.598 + ], + "angle": 0, + "content": "Figure 1. Our proposed registration paradigm for collaborative SfM reconstructions (ColabSfM). Given two input SfM reconstructions \\(\\mathcal{P},\\mathcal{Q}\\) of the same scene, the task is to estimate the relative similarity transform \\((s,R,t)\\) between them. Our first contribution is to address this as a point cloud registration problem, using only 3D SfM tracks. For this, we do not rely on the visual descriptors, but on the 3D coordinates of the points \\(\\mathbf{P},\\mathbf{Q}\\), their normals \\(\\mathbf{N},\\mathbf{M}\\) and, optionally, but not necessarily, features \\(\\mathbf{X},\\mathbf{Y}\\). To make point cloud registration methods perform well on this task, we propose as our second contribution a scalable pipeline to construct synthetic training datasets for SfM registration. Finally, we propose an improved version of RoITr [64] as registration method \\(f_{\\theta}\\)." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.604, + 0.329, + 0.62 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.631, + 0.486, + 0.875 + ], + "angle": 0, + "content": "Structure-from-Motion (SfM) is the task of estimating 3D structure and camera poses from images. We define Collaborative SfM (ColabSfM) as sharing distributed SfM reconstructions. Sharing maps requires estimating a joint reference frame, which is typically referred to as registration. However, there is a lack of scalable methods and training datasets for registering SfM reconstructions. In this paper, we tackle this challenge by proposing the scalable task of point cloud registration for SfM reconstructions. We find that current registration methods cannot register SfM point clouds when trained on existing datasets. To this end, we propose a SfM registration dataset generation pipeline, leveraging partial reconstructions from synthetically generated camera trajectories for each scene. Finally, we propose a simple but impactful neural refiner on top of the SotA registration method RoITr that yields significant improve" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.605, + 0.909, + 0.667 + ], + "angle": 0, + "content": "ments, which we call RefineRoITr. Our extensive experimental evaluation shows that our proposed pipeline and model enables ColabSfM. Code is available at https://github.com/EricssonResearch/ColabSfM" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.692, + 0.646, + 0.709 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.72, + 0.909, + 0.903 + ], + "angle": 0, + "content": "For robots and extended reality (XR) devices to become ubiquitous in our lives, they must be able to accurately build and localize in maps of the environment. Most available solutions rely on visual data and onboard computing, which can limit the accuracy in the context of small devices. Thus, in the last years there has been a push from industry to move computation to the cloud (e.g. Google's Visual Positioning System [40], Microsoft's Azure Spatial Anchors [27], and Niantic's Lightship [47]). However, the heterogeneous nature of these systems prevents devices of different vendors from interoperating, i.e., sharing their maps and localizing in others' maps. If we are to make visual localization in-" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.887, + 0.401, + 0.901 + ], + "angle": 0, + "content": "*Work done during an internship at Ericsson Research." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.958 + ], + "angle": 0, + "content": "6573" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.107, + 0.086, + 0.891, + 0.407 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.423, + 0.908, + 0.496 + ], + "angle": 0, + "content": "Figure 2. Qualitative comparison. We compare our approach to the previous point cloud registration method OverlapPredator [24] on the St. Peter's Basilica test scene. Without training on our proposed SfM registration dataset (column 1), previous methods are unable to produce sufficiently good matches (top row) and accurate relative pose estimation results (bottom row). In contrast, our proposed model RefineRoITr, trained on the proposed dataset, is able to find better matches and hence register the scenes well. The source and target point clouds are depicted in yellow and blue, respectively." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.52, + 0.482, + 0.686 + ], + "angle": 0, + "content": "teroperable, we need ways to merge such maps. Structurefrom-Motion (SfM) [44] is the most used method to process images and produce 3D maps (scene reconstructions consisting of a set of 3D points and the poses of the images expressed in a common coordinate system). To enable localization, a visual descriptor is stored for each map point, and for localization, 2D-3D matches are found by matching the descriptors of the query image with those of the map. Typically, SfM maps are merged using these visual descriptors, by finding 3D-3D matches [22, 33], or inserting new sub-maps based on 2D-3D correspondences [50]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.483, + 0.901 + ], + "angle": 0, + "content": "However, these methods have limitations. First, different SfM pipelines tend to use different feature extractors, which yield incompatible descriptors. This is partially addressed by the cross-descriptor approach proposed in [17], but the assumption that the descriptors are computed from the same keypoints is unrealistic. Additionally, exposing visual descriptors may allow feature inversion attacks [12, 15, 38, 58], which can recover private elements from the images. The second limitation concerns scalability, since storing visual descriptors leads to a map size increase between two and three orders of magnitude [67]; and matching using a traditional pipeline [35] may suffer in performance; which can be enhanced by using learned methods [31, 43] at the cost of excessive runtimes. Visual" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.52, + 0.907, + 0.657 + ], + "angle": 0, + "content": "localization without descriptors was tackled in [55, 66, 67], demonstrating competitive performance, but since they rely solely on 2D-3D geometry, which suffers from perspective distortions, their accuracy is not as high as their descriptor-based counterparts. Furthermore, they still rely on image retrieval, which again raises privacy, scalability, and interoperability concerns. Nevertheless, their results show that geometry alone could be sufficient for matching. Based on that observation, we pose the following question" + }, + { + "type": "title", + "bbox": [ + 0.549, + 0.666, + 0.885, + 0.681 + ], + "angle": 0, + "content": "Can we merge SfM maps using only 3D geometry?" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.907, + 0.795 + ], + "angle": 0, + "content": "To this end, we propose to merge SfM 3D maps through 3D point cloud registration methods [1, 65], which rely solely on geometric information, see Fig. 1. We call our proposed paradigm Collaborative SfM (ColabSfM). In contrast to visual localization, which aims to localize a single image to a map, our proposed task involves registering a map to a map, and is image retrieval-free." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.909, + 0.901 + ], + "angle": 0, + "content": "State of the art methods for 3D point cloud registration rely on learned features [24, 39, 64], therefore it is necessary to provide suitable datasets to train such networks. There is a wide range of datasets available [34] for the point cloud registration task. However, those mostly consist of 3D sensor data, e.g., RGB-D [65] and/or LiDAR scans [20], which differ from SfM point clouds. To the best of our knowl" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6574" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.287 + ], + "angle": 0, + "content": "edge, there are no large datasets for SfM point cloud registration. SfM datasets [30] provide a 3D point cloud for each scene, but since each scene is disconnected, no matching is possible between different scenes. To tackle this issue, we propose a pipeline to generate synthetic datasets for SfM point cloud registration, which generates multiple synthetic trajectories for each scene. The synthetic trajectories provide a trade-off between random internet images and sequential images (e.g., Cambridge Landmarks Dataset [28]). We apply our pipeline and generate training and benchmarks datasets from MegaDepth [30] and a benchmark dataset from Quad6k [11]. Our pipeline is presented in detail in Sec. 3." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.288, + 0.482, + 0.439 + ], + "angle": 0, + "content": "SfM reference frames may be arbitrarily aligned relative to each other. As such, using a registration model invariant to rotation and translation SE(3), and robust to scale Sim(3), is important. To this end, we build on top of the recently proposed SotA rotation invariant model RoITr [64]. While this method already produces good results when trained on our dataset, we find that we can further improve results on ColabSfM by introducing a neural refinement stage to the model. We introduce our baseline model and our improved version, which we call RefineRoITr, in Sec. 4." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.44, + 0.357, + 0.453 + ], + "angle": 0, + "content": "To summarize, our contributions are:" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.454, + 0.483, + 0.484 + ], + "angle": 0, + "content": "1. A simple but powerful pipeline to synthetically generate SfM registration datasets, detailed in Sec. 3." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.484, + 0.482, + 0.514 + ], + "angle": 0, + "content": "2. An improved model for SfM Registration, detailed in Sec. 4." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.514, + 0.482, + 0.545 + ], + "angle": 0, + "content": "3. The task of SfM point cloud registration, with extensive experiments in Sec. 5 demonstrating its potential." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.454, + 0.483, + 0.545 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.559, + 0.233, + 0.574 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.584, + 0.483, + 0.795 + ], + "angle": 0, + "content": "Point cloud registration is typically defined as estimating the rigid transform that aligns two 3D point clouds to a joint reference frame. The classic approach is the Iterative Closest Point (ICP) algorithm [3, 6], which estimates the pose by alternating between computing point correspondences using the current pose and estimating the pose using those correspondences. Several improvements have been proposed over the years to handle outliers and to attempt to find a global minimum [29, 60]. More recently, point cloud registration has been addressed with neural networks, such as DCP [56] and its iterative extension [57], with similar methods presented in [19, 62]. Another class of iterative methods [1, 25, 59], extract global features from each point cloud, and exploits them to regress the pose." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.483, + 0.902 + ], + "angle": 0, + "content": "An alternative approach, which will be our main focus, consists of finding and extracting correspondences between the two point clouds, being those points [37, 45], planes and lines [36], and finding the pose using robust estimators such as RANSAC [18]. Several point correspondence methods have been proposed, both handcrafted [16, 41] and learned [7, 14, 21]. In [24] an overlap-attention block is" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.244 + ], + "angle": 0, + "content": "proposed to improve matching of point clouds with low overlap. Overlap-aware methods are also presented in [26, 39, 63], which focus on improving the matching accuracy and thus avoid the use of RANSAC to filter outliers. The previous methods are not robust to large rotations, which are common in SfM. This issue has been tackled by proposing methods that are rotation invariant [4, 13, 64]. In this paper we will mainly consider point pair features (PPFs) [16] to ensure invariance to global rigid transformations, which have recently shown SotA performance [64]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.247, + 0.905, + 0.565 + ], + "angle": 0, + "content": "Localization without descriptors has recently received renewed attention as a way to reduce computation, and preserve privacy. BPnPNet [5] jointly estimates the pose and the correspondences between 2D keypoints on a query image and 3D points from a scene point cloud. However, it is susceptible to outlier correspondences. This issue was tackled by Zhou et al. [67], who use only 2D keypoints and the 3D map point cloud to localize a query image to a 3D map. They do this by projecting the map point cloud into retrieved map images and estimating 2D-2D correspondences with a graph neural network. A similar method DGC-GNN [55] proposes to exploit geometry and color information to guide point matching by computing a geometric embedding of Euclidean and angular relations. These methods have the advantage of not requiring storage of descriptors. However, they still require image retrieval to get sufficiently close map images to the query. This is problematic if the map is sparse in the region of the query, and introduces additional complexity. Furthermore, the projection from 3D into 2D loses important geometrical information, which we retain by directly registering the point clouds." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.569, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Collaborative mapping is a general term, typically referring to fusion of partial 3D reconstructions. Cohen et al. [8] use the Manhattan world assumption [10] to derive geometric constraints on partial reconstructions to enable model fusion. Cohen et al. [9] fuse indoor and outdoor reconstructions by first leveraging images to obtain semantic masks with window positions. They then match the windows between the two models to solve for the pose with RANSAC [18]. This approach requires images to be stored alongside the point clouds which brings additional space constraints [67]. Strecha et al. [48] and Untzelmann et al. [54] merge multiple building reconstructions to a joint map by combining GPS coordinates and known building outlines to construct an alignment optimization objective. This has the downside of restricting the reconstructions to buildings that have previously been mapped and requires geo-tags. Dusmanu et al. [17] investigated collaborative mapping and localization by translating descriptors to a joint embedding space. They consider a closed set of descriptors and learn a common embedding space for them. Collaborative mapping is achieved through embedding the descriptors into the joint space and running feature match-" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6575" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.304 + ], + "angle": 0, + "content": "ing on the joint space. This requires access to all features for all images, which is expensive, and the required compute scales as standard SfM with the number of images. In contrast to these methods, we make far fewer assumptions on the data available. Namely, we do not store any descriptors (neither local nor global), do not make assumptions on the structure of the scenes, and do not use any topological maps or data. Thus, we constrain ourselves to using only the geometry. Closest to our work Liu et al. [32] align line reconstructions and simulate partial reconstructions by adding noise and dropout of lines to a single reconstruction. In contrast we consider point-based reconstructions and propose a more realistic way to generate partial reconstructions through synthetic camera trajectories and retriangulation." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.315, + 0.484, + 0.348 + ], + "angle": 0, + "content": "3. Pipeline for SfM Registration Dataset Generation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.358, + 0.484, + 0.525 + ], + "angle": 0, + "content": "We find that current SotA point cloud registration models trained on 3DMatch [65], the most commonly used training set for 3D registration, are not able to register SfM reconstruction out of the box (cf. Tabs. 1, 2 and 4). Intuitively this is due to the large distribution shift between registration of RGB-D and/or LiDAR scans, and the much larger scenes covered by SfM reconstructions. However, it is not obvious how to generate a diverse dataset with multiple aligned reconstructions. Manually collecting such datasets is prohibitively expensive, and publicly available datasets typically contain a single reconstruction per scene." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.525, + 0.484, + 0.645 + ], + "angle": 0, + "content": "To tackle this issue, we propose a pipeline to generate a large synthetic dataset for registration between SfM models from random image collection datasets, e.g., MegaDepth [30]. Our key insight is that we can leverage these models and for each scene to construct partial reconstructions by sampling subsets of images/c cameras. An overview of our approach is presented in Fig. 3. In the next sections we describe our approach in more detail." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.653, + 0.382, + 0.669 + ], + "angle": 0, + "content": "3.1. Creating Partial Reconstructions" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.675, + 0.484, + 0.735 + ], + "angle": 0, + "content": "We use hloc [42] with COLMAP [44] to create partial reconstructions from scenes. We use SIFT [35] and SOSNet [51] retriangulations for training and evaluation. We consider two approaches to create partial reconstructions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.736, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Random point sampling: In the simplest approach, we sample a set of 3D points from the scene for each partial reconstruction, and add the images in which they are visible until reaching approximately 200 images. For each scene we create 10 such partial reconstructions. We found that while this approach works well when registering reconstructions from collections of random images, it does not generalize as well to reconstructions from a single camera trajectory, which are common in real-life scenarios, e.g., video. We believe that this is due to the point clouds exhibiting more significant variance when the scene is cap-" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.182 + ], + "angle": 0, + "content": "tured from different viewpoints. For example, keypoints are likely to be denser for nearby cameras, and natural occlusions will depend on viewpoints. As random image based datasets are not naturally divisible into videos, it is not obvious how to bridge this gap. We next describe our approach addressing this." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.183, + 0.907, + 0.426 + ], + "angle": 0, + "content": "Partial Trajectories: Our approach to bridge the random image \\(\\rightarrow\\) video gap is to generate synthetic trajectories from reconstructions of random images. Concretely, we randomly pick a starting image, and then sequentially (without replacement) pick the nearest neighbour of the image using a distance metric (which we refer to as dist). In practice we found that, for each trajectory, using a randomly weighted combination of geodesic rotation distance and Euclidean distance on the translation produced satisfactory results. Each trajectory has a size between \\(n_{\\mathrm{low}} = 75\\) and \\(n_{\\mathrm{high}} = 300\\). The algorithm is presented in Algorithm 1. We show a qualitative example of three generated synthetic trajectories in Fig. 4. Running this on all scenes gave us between 5 to 20 synthetic trajectories per scene. Training is conducted on an equally weighted combination of the random point sampling approach and the synthetic trajectories." + }, + { + "type": "code_caption", + "bbox": [ + 0.515, + 0.451, + 0.907, + 0.482 + ], + "angle": 0, + "content": "Algorithm 1 Generation of synthetic trajectory from a set of posed images" + }, + { + "type": "algorithm", + "bbox": [ + 0.516, + 0.486, + 0.906, + 0.743 + ], + "angle": 0, + "content": "Precondition: \\(\\mathbb{T}\\) set of posed images, \\(\\mathbb{I}\\) remaining index set, distance function dist \n1: function GENERATETRAJECTORY(T, I) \n2: \\(\\tau \\gets \\emptyset\\) \n3: i \\(\\leftarrow\\) Random sample from I \n4: n \\(\\leftarrow\\) Random sample from \\(\\{n_{\\mathrm{low}},\\dots ,n_{\\mathrm{high}}\\}\\) \n5: \\(\\mathbb{I}\\gets \\mathbb{I}\\backslash i\\) \n6: \\(\\tau \\gets \\tau \\cup i\\) \n7: for \\(k\\gets 1\\) to \\(n\\) do \n8: \\(i^{\\prime}\\gets \\arg \\min_{i^{\\prime}\\in \\mathbb{I}}\\mathrm{dist}(\\mathbb{T}_{i},\\mathbb{T}_{i^{\\prime}})\\) \n9: \\(\\tau \\gets \\tau \\cup i^{\\prime}\\) \n10: \\(\\mathbb{I}\\gets \\mathbb{I}\\backslash i^{\\prime}\\) \n11: \\(i\\gets i^{\\prime}\\) \n12: end for \n13: return \\(\\tau ,\\mathbb{I}\\) \n14: end function" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.773, + 0.71, + 0.789 + ], + "angle": 0, + "content": "3.2. Pipeline and Dataset" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Reconstruction Pipeline: We are interested in having good ground truth correspondences between the partial reconstructions. As the SfM pipeline accuracy usually degrades with fewer available cameras, we use fixed camera poses from the joint reconstruction using all images and simply run triangulation from feature matching. This also has the benefit of being significantly less computationally" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6576" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.174, + 0.088, + 0.825, + 0.281 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.289, + 0.908, + 0.334 + ], + "angle": 0, + "content": "Figure 3. Overview of our pipeline. For each scene, from a random image SfM dataset, e.g., MegaDepth [30], we retriangulate partial reconstructions using partial trajectories from the scene. Since these trajectories are in the global SfM reference frame, the relative transformation is simply the identity mapping." + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.354, + 0.473, + 0.479 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.489, + 0.487, + 0.589 + ], + "angle": 0, + "content": "Figure 4. Example of synthetic trajectories in our proposed dataset. We start with a large scale scene, consisting of hundreds of cameras (left). From this set of cameras we run Algorithm 1 until the remaining set of cameras is small. Plotting the trajectories shows that plausible camera motion is achieved by this procedure (right). Sampling in this way bridges the gap between random image collections and video-based trajectories." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.616, + 0.414, + 0.631 + ], + "angle": 0, + "content": "demanding than running the full reconstructions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.632, + 0.484, + 0.798 + ], + "angle": 0, + "content": "Dataset Construction: We run our pipeline on all 196 MegaDepth scenes where 10 of them are reserved for the test benchmark. We use either SIFT [35] or SOSNet [51] features for the retriangulation, however any local features could be used. We compute the overlap between all pairs of partial reconstructions. To train models we enforce the overlap to be larger than \\(30\\%\\). In total, after enforcing sufficient overlap, we are left with about 22000 pairs (\\(\\sim\\)20000 for training and \\(\\sim\\)2000 for testing), which constitutes our dataset. Details on how point cloud overlap is computed are presented in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.814, + 0.436, + 0.831 + ], + "angle": 0, + "content": "4. RoITr and RefineRoITr for ColabSfM" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.485, + 0.903 + ], + "angle": 0, + "content": "In this section, we discuss our baseline and our proposed improvements for ColabSfM. An overview of the proposed registration model \\( f_{\\theta} \\) is presented in Figure 5. For consistency we follow the notation of Yu et al. [64]." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.357, + 0.655, + 0.372 + ], + "angle": 0, + "content": "4.1. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.38, + 0.907, + 0.532 + ], + "angle": 0, + "content": "We are interested in registration methods that are robust to the challenges posed by SfM reconstructions. In particular, one of the main challenges is that SfM reconstructions are defined up-to-scale, and the rotation and origin of the reference frame is arbitrary1. Hence, we would like the registration model to be approximately invariant to the action of the similarity group, i.e., \\(\\mathrm{Sim}(3) = \\mathrm{SE}(3) + \\mathrm{scale}\\). In practice we will achieve this by using an SE(3) invariant model, and learn approximate scale invariance. We next describe our baseline, and then our proposed improvements." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.543, + 0.669, + 0.557 + ], + "angle": 0, + "content": "4.2. Baseline Model" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.566, + 0.907, + 0.641 + ], + "angle": 0, + "content": "As a strong baseline, we base our approach on the SotA SE(3) invariant registration method RoITr [64]. We next give an overview of our baseline. An more thorough description of the architecture is given in the supplementary and by Yu et al. [64]." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.642, + 0.909, + 0.763 + ], + "angle": 0, + "content": "Rotation Invariance through PPFs: RoITr [64] is based on point pair features (PPFs) [16], which computes a 4D feature for point pairs by the angles between their relative position and their normals. Mathematically, the PPF is defined as \\(\\mathrm{PPF}(p_1,p_2,n_1,n_2) =\\) \\([\\| d\\| ,\\angle (n_1,d),\\angle (n_2,d),\\angle (n_1,n_2)]\\) , where \\(d = p_{2} - p_{1}\\) It can easily be shown that this construction is invariant to rotations," + }, + { + "type": "equation", + "bbox": [ + 0.529, + 0.78, + 0.905, + 0.835 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\operatorname {P P F} \\left(R p _ {1}, R p _ {2}, R n _ {1}, R n _ {2}\\right) = [ \\| R d \\|, \\angle \\left(R n _ {1}, R d\\right), \\\\ \\angle (R n _ {2}, R d), \\angle (R n _ {1}, R n _ {2}) ] = [ \\| d \\|, \\angle (n _ {1}, d), \\angle (n _ {2}, d), \\\\ \\left. \\angle \\left(n _ {1}, n _ {2}\\right) \\right] = \\operatorname {P P F} \\left(p _ {1}, p _ {2}, n _ {1}, n _ {2}\\right), \\tag {1} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.848, + 0.838, + 0.862 + ], + "angle": 0, + "content": "since norms and angles are invariant to rotations." + }, + { + "type": "page_footnote", + "bbox": [ + 0.513, + 0.875, + 0.907, + 0.903 + ], + "angle": 0, + "content": "1Although in practice, approximate scale can be obtained by, e.g., standardization, objects with known size, or additional sensors." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6577" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.132, + 0.098, + 0.868, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.253, + 0.908, + 0.366 + ], + "angle": 0, + "content": "Figure 5. Overview of our proposed model RefineRoITr. As input we take two point clouds \\(\\mathcal{P} = (\\mathbf{P},\\mathbf{N},\\mathbf{X})\\) \\(\\mathcal{Q} = (\\mathbf{Q},\\mathbf{M},\\mathbf{Y})\\) consisting of 3D points \\(\\mathbf{P} / \\mathbf{Q}\\in \\mathbb{R}^{n / m\\times 3}\\), and normals \\(\\mathbf{N} / \\mathbf{M}\\in \\mathbb{R}^{n / m\\times 3}\\). In this work the features \\(\\mathbf{X} / \\mathbf{Y}\\) are always assumed to be constant. We ensure rotation invariance by Point Pair Feature (PPF) encoding. The PPFs are fed together with the features into an encoder \\(e_{\\theta}\\), producing \\(N^{\\prime}\\) coarse superpoints \\(\\mathcal{P}'/\\mathcal{Q}'\\). These are passed through a global Transformer \\(g_{\\theta}\\), from which the top-\\(k\\) coarse correspondences \\(\\mathcal{C}'\\in \\mathbb{R}^{k\\times 2}\\) are extracted. A decoder \\(d_{\\theta}\\) takes the coarse features from \\(g_{\\theta}\\) and the finer features from \\(e_{\\theta}\\) and produces fine features \\(\\hat{\\mathbf{X}},\\hat{\\mathbf{Y}}\\). Using the coarse correspondences we extract neighbourhoods \\(\\hat{\\mathbf{G}}^X,\\hat{\\mathbf{G}}^Y\\in \\mathbb{R}^{k\\times 64\\times c}\\), which we feed into our proposed refinement Transformer. The refined features from the Transformer are used to construct a cost matrix, where the Sinkhorn algorithm [43, 54, 64] is used to solve the optimal transport (OT) problem, producing our final matches \\(\\mathcal{C}\\). The network is trained using \\(\\mathcal{L}_s + \\mathcal{L}_p\\)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.391, + 0.482, + 0.542 + ], + "angle": 0, + "content": "Normalization Procedure: SfM reconstructions do not in general come with metric scale. This poses several challenges for registration methods, that typically assume both point clouds to share the same scale. We tackle this by normalization. Specifically, for Sim(3) training we normalize the point clouds \\(\\mathbf{P}\\), \\(\\mathbf{Q}\\) with \\(\\sigma_{\\mathbf{P}} / \\sqrt{2}\\), \\(\\sigma_{\\mathbf{Q}} / \\sqrt{2}\\), where \\(\\sigma\\) is the largest singular value of the centered point cloud. For SE(3) training, we simply use \\(\\sigma_{\\mathbf{P}}\\) to normalize both \\(\\mathbf{P}\\) and \\(\\mathbf{Q}\\). We find that this simple approach leads to models robust to scale variation, as we demonstrate in Tab. 1." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.546, + 0.483, + 0.727 + ], + "angle": 0, + "content": "Normal Estimation: We follow RoITr [64] and use Open3D [68] with a neighborhood size of 33 for estimating normals, we found that this works well in practice. In contrast to the RGB-D scans of 3DMatch, our dataset consists of point clouds from multiple sensors. This means that the simple approach of aligning the normal orientation towards the origin of the coordinate system, as done in RoITr, will not be consistent in general. To remedy this, we use the fact that all cameras observing a 3D track will yield consistent orientation estimates. In practice, we align the orientation towards the camera center of a random image observing the track, which we found yields consistent orientation." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.731, + 0.483, + 0.866 + ], + "angle": 0, + "content": "Overview of Architecture: RoITr consists of an encoder \\( e_{\\theta} \\), a global Transformer \\( g_{\\theta} \\), and a decoder \\( d_{\\theta} \\). \\( e_{\\theta} \\) and \\( d_{\\theta} \\) take as input sets of PPF encoded local neighbourhoods (typically around 8/16 points) around each point/superpoint. In contrast, \\( g_{\\theta} \\) performs global attention on the coarse point cloud \\( \\hat{\\mathcal{P}} \\). \\( e_{\\theta}, d_{\\theta}, g_{\\theta} \\) all use the Transformer architectures as described by Yu et al. [64]. Note that RoITr does not include the refining Transformer \\( r_{\\theta} \\), which will be detailed in our RefineRoITr." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.871, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Losses: RoITr uses a superpoint matching loss \\(\\mathcal{L}_s\\) and a point matching \\(\\mathcal{L}_p\\) loss for training. We follow RoITr" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.391, + 0.905, + 0.452 + ], + "angle": 0, + "content": "and use their overlap-aware [39] circle loss [49] as \\(\\mathcal{L}_s\\). The point matching loss \\(\\mathcal{L}_p\\) is the Negative Log Likelihood (NLL) of the GT correspondences after running the Sinkhorn algorithm [43] on the fine feature similarities \\(\\tilde{S}\\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.461, + 0.704, + 0.476 + ], + "angle": 0, + "content": "4.3. RefineRoITr (Ours)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.483, + 0.906, + 0.633 + ], + "angle": 0, + "content": "Refinement Transformer \\(r_{\\theta}\\): The refinement in RoITr is performed by optimizing the Sinkhorn [54] distance between shallow features in a local neighbourhood around the coarse matches, denoted \\(\\hat{\\mathbf{G}}^X, \\hat{\\mathbf{G}}^Y\\). While this provides satisfactory results on 3DMatch, we found that we could improve results by incorporating a joint local Transformer \\(r_{\\theta}\\). This Transformer takes \\(\\hat{\\mathbf{G}}^X, \\hat{\\mathbf{G}}^Y\\) as inputs and outputs enriched neighbourhoods \\(\\hat{\\mathbf{G}}^X, \\hat{\\mathbf{G}}^Y\\). We found that these cross-point cloud enriched fine features from \\(r_{\\theta}\\) yielded improved results." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.634, + 0.907, + 0.876 + ], + "angle": 0, + "content": "Architectural Details of \\( r_{\\theta} \\): \\( r_{\\theta} \\) takes in local neighbourhoods of shape \\( m \\times 64 \\times c \\), where \\( m \\) is the number of neighbourhoods, 64 is the number of neighbours (identical to RoITr), and \\( c \\) is the feature dimensionality. We use layers of interleaved self-attention and cross-attention as in LightGlue [31]. Note however, that in contrast to LightGlue that performs global attention, our refinement Transformer performs attention locally, which is significantly cheaper. We use an input/output dimensionality of \\( c_{\\mathrm{in}} = c_{\\mathrm{out}} = c = 64 \\), and 4 self+cross attention layers. Each self/cross attention layer consists of single head attention, using dim \\( c \\), which is stacked with the input into a 1 hidden layer MLP (input dim \\( 2c \\), hidden dim \\( 2c \\), output dim \\( c \\)) using the GELU [23] activation function, with layer normalization [2] applied before GELU. Each self/cross block is residual, i.e., the output of the block is the output of the MLP plus the input." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6578" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.483, + 0.187 + ], + "angle": 0, + "content": "Table 1. Results of SfM registration on MegaDepth. We evaluate 2 different scenarios. Unknown relative orientation SO(3) + unknown relative position = SE(3), SE(3) + unknown relative scale = Sim(3). The top portion contains methods only trained on the 3DMatch (3DM) dataset, the middle portion methods trained only on our proposed dataset (Mega), while the lower portion contains methods trained on the former and fine-tuned on the latter." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.198, + 0.481, + 0.337 + ], + "angle": 0, + "content": "
MethodSE(3)Sim(3)
IRFMRRRIRFMRRR
OverlapPredator [24] (3DM)6.135.510.03.621.32.1
GeoTransformer [39] (3DM)2.114.63.31.38.60.4
PareNet [61] (3DM)7.338.65.24.727.70.8
RoITr [64] (3DM)3.012.60.01.67.00.8
RefineRoITr (3DM)10.032.813.96.222.83.5
RefineRoITr (Mega)51.096.570.244.692.842.7
OverlapPredator [24] (3DM + Mega)21.374.935.411.456.210.7
RoITr [64] (3DM + Mega)44.690.960.038.486.138.7
RefineRoITr (3DM + Mega)48.795.167.744.693.944.3
" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.38, + 0.223, + 0.398 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.414, + 0.483, + 0.596 + ], + "angle": 0, + "content": "In this section, we evaluate the RefineRoITr model presented in Sec. 4. Three different model instances were trained: 1) pre-trained RefineRoITr on 3DMatch [65]; 2) the model in 1) fine-tuned on the dataset in Sec. 3 and; 3) RefineRoITr trained from scratch on our dataset. For comparison, we used RoITr [64] and OverlapPredator [24], both with pre-trained weights from 3DMatch, and fine-tuned weights trained on our dataset. Additionally, we tested GeoTransformer [39] also pre-trained on 3DMatch [65]. As we found that the performance, of fine-tuning 3DMatch trained models, on our dataset was similar for our proposed model, we did not train other models in our dataset from scratch." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.605, + 0.483, + 0.802 + ], + "angle": 0, + "content": "Metrics: We use the standard registration metric Inlier Ratio (IR). This measures the percentage of estimated correspondences lying within a threshold distance \\(\\tau_{\\mathrm{IR}} = 0.1\\) from each other. We additionally measure the Feature Matching Recall (FMR), which is defined as the percentage of point cloud pairs whose IR is above a threshold \\(\\tau_{\\mathrm{FMR}} = 5\\%\\). Finally, we use Registration Recall (RR), which consists of the percentage of pairs for which the rotation error and translation error are below \\(5^{\\circ}\\) and \\(0.05\\), respectively. We solve for the pose using the RANSAC [18] implementation from Open3D [68] on the matches retrieved by the neural networks. Details on error computation and RANSAC can be found in the Supplemental material." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.483, + 0.9 + ], + "angle": 0, + "content": "Computational cost: On a desktop machine equipped with an i9-13900K @ 3.00GHz CPU and an NVIDIA 4090 GPU, the median runtime for RoITr is 99.94 ms and for RefineRoITris 102.84, which shows that the computational burden added by the modifications described in Sec. 4.3 is around \\(3\\%\\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.091, + 0.871, + 0.108 + ], + "angle": 0, + "content": "5.1. MegaDepth SfM Registration Benchmark" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.114, + 0.905, + 0.266 + ], + "angle": 0, + "content": "We select 10 scenes from MegaDepth [30] dataset, which consists of 196 large scale 3D scenes reconstructed from internet images for our benchmark. The test scenes are, in order, [0008, 0015, 0019, 0021, 0022, 0024, 0025, 0032, 0063, 1589]. The average size of the point clouds is \\(15.3\\mathrm{K}\\) points, using either SIFT or SOSNet features for the triangulation. To train and evaluate the network presented in Sec. 4 on our dataset, we define two different scenarios SE(3) and Sim(3). We define a random rotation by sampling three Euler angles \\(\\alpha, \\beta, \\gamma \\in [0,2\\pi]\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.267, + 0.907, + 0.494 + ], + "angle": 0, + "content": "The results of the RefineRoITr method and the baselines are presented in Tab. 1. As expected, the methods which are not fine-tuned in our data perform the worst in both SE(3) and Sim(3) data. This shows the lack of generalization of current registration methods on SfM point clouds. When considering FMR, the methods trained from scratch or fine-tuned on our data obtain high scores (\\(>90\\%\\)) on SE(3) data with slightly lower scores on Sim(3). When looking at the three variations of RefineRoITr, we see that while the version trained on 3DMatch outperforms vanilla RoITr, the major boost in performance is due to fine-tuning the model on our dataset. The best overall version is the one trained from scratch on our dataset, even though the improvement is not as significant. Qualitative results are presented in Fig. 2 and the Supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.508, + 0.905, + 0.525 + ], + "angle": 0, + "content": "5.2. Standard Benchmark: Cambridge Landmarks" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.531, + 0.907, + 0.758 + ], + "angle": 0, + "content": "The Cambridge Landmarks [28] dataset consists of 5 large-scale 3D scenes recorded with a smartphone by a pedestrian in Cambridge, each of which consisting of a train and a test trajectories with camera poses. We retriangulate the train/test sequences independently, which results in two reconstructions for each scene. In this case, we evaluate the IR and number of matches. Here we only test the SE(3) case. Rotation and translation errors and qualitative results for this dataset are reported in the supplementary material. The results are presented in Tab. 2. Similar to the evaluation results in the previous section, the models trained only on 3DMatch [65] perform poorly, while the ones fine-tuned and or trained from scratch on our dataset present the highest IR and number of matches, with small differences in favor of the version trained from scratch." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.772, + 0.857, + 0.788 + ], + "angle": 0, + "content": "5.3. Noisy Indoor Reconstructions: 7-Scenes" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.905, + 0.903 + ], + "angle": 0, + "content": "The 7-Scenes [46] dataset is an indoor RGB-D dataset. This dataset is challenging for SfM, since textured regions lead to significantly more noisy point clouds, and less geometric features. We present results in Tab. 3. Encouragingly, we find that models trained using our pipeline perform well also on these indoor scenes, showing the generalizability of our approach." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6579" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.907, + 0.132 + ], + "angle": 0, + "content": "Table 2. Results of SfM registration on Cambridge Landmarks [28]. We evaluate unknown relative orientation \\( \\mathrm{SO}(3) + \\) unknown relative position \\( = \\mathrm{{SE}}\\left( 3\\right) \\) . The top portion contains methods only trained on the 3DMatch (3DM) dataset, the middle portion methods trained only on our proposed dataset (Mega), while the lower portion contains methods trained on the former and fine-tuned on the latter." + }, + { + "type": "table", + "bbox": [ + 0.212, + 0.142, + 0.784, + 0.272 + ], + "angle": 0, + "content": "
MethodGreat CourtKings CollegeOld HospitalShop FacadeSt Mary's Church
IRMatchesIRMatchesIRMatchesIRMatchesIRMatches
OverlapPredator [24] (3DM)1.43560.53720.03632.53540.2382
GeoTransformer [39] (3DM)0.03820.17850.04380.113680.0431
PureNet [61] (3DM)4.320001.62000020001.220006.72000
RoiTr [64] (3DM)0.02850.04470.03210.39591.5407
RefineRoiTr (3DM)0.24122.24080.05440.97372.8567
RefineRoiTr (Mega)70.9453857.6397731.5221441.6132581.85167
RoiTr [64] (3DM + Mega)52.1142639.6137721.960328.070064.52481
RefineRoiTr (3DM + Mega)69.4455154.0324628.0159139.1134277.85168
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.283, + 0.907, + 0.327 + ], + "angle": 0, + "content": "Table 3. Results of SfM registration on 7-Scenes. We evaluate unknown relative orientation \\( \\mathrm{SO}(3) + \\) unknown relative position \\( = \\mathrm{{SE}}\\left( 3\\right) \\) . The top portion contains methods only trained on the 3DMatch (3DM) dataset, the middle portion methods trained only on our proposed dataset (Mega), while the lower portion contains methods trained on the former and fine-tuned on the latter." + }, + { + "type": "table", + "bbox": [ + 0.151, + 0.336, + 0.846, + 0.432 + ], + "angle": 0, + "content": "
MethodChessFireHeadsOfficePumpkinRed KitchenStairs
IRMatchesIRMatchesIRMatchesIRMatchesIRMatchesIRMatchesIRMatches
RoITr [64] (3DM)0.03070.010040.016990.013220.03270.011630.0730
RefineRoITr (3DM)0.08420.022040.011040.010730.010160.017210.01043
RefineRoITr (Mega)23.4671738.8904350.6541968.0946114.9289327.980506.23230
RoITr [64] (3DM+Mega)20.4209830.5416536.4381250.047527.9127118.524389.11883
RefineRoITr (3DM+Mega)23.6563233.3833950.5599662.788878.9226826.866166.42863
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.453, + 0.483, + 0.537 + ], + "angle": 0, + "content": "Table 4. Results of SfM registration on Quad6k [11]. We evaluate unknown relative orientation \\(\\mathrm{SO}(3) +\\) unknown relative position \\(=\\) SE(3). The top portion contains methods only trained on the 3DMatch (3DM) dataset, the middle portion methods trained only on our proposed dataset (Mega), while the lower portion contains methods trained on the former and fine-tuned on the latter." + }, + { + "type": "table", + "bbox": [ + 0.14, + 0.548, + 0.432, + 0.66 + ], + "angle": 0, + "content": "
MethodIRFMRMatchesRR
OverlapPredator [24] (3DM)1.610.21210.0
GeoTransformer [39] (3DM)0.60.019360.0
PareNet [61] (3DM)1.610.420000.0
RoITr [64] (3DM)0.40.05740.0
RefineRoITr (3DM)1.76.16172.0
RefineRoITr (Mega)17.267.487822.4
RoITr [64] (3DM + Mega)14.263.349224.5
RefineRoITr (3DM + Mega)18.165.371516.3
" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.693, + 0.303, + 0.71 + ], + "angle": 0, + "content": "5.4. Low Overlap: Quad6k" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.484, + 0.901 + ], + "angle": 0, + "content": "The Quad6k dataset [11] consists of about 6500 images from the Arts Quad at Cornell University. To create partial reconstructions for testing, we first perform a full scene retriangulation using the refined poses provided by [52] and then follow the procedure in Sec. 3.1, generating a total of 24 trajectories, and 49 pairs. The results are presented in Tab. 4. Similar to the previous datasets, the performance of the methods has a substantial boost after training on the dataset presented in Sec. 3. GeoTransformer [39] trained only on 3DMatch [65] presents the most matches on average, but only a small fraction (0.6%) are inliers, and its FMR and RR are both zero. Our Quad6k benchmark is" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.455, + 0.907, + 0.517 + ], + "angle": 0, + "content": "more challenging and thus the performance is lower than in Tabs. 1 and 2, due to a low overlap of \\(16.5\\%\\), in comparison with the \\(61.2\\%\\) of MegaDepth. Qualitative results for this dataset are shown in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.53, + 0.634, + 0.545 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.555, + 0.907, + 0.827 + ], + "angle": 0, + "content": "We have proposed Collaborative Structure-from-Motion by Point Cloud Registration, ColabSfM, a challenging new task to enable scalable collaborative mapping and localization. By extensive experiments we showed that with our proposed dataset generation pipeline and SE(3) invariant model, this task is solvable, and showed that our model, while trained only on MegaDepth, is able to generalize to several challenging scenarios, such as noisy indoor reconstructions, and low-overlap registration. We believe our initial results will encourage further research on collaborative 3D reconstruction. Limitations: a) Our proposed model RefineRoITr is E(3) invariant like RoITr, and therefore can struggle with estimating correct matches in symmetric scenes. b) Our models are trained and evaluated on retriangulated partial reconstructions. In practice, issues such as drift occur, making global alignment more difficult, however we find that our model still performs well in this setting in Table 7." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.841, + 0.671, + 0.858 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.866, + 0.906, + 0.897 + ], + "angle": 0, + "content": "This work was supported by the strategic research environment ELLIIT, funded by the Swedish government." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6580" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.116, + 0.484, + 0.171 + ], + "angle": 0, + "content": "[1] Yasuhiro Aoki, Hunter Goforth, Rangaprasad Arun Srivatsan, and Simon Lucey. Pointnetlk: Robust & efficient point cloud registration using pointnet. In IEEE Conf. Comput. Vis. Pattern Recog., pages 7163-7172, 2019. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.172, + 0.484, + 0.212 + ], + "angle": 0, + "content": "[2] Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. arXiv preprint arXiv:1607.06450, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.215, + 0.483, + 0.255 + ], + "angle": 0, + "content": "[3] Paul J. Besl and Neil D. McKay. A method for registration of 3-d shapes. IEEE Trans. Pattern Anal. Mach. Intell., 14 (2):239-256, 1992. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.257, + 0.483, + 0.31 + ], + "angle": 0, + "content": "[4] Georg Bokman and Fredrik Kahl. A case for using rotation invariant features in state of the art feature matchers. In IEEE Conf. Comput. Vis. Pattern Recog., pages 5110-5119, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.312, + 0.483, + 0.367 + ], + "angle": 0, + "content": "[5] Dylan Campbell, Liu Liu, and Stephen Gould. Solving the blind perspective-n-point problem end-to-end with robust differentiable geometric optimization. In Eur. Conf. Comput. Vis., pages 244-261, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.368, + 0.483, + 0.41 + ], + "angle": 0, + "content": "[6] Yang Chen and Gerard Medioni. Object modeling by registration of multiple range images. In IEEE Int'l Conf. Robotics and Auto., pages 2724-2729, 1991. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.411, + 0.483, + 0.451 + ], + "angle": 0, + "content": "[7] Christopher Choy, Jaesik Park, and Vladlen Koltun. Fully convolutional geometric features. In Int. Conf. Comput. Vis., pages 8958-8966, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.453, + 0.483, + 0.494 + ], + "angle": 0, + "content": "[8] Andrea Cohen, Torsten Sattler, and Marc Pollefeys. Merging the unmatched: Stitching visually disconnected sfm models. In Int. Conf. Comput. Vis., pages 2129-2137, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.496, + 0.483, + 0.55 + ], + "angle": 0, + "content": "[9] Andrea Cohen, Johannes L Schonberger, Pablo Speciale, Torsten Sattler, Jan-Michael Frahm, and Marc Pollefeys. Indoor-outdoor 3d reconstruction alignment. In Eur. Conf. Comput. Vis., pages 285-300, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.551, + 0.483, + 0.605 + ], + "angle": 0, + "content": "[10] James Coughlan and Alan L Yuille. The manhattan world assumption: Regularities in scene statistics which enable bayesian inference. Adv. Neural Inform. Process. Syst., 13, 2000. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.607, + 0.483, + 0.662 + ], + "angle": 0, + "content": "[11] David Crandall, Andrew Owens, Noah Snavely, and Dan Huttenlocher. Discrete-continuous optimization for large-scale structure from motion. In IEEE Conf. Comput. Vis. Pattern Recog., pages 3001–3008, 2011. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.663, + 0.483, + 0.732 + ], + "angle": 0, + "content": "[12] Deeksha Dangwal, Vincent T Lee, Hyo Jin Kim, Tianwei Shen, Meghan Cowan, Rajvi Shah, Caroline Trippel, Brandon Reagen, Timothy Sherwood, Vasileios Balntas, et al. Mitigating reverse engineering attacks on local feature descriptors. In Brit. Mach. Vis. Conf., 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.733, + 0.483, + 0.774 + ], + "angle": 0, + "content": "[13] Haowen Deng, Tolga Birdal, and Slobodan Ilic. Ppf-foldnet: Unsupervised learning of rotation invariant 3d local descriptors. In Eur. Conf. Comput. Vis., pages 602-618, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.775, + 0.483, + 0.829 + ], + "angle": 0, + "content": "[14] Haowen Deng, Tolga Birdal, and Slobodan Ilic. Ppfnet: Global context aware local features for robust 3d point matching. In IEEE Conf. Comput. Vis. Pattern Recog., pages 195-205, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.831, + 0.483, + 0.873 + ], + "angle": 0, + "content": "[15] Alexey Dosovitskiy and Thomas Brox. Inverting visual representations with convolutional networks. In IEEE Conf. Comput. Vis. Pattern Recog., pages 4829-4837, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.874, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[16] Bertram Drost, Markus Ulrich, Nassir Navab, and Slobodan Ilic. Model globally, match locally: Efficient and robust" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.116, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "3d object recognition. In IEEE Conf. Comput. Vis. Pattern Recog., pages 998-1005, 2010. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.177 + ], + "angle": 0, + "content": "[17] Mihai Dusmanu, Ondrej Miksik, Johannes L Schonberger, and Marc Pollefeys. Cross-descriptor visual localization and mapping. In Int. Conf. Comput. Vis., pages 6058-6067, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.905, + 0.234 + ], + "angle": 0, + "content": "[18] Martin A Fischler and Robert C Bolles. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM, 24(6):381-395, 1981. 3, 7, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.235, + 0.905, + 0.291 + ], + "angle": 0, + "content": "[19] Kexue Fu, Shaolei Liu, Xiaoyuan Luo, and Manning Wang. Robust point cloud registration framework based on deep graph matching. In IEEE Conf. Comput. Vis. Pattern Recog., pages 8893-8902, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.293, + 0.905, + 0.347 + ], + "angle": 0, + "content": "[20] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In IEEE Conf. Comput. Vis. Pattern Recog., pages 3354-3361, 2012. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.349, + 0.905, + 0.403 + ], + "angle": 0, + "content": "[21] Zan Gojcic, Caifa Zhou, Jan D Wegner, and Andreas Wieser. The perfect match: 3d point cloud matching with smoothed densities. In IEEE Conf. Comput. Vis. Pattern Recog., pages 5545-5554, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.406, + 0.905, + 0.448 + ], + "angle": 0, + "content": "[22] Michal Havlena, Akihiko Torii, and Tomáš Pajdla. Efficient structure from motion by graph optimization. In Eur. Conf. Comput. Vis., pages 100-113, 2010. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.449, + 0.905, + 0.476 + ], + "angle": 0, + "content": "[23] Dan Hendrycks and Kevin Gimpel. Gaussian error linear units (gelus). arXiv preprint arXiv:1606.08415, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.478, + 0.905, + 0.533 + ], + "angle": 0, + "content": "[24] Shengyu Huang, Zan Gojcic, Mikhail Usvyatsov, Andreas Wieser, and Konrad Schindler. Predator: Registration of 3d point clouds with low overlap. In IEEE Conf. Comput. Vis. Pattern Recog., pages 4267-4276, 2021. 2, 3, 7, 8, 1, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.535, + 0.905, + 0.603 + ], + "angle": 0, + "content": "[25] Xiaoshui Huang, Guofeng Mei, and Jian Zhang. Feature-metric registration: A fast semi-supervised approach for robust point cloud registration without correspondences. In IEEE Conf. Comput. Vis. Pattern Recog., pages 11366-11374, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.605, + 0.905, + 0.66 + ], + "angle": 0, + "content": "[26] Shengze Jin, Iro Armeni, Marc Pollefeys, and Daniel Barath. Multiway point cloud mosaicking with diffusion and global optimization. In IEEE Conf. Comput. Vis. Pattern Recog., pages 20838-20849, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.661, + 0.905, + 0.745 + ], + "angle": 0, + "content": "[27] Neena Kamath. Announcing azure spatial anchors for collaborative, cross-platform mixed reality apps, 2019. URL https://azure.microsoft.com/en-us/blog/announcing-azure-spatial-anchors-for-collaborative-cross-platform-mixed-reality-apps/.1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.746, + 0.905, + 0.801 + ], + "angle": 0, + "content": "[28] Alex Kendall, Matthew Grimes, and Roberto Cipolla. Posenet: A convolutional network for real-time 6-dof camera relocalization. In Int. Conf. Comput. Vis., pages 2938-2946, 2015. 3, 7, 8, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.802, + 0.905, + 0.843 + ], + "angle": 0, + "content": "[29] Hongdong Li and Richard Hartley. The 3d-3d registration problem revisited. In Int. Conf. Comput. Vis., pages 1-8, 2007. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.845, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[30] Zhengqi Li and Noah Snavely. Megadepth: Learning single-view depth prediction from internet photos. In IEEE Conf. Comput. Vis. Pattern Recog., pages 2041-2050, 2018. 3, 4, 5, 7" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.514, + 0.957 + ], + "angle": 0, + "content": "6581" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.134 + ], + "angle": 0, + "content": "[31] Philipp Lindenberger, Paul-Edouard Sarlin, and Marc Pollefeys. LightGlue: Local Feature Matching at Light Speed. In Int. Conf. Comput. Vis., 2023. 2, 6, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.134, + 0.483, + 0.19 + ], + "angle": 0, + "content": "[32] Liu Liu, Hongdong Li, Haodong Yao, and Ruyi Zha. Pluckernet: Learn to register 3d line reconstructions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1842-1852, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.19, + 0.482, + 0.232 + ], + "angle": 0, + "content": "[33] Alex Locher, Michal Havlena, and Luc Van Gool. Progressive structure from motion. In Eur. Conf. Comput. Vis., pages 20-35, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.232, + 0.482, + 0.273 + ], + "angle": 0, + "content": "[34] Alexandre Lopes, Roberto Souza, and Helio Pedrini. A survey on rgb-d datasets. Comput. Vis. Img. Underst., 222: 103489, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.274, + 0.482, + 0.315 + ], + "angle": 0, + "content": "[35] David G Lowe. Distinctive image features from scale-invariant keypoints. Int. J. Comput. Vis., 60:91-110, 2004. 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.316, + 0.482, + 0.37 + ], + "angle": 0, + "content": "[36] André Mateus, Siddhant Ranade, Srikumar Ramalingam, and Pedro Miraldo. Fast and accurate 3d registration from line intersection constraints. Int. J. Comput. Vis., pages 1-26, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.371, + 0.482, + 0.425 + ], + "angle": 0, + "content": "[37] Pedro Miraldo, Surojit Saha, and Srikumar Ramalingam. Minimal solvers for mini-loop closures in 3d multi-scan alignment. In IEEE Conf. Comput. Vis. Pattern Recog., 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.427, + 0.482, + 0.482 + ], + "angle": 0, + "content": "[38] Francesco Pittaluga, Sanjeev J Koppal, Sing Bing Kang, and Sudipta N Sinha. Revealing scenes by inverting structure from motion reconstructions. In IEEE Conf. Comput. Vis. Pattern Recog., pages 145-154, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.484, + 0.482, + 0.538 + ], + "angle": 0, + "content": "[39] Zheng Qin, Hao Yu, Changjian Wang, Yulan Guo, Yuxing Peng, and Kai Xu. Geometric transformer for fast and robust point cloud registration. In IEEE Conf. Comput. Vis. Pattern Recog., pages 11143-11152, 2022. 2, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.539, + 0.482, + 0.594 + ], + "angle": 0, + "content": "[40] Tilman Reinhardt. Using global localization to improve navigation, 2019. URL https://blog.research.google/2019/02/using-global-localization-to-improve.html.1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.595, + 0.482, + 0.637 + ], + "angle": 0, + "content": "[41] Radu Bogdan Rusu, Nico Blodow, and Michael Beetz. Fast point feature histograms (fpfh) for 3d registration. In IEEE Int'l Conf. Robotics and Auto., pages 3212-3217, 2009. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.637, + 0.482, + 0.692 + ], + "angle": 0, + "content": "[42] Paul-Edouard Sarlin, Cesar Cadena, Roland Siegwart, and Marcin Dymczyk. From coarse to fine: Robust hierarchical localization at large scale. In IEEE Conf. Comput. Vis. Pattern Recog., pages 12716–12725, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.692, + 0.482, + 0.748 + ], + "angle": 0, + "content": "[43] Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. SuperGlue: Learning feature matching with graph neural networks. In IEEE Conf. Comput. Vis. Pattern Recog., 2020. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.748, + 0.482, + 0.789 + ], + "angle": 0, + "content": "[44] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In IEEE Conf. Comput. Vis. Pattern Recog., 2016. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.79, + 0.482, + 0.83 + ], + "angle": 0, + "content": "[45] Peter H. Schonemann. A generalized solution of the orthogonal procrustes problem. Psychometrika, 31(1):1-10, 1966. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[46] Jamie Shotton, Ben Glocker, Christopher Zach, Shahram Izadi, Antonio Criminisi, and Andrew Fitzgibbon. Scene coordinate regression forests for camera relocalization in rgb-d images. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2930-2937, 2013. 7" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.135 + ], + "angle": 0, + "content": "[47] Tory Smith. Introducing lightship vps, 2022. URL https://lightship.dev/blog/introducing-lightship-vps/.1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.178 + ], + "angle": 0, + "content": "[48] Christoph Strecha, Timo Pylvänäinen, and Pascal Fua. Dynamic and scalable large scale image reconstruction. In IEEE Conf. Comput. Vis. Pattern Recog., pages 406-413, 2010. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.18, + 0.905, + 0.248 + ], + "angle": 0, + "content": "[49] Yifan Sun, Changmao Cheng, Yuhan Zhang, Chi Zhang, Liang Zheng, Zhongdao Wang, and Yichen Wei. Circle loss: A unified perspective of pair similarity optimization. In IEEE Conf. Comput. Vis. Pattern Recog., pages 6398-6407, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.251, + 0.905, + 0.306 + ], + "angle": 0, + "content": "[50] Chris Sweeney, Victor Fragoso, Tobias Höllerer, and Matthew Turk. Large scale sfm with the distributed camera model. In Int'l Conf. 3D Vision, pages 230-238, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.309, + 0.905, + 0.366 + ], + "angle": 0, + "content": "[51] Yurun Tian, Xin Yu, Bin Fan, Fuchao Wu, Huub Heijnen, and Vassileios Balntas. Sosnet: Second order similarity regularization for local descriptor learning. In IEEE Conf. Comput. Vis. Pattern Recog., pages 11016-11025, 2019. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.367, + 0.905, + 0.422 + ], + "angle": 0, + "content": "[52] Haithem Turki, Deva Ramanan, and Mahadev Satyanarayanan. Mega-nerf: Scalable construction of large-scale nerfs for virtual fly-throughs. In IEEE Conf. Comput. Vis. Pattern Recog., pages 12922–12931, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.424, + 0.905, + 0.467 + ], + "angle": 0, + "content": "[53] Michal Tyszkiewicz, Pascal Fua, and Eduard Trulls. Disk: Learning local features with policy gradient. Adv. Neural Inform. Process. Syst., 33:14254-14265, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.469, + 0.905, + 0.537 + ], + "angle": 0, + "content": "[54] Ole Untzelmann, Torsten Sattler, Sven Middelberg, and Leif Kobbelt. A scalable collaborative online system for city reconstruction. In Proceedings of the IEEE International Conference on Computer Vision Workshops, pages 644-651, 2013. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.54, + 0.905, + 0.609 + ], + "angle": 0, + "content": "[55] Shuzhe Wang, Juho Kannala, and Daniel Barath. Dgc-gnn: Leveraging geometry and color cues for visual descriptor-free 2d-3d matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20881–20891, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.612, + 0.905, + 0.654 + ], + "angle": 0, + "content": "[56] Yue Wang and Justin M Solomon. Deep closest point: Learning representations for point cloud registration. In Int. Conf. Comput. Vis., pages 3523-3532, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.655, + 0.905, + 0.698 + ], + "angle": 0, + "content": "[57] Yue Wang and Justin M Solomon. Prnet: Self-supervised learning for partial-to-partial registration. Adv. Neural Inform. Process. Syst., 32, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.7, + 0.905, + 0.742 + ], + "angle": 0, + "content": "[58] Philippe Weinzaepfel, Hervé Jégou, and Patrick Pérez. Reconstructing an image from its local descriptors. In IEEE Conf. Comput. Vis. Pattern Recog., pages 337-344, 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.743, + 0.905, + 0.799 + ], + "angle": 0, + "content": "[59] Hao Xu, Shuaicheng Liu, Guangfu Wang, Guanghui Liu, and Bing Zeng. Omnet: Learning overlapping mask for partial-to-partial point cloud registration. In Int. Conf. Comput. Vis., pages 3132-3141, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.801, + 0.905, + 0.843 + ], + "angle": 0, + "content": "[60] Jiaolong Yang, Hongdong Li, and Yunde Jia. Go-icp: Solving 3d registration efficiently and globally optimally. In Int. Conf. Comput. Vis., pages 1457–1464, 2013. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.845, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[61] Runzhao Yao, Shaoyi Du, Wenting Cui, Canhui Tang, and Chengwu Yang. Pare-net: Position-aware rotation-equivariant networks for robust point cloud registration. In ECCV, 2024. 7, 8, 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6582" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.134 + ], + "angle": 0, + "content": "[62] Zi Jian Yew and Gim Hee Lee. RPM-net: Robust point matching using learned features. In IEEE Conf. Comput. Vis. Pattern Recog., pages 11824-11833, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.483, + 0.177 + ], + "angle": 0, + "content": "[63] Zi Jian Yew and Gim Hee Lee. Regtr: End-to-end point cloud correspondences with transformers. In IEEE Conf. Comput. Vis. Pattern Recog., pages 6677-6686, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.179, + 0.482, + 0.247 + ], + "angle": 0, + "content": "[64] Hao Yu, Zheng Qin, Ji Hou, Mahdi Saleh, Dongsheng Li, Benjamin Busam, and Slobodan Ilic. Rotation-invariant transformer for point cloud matching. In IEEE Conf. Comput. Vis. Pattern Recog., pages 5384-5393, 2023. 1, 2, 3, 5, 6, 7, 8, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.249, + 0.482, + 0.317 + ], + "angle": 0, + "content": "[65] Andy Zeng, Shuran Song, Matthias Nießner, Matthew Fisher, Jianxiong Xiao, and Thomas Funkhouser. 3dmatch: Learning local geometric descriptors from rgb-d reconstructions. In IEEE Conf. Comput. Vis. Pattern Recog., pages 1802-1811, 2017. 2, 4, 7, 8, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.32, + 0.482, + 0.375 + ], + "angle": 0, + "content": "[66] Yejun Zhang, Shuzhe Wang, and Juho Kannala. A2-gnn: Angle-annular gnn for visual descriptor-free camera relocalization. In International Conference on 3D Vision (3DV), 2025. arXiv preprint arXiv:2502.20036. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.376, + 0.482, + 0.431 + ], + "angle": 0, + "content": "[67] Qunjie Zhou, Sérgio Agostinho, Aljosa Osep, and Laura Leal-Taixe. Is geometry enough for matching in visual localization? In Eur. Conf. Comput. Vis., pages 407-425, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.433, + 0.482, + 0.473 + ], + "angle": 0, + "content": "[68] Qian-Yi Zhou, Jaesik Park, and Vladlen Koltun. Open3D: A modern library for 3D data processing. arXiv:1801.09847, 2018. 6, 7, 1" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.473 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.515, + 0.956 + ], + "angle": 0, + "content": "6583" + } + ] +] \ No newline at end of file diff --git a/2025/ColabSfM_ Collaborative Structure-from-Motion by Point Cloud Registration/52ce366e-9afd-441c-9c7e-e1d8e8bedeb7_origin.pdf b/2025/ColabSfM_ Collaborative Structure-from-Motion by Point Cloud Registration/52ce366e-9afd-441c-9c7e-e1d8e8bedeb7_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..8315c553741c409434fcbde5ce197aff777f2903 --- /dev/null +++ b/2025/ColabSfM_ Collaborative Structure-from-Motion by Point Cloud Registration/52ce366e-9afd-441c-9c7e-e1d8e8bedeb7_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:893a0c2ed375bfb88ce49c1e88851f5809dd0ed9d9da664897ed28d570d1e432 +size 7378016 diff --git a/2025/ColabSfM_ Collaborative Structure-from-Motion by Point Cloud Registration/full.md b/2025/ColabSfM_ Collaborative Structure-from-Motion by Point Cloud Registration/full.md new file mode 100644 index 0000000000000000000000000000000000000000..e7021dc66464d47bfc8d8e05e66eedccd0cf3385 --- /dev/null +++ b/2025/ColabSfM_ Collaborative Structure-from-Motion by Point Cloud Registration/full.md @@ -0,0 +1,271 @@ +# ColabSfM: Collaborative Structure-from-Motion by Point Cloud Registration + +Johan Edstedt\*1 André Mateus2 Alberto Jaenal2 + +$^{1}$ Linköping University $^{2}$ Ericsson Research, Sweden + +![](images/63d8556a7d3a7ee377336bc1d92af36db15e3b370e54ae4c18f88e5a2d2a60ec.jpg) +Figure 1. Our proposed registration paradigm for collaborative SfM reconstructions (ColabSfM). Given two input SfM reconstructions $\mathcal{P},\mathcal{Q}$ of the same scene, the task is to estimate the relative similarity transform $(s,R,t)$ between them. Our first contribution is to address this as a point cloud registration problem, using only 3D SfM tracks. For this, we do not rely on the visual descriptors, but on the 3D coordinates of the points $\mathbf{P},\mathbf{Q}$ , their normals $\mathbf{N},\mathbf{M}$ and, optionally, but not necessarily, features $\mathbf{X},\mathbf{Y}$ . To make point cloud registration methods perform well on this task, we propose as our second contribution a scalable pipeline to construct synthetic training datasets for SfM registration. Finally, we propose an improved version of RoITr [64] as registration method $f_{\theta}$ . + +# Abstract + +Structure-from-Motion (SfM) is the task of estimating 3D structure and camera poses from images. We define Collaborative SfM (ColabSfM) as sharing distributed SfM reconstructions. Sharing maps requires estimating a joint reference frame, which is typically referred to as registration. However, there is a lack of scalable methods and training datasets for registering SfM reconstructions. In this paper, we tackle this challenge by proposing the scalable task of point cloud registration for SfM reconstructions. We find that current registration methods cannot register SfM point clouds when trained on existing datasets. To this end, we propose a SfM registration dataset generation pipeline, leveraging partial reconstructions from synthetically generated camera trajectories for each scene. Finally, we propose a simple but impactful neural refiner on top of the SotA registration method RoITr that yields significant improve + +ments, which we call RefineRoITr. Our extensive experimental evaluation shows that our proposed pipeline and model enables ColabSfM. Code is available at https://github.com/EricssonResearch/ColabSfM + +# 1. Introduction + +For robots and extended reality (XR) devices to become ubiquitous in our lives, they must be able to accurately build and localize in maps of the environment. Most available solutions rely on visual data and onboard computing, which can limit the accuracy in the context of small devices. Thus, in the last years there has been a push from industry to move computation to the cloud (e.g. Google's Visual Positioning System [40], Microsoft's Azure Spatial Anchors [27], and Niantic's Lightship [47]). However, the heterogeneous nature of these systems prevents devices of different vendors from interoperating, i.e., sharing their maps and localizing in others' maps. If we are to make visual localization in- + +![](images/70d2831a980b6a8bb1a51e38bb366823f1fde1599554af37c6432bab8b1c5471.jpg) +Figure 2. Qualitative comparison. We compare our approach to the previous point cloud registration method OverlapPredator [24] on the St. Peter's Basilica test scene. Without training on our proposed SfM registration dataset (column 1), previous methods are unable to produce sufficiently good matches (top row) and accurate relative pose estimation results (bottom row). In contrast, our proposed model RefineRoITr, trained on the proposed dataset, is able to find better matches and hence register the scenes well. The source and target point clouds are depicted in yellow and blue, respectively. + +teroperable, we need ways to merge such maps. Structurefrom-Motion (SfM) [44] is the most used method to process images and produce 3D maps (scene reconstructions consisting of a set of 3D points and the poses of the images expressed in a common coordinate system). To enable localization, a visual descriptor is stored for each map point, and for localization, 2D-3D matches are found by matching the descriptors of the query image with those of the map. Typically, SfM maps are merged using these visual descriptors, by finding 3D-3D matches [22, 33], or inserting new sub-maps based on 2D-3D correspondences [50]. + +However, these methods have limitations. First, different SfM pipelines tend to use different feature extractors, which yield incompatible descriptors. This is partially addressed by the cross-descriptor approach proposed in [17], but the assumption that the descriptors are computed from the same keypoints is unrealistic. Additionally, exposing visual descriptors may allow feature inversion attacks [12, 15, 38, 58], which can recover private elements from the images. The second limitation concerns scalability, since storing visual descriptors leads to a map size increase between two and three orders of magnitude [67]; and matching using a traditional pipeline [35] may suffer in performance; which can be enhanced by using learned methods [31, 43] at the cost of excessive runtimes. Visual + +localization without descriptors was tackled in [55, 66, 67], demonstrating competitive performance, but since they rely solely on 2D-3D geometry, which suffers from perspective distortions, their accuracy is not as high as their descriptor-based counterparts. Furthermore, they still rely on image retrieval, which again raises privacy, scalability, and interoperability concerns. Nevertheless, their results show that geometry alone could be sufficient for matching. Based on that observation, we pose the following question + +# Can we merge SfM maps using only 3D geometry? + +To this end, we propose to merge SfM 3D maps through 3D point cloud registration methods [1, 65], which rely solely on geometric information, see Fig. 1. We call our proposed paradigm Collaborative SfM (ColabSfM). In contrast to visual localization, which aims to localize a single image to a map, our proposed task involves registering a map to a map, and is image retrieval-free. + +State of the art methods for 3D point cloud registration rely on learned features [24, 39, 64], therefore it is necessary to provide suitable datasets to train such networks. There is a wide range of datasets available [34] for the point cloud registration task. However, those mostly consist of 3D sensor data, e.g., RGB-D [65] and/or LiDAR scans [20], which differ from SfM point clouds. To the best of our knowl + +edge, there are no large datasets for SfM point cloud registration. SfM datasets [30] provide a 3D point cloud for each scene, but since each scene is disconnected, no matching is possible between different scenes. To tackle this issue, we propose a pipeline to generate synthetic datasets for SfM point cloud registration, which generates multiple synthetic trajectories for each scene. The synthetic trajectories provide a trade-off between random internet images and sequential images (e.g., Cambridge Landmarks Dataset [28]). We apply our pipeline and generate training and benchmarks datasets from MegaDepth [30] and a benchmark dataset from Quad6k [11]. Our pipeline is presented in detail in Sec. 3. + +SfM reference frames may be arbitrarily aligned relative to each other. As such, using a registration model invariant to rotation and translation SE(3), and robust to scale Sim(3), is important. To this end, we build on top of the recently proposed SotA rotation invariant model RoITr [64]. While this method already produces good results when trained on our dataset, we find that we can further improve results on ColabSfM by introducing a neural refinement stage to the model. We introduce our baseline model and our improved version, which we call RefineRoITr, in Sec. 4. + +To summarize, our contributions are: + +1. A simple but powerful pipeline to synthetically generate SfM registration datasets, detailed in Sec. 3. +2. An improved model for SfM Registration, detailed in Sec. 4. +3. The task of SfM point cloud registration, with extensive experiments in Sec. 5 demonstrating its potential. + +# 2. Related Work + +Point cloud registration is typically defined as estimating the rigid transform that aligns two 3D point clouds to a joint reference frame. The classic approach is the Iterative Closest Point (ICP) algorithm [3, 6], which estimates the pose by alternating between computing point correspondences using the current pose and estimating the pose using those correspondences. Several improvements have been proposed over the years to handle outliers and to attempt to find a global minimum [29, 60]. More recently, point cloud registration has been addressed with neural networks, such as DCP [56] and its iterative extension [57], with similar methods presented in [19, 62]. Another class of iterative methods [1, 25, 59], extract global features from each point cloud, and exploits them to regress the pose. + +An alternative approach, which will be our main focus, consists of finding and extracting correspondences between the two point clouds, being those points [37, 45], planes and lines [36], and finding the pose using robust estimators such as RANSAC [18]. Several point correspondence methods have been proposed, both handcrafted [16, 41] and learned [7, 14, 21]. In [24] an overlap-attention block is + +proposed to improve matching of point clouds with low overlap. Overlap-aware methods are also presented in [26, 39, 63], which focus on improving the matching accuracy and thus avoid the use of RANSAC to filter outliers. The previous methods are not robust to large rotations, which are common in SfM. This issue has been tackled by proposing methods that are rotation invariant [4, 13, 64]. In this paper we will mainly consider point pair features (PPFs) [16] to ensure invariance to global rigid transformations, which have recently shown SotA performance [64]. + +Localization without descriptors has recently received renewed attention as a way to reduce computation, and preserve privacy. BPnPNet [5] jointly estimates the pose and the correspondences between 2D keypoints on a query image and 3D points from a scene point cloud. However, it is susceptible to outlier correspondences. This issue was tackled by Zhou et al. [67], who use only 2D keypoints and the 3D map point cloud to localize a query image to a 3D map. They do this by projecting the map point cloud into retrieved map images and estimating 2D-2D correspondences with a graph neural network. A similar method DGC-GNN [55] proposes to exploit geometry and color information to guide point matching by computing a geometric embedding of Euclidean and angular relations. These methods have the advantage of not requiring storage of descriptors. However, they still require image retrieval to get sufficiently close map images to the query. This is problematic if the map is sparse in the region of the query, and introduces additional complexity. Furthermore, the projection from 3D into 2D loses important geometrical information, which we retain by directly registering the point clouds. + +Collaborative mapping is a general term, typically referring to fusion of partial 3D reconstructions. Cohen et al. [8] use the Manhattan world assumption [10] to derive geometric constraints on partial reconstructions to enable model fusion. Cohen et al. [9] fuse indoor and outdoor reconstructions by first leveraging images to obtain semantic masks with window positions. They then match the windows between the two models to solve for the pose with RANSAC [18]. This approach requires images to be stored alongside the point clouds which brings additional space constraints [67]. Strecha et al. [48] and Untzelmann et al. [54] merge multiple building reconstructions to a joint map by combining GPS coordinates and known building outlines to construct an alignment optimization objective. This has the downside of restricting the reconstructions to buildings that have previously been mapped and requires geo-tags. Dusmanu et al. [17] investigated collaborative mapping and localization by translating descriptors to a joint embedding space. They consider a closed set of descriptors and learn a common embedding space for them. Collaborative mapping is achieved through embedding the descriptors into the joint space and running feature match- + +ing on the joint space. This requires access to all features for all images, which is expensive, and the required compute scales as standard SfM with the number of images. In contrast to these methods, we make far fewer assumptions on the data available. Namely, we do not store any descriptors (neither local nor global), do not make assumptions on the structure of the scenes, and do not use any topological maps or data. Thus, we constrain ourselves to using only the geometry. Closest to our work Liu et al. [32] align line reconstructions and simulate partial reconstructions by adding noise and dropout of lines to a single reconstruction. In contrast we consider point-based reconstructions and propose a more realistic way to generate partial reconstructions through synthetic camera trajectories and retriangulation. + +# 3. Pipeline for SfM Registration Dataset Generation + +We find that current SotA point cloud registration models trained on 3DMatch [65], the most commonly used training set for 3D registration, are not able to register SfM reconstruction out of the box (cf. Tabs. 1, 2 and 4). Intuitively this is due to the large distribution shift between registration of RGB-D and/or LiDAR scans, and the much larger scenes covered by SfM reconstructions. However, it is not obvious how to generate a diverse dataset with multiple aligned reconstructions. Manually collecting such datasets is prohibitively expensive, and publicly available datasets typically contain a single reconstruction per scene. + +To tackle this issue, we propose a pipeline to generate a large synthetic dataset for registration between SfM models from random image collection datasets, e.g., MegaDepth [30]. Our key insight is that we can leverage these models and for each scene to construct partial reconstructions by sampling subsets of images/c cameras. An overview of our approach is presented in Fig. 3. In the next sections we describe our approach in more detail. + +# 3.1. Creating Partial Reconstructions + +We use hloc [42] with COLMAP [44] to create partial reconstructions from scenes. We use SIFT [35] and SOSNet [51] retriangulations for training and evaluation. We consider two approaches to create partial reconstructions. + +Random point sampling: In the simplest approach, we sample a set of 3D points from the scene for each partial reconstruction, and add the images in which they are visible until reaching approximately 200 images. For each scene we create 10 such partial reconstructions. We found that while this approach works well when registering reconstructions from collections of random images, it does not generalize as well to reconstructions from a single camera trajectory, which are common in real-life scenarios, e.g., video. We believe that this is due to the point clouds exhibiting more significant variance when the scene is cap- + +tured from different viewpoints. For example, keypoints are likely to be denser for nearby cameras, and natural occlusions will depend on viewpoints. As random image based datasets are not naturally divisible into videos, it is not obvious how to bridge this gap. We next describe our approach addressing this. + +Partial Trajectories: Our approach to bridge the random image $\rightarrow$ video gap is to generate synthetic trajectories from reconstructions of random images. Concretely, we randomly pick a starting image, and then sequentially (without replacement) pick the nearest neighbour of the image using a distance metric (which we refer to as dist). In practice we found that, for each trajectory, using a randomly weighted combination of geodesic rotation distance and Euclidean distance on the translation produced satisfactory results. Each trajectory has a size between $n_{\mathrm{low}} = 75$ and $n_{\mathrm{high}} = 300$ . The algorithm is presented in Algorithm 1. We show a qualitative example of three generated synthetic trajectories in Fig. 4. Running this on all scenes gave us between 5 to 20 synthetic trajectories per scene. Training is conducted on an equally weighted combination of the random point sampling approach and the synthetic trajectories. + +Algorithm 1 Generation of synthetic trajectory from a set of posed images +Precondition: $\mathbb{T}$ set of posed images, $\mathbb{I}$ remaining index set, distance function dist +1: function GENERATETRAJECTORY(T, I) +2: $\tau \gets \emptyset$ +3: i $\leftarrow$ Random sample from I +4: n $\leftarrow$ Random sample from $\{n_{\mathrm{low}},\dots ,n_{\mathrm{high}}\}$ +5: $\mathbb{I}\gets \mathbb{I}\backslash i$ +6: $\tau \gets \tau \cup i$ +7: for $k\gets 1$ to $n$ do +8: $i^{\prime}\gets \arg \min_{i^{\prime}\in \mathbb{I}}\mathrm{dist}(\mathbb{T}_{i},\mathbb{T}_{i^{\prime}})$ +9: $\tau \gets \tau \cup i^{\prime}$ +10: $\mathbb{I}\gets \mathbb{I}\backslash i^{\prime}$ +11: $i\gets i^{\prime}$ +12: end for +13: return $\tau ,\mathbb{I}$ +14: end function + +# 3.2. Pipeline and Dataset + +Reconstruction Pipeline: We are interested in having good ground truth correspondences between the partial reconstructions. As the SfM pipeline accuracy usually degrades with fewer available cameras, we use fixed camera poses from the joint reconstruction using all images and simply run triangulation from feature matching. This also has the benefit of being significantly less computationally + +![](images/8c6176577430faea20741dd3474d55eae1e91e84627eb5f3d7ddd10f6a8938ce.jpg) +Figure 3. Overview of our pipeline. For each scene, from a random image SfM dataset, e.g., MegaDepth [30], we retriangulate partial reconstructions using partial trajectories from the scene. Since these trajectories are in the global SfM reference frame, the relative transformation is simply the identity mapping. + +![](images/5e26daff9cf077a7f0594b642cfd5aa9131102f539201cfac5a45660b73d5bd7.jpg) +Figure 4. Example of synthetic trajectories in our proposed dataset. We start with a large scale scene, consisting of hundreds of cameras (left). From this set of cameras we run Algorithm 1 until the remaining set of cameras is small. Plotting the trajectories shows that plausible camera motion is achieved by this procedure (right). Sampling in this way bridges the gap between random image collections and video-based trajectories. + +demanding than running the full reconstructions. + +Dataset Construction: We run our pipeline on all 196 MegaDepth scenes where 10 of them are reserved for the test benchmark. We use either SIFT [35] or SOSNet [51] features for the retriangulation, however any local features could be used. We compute the overlap between all pairs of partial reconstructions. To train models we enforce the overlap to be larger than $30\%$ . In total, after enforcing sufficient overlap, we are left with about 22000 pairs ( $\sim$ 20000 for training and $\sim$ 2000 for testing), which constitutes our dataset. Details on how point cloud overlap is computed are presented in the supplementary material. + +# 4. RoITr and RefineRoITr for ColabSfM + +In this section, we discuss our baseline and our proposed improvements for ColabSfM. An overview of the proposed registration model $f_{\theta}$ is presented in Figure 5. For consistency we follow the notation of Yu et al. [64]. + +# 4.1. Preliminaries + +We are interested in registration methods that are robust to the challenges posed by SfM reconstructions. In particular, one of the main challenges is that SfM reconstructions are defined up-to-scale, and the rotation and origin of the reference frame is arbitrary1. Hence, we would like the registration model to be approximately invariant to the action of the similarity group, i.e., $\mathrm{Sim}(3) = \mathrm{SE}(3) + \mathrm{scale}$ . In practice we will achieve this by using an SE(3) invariant model, and learn approximate scale invariance. We next describe our baseline, and then our proposed improvements. + +# 4.2. Baseline Model + +As a strong baseline, we base our approach on the SotA SE(3) invariant registration method RoITr [64]. We next give an overview of our baseline. An more thorough description of the architecture is given in the supplementary and by Yu et al. [64]. + +Rotation Invariance through PPFs: RoITr [64] is based on point pair features (PPFs) [16], which computes a 4D feature for point pairs by the angles between their relative position and their normals. Mathematically, the PPF is defined as $\mathrm{PPF}(p_1,p_2,n_1,n_2) =$ $[\| d\| ,\angle (n_1,d),\angle (n_2,d),\angle (n_1,n_2)]$ , where $d = p_{2} - p_{1}$ It can easily be shown that this construction is invariant to rotations, + +$$ +\begin{array}{l} \operatorname {P P F} \left(R p _ {1}, R p _ {2}, R n _ {1}, R n _ {2}\right) = [ \| R d \|, \angle \left(R n _ {1}, R d\right), \\ \angle (R n _ {2}, R d), \angle (R n _ {1}, R n _ {2}) ] = [ \| d \|, \angle (n _ {1}, d), \angle (n _ {2}, d), \\ \left. \angle \left(n _ {1}, n _ {2}\right) \right] = \operatorname {P P F} \left(p _ {1}, p _ {2}, n _ {1}, n _ {2}\right), \tag {1} \\ \end{array} +$$ + +since norms and angles are invariant to rotations. + +![](images/624943ce69130c89c89e4623403e4a9cae83d8a5ef7844f1d3cfc722152ba5d4.jpg) +Figure 5. Overview of our proposed model RefineRoITr. As input we take two point clouds $\mathcal{P} = (\mathbf{P},\mathbf{N},\mathbf{X})$ $\mathcal{Q} = (\mathbf{Q},\mathbf{M},\mathbf{Y})$ consisting of 3D points $\mathbf{P} / \mathbf{Q}\in \mathbb{R}^{n / m\times 3}$ , and normals $\mathbf{N} / \mathbf{M}\in \mathbb{R}^{n / m\times 3}$ . In this work the features $\mathbf{X} / \mathbf{Y}$ are always assumed to be constant. We ensure rotation invariance by Point Pair Feature (PPF) encoding. The PPFs are fed together with the features into an encoder $e_{\theta}$ , producing $N^{\prime}$ coarse superpoints $\mathcal{P}'/\mathcal{Q}'$ . These are passed through a global Transformer $g_{\theta}$ , from which the top- $k$ coarse correspondences $\mathcal{C}'\in \mathbb{R}^{k\times 2}$ are extracted. A decoder $d_{\theta}$ takes the coarse features from $g_{\theta}$ and the finer features from $e_{\theta}$ and produces fine features $\hat{\mathbf{X}},\hat{\mathbf{Y}}$ . Using the coarse correspondences we extract neighbourhoods $\hat{\mathbf{G}}^X,\hat{\mathbf{G}}^Y\in \mathbb{R}^{k\times 64\times c}$ , which we feed into our proposed refinement Transformer. The refined features from the Transformer are used to construct a cost matrix, where the Sinkhorn algorithm [43, 54, 64] is used to solve the optimal transport (OT) problem, producing our final matches $\mathcal{C}$ . The network is trained using $\mathcal{L}_s + \mathcal{L}_p$ . + +Normalization Procedure: SfM reconstructions do not in general come with metric scale. This poses several challenges for registration methods, that typically assume both point clouds to share the same scale. We tackle this by normalization. Specifically, for Sim(3) training we normalize the point clouds $\mathbf{P}$ , $\mathbf{Q}$ with $\sigma_{\mathbf{P}} / \sqrt{2}$ , $\sigma_{\mathbf{Q}} / \sqrt{2}$ , where $\sigma$ is the largest singular value of the centered point cloud. For SE(3) training, we simply use $\sigma_{\mathbf{P}}$ to normalize both $\mathbf{P}$ and $\mathbf{Q}$ . We find that this simple approach leads to models robust to scale variation, as we demonstrate in Tab. 1. + +Normal Estimation: We follow RoITr [64] and use Open3D [68] with a neighborhood size of 33 for estimating normals, we found that this works well in practice. In contrast to the RGB-D scans of 3DMatch, our dataset consists of point clouds from multiple sensors. This means that the simple approach of aligning the normal orientation towards the origin of the coordinate system, as done in RoITr, will not be consistent in general. To remedy this, we use the fact that all cameras observing a 3D track will yield consistent orientation estimates. In practice, we align the orientation towards the camera center of a random image observing the track, which we found yields consistent orientation. + +Overview of Architecture: RoITr consists of an encoder $e_{\theta}$ , a global Transformer $g_{\theta}$ , and a decoder $d_{\theta}$ . $e_{\theta}$ and $d_{\theta}$ take as input sets of PPF encoded local neighbourhoods (typically around 8/16 points) around each point/superpoint. In contrast, $g_{\theta}$ performs global attention on the coarse point cloud $\hat{\mathcal{P}}$ . $e_{\theta}, d_{\theta}, g_{\theta}$ all use the Transformer architectures as described by Yu et al. [64]. Note that RoITr does not include the refining Transformer $r_{\theta}$ , which will be detailed in our RefineRoITr. + +Losses: RoITr uses a superpoint matching loss $\mathcal{L}_s$ and a point matching $\mathcal{L}_p$ loss for training. We follow RoITr + +and use their overlap-aware [39] circle loss [49] as $\mathcal{L}_s$ . The point matching loss $\mathcal{L}_p$ is the Negative Log Likelihood (NLL) of the GT correspondences after running the Sinkhorn algorithm [43] on the fine feature similarities $\tilde{S}$ . + +# 4.3. RefineRoITr (Ours) + +Refinement Transformer $r_{\theta}$ : The refinement in RoITr is performed by optimizing the Sinkhorn [54] distance between shallow features in a local neighbourhood around the coarse matches, denoted $\hat{\mathbf{G}}^X, \hat{\mathbf{G}}^Y$ . While this provides satisfactory results on 3DMatch, we found that we could improve results by incorporating a joint local Transformer $r_{\theta}$ . This Transformer takes $\hat{\mathbf{G}}^X, \hat{\mathbf{G}}^Y$ as inputs and outputs enriched neighbourhoods $\hat{\mathbf{G}}^X, \hat{\mathbf{G}}^Y$ . We found that these cross-point cloud enriched fine features from $r_{\theta}$ yielded improved results. + +Architectural Details of $r_{\theta}$ : $r_{\theta}$ takes in local neighbourhoods of shape $m \times 64 \times c$ , where $m$ is the number of neighbourhoods, 64 is the number of neighbours (identical to RoITr), and $c$ is the feature dimensionality. We use layers of interleaved self-attention and cross-attention as in LightGlue [31]. Note however, that in contrast to LightGlue that performs global attention, our refinement Transformer performs attention locally, which is significantly cheaper. We use an input/output dimensionality of $c_{\mathrm{in}} = c_{\mathrm{out}} = c = 64$ , and 4 self+cross attention layers. Each self/cross attention layer consists of single head attention, using dim $c$ , which is stacked with the input into a 1 hidden layer MLP (input dim $2c$ , hidden dim $2c$ , output dim $c$ ) using the GELU [23] activation function, with layer normalization [2] applied before GELU. Each self/cross block is residual, i.e., the output of the block is the output of the MLP plus the input. + +Table 1. Results of SfM registration on MegaDepth. We evaluate 2 different scenarios. Unknown relative orientation SO(3) + unknown relative position = SE(3), SE(3) + unknown relative scale = Sim(3). The top portion contains methods only trained on the 3DMatch (3DM) dataset, the middle portion methods trained only on our proposed dataset (Mega), while the lower portion contains methods trained on the former and fine-tuned on the latter. + +
MethodSE(3)Sim(3)
IRFMRRRIRFMRRR
OverlapPredator [24] (3DM)6.135.510.03.621.32.1
GeoTransformer [39] (3DM)2.114.63.31.38.60.4
PareNet [61] (3DM)7.338.65.24.727.70.8
RoITr [64] (3DM)3.012.60.01.67.00.8
RefineRoITr (3DM)10.032.813.96.222.83.5
RefineRoITr (Mega)51.096.570.244.692.842.7
OverlapPredator [24] (3DM + Mega)21.374.935.411.456.210.7
RoITr [64] (3DM + Mega)44.690.960.038.486.138.7
RefineRoITr (3DM + Mega)48.795.167.744.693.944.3
+ +# 5. Experiments + +In this section, we evaluate the RefineRoITr model presented in Sec. 4. Three different model instances were trained: 1) pre-trained RefineRoITr on 3DMatch [65]; 2) the model in 1) fine-tuned on the dataset in Sec. 3 and; 3) RefineRoITr trained from scratch on our dataset. For comparison, we used RoITr [64] and OverlapPredator [24], both with pre-trained weights from 3DMatch, and fine-tuned weights trained on our dataset. Additionally, we tested GeoTransformer [39] also pre-trained on 3DMatch [65]. As we found that the performance, of fine-tuning 3DMatch trained models, on our dataset was similar for our proposed model, we did not train other models in our dataset from scratch. + +Metrics: We use the standard registration metric Inlier Ratio (IR). This measures the percentage of estimated correspondences lying within a threshold distance $\tau_{\mathrm{IR}} = 0.1$ from each other. We additionally measure the Feature Matching Recall (FMR), which is defined as the percentage of point cloud pairs whose IR is above a threshold $\tau_{\mathrm{FMR}} = 5\%$ . Finally, we use Registration Recall (RR), which consists of the percentage of pairs for which the rotation error and translation error are below $5^{\circ}$ and $0.05$ , respectively. We solve for the pose using the RANSAC [18] implementation from Open3D [68] on the matches retrieved by the neural networks. Details on error computation and RANSAC can be found in the Supplemental material. + +Computational cost: On a desktop machine equipped with an i9-13900K @ 3.00GHz CPU and an NVIDIA 4090 GPU, the median runtime for RoITr is 99.94 ms and for RefineRoITris 102.84, which shows that the computational burden added by the modifications described in Sec. 4.3 is around $3\%$ . + +# 5.1. MegaDepth SfM Registration Benchmark + +We select 10 scenes from MegaDepth [30] dataset, which consists of 196 large scale 3D scenes reconstructed from internet images for our benchmark. The test scenes are, in order, [0008, 0015, 0019, 0021, 0022, 0024, 0025, 0032, 0063, 1589]. The average size of the point clouds is $15.3\mathrm{K}$ points, using either SIFT or SOSNet features for the triangulation. To train and evaluate the network presented in Sec. 4 on our dataset, we define two different scenarios SE(3) and Sim(3). We define a random rotation by sampling three Euler angles $\alpha, \beta, \gamma \in [0,2\pi]$ . + +The results of the RefineRoITr method and the baselines are presented in Tab. 1. As expected, the methods which are not fine-tuned in our data perform the worst in both SE(3) and Sim(3) data. This shows the lack of generalization of current registration methods on SfM point clouds. When considering FMR, the methods trained from scratch or fine-tuned on our data obtain high scores ( $>90\%$ ) on SE(3) data with slightly lower scores on Sim(3). When looking at the three variations of RefineRoITr, we see that while the version trained on 3DMatch outperforms vanilla RoITr, the major boost in performance is due to fine-tuning the model on our dataset. The best overall version is the one trained from scratch on our dataset, even though the improvement is not as significant. Qualitative results are presented in Fig. 2 and the Supplementary material. + +# 5.2. Standard Benchmark: Cambridge Landmarks + +The Cambridge Landmarks [28] dataset consists of 5 large-scale 3D scenes recorded with a smartphone by a pedestrian in Cambridge, each of which consisting of a train and a test trajectories with camera poses. We retriangulate the train/test sequences independently, which results in two reconstructions for each scene. In this case, we evaluate the IR and number of matches. Here we only test the SE(3) case. Rotation and translation errors and qualitative results for this dataset are reported in the supplementary material. The results are presented in Tab. 2. Similar to the evaluation results in the previous section, the models trained only on 3DMatch [65] perform poorly, while the ones fine-tuned and or trained from scratch on our dataset present the highest IR and number of matches, with small differences in favor of the version trained from scratch. + +# 5.3. Noisy Indoor Reconstructions: 7-Scenes + +The 7-Scenes [46] dataset is an indoor RGB-D dataset. This dataset is challenging for SfM, since textured regions lead to significantly more noisy point clouds, and less geometric features. We present results in Tab. 3. Encouragingly, we find that models trained using our pipeline perform well also on these indoor scenes, showing the generalizability of our approach. + +Table 2. Results of SfM registration on Cambridge Landmarks [28]. We evaluate unknown relative orientation $\mathrm{SO}(3) +$ unknown relative position $= \mathrm{{SE}}\left( 3\right)$ . The top portion contains methods only trained on the 3DMatch (3DM) dataset, the middle portion methods trained only on our proposed dataset (Mega), while the lower portion contains methods trained on the former and fine-tuned on the latter. + +
MethodGreat CourtKings CollegeOld HospitalShop FacadeSt Mary's Church
IRMatchesIRMatchesIRMatchesIRMatchesIRMatches
OverlapPredator [24] (3DM)1.43560.53720.03632.53540.2382
GeoTransformer [39] (3DM)0.03820.17850.04380.113680.0431
PureNet [61] (3DM)4.320001.62000020001.220006.72000
RoiTr [64] (3DM)0.02850.04470.03210.39591.5407
RefineRoiTr (3DM)0.24122.24080.05440.97372.8567
RefineRoiTr (Mega)70.9453857.6397731.5221441.6132581.85167
RoiTr [64] (3DM + Mega)52.1142639.6137721.960328.070064.52481
RefineRoiTr (3DM + Mega)69.4455154.0324628.0159139.1134277.85168
+ +Table 3. Results of SfM registration on 7-Scenes. We evaluate unknown relative orientation $\mathrm{SO}(3) +$ unknown relative position $= \mathrm{{SE}}\left( 3\right)$ . The top portion contains methods only trained on the 3DMatch (3DM) dataset, the middle portion methods trained only on our proposed dataset (Mega), while the lower portion contains methods trained on the former and fine-tuned on the latter. + +
MethodChessFireHeadsOfficePumpkinRed KitchenStairs
IRMatchesIRMatchesIRMatchesIRMatchesIRMatchesIRMatchesIRMatches
RoITr [64] (3DM)0.03070.010040.016990.013220.03270.011630.0730
RefineRoITr (3DM)0.08420.022040.011040.010730.010160.017210.01043
RefineRoITr (Mega)23.4671738.8904350.6541968.0946114.9289327.980506.23230
RoITr [64] (3DM+Mega)20.4209830.5416536.4381250.047527.9127118.524389.11883
RefineRoITr (3DM+Mega)23.6563233.3833950.5599662.788878.9226826.866166.42863
+ +Table 4. Results of SfM registration on Quad6k [11]. We evaluate unknown relative orientation $\mathrm{SO}(3) +$ unknown relative position $=$ SE(3). The top portion contains methods only trained on the 3DMatch (3DM) dataset, the middle portion methods trained only on our proposed dataset (Mega), while the lower portion contains methods trained on the former and fine-tuned on the latter. + +
MethodIRFMRMatchesRR
OverlapPredator [24] (3DM)1.610.21210.0
GeoTransformer [39] (3DM)0.60.019360.0
PareNet [61] (3DM)1.610.420000.0
RoITr [64] (3DM)0.40.05740.0
RefineRoITr (3DM)1.76.16172.0
RefineRoITr (Mega)17.267.487822.4
RoITr [64] (3DM + Mega)14.263.349224.5
RefineRoITr (3DM + Mega)18.165.371516.3
+ +# 5.4. Low Overlap: Quad6k + +The Quad6k dataset [11] consists of about 6500 images from the Arts Quad at Cornell University. To create partial reconstructions for testing, we first perform a full scene retriangulation using the refined poses provided by [52] and then follow the procedure in Sec. 3.1, generating a total of 24 trajectories, and 49 pairs. The results are presented in Tab. 4. Similar to the previous datasets, the performance of the methods has a substantial boost after training on the dataset presented in Sec. 3. GeoTransformer [39] trained only on 3DMatch [65] presents the most matches on average, but only a small fraction (0.6%) are inliers, and its FMR and RR are both zero. Our Quad6k benchmark is + +more challenging and thus the performance is lower than in Tabs. 1 and 2, due to a low overlap of $16.5\%$ , in comparison with the $61.2\%$ of MegaDepth. Qualitative results for this dataset are shown in the supplementary material. + +# 6. Conclusion + +We have proposed Collaborative Structure-from-Motion by Point Cloud Registration, ColabSfM, a challenging new task to enable scalable collaborative mapping and localization. By extensive experiments we showed that with our proposed dataset generation pipeline and SE(3) invariant model, this task is solvable, and showed that our model, while trained only on MegaDepth, is able to generalize to several challenging scenarios, such as noisy indoor reconstructions, and low-overlap registration. We believe our initial results will encourage further research on collaborative 3D reconstruction. Limitations: a) Our proposed model RefineRoITr is E(3) invariant like RoITr, and therefore can struggle with estimating correct matches in symmetric scenes. b) Our models are trained and evaluated on retriangulated partial reconstructions. In practice, issues such as drift occur, making global alignment more difficult, however we find that our model still performs well in this setting in Table 7. + +# Acknowledgements + +This work was supported by the strategic research environment ELLIIT, funded by the Swedish government. + +# References + +[1] Yasuhiro Aoki, Hunter Goforth, Rangaprasad Arun Srivatsan, and Simon Lucey. Pointnetlk: Robust & efficient point cloud registration using pointnet. In IEEE Conf. Comput. Vis. Pattern Recog., pages 7163-7172, 2019. 2, 3 +[2] Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. arXiv preprint arXiv:1607.06450, 2016. 6 +[3] Paul J. Besl and Neil D. McKay. A method for registration of 3-d shapes. IEEE Trans. Pattern Anal. Mach. Intell., 14 (2):239-256, 1992. 3 +[4] Georg Bokman and Fredrik Kahl. A case for using rotation invariant features in state of the art feature matchers. In IEEE Conf. Comput. Vis. Pattern Recog., pages 5110-5119, 2022. 3 +[5] Dylan Campbell, Liu Liu, and Stephen Gould. Solving the blind perspective-n-point problem end-to-end with robust differentiable geometric optimization. In Eur. Conf. Comput. Vis., pages 244-261, 2020. 3 +[6] Yang Chen and Gerard Medioni. Object modeling by registration of multiple range images. In IEEE Int'l Conf. Robotics and Auto., pages 2724-2729, 1991. 3 +[7] Christopher Choy, Jaesik Park, and Vladlen Koltun. Fully convolutional geometric features. In Int. Conf. Comput. Vis., pages 8958-8966, 2019. 3 +[8] Andrea Cohen, Torsten Sattler, and Marc Pollefeys. Merging the unmatched: Stitching visually disconnected sfm models. In Int. Conf. Comput. Vis., pages 2129-2137, 2015. 3 +[9] Andrea Cohen, Johannes L Schonberger, Pablo Speciale, Torsten Sattler, Jan-Michael Frahm, and Marc Pollefeys. Indoor-outdoor 3d reconstruction alignment. In Eur. Conf. Comput. Vis., pages 285-300, 2016. 3 +[10] James Coughlan and Alan L Yuille. The manhattan world assumption: Regularities in scene statistics which enable bayesian inference. Adv. Neural Inform. Process. Syst., 13, 2000. 3 +[11] David Crandall, Andrew Owens, Noah Snavely, and Dan Huttenlocher. Discrete-continuous optimization for large-scale structure from motion. In IEEE Conf. Comput. Vis. Pattern Recog., pages 3001–3008, 2011. 3, 8 +[12] Deeksha Dangwal, Vincent T Lee, Hyo Jin Kim, Tianwei Shen, Meghan Cowan, Rajvi Shah, Caroline Trippel, Brandon Reagen, Timothy Sherwood, Vasileios Balntas, et al. Mitigating reverse engineering attacks on local feature descriptors. In Brit. Mach. Vis. Conf., 2021. 2 +[13] Haowen Deng, Tolga Birdal, and Slobodan Ilic. Ppf-foldnet: Unsupervised learning of rotation invariant 3d local descriptors. In Eur. Conf. Comput. Vis., pages 602-618, 2018. 3 +[14] Haowen Deng, Tolga Birdal, and Slobodan Ilic. Ppfnet: Global context aware local features for robust 3d point matching. In IEEE Conf. Comput. Vis. Pattern Recog., pages 195-205, 2018. 3 +[15] Alexey Dosovitskiy and Thomas Brox. Inverting visual representations with convolutional networks. In IEEE Conf. Comput. Vis. Pattern Recog., pages 4829-4837, 2016. 2 +[16] Bertram Drost, Markus Ulrich, Nassir Navab, and Slobodan Ilic. Model globally, match locally: Efficient and robust + +3d object recognition. In IEEE Conf. Comput. Vis. Pattern Recog., pages 998-1005, 2010. 3, 5 +[17] Mihai Dusmanu, Ondrej Miksik, Johannes L Schonberger, and Marc Pollefeys. Cross-descriptor visual localization and mapping. In Int. Conf. Comput. Vis., pages 6058-6067, 2021. 2, 3 +[18] Martin A Fischler and Robert C Bolles. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM, 24(6):381-395, 1981. 3, 7, 1 +[19] Kexue Fu, Shaolei Liu, Xiaoyuan Luo, and Manning Wang. Robust point cloud registration framework based on deep graph matching. In IEEE Conf. Comput. Vis. Pattern Recog., pages 8893-8902, 2021. 3 +[20] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In IEEE Conf. Comput. Vis. Pattern Recog., pages 3354-3361, 2012. 2 +[21] Zan Gojcic, Caifa Zhou, Jan D Wegner, and Andreas Wieser. The perfect match: 3d point cloud matching with smoothed densities. In IEEE Conf. Comput. Vis. Pattern Recog., pages 5545-5554, 2019. 3 +[22] Michal Havlena, Akihiko Torii, and Tomáš Pajdla. Efficient structure from motion by graph optimization. In Eur. Conf. Comput. Vis., pages 100-113, 2010. 2 +[23] Dan Hendrycks and Kevin Gimpel. Gaussian error linear units (gelus). arXiv preprint arXiv:1606.08415, 2016. 6 +[24] Shengyu Huang, Zan Gojcic, Mikhail Usvyatsov, Andreas Wieser, and Konrad Schindler. Predator: Registration of 3d point clouds with low overlap. In IEEE Conf. Comput. Vis. Pattern Recog., pages 4267-4276, 2021. 2, 3, 7, 8, 1, 4, 5 +[25] Xiaoshui Huang, Guofeng Mei, and Jian Zhang. Feature-metric registration: A fast semi-supervised approach for robust point cloud registration without correspondences. In IEEE Conf. Comput. Vis. Pattern Recog., pages 11366-11374, 2020. 3 +[26] Shengze Jin, Iro Armeni, Marc Pollefeys, and Daniel Barath. Multiway point cloud mosaicking with diffusion and global optimization. In IEEE Conf. Comput. Vis. Pattern Recog., pages 20838-20849, 2024. 3 +[27] Neena Kamath. Announcing azure spatial anchors for collaborative, cross-platform mixed reality apps, 2019. URL https://azure.microsoft.com/en-us/blog/announcing-azure-spatial-anchors-for-collaborative-cross-platform-mixed-reality-apps/.1 +[28] Alex Kendall, Matthew Grimes, and Roberto Cipolla. Posenet: A convolutional network for real-time 6-dof camera relocalization. In Int. Conf. Comput. Vis., pages 2938-2946, 2015. 3, 7, 8, 2, 6 +[29] Hongdong Li and Richard Hartley. The 3d-3d registration problem revisited. In Int. Conf. Comput. Vis., pages 1-8, 2007. 3 +[30] Zhengqi Li and Noah Snavely. Megadepth: Learning single-view depth prediction from internet photos. In IEEE Conf. Comput. Vis. Pattern Recog., pages 2041-2050, 2018. 3, 4, 5, 7 + +[31] Philipp Lindenberger, Paul-Edouard Sarlin, and Marc Pollefeys. LightGlue: Local Feature Matching at Light Speed. In Int. Conf. Comput. Vis., 2023. 2, 6, 1 +[32] Liu Liu, Hongdong Li, Haodong Yao, and Ruyi Zha. Pluckernet: Learn to register 3d line reconstructions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1842-1852, 2021. 4 +[33] Alex Locher, Michal Havlena, and Luc Van Gool. Progressive structure from motion. In Eur. Conf. Comput. Vis., pages 20-35, 2018. 2 +[34] Alexandre Lopes, Roberto Souza, and Helio Pedrini. A survey on rgb-d datasets. Comput. Vis. Img. Underst., 222: 103489, 2022. 2 +[35] David G Lowe. Distinctive image features from scale-invariant keypoints. Int. J. Comput. Vis., 60:91-110, 2004. 2, 4, 5 +[36] André Mateus, Siddhant Ranade, Srikumar Ramalingam, and Pedro Miraldo. Fast and accurate 3d registration from line intersection constraints. Int. J. Comput. Vis., pages 1-26, 2023. 3 +[37] Pedro Miraldo, Surojit Saha, and Srikumar Ramalingam. Minimal solvers for mini-loop closures in 3d multi-scan alignment. In IEEE Conf. Comput. Vis. Pattern Recog., 2019. 3 +[38] Francesco Pittaluga, Sanjeev J Koppal, Sing Bing Kang, and Sudipta N Sinha. Revealing scenes by inverting structure from motion reconstructions. In IEEE Conf. Comput. Vis. Pattern Recog., pages 145-154, 2019. 2 +[39] Zheng Qin, Hao Yu, Changjian Wang, Yulan Guo, Yuxing Peng, and Kai Xu. Geometric transformer for fast and robust point cloud registration. In IEEE Conf. Comput. Vis. Pattern Recog., pages 11143-11152, 2022. 2, 3, 6, 7, 8 +[40] Tilman Reinhardt. Using global localization to improve navigation, 2019. URL https://blog.research.google/2019/02/using-global-localization-to-improve.html.1 +[41] Radu Bogdan Rusu, Nico Blodow, and Michael Beetz. Fast point feature histograms (fpfh) for 3d registration. In IEEE Int'l Conf. Robotics and Auto., pages 3212-3217, 2009. 3 +[42] Paul-Edouard Sarlin, Cesar Cadena, Roland Siegwart, and Marcin Dymczyk. From coarse to fine: Robust hierarchical localization at large scale. In IEEE Conf. Comput. Vis. Pattern Recog., pages 12716–12725, 2019. 4 +[43] Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. SuperGlue: Learning feature matching with graph neural networks. In IEEE Conf. Comput. Vis. Pattern Recog., 2020. 2, 6 +[44] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In IEEE Conf. Comput. Vis. Pattern Recog., 2016. 2, 4 +[45] Peter H. Schonemann. A generalized solution of the orthogonal procrustes problem. Psychometrika, 31(1):1-10, 1966. 3 +[46] Jamie Shotton, Ben Glocker, Christopher Zach, Shahram Izadi, Antonio Criminisi, and Andrew Fitzgibbon. Scene coordinate regression forests for camera relocalization in rgb-d images. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2930-2937, 2013. 7 + +[47] Tory Smith. Introducing lightship vps, 2022. URL https://lightship.dev/blog/introducing-lightship-vps/.1 +[48] Christoph Strecha, Timo Pylvänäinen, and Pascal Fua. Dynamic and scalable large scale image reconstruction. In IEEE Conf. Comput. Vis. Pattern Recog., pages 406-413, 2010. 3 +[49] Yifan Sun, Changmao Cheng, Yuhan Zhang, Chi Zhang, Liang Zheng, Zhongdao Wang, and Yichen Wei. Circle loss: A unified perspective of pair similarity optimization. In IEEE Conf. Comput. Vis. Pattern Recog., pages 6398-6407, 2020. 6 +[50] Chris Sweeney, Victor Fragoso, Tobias Höllerer, and Matthew Turk. Large scale sfm with the distributed camera model. In Int'l Conf. 3D Vision, pages 230-238, 2016. 2 +[51] Yurun Tian, Xin Yu, Bin Fan, Fuchao Wu, Huub Heijnen, and Vassileios Balntas. Sosnet: Second order similarity regularization for local descriptor learning. In IEEE Conf. Comput. Vis. Pattern Recog., pages 11016-11025, 2019. 4, 5 +[52] Haithem Turki, Deva Ramanan, and Mahadev Satyanarayanan. Mega-nerf: Scalable construction of large-scale nerfs for virtual fly-throughs. In IEEE Conf. Comput. Vis. Pattern Recog., pages 12922–12931, 2022. 8 +[53] Michal Tyszkiewicz, Pascal Fua, and Eduard Trulls. Disk: Learning local features with policy gradient. Adv. Neural Inform. Process. Syst., 33:14254-14265, 2020. 1 +[54] Ole Untzelmann, Torsten Sattler, Sven Middelberg, and Leif Kobbelt. A scalable collaborative online system for city reconstruction. In Proceedings of the IEEE International Conference on Computer Vision Workshops, pages 644-651, 2013. 3, 6 +[55] Shuzhe Wang, Juho Kannala, and Daniel Barath. Dgc-gnn: Leveraging geometry and color cues for visual descriptor-free 2d-3d matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20881–20891, 2024. 2, 3 +[56] Yue Wang and Justin M Solomon. Deep closest point: Learning representations for point cloud registration. In Int. Conf. Comput. Vis., pages 3523-3532, 2019. 3 +[57] Yue Wang and Justin M Solomon. Prnet: Self-supervised learning for partial-to-partial registration. Adv. Neural Inform. Process. Syst., 32, 2019. 3 +[58] Philippe Weinzaepfel, Hervé Jégou, and Patrick Pérez. Reconstructing an image from its local descriptors. In IEEE Conf. Comput. Vis. Pattern Recog., pages 337-344, 2011. 2 +[59] Hao Xu, Shuaicheng Liu, Guangfu Wang, Guanghui Liu, and Bing Zeng. Omnet: Learning overlapping mask for partial-to-partial point cloud registration. In Int. Conf. Comput. Vis., pages 3132-3141, 2021. 3 +[60] Jiaolong Yang, Hongdong Li, and Yunde Jia. Go-icp: Solving 3d registration efficiently and globally optimally. In Int. Conf. Comput. Vis., pages 1457–1464, 2013. 3 +[61] Runzhao Yao, Shaoyi Du, Wenting Cui, Canhui Tang, and Chengwu Yang. Pare-net: Position-aware rotation-equivariant networks for robust point cloud registration. In ECCV, 2024. 7, 8, 2 + +[62] Zi Jian Yew and Gim Hee Lee. RPM-net: Robust point matching using learned features. In IEEE Conf. Comput. Vis. Pattern Recog., pages 11824-11833, 2020. 3 +[63] Zi Jian Yew and Gim Hee Lee. Regtr: End-to-end point cloud correspondences with transformers. In IEEE Conf. Comput. Vis. Pattern Recog., pages 6677-6686, 2022. 3 +[64] Hao Yu, Zheng Qin, Ji Hou, Mahdi Saleh, Dongsheng Li, Benjamin Busam, and Slobodan Ilic. Rotation-invariant transformer for point cloud matching. In IEEE Conf. Comput. Vis. Pattern Recog., pages 5384-5393, 2023. 1, 2, 3, 5, 6, 7, 8, 4 +[65] Andy Zeng, Shuran Song, Matthias Nießner, Matthew Fisher, Jianxiong Xiao, and Thomas Funkhouser. 3dmatch: Learning local geometric descriptors from rgb-d reconstructions. In IEEE Conf. Comput. Vis. Pattern Recog., pages 1802-1811, 2017. 2, 4, 7, 8, 1 +[66] Yejun Zhang, Shuzhe Wang, and Juho Kannala. A2-gnn: Angle-annular gnn for visual descriptor-free camera relocalization. In International Conference on 3D Vision (3DV), 2025. arXiv preprint arXiv:2502.20036. 2 +[67] Qunjie Zhou, Sérgio Agostinho, Aljosa Osep, and Laura Leal-Taixe. Is geometry enough for matching in visual localization? In Eur. Conf. Comput. Vis., pages 407-425, 2022. 2, 3 +[68] Qian-Yi Zhou, Jaesik Park, and Vladlen Koltun. Open3D: A modern library for 3D data processing. arXiv:1801.09847, 2018. 6, 7, 1 \ No newline at end of file diff --git a/2025/ColabSfM_ Collaborative Structure-from-Motion by Point Cloud Registration/images.zip b/2025/ColabSfM_ Collaborative Structure-from-Motion by Point Cloud Registration/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..729e152b19da69f40161f301b435e6284977a3ed --- /dev/null +++ b/2025/ColabSfM_ Collaborative Structure-from-Motion by Point Cloud Registration/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d5d55b804f47a49c39ac076caea0aff1eba15428cbe5cdc101f4684ffe2931f3 +size 504682 diff --git a/2025/ColabSfM_ Collaborative Structure-from-Motion by Point Cloud Registration/layout.json b/2025/ColabSfM_ Collaborative Structure-from-Motion by Point Cloud Registration/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..e97d56a19da14ff38746781eeba661666a37c7ae --- /dev/null +++ b/2025/ColabSfM_ Collaborative Structure-from-Motion by Point Cloud Registration/layout.json @@ -0,0 +1,8334 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 66, + 103, + 545, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 103, + 545, + 121 + ], + "spans": [ + { + "bbox": [ + 66, + 103, + 545, + 121 + ], + "type": "text", + "content": "ColabSfM: Collaborative Structure-from-Motion by Point Cloud Registration" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 179, + 147, + 430, + 162 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 147, + 430, + 162 + ], + "spans": [ + { + "bbox": [ + 179, + 147, + 430, + 162 + ], + "type": "text", + "content": "Johan Edstedt\\*1 André Mateus2 Alberto Jaenal2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 193, + 168, + 414, + 182 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 193, + 168, + 414, + 182 + ], + "spans": [ + { + "bbox": [ + 193, + 168, + 414, + 182 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 193, + 168, + 414, + 182 + ], + "type": "text", + "content": "Linköping University " + }, + { + "bbox": [ + 193, + 168, + 414, + 182 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 193, + 168, + 414, + 182 + ], + "type": "text", + "content": "Ericsson Research, Sweden" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 67, + 207, + 544, + 396 + ], + "blocks": [ + { + "bbox": [ + 67, + 207, + 544, + 396 + ], + "lines": [ + { + "bbox": [ + 67, + 207, + 544, + 396 + ], + "spans": [ + { + "bbox": [ + 67, + 207, + 544, + 396 + ], + "type": "image", + "image_path": "63d8556a7d3a7ee377336bc1d92af36db15e3b370e54ae4c18f88e5a2d2a60ec.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 405, + 555, + 473 + ], + "lines": [ + { + "bbox": [ + 54, + 405, + 555, + 473 + ], + "spans": [ + { + "bbox": [ + 54, + 405, + 555, + 473 + ], + "type": "text", + "content": "Figure 1. Our proposed registration paradigm for collaborative SfM reconstructions (ColabSfM). Given two input SfM reconstructions " + }, + { + "bbox": [ + 54, + 405, + 555, + 473 + ], + "type": "inline_equation", + "content": "\\mathcal{P},\\mathcal{Q}" + }, + { + "bbox": [ + 54, + 405, + 555, + 473 + ], + "type": "text", + "content": " of the same scene, the task is to estimate the relative similarity transform " + }, + { + "bbox": [ + 54, + 405, + 555, + 473 + ], + "type": "inline_equation", + "content": "(s,R,t)" + }, + { + "bbox": [ + 54, + 405, + 555, + 473 + ], + "type": "text", + "content": " between them. Our first contribution is to address this as a point cloud registration problem, using only 3D SfM tracks. For this, we do not rely on the visual descriptors, but on the 3D coordinates of the points " + }, + { + "bbox": [ + 54, + 405, + 555, + 473 + ], + "type": "inline_equation", + "content": "\\mathbf{P},\\mathbf{Q}" + }, + { + "bbox": [ + 54, + 405, + 555, + 473 + ], + "type": "text", + "content": ", their normals " + }, + { + "bbox": [ + 54, + 405, + 555, + 473 + ], + "type": "inline_equation", + "content": "\\mathbf{N},\\mathbf{M}" + }, + { + "bbox": [ + 54, + 405, + 555, + 473 + ], + "type": "text", + "content": " and, optionally, but not necessarily, features " + }, + { + "bbox": [ + 54, + 405, + 555, + 473 + ], + "type": "inline_equation", + "content": "\\mathbf{X},\\mathbf{Y}" + }, + { + "bbox": [ + 54, + 405, + 555, + 473 + ], + "type": "text", + "content": ". To make point cloud registration methods perform well on this task, we propose as our second contribution a scalable pipeline to construct synthetic training datasets for SfM registration. Finally, we propose an improved version of RoITr [64] as registration method " + }, + { + "bbox": [ + 54, + 405, + 555, + 473 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 54, + 405, + 555, + 473 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 151, + 478, + 201, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 478, + 201, + 491 + ], + "spans": [ + { + "bbox": [ + 151, + 478, + 201, + 491 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 499, + 297, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 499, + 297, + 693 + ], + "spans": [ + { + "bbox": [ + 54, + 499, + 297, + 693 + ], + "type": "text", + "content": "Structure-from-Motion (SfM) is the task of estimating 3D structure and camera poses from images. We define Collaborative SfM (ColabSfM) as sharing distributed SfM reconstructions. Sharing maps requires estimating a joint reference frame, which is typically referred to as registration. However, there is a lack of scalable methods and training datasets for registering SfM reconstructions. In this paper, we tackle this challenge by proposing the scalable task of point cloud registration for SfM reconstructions. We find that current registration methods cannot register SfM point clouds when trained on existing datasets. To this end, we propose a SfM registration dataset generation pipeline, leveraging partial reconstructions from synthetically generated camera trajectories for each scene. Finally, we propose a simple but impactful neural refiner on top of the SotA registration method RoITr that yields significant improve" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 479, + 556, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 479, + 556, + 528 + ], + "spans": [ + { + "bbox": [ + 313, + 479, + 556, + 528 + ], + "type": "text", + "content": "ments, which we call RefineRoITr. Our extensive experimental evaluation shows that our proposed pipeline and model enables ColabSfM. Code is available at https://github.com/EricssonResearch/ColabSfM" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 548, + 395, + 561 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 548, + 395, + 561 + ], + "spans": [ + { + "bbox": [ + 314, + 548, + 395, + 561 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 312, + 570, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 570, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 312, + 570, + 556, + 715 + ], + "type": "text", + "content": "For robots and extended reality (XR) devices to become ubiquitous in our lives, they must be able to accurately build and localize in maps of the environment. Most available solutions rely on visual data and onboard computing, which can limit the accuracy in the context of small devices. Thus, in the last years there has been a push from industry to move computation to the cloud (e.g. Google's Visual Positioning System [40], Microsoft's Azure Spatial Anchors [27], and Niantic's Lightship [47]). However, the heterogeneous nature of these systems prevents devices of different vendors from interoperating, i.e., sharing their maps and localizing in others' maps. If we are to make visual localization in-" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 702, + 245, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 702, + 245, + 713 + ], + "spans": [ + { + "bbox": [ + 66, + 702, + 245, + 713 + ], + "type": "text", + "content": "*Work done during an internship at Ericsson Research." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "type": "text", + "content": "6573" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 65, + 68, + 545, + 322 + ], + "blocks": [ + { + "bbox": [ + 65, + 68, + 545, + 322 + ], + "lines": [ + { + "bbox": [ + 65, + 68, + 545, + 322 + ], + "spans": [ + { + "bbox": [ + 65, + 68, + 545, + 322 + ], + "type": "image", + "image_path": "70d2831a980b6a8bb1a51e38bb366823f1fde1599554af37c6432bab8b1c5471.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 335, + 555, + 392 + ], + "lines": [ + { + "bbox": [ + 54, + 335, + 555, + 392 + ], + "spans": [ + { + "bbox": [ + 54, + 335, + 555, + 392 + ], + "type": "text", + "content": "Figure 2. Qualitative comparison. We compare our approach to the previous point cloud registration method OverlapPredator [24] on the St. Peter's Basilica test scene. Without training on our proposed SfM registration dataset (column 1), previous methods are unable to produce sufficiently good matches (top row) and accurate relative pose estimation results (bottom row). In contrast, our proposed model RefineRoITr, trained on the proposed dataset, is able to find better matches and hence register the scenes well. The source and target point clouds are depicted in yellow and blue, respectively." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 411, + 294, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 411, + 294, + 543 + ], + "spans": [ + { + "bbox": [ + 54, + 411, + 294, + 543 + ], + "type": "text", + "content": "teroperable, we need ways to merge such maps. Structurefrom-Motion (SfM) [44] is the most used method to process images and produce 3D maps (scene reconstructions consisting of a set of 3D points and the poses of the images expressed in a common coordinate system). To enable localization, a visual descriptor is stored for each map point, and for localization, 2D-3D matches are found by matching the descriptors of the query image with those of the map. Typically, SfM maps are merged using these visual descriptors, by finding 3D-3D matches [22, 33], or inserting new sub-maps based on 2D-3D correspondences [50]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "text", + "content": "However, these methods have limitations. First, different SfM pipelines tend to use different feature extractors, which yield incompatible descriptors. This is partially addressed by the cross-descriptor approach proposed in [17], but the assumption that the descriptors are computed from the same keypoints is unrealistic. Additionally, exposing visual descriptors may allow feature inversion attacks [12, 15, 38, 58], which can recover private elements from the images. The second limitation concerns scalability, since storing visual descriptors leads to a map size increase between two and three orders of magnitude [67]; and matching using a traditional pipeline [35] may suffer in performance; which can be enhanced by using learned methods [31, 43] at the cost of excessive runtimes. Visual" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 411, + 555, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 411, + 555, + 520 + ], + "spans": [ + { + "bbox": [ + 313, + 411, + 555, + 520 + ], + "type": "text", + "content": "localization without descriptors was tackled in [55, 66, 67], demonstrating competitive performance, but since they rely solely on 2D-3D geometry, which suffers from perspective distortions, their accuracy is not as high as their descriptor-based counterparts. Furthermore, they still rely on image retrieval, which again raises privacy, scalability, and interoperability concerns. Nevertheless, their results show that geometry alone could be sufficient for matching. Based on that observation, we pose the following question" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 335, + 527, + 541, + 539 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 527, + 541, + 539 + ], + "spans": [ + { + "bbox": [ + 335, + 527, + 541, + 539 + ], + "type": "text", + "content": "Can we merge SfM maps using only 3D geometry?" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 545, + 555, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 555, + 629 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 555, + 629 + ], + "type": "text", + "content": "To this end, we propose to merge SfM 3D maps through 3D point cloud registration methods [1, 65], which rely solely on geometric information, see Fig. 1. We call our proposed paradigm Collaborative SfM (ColabSfM). In contrast to visual localization, which aims to localize a single image to a map, our proposed task involves registering a map to a map, and is image retrieval-free." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 630, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 556, + 713 + ], + "type": "text", + "content": "State of the art methods for 3D point cloud registration rely on learned features [24, 39, 64], therefore it is necessary to provide suitable datasets to train such networks. There is a wide range of datasets available [34] for the point cloud registration task. However, those mostly consist of 3D sensor data, e.g., RGB-D [65] and/or LiDAR scans [20], which differ from SfM point clouds. To the best of our knowl" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6574" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 227 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 227 + ], + "type": "text", + "content": "edge, there are no large datasets for SfM point cloud registration. SfM datasets [30] provide a 3D point cloud for each scene, but since each scene is disconnected, no matching is possible between different scenes. To tackle this issue, we propose a pipeline to generate synthetic datasets for SfM point cloud registration, which generates multiple synthetic trajectories for each scene. The synthetic trajectories provide a trade-off between random internet images and sequential images (e.g., Cambridge Landmarks Dataset [28]). We apply our pipeline and generate training and benchmarks datasets from MegaDepth [30] and a benchmark dataset from Quad6k [11]. Our pipeline is presented in detail in Sec. 3." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 228, + 294, + 347 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 228, + 294, + 347 + ], + "spans": [ + { + "bbox": [ + 55, + 228, + 294, + 347 + ], + "type": "text", + "content": "SfM reference frames may be arbitrarily aligned relative to each other. As such, using a registration model invariant to rotation and translation SE(3), and robust to scale Sim(3), is important. To this end, we build on top of the recently proposed SotA rotation invariant model RoITr [64]. While this method already produces good results when trained on our dataset, we find that we can further improve results on ColabSfM by introducing a neural refinement stage to the model. We introduce our baseline model and our improved version, which we call RefineRoITr, in Sec. 4." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 348, + 218, + 358 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 348, + 218, + 358 + ], + "spans": [ + { + "bbox": [ + 67, + 348, + 218, + 358 + ], + "type": "text", + "content": "To summarize, our contributions are:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 359, + 295, + 431 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 56, + 359, + 295, + 383 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 359, + 295, + 383 + ], + "spans": [ + { + "bbox": [ + 56, + 359, + 295, + 383 + ], + "type": "text", + "content": "1. A simple but powerful pipeline to synthetically generate SfM registration datasets, detailed in Sec. 3." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 383, + 294, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 383, + 294, + 407 + ], + "spans": [ + { + "bbox": [ + 55, + 383, + 294, + 407 + ], + "type": "text", + "content": "2. An improved model for SfM Registration, detailed in Sec. 4." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 407, + 294, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 407, + 294, + 431 + ], + "spans": [ + { + "bbox": [ + 55, + 407, + 294, + 431 + ], + "type": "text", + "content": "3. The task of SfM point cloud registration, with extensive experiments in Sec. 5 demonstrating its potential." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 442, + 142, + 454 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 442, + 142, + 454 + ], + "spans": [ + { + "bbox": [ + 55, + 442, + 142, + 454 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 462, + 295, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 462, + 295, + 629 + ], + "spans": [ + { + "bbox": [ + 55, + 462, + 295, + 629 + ], + "type": "text", + "content": "Point cloud registration is typically defined as estimating the rigid transform that aligns two 3D point clouds to a joint reference frame. The classic approach is the Iterative Closest Point (ICP) algorithm [3, 6], which estimates the pose by alternating between computing point correspondences using the current pose and estimating the pose using those correspondences. Several improvements have been proposed over the years to handle outliers and to attempt to find a global minimum [29, 60]. More recently, point cloud registration has been addressed with neural networks, such as DCP [56] and its iterative extension [57], with similar methods presented in [19, 62]. Another class of iterative methods [1, 25, 59], extract global features from each point cloud, and exploits them to regress the pose." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 630, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 295, + 714 + ], + "type": "text", + "content": "An alternative approach, which will be our main focus, consists of finding and extracting correspondences between the two point clouds, being those points [37, 45], planes and lines [36], and finding the pose using robust estimators such as RANSAC [18]. Several point correspondence methods have been proposed, both handcrafted [16, 41] and learned [7, 14, 21]. In [24] an overlap-attention block is" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 553, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 193 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 193 + ], + "type": "text", + "content": "proposed to improve matching of point clouds with low overlap. Overlap-aware methods are also presented in [26, 39, 63], which focus on improving the matching accuracy and thus avoid the use of RANSAC to filter outliers. The previous methods are not robust to large rotations, which are common in SfM. This issue has been tackled by proposing methods that are rotation invariant [4, 13, 64]. In this paper we will mainly consider point pair features (PPFs) [16] to ensure invariance to global rigid transformations, which have recently shown SotA performance [64]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 195, + 553, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 195, + 553, + 447 + ], + "spans": [ + { + "bbox": [ + 313, + 195, + 553, + 447 + ], + "type": "text", + "content": "Localization without descriptors has recently received renewed attention as a way to reduce computation, and preserve privacy. BPnPNet [5] jointly estimates the pose and the correspondences between 2D keypoints on a query image and 3D points from a scene point cloud. However, it is susceptible to outlier correspondences. This issue was tackled by Zhou et al. [67], who use only 2D keypoints and the 3D map point cloud to localize a query image to a 3D map. They do this by projecting the map point cloud into retrieved map images and estimating 2D-2D correspondences with a graph neural network. A similar method DGC-GNN [55] proposes to exploit geometry and color information to guide point matching by computing a geometric embedding of Euclidean and angular relations. These methods have the advantage of not requiring storage of descriptors. However, they still require image retrieval to get sufficiently close map images to the query. This is problematic if the map is sparse in the region of the query, and introduces additional complexity. Furthermore, the projection from 3D into 2D loses important geometrical information, which we retain by directly registering the point clouds." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 450, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 450, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 450, + 555, + 715 + ], + "type": "text", + "content": "Collaborative mapping is a general term, typically referring to fusion of partial 3D reconstructions. Cohen et al. [8] use the Manhattan world assumption [10] to derive geometric constraints on partial reconstructions to enable model fusion. Cohen et al. [9] fuse indoor and outdoor reconstructions by first leveraging images to obtain semantic masks with window positions. They then match the windows between the two models to solve for the pose with RANSAC [18]. This approach requires images to be stored alongside the point clouds which brings additional space constraints [67]. Strecha et al. [48] and Untzelmann et al. [54] merge multiple building reconstructions to a joint map by combining GPS coordinates and known building outlines to construct an alignment optimization objective. This has the downside of restricting the reconstructions to buildings that have previously been mapped and requires geo-tags. Dusmanu et al. [17] investigated collaborative mapping and localization by translating descriptors to a joint embedding space. They consider a closed set of descriptors and learn a common embedding space for them. Collaborative mapping is achieved through embedding the descriptors into the joint space and running feature match-" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6575" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 240 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 240 + ], + "type": "text", + "content": "ing on the joint space. This requires access to all features for all images, which is expensive, and the required compute scales as standard SfM with the number of images. In contrast to these methods, we make far fewer assumptions on the data available. Namely, we do not store any descriptors (neither local nor global), do not make assumptions on the structure of the scenes, and do not use any topological maps or data. Thus, we constrain ourselves to using only the geometry. Closest to our work Liu et al. [32] align line reconstructions and simulate partial reconstructions by adding noise and dropout of lines to a single reconstruction. In contrast we consider point-based reconstructions and propose a more realistic way to generate partial reconstructions through synthetic camera trajectories and retriangulation." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 249, + 296, + 275 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 249, + 296, + 275 + ], + "spans": [ + { + "bbox": [ + 55, + 249, + 296, + 275 + ], + "type": "text", + "content": "3. Pipeline for SfM Registration Dataset Generation" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 283, + 296, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 283, + 296, + 415 + ], + "spans": [ + { + "bbox": [ + 55, + 283, + 296, + 415 + ], + "type": "text", + "content": "We find that current SotA point cloud registration models trained on 3DMatch [65], the most commonly used training set for 3D registration, are not able to register SfM reconstruction out of the box (cf. Tabs. 1, 2 and 4). Intuitively this is due to the large distribution shift between registration of RGB-D and/or LiDAR scans, and the much larger scenes covered by SfM reconstructions. However, it is not obvious how to generate a diverse dataset with multiple aligned reconstructions. Manually collecting such datasets is prohibitively expensive, and publicly available datasets typically contain a single reconstruction per scene." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 415, + 296, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 415, + 296, + 510 + ], + "spans": [ + { + "bbox": [ + 55, + 415, + 296, + 510 + ], + "type": "text", + "content": "To tackle this issue, we propose a pipeline to generate a large synthetic dataset for registration between SfM models from random image collection datasets, e.g., MegaDepth [30]. Our key insight is that we can leverage these models and for each scene to construct partial reconstructions by sampling subsets of images/c cameras. An overview of our approach is presented in Fig. 3. In the next sections we describe our approach in more detail." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 517, + 233, + 529 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 517, + 233, + 529 + ], + "spans": [ + { + "bbox": [ + 55, + 517, + 233, + 529 + ], + "type": "text", + "content": "3.1. Creating Partial Reconstructions" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 534, + 296, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 296, + 582 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 296, + 582 + ], + "type": "text", + "content": "We use hloc [42] with COLMAP [44] to create partial reconstructions from scenes. We use SIFT [35] and SOSNet [51] retriangulations for training and evaluation. We consider two approaches to create partial reconstructions." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "content": "Random point sampling: In the simplest approach, we sample a set of 3D points from the scene for each partial reconstruction, and add the images in which they are visible until reaching approximately 200 images. For each scene we create 10 such partial reconstructions. We found that while this approach works well when registering reconstructions from collections of random images, it does not generalize as well to reconstructions from a single camera trajectory, which are common in real-life scenarios, e.g., video. We believe that this is due to the point clouds exhibiting more significant variance when the scene is cap-" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": "tured from different viewpoints. For example, keypoints are likely to be denser for nearby cameras, and natural occlusions will depend on viewpoints. As random image based datasets are not naturally divisible into videos, it is not obvious how to bridge this gap. We next describe our approach addressing this." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 144, + 555, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 144, + 555, + 337 + ], + "spans": [ + { + "bbox": [ + 313, + 144, + 555, + 337 + ], + "type": "text", + "content": "Partial Trajectories: Our approach to bridge the random image " + }, + { + "bbox": [ + 313, + 144, + 555, + 337 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 313, + 144, + 555, + 337 + ], + "type": "text", + "content": " video gap is to generate synthetic trajectories from reconstructions of random images. Concretely, we randomly pick a starting image, and then sequentially (without replacement) pick the nearest neighbour of the image using a distance metric (which we refer to as dist). In practice we found that, for each trajectory, using a randomly weighted combination of geodesic rotation distance and Euclidean distance on the translation produced satisfactory results. Each trajectory has a size between " + }, + { + "bbox": [ + 313, + 144, + 555, + 337 + ], + "type": "inline_equation", + "content": "n_{\\mathrm{low}} = 75" + }, + { + "bbox": [ + 313, + 144, + 555, + 337 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 144, + 555, + 337 + ], + "type": "inline_equation", + "content": "n_{\\mathrm{high}} = 300" + }, + { + "bbox": [ + 313, + 144, + 555, + 337 + ], + "type": "text", + "content": ". The algorithm is presented in Algorithm 1. We show a qualitative example of three generated synthetic trajectories in Fig. 4. Running this on all scenes gave us between 5 to 20 synthetic trajectories per scene. Training is conducted on an equally weighted combination of the random point sampling approach and the synthetic trajectories." + } + ] + } + ], + "index": 8 + }, + { + "type": "code", + "bbox": [ + 315, + 384, + 554, + 588 + ], + "blocks": [ + { + "bbox": [ + 315, + 357, + 555, + 381 + ], + "lines": [ + { + "bbox": [ + 315, + 357, + 555, + 381 + ], + "spans": [ + { + "bbox": [ + 315, + 357, + 555, + 381 + ], + "type": "text", + "content": "Algorithm 1 Generation of synthetic trajectory from a set of posed images" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 315, + 384, + 554, + 588 + ], + "lines": [ + { + "bbox": [ + 315, + 384, + 554, + 588 + ], + "spans": [ + { + "bbox": [ + 315, + 384, + 554, + 588 + ], + "type": "text", + "content": "Precondition: " + }, + { + "bbox": [ + 315, + 384, + 554, + 588 + ], + "type": "inline_equation", + "content": "\\mathbb{T}" + }, + { + "bbox": [ + 315, + 384, + 554, + 588 + ], + "type": "text", + "content": " set of posed images, " + }, + { + "bbox": [ + 315, + 384, + 554, + 588 + ], + "type": "inline_equation", + "content": "\\mathbb{I}" + }, + { + "bbox": [ + 315, + 384, + 554, + 588 + ], + "type": "text", + "content": " remaining index set, distance function dist \n1: function GENERATETRAJECTORY(T, I) \n2: " + }, + { + "bbox": [ + 315, + 384, + 554, + 588 + ], + "type": "inline_equation", + "content": "\\tau \\gets \\emptyset" + }, + { + "bbox": [ + 315, + 384, + 554, + 588 + ], + "type": "text", + "content": " \n3: i " + }, + { + "bbox": [ + 315, + 384, + 554, + 588 + ], + "type": "inline_equation", + "content": "\\leftarrow" + }, + { + "bbox": [ + 315, + 384, + 554, + 588 + ], + "type": "text", + "content": " Random sample from I \n4: n " + }, + { + "bbox": [ + 315, + 384, + 554, + 588 + ], + "type": "inline_equation", + "content": "\\leftarrow" + }, + { + "bbox": [ + 315, + 384, + 554, + 588 + ], + "type": "text", + "content": " Random sample from " + }, + { + "bbox": [ + 315, + 384, + 554, + 588 + ], + "type": "inline_equation", + "content": "\\{n_{\\mathrm{low}},\\dots ,n_{\\mathrm{high}}\\}" + }, + { + "bbox": [ + 315, + 384, + 554, + 588 + ], + "type": "text", + "content": " \n5: " + }, + { + "bbox": [ + 315, + 384, + 554, + 588 + ], + "type": "inline_equation", + "content": "\\mathbb{I}\\gets \\mathbb{I}\\backslash i" + }, + { + "bbox": [ + 315, + 384, + 554, + 588 + ], + "type": "text", + "content": " \n6: " + }, + { + "bbox": [ + 315, + 384, + 554, + 588 + ], + "type": "inline_equation", + "content": "\\tau \\gets \\tau \\cup i" + }, + { + "bbox": [ + 315, + 384, + 554, + 588 + ], + "type": "text", + "content": " \n7: for " + }, + { + "bbox": [ + 315, + 384, + 554, + 588 + ], + "type": "inline_equation", + "content": "k\\gets 1" + }, + { + "bbox": [ + 315, + 384, + 554, + 588 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 315, + 384, + 554, + 588 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 315, + 384, + 554, + 588 + ], + "type": "text", + "content": " do \n8: " + }, + { + "bbox": [ + 315, + 384, + 554, + 588 + ], + "type": "inline_equation", + "content": "i^{\\prime}\\gets \\arg \\min_{i^{\\prime}\\in \\mathbb{I}}\\mathrm{dist}(\\mathbb{T}_{i},\\mathbb{T}_{i^{\\prime}})" + }, + { + "bbox": [ + 315, + 384, + 554, + 588 + ], + "type": "text", + "content": " \n9: " + }, + { + "bbox": [ + 315, + 384, + 554, + 588 + ], + "type": "inline_equation", + "content": "\\tau \\gets \\tau \\cup i^{\\prime}" + }, + { + "bbox": [ + 315, + 384, + 554, + 588 + ], + "type": "text", + "content": " \n10: " + }, + { + "bbox": [ + 315, + 384, + 554, + 588 + ], + "type": "inline_equation", + "content": "\\mathbb{I}\\gets \\mathbb{I}\\backslash i^{\\prime}" + }, + { + "bbox": [ + 315, + 384, + 554, + 588 + ], + "type": "text", + "content": " \n11: " + }, + { + "bbox": [ + 315, + 384, + 554, + 588 + ], + "type": "inline_equation", + "content": "i\\gets i^{\\prime}" + }, + { + "bbox": [ + 315, + 384, + 554, + 588 + ], + "type": "text", + "content": " \n12: end for \n13: return " + }, + { + "bbox": [ + 315, + 384, + 554, + 588 + ], + "type": "inline_equation", + "content": "\\tau ,\\mathbb{I}" + }, + { + "bbox": [ + 315, + 384, + 554, + 588 + ], + "type": "text", + "content": " \n14: end function" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "code_body" + } + ], + "index": 10, + "sub_type": "algorithm" + }, + { + "bbox": [ + 314, + 612, + 434, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 612, + 434, + 624 + ], + "spans": [ + { + "bbox": [ + 314, + 612, + 434, + 624 + ], + "type": "text", + "content": "3.2. Pipeline and Dataset" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "type": "text", + "content": "Reconstruction Pipeline: We are interested in having good ground truth correspondences between the partial reconstructions. As the SfM pipeline accuracy usually degrades with fewer available cameras, we use fixed camera poses from the joint reconstruction using all images and simply run triangulation from feature matching. This also has the benefit of being significantly less computationally" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6576" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 106, + 69, + 504, + 222 + ], + "blocks": [ + { + "bbox": [ + 106, + 69, + 504, + 222 + ], + "lines": [ + { + "bbox": [ + 106, + 69, + 504, + 222 + ], + "spans": [ + { + "bbox": [ + 106, + 69, + 504, + 222 + ], + "type": "image", + "image_path": "8c6176577430faea20741dd3474d55eae1e91e84627eb5f3d7ddd10f6a8938ce.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 228, + 555, + 264 + ], + "lines": [ + { + "bbox": [ + 54, + 228, + 555, + 264 + ], + "spans": [ + { + "bbox": [ + 54, + 228, + 555, + 264 + ], + "type": "text", + "content": "Figure 3. Overview of our pipeline. For each scene, from a random image SfM dataset, e.g., MegaDepth [30], we retriangulate partial reconstructions using partial trajectories from the scene. Since these trajectories are in the global SfM reference frame, the relative transformation is simply the identity mapping." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 58, + 280, + 289, + 379 + ], + "blocks": [ + { + "bbox": [ + 58, + 280, + 289, + 379 + ], + "lines": [ + { + "bbox": [ + 58, + 280, + 289, + 379 + ], + "spans": [ + { + "bbox": [ + 58, + 280, + 289, + 379 + ], + "type": "image", + "image_path": "5e26daff9cf077a7f0594b642cfd5aa9131102f539201cfac5a45660b73d5bd7.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 387, + 298, + 466 + ], + "lines": [ + { + "bbox": [ + 54, + 387, + 298, + 466 + ], + "spans": [ + { + "bbox": [ + 54, + 387, + 298, + 466 + ], + "type": "text", + "content": "Figure 4. Example of synthetic trajectories in our proposed dataset. We start with a large scale scene, consisting of hundreds of cameras (left). From this set of cameras we run Algorithm 1 until the remaining set of cameras is small. Plotting the trajectories shows that plausible camera motion is achieved by this procedure (right). Sampling in this way bridges the gap between random image collections and video-based trajectories." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 487, + 253, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 487, + 253, + 499 + ], + "spans": [ + { + "bbox": [ + 55, + 487, + 253, + 499 + ], + "type": "text", + "content": "demanding than running the full reconstructions." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 500, + 296, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 500, + 296, + 632 + ], + "spans": [ + { + "bbox": [ + 55, + 500, + 296, + 632 + ], + "type": "text", + "content": "Dataset Construction: We run our pipeline on all 196 MegaDepth scenes where 10 of them are reserved for the test benchmark. We use either SIFT [35] or SOSNet [51] features for the retriangulation, however any local features could be used. We compute the overlap between all pairs of partial reconstructions. To train models we enforce the overlap to be larger than " + }, + { + "bbox": [ + 55, + 500, + 296, + 632 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 55, + 500, + 296, + 632 + ], + "type": "text", + "content": ". In total, after enforcing sufficient overlap, we are left with about 22000 pairs (" + }, + { + "bbox": [ + 55, + 500, + 296, + 632 + ], + "type": "inline_equation", + "content": "\\sim" + }, + { + "bbox": [ + 55, + 500, + 296, + 632 + ], + "type": "text", + "content": "20000 for training and " + }, + { + "bbox": [ + 55, + 500, + 296, + 632 + ], + "type": "inline_equation", + "content": "\\sim" + }, + { + "bbox": [ + 55, + 500, + 296, + 632 + ], + "type": "text", + "content": "2000 for testing), which constitutes our dataset. Details on how point cloud overlap is computed are presented in the supplementary material." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 644, + 266, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 644, + 266, + 658 + ], + "spans": [ + { + "bbox": [ + 55, + 644, + 266, + 658 + ], + "type": "text", + "content": "4. RoITr and RefineRoITr for ColabSfM" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "content": "In this section, we discuss our baseline and our proposed improvements for ColabSfM. An overview of the proposed registration model " + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "content": " is presented in Figure 5. For consistency we follow the notation of Yu et al. [64]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 282, + 400, + 294 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 282, + 400, + 294 + ], + "spans": [ + { + "bbox": [ + 313, + 282, + 400, + 294 + ], + "type": "text", + "content": "4.1. Preliminaries" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 300, + 555, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 300, + 555, + 421 + ], + "spans": [ + { + "bbox": [ + 313, + 300, + 555, + 421 + ], + "type": "text", + "content": "We are interested in registration methods that are robust to the challenges posed by SfM reconstructions. In particular, one of the main challenges is that SfM reconstructions are defined up-to-scale, and the rotation and origin of the reference frame is arbitrary1. Hence, we would like the registration model to be approximately invariant to the action of the similarity group, i.e., " + }, + { + "bbox": [ + 313, + 300, + 555, + 421 + ], + "type": "inline_equation", + "content": "\\mathrm{Sim}(3) = \\mathrm{SE}(3) + \\mathrm{scale}" + }, + { + "bbox": [ + 313, + 300, + 555, + 421 + ], + "type": "text", + "content": ". In practice we will achieve this by using an SE(3) invariant model, and learn approximate scale invariance. We next describe our baseline, and then our proposed improvements." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 430, + 409, + 441 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 430, + 409, + 441 + ], + "spans": [ + { + "bbox": [ + 313, + 430, + 409, + 441 + ], + "type": "text", + "content": "4.2. Baseline Model" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 448, + 555, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 448, + 555, + 507 + ], + "spans": [ + { + "bbox": [ + 313, + 448, + 555, + 507 + ], + "type": "text", + "content": "As a strong baseline, we base our approach on the SotA SE(3) invariant registration method RoITr [64]. We next give an overview of our baseline. An more thorough description of the architecture is given in the supplementary and by Yu et al. [64]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 508, + 556, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 508, + 556, + 604 + ], + "spans": [ + { + "bbox": [ + 313, + 508, + 556, + 604 + ], + "type": "text", + "content": "Rotation Invariance through PPFs: RoITr [64] is based on point pair features (PPFs) [16], which computes a 4D feature for point pairs by the angles between their relative position and their normals. Mathematically, the PPF is defined as " + }, + { + "bbox": [ + 313, + 508, + 556, + 604 + ], + "type": "inline_equation", + "content": "\\mathrm{PPF}(p_1,p_2,n_1,n_2) =" + }, + { + "bbox": [ + 313, + 508, + 556, + 604 + ], + "type": "inline_equation", + "content": "[\\| d\\| ,\\angle (n_1,d),\\angle (n_2,d),\\angle (n_1,n_2)]" + }, + { + "bbox": [ + 313, + 508, + 556, + 604 + ], + "type": "text", + "content": " , where " + }, + { + "bbox": [ + 313, + 508, + 556, + 604 + ], + "type": "inline_equation", + "content": "d = p_{2} - p_{1}" + }, + { + "bbox": [ + 313, + 508, + 556, + 604 + ], + "type": "text", + "content": " It can easily be shown that this construction is invariant to rotations," + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 323, + 617, + 553, + 661 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 617, + 553, + 661 + ], + "spans": [ + { + "bbox": [ + 323, + 617, + 553, + 661 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\operatorname {P P F} \\left(R p _ {1}, R p _ {2}, R n _ {1}, R n _ {2}\\right) = [ \\| R d \\|, \\angle \\left(R n _ {1}, R d\\right), \\\\ \\angle (R n _ {2}, R d), \\angle (R n _ {1}, R n _ {2}) ] = [ \\| d \\|, \\angle (n _ {1}, d), \\angle (n _ {2}, d), \\\\ \\left. \\angle \\left(n _ {1}, n _ {2}\\right) \\right] = \\operatorname {P P F} \\left(p _ {1}, p _ {2}, n _ {1}, n _ {2}\\right), \\tag {1} \\\\ \\end{array}", + "image_path": "24a4dc08df9cb53b562e291deb27ff645bbd94ee737f6ca471dcf814b94429c1.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 671, + 512, + 682 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 671, + 512, + 682 + ], + "spans": [ + { + "bbox": [ + 313, + 671, + 512, + 682 + ], + "type": "text", + "content": "since norms and angles are invariant to rotations." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 313, + 693, + 555, + 715 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 693, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 693, + 555, + 715 + ], + "type": "text", + "content": "1Although in practice, approximate scale can be obtained by, e.g., standardization, objects with known size, or additional sensors." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6577" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 80, + 77, + 531, + 194 + ], + "blocks": [ + { + "bbox": [ + 80, + 77, + 531, + 194 + ], + "lines": [ + { + "bbox": [ + 80, + 77, + 531, + 194 + ], + "spans": [ + { + "bbox": [ + 80, + 77, + 531, + 194 + ], + "type": "image", + "image_path": "624943ce69130c89c89e4623403e4a9cae83d8a5ef7844f1d3cfc722152ba5d4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 200, + 555, + 289 + ], + "lines": [ + { + "bbox": [ + 54, + 200, + 555, + 289 + ], + "spans": [ + { + "bbox": [ + 54, + 200, + 555, + 289 + ], + "type": "text", + "content": "Figure 5. Overview of our proposed model RefineRoITr. As input we take two point clouds " + }, + { + "bbox": [ + 54, + 200, + 555, + 289 + ], + "type": "inline_equation", + "content": "\\mathcal{P} = (\\mathbf{P},\\mathbf{N},\\mathbf{X})" + }, + { + "bbox": [ + 54, + 200, + 555, + 289 + ], + "type": "inline_equation", + "content": "\\mathcal{Q} = (\\mathbf{Q},\\mathbf{M},\\mathbf{Y})" + }, + { + "bbox": [ + 54, + 200, + 555, + 289 + ], + "type": "text", + "content": " consisting of 3D points " + }, + { + "bbox": [ + 54, + 200, + 555, + 289 + ], + "type": "inline_equation", + "content": "\\mathbf{P} / \\mathbf{Q}\\in \\mathbb{R}^{n / m\\times 3}" + }, + { + "bbox": [ + 54, + 200, + 555, + 289 + ], + "type": "text", + "content": ", and normals " + }, + { + "bbox": [ + 54, + 200, + 555, + 289 + ], + "type": "inline_equation", + "content": "\\mathbf{N} / \\mathbf{M}\\in \\mathbb{R}^{n / m\\times 3}" + }, + { + "bbox": [ + 54, + 200, + 555, + 289 + ], + "type": "text", + "content": ". In this work the features " + }, + { + "bbox": [ + 54, + 200, + 555, + 289 + ], + "type": "inline_equation", + "content": "\\mathbf{X} / \\mathbf{Y}" + }, + { + "bbox": [ + 54, + 200, + 555, + 289 + ], + "type": "text", + "content": " are always assumed to be constant. We ensure rotation invariance by Point Pair Feature (PPF) encoding. The PPFs are fed together with the features into an encoder " + }, + { + "bbox": [ + 54, + 200, + 555, + 289 + ], + "type": "inline_equation", + "content": "e_{\\theta}" + }, + { + "bbox": [ + 54, + 200, + 555, + 289 + ], + "type": "text", + "content": ", producing " + }, + { + "bbox": [ + 54, + 200, + 555, + 289 + ], + "type": "inline_equation", + "content": "N^{\\prime}" + }, + { + "bbox": [ + 54, + 200, + 555, + 289 + ], + "type": "text", + "content": " coarse superpoints " + }, + { + "bbox": [ + 54, + 200, + 555, + 289 + ], + "type": "inline_equation", + "content": "\\mathcal{P}'/\\mathcal{Q}'" + }, + { + "bbox": [ + 54, + 200, + 555, + 289 + ], + "type": "text", + "content": ". These are passed through a global Transformer " + }, + { + "bbox": [ + 54, + 200, + 555, + 289 + ], + "type": "inline_equation", + "content": "g_{\\theta}" + }, + { + "bbox": [ + 54, + 200, + 555, + 289 + ], + "type": "text", + "content": ", from which the top-" + }, + { + "bbox": [ + 54, + 200, + 555, + 289 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 54, + 200, + 555, + 289 + ], + "type": "text", + "content": " coarse correspondences " + }, + { + "bbox": [ + 54, + 200, + 555, + 289 + ], + "type": "inline_equation", + "content": "\\mathcal{C}'\\in \\mathbb{R}^{k\\times 2}" + }, + { + "bbox": [ + 54, + 200, + 555, + 289 + ], + "type": "text", + "content": " are extracted. A decoder " + }, + { + "bbox": [ + 54, + 200, + 555, + 289 + ], + "type": "inline_equation", + "content": "d_{\\theta}" + }, + { + "bbox": [ + 54, + 200, + 555, + 289 + ], + "type": "text", + "content": " takes the coarse features from " + }, + { + "bbox": [ + 54, + 200, + 555, + 289 + ], + "type": "inline_equation", + "content": "g_{\\theta}" + }, + { + "bbox": [ + 54, + 200, + 555, + 289 + ], + "type": "text", + "content": " and the finer features from " + }, + { + "bbox": [ + 54, + 200, + 555, + 289 + ], + "type": "inline_equation", + "content": "e_{\\theta}" + }, + { + "bbox": [ + 54, + 200, + 555, + 289 + ], + "type": "text", + "content": " and produces fine features " + }, + { + "bbox": [ + 54, + 200, + 555, + 289 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{X}},\\hat{\\mathbf{Y}}" + }, + { + "bbox": [ + 54, + 200, + 555, + 289 + ], + "type": "text", + "content": ". Using the coarse correspondences we extract neighbourhoods " + }, + { + "bbox": [ + 54, + 200, + 555, + 289 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{G}}^X,\\hat{\\mathbf{G}}^Y\\in \\mathbb{R}^{k\\times 64\\times c}" + }, + { + "bbox": [ + 54, + 200, + 555, + 289 + ], + "type": "text", + "content": ", which we feed into our proposed refinement Transformer. The refined features from the Transformer are used to construct a cost matrix, where the Sinkhorn algorithm [43, 54, 64] is used to solve the optimal transport (OT) problem, producing our final matches " + }, + { + "bbox": [ + 54, + 200, + 555, + 289 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 54, + 200, + 555, + 289 + ], + "type": "text", + "content": ". The network is trained using " + }, + { + "bbox": [ + 54, + 200, + 555, + 289 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_s + \\mathcal{L}_p" + }, + { + "bbox": [ + 54, + 200, + 555, + 289 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 309, + 294, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 309, + 294, + 429 + ], + "spans": [ + { + "bbox": [ + 54, + 309, + 294, + 429 + ], + "type": "text", + "content": "Normalization Procedure: SfM reconstructions do not in general come with metric scale. This poses several challenges for registration methods, that typically assume both point clouds to share the same scale. We tackle this by normalization. Specifically, for Sim(3) training we normalize the point clouds " + }, + { + "bbox": [ + 54, + 309, + 294, + 429 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 54, + 309, + 294, + 429 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 54, + 309, + 294, + 429 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}" + }, + { + "bbox": [ + 54, + 309, + 294, + 429 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 54, + 309, + 294, + 429 + ], + "type": "inline_equation", + "content": "\\sigma_{\\mathbf{P}} / \\sqrt{2}" + }, + { + "bbox": [ + 54, + 309, + 294, + 429 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 54, + 309, + 294, + 429 + ], + "type": "inline_equation", + "content": "\\sigma_{\\mathbf{Q}} / \\sqrt{2}" + }, + { + "bbox": [ + 54, + 309, + 294, + 429 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 54, + 309, + 294, + 429 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 54, + 309, + 294, + 429 + ], + "type": "text", + "content": " is the largest singular value of the centered point cloud. For SE(3) training, we simply use " + }, + { + "bbox": [ + 54, + 309, + 294, + 429 + ], + "type": "inline_equation", + "content": "\\sigma_{\\mathbf{P}}" + }, + { + "bbox": [ + 54, + 309, + 294, + 429 + ], + "type": "text", + "content": " to normalize both " + }, + { + "bbox": [ + 54, + 309, + 294, + 429 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 54, + 309, + 294, + 429 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 309, + 294, + 429 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}" + }, + { + "bbox": [ + 54, + 309, + 294, + 429 + ], + "type": "text", + "content": ". We find that this simple approach leads to models robust to scale variation, as we demonstrate in Tab. 1." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 432, + 295, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 432, + 295, + 575 + ], + "spans": [ + { + "bbox": [ + 54, + 432, + 295, + 575 + ], + "type": "text", + "content": "Normal Estimation: We follow RoITr [64] and use Open3D [68] with a neighborhood size of 33 for estimating normals, we found that this works well in practice. In contrast to the RGB-D scans of 3DMatch, our dataset consists of point clouds from multiple sensors. This means that the simple approach of aligning the normal orientation towards the origin of the coordinate system, as done in RoITr, will not be consistent in general. To remedy this, we use the fact that all cameras observing a 3D track will yield consistent orientation estimates. In practice, we align the orientation towards the camera center of a random image observing the track, which we found yields consistent orientation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 578, + 295, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 578, + 295, + 685 + ], + "spans": [ + { + "bbox": [ + 54, + 578, + 295, + 685 + ], + "type": "text", + "content": "Overview of Architecture: RoITr consists of an encoder " + }, + { + "bbox": [ + 54, + 578, + 295, + 685 + ], + "type": "inline_equation", + "content": "e_{\\theta}" + }, + { + "bbox": [ + 54, + 578, + 295, + 685 + ], + "type": "text", + "content": ", a global Transformer " + }, + { + "bbox": [ + 54, + 578, + 295, + 685 + ], + "type": "inline_equation", + "content": "g_{\\theta}" + }, + { + "bbox": [ + 54, + 578, + 295, + 685 + ], + "type": "text", + "content": ", and a decoder " + }, + { + "bbox": [ + 54, + 578, + 295, + 685 + ], + "type": "inline_equation", + "content": "d_{\\theta}" + }, + { + "bbox": [ + 54, + 578, + 295, + 685 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 54, + 578, + 295, + 685 + ], + "type": "inline_equation", + "content": "e_{\\theta}" + }, + { + "bbox": [ + 54, + 578, + 295, + 685 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 578, + 295, + 685 + ], + "type": "inline_equation", + "content": "d_{\\theta}" + }, + { + "bbox": [ + 54, + 578, + 295, + 685 + ], + "type": "text", + "content": " take as input sets of PPF encoded local neighbourhoods (typically around 8/16 points) around each point/superpoint. In contrast, " + }, + { + "bbox": [ + 54, + 578, + 295, + 685 + ], + "type": "inline_equation", + "content": "g_{\\theta}" + }, + { + "bbox": [ + 54, + 578, + 295, + 685 + ], + "type": "text", + "content": " performs global attention on the coarse point cloud " + }, + { + "bbox": [ + 54, + 578, + 295, + 685 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{P}}" + }, + { + "bbox": [ + 54, + 578, + 295, + 685 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 54, + 578, + 295, + 685 + ], + "type": "inline_equation", + "content": "e_{\\theta}, d_{\\theta}, g_{\\theta}" + }, + { + "bbox": [ + 54, + 578, + 295, + 685 + ], + "type": "text", + "content": " all use the Transformer architectures as described by Yu et al. [64]. Note that RoITr does not include the refining Transformer " + }, + { + "bbox": [ + 54, + 578, + 295, + 685 + ], + "type": "inline_equation", + "content": "r_{\\theta}" + }, + { + "bbox": [ + 54, + 578, + 295, + 685 + ], + "type": "text", + "content": ", which will be detailed in our RefineRoITr." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 689, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 689, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 689, + 295, + 714 + ], + "type": "text", + "content": "Losses: RoITr uses a superpoint matching loss " + }, + { + "bbox": [ + 54, + 689, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_s" + }, + { + "bbox": [ + 54, + 689, + 295, + 714 + ], + "type": "text", + "content": " and a point matching " + }, + { + "bbox": [ + 54, + 689, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_p" + }, + { + "bbox": [ + 54, + 689, + 295, + 714 + ], + "type": "text", + "content": " loss for training. We follow RoITr" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 309, + 553, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 309, + 553, + 357 + ], + "spans": [ + { + "bbox": [ + 313, + 309, + 553, + 357 + ], + "type": "text", + "content": "and use their overlap-aware [39] circle loss [49] as " + }, + { + "bbox": [ + 313, + 309, + 553, + 357 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_s" + }, + { + "bbox": [ + 313, + 309, + 553, + 357 + ], + "type": "text", + "content": ". The point matching loss " + }, + { + "bbox": [ + 313, + 309, + 553, + 357 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_p" + }, + { + "bbox": [ + 313, + 309, + 553, + 357 + ], + "type": "text", + "content": " is the Negative Log Likelihood (NLL) of the GT correspondences after running the Sinkhorn algorithm [43] on the fine feature similarities " + }, + { + "bbox": [ + 313, + 309, + 553, + 357 + ], + "type": "inline_equation", + "content": "\\tilde{S}" + }, + { + "bbox": [ + 313, + 309, + 553, + 357 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 365, + 430, + 376 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 365, + 430, + 376 + ], + "spans": [ + { + "bbox": [ + 313, + 365, + 430, + 376 + ], + "type": "text", + "content": "4.3. RefineRoITr (Ours)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 382, + 554, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 382, + 554, + 501 + ], + "spans": [ + { + "bbox": [ + 313, + 382, + 554, + 501 + ], + "type": "text", + "content": "Refinement Transformer " + }, + { + "bbox": [ + 313, + 382, + 554, + 501 + ], + "type": "inline_equation", + "content": "r_{\\theta}" + }, + { + "bbox": [ + 313, + 382, + 554, + 501 + ], + "type": "text", + "content": ": The refinement in RoITr is performed by optimizing the Sinkhorn [54] distance between shallow features in a local neighbourhood around the coarse matches, denoted " + }, + { + "bbox": [ + 313, + 382, + 554, + 501 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{G}}^X, \\hat{\\mathbf{G}}^Y" + }, + { + "bbox": [ + 313, + 382, + 554, + 501 + ], + "type": "text", + "content": ". While this provides satisfactory results on 3DMatch, we found that we could improve results by incorporating a joint local Transformer " + }, + { + "bbox": [ + 313, + 382, + 554, + 501 + ], + "type": "inline_equation", + "content": "r_{\\theta}" + }, + { + "bbox": [ + 313, + 382, + 554, + 501 + ], + "type": "text", + "content": ". This Transformer takes " + }, + { + "bbox": [ + 313, + 382, + 554, + 501 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{G}}^X, \\hat{\\mathbf{G}}^Y" + }, + { + "bbox": [ + 313, + 382, + 554, + 501 + ], + "type": "text", + "content": " as inputs and outputs enriched neighbourhoods " + }, + { + "bbox": [ + 313, + 382, + 554, + 501 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{G}}^X, \\hat{\\mathbf{G}}^Y" + }, + { + "bbox": [ + 313, + 382, + 554, + 501 + ], + "type": "text", + "content": ". We found that these cross-point cloud enriched fine features from " + }, + { + "bbox": [ + 313, + 382, + 554, + 501 + ], + "type": "inline_equation", + "content": "r_{\\theta}" + }, + { + "bbox": [ + 313, + 382, + 554, + 501 + ], + "type": "text", + "content": " yielded improved results." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 502, + 555, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 502, + 555, + 693 + ], + "spans": [ + { + "bbox": [ + 313, + 502, + 555, + 693 + ], + "type": "text", + "content": "Architectural Details of " + }, + { + "bbox": [ + 313, + 502, + 555, + 693 + ], + "type": "inline_equation", + "content": "r_{\\theta}" + }, + { + "bbox": [ + 313, + 502, + 555, + 693 + ], + "type": "text", + "content": ": " + }, + { + "bbox": [ + 313, + 502, + 555, + 693 + ], + "type": "inline_equation", + "content": "r_{\\theta}" + }, + { + "bbox": [ + 313, + 502, + 555, + 693 + ], + "type": "text", + "content": " takes in local neighbourhoods of shape " + }, + { + "bbox": [ + 313, + 502, + 555, + 693 + ], + "type": "inline_equation", + "content": "m \\times 64 \\times c" + }, + { + "bbox": [ + 313, + 502, + 555, + 693 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 502, + 555, + 693 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 502, + 555, + 693 + ], + "type": "text", + "content": " is the number of neighbourhoods, 64 is the number of neighbours (identical to RoITr), and " + }, + { + "bbox": [ + 313, + 502, + 555, + 693 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 502, + 555, + 693 + ], + "type": "text", + "content": " is the feature dimensionality. We use layers of interleaved self-attention and cross-attention as in LightGlue [31]. Note however, that in contrast to LightGlue that performs global attention, our refinement Transformer performs attention locally, which is significantly cheaper. We use an input/output dimensionality of " + }, + { + "bbox": [ + 313, + 502, + 555, + 693 + ], + "type": "inline_equation", + "content": "c_{\\mathrm{in}} = c_{\\mathrm{out}} = c = 64" + }, + { + "bbox": [ + 313, + 502, + 555, + 693 + ], + "type": "text", + "content": ", and 4 self+cross attention layers. Each self/cross attention layer consists of single head attention, using dim " + }, + { + "bbox": [ + 313, + 502, + 555, + 693 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 502, + 555, + 693 + ], + "type": "text", + "content": ", which is stacked with the input into a 1 hidden layer MLP (input dim " + }, + { + "bbox": [ + 313, + 502, + 555, + 693 + ], + "type": "inline_equation", + "content": "2c" + }, + { + "bbox": [ + 313, + 502, + 555, + 693 + ], + "type": "text", + "content": ", hidden dim " + }, + { + "bbox": [ + 313, + 502, + 555, + 693 + ], + "type": "inline_equation", + "content": "2c" + }, + { + "bbox": [ + 313, + 502, + 555, + 693 + ], + "type": "text", + "content": ", output dim " + }, + { + "bbox": [ + 313, + 502, + 555, + 693 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 502, + 555, + 693 + ], + "type": "text", + "content": ") using the GELU [23] activation function, with layer normalization [2] applied before GELU. Each self/cross block is residual, i.e., the output of the block is the output of the MLP plus the input." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6578" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 156, + 294, + 266 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 295, + 148 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 295, + 148 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 295, + 148 + ], + "type": "text", + "content": "Table 1. Results of SfM registration on MegaDepth. We evaluate 2 different scenarios. Unknown relative orientation SO(3) + unknown relative position = SE(3), SE(3) + unknown relative scale = Sim(3). The top portion contains methods only trained on the 3DMatch (3DM) dataset, the middle portion methods trained only on our proposed dataset (Mega), while the lower portion contains methods trained on the former and fine-tuned on the latter." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 156, + 294, + 266 + ], + "lines": [ + { + "bbox": [ + 56, + 156, + 294, + 266 + ], + "spans": [ + { + "bbox": [ + 56, + 156, + 294, + 266 + ], + "type": "table", + "html": "
MethodSE(3)Sim(3)
IRFMRRRIRFMRRR
OverlapPredator [24] (3DM)6.135.510.03.621.32.1
GeoTransformer [39] (3DM)2.114.63.31.38.60.4
PareNet [61] (3DM)7.338.65.24.727.70.8
RoITr [64] (3DM)3.012.60.01.67.00.8
RefineRoITr (3DM)10.032.813.96.222.83.5
RefineRoITr (Mega)51.096.570.244.692.842.7
OverlapPredator [24] (3DM + Mega)21.374.935.411.456.210.7
RoITr [64] (3DM + Mega)44.690.960.038.486.138.7
RefineRoITr (3DM + Mega)48.795.167.744.693.944.3
", + "image_path": "d32433905fba16354b8e2537aafce4887ca5a4fe98694837d0a270a7181d5491.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 300, + 136, + 315 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 300, + 136, + 315 + ], + "spans": [ + { + "bbox": [ + 55, + 300, + 136, + 315 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 327, + 295, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 327, + 295, + 472 + ], + "spans": [ + { + "bbox": [ + 54, + 327, + 295, + 472 + ], + "type": "text", + "content": "In this section, we evaluate the RefineRoITr model presented in Sec. 4. Three different model instances were trained: 1) pre-trained RefineRoITr on 3DMatch [65]; 2) the model in 1) fine-tuned on the dataset in Sec. 3 and; 3) RefineRoITr trained from scratch on our dataset. For comparison, we used RoITr [64] and OverlapPredator [24], both with pre-trained weights from 3DMatch, and fine-tuned weights trained on our dataset. Additionally, we tested GeoTransformer [39] also pre-trained on 3DMatch [65]. As we found that the performance, of fine-tuning 3DMatch trained models, on our dataset was similar for our proposed model, we did not train other models in our dataset from scratch." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 479, + 295, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 479, + 295, + 635 + ], + "spans": [ + { + "bbox": [ + 55, + 479, + 295, + 635 + ], + "type": "text", + "content": "Metrics: We use the standard registration metric Inlier Ratio (IR). This measures the percentage of estimated correspondences lying within a threshold distance " + }, + { + "bbox": [ + 55, + 479, + 295, + 635 + ], + "type": "inline_equation", + "content": "\\tau_{\\mathrm{IR}} = 0.1" + }, + { + "bbox": [ + 55, + 479, + 295, + 635 + ], + "type": "text", + "content": " from each other. We additionally measure the Feature Matching Recall (FMR), which is defined as the percentage of point cloud pairs whose IR is above a threshold " + }, + { + "bbox": [ + 55, + 479, + 295, + 635 + ], + "type": "inline_equation", + "content": "\\tau_{\\mathrm{FMR}} = 5\\%" + }, + { + "bbox": [ + 55, + 479, + 295, + 635 + ], + "type": "text", + "content": ". Finally, we use Registration Recall (RR), which consists of the percentage of pairs for which the rotation error and translation error are below " + }, + { + "bbox": [ + 55, + 479, + 295, + 635 + ], + "type": "inline_equation", + "content": "5^{\\circ}" + }, + { + "bbox": [ + 55, + 479, + 295, + 635 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 479, + 295, + 635 + ], + "type": "inline_equation", + "content": "0.05" + }, + { + "bbox": [ + 55, + 479, + 295, + 635 + ], + "type": "text", + "content": ", respectively. We solve for the pose using the RANSAC [18] implementation from Open3D [68] on the matches retrieved by the neural networks. Details on error computation and RANSAC can be found in the Supplemental material." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 641, + 295, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 295, + 712 + ], + "type": "text", + "content": "Computational cost: On a desktop machine equipped with an i9-13900K @ 3.00GHz CPU and an NVIDIA 4090 GPU, the median runtime for RoITr is 99.94 ms and for RefineRoITris 102.84, which shows that the computational burden added by the modifications described in Sec. 4.3 is around " + }, + { + "bbox": [ + 55, + 641, + 295, + 712 + ], + "type": "inline_equation", + "content": "3\\%" + }, + { + "bbox": [ + 55, + 641, + 295, + 712 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 533, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 533, + 85 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 533, + 85 + ], + "type": "text", + "content": "5.1. MegaDepth SfM Registration Benchmark" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 312, + 90, + 553, + 210 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 90, + 553, + 210 + ], + "spans": [ + { + "bbox": [ + 312, + 90, + 553, + 210 + ], + "type": "text", + "content": "We select 10 scenes from MegaDepth [30] dataset, which consists of 196 large scale 3D scenes reconstructed from internet images for our benchmark. The test scenes are, in order, [0008, 0015, 0019, 0021, 0022, 0024, 0025, 0032, 0063, 1589]. The average size of the point clouds is " + }, + { + "bbox": [ + 312, + 90, + 553, + 210 + ], + "type": "inline_equation", + "content": "15.3\\mathrm{K}" + }, + { + "bbox": [ + 312, + 90, + 553, + 210 + ], + "type": "text", + "content": " points, using either SIFT or SOSNet features for the triangulation. To train and evaluate the network presented in Sec. 4 on our dataset, we define two different scenarios SE(3) and Sim(3). We define a random rotation by sampling three Euler angles " + }, + { + "bbox": [ + 312, + 90, + 553, + 210 + ], + "type": "inline_equation", + "content": "\\alpha, \\beta, \\gamma \\in [0,2\\pi]" + }, + { + "bbox": [ + 312, + 90, + 553, + 210 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 211, + 555, + 391 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 211, + 555, + 391 + ], + "spans": [ + { + "bbox": [ + 313, + 211, + 555, + 391 + ], + "type": "text", + "content": "The results of the RefineRoITr method and the baselines are presented in Tab. 1. As expected, the methods which are not fine-tuned in our data perform the worst in both SE(3) and Sim(3) data. This shows the lack of generalization of current registration methods on SfM point clouds. When considering FMR, the methods trained from scratch or fine-tuned on our data obtain high scores (" + }, + { + "bbox": [ + 313, + 211, + 555, + 391 + ], + "type": "inline_equation", + "content": ">90\\%" + }, + { + "bbox": [ + 313, + 211, + 555, + 391 + ], + "type": "text", + "content": ") on SE(3) data with slightly lower scores on Sim(3). When looking at the three variations of RefineRoITr, we see that while the version trained on 3DMatch outperforms vanilla RoITr, the major boost in performance is due to fine-tuning the model on our dataset. The best overall version is the one trained from scratch on our dataset, even though the improvement is not as significant. Qualitative results are presented in Fig. 2 and the Supplementary material." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 402, + 553, + 415 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 402, + 553, + 415 + ], + "spans": [ + { + "bbox": [ + 313, + 402, + 553, + 415 + ], + "type": "text", + "content": "5.2. Standard Benchmark: Cambridge Landmarks" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 420, + 555, + 600 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 420, + 555, + 600 + ], + "spans": [ + { + "bbox": [ + 313, + 420, + 555, + 600 + ], + "type": "text", + "content": "The Cambridge Landmarks [28] dataset consists of 5 large-scale 3D scenes recorded with a smartphone by a pedestrian in Cambridge, each of which consisting of a train and a test trajectories with camera poses. We retriangulate the train/test sequences independently, which results in two reconstructions for each scene. In this case, we evaluate the IR and number of matches. Here we only test the SE(3) case. Rotation and translation errors and qualitative results for this dataset are reported in the supplementary material. The results are presented in Tab. 2. Similar to the evaluation results in the previous section, the models trained only on 3DMatch [65] perform poorly, while the ones fine-tuned and or trained from scratch on our dataset present the highest IR and number of matches, with small differences in favor of the version trained from scratch." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 611, + 524, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 611, + 524, + 624 + ], + "spans": [ + { + "bbox": [ + 313, + 611, + 524, + 624 + ], + "type": "text", + "content": "5.3. Noisy Indoor Reconstructions: 7-Scenes" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 629, + 553, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 553, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 553, + 715 + ], + "type": "text", + "content": "The 7-Scenes [46] dataset is an indoor RGB-D dataset. This dataset is challenging for SfM, since textured regions lead to significantly more noisy point clouds, and less geometric features. We present results in Tab. 3. Encouragingly, we find that models trained using our pipeline perform well also on these indoor scenes, showing the generalizability of our approach." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "6579" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 129, + 112, + 479, + 215 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 104 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 104 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 104 + ], + "type": "text", + "content": "Table 2. Results of SfM registration on Cambridge Landmarks [28]. We evaluate unknown relative orientation " + }, + { + "bbox": [ + 55, + 70, + 555, + 104 + ], + "type": "inline_equation", + "content": "\\mathrm{SO}(3) +" + }, + { + "bbox": [ + 55, + 70, + 555, + 104 + ], + "type": "text", + "content": " unknown relative position " + }, + { + "bbox": [ + 55, + 70, + 555, + 104 + ], + "type": "inline_equation", + "content": "= \\mathrm{{SE}}\\left( 3\\right)" + }, + { + "bbox": [ + 55, + 70, + 555, + 104 + ], + "type": "text", + "content": " . The top portion contains methods only trained on the 3DMatch (3DM) dataset, the middle portion methods trained only on our proposed dataset (Mega), while the lower portion contains methods trained on the former and fine-tuned on the latter." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 129, + 112, + 479, + 215 + ], + "lines": [ + { + "bbox": [ + 129, + 112, + 479, + 215 + ], + "spans": [ + { + "bbox": [ + 129, + 112, + 479, + 215 + ], + "type": "table", + "html": "
MethodGreat CourtKings CollegeOld HospitalShop FacadeSt Mary's Church
IRMatchesIRMatchesIRMatchesIRMatchesIRMatches
OverlapPredator [24] (3DM)1.43560.53720.03632.53540.2382
GeoTransformer [39] (3DM)0.03820.17850.04380.113680.0431
PureNet [61] (3DM)4.320001.62000020001.220006.72000
RoiTr [64] (3DM)0.02850.04470.03210.39591.5407
RefineRoiTr (3DM)0.24122.24080.05440.97372.8567
RefineRoiTr (Mega)70.9453857.6397731.5221441.6132581.85167
RoiTr [64] (3DM + Mega)52.1142639.6137721.960328.070064.52481
RefineRoiTr (3DM + Mega)69.4455154.0324628.0159139.1134277.85168
", + "image_path": "b869fec888bc2ba0f26af053082366cbd8a9173a70d33b669767992ce130c00f.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 92, + 266, + 517, + 342 + ], + "blocks": [ + { + "bbox": [ + 55, + 224, + 555, + 258 + ], + "lines": [ + { + "bbox": [ + 55, + 224, + 555, + 258 + ], + "spans": [ + { + "bbox": [ + 55, + 224, + 555, + 258 + ], + "type": "text", + "content": "Table 3. Results of SfM registration on 7-Scenes. We evaluate unknown relative orientation " + }, + { + "bbox": [ + 55, + 224, + 555, + 258 + ], + "type": "inline_equation", + "content": "\\mathrm{SO}(3) +" + }, + { + "bbox": [ + 55, + 224, + 555, + 258 + ], + "type": "text", + "content": " unknown relative position " + }, + { + "bbox": [ + 55, + 224, + 555, + 258 + ], + "type": "inline_equation", + "content": "= \\mathrm{{SE}}\\left( 3\\right)" + }, + { + "bbox": [ + 55, + 224, + 555, + 258 + ], + "type": "text", + "content": " . The top portion contains methods only trained on the 3DMatch (3DM) dataset, the middle portion methods trained only on our proposed dataset (Mega), while the lower portion contains methods trained on the former and fine-tuned on the latter." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 92, + 266, + 517, + 342 + ], + "lines": [ + { + "bbox": [ + 92, + 266, + 517, + 342 + ], + "spans": [ + { + "bbox": [ + 92, + 266, + 517, + 342 + ], + "type": "table", + "html": "
MethodChessFireHeadsOfficePumpkinRed KitchenStairs
IRMatchesIRMatchesIRMatchesIRMatchesIRMatchesIRMatchesIRMatches
RoITr [64] (3DM)0.03070.010040.016990.013220.03270.011630.0730
RefineRoITr (3DM)0.08420.022040.011040.010730.010160.017210.01043
RefineRoITr (Mega)23.4671738.8904350.6541968.0946114.9289327.980506.23230
RoITr [64] (3DM+Mega)20.4209830.5416536.4381250.047527.9127118.524389.11883
RefineRoITr (3DM+Mega)23.6563233.3833950.5599662.788878.9226826.866166.42863
", + "image_path": "1432871e0f8a88c437e493c8814857d0d950cc8d539fa4ef98ade6d799f47c39.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 85, + 434, + 264, + 522 + ], + "blocks": [ + { + "bbox": [ + 55, + 358, + 295, + 425 + ], + "lines": [ + { + "bbox": [ + 55, + 358, + 295, + 425 + ], + "spans": [ + { + "bbox": [ + 55, + 358, + 295, + 425 + ], + "type": "text", + "content": "Table 4. Results of SfM registration on Quad6k [11]. We evaluate unknown relative orientation " + }, + { + "bbox": [ + 55, + 358, + 295, + 425 + ], + "type": "inline_equation", + "content": "\\mathrm{SO}(3) +" + }, + { + "bbox": [ + 55, + 358, + 295, + 425 + ], + "type": "text", + "content": " unknown relative position " + }, + { + "bbox": [ + 55, + 358, + 295, + 425 + ], + "type": "inline_equation", + "content": "=" + }, + { + "bbox": [ + 55, + 358, + 295, + 425 + ], + "type": "text", + "content": " SE(3). The top portion contains methods only trained on the 3DMatch (3DM) dataset, the middle portion methods trained only on our proposed dataset (Mega), while the lower portion contains methods trained on the former and fine-tuned on the latter." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 85, + 434, + 264, + 522 + ], + "lines": [ + { + "bbox": [ + 85, + 434, + 264, + 522 + ], + "spans": [ + { + "bbox": [ + 85, + 434, + 264, + 522 + ], + "type": "table", + "html": "
MethodIRFMRMatchesRR
OverlapPredator [24] (3DM)1.610.21210.0
GeoTransformer [39] (3DM)0.60.019360.0
PareNet [61] (3DM)1.610.420000.0
RoITr [64] (3DM)0.40.05740.0
RefineRoITr (3DM)1.76.16172.0
RefineRoITr (Mega)17.267.487822.4
RoITr [64] (3DM + Mega)14.263.349224.5
RefineRoITr (3DM + Mega)18.165.371516.3
", + "image_path": "f4e78ae21f432629b3d687b8901527829f02b2f93ef38342a4239b77194c1f57.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 548, + 185, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 548, + 185, + 562 + ], + "spans": [ + { + "bbox": [ + 55, + 548, + 185, + 562 + ], + "type": "text", + "content": "5.4. Low Overlap: Quad6k" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "content": "The Quad6k dataset [11] consists of about 6500 images from the Arts Quad at Cornell University. To create partial reconstructions for testing, we first perform a full scene retriangulation using the refined poses provided by [52] and then follow the procedure in Sec. 3.1, generating a total of 24 trajectories, and 49 pairs. The results are presented in Tab. 4. Similar to the previous datasets, the performance of the methods has a substantial boost after training on the dataset presented in Sec. 3. GeoTransformer [39] trained only on 3DMatch [65] presents the most matches on average, but only a small fraction (0.6%) are inliers, and its FMR and RR are both zero. Our Quad6k benchmark is" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 360, + 555, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 360, + 555, + 409 + ], + "spans": [ + { + "bbox": [ + 313, + 360, + 555, + 409 + ], + "type": "text", + "content": "more challenging and thus the performance is lower than in Tabs. 1 and 2, due to a low overlap of " + }, + { + "bbox": [ + 313, + 360, + 555, + 409 + ], + "type": "inline_equation", + "content": "16.5\\%" + }, + { + "bbox": [ + 313, + 360, + 555, + 409 + ], + "type": "text", + "content": ", in comparison with the " + }, + { + "bbox": [ + 313, + 360, + 555, + 409 + ], + "type": "inline_equation", + "content": "61.2\\%" + }, + { + "bbox": [ + 313, + 360, + 555, + 409 + ], + "type": "text", + "content": " of MegaDepth. Qualitative results for this dataset are shown in the supplementary material." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 419, + 388, + 431 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 419, + 388, + 431 + ], + "spans": [ + { + "bbox": [ + 313, + 419, + 388, + 431 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 439, + 555, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 439, + 555, + 654 + ], + "spans": [ + { + "bbox": [ + 313, + 439, + 555, + 654 + ], + "type": "text", + "content": "We have proposed Collaborative Structure-from-Motion by Point Cloud Registration, ColabSfM, a challenging new task to enable scalable collaborative mapping and localization. By extensive experiments we showed that with our proposed dataset generation pipeline and SE(3) invariant model, this task is solvable, and showed that our model, while trained only on MegaDepth, is able to generalize to several challenging scenarios, such as noisy indoor reconstructions, and low-overlap registration. We believe our initial results will encourage further research on collaborative 3D reconstruction. Limitations: a) Our proposed model RefineRoITr is E(3) invariant like RoITr, and therefore can struggle with estimating correct matches in symmetric scenes. b) Our models are trained and evaluated on retriangulated partial reconstructions. In practice, issues such as drift occur, making global alignment more difficult, however we find that our model still performs well in this setting in Table 7." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 666, + 410, + 679 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 666, + 410, + 679 + ], + "spans": [ + { + "bbox": [ + 313, + 666, + 410, + 679 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 685, + 554, + 710 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 685, + 554, + 710 + ], + "spans": [ + { + "bbox": [ + 313, + 685, + 554, + 710 + ], + "type": "text", + "content": "This work was supported by the strategic research environment ELLIIT, funded by the Swedish government." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "6580" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "type": "text", + "content": "[1] Yasuhiro Aoki, Hunter Goforth, Rangaprasad Arun Srivatsan, and Simon Lucey. Pointnetlk: Robust & efficient point cloud registration using pointnet. In IEEE Conf. Comput. Vis. Pattern Recog., pages 7163-7172, 2019. 2, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 136, + 296, + 167 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 136, + 296, + 167 + ], + "spans": [ + { + "bbox": [ + 61, + 136, + 296, + 167 + ], + "type": "text", + "content": "[2] Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. arXiv preprint arXiv:1607.06450, 2016. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 170, + 295, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 170, + 295, + 201 + ], + "spans": [ + { + "bbox": [ + 62, + 170, + 295, + 201 + ], + "type": "text", + "content": "[3] Paul J. Besl and Neil D. McKay. A method for registration of 3-d shapes. IEEE Trans. Pattern Anal. Mach. Intell., 14 (2):239-256, 1992. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 203, + 295, + 245 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 203, + 295, + 245 + ], + "spans": [ + { + "bbox": [ + 62, + 203, + 295, + 245 + ], + "type": "text", + "content": "[4] Georg Bokman and Fredrik Kahl. A case for using rotation invariant features in state of the art feature matchers. In IEEE Conf. Comput. Vis. Pattern Recog., pages 5110-5119, 2022. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 247, + 295, + 290 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 247, + 295, + 290 + ], + "spans": [ + { + "bbox": [ + 62, + 247, + 295, + 290 + ], + "type": "text", + "content": "[5] Dylan Campbell, Liu Liu, and Stephen Gould. Solving the blind perspective-n-point problem end-to-end with robust differentiable geometric optimization. In Eur. Conf. Comput. Vis., pages 244-261, 2020. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 291, + 295, + 324 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 291, + 295, + 324 + ], + "spans": [ + { + "bbox": [ + 62, + 291, + 295, + 324 + ], + "type": "text", + "content": "[6] Yang Chen and Gerard Medioni. Object modeling by registration of multiple range images. In IEEE Int'l Conf. Robotics and Auto., pages 2724-2729, 1991. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 325, + 295, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 325, + 295, + 357 + ], + "spans": [ + { + "bbox": [ + 62, + 325, + 295, + 357 + ], + "type": "text", + "content": "[7] Christopher Choy, Jaesik Park, and Vladlen Koltun. Fully convolutional geometric features. In Int. Conf. Comput. Vis., pages 8958-8966, 2019. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 358, + 295, + 391 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 358, + 295, + 391 + ], + "spans": [ + { + "bbox": [ + 62, + 358, + 295, + 391 + ], + "type": "text", + "content": "[8] Andrea Cohen, Torsten Sattler, and Marc Pollefeys. Merging the unmatched: Stitching visually disconnected sfm models. In Int. Conf. Comput. Vis., pages 2129-2137, 2015. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 392, + 295, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 392, + 295, + 435 + ], + "spans": [ + { + "bbox": [ + 62, + 392, + 295, + 435 + ], + "type": "text", + "content": "[9] Andrea Cohen, Johannes L Schonberger, Pablo Speciale, Torsten Sattler, Jan-Michael Frahm, and Marc Pollefeys. Indoor-outdoor 3d reconstruction alignment. In Eur. Conf. Comput. Vis., pages 285-300, 2016. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 436, + 295, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 436, + 295, + 479 + ], + "spans": [ + { + "bbox": [ + 57, + 436, + 295, + 479 + ], + "type": "text", + "content": "[10] James Coughlan and Alan L Yuille. The manhattan world assumption: Regularities in scene statistics which enable bayesian inference. Adv. Neural Inform. Process. Syst., 13, 2000. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 480, + 295, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 480, + 295, + 524 + ], + "spans": [ + { + "bbox": [ + 57, + 480, + 295, + 524 + ], + "type": "text", + "content": "[11] David Crandall, Andrew Owens, Noah Snavely, and Dan Huttenlocher. Discrete-continuous optimization for large-scale structure from motion. In IEEE Conf. Comput. Vis. Pattern Recog., pages 3001–3008, 2011. 3, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 525, + 295, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 525, + 295, + 579 + ], + "spans": [ + { + "bbox": [ + 57, + 525, + 295, + 579 + ], + "type": "text", + "content": "[12] Deeksha Dangwal, Vincent T Lee, Hyo Jin Kim, Tianwei Shen, Meghan Cowan, Rajvi Shah, Caroline Trippel, Brandon Reagen, Timothy Sherwood, Vasileios Balntas, et al. Mitigating reverse engineering attacks on local feature descriptors. In Brit. Mach. Vis. Conf., 2021. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 580, + 295, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 580, + 295, + 613 + ], + "spans": [ + { + "bbox": [ + 57, + 580, + 295, + 613 + ], + "type": "text", + "content": "[13] Haowen Deng, Tolga Birdal, and Slobodan Ilic. Ppf-foldnet: Unsupervised learning of rotation invariant 3d local descriptors. In Eur. Conf. Comput. Vis., pages 602-618, 2018. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 613, + 295, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 613, + 295, + 656 + ], + "spans": [ + { + "bbox": [ + 57, + 613, + 295, + 656 + ], + "type": "text", + "content": "[14] Haowen Deng, Tolga Birdal, and Slobodan Ilic. Ppfnet: Global context aware local features for robust 3d point matching. In IEEE Conf. Comput. Vis. Pattern Recog., pages 195-205, 2018. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 658, + 295, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 658, + 295, + 691 + ], + "spans": [ + { + "bbox": [ + 57, + 658, + 295, + 691 + ], + "type": "text", + "content": "[15] Alexey Dosovitskiy and Thomas Brox. Inverting visual representations with convolutional networks. In IEEE Conf. Comput. Vis. Pattern Recog., pages 4829-4837, 2016. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 57, + 692, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 692, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 692, + 295, + 713 + ], + "type": "text", + "content": "[16] Bertram Drost, Markus Ulrich, Nassir Navab, and Slobodan Ilic. Model globally, match locally: Efficient and robust" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "text", + "content": "3d object recognition. In IEEE Conf. Comput. Vis. Pattern Recog., pages 998-1005, 2010. 3, 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "type": "text", + "content": "[17] Mihai Dusmanu, Ondrej Miksik, Johannes L Schonberger, and Marc Pollefeys. Cross-descriptor visual localization and mapping. In Int. Conf. Comput. Vis., pages 6058-6067, 2021. 2, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 141, + 553, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 553, + 185 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 553, + 185 + ], + "type": "text", + "content": "[18] Martin A Fischler and Robert C Bolles. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM, 24(6):381-395, 1981. 3, 7, 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 186, + 553, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 186, + 553, + 230 + ], + "spans": [ + { + "bbox": [ + 317, + 186, + 553, + 230 + ], + "type": "text", + "content": "[19] Kexue Fu, Shaolei Liu, Xiaoyuan Luo, and Manning Wang. Robust point cloud registration framework based on deep graph matching. In IEEE Conf. Comput. Vis. Pattern Recog., pages 8893-8902, 2021. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 232, + 553, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 232, + 553, + 274 + ], + "spans": [ + { + "bbox": [ + 317, + 232, + 553, + 274 + ], + "type": "text", + "content": "[20] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In IEEE Conf. Comput. Vis. Pattern Recog., pages 3354-3361, 2012. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 276, + 553, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 276, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 317, + 276, + 553, + 319 + ], + "type": "text", + "content": "[21] Zan Gojcic, Caifa Zhou, Jan D Wegner, and Andreas Wieser. The perfect match: 3d point cloud matching with smoothed densities. In IEEE Conf. Comput. Vis. Pattern Recog., pages 5545-5554, 2019. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 321, + 553, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 321, + 553, + 354 + ], + "spans": [ + { + "bbox": [ + 317, + 321, + 553, + 354 + ], + "type": "text", + "content": "[22] Michal Havlena, Akihiko Torii, and Tomáš Pajdla. Efficient structure from motion by graph optimization. In Eur. Conf. Comput. Vis., pages 100-113, 2010. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 355, + 553, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 355, + 553, + 376 + ], + "spans": [ + { + "bbox": [ + 316, + 355, + 553, + 376 + ], + "type": "text", + "content": "[23] Dan Hendrycks and Kevin Gimpel. Gaussian error linear units (gelus). arXiv preprint arXiv:1606.08415, 2016. 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 378, + 553, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 378, + 553, + 422 + ], + "spans": [ + { + "bbox": [ + 317, + 378, + 553, + 422 + ], + "type": "text", + "content": "[24] Shengyu Huang, Zan Gojcic, Mikhail Usvyatsov, Andreas Wieser, and Konrad Schindler. Predator: Registration of 3d point clouds with low overlap. In IEEE Conf. Comput. Vis. Pattern Recog., pages 4267-4276, 2021. 2, 3, 7, 8, 1, 4, 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 423, + 553, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 423, + 553, + 477 + ], + "spans": [ + { + "bbox": [ + 317, + 423, + 553, + 477 + ], + "type": "text", + "content": "[25] Xiaoshui Huang, Guofeng Mei, and Jian Zhang. Feature-metric registration: A fast semi-supervised approach for robust point cloud registration without correspondences. In IEEE Conf. Comput. Vis. Pattern Recog., pages 11366-11374, 2020. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 479, + 553, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 479, + 553, + 522 + ], + "spans": [ + { + "bbox": [ + 317, + 479, + 553, + 522 + ], + "type": "text", + "content": "[26] Shengze Jin, Iro Armeni, Marc Pollefeys, and Daniel Barath. Multiway point cloud mosaicking with diffusion and global optimization. In IEEE Conf. Comput. Vis. Pattern Recog., pages 20838-20849, 2024. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 317, + 523, + 553, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 523, + 553, + 590 + ], + "spans": [ + { + "bbox": [ + 317, + 523, + 553, + 590 + ], + "type": "text", + "content": "[27] Neena Kamath. Announcing azure spatial anchors for collaborative, cross-platform mixed reality apps, 2019. URL https://azure.microsoft.com/en-us/blog/announcing-azure-spatial-anchors-for-collaborative-cross-platform-mixed-reality-apps/.1" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 317, + 590, + 553, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 590, + 553, + 634 + ], + "spans": [ + { + "bbox": [ + 317, + 590, + 553, + 634 + ], + "type": "text", + "content": "[28] Alex Kendall, Matthew Grimes, and Roberto Cipolla. Posenet: A convolutional network for real-time 6-dof camera relocalization. In Int. Conf. Comput. Vis., pages 2938-2946, 2015. 3, 7, 8, 2, 6" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 317, + 635, + 553, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 635, + 553, + 667 + ], + "spans": [ + { + "bbox": [ + 317, + 635, + 553, + 667 + ], + "type": "text", + "content": "[29] Hongdong Li and Richard Hartley. The 3d-3d registration problem revisited. In Int. Conf. Comput. Vis., pages 1-8, 2007. 3" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 316, + 669, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 669, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 669, + 553, + 712 + ], + "type": "text", + "content": "[30] Zhengqi Li and Noah Snavely. Megadepth: Learning single-view depth prediction from internet photos. In IEEE Conf. Comput. Vis. Pattern Recog., pages 2041-2050, 2018. 3, 4, 5, 7" + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "text", + "content": "6581" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "type": "text", + "content": "[31] Philipp Lindenberger, Paul-Edouard Sarlin, and Marc Pollefeys. LightGlue: Local Feature Matching at Light Speed. In Int. Conf. Comput. Vis., 2023. 2, 6, 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 106, + 295, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 106, + 295, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 106, + 295, + 150 + ], + "type": "text", + "content": "[32] Liu Liu, Hongdong Li, Haodong Yao, and Ruyi Zha. Pluckernet: Learn to register 3d line reconstructions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1842-1852, 2021. 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 150, + 294, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 150, + 294, + 183 + ], + "spans": [ + { + "bbox": [ + 56, + 150, + 294, + 183 + ], + "type": "text", + "content": "[33] Alex Locher, Michal Havlena, and Luc Van Gool. Progressive structure from motion. In Eur. Conf. Comput. Vis., pages 20-35, 2018. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 183, + 294, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 183, + 294, + 216 + ], + "spans": [ + { + "bbox": [ + 56, + 183, + 294, + 216 + ], + "type": "text", + "content": "[34] Alexandre Lopes, Roberto Souza, and Helio Pedrini. A survey on rgb-d datasets. Comput. Vis. Img. Underst., 222: 103489, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 217, + 294, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 217, + 294, + 249 + ], + "spans": [ + { + "bbox": [ + 56, + 217, + 294, + 249 + ], + "type": "text", + "content": "[35] David G Lowe. Distinctive image features from scale-invariant keypoints. Int. J. Comput. Vis., 60:91-110, 2004. 2, 4, 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 250, + 294, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 250, + 294, + 293 + ], + "spans": [ + { + "bbox": [ + 56, + 250, + 294, + 293 + ], + "type": "text", + "content": "[36] André Mateus, Siddhant Ranade, Srikumar Ramalingam, and Pedro Miraldo. Fast and accurate 3d registration from line intersection constraints. Int. J. Comput. Vis., pages 1-26, 2023. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 293, + 294, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 293, + 294, + 336 + ], + "spans": [ + { + "bbox": [ + 56, + 293, + 294, + 336 + ], + "type": "text", + "content": "[37] Pedro Miraldo, Surojit Saha, and Srikumar Ramalingam. Minimal solvers for mini-loop closures in 3d multi-scan alignment. In IEEE Conf. Comput. Vis. Pattern Recog., 2019. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 338, + 294, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 338, + 294, + 381 + ], + "spans": [ + { + "bbox": [ + 56, + 338, + 294, + 381 + ], + "type": "text", + "content": "[38] Francesco Pittaluga, Sanjeev J Koppal, Sing Bing Kang, and Sudipta N Sinha. Revealing scenes by inverting structure from motion reconstructions. In IEEE Conf. Comput. Vis. Pattern Recog., pages 145-154, 2019. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 383, + 294, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 383, + 294, + 426 + ], + "spans": [ + { + "bbox": [ + 56, + 383, + 294, + 426 + ], + "type": "text", + "content": "[39] Zheng Qin, Hao Yu, Changjian Wang, Yulan Guo, Yuxing Peng, and Kai Xu. Geometric transformer for fast and robust point cloud registration. In IEEE Conf. Comput. Vis. Pattern Recog., pages 11143-11152, 2022. 2, 3, 6, 7, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 426, + 294, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 426, + 294, + 470 + ], + "spans": [ + { + "bbox": [ + 56, + 426, + 294, + 470 + ], + "type": "text", + "content": "[40] Tilman Reinhardt. Using global localization to improve navigation, 2019. URL https://blog.research.google/2019/02/using-global-localization-to-improve.html.1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 471, + 294, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 471, + 294, + 504 + ], + "spans": [ + { + "bbox": [ + 56, + 471, + 294, + 504 + ], + "type": "text", + "content": "[41] Radu Bogdan Rusu, Nico Blodow, and Michael Beetz. Fast point feature histograms (fpfh) for 3d registration. In IEEE Int'l Conf. Robotics and Auto., pages 3212-3217, 2009. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 504, + 294, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 504, + 294, + 548 + ], + "spans": [ + { + "bbox": [ + 56, + 504, + 294, + 548 + ], + "type": "text", + "content": "[42] Paul-Edouard Sarlin, Cesar Cadena, Roland Siegwart, and Marcin Dymczyk. From coarse to fine: Robust hierarchical localization at large scale. In IEEE Conf. Comput. Vis. Pattern Recog., pages 12716–12725, 2019. 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 548, + 294, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 548, + 294, + 592 + ], + "spans": [ + { + "bbox": [ + 56, + 548, + 294, + 592 + ], + "type": "text", + "content": "[43] Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. SuperGlue: Learning feature matching with graph neural networks. In IEEE Conf. Comput. Vis. Pattern Recog., 2020. 2, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 592, + 294, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 592, + 294, + 624 + ], + "spans": [ + { + "bbox": [ + 56, + 592, + 294, + 624 + ], + "type": "text", + "content": "[44] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In IEEE Conf. Comput. Vis. Pattern Recog., 2016. 2, 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 625, + 294, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 625, + 294, + 657 + ], + "spans": [ + { + "bbox": [ + 56, + 625, + 294, + 657 + ], + "type": "text", + "content": "[45] Peter H. Schonemann. A generalized solution of the orthogonal procrustes problem. Psychometrika, 31(1):1-10, 1966. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 56, + 658, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 294, + 713 + ], + "type": "text", + "content": "[46] Jamie Shotton, Ben Glocker, Christopher Zach, Shahram Izadi, Antonio Criminisi, and Andrew Fitzgibbon. Scene coordinate regression forests for camera relocalization in rgb-d images. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2930-2937, 2013. 7" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 106 + ], + "type": "text", + "content": "[47] Tory Smith. Introducing lightship vps, 2022. URL https://lightship.dev/blog/introducing-lightship-vps/.1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 107, + 553, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 140 + ], + "type": "text", + "content": "[48] Christoph Strecha, Timo Pylvänäinen, and Pascal Fua. Dynamic and scalable large scale image reconstruction. In IEEE Conf. Comput. Vis. Pattern Recog., pages 406-413, 2010. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 142, + 553, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 142, + 553, + 196 + ], + "spans": [ + { + "bbox": [ + 316, + 142, + 553, + 196 + ], + "type": "text", + "content": "[49] Yifan Sun, Changmao Cheng, Yuhan Zhang, Chi Zhang, Liang Zheng, Zhongdao Wang, and Yichen Wei. Circle loss: A unified perspective of pair similarity optimization. In IEEE Conf. Comput. Vis. Pattern Recog., pages 6398-6407, 2020. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 198, + 553, + 242 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 198, + 553, + 242 + ], + "spans": [ + { + "bbox": [ + 316, + 198, + 553, + 242 + ], + "type": "text", + "content": "[50] Chris Sweeney, Victor Fragoso, Tobias Höllerer, and Matthew Turk. Large scale sfm with the distributed camera model. In Int'l Conf. 3D Vision, pages 230-238, 2016. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 244, + 553, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 244, + 553, + 289 + ], + "spans": [ + { + "bbox": [ + 316, + 244, + 553, + 289 + ], + "type": "text", + "content": "[51] Yurun Tian, Xin Yu, Bin Fan, Fuchao Wu, Huub Heijnen, and Vassileios Balntas. Sosnet: Second order similarity regularization for local descriptor learning. In IEEE Conf. Comput. Vis. Pattern Recog., pages 11016-11025, 2019. 4, 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 290, + 553, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 290, + 553, + 334 + ], + "spans": [ + { + "bbox": [ + 316, + 290, + 553, + 334 + ], + "type": "text", + "content": "[52] Haithem Turki, Deva Ramanan, and Mahadev Satyanarayanan. Mega-nerf: Scalable construction of large-scale nerfs for virtual fly-throughs. In IEEE Conf. Comput. Vis. Pattern Recog., pages 12922–12931, 2022. 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 335, + 553, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 335, + 553, + 369 + ], + "spans": [ + { + "bbox": [ + 316, + 335, + 553, + 369 + ], + "type": "text", + "content": "[53] Michal Tyszkiewicz, Pascal Fua, and Eduard Trulls. Disk: Learning local features with policy gradient. Adv. Neural Inform. Process. Syst., 33:14254-14265, 2020. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 371, + 553, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 371, + 553, + 425 + ], + "spans": [ + { + "bbox": [ + 316, + 371, + 553, + 425 + ], + "type": "text", + "content": "[54] Ole Untzelmann, Torsten Sattler, Sven Middelberg, and Leif Kobbelt. A scalable collaborative online system for city reconstruction. In Proceedings of the IEEE International Conference on Computer Vision Workshops, pages 644-651, 2013. 3, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 427, + 553, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 427, + 553, + 482 + ], + "spans": [ + { + "bbox": [ + 316, + 427, + 553, + 482 + ], + "type": "text", + "content": "[55] Shuzhe Wang, Juho Kannala, and Daniel Barath. Dgc-gnn: Leveraging geometry and color cues for visual descriptor-free 2d-3d matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20881–20891, 2024. 2, 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 484, + 553, + 517 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 484, + 553, + 517 + ], + "spans": [ + { + "bbox": [ + 316, + 484, + 553, + 517 + ], + "type": "text", + "content": "[56] Yue Wang and Justin M Solomon. Deep closest point: Learning representations for point cloud registration. In Int. Conf. Comput. Vis., pages 3523-3532, 2019. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 518, + 553, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 518, + 553, + 552 + ], + "spans": [ + { + "bbox": [ + 316, + 518, + 553, + 552 + ], + "type": "text", + "content": "[57] Yue Wang and Justin M Solomon. Prnet: Self-supervised learning for partial-to-partial registration. Adv. Neural Inform. Process. Syst., 32, 2019. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 554, + 553, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 554, + 553, + 587 + ], + "spans": [ + { + "bbox": [ + 316, + 554, + 553, + 587 + ], + "type": "text", + "content": "[58] Philippe Weinzaepfel, Hervé Jégou, and Patrick Pérez. Reconstructing an image from its local descriptors. In IEEE Conf. Comput. Vis. Pattern Recog., pages 337-344, 2011. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 588, + 553, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 588, + 553, + 632 + ], + "spans": [ + { + "bbox": [ + 316, + 588, + 553, + 632 + ], + "type": "text", + "content": "[59] Hao Xu, Shuaicheng Liu, Guangfu Wang, Guanghui Liu, and Bing Zeng. Omnet: Learning overlapping mask for partial-to-partial point cloud registration. In Int. Conf. Comput. Vis., pages 3132-3141, 2021. 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 634, + 553, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 634, + 553, + 667 + ], + "spans": [ + { + "bbox": [ + 316, + 634, + 553, + 667 + ], + "type": "text", + "content": "[60] Jiaolong Yang, Hongdong Li, and Yunde Jia. Go-icp: Solving 3d registration efficiently and globally optimally. In Int. Conf. Comput. Vis., pages 1457–1464, 2013. 3" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 669, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 669, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 669, + 553, + 713 + ], + "type": "text", + "content": "[61] Runzhao Yao, Shaoyi Du, Wenting Cui, Canhui Tang, and Chengwu Yang. Pare-net: Position-aware rotation-equivariant networks for robust point cloud registration. In ECCV, 2024. 7, 8, 2" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "6582" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 374 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "type": "text", + "content": "[62] Zi Jian Yew and Gim Hee Lee. RPM-net: Robust point matching using learned features. In IEEE Conf. Comput. Vis. Pattern Recog., pages 11824-11833, 2020. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 295, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 295, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 295, + 140 + ], + "type": "text", + "content": "[63] Zi Jian Yew and Gim Hee Lee. Regtr: End-to-end point cloud correspondences with transformers. In IEEE Conf. Comput. Vis. Pattern Recog., pages 6677-6686, 2022. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 141, + 294, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 141, + 294, + 195 + ], + "spans": [ + { + "bbox": [ + 56, + 141, + 294, + 195 + ], + "type": "text", + "content": "[64] Hao Yu, Zheng Qin, Ji Hou, Mahdi Saleh, Dongsheng Li, Benjamin Busam, and Slobodan Ilic. Rotation-invariant transformer for point cloud matching. In IEEE Conf. Comput. Vis. Pattern Recog., pages 5384-5393, 2023. 1, 2, 3, 5, 6, 7, 8, 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 197, + 294, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 197, + 294, + 251 + ], + "spans": [ + { + "bbox": [ + 56, + 197, + 294, + 251 + ], + "type": "text", + "content": "[65] Andy Zeng, Shuran Song, Matthias Nießner, Matthew Fisher, Jianxiong Xiao, and Thomas Funkhouser. 3dmatch: Learning local geometric descriptors from rgb-d reconstructions. In IEEE Conf. Comput. Vis. Pattern Recog., pages 1802-1811, 2017. 2, 4, 7, 8, 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 253, + 294, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 253, + 294, + 297 + ], + "spans": [ + { + "bbox": [ + 56, + 253, + 294, + 297 + ], + "type": "text", + "content": "[66] Yejun Zhang, Shuzhe Wang, and Juho Kannala. A2-gnn: Angle-annular gnn for visual descriptor-free camera relocalization. In International Conference on 3D Vision (3DV), 2025. arXiv preprint arXiv:2502.20036. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 297, + 294, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 297, + 294, + 341 + ], + "spans": [ + { + "bbox": [ + 56, + 297, + 294, + 341 + ], + "type": "text", + "content": "[67] Qunjie Zhou, Sérgio Agostinho, Aljosa Osep, and Laura Leal-Taixe. Is geometry enough for matching in visual localization? In Eur. Conf. Comput. Vis., pages 407-425, 2022. 2, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 342, + 294, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 342, + 294, + 374 + ], + "spans": [ + { + "bbox": [ + 56, + 342, + 294, + 374 + ], + "type": "text", + "content": "[68] Qian-Yi Zhou, Jaesik Park, and Vladlen Koltun. Open3D: A modern library for 3D data processing. arXiv:1801.09847, 2018. 6, 7, 1" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "6583" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Collaborative Decoding Makes Visual Auto-Regressive Modeling Efficient/9a935c24-e29c-43fe-a54b-aa3325d4f954_content_list.json b/2025/Collaborative Decoding Makes Visual Auto-Regressive Modeling Efficient/9a935c24-e29c-43fe-a54b-aa3325d4f954_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..2435f8e6dbb00395e2cd3d9cc06aac2025c9e951 --- /dev/null +++ b/2025/Collaborative Decoding Makes Visual Auto-Regressive Modeling Efficient/9a935c24-e29c-43fe-a54b-aa3325d4f954_content_list.json @@ -0,0 +1,1427 @@ +[ + { + "type": "text", + "text": "Collaborative Decoding Makes Visual Auto-Regressive Modeling Efficient", + "text_level": 1, + "bbox": [ + 127, + 130, + 870, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zigeng Chen, Xinyin Ma, Gongfan Fang, Xinchao Wang* National University of Singapore", + "bbox": [ + 259, + 180, + 743, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{zigeng99, maxinyin, gongfan}@u.nus.edu, xinchao@nus.edu.sg", + "bbox": [ + 238, + 219, + 759, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "https://github.com/czg1225/CoDe", + "bbox": [ + 310, + 234, + 679, + 251 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 286, + 326, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In the rapidly advancing field of image generation, Visual Auto-Regressive (VAR) modeling has garnered considerable attention for its innovative next-scale prediction approach. This paradigm offers substantial improvements in efficiency, scalability, and zero-shot generalization. Yet, the inherently coarse-to-fine nature of VAR introduces a prolonged token sequence, leading to prohibitive memory consumption and computational redundancies. To address these bottlenecks, we propose Collaborative Decoding (CoDe), a novel efficient decoding strategy tailored for the VAR framework. CoDe capitalizes on two critical observations: the substantially reduced parameter demands at larger scales and the exclusive generation patterns across different scales. Based on these insights, we partition the multi-scale inference process into a seamless collaboration between a large model and a small model. The large model serves as the 'drafter', specializing in generating low-frequency content at smaller scales, while the smaller model serves as the 'refiner', solely focusing on predicting high-frequency details at larger scales. This collaboration yields remarkable efficiency with minimal impact on quality: CoDe achieves a $1.7x$ speedup, slashes memory usage by around $50\\%$ , and preserves image quality with only a negligible FID increase from 1.95 to 1.98. When drafting steps are further decreased, CoDe can achieve an impressive $2.9x$ acceleration ratio, reaching 41 images/s at 256x256 resolution on a single NVIDIA 4090 GPU, while preserving a commendable FID of 2.27.", + "bbox": [ + 89, + 319, + 483, + 743 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 773, + 220, + 789 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The past year has witnessed significant advancements in Auto-Regressive (AR) models for image generation [17, 22, 26, 31, 48, 59], driven by their proven scalability and strong generalization capabilities [1, 9, 16, 52]. Leveraging these strengths, AR approaches have demonstrated re", + "bbox": [ + 89, + 799, + 483, + 876 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9c935d98139a7d1b133241f4ce75e3c756050044b7aed6ca6d0f7ef3548bd591.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 522, + 286, + 898, + 414 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/40927f390e85beaf727d352c7ff6eb69fa1b9634effca101de8ac3c6a469dd37.jpg", + "image_caption": [ + "Figure 1. We partition the next-scale prediction process into the efficient collaboration between large and small VAR models." + ], + "image_footnote": [], + "bbox": [ + 522, + 412, + 895, + 542 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "markable image quality and promising multi-modal potential [32, 56, 70]. However, the next-token prediction process inherent to conventional AR models requires numerous decoding steps, leading to considerable generation latency. Visual Auto-Regressive (VAR) Modeling [51] replaces the GPT-style next-token prediction with a next-scale prediction strategy. It generates content in a multi-level, coarse-to-fine progression, enabling the model to decode multiple tokens in parallel [29, 42, 49, 63, 65], thus considerably reducing the inference steps.", + "bbox": [ + 509, + 612, + 906, + 763 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite less decoding steps, VAR's progressive scaling approach significantly increases the overall sequence length. To generate a 16x16 token image, VAR hierarchically decodes up to 680 tokens across 10 scales—2.7 times the sequence length required by conventional AR models. As the VAR model must store the KV cache accumulated from all previous tokens, this prolonged sequence results in a substantial memory overhead, especially in the final scales. During inference, the KV cache consumes approx-", + "bbox": [ + 511, + 763, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author", + "bbox": [ + 109, + 887, + 233, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "23334", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/94ce9a46c2690dcbe99efb7d4bc688229150a02504b85c5955cbe9f55ef54bac.jpg", + "image_caption": [ + "Figure 2. Comparison of generation results between original VAR-d30 (up) and our VAR-CoDE (bottom) for ImageNet $256 \\times 256$ . Our method achieves 1.7x speedup (3.62s to 2.11s), and needs only 0.5x memory space (40GB to 20GB), with negligible quality degradation." + ], + "image_footnote": [], + "bbox": [ + 125, + 90, + 903, + 280 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "imately 12 times more memory than the forward computation itself. For instance, generating images using VAR-d30 with a batch size of 128 demands 70 GB of memory, with 57 GB dedicated solely to KV caching, becoming the primary bottleneck. Moreover, the extended sequence length exacerbates the computational cost of self-attention, given the quadratic growth in attention map calculations.", + "bbox": [ + 88, + 343, + 480, + 450 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address these inefficiencies, we begin by analyzing the specific properties of VAR's next-scale prediction paradigm. First, we observe that as generated scales increase, the parameter demands for high-quality token generation drop substantially, leading to considerable computational redundancy at most of the tokens in the long sequence. Next, we find that the generation patterns at small and large scales are exclusive, resulting in mutual interference between learning low- and high-frequency modeling capability, thus hindering efficient parameter utilization.", + "bbox": [ + 88, + 455, + 482, + 607 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our Approach. Inspired by the above observations, we propose Collaborative Decoding (CoDe), a simple yet highly effective method that significantly boosts the inference efficiency of VAR models while preserving generation quality comparable to the original. Our core idea is to decompose the long-sequence scaling-up process into a collaboration between two VAR models with different sizes and specialized roles to enhance efficiency. Figure 1 presents the overview of our method. We use a large VAR model as a drafter for the initial small scales, where model capacity demands are high but computations and KV cache are sparse. Conversely, we use a small VAR model as a refiner at the remaining large scales, where computations and KV cache are intensive but fewer parameters are needed. Next, we apply specialized fine-tuning to both models to address the optimization interference between scales encountered during the pre-training stage. Each model is fine-tuned exclusively on the specific scales it handles, leading to a notable performance boost with limited additional cost.", + "bbox": [ + 91, + 613, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Extensive experiments validate the effectiveness of our method. As shown in Figure 2, compared to the original VAR-d30 model, CoDe achieves a $1.7\\mathrm{x}$ speedup with merely $0.5\\mathrm{x}$ GPU memory consumption and negligible quality loss (FID [18] slightly increases from 1.95 to 1.98). Notably, our approach can reach up to a $2.9\\mathrm{x}$ speedup, generating 41 images (256x256) per second on a single NVIDIA 4090 GPU, while still maintaining an FID of 2.27.", + "bbox": [ + 511, + 343, + 903, + 464 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In conclusion, we introduce CoDe, a novel decoding framework for visual auto-regressive modeling that significantly enhances speed and reduces memory overhead with a negligible impact on quality. CoDe divides the hierarchical sequence modeling of VAR into a collaborative process, utilizing a large VAR model and a small VAR model in a progressive partnership. Additionally, we propose specialized fine-tuning to optimize each model for its specific role, effectively mitigating training interference and maximizing parameter utilization. To the best of our knowledge, CoDe is the fastest method available to achieve an FID below 2, making it a clear advancement in efficient image generation.", + "bbox": [ + 511, + 465, + 906, + 647 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 511, + 662, + 663, + 679 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Auto-regressive image generation. Early works [4, 53] pioneered image generation by generating pixels in raster-scan order. Later, VQVAE [54] and VQGAN [10] improved this approach by quantizing image patches into discrete tokens, using a transformer in a decoder-only setup to generate these tokens in a raster-scan manner. Building on these foundations, recent efforts have focused on enhancing autoregressive (AR) models for image generation. LlamaGEN [48] and Lumina-mGPT [32], for example, use a GPT-style next-token prediction strategy to achieve high-quality image generation with good scalability. AiM [22] and MARS [17] further improve this paradigm by introducing mixture-of-experts and linear attention mechanisms [14]. Methods like SHOW-O [56], Transfusion [70], and DART [15]", + "bbox": [ + 511, + 688, + 906, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "23335", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/5bcc327a07b0252b3123f285d8b87697801a6ba07fb16dd8579f9674e2dc2efa.jpg", + "image_caption": [ + "(a) Quality vs. Params at different scales", + "Figure 3. (a) Effectiveness of increasing parameters at the $k$ -th scale is evaluated by predicting token map $r_k$ using four VAR models with different parameter sizes (2B, 1B, 0.6B, and 0.3B), while other scales $(r_1, r_2, \\ldots, r_{k-1}, r_{k+1}, \\ldots, r_{10})$ are generated using the largest VAR-d30 model. (b) Fourier spectrum analysis is conducted on generated content at the first 3 scales and the last 3 scales. (c) Training-free performance comparison of model collaboration decoding across various settings of draft tokens $M$ and refiner tokens $680 - M$ ." + ], + "image_footnote": [], + "bbox": [ + 94, + 89, + 351, + 220 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/832ed1fa91c44ad4b066fbf183c7d73540ac572e01d7cfd12114957736abb29d.jpg", + "image_caption": [ + "(b) Spectrum analysis" + ], + "image_footnote": [], + "bbox": [ + 361, + 88, + 614, + 220 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/8ad5650c2f3680e6367052a46b7be781d5af6cbab821c8646a78087076862ebf.jpg", + "image_caption": [ + "(c) Training-free model collaboration" + ], + "image_footnote": [], + "bbox": [ + 617, + 88, + 870, + 219 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "combine diffusion processes with autoregressive modeling, while [3, 26, 36] introduces masked autoregression to generate images. However, all these approaches suffer from high latency due to the large number of forward steps. VAR [51] addresses this by using hierarchical parallel decoding, which greatly reduces the number of steps, leading to significantly lower inference latency without sacrificing quality. Building on VAR's progress, many recent works have aimed to improve next-scale prediction across multiple tasks, including text-to-image [28, 39, 49, 65], controllable generation [29, 31], audio generation [42], and 3D generation [63].", + "bbox": [ + 88, + 328, + 483, + 494 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Efficient Image Generation Models. For diffusion models, acceleration techniques are already well-developed. [35, 43, 44, 60, 61] focuses on reducing sampling steps while [12, 30, 58, 62, 68] optimize model architectures through pruning [5, 11] or knowledge distillation [19]. To avoid the high costs of training, some training-free methods are proposed. Some approaches develop fast solvers for stochastic differential equations (SDE) or ordinary differential equations (ODE) to enhance sampling efficiency [2, 33, 34, 64, 69]. Other works [25, 37, 38, 46, 55, 57, 66, 67] exploit specific characteristics of diffusion models to skip redundant computations during the denoising process. [6, 13, 23] reduce latency to another level through distributed computing. Quantization methods [24, 27, 45] have also shown great potential. In contrast, research into accelerating AR image generation is still in its early stages. LANTERN [20] and SJD [50] employ speculative decoding to speed up next-token prediction, achieving notable speedups but not suitable for the innovative next-scale prediction paradigm introduced by VAR.", + "bbox": [ + 91, + 500, + 483, + 803 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this paper, we introduce CoDe, a novel and efficient decoding method tailored for the next-scale prediction paradigm. CoDe effectively addresses the inefficiencies associated with the long sequence structure of VAR models while maintaining high generation quality. Notably, CoDe stands out as the fastest method capable of achieving an FID", + "bbox": [ + 89, + 810, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "below 2, making it a clear advancement.", + "bbox": [ + 511, + 328, + 781, + 343 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 511, + 354, + 604, + 371 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminary", + "text_level": 1, + "bbox": [ + 511, + 380, + 643, + 397 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Visual auto-regressive modeling [51] redefines the traditional AR by shifting from a \"next-token\" prediction to a \"next-scale\" prediction. In this framework, each auto-regressive unit is a token map at varying scales, rather than a single token. For a given image feature map $f \\in \\mathbb{R}^{h \\times w \\times C}$ , VAR quantizes it into $K$ multi-scale token maps $R = (r_1, r_2, \\ldots, r_K)$ at progressively finer resolutions, with the final token map $r_K$ matching the original feature map's resolution. The probability distribution is reformulated as:", + "bbox": [ + 511, + 402, + 906, + 539 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\np \\left(r _ {1}, r _ {2}, \\dots , r _ {K}\\right) = \\prod_ {k = 1} ^ {K} p \\left(r _ {k} \\mid r _ {1}, r _ {2}, \\dots , r _ {k - 1}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 534, + 547, + 906, + 589 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where each token map $r_k \\in [V]^{h_k \\times w_k}$ consists of $h_k \\times w_k$ tokens at scale $k$ , and the sequence $(r_1, r_2, \\ldots, r_{k-1})$ serves as the \"prefix\" for $r_k$ . In this paradigm, during each autoregressive step $k$ , the model predicts all $h_k \\times w_k$ tokens in $r_k$ in parallel, conditioned on prior scales and position embeddings. This approach aligns with a coarse-to-fine generation pattern, enabling parallel decoding within each scale. VAR effectively improves inference speed and generation quality, but it also considerably increases sequence length.", + "bbox": [ + 511, + 598, + 908, + 736 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Key Observations", + "text_level": 1, + "bbox": [ + 511, + 742, + 687, + 758 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "VAR represents an innovative paradigm specific for visual generation, distinct from traditional autoregressive approaches and introducing many yet unexplored characteristics. In this work, we revisit the entire next-scale prediction process to uncover specific properties that can be optimized to reduce computational redundancy and improve inference efficiency.", + "bbox": [ + 511, + 763, + 906, + 869 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Observation 1: As the predicted scale becomes larger, the need for parameters reduces significantly. While VAR", + "bbox": [ + 511, + 869, + 908, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "23336", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/091c168ef7e121154f6d60de6c1db09622d48246121795b963374b64f34389f9.jpg", + "image_caption": [ + "Figure 4. Overview of the collaborative decoding process, we use a drafting step $N = 6$ for instance. CoDe uses a large VAR model as the drafter $\\epsilon_{\\theta_d}$ to generate the token maps $R_L = (r_1, r_2, \\dots, r_N)$ at smaller scales. The small refiner model $\\epsilon_{\\theta_r}$ then uses $R_L$ as an initial prefix to efficiently predict the remaining token maps $R_H = (r_{N+1}, r_{N+2}, \\dots, r_K)$ at larger scales. Both models are fine-tuned on their designated predictive scales using ground truth labels $r_k^*$ and teacher logits $p_{\\text{teacher}}(r_k)$ , respectively." + ], + "image_footnote": [], + "bbox": [ + 96, + 89, + 612, + 308 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/7343d83148610ce5b6f70d0c9222955db4eb4eede9d0c4b62e016674eb7b9a67.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 627, + 128, + 903, + 299 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "models exhibit strong scalability, the effectiveness of increasing parameter counts varies greatly in predicting various scales. In Figure 3. (a), we present the differences in class-conditional generation (ImageNet-256 [8]) quality when using VAR models of different sizes at each scale. At the initial small scales, increasing model parameters leads to noticeable improvements in generation quality. However, as the predicted scale gradually progresses, the impact of additional parameters becomes minimal. At the final scale—which accounts for $38\\%$ tokens of the entire sequence—we observe that the performance of the 2B model is even nearly equivalent to that of the 0.3B model. These findings suggest that as the predicted scale increases, the parameter demand for accurate token predictions declines markedly, revealing substantial computational redundancy in the current VAR inference process at large scales.", + "bbox": [ + 88, + 402, + 485, + 643 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Observation 2: The generative patterns are exclusive between small and large scales. For VAR models, the content generated at small and large scales varies significantly. As demonstrated by the Fourier analysis in Figure 3 (b), the feature maps produced by the first three scales contain primarily low-frequency components, while the last three scales focus on high-frequency components. To further validate these distinct generative patterns, we conducted a perturbation fine-tuning experiment. Using a pre-trained VARd16 model, we applied CSE loss only to tokens in the largest three scales and fine-tuned for just $1\\%$ of the original training epochs. This minor fine-tuning at large scales led to a complete collapse of the model's global modeling capacity at small scales, with the FID increasing from 3.30 to 21.93 and the IS score dropping from 277 to 88. These results demonstrate that VAR models perform entirely dis", + "bbox": [ + 89, + 659, + 485, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "tinct generative tasks at small and large scales, with minimal overlap. Training a single VAR model to predict across all scales results in significant mutual interference between learning low-frequency and high-frequency modeling capabilities, making parameter optimization challenging.", + "bbox": [ + 511, + 402, + 906, + 477 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In sum, the key insights from our observations can be concluded as follows: (1) At large scales, VAR models require significantly fewer parameters for accurate token prediction, leading to considerable computational redundancy. (2) Training a single model to generate across all scales causes mutual interference between small and large scales, hindering parameter optimization due to conflicting low-frequency and high-frequency learning tendencies.", + "bbox": [ + 511, + 478, + 908, + 602 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Collaborative Decoding", + "text_level": 1, + "bbox": [ + 511, + 606, + 732, + 623 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Based on the above insights, we propose a simple yet pretty efficient decoding method for next-scale prediction called collaborative decoding. As presented in Figure 4, CoDe decomposes the next-scale prediction process into a collaboration between a large drafter model and a small refiner model. The drafter is the original large VAR model with 2B parameters, while the refiner is a lightweight 0.3B VAR model, sharing the same transformer-based architecture but with reduced width and depth. The drafter is responsible for generating coarse, low-frequency global structures at smaller scales to draft the image, while the refiner continually predicts the high-frequency details at larger scales to refine the image.", + "bbox": [ + 511, + 628, + 908, + 824 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Model Collaboration. To generate an image $im$ , original VAR model needs to auto-regressively generate $K$ multiscale token maps $R = (r_1, r_2, \\ldots, r_K)$ with progressively finer resolutions, where each token map $r_k \\in [V]^{h_k \\times w_k}$ represents the token map at scale $k$ . In CoDe, we decom", + "bbox": [ + 511, + 825, + 908, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "23337", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "pose the auto-regressive steps into $N$ drafting steps and $K - N$ refining steps. In the initial drafting phase where computation is sparse but additional parameters provide significant benefits, we serve a large VAR model as the drafter $\\epsilon_{\\theta_d}$ , generating the initial set of low-frequency token maps $R_L = (r_1,r_2,\\dots ,r_N)$ up to scale $N$ , where $N < K$ . These token maps predict the global structure of the image, serving as a prefix for the remaining finer-scale token maps. The drafting process can be represented as:", + "bbox": [ + 89, + 90, + 483, + 227 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\np _ {\\theta_ {d}} \\left(r _ {1}, r _ {2}, \\dots , r _ {N}\\right) = \\prod_ {k = 1} ^ {N} p _ {\\theta_ {d}} \\left(r _ {k} \\mid r _ {1}, r _ {2}, \\dots , r _ {k - 1}\\right). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 233, + 483, + 273 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After drafting, the KV cache of the large model is released, significantly optimizing memory usage.", + "bbox": [ + 89, + 279, + 482, + 309 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the refining stage where the computation is intensive but the parameter demand is low, we serve a small VAR model as the refiner $\\epsilon_{\\theta_r}$ , employing the drafter's predictions $R_L$ as a prefix and focus on refining the details by generating the remaining high-frequency token maps $R_H = (r_{N + 1}, r_{N + 2}, \\ldots, r_K)$ up to scale $K$ , where $r_K$ matches the resolution of the original feature map. Notably, predicting $r_{N + 1}$ requires an attention mask, as the refiner model lacks the KV cache for the first $N$ scales. The probability distribution of refining steps is formulated as:", + "bbox": [ + 89, + 310, + 483, + 460 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} p \\left(r _ {N + 1}, r _ {N + 2}, \\dots , r _ {K} \\mid R _ {L}\\right) \\\\ = \\prod_ {k = N + 1} ^ {K} p \\left(r _ {k} \\mid R _ {L}, r _ {N + 1}, r _ {N + 2}, \\dots , r _ {k - 1}\\right). \\tag {3} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 138, + 465, + 482, + 527 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Finally, the generated images $im$ are reconstructed from both drafting maps $R_{L}$ and refining maps $R_{H}$ as follows:", + "bbox": [ + 89, + 534, + 483, + 565 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\ni m = D \\left(Q \\left(R _ {L}, R _ {H}\\right), \\right. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 210, + 578, + 482, + 595 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $Q(.)$ is a residual-style quantization function, and $D(.)$ is a multi-scale VQVAE decoder. Figure 3. (c) demonstrates the remarkable effectiveness of model collaboration under training-free conditions. With a large model as the drafter, the small refiner only results in a slight and gentle increase in FID. Conversely, with a small model as the drafter, even assigning the last $80\\%$ tokens to a large refiner fails to improve performance. This also verifies Observation 1: small scales require more model capacity, whereas the parameter demands for large scales are low.", + "bbox": [ + 89, + 598, + 482, + 748 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The model collaboration method effectively reduces the computational redundancy of VAR and maintains generation quality comparable to the original model, while significantly offering faster speed and reduced memory usage.", + "bbox": [ + 89, + 750, + 482, + 810 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Specialized Fine-Tuning. Given the exclusive generation patterns between drafting scales and refining scales, we propose the specialized fine-tuning to further specialize drafter and refiner models by fine-tuning them solely on their respective predictive scales, thereby avoiding training interference and enhancing generation quality.", + "bbox": [ + 89, + 810, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For the drafter model $\\epsilon_{\\theta_d}$ , we fine-tune the drafting steps to improve the generation of the initial low-frequency token maps $R_L$ . The drafter model is trained using a CSE loss $\\mathrm{CSE}(\\cdot, \\cdot)$ between its generated token distribution $p_{\\theta_d}(r_k)$ and the ground truth labels $r_k^*$ , defined as:", + "bbox": [ + 511, + 90, + 905, + 167 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {d r a f t e r}} = \\sum_ {k = 1} ^ {N} \\mathrm {C S E} \\left(p _ {\\theta_ {d}} \\left(r _ {k}\\right), r _ {k} ^ {*}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 601, + 176, + 905, + 217 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "which encourages the drafter to closely match the target global structure at each drafting step.", + "bbox": [ + 511, + 226, + 905, + 255 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For the refiner model $\\epsilon_{\\theta_r}$ , we adopt a Knowledge Distillation (KD) [19] approach to partially transfer the knowledge from a larger pre-trained VAR model. The KD loss function for epoch $ep$ can be expressed as:", + "bbox": [ + 511, + 256, + 905, + 316 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {r e f i n e r}} ^ {(e p)} = \\sum_ {k = 1} ^ {K} \\left(\\lambda_ {e p} \\cdot \\mathbf {1} _ {[ k \\leq N ]} + \\mathbf {1} _ {[ k > N ]}\\right) \\mathrm {K L} \\left(p _ {\\theta_ {r}} \\left(r _ {k}\\right) \\| p _ {\\text {t e a c h e r}} \\left(r _ {k}\\right)\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 513, + 324, + 908, + 372 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $p_{\\mathrm{teacher}}(\\cdot)$ is the distribution predicted by the larger teacher VAR model, $\\mathrm{KL}(\\cdot, \\cdot)$ is the standard KL-divergence loss aligns refiner's output distribution with that of the teacher model, and $\\lambda_{ep}$ is a dynamic weighting factor. $\\lambda_t$ linearly decreases from one to zero from the initial to the end of finetuning, gradually shifting the KD emphasis from all tokens to the specific refining tokens. This dynamic weighting allows the refiner to specialize in refining the details with a smoother and more stable training process.", + "bbox": [ + 511, + 373, + 905, + 508 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The specialized fine-tuning of the drafter and refiner models ensures that each model becomes highly proficient in its respective task. By minimizing interference between small and large scales, this approach allows for more sufficient parameter optimization, resulting in enhanced image generation quality with limited training costs.", + "bbox": [ + 511, + 508, + 905, + 599 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experimental Results", + "text_level": 1, + "bbox": [ + 511, + 612, + 718, + 628 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Setup.", + "text_level": 1, + "bbox": [ + 511, + 637, + 707, + 652 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details. We evaluated our method's effectiveness on the ImageNet1K [8] class-conditional generation benchmarks. The proposed CoDe framework involves collaboration between a large and a small VAR model. Specifically, we use the pre-trained VAR-d30 model as the drafter and the pre-trained VAR-d16 model as the refiner. Each model undergoes specialized fine-tuning focused exclusively on its respective predictive scales. For the drafter, we fine-tune it for $5\\%$ of its original training epochs with a base learning rate of 1e-6 and a weight decay of 0.08. The refiner model, on the other hand, is fine-tuned for $25\\%$ of its original training epochs with a base learning rate of 1e-5 and no weight decay. Both models are optimized using the AdamW [21] optimizer with a batch size of 1024, achieved through gradient accumulation. The fine-tuning was conducted on 4 NVIDIA L20 GPUs.", + "bbox": [ + 511, + 657, + 906, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "23338", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/dc55050fcdac980dc9e1d9d489ef7462f9451c1081fd7f37592f82b10ecd2d99.jpg", + "image_caption": [ + "(a) Efficiency & Quality trade-off" + ], + "image_footnote": [], + "bbox": [ + 93, + 88, + 357, + 223 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4a5834facc896802db97a7dc744ebe5cbdb464c67ba2bb8e5a1055abf5330c48.jpg", + "image_caption": [ + "Figure 5. (a) Our CoDe demonstrates the optimal efficiency-quality trade-off among all evaluated methods. (b) Inference latency is measured across varying batch sizes for the original VAR-d30, our CoDe $(N = 6)$ , and the VQVAE decoder. (c) We analyze the time cost associated with parallel decoding at each scale, showing that the refiner model is significantly more efficient than the drafter at larger scales." + ], + "image_footnote": [], + "bbox": [ + 366, + 88, + 630, + 222 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c7c8080d11d76784bfaf5c2250f05668741b1712e2d7b5cd25d961cd1d874316.jpg", + "image_caption": [ + "(b) Time cost with different batchsizes", + "(c) Time cost at each scale" + ], + "image_footnote": [], + "bbox": [ + 632, + 88, + 895, + 222 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/7c703a633b6ae10e5aeb4abf6571a1442e3f61874087ae216a9692e431a1c9e4.jpg", + "table_caption": [ + "Table 1. Quantitative assessment of the efficiency-quality trade-off across various methods. Inference efficiency is evaluated with a batch size of 64 on NVIDIA L20 GPU, with latency measured excluding VQVAE or VQGAN as it incurs a shared time cost across all methods." + ], + "table_footnote": [], + "table_body": "
MethodInference EfficiencyGeneration Quality
#StepsSpeedup↑Latency↓Throughput↑#ParamMemory↓FID ↓IS ↑Precision↑Recall↑
DiT-XL/2500.2x19.20s3.33it/s675M11369MB2.262390.800.60
MAR-B1000.1x29.80s2.15it/s208M8725MB2.312820.820.57
AiM-XL2560.4x9.32s6.87it/s763M20983MB2.562570.820.57
LlamaGen-XXL384<0.1x73.97s0.87it/s1.4B42632MB2.342540.800.59
VAR-d30101.0x3.62s17.71it/s2.0B39228MB1.953010.810.59
VAR-d24101.7x2.07s30.92it/s1.0B25093MB2.113110.820.59
VAR-d20102.8x1.29s49.62it/s600M17814MB2.613010.830.56
VAR-CoDe N=99+11.2x2.97s21.54it/s2.0+0.3B28803MB1.942960.800.61
VAR-CoDe N=88+21.7x2.11s30.33it/s2.0+0.3B21019MB1.983020.810.60
VAR-CoDe N=77+32.3x1.60s40.00it/s2.0+0.3B19943MB2.113030.820.59
VAR-CoDe N=66+42.9x1.27s50.39it/s2.0+0.3B19943MB2.272970.820.58
", + "bbox": [ + 91, + 345, + 906, + 563 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/78682c545e01d162f00410402f9a09f59115e1d3c193206ffe00f66f05285bdc.jpg", + "table_caption": [ + "Table 2. Effect of specialized fine-tuning" + ], + "table_footnote": [], + "table_body": "
DiscriptionN=6N=7N=8N=9
CoDe Training-free2.422.262.101.99
CoDe Fine-tuning2.272.111.981.94
", + "bbox": [ + 102, + 612, + 467, + 671 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "**Evaluations.** We evaluated our method on ImageNet1K-256 generation, focusing on both quality and efficiency metrics. For quality assessment, we used standard metrics such as FID, Inception Score (IS), Precision, and Recall. For other baselines, we use their default sampling methods. For CoDe, we reduced the sampled top-k from 900 to 600 due to improved token prediction accuracy after fine-tuning, while keeping the default top-p at 0.96. To compensate for reduced diversity, we introduced a temperature coefficient of $t = 1.1$ for sampling on the smallest 7 scales. Efficiency was measured through inference latency, throughput, memory consumption, and parameter count to provide a comprehensive comparison. Notably, the speed measurements", + "bbox": [ + 89, + 703, + 485, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "exclude the VQGAN decoder, which contributes minimally to overall runtime and is a shared component across all methods. All efficiency tests were conducted on a single NVIDIA L20 GPU without additional optimizations such as FlashAttention [7], using PyTorch 2.1 [40] and FP16.", + "bbox": [ + 511, + 588, + 906, + 664 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Main Results.", + "text_level": 1, + "bbox": [ + 511, + 678, + 656, + 693 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Quality-Efficiency Trade-off. We evaluated the quality-efficiency trade-off of our proposed method (CoDe) against original VAR models [51], state-of-the-art AR models, and diffusion models. The AR models included GPT-style LlamaGEN [48], Mamba-based Aim [22], and MAE-style MAR [26]. For diffusion models, we employed the widely used DiT models [41]. To measure the trade-offs, we used series models with different parameter counts for VAR, LlamaGEN, Aim, and MAR, while for DiT we varied the DDIM [47] steps to generate the efficiency-quality curve.", + "bbox": [ + 511, + 702, + 906, + 853 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As illustrated in Table 1, We analyzed CoDe's performance under different drafting steps N. Compared to traditional AR models, our proposed VAR-CoDe $(N = 6)$ achieves", + "bbox": [ + 511, + 854, + 905, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "23339", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/aa9b79272d373c128c4d3be3d5a6b11633b164ec4a9e1aa8863fe6d5dbc50c31.jpg", + "image_caption": [ + "Figure 6. Qualitative comparison between the original VAR-d30 model and our proposed CoDe model, with different drafting steps." + ], + "image_footnote": [], + "bbox": [ + 93, + 89, + 903, + 340 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "60 times faster inference than LlamaGEN-XXL, while also surpassing its generation quality. Additionally, in comparison to diffusion models, VAR-CoDe $(N = 6)$ is 15 times faster than DiT-XL/2, while maintaining the same level of quality. Compared to the original VAR-d30, CoDe $(N = 8)$ achieves a $1.7\\mathrm{x}$ speedup and reduces memory consumption by $50\\%$ , with only a negligible FID increase from 1.95 to 1.98. When the drafting stage is reduced to just 6 steps, CoDe $(N = 6)$ achieves a $2.9\\mathrm{x}$ speedup, reaching a throughput of over 50it/s, while maintaining a low FID of 2.27. This is a speed unmatched by any other existing methods. Compared to other VAR models with fewer parameters (VAR-d24 and VAR-d20), our method achieves significantly better quality while maintaining the same speedup ratio.", + "bbox": [ + 88, + 393, + 482, + 604 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As shown in Figure 5 (a), VAR-CoDe achieves the best efficiency-quality trade-off compared to all other methods, effectively solving the inefficiencies introduced by the extended sequence length in the original VAR paradigm.", + "bbox": [ + 89, + 607, + 482, + 670 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Training-Free vs Specialized Finetuning. Our CoDe framework can operate in a training-free manner by directly using pre-trained VAR-d30 and VAR-d16 models as the drafter and refiner, respectively. Alternatively, we can perform specialized fine-tuning to further enhance the models' performance at their respective scales, with limited training cost. Table 2 compares the results of training-free CoDe with the specialized fine-tuned version. Even without additional training, CoDe demonstrates competitive performance, outperforming VAR-d24 and VAR-d20 models with the same speedup ratio. With specialized fine-tuning, CoDe's performance improves significantly, even achieving a slight quality enhancement over the original model at a $1.2\\mathrm{x}$ acceleration ratio.", + "bbox": [ + 89, + 670, + 482, + 882 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Qualitative Results. We provide an extensive qualitative", + "bbox": [ + 89, + 885, + 482, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "comparison between the original VAR-d30 model and our proposed CoDe, with varying drafting steps $N = \\{6,7,8\\}$ . As illustrated in Figure 6, our approach achieves significant speedup and substantial memory optimization, with only minimal quality degradation that is nearly imperceptible to the human eye. Even at a speedup rate of 2.9 times, the generated images maintain high quality and accurate semantic information. Through specialized fine-tuning, CoDe's draft model exhibits higher predictive accuracy compared to the original model. This sometimes results in a different global structure from the original output, yet the image quality remains consistently high or even better.", + "bbox": [ + 511, + 393, + 906, + 575 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Zero-shot Task Generalization. To evaluate the zero-shot generalization capability of CoDe, we conducted additional experiments on zero-shot class-conditional inpainting and image editing. During image inpainting, we applied teacher forcing by providing the ground truth tokens outside the masked area, allowing the model to generate tokens solely within the mask. Notably, class-conditional information was introduced to the model. In the image editing task, CoDe was restricted to generating tokens within a given bounding box based on a specific class label. Figure 7 illustrates the qualitative results of CoDe's zero-shot performance $(N = 8)$ . Our approach demonstrates strong zero-shot generalization without any additional training on these downstream tasks.", + "bbox": [ + 511, + 575, + 908, + 787 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Efficiency Analysis.", + "text_level": 1, + "bbox": [ + 511, + 801, + 699, + 816 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Time Cost Analysis. We first present the time cost (bs=64) for each decoding step of the VAR-d30 model in Figure 5 (c). Notably, the computational complexity across the ten decoding steps of VAR is highly non-uniform. The last three decoding steps alone account for $64\\%$ of the total in", + "bbox": [ + 511, + 824, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "23340", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1717531237c044669f2b369a0fceacaf34053e4113ea17db356a0cde4a2088fc.jpg", + "image_caption": [ + "Figure 7. Qualitative results of CoDe's zero-shot generalization on image inpainting and image editing." + ], + "image_footnote": [], + "bbox": [ + 99, + 90, + 910, + 315 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/34d9c1a3addc044bc941aeb520366c11ac91d6e44f332c5d5a1a2763fab0764c.jpg", + "table_caption": [ + "Table 3. Memory usage comparison across different batch sizes" + ], + "table_footnote": [], + "table_body": "
MethodMemory Consumption↓
RunningKV CacheParamsTotal
VAR (bs=8)314MB3595MB8089MB12002MB
+CoDe284MB1023MB9275MB10619MB
VAR (bs=16)615MB7191MB8089MB15901MB
+CoDe557MB2056MB9275MB11951MB
VAR (bs=32)1216MB14345MB8089MB23662MB
+CoDe1103MB4083MB9275MB14614MB
VAR (bs=64)2420MB28707MB8089MB39228MB
+CoDe2195MB8160MB9275MB19943MB
VAR (bs=128)OOM(0.48GB)OOM(57GB)OOM(0.80GB)OOM(70GB)
+CoDe4380MB16320MB9275MB30598MB
", + "bbox": [ + 91, + 393, + 491, + 563 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ference time, with latency increasing quadratically as token map resolution grows. Our solution is to replace the large VAR model with a smaller one for the last few steps, as fewer parameters are needed for large-scale token maps. This small refiner is substantially faster than the original large drafter at these scales, achieving a 4.6x speed improvement in the final and most computationally demanding step. This change dramatically accelerates the entire inference process compared to the original VAR model.", + "bbox": [ + 89, + 595, + 482, + 731 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Next, we present the speedup effect of CoDe $(N = 6)$ across different batch sizes. As shown in Figure 5.(b), CoDe achieves a speedup of $1.6\\mathrm{x}$ for batch size 1 but reaches up to $2.9\\mathrm{x}$ for batch size 64. The reduced speedup at smaller batch sizes is due to low GPU utilization when generating $256\\times 256$ images, which limits the acceleration potential. However, as GPU utilization increases with larger batch sizes, the speedup of CoDe also improves significantly. This indicates that CoDe is well-suited for computationally intensive image generation tasks, and holds great potential for enhancing efficiency in high-resolution image genera", + "bbox": [ + 89, + 734, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "tion under the VAR paradigm. Additionally, we report the time cost of the VQVAE decoder in the figure.", + "bbox": [ + 511, + 369, + 903, + 400 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Memory Consumption Analysis. In Table 3, we provide a detailed analysis of memory usage during the VAR inference process. The KV cache of the VAR model is the largest memory consumer, requiring 12 times the memory needed for the activations due to the significantly extended sequence length in the VAR paradigm. Our proposed CoDe effectively addresses this KV cache memory challenge. Specifically, we use the large VAR model only for predicting the first 91 tokens, which represent just $13\\%$ of the total sequence. After this, the KV cache in the dafter is released, and the remaining computation is handled by a smaller model. This approach drastically reduces the KV cache memory requirements, compressing it to approximately $28\\%$ of the original VAR-d30. Although the refiner model adds a small number of additional parameters, this overhead is minimal compared to the major optimization achieved in KV cache storage. As a result, the total memory usage of CoDe is significantly lower than the original model. Moreover, as batch size increases, the memory savings with CoDe become even more pronounced, underscoring its potential to support efficient high-resolution image generation within the VAR paradigm.", + "bbox": [ + 511, + 402, + 906, + 734 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 752, + 633, + 768 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work presents CoDe, a novel method designed for efficient decoding in visual auto-regressive modeling. CoDe effectively mitigates the significant memory overhead and computational redundancy typically associated with the prolonged sequences of next-scale predictions. Through extensive experimentation, our method demonstrates a superior efficiency-quality trade-off, establishing a new benchmark for efficient, high-quality image generation.", + "bbox": [ + 511, + 779, + 905, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "23341", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgment", + "text_level": 1, + "bbox": [ + 91, + 90, + 240, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This project is supported by the Ministry of Education, Singapore, under its Academic Research Fund Tier 2 (Award Number: MOE-T2EP20122-0006).", + "bbox": [ + 89, + 114, + 482, + 160 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 172, + 186, + 188 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, et al. Qwen technical report. arXiv preprint arXiv:2309.16609, 2023. 1", + "[2] Fan Bao, Chongxuan Li, Jun Zhu, and Bo Zhang. Analytic-dpm: an analytic estimate of the optimal reverse variance in diffusion probabilistic models. arXiv preprint arXiv:2201.06503, 2022. 3", + "[3] Huiwen Chang, Han Zhang, Lu Jiang, Ce Liu, and William T Freeman. Maskgit: Masked generative image transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11315-11325, 2022. 3", + "[4] Mark Chen, Alec Radford, Rewon Child, Jeffrey Wu, Heewoo Jun, David Luan, and Ilya Sutskever. Generative pretraining from pixels. In International conference on machine learning, pages 1691-1703. PMLR, 2020. 2", + "[5] Zigeng Chen, Gongfan Fang, Xinyin Ma, and Xinchao Wang. $0.1\\%$ data makes segment anything slim. arXiv preprint arXiv:2312.05284, 2023. 3", + "[6] Zigeng Chen, Xinyin Ma, Gongfan Fang, Zhenxiong Tan, and Xinchao Wang. Asyncdiff: Parallelizing diffusion models by asynchronous denoising. arXiv preprint arXiv:2406.06911, 2024.3", + "[7] Tri Dao, Dan Fu, Stefano Ermon, Atri Rudra, and Christopher Ré. Flashattention: Fast and memory-efficient exact attention with io-awareness. Advances in Neural Information Processing Systems, 35:16344-16359, 2022. 6", + "[8] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 4, 5", + "[9] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 1", + "[10] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12873-12883, 2021. 2", + "[11] Gongfan Fang, Xinyin Ma, Mingli Song, Michael Bi Mi, and Xinchao Wang. Depgraph: Towards any structural pruning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16091-16101, 2023. 3", + "[12] Gongfan Fang, Xinyin Ma, and Xinchao Wang. Structural pruning for diffusion models. Advances in neural information processing systems, 36, 2024. 3", + "[13] Jiarui Fang, Jinzhe Pan, Xibo Sun, Aoyu Li, and Jiannan Wang. xedit: an inference engine for diffusion trans" + ], + "bbox": [ + 93, + 198, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "formers (dits) with massive parallelism. arXiv preprint arXiv:2411.01738, 2024. 3", + "[14] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023. 2", + "[15] Jiatao Gu, Yuyang Wang, Yizhe Zhang, Qihang Zhang, Dinghuai Zhang, Navdeep Jaitly, Josh Susskind, and Shuangfei Zhai. Dart: Denoising autoregressive transformer for scalable text-to-image generation. arXiv preprint arXiv:2410.08159, 2024. 2", + "[16] Daya Guo, Qihao Zhu, Dejian Yang, Zhenda Xie, Kai Dong, Wentao Zhang, Guanting Chen, Xiao Bi, Yu Wu, YK Li, et al. Deepseek-coder: When the large language model meets programming-the rise of code intelligence. arXiv preprint arXiv:2401.14196, 2024. 1", + "[17] Wanggui He, Siming Fu, Mushui Liu, Xierui Wang, Wenyi Xiao, Fangxun Shu, Yi Wang, Lei Zhang, Zhelun Yu, Haoyuan Li, et al. Mars: Mixture of auto-regressive models for fine-grained text-to-image synthesis. arXiv preprint arXiv:2407.07614, 2024. 1, 2", + "[18] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 2", + "[19] Geoffrey Hinton. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 3, 5", + "[20] Doohyuk Jang, Sihwan Park, June Yong Yang, Yeonsung Jung, Jihun Yun, Souvik Kundu, Sung-Yub Kim, and Eunho Yang. *Lantern: Accelerating visual autoregressive models with relaxed speculative decoding.* arXiv preprint arXiv:2410.03355, 2024. 3", + "[21] Diederik P Kingma. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 5", + "[22] Haopeng Li, Jinyue Yang, Kexin Wang, Xuerui Qiu, Yuhong Chou, Xin Li, and Guoqi Li. Scalable autoregressive image generation with mamba. arXiv preprint arXiv:2408.12245, 2024. 1, 2, 6", + "[23] Muyang Li, Tianle Cai, Jiaxin Cao, Qinsheng Zhang, Han Cai, Junjie Bai, Yangqing Jia, Kai Li, and Song Han. Distributed parallel inference for high-resolution diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7183-7193, 2024. 3", + "[24] Muyang Li*, Yujun Lin*, Zhekai Zhang*, Tianle Cai, Xiuyu Li, Junxian Guo, Enze Xie, Chenlin Meng, Jun-Yan Zhu, and Song Han. Svdquant: Absorbing outliers by low-rank components for 4-bit diffusion models. arXiv preprint arXiv:2411.05007, 2024.3", + "[25] Senmao Li, Taihang Hu, Fahad Shahbaz Khan, Linxuan Li, Shiqi Yang, Yaxing Wang, Ming-Ming Cheng, and Jian Yang. Faster diffusion: Rethinking the role of unet encoder in diffusion models. arXiv preprint arXiv:2312.09608, 2023. 3", + "[26] Tianhong Li, Yonglong Tian, He Li, Mingyang Deng, and Kaiming He. Autoregressive image generation without vector quantization. arXiv preprint arXiv:2406.11838, 2024. 1, 3, 6" + ], + "bbox": [ + 516, + 92, + 905, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "23342", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[27] Xiuyu Li, Yijiang Liu, Long Lian, Huanrui Yang, Zhen Dong, Daniel Kang, Shanghang Zhang, and Kurt Keutzer. Q-diffusion: Quantizing diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17535-17545, 2023. 3", + "[28] Xiang Li, Hao Chen, Kai Qiu, Jason Kuen, Jiuming Gu, Bhiksha Raj, and Zhe Lin. Imagefolder: Autoregressive image generation with folded tokens. arXiv preprint arXiv:2410.01756, 2024.3", + "[29] Xiang Li, Kai Qiu, Hao Chen, Jason Kuen, Zhe Lin, Rita Singh, and Bhiksha Raj. Controlvar: Exploring controllable visual autoregressive modeling. arXiv preprint arXiv:2406.09750, 2024. 1, 3", + "[30] Yanyu Li, Huan Wang, Qing Jin, Ju Hu, Pavlo Chemerys, Yun Fu, Yanzhi Wang, Sergey Tulyakov, and Jian Ren. Snapfusion: Text-to-image diffusion model on mobile devices within two seconds. Advances in Neural Information Processing Systems, 36, 2024. 3", + "[31] Zongming Li, Tianheng Cheng, Shoufa Chen, Peize Sun, Haocheng Shen, Longjin Ran, Xiaoxin Chen, Wenyu Liu, and Xinggang Wang. Controlar: Controllable image generation with autoregressive models. arXiv preprint arXiv:2410.02705, 2024. 1, 3", + "[32] Dongyang Liu, Shitian Zhao, Le Zhuo, Weifeng Lin, Yu Qiao, Hongsheng Li, and Peng Gao. Lumina-mgpt: Illuminate flexible photorealistic text-to-image generation with multimodal generative pretraining. arXiv preprint arXiv:2408.02657, 2024. 1, 2", + "[33] Luping Liu, Yi Ren, Zhijie Lin, and Zhou Zhao. Pseudo numerical methods for diffusion models on manifolds. arXiv preprint arXiv:2202.09778, 2022. 3", + "[34] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps. Advances in Neural Information Processing Systems, 35:5775-5787, 2022. 3", + "[35] Simian Luo, Yiqin Tan, Longbo Huang, Jian Li, and Hang Zhao. Latent consistency models: Synthesizing high-resolution images with few-step inference. arXiv preprint arXiv:2310.04378, 2023. 3", + "[36] Zhuoyan Luo, Fengyuan Shi, Yixiao Ge, Yujiu Yang, Limin Wang, and Ying Shan. Open-magvit2: An open-source project toward democratizing auto-regressive visual generation. arXiv preprint arXiv:2409.04410, 2024. 3", + "[37] Zhaoyang Lyu, Xudong Xu, Ceyuan Yang, Dahua Lin, and Bo Dai. Accelerating diffusion models via early stop of the diffusion process. arXiv preprint arXiv:2205.12524, 2022. 3", + "[38] Xinyin Ma, Gongfan Fang, and Xinchao Wang. Deepcache: Accelerating diffusion models for free. arXiv preprint arXiv:2312.00858, 2023. 3", + "[39] Xiaoxiao Ma, Mohan Zhou, Tao Liang, Yalong Bai, Tiejun Zhao, Huaian Chen, and Yi Jin. Star: Scale-wise text-to-image generation via auto-regressive representations. arXiv preprint arXiv:2406.10797, 2024. 3", + "[40] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An im" + ], + "bbox": [ + 91, + 90, + 480, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "perative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019. 6", + "[41] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4195-4205, 2023. 6", + "[42] Kai Qiu, Xiang Li, Hao Chen, Jie Sun, Jinglu Wang, Zhe Lin, Marios Savvides, and Bhiksha Raj. Efficient autoregressive audio modeling via next-scale prediction. arXiv preprint arXiv:2408.09027, 2024. 1, 3", + "[43] Tim Salimans and Jonathan Ho. Progressive distillation for fast sampling of diffusion models. arXiv preprint arXiv:2202.00512, 2022. 3", + "[44] Axel Sauer, Dominik Lorenz, Andreas Blattmann, and Robin Rombach. Adversarial diffusion distillation. arXiv preprint arXiv:2311.17042, 2023. 3", + "[45] Yuzhang Shang, Zhihang Yuan, Bin Xie, Bingzhe Wu, and Yan Yan. Post-training quantization on diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1972-1981, 2023. 3", + "[46] Junhyuk So, Jungwon Lee, and Eunhyeok Park. Frdiff: Feature reuse for exquisite zero-shot acceleration of diffusion models. arXiv preprint arXiv:2312.03517, 2023. 3", + "[47] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 6", + "[48] Peize Sun, Yi Jiang, Shoufa Chen, Shilong Zhang, Bingyue Peng, Ping Luo, and Zehuan Yuan. Autoregressive model beats diffusion: Llama for scalable image generation. arXiv preprint arXiv:2406.06525, 2024. 1, 2, 6", + "[49] Haotian Tang, Yecheng Wu, Shang Yang, Enze Xie, Junsong Chen, Junyu Chen, Zhuoyang Zhang, Han Cai, Yao Lu, and Song Han. Hart: Efficient visual generation with hybrid autoregressive transformer. arXiv preprint arXiv:2410.10812, 2024. 1, 3", + "[50] Yao Teng, Han Shi, Xian Liu, Xuefei Ning, Guohao Dai, Yu Wang, Zhenguo Li, and Xihui Liu. Accelerating autoregressive text-to-image generation with training-free speculative jacobi decoding. arXiv preprint arXiv:2410.01699, 2024.3", + "[51] Keyu Tian, Yi Jiang, Zehuan Yuan, Bingyue Peng, and Liwei Wang. Visual autoregressive modeling: Scalable image generation via next-scale prediction. arXiv preprint arXiv:2404.02905, 2024. 1, 3, 6", + "[52] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 1", + "[53] Aaron Van den Oord, Nal Kalchbrenner, Lasse Espeholt, Oriol Vinyals, Alex Graves, et al. Conditional image generation with pixelCNN decoders. Advances in neural information processing systems, 29, 2016. 2", + "[54] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017. 2" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "23343", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[55] Felix Wimbauer, Bichen Wu, Edgar Schoenfeld, Xiaoliang Dai, Ji Hou, Zijian He, Artsiom Sanakoyeu, Peizhao Zhang, Sam Tsai, Jonas Kohler, et al. Cache me if you can: Accelerating diffusion models through block caching. arXiv preprint arXiv:2312.03209, 2023. 3", + "[56] Jinheng Xie, Weijia Mao, Zechen Bai, David Junhao Zhang, Weihao Wang, Kevin Qinghong Lin, Yuchao Gu, Zhijie Chen, Zhenheng Yang, and Mike Zheng Shou. Show-o: One single transformer to unify multimodal understanding and generation. arXiv preprint arXiv:2408.12528, 2024. 1, 2", + "[57] Xingyi Yang and Xinchao Wang. Hash3d: Training-free acceleration for 3d generation. arXiv preprint arXiv:2404.06091, 2024. 3", + "[58] Xingyi Yang, Daquan Zhou, Jiashi Feng, and Xinchao Wang. Diffusion probabilistic model made slim. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22552-22562, 2023. 3", + "[59] Ziyu Yao, Jialin Li, Yifeng Zhou, Yong Liu, Xi Jiang, Chengjie Wang, Feng Zheng, Yuexian Zou, and Lei Li. Car: Controllable autoregressive modeling for visual generation. arXiv preprint arXiv:2410.04671, 2024. 1", + "[60] Tianwei Yin, Michael Gharbi, Richard Zhang, Eli Shechtman, Fredo Durand, William T Freeman, and Taesung Park. One-step diffusion with distribution matching distillation. arXiv preprint arXiv:2311.18828, 2023. 3", + "[61] Zongsheng Yue, Jianyi Wang, and Chen Change Loy. Reshift: Efficient diffusion model for image superresolution by residual shifting. Advances in Neural Information Processing Systems, 36, 2024. 3", + "[62] Dingkun Zhang, Sijia Li, Chen Chen, Qingsong Xie, and Haonan Lu. Laptop-diff: Layer pruning and normalized distillation for compressing diffusion models. arXiv preprint arXiv:2404.11098, 2024. 3", + "[63] Jinzhi Zhang, Feng Xiong, and Mu Xu. G3pt: Unleash the power of autoregressive modeling in 3d generation via cross-scale querying transformer. arXiv preprint arXiv:2409.06322, 2024. 1, 3", + "[64] Qinsheng Zhang and Yongxin Chen. Fast sampling of diffusion models with exponential integrator. arXiv preprint arXiv:2204.13902, 2022. 3", + "[65] Qian Zhang, Xiangzi Dai, Ninghua Yang, Xiang An, Ziyong Feng, and Xingyu Ren. Var-clip: Text-to-image generator with visual auto-regressive modeling. arXiv preprint arXiv:2408.01181, 2024. 1, 3", + "[66] Wentian Zhang, Haozhe Liu, Jinheng Xie, Francesco Faccio, Mike Zheng Shou, and Jürgen Schmidhuber. Cross-attention makes inference cumbersome in text-to-image diffusion models. arXiv preprint arXiv:2404.02747, 2024. 3", + "[67] Xuanlei Zhao, Xiaolong Jin, Kai Wang, and Yang You. Real-time video generation with pyramid attention broadcast. arXiv preprint arXiv:2408.12588, 2024. 3", + "[68] Yang Zhao, Yanwu Xu, Zhisheng Xiao, and Tingbo Hou. Mobilediffusion: Subsecond text-to-image generation on mobile devices. arXiv preprint arXiv:2311.16567, 2023. 3", + "[69] Kaiwen Zheng, Cheng Lu, Jianfei Chen, and Jun Zhu. Dpm solver-v3: Improved diffusion ode solver with empirical" + ], + "bbox": [ + 91, + 90, + 482, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "model statistics. Advances in Neural Information Processing Systems, 36, 2024. 3", + "[70] Chunting Zhou, Lili Yu, Arun Babu, Kushal Tirumala, Michihiro Yasunaga, Leonid Shamis, Jacob Kahn, Xuezhe Ma, Luke Zettlemoyer, and Omer Levy. Transfusion: Predict the next token and diffuse images with one multi-modal model. arXiv preprint arXiv:2408.11039, 2024. 1, 2" + ], + "bbox": [ + 516, + 92, + 903, + 190 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "23344", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Collaborative Decoding Makes Visual Auto-Regressive Modeling Efficient/9a935c24-e29c-43fe-a54b-aa3325d4f954_model.json b/2025/Collaborative Decoding Makes Visual Auto-Regressive Modeling Efficient/9a935c24-e29c-43fe-a54b-aa3325d4f954_model.json new file mode 100644 index 0000000000000000000000000000000000000000..be359a951758d2d55c72c835edb3357f7c44c95e --- /dev/null +++ b/2025/Collaborative Decoding Makes Visual Auto-Regressive Modeling Efficient/9a935c24-e29c-43fe-a54b-aa3325d4f954_model.json @@ -0,0 +1,2257 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.129, + 0.131, + 0.872, + 0.154 + ], + "angle": 0, + "content": "Collaborative Decoding Makes Visual Auto-Regressive Modeling Efficient" + }, + { + "type": "text", + "bbox": [ + 0.26, + 0.181, + 0.744, + 0.218 + ], + "angle": 0, + "content": "Zigeng Chen, Xinyin Ma, Gongfan Fang, Xinchao Wang* National University of Singapore" + }, + { + "type": "text", + "bbox": [ + 0.239, + 0.22, + 0.76, + 0.235 + ], + "angle": 0, + "content": "{zigeng99, maxinyin, gongfan}@u.nus.edu, xinchao@nus.edu.sg" + }, + { + "type": "text", + "bbox": [ + 0.312, + 0.236, + 0.68, + 0.252 + ], + "angle": 0, + "content": "https://github.com/czg1225/CoDe" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.287, + 0.327, + 0.303 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.32, + 0.485, + 0.744 + ], + "angle": 0, + "content": "In the rapidly advancing field of image generation, Visual Auto-Regressive (VAR) modeling has garnered considerable attention for its innovative next-scale prediction approach. This paradigm offers substantial improvements in efficiency, scalability, and zero-shot generalization. Yet, the inherently coarse-to-fine nature of VAR introduces a prolonged token sequence, leading to prohibitive memory consumption and computational redundancies. To address these bottlenecks, we propose Collaborative Decoding (CoDe), a novel efficient decoding strategy tailored for the VAR framework. CoDe capitalizes on two critical observations: the substantially reduced parameter demands at larger scales and the exclusive generation patterns across different scales. Based on these insights, we partition the multi-scale inference process into a seamless collaboration between a large model and a small model. The large model serves as the 'drafter', specializing in generating low-frequency content at smaller scales, while the smaller model serves as the 'refiner', solely focusing on predicting high-frequency details at larger scales. This collaboration yields remarkable efficiency with minimal impact on quality: CoDe achieves a \\(1.7x\\) speedup, slashes memory usage by around \\(50\\%\\), and preserves image quality with only a negligible FID increase from 1.95 to 1.98. When drafting steps are further decreased, CoDe can achieve an impressive \\(2.9x\\) acceleration ratio, reaching 41 images/s at 256x256 resolution on a single NVIDIA 4090 GPU, while preserving a commendable FID of 2.27." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.774, + 0.222, + 0.79 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.8, + 0.484, + 0.877 + ], + "angle": 0, + "content": "The past year has witnessed significant advancements in Auto-Regressive (AR) models for image generation [17, 22, 26, 31, 48, 59], driven by their proven scalability and strong generalization capabilities [1, 9, 16, 52]. Leveraging these strengths, AR approaches have demonstrated re" + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.287, + 0.9, + 0.415 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.524, + 0.414, + 0.897, + 0.543 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.556, + 0.908, + 0.584 + ], + "angle": 0, + "content": "Figure 1. We partition the next-scale prediction process into the efficient collaboration between large and small VAR models." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.613, + 0.907, + 0.764 + ], + "angle": 0, + "content": "markable image quality and promising multi-modal potential [32, 56, 70]. However, the next-token prediction process inherent to conventional AR models requires numerous decoding steps, leading to considerable generation latency. Visual Auto-Regressive (VAR) Modeling [51] replaces the GPT-style next-token prediction with a next-scale prediction strategy. It generates content in a multi-level, coarse-to-fine progression, enabling the model to decode multiple tokens in parallel [29, 42, 49, 63, 65], thus considerably reducing the inference steps." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Despite less decoding steps, VAR's progressive scaling approach significantly increases the overall sequence length. To generate a 16x16 token image, VAR hierarchically decodes up to 680 tokens across 10 scales—2.7 times the sequence length required by conventional AR models. As the VAR model must store the KV cache accumulated from all previous tokens, this prolonged sequence results in a substantial memory overhead, especially in the final scales. During inference, the KV cache consumes approx-" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.235, + 0.901 + ], + "angle": 0, + "content": "*Corresponding author" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23334" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.127, + 0.092, + 0.905, + 0.281 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.289, + 0.908, + 0.32 + ], + "angle": 0, + "content": "Figure 2. Comparison of generation results between original VAR-d30 (up) and our VAR-CoDE (bottom) for ImageNet \\(256 \\times 256\\). Our method achieves 1.7x speedup (3.62s to 2.11s), and needs only 0.5x memory space (40GB to 20GB), with negligible quality degradation." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.344, + 0.482, + 0.451 + ], + "angle": 0, + "content": "imately 12 times more memory than the forward computation itself. For instance, generating images using VAR-d30 with a batch size of 128 demands 70 GB of memory, with 57 GB dedicated solely to KV caching, becoming the primary bottleneck. Moreover, the extended sequence length exacerbates the computational cost of self-attention, given the quadratic growth in attention map calculations." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.457, + 0.483, + 0.608 + ], + "angle": 0, + "content": "To address these inefficiencies, we begin by analyzing the specific properties of VAR's next-scale prediction paradigm. First, we observe that as generated scales increase, the parameter demands for high-quality token generation drop substantially, leading to considerable computational redundancy at most of the tokens in the long sequence. Next, we find that the generation patterns at small and large scales are exclusive, resulting in mutual interference between learning low- and high-frequency modeling capability, thus hindering efficient parameter utilization." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.614, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Our Approach. Inspired by the above observations, we propose Collaborative Decoding (CoDe), a simple yet highly effective method that significantly boosts the inference efficiency of VAR models while preserving generation quality comparable to the original. Our core idea is to decompose the long-sequence scaling-up process into a collaboration between two VAR models with different sizes and specialized roles to enhance efficiency. Figure 1 presents the overview of our method. We use a large VAR model as a drafter for the initial small scales, where model capacity demands are high but computations and KV cache are sparse. Conversely, we use a small VAR model as a refiner at the remaining large scales, where computations and KV cache are intensive but fewer parameters are needed. Next, we apply specialized fine-tuning to both models to address the optimization interference between scales encountered during the pre-training stage. Each model is fine-tuned exclusively on the specific scales it handles, leading to a notable performance boost with limited additional cost." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.344, + 0.905, + 0.465 + ], + "angle": 0, + "content": "Extensive experiments validate the effectiveness of our method. As shown in Figure 2, compared to the original VAR-d30 model, CoDe achieves a \\(1.7\\mathrm{x}\\) speedup with merely \\(0.5\\mathrm{x}\\) GPU memory consumption and negligible quality loss (FID [18] slightly increases from 1.95 to 1.98). Notably, our approach can reach up to a \\(2.9\\mathrm{x}\\) speedup, generating 41 images (256x256) per second on a single NVIDIA 4090 GPU, while still maintaining an FID of 2.27." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.466, + 0.907, + 0.648 + ], + "angle": 0, + "content": "In conclusion, we introduce CoDe, a novel decoding framework for visual auto-regressive modeling that significantly enhances speed and reduces memory overhead with a negligible impact on quality. CoDe divides the hierarchical sequence modeling of VAR into a collaborative process, utilizing a large VAR model and a small VAR model in a progressive partnership. Additionally, we propose specialized fine-tuning to optimize each model for its specific role, effectively mitigating training interference and maximizing parameter utilization. To the best of our knowledge, CoDe is the fastest method available to achieve an FID below 2, making it a clear advancement in efficient image generation." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.663, + 0.665, + 0.68 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Auto-regressive image generation. Early works [4, 53] pioneered image generation by generating pixels in raster-scan order. Later, VQVAE [54] and VQGAN [10] improved this approach by quantizing image patches into discrete tokens, using a transformer in a decoder-only setup to generate these tokens in a raster-scan manner. Building on these foundations, recent efforts have focused on enhancing autoregressive (AR) models for image generation. LlamaGEN [48] and Lumina-mGPT [32], for example, use a GPT-style next-token prediction strategy to achieve high-quality image generation with good scalability. AiM [22] and MARS [17] further improve this paradigm by introducing mixture-of-experts and linear attention mechanisms [14]. Methods like SHOW-O [56], Transfusion [70], and DART [15]" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23335" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.09, + 0.352, + 0.221 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.119, + 0.222, + 0.355, + 0.235 + ], + "angle": 0, + "content": "(a) Quality vs. Params at different scales" + }, + { + "type": "image", + "bbox": [ + 0.362, + 0.089, + 0.616, + 0.221 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.439, + 0.223, + 0.566, + 0.236 + ], + "angle": 0, + "content": "(b) Spectrum analysis" + }, + { + "type": "image", + "bbox": [ + 0.619, + 0.089, + 0.871, + 0.22 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.655, + 0.222, + 0.871, + 0.236 + ], + "angle": 0, + "content": "(c) Training-free model collaboration" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.246, + 0.908, + 0.304 + ], + "angle": 0, + "content": "Figure 3. (a) Effectiveness of increasing parameters at the \\(k\\)-th scale is evaluated by predicting token map \\(r_k\\) using four VAR models with different parameter sizes (2B, 1B, 0.6B, and 0.3B), while other scales \\((r_1, r_2, \\ldots, r_{k-1}, r_{k+1}, \\ldots, r_{10})\\) are generated using the largest VAR-d30 model. (b) Fourier spectrum analysis is conducted on generated content at the first 3 scales and the last 3 scales. (c) Training-free performance comparison of model collaboration decoding across various settings of draft tokens \\(M\\) and refiner tokens \\(680 - M\\)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.329, + 0.484, + 0.496 + ], + "angle": 0, + "content": "combine diffusion processes with autoregressive modeling, while [3, 26, 36] introduces masked autoregression to generate images. However, all these approaches suffer from high latency due to the large number of forward steps. VAR [51] addresses this by using hierarchical parallel decoding, which greatly reduces the number of steps, leading to significantly lower inference latency without sacrificing quality. Building on VAR's progress, many recent works have aimed to improve next-scale prediction across multiple tasks, including text-to-image [28, 39, 49, 65], controllable generation [29, 31], audio generation [42], and 3D generation [63]." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.501, + 0.484, + 0.804 + ], + "angle": 0, + "content": "Efficient Image Generation Models. For diffusion models, acceleration techniques are already well-developed. [35, 43, 44, 60, 61] focuses on reducing sampling steps while [12, 30, 58, 62, 68] optimize model architectures through pruning [5, 11] or knowledge distillation [19]. To avoid the high costs of training, some training-free methods are proposed. Some approaches develop fast solvers for stochastic differential equations (SDE) or ordinary differential equations (ODE) to enhance sampling efficiency [2, 33, 34, 64, 69]. Other works [25, 37, 38, 46, 55, 57, 66, 67] exploit specific characteristics of diffusion models to skip redundant computations during the denoising process. [6, 13, 23] reduce latency to another level through distributed computing. Quantization methods [24, 27, 45] have also shown great potential. In contrast, research into accelerating AR image generation is still in its early stages. LANTERN [20] and SJD [50] employ speculative decoding to speed up next-token prediction, achieving notable speedups but not suitable for the innovative next-scale prediction paradigm introduced by VAR." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.484, + 0.902 + ], + "angle": 0, + "content": "In this paper, we introduce CoDe, a novel and efficient decoding method tailored for the next-scale prediction paradigm. CoDe effectively addresses the inefficiencies associated with the long sequence structure of VAR models while maintaining high generation quality. Notably, CoDe stands out as the fastest method capable of achieving an FID" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.329, + 0.782, + 0.344 + ], + "angle": 0, + "content": "below 2, making it a clear advancement." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.356, + 0.606, + 0.372 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.381, + 0.644, + 0.398 + ], + "angle": 0, + "content": "3.1. Preliminary" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.403, + 0.907, + 0.54 + ], + "angle": 0, + "content": "Visual auto-regressive modeling [51] redefines the traditional AR by shifting from a \"next-token\" prediction to a \"next-scale\" prediction. In this framework, each auto-regressive unit is a token map at varying scales, rather than a single token. For a given image feature map \\( f \\in \\mathbb{R}^{h \\times w \\times C} \\), VAR quantizes it into \\( K \\) multi-scale token maps \\( R = (r_1, r_2, \\ldots, r_K) \\) at progressively finer resolutions, with the final token map \\( r_K \\) matching the original feature map's resolution. The probability distribution is reformulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.535, + 0.548, + 0.907, + 0.59 + ], + "angle": 0, + "content": "\\[\np \\left(r _ {1}, r _ {2}, \\dots , r _ {K}\\right) = \\prod_ {k = 1} ^ {K} p \\left(r _ {k} \\mid r _ {1}, r _ {2}, \\dots , r _ {k - 1}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.599, + 0.909, + 0.737 + ], + "angle": 0, + "content": "where each token map \\( r_k \\in [V]^{h_k \\times w_k} \\) consists of \\( h_k \\times w_k \\) tokens at scale \\( k \\), and the sequence \\( (r_1, r_2, \\ldots, r_{k-1}) \\) serves as the \"prefix\" for \\( r_k \\). In this paradigm, during each autoregressive step \\( k \\), the model predicts all \\( h_k \\times w_k \\) tokens in \\( r_k \\) in parallel, conditioned on prior scales and position embeddings. This approach aligns with a coarse-to-fine generation pattern, enabling parallel decoding within each scale. VAR effectively improves inference speed and generation quality, but it also considerably increases sequence length." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.743, + 0.688, + 0.759 + ], + "angle": 0, + "content": "3.2. Key Observations" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.907, + 0.87 + ], + "angle": 0, + "content": "VAR represents an innovative paradigm specific for visual generation, distinct from traditional autoregressive approaches and introducing many yet unexplored characteristics. In this work, we revisit the entire next-scale prediction process to uncover specific properties that can be optimized to reduce computational redundancy and improve inference efficiency." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Observation 1: As the predicted scale becomes larger, the need for parameters reduces significantly. While VAR" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "23336" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.098, + 0.09, + 0.614, + 0.309 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.629, + 0.129, + 0.904, + 0.3 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.32, + 0.908, + 0.381 + ], + "angle": 0, + "content": "Figure 4. Overview of the collaborative decoding process, we use a drafting step \\( N = 6 \\) for instance. CoDe uses a large VAR model as the drafter \\( \\epsilon_{\\theta_d} \\) to generate the token maps \\( R_L = (r_1, r_2, \\dots, r_N) \\) at smaller scales. The small refiner model \\( \\epsilon_{\\theta_r} \\) then uses \\( R_L \\) as an initial prefix to efficiently predict the remaining token maps \\( R_H = (r_{N+1}, r_{N+2}, \\dots, r_K) \\) at larger scales. Both models are fine-tuned on their designated predictive scales using ground truth labels \\( r_k^* \\) and teacher logits \\( p_{\\text{teacher}}(r_k) \\), respectively." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.403, + 0.486, + 0.645 + ], + "angle": 0, + "content": "models exhibit strong scalability, the effectiveness of increasing parameter counts varies greatly in predicting various scales. In Figure 3. (a), we present the differences in class-conditional generation (ImageNet-256 [8]) quality when using VAR models of different sizes at each scale. At the initial small scales, increasing model parameters leads to noticeable improvements in generation quality. However, as the predicted scale gradually progresses, the impact of additional parameters becomes minimal. At the final scale—which accounts for \\(38\\%\\) tokens of the entire sequence—we observe that the performance of the 2B model is even nearly equivalent to that of the 0.3B model. These findings suggest that as the predicted scale increases, the parameter demand for accurate token predictions declines markedly, revealing substantial computational redundancy in the current VAR inference process at large scales." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.66, + 0.486, + 0.902 + ], + "angle": 0, + "content": "Observation 2: The generative patterns are exclusive between small and large scales. For VAR models, the content generated at small and large scales varies significantly. As demonstrated by the Fourier analysis in Figure 3 (b), the feature maps produced by the first three scales contain primarily low-frequency components, while the last three scales focus on high-frequency components. To further validate these distinct generative patterns, we conducted a perturbation fine-tuning experiment. Using a pre-trained VARd16 model, we applied CSE loss only to tokens in the largest three scales and fine-tuned for just \\(1\\%\\) of the original training epochs. This minor fine-tuning at large scales led to a complete collapse of the model's global modeling capacity at small scales, with the FID increasing from 3.30 to 21.93 and the IS score dropping from 277 to 88. These results demonstrate that VAR models perform entirely dis" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.403, + 0.907, + 0.478 + ], + "angle": 0, + "content": "tinct generative tasks at small and large scales, with minimal overlap. Training a single VAR model to predict across all scales results in significant mutual interference between learning low-frequency and high-frequency modeling capabilities, making parameter optimization challenging." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.479, + 0.909, + 0.603 + ], + "angle": 0, + "content": "In sum, the key insights from our observations can be concluded as follows: (1) At large scales, VAR models require significantly fewer parameters for accurate token prediction, leading to considerable computational redundancy. (2) Training a single model to generate across all scales causes mutual interference between small and large scales, hindering parameter optimization due to conflicting low-frequency and high-frequency learning tendencies." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.607, + 0.733, + 0.624 + ], + "angle": 0, + "content": "3.3. Collaborative Decoding" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.629, + 0.909, + 0.825 + ], + "angle": 0, + "content": "Based on the above insights, we propose a simple yet pretty efficient decoding method for next-scale prediction called collaborative decoding. As presented in Figure 4, CoDe decomposes the next-scale prediction process into a collaboration between a large drafter model and a small refiner model. The drafter is the original large VAR model with 2B parameters, while the refiner is a lightweight 0.3B VAR model, sharing the same transformer-based architecture but with reduced width and depth. The drafter is responsible for generating coarse, low-frequency global structures at smaller scales to draft the image, while the refiner continually predicts the high-frequency details at larger scales to refine the image." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Model Collaboration. To generate an image \\(im\\), original VAR model needs to auto-regressively generate \\(K\\) multiscale token maps \\(R = (r_1, r_2, \\ldots, r_K)\\) with progressively finer resolutions, where each token map \\(r_k \\in [V]^{h_k \\times w_k}\\) represents the token map at scale \\(k\\). In CoDe, we decom" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23337" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.228 + ], + "angle": 0, + "content": "pose the auto-regressive steps into \\(N\\) drafting steps and \\(K - N\\) refining steps. In the initial drafting phase where computation is sparse but additional parameters provide significant benefits, we serve a large VAR model as the drafter \\(\\epsilon_{\\theta_d}\\), generating the initial set of low-frequency token maps \\(R_L = (r_1,r_2,\\dots ,r_N)\\) up to scale \\(N\\), where \\(N < K\\). These token maps predict the global structure of the image, serving as a prefix for the remaining finer-scale token maps. The drafting process can be represented as:" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.234, + 0.484, + 0.275 + ], + "angle": 0, + "content": "\\[\np _ {\\theta_ {d}} \\left(r _ {1}, r _ {2}, \\dots , r _ {N}\\right) = \\prod_ {k = 1} ^ {N} p _ {\\theta_ {d}} \\left(r _ {k} \\mid r _ {1}, r _ {2}, \\dots , r _ {k - 1}\\right). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.28, + 0.483, + 0.31 + ], + "angle": 0, + "content": "After drafting, the KV cache of the large model is released, significantly optimizing memory usage." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.311, + 0.484, + 0.462 + ], + "angle": 0, + "content": "In the refining stage where the computation is intensive but the parameter demand is low, we serve a small VAR model as the refiner \\(\\epsilon_{\\theta_r}\\), employing the drafter's predictions \\(R_L\\) as a prefix and focus on refining the details by generating the remaining high-frequency token maps \\(R_H = (r_{N + 1}, r_{N + 2}, \\ldots, r_K)\\) up to scale \\(K\\), where \\(r_K\\) matches the resolution of the original feature map. Notably, predicting \\(r_{N + 1}\\) requires an attention mask, as the refiner model lacks the KV cache for the first \\(N\\) scales. The probability distribution of refining steps is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.139, + 0.467, + 0.483, + 0.529 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} p \\left(r _ {N + 1}, r _ {N + 2}, \\dots , r _ {K} \\mid R _ {L}\\right) \\\\ = \\prod_ {k = N + 1} ^ {K} p \\left(r _ {k} \\mid R _ {L}, r _ {N + 1}, r _ {N + 2}, \\dots , r _ {k - 1}\\right). \\tag {3} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.535, + 0.484, + 0.566 + ], + "angle": 0, + "content": "Finally, the generated images \\(im\\) are reconstructed from both drafting maps \\(R_{L}\\) and refining maps \\(R_{H}\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.211, + 0.579, + 0.483, + 0.596 + ], + "angle": 0, + "content": "\\[\ni m = D \\left(Q \\left(R _ {L}, R _ {H}\\right), \\right. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.599, + 0.483, + 0.749 + ], + "angle": 0, + "content": "where \\( Q(.) \\) is a residual-style quantization function, and \\( D(.) \\) is a multi-scale VQVAE decoder. Figure 3. (c) demonstrates the remarkable effectiveness of model collaboration under training-free conditions. With a large model as the drafter, the small refiner only results in a slight and gentle increase in FID. Conversely, with a small model as the drafter, even assigning the last \\( 80\\% \\) tokens to a large refiner fails to improve performance. This also verifies Observation 1: small scales require more model capacity, whereas the parameter demands for large scales are low." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.811 + ], + "angle": 0, + "content": "The model collaboration method effectively reduces the computational redundancy of VAR and maintains generation quality comparable to the original model, while significantly offering faster speed and reduced memory usage." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Specialized Fine-Tuning. Given the exclusive generation patterns between drafting scales and refining scales, we propose the specialized fine-tuning to further specialize drafter and refiner models by fine-tuning them solely on their respective predictive scales, thereby avoiding training interference and enhancing generation quality." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.169 + ], + "angle": 0, + "content": "For the drafter model \\(\\epsilon_{\\theta_d}\\), we fine-tune the drafting steps to improve the generation of the initial low-frequency token maps \\(R_L\\). The drafter model is trained using a CSE loss \\(\\mathrm{CSE}(\\cdot, \\cdot)\\) between its generated token distribution \\(p_{\\theta_d}(r_k)\\) and the ground truth labels \\(r_k^*\\), defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.602, + 0.177, + 0.906, + 0.218 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {d r a f t e r}} = \\sum_ {k = 1} ^ {N} \\mathrm {C S E} \\left(p _ {\\theta_ {d}} \\left(r _ {k}\\right), r _ {k} ^ {*}\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.227, + 0.906, + 0.256 + ], + "angle": 0, + "content": "which encourages the drafter to closely match the target global structure at each drafting step." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.257, + 0.906, + 0.317 + ], + "angle": 0, + "content": "For the refiner model \\(\\epsilon_{\\theta_r}\\), we adopt a Knowledge Distillation (KD) [19] approach to partially transfer the knowledge from a larger pre-trained VAR model. The KD loss function for epoch \\(ep\\) can be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.514, + 0.325, + 0.909, + 0.373 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {r e f i n e r}} ^ {(e p)} = \\sum_ {k = 1} ^ {K} \\left(\\lambda_ {e p} \\cdot \\mathbf {1} _ {[ k \\leq N ]} + \\mathbf {1} _ {[ k > N ]}\\right) \\mathrm {K L} \\left(p _ {\\theta_ {r}} \\left(r _ {k}\\right) \\| p _ {\\text {t e a c h e r}} \\left(r _ {k}\\right)\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.374, + 0.906, + 0.51 + ], + "angle": 0, + "content": "where \\( p_{\\mathrm{teacher}}(\\cdot) \\) is the distribution predicted by the larger teacher VAR model, \\( \\mathrm{KL}(\\cdot, \\cdot) \\) is the standard KL-divergence loss aligns refiner's output distribution with that of the teacher model, and \\( \\lambda_{ep} \\) is a dynamic weighting factor. \\( \\lambda_t \\) linearly decreases from one to zero from the initial to the end of finetuning, gradually shifting the KD emphasis from all tokens to the specific refining tokens. This dynamic weighting allows the refiner to specialize in refining the details with a smoother and more stable training process." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.51, + 0.906, + 0.601 + ], + "angle": 0, + "content": "The specialized fine-tuning of the drafter and refiner models ensures that each model becomes highly proficient in its respective task. By minimizing interference between small and large scales, this approach allows for more sufficient parameter optimization, resulting in enhanced image generation quality with limited training costs." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.613, + 0.72, + 0.63 + ], + "angle": 0, + "content": "4. Experimental Results" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.638, + 0.709, + 0.654 + ], + "angle": 0, + "content": "4.1. Experimental Setup." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.659, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Implementation Details. We evaluated our method's effectiveness on the ImageNet1K [8] class-conditional generation benchmarks. The proposed CoDe framework involves collaboration between a large and a small VAR model. Specifically, we use the pre-trained VAR-d30 model as the drafter and the pre-trained VAR-d16 model as the refiner. Each model undergoes specialized fine-tuning focused exclusively on its respective predictive scales. For the drafter, we fine-tune it for \\(5\\%\\) of its original training epochs with a base learning rate of 1e-6 and a weight decay of 0.08. The refiner model, on the other hand, is fine-tuned for \\(25\\%\\) of its original training epochs with a base learning rate of 1e-5 and no weight decay. Both models are optimized using the AdamW [21] optimizer with a batch size of 1024, achieved through gradient accumulation. The fine-tuning was conducted on 4 NVIDIA L20 GPUs." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "23338" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.089, + 0.359, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.14, + 0.226, + 0.343, + 0.239 + ], + "angle": 0, + "content": "(a) Efficiency & Quality trade-off" + }, + { + "type": "image", + "bbox": [ + 0.367, + 0.089, + 0.632, + 0.223 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.397, + 0.225, + 0.625, + 0.239 + ], + "angle": 0, + "content": "(b) Time cost with different batchsizes" + }, + { + "type": "image", + "bbox": [ + 0.633, + 0.089, + 0.896, + 0.223 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.699, + 0.225, + 0.858, + 0.238 + ], + "angle": 0, + "content": "(c) Time cost at each scale" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.25, + 0.907, + 0.293 + ], + "angle": 0, + "content": "Figure 5. (a) Our CoDe demonstrates the optimal efficiency-quality trade-off among all evaluated methods. (b) Inference latency is measured across varying batch sizes for the original VAR-d30, our CoDe \\((N = 6)\\), and the VQVAE decoder. (c) We analyze the time cost associated with parallel decoding at each scale, showing that the refiner model is significantly more efficient than the drafter at larger scales." + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.307, + 0.907, + 0.336 + ], + "angle": 0, + "content": "Table 1. Quantitative assessment of the efficiency-quality trade-off across various methods. Inference efficiency is evaluated with a batch size of 64 on NVIDIA L20 GPU, with latency measured excluding VQVAE or VQGAN as it incurs a shared time cost across all methods." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.346, + 0.907, + 0.564 + ], + "angle": 0, + "content": "
MethodInference EfficiencyGeneration Quality
#StepsSpeedup↑Latency↓Throughput↑#ParamMemory↓FID ↓IS ↑Precision↑Recall↑
DiT-XL/2500.2x19.20s3.33it/s675M11369MB2.262390.800.60
MAR-B1000.1x29.80s2.15it/s208M8725MB2.312820.820.57
AiM-XL2560.4x9.32s6.87it/s763M20983MB2.562570.820.57
LlamaGen-XXL384<0.1x73.97s0.87it/s1.4B42632MB2.342540.800.59
VAR-d30101.0x3.62s17.71it/s2.0B39228MB1.953010.810.59
VAR-d24101.7x2.07s30.92it/s1.0B25093MB2.113110.820.59
VAR-d20102.8x1.29s49.62it/s600M17814MB2.613010.830.56
VAR-CoDe N=99+11.2x2.97s21.54it/s2.0+0.3B28803MB1.942960.800.61
VAR-CoDe N=88+21.7x2.11s30.33it/s2.0+0.3B21019MB1.983020.810.60
VAR-CoDe N=77+32.3x1.60s40.00it/s2.0+0.3B19943MB2.113030.820.59
VAR-CoDe N=66+42.9x1.27s50.39it/s2.0+0.3B19943MB2.272970.820.58
" + }, + { + "type": "table_caption", + "bbox": [ + 0.165, + 0.587, + 0.411, + 0.602 + ], + "angle": 0, + "content": "Table 2. Effect of specialized fine-tuning" + }, + { + "type": "table", + "bbox": [ + 0.104, + 0.613, + 0.468, + 0.672 + ], + "angle": 0, + "content": "
DiscriptionN=6N=7N=8N=9
CoDe Training-free2.422.262.101.99
CoDe Fine-tuning2.272.111.981.94
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.704, + 0.486, + 0.901 + ], + "angle": 0, + "content": "**Evaluations.** We evaluated our method on ImageNet1K-256 generation, focusing on both quality and efficiency metrics. For quality assessment, we used standard metrics such as FID, Inception Score (IS), Precision, and Recall. For other baselines, we use their default sampling methods. For CoDe, we reduced the sampled top-k from 900 to 600 due to improved token prediction accuracy after fine-tuning, while keeping the default top-p at 0.96. To compensate for reduced diversity, we introduced a temperature coefficient of \\( t = 1.1 \\) for sampling on the smallest 7 scales. Efficiency was measured through inference latency, throughput, memory consumption, and parameter count to provide a comprehensive comparison. Notably, the speed measurements" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.589, + 0.907, + 0.665 + ], + "angle": 0, + "content": "exclude the VQGAN decoder, which contributes minimally to overall runtime and is a shared component across all methods. All efficiency tests were conducted on a single NVIDIA L20 GPU without additional optimizations such as FlashAttention [7], using PyTorch 2.1 [40] and FP16." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.679, + 0.657, + 0.694 + ], + "angle": 0, + "content": "4.2. Main Results." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.703, + 0.907, + 0.854 + ], + "angle": 0, + "content": "Quality-Efficiency Trade-off. We evaluated the quality-efficiency trade-off of our proposed method (CoDe) against original VAR models [51], state-of-the-art AR models, and diffusion models. The AR models included GPT-style LlamaGEN [48], Mamba-based Aim [22], and MAE-style MAR [26]. For diffusion models, we employed the widely used DiT models [41]. To measure the trade-offs, we used series models with different parameter counts for VAR, LlamaGEN, Aim, and MAR, while for DiT we varied the DDIM [47] steps to generate the efficiency-quality curve." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.906, + 0.901 + ], + "angle": 0, + "content": "As illustrated in Table 1, We analyzed CoDe's performance under different drafting steps N. Compared to traditional AR models, our proposed VAR-CoDe \\((N = 6)\\) achieves" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "23339" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.09, + 0.904, + 0.341 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.108, + 0.353, + 0.888, + 0.369 + ], + "angle": 0, + "content": "Figure 6. Qualitative comparison between the original VAR-d30 model and our proposed CoDe model, with different drafting steps." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.394, + 0.483, + 0.606 + ], + "angle": 0, + "content": "60 times faster inference than LlamaGEN-XXL, while also surpassing its generation quality. Additionally, in comparison to diffusion models, VAR-CoDe \\((N = 6)\\) is 15 times faster than DiT-XL/2, while maintaining the same level of quality. Compared to the original VAR-d30, CoDe \\((N = 8)\\) achieves a \\(1.7\\mathrm{x}\\) speedup and reduces memory consumption by \\(50\\%\\), with only a negligible FID increase from 1.95 to 1.98. When the drafting stage is reduced to just 6 steps, CoDe \\((N = 6)\\) achieves a \\(2.9\\mathrm{x}\\) speedup, reaching a throughput of over 50it/s, while maintaining a low FID of 2.27. This is a speed unmatched by any other existing methods. Compared to other VAR models with fewer parameters (VAR-d24 and VAR-d20), our method achieves significantly better quality while maintaining the same speedup ratio." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.608, + 0.483, + 0.671 + ], + "angle": 0, + "content": "As shown in Figure 5 (a), VAR-CoDe achieves the best efficiency-quality trade-off compared to all other methods, effectively solving the inefficiencies introduced by the extended sequence length in the original VAR paradigm." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.671, + 0.483, + 0.883 + ], + "angle": 0, + "content": "Training-Free vs Specialized Finetuning. Our CoDe framework can operate in a training-free manner by directly using pre-trained VAR-d30 and VAR-d16 models as the drafter and refiner, respectively. Alternatively, we can perform specialized fine-tuning to further enhance the models' performance at their respective scales, with limited training cost. Table 2 compares the results of training-free CoDe with the specialized fine-tuned version. Even without additional training, CoDe demonstrates competitive performance, outperforming VAR-d24 and VAR-d20 models with the same speedup ratio. With specialized fine-tuning, CoDe's performance improves significantly, even achieving a slight quality enhancement over the original model at a \\(1.2\\mathrm{x}\\) acceleration ratio." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.886, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Qualitative Results. We provide an extensive qualitative" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.394, + 0.907, + 0.576 + ], + "angle": 0, + "content": "comparison between the original VAR-d30 model and our proposed CoDe, with varying drafting steps \\( N = \\{6,7,8\\} \\). As illustrated in Figure 6, our approach achieves significant speedup and substantial memory optimization, with only minimal quality degradation that is nearly imperceptible to the human eye. Even at a speedup rate of 2.9 times, the generated images maintain high quality and accurate semantic information. Through specialized fine-tuning, CoDe's draft model exhibits higher predictive accuracy compared to the original model. This sometimes results in a different global structure from the original output, yet the image quality remains consistently high or even better." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.577, + 0.909, + 0.789 + ], + "angle": 0, + "content": "Zero-shot Task Generalization. To evaluate the zero-shot generalization capability of CoDe, we conducted additional experiments on zero-shot class-conditional inpainting and image editing. During image inpainting, we applied teacher forcing by providing the ground truth tokens outside the masked area, allowing the model to generate tokens solely within the mask. Notably, class-conditional information was introduced to the model. In the image editing task, CoDe was restricted to generating tokens within a given bounding box based on a specific class label. Figure 7 illustrates the qualitative results of CoDe's zero-shot performance \\((N = 8)\\). Our approach demonstrates strong zero-shot generalization without any additional training on these downstream tasks." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.802, + 0.7, + 0.818 + ], + "angle": 0, + "content": "4.3. Efficiency Analysis." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Time Cost Analysis. We first present the time cost (bs=64) for each decoding step of the VAR-d30 model in Figure 5 (c). Notably, the computational complexity across the ten decoding steps of VAR is highly non-uniform. The last three decoding steps alone account for \\(64\\%\\) of the total in" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23340" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.1, + 0.091, + 0.911, + 0.316 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.192, + 0.33, + 0.804, + 0.344 + ], + "angle": 0, + "content": "Figure 7. Qualitative results of CoDe's zero-shot generalization on image inpainting and image editing." + }, + { + "type": "table_caption", + "bbox": [ + 0.098, + 0.369, + 0.476, + 0.383 + ], + "angle": 0, + "content": "Table 3. Memory usage comparison across different batch sizes" + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.395, + 0.492, + 0.564 + ], + "angle": 0, + "content": "
MethodMemory Consumption↓
RunningKV CacheParamsTotal
VAR (bs=8)314MB3595MB8089MB12002MB
+CoDe284MB1023MB9275MB10619MB
VAR (bs=16)615MB7191MB8089MB15901MB
+CoDe557MB2056MB9275MB11951MB
VAR (bs=32)1216MB14345MB8089MB23662MB
+CoDe1103MB4083MB9275MB14614MB
VAR (bs=64)2420MB28707MB8089MB39228MB
+CoDe2195MB8160MB9275MB19943MB
VAR (bs=128)OOM(0.48GB)OOM(57GB)OOM(0.80GB)OOM(70GB)
+CoDe4380MB16320MB9275MB30598MB
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.597, + 0.483, + 0.732 + ], + "angle": 0, + "content": "ference time, with latency increasing quadratically as token map resolution grows. Our solution is to replace the large VAR model with a smaller one for the last few steps, as fewer parameters are needed for large-scale token maps. This small refiner is substantially faster than the original large drafter at these scales, achieving a 4.6x speed improvement in the final and most computationally demanding step. This change dramatically accelerates the entire inference process compared to the original VAR model." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Next, we present the speedup effect of CoDe \\((N = 6)\\) across different batch sizes. As shown in Figure 5.(b), CoDe achieves a speedup of \\(1.6\\mathrm{x}\\) for batch size 1 but reaches up to \\(2.9\\mathrm{x}\\) for batch size 64. The reduced speedup at smaller batch sizes is due to low GPU utilization when generating \\(256\\times 256\\) images, which limits the acceleration potential. However, as GPU utilization increases with larger batch sizes, the speedup of CoDe also improves significantly. This indicates that CoDe is well-suited for computationally intensive image generation tasks, and holds great potential for enhancing efficiency in high-resolution image genera" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.371, + 0.905, + 0.401 + ], + "angle": 0, + "content": "tion under the VAR paradigm. Additionally, we report the time cost of the VQVAE decoder in the figure." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.403, + 0.907, + 0.735 + ], + "angle": 0, + "content": "Memory Consumption Analysis. In Table 3, we provide a detailed analysis of memory usage during the VAR inference process. The KV cache of the VAR model is the largest memory consumer, requiring 12 times the memory needed for the activations due to the significantly extended sequence length in the VAR paradigm. Our proposed CoDe effectively addresses this KV cache memory challenge. Specifically, we use the large VAR model only for predicting the first 91 tokens, which represent just \\(13\\%\\) of the total sequence. After this, the KV cache in the dafter is released, and the remaining computation is handled by a smaller model. This approach drastically reduces the KV cache memory requirements, compressing it to approximately \\(28\\%\\) of the original VAR-d30. Although the refiner model adds a small number of additional parameters, this overhead is minimal compared to the major optimization achieved in KV cache storage. As a result, the total memory usage of CoDe is significantly lower than the original model. Moreover, as batch size increases, the memory savings with CoDe become even more pronounced, underscoring its potential to support efficient high-resolution image generation within the VAR paradigm." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.753, + 0.634, + 0.769 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.906, + 0.901 + ], + "angle": 0, + "content": "This work presents CoDe, a novel method designed for efficient decoding in visual auto-regressive modeling. CoDe effectively mitigates the significant memory overhead and computational redundancy typically associated with the prolonged sequences of next-scale predictions. Through extensive experimentation, our method demonstrates a superior efficiency-quality trade-off, establishing a new benchmark for efficient, high-quality image generation." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "23341" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.091, + 0.241, + 0.108 + ], + "angle": 0, + "content": "Acknowledgment" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.115, + 0.483, + 0.161 + ], + "angle": 0, + "content": "This project is supported by the Ministry of Education, Singapore, under its Academic Research Fund Tier 2 (Award Number: MOE-T2EP20122-0006)." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.173, + 0.187, + 0.189 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.199, + 0.484, + 0.253 + ], + "angle": 0, + "content": "[1] Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, et al. Qwen technical report. arXiv preprint arXiv:2309.16609, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.256, + 0.484, + 0.31 + ], + "angle": 0, + "content": "[2] Fan Bao, Chongxuan Li, Jun Zhu, and Bo Zhang. Analytic-dpm: an analytic estimate of the optimal reverse variance in diffusion probabilistic models. arXiv preprint arXiv:2201.06503, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.311, + 0.484, + 0.379 + ], + "angle": 0, + "content": "[3] Huiwen Chang, Han Zhang, Lu Jiang, Ce Liu, and William T Freeman. Maskgit: Masked generative image transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11315-11325, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.381, + 0.483, + 0.436 + ], + "angle": 0, + "content": "[4] Mark Chen, Alec Radford, Rewon Child, Jeffrey Wu, Heewoo Jun, David Luan, and Ilya Sutskever. Generative pretraining from pixels. In International conference on machine learning, pages 1691-1703. PMLR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.438, + 0.483, + 0.478 + ], + "angle": 0, + "content": "[5] Zigeng Chen, Gongfan Fang, Xinyin Ma, and Xinchao Wang. \\(0.1\\%\\) data makes segment anything slim. arXiv preprint arXiv:2312.05284, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.48, + 0.483, + 0.535 + ], + "angle": 0, + "content": "[6] Zigeng Chen, Xinyin Ma, Gongfan Fang, Zhenxiong Tan, and Xinchao Wang. Asyncdiff: Parallelizing diffusion models by asynchronous denoising. arXiv preprint arXiv:2406.06911, 2024.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.537, + 0.483, + 0.59 + ], + "angle": 0, + "content": "[7] Tri Dao, Dan Fu, Stefano Ermon, Atri Rudra, and Christopher Ré. Flashattention: Fast and memory-efficient exact attention with io-awareness. Advances in Neural Information Processing Systems, 35:16344-16359, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.592, + 0.483, + 0.648 + ], + "angle": 0, + "content": "[8] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.649, + 0.483, + 0.716 + ], + "angle": 0, + "content": "[9] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.719, + 0.483, + 0.773 + ], + "angle": 0, + "content": "[10] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12873-12883, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.775, + 0.483, + 0.83 + ], + "angle": 0, + "content": "[11] Gongfan Fang, Xinyin Ma, Mingli Song, Michael Bi Mi, and Xinchao Wang. Depgraph: Towards any structural pruning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16091-16101, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.831, + 0.483, + 0.872 + ], + "angle": 0, + "content": "[12] Gongfan Fang, Xinyin Ma, and Xinchao Wang. Structural pruning for diffusion models. Advances in neural information processing systems, 36, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[13] Jiarui Fang, Jinzhe Pan, Xibo Sun, Aoyu Li, and Jiannan Wang. xedit: an inference engine for diffusion trans" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.199, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.906, + 0.12 + ], + "angle": 0, + "content": "formers (dits) with massive parallelism. arXiv preprint arXiv:2411.01738, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.121, + 0.906, + 0.161 + ], + "angle": 0, + "content": "[14] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.163, + 0.906, + 0.232 + ], + "angle": 0, + "content": "[15] Jiatao Gu, Yuyang Wang, Yizhe Zhang, Qihang Zhang, Dinghuai Zhang, Navdeep Jaitly, Josh Susskind, and Shuangfei Zhai. Dart: Denoising autoregressive transformer for scalable text-to-image generation. arXiv preprint arXiv:2410.08159, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.233, + 0.906, + 0.3 + ], + "angle": 0, + "content": "[16] Daya Guo, Qihao Zhu, Dejian Yang, Zhenda Xie, Kai Dong, Wentao Zhang, Guanting Chen, Xiao Bi, Yu Wu, YK Li, et al. Deepseek-coder: When the large language model meets programming-the rise of code intelligence. arXiv preprint arXiv:2401.14196, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.302, + 0.906, + 0.37 + ], + "angle": 0, + "content": "[17] Wanggui He, Siming Fu, Mushui Liu, Xierui Wang, Wenyi Xiao, Fangxun Shu, Yi Wang, Lei Zhang, Zhelun Yu, Haoyuan Li, et al. Mars: Mixture of auto-regressive models for fine-grained text-to-image synthesis. arXiv preprint arXiv:2407.07614, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.372, + 0.906, + 0.44 + ], + "angle": 0, + "content": "[18] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.442, + 0.905, + 0.469 + ], + "angle": 0, + "content": "[19] Geoffrey Hinton. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.47, + 0.906, + 0.537 + ], + "angle": 0, + "content": "[20] Doohyuk Jang, Sihwan Park, June Yong Yang, Yeonsung Jung, Jihun Yun, Souvik Kundu, Sung-Yub Kim, and Eunho Yang. *Lantern: Accelerating visual autoregressive models with relaxed speculative decoding.* arXiv preprint arXiv:2410.03355, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.539, + 0.905, + 0.566 + ], + "angle": 0, + "content": "[21] Diederik P Kingma. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.568, + 0.906, + 0.621 + ], + "angle": 0, + "content": "[22] Haopeng Li, Jinyue Yang, Kexin Wang, Xuerui Qiu, Yuhong Chou, Xin Li, and Guoqi Li. Scalable autoregressive image generation with mamba. arXiv preprint arXiv:2408.12245, 2024. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.623, + 0.906, + 0.704 + ], + "angle": 0, + "content": "[23] Muyang Li, Tianle Cai, Jiaxin Cao, Qinsheng Zhang, Han Cai, Junjie Bai, Yangqing Jia, Kai Li, and Song Han. Distributed parallel inference for high-resolution diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7183-7193, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.707, + 0.906, + 0.774 + ], + "angle": 0, + "content": "[24] Muyang Li*, Yujun Lin*, Zhekai Zhang*, Tianle Cai, Xiuyu Li, Junxian Guo, Enze Xie, Chenlin Meng, Jun-Yan Zhu, and Song Han. Svdquant: Absorbing outliers by low-rank components for 4-bit diffusion models. arXiv preprint arXiv:2411.05007, 2024.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.776, + 0.906, + 0.843 + ], + "angle": 0, + "content": "[25] Senmao Li, Taihang Hu, Fahad Shahbaz Khan, Linxuan Li, Shiqi Yang, Yaxing Wang, Ming-Ming Cheng, and Jian Yang. Faster diffusion: Rethinking the role of unet encoder in diffusion models. arXiv preprint arXiv:2312.09608, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.906, + 0.9 + ], + "angle": 0, + "content": "[26] Tianhong Li, Yonglong Tian, He Li, Mingyang Deng, and Kaiming He. Autoregressive image generation without vector quantization. arXiv preprint arXiv:2406.11838, 2024. 1, 3, 6" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23342" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.161 + ], + "angle": 0, + "content": "[27] Xiuyu Li, Yijiang Liu, Long Lian, Huanrui Yang, Zhen Dong, Daniel Kang, Shanghang Zhang, and Kurt Keutzer. Q-diffusion: Quantizing diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17535-17545, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.164, + 0.482, + 0.218 + ], + "angle": 0, + "content": "[28] Xiang Li, Hao Chen, Kai Qiu, Jason Kuen, Jiuming Gu, Bhiksha Raj, and Zhe Lin. Imagefolder: Autoregressive image generation with folded tokens. arXiv preprint arXiv:2410.01756, 2024.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.221, + 0.482, + 0.275 + ], + "angle": 0, + "content": "[29] Xiang Li, Kai Qiu, Hao Chen, Jason Kuen, Zhe Lin, Rita Singh, and Bhiksha Raj. Controlvar: Exploring controllable visual autoregressive modeling. arXiv preprint arXiv:2406.09750, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.278, + 0.482, + 0.346 + ], + "angle": 0, + "content": "[30] Yanyu Li, Huan Wang, Qing Jin, Ju Hu, Pavlo Chemerys, Yun Fu, Yanzhi Wang, Sergey Tulyakov, and Jian Ren. Snapfusion: Text-to-image diffusion model on mobile devices within two seconds. Advances in Neural Information Processing Systems, 36, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.349, + 0.482, + 0.416 + ], + "angle": 0, + "content": "[31] Zongming Li, Tianheng Cheng, Shoufa Chen, Peize Sun, Haocheng Shen, Longjin Ran, Xiaoxin Chen, Wenyu Liu, and Xinggang Wang. Controlar: Controllable image generation with autoregressive models. arXiv preprint arXiv:2410.02705, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.419, + 0.482, + 0.487 + ], + "angle": 0, + "content": "[32] Dongyang Liu, Shitian Zhao, Le Zhuo, Weifeng Lin, Yu Qiao, Hongsheng Li, and Peng Gao. Lumina-mgpt: Illuminate flexible photorealistic text-to-image generation with multimodal generative pretraining. arXiv preprint arXiv:2408.02657, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.49, + 0.482, + 0.531 + ], + "angle": 0, + "content": "[33] Luping Liu, Yi Ren, Zhijie Lin, and Zhou Zhao. Pseudo numerical methods for diffusion models on manifolds. arXiv preprint arXiv:2202.09778, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.533, + 0.482, + 0.6 + ], + "angle": 0, + "content": "[34] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps. Advances in Neural Information Processing Systems, 35:5775-5787, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.603, + 0.482, + 0.657 + ], + "angle": 0, + "content": "[35] Simian Luo, Yiqin Tan, Longbo Huang, Jian Li, and Hang Zhao. Latent consistency models: Synthesizing high-resolution images with few-step inference. arXiv preprint arXiv:2310.04378, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.66, + 0.482, + 0.715 + ], + "angle": 0, + "content": "[36] Zhuoyan Luo, Fengyuan Shi, Yixiao Ge, Yujiu Yang, Limin Wang, and Ying Shan. Open-magvit2: An open-source project toward democratizing auto-regressive visual generation. arXiv preprint arXiv:2409.04410, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.717, + 0.482, + 0.758 + ], + "angle": 0, + "content": "[37] Zhaoyang Lyu, Xudong Xu, Ceyuan Yang, Dahua Lin, and Bo Dai. Accelerating diffusion models via early stop of the diffusion process. arXiv preprint arXiv:2205.12524, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.76, + 0.482, + 0.8 + ], + "angle": 0, + "content": "[38] Xinyin Ma, Gongfan Fang, and Xinchao Wang. Deepcache: Accelerating diffusion models for free. arXiv preprint arXiv:2312.00858, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.803, + 0.482, + 0.858 + ], + "angle": 0, + "content": "[39] Xiaoxiao Ma, Mohan Zhou, Tao Liang, Yalong Bai, Tiejun Zhao, Huaian Chen, and Yi Jin. Star: Scale-wise text-to-image generation via auto-regressive representations. arXiv preprint arXiv:2406.10797, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.86, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[40] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An im" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.093, + 0.905, + 0.133 + ], + "angle": 0, + "content": "perative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[41] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4195-4205, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.193, + 0.905, + 0.247 + ], + "angle": 0, + "content": "[42] Kai Qiu, Xiang Li, Hao Chen, Jie Sun, Jinglu Wang, Zhe Lin, Marios Savvides, and Bhiksha Raj. Efficient autoregressive audio modeling via next-scale prediction. arXiv preprint arXiv:2408.09027, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.25, + 0.905, + 0.29 + ], + "angle": 0, + "content": "[43] Tim Salimans and Jonathan Ho. Progressive distillation for fast sampling of diffusion models. arXiv preprint arXiv:2202.00512, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.293, + 0.905, + 0.332 + ], + "angle": 0, + "content": "[44] Axel Sauer, Dominik Lorenz, Andreas Blattmann, and Robin Rombach. Adversarial diffusion distillation. arXiv preprint arXiv:2311.17042, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.335, + 0.905, + 0.39 + ], + "angle": 0, + "content": "[45] Yuzhang Shang, Zhihang Yuan, Bin Xie, Bingzhe Wu, and Yan Yan. Post-training quantization on diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1972-1981, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.392, + 0.905, + 0.433 + ], + "angle": 0, + "content": "[46] Junhyuk So, Jungwon Lee, and Eunhyeok Park. Frdiff: Feature reuse for exquisite zero-shot acceleration of diffusion models. arXiv preprint arXiv:2312.03517, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.435, + 0.905, + 0.475 + ], + "angle": 0, + "content": "[47] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.478, + 0.905, + 0.533 + ], + "angle": 0, + "content": "[48] Peize Sun, Yi Jiang, Shoufa Chen, Shilong Zhang, Bingyue Peng, Ping Luo, and Zehuan Yuan. Autoregressive model beats diffusion: Llama for scalable image generation. arXiv preprint arXiv:2406.06525, 2024. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.535, + 0.905, + 0.603 + ], + "angle": 0, + "content": "[49] Haotian Tang, Yecheng Wu, Shang Yang, Enze Xie, Junsong Chen, Junyu Chen, Zhuoyang Zhang, Han Cai, Yao Lu, and Song Han. Hart: Efficient visual generation with hybrid autoregressive transformer. arXiv preprint arXiv:2410.10812, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.605, + 0.905, + 0.673 + ], + "angle": 0, + "content": "[50] Yao Teng, Han Shi, Xian Liu, Xuefei Ning, Guohao Dai, Yu Wang, Zhenguo Li, and Xihui Liu. Accelerating autoregressive text-to-image generation with training-free speculative jacobi decoding. arXiv preprint arXiv:2410.01699, 2024.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.676, + 0.905, + 0.731 + ], + "angle": 0, + "content": "[51] Keyu Tian, Yi Jiang, Zehuan Yuan, Bingyue Peng, and Liwei Wang. Visual autoregressive modeling: Scalable image generation via next-scale prediction. arXiv preprint arXiv:2404.02905, 2024. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.733, + 0.905, + 0.801 + ], + "angle": 0, + "content": "[52] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.803, + 0.905, + 0.858 + ], + "angle": 0, + "content": "[53] Aaron Van den Oord, Nal Kalchbrenner, Lasse Espeholt, Oriol Vinyals, Alex Graves, et al. Conditional image generation with pixelCNN decoders. Advances in neural information processing systems, 29, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[54] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23343" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.161 + ], + "angle": 0, + "content": "[55] Felix Wimbauer, Bichen Wu, Edgar Schoenfeld, Xiaoliang Dai, Ji Hou, Zijian He, Artsiom Sanakoyeu, Peizhao Zhang, Sam Tsai, Jonas Kohler, et al. Cache me if you can: Accelerating diffusion models through block caching. arXiv preprint arXiv:2312.03209, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.164, + 0.482, + 0.245 + ], + "angle": 0, + "content": "[56] Jinheng Xie, Weijia Mao, Zechen Bai, David Junhao Zhang, Weihao Wang, Kevin Qinghong Lin, Yuchao Gu, Zhijie Chen, Zhenheng Yang, and Mike Zheng Shou. Show-o: One single transformer to unify multimodal understanding and generation. arXiv preprint arXiv:2408.12528, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.249, + 0.482, + 0.288 + ], + "angle": 0, + "content": "[57] Xingyi Yang and Xinchao Wang. Hash3d: Training-free acceleration for 3d generation. arXiv preprint arXiv:2404.06091, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.291, + 0.483, + 0.346 + ], + "angle": 0, + "content": "[58] Xingyi Yang, Daquan Zhou, Jiashi Feng, and Xinchao Wang. Diffusion probabilistic model made slim. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22552-22562, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.348, + 0.482, + 0.403 + ], + "angle": 0, + "content": "[59] Ziyu Yao, Jialin Li, Yifeng Zhou, Yong Liu, Xi Jiang, Chengjie Wang, Feng Zheng, Yuexian Zou, and Lei Li. Car: Controllable autoregressive modeling for visual generation. arXiv preprint arXiv:2410.04671, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.405, + 0.482, + 0.459 + ], + "angle": 0, + "content": "[60] Tianwei Yin, Michael Gharbi, Richard Zhang, Eli Shechtman, Fredo Durand, William T Freeman, and Taesung Park. One-step diffusion with distribution matching distillation. arXiv preprint arXiv:2311.18828, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.461, + 0.482, + 0.516 + ], + "angle": 0, + "content": "[61] Zongsheng Yue, Jianyi Wang, and Chen Change Loy. Reshift: Efficient diffusion model for image superresolution by residual shifting. Advances in Neural Information Processing Systems, 36, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.518, + 0.482, + 0.572 + ], + "angle": 0, + "content": "[62] Dingkun Zhang, Sijia Li, Chen Chen, Qingsong Xie, and Haonan Lu. Laptop-diff: Layer pruning and normalized distillation for compressing diffusion models. arXiv preprint arXiv:2404.11098, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.575, + 0.482, + 0.628 + ], + "angle": 0, + "content": "[63] Jinzhi Zhang, Feng Xiong, and Mu Xu. G3pt: Unleash the power of autoregressive modeling in 3d generation via cross-scale querying transformer. arXiv preprint arXiv:2409.06322, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.632, + 0.482, + 0.672 + ], + "angle": 0, + "content": "[64] Qinsheng Zhang and Yongxin Chen. Fast sampling of diffusion models with exponential integrator. arXiv preprint arXiv:2204.13902, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.675, + 0.482, + 0.729 + ], + "angle": 0, + "content": "[65] Qian Zhang, Xiangzi Dai, Ninghua Yang, Xiang An, Ziyong Feng, and Xingyu Ren. Var-clip: Text-to-image generator with visual auto-regressive modeling. arXiv preprint arXiv:2408.01181, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.732, + 0.482, + 0.786 + ], + "angle": 0, + "content": "[66] Wentian Zhang, Haozhe Liu, Jinheng Xie, Francesco Faccio, Mike Zheng Shou, and Jürgen Schmidhuber. Cross-attention makes inference cumbersome in text-to-image diffusion models. arXiv preprint arXiv:2404.02747, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.788, + 0.482, + 0.829 + ], + "angle": 0, + "content": "[67] Xuanlei Zhao, Xiaolong Jin, Kai Wang, and Yang You. Real-time video generation with pyramid attention broadcast. arXiv preprint arXiv:2408.12588, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.831, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[68] Yang Zhao, Yanwu Xu, Zhisheng Xiao, and Tingbo Hou. Mobilediffusion: Subsecond text-to-image generation on mobile devices. arXiv preprint arXiv:2311.16567, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.874, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[69] Kaiwen Zheng, Cheng Lu, Jianfei Chen, and Jun Zhu. Dpm solver-v3: Improved diffusion ode solver with empirical" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.548, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "model statistics. Advances in Neural Information Processing Systems, 36, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.191 + ], + "angle": 0, + "content": "[70] Chunting Zhou, Lili Yu, Arun Babu, Kushal Tirumala, Michihiro Yasunaga, Leonid Shamis, Jacob Kahn, Xuezhe Ma, Luke Zettlemoyer, and Omer Levy. Transfusion: Predict the next token and diffuse images with one multi-modal model. arXiv preprint arXiv:2408.11039, 2024. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.191 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23344" + } + ] +] \ No newline at end of file diff --git a/2025/Collaborative Decoding Makes Visual Auto-Regressive Modeling Efficient/9a935c24-e29c-43fe-a54b-aa3325d4f954_origin.pdf b/2025/Collaborative Decoding Makes Visual Auto-Regressive Modeling Efficient/9a935c24-e29c-43fe-a54b-aa3325d4f954_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..3f0fe5e04a14a444ef3bf3a010a328df582d3497 --- /dev/null +++ b/2025/Collaborative Decoding Makes Visual Auto-Regressive Modeling Efficient/9a935c24-e29c-43fe-a54b-aa3325d4f954_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1bd1e926edc4e7ba5d96c706dcc4bc8ce9c77dcda9d052f471a1aa5efbd5d46a +size 6802994 diff --git a/2025/Collaborative Decoding Makes Visual Auto-Regressive Modeling Efficient/full.md b/2025/Collaborative Decoding Makes Visual Auto-Regressive Modeling Efficient/full.md new file mode 100644 index 0000000000000000000000000000000000000000..314d27f15ce3460700c914435b00797ddae550e9 --- /dev/null +++ b/2025/Collaborative Decoding Makes Visual Auto-Regressive Modeling Efficient/full.md @@ -0,0 +1,297 @@ +# Collaborative Decoding Makes Visual Auto-Regressive Modeling Efficient + +Zigeng Chen, Xinyin Ma, Gongfan Fang, Xinchao Wang* National University of Singapore + +{zigeng99, maxinyin, gongfan}@u.nus.edu, xinchao@nus.edu.sg + +https://github.com/czg1225/CoDe + +# Abstract + +In the rapidly advancing field of image generation, Visual Auto-Regressive (VAR) modeling has garnered considerable attention for its innovative next-scale prediction approach. This paradigm offers substantial improvements in efficiency, scalability, and zero-shot generalization. Yet, the inherently coarse-to-fine nature of VAR introduces a prolonged token sequence, leading to prohibitive memory consumption and computational redundancies. To address these bottlenecks, we propose Collaborative Decoding (CoDe), a novel efficient decoding strategy tailored for the VAR framework. CoDe capitalizes on two critical observations: the substantially reduced parameter demands at larger scales and the exclusive generation patterns across different scales. Based on these insights, we partition the multi-scale inference process into a seamless collaboration between a large model and a small model. The large model serves as the 'drafter', specializing in generating low-frequency content at smaller scales, while the smaller model serves as the 'refiner', solely focusing on predicting high-frequency details at larger scales. This collaboration yields remarkable efficiency with minimal impact on quality: CoDe achieves a $1.7x$ speedup, slashes memory usage by around $50\%$ , and preserves image quality with only a negligible FID increase from 1.95 to 1.98. When drafting steps are further decreased, CoDe can achieve an impressive $2.9x$ acceleration ratio, reaching 41 images/s at 256x256 resolution on a single NVIDIA 4090 GPU, while preserving a commendable FID of 2.27. + +# 1. Introduction + +The past year has witnessed significant advancements in Auto-Regressive (AR) models for image generation [17, 22, 26, 31, 48, 59], driven by their proven scalability and strong generalization capabilities [1, 9, 16, 52]. Leveraging these strengths, AR approaches have demonstrated re + +![](images/9c935d98139a7d1b133241f4ce75e3c756050044b7aed6ca6d0f7ef3548bd591.jpg) + +![](images/40927f390e85beaf727d352c7ff6eb69fa1b9634effca101de8ac3c6a469dd37.jpg) +Figure 1. We partition the next-scale prediction process into the efficient collaboration between large and small VAR models. + +markable image quality and promising multi-modal potential [32, 56, 70]. However, the next-token prediction process inherent to conventional AR models requires numerous decoding steps, leading to considerable generation latency. Visual Auto-Regressive (VAR) Modeling [51] replaces the GPT-style next-token prediction with a next-scale prediction strategy. It generates content in a multi-level, coarse-to-fine progression, enabling the model to decode multiple tokens in parallel [29, 42, 49, 63, 65], thus considerably reducing the inference steps. + +Despite less decoding steps, VAR's progressive scaling approach significantly increases the overall sequence length. To generate a 16x16 token image, VAR hierarchically decodes up to 680 tokens across 10 scales—2.7 times the sequence length required by conventional AR models. As the VAR model must store the KV cache accumulated from all previous tokens, this prolonged sequence results in a substantial memory overhead, especially in the final scales. During inference, the KV cache consumes approx- + +![](images/94ce9a46c2690dcbe99efb7d4bc688229150a02504b85c5955cbe9f55ef54bac.jpg) +Figure 2. Comparison of generation results between original VAR-d30 (up) and our VAR-CoDE (bottom) for ImageNet $256 \times 256$ . Our method achieves 1.7x speedup (3.62s to 2.11s), and needs only 0.5x memory space (40GB to 20GB), with negligible quality degradation. + +imately 12 times more memory than the forward computation itself. For instance, generating images using VAR-d30 with a batch size of 128 demands 70 GB of memory, with 57 GB dedicated solely to KV caching, becoming the primary bottleneck. Moreover, the extended sequence length exacerbates the computational cost of self-attention, given the quadratic growth in attention map calculations. + +To address these inefficiencies, we begin by analyzing the specific properties of VAR's next-scale prediction paradigm. First, we observe that as generated scales increase, the parameter demands for high-quality token generation drop substantially, leading to considerable computational redundancy at most of the tokens in the long sequence. Next, we find that the generation patterns at small and large scales are exclusive, resulting in mutual interference between learning low- and high-frequency modeling capability, thus hindering efficient parameter utilization. + +Our Approach. Inspired by the above observations, we propose Collaborative Decoding (CoDe), a simple yet highly effective method that significantly boosts the inference efficiency of VAR models while preserving generation quality comparable to the original. Our core idea is to decompose the long-sequence scaling-up process into a collaboration between two VAR models with different sizes and specialized roles to enhance efficiency. Figure 1 presents the overview of our method. We use a large VAR model as a drafter for the initial small scales, where model capacity demands are high but computations and KV cache are sparse. Conversely, we use a small VAR model as a refiner at the remaining large scales, where computations and KV cache are intensive but fewer parameters are needed. Next, we apply specialized fine-tuning to both models to address the optimization interference between scales encountered during the pre-training stage. Each model is fine-tuned exclusively on the specific scales it handles, leading to a notable performance boost with limited additional cost. + +Extensive experiments validate the effectiveness of our method. As shown in Figure 2, compared to the original VAR-d30 model, CoDe achieves a $1.7\mathrm{x}$ speedup with merely $0.5\mathrm{x}$ GPU memory consumption and negligible quality loss (FID [18] slightly increases from 1.95 to 1.98). Notably, our approach can reach up to a $2.9\mathrm{x}$ speedup, generating 41 images (256x256) per second on a single NVIDIA 4090 GPU, while still maintaining an FID of 2.27. + +In conclusion, we introduce CoDe, a novel decoding framework for visual auto-regressive modeling that significantly enhances speed and reduces memory overhead with a negligible impact on quality. CoDe divides the hierarchical sequence modeling of VAR into a collaborative process, utilizing a large VAR model and a small VAR model in a progressive partnership. Additionally, we propose specialized fine-tuning to optimize each model for its specific role, effectively mitigating training interference and maximizing parameter utilization. To the best of our knowledge, CoDe is the fastest method available to achieve an FID below 2, making it a clear advancement in efficient image generation. + +# 2. Related Works + +Auto-regressive image generation. Early works [4, 53] pioneered image generation by generating pixels in raster-scan order. Later, VQVAE [54] and VQGAN [10] improved this approach by quantizing image patches into discrete tokens, using a transformer in a decoder-only setup to generate these tokens in a raster-scan manner. Building on these foundations, recent efforts have focused on enhancing autoregressive (AR) models for image generation. LlamaGEN [48] and Lumina-mGPT [32], for example, use a GPT-style next-token prediction strategy to achieve high-quality image generation with good scalability. AiM [22] and MARS [17] further improve this paradigm by introducing mixture-of-experts and linear attention mechanisms [14]. Methods like SHOW-O [56], Transfusion [70], and DART [15] + +![](images/5bcc327a07b0252b3123f285d8b87697801a6ba07fb16dd8579f9674e2dc2efa.jpg) +(a) Quality vs. Params at different scales +Figure 3. (a) Effectiveness of increasing parameters at the $k$ -th scale is evaluated by predicting token map $r_k$ using four VAR models with different parameter sizes (2B, 1B, 0.6B, and 0.3B), while other scales $(r_1, r_2, \ldots, r_{k-1}, r_{k+1}, \ldots, r_{10})$ are generated using the largest VAR-d30 model. (b) Fourier spectrum analysis is conducted on generated content at the first 3 scales and the last 3 scales. (c) Training-free performance comparison of model collaboration decoding across various settings of draft tokens $M$ and refiner tokens $680 - M$ . + +![](images/832ed1fa91c44ad4b066fbf183c7d73540ac572e01d7cfd12114957736abb29d.jpg) +(b) Spectrum analysis + +![](images/8ad5650c2f3680e6367052a46b7be781d5af6cbab821c8646a78087076862ebf.jpg) +(c) Training-free model collaboration + +combine diffusion processes with autoregressive modeling, while [3, 26, 36] introduces masked autoregression to generate images. However, all these approaches suffer from high latency due to the large number of forward steps. VAR [51] addresses this by using hierarchical parallel decoding, which greatly reduces the number of steps, leading to significantly lower inference latency without sacrificing quality. Building on VAR's progress, many recent works have aimed to improve next-scale prediction across multiple tasks, including text-to-image [28, 39, 49, 65], controllable generation [29, 31], audio generation [42], and 3D generation [63]. + +Efficient Image Generation Models. For diffusion models, acceleration techniques are already well-developed. [35, 43, 44, 60, 61] focuses on reducing sampling steps while [12, 30, 58, 62, 68] optimize model architectures through pruning [5, 11] or knowledge distillation [19]. To avoid the high costs of training, some training-free methods are proposed. Some approaches develop fast solvers for stochastic differential equations (SDE) or ordinary differential equations (ODE) to enhance sampling efficiency [2, 33, 34, 64, 69]. Other works [25, 37, 38, 46, 55, 57, 66, 67] exploit specific characteristics of diffusion models to skip redundant computations during the denoising process. [6, 13, 23] reduce latency to another level through distributed computing. Quantization methods [24, 27, 45] have also shown great potential. In contrast, research into accelerating AR image generation is still in its early stages. LANTERN [20] and SJD [50] employ speculative decoding to speed up next-token prediction, achieving notable speedups but not suitable for the innovative next-scale prediction paradigm introduced by VAR. + +In this paper, we introduce CoDe, a novel and efficient decoding method tailored for the next-scale prediction paradigm. CoDe effectively addresses the inefficiencies associated with the long sequence structure of VAR models while maintaining high generation quality. Notably, CoDe stands out as the fastest method capable of achieving an FID + +below 2, making it a clear advancement. + +# 3. Method + +# 3.1. Preliminary + +Visual auto-regressive modeling [51] redefines the traditional AR by shifting from a "next-token" prediction to a "next-scale" prediction. In this framework, each auto-regressive unit is a token map at varying scales, rather than a single token. For a given image feature map $f \in \mathbb{R}^{h \times w \times C}$ , VAR quantizes it into $K$ multi-scale token maps $R = (r_1, r_2, \ldots, r_K)$ at progressively finer resolutions, with the final token map $r_K$ matching the original feature map's resolution. The probability distribution is reformulated as: + +$$ +p \left(r _ {1}, r _ {2}, \dots , r _ {K}\right) = \prod_ {k = 1} ^ {K} p \left(r _ {k} \mid r _ {1}, r _ {2}, \dots , r _ {k - 1}\right), \tag {1} +$$ + +where each token map $r_k \in [V]^{h_k \times w_k}$ consists of $h_k \times w_k$ tokens at scale $k$ , and the sequence $(r_1, r_2, \ldots, r_{k-1})$ serves as the "prefix" for $r_k$ . In this paradigm, during each autoregressive step $k$ , the model predicts all $h_k \times w_k$ tokens in $r_k$ in parallel, conditioned on prior scales and position embeddings. This approach aligns with a coarse-to-fine generation pattern, enabling parallel decoding within each scale. VAR effectively improves inference speed and generation quality, but it also considerably increases sequence length. + +# 3.2. Key Observations + +VAR represents an innovative paradigm specific for visual generation, distinct from traditional autoregressive approaches and introducing many yet unexplored characteristics. In this work, we revisit the entire next-scale prediction process to uncover specific properties that can be optimized to reduce computational redundancy and improve inference efficiency. + +Observation 1: As the predicted scale becomes larger, the need for parameters reduces significantly. While VAR + +![](images/091c168ef7e121154f6d60de6c1db09622d48246121795b963374b64f34389f9.jpg) +Figure 4. Overview of the collaborative decoding process, we use a drafting step $N = 6$ for instance. CoDe uses a large VAR model as the drafter $\epsilon_{\theta_d}$ to generate the token maps $R_L = (r_1, r_2, \dots, r_N)$ at smaller scales. The small refiner model $\epsilon_{\theta_r}$ then uses $R_L$ as an initial prefix to efficiently predict the remaining token maps $R_H = (r_{N+1}, r_{N+2}, \dots, r_K)$ at larger scales. Both models are fine-tuned on their designated predictive scales using ground truth labels $r_k^*$ and teacher logits $p_{\text{teacher}}(r_k)$ , respectively. + +![](images/7343d83148610ce5b6f70d0c9222955db4eb4eede9d0c4b62e016674eb7b9a67.jpg) + +models exhibit strong scalability, the effectiveness of increasing parameter counts varies greatly in predicting various scales. In Figure 3. (a), we present the differences in class-conditional generation (ImageNet-256 [8]) quality when using VAR models of different sizes at each scale. At the initial small scales, increasing model parameters leads to noticeable improvements in generation quality. However, as the predicted scale gradually progresses, the impact of additional parameters becomes minimal. At the final scale—which accounts for $38\%$ tokens of the entire sequence—we observe that the performance of the 2B model is even nearly equivalent to that of the 0.3B model. These findings suggest that as the predicted scale increases, the parameter demand for accurate token predictions declines markedly, revealing substantial computational redundancy in the current VAR inference process at large scales. + +Observation 2: The generative patterns are exclusive between small and large scales. For VAR models, the content generated at small and large scales varies significantly. As demonstrated by the Fourier analysis in Figure 3 (b), the feature maps produced by the first three scales contain primarily low-frequency components, while the last three scales focus on high-frequency components. To further validate these distinct generative patterns, we conducted a perturbation fine-tuning experiment. Using a pre-trained VARd16 model, we applied CSE loss only to tokens in the largest three scales and fine-tuned for just $1\%$ of the original training epochs. This minor fine-tuning at large scales led to a complete collapse of the model's global modeling capacity at small scales, with the FID increasing from 3.30 to 21.93 and the IS score dropping from 277 to 88. These results demonstrate that VAR models perform entirely dis + +tinct generative tasks at small and large scales, with minimal overlap. Training a single VAR model to predict across all scales results in significant mutual interference between learning low-frequency and high-frequency modeling capabilities, making parameter optimization challenging. + +In sum, the key insights from our observations can be concluded as follows: (1) At large scales, VAR models require significantly fewer parameters for accurate token prediction, leading to considerable computational redundancy. (2) Training a single model to generate across all scales causes mutual interference between small and large scales, hindering parameter optimization due to conflicting low-frequency and high-frequency learning tendencies. + +# 3.3. Collaborative Decoding + +Based on the above insights, we propose a simple yet pretty efficient decoding method for next-scale prediction called collaborative decoding. As presented in Figure 4, CoDe decomposes the next-scale prediction process into a collaboration between a large drafter model and a small refiner model. The drafter is the original large VAR model with 2B parameters, while the refiner is a lightweight 0.3B VAR model, sharing the same transformer-based architecture but with reduced width and depth. The drafter is responsible for generating coarse, low-frequency global structures at smaller scales to draft the image, while the refiner continually predicts the high-frequency details at larger scales to refine the image. + +Model Collaboration. To generate an image $im$ , original VAR model needs to auto-regressively generate $K$ multiscale token maps $R = (r_1, r_2, \ldots, r_K)$ with progressively finer resolutions, where each token map $r_k \in [V]^{h_k \times w_k}$ represents the token map at scale $k$ . In CoDe, we decom + +pose the auto-regressive steps into $N$ drafting steps and $K - N$ refining steps. In the initial drafting phase where computation is sparse but additional parameters provide significant benefits, we serve a large VAR model as the drafter $\epsilon_{\theta_d}$ , generating the initial set of low-frequency token maps $R_L = (r_1,r_2,\dots ,r_N)$ up to scale $N$ , where $N < K$ . These token maps predict the global structure of the image, serving as a prefix for the remaining finer-scale token maps. The drafting process can be represented as: + +$$ +p _ {\theta_ {d}} \left(r _ {1}, r _ {2}, \dots , r _ {N}\right) = \prod_ {k = 1} ^ {N} p _ {\theta_ {d}} \left(r _ {k} \mid r _ {1}, r _ {2}, \dots , r _ {k - 1}\right). \tag {2} +$$ + +After drafting, the KV cache of the large model is released, significantly optimizing memory usage. + +In the refining stage where the computation is intensive but the parameter demand is low, we serve a small VAR model as the refiner $\epsilon_{\theta_r}$ , employing the drafter's predictions $R_L$ as a prefix and focus on refining the details by generating the remaining high-frequency token maps $R_H = (r_{N + 1}, r_{N + 2}, \ldots, r_K)$ up to scale $K$ , where $r_K$ matches the resolution of the original feature map. Notably, predicting $r_{N + 1}$ requires an attention mask, as the refiner model lacks the KV cache for the first $N$ scales. The probability distribution of refining steps is formulated as: + +$$ +\begin{array}{l} p \left(r _ {N + 1}, r _ {N + 2}, \dots , r _ {K} \mid R _ {L}\right) \\ = \prod_ {k = N + 1} ^ {K} p \left(r _ {k} \mid R _ {L}, r _ {N + 1}, r _ {N + 2}, \dots , r _ {k - 1}\right). \tag {3} \\ \end{array} +$$ + +Finally, the generated images $im$ are reconstructed from both drafting maps $R_{L}$ and refining maps $R_{H}$ as follows: + +$$ +i m = D \left(Q \left(R _ {L}, R _ {H}\right), \right. \tag {4} +$$ + +where $Q(.)$ is a residual-style quantization function, and $D(.)$ is a multi-scale VQVAE decoder. Figure 3. (c) demonstrates the remarkable effectiveness of model collaboration under training-free conditions. With a large model as the drafter, the small refiner only results in a slight and gentle increase in FID. Conversely, with a small model as the drafter, even assigning the last $80\%$ tokens to a large refiner fails to improve performance. This also verifies Observation 1: small scales require more model capacity, whereas the parameter demands for large scales are low. + +The model collaboration method effectively reduces the computational redundancy of VAR and maintains generation quality comparable to the original model, while significantly offering faster speed and reduced memory usage. + +Specialized Fine-Tuning. Given the exclusive generation patterns between drafting scales and refining scales, we propose the specialized fine-tuning to further specialize drafter and refiner models by fine-tuning them solely on their respective predictive scales, thereby avoiding training interference and enhancing generation quality. + +For the drafter model $\epsilon_{\theta_d}$ , we fine-tune the drafting steps to improve the generation of the initial low-frequency token maps $R_L$ . The drafter model is trained using a CSE loss $\mathrm{CSE}(\cdot, \cdot)$ between its generated token distribution $p_{\theta_d}(r_k)$ and the ground truth labels $r_k^*$ , defined as: + +$$ +\mathcal {L} _ {\text {d r a f t e r}} = \sum_ {k = 1} ^ {N} \mathrm {C S E} \left(p _ {\theta_ {d}} \left(r _ {k}\right), r _ {k} ^ {*}\right), \tag {5} +$$ + +which encourages the drafter to closely match the target global structure at each drafting step. + +For the refiner model $\epsilon_{\theta_r}$ , we adopt a Knowledge Distillation (KD) [19] approach to partially transfer the knowledge from a larger pre-trained VAR model. The KD loss function for epoch $ep$ can be expressed as: + +$$ +\mathcal {L} _ {\text {r e f i n e r}} ^ {(e p)} = \sum_ {k = 1} ^ {K} \left(\lambda_ {e p} \cdot \mathbf {1} _ {[ k \leq N ]} + \mathbf {1} _ {[ k > N ]}\right) \mathrm {K L} \left(p _ {\theta_ {r}} \left(r _ {k}\right) \| p _ {\text {t e a c h e r}} \left(r _ {k}\right)\right), \tag {6} +$$ + +where $p_{\mathrm{teacher}}(\cdot)$ is the distribution predicted by the larger teacher VAR model, $\mathrm{KL}(\cdot, \cdot)$ is the standard KL-divergence loss aligns refiner's output distribution with that of the teacher model, and $\lambda_{ep}$ is a dynamic weighting factor. $\lambda_t$ linearly decreases from one to zero from the initial to the end of finetuning, gradually shifting the KD emphasis from all tokens to the specific refining tokens. This dynamic weighting allows the refiner to specialize in refining the details with a smoother and more stable training process. + +The specialized fine-tuning of the drafter and refiner models ensures that each model becomes highly proficient in its respective task. By minimizing interference between small and large scales, this approach allows for more sufficient parameter optimization, resulting in enhanced image generation quality with limited training costs. + +# 4. Experimental Results + +# 4.1. Experimental Setup. + +Implementation Details. We evaluated our method's effectiveness on the ImageNet1K [8] class-conditional generation benchmarks. The proposed CoDe framework involves collaboration between a large and a small VAR model. Specifically, we use the pre-trained VAR-d30 model as the drafter and the pre-trained VAR-d16 model as the refiner. Each model undergoes specialized fine-tuning focused exclusively on its respective predictive scales. For the drafter, we fine-tune it for $5\%$ of its original training epochs with a base learning rate of 1e-6 and a weight decay of 0.08. The refiner model, on the other hand, is fine-tuned for $25\%$ of its original training epochs with a base learning rate of 1e-5 and no weight decay. Both models are optimized using the AdamW [21] optimizer with a batch size of 1024, achieved through gradient accumulation. The fine-tuning was conducted on 4 NVIDIA L20 GPUs. + +![](images/dc55050fcdac980dc9e1d9d489ef7462f9451c1081fd7f37592f82b10ecd2d99.jpg) +(a) Efficiency & Quality trade-off + +![](images/4a5834facc896802db97a7dc744ebe5cbdb464c67ba2bb8e5a1055abf5330c48.jpg) +Figure 5. (a) Our CoDe demonstrates the optimal efficiency-quality trade-off among all evaluated methods. (b) Inference latency is measured across varying batch sizes for the original VAR-d30, our CoDe $(N = 6)$ , and the VQVAE decoder. (c) We analyze the time cost associated with parallel decoding at each scale, showing that the refiner model is significantly more efficient than the drafter at larger scales. + +![](images/c7c8080d11d76784bfaf5c2250f05668741b1712e2d7b5cd25d961cd1d874316.jpg) +(b) Time cost with different batchsizes +(c) Time cost at each scale + +Table 1. Quantitative assessment of the efficiency-quality trade-off across various methods. Inference efficiency is evaluated with a batch size of 64 on NVIDIA L20 GPU, with latency measured excluding VQVAE or VQGAN as it incurs a shared time cost across all methods. + +
MethodInference EfficiencyGeneration Quality
#StepsSpeedup↑Latency↓Throughput↑#ParamMemory↓FID ↓IS ↑Precision↑Recall↑
DiT-XL/2500.2x19.20s3.33it/s675M11369MB2.262390.800.60
MAR-B1000.1x29.80s2.15it/s208M8725MB2.312820.820.57
AiM-XL2560.4x9.32s6.87it/s763M20983MB2.562570.820.57
LlamaGen-XXL384<0.1x73.97s0.87it/s1.4B42632MB2.342540.800.59
VAR-d30101.0x3.62s17.71it/s2.0B39228MB1.953010.810.59
VAR-d24101.7x2.07s30.92it/s1.0B25093MB2.113110.820.59
VAR-d20102.8x1.29s49.62it/s600M17814MB2.613010.830.56
VAR-CoDe N=99+11.2x2.97s21.54it/s2.0+0.3B28803MB1.942960.800.61
VAR-CoDe N=88+21.7x2.11s30.33it/s2.0+0.3B21019MB1.983020.810.60
VAR-CoDe N=77+32.3x1.60s40.00it/s2.0+0.3B19943MB2.113030.820.59
VAR-CoDe N=66+42.9x1.27s50.39it/s2.0+0.3B19943MB2.272970.820.58
+ +Table 2. Effect of specialized fine-tuning + +
DiscriptionN=6N=7N=8N=9
CoDe Training-free2.422.262.101.99
CoDe Fine-tuning2.272.111.981.94
+ +**Evaluations.** We evaluated our method on ImageNet1K-256 generation, focusing on both quality and efficiency metrics. For quality assessment, we used standard metrics such as FID, Inception Score (IS), Precision, and Recall. For other baselines, we use their default sampling methods. For CoDe, we reduced the sampled top-k from 900 to 600 due to improved token prediction accuracy after fine-tuning, while keeping the default top-p at 0.96. To compensate for reduced diversity, we introduced a temperature coefficient of $t = 1.1$ for sampling on the smallest 7 scales. Efficiency was measured through inference latency, throughput, memory consumption, and parameter count to provide a comprehensive comparison. Notably, the speed measurements + +exclude the VQGAN decoder, which contributes minimally to overall runtime and is a shared component across all methods. All efficiency tests were conducted on a single NVIDIA L20 GPU without additional optimizations such as FlashAttention [7], using PyTorch 2.1 [40] and FP16. + +# 4.2. Main Results. + +Quality-Efficiency Trade-off. We evaluated the quality-efficiency trade-off of our proposed method (CoDe) against original VAR models [51], state-of-the-art AR models, and diffusion models. The AR models included GPT-style LlamaGEN [48], Mamba-based Aim [22], and MAE-style MAR [26]. For diffusion models, we employed the widely used DiT models [41]. To measure the trade-offs, we used series models with different parameter counts for VAR, LlamaGEN, Aim, and MAR, while for DiT we varied the DDIM [47] steps to generate the efficiency-quality curve. + +As illustrated in Table 1, We analyzed CoDe's performance under different drafting steps N. Compared to traditional AR models, our proposed VAR-CoDe $(N = 6)$ achieves + +![](images/aa9b79272d373c128c4d3be3d5a6b11633b164ec4a9e1aa8863fe6d5dbc50c31.jpg) +Figure 6. Qualitative comparison between the original VAR-d30 model and our proposed CoDe model, with different drafting steps. + +60 times faster inference than LlamaGEN-XXL, while also surpassing its generation quality. Additionally, in comparison to diffusion models, VAR-CoDe $(N = 6)$ is 15 times faster than DiT-XL/2, while maintaining the same level of quality. Compared to the original VAR-d30, CoDe $(N = 8)$ achieves a $1.7\mathrm{x}$ speedup and reduces memory consumption by $50\%$ , with only a negligible FID increase from 1.95 to 1.98. When the drafting stage is reduced to just 6 steps, CoDe $(N = 6)$ achieves a $2.9\mathrm{x}$ speedup, reaching a throughput of over 50it/s, while maintaining a low FID of 2.27. This is a speed unmatched by any other existing methods. Compared to other VAR models with fewer parameters (VAR-d24 and VAR-d20), our method achieves significantly better quality while maintaining the same speedup ratio. + +As shown in Figure 5 (a), VAR-CoDe achieves the best efficiency-quality trade-off compared to all other methods, effectively solving the inefficiencies introduced by the extended sequence length in the original VAR paradigm. + +Training-Free vs Specialized Finetuning. Our CoDe framework can operate in a training-free manner by directly using pre-trained VAR-d30 and VAR-d16 models as the drafter and refiner, respectively. Alternatively, we can perform specialized fine-tuning to further enhance the models' performance at their respective scales, with limited training cost. Table 2 compares the results of training-free CoDe with the specialized fine-tuned version. Even without additional training, CoDe demonstrates competitive performance, outperforming VAR-d24 and VAR-d20 models with the same speedup ratio. With specialized fine-tuning, CoDe's performance improves significantly, even achieving a slight quality enhancement over the original model at a $1.2\mathrm{x}$ acceleration ratio. + +Qualitative Results. We provide an extensive qualitative + +comparison between the original VAR-d30 model and our proposed CoDe, with varying drafting steps $N = \{6,7,8\}$ . As illustrated in Figure 6, our approach achieves significant speedup and substantial memory optimization, with only minimal quality degradation that is nearly imperceptible to the human eye. Even at a speedup rate of 2.9 times, the generated images maintain high quality and accurate semantic information. Through specialized fine-tuning, CoDe's draft model exhibits higher predictive accuracy compared to the original model. This sometimes results in a different global structure from the original output, yet the image quality remains consistently high or even better. + +Zero-shot Task Generalization. To evaluate the zero-shot generalization capability of CoDe, we conducted additional experiments on zero-shot class-conditional inpainting and image editing. During image inpainting, we applied teacher forcing by providing the ground truth tokens outside the masked area, allowing the model to generate tokens solely within the mask. Notably, class-conditional information was introduced to the model. In the image editing task, CoDe was restricted to generating tokens within a given bounding box based on a specific class label. Figure 7 illustrates the qualitative results of CoDe's zero-shot performance $(N = 8)$ . Our approach demonstrates strong zero-shot generalization without any additional training on these downstream tasks. + +# 4.3. Efficiency Analysis. + +Time Cost Analysis. We first present the time cost (bs=64) for each decoding step of the VAR-d30 model in Figure 5 (c). Notably, the computational complexity across the ten decoding steps of VAR is highly non-uniform. The last three decoding steps alone account for $64\%$ of the total in + +![](images/1717531237c044669f2b369a0fceacaf34053e4113ea17db356a0cde4a2088fc.jpg) +Figure 7. Qualitative results of CoDe's zero-shot generalization on image inpainting and image editing. + +Table 3. Memory usage comparison across different batch sizes + +
MethodMemory Consumption↓
RunningKV CacheParamsTotal
VAR (bs=8)314MB3595MB8089MB12002MB
+CoDe284MB1023MB9275MB10619MB
VAR (bs=16)615MB7191MB8089MB15901MB
+CoDe557MB2056MB9275MB11951MB
VAR (bs=32)1216MB14345MB8089MB23662MB
+CoDe1103MB4083MB9275MB14614MB
VAR (bs=64)2420MB28707MB8089MB39228MB
+CoDe2195MB8160MB9275MB19943MB
VAR (bs=128)OOM(0.48GB)OOM(57GB)OOM(0.80GB)OOM(70GB)
+CoDe4380MB16320MB9275MB30598MB
+ +ference time, with latency increasing quadratically as token map resolution grows. Our solution is to replace the large VAR model with a smaller one for the last few steps, as fewer parameters are needed for large-scale token maps. This small refiner is substantially faster than the original large drafter at these scales, achieving a 4.6x speed improvement in the final and most computationally demanding step. This change dramatically accelerates the entire inference process compared to the original VAR model. + +Next, we present the speedup effect of CoDe $(N = 6)$ across different batch sizes. As shown in Figure 5.(b), CoDe achieves a speedup of $1.6\mathrm{x}$ for batch size 1 but reaches up to $2.9\mathrm{x}$ for batch size 64. The reduced speedup at smaller batch sizes is due to low GPU utilization when generating $256\times 256$ images, which limits the acceleration potential. However, as GPU utilization increases with larger batch sizes, the speedup of CoDe also improves significantly. This indicates that CoDe is well-suited for computationally intensive image generation tasks, and holds great potential for enhancing efficiency in high-resolution image genera + +tion under the VAR paradigm. Additionally, we report the time cost of the VQVAE decoder in the figure. + +Memory Consumption Analysis. In Table 3, we provide a detailed analysis of memory usage during the VAR inference process. The KV cache of the VAR model is the largest memory consumer, requiring 12 times the memory needed for the activations due to the significantly extended sequence length in the VAR paradigm. Our proposed CoDe effectively addresses this KV cache memory challenge. Specifically, we use the large VAR model only for predicting the first 91 tokens, which represent just $13\%$ of the total sequence. After this, the KV cache in the dafter is released, and the remaining computation is handled by a smaller model. This approach drastically reduces the KV cache memory requirements, compressing it to approximately $28\%$ of the original VAR-d30. Although the refiner model adds a small number of additional parameters, this overhead is minimal compared to the major optimization achieved in KV cache storage. As a result, the total memory usage of CoDe is significantly lower than the original model. Moreover, as batch size increases, the memory savings with CoDe become even more pronounced, underscoring its potential to support efficient high-resolution image generation within the VAR paradigm. + +# 5. Conclusion + +This work presents CoDe, a novel method designed for efficient decoding in visual auto-regressive modeling. CoDe effectively mitigates the significant memory overhead and computational redundancy typically associated with the prolonged sequences of next-scale predictions. Through extensive experimentation, our method demonstrates a superior efficiency-quality trade-off, establishing a new benchmark for efficient, high-quality image generation. + +# Acknowledgment + +This project is supported by the Ministry of Education, Singapore, under its Academic Research Fund Tier 2 (Award Number: MOE-T2EP20122-0006). + +# References + +[1] Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, et al. Qwen technical report. arXiv preprint arXiv:2309.16609, 2023. 1 +[2] Fan Bao, Chongxuan Li, Jun Zhu, and Bo Zhang. Analytic-dpm: an analytic estimate of the optimal reverse variance in diffusion probabilistic models. arXiv preprint arXiv:2201.06503, 2022. 3 +[3] Huiwen Chang, Han Zhang, Lu Jiang, Ce Liu, and William T Freeman. Maskgit: Masked generative image transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11315-11325, 2022. 3 +[4] Mark Chen, Alec Radford, Rewon Child, Jeffrey Wu, Heewoo Jun, David Luan, and Ilya Sutskever. Generative pretraining from pixels. In International conference on machine learning, pages 1691-1703. PMLR, 2020. 2 +[5] Zigeng Chen, Gongfan Fang, Xinyin Ma, and Xinchao Wang. $0.1\%$ data makes segment anything slim. arXiv preprint arXiv:2312.05284, 2023. 3 +[6] Zigeng Chen, Xinyin Ma, Gongfan Fang, Zhenxiong Tan, and Xinchao Wang. Asyncdiff: Parallelizing diffusion models by asynchronous denoising. arXiv preprint arXiv:2406.06911, 2024.3 +[7] Tri Dao, Dan Fu, Stefano Ermon, Atri Rudra, and Christopher Ré. Flashattention: Fast and memory-efficient exact attention with io-awareness. Advances in Neural Information Processing Systems, 35:16344-16359, 2022. 6 +[8] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 4, 5 +[9] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 1 +[10] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12873-12883, 2021. 2 +[11] Gongfan Fang, Xinyin Ma, Mingli Song, Michael Bi Mi, and Xinchao Wang. Depgraph: Towards any structural pruning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16091-16101, 2023. 3 +[12] Gongfan Fang, Xinyin Ma, and Xinchao Wang. Structural pruning for diffusion models. Advances in neural information processing systems, 36, 2024. 3 +[13] Jiarui Fang, Jinzhe Pan, Xibo Sun, Aoyu Li, and Jiannan Wang. xedit: an inference engine for diffusion trans + +formers (dits) with massive parallelism. arXiv preprint arXiv:2411.01738, 2024. 3 +[14] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023. 2 +[15] Jiatao Gu, Yuyang Wang, Yizhe Zhang, Qihang Zhang, Dinghuai Zhang, Navdeep Jaitly, Josh Susskind, and Shuangfei Zhai. Dart: Denoising autoregressive transformer for scalable text-to-image generation. arXiv preprint arXiv:2410.08159, 2024. 2 +[16] Daya Guo, Qihao Zhu, Dejian Yang, Zhenda Xie, Kai Dong, Wentao Zhang, Guanting Chen, Xiao Bi, Yu Wu, YK Li, et al. Deepseek-coder: When the large language model meets programming-the rise of code intelligence. arXiv preprint arXiv:2401.14196, 2024. 1 +[17] Wanggui He, Siming Fu, Mushui Liu, Xierui Wang, Wenyi Xiao, Fangxun Shu, Yi Wang, Lei Zhang, Zhelun Yu, Haoyuan Li, et al. Mars: Mixture of auto-regressive models for fine-grained text-to-image synthesis. arXiv preprint arXiv:2407.07614, 2024. 1, 2 +[18] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 2 +[19] Geoffrey Hinton. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 3, 5 +[20] Doohyuk Jang, Sihwan Park, June Yong Yang, Yeonsung Jung, Jihun Yun, Souvik Kundu, Sung-Yub Kim, and Eunho Yang. *Lantern: Accelerating visual autoregressive models with relaxed speculative decoding.* arXiv preprint arXiv:2410.03355, 2024. 3 +[21] Diederik P Kingma. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 5 +[22] Haopeng Li, Jinyue Yang, Kexin Wang, Xuerui Qiu, Yuhong Chou, Xin Li, and Guoqi Li. Scalable autoregressive image generation with mamba. arXiv preprint arXiv:2408.12245, 2024. 1, 2, 6 +[23] Muyang Li, Tianle Cai, Jiaxin Cao, Qinsheng Zhang, Han Cai, Junjie Bai, Yangqing Jia, Kai Li, and Song Han. Distributed parallel inference for high-resolution diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7183-7193, 2024. 3 +[24] Muyang Li*, Yujun Lin*, Zhekai Zhang*, Tianle Cai, Xiuyu Li, Junxian Guo, Enze Xie, Chenlin Meng, Jun-Yan Zhu, and Song Han. Svdquant: Absorbing outliers by low-rank components for 4-bit diffusion models. arXiv preprint arXiv:2411.05007, 2024.3 +[25] Senmao Li, Taihang Hu, Fahad Shahbaz Khan, Linxuan Li, Shiqi Yang, Yaxing Wang, Ming-Ming Cheng, and Jian Yang. Faster diffusion: Rethinking the role of unet encoder in diffusion models. arXiv preprint arXiv:2312.09608, 2023. 3 +[26] Tianhong Li, Yonglong Tian, He Li, Mingyang Deng, and Kaiming He. Autoregressive image generation without vector quantization. arXiv preprint arXiv:2406.11838, 2024. 1, 3, 6 + +[27] Xiuyu Li, Yijiang Liu, Long Lian, Huanrui Yang, Zhen Dong, Daniel Kang, Shanghang Zhang, and Kurt Keutzer. Q-diffusion: Quantizing diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17535-17545, 2023. 3 +[28] Xiang Li, Hao Chen, Kai Qiu, Jason Kuen, Jiuming Gu, Bhiksha Raj, and Zhe Lin. Imagefolder: Autoregressive image generation with folded tokens. arXiv preprint arXiv:2410.01756, 2024.3 +[29] Xiang Li, Kai Qiu, Hao Chen, Jason Kuen, Zhe Lin, Rita Singh, and Bhiksha Raj. Controlvar: Exploring controllable visual autoregressive modeling. arXiv preprint arXiv:2406.09750, 2024. 1, 3 +[30] Yanyu Li, Huan Wang, Qing Jin, Ju Hu, Pavlo Chemerys, Yun Fu, Yanzhi Wang, Sergey Tulyakov, and Jian Ren. Snapfusion: Text-to-image diffusion model on mobile devices within two seconds. Advances in Neural Information Processing Systems, 36, 2024. 3 +[31] Zongming Li, Tianheng Cheng, Shoufa Chen, Peize Sun, Haocheng Shen, Longjin Ran, Xiaoxin Chen, Wenyu Liu, and Xinggang Wang. Controlar: Controllable image generation with autoregressive models. arXiv preprint arXiv:2410.02705, 2024. 1, 3 +[32] Dongyang Liu, Shitian Zhao, Le Zhuo, Weifeng Lin, Yu Qiao, Hongsheng Li, and Peng Gao. Lumina-mgpt: Illuminate flexible photorealistic text-to-image generation with multimodal generative pretraining. arXiv preprint arXiv:2408.02657, 2024. 1, 2 +[33] Luping Liu, Yi Ren, Zhijie Lin, and Zhou Zhao. Pseudo numerical methods for diffusion models on manifolds. arXiv preprint arXiv:2202.09778, 2022. 3 +[34] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps. Advances in Neural Information Processing Systems, 35:5775-5787, 2022. 3 +[35] Simian Luo, Yiqin Tan, Longbo Huang, Jian Li, and Hang Zhao. Latent consistency models: Synthesizing high-resolution images with few-step inference. arXiv preprint arXiv:2310.04378, 2023. 3 +[36] Zhuoyan Luo, Fengyuan Shi, Yixiao Ge, Yujiu Yang, Limin Wang, and Ying Shan. Open-magvit2: An open-source project toward democratizing auto-regressive visual generation. arXiv preprint arXiv:2409.04410, 2024. 3 +[37] Zhaoyang Lyu, Xudong Xu, Ceyuan Yang, Dahua Lin, and Bo Dai. Accelerating diffusion models via early stop of the diffusion process. arXiv preprint arXiv:2205.12524, 2022. 3 +[38] Xinyin Ma, Gongfan Fang, and Xinchao Wang. Deepcache: Accelerating diffusion models for free. arXiv preprint arXiv:2312.00858, 2023. 3 +[39] Xiaoxiao Ma, Mohan Zhou, Tao Liang, Yalong Bai, Tiejun Zhao, Huaian Chen, and Yi Jin. Star: Scale-wise text-to-image generation via auto-regressive representations. arXiv preprint arXiv:2406.10797, 2024. 3 +[40] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An im + +perative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019. 6 +[41] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4195-4205, 2023. 6 +[42] Kai Qiu, Xiang Li, Hao Chen, Jie Sun, Jinglu Wang, Zhe Lin, Marios Savvides, and Bhiksha Raj. Efficient autoregressive audio modeling via next-scale prediction. arXiv preprint arXiv:2408.09027, 2024. 1, 3 +[43] Tim Salimans and Jonathan Ho. Progressive distillation for fast sampling of diffusion models. arXiv preprint arXiv:2202.00512, 2022. 3 +[44] Axel Sauer, Dominik Lorenz, Andreas Blattmann, and Robin Rombach. Adversarial diffusion distillation. arXiv preprint arXiv:2311.17042, 2023. 3 +[45] Yuzhang Shang, Zhihang Yuan, Bin Xie, Bingzhe Wu, and Yan Yan. Post-training quantization on diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1972-1981, 2023. 3 +[46] Junhyuk So, Jungwon Lee, and Eunhyeok Park. Frdiff: Feature reuse for exquisite zero-shot acceleration of diffusion models. arXiv preprint arXiv:2312.03517, 2023. 3 +[47] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 6 +[48] Peize Sun, Yi Jiang, Shoufa Chen, Shilong Zhang, Bingyue Peng, Ping Luo, and Zehuan Yuan. Autoregressive model beats diffusion: Llama for scalable image generation. arXiv preprint arXiv:2406.06525, 2024. 1, 2, 6 +[49] Haotian Tang, Yecheng Wu, Shang Yang, Enze Xie, Junsong Chen, Junyu Chen, Zhuoyang Zhang, Han Cai, Yao Lu, and Song Han. Hart: Efficient visual generation with hybrid autoregressive transformer. arXiv preprint arXiv:2410.10812, 2024. 1, 3 +[50] Yao Teng, Han Shi, Xian Liu, Xuefei Ning, Guohao Dai, Yu Wang, Zhenguo Li, and Xihui Liu. Accelerating autoregressive text-to-image generation with training-free speculative jacobi decoding. arXiv preprint arXiv:2410.01699, 2024.3 +[51] Keyu Tian, Yi Jiang, Zehuan Yuan, Bingyue Peng, and Liwei Wang. Visual autoregressive modeling: Scalable image generation via next-scale prediction. arXiv preprint arXiv:2404.02905, 2024. 1, 3, 6 +[52] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 1 +[53] Aaron Van den Oord, Nal Kalchbrenner, Lasse Espeholt, Oriol Vinyals, Alex Graves, et al. Conditional image generation with pixelCNN decoders. Advances in neural information processing systems, 29, 2016. 2 +[54] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017. 2 + +[55] Felix Wimbauer, Bichen Wu, Edgar Schoenfeld, Xiaoliang Dai, Ji Hou, Zijian He, Artsiom Sanakoyeu, Peizhao Zhang, Sam Tsai, Jonas Kohler, et al. Cache me if you can: Accelerating diffusion models through block caching. arXiv preprint arXiv:2312.03209, 2023. 3 +[56] Jinheng Xie, Weijia Mao, Zechen Bai, David Junhao Zhang, Weihao Wang, Kevin Qinghong Lin, Yuchao Gu, Zhijie Chen, Zhenheng Yang, and Mike Zheng Shou. Show-o: One single transformer to unify multimodal understanding and generation. arXiv preprint arXiv:2408.12528, 2024. 1, 2 +[57] Xingyi Yang and Xinchao Wang. Hash3d: Training-free acceleration for 3d generation. arXiv preprint arXiv:2404.06091, 2024. 3 +[58] Xingyi Yang, Daquan Zhou, Jiashi Feng, and Xinchao Wang. Diffusion probabilistic model made slim. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22552-22562, 2023. 3 +[59] Ziyu Yao, Jialin Li, Yifeng Zhou, Yong Liu, Xi Jiang, Chengjie Wang, Feng Zheng, Yuexian Zou, and Lei Li. Car: Controllable autoregressive modeling for visual generation. arXiv preprint arXiv:2410.04671, 2024. 1 +[60] Tianwei Yin, Michael Gharbi, Richard Zhang, Eli Shechtman, Fredo Durand, William T Freeman, and Taesung Park. One-step diffusion with distribution matching distillation. arXiv preprint arXiv:2311.18828, 2023. 3 +[61] Zongsheng Yue, Jianyi Wang, and Chen Change Loy. Reshift: Efficient diffusion model for image superresolution by residual shifting. Advances in Neural Information Processing Systems, 36, 2024. 3 +[62] Dingkun Zhang, Sijia Li, Chen Chen, Qingsong Xie, and Haonan Lu. Laptop-diff: Layer pruning and normalized distillation for compressing diffusion models. arXiv preprint arXiv:2404.11098, 2024. 3 +[63] Jinzhi Zhang, Feng Xiong, and Mu Xu. G3pt: Unleash the power of autoregressive modeling in 3d generation via cross-scale querying transformer. arXiv preprint arXiv:2409.06322, 2024. 1, 3 +[64] Qinsheng Zhang and Yongxin Chen. Fast sampling of diffusion models with exponential integrator. arXiv preprint arXiv:2204.13902, 2022. 3 +[65] Qian Zhang, Xiangzi Dai, Ninghua Yang, Xiang An, Ziyong Feng, and Xingyu Ren. Var-clip: Text-to-image generator with visual auto-regressive modeling. arXiv preprint arXiv:2408.01181, 2024. 1, 3 +[66] Wentian Zhang, Haozhe Liu, Jinheng Xie, Francesco Faccio, Mike Zheng Shou, and Jürgen Schmidhuber. Cross-attention makes inference cumbersome in text-to-image diffusion models. arXiv preprint arXiv:2404.02747, 2024. 3 +[67] Xuanlei Zhao, Xiaolong Jin, Kai Wang, and Yang You. Real-time video generation with pyramid attention broadcast. arXiv preprint arXiv:2408.12588, 2024. 3 +[68] Yang Zhao, Yanwu Xu, Zhisheng Xiao, and Tingbo Hou. Mobilediffusion: Subsecond text-to-image generation on mobile devices. arXiv preprint arXiv:2311.16567, 2023. 3 +[69] Kaiwen Zheng, Cheng Lu, Jianfei Chen, and Jun Zhu. Dpm solver-v3: Improved diffusion ode solver with empirical + +model statistics. Advances in Neural Information Processing Systems, 36, 2024. 3 +[70] Chunting Zhou, Lili Yu, Arun Babu, Kushal Tirumala, Michihiro Yasunaga, Leonid Shamis, Jacob Kahn, Xuezhe Ma, Luke Zettlemoyer, and Omer Levy. Transfusion: Predict the next token and diffuse images with one multi-modal model. arXiv preprint arXiv:2408.11039, 2024. 1, 2 \ No newline at end of file diff --git a/2025/Collaborative Decoding Makes Visual Auto-Regressive Modeling Efficient/images.zip b/2025/Collaborative Decoding Makes Visual Auto-Regressive Modeling Efficient/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..7c52d521d3039307e22c53e32f69f050d0391d63 --- /dev/null +++ b/2025/Collaborative Decoding Makes Visual Auto-Regressive Modeling Efficient/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c777700121b055f1f56a7a428476231604ac70792c59d02592ffb0d3215670d0 +size 848615 diff --git a/2025/Collaborative Decoding Makes Visual Auto-Regressive Modeling Efficient/layout.json b/2025/Collaborative Decoding Makes Visual Auto-Regressive Modeling Efficient/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..51ec4a607dda564e68f0884ce2b5d495e999f6f4 --- /dev/null +++ b/2025/Collaborative Decoding Makes Visual Auto-Regressive Modeling Efficient/layout.json @@ -0,0 +1,8733 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 78, + 103, + 533, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 103, + 533, + 121 + ], + "spans": [ + { + "bbox": [ + 78, + 103, + 533, + 121 + ], + "type": "text", + "content": "Collaborative Decoding Makes Visual Auto-Regressive Modeling Efficient" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 159, + 143, + 455, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 143, + 455, + 172 + ], + "spans": [ + { + "bbox": [ + 159, + 143, + 455, + 172 + ], + "type": "text", + "content": "Zigeng Chen, Xinyin Ma, Gongfan Fang, Xinchao Wang* National University of Singapore" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 146, + 174, + 465, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 174, + 465, + 186 + ], + "spans": [ + { + "bbox": [ + 146, + 174, + 465, + 186 + ], + "type": "text", + "content": "{zigeng99, maxinyin, gongfan}@u.nus.edu, xinchao@nus.edu.sg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 190, + 186, + 416, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 190, + 186, + 416, + 199 + ], + "spans": [ + { + "bbox": [ + 190, + 186, + 416, + 199 + ], + "type": "text", + "content": "https://github.com/czg1225/CoDe" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "spans": [ + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 253, + 296, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 253, + 296, + 589 + ], + "spans": [ + { + "bbox": [ + 55, + 253, + 296, + 589 + ], + "type": "text", + "content": "In the rapidly advancing field of image generation, Visual Auto-Regressive (VAR) modeling has garnered considerable attention for its innovative next-scale prediction approach. This paradigm offers substantial improvements in efficiency, scalability, and zero-shot generalization. Yet, the inherently coarse-to-fine nature of VAR introduces a prolonged token sequence, leading to prohibitive memory consumption and computational redundancies. To address these bottlenecks, we propose Collaborative Decoding (CoDe), a novel efficient decoding strategy tailored for the VAR framework. CoDe capitalizes on two critical observations: the substantially reduced parameter demands at larger scales and the exclusive generation patterns across different scales. Based on these insights, we partition the multi-scale inference process into a seamless collaboration between a large model and a small model. The large model serves as the 'drafter', specializing in generating low-frequency content at smaller scales, while the smaller model serves as the 'refiner', solely focusing on predicting high-frequency details at larger scales. This collaboration yields remarkable efficiency with minimal impact on quality: CoDe achieves a " + }, + { + "bbox": [ + 55, + 253, + 296, + 589 + ], + "type": "inline_equation", + "content": "1.7x" + }, + { + "bbox": [ + 55, + 253, + 296, + 589 + ], + "type": "text", + "content": " speedup, slashes memory usage by around " + }, + { + "bbox": [ + 55, + 253, + 296, + 589 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 55, + 253, + 296, + 589 + ], + "type": "text", + "content": ", and preserves image quality with only a negligible FID increase from 1.95 to 1.98. When drafting steps are further decreased, CoDe can achieve an impressive " + }, + { + "bbox": [ + 55, + 253, + 296, + 589 + ], + "type": "inline_equation", + "content": "2.9x" + }, + { + "bbox": [ + 55, + 253, + 296, + 589 + ], + "type": "text", + "content": " acceleration ratio, reaching 41 images/s at 256x256 resolution on a single NVIDIA 4090 GPU, while preserving a commendable FID of 2.27." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 613, + 135, + 625 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 613, + 135, + 625 + ], + "spans": [ + { + "bbox": [ + 55, + 613, + 135, + 625 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 633, + 296, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 633, + 296, + 694 + ], + "spans": [ + { + "bbox": [ + 55, + 633, + 296, + 694 + ], + "type": "text", + "content": "The past year has witnessed significant advancements in Auto-Regressive (AR) models for image generation [17, 22, 26, 31, 48, 59], driven by their proven scalability and strong generalization capabilities [1, 9, 16, 52]. Leveraging these strengths, AR approaches have demonstrated re" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 320, + 227, + 550, + 328 + ], + "blocks": [ + { + "bbox": [ + 320, + 227, + 550, + 328 + ], + "lines": [ + { + "bbox": [ + 320, + 227, + 550, + 328 + ], + "spans": [ + { + "bbox": [ + 320, + 227, + 550, + 328 + ], + "type": "image", + "image_path": "9c935d98139a7d1b133241f4ce75e3c756050044b7aed6ca6d0f7ef3548bd591.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 320, + 327, + 548, + 430 + ], + "blocks": [ + { + "bbox": [ + 320, + 327, + 548, + 430 + ], + "lines": [ + { + "bbox": [ + 320, + 327, + 548, + 430 + ], + "spans": [ + { + "bbox": [ + 320, + 327, + 548, + 430 + ], + "type": "image", + "image_path": "40927f390e85beaf727d352c7ff6eb69fa1b9634effca101de8ac3c6a469dd37.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 440, + 555, + 462 + ], + "lines": [ + { + "bbox": [ + 313, + 440, + 555, + 462 + ], + "spans": [ + { + "bbox": [ + 313, + 440, + 555, + 462 + ], + "type": "text", + "content": "Figure 1. We partition the next-scale prediction process into the efficient collaboration between large and small VAR models." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 312, + 485, + 555, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 485, + 555, + 605 + ], + "spans": [ + { + "bbox": [ + 312, + 485, + 555, + 605 + ], + "type": "text", + "content": "markable image quality and promising multi-modal potential [32, 56, 70]. However, the next-token prediction process inherent to conventional AR models requires numerous decoding steps, leading to considerable generation latency. Visual Auto-Regressive (VAR) Modeling [51] replaces the GPT-style next-token prediction with a next-scale prediction strategy. It generates content in a multi-level, coarse-to-fine progression, enabling the model to decode multiple tokens in parallel [29, 42, 49, 63, 65], thus considerably reducing the inference steps." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "content": "Despite less decoding steps, VAR's progressive scaling approach significantly increases the overall sequence length. To generate a 16x16 token image, VAR hierarchically decodes up to 680 tokens across 10 scales—2.7 times the sequence length required by conventional AR models. As the VAR model must store the KV cache accumulated from all previous tokens, this prolonged sequence results in a substantial memory overhead, especially in the final scales. During inference, the KV cache consumes approx-" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 703, + 143, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 143, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 143, + 713 + ], + "type": "text", + "content": "*Corresponding author" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23334" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 77, + 72, + 553, + 222 + ], + "blocks": [ + { + "bbox": [ + 77, + 72, + 553, + 222 + ], + "lines": [ + { + "bbox": [ + 77, + 72, + 553, + 222 + ], + "spans": [ + { + "bbox": [ + 77, + 72, + 553, + 222 + ], + "type": "image", + "image_path": "94ce9a46c2690dcbe99efb7d4bc688229150a02504b85c5955cbe9f55ef54bac.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 228, + 555, + 253 + ], + "lines": [ + { + "bbox": [ + 55, + 228, + 555, + 253 + ], + "spans": [ + { + "bbox": [ + 55, + 228, + 555, + 253 + ], + "type": "text", + "content": "Figure 2. Comparison of generation results between original VAR-d30 (up) and our VAR-CoDE (bottom) for ImageNet " + }, + { + "bbox": [ + 55, + 228, + 555, + 253 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 55, + 228, + 555, + 253 + ], + "type": "text", + "content": ". Our method achieves 1.7x speedup (3.62s to 2.11s), and needs only 0.5x memory space (40GB to 20GB), with negligible quality degradation." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 272, + 294, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 272, + 294, + 357 + ], + "spans": [ + { + "bbox": [ + 54, + 272, + 294, + 357 + ], + "type": "text", + "content": "imately 12 times more memory than the forward computation itself. For instance, generating images using VAR-d30 with a batch size of 128 demands 70 GB of memory, with 57 GB dedicated solely to KV caching, becoming the primary bottleneck. Moreover, the extended sequence length exacerbates the computational cost of self-attention, given the quadratic growth in attention map calculations." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 361, + 295, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 361, + 295, + 481 + ], + "spans": [ + { + "bbox": [ + 54, + 361, + 295, + 481 + ], + "type": "text", + "content": "To address these inefficiencies, we begin by analyzing the specific properties of VAR's next-scale prediction paradigm. First, we observe that as generated scales increase, the parameter demands for high-quality token generation drop substantially, leading to considerable computational redundancy at most of the tokens in the long sequence. Next, we find that the generation patterns at small and large scales are exclusive, resulting in mutual interference between learning low- and high-frequency modeling capability, thus hindering efficient parameter utilization." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 486, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 486, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 486, + 295, + 713 + ], + "type": "text", + "content": "Our Approach. Inspired by the above observations, we propose Collaborative Decoding (CoDe), a simple yet highly effective method that significantly boosts the inference efficiency of VAR models while preserving generation quality comparable to the original. Our core idea is to decompose the long-sequence scaling-up process into a collaboration between two VAR models with different sizes and specialized roles to enhance efficiency. Figure 1 presents the overview of our method. We use a large VAR model as a drafter for the initial small scales, where model capacity demands are high but computations and KV cache are sparse. Conversely, we use a small VAR model as a refiner at the remaining large scales, where computations and KV cache are intensive but fewer parameters are needed. Next, we apply specialized fine-tuning to both models to address the optimization interference between scales encountered during the pre-training stage. Each model is fine-tuned exclusively on the specific scales it handles, leading to a notable performance boost with limited additional cost." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 272, + 553, + 368 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 272, + 553, + 368 + ], + "spans": [ + { + "bbox": [ + 313, + 272, + 553, + 368 + ], + "type": "text", + "content": "Extensive experiments validate the effectiveness of our method. As shown in Figure 2, compared to the original VAR-d30 model, CoDe achieves a " + }, + { + "bbox": [ + 313, + 272, + 553, + 368 + ], + "type": "inline_equation", + "content": "1.7\\mathrm{x}" + }, + { + "bbox": [ + 313, + 272, + 553, + 368 + ], + "type": "text", + "content": " speedup with merely " + }, + { + "bbox": [ + 313, + 272, + 553, + 368 + ], + "type": "inline_equation", + "content": "0.5\\mathrm{x}" + }, + { + "bbox": [ + 313, + 272, + 553, + 368 + ], + "type": "text", + "content": " GPU memory consumption and negligible quality loss (FID [18] slightly increases from 1.95 to 1.98). Notably, our approach can reach up to a " + }, + { + "bbox": [ + 313, + 272, + 553, + 368 + ], + "type": "inline_equation", + "content": "2.9\\mathrm{x}" + }, + { + "bbox": [ + 313, + 272, + 553, + 368 + ], + "type": "text", + "content": " speedup, generating 41 images (256x256) per second on a single NVIDIA 4090 GPU, while still maintaining an FID of 2.27." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 369, + 555, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 369, + 555, + 513 + ], + "spans": [ + { + "bbox": [ + 313, + 369, + 555, + 513 + ], + "type": "text", + "content": "In conclusion, we introduce CoDe, a novel decoding framework for visual auto-regressive modeling that significantly enhances speed and reduces memory overhead with a negligible impact on quality. CoDe divides the hierarchical sequence modeling of VAR into a collaborative process, utilizing a large VAR model and a small VAR model in a progressive partnership. Additionally, we propose specialized fine-tuning to optimize each model for its specific role, effectively mitigating training interference and maximizing parameter utilization. To the best of our knowledge, CoDe is the fastest method available to achieve an FID below 2, making it a clear advancement in efficient image generation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 525, + 406, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 525, + 406, + 538 + ], + "spans": [ + { + "bbox": [ + 313, + 525, + 406, + 538 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "type": "text", + "content": "Auto-regressive image generation. Early works [4, 53] pioneered image generation by generating pixels in raster-scan order. Later, VQVAE [54] and VQGAN [10] improved this approach by quantizing image patches into discrete tokens, using a transformer in a decoder-only setup to generate these tokens in a raster-scan manner. Building on these foundations, recent efforts have focused on enhancing autoregressive (AR) models for image generation. LlamaGEN [48] and Lumina-mGPT [32], for example, use a GPT-style next-token prediction strategy to achieve high-quality image generation with good scalability. AiM [22] and MARS [17] further improve this paradigm by introducing mixture-of-experts and linear attention mechanisms [14]. Methods like SHOW-O [56], Transfusion [70], and DART [15]" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23335" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 71, + 215, + 175 + ], + "blocks": [ + { + "bbox": [ + 58, + 71, + 215, + 175 + ], + "lines": [ + { + "bbox": [ + 58, + 71, + 215, + 175 + ], + "spans": [ + { + "bbox": [ + 58, + 71, + 215, + 175 + ], + "type": "image", + "image_path": "5bcc327a07b0252b3123f285d8b87697801a6ba07fb16dd8579f9674e2dc2efa.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 72, + 175, + 217, + 186 + ], + "lines": [ + { + "bbox": [ + 72, + 175, + 217, + 186 + ], + "spans": [ + { + "bbox": [ + 72, + 175, + 217, + 186 + ], + "type": "text", + "content": "(a) Quality vs. Params at different scales" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 194, + 555, + 240 + ], + "lines": [ + { + "bbox": [ + 55, + 194, + 555, + 240 + ], + "spans": [ + { + "bbox": [ + 55, + 194, + 555, + 240 + ], + "type": "text", + "content": "Figure 3. (a) Effectiveness of increasing parameters at the " + }, + { + "bbox": [ + 55, + 194, + 555, + 240 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 194, + 555, + 240 + ], + "type": "text", + "content": "-th scale is evaluated by predicting token map " + }, + { + "bbox": [ + 55, + 194, + 555, + 240 + ], + "type": "inline_equation", + "content": "r_k" + }, + { + "bbox": [ + 55, + 194, + 555, + 240 + ], + "type": "text", + "content": " using four VAR models with different parameter sizes (2B, 1B, 0.6B, and 0.3B), while other scales " + }, + { + "bbox": [ + 55, + 194, + 555, + 240 + ], + "type": "inline_equation", + "content": "(r_1, r_2, \\ldots, r_{k-1}, r_{k+1}, \\ldots, r_{10})" + }, + { + "bbox": [ + 55, + 194, + 555, + 240 + ], + "type": "text", + "content": " are generated using the largest VAR-d30 model. (b) Fourier spectrum analysis is conducted on generated content at the first 3 scales and the last 3 scales. (c) Training-free performance comparison of model collaboration decoding across various settings of draft tokens " + }, + { + "bbox": [ + 55, + 194, + 555, + 240 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 194, + 555, + 240 + ], + "type": "text", + "content": " and refiner tokens " + }, + { + "bbox": [ + 55, + 194, + 555, + 240 + ], + "type": "inline_equation", + "content": "680 - M" + }, + { + "bbox": [ + 55, + 194, + 555, + 240 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 221, + 70, + 376, + 175 + ], + "blocks": [ + { + "bbox": [ + 221, + 70, + 376, + 175 + ], + "lines": [ + { + "bbox": [ + 221, + 70, + 376, + 175 + ], + "spans": [ + { + "bbox": [ + 221, + 70, + 376, + 175 + ], + "type": "image", + "image_path": "832ed1fa91c44ad4b066fbf183c7d73540ac572e01d7cfd12114957736abb29d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 268, + 176, + 346, + 186 + ], + "lines": [ + { + "bbox": [ + 268, + 176, + 346, + 186 + ], + "spans": [ + { + "bbox": [ + 268, + 176, + 346, + 186 + ], + "type": "text", + "content": "(b) Spectrum analysis" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 378, + 70, + 533, + 174 + ], + "blocks": [ + { + "bbox": [ + 378, + 70, + 533, + 174 + ], + "lines": [ + { + "bbox": [ + 378, + 70, + 533, + 174 + ], + "spans": [ + { + "bbox": [ + 378, + 70, + 533, + 174 + ], + "type": "image", + "image_path": "8ad5650c2f3680e6367052a46b7be781d5af6cbab821c8646a78087076862ebf.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 400, + 175, + 533, + 186 + ], + "lines": [ + { + "bbox": [ + 400, + 175, + 533, + 186 + ], + "spans": [ + { + "bbox": [ + 400, + 175, + 533, + 186 + ], + "type": "text", + "content": "(c) Training-free model collaboration" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 260, + 296, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 260, + 296, + 392 + ], + "spans": [ + { + "bbox": [ + 54, + 260, + 296, + 392 + ], + "type": "text", + "content": "combine diffusion processes with autoregressive modeling, while [3, 26, 36] introduces masked autoregression to generate images. However, all these approaches suffer from high latency due to the large number of forward steps. VAR [51] addresses this by using hierarchical parallel decoding, which greatly reduces the number of steps, leading to significantly lower inference latency without sacrificing quality. Building on VAR's progress, many recent works have aimed to improve next-scale prediction across multiple tasks, including text-to-image [28, 39, 49, 65], controllable generation [29, 31], audio generation [42], and 3D generation [63]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 396, + 296, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 396, + 296, + 636 + ], + "spans": [ + { + "bbox": [ + 56, + 396, + 296, + 636 + ], + "type": "text", + "content": "Efficient Image Generation Models. For diffusion models, acceleration techniques are already well-developed. [35, 43, 44, 60, 61] focuses on reducing sampling steps while [12, 30, 58, 62, 68] optimize model architectures through pruning [5, 11] or knowledge distillation [19]. To avoid the high costs of training, some training-free methods are proposed. Some approaches develop fast solvers for stochastic differential equations (SDE) or ordinary differential equations (ODE) to enhance sampling efficiency [2, 33, 34, 64, 69]. Other works [25, 37, 38, 46, 55, 57, 66, 67] exploit specific characteristics of diffusion models to skip redundant computations during the denoising process. [6, 13, 23] reduce latency to another level through distributed computing. Quantization methods [24, 27, 45] have also shown great potential. In contrast, research into accelerating AR image generation is still in its early stages. LANTERN [20] and SJD [50] employ speculative decoding to speed up next-token prediction, achieving notable speedups but not suitable for the innovative next-scale prediction paradigm introduced by VAR." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": "In this paper, we introduce CoDe, a novel and efficient decoding method tailored for the next-scale prediction paradigm. CoDe effectively addresses the inefficiencies associated with the long sequence structure of VAR models while maintaining high generation quality. Notably, CoDe stands out as the fastest method capable of achieving an FID" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 260, + 478, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 260, + 478, + 272 + ], + "spans": [ + { + "bbox": [ + 313, + 260, + 478, + 272 + ], + "type": "text", + "content": "below 2, making it a clear advancement." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 281, + 370, + 294 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 281, + 370, + 294 + ], + "spans": [ + { + "bbox": [ + 313, + 281, + 370, + 294 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 301, + 394, + 315 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 301, + 394, + 315 + ], + "spans": [ + { + "bbox": [ + 313, + 301, + 394, + 315 + ], + "type": "text", + "content": "3.1. Preliminary" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 319, + 555, + 427 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 319, + 555, + 427 + ], + "spans": [ + { + "bbox": [ + 313, + 319, + 555, + 427 + ], + "type": "text", + "content": "Visual auto-regressive modeling [51] redefines the traditional AR by shifting from a \"next-token\" prediction to a \"next-scale\" prediction. In this framework, each auto-regressive unit is a token map at varying scales, rather than a single token. For a given image feature map " + }, + { + "bbox": [ + 313, + 319, + 555, + 427 + ], + "type": "inline_equation", + "content": "f \\in \\mathbb{R}^{h \\times w \\times C}" + }, + { + "bbox": [ + 313, + 319, + 555, + 427 + ], + "type": "text", + "content": ", VAR quantizes it into " + }, + { + "bbox": [ + 313, + 319, + 555, + 427 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 319, + 555, + 427 + ], + "type": "text", + "content": " multi-scale token maps " + }, + { + "bbox": [ + 313, + 319, + 555, + 427 + ], + "type": "inline_equation", + "content": "R = (r_1, r_2, \\ldots, r_K)" + }, + { + "bbox": [ + 313, + 319, + 555, + 427 + ], + "type": "text", + "content": " at progressively finer resolutions, with the final token map " + }, + { + "bbox": [ + 313, + 319, + 555, + 427 + ], + "type": "inline_equation", + "content": "r_K" + }, + { + "bbox": [ + 313, + 319, + 555, + 427 + ], + "type": "text", + "content": " matching the original feature map's resolution. The probability distribution is reformulated as:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 327, + 434, + 555, + 467 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 434, + 555, + 467 + ], + "spans": [ + { + "bbox": [ + 327, + 434, + 555, + 467 + ], + "type": "interline_equation", + "content": "p \\left(r _ {1}, r _ {2}, \\dots , r _ {K}\\right) = \\prod_ {k = 1} ^ {K} p \\left(r _ {k} \\mid r _ {1}, r _ {2}, \\dots , r _ {k - 1}\\right), \\tag {1}", + "image_path": "c87207a31fe9c7c5152df3008b4e1c6ff45569208ce5f9a89d35912f5d51c68b.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 474, + 556, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 474, + 556, + 583 + ], + "spans": [ + { + "bbox": [ + 313, + 474, + 556, + 583 + ], + "type": "text", + "content": "where each token map " + }, + { + "bbox": [ + 313, + 474, + 556, + 583 + ], + "type": "inline_equation", + "content": "r_k \\in [V]^{h_k \\times w_k}" + }, + { + "bbox": [ + 313, + 474, + 556, + 583 + ], + "type": "text", + "content": " consists of " + }, + { + "bbox": [ + 313, + 474, + 556, + 583 + ], + "type": "inline_equation", + "content": "h_k \\times w_k" + }, + { + "bbox": [ + 313, + 474, + 556, + 583 + ], + "type": "text", + "content": " tokens at scale " + }, + { + "bbox": [ + 313, + 474, + 556, + 583 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 474, + 556, + 583 + ], + "type": "text", + "content": ", and the sequence " + }, + { + "bbox": [ + 313, + 474, + 556, + 583 + ], + "type": "inline_equation", + "content": "(r_1, r_2, \\ldots, r_{k-1})" + }, + { + "bbox": [ + 313, + 474, + 556, + 583 + ], + "type": "text", + "content": " serves as the \"prefix\" for " + }, + { + "bbox": [ + 313, + 474, + 556, + 583 + ], + "type": "inline_equation", + "content": "r_k" + }, + { + "bbox": [ + 313, + 474, + 556, + 583 + ], + "type": "text", + "content": ". In this paradigm, during each autoregressive step " + }, + { + "bbox": [ + 313, + 474, + 556, + 583 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 474, + 556, + 583 + ], + "type": "text", + "content": ", the model predicts all " + }, + { + "bbox": [ + 313, + 474, + 556, + 583 + ], + "type": "inline_equation", + "content": "h_k \\times w_k" + }, + { + "bbox": [ + 313, + 474, + 556, + 583 + ], + "type": "text", + "content": " tokens in " + }, + { + "bbox": [ + 313, + 474, + 556, + 583 + ], + "type": "inline_equation", + "content": "r_k" + }, + { + "bbox": [ + 313, + 474, + 556, + 583 + ], + "type": "text", + "content": " in parallel, conditioned on prior scales and position embeddings. This approach aligns with a coarse-to-fine generation pattern, enabling parallel decoding within each scale. VAR effectively improves inference speed and generation quality, but it also considerably increases sequence length." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 588, + 421, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 588, + 421, + 601 + ], + "spans": [ + { + "bbox": [ + 313, + 588, + 421, + 601 + ], + "type": "text", + "content": "3.2. Key Observations" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 605, + 555, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 555, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 555, + 689 + ], + "type": "text", + "content": "VAR represents an innovative paradigm specific for visual generation, distinct from traditional autoregressive approaches and introducing many yet unexplored characteristics. In this work, we revisit the entire next-scale prediction process to uncover specific properties that can be optimized to reduce computational redundancy and improve inference efficiency." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 689, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 556, + 714 + ], + "type": "text", + "content": "Observation 1: As the predicted scale becomes larger, the need for parameters reduces significantly. While VAR" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "23336" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 71, + 375, + 244 + ], + "blocks": [ + { + "bbox": [ + 59, + 71, + 375, + 244 + ], + "lines": [ + { + "bbox": [ + 59, + 71, + 375, + 244 + ], + "spans": [ + { + "bbox": [ + 59, + 71, + 375, + 244 + ], + "type": "image", + "image_path": "091c168ef7e121154f6d60de6c1db09622d48246121795b963374b64f34389f9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 253, + 555, + 301 + ], + "lines": [ + { + "bbox": [ + 54, + 253, + 555, + 301 + ], + "spans": [ + { + "bbox": [ + 54, + 253, + 555, + 301 + ], + "type": "text", + "content": "Figure 4. Overview of the collaborative decoding process, we use a drafting step " + }, + { + "bbox": [ + 54, + 253, + 555, + 301 + ], + "type": "inline_equation", + "content": "N = 6" + }, + { + "bbox": [ + 54, + 253, + 555, + 301 + ], + "type": "text", + "content": " for instance. CoDe uses a large VAR model as the drafter " + }, + { + "bbox": [ + 54, + 253, + 555, + 301 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta_d}" + }, + { + "bbox": [ + 54, + 253, + 555, + 301 + ], + "type": "text", + "content": " to generate the token maps " + }, + { + "bbox": [ + 54, + 253, + 555, + 301 + ], + "type": "inline_equation", + "content": "R_L = (r_1, r_2, \\dots, r_N)" + }, + { + "bbox": [ + 54, + 253, + 555, + 301 + ], + "type": "text", + "content": " at smaller scales. The small refiner model " + }, + { + "bbox": [ + 54, + 253, + 555, + 301 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta_r}" + }, + { + "bbox": [ + 54, + 253, + 555, + 301 + ], + "type": "text", + "content": " then uses " + }, + { + "bbox": [ + 54, + 253, + 555, + 301 + ], + "type": "inline_equation", + "content": "R_L" + }, + { + "bbox": [ + 54, + 253, + 555, + 301 + ], + "type": "text", + "content": " as an initial prefix to efficiently predict the remaining token maps " + }, + { + "bbox": [ + 54, + 253, + 555, + 301 + ], + "type": "inline_equation", + "content": "R_H = (r_{N+1}, r_{N+2}, \\dots, r_K)" + }, + { + "bbox": [ + 54, + 253, + 555, + 301 + ], + "type": "text", + "content": " at larger scales. Both models are fine-tuned on their designated predictive scales using ground truth labels " + }, + { + "bbox": [ + 54, + 253, + 555, + 301 + ], + "type": "inline_equation", + "content": "r_k^*" + }, + { + "bbox": [ + 54, + 253, + 555, + 301 + ], + "type": "text", + "content": " and teacher logits " + }, + { + "bbox": [ + 54, + 253, + 555, + 301 + ], + "type": "inline_equation", + "content": "p_{\\text{teacher}}(r_k)" + }, + { + "bbox": [ + 54, + 253, + 555, + 301 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 384, + 102, + 553, + 237 + ], + "blocks": [ + { + "bbox": [ + 384, + 102, + 553, + 237 + ], + "lines": [ + { + "bbox": [ + 384, + 102, + 553, + 237 + ], + "spans": [ + { + "bbox": [ + 384, + 102, + 553, + 237 + ], + "type": "image", + "image_path": "7343d83148610ce5b6f70d0c9222955db4eb4eede9d0c4b62e016674eb7b9a67.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 319, + 297, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 319, + 297, + 510 + ], + "spans": [ + { + "bbox": [ + 54, + 319, + 297, + 510 + ], + "type": "text", + "content": "models exhibit strong scalability, the effectiveness of increasing parameter counts varies greatly in predicting various scales. In Figure 3. (a), we present the differences in class-conditional generation (ImageNet-256 [8]) quality when using VAR models of different sizes at each scale. At the initial small scales, increasing model parameters leads to noticeable improvements in generation quality. However, as the predicted scale gradually progresses, the impact of additional parameters becomes minimal. At the final scale—which accounts for " + }, + { + "bbox": [ + 54, + 319, + 297, + 510 + ], + "type": "inline_equation", + "content": "38\\%" + }, + { + "bbox": [ + 54, + 319, + 297, + 510 + ], + "type": "text", + "content": " tokens of the entire sequence—we observe that the performance of the 2B model is even nearly equivalent to that of the 0.3B model. These findings suggest that as the predicted scale increases, the parameter demand for accurate token predictions declines markedly, revealing substantial computational redundancy in the current VAR inference process at large scales." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 522, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 522, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 522, + 297, + 714 + ], + "type": "text", + "content": "Observation 2: The generative patterns are exclusive between small and large scales. For VAR models, the content generated at small and large scales varies significantly. As demonstrated by the Fourier analysis in Figure 3 (b), the feature maps produced by the first three scales contain primarily low-frequency components, while the last three scales focus on high-frequency components. To further validate these distinct generative patterns, we conducted a perturbation fine-tuning experiment. Using a pre-trained VARd16 model, we applied CSE loss only to tokens in the largest three scales and fine-tuned for just " + }, + { + "bbox": [ + 55, + 522, + 297, + 714 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 55, + 522, + 297, + 714 + ], + "type": "text", + "content": " of the original training epochs. This minor fine-tuning at large scales led to a complete collapse of the model's global modeling capacity at small scales, with the FID increasing from 3.30 to 21.93 and the IS score dropping from 277 to 88. These results demonstrate that VAR models perform entirely dis" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 319, + 555, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 319, + 555, + 378 + ], + "spans": [ + { + "bbox": [ + 313, + 319, + 555, + 378 + ], + "type": "text", + "content": "tinct generative tasks at small and large scales, with minimal overlap. Training a single VAR model to predict across all scales results in significant mutual interference between learning low-frequency and high-frequency modeling capabilities, making parameter optimization challenging." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 379, + 556, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 379, + 556, + 477 + ], + "spans": [ + { + "bbox": [ + 313, + 379, + 556, + 477 + ], + "type": "text", + "content": "In sum, the key insights from our observations can be concluded as follows: (1) At large scales, VAR models require significantly fewer parameters for accurate token prediction, leading to considerable computational redundancy. (2) Training a single model to generate across all scales causes mutual interference between small and large scales, hindering parameter optimization due to conflicting low-frequency and high-frequency learning tendencies." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 480, + 448, + 494 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 480, + 448, + 494 + ], + "spans": [ + { + "bbox": [ + 313, + 480, + 448, + 494 + ], + "type": "text", + "content": "3.3. Collaborative Decoding" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 498, + 556, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 556, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 556, + 653 + ], + "type": "text", + "content": "Based on the above insights, we propose a simple yet pretty efficient decoding method for next-scale prediction called collaborative decoding. As presented in Figure 4, CoDe decomposes the next-scale prediction process into a collaboration between a large drafter model and a small refiner model. The drafter is the original large VAR model with 2B parameters, while the refiner is a lightweight 0.3B VAR model, sharing the same transformer-based architecture but with reduced width and depth. The drafter is responsible for generating coarse, low-frequency global structures at smaller scales to draft the image, while the refiner continually predicts the high-frequency details at larger scales to refine the image." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "text", + "content": "Model Collaboration. To generate an image " + }, + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "inline_equation", + "content": "im" + }, + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "text", + "content": ", original VAR model needs to auto-regressively generate " + }, + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "text", + "content": " multiscale token maps " + }, + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "inline_equation", + "content": "R = (r_1, r_2, \\ldots, r_K)" + }, + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "text", + "content": " with progressively finer resolutions, where each token map " + }, + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "inline_equation", + "content": "r_k \\in [V]^{h_k \\times w_k}" + }, + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "text", + "content": " represents the token map at scale " + }, + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "text", + "content": ". In CoDe, we decom" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23337" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "content": "pose the auto-regressive steps into " + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "content": " drafting steps and " + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "inline_equation", + "content": "K - N" + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "content": " refining steps. In the initial drafting phase where computation is sparse but additional parameters provide significant benefits, we serve a large VAR model as the drafter " + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta_d}" + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "content": ", generating the initial set of low-frequency token maps " + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "inline_equation", + "content": "R_L = (r_1,r_2,\\dots ,r_N)" + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "content": " up to scale " + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "inline_equation", + "content": "N < K" + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "content": ". These token maps predict the global structure of the image, serving as a prefix for the remaining finer-scale token maps. The drafting process can be represented as:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 185, + 296, + 217 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 185, + 296, + 217 + ], + "spans": [ + { + "bbox": [ + 61, + 185, + 296, + 217 + ], + "type": "interline_equation", + "content": "p _ {\\theta_ {d}} \\left(r _ {1}, r _ {2}, \\dots , r _ {N}\\right) = \\prod_ {k = 1} ^ {N} p _ {\\theta_ {d}} \\left(r _ {k} \\mid r _ {1}, r _ {2}, \\dots , r _ {k - 1}\\right). \\tag {2}", + "image_path": "d9ecdb5c456611d57243aabd2128133674711609fbdc016d06769aab403ec9fc.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 221, + 295, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 221, + 295, + 245 + ], + "spans": [ + { + "bbox": [ + 55, + 221, + 295, + 245 + ], + "type": "text", + "content": "After drafting, the KV cache of the large model is released, significantly optimizing memory usage." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 246, + 296, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 246, + 296, + 365 + ], + "spans": [ + { + "bbox": [ + 55, + 246, + 296, + 365 + ], + "type": "text", + "content": "In the refining stage where the computation is intensive but the parameter demand is low, we serve a small VAR model as the refiner " + }, + { + "bbox": [ + 55, + 246, + 296, + 365 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta_r}" + }, + { + "bbox": [ + 55, + 246, + 296, + 365 + ], + "type": "text", + "content": ", employing the drafter's predictions " + }, + { + "bbox": [ + 55, + 246, + 296, + 365 + ], + "type": "inline_equation", + "content": "R_L" + }, + { + "bbox": [ + 55, + 246, + 296, + 365 + ], + "type": "text", + "content": " as a prefix and focus on refining the details by generating the remaining high-frequency token maps " + }, + { + "bbox": [ + 55, + 246, + 296, + 365 + ], + "type": "inline_equation", + "content": "R_H = (r_{N + 1}, r_{N + 2}, \\ldots, r_K)" + }, + { + "bbox": [ + 55, + 246, + 296, + 365 + ], + "type": "text", + "content": " up to scale " + }, + { + "bbox": [ + 55, + 246, + 296, + 365 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 246, + 296, + 365 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 246, + 296, + 365 + ], + "type": "inline_equation", + "content": "r_K" + }, + { + "bbox": [ + 55, + 246, + 296, + 365 + ], + "type": "text", + "content": " matches the resolution of the original feature map. Notably, predicting " + }, + { + "bbox": [ + 55, + 246, + 296, + 365 + ], + "type": "inline_equation", + "content": "r_{N + 1}" + }, + { + "bbox": [ + 55, + 246, + 296, + 365 + ], + "type": "text", + "content": " requires an attention mask, as the refiner model lacks the KV cache for the first " + }, + { + "bbox": [ + 55, + 246, + 296, + 365 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 246, + 296, + 365 + ], + "type": "text", + "content": " scales. The probability distribution of refining steps is formulated as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 85, + 369, + 295, + 418 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 369, + 295, + 418 + ], + "spans": [ + { + "bbox": [ + 85, + 369, + 295, + 418 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} p \\left(r _ {N + 1}, r _ {N + 2}, \\dots , r _ {K} \\mid R _ {L}\\right) \\\\ = \\prod_ {k = N + 1} ^ {K} p \\left(r _ {k} \\mid R _ {L}, r _ {N + 1}, r _ {N + 2}, \\dots , r _ {k - 1}\\right). \\tag {3} \\\\ \\end{array}", + "image_path": "9f4c6b8d04933752c3ccde4656d6c103560b382f0428086bac27b1b227235037.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 423, + 296, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 423, + 296, + 448 + ], + "spans": [ + { + "bbox": [ + 55, + 423, + 296, + 448 + ], + "type": "text", + "content": "Finally, the generated images " + }, + { + "bbox": [ + 55, + 423, + 296, + 448 + ], + "type": "inline_equation", + "content": "im" + }, + { + "bbox": [ + 55, + 423, + 296, + 448 + ], + "type": "text", + "content": " are reconstructed from both drafting maps " + }, + { + "bbox": [ + 55, + 423, + 296, + 448 + ], + "type": "inline_equation", + "content": "R_{L}" + }, + { + "bbox": [ + 55, + 423, + 296, + 448 + ], + "type": "text", + "content": " and refining maps " + }, + { + "bbox": [ + 55, + 423, + 296, + 448 + ], + "type": "inline_equation", + "content": "R_{H}" + }, + { + "bbox": [ + 55, + 423, + 296, + 448 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 129, + 458, + 295, + 472 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 458, + 295, + 472 + ], + "spans": [ + { + "bbox": [ + 129, + 458, + 295, + 472 + ], + "type": "interline_equation", + "content": "i m = D \\left(Q \\left(R _ {L}, R _ {H}\\right), \\right. \\tag {4}", + "image_path": "4cd1557444c1d7665eabe3852d3a2c53b5e04e1c50b5713b6caced7effb64d2f.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 474, + 295, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 474, + 295, + 593 + ], + "spans": [ + { + "bbox": [ + 55, + 474, + 295, + 593 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 474, + 295, + 593 + ], + "type": "inline_equation", + "content": "Q(.)" + }, + { + "bbox": [ + 55, + 474, + 295, + 593 + ], + "type": "text", + "content": " is a residual-style quantization function, and " + }, + { + "bbox": [ + 55, + 474, + 295, + 593 + ], + "type": "inline_equation", + "content": "D(.)" + }, + { + "bbox": [ + 55, + 474, + 295, + 593 + ], + "type": "text", + "content": " is a multi-scale VQVAE decoder. Figure 3. (c) demonstrates the remarkable effectiveness of model collaboration under training-free conditions. With a large model as the drafter, the small refiner only results in a slight and gentle increase in FID. Conversely, with a small model as the drafter, even assigning the last " + }, + { + "bbox": [ + 55, + 474, + 295, + 593 + ], + "type": "inline_equation", + "content": "80\\%" + }, + { + "bbox": [ + 55, + 474, + 295, + 593 + ], + "type": "text", + "content": " tokens to a large refiner fails to improve performance. This also verifies Observation 1: small scales require more model capacity, whereas the parameter demands for large scales are low." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 594, + 295, + 642 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 642 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 642 + ], + "type": "text", + "content": "The model collaboration method effectively reduces the computational redundancy of VAR and maintains generation quality comparable to the original model, while significantly offering faster speed and reduced memory usage." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": "Specialized Fine-Tuning. Given the exclusive generation patterns between drafting scales and refining scales, we propose the specialized fine-tuning to further specialize drafter and refiner models by fine-tuning them solely on their respective predictive scales, thereby avoiding training interference and enhancing generation quality." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "text", + "content": "For the drafter model " + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta_d}" + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "text", + "content": ", we fine-tune the drafting steps to improve the generation of the initial low-frequency token maps " + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "inline_equation", + "content": "R_L" + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "text", + "content": ". The drafter model is trained using a CSE loss " + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "inline_equation", + "content": "\\mathrm{CSE}(\\cdot, \\cdot)" + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "text", + "content": " between its generated token distribution " + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "inline_equation", + "content": "p_{\\theta_d}(r_k)" + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "text", + "content": " and the ground truth labels " + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "inline_equation", + "content": "r_k^*" + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "text", + "content": ", defined as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 368, + 140, + 554, + 172 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 368, + 140, + 554, + 172 + ], + "spans": [ + { + "bbox": [ + 368, + 140, + 554, + 172 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {d r a f t e r}} = \\sum_ {k = 1} ^ {N} \\mathrm {C S E} \\left(p _ {\\theta_ {d}} \\left(r _ {k}\\right), r _ {k} ^ {*}\\right), \\tag {5}", + "image_path": "d91de98c5b2c1c5aaddef29168861857d82cbf2756a4124080038125097441dd.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 179, + 554, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 179, + 554, + 202 + ], + "spans": [ + { + "bbox": [ + 313, + 179, + 554, + 202 + ], + "type": "text", + "content": "which encourages the drafter to closely match the target global structure at each drafting step." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 203, + 554, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 203, + 554, + 251 + ], + "spans": [ + { + "bbox": [ + 313, + 203, + 554, + 251 + ], + "type": "text", + "content": "For the refiner model " + }, + { + "bbox": [ + 313, + 203, + 554, + 251 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta_r}" + }, + { + "bbox": [ + 313, + 203, + 554, + 251 + ], + "type": "text", + "content": ", we adopt a Knowledge Distillation (KD) [19] approach to partially transfer the knowledge from a larger pre-trained VAR model. The KD loss function for epoch " + }, + { + "bbox": [ + 313, + 203, + 554, + 251 + ], + "type": "inline_equation", + "content": "ep" + }, + { + "bbox": [ + 313, + 203, + 554, + 251 + ], + "type": "text", + "content": " can be expressed as:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 314, + 257, + 556, + 295 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 257, + 556, + 295 + ], + "spans": [ + { + "bbox": [ + 314, + 257, + 556, + 295 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {r e f i n e r}} ^ {(e p)} = \\sum_ {k = 1} ^ {K} \\left(\\lambda_ {e p} \\cdot \\mathbf {1} _ {[ k \\leq N ]} + \\mathbf {1} _ {[ k > N ]}\\right) \\mathrm {K L} \\left(p _ {\\theta_ {r}} \\left(r _ {k}\\right) \\| p _ {\\text {t e a c h e r}} \\left(r _ {k}\\right)\\right), \\tag {6}", + "image_path": "9ac7104932222ed42b8aa2fd7b84ea0047a9ad3190b948578745e315eaa249d6.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 296, + 554, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 296, + 554, + 403 + ], + "spans": [ + { + "bbox": [ + 313, + 296, + 554, + 403 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 296, + 554, + 403 + ], + "type": "inline_equation", + "content": "p_{\\mathrm{teacher}}(\\cdot)" + }, + { + "bbox": [ + 313, + 296, + 554, + 403 + ], + "type": "text", + "content": " is the distribution predicted by the larger teacher VAR model, " + }, + { + "bbox": [ + 313, + 296, + 554, + 403 + ], + "type": "inline_equation", + "content": "\\mathrm{KL}(\\cdot, \\cdot)" + }, + { + "bbox": [ + 313, + 296, + 554, + 403 + ], + "type": "text", + "content": " is the standard KL-divergence loss aligns refiner's output distribution with that of the teacher model, and " + }, + { + "bbox": [ + 313, + 296, + 554, + 403 + ], + "type": "inline_equation", + "content": "\\lambda_{ep}" + }, + { + "bbox": [ + 313, + 296, + 554, + 403 + ], + "type": "text", + "content": " is a dynamic weighting factor. " + }, + { + "bbox": [ + 313, + 296, + 554, + 403 + ], + "type": "inline_equation", + "content": "\\lambda_t" + }, + { + "bbox": [ + 313, + 296, + 554, + 403 + ], + "type": "text", + "content": " linearly decreases from one to zero from the initial to the end of finetuning, gradually shifting the KD emphasis from all tokens to the specific refining tokens. This dynamic weighting allows the refiner to specialize in refining the details with a smoother and more stable training process." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 403, + 554, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 403, + 554, + 475 + ], + "spans": [ + { + "bbox": [ + 313, + 403, + 554, + 475 + ], + "type": "text", + "content": "The specialized fine-tuning of the drafter and refiner models ensures that each model becomes highly proficient in its respective task. By minimizing interference between small and large scales, this approach allows for more sufficient parameter optimization, resulting in enhanced image generation quality with limited training costs." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 485, + 440, + 498 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 485, + 440, + 498 + ], + "spans": [ + { + "bbox": [ + 313, + 485, + 440, + 498 + ], + "type": "text", + "content": "4. Experimental Results" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 505, + 433, + 517 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 505, + 433, + 517 + ], + "spans": [ + { + "bbox": [ + 313, + 505, + 433, + 517 + ], + "type": "text", + "content": "4.1. Experimental Setup." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "type": "text", + "content": "Implementation Details. We evaluated our method's effectiveness on the ImageNet1K [8] class-conditional generation benchmarks. The proposed CoDe framework involves collaboration between a large and a small VAR model. Specifically, we use the pre-trained VAR-d30 model as the drafter and the pre-trained VAR-d16 model as the refiner. Each model undergoes specialized fine-tuning focused exclusively on its respective predictive scales. For the drafter, we fine-tune it for " + }, + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "type": "text", + "content": " of its original training epochs with a base learning rate of 1e-6 and a weight decay of 0.08. The refiner model, on the other hand, is fine-tuned for " + }, + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "type": "text", + "content": " of its original training epochs with a base learning rate of 1e-5 and no weight decay. Both models are optimized using the AdamW [21] optimizer with a batch size of 1024, achieved through gradient accumulation. The fine-tuning was conducted on 4 NVIDIA L20 GPUs." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "23338" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 70, + 219, + 177 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 219, + 177 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 219, + 177 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 219, + 177 + ], + "type": "image", + "image_path": "dc55050fcdac980dc9e1d9d489ef7462f9451c1081fd7f37592f82b10ecd2d99.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 85, + 178, + 209, + 189 + ], + "lines": [ + { + "bbox": [ + 85, + 178, + 209, + 189 + ], + "spans": [ + { + "bbox": [ + 85, + 178, + 209, + 189 + ], + "type": "text", + "content": "(a) Efficiency & Quality trade-off" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 224, + 70, + 386, + 176 + ], + "blocks": [ + { + "bbox": [ + 224, + 70, + 386, + 176 + ], + "lines": [ + { + "bbox": [ + 224, + 70, + 386, + 176 + ], + "spans": [ + { + "bbox": [ + 224, + 70, + 386, + 176 + ], + "type": "image", + "image_path": "4a5834facc896802db97a7dc744ebe5cbdb464c67ba2bb8e5a1055abf5330c48.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 198, + 555, + 232 + ], + "lines": [ + { + "bbox": [ + 55, + 198, + 555, + 232 + ], + "spans": [ + { + "bbox": [ + 55, + 198, + 555, + 232 + ], + "type": "text", + "content": "Figure 5. (a) Our CoDe demonstrates the optimal efficiency-quality trade-off among all evaluated methods. (b) Inference latency is measured across varying batch sizes for the original VAR-d30, our CoDe " + }, + { + "bbox": [ + 55, + 198, + 555, + 232 + ], + "type": "inline_equation", + "content": "(N = 6)" + }, + { + "bbox": [ + 55, + 198, + 555, + 232 + ], + "type": "text", + "content": ", and the VQVAE decoder. (c) We analyze the time cost associated with parallel decoding at each scale, showing that the refiner model is significantly more efficient than the drafter at larger scales." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 387, + 70, + 548, + 176 + ], + "blocks": [ + { + "bbox": [ + 242, + 178, + 382, + 189 + ], + "lines": [ + { + "bbox": [ + 242, + 178, + 382, + 189 + ], + "spans": [ + { + "bbox": [ + 242, + 178, + 382, + 189 + ], + "type": "text", + "content": "(b) Time cost with different batchsizes" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 387, + 70, + 548, + 176 + ], + "lines": [ + { + "bbox": [ + 387, + 70, + 548, + 176 + ], + "spans": [ + { + "bbox": [ + 387, + 70, + 548, + 176 + ], + "type": "image", + "image_path": "c7c8080d11d76784bfaf5c2250f05668741b1712e2d7b5cd25d961cd1d874316.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 427, + 178, + 525, + 188 + ], + "lines": [ + { + "bbox": [ + 427, + 178, + 525, + 188 + ], + "spans": [ + { + "bbox": [ + 427, + 178, + 525, + 188 + ], + "type": "text", + "content": "(c) Time cost at each scale" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 56, + 274, + 555, + 446 + ], + "blocks": [ + { + "bbox": [ + 55, + 243, + 555, + 266 + ], + "lines": [ + { + "bbox": [ + 55, + 243, + 555, + 266 + ], + "spans": [ + { + "bbox": [ + 55, + 243, + 555, + 266 + ], + "type": "text", + "content": "Table 1. Quantitative assessment of the efficiency-quality trade-off across various methods. Inference efficiency is evaluated with a batch size of 64 on NVIDIA L20 GPU, with latency measured excluding VQVAE or VQGAN as it incurs a shared time cost across all methods." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 274, + 555, + 446 + ], + "lines": [ + { + "bbox": [ + 56, + 274, + 555, + 446 + ], + "spans": [ + { + "bbox": [ + 56, + 274, + 555, + 446 + ], + "type": "table", + "html": "
MethodInference EfficiencyGeneration Quality
#StepsSpeedup↑Latency↓Throughput↑#ParamMemory↓FID ↓IS ↑Precision↑Recall↑
DiT-XL/2500.2x19.20s3.33it/s675M11369MB2.262390.800.60
MAR-B1000.1x29.80s2.15it/s208M8725MB2.312820.820.57
AiM-XL2560.4x9.32s6.87it/s763M20983MB2.562570.820.57
LlamaGen-XXL384<0.1x73.97s0.87it/s1.4B42632MB2.342540.800.59
VAR-d30101.0x3.62s17.71it/s2.0B39228MB1.953010.810.59
VAR-d24101.7x2.07s30.92it/s1.0B25093MB2.113110.820.59
VAR-d20102.8x1.29s49.62it/s600M17814MB2.613010.830.56
VAR-CoDe N=99+11.2x2.97s21.54it/s2.0+0.3B28803MB1.942960.800.61
VAR-CoDe N=88+21.7x2.11s30.33it/s2.0+0.3B21019MB1.983020.810.60
VAR-CoDe N=77+32.3x1.60s40.00it/s2.0+0.3B19943MB2.113030.820.59
VAR-CoDe N=66+42.9x1.27s50.39it/s2.0+0.3B19943MB2.272970.820.58
", + "image_path": "7c703a633b6ae10e5aeb4abf6571a1442e3f61874087ae216a9692e431a1c9e4.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 63, + 485, + 286, + 532 + ], + "blocks": [ + { + "bbox": [ + 100, + 464, + 251, + 476 + ], + "lines": [ + { + "bbox": [ + 100, + 464, + 251, + 476 + ], + "spans": [ + { + "bbox": [ + 100, + 464, + 251, + 476 + ], + "type": "text", + "content": "Table 2. Effect of specialized fine-tuning" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 63, + 485, + 286, + 532 + ], + "lines": [ + { + "bbox": [ + 63, + 485, + 286, + 532 + ], + "spans": [ + { + "bbox": [ + 63, + 485, + 286, + 532 + ], + "type": "table", + "html": "
DiscriptionN=6N=7N=8N=9
CoDe Training-free2.422.262.101.99
CoDe Fine-tuning2.272.111.981.94
", + "image_path": "78682c545e01d162f00410402f9a09f59115e1d3c193206ffe00f66f05285bdc.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 557, + 297, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 557, + 297, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 557, + 297, + 713 + ], + "type": "text", + "content": "**Evaluations.** We evaluated our method on ImageNet1K-256 generation, focusing on both quality and efficiency metrics. For quality assessment, we used standard metrics such as FID, Inception Score (IS), Precision, and Recall. For other baselines, we use their default sampling methods. For CoDe, we reduced the sampled top-k from 900 to 600 due to improved token prediction accuracy after fine-tuning, while keeping the default top-p at 0.96. To compensate for reduced diversity, we introduced a temperature coefficient of " + }, + { + "bbox": [ + 55, + 557, + 297, + 713 + ], + "type": "inline_equation", + "content": "t = 1.1" + }, + { + "bbox": [ + 55, + 557, + 297, + 713 + ], + "type": "text", + "content": " for sampling on the smallest 7 scales. Efficiency was measured through inference latency, throughput, memory consumption, and parameter count to provide a comprehensive comparison. Notably, the speed measurements" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 466, + 555, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 466, + 555, + 526 + ], + "spans": [ + { + "bbox": [ + 313, + 466, + 555, + 526 + ], + "type": "text", + "content": "exclude the VQGAN decoder, which contributes minimally to overall runtime and is a shared component across all methods. All efficiency tests were conducted on a single NVIDIA L20 GPU without additional optimizations such as FlashAttention [7], using PyTorch 2.1 [40] and FP16." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 537, + 402, + 549 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 537, + 402, + 549 + ], + "spans": [ + { + "bbox": [ + 313, + 537, + 402, + 549 + ], + "type": "text", + "content": "4.2. Main Results." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 556, + 555, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 556, + 555, + 676 + ], + "spans": [ + { + "bbox": [ + 313, + 556, + 555, + 676 + ], + "type": "text", + "content": "Quality-Efficiency Trade-off. We evaluated the quality-efficiency trade-off of our proposed method (CoDe) against original VAR models [51], state-of-the-art AR models, and diffusion models. The AR models included GPT-style LlamaGEN [48], Mamba-based Aim [22], and MAE-style MAR [26]. For diffusion models, we employed the widely used DiT models [41]. To measure the trade-offs, we used series models with different parameter counts for VAR, LlamaGEN, Aim, and MAR, while for DiT we varied the DDIM [47] steps to generate the efficiency-quality curve." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "content": "As illustrated in Table 1, We analyzed CoDe's performance under different drafting steps N. Compared to traditional AR models, our proposed VAR-CoDe " + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "inline_equation", + "content": "(N = 6)" + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "content": " achieves" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "23339" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 71, + 553, + 270 + ], + "blocks": [ + { + "bbox": [ + 57, + 71, + 553, + 270 + ], + "lines": [ + { + "bbox": [ + 57, + 71, + 553, + 270 + ], + "spans": [ + { + "bbox": [ + 57, + 71, + 553, + 270 + ], + "type": "image", + "image_path": "aa9b79272d373c128c4d3be3d5a6b11633b164ec4a9e1aa8863fe6d5dbc50c31.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 66, + 279, + 543, + 292 + ], + "lines": [ + { + "bbox": [ + 66, + 279, + 543, + 292 + ], + "spans": [ + { + "bbox": [ + 66, + 279, + 543, + 292 + ], + "type": "text", + "content": "Figure 6. Qualitative comparison between the original VAR-d30 model and our proposed CoDe model, with different drafting steps." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 312, + 295, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 312, + 295, + 479 + ], + "spans": [ + { + "bbox": [ + 54, + 312, + 295, + 479 + ], + "type": "text", + "content": "60 times faster inference than LlamaGEN-XXL, while also surpassing its generation quality. Additionally, in comparison to diffusion models, VAR-CoDe " + }, + { + "bbox": [ + 54, + 312, + 295, + 479 + ], + "type": "inline_equation", + "content": "(N = 6)" + }, + { + "bbox": [ + 54, + 312, + 295, + 479 + ], + "type": "text", + "content": " is 15 times faster than DiT-XL/2, while maintaining the same level of quality. Compared to the original VAR-d30, CoDe " + }, + { + "bbox": [ + 54, + 312, + 295, + 479 + ], + "type": "inline_equation", + "content": "(N = 8)" + }, + { + "bbox": [ + 54, + 312, + 295, + 479 + ], + "type": "text", + "content": " achieves a " + }, + { + "bbox": [ + 54, + 312, + 295, + 479 + ], + "type": "inline_equation", + "content": "1.7\\mathrm{x}" + }, + { + "bbox": [ + 54, + 312, + 295, + 479 + ], + "type": "text", + "content": " speedup and reduces memory consumption by " + }, + { + "bbox": [ + 54, + 312, + 295, + 479 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 54, + 312, + 295, + 479 + ], + "type": "text", + "content": ", with only a negligible FID increase from 1.95 to 1.98. When the drafting stage is reduced to just 6 steps, CoDe " + }, + { + "bbox": [ + 54, + 312, + 295, + 479 + ], + "type": "inline_equation", + "content": "(N = 6)" + }, + { + "bbox": [ + 54, + 312, + 295, + 479 + ], + "type": "text", + "content": " achieves a " + }, + { + "bbox": [ + 54, + 312, + 295, + 479 + ], + "type": "inline_equation", + "content": "2.9\\mathrm{x}" + }, + { + "bbox": [ + 54, + 312, + 295, + 479 + ], + "type": "text", + "content": " speedup, reaching a throughput of over 50it/s, while maintaining a low FID of 2.27. This is a speed unmatched by any other existing methods. Compared to other VAR models with fewer parameters (VAR-d24 and VAR-d20), our method achieves significantly better quality while maintaining the same speedup ratio." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 481, + 295, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 481, + 295, + 531 + ], + "spans": [ + { + "bbox": [ + 55, + 481, + 295, + 531 + ], + "type": "text", + "content": "As shown in Figure 5 (a), VAR-CoDe achieves the best efficiency-quality trade-off compared to all other methods, effectively solving the inefficiencies introduced by the extended sequence length in the original VAR paradigm." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 531, + 295, + 699 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 531, + 295, + 699 + ], + "spans": [ + { + "bbox": [ + 55, + 531, + 295, + 699 + ], + "type": "text", + "content": "Training-Free vs Specialized Finetuning. Our CoDe framework can operate in a training-free manner by directly using pre-trained VAR-d30 and VAR-d16 models as the drafter and refiner, respectively. Alternatively, we can perform specialized fine-tuning to further enhance the models' performance at their respective scales, with limited training cost. Table 2 compares the results of training-free CoDe with the specialized fine-tuned version. Even without additional training, CoDe demonstrates competitive performance, outperforming VAR-d24 and VAR-d20 models with the same speedup ratio. With specialized fine-tuning, CoDe's performance improves significantly, even achieving a slight quality enhancement over the original model at a " + }, + { + "bbox": [ + 55, + 531, + 295, + 699 + ], + "type": "inline_equation", + "content": "1.2\\mathrm{x}" + }, + { + "bbox": [ + 55, + 531, + 295, + 699 + ], + "type": "text", + "content": " acceleration ratio." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 701, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 701, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 701, + 295, + 714 + ], + "type": "text", + "content": "Qualitative Results. We provide an extensive qualitative" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 312, + 555, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 312, + 555, + 456 + ], + "spans": [ + { + "bbox": [ + 313, + 312, + 555, + 456 + ], + "type": "text", + "content": "comparison between the original VAR-d30 model and our proposed CoDe, with varying drafting steps " + }, + { + "bbox": [ + 313, + 312, + 555, + 456 + ], + "type": "inline_equation", + "content": "N = \\{6,7,8\\}" + }, + { + "bbox": [ + 313, + 312, + 555, + 456 + ], + "type": "text", + "content": ". As illustrated in Figure 6, our approach achieves significant speedup and substantial memory optimization, with only minimal quality degradation that is nearly imperceptible to the human eye. Even at a speedup rate of 2.9 times, the generated images maintain high quality and accurate semantic information. Through specialized fine-tuning, CoDe's draft model exhibits higher predictive accuracy compared to the original model. This sometimes results in a different global structure from the original output, yet the image quality remains consistently high or even better." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 456, + 556, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 456, + 556, + 624 + ], + "spans": [ + { + "bbox": [ + 313, + 456, + 556, + 624 + ], + "type": "text", + "content": "Zero-shot Task Generalization. To evaluate the zero-shot generalization capability of CoDe, we conducted additional experiments on zero-shot class-conditional inpainting and image editing. During image inpainting, we applied teacher forcing by providing the ground truth tokens outside the masked area, allowing the model to generate tokens solely within the mask. Notably, class-conditional information was introduced to the model. In the image editing task, CoDe was restricted to generating tokens within a given bounding box based on a specific class label. Figure 7 illustrates the qualitative results of CoDe's zero-shot performance " + }, + { + "bbox": [ + 313, + 456, + 556, + 624 + ], + "type": "inline_equation", + "content": "(N = 8)" + }, + { + "bbox": [ + 313, + 456, + 556, + 624 + ], + "type": "text", + "content": ". Our approach demonstrates strong zero-shot generalization without any additional training on these downstream tasks." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 635, + 428, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 635, + 428, + 647 + ], + "spans": [ + { + "bbox": [ + 313, + 635, + 428, + 647 + ], + "type": "text", + "content": "4.3. Efficiency Analysis." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 653, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 555, + 713 + ], + "type": "text", + "content": "Time Cost Analysis. We first present the time cost (bs=64) for each decoding step of the VAR-d30 model in Figure 5 (c). Notably, the computational complexity across the ten decoding steps of VAR is highly non-uniform. The last three decoding steps alone account for " + }, + { + "bbox": [ + 313, + 653, + 555, + 713 + ], + "type": "inline_equation", + "content": "64\\%" + }, + { + "bbox": [ + 313, + 653, + 555, + 713 + ], + "type": "text", + "content": " of the total in" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23340" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 72, + 557, + 250 + ], + "blocks": [ + { + "bbox": [ + 61, + 72, + 557, + 250 + ], + "lines": [ + { + "bbox": [ + 61, + 72, + 557, + 250 + ], + "spans": [ + { + "bbox": [ + 61, + 72, + 557, + 250 + ], + "type": "image", + "image_path": "1717531237c044669f2b369a0fceacaf34053e4113ea17db356a0cde4a2088fc.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 117, + 261, + 492, + 272 + ], + "lines": [ + { + "bbox": [ + 117, + 261, + 492, + 272 + ], + "spans": [ + { + "bbox": [ + 117, + 261, + 492, + 272 + ], + "type": "text", + "content": "Figure 7. Qualitative results of CoDe's zero-shot generalization on image inpainting and image editing." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 56, + 312, + 301, + 446 + ], + "blocks": [ + { + "bbox": [ + 59, + 292, + 291, + 303 + ], + "lines": [ + { + "bbox": [ + 59, + 292, + 291, + 303 + ], + "spans": [ + { + "bbox": [ + 59, + 292, + 291, + 303 + ], + "type": "text", + "content": "Table 3. Memory usage comparison across different batch sizes" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 312, + 301, + 446 + ], + "lines": [ + { + "bbox": [ + 56, + 312, + 301, + 446 + ], + "spans": [ + { + "bbox": [ + 56, + 312, + 301, + 446 + ], + "type": "table", + "html": "
MethodMemory Consumption↓
RunningKV CacheParamsTotal
VAR (bs=8)314MB3595MB8089MB12002MB
+CoDe284MB1023MB9275MB10619MB
VAR (bs=16)615MB7191MB8089MB15901MB
+CoDe557MB2056MB9275MB11951MB
VAR (bs=32)1216MB14345MB8089MB23662MB
+CoDe1103MB4083MB9275MB14614MB
VAR (bs=64)2420MB28707MB8089MB39228MB
+CoDe2195MB8160MB9275MB19943MB
VAR (bs=128)OOM(0.48GB)OOM(57GB)OOM(0.80GB)OOM(70GB)
+CoDe4380MB16320MB9275MB30598MB
", + "image_path": "34d9c1a3addc044bc941aeb520366c11ac91d6e44f332c5d5a1a2763fab0764c.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 472, + 295, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 472, + 295, + 579 + ], + "spans": [ + { + "bbox": [ + 55, + 472, + 295, + 579 + ], + "type": "text", + "content": "ference time, with latency increasing quadratically as token map resolution grows. Our solution is to replace the large VAR model with a smaller one for the last few steps, as fewer parameters are needed for large-scale token maps. This small refiner is substantially faster than the original large drafter at these scales, achieving a 4.6x speed improvement in the final and most computationally demanding step. This change dramatically accelerates the entire inference process compared to the original VAR model." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "text", + "content": "Next, we present the speedup effect of CoDe " + }, + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "inline_equation", + "content": "(N = 6)" + }, + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "text", + "content": " across different batch sizes. As shown in Figure 5.(b), CoDe achieves a speedup of " + }, + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "inline_equation", + "content": "1.6\\mathrm{x}" + }, + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "text", + "content": " for batch size 1 but reaches up to " + }, + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "inline_equation", + "content": "2.9\\mathrm{x}" + }, + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "text", + "content": " for batch size 64. The reduced speedup at smaller batch sizes is due to low GPU utilization when generating " + }, + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "inline_equation", + "content": "256\\times 256" + }, + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "text", + "content": " images, which limits the acceleration potential. However, as GPU utilization increases with larger batch sizes, the speedup of CoDe also improves significantly. This indicates that CoDe is well-suited for computationally intensive image generation tasks, and holds great potential for enhancing efficiency in high-resolution image genera" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 293, + 553, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 293, + 553, + 317 + ], + "spans": [ + { + "bbox": [ + 313, + 293, + 553, + 317 + ], + "type": "text", + "content": "tion under the VAR paradigm. Additionally, we report the time cost of the VQVAE decoder in the figure." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 319, + 555, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 319, + 555, + 582 + ], + "spans": [ + { + "bbox": [ + 313, + 319, + 555, + 582 + ], + "type": "text", + "content": "Memory Consumption Analysis. In Table 3, we provide a detailed analysis of memory usage during the VAR inference process. The KV cache of the VAR model is the largest memory consumer, requiring 12 times the memory needed for the activations due to the significantly extended sequence length in the VAR paradigm. Our proposed CoDe effectively addresses this KV cache memory challenge. Specifically, we use the large VAR model only for predicting the first 91 tokens, which represent just " + }, + { + "bbox": [ + 313, + 319, + 555, + 582 + ], + "type": "inline_equation", + "content": "13\\%" + }, + { + "bbox": [ + 313, + 319, + 555, + 582 + ], + "type": "text", + "content": " of the total sequence. After this, the KV cache in the dafter is released, and the remaining computation is handled by a smaller model. This approach drastically reduces the KV cache memory requirements, compressing it to approximately " + }, + { + "bbox": [ + 313, + 319, + 555, + 582 + ], + "type": "inline_equation", + "content": "28\\%" + }, + { + "bbox": [ + 313, + 319, + 555, + 582 + ], + "type": "text", + "content": " of the original VAR-d30. Although the refiner model adds a small number of additional parameters, this overhead is minimal compared to the major optimization achieved in KV cache storage. As a result, the total memory usage of CoDe is significantly lower than the original model. Moreover, as batch size increases, the memory savings with CoDe become even more pronounced, underscoring its potential to support efficient high-resolution image generation within the VAR paradigm." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 596, + 388, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 596, + 388, + 609 + ], + "spans": [ + { + "bbox": [ + 313, + 596, + 388, + 609 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 617, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 554, + 713 + ], + "type": "text", + "content": "This work presents CoDe, a novel method designed for efficient decoding in visual auto-regressive modeling. CoDe effectively mitigates the significant memory overhead and computational redundancy typically associated with the prolonged sequences of next-scale predictions. Through extensive experimentation, our method demonstrates a superior efficiency-quality trade-off, establishing a new benchmark for efficient, high-quality image generation." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "23341" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "type": "text", + "content": "Acknowledgment" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 295, + 127 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 295, + 127 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 295, + 127 + ], + "type": "text", + "content": "This project is supported by the Ministry of Education, Singapore, under its Academic Research Fund Tier 2 (Award Number: MOE-T2EP20122-0006)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 137, + 114, + 149 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 137, + 114, + 149 + ], + "spans": [ + { + "bbox": [ + 56, + 137, + 114, + 149 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 157, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 61, + 157, + 296, + 200 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 157, + 296, + 200 + ], + "spans": [ + { + "bbox": [ + 61, + 157, + 296, + 200 + ], + "type": "text", + "content": "[1] Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, et al. Qwen technical report. arXiv preprint arXiv:2309.16609, 2023. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 202, + 296, + 245 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 202, + 296, + 245 + ], + "spans": [ + { + "bbox": [ + 62, + 202, + 296, + 245 + ], + "type": "text", + "content": "[2] Fan Bao, Chongxuan Li, Jun Zhu, and Bo Zhang. Analytic-dpm: an analytic estimate of the optimal reverse variance in diffusion probabilistic models. arXiv preprint arXiv:2201.06503, 2022. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 246, + 296, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 246, + 296, + 300 + ], + "spans": [ + { + "bbox": [ + 62, + 246, + 296, + 300 + ], + "type": "text", + "content": "[3] Huiwen Chang, Han Zhang, Lu Jiang, Ce Liu, and William T Freeman. Maskgit: Masked generative image transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11315-11325, 2022. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 301, + 295, + 345 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 301, + 295, + 345 + ], + "spans": [ + { + "bbox": [ + 62, + 301, + 295, + 345 + ], + "type": "text", + "content": "[4] Mark Chen, Alec Radford, Rewon Child, Jeffrey Wu, Heewoo Jun, David Luan, and Ilya Sutskever. Generative pretraining from pixels. In International conference on machine learning, pages 1691-1703. PMLR, 2020. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 346, + 295, + 378 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 346, + 295, + 378 + ], + "spans": [ + { + "bbox": [ + 62, + 346, + 295, + 378 + ], + "type": "text", + "content": "[5] Zigeng Chen, Gongfan Fang, Xinyin Ma, and Xinchao Wang. " + }, + { + "bbox": [ + 62, + 346, + 295, + 378 + ], + "type": "inline_equation", + "content": "0.1\\%" + }, + { + "bbox": [ + 62, + 346, + 295, + 378 + ], + "type": "text", + "content": " data makes segment anything slim. arXiv preprint arXiv:2312.05284, 2023. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 380, + 295, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 380, + 295, + 423 + ], + "spans": [ + { + "bbox": [ + 62, + 380, + 295, + 423 + ], + "type": "text", + "content": "[6] Zigeng Chen, Xinyin Ma, Gongfan Fang, Zhenxiong Tan, and Xinchao Wang. Asyncdiff: Parallelizing diffusion models by asynchronous denoising. arXiv preprint arXiv:2406.06911, 2024.3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 425, + 295, + 467 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 425, + 295, + 467 + ], + "spans": [ + { + "bbox": [ + 62, + 425, + 295, + 467 + ], + "type": "text", + "content": "[7] Tri Dao, Dan Fu, Stefano Ermon, Atri Rudra, and Christopher Ré. Flashattention: Fast and memory-efficient exact attention with io-awareness. Advances in Neural Information Processing Systems, 35:16344-16359, 2022. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 468, + 295, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 468, + 295, + 513 + ], + "spans": [ + { + "bbox": [ + 62, + 468, + 295, + 513 + ], + "type": "text", + "content": "[8] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 4, 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 514, + 295, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 514, + 295, + 567 + ], + "spans": [ + { + "bbox": [ + 62, + 514, + 295, + 567 + ], + "type": "text", + "content": "[9] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 569, + 295, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 569, + 295, + 612 + ], + "spans": [ + { + "bbox": [ + 57, + 569, + 295, + 612 + ], + "type": "text", + "content": "[10] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12873-12883, 2021. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 613, + 295, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 613, + 295, + 657 + ], + "spans": [ + { + "bbox": [ + 57, + 613, + 295, + 657 + ], + "type": "text", + "content": "[11] Gongfan Fang, Xinyin Ma, Mingli Song, Michael Bi Mi, and Xinchao Wang. Depgraph: Towards any structural pruning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16091-16101, 2023. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 658, + 295, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 658, + 295, + 690 + ], + "spans": [ + { + "bbox": [ + 57, + 658, + 295, + 690 + ], + "type": "text", + "content": "[12] Gongfan Fang, Xinyin Ma, and Xinchao Wang. Structural pruning for diffusion models. Advances in neural information processing systems, 36, 2024. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "type": "text", + "content": "[13] Jiarui Fang, Jinzhe Pan, Xibo Sun, Aoyu Li, and Jiannan Wang. xedit: an inference engine for diffusion trans" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 554, + 712 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 333, + 73, + 554, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 554, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 554, + 95 + ], + "type": "text", + "content": "formers (dits) with massive parallelism. arXiv preprint arXiv:2411.01738, 2024. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 95, + 554, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 95, + 554, + 127 + ], + "spans": [ + { + "bbox": [ + 316, + 95, + 554, + 127 + ], + "type": "text", + "content": "[14] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 129, + 554, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 129, + 554, + 183 + ], + "spans": [ + { + "bbox": [ + 317, + 129, + 554, + 183 + ], + "type": "text", + "content": "[15] Jiatao Gu, Yuyang Wang, Yizhe Zhang, Qihang Zhang, Dinghuai Zhang, Navdeep Jaitly, Josh Susskind, and Shuangfei Zhai. Dart: Denoising autoregressive transformer for scalable text-to-image generation. arXiv preprint arXiv:2410.08159, 2024. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 184, + 554, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 184, + 554, + 237 + ], + "spans": [ + { + "bbox": [ + 317, + 184, + 554, + 237 + ], + "type": "text", + "content": "[16] Daya Guo, Qihao Zhu, Dejian Yang, Zhenda Xie, Kai Dong, Wentao Zhang, Guanting Chen, Xiao Bi, Yu Wu, YK Li, et al. Deepseek-coder: When the large language model meets programming-the rise of code intelligence. arXiv preprint arXiv:2401.14196, 2024. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 239, + 554, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 239, + 554, + 293 + ], + "spans": [ + { + "bbox": [ + 316, + 239, + 554, + 293 + ], + "type": "text", + "content": "[17] Wanggui He, Siming Fu, Mushui Liu, Xierui Wang, Wenyi Xiao, Fangxun Shu, Yi Wang, Lei Zhang, Zhelun Yu, Haoyuan Li, et al. Mars: Mixture of auto-regressive models for fine-grained text-to-image synthesis. arXiv preprint arXiv:2407.07614, 2024. 1, 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 294, + 554, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 294, + 554, + 348 + ], + "spans": [ + { + "bbox": [ + 317, + 294, + 554, + 348 + ], + "type": "text", + "content": "[18] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 350, + 553, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 350, + 553, + 371 + ], + "spans": [ + { + "bbox": [ + 316, + 350, + 553, + 371 + ], + "type": "text", + "content": "[19] Geoffrey Hinton. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 3, 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 372, + 554, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 372, + 554, + 425 + ], + "spans": [ + { + "bbox": [ + 317, + 372, + 554, + 425 + ], + "type": "text", + "content": "[20] Doohyuk Jang, Sihwan Park, June Yong Yang, Yeonsung Jung, Jihun Yun, Souvik Kundu, Sung-Yub Kim, and Eunho Yang. *Lantern: Accelerating visual autoregressive models with relaxed speculative decoding.* arXiv preprint arXiv:2410.03355, 2024. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 426, + 553, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 426, + 553, + 448 + ], + "spans": [ + { + "bbox": [ + 316, + 426, + 553, + 448 + ], + "type": "text", + "content": "[21] Diederik P Kingma. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 449, + 554, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 449, + 554, + 491 + ], + "spans": [ + { + "bbox": [ + 317, + 449, + 554, + 491 + ], + "type": "text", + "content": "[22] Haopeng Li, Jinyue Yang, Kexin Wang, Xuerui Qiu, Yuhong Chou, Xin Li, and Guoqi Li. Scalable autoregressive image generation with mamba. arXiv preprint arXiv:2408.12245, 2024. 1, 2, 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 493, + 554, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 493, + 554, + 557 + ], + "spans": [ + { + "bbox": [ + 317, + 493, + 554, + 557 + ], + "type": "text", + "content": "[23] Muyang Li, Tianle Cai, Jiaxin Cao, Qinsheng Zhang, Han Cai, Junjie Bai, Yangqing Jia, Kai Li, and Song Han. Distributed parallel inference for high-resolution diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7183-7193, 2024. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 559, + 554, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 559, + 554, + 613 + ], + "spans": [ + { + "bbox": [ + 317, + 559, + 554, + 613 + ], + "type": "text", + "content": "[24] Muyang Li*, Yujun Lin*, Zhekai Zhang*, Tianle Cai, Xiuyu Li, Junxian Guo, Enze Xie, Chenlin Meng, Jun-Yan Zhu, and Song Han. Svdquant: Absorbing outliers by low-rank components for 4-bit diffusion models. arXiv preprint arXiv:2411.05007, 2024.3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 317, + 614, + 554, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 614, + 554, + 667 + ], + "spans": [ + { + "bbox": [ + 317, + 614, + 554, + 667 + ], + "type": "text", + "content": "[25] Senmao Li, Taihang Hu, Fahad Shahbaz Khan, Linxuan Li, Shiqi Yang, Yaxing Wang, Ming-Ming Cheng, and Jian Yang. Faster diffusion: Rethinking the role of unet encoder in diffusion models. arXiv preprint arXiv:2312.09608, 2023. 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 670, + 554, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 554, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 554, + 712 + ], + "type": "text", + "content": "[26] Tianhong Li, Yonglong Tian, He Li, Mingyang Deng, and Kaiming He. Autoregressive image generation without vector quantization. arXiv preprint arXiv:2406.11838, 2024. 1, 3, 6" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23342" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "text", + "content": "[27] Xiuyu Li, Yijiang Liu, Long Lian, Huanrui Yang, Zhen Dong, Daniel Kang, Shanghang Zhang, and Kurt Keutzer. Q-diffusion: Quantizing diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17535-17545, 2023. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 294, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 294, + 172 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 294, + 172 + ], + "type": "text", + "content": "[28] Xiang Li, Hao Chen, Kai Qiu, Jason Kuen, Jiuming Gu, Bhiksha Raj, and Zhe Lin. Imagefolder: Autoregressive image generation with folded tokens. arXiv preprint arXiv:2410.01756, 2024.3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 175, + 294, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 294, + 217 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 294, + 217 + ], + "type": "text", + "content": "[29] Xiang Li, Kai Qiu, Hao Chen, Jason Kuen, Zhe Lin, Rita Singh, and Bhiksha Raj. Controlvar: Exploring controllable visual autoregressive modeling. arXiv preprint arXiv:2406.09750, 2024. 1, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 220, + 294, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 220, + 294, + 274 + ], + "spans": [ + { + "bbox": [ + 56, + 220, + 294, + 274 + ], + "type": "text", + "content": "[30] Yanyu Li, Huan Wang, Qing Jin, Ju Hu, Pavlo Chemerys, Yun Fu, Yanzhi Wang, Sergey Tulyakov, and Jian Ren. Snapfusion: Text-to-image diffusion model on mobile devices within two seconds. Advances in Neural Information Processing Systems, 36, 2024. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 276, + 294, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 276, + 294, + 329 + ], + "spans": [ + { + "bbox": [ + 56, + 276, + 294, + 329 + ], + "type": "text", + "content": "[31] Zongming Li, Tianheng Cheng, Shoufa Chen, Peize Sun, Haocheng Shen, Longjin Ran, Xiaoxin Chen, Wenyu Liu, and Xinggang Wang. Controlar: Controllable image generation with autoregressive models. arXiv preprint arXiv:2410.02705, 2024. 1, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 331, + 294, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 331, + 294, + 385 + ], + "spans": [ + { + "bbox": [ + 56, + 331, + 294, + 385 + ], + "type": "text", + "content": "[32] Dongyang Liu, Shitian Zhao, Le Zhuo, Weifeng Lin, Yu Qiao, Hongsheng Li, and Peng Gao. Lumina-mgpt: Illuminate flexible photorealistic text-to-image generation with multimodal generative pretraining. arXiv preprint arXiv:2408.02657, 2024. 1, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 388, + 294, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 388, + 294, + 420 + ], + "spans": [ + { + "bbox": [ + 56, + 388, + 294, + 420 + ], + "type": "text", + "content": "[33] Luping Liu, Yi Ren, Zhijie Lin, and Zhou Zhao. Pseudo numerical methods for diffusion models on manifolds. arXiv preprint arXiv:2202.09778, 2022. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 422, + 294, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 422, + 294, + 475 + ], + "spans": [ + { + "bbox": [ + 56, + 422, + 294, + 475 + ], + "type": "text", + "content": "[34] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps. Advances in Neural Information Processing Systems, 35:5775-5787, 2022. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 477, + 294, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 477, + 294, + 520 + ], + "spans": [ + { + "bbox": [ + 56, + 477, + 294, + 520 + ], + "type": "text", + "content": "[35] Simian Luo, Yiqin Tan, Longbo Huang, Jian Li, and Hang Zhao. Latent consistency models: Synthesizing high-resolution images with few-step inference. arXiv preprint arXiv:2310.04378, 2023. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 522, + 294, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 522, + 294, + 566 + ], + "spans": [ + { + "bbox": [ + 56, + 522, + 294, + 566 + ], + "type": "text", + "content": "[36] Zhuoyan Luo, Fengyuan Shi, Yixiao Ge, Yujiu Yang, Limin Wang, and Ying Shan. Open-magvit2: An open-source project toward democratizing auto-regressive visual generation. arXiv preprint arXiv:2409.04410, 2024. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 567, + 294, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 567, + 294, + 600 + ], + "spans": [ + { + "bbox": [ + 56, + 567, + 294, + 600 + ], + "type": "text", + "content": "[37] Zhaoyang Lyu, Xudong Xu, Ceyuan Yang, Dahua Lin, and Bo Dai. Accelerating diffusion models via early stop of the diffusion process. arXiv preprint arXiv:2205.12524, 2022. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 601, + 294, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 601, + 294, + 633 + ], + "spans": [ + { + "bbox": [ + 56, + 601, + 294, + 633 + ], + "type": "text", + "content": "[38] Xinyin Ma, Gongfan Fang, and Xinchao Wang. Deepcache: Accelerating diffusion models for free. arXiv preprint arXiv:2312.00858, 2023. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 635, + 294, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 294, + 679 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 294, + 679 + ], + "type": "text", + "content": "[39] Xiaoxiao Ma, Mohan Zhou, Tao Liang, Yalong Bai, Tiejun Zhao, Huaian Chen, and Yi Jin. Star: Scale-wise text-to-image generation via auto-regressive representations. arXiv preprint arXiv:2406.10797, 2024. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 681, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 681, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 681, + 294, + 713 + ], + "type": "text", + "content": "[40] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An im" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 105 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 105 + ], + "type": "text", + "content": "perative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019. 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 107, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 150 + ], + "type": "text", + "content": "[41] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4195-4205, 2023. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 152, + 553, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 553, + 195 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 553, + 195 + ], + "type": "text", + "content": "[42] Kai Qiu, Xiang Li, Hao Chen, Jie Sun, Jinglu Wang, Zhe Lin, Marios Savvides, and Bhiksha Raj. Efficient autoregressive audio modeling via next-scale prediction. arXiv preprint arXiv:2408.09027, 2024. 1, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 198, + 553, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 198, + 553, + 229 + ], + "spans": [ + { + "bbox": [ + 316, + 198, + 553, + 229 + ], + "type": "text", + "content": "[43] Tim Salimans and Jonathan Ho. Progressive distillation for fast sampling of diffusion models. arXiv preprint arXiv:2202.00512, 2022. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 232, + 553, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 232, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 316, + 232, + 553, + 262 + ], + "type": "text", + "content": "[44] Axel Sauer, Dominik Lorenz, Andreas Blattmann, and Robin Rombach. Adversarial diffusion distillation. arXiv preprint arXiv:2311.17042, 2023. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 265, + 553, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 265, + 553, + 308 + ], + "spans": [ + { + "bbox": [ + 316, + 265, + 553, + 308 + ], + "type": "text", + "content": "[45] Yuzhang Shang, Zhihang Yuan, Bin Xie, Bingzhe Wu, and Yan Yan. Post-training quantization on diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1972-1981, 2023. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 310, + 553, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 310, + 553, + 342 + ], + "spans": [ + { + "bbox": [ + 316, + 310, + 553, + 342 + ], + "type": "text", + "content": "[46] Junhyuk So, Jungwon Lee, and Eunhyeok Park. Frdiff: Feature reuse for exquisite zero-shot acceleration of diffusion models. arXiv preprint arXiv:2312.03517, 2023. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 344, + 553, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 344, + 553, + 376 + ], + "spans": [ + { + "bbox": [ + 316, + 344, + 553, + 376 + ], + "type": "text", + "content": "[47] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 378, + 553, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 378, + 553, + 422 + ], + "spans": [ + { + "bbox": [ + 316, + 378, + 553, + 422 + ], + "type": "text", + "content": "[48] Peize Sun, Yi Jiang, Shoufa Chen, Shilong Zhang, Bingyue Peng, Ping Luo, and Zehuan Yuan. Autoregressive model beats diffusion: Llama for scalable image generation. arXiv preprint arXiv:2406.06525, 2024. 1, 2, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 423, + 553, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 423, + 553, + 477 + ], + "spans": [ + { + "bbox": [ + 316, + 423, + 553, + 477 + ], + "type": "text", + "content": "[49] Haotian Tang, Yecheng Wu, Shang Yang, Enze Xie, Junsong Chen, Junyu Chen, Zhuoyang Zhang, Han Cai, Yao Lu, and Song Han. Hart: Efficient visual generation with hybrid autoregressive transformer. arXiv preprint arXiv:2410.10812, 2024. 1, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 479, + 553, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 479, + 553, + 533 + ], + "spans": [ + { + "bbox": [ + 316, + 479, + 553, + 533 + ], + "type": "text", + "content": "[50] Yao Teng, Han Shi, Xian Liu, Xuefei Ning, Guohao Dai, Yu Wang, Zhenguo Li, and Xihui Liu. Accelerating autoregressive text-to-image generation with training-free speculative jacobi decoding. arXiv preprint arXiv:2410.01699, 2024.3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 535, + 553, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 535, + 553, + 578 + ], + "spans": [ + { + "bbox": [ + 316, + 535, + 553, + 578 + ], + "type": "text", + "content": "[51] Keyu Tian, Yi Jiang, Zehuan Yuan, Bingyue Peng, and Liwei Wang. Visual autoregressive modeling: Scalable image generation via next-scale prediction. arXiv preprint arXiv:2404.02905, 2024. 1, 3, 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 580, + 553, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 580, + 553, + 634 + ], + "spans": [ + { + "bbox": [ + 316, + 580, + 553, + 634 + ], + "type": "text", + "content": "[52] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 635, + 553, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 553, + 679 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 553, + 679 + ], + "type": "text", + "content": "[53] Aaron Van den Oord, Nal Kalchbrenner, Lasse Espeholt, Oriol Vinyals, Alex Graves, et al. Conditional image generation with pixelCNN decoders. Advances in neural information processing systems, 29, 2016. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "text", + "content": "[54] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017. 2" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "23343" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "text", + "content": "[55] Felix Wimbauer, Bichen Wu, Edgar Schoenfeld, Xiaoliang Dai, Ji Hou, Zijian He, Artsiom Sanakoyeu, Peizhao Zhang, Sam Tsai, Jonas Kohler, et al. Cache me if you can: Accelerating diffusion models through block caching. arXiv preprint arXiv:2312.03209, 2023. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 294, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 294, + 194 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 294, + 194 + ], + "type": "text", + "content": "[56] Jinheng Xie, Weijia Mao, Zechen Bai, David Junhao Zhang, Weihao Wang, Kevin Qinghong Lin, Yuchao Gu, Zhijie Chen, Zhenheng Yang, and Mike Zheng Shou. Show-o: One single transformer to unify multimodal understanding and generation. arXiv preprint arXiv:2408.12528, 2024. 1, 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 197, + 294, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 197, + 294, + 228 + ], + "spans": [ + { + "bbox": [ + 56, + 197, + 294, + 228 + ], + "type": "text", + "content": "[57] Xingyi Yang and Xinchao Wang. Hash3d: Training-free acceleration for 3d generation. arXiv preprint arXiv:2404.06091, 2024. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 230, + 295, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 230, + 295, + 274 + ], + "spans": [ + { + "bbox": [ + 56, + 230, + 295, + 274 + ], + "type": "text", + "content": "[58] Xingyi Yang, Daquan Zhou, Jiashi Feng, and Xinchao Wang. Diffusion probabilistic model made slim. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22552-22562, 2023. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 275, + 294, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 275, + 294, + 319 + ], + "spans": [ + { + "bbox": [ + 56, + 275, + 294, + 319 + ], + "type": "text", + "content": "[59] Ziyu Yao, Jialin Li, Yifeng Zhou, Yong Liu, Xi Jiang, Chengjie Wang, Feng Zheng, Yuexian Zou, and Lei Li. Car: Controllable autoregressive modeling for visual generation. arXiv preprint arXiv:2410.04671, 2024. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 320, + 294, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 320, + 294, + 363 + ], + "spans": [ + { + "bbox": [ + 56, + 320, + 294, + 363 + ], + "type": "text", + "content": "[60] Tianwei Yin, Michael Gharbi, Richard Zhang, Eli Shechtman, Fredo Durand, William T Freeman, and Taesung Park. One-step diffusion with distribution matching distillation. arXiv preprint arXiv:2311.18828, 2023. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 365, + 294, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 365, + 294, + 408 + ], + "spans": [ + { + "bbox": [ + 56, + 365, + 294, + 408 + ], + "type": "text", + "content": "[61] Zongsheng Yue, Jianyi Wang, and Chen Change Loy. Reshift: Efficient diffusion model for image superresolution by residual shifting. Advances in Neural Information Processing Systems, 36, 2024. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 410, + 294, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 410, + 294, + 453 + ], + "spans": [ + { + "bbox": [ + 56, + 410, + 294, + 453 + ], + "type": "text", + "content": "[62] Dingkun Zhang, Sijia Li, Chen Chen, Qingsong Xie, and Haonan Lu. Laptop-diff: Layer pruning and normalized distillation for compressing diffusion models. arXiv preprint arXiv:2404.11098, 2024. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 455, + 294, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 455, + 294, + 497 + ], + "spans": [ + { + "bbox": [ + 56, + 455, + 294, + 497 + ], + "type": "text", + "content": "[63] Jinzhi Zhang, Feng Xiong, and Mu Xu. G3pt: Unleash the power of autoregressive modeling in 3d generation via cross-scale querying transformer. arXiv preprint arXiv:2409.06322, 2024. 1, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 500, + 294, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 500, + 294, + 532 + ], + "spans": [ + { + "bbox": [ + 56, + 500, + 294, + 532 + ], + "type": "text", + "content": "[64] Qinsheng Zhang and Yongxin Chen. Fast sampling of diffusion models with exponential integrator. arXiv preprint arXiv:2204.13902, 2022. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 534, + 294, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 534, + 294, + 577 + ], + "spans": [ + { + "bbox": [ + 56, + 534, + 294, + 577 + ], + "type": "text", + "content": "[65] Qian Zhang, Xiangzi Dai, Ninghua Yang, Xiang An, Ziyong Feng, and Xingyu Ren. Var-clip: Text-to-image generator with visual auto-regressive modeling. arXiv preprint arXiv:2408.01181, 2024. 1, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 579, + 294, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 579, + 294, + 622 + ], + "spans": [ + { + "bbox": [ + 56, + 579, + 294, + 622 + ], + "type": "text", + "content": "[66] Wentian Zhang, Haozhe Liu, Jinheng Xie, Francesco Faccio, Mike Zheng Shou, and Jürgen Schmidhuber. Cross-attention makes inference cumbersome in text-to-image diffusion models. arXiv preprint arXiv:2404.02747, 2024. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 624, + 294, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 294, + 656 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 294, + 656 + ], + "type": "text", + "content": "[67] Xuanlei Zhao, Xiaolong Jin, Kai Wang, and Yang You. Real-time video generation with pyramid attention broadcast. arXiv preprint arXiv:2408.12588, 2024. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 658, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 294, + 689 + ], + "type": "text", + "content": "[68] Yang Zhao, Yanwu Xu, Zhisheng Xiao, and Tingbo Hou. Mobilediffusion: Subsecond text-to-image generation on mobile devices. arXiv preprint arXiv:2311.16567, 2023. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 692, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 692, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 692, + 294, + 713 + ], + "type": "text", + "content": "[69] Kaiwen Zheng, Cheng Lu, Jianfei Chen, and Jun Zhu. Dpm solver-v3: Improved diffusion ode solver with empirical" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 151 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "text", + "content": "model statistics. Advances in Neural Information Processing Systems, 36, 2024. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 96, + 553, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 151 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 151 + ], + "type": "text", + "content": "[70] Chunting Zhou, Lili Yu, Arun Babu, Kushal Tirumala, Michihiro Yasunaga, Leonid Shamis, Jacob Kahn, Xuezhe Ma, Luke Zettlemoyer, and Omer Levy. Transfusion: Predict the next token and diffuse images with one multi-modal model. arXiv preprint arXiv:2408.11039, 2024. 1, 2" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "23344" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Collaborative Tree Search for Enhancing Embodied Multi-Agent Collaboration/509a3bd5-5db7-45e0-bb5b-ba942b10cc92_content_list.json b/2025/Collaborative Tree Search for Enhancing Embodied Multi-Agent Collaboration/509a3bd5-5db7-45e0-bb5b-ba942b10cc92_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..fc20dd986200687b509fa2ebf7f54bc16bd31c28 --- /dev/null +++ b/2025/Collaborative Tree Search for Enhancing Embodied Multi-Agent Collaboration/509a3bd5-5db7-45e0-bb5b-ba942b10cc92_content_list.json @@ -0,0 +1,1691 @@ +[ + { + "type": "text", + "text": "Collaborative Tree Search for Enhancing Embodied Multi-Agent Collaboration", + "text_level": 1, + "bbox": [ + 99, + 128, + 898, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Lizheng $\\mathsf{Zu}^{1,2}$ , Lin Lin $^{1*}$ , Song Fu $^{1*}$ , Na Zhao $^{3}$ , Pan Zhou $^{4}$ $^{1}$ Harbin Institute of Technology, $^{2}$ Nanyang Technological University, \n $^{3}$ Singapore University of Technology and Design, $^{4}$ Singapore Management University \nzuli_edu@outlook.com, waiwaiyl@hit.edu.cn, fusong@hit.edu.cn \nna_zhao@sutd.edu.sg, panzhou@smu.edu.sg", + "bbox": [ + 158, + 179, + 839, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 303, + 326, + 320 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Embodied agents based on large language models (LLMs) face significant challenges in collaborative tasks, requiring effective communication and reasonable division of labor to ensure efficient and correct task completion. Previous approaches with simple communication patterns carry erroneous or incoherent agent actions, which can lead to additional risks. To address these problems, we propose Cooperative Tree Search (CoTS), a framework designed to significantly improve collaborative planning and task execution efficiency among embodied agents. CoTS guides multi-agents to discuss long-term strategic plans within a modified Monte Carlo tree, searching along LLM-driven reward functions to provide a more thoughtful and promising approach to cooperation. Another key feature of our method is the introduction of a plan evaluation module, which not only prevents agent action confusion caused by frequent plan updates but also ensures plan updates when the current plan becomes unsuitable. Experimental results show that the proposed method performs excellently in planning, communication, and collaboration on embodied environments (CWAH and TDW-MAT), efficiently completing long-term, complex tasks and significantly outperforming existing methods.", + "bbox": [ + 88, + 335, + 485, + 684 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 714, + 220, + 729 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Embodied agents are increasingly capable of perceiving and interacting with their environments like humans, autonomously managing planning, decision-making, actions, and execution tasks [23, 36]. Recent advances have generated substantial interest in these agents, especially for applications requiring complex task execution in dynamic environments, such as robotic operations in warehouses or domestic settings. In particular, collaborative task completion among multiple agents enhances both efficiency and adapt", + "bbox": [ + 89, + 739, + 483, + 877 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ability, especially in dynamic scenarios [5, 10]. However, collaboration among embodied agents introduces significant challenges: agents must not only perceive and understand their environment but also communicate, share information, divide tasks, and coordinate actions responsively. For example, consider autonomous robots in a home responding to a request: \"Bring me the iPad and apple, and put the milk in the refrigerator.\" Achieving this requires multi-level coordination, including optimal search strategies, task prioritization, and efficient movement planning.", + "bbox": [ + 511, + 305, + 906, + 457 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Large language models (LLMs) have recently provided embodied agents with advanced natural language understanding, dialogue, and reasoning abilities [1, 3, 38, 40]. These capabilities allow LLMs to decompose complex and long-term tasks into a sequence of manageable sub-goals, making LLM-driven agents a promising alternative to traditional reinforcement learning models [14, 34] which are difficult to train and often generalize poorly [7, 21, 39].", + "bbox": [ + 511, + 458, + 908, + 580 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "However, enabling embodied agents to work together in decentralized environments remains a significant and underexplored challenge, as it requires long-term planning and coherent decision-making to coordinate actions efficiently. Early attempts at multi-agent collaboration, such as CoELA [43] and RoCo [24], demonstrate progress but also reveal limitations. For instance, as shown in Fig. 1 (b), CoELA facilitates collaboration by sharing updates through natural language when sub-tasks are completed, yet each agent's decision-making process remains independent, resulting in suboptimal coordination. In contrast, as illustrated in Fig. 1 (c), RoCo develops multi-agent work plans through agent discussions and environmental interactions. However, RoCo relies on a single reasoning path, which is susceptible to the randomness and unpredictability of LLM outputs. This can lead to inefficient or incorrect planning, particularly in applications requiring precision, where such errors can have serious consequences.", + "bbox": [ + 511, + 580, + 908, + 852 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Contributions. To address these limitations, we propose Cooperative Tree Search (CoTS), a framework designed to significantly improve collaborative planning and task ex", + "bbox": [ + 511, + 854, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author.", + "bbox": [ + 109, + 887, + 236, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "29513", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Task: transport 2 oranges, 3 apples, 1 banana, 3 breads, 1 burger to the bed.", + "bbox": [ + 99, + 80, + 509, + 93 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Background: Agents randomly appear in a home environment with a total of six rooms and need to discover objects through exploration. Each agent can carry up to two objects. Finding and using containers can improve efficiency.", + "bbox": [ + 98, + 93, + 872, + 119 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/51f8faae0a245a38a0bcbb01d92263bdf6aa223be74feae5bfb9102f14896f94.jpg", + "image_caption": [ + "(a) The Embodied Environment", + "(b) CoELA (Communication but independent decision-making.)" + ], + "image_footnote": [], + "bbox": [ + 93, + 125, + 369, + 277 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/0b6e08abed85d204b9d2583ea6cc8aba813591ec8e0cc70b32599d38f8cb576f.jpg", + "image_caption": [ + "(c) RoCo (Discussion, Planning and Execution)" + ], + "image_footnote": [], + "bbox": [ + 372, + 125, + 625, + 277 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/822a2c81f6949651bd472a1ffd8758b0757df6f8bb5e0a26c80dec6d849e2ba0.jpg", + "image_caption": [ + "(d) Ours (Plan Searching, Evaluation, and Plan-Guided Execution)", + "Figure 1. The task execution processes of CoELA, RoCo, and our CoTS. (1) In CoELA, agents update each other on sub-task completion and individually determine their next actions without long-term strategic planning. (2) RoCo uses multi-round dialogues and summarizes the plan at the end, but its single-dialogue chain can lead to suboptimal plans and even errors due to the randomness and unpredictability of LLMs. (3) Within a modified Monte Carlo Tree Search framework, CoTS allows agents to evaluate multiple planning paths, enabling them to collaboratively deliberate and identify the most effective course of action based on task progress and environmental context." + ], + "image_footnote": [], + "bbox": [ + 627, + 125, + 893, + 282 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "cution efficiency among embodied agents. Inspired by human decision-making, CoTS allows agents to search multiple planning paths by combining exploration and backtracking, enabling them to collaboratively deliberate and identify the most effective course of action based on task progress and environmental context.", + "bbox": [ + 88, + 401, + 480, + 491 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "As illustrated in Fig. 1 (d), CoTS advances agent collaboration in embodied environments through two core innovations. First, we introduce a collaborative planning module based on a modified Monte Carlo Tree Search (MCTS) framework, which enables agents to explore multiple planning paths and select the most coherent and efficient long-term plan. Unlike traditional MCTS which relies on simulation-based rewards, CoTS incorporates a non-simulated LLM-based reward function suited for high-cost physical movement interactions in embodied environments. In this way, CoTS transforms the traditional linear sequence of action steps into a search process, exploring potential paths within a shared reasoning and action space to develop robust collaborative strategies. Second, CoTS features a plan evaluation module, that actively monitors the relevance and feasibility of the selected plan during execution. This mechanism allows agents to adapt plans as needed, reducing unnecessary disruptions from frequent plan updates while still accommodating environmental changes.", + "bbox": [ + 89, + 492, + 482, + 779 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our experiments on two long-term multi-objective planning tasks — TDW-MAT [43] and C-WAH [8] — demonstrate that CoTS significantly outperforms current state-of-the-art methods, e.g., CoELA [43] and RoCo [24], in terms of task completion rate and efficiency. These results highlight the effectiveness and superiority of CoTS, underscoring its potential for high-stakes applications that demand reliable multi-agent collaboration.", + "bbox": [ + 89, + 780, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 398, + 653, + 415 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "LLM-based Agents. Recent advances in LLMs have unlocked near-human reasoning and planning capabilities [1, 27, 37]. Researchers have leveraged LLMs as the \"brain\" or controller of agents, expanding their perceptual and action capacities through multimodal perception and tool use [6, 25, 32, 38, 40, 45]. Multi-agent systems [16, 35] show particular promise, enabling agents to collaborate on planning, discussion, and decision-making—similar to how humans solve complex problems in teams [26, 33]. Works such as MetaGPT [13], CAMEL [19], and ChatDev [31] decompose complex tasks into simpler, manageable subtasks to reduce hallucinations and improve task-solving capabilities. However, embodied multi-agent systems [20, 44] require advanced communication and face significant challenges in navigating a vast action search space in complex, dynamic environments, especially for long-term tasks.", + "bbox": [ + 511, + 429, + 906, + 671 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Embodied Multi-Agent Cooperation. Research on embodied multi-agent systems is still emerging. Habitat [30] explores agents' social awareness, enabling adaptive assistance strategies in collaboration. Co-NavGPT [41] uses LLMs for multi-robot collaborative goal navigation, encoding environmental data to improve scene understanding. Organized-LLM-Agents [11] reduces redundancy in multiagent collaboration through structured prompts, enhancing coordination. CoELA [43] enables collaborative agents to communicate and plan through natural language and chain-of-thought processes, yet lacks strategic long-term planning. RoCo [24] and CaPo [22] advance multi-agent collaboration with high-level communication and path planning, but their single reasoning path often leads to suboptimal plans and even errors. Although some existing work", + "bbox": [ + 511, + 674, + 908, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "29514", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/232b84dc39248ff7fdd9552a6c86729adc9372537500dbd2137b559f4f78a0bb.jpg", + "image_caption": [ + "Figure 2. The framework of CoTS. Upon receiving a new task in the embodied environment, each embodied agent initially employs the Perception Module to gather raw observational data, which is then stored in its Memory Module. In the collaborative planning module, agents guide the exploration of collaborative plans toward more promising directions through the collaborative tree search. The explored plans are retained within the memory module as a special 'extended memory' and are periodically updated based on assessments from the planning evaluation module. Finally, agents perform actions aligned with the collaborative plan and tasks." + ], + "image_footnote": [], + "bbox": [ + 138, + 88, + 846, + 324 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "on LLMs involves multi-branch reasoning [2, 17, 40, 42], they are inadequate for complex embodied environments and agent collaboration. Our goal is to develop embodied agents capable of searching long-term strategic planning.", + "bbox": [ + 89, + 407, + 483, + 468 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Preliminaries", + "text_level": 1, + "bbox": [ + 89, + 479, + 230, + 496 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Task Assumption. As introduced by previous work [9, 24], multi-agent collaborative tasks can be viewed as a Markov game, which is defined by a tuple $(\\mathcal{N},\\mathcal{S},\\mathcal{A},P,r,\\gamma)$ , in which $\\mathcal{N}$ is the set of agents (e.g. Alice and Bob), $\\mathcal{S}$ is the state space, $\\mathcal{A} = \\prod_{i=1}^{n}\\mathcal{A}^i$ is the combination of all action spaces, $P: \\mathcal{S} \\times \\mathcal{A} \\times \\mathcal{S} \\to [0,1]$ is the transition probability function, $r: \\mathcal{S} \\times \\mathcal{A} \\to \\mathbb{R}$ is the reward score, and $\\gamma \\in [0,1)$ is the discount factor. In the task, each agent at time step $t \\in \\mathbb{N}$ obtains the information by perception module from the embodied environment $s_t \\in \\mathcal{S}$ , chooses and executes an action $a_t^i \\in \\mathcal{A}^i$ from its available policy $\\pi^i(\\cdot|s_t)$ . This will along with the actions of other agents, constitutes a joint action $\\mathbf{a}_t = (a_t^1,a_t^2,\\dots,a_t^n) \\in \\mathcal{A}$ derived from the joint policy $\\pi(\\cdot|s_t) = \\prod_{i=1}^{n}\\pi^i(\\cdot|s_t)$ . Then agents perceive a new state $s_{t+1}$ with probability $P(s_{t+1}|s_t,\\mathbf{a}_t)$ . We seek to determine a collaborative plan that maximizes the $J(\\pi) \\triangleq \\mathbb{E}_{s_0,\\infty} \\sim P,a_0,\\infty \\sim \\pi[\\sum_{i=0}^{\\infty}\\gamma^i r_i]$ . For this paper, the goal is to coordinate the distribution of sub-tasks among agents so as to complete tasks more efficiently.", + "bbox": [ + 91, + 507, + 482, + 792 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "CoELA [43]. It contains five modules: 1) Perception, 2) Memory, 3) Communication, 4) Planning and 5) Execution. The Perception module leverages a trained Mask-RCNN [12] to segment RGB images and detect objects and obstacles, creating 3D point clouds from RGB-D data and extracting essential high-level information like key object states. The Memory module manages short-term and long-", + "bbox": [ + 89, + 795, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "term memory, with short-term memory capturing immediate details such as semantic maps and agent states, and long-term memory storing persistent information like task descriptions. In the Communication module, CoELA retrieves relevant information from memory, transforming it into template-based textual descriptions for agents to aid in action selection. However, each agent plans independently, with no unified action coordination. In the Planning and Execution modules, CoELA retrieves agent-related information from memory, generates potential actions for the LLM to consider, and applies chain-of-thought [38] reasoning for more in-depth decision-making. Each agent then executes its chosen action accordingly.", + "bbox": [ + 511, + 407, + 906, + 604 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Collaborative Tree Search for Multi-Agents", + "text_level": 1, + "bbox": [ + 511, + 622, + 901, + 640 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1. Overview of CoTS", + "text_level": 1, + "bbox": [ + 511, + 648, + 692, + 664 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Now we introduce an overview of our Collaborative Tree Search (CoTS), which is designed for multi-agent cooperation in embodied environments. As shown in Fig. 2, CoTS contains four key modules: 1) Perception Module, 2) Memory Module, 3) Collaborative Planning Module, and 4) Plan Parsing and Execution Module. This work primarily enhances CoELA in its planning by designing the Collaborative Planning Module. The other three modules are largely borrowed from CoELA and introduced in Sec. 3. Here we briefly introduce our enhancements to CoEAL below.", + "bbox": [ + 511, + 672, + 906, + 823 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Perception Module. Unlike CoELA, we introduce distance as a key perceptual factor, leveraging fixed spatial layouts in real-world settings to enhance agents' room-based understanding and improve planning efficiency by using LLM to access the moving distance of agents.", + "bbox": [ + 511, + 825, + 908, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "29515", + "bbox": [ + 478, + 944, + 517, + 957 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Task: transport 2 oranges, 3 apples, 1 banana, 3 breads, 1 burger to the bed", + "bbox": [ + 102, + 93, + 493, + 104 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The Process in Each Node: Alice makes plans, Bob provides suggestions and rewards.", + "bbox": [ + 102, + 109, + 557, + 121 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Plan from Alice: Step 1: Alice grasps apple. Bob explore his current room.", + "bbox": [ + 102, + 127, + 491, + 138 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Step 2: Alice explore the Living room, if they find any items, transport them to the bed.", + "bbox": [ + 102, + 140, + 555, + 150 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Message to Bob: Hi Bob, I am in Livingroom where I found an apple. Do you have suggestions.", + "bbox": [ + 102, + 152, + 599, + 162 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Message to Alice: Hi Alice, I am in Livingroom and found bread. As you said, I am just about to explore my current room. Reward: 0.3", + "bbox": [ + 102, + 167, + 534, + 190 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Plan from Alice: Step 1: Alice moves to Bedroom. Bob grasps bread and apple.", + "bbox": [ + 102, + 199, + 514, + 210 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Step 2: Alice explores Bedroom. Bob explores Living room.", + "bbox": [ + 102, + 210, + 413, + 222 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Step 3: If any target item are found, transport them to the bed.", + "bbox": [ + 102, + 223, + 424, + 234 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Message to Bob: Hi Bob, Please grasp bread and apple and I will now shift to the Bedroom because we should not explore the same room.", + "bbox": [ + 102, + 234, + 575, + 258 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Message to Alice: Hi Alice, but I think you could head to kitchen, it is the nearest. Reward: 0.6", + "bbox": [ + 102, + 258, + 598, + 268 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Plan from Alice: Step 1: Alice grasps the apple. Bob grasps the bread.", + "bbox": [ + 102, + 277, + 465, + 289 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Step 2: Alice moves to Kitchen. Bob explores Living room.", + "bbox": [ + 102, + 290, + 406, + 301 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Step 3: Alice explore the Kitchen, if they find any items, transport them to the bed.", + "bbox": [ + 102, + 301, + 532, + 313 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Message to Bob: Hi Bob, I agree with you. I will grasp the apple before I mo", + "bbox": [ + 102, + 315, + 501, + 325 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Kitchen while you continue in Livingroom. Please keep me updated on your progress", + "bbox": [ + 102, + 327, + 544, + 339 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Message to Alice: Hi, Alice, your plan looks good. Let's get moving. Reward: 1", + "bbox": [ + 102, + 340, + 517, + 353 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/49f12660ec44c9334be6776b9704c999105e257b7f8ddb2a24ea2df01f7b0f32.jpg", + "image_caption": [ + "Figure 3. Overview of the tree search in CoTS. Each node in the tree contains content generated by Alice and Bob. Alice generates the collaborative plans and messages to Bob, while Bob generates messages to Alice and rewards for the plans. These rewards are used for backpropagation to guide the tree search toward more promising collaborative plans. The figure also shows examples of three nodes, illustrating the stepwise optimization of collaborative planning and the dialogue process between agents." + ], + "image_footnote": [], + "bbox": [ + 612, + 103, + 901, + 339 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Memory Module. Compared with CoELA, CoTS introduces \"extended memory\" (Fig. 2) to manage collaboration plans—information that needs to be retained for a moderate period and updated periodically. This is because collaboration plans must balance retention for a certain period and timely updates when they become unsuitable for the current embodied environment. This extended memory mimics the medium-term storage of the human memory system, bridging short-term recall and long-term consolidation.", + "bbox": [ + 88, + 431, + 482, + 568 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Collaborative Planning Module is the core of CoTS, where MCTS is modified and employed to develop long-term strategic collaboration plans. Inspired by human problem-solving through iterative discussion and cooperation, it avoids error accumulation caused by a single path approach. This module represents the first integration of embodied agent collaboration with MCTS, enhancing agents' capability to tackle complex, long-term tasks through tree-based search and backtracking. Details of this module are provided in Sec. 4.2.", + "bbox": [ + 88, + 569, + 482, + 718 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Plan Parsing and Execution Module translates collaborative plans into executable actions, enabling agents to complete tasks in a coordinated manner. See details in Sec. 4.3.", + "bbox": [ + 89, + 719, + 482, + 763 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Below we will delve deeper into two core modules, Collaborative Planning Module and Plan Parsing and Execution Module. Additionally, we will highlight CoTS's integration of spatial awareness and extended memory to enhance overall functionality and efficiency.", + "bbox": [ + 89, + 765, + 482, + 840 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Collaborative Planning Module", + "text_level": 1, + "bbox": [ + 89, + 848, + 367, + 864 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our Collaborative Planning Module, central to our CoTS framework, is designed to search for long-term, efficient", + "bbox": [ + 89, + 869, + 482, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "and coherent collaborative plans for multi-agent tasks in dynamic environments. It comprises two main sub-modules: a collaborative tree search module for generating collaborative plans (Sec. 4.2.1) and a plan evaluation module to analyze when to update the plan (Sec. 4.2.2).", + "bbox": [ + 511, + 431, + 906, + 508 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2.1. Monte Carlo Tree Search for Collaborative Plan", + "text_level": 1, + "bbox": [ + 511, + 518, + 893, + 532 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We adapt MCTS to convert multi-agent collaborative planning into a tree search, identifying an optimal pathway from multiple reasoning paths for cohesive strategies. As shown in Fig. 3, each node in this tree stores a proposed plan, agent feedback, and a performance reward. In this setup, one agent (like Alice) functions as the \"planner\" to propose plans, while other agents (like Bob) act as \"critics\". Alice drafts several initial collaborative plans or extends existing plan based on task goals and current task progress. Then Bob evaluates them, considering factors like task allocation and movement efficiency, to generate a reward for the current plan. Once feedback is provided, each plan's reward is propagated back through the collaboration tree, strengthening successful strategies and adjusting future planning pathways. Accordingly, they will select the most efficient collaborative plan for further reflections and refinements. As shown in Fig. 3, this process, including plan extension (drafting), evaluation, backpropagation, and selection, will repeat until agents explore all potential collaborative plans and select the most promising one. In the following, for these four steps, we will elaborate on each of them.", + "bbox": [ + 511, + 537, + 906, + 853 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Selection Similar to traditional MCTS [4, 15], CoTS starts from the root node and traverses down the tree by selecting nodes with the highest UCT values. This process is repeated", + "bbox": [ + 511, + 854, + 906, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "29516", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "until a leaf node is reached, and is illustrated in Fig. 3 where we mark the selected nodes with gold cups. For UCT of each node, it is computed as follows:", + "bbox": [ + 89, + 90, + 485, + 137 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {U C T} (s _ {i}) = V (s _ {i}) + w \\sqrt {\\ln N (p _ {i}) / N (s _ {i})}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 143, + 145, + 482, + 165 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Where $N(s_{i})$ is visiting number of node $s_i$ , $V(s_{i})$ is the reward of the subtree rooted at $s_i$ , $w$ is the exploration coefficient, and $p_i$ is the parent node of $s_i$ . After an episode, backpropagation updates each node's $V(s_{i})$ along the path: $V(s_{i}) = \\frac{V(s_{i - 1})\\cdot(N(s_{i}) - 1) + r}{N(s_{i})}$ , where $V(s_{i - 1})$ is the previous reward, and $r$ is the reward used to update each $V(s_{i})$ .", + "bbox": [ + 89, + 172, + 482, + 268 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Expansion. Once a node is selected, we expand it by sampling $n$ new discussion paths, each building on the plan (from \"planner\", Alice) and suggestions (from \"critic\", Bob) in the parent node. This generates $n$ new child nodes, representing potential planning routes. The tree, along with its nodes, is stored in an external long-term memory to retain valuable information and facilitate consistent planning across episodes.", + "bbox": [ + 89, + 268, + 482, + 387 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation. Each newly created child node receives a reward $r$ to quantify the quality of the collaboration plan and to guide the search algorithm towards the most promising areas of the tree. Here CoTS incorporates a non-simulated LLM-based reward function to estimate the reward $r$ of a plan in a node. In traditional MCTS, reward estimation requires agents to execute their actions for receiving environment feedback, like completing a subtask, and thus is high-cost. So CoTS introduces an LLM-based reward function based on two components: (1) an allocation reward assessing coherence and efficiency, and (2) a cost reward evaluating the spatial movement cost of agents. To ensure consistency, we standardize the evaluation criteria to provide clear guidelines to minimize ambiguity. Each criterion is scored on a scale of 1 to 5, with the final reward normalized between 0 and 1. See examples in Appendix B.1.", + "bbox": [ + 89, + 388, + 482, + 628 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Backpropagation. Each evaluation reward is propagated back along the path to update node rewards, with formulas ensuring that parent nodes integrate the feedback, guiding future selections. Specifically, the back-propagation updates node rewards $N(s_{i}) = N(s_{i - 1}) + 1$ and $V(s_{i}) = \\frac{V(s_{i - 1})N(s_{i - 1}) + r}{N(s_i)}$ for each node $s_0,s_1,\\ldots ,s_p$ along the path from the root (initial state $s_0$ ) to leaf (terminal state $s_p$ ). Accordingly these estimated rewards and visiting number are used in \"Expansion\" step for node selection, directing the search toward the most promising areas of the tree.", + "bbox": [ + 89, + 630, + 482, + 785 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2.2. Collaborative Plan Evaluation and Adaptation", + "text_level": 1, + "bbox": [ + 89, + 791, + 460, + 806 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we will explain why we introduced the planning evaluation module (a sub-module in the collaborative planning module, as shown in Fig. 2) and its function. The collaborative plan provides high-level embodied agent action guidance, but as the task is executed and agents act, the embodied environment gradually changes (e.g. the agents'", + "bbox": [ + 89, + 810, + 482, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "positions, the agents' progress, and the current status, etc.). Therefore, the collaborative plan needs to be updated based on the latest environmental information.", + "bbox": [ + 511, + 90, + 903, + 135 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "At this point, two scenarios arise: If the plan is updated too slowly or not updated at all, the current plan becomes outdated and cannot effectively guide the embodied agents in executing tasks. However, in a multi-agent system, instantly updating the plan with every progress can be inefficient and counterproductive. It may lead to excessive queries to the LLM and environments. Furthermore, frequent plan revisions may introduce conflicting strategies due to minor environmental shifts, potentially disrupting agent actions and lowering overall efficiency. This is illustrated in Fig. 4 (c): if the updated plan conflicts with the action that Bob is currently performing, it may interrupt Bob's action which even almost finishes the subtask.", + "bbox": [ + 511, + 136, + 906, + 330 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To address this, we introduce a plan evaluation module within CoTS to continuously evaluate the viability of the current collaborative plan. When an agent makes progress, the module determines if the existing plan remains effective based on the status of task execution. If the evaluation metric meets a defined threshold, indicating that the plan is still viable, it is retained. Otherwise, CoTS will call the collaborative planning module to re-search and update the plan. By focusing on the feasibility and long-term stability of the plan rather than continuous recalibration, this module supports efficient, coherent collaboration, enabling agents to complete tasks reliably in a dynamic environment.", + "bbox": [ + 511, + 333, + 908, + 513 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Plan Parsing and Execution Module", + "text_level": 1, + "bbox": [ + 511, + 523, + 826, + 539 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To transform the collaborative plan generated in Sec. 4.2 into executable actions, CoTS incorporates an action parsing module and a plan decomposition module. Action parsing module bridges high-level planning with real-world actions, generating a set of feasible actions for each agent based on current memory. For example, if an agent's hands are free, it will list objects that can be grabbed; if the agent is in an unknown room, it will list movement options. This module ensures that all sub-plans are formatted into natural language and returned for subsequent process.", + "bbox": [ + 511, + 544, + 906, + 696 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Plan decomposition module understands and breaks down high-level tasks from the collaboration plan into specific and prioritized actions. The agent interprets the task arrangement, selects actions from the parsed options that best advance the collaborative plan's goals, and finally translates each chosen action into real-world execution.", + "bbox": [ + 511, + 696, + 905, + 785 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 511, + 800, + 645, + 816 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "**Benchmarks. Following CoELA [43], we evaluate the performance of CoTS in two highly challenging embodied-world multi-agent environments: (1) TDW-MAT (ThreeD World Multi-Agent Transport) [8] extends the ThreeD World Transport Challenge by simulating collaborative", + "bbox": [ + 511, + 825, + 905, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "29517", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/a61b2d34aedad11f1ce58f63840f94b2af4f9dc4112254e58c4318cf587bad5d.jpg", + "table_caption": [ + "Table 1. Quantitative results on TDW-MAT. Consistent with CoELA [43], TDW-MAT has two environment settings: with oracle perception and without oracle perception. The results marked with * are from CoELA [43]. MAT uses central observation and oracle perception. $91_{+9.64\\%}$ means the relative improvement when comparing with CoELA is $9.64\\%$ ." + ], + "table_footnote": [], + "table_body": "
Transport Rate (%)Base AgentRL AgentLLM Agent
RHP*CoRHP*MAT*GPT 3.5GPT 4
CoELARoCoCaPoCoTS(ours)CoELA*RoCoCaPoCoTS(ours)
TDW-MAT w/ Oracle Perception
Stuff(↑)4974177380+9.68%84+15.07%81+10.96%8385+2.41%87+4.82%91+9.64%
Food(↑)5276137274+2.78%85+18.06%83+15.28%8789+2.30%90+3.45%94+8.05%
Average(↑)5075157377+5.48%84+15.07%82+12.33%8587+2.35%89+4.71%93+9.41%
TDW-MAT w/o Oracle Perception
Stuff(↑)3654-3947+20.51%45+15.38%50+28.21%6164+4.92%64+4.92%77+26.22%
Food(↑)4967-6764-4.48%70+4.48%67+0.00%8280-2.44%85+3.65%88+7.32%
Average(↑)4361-5255+5.77%57+9.62%59+13.46%7172+1.41%74+4.23%83+16.90%
", + "bbox": [ + 133, + 137, + 864, + 291 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/3b496952cb4010d9fd9cc36461b2345cb286a40b77a6380041743e27666a57be.jpg", + "table_caption": [ + "Table 2. Quantitative results on C-WAH. LLM-based agents are driven by GPT 4. Results marked with * are from CoELA [43]." + ], + "table_footnote": [], + "table_body": "
Average StepVisual Obs(↓)Symbolic Obs(↓)
Base AgentMHP*141111
CoMHP*10375
LLM AgentCoELA*9257
RoCo89+3.26%57+0.00%
CaPo83+9.78%51+10.53%
CoTS(ours)81+11.96%49+14.04%
", + "bbox": [ + 112, + 345, + 457, + 454 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "transportation tasks in a physics-based household environment. We select 24 tasks, split evenly between food and stuff transportation. (2) C-WAH (Communicative Watch-And-Help) [29] extends the Watch-And-Help challenge within the VirtualHome-Social platform [28]. It features five common household scenarios. Our test set includes 10 tasks, with two tasks per scenario. The objective is to complete all sub-goals, prioritizing faster completion.", + "bbox": [ + 88, + 472, + 482, + 593 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metrics. On TDW-MAT, transport rate is evaluated by the percentage of target objects successfully delivered to their designated locations within a 3,000-frame timeframe. For C-WAH, efficiency is measured by the number of time steps needed to complete all assigned tasks, where fewer steps indicate higher efficiency.", + "bbox": [ + 89, + 595, + 482, + 686 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baseline Methods. The baselines include three types of embodied agents: (1) Base Agent: RHP [8] uses heuristic rules for navigation and $\\mathbf{A}^*$ for path planning. MHP [18] employs Monte Carlo Tree Search (MCTS) for high-level planning coupled with regression planning for task execution (2) RL-Based Agent: MAT [39] is a multi-agent reinforcement learning framework, featuring a centralized decision transformer that synthesizes actions from shared observations. (3) LLM-Based Agent: CoELA [43] is a leading collaboration framework for embodied agents that integrates the reasoning capabilities of LLMs. It facilitates communication and interaction between agents through natural language. RoCo [24] and CaPo [22] are multi-agent collaboration methods that enhance collaboration by gener", + "bbox": [ + 89, + 688, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ating coordinated plans through structured agent dialogues. Setup. We conduct all experiments on an Intel Core i9- 10980XE CPU and an NVIDIA GeForce RTX 3090 GPU. We utilize both GPT-4 and GPT-3.5-Turbo [1], configuring them with a temperature setting of 0.7 and top-1 sampling to produce outputs up to 512 tokens.", + "bbox": [ + 511, + 315, + 906, + 409 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Main Results", + "text_level": 1, + "bbox": [ + 511, + 420, + 653, + 436 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "CoTS Outperforms the Strong Baselines. Following the experimental frameworks of CoELA [43] and RoCo [24], we configure two-agent cooperation, excluding RHP. As demonstrated in Table 1, CoTS achieves best performance on the TDW-MAT task. Specifically, when powered by GPT-4 within the Oracle Perception environment, CoTS surpasses CoELA, achieving a $9.41\\%$ average relative improvement in task completion, and a $16.9\\%$ improvement in the absence of Oracle Perception. The primary strength of CoTS is its ability to generate superior long-term strategies through collaborative tree search.", + "bbox": [ + 511, + 444, + 906, + 609 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Conversely, CoELA focuses on short-term actions derived from immediate information, which limits its strategic depth and adaptability to future scenarios and environmental shifts. RoCo and CaPo are similar. Although they are capable of developing plans, they rely on a single-branch discussion method, which may result in suboptimal or inaccurate strategies. Moreover, Their frequent plan revisions can interfere with agent coordination. Therefore, while RoCo and CaPo surpass CoELA, they are inferior to CoTS. When driven by GPT-3.5, CaPo slightly outperforms CoTS on some metrics. The main reason for this is that GPT-3.5 has weaker evaluation and scoring capabilities. However, when powered by GPT-4, CoTS significantly outperforms CaPo. Our findings underscore that multi-agent collaboration generally surpasses single-agent strategies, exemplified by RHP. The performance of reinforcement learning-based MAT is poor, mainly because TDW-MAT is too difficult for it with long time domain and sparse reward signals.", + "bbox": [ + 511, + 612, + 908, + 883 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The performance of LLM-based agents primarily de", + "bbox": [ + 532, + 885, + 903, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "29518", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "(a) CoELA leads to short-sighted action plans that ignore precious action steps", + "text_level": 1, + "bbox": [ + 102, + 93, + 334, + 133 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Background: Alice is in living room holding a container. Bob finds two target objects in living room $< 1000>$ .", + "bbox": [ + 102, + 137, + 321, + 176 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7ec545d557c91cb677e0ac6a1a31ce96a613fb553992580942afa15368db28e0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 99, + 179, + 346, + 301 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "CoELA's plan : Alice goes to the living room to help Bob, because Alice can take these objects.", + "bbox": [ + 101, + 301, + 331, + 338 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our plan: Alice explores the current room, find objects, then transport them to the bedroom. Then go to the living room $< 1000>$ , no actions are wasted.", + "bbox": [ + 99, + 339, + 328, + 388 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "(b) RoCo's single-chain dialogue leads to suboptimal collaboration plan", + "text_level": 1, + "bbox": [ + 377, + 92, + 606, + 133 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Background: Alice and Bob appear in the same living room and now need to explore some rooms to find objects.", + "bbox": [ + 375, + 138, + 602, + 176 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/df07c36f0467d2e805c8a4847c030012552c8294be636c605e3c962eddf7ad6e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 374, + 176, + 625, + 301 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "RoCo's plan : Alice explore the current room. Bob explore the current room.", + "bbox": [ + 375, + 301, + 609, + 327 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our plan: Alice goes to the nearest room, Bedroom $< 2000>$ , while Bob explores the current room. Exploring rooms separately would improve efficiency.", + "bbox": [ + 372, + 338, + 622, + 391 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "(c) Frequent plan updates may hinder completion of current actions", + "text_level": 1, + "bbox": [ + 658, + 94, + 879, + 133 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Background: Bob was transporting objects to the bed, Alice found a bread in the living room<6000>", + "bbox": [ + 658, + 137, + 870, + 176 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/061d1fe605d5d158f3c45e816e2ca1490761704a4290c947d3a587adc4e3c613.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 653, + 176, + 895, + 301 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "RoCo's Plan: Bob goes to the living room $< 6000>$ to help Alice grasp the bread.", + "bbox": [ + 653, + 301, + 893, + 338 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our plan: The current plan is still valid. Bob is close to the bedroom. He can first deliver the objects in his hand to the bed, and then help Alice.", + "bbox": [ + 653, + 338, + 890, + 390 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/416f57abde0d9bc8ea29cbe41eebe0e3a9a49735b8e634b2cf408935b176fec2.jpg", + "image_caption": [ + "Figure 4. Some interesting example cases demonstrate that our CoTS achieves better collaboration than RoCo and CoELA. These scenes are taken from the experiments in TDW-MAT." + ], + "image_footnote": [ + "There seems to be no plan. Both Alice and Bob are not holding anything. Bob in bedroom and has found an container, while Alice is in the. A collaboration plan needs to be updated. Reward: 1" + ], + "bbox": [ + 91, + 441, + 272, + 542 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/67ebbf0492b3fb860a8c8644426bd3bed0427552740056ca7d3d2f47705ecba0.jpg", + "image_caption": [], + "image_footnote": [ + "Alice have found an apple in the fridge and Bob is holding a cupcake. Current progress is consistent with the collaborative plan and future actions are outlined in the plan. There is no need to update the plan. Reward: 4" + ], + "bbox": [ + 282, + 441, + 473, + 542 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "pends on their reasoning abilities and the depth of embedded world knowledge. Agents powered by GPT-4, as demonstrated by CoTS, significantly surpass those based on GPT-3.5, highlighting the advancements in LLM technologies and their impact on agent proficiency. As shown in Table 2, CoTS achieves superior efficiency, improving by $11.96\\%$ and $14.04\\%$ over CoELA in Visual Obs and Symbolic Obs settings, respectively. These results underscore CoTS's robust adaptability across various embodied environments and its potential for real-world applications.", + "bbox": [ + 89, + 683, + 482, + 835 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Qualitative Analysis", + "text_level": 1, + "bbox": [ + 89, + 847, + 282, + 863 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "CoTS Can Develop Long-term Strategic Plans. Unlike CoELA, which is adept at responding to immediate tasks", + "bbox": [ + 89, + 869, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/51da022b49258ffbf5632dc1659b5bacbabb95bdc0c5f68b8ea8931500179902.jpg", + "image_caption": [ + "Figure 5. Plan Evaluation and Update. The plan evaluation module gives a lower reward when it determines that the current plan needs to be updated, and a higher reward when it believes the current plan is still valid. The reward range is from 1 to 5.", + "Figure 6. Comparison of transportation rates for GPT-4-based CoTS and CoELA at different time steps on TDW-MAT." + ], + "image_footnote": [], + "bbox": [ + 558, + 444, + 856, + 569 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "but falters in sustained coordination, CoTS excels in long-term strategic planning. As illustrated in Fig. 4 (a), when Bob discovers a new target that requires assistance, Alice, under CoTS's direction, continues to thoroughly explore her current room before moving the item to the bed. This systematic approach not only supports Bob but also adheres to an overarching collaborative strategy, optimizing both immediate and future task alignments.", + "bbox": [ + 511, + 628, + 906, + 750 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "CoTS Generates More Optimal Collaboration Plans. Unlike RoCo to generate plans sequentially through a single-chain dialogue, CoTS makes more optimized collaboration plans. Due to the randomness of LLMs and the embodied environment's complexity, RoCo's plans may sometimes be suboptimal. For example, in Fig. 4 (b), RoCo instructs both Alice and Bob to explore their respective rooms. However, they are currently in the same room, making joint exploration redundant. CoTS effectively identifies this and assigns Alice to explore the nearest room instead.", + "bbox": [ + 511, + 750, + 908, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "29519", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/cbda7706dce07bdf1b38b22867ba2e33c5cfaf3247cd8abf647981a304ef0b4c.jpg", + "image_caption": [ + "Figure 7. Comparison of transportation rates with different branching factors. More branching factors means more new nodes are generated each time in the expansion phase." + ], + "image_footnote": [], + "bbox": [ + 101, + 95, + 475, + 234 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "CoTS Ensures Smooth Completion of Agent Actions. In RoCo, new progress often prompt discussions and lead to new plans that can override existing ones. As illustrated in Fig. 4 (c), Bob, tasked with transporting items to the bed, must halt his activity to help Alice following a revised plan. Conversely, CoTS incorporates a plan evaluation module that validates the feasibility of plans before any adjustments are made. In Fig. 4 (c), CoTS allows Bob to finish transporting before next action. This approach avoids chaotic agent actions and ensures timely plan updates when necessary.", + "bbox": [ + 89, + 295, + 482, + 446 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "CoTS Dynamically Manages Collaborative Plans. As shown in Fig. 5, the assessment module continuously evaluates the viability of existing plans. It assigns a lower reward (e.g. 1) if collaboration necessitates devising a new plan or modifying an existing one, prompting agents to adapt their strategies accordingly. Conversely, when the current progress aligns with planned tasks and their distribution among agents is deemed efficient, the module rewards this alignment with a higher value (e.g. 4). This incentivizes agents to persist with the established collaborative plan, avoiding unnecessary replanning.", + "bbox": [ + 89, + 446, + 482, + 613 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.3. Ablation Study and Additional Analysis", + "text_level": 1, + "bbox": [ + 89, + 622, + 431, + 638 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The Core Modules of CoTS Have Important Impacts. The Table 3 shows the ablation results of CoTS. To examine the impact of tree search, we modify it to a single-branch structure, where agents can only communicate along one path. The experimental results indicate a $5.38\\%$ relative decrease in the average transportation rate, mainly because the limited communication does not consistently yield correct conclusions and diverges from typical human discussion patterns. Furthermore, eliminating the reward function and plan update evaluation components from CoTS leads to variable performance declines, highlighting the critical roles these elements play in the system's efficacy.", + "bbox": [ + 89, + 643, + 482, + 824 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Analysis of Agent Collaboration Progress at Different Time Steps. We analyze the transportation rates of CoELA and CoTS on TDW-MAT at various time steps, as depicted in Fig. 6. An interesting observation emerges at 500 steps, where CoTS exhibits a lower performance compared to", + "bbox": [ + 89, + 825, + 482, + 901 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/0e32c318cfdd26fe09e10f2a61d9265eca0fb6931bd4f3f4cdf12ed20de707db.jpg", + "table_caption": [ + "Table 3. Effects of the reward function, collaboration tree search, and plan evaluation module in GPT-4-driven CoTS on TDW-MAT." + ], + "table_footnote": [], + "table_body": "
Transport RateStuff(↑)Food(↑)Average(↑)
CoTS w/o Reward87-4.40%90-4.26%89-4.30%
CoTS w/o MCTS Plan85-6.59%91-3.19%88-5.38%
CoTS w/o E&U88-3.30%92-2.13%90-3.23%
CoTS919493
", + "bbox": [ + 535, + 122, + 883, + 208 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "CoELA, attributed to CoTS's strategy of prioritizing item collection early to enhance transport capacity. However, this initial disadvantage is counterbalanced by a significant improvement at 2000 steps, where CoTS markedly surpasses CoELA. This shift underscores CoTS's effectiveness in optimizing agent collaboration for task execution.", + "bbox": [ + 511, + 232, + 903, + 321 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Impact of Different Branching Factors on Results. During the expansion phase of MCTS, we define the number of new candidate nodes generated at each step as the \"branching factors.\" As shown in Fig. 7, increasing the number of branching factors expands the selection space, leading to a better plan. Adjusting the branching factor allows for a balance between performance and efficiency.", + "bbox": [ + 511, + 323, + 903, + 428 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Besides, the human-agent collaboration and the cost of CoTS are discussed in Appendix A.1 and Appendix A.2, respectively. CoTS also demonstrates significant advantages.", + "bbox": [ + 511, + 429, + 903, + 474 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusions", + "text_level": 1, + "bbox": [ + 511, + 488, + 638, + 503 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper introduces CoTS, a collaborative framework that leverages large language models with a modified Monte Carlo tree search to enhance multi-agent task planning in complex and dynamic environments. By enabling agents to evaluate multiple potential plans and refine strategies based on progress, CoTS significantly advances over prior methods like CoELA and RoCo. The addition of a plan feasibility module ensures smooth task execution by adapting plans only when necessary. Experimental results on multi-objective tasks confirm CoTS's superior performance.", + "bbox": [ + 511, + 513, + 903, + 664 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitation. CoTS relies on LLMs' reasoning abilities. We find that GPT-4 surpasses GPT-3.5-turbo in instruction-following and response quality, while GPT-3.5-turbo can be verbose and occasionally inaccurate. Despite this, even for GPT-3.5, CoTS improves agent collaboration through collaborative planning search in the collaboration tree.", + "bbox": [ + 511, + 665, + 903, + 755 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Acknowledgments", + "text_level": 1, + "bbox": [ + 511, + 768, + 691, + 787 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work is supported by National Natural Science Foundation of China (U2133202, 52305570), China Postdoctoral Science Foundation (2022M720955), and Outstanding Doctoral Dissertation Funding Project of Heilongjiang Province (LJYXL2022-011). Lizheng Zu also acknowledges the support of the China Scholarship Council program (202406120111).", + "bbox": [ + 511, + 794, + 903, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "29520", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 1, 2, 6", + "[2] David Brandfonbrener, Simon Henniger, Sibi Raja, Tarun Prasad, Chloe R Loughridge, Federico Cassano, Sabrina Ruixin Hu, Jianang Yang, William E Byrd, Robert Zinkov, et al. Vermcts: Synthesizing multi-step programs using a verifier, a large language model, and tree search. In The 4th Workshop on Mathematical Reasoning and AI at NeurIPS'24, 2024. 3", + "[3] Tom B Brown. Language models are few-shot learners. arXiv preprint arXiv:2005.14165, 2020. 1", + "[4] Cameron B Browne, Edward Powley, Daniel Whitehouse, Simon M Lucas, Peter I Cowling, Philipp Rohlfshagen, Stephen Tavener, Diego Perez, Spyridon Samothrakis, and Simon Colton. A survey of monte carlo tree search methods. IEEE Transactions on Computational Intelligence and AI in games, 4(1):1-43, 2012. 4, 3", + "[5] Abhinav Dahiya, Alexander M Aroyo, Kerstin Dautenhahn, and Stephen L Smith. A survey of multi-agent human-robot interaction systems. Robotics and Autonomous Systems, 161: 104335, 2023. 1", + "[6] Norman Di Palo, Arunkumar Byravan, Leonard Hasenclever, Markus Wulfmeier, Nicolas Heess, and Martin Riedmiller. Towards a unified agent with foundation models. arXiv preprint arXiv:2307.09668, 2023. 2", + "[7] Amal Feriani and Ekram Hossain. Single and multi-agent deep reinforcement learning for ai-enabled wireless networks: A tutorial. IEEE Communications Surveys & Tutorial, 23(2):1226-1252, 2021. 1", + "[8] Chuang Gan, Siyuan Zhou, Jeremy Schwartz, Seth Alter, Abhishek Bhandwaldar, Dan Gutfreund, Daniel LK Yamins, James J DiCarlo, Josh McDermott, Antonio Torralba, et al. The threadworld transport challenge: A visually guided task-and-motion planning benchmark towards physically realistic embodied ai. In 2022 International conference on robotics and automation (ICRA), pages 8847-8854. IEEE, 2022. 2, 5, 6", + "[9] Ran Gong, Qiuyuan Huang, Xiaojian Ma, Yusuke Noda, Zane Durante, Zilong Zheng, Demetri Terzopoulos, Li Fei-Fei, Jianfeng Gao, and Hoi Vo. Mindagent: Emergent gaming interaction. In Findings of the Association for Computational Linguistics: NAACL 2024, pages 3154–3183, 2024. 3", + "[10] Taicheng Guo, Xiuying Chen, Yaqi Wang, Ruidi Chang, Shichao Pei, Nitesh V Chawla, Olaf Wiest, and Xiangliang Zhang. Large language model based multi-agents: A survey of progress and challenges. arXiv preprint arXiv:2402.01680, 2024. 1", + "[11] Xudong Guo, Kaixuan Huang, Jiale Liu, Wenhui Fan, Natalia Vélez, Qingyun Wu, Huazheng Wang, Thomas L Griffiths, and Mengdi Wang. Embodied llm agents learn to cooperate in organized teams. arXiv preprint arXiv:2403.12482, 2024. 2" + ], + "bbox": [ + 93, + 114, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[12] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 3", + "[13] Sirui Hong, Xiawu Zheng, Jonathan Chen, Yuheng Cheng, Jinlin Wang, Ceyao Zhang, Zili Wang, Steven Ka Shing Yau, Zijuan Lin, Liyang Zhou, et al. Metagpt: Meta programming for multi-agent collaborative framework. arXiv preprint arXiv:2308.00352, 2023. 2", + "[14] Wenlong Huang, Pieter Abbeel, Deepak Pathak, and Igor Mordatch. Language models as zero-shot planners: Extracting actionable knowledge for embodied agents. In International conference on machine learning, pages 9118-9147. PMLR, 2022. 1", + "[15] Steven James, George Konidaris, and Benjamin Rosman. An analysis of monte carlo tree search. In Proceedings of the AAAI Conference on Artificial Intelligence, 2017. 4", + "[16] Feibo Jiang, Yubo Peng, Li Dong, Kezhi Wang, Kun Yang, Cunhua Pan, Dusit Niyato, and Octavia A Dobre. Large language model enhanced multi-agent systems for 6g communications. IEEE Wireless Communications, 2024. 2", + "[17] Bowen Jin, Chulin Xie, Jiawei Zhang, Kashob Kumar Roy, Yu Zhang, Zheng Li, Ruirui Li, Xianfeng Tang, Suhang Wang, Yu Meng, et al. Graph chain-of-thought: Augmenting large language models by reasoning on graphs. arXiv preprint arXiv:2404.07103, 2024. 3", + "[18] Richard E Korf. Planning as search: A quantitative approach. Artificial intelligence, 33(1):65-88, 1987. 6, 5", + "[19] Guohao Li, Hasan Hammoud, Hani Itani, Dmitrii Khizbullin, and Bernard Ghanem. Camel: Communicative agents for\" mind\" exploration of large language model society. Advances in Neural Information Processing Systems, 36: 51991-52008, 2023. 2", + "[20] Xin Li. Positioning is all you need. arXiv preprint arXiv:2404.01183, 2024. 2", + "[21] Iou-Jen Liu, Unnat Jain, Raymond A Yeh, and Alexander Schwing. Cooperative exploration for multi-agent deep reinforcement learning. In International conference on machine learning, pages 6826-6836. PMLR, 2021. 1", + "[22] Jie Liu, Pan Zhou, Yingjun Du, Ah-Hwee Tan, Cees GM Snoek, Jan-Jakob Sonke, and Efstratios Gavves. Capo: Cooperative plan optimization for efficient embodied multiagent cooperation. arXiv preprint arXiv:2411.04679, 2024. 2, 6, 5", + "[23] Yang Liu, Weixing Chen, Yongjie Bai, Jingzhou Luo, Xinshuai Song, Kaixuan Jiang, Zhida Li, Ganlong Zhao, Junyi Lin, Guanbin Li, et al. Aligning cyber space with physical world: A comprehensive survey on embodied ai. arXiv preprint arXiv:2407.06886, 2024. 1", + "[24] Zhao Mandi, Shreeya Jain, and Shuran Song. Roco: Dialectic multi-robot collaboration with large language models. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 286–299. IEEE, 2024. 1, 2, 3, 6, 5", + "[25] Reiichiro Nakano, Jacob Hilton, Suchir Balaji, Jeff Wu, Long Ouyang, Christina Kim, Christopher Hesse, Shantanu Jain, Vineet Kosaraju, William Saunders, et al. Webgpt: Browser-assisted question-answering with human feedback. arXiv preprint arXiv:2112.09332, 2021. 2" + ], + "bbox": [ + 516, + 92, + 905, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "29521", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[26] Bo Ni and Markus J Buehler. Mechagents: Large language model multi-agent collaborations can solve mechanics problems, generate new data, and integrate knowledge. Extreme Mechanics Letters, 67:102131, 2024. 2", + "[27] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. Advances in neural information processing systems, 35:27730-27744, 2022. 2", + "[28] Xavier Puig, Kevin Ra, Marko Boben, Jiaman Li, Tingwu Wang, Sanja Fidler, and Antonio Torralba. Virtualhome: Simulating household activities via programs. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8494-8502, 2018. 6", + "[29] Xavier Puig, Tianmin Shu, Shuang Li, Zilin Wang, Yuan-Hong Liao, Joshua B Tenenbaum, Sanja Fidler, and Antonio Torralba. Watch-and-help: A challenge for social perception and human-ai collaboration. arXiv preprint arXiv:2010.09890, 2020. 6", + "[30] Xavier Puig, Eric Undersander, Andrew Szot, Mikael Dallaire Cote, Tsung-Yen Yang, Ruslan Partsey, Ruta Desai, Alexander William Clegg, Michal Hlavac, So Yeon Min, et al. Habitat 3.0: A co-habitat for humans, avatars and robots. arXiv preprint arXiv:2310.13724, 2023. 2", + "[31] Chen Qian, Wei Liu, Hongzhang Liu, Nuo Chen, Yufan Dang, Jiahao Li, Cheng Yang, Weize Chen, Yusheng Su, Xin Cong, et al. Chatdev: Communicative agents for software development. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 15174-15186, 2024. 2", + "[32] Timo Schick, Jane Dwivedi-Yu, Roberto Dessì, Roberta Raileanu, Maria Lomeli, Eric Hambro, Luke Zettlemoyer, Nicola Cancedda, and Thomas Scialom. Toolformer: Language models can teach themselves to use tools. Advances in Neural Information Processing Systems, 36, 2024. 2", + "[33] Paul E Smaldino, Cody Moser, Alejandro Perez Velilla, and Mikkel Werling. Maintaining transient diversity is a general principle for improving collective problem solving. Perspectives on Psychological Science, 19(2):454-464, 2024. 2", + "[34] Chan Hee Song, Jiaman Wu, Clayton Washington, Brian M Sadler, Wei-Lun Chao, and Yu Su. Llm-planner: Few-shot grounded planning for embodied agents with large language models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2998-3009, 2023. 1", + "[35] Wiebe Van der Hoek and Michael Wooldridge. Multi-agent systems. Foundations of Artificial Intelligence, 3:887-928, 2008. 2", + "[36] Tai Wang, Xiaohan Mao, Chenming Zhu, Runsen Xu, Ruiyuan Lyu, Peisen Li, Xiao Chen, Wenwei Zhang, Kai Chen, Tianfan Xue, et al. Embodiedscan: A holistic multimodal 3d perception suite towards embodied ai. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19757-19767, 2024. 1", + "[37] Jason Wei, Yi Tay, Rishi Bommasani, Colin Raffel, Barret Zoph, Sebastian Borgeaud, Dani Yogatama, Maarten Bosma, Denny Zhou, Donald Metzler, et al. Emergent abilities of" + ], + "bbox": [ + 91, + 92, + 480, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "large language models. arXiv preprint arXiv:2206.07682, 2022. 2", + "[38] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824-24837, 2022. 1, 2, 3", + "[39] Muning Wen, Jakub Kuba, Runji Lin, Weinan Zhang, Ying Wen, Jun Wang, and Yaodong Yang. Multi-agent reinforcement learning is a sequence modeling problem. Advances in Neural Information Processing Systems, 35:16509-16521, 2022. 1, 6, 5", + "[40] Shunyu Yao, Dian Yu, Jeffrey Zhao, Izhak Shafran, Tom Griffiths, Yuan Cao, and Karthik Narasimhan. Tree of thoughts: Deliberate problem solving with large language models. Advances in Neural Information Processing Systems, 36, 2024. 1, 2, 3", + "[41] Bangguo Yu, Hamidreza Kasaei, and Ming Cao. Co-navgpt: Multi-robot cooperative visual semantic navigation using large language models. arXiv preprint arXiv:2310.07937, 2023. 2", + "[42] Di Zhang, Xiaoshui Huang, Dongzhan Zhou, Yuqiang Li, and Wanli Ouyang. Accessing gpt-4 level mathematical olympiad solutions via monte carlo tree self-refine with llama-3 8b: A technical report. arXiv preprint arXiv:2406.07394, 2024. 3", + "[43] Hongxin Zhang, Weihua Du, Jiaming Shan, Qinhong Zhou, Yilun Du, Joshua B Tenenbaum, Tianmin Shu, and Chuang Gan. Building cooperative embodied agents modularly with large language models. arXiv preprint arXiv:2307.02485, 2023. 1, 2, 3, 5, 6", + "[44] Zhonghan Zhao, Ke Ma, Wenhao Chai, Xuan Wang, Kewei Chen, Dongxu Guo, Yanting Zhang, Hongwei Wang, and Gaoang Wang. Do we really need a complex agent system? distill embodied agent into a single model. arXiv preprint arXiv:2404.04619, 2024. 2", + "[45] Andy Zhou, Kai Yan, Michal Shlapentokh-Rothman, Haohan Wang, and Yu-Xiong Wang. Language agent tree search unifies reasoning acting and planning in language models. arXiv preprint arXiv:2310.04406, 2023. 2" + ], + "bbox": [ + 516, + 92, + 903, + 655 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "29522", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Collaborative Tree Search for Enhancing Embodied Multi-Agent Collaboration/509a3bd5-5db7-45e0-bb5b-ba942b10cc92_model.json b/2025/Collaborative Tree Search for Enhancing Embodied Multi-Agent Collaboration/509a3bd5-5db7-45e0-bb5b-ba942b10cc92_model.json new file mode 100644 index 0000000000000000000000000000000000000000..0f1716b3418e926148042887d025933d830a2126 --- /dev/null +++ b/2025/Collaborative Tree Search for Enhancing Embodied Multi-Agent Collaboration/509a3bd5-5db7-45e0-bb5b-ba942b10cc92_model.json @@ -0,0 +1,2255 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.1, + 0.13, + 0.9, + 0.153 + ], + "angle": 0, + "content": "Collaborative Tree Search for Enhancing Embodied Multi-Agent Collaboration" + }, + { + "type": "text", + "bbox": [ + 0.159, + 0.18, + 0.841, + 0.269 + ], + "angle": 0, + "content": "Lizheng \\(\\mathsf{Zu}^{1,2}\\), Lin Lin\\(^{1*}\\), Song Fu\\(^{1*}\\), Na Zhao\\(^{3}\\), Pan Zhou\\(^{4}\\) \n\\(^{1}\\)Harbin Institute of Technology, \\(^{2}\\)Nanyang Technological University, \n\\(^{3}\\)Singapore University of Technology and Design, \\(^{4}\\)Singapore Management University \nzuli_edu@outlook.com, waiwaiyl@hit.edu.cn, fusong@hit.edu.cn \nna_zhao@sutd.edu.sg, panzhou@smu.edu.sg" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.304, + 0.328, + 0.321 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.337, + 0.486, + 0.685 + ], + "angle": 0, + "content": "Embodied agents based on large language models (LLMs) face significant challenges in collaborative tasks, requiring effective communication and reasonable division of labor to ensure efficient and correct task completion. Previous approaches with simple communication patterns carry erroneous or incoherent agent actions, which can lead to additional risks. To address these problems, we propose Cooperative Tree Search (CoTS), a framework designed to significantly improve collaborative planning and task execution efficiency among embodied agents. CoTS guides multi-agents to discuss long-term strategic plans within a modified Monte Carlo tree, searching along LLM-driven reward functions to provide a more thoughtful and promising approach to cooperation. Another key feature of our method is the introduction of a plan evaluation module, which not only prevents agent action confusion caused by frequent plan updates but also ensures plan updates when the current plan becomes unsuitable. Experimental results show that the proposed method performs excellently in planning, communication, and collaboration on embodied environments (CWAH and TDW-MAT), efficiently completing long-term, complex tasks and significantly outperforming existing methods." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.715, + 0.222, + 0.731 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.74, + 0.485, + 0.878 + ], + "angle": 0, + "content": "Embodied agents are increasingly capable of perceiving and interacting with their environments like humans, autonomously managing planning, decision-making, actions, and execution tasks [23, 36]. Recent advances have generated substantial interest in these agents, especially for applications requiring complex task execution in dynamic environments, such as robotic operations in warehouses or domestic settings. In particular, collaborative task completion among multiple agents enhances both efficiency and adapt" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.306, + 0.907, + 0.458 + ], + "angle": 0, + "content": "ability, especially in dynamic scenarios [5, 10]. However, collaboration among embodied agents introduces significant challenges: agents must not only perceive and understand their environment but also communicate, share information, divide tasks, and coordinate actions responsively. For example, consider autonomous robots in a home responding to a request: \"Bring me the iPad and apple, and put the milk in the refrigerator.\" Achieving this requires multi-level coordination, including optimal search strategies, task prioritization, and efficient movement planning." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.459, + 0.909, + 0.581 + ], + "angle": 0, + "content": "Large language models (LLMs) have recently provided embodied agents with advanced natural language understanding, dialogue, and reasoning abilities [1, 3, 38, 40]. These capabilities allow LLMs to decompose complex and long-term tasks into a sequence of manageable sub-goals, making LLM-driven agents a promising alternative to traditional reinforcement learning models [14, 34] which are difficult to train and often generalize poorly [7, 21, 39]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.582, + 0.909, + 0.853 + ], + "angle": 0, + "content": "However, enabling embodied agents to work together in decentralized environments remains a significant and underexplored challenge, as it requires long-term planning and coherent decision-making to coordinate actions efficiently. Early attempts at multi-agent collaboration, such as CoELA [43] and RoCo [24], demonstrate progress but also reveal limitations. For instance, as shown in Fig. 1 (b), CoELA facilitates collaboration by sharing updates through natural language when sub-tasks are completed, yet each agent's decision-making process remains independent, resulting in suboptimal coordination. In contrast, as illustrated in Fig. 1 (c), RoCo develops multi-agent work plans through agent discussions and environmental interactions. However, RoCo relies on a single reasoning path, which is susceptible to the randomness and unpredictability of LLM outputs. This can lead to inefficient or incorrect planning, particularly in applications requiring precision, where such errors can have serious consequences." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.856, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Contributions. To address these limitations, we propose Cooperative Tree Search (CoTS), a framework designed to significantly improve collaborative planning and task ex" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.238, + 0.901 + ], + "angle": 0, + "content": "*Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "29513" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.1, + 0.081, + 0.51, + 0.094 + ], + "angle": 0, + "content": "Task: transport 2 oranges, 3 apples, 1 banana, 3 breads, 1 burger to the bed." + }, + { + "type": "text", + "bbox": [ + 0.099, + 0.094, + 0.873, + 0.12 + ], + "angle": 0, + "content": "Background: Agents randomly appear in a home environment with a total of six rooms and need to discover objects through exploration. Each agent can carry up to two objects. Finding and using containers can improve efficiency." + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.125, + 0.37, + 0.278 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.12, + 0.286, + 0.208, + 0.307 + ], + "angle": 0, + "content": "(a) The Embodied Environment" + }, + { + "type": "image_caption", + "bbox": [ + 0.222, + 0.284, + 0.387, + 0.306 + ], + "angle": 0, + "content": "(b) CoELA (Communication but independent decision-making.)" + }, + { + "type": "image", + "bbox": [ + 0.373, + 0.125, + 0.626, + 0.278 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.418, + 0.286, + 0.589, + 0.308 + ], + "angle": 0, + "content": "(c) RoCo (Discussion, Planning and Execution)" + }, + { + "type": "image", + "bbox": [ + 0.629, + 0.125, + 0.895, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.647, + 0.285, + 0.873, + 0.306 + ], + "angle": 0, + "content": "(d) Ours (Plan Searching, Evaluation, and Plan-Guided Execution)" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.318, + 0.907, + 0.388 + ], + "angle": 0, + "content": "Figure 1. The task execution processes of CoELA, RoCo, and our CoTS. (1) In CoELA, agents update each other on sub-task completion and individually determine their next actions without long-term strategic planning. (2) RoCo uses multi-round dialogues and summarizes the plan at the end, but its single-dialogue chain can lead to suboptimal plans and even errors due to the randomness and unpredictability of LLMs. (3) Within a modified Monte Carlo Tree Search framework, CoTS allows agents to evaluate multiple planning paths, enabling them to collaboratively deliberate and identify the most effective course of action based on task progress and environmental context." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.402, + 0.482, + 0.492 + ], + "angle": 0, + "content": "cution efficiency among embodied agents. Inspired by human decision-making, CoTS allows agents to search multiple planning paths by combining exploration and backtracking, enabling them to collaboratively deliberate and identify the most effective course of action based on task progress and environmental context." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.493, + 0.483, + 0.78 + ], + "angle": 0, + "content": "As illustrated in Fig. 1 (d), CoTS advances agent collaboration in embodied environments through two core innovations. First, we introduce a collaborative planning module based on a modified Monte Carlo Tree Search (MCTS) framework, which enables agents to explore multiple planning paths and select the most coherent and efficient long-term plan. Unlike traditional MCTS which relies on simulation-based rewards, CoTS incorporates a non-simulated LLM-based reward function suited for high-cost physical movement interactions in embodied environments. In this way, CoTS transforms the traditional linear sequence of action steps into a search process, exploring potential paths within a shared reasoning and action space to develop robust collaborative strategies. Second, CoTS features a plan evaluation module, that actively monitors the relevance and feasibility of the selected plan during execution. This mechanism allows agents to adapt plans as needed, reducing unnecessary disruptions from frequent plan updates while still accommodating environmental changes." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Our experiments on two long-term multi-objective planning tasks — TDW-MAT [43] and C-WAH [8] — demonstrate that CoTS significantly outperforms current state-of-the-art methods, e.g., CoELA [43] and RoCo [24], in terms of task completion rate and efficiency. These results highlight the effectiveness and superiority of CoTS, underscoring its potential for high-stakes applications that demand reliable multi-agent collaboration." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.4, + 0.655, + 0.416 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.43, + 0.907, + 0.672 + ], + "angle": 0, + "content": "LLM-based Agents. Recent advances in LLMs have unlocked near-human reasoning and planning capabilities [1, 27, 37]. Researchers have leveraged LLMs as the \"brain\" or controller of agents, expanding their perceptual and action capacities through multimodal perception and tool use [6, 25, 32, 38, 40, 45]. Multi-agent systems [16, 35] show particular promise, enabling agents to collaborate on planning, discussion, and decision-making—similar to how humans solve complex problems in teams [26, 33]. Works such as MetaGPT [13], CAMEL [19], and ChatDev [31] decompose complex tasks into simpler, manageable subtasks to reduce hallucinations and improve task-solving capabilities. However, embodied multi-agent systems [20, 44] require advanced communication and face significant challenges in navigating a vast action search space in complex, dynamic environments, especially for long-term tasks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Embodied Multi-Agent Cooperation. Research on embodied multi-agent systems is still emerging. Habitat [30] explores agents' social awareness, enabling adaptive assistance strategies in collaboration. Co-NavGPT [41] uses LLMs for multi-robot collaborative goal navigation, encoding environmental data to improve scene understanding. Organized-LLM-Agents [11] reduces redundancy in multiagent collaboration through structured prompts, enhancing coordination. CoELA [43] enables collaborative agents to communicate and plan through natural language and chain-of-thought processes, yet lacks strategic long-term planning. RoCo [24] and CaPo [22] advance multi-agent collaboration with high-level communication and path planning, but their single reasoning path often leads to suboptimal plans and even errors. Although some existing work" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "29514" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.139, + 0.089, + 0.848, + 0.325 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.327, + 0.908, + 0.397 + ], + "angle": 0, + "content": "Figure 2. The framework of CoTS. Upon receiving a new task in the embodied environment, each embodied agent initially employs the Perception Module to gather raw observational data, which is then stored in its Memory Module. In the collaborative planning module, agents guide the exploration of collaborative plans toward more promising directions through the collaborative tree search. The explored plans are retained within the memory module as a special 'extended memory' and are periodically updated based on assessments from the planning evaluation module. Finally, agents perform actions aligned with the collaborative plan and tasks." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.408, + 0.484, + 0.469 + ], + "angle": 0, + "content": "on LLMs involves multi-branch reasoning [2, 17, 40, 42], they are inadequate for complex embodied environments and agent collaboration. Our goal is to develop embodied agents capable of searching long-term strategic planning." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.481, + 0.231, + 0.497 + ], + "angle": 0, + "content": "3. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.508, + 0.483, + 0.794 + ], + "angle": 0, + "content": "Task Assumption. As introduced by previous work [9, 24], multi-agent collaborative tasks can be viewed as a Markov game, which is defined by a tuple \\((\\mathcal{N},\\mathcal{S},\\mathcal{A},P,r,\\gamma)\\), in which \\(\\mathcal{N}\\) is the set of agents (e.g. Alice and Bob), \\(\\mathcal{S}\\) is the state space, \\(\\mathcal{A} = \\prod_{i=1}^{n}\\mathcal{A}^i\\) is the combination of all action spaces, \\(P: \\mathcal{S} \\times \\mathcal{A} \\times \\mathcal{S} \\to [0,1]\\) is the transition probability function, \\(r: \\mathcal{S} \\times \\mathcal{A} \\to \\mathbb{R}\\) is the reward score, and \\(\\gamma \\in [0,1)\\) is the discount factor. In the task, each agent at time step \\(t \\in \\mathbb{N}\\) obtains the information by perception module from the embodied environment \\(s_t \\in \\mathcal{S}\\), chooses and executes an action \\(a_t^i \\in \\mathcal{A}^i\\) from its available policy \\(\\pi^i(\\cdot|s_t)\\). This will along with the actions of other agents, constitutes a joint action \\(\\mathbf{a}_t = (a_t^1,a_t^2,\\dots,a_t^n) \\in \\mathcal{A}\\) derived from the joint policy \\(\\pi(\\cdot|s_t) = \\prod_{i=1}^{n}\\pi^i(\\cdot|s_t)\\). Then agents perceive a new state \\(s_{t+1}\\) with probability \\(P(s_{t+1}|s_t,\\mathbf{a}_t)\\). We seek to determine a collaborative plan that maximizes the \\(J(\\pi) \\triangleq \\mathbb{E}_{s_0,\\infty} \\sim P,a_0,\\infty \\sim \\pi[\\sum_{i=0}^{\\infty}\\gamma^i r_i]\\). For this paper, the goal is to coordinate the distribution of sub-tasks among agents so as to complete tasks more efficiently." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.484, + 0.902 + ], + "angle": 0, + "content": "CoELA [43]. It contains five modules: 1) Perception, 2) Memory, 3) Communication, 4) Planning and 5) Execution. The Perception module leverages a trained Mask-RCNN [12] to segment RGB images and detect objects and obstacles, creating 3D point clouds from RGB-D data and extracting essential high-level information like key object states. The Memory module manages short-term and long-" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.409, + 0.907, + 0.605 + ], + "angle": 0, + "content": "term memory, with short-term memory capturing immediate details such as semantic maps and agent states, and long-term memory storing persistent information like task descriptions. In the Communication module, CoELA retrieves relevant information from memory, transforming it into template-based textual descriptions for agents to aid in action selection. However, each agent plans independently, with no unified action coordination. In the Planning and Execution modules, CoELA retrieves agent-related information from memory, generates potential actions for the LLM to consider, and applies chain-of-thought [38] reasoning for more in-depth decision-making. Each agent then executes its chosen action accordingly." + }, + { + "type": "title", + "bbox": [ + 0.512, + 0.623, + 0.902, + 0.641 + ], + "angle": 0, + "content": "4. Collaborative Tree Search for Multi-Agents" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.649, + 0.693, + 0.665 + ], + "angle": 0, + "content": "4.1. Overview of CoTS" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.673, + 0.907, + 0.824 + ], + "angle": 0, + "content": "Now we introduce an overview of our Collaborative Tree Search (CoTS), which is designed for multi-agent cooperation in embodied environments. As shown in Fig. 2, CoTS contains four key modules: 1) Perception Module, 2) Memory Module, 3) Collaborative Planning Module, and 4) Plan Parsing and Execution Module. This work primarily enhances CoELA in its planning by designing the Collaborative Planning Module. The other three modules are largely borrowed from CoELA and introduced in Sec. 3. Here we briefly introduce our enhancements to CoEAL below." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Perception Module. Unlike CoELA, we introduce distance as a key perceptual factor, leveraging fixed spatial layouts in real-world settings to enhance agents' room-based understanding and improve planning efficiency by using LLM to access the moving distance of agents." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.958 + ], + "angle": 0, + "content": "29515" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.104, + 0.094, + 0.494, + 0.106 + ], + "angle": 0, + "content": "Task: transport 2 oranges, 3 apples, 1 banana, 3 breads, 1 burger to the bed" + }, + { + "type": "text", + "bbox": [ + 0.104, + 0.11, + 0.558, + 0.122 + ], + "angle": 0, + "content": "The Process in Each Node: Alice makes plans, Bob provides suggestions and rewards." + }, + { + "type": "text", + "bbox": [ + 0.104, + 0.128, + 0.493, + 0.139 + ], + "angle": 0, + "content": "Plan from Alice: Step 1: Alice grasps apple. Bob explore his current room." + }, + { + "type": "text", + "bbox": [ + 0.104, + 0.141, + 0.557, + 0.151 + ], + "angle": 0, + "content": "Step 2: Alice explore the Living room, if they find any items, transport them to the bed." + }, + { + "type": "text", + "bbox": [ + 0.104, + 0.153, + 0.601, + 0.164 + ], + "angle": 0, + "content": "Message to Bob: Hi Bob, I am in Livingroom where I found an apple. Do you have suggestions." + }, + { + "type": "text", + "bbox": [ + 0.103, + 0.168, + 0.535, + 0.191 + ], + "angle": 0, + "content": "Message to Alice: Hi Alice, I am in Livingroom and found bread. As you said, I am just about to explore my current room. Reward: 0.3" + }, + { + "type": "text", + "bbox": [ + 0.104, + 0.2, + 0.516, + 0.211 + ], + "angle": 0, + "content": "Plan from Alice: Step 1: Alice moves to Bedroom. Bob grasps bread and apple." + }, + { + "type": "text", + "bbox": [ + 0.103, + 0.212, + 0.414, + 0.223 + ], + "angle": 0, + "content": "Step 2: Alice explores Bedroom. Bob explores Living room." + }, + { + "type": "text", + "bbox": [ + 0.103, + 0.224, + 0.425, + 0.235 + ], + "angle": 0, + "content": "Step 3: If any target item are found, transport them to the bed." + }, + { + "type": "text", + "bbox": [ + 0.103, + 0.236, + 0.576, + 0.259 + ], + "angle": 0, + "content": "Message to Bob: Hi Bob, Please grasp bread and apple and I will now shift to the Bedroom because we should not explore the same room." + }, + { + "type": "text", + "bbox": [ + 0.103, + 0.26, + 0.599, + 0.27 + ], + "angle": 0, + "content": "Message to Alice: Hi Alice, but I think you could head to kitchen, it is the nearest. Reward: 0.6" + }, + { + "type": "text", + "bbox": [ + 0.104, + 0.278, + 0.467, + 0.29 + ], + "angle": 0, + "content": "Plan from Alice: Step 1: Alice grasps the apple. Bob grasps the bread." + }, + { + "type": "text", + "bbox": [ + 0.103, + 0.291, + 0.408, + 0.302 + ], + "angle": 0, + "content": "Step 2: Alice moves to Kitchen. Bob explores Living room." + }, + { + "type": "text", + "bbox": [ + 0.103, + 0.303, + 0.534, + 0.314 + ], + "angle": 0, + "content": "Step 3: Alice explore the Kitchen, if they find any items, transport them to the bed." + }, + { + "type": "text", + "bbox": [ + 0.103, + 0.316, + 0.503, + 0.327 + ], + "angle": 0, + "content": "Message to Bob: Hi Bob, I agree with you. I will grasp the apple before I mo" + }, + { + "type": "text", + "bbox": [ + 0.103, + 0.328, + 0.545, + 0.34 + ], + "angle": 0, + "content": "Kitchen while you continue in Livingroom. Please keep me updated on your progress" + }, + { + "type": "text", + "bbox": [ + 0.103, + 0.342, + 0.519, + 0.354 + ], + "angle": 0, + "content": "Message to Alice: Hi, Alice, your plan looks good. Let's get moving. Reward: 1" + }, + { + "type": "image", + "bbox": [ + 0.613, + 0.104, + 0.902, + 0.34 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.363, + 0.907, + 0.42 + ], + "angle": 0, + "content": "Figure 3. Overview of the tree search in CoTS. Each node in the tree contains content generated by Alice and Bob. Alice generates the collaborative plans and messages to Bob, while Bob generates messages to Alice and rewards for the plans. These rewards are used for backpropagation to guide the tree search toward more promising collaborative plans. The figure also shows examples of three nodes, illustrating the stepwise optimization of collaborative planning and the dialogue process between agents." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.433, + 0.483, + 0.569 + ], + "angle": 0, + "content": "Memory Module. Compared with CoELA, CoTS introduces \"extended memory\" (Fig. 2) to manage collaboration plans—information that needs to be retained for a moderate period and updated periodically. This is because collaboration plans must balance retention for a certain period and timely updates when they become unsuitable for the current embodied environment. This extended memory mimics the medium-term storage of the human memory system, bridging short-term recall and long-term consolidation." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.57, + 0.483, + 0.719 + ], + "angle": 0, + "content": "Collaborative Planning Module is the core of CoTS, where MCTS is modified and employed to develop long-term strategic collaboration plans. Inspired by human problem-solving through iterative discussion and cooperation, it avoids error accumulation caused by a single path approach. This module represents the first integration of embodied agent collaboration with MCTS, enhancing agents' capability to tackle complex, long-term tasks through tree-based search and backtracking. Details of this module are provided in Sec. 4.2." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.483, + 0.765 + ], + "angle": 0, + "content": "Plan Parsing and Execution Module translates collaborative plans into executable actions, enabling agents to complete tasks in a coordinated manner. See details in Sec. 4.3." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.766, + 0.483, + 0.841 + ], + "angle": 0, + "content": "Below we will delve deeper into two core modules, Collaborative Planning Module and Plan Parsing and Execution Module. Additionally, we will highlight CoTS's integration of spatial awareness and extended memory to enhance overall functionality and efficiency." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.849, + 0.369, + 0.865 + ], + "angle": 0, + "content": "4.2. Collaborative Planning Module" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Our Collaborative Planning Module, central to our CoTS framework, is designed to search for long-term, efficient" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.433, + 0.907, + 0.51 + ], + "angle": 0, + "content": "and coherent collaborative plans for multi-agent tasks in dynamic environments. It comprises two main sub-modules: a collaborative tree search module for generating collaborative plans (Sec. 4.2.1) and a plan evaluation module to analyze when to update the plan (Sec. 4.2.2)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.519, + 0.895, + 0.533 + ], + "angle": 0, + "content": "4.2.1. Monte Carlo Tree Search for Collaborative Plan" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.538, + 0.907, + 0.854 + ], + "angle": 0, + "content": "We adapt MCTS to convert multi-agent collaborative planning into a tree search, identifying an optimal pathway from multiple reasoning paths for cohesive strategies. As shown in Fig. 3, each node in this tree stores a proposed plan, agent feedback, and a performance reward. In this setup, one agent (like Alice) functions as the \"planner\" to propose plans, while other agents (like Bob) act as \"critics\". Alice drafts several initial collaborative plans or extends existing plan based on task goals and current task progress. Then Bob evaluates them, considering factors like task allocation and movement efficiency, to generate a reward for the current plan. Once feedback is provided, each plan's reward is propagated back through the collaboration tree, strengthening successful strategies and adjusting future planning pathways. Accordingly, they will select the most efficient collaborative plan for further reflections and refinements. As shown in Fig. 3, this process, including plan extension (drafting), evaluation, backpropagation, and selection, will repeat until agents explore all potential collaborative plans and select the most promising one. In the following, for these four steps, we will elaborate on each of them." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Selection Similar to traditional MCTS [4, 15], CoTS starts from the root node and traverses down the tree by selecting nodes with the highest UCT values. This process is repeated" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "29516" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.138 + ], + "angle": 0, + "content": "until a leaf node is reached, and is illustrated in Fig. 3 where we mark the selected nodes with gold cups. For UCT of each node, it is computed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.145, + 0.146, + 0.483, + 0.166 + ], + "angle": 0, + "content": "\\[\n\\mathrm {U C T} (s _ {i}) = V (s _ {i}) + w \\sqrt {\\ln N (p _ {i}) / N (s _ {i})}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.173, + 0.483, + 0.269 + ], + "angle": 0, + "content": "Where \\( N(s_{i}) \\) is visiting number of node \\( s_i \\), \\( V(s_{i}) \\) is the reward of the subtree rooted at \\( s_i \\), \\( w \\) is the exploration coefficient, and \\( p_i \\) is the parent node of \\( s_i \\). After an episode, backpropagation updates each node's \\( V(s_{i}) \\) along the path: \\( V(s_{i}) = \\frac{V(s_{i - 1})\\cdot(N(s_{i}) - 1) + r}{N(s_{i})} \\), where \\( V(s_{i - 1}) \\) is the previous reward, and \\( r \\) is the reward used to update each \\( V(s_{i}) \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.269, + 0.483, + 0.388 + ], + "angle": 0, + "content": "Expansion. Once a node is selected, we expand it by sampling \\( n \\) new discussion paths, each building on the plan (from \"planner\", Alice) and suggestions (from \"critic\", Bob) in the parent node. This generates \\( n \\) new child nodes, representing potential planning routes. The tree, along with its nodes, is stored in an external long-term memory to retain valuable information and facilitate consistent planning across episodes." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.389, + 0.483, + 0.63 + ], + "angle": 0, + "content": "Evaluation. Each newly created child node receives a reward \\( r \\) to quantify the quality of the collaboration plan and to guide the search algorithm towards the most promising areas of the tree. Here CoTS incorporates a non-simulated LLM-based reward function to estimate the reward \\( r \\) of a plan in a node. In traditional MCTS, reward estimation requires agents to execute their actions for receiving environment feedback, like completing a subtask, and thus is high-cost. So CoTS introduces an LLM-based reward function based on two components: (1) an allocation reward assessing coherence and efficiency, and (2) a cost reward evaluating the spatial movement cost of agents. To ensure consistency, we standardize the evaluation criteria to provide clear guidelines to minimize ambiguity. Each criterion is scored on a scale of 1 to 5, with the final reward normalized between 0 and 1. See examples in Appendix B.1." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.631, + 0.483, + 0.786 + ], + "angle": 0, + "content": "Backpropagation. Each evaluation reward is propagated back along the path to update node rewards, with formulas ensuring that parent nodes integrate the feedback, guiding future selections. Specifically, the back-propagation updates node rewards \\( N(s_{i}) = N(s_{i - 1}) + 1 \\) and \\( V(s_{i}) = \\frac{V(s_{i - 1})N(s_{i - 1}) + r}{N(s_i)} \\) for each node \\( s_0,s_1,\\ldots ,s_p \\) along the path from the root (initial state \\( s_0 \\)) to leaf (terminal state \\( s_p \\)). Accordingly these estimated rewards and visiting number are used in \"Expansion\" step for node selection, directing the search toward the most promising areas of the tree." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.792, + 0.462, + 0.807 + ], + "angle": 0, + "content": "4.2.2. Collaborative Plan Evaluation and Adaptation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.483, + 0.902 + ], + "angle": 0, + "content": "In this section, we will explain why we introduced the planning evaluation module (a sub-module in the collaborative planning module, as shown in Fig. 2) and its function. The collaborative plan provides high-level embodied agent action guidance, but as the task is executed and agents act, the embodied environment gradually changes (e.g. the agents'" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.136 + ], + "angle": 0, + "content": "positions, the agents' progress, and the current status, etc.). Therefore, the collaborative plan needs to be updated based on the latest environmental information." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.137, + 0.907, + 0.332 + ], + "angle": 0, + "content": "At this point, two scenarios arise: If the plan is updated too slowly or not updated at all, the current plan becomes outdated and cannot effectively guide the embodied agents in executing tasks. However, in a multi-agent system, instantly updating the plan with every progress can be inefficient and counterproductive. It may lead to excessive queries to the LLM and environments. Furthermore, frequent plan revisions may introduce conflicting strategies due to minor environmental shifts, potentially disrupting agent actions and lowering overall efficiency. This is illustrated in Fig. 4 (c): if the updated plan conflicts with the action that Bob is currently performing, it may interrupt Bob's action which even almost finishes the subtask." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.334, + 0.909, + 0.515 + ], + "angle": 0, + "content": "To address this, we introduce a plan evaluation module within CoTS to continuously evaluate the viability of the current collaborative plan. When an agent makes progress, the module determines if the existing plan remains effective based on the status of task execution. If the evaluation metric meets a defined threshold, indicating that the plan is still viable, it is retained. Otherwise, CoTS will call the collaborative planning module to re-search and update the plan. By focusing on the feasibility and long-term stability of the plan rather than continuous recalibration, this module supports efficient, coherent collaboration, enabling agents to complete tasks reliably in a dynamic environment." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.524, + 0.828, + 0.54 + ], + "angle": 0, + "content": "4.3. Plan Parsing and Execution Module" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.545, + 0.907, + 0.697 + ], + "angle": 0, + "content": "To transform the collaborative plan generated in Sec. 4.2 into executable actions, CoTS incorporates an action parsing module and a plan decomposition module. Action parsing module bridges high-level planning with real-world actions, generating a set of feasible actions for each agent based on current memory. For example, if an agent's hands are free, it will list objects that can be grabbed; if the agent is in an unknown room, it will list movement options. This module ensures that all sub-plans are formatted into natural language and returned for subsequent process." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.697, + 0.906, + 0.786 + ], + "angle": 0, + "content": "Plan decomposition module understands and breaks down high-level tasks from the collaboration plan into specific and prioritized actions. The agent interprets the task arrangement, selects actions from the parsed options that best advance the collaborative plan's goals, and finally translates each chosen action into real-world execution." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.801, + 0.646, + 0.817 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.906, + 0.902 + ], + "angle": 0, + "content": "**Benchmarks. Following CoELA [43], we evaluate the performance of CoTS in two highly challenging embodied-world multi-agent environments: (1) TDW-MAT (ThreeD World Multi-Agent Transport) [8] extends the ThreeD World Transport Challenge by simulating collaborative" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "29517" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.908, + 0.133 + ], + "angle": 0, + "content": "Table 1. Quantitative results on TDW-MAT. Consistent with CoELA [43], TDW-MAT has two environment settings: with oracle perception and without oracle perception. The results marked with * are from CoELA [43]. MAT uses central observation and oracle perception. \\(91_{+9.64\\%}\\) means the relative improvement when comparing with CoELA is \\(9.64\\%\\)." + }, + { + "type": "table", + "bbox": [ + 0.134, + 0.138, + 0.865, + 0.292 + ], + "angle": 0, + "content": "
Transport Rate (%)Base AgentRL AgentLLM Agent
RHP*CoRHP*MAT*GPT 3.5GPT 4
CoELARoCoCaPoCoTS(ours)CoELA*RoCoCaPoCoTS(ours)
TDW-MAT w/ Oracle Perception
Stuff(↑)4974177380+9.68%84+15.07%81+10.96%8385+2.41%87+4.82%91+9.64%
Food(↑)5276137274+2.78%85+18.06%83+15.28%8789+2.30%90+3.45%94+8.05%
Average(↑)5075157377+5.48%84+15.07%82+12.33%8587+2.35%89+4.71%93+9.41%
TDW-MAT w/o Oracle Perception
Stuff(↑)3654-3947+20.51%45+15.38%50+28.21%6164+4.92%64+4.92%77+26.22%
Food(↑)4967-6764-4.48%70+4.48%67+0.00%8280-2.44%85+3.65%88+7.32%
Average(↑)4361-5255+5.77%57+9.62%59+13.46%7172+1.41%74+4.23%83+16.90%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.314, + 0.483, + 0.342 + ], + "angle": 0, + "content": "Table 2. Quantitative results on C-WAH. LLM-based agents are driven by GPT 4. Results marked with * are from CoELA [43]." + }, + { + "type": "table", + "bbox": [ + 0.113, + 0.347, + 0.459, + 0.455 + ], + "angle": 0, + "content": "
Average StepVisual Obs(↓)Symbolic Obs(↓)
Base AgentMHP*141111
CoMHP*10375
LLM AgentCoELA*9257
RoCo89+3.26%57+0.00%
CaPo83+9.78%51+10.53%
CoTS(ours)81+11.96%49+14.04%
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.473, + 0.483, + 0.594 + ], + "angle": 0, + "content": "transportation tasks in a physics-based household environment. We select 24 tasks, split evenly between food and stuff transportation. (2) C-WAH (Communicative Watch-And-Help) [29] extends the Watch-And-Help challenge within the VirtualHome-Social platform [28]. It features five common household scenarios. Our test set includes 10 tasks, with two tasks per scenario. The objective is to complete all sub-goals, prioritizing faster completion." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.597, + 0.483, + 0.688 + ], + "angle": 0, + "content": "Evaluation Metrics. On TDW-MAT, transport rate is evaluated by the percentage of target objects successfully delivered to their designated locations within a 3,000-frame timeframe. For C-WAH, efficiency is measured by the number of time steps needed to complete all assigned tasks, where fewer steps indicate higher efficiency." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Baseline Methods. The baselines include three types of embodied agents: (1) Base Agent: RHP [8] uses heuristic rules for navigation and \\(\\mathbf{A}^*\\) for path planning. MHP [18] employs Monte Carlo Tree Search (MCTS) for high-level planning coupled with regression planning for task execution (2) RL-Based Agent: MAT [39] is a multi-agent reinforcement learning framework, featuring a centralized decision transformer that synthesizes actions from shared observations. (3) LLM-Based Agent: CoELA [43] is a leading collaboration framework for embodied agents that integrates the reasoning capabilities of LLMs. It facilitates communication and interaction between agents through natural language. RoCo [24] and CaPo [22] are multi-agent collaboration methods that enhance collaboration by gener" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.316, + 0.907, + 0.41 + ], + "angle": 0, + "content": "ating coordinated plans through structured agent dialogues. Setup. We conduct all experiments on an Intel Core i9- 10980XE CPU and an NVIDIA GeForce RTX 3090 GPU. We utilize both GPT-4 and GPT-3.5-Turbo [1], configuring them with a temperature setting of 0.7 and top-1 sampling to produce outputs up to 512 tokens." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.421, + 0.655, + 0.437 + ], + "angle": 0, + "content": "5.1. Main Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.445, + 0.907, + 0.61 + ], + "angle": 0, + "content": "CoTS Outperforms the Strong Baselines. Following the experimental frameworks of CoELA [43] and RoCo [24], we configure two-agent cooperation, excluding RHP. As demonstrated in Table 1, CoTS achieves best performance on the TDW-MAT task. Specifically, when powered by GPT-4 within the Oracle Perception environment, CoTS surpasses CoELA, achieving a \\(9.41\\%\\) average relative improvement in task completion, and a \\(16.9\\%\\) improvement in the absence of Oracle Perception. The primary strength of CoTS is its ability to generate superior long-term strategies through collaborative tree search." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.613, + 0.909, + 0.884 + ], + "angle": 0, + "content": "Conversely, CoELA focuses on short-term actions derived from immediate information, which limits its strategic depth and adaptability to future scenarios and environmental shifts. RoCo and CaPo are similar. Although they are capable of developing plans, they rely on a single-branch discussion method, which may result in suboptimal or inaccurate strategies. Moreover, Their frequent plan revisions can interfere with agent coordination. Therefore, while RoCo and CaPo surpass CoELA, they are inferior to CoTS. When driven by GPT-3.5, CaPo slightly outperforms CoTS on some metrics. The main reason for this is that GPT-3.5 has weaker evaluation and scoring capabilities. However, when powered by GPT-4, CoTS significantly outperforms CaPo. Our findings underscore that multi-agent collaboration generally surpasses single-agent strategies, exemplified by RHP. The performance of reinforcement learning-based MAT is poor, mainly because TDW-MAT is too difficult for it with long time domain and sparse reward signals." + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.886, + 0.905, + 0.902 + ], + "angle": 0, + "content": "The performance of LLM-based agents primarily de" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "29518" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.103, + 0.094, + 0.335, + 0.134 + ], + "angle": 0, + "content": "(a) CoELA leads to short-sighted action plans that ignore precious action steps" + }, + { + "type": "text", + "bbox": [ + 0.104, + 0.138, + 0.323, + 0.178 + ], + "angle": 0, + "content": "Background: Alice is in living room holding a container. Bob finds two target objects in living room \\(< 1000>\\)." + }, + { + "type": "image", + "bbox": [ + 0.1, + 0.18, + 0.348, + 0.302 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.303, + 0.332, + 0.339 + ], + "angle": 0, + "content": "CoELA's plan : Alice goes to the living room to help Bob, because Alice can take these objects." + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.34, + 0.329, + 0.39 + ], + "angle": 0, + "content": "Our plan: Alice explores the current room, find objects, then transport them to the bedroom. Then go to the living room \\(< 1000>\\), no actions are wasted." + }, + { + "type": "title", + "bbox": [ + 0.379, + 0.093, + 0.607, + 0.134 + ], + "angle": 0, + "content": "(b) RoCo's single-chain dialogue leads to suboptimal collaboration plan" + }, + { + "type": "text", + "bbox": [ + 0.377, + 0.139, + 0.603, + 0.178 + ], + "angle": 0, + "content": "Background: Alice and Bob appear in the same living room and now need to explore some rooms to find objects." + }, + { + "type": "image", + "bbox": [ + 0.375, + 0.178, + 0.627, + 0.302 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.376, + 0.303, + 0.61, + 0.328 + ], + "angle": 0, + "content": "RoCo's plan : Alice explore the current room. Bob explore the current room." + }, + { + "type": "text", + "bbox": [ + 0.374, + 0.339, + 0.624, + 0.392 + ], + "angle": 0, + "content": "Our plan: Alice goes to the nearest room, Bedroom \\(< 2000>\\), while Bob explores the current room. Exploring rooms separately would improve efficiency." + }, + { + "type": "title", + "bbox": [ + 0.66, + 0.095, + 0.88, + 0.135 + ], + "angle": 0, + "content": "(c) Frequent plan updates may hinder completion of current actions" + }, + { + "type": "text", + "bbox": [ + 0.659, + 0.138, + 0.872, + 0.177 + ], + "angle": 0, + "content": "Background: Bob was transporting objects to the bed, Alice found a bread in the living room<6000>" + }, + { + "type": "image", + "bbox": [ + 0.654, + 0.177, + 0.897, + 0.302 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.655, + 0.302, + 0.894, + 0.339 + ], + "angle": 0, + "content": "RoCo's Plan: Bob goes to the living room \\(< 6000>\\) to help Alice grasp the bread." + }, + { + "type": "text", + "bbox": [ + 0.655, + 0.339, + 0.892, + 0.391 + ], + "angle": 0, + "content": "Our plan: The current plan is still valid. Bob is close to the bedroom. He can first deliver the objects in his hand to the bed, and then help Alice." + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.4, + 0.907, + 0.428 + ], + "angle": 0, + "content": "Figure 4. Some interesting example cases demonstrate that our CoTS achieves better collaboration than RoCo and CoELA. These scenes are taken from the experiments in TDW-MAT." + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.443, + 0.273, + 0.543 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.098, + 0.545, + 0.268, + 0.594 + ], + "angle": 0, + "content": "There seems to be no plan. Both Alice and Bob are not holding anything. Bob in bedroom and has found an container, while Alice is in the. A collaboration plan needs to be updated. Reward: 1" + }, + { + "type": "image", + "bbox": [ + 0.284, + 0.443, + 0.474, + 0.544 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.283, + 0.545, + 0.475, + 0.595 + ], + "angle": 0, + "content": "Alice have found an apple in the fridge and Bob is holding a cupcake. Current progress is consistent with the collaborative plan and future actions are outlined in the plan. There is no need to update the plan. Reward: 4" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.6, + 0.483, + 0.656 + ], + "angle": 0, + "content": "Figure 5. Plan Evaluation and Update. The plan evaluation module gives a lower reward when it determines that the current plan needs to be updated, and a higher reward when it believes the current plan is still valid. The reward range is from 1 to 5." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.684, + 0.483, + 0.837 + ], + "angle": 0, + "content": "pends on their reasoning abilities and the depth of embedded world knowledge. Agents powered by GPT-4, as demonstrated by CoTS, significantly surpass those based on GPT-3.5, highlighting the advancements in LLM technologies and their impact on agent proficiency. As shown in Table 2, CoTS achieves superior efficiency, improving by \\(11.96\\%\\) and \\(14.04\\%\\) over CoELA in Visual Obs and Symbolic Obs settings, respectively. These results underscore CoTS's robust adaptability across various embodied environments and its potential for real-world applications." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.848, + 0.283, + 0.864 + ], + "angle": 0, + "content": "5.2. Qualitative Analysis" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.901 + ], + "angle": 0, + "content": "CoTS Can Develop Long-term Strategic Plans. Unlike CoELA, which is adept at responding to immediate tasks" + }, + { + "type": "image", + "bbox": [ + 0.56, + 0.445, + 0.857, + 0.57 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.574, + 0.907, + 0.603 + ], + "angle": 0, + "content": "Figure 6. Comparison of transportation rates for GPT-4-based CoTS and CoELA at different time steps on TDW-MAT." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.629, + 0.907, + 0.75 + ], + "angle": 0, + "content": "but falters in sustained coordination, CoTS excels in long-term strategic planning. As illustrated in Fig. 4 (a), when Bob discovers a new target that requires assistance, Alice, under CoTS's direction, continues to thoroughly explore her current room before moving the item to the bed. This systematic approach not only supports Bob but also adheres to an overarching collaborative strategy, optimizing both immediate and future task alignments." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.909, + 0.902 + ], + "angle": 0, + "content": "CoTS Generates More Optimal Collaboration Plans. Unlike RoCo to generate plans sequentially through a single-chain dialogue, CoTS makes more optimized collaboration plans. Due to the randomness of LLMs and the embodied environment's complexity, RoCo's plans may sometimes be suboptimal. For example, in Fig. 4 (b), RoCo instructs both Alice and Bob to explore their respective rooms. However, they are currently in the same room, making joint exploration redundant. CoTS effectively identifies this and assigns Alice to explore the nearest room instead." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "29519" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.102, + 0.096, + 0.476, + 0.235 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.242, + 0.483, + 0.286 + ], + "angle": 0, + "content": "Figure 7. Comparison of transportation rates with different branching factors. More branching factors means more new nodes are generated each time in the expansion phase." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.296, + 0.483, + 0.447 + ], + "angle": 0, + "content": "CoTS Ensures Smooth Completion of Agent Actions. In RoCo, new progress often prompt discussions and lead to new plans that can override existing ones. As illustrated in Fig. 4 (c), Bob, tasked with transporting items to the bed, must halt his activity to help Alice following a revised plan. Conversely, CoTS incorporates a plan evaluation module that validates the feasibility of plans before any adjustments are made. In Fig. 4 (c), CoTS allows Bob to finish transporting before next action. This approach avoids chaotic agent actions and ensures timely plan updates when necessary." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.448, + 0.483, + 0.614 + ], + "angle": 0, + "content": "CoTS Dynamically Manages Collaborative Plans. As shown in Fig. 5, the assessment module continuously evaluates the viability of existing plans. It assigns a lower reward (e.g. 1) if collaboration necessitates devising a new plan or modifying an existing one, prompting agents to adapt their strategies accordingly. Conversely, when the current progress aligns with planned tasks and their distribution among agents is deemed efficient, the module rewards this alignment with a higher value (e.g. 4). This incentivizes agents to persist with the established collaborative plan, avoiding unnecessary replanning." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.623, + 0.433, + 0.639 + ], + "angle": 0, + "content": "5.3. Ablation Study and Additional Analysis" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.645, + 0.483, + 0.825 + ], + "angle": 0, + "content": "The Core Modules of CoTS Have Important Impacts. The Table 3 shows the ablation results of CoTS. To examine the impact of tree search, we modify it to a single-branch structure, where agents can only communicate along one path. The experimental results indicate a \\(5.38\\%\\) relative decrease in the average transportation rate, mainly because the limited communication does not consistently yield correct conclusions and diverges from typical human discussion patterns. Furthermore, eliminating the reward function and plan update evaluation components from CoTS leads to variable performance declines, highlighting the critical roles these elements play in the system's efficacy." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Analysis of Agent Collaboration Progress at Different Time Steps. We analyze the transportation rates of CoELA and CoTS on TDW-MAT at various time steps, as depicted in Fig. 6. An interesting observation emerges at 500 steps, where CoTS exhibits a lower performance compared to" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.09, + 0.905, + 0.118 + ], + "angle": 0, + "content": "Table 3. Effects of the reward function, collaboration tree search, and plan evaluation module in GPT-4-driven CoTS on TDW-MAT." + }, + { + "type": "table", + "bbox": [ + 0.537, + 0.123, + 0.885, + 0.209 + ], + "angle": 0, + "content": "
Transport RateStuff(↑)Food(↑)Average(↑)
CoTS w/o Reward87-4.40%90-4.26%89-4.30%
CoTS w/o MCTS Plan85-6.59%91-3.19%88-5.38%
CoTS w/o E&U88-3.30%92-2.13%90-3.23%
CoTS919493
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.233, + 0.905, + 0.323 + ], + "angle": 0, + "content": "CoELA, attributed to CoTS's strategy of prioritizing item collection early to enhance transport capacity. However, this initial disadvantage is counterbalanced by a significant improvement at 2000 steps, where CoTS markedly surpasses CoELA. This shift underscores CoTS's effectiveness in optimizing agent collaboration for task execution." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.324, + 0.905, + 0.429 + ], + "angle": 0, + "content": "Impact of Different Branching Factors on Results. During the expansion phase of MCTS, we define the number of new candidate nodes generated at each step as the \"branching factors.\" As shown in Fig. 7, increasing the number of branching factors expands the selection space, leading to a better plan. Adjusting the branching factor allows for a balance between performance and efficiency." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.43, + 0.905, + 0.475 + ], + "angle": 0, + "content": "Besides, the human-agent collaboration and the cost of CoTS are discussed in Appendix A.1 and Appendix A.2, respectively. CoTS also demonstrates significant advantages." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.489, + 0.64, + 0.505 + ], + "angle": 0, + "content": "6. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.514, + 0.905, + 0.665 + ], + "angle": 0, + "content": "This paper introduces CoTS, a collaborative framework that leverages large language models with a modified Monte Carlo tree search to enhance multi-agent task planning in complex and dynamic environments. By enabling agents to evaluate multiple potential plans and refine strategies based on progress, CoTS significantly advances over prior methods like CoELA and RoCo. The addition of a plan feasibility module ensures smooth task execution by adapting plans only when necessary. Experimental results on multi-objective tasks confirm CoTS's superior performance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.666, + 0.905, + 0.756 + ], + "angle": 0, + "content": "Limitation. CoTS relies on LLMs' reasoning abilities. We find that GPT-4 surpasses GPT-3.5-turbo in instruction-following and response quality, while GPT-3.5-turbo can be verbose and occasionally inaccurate. Despite this, even for GPT-3.5, CoTS improves agent collaboration through collaborative planning search in the collaboration tree." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.77, + 0.692, + 0.788 + ], + "angle": 0, + "content": "7. Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.905, + 0.901 + ], + "angle": 0, + "content": "This work is supported by National Natural Science Foundation of China (U2133202, 52305570), China Postdoctoral Science Foundation (2022M720955), and Outstanding Doctoral Dissertation Funding Project of Heilongjiang Province (LJYXL2022-011). Lizheng Zu also acknowledges the support of the China Scholarship Council program (202406120111)." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "29520" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.184 + ], + "angle": 0, + "content": "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.185, + 0.484, + 0.282 + ], + "angle": 0, + "content": "[2] David Brandfonbrener, Simon Henniger, Sibi Raja, Tarun Prasad, Chloe R Loughridge, Federico Cassano, Sabrina Ruixin Hu, Jianang Yang, William E Byrd, Robert Zinkov, et al. Vermcts: Synthesizing multi-step programs using a verifier, a large language model, and tree search. In The 4th Workshop on Mathematical Reasoning and AI at NeurIPS'24, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.283, + 0.482, + 0.311 + ], + "angle": 0, + "content": "[3] Tom B Brown. Language models are few-shot learners. arXiv preprint arXiv:2005.14165, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.313, + 0.483, + 0.395 + ], + "angle": 0, + "content": "[4] Cameron B Browne, Edward Powley, Daniel Whitehouse, Simon M Lucas, Peter I Cowling, Philipp Rohlfshagen, Stephen Tavener, Diego Perez, Spyridon Samothrakis, and Simon Colton. A survey of monte carlo tree search methods. IEEE Transactions on Computational Intelligence and AI in games, 4(1):1-43, 2012. 4, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.397, + 0.483, + 0.451 + ], + "angle": 0, + "content": "[5] Abhinav Dahiya, Alexander M Aroyo, Kerstin Dautenhahn, and Stephen L Smith. A survey of multi-agent human-robot interaction systems. Robotics and Autonomous Systems, 161: 104335, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.453, + 0.483, + 0.508 + ], + "angle": 0, + "content": "[6] Norman Di Palo, Arunkumar Byravan, Leonard Hasenclever, Markus Wulfmeier, Nicolas Heess, and Martin Riedmiller. Towards a unified agent with foundation models. arXiv preprint arXiv:2307.09668, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.51, + 0.483, + 0.564 + ], + "angle": 0, + "content": "[7] Amal Feriani and Ekram Hossain. Single and multi-agent deep reinforcement learning for ai-enabled wireless networks: A tutorial. IEEE Communications Surveys & Tutorial, 23(2):1226-1252, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.566, + 0.483, + 0.675 + ], + "angle": 0, + "content": "[8] Chuang Gan, Siyuan Zhou, Jeremy Schwartz, Seth Alter, Abhishek Bhandwaldar, Dan Gutfreund, Daniel LK Yamins, James J DiCarlo, Josh McDermott, Antonio Torralba, et al. The threadworld transport challenge: A visually guided task-and-motion planning benchmark towards physically realistic embodied ai. In 2022 International conference on robotics and automation (ICRA), pages 8847-8854. IEEE, 2022. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.678, + 0.483, + 0.759 + ], + "angle": 0, + "content": "[9] Ran Gong, Qiuyuan Huang, Xiaojian Ma, Yusuke Noda, Zane Durante, Zilong Zheng, Demetri Terzopoulos, Li Fei-Fei, Jianfeng Gao, and Hoi Vo. Mindagent: Emergent gaming interaction. In Findings of the Association for Computational Linguistics: NAACL 2024, pages 3154–3183, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.762, + 0.483, + 0.829 + ], + "angle": 0, + "content": "[10] Taicheng Guo, Xiuying Chen, Yaqi Wang, Ruidi Chang, Shichao Pei, Nitesh V Chawla, Olaf Wiest, and Xiangliang Zhang. Large language model based multi-agents: A survey of progress and challenges. arXiv preprint arXiv:2402.01680, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.831, + 0.483, + 0.899 + ], + "angle": 0, + "content": "[11] Xudong Guo, Kaixuan Huang, Jiale Liu, Wenhui Fan, Natalia Vélez, Qingyun Wu, Huazheng Wang, Thomas L Griffiths, and Mengdi Wang. Embodied llm agents learn to cooperate in organized teams. arXiv preprint arXiv:2403.12482, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.135 + ], + "angle": 0, + "content": "[12] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.136, + 0.906, + 0.204 + ], + "angle": 0, + "content": "[13] Sirui Hong, Xiawu Zheng, Jonathan Chen, Yuheng Cheng, Jinlin Wang, Ceyao Zhang, Zili Wang, Steven Ka Shing Yau, Zijuan Lin, Liyang Zhou, et al. Metagpt: Meta programming for multi-agent collaborative framework. arXiv preprint arXiv:2308.00352, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.205, + 0.906, + 0.272 + ], + "angle": 0, + "content": "[14] Wenlong Huang, Pieter Abbeel, Deepak Pathak, and Igor Mordatch. Language models as zero-shot planners: Extracting actionable knowledge for embodied agents. In International conference on machine learning, pages 9118-9147. PMLR, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.274, + 0.905, + 0.315 + ], + "angle": 0, + "content": "[15] Steven James, George Konidaris, and Benjamin Rosman. An analysis of monte carlo tree search. In Proceedings of the AAAI Conference on Artificial Intelligence, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.316, + 0.905, + 0.371 + ], + "angle": 0, + "content": "[16] Feibo Jiang, Yubo Peng, Li Dong, Kezhi Wang, Kun Yang, Cunhua Pan, Dusit Niyato, and Octavia A Dobre. Large language model enhanced multi-agent systems for 6g communications. IEEE Wireless Communications, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.372, + 0.905, + 0.44 + ], + "angle": 0, + "content": "[17] Bowen Jin, Chulin Xie, Jiawei Zhang, Kashob Kumar Roy, Yu Zhang, Zheng Li, Ruirui Li, Xianfeng Tang, Suhang Wang, Yu Meng, et al. Graph chain-of-thought: Augmenting large language models by reasoning on graphs. arXiv preprint arXiv:2404.07103, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.442, + 0.904, + 0.469 + ], + "angle": 0, + "content": "[18] Richard E Korf. Planning as search: A quantitative approach. Artificial intelligence, 33(1):65-88, 1987. 6, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.47, + 0.905, + 0.537 + ], + "angle": 0, + "content": "[19] Guohao Li, Hasan Hammoud, Hani Itani, Dmitrii Khizbullin, and Bernard Ghanem. Camel: Communicative agents for\" mind\" exploration of large language model society. Advances in Neural Information Processing Systems, 36: 51991-52008, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.539, + 0.905, + 0.565 + ], + "angle": 0, + "content": "[20] Xin Li. Positioning is all you need. arXiv preprint arXiv:2404.01183, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.567, + 0.905, + 0.622 + ], + "angle": 0, + "content": "[21] Iou-Jen Liu, Unnat Jain, Raymond A Yeh, and Alexander Schwing. Cooperative exploration for multi-agent deep reinforcement learning. In International conference on machine learning, pages 6826-6836. PMLR, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.623, + 0.905, + 0.691 + ], + "angle": 0, + "content": "[22] Jie Liu, Pan Zhou, Yingjun Du, Ah-Hwee Tan, Cees GM Snoek, Jan-Jakob Sonke, and Efstratios Gavves. Capo: Cooperative plan optimization for efficient embodied multiagent cooperation. arXiv preprint arXiv:2411.04679, 2024. 2, 6, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.692, + 0.905, + 0.762 + ], + "angle": 0, + "content": "[23] Yang Liu, Weixing Chen, Yongjie Bai, Jingzhou Luo, Xinshuai Song, Kaixuan Jiang, Zhida Li, Ganlong Zhao, Junyi Lin, Guanbin Li, et al. Aligning cyber space with physical world: A comprehensive survey on embodied ai. arXiv preprint arXiv:2407.06886, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.763, + 0.905, + 0.829 + ], + "angle": 0, + "content": "[24] Zhao Mandi, Shreeya Jain, and Shuran Song. Roco: Dialectic multi-robot collaboration with large language models. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 286–299. IEEE, 2024. 1, 2, 3, 6, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.831, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[25] Reiichiro Nakano, Jacob Hilton, Suchir Balaji, Jeff Wu, Long Ouyang, Christina Kim, Christopher Hesse, Shantanu Jain, Vineet Kosaraju, William Saunders, et al. Webgpt: Browser-assisted question-answering with human feedback. arXiv preprint arXiv:2112.09332, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "29521" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[26] Bo Ni and Markus J Buehler. Mechagents: Large language model multi-agent collaborations can solve mechanics problems, generate new data, and integrate knowledge. Extreme Mechanics Letters, 67:102131, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.15, + 0.482, + 0.233 + ], + "angle": 0, + "content": "[27] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. Advances in neural information processing systems, 35:27730-27744, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.235, + 0.482, + 0.303 + ], + "angle": 0, + "content": "[28] Xavier Puig, Kevin Ra, Marko Boben, Jiaman Li, Tingwu Wang, Sanja Fidler, and Antonio Torralba. Virtualhome: Simulating household activities via programs. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8494-8502, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.306, + 0.482, + 0.374 + ], + "angle": 0, + "content": "[29] Xavier Puig, Tianmin Shu, Shuang Li, Zilin Wang, Yuan-Hong Liao, Joshua B Tenenbaum, Sanja Fidler, and Antonio Torralba. Watch-and-help: A challenge for social perception and human-ai collaboration. arXiv preprint arXiv:2010.09890, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.377, + 0.482, + 0.445 + ], + "angle": 0, + "content": "[30] Xavier Puig, Eric Undersander, Andrew Szot, Mikael Dallaire Cote, Tsung-Yen Yang, Ruslan Partsey, Ruta Desai, Alexander William Clegg, Michal Hlavac, So Yeon Min, et al. Habitat 3.0: A co-habitat for humans, avatars and robots. arXiv preprint arXiv:2310.13724, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.448, + 0.482, + 0.53 + ], + "angle": 0, + "content": "[31] Chen Qian, Wei Liu, Hongzhang Liu, Nuo Chen, Yufan Dang, Jiahao Li, Cheng Yang, Weize Chen, Yusheng Su, Xin Cong, et al. Chatdev: Communicative agents for software development. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 15174-15186, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.533, + 0.482, + 0.601 + ], + "angle": 0, + "content": "[32] Timo Schick, Jane Dwivedi-Yu, Roberto Dessì, Roberta Raileanu, Maria Lomeli, Eric Hambro, Luke Zettlemoyer, Nicola Cancedda, and Thomas Scialom. Toolformer: Language models can teach themselves to use tools. Advances in Neural Information Processing Systems, 36, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.604, + 0.482, + 0.658 + ], + "angle": 0, + "content": "[33] Paul E Smaldino, Cody Moser, Alejandro Perez Velilla, and Mikkel Werling. Maintaining transient diversity is a general principle for improving collective problem solving. Perspectives on Psychological Science, 19(2):454-464, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.661, + 0.482, + 0.73 + ], + "angle": 0, + "content": "[34] Chan Hee Song, Jiaman Wu, Clayton Washington, Brian M Sadler, Wei-Lun Chao, and Yu Su. Llm-planner: Few-shot grounded planning for embodied agents with large language models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2998-3009, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.732, + 0.482, + 0.771 + ], + "angle": 0, + "content": "[35] Wiebe Van der Hoek and Michael Wooldridge. Multi-agent systems. Foundations of Artificial Intelligence, 3:887-928, 2008. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.775, + 0.482, + 0.857 + ], + "angle": 0, + "content": "[36] Tai Wang, Xiaohan Mao, Chenming Zhu, Runsen Xu, Ruiyuan Lyu, Peisen Li, Xiao Chen, Wenwei Zhang, Kai Chen, Tianfan Xue, et al. Embodiedscan: A holistic multimodal 3d perception suite towards embodied ai. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19757-19767, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.86, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[37] Jason Wei, Yi Tay, Rishi Bommasani, Colin Raffel, Barret Zoph, Sebastian Borgeaud, Dani Yogatama, Maarten Bosma, Denny Zhou, Donald Metzler, et al. Emergent abilities of" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.119 + ], + "angle": 0, + "content": "large language models. arXiv preprint arXiv:2206.07682, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[38] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824-24837, 2022. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.193, + 0.905, + 0.261 + ], + "angle": 0, + "content": "[39] Muning Wen, Jakub Kuba, Runji Lin, Weinan Zhang, Ying Wen, Jun Wang, and Yaodong Yang. Multi-agent reinforcement learning is a sequence modeling problem. Advances in Neural Information Processing Systems, 35:16509-16521, 2022. 1, 6, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.263, + 0.905, + 0.331 + ], + "angle": 0, + "content": "[40] Shunyu Yao, Dian Yu, Jeffrey Zhao, Izhak Shafran, Tom Griffiths, Yuan Cao, and Karthik Narasimhan. Tree of thoughts: Deliberate problem solving with large language models. Advances in Neural Information Processing Systems, 36, 2024. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.334, + 0.905, + 0.386 + ], + "angle": 0, + "content": "[41] Bangguo Yu, Hamidreza Kasaei, and Ming Cao. Co-navgpt: Multi-robot cooperative visual semantic navigation using large language models. arXiv preprint arXiv:2310.07937, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.39, + 0.905, + 0.458 + ], + "angle": 0, + "content": "[42] Di Zhang, Xiaoshui Huang, Dongzhan Zhou, Yuqiang Li, and Wanli Ouyang. Accessing gpt-4 level mathematical olympiad solutions via monte carlo tree self-refine with llama-3 8b: A technical report. arXiv preprint arXiv:2406.07394, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.461, + 0.905, + 0.529 + ], + "angle": 0, + "content": "[43] Hongxin Zhang, Weihua Du, Jiaming Shan, Qinhong Zhou, Yilun Du, Joshua B Tenenbaum, Tianmin Shu, and Chuang Gan. Building cooperative embodied agents modularly with large language models. arXiv preprint arXiv:2307.02485, 2023. 1, 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.531, + 0.905, + 0.598 + ], + "angle": 0, + "content": "[44] Zhonghan Zhao, Ke Ma, Wenhao Chai, Xuan Wang, Kewei Chen, Dongxu Guo, Yanting Zhang, Hongwei Wang, and Gaoang Wang. Do we really need a complex agent system? distill embodied agent into a single model. arXiv preprint arXiv:2404.04619, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.601, + 0.905, + 0.656 + ], + "angle": 0, + "content": "[45] Andy Zhou, Kai Yan, Michal Shlapentokh-Rothman, Haohan Wang, and Yu-Xiong Wang. Language agent tree search unifies reasoning acting and planning in language models. arXiv preprint arXiv:2310.04406, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.656 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "29522" + } + ] +] \ No newline at end of file diff --git a/2025/Collaborative Tree Search for Enhancing Embodied Multi-Agent Collaboration/509a3bd5-5db7-45e0-bb5b-ba942b10cc92_origin.pdf b/2025/Collaborative Tree Search for Enhancing Embodied Multi-Agent Collaboration/509a3bd5-5db7-45e0-bb5b-ba942b10cc92_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..3203231a555e6648ab321526fcb5772af346fbd2 --- /dev/null +++ b/2025/Collaborative Tree Search for Enhancing Embodied Multi-Agent Collaboration/509a3bd5-5db7-45e0-bb5b-ba942b10cc92_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4421faf81f73037771ba5a3f9322e373ef0124db4d342a2f3c313b88cdd81500 +size 1077067 diff --git a/2025/Collaborative Tree Search for Enhancing Embodied Multi-Agent Collaboration/full.md b/2025/Collaborative Tree Search for Enhancing Embodied Multi-Agent Collaboration/full.md new file mode 100644 index 0000000000000000000000000000000000000000..9830d7f87f16d42d9a3b1f139905dbe1f69c0e05 --- /dev/null +++ b/2025/Collaborative Tree Search for Enhancing Embodied Multi-Agent Collaboration/full.md @@ -0,0 +1,321 @@ +# Collaborative Tree Search for Enhancing Embodied Multi-Agent Collaboration + +Lizheng $\mathsf{Zu}^{1,2}$ , Lin Lin $^{1*}$ , Song Fu $^{1*}$ , Na Zhao $^{3}$ , Pan Zhou $^{4}$ $^{1}$ Harbin Institute of Technology, $^{2}$ Nanyang Technological University, + $^{3}$ Singapore University of Technology and Design, $^{4}$ Singapore Management University +zuli_edu@outlook.com, waiwaiyl@hit.edu.cn, fusong@hit.edu.cn +na_zhao@sutd.edu.sg, panzhou@smu.edu.sg + +# Abstract + +Embodied agents based on large language models (LLMs) face significant challenges in collaborative tasks, requiring effective communication and reasonable division of labor to ensure efficient and correct task completion. Previous approaches with simple communication patterns carry erroneous or incoherent agent actions, which can lead to additional risks. To address these problems, we propose Cooperative Tree Search (CoTS), a framework designed to significantly improve collaborative planning and task execution efficiency among embodied agents. CoTS guides multi-agents to discuss long-term strategic plans within a modified Monte Carlo tree, searching along LLM-driven reward functions to provide a more thoughtful and promising approach to cooperation. Another key feature of our method is the introduction of a plan evaluation module, which not only prevents agent action confusion caused by frequent plan updates but also ensures plan updates when the current plan becomes unsuitable. Experimental results show that the proposed method performs excellently in planning, communication, and collaboration on embodied environments (CWAH and TDW-MAT), efficiently completing long-term, complex tasks and significantly outperforming existing methods. + +# 1. Introduction + +Embodied agents are increasingly capable of perceiving and interacting with their environments like humans, autonomously managing planning, decision-making, actions, and execution tasks [23, 36]. Recent advances have generated substantial interest in these agents, especially for applications requiring complex task execution in dynamic environments, such as robotic operations in warehouses or domestic settings. In particular, collaborative task completion among multiple agents enhances both efficiency and adapt + +ability, especially in dynamic scenarios [5, 10]. However, collaboration among embodied agents introduces significant challenges: agents must not only perceive and understand their environment but also communicate, share information, divide tasks, and coordinate actions responsively. For example, consider autonomous robots in a home responding to a request: "Bring me the iPad and apple, and put the milk in the refrigerator." Achieving this requires multi-level coordination, including optimal search strategies, task prioritization, and efficient movement planning. + +Large language models (LLMs) have recently provided embodied agents with advanced natural language understanding, dialogue, and reasoning abilities [1, 3, 38, 40]. These capabilities allow LLMs to decompose complex and long-term tasks into a sequence of manageable sub-goals, making LLM-driven agents a promising alternative to traditional reinforcement learning models [14, 34] which are difficult to train and often generalize poorly [7, 21, 39]. + +However, enabling embodied agents to work together in decentralized environments remains a significant and underexplored challenge, as it requires long-term planning and coherent decision-making to coordinate actions efficiently. Early attempts at multi-agent collaboration, such as CoELA [43] and RoCo [24], demonstrate progress but also reveal limitations. For instance, as shown in Fig. 1 (b), CoELA facilitates collaboration by sharing updates through natural language when sub-tasks are completed, yet each agent's decision-making process remains independent, resulting in suboptimal coordination. In contrast, as illustrated in Fig. 1 (c), RoCo develops multi-agent work plans through agent discussions and environmental interactions. However, RoCo relies on a single reasoning path, which is susceptible to the randomness and unpredictability of LLM outputs. This can lead to inefficient or incorrect planning, particularly in applications requiring precision, where such errors can have serious consequences. + +Contributions. To address these limitations, we propose Cooperative Tree Search (CoTS), a framework designed to significantly improve collaborative planning and task ex + +Task: transport 2 oranges, 3 apples, 1 banana, 3 breads, 1 burger to the bed. + +Background: Agents randomly appear in a home environment with a total of six rooms and need to discover objects through exploration. Each agent can carry up to two objects. Finding and using containers can improve efficiency. + +![](images/51f8faae0a245a38a0bcbb01d92263bdf6aa223be74feae5bfb9102f14896f94.jpg) +(a) The Embodied Environment +(b) CoELA (Communication but independent decision-making.) + +![](images/0b6e08abed85d204b9d2583ea6cc8aba813591ec8e0cc70b32599d38f8cb576f.jpg) +(c) RoCo (Discussion, Planning and Execution) + +![](images/822a2c81f6949651bd472a1ffd8758b0757df6f8bb5e0a26c80dec6d849e2ba0.jpg) +(d) Ours (Plan Searching, Evaluation, and Plan-Guided Execution) +Figure 1. The task execution processes of CoELA, RoCo, and our CoTS. (1) In CoELA, agents update each other on sub-task completion and individually determine their next actions without long-term strategic planning. (2) RoCo uses multi-round dialogues and summarizes the plan at the end, but its single-dialogue chain can lead to suboptimal plans and even errors due to the randomness and unpredictability of LLMs. (3) Within a modified Monte Carlo Tree Search framework, CoTS allows agents to evaluate multiple planning paths, enabling them to collaboratively deliberate and identify the most effective course of action based on task progress and environmental context. + +cution efficiency among embodied agents. Inspired by human decision-making, CoTS allows agents to search multiple planning paths by combining exploration and backtracking, enabling them to collaboratively deliberate and identify the most effective course of action based on task progress and environmental context. + +As illustrated in Fig. 1 (d), CoTS advances agent collaboration in embodied environments through two core innovations. First, we introduce a collaborative planning module based on a modified Monte Carlo Tree Search (MCTS) framework, which enables agents to explore multiple planning paths and select the most coherent and efficient long-term plan. Unlike traditional MCTS which relies on simulation-based rewards, CoTS incorporates a non-simulated LLM-based reward function suited for high-cost physical movement interactions in embodied environments. In this way, CoTS transforms the traditional linear sequence of action steps into a search process, exploring potential paths within a shared reasoning and action space to develop robust collaborative strategies. Second, CoTS features a plan evaluation module, that actively monitors the relevance and feasibility of the selected plan during execution. This mechanism allows agents to adapt plans as needed, reducing unnecessary disruptions from frequent plan updates while still accommodating environmental changes. + +Our experiments on two long-term multi-objective planning tasks — TDW-MAT [43] and C-WAH [8] — demonstrate that CoTS significantly outperforms current state-of-the-art methods, e.g., CoELA [43] and RoCo [24], in terms of task completion rate and efficiency. These results highlight the effectiveness and superiority of CoTS, underscoring its potential for high-stakes applications that demand reliable multi-agent collaboration. + +# 2. Related Work + +LLM-based Agents. Recent advances in LLMs have unlocked near-human reasoning and planning capabilities [1, 27, 37]. Researchers have leveraged LLMs as the "brain" or controller of agents, expanding their perceptual and action capacities through multimodal perception and tool use [6, 25, 32, 38, 40, 45]. Multi-agent systems [16, 35] show particular promise, enabling agents to collaborate on planning, discussion, and decision-making—similar to how humans solve complex problems in teams [26, 33]. Works such as MetaGPT [13], CAMEL [19], and ChatDev [31] decompose complex tasks into simpler, manageable subtasks to reduce hallucinations and improve task-solving capabilities. However, embodied multi-agent systems [20, 44] require advanced communication and face significant challenges in navigating a vast action search space in complex, dynamic environments, especially for long-term tasks. + +Embodied Multi-Agent Cooperation. Research on embodied multi-agent systems is still emerging. Habitat [30] explores agents' social awareness, enabling adaptive assistance strategies in collaboration. Co-NavGPT [41] uses LLMs for multi-robot collaborative goal navigation, encoding environmental data to improve scene understanding. Organized-LLM-Agents [11] reduces redundancy in multiagent collaboration through structured prompts, enhancing coordination. CoELA [43] enables collaborative agents to communicate and plan through natural language and chain-of-thought processes, yet lacks strategic long-term planning. RoCo [24] and CaPo [22] advance multi-agent collaboration with high-level communication and path planning, but their single reasoning path often leads to suboptimal plans and even errors. Although some existing work + +![](images/232b84dc39248ff7fdd9552a6c86729adc9372537500dbd2137b559f4f78a0bb.jpg) +Figure 2. The framework of CoTS. Upon receiving a new task in the embodied environment, each embodied agent initially employs the Perception Module to gather raw observational data, which is then stored in its Memory Module. In the collaborative planning module, agents guide the exploration of collaborative plans toward more promising directions through the collaborative tree search. The explored plans are retained within the memory module as a special 'extended memory' and are periodically updated based on assessments from the planning evaluation module. Finally, agents perform actions aligned with the collaborative plan and tasks. + +on LLMs involves multi-branch reasoning [2, 17, 40, 42], they are inadequate for complex embodied environments and agent collaboration. Our goal is to develop embodied agents capable of searching long-term strategic planning. + +# 3. Preliminaries + +Task Assumption. As introduced by previous work [9, 24], multi-agent collaborative tasks can be viewed as a Markov game, which is defined by a tuple $(\mathcal{N},\mathcal{S},\mathcal{A},P,r,\gamma)$ , in which $\mathcal{N}$ is the set of agents (e.g. Alice and Bob), $\mathcal{S}$ is the state space, $\mathcal{A} = \prod_{i=1}^{n}\mathcal{A}^i$ is the combination of all action spaces, $P: \mathcal{S} \times \mathcal{A} \times \mathcal{S} \to [0,1]$ is the transition probability function, $r: \mathcal{S} \times \mathcal{A} \to \mathbb{R}$ is the reward score, and $\gamma \in [0,1)$ is the discount factor. In the task, each agent at time step $t \in \mathbb{N}$ obtains the information by perception module from the embodied environment $s_t \in \mathcal{S}$ , chooses and executes an action $a_t^i \in \mathcal{A}^i$ from its available policy $\pi^i(\cdot|s_t)$ . This will along with the actions of other agents, constitutes a joint action $\mathbf{a}_t = (a_t^1,a_t^2,\dots,a_t^n) \in \mathcal{A}$ derived from the joint policy $\pi(\cdot|s_t) = \prod_{i=1}^{n}\pi^i(\cdot|s_t)$ . Then agents perceive a new state $s_{t+1}$ with probability $P(s_{t+1}|s_t,\mathbf{a}_t)$ . We seek to determine a collaborative plan that maximizes the $J(\pi) \triangleq \mathbb{E}_{s_0,\infty} \sim P,a_0,\infty \sim \pi[\sum_{i=0}^{\infty}\gamma^i r_i]$ . For this paper, the goal is to coordinate the distribution of sub-tasks among agents so as to complete tasks more efficiently. + +CoELA [43]. It contains five modules: 1) Perception, 2) Memory, 3) Communication, 4) Planning and 5) Execution. The Perception module leverages a trained Mask-RCNN [12] to segment RGB images and detect objects and obstacles, creating 3D point clouds from RGB-D data and extracting essential high-level information like key object states. The Memory module manages short-term and long- + +term memory, with short-term memory capturing immediate details such as semantic maps and agent states, and long-term memory storing persistent information like task descriptions. In the Communication module, CoELA retrieves relevant information from memory, transforming it into template-based textual descriptions for agents to aid in action selection. However, each agent plans independently, with no unified action coordination. In the Planning and Execution modules, CoELA retrieves agent-related information from memory, generates potential actions for the LLM to consider, and applies chain-of-thought [38] reasoning for more in-depth decision-making. Each agent then executes its chosen action accordingly. + +# 4. Collaborative Tree Search for Multi-Agents + +# 4.1. Overview of CoTS + +Now we introduce an overview of our Collaborative Tree Search (CoTS), which is designed for multi-agent cooperation in embodied environments. As shown in Fig. 2, CoTS contains four key modules: 1) Perception Module, 2) Memory Module, 3) Collaborative Planning Module, and 4) Plan Parsing and Execution Module. This work primarily enhances CoELA in its planning by designing the Collaborative Planning Module. The other three modules are largely borrowed from CoELA and introduced in Sec. 3. Here we briefly introduce our enhancements to CoEAL below. + +Perception Module. Unlike CoELA, we introduce distance as a key perceptual factor, leveraging fixed spatial layouts in real-world settings to enhance agents' room-based understanding and improve planning efficiency by using LLM to access the moving distance of agents. + +Task: transport 2 oranges, 3 apples, 1 banana, 3 breads, 1 burger to the bed + +The Process in Each Node: Alice makes plans, Bob provides suggestions and rewards. + +Plan from Alice: Step 1: Alice grasps apple. Bob explore his current room. + +Step 2: Alice explore the Living room, if they find any items, transport them to the bed. + +Message to Bob: Hi Bob, I am in Livingroom where I found an apple. Do you have suggestions. + +Message to Alice: Hi Alice, I am in Livingroom and found bread. As you said, I am just about to explore my current room. Reward: 0.3 + +Plan from Alice: Step 1: Alice moves to Bedroom. Bob grasps bread and apple. + +Step 2: Alice explores Bedroom. Bob explores Living room. + +Step 3: If any target item are found, transport them to the bed. + +Message to Bob: Hi Bob, Please grasp bread and apple and I will now shift to the Bedroom because we should not explore the same room. + +Message to Alice: Hi Alice, but I think you could head to kitchen, it is the nearest. Reward: 0.6 + +Plan from Alice: Step 1: Alice grasps the apple. Bob grasps the bread. + +Step 2: Alice moves to Kitchen. Bob explores Living room. + +Step 3: Alice explore the Kitchen, if they find any items, transport them to the bed. + +Message to Bob: Hi Bob, I agree with you. I will grasp the apple before I mo + +Kitchen while you continue in Livingroom. Please keep me updated on your progress + +Message to Alice: Hi, Alice, your plan looks good. Let's get moving. Reward: 1 + +![](images/49f12660ec44c9334be6776b9704c999105e257b7f8ddb2a24ea2df01f7b0f32.jpg) +Figure 3. Overview of the tree search in CoTS. Each node in the tree contains content generated by Alice and Bob. Alice generates the collaborative plans and messages to Bob, while Bob generates messages to Alice and rewards for the plans. These rewards are used for backpropagation to guide the tree search toward more promising collaborative plans. The figure also shows examples of three nodes, illustrating the stepwise optimization of collaborative planning and the dialogue process between agents. + +Memory Module. Compared with CoELA, CoTS introduces "extended memory" (Fig. 2) to manage collaboration plans—information that needs to be retained for a moderate period and updated periodically. This is because collaboration plans must balance retention for a certain period and timely updates when they become unsuitable for the current embodied environment. This extended memory mimics the medium-term storage of the human memory system, bridging short-term recall and long-term consolidation. + +Collaborative Planning Module is the core of CoTS, where MCTS is modified and employed to develop long-term strategic collaboration plans. Inspired by human problem-solving through iterative discussion and cooperation, it avoids error accumulation caused by a single path approach. This module represents the first integration of embodied agent collaboration with MCTS, enhancing agents' capability to tackle complex, long-term tasks through tree-based search and backtracking. Details of this module are provided in Sec. 4.2. + +Plan Parsing and Execution Module translates collaborative plans into executable actions, enabling agents to complete tasks in a coordinated manner. See details in Sec. 4.3. + +Below we will delve deeper into two core modules, Collaborative Planning Module and Plan Parsing and Execution Module. Additionally, we will highlight CoTS's integration of spatial awareness and extended memory to enhance overall functionality and efficiency. + +# 4.2. Collaborative Planning Module + +Our Collaborative Planning Module, central to our CoTS framework, is designed to search for long-term, efficient + +and coherent collaborative plans for multi-agent tasks in dynamic environments. It comprises two main sub-modules: a collaborative tree search module for generating collaborative plans (Sec. 4.2.1) and a plan evaluation module to analyze when to update the plan (Sec. 4.2.2). + +# 4.2.1. Monte Carlo Tree Search for Collaborative Plan + +We adapt MCTS to convert multi-agent collaborative planning into a tree search, identifying an optimal pathway from multiple reasoning paths for cohesive strategies. As shown in Fig. 3, each node in this tree stores a proposed plan, agent feedback, and a performance reward. In this setup, one agent (like Alice) functions as the "planner" to propose plans, while other agents (like Bob) act as "critics". Alice drafts several initial collaborative plans or extends existing plan based on task goals and current task progress. Then Bob evaluates them, considering factors like task allocation and movement efficiency, to generate a reward for the current plan. Once feedback is provided, each plan's reward is propagated back through the collaboration tree, strengthening successful strategies and adjusting future planning pathways. Accordingly, they will select the most efficient collaborative plan for further reflections and refinements. As shown in Fig. 3, this process, including plan extension (drafting), evaluation, backpropagation, and selection, will repeat until agents explore all potential collaborative plans and select the most promising one. In the following, for these four steps, we will elaborate on each of them. + +Selection Similar to traditional MCTS [4, 15], CoTS starts from the root node and traverses down the tree by selecting nodes with the highest UCT values. This process is repeated + +until a leaf node is reached, and is illustrated in Fig. 3 where we mark the selected nodes with gold cups. For UCT of each node, it is computed as follows: + +$$ +\mathrm {U C T} (s _ {i}) = V (s _ {i}) + w \sqrt {\ln N (p _ {i}) / N (s _ {i})}, \tag {1} +$$ + +Where $N(s_{i})$ is visiting number of node $s_i$ , $V(s_{i})$ is the reward of the subtree rooted at $s_i$ , $w$ is the exploration coefficient, and $p_i$ is the parent node of $s_i$ . After an episode, backpropagation updates each node's $V(s_{i})$ along the path: $V(s_{i}) = \frac{V(s_{i - 1})\cdot(N(s_{i}) - 1) + r}{N(s_{i})}$ , where $V(s_{i - 1})$ is the previous reward, and $r$ is the reward used to update each $V(s_{i})$ . + +Expansion. Once a node is selected, we expand it by sampling $n$ new discussion paths, each building on the plan (from "planner", Alice) and suggestions (from "critic", Bob) in the parent node. This generates $n$ new child nodes, representing potential planning routes. The tree, along with its nodes, is stored in an external long-term memory to retain valuable information and facilitate consistent planning across episodes. + +Evaluation. Each newly created child node receives a reward $r$ to quantify the quality of the collaboration plan and to guide the search algorithm towards the most promising areas of the tree. Here CoTS incorporates a non-simulated LLM-based reward function to estimate the reward $r$ of a plan in a node. In traditional MCTS, reward estimation requires agents to execute their actions for receiving environment feedback, like completing a subtask, and thus is high-cost. So CoTS introduces an LLM-based reward function based on two components: (1) an allocation reward assessing coherence and efficiency, and (2) a cost reward evaluating the spatial movement cost of agents. To ensure consistency, we standardize the evaluation criteria to provide clear guidelines to minimize ambiguity. Each criterion is scored on a scale of 1 to 5, with the final reward normalized between 0 and 1. See examples in Appendix B.1. + +Backpropagation. Each evaluation reward is propagated back along the path to update node rewards, with formulas ensuring that parent nodes integrate the feedback, guiding future selections. Specifically, the back-propagation updates node rewards $N(s_{i}) = N(s_{i - 1}) + 1$ and $V(s_{i}) = \frac{V(s_{i - 1})N(s_{i - 1}) + r}{N(s_i)}$ for each node $s_0,s_1,\ldots ,s_p$ along the path from the root (initial state $s_0$ ) to leaf (terminal state $s_p$ ). Accordingly these estimated rewards and visiting number are used in "Expansion" step for node selection, directing the search toward the most promising areas of the tree. + +# 4.2.2. Collaborative Plan Evaluation and Adaptation + +In this section, we will explain why we introduced the planning evaluation module (a sub-module in the collaborative planning module, as shown in Fig. 2) and its function. The collaborative plan provides high-level embodied agent action guidance, but as the task is executed and agents act, the embodied environment gradually changes (e.g. the agents' + +positions, the agents' progress, and the current status, etc.). Therefore, the collaborative plan needs to be updated based on the latest environmental information. + +At this point, two scenarios arise: If the plan is updated too slowly or not updated at all, the current plan becomes outdated and cannot effectively guide the embodied agents in executing tasks. However, in a multi-agent system, instantly updating the plan with every progress can be inefficient and counterproductive. It may lead to excessive queries to the LLM and environments. Furthermore, frequent plan revisions may introduce conflicting strategies due to minor environmental shifts, potentially disrupting agent actions and lowering overall efficiency. This is illustrated in Fig. 4 (c): if the updated plan conflicts with the action that Bob is currently performing, it may interrupt Bob's action which even almost finishes the subtask. + +To address this, we introduce a plan evaluation module within CoTS to continuously evaluate the viability of the current collaborative plan. When an agent makes progress, the module determines if the existing plan remains effective based on the status of task execution. If the evaluation metric meets a defined threshold, indicating that the plan is still viable, it is retained. Otherwise, CoTS will call the collaborative planning module to re-search and update the plan. By focusing on the feasibility and long-term stability of the plan rather than continuous recalibration, this module supports efficient, coherent collaboration, enabling agents to complete tasks reliably in a dynamic environment. + +# 4.3. Plan Parsing and Execution Module + +To transform the collaborative plan generated in Sec. 4.2 into executable actions, CoTS incorporates an action parsing module and a plan decomposition module. Action parsing module bridges high-level planning with real-world actions, generating a set of feasible actions for each agent based on current memory. For example, if an agent's hands are free, it will list objects that can be grabbed; if the agent is in an unknown room, it will list movement options. This module ensures that all sub-plans are formatted into natural language and returned for subsequent process. + +Plan decomposition module understands and breaks down high-level tasks from the collaboration plan into specific and prioritized actions. The agent interprets the task arrangement, selects actions from the parsed options that best advance the collaborative plan's goals, and finally translates each chosen action into real-world execution. + +# 5. Experiments + +**Benchmarks. Following CoELA [43], we evaluate the performance of CoTS in two highly challenging embodied-world multi-agent environments: (1) TDW-MAT (ThreeD World Multi-Agent Transport) [8] extends the ThreeD World Transport Challenge by simulating collaborative + +Table 1. Quantitative results on TDW-MAT. Consistent with CoELA [43], TDW-MAT has two environment settings: with oracle perception and without oracle perception. The results marked with * are from CoELA [43]. MAT uses central observation and oracle perception. $91_{+9.64\%}$ means the relative improvement when comparing with CoELA is $9.64\%$ . + +
Transport Rate (%)Base AgentRL AgentLLM Agent
RHP*CoRHP*MAT*GPT 3.5GPT 4
CoELARoCoCaPoCoTS(ours)CoELA*RoCoCaPoCoTS(ours)
TDW-MAT w/ Oracle Perception
Stuff(↑)4974177380+9.68%84+15.07%81+10.96%8385+2.41%87+4.82%91+9.64%
Food(↑)5276137274+2.78%85+18.06%83+15.28%8789+2.30%90+3.45%94+8.05%
Average(↑)5075157377+5.48%84+15.07%82+12.33%8587+2.35%89+4.71%93+9.41%
TDW-MAT w/o Oracle Perception
Stuff(↑)3654-3947+20.51%45+15.38%50+28.21%6164+4.92%64+4.92%77+26.22%
Food(↑)4967-6764-4.48%70+4.48%67+0.00%8280-2.44%85+3.65%88+7.32%
Average(↑)4361-5255+5.77%57+9.62%59+13.46%7172+1.41%74+4.23%83+16.90%
+ +Table 2. Quantitative results on C-WAH. LLM-based agents are driven by GPT 4. Results marked with * are from CoELA [43]. + +
Average StepVisual Obs(↓)Symbolic Obs(↓)
Base AgentMHP*141111
CoMHP*10375
LLM AgentCoELA*9257
RoCo89+3.26%57+0.00%
CaPo83+9.78%51+10.53%
CoTS(ours)81+11.96%49+14.04%
+ +transportation tasks in a physics-based household environment. We select 24 tasks, split evenly between food and stuff transportation. (2) C-WAH (Communicative Watch-And-Help) [29] extends the Watch-And-Help challenge within the VirtualHome-Social platform [28]. It features five common household scenarios. Our test set includes 10 tasks, with two tasks per scenario. The objective is to complete all sub-goals, prioritizing faster completion. + +Evaluation Metrics. On TDW-MAT, transport rate is evaluated by the percentage of target objects successfully delivered to their designated locations within a 3,000-frame timeframe. For C-WAH, efficiency is measured by the number of time steps needed to complete all assigned tasks, where fewer steps indicate higher efficiency. + +Baseline Methods. The baselines include three types of embodied agents: (1) Base Agent: RHP [8] uses heuristic rules for navigation and $\mathbf{A}^*$ for path planning. MHP [18] employs Monte Carlo Tree Search (MCTS) for high-level planning coupled with regression planning for task execution (2) RL-Based Agent: MAT [39] is a multi-agent reinforcement learning framework, featuring a centralized decision transformer that synthesizes actions from shared observations. (3) LLM-Based Agent: CoELA [43] is a leading collaboration framework for embodied agents that integrates the reasoning capabilities of LLMs. It facilitates communication and interaction between agents through natural language. RoCo [24] and CaPo [22] are multi-agent collaboration methods that enhance collaboration by gener + +ating coordinated plans through structured agent dialogues. Setup. We conduct all experiments on an Intel Core i9- 10980XE CPU and an NVIDIA GeForce RTX 3090 GPU. We utilize both GPT-4 and GPT-3.5-Turbo [1], configuring them with a temperature setting of 0.7 and top-1 sampling to produce outputs up to 512 tokens. + +# 5.1. Main Results + +CoTS Outperforms the Strong Baselines. Following the experimental frameworks of CoELA [43] and RoCo [24], we configure two-agent cooperation, excluding RHP. As demonstrated in Table 1, CoTS achieves best performance on the TDW-MAT task. Specifically, when powered by GPT-4 within the Oracle Perception environment, CoTS surpasses CoELA, achieving a $9.41\%$ average relative improvement in task completion, and a $16.9\%$ improvement in the absence of Oracle Perception. The primary strength of CoTS is its ability to generate superior long-term strategies through collaborative tree search. + +Conversely, CoELA focuses on short-term actions derived from immediate information, which limits its strategic depth and adaptability to future scenarios and environmental shifts. RoCo and CaPo are similar. Although they are capable of developing plans, they rely on a single-branch discussion method, which may result in suboptimal or inaccurate strategies. Moreover, Their frequent plan revisions can interfere with agent coordination. Therefore, while RoCo and CaPo surpass CoELA, they are inferior to CoTS. When driven by GPT-3.5, CaPo slightly outperforms CoTS on some metrics. The main reason for this is that GPT-3.5 has weaker evaluation and scoring capabilities. However, when powered by GPT-4, CoTS significantly outperforms CaPo. Our findings underscore that multi-agent collaboration generally surpasses single-agent strategies, exemplified by RHP. The performance of reinforcement learning-based MAT is poor, mainly because TDW-MAT is too difficult for it with long time domain and sparse reward signals. + +The performance of LLM-based agents primarily de + +# (a) CoELA leads to short-sighted action plans that ignore precious action steps + +Background: Alice is in living room holding a container. Bob finds two target objects in living room $< 1000>$ . + +![](images/7ec545d557c91cb677e0ac6a1a31ce96a613fb553992580942afa15368db28e0.jpg) + +CoELA's plan : Alice goes to the living room to help Bob, because Alice can take these objects. + +Our plan: Alice explores the current room, find objects, then transport them to the bedroom. Then go to the living room $< 1000>$ , no actions are wasted. + +# (b) RoCo's single-chain dialogue leads to suboptimal collaboration plan + +Background: Alice and Bob appear in the same living room and now need to explore some rooms to find objects. + +![](images/df07c36f0467d2e805c8a4847c030012552c8294be636c605e3c962eddf7ad6e.jpg) + +RoCo's plan : Alice explore the current room. Bob explore the current room. + +Our plan: Alice goes to the nearest room, Bedroom $< 2000>$ , while Bob explores the current room. Exploring rooms separately would improve efficiency. + +# (c) Frequent plan updates may hinder completion of current actions + +Background: Bob was transporting objects to the bed, Alice found a bread in the living room<6000> + +![](images/061d1fe605d5d158f3c45e816e2ca1490761704a4290c947d3a587adc4e3c613.jpg) + +RoCo's Plan: Bob goes to the living room $< 6000>$ to help Alice grasp the bread. + +Our plan: The current plan is still valid. Bob is close to the bedroom. He can first deliver the objects in his hand to the bed, and then help Alice. + +Figure 4. Some interesting example cases demonstrate that our CoTS achieves better collaboration than RoCo and CoELA. These scenes are taken from the experiments in TDW-MAT. +![](images/416f57abde0d9bc8ea29cbe41eebe0e3a9a49735b8e634b2cf408935b176fec2.jpg) +There seems to be no plan. Both Alice and Bob are not holding anything. Bob in bedroom and has found an container, while Alice is in the. A collaboration plan needs to be updated. Reward: 1 + +![](images/67ebbf0492b3fb860a8c8644426bd3bed0427552740056ca7d3d2f47705ecba0.jpg) +Alice have found an apple in the fridge and Bob is holding a cupcake. Current progress is consistent with the collaborative plan and future actions are outlined in the plan. There is no need to update the plan. Reward: 4 + +pends on their reasoning abilities and the depth of embedded world knowledge. Agents powered by GPT-4, as demonstrated by CoTS, significantly surpass those based on GPT-3.5, highlighting the advancements in LLM technologies and their impact on agent proficiency. As shown in Table 2, CoTS achieves superior efficiency, improving by $11.96\%$ and $14.04\%$ over CoELA in Visual Obs and Symbolic Obs settings, respectively. These results underscore CoTS's robust adaptability across various embodied environments and its potential for real-world applications. + +# 5.2. Qualitative Analysis + +CoTS Can Develop Long-term Strategic Plans. Unlike CoELA, which is adept at responding to immediate tasks + +![](images/51da022b49258ffbf5632dc1659b5bacbabb95bdc0c5f68b8ea8931500179902.jpg) +Figure 5. Plan Evaluation and Update. The plan evaluation module gives a lower reward when it determines that the current plan needs to be updated, and a higher reward when it believes the current plan is still valid. The reward range is from 1 to 5. +Figure 6. Comparison of transportation rates for GPT-4-based CoTS and CoELA at different time steps on TDW-MAT. + +but falters in sustained coordination, CoTS excels in long-term strategic planning. As illustrated in Fig. 4 (a), when Bob discovers a new target that requires assistance, Alice, under CoTS's direction, continues to thoroughly explore her current room before moving the item to the bed. This systematic approach not only supports Bob but also adheres to an overarching collaborative strategy, optimizing both immediate and future task alignments. + +CoTS Generates More Optimal Collaboration Plans. Unlike RoCo to generate plans sequentially through a single-chain dialogue, CoTS makes more optimized collaboration plans. Due to the randomness of LLMs and the embodied environment's complexity, RoCo's plans may sometimes be suboptimal. For example, in Fig. 4 (b), RoCo instructs both Alice and Bob to explore their respective rooms. However, they are currently in the same room, making joint exploration redundant. CoTS effectively identifies this and assigns Alice to explore the nearest room instead. + +![](images/cbda7706dce07bdf1b38b22867ba2e33c5cfaf3247cd8abf647981a304ef0b4c.jpg) +Figure 7. Comparison of transportation rates with different branching factors. More branching factors means more new nodes are generated each time in the expansion phase. + +CoTS Ensures Smooth Completion of Agent Actions. In RoCo, new progress often prompt discussions and lead to new plans that can override existing ones. As illustrated in Fig. 4 (c), Bob, tasked with transporting items to the bed, must halt his activity to help Alice following a revised plan. Conversely, CoTS incorporates a plan evaluation module that validates the feasibility of plans before any adjustments are made. In Fig. 4 (c), CoTS allows Bob to finish transporting before next action. This approach avoids chaotic agent actions and ensures timely plan updates when necessary. + +CoTS Dynamically Manages Collaborative Plans. As shown in Fig. 5, the assessment module continuously evaluates the viability of existing plans. It assigns a lower reward (e.g. 1) if collaboration necessitates devising a new plan or modifying an existing one, prompting agents to adapt their strategies accordingly. Conversely, when the current progress aligns with planned tasks and their distribution among agents is deemed efficient, the module rewards this alignment with a higher value (e.g. 4). This incentivizes agents to persist with the established collaborative plan, avoiding unnecessary replanning. + +# 5.3. Ablation Study and Additional Analysis + +The Core Modules of CoTS Have Important Impacts. The Table 3 shows the ablation results of CoTS. To examine the impact of tree search, we modify it to a single-branch structure, where agents can only communicate along one path. The experimental results indicate a $5.38\%$ relative decrease in the average transportation rate, mainly because the limited communication does not consistently yield correct conclusions and diverges from typical human discussion patterns. Furthermore, eliminating the reward function and plan update evaluation components from CoTS leads to variable performance declines, highlighting the critical roles these elements play in the system's efficacy. + +Analysis of Agent Collaboration Progress at Different Time Steps. We analyze the transportation rates of CoELA and CoTS on TDW-MAT at various time steps, as depicted in Fig. 6. An interesting observation emerges at 500 steps, where CoTS exhibits a lower performance compared to + +Table 3. Effects of the reward function, collaboration tree search, and plan evaluation module in GPT-4-driven CoTS on TDW-MAT. + +
Transport RateStuff(↑)Food(↑)Average(↑)
CoTS w/o Reward87-4.40%90-4.26%89-4.30%
CoTS w/o MCTS Plan85-6.59%91-3.19%88-5.38%
CoTS w/o E&U88-3.30%92-2.13%90-3.23%
CoTS919493
+ +CoELA, attributed to CoTS's strategy of prioritizing item collection early to enhance transport capacity. However, this initial disadvantage is counterbalanced by a significant improvement at 2000 steps, where CoTS markedly surpasses CoELA. This shift underscores CoTS's effectiveness in optimizing agent collaboration for task execution. + +Impact of Different Branching Factors on Results. During the expansion phase of MCTS, we define the number of new candidate nodes generated at each step as the "branching factors." As shown in Fig. 7, increasing the number of branching factors expands the selection space, leading to a better plan. Adjusting the branching factor allows for a balance between performance and efficiency. + +Besides, the human-agent collaboration and the cost of CoTS are discussed in Appendix A.1 and Appendix A.2, respectively. CoTS also demonstrates significant advantages. + +# 6. Conclusions + +This paper introduces CoTS, a collaborative framework that leverages large language models with a modified Monte Carlo tree search to enhance multi-agent task planning in complex and dynamic environments. By enabling agents to evaluate multiple potential plans and refine strategies based on progress, CoTS significantly advances over prior methods like CoELA and RoCo. The addition of a plan feasibility module ensures smooth task execution by adapting plans only when necessary. Experimental results on multi-objective tasks confirm CoTS's superior performance. + +Limitation. CoTS relies on LLMs' reasoning abilities. We find that GPT-4 surpasses GPT-3.5-turbo in instruction-following and response quality, while GPT-3.5-turbo can be verbose and occasionally inaccurate. Despite this, even for GPT-3.5, CoTS improves agent collaboration through collaborative planning search in the collaboration tree. + +# 7. Acknowledgments + +This work is supported by National Natural Science Foundation of China (U2133202, 52305570), China Postdoctoral Science Foundation (2022M720955), and Outstanding Doctoral Dissertation Funding Project of Heilongjiang Province (LJYXL2022-011). Lizheng Zu also acknowledges the support of the China Scholarship Council program (202406120111). + +# References + +[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 1, 2, 6 +[2] David Brandfonbrener, Simon Henniger, Sibi Raja, Tarun Prasad, Chloe R Loughridge, Federico Cassano, Sabrina Ruixin Hu, Jianang Yang, William E Byrd, Robert Zinkov, et al. Vermcts: Synthesizing multi-step programs using a verifier, a large language model, and tree search. In The 4th Workshop on Mathematical Reasoning and AI at NeurIPS'24, 2024. 3 +[3] Tom B Brown. Language models are few-shot learners. arXiv preprint arXiv:2005.14165, 2020. 1 +[4] Cameron B Browne, Edward Powley, Daniel Whitehouse, Simon M Lucas, Peter I Cowling, Philipp Rohlfshagen, Stephen Tavener, Diego Perez, Spyridon Samothrakis, and Simon Colton. A survey of monte carlo tree search methods. IEEE Transactions on Computational Intelligence and AI in games, 4(1):1-43, 2012. 4, 3 +[5] Abhinav Dahiya, Alexander M Aroyo, Kerstin Dautenhahn, and Stephen L Smith. A survey of multi-agent human-robot interaction systems. Robotics and Autonomous Systems, 161: 104335, 2023. 1 +[6] Norman Di Palo, Arunkumar Byravan, Leonard Hasenclever, Markus Wulfmeier, Nicolas Heess, and Martin Riedmiller. Towards a unified agent with foundation models. arXiv preprint arXiv:2307.09668, 2023. 2 +[7] Amal Feriani and Ekram Hossain. Single and multi-agent deep reinforcement learning for ai-enabled wireless networks: A tutorial. IEEE Communications Surveys & Tutorial, 23(2):1226-1252, 2021. 1 +[8] Chuang Gan, Siyuan Zhou, Jeremy Schwartz, Seth Alter, Abhishek Bhandwaldar, Dan Gutfreund, Daniel LK Yamins, James J DiCarlo, Josh McDermott, Antonio Torralba, et al. The threadworld transport challenge: A visually guided task-and-motion planning benchmark towards physically realistic embodied ai. In 2022 International conference on robotics and automation (ICRA), pages 8847-8854. IEEE, 2022. 2, 5, 6 +[9] Ran Gong, Qiuyuan Huang, Xiaojian Ma, Yusuke Noda, Zane Durante, Zilong Zheng, Demetri Terzopoulos, Li Fei-Fei, Jianfeng Gao, and Hoi Vo. Mindagent: Emergent gaming interaction. In Findings of the Association for Computational Linguistics: NAACL 2024, pages 3154–3183, 2024. 3 +[10] Taicheng Guo, Xiuying Chen, Yaqi Wang, Ruidi Chang, Shichao Pei, Nitesh V Chawla, Olaf Wiest, and Xiangliang Zhang. Large language model based multi-agents: A survey of progress and challenges. arXiv preprint arXiv:2402.01680, 2024. 1 +[11] Xudong Guo, Kaixuan Huang, Jiale Liu, Wenhui Fan, Natalia Vélez, Qingyun Wu, Huazheng Wang, Thomas L Griffiths, and Mengdi Wang. Embodied llm agents learn to cooperate in organized teams. arXiv preprint arXiv:2403.12482, 2024. 2 + +[12] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 3 +[13] Sirui Hong, Xiawu Zheng, Jonathan Chen, Yuheng Cheng, Jinlin Wang, Ceyao Zhang, Zili Wang, Steven Ka Shing Yau, Zijuan Lin, Liyang Zhou, et al. Metagpt: Meta programming for multi-agent collaborative framework. arXiv preprint arXiv:2308.00352, 2023. 2 +[14] Wenlong Huang, Pieter Abbeel, Deepak Pathak, and Igor Mordatch. Language models as zero-shot planners: Extracting actionable knowledge for embodied agents. In International conference on machine learning, pages 9118-9147. PMLR, 2022. 1 +[15] Steven James, George Konidaris, and Benjamin Rosman. An analysis of monte carlo tree search. In Proceedings of the AAAI Conference on Artificial Intelligence, 2017. 4 +[16] Feibo Jiang, Yubo Peng, Li Dong, Kezhi Wang, Kun Yang, Cunhua Pan, Dusit Niyato, and Octavia A Dobre. Large language model enhanced multi-agent systems for 6g communications. IEEE Wireless Communications, 2024. 2 +[17] Bowen Jin, Chulin Xie, Jiawei Zhang, Kashob Kumar Roy, Yu Zhang, Zheng Li, Ruirui Li, Xianfeng Tang, Suhang Wang, Yu Meng, et al. Graph chain-of-thought: Augmenting large language models by reasoning on graphs. arXiv preprint arXiv:2404.07103, 2024. 3 +[18] Richard E Korf. Planning as search: A quantitative approach. Artificial intelligence, 33(1):65-88, 1987. 6, 5 +[19] Guohao Li, Hasan Hammoud, Hani Itani, Dmitrii Khizbullin, and Bernard Ghanem. Camel: Communicative agents for" mind" exploration of large language model society. Advances in Neural Information Processing Systems, 36: 51991-52008, 2023. 2 +[20] Xin Li. Positioning is all you need. arXiv preprint arXiv:2404.01183, 2024. 2 +[21] Iou-Jen Liu, Unnat Jain, Raymond A Yeh, and Alexander Schwing. Cooperative exploration for multi-agent deep reinforcement learning. In International conference on machine learning, pages 6826-6836. PMLR, 2021. 1 +[22] Jie Liu, Pan Zhou, Yingjun Du, Ah-Hwee Tan, Cees GM Snoek, Jan-Jakob Sonke, and Efstratios Gavves. Capo: Cooperative plan optimization for efficient embodied multiagent cooperation. arXiv preprint arXiv:2411.04679, 2024. 2, 6, 5 +[23] Yang Liu, Weixing Chen, Yongjie Bai, Jingzhou Luo, Xinshuai Song, Kaixuan Jiang, Zhida Li, Ganlong Zhao, Junyi Lin, Guanbin Li, et al. Aligning cyber space with physical world: A comprehensive survey on embodied ai. arXiv preprint arXiv:2407.06886, 2024. 1 +[24] Zhao Mandi, Shreeya Jain, and Shuran Song. Roco: Dialectic multi-robot collaboration with large language models. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 286–299. IEEE, 2024. 1, 2, 3, 6, 5 +[25] Reiichiro Nakano, Jacob Hilton, Suchir Balaji, Jeff Wu, Long Ouyang, Christina Kim, Christopher Hesse, Shantanu Jain, Vineet Kosaraju, William Saunders, et al. Webgpt: Browser-assisted question-answering with human feedback. arXiv preprint arXiv:2112.09332, 2021. 2 + +[26] Bo Ni and Markus J Buehler. Mechagents: Large language model multi-agent collaborations can solve mechanics problems, generate new data, and integrate knowledge. Extreme Mechanics Letters, 67:102131, 2024. 2 +[27] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. Advances in neural information processing systems, 35:27730-27744, 2022. 2 +[28] Xavier Puig, Kevin Ra, Marko Boben, Jiaman Li, Tingwu Wang, Sanja Fidler, and Antonio Torralba. Virtualhome: Simulating household activities via programs. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8494-8502, 2018. 6 +[29] Xavier Puig, Tianmin Shu, Shuang Li, Zilin Wang, Yuan-Hong Liao, Joshua B Tenenbaum, Sanja Fidler, and Antonio Torralba. Watch-and-help: A challenge for social perception and human-ai collaboration. arXiv preprint arXiv:2010.09890, 2020. 6 +[30] Xavier Puig, Eric Undersander, Andrew Szot, Mikael Dallaire Cote, Tsung-Yen Yang, Ruslan Partsey, Ruta Desai, Alexander William Clegg, Michal Hlavac, So Yeon Min, et al. Habitat 3.0: A co-habitat for humans, avatars and robots. arXiv preprint arXiv:2310.13724, 2023. 2 +[31] Chen Qian, Wei Liu, Hongzhang Liu, Nuo Chen, Yufan Dang, Jiahao Li, Cheng Yang, Weize Chen, Yusheng Su, Xin Cong, et al. Chatdev: Communicative agents for software development. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 15174-15186, 2024. 2 +[32] Timo Schick, Jane Dwivedi-Yu, Roberto Dessì, Roberta Raileanu, Maria Lomeli, Eric Hambro, Luke Zettlemoyer, Nicola Cancedda, and Thomas Scialom. Toolformer: Language models can teach themselves to use tools. Advances in Neural Information Processing Systems, 36, 2024. 2 +[33] Paul E Smaldino, Cody Moser, Alejandro Perez Velilla, and Mikkel Werling. Maintaining transient diversity is a general principle for improving collective problem solving. Perspectives on Psychological Science, 19(2):454-464, 2024. 2 +[34] Chan Hee Song, Jiaman Wu, Clayton Washington, Brian M Sadler, Wei-Lun Chao, and Yu Su. Llm-planner: Few-shot grounded planning for embodied agents with large language models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2998-3009, 2023. 1 +[35] Wiebe Van der Hoek and Michael Wooldridge. Multi-agent systems. Foundations of Artificial Intelligence, 3:887-928, 2008. 2 +[36] Tai Wang, Xiaohan Mao, Chenming Zhu, Runsen Xu, Ruiyuan Lyu, Peisen Li, Xiao Chen, Wenwei Zhang, Kai Chen, Tianfan Xue, et al. Embodiedscan: A holistic multimodal 3d perception suite towards embodied ai. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19757-19767, 2024. 1 +[37] Jason Wei, Yi Tay, Rishi Bommasani, Colin Raffel, Barret Zoph, Sebastian Borgeaud, Dani Yogatama, Maarten Bosma, Denny Zhou, Donald Metzler, et al. Emergent abilities of + +large language models. arXiv preprint arXiv:2206.07682, 2022. 2 +[38] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824-24837, 2022. 1, 2, 3 +[39] Muning Wen, Jakub Kuba, Runji Lin, Weinan Zhang, Ying Wen, Jun Wang, and Yaodong Yang. Multi-agent reinforcement learning is a sequence modeling problem. Advances in Neural Information Processing Systems, 35:16509-16521, 2022. 1, 6, 5 +[40] Shunyu Yao, Dian Yu, Jeffrey Zhao, Izhak Shafran, Tom Griffiths, Yuan Cao, and Karthik Narasimhan. Tree of thoughts: Deliberate problem solving with large language models. Advances in Neural Information Processing Systems, 36, 2024. 1, 2, 3 +[41] Bangguo Yu, Hamidreza Kasaei, and Ming Cao. Co-navgpt: Multi-robot cooperative visual semantic navigation using large language models. arXiv preprint arXiv:2310.07937, 2023. 2 +[42] Di Zhang, Xiaoshui Huang, Dongzhan Zhou, Yuqiang Li, and Wanli Ouyang. Accessing gpt-4 level mathematical olympiad solutions via monte carlo tree self-refine with llama-3 8b: A technical report. arXiv preprint arXiv:2406.07394, 2024. 3 +[43] Hongxin Zhang, Weihua Du, Jiaming Shan, Qinhong Zhou, Yilun Du, Joshua B Tenenbaum, Tianmin Shu, and Chuang Gan. Building cooperative embodied agents modularly with large language models. arXiv preprint arXiv:2307.02485, 2023. 1, 2, 3, 5, 6 +[44] Zhonghan Zhao, Ke Ma, Wenhao Chai, Xuan Wang, Kewei Chen, Dongxu Guo, Yanting Zhang, Hongwei Wang, and Gaoang Wang. Do we really need a complex agent system? distill embodied agent into a single model. arXiv preprint arXiv:2404.04619, 2024. 2 +[45] Andy Zhou, Kai Yan, Michal Shlapentokh-Rothman, Haohan Wang, and Yu-Xiong Wang. Language agent tree search unifies reasoning acting and planning in language models. arXiv preprint arXiv:2310.04406, 2023. 2 \ No newline at end of file diff --git a/2025/Collaborative Tree Search for Enhancing Embodied Multi-Agent Collaboration/images.zip b/2025/Collaborative Tree Search for Enhancing Embodied Multi-Agent Collaboration/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..6fc3c8dd46b72c0c27e82e6e504671cf0c33251c --- /dev/null +++ b/2025/Collaborative Tree Search for Enhancing Embodied Multi-Agent Collaboration/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:035330dd57af07f45f07b9778af646e73fea2f786851f3a9b6c804a520877a0f +size 443777 diff --git a/2025/Collaborative Tree Search for Enhancing Embodied Multi-Agent Collaboration/layout.json b/2025/Collaborative Tree Search for Enhancing Embodied Multi-Agent Collaboration/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..56dc6d4f86c7bd463f5b2297dd2a1ab2a07221bc --- /dev/null +++ b/2025/Collaborative Tree Search for Enhancing Embodied Multi-Agent Collaboration/layout.json @@ -0,0 +1,8073 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 61, + 102, + 550, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 102, + 550, + 121 + ], + "spans": [ + { + "bbox": [ + 61, + 102, + 550, + 121 + ], + "type": "text", + "content": "Collaborative Tree Search for Enhancing Embodied Multi-Agent Collaboration" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 97, + 142, + 514, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 142, + 514, + 213 + ], + "spans": [ + { + "bbox": [ + 97, + 142, + 514, + 213 + ], + "type": "text", + "content": "Lizheng " + }, + { + "bbox": [ + 97, + 142, + 514, + 213 + ], + "type": "inline_equation", + "content": "\\mathsf{Zu}^{1,2}" + }, + { + "bbox": [ + 97, + 142, + 514, + 213 + ], + "type": "text", + "content": ", Lin Lin" + }, + { + "bbox": [ + 97, + 142, + 514, + 213 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 97, + 142, + 514, + 213 + ], + "type": "text", + "content": ", Song Fu" + }, + { + "bbox": [ + 97, + 142, + 514, + 213 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 97, + 142, + 514, + 213 + ], + "type": "text", + "content": ", Na Zhao" + }, + { + "bbox": [ + 97, + 142, + 514, + 213 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 97, + 142, + 514, + 213 + ], + "type": "text", + "content": ", Pan Zhou" + }, + { + "bbox": [ + 97, + 142, + 514, + 213 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 97, + 142, + 514, + 213 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 97, + 142, + 514, + 213 + ], + "type": "text", + "content": "Harbin Institute of Technology, " + }, + { + "bbox": [ + 97, + 142, + 514, + 213 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 97, + 142, + 514, + 213 + ], + "type": "text", + "content": "Nanyang Technological University, \n" + }, + { + "bbox": [ + 97, + 142, + 514, + 213 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 97, + 142, + 514, + 213 + ], + "type": "text", + "content": "Singapore University of Technology and Design, " + }, + { + "bbox": [ + 97, + 142, + 514, + 213 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 97, + 142, + 514, + 213 + ], + "type": "text", + "content": "Singapore Management University \nzuli_edu@outlook.com, waiwaiyl@hit.edu.cn, fusong@hit.edu.cn \nna_zhao@sutd.edu.sg, panzhou@smu.edu.sg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 151, + 240, + 200, + 254 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 240, + 200, + 254 + ], + "spans": [ + { + "bbox": [ + 151, + 240, + 200, + 254 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 266, + 297, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 266, + 297, + 542 + ], + "spans": [ + { + "bbox": [ + 54, + 266, + 297, + 542 + ], + "type": "text", + "content": "Embodied agents based on large language models (LLMs) face significant challenges in collaborative tasks, requiring effective communication and reasonable division of labor to ensure efficient and correct task completion. Previous approaches with simple communication patterns carry erroneous or incoherent agent actions, which can lead to additional risks. To address these problems, we propose Cooperative Tree Search (CoTS), a framework designed to significantly improve collaborative planning and task execution efficiency among embodied agents. CoTS guides multi-agents to discuss long-term strategic plans within a modified Monte Carlo tree, searching along LLM-driven reward functions to provide a more thoughtful and promising approach to cooperation. Another key feature of our method is the introduction of a plan evaluation module, which not only prevents agent action confusion caused by frequent plan updates but also ensures plan updates when the current plan becomes unsuitable. Experimental results show that the proposed method performs excellently in planning, communication, and collaboration on embodied environments (CWAH and TDW-MAT), efficiently completing long-term, complex tasks and significantly outperforming existing methods." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 566, + 135, + 578 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 566, + 135, + 578 + ], + "spans": [ + { + "bbox": [ + 55, + 566, + 135, + 578 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 586, + 296, + 695 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 586, + 296, + 695 + ], + "spans": [ + { + "bbox": [ + 55, + 586, + 296, + 695 + ], + "type": "text", + "content": "Embodied agents are increasingly capable of perceiving and interacting with their environments like humans, autonomously managing planning, decision-making, actions, and execution tasks [23, 36]. Recent advances have generated substantial interest in these agents, especially for applications requiring complex task execution in dynamic environments, such as robotic operations in warehouses or domestic settings. In particular, collaborative task completion among multiple agents enhances both efficiency and adapt" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 242, + 555, + 362 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 242, + 555, + 362 + ], + "spans": [ + { + "bbox": [ + 313, + 242, + 555, + 362 + ], + "type": "text", + "content": "ability, especially in dynamic scenarios [5, 10]. However, collaboration among embodied agents introduces significant challenges: agents must not only perceive and understand their environment but also communicate, share information, divide tasks, and coordinate actions responsively. For example, consider autonomous robots in a home responding to a request: \"Bring me the iPad and apple, and put the milk in the refrigerator.\" Achieving this requires multi-level coordination, including optimal search strategies, task prioritization, and efficient movement planning." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 363, + 556, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 363, + 556, + 460 + ], + "spans": [ + { + "bbox": [ + 313, + 363, + 556, + 460 + ], + "type": "text", + "content": "Large language models (LLMs) have recently provided embodied agents with advanced natural language understanding, dialogue, and reasoning abilities [1, 3, 38, 40]. These capabilities allow LLMs to decompose complex and long-term tasks into a sequence of manageable sub-goals, making LLM-driven agents a promising alternative to traditional reinforcement learning models [14, 34] which are difficult to train and often generalize poorly [7, 21, 39]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 460, + 556, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 460, + 556, + 675 + ], + "spans": [ + { + "bbox": [ + 313, + 460, + 556, + 675 + ], + "type": "text", + "content": "However, enabling embodied agents to work together in decentralized environments remains a significant and underexplored challenge, as it requires long-term planning and coherent decision-making to coordinate actions efficiently. Early attempts at multi-agent collaboration, such as CoELA [43] and RoCo [24], demonstrate progress but also reveal limitations. For instance, as shown in Fig. 1 (b), CoELA facilitates collaboration by sharing updates through natural language when sub-tasks are completed, yet each agent's decision-making process remains independent, resulting in suboptimal coordination. In contrast, as illustrated in Fig. 1 (c), RoCo develops multi-agent work plans through agent discussions and environmental interactions. However, RoCo relies on a single reasoning path, which is susceptible to the randomness and unpredictability of LLM outputs. This can lead to inefficient or incorrect planning, particularly in applications requiring precision, where such errors can have serious consequences." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "text", + "content": "Contributions. To address these limitations, we propose Cooperative Tree Search (CoTS), a framework designed to significantly improve collaborative planning and task ex" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 703, + 145, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 145, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 145, + 713 + ], + "type": "text", + "content": "*Corresponding author." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "29513" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 61, + 64, + 312, + 74 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 64, + 312, + 74 + ], + "spans": [ + { + "bbox": [ + 61, + 64, + 312, + 74 + ], + "type": "text", + "content": "Task: transport 2 oranges, 3 apples, 1 banana, 3 breads, 1 burger to the bed." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 60, + 74, + 534, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 74, + 534, + 95 + ], + "spans": [ + { + "bbox": [ + 60, + 74, + 534, + 95 + ], + "type": "text", + "content": "Background: Agents randomly appear in a home environment with a total of six rooms and need to discover objects through exploration. Each agent can carry up to two objects. Finding and using containers can improve efficiency." + } + ] + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 57, + 99, + 226, + 220 + ], + "blocks": [ + { + "bbox": [ + 57, + 99, + 226, + 220 + ], + "lines": [ + { + "bbox": [ + 57, + 99, + 226, + 220 + ], + "spans": [ + { + "bbox": [ + 57, + 99, + 226, + 220 + ], + "type": "image", + "image_path": "51f8faae0a245a38a0bcbb01d92263bdf6aa223be74feae5bfb9102f14896f94.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 73, + 226, + 127, + 243 + ], + "lines": [ + { + "bbox": [ + 73, + 226, + 127, + 243 + ], + "spans": [ + { + "bbox": [ + 73, + 226, + 127, + 243 + ], + "type": "text", + "content": "(a) The Embodied Environment" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 135, + 224, + 236, + 242 + ], + "lines": [ + { + "bbox": [ + 135, + 224, + 236, + 242 + ], + "spans": [ + { + "bbox": [ + 135, + 224, + 236, + 242 + ], + "type": "text", + "content": "(b) CoELA (Communication but independent decision-making.)" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 228, + 99, + 383, + 220 + ], + "blocks": [ + { + "bbox": [ + 228, + 99, + 383, + 220 + ], + "lines": [ + { + "bbox": [ + 228, + 99, + 383, + 220 + ], + "spans": [ + { + "bbox": [ + 228, + 99, + 383, + 220 + ], + "type": "image", + "image_path": "0b6e08abed85d204b9d2583ea6cc8aba813591ec8e0cc70b32599d38f8cb576f.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 255, + 226, + 360, + 243 + ], + "lines": [ + { + "bbox": [ + 255, + 226, + 360, + 243 + ], + "spans": [ + { + "bbox": [ + 255, + 226, + 360, + 243 + ], + "type": "text", + "content": "(c) RoCo (Discussion, Planning and Execution)" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 384, + 99, + 547, + 224 + ], + "blocks": [ + { + "bbox": [ + 384, + 99, + 547, + 224 + ], + "lines": [ + { + "bbox": [ + 384, + 99, + 547, + 224 + ], + "spans": [ + { + "bbox": [ + 384, + 99, + 547, + 224 + ], + "type": "image", + "image_path": "822a2c81f6949651bd472a1ffd8758b0757df6f8bb5e0a26c80dec6d849e2ba0.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 395, + 225, + 534, + 242 + ], + "lines": [ + { + "bbox": [ + 395, + 225, + 534, + 242 + ], + "spans": [ + { + "bbox": [ + 395, + 225, + 534, + 242 + ], + "type": "text", + "content": "(d) Ours (Plan Searching, Evaluation, and Plan-Guided Execution)" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 251, + 555, + 307 + ], + "lines": [ + { + "bbox": [ + 55, + 251, + 555, + 307 + ], + "spans": [ + { + "bbox": [ + 55, + 251, + 555, + 307 + ], + "type": "text", + "content": "Figure 1. The task execution processes of CoELA, RoCo, and our CoTS. (1) In CoELA, agents update each other on sub-task completion and individually determine their next actions without long-term strategic planning. (2) RoCo uses multi-round dialogues and summarizes the plan at the end, but its single-dialogue chain can lead to suboptimal plans and even errors due to the randomness and unpredictability of LLMs. (3) Within a modified Monte Carlo Tree Search framework, CoTS allows agents to evaluate multiple planning paths, enabling them to collaboratively deliberate and identify the most effective course of action based on task progress and environmental context." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 318, + 294, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 318, + 294, + 389 + ], + "spans": [ + { + "bbox": [ + 54, + 318, + 294, + 389 + ], + "type": "text", + "content": "cution efficiency among embodied agents. Inspired by human decision-making, CoTS allows agents to search multiple planning paths by combining exploration and backtracking, enabling them to collaboratively deliberate and identify the most effective course of action based on task progress and environmental context." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 390, + 295, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 390, + 295, + 617 + ], + "spans": [ + { + "bbox": [ + 55, + 390, + 295, + 617 + ], + "type": "text", + "content": "As illustrated in Fig. 1 (d), CoTS advances agent collaboration in embodied environments through two core innovations. First, we introduce a collaborative planning module based on a modified Monte Carlo Tree Search (MCTS) framework, which enables agents to explore multiple planning paths and select the most coherent and efficient long-term plan. Unlike traditional MCTS which relies on simulation-based rewards, CoTS incorporates a non-simulated LLM-based reward function suited for high-cost physical movement interactions in embodied environments. In this way, CoTS transforms the traditional linear sequence of action steps into a search process, exploring potential paths within a shared reasoning and action space to develop robust collaborative strategies. Second, CoTS features a plan evaluation module, that actively monitors the relevance and feasibility of the selected plan during execution. This mechanism allows agents to adapt plans as needed, reducing unnecessary disruptions from frequent plan updates while still accommodating environmental changes." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 618, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 295, + 714 + ], + "type": "text", + "content": "Our experiments on two long-term multi-objective planning tasks — TDW-MAT [43] and C-WAH [8] — demonstrate that CoTS significantly outperforms current state-of-the-art methods, e.g., CoELA [43] and RoCo [24], in terms of task completion rate and efficiency. These results highlight the effectiveness and superiority of CoTS, underscoring its potential for high-stakes applications that demand reliable multi-agent collaboration." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 316, + 400, + 329 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 316, + 400, + 329 + ], + "spans": [ + { + "bbox": [ + 314, + 316, + 400, + 329 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 340, + 555, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 340, + 555, + 532 + ], + "spans": [ + { + "bbox": [ + 313, + 340, + 555, + 532 + ], + "type": "text", + "content": "LLM-based Agents. Recent advances in LLMs have unlocked near-human reasoning and planning capabilities [1, 27, 37]. Researchers have leveraged LLMs as the \"brain\" or controller of agents, expanding their perceptual and action capacities through multimodal perception and tool use [6, 25, 32, 38, 40, 45]. Multi-agent systems [16, 35] show particular promise, enabling agents to collaborate on planning, discussion, and decision-making—similar to how humans solve complex problems in teams [26, 33]. Works such as MetaGPT [13], CAMEL [19], and ChatDev [31] decompose complex tasks into simpler, manageable subtasks to reduce hallucinations and improve task-solving capabilities. However, embodied multi-agent systems [20, 44] require advanced communication and face significant challenges in navigating a vast action search space in complex, dynamic environments, especially for long-term tasks." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "type": "text", + "content": "Embodied Multi-Agent Cooperation. Research on embodied multi-agent systems is still emerging. Habitat [30] explores agents' social awareness, enabling adaptive assistance strategies in collaboration. Co-NavGPT [41] uses LLMs for multi-robot collaborative goal navigation, encoding environmental data to improve scene understanding. Organized-LLM-Agents [11] reduces redundancy in multiagent collaboration through structured prompts, enhancing coordination. CoELA [43] enables collaborative agents to communicate and plan through natural language and chain-of-thought processes, yet lacks strategic long-term planning. RoCo [24] and CaPo [22] advance multi-agent collaboration with high-level communication and path planning, but their single reasoning path often leads to suboptimal plans and even errors. Although some existing work" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "29514" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 85, + 70, + 518, + 257 + ], + "blocks": [ + { + "bbox": [ + 85, + 70, + 518, + 257 + ], + "lines": [ + { + "bbox": [ + 85, + 70, + 518, + 257 + ], + "spans": [ + { + "bbox": [ + 85, + 70, + 518, + 257 + ], + "type": "image", + "image_path": "232b84dc39248ff7fdd9552a6c86729adc9372537500dbd2137b559f4f78a0bb.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 258, + 555, + 314 + ], + "lines": [ + { + "bbox": [ + 54, + 258, + 555, + 314 + ], + "spans": [ + { + "bbox": [ + 54, + 258, + 555, + 314 + ], + "type": "text", + "content": "Figure 2. The framework of CoTS. Upon receiving a new task in the embodied environment, each embodied agent initially employs the Perception Module to gather raw observational data, which is then stored in its Memory Module. In the collaborative planning module, agents guide the exploration of collaborative plans toward more promising directions through the collaborative tree search. The explored plans are retained within the memory module as a special 'extended memory' and are periodically updated based on assessments from the planning evaluation module. Finally, agents perform actions aligned with the collaborative plan and tasks." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 323, + 296, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 323, + 296, + 371 + ], + "spans": [ + { + "bbox": [ + 55, + 323, + 296, + 371 + ], + "type": "text", + "content": "on LLMs involves multi-branch reasoning [2, 17, 40, 42], they are inadequate for complex embodied environments and agent collaboration. Our goal is to develop embodied agents capable of searching long-term strategic planning." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 380, + 141, + 393 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 380, + 141, + 393 + ], + "spans": [ + { + "bbox": [ + 55, + 380, + 141, + 393 + ], + "type": "text", + "content": "3. Preliminaries" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 402, + 295, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 402, + 295, + 628 + ], + "spans": [ + { + "bbox": [ + 56, + 402, + 295, + 628 + ], + "type": "text", + "content": "Task Assumption. As introduced by previous work [9, 24], multi-agent collaborative tasks can be viewed as a Markov game, which is defined by a tuple " + }, + { + "bbox": [ + 56, + 402, + 295, + 628 + ], + "type": "inline_equation", + "content": "(\\mathcal{N},\\mathcal{S},\\mathcal{A},P,r,\\gamma)" + }, + { + "bbox": [ + 56, + 402, + 295, + 628 + ], + "type": "text", + "content": ", in which " + }, + { + "bbox": [ + 56, + 402, + 295, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{N}" + }, + { + "bbox": [ + 56, + 402, + 295, + 628 + ], + "type": "text", + "content": " is the set of agents (e.g. Alice and Bob), " + }, + { + "bbox": [ + 56, + 402, + 295, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{S}" + }, + { + "bbox": [ + 56, + 402, + 295, + 628 + ], + "type": "text", + "content": " is the state space, " + }, + { + "bbox": [ + 56, + 402, + 295, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{A} = \\prod_{i=1}^{n}\\mathcal{A}^i" + }, + { + "bbox": [ + 56, + 402, + 295, + 628 + ], + "type": "text", + "content": " is the combination of all action spaces, " + }, + { + "bbox": [ + 56, + 402, + 295, + 628 + ], + "type": "inline_equation", + "content": "P: \\mathcal{S} \\times \\mathcal{A} \\times \\mathcal{S} \\to [0,1]" + }, + { + "bbox": [ + 56, + 402, + 295, + 628 + ], + "type": "text", + "content": " is the transition probability function, " + }, + { + "bbox": [ + 56, + 402, + 295, + 628 + ], + "type": "inline_equation", + "content": "r: \\mathcal{S} \\times \\mathcal{A} \\to \\mathbb{R}" + }, + { + "bbox": [ + 56, + 402, + 295, + 628 + ], + "type": "text", + "content": " is the reward score, and " + }, + { + "bbox": [ + 56, + 402, + 295, + 628 + ], + "type": "inline_equation", + "content": "\\gamma \\in [0,1)" + }, + { + "bbox": [ + 56, + 402, + 295, + 628 + ], + "type": "text", + "content": " is the discount factor. In the task, each agent at time step " + }, + { + "bbox": [ + 56, + 402, + 295, + 628 + ], + "type": "inline_equation", + "content": "t \\in \\mathbb{N}" + }, + { + "bbox": [ + 56, + 402, + 295, + 628 + ], + "type": "text", + "content": " obtains the information by perception module from the embodied environment " + }, + { + "bbox": [ + 56, + 402, + 295, + 628 + ], + "type": "inline_equation", + "content": "s_t \\in \\mathcal{S}" + }, + { + "bbox": [ + 56, + 402, + 295, + 628 + ], + "type": "text", + "content": ", chooses and executes an action " + }, + { + "bbox": [ + 56, + 402, + 295, + 628 + ], + "type": "inline_equation", + "content": "a_t^i \\in \\mathcal{A}^i" + }, + { + "bbox": [ + 56, + 402, + 295, + 628 + ], + "type": "text", + "content": " from its available policy " + }, + { + "bbox": [ + 56, + 402, + 295, + 628 + ], + "type": "inline_equation", + "content": "\\pi^i(\\cdot|s_t)" + }, + { + "bbox": [ + 56, + 402, + 295, + 628 + ], + "type": "text", + "content": ". This will along with the actions of other agents, constitutes a joint action " + }, + { + "bbox": [ + 56, + 402, + 295, + 628 + ], + "type": "inline_equation", + "content": "\\mathbf{a}_t = (a_t^1,a_t^2,\\dots,a_t^n) \\in \\mathcal{A}" + }, + { + "bbox": [ + 56, + 402, + 295, + 628 + ], + "type": "text", + "content": " derived from the joint policy " + }, + { + "bbox": [ + 56, + 402, + 295, + 628 + ], + "type": "inline_equation", + "content": "\\pi(\\cdot|s_t) = \\prod_{i=1}^{n}\\pi^i(\\cdot|s_t)" + }, + { + "bbox": [ + 56, + 402, + 295, + 628 + ], + "type": "text", + "content": ". Then agents perceive a new state " + }, + { + "bbox": [ + 56, + 402, + 295, + 628 + ], + "type": "inline_equation", + "content": "s_{t+1}" + }, + { + "bbox": [ + 56, + 402, + 295, + 628 + ], + "type": "text", + "content": " with probability " + }, + { + "bbox": [ + 56, + 402, + 295, + 628 + ], + "type": "inline_equation", + "content": "P(s_{t+1}|s_t,\\mathbf{a}_t)" + }, + { + "bbox": [ + 56, + 402, + 295, + 628 + ], + "type": "text", + "content": ". We seek to determine a collaborative plan that maximizes the " + }, + { + "bbox": [ + 56, + 402, + 295, + 628 + ], + "type": "inline_equation", + "content": "J(\\pi) \\triangleq \\mathbb{E}_{s_0,\\infty} \\sim P,a_0,\\infty \\sim \\pi[\\sum_{i=0}^{\\infty}\\gamma^i r_i]" + }, + { + "bbox": [ + 56, + 402, + 295, + 628 + ], + "type": "text", + "content": ". For this paper, the goal is to coordinate the distribution of sub-tasks among agents so as to complete tasks more efficiently." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "text", + "content": "CoELA [43]. It contains five modules: 1) Perception, 2) Memory, 3) Communication, 4) Planning and 5) Execution. The Perception module leverages a trained Mask-RCNN [12] to segment RGB images and detect objects and obstacles, creating 3D point clouds from RGB-D data and extracting essential high-level information like key object states. The Memory module manages short-term and long-" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 323, + 555, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 323, + 555, + 479 + ], + "spans": [ + { + "bbox": [ + 313, + 323, + 555, + 479 + ], + "type": "text", + "content": "term memory, with short-term memory capturing immediate details such as semantic maps and agent states, and long-term memory storing persistent information like task descriptions. In the Communication module, CoELA retrieves relevant information from memory, transforming it into template-based textual descriptions for agents to aid in action selection. However, each agent plans independently, with no unified action coordination. In the Planning and Execution modules, CoELA retrieves agent-related information from memory, generates potential actions for the LLM to consider, and applies chain-of-thought [38] reasoning for more in-depth decision-making. Each agent then executes its chosen action accordingly." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 493, + 552, + 507 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 493, + 552, + 507 + ], + "spans": [ + { + "bbox": [ + 313, + 493, + 552, + 507 + ], + "type": "text", + "content": "4. Collaborative Tree Search for Multi-Agents" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 514, + 424, + 526 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 514, + 424, + 526 + ], + "spans": [ + { + "bbox": [ + 313, + 514, + 424, + 526 + ], + "type": "text", + "content": "4.1. Overview of CoTS" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 533, + 555, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 555, + 652 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 555, + 652 + ], + "type": "text", + "content": "Now we introduce an overview of our Collaborative Tree Search (CoTS), which is designed for multi-agent cooperation in embodied environments. As shown in Fig. 2, CoTS contains four key modules: 1) Perception Module, 2) Memory Module, 3) Collaborative Planning Module, and 4) Plan Parsing and Execution Module. This work primarily enhances CoELA in its planning by designing the Collaborative Planning Module. The other three modules are largely borrowed from CoELA and introduced in Sec. 3. Here we briefly introduce our enhancements to CoEAL below." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "text", + "content": "Perception Module. Unlike CoELA, we introduce distance as a key perceptual factor, leveraging fixed spatial layouts in real-world settings to enhance agents' room-based understanding and improve planning efficiency by using LLM to access the moving distance of agents." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "text", + "content": "29515" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 63, + 74, + 302, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 74, + 302, + 83 + ], + "spans": [ + { + "bbox": [ + 63, + 74, + 302, + 83 + ], + "type": "text", + "content": "Task: transport 2 oranges, 3 apples, 1 banana, 3 breads, 1 burger to the bed" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 63, + 87, + 341, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 87, + 341, + 96 + ], + "spans": [ + { + "bbox": [ + 63, + 87, + 341, + 96 + ], + "type": "text", + "content": "The Process in Each Node: Alice makes plans, Bob provides suggestions and rewards." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 63, + 101, + 301, + 110 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 101, + 301, + 110 + ], + "spans": [ + { + "bbox": [ + 63, + 101, + 301, + 110 + ], + "type": "text", + "content": "Plan from Alice: Step 1: Alice grasps apple. Bob explore his current room." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 63, + 111, + 340, + 119 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 111, + 340, + 119 + ], + "spans": [ + { + "bbox": [ + 63, + 111, + 340, + 119 + ], + "type": "text", + "content": "Step 2: Alice explore the Living room, if they find any items, transport them to the bed." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 63, + 121, + 367, + 129 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 121, + 367, + 129 + ], + "spans": [ + { + "bbox": [ + 63, + 121, + 367, + 129 + ], + "type": "text", + "content": "Message to Bob: Hi Bob, I am in Livingroom where I found an apple. Do you have suggestions." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 63, + 133, + 327, + 151 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 133, + 327, + 151 + ], + "spans": [ + { + "bbox": [ + 63, + 133, + 327, + 151 + ], + "type": "text", + "content": "Message to Alice: Hi Alice, I am in Livingroom and found bread. As you said, I am just about to explore my current room. Reward: 0.3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 63, + 158, + 315, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 158, + 315, + 167 + ], + "spans": [ + { + "bbox": [ + 63, + 158, + 315, + 167 + ], + "type": "text", + "content": "Plan from Alice: Step 1: Alice moves to Bedroom. Bob grasps bread and apple." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 63, + 167, + 253, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 167, + 253, + 176 + ], + "spans": [ + { + "bbox": [ + 63, + 167, + 253, + 176 + ], + "type": "text", + "content": "Step 2: Alice explores Bedroom. Bob explores Living room." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 63, + 177, + 260, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 177, + 260, + 186 + ], + "spans": [ + { + "bbox": [ + 63, + 177, + 260, + 186 + ], + "type": "text", + "content": "Step 3: If any target item are found, transport them to the bed." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 63, + 186, + 352, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 186, + 352, + 205 + ], + "spans": [ + { + "bbox": [ + 63, + 186, + 352, + 205 + ], + "type": "text", + "content": "Message to Bob: Hi Bob, Please grasp bread and apple and I will now shift to the Bedroom because we should not explore the same room." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 63, + 205, + 366, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 205, + 366, + 213 + ], + "spans": [ + { + "bbox": [ + 63, + 205, + 366, + 213 + ], + "type": "text", + "content": "Message to Alice: Hi Alice, but I think you could head to kitchen, it is the nearest. Reward: 0.6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 63, + 220, + 285, + 229 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 220, + 285, + 229 + ], + "spans": [ + { + "bbox": [ + 63, + 220, + 285, + 229 + ], + "type": "text", + "content": "Plan from Alice: Step 1: Alice grasps the apple. Bob grasps the bread." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 63, + 230, + 249, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 230, + 249, + 239 + ], + "spans": [ + { + "bbox": [ + 63, + 230, + 249, + 239 + ], + "type": "text", + "content": "Step 2: Alice moves to Kitchen. Bob explores Living room." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 63, + 239, + 326, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 239, + 326, + 248 + ], + "spans": [ + { + "bbox": [ + 63, + 239, + 326, + 248 + ], + "type": "text", + "content": "Step 3: Alice explore the Kitchen, if they find any items, transport them to the bed." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 63, + 250, + 307, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 250, + 307, + 258 + ], + "spans": [ + { + "bbox": [ + 63, + 250, + 307, + 258 + ], + "type": "text", + "content": "Message to Bob: Hi Bob, I agree with you. I will grasp the apple before I mo" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 63, + 259, + 333, + 269 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 259, + 333, + 269 + ], + "spans": [ + { + "bbox": [ + 63, + 259, + 333, + 269 + ], + "type": "text", + "content": "Kitchen while you continue in Livingroom. Please keep me updated on your progress" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 63, + 270, + 317, + 280 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 270, + 317, + 280 + ], + "spans": [ + { + "bbox": [ + 63, + 270, + 317, + 280 + ], + "type": "text", + "content": "Message to Alice: Hi, Alice, your plan looks good. Let's get moving. Reward: 1" + } + ] + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 375, + 82, + 552, + 269 + ], + "blocks": [ + { + "bbox": [ + 375, + 82, + 552, + 269 + ], + "lines": [ + { + "bbox": [ + 375, + 82, + 552, + 269 + ], + "spans": [ + { + "bbox": [ + 375, + 82, + 552, + 269 + ], + "type": "image", + "image_path": "49f12660ec44c9334be6776b9704c999105e257b7f8ddb2a24ea2df01f7b0f32.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 287, + 555, + 332 + ], + "lines": [ + { + "bbox": [ + 55, + 287, + 555, + 332 + ], + "spans": [ + { + "bbox": [ + 55, + 287, + 555, + 332 + ], + "type": "text", + "content": "Figure 3. Overview of the tree search in CoTS. Each node in the tree contains content generated by Alice and Bob. Alice generates the collaborative plans and messages to Bob, while Bob generates messages to Alice and rewards for the plans. These rewards are used for backpropagation to guide the tree search toward more promising collaborative plans. The figure also shows examples of three nodes, illustrating the stepwise optimization of collaborative planning and the dialogue process between agents." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "bbox": [ + 54, + 342, + 295, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 342, + 295, + 450 + ], + "spans": [ + { + "bbox": [ + 54, + 342, + 295, + 450 + ], + "type": "text", + "content": "Memory Module. Compared with CoELA, CoTS introduces \"extended memory\" (Fig. 2) to manage collaboration plans—information that needs to be retained for a moderate period and updated periodically. This is because collaboration plans must balance retention for a certain period and timely updates when they become unsuitable for the current embodied environment. This extended memory mimics the medium-term storage of the human memory system, bridging short-term recall and long-term consolidation." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 54, + 451, + 295, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 451, + 295, + 569 + ], + "spans": [ + { + "bbox": [ + 54, + 451, + 295, + 569 + ], + "type": "text", + "content": "Collaborative Planning Module is the core of CoTS, where MCTS is modified and employed to develop long-term strategic collaboration plans. Inspired by human problem-solving through iterative discussion and cooperation, it avoids error accumulation caused by a single path approach. This module represents the first integration of embodied agent collaboration with MCTS, enhancing agents' capability to tackle complex, long-term tasks through tree-based search and backtracking. Details of this module are provided in Sec. 4.2." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 55, + 570, + 295, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 295, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 295, + 605 + ], + "type": "text", + "content": "Plan Parsing and Execution Module translates collaborative plans into executable actions, enabling agents to complete tasks in a coordinated manner. See details in Sec. 4.3." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 55, + 606, + 295, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 606, + 295, + 666 + ], + "spans": [ + { + "bbox": [ + 55, + 606, + 295, + 666 + ], + "type": "text", + "content": "Below we will delve deeper into two core modules, Collaborative Planning Module and Plan Parsing and Execution Module. Additionally, we will highlight CoTS's integration of spatial awareness and extended memory to enhance overall functionality and efficiency." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 55, + 672, + 225, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 672, + 225, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 672, + 225, + 685 + ], + "type": "text", + "content": "4.2. Collaborative Planning Module" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "content": "Our Collaborative Planning Module, central to our CoTS framework, is designed to search for long-term, efficient" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 342, + 555, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 342, + 555, + 403 + ], + "spans": [ + { + "bbox": [ + 313, + 342, + 555, + 403 + ], + "type": "text", + "content": "and coherent collaborative plans for multi-agent tasks in dynamic environments. It comprises two main sub-modules: a collaborative tree search module for generating collaborative plans (Sec. 4.2.1) and a plan evaluation module to analyze when to update the plan (Sec. 4.2.2)." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 411, + 547, + 422 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 411, + 547, + 422 + ], + "spans": [ + { + "bbox": [ + 313, + 411, + 547, + 422 + ], + "type": "text", + "content": "4.2.1. Monte Carlo Tree Search for Collaborative Plan" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 426, + 555, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 555, + 676 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 555, + 676 + ], + "type": "text", + "content": "We adapt MCTS to convert multi-agent collaborative planning into a tree search, identifying an optimal pathway from multiple reasoning paths for cohesive strategies. As shown in Fig. 3, each node in this tree stores a proposed plan, agent feedback, and a performance reward. In this setup, one agent (like Alice) functions as the \"planner\" to propose plans, while other agents (like Bob) act as \"critics\". Alice drafts several initial collaborative plans or extends existing plan based on task goals and current task progress. Then Bob evaluates them, considering factors like task allocation and movement efficiency, to generate a reward for the current plan. Once feedback is provided, each plan's reward is propagated back through the collaboration tree, strengthening successful strategies and adjusting future planning pathways. Accordingly, they will select the most efficient collaborative plan for further reflections and refinements. As shown in Fig. 3, this process, including plan extension (drafting), evaluation, backpropagation, and selection, will repeat until agents explore all potential collaborative plans and select the most promising one. In the following, for these four steps, we will elaborate on each of them." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "content": "Selection Similar to traditional MCTS [4, 15], CoTS starts from the root node and traverses down the tree by selecting nodes with the highest UCT values. This process is repeated" + } + ] + } + ], + "index": 28 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "29516" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 109 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 109 + ], + "type": "text", + "content": "until a leaf node is reached, and is illustrated in Fig. 3 where we mark the selected nodes with gold cups. For UCT of each node, it is computed as follows:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 88, + 115, + 295, + 131 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 115, + 295, + 131 + ], + "spans": [ + { + "bbox": [ + 88, + 115, + 295, + 131 + ], + "type": "interline_equation", + "content": "\\mathrm {U C T} (s _ {i}) = V (s _ {i}) + w \\sqrt {\\ln N (p _ {i}) / N (s _ {i})}, \\tag {1}", + "image_path": "c9ce998111deb8b9ffa481bab5f610ff35f49d6dac1e455ec741afec6e291e4b.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 137, + 295, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 137, + 295, + 213 + ], + "spans": [ + { + "bbox": [ + 55, + 137, + 295, + 213 + ], + "type": "text", + "content": "Where " + }, + { + "bbox": [ + 55, + 137, + 295, + 213 + ], + "type": "inline_equation", + "content": "N(s_{i})" + }, + { + "bbox": [ + 55, + 137, + 295, + 213 + ], + "type": "text", + "content": " is visiting number of node " + }, + { + "bbox": [ + 55, + 137, + 295, + 213 + ], + "type": "inline_equation", + "content": "s_i" + }, + { + "bbox": [ + 55, + 137, + 295, + 213 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 137, + 295, + 213 + ], + "type": "inline_equation", + "content": "V(s_{i})" + }, + { + "bbox": [ + 55, + 137, + 295, + 213 + ], + "type": "text", + "content": " is the reward of the subtree rooted at " + }, + { + "bbox": [ + 55, + 137, + 295, + 213 + ], + "type": "inline_equation", + "content": "s_i" + }, + { + "bbox": [ + 55, + 137, + 295, + 213 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 137, + 295, + 213 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 55, + 137, + 295, + 213 + ], + "type": "text", + "content": " is the exploration coefficient, and " + }, + { + "bbox": [ + 55, + 137, + 295, + 213 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 55, + 137, + 295, + 213 + ], + "type": "text", + "content": " is the parent node of " + }, + { + "bbox": [ + 55, + 137, + 295, + 213 + ], + "type": "inline_equation", + "content": "s_i" + }, + { + "bbox": [ + 55, + 137, + 295, + 213 + ], + "type": "text", + "content": ". After an episode, backpropagation updates each node's " + }, + { + "bbox": [ + 55, + 137, + 295, + 213 + ], + "type": "inline_equation", + "content": "V(s_{i})" + }, + { + "bbox": [ + 55, + 137, + 295, + 213 + ], + "type": "text", + "content": " along the path: " + }, + { + "bbox": [ + 55, + 137, + 295, + 213 + ], + "type": "inline_equation", + "content": "V(s_{i}) = \\frac{V(s_{i - 1})\\cdot(N(s_{i}) - 1) + r}{N(s_{i})}" + }, + { + "bbox": [ + 55, + 137, + 295, + 213 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 137, + 295, + 213 + ], + "type": "inline_equation", + "content": "V(s_{i - 1})" + }, + { + "bbox": [ + 55, + 137, + 295, + 213 + ], + "type": "text", + "content": " is the previous reward, and " + }, + { + "bbox": [ + 55, + 137, + 295, + 213 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 55, + 137, + 295, + 213 + ], + "type": "text", + "content": " is the reward used to update each " + }, + { + "bbox": [ + 55, + 137, + 295, + 213 + ], + "type": "inline_equation", + "content": "V(s_{i})" + }, + { + "bbox": [ + 55, + 137, + 295, + 213 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 213, + 295, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 213, + 295, + 307 + ], + "spans": [ + { + "bbox": [ + 55, + 213, + 295, + 307 + ], + "type": "text", + "content": "Expansion. Once a node is selected, we expand it by sampling " + }, + { + "bbox": [ + 55, + 213, + 295, + 307 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 213, + 295, + 307 + ], + "type": "text", + "content": " new discussion paths, each building on the plan (from \"planner\", Alice) and suggestions (from \"critic\", Bob) in the parent node. This generates " + }, + { + "bbox": [ + 55, + 213, + 295, + 307 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 213, + 295, + 307 + ], + "type": "text", + "content": " new child nodes, representing potential planning routes. The tree, along with its nodes, is stored in an external long-term memory to retain valuable information and facilitate consistent planning across episodes." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 308, + 295, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 308, + 295, + 498 + ], + "spans": [ + { + "bbox": [ + 55, + 308, + 295, + 498 + ], + "type": "text", + "content": "Evaluation. Each newly created child node receives a reward " + }, + { + "bbox": [ + 55, + 308, + 295, + 498 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 55, + 308, + 295, + 498 + ], + "type": "text", + "content": " to quantify the quality of the collaboration plan and to guide the search algorithm towards the most promising areas of the tree. Here CoTS incorporates a non-simulated LLM-based reward function to estimate the reward " + }, + { + "bbox": [ + 55, + 308, + 295, + 498 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 55, + 308, + 295, + 498 + ], + "type": "text", + "content": " of a plan in a node. In traditional MCTS, reward estimation requires agents to execute their actions for receiving environment feedback, like completing a subtask, and thus is high-cost. So CoTS introduces an LLM-based reward function based on two components: (1) an allocation reward assessing coherence and efficiency, and (2) a cost reward evaluating the spatial movement cost of agents. To ensure consistency, we standardize the evaluation criteria to provide clear guidelines to minimize ambiguity. Each criterion is scored on a scale of 1 to 5, with the final reward normalized between 0 and 1. See examples in Appendix B.1." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 499, + 295, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 499, + 295, + 622 + ], + "spans": [ + { + "bbox": [ + 55, + 499, + 295, + 622 + ], + "type": "text", + "content": "Backpropagation. Each evaluation reward is propagated back along the path to update node rewards, with formulas ensuring that parent nodes integrate the feedback, guiding future selections. Specifically, the back-propagation updates node rewards " + }, + { + "bbox": [ + 55, + 499, + 295, + 622 + ], + "type": "inline_equation", + "content": "N(s_{i}) = N(s_{i - 1}) + 1" + }, + { + "bbox": [ + 55, + 499, + 295, + 622 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 499, + 295, + 622 + ], + "type": "inline_equation", + "content": "V(s_{i}) = \\frac{V(s_{i - 1})N(s_{i - 1}) + r}{N(s_i)}" + }, + { + "bbox": [ + 55, + 499, + 295, + 622 + ], + "type": "text", + "content": " for each node " + }, + { + "bbox": [ + 55, + 499, + 295, + 622 + ], + "type": "inline_equation", + "content": "s_0,s_1,\\ldots ,s_p" + }, + { + "bbox": [ + 55, + 499, + 295, + 622 + ], + "type": "text", + "content": " along the path from the root (initial state " + }, + { + "bbox": [ + 55, + 499, + 295, + 622 + ], + "type": "inline_equation", + "content": "s_0" + }, + { + "bbox": [ + 55, + 499, + 295, + 622 + ], + "type": "text", + "content": ") to leaf (terminal state " + }, + { + "bbox": [ + 55, + 499, + 295, + 622 + ], + "type": "inline_equation", + "content": "s_p" + }, + { + "bbox": [ + 55, + 499, + 295, + 622 + ], + "type": "text", + "content": "). Accordingly these estimated rewards and visiting number are used in \"Expansion\" step for node selection, directing the search toward the most promising areas of the tree." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 627, + 282, + 639 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 627, + 282, + 639 + ], + "spans": [ + { + "bbox": [ + 55, + 627, + 282, + 639 + ], + "type": "text", + "content": "4.2.2. Collaborative Plan Evaluation and Adaptation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "type": "text", + "content": "In this section, we will explain why we introduced the planning evaluation module (a sub-module in the collaborative planning module, as shown in Fig. 2) and its function. The collaborative plan provides high-level embodied agent action guidance, but as the task is executed and agents act, the embodied environment gradually changes (e.g. the agents'" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 553, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 107 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 107 + ], + "type": "text", + "content": "positions, the agents' progress, and the current status, etc.). Therefore, the collaborative plan needs to be updated based on the latest environmental information." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 108, + 555, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 108, + 555, + 262 + ], + "spans": [ + { + "bbox": [ + 313, + 108, + 555, + 262 + ], + "type": "text", + "content": "At this point, two scenarios arise: If the plan is updated too slowly or not updated at all, the current plan becomes outdated and cannot effectively guide the embodied agents in executing tasks. However, in a multi-agent system, instantly updating the plan with every progress can be inefficient and counterproductive. It may lead to excessive queries to the LLM and environments. Furthermore, frequent plan revisions may introduce conflicting strategies due to minor environmental shifts, potentially disrupting agent actions and lowering overall efficiency. This is illustrated in Fig. 4 (c): if the updated plan conflicts with the action that Bob is currently performing, it may interrupt Bob's action which even almost finishes the subtask." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 264, + 556, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 264, + 556, + 407 + ], + "spans": [ + { + "bbox": [ + 313, + 264, + 556, + 407 + ], + "type": "text", + "content": "To address this, we introduce a plan evaluation module within CoTS to continuously evaluate the viability of the current collaborative plan. When an agent makes progress, the module determines if the existing plan remains effective based on the status of task execution. If the evaluation metric meets a defined threshold, indicating that the plan is still viable, it is retained. Otherwise, CoTS will call the collaborative planning module to re-search and update the plan. By focusing on the feasibility and long-term stability of the plan rather than continuous recalibration, this module supports efficient, coherent collaboration, enabling agents to complete tasks reliably in a dynamic environment." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 415, + 506, + 427 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 415, + 506, + 427 + ], + "spans": [ + { + "bbox": [ + 313, + 415, + 506, + 427 + ], + "type": "text", + "content": "4.3. Plan Parsing and Execution Module" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 431, + 555, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 431, + 555, + 552 + ], + "spans": [ + { + "bbox": [ + 313, + 431, + 555, + 552 + ], + "type": "text", + "content": "To transform the collaborative plan generated in Sec. 4.2 into executable actions, CoTS incorporates an action parsing module and a plan decomposition module. Action parsing module bridges high-level planning with real-world actions, generating a set of feasible actions for each agent based on current memory. For example, if an agent's hands are free, it will list objects that can be grabbed; if the agent is in an unknown room, it will list movement options. This module ensures that all sub-plans are formatted into natural language and returned for subsequent process." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 552, + 554, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 552, + 554, + 622 + ], + "spans": [ + { + "bbox": [ + 313, + 552, + 554, + 622 + ], + "type": "text", + "content": "Plan decomposition module understands and breaks down high-level tasks from the collaboration plan into specific and prioritized actions. The agent interprets the task arrangement, selects actions from the parsed options that best advance the collaborative plan's goals, and finally translates each chosen action into real-world execution." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 634, + 395, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 634, + 395, + 647 + ], + "spans": [ + { + "bbox": [ + 313, + 634, + 395, + 647 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 654, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 554, + 714 + ], + "type": "text", + "content": "**Benchmarks. Following CoELA [43], we evaluate the performance of CoTS in two highly challenging embodied-world multi-agent environments: (1) TDW-MAT (ThreeD World Multi-Agent Transport) [8] extends the ThreeD World Transport Challenge by simulating collaborative" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "29517" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 82, + 109, + 529, + 231 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 105 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 105 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 105 + ], + "type": "text", + "content": "Table 1. Quantitative results on TDW-MAT. Consistent with CoELA [43], TDW-MAT has two environment settings: with oracle perception and without oracle perception. The results marked with * are from CoELA [43]. MAT uses central observation and oracle perception. " + }, + { + "bbox": [ + 55, + 70, + 555, + 105 + ], + "type": "inline_equation", + "content": "91_{+9.64\\%}" + }, + { + "bbox": [ + 55, + 70, + 555, + 105 + ], + "type": "text", + "content": " means the relative improvement when comparing with CoELA is " + }, + { + "bbox": [ + 55, + 70, + 555, + 105 + ], + "type": "inline_equation", + "content": "9.64\\%" + }, + { + "bbox": [ + 55, + 70, + 555, + 105 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 82, + 109, + 529, + 231 + ], + "lines": [ + { + "bbox": [ + 82, + 109, + 529, + 231 + ], + "spans": [ + { + "bbox": [ + 82, + 109, + 529, + 231 + ], + "type": "table", + "html": "
Transport Rate (%)Base AgentRL AgentLLM Agent
RHP*CoRHP*MAT*GPT 3.5GPT 4
CoELARoCoCaPoCoTS(ours)CoELA*RoCoCaPoCoTS(ours)
TDW-MAT w/ Oracle Perception
Stuff(↑)4974177380+9.68%84+15.07%81+10.96%8385+2.41%87+4.82%91+9.64%
Food(↑)5276137274+2.78%85+18.06%83+15.28%8789+2.30%90+3.45%94+8.05%
Average(↑)5075157377+5.48%84+15.07%82+12.33%8587+2.35%89+4.71%93+9.41%
TDW-MAT w/o Oracle Perception
Stuff(↑)3654-3947+20.51%45+15.38%50+28.21%6164+4.92%64+4.92%77+26.22%
Food(↑)4967-6764-4.48%70+4.48%67+0.00%8280-2.44%85+3.65%88+7.32%
Average(↑)4361-5255+5.77%57+9.62%59+13.46%7172+1.41%74+4.23%83+16.90%
", + "image_path": "a61b2d34aedad11f1ce58f63840f94b2af4f9dc4112254e58c4318cf587bad5d.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 69, + 274, + 280, + 360 + ], + "blocks": [ + { + "bbox": [ + 55, + 248, + 295, + 270 + ], + "lines": [ + { + "bbox": [ + 55, + 248, + 295, + 270 + ], + "spans": [ + { + "bbox": [ + 55, + 248, + 295, + 270 + ], + "type": "text", + "content": "Table 2. Quantitative results on C-WAH. LLM-based agents are driven by GPT 4. Results marked with * are from CoELA [43]." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 69, + 274, + 280, + 360 + ], + "lines": [ + { + "bbox": [ + 69, + 274, + 280, + 360 + ], + "spans": [ + { + "bbox": [ + 69, + 274, + 280, + 360 + ], + "type": "table", + "html": "
Average StepVisual Obs(↓)Symbolic Obs(↓)
Base AgentMHP*141111
CoMHP*10375
LLM AgentCoELA*9257
RoCo89+3.26%57+0.00%
CaPo83+9.78%51+10.53%
CoTS(ours)81+11.96%49+14.04%
", + "image_path": "3b496952cb4010d9fd9cc36461b2345cb286a40b77a6380041743e27666a57be.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 374, + 295, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 374, + 295, + 470 + ], + "spans": [ + { + "bbox": [ + 54, + 374, + 295, + 470 + ], + "type": "text", + "content": "transportation tasks in a physics-based household environment. We select 24 tasks, split evenly between food and stuff transportation. (2) C-WAH (Communicative Watch-And-Help) [29] extends the Watch-And-Help challenge within the VirtualHome-Social platform [28]. It features five common household scenarios. Our test set includes 10 tasks, with two tasks per scenario. The objective is to complete all sub-goals, prioritizing faster completion." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 472, + 295, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 472, + 295, + 544 + ], + "spans": [ + { + "bbox": [ + 55, + 472, + 295, + 544 + ], + "type": "text", + "content": "Evaluation Metrics. On TDW-MAT, transport rate is evaluated by the percentage of target objects successfully delivered to their designated locations within a 3,000-frame timeframe. For C-WAH, efficiency is measured by the number of time steps needed to complete all assigned tasks, where fewer steps indicate higher efficiency." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "text", + "content": "Baseline Methods. The baselines include three types of embodied agents: (1) Base Agent: RHP [8] uses heuristic rules for navigation and " + }, + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{A}^*" + }, + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "text", + "content": " for path planning. MHP [18] employs Monte Carlo Tree Search (MCTS) for high-level planning coupled with regression planning for task execution (2) RL-Based Agent: MAT [39] is a multi-agent reinforcement learning framework, featuring a centralized decision transformer that synthesizes actions from shared observations. (3) LLM-Based Agent: CoELA [43] is a leading collaboration framework for embodied agents that integrates the reasoning capabilities of LLMs. It facilitates communication and interaction between agents through natural language. RoCo [24] and CaPo [22] are multi-agent collaboration methods that enhance collaboration by gener" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 250, + 555, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 250, + 555, + 324 + ], + "spans": [ + { + "bbox": [ + 313, + 250, + 555, + 324 + ], + "type": "text", + "content": "ating coordinated plans through structured agent dialogues. Setup. We conduct all experiments on an Intel Core i9- 10980XE CPU and an NVIDIA GeForce RTX 3090 GPU. We utilize both GPT-4 and GPT-3.5-Turbo [1], configuring them with a temperature setting of 0.7 and top-1 sampling to produce outputs up to 512 tokens." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 333, + 400, + 346 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 333, + 400, + 346 + ], + "spans": [ + { + "bbox": [ + 313, + 333, + 400, + 346 + ], + "type": "text", + "content": "5.1. Main Results" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 352, + 555, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 352, + 555, + 483 + ], + "spans": [ + { + "bbox": [ + 313, + 352, + 555, + 483 + ], + "type": "text", + "content": "CoTS Outperforms the Strong Baselines. Following the experimental frameworks of CoELA [43] and RoCo [24], we configure two-agent cooperation, excluding RHP. As demonstrated in Table 1, CoTS achieves best performance on the TDW-MAT task. Specifically, when powered by GPT-4 within the Oracle Perception environment, CoTS surpasses CoELA, achieving a " + }, + { + "bbox": [ + 313, + 352, + 555, + 483 + ], + "type": "inline_equation", + "content": "9.41\\%" + }, + { + "bbox": [ + 313, + 352, + 555, + 483 + ], + "type": "text", + "content": " average relative improvement in task completion, and a " + }, + { + "bbox": [ + 313, + 352, + 555, + 483 + ], + "type": "inline_equation", + "content": "16.9\\%" + }, + { + "bbox": [ + 313, + 352, + 555, + 483 + ], + "type": "text", + "content": " improvement in the absence of Oracle Perception. The primary strength of CoTS is its ability to generate superior long-term strategies through collaborative tree search." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 485, + 556, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 485, + 556, + 700 + ], + "spans": [ + { + "bbox": [ + 313, + 485, + 556, + 700 + ], + "type": "text", + "content": "Conversely, CoELA focuses on short-term actions derived from immediate information, which limits its strategic depth and adaptability to future scenarios and environmental shifts. RoCo and CaPo are similar. Although they are capable of developing plans, they rely on a single-branch discussion method, which may result in suboptimal or inaccurate strategies. Moreover, Their frequent plan revisions can interfere with agent coordination. Therefore, while RoCo and CaPo surpass CoELA, they are inferior to CoTS. When driven by GPT-3.5, CaPo slightly outperforms CoTS on some metrics. The main reason for this is that GPT-3.5 has weaker evaluation and scoring capabilities. However, when powered by GPT-4, CoTS significantly outperforms CaPo. Our findings underscore that multi-agent collaboration generally surpasses single-agent strategies, exemplified by RHP. The performance of reinforcement learning-based MAT is poor, mainly because TDW-MAT is too difficult for it with long time domain and sparse reward signals." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 326, + 701, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 701, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 326, + 701, + 553, + 714 + ], + "type": "text", + "content": "The performance of LLM-based agents primarily de" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "29518" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 63, + 74, + 205, + 106 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 74, + 205, + 106 + ], + "spans": [ + { + "bbox": [ + 63, + 74, + 205, + 106 + ], + "type": "text", + "content": "(a) CoELA leads to short-sighted action plans that ignore precious action steps" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 63, + 109, + 197, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 109, + 197, + 140 + ], + "spans": [ + { + "bbox": [ + 63, + 109, + 197, + 140 + ], + "type": "text", + "content": "Background: Alice is in living room holding a container. Bob finds two target objects in living room " + }, + { + "bbox": [ + 63, + 109, + 197, + 140 + ], + "type": "inline_equation", + "content": "< 1000>" + }, + { + "bbox": [ + 63, + 109, + 197, + 140 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 61, + 142, + 212, + 239 + ], + "blocks": [ + { + "bbox": [ + 61, + 142, + 212, + 239 + ], + "lines": [ + { + "bbox": [ + 61, + 142, + 212, + 239 + ], + "spans": [ + { + "bbox": [ + 61, + 142, + 212, + 239 + ], + "type": "image", + "image_path": "7ec545d557c91cb677e0ac6a1a31ce96a613fb553992580942afa15368db28e0.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 239, + 203, + 268 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 239, + 203, + 268 + ], + "spans": [ + { + "bbox": [ + 62, + 239, + 203, + 268 + ], + "type": "text", + "content": "CoELA's plan : Alice goes to the living room to help Bob, because Alice can take these objects." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 269, + 201, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 269, + 201, + 308 + ], + "spans": [ + { + "bbox": [ + 61, + 269, + 201, + 308 + ], + "type": "text", + "content": "Our plan: Alice explores the current room, find objects, then transport them to the bedroom. Then go to the living room " + }, + { + "bbox": [ + 61, + 269, + 201, + 308 + ], + "type": "inline_equation", + "content": "< 1000>" + }, + { + "bbox": [ + 61, + 269, + 201, + 308 + ], + "type": "text", + "content": ", no actions are wasted." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 231, + 73, + 371, + 106 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 231, + 73, + 371, + 106 + ], + "spans": [ + { + "bbox": [ + 231, + 73, + 371, + 106 + ], + "type": "text", + "content": "(b) RoCo's single-chain dialogue leads to suboptimal collaboration plan" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 230, + 110, + 369, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 230, + 110, + 369, + 140 + ], + "spans": [ + { + "bbox": [ + 230, + 110, + 369, + 140 + ], + "type": "text", + "content": "Background: Alice and Bob appear in the same living room and now need to explore some rooms to find objects." + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 229, + 140, + 383, + 239 + ], + "blocks": [ + { + "bbox": [ + 229, + 140, + 383, + 239 + ], + "lines": [ + { + "bbox": [ + 229, + 140, + 383, + 239 + ], + "spans": [ + { + "bbox": [ + 229, + 140, + 383, + 239 + ], + "type": "image", + "image_path": "df07c36f0467d2e805c8a4847c030012552c8294be636c605e3c962eddf7ad6e.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 230, + 239, + 373, + 259 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 230, + 239, + 373, + 259 + ], + "spans": [ + { + "bbox": [ + 230, + 239, + 373, + 259 + ], + "type": "text", + "content": "RoCo's plan : Alice explore the current room. Bob explore the current room." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 228, + 268, + 381, + 310 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 228, + 268, + 381, + 310 + ], + "spans": [ + { + "bbox": [ + 228, + 268, + 381, + 310 + ], + "type": "text", + "content": "Our plan: Alice goes to the nearest room, Bedroom " + }, + { + "bbox": [ + 228, + 268, + 381, + 310 + ], + "type": "inline_equation", + "content": "< 2000>" + }, + { + "bbox": [ + 228, + 268, + 381, + 310 + ], + "type": "text", + "content": ", while Bob explores the current room. Exploring rooms separately would improve efficiency." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 403, + 75, + 538, + 106 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 403, + 75, + 538, + 106 + ], + "spans": [ + { + "bbox": [ + 403, + 75, + 538, + 106 + ], + "type": "text", + "content": "(c) Frequent plan updates may hinder completion of current actions" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 403, + 109, + 533, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 403, + 109, + 533, + 140 + ], + "spans": [ + { + "bbox": [ + 403, + 109, + 533, + 140 + ], + "type": "text", + "content": "Background: Bob was transporting objects to the bed, Alice found a bread in the living room<6000>" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 400, + 140, + 548, + 239 + ], + "blocks": [ + { + "bbox": [ + 400, + 140, + 548, + 239 + ], + "lines": [ + { + "bbox": [ + 400, + 140, + 548, + 239 + ], + "spans": [ + { + "bbox": [ + 400, + 140, + 548, + 239 + ], + "type": "image", + "image_path": "061d1fe605d5d158f3c45e816e2ca1490761704a4290c947d3a587adc4e3c613.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 400, + 239, + 547, + 268 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 400, + 239, + 547, + 268 + ], + "spans": [ + { + "bbox": [ + 400, + 239, + 547, + 268 + ], + "type": "text", + "content": "RoCo's Plan: Bob goes to the living room " + }, + { + "bbox": [ + 400, + 239, + 547, + 268 + ], + "type": "inline_equation", + "content": "< 6000>" + }, + { + "bbox": [ + 400, + 239, + 547, + 268 + ], + "type": "text", + "content": " to help Alice grasp the bread." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 400, + 268, + 545, + 309 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 400, + 268, + 545, + 309 + ], + "spans": [ + { + "bbox": [ + 400, + 268, + 545, + 309 + ], + "type": "text", + "content": "Our plan: The current plan is still valid. Bob is close to the bedroom. He can first deliver the objects in his hand to the bed, and then help Alice." + } + ] + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 56, + 350, + 167, + 430 + ], + "blocks": [ + { + "bbox": [ + 55, + 316, + 555, + 338 + ], + "lines": [ + { + "bbox": [ + 55, + 316, + 555, + 338 + ], + "spans": [ + { + "bbox": [ + 55, + 316, + 555, + 338 + ], + "type": "text", + "content": "Figure 4. Some interesting example cases demonstrate that our CoTS achieves better collaboration than RoCo and CoELA. These scenes are taken from the experiments in TDW-MAT." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 56, + 350, + 167, + 430 + ], + "lines": [ + { + "bbox": [ + 56, + 350, + 167, + 430 + ], + "spans": [ + { + "bbox": [ + 56, + 350, + 167, + 430 + ], + "type": "image", + "image_path": "416f57abde0d9bc8ea29cbe41eebe0e3a9a49735b8e634b2cf408935b176fec2.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 59, + 431, + 164, + 470 + ], + "lines": [ + { + "bbox": [ + 59, + 431, + 164, + 470 + ], + "spans": [ + { + "bbox": [ + 59, + 431, + 164, + 470 + ], + "type": "text", + "content": "There seems to be no plan. Both Alice and Bob are not holding anything. Bob in bedroom and has found an container, while Alice is in the. A collaboration plan needs to be updated. Reward: 1" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 173, + 350, + 290, + 430 + ], + "blocks": [ + { + "bbox": [ + 173, + 350, + 290, + 430 + ], + "lines": [ + { + "bbox": [ + 173, + 350, + 290, + 430 + ], + "spans": [ + { + "bbox": [ + 173, + 350, + 290, + 430 + ], + "type": "image", + "image_path": "67ebbf0492b3fb860a8c8644426bd3bed0427552740056ca7d3d2f47705ecba0.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 173, + 431, + 290, + 471 + ], + "lines": [ + { + "bbox": [ + 173, + 431, + 290, + 471 + ], + "spans": [ + { + "bbox": [ + 173, + 431, + 290, + 471 + ], + "type": "text", + "content": "Alice have found an apple in the fridge and Bob is holding a cupcake. Current progress is consistent with the collaborative plan and future actions are outlined in the plan. There is no need to update the plan. Reward: 4" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 18 + }, + { + "bbox": [ + 55, + 541, + 295, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 541, + 295, + 662 + ], + "spans": [ + { + "bbox": [ + 55, + 541, + 295, + 662 + ], + "type": "text", + "content": "pends on their reasoning abilities and the depth of embedded world knowledge. Agents powered by GPT-4, as demonstrated by CoTS, significantly surpass those based on GPT-3.5, highlighting the advancements in LLM technologies and their impact on agent proficiency. As shown in Table 2, CoTS achieves superior efficiency, improving by " + }, + { + "bbox": [ + 55, + 541, + 295, + 662 + ], + "type": "inline_equation", + "content": "11.96\\%" + }, + { + "bbox": [ + 55, + 541, + 295, + 662 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 541, + 295, + 662 + ], + "type": "inline_equation", + "content": "14.04\\%" + }, + { + "bbox": [ + 55, + 541, + 295, + 662 + ], + "type": "text", + "content": " over CoELA in Visual Obs and Symbolic Obs settings, respectively. These results underscore CoTS's robust adaptability across various embodied environments and its potential for real-world applications." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 55, + 671, + 173, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 671, + 173, + 684 + ], + "spans": [ + { + "bbox": [ + 55, + 671, + 173, + 684 + ], + "type": "text", + "content": "5.2. Qualitative Analysis" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "content": "CoTS Can Develop Long-term Strategic Plans. Unlike CoELA, which is adept at responding to immediate tasks" + } + ] + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 342, + 352, + 524, + 451 + ], + "blocks": [ + { + "bbox": [ + 55, + 475, + 295, + 519 + ], + "lines": [ + { + "bbox": [ + 55, + 475, + 295, + 519 + ], + "spans": [ + { + "bbox": [ + 55, + 475, + 295, + 519 + ], + "type": "text", + "content": "Figure 5. Plan Evaluation and Update. The plan evaluation module gives a lower reward when it determines that the current plan needs to be updated, and a higher reward when it believes the current plan is still valid. The reward range is from 1 to 5." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 342, + 352, + 524, + 451 + ], + "lines": [ + { + "bbox": [ + 342, + 352, + 524, + 451 + ], + "spans": [ + { + "bbox": [ + 342, + 352, + 524, + 451 + ], + "type": "image", + "image_path": "51da022b49258ffbf5632dc1659b5bacbabb95bdc0c5f68b8ea8931500179902.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 454, + 555, + 477 + ], + "lines": [ + { + "bbox": [ + 313, + 454, + 555, + 477 + ], + "spans": [ + { + "bbox": [ + 313, + 454, + 555, + 477 + ], + "type": "text", + "content": "Figure 6. Comparison of transportation rates for GPT-4-based CoTS and CoELA at different time steps on TDW-MAT." + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 498, + 555, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 555, + 594 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 555, + 594 + ], + "type": "text", + "content": "but falters in sustained coordination, CoTS excels in long-term strategic planning. As illustrated in Fig. 4 (a), when Bob discovers a new target that requires assistance, Alice, under CoTS's direction, continues to thoroughly explore her current room before moving the item to the bed. This systematic approach not only supports Bob but also adheres to an overarching collaborative strategy, optimizing both immediate and future task alignments." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 594, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 556, + 714 + ], + "type": "text", + "content": "CoTS Generates More Optimal Collaboration Plans. Unlike RoCo to generate plans sequentially through a single-chain dialogue, CoTS makes more optimized collaboration plans. Due to the randomness of LLMs and the embodied environment's complexity, RoCo's plans may sometimes be suboptimal. For example, in Fig. 4 (b), RoCo instructs both Alice and Bob to explore their respective rooms. However, they are currently in the same room, making joint exploration redundant. CoTS effectively identifies this and assigns Alice to explore the nearest room instead." + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "29519" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 76, + 291, + 186 + ], + "blocks": [ + { + "bbox": [ + 62, + 76, + 291, + 186 + ], + "lines": [ + { + "bbox": [ + 62, + 76, + 291, + 186 + ], + "spans": [ + { + "bbox": [ + 62, + 76, + 291, + 186 + ], + "type": "image", + "image_path": "cbda7706dce07bdf1b38b22867ba2e33c5cfaf3247cd8abf647981a304ef0b4c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 191, + 295, + 226 + ], + "lines": [ + { + "bbox": [ + 55, + 191, + 295, + 226 + ], + "spans": [ + { + "bbox": [ + 55, + 191, + 295, + 226 + ], + "type": "text", + "content": "Figure 7. Comparison of transportation rates with different branching factors. More branching factors means more new nodes are generated each time in the expansion phase." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 234, + 295, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 234, + 295, + 354 + ], + "spans": [ + { + "bbox": [ + 55, + 234, + 295, + 354 + ], + "type": "text", + "content": "CoTS Ensures Smooth Completion of Agent Actions. In RoCo, new progress often prompt discussions and lead to new plans that can override existing ones. As illustrated in Fig. 4 (c), Bob, tasked with transporting items to the bed, must halt his activity to help Alice following a revised plan. Conversely, CoTS incorporates a plan evaluation module that validates the feasibility of plans before any adjustments are made. In Fig. 4 (c), CoTS allows Bob to finish transporting before next action. This approach avoids chaotic agent actions and ensures timely plan updates when necessary." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 354, + 295, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 354, + 295, + 486 + ], + "spans": [ + { + "bbox": [ + 55, + 354, + 295, + 486 + ], + "type": "text", + "content": "CoTS Dynamically Manages Collaborative Plans. As shown in Fig. 5, the assessment module continuously evaluates the viability of existing plans. It assigns a lower reward (e.g. 1) if collaboration necessitates devising a new plan or modifying an existing one, prompting agents to adapt their strategies accordingly. Conversely, when the current progress aligns with planned tasks and their distribution among agents is deemed efficient, the module rewards this alignment with a higher value (e.g. 4). This incentivizes agents to persist with the established collaborative plan, avoiding unnecessary replanning." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 493, + 264, + 506 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 493, + 264, + 506 + ], + "spans": [ + { + "bbox": [ + 55, + 493, + 264, + 506 + ], + "type": "text", + "content": "5.3. Ablation Study and Additional Analysis" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 510, + 295, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 295, + 653 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 295, + 653 + ], + "type": "text", + "content": "The Core Modules of CoTS Have Important Impacts. The Table 3 shows the ablation results of CoTS. To examine the impact of tree search, we modify it to a single-branch structure, where agents can only communicate along one path. The experimental results indicate a " + }, + { + "bbox": [ + 55, + 510, + 295, + 653 + ], + "type": "inline_equation", + "content": "5.38\\%" + }, + { + "bbox": [ + 55, + 510, + 295, + 653 + ], + "type": "text", + "content": " relative decrease in the average transportation rate, mainly because the limited communication does not consistently yield correct conclusions and diverges from typical human discussion patterns. Furthermore, eliminating the reward function and plan update evaluation components from CoTS leads to variable performance declines, highlighting the critical roles these elements play in the system's efficacy." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 654, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 295, + 714 + ], + "type": "text", + "content": "Analysis of Agent Collaboration Progress at Different Time Steps. We analyze the transportation rates of CoELA and CoTS on TDW-MAT at various time steps, as depicted in Fig. 6. An interesting observation emerges at 500 steps, where CoTS exhibits a lower performance compared to" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 328, + 97, + 541, + 165 + ], + "blocks": [ + { + "bbox": [ + 313, + 71, + 553, + 93 + ], + "lines": [ + { + "bbox": [ + 313, + 71, + 553, + 93 + ], + "spans": [ + { + "bbox": [ + 313, + 71, + 553, + 93 + ], + "type": "text", + "content": "Table 3. Effects of the reward function, collaboration tree search, and plan evaluation module in GPT-4-driven CoTS on TDW-MAT." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 328, + 97, + 541, + 165 + ], + "lines": [ + { + "bbox": [ + 328, + 97, + 541, + 165 + ], + "spans": [ + { + "bbox": [ + 328, + 97, + 541, + 165 + ], + "type": "table", + "html": "
Transport RateStuff(↑)Food(↑)Average(↑)
CoTS w/o Reward87-4.40%90-4.26%89-4.30%
CoTS w/o MCTS Plan85-6.59%91-3.19%88-5.38%
CoTS w/o E&U88-3.30%92-2.13%90-3.23%
CoTS919493
", + "image_path": "0e32c318cfdd26fe09e10f2a61d9265eca0fb6931bd4f3f4cdf12ed20de707db.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 184, + 553, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 184, + 553, + 255 + ], + "spans": [ + { + "bbox": [ + 313, + 184, + 553, + 255 + ], + "type": "text", + "content": "CoELA, attributed to CoTS's strategy of prioritizing item collection early to enhance transport capacity. However, this initial disadvantage is counterbalanced by a significant improvement at 2000 steps, where CoTS markedly surpasses CoELA. This shift underscores CoTS's effectiveness in optimizing agent collaboration for task execution." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 256, + 553, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 256, + 553, + 339 + ], + "spans": [ + { + "bbox": [ + 313, + 256, + 553, + 339 + ], + "type": "text", + "content": "Impact of Different Branching Factors on Results. During the expansion phase of MCTS, we define the number of new candidate nodes generated at each step as the \"branching factors.\" As shown in Fig. 7, increasing the number of branching factors expands the selection space, leading to a better plan. Adjusting the branching factor allows for a balance between performance and efficiency." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 340, + 553, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 340, + 553, + 376 + ], + "spans": [ + { + "bbox": [ + 313, + 340, + 553, + 376 + ], + "type": "text", + "content": "Besides, the human-agent collaboration and the cost of CoTS are discussed in Appendix A.1 and Appendix A.2, respectively. CoTS also demonstrates significant advantages." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 387, + 391, + 399 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 387, + 391, + 399 + ], + "spans": [ + { + "bbox": [ + 313, + 387, + 391, + 399 + ], + "type": "text", + "content": "6. Conclusions" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 407, + 553, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 407, + 553, + 526 + ], + "spans": [ + { + "bbox": [ + 313, + 407, + 553, + 526 + ], + "type": "text", + "content": "This paper introduces CoTS, a collaborative framework that leverages large language models with a modified Monte Carlo tree search to enhance multi-agent task planning in complex and dynamic environments. By enabling agents to evaluate multiple potential plans and refine strategies based on progress, CoTS significantly advances over prior methods like CoELA and RoCo. The addition of a plan feasibility module ensures smooth task execution by adapting plans only when necessary. Experimental results on multi-objective tasks confirm CoTS's superior performance." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 527, + 553, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 527, + 553, + 598 + ], + "spans": [ + { + "bbox": [ + 313, + 527, + 553, + 598 + ], + "type": "text", + "content": "Limitation. CoTS relies on LLMs' reasoning abilities. We find that GPT-4 surpasses GPT-3.5-turbo in instruction-following and response quality, while GPT-3.5-turbo can be verbose and occasionally inaccurate. Despite this, even for GPT-3.5, CoTS improves agent collaboration through collaborative planning search in the collaboration tree." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 609, + 423, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 609, + 423, + 624 + ], + "spans": [ + { + "bbox": [ + 313, + 609, + 423, + 624 + ], + "type": "text", + "content": "7. Acknowledgments" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 629, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 553, + 713 + ], + "type": "text", + "content": "This work is supported by National Natural Science Foundation of China (U2133202, 52305570), China Postdoctoral Science Foundation (2022M720955), and Outstanding Doctoral Dissertation Funding Project of Heilongjiang Province (LJYXL2022-011). Lizheng Zu also acknowledges the support of the China Scholarship Council program (202406120111)." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "29520" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "text", + "content": "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 1, 2, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 146, + 296, + 223 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 146, + 296, + 223 + ], + "spans": [ + { + "bbox": [ + 61, + 146, + 296, + 223 + ], + "type": "text", + "content": "[2] David Brandfonbrener, Simon Henniger, Sibi Raja, Tarun Prasad, Chloe R Loughridge, Federico Cassano, Sabrina Ruixin Hu, Jianang Yang, William E Byrd, Robert Zinkov, et al. Vermcts: Synthesizing multi-step programs using a verifier, a large language model, and tree search. In The 4th Workshop on Mathematical Reasoning and AI at NeurIPS'24, 2024. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 224, + 294, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 224, + 294, + 246 + ], + "spans": [ + { + "bbox": [ + 62, + 224, + 294, + 246 + ], + "type": "text", + "content": "[3] Tom B Brown. Language models are few-shot learners. arXiv preprint arXiv:2005.14165, 2020. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 247, + 295, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 247, + 295, + 312 + ], + "spans": [ + { + "bbox": [ + 62, + 247, + 295, + 312 + ], + "type": "text", + "content": "[4] Cameron B Browne, Edward Powley, Daniel Whitehouse, Simon M Lucas, Peter I Cowling, Philipp Rohlfshagen, Stephen Tavener, Diego Perez, Spyridon Samothrakis, and Simon Colton. A survey of monte carlo tree search methods. IEEE Transactions on Computational Intelligence and AI in games, 4(1):1-43, 2012. 4, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 314, + 295, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 314, + 295, + 357 + ], + "spans": [ + { + "bbox": [ + 62, + 314, + 295, + 357 + ], + "type": "text", + "content": "[5] Abhinav Dahiya, Alexander M Aroyo, Kerstin Dautenhahn, and Stephen L Smith. A survey of multi-agent human-robot interaction systems. Robotics and Autonomous Systems, 161: 104335, 2023. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 358, + 295, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 358, + 295, + 402 + ], + "spans": [ + { + "bbox": [ + 62, + 358, + 295, + 402 + ], + "type": "text", + "content": "[6] Norman Di Palo, Arunkumar Byravan, Leonard Hasenclever, Markus Wulfmeier, Nicolas Heess, and Martin Riedmiller. Towards a unified agent with foundation models. arXiv preprint arXiv:2307.09668, 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 403, + 295, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 403, + 295, + 446 + ], + "spans": [ + { + "bbox": [ + 62, + 403, + 295, + 446 + ], + "type": "text", + "content": "[7] Amal Feriani and Ekram Hossain. Single and multi-agent deep reinforcement learning for ai-enabled wireless networks: A tutorial. IEEE Communications Surveys & Tutorial, 23(2):1226-1252, 2021. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 448, + 295, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 448, + 295, + 534 + ], + "spans": [ + { + "bbox": [ + 62, + 448, + 295, + 534 + ], + "type": "text", + "content": "[8] Chuang Gan, Siyuan Zhou, Jeremy Schwartz, Seth Alter, Abhishek Bhandwaldar, Dan Gutfreund, Daniel LK Yamins, James J DiCarlo, Josh McDermott, Antonio Torralba, et al. The threadworld transport challenge: A visually guided task-and-motion planning benchmark towards physically realistic embodied ai. In 2022 International conference on robotics and automation (ICRA), pages 8847-8854. IEEE, 2022. 2, 5, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 536, + 295, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 536, + 295, + 601 + ], + "spans": [ + { + "bbox": [ + 62, + 536, + 295, + 601 + ], + "type": "text", + "content": "[9] Ran Gong, Qiuyuan Huang, Xiaojian Ma, Yusuke Noda, Zane Durante, Zilong Zheng, Demetri Terzopoulos, Li Fei-Fei, Jianfeng Gao, and Hoi Vo. Mindagent: Emergent gaming interaction. In Findings of the Association for Computational Linguistics: NAACL 2024, pages 3154–3183, 2024. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 603, + 295, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 603, + 295, + 656 + ], + "spans": [ + { + "bbox": [ + 57, + 603, + 295, + 656 + ], + "type": "text", + "content": "[10] Taicheng Guo, Xiuying Chen, Yaqi Wang, Ruidi Chang, Shichao Pei, Nitesh V Chawla, Olaf Wiest, and Xiangliang Zhang. Large language model based multi-agents: A survey of progress and challenges. arXiv preprint arXiv:2402.01680, 2024. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 658, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 658, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 658, + 295, + 712 + ], + "type": "text", + "content": "[11] Xudong Guo, Kaixuan Huang, Jiale Liu, Wenhui Fan, Natalia Vélez, Qingyun Wu, Huazheng Wang, Thomas L Griffiths, and Mengdi Wang. Embodied llm agents learn to cooperate in organized teams. arXiv preprint arXiv:2403.12482, 2024. 2" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 554, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 106 + ], + "type": "text", + "content": "[12] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 317, + 107, + 554, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 107, + 554, + 161 + ], + "spans": [ + { + "bbox": [ + 317, + 107, + 554, + 161 + ], + "type": "text", + "content": "[13] Sirui Hong, Xiawu Zheng, Jonathan Chen, Yuheng Cheng, Jinlin Wang, Ceyao Zhang, Zili Wang, Steven Ka Shing Yau, Zijuan Lin, Liyang Zhou, et al. Metagpt: Meta programming for multi-agent collaborative framework. arXiv preprint arXiv:2308.00352, 2023. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 162, + 554, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 162, + 554, + 215 + ], + "spans": [ + { + "bbox": [ + 317, + 162, + 554, + 215 + ], + "type": "text", + "content": "[14] Wenlong Huang, Pieter Abbeel, Deepak Pathak, and Igor Mordatch. Language models as zero-shot planners: Extracting actionable knowledge for embodied agents. In International conference on machine learning, pages 9118-9147. PMLR, 2022. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 217, + 553, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 217, + 553, + 249 + ], + "spans": [ + { + "bbox": [ + 316, + 217, + 553, + 249 + ], + "type": "text", + "content": "[15] Steven James, George Konidaris, and Benjamin Rosman. An analysis of monte carlo tree search. In Proceedings of the AAAI Conference on Artificial Intelligence, 2017. 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 250, + 553, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 250, + 553, + 293 + ], + "spans": [ + { + "bbox": [ + 317, + 250, + 553, + 293 + ], + "type": "text", + "content": "[16] Feibo Jiang, Yubo Peng, Li Dong, Kezhi Wang, Kun Yang, Cunhua Pan, Dusit Niyato, and Octavia A Dobre. Large language model enhanced multi-agent systems for 6g communications. IEEE Wireless Communications, 2024. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 294, + 553, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 294, + 553, + 348 + ], + "spans": [ + { + "bbox": [ + 317, + 294, + 553, + 348 + ], + "type": "text", + "content": "[17] Bowen Jin, Chulin Xie, Jiawei Zhang, Kashob Kumar Roy, Yu Zhang, Zheng Li, Ruirui Li, Xianfeng Tang, Suhang Wang, Yu Meng, et al. Graph chain-of-thought: Augmenting large language models by reasoning on graphs. arXiv preprint arXiv:2404.07103, 2024. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 350, + 553, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 350, + 553, + 371 + ], + "spans": [ + { + "bbox": [ + 316, + 350, + 553, + 371 + ], + "type": "text", + "content": "[18] Richard E Korf. Planning as search: A quantitative approach. Artificial intelligence, 33(1):65-88, 1987. 6, 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 372, + 553, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 372, + 553, + 425 + ], + "spans": [ + { + "bbox": [ + 317, + 372, + 553, + 425 + ], + "type": "text", + "content": "[19] Guohao Li, Hasan Hammoud, Hani Itani, Dmitrii Khizbullin, and Bernard Ghanem. Camel: Communicative agents for\" mind\" exploration of large language model society. Advances in Neural Information Processing Systems, 36: 51991-52008, 2023. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 426, + 553, + 447 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 426, + 553, + 447 + ], + "spans": [ + { + "bbox": [ + 316, + 426, + 553, + 447 + ], + "type": "text", + "content": "[20] Xin Li. Positioning is all you need. arXiv preprint arXiv:2404.01183, 2024. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 449, + 553, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 449, + 553, + 492 + ], + "spans": [ + { + "bbox": [ + 317, + 449, + 553, + 492 + ], + "type": "text", + "content": "[21] Iou-Jen Liu, Unnat Jain, Raymond A Yeh, and Alexander Schwing. Cooperative exploration for multi-agent deep reinforcement learning. In International conference on machine learning, pages 6826-6836. PMLR, 2021. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 493, + 553, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 493, + 553, + 547 + ], + "spans": [ + { + "bbox": [ + 317, + 493, + 553, + 547 + ], + "type": "text", + "content": "[22] Jie Liu, Pan Zhou, Yingjun Du, Ah-Hwee Tan, Cees GM Snoek, Jan-Jakob Sonke, and Efstratios Gavves. Capo: Cooperative plan optimization for efficient embodied multiagent cooperation. arXiv preprint arXiv:2411.04679, 2024. 2, 6, 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 548, + 553, + 603 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 548, + 553, + 603 + ], + "spans": [ + { + "bbox": [ + 317, + 548, + 553, + 603 + ], + "type": "text", + "content": "[23] Yang Liu, Weixing Chen, Yongjie Bai, Jingzhou Luo, Xinshuai Song, Kaixuan Jiang, Zhida Li, Ganlong Zhao, Junyi Lin, Guanbin Li, et al. Aligning cyber space with physical world: A comprehensive survey on embodied ai. arXiv preprint arXiv:2407.06886, 2024. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 604, + 553, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 604, + 553, + 656 + ], + "spans": [ + { + "bbox": [ + 317, + 604, + 553, + 656 + ], + "type": "text", + "content": "[24] Zhao Mandi, Shreeya Jain, and Shuran Song. Roco: Dialectic multi-robot collaboration with large language models. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 286–299. IEEE, 2024. 1, 2, 3, 6, 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 658, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 658, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 658, + 553, + 713 + ], + "type": "text", + "content": "[25] Reiichiro Nakano, Jacob Hilton, Suchir Balaji, Jeff Wu, Long Ouyang, Christina Kim, Christopher Hesse, Shantanu Jain, Vineet Kosaraju, William Saunders, et al. Webgpt: Browser-assisted question-answering with human feedback. arXiv preprint arXiv:2112.09332, 2021. 2" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "29521" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "text", + "content": "[26] Bo Ni and Markus J Buehler. Mechagents: Large language model multi-agent collaborations can solve mechanics problems, generate new data, and integrate knowledge. Extreme Mechanics Letters, 67:102131, 2024. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 184 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 184 + ], + "type": "text", + "content": "[27] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. Advances in neural information processing systems, 35:27730-27744, 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 186, + 294, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 186, + 294, + 239 + ], + "spans": [ + { + "bbox": [ + 56, + 186, + 294, + 239 + ], + "type": "text", + "content": "[28] Xavier Puig, Kevin Ra, Marko Boben, Jiaman Li, Tingwu Wang, Sanja Fidler, and Antonio Torralba. Virtualhome: Simulating household activities via programs. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8494-8502, 2018. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 242, + 294, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 242, + 294, + 296 + ], + "spans": [ + { + "bbox": [ + 56, + 242, + 294, + 296 + ], + "type": "text", + "content": "[29] Xavier Puig, Tianmin Shu, Shuang Li, Zilin Wang, Yuan-Hong Liao, Joshua B Tenenbaum, Sanja Fidler, and Antonio Torralba. Watch-and-help: A challenge for social perception and human-ai collaboration. arXiv preprint arXiv:2010.09890, 2020. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 298, + 294, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 298, + 294, + 352 + ], + "spans": [ + { + "bbox": [ + 56, + 298, + 294, + 352 + ], + "type": "text", + "content": "[30] Xavier Puig, Eric Undersander, Andrew Szot, Mikael Dallaire Cote, Tsung-Yen Yang, Ruslan Partsey, Ruta Desai, Alexander William Clegg, Michal Hlavac, So Yeon Min, et al. Habitat 3.0: A co-habitat for humans, avatars and robots. arXiv preprint arXiv:2310.13724, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 354, + 294, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 354, + 294, + 419 + ], + "spans": [ + { + "bbox": [ + 56, + 354, + 294, + 419 + ], + "type": "text", + "content": "[31] Chen Qian, Wei Liu, Hongzhang Liu, Nuo Chen, Yufan Dang, Jiahao Li, Cheng Yang, Weize Chen, Yusheng Su, Xin Cong, et al. Chatdev: Communicative agents for software development. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 15174-15186, 2024. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 422, + 294, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 422, + 294, + 475 + ], + "spans": [ + { + "bbox": [ + 56, + 422, + 294, + 475 + ], + "type": "text", + "content": "[32] Timo Schick, Jane Dwivedi-Yu, Roberto Dessì, Roberta Raileanu, Maria Lomeli, Eric Hambro, Luke Zettlemoyer, Nicola Cancedda, and Thomas Scialom. Toolformer: Language models can teach themselves to use tools. Advances in Neural Information Processing Systems, 36, 2024. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 478, + 294, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 478, + 294, + 521 + ], + "spans": [ + { + "bbox": [ + 56, + 478, + 294, + 521 + ], + "type": "text", + "content": "[33] Paul E Smaldino, Cody Moser, Alejandro Perez Velilla, and Mikkel Werling. Maintaining transient diversity is a general principle for improving collective problem solving. Perspectives on Psychological Science, 19(2):454-464, 2024. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 523, + 294, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 523, + 294, + 578 + ], + "spans": [ + { + "bbox": [ + 56, + 523, + 294, + 578 + ], + "type": "text", + "content": "[34] Chan Hee Song, Jiaman Wu, Clayton Washington, Brian M Sadler, Wei-Lun Chao, and Yu Su. Llm-planner: Few-shot grounded planning for embodied agents with large language models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2998-3009, 2023. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 579, + 294, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 579, + 294, + 610 + ], + "spans": [ + { + "bbox": [ + 56, + 579, + 294, + 610 + ], + "type": "text", + "content": "[35] Wiebe Van der Hoek and Michael Wooldridge. Multi-agent systems. Foundations of Artificial Intelligence, 3:887-928, 2008. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 613, + 294, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 613, + 294, + 678 + ], + "spans": [ + { + "bbox": [ + 56, + 613, + 294, + 678 + ], + "type": "text", + "content": "[36] Tai Wang, Xiaohan Mao, Chenming Zhu, Runsen Xu, Ruiyuan Lyu, Peisen Li, Xiao Chen, Wenwei Zhang, Kai Chen, Tianfan Xue, et al. Embodiedscan: A holistic multimodal 3d perception suite towards embodied ai. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19757-19767, 2024. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 681, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 681, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 681, + 294, + 712 + ], + "type": "text", + "content": "[37] Jason Wei, Yi Tay, Rishi Bommasani, Colin Raffel, Barret Zoph, Sebastian Borgeaud, Dani Yogatama, Maarten Bosma, Denny Zhou, Donald Metzler, et al. Emergent abilities of" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 519 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 94 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 94 + ], + "type": "text", + "content": "large language models. arXiv preprint arXiv:2206.07682, 2022. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 96, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 150 + ], + "type": "text", + "content": "[38] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824-24837, 2022. 1, 2, 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 152, + 553, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 553, + 206 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 553, + 206 + ], + "type": "text", + "content": "[39] Muning Wen, Jakub Kuba, Runji Lin, Weinan Zhang, Ying Wen, Jun Wang, and Yaodong Yang. Multi-agent reinforcement learning is a sequence modeling problem. Advances in Neural Information Processing Systems, 35:16509-16521, 2022. 1, 6, 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 208, + 553, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 208, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 316, + 208, + 553, + 262 + ], + "type": "text", + "content": "[40] Shunyu Yao, Dian Yu, Jeffrey Zhao, Izhak Shafran, Tom Griffiths, Yuan Cao, and Karthik Narasimhan. Tree of thoughts: Deliberate problem solving with large language models. Advances in Neural Information Processing Systems, 36, 2024. 1, 2, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 264, + 553, + 305 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 264, + 553, + 305 + ], + "spans": [ + { + "bbox": [ + 316, + 264, + 553, + 305 + ], + "type": "text", + "content": "[41] Bangguo Yu, Hamidreza Kasaei, and Ming Cao. Co-navgpt: Multi-robot cooperative visual semantic navigation using large language models. arXiv preprint arXiv:2310.07937, 2023. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 308, + 553, + 362 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 308, + 553, + 362 + ], + "spans": [ + { + "bbox": [ + 316, + 308, + 553, + 362 + ], + "type": "text", + "content": "[42] Di Zhang, Xiaoshui Huang, Dongzhan Zhou, Yuqiang Li, and Wanli Ouyang. Accessing gpt-4 level mathematical olympiad solutions via monte carlo tree self-refine with llama-3 8b: A technical report. arXiv preprint arXiv:2406.07394, 2024. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 365, + 553, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 365, + 553, + 418 + ], + "spans": [ + { + "bbox": [ + 316, + 365, + 553, + 418 + ], + "type": "text", + "content": "[43] Hongxin Zhang, Weihua Du, Jiaming Shan, Qinhong Zhou, Yilun Du, Joshua B Tenenbaum, Tianmin Shu, and Chuang Gan. Building cooperative embodied agents modularly with large language models. arXiv preprint arXiv:2307.02485, 2023. 1, 2, 3, 5, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 420, + 553, + 473 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 420, + 553, + 473 + ], + "spans": [ + { + "bbox": [ + 316, + 420, + 553, + 473 + ], + "type": "text", + "content": "[44] Zhonghan Zhao, Ke Ma, Wenhao Chai, Xuan Wang, Kewei Chen, Dongxu Guo, Yanting Zhang, Hongwei Wang, and Gaoang Wang. Do we really need a complex agent system? distill embodied agent into a single model. arXiv preprint arXiv:2404.04619, 2024. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 475, + 553, + 519 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 475, + 553, + 519 + ], + "spans": [ + { + "bbox": [ + 316, + 475, + 553, + 519 + ], + "type": "text", + "content": "[45] Andy Zhou, Kai Yan, Michal Shlapentokh-Rothman, Haohan Wang, and Yu-Xiong Wang. Language agent tree search unifies reasoning acting and planning in language models. arXiv preprint arXiv:2310.04406, 2023. 2" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "29522" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Color Alignment in Diffusion/4b8bcc3c-a5c5-4a38-89da-b85960cd1b00_content_list.json b/2025/Color Alignment in Diffusion/4b8bcc3c-a5c5-4a38-89da-b85960cd1b00_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..25b96f48809a29a8687b37a03686ba814511b062 --- /dev/null +++ b/2025/Color Alignment in Diffusion/4b8bcc3c-a5c5-4a38-89da-b85960cd1b00_content_list.json @@ -0,0 +1,1988 @@ +[ + { + "type": "text", + "text": "Color Alignment in Diffusion", + "text_level": 1, + "bbox": [ + 349, + 130, + 647, + 151 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ka Chun Shum1 Binh-Son Hua2 Duc Thanh Nguyen3 Sai-Kit Yeung1", + "bbox": [ + 199, + 191, + 795, + 212 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Hong Kong University of Science and Technology $^{2}$ Trinity College Dublin $^{3}$ Deakin University", + "bbox": [ + 117, + 222, + 877, + 242 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 286, + 326, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Diffusion models have shown great promise in synthesizing visually appealing images. However, it remains challenging to condition the synthesis at a fine-grained level, for instance, synthesizing image pixels following some generic color pattern. Existing image synthesis methods often produce contents that fall outside the desired pixel conditions. To address this, we introduce a novel color alignment algorithm that confines the generative process in diffusion models within a given color pattern. Specifically, we project diffusion terms, either imagery samples or latent representations, into a conditional color space to align with the input color distribution. This strategy simplifies the prediction in diffusion models within a color manifold while still allowing plausible structures in generated contents, thus enabling the generation of diverse contents that comply with the target color pattern. Experimental results demonstrate our state-of-the-art performance in conditioning and controlling of color pixels, while maintaining on-par generation quality and diversity in comparison with regular diffusion models.", + "bbox": [ + 89, + 318, + 485, + 604 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 633, + 220, + 648 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, diffusion models [18] and their large-scale text-conditioned variants [36, 39] have demonstrated impressive quality and diversity of generated contents in image synthesis. Subsequent works have explored leveraging this capability for possible controllability, incorporating various constraints, such as structural [30, 35, 48] and reference constraints [14, 46], which typically serve as additional inputs. However, due to the inherent randomness in the diffusion process, it remains challenging to control generated contents at a fine-grained level, for instance, synthesizing image pixels following some given content and generic color distribution.", + "bbox": [ + 89, + 657, + 483, + 824 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we focus on such a fine-grained control over pixel colors in synthesized images to meet with three color-conditioned generation criteria: the accuracy in adhering to given color values; the completeness of covering proportions of the specified colors; and the disentanglement of the colors", + "bbox": [ + 89, + 825, + 483, + 900 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/85cf9bc14882995fa8308b682e746006db8a7e587eb076f93a95b3d7515154ad.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 514, + 285, + 612, + 359 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c24737b727e83432a2209be3aa2222992e90877ade8b8acab3a286094958b777.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 516, + 359, + 609, + 431 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1f3294cc6f0c5c27e2fa59fdf44c611d5dee03ed0715abfe767a1615f6c17edb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 614, + 285, + 709, + 358 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7cf7ee32d69fbe1e881cc15aa9090585d5754542f3b20627dc66d065db4a9752.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 614, + 359, + 707, + 431 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/78701179e89d71a22f1ec49707e734a0038432957f6775df1d54b8c6de66caff.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 710, + 285, + 803, + 358 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/06dc785169846cfd8d5c373b8698d4ffc2457e433af540f531d4b57d0f829a1f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 710, + 359, + 803, + 431 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/94802e72e417ea78c39f53cace29debadff122c963db708f69e04b111fc4ffc7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 808, + 285, + 903, + 358 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3473587401498fe280f8d76908cf995f24de07b5117a072217a90f739d9a5486.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 808, + 359, + 903, + 431 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a85619011eaae6615460963d82b003da2de87ed01e3e0ad0b77509fd40d19ff6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 516, + 439, + 609, + 512 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6d5192fb92440f9f7a9f5f9e8b48e6c3d95d8871ed062fb5ff5a4600ee436db1.jpg", + "image_caption": [ + "(a) Condition" + ], + "image_footnote": [], + "bbox": [ + 516, + 512, + 609, + 585 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/48e2934d33a138d09618621ae0b10f915b65955eb2195a2dd7e41c99c0523e29.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 616, + 439, + 707, + 511 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d852296cb83b6b41f5a1a0d3d22cf26c74df25ac9a4dbaf3c20611be1273d50c.jpg", + "image_caption": [ + "(b) Ours" + ], + "image_footnote": [], + "bbox": [ + 616, + 512, + 707, + 585 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/cdacc8f52e07ac25f4fa920dcb0b4f4703c4445f3e58bb534556716f5b49ac29.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 710, + 439, + 803, + 511 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7a529dcb871fe764f313a2b0a9dcf845c3eebd2ba3d616624e6d2916ae85312f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 710, + 512, + 803, + 585 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6ca033b21327cd259be2f7120b59788f61b4b52142988f95d5a5a38a4b1a770b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 810, + 439, + 903, + 511 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d80b7b18f4d9ccadc468d0e6cf3eeb379f0f195ffc81f02a1728d5ea58c9aec4.jpg", + "image_caption": [ + "(c)Baselines", + "Figure 1. Examples of color-conditioned generation from different methods. Color conditions (a) can be derived from imagery (top two rows) or manual drawing (bottom two rows). Input text prompts are \"painting\" and \"pen\" (1st & 3rd rows), and \"boat\" (2nd & 4th rows). Our method (b) can generate pixels closely aligned with the color conditions and effectively disentangle the colors into different structures. Existing baselines [14, 36, 46, 48] (from top to bottom) (c) struggle with either the semantics (the first two rows) or colors (the last two rows) of generated contents." + ], + "image_footnote": [], + "bbox": [ + 810, + 512, + 903, + 585 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "into various structures. Such a color-conditioned generation ability would be a great aid to numerous downstream tasks in creative artwork and graphic design [10].", + "bbox": [ + 511, + 763, + 903, + 809 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Different attempts of color-conditioned image generation are visualized in Fig. 1, where previous works often produce suboptimal results that do not align with the aforementioned criteria. In this paper, we propose an effective yet control-easy method facilitating color alignment in diffusion models for color-conditioned image synthesis. Specifically,", + "bbox": [ + 511, + 810, + 908, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "28446", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "our method accepts a color condition via a color pattern, determining color values and their proportions in a synthesized image. Different from the regular diffusion approach, we perform the color alignment on intermediate samples or latent representations during the diffusion process prior to inputting them into a denoising model for image synthesis, thereby facilitating the perception of the conditional colors in diffusion models while preserving the overall attribution of the diffusion process. Furthermore, our method can support various settings: re-training of a color-aligned diffusion model; fine-tuning of a color-aligned diffusion model from a pre-trained regular diffusion model; and zero-shot approximation (training-free) of the color alignment process with a pre-trained diffusion model. We validate our method by demonstrating its capabilities across various color conditions and compare it with existing baselines on benchmark datasets. In summary, we make the following contributions:", + "bbox": [ + 89, + 90, + 485, + 348 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We address a challenging task: fine-grained color-conditioned image synthesis. We aim to generate images that closely follow a specified color pattern that defines color values and proportions without requiring an explicit structure. Our approach thus enhances the controllability and flexibility of image generation.", + "- We propose a novel color alignment method that operates on the diffusion process of diffusion models, allowing the diffusion to effectively perceive color conditions while maintaining the quality and diversity of generated contents. We derive various settings including re-training, fine-tuning, and a zero-shot approximation, to adapt our color alignment method into different down-stream needs.", + "- We conduct extensive experiments to validate our proposed method in various scenarios and controllability settings. Our results demonstrate state-of-the-art performance compared with existing baselines." + ], + "bbox": [ + 89, + 348, + 483, + 604 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 89, + 616, + 228, + 631 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Neural image synthesis has started with Variational Autoencoders (VAEs) [22, 41, 43] and Generative Adversarial Networks (GANs) [12, 19, 20, 26], followed then by transformer-based methods [9, 23] and diffusion models [18, 31, 42]. Among these, Denoising Diffusion Probabilistic Models (DDPM) [18] have attracted considerable attention due to their generation quality and ease of extension. These models have been further developed into text-to-image (T2I) diffusion techniques [35, 36, 39], which have been boosted recently to achieve greater controllability through incorporation of extra conditions [5, 37, 38, 48]. Our method follows the general diffusion framework with a focus on fine-grained color-conditioned image synthesis.", + "bbox": [ + 89, + 641, + 485, + 838 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Reference-based conditional diffusion. An intuitive condition for diffusion models is the use of an image reference. Structural conditions, such as edge maps and semantic maps,", + "bbox": [ + 89, + 854, + 485, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "have been recently studied in [24, 25, 30, 38, 45, 48]. These methods strictly apply spatial constraints given in the reference to align generated pixels. A more relaxed approach is image editing [5, 6, 13, 21, 27], where the structure and appearance in the reference loosely guide the generation of target images. Those works aim to achieve shifts in appearance while permitting a small range of spatial edits, such as changes in the size and pose of objects. Several methods [14, 37, 46] condition only on abstract concepts in the reference, such as image/object style, enabling more creative synthesis. In this work, we disentangle colors in the reference from their original spatial structure while preserving the color values and proportions. Our method allows a variety of user-provided color materials to construct diverse structures.", + "bbox": [ + 511, + 90, + 908, + 303 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Sample altering in diffusion. As indicated in [25], diffusion from the same noisy sample can result in similar outputs. Diffusion inversion [27] is commonly used to construct samples, used directly [27, 33] or mapped to another diffusion process [7, 29, 49] to diverge outputs. Several methods update intermediate latent representations with gradients calculated from external modules, e.g., other networks [47], images [40], features [28]. The Gaussian noise is defined [11] using estimated mean and variance to improve model convergence. Cold Diffusion [4] replaces the noising process with image degradations such as blurring and masking.", + "bbox": [ + 511, + 329, + 910, + 497 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Proposed method", + "text_level": 1, + "bbox": [ + 511, + 513, + 684, + 532 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We first in Sec. 3.1 review the regular diffusion process. In Sec. 3.2, we introduce our color alignment technique under its simplest setting, i.e., re-training of a diffusion model with color alignment. We then in Sec. 3.3 extend our technique to latent diffusion models, where we address typical challenges in latent space diffusion and demonstrate our method's ability via fine-tuning of a pre-trained latent diffusion model to a color-aligned one. Finally, in Sec. 3.4, we present a zero-shot variant of our method that achieves similar generation quality without extra training. We provide an overview of our pipeline and the regular diffusion in Fig. 2.", + "bbox": [ + 511, + 541, + 908, + 708 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3.1. Diffusion preliminaries", + "text_level": 1, + "bbox": [ + 511, + 722, + 728, + 738 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Diffusion models [18] could be seen as a variant of VAEs [22] that map a data distribution $\\mathcal{X}(\\mathbf{x}_0)$ from a Gaussian distribution $\\mathcal{N}(\\mathbf{0},\\mathbf{I})$ . In the forward process, a sample $\\mathbf{x}_t$ is encoded by adding a Gaussian noise to $\\mathbf{x}_0$ :", + "bbox": [ + 511, + 746, + 906, + 806 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {x} _ {t} = \\sqrt {\\bar {\\alpha} _ {t}} \\mathbf {x} _ {0} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\boldsymbol {\\epsilon}, \\quad \\boldsymbol {\\epsilon} \\sim \\mathcal {N} (\\mathbf {0}, \\mathbf {I}) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 821, + 906, + 838 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "where $t \\in [1, T]$ is the time step associated with a noise strength $\\bar{\\alpha}_t$ where $(1 - \\bar{\\alpha}_1) \\approx 0$ and $(1 - \\bar{\\alpha}_T) \\approx 1$ .", + "bbox": [ + 511, + 853, + 903, + 883 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Then, a denoising model $\\epsilon_{\\theta}$ with parameters $\\theta$ learns to", + "bbox": [ + 532, + 885, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "28447", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "reverse the process in Eq. 1 by training with a loss $\\mathcal{L}$ as,", + "bbox": [ + 89, + 90, + 455, + 107 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\nabla_ {\\theta} \\| \\epsilon - \\epsilon_ {\\theta} (\\mathbf {x} _ {t}, t) \\| _ {2} ^ {2} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 133, + 119, + 483, + 138 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {x}} _ {0} = \\frac {1}{\\sqrt {\\bar {\\alpha} _ {t}}} \\left(\\mathbf {x} _ {t} - \\sqrt {1 - \\bar {\\alpha} _ {t}} \\boldsymbol {\\epsilon} _ {\\theta} \\left(\\mathbf {x} _ {t}, t\\right)\\right) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 129, + 140, + 483, + 171 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {x} _ {t - 1} = \\frac {\\sqrt {\\bar {\\alpha} _ {t - 1}} \\beta_ {t}}{1 - \\bar {\\alpha} _ {t}} \\hat {\\mathbf {x}} _ {0} + \\frac {\\sqrt {\\alpha_ {t}} (1 - \\bar {\\alpha} _ {t - 1})}{1 - \\bar {\\alpha} _ {t}} \\mathbf {x} _ {t} + \\sigma_ {t} \\boldsymbol {\\varepsilon} \\quad (4)\n$$\n", + "text_format": "latex", + "bbox": [ + 112, + 172, + 483, + 205 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\hat{\\mathbf{x}}_0$ is the predicted sample. $\\alpha ,\\beta ,\\sigma \\in \\mathbb{R}$ and $\\varepsilon \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})$ are scheduler-related parameters [18].", + "bbox": [ + 89, + 218, + 483, + 250 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Color-aligned diffusion in image space", + "text_level": 1, + "bbox": [ + 89, + 261, + 421, + 277 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As presented in Sec. 3.1, the regular diffusion model learns to reconstruct $\\mathbf{x}_0$ from $\\mathcal{N}(\\mathbf{0},\\mathbf{I})$ during training and allows sampling-free generation in testing, i.e., unconditional generation. To condition the diffusion process with color conditions, existing methods pass color text prompts [36, 46] or referenced images [14, 46, 48] in parallel with $\\mathbf{x}_t$ untouched. We found this approach loosely constrains the sampled data manifold, i.e., $\\hat{\\mathbf{x}}_0$ can be sampled with out-of-bound colors. This is because the conditions are not involved in the sampling process (in construction of $\\mathbf{x}_t$ ) and the denoising model $\\epsilon_{\\theta}$ hence leans towards the unconditional generation. Moreover, colors are learned in conjunction with spatial information given in referenced images, limiting the diversity of generated contents (see Fig. 1).", + "bbox": [ + 89, + 284, + 483, + 494 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this paper, we take a color pattern $\\mathbf{c}$ , determining color values and their proportions in synthesis, as a color condition. The color pattern can be an imagery or hand-drawing (see Fig. 1). To effectively constrain the sampling manifold with color condition, we involve the color pattern $\\mathbf{c}$ into the diffusion process via a new denoising model $\\epsilon_{\\theta}(\\mathbf{x}_t,t,\\mathbf{c})$ . In addition, we relax the spatial information in $\\mathbf{c}$ by using a color alignment function $f(\\mathbf{x}_t,\\mathbf{c})$ that produces a color-aligned image from an image $\\mathbf{x}_t$ and color pattern $\\mathbf{c}$ as,", + "bbox": [ + 89, + 497, + 485, + 633 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nf \\left(\\mathbf {x} _ {t}, \\mathbf {c}\\right) [ p ] = \\underset {\\mathbf {c} [ q ]} {\\arg \\min } \\| \\mathbf {x} _ {t} [ p ] - \\mathbf {c} [ q ] \\| _ {2} ^ {2} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 158, + 646, + 483, + 674 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbf{x}_t[p]$ and $\\mathbf{c}[q]$ are colors at pixels $p$ and $q$ .", + "bbox": [ + 89, + 686, + 411, + 704 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We visualize the diffusion process in Fig. 3. Our color alignment technique benefits the sampling in several aspects. First, it enforces the generated pixel colors within the target space, guaranteeing the accuracy. Second, its non-spatial manner allows pixels distribute more freely, fulfilling the color disentanglement. For the sake of completeness, we further concatenate $\\mathbf{c}$ to the model's input as an extra hint. However, to completely break down the spatial information, we opt to randomly permute the pixel locations in $\\mathbf{c}$ , resulting in $\\psi(\\mathbf{c})$ with $\\psi$ being an image permutation operator. Such a permutation allows us to directly take $\\mathbf{x}_0$ as color condition in training time, avoiding extra data requirement and data leakage. Finally, training and sampling of our color", + "bbox": [ + 89, + 704, + 485, + 901 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/f740dcf1b938da0a080229c45ea47377d6e35fcf57f37b87a3f145339caa1c82.jpg", + "image_caption": [ + "(a) Regular diffusion [36]" + ], + "image_footnote": [], + "bbox": [ + 516, + 89, + 638, + 239 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/6cda9bde6cc68309455310737b2727d8902357514ecd19845311bbf3b14c9c11.jpg", + "image_caption": [ + "(b) Regular diffusion with condition" + ], + "image_footnote": [], + "bbox": [ + 709, + 89, + 903, + 239 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/db639651312e35320c3d57121c13e2a6fe3fe026c512b2f1a6bd5647e94ffae7.jpg", + "image_caption": [ + "(c) Our color-aligned diffusion for retraining or fine-tuning", + "Figure 2. Pipelines of the regular diffusion (a) and its color-conditioned version (b), our color-aligned diffusion (c), where noisy samples $\\mathbf{x}_t$ are aligned with conditioned colors, and our zero-shot version (d), where color alignment is directly applied to predicted sample $\\hat{\\mathbf{x}}_0$ ." + ], + "image_footnote": [], + "bbox": [ + 516, + 268, + 707, + 468 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/a270035bdc2014badd0926ff6f48f34a6a54ce077f59c3352ac33d258c09b61e.jpg", + "image_caption": [ + "(d) Our zero-shot approximation of the color-aligned diffusion" + ], + "image_footnote": [], + "bbox": [ + 712, + 268, + 903, + 468 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "aligned diffusion process (Fig. 2c) are formulated as:", + "bbox": [ + 511, + 608, + 864, + 625 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\nabla_ {\\boldsymbol {\\theta}} \\| \\boldsymbol {\\epsilon} ^ {\\prime} - \\boldsymbol {\\epsilon} _ {\\boldsymbol {\\theta}} \\big (f (\\mathbf {x} _ {t}, \\mathbf {c}), t, \\mathbf {c} \\big) \\| _ {2} ^ {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 539, + 633, + 774, + 652 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\epsilon} ^ {\\prime} = \\frac {f (\\mathbf {x} _ {t} , \\mathbf {c}) - \\sqrt {\\bar {\\alpha} _ {t}} \\mathbf {x} _ {0}}{\\sqrt {1 - \\bar {\\alpha} _ {t}}}, \\quad \\mathbf {c} = \\psi (\\mathbf {x} _ {0}) \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 539, + 654, + 906, + 688 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {x}} _ {0} = \\frac {1}{\\sqrt {\\bar {\\alpha} _ {t}}} \\left(f \\left(\\mathbf {x} _ {t}, \\mathbf {c}\\right) - \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon_ {\\theta} \\left(f \\left(\\mathbf {x} _ {t}, \\mathbf {c}\\right), t, \\mathbf {c}\\right)\\right) \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 534, + 689, + 906, + 722 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\epsilon^{\\prime}$ is the adapted noise complemented with the effect of $f$ . Notably, we alter only what $\\epsilon_{\\theta}$ sees and predicts. The parameters $\\theta$ thus can be learned more easily with color condition compared with the regular one. All other nonlearnable steps, e.g., the regular forward process and Eq. 4, remain unchanged for the ease of sampling. For more technical visualizations, please refer to the supplementary material.", + "bbox": [ + 511, + 732, + 908, + 839 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.3. Color-aligned diffusion in latent space", + "text_level": 1, + "bbox": [ + 511, + 848, + 841, + 864 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Training of diffusion models in their latent space is a common practice. In this section, we extend our color alignment", + "bbox": [ + 511, + 869, + 906, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "28448", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "technique to latent diffusion models via fine-tuning of a pre-trained latent model to achieve color alignment.", + "bbox": [ + 89, + 90, + 482, + 119 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Following the Stable Diffusion in [36], we encode $\\mathbf{x}_0$ into its latent representation $\\mathbf{z}_0$ . We then train our model with alignment in the latent space, i.e., replacing $\\mathbf{x}$ by $\\mathbf{z}$ in Eqs. 6 and 7. However, when high-frequency structural information is present in $\\mathbf{x}_0$ , unintended colors and local structural information can be encoded in $\\mathbf{z}_0$ . We discovered that blurring $\\mathbf{x}_0$ prior to encoding it enhances the color disentanglement and accuracy, with minimal loss of the completeness in generated results. Also, we found that suspending color alignment at late time steps (e.g., $t < 0.2T$ where diffusion model is fixing details) allows a proper refinement of the latent, making latent decoding towards more photo-realistic details.", + "bbox": [ + 89, + 121, + 483, + 303 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Zero-shot color-aligned diffusion", + "text_level": 1, + "bbox": [ + 89, + 313, + 379, + 329 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To mitigate the computational burden of training, we introduce a zero-shot approximation of our color alignment method. As shown in Eqs. 6 and 7, the training aims to reconstruct $\\hat{\\mathbf{x}}_0$ whose color distribution follows the color condition $\\mathbf{c} = \\psi (\\mathbf{x}_0)$ . Under the zero-shot setting, to achieve this ability without learning the parameters $\\theta$ , we directly map an unconditionally generated $\\hat{\\mathbf{x}}_0$ (from Eq. 3) to $\\mathbf{c}$ during the sampling process using a color alignment function $g(\\hat{\\mathbf{x}}_0,\\mathbf{c})$ :", + "bbox": [ + 89, + 335, + 483, + 457 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ng \\left(\\hat {\\mathbf {x}} _ {0}, \\mathbf {c}\\right) = \\underset {\\psi (\\mathbf {c})} {\\arg \\min } \\| \\hat {\\mathbf {x}} _ {0} - \\psi (\\mathbf {c}) \\| _ {2} ^ {2} \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 169, + 483, + 483, + 510 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where we map the unconditional $\\hat{\\mathbf{x}}_0$ to a closest reorganization of the conditional colors $\\psi (\\mathbf{c})$ for color alignment.", + "bbox": [ + 89, + 522, + 483, + 551 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Unlike $f$ defined in Eq. 5, $g$ in Eq. 8 performs one-to-one mapping, i.e., each location in $\\mathbf{c}$ is used for color alignment exactly once. Note that, like $f$ , our $g$ also applies non-spatial mapping. This would ensure desired properties (color accuracy, completeness, and disentanglement) from the zero-shot color-aligned diffusion. However, since $g$ is not involved in the learning process, the color-aligned image $g(\\hat{\\mathbf{x}}_0,\\mathbf{c})$ could be messy, showing invalid structures and outliers (see Fig. 2d). Fortunately, we observed that pre-trained diffusion models have capability of clustering semantics and eliminating outliers at early sampling steps (e.g., $t > 0.2T$ ). We also suspend the color alignment at late time steps to prevent excessive color changes and allow the model to refine details. Additionally, such zero-shot approximation is extendable to latent diffusion models. We found the same conclusion for using the blurring strategy as in Sec. 3.3 applies.", + "bbox": [ + 89, + 553, + 483, + 809 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Another interest could be replacing $f$ by $g$ in learning. However, we find this not applicable because $g(\\mathbf{x}_t,\\mathbf{c})$ corrupts the distribution of $\\mathbf{x}_t$ too heavily. Its subsequent adapted noise $\\epsilon^{\\prime}$ (the learning target in Eq. 6) deviates too far from the standard Gaussian $\\mathcal{N}(\\mathbf{0},\\mathbf{I})$ , eliminating the effect of a proper diffusion process.", + "bbox": [ + 89, + 810, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 513, + 90, + 645, + 107 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Datasets", + "text_level": 1, + "bbox": [ + 513, + 114, + 614, + 128 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We experimented our method in two scenarios: color-aligned diffusion in image space and color-aligned diffusion in latent space. For diffusion in image space, we ran experiments on Oxford-flower [32] (7.17k training and 1.02k test images) and Microsoft-emoji [8] datasets (6.80k training and 0.76k test images) at resolution $64 \\times 64$ . For latent diffusion, we experimented on a sub-set of a high-quality generic dataset Text-to-image-2M [1] (300k training and 20k test images) at resolution $512 \\times 512$ . To validate the color disentanglement ability of our method, we replaced original prompts in Text-to-image-2M [1] with 50 daily object prompts randomly generated by ChatGPT [2] (e.g., those in Fig. 4).", + "bbox": [ + 511, + 137, + 906, + 319 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Baselines", + "text_level": 1, + "bbox": [ + 513, + 329, + 620, + 343 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our color-aligned image diffusion version is adapted from the DDPM [18]. Therefore, we compare our method with the DDPM to demonstrate its extra color-conditioned ability without requiring additional data creation or labeling. For the same reason, we also compare our latent diffusion version with the pre-trained Stable Diffusion in [36].", + "bbox": [ + 511, + 351, + 905, + 441 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given that our color reference can be an in-the-wild image or a manual drawing, we compare our method with SOTA reference-based generation methods. IP-Adapter [46] and T2I-Adapter [30] convert an image reference into features, which are subsequently combined with the model features for image generation. Style-Aligned [14] transfers source image information from its attention layer to the target generation attention layer, thereby preserving spatial and style features. ControlNet-Color is a variant of ControlNet [48] that accepts a predefined pixelized color image as input and generates a refined version in which the colors are approximately aligned with the original image spatially. IP-Adapter [46] and ControlNet-Color [48] require fine-tuning, while Style-Aligned [14] operates as a zero-shot approach.", + "bbox": [ + 511, + 441, + 908, + 654 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.3. Implementation details", + "text_level": 1, + "bbox": [ + 511, + 664, + 725, + 679 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We follow default settings in the DDPM [18] and the Stable Diffusion [36] for network architectures and diffusion scheduling, specifically, from the huggingface-diffusers [44]. Further details, e.g., DDPM re-training, Stable Diffusion fine-tuning are presented in our supplementary material.", + "bbox": [ + 511, + 686, + 906, + 762 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We trained our image diffusion version for 100 epochs and latent diffusion version for $160\\mathrm{k}$ iterations. We performed inference at 50 time steps. We used classifier-free guidance [17] with guidance scale 5. All experiments were executed on two RTX 3090 GPUs with batch size 4.", + "bbox": [ + 511, + 762, + 906, + 835 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.4. Qualitative study", + "text_level": 1, + "bbox": [ + 511, + 848, + 681, + 864 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Color-aligned image diffusion. We visually compare our method with the regular diffusion in DDPM [18] in Fig. 3.", + "bbox": [ + 511, + 869, + 906, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "28449", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/f7608774d3e9bb2970cd84dd5ea576b611c4e671295186736580e59440c6ecba.jpg", + "image_caption": [ + "(a) Our color-aligned image diffusion process. The top row shows our training inputs, created from early to late steps in the forward process. The bottom row is a test case from our reverse process, sampled from pure noise." + ], + "image_footnote": [], + "bbox": [ + 93, + 90, + 475, + 191 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/17c9cf1f6efbf645cb14f5c2a30f4a4c8575dfca3ed7a9ed2ed01ccce09ac9c2.jpg", + "image_caption": [ + "(b) Regular image diffusion process in DDPM [18] in same order as (a)." + ], + "image_footnote": [], + "bbox": [ + 93, + 233, + 475, + 334 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/8f64d93754bd2496078cd957abd57ba6779b959093f5fc401ddec9697730a72e.jpg", + "image_caption": [ + "(c) Several color-aligned image synthesis results. From top to bottom rows are condition, our generated results, and regular diffusion's results.", + "Figure 3. Qualitative results of color-aligned image diffusion. (a)-(b) Visualization of the diffusion process by our method and DDPM [18]. (c) Generation results of our method and DDPM [18]." + ], + "image_footnote": [], + "bbox": [ + 93, + 351, + 475, + 503 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As shown, our method (Fig. 3a) well restricts the diffusion process to the conditional color space, wherein only whites, yellows, and background blacks are allowed. In contrast, the regular diffusion in DDPM (Fig. 3b) generates out-of-bounds colors during the diffusion process, leading to an unintended reconstruction result.", + "bbox": [ + 89, + 613, + 483, + 702 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Fig. 3c illustrates an immediate application of our color-aligned image diffusion in image rephrasing, i.e., colors in an input image are reorganized by our model to construct new structures. More importantly, this conditional process does not require any additional data (e.g., labeled data or paired data) or pre-defined spatial information (e.g., semantic maps or masks) in both training or inference. This makes our technique scalable to larger datasets.", + "bbox": [ + 89, + 704, + 483, + 824 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Color-aligned latent diffusion. We present our coloraligned latent diffusion results in comparison with existing baselines in Fig. 4. As demonstrated, our method (Figs. 4b and 4c) effectively conditions the synthesis on target color inputs. Color values and proportions are aligned, i.e., color", + "bbox": [ + 89, + 825, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "accuracy and completeness are maintained. Additionally, our method generates contents that align well with target text prompts, showcasing the ability for color disentangle-ment where colors are rearranged into new structures. Furthermore, with these color-conditioned generation criteria preserved, our method achieves image quality and diversity on par with the original diffusion model.", + "bbox": [ + 511, + 90, + 906, + 196 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In contrast, the baselines produce suboptimal results in various aspects. IP-Adapter [46] (Fig. 4d), T2I-Adapter [30] (Fig. 4e), and Style-Aligned [14] (Fig. 4f) construct target contents on top of an input image reference, thus often struggle to disentangle colors. Since the spatial information of colors is taken into account, these methods likely to fail completely if input reference and target text prompt belong to significantly different domains. ControlNet-Color [48] (Fig. 4g) treats input reference as a weak guidance, leading to failures to accurately preserve color values and proportions.", + "bbox": [ + 511, + 198, + 908, + 349 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Generalizability and controllability. Our method also works with manual drawings, for example, a rough color pattern. As shown in Fig. 5, our method (the fine-tuned version) can generate images aligned with the color values and proportions specified in input color patterns. Moreover, even if only a few color values presented, our method can adjust these values to achieve appropriate smoothness, brightness, and shadow in generated contents.", + "bbox": [ + 511, + 349, + 906, + 470 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "These manual color conditions can be easily modified to influence the generation process. For example, the amount of each color value can be scaled (Fig. 5a). The color values can be tuned intuitively without the need for careful prompt engineering (Fig. 5b). Additional colors can be incorporated into conditions along with user ideation (Fig. 5c). By controlling target text prompt, we can specify the placement of colors (Fig. 5d). Note that, like the regular diffusion, our method can generate diverse results from the same input condition by sampling different noises.", + "bbox": [ + 511, + 472, + 908, + 623 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Zero-shot diffusion. We showcase the ability of our method in zero-shot diffusion in Fig. 6. Our experiments indicate that, compared with our fine-tuned version, the zero-shot version tends to generate images with less fine-grained details, resulting in issues such as blurry lighting, uniform texture, and shadow missing. Since our zero-shot approach mechanically maps colors without model training, it performs more comparably to the fine-tuned version when applied to well-sampled color conditions, e.g., in-the-wild images in the fifth row in Fig. 6 and those in Fig. 4.", + "bbox": [ + 511, + 623, + 908, + 776 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.5.Quantitative study", + "text_level": 1, + "bbox": [ + 511, + 787, + 692, + 803 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Metrics. We evaluate the quality of generated images using the widely adopted FID score [16]. To assess the color disentanglement, we use the CLIP-Score [15] measuring the similarity between a generated image and a target text prompt within the CLIP [34] embedding space. Since the color accuracy and completeness can be evaluated independ-", + "bbox": [ + 511, + 809, + 908, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "28450", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/a7d6e88669e9438c366231e09cde4d8b0d81066ff8fa2d175dce681ab1e80c8f.jpg", + "image_caption": [ + "(g) ControlNet-Color [48]", + "Figure 4. Qualitative results of color-aligned latent diffusion. Each input (first row) includes an in-the-wild image as color condition and a target text prompt. Each column presents results of experimented methods using the same input." + ], + "image_footnote": [], + "bbox": [ + 89, + 85, + 906, + 758 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "dently of spatial relations, we measure the proximity between the color distributions of a generated image and input color condition. We employ the Chamfer Distance (CD) [3], a commonly used metric for measuring the distance between", + "bbox": [ + 89, + 839, + 485, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "two point sets, to quantify the distance between generated RGB points $\\hat{\\mathbf{x}}_0$ and input condition RGB points $\\mathbf{c}$ :", + "bbox": [ + 511, + 839, + 906, + 869 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "28451", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/adaad4824b3da68008f02200840b931fd9d7feccb871eb2e3ac0c73114301c19.jpg", + "image_caption": [ + "Figure 5. Qualitative results of sampling and editing of input color conditions. All results are generated by our fine-tuned color-aligned latent model. Red arrows represent the generation process. Blue arrows represent the editing of color conditions." + ], + "image_footnote": [], + "bbox": [ + 94, + 88, + 908, + 556 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/8b756067626fe858ea710cc80de10258b46899b709d454cbb52cf78c4af58060.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodFID ↓CD-A ↓CD-C ↓Re-training Cost ↓Inference Cost ↓
Image Diffusion (Re-train) [18]63.9; 14940.6; 42.442.6; 39.01.001.00
Ours (Image; Re-train)57.5; 45.70.00; 0.004.12; 3.651.061.13
Ours (Image; Zero-shot)50.0; 84.55.59; 5.468.28; 10.40.006.76
", + "bbox": [ + 151, + 608, + 844, + 686 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 1. Quantitative results of image diffusion. In each metric column, the left and right numbers correspond to the scores from the Oxford-flower dataset [32] and Microsoft-emoji [8] dataset, respectively.", + "bbox": [ + 89, + 694, + 906, + 722 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {C D - A} \\left(\\hat {\\mathbf {x}} _ {0}, \\mathbf {c}\\right) = \\frac {\\sum_ {\\hat {\\mathbf {x}} _ {0} [ p ]} \\min _ {\\mathbf {c} [ q ]} \\| \\hat {\\mathbf {x}} _ {0} [ p ] - \\mathbf {c} [ q ] \\| _ {2} ^ {2}}{| \\hat {\\mathbf {x}} _ {0} |} \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 112, + 771, + 483, + 806 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {C D - C} \\left(\\hat {\\mathbf {x}} _ {0}, \\mathbf {c}\\right) = \\frac {\\sum_ {\\mathbf {c} [ q ]} \\min _ {\\hat {\\mathbf {x}} _ {0} [ p ]} \\| \\hat {\\mathbf {x}} _ {0} [ p ] - \\mathbf {c} [ q ] \\| _ {2} ^ {2}}{| \\mathbf {c} |} \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 112, + 809, + 483, + 845 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\mathrm{CD} = \\mathrm{CD - A} + \\mathrm{CD - C}$ consists of an accuracy term and a completeness term that measure the set distance from one to another; $|\\cdot |$ denotes set cardinality. Lastly, we evaluate", + "bbox": [ + 89, + 854, + 483, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "the inference cost of different methods, using the regular diffusion as the standard reference.", + "bbox": [ + 511, + 750, + 906, + 777 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Baseline comparison. We first present the performance scores for the image diffusion case study in Tab. 1. The scores clearly indicate that our technique effectively preserves conditioned-colors. Note that, for inference, the color alignment is performed on RGB pixels until the final diffusion step, making the CD-A approach zero. We found that our fine-tuned version requires nearly the same training and inference costs as the regular diffusion model.", + "bbox": [ + 511, + 779, + 908, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "28452", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/568e4668110ff5d15f34aa3c6247214ae0405a0f315fefc16b739c926b033b7a.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodFID ↓CLIPScore ↑CD-A ↓CD-C ↓Inference Cost ↓
Stable Diffusion (Pre-train) [36]72.6; 72.627.3; 27.3166; 27.865.3; 16.01.00
Ours (Latent; Fine-tune)86.7; 69.429.0; 27.473.9; 4.9815.8; 4.871.03
Ours (Latent; Zero-shot)104; 77.928.3; 27.435.2; 2.683.19; 3.933.52
Baseline (IP-Adapter [46])202; 50.925.5; 22.0145; 10.1106; 13.91.44
Baseline (T2I-Adapter [30])134; 77.124.3; 24.9143; 8.3424.7; 7.011.56
Baseline (Style-Aligned [14])177; 73.123.5; 22.945.1; 6.1832.5; 8.505.69
Baseline (ControlNet-Color [48])105; 81.022.9; 23.4129; 13.149.8; 6.541.78
", + "bbox": [ + 163, + 88, + 834, + 227 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a6166db6e7cd9f137e37587475e03cadabb97ff0a257cafbdb37166b024cbe0d.jpg", + "image_caption": [ + "\"A professional photo of a bar.\"" + ], + "image_footnote": [], + "bbox": [ + 93, + 295, + 187, + 369 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5f01b718d4c505a1d9183a0b728ef45d6bdbf7bdd393d14148fd6ec9378f7673.jpg", + "image_caption": [ + "of a bar:" + ], + "image_footnote": [], + "bbox": [ + 192, + 295, + 287, + 369 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6dc44c9b58ecc010a2291884d57a229b439b69ac24a25f3fe8be93bd8dabb06d.jpg", + "image_caption": [ + "" + ], + "image_footnote": [], + "bbox": [ + 388, + 295, + 482, + 369 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/39f889f61f5d4b8f7638cc69f76b58acc3258807476ddd9b76ee01f6715697cf.jpg", + "image_caption": [ + "\"Gray modern house, photograph, highly detailed." + ], + "image_footnote": [], + "bbox": [ + 93, + 382, + 187, + 455 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f36b925e4ea021276383c4da6f0c1395f7d57a694993589b763ec5939c6fa83f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 192, + 382, + 287, + 455 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "photograph, highly detailed.", + "bbox": [ + 192, + 455, + 316, + 465 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7a47f65f1c0dc17a9c09d117ddaa62b337fa887e3a97d5ab43cdb3c86c3ed3d2.jpg", + "image_caption": [ + "tailed." + ], + "image_footnote": [], + "bbox": [ + 289, + 382, + 382, + 455 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/59069fe124565ac81a834c0ac606fba67550fa011f685fd79892cef13ea44269.jpg", + "image_caption": [ + "" + ], + "image_footnote": [], + "bbox": [ + 387, + 382, + 480, + 455 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/76562308239e8646d2b5d78dcc8e42dda1bb757afbe4de247708d78c75d466f1.jpg", + "image_caption": [ + "\"An animated owl looking at the camera." + ], + "image_footnote": [], + "bbox": [ + 94, + 468, + 187, + 541 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d390afd7ad3750eddf15560afd5488e38a1a6d29f80b26844ccc03ba7895f4d2.jpg", + "image_caption": [ + "king at the camera." + ], + "image_footnote": [], + "bbox": [ + 192, + 468, + 287, + 541 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/274de5eaa47b6b333bad1c8e8f0a6e870568274c42a499a86e96813351a26fe2.jpg", + "image_caption": [ + "$\\because {AD} = {AQ},\\angle {PAQ} = {60}^{ \\circ }\\therefore {\\Delta PQA}$ 为等边编形" + ], + "image_footnote": [], + "bbox": [ + 289, + 468, + 382, + 541 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/cc65c13d277e9ea68905a1827188e40a0b1a26c22d56782f69ae34435bd7199c.jpg", + "image_caption": [ + "" + ], + "image_footnote": [], + "bbox": [ + 387, + 468, + 480, + 541 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1d995c919204a04b802081d68b7fe699b7fa05a56b5944bc9866fb6c096f8bbd.jpg", + "image_caption": [ + "\"An artistic flat design poster featuring sushi on a plate." + ], + "image_footnote": [], + "bbox": [ + 94, + 555, + 187, + 628 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9f4415aee014082399b309a3357adcbaae9b40c90e3c1a50af89c91599bcdeb9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 192, + 555, + 287, + 628 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "n poster featuring su", + "bbox": [ + 192, + 628, + 277, + 638 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/41ea7c2a5cdce56be20445b35bcd686de828e4963016efce7e7d389722e8354a.jpg", + "image_caption": [ + "ni on a plate." + ], + "image_footnote": [], + "bbox": [ + 289, + 555, + 382, + 628 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ea342a221aa9de8de74a6125ad4a9224a380bcba81a71f68ae23013e4aa89621.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 387, + 555, + 480, + 628 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "", + "bbox": [ + 387, + 641, + 480, + 715 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "(c) Fine-tune", + "bbox": [ + 398, + 726, + 468, + 737 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b9d11ffd8ec5106340840ed8a63fbac74871e7f9beb57d89a12b4bf32e64a5da.jpg", + "image_caption": [ + "\"Thundery sky.\"", + "(a) Condition" + ], + "image_footnote": [], + "bbox": [ + 94, + 641, + 187, + 714 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d872e269915874e5e2043056acbdab71d685bf440b2834f2bdf306c680f3485e.jpg", + "image_caption": [ + "(b) Zero-shot", + "Figure 6. Qualitative results of our zero-shot approach. The first four rows show results from manual color conditions, while the fifth row presents results from in-the-wild imagery condition." + ], + "image_footnote": [], + "bbox": [ + 192, + 641, + 287, + 715 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e0938ebcd9865118665b929cc54306f6a28fea8f5ceea9ea8b3d330d0d69a847.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 289, + 641, + 382, + 715 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "o-shot", + "bbox": [ + 289, + 726, + 321, + 737 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "", + "bbox": [ + 387, + 641, + 480, + 715 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "(c) Fine-tune", + "bbox": [ + 398, + 726, + 468, + 737 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We report the performance scores for the latent diffusion case study in Tab. 2. As shown, our method generally outperforms the baselines in most of the metrics. The baselines exhibit significant failures in metrics reflecting the disentanglement (CLIP-score), accuracy and completeness (CD-A", + "bbox": [ + 89, + 825, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/dd591665348ea62800200938301f17c91a642ce8638599f00c7df177920d99ac.jpg", + "table_caption": [ + "Table 2. Quantitative results of latent diffusion. In each metric column, the left and right numbers correspond to the scores from the settings of manually sampled color condition and in-the-wild image color condition, respectively." + ], + "table_footnote": [], + "table_body": "
MappingConditionFID ↓CD-A ↓CD-C ↓
XX63.9; 14940.6; 42.442.6; 39.0
X✓57.9; 88.115.4; 10.618.4; 15.2
✓X59.8; 45.70.00; 0.005.10; 4.62
✓✓57.5; 45.60.00; 0.004.12; 3.65
", + "bbox": [ + 514, + 291, + 906, + 377 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3. Ablation study on color mapping and color condition. For each metric, the left and right numbers are from the Oxford-flower [32] and Microsoft-emoji [8] datasets, respectively.", + "bbox": [ + 511, + 387, + 906, + 430 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "and CD-C scores) of generated colors in synthesized images. When provided with manual color conditions, our zero-shot version achieves worse FID score than the fine-tuned version does. This aligns with our observations in Sec. 4.4 indicating imagery details lacking caused by the zero-shot version. The zero-shot setting's very low CD scores also indicate color overfitting caused by mechanically mapping the manual color conditions.", + "bbox": [ + 511, + 455, + 906, + 575 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ablation study. We validate the effectiveness of important components in our method in Tab. 3. Specifically, we investigate the role of the color mappings $f$ , and the inputting of the color condition $c$ (those in Fig. 2c). When mappings are not used, noisy samples are not altered and only color conditions are fed to the diffusion model. As shown in Tab. 3, the mappings and color conditions significantly boost up the performance of our method.", + "bbox": [ + 511, + 577, + 908, + 699 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion and Discussion", + "text_level": 1, + "bbox": [ + 511, + 709, + 761, + 724 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We propose a color alignment method that operates on the diffusion process of diffusion models. Diffused colors are mapped to conditional colors across diffusion steps, enabling a continuous pathway for synthesizing target color patterns while maintaining the creativity of diffusion models.", + "bbox": [ + 511, + 734, + 906, + 808 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "There are issues that would be addressed in our future work. First, while our method can condition on color values and proportions, it is worthwhile to understand the influence of the spatial information in the color conditions. Second, it is of great interest to extend our color alignment to video and 3D data, potentially with integration of multi-view attributes.", + "bbox": [ + 511, + 810, + 908, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "28453", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 91, + 90, + 250, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Binh-Son Hua is supported by Research Ireland under the Research Ireland Frontiers for the Future Programme - Project, award number 22/FFP-P/11522.", + "bbox": [ + 89, + 114, + 485, + 160 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 176, + 187, + 193 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Text-to-image-2m: A high-quality, diverse text-to-image training dataset. https://huggingface.co/datasets/jackyhate/text-to-image-2M.4", + "[2] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 4", + "[3] Vassilis Athitsos and Stan Sclaroff. Estimating 3d hand pose from a cluttered image. In 2003 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, 2003. Proceedings., pages II-432. IEEE, 2003. 6", + "[4] Arpit Bansal, Eitan Borgnia, Hong-Min Chu, Jie Li, Hamid Kazemi, Furong Huang, Micah Goldblum, Jonas Geiping, and Tom Goldstein. Cold diffusion: Inverting arbitrary image transforms without noise. Advances in Neural Information Processing Systems, 36, 2024. 2", + "[5] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18392-18402, 2023. 2", + "[6] Mingdeng Cao, Xintao Wang, Zhongang Qi, Ying Shan, Xiaohu Qie, and Yinqiang Zheng. Masactrl: Tuning-free mutual self-attention control for consistent image synthesis and editing. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22560-22570, 2023. 2", + "[7] Jiwoo Chung, Sangeek Hyun, and Jae-Pil Heo. Style injection in diffusion: A training-free approach for adapting large-scale diffusion models for style transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8795-8805, 2024. 2", + "[8] Jason Custer and Spencer Nelson. Microsoft fluent emoji. https://github.com/microsoft/fluentui-emoji.4,7,8", + "[9] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2", + "[10] Ziv Epstein, Aaron Hertzmann, Laura Herman, Robert Mahari, Morgan R. Frank, Matthew Groh, Hope Schroeder, Amy Smith, Memo Akten, Jessica Fjeld, Hany Farid, Neil Leach, Alex \"Sandy\" Pentland, and Olga Russakovsk. Art and the science of generative ai. Science, 380(6650):1110-1111, 2023. 1", + "[11] Martin Nicolas Everaert, Marco Bocchio, Sami Arpa, Sabine Susstrunk, and Radhakrishna Achanta. Diffusion in style. In" + ], + "bbox": [ + 93, + 202, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2251-2261, 2023. 2", + "[12] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139-144, 2020. 2", + "[13] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 2", + "[14] Amir Hertz, Andrey Voynov, Shlomi Fruchter, and Daniel Cohen-Or. Style aligned image generation via shared attention. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4775–4785, 2024. 1, 2, 3, 4, 5, 6, 8", + "[15] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning. arXiv preprint arXiv:2104.08718, 2021. 5", + "[16] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 5", + "[17] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 4", + "[18] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1, 2, 3, 4, 5, 7", + "[19] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1125-1134, 2017. 2", + "[20] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 2", + "[21] Bahjat Kawar, Shiran Zada, Oran Lang, Omer Tov, Huiwen Chang, Tali Dekel, Inbar Mosseri, and Michal Irani. Imag: Text-based real image editing with diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6007-6017, 2023. 2", + "[22] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 2", + "[23] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 2", + "[24] Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11461-11471, 2022. 2", + "[25] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jiajun Wu, Jun-Yan Zhu, and Stefano Ermon. Sdedit: Guided image" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "28454", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "synthesis and editing with stochastic differential equations. arXiv preprint arXiv:2108.01073, 2021. 2", + "[26] Mehdi Mirza and Simon Osindero. Conditional generative adversarial nets. arXiv preprint arXiv:1411.1784, 2014. 2", + "[27] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6038–6047, 2023. 2", + "[28] Chong Mou, Xintao Wang, Jiechong Song, Ying Shan, and Jian Zhang. *Dragondiffusion: Enabling drag-style manipulation on diffusion models.* arXiv preprint arXiv:2307.02421, 2023. 2", + "[29] Chong Mou, Xintao Wang, Jiechong Song, Ying Shan, and Jian Zhang. Diffeditor: Boosting accuracy and flexibility on diffusion-based image editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8488-8497, 2024. 2", + "[30] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, and Ying Shan. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 4296–4304, 2024. 1, 2, 4, 5, 6, 8", + "[31] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International conference on machine learning, pages 8162-8171. PMLR, 2021. 2", + "[32] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In 2008 Sixth Indian conference on computer vision, graphics & image processing, pages 722-729. IEEE, 2008. 4, 7, 8", + "[33] Gaurav Parmar, Krishna Kumar Singh, Richard Zhang, Yijun Li, Jingwan Lu, and Jun-Yan Zhu. Zero-shot image-to-image translation. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-11, 2023. 2", + "[34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 5", + "[35] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1(2): 3, 2022. 1, 2", + "[36] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 2, 3, 4, 8", + "[37] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22500-22510, 2023. 2" + ], + "bbox": [ + 91, + 92, + 483, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[38] Chitwan Sahara, William Chan, Huiwen Chang, Chris Lee, Jonathan Ho, Tim Salimans, David Fleet, and Mohammad Norouzi. Palette: Image-to-image diffusion models. In ACM SIGGRAPH 2022 conference proceedings, pages 1-10, 2022. 2", + "[39] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural information processing systems, 35:36479-36494, 2022. 1, 2", + "[40] Yujun Shi, Chuhui Xue, Jun Hao Liew, Jiachun Pan, Hanshu Yan, Wenqing Zhang, Vincent YF Tan, and Song Bai. Dragdiffusion: Harnessing diffusion models for interactive point-based image editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8839-8849, 2024. 2", + "[41] Kihyuk Sohn, Honglak Lee, and Xinchen Yan. Learning structured output representation using deep conditional generative models. Advances in neural information processing systems, 28, 2015. 2", + "[42] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020.2", + "[43] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017. 2", + "[44] Patrick von Platen, Suraj Patil, Anton Lozhkov, Pedro Cuenca, Nathan Lambert, Kashif Rasul, Mishig Davaadorj, Dhruv Nair, Sayak Paul, Steven Liu, William Berman, Yiyi Xu, and Thomas Wolf. Diffusers: State-of-the-art diffusion models. 4", + "[45] Shaoan Xie, Zhifei Zhang, Zhe Lin, Tobias Hinz, and Kun Zhang. Smartbrush: Text and shape guided object inpainting with diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22428-22437, 2023. 2", + "[46] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023. 1, 2, 3, 4, 5, 6, 8", + "[47] Jiwen Yu, Yinhuai Wang, Chen Zhao, Bernard Ghanem, and Jian Zhang. Freedom: Training-free energy-guided conditional diffusion model. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 23174-23184, 2023. 2", + "[48] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 1, 2, 3, 4, 5, 6, 8", + "[49] Yuxin Zhang, Nisha Huang, Fan Tang, Haibin Huang, Chongyang Ma, Weiming Dong, and Changsheng Xu. Inversion-based style transfer with diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10146-10156, 2023. 2" + ], + "bbox": [ + 516, + 92, + 906, + 864 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "28455", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Color Alignment in Diffusion/4b8bcc3c-a5c5-4a38-89da-b85960cd1b00_model.json b/2025/Color Alignment in Diffusion/4b8bcc3c-a5c5-4a38-89da-b85960cd1b00_model.json new file mode 100644 index 0000000000000000000000000000000000000000..a8607dbe10ea9afa0da06df12373695a4ca97ca9 --- /dev/null +++ b/2025/Color Alignment in Diffusion/4b8bcc3c-a5c5-4a38-89da-b85960cd1b00_model.json @@ -0,0 +1,2717 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.351, + 0.131, + 0.648, + 0.152 + ], + "angle": 0, + "content": "Color Alignment in Diffusion" + }, + { + "type": "text", + "bbox": [ + 0.2, + 0.193, + 0.796, + 0.213 + ], + "angle": 0, + "content": "Ka Chun Shum1 Binh-Son Hua2 Duc Thanh Nguyen3 Sai-Kit Yeung1" + }, + { + "type": "text", + "bbox": [ + 0.118, + 0.223, + 0.879, + 0.243 + ], + "angle": 0, + "content": "\\(^{1}\\)Hong Kong University of Science and Technology \\(^{2}\\)Trinity College Dublin \\(^{3}\\)Deakin University" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.287, + 0.327, + 0.303 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.319, + 0.486, + 0.606 + ], + "angle": 0, + "content": "Diffusion models have shown great promise in synthesizing visually appealing images. However, it remains challenging to condition the synthesis at a fine-grained level, for instance, synthesizing image pixels following some generic color pattern. Existing image synthesis methods often produce contents that fall outside the desired pixel conditions. To address this, we introduce a novel color alignment algorithm that confines the generative process in diffusion models within a given color pattern. Specifically, we project diffusion terms, either imagery samples or latent representations, into a conditional color space to align with the input color distribution. This strategy simplifies the prediction in diffusion models within a color manifold while still allowing plausible structures in generated contents, thus enabling the generation of diverse contents that comply with the target color pattern. Experimental results demonstrate our state-of-the-art performance in conditioning and controlling of color pixels, while maintaining on-par generation quality and diversity in comparison with regular diffusion models." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.634, + 0.222, + 0.649 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.659, + 0.485, + 0.825 + ], + "angle": 0, + "content": "Recently, diffusion models [18] and their large-scale text-conditioned variants [36, 39] have demonstrated impressive quality and diversity of generated contents in image synthesis. Subsequent works have explored leveraging this capability for possible controllability, incorporating various constraints, such as structural [30, 35, 48] and reference constraints [14, 46], which typically serve as additional inputs. However, due to the inherent randomness in the diffusion process, it remains challenging to control generated contents at a fine-grained level, for instance, synthesizing image pixels following some given content and generic color distribution." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.484, + 0.901 + ], + "angle": 0, + "content": "In this paper, we focus on such a fine-grained control over pixel colors in synthesized images to meet with three color-conditioned generation criteria: the accuracy in adhering to given color values; the completeness of covering proportions of the specified colors; and the disentanglement of the colors" + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.286, + 0.613, + 0.36 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.361, + 0.61, + 0.433 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.616, + 0.286, + 0.71, + 0.359 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.616, + 0.36, + 0.709, + 0.433 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.712, + 0.286, + 0.805, + 0.359 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.712, + 0.36, + 0.805, + 0.433 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.81, + 0.286, + 0.905, + 0.359 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.81, + 0.36, + 0.904, + 0.433 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.44, + 0.611, + 0.513 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.513, + 0.611, + 0.587 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.527, + 0.589, + 0.599, + 0.6 + ], + "angle": 0, + "content": "(a) Condition" + }, + { + "type": "image", + "bbox": [ + 0.617, + 0.44, + 0.709, + 0.512 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.617, + 0.513, + 0.709, + 0.586 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.687, + 0.589, + 0.733, + 0.6 + ], + "angle": 0, + "content": "(b) Ours" + }, + { + "type": "image", + "bbox": [ + 0.712, + 0.44, + 0.805, + 0.512 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.712, + 0.513, + 0.805, + 0.586 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.811, + 0.44, + 0.904, + 0.512 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.811, + 0.513, + 0.904, + 0.586 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.822, + 0.589, + 0.893, + 0.6 + ], + "angle": 0, + "content": "(c)Baselines" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.613, + 0.907, + 0.738 + ], + "angle": 0, + "content": "Figure 1. Examples of color-conditioned generation from different methods. Color conditions (a) can be derived from imagery (top two rows) or manual drawing (bottom two rows). Input text prompts are \"painting\" and \"pen\" (1st & 3rd rows), and \"boat\" (2nd & 4th rows). Our method (b) can generate pixels closely aligned with the color conditions and effectively disentangle the colors into different structures. Existing baselines [14, 36, 46, 48] (from top to bottom) (c) struggle with either the semantics (the first two rows) or colors (the last two rows) of generated contents." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.905, + 0.81 + ], + "angle": 0, + "content": "into various structures. Such a color-conditioned generation ability would be a great aid to numerous downstream tasks in creative artwork and graphic design [10]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Different attempts of color-conditioned image generation are visualized in Fig. 1, where previous works often produce suboptimal results that do not align with the aforementioned criteria. In this paper, we propose an effective yet control-easy method facilitating color alignment in diffusion models for color-conditioned image synthesis. Specifically," + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "28446" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.349 + ], + "angle": 0, + "content": "our method accepts a color condition via a color pattern, determining color values and their proportions in a synthesized image. Different from the regular diffusion approach, we perform the color alignment on intermediate samples or latent representations during the diffusion process prior to inputting them into a denoising model for image synthesis, thereby facilitating the perception of the conditional colors in diffusion models while preserving the overall attribution of the diffusion process. Furthermore, our method can support various settings: re-training of a color-aligned diffusion model; fine-tuning of a color-aligned diffusion model from a pre-trained regular diffusion model; and zero-shot approximation (training-free) of the color alignment process with a pre-trained diffusion model. We validate our method by demonstrating its capabilities across various color conditions and compare it with existing baselines on benchmark datasets. In summary, we make the following contributions:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.349, + 0.485, + 0.439 + ], + "angle": 0, + "content": "- We address a challenging task: fine-grained color-conditioned image synthesis. We aim to generate images that closely follow a specified color pattern that defines color values and proportions without requiring an explicit structure. Our approach thus enhances the controllability and flexibility of image generation." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.439, + 0.485, + 0.544 + ], + "angle": 0, + "content": "- We propose a novel color alignment method that operates on the diffusion process of diffusion models, allowing the diffusion to effectively perceive color conditions while maintaining the quality and diversity of generated contents. We derive various settings including re-training, fine-tuning, and a zero-shot approximation, to adapt our color alignment method into different down-stream needs." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.545, + 0.485, + 0.605 + ], + "angle": 0, + "content": "- We conduct extensive experiments to validate our proposed method in various scenarios and controllability settings. Our results demonstrate state-of-the-art performance compared with existing baselines." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.349, + 0.485, + 0.605 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.617, + 0.229, + 0.632 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.642, + 0.486, + 0.839 + ], + "angle": 0, + "content": "Neural image synthesis has started with Variational Autoencoders (VAEs) [22, 41, 43] and Generative Adversarial Networks (GANs) [12, 19, 20, 26], followed then by transformer-based methods [9, 23] and diffusion models [18, 31, 42]. Among these, Denoising Diffusion Probabilistic Models (DDPM) [18] have attracted considerable attention due to their generation quality and ease of extension. These models have been further developed into text-to-image (T2I) diffusion techniques [35, 36, 39], which have been boosted recently to achieve greater controllability through incorporation of extra conditions [5, 37, 38, 48]. Our method follows the general diffusion framework with a focus on fine-grained color-conditioned image synthesis." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.486, + 0.902 + ], + "angle": 0, + "content": "Reference-based conditional diffusion. An intuitive condition for diffusion models is the use of an image reference. Structural conditions, such as edge maps and semantic maps," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.909, + 0.304 + ], + "angle": 0, + "content": "have been recently studied in [24, 25, 30, 38, 45, 48]. These methods strictly apply spatial constraints given in the reference to align generated pixels. A more relaxed approach is image editing [5, 6, 13, 21, 27], where the structure and appearance in the reference loosely guide the generation of target images. Those works aim to achieve shifts in appearance while permitting a small range of spatial edits, such as changes in the size and pose of objects. Several methods [14, 37, 46] condition only on abstract concepts in the reference, such as image/object style, enabling more creative synthesis. In this work, we disentangle colors in the reference from their original spatial structure while preserving the color values and proportions. Our method allows a variety of user-provided color materials to construct diverse structures." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.33, + 0.911, + 0.498 + ], + "angle": 0, + "content": "Sample altering in diffusion. As indicated in [25], diffusion from the same noisy sample can result in similar outputs. Diffusion inversion [27] is commonly used to construct samples, used directly [27, 33] or mapped to another diffusion process [7, 29, 49] to diverge outputs. Several methods update intermediate latent representations with gradients calculated from external modules, e.g., other networks [47], images [40], features [28]. The Gaussian noise is defined [11] using estimated mean and variance to improve model convergence. Cold Diffusion [4] replaces the noising process with image degradations such as blurring and masking." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.515, + 0.686, + 0.533 + ], + "angle": 0, + "content": "3. Proposed method" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.542, + 0.909, + 0.709 + ], + "angle": 0, + "content": "We first in Sec. 3.1 review the regular diffusion process. In Sec. 3.2, we introduce our color alignment technique under its simplest setting, i.e., re-training of a diffusion model with color alignment. We then in Sec. 3.3 extend our technique to latent diffusion models, where we address typical challenges in latent space diffusion and demonstrate our method's ability via fine-tuning of a pre-trained latent diffusion model to a color-aligned one. Finally, in Sec. 3.4, we present a zero-shot variant of our method that achieves similar generation quality without extra training. We provide an overview of our pipeline and the regular diffusion in Fig. 2." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.723, + 0.729, + 0.739 + ], + "angle": 0, + "content": "3.1. Diffusion preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.747, + 0.908, + 0.808 + ], + "angle": 0, + "content": "Diffusion models [18] could be seen as a variant of VAEs [22] that map a data distribution \\(\\mathcal{X}(\\mathbf{x}_0)\\) from a Gaussian distribution \\(\\mathcal{N}(\\mathbf{0},\\mathbf{I})\\). In the forward process, a sample \\(\\mathbf{x}_t\\) is encoded by adding a Gaussian noise to \\(\\mathbf{x}_0\\):" + }, + { + "type": "equation", + "bbox": [ + 0.568, + 0.822, + 0.907, + 0.839 + ], + "angle": 0, + "content": "\\[\n\\mathbf {x} _ {t} = \\sqrt {\\bar {\\alpha} _ {t}} \\mathbf {x} _ {0} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\boldsymbol {\\epsilon}, \\quad \\boldsymbol {\\epsilon} \\sim \\mathcal {N} (\\mathbf {0}, \\mathbf {I}) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.854, + 0.905, + 0.884 + ], + "angle": 0, + "content": "where \\(t \\in [1, T]\\) is the time step associated with a noise strength \\(\\bar{\\alpha}_t\\) where \\((1 - \\bar{\\alpha}_1) \\approx 0\\) and \\((1 - \\bar{\\alpha}_T) \\approx 1\\)." + }, + { + "type": "text", + "bbox": [ + 0.534, + 0.886, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Then, a denoising model \\(\\epsilon_{\\theta}\\) with parameters \\(\\theta\\) learns to" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28447" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.457, + 0.108 + ], + "angle": 0, + "content": "reverse the process in Eq. 1 by training with a loss \\(\\mathcal{L}\\) as," + }, + { + "type": "equation", + "bbox": [ + 0.135, + 0.12, + 0.484, + 0.139 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\nabla_ {\\theta} \\| \\epsilon - \\epsilon_ {\\theta} (\\mathbf {x} _ {t}, t) \\| _ {2} ^ {2} \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.13, + 0.141, + 0.484, + 0.172 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {x}} _ {0} = \\frac {1}{\\sqrt {\\bar {\\alpha} _ {t}}} \\left(\\mathbf {x} _ {t} - \\sqrt {1 - \\bar {\\alpha} _ {t}} \\boldsymbol {\\epsilon} _ {\\theta} \\left(\\mathbf {x} _ {t}, t\\right)\\right) \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.114, + 0.174, + 0.484, + 0.207 + ], + "angle": 0, + "content": "\\[\n\\mathbf {x} _ {t - 1} = \\frac {\\sqrt {\\bar {\\alpha} _ {t - 1}} \\beta_ {t}}{1 - \\bar {\\alpha} _ {t}} \\hat {\\mathbf {x}} _ {0} + \\frac {\\sqrt {\\alpha_ {t}} (1 - \\bar {\\alpha} _ {t - 1})}{1 - \\bar {\\alpha} _ {t}} \\mathbf {x} _ {t} + \\sigma_ {t} \\boldsymbol {\\varepsilon} \\quad (4)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.219, + 0.484, + 0.25 + ], + "angle": 0, + "content": "where \\(\\hat{\\mathbf{x}}_0\\) is the predicted sample. \\(\\alpha ,\\beta ,\\sigma \\in \\mathbb{R}\\) and \\(\\varepsilon \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})\\) are scheduler-related parameters [18]." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.262, + 0.422, + 0.278 + ], + "angle": 0, + "content": "3.2. Color-aligned diffusion in image space" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.285, + 0.485, + 0.496 + ], + "angle": 0, + "content": "As presented in Sec. 3.1, the regular diffusion model learns to reconstruct \\(\\mathbf{x}_0\\) from \\(\\mathcal{N}(\\mathbf{0},\\mathbf{I})\\) during training and allows sampling-free generation in testing, i.e., unconditional generation. To condition the diffusion process with color conditions, existing methods pass color text prompts [36, 46] or referenced images [14, 46, 48] in parallel with \\(\\mathbf{x}_t\\) untouched. We found this approach loosely constrains the sampled data manifold, i.e., \\(\\hat{\\mathbf{x}}_0\\) can be sampled with out-of-bound colors. This is because the conditions are not involved in the sampling process (in construction of \\(\\mathbf{x}_t\\)) and the denoising model \\(\\epsilon_{\\theta}\\) hence leans towards the unconditional generation. Moreover, colors are learned in conjunction with spatial information given in referenced images, limiting the diversity of generated contents (see Fig. 1)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.498, + 0.486, + 0.634 + ], + "angle": 0, + "content": "In this paper, we take a color pattern \\(\\mathbf{c}\\), determining color values and their proportions in synthesis, as a color condition. The color pattern can be an imagery or hand-drawing (see Fig. 1). To effectively constrain the sampling manifold with color condition, we involve the color pattern \\(\\mathbf{c}\\) into the diffusion process via a new denoising model \\(\\epsilon_{\\theta}(\\mathbf{x}_t,t,\\mathbf{c})\\). In addition, we relax the spatial information in \\(\\mathbf{c}\\) by using a color alignment function \\(f(\\mathbf{x}_t,\\mathbf{c})\\) that produces a color-aligned image from an image \\(\\mathbf{x}_t\\) and color pattern \\(\\mathbf{c}\\) as," + }, + { + "type": "equation", + "bbox": [ + 0.159, + 0.647, + 0.484, + 0.675 + ], + "angle": 0, + "content": "\\[\nf \\left(\\mathbf {x} _ {t}, \\mathbf {c}\\right) [ p ] = \\underset {\\mathbf {c} [ q ]} {\\arg \\min } \\| \\mathbf {x} _ {t} [ p ] - \\mathbf {c} [ q ] \\| _ {2} ^ {2} \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.688, + 0.412, + 0.705 + ], + "angle": 0, + "content": "where \\(\\mathbf{x}_t[p]\\) and \\(\\mathbf{c}[q]\\) are colors at pixels \\(p\\) and \\(q\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.486, + 0.902 + ], + "angle": 0, + "content": "We visualize the diffusion process in Fig. 3. Our color alignment technique benefits the sampling in several aspects. First, it enforces the generated pixel colors within the target space, guaranteeing the accuracy. Second, its non-spatial manner allows pixels distribute more freely, fulfilling the color disentanglement. For the sake of completeness, we further concatenate \\(\\mathbf{c}\\) to the model's input as an extra hint. However, to completely break down the spatial information, we opt to randomly permute the pixel locations in \\(\\mathbf{c}\\), resulting in \\(\\psi(\\mathbf{c})\\) with \\(\\psi\\) being an image permutation operator. Such a permutation allows us to directly take \\(\\mathbf{x}_0\\) as color condition in training time, avoiding extra data requirement and data leakage. Finally, training and sampling of our color" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.09, + 0.639, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.543, + 0.245, + 0.681, + 0.258 + ], + "angle": 0, + "content": "(a) Regular diffusion [36]" + }, + { + "type": "image", + "bbox": [ + 0.71, + 0.09, + 0.904, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.713, + 0.245, + 0.905, + 0.258 + ], + "angle": 0, + "content": "(b) Regular diffusion with condition" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.269, + 0.709, + 0.469 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.514, + 0.475, + 0.71, + 0.501 + ], + "angle": 0, + "content": "(c) Our color-aligned diffusion for retraining or fine-tuning" + }, + { + "type": "image", + "bbox": [ + 0.714, + 0.269, + 0.904, + 0.469 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.712, + 0.475, + 0.907, + 0.501 + ], + "angle": 0, + "content": "(d) Our zero-shot approximation of the color-aligned diffusion" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.512, + 0.908, + 0.582 + ], + "angle": 0, + "content": "Figure 2. Pipelines of the regular diffusion (a) and its color-conditioned version (b), our color-aligned diffusion (c), where noisy samples \\(\\mathbf{x}_t\\) are aligned with conditioned colors, and our zero-shot version (d), where color alignment is directly applied to predicted sample \\(\\hat{\\mathbf{x}}_0\\)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.609, + 0.866, + 0.625 + ], + "angle": 0, + "content": "aligned diffusion process (Fig. 2c) are formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.54, + 0.635, + 0.775, + 0.654 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\nabla_ {\\boldsymbol {\\theta}} \\| \\boldsymbol {\\epsilon} ^ {\\prime} - \\boldsymbol {\\epsilon} _ {\\boldsymbol {\\theta}} \\big (f (\\mathbf {x} _ {t}, \\mathbf {c}), t, \\mathbf {c} \\big) \\| _ {2} ^ {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.54, + 0.655, + 0.907, + 0.689 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\epsilon} ^ {\\prime} = \\frac {f (\\mathbf {x} _ {t} , \\mathbf {c}) - \\sqrt {\\bar {\\alpha} _ {t}} \\mathbf {x} _ {0}}{\\sqrt {1 - \\bar {\\alpha} _ {t}}}, \\quad \\mathbf {c} = \\psi (\\mathbf {x} _ {0}) \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.535, + 0.69, + 0.907, + 0.723 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {x}} _ {0} = \\frac {1}{\\sqrt {\\bar {\\alpha} _ {t}}} \\left(f \\left(\\mathbf {x} _ {t}, \\mathbf {c}\\right) - \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon_ {\\theta} \\left(f \\left(\\mathbf {x} _ {t}, \\mathbf {c}\\right), t, \\mathbf {c}\\right)\\right) \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.733, + 0.909, + 0.84 + ], + "angle": 0, + "content": "where \\(\\epsilon^{\\prime}\\) is the adapted noise complemented with the effect of \\(f\\). Notably, we alter only what \\(\\epsilon_{\\theta}\\) sees and predicts. The parameters \\(\\theta\\) thus can be learned more easily with color condition compared with the regular one. All other nonlearnable steps, e.g., the regular forward process and Eq. 4, remain unchanged for the ease of sampling. For more technical visualizations, please refer to the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.849, + 0.843, + 0.865 + ], + "angle": 0, + "content": "3.3. Color-aligned diffusion in latent space" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.908, + 0.901 + ], + "angle": 0, + "content": "Training of diffusion models in their latent space is a common practice. In this section, we extend our color alignment" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28448" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.121 + ], + "angle": 0, + "content": "technique to latent diffusion models via fine-tuning of a pre-trained latent model to achieve color alignment." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.122, + 0.484, + 0.304 + ], + "angle": 0, + "content": "Following the Stable Diffusion in [36], we encode \\(\\mathbf{x}_0\\) into its latent representation \\(\\mathbf{z}_0\\). We then train our model with alignment in the latent space, i.e., replacing \\(\\mathbf{x}\\) by \\(\\mathbf{z}\\) in Eqs. 6 and 7. However, when high-frequency structural information is present in \\(\\mathbf{x}_0\\), unintended colors and local structural information can be encoded in \\(\\mathbf{z}_0\\). We discovered that blurring \\(\\mathbf{x}_0\\) prior to encoding it enhances the color disentanglement and accuracy, with minimal loss of the completeness in generated results. Also, we found that suspending color alignment at late time steps (e.g., \\(t < 0.2T\\) where diffusion model is fixing details) allows a proper refinement of the latent, making latent decoding towards more photo-realistic details." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.314, + 0.38, + 0.33 + ], + "angle": 0, + "content": "3.4. Zero-shot color-aligned diffusion" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.336, + 0.484, + 0.458 + ], + "angle": 0, + "content": "To mitigate the computational burden of training, we introduce a zero-shot approximation of our color alignment method. As shown in Eqs. 6 and 7, the training aims to reconstruct \\(\\hat{\\mathbf{x}}_0\\) whose color distribution follows the color condition \\(\\mathbf{c} = \\psi (\\mathbf{x}_0)\\). Under the zero-shot setting, to achieve this ability without learning the parameters \\(\\theta\\), we directly map an unconditionally generated \\(\\hat{\\mathbf{x}}_0\\) (from Eq. 3) to \\(\\mathbf{c}\\) during the sampling process using a color alignment function \\(g(\\hat{\\mathbf{x}}_0,\\mathbf{c})\\):" + }, + { + "type": "equation", + "bbox": [ + 0.171, + 0.484, + 0.484, + 0.511 + ], + "angle": 0, + "content": "\\[\ng \\left(\\hat {\\mathbf {x}} _ {0}, \\mathbf {c}\\right) = \\underset {\\psi (\\mathbf {c})} {\\arg \\min } \\| \\hat {\\mathbf {x}} _ {0} - \\psi (\\mathbf {c}) \\| _ {2} ^ {2} \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.523, + 0.484, + 0.553 + ], + "angle": 0, + "content": "where we map the unconditional \\(\\hat{\\mathbf{x}}_0\\) to a closest reorganization of the conditional colors \\(\\psi (\\mathbf{c})\\) for color alignment." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.554, + 0.484, + 0.81 + ], + "angle": 0, + "content": "Unlike \\( f \\) defined in Eq. 5, \\( g \\) in Eq. 8 performs one-to-one mapping, i.e., each location in \\( \\mathbf{c} \\) is used for color alignment exactly once. Note that, like \\( f \\), our \\( g \\) also applies non-spatial mapping. This would ensure desired properties (color accuracy, completeness, and disentanglement) from the zero-shot color-aligned diffusion. However, since \\( g \\) is not involved in the learning process, the color-aligned image \\( g(\\hat{\\mathbf{x}}_0,\\mathbf{c}) \\) could be messy, showing invalid structures and outliers (see Fig. 2d). Fortunately, we observed that pre-trained diffusion models have capability of clustering semantics and eliminating outliers at early sampling steps (e.g., \\( t > 0.2T \\)). We also suspend the color alignment at late time steps to prevent excessive color changes and allow the model to refine details. Additionally, such zero-shot approximation is extendable to latent diffusion models. We found the same conclusion for using the blurring strategy as in Sec. 3.3 applies." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Another interest could be replacing \\( f \\) by \\( g \\) in learning. However, we find this not applicable because \\( g(\\mathbf{x}_t,\\mathbf{c}) \\) corrupts the distribution of \\( \\mathbf{x}_t \\) too heavily. Its subsequent adapted noise \\( \\epsilon^{\\prime} \\) (the learning target in Eq. 6) deviates too far from the standard Gaussian \\( \\mathcal{N}(\\mathbf{0},\\mathbf{I}) \\), eliminating the effect of a proper diffusion process." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.091, + 0.646, + 0.108 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.115, + 0.616, + 0.13 + ], + "angle": 0, + "content": "4.1. Datasets" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.138, + 0.907, + 0.32 + ], + "angle": 0, + "content": "We experimented our method in two scenarios: color-aligned diffusion in image space and color-aligned diffusion in latent space. For diffusion in image space, we ran experiments on Oxford-flower [32] (7.17k training and 1.02k test images) and Microsoft-emoji [8] datasets (6.80k training and 0.76k test images) at resolution \\(64 \\times 64\\). For latent diffusion, we experimented on a sub-set of a high-quality generic dataset Text-to-image-2M [1] (300k training and 20k test images) at resolution \\(512 \\times 512\\). To validate the color disentanglement ability of our method, we replaced original prompts in Text-to-image-2M [1] with 50 daily object prompts randomly generated by ChatGPT [2] (e.g., those in Fig. 4)." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.33, + 0.622, + 0.344 + ], + "angle": 0, + "content": "4.2. Baselines" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.352, + 0.906, + 0.442 + ], + "angle": 0, + "content": "Our color-aligned image diffusion version is adapted from the DDPM [18]. Therefore, we compare our method with the DDPM to demonstrate its extra color-conditioned ability without requiring additional data creation or labeling. For the same reason, we also compare our latent diffusion version with the pre-trained Stable Diffusion in [36]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.443, + 0.909, + 0.655 + ], + "angle": 0, + "content": "Given that our color reference can be an in-the-wild image or a manual drawing, we compare our method with SOTA reference-based generation methods. IP-Adapter [46] and T2I-Adapter [30] convert an image reference into features, which are subsequently combined with the model features for image generation. Style-Aligned [14] transfers source image information from its attention layer to the target generation attention layer, thereby preserving spatial and style features. ControlNet-Color is a variant of ControlNet [48] that accepts a predefined pixelized color image as input and generates a refined version in which the colors are approximately aligned with the original image spatially. IP-Adapter [46] and ControlNet-Color [48] require fine-tuning, while Style-Aligned [14] operates as a zero-shot approach." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.665, + 0.727, + 0.68 + ], + "angle": 0, + "content": "4.3. Implementation details" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.687, + 0.907, + 0.763 + ], + "angle": 0, + "content": "We follow default settings in the DDPM [18] and the Stable Diffusion [36] for network architectures and diffusion scheduling, specifically, from the huggingface-diffusers [44]. Further details, e.g., DDPM re-training, Stable Diffusion fine-tuning are presented in our supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.763, + 0.907, + 0.837 + ], + "angle": 0, + "content": "We trained our image diffusion version for 100 epochs and latent diffusion version for \\(160\\mathrm{k}\\) iterations. We performed inference at 50 time steps. We used classifier-free guidance [17] with guidance scale 5. All experiments were executed on two RTX 3090 GPUs with batch size 4." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.849, + 0.682, + 0.865 + ], + "angle": 0, + "content": "4.4. Qualitative study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Color-aligned image diffusion. We visually compare our method with the regular diffusion in DDPM [18] in Fig. 3." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28449" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.092, + 0.477, + 0.193 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.194, + 0.483, + 0.231 + ], + "angle": 0, + "content": "(a) Our color-aligned image diffusion process. The top row shows our training inputs, created from early to late steps in the forward process. The bottom row is a test case from our reverse process, sampled from pure noise." + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.234, + 0.476, + 0.335 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.097, + 0.336, + 0.477, + 0.349 + ], + "angle": 0, + "content": "(b) Regular image diffusion process in DDPM [18] in same order as (a)." + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.352, + 0.476, + 0.504 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.507, + 0.483, + 0.533 + ], + "angle": 0, + "content": "(c) Several color-aligned image synthesis results. From top to bottom rows are condition, our generated results, and regular diffusion's results." + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.544, + 0.485, + 0.587 + ], + "angle": 0, + "content": "Figure 3. Qualitative results of color-aligned image diffusion. (a)-(b) Visualization of the diffusion process by our method and DDPM [18]. (c) Generation results of our method and DDPM [18]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.614, + 0.484, + 0.703 + ], + "angle": 0, + "content": "As shown, our method (Fig. 3a) well restricts the diffusion process to the conditional color space, wherein only whites, yellows, and background blacks are allowed. In contrast, the regular diffusion in DDPM (Fig. 3b) generates out-of-bounds colors during the diffusion process, leading to an unintended reconstruction result." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.484, + 0.825 + ], + "angle": 0, + "content": "Fig. 3c illustrates an immediate application of our color-aligned image diffusion in image rephrasing, i.e., colors in an input image are reorganized by our model to construct new structures. More importantly, this conditional process does not require any additional data (e.g., labeled data or paired data) or pre-defined spatial information (e.g., semantic maps or masks) in both training or inference. This makes our technique scalable to larger datasets." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Color-aligned latent diffusion. We present our coloraligned latent diffusion results in comparison with existing baselines in Fig. 4. As demonstrated, our method (Figs. 4b and 4c) effectively conditions the synthesis on target color inputs. Color values and proportions are aligned, i.e., color" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.198 + ], + "angle": 0, + "content": "accuracy and completeness are maintained. Additionally, our method generates contents that align well with target text prompts, showcasing the ability for color disentangle-ment where colors are rearranged into new structures. Furthermore, with these color-conditioned generation criteria preserved, our method achieves image quality and diversity on par with the original diffusion model." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.199, + 0.909, + 0.35 + ], + "angle": 0, + "content": "In contrast, the baselines produce suboptimal results in various aspects. IP-Adapter [46] (Fig. 4d), T2I-Adapter [30] (Fig. 4e), and Style-Aligned [14] (Fig. 4f) construct target contents on top of an input image reference, thus often struggle to disentangle colors. Since the spatial information of colors is taken into account, these methods likely to fail completely if input reference and target text prompt belong to significantly different domains. ControlNet-Color [48] (Fig. 4g) treats input reference as a weak guidance, leading to failures to accurately preserve color values and proportions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.35, + 0.908, + 0.472 + ], + "angle": 0, + "content": "Generalizability and controllability. Our method also works with manual drawings, for example, a rough color pattern. As shown in Fig. 5, our method (the fine-tuned version) can generate images aligned with the color values and proportions specified in input color patterns. Moreover, even if only a few color values presented, our method can adjust these values to achieve appropriate smoothness, brightness, and shadow in generated contents." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.473, + 0.909, + 0.624 + ], + "angle": 0, + "content": "These manual color conditions can be easily modified to influence the generation process. For example, the amount of each color value can be scaled (Fig. 5a). The color values can be tuned intuitively without the need for careful prompt engineering (Fig. 5b). Additional colors can be incorporated into conditions along with user ideation (Fig. 5c). By controlling target text prompt, we can specify the placement of colors (Fig. 5d). Note that, like the regular diffusion, our method can generate diverse results from the same input condition by sampling different noises." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.624, + 0.909, + 0.777 + ], + "angle": 0, + "content": "Zero-shot diffusion. We showcase the ability of our method in zero-shot diffusion in Fig. 6. Our experiments indicate that, compared with our fine-tuned version, the zero-shot version tends to generate images with less fine-grained details, resulting in issues such as blurry lighting, uniform texture, and shadow missing. Since our zero-shot approach mechanically maps colors without model training, it performs more comparably to the fine-tuned version when applied to well-sampled color conditions, e.g., in-the-wild images in the fifth row in Fig. 6 and those in Fig. 4." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.788, + 0.693, + 0.804 + ], + "angle": 0, + "content": "4.5.Quantitative study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Metrics. We evaluate the quality of generated images using the widely adopted FID score [16]. To assess the color disentanglement, we use the CLIP-Score [15] measuring the similarity between a generated image and a target text prompt within the CLIP [34] embedding space. Since the color accuracy and completeness can be evaluated independ-" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28450" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.091, + 0.087, + 0.907, + 0.759 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.429, + 0.761, + 0.568, + 0.774 + ], + "angle": 0, + "content": "(g) ControlNet-Color [48]" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.785, + 0.908, + 0.814 + ], + "angle": 0, + "content": "Figure 4. Qualitative results of color-aligned latent diffusion. Each input (first row) includes an in-the-wild image as color condition and a target text prompt. Each column presents results of experimented methods using the same input." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.486, + 0.901 + ], + "angle": 0, + "content": "dently of spatial relations, we measure the proximity between the color distributions of a generated image and input color condition. We employ the Chamfer Distance (CD) [3], a commonly used metric for measuring the distance between" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.84, + 0.907, + 0.871 + ], + "angle": 0, + "content": "two point sets, to quantify the distance between generated RGB points \\(\\hat{\\mathbf{x}}_0\\) and input condition RGB points \\(\\mathbf{c}\\):" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "28451" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.089, + 0.91, + 0.557 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.567, + 0.908, + 0.596 + ], + "angle": 0, + "content": "Figure 5. Qualitative results of sampling and editing of input color conditions. All results are generated by our fine-tuned color-aligned latent model. Red arrows represent the generation process. Blue arrows represent the editing of color conditions." + }, + { + "type": "table", + "bbox": [ + 0.153, + 0.609, + 0.846, + 0.687 + ], + "angle": 0, + "content": "
MethodFID ↓CD-A ↓CD-C ↓Re-training Cost ↓Inference Cost ↓
Image Diffusion (Re-train) [18]63.9; 14940.6; 42.442.6; 39.01.001.00
Ours (Image; Re-train)57.5; 45.70.00; 0.004.12; 3.651.061.13
Ours (Image; Zero-shot)50.0; 84.55.59; 5.468.28; 10.40.006.76
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.695, + 0.908, + 0.723 + ], + "angle": 0, + "content": "Table 1. Quantitative results of image diffusion. In each metric column, the left and right numbers correspond to the scores from the Oxford-flower dataset [32] and Microsoft-emoji [8] dataset, respectively." + }, + { + "type": "equation", + "bbox": [ + 0.113, + 0.772, + 0.484, + 0.808 + ], + "angle": 0, + "content": "\\[\n\\operatorname {C D - A} \\left(\\hat {\\mathbf {x}} _ {0}, \\mathbf {c}\\right) = \\frac {\\sum_ {\\hat {\\mathbf {x}} _ {0} [ p ]} \\min _ {\\mathbf {c} [ q ]} \\| \\hat {\\mathbf {x}} _ {0} [ p ] - \\mathbf {c} [ q ] \\| _ {2} ^ {2}}{| \\hat {\\mathbf {x}} _ {0} |} \\tag {9}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.114, + 0.81, + 0.484, + 0.846 + ], + "angle": 0, + "content": "\\[\n\\operatorname {C D - C} \\left(\\hat {\\mathbf {x}} _ {0}, \\mathbf {c}\\right) = \\frac {\\sum_ {\\mathbf {c} [ q ]} \\min _ {\\hat {\\mathbf {x}} _ {0} [ p ]} \\| \\hat {\\mathbf {x}} _ {0} [ p ] - \\mathbf {c} [ q ] \\| _ {2} ^ {2}}{| \\mathbf {c} |} \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.902 + ], + "angle": 0, + "content": "where \\(\\mathrm{CD} = \\mathrm{CD - A} + \\mathrm{CD - C}\\) consists of an accuracy term and a completeness term that measure the set distance from one to another; \\(|\\cdot |\\) denotes set cardinality. Lastly, we evaluate" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.907, + 0.779 + ], + "angle": 0, + "content": "the inference cost of different methods, using the regular diffusion as the standard reference." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Baseline comparison. We first present the performance scores for the image diffusion case study in Tab. 1. The scores clearly indicate that our technique effectively preserves conditioned-colors. Note that, for inference, the color alignment is performed on RGB pixels until the final diffusion step, making the CD-A approach zero. We found that our fine-tuned version requires nearly the same training and inference costs as the regular diffusion model." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "28452" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.164, + 0.089, + 0.835, + 0.228 + ], + "angle": 0, + "content": "
MethodFID ↓CLIPScore ↑CD-A ↓CD-C ↓Inference Cost ↓
Stable Diffusion (Pre-train) [36]72.6; 72.627.3; 27.3166; 27.865.3; 16.01.00
Ours (Latent; Fine-tune)86.7; 69.429.0; 27.473.9; 4.9815.8; 4.871.03
Ours (Latent; Zero-shot)104; 77.928.3; 27.435.2; 2.683.19; 3.933.52
Baseline (IP-Adapter [46])202; 50.925.5; 22.0145; 10.1106; 13.91.44
Baseline (T2I-Adapter [30])134; 77.124.3; 24.9143; 8.3424.7; 7.011.56
Baseline (Style-Aligned [14])177; 73.123.5; 22.945.1; 6.1832.5; 8.505.69
Baseline (ControlNet-Color [48])105; 81.022.9; 23.4129; 13.149.8; 6.541.78
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.239, + 0.907, + 0.269 + ], + "angle": 0, + "content": "Table 2. Quantitative results of latent diffusion. In each metric column, the left and right numbers correspond to the scores from the settings of manually sampled color condition and in-the-wild image color condition, respectively." + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.296, + 0.188, + 0.37 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.095, + 0.37, + 0.245, + 0.38 + ], + "angle": 0, + "content": "\"A professional photo of a bar.\"" + }, + { + "type": "image", + "bbox": [ + 0.193, + 0.296, + 0.288, + 0.37 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.193, + 0.37, + 0.383, + 0.379 + ], + "angle": 0, + "content": "of a bar:" + }, + { + "type": "image", + "bbox": [ + 0.389, + 0.296, + 0.483, + 0.37 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.389, + 0.37, + 0.482, + 0.379 + ], + "angle": 0, + "content": "" + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.383, + 0.188, + 0.456 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.094, + 0.457, + 0.317, + 0.467 + ], + "angle": 0, + "content": "\"Gray modern house, photograph, highly detailed." + }, + { + "type": "image", + "bbox": [ + 0.193, + 0.383, + 0.288, + 0.456 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.193, + 0.457, + 0.317, + 0.467 + ], + "angle": 0, + "content": "photograph, highly detailed." + }, + { + "type": "image", + "bbox": [ + 0.29, + 0.383, + 0.383, + 0.456 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.29, + 0.457, + 0.335, + 0.466 + ], + "angle": 0, + "content": "tailed." + }, + { + "type": "image", + "bbox": [ + 0.388, + 0.383, + 0.482, + 0.456 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.388, + 0.457, + 0.482, + 0.466 + ], + "angle": 0, + "content": "" + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.469, + 0.188, + 0.542 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.094, + 0.543, + 0.277, + 0.553 + ], + "angle": 0, + "content": "\"An animated owl looking at the camera." + }, + { + "type": "image", + "bbox": [ + 0.193, + 0.469, + 0.288, + 0.542 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.193, + 0.543, + 0.278, + 0.553 + ], + "angle": 0, + "content": "king at the camera." + }, + { + "type": "image", + "bbox": [ + 0.29, + 0.469, + 0.383, + 0.542 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.29, + 0.543, + 0.383, + 0.552 + ], + "angle": 0, + "content": "\\( \\because {AD} = {AQ},\\angle {PAQ} = {60}^{ \\circ }\\therefore {\\Delta PQA} \\) 为等边编形" + }, + { + "type": "image", + "bbox": [ + 0.388, + 0.469, + 0.482, + 0.542 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.388, + 0.543, + 0.482, + 0.553 + ], + "angle": 0, + "content": "" + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.556, + 0.188, + 0.629 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.095, + 0.63, + 0.343, + 0.64 + ], + "angle": 0, + "content": "\"An artistic flat design poster featuring sushi on a plate." + }, + { + "type": "image", + "bbox": [ + 0.193, + 0.556, + 0.288, + 0.629 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.193, + 0.63, + 0.278, + 0.639 + ], + "angle": 0, + "content": "n poster featuring su" + }, + { + "type": "image", + "bbox": [ + 0.29, + 0.556, + 0.383, + 0.629 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.29, + 0.63, + 0.383, + 0.639 + ], + "angle": 0, + "content": "ni on a plate." + }, + { + "type": "image", + "bbox": [ + 0.388, + 0.556, + 0.482, + 0.629 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.388, + 0.642, + 0.482, + 0.716 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.399, + 0.727, + 0.47, + 0.738 + ], + "angle": 0, + "content": "(c) Fine-tune" + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.642, + 0.188, + 0.715 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.094, + 0.716, + 0.175, + 0.726 + ], + "angle": 0, + "content": "\"Thundery sky.\"" + }, + { + "type": "image_caption", + "bbox": [ + 0.103, + 0.728, + 0.175, + 0.738 + ], + "angle": 0, + "content": "(a) Condition" + }, + { + "type": "image", + "bbox": [ + 0.193, + 0.642, + 0.288, + 0.716 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.25, + 0.727, + 0.322, + 0.738 + ], + "angle": 0, + "content": "(b) Zero-shot" + }, + { + "type": "image", + "bbox": [ + 0.29, + 0.642, + 0.383, + 0.716 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.29, + 0.727, + 0.322, + 0.738 + ], + "angle": 0, + "content": "o-shot" + }, + { + "type": "image", + "bbox": [ + 0.388, + 0.642, + 0.482, + 0.716 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.399, + 0.727, + 0.469, + 0.738 + ], + "angle": 0, + "content": "(c) Fine-tune" + }, + { + "type": "image_caption", + "bbox": [ + 0.091, + 0.751, + 0.483, + 0.794 + ], + "angle": 0, + "content": "Figure 6. Qualitative results of our zero-shot approach. The first four rows show results from manual color conditions, while the fifth row presents results from in-the-wild imagery condition." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.485, + 0.901 + ], + "angle": 0, + "content": "We report the performance scores for the latent diffusion case study in Tab. 2. As shown, our method generally outperforms the baselines in most of the metrics. The baselines exhibit significant failures in metrics reflecting the disentanglement (CLIP-score), accuracy and completeness (CD-A" + }, + { + "type": "table", + "bbox": [ + 0.515, + 0.292, + 0.908, + 0.378 + ], + "angle": 0, + "content": "
MappingConditionFID ↓CD-A ↓CD-C ↓
XX63.9; 14940.6; 42.442.6; 39.0
X✓57.9; 88.115.4; 10.618.4; 15.2
✓X59.8; 45.70.00; 0.005.10; 4.62
✓✓57.5; 45.60.00; 0.004.12; 3.65
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.388, + 0.908, + 0.431 + ], + "angle": 0, + "content": "Table 3. Ablation study on color mapping and color condition. For each metric, the left and right numbers are from the Oxford-flower [32] and Microsoft-emoji [8] datasets, respectively." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.457, + 0.907, + 0.576 + ], + "angle": 0, + "content": "and CD-C scores) of generated colors in synthesized images. When provided with manual color conditions, our zero-shot version achieves worse FID score than the fine-tuned version does. This aligns with our observations in Sec. 4.4 indicating imagery details lacking caused by the zero-shot version. The zero-shot setting's very low CD scores also indicate color overfitting caused by mechanically mapping the manual color conditions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.578, + 0.909, + 0.7 + ], + "angle": 0, + "content": "Ablation study. We validate the effectiveness of important components in our method in Tab. 3. Specifically, we investigate the role of the color mappings \\( f \\), and the inputting of the color condition \\( c \\) (those in Fig. 2c). When mappings are not used, noisy samples are not altered and only color conditions are fed to the diffusion model. As shown in Tab. 3, the mappings and color conditions significantly boost up the performance of our method." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.71, + 0.763, + 0.725 + ], + "angle": 0, + "content": "5. Conclusion and Discussion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.907, + 0.809 + ], + "angle": 0, + "content": "We propose a color alignment method that operates on the diffusion process of diffusion models. Diffused colors are mapped to conditional colors across diffusion steps, enabling a continuous pathway for synthesizing target color patterns while maintaining the creativity of diffusion models." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.909, + 0.901 + ], + "angle": 0, + "content": "There are issues that would be addressed in our future work. First, while our method can condition on color values and proportions, it is worthwhile to understand the influence of the spatial information in the color conditions. Second, it is of great interest to extend our color alignment to video and 3D data, potentially with integration of multi-view attributes." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28453" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.091, + 0.251, + 0.108 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.116, + 0.486, + 0.161 + ], + "angle": 0, + "content": "Binh-Son Hua is supported by Research Ireland under the Research Ireland Frontiers for the Future Programme - Project, award number 22/FFP-P/11522." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.178, + 0.188, + 0.194 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.203, + 0.484, + 0.244 + ], + "angle": 0, + "content": "[1] Text-to-image-2m: A high-quality, diverse text-to-image training dataset. https://huggingface.co/datasets/jackyhate/text-to-image-2M.4" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.246, + 0.484, + 0.302 + ], + "angle": 0, + "content": "[2] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.304, + 0.484, + 0.358 + ], + "angle": 0, + "content": "[3] Vassilis Athitsos and Stan Sclaroff. Estimating 3d hand pose from a cluttered image. In 2003 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, 2003. Proceedings., pages II-432. IEEE, 2003. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.361, + 0.484, + 0.43 + ], + "angle": 0, + "content": "[4] Arpit Bansal, Eitan Borgnia, Hong-Min Chu, Jie Li, Hamid Kazemi, Furong Huang, Micah Goldblum, Jonas Geiping, and Tom Goldstein. Cold diffusion: Inverting arbitrary image transforms without noise. Advances in Neural Information Processing Systems, 36, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.433, + 0.485, + 0.5 + ], + "angle": 0, + "content": "[5] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18392-18402, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.504, + 0.484, + 0.585 + ], + "angle": 0, + "content": "[6] Mingdeng Cao, Xintao Wang, Zhongang Qi, Ying Shan, Xiaohu Qie, and Yinqiang Zheng. Masactrl: Tuning-free mutual self-attention control for consistent image synthesis and editing. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22560-22570, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.589, + 0.484, + 0.657 + ], + "angle": 0, + "content": "[7] Jiwoo Chung, Sangeek Hyun, and Jae-Pil Heo. Style injection in diffusion: A training-free approach for adapting large-scale diffusion models for style transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8795-8805, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.66, + 0.485, + 0.701 + ], + "angle": 0, + "content": "[8] Jason Custer and Spencer Nelson. Microsoft fluent emoji. https://github.com/microsoft/fluentui-emoji.4,7,8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.704, + 0.484, + 0.786 + ], + "angle": 0, + "content": "[9] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.789, + 0.484, + 0.87 + ], + "angle": 0, + "content": "[10] Ziv Epstein, Aaron Hertzmann, Laura Herman, Robert Mahari, Morgan R. Frank, Matthew Groh, Hope Schroeder, Amy Smith, Memo Akten, Jessica Fjeld, Hany Farid, Neil Leach, Alex \"Sandy\" Pentland, and Olga Russakovsk. Art and the science of generative ai. Science, 380(6650):1110-1111, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[11] Martin Nicolas Everaert, Marco Bocchio, Sami Arpa, Sabine Susstrunk, and Radhakrishna Achanta. Diffusion in style. In" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.203, + 0.485, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.548, + 0.093, + 0.906, + 0.121 + ], + "angle": 0, + "content": "Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2251-2261, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.123, + 0.908, + 0.177 + ], + "angle": 0, + "content": "[12] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139-144, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.907, + 0.234 + ], + "angle": 0, + "content": "[13] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.236, + 0.907, + 0.304 + ], + "angle": 0, + "content": "[14] Amir Hertz, Andrey Voynov, Shlomi Fruchter, and Daniel Cohen-Or. Style aligned image generation via shared attention. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4775–4785, 2024. 1, 2, 3, 4, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.305, + 0.907, + 0.359 + ], + "angle": 0, + "content": "[15] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning. arXiv preprint arXiv:2104.08718, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.362, + 0.907, + 0.431 + ], + "angle": 0, + "content": "[16] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.433, + 0.906, + 0.461 + ], + "angle": 0, + "content": "[17] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.463, + 0.907, + 0.504 + ], + "angle": 0, + "content": "[18] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1, 2, 3, 4, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.506, + 0.907, + 0.56 + ], + "angle": 0, + "content": "[19] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1125-1134, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.562, + 0.907, + 0.617 + ], + "angle": 0, + "content": "[20] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.619, + 0.908, + 0.688 + ], + "angle": 0, + "content": "[21] Bahjat Kawar, Shiran Zada, Oran Lang, Omer Tov, Huiwen Chang, Tali Dekel, Inbar Mosseri, and Michal Irani. Imag: Text-based real image editing with diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6007-6017, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.69, + 0.907, + 0.718 + ], + "angle": 0, + "content": "[22] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.719, + 0.907, + 0.788 + ], + "angle": 0, + "content": "[23] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.79, + 0.908, + 0.87 + ], + "angle": 0, + "content": "[24] Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11461-11471, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[25] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jiajun Wu, Jun-Yan Zhu, and Stefano Ermon. Sdedit: Guided image" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "28454" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.126, + 0.093, + 0.484, + 0.12 + ], + "angle": 0, + "content": "synthesis and editing with stochastic differential equations. arXiv preprint arXiv:2108.01073, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.122, + 0.483, + 0.15 + ], + "angle": 0, + "content": "[26] Mehdi Mirza and Simon Osindero. Conditional generative adversarial nets. arXiv preprint arXiv:1411.1784, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.151, + 0.484, + 0.22 + ], + "angle": 0, + "content": "[27] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6038–6047, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.222, + 0.484, + 0.276 + ], + "angle": 0, + "content": "[28] Chong Mou, Xintao Wang, Jiechong Song, Ying Shan, and Jian Zhang. *Dragondiffusion: Enabling drag-style manipulation on diffusion models.* arXiv preprint arXiv:2307.02421, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.279, + 0.484, + 0.348 + ], + "angle": 0, + "content": "[29] Chong Mou, Xintao Wang, Jiechong Song, Ying Shan, and Jian Zhang. Diffeditor: Boosting accuracy and flexibility on diffusion-based image editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8488-8497, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.35, + 0.484, + 0.432 + ], + "angle": 0, + "content": "[30] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, and Ying Shan. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 4296–4304, 2024. 1, 2, 4, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.434, + 0.484, + 0.489 + ], + "angle": 0, + "content": "[31] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International conference on machine learning, pages 8162-8171. PMLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.491, + 0.484, + 0.547 + ], + "angle": 0, + "content": "[32] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In 2008 Sixth Indian conference on computer vision, graphics & image processing, pages 722-729. IEEE, 2008. 4, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.548, + 0.484, + 0.604 + ], + "angle": 0, + "content": "[33] Gaurav Parmar, Krishna Kumar Singh, Richard Zhang, Yijun Li, Jingwan Lu, and Jun-Yan Zhu. Zero-shot image-to-image translation. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-11, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.606, + 0.484, + 0.687 + ], + "angle": 0, + "content": "[34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.69, + 0.484, + 0.744 + ], + "angle": 0, + "content": "[35] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1(2): 3, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.747, + 0.484, + 0.816 + ], + "angle": 0, + "content": "[36] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 2, 3, 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.818, + 0.484, + 0.899 + ], + "angle": 0, + "content": "[37] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22500-22510, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.484, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.16 + ], + "angle": 0, + "content": "[38] Chitwan Sahara, William Chan, Huiwen Chang, Chris Lee, Jonathan Ho, Tim Salimans, David Fleet, and Mohammad Norouzi. Palette: Image-to-image diffusion models. In ACM SIGGRAPH 2022 conference proceedings, pages 1-10, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.164, + 0.908, + 0.245 + ], + "angle": 0, + "content": "[39] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural information processing systems, 35:36479-36494, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.248, + 0.908, + 0.33 + ], + "angle": 0, + "content": "[40] Yujun Shi, Chuhui Xue, Jun Hao Liew, Jiachun Pan, Hanshu Yan, Wenqing Zhang, Vincent YF Tan, and Song Bai. Dragdiffusion: Harnessing diffusion models for interactive point-based image editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8839-8849, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.332, + 0.908, + 0.386 + ], + "angle": 0, + "content": "[41] Kihyuk Sohn, Honglak Lee, and Xinchen Yan. Learning structured output representation using deep conditional generative models. Advances in neural information processing systems, 28, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.388, + 0.907, + 0.428 + ], + "angle": 0, + "content": "[42] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020.2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.431, + 0.908, + 0.472 + ], + "angle": 0, + "content": "[43] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.474, + 0.907, + 0.529 + ], + "angle": 0, + "content": "[44] Patrick von Platen, Suraj Patil, Anton Lozhkov, Pedro Cuenca, Nathan Lambert, Kashif Rasul, Mishig Davaadorj, Dhruv Nair, Sayak Paul, Steven Liu, William Berman, Yiyi Xu, and Thomas Wolf. Diffusers: State-of-the-art diffusion models. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.531, + 0.908, + 0.599 + ], + "angle": 0, + "content": "[45] Shaoan Xie, Zhifei Zhang, Zhe Lin, Tobias Hinz, and Kun Zhang. Smartbrush: Text and shape guided object inpainting with diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22428-22437, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.601, + 0.908, + 0.655 + ], + "angle": 0, + "content": "[46] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023. 1, 2, 3, 4, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.657, + 0.908, + 0.726 + ], + "angle": 0, + "content": "[47] Jiwen Yu, Yinhuai Wang, Chen Zhao, Bernard Ghanem, and Jian Zhang. Freedom: Training-free energy-guided conditional diffusion model. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 23174-23184, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.728, + 0.908, + 0.783 + ], + "angle": 0, + "content": "[48] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 1, 2, 3, 4, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.785, + 0.908, + 0.865 + ], + "angle": 0, + "content": "[49] Yuxin Zhang, Nisha Huang, Fan Tang, Haibin Huang, Chongyang Ma, Weiming Dong, and Changsheng Xu. Inversion-based style transfer with diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10146-10156, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.865 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28455" + } + ] +] \ No newline at end of file diff --git a/2025/Color Alignment in Diffusion/4b8bcc3c-a5c5-4a38-89da-b85960cd1b00_origin.pdf b/2025/Color Alignment in Diffusion/4b8bcc3c-a5c5-4a38-89da-b85960cd1b00_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..431de3a9261850e6d2e33d02dc86c9213312b0c0 --- /dev/null +++ b/2025/Color Alignment in Diffusion/4b8bcc3c-a5c5-4a38-89da-b85960cd1b00_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d8e8305afe902ffb0c994d7442d74a1739588683a6a3ec4bd8b89e51febfe9f9 +size 4639135 diff --git a/2025/Color Alignment in Diffusion/full.md b/2025/Color Alignment in Diffusion/full.md new file mode 100644 index 0000000000000000000000000000000000000000..d4f3aa53c1268c23200f35fe7c958d5b71f33c13 --- /dev/null +++ b/2025/Color Alignment in Diffusion/full.md @@ -0,0 +1,389 @@ +# Color Alignment in Diffusion + +Ka Chun Shum1 Binh-Son Hua2 Duc Thanh Nguyen3 Sai-Kit Yeung1 + +$^{1}$ Hong Kong University of Science and Technology $^{2}$ Trinity College Dublin $^{3}$ Deakin University + +# Abstract + +Diffusion models have shown great promise in synthesizing visually appealing images. However, it remains challenging to condition the synthesis at a fine-grained level, for instance, synthesizing image pixels following some generic color pattern. Existing image synthesis methods often produce contents that fall outside the desired pixel conditions. To address this, we introduce a novel color alignment algorithm that confines the generative process in diffusion models within a given color pattern. Specifically, we project diffusion terms, either imagery samples or latent representations, into a conditional color space to align with the input color distribution. This strategy simplifies the prediction in diffusion models within a color manifold while still allowing plausible structures in generated contents, thus enabling the generation of diverse contents that comply with the target color pattern. Experimental results demonstrate our state-of-the-art performance in conditioning and controlling of color pixels, while maintaining on-par generation quality and diversity in comparison with regular diffusion models. + +# 1. Introduction + +Recently, diffusion models [18] and their large-scale text-conditioned variants [36, 39] have demonstrated impressive quality and diversity of generated contents in image synthesis. Subsequent works have explored leveraging this capability for possible controllability, incorporating various constraints, such as structural [30, 35, 48] and reference constraints [14, 46], which typically serve as additional inputs. However, due to the inherent randomness in the diffusion process, it remains challenging to control generated contents at a fine-grained level, for instance, synthesizing image pixels following some given content and generic color distribution. + +In this paper, we focus on such a fine-grained control over pixel colors in synthesized images to meet with three color-conditioned generation criteria: the accuracy in adhering to given color values; the completeness of covering proportions of the specified colors; and the disentanglement of the colors + +![](images/85cf9bc14882995fa8308b682e746006db8a7e587eb076f93a95b3d7515154ad.jpg) + +![](images/c24737b727e83432a2209be3aa2222992e90877ade8b8acab3a286094958b777.jpg) + +![](images/1f3294cc6f0c5c27e2fa59fdf44c611d5dee03ed0715abfe767a1615f6c17edb.jpg) + +![](images/7cf7ee32d69fbe1e881cc15aa9090585d5754542f3b20627dc66d065db4a9752.jpg) + +![](images/78701179e89d71a22f1ec49707e734a0038432957f6775df1d54b8c6de66caff.jpg) + +![](images/06dc785169846cfd8d5c373b8698d4ffc2457e433af540f531d4b57d0f829a1f.jpg) + +![](images/94802e72e417ea78c39f53cace29debadff122c963db708f69e04b111fc4ffc7.jpg) + +![](images/3473587401498fe280f8d76908cf995f24de07b5117a072217a90f739d9a5486.jpg) + +![](images/a85619011eaae6615460963d82b003da2de87ed01e3e0ad0b77509fd40d19ff6.jpg) + +![](images/6d5192fb92440f9f7a9f5f9e8b48e6c3d95d8871ed062fb5ff5a4600ee436db1.jpg) +(a) Condition + +![](images/48e2934d33a138d09618621ae0b10f915b65955eb2195a2dd7e41c99c0523e29.jpg) + +![](images/d852296cb83b6b41f5a1a0d3d22cf26c74df25ac9a4dbaf3c20611be1273d50c.jpg) +(b) Ours + +![](images/cdacc8f52e07ac25f4fa920dcb0b4f4703c4445f3e58bb534556716f5b49ac29.jpg) + +![](images/7a529dcb871fe764f313a2b0a9dcf845c3eebd2ba3d616624e6d2916ae85312f.jpg) + +![](images/6ca033b21327cd259be2f7120b59788f61b4b52142988f95d5a5a38a4b1a770b.jpg) + +![](images/d80b7b18f4d9ccadc468d0e6cf3eeb379f0f195ffc81f02a1728d5ea58c9aec4.jpg) +(c)Baselines +Figure 1. Examples of color-conditioned generation from different methods. Color conditions (a) can be derived from imagery (top two rows) or manual drawing (bottom two rows). Input text prompts are "painting" and "pen" (1st & 3rd rows), and "boat" (2nd & 4th rows). Our method (b) can generate pixels closely aligned with the color conditions and effectively disentangle the colors into different structures. Existing baselines [14, 36, 46, 48] (from top to bottom) (c) struggle with either the semantics (the first two rows) or colors (the last two rows) of generated contents. + +into various structures. Such a color-conditioned generation ability would be a great aid to numerous downstream tasks in creative artwork and graphic design [10]. + +Different attempts of color-conditioned image generation are visualized in Fig. 1, where previous works often produce suboptimal results that do not align with the aforementioned criteria. In this paper, we propose an effective yet control-easy method facilitating color alignment in diffusion models for color-conditioned image synthesis. Specifically, + +our method accepts a color condition via a color pattern, determining color values and their proportions in a synthesized image. Different from the regular diffusion approach, we perform the color alignment on intermediate samples or latent representations during the diffusion process prior to inputting them into a denoising model for image synthesis, thereby facilitating the perception of the conditional colors in diffusion models while preserving the overall attribution of the diffusion process. Furthermore, our method can support various settings: re-training of a color-aligned diffusion model; fine-tuning of a color-aligned diffusion model from a pre-trained regular diffusion model; and zero-shot approximation (training-free) of the color alignment process with a pre-trained diffusion model. We validate our method by demonstrating its capabilities across various color conditions and compare it with existing baselines on benchmark datasets. In summary, we make the following contributions: + +- We address a challenging task: fine-grained color-conditioned image synthesis. We aim to generate images that closely follow a specified color pattern that defines color values and proportions without requiring an explicit structure. Our approach thus enhances the controllability and flexibility of image generation. +- We propose a novel color alignment method that operates on the diffusion process of diffusion models, allowing the diffusion to effectively perceive color conditions while maintaining the quality and diversity of generated contents. We derive various settings including re-training, fine-tuning, and a zero-shot approximation, to adapt our color alignment method into different down-stream needs. +- We conduct extensive experiments to validate our proposed method in various scenarios and controllability settings. Our results demonstrate state-of-the-art performance compared with existing baselines. + +# 2. Related work + +Neural image synthesis has started with Variational Autoencoders (VAEs) [22, 41, 43] and Generative Adversarial Networks (GANs) [12, 19, 20, 26], followed then by transformer-based methods [9, 23] and diffusion models [18, 31, 42]. Among these, Denoising Diffusion Probabilistic Models (DDPM) [18] have attracted considerable attention due to their generation quality and ease of extension. These models have been further developed into text-to-image (T2I) diffusion techniques [35, 36, 39], which have been boosted recently to achieve greater controllability through incorporation of extra conditions [5, 37, 38, 48]. Our method follows the general diffusion framework with a focus on fine-grained color-conditioned image synthesis. + +Reference-based conditional diffusion. An intuitive condition for diffusion models is the use of an image reference. Structural conditions, such as edge maps and semantic maps, + +have been recently studied in [24, 25, 30, 38, 45, 48]. These methods strictly apply spatial constraints given in the reference to align generated pixels. A more relaxed approach is image editing [5, 6, 13, 21, 27], where the structure and appearance in the reference loosely guide the generation of target images. Those works aim to achieve shifts in appearance while permitting a small range of spatial edits, such as changes in the size and pose of objects. Several methods [14, 37, 46] condition only on abstract concepts in the reference, such as image/object style, enabling more creative synthesis. In this work, we disentangle colors in the reference from their original spatial structure while preserving the color values and proportions. Our method allows a variety of user-provided color materials to construct diverse structures. + +Sample altering in diffusion. As indicated in [25], diffusion from the same noisy sample can result in similar outputs. Diffusion inversion [27] is commonly used to construct samples, used directly [27, 33] or mapped to another diffusion process [7, 29, 49] to diverge outputs. Several methods update intermediate latent representations with gradients calculated from external modules, e.g., other networks [47], images [40], features [28]. The Gaussian noise is defined [11] using estimated mean and variance to improve model convergence. Cold Diffusion [4] replaces the noising process with image degradations such as blurring and masking. + +# 3. Proposed method + +We first in Sec. 3.1 review the regular diffusion process. In Sec. 3.2, we introduce our color alignment technique under its simplest setting, i.e., re-training of a diffusion model with color alignment. We then in Sec. 3.3 extend our technique to latent diffusion models, where we address typical challenges in latent space diffusion and demonstrate our method's ability via fine-tuning of a pre-trained latent diffusion model to a color-aligned one. Finally, in Sec. 3.4, we present a zero-shot variant of our method that achieves similar generation quality without extra training. We provide an overview of our pipeline and the regular diffusion in Fig. 2. + +# 3.1. Diffusion preliminaries + +Diffusion models [18] could be seen as a variant of VAEs [22] that map a data distribution $\mathcal{X}(\mathbf{x}_0)$ from a Gaussian distribution $\mathcal{N}(\mathbf{0},\mathbf{I})$ . In the forward process, a sample $\mathbf{x}_t$ is encoded by adding a Gaussian noise to $\mathbf{x}_0$ : + +$$ +\mathbf {x} _ {t} = \sqrt {\bar {\alpha} _ {t}} \mathbf {x} _ {0} + \sqrt {1 - \bar {\alpha} _ {t}} \boldsymbol {\epsilon}, \quad \boldsymbol {\epsilon} \sim \mathcal {N} (\mathbf {0}, \mathbf {I}) \tag {1} +$$ + +where $t \in [1, T]$ is the time step associated with a noise strength $\bar{\alpha}_t$ where $(1 - \bar{\alpha}_1) \approx 0$ and $(1 - \bar{\alpha}_T) \approx 1$ . + +Then, a denoising model $\epsilon_{\theta}$ with parameters $\theta$ learns to + +reverse the process in Eq. 1 by training with a loss $\mathcal{L}$ as, + +$$ +\mathcal {L} = \nabla_ {\theta} \| \epsilon - \epsilon_ {\theta} (\mathbf {x} _ {t}, t) \| _ {2} ^ {2} \tag {2} +$$ + +$$ +\hat {\mathbf {x}} _ {0} = \frac {1}{\sqrt {\bar {\alpha} _ {t}}} \left(\mathbf {x} _ {t} - \sqrt {1 - \bar {\alpha} _ {t}} \boldsymbol {\epsilon} _ {\theta} \left(\mathbf {x} _ {t}, t\right)\right) \tag {3} +$$ + +$$ +\mathbf {x} _ {t - 1} = \frac {\sqrt {\bar {\alpha} _ {t - 1}} \beta_ {t}}{1 - \bar {\alpha} _ {t}} \hat {\mathbf {x}} _ {0} + \frac {\sqrt {\alpha_ {t}} (1 - \bar {\alpha} _ {t - 1})}{1 - \bar {\alpha} _ {t}} \mathbf {x} _ {t} + \sigma_ {t} \boldsymbol {\varepsilon} \quad (4) +$$ + +where $\hat{\mathbf{x}}_0$ is the predicted sample. $\alpha ,\beta ,\sigma \in \mathbb{R}$ and $\varepsilon \sim \mathcal{N}(\mathbf{0},\mathbf{I})$ are scheduler-related parameters [18]. + +# 3.2. Color-aligned diffusion in image space + +As presented in Sec. 3.1, the regular diffusion model learns to reconstruct $\mathbf{x}_0$ from $\mathcal{N}(\mathbf{0},\mathbf{I})$ during training and allows sampling-free generation in testing, i.e., unconditional generation. To condition the diffusion process with color conditions, existing methods pass color text prompts [36, 46] or referenced images [14, 46, 48] in parallel with $\mathbf{x}_t$ untouched. We found this approach loosely constrains the sampled data manifold, i.e., $\hat{\mathbf{x}}_0$ can be sampled with out-of-bound colors. This is because the conditions are not involved in the sampling process (in construction of $\mathbf{x}_t$ ) and the denoising model $\epsilon_{\theta}$ hence leans towards the unconditional generation. Moreover, colors are learned in conjunction with spatial information given in referenced images, limiting the diversity of generated contents (see Fig. 1). + +In this paper, we take a color pattern $\mathbf{c}$ , determining color values and their proportions in synthesis, as a color condition. The color pattern can be an imagery or hand-drawing (see Fig. 1). To effectively constrain the sampling manifold with color condition, we involve the color pattern $\mathbf{c}$ into the diffusion process via a new denoising model $\epsilon_{\theta}(\mathbf{x}_t,t,\mathbf{c})$ . In addition, we relax the spatial information in $\mathbf{c}$ by using a color alignment function $f(\mathbf{x}_t,\mathbf{c})$ that produces a color-aligned image from an image $\mathbf{x}_t$ and color pattern $\mathbf{c}$ as, + +$$ +f \left(\mathbf {x} _ {t}, \mathbf {c}\right) [ p ] = \underset {\mathbf {c} [ q ]} {\arg \min } \| \mathbf {x} _ {t} [ p ] - \mathbf {c} [ q ] \| _ {2} ^ {2} \tag {5} +$$ + +where $\mathbf{x}_t[p]$ and $\mathbf{c}[q]$ are colors at pixels $p$ and $q$ . + +We visualize the diffusion process in Fig. 3. Our color alignment technique benefits the sampling in several aspects. First, it enforces the generated pixel colors within the target space, guaranteeing the accuracy. Second, its non-spatial manner allows pixels distribute more freely, fulfilling the color disentanglement. For the sake of completeness, we further concatenate $\mathbf{c}$ to the model's input as an extra hint. However, to completely break down the spatial information, we opt to randomly permute the pixel locations in $\mathbf{c}$ , resulting in $\psi(\mathbf{c})$ with $\psi$ being an image permutation operator. Such a permutation allows us to directly take $\mathbf{x}_0$ as color condition in training time, avoiding extra data requirement and data leakage. Finally, training and sampling of our color + +![](images/f740dcf1b938da0a080229c45ea47377d6e35fcf57f37b87a3f145339caa1c82.jpg) +(a) Regular diffusion [36] + +![](images/6cda9bde6cc68309455310737b2727d8902357514ecd19845311bbf3b14c9c11.jpg) +(b) Regular diffusion with condition + +![](images/db639651312e35320c3d57121c13e2a6fe3fe026c512b2f1a6bd5647e94ffae7.jpg) +(c) Our color-aligned diffusion for retraining or fine-tuning +Figure 2. Pipelines of the regular diffusion (a) and its color-conditioned version (b), our color-aligned diffusion (c), where noisy samples $\mathbf{x}_t$ are aligned with conditioned colors, and our zero-shot version (d), where color alignment is directly applied to predicted sample $\hat{\mathbf{x}}_0$ . + +![](images/a270035bdc2014badd0926ff6f48f34a6a54ce077f59c3352ac33d258c09b61e.jpg) +(d) Our zero-shot approximation of the color-aligned diffusion + +aligned diffusion process (Fig. 2c) are formulated as: + +$$ +\mathcal {L} = \nabla_ {\boldsymbol {\theta}} \| \boldsymbol {\epsilon} ^ {\prime} - \boldsymbol {\epsilon} _ {\boldsymbol {\theta}} \big (f (\mathbf {x} _ {t}, \mathbf {c}), t, \mathbf {c} \big) \| _ {2} ^ {2} +$$ + +$$ +\boldsymbol {\epsilon} ^ {\prime} = \frac {f (\mathbf {x} _ {t} , \mathbf {c}) - \sqrt {\bar {\alpha} _ {t}} \mathbf {x} _ {0}}{\sqrt {1 - \bar {\alpha} _ {t}}}, \quad \mathbf {c} = \psi (\mathbf {x} _ {0}) \tag {6} +$$ + +$$ +\hat {\mathbf {x}} _ {0} = \frac {1}{\sqrt {\bar {\alpha} _ {t}}} \left(f \left(\mathbf {x} _ {t}, \mathbf {c}\right) - \sqrt {1 - \bar {\alpha} _ {t}} \epsilon_ {\theta} \left(f \left(\mathbf {x} _ {t}, \mathbf {c}\right), t, \mathbf {c}\right)\right) \tag {7} +$$ + +where $\epsilon^{\prime}$ is the adapted noise complemented with the effect of $f$ . Notably, we alter only what $\epsilon_{\theta}$ sees and predicts. The parameters $\theta$ thus can be learned more easily with color condition compared with the regular one. All other nonlearnable steps, e.g., the regular forward process and Eq. 4, remain unchanged for the ease of sampling. For more technical visualizations, please refer to the supplementary material. + +# 3.3. Color-aligned diffusion in latent space + +Training of diffusion models in their latent space is a common practice. In this section, we extend our color alignment + +technique to latent diffusion models via fine-tuning of a pre-trained latent model to achieve color alignment. + +Following the Stable Diffusion in [36], we encode $\mathbf{x}_0$ into its latent representation $\mathbf{z}_0$ . We then train our model with alignment in the latent space, i.e., replacing $\mathbf{x}$ by $\mathbf{z}$ in Eqs. 6 and 7. However, when high-frequency structural information is present in $\mathbf{x}_0$ , unintended colors and local structural information can be encoded in $\mathbf{z}_0$ . We discovered that blurring $\mathbf{x}_0$ prior to encoding it enhances the color disentanglement and accuracy, with minimal loss of the completeness in generated results. Also, we found that suspending color alignment at late time steps (e.g., $t < 0.2T$ where diffusion model is fixing details) allows a proper refinement of the latent, making latent decoding towards more photo-realistic details. + +# 3.4. Zero-shot color-aligned diffusion + +To mitigate the computational burden of training, we introduce a zero-shot approximation of our color alignment method. As shown in Eqs. 6 and 7, the training aims to reconstruct $\hat{\mathbf{x}}_0$ whose color distribution follows the color condition $\mathbf{c} = \psi (\mathbf{x}_0)$ . Under the zero-shot setting, to achieve this ability without learning the parameters $\theta$ , we directly map an unconditionally generated $\hat{\mathbf{x}}_0$ (from Eq. 3) to $\mathbf{c}$ during the sampling process using a color alignment function $g(\hat{\mathbf{x}}_0,\mathbf{c})$ : + +$$ +g \left(\hat {\mathbf {x}} _ {0}, \mathbf {c}\right) = \underset {\psi (\mathbf {c})} {\arg \min } \| \hat {\mathbf {x}} _ {0} - \psi (\mathbf {c}) \| _ {2} ^ {2} \tag {8} +$$ + +where we map the unconditional $\hat{\mathbf{x}}_0$ to a closest reorganization of the conditional colors $\psi (\mathbf{c})$ for color alignment. + +Unlike $f$ defined in Eq. 5, $g$ in Eq. 8 performs one-to-one mapping, i.e., each location in $\mathbf{c}$ is used for color alignment exactly once. Note that, like $f$ , our $g$ also applies non-spatial mapping. This would ensure desired properties (color accuracy, completeness, and disentanglement) from the zero-shot color-aligned diffusion. However, since $g$ is not involved in the learning process, the color-aligned image $g(\hat{\mathbf{x}}_0,\mathbf{c})$ could be messy, showing invalid structures and outliers (see Fig. 2d). Fortunately, we observed that pre-trained diffusion models have capability of clustering semantics and eliminating outliers at early sampling steps (e.g., $t > 0.2T$ ). We also suspend the color alignment at late time steps to prevent excessive color changes and allow the model to refine details. Additionally, such zero-shot approximation is extendable to latent diffusion models. We found the same conclusion for using the blurring strategy as in Sec. 3.3 applies. + +Another interest could be replacing $f$ by $g$ in learning. However, we find this not applicable because $g(\mathbf{x}_t,\mathbf{c})$ corrupts the distribution of $\mathbf{x}_t$ too heavily. Its subsequent adapted noise $\epsilon^{\prime}$ (the learning target in Eq. 6) deviates too far from the standard Gaussian $\mathcal{N}(\mathbf{0},\mathbf{I})$ , eliminating the effect of a proper diffusion process. + +# 4. Experiments + +# 4.1. Datasets + +We experimented our method in two scenarios: color-aligned diffusion in image space and color-aligned diffusion in latent space. For diffusion in image space, we ran experiments on Oxford-flower [32] (7.17k training and 1.02k test images) and Microsoft-emoji [8] datasets (6.80k training and 0.76k test images) at resolution $64 \times 64$ . For latent diffusion, we experimented on a sub-set of a high-quality generic dataset Text-to-image-2M [1] (300k training and 20k test images) at resolution $512 \times 512$ . To validate the color disentanglement ability of our method, we replaced original prompts in Text-to-image-2M [1] with 50 daily object prompts randomly generated by ChatGPT [2] (e.g., those in Fig. 4). + +# 4.2. Baselines + +Our color-aligned image diffusion version is adapted from the DDPM [18]. Therefore, we compare our method with the DDPM to demonstrate its extra color-conditioned ability without requiring additional data creation or labeling. For the same reason, we also compare our latent diffusion version with the pre-trained Stable Diffusion in [36]. + +Given that our color reference can be an in-the-wild image or a manual drawing, we compare our method with SOTA reference-based generation methods. IP-Adapter [46] and T2I-Adapter [30] convert an image reference into features, which are subsequently combined with the model features for image generation. Style-Aligned [14] transfers source image information from its attention layer to the target generation attention layer, thereby preserving spatial and style features. ControlNet-Color is a variant of ControlNet [48] that accepts a predefined pixelized color image as input and generates a refined version in which the colors are approximately aligned with the original image spatially. IP-Adapter [46] and ControlNet-Color [48] require fine-tuning, while Style-Aligned [14] operates as a zero-shot approach. + +# 4.3. Implementation details + +We follow default settings in the DDPM [18] and the Stable Diffusion [36] for network architectures and diffusion scheduling, specifically, from the huggingface-diffusers [44]. Further details, e.g., DDPM re-training, Stable Diffusion fine-tuning are presented in our supplementary material. + +We trained our image diffusion version for 100 epochs and latent diffusion version for $160\mathrm{k}$ iterations. We performed inference at 50 time steps. We used classifier-free guidance [17] with guidance scale 5. All experiments were executed on two RTX 3090 GPUs with batch size 4. + +# 4.4. Qualitative study + +Color-aligned image diffusion. We visually compare our method with the regular diffusion in DDPM [18] in Fig. 3. + +![](images/f7608774d3e9bb2970cd84dd5ea576b611c4e671295186736580e59440c6ecba.jpg) +(a) Our color-aligned image diffusion process. The top row shows our training inputs, created from early to late steps in the forward process. The bottom row is a test case from our reverse process, sampled from pure noise. + +![](images/17c9cf1f6efbf645cb14f5c2a30f4a4c8575dfca3ed7a9ed2ed01ccce09ac9c2.jpg) +(b) Regular image diffusion process in DDPM [18] in same order as (a). + +![](images/8f64d93754bd2496078cd957abd57ba6779b959093f5fc401ddec9697730a72e.jpg) +(c) Several color-aligned image synthesis results. From top to bottom rows are condition, our generated results, and regular diffusion's results. +Figure 3. Qualitative results of color-aligned image diffusion. (a)-(b) Visualization of the diffusion process by our method and DDPM [18]. (c) Generation results of our method and DDPM [18]. + +As shown, our method (Fig. 3a) well restricts the diffusion process to the conditional color space, wherein only whites, yellows, and background blacks are allowed. In contrast, the regular diffusion in DDPM (Fig. 3b) generates out-of-bounds colors during the diffusion process, leading to an unintended reconstruction result. + +Fig. 3c illustrates an immediate application of our color-aligned image diffusion in image rephrasing, i.e., colors in an input image are reorganized by our model to construct new structures. More importantly, this conditional process does not require any additional data (e.g., labeled data or paired data) or pre-defined spatial information (e.g., semantic maps or masks) in both training or inference. This makes our technique scalable to larger datasets. + +Color-aligned latent diffusion. We present our coloraligned latent diffusion results in comparison with existing baselines in Fig. 4. As demonstrated, our method (Figs. 4b and 4c) effectively conditions the synthesis on target color inputs. Color values and proportions are aligned, i.e., color + +accuracy and completeness are maintained. Additionally, our method generates contents that align well with target text prompts, showcasing the ability for color disentangle-ment where colors are rearranged into new structures. Furthermore, with these color-conditioned generation criteria preserved, our method achieves image quality and diversity on par with the original diffusion model. + +In contrast, the baselines produce suboptimal results in various aspects. IP-Adapter [46] (Fig. 4d), T2I-Adapter [30] (Fig. 4e), and Style-Aligned [14] (Fig. 4f) construct target contents on top of an input image reference, thus often struggle to disentangle colors. Since the spatial information of colors is taken into account, these methods likely to fail completely if input reference and target text prompt belong to significantly different domains. ControlNet-Color [48] (Fig. 4g) treats input reference as a weak guidance, leading to failures to accurately preserve color values and proportions. + +Generalizability and controllability. Our method also works with manual drawings, for example, a rough color pattern. As shown in Fig. 5, our method (the fine-tuned version) can generate images aligned with the color values and proportions specified in input color patterns. Moreover, even if only a few color values presented, our method can adjust these values to achieve appropriate smoothness, brightness, and shadow in generated contents. + +These manual color conditions can be easily modified to influence the generation process. For example, the amount of each color value can be scaled (Fig. 5a). The color values can be tuned intuitively without the need for careful prompt engineering (Fig. 5b). Additional colors can be incorporated into conditions along with user ideation (Fig. 5c). By controlling target text prompt, we can specify the placement of colors (Fig. 5d). Note that, like the regular diffusion, our method can generate diverse results from the same input condition by sampling different noises. + +Zero-shot diffusion. We showcase the ability of our method in zero-shot diffusion in Fig. 6. Our experiments indicate that, compared with our fine-tuned version, the zero-shot version tends to generate images with less fine-grained details, resulting in issues such as blurry lighting, uniform texture, and shadow missing. Since our zero-shot approach mechanically maps colors without model training, it performs more comparably to the fine-tuned version when applied to well-sampled color conditions, e.g., in-the-wild images in the fifth row in Fig. 6 and those in Fig. 4. + +# 4.5.Quantitative study + +Metrics. We evaluate the quality of generated images using the widely adopted FID score [16]. To assess the color disentanglement, we use the CLIP-Score [15] measuring the similarity between a generated image and a target text prompt within the CLIP [34] embedding space. Since the color accuracy and completeness can be evaluated independ- + +![](images/a7d6e88669e9438c366231e09cde4d8b0d81066ff8fa2d175dce681ab1e80c8f.jpg) +(g) ControlNet-Color [48] +Figure 4. Qualitative results of color-aligned latent diffusion. Each input (first row) includes an in-the-wild image as color condition and a target text prompt. Each column presents results of experimented methods using the same input. + +dently of spatial relations, we measure the proximity between the color distributions of a generated image and input color condition. We employ the Chamfer Distance (CD) [3], a commonly used metric for measuring the distance between + +two point sets, to quantify the distance between generated RGB points $\hat{\mathbf{x}}_0$ and input condition RGB points $\mathbf{c}$ : + +![](images/adaad4824b3da68008f02200840b931fd9d7feccb871eb2e3ac0c73114301c19.jpg) +Figure 5. Qualitative results of sampling and editing of input color conditions. All results are generated by our fine-tuned color-aligned latent model. Red arrows represent the generation process. Blue arrows represent the editing of color conditions. + +
MethodFID ↓CD-A ↓CD-C ↓Re-training Cost ↓Inference Cost ↓
Image Diffusion (Re-train) [18]63.9; 14940.6; 42.442.6; 39.01.001.00
Ours (Image; Re-train)57.5; 45.70.00; 0.004.12; 3.651.061.13
Ours (Image; Zero-shot)50.0; 84.55.59; 5.468.28; 10.40.006.76
+ +Table 1. Quantitative results of image diffusion. In each metric column, the left and right numbers correspond to the scores from the Oxford-flower dataset [32] and Microsoft-emoji [8] dataset, respectively. + +$$ +\operatorname {C D - A} \left(\hat {\mathbf {x}} _ {0}, \mathbf {c}\right) = \frac {\sum_ {\hat {\mathbf {x}} _ {0} [ p ]} \min _ {\mathbf {c} [ q ]} \| \hat {\mathbf {x}} _ {0} [ p ] - \mathbf {c} [ q ] \| _ {2} ^ {2}}{| \hat {\mathbf {x}} _ {0} |} \tag {9} +$$ + +$$ +\operatorname {C D - C} \left(\hat {\mathbf {x}} _ {0}, \mathbf {c}\right) = \frac {\sum_ {\mathbf {c} [ q ]} \min _ {\hat {\mathbf {x}} _ {0} [ p ]} \| \hat {\mathbf {x}} _ {0} [ p ] - \mathbf {c} [ q ] \| _ {2} ^ {2}}{| \mathbf {c} |} \tag {10} +$$ + +where $\mathrm{CD} = \mathrm{CD - A} + \mathrm{CD - C}$ consists of an accuracy term and a completeness term that measure the set distance from one to another; $|\cdot |$ denotes set cardinality. Lastly, we evaluate + +the inference cost of different methods, using the regular diffusion as the standard reference. + +Baseline comparison. We first present the performance scores for the image diffusion case study in Tab. 1. The scores clearly indicate that our technique effectively preserves conditioned-colors. Note that, for inference, the color alignment is performed on RGB pixels until the final diffusion step, making the CD-A approach zero. We found that our fine-tuned version requires nearly the same training and inference costs as the regular diffusion model. + +
MethodFID ↓CLIPScore ↑CD-A ↓CD-C ↓Inference Cost ↓
Stable Diffusion (Pre-train) [36]72.6; 72.627.3; 27.3166; 27.865.3; 16.01.00
Ours (Latent; Fine-tune)86.7; 69.429.0; 27.473.9; 4.9815.8; 4.871.03
Ours (Latent; Zero-shot)104; 77.928.3; 27.435.2; 2.683.19; 3.933.52
Baseline (IP-Adapter [46])202; 50.925.5; 22.0145; 10.1106; 13.91.44
Baseline (T2I-Adapter [30])134; 77.124.3; 24.9143; 8.3424.7; 7.011.56
Baseline (Style-Aligned [14])177; 73.123.5; 22.945.1; 6.1832.5; 8.505.69
Baseline (ControlNet-Color [48])105; 81.022.9; 23.4129; 13.149.8; 6.541.78
+ +![](images/a6166db6e7cd9f137e37587475e03cadabb97ff0a257cafbdb37166b024cbe0d.jpg) +"A professional photo of a bar." + +![](images/5f01b718d4c505a1d9183a0b728ef45d6bdbf7bdd393d14148fd6ec9378f7673.jpg) +of a bar: + +![](images/6dc44c9b58ecc010a2291884d57a229b439b69ac24a25f3fe8be93bd8dabb06d.jpg) + +![](images/39f889f61f5d4b8f7638cc69f76b58acc3258807476ddd9b76ee01f6715697cf.jpg) +"Gray modern house, photograph, highly detailed. + +![](images/f36b925e4ea021276383c4da6f0c1395f7d57a694993589b763ec5939c6fa83f.jpg) + +photograph, highly detailed. + +![](images/7a47f65f1c0dc17a9c09d117ddaa62b337fa887e3a97d5ab43cdb3c86c3ed3d2.jpg) +tailed. + +![](images/59069fe124565ac81a834c0ac606fba67550fa011f685fd79892cef13ea44269.jpg) + +![](images/76562308239e8646d2b5d78dcc8e42dda1bb757afbe4de247708d78c75d466f1.jpg) +"An animated owl looking at the camera. + +![](images/d390afd7ad3750eddf15560afd5488e38a1a6d29f80b26844ccc03ba7895f4d2.jpg) +king at the camera. + +![](images/274de5eaa47b6b333bad1c8e8f0a6e870568274c42a499a86e96813351a26fe2.jpg) +$\because {AD} = {AQ},\angle {PAQ} = {60}^{ \circ }\therefore {\Delta PQA}$ 为等边编形 + +![](images/cc65c13d277e9ea68905a1827188e40a0b1a26c22d56782f69ae34435bd7199c.jpg) + +![](images/1d995c919204a04b802081d68b7fe699b7fa05a56b5944bc9866fb6c096f8bbd.jpg) +"An artistic flat design poster featuring sushi on a plate. + +![](images/9f4415aee014082399b309a3357adcbaae9b40c90e3c1a50af89c91599bcdeb9.jpg) + +n poster featuring su + +![](images/41ea7c2a5cdce56be20445b35bcd686de828e4963016efce7e7d389722e8354a.jpg) +ni on a plate. + +![](images/ea342a221aa9de8de74a6125ad4a9224a380bcba81a71f68ae23013e4aa89621.jpg) + +(c) Fine-tune + +![](images/b9d11ffd8ec5106340840ed8a63fbac74871e7f9beb57d89a12b4bf32e64a5da.jpg) +"Thundery sky." +(a) Condition + +![](images/d872e269915874e5e2043056acbdab71d685bf440b2834f2bdf306c680f3485e.jpg) +(b) Zero-shot +Figure 6. Qualitative results of our zero-shot approach. The first four rows show results from manual color conditions, while the fifth row presents results from in-the-wild imagery condition. + +![](images/e0938ebcd9865118665b929cc54306f6a28fea8f5ceea9ea8b3d330d0d69a847.jpg) + +o-shot + +(c) Fine-tune + +We report the performance scores for the latent diffusion case study in Tab. 2. As shown, our method generally outperforms the baselines in most of the metrics. The baselines exhibit significant failures in metrics reflecting the disentanglement (CLIP-score), accuracy and completeness (CD-A + +Table 2. Quantitative results of latent diffusion. In each metric column, the left and right numbers correspond to the scores from the settings of manually sampled color condition and in-the-wild image color condition, respectively. + +
MappingConditionFID ↓CD-A ↓CD-C ↓
XX63.9; 14940.6; 42.442.6; 39.0
X✓57.9; 88.115.4; 10.618.4; 15.2
✓X59.8; 45.70.00; 0.005.10; 4.62
✓✓57.5; 45.60.00; 0.004.12; 3.65
+ +Table 3. Ablation study on color mapping and color condition. For each metric, the left and right numbers are from the Oxford-flower [32] and Microsoft-emoji [8] datasets, respectively. + +and CD-C scores) of generated colors in synthesized images. When provided with manual color conditions, our zero-shot version achieves worse FID score than the fine-tuned version does. This aligns with our observations in Sec. 4.4 indicating imagery details lacking caused by the zero-shot version. The zero-shot setting's very low CD scores also indicate color overfitting caused by mechanically mapping the manual color conditions. + +Ablation study. We validate the effectiveness of important components in our method in Tab. 3. Specifically, we investigate the role of the color mappings $f$ , and the inputting of the color condition $c$ (those in Fig. 2c). When mappings are not used, noisy samples are not altered and only color conditions are fed to the diffusion model. As shown in Tab. 3, the mappings and color conditions significantly boost up the performance of our method. + +# 5. Conclusion and Discussion + +We propose a color alignment method that operates on the diffusion process of diffusion models. Diffused colors are mapped to conditional colors across diffusion steps, enabling a continuous pathway for synthesizing target color patterns while maintaining the creativity of diffusion models. + +There are issues that would be addressed in our future work. First, while our method can condition on color values and proportions, it is worthwhile to understand the influence of the spatial information in the color conditions. Second, it is of great interest to extend our color alignment to video and 3D data, potentially with integration of multi-view attributes. + +# Acknowledgments + +Binh-Son Hua is supported by Research Ireland under the Research Ireland Frontiers for the Future Programme - Project, award number 22/FFP-P/11522. + +# References + +[1] Text-to-image-2m: A high-quality, diverse text-to-image training dataset. https://huggingface.co/datasets/jackyhate/text-to-image-2M.4 +[2] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 4 +[3] Vassilis Athitsos and Stan Sclaroff. Estimating 3d hand pose from a cluttered image. In 2003 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, 2003. Proceedings., pages II-432. IEEE, 2003. 6 +[4] Arpit Bansal, Eitan Borgnia, Hong-Min Chu, Jie Li, Hamid Kazemi, Furong Huang, Micah Goldblum, Jonas Geiping, and Tom Goldstein. Cold diffusion: Inverting arbitrary image transforms without noise. Advances in Neural Information Processing Systems, 36, 2024. 2 +[5] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18392-18402, 2023. 2 +[6] Mingdeng Cao, Xintao Wang, Zhongang Qi, Ying Shan, Xiaohu Qie, and Yinqiang Zheng. Masactrl: Tuning-free mutual self-attention control for consistent image synthesis and editing. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22560-22570, 2023. 2 +[7] Jiwoo Chung, Sangeek Hyun, and Jae-Pil Heo. Style injection in diffusion: A training-free approach for adapting large-scale diffusion models for style transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8795-8805, 2024. 2 +[8] Jason Custer and Spencer Nelson. Microsoft fluent emoji. https://github.com/microsoft/fluentui-emoji.4,7,8 +[9] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2 +[10] Ziv Epstein, Aaron Hertzmann, Laura Herman, Robert Mahari, Morgan R. Frank, Matthew Groh, Hope Schroeder, Amy Smith, Memo Akten, Jessica Fjeld, Hany Farid, Neil Leach, Alex "Sandy" Pentland, and Olga Russakovsk. Art and the science of generative ai. Science, 380(6650):1110-1111, 2023. 1 +[11] Martin Nicolas Everaert, Marco Bocchio, Sami Arpa, Sabine Susstrunk, and Radhakrishna Achanta. Diffusion in style. In + +Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2251-2261, 2023. 2 +[12] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139-144, 2020. 2 +[13] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 2 +[14] Amir Hertz, Andrey Voynov, Shlomi Fruchter, and Daniel Cohen-Or. Style aligned image generation via shared attention. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4775–4785, 2024. 1, 2, 3, 4, 5, 6, 8 +[15] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning. arXiv preprint arXiv:2104.08718, 2021. 5 +[16] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 5 +[17] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 4 +[18] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1, 2, 3, 4, 5, 7 +[19] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1125-1134, 2017. 2 +[20] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 2 +[21] Bahjat Kawar, Shiran Zada, Oran Lang, Omer Tov, Huiwen Chang, Tali Dekel, Inbar Mosseri, and Michal Irani. Imag: Text-based real image editing with diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6007-6017, 2023. 2 +[22] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 2 +[23] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 2 +[24] Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11461-11471, 2022. 2 +[25] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jiajun Wu, Jun-Yan Zhu, and Stefano Ermon. Sdedit: Guided image + +synthesis and editing with stochastic differential equations. arXiv preprint arXiv:2108.01073, 2021. 2 +[26] Mehdi Mirza and Simon Osindero. Conditional generative adversarial nets. arXiv preprint arXiv:1411.1784, 2014. 2 +[27] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6038–6047, 2023. 2 +[28] Chong Mou, Xintao Wang, Jiechong Song, Ying Shan, and Jian Zhang. *Dragondiffusion: Enabling drag-style manipulation on diffusion models.* arXiv preprint arXiv:2307.02421, 2023. 2 +[29] Chong Mou, Xintao Wang, Jiechong Song, Ying Shan, and Jian Zhang. Diffeditor: Boosting accuracy and flexibility on diffusion-based image editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8488-8497, 2024. 2 +[30] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, and Ying Shan. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 4296–4304, 2024. 1, 2, 4, 5, 6, 8 +[31] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International conference on machine learning, pages 8162-8171. PMLR, 2021. 2 +[32] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In 2008 Sixth Indian conference on computer vision, graphics & image processing, pages 722-729. IEEE, 2008. 4, 7, 8 +[33] Gaurav Parmar, Krishna Kumar Singh, Richard Zhang, Yijun Li, Jingwan Lu, and Jun-Yan Zhu. Zero-shot image-to-image translation. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-11, 2023. 2 +[34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 5 +[35] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1(2): 3, 2022. 1, 2 +[36] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 2, 3, 4, 8 +[37] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22500-22510, 2023. 2 + +[38] Chitwan Sahara, William Chan, Huiwen Chang, Chris Lee, Jonathan Ho, Tim Salimans, David Fleet, and Mohammad Norouzi. Palette: Image-to-image diffusion models. In ACM SIGGRAPH 2022 conference proceedings, pages 1-10, 2022. 2 +[39] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural information processing systems, 35:36479-36494, 2022. 1, 2 +[40] Yujun Shi, Chuhui Xue, Jun Hao Liew, Jiachun Pan, Hanshu Yan, Wenqing Zhang, Vincent YF Tan, and Song Bai. Dragdiffusion: Harnessing diffusion models for interactive point-based image editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8839-8849, 2024. 2 +[41] Kihyuk Sohn, Honglak Lee, and Xinchen Yan. Learning structured output representation using deep conditional generative models. Advances in neural information processing systems, 28, 2015. 2 +[42] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020.2 +[43] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017. 2 +[44] Patrick von Platen, Suraj Patil, Anton Lozhkov, Pedro Cuenca, Nathan Lambert, Kashif Rasul, Mishig Davaadorj, Dhruv Nair, Sayak Paul, Steven Liu, William Berman, Yiyi Xu, and Thomas Wolf. Diffusers: State-of-the-art diffusion models. 4 +[45] Shaoan Xie, Zhifei Zhang, Zhe Lin, Tobias Hinz, and Kun Zhang. Smartbrush: Text and shape guided object inpainting with diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22428-22437, 2023. 2 +[46] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023. 1, 2, 3, 4, 5, 6, 8 +[47] Jiwen Yu, Yinhuai Wang, Chen Zhao, Bernard Ghanem, and Jian Zhang. Freedom: Training-free energy-guided conditional diffusion model. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 23174-23184, 2023. 2 +[48] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 1, 2, 3, 4, 5, 6, 8 +[49] Yuxin Zhang, Nisha Huang, Fan Tang, Haibin Huang, Chongyang Ma, Weiming Dong, and Changsheng Xu. Inversion-based style transfer with diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10146-10156, 2023. 2 \ No newline at end of file diff --git a/2025/Color Alignment in Diffusion/images.zip b/2025/Color Alignment in Diffusion/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..94e55232083d4e60ceb0a87b7d72e3fef8644650 --- /dev/null +++ b/2025/Color Alignment in Diffusion/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:37e42f3405854c2a03136b0f83c882b397022a1571e5fecb0f0e7631fd7650f4 +size 1136672 diff --git a/2025/Color Alignment in Diffusion/layout.json b/2025/Color Alignment in Diffusion/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..2c0b9dd03cfaec6e0603388654a168919adb5499 --- /dev/null +++ b/2025/Color Alignment in Diffusion/layout.json @@ -0,0 +1,10352 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 214, + 103, + 396, + 120 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 214, + 103, + 396, + 120 + ], + "spans": [ + { + "bbox": [ + 214, + 103, + 396, + 120 + ], + "type": "text", + "content": "Color Alignment in Diffusion" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 122, + 152, + 487, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 122, + 152, + 487, + 168 + ], + "spans": [ + { + "bbox": [ + 122, + 152, + 487, + 168 + ], + "type": "text", + "content": "Ka Chun Shum1 Binh-Son Hua2 Duc Thanh Nguyen3 Sai-Kit Yeung1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 72, + 176, + 537, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 176, + 537, + 192 + ], + "spans": [ + { + "bbox": [ + 72, + 176, + 537, + 192 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 72, + 176, + 537, + 192 + ], + "type": "text", + "content": "Hong Kong University of Science and Technology " + }, + { + "bbox": [ + 72, + 176, + 537, + 192 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 72, + 176, + 537, + 192 + ], + "type": "text", + "content": "Trinity College Dublin " + }, + { + "bbox": [ + 72, + 176, + 537, + 192 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 72, + 176, + 537, + 192 + ], + "type": "text", + "content": "Deakin University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "spans": [ + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 252, + 297, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 252, + 297, + 479 + ], + "spans": [ + { + "bbox": [ + 55, + 252, + 297, + 479 + ], + "type": "text", + "content": "Diffusion models have shown great promise in synthesizing visually appealing images. However, it remains challenging to condition the synthesis at a fine-grained level, for instance, synthesizing image pixels following some generic color pattern. Existing image synthesis methods often produce contents that fall outside the desired pixel conditions. To address this, we introduce a novel color alignment algorithm that confines the generative process in diffusion models within a given color pattern. Specifically, we project diffusion terms, either imagery samples or latent representations, into a conditional color space to align with the input color distribution. This strategy simplifies the prediction in diffusion models within a color manifold while still allowing plausible structures in generated contents, thus enabling the generation of diverse contents that comply with the target color pattern. Experimental results demonstrate our state-of-the-art performance in conditioning and controlling of color pixels, while maintaining on-par generation quality and diversity in comparison with regular diffusion models." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 502, + 135, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 502, + 135, + 514 + ], + "spans": [ + { + "bbox": [ + 56, + 502, + 135, + 514 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 521, + 296, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 296, + 653 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 296, + 653 + ], + "type": "text", + "content": "Recently, diffusion models [18] and their large-scale text-conditioned variants [36, 39] have demonstrated impressive quality and diversity of generated contents in image synthesis. Subsequent works have explored leveraging this capability for possible controllability, incorporating various constraints, such as structural [30, 35, 48] and reference constraints [14, 46], which typically serve as additional inputs. However, due to the inherent randomness in the diffusion process, it remains challenging to control generated contents at a fine-grained level, for instance, synthesizing image pixels following some given content and generic color distribution." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "content": "In this paper, we focus on such a fine-grained control over pixel colors in synthesized images to meet with three color-conditioned generation criteria: the accuracy in adhering to given color values; the completeness of covering proportions of the specified colors; and the disentanglement of the colors" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 315, + 226, + 375, + 285 + ], + "blocks": [ + { + "bbox": [ + 315, + 226, + 375, + 285 + ], + "lines": [ + { + "bbox": [ + 315, + 226, + 375, + 285 + ], + "spans": [ + { + "bbox": [ + 315, + 226, + 375, + 285 + ], + "type": "image", + "image_path": "85cf9bc14882995fa8308b682e746006db8a7e587eb076f93a95b3d7515154ad.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 316, + 285, + 373, + 342 + ], + "blocks": [ + { + "bbox": [ + 316, + 285, + 373, + 342 + ], + "lines": [ + { + "bbox": [ + 316, + 285, + 373, + 342 + ], + "spans": [ + { + "bbox": [ + 316, + 285, + 373, + 342 + ], + "type": "image", + "image_path": "c24737b727e83432a2209be3aa2222992e90877ade8b8acab3a286094958b777.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 376, + 226, + 434, + 284 + ], + "blocks": [ + { + "bbox": [ + 376, + 226, + 434, + 284 + ], + "lines": [ + { + "bbox": [ + 376, + 226, + 434, + 284 + ], + "spans": [ + { + "bbox": [ + 376, + 226, + 434, + 284 + ], + "type": "image", + "image_path": "1f3294cc6f0c5c27e2fa59fdf44c611d5dee03ed0715abfe767a1615f6c17edb.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 376, + 285, + 433, + 342 + ], + "blocks": [ + { + "bbox": [ + 376, + 285, + 433, + 342 + ], + "lines": [ + { + "bbox": [ + 376, + 285, + 433, + 342 + ], + "spans": [ + { + "bbox": [ + 376, + 285, + 433, + 342 + ], + "type": "image", + "image_path": "7cf7ee32d69fbe1e881cc15aa9090585d5754542f3b20627dc66d065db4a9752.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 435, + 226, + 492, + 284 + ], + "blocks": [ + { + "bbox": [ + 435, + 226, + 492, + 284 + ], + "lines": [ + { + "bbox": [ + 435, + 226, + 492, + 284 + ], + "spans": [ + { + "bbox": [ + 435, + 226, + 492, + 284 + ], + "type": "image", + "image_path": "78701179e89d71a22f1ec49707e734a0038432957f6775df1d54b8c6de66caff.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 435, + 285, + 492, + 342 + ], + "blocks": [ + { + "bbox": [ + 435, + 285, + 492, + 342 + ], + "lines": [ + { + "bbox": [ + 435, + 285, + 492, + 342 + ], + "spans": [ + { + "bbox": [ + 435, + 285, + 492, + 342 + ], + "type": "image", + "image_path": "06dc785169846cfd8d5c373b8698d4ffc2457e433af540f531d4b57d0f829a1f.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 495, + 226, + 553, + 284 + ], + "blocks": [ + { + "bbox": [ + 495, + 226, + 553, + 284 + ], + "lines": [ + { + "bbox": [ + 495, + 226, + 553, + 284 + ], + "spans": [ + { + "bbox": [ + 495, + 226, + 553, + 284 + ], + "type": "image", + "image_path": "94802e72e417ea78c39f53cace29debadff122c963db708f69e04b111fc4ffc7.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 495, + 285, + 553, + 342 + ], + "blocks": [ + { + "bbox": [ + 495, + 285, + 553, + 342 + ], + "lines": [ + { + "bbox": [ + 495, + 285, + 553, + 342 + ], + "spans": [ + { + "bbox": [ + 495, + 285, + 553, + 342 + ], + "type": "image", + "image_path": "3473587401498fe280f8d76908cf995f24de07b5117a072217a90f739d9a5486.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 316, + 348, + 373, + 406 + ], + "blocks": [ + { + "bbox": [ + 316, + 348, + 373, + 406 + ], + "lines": [ + { + "bbox": [ + 316, + 348, + 373, + 406 + ], + "spans": [ + { + "bbox": [ + 316, + 348, + 373, + 406 + ], + "type": "image", + "image_path": "a85619011eaae6615460963d82b003da2de87ed01e3e0ad0b77509fd40d19ff6.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 316, + 406, + 373, + 464 + ], + "blocks": [ + { + "bbox": [ + 316, + 406, + 373, + 464 + ], + "lines": [ + { + "bbox": [ + 316, + 406, + 373, + 464 + ], + "spans": [ + { + "bbox": [ + 316, + 406, + 373, + 464 + ], + "type": "image", + "image_path": "6d5192fb92440f9f7a9f5f9e8b48e6c3d95d8871ed062fb5ff5a4600ee436db1.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 322, + 466, + 366, + 475 + ], + "lines": [ + { + "bbox": [ + 322, + 466, + 366, + 475 + ], + "spans": [ + { + "bbox": [ + 322, + 466, + 366, + 475 + ], + "type": "text", + "content": "(a) Condition" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 377, + 348, + 433, + 405 + ], + "blocks": [ + { + "bbox": [ + 377, + 348, + 433, + 405 + ], + "lines": [ + { + "bbox": [ + 377, + 348, + 433, + 405 + ], + "spans": [ + { + "bbox": [ + 377, + 348, + 433, + 405 + ], + "type": "image", + "image_path": "48e2934d33a138d09618621ae0b10f915b65955eb2195a2dd7e41c99c0523e29.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 377, + 406, + 433, + 464 + ], + "blocks": [ + { + "bbox": [ + 377, + 406, + 433, + 464 + ], + "lines": [ + { + "bbox": [ + 377, + 406, + 433, + 464 + ], + "spans": [ + { + "bbox": [ + 377, + 406, + 433, + 464 + ], + "type": "image", + "image_path": "d852296cb83b6b41f5a1a0d3d22cf26c74df25ac9a4dbaf3c20611be1273d50c.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 420, + 466, + 448, + 475 + ], + "lines": [ + { + "bbox": [ + 420, + 466, + 448, + 475 + ], + "spans": [ + { + "bbox": [ + 420, + 466, + 448, + 475 + ], + "type": "text", + "content": "(b) Ours" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 435, + 348, + 492, + 405 + ], + "blocks": [ + { + "bbox": [ + 435, + 348, + 492, + 405 + ], + "lines": [ + { + "bbox": [ + 435, + 348, + 492, + 405 + ], + "spans": [ + { + "bbox": [ + 435, + 348, + 492, + 405 + ], + "type": "image", + "image_path": "cdacc8f52e07ac25f4fa920dcb0b4f4703c4445f3e58bb534556716f5b49ac29.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 435, + 406, + 492, + 464 + ], + "blocks": [ + { + "bbox": [ + 435, + 406, + 492, + 464 + ], + "lines": [ + { + "bbox": [ + 435, + 406, + 492, + 464 + ], + "spans": [ + { + "bbox": [ + 435, + 406, + 492, + 464 + ], + "type": "image", + "image_path": "7a529dcb871fe764f313a2b0a9dcf845c3eebd2ba3d616624e6d2916ae85312f.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 496, + 348, + 553, + 405 + ], + "blocks": [ + { + "bbox": [ + 496, + 348, + 553, + 405 + ], + "lines": [ + { + "bbox": [ + 496, + 348, + 553, + 405 + ], + "spans": [ + { + "bbox": [ + 496, + 348, + 553, + 405 + ], + "type": "image", + "image_path": "6ca033b21327cd259be2f7120b59788f61b4b52142988f95d5a5a38a4b1a770b.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 496, + 406, + 553, + 464 + ], + "blocks": [ + { + "bbox": [ + 496, + 406, + 553, + 464 + ], + "lines": [ + { + "bbox": [ + 496, + 406, + 553, + 464 + ], + "spans": [ + { + "bbox": [ + 496, + 406, + 553, + 464 + ], + "type": "image", + "image_path": "d80b7b18f4d9ccadc468d0e6cf3eeb379f0f195ffc81f02a1728d5ea58c9aec4.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 503, + 466, + 546, + 475 + ], + "lines": [ + { + "bbox": [ + 503, + 466, + 546, + 475 + ], + "spans": [ + { + "bbox": [ + 503, + 466, + 546, + 475 + ], + "type": "text", + "content": "(c)Baselines" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 485, + 555, + 584 + ], + "lines": [ + { + "bbox": [ + 313, + 485, + 555, + 584 + ], + "spans": [ + { + "bbox": [ + 313, + 485, + 555, + 584 + ], + "type": "text", + "content": "Figure 1. Examples of color-conditioned generation from different methods. Color conditions (a) can be derived from imagery (top two rows) or manual drawing (bottom two rows). Input text prompts are \"painting\" and \"pen\" (1st & 3rd rows), and \"boat\" (2nd & 4th rows). Our method (b) can generate pixels closely aligned with the color conditions and effectively disentangle the colors into different structures. Existing baselines [14, 36, 46, 48] (from top to bottom) (c) struggle with either the semantics (the first two rows) or colors (the last two rows) of generated contents." + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 27 + }, + { + "bbox": [ + 313, + 605, + 553, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 553, + 641 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 553, + 641 + ], + "type": "text", + "content": "into various structures. Such a color-conditioned generation ability would be a great aid to numerous downstream tasks in creative artwork and graphic design [10]." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "content": "Different attempts of color-conditioned image generation are visualized in Fig. 1, where previous works often produce suboptimal results that do not align with the aforementioned criteria. In this paper, we propose an effective yet control-easy method facilitating color alignment in diffusion models for color-conditioned image synthesis. Specifically," + } + ] + } + ], + "index": 31 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "28446" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 276 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 276 + ], + "type": "text", + "content": "our method accepts a color condition via a color pattern, determining color values and their proportions in a synthesized image. Different from the regular diffusion approach, we perform the color alignment on intermediate samples or latent representations during the diffusion process prior to inputting them into a denoising model for image synthesis, thereby facilitating the perception of the conditional colors in diffusion models while preserving the overall attribution of the diffusion process. Furthermore, our method can support various settings: re-training of a color-aligned diffusion model; fine-tuning of a color-aligned diffusion model from a pre-trained regular diffusion model; and zero-shot approximation (training-free) of the color alignment process with a pre-trained diffusion model. We validate our method by demonstrating its capabilities across various color conditions and compare it with existing baselines on benchmark datasets. In summary, we make the following contributions:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 276, + 296, + 479 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 55, + 276, + 296, + 347 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 276, + 296, + 347 + ], + "spans": [ + { + "bbox": [ + 55, + 276, + 296, + 347 + ], + "type": "text", + "content": "- We address a challenging task: fine-grained color-conditioned image synthesis. We aim to generate images that closely follow a specified color pattern that defines color values and proportions without requiring an explicit structure. Our approach thus enhances the controllability and flexibility of image generation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 347, + 296, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 347, + 296, + 430 + ], + "spans": [ + { + "bbox": [ + 55, + 347, + 296, + 430 + ], + "type": "text", + "content": "- We propose a novel color alignment method that operates on the diffusion process of diffusion models, allowing the diffusion to effectively perceive color conditions while maintaining the quality and diversity of generated contents. We derive various settings including re-training, fine-tuning, and a zero-shot approximation, to adapt our color alignment method into different down-stream needs." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 431, + 296, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 431, + 296, + 479 + ], + "spans": [ + { + "bbox": [ + 55, + 431, + 296, + 479 + ], + "type": "text", + "content": "- We conduct extensive experiments to validate our proposed method in various scenarios and controllability settings. Our results demonstrate state-of-the-art performance compared with existing baselines." + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 488, + 140, + 500 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 488, + 140, + 500 + ], + "spans": [ + { + "bbox": [ + 55, + 488, + 140, + 500 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 508, + 297, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 508, + 297, + 664 + ], + "spans": [ + { + "bbox": [ + 55, + 508, + 297, + 664 + ], + "type": "text", + "content": "Neural image synthesis has started with Variational Autoencoders (VAEs) [22, 41, 43] and Generative Adversarial Networks (GANs) [12, 19, 20, 26], followed then by transformer-based methods [9, 23] and diffusion models [18, 31, 42]. Among these, Denoising Diffusion Probabilistic Models (DDPM) [18] have attracted considerable attention due to their generation quality and ease of extension. These models have been further developed into text-to-image (T2I) diffusion techniques [35, 36, 39], which have been boosted recently to achieve greater controllability through incorporation of extra conditions [5, 37, 38, 48]. Our method follows the general diffusion framework with a focus on fine-grained color-conditioned image synthesis." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 677, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 297, + 714 + ], + "type": "text", + "content": "Reference-based conditional diffusion. An intuitive condition for diffusion models is the use of an image reference. Structural conditions, such as edge maps and semantic maps," + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 556, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 556, + 240 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 556, + 240 + ], + "type": "text", + "content": "have been recently studied in [24, 25, 30, 38, 45, 48]. These methods strictly apply spatial constraints given in the reference to align generated pixels. A more relaxed approach is image editing [5, 6, 13, 21, 27], where the structure and appearance in the reference loosely guide the generation of target images. Those works aim to achieve shifts in appearance while permitting a small range of spatial edits, such as changes in the size and pose of objects. Several methods [14, 37, 46] condition only on abstract concepts in the reference, such as image/object style, enabling more creative synthesis. In this work, we disentangle colors in the reference from their original spatial structure while preserving the color values and proportions. Our method allows a variety of user-provided color materials to construct diverse structures." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 261, + 557, + 394 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 261, + 557, + 394 + ], + "spans": [ + { + "bbox": [ + 313, + 261, + 557, + 394 + ], + "type": "text", + "content": "Sample altering in diffusion. As indicated in [25], diffusion from the same noisy sample can result in similar outputs. Diffusion inversion [27] is commonly used to construct samples, used directly [27, 33] or mapped to another diffusion process [7, 29, 49] to diverge outputs. Several methods update intermediate latent representations with gradients calculated from external modules, e.g., other networks [47], images [40], features [28]. The Gaussian noise is defined [11] using estimated mean and variance to improve model convergence. Cold Diffusion [4] replaces the noising process with image degradations such as blurring and masking." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 407, + 419, + 422 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 407, + 419, + 422 + ], + "spans": [ + { + "bbox": [ + 313, + 407, + 419, + 422 + ], + "type": "text", + "content": "3. Proposed method" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 429, + 556, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 429, + 556, + 561 + ], + "spans": [ + { + "bbox": [ + 313, + 429, + 556, + 561 + ], + "type": "text", + "content": "We first in Sec. 3.1 review the regular diffusion process. In Sec. 3.2, we introduce our color alignment technique under its simplest setting, i.e., re-training of a diffusion model with color alignment. We then in Sec. 3.3 extend our technique to latent diffusion models, where we address typical challenges in latent space diffusion and demonstrate our method's ability via fine-tuning of a pre-trained latent diffusion model to a color-aligned one. Finally, in Sec. 3.4, we present a zero-shot variant of our method that achieves similar generation quality without extra training. We provide an overview of our pipeline and the regular diffusion in Fig. 2." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 572, + 446, + 585 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 572, + 446, + 585 + ], + "spans": [ + { + "bbox": [ + 313, + 572, + 446, + 585 + ], + "type": "text", + "content": "3.1. Diffusion preliminaries" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 591, + 555, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 591, + 555, + 639 + ], + "spans": [ + { + "bbox": [ + 313, + 591, + 555, + 639 + ], + "type": "text", + "content": "Diffusion models [18] could be seen as a variant of VAEs [22] that map a data distribution " + }, + { + "bbox": [ + 313, + 591, + 555, + 639 + ], + "type": "inline_equation", + "content": "\\mathcal{X}(\\mathbf{x}_0)" + }, + { + "bbox": [ + 313, + 591, + 555, + 639 + ], + "type": "text", + "content": " from a Gaussian distribution " + }, + { + "bbox": [ + 313, + 591, + 555, + 639 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(\\mathbf{0},\\mathbf{I})" + }, + { + "bbox": [ + 313, + 591, + 555, + 639 + ], + "type": "text", + "content": ". In the forward process, a sample " + }, + { + "bbox": [ + 313, + 591, + 555, + 639 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t" + }, + { + "bbox": [ + 313, + 591, + 555, + 639 + ], + "type": "text", + "content": " is encoded by adding a Gaussian noise to " + }, + { + "bbox": [ + 313, + 591, + 555, + 639 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0" + }, + { + "bbox": [ + 313, + 591, + 555, + 639 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 347, + 651, + 555, + 664 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 651, + 555, + 664 + ], + "spans": [ + { + "bbox": [ + 347, + 651, + 555, + 664 + ], + "type": "interline_equation", + "content": "\\mathbf {x} _ {t} = \\sqrt {\\bar {\\alpha} _ {t}} \\mathbf {x} _ {0} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\boldsymbol {\\epsilon}, \\quad \\boldsymbol {\\epsilon} \\sim \\mathcal {N} (\\mathbf {0}, \\mathbf {I}) \\tag {1}", + "image_path": "bc84d539833b39a32bd2cc82a4ec2af62e5c4e58d6afa271002b865a38b8891b.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 676, + 553, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 676, + 553, + 700 + ], + "spans": [ + { + "bbox": [ + 313, + 676, + 553, + 700 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 676, + 553, + 700 + ], + "type": "inline_equation", + "content": "t \\in [1, T]" + }, + { + "bbox": [ + 313, + 676, + 553, + 700 + ], + "type": "text", + "content": " is the time step associated with a noise strength " + }, + { + "bbox": [ + 313, + 676, + 553, + 700 + ], + "type": "inline_equation", + "content": "\\bar{\\alpha}_t" + }, + { + "bbox": [ + 313, + 676, + 553, + 700 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 313, + 676, + 553, + 700 + ], + "type": "inline_equation", + "content": "(1 - \\bar{\\alpha}_1) \\approx 0" + }, + { + "bbox": [ + 313, + 676, + 553, + 700 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 676, + 553, + 700 + ], + "type": "inline_equation", + "content": "(1 - \\bar{\\alpha}_T) \\approx 1" + }, + { + "bbox": [ + 313, + 676, + 553, + 700 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 326, + 701, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 701, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 326, + 701, + 553, + 713 + ], + "type": "text", + "content": "Then, a denoising model " + }, + { + "bbox": [ + 326, + 701, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 326, + 701, + 553, + 713 + ], + "type": "text", + "content": " with parameters " + }, + { + "bbox": [ + 326, + 701, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 326, + 701, + 553, + 713 + ], + "type": "text", + "content": " learns to" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28447" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 279, + 85 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 279, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 279, + 85 + ], + "type": "text", + "content": "reverse the process in Eq. 1 by training with a loss " + }, + { + "bbox": [ + 55, + 72, + 279, + 85 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 55, + 72, + 279, + 85 + ], + "type": "text", + "content": " as," + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 82, + 95, + 296, + 110 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 95, + 296, + 110 + ], + "spans": [ + { + "bbox": [ + 82, + 95, + 296, + 110 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\nabla_ {\\theta} \\| \\epsilon - \\epsilon_ {\\theta} (\\mathbf {x} _ {t}, t) \\| _ {2} ^ {2} \\tag {2}", + "image_path": "add303cc5e62ebdbfbc32c577e9f15e41b9d1f3ffb0473ba98de3a5427cf51d2.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 79, + 111, + 296, + 136 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 111, + 296, + 136 + ], + "spans": [ + { + "bbox": [ + 79, + 111, + 296, + 136 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {x}} _ {0} = \\frac {1}{\\sqrt {\\bar {\\alpha} _ {t}}} \\left(\\mathbf {x} _ {t} - \\sqrt {1 - \\bar {\\alpha} _ {t}} \\boldsymbol {\\epsilon} _ {\\theta} \\left(\\mathbf {x} _ {t}, t\\right)\\right) \\tag {3}", + "image_path": "2718e9580fb57c2ec708a1737a4d72d71e8fdc60a0524366c3e9ccf3e4381dfc.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 69, + 137, + 296, + 163 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 137, + 296, + 163 + ], + "spans": [ + { + "bbox": [ + 69, + 137, + 296, + 163 + ], + "type": "interline_equation", + "content": "\\mathbf {x} _ {t - 1} = \\frac {\\sqrt {\\bar {\\alpha} _ {t - 1}} \\beta_ {t}}{1 - \\bar {\\alpha} _ {t}} \\hat {\\mathbf {x}} _ {0} + \\frac {\\sqrt {\\alpha_ {t}} (1 - \\bar {\\alpha} _ {t - 1})}{1 - \\bar {\\alpha} _ {t}} \\mathbf {x} _ {t} + \\sigma_ {t} \\boldsymbol {\\varepsilon} \\quad (4)", + "image_path": "d4e1845ceff5d492fdd9bb076005f3be940050e7e38605f3389c243d214a2439.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 173, + 296, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 173, + 296, + 198 + ], + "spans": [ + { + "bbox": [ + 55, + 173, + 296, + 198 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 173, + 296, + 198 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}_0" + }, + { + "bbox": [ + 55, + 173, + 296, + 198 + ], + "type": "text", + "content": " is the predicted sample. " + }, + { + "bbox": [ + 55, + 173, + 296, + 198 + ], + "type": "inline_equation", + "content": "\\alpha ,\\beta ,\\sigma \\in \\mathbb{R}" + }, + { + "bbox": [ + 55, + 173, + 296, + 198 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 173, + 296, + 198 + ], + "type": "inline_equation", + "content": "\\varepsilon \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})" + }, + { + "bbox": [ + 55, + 173, + 296, + 198 + ], + "type": "text", + "content": " are scheduler-related parameters [18]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 207, + 258, + 220 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 207, + 258, + 220 + ], + "spans": [ + { + "bbox": [ + 55, + 207, + 258, + 220 + ], + "type": "text", + "content": "3.2. Color-aligned diffusion in image space" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 225, + 296, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 225, + 296, + 392 + ], + "spans": [ + { + "bbox": [ + 55, + 225, + 296, + 392 + ], + "type": "text", + "content": "As presented in Sec. 3.1, the regular diffusion model learns to reconstruct " + }, + { + "bbox": [ + 55, + 225, + 296, + 392 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0" + }, + { + "bbox": [ + 55, + 225, + 296, + 392 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 55, + 225, + 296, + 392 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(\\mathbf{0},\\mathbf{I})" + }, + { + "bbox": [ + 55, + 225, + 296, + 392 + ], + "type": "text", + "content": " during training and allows sampling-free generation in testing, i.e., unconditional generation. To condition the diffusion process with color conditions, existing methods pass color text prompts [36, 46] or referenced images [14, 46, 48] in parallel with " + }, + { + "bbox": [ + 55, + 225, + 296, + 392 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t" + }, + { + "bbox": [ + 55, + 225, + 296, + 392 + ], + "type": "text", + "content": " untouched. We found this approach loosely constrains the sampled data manifold, i.e., " + }, + { + "bbox": [ + 55, + 225, + 296, + 392 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}_0" + }, + { + "bbox": [ + 55, + 225, + 296, + 392 + ], + "type": "text", + "content": " can be sampled with out-of-bound colors. This is because the conditions are not involved in the sampling process (in construction of " + }, + { + "bbox": [ + 55, + 225, + 296, + 392 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t" + }, + { + "bbox": [ + 55, + 225, + 296, + 392 + ], + "type": "text", + "content": ") and the denoising model " + }, + { + "bbox": [ + 55, + 225, + 296, + 392 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 55, + 225, + 296, + 392 + ], + "type": "text", + "content": " hence leans towards the unconditional generation. Moreover, colors are learned in conjunction with spatial information given in referenced images, limiting the diversity of generated contents (see Fig. 1)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 394, + 297, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 394, + 297, + 502 + ], + "spans": [ + { + "bbox": [ + 55, + 394, + 297, + 502 + ], + "type": "text", + "content": "In this paper, we take a color pattern " + }, + { + "bbox": [ + 55, + 394, + 297, + 502 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 55, + 394, + 297, + 502 + ], + "type": "text", + "content": ", determining color values and their proportions in synthesis, as a color condition. The color pattern can be an imagery or hand-drawing (see Fig. 1). To effectively constrain the sampling manifold with color condition, we involve the color pattern " + }, + { + "bbox": [ + 55, + 394, + 297, + 502 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 55, + 394, + 297, + 502 + ], + "type": "text", + "content": " into the diffusion process via a new denoising model " + }, + { + "bbox": [ + 55, + 394, + 297, + 502 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(\\mathbf{x}_t,t,\\mathbf{c})" + }, + { + "bbox": [ + 55, + 394, + 297, + 502 + ], + "type": "text", + "content": ". In addition, we relax the spatial information in " + }, + { + "bbox": [ + 55, + 394, + 297, + 502 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 55, + 394, + 297, + 502 + ], + "type": "text", + "content": " by using a color alignment function " + }, + { + "bbox": [ + 55, + 394, + 297, + 502 + ], + "type": "inline_equation", + "content": "f(\\mathbf{x}_t,\\mathbf{c})" + }, + { + "bbox": [ + 55, + 394, + 297, + 502 + ], + "type": "text", + "content": " that produces a color-aligned image from an image " + }, + { + "bbox": [ + 55, + 394, + 297, + 502 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t" + }, + { + "bbox": [ + 55, + 394, + 297, + 502 + ], + "type": "text", + "content": " and color pattern " + }, + { + "bbox": [ + 55, + 394, + 297, + 502 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 55, + 394, + 297, + 502 + ], + "type": "text", + "content": " as," + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 97, + 512, + 296, + 534 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 512, + 296, + 534 + ], + "spans": [ + { + "bbox": [ + 97, + 512, + 296, + 534 + ], + "type": "interline_equation", + "content": "f \\left(\\mathbf {x} _ {t}, \\mathbf {c}\\right) [ p ] = \\underset {\\mathbf {c} [ q ]} {\\arg \\min } \\| \\mathbf {x} _ {t} [ p ] - \\mathbf {c} [ q ] \\| _ {2} ^ {2} \\tag {5}", + "image_path": "9e285fd3f24a5871e6d1ff7914582dca0ccbc140f897c04290ffc51eb4186356.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 544, + 252, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 544, + 252, + 558 + ], + "spans": [ + { + "bbox": [ + 55, + 544, + 252, + 558 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 544, + 252, + 558 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t[p]" + }, + { + "bbox": [ + 55, + 544, + 252, + 558 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 544, + 252, + 558 + ], + "type": "inline_equation", + "content": "\\mathbf{c}[q]" + }, + { + "bbox": [ + 55, + 544, + 252, + 558 + ], + "type": "text", + "content": " are colors at pixels " + }, + { + "bbox": [ + 55, + 544, + 252, + 558 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 55, + 544, + 252, + 558 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 544, + 252, + 558 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 55, + 544, + 252, + 558 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 558, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 297, + 714 + ], + "type": "text", + "content": "We visualize the diffusion process in Fig. 3. Our color alignment technique benefits the sampling in several aspects. First, it enforces the generated pixel colors within the target space, guaranteeing the accuracy. Second, its non-spatial manner allows pixels distribute more freely, fulfilling the color disentanglement. For the sake of completeness, we further concatenate " + }, + { + "bbox": [ + 55, + 558, + 297, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 55, + 558, + 297, + 714 + ], + "type": "text", + "content": " to the model's input as an extra hint. However, to completely break down the spatial information, we opt to randomly permute the pixel locations in " + }, + { + "bbox": [ + 55, + 558, + 297, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 55, + 558, + 297, + 714 + ], + "type": "text", + "content": ", resulting in " + }, + { + "bbox": [ + 55, + 558, + 297, + 714 + ], + "type": "inline_equation", + "content": "\\psi(\\mathbf{c})" + }, + { + "bbox": [ + 55, + 558, + 297, + 714 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 55, + 558, + 297, + 714 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 55, + 558, + 297, + 714 + ], + "type": "text", + "content": " being an image permutation operator. Such a permutation allows us to directly take " + }, + { + "bbox": [ + 55, + 558, + 297, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0" + }, + { + "bbox": [ + 55, + 558, + 297, + 714 + ], + "type": "text", + "content": " as color condition in training time, avoiding extra data requirement and data leakage. Finally, training and sampling of our color" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 316, + 71, + 391, + 190 + ], + "blocks": [ + { + "bbox": [ + 316, + 71, + 391, + 190 + ], + "lines": [ + { + "bbox": [ + 316, + 71, + 391, + 190 + ], + "spans": [ + { + "bbox": [ + 316, + 71, + 391, + 190 + ], + "type": "image", + "image_path": "f740dcf1b938da0a080229c45ea47377d6e35fcf57f37b87a3f145339caa1c82.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 332, + 194, + 416, + 204 + ], + "lines": [ + { + "bbox": [ + 332, + 194, + 416, + 204 + ], + "spans": [ + { + "bbox": [ + 332, + 194, + 416, + 204 + ], + "type": "text", + "content": "(a) Regular diffusion [36]" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 434, + 71, + 553, + 190 + ], + "blocks": [ + { + "bbox": [ + 434, + 71, + 553, + 190 + ], + "lines": [ + { + "bbox": [ + 434, + 71, + 553, + 190 + ], + "spans": [ + { + "bbox": [ + 434, + 71, + 553, + 190 + ], + "type": "image", + "image_path": "6cda9bde6cc68309455310737b2727d8902357514ecd19845311bbf3b14c9c11.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 436, + 194, + 553, + 204 + ], + "lines": [ + { + "bbox": [ + 436, + 194, + 553, + 204 + ], + "spans": [ + { + "bbox": [ + 436, + 194, + 553, + 204 + ], + "type": "text", + "content": "(b) Regular diffusion with condition" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 316, + 213, + 433, + 371 + ], + "blocks": [ + { + "bbox": [ + 316, + 213, + 433, + 371 + ], + "lines": [ + { + "bbox": [ + 316, + 213, + 433, + 371 + ], + "spans": [ + { + "bbox": [ + 316, + 213, + 433, + 371 + ], + "type": "image", + "image_path": "db639651312e35320c3d57121c13e2a6fe3fe026c512b2f1a6bd5647e94ffae7.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 314, + 376, + 434, + 396 + ], + "lines": [ + { + "bbox": [ + 314, + 376, + 434, + 396 + ], + "spans": [ + { + "bbox": [ + 314, + 376, + 434, + 396 + ], + "type": "text", + "content": "(c) Our color-aligned diffusion for retraining or fine-tuning" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 405, + 555, + 460 + ], + "lines": [ + { + "bbox": [ + 313, + 405, + 555, + 460 + ], + "spans": [ + { + "bbox": [ + 313, + 405, + 555, + 460 + ], + "type": "text", + "content": "Figure 2. Pipelines of the regular diffusion (a) and its color-conditioned version (b), our color-aligned diffusion (c), where noisy samples " + }, + { + "bbox": [ + 313, + 405, + 555, + 460 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t" + }, + { + "bbox": [ + 313, + 405, + 555, + 460 + ], + "type": "text", + "content": " are aligned with conditioned colors, and our zero-shot version (d), where color alignment is directly applied to predicted sample " + }, + { + "bbox": [ + 313, + 405, + 555, + 460 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}_0" + }, + { + "bbox": [ + 313, + 405, + 555, + 460 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 436, + 213, + 553, + 371 + ], + "blocks": [ + { + "bbox": [ + 436, + 213, + 553, + 371 + ], + "lines": [ + { + "bbox": [ + 436, + 213, + 553, + 371 + ], + "spans": [ + { + "bbox": [ + 436, + 213, + 553, + 371 + ], + "type": "image", + "image_path": "a270035bdc2014badd0926ff6f48f34a6a54ce077f59c3352ac33d258c09b61e.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 435, + 376, + 555, + 396 + ], + "lines": [ + { + "bbox": [ + 435, + 376, + 555, + 396 + ], + "spans": [ + { + "bbox": [ + 435, + 376, + 555, + 396 + ], + "type": "text", + "content": "(d) Our zero-shot approximation of the color-aligned diffusion" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 482, + 529, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 482, + 529, + 495 + ], + "spans": [ + { + "bbox": [ + 313, + 482, + 529, + 495 + ], + "type": "text", + "content": "aligned diffusion process (Fig. 2c) are formulated as:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 330, + 502, + 474, + 517 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 330, + 502, + 474, + 517 + ], + "spans": [ + { + "bbox": [ + 330, + 502, + 474, + 517 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\nabla_ {\\boldsymbol {\\theta}} \\| \\boldsymbol {\\epsilon} ^ {\\prime} - \\boldsymbol {\\epsilon} _ {\\boldsymbol {\\theta}} \\big (f (\\mathbf {x} _ {t}, \\mathbf {c}), t, \\mathbf {c} \\big) \\| _ {2} ^ {2}", + "image_path": "0f57ff6f3680d4e383dfce7fb27638e0b3ba9868dd89d14eb85cec78c3db5393.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 330, + 518, + 555, + 545 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 330, + 518, + 555, + 545 + ], + "spans": [ + { + "bbox": [ + 330, + 518, + 555, + 545 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\epsilon} ^ {\\prime} = \\frac {f (\\mathbf {x} _ {t} , \\mathbf {c}) - \\sqrt {\\bar {\\alpha} _ {t}} \\mathbf {x} _ {0}}{\\sqrt {1 - \\bar {\\alpha} _ {t}}}, \\quad \\mathbf {c} = \\psi (\\mathbf {x} _ {0}) \\tag {6}", + "image_path": "94da3c655f1ca36735f2170051f2db3f0ed3f280ff9373f73060fbe59d2c9030.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 327, + 546, + 555, + 572 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 546, + 555, + 572 + ], + "spans": [ + { + "bbox": [ + 327, + 546, + 555, + 572 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {x}} _ {0} = \\frac {1}{\\sqrt {\\bar {\\alpha} _ {t}}} \\left(f \\left(\\mathbf {x} _ {t}, \\mathbf {c}\\right) - \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon_ {\\theta} \\left(f \\left(\\mathbf {x} _ {t}, \\mathbf {c}\\right), t, \\mathbf {c}\\right)\\right) \\tag {7}", + "image_path": "78b769ee34083d8aff039d9a08cade9227e719a63fea1b29d9ebc80fd0ddb367.jpg" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 580, + 556, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 580, + 556, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 580, + 556, + 665 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 580, + 556, + 665 + ], + "type": "inline_equation", + "content": "\\epsilon^{\\prime}" + }, + { + "bbox": [ + 313, + 580, + 556, + 665 + ], + "type": "text", + "content": " is the adapted noise complemented with the effect of " + }, + { + "bbox": [ + 313, + 580, + 556, + 665 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 313, + 580, + 556, + 665 + ], + "type": "text", + "content": ". Notably, we alter only what " + }, + { + "bbox": [ + 313, + 580, + 556, + 665 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 313, + 580, + 556, + 665 + ], + "type": "text", + "content": " sees and predicts. The parameters " + }, + { + "bbox": [ + 313, + 580, + 556, + 665 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 580, + 556, + 665 + ], + "type": "text", + "content": " thus can be learned more easily with color condition compared with the regular one. All other nonlearnable steps, e.g., the regular forward process and Eq. 4, remain unchanged for the ease of sampling. For more technical visualizations, please refer to the supplementary material." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 672, + 515, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 672, + 515, + 685 + ], + "spans": [ + { + "bbox": [ + 313, + 672, + 515, + 685 + ], + "type": "text", + "content": "3.3. Color-aligned diffusion in latent space" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "content": "Training of diffusion models in their latent space is a common practice. In this section, we extend our color alignment" + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28448" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "type": "text", + "content": "technique to latent diffusion models via fine-tuning of a pre-trained latent model to achieve color alignment." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 96, + 296, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 96, + 296, + 240 + ], + "spans": [ + { + "bbox": [ + 55, + 96, + 296, + 240 + ], + "type": "text", + "content": "Following the Stable Diffusion in [36], we encode " + }, + { + "bbox": [ + 55, + 96, + 296, + 240 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0" + }, + { + "bbox": [ + 55, + 96, + 296, + 240 + ], + "type": "text", + "content": " into its latent representation " + }, + { + "bbox": [ + 55, + 96, + 296, + 240 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_0" + }, + { + "bbox": [ + 55, + 96, + 296, + 240 + ], + "type": "text", + "content": ". We then train our model with alignment in the latent space, i.e., replacing " + }, + { + "bbox": [ + 55, + 96, + 296, + 240 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 55, + 96, + 296, + 240 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 55, + 96, + 296, + 240 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 55, + 96, + 296, + 240 + ], + "type": "text", + "content": " in Eqs. 6 and 7. However, when high-frequency structural information is present in " + }, + { + "bbox": [ + 55, + 96, + 296, + 240 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0" + }, + { + "bbox": [ + 55, + 96, + 296, + 240 + ], + "type": "text", + "content": ", unintended colors and local structural information can be encoded in " + }, + { + "bbox": [ + 55, + 96, + 296, + 240 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_0" + }, + { + "bbox": [ + 55, + 96, + 296, + 240 + ], + "type": "text", + "content": ". We discovered that blurring " + }, + { + "bbox": [ + 55, + 96, + 296, + 240 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0" + }, + { + "bbox": [ + 55, + 96, + 296, + 240 + ], + "type": "text", + "content": " prior to encoding it enhances the color disentanglement and accuracy, with minimal loss of the completeness in generated results. Also, we found that suspending color alignment at late time steps (e.g., " + }, + { + "bbox": [ + 55, + 96, + 296, + 240 + ], + "type": "inline_equation", + "content": "t < 0.2T" + }, + { + "bbox": [ + 55, + 96, + 296, + 240 + ], + "type": "text", + "content": " where diffusion model is fixing details) allows a proper refinement of the latent, making latent decoding towards more photo-realistic details." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 248, + 232, + 261 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 248, + 232, + 261 + ], + "spans": [ + { + "bbox": [ + 55, + 248, + 232, + 261 + ], + "type": "text", + "content": "3.4. Zero-shot color-aligned diffusion" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 266, + 296, + 362 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 266, + 296, + 362 + ], + "spans": [ + { + "bbox": [ + 55, + 266, + 296, + 362 + ], + "type": "text", + "content": "To mitigate the computational burden of training, we introduce a zero-shot approximation of our color alignment method. As shown in Eqs. 6 and 7, the training aims to reconstruct " + }, + { + "bbox": [ + 55, + 266, + 296, + 362 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}_0" + }, + { + "bbox": [ + 55, + 266, + 296, + 362 + ], + "type": "text", + "content": " whose color distribution follows the color condition " + }, + { + "bbox": [ + 55, + 266, + 296, + 362 + ], + "type": "inline_equation", + "content": "\\mathbf{c} = \\psi (\\mathbf{x}_0)" + }, + { + "bbox": [ + 55, + 266, + 296, + 362 + ], + "type": "text", + "content": ". Under the zero-shot setting, to achieve this ability without learning the parameters " + }, + { + "bbox": [ + 55, + 266, + 296, + 362 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 55, + 266, + 296, + 362 + ], + "type": "text", + "content": ", we directly map an unconditionally generated " + }, + { + "bbox": [ + 55, + 266, + 296, + 362 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}_0" + }, + { + "bbox": [ + 55, + 266, + 296, + 362 + ], + "type": "text", + "content": " (from Eq. 3) to " + }, + { + "bbox": [ + 55, + 266, + 296, + 362 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 55, + 266, + 296, + 362 + ], + "type": "text", + "content": " during the sampling process using a color alignment function " + }, + { + "bbox": [ + 55, + 266, + 296, + 362 + ], + "type": "inline_equation", + "content": "g(\\hat{\\mathbf{x}}_0,\\mathbf{c})" + }, + { + "bbox": [ + 55, + 266, + 296, + 362 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 104, + 383, + 296, + 404 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 383, + 296, + 404 + ], + "spans": [ + { + "bbox": [ + 104, + 383, + 296, + 404 + ], + "type": "interline_equation", + "content": "g \\left(\\hat {\\mathbf {x}} _ {0}, \\mathbf {c}\\right) = \\underset {\\psi (\\mathbf {c})} {\\arg \\min } \\| \\hat {\\mathbf {x}} _ {0} - \\psi (\\mathbf {c}) \\| _ {2} ^ {2} \\tag {8}", + "image_path": "772977c39607a3598a831922a071c9e0b505a0ccafaceb7013ba55d9725541d6.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 414, + 296, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 414, + 296, + 437 + ], + "spans": [ + { + "bbox": [ + 55, + 414, + 296, + 437 + ], + "type": "text", + "content": "where we map the unconditional " + }, + { + "bbox": [ + 55, + 414, + 296, + 437 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}_0" + }, + { + "bbox": [ + 55, + 414, + 296, + 437 + ], + "type": "text", + "content": " to a closest reorganization of the conditional colors " + }, + { + "bbox": [ + 55, + 414, + 296, + 437 + ], + "type": "inline_equation", + "content": "\\psi (\\mathbf{c})" + }, + { + "bbox": [ + 55, + 414, + 296, + 437 + ], + "type": "text", + "content": " for color alignment." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 438, + 296, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 438, + 296, + 641 + ], + "spans": [ + { + "bbox": [ + 55, + 438, + 296, + 641 + ], + "type": "text", + "content": "Unlike " + }, + { + "bbox": [ + 55, + 438, + 296, + 641 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 55, + 438, + 296, + 641 + ], + "type": "text", + "content": " defined in Eq. 5, " + }, + { + "bbox": [ + 55, + 438, + 296, + 641 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 55, + 438, + 296, + 641 + ], + "type": "text", + "content": " in Eq. 8 performs one-to-one mapping, i.e., each location in " + }, + { + "bbox": [ + 55, + 438, + 296, + 641 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 55, + 438, + 296, + 641 + ], + "type": "text", + "content": " is used for color alignment exactly once. Note that, like " + }, + { + "bbox": [ + 55, + 438, + 296, + 641 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 55, + 438, + 296, + 641 + ], + "type": "text", + "content": ", our " + }, + { + "bbox": [ + 55, + 438, + 296, + 641 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 55, + 438, + 296, + 641 + ], + "type": "text", + "content": " also applies non-spatial mapping. This would ensure desired properties (color accuracy, completeness, and disentanglement) from the zero-shot color-aligned diffusion. However, since " + }, + { + "bbox": [ + 55, + 438, + 296, + 641 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 55, + 438, + 296, + 641 + ], + "type": "text", + "content": " is not involved in the learning process, the color-aligned image " + }, + { + "bbox": [ + 55, + 438, + 296, + 641 + ], + "type": "inline_equation", + "content": "g(\\hat{\\mathbf{x}}_0,\\mathbf{c})" + }, + { + "bbox": [ + 55, + 438, + 296, + 641 + ], + "type": "text", + "content": " could be messy, showing invalid structures and outliers (see Fig. 2d). Fortunately, we observed that pre-trained diffusion models have capability of clustering semantics and eliminating outliers at early sampling steps (e.g., " + }, + { + "bbox": [ + 55, + 438, + 296, + 641 + ], + "type": "inline_equation", + "content": "t > 0.2T" + }, + { + "bbox": [ + 55, + 438, + 296, + 641 + ], + "type": "text", + "content": "). We also suspend the color alignment at late time steps to prevent excessive color changes and allow the model to refine details. Additionally, such zero-shot approximation is extendable to latent diffusion models. We found the same conclusion for using the blurring strategy as in Sec. 3.3 applies." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": "Another interest could be replacing " + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": " in learning. However, we find this not applicable because " + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "inline_equation", + "content": "g(\\mathbf{x}_t,\\mathbf{c})" + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": " corrupts the distribution of " + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t" + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": " too heavily. Its subsequent adapted noise " + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\epsilon^{\\prime}" + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": " (the learning target in Eq. 6) deviates too far from the standard Gaussian " + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(\\mathbf{0},\\mathbf{I})" + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": ", eliminating the effect of a proper diffusion process." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 72, + 395, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 395, + 85 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 395, + 85 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 91, + 376, + 102 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 91, + 376, + 102 + ], + "spans": [ + { + "bbox": [ + 314, + 91, + 376, + 102 + ], + "type": "text", + "content": "4.1. Datasets" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 109, + 555, + 253 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 109, + 555, + 253 + ], + "spans": [ + { + "bbox": [ + 313, + 109, + 555, + 253 + ], + "type": "text", + "content": "We experimented our method in two scenarios: color-aligned diffusion in image space and color-aligned diffusion in latent space. For diffusion in image space, we ran experiments on Oxford-flower [32] (7.17k training and 1.02k test images) and Microsoft-emoji [8] datasets (6.80k training and 0.76k test images) at resolution " + }, + { + "bbox": [ + 313, + 109, + 555, + 253 + ], + "type": "inline_equation", + "content": "64 \\times 64" + }, + { + "bbox": [ + 313, + 109, + 555, + 253 + ], + "type": "text", + "content": ". For latent diffusion, we experimented on a sub-set of a high-quality generic dataset Text-to-image-2M [1] (300k training and 20k test images) at resolution " + }, + { + "bbox": [ + 313, + 109, + 555, + 253 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 313, + 109, + 555, + 253 + ], + "type": "text", + "content": ". To validate the color disentanglement ability of our method, we replaced original prompts in Text-to-image-2M [1] with 50 daily object prompts randomly generated by ChatGPT [2] (e.g., those in Fig. 4)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 261, + 380, + 272 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 261, + 380, + 272 + ], + "spans": [ + { + "bbox": [ + 314, + 261, + 380, + 272 + ], + "type": "text", + "content": "4.2. Baselines" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 278, + 554, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 278, + 554, + 350 + ], + "spans": [ + { + "bbox": [ + 313, + 278, + 554, + 350 + ], + "type": "text", + "content": "Our color-aligned image diffusion version is adapted from the DDPM [18]. Therefore, we compare our method with the DDPM to demonstrate its extra color-conditioned ability without requiring additional data creation or labeling. For the same reason, we also compare our latent diffusion version with the pre-trained Stable Diffusion in [36]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 350, + 556, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 350, + 556, + 518 + ], + "spans": [ + { + "bbox": [ + 313, + 350, + 556, + 518 + ], + "type": "text", + "content": "Given that our color reference can be an in-the-wild image or a manual drawing, we compare our method with SOTA reference-based generation methods. IP-Adapter [46] and T2I-Adapter [30] convert an image reference into features, which are subsequently combined with the model features for image generation. Style-Aligned [14] transfers source image information from its attention layer to the target generation attention layer, thereby preserving spatial and style features. ControlNet-Color is a variant of ControlNet [48] that accepts a predefined pixelized color image as input and generates a refined version in which the colors are approximately aligned with the original image spatially. IP-Adapter [46] and ControlNet-Color [48] require fine-tuning, while Style-Aligned [14] operates as a zero-shot approach." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 526, + 444, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 526, + 444, + 538 + ], + "spans": [ + { + "bbox": [ + 313, + 526, + 444, + 538 + ], + "type": "text", + "content": "4.3. Implementation details" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 544, + 555, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 544, + 555, + 604 + ], + "spans": [ + { + "bbox": [ + 313, + 544, + 555, + 604 + ], + "type": "text", + "content": "We follow default settings in the DDPM [18] and the Stable Diffusion [36] for network architectures and diffusion scheduling, specifically, from the huggingface-diffusers [44]. Further details, e.g., DDPM re-training, Stable Diffusion fine-tuning are presented in our supplementary material." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 604, + 555, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 604, + 555, + 662 + ], + "spans": [ + { + "bbox": [ + 313, + 604, + 555, + 662 + ], + "type": "text", + "content": "We trained our image diffusion version for 100 epochs and latent diffusion version for " + }, + { + "bbox": [ + 313, + 604, + 555, + 662 + ], + "type": "inline_equation", + "content": "160\\mathrm{k}" + }, + { + "bbox": [ + 313, + 604, + 555, + 662 + ], + "type": "text", + "content": " iterations. We performed inference at 50 time steps. We used classifier-free guidance [17] with guidance scale 5. All experiments were executed on two RTX 3090 GPUs with batch size 4." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 672, + 417, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 672, + 417, + 685 + ], + "spans": [ + { + "bbox": [ + 313, + 672, + 417, + 685 + ], + "type": "text", + "content": "4.4. Qualitative study" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": "Color-aligned image diffusion. We visually compare our method with the regular diffusion in DDPM [18] in Fig. 3." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28449" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 72, + 291, + 152 + ], + "blocks": [ + { + "bbox": [ + 57, + 72, + 291, + 152 + ], + "lines": [ + { + "bbox": [ + 57, + 72, + 291, + 152 + ], + "spans": [ + { + "bbox": [ + 57, + 72, + 291, + 152 + ], + "type": "image", + "image_path": "f7608774d3e9bb2970cd84dd5ea576b611c4e671295186736580e59440c6ecba.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 153, + 295, + 182 + ], + "lines": [ + { + "bbox": [ + 55, + 153, + 295, + 182 + ], + "spans": [ + { + "bbox": [ + 55, + 153, + 295, + 182 + ], + "type": "text", + "content": "(a) Our color-aligned image diffusion process. The top row shows our training inputs, created from early to late steps in the forward process. The bottom row is a test case from our reverse process, sampled from pure noise." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 57, + 185, + 291, + 265 + ], + "blocks": [ + { + "bbox": [ + 57, + 185, + 291, + 265 + ], + "lines": [ + { + "bbox": [ + 57, + 185, + 291, + 265 + ], + "spans": [ + { + "bbox": [ + 57, + 185, + 291, + 265 + ], + "type": "image", + "image_path": "17c9cf1f6efbf645cb14f5c2a30f4a4c8575dfca3ed7a9ed2ed01ccce09ac9c2.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 59, + 266, + 291, + 276 + ], + "lines": [ + { + "bbox": [ + 59, + 266, + 291, + 276 + ], + "spans": [ + { + "bbox": [ + 59, + 266, + 291, + 276 + ], + "type": "text", + "content": "(b) Regular image diffusion process in DDPM [18] in same order as (a)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 57, + 278, + 291, + 399 + ], + "blocks": [ + { + "bbox": [ + 57, + 278, + 291, + 399 + ], + "lines": [ + { + "bbox": [ + 57, + 278, + 291, + 399 + ], + "spans": [ + { + "bbox": [ + 57, + 278, + 291, + 399 + ], + "type": "image", + "image_path": "8f64d93754bd2496078cd957abd57ba6779b959093f5fc401ddec9697730a72e.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 401, + 295, + 422 + ], + "lines": [ + { + "bbox": [ + 55, + 401, + 295, + 422 + ], + "spans": [ + { + "bbox": [ + 55, + 401, + 295, + 422 + ], + "type": "text", + "content": "(c) Several color-aligned image synthesis results. From top to bottom rows are condition, our generated results, and regular diffusion's results." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 430, + 296, + 464 + ], + "lines": [ + { + "bbox": [ + 55, + 430, + 296, + 464 + ], + "spans": [ + { + "bbox": [ + 55, + 430, + 296, + 464 + ], + "type": "text", + "content": "Figure 3. Qualitative results of color-aligned image diffusion. (a)-(b) Visualization of the diffusion process by our method and DDPM [18]. (c) Generation results of our method and DDPM [18]." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 486, + 296, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 486, + 296, + 556 + ], + "spans": [ + { + "bbox": [ + 55, + 486, + 296, + 556 + ], + "type": "text", + "content": "As shown, our method (Fig. 3a) well restricts the diffusion process to the conditional color space, wherein only whites, yellows, and background blacks are allowed. In contrast, the regular diffusion in DDPM (Fig. 3b) generates out-of-bounds colors during the diffusion process, leading to an unintended reconstruction result." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 558, + 296, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 296, + 653 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 296, + 653 + ], + "type": "text", + "content": "Fig. 3c illustrates an immediate application of our color-aligned image diffusion in image rephrasing, i.e., colors in an input image are reorganized by our model to construct new structures. More importantly, this conditional process does not require any additional data (e.g., labeled data or paired data) or pre-defined spatial information (e.g., semantic maps or masks) in both training or inference. This makes our technique scalable to larger datasets." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "content": "Color-aligned latent diffusion. We present our coloraligned latent diffusion results in comparison with existing baselines in Fig. 4. As demonstrated, our method (Figs. 4b and 4c) effectively conditions the synthesis on target color inputs. Color values and proportions are aligned, i.e., color" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "type": "text", + "content": "accuracy and completeness are maintained. Additionally, our method generates contents that align well with target text prompts, showcasing the ability for color disentangle-ment where colors are rearranged into new structures. Furthermore, with these color-conditioned generation criteria preserved, our method achieves image quality and diversity on par with the original diffusion model." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 157, + 556, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 157, + 556, + 277 + ], + "spans": [ + { + "bbox": [ + 313, + 157, + 556, + 277 + ], + "type": "text", + "content": "In contrast, the baselines produce suboptimal results in various aspects. IP-Adapter [46] (Fig. 4d), T2I-Adapter [30] (Fig. 4e), and Style-Aligned [14] (Fig. 4f) construct target contents on top of an input image reference, thus often struggle to disentangle colors. Since the spatial information of colors is taken into account, these methods likely to fail completely if input reference and target text prompt belong to significantly different domains. ControlNet-Color [48] (Fig. 4g) treats input reference as a weak guidance, leading to failures to accurately preserve color values and proportions." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 277, + 555, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 277, + 555, + 373 + ], + "spans": [ + { + "bbox": [ + 313, + 277, + 555, + 373 + ], + "type": "text", + "content": "Generalizability and controllability. Our method also works with manual drawings, for example, a rough color pattern. As shown in Fig. 5, our method (the fine-tuned version) can generate images aligned with the color values and proportions specified in input color patterns. Moreover, even if only a few color values presented, our method can adjust these values to achieve appropriate smoothness, brightness, and shadow in generated contents." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 374, + 556, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 374, + 556, + 494 + ], + "spans": [ + { + "bbox": [ + 313, + 374, + 556, + 494 + ], + "type": "text", + "content": "These manual color conditions can be easily modified to influence the generation process. For example, the amount of each color value can be scaled (Fig. 5a). The color values can be tuned intuitively without the need for careful prompt engineering (Fig. 5b). Additional colors can be incorporated into conditions along with user ideation (Fig. 5c). By controlling target text prompt, we can specify the placement of colors (Fig. 5d). Note that, like the regular diffusion, our method can generate diverse results from the same input condition by sampling different noises." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 494, + 556, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 494, + 556, + 615 + ], + "spans": [ + { + "bbox": [ + 313, + 494, + 556, + 615 + ], + "type": "text", + "content": "Zero-shot diffusion. We showcase the ability of our method in zero-shot diffusion in Fig. 6. Our experiments indicate that, compared with our fine-tuned version, the zero-shot version tends to generate images with less fine-grained details, resulting in issues such as blurry lighting, uniform texture, and shadow missing. Since our zero-shot approach mechanically maps colors without model training, it performs more comparably to the fine-tuned version when applied to well-sampled color conditions, e.g., in-the-wild images in the fifth row in Fig. 6 and those in Fig. 4." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 624, + 424, + 636 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 624, + 424, + 636 + ], + "spans": [ + { + "bbox": [ + 313, + 624, + 424, + 636 + ], + "type": "text", + "content": "4.5.Quantitative study" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "text", + "content": "Metrics. We evaluate the quality of generated images using the widely adopted FID score [16]. To assess the color disentanglement, we use the CLIP-Score [15] measuring the similarity between a generated image and a target text prompt within the CLIP [34] embedding space. Since the color accuracy and completeness can be evaluated independ-" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28450" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 55, + 68, + 555, + 601 + ], + "blocks": [ + { + "bbox": [ + 55, + 68, + 555, + 601 + ], + "lines": [ + { + "bbox": [ + 55, + 68, + 555, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 68, + 555, + 601 + ], + "type": "image", + "image_path": "a7d6e88669e9438c366231e09cde4d8b0d81066ff8fa2d175dce681ab1e80c8f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 262, + 602, + 347, + 613 + ], + "lines": [ + { + "bbox": [ + 262, + 602, + 347, + 613 + ], + "spans": [ + { + "bbox": [ + 262, + 602, + 347, + 613 + ], + "type": "text", + "content": "(g) ControlNet-Color [48]" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 621, + 555, + 644 + ], + "lines": [ + { + "bbox": [ + 55, + 621, + 555, + 644 + ], + "spans": [ + { + "bbox": [ + 55, + 621, + 555, + 644 + ], + "type": "text", + "content": "Figure 4. Qualitative results of color-aligned latent diffusion. Each input (first row) includes an in-the-wild image as color condition and a target text prompt. Each column presents results of experimented methods using the same input." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 665, + 297, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 297, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 297, + 713 + ], + "type": "text", + "content": "dently of spatial relations, we measure the proximity between the color distributions of a generated image and input color condition. We employ the Chamfer Distance (CD) [3], a commonly used metric for measuring the distance between" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 665, + 555, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 555, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 555, + 689 + ], + "type": "text", + "content": "two point sets, to quantify the distance between generated RGB points " + }, + { + "bbox": [ + 313, + 665, + 555, + 689 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}_0" + }, + { + "bbox": [ + 313, + 665, + 555, + 689 + ], + "type": "text", + "content": " and input condition RGB points " + }, + { + "bbox": [ + 313, + 665, + 555, + 689 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 313, + 665, + 555, + 689 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 4 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28451" + } + ] + } + ], + "index": 5 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 70, + 556, + 441 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 556, + 441 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 556, + 441 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 556, + 441 + ], + "type": "image", + "image_path": "adaad4824b3da68008f02200840b931fd9d7feccb871eb2e3ac0c73114301c19.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 449, + 555, + 472 + ], + "lines": [ + { + "bbox": [ + 55, + 449, + 555, + 472 + ], + "spans": [ + { + "bbox": [ + 55, + 449, + 555, + 472 + ], + "type": "text", + "content": "Figure 5. Qualitative results of sampling and editing of input color conditions. All results are generated by our fine-tuned color-aligned latent model. Red arrows represent the generation process. Blue arrows represent the editing of color conditions." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 93, + 482, + 517, + 544 + ], + "blocks": [ + { + "bbox": [ + 93, + 482, + 517, + 544 + ], + "lines": [ + { + "bbox": [ + 93, + 482, + 517, + 544 + ], + "spans": [ + { + "bbox": [ + 93, + 482, + 517, + 544 + ], + "type": "table", + "html": "
MethodFID ↓CD-A ↓CD-C ↓Re-training Cost ↓Inference Cost ↓
Image Diffusion (Re-train) [18]63.9; 14940.6; 42.442.6; 39.01.001.00
Ours (Image; Re-train)57.5; 45.70.00; 0.004.12; 3.651.061.13
Ours (Image; Zero-shot)50.0; 84.55.59; 5.468.28; 10.40.006.76
", + "image_path": "8b756067626fe858ea710cc80de10258b46899b709d454cbb52cf78c4af58060.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 550, + 555, + 572 + ], + "lines": [ + { + "bbox": [ + 55, + 550, + 555, + 572 + ], + "spans": [ + { + "bbox": [ + 55, + 550, + 555, + 572 + ], + "type": "text", + "content": "Table 1. Quantitative results of image diffusion. In each metric column, the left and right numbers correspond to the scores from the Oxford-flower dataset [32] and Microsoft-emoji [8] dataset, respectively." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 69, + 611, + 296, + 639 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 611, + 296, + 639 + ], + "spans": [ + { + "bbox": [ + 69, + 611, + 296, + 639 + ], + "type": "interline_equation", + "content": "\\operatorname {C D - A} \\left(\\hat {\\mathbf {x}} _ {0}, \\mathbf {c}\\right) = \\frac {\\sum_ {\\hat {\\mathbf {x}} _ {0} [ p ]} \\min _ {\\mathbf {c} [ q ]} \\| \\hat {\\mathbf {x}} _ {0} [ p ] - \\mathbf {c} [ q ] \\| _ {2} ^ {2}}{| \\hat {\\mathbf {x}} _ {0} |} \\tag {9}", + "image_path": "0a6100c47891ea165d49298f765725fa4226e20be7603935fe5691b94c87b2c9.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 69, + 641, + 296, + 670 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 641, + 296, + 670 + ], + "spans": [ + { + "bbox": [ + 69, + 641, + 296, + 670 + ], + "type": "interline_equation", + "content": "\\operatorname {C D - C} \\left(\\hat {\\mathbf {x}} _ {0}, \\mathbf {c}\\right) = \\frac {\\sum_ {\\mathbf {c} [ q ]} \\min _ {\\hat {\\mathbf {x}} _ {0} [ p ]} \\| \\hat {\\mathbf {x}} _ {0} [ p ] - \\mathbf {c} [ q ] \\| _ {2} ^ {2}}{| \\mathbf {c} |} \\tag {10}", + "image_path": "3ae175e986d089520ccfce2011a988ae2b62937805b227c9aa046cb9a1af9519.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{CD} = \\mathrm{CD - A} + \\mathrm{CD - C}" + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": " consists of an accuracy term and a completeness term that measure the set distance from one to another; " + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "inline_equation", + "content": "|\\cdot |" + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": " denotes set cardinality. Lastly, we evaluate" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 594, + 555, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 616 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 616 + ], + "type": "text", + "content": "the inference cost of different methods, using the regular diffusion as the standard reference." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 617, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 556, + 713 + ], + "type": "text", + "content": "Baseline comparison. We first present the performance scores for the image diffusion case study in Tab. 1. The scores clearly indicate that our technique effectively preserves conditioned-colors. Note that, for inference, the color alignment is performed on RGB pixels until the final diffusion step, making the CD-A approach zero. We found that our fine-tuned version requires nearly the same training and inference costs as the regular diffusion model." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "28452" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 100, + 70, + 511, + 180 + ], + "blocks": [ + { + "bbox": [ + 100, + 70, + 511, + 180 + ], + "lines": [ + { + "bbox": [ + 100, + 70, + 511, + 180 + ], + "spans": [ + { + "bbox": [ + 100, + 70, + 511, + 180 + ], + "type": "table", + "html": "
MethodFID ↓CLIPScore ↑CD-A ↓CD-C ↓Inference Cost ↓
Stable Diffusion (Pre-train) [36]72.6; 72.627.3; 27.3166; 27.865.3; 16.01.00
Ours (Latent; Fine-tune)86.7; 69.429.0; 27.473.9; 4.9815.8; 4.871.03
Ours (Latent; Zero-shot)104; 77.928.3; 27.435.2; 2.683.19; 3.933.52
Baseline (IP-Adapter [46])202; 50.925.5; 22.0145; 10.1106; 13.91.44
Baseline (T2I-Adapter [30])134; 77.124.3; 24.9143; 8.3424.7; 7.011.56
Baseline (Style-Aligned [14])177; 73.123.5; 22.945.1; 6.1832.5; 8.505.69
Baseline (ControlNet-Color [48])105; 81.022.9; 23.4129; 13.149.8; 6.541.78
", + "image_path": "568e4668110ff5d15f34aa3c6247214ae0405a0f315fefc16b739c926b033b7a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 57, + 234, + 115, + 293 + ], + "blocks": [ + { + "bbox": [ + 57, + 234, + 115, + 293 + ], + "lines": [ + { + "bbox": [ + 57, + 234, + 115, + 293 + ], + "spans": [ + { + "bbox": [ + 57, + 234, + 115, + 293 + ], + "type": "image", + "image_path": "a6166db6e7cd9f137e37587475e03cadabb97ff0a257cafbdb37166b024cbe0d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 58, + 293, + 149, + 300 + ], + "lines": [ + { + "bbox": [ + 58, + 293, + 149, + 300 + ], + "spans": [ + { + "bbox": [ + 58, + 293, + 149, + 300 + ], + "type": "text", + "content": "\"A professional photo of a bar.\"" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 118, + 234, + 176, + 293 + ], + "blocks": [ + { + "bbox": [ + 118, + 234, + 176, + 293 + ], + "lines": [ + { + "bbox": [ + 118, + 234, + 176, + 293 + ], + "spans": [ + { + "bbox": [ + 118, + 234, + 176, + 293 + ], + "type": "image", + "image_path": "5f01b718d4c505a1d9183a0b728ef45d6bdbf7bdd393d14148fd6ec9378f7673.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 118, + 293, + 234, + 300 + ], + "lines": [ + { + "bbox": [ + 118, + 293, + 234, + 300 + ], + "spans": [ + { + "bbox": [ + 118, + 293, + 234, + 300 + ], + "type": "text", + "content": "of a bar:" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 238, + 234, + 295, + 293 + ], + "blocks": [ + { + "bbox": [ + 238, + 234, + 295, + 293 + ], + "lines": [ + { + "bbox": [ + 238, + 234, + 295, + 293 + ], + "spans": [ + { + "bbox": [ + 238, + 234, + 295, + 293 + ], + "type": "image", + "image_path": "6dc44c9b58ecc010a2291884d57a229b439b69ac24a25f3fe8be93bd8dabb06d.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 238, + 293, + 294, + 300 + ], + "lines": [ + { + "bbox": [ + 238, + 293, + 294, + 300 + ], + "spans": [ + { + "bbox": [ + 238, + 293, + 294, + 300 + ], + "type": "text", + "content": "" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 57, + 303, + 115, + 361 + ], + "blocks": [ + { + "bbox": [ + 57, + 303, + 115, + 361 + ], + "lines": [ + { + "bbox": [ + 57, + 303, + 115, + 361 + ], + "spans": [ + { + "bbox": [ + 57, + 303, + 115, + 361 + ], + "type": "image", + "image_path": "39f889f61f5d4b8f7638cc69f76b58acc3258807476ddd9b76ee01f6715697cf.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 57, + 361, + 194, + 369 + ], + "lines": [ + { + "bbox": [ + 57, + 361, + 194, + 369 + ], + "spans": [ + { + "bbox": [ + 57, + 361, + 194, + 369 + ], + "type": "text", + "content": "\"Gray modern house, photograph, highly detailed." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 118, + 303, + 176, + 361 + ], + "blocks": [ + { + "bbox": [ + 118, + 303, + 176, + 361 + ], + "lines": [ + { + "bbox": [ + 118, + 303, + 176, + 361 + ], + "spans": [ + { + "bbox": [ + 118, + 303, + 176, + 361 + ], + "type": "image", + "image_path": "f36b925e4ea021276383c4da6f0c1395f7d57a694993589b763ec5939c6fa83f.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 118, + 361, + 194, + 369 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 361, + 194, + 369 + ], + "spans": [ + { + "bbox": [ + 118, + 361, + 194, + 369 + ], + "type": "text", + "content": "photograph, highly detailed." + } + ] + } + ], + "index": 11, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 177, + 303, + 234, + 361 + ], + "blocks": [ + { + "bbox": [ + 177, + 303, + 234, + 361 + ], + "lines": [ + { + "bbox": [ + 177, + 303, + 234, + 361 + ], + "spans": [ + { + "bbox": [ + 177, + 303, + 234, + 361 + ], + "type": "image", + "image_path": "7a47f65f1c0dc17a9c09d117ddaa62b337fa887e3a97d5ab43cdb3c86c3ed3d2.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 177, + 361, + 205, + 369 + ], + "lines": [ + { + "bbox": [ + 177, + 361, + 205, + 369 + ], + "spans": [ + { + "bbox": [ + 177, + 361, + 205, + 369 + ], + "type": "text", + "content": "tailed." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 237, + 303, + 294, + 361 + ], + "blocks": [ + { + "bbox": [ + 237, + 303, + 294, + 361 + ], + "lines": [ + { + "bbox": [ + 237, + 303, + 294, + 361 + ], + "spans": [ + { + "bbox": [ + 237, + 303, + 294, + 361 + ], + "type": "image", + "image_path": "59069fe124565ac81a834c0ac606fba67550fa011f685fd79892cef13ea44269.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 237, + 361, + 294, + 369 + ], + "lines": [ + { + "bbox": [ + 237, + 361, + 294, + 369 + ], + "spans": [ + { + "bbox": [ + 237, + 361, + 294, + 369 + ], + "type": "text", + "content": "" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 58, + 371, + 115, + 429 + ], + "blocks": [ + { + "bbox": [ + 58, + 371, + 115, + 429 + ], + "lines": [ + { + "bbox": [ + 58, + 371, + 115, + 429 + ], + "spans": [ + { + "bbox": [ + 58, + 371, + 115, + 429 + ], + "type": "image", + "image_path": "76562308239e8646d2b5d78dcc8e42dda1bb757afbe4de247708d78c75d466f1.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 57, + 430, + 169, + 437 + ], + "lines": [ + { + "bbox": [ + 57, + 430, + 169, + 437 + ], + "spans": [ + { + "bbox": [ + 57, + 430, + 169, + 437 + ], + "type": "text", + "content": "\"An animated owl looking at the camera." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 118, + 371, + 176, + 429 + ], + "blocks": [ + { + "bbox": [ + 118, + 371, + 176, + 429 + ], + "lines": [ + { + "bbox": [ + 118, + 371, + 176, + 429 + ], + "spans": [ + { + "bbox": [ + 118, + 371, + 176, + 429 + ], + "type": "image", + "image_path": "d390afd7ad3750eddf15560afd5488e38a1a6d29f80b26844ccc03ba7895f4d2.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 118, + 430, + 170, + 437 + ], + "lines": [ + { + "bbox": [ + 118, + 430, + 170, + 437 + ], + "spans": [ + { + "bbox": [ + 118, + 430, + 170, + 437 + ], + "type": "text", + "content": "king at the camera." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 177, + 371, + 234, + 429 + ], + "blocks": [ + { + "bbox": [ + 177, + 371, + 234, + 429 + ], + "lines": [ + { + "bbox": [ + 177, + 371, + 234, + 429 + ], + "spans": [ + { + "bbox": [ + 177, + 371, + 234, + 429 + ], + "type": "image", + "image_path": "274de5eaa47b6b333bad1c8e8f0a6e870568274c42a499a86e96813351a26fe2.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 177, + 430, + 234, + 437 + ], + "lines": [ + { + "bbox": [ + 177, + 430, + 234, + 437 + ], + "spans": [ + { + "bbox": [ + 177, + 430, + 234, + 437 + ], + "type": "inline_equation", + "content": "\\because {AD} = {AQ},\\angle {PAQ} = {60}^{ \\circ }\\therefore {\\Delta PQA}" + }, + { + "bbox": [ + 177, + 430, + 234, + 437 + ], + "type": "text", + "content": " 为等边编形" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 237, + 371, + 294, + 429 + ], + "blocks": [ + { + "bbox": [ + 237, + 371, + 294, + 429 + ], + "lines": [ + { + "bbox": [ + 237, + 371, + 294, + 429 + ], + "spans": [ + { + "bbox": [ + 237, + 371, + 294, + 429 + ], + "type": "image", + "image_path": "cc65c13d277e9ea68905a1827188e40a0b1a26c22d56782f69ae34435bd7199c.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 237, + 430, + 294, + 437 + ], + "lines": [ + { + "bbox": [ + 237, + 430, + 294, + 437 + ], + "spans": [ + { + "bbox": [ + 237, + 430, + 294, + 437 + ], + "type": "text", + "content": "" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 58, + 440, + 115, + 498 + ], + "blocks": [ + { + "bbox": [ + 58, + 440, + 115, + 498 + ], + "lines": [ + { + "bbox": [ + 58, + 440, + 115, + 498 + ], + "spans": [ + { + "bbox": [ + 58, + 440, + 115, + 498 + ], + "type": "image", + "image_path": "1d995c919204a04b802081d68b7fe699b7fa05a56b5944bc9866fb6c096f8bbd.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 58, + 498, + 209, + 506 + ], + "lines": [ + { + "bbox": [ + 58, + 498, + 209, + 506 + ], + "spans": [ + { + "bbox": [ + 58, + 498, + 209, + 506 + ], + "type": "text", + "content": "\"An artistic flat design poster featuring sushi on a plate." + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 118, + 440, + 176, + 498 + ], + "blocks": [ + { + "bbox": [ + 118, + 440, + 176, + 498 + ], + "lines": [ + { + "bbox": [ + 118, + 440, + 176, + 498 + ], + "spans": [ + { + "bbox": [ + 118, + 440, + 176, + 498 + ], + "type": "image", + "image_path": "9f4415aee014082399b309a3357adcbaae9b40c90e3c1a50af89c91599bcdeb9.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "bbox": [ + 118, + 498, + 170, + 506 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 498, + 170, + 506 + ], + "spans": [ + { + "bbox": [ + 118, + 498, + 170, + 506 + ], + "type": "text", + "content": "n poster featuring su" + } + ] + } + ], + "index": 27, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 177, + 440, + 234, + 498 + ], + "blocks": [ + { + "bbox": [ + 177, + 440, + 234, + 498 + ], + "lines": [ + { + "bbox": [ + 177, + 440, + 234, + 498 + ], + "spans": [ + { + "bbox": [ + 177, + 440, + 234, + 498 + ], + "type": "image", + "image_path": "41ea7c2a5cdce56be20445b35bcd686de828e4963016efce7e7d389722e8354a.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 177, + 498, + 234, + 506 + ], + "lines": [ + { + "bbox": [ + 177, + 498, + 234, + 506 + ], + "spans": [ + { + "bbox": [ + 177, + 498, + 234, + 506 + ], + "type": "text", + "content": "ni on a plate." + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 237, + 440, + 294, + 498 + ], + "blocks": [ + { + "bbox": [ + 237, + 440, + 294, + 498 + ], + "lines": [ + { + "bbox": [ + 237, + 440, + 294, + 498 + ], + "spans": [ + { + "bbox": [ + 237, + 440, + 294, + 498 + ], + "type": "image", + "image_path": "ea342a221aa9de8de74a6125ad4a9224a380bcba81a71f68ae23013e4aa89621.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + } + ], + "index": 30 + }, + { + "bbox": [ + 237, + 508, + 294, + 567 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 237, + 508, + 294, + 567 + ], + "spans": [ + { + "bbox": [ + 237, + 508, + 294, + 567 + ], + "type": "image", + "image_path": "a46f3e58a52dd25e6c51232dd516f82a9d438f6cb6086fcd7b470ac6c31acd2b.jpg" + } + ] + } + ], + "index": 31, + "type": "text" + }, + { + "bbox": [ + 244, + 575, + 287, + 584 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 244, + 575, + 287, + 584 + ], + "spans": [ + { + "bbox": [ + 244, + 575, + 287, + 584 + ], + "type": "text", + "content": "(c) Fine-tune" + } + ] + } + ], + "index": 32, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 58, + 508, + 115, + 566 + ], + "blocks": [ + { + "bbox": [ + 58, + 508, + 115, + 566 + ], + "lines": [ + { + "bbox": [ + 58, + 508, + 115, + 566 + ], + "spans": [ + { + "bbox": [ + 58, + 508, + 115, + 566 + ], + "type": "image", + "image_path": "b9d11ffd8ec5106340840ed8a63fbac74871e7f9beb57d89a12b4bf32e64a5da.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 57, + 567, + 107, + 574 + ], + "lines": [ + { + "bbox": [ + 57, + 567, + 107, + 574 + ], + "spans": [ + { + "bbox": [ + 57, + 567, + 107, + 574 + ], + "type": "text", + "content": "\"Thundery sky.\"" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 63, + 576, + 107, + 584 + ], + "lines": [ + { + "bbox": [ + 63, + 576, + 107, + 584 + ], + "spans": [ + { + "bbox": [ + 63, + 576, + 107, + 584 + ], + "type": "text", + "content": "(a) Condition" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_caption" + } + ], + "index": 33 + }, + { + "type": "image", + "bbox": [ + 118, + 508, + 176, + 567 + ], + "blocks": [ + { + "bbox": [ + 118, + 508, + 176, + 567 + ], + "lines": [ + { + "bbox": [ + 118, + 508, + 176, + 567 + ], + "spans": [ + { + "bbox": [ + 118, + 508, + 176, + 567 + ], + "type": "image", + "image_path": "d872e269915874e5e2043056acbdab71d685bf440b2834f2bdf306c680f3485e.jpg" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 153, + 575, + 197, + 584 + ], + "lines": [ + { + "bbox": [ + 153, + 575, + 197, + 584 + ], + "spans": [ + { + "bbox": [ + 153, + 575, + 197, + 584 + ], + "type": "text", + "content": "(b) Zero-shot" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 594, + 295, + 628 + ], + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 628 + ], + "type": "text", + "content": "Figure 6. Qualitative results of our zero-shot approach. The first four rows show results from manual color conditions, while the fifth row presents results from in-the-wild imagery condition." + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_caption" + } + ], + "index": 36 + }, + { + "type": "image", + "bbox": [ + 177, + 508, + 234, + 567 + ], + "blocks": [ + { + "bbox": [ + 177, + 508, + 234, + 567 + ], + "lines": [ + { + "bbox": [ + 177, + 508, + 234, + 567 + ], + "spans": [ + { + "bbox": [ + 177, + 508, + 234, + 567 + ], + "type": "image", + "image_path": "e0938ebcd9865118665b929cc54306f6a28fea8f5ceea9ea8b3d330d0d69a847.jpg" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_body" + } + ], + "index": 38 + }, + { + "bbox": [ + 177, + 575, + 197, + 584 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 177, + 575, + 197, + 584 + ], + "spans": [ + { + "bbox": [ + 177, + 575, + 197, + 584 + ], + "type": "text", + "content": "o-shot" + } + ] + } + ], + "index": 39, + "type": "text" + }, + { + "bbox": [ + 237, + 508, + 294, + 567 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 237, + 508, + 294, + 567 + ], + "spans": [ + { + "bbox": [ + 237, + 508, + 294, + 567 + ], + "type": "image", + "image_path": "a46f3e58a52dd25e6c51232dd516f82a9d438f6cb6086fcd7b470ac6c31acd2b.jpg" + } + ] + } + ], + "index": 40, + "type": "text" + }, + { + "bbox": [ + 244, + 575, + 287, + 584 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 244, + 575, + 287, + 584 + ], + "spans": [ + { + "bbox": [ + 244, + 575, + 287, + 584 + ], + "type": "text", + "content": "(c) Fine-tune" + } + ] + } + ], + "index": 41, + "type": "text" + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "content": "We report the performance scores for the latent diffusion case study in Tab. 2. As shown, our method generally outperforms the baselines in most of the metrics. The baselines exhibit significant failures in metrics reflecting the disentanglement (CLIP-score), accuracy and completeness (CD-A" + } + ] + } + ], + "index": 43 + }, + { + "type": "table", + "bbox": [ + 315, + 231, + 555, + 299 + ], + "blocks": [ + { + "bbox": [ + 55, + 189, + 555, + 213 + ], + "lines": [ + { + "bbox": [ + 55, + 189, + 555, + 213 + ], + "spans": [ + { + "bbox": [ + 55, + 189, + 555, + 213 + ], + "type": "text", + "content": "Table 2. Quantitative results of latent diffusion. In each metric column, the left and right numbers correspond to the scores from the settings of manually sampled color condition and in-the-wild image color condition, respectively." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 231, + 555, + 299 + ], + "lines": [ + { + "bbox": [ + 315, + 231, + 555, + 299 + ], + "spans": [ + { + "bbox": [ + 315, + 231, + 555, + 299 + ], + "type": "table", + "html": "
MappingConditionFID ↓CD-A ↓CD-C ↓
XX63.9; 14940.6; 42.442.6; 39.0
X✓57.9; 88.115.4; 10.618.4; 15.2
✓X59.8; 45.70.00; 0.005.10; 4.62
✓✓57.5; 45.60.00; 0.004.12; 3.65
", + "image_path": "dd591665348ea62800200938301f17c91a642ce8638599f00c7df177920d99ac.jpg" + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "table_body" + } + ], + "index": 44 + }, + { + "bbox": [ + 313, + 307, + 555, + 341 + ], + "lines": [ + { + "bbox": [ + 313, + 307, + 555, + 341 + ], + "spans": [ + { + "bbox": [ + 313, + 307, + 555, + 341 + ], + "type": "text", + "content": "Table 3. Ablation study on color mapping and color condition. For each metric, the left and right numbers are from the Oxford-flower [32] and Microsoft-emoji [8] datasets, respectively." + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 361, + 555, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 361, + 555, + 456 + ], + "spans": [ + { + "bbox": [ + 313, + 361, + 555, + 456 + ], + "type": "text", + "content": "and CD-C scores) of generated colors in synthesized images. When provided with manual color conditions, our zero-shot version achieves worse FID score than the fine-tuned version does. This aligns with our observations in Sec. 4.4 indicating imagery details lacking caused by the zero-shot version. The zero-shot setting's very low CD scores also indicate color overfitting caused by mechanically mapping the manual color conditions." + } + ] + } + ], + "index": 46 + }, + { + "bbox": [ + 313, + 457, + 556, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 457, + 556, + 554 + ], + "spans": [ + { + "bbox": [ + 313, + 457, + 556, + 554 + ], + "type": "text", + "content": "Ablation study. We validate the effectiveness of important components in our method in Tab. 3. Specifically, we investigate the role of the color mappings " + }, + { + "bbox": [ + 313, + 457, + 556, + 554 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 313, + 457, + 556, + 554 + ], + "type": "text", + "content": ", and the inputting of the color condition " + }, + { + "bbox": [ + 313, + 457, + 556, + 554 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 457, + 556, + 554 + ], + "type": "text", + "content": " (those in Fig. 2c). When mappings are not used, noisy samples are not altered and only color conditions are fed to the diffusion model. As shown in Tab. 3, the mappings and color conditions significantly boost up the performance of our method." + } + ] + } + ], + "index": 47 + }, + { + "bbox": [ + 313, + 562, + 466, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 562, + 466, + 574 + ], + "spans": [ + { + "bbox": [ + 313, + 562, + 466, + 574 + ], + "type": "text", + "content": "5. Conclusion and Discussion" + } + ] + } + ], + "index": 48 + }, + { + "bbox": [ + 313, + 582, + 555, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 555, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 555, + 640 + ], + "type": "text", + "content": "We propose a color alignment method that operates on the diffusion process of diffusion models. Diffused colors are mapped to conditional colors across diffusion steps, enabling a continuous pathway for synthesizing target color patterns while maintaining the creativity of diffusion models." + } + ] + } + ], + "index": 49 + }, + { + "bbox": [ + 313, + 642, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 556, + 713 + ], + "type": "text", + "content": "There are issues that would be addressed in our future work. First, while our method can condition on color values and proportions, it is worthwhile to understand the influence of the spatial information in the color conditions. Second, it is of great interest to extend our color alignment to video and 3D data, potentially with integration of multi-view attributes." + } + ] + } + ], + "index": 50 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28453" + } + ] + } + ], + "index": 51 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 297, + 127 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 297, + 127 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 297, + 127 + ], + "type": "text", + "content": "Binh-Son Hua is supported by Research Ireland under the Research Ireland Frontiers for the Future Programme - Project, award number 22/FFP-P/11522." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 140, + 115, + 153 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 140, + 115, + 153 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 115, + 153 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 160, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 160, + 296, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 160, + 296, + 193 + ], + "spans": [ + { + "bbox": [ + 61, + 160, + 296, + 193 + ], + "type": "text", + "content": "[1] Text-to-image-2m: A high-quality, diverse text-to-image training dataset. https://huggingface.co/datasets/jackyhate/text-to-image-2M.4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 194, + 296, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 194, + 296, + 239 + ], + "spans": [ + { + "bbox": [ + 61, + 194, + 296, + 239 + ], + "type": "text", + "content": "[2] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 240, + 296, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 240, + 296, + 283 + ], + "spans": [ + { + "bbox": [ + 61, + 240, + 296, + 283 + ], + "type": "text", + "content": "[3] Vassilis Athitsos and Stan Sclaroff. Estimating 3d hand pose from a cluttered image. In 2003 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, 2003. Proceedings., pages II-432. IEEE, 2003. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 285, + 296, + 340 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 285, + 296, + 340 + ], + "spans": [ + { + "bbox": [ + 62, + 285, + 296, + 340 + ], + "type": "text", + "content": "[4] Arpit Bansal, Eitan Borgnia, Hong-Min Chu, Jie Li, Hamid Kazemi, Furong Huang, Micah Goldblum, Jonas Geiping, and Tom Goldstein. Cold diffusion: Inverting arbitrary image transforms without noise. Advances in Neural Information Processing Systems, 36, 2024. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 342, + 296, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 342, + 296, + 396 + ], + "spans": [ + { + "bbox": [ + 62, + 342, + 296, + 396 + ], + "type": "text", + "content": "[5] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18392-18402, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 399, + 296, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 399, + 296, + 463 + ], + "spans": [ + { + "bbox": [ + 62, + 399, + 296, + 463 + ], + "type": "text", + "content": "[6] Mingdeng Cao, Xintao Wang, Zhongang Qi, Ying Shan, Xiaohu Qie, and Yinqiang Zheng. Masactrl: Tuning-free mutual self-attention control for consistent image synthesis and editing. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22560-22570, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 466, + 296, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 466, + 296, + 520 + ], + "spans": [ + { + "bbox": [ + 62, + 466, + 296, + 520 + ], + "type": "text", + "content": "[7] Jiwoo Chung, Sangeek Hyun, and Jae-Pil Heo. Style injection in diffusion: A training-free approach for adapting large-scale diffusion models for style transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8795-8805, 2024. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 522, + 296, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 522, + 296, + 555 + ], + "spans": [ + { + "bbox": [ + 62, + 522, + 296, + 555 + ], + "type": "text", + "content": "[8] Jason Custer and Spencer Nelson. Microsoft fluent emoji. https://github.com/microsoft/fluentui-emoji.4,7,8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 557, + 296, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 557, + 296, + 622 + ], + "spans": [ + { + "bbox": [ + 62, + 557, + 296, + 622 + ], + "type": "text", + "content": "[9] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 624, + 296, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 624, + 296, + 689 + ], + "spans": [ + { + "bbox": [ + 57, + 624, + 296, + 689 + ], + "type": "text", + "content": "[10] Ziv Epstein, Aaron Hertzmann, Laura Herman, Robert Mahari, Morgan R. Frank, Matthew Groh, Hope Schroeder, Amy Smith, Memo Akten, Jessica Fjeld, Hany Farid, Neil Leach, Alex \"Sandy\" Pentland, and Olga Russakovsk. Art and the science of generative ai. Science, 380(6650):1110-1111, 2023. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 691, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 296, + 713 + ], + "type": "text", + "content": "[11] Martin Nicolas Everaert, Marco Bocchio, Sami Arpa, Sabine Susstrunk, and Radhakrishna Achanta. Diffusion in style. In" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 335, + 73, + 554, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 554, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 554, + 95 + ], + "type": "text", + "content": "Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2251-2261, 2023. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 97, + 555, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 97, + 555, + 140 + ], + "spans": [ + { + "bbox": [ + 316, + 97, + 555, + 140 + ], + "type": "text", + "content": "[12] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139-144, 2020. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 141, + 555, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 555, + 185 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 555, + 185 + ], + "type": "text", + "content": "[13] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 186, + 555, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 186, + 555, + 240 + ], + "spans": [ + { + "bbox": [ + 317, + 186, + 555, + 240 + ], + "type": "text", + "content": "[14] Amir Hertz, Andrey Voynov, Shlomi Fruchter, and Daniel Cohen-Or. Style aligned image generation via shared attention. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4775–4785, 2024. 1, 2, 3, 4, 5, 6, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 241, + 555, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 241, + 555, + 284 + ], + "spans": [ + { + "bbox": [ + 316, + 241, + 555, + 284 + ], + "type": "text", + "content": "[15] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning. arXiv preprint arXiv:2104.08718, 2021. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 286, + 555, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 286, + 555, + 341 + ], + "spans": [ + { + "bbox": [ + 316, + 286, + 555, + 341 + ], + "type": "text", + "content": "[16] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 342, + 554, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 342, + 554, + 365 + ], + "spans": [ + { + "bbox": [ + 316, + 342, + 554, + 365 + ], + "type": "text", + "content": "[17] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 366, + 555, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 366, + 555, + 399 + ], + "spans": [ + { + "bbox": [ + 316, + 366, + 555, + 399 + ], + "type": "text", + "content": "[18] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1, 2, 3, 4, 5, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 400, + 555, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 400, + 555, + 443 + ], + "spans": [ + { + "bbox": [ + 317, + 400, + 555, + 443 + ], + "type": "text", + "content": "[19] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1125-1134, 2017. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 445, + 555, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 445, + 555, + 488 + ], + "spans": [ + { + "bbox": [ + 317, + 445, + 555, + 488 + ], + "type": "text", + "content": "[20] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 490, + 555, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 490, + 555, + 544 + ], + "spans": [ + { + "bbox": [ + 317, + 490, + 555, + 544 + ], + "type": "text", + "content": "[21] Bahjat Kawar, Shiran Zada, Oran Lang, Omer Tov, Huiwen Chang, Tali Dekel, Inbar Mosseri, and Michal Irani. Imag: Text-based real image editing with diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6007-6017, 2023. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 546, + 555, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 546, + 555, + 568 + ], + "spans": [ + { + "bbox": [ + 316, + 546, + 555, + 568 + ], + "type": "text", + "content": "[22] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 569, + 555, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 569, + 555, + 624 + ], + "spans": [ + { + "bbox": [ + 317, + 569, + 555, + 624 + ], + "type": "text", + "content": "[23] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 625, + 555, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 625, + 555, + 689 + ], + "spans": [ + { + "bbox": [ + 317, + 625, + 555, + 689 + ], + "type": "text", + "content": "[24] Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11461-11471, 2022. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 691, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 554, + 713 + ], + "type": "text", + "content": "[25] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jiajun Wu, Jun-Yan Zhu, and Stefano Ermon. Sdedit: Guided image" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "28454" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 77, + 73, + 296, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 73, + 296, + 95 + ], + "spans": [ + { + "bbox": [ + 77, + 73, + 296, + 95 + ], + "type": "text", + "content": "synthesis and editing with stochastic differential equations. arXiv preprint arXiv:2108.01073, 2021. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 96, + 295, + 118 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 96, + 295, + 118 + ], + "spans": [ + { + "bbox": [ + 57, + 96, + 295, + 118 + ], + "type": "text", + "content": "[26] Mehdi Mirza and Simon Osindero. Conditional generative adversarial nets. arXiv preprint arXiv:1411.1784, 2014. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 119, + 296, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 119, + 296, + 174 + ], + "spans": [ + { + "bbox": [ + 56, + 119, + 296, + 174 + ], + "type": "text", + "content": "[27] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6038–6047, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 175, + 296, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 175, + 296, + 218 + ], + "spans": [ + { + "bbox": [ + 57, + 175, + 296, + 218 + ], + "type": "text", + "content": "[28] Chong Mou, Xintao Wang, Jiechong Song, Ying Shan, and Jian Zhang. *Dragondiffusion: Enabling drag-style manipulation on diffusion models.* arXiv preprint arXiv:2307.02421, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 57, + 220, + 296, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 220, + 296, + 275 + ], + "spans": [ + { + "bbox": [ + 57, + 220, + 296, + 275 + ], + "type": "text", + "content": "[29] Chong Mou, Xintao Wang, Jiechong Song, Ying Shan, and Jian Zhang. Diffeditor: Boosting accuracy and flexibility on diffusion-based image editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8488-8497, 2024. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 277, + 296, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 277, + 296, + 342 + ], + "spans": [ + { + "bbox": [ + 56, + 277, + 296, + 342 + ], + "type": "text", + "content": "[30] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, and Ying Shan. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 4296–4304, 2024. 1, 2, 4, 5, 6, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 343, + 296, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 343, + 296, + 387 + ], + "spans": [ + { + "bbox": [ + 56, + 343, + 296, + 387 + ], + "type": "text", + "content": "[31] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International conference on machine learning, pages 8162-8171. PMLR, 2021. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 388, + 296, + 433 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 388, + 296, + 433 + ], + "spans": [ + { + "bbox": [ + 56, + 388, + 296, + 433 + ], + "type": "text", + "content": "[32] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In 2008 Sixth Indian conference on computer vision, graphics & image processing, pages 722-729. IEEE, 2008. 4, 7, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 434, + 296, + 478 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 434, + 296, + 478 + ], + "spans": [ + { + "bbox": [ + 56, + 434, + 296, + 478 + ], + "type": "text", + "content": "[33] Gaurav Parmar, Krishna Kumar Singh, Richard Zhang, Yijun Li, Jingwan Lu, and Jun-Yan Zhu. Zero-shot image-to-image translation. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-11, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 479, + 296, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 479, + 296, + 544 + ], + "spans": [ + { + "bbox": [ + 56, + 479, + 296, + 544 + ], + "type": "text", + "content": "[34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 546, + 296, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 546, + 296, + 589 + ], + "spans": [ + { + "bbox": [ + 56, + 546, + 296, + 589 + ], + "type": "text", + "content": "[35] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1(2): 3, 2022. 1, 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 591, + 296, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 591, + 296, + 646 + ], + "spans": [ + { + "bbox": [ + 56, + 591, + 296, + 646 + ], + "type": "text", + "content": "[36] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 2, 3, 4, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 647, + 296, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 296, + 712 + ], + "type": "text", + "content": "[37] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22500-22510, 2023. 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 685 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 316, + 73, + 555, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 555, + 126 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 555, + 126 + ], + "type": "text", + "content": "[38] Chitwan Sahara, William Chan, Huiwen Chang, Chris Lee, Jonathan Ho, Tim Salimans, David Fleet, and Mohammad Norouzi. Palette: Image-to-image diffusion models. In ACM SIGGRAPH 2022 conference proceedings, pages 1-10, 2022. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 129, + 555, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 555, + 194 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 555, + 194 + ], + "type": "text", + "content": "[39] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural information processing systems, 35:36479-36494, 2022. 1, 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 196, + 555, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 196, + 555, + 261 + ], + "spans": [ + { + "bbox": [ + 316, + 196, + 555, + 261 + ], + "type": "text", + "content": "[40] Yujun Shi, Chuhui Xue, Jun Hao Liew, Jiachun Pan, Hanshu Yan, Wenqing Zhang, Vincent YF Tan, and Song Bai. Dragdiffusion: Harnessing diffusion models for interactive point-based image editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8839-8849, 2024. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 262, + 555, + 305 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 262, + 555, + 305 + ], + "spans": [ + { + "bbox": [ + 316, + 262, + 555, + 305 + ], + "type": "text", + "content": "[41] Kihyuk Sohn, Honglak Lee, and Xinchen Yan. Learning structured output representation using deep conditional generative models. Advances in neural information processing systems, 28, 2015. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 307, + 555, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 307, + 555, + 338 + ], + "spans": [ + { + "bbox": [ + 316, + 307, + 555, + 338 + ], + "type": "text", + "content": "[42] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020.2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 341, + 555, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 341, + 555, + 373 + ], + "spans": [ + { + "bbox": [ + 316, + 341, + 555, + 373 + ], + "type": "text", + "content": "[43] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 375, + 555, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 375, + 555, + 418 + ], + "spans": [ + { + "bbox": [ + 316, + 375, + 555, + 418 + ], + "type": "text", + "content": "[44] Patrick von Platen, Suraj Patil, Anton Lozhkov, Pedro Cuenca, Nathan Lambert, Kashif Rasul, Mishig Davaadorj, Dhruv Nair, Sayak Paul, Steven Liu, William Berman, Yiyi Xu, and Thomas Wolf. Diffusers: State-of-the-art diffusion models. 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 420, + 555, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 420, + 555, + 474 + ], + "spans": [ + { + "bbox": [ + 316, + 420, + 555, + 474 + ], + "type": "text", + "content": "[45] Shaoan Xie, Zhifei Zhang, Zhe Lin, Tobias Hinz, and Kun Zhang. Smartbrush: Text and shape guided object inpainting with diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22428-22437, 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 475, + 555, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 475, + 555, + 518 + ], + "spans": [ + { + "bbox": [ + 316, + 475, + 555, + 518 + ], + "type": "text", + "content": "[46] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023. 1, 2, 3, 4, 5, 6, 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 520, + 555, + 574 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 520, + 555, + 574 + ], + "spans": [ + { + "bbox": [ + 316, + 520, + 555, + 574 + ], + "type": "text", + "content": "[47] Jiwen Yu, Yinhuai Wang, Chen Zhao, Bernard Ghanem, and Jian Zhang. Freedom: Training-free energy-guided conditional diffusion model. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 23174-23184, 2023. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 576, + 555, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 576, + 555, + 620 + ], + "spans": [ + { + "bbox": [ + 316, + 576, + 555, + 620 + ], + "type": "text", + "content": "[48] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 1, 2, 3, 4, 5, 6, 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 621, + 555, + 685 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 621, + 555, + 685 + ], + "spans": [ + { + "bbox": [ + 316, + 621, + 555, + 685 + ], + "type": "text", + "content": "[49] Yuxin Zhang, Nisha Huang, Fan Tang, Haibin Huang, Chongyang Ma, Weiming Dong, and Changsheng Xu. Inversion-based style transfer with diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10146-10156, 2023. 2" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "28455" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/ComRoPE_ Scalable and Robust Rotary Position Embedding Parameterized by Trainable Commuting Angle Matrices/287888de-80a8-43c2-8d0f-f61f04d80e4e_content_list.json b/2025/ComRoPE_ Scalable and Robust Rotary Position Embedding Parameterized by Trainable Commuting Angle Matrices/287888de-80a8-43c2-8d0f-f61f04d80e4e_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..bfa0f2e2ddb1722bcdb4cfe41f3c184fa4f38f90 --- /dev/null +++ b/2025/ComRoPE_ Scalable and Robust Rotary Position Embedding Parameterized by Trainable Commuting Angle Matrices/287888de-80a8-43c2-8d0f-f61f04d80e4e_content_list.json @@ -0,0 +1,1859 @@ +[ + { + "type": "text", + "text": "ComRoPE: Scalable and Robust Rotary Position Embedding Parameterized by Trainable Commuting Angle Matrices", + "text_level": 1, + "bbox": [ + 99, + 128, + 898, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hao $\\mathbf{Y}\\mathbf{u}^{1}$ Tangyu Jiang $^{1\\dagger}$ Shuning Jia $^{1,2}$ Shannon Yan $^{1}$ Shunning Liu $^{1}$", + "bbox": [ + 127, + 200, + 867, + 222 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Haolong Qian1 Guanghao Li1 Shuting Dong1 Chun Yuan1†", + "bbox": [ + 179, + 227, + 816, + 247 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Tsinghua University $^{2}$ Shenzhen University", + "bbox": [ + 315, + 255, + 681, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "longinyh@gmail.com, jiangtangyu@sz.tsinghua.edu.cn, yuanc@sz.tsinghua.edu.cn", + "bbox": [ + 153, + 277, + 826, + 292 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 318, + 326, + 334 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The Transformer architecture has revolutionized various fields since it was proposed, where positional encoding plays an essential role in effectively capturing sequential order and context. Therefore, Rotary Positional Encoding (RoPE) was proposed to alleviate these issues, which integrates positional information by rotating the embeddings in the attention mechanism. However, RoPE utilizes manually defined rotation matrices, a design choice that favors computational efficiency but limits the model's flexibility and adaptability. In this work, we propose ComRoPE, which generalizes RoPE by defining it in terms of trainable commuting angle matrices. Specifically, we demonstrate that pairwise commutativity of these matrices is essential for RoPE to achieve scalability and positional robustness. We formally define the RoPE Equation, which is an essential condition that ensures consistent performance with position offsets. Based on the theoretical analysis, we present two types of trainable commuting angle matrices as sufficient solutions to the RoPE equation, which significantly improve performance, surpassing the current state-of-the-art method by $1.6\\%$ at training resolution and $2.9\\%$ at higher resolution on the ImageNet-1K dataset. Furthermore, our framework shows versatility in generalizing to existing RoPE formulations and offering new insights for future positional encoding research. To ensure reproducibility, the source code and instructions are available at https://github.com/Longin-Yu/ComRoPE.", + "bbox": [ + 88, + 349, + 485, + 760 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 766, + 220, + 781 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The Transformer architecture [30] has been widely adopted across various fields, including Natural Language Processing (NLP) [1, 5, 8, 22, 23] and Computer Vision (CV) [6]. Moreover, an increasing number of Transformer-based applications [2, 12, 13, 15, 16, 33-36] continue to demonstrate", + "bbox": [ + 89, + 791, + 483, + 868 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "their effectiveness across a wide range of domains. At the core of the Transformer model lies the attention mechanism, which enables the model to selectively focus on different parts of the input based on relevance, thereby effectively capturing long-range dependencies and contextual relationships.", + "bbox": [ + 511, + 330, + 906, + 421 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "However, since the attention mechanism is insensitive to the fundamental position information, it cannot inherently capture the order of elements in the data. In order to alleviate this issue, positional embeddings are added to the input representation, providing the model with the necessary positional information to process the sequence order and structure effectively.", + "bbox": [ + 511, + 424, + 908, + 530 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "A variety of positional encoding methods have been proposed, which can be divided into two categories: Absolute Positional Encoding (APE) and Relative Positional Encoding (RPE). APE explicitly encodes the absolute position of each token by generating fixed positional embeddings, which are directly added to the input embeddings at the beginning of training [30]. However, APE struggles with handling long sequences and exhibits high sensitivity to positional shifts, limiting the scalability and robustness of models [11]. In contrast, RPE does not directly modify input embeddings; instead, it incorporates relative positional information within the attention mechanism, enabling the model to effectively capture positional relationships between tokens. Among all the existing works, Rotary Position Embedding (RoPE) [27] has gained significant attention due to the advantage of applying a rotational transformation to token embeddings. It encodes relative positions by treating each pair of features as coordinates and rotating them by an angle proportional to their position, enabling all tokens to interact within the attention mechanism regardless of distance.", + "bbox": [ + 509, + 534, + 908, + 851 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "However, existing RoPE methods face several key challenges: i) The essential components (i.e., the RoPE matrices) of previous RoPE approaches rely on 2D rotation", + "bbox": [ + 511, + 854, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "†Corresponding author. This work was done when Shuning Jia was an intern at Tsinghua University.", + "bbox": [ + 89, + 875, + 482, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "4508", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "groups, which simplify computations but consequently restrict their feature projection capabilities, especially in high-dimensional spaces [27]. ii) Moreover, the majority of the rotation matrices of RoPE require to be manually designed, leading to insufficient capability and suboptimal performance[21, 25]. iii) Finally, previous attempts [20] to extend the rotation group often prioritize design simplicity, making it difficult to consistently satisfy relative position dependency—a critical property of RoPE that ensures positional robustness against offsets.", + "bbox": [ + 89, + 90, + 480, + 242 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our objectives. This work aims to develop a novel RoPE method for Transformers that is both scalable and robust. Specifically, we seek to extend the rotation group from the existing 2D representation to a larger subgroup of the special orthogonal group, allowing for higher degrees of freedom while preserving consistent behavior with respect to position offsets. Unlike existing methods that rely on manually designed non-trainable rotation matrices, which suffer from limited expressiveness and reduced robustness, our approach is designed to offer richer feature representation capabilities. This framework addresses the scalability limitations of current approaches and enhances their robustness against positional transformations.", + "bbox": [ + 89, + 253, + 480, + 450 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions. This work introduces ComRoPE, a novel framework that significantly enhances positional encoding in Transformers. ComRoPE leverages trainable angle matrices, extending the RoPE mechanism with higher scalability and robustness. We identify the pairwise commutativity of these matrices as a necessary and sufficient condition for effective positional encoding, thereby unifying various existing RoPE formulations under a single theoretical framework. The contributions are summarized as follows:", + "bbox": [ + 89, + 460, + 480, + 609 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We formally define the RoPE function parameterized by angle matrices and prove that pairwise commutativity is a necessary and sufficient condition, offering a unified theory that encompasses several existing RoPE variants.", + "- We introduce ComRoPE, a scalable and robust solution that leverages two types of trainable commuting angle matrix sets as sufficient solutions to the RoPE equation, capturing richer positional representations without the need for manual design.", + "- Our extensive experiments show that ComRoPE surpasses the current state-of-the-art LieRE, achieving a performance increase of $1.6\\%$ at training resolution and $2.9\\%$ at higher resolutions on the ImageNet-1K classification task while delivering strong results across other benchmarks.", + "- We explore further applications of ComRoPE, providing valuable insights for advancing position encoding techniques in future Transformer-based models." + ], + "bbox": [ + 91, + 623, + 478, + 892 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 513, + 89, + 651, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Position information in attention", + "text_level": 1, + "bbox": [ + 513, + 117, + 799, + 132 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Transformers [30] utilize the attention mechanism to capture similarities within sequences. However, they lack inherent sequential information and cannot capture the positional information of each token. To address this limitation, positional encoding [5, 24, 26] was introduced. As research progressed, positional encoding generally evolved into two types: APE [5, 28, 30] and RPE [10, 18, 26, 27]. [30] first proposed using APE in the form of sine and cosine functions, effectively representing the positional relationships within the input sequence. This positional encoding method achieved remarkable results in natural language processing, becoming a foundational component for many NLP tasks. However, the fixed nature of positional encoding limited the model's ability to generalize to longer input sequences. In response, subsequent research introduced learnable absolute positional encoding. [5] proposed further enhancing the model's expressive power by incorporating learnable position embeddings, particularly excelling in tasks such as sentence alignment and context representation. Although APE provides positional information to enhance the model's understanding of sequences, it shows limitations in handling long sequences and cross-sequence scenarios. Hierarchical ViT such as Swin Transformer [17], introduced Relative Position Bias (RPB) [17, 25, 26] to handle large numbers of tokens with limited positional embeddings [9]. Related research has explored alternative encoding methods, such as RPE, as a replacement for APE to better capture complex dependency structures. iRPE [31] proposed an improved RPB by combining relative position embedding.", + "bbox": [ + 511, + 143, + 903, + 583 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Rotary position embedding", + "text_level": 1, + "bbox": [ + 513, + 602, + 759, + 618 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Building on these developments, RoFormer [27] combined the advantages of APE and RPE, proposing RoPE. Currently, RoPE is widely used in Large Language Models (LLMs), such as LLaMA [29] and Vicuna [3]. This approach enhances model performance on tasks involving long-text semantics and multi-turn dialogues, improving extrapolation capability. To better adapt to the characteristics of two-dimensional data such as images, researchers extended RoPE to two-dimensional sequences (2D-RoPE). LieRE [20] extends RoPE to a more generalized form by introducing a rotation-based positional encoding method grounded in Lie group theory. Unified-IO 2 [19] applies 2D-RoPE within its multimodal architecture; EVA-02 [7], FiT [7] these pioneering works used 2D RoPE with axial frequencies (2D Axial RoPE), but had limitations in processing in the diagonal direction. Therefore, RoPE for ViT [9] proposes to use mixed axial frequency for 2D RoPE, named RoPE-Mixed.", + "bbox": [ + 511, + 628, + 903, + 898 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "4509", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 89, + 181, + 104 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we begin by introducing key definitions and reformulating the RoPE paradigm within multi-axial attention mechanisms, which we collectively refer to as the RoPE Equations. Next, we present our main theorem, which establishes the necessary and sufficient conditions for RoPE functions parameterized by angle matrices. Finally, we provide solutions to the RoPE Equations based on the propositions that outline the sufficient conditions.", + "bbox": [ + 89, + 114, + 483, + 236 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminaries", + "text_level": 1, + "bbox": [ + 89, + 243, + 230, + 258 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We use $\\mathbf{R}(\\cdot):\\mathbb{R}^N\\to \\mathbb{R}^{d\\times d}$ to denote a matrix-value function. We denote the vectors and matrices as the lower and uppercase bold font, respectively. We first recall two fundamental definitions of matrices as follows:", + "bbox": [ + 89, + 266, + 483, + 327 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Definition 1 (Matrix Exponential). The exponential of a square matrix $\\mathbf{A}$ , denoted as $e^{\\mathbf{A}}$ or $\\exp (\\mathbf{A})$ , is defined using the matrix exponential series such that:", + "bbox": [ + 89, + 334, + 483, + 380 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\ne ^ {\\mathbf {A}} = \\exp (\\mathbf {A}) = \\sum_ {k = 0} ^ {\\infty} \\frac {\\mathbf {A} ^ {k}}{k !} = \\mathbf {I} + \\mathbf {A} + \\frac {\\mathbf {A} ^ {2}}{2 !} + \\frac {\\mathbf {A} ^ {3}}{3 !} + \\dots ,\n$$\n", + "text_format": "latex", + "bbox": [ + 98, + 387, + 475, + 426 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This series converges for any square matrix $\\mathbf{A}$ .", + "bbox": [ + 89, + 433, + 401, + 449 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Definition 2 (Commuting Matrices). Two square matrices $\\mathbf{A}_1$ and $\\mathbf{A}_2$ are said to commute if their product is independent of the order of multiplication, i.e.,", + "bbox": [ + 89, + 455, + 483, + 501 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {A} _ {1} \\mathbf {A} _ {2} = \\mathbf {A} _ {2} \\mathbf {A} _ {1}.\n$$\n", + "text_format": "latex", + "bbox": [ + 228, + 508, + 344, + 525 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "A set of square matrices $\\{\\mathbf{A}_1,\\mathbf{A}_2,\\dots ,\\mathbf{A}_N\\}$ is said to pairwise commute if every pair of matrices within the set commutes with each other. That is, for all $i,j$ such that $1\\leq i,j\\leq N$", + "bbox": [ + 89, + 532, + 483, + 593 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {A} _ {i} \\mathbf {A} _ {j} = \\mathbf {A} _ {j} \\mathbf {A} _ {i}.\n$$\n", + "text_format": "latex", + "bbox": [ + 230, + 594, + 341, + 609 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To clarify and better illustrate the main theorems presented in the following sections, we first reformulate and unify the definitions of the RoPE paradigm, specifically in the context of multi-axial attention mechanisms. RoPE was initially proposed by Su et al. [27] as a positional encoding method based on relative position dependencies. However, previous work provided only conceptual and descriptive descriptions of RPE and RoPE without offering a rigorous formal definition. In this work, we provide the formal definitions of both RPE and RoPE.", + "bbox": [ + 89, + 616, + 483, + 767 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Definition 3 (RPE Equation). Let $f: \\mathbb{R}^d \\times \\mathbb{R}^N \\to \\mathbb{R}^d$ be a positional encoding function. $f$ is said to be a RPE function if and only if there exists a function $\\rho: \\mathbb{R}^d \\times \\mathbb{R}^d \\to \\mathbb{R}$ , and a function $g: \\mathbb{R}^d \\times \\mathbb{R}^d \\times \\mathbb{R}^N \\to \\mathbb{R}$ such that the following conditions hold for all $x, y \\in \\mathbb{R}^N$ and $q, k \\in \\mathbb{R}^d$ :", + "bbox": [ + 86, + 773, + 483, + 851 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\ng (\\boldsymbol {q}, \\boldsymbol {k}, \\boldsymbol {x} - \\boldsymbol {y}) = \\rho (f (\\boldsymbol {q}, \\boldsymbol {x}), f (\\boldsymbol {k}, \\boldsymbol {y})), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 158, + 864, + 482, + 881 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We refer to Eq. (1) as the RPE Equation.", + "bbox": [ + 89, + 885, + 364, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Definition 4 (RoPE Equation). Let $f: \\mathbb{R}^d \\times \\mathbb{R}^N \\to \\mathbb{R}^d$ be an RPE function. $f$ is said to be a RoPE function if and only if there exists a matrix-valued function $\\mathbf{R}_f(\\cdot)$ such that the following conditions hold for all $\\mathbf{x}, \\mathbf{y} \\in \\mathbb{R}^N$ and $\\mathbf{k}, \\mathbf{k} \\in \\mathbb{R}^d$ in RPE Equation:", + "bbox": [ + 511, + 89, + 906, + 167 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{ \\begin{array}{l} f (\\boldsymbol {q}, \\boldsymbol {x}) = \\mathbf {R} _ {f} (\\boldsymbol {x}) \\boldsymbol {q} \\\\ \\rho (\\boldsymbol {q}, \\boldsymbol {k}) = \\boldsymbol {q} ^ {\\top} \\boldsymbol {k} \\\\ g (\\boldsymbol {q}, \\boldsymbol {k}, \\boldsymbol {x} - \\boldsymbol {y}) = \\boldsymbol {q} ^ {\\top} \\mathbf {R} _ {f} (\\boldsymbol {y} - \\boldsymbol {x}) \\boldsymbol {k} \\end{array} \\right. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 176, + 906, + 237 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We refer to Eq. (2) as the RoPE Equation.", + "bbox": [ + 513, + 246, + 795, + 262 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "By substituting the RoPE equation into the RPE equation, we obtain that the RoPE function satisfies the following property.", + "bbox": [ + 511, + 271, + 905, + 316 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Proposition 1. $f$ is said to be a RoPE function if and only if the matrix-valued function $\\mathbf{R}_f(\\cdot)$ satisfies:", + "bbox": [ + 511, + 325, + 905, + 358 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {R} _ {f} (\\boldsymbol {x}) ^ {\\top} \\mathbf {R} _ {f} (\\boldsymbol {y}) = \\mathbf {R} _ {f} (\\boldsymbol {y} - \\boldsymbol {x}),\n$$\n", + "text_format": "latex", + "bbox": [ + 602, + 367, + 813, + 386 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "for all $\\pmb {x},\\pmb {y}\\in \\mathbb{R}^{N}$", + "bbox": [ + 511, + 396, + 638, + 412 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Note that the definition of RoPE Equation demonstrates that the position encoding should only be dependent on the relative position of the tokens, which thus be robust against the offset operations. We further expand the definition of the RoPE function to a parameterized one (i.e., Definition 6) via rotation matrices (i.e., Definition 5) as follows1.", + "bbox": [ + 511, + 421, + 905, + 513 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Definition 5 (Parameterized Rotation Matrix). Let $\\mathbf{R}(\\cdot ;\\mathcal{A}): \\mathbb{R}^N \\to \\mathbb{R}^{d\\times d}$ be a matrix-valued function parameterized by $N$ skew-symmetric matrices $\\mathcal{A} = \\{\\mathbf{A}_1,\\mathbf{A}_2,\\dots ,\\mathbf{A}_N\\}$ . We say that $\\mathbf{R}(\\cdot ;\\mathcal{A})$ is a rotation matrix function parameterized by angle matrices $\\mathcal{A}$ if it can be expressed as:", + "bbox": [ + 511, + 522, + 906, + 614 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {R} (\\boldsymbol {x}; \\mathcal {A}) = \\exp \\left(\\sum_ {i = 1} ^ {N} \\mathbf {A} _ {i} x _ {i}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 604, + 623, + 905, + 666 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\pmb{x} = (x_{1}, x_{2}, \\ldots, x_{N}) \\in \\mathbb{R}^{N}$ and $\\exp(\\cdot)$ denotes the matrix exponential.", + "bbox": [ + 511, + 676, + 905, + 708 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Definition 6 (Parameterized RoPE Function). Let $f: \\mathbb{R}^d \\times \\mathbb{R}^N \\to \\mathbb{R}^d$ be a RoPE function. $f$ is said to be parameterized by angle matrices if and only if there exists a rotation matrix function $\\mathbf{R}_f(\\cdot; \\mathcal{A})$ parameterized by angle matrices $\\mathcal{A}$ such that the RoPE Equation (i.e., Eq. (2)) holds.", + "bbox": [ + 511, + 715, + 905, + 792 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this case, $\\mathbf{R}_f(\\cdot ;\\mathcal{A})$ is referred to as the rotation matrix function of RoPE function $f(\\cdot ;\\mathcal{A})$ parameterized by angle matrices $\\mathcal{A}$ . For simplicity, we slightly abuse the notation and refer to $\\mathbf{R}$ as the rotation matrix of RoPE function $f$ .", + "bbox": [ + 511, + 792, + 905, + 868 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "For clarification, the definitions of the rotation matrix and its exponential representation can be found in Appendix A.", + "bbox": [ + 511, + 875, + 905, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "4510", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/65b5ac4bfd70910ee3a5a462d8ee4798cac3086e6d8b276c4d6a36bb8a61650f.jpg", + "image_caption": [ + "Figure 1. Overview of ComRoPE. Features are arranged into several blocks, each representing a distinct point in the feature space. The positions, along with the angle matrices, define the rotation matrix, which is an element of the special orthogonal group. The rotation transformation projects a feature point onto another point on the surface of the same hypersphere." + ], + "image_footnote": [], + "bbox": [ + 151, + 85, + 848, + 393 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Main theorems", + "text_level": 1, + "bbox": [ + 89, + 474, + 243, + 489 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Based on the formal definitions above, we present our key theoretical results in the following theorem.", + "bbox": [ + 89, + 500, + 482, + 529 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Theorem 1. Let $\\mathbf{R}(\\cdot ;\\mathcal{A}): \\mathbb{R}^N \\to \\mathbb{R}^{d\\times d}$ ( $N > 1$ ) be a rotation matrix function parameterized by angle matrices $\\mathcal{A}$ . The rotation difference $\\mathbf{R}(\\pmb{x})^\\top \\mathbf{R}(\\pmb{y})$ can be represented by the location difference $\\pmb{y} - \\pmb{x}$ if and only if $\\mathcal{A}$ pairwise commute.", + "bbox": [ + 89, + 534, + 483, + 609 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The proof of Theorem 1 is shown in Appendix A.", + "bbox": [ + 109, + 616, + 439, + 631 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Theorem 1 together with Proposition 1 demonstrate that a function $f$ is a RoPE function parameterized by angle matrices if and only if the angle matrices $\\mathcal{A}$ pairwise commute. In other words, to construct a relative positional encoding method for an attention mechanism that is robust to offset, it suffices to establish an angle matrix set that pairwise commutes. Thus, in the following section, we focus on the construction of $\\mathcal{A}$ to satisfy this particular requirement. We call our method ComRoPE to indicate the commutativity of angle matrices in RoPE function. The overview of ComRoPE is shown in Figure 1.", + "bbox": [ + 89, + 631, + 483, + 797 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Remark 1. Among all the previous RoPE methods, LieRE [20] is the most related one to ours which solves Equation 3 by directly training the skew-symmetric matrix set $\\mathcal{A}$ . However, it is worth noting that in the following equation of their implementation", + "bbox": [ + 89, + 801, + 483, + 878 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left(\\mathbf {R} (\\boldsymbol {x} _ {q}; \\mathcal {A}) \\boldsymbol {q}\\right) ^ {\\top} \\left(\\mathbf {R} (\\boldsymbol {x} _ {k}; \\mathcal {A}) \\boldsymbol {k}\\right) = \\boldsymbol {q} ^ {\\top} \\mathbf {R} (\\boldsymbol {x} _ {q}; \\mathcal {A}) ^ {\\top} \\mathbf {R} (\\boldsymbol {x} _ {k}; \\mathcal {A}) \\boldsymbol {k}, (4)\n$$\n", + "text_format": "latex", + "bbox": [ + 102, + 883, + 483, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "sometimes $\\mathbf{R}(\\boldsymbol{x}_q; \\mathcal{A})^\\top \\mathbf{R}(\\boldsymbol{x}_k; \\mathcal{A}) \\neq \\mathbf{R}(\\boldsymbol{x}_q - \\boldsymbol{x}_k; \\mathcal{A})$ because the probability that two random matrices commute is small, where $\\boldsymbol{x}_q$ and $\\boldsymbol{x}_k$ are two arbitrary coordinates. Thus, $\\mathbf{R}$ proposed by LieRe does not consistently satisfy the requirements of the RoPE Equation.", + "bbox": [ + 511, + 474, + 905, + 551 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Construction of pairwise commuting matrices", + "text_level": 1, + "bbox": [ + 511, + 558, + 901, + 574 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we elaborate on concrete ways to construct the pairwise commuting matrices to solve the RoPE Equation parameterized by angle matrices. Note that if two matrices are both block diagonal with the same block sizes, where the corresponding blocks are commutative, then these two matrices are commutative. Formally speaking, for two matrices $\\mathbf{M},\\mathbf{N}\\in \\mathbb{R}^{d\\times d}$ , they commute if", + "bbox": [ + 511, + 580, + 906, + 686 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{ \\begin{array}{l} \\mathbf {M} = \\operatorname {d i a g} \\left(\\mathbf {M} _ {1}, \\mathbf {M} _ {2}, \\dots , \\mathbf {M} _ {b}\\right) \\\\ \\mathbf {N} = \\operatorname {d i a g} \\left(\\mathbf {N} _ {1}, \\mathbf {N} _ {2}, \\dots , \\mathbf {N} _ {b}\\right) \\\\ \\mathbf {M} _ {i}, \\mathbf {N} _ {i} \\in \\mathbb {R} ^ {b \\times b} \\quad \\forall i \\in \\{1, 2, \\dots , m \\} \\\\ \\mathbf {M} _ {i} \\mathbf {N} _ {i} = \\mathbf {N} _ {i} \\mathbf {M} _ {i} \\quad \\forall i \\in \\{1, 2, \\dots , m \\} \\end{array} \\right. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 557, + 694, + 905, + 776 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $b$ denotes block size that is a factor of $d$ and $m = \\frac{d}{b}$ .", + "bbox": [ + 511, + 782, + 901, + 799 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Thus, to present solutions to RoPE Equation parameterized by angle matrices, it suffices to partition the angle matrices $\\mathbf{A}_i$ in Eq. (3) into $m$ blocks $\\mathbf{B}_{i1},\\mathbf{B}_{i2},\\dots ,\\mathbf{B}_{im}$ where:", + "bbox": [ + 511, + 799, + 903, + 857 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{ \\begin{array}{l} \\mathbf {A} _ {i} = \\operatorname {d i a g} \\left(\\mathbf {B} _ {i 1}, \\mathbf {B} _ {i 2}, \\dots , \\mathbf {B} _ {i m}\\right) \\\\ \\mathbf {B} _ {i j} \\in \\mathbb {R} ^ {b \\times b} \\quad \\forall j \\in \\{1, 2, \\dots , m \\} \\end{array} \\right. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 571, + 863, + 903, + 904 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4511", + "bbox": [ + 480, + 944, + 513, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Defining $\\mathcal{B}_j = \\{\\mathbf{B}_{1j},\\mathbf{B}_{2j},\\dots ,\\mathbf{B}_{Nj}\\}$ , if $\\mathcal{B}_j$ pairwise commutes for all $j\\in \\{1,2,\\dots ,m\\}$ , then $\\mathcal{A}$ pairwise commutes. Note that $\\mathcal{B}_1,\\mathcal{B}_2,\\dots ,\\mathcal{B}_m$ are equivalent. Thus, without causing confusion, we will uniformly use $\\mathcal{B} = \\{\\mathbf{B}_1,\\mathbf{B}_2,\\dots ,\\mathbf{B}_N\\}$ to represent $\\mathcal{B}_j = \\{\\mathbf{B}_{1j},\\mathbf{B}_{2j},\\dots ,\\mathbf{B}_{Nj}\\}$ in the following.", + "bbox": [ + 89, + 90, + 482, + 181 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Currently, there is no general method that provides the necessary and sufficient conditions for ensuring that arbitrary trainable skew-symmetric matrices commute. Therefore, we aim to establish sufficient conditions for enforcing this constraint. More concretely, we present Proposition 2 and Proposition 3 to construct two different variants of matrices, respectively.", + "bbox": [ + 89, + 181, + 483, + 287 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Proposition 2. A set of matrices $\\mathcal{B} = \\{\\mathbf{B}_1, \\mathbf{B}_2, \\dots, \\mathbf{B}_N\\}$ pairwise commutes if all but one of them are zero matrices, i.e., $\\exists k \\in \\{1, 2, \\dots, N\\}$ s.t.", + "bbox": [ + 88, + 295, + 483, + 342 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {B} _ {i} = \\mathbf {O} (\\forall i \\neq k) \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 223, + 351, + 480, + 368 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Based on Proposition 2, we propose Trainable RoPE parameterized by Axial-Partition Angle Matrices (ComRoPE-AP). We divide the $m$ blocks into $N$ parts, such that each part is responsible for the rotation of a specific axis. In this case, $m$ should be a multiple of $N$ . Specifically,", + "bbox": [ + 89, + 378, + 483, + 468 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {B} _ {i j} = \\left\\{ \\begin{array}{l l} \\mathbf {P} _ {j} - \\mathbf {P} _ {j} ^ {\\top}, & \\text {i f} j \\equiv i (\\mathrm {m o d} N) \\\\ \\mathbf {O}, & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 112, + 477, + 480, + 510 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\{\\mathbf{P}_j\\}_{j=1}^m$ represents a set of trainable matrices, $\\mathbf{O}$ denotes a zero matrix, and $\\equiv$ indicates congruence modulo.", + "bbox": [ + 89, + 520, + 482, + 551 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Proposition 3. A set of matrices $\\mathcal{B} = \\{\\mathbf{B}_1, \\mathbf{B}_2, \\dots, \\mathbf{B}_N\\}$ pairwise commutes if they pairwise linearly dependent, i.e., $\\exists \\lambda_1, \\lambda_2, \\dots, \\lambda_N \\in \\mathbb{R}$ s.t.", + "bbox": [ + 89, + 559, + 483, + 604 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\lambda_ {1} \\mathbf {B} _ {1} = \\lambda_ {2} \\mathbf {B} _ {2} = \\dots = \\lambda_ {N} \\mathbf {B} _ {N} \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 179, + 614, + 480, + 631 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Based on Proposition 3, we propose Trainable RoPE parameterized by Linearly-Dependent Angle Matrices (ComRoPE-LD). Specifically, We train a base matrix $\\mathbf{P}$ with scaling factors $\\{\\theta_i\\}_{i=1}^N$ . Then we obtain:", + "bbox": [ + 89, + 641, + 483, + 704 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {B} = \\left\\{\\mathbf {B} _ {i} = \\theta_ {i} \\left(\\mathbf {P} - \\mathbf {P} ^ {\\top}\\right) \\mid i = 1, 2, \\dots , N \\right\\} \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 124, + 710, + 480, + 729 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Implementation details of coordinates and improvements", + "text_level": 1, + "bbox": [ + 89, + 747, + 482, + 779 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4.1 Relative scaling and center offset", + "text_level": 1, + "bbox": [ + 89, + 786, + 374, + 801 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Relative scaling. In language models, positions are typically treated as discrete. Additionally, due to the relationship between tokens and the inherent uncertainty in sequence length, there is no need to scale these positions into a specified range. However, in the case of images, positions are continuous. When using different patch sizes, it", + "bbox": [ + 89, + 810, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/39ff5b1cf2b7367ee8bd2e68d6c3653f0b289ba70dedd64f332c698644271be7.jpg", + "image_caption": [ + "Figure 2. Different patch sizes result in different relative relationships." + ], + "image_footnote": [], + "bbox": [ + 516, + 71, + 906, + 325 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "becomes unreasonable to define their relative relationships solely based on the span of patches. As shown in Figure 2, the same locations with varying patch sizes can result in significantly different relative relationships.", + "bbox": [ + 511, + 402, + 905, + 463 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As a result, applying relative coordinates is a better method for measuring relative relationships in images. For an image with shape $H \\times W$ , we scale both the height $H$ and the width $W$ to 1. Therefore, for a pixel located at $(h, w)$ in the raw image, its coordinate is treated as $\\left(\\frac{h}{H}, \\frac{w}{W}\\right)$ . For high-dimensional coordinates, we perform the same operation, i.e., transform a raw coordinate $(x_1, x_2, \\dots, x_N)$ in multi-axial canvas with shape $(X_1, X_2, \\dots, X_N)$ into $\\left(\\frac{x_1}{X_1}, \\frac{x_2}{X_2}, \\dots, \\frac{x_N}{X_N}\\right)$ .", + "bbox": [ + 511, + 464, + 906, + 604 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Center offset. When we project a patch to a feature tensor with a coordinate, we aggregate all information of the patch into a specified location. Simply, we adopt the center point of a patch as the aggregation location.", + "bbox": [ + 511, + 609, + 905, + 671 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4.2 Position perturbation", + "text_level": 1, + "bbox": [ + 511, + 693, + 718, + 708 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To achieve better robustness and excellent performance across different scales during inference, we add perturbations to the coordinates of the patches. Specifically, for a patch with center $(x_{1},\\dots ,x_{N})$ and size $(\\Delta X_{1},\\dots ,\\Delta X_{N})$ , during training, we formulate its location as:", + "bbox": [ + 511, + 718, + 905, + 808 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left. \\mathcal {N} \\left(\\left(x _ {1}, \\dots , x _ {N}\\right) ^ {\\top}; \\operatorname {d i a g} \\left(\\sigma \\Delta X _ {1}, \\dots , \\sigma \\Delta X _ {N}\\right) ^ {2}\\right) \\right. \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 526, + 816, + 903, + 844 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\sigma$ is a hyper-parameter called perturbation intensity. Additionally, we truncate the location within the patch area, i.e., $\\left[x_k - \\frac{\\Delta X_k}{2}, x_k + \\frac{\\Delta X_k}{2}\\right]$ .", + "bbox": [ + 511, + 854, + 905, + 904 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "4512", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/d0027c8cd1ec53bbbd96fb9d30bd7c2469d3efde9c8bb941dd604677fa2a38fb.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Position Encoding MethodPerturbation IntensityEvaluation Resolution
112128192224256320384448512
APE130.0438.6956.458.7660.0259.2757.0454.1050.99
019.7133.4355.9758.6255.3152.6349.6846.3942.66
Vanilla RoPE136.9445.4860.7263.0963.2462.1259.2455.5151.11
036.4144.4859.9762.0362.5461.3658.5654.7950.58
LieRE138.0346.9762.2264.3664.9963.7861.1557.9253.74
038.2246.8562.0164.0764.5463.4660.7456.8952.51
ComRoPE-AP (ours)135.7546.1862.8265.3265.8364.7861.8858.2154.02
037.1746.9562.6364.7665.1164.3561.6258.1053.81
ComRoPE-LD (ours)138.3047.2863.5365.4965.9565.2762.6259.1155.29
036.8846.5462.8965.2765.6664.2761.8357.8753.64
", + "bbox": [ + 102, + 88, + 893, + 309 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Accuracy of 2D classification on ImageNet. Models are trained at a resolution of ${224} \\times {224}$ and evaluated at varying resolutions.", + "bbox": [ + 89, + 311, + 903, + 325 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 351, + 222, + 369 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we evaluate the performance of various positional encoding methods on classic vision tasks. We first assess their scalability in 2D image classification across different resolutions. Additionally, we conduct object detection experiments to demonstrate the generalizability of our approach. To further examine the ability to handle higher-dimensional data, we perform 3D classification experiments, which are detailed in Appendix B.", + "bbox": [ + 88, + 376, + 482, + 498 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. 2D classification", + "text_level": 1, + "bbox": [ + 89, + 506, + 251, + 522 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1.1 Setup", + "text_level": 1, + "bbox": [ + 89, + 530, + 186, + 545 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines and model architecture. We evaluate our proposed methods (ComRoPE-LD and ComRoPE-AP) against APE, vanilla RoPE (as introduced by RoFormer), and LieRE. To isolate the effects of positional encoding, we utilize a standard Vision Transformer (ViT-B/16) architecture with minimal modifications. The APE codebook is removed for methods that do not employ APE, and self-attention layers are replaced with RoPE self-attention parameterized by angle matrices. This design highlights the performance differences among various positional encoding methods. A block size of 8 is used in practice. More details are provided in Appendix D.", + "bbox": [ + 88, + 551, + 482, + 733 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Training and evaluation protocol. All models are trained at a standard resolution of $224 \\times 224$ and evaluated across multiple resolutions to test their robustness and scalability on the ImageNet-1K dataset [4]. The models are trained from scratch using randomly initialized parameters, ensuring no influence from pre-trained weights or external priors. To maintain fairness and reproducibility, we apply only basic data augmentation techniques, such as resizing and random cropping, focusing on relative performance comparisons rather than achieving absolute accuracy. The primary evaluation metric is accuracy on the test set across", + "bbox": [ + 88, + 734, + 482, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "various resolutions. Since APE is inherently fixed and discrete, bilinear interpolation is applied to adapt it to different resolutions during evaluation. Optimization methods and hyper-parameters are detailed in Appendix C.", + "bbox": [ + 511, + 352, + 906, + 414 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/91d57c34afde96ea9cce5466b8f19db5e9a5705309c57944613b82b58b5afcf8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 521, + 431, + 895, + 523 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1f72333d708d59e72203137d1ef3d51564cf45c13c96e0e5c875c35e811c6287.jpg", + "image_caption": [ + "Figure 3. Accuracy on ImageNet-1K for various positional encoding methods. The results for the same perturbation intensity are presented together for better comparison. For better visualization, evaluation resolution at $112 \\times 112$ is not included in these figures." + ], + "image_footnote": [], + "bbox": [ + 522, + 532, + 895, + 623 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1.2 Main results", + "text_level": 1, + "bbox": [ + 511, + 724, + 655, + 738 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The accuracy metrics for each positional encoding method across various resolutions are summarized in Table 1. Additionally, Figure 3 visually compares performance under different levels of perturbation intensity separately.", + "bbox": [ + 511, + 750, + 905, + 808 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Overall performance. Across all evaluations, APE consistently exhibits the lowest accuracy, corroborating previous findings regarding its limitations in dynamic contexts. The vanilla RoPE shows a modest improvement over APE but remains less effective. In contrast, the trainable angle matrices, namely LieRE, ComRoPE-AP, and ComRoPE-LD,", + "bbox": [ + 511, + 810, + 905, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "4513", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "demonstrate significantly higher accuracy across all resolutions. Notably, ComRoPE-LD achieves the best performance among the three, suggesting that its inherent linear dependencies may enhance flexibility and structural learning capabilities.", + "bbox": [ + 89, + 90, + 480, + 167 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Accuracy at training resolution. At the training resolution of $224 \\times 224$ , all three methods with trainable angle matrices (ComRoPE-LD, ComRoPE-AP, and LieRE) achieve comparable accuracy, significantly outperforming both APE and the standard RoPE, which underscores the effectiveness of RoPE parameterized by trainable angle matrices. Notably, ComRoPE-LD surpasses the current state-of-the-art LieRE by $1.6\\%$ .", + "bbox": [ + 89, + 169, + 480, + 290 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Scaling to higher resolution. At resolutions beyond the training size, LieRE shows the steepest decline in accuracy among the three trainable RoPE variants, indicating greater sensitivity to resolution changes. In contrast, ComRoPE-LD and ComRoPE-AP exhibit a more gradual decrease in performance, thanks to their commutative properties that enhance positional robustness. Specifically, ComRoPE-LD outperforms LieRE by $2.9\\%$ at a resolution of $512 \\times 512$ .", + "bbox": [ + 89, + 292, + 480, + 414 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Significance of commutativity. These findings illustrate the effectiveness of trainable commutative angle matrices, particularly ComRoPE-LD, in maintaining accuracy and scalability across diverse resolutions. The results underscore the importance of commutativity in ensuring robust RoPE parameterized by trainable angle matrices for vision tasks. Furthermore, to better understand the role of commutativity, we conduct additional experiments by introducing coordinate offsets in LieRE (see Section 4.3.1 for details).", + "bbox": [ + 89, + 417, + 480, + 551 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Object detection", + "text_level": 1, + "bbox": [ + 89, + 571, + 253, + 585 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To demonstrate the generalizability of our approach, we conduct object detection experiments using the framework from [32]. We adopt ViT-S as our backbone and apply ComRoPE to the attention layers. To ensure consistency with the pre-trained model, we initialize the angle matrix to zero.", + "bbox": [ + 89, + 597, + 480, + 672 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We evaluate ComRoPE-LD, LieRE, and APE on the MS COCO dataset [14]. As summarized in Table 2, both ComRoPE-LD and LieRE outperform APE, with ComRoPE-LD achieving slightly better performance than LieRE while only using nearly half the number of extra parameters.", + "bbox": [ + 89, + 675, + 480, + 765 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To compare training efficiency, we plot the results for each epoch in Figure 4. Our findings indicate that both ComRoPE-LD and LieRE converge faster than APE, requiring $29\\%$ fewer iterations to achieve the same results as APE.", + "bbox": [ + 89, + 768, + 480, + 829 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/07d37257ac979a582a3ee1b3db2aa587ffea0436a53511bbb3db0080040f7838.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
PE MethodAPAP50AP75APsAPmAPl
APE44.066.647.728.246.858.4
LieRE44.567.348.429.046.958.7
ComRoPE-LD44.767.648.529.247.160.0
", + "bbox": [ + 519, + 88, + 898, + 157 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Results of object detection on MS COCO.", + "bbox": [ + 557, + 160, + 859, + 172 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2a7590cffbb7ff400373f87e79574f869a6c3e1cdf531766c8f4a02fb629294c.jpg", + "image_caption": [ + "Figure 4. Results over the whole training procedure." + ], + "image_footnote": [], + "bbox": [ + 522, + 189, + 867, + 305 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation study", + "text_level": 1, + "bbox": [ + 511, + 335, + 663, + 352 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3.1 Impact of commutativity", + "text_level": 1, + "bbox": [ + 511, + 359, + 740, + 375 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To evaluate the significance of the commutativity of angle matrices, we conduct experiments on the LieRE by introducing a coordinate offset. Specifically, before inference, a random offset is applied uniformly across all coordinates within the image. The offset is sampled from a Gaussian distribution as follows:", + "bbox": [ + 511, + 382, + 903, + 470 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {N} (0; \\rho^ {2} \\cdot \\mathbf {I} _ {N \\times N}) \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 643, + 483, + 903, + 502 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\rho$ represents the standard deviation of the random offset, and $\\mathbf{I}_{N\\times N}$ is the identity matrix of size $N\\times N$ . It is important to note that applying the same offset to all coordinates does not influence the relative positional dependencies between the patches. The other experimental settings remain consistent with those in the 2D classification tasks.", + "bbox": [ + 511, + 511, + 903, + 601 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The results shown in Figure 5 demonstrate that the baseline model's performance deteriorates significantly as the standard deviation of the offset increases. In contrast, our proposed model, ComRoPE, maintains consistent performance across all offset values. This is due to the commutativity of the angle matrices, which allows the model to remain invariant to such coordinate shifts. The robustness of ComRoPE to this type of perturbation highlights its capacity to preserve relative positional information, even in the presence of modification introduced by coordinate offsets.", + "bbox": [ + 511, + 602, + 905, + 753 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3.2 Impact of block size", + "text_level": 1, + "bbox": [ + 511, + 771, + 705, + 786 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we examine the impact of block size on 2D classification using the ImageNet dataset. We maintain the same experimental setup as in our primary experiments, varying the block size from 2 to 8. The results are presented in Figure 6. Our findings indicate that larger block sizes consistently improve performance by extending the rotation transformation space to a more significant subgroup of the", + "bbox": [ + 511, + 794, + 905, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_footnote", + "text": "2Relative improvement is calculated as target $=$ baseline $\\times$ (1 + improvement), where the improvement denotes the percentage increase over the baseline performance. This formula applies throughout the following sections.", + "bbox": [ + 89, + 849, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "4514", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c74fed46cd6519d96090420608f1d5304df3ea1b11b005f7cde938dbb1603293.jpg", + "image_caption": [ + "Figure 5. Effect of coordinate offset on LieRE. As the standard deviation of the offset increases, the performance of the baseline model deteriorates, while ComRoPE remains unaffected (un-painted)." + ], + "image_footnote": [], + "bbox": [ + 99, + 95, + 473, + 210 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/61d78426ebfcb71393d359a0855521f0a4562974f5b426f6b6cd0f3f38efac08.jpg", + "image_caption": [ + "(a) LieRE" + ], + "image_footnote": [], + "bbox": [ + 96, + 286, + 475, + 347 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8efe65594a14fec2220edd8e5275cbe83f78697ec31e930a67846059716ef106.jpg", + "image_caption": [ + "(b) ComRoPE-AP" + ], + "image_footnote": [], + "bbox": [ + 96, + 369, + 475, + 429 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5c5df1dcd0c3dea495c46bef9c13c99829a7e4102f564457c9e6d7209d24d469.jpg", + "image_caption": [ + "(c) ComRoPE-LD", + "Figure 6. Accuracy on ImageNet for various block sizes. Larger block size results in better performance." + ], + "image_footnote": [], + "bbox": [ + 96, + 450, + 475, + 511 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "particular orthogonal group by introducing additional parameters and computation time. When the block size is small, the associated costs are minimal. However, as the block size increases, the primary term of time complexity grows to $O(Lndb^{2})$ from $O(Lnd\\cdot \\frac{d}{h})$ , which becomes significant. Therefore, we limit the block size to a maximum of 8 to balance performance with additional costs. In other words, we select a block size of 8 to optimize performance while keeping the extra computational cost manageable.", + "bbox": [ + 89, + 569, + 482, + 705 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3.3 Utility of position perturbation", + "text_level": 1, + "bbox": [ + 89, + 724, + 359, + 739 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this section, we explore the impact of positional perturbations. We conducted experiments on ComRoPE-LD and APE using the ImageNet dataset, with the results presented in Figure 7.", + "bbox": [ + 89, + 750, + 482, + 809 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As shown in Figure 7a, APE is highly sensitive to positional perturbations, leading to significant performance improvement $(+19.5\\%)$ when increasing intensity from 0 to 1) when these perturbations are introduced. For RoPE with angle matrices shown in Figure 7b, positional perturbations also resulted in some performance gains $(+2.9\\%)$ , though", + "bbox": [ + 89, + 810, + 482, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8d171c91a600ba24a9c12ce64527510a50312b8948113ffd0f1c1b5385daca9b.jpg", + "image_caption": [ + "(a) APE" + ], + "image_footnote": [], + "bbox": [ + 522, + 94, + 895, + 169 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3380304f09d22faee65a8eb2bfb9c163ad7a0ec5e552cb5210a31e9cd3177d53.jpg", + "image_caption": [ + "(b) ComRoPE-LD", + "Figure 7. Accuracy on ImageNet over varying position perturbation intensity." + ], + "image_footnote": [], + "bbox": [ + 521, + 193, + 895, + 268 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "the improvement was relatively modest. This is likely due to the inherent robustness of the RoPE design with angle matrices, which is already well-equipped to handle variations in position.", + "bbox": [ + 511, + 328, + 903, + 388 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Applications", + "text_level": 1, + "bbox": [ + 511, + 398, + 645, + 414 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In our approach, when the angle matrix is an all-zero matrix, the rotation matrix becomes the identity matrix, causing RoPE Attention to reduce to the standard Attention mechanism. When the block size of the angle matrix is set to 2, ComRoPE-AP effectively reduces to the commonly used RoPE Attention in language models. This demonstrates that our method can represent standard Attention and various common variants of RoPE Attention. Therefore, during the fine-tuning stage, we can replace standard Attention with our method, load the pre-trained weights, and fine-tune them under the new paradigm. In other words, ComRoPE can be seamlessly integrated into the fine-tuning process, even if it was not applied during pre-training. Additional experiments can be found in Appendix B.", + "bbox": [ + 511, + 421, + 906, + 633 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 646, + 633, + 661 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we proposed ComRoPE, a novel and adaptive framework for Rotary Position Embedding (RoPE) parameterized by trainable angle matrices. We rigorously formulate the RoPE Equation and then establish a necessary and sufficient condition for its solution. Our approach effectively overcomes the scalability and robustness limitations of existing RoPE methods by eliminating the need for manually designed rotation matrices and introducing a more flexible, scalable solution. Extensive experimental results show that ComRoPE outperforms the existing positional encoding methods across various tasks. Furthermore, our framework generalizes existing RoPE formulations and demonstrates the potential for broader application in Transformer models, offering insights and a solid foundation for future research in positional encoding techniques.", + "bbox": [ + 511, + 671, + 906, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "4515", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgment", + "text_level": 1, + "bbox": [ + 91, + 90, + 240, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work is supported by the National Key R&D Program of China (2022YFB4701400/4701402), SSTIC Grant (KJZD20230923115106012, KJZD20230923114916032, GJHZ20240218113604008) and Beijing Key Lab of Networked Multimedia.", + "bbox": [ + 89, + 114, + 480, + 190 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 205, + 186, + 220 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Tom B Brown. Language models are few-shot learners. arXiv preprint arXiv:2005.14165, 2020. 1", + "[2] Yirui Chen, Xudong Huang, Quan Zhang, Wei Li, Mingjian Zhu, Qiangyu Yan, Simiao Li, Hanting Chen, Hailin Hu, Jie Yang, et al. Gim: A million-scale benchmark for generative image manipulation detection and localization. arXiv preprint arXiv:2406.16531, 2024. 1", + "[3] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhanghao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E Gonzalez, et al. Vicuna: An open-source chatbot impressing gpt-4 with $90\\%$ * chatgpt quality, march 2023. URL https://lmsys.org/blog/2023-03-30-vicuna, 3(5), 2023. 2", + "[4] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, K. Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 248-255, 2009. 6", + "[5] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In North American Chapter of the Association for Computational Linguistics, 2019. 1, 2", + "[6] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 1", + "[7] Yuxin Fang, Quan Sun, Xinggang Wang, Tiejun Huang, Xinlong Wang, and Yue Cao. Eva-02: A visual representation for neon genesis. Image and Vision Computing, 149:105171, 2024. 2", + "[8] Team GLM, Aohan Zeng, Bin Xu, Bowen Wang, Chenhui Zhang, Da Yin, Dan Zhang, Diego Rojas, Guanyu Feng, Hanlin Zhao, et al. Chatglm: A family of large language models from glm-130b to glm-4 all tools. arXiv preprint arXiv:2406.12793, 2024. 1", + "[9] Byeongho Heo, Song Park, Dongyoon Han, and Sangwoo Yun. Rotary position embedding for vision transformer. arXiv preprint arXiv:2403.13298, 2024. 2", + "[10] Max Horn, Kumar Shridhar, Elrich Groenewald, and Philipp FM Baumann. Translational equivariance in kernelizable attention. arXiv preprint arXiv:2102.07680, 2021. 2", + "[11] Amirhossein Kazemnejad, Inkit Padhi, Karthikeyan Natesan Ramamurthy, Payel Das, and Siva Reddy. The impact of positional encoding on length generalization in transformers. Advances in Neural Information Processing Systems, 36: 24892-24928, 2023. 1", + "[12] Hanyu Lai, Xiao Liu, Iat Long Iong, Shuntian Yao, Yuxuan Chen, Pengbo Shen, Hao Yu, Hanchen Zhang, Xiaohan" + ], + "bbox": [ + 93, + 231, + 482, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Zhang, Yuxiao Dong, et al. Autowebglm: A large language model-based web navigating agent. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 5295-5306, 2024. 1", + "[13] Hongyu Li, Jinyu Chen, Ziyu Wei, Shaofei Huang, Tianrui Hui, Jialin Gao, Xiaoming Wei, and Si Liu. Llava-st: A multimodal large language model for fine-grained spatial-temporal understanding. arXiv preprint arXiv:2501.08282, 2025. 1", + "[14] Tsung-Yi Lin, Michael Maire, Serge Belongie, Lubomir Bourdev, Ross Girshick, James Hays, Pietro Perona, Deva Ramanan, C. Lawrence Zitnick, and Piotr Dollar. Microsoft coco: Common objects in context. arXiv preprint arXiv: 1405.0312, 2014. 7", + "[15] Xiao Liu, Hanyu Lai, Hao Yu, Yifan Xu, Aohan Zeng, Zhengxiao Du, Peng Zhang, Yuxiao Dong, and Jie Tang. Webglm: towards an efficient web-enhanced question answering system with human preferences. In Proceedings of the 29th ACM SIGKDD conference on knowledge discovery and data mining, pages 4549-4560, 2023. 1", + "[16] Xiao Liu, Hao Yu, Hanchen Zhang, Yifan Xu, Xuanyu Lei, Hanyu Lai, Yu Gu, Hangliang Ding, Kaiwen Men, Kejuan Yang, et al. Agentbench: Evaluating llms as agents. In The Twelfth International Conference on Learning Representations, 2024. 1", + "[17] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 2", + "[18] Antoine Liutkus, Ondrej Cifka, Shih-Lun Wu, Umut Simsekli, Yi-Hsuan Yang, and Gael Richard. Relative positional encoding for transformers with linear complexity. In International Conference on Machine Learning, pages 7067-7079. PMLR, 2021. 2", + "[19] Jiasen Lu, Christopher Clark, Sangho Lee, Zichen Zhang, Savya Khosla, Ryan Marten, Derek Hoiem, and Aniruddha Kembhavi. Unified-io 2: Scaling autoregressive multimodal models with vision language audio and action. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26439–26455, 2024. 2", + "[20] Sophie Ostmeier, Brian Axelrod, Michael E Moseley, Akshay Chaudhari, and Curtis Langlotz. Liere: Generalizing rotary position encodings. arXiv preprint arXiv:2406.10322, 2024. 2, 4", + "[21] Ofir Press, Noah A Smith, and Mike Lewis. Train short, test long: Attention with linear biases enables input length extrapolation. arXiv preprint arXiv:2108.12409, 2021. 2", + "[22] A Radford. Improving language understanding by generative pre-training. 2018. 1", + "[23] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. 1", + "[24] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of machine learning research, 21(140):1-67, 2020. 2" + ], + "bbox": [ + 517, + 92, + 903, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "4516", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[25] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of machine learning research, 21(140):1-67, 2020. 2", + "[26] Peter Shaw, Jakob Uszkoreit, and Ashish Vaswani. Self-attention with relative position representations. arXiv preprint arXiv:1803.02155, 2018. 2", + "[27] Jianlin Su, Murtadha Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu. Roformer: Enhanced transformer with rotary position embedding. Neurocomputing, 568:127063, 2024. 1, 2, 3", + "[28] Sainbayar Sukhbaatar, Jason Weston, Rob Fergus, et al. End-to-end memory networks. Advances in neural information processing systems, 28, 2015. 2", + "[29] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 2", + "[30] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 1, 2", + "[31] Kan Wu, Houwen Peng, Minghao Chen, Jianlong Fu, and Hongyang Chao. Rethinking and improving relative position encoding for vision transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10033-10041, 2021. 2", + "[32] Chunlong Xia, Xinliang Wang, Feng Lv, Xin Hao, and Yifeng Shi. Vit-comer: Vision transformer with convolutional multi-scale feature interaction for dense predictions. Computer Vision and Pattern Recognition, 2024. 7", + "[33] Quan Zhang and Yuxin Qi. Can mllms guide weakly-supervised temporal action localization tasks? arXiv preprint arXiv:2411.08466, 2024. 1", + "[34] Quan Zhang, Xiaoyu Liu, Wei Li, Hanting Chen, Junchao Liu, Jie Hu, Zhiwei Xiong, Chun Yuan, and Yunhe Wang. Distilling semantic priors from sam to efficient image restoration models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25409-25419, 2024.", + "[35] Quan Zhang, Yuxin Qi, Xi Tang, Jinwei Fang, Xi Lin, Ke Zhang, and Chun Yuan. IMDPrompter: Adapting SAM to image manipulation detection by cross-view automated prompt learning. In The Thirteenth International Conference on Learning Representations, 2025.", + "[36] Quan Zhang, Yuxin Qi, Xi Tang, Rui Yuan, Xi Lin, Ke Zhang, and Chun Yuan. Rethinking pseudo-label guided learning for weakly supervised temporal action localization from the perspective of noise correction. arXiv preprint arXiv:2501.11124, 2025. 1" + ], + "bbox": [ + 91, + 92, + 480, + 797 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "4517", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/ComRoPE_ Scalable and Robust Rotary Position Embedding Parameterized by Trainable Commuting Angle Matrices/287888de-80a8-43c2-8d0f-f61f04d80e4e_model.json b/2025/ComRoPE_ Scalable and Robust Rotary Position Embedding Parameterized by Trainable Commuting Angle Matrices/287888de-80a8-43c2-8d0f-f61f04d80e4e_model.json new file mode 100644 index 0000000000000000000000000000000000000000..d4f62c53f07fc0e452f08345b97d76fe0f721ecb --- /dev/null +++ b/2025/ComRoPE_ Scalable and Robust Rotary Position Embedding Parameterized by Trainable Commuting Angle Matrices/287888de-80a8-43c2-8d0f-f61f04d80e4e_model.json @@ -0,0 +1,2321 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.101, + 0.13, + 0.9, + 0.177 + ], + "angle": 0, + "content": "ComRoPE: Scalable and Robust Rotary Position Embedding Parameterized by Trainable Commuting Angle Matrices" + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.202, + 0.868, + 0.223 + ], + "angle": 0, + "content": "Hao \\(\\mathbf{Y}\\mathbf{u}^{1}\\) Tangyu Jiang\\(^{1\\dagger}\\) Shuning Jia\\(^{1,2}\\) Shannon Yan\\(^{1}\\) Shunning Liu\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.18, + 0.228, + 0.817, + 0.248 + ], + "angle": 0, + "content": "Haolong Qian1 Guanghao Li1 Shuting Dong1 Chun Yuan1†" + }, + { + "type": "text", + "bbox": [ + 0.316, + 0.256, + 0.682, + 0.275 + ], + "angle": 0, + "content": "\\(^{1}\\)Tsinghua University \\(^{2}\\)Shenzhen University" + }, + { + "type": "text", + "bbox": [ + 0.155, + 0.279, + 0.828, + 0.293 + ], + "angle": 0, + "content": "longinyh@gmail.com, jiangtangyu@sz.tsinghua.edu.cn, yuanc@sz.tsinghua.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.319, + 0.327, + 0.335 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.351, + 0.486, + 0.761 + ], + "angle": 0, + "content": "The Transformer architecture has revolutionized various fields since it was proposed, where positional encoding plays an essential role in effectively capturing sequential order and context. Therefore, Rotary Positional Encoding (RoPE) was proposed to alleviate these issues, which integrates positional information by rotating the embeddings in the attention mechanism. However, RoPE utilizes manually defined rotation matrices, a design choice that favors computational efficiency but limits the model's flexibility and adaptability. In this work, we propose ComRoPE, which generalizes RoPE by defining it in terms of trainable commuting angle matrices. Specifically, we demonstrate that pairwise commutativity of these matrices is essential for RoPE to achieve scalability and positional robustness. We formally define the RoPE Equation, which is an essential condition that ensures consistent performance with position offsets. Based on the theoretical analysis, we present two types of trainable commuting angle matrices as sufficient solutions to the RoPE equation, which significantly improve performance, surpassing the current state-of-the-art method by \\(1.6\\%\\) at training resolution and \\(2.9\\%\\) at higher resolution on the ImageNet-1K dataset. Furthermore, our framework shows versatility in generalizing to existing RoPE formulations and offering new insights for future positional encoding research. To ensure reproducibility, the source code and instructions are available at https://github.com/Longin-Yu/ComRoPE." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.767, + 0.222, + 0.782 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.792, + 0.484, + 0.869 + ], + "angle": 0, + "content": "The Transformer architecture [30] has been widely adopted across various fields, including Natural Language Processing (NLP) [1, 5, 8, 22, 23] and Computer Vision (CV) [6]. Moreover, an increasing number of Transformer-based applications [2, 12, 13, 15, 16, 33-36] continue to demonstrate" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.331, + 0.907, + 0.422 + ], + "angle": 0, + "content": "their effectiveness across a wide range of domains. At the core of the Transformer model lies the attention mechanism, which enables the model to selectively focus on different parts of the input based on relevance, thereby effectively capturing long-range dependencies and contextual relationships." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.425, + 0.909, + 0.531 + ], + "angle": 0, + "content": "However, since the attention mechanism is insensitive to the fundamental position information, it cannot inherently capture the order of elements in the data. In order to alleviate this issue, positional embeddings are added to the input representation, providing the model with the necessary positional information to process the sequence order and structure effectively." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.535, + 0.909, + 0.852 + ], + "angle": 0, + "content": "A variety of positional encoding methods have been proposed, which can be divided into two categories: Absolute Positional Encoding (APE) and Relative Positional Encoding (RPE). APE explicitly encodes the absolute position of each token by generating fixed positional embeddings, which are directly added to the input embeddings at the beginning of training [30]. However, APE struggles with handling long sequences and exhibits high sensitivity to positional shifts, limiting the scalability and robustness of models [11]. In contrast, RPE does not directly modify input embeddings; instead, it incorporates relative positional information within the attention mechanism, enabling the model to effectively capture positional relationships between tokens. Among all the existing works, Rotary Position Embedding (RoPE) [27] has gained significant attention due to the advantage of applying a rotational transformation to token embeddings. It encodes relative positions by treating each pair of features as coordinates and rotating them by an angle proportional to their position, enabling all tokens to interact within the attention mechanism regardless of distance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.909, + 0.903 + ], + "angle": 0, + "content": "However, existing RoPE methods face several key challenges: i) The essential components (i.e., the RoPE matrices) of previous RoPE approaches rely on 2D rotation" + }, + { + "type": "page_footnote", + "bbox": [ + 0.09, + 0.875, + 0.483, + 0.901 + ], + "angle": 0, + "content": "†Corresponding author. This work was done when Shuning Jia was an intern at Tsinghua University." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4508" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.243 + ], + "angle": 0, + "content": "groups, which simplify computations but consequently restrict their feature projection capabilities, especially in high-dimensional spaces [27]. ii) Moreover, the majority of the rotation matrices of RoPE require to be manually designed, leading to insufficient capability and suboptimal performance[21, 25]. iii) Finally, previous attempts [20] to extend the rotation group often prioritize design simplicity, making it difficult to consistently satisfy relative position dependency—a critical property of RoPE that ensures positional robustness against offsets." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.254, + 0.482, + 0.451 + ], + "angle": 0, + "content": "Our objectives. This work aims to develop a novel RoPE method for Transformers that is both scalable and robust. Specifically, we seek to extend the rotation group from the existing 2D representation to a larger subgroup of the special orthogonal group, allowing for higher degrees of freedom while preserving consistent behavior with respect to position offsets. Unlike existing methods that rely on manually designed non-trainable rotation matrices, which suffer from limited expressiveness and reduced robustness, our approach is designed to offer richer feature representation capabilities. This framework addresses the scalability limitations of current approaches and enhances their robustness against positional transformations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.461, + 0.482, + 0.611 + ], + "angle": 0, + "content": "Our contributions. This work introduces ComRoPE, a novel framework that significantly enhances positional encoding in Transformers. ComRoPE leverages trainable angle matrices, extending the RoPE mechanism with higher scalability and robustness. We identify the pairwise commutativity of these matrices as a necessary and sufficient condition for effective positional encoding, thereby unifying various existing RoPE formulations under a single theoretical framework. The contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.624, + 0.48, + 0.68 + ], + "angle": 0, + "content": "- We formally define the RoPE function parameterized by angle matrices and prove that pairwise commutativity is a necessary and sufficient condition, offering a unified theory that encompasses several existing RoPE variants." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.685, + 0.48, + 0.756 + ], + "angle": 0, + "content": "- We introduce ComRoPE, a scalable and robust solution that leverages two types of trainable commuting angle matrix sets as sufficient solutions to the RoPE equation, capturing richer positional representations without the need for manual design." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.76, + 0.48, + 0.845 + ], + "angle": 0, + "content": "- Our extensive experiments show that ComRoPE surpasses the current state-of-the-art LieRE, achieving a performance increase of \\(1.6\\%\\) at training resolution and \\(2.9\\%\\) at higher resolutions on the ImageNet-1K classification task while delivering strong results across other benchmarks." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.85, + 0.48, + 0.893 + ], + "angle": 0, + "content": "- We explore further applications of ComRoPE, providing valuable insights for advancing position encoding techniques in future Transformer-based models." + }, + { + "type": "list", + "bbox": [ + 0.092, + 0.624, + 0.48, + 0.893 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.09, + 0.652, + 0.106 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.118, + 0.8, + 0.133 + ], + "angle": 0, + "content": "2.1. Position information in attention" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.145, + 0.905, + 0.584 + ], + "angle": 0, + "content": "Transformers [30] utilize the attention mechanism to capture similarities within sequences. However, they lack inherent sequential information and cannot capture the positional information of each token. To address this limitation, positional encoding [5, 24, 26] was introduced. As research progressed, positional encoding generally evolved into two types: APE [5, 28, 30] and RPE [10, 18, 26, 27]. [30] first proposed using APE in the form of sine and cosine functions, effectively representing the positional relationships within the input sequence. This positional encoding method achieved remarkable results in natural language processing, becoming a foundational component for many NLP tasks. However, the fixed nature of positional encoding limited the model's ability to generalize to longer input sequences. In response, subsequent research introduced learnable absolute positional encoding. [5] proposed further enhancing the model's expressive power by incorporating learnable position embeddings, particularly excelling in tasks such as sentence alignment and context representation. Although APE provides positional information to enhance the model's understanding of sequences, it shows limitations in handling long sequences and cross-sequence scenarios. Hierarchical ViT such as Swin Transformer [17], introduced Relative Position Bias (RPB) [17, 25, 26] to handle large numbers of tokens with limited positional embeddings [9]. Related research has explored alternative encoding methods, such as RPE, as a replacement for APE to better capture complex dependency structures. iRPE [31] proposed an improved RPB by combining relative position embedding." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.603, + 0.76, + 0.619 + ], + "angle": 0, + "content": "2.2. Rotary position embedding" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.629, + 0.905, + 0.9 + ], + "angle": 0, + "content": "Building on these developments, RoFormer [27] combined the advantages of APE and RPE, proposing RoPE. Currently, RoPE is widely used in Large Language Models (LLMs), such as LLaMA [29] and Vicuna [3]. This approach enhances model performance on tasks involving long-text semantics and multi-turn dialogues, improving extrapolation capability. To better adapt to the characteristics of two-dimensional data such as images, researchers extended RoPE to two-dimensional sequences (2D-RoPE). LieRE [20] extends RoPE to a more generalized form by introducing a rotation-based positional encoding method grounded in Lie group theory. Unified-IO 2 [19] applies 2D-RoPE within its multimodal architecture; EVA-02 [7], FiT [7] these pioneering works used 2D RoPE with axial frequencies (2D Axial RoPE), but had limitations in processing in the diagonal direction. Therefore, RoPE for ViT [9] proposes to use mixed axial frequency for 2D RoPE, named RoPE-Mixed." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4509" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.09, + 0.182, + 0.105 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.484, + 0.237 + ], + "angle": 0, + "content": "In this section, we begin by introducing key definitions and reformulating the RoPE paradigm within multi-axial attention mechanisms, which we collectively refer to as the RoPE Equations. Next, we present our main theorem, which establishes the necessary and sufficient conditions for RoPE functions parameterized by angle matrices. Finally, we provide solutions to the RoPE Equations based on the propositions that outline the sufficient conditions." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.244, + 0.232, + 0.259 + ], + "angle": 0, + "content": "3.1. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.267, + 0.484, + 0.328 + ], + "angle": 0, + "content": "We use \\(\\mathbf{R}(\\cdot):\\mathbb{R}^N\\to \\mathbb{R}^{d\\times d}\\) to denote a matrix-value function. We denote the vectors and matrices as the lower and uppercase bold font, respectively. We first recall two fundamental definitions of matrices as follows:" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.335, + 0.484, + 0.381 + ], + "angle": 0, + "content": "Definition 1 (Matrix Exponential). The exponential of a square matrix \\(\\mathbf{A}\\), denoted as \\(e^{\\mathbf{A}}\\) or \\(\\exp (\\mathbf{A})\\), is defined using the matrix exponential series such that:" + }, + { + "type": "equation", + "bbox": [ + 0.099, + 0.388, + 0.476, + 0.428 + ], + "angle": 0, + "content": "\\[\ne ^ {\\mathbf {A}} = \\exp (\\mathbf {A}) = \\sum_ {k = 0} ^ {\\infty} \\frac {\\mathbf {A} ^ {k}}{k !} = \\mathbf {I} + \\mathbf {A} + \\frac {\\mathbf {A} ^ {2}}{2 !} + \\frac {\\mathbf {A} ^ {3}}{3 !} + \\dots ,\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.434, + 0.402, + 0.45 + ], + "angle": 0, + "content": "This series converges for any square matrix \\(\\mathbf{A}\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.456, + 0.484, + 0.502 + ], + "angle": 0, + "content": "Definition 2 (Commuting Matrices). Two square matrices \\(\\mathbf{A}_1\\) and \\(\\mathbf{A}_2\\) are said to commute if their product is independent of the order of multiplication, i.e.," + }, + { + "type": "equation", + "bbox": [ + 0.229, + 0.51, + 0.345, + 0.526 + ], + "angle": 0, + "content": "\\[\n\\mathbf {A} _ {1} \\mathbf {A} _ {2} = \\mathbf {A} _ {2} \\mathbf {A} _ {1}.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.534, + 0.484, + 0.594 + ], + "angle": 0, + "content": "A set of square matrices \\(\\{\\mathbf{A}_1,\\mathbf{A}_2,\\dots ,\\mathbf{A}_N\\}\\) is said to pairwise commute if every pair of matrices within the set commutes with each other. That is, for all \\(i,j\\) such that \\(1\\leq i,j\\leq N\\)" + }, + { + "type": "equation", + "bbox": [ + 0.232, + 0.595, + 0.343, + 0.611 + ], + "angle": 0, + "content": "\\[\n\\mathbf {A} _ {i} \\mathbf {A} _ {j} = \\mathbf {A} _ {j} \\mathbf {A} _ {i}.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.617, + 0.484, + 0.768 + ], + "angle": 0, + "content": "To clarify and better illustrate the main theorems presented in the following sections, we first reformulate and unify the definitions of the RoPE paradigm, specifically in the context of multi-axial attention mechanisms. RoPE was initially proposed by Su et al. [27] as a positional encoding method based on relative position dependencies. However, previous work provided only conceptual and descriptive descriptions of RPE and RoPE without offering a rigorous formal definition. In this work, we provide the formal definitions of both RPE and RoPE." + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.774, + 0.485, + 0.852 + ], + "angle": 0, + "content": "Definition 3 (RPE Equation). Let \\( f: \\mathbb{R}^d \\times \\mathbb{R}^N \\to \\mathbb{R}^d \\) be a positional encoding function. \\( f \\) is said to be a RPE function if and only if there exists a function \\( \\rho: \\mathbb{R}^d \\times \\mathbb{R}^d \\to \\mathbb{R} \\), and a function \\( g: \\mathbb{R}^d \\times \\mathbb{R}^d \\times \\mathbb{R}^N \\to \\mathbb{R} \\) such that the following conditions hold for all \\( x, y \\in \\mathbb{R}^N \\) and \\( q, k \\in \\mathbb{R}^d \\):" + }, + { + "type": "equation", + "bbox": [ + 0.159, + 0.865, + 0.483, + 0.882 + ], + "angle": 0, + "content": "\\[\ng (\\boldsymbol {q}, \\boldsymbol {k}, \\boldsymbol {x} - \\boldsymbol {y}) = \\rho (f (\\boldsymbol {q}, \\boldsymbol {x}), f (\\boldsymbol {k}, \\boldsymbol {y})), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.886, + 0.366, + 0.902 + ], + "angle": 0, + "content": "We refer to Eq. (1) as the RPE Equation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.09, + 0.907, + 0.168 + ], + "angle": 0, + "content": "Definition 4 (RoPE Equation). Let \\( f: \\mathbb{R}^d \\times \\mathbb{R}^N \\to \\mathbb{R}^d \\) be an RPE function. \\( f \\) is said to be a RoPE function if and only if there exists a matrix-valued function \\( \\mathbf{R}_f(\\cdot) \\) such that the following conditions hold for all \\( \\mathbf{x}, \\mathbf{y} \\in \\mathbb{R}^N \\) and \\( \\mathbf{k}, \\mathbf{k} \\in \\mathbb{R}^d \\) in RPE Equation:" + }, + { + "type": "equation", + "bbox": [ + 0.577, + 0.178, + 0.907, + 0.238 + ], + "angle": 0, + "content": "\\[\n\\left\\{ \\begin{array}{l} f (\\boldsymbol {q}, \\boldsymbol {x}) = \\mathbf {R} _ {f} (\\boldsymbol {x}) \\boldsymbol {q} \\\\ \\rho (\\boldsymbol {q}, \\boldsymbol {k}) = \\boldsymbol {q} ^ {\\top} \\boldsymbol {k} \\\\ g (\\boldsymbol {q}, \\boldsymbol {k}, \\boldsymbol {x} - \\boldsymbol {y}) = \\boldsymbol {q} ^ {\\top} \\mathbf {R} _ {f} (\\boldsymbol {y} - \\boldsymbol {x}) \\boldsymbol {k} \\end{array} \\right. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.247, + 0.796, + 0.263 + ], + "angle": 0, + "content": "We refer to Eq. (2) as the RoPE Equation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.272, + 0.906, + 0.318 + ], + "angle": 0, + "content": "By substituting the RoPE equation into the RPE equation, we obtain that the RoPE function satisfies the following property." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.327, + 0.906, + 0.359 + ], + "angle": 0, + "content": "Proposition 1. \\( f \\) is said to be a RoPE function if and only if the matrix-valued function \\( \\mathbf{R}_f(\\cdot) \\) satisfies:" + }, + { + "type": "equation", + "bbox": [ + 0.604, + 0.368, + 0.815, + 0.387 + ], + "angle": 0, + "content": "\\[\n\\mathbf {R} _ {f} (\\boldsymbol {x}) ^ {\\top} \\mathbf {R} _ {f} (\\boldsymbol {y}) = \\mathbf {R} _ {f} (\\boldsymbol {y} - \\boldsymbol {x}),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.397, + 0.64, + 0.413 + ], + "angle": 0, + "content": "for all \\(\\pmb {x},\\pmb {y}\\in \\mathbb{R}^{N}\\)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.422, + 0.906, + 0.514 + ], + "angle": 0, + "content": "Note that the definition of RoPE Equation demonstrates that the position encoding should only be dependent on the relative position of the tokens, which thus be robust against the offset operations. We further expand the definition of the RoPE function to a parameterized one (i.e., Definition 6) via rotation matrices (i.e., Definition 5) as follows1." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.523, + 0.907, + 0.615 + ], + "angle": 0, + "content": "Definition 5 (Parameterized Rotation Matrix). Let \\(\\mathbf{R}(\\cdot ;\\mathcal{A}): \\mathbb{R}^N \\to \\mathbb{R}^{d\\times d}\\) be a matrix-valued function parameterized by \\(N\\) skew-symmetric matrices \\(\\mathcal{A} = \\{\\mathbf{A}_1,\\mathbf{A}_2,\\dots ,\\mathbf{A}_N\\}\\). We say that \\(\\mathbf{R}(\\cdot ;\\mathcal{A})\\) is a rotation matrix function parameterized by angle matrices \\(\\mathcal{A}\\) if it can be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.606, + 0.624, + 0.906, + 0.667 + ], + "angle": 0, + "content": "\\[\n\\mathbf {R} (\\boldsymbol {x}; \\mathcal {A}) = \\exp \\left(\\sum_ {i = 1} ^ {N} \\mathbf {A} _ {i} x _ {i}\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.678, + 0.906, + 0.709 + ], + "angle": 0, + "content": "where \\(\\pmb{x} = (x_{1}, x_{2}, \\ldots, x_{N}) \\in \\mathbb{R}^{N}\\) and \\(\\exp(\\cdot)\\) denotes the matrix exponential." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.717, + 0.906, + 0.794 + ], + "angle": 0, + "content": "Definition 6 (Parameterized RoPE Function). Let \\( f: \\mathbb{R}^d \\times \\mathbb{R}^N \\to \\mathbb{R}^d \\) be a RoPE function. \\( f \\) is said to be parameterized by angle matrices if and only if there exists a rotation matrix function \\( \\mathbf{R}_f(\\cdot; \\mathcal{A}) \\) parameterized by angle matrices \\( \\mathcal{A} \\) such that the RoPE Equation (i.e., Eq. (2)) holds." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.794, + 0.906, + 0.869 + ], + "angle": 0, + "content": "In this case, \\(\\mathbf{R}_f(\\cdot ;\\mathcal{A})\\) is referred to as the rotation matrix function of RoPE function \\(f(\\cdot ;\\mathcal{A})\\) parameterized by angle matrices \\(\\mathcal{A}\\). For simplicity, we slightly abuse the notation and refer to \\(\\mathbf{R}\\) as the rotation matrix of RoPE function \\(f\\)." + }, + { + "type": "page_footnote", + "bbox": [ + 0.513, + 0.875, + 0.906, + 0.901 + ], + "angle": 0, + "content": "For clarification, the definitions of the rotation matrix and its exponential representation can be found in Appendix A." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4510" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.152, + 0.086, + 0.849, + 0.395 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.408, + 0.908, + 0.452 + ], + "angle": 0, + "content": "Figure 1. Overview of ComRoPE. Features are arranged into several blocks, each representing a distinct point in the feature space. The positions, along with the angle matrices, define the rotation matrix, which is an element of the special orthogonal group. The rotation transformation projects a feature point onto another point on the surface of the same hypersphere." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.476, + 0.245, + 0.491 + ], + "angle": 0, + "content": "3.2. Main theorems" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.5, + 0.483, + 0.53 + ], + "angle": 0, + "content": "Based on the formal definitions above, we present our key theoretical results in the following theorem." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.535, + 0.484, + 0.61 + ], + "angle": 0, + "content": "Theorem 1. Let \\(\\mathbf{R}(\\cdot ;\\mathcal{A}): \\mathbb{R}^N \\to \\mathbb{R}^{d\\times d}\\) (\\(N > 1\\)) be a rotation matrix function parameterized by angle matrices \\(\\mathcal{A}\\). The rotation difference \\(\\mathbf{R}(\\pmb{x})^\\top \\mathbf{R}(\\pmb{y})\\) can be represented by the location difference \\(\\pmb{y} - \\pmb{x}\\) if and only if \\(\\mathcal{A}\\) pairwise commute." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.617, + 0.441, + 0.632 + ], + "angle": 0, + "content": "The proof of Theorem 1 is shown in Appendix A." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.632, + 0.484, + 0.798 + ], + "angle": 0, + "content": "Theorem 1 together with Proposition 1 demonstrate that a function \\( f \\) is a RoPE function parameterized by angle matrices if and only if the angle matrices \\( \\mathcal{A} \\) pairwise commute. In other words, to construct a relative positional encoding method for an attention mechanism that is robust to offset, it suffices to establish an angle matrix set that pairwise commutes. Thus, in the following section, we focus on the construction of \\( \\mathcal{A} \\) to satisfy this particular requirement. We call our method ComRoPE to indicate the commutativity of angle matrices in RoPE function. The overview of ComRoPE is shown in Figure 1." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.803, + 0.485, + 0.879 + ], + "angle": 0, + "content": "Remark 1. Among all the previous RoPE methods, LieRE [20] is the most related one to ours which solves Equation 3 by directly training the skew-symmetric matrix set \\(\\mathcal{A}\\). However, it is worth noting that in the following equation of their implementation" + }, + { + "type": "equation", + "bbox": [ + 0.103, + 0.885, + 0.484, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\left(\\mathbf {R} (\\boldsymbol {x} _ {q}; \\mathcal {A}) \\boldsymbol {q}\\right) ^ {\\top} \\left(\\mathbf {R} (\\boldsymbol {x} _ {k}; \\mathcal {A}) \\boldsymbol {k}\\right) = \\boldsymbol {q} ^ {\\top} \\mathbf {R} (\\boldsymbol {x} _ {q}; \\mathcal {A}) ^ {\\top} \\mathbf {R} (\\boldsymbol {x} _ {k}; \\mathcal {A}) \\boldsymbol {k}, (4)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.476, + 0.906, + 0.553 + ], + "angle": 0, + "content": "sometimes \\(\\mathbf{R}(\\boldsymbol{x}_q; \\mathcal{A})^\\top \\mathbf{R}(\\boldsymbol{x}_k; \\mathcal{A}) \\neq \\mathbf{R}(\\boldsymbol{x}_q - \\boldsymbol{x}_k; \\mathcal{A})\\) because the probability that two random matrices commute is small, where \\(\\boldsymbol{x}_q\\) and \\(\\boldsymbol{x}_k\\) are two arbitrary coordinates. Thus, \\(\\mathbf{R}\\) proposed by LieRe does not consistently satisfy the requirements of the RoPE Equation." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.559, + 0.903, + 0.575 + ], + "angle": 0, + "content": "3.3. Construction of pairwise commuting matrices" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.582, + 0.907, + 0.688 + ], + "angle": 0, + "content": "In this section, we elaborate on concrete ways to construct the pairwise commuting matrices to solve the RoPE Equation parameterized by angle matrices. Note that if two matrices are both block diagonal with the same block sizes, where the corresponding blocks are commutative, then these two matrices are commutative. Formally speaking, for two matrices \\(\\mathbf{M},\\mathbf{N}\\in \\mathbb{R}^{d\\times d}\\), they commute if" + }, + { + "type": "equation", + "bbox": [ + 0.558, + 0.695, + 0.906, + 0.777 + ], + "angle": 0, + "content": "\\[\n\\left\\{ \\begin{array}{l} \\mathbf {M} = \\operatorname {d i a g} \\left(\\mathbf {M} _ {1}, \\mathbf {M} _ {2}, \\dots , \\mathbf {M} _ {b}\\right) \\\\ \\mathbf {N} = \\operatorname {d i a g} \\left(\\mathbf {N} _ {1}, \\mathbf {N} _ {2}, \\dots , \\mathbf {N} _ {b}\\right) \\\\ \\mathbf {M} _ {i}, \\mathbf {N} _ {i} \\in \\mathbb {R} ^ {b \\times b} \\quad \\forall i \\in \\{1, 2, \\dots , m \\} \\\\ \\mathbf {M} _ {i} \\mathbf {N} _ {i} = \\mathbf {N} _ {i} \\mathbf {M} _ {i} \\quad \\forall i \\in \\{1, 2, \\dots , m \\} \\end{array} \\right. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.783, + 0.903, + 0.8 + ], + "angle": 0, + "content": "where \\( b \\) denotes block size that is a factor of \\( d \\) and \\( m = \\frac{d}{b} \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.8, + 0.905, + 0.858 + ], + "angle": 0, + "content": "Thus, to present solutions to RoPE Equation parameterized by angle matrices, it suffices to partition the angle matrices \\(\\mathbf{A}_i\\) in Eq. (3) into \\(m\\) blocks \\(\\mathbf{B}_{i1},\\mathbf{B}_{i2},\\dots ,\\mathbf{B}_{im}\\) where:" + }, + { + "type": "equation", + "bbox": [ + 0.572, + 0.864, + 0.905, + 0.905 + ], + "angle": 0, + "content": "\\[\n\\left\\{ \\begin{array}{l} \\mathbf {A} _ {i} = \\operatorname {d i a g} \\left(\\mathbf {B} _ {i 1}, \\mathbf {B} _ {i 2}, \\dots , \\mathbf {B} _ {i m}\\right) \\\\ \\mathbf {B} _ {i j} \\in \\mathbb {R} ^ {b \\times b} \\quad \\forall j \\in \\{1, 2, \\dots , m \\} \\end{array} \\right. \\tag {6}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "4511" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.091, + 0.483, + 0.183 + ], + "angle": 0, + "content": "Defining \\(\\mathcal{B}_j = \\{\\mathbf{B}_{1j},\\mathbf{B}_{2j},\\dots ,\\mathbf{B}_{Nj}\\}\\), if \\(\\mathcal{B}_j\\) pairwise commutes for all \\(j\\in \\{1,2,\\dots ,m\\}\\), then \\(\\mathcal{A}\\) pairwise commutes. Note that \\(\\mathcal{B}_1,\\mathcal{B}_2,\\dots ,\\mathcal{B}_m\\) are equivalent. Thus, without causing confusion, we will uniformly use \\(\\mathcal{B} = \\{\\mathbf{B}_1,\\mathbf{B}_2,\\dots ,\\mathbf{B}_N\\}\\) to represent \\(\\mathcal{B}_j = \\{\\mathbf{B}_{1j},\\mathbf{B}_{2j},\\dots ,\\mathbf{B}_{Nj}\\}\\) in the following." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.183, + 0.484, + 0.289 + ], + "angle": 0, + "content": "Currently, there is no general method that provides the necessary and sufficient conditions for ensuring that arbitrary trainable skew-symmetric matrices commute. Therefore, we aim to establish sufficient conditions for enforcing this constraint. More concretely, we present Proposition 2 and Proposition 3 to construct two different variants of matrices, respectively." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.296, + 0.484, + 0.343 + ], + "angle": 0, + "content": "Proposition 2. A set of matrices \\(\\mathcal{B} = \\{\\mathbf{B}_1, \\mathbf{B}_2, \\dots, \\mathbf{B}_N\\}\\) pairwise commutes if all but one of them are zero matrices, i.e., \\(\\exists k \\in \\{1, 2, \\dots, N\\}\\) s.t." + }, + { + "type": "equation", + "bbox": [ + 0.225, + 0.352, + 0.482, + 0.369 + ], + "angle": 0, + "content": "\\[\n\\mathbf {B} _ {i} = \\mathbf {O} (\\forall i \\neq k) \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.379, + 0.484, + 0.469 + ], + "angle": 0, + "content": "Based on Proposition 2, we propose Trainable RoPE parameterized by Axial-Partition Angle Matrices (ComRoPE-AP). We divide the \\(m\\) blocks into \\(N\\) parts, such that each part is responsible for the rotation of a specific axis. In this case, \\(m\\) should be a multiple of \\(N\\). Specifically," + }, + { + "type": "equation", + "bbox": [ + 0.114, + 0.478, + 0.482, + 0.511 + ], + "angle": 0, + "content": "\\[\n\\mathbf {B} _ {i j} = \\left\\{ \\begin{array}{l l} \\mathbf {P} _ {j} - \\mathbf {P} _ {j} ^ {\\top}, & \\text {i f} j \\equiv i (\\mathrm {m o d} N) \\\\ \\mathbf {O}, & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.521, + 0.483, + 0.552 + ], + "angle": 0, + "content": "where \\(\\{\\mathbf{P}_j\\}_{j=1}^m\\) represents a set of trainable matrices, \\(\\mathbf{O}\\) denotes a zero matrix, and \\(\\equiv\\) indicates congruence modulo." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.56, + 0.484, + 0.606 + ], + "angle": 0, + "content": "Proposition 3. A set of matrices \\(\\mathcal{B} = \\{\\mathbf{B}_1, \\mathbf{B}_2, \\dots, \\mathbf{B}_N\\}\\) pairwise commutes if they pairwise linearly dependent, i.e., \\(\\exists \\lambda_1, \\lambda_2, \\dots, \\lambda_N \\in \\mathbb{R}\\) s.t." + }, + { + "type": "equation", + "bbox": [ + 0.18, + 0.616, + 0.482, + 0.632 + ], + "angle": 0, + "content": "\\[\n\\lambda_ {1} \\mathbf {B} _ {1} = \\lambda_ {2} \\mathbf {B} _ {2} = \\dots = \\lambda_ {N} \\mathbf {B} _ {N} \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.642, + 0.484, + 0.705 + ], + "angle": 0, + "content": "Based on Proposition 3, we propose Trainable RoPE parameterized by Linearly-Dependent Angle Matrices (ComRoPE-LD). Specifically, We train a base matrix \\(\\mathbf{P}\\) with scaling factors \\(\\{\\theta_i\\}_{i=1}^N\\). Then we obtain:" + }, + { + "type": "equation", + "bbox": [ + 0.125, + 0.712, + 0.482, + 0.73 + ], + "angle": 0, + "content": "\\[\n\\mathcal {B} = \\left\\{\\mathbf {B} _ {i} = \\theta_ {i} \\left(\\mathbf {P} - \\mathbf {P} ^ {\\top}\\right) \\mid i = 1, 2, \\dots , N \\right\\} \\tag {10}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.748, + 0.483, + 0.78 + ], + "angle": 0, + "content": "3.4. Implementation details of coordinates and improvements" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.787, + 0.375, + 0.802 + ], + "angle": 0, + "content": "3.4.1 Relative scaling and center offset" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Relative scaling. In language models, positions are typically treated as discrete. Additionally, due to the relationship between tokens and the inherent uncertainty in sequence length, there is no need to scale these positions into a specified range. However, in the case of images, positions are continuous. When using different patch sizes, it" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.073, + 0.907, + 0.327 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.347, + 0.907, + 0.378 + ], + "angle": 0, + "content": "Figure 2. Different patch sizes result in different relative relationships." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.404, + 0.906, + 0.464 + ], + "angle": 0, + "content": "becomes unreasonable to define their relative relationships solely based on the span of patches. As shown in Figure 2, the same locations with varying patch sizes can result in significantly different relative relationships." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.465, + 0.907, + 0.606 + ], + "angle": 0, + "content": "As a result, applying relative coordinates is a better method for measuring relative relationships in images. For an image with shape \\( H \\times W \\), we scale both the height \\( H \\) and the width \\( W \\) to 1. Therefore, for a pixel located at \\( (h, w) \\) in the raw image, its coordinate is treated as \\( \\left(\\frac{h}{H}, \\frac{w}{W}\\right) \\). For high-dimensional coordinates, we perform the same operation, i.e., transform a raw coordinate \\( (x_1, x_2, \\dots, x_N) \\) in multi-axial canvas with shape \\( (X_1, X_2, \\dots, X_N) \\) into \\( \\left(\\frac{x_1}{X_1}, \\frac{x_2}{X_2}, \\dots, \\frac{x_N}{X_N}\\right) \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.61, + 0.906, + 0.672 + ], + "angle": 0, + "content": "Center offset. When we project a patch to a feature tensor with a coordinate, we aggregate all information of the patch into a specified location. Simply, we adopt the center point of a patch as the aggregation location." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.694, + 0.719, + 0.709 + ], + "angle": 0, + "content": "3.4.2 Position perturbation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.719, + 0.906, + 0.809 + ], + "angle": 0, + "content": "To achieve better robustness and excellent performance across different scales during inference, we add perturbations to the coordinates of the patches. Specifically, for a patch with center \\((x_{1},\\dots ,x_{N})\\) and size \\((\\Delta X_{1},\\dots ,\\Delta X_{N})\\), during training, we formulate its location as:" + }, + { + "type": "equation", + "bbox": [ + 0.527, + 0.818, + 0.905, + 0.845 + ], + "angle": 0, + "content": "\\[\n\\left. \\mathcal {N} \\left(\\left(x _ {1}, \\dots , x _ {N}\\right) ^ {\\top}; \\operatorname {d i a g} \\left(\\sigma \\Delta X _ {1}, \\dots , \\sigma \\Delta X _ {N}\\right) ^ {2}\\right) \\right. \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.906, + 0.905 + ], + "angle": 0, + "content": "where \\(\\sigma\\) is a hyper-parameter called perturbation intensity. Additionally, we truncate the location within the patch area, i.e., \\(\\left[x_k - \\frac{\\Delta X_k}{2}, x_k + \\frac{\\Delta X_k}{2}\\right]\\)." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4512" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.104, + 0.089, + 0.894, + 0.31 + ], + "angle": 0, + "content": "
Position Encoding MethodPerturbation IntensityEvaluation Resolution
112128192224256320384448512
APE130.0438.6956.458.7660.0259.2757.0454.1050.99
019.7133.4355.9758.6255.3152.6349.6846.3942.66
Vanilla RoPE136.9445.4860.7263.0963.2462.1259.2455.5151.11
036.4144.4859.9762.0362.5461.3658.5654.7950.58
LieRE138.0346.9762.2264.3664.9963.7861.1557.9253.74
038.2246.8562.0164.0764.5463.4660.7456.8952.51
ComRoPE-AP (ours)135.7546.1862.8265.3265.8364.7861.8858.2154.02
037.1746.9562.6364.7665.1164.3561.6258.1053.81
ComRoPE-LD (ours)138.3047.2863.5365.4965.9565.2762.6259.1155.29
036.8846.5462.8965.2765.6664.2761.8357.8753.64
" + }, + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.313, + 0.904, + 0.327 + ], + "angle": 0, + "content": "Table 1. Accuracy of 2D classification on ImageNet. Models are trained at a resolution of \\( {224} \\times {224} \\) and evaluated at varying resolutions." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.352, + 0.223, + 0.37 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.377, + 0.483, + 0.499 + ], + "angle": 0, + "content": "In this section, we evaluate the performance of various positional encoding methods on classic vision tasks. We first assess their scalability in 2D image classification across different resolutions. Additionally, we conduct object detection experiments to demonstrate the generalizability of our approach. To further examine the ability to handle higher-dimensional data, we perform 3D classification experiments, which are detailed in Appendix B." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.507, + 0.253, + 0.523 + ], + "angle": 0, + "content": "4.1. 2D classification" + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.531, + 0.187, + 0.546 + ], + "angle": 0, + "content": "4.1.1 Setup" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.553, + 0.483, + 0.734 + ], + "angle": 0, + "content": "Baselines and model architecture. We evaluate our proposed methods (ComRoPE-LD and ComRoPE-AP) against APE, vanilla RoPE (as introduced by RoFormer), and LieRE. To isolate the effects of positional encoding, we utilize a standard Vision Transformer (ViT-B/16) architecture with minimal modifications. The APE codebook is removed for methods that do not employ APE, and self-attention layers are replaced with RoPE self-attention parameterized by angle matrices. This design highlights the performance differences among various positional encoding methods. A block size of 8 is used in practice. More details are provided in Appendix D." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.735, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Training and evaluation protocol. All models are trained at a standard resolution of \\(224 \\times 224\\) and evaluated across multiple resolutions to test their robustness and scalability on the ImageNet-1K dataset [4]. The models are trained from scratch using randomly initialized parameters, ensuring no influence from pre-trained weights or external priors. To maintain fairness and reproducibility, we apply only basic data augmentation techniques, such as resizing and random cropping, focusing on relative performance comparisons rather than achieving absolute accuracy. The primary evaluation metric is accuracy on the test set across" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.353, + 0.907, + 0.415 + ], + "angle": 0, + "content": "various resolutions. Since APE is inherently fixed and discrete, bilinear interpolation is applied to adapt it to different resolutions during evaluation. Optimization methods and hyper-parameters are detailed in Appendix C." + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.433, + 0.897, + 0.525 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.533, + 0.897, + 0.624 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.629, + 0.907, + 0.687 + ], + "angle": 0, + "content": "Figure 3. Accuracy on ImageNet-1K for various positional encoding methods. The results for the same perturbation intensity are presented together for better comparison. For better visualization, evaluation resolution at \\(112 \\times 112\\) is not included in these figures." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.725, + 0.656, + 0.739 + ], + "angle": 0, + "content": "4.1.2 Main results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.906, + 0.809 + ], + "angle": 0, + "content": "The accuracy metrics for each positional encoding method across various resolutions are summarized in Table 1. Additionally, Figure 3 visually compares performance under different levels of perturbation intensity separately." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Overall performance. Across all evaluations, APE consistently exhibits the lowest accuracy, corroborating previous findings regarding its limitations in dynamic contexts. The vanilla RoPE shows a modest improvement over APE but remains less effective. In contrast, the trainable angle matrices, namely LieRE, ComRoPE-AP, and ComRoPE-LD," + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4513" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.168 + ], + "angle": 0, + "content": "demonstrate significantly higher accuracy across all resolutions. Notably, ComRoPE-LD achieves the best performance among the three, suggesting that its inherent linear dependencies may enhance flexibility and structural learning capabilities." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.17, + 0.482, + 0.291 + ], + "angle": 0, + "content": "Accuracy at training resolution. At the training resolution of \\(224 \\times 224\\), all three methods with trainable angle matrices (ComRoPE-LD, ComRoPE-AP, and LieRE) achieve comparable accuracy, significantly outperforming both APE and the standard RoPE, which underscores the effectiveness of RoPE parameterized by trainable angle matrices. Notably, ComRoPE-LD surpasses the current state-of-the-art LieRE by \\(1.6\\%\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.294, + 0.482, + 0.415 + ], + "angle": 0, + "content": "Scaling to higher resolution. At resolutions beyond the training size, LieRE shows the steepest decline in accuracy among the three trainable RoPE variants, indicating greater sensitivity to resolution changes. In contrast, ComRoPE-LD and ComRoPE-AP exhibit a more gradual decrease in performance, thanks to their commutative properties that enhance positional robustness. Specifically, ComRoPE-LD outperforms LieRE by \\(2.9\\%\\) at a resolution of \\(512 \\times 512\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.418, + 0.482, + 0.552 + ], + "angle": 0, + "content": "Significance of commutativity. These findings illustrate the effectiveness of trainable commutative angle matrices, particularly ComRoPE-LD, in maintaining accuracy and scalability across diverse resolutions. The results underscore the importance of commutativity in ensuring robust RoPE parameterized by trainable angle matrices for vision tasks. Furthermore, to better understand the role of commutativity, we conduct additional experiments by introducing coordinate offsets in LieRE (see Section 4.3.1 for details)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.572, + 0.254, + 0.587 + ], + "angle": 0, + "content": "4.2. Object detection" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.598, + 0.482, + 0.674 + ], + "angle": 0, + "content": "To demonstrate the generalizability of our approach, we conduct object detection experiments using the framework from [32]. We adopt ViT-S as our backbone and apply ComRoPE to the attention layers. To ensure consistency with the pre-trained model, we initialize the angle matrix to zero." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.676, + 0.482, + 0.766 + ], + "angle": 0, + "content": "We evaluate ComRoPE-LD, LieRE, and APE on the MS COCO dataset [14]. As summarized in Table 2, both ComRoPE-LD and LieRE outperform APE, with ComRoPE-LD achieving slightly better performance than LieRE while only using nearly half the number of extra parameters." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.77, + 0.482, + 0.83 + ], + "angle": 0, + "content": "To compare training efficiency, we plot the results for each epoch in Figure 4. Our findings indicate that both ComRoPE-LD and LieRE converge faster than APE, requiring \\(29\\%\\) fewer iterations to achieve the same results as APE." + }, + { + "type": "table", + "bbox": [ + 0.52, + 0.089, + 0.9, + 0.158 + ], + "angle": 0, + "content": "
PE MethodAPAP50AP75APsAPmAPl
APE44.066.647.728.246.858.4
LieRE44.567.348.429.046.958.7
ComRoPE-LD44.767.648.529.247.160.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.558, + 0.161, + 0.861, + 0.174 + ], + "angle": 0, + "content": "Table 2. Results of object detection on MS COCO." + }, + { + "type": "image", + "bbox": [ + 0.524, + 0.19, + 0.868, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.553, + 0.31, + 0.865, + 0.324 + ], + "angle": 0, + "content": "Figure 4. Results over the whole training procedure." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.337, + 0.664, + 0.353 + ], + "angle": 0, + "content": "4.3. Ablation study" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.36, + 0.741, + 0.376 + ], + "angle": 0, + "content": "4.3.1 Impact of commutativity" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.383, + 0.905, + 0.472 + ], + "angle": 0, + "content": "To evaluate the significance of the commutativity of angle matrices, we conduct experiments on the LieRE by introducing a coordinate offset. Specifically, before inference, a random offset is applied uniformly across all coordinates within the image. The offset is sampled from a Gaussian distribution as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.644, + 0.484, + 0.905, + 0.503 + ], + "angle": 0, + "content": "\\[\n\\mathcal {N} (0; \\rho^ {2} \\cdot \\mathbf {I} _ {N \\times N}) \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.512, + 0.905, + 0.602 + ], + "angle": 0, + "content": "where \\(\\rho\\) represents the standard deviation of the random offset, and \\(\\mathbf{I}_{N\\times N}\\) is the identity matrix of size \\(N\\times N\\). It is important to note that applying the same offset to all coordinates does not influence the relative positional dependencies between the patches. The other experimental settings remain consistent with those in the 2D classification tasks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.603, + 0.906, + 0.754 + ], + "angle": 0, + "content": "The results shown in Figure 5 demonstrate that the baseline model's performance deteriorates significantly as the standard deviation of the offset increases. In contrast, our proposed model, ComRoPE, maintains consistent performance across all offset values. This is due to the commutativity of the angle matrices, which allows the model to remain invariant to such coordinate shifts. The robustness of ComRoPE to this type of perturbation highlights its capacity to preserve relative positional information, even in the presence of modification introduced by coordinate offsets." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.772, + 0.707, + 0.787 + ], + "angle": 0, + "content": "4.3.2 Impact of block size" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.906, + 0.901 + ], + "angle": 0, + "content": "In this section, we examine the impact of block size on 2D classification using the ImageNet dataset. We maintain the same experimental setup as in our primary experiments, varying the block size from 2 to 8. The results are presented in Figure 6. Our findings indicate that larger block sizes consistently improve performance by extending the rotation transformation space to a more significant subgroup of the" + }, + { + "type": "page_footnote", + "bbox": [ + 0.09, + 0.85, + 0.483, + 0.901 + ], + "angle": 0, + "content": "2Relative improvement is calculated as target \\(=\\) baseline \\(\\times\\) (1 + improvement), where the improvement denotes the percentage increase over the baseline performance. This formula applies throughout the following sections." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4514" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.101, + 0.096, + 0.474, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.22, + 0.483, + 0.276 + ], + "angle": 0, + "content": "Figure 5. Effect of coordinate offset on LieRE. As the standard deviation of the offset increases, the performance of the baseline model deteriorates, while ComRoPE remains unaffected (un-painted)." + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.287, + 0.476, + 0.348 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.26, + 0.35, + 0.315, + 0.361 + ], + "angle": 0, + "content": "(a) LieRE" + }, + { + "type": "image", + "bbox": [ + 0.098, + 0.37, + 0.476, + 0.43 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.239, + 0.432, + 0.337, + 0.442 + ], + "angle": 0, + "content": "(b) ComRoPE-AP" + }, + { + "type": "image", + "bbox": [ + 0.098, + 0.452, + 0.476, + 0.512 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.239, + 0.513, + 0.337, + 0.524 + ], + "angle": 0, + "content": "(c) ComRoPE-LD" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.527, + 0.483, + 0.556 + ], + "angle": 0, + "content": "Figure 6. Accuracy on ImageNet for various block sizes. Larger block size results in better performance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.57, + 0.483, + 0.706 + ], + "angle": 0, + "content": "particular orthogonal group by introducing additional parameters and computation time. When the block size is small, the associated costs are minimal. However, as the block size increases, the primary term of time complexity grows to \\( O(Lndb^{2}) \\) from \\( O(Lnd\\cdot \\frac{d}{h}) \\), which becomes significant. Therefore, we limit the block size to a maximum of 8 to balance performance with additional costs. In other words, we select a block size of 8 to optimize performance while keeping the extra computational cost manageable." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.726, + 0.36, + 0.741 + ], + "angle": 0, + "content": "4.3.3 Utility of position perturbation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.81 + ], + "angle": 0, + "content": "In this section, we explore the impact of positional perturbations. We conducted experiments on ComRoPE-LD and APE using the ImageNet dataset, with the results presented in Figure 7." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.483, + 0.902 + ], + "angle": 0, + "content": "As shown in Figure 7a, APE is highly sensitive to positional perturbations, leading to significant performance improvement \\((+19.5\\%)\\) when increasing intensity from 0 to 1) when these perturbations are introduced. For RoPE with angle matrices shown in Figure 7b, positional perturbations also resulted in some performance gains \\((+2.9\\%)\\), though" + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.095, + 0.897, + 0.17 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.688, + 0.172, + 0.734, + 0.184 + ], + "angle": 0, + "content": "(a) APE" + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.194, + 0.897, + 0.27 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.662, + 0.272, + 0.761, + 0.284 + ], + "angle": 0, + "content": "(b) ComRoPE-LD" + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.287, + 0.905, + 0.315 + ], + "angle": 0, + "content": "Figure 7. Accuracy on ImageNet over varying position perturbation intensity." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.329, + 0.905, + 0.39 + ], + "angle": 0, + "content": "the improvement was relatively modest. This is likely due to the inherent robustness of the RoPE design with angle matrices, which is already well-equipped to handle variations in position." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.399, + 0.647, + 0.415 + ], + "angle": 0, + "content": "4.4. Applications" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.422, + 0.907, + 0.634 + ], + "angle": 0, + "content": "In our approach, when the angle matrix is an all-zero matrix, the rotation matrix becomes the identity matrix, causing RoPE Attention to reduce to the standard Attention mechanism. When the block size of the angle matrix is set to 2, ComRoPE-AP effectively reduces to the commonly used RoPE Attention in language models. This demonstrates that our method can represent standard Attention and various common variants of RoPE Attention. Therefore, during the fine-tuning stage, we can replace standard Attention with our method, load the pre-trained weights, and fine-tune them under the new paradigm. In other words, ComRoPE can be seamlessly integrated into the fine-tuning process, even if it was not applied during pre-training. Additional experiments can be found in Appendix B." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.647, + 0.634, + 0.662 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.672, + 0.907, + 0.899 + ], + "angle": 0, + "content": "In this work, we proposed ComRoPE, a novel and adaptive framework for Rotary Position Embedding (RoPE) parameterized by trainable angle matrices. We rigorously formulate the RoPE Equation and then establish a necessary and sufficient condition for its solution. Our approach effectively overcomes the scalability and robustness limitations of existing RoPE methods by eliminating the need for manually designed rotation matrices and introducing a more flexible, scalable solution. Extensive experimental results show that ComRoPE outperforms the existing positional encoding methods across various tasks. Furthermore, our framework generalizes existing RoPE formulations and demonstrates the potential for broader application in Transformer models, offering insights and a solid foundation for future research in positional encoding techniques." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4515" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.091, + 0.241, + 0.108 + ], + "angle": 0, + "content": "Acknowledgment" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.116, + 0.482, + 0.191 + ], + "angle": 0, + "content": "This work is supported by the National Key R&D Program of China (2022YFB4701400/4701402), SSTIC Grant (KJZD20230923115106012, KJZD20230923114916032, GJHZ20240218113604008) and Beijing Key Lab of Networked Multimedia." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.206, + 0.187, + 0.222 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.232, + 0.482, + 0.259 + ], + "angle": 0, + "content": "[1] Tom B Brown. Language models are few-shot learners. arXiv preprint arXiv:2005.14165, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.262, + 0.483, + 0.33 + ], + "angle": 0, + "content": "[2] Yirui Chen, Xudong Huang, Quan Zhang, Wei Li, Mingjian Zhu, Qiangyu Yan, Simiao Li, Hanting Chen, Hailin Hu, Jie Yang, et al. Gim: A million-scale benchmark for generative image manipulation detection and localization. arXiv preprint arXiv:2406.16531, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.332, + 0.483, + 0.413 + ], + "angle": 0, + "content": "[3] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhanghao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E Gonzalez, et al. Vicuna: An open-source chatbot impressing gpt-4 with \\(90\\%\\) * chatgpt quality, march 2023. URL https://lmsys.org/blog/2023-03-30-vicuna, 3(5), 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.417, + 0.483, + 0.472 + ], + "angle": 0, + "content": "[4] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, K. Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 248-255, 2009. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.475, + 0.482, + 0.543 + ], + "angle": 0, + "content": "[5] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In North American Chapter of the Association for Computational Linguistics, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.546, + 0.482, + 0.585 + ], + "angle": 0, + "content": "[6] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.589, + 0.482, + 0.641 + ], + "angle": 0, + "content": "[7] Yuxin Fang, Quan Sun, Xinggang Wang, Tiejun Huang, Xinlong Wang, and Yue Cao. Eva-02: A visual representation for neon genesis. Image and Vision Computing, 149:105171, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.646, + 0.482, + 0.713 + ], + "angle": 0, + "content": "[8] Team GLM, Aohan Zeng, Bin Xu, Bowen Wang, Chenhui Zhang, Da Yin, Dan Zhang, Diego Rojas, Guanyu Feng, Hanlin Zhao, et al. Chatglm: A family of large language models from glm-130b to glm-4 all tools. arXiv preprint arXiv:2406.12793, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.717, + 0.482, + 0.757 + ], + "angle": 0, + "content": "[9] Byeongho Heo, Song Park, Dongyoon Han, and Sangwoo Yun. Rotary position embedding for vision transformer. arXiv preprint arXiv:2403.13298, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.76, + 0.482, + 0.801 + ], + "angle": 0, + "content": "[10] Max Horn, Kumar Shridhar, Elrich Groenewald, and Philipp FM Baumann. Translational equivariance in kernelizable attention. arXiv preprint arXiv:2102.07680, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.804, + 0.482, + 0.87 + ], + "angle": 0, + "content": "[11] Amirhossein Kazemnejad, Inkit Padhi, Karthikeyan Natesan Ramamurthy, Payel Das, and Siva Reddy. The impact of positional encoding on length generalization in transformers. Advances in Neural Information Processing Systems, 36: 24892-24928, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.874, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[12] Hanyu Lai, Xiao Liu, Iat Long Iong, Shuntian Yao, Yuxuan Chen, Pengbo Shen, Hao Yu, Hanchen Zhang, Xiaohan" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.232, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.548, + 0.093, + 0.905, + 0.148 + ], + "angle": 0, + "content": "Zhang, Yuxiao Dong, et al. Autowebglm: A large language model-based web navigating agent. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 5295-5306, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.149, + 0.905, + 0.216 + ], + "angle": 0, + "content": "[13] Hongyu Li, Jinyu Chen, Ziyu Wei, Shaofei Huang, Tianrui Hui, Jialin Gao, Xiaoming Wei, and Si Liu. Llava-st: A multimodal large language model for fine-grained spatial-temporal understanding. arXiv preprint arXiv:2501.08282, 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.219, + 0.905, + 0.286 + ], + "angle": 0, + "content": "[14] Tsung-Yi Lin, Michael Maire, Serge Belongie, Lubomir Bourdev, Ross Girshick, James Hays, Pietro Perona, Deva Ramanan, C. Lawrence Zitnick, and Piotr Dollar. Microsoft coco: Common objects in context. arXiv preprint arXiv: 1405.0312, 2014. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.289, + 0.905, + 0.37 + ], + "angle": 0, + "content": "[15] Xiao Liu, Hanyu Lai, Hao Yu, Yifan Xu, Aohan Zeng, Zhengxiao Du, Peng Zhang, Yuxiao Dong, and Jie Tang. Webglm: towards an efficient web-enhanced question answering system with human preferences. In Proceedings of the 29th ACM SIGKDD conference on knowledge discovery and data mining, pages 4549-4560, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.372, + 0.905, + 0.439 + ], + "angle": 0, + "content": "[16] Xiao Liu, Hao Yu, Hanchen Zhang, Yifan Xu, Xuanyu Lei, Hanyu Lai, Yu Gu, Hangliang Ding, Kaiwen Men, Kejuan Yang, et al. Agentbench: Evaluating llms as agents. In The Twelfth International Conference on Learning Representations, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.442, + 0.905, + 0.509 + ], + "angle": 0, + "content": "[17] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.512, + 0.905, + 0.579 + ], + "angle": 0, + "content": "[18] Antoine Liutkus, Ondrej Cifka, Shih-Lun Wu, Umut Simsekli, Yi-Hsuan Yang, and Gael Richard. Relative positional encoding for transformers with linear complexity. In International Conference on Machine Learning, pages 7067-7079. PMLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.581, + 0.905, + 0.663 + ], + "angle": 0, + "content": "[19] Jiasen Lu, Christopher Clark, Sangho Lee, Zichen Zhang, Savya Khosla, Ryan Marten, Derek Hoiem, and Aniruddha Kembhavi. Unified-io 2: Scaling autoregressive multimodal models with vision language audio and action. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26439–26455, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.665, + 0.905, + 0.718 + ], + "angle": 0, + "content": "[20] Sophie Ostmeier, Brian Axelrod, Michael E Moseley, Akshay Chaudhari, and Curtis Langlotz. Liere: Generalizing rotary position encodings. arXiv preprint arXiv:2406.10322, 2024. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.721, + 0.905, + 0.761 + ], + "angle": 0, + "content": "[21] Ofir Press, Noah A Smith, and Mike Lewis. Train short, test long: Attention with linear biases enables input length extrapolation. arXiv preprint arXiv:2108.12409, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.763, + 0.905, + 0.789 + ], + "angle": 0, + "content": "[22] A Radford. Improving language understanding by generative pre-training. 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.79, + 0.905, + 0.831 + ], + "angle": 0, + "content": "[23] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.833, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[24] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of machine learning research, 21(140):1-67, 2020. 2" + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.093, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4516" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.161 + ], + "angle": 0, + "content": "[25] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of machine learning research, 21(140):1-67, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.164, + 0.482, + 0.207 + ], + "angle": 0, + "content": "[26] Peter Shaw, Jakob Uszkoreit, and Ashish Vaswani. Self-attention with relative position representations. arXiv preprint arXiv:1803.02155, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.208, + 0.482, + 0.261 + ], + "angle": 0, + "content": "[27] Jianlin Su, Murtadha Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu. Roformer: Enhanced transformer with rotary position embedding. Neurocomputing, 568:127063, 2024. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.263, + 0.482, + 0.304 + ], + "angle": 0, + "content": "[28] Sainbayar Sukhbaatar, Jason Weston, Rob Fergus, et al. End-to-end memory networks. Advances in neural information processing systems, 28, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.305, + 0.482, + 0.375 + ], + "angle": 0, + "content": "[29] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.376, + 0.482, + 0.404 + ], + "angle": 0, + "content": "[30] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.405, + 0.482, + 0.473 + ], + "angle": 0, + "content": "[31] Kan Wu, Houwen Peng, Minghao Chen, Jianlong Fu, and Hongyang Chao. Rethinking and improving relative position encoding for vision transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10033-10041, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.476, + 0.482, + 0.532 + ], + "angle": 0, + "content": "[32] Chunlong Xia, Xinliang Wang, Feng Lv, Xin Hao, and Yifeng Shi. Vit-comer: Vision transformer with convolutional multi-scale feature interaction for dense predictions. Computer Vision and Pattern Recognition, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.532, + 0.482, + 0.573 + ], + "angle": 0, + "content": "[33] Quan Zhang and Yuxin Qi. Can mllms guide weakly-supervised temporal action localization tasks? arXiv preprint arXiv:2411.08466, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.575, + 0.482, + 0.657 + ], + "angle": 0, + "content": "[34] Quan Zhang, Xiaoyu Liu, Wei Li, Hanting Chen, Junchao Liu, Jie Hu, Zhiwei Xiong, Chun Yuan, and Yunhe Wang. Distilling semantic priors from sam to efficient image restoration models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25409-25419, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.659, + 0.482, + 0.729 + ], + "angle": 0, + "content": "[35] Quan Zhang, Yuxin Qi, Xi Tang, Jinwei Fang, Xi Lin, Ke Zhang, and Chun Yuan. IMDPrompter: Adapting SAM to image manipulation detection by cross-view automated prompt learning. In The Thirteenth International Conference on Learning Representations, 2025." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.729, + 0.482, + 0.798 + ], + "angle": 0, + "content": "[36] Quan Zhang, Yuxin Qi, Xi Tang, Rui Yuan, Xi Lin, Ke Zhang, and Chun Yuan. Rethinking pseudo-label guided learning for weakly supervised temporal action localization from the perspective of noise correction. arXiv preprint arXiv:2501.11124, 2025. 1" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.798 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4517" + } + ] +] \ No newline at end of file diff --git a/2025/ComRoPE_ Scalable and Robust Rotary Position Embedding Parameterized by Trainable Commuting Angle Matrices/287888de-80a8-43c2-8d0f-f61f04d80e4e_origin.pdf b/2025/ComRoPE_ Scalable and Robust Rotary Position Embedding Parameterized by Trainable Commuting Angle Matrices/287888de-80a8-43c2-8d0f-f61f04d80e4e_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..8039f25e72d671dc0f60183585dcc9b555b23684 --- /dev/null +++ b/2025/ComRoPE_ Scalable and Robust Rotary Position Embedding Parameterized by Trainable Commuting Angle Matrices/287888de-80a8-43c2-8d0f-f61f04d80e4e_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:db310114861a87a90f9d6013b81bfd3a4aac2761714568bdfdb4c26fa0103e40 +size 2983343 diff --git a/2025/ComRoPE_ Scalable and Robust Rotary Position Embedding Parameterized by Trainable Commuting Angle Matrices/full.md b/2025/ComRoPE_ Scalable and Robust Rotary Position Embedding Parameterized by Trainable Commuting Angle Matrices/full.md new file mode 100644 index 0000000000000000000000000000000000000000..0cf5cef6b794580f2f782f000f69ff37f78e3d42 --- /dev/null +++ b/2025/ComRoPE_ Scalable and Robust Rotary Position Embedding Parameterized by Trainable Commuting Angle Matrices/full.md @@ -0,0 +1,362 @@ +# ComRoPE: Scalable and Robust Rotary Position Embedding Parameterized by Trainable Commuting Angle Matrices + +Hao $\mathbf{Y}\mathbf{u}^{1}$ Tangyu Jiang $^{1\dagger}$ Shuning Jia $^{1,2}$ Shannon Yan $^{1}$ Shunning Liu $^{1}$ + +Haolong Qian1 Guanghao Li1 Shuting Dong1 Chun Yuan1† + +$^{1}$ Tsinghua University $^{2}$ Shenzhen University + +longinyh@gmail.com, jiangtangyu@sz.tsinghua.edu.cn, yuanc@sz.tsinghua.edu.cn + +# Abstract + +The Transformer architecture has revolutionized various fields since it was proposed, where positional encoding plays an essential role in effectively capturing sequential order and context. Therefore, Rotary Positional Encoding (RoPE) was proposed to alleviate these issues, which integrates positional information by rotating the embeddings in the attention mechanism. However, RoPE utilizes manually defined rotation matrices, a design choice that favors computational efficiency but limits the model's flexibility and adaptability. In this work, we propose ComRoPE, which generalizes RoPE by defining it in terms of trainable commuting angle matrices. Specifically, we demonstrate that pairwise commutativity of these matrices is essential for RoPE to achieve scalability and positional robustness. We formally define the RoPE Equation, which is an essential condition that ensures consistent performance with position offsets. Based on the theoretical analysis, we present two types of trainable commuting angle matrices as sufficient solutions to the RoPE equation, which significantly improve performance, surpassing the current state-of-the-art method by $1.6\%$ at training resolution and $2.9\%$ at higher resolution on the ImageNet-1K dataset. Furthermore, our framework shows versatility in generalizing to existing RoPE formulations and offering new insights for future positional encoding research. To ensure reproducibility, the source code and instructions are available at https://github.com/Longin-Yu/ComRoPE. + +# 1. Introduction + +The Transformer architecture [30] has been widely adopted across various fields, including Natural Language Processing (NLP) [1, 5, 8, 22, 23] and Computer Vision (CV) [6]. Moreover, an increasing number of Transformer-based applications [2, 12, 13, 15, 16, 33-36] continue to demonstrate + +their effectiveness across a wide range of domains. At the core of the Transformer model lies the attention mechanism, which enables the model to selectively focus on different parts of the input based on relevance, thereby effectively capturing long-range dependencies and contextual relationships. + +However, since the attention mechanism is insensitive to the fundamental position information, it cannot inherently capture the order of elements in the data. In order to alleviate this issue, positional embeddings are added to the input representation, providing the model with the necessary positional information to process the sequence order and structure effectively. + +A variety of positional encoding methods have been proposed, which can be divided into two categories: Absolute Positional Encoding (APE) and Relative Positional Encoding (RPE). APE explicitly encodes the absolute position of each token by generating fixed positional embeddings, which are directly added to the input embeddings at the beginning of training [30]. However, APE struggles with handling long sequences and exhibits high sensitivity to positional shifts, limiting the scalability and robustness of models [11]. In contrast, RPE does not directly modify input embeddings; instead, it incorporates relative positional information within the attention mechanism, enabling the model to effectively capture positional relationships between tokens. Among all the existing works, Rotary Position Embedding (RoPE) [27] has gained significant attention due to the advantage of applying a rotational transformation to token embeddings. It encodes relative positions by treating each pair of features as coordinates and rotating them by an angle proportional to their position, enabling all tokens to interact within the attention mechanism regardless of distance. + +However, existing RoPE methods face several key challenges: i) The essential components (i.e., the RoPE matrices) of previous RoPE approaches rely on 2D rotation + +groups, which simplify computations but consequently restrict their feature projection capabilities, especially in high-dimensional spaces [27]. ii) Moreover, the majority of the rotation matrices of RoPE require to be manually designed, leading to insufficient capability and suboptimal performance[21, 25]. iii) Finally, previous attempts [20] to extend the rotation group often prioritize design simplicity, making it difficult to consistently satisfy relative position dependency—a critical property of RoPE that ensures positional robustness against offsets. + +Our objectives. This work aims to develop a novel RoPE method for Transformers that is both scalable and robust. Specifically, we seek to extend the rotation group from the existing 2D representation to a larger subgroup of the special orthogonal group, allowing for higher degrees of freedom while preserving consistent behavior with respect to position offsets. Unlike existing methods that rely on manually designed non-trainable rotation matrices, which suffer from limited expressiveness and reduced robustness, our approach is designed to offer richer feature representation capabilities. This framework addresses the scalability limitations of current approaches and enhances their robustness against positional transformations. + +Our contributions. This work introduces ComRoPE, a novel framework that significantly enhances positional encoding in Transformers. ComRoPE leverages trainable angle matrices, extending the RoPE mechanism with higher scalability and robustness. We identify the pairwise commutativity of these matrices as a necessary and sufficient condition for effective positional encoding, thereby unifying various existing RoPE formulations under a single theoretical framework. The contributions are summarized as follows: + +- We formally define the RoPE function parameterized by angle matrices and prove that pairwise commutativity is a necessary and sufficient condition, offering a unified theory that encompasses several existing RoPE variants. +- We introduce ComRoPE, a scalable and robust solution that leverages two types of trainable commuting angle matrix sets as sufficient solutions to the RoPE equation, capturing richer positional representations without the need for manual design. +- Our extensive experiments show that ComRoPE surpasses the current state-of-the-art LieRE, achieving a performance increase of $1.6\%$ at training resolution and $2.9\%$ at higher resolutions on the ImageNet-1K classification task while delivering strong results across other benchmarks. +- We explore further applications of ComRoPE, providing valuable insights for advancing position encoding techniques in future Transformer-based models. + +# 2. Related work + +# 2.1. Position information in attention + +Transformers [30] utilize the attention mechanism to capture similarities within sequences. However, they lack inherent sequential information and cannot capture the positional information of each token. To address this limitation, positional encoding [5, 24, 26] was introduced. As research progressed, positional encoding generally evolved into two types: APE [5, 28, 30] and RPE [10, 18, 26, 27]. [30] first proposed using APE in the form of sine and cosine functions, effectively representing the positional relationships within the input sequence. This positional encoding method achieved remarkable results in natural language processing, becoming a foundational component for many NLP tasks. However, the fixed nature of positional encoding limited the model's ability to generalize to longer input sequences. In response, subsequent research introduced learnable absolute positional encoding. [5] proposed further enhancing the model's expressive power by incorporating learnable position embeddings, particularly excelling in tasks such as sentence alignment and context representation. Although APE provides positional information to enhance the model's understanding of sequences, it shows limitations in handling long sequences and cross-sequence scenarios. Hierarchical ViT such as Swin Transformer [17], introduced Relative Position Bias (RPB) [17, 25, 26] to handle large numbers of tokens with limited positional embeddings [9]. Related research has explored alternative encoding methods, such as RPE, as a replacement for APE to better capture complex dependency structures. iRPE [31] proposed an improved RPB by combining relative position embedding. + +# 2.2. Rotary position embedding + +Building on these developments, RoFormer [27] combined the advantages of APE and RPE, proposing RoPE. Currently, RoPE is widely used in Large Language Models (LLMs), such as LLaMA [29] and Vicuna [3]. This approach enhances model performance on tasks involving long-text semantics and multi-turn dialogues, improving extrapolation capability. To better adapt to the characteristics of two-dimensional data such as images, researchers extended RoPE to two-dimensional sequences (2D-RoPE). LieRE [20] extends RoPE to a more generalized form by introducing a rotation-based positional encoding method grounded in Lie group theory. Unified-IO 2 [19] applies 2D-RoPE within its multimodal architecture; EVA-02 [7], FiT [7] these pioneering works used 2D RoPE with axial frequencies (2D Axial RoPE), but had limitations in processing in the diagonal direction. Therefore, RoPE for ViT [9] proposes to use mixed axial frequency for 2D RoPE, named RoPE-Mixed. + +# 3. Method + +In this section, we begin by introducing key definitions and reformulating the RoPE paradigm within multi-axial attention mechanisms, which we collectively refer to as the RoPE Equations. Next, we present our main theorem, which establishes the necessary and sufficient conditions for RoPE functions parameterized by angle matrices. Finally, we provide solutions to the RoPE Equations based on the propositions that outline the sufficient conditions. + +# 3.1. Preliminaries + +We use $\mathbf{R}(\cdot):\mathbb{R}^N\to \mathbb{R}^{d\times d}$ to denote a matrix-value function. We denote the vectors and matrices as the lower and uppercase bold font, respectively. We first recall two fundamental definitions of matrices as follows: + +Definition 1 (Matrix Exponential). The exponential of a square matrix $\mathbf{A}$ , denoted as $e^{\mathbf{A}}$ or $\exp (\mathbf{A})$ , is defined using the matrix exponential series such that: + +$$ +e ^ {\mathbf {A}} = \exp (\mathbf {A}) = \sum_ {k = 0} ^ {\infty} \frac {\mathbf {A} ^ {k}}{k !} = \mathbf {I} + \mathbf {A} + \frac {\mathbf {A} ^ {2}}{2 !} + \frac {\mathbf {A} ^ {3}}{3 !} + \dots , +$$ + +This series converges for any square matrix $\mathbf{A}$ . + +Definition 2 (Commuting Matrices). Two square matrices $\mathbf{A}_1$ and $\mathbf{A}_2$ are said to commute if their product is independent of the order of multiplication, i.e., + +$$ +\mathbf {A} _ {1} \mathbf {A} _ {2} = \mathbf {A} _ {2} \mathbf {A} _ {1}. +$$ + +A set of square matrices $\{\mathbf{A}_1,\mathbf{A}_2,\dots ,\mathbf{A}_N\}$ is said to pairwise commute if every pair of matrices within the set commutes with each other. That is, for all $i,j$ such that $1\leq i,j\leq N$ + +$$ +\mathbf {A} _ {i} \mathbf {A} _ {j} = \mathbf {A} _ {j} \mathbf {A} _ {i}. +$$ + +To clarify and better illustrate the main theorems presented in the following sections, we first reformulate and unify the definitions of the RoPE paradigm, specifically in the context of multi-axial attention mechanisms. RoPE was initially proposed by Su et al. [27] as a positional encoding method based on relative position dependencies. However, previous work provided only conceptual and descriptive descriptions of RPE and RoPE without offering a rigorous formal definition. In this work, we provide the formal definitions of both RPE and RoPE. + +Definition 3 (RPE Equation). Let $f: \mathbb{R}^d \times \mathbb{R}^N \to \mathbb{R}^d$ be a positional encoding function. $f$ is said to be a RPE function if and only if there exists a function $\rho: \mathbb{R}^d \times \mathbb{R}^d \to \mathbb{R}$ , and a function $g: \mathbb{R}^d \times \mathbb{R}^d \times \mathbb{R}^N \to \mathbb{R}$ such that the following conditions hold for all $x, y \in \mathbb{R}^N$ and $q, k \in \mathbb{R}^d$ : + +$$ +g (\boldsymbol {q}, \boldsymbol {k}, \boldsymbol {x} - \boldsymbol {y}) = \rho (f (\boldsymbol {q}, \boldsymbol {x}), f (\boldsymbol {k}, \boldsymbol {y})), \tag {1} +$$ + +We refer to Eq. (1) as the RPE Equation. + +Definition 4 (RoPE Equation). Let $f: \mathbb{R}^d \times \mathbb{R}^N \to \mathbb{R}^d$ be an RPE function. $f$ is said to be a RoPE function if and only if there exists a matrix-valued function $\mathbf{R}_f(\cdot)$ such that the following conditions hold for all $\mathbf{x}, \mathbf{y} \in \mathbb{R}^N$ and $\mathbf{k}, \mathbf{k} \in \mathbb{R}^d$ in RPE Equation: + +$$ +\left\{ \begin{array}{l} f (\boldsymbol {q}, \boldsymbol {x}) = \mathbf {R} _ {f} (\boldsymbol {x}) \boldsymbol {q} \\ \rho (\boldsymbol {q}, \boldsymbol {k}) = \boldsymbol {q} ^ {\top} \boldsymbol {k} \\ g (\boldsymbol {q}, \boldsymbol {k}, \boldsymbol {x} - \boldsymbol {y}) = \boldsymbol {q} ^ {\top} \mathbf {R} _ {f} (\boldsymbol {y} - \boldsymbol {x}) \boldsymbol {k} \end{array} \right. \tag {2} +$$ + +We refer to Eq. (2) as the RoPE Equation. + +By substituting the RoPE equation into the RPE equation, we obtain that the RoPE function satisfies the following property. + +Proposition 1. $f$ is said to be a RoPE function if and only if the matrix-valued function $\mathbf{R}_f(\cdot)$ satisfies: + +$$ +\mathbf {R} _ {f} (\boldsymbol {x}) ^ {\top} \mathbf {R} _ {f} (\boldsymbol {y}) = \mathbf {R} _ {f} (\boldsymbol {y} - \boldsymbol {x}), +$$ + +for all $\pmb {x},\pmb {y}\in \mathbb{R}^{N}$ + +Note that the definition of RoPE Equation demonstrates that the position encoding should only be dependent on the relative position of the tokens, which thus be robust against the offset operations. We further expand the definition of the RoPE function to a parameterized one (i.e., Definition 6) via rotation matrices (i.e., Definition 5) as follows1. + +Definition 5 (Parameterized Rotation Matrix). Let $\mathbf{R}(\cdot ;\mathcal{A}): \mathbb{R}^N \to \mathbb{R}^{d\times d}$ be a matrix-valued function parameterized by $N$ skew-symmetric matrices $\mathcal{A} = \{\mathbf{A}_1,\mathbf{A}_2,\dots ,\mathbf{A}_N\}$ . We say that $\mathbf{R}(\cdot ;\mathcal{A})$ is a rotation matrix function parameterized by angle matrices $\mathcal{A}$ if it can be expressed as: + +$$ +\mathbf {R} (\boldsymbol {x}; \mathcal {A}) = \exp \left(\sum_ {i = 1} ^ {N} \mathbf {A} _ {i} x _ {i}\right), \tag {3} +$$ + +where $\pmb{x} = (x_{1}, x_{2}, \ldots, x_{N}) \in \mathbb{R}^{N}$ and $\exp(\cdot)$ denotes the matrix exponential. + +Definition 6 (Parameterized RoPE Function). Let $f: \mathbb{R}^d \times \mathbb{R}^N \to \mathbb{R}^d$ be a RoPE function. $f$ is said to be parameterized by angle matrices if and only if there exists a rotation matrix function $\mathbf{R}_f(\cdot; \mathcal{A})$ parameterized by angle matrices $\mathcal{A}$ such that the RoPE Equation (i.e., Eq. (2)) holds. + +In this case, $\mathbf{R}_f(\cdot ;\mathcal{A})$ is referred to as the rotation matrix function of RoPE function $f(\cdot ;\mathcal{A})$ parameterized by angle matrices $\mathcal{A}$ . For simplicity, we slightly abuse the notation and refer to $\mathbf{R}$ as the rotation matrix of RoPE function $f$ . + +![](images/65b5ac4bfd70910ee3a5a462d8ee4798cac3086e6d8b276c4d6a36bb8a61650f.jpg) +Figure 1. Overview of ComRoPE. Features are arranged into several blocks, each representing a distinct point in the feature space. The positions, along with the angle matrices, define the rotation matrix, which is an element of the special orthogonal group. The rotation transformation projects a feature point onto another point on the surface of the same hypersphere. + +# 3.2. Main theorems + +Based on the formal definitions above, we present our key theoretical results in the following theorem. + +Theorem 1. Let $\mathbf{R}(\cdot ;\mathcal{A}): \mathbb{R}^N \to \mathbb{R}^{d\times d}$ ( $N > 1$ ) be a rotation matrix function parameterized by angle matrices $\mathcal{A}$ . The rotation difference $\mathbf{R}(\pmb{x})^\top \mathbf{R}(\pmb{y})$ can be represented by the location difference $\pmb{y} - \pmb{x}$ if and only if $\mathcal{A}$ pairwise commute. + +The proof of Theorem 1 is shown in Appendix A. + +Theorem 1 together with Proposition 1 demonstrate that a function $f$ is a RoPE function parameterized by angle matrices if and only if the angle matrices $\mathcal{A}$ pairwise commute. In other words, to construct a relative positional encoding method for an attention mechanism that is robust to offset, it suffices to establish an angle matrix set that pairwise commutes. Thus, in the following section, we focus on the construction of $\mathcal{A}$ to satisfy this particular requirement. We call our method ComRoPE to indicate the commutativity of angle matrices in RoPE function. The overview of ComRoPE is shown in Figure 1. + +Remark 1. Among all the previous RoPE methods, LieRE [20] is the most related one to ours which solves Equation 3 by directly training the skew-symmetric matrix set $\mathcal{A}$ . However, it is worth noting that in the following equation of their implementation + +$$ +\left(\mathbf {R} (\boldsymbol {x} _ {q}; \mathcal {A}) \boldsymbol {q}\right) ^ {\top} \left(\mathbf {R} (\boldsymbol {x} _ {k}; \mathcal {A}) \boldsymbol {k}\right) = \boldsymbol {q} ^ {\top} \mathbf {R} (\boldsymbol {x} _ {q}; \mathcal {A}) ^ {\top} \mathbf {R} (\boldsymbol {x} _ {k}; \mathcal {A}) \boldsymbol {k}, (4) +$$ + +sometimes $\mathbf{R}(\boldsymbol{x}_q; \mathcal{A})^\top \mathbf{R}(\boldsymbol{x}_k; \mathcal{A}) \neq \mathbf{R}(\boldsymbol{x}_q - \boldsymbol{x}_k; \mathcal{A})$ because the probability that two random matrices commute is small, where $\boldsymbol{x}_q$ and $\boldsymbol{x}_k$ are two arbitrary coordinates. Thus, $\mathbf{R}$ proposed by LieRe does not consistently satisfy the requirements of the RoPE Equation. + +# 3.3. Construction of pairwise commuting matrices + +In this section, we elaborate on concrete ways to construct the pairwise commuting matrices to solve the RoPE Equation parameterized by angle matrices. Note that if two matrices are both block diagonal with the same block sizes, where the corresponding blocks are commutative, then these two matrices are commutative. Formally speaking, for two matrices $\mathbf{M},\mathbf{N}\in \mathbb{R}^{d\times d}$ , they commute if + +$$ +\left\{ \begin{array}{l} \mathbf {M} = \operatorname {d i a g} \left(\mathbf {M} _ {1}, \mathbf {M} _ {2}, \dots , \mathbf {M} _ {b}\right) \\ \mathbf {N} = \operatorname {d i a g} \left(\mathbf {N} _ {1}, \mathbf {N} _ {2}, \dots , \mathbf {N} _ {b}\right) \\ \mathbf {M} _ {i}, \mathbf {N} _ {i} \in \mathbb {R} ^ {b \times b} \quad \forall i \in \{1, 2, \dots , m \} \\ \mathbf {M} _ {i} \mathbf {N} _ {i} = \mathbf {N} _ {i} \mathbf {M} _ {i} \quad \forall i \in \{1, 2, \dots , m \} \end{array} \right. \tag {5} +$$ + +where $b$ denotes block size that is a factor of $d$ and $m = \frac{d}{b}$ . + +Thus, to present solutions to RoPE Equation parameterized by angle matrices, it suffices to partition the angle matrices $\mathbf{A}_i$ in Eq. (3) into $m$ blocks $\mathbf{B}_{i1},\mathbf{B}_{i2},\dots ,\mathbf{B}_{im}$ where: + +$$ +\left\{ \begin{array}{l} \mathbf {A} _ {i} = \operatorname {d i a g} \left(\mathbf {B} _ {i 1}, \mathbf {B} _ {i 2}, \dots , \mathbf {B} _ {i m}\right) \\ \mathbf {B} _ {i j} \in \mathbb {R} ^ {b \times b} \quad \forall j \in \{1, 2, \dots , m \} \end{array} \right. \tag {6} +$$ + +Defining $\mathcal{B}_j = \{\mathbf{B}_{1j},\mathbf{B}_{2j},\dots ,\mathbf{B}_{Nj}\}$ , if $\mathcal{B}_j$ pairwise commutes for all $j\in \{1,2,\dots ,m\}$ , then $\mathcal{A}$ pairwise commutes. Note that $\mathcal{B}_1,\mathcal{B}_2,\dots ,\mathcal{B}_m$ are equivalent. Thus, without causing confusion, we will uniformly use $\mathcal{B} = \{\mathbf{B}_1,\mathbf{B}_2,\dots ,\mathbf{B}_N\}$ to represent $\mathcal{B}_j = \{\mathbf{B}_{1j},\mathbf{B}_{2j},\dots ,\mathbf{B}_{Nj}\}$ in the following. + +Currently, there is no general method that provides the necessary and sufficient conditions for ensuring that arbitrary trainable skew-symmetric matrices commute. Therefore, we aim to establish sufficient conditions for enforcing this constraint. More concretely, we present Proposition 2 and Proposition 3 to construct two different variants of matrices, respectively. + +Proposition 2. A set of matrices $\mathcal{B} = \{\mathbf{B}_1, \mathbf{B}_2, \dots, \mathbf{B}_N\}$ pairwise commutes if all but one of them are zero matrices, i.e., $\exists k \in \{1, 2, \dots, N\}$ s.t. + +$$ +\mathbf {B} _ {i} = \mathbf {O} (\forall i \neq k) \tag {7} +$$ + +Based on Proposition 2, we propose Trainable RoPE parameterized by Axial-Partition Angle Matrices (ComRoPE-AP). We divide the $m$ blocks into $N$ parts, such that each part is responsible for the rotation of a specific axis. In this case, $m$ should be a multiple of $N$ . Specifically, + +$$ +\mathbf {B} _ {i j} = \left\{ \begin{array}{l l} \mathbf {P} _ {j} - \mathbf {P} _ {j} ^ {\top}, & \text {i f} j \equiv i (\mathrm {m o d} N) \\ \mathbf {O}, & \text {o t h e r w i s e} \end{array} \right. \tag {8} +$$ + +where $\{\mathbf{P}_j\}_{j=1}^m$ represents a set of trainable matrices, $\mathbf{O}$ denotes a zero matrix, and $\equiv$ indicates congruence modulo. + +Proposition 3. A set of matrices $\mathcal{B} = \{\mathbf{B}_1, \mathbf{B}_2, \dots, \mathbf{B}_N\}$ pairwise commutes if they pairwise linearly dependent, i.e., $\exists \lambda_1, \lambda_2, \dots, \lambda_N \in \mathbb{R}$ s.t. + +$$ +\lambda_ {1} \mathbf {B} _ {1} = \lambda_ {2} \mathbf {B} _ {2} = \dots = \lambda_ {N} \mathbf {B} _ {N} \tag {9} +$$ + +Based on Proposition 3, we propose Trainable RoPE parameterized by Linearly-Dependent Angle Matrices (ComRoPE-LD). Specifically, We train a base matrix $\mathbf{P}$ with scaling factors $\{\theta_i\}_{i=1}^N$ . Then we obtain: + +$$ +\mathcal {B} = \left\{\mathbf {B} _ {i} = \theta_ {i} \left(\mathbf {P} - \mathbf {P} ^ {\top}\right) \mid i = 1, 2, \dots , N \right\} \tag {10} +$$ + +# 3.4. Implementation details of coordinates and improvements + +# 3.4.1 Relative scaling and center offset + +Relative scaling. In language models, positions are typically treated as discrete. Additionally, due to the relationship between tokens and the inherent uncertainty in sequence length, there is no need to scale these positions into a specified range. However, in the case of images, positions are continuous. When using different patch sizes, it + +![](images/39ff5b1cf2b7367ee8bd2e68d6c3653f0b289ba70dedd64f332c698644271be7.jpg) +Figure 2. Different patch sizes result in different relative relationships. + +becomes unreasonable to define their relative relationships solely based on the span of patches. As shown in Figure 2, the same locations with varying patch sizes can result in significantly different relative relationships. + +As a result, applying relative coordinates is a better method for measuring relative relationships in images. For an image with shape $H \times W$ , we scale both the height $H$ and the width $W$ to 1. Therefore, for a pixel located at $(h, w)$ in the raw image, its coordinate is treated as $\left(\frac{h}{H}, \frac{w}{W}\right)$ . For high-dimensional coordinates, we perform the same operation, i.e., transform a raw coordinate $(x_1, x_2, \dots, x_N)$ in multi-axial canvas with shape $(X_1, X_2, \dots, X_N)$ into $\left(\frac{x_1}{X_1}, \frac{x_2}{X_2}, \dots, \frac{x_N}{X_N}\right)$ . + +Center offset. When we project a patch to a feature tensor with a coordinate, we aggregate all information of the patch into a specified location. Simply, we adopt the center point of a patch as the aggregation location. + +# 3.4.2 Position perturbation + +To achieve better robustness and excellent performance across different scales during inference, we add perturbations to the coordinates of the patches. Specifically, for a patch with center $(x_{1},\dots ,x_{N})$ and size $(\Delta X_{1},\dots ,\Delta X_{N})$ , during training, we formulate its location as: + +$$ +\left. \mathcal {N} \left(\left(x _ {1}, \dots , x _ {N}\right) ^ {\top}; \operatorname {d i a g} \left(\sigma \Delta X _ {1}, \dots , \sigma \Delta X _ {N}\right) ^ {2}\right) \right. \tag {11} +$$ + +where $\sigma$ is a hyper-parameter called perturbation intensity. Additionally, we truncate the location within the patch area, i.e., $\left[x_k - \frac{\Delta X_k}{2}, x_k + \frac{\Delta X_k}{2}\right]$ . + +
Position Encoding MethodPerturbation IntensityEvaluation Resolution
112128192224256320384448512
APE130.0438.6956.458.7660.0259.2757.0454.1050.99
019.7133.4355.9758.6255.3152.6349.6846.3942.66
Vanilla RoPE136.9445.4860.7263.0963.2462.1259.2455.5151.11
036.4144.4859.9762.0362.5461.3658.5654.7950.58
LieRE138.0346.9762.2264.3664.9963.7861.1557.9253.74
038.2246.8562.0164.0764.5463.4660.7456.8952.51
ComRoPE-AP (ours)135.7546.1862.8265.3265.8364.7861.8858.2154.02
037.1746.9562.6364.7665.1164.3561.6258.1053.81
ComRoPE-LD (ours)138.3047.2863.5365.4965.9565.2762.6259.1155.29
036.8846.5462.8965.2765.6664.2761.8357.8753.64
+ +Table 1. Accuracy of 2D classification on ImageNet. Models are trained at a resolution of ${224} \times {224}$ and evaluated at varying resolutions. + +# 4. Experiments + +In this section, we evaluate the performance of various positional encoding methods on classic vision tasks. We first assess their scalability in 2D image classification across different resolutions. Additionally, we conduct object detection experiments to demonstrate the generalizability of our approach. To further examine the ability to handle higher-dimensional data, we perform 3D classification experiments, which are detailed in Appendix B. + +# 4.1. 2D classification + +# 4.1.1 Setup + +Baselines and model architecture. We evaluate our proposed methods (ComRoPE-LD and ComRoPE-AP) against APE, vanilla RoPE (as introduced by RoFormer), and LieRE. To isolate the effects of positional encoding, we utilize a standard Vision Transformer (ViT-B/16) architecture with minimal modifications. The APE codebook is removed for methods that do not employ APE, and self-attention layers are replaced with RoPE self-attention parameterized by angle matrices. This design highlights the performance differences among various positional encoding methods. A block size of 8 is used in practice. More details are provided in Appendix D. + +Training and evaluation protocol. All models are trained at a standard resolution of $224 \times 224$ and evaluated across multiple resolutions to test their robustness and scalability on the ImageNet-1K dataset [4]. The models are trained from scratch using randomly initialized parameters, ensuring no influence from pre-trained weights or external priors. To maintain fairness and reproducibility, we apply only basic data augmentation techniques, such as resizing and random cropping, focusing on relative performance comparisons rather than achieving absolute accuracy. The primary evaluation metric is accuracy on the test set across + +various resolutions. Since APE is inherently fixed and discrete, bilinear interpolation is applied to adapt it to different resolutions during evaluation. Optimization methods and hyper-parameters are detailed in Appendix C. + +![](images/91d57c34afde96ea9cce5466b8f19db5e9a5705309c57944613b82b58b5afcf8.jpg) + +![](images/1f72333d708d59e72203137d1ef3d51564cf45c13c96e0e5c875c35e811c6287.jpg) +Figure 3. Accuracy on ImageNet-1K for various positional encoding methods. The results for the same perturbation intensity are presented together for better comparison. For better visualization, evaluation resolution at $112 \times 112$ is not included in these figures. + +# 4.1.2 Main results + +The accuracy metrics for each positional encoding method across various resolutions are summarized in Table 1. Additionally, Figure 3 visually compares performance under different levels of perturbation intensity separately. + +Overall performance. Across all evaluations, APE consistently exhibits the lowest accuracy, corroborating previous findings regarding its limitations in dynamic contexts. The vanilla RoPE shows a modest improvement over APE but remains less effective. In contrast, the trainable angle matrices, namely LieRE, ComRoPE-AP, and ComRoPE-LD, + +demonstrate significantly higher accuracy across all resolutions. Notably, ComRoPE-LD achieves the best performance among the three, suggesting that its inherent linear dependencies may enhance flexibility and structural learning capabilities. + +Accuracy at training resolution. At the training resolution of $224 \times 224$ , all three methods with trainable angle matrices (ComRoPE-LD, ComRoPE-AP, and LieRE) achieve comparable accuracy, significantly outperforming both APE and the standard RoPE, which underscores the effectiveness of RoPE parameterized by trainable angle matrices. Notably, ComRoPE-LD surpasses the current state-of-the-art LieRE by $1.6\%$ . + +Scaling to higher resolution. At resolutions beyond the training size, LieRE shows the steepest decline in accuracy among the three trainable RoPE variants, indicating greater sensitivity to resolution changes. In contrast, ComRoPE-LD and ComRoPE-AP exhibit a more gradual decrease in performance, thanks to their commutative properties that enhance positional robustness. Specifically, ComRoPE-LD outperforms LieRE by $2.9\%$ at a resolution of $512 \times 512$ . + +Significance of commutativity. These findings illustrate the effectiveness of trainable commutative angle matrices, particularly ComRoPE-LD, in maintaining accuracy and scalability across diverse resolutions. The results underscore the importance of commutativity in ensuring robust RoPE parameterized by trainable angle matrices for vision tasks. Furthermore, to better understand the role of commutativity, we conduct additional experiments by introducing coordinate offsets in LieRE (see Section 4.3.1 for details). + +# 4.2. Object detection + +To demonstrate the generalizability of our approach, we conduct object detection experiments using the framework from [32]. We adopt ViT-S as our backbone and apply ComRoPE to the attention layers. To ensure consistency with the pre-trained model, we initialize the angle matrix to zero. + +We evaluate ComRoPE-LD, LieRE, and APE on the MS COCO dataset [14]. As summarized in Table 2, both ComRoPE-LD and LieRE outperform APE, with ComRoPE-LD achieving slightly better performance than LieRE while only using nearly half the number of extra parameters. + +To compare training efficiency, we plot the results for each epoch in Figure 4. Our findings indicate that both ComRoPE-LD and LieRE converge faster than APE, requiring $29\%$ fewer iterations to achieve the same results as APE. + +
PE MethodAPAP50AP75APsAPmAPl
APE44.066.647.728.246.858.4
LieRE44.567.348.429.046.958.7
ComRoPE-LD44.767.648.529.247.160.0
+ +Table 2. Results of object detection on MS COCO. + +![](images/2a7590cffbb7ff400373f87e79574f869a6c3e1cdf531766c8f4a02fb629294c.jpg) +Figure 4. Results over the whole training procedure. + +# 4.3. Ablation study + +# 4.3.1 Impact of commutativity + +To evaluate the significance of the commutativity of angle matrices, we conduct experiments on the LieRE by introducing a coordinate offset. Specifically, before inference, a random offset is applied uniformly across all coordinates within the image. The offset is sampled from a Gaussian distribution as follows: + +$$ +\mathcal {N} (0; \rho^ {2} \cdot \mathbf {I} _ {N \times N}) \tag {12} +$$ + +where $\rho$ represents the standard deviation of the random offset, and $\mathbf{I}_{N\times N}$ is the identity matrix of size $N\times N$ . It is important to note that applying the same offset to all coordinates does not influence the relative positional dependencies between the patches. The other experimental settings remain consistent with those in the 2D classification tasks. + +The results shown in Figure 5 demonstrate that the baseline model's performance deteriorates significantly as the standard deviation of the offset increases. In contrast, our proposed model, ComRoPE, maintains consistent performance across all offset values. This is due to the commutativity of the angle matrices, which allows the model to remain invariant to such coordinate shifts. The robustness of ComRoPE to this type of perturbation highlights its capacity to preserve relative positional information, even in the presence of modification introduced by coordinate offsets. + +# 4.3.2 Impact of block size + +In this section, we examine the impact of block size on 2D classification using the ImageNet dataset. We maintain the same experimental setup as in our primary experiments, varying the block size from 2 to 8. The results are presented in Figure 6. Our findings indicate that larger block sizes consistently improve performance by extending the rotation transformation space to a more significant subgroup of the + +![](images/c74fed46cd6519d96090420608f1d5304df3ea1b11b005f7cde938dbb1603293.jpg) +Figure 5. Effect of coordinate offset on LieRE. As the standard deviation of the offset increases, the performance of the baseline model deteriorates, while ComRoPE remains unaffected (un-painted). + +![](images/61d78426ebfcb71393d359a0855521f0a4562974f5b426f6b6cd0f3f38efac08.jpg) +(a) LieRE + +![](images/8efe65594a14fec2220edd8e5275cbe83f78697ec31e930a67846059716ef106.jpg) +(b) ComRoPE-AP + +![](images/5c5df1dcd0c3dea495c46bef9c13c99829a7e4102f564457c9e6d7209d24d469.jpg) +(c) ComRoPE-LD +Figure 6. Accuracy on ImageNet for various block sizes. Larger block size results in better performance. + +particular orthogonal group by introducing additional parameters and computation time. When the block size is small, the associated costs are minimal. However, as the block size increases, the primary term of time complexity grows to $O(Lndb^{2})$ from $O(Lnd\cdot \frac{d}{h})$ , which becomes significant. Therefore, we limit the block size to a maximum of 8 to balance performance with additional costs. In other words, we select a block size of 8 to optimize performance while keeping the extra computational cost manageable. + +# 4.3.3 Utility of position perturbation + +In this section, we explore the impact of positional perturbations. We conducted experiments on ComRoPE-LD and APE using the ImageNet dataset, with the results presented in Figure 7. + +As shown in Figure 7a, APE is highly sensitive to positional perturbations, leading to significant performance improvement $(+19.5\%)$ when increasing intensity from 0 to 1) when these perturbations are introduced. For RoPE with angle matrices shown in Figure 7b, positional perturbations also resulted in some performance gains $(+2.9\%)$ , though + +![](images/8d171c91a600ba24a9c12ce64527510a50312b8948113ffd0f1c1b5385daca9b.jpg) +(a) APE + +![](images/3380304f09d22faee65a8eb2bfb9c163ad7a0ec5e552cb5210a31e9cd3177d53.jpg) +(b) ComRoPE-LD +Figure 7. Accuracy on ImageNet over varying position perturbation intensity. + +the improvement was relatively modest. This is likely due to the inherent robustness of the RoPE design with angle matrices, which is already well-equipped to handle variations in position. + +# 4.4. Applications + +In our approach, when the angle matrix is an all-zero matrix, the rotation matrix becomes the identity matrix, causing RoPE Attention to reduce to the standard Attention mechanism. When the block size of the angle matrix is set to 2, ComRoPE-AP effectively reduces to the commonly used RoPE Attention in language models. This demonstrates that our method can represent standard Attention and various common variants of RoPE Attention. Therefore, during the fine-tuning stage, we can replace standard Attention with our method, load the pre-trained weights, and fine-tune them under the new paradigm. In other words, ComRoPE can be seamlessly integrated into the fine-tuning process, even if it was not applied during pre-training. Additional experiments can be found in Appendix B. + +# 5. Conclusion + +In this work, we proposed ComRoPE, a novel and adaptive framework for Rotary Position Embedding (RoPE) parameterized by trainable angle matrices. We rigorously formulate the RoPE Equation and then establish a necessary and sufficient condition for its solution. Our approach effectively overcomes the scalability and robustness limitations of existing RoPE methods by eliminating the need for manually designed rotation matrices and introducing a more flexible, scalable solution. Extensive experimental results show that ComRoPE outperforms the existing positional encoding methods across various tasks. Furthermore, our framework generalizes existing RoPE formulations and demonstrates the potential for broader application in Transformer models, offering insights and a solid foundation for future research in positional encoding techniques. + +# Acknowledgment + +This work is supported by the National Key R&D Program of China (2022YFB4701400/4701402), SSTIC Grant (KJZD20230923115106012, KJZD20230923114916032, GJHZ20240218113604008) and Beijing Key Lab of Networked Multimedia. + +# References + +[1] Tom B Brown. Language models are few-shot learners. arXiv preprint arXiv:2005.14165, 2020. 1 +[2] Yirui Chen, Xudong Huang, Quan Zhang, Wei Li, Mingjian Zhu, Qiangyu Yan, Simiao Li, Hanting Chen, Hailin Hu, Jie Yang, et al. Gim: A million-scale benchmark for generative image manipulation detection and localization. arXiv preprint arXiv:2406.16531, 2024. 1 +[3] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhanghao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E Gonzalez, et al. Vicuna: An open-source chatbot impressing gpt-4 with $90\%$ * chatgpt quality, march 2023. URL https://lmsys.org/blog/2023-03-30-vicuna, 3(5), 2023. 2 +[4] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, K. Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 248-255, 2009. 6 +[5] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In North American Chapter of the Association for Computational Linguistics, 2019. 1, 2 +[6] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 1 +[7] Yuxin Fang, Quan Sun, Xinggang Wang, Tiejun Huang, Xinlong Wang, and Yue Cao. Eva-02: A visual representation for neon genesis. Image and Vision Computing, 149:105171, 2024. 2 +[8] Team GLM, Aohan Zeng, Bin Xu, Bowen Wang, Chenhui Zhang, Da Yin, Dan Zhang, Diego Rojas, Guanyu Feng, Hanlin Zhao, et al. Chatglm: A family of large language models from glm-130b to glm-4 all tools. arXiv preprint arXiv:2406.12793, 2024. 1 +[9] Byeongho Heo, Song Park, Dongyoon Han, and Sangwoo Yun. Rotary position embedding for vision transformer. arXiv preprint arXiv:2403.13298, 2024. 2 +[10] Max Horn, Kumar Shridhar, Elrich Groenewald, and Philipp FM Baumann. Translational equivariance in kernelizable attention. arXiv preprint arXiv:2102.07680, 2021. 2 +[11] Amirhossein Kazemnejad, Inkit Padhi, Karthikeyan Natesan Ramamurthy, Payel Das, and Siva Reddy. The impact of positional encoding on length generalization in transformers. Advances in Neural Information Processing Systems, 36: 24892-24928, 2023. 1 +[12] Hanyu Lai, Xiao Liu, Iat Long Iong, Shuntian Yao, Yuxuan Chen, Pengbo Shen, Hao Yu, Hanchen Zhang, Xiaohan + +Zhang, Yuxiao Dong, et al. Autowebglm: A large language model-based web navigating agent. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 5295-5306, 2024. 1 +[13] Hongyu Li, Jinyu Chen, Ziyu Wei, Shaofei Huang, Tianrui Hui, Jialin Gao, Xiaoming Wei, and Si Liu. Llava-st: A multimodal large language model for fine-grained spatial-temporal understanding. arXiv preprint arXiv:2501.08282, 2025. 1 +[14] Tsung-Yi Lin, Michael Maire, Serge Belongie, Lubomir Bourdev, Ross Girshick, James Hays, Pietro Perona, Deva Ramanan, C. Lawrence Zitnick, and Piotr Dollar. Microsoft coco: Common objects in context. arXiv preprint arXiv: 1405.0312, 2014. 7 +[15] Xiao Liu, Hanyu Lai, Hao Yu, Yifan Xu, Aohan Zeng, Zhengxiao Du, Peng Zhang, Yuxiao Dong, and Jie Tang. Webglm: towards an efficient web-enhanced question answering system with human preferences. In Proceedings of the 29th ACM SIGKDD conference on knowledge discovery and data mining, pages 4549-4560, 2023. 1 +[16] Xiao Liu, Hao Yu, Hanchen Zhang, Yifan Xu, Xuanyu Lei, Hanyu Lai, Yu Gu, Hangliang Ding, Kaiwen Men, Kejuan Yang, et al. Agentbench: Evaluating llms as agents. In The Twelfth International Conference on Learning Representations, 2024. 1 +[17] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 2 +[18] Antoine Liutkus, Ondrej Cifka, Shih-Lun Wu, Umut Simsekli, Yi-Hsuan Yang, and Gael Richard. Relative positional encoding for transformers with linear complexity. In International Conference on Machine Learning, pages 7067-7079. PMLR, 2021. 2 +[19] Jiasen Lu, Christopher Clark, Sangho Lee, Zichen Zhang, Savya Khosla, Ryan Marten, Derek Hoiem, and Aniruddha Kembhavi. Unified-io 2: Scaling autoregressive multimodal models with vision language audio and action. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26439–26455, 2024. 2 +[20] Sophie Ostmeier, Brian Axelrod, Michael E Moseley, Akshay Chaudhari, and Curtis Langlotz. Liere: Generalizing rotary position encodings. arXiv preprint arXiv:2406.10322, 2024. 2, 4 +[21] Ofir Press, Noah A Smith, and Mike Lewis. Train short, test long: Attention with linear biases enables input length extrapolation. arXiv preprint arXiv:2108.12409, 2021. 2 +[22] A Radford. Improving language understanding by generative pre-training. 2018. 1 +[23] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. 1 +[24] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of machine learning research, 21(140):1-67, 2020. 2 + +[25] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of machine learning research, 21(140):1-67, 2020. 2 +[26] Peter Shaw, Jakob Uszkoreit, and Ashish Vaswani. Self-attention with relative position representations. arXiv preprint arXiv:1803.02155, 2018. 2 +[27] Jianlin Su, Murtadha Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu. Roformer: Enhanced transformer with rotary position embedding. Neurocomputing, 568:127063, 2024. 1, 2, 3 +[28] Sainbayar Sukhbaatar, Jason Weston, Rob Fergus, et al. End-to-end memory networks. Advances in neural information processing systems, 28, 2015. 2 +[29] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 2 +[30] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 1, 2 +[31] Kan Wu, Houwen Peng, Minghao Chen, Jianlong Fu, and Hongyang Chao. Rethinking and improving relative position encoding for vision transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10033-10041, 2021. 2 +[32] Chunlong Xia, Xinliang Wang, Feng Lv, Xin Hao, and Yifeng Shi. Vit-comer: Vision transformer with convolutional multi-scale feature interaction for dense predictions. Computer Vision and Pattern Recognition, 2024. 7 +[33] Quan Zhang and Yuxin Qi. Can mllms guide weakly-supervised temporal action localization tasks? arXiv preprint arXiv:2411.08466, 2024. 1 +[34] Quan Zhang, Xiaoyu Liu, Wei Li, Hanting Chen, Junchao Liu, Jie Hu, Zhiwei Xiong, Chun Yuan, and Yunhe Wang. Distilling semantic priors from sam to efficient image restoration models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25409-25419, 2024. +[35] Quan Zhang, Yuxin Qi, Xi Tang, Jinwei Fang, Xi Lin, Ke Zhang, and Chun Yuan. IMDPrompter: Adapting SAM to image manipulation detection by cross-view automated prompt learning. In The Thirteenth International Conference on Learning Representations, 2025. +[36] Quan Zhang, Yuxin Qi, Xi Tang, Rui Yuan, Xi Lin, Ke Zhang, and Chun Yuan. Rethinking pseudo-label guided learning for weakly supervised temporal action localization from the perspective of noise correction. arXiv preprint arXiv:2501.11124, 2025. 1 \ No newline at end of file diff --git a/2025/ComRoPE_ Scalable and Robust Rotary Position Embedding Parameterized by Trainable Commuting Angle Matrices/images.zip b/2025/ComRoPE_ Scalable and Robust Rotary Position Embedding Parameterized by Trainable Commuting Angle Matrices/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..dfd584c2fa33418792ed3f347a15ffffe76d1d9d --- /dev/null +++ b/2025/ComRoPE_ Scalable and Robust Rotary Position Embedding Parameterized by Trainable Commuting Angle Matrices/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:77dc4ffa83499eccfe8dfa519c0afbdec63276d4fe13474f1d609d167fecf989 +size 485913 diff --git a/2025/ComRoPE_ Scalable and Robust Rotary Position Embedding Parameterized by Trainable Commuting Angle Matrices/layout.json b/2025/ComRoPE_ Scalable and Robust Rotary Position Embedding Parameterized by Trainable Commuting Angle Matrices/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..5ca37a3bcbc7da22330e6b1abb221c76ea6eeb76 --- /dev/null +++ b/2025/ComRoPE_ Scalable and Robust Rotary Position Embedding Parameterized by Trainable Commuting Angle Matrices/layout.json @@ -0,0 +1,9511 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 61, + 102, + 550, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 102, + 550, + 140 + ], + "spans": [ + { + "bbox": [ + 61, + 102, + 550, + 140 + ], + "type": "text", + "content": "ComRoPE: Scalable and Robust Rotary Position Embedding Parameterized by Trainable Commuting Angle Matrices" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 78, + 159, + 531, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 159, + 531, + 176 + ], + "spans": [ + { + "bbox": [ + 78, + 159, + 531, + 176 + ], + "type": "text", + "content": "Hao " + }, + { + "bbox": [ + 78, + 159, + 531, + 176 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}\\mathbf{u}^{1}" + }, + { + "bbox": [ + 78, + 159, + 531, + 176 + ], + "type": "text", + "content": " Tangyu Jiang" + }, + { + "bbox": [ + 78, + 159, + 531, + 176 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 78, + 159, + 531, + 176 + ], + "type": "text", + "content": " Shuning Jia" + }, + { + "bbox": [ + 78, + 159, + 531, + 176 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 78, + 159, + 531, + 176 + ], + "type": "text", + "content": " Shannon Yan" + }, + { + "bbox": [ + 78, + 159, + 531, + 176 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 78, + 159, + 531, + 176 + ], + "type": "text", + "content": " Shunning Liu" + }, + { + "bbox": [ + 78, + 159, + 531, + 176 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 110, + 180, + 500, + 196 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 180, + 500, + 196 + ], + "spans": [ + { + "bbox": [ + 110, + 180, + 500, + 196 + ], + "type": "text", + "content": "Haolong Qian1 Guanghao Li1 Shuting Dong1 Chun Yuan1†" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 193, + 202, + 417, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 193, + 202, + 417, + 217 + ], + "spans": [ + { + "bbox": [ + 193, + 202, + 417, + 217 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 193, + 202, + 417, + 217 + ], + "type": "text", + "content": "Tsinghua University " + }, + { + "bbox": [ + 193, + 202, + 417, + 217 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 193, + 202, + 417, + 217 + ], + "type": "text", + "content": "Shenzhen University" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 94, + 220, + 506, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 220, + 506, + 232 + ], + "spans": [ + { + "bbox": [ + 94, + 220, + 506, + 232 + ], + "type": "text", + "content": "longinyh@gmail.com, jiangtangyu@sz.tsinghua.edu.cn, yuanc@sz.tsinghua.edu.cn" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 151, + 252, + 200, + 265 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 252, + 200, + 265 + ], + "spans": [ + { + "bbox": [ + 151, + 252, + 200, + 265 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 277, + 297, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 277, + 297, + 602 + ], + "spans": [ + { + "bbox": [ + 54, + 277, + 297, + 602 + ], + "type": "text", + "content": "The Transformer architecture has revolutionized various fields since it was proposed, where positional encoding plays an essential role in effectively capturing sequential order and context. Therefore, Rotary Positional Encoding (RoPE) was proposed to alleviate these issues, which integrates positional information by rotating the embeddings in the attention mechanism. However, RoPE utilizes manually defined rotation matrices, a design choice that favors computational efficiency but limits the model's flexibility and adaptability. In this work, we propose ComRoPE, which generalizes RoPE by defining it in terms of trainable commuting angle matrices. Specifically, we demonstrate that pairwise commutativity of these matrices is essential for RoPE to achieve scalability and positional robustness. We formally define the RoPE Equation, which is an essential condition that ensures consistent performance with position offsets. Based on the theoretical analysis, we present two types of trainable commuting angle matrices as sufficient solutions to the RoPE equation, which significantly improve performance, surpassing the current state-of-the-art method by " + }, + { + "bbox": [ + 54, + 277, + 297, + 602 + ], + "type": "inline_equation", + "content": "1.6\\%" + }, + { + "bbox": [ + 54, + 277, + 297, + 602 + ], + "type": "text", + "content": " at training resolution and " + }, + { + "bbox": [ + 54, + 277, + 297, + 602 + ], + "type": "inline_equation", + "content": "2.9\\%" + }, + { + "bbox": [ + 54, + 277, + 297, + 602 + ], + "type": "text", + "content": " at higher resolution on the ImageNet-1K dataset. Furthermore, our framework shows versatility in generalizing to existing RoPE formulations and offering new insights for future positional encoding research. To ensure reproducibility, the source code and instructions are available at https://github.com/Longin-Yu/ComRoPE." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 607, + 135, + 619 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 607, + 135, + 619 + ], + "spans": [ + { + "bbox": [ + 56, + 607, + 135, + 619 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 627, + 296, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 627, + 296, + 688 + ], + "spans": [ + { + "bbox": [ + 55, + 627, + 296, + 688 + ], + "type": "text", + "content": "The Transformer architecture [30] has been widely adopted across various fields, including Natural Language Processing (NLP) [1, 5, 8, 22, 23] and Computer Vision (CV) [6]. Moreover, an increasing number of Transformer-based applications [2, 12, 13, 15, 16, 33-36] continue to demonstrate" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 262, + 555, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 262, + 555, + 334 + ], + "spans": [ + { + "bbox": [ + 313, + 262, + 555, + 334 + ], + "type": "text", + "content": "their effectiveness across a wide range of domains. At the core of the Transformer model lies the attention mechanism, which enables the model to selectively focus on different parts of the input based on relevance, thereby effectively capturing long-range dependencies and contextual relationships." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 336, + 556, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 336, + 556, + 420 + ], + "spans": [ + { + "bbox": [ + 313, + 336, + 556, + 420 + ], + "type": "text", + "content": "However, since the attention mechanism is insensitive to the fundamental position information, it cannot inherently capture the order of elements in the data. In order to alleviate this issue, positional embeddings are added to the input representation, providing the model with the necessary positional information to process the sequence order and structure effectively." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 312, + 423, + 556, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 423, + 556, + 674 + ], + "spans": [ + { + "bbox": [ + 312, + 423, + 556, + 674 + ], + "type": "text", + "content": "A variety of positional encoding methods have been proposed, which can be divided into two categories: Absolute Positional Encoding (APE) and Relative Positional Encoding (RPE). APE explicitly encodes the absolute position of each token by generating fixed positional embeddings, which are directly added to the input embeddings at the beginning of training [30]. However, APE struggles with handling long sequences and exhibits high sensitivity to positional shifts, limiting the scalability and robustness of models [11]. In contrast, RPE does not directly modify input embeddings; instead, it incorporates relative positional information within the attention mechanism, enabling the model to effectively capture positional relationships between tokens. Among all the existing works, Rotary Position Embedding (RoPE) [27] has gained significant attention due to the advantage of applying a rotational transformation to token embeddings. It encodes relative positions by treating each pair of features as coordinates and rotating them by an angle proportional to their position, enabling all tokens to interact within the attention mechanism regardless of distance." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "text", + "content": "However, existing RoPE methods face several key challenges: i) The essential components (i.e., the RoPE matrices) of previous RoPE approaches rely on 2D rotation" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "type": "text", + "content": "†Corresponding author. This work was done when Shuning Jia was an intern at Tsinghua University." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4508" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 192 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 192 + ], + "type": "text", + "content": "groups, which simplify computations but consequently restrict their feature projection capabilities, especially in high-dimensional spaces [27]. ii) Moreover, the majority of the rotation matrices of RoPE require to be manually designed, leading to insufficient capability and suboptimal performance[21, 25]. iii) Finally, previous attempts [20] to extend the rotation group often prioritize design simplicity, making it difficult to consistently satisfy relative position dependency—a critical property of RoPE that ensures positional robustness against offsets." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 201, + 294, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 201, + 294, + 357 + ], + "spans": [ + { + "bbox": [ + 55, + 201, + 294, + 357 + ], + "type": "text", + "content": "Our objectives. This work aims to develop a novel RoPE method for Transformers that is both scalable and robust. Specifically, we seek to extend the rotation group from the existing 2D representation to a larger subgroup of the special orthogonal group, allowing for higher degrees of freedom while preserving consistent behavior with respect to position offsets. Unlike existing methods that rely on manually designed non-trainable rotation matrices, which suffer from limited expressiveness and reduced robustness, our approach is designed to offer richer feature representation capabilities. This framework addresses the scalability limitations of current approaches and enhances their robustness against positional transformations." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 365, + 294, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 365, + 294, + 483 + ], + "spans": [ + { + "bbox": [ + 55, + 365, + 294, + 483 + ], + "type": "text", + "content": "Our contributions. This work introduces ComRoPE, a novel framework that significantly enhances positional encoding in Transformers. ComRoPE leverages trainable angle matrices, extending the RoPE mechanism with higher scalability and robustness. We identify the pairwise commutativity of these matrices as a necessary and sufficient condition for effective positional encoding, thereby unifying various existing RoPE formulations under a single theoretical framework. The contributions are summarized as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 494, + 293, + 707 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 56, + 494, + 293, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 494, + 293, + 538 + ], + "spans": [ + { + "bbox": [ + 56, + 494, + 293, + 538 + ], + "type": "text", + "content": "- We formally define the RoPE function parameterized by angle matrices and prove that pairwise commutativity is a necessary and sufficient condition, offering a unified theory that encompasses several existing RoPE variants." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 542, + 293, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 542, + 293, + 598 + ], + "spans": [ + { + "bbox": [ + 56, + 542, + 293, + 598 + ], + "type": "text", + "content": "- We introduce ComRoPE, a scalable and robust solution that leverages two types of trainable commuting angle matrix sets as sufficient solutions to the RoPE equation, capturing richer positional representations without the need for manual design." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 601, + 293, + 669 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 601, + 293, + 669 + ], + "spans": [ + { + "bbox": [ + 56, + 601, + 293, + 669 + ], + "type": "text", + "content": "- Our extensive experiments show that ComRoPE surpasses the current state-of-the-art LieRE, achieving a performance increase of " + }, + { + "bbox": [ + 56, + 601, + 293, + 669 + ], + "type": "inline_equation", + "content": "1.6\\%" + }, + { + "bbox": [ + 56, + 601, + 293, + 669 + ], + "type": "text", + "content": " at training resolution and " + }, + { + "bbox": [ + 56, + 601, + 293, + 669 + ], + "type": "inline_equation", + "content": "2.9\\%" + }, + { + "bbox": [ + 56, + 601, + 293, + 669 + ], + "type": "text", + "content": " at higher resolutions on the ImageNet-1K classification task while delivering strong results across other benchmarks." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 673, + 293, + 707 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 673, + 293, + 707 + ], + "spans": [ + { + "bbox": [ + 56, + 673, + 293, + 707 + ], + "type": "text", + "content": "- We explore further applications of ComRoPE, providing valuable insights for advancing position encoding techniques in future Transformer-based models." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 71, + 399, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 71, + 399, + 83 + ], + "spans": [ + { + "bbox": [ + 314, + 71, + 399, + 83 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 93, + 489, + 105 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 93, + 489, + 105 + ], + "spans": [ + { + "bbox": [ + 314, + 93, + 489, + 105 + ], + "type": "text", + "content": "2.1. Position information in attention" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 114, + 553, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 114, + 553, + 462 + ], + "spans": [ + { + "bbox": [ + 313, + 114, + 553, + 462 + ], + "type": "text", + "content": "Transformers [30] utilize the attention mechanism to capture similarities within sequences. However, they lack inherent sequential information and cannot capture the positional information of each token. To address this limitation, positional encoding [5, 24, 26] was introduced. As research progressed, positional encoding generally evolved into two types: APE [5, 28, 30] and RPE [10, 18, 26, 27]. [30] first proposed using APE in the form of sine and cosine functions, effectively representing the positional relationships within the input sequence. This positional encoding method achieved remarkable results in natural language processing, becoming a foundational component for many NLP tasks. However, the fixed nature of positional encoding limited the model's ability to generalize to longer input sequences. In response, subsequent research introduced learnable absolute positional encoding. [5] proposed further enhancing the model's expressive power by incorporating learnable position embeddings, particularly excelling in tasks such as sentence alignment and context representation. Although APE provides positional information to enhance the model's understanding of sequences, it shows limitations in handling long sequences and cross-sequence scenarios. Hierarchical ViT such as Swin Transformer [17], introduced Relative Position Bias (RPB) [17, 25, 26] to handle large numbers of tokens with limited positional embeddings [9]. Related research has explored alternative encoding methods, such as RPE, as a replacement for APE to better capture complex dependency structures. iRPE [31] proposed an improved RPB by combining relative position embedding." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 477, + 465, + 490 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 477, + 465, + 490 + ], + "spans": [ + { + "bbox": [ + 314, + 477, + 465, + 490 + ], + "type": "text", + "content": "2.2. Rotary position embedding" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 498, + 553, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 553, + 712 + ], + "type": "text", + "content": "Building on these developments, RoFormer [27] combined the advantages of APE and RPE, proposing RoPE. Currently, RoPE is widely used in Large Language Models (LLMs), such as LLaMA [29] and Vicuna [3]. This approach enhances model performance on tasks involving long-text semantics and multi-turn dialogues, improving extrapolation capability. To better adapt to the characteristics of two-dimensional data such as images, researchers extended RoPE to two-dimensional sequences (2D-RoPE). LieRE [20] extends RoPE to a more generalized form by introducing a rotation-based positional encoding method grounded in Lie group theory. Unified-IO 2 [19] applies 2D-RoPE within its multimodal architecture; EVA-02 [7], FiT [7] these pioneering works used 2D RoPE with axial frequencies (2D Axial RoPE), but had limitations in processing in the diagonal direction. Therefore, RoPE for ViT [9] proposes to use mixed axial frequency for 2D RoPE, named RoPE-Mixed." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "4509" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 71, + 111, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 71, + 111, + 83 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 111, + 83 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 187 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 187 + ], + "type": "text", + "content": "In this section, we begin by introducing key definitions and reformulating the RoPE paradigm within multi-axial attention mechanisms, which we collectively refer to as the RoPE Equations. Next, we present our main theorem, which establishes the necessary and sufficient conditions for RoPE functions parameterized by angle matrices. Finally, we provide solutions to the RoPE Equations based on the propositions that outline the sufficient conditions." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 193, + 141, + 205 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 193, + 141, + 205 + ], + "spans": [ + { + "bbox": [ + 55, + 193, + 141, + 205 + ], + "type": "text", + "content": "3.1. Preliminaries" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 211, + 296, + 259 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 211, + 296, + 259 + ], + "spans": [ + { + "bbox": [ + 55, + 211, + 296, + 259 + ], + "type": "text", + "content": "We use " + }, + { + "bbox": [ + 55, + 211, + 296, + 259 + ], + "type": "inline_equation", + "content": "\\mathbf{R}(\\cdot):\\mathbb{R}^N\\to \\mathbb{R}^{d\\times d}" + }, + { + "bbox": [ + 55, + 211, + 296, + 259 + ], + "type": "text", + "content": " to denote a matrix-value function. We denote the vectors and matrices as the lower and uppercase bold font, respectively. We first recall two fundamental definitions of matrices as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 265, + 296, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 265, + 296, + 301 + ], + "spans": [ + { + "bbox": [ + 55, + 265, + 296, + 301 + ], + "type": "text", + "content": "Definition 1 (Matrix Exponential). The exponential of a square matrix " + }, + { + "bbox": [ + 55, + 265, + 296, + 301 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 55, + 265, + 296, + 301 + ], + "type": "text", + "content": ", denoted as " + }, + { + "bbox": [ + 55, + 265, + 296, + 301 + ], + "type": "inline_equation", + "content": "e^{\\mathbf{A}}" + }, + { + "bbox": [ + 55, + 265, + 296, + 301 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 55, + 265, + 296, + 301 + ], + "type": "inline_equation", + "content": "\\exp (\\mathbf{A})" + }, + { + "bbox": [ + 55, + 265, + 296, + 301 + ], + "type": "text", + "content": ", is defined using the matrix exponential series such that:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 60, + 307, + 291, + 338 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 307, + 291, + 338 + ], + "spans": [ + { + "bbox": [ + 60, + 307, + 291, + 338 + ], + "type": "interline_equation", + "content": "e ^ {\\mathbf {A}} = \\exp (\\mathbf {A}) = \\sum_ {k = 0} ^ {\\infty} \\frac {\\mathbf {A} ^ {k}}{k !} = \\mathbf {I} + \\mathbf {A} + \\frac {\\mathbf {A} ^ {2}}{2 !} + \\frac {\\mathbf {A} ^ {3}}{3 !} + \\dots ,", + "image_path": "b82ee36214e31dccf1a9b647cdbbae5dfee785a7bd78d16f42f87b1aabbf96ed.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 343, + 246, + 356 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 343, + 246, + 356 + ], + "spans": [ + { + "bbox": [ + 55, + 343, + 246, + 356 + ], + "type": "text", + "content": "This series converges for any square matrix " + }, + { + "bbox": [ + 55, + 343, + 246, + 356 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 55, + 343, + 246, + 356 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 361, + 296, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 361, + 296, + 397 + ], + "spans": [ + { + "bbox": [ + 55, + 361, + 296, + 397 + ], + "type": "text", + "content": "Definition 2 (Commuting Matrices). Two square matrices " + }, + { + "bbox": [ + 55, + 361, + 296, + 397 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_1" + }, + { + "bbox": [ + 55, + 361, + 296, + 397 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 361, + 296, + 397 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_2" + }, + { + "bbox": [ + 55, + 361, + 296, + 397 + ], + "type": "text", + "content": " are said to commute if their product is independent of the order of multiplication, i.e.," + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 140, + 403, + 211, + 416 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 403, + 211, + 416 + ], + "spans": [ + { + "bbox": [ + 140, + 403, + 211, + 416 + ], + "type": "interline_equation", + "content": "\\mathbf {A} _ {1} \\mathbf {A} _ {2} = \\mathbf {A} _ {2} \\mathbf {A} _ {1}.", + "image_path": "f4c252d95cb37dfe01b8b8699de6f37a54f774f2c89ab83c46d6cdd007d9ed3b.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 422, + 296, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 422, + 296, + 470 + ], + "spans": [ + { + "bbox": [ + 55, + 422, + 296, + 470 + ], + "type": "text", + "content": "A set of square matrices " + }, + { + "bbox": [ + 55, + 422, + 296, + 470 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{A}_1,\\mathbf{A}_2,\\dots ,\\mathbf{A}_N\\}" + }, + { + "bbox": [ + 55, + 422, + 296, + 470 + ], + "type": "text", + "content": " is said to pairwise commute if every pair of matrices within the set commutes with each other. That is, for all " + }, + { + "bbox": [ + 55, + 422, + 296, + 470 + ], + "type": "inline_equation", + "content": "i,j" + }, + { + "bbox": [ + 55, + 422, + 296, + 470 + ], + "type": "text", + "content": " such that " + }, + { + "bbox": [ + 55, + 422, + 296, + 470 + ], + "type": "inline_equation", + "content": "1\\leq i,j\\leq N" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 141, + 471, + 209, + 483 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 471, + 209, + 483 + ], + "spans": [ + { + "bbox": [ + 141, + 471, + 209, + 483 + ], + "type": "interline_equation", + "content": "\\mathbf {A} _ {i} \\mathbf {A} _ {j} = \\mathbf {A} _ {j} \\mathbf {A} _ {i}.", + "image_path": "c4ae3b6800f5dd42da25cd97ddde33b34a8243d5795305117f696d10f547e88b.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 488, + 296, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 488, + 296, + 608 + ], + "spans": [ + { + "bbox": [ + 55, + 488, + 296, + 608 + ], + "type": "text", + "content": "To clarify and better illustrate the main theorems presented in the following sections, we first reformulate and unify the definitions of the RoPE paradigm, specifically in the context of multi-axial attention mechanisms. RoPE was initially proposed by Su et al. [27] as a positional encoding method based on relative position dependencies. However, previous work provided only conceptual and descriptive descriptions of RPE and RoPE without offering a rigorous formal definition. In this work, we provide the formal definitions of both RPE and RoPE." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 53, + 613, + 296, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 613, + 296, + 674 + ], + "spans": [ + { + "bbox": [ + 53, + 613, + 296, + 674 + ], + "type": "text", + "content": "Definition 3 (RPE Equation). Let " + }, + { + "bbox": [ + 53, + 613, + 296, + 674 + ], + "type": "inline_equation", + "content": "f: \\mathbb{R}^d \\times \\mathbb{R}^N \\to \\mathbb{R}^d" + }, + { + "bbox": [ + 53, + 613, + 296, + 674 + ], + "type": "text", + "content": " be a positional encoding function. " + }, + { + "bbox": [ + 53, + 613, + 296, + 674 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 53, + 613, + 296, + 674 + ], + "type": "text", + "content": " is said to be a RPE function if and only if there exists a function " + }, + { + "bbox": [ + 53, + 613, + 296, + 674 + ], + "type": "inline_equation", + "content": "\\rho: \\mathbb{R}^d \\times \\mathbb{R}^d \\to \\mathbb{R}" + }, + { + "bbox": [ + 53, + 613, + 296, + 674 + ], + "type": "text", + "content": ", and a function " + }, + { + "bbox": [ + 53, + 613, + 296, + 674 + ], + "type": "inline_equation", + "content": "g: \\mathbb{R}^d \\times \\mathbb{R}^d \\times \\mathbb{R}^N \\to \\mathbb{R}" + }, + { + "bbox": [ + 53, + 613, + 296, + 674 + ], + "type": "text", + "content": " such that the following conditions hold for all " + }, + { + "bbox": [ + 53, + 613, + 296, + 674 + ], + "type": "inline_equation", + "content": "x, y \\in \\mathbb{R}^N" + }, + { + "bbox": [ + 53, + 613, + 296, + 674 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 53, + 613, + 296, + 674 + ], + "type": "inline_equation", + "content": "q, k \\in \\mathbb{R}^d" + }, + { + "bbox": [ + 53, + 613, + 296, + 674 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 97, + 685, + 295, + 698 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 685, + 295, + 698 + ], + "spans": [ + { + "bbox": [ + 97, + 685, + 295, + 698 + ], + "type": "interline_equation", + "content": "g (\\boldsymbol {q}, \\boldsymbol {k}, \\boldsymbol {x} - \\boldsymbol {y}) = \\rho (f (\\boldsymbol {q}, \\boldsymbol {x}), f (\\boldsymbol {k}, \\boldsymbol {y})), \\tag {1}", + "image_path": "7a21d4b1eb90cf32de30aec2ff7641f1b6e233d91c901c93c8ec970c3099b2b6.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 701, + 223, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 701, + 223, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 701, + 223, + 714 + ], + "type": "text", + "content": "We refer to Eq. (1) as the RPE Equation." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 71, + 555, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 71, + 555, + 133 + ], + "spans": [ + { + "bbox": [ + 313, + 71, + 555, + 133 + ], + "type": "text", + "content": "Definition 4 (RoPE Equation). Let " + }, + { + "bbox": [ + 313, + 71, + 555, + 133 + ], + "type": "inline_equation", + "content": "f: \\mathbb{R}^d \\times \\mathbb{R}^N \\to \\mathbb{R}^d" + }, + { + "bbox": [ + 313, + 71, + 555, + 133 + ], + "type": "text", + "content": " be an RPE function. " + }, + { + "bbox": [ + 313, + 71, + 555, + 133 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 313, + 71, + 555, + 133 + ], + "type": "text", + "content": " is said to be a RoPE function if and only if there exists a matrix-valued function " + }, + { + "bbox": [ + 313, + 71, + 555, + 133 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_f(\\cdot)" + }, + { + "bbox": [ + 313, + 71, + 555, + 133 + ], + "type": "text", + "content": " such that the following conditions hold for all " + }, + { + "bbox": [ + 313, + 71, + 555, + 133 + ], + "type": "inline_equation", + "content": "\\mathbf{x}, \\mathbf{y} \\in \\mathbb{R}^N" + }, + { + "bbox": [ + 313, + 71, + 555, + 133 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 71, + 555, + 133 + ], + "type": "inline_equation", + "content": "\\mathbf{k}, \\mathbf{k} \\in \\mathbb{R}^d" + }, + { + "bbox": [ + 313, + 71, + 555, + 133 + ], + "type": "text", + "content": " in RPE Equation:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 353, + 140, + 555, + 188 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 140, + 555, + 188 + ], + "spans": [ + { + "bbox": [ + 353, + 140, + 555, + 188 + ], + "type": "interline_equation", + "content": "\\left\\{ \\begin{array}{l} f (\\boldsymbol {q}, \\boldsymbol {x}) = \\mathbf {R} _ {f} (\\boldsymbol {x}) \\boldsymbol {q} \\\\ \\rho (\\boldsymbol {q}, \\boldsymbol {k}) = \\boldsymbol {q} ^ {\\top} \\boldsymbol {k} \\\\ g (\\boldsymbol {q}, \\boldsymbol {k}, \\boldsymbol {x} - \\boldsymbol {y}) = \\boldsymbol {q} ^ {\\top} \\mathbf {R} _ {f} (\\boldsymbol {y} - \\boldsymbol {x}) \\boldsymbol {k} \\end{array} \\right. \\tag {2}", + "image_path": "68d2bcba81c566139e80cff970107915522a490d299656553b0f1004e8fa9686.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 314, + 195, + 487, + 208 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 195, + 487, + 208 + ], + "spans": [ + { + "bbox": [ + 314, + 195, + 487, + 208 + ], + "type": "text", + "content": "We refer to Eq. (2) as the RoPE Equation." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 215, + 554, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 215, + 554, + 251 + ], + "spans": [ + { + "bbox": [ + 313, + 215, + 554, + 251 + ], + "type": "text", + "content": "By substituting the RoPE equation into the RPE equation, we obtain that the RoPE function satisfies the following property." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 258, + 554, + 284 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 258, + 554, + 284 + ], + "spans": [ + { + "bbox": [ + 313, + 258, + 554, + 284 + ], + "type": "text", + "content": "Proposition 1. " + }, + { + "bbox": [ + 313, + 258, + 554, + 284 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 313, + 258, + 554, + 284 + ], + "type": "text", + "content": " is said to be a RoPE function if and only if the matrix-valued function " + }, + { + "bbox": [ + 313, + 258, + 554, + 284 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_f(\\cdot)" + }, + { + "bbox": [ + 313, + 258, + 554, + 284 + ], + "type": "text", + "content": " satisfies:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 369, + 291, + 498, + 306 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 369, + 291, + 498, + 306 + ], + "spans": [ + { + "bbox": [ + 369, + 291, + 498, + 306 + ], + "type": "interline_equation", + "content": "\\mathbf {R} _ {f} (\\boldsymbol {x}) ^ {\\top} \\mathbf {R} _ {f} (\\boldsymbol {y}) = \\mathbf {R} _ {f} (\\boldsymbol {y} - \\boldsymbol {x}),", + "image_path": "1969fef67c4ca388035073f75e5b03a88d76d2897ce84792d75be242ffcf5dc7.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 314, + 391, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 314, + 391, + 327 + ], + "spans": [ + { + "bbox": [ + 313, + 314, + 391, + 327 + ], + "type": "text", + "content": "for all " + }, + { + "bbox": [ + 313, + 314, + 391, + 327 + ], + "type": "inline_equation", + "content": "\\pmb {x},\\pmb {y}\\in \\mathbb{R}^{N}" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 334, + 554, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 334, + 554, + 407 + ], + "spans": [ + { + "bbox": [ + 313, + 334, + 554, + 407 + ], + "type": "text", + "content": "Note that the definition of RoPE Equation demonstrates that the position encoding should only be dependent on the relative position of the tokens, which thus be robust against the offset operations. We further expand the definition of the RoPE function to a parameterized one (i.e., Definition 6) via rotation matrices (i.e., Definition 5) as follows1." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 414, + 555, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 414, + 555, + 487 + ], + "spans": [ + { + "bbox": [ + 313, + 414, + 555, + 487 + ], + "type": "text", + "content": "Definition 5 (Parameterized Rotation Matrix). Let " + }, + { + "bbox": [ + 313, + 414, + 555, + 487 + ], + "type": "inline_equation", + "content": "\\mathbf{R}(\\cdot ;\\mathcal{A}): \\mathbb{R}^N \\to \\mathbb{R}^{d\\times d}" + }, + { + "bbox": [ + 313, + 414, + 555, + 487 + ], + "type": "text", + "content": " be a matrix-valued function parameterized by " + }, + { + "bbox": [ + 313, + 414, + 555, + 487 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 414, + 555, + 487 + ], + "type": "text", + "content": " skew-symmetric matrices " + }, + { + "bbox": [ + 313, + 414, + 555, + 487 + ], + "type": "inline_equation", + "content": "\\mathcal{A} = \\{\\mathbf{A}_1,\\mathbf{A}_2,\\dots ,\\mathbf{A}_N\\}" + }, + { + "bbox": [ + 313, + 414, + 555, + 487 + ], + "type": "text", + "content": ". We say that " + }, + { + "bbox": [ + 313, + 414, + 555, + 487 + ], + "type": "inline_equation", + "content": "\\mathbf{R}(\\cdot ;\\mathcal{A})" + }, + { + "bbox": [ + 313, + 414, + 555, + 487 + ], + "type": "text", + "content": " is a rotation matrix function parameterized by angle matrices " + }, + { + "bbox": [ + 313, + 414, + 555, + 487 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 313, + 414, + 555, + 487 + ], + "type": "text", + "content": " if it can be expressed as:" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 370, + 494, + 554, + 528 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 494, + 554, + 528 + ], + "spans": [ + { + "bbox": [ + 370, + 494, + 554, + 528 + ], + "type": "interline_equation", + "content": "\\mathbf {R} (\\boldsymbol {x}; \\mathcal {A}) = \\exp \\left(\\sum_ {i = 1} ^ {N} \\mathbf {A} _ {i} x _ {i}\\right), \\tag {3}", + "image_path": "a09c27207b1f3cb07ef83a3c0e5852957734d45c040494c7386cc8def1741a0e.jpg" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 536, + 554, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 536, + 554, + 561 + ], + "spans": [ + { + "bbox": [ + 313, + 536, + 554, + 561 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 536, + 554, + 561 + ], + "type": "inline_equation", + "content": "\\pmb{x} = (x_{1}, x_{2}, \\ldots, x_{N}) \\in \\mathbb{R}^{N}" + }, + { + "bbox": [ + 313, + 536, + 554, + 561 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 536, + 554, + 561 + ], + "type": "inline_equation", + "content": "\\exp(\\cdot)" + }, + { + "bbox": [ + 313, + 536, + 554, + 561 + ], + "type": "text", + "content": " denotes the matrix exponential." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 567, + 554, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 567, + 554, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 567, + 554, + 628 + ], + "type": "text", + "content": "Definition 6 (Parameterized RoPE Function). Let " + }, + { + "bbox": [ + 313, + 567, + 554, + 628 + ], + "type": "inline_equation", + "content": "f: \\mathbb{R}^d \\times \\mathbb{R}^N \\to \\mathbb{R}^d" + }, + { + "bbox": [ + 313, + 567, + 554, + 628 + ], + "type": "text", + "content": " be a RoPE function. " + }, + { + "bbox": [ + 313, + 567, + 554, + 628 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 313, + 567, + 554, + 628 + ], + "type": "text", + "content": " is said to be parameterized by angle matrices if and only if there exists a rotation matrix function " + }, + { + "bbox": [ + 313, + 567, + 554, + 628 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_f(\\cdot; \\mathcal{A})" + }, + { + "bbox": [ + 313, + 567, + 554, + 628 + ], + "type": "text", + "content": " parameterized by angle matrices " + }, + { + "bbox": [ + 313, + 567, + 554, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 313, + 567, + 554, + 628 + ], + "type": "text", + "content": " such that the RoPE Equation (i.e., Eq. (2)) holds." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 628, + 554, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 628, + 554, + 688 + ], + "spans": [ + { + "bbox": [ + 313, + 628, + 554, + 688 + ], + "type": "text", + "content": "In this case, " + }, + { + "bbox": [ + 313, + 628, + 554, + 688 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_f(\\cdot ;\\mathcal{A})" + }, + { + "bbox": [ + 313, + 628, + 554, + 688 + ], + "type": "text", + "content": " is referred to as the rotation matrix function of RoPE function " + }, + { + "bbox": [ + 313, + 628, + 554, + 688 + ], + "type": "inline_equation", + "content": "f(\\cdot ;\\mathcal{A})" + }, + { + "bbox": [ + 313, + 628, + 554, + 688 + ], + "type": "text", + "content": " parameterized by angle matrices " + }, + { + "bbox": [ + 313, + 628, + 554, + 688 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 313, + 628, + 554, + 688 + ], + "type": "text", + "content": ". For simplicity, we slightly abuse the notation and refer to " + }, + { + "bbox": [ + 313, + 628, + 554, + 688 + ], + "type": "inline_equation", + "content": "\\mathbf{R}" + }, + { + "bbox": [ + 313, + 628, + 554, + 688 + ], + "type": "text", + "content": " as the rotation matrix of RoPE function " + }, + { + "bbox": [ + 313, + 628, + 554, + 688 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 313, + 628, + 554, + 688 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 313, + 693, + 554, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 693, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 693, + 554, + 713 + ], + "type": "text", + "content": "For clarification, the definitions of the rotation matrix and its exponential representation can be found in Appendix A." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "4510" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 93, + 68, + 519, + 312 + ], + "blocks": [ + { + "bbox": [ + 93, + 68, + 519, + 312 + ], + "lines": [ + { + "bbox": [ + 93, + 68, + 519, + 312 + ], + "spans": [ + { + "bbox": [ + 93, + 68, + 519, + 312 + ], + "type": "image", + "image_path": "65b5ac4bfd70910ee3a5a462d8ee4798cac3086e6d8b276c4d6a36bb8a61650f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 323, + 555, + 357 + ], + "lines": [ + { + "bbox": [ + 54, + 323, + 555, + 357 + ], + "spans": [ + { + "bbox": [ + 54, + 323, + 555, + 357 + ], + "type": "text", + "content": "Figure 1. Overview of ComRoPE. Features are arranged into several blocks, each representing a distinct point in the feature space. The positions, along with the angle matrices, define the rotation matrix, which is an element of the special orthogonal group. The rotation transformation projects a feature point onto another point on the surface of the same hypersphere." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 376, + 149, + 388 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 376, + 149, + 388 + ], + "spans": [ + { + "bbox": [ + 55, + 376, + 149, + 388 + ], + "type": "text", + "content": "3.2. Main theorems" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 396, + 295, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 396, + 295, + 419 + ], + "spans": [ + { + "bbox": [ + 55, + 396, + 295, + 419 + ], + "type": "text", + "content": "Based on the formal definitions above, we present our key theoretical results in the following theorem." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 423, + 296, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 423, + 296, + 483 + ], + "spans": [ + { + "bbox": [ + 55, + 423, + 296, + 483 + ], + "type": "text", + "content": "Theorem 1. Let " + }, + { + "bbox": [ + 55, + 423, + 296, + 483 + ], + "type": "inline_equation", + "content": "\\mathbf{R}(\\cdot ;\\mathcal{A}): \\mathbb{R}^N \\to \\mathbb{R}^{d\\times d}" + }, + { + "bbox": [ + 55, + 423, + 296, + 483 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 55, + 423, + 296, + 483 + ], + "type": "inline_equation", + "content": "N > 1" + }, + { + "bbox": [ + 55, + 423, + 296, + 483 + ], + "type": "text", + "content": ") be a rotation matrix function parameterized by angle matrices " + }, + { + "bbox": [ + 55, + 423, + 296, + 483 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 55, + 423, + 296, + 483 + ], + "type": "text", + "content": ". The rotation difference " + }, + { + "bbox": [ + 55, + 423, + 296, + 483 + ], + "type": "inline_equation", + "content": "\\mathbf{R}(\\pmb{x})^\\top \\mathbf{R}(\\pmb{y})" + }, + { + "bbox": [ + 55, + 423, + 296, + 483 + ], + "type": "text", + "content": " can be represented by the location difference " + }, + { + "bbox": [ + 55, + 423, + 296, + 483 + ], + "type": "inline_equation", + "content": "\\pmb{y} - \\pmb{x}" + }, + { + "bbox": [ + 55, + 423, + 296, + 483 + ], + "type": "text", + "content": " if and only if " + }, + { + "bbox": [ + 55, + 423, + 296, + 483 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 55, + 423, + 296, + 483 + ], + "type": "text", + "content": " pairwise commute." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 488, + 269, + 500 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 488, + 269, + 500 + ], + "spans": [ + { + "bbox": [ + 67, + 488, + 269, + 500 + ], + "type": "text", + "content": "The proof of Theorem 1 is shown in Appendix A." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 500, + 296, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 500, + 296, + 632 + ], + "spans": [ + { + "bbox": [ + 55, + 500, + 296, + 632 + ], + "type": "text", + "content": "Theorem 1 together with Proposition 1 demonstrate that a function " + }, + { + "bbox": [ + 55, + 500, + 296, + 632 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 55, + 500, + 296, + 632 + ], + "type": "text", + "content": " is a RoPE function parameterized by angle matrices if and only if the angle matrices " + }, + { + "bbox": [ + 55, + 500, + 296, + 632 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 55, + 500, + 296, + 632 + ], + "type": "text", + "content": " pairwise commute. In other words, to construct a relative positional encoding method for an attention mechanism that is robust to offset, it suffices to establish an angle matrix set that pairwise commutes. Thus, in the following section, we focus on the construction of " + }, + { + "bbox": [ + 55, + 500, + 296, + 632 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 55, + 500, + 296, + 632 + ], + "type": "text", + "content": " to satisfy this particular requirement. We call our method ComRoPE to indicate the commutativity of angle matrices in RoPE function. The overview of ComRoPE is shown in Figure 1." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 635, + 296, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 635, + 296, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 635, + 296, + 696 + ], + "type": "text", + "content": "Remark 1. Among all the previous RoPE methods, LieRE [20] is the most related one to ours which solves Equation 3 by directly training the skew-symmetric matrix set " + }, + { + "bbox": [ + 55, + 635, + 296, + 696 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 55, + 635, + 296, + 696 + ], + "type": "text", + "content": ". However, it is worth noting that in the following equation of their implementation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 63, + 700, + 296, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 700, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 63, + 700, + 296, + 715 + ], + "type": "interline_equation", + "content": "\\left(\\mathbf {R} (\\boldsymbol {x} _ {q}; \\mathcal {A}) \\boldsymbol {q}\\right) ^ {\\top} \\left(\\mathbf {R} (\\boldsymbol {x} _ {k}; \\mathcal {A}) \\boldsymbol {k}\\right) = \\boldsymbol {q} ^ {\\top} \\mathbf {R} (\\boldsymbol {x} _ {q}; \\mathcal {A}) ^ {\\top} \\mathbf {R} (\\boldsymbol {x} _ {k}; \\mathcal {A}) \\boldsymbol {k}, (4)", + "image_path": "94575ec0fe5ac0bc696ed9912d94384fe3ddb73ff7907bc221e0e28eb1565acc.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 376, + 554, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 376, + 554, + 437 + ], + "spans": [ + { + "bbox": [ + 313, + 376, + 554, + 437 + ], + "type": "text", + "content": "sometimes " + }, + { + "bbox": [ + 313, + 376, + 554, + 437 + ], + "type": "inline_equation", + "content": "\\mathbf{R}(\\boldsymbol{x}_q; \\mathcal{A})^\\top \\mathbf{R}(\\boldsymbol{x}_k; \\mathcal{A}) \\neq \\mathbf{R}(\\boldsymbol{x}_q - \\boldsymbol{x}_k; \\mathcal{A})" + }, + { + "bbox": [ + 313, + 376, + 554, + 437 + ], + "type": "text", + "content": " because the probability that two random matrices commute is small, where " + }, + { + "bbox": [ + 313, + 376, + 554, + 437 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}_q" + }, + { + "bbox": [ + 313, + 376, + 554, + 437 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 376, + 554, + 437 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}_k" + }, + { + "bbox": [ + 313, + 376, + 554, + 437 + ], + "type": "text", + "content": " are two arbitrary coordinates. Thus, " + }, + { + "bbox": [ + 313, + 376, + 554, + 437 + ], + "type": "inline_equation", + "content": "\\mathbf{R}" + }, + { + "bbox": [ + 313, + 376, + 554, + 437 + ], + "type": "text", + "content": " proposed by LieRe does not consistently satisfy the requirements of the RoPE Equation." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 442, + 552, + 455 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 442, + 552, + 455 + ], + "spans": [ + { + "bbox": [ + 313, + 442, + 552, + 455 + ], + "type": "text", + "content": "3.3. Construction of pairwise commuting matrices" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 460, + 555, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 460, + 555, + 544 + ], + "spans": [ + { + "bbox": [ + 313, + 460, + 555, + 544 + ], + "type": "text", + "content": "In this section, we elaborate on concrete ways to construct the pairwise commuting matrices to solve the RoPE Equation parameterized by angle matrices. Note that if two matrices are both block diagonal with the same block sizes, where the corresponding blocks are commutative, then these two matrices are commutative. Formally speaking, for two matrices " + }, + { + "bbox": [ + 313, + 460, + 555, + 544 + ], + "type": "inline_equation", + "content": "\\mathbf{M},\\mathbf{N}\\in \\mathbb{R}^{d\\times d}" + }, + { + "bbox": [ + 313, + 460, + 555, + 544 + ], + "type": "text", + "content": ", they commute if" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 341, + 550, + 554, + 615 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 341, + 550, + 554, + 615 + ], + "spans": [ + { + "bbox": [ + 341, + 550, + 554, + 615 + ], + "type": "interline_equation", + "content": "\\left\\{ \\begin{array}{l} \\mathbf {M} = \\operatorname {d i a g} \\left(\\mathbf {M} _ {1}, \\mathbf {M} _ {2}, \\dots , \\mathbf {M} _ {b}\\right) \\\\ \\mathbf {N} = \\operatorname {d i a g} \\left(\\mathbf {N} _ {1}, \\mathbf {N} _ {2}, \\dots , \\mathbf {N} _ {b}\\right) \\\\ \\mathbf {M} _ {i}, \\mathbf {N} _ {i} \\in \\mathbb {R} ^ {b \\times b} \\quad \\forall i \\in \\{1, 2, \\dots , m \\} \\\\ \\mathbf {M} _ {i} \\mathbf {N} _ {i} = \\mathbf {N} _ {i} \\mathbf {M} _ {i} \\quad \\forall i \\in \\{1, 2, \\dots , m \\} \\end{array} \\right. \\tag {5}", + "image_path": "2cf1485691e485315ec44ad62cb8ca7fac036ba1f3d5c0862d711ec960a7ab09.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 620, + 552, + 633 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 620, + 552, + 633 + ], + "spans": [ + { + "bbox": [ + 313, + 620, + 552, + 633 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 620, + 552, + 633 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 313, + 620, + 552, + 633 + ], + "type": "text", + "content": " denotes block size that is a factor of " + }, + { + "bbox": [ + 313, + 620, + 552, + 633 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 313, + 620, + 552, + 633 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 620, + 552, + 633 + ], + "type": "inline_equation", + "content": "m = \\frac{d}{b}" + }, + { + "bbox": [ + 313, + 620, + 552, + 633 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 633, + 553, + 679 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 633, + 553, + 679 + ], + "spans": [ + { + "bbox": [ + 313, + 633, + 553, + 679 + ], + "type": "text", + "content": "Thus, to present solutions to RoPE Equation parameterized by angle matrices, it suffices to partition the angle matrices " + }, + { + "bbox": [ + 313, + 633, + 553, + 679 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_i" + }, + { + "bbox": [ + 313, + 633, + 553, + 679 + ], + "type": "text", + "content": " in Eq. (3) into " + }, + { + "bbox": [ + 313, + 633, + 553, + 679 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 633, + 553, + 679 + ], + "type": "text", + "content": " blocks " + }, + { + "bbox": [ + 313, + 633, + 553, + 679 + ], + "type": "inline_equation", + "content": "\\mathbf{B}_{i1},\\mathbf{B}_{i2},\\dots ,\\mathbf{B}_{im}" + }, + { + "bbox": [ + 313, + 633, + 553, + 679 + ], + "type": "text", + "content": " where:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 350, + 684, + 553, + 716 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 684, + 553, + 716 + ], + "spans": [ + { + "bbox": [ + 350, + 684, + 553, + 716 + ], + "type": "interline_equation", + "content": "\\left\\{ \\begin{array}{l} \\mathbf {A} _ {i} = \\operatorname {d i a g} \\left(\\mathbf {B} _ {i 1}, \\mathbf {B} _ {i 2}, \\dots , \\mathbf {B} _ {i m}\\right) \\\\ \\mathbf {B} _ {i j} \\in \\mathbb {R} ^ {b \\times b} \\quad \\forall j \\in \\{1, 2, \\dots , m \\} \\end{array} \\right. \\tag {6}", + "image_path": "e3ba778483d5c1eeb2d420e4c8bdf80f2d5fa306d30eeac1b10afd3acbfc3718.jpg" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 314, + 757 + ], + "type": "text", + "content": "4511" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 144 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 144 + ], + "type": "text", + "content": "Defining " + }, + { + "bbox": [ + 55, + 72, + 295, + 144 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_j = \\{\\mathbf{B}_{1j},\\mathbf{B}_{2j},\\dots ,\\mathbf{B}_{Nj}\\}" + }, + { + "bbox": [ + 55, + 72, + 295, + 144 + ], + "type": "text", + "content": ", if " + }, + { + "bbox": [ + 55, + 72, + 295, + 144 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_j" + }, + { + "bbox": [ + 55, + 72, + 295, + 144 + ], + "type": "text", + "content": " pairwise commutes for all " + }, + { + "bbox": [ + 55, + 72, + 295, + 144 + ], + "type": "inline_equation", + "content": "j\\in \\{1,2,\\dots ,m\\}" + }, + { + "bbox": [ + 55, + 72, + 295, + 144 + ], + "type": "text", + "content": ", then " + }, + { + "bbox": [ + 55, + 72, + 295, + 144 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 55, + 72, + 295, + 144 + ], + "type": "text", + "content": " pairwise commutes. Note that " + }, + { + "bbox": [ + 55, + 72, + 295, + 144 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_1,\\mathcal{B}_2,\\dots ,\\mathcal{B}_m" + }, + { + "bbox": [ + 55, + 72, + 295, + 144 + ], + "type": "text", + "content": " are equivalent. Thus, without causing confusion, we will uniformly use " + }, + { + "bbox": [ + 55, + 72, + 295, + 144 + ], + "type": "inline_equation", + "content": "\\mathcal{B} = \\{\\mathbf{B}_1,\\mathbf{B}_2,\\dots ,\\mathbf{B}_N\\}" + }, + { + "bbox": [ + 55, + 72, + 295, + 144 + ], + "type": "text", + "content": " to represent " + }, + { + "bbox": [ + 55, + 72, + 295, + 144 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_j = \\{\\mathbf{B}_{1j},\\mathbf{B}_{2j},\\dots ,\\mathbf{B}_{Nj}\\}" + }, + { + "bbox": [ + 55, + 72, + 295, + 144 + ], + "type": "text", + "content": " in the following." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 144, + 296, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 144, + 296, + 228 + ], + "spans": [ + { + "bbox": [ + 55, + 144, + 296, + 228 + ], + "type": "text", + "content": "Currently, there is no general method that provides the necessary and sufficient conditions for ensuring that arbitrary trainable skew-symmetric matrices commute. Therefore, we aim to establish sufficient conditions for enforcing this constraint. More concretely, we present Proposition 2 and Proposition 3 to construct two different variants of matrices, respectively." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 234, + 296, + 271 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 234, + 296, + 271 + ], + "spans": [ + { + "bbox": [ + 54, + 234, + 296, + 271 + ], + "type": "text", + "content": "Proposition 2. A set of matrices " + }, + { + "bbox": [ + 54, + 234, + 296, + 271 + ], + "type": "inline_equation", + "content": "\\mathcal{B} = \\{\\mathbf{B}_1, \\mathbf{B}_2, \\dots, \\mathbf{B}_N\\}" + }, + { + "bbox": [ + 54, + 234, + 296, + 271 + ], + "type": "text", + "content": " pairwise commutes if all but one of them are zero matrices, i.e., " + }, + { + "bbox": [ + 54, + 234, + 296, + 271 + ], + "type": "inline_equation", + "content": "\\exists k \\in \\{1, 2, \\dots, N\\}" + }, + { + "bbox": [ + 54, + 234, + 296, + 271 + ], + "type": "text", + "content": " s.t." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 137, + 278, + 294, + 292 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 278, + 294, + 292 + ], + "spans": [ + { + "bbox": [ + 137, + 278, + 294, + 292 + ], + "type": "interline_equation", + "content": "\\mathbf {B} _ {i} = \\mathbf {O} (\\forall i \\neq k) \\tag {7}", + "image_path": "5e20a9d8edda5e3276d82313120d9a4970d53685388d21bced55fd845e75ffdb.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 300, + 296, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 300, + 296, + 371 + ], + "spans": [ + { + "bbox": [ + 55, + 300, + 296, + 371 + ], + "type": "text", + "content": "Based on Proposition 2, we propose Trainable RoPE parameterized by Axial-Partition Angle Matrices (ComRoPE-AP). We divide the " + }, + { + "bbox": [ + 55, + 300, + 296, + 371 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 55, + 300, + 296, + 371 + ], + "type": "text", + "content": " blocks into " + }, + { + "bbox": [ + 55, + 300, + 296, + 371 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 300, + 296, + 371 + ], + "type": "text", + "content": " parts, such that each part is responsible for the rotation of a specific axis. In this case, " + }, + { + "bbox": [ + 55, + 300, + 296, + 371 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 55, + 300, + 296, + 371 + ], + "type": "text", + "content": " should be a multiple of " + }, + { + "bbox": [ + 55, + 300, + 296, + 371 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 300, + 296, + 371 + ], + "type": "text", + "content": ". Specifically," + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 69, + 378, + 294, + 404 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 378, + 294, + 404 + ], + "spans": [ + { + "bbox": [ + 69, + 378, + 294, + 404 + ], + "type": "interline_equation", + "content": "\\mathbf {B} _ {i j} = \\left\\{ \\begin{array}{l l} \\mathbf {P} _ {j} - \\mathbf {P} _ {j} ^ {\\top}, & \\text {i f} j \\equiv i (\\mathrm {m o d} N) \\\\ \\mathbf {O}, & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {8}", + "image_path": "f14740a45764077b94c27f3f4446df8c365b2d56356696ec1ba92985db82324b.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 412, + 295, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 412, + 295, + 437 + ], + "spans": [ + { + "bbox": [ + 55, + 412, + 295, + 437 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 412, + 295, + 437 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{P}_j\\}_{j=1}^m" + }, + { + "bbox": [ + 55, + 412, + 295, + 437 + ], + "type": "text", + "content": " represents a set of trainable matrices, " + }, + { + "bbox": [ + 55, + 412, + 295, + 437 + ], + "type": "inline_equation", + "content": "\\mathbf{O}" + }, + { + "bbox": [ + 55, + 412, + 295, + 437 + ], + "type": "text", + "content": " denotes a zero matrix, and " + }, + { + "bbox": [ + 55, + 412, + 295, + 437 + ], + "type": "inline_equation", + "content": "\\equiv" + }, + { + "bbox": [ + 55, + 412, + 295, + 437 + ], + "type": "text", + "content": " indicates congruence modulo." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 443, + 296, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 443, + 296, + 479 + ], + "spans": [ + { + "bbox": [ + 55, + 443, + 296, + 479 + ], + "type": "text", + "content": "Proposition 3. A set of matrices " + }, + { + "bbox": [ + 55, + 443, + 296, + 479 + ], + "type": "inline_equation", + "content": "\\mathcal{B} = \\{\\mathbf{B}_1, \\mathbf{B}_2, \\dots, \\mathbf{B}_N\\}" + }, + { + "bbox": [ + 55, + 443, + 296, + 479 + ], + "type": "text", + "content": " pairwise commutes if they pairwise linearly dependent, i.e., " + }, + { + "bbox": [ + 55, + 443, + 296, + 479 + ], + "type": "inline_equation", + "content": "\\exists \\lambda_1, \\lambda_2, \\dots, \\lambda_N \\in \\mathbb{R}" + }, + { + "bbox": [ + 55, + 443, + 296, + 479 + ], + "type": "text", + "content": " s.t." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 110, + 487, + 294, + 500 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 487, + 294, + 500 + ], + "spans": [ + { + "bbox": [ + 110, + 487, + 294, + 500 + ], + "type": "interline_equation", + "content": "\\lambda_ {1} \\mathbf {B} _ {1} = \\lambda_ {2} \\mathbf {B} _ {2} = \\dots = \\lambda_ {N} \\mathbf {B} _ {N} \\tag {9}", + "image_path": "ac3534aacf3d4bba9523d4da0e5d34dee587fb39be5bcce51718da3a96bc6194.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 508, + 296, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 508, + 296, + 558 + ], + "spans": [ + { + "bbox": [ + 55, + 508, + 296, + 558 + ], + "type": "text", + "content": "Based on Proposition 3, we propose Trainable RoPE parameterized by Linearly-Dependent Angle Matrices (ComRoPE-LD). Specifically, We train a base matrix " + }, + { + "bbox": [ + 55, + 508, + 296, + 558 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 55, + 508, + 296, + 558 + ], + "type": "text", + "content": " with scaling factors " + }, + { + "bbox": [ + 55, + 508, + 296, + 558 + ], + "type": "inline_equation", + "content": "\\{\\theta_i\\}_{i=1}^N" + }, + { + "bbox": [ + 55, + 508, + 296, + 558 + ], + "type": "text", + "content": ". Then we obtain:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 76, + 563, + 294, + 578 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 563, + 294, + 578 + ], + "spans": [ + { + "bbox": [ + 76, + 563, + 294, + 578 + ], + "type": "interline_equation", + "content": "\\mathcal {B} = \\left\\{\\mathbf {B} _ {i} = \\theta_ {i} \\left(\\mathbf {P} - \\mathbf {P} ^ {\\top}\\right) \\mid i = 1, 2, \\dots , N \\right\\} \\tag {10}", + "image_path": "2b4cc42718112f079e34d07e9d306afbd01f8735aa4f63c33c88a4f716a71206.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 592, + 295, + 617 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 592, + 295, + 617 + ], + "spans": [ + { + "bbox": [ + 55, + 592, + 295, + 617 + ], + "type": "text", + "content": "3.4. Implementation details of coordinates and improvements" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 623, + 229, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 623, + 229, + 635 + ], + "spans": [ + { + "bbox": [ + 55, + 623, + 229, + 635 + ], + "type": "text", + "content": "3.4.1 Relative scaling and center offset" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "type": "text", + "content": "Relative scaling. In language models, positions are typically treated as discrete. Additionally, due to the relationship between tokens and the inherent uncertainty in sequence length, there is no need to scale these positions into a specified range. However, in the case of images, positions are continuous. When using different patch sizes, it" + } + ] + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 316, + 57, + 555, + 258 + ], + "blocks": [ + { + "bbox": [ + 316, + 57, + 555, + 258 + ], + "lines": [ + { + "bbox": [ + 316, + 57, + 555, + 258 + ], + "spans": [ + { + "bbox": [ + 316, + 57, + 555, + 258 + ], + "type": "image", + "image_path": "39ff5b1cf2b7367ee8bd2e68d6c3653f0b289ba70dedd64f332c698644271be7.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 274, + 555, + 299 + ], + "lines": [ + { + "bbox": [ + 313, + 274, + 555, + 299 + ], + "spans": [ + { + "bbox": [ + 313, + 274, + 555, + 299 + ], + "type": "text", + "content": "Figure 2. Different patch sizes result in different relative relationships." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 319, + 554, + 367 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 319, + 554, + 367 + ], + "spans": [ + { + "bbox": [ + 313, + 319, + 554, + 367 + ], + "type": "text", + "content": "becomes unreasonable to define their relative relationships solely based on the span of patches. As shown in Figure 2, the same locations with varying patch sizes can result in significantly different relative relationships." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 368, + 555, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 368, + 555, + 479 + ], + "spans": [ + { + "bbox": [ + 313, + 368, + 555, + 479 + ], + "type": "text", + "content": "As a result, applying relative coordinates is a better method for measuring relative relationships in images. For an image with shape " + }, + { + "bbox": [ + 313, + 368, + 555, + 479 + ], + "type": "inline_equation", + "content": "H \\times W" + }, + { + "bbox": [ + 313, + 368, + 555, + 479 + ], + "type": "text", + "content": ", we scale both the height " + }, + { + "bbox": [ + 313, + 368, + 555, + 479 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 313, + 368, + 555, + 479 + ], + "type": "text", + "content": " and the width " + }, + { + "bbox": [ + 313, + 368, + 555, + 479 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 313, + 368, + 555, + 479 + ], + "type": "text", + "content": " to 1. Therefore, for a pixel located at " + }, + { + "bbox": [ + 313, + 368, + 555, + 479 + ], + "type": "inline_equation", + "content": "(h, w)" + }, + { + "bbox": [ + 313, + 368, + 555, + 479 + ], + "type": "text", + "content": " in the raw image, its coordinate is treated as " + }, + { + "bbox": [ + 313, + 368, + 555, + 479 + ], + "type": "inline_equation", + "content": "\\left(\\frac{h}{H}, \\frac{w}{W}\\right)" + }, + { + "bbox": [ + 313, + 368, + 555, + 479 + ], + "type": "text", + "content": ". For high-dimensional coordinates, we perform the same operation, i.e., transform a raw coordinate " + }, + { + "bbox": [ + 313, + 368, + 555, + 479 + ], + "type": "inline_equation", + "content": "(x_1, x_2, \\dots, x_N)" + }, + { + "bbox": [ + 313, + 368, + 555, + 479 + ], + "type": "text", + "content": " in multi-axial canvas with shape " + }, + { + "bbox": [ + 313, + 368, + 555, + 479 + ], + "type": "inline_equation", + "content": "(X_1, X_2, \\dots, X_N)" + }, + { + "bbox": [ + 313, + 368, + 555, + 479 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 313, + 368, + 555, + 479 + ], + "type": "inline_equation", + "content": "\\left(\\frac{x_1}{X_1}, \\frac{x_2}{X_2}, \\dots, \\frac{x_N}{X_N}\\right)" + }, + { + "bbox": [ + 313, + 368, + 555, + 479 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 483, + 554, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 483, + 554, + 532 + ], + "spans": [ + { + "bbox": [ + 313, + 483, + 554, + 532 + ], + "type": "text", + "content": "Center offset. When we project a patch to a feature tensor with a coordinate, we aggregate all information of the patch into a specified location. Simply, we adopt the center point of a patch as the aggregation location." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 549, + 440, + 561 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 549, + 440, + 561 + ], + "spans": [ + { + "bbox": [ + 313, + 549, + 440, + 561 + ], + "type": "text", + "content": "3.4.2 Position perturbation" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 569, + 554, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 569, + 554, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 569, + 554, + 640 + ], + "type": "text", + "content": "To achieve better robustness and excellent performance across different scales during inference, we add perturbations to the coordinates of the patches. Specifically, for a patch with center " + }, + { + "bbox": [ + 313, + 569, + 554, + 640 + ], + "type": "inline_equation", + "content": "(x_{1},\\dots ,x_{N})" + }, + { + "bbox": [ + 313, + 569, + 554, + 640 + ], + "type": "text", + "content": " and size " + }, + { + "bbox": [ + 313, + 569, + 554, + 640 + ], + "type": "inline_equation", + "content": "(\\Delta X_{1},\\dots ,\\Delta X_{N})" + }, + { + "bbox": [ + 313, + 569, + 554, + 640 + ], + "type": "text", + "content": ", during training, we formulate its location as:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 322, + 647, + 553, + 669 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 647, + 553, + 669 + ], + "spans": [ + { + "bbox": [ + 322, + 647, + 553, + 669 + ], + "type": "interline_equation", + "content": "\\left. \\mathcal {N} \\left(\\left(x _ {1}, \\dots , x _ {N}\\right) ^ {\\top}; \\operatorname {d i a g} \\left(\\sigma \\Delta X _ {1}, \\dots , \\sigma \\Delta X _ {N}\\right) ^ {2}\\right) \\right. \\tag {11}", + "image_path": "ee4558cd98920c5cd69cc9ca31260f83c9a08325366f28313d39c1b0141228c5.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 677, + 554, + 716 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 554, + 716 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 554, + 716 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 677, + 554, + 716 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 313, + 677, + 554, + 716 + ], + "type": "text", + "content": " is a hyper-parameter called perturbation intensity. Additionally, we truncate the location within the patch area, i.e., " + }, + { + "bbox": [ + 313, + 677, + 554, + 716 + ], + "type": "inline_equation", + "content": "\\left[x_k - \\frac{\\Delta X_k}{2}, x_k + \\frac{\\Delta X_k}{2}\\right]" + }, + { + "bbox": [ + 313, + 677, + 554, + 716 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "4512" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 63, + 70, + 547, + 245 + ], + "blocks": [ + { + "bbox": [ + 63, + 70, + 547, + 245 + ], + "lines": [ + { + "bbox": [ + 63, + 70, + 547, + 245 + ], + "spans": [ + { + "bbox": [ + 63, + 70, + 547, + 245 + ], + "type": "table", + "html": "
Position Encoding MethodPerturbation IntensityEvaluation Resolution
112128192224256320384448512
APE130.0438.6956.458.7660.0259.2757.0454.1050.99
019.7133.4355.9758.6255.3152.6349.6846.3942.66
Vanilla RoPE136.9445.4860.7263.0963.2462.1259.2455.5151.11
036.4144.4859.9762.0362.5461.3658.5654.7950.58
LieRE138.0346.9762.2264.3664.9963.7861.1557.9253.74
038.2246.8562.0164.0764.5463.4660.7456.8952.51
ComRoPE-AP (ours)135.7546.1862.8265.3265.8364.7861.8858.2154.02
037.1746.9562.6364.7665.1164.3561.6258.1053.81
ComRoPE-LD (ours)138.3047.2863.5365.4965.9565.2762.6259.1155.29
036.8846.5462.8965.2765.6664.2761.8357.8753.64
", + "image_path": "d0027c8cd1ec53bbbd96fb9d30bd7c2469d3efde9c8bb941dd604677fa2a38fb.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 247, + 553, + 258 + ], + "lines": [ + { + "bbox": [ + 55, + 247, + 553, + 258 + ], + "spans": [ + { + "bbox": [ + 55, + 247, + 553, + 258 + ], + "type": "text", + "content": "Table 1. Accuracy of 2D classification on ImageNet. Models are trained at a resolution of " + }, + { + "bbox": [ + 55, + 247, + 553, + 258 + ], + "type": "inline_equation", + "content": "{224} \\times {224}" + }, + { + "bbox": [ + 55, + 247, + 553, + 258 + ], + "type": "text", + "content": " and evaluated at varying resolutions." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 278, + 136, + 293 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 278, + 136, + 293 + ], + "spans": [ + { + "bbox": [ + 55, + 278, + 136, + 293 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 298, + 295, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 298, + 295, + 395 + ], + "spans": [ + { + "bbox": [ + 54, + 298, + 295, + 395 + ], + "type": "text", + "content": "In this section, we evaluate the performance of various positional encoding methods on classic vision tasks. We first assess their scalability in 2D image classification across different resolutions. Additionally, we conduct object detection experiments to demonstrate the generalizability of our approach. To further examine the ability to handle higher-dimensional data, we perform 3D classification experiments, which are detailed in Appendix B." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 401, + 154, + 414 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 401, + 154, + 414 + ], + "spans": [ + { + "bbox": [ + 55, + 401, + 154, + 414 + ], + "type": "text", + "content": "4.1. 2D classification" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 420, + 114, + 432 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 420, + 114, + 432 + ], + "spans": [ + { + "bbox": [ + 55, + 420, + 114, + 432 + ], + "type": "text", + "content": "4.1.1 Setup" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 437, + 295, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 437, + 295, + 581 + ], + "spans": [ + { + "bbox": [ + 54, + 437, + 295, + 581 + ], + "type": "text", + "content": "Baselines and model architecture. We evaluate our proposed methods (ComRoPE-LD and ComRoPE-AP) against APE, vanilla RoPE (as introduced by RoFormer), and LieRE. To isolate the effects of positional encoding, we utilize a standard Vision Transformer (ViT-B/16) architecture with minimal modifications. The APE codebook is removed for methods that do not employ APE, and self-attention layers are replaced with RoPE self-attention parameterized by angle matrices. This design highlights the performance differences among various positional encoding methods. A block size of 8 is used in practice. More details are provided in Appendix D." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 582, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 582, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 582, + 295, + 714 + ], + "type": "text", + "content": "Training and evaluation protocol. All models are trained at a standard resolution of " + }, + { + "bbox": [ + 54, + 582, + 295, + 714 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 54, + 582, + 295, + 714 + ], + "type": "text", + "content": " and evaluated across multiple resolutions to test their robustness and scalability on the ImageNet-1K dataset [4]. The models are trained from scratch using randomly initialized parameters, ensuring no influence from pre-trained weights or external priors. To maintain fairness and reproducibility, we apply only basic data augmentation techniques, such as resizing and random cropping, focusing on relative performance comparisons rather than achieving absolute accuracy. The primary evaluation metric is accuracy on the test set across" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 279, + 555, + 328 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 279, + 555, + 328 + ], + "spans": [ + { + "bbox": [ + 313, + 279, + 555, + 328 + ], + "type": "text", + "content": "various resolutions. Since APE is inherently fixed and discrete, bilinear interpolation is applied to adapt it to different resolutions during evaluation. Optimization methods and hyper-parameters are detailed in Appendix C." + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 319, + 342, + 548, + 415 + ], + "blocks": [ + { + "bbox": [ + 319, + 342, + 548, + 415 + ], + "lines": [ + { + "bbox": [ + 319, + 342, + 548, + 415 + ], + "spans": [ + { + "bbox": [ + 319, + 342, + 548, + 415 + ], + "type": "image", + "image_path": "91d57c34afde96ea9cce5466b8f19db5e9a5705309c57944613b82b58b5afcf8.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 320, + 422, + 548, + 494 + ], + "blocks": [ + { + "bbox": [ + 320, + 422, + 548, + 494 + ], + "lines": [ + { + "bbox": [ + 320, + 422, + 548, + 494 + ], + "spans": [ + { + "bbox": [ + 320, + 422, + 548, + 494 + ], + "type": "image", + "image_path": "1f72333d708d59e72203137d1ef3d51564cf45c13c96e0e5c875c35e811c6287.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 498, + 555, + 544 + ], + "lines": [ + { + "bbox": [ + 313, + 498, + 555, + 544 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 555, + 544 + ], + "type": "text", + "content": "Figure 3. Accuracy on ImageNet-1K for various positional encoding methods. The results for the same perturbation intensity are presented together for better comparison. For better visualization, evaluation resolution at " + }, + { + "bbox": [ + 313, + 498, + 555, + 544 + ], + "type": "inline_equation", + "content": "112 \\times 112" + }, + { + "bbox": [ + 313, + 498, + 555, + 544 + ], + "type": "text", + "content": " is not included in these figures." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 574, + 401, + 585 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 574, + 401, + 585 + ], + "spans": [ + { + "bbox": [ + 313, + 574, + 401, + 585 + ], + "type": "text", + "content": "4.1.2 Main results" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 594, + 554, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 554, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 554, + 640 + ], + "type": "text", + "content": "The accuracy metrics for each positional encoding method across various resolutions are summarized in Table 1. Additionally, Figure 3 visually compares performance under different levels of perturbation intensity separately." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 642, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 554, + 713 + ], + "type": "text", + "content": "Overall performance. Across all evaluations, APE consistently exhibits the lowest accuracy, corroborating previous findings regarding its limitations in dynamic contexts. The vanilla RoPE shows a modest improvement over APE but remains less effective. In contrast, the trainable angle matrices, namely LieRE, ComRoPE-AP, and ComRoPE-LD," + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "4513" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "type": "text", + "content": "demonstrate significantly higher accuracy across all resolutions. Notably, ComRoPE-LD achieves the best performance among the three, suggesting that its inherent linear dependencies may enhance flexibility and structural learning capabilities." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 134, + 294, + 230 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 134, + 294, + 230 + ], + "spans": [ + { + "bbox": [ + 55, + 134, + 294, + 230 + ], + "type": "text", + "content": "Accuracy at training resolution. At the training resolution of " + }, + { + "bbox": [ + 55, + 134, + 294, + 230 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 55, + 134, + 294, + 230 + ], + "type": "text", + "content": ", all three methods with trainable angle matrices (ComRoPE-LD, ComRoPE-AP, and LieRE) achieve comparable accuracy, significantly outperforming both APE and the standard RoPE, which underscores the effectiveness of RoPE parameterized by trainable angle matrices. Notably, ComRoPE-LD surpasses the current state-of-the-art LieRE by " + }, + { + "bbox": [ + 55, + 134, + 294, + 230 + ], + "type": "inline_equation", + "content": "1.6\\%" + }, + { + "bbox": [ + 55, + 134, + 294, + 230 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 232, + 294, + 328 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 232, + 294, + 328 + ], + "spans": [ + { + "bbox": [ + 55, + 232, + 294, + 328 + ], + "type": "text", + "content": "Scaling to higher resolution. At resolutions beyond the training size, LieRE shows the steepest decline in accuracy among the three trainable RoPE variants, indicating greater sensitivity to resolution changes. In contrast, ComRoPE-LD and ComRoPE-AP exhibit a more gradual decrease in performance, thanks to their commutative properties that enhance positional robustness. Specifically, ComRoPE-LD outperforms LieRE by " + }, + { + "bbox": [ + 55, + 232, + 294, + 328 + ], + "type": "inline_equation", + "content": "2.9\\%" + }, + { + "bbox": [ + 55, + 232, + 294, + 328 + ], + "type": "text", + "content": " at a resolution of " + }, + { + "bbox": [ + 55, + 232, + 294, + 328 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 55, + 232, + 294, + 328 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 331, + 294, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 331, + 294, + 437 + ], + "spans": [ + { + "bbox": [ + 55, + 331, + 294, + 437 + ], + "type": "text", + "content": "Significance of commutativity. These findings illustrate the effectiveness of trainable commutative angle matrices, particularly ComRoPE-LD, in maintaining accuracy and scalability across diverse resolutions. The results underscore the importance of commutativity in ensuring robust RoPE parameterized by trainable angle matrices for vision tasks. Furthermore, to better understand the role of commutativity, we conduct additional experiments by introducing coordinate offsets in LieRE (see Section 4.3.1 for details)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 453, + 155, + 464 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 453, + 155, + 464 + ], + "spans": [ + { + "bbox": [ + 55, + 453, + 155, + 464 + ], + "type": "text", + "content": "4.2. Object detection" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 473, + 294, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 473, + 294, + 533 + ], + "spans": [ + { + "bbox": [ + 55, + 473, + 294, + 533 + ], + "type": "text", + "content": "To demonstrate the generalizability of our approach, we conduct object detection experiments using the framework from [32]. We adopt ViT-S as our backbone and apply ComRoPE to the attention layers. To ensure consistency with the pre-trained model, we initialize the angle matrix to zero." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 535, + 294, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 535, + 294, + 606 + ], + "spans": [ + { + "bbox": [ + 55, + 535, + 294, + 606 + ], + "type": "text", + "content": "We evaluate ComRoPE-LD, LieRE, and APE on the MS COCO dataset [14]. As summarized in Table 2, both ComRoPE-LD and LieRE outperform APE, with ComRoPE-LD achieving slightly better performance than LieRE while only using nearly half the number of extra parameters." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 609, + 294, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 609, + 294, + 657 + ], + "spans": [ + { + "bbox": [ + 55, + 609, + 294, + 657 + ], + "type": "text", + "content": "To compare training efficiency, we plot the results for each epoch in Figure 4. Our findings indicate that both ComRoPE-LD and LieRE converge faster than APE, requiring " + }, + { + "bbox": [ + 55, + 609, + 294, + 657 + ], + "type": "inline_equation", + "content": "29\\%" + }, + { + "bbox": [ + 55, + 609, + 294, + 657 + ], + "type": "text", + "content": " fewer iterations to achieve the same results as APE." + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 318, + 70, + 550, + 125 + ], + "blocks": [ + { + "bbox": [ + 318, + 70, + 550, + 125 + ], + "lines": [ + { + "bbox": [ + 318, + 70, + 550, + 125 + ], + "spans": [ + { + "bbox": [ + 318, + 70, + 550, + 125 + ], + "type": "table", + "html": "
PE MethodAPAP50AP75APsAPmAPl
APE44.066.647.728.246.858.4
LieRE44.567.348.429.046.958.7
ComRoPE-LD44.767.648.529.247.160.0
", + "image_path": "07d37257ac979a582a3ee1b3db2aa587ffea0436a53511bbb3db0080040f7838.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 341, + 127, + 526, + 137 + ], + "lines": [ + { + "bbox": [ + 341, + 127, + 526, + 137 + ], + "spans": [ + { + "bbox": [ + 341, + 127, + 526, + 137 + ], + "type": "text", + "content": "Table 2. Results of object detection on MS COCO." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 320, + 150, + 531, + 242 + ], + "blocks": [ + { + "bbox": [ + 320, + 150, + 531, + 242 + ], + "lines": [ + { + "bbox": [ + 320, + 150, + 531, + 242 + ], + "spans": [ + { + "bbox": [ + 320, + 150, + 531, + 242 + ], + "type": "image", + "image_path": "2a7590cffbb7ff400373f87e79574f869a6c3e1cdf531766c8f4a02fb629294c.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 338, + 245, + 529, + 256 + ], + "lines": [ + { + "bbox": [ + 338, + 245, + 529, + 256 + ], + "spans": [ + { + "bbox": [ + 338, + 245, + 529, + 256 + ], + "type": "text", + "content": "Figure 4. Results over the whole training procedure." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 266, + 406, + 279 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 266, + 406, + 279 + ], + "spans": [ + { + "bbox": [ + 313, + 266, + 406, + 279 + ], + "type": "text", + "content": "4.3. Ablation study" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 285, + 453, + 297 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 285, + 453, + 297 + ], + "spans": [ + { + "bbox": [ + 313, + 285, + 453, + 297 + ], + "type": "text", + "content": "4.3.1 Impact of commutativity" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 303, + 553, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 303, + 553, + 373 + ], + "spans": [ + { + "bbox": [ + 313, + 303, + 553, + 373 + ], + "type": "text", + "content": "To evaluate the significance of the commutativity of angle matrices, we conduct experiments on the LieRE by introducing a coordinate offset. Specifically, before inference, a random offset is applied uniformly across all coordinates within the image. The offset is sampled from a Gaussian distribution as follows:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 394, + 383, + 553, + 398 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 394, + 383, + 553, + 398 + ], + "spans": [ + { + "bbox": [ + 394, + 383, + 553, + 398 + ], + "type": "interline_equation", + "content": "\\mathcal {N} (0; \\rho^ {2} \\cdot \\mathbf {I} _ {N \\times N}) \\tag {12}", + "image_path": "01ace9010723fe5cf5e15e96c8f2dffff9f9fe6bdb4a02bb1236660c0ddca947.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 405, + 553, + 476 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 405, + 553, + 476 + ], + "spans": [ + { + "bbox": [ + 313, + 405, + 553, + 476 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 405, + 553, + 476 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 313, + 405, + 553, + 476 + ], + "type": "text", + "content": " represents the standard deviation of the random offset, and " + }, + { + "bbox": [ + 313, + 405, + 553, + 476 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{N\\times N}" + }, + { + "bbox": [ + 313, + 405, + 553, + 476 + ], + "type": "text", + "content": " is the identity matrix of size " + }, + { + "bbox": [ + 313, + 405, + 553, + 476 + ], + "type": "inline_equation", + "content": "N\\times N" + }, + { + "bbox": [ + 313, + 405, + 553, + 476 + ], + "type": "text", + "content": ". It is important to note that applying the same offset to all coordinates does not influence the relative positional dependencies between the patches. The other experimental settings remain consistent with those in the 2D classification tasks." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 477, + 554, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 477, + 554, + 597 + ], + "spans": [ + { + "bbox": [ + 313, + 477, + 554, + 597 + ], + "type": "text", + "content": "The results shown in Figure 5 demonstrate that the baseline model's performance deteriorates significantly as the standard deviation of the offset increases. In contrast, our proposed model, ComRoPE, maintains consistent performance across all offset values. This is due to the commutativity of the angle matrices, which allows the model to remain invariant to such coordinate shifts. The robustness of ComRoPE to this type of perturbation highlights its capacity to preserve relative positional information, even in the presence of modification introduced by coordinate offsets." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 611, + 432, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 611, + 432, + 623 + ], + "spans": [ + { + "bbox": [ + 313, + 611, + 432, + 623 + ], + "type": "text", + "content": "4.3.2 Impact of block size" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 629, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 554, + 713 + ], + "type": "text", + "content": "In this section, we examine the impact of block size on 2D classification using the ImageNet dataset. We maintain the same experimental setup as in our primary experiments, varying the block size from 2 to 8. The results are presented in Figure 6. Our findings indicate that larger block sizes consistently improve performance by extending the rotation transformation space to a more significant subgroup of the" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 55, + 673, + 295, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 673, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 673, + 295, + 713 + ], + "type": "text", + "content": "2Relative improvement is calculated as target " + }, + { + "bbox": [ + 55, + 673, + 295, + 713 + ], + "type": "inline_equation", + "content": "=" + }, + { + "bbox": [ + 55, + 673, + 295, + 713 + ], + "type": "text", + "content": " baseline " + }, + { + "bbox": [ + 55, + 673, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\times" + }, + { + "bbox": [ + 55, + 673, + 295, + 713 + ], + "type": "text", + "content": " (1 + improvement), where the improvement denotes the percentage increase over the baseline performance. This formula applies throughout the following sections." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4514" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 76, + 290, + 167 + ], + "blocks": [ + { + "bbox": [ + 61, + 76, + 290, + 167 + ], + "lines": [ + { + "bbox": [ + 61, + 76, + 290, + 167 + ], + "spans": [ + { + "bbox": [ + 61, + 76, + 290, + 167 + ], + "type": "image", + "image_path": "c74fed46cd6519d96090420608f1d5304df3ea1b11b005f7cde938dbb1603293.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 174, + 295, + 218 + ], + "lines": [ + { + "bbox": [ + 55, + 174, + 295, + 218 + ], + "spans": [ + { + "bbox": [ + 55, + 174, + 295, + 218 + ], + "type": "text", + "content": "Figure 5. Effect of coordinate offset on LieRE. As the standard deviation of the offset increases, the performance of the baseline model deteriorates, while ComRoPE remains unaffected (un-painted)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 59, + 227, + 291, + 275 + ], + "blocks": [ + { + "bbox": [ + 59, + 227, + 291, + 275 + ], + "lines": [ + { + "bbox": [ + 59, + 227, + 291, + 275 + ], + "spans": [ + { + "bbox": [ + 59, + 227, + 291, + 275 + ], + "type": "image", + "image_path": "61d78426ebfcb71393d359a0855521f0a4562974f5b426f6b6cd0f3f38efac08.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 159, + 277, + 192, + 285 + ], + "lines": [ + { + "bbox": [ + 159, + 277, + 192, + 285 + ], + "spans": [ + { + "bbox": [ + 159, + 277, + 192, + 285 + ], + "type": "text", + "content": "(a) LieRE" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 59, + 293, + 291, + 340 + ], + "blocks": [ + { + "bbox": [ + 59, + 293, + 291, + 340 + ], + "lines": [ + { + "bbox": [ + 59, + 293, + 291, + 340 + ], + "spans": [ + { + "bbox": [ + 59, + 293, + 291, + 340 + ], + "type": "image", + "image_path": "8efe65594a14fec2220edd8e5275cbe83f78697ec31e930a67846059716ef106.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 146, + 342, + 206, + 350 + ], + "lines": [ + { + "bbox": [ + 146, + 342, + 206, + 350 + ], + "spans": [ + { + "bbox": [ + 146, + 342, + 206, + 350 + ], + "type": "text", + "content": "(b) ComRoPE-AP" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 59, + 357, + 291, + 405 + ], + "blocks": [ + { + "bbox": [ + 59, + 357, + 291, + 405 + ], + "lines": [ + { + "bbox": [ + 59, + 357, + 291, + 405 + ], + "spans": [ + { + "bbox": [ + 59, + 357, + 291, + 405 + ], + "type": "image", + "image_path": "5c5df1dcd0c3dea495c46bef9c13c99829a7e4102f564457c9e6d7209d24d469.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 146, + 406, + 206, + 415 + ], + "lines": [ + { + "bbox": [ + 146, + 406, + 206, + 415 + ], + "spans": [ + { + "bbox": [ + 146, + 406, + 206, + 415 + ], + "type": "text", + "content": "(c) ComRoPE-LD" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 417, + 295, + 440 + ], + "lines": [ + { + "bbox": [ + 55, + 417, + 295, + 440 + ], + "spans": [ + { + "bbox": [ + 55, + 417, + 295, + 440 + ], + "type": "text", + "content": "Figure 6. Accuracy on ImageNet for various block sizes. Larger block size results in better performance." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 451, + 295, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 451, + 295, + 559 + ], + "spans": [ + { + "bbox": [ + 55, + 451, + 295, + 559 + ], + "type": "text", + "content": "particular orthogonal group by introducing additional parameters and computation time. When the block size is small, the associated costs are minimal. However, as the block size increases, the primary term of time complexity grows to " + }, + { + "bbox": [ + 55, + 451, + 295, + 559 + ], + "type": "inline_equation", + "content": "O(Lndb^{2})" + }, + { + "bbox": [ + 55, + 451, + 295, + 559 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 55, + 451, + 295, + 559 + ], + "type": "inline_equation", + "content": "O(Lnd\\cdot \\frac{d}{h})" + }, + { + "bbox": [ + 55, + 451, + 295, + 559 + ], + "type": "text", + "content": ", which becomes significant. Therefore, we limit the block size to a maximum of 8 to balance performance with additional costs. In other words, we select a block size of 8 to optimize performance while keeping the extra computational cost manageable." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 574, + 220, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 574, + 220, + 586 + ], + "spans": [ + { + "bbox": [ + 55, + 574, + 220, + 586 + ], + "type": "text", + "content": "4.3.3 Utility of position perturbation" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 594, + 295, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 641 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 641 + ], + "type": "text", + "content": "In this section, we explore the impact of positional perturbations. We conducted experiments on ComRoPE-LD and APE using the ImageNet dataset, with the results presented in Figure 7." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "type": "text", + "content": "As shown in Figure 7a, APE is highly sensitive to positional perturbations, leading to significant performance improvement " + }, + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "type": "inline_equation", + "content": "(+19.5\\%)" + }, + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "type": "text", + "content": " when increasing intensity from 0 to 1) when these perturbations are introduced. For RoPE with angle matrices shown in Figure 7b, positional perturbations also resulted in some performance gains " + }, + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "type": "inline_equation", + "content": "(+2.9\\%)" + }, + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "type": "text", + "content": ", though" + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 320, + 75, + 548, + 134 + ], + "blocks": [ + { + "bbox": [ + 320, + 75, + 548, + 134 + ], + "lines": [ + { + "bbox": [ + 320, + 75, + 548, + 134 + ], + "spans": [ + { + "bbox": [ + 320, + 75, + 548, + 134 + ], + "type": "image", + "image_path": "8d171c91a600ba24a9c12ce64527510a50312b8948113ffd0f1c1b5385daca9b.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 421, + 136, + 449, + 145 + ], + "lines": [ + { + "bbox": [ + 421, + 136, + 449, + 145 + ], + "spans": [ + { + "bbox": [ + 421, + 136, + 449, + 145 + ], + "type": "text", + "content": "(a) APE" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 319, + 153, + 548, + 213 + ], + "blocks": [ + { + "bbox": [ + 319, + 153, + 548, + 213 + ], + "lines": [ + { + "bbox": [ + 319, + 153, + 548, + 213 + ], + "spans": [ + { + "bbox": [ + 319, + 153, + 548, + 213 + ], + "type": "image", + "image_path": "3380304f09d22faee65a8eb2bfb9c163ad7a0ec5e552cb5210a31e9cd3177d53.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 405, + 215, + 465, + 224 + ], + "lines": [ + { + "bbox": [ + 405, + 215, + 465, + 224 + ], + "spans": [ + { + "bbox": [ + 405, + 215, + 465, + 224 + ], + "type": "text", + "content": "(b) ComRoPE-LD" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 227, + 553, + 249 + ], + "lines": [ + { + "bbox": [ + 313, + 227, + 553, + 249 + ], + "spans": [ + { + "bbox": [ + 313, + 227, + 553, + 249 + ], + "type": "text", + "content": "Figure 7. Accuracy on ImageNet over varying position perturbation intensity." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 260, + 553, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 260, + 553, + 308 + ], + "spans": [ + { + "bbox": [ + 313, + 260, + 553, + 308 + ], + "type": "text", + "content": "the improvement was relatively modest. This is likely due to the inherent robustness of the RoPE design with angle matrices, which is already well-equipped to handle variations in position." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 316, + 395, + 328 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 316, + 395, + 328 + ], + "spans": [ + { + "bbox": [ + 313, + 316, + 395, + 328 + ], + "type": "text", + "content": "4.4. Applications" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 334, + 555, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 334, + 555, + 502 + ], + "spans": [ + { + "bbox": [ + 313, + 334, + 555, + 502 + ], + "type": "text", + "content": "In our approach, when the angle matrix is an all-zero matrix, the rotation matrix becomes the identity matrix, causing RoPE Attention to reduce to the standard Attention mechanism. When the block size of the angle matrix is set to 2, ComRoPE-AP effectively reduces to the commonly used RoPE Attention in language models. This demonstrates that our method can represent standard Attention and various common variants of RoPE Attention. Therefore, during the fine-tuning stage, we can replace standard Attention with our method, load the pre-trained weights, and fine-tune them under the new paradigm. In other words, ComRoPE can be seamlessly integrated into the fine-tuning process, even if it was not applied during pre-training. Additional experiments can be found in Appendix B." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 512, + 388, + 524 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 512, + 388, + 524 + ], + "spans": [ + { + "bbox": [ + 313, + 512, + 388, + 524 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 532, + 555, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 532, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 532, + 555, + 712 + ], + "type": "text", + "content": "In this work, we proposed ComRoPE, a novel and adaptive framework for Rotary Position Embedding (RoPE) parameterized by trainable angle matrices. We rigorously formulate the RoPE Equation and then establish a necessary and sufficient condition for its solution. Our approach effectively overcomes the scalability and robustness limitations of existing RoPE methods by eliminating the need for manually designed rotation matrices and introducing a more flexible, scalable solution. Extensive experimental results show that ComRoPE outperforms the existing positional encoding methods across various tasks. Furthermore, our framework generalizes existing RoPE formulations and demonstrates the potential for broader application in Transformer models, offering insights and a solid foundation for future research in positional encoding techniques." + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4515" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "type": "text", + "content": "Acknowledgment" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 294, + 151 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 294, + 151 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 294, + 151 + ], + "type": "text", + "content": "This work is supported by the National Key R&D Program of China (2022YFB4701400/4701402), SSTIC Grant (KJZD20230923115106012, KJZD20230923114916032, GJHZ20240218113604008) and Beijing Key Lab of Networked Multimedia." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 114, + 175 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 114, + 175 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 114, + 175 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 183, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 183, + 294, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 183, + 294, + 205 + ], + "spans": [ + { + "bbox": [ + 61, + 183, + 294, + 205 + ], + "type": "text", + "content": "[1] Tom B Brown. Language models are few-shot learners. arXiv preprint arXiv:2005.14165, 2020. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 207, + 295, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 207, + 295, + 261 + ], + "spans": [ + { + "bbox": [ + 62, + 207, + 295, + 261 + ], + "type": "text", + "content": "[2] Yirui Chen, Xudong Huang, Quan Zhang, Wei Li, Mingjian Zhu, Qiangyu Yan, Simiao Li, Hanting Chen, Hailin Hu, Jie Yang, et al. Gim: A million-scale benchmark for generative image manipulation detection and localization. arXiv preprint arXiv:2406.16531, 2024. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 262, + 295, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 262, + 295, + 327 + ], + "spans": [ + { + "bbox": [ + 62, + 262, + 295, + 327 + ], + "type": "text", + "content": "[3] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhanghao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E Gonzalez, et al. Vicuna: An open-source chatbot impressing gpt-4 with " + }, + { + "bbox": [ + 62, + 262, + 295, + 327 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 62, + 262, + 295, + 327 + ], + "type": "text", + "content": " * chatgpt quality, march 2023. URL https://lmsys.org/blog/2023-03-30-vicuna, 3(5), 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 330, + 295, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 330, + 295, + 373 + ], + "spans": [ + { + "bbox": [ + 62, + 330, + 295, + 373 + ], + "type": "text", + "content": "[4] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, K. Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 248-255, 2009. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 376, + 294, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 376, + 294, + 430 + ], + "spans": [ + { + "bbox": [ + 62, + 376, + 294, + 430 + ], + "type": "text", + "content": "[5] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In North American Chapter of the Association for Computational Linguistics, 2019. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 432, + 294, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 432, + 294, + 463 + ], + "spans": [ + { + "bbox": [ + 62, + 432, + 294, + 463 + ], + "type": "text", + "content": "[6] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 466, + 294, + 507 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 466, + 294, + 507 + ], + "spans": [ + { + "bbox": [ + 62, + 466, + 294, + 507 + ], + "type": "text", + "content": "[7] Yuxin Fang, Quan Sun, Xinggang Wang, Tiejun Huang, Xinlong Wang, and Yue Cao. Eva-02: A visual representation for neon genesis. Image and Vision Computing, 149:105171, 2024. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 511, + 294, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 511, + 294, + 564 + ], + "spans": [ + { + "bbox": [ + 62, + 511, + 294, + 564 + ], + "type": "text", + "content": "[8] Team GLM, Aohan Zeng, Bin Xu, Bowen Wang, Chenhui Zhang, Da Yin, Dan Zhang, Diego Rojas, Guanyu Feng, Hanlin Zhao, et al. Chatglm: A family of large language models from glm-130b to glm-4 all tools. arXiv preprint arXiv:2406.12793, 2024. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 567, + 294, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 567, + 294, + 599 + ], + "spans": [ + { + "bbox": [ + 62, + 567, + 294, + 599 + ], + "type": "text", + "content": "[9] Byeongho Heo, Song Park, Dongyoon Han, and Sangwoo Yun. Rotary position embedding for vision transformer. arXiv preprint arXiv:2403.13298, 2024. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 601, + 294, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 601, + 294, + 634 + ], + "spans": [ + { + "bbox": [ + 57, + 601, + 294, + 634 + ], + "type": "text", + "content": "[10] Max Horn, Kumar Shridhar, Elrich Groenewald, and Philipp FM Baumann. Translational equivariance in kernelizable attention. arXiv preprint arXiv:2102.07680, 2021. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 636, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 636, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 57, + 636, + 294, + 689 + ], + "type": "text", + "content": "[11] Amirhossein Kazemnejad, Inkit Padhi, Karthikeyan Natesan Ramamurthy, Payel Das, and Siva Reddy. The impact of positional encoding on length generalization in transformers. Advances in Neural Information Processing Systems, 36: 24892-24928, 2023. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 692, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 692, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 692, + 294, + 713 + ], + "type": "text", + "content": "[12] Hanyu Lai, Xiao Liu, Iat Long Iong, Shuntian Yao, Yuxuan Chen, Pengbo Shen, Hao Yu, Hanchen Zhang, Xiaohan" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 317, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 117 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 117 + ], + "type": "text", + "content": "Zhang, Yuxiao Dong, et al. Autowebglm: A large language model-based web navigating agent. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 5295-5306, 2024. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 118, + 553, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 118, + 553, + 171 + ], + "spans": [ + { + "bbox": [ + 317, + 118, + 553, + 171 + ], + "type": "text", + "content": "[13] Hongyu Li, Jinyu Chen, Ziyu Wei, Shaofei Huang, Tianrui Hui, Jialin Gao, Xiaoming Wei, and Si Liu. Llava-st: A multimodal large language model for fine-grained spatial-temporal understanding. arXiv preprint arXiv:2501.08282, 2025. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 173, + 553, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 173, + 553, + 226 + ], + "spans": [ + { + "bbox": [ + 317, + 173, + 553, + 226 + ], + "type": "text", + "content": "[14] Tsung-Yi Lin, Michael Maire, Serge Belongie, Lubomir Bourdev, Ross Girshick, James Hays, Pietro Perona, Deva Ramanan, C. Lawrence Zitnick, and Piotr Dollar. Microsoft coco: Common objects in context. arXiv preprint arXiv: 1405.0312, 2014. 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 228, + 553, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 228, + 553, + 293 + ], + "spans": [ + { + "bbox": [ + 317, + 228, + 553, + 293 + ], + "type": "text", + "content": "[15] Xiao Liu, Hanyu Lai, Hao Yu, Yifan Xu, Aohan Zeng, Zhengxiao Du, Peng Zhang, Yuxiao Dong, and Jie Tang. Webglm: towards an efficient web-enhanced question answering system with human preferences. In Proceedings of the 29th ACM SIGKDD conference on knowledge discovery and data mining, pages 4549-4560, 2023. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 294, + 553, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 294, + 553, + 347 + ], + "spans": [ + { + "bbox": [ + 317, + 294, + 553, + 347 + ], + "type": "text", + "content": "[16] Xiao Liu, Hao Yu, Hanchen Zhang, Yifan Xu, Xuanyu Lei, Hanyu Lai, Yu Gu, Hangliang Ding, Kaiwen Men, Kejuan Yang, et al. Agentbench: Evaluating llms as agents. In The Twelfth International Conference on Learning Representations, 2024. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 350, + 553, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 350, + 553, + 403 + ], + "spans": [ + { + "bbox": [ + 317, + 350, + 553, + 403 + ], + "type": "text", + "content": "[17] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 405, + 553, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 405, + 553, + 458 + ], + "spans": [ + { + "bbox": [ + 317, + 405, + 553, + 458 + ], + "type": "text", + "content": "[18] Antoine Liutkus, Ondrej Cifka, Shih-Lun Wu, Umut Simsekli, Yi-Hsuan Yang, and Gael Richard. Relative positional encoding for transformers with linear complexity. In International Conference on Machine Learning, pages 7067-7079. PMLR, 2021. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 460, + 553, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 460, + 553, + 525 + ], + "spans": [ + { + "bbox": [ + 317, + 460, + 553, + 525 + ], + "type": "text", + "content": "[19] Jiasen Lu, Christopher Clark, Sangho Lee, Zichen Zhang, Savya Khosla, Ryan Marten, Derek Hoiem, and Aniruddha Kembhavi. Unified-io 2: Scaling autoregressive multimodal models with vision language audio and action. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26439–26455, 2024. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 526, + 553, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 526, + 553, + 568 + ], + "spans": [ + { + "bbox": [ + 317, + 526, + 553, + 568 + ], + "type": "text", + "content": "[20] Sophie Ostmeier, Brian Axelrod, Michael E Moseley, Akshay Chaudhari, and Curtis Langlotz. Liere: Generalizing rotary position encodings. arXiv preprint arXiv:2406.10322, 2024. 2, 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 571, + 553, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 571, + 553, + 602 + ], + "spans": [ + { + "bbox": [ + 317, + 571, + 553, + 602 + ], + "type": "text", + "content": "[21] Ofir Press, Noah A Smith, and Mike Lewis. Train short, test long: Attention with linear biases enables input length extrapolation. arXiv preprint arXiv:2108.12409, 2021. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 604, + 553, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 604, + 553, + 624 + ], + "spans": [ + { + "bbox": [ + 317, + 604, + 553, + 624 + ], + "type": "text", + "content": "[22] A Radford. Improving language understanding by generative pre-training. 2018. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 625, + 553, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 625, + 553, + 658 + ], + "spans": [ + { + "bbox": [ + 317, + 625, + 553, + 658 + ], + "type": "text", + "content": "[23] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 659, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 659, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 317, + 659, + 553, + 712 + ], + "type": "text", + "content": "[24] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of machine learning research, 21(140):1-67, 2020. 2" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4516" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 632 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 127 + ], + "type": "text", + "content": "[25] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of machine learning research, 21(140):1-67, 2020. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 294, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 294, + 163 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 294, + 163 + ], + "type": "text", + "content": "[26] Peter Shaw, Jakob Uszkoreit, and Ashish Vaswani. Self-attention with relative position representations. arXiv preprint arXiv:1803.02155, 2018. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 164, + 294, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 164, + 294, + 206 + ], + "spans": [ + { + "bbox": [ + 56, + 164, + 294, + 206 + ], + "type": "text", + "content": "[27] Jianlin Su, Murtadha Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu. Roformer: Enhanced transformer with rotary position embedding. Neurocomputing, 568:127063, 2024. 1, 2, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 208, + 294, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 208, + 294, + 240 + ], + "spans": [ + { + "bbox": [ + 56, + 208, + 294, + 240 + ], + "type": "text", + "content": "[28] Sainbayar Sukhbaatar, Jason Weston, Rob Fergus, et al. End-to-end memory networks. Advances in neural information processing systems, 28, 2015. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 241, + 294, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 241, + 294, + 297 + ], + "spans": [ + { + "bbox": [ + 56, + 241, + 294, + 297 + ], + "type": "text", + "content": "[29] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 297, + 294, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 297, + 294, + 319 + ], + "spans": [ + { + "bbox": [ + 56, + 297, + 294, + 319 + ], + "type": "text", + "content": "[30] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 1, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 320, + 294, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 320, + 294, + 374 + ], + "spans": [ + { + "bbox": [ + 56, + 320, + 294, + 374 + ], + "type": "text", + "content": "[31] Kan Wu, Houwen Peng, Minghao Chen, Jianlong Fu, and Hongyang Chao. Rethinking and improving relative position encoding for vision transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10033-10041, 2021. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 376, + 294, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 376, + 294, + 421 + ], + "spans": [ + { + "bbox": [ + 56, + 376, + 294, + 421 + ], + "type": "text", + "content": "[32] Chunlong Xia, Xinliang Wang, Feng Lv, Xin Hao, and Yifeng Shi. Vit-comer: Vision transformer with convolutional multi-scale feature interaction for dense predictions. Computer Vision and Pattern Recognition, 2024. 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 421, + 294, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 421, + 294, + 453 + ], + "spans": [ + { + "bbox": [ + 56, + 421, + 294, + 453 + ], + "type": "text", + "content": "[33] Quan Zhang and Yuxin Qi. Can mllms guide weakly-supervised temporal action localization tasks? arXiv preprint arXiv:2411.08466, 2024. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 455, + 294, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 455, + 294, + 520 + ], + "spans": [ + { + "bbox": [ + 56, + 455, + 294, + 520 + ], + "type": "text", + "content": "[34] Quan Zhang, Xiaoyu Liu, Wei Li, Hanting Chen, Junchao Liu, Jie Hu, Zhiwei Xiong, Chun Yuan, and Yunhe Wang. Distilling semantic priors from sam to efficient image restoration models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25409-25419, 2024." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 521, + 294, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 521, + 294, + 577 + ], + "spans": [ + { + "bbox": [ + 56, + 521, + 294, + 577 + ], + "type": "text", + "content": "[35] Quan Zhang, Yuxin Qi, Xi Tang, Jinwei Fang, Xi Lin, Ke Zhang, and Chun Yuan. IMDPrompter: Adapting SAM to image manipulation detection by cross-view automated prompt learning. In The Thirteenth International Conference on Learning Representations, 2025." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 577, + 294, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 577, + 294, + 632 + ], + "spans": [ + { + "bbox": [ + 56, + 577, + 294, + 632 + ], + "type": "text", + "content": "[36] Quan Zhang, Yuxin Qi, Xi Tang, Rui Yuan, Xi Lin, Ke Zhang, and Chun Yuan. Rethinking pseudo-label guided learning for weakly supervised temporal action localization from the perspective of noise correction. arXiv preprint arXiv:2501.11124, 2025. 1" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4517" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/ComfyBench_ Benchmarking LLM-based Agents in ComfyUI for Autonomously Designing Collaborative AI Systems/9eeb1c75-c4c6-45dd-9896-e14c13fb270d_content_list.json b/2025/ComfyBench_ Benchmarking LLM-based Agents in ComfyUI for Autonomously Designing Collaborative AI Systems/9eeb1c75-c4c6-45dd-9896-e14c13fb270d_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..6da664f0103bffca1f832b39fb507f1b81fea1e0 --- /dev/null +++ b/2025/ComfyBench_ Benchmarking LLM-based Agents in ComfyUI for Autonomously Designing Collaborative AI Systems/9eeb1c75-c4c6-45dd-9896-e14c13fb270d_content_list.json @@ -0,0 +1,1669 @@ +[ + { + "type": "text", + "text": "ComfyBench: Benchmarking LLM-based Agents in ComfyUI for Autonomous Designing Collaborative AI Systems", + "text_level": 1, + "bbox": [ + 94, + 128, + 901, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xiangyuan Xue $^{1,2*}$ Zeyu Lu $^{1,2}$ Di Huang $^{3}$ Zidong Wang $^{1,4}$ Wanli Ouyang $^{1,4}$ Lei Bai $^{1\\boxtimes}$ $^{1}$ Shanghai Artificial Intelligence Laboratory \n $^{2}$ Shanghai Jiao Tong University \n $^{3}$ The University of Sydney \n $^{4}$ The Chinese University of Hong Kong", + "bbox": [ + 107, + 200, + 888, + 258 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 290, + 326, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Much previous AI research has focused on developing monolithic models to maximize their intelligence, with the primary goal of enhancing performance on specific tasks. In contrast, this work attempts to study using LLM-based agents to design collaborative AI systems autonomously. To explore this problem, we first introduce ComfyBench to evaluate agents' ability to design collaborative AI systems in ComfyUI. ComfyBench is a comprehensive benchmark comprising 200 diverse tasks covering various instruction-following generation challenges, along with detailed annotations for 3,205 nodes and 20 workflows. Based on ComfyBench, we further develop ComfyAgent, a novel framework that empowers LLM-based agents to autonomously design collaborative AI systems by generating workflows. ComfyAgent is based on two core concepts. First, it represents workflows with code, which can be reversibly converted into workflows and executed as collaborative systems by the interpreter. Second, it constructs a multi-agent system that cooperates to learn from existing workflows and generate new workflows for a given task. While experimental results demonstrate that ComfyAgent achieves a comparable resolve rate to o1-preview and significantly surpasses other agents on ComfyBench, ComfyAgent has resolved only $15\\%$ of creative tasks. LLM-based agents still have a long way to go in autonomously designing collaborative AI systems. Progress with ComfyBench is paving the way for more intelligent and autonomous collaborative AI systems. Our code is available at: https://github.com/xyxQwQ/ComfyBench.", + "bbox": [ + 88, + 323, + 485, + 748 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 776, + 220, + 792 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The recent evolution of AI is defined by the growing importance of collaborative AI systems, which integrate multiple models and tools to work as a whole collaborative system [74]. The success of ChatGPT [37] presents the pos", + "bbox": [ + 89, + 801, + 483, + 864 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "sibility of integrating a wide range of tasks, such as web browsing [35], image generation [43], and code execution, into a single conversational agent. Unlike traditional AI models that function as single entities, collaborative AI systems integrate multiple AI components, each contributing unique capabilities to solve complex problems. This shift towards integration has become crucial for achieving state-of-the-art results, as it leverages the combined strengths of diverse AI functionalities within a unified framework.", + "bbox": [ + 511, + 292, + 906, + 428 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Unfortunately, the design space and optimization of collaborative systems often require significant human expertise [31]. For instance, AlphaCode 2 [2] utilizes an intricate AI system to achieve expert-level performance in competitive programming, which involves fine-tuned LLMs for coding and scoring, expert models for clustering, and modules for code execution and filtering. These components collaborate in a precise manner carefully designed by human experts to achieve state-of-the-art performance, which is not achievable by any single component or their casual combination. This brings us to a pivotal question: Can we develop an agent, akin to a human expert, capable of autonomously designing collaborative AI systems?", + "bbox": [ + 511, + 429, + 908, + 626 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this work, we explore this question in ComfyUI [13], a trending and open-source platform that supports various generative tools and models in terms of nodes. Within ComfyUI, users can flexibly construct and connect nodes to build the workflow of the collaborative AI system, which is capable of producing highly customized artwork.", + "bbox": [ + 511, + 627, + 908, + 718 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To evaluate agents's ability to design collaborative AI systems in ComfyUI, we introduce ComfyBench, a comprehensive benchmark shown in Figure 1, which comprises 200 diverse tasks covering various instruction-following generation challenges, along with detailed annotations for 3,205 nodes and 20 workflows. We categorize tasks into three levels of difficulty: vanilla, complex, and creative. In each task, agents are required to create workflows that describe collaborative AI systems within ComfyUI, which can then be executed to produce the desired outcomes specified in the task instructions. We assess the performance of agents using two evaluation metrics: pass rate, showing the", + "bbox": [ + 511, + 719, + 910, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*This work was done during his internship at Shanghai Artificial Intelligence Laboratory. Corresponding author, baisanshi@gmail.com.", + "bbox": [ + 89, + 875, + 482, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "24614", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/306174134a52a2729f8f0f05a8100caea5e4cd467a0c3d888cebccad81ba374a.jpg", + "image_caption": [ + "(a) ComfyBench" + ], + "image_footnote": [], + "bbox": [ + 94, + 106, + 903, + 253 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/164f6d4b2a57e631b2ce3976504a198483319364e23deda3dbe4ff97aa8b596f.jpg", + "image_caption": [ + "(b) ComfyAgent", + "Figure 1. (a) ComfyBench is a comprehensive benchmark to evaluate agents's ability to design collaborative AI systems in ComfyUI. Given the task instruction, agents are required to learn from documents and create workflows to describe collaborative AI systems. The performance is measured by pass rate and resolve rate, reflecting whether the workflow can be correctly executed and whether the task requirements are realized. (b) ComfyAgent builds collaborative AI systems in ComfyUI by generating workflows. The workflows are converted into equivalent code so that LLMs can better understand them. ComfyAgent can learn from existing workflows and autonomously design new ones. The generated workflows can be interpreted as collaborative AI systems to complete given tasks." + ], + "image_footnote": [], + "bbox": [ + 93, + 275, + 906, + 359 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "success of workflow execution, and resovle rate, reflecting the effectiveness in meeting task requirements.", + "bbox": [ + 89, + 482, + 482, + 512 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Based on ComfyBench, we find such tasks present great challenges for agents. First, the native JSON representation of workflows is inadequate for LLM-based agents to fully comprehend the intrinsic logic and dependencies within the systems [65]. Second, properly organizing the pipeline and utilizing AI-related modules require substantial expert knowledge and complex cognitive processes [21]. Traditional methods, such as Chain-of-Thought (CoT) [57] or Retrieval-Augmented Generation (RAG) [29], suffer from context limitation and hallucination problems, thus failing to generate complex workflows reliably.", + "bbox": [ + 88, + 517, + 482, + 684 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To tackle the aforementioned problems, we introduce ComfyAgent, a novel framework shown in Figure 1, which empowers LLM-based agents to autonomously design collaborative AI systems by generating workflows. The core concepts behind ComfyAgent are two-fold. First, it represents workflows with code, which can be reversibly converted into workflows and executed as collaborative systems by the interpreter. Second, it constructs a multi-agent system [19] which cooperates to learn from existing workflows and generate new workflows for a given task. The core of the multi-agent system is PlanAgent, responsible for the global planning of workflows by task instruction. After collecting document descriptions of all components and code examples of collected workflows, RetrievalAgent au", + "bbox": [ + 89, + 689, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tomatically retrieves and learns key information about the given task from code. CombineAgent and AdaptAgent are responsible for code modification to construct a workflow that can fulfill the specified task. Upon task completion, PlanAgent finishes the task to form the final workflow.", + "bbox": [ + 511, + 482, + 906, + 556 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We evaluate common baseline agents [9, 29, 53, 57], as well as our proposed ComfyAgent, on ComfyBench. The experimental results indicate that ComfyAgent achieves comparable pass rate and resolve rate to o1-preview [39] and significantly surpasses other agents on ComfyBench. However, ComfyAgent has resolved only $15\\%$ of creative tasks, which is still far from satisfactory. We believe that LLM-based agents still have a long way to go for autonomously designing collaborative AI systems.", + "bbox": [ + 511, + 559, + 908, + 696 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions can be summarized as follows:", + "bbox": [ + 531, + 699, + 857, + 712 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce ComfyBench, the first-of-its-kind comprehensive benchmark for the development of agents capable of designing and executing a wide range of collaborative AI systems in ComfyUI.", + "- We propose ComfyAgent, which represents workflows with code and constructs a novel multi-agent framework to empower LLM-based agents in autonomously designing collaborative AI systems.", + "- Experimental results demonstrate that with our intricate multi-agent design, ComfyAgent achieves a comparable pass rate and resolve rate to o1-preview and significantly surpasses other agents on ComfyBench." + ], + "bbox": [ + 511, + 715, + 903, + 897 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "24615", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 91, + 89, + 233, + 104 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1. LLM-based Agents", + "text_level": 1, + "bbox": [ + 89, + 114, + 274, + 131 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "LLM-based agents leverage LLMs to interact with external tools and solve real-world problems [52, 60]. Agents can exploit the powerful capabilities of LLMs in cognition and creativity, adapting to various novel tasks without relying on human intervention [30, 34]. Besides, agents have access to a wide range of external tools, making up for the natural limitation of processing only natural language [12, 77]. Researchers have been devoted to enhancing the capabilities of LLM-based agents. Many prompt-based methods [7, 33, 47, 53, 54, 57, 70] are proven effective for improving agents' planning and reasoning. Memory mechanism [78, 79] and RAG [16, 29] further extend agents' working context and knowledge base. Reinforcement learning [6, 23, 41], serving as a post-training paradigm, allowing even stronger enhancement and alignment. These various methods hasten the advent of advanced theories for agents, such as self-evolution [61, 64, 82] and unified architecture [48, 81].", + "bbox": [ + 89, + 138, + 482, + 409 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Based on well-designed module architectures and interaction mechanisms, LLM-based agents can be applied to solve a wide range of complex tasks, such as web navigation [14, 27, 35, 69, 80], interface operation [10, 59, 62, 63, 68], code generation [55, 56, 66, 72], and other tasks in specified domains [17, 22, 46, 50, 67]. These tasks, previously considered to require expert knowledge and human intervention, can now be accomplished by agents reliably.", + "bbox": [ + 89, + 411, + 482, + 532 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Collaborative AI Systems", + "text_level": 1, + "bbox": [ + 89, + 544, + 323, + 559 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Zaharia et al. [74] points out that state-of-the-art results are increasingly obtained by collaborative AI systems instead of monolithic models, which has been verified by many previous works [2, 29, 36, 49, 51]. Yuan et al. [73] even designs a multi-agent framework that incorporates several advanced visual agents to replicate generalist video generation demonstrated by Sora [8]. The advantages of collaborative AI systems are prominent. Dynamic systems make it possible to fetch external knowledge with components such as retrievers. Besides, extra modules such as verifier make the results controllable and trustworthy.", + "bbox": [ + 89, + 566, + 482, + 733 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Compared to monolithic models, collaborative AI systems have posed new challenges in design and optimization. Novel paradigms are emerging to tackle these challenges, of which the most popular is composition frameworks, which build applications out of calls to models and other components [11, 20, 28, 32, 58]. DSPy [24] automatically optimizes the prompt instructions, few-shot examples and other parameters in the pipeline to maximize the end-to-end performance. GPTSwarm [83] supports building agents in the form of graphs and optimizing the parameters and connections as an entire system.", + "bbox": [ + 89, + 734, + 482, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. ComfyBench", + "text_level": 1, + "bbox": [ + 513, + 89, + 648, + 106 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We propose ComfyBench, a new benchmark for evaluating the capability of LLM-based agents to autonomously design collaborative AI systems within the ComfyUI [13] platform. Given the task instruction, ComfyBench requires agents to generate workflows to describe collaborative AI systems in ComfyUI, producing the desired results after execution.", + "bbox": [ + 511, + 114, + 906, + 205 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. ComfyUI Platform", + "text_level": 1, + "bbox": [ + 511, + 213, + 694, + 229 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Overview. ComfyUI is an open-source platform for designing and executing generative pipelines, where users can design their workflows by creating and connecting a series of nodes. Each node stands for a module, which can be a model, a tool, or a widget. On the user side, the workflows are visualized as directed acyclic graphs (DAGs), so that users can operate them intuitively. On the server side, however, the workflows are saved in a JSON object, which describes the nodes and connectivity in a structured manner. ComfyUI executes the JSON format workflows at the granularity of nodes, automatically loading models, conducting inference, passing variables, and returning results to users.", + "bbox": [ + 511, + 234, + 906, + 416 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Features. ComfyUI provides native support for common models and tools related to Stable Diffusion [44] [44], including but not limited to VAEs [25], CLIPs [42], LoRAs [45], IP-Adapters [71], and ControlNets [75]. With the joint efforts of the open-source community, hundreds of ComfyUI extensions are developed to support various latest models and tools, such as InsightFace [18], IC-Light [76], and Segment Anything [26], which bring about immense potential for realizing flexible and controllable generation tasks. By designing intricate workflows, users can produce fantastic artwork that cannot be achieved by common tools. Limitations. Despite the fancy effects that ComfyUI can achieve, the workflows are always manually crafted, which can be quite expensive. On the one hand, expert-level skills are required to figure out the appropriate nodes and parameters that can produce the desired results. On the other hand, workflows are designed to solve a specific task, while creating new workflows can be laborious and time-consuming.", + "bbox": [ + 511, + 417, + 908, + 690 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Benchmark Contents", + "text_level": 1, + "bbox": [ + 511, + 696, + 714, + 712 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ComfyBench collects annotations for 3205 nodes and 20 workflows and construct 200 tasks as is shown in Figure 2. More examples can be found in the supplementary material. Node document provides complete documentation for the 3205 nodes registered in ComfyUI until September 2024, serving as an important manual for agents to utilize the nodes properly. The source code is cloned from Salt AI Documentation [1], which is then cleaned and arranged into documentation, describing the usage of each node and their inputs and outputs. This structured documentation serves as a solid reference resource for understanding the complex node ecosystem within the ComfyUI framework.", + "bbox": [ + 511, + 719, + 906, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "24616", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/e936fb0a80897dd3fc39df4e872a7f0f0639304d6354931290ff1361e194c04a.jpg", + "image_caption": [ + "Figure 2. ComfyBench provides annotations for 3205 nodes and 20 workflows, together with 200 task instructions categorized into three difficulty levels: vanilla, complex, and creative." + ], + "image_footnote": [], + "bbox": [ + 93, + 88, + 305, + 324 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/758c9e14cecc1cbc7c36afddb1d449e623eca026940e70b3083ba616f2ac9440.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 310, + 88, + 480, + 324 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Curriculum workflow is a set of 20 tutorial workflows in ComfyUI, which introduces some common nodes and routines to build workflows for solving basic tasks. All the workflows are verified feasible with common nodes and models in ComfyUI and correctly annotated with their purposes and principles. The curriculum workflow allows agents to master useful skills in a human-like manner.", + "bbox": [ + 88, + 395, + 482, + 500 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Task instruction serves as the core part of the benchmark, containing 200 task instructions. Each instruction defines a specific task, which should be solved by designing and executing a ComfyUI workflow. To reflect the generalization ability of agents, the tasks are manually categorized into three difficulty levels: vanilla, complex, and creative.", + "bbox": [ + 89, + 501, + 483, + 592 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Vanilla includes 100 tasks, which can be solved by learning from a single curriculum workflow, involving minor modifications on several nodes and their parameters.", + "- Complex includes 60 tasks, which necessitates learning from more than one curriculum workflow, involving functional combinations of multiple pipelines and minor adjustments on nodes and parameters.", + "- Creative includes 40 tasks, which cannot be solved by directly imitating one or more curriculum workflows, requiring agents to understand the underlying principles and apply the learned skills in a novel way." + ], + "bbox": [ + 89, + 593, + 482, + 758 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Besides, we provide an auxiliary dataset as the input for some of the tasks. If an image or video is involved in the task, it will be specified in the instruction.", + "bbox": [ + 89, + 760, + 483, + 805 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Evaluation Metrics", + "text_level": 1, + "bbox": [ + 89, + 816, + 274, + 832 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Different from the common benchmarks for image or video generation, we are evaluating the generated workflows, so traditional metrics are not applicable. We divide the evaluation into two progressive stages. The first stage is to check", + "bbox": [ + 89, + 839, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "whether the generated workflows are syntactically and semantically correct, and the second stage is to check whether the generated workflows can produce the expected results in the task instructions. We design two metrics: pass rate and resolve rate, corresponding to the two stages respectively.", + "bbox": [ + 511, + 90, + 903, + 167 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Pass rate measures the ratio of tasks where generated workflows are syntactically and semantically correct. In practice, each workflow is sent to the server of ComfyUI, which automatically conducts validation and execution. A task will be marked as passed only if the server finishes the execution process and returns a success message.", + "bbox": [ + 511, + 167, + 903, + 257 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Resolve rate measures the ratio of tasks where generated workflows can produce the expected results in the task instruction. In other words, we need to automatically determine the consistency between the produced images or videos and the task requirements. In practice, we apply Visual Language Models (VLMs), where different prompt templates are written according to the input and output modalities. Specifically, we adopt GPT-4o for evaluation, as it demonstrates strong capabilities for understanding language and analyzing visual content. For images, we directly encode them into the prompt. For videos, we uniformly sample up to 10 frames before encoding, which compresses the context length within a reasonable range. A task will be marked as resolved only if the GPT-4o model confirms that all the requirements are satisfied.", + "bbox": [ + 511, + 258, + 906, + 484 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Human Evaluation", + "text_level": 1, + "bbox": [ + 511, + 494, + 694, + 510 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The computation of the resolve rate involves using VLMs to determine the consistency between the produced images or videos and the task requirements, which unavoidably involves subjectivity. To verify the reliability of VLM-based evaluation, we conduct a human evaluation to analyze the agreement between human evaluators and GPT-4o.", + "bbox": [ + 511, + 517, + 905, + 608 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We randomly select 70 tasks that are either completed by ComfyAgent proposed in Section 4 or by a human expert. Human evaluators and the latest GPT-4o model are asked to provide their judgments respectively. We compute the average scores for each question under a sample size of 560 and conduct a correlation analysis. Results are presented in Table 1, where the statistics indicate a strong agreement between human evaluators and GPT-4o, which verifies the reliability of our proposed VLM-based evaluation. More details can be found in the supplementary material.", + "bbox": [ + 511, + 609, + 905, + 761 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/a493af92df48a2aa7bd3ad3821a9b3842eda4ead0aaea2dd2ab82de9900059c7.jpg", + "table_caption": [ + "Table 1. Correlation analysis between the average scores given by human evaluators and GPT-4o indicates a strong agreement." + ], + "table_footnote": [], + "table_body": "
MetricStatisticP-value
Kendall's τ0.6455.427 × 10-11
Pearson's r0.7641.482 × 10-14
Spearman's ρ0.7373.672 × 10-13
", + "bbox": [ + 557, + 813, + 862, + 891 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "24617", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/03ca6b663ec4ba2105a522171c52d968b5127a0306d822290786268bbf4b60b4.jpg", + "image_caption": [ + "Figure 3. The architecture of the ComfyAgent framework. Multiple agents cooperate to design workflows in a step-by-step manner. Given the task instruction, the planner initializes the memory and produces a plan. For each step, the planner updates the plan and forms an action. Different actions, including combine, adapt, and retrieve, are then handled by corresponding agents. After combination or adaptation, refine action will be conducted to ensure the correctness. All the agents can interact with the memory, which consists of history, reference, and workspace. Once the task is deemed completed, the planner will finish the procedure and save the workflow." + ], + "image_footnote": [], + "bbox": [ + 93, + 88, + 906, + 256 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. ComfyAgent", + "text_level": 1, + "bbox": [ + 89, + 361, + 225, + 378 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Workflow Representation", + "text_level": 1, + "bbox": [ + 89, + 386, + 323, + 402 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/340be82b2a1bbcc03409b3e02d0b87f2bf68485fb68ef616563355887aad67fb.jpg", + "image_caption": [ + "Figure 4. Examples of four common formats to represent workflows: flow graph, JSON, element list, and code." + ], + "image_footnote": [], + "bbox": [ + 93, + 422, + 480, + 661 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Collaborative AI systems coordinate multiple interacting components, such as AI models and external tools, using workflows. These workflows are essentially high-level diagrams that outline the operational logic. They can be naturally represented as directed acyclic graphs (DAGs) consisting of two main elements: vertices and directed edges. Vertices represent information processing components with input and output interfaces, while directed edges represent information flow between components. Specifically, the incoming edges of a vertex represent the external information it receives, while the outgoing edges signify the information it outputs, with multiple edges illustrating various", + "bbox": [ + 89, + 719, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "information paths. This abstraction allows efficient construction and modification of collaborative AI systems, as demonstrated by platforms like ComfyUI [13].", + "bbox": [ + 511, + 362, + 903, + 407 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "There are four common formats to represent workflows: flow graph, JSON, element list, and code, as shown in Figure 4. Flow graph [13] provides intuitive visualization for humans but is unsuitable for LLM or VLM processing. JSON, typically used in [77], provides a structured representation but is unwieldy for complex workflows due to redundant information and LLMs' context limitations. Element list, typically used in [31], is closer to natural language and provides a more compact representation, but it lacks explicit topological relationships, hindering LLMs from correctly processing complex workflows.", + "bbox": [ + 511, + 409, + 905, + 575 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In contrast, code emerges as the most effective representation, offering various advantages including Turing completeness, rich semantic information, and natural compatibility with LLMs' code generation capabilities. We implement code representation using a restricted subset of Python-like syntax, involving basic operations and control structures while excluding advanced features. We also conduct an ablation study in Section 5.4 to demonstrate the superiority of code representation over other representations.", + "bbox": [ + 511, + 577, + 905, + 712 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Multi-Agent Framework", + "text_level": 1, + "bbox": [ + 511, + 724, + 740, + 739 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Through code representation, we can effectively prompt LLMs to generate workflows for the expected collaborative AI systems. However, with the complexity of workflows increasing, traditional agents suffer from context limitation and hallucination problems, failing to generate workflows correctly and stably, which is proved in Section 5.3.", + "bbox": [ + 511, + 750, + 905, + 839 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To achieve reliable generation of ComfyUI workflows, we propose ComfyAgent, a multi-agent framework to design workflows in a step-by-step manner. As demonstrated in Figure 3, ComfyAgent consists of three independent", + "bbox": [ + 511, + 840, + 905, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "24618", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/3d91971c2a74b4f6cee354a1a71d868971f9d1de174d1a2c048aad04a6d01cb3.jpg", + "table_caption": [ + "Table 2. Evaluation results of all the agents on ComfyBench. The average task-wise token consumption, the pass rate and resolve rate of every task category denoted by %P and %R, and the overall result on ComfyBench, are reported. The best results are highlighted in bold." + ], + "table_footnote": [], + "table_body": "
AgentCostVanillaComplexCreativeTotal
#Tokens (k)%P%R%P%R%P%R%P%R
GPT-4o + Zero-shot0.500.00.00.00.00.00.00.00.0
GPT-4o + Few-shot2.5532.027.016.78.37.50.022.516.0
GPT-4o + CoT2.5044.029.011.78.312.50.028.017.0
GPT-4o + CoT-SC11.0345.034.011.75.015.00.029.018.5
Claude-3.5-Sonnet + RAG5.7327.013.023.06.77.50.022.08.5
Llama-3.1-70B + RAG5.0458.032.023.010.015.05.039.020.0
GPT-4o + RAG4.9762.041.045.021.740.07.552.023.0
o1-mini + RAG6.0532.016.021.78.312.57.525.012.0
o1-preview + RAG9.4670.046.048.323.330.012.555.532.5
Llama-3.1-70B + ComfyAgent24.2263.035.026.718.320.05.043.524.0
GPT-4o + ComfyAgent14.9667.046.048.321.740.015.056.032.5
", + "bbox": [ + 91, + 128, + 906, + 354 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "modules: Memory, Planner, and Actions.", + "bbox": [ + 89, + 378, + 364, + 393 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Memory stores the recent state of ComfyAgent, which is built on psychological theories [4]. Specifically, it refers to the working memory [5], reflecting the current circumstances, namely the recent history behaviors, results from intermediate steps, internal reasoning, and external knowledge. We formulate them into three parts.", + "bbox": [ + 89, + 395, + 482, + 487 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- History maintains recent plans and actions of Planner, enabling action review for subsequent planning.", + "- Reference stores information retrieved from the knowledge base, and can be updated through active retrieval.", + "- **Workspace contains the current workflow together with its natural language annotation, which maintains a complete record of the current working status.**" + ], + "bbox": [ + 89, + 487, + 482, + 593 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Planner, which can be specified as PlanAgent, serves as the core of ComfyAgent, providing the global scheme to design and modify workflows under the task instruction.", + "bbox": [ + 89, + 594, + 482, + 638 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- At the beginning of the task, PlanAgent selects an existing workflow to initialize the memory and produces a thorough multi-step plan based on the task instruction.", + "- For each step, PlanAgent produces a high-level plan, together with an action based on the current memory, which will then be processed by other agents.", + "- For each step, PlanAgent evaluates the completion status of the task. Once the task is deemed completed, PlanAgent will finish the procedure and save the workflow." + ], + "bbox": [ + 89, + 641, + 482, + 776 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Actions represent different actions that can be selected by PlanAgent. All the actions are meant to modify the current memory. In ComfyAgent, we have three actions.", + "bbox": [ + 89, + 777, + 482, + 823 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Combine is processed by CombineAgent, which aims to combine the current workflow with another workflow from references specified by PlanAgent.", + "- Adapt is processed by AdaptAgent, which is responsible for adapting the details (e.g. parameters) of the current" + ], + "bbox": [ + 89, + 825, + 482, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "workflow based on the prompt given by PlanAgent.", + "bbox": [ + 526, + 378, + 867, + 393 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "- Retrieve is processed by RetrieveAgent, which follows the prompt given by PlanAgent to retrieve relevant information and update references accordingly.", + "bbox": [ + 511, + 393, + 903, + 439 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "After combination or adaptation, the updated workflow will be checked and refined by RefineAgent before entering the workspace, which avoids the accumulation of errors.", + "bbox": [ + 511, + 440, + 903, + 484 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "After the action is processed, ComfyAgent will enter a new step, where PlanAgent updates the existing plan and forms a new action. In such a way, ComfyAgent gradually generates the workflow. When the procedure is finished, the code representation will be converted into the standard format, which exactly describes the collaborative AI system.", + "bbox": [ + 511, + 486, + 905, + 577 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 513, + 593, + 645, + 611 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Baseline Agents", + "text_level": 1, + "bbox": [ + 513, + 619, + 669, + 635 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ComfyBench places strong requirements on the learning and generalization ability of agents, necessitating LLMs with powerful reasoning capabilities and sufficient context length. Therefore, we conduct experiments on the advanced models, involving both commercial and open-source LLMs.", + "bbox": [ + 511, + 642, + 905, + 718 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Llama-3.1 (llama-3.1-70b-instruct) [15] is an LLM developed by Meta, achieving state-of-the-art performance in the open-source community.", + "- Claude-3.5 (claude-3.5-sonnet-20240620) [3] is a commercial LLM developed by Anthropic, with a reputation for its creativity and interaction quality.", + "- GPT-4o (gpt-4o-2024-08-06) [38] is currently the mainstream model provided by OpenAI and most widely used by the community. It still maintains state-of-the-art performance in leaderboards of various domains.", + "- o1-mini (o1-mini-2024-09-12) and o1-preview (o1-preview-2024-09-12) [39] are the latest mod" + ], + "bbox": [ + 511, + 719, + 903, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "24619", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "els released by OpenAI, which focus on complex reasoning tasks such as mathematics and programming.", + "bbox": [ + 102, + 90, + 482, + 121 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We adopt five methods that have proved to be universally effective by the community and can be conveniently adapted to solve the tasks in ComfyBench.", + "bbox": [ + 89, + 122, + 482, + 167 + ], + "page_idx": 6 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Zero-shot Learning directly feeds LLMs with the task instruction to conduct inference.", + "- Few-shot Learning [9] provides a set of demonstrations in the prompt, which utilizes the in-context learning ability of LLMs and facilitates the correct generation.", + "- Chain-of-Thought (CoT) [57] is improved based on few-shot learning, which instructs the agent to articulate the reasoning process before providing the final answer.", + "- CoT with Self-consistency (CoT-SC) [53] is an improved version of CoT, which ensembles parallel trajectories and then selects the most consistent answer.", + "- Retrieval-Augmented Generation (RAG) [29] learns from the retrieved demonstrations, which efficiently utilizes the most relevant information." + ], + "bbox": [ + 89, + 167, + 482, + 378 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Considering the limited experimental budget, we evaluate the methods mainly on GPT-4o. ComfyAgent is also evaluated on Llama-3.1 to verify its performance on the open-source model. RAG is evaluated on all the models to provide a horizontal comparison of their capabilities.", + "bbox": [ + 89, + 381, + 483, + 455 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Implementation Details", + "text_level": 1, + "bbox": [ + 89, + 468, + 307, + 484 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ComfyUI Settings. The experiments are conducted on a fixed version of ComfyUI with necessary extensions. A set of common models is installed to make the workflows runnable and the results reproducible. In ComfyUI, the workflows are natively represented in JSON format. We implement a parser to extract the entire DAG described in the JSON object, which is then converted into a sequence of Python-like function calls in topological order. Another parser is implemented to restore this conversion. This ensures that workflows can be effectively generated by agents through code representation while remaining fully compatible with the ComfyUI execution environment. Note that the node documents are not fed into the agents due to context limitations, even though this technique would potentially improve their performance.", + "bbox": [ + 88, + 491, + 482, + 718 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Agent Parameters. We follow CodeAct [54] to format the interaction with LLMs. The parameters of the agents are manually set to exploit their largest potential within a limited context length. For few-shot learning, CoT, and CoT-SC, the number of demonstrations is consistently set to 3, but CoT-SC will ensemble 3 trajectories in parallel to form the final answer. For RAG and ComfyAgent, we adopt an advanced embedding model (text-embedding-3-large) [40] released by OpenAI to encode the texts for retrieval, and the number of retrieved demonstrations is set to 5. For ComfyAgent, the modules are implemented as different roles played by the", + "bbox": [ + 88, + 719, + 482, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/c874ec8ed6c149495edaf30f90f06e8ef484ec27c0c34f2cb964600f4c8e164a.jpg", + "table_caption": [ + "Table 3. Ablation results of RAG on GPT-4o with different workflow representations on ComfyBench. The pass rate and resolve rate are reported, denoted by %P and %R respectively." + ], + "table_footnote": [], + "table_body": "
RepresentationVanillaComplexCreative
%P%R%P%R%P%R
Flow Graph------
JSON52.033.025.08.317.57.5
Element List51.030.023.35.020.010.0
Code62.041.045.015.030.07.5
", + "bbox": [ + 514, + 142, + 903, + 247 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/607da0b1567adb3d1f37bd70c7c5715c0d3eca0bb03de12a7d73a4a75b8fbaa1.jpg", + "table_caption": [ + "Table 4. Ablation results of ComfyAgent on GPT-4o with different architectures on ComfyBench. Each variant is implemented by removing one agent from the multi-agent framework. The pass rate and resolve rate are reported, denoted by %P and %R respectively." + ], + "table_footnote": [], + "table_body": "
ArchitectureVanillaComplexCreative
%P%R%P%R%P%R
ComfyAgent67.046.048.321.740.015.0
w/o Combine69.043.015.05.035.015.0
w/o Adapt52.026.020.010.017.55.0
w/o Retrieve27.016.020.05.07.57.5
w/o Refine65.042.043.321.735.012.5
", + "bbox": [ + 522, + 329, + 898, + 449 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "same LLM. We provide the prompt implementation and an example trajectory in the supplementary material.", + "bbox": [ + 511, + 474, + 906, + 506 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Evaluation Results", + "text_level": 1, + "bbox": [ + 511, + 515, + 694, + 529 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We report the evaluation results of different agents on ComfyBench in Table 2. All the agents follow the parameter settings in Section 5.2 and adopt code representation in Section 4.1. We also present some examples generated by ComfyAgent on ComfyBench in Table 5. More examples can be found in the supplementary material.", + "bbox": [ + 511, + 537, + 906, + 628 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In terms of methods, zero-shot learning fails due to a lack of domain knowledge. CoT achieves a better performance than few-shot learning, and CoT-SC shows a slight improvement over CoT. However, these agents fail to achieve comparable performance with retrieval-based agents, because they cannot fully exploit the external knowledge provided by ComfyBench. ComfyAgent outperforms all the baselines in every task category, achieving an overall pass rate of $56\\%$ and resolve rate of $32.5\\%$ , which indicates that the multi-agent framework applied in ComfyAgent is capable of alleviating the hallucination problem and improving the generation quality and stability. This improvement is consistent on both open-source and commercial models.", + "bbox": [ + 511, + 628, + 905, + 824 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In terms of models, Claude-3.5 and o1-mini show poor performance under RAG, which is mainly caused by the hallucination problem, where models modify the nodes or parameters without verifying their validity. In contrast, Llama-3.1 and GPT-4o show less hallucination and achieve", + "bbox": [ + 511, + 825, + 906, + 898 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "24620", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 5. Some examples generated by ComfyAgent on ComfyBench. Given the task instruction, ComfyAgent designs a workflow to describe the collaborative AI system in ComfyUI, which is then executed to generate the result required by the instruction.", + "bbox": [ + 89, + 88, + 906, + 118 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Task Instruction", + "text_level": 1, + "bbox": [ + 176, + 135, + 282, + 147 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "You are given an image of a male celebrity. Transform the man in the image into a beautiful woman with ponytail hair while preserving her facial identity.", + "bbox": [ + 101, + 152, + 349, + 208 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9850924b40131c41758c93ce99995792158edeae95609ce6e8686df0f16391ce.jpg", + "image_caption": [ + "Input" + ], + "image_footnote": [], + "bbox": [ + 102, + 231, + 225, + 327 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e5c87d670caae48f57e25f3c959eab525508c1c4f2922472c15c79247d0bd3d9.jpg", + "image_caption": [ + "Result" + ], + "image_footnote": [], + "bbox": [ + 230, + 231, + 356, + 327 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Task Instruction", + "text_level": 1, + "bbox": [ + 444, + 135, + 552, + 147 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "You are given an image of a standing cat. Replace the background with a scene of a cozy living room while keeping the lighting and shadows consistent.", + "bbox": [ + 370, + 152, + 619, + 208 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/864ec1fe4c53ac904d443498c676c50f83e2fea4928a0a1b0717d3d9b669389a.jpg", + "image_caption": [ + "Input" + ], + "image_footnote": [], + "bbox": [ + 372, + 231, + 496, + 327 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6766d351c96778144f5b08142df8f83fe284d52258f533f9ecc74f4f4eae225d.jpg", + "image_caption": [ + "Result" + ], + "image_footnote": [], + "bbox": [ + 501, + 231, + 625, + 327 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Task Instruction", + "text_level": 1, + "bbox": [ + 715, + 135, + 821, + 147 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "You are given an image of a red bridge with a person standing on it. Remove the person from the image while maintaining the original appearance of the bridge.", + "bbox": [ + 640, + 152, + 888, + 208 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7851733a1bc670c7d7e0c418aa23d2104fe11fd8035cb220e3ddb13a69d5a6e4.jpg", + "image_caption": [ + "Input" + ], + "image_footnote": [], + "bbox": [ + 642, + 231, + 766, + 327 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9b3d6606dda2e2a66e4db56d8c3891be29311c94e680b0b935dd1e934ae44cf0.jpg", + "image_caption": [ + "Result" + ], + "image_footnote": [], + "bbox": [ + 769, + 231, + 893, + 327 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "better performance, and o1-preview achieves an outstanding performance comparable to ComfyAgent on GPT-4o.", + "bbox": [ + 89, + 359, + 482, + 391 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.4. Ablation Studies", + "text_level": 1, + "bbox": [ + 89, + 406, + 254, + 420 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Representation. We conduct an ablation study to verify the advantages of representing workflows with code introduced in Section 4.1. We implement three variants of RAG on GPT-4o, where the workflows are respectively represented in JSON, element list, and code. The variants are evaluated on ComfyBench with the same parameter settings as in Section 5.3. Flow graph is not implemented because it is not understandable by LLM-based agents. We report the ablation results in Table 3, where RAG with code representation shows superior performance over other representations, which verifies our claim. RAG with JSON representation is expected to perform even worse without the unique feature of GPT-4o to generate structured JSON objects.", + "bbox": [ + 89, + 430, + 483, + 626 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Architecture. We conduct another ablation to demonstrate the rationality of the multi-agent framework introduced in Section 4.2, where four variants of ComfyAgent on GPT-4o are implemented, each removing one agent from the original framework. We report the ablation results in Table 4. We find that removing any of the four agents will lead to a performance drop. Removing CombineAgent alleviates hallucination in the vanilla category, but leads to failure in the complex category, demonstrating the necessity of integrating multiple demonstrations. Removing AdaptAgent or RetrieveAgent leads to a significant performance drop in all the categories, which underscores the critical role of detail modification and knowledge utilization. Removing RefineAgent leads to a minor but noticeable performance drop in all the categories, indicating its effectiveness in mitigating error accumulation. Therefore, we can conclude that our meticulously-designed multi-agent framework contributes to the superior performance of ComfyAgent.", + "bbox": [ + 89, + 628, + 483, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 359, + 633, + 375 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we investigate the potential of LLM-based agents to autonomously design collaborative AI systems. Building upon the open-source ComfyUI [13] platform, we present ComfyBench, an autonomous workflow design environment serving as both a benchmark for evaluating the capabilities of agents in designing collaborative systems and a means to extend the applicability of ComfyUI in real-world, language-instruction-based generative applications. We further develop ComfyAgent, which achieves a comparable resolve rate to o1-preview [39] and significantly surpasses other agents on ComfyBench. We point out two key insights underpinning ComfyAgent. First, representing workflows through code is efficient, which provides rich semantics for agents to better understand complex systems. Second, the multi-agent framework of ComfyAgent highlights the substantial benefits of task decomposition and collaborative generation. However, the top-performing agent, ComfyAgent, has achieved a resolve rate of only $15\\%$ for creative tasks on ComfyBench. LLM-based agents still have a long way to go in designing collaborative AI systems. We hope that ComfyBench and ComfyAgent can serve as valuable resources in advancing the development of intelligent and autonomous collaborative AI systems.", + "bbox": [ + 511, + 386, + 906, + 734 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations. Well-designed fine-tuning methods could potentially elevate agents to an expert level on ComfyBench, but they are quite expensive and thus not involved in our work. In addition, ComfyAgent operates in an open-loop manner. An alternative approach would involve using VLMs as evaluators to provide feedback [21], enabling a closed-loop system where agents can refine their workflows based on evaluations, thus leading to better results.", + "bbox": [ + 511, + 734, + 906, + 856 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments. This work is supported by the Shanghai Municipal Science and Technology Major Project.", + "bbox": [ + 511, + 857, + 906, + 887 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "24621", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 90, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Salt AI. Salt node documentation. https://docs. getsalt.ai/md/, 2024.3", + "[2] Google DeepMind AlphaCode Team. Alphacode 2 technical report. https://storage.googleapis.com/deepmind-media/AlphaCode2/AlphaCode2_Tech_Report.pdf, 2024.1,3", + "[3] Anthropic. Claude 3.5 sonnet. https://www.anthropic.com/news/claude-3-5-sonnet, 2024.6", + "[4] Richard C Atkinson. Human memory: A proposed system and its control processes. The psychology of learning and motivation, 1968. 6", + "[5] Alan Baddeley. Working memory. Science, 1992. 6", + "[6] Yuntao Bai, Andy Jones, Kamal Ndousse, Amanda Askell, Anna Chen, Nova DasSarma, Dawn Drain, Stanislav Fort, Deep Ganguli, Tom Henighan, et al. Training a helpful and harmless assistant with reinforcement learning from human feedback. arXiv preprint arXiv:2204.05862, 2022. 3", + "[7] Maciej Besta, Nils Blach, Ales Kubicek, Robert Gerstenberger, Michal Podstawski, Lukas Gianinazzi, Joanna Gajda, Tomasz Lehmann, Hubert Niewiadomski, Piotr Nczyk, et al. Graph of thoughts: Solving elaborate problems with large language models. In AAAI, 2024. 3", + "[8] Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Li Jing, David Schnurr, Joe Taylor, Troy Luhman, Eric Luhman, Clarence Ng, Ricky Wang, and Aditya Ramesh. Video generation models as world simulators. https://openai.com/research/video-generation-models-as-world-simulators, 2024.3", + "[9] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. In NeurIPS, 2020. 2, 7", + "[10] Ruisheng Cao, Fangyu Lei, Haoyuan Wu, Jixuan Chen, Yeqiao Fu, Hongcheng Gao, Xinzhuang Xiong, Hanchong Zhang, Yuchen Mao, Wenjing Hu, et al. Spider2-v: How far are multimodal agents from automating data science and engineering workflows? arXiv preprint arXiv:2407.10956, 2024.3", + "[11] Weize Chen, Yusheng Su, Jingwei Zuo, Cheng Yang, Chenfei Yuan, Chen Qian, Chi-Min Chan, Yujia Qin, Yaxi Lu, Ruobing Xie, et al. Agentverse: Facilitating multi-agent collaboration and exploring emergent behaviors in agents. arXiv preprint arXiv:2308.10848, 2023. 3", + "[12] Zehui Chen, Kuikun Liu, Qiuchen Wang, Jiangning Liu, Wenwei Zhang, Kai Chen, and Feng Zhao. Mindsearch: Mimicking human minds elicits deep ai searcher. arXiv preprint arXiv:2407.20183, 2024. 3", + "[13] comfyanonymous. Comfyui. https://github.com/comfyanonymous/ComfyUI, 2023. 1, 3, 5, 8", + "[14] Xiang Deng, Yu Gu, Boyuan Zheng, Shijie Chen, Sam Stevens, Boshi Wang, Huan Sun, and Yu Su. Mind2web: Towards a generalist agent for the web. In NeurIPS, 2024. 3", + "[15] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil" + ], + "bbox": [ + 93, + 114, + 482, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 6", + "[16] Yunfan Gao, Yun Xiong, Xinyu Gao, Kangxiang Jia, Jinliu Pan, Yuxi Bi, Yi Dai, Jiawei Sun, and Haofen Wang. Retrieval-augmented generation for large language models: A survey. arXiv preprint arXiv:2312.10997, 2023. 3", + "[17] Yingqiang Ge, Wenyue Hua, Kai Mei, Juntao Tan, Shuyuan Xu, Zelong Li, Yongfeng Zhang, et al. Openagi: When llm meets domain experts. In NeurlPS, 2024. 3", + "[18] Jia Guo, Jiankang Deng, Xiang An, Jack Yu, and Baris Gecer. Insightface: 2d and 3d face analysis project. https://github.com/deepinsight/insightface, 2023.3", + "[19] Taicheng Guo, Xiuying Chen, Yaqi Wang, Ruidi Chang, Shichao Pei, Nitesh V Chawla, Olaf Wiest, and Xiangliang Zhang. Large language model based multi-agents: A survey of progress and challenges. arXiv preprint arXiv:2402.01680, 2024. 2", + "[20] Sirui Hong, Xiawu Zheng, Jonathan Chen, Yuheng Cheng, Jinlin Wang, Ceyao Zhang, Zili Wang, Steven Ka Shing Yau, Zijuan Lin, Liyang Zhou, et al. Metagpt: Meta programming for multi-agent collaborative framework. arXiv preprint arXiv:2308.00352, 2023. 3", + "[21] Shengran Hu, Cong Lu, and Jeff Clune. Automated design of agentic systems. arXiv preprint arXiv:2408.08435, 2024. 2, 8", + "[22] Ian Huang, Guandao Yang, and Leonidas Guibas. Blender alchemy: Editing 3d graphics with vision-language models. arXiv preprint arXiv:2404.17672, 2024. 3", + "[23] Timo Kaufmann, Paul Weng, Viktor Bengs, and Eyke Hüllermeier. A survey of reinforcement learning from human feedback. arXiv preprint arXiv:2312.14925, 2023. 3", + "[24] Omar Khattab, Arnav Singhvi, Paridhi Maheshwari, Zhiyuan Zhang, Keshav Santhanam, Sri Vardhamanan, Saiful Haq, Ashutosh Sharma, Thomas T Joshi, Hanna Moazam, et al. Dspy: Compiling declarative language model calls into self-improving pipelines. arXiv preprint arXiv:2310.03714, 2023. 3", + "[25] Diederik P Kingma. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 3", + "[26] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dólar, and Ross Girshick. Segment anything. arXiv preprint arXiv:2304.02643, 2023. 3", + "[27] Jing Yu Koh, Robert Lo, Lawrence Jang, Vikram Duvvur, Ming Chong Lim, Po-Yu Huang, Graham Neubig, Shuyan Zhou, Ruslan Salakhutdinov, and Daniel Fried. Visualwebarena: Evaluating multimodal agents on realistic visual web tasks. arXiv preprint arXiv:2401.13649, 2024. 3", + "[28] LangChain. Applications that can reason. powered by langchain. https://www.langchain.com/, 2023.3", + "[29] Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Kuttler, Mike Lewis, Wen-tau Yih, Tim Rocktäschel, et al. Retrieval-augmented generation for knowledge-intensive nlp tasks. In NeurIPS, 2020. 2, 3, 7" + ], + "bbox": [ + 516, + 93, + 903, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "24622", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[30] Guohao Li, Hasan Hammoud, Hani Itani, Dmitrii Khizbullin, and Bernard Ghanem. Camel: Communicative agents for\" mind\" exploration of large language model society. In NeurIPS, 2023. 3", + "[31] Zelong Li, Shuyuan Xu, Kai Mei, Wenyue Hua, Balaji Rama, Om Raheja, Hao Wang, He Zhu, and Yongfeng Zhang. Autoflow: Automated workflow generation for large language model agents. arXiv preprint arXiv:2407.12821, 2024. 1, 5", + "[32] Tian Liang, Zhiwei He, Wenxiang Jiao, Xing Wang, Yan Wang, Rui Wang, Yujiu Yang, Shuming Shi, and Zhaopeng Tu. Encouraging divergent thinking in large language models through multi-agent debate. arXiv preprint arXiv:2305.19118, 2023. 3", + "[33] Aman Madaan, Niket Tandon, Prakhar Gupta, Skyler Hallinan, Luyu Gao, Sarah Wiegreffe, Uri Alon, Nouha Dziri, Shrimai Prabhumoye, Yiming Yang, et al. Self-refine: Iterative refinement with self-feedback. In NeurIPS, 2024. 3", + "[34] Kai Mei, Zelong Li, Shuyuan Xu, Ruosong Ye, Yingqiang Ge, and Yongfeng Zhang. LIm agent operating system. arXiv preprint arXiv:2403.16971, 2024. 3", + "[35] Reiichiro Nakano, Jacob Hilton, Suchir Balaji, Jeff Wu, Long Ouyang, Christina Kim, Christopher Hesse, Shantanu Jain, Vineet Kosaraju, William Saunders, et al. Webgpt: Browser-assisted question-answering with human feedback. arXiv preprint arXiv:2112.09332, 2021. 1, 3", + "[36] Harsha Nori, Yin Tat Lee, Sheng Zhang, Dean Carignan, Richard Edgar, Nicolo Fusi, Nicholas King, Jonathan Larson, Yuanzhi Li, Weishung Liu, et al. Can generalist foundation models outcompete special-purpose tuning? case study in medicine. arXiv preprint arXiv:2311.16452, 2023. 3", + "[37] OpenAI. Chatgpt. https://openai.com/chatgpt/overview/, 2024.1", + "[38] OpenAI. Hello gpt-4o. https://openai.com/index/hello-gpt-4o/, 2024.6", + "[39] OpenAI. Introducing openai ol-preview. https://openai.com/index/introducing-openai-ol-preview/, 2024.2,6,8", + "[40] OpenAI. New embedding models and api updates. https://openai.com/index/new-embedding-models-and-api-updates/, 2024.7", + "[41] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. In NeurIPS, 2022. 3", + "[42] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 3", + "[43] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In ICML, 2021. 1", + "[44] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 3" + ], + "bbox": [ + 91, + 90, + 480, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[45] Simo Ryu. Low-rank adaptation for fast text-to-image diffusion fine-tuning. https://github.com/cloneofsimo/lora, 2023.3", + "[46] Yongliang Shen, Kaitao Song, Xu Tan, Dongsheng Li, Weiming Lu, and Yueting Zhuang. Hugginggpt: Solving ai tasks with chatgpt and its friends in hugging face. In NeurIPS, 2024. 3", + "[47] Chan Hee Song, Jiaman Wu, Clayton Washington, Brian M Sadler, Wei-Lun Chao, and Yu Su. Llm-planner: Few-shot grounded planning for embodied agents with large language models. In CVPR, 2023. 3", + "[48] Theodore R Sumers, Shunyu Yao, Karthik Narasimhan, and Thomas L Griffiths. Cognitive architectures for language agents. arXiv preprint arXiv:2309.02427, 2023. 3", + "[49] Gemini Team, Rohan Anil, Sebastian Borgeaud, Yonghui Wu, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, et al. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805, 2023. 3", + "[50] Amitayush Thakur, Yeming Wen, and Swarat Chaudhuri. A language-agent approach to formal theorem-proving. arXiv preprint arXiv:2310.04353, 2023. 3", + "[51] Trieu H Trinh, Yuhuai Wu, Quoc V Le, He He, and Thang Luong. Solving olympiad geometry without human demonstrations. Nature, 2024. 3", + "[52] Lei Wang, Chen Ma, Xueyang Feng, Zeyu Zhang, Hao Yang, Jingsen Zhang, Zhiyuan Chen, Jiakai Tang, Xu Chen, Yankai Lin, et al. A survey on large language model based autonomous agents. Frontiers of Computer Science, 2024. 3", + "[53] Xuezhi Wang, Jason Wei, Dale Schuurmans, Quoc Le, Ed Chi, Sharan Narang, Aakanksha Chowdhery, and Denny Zhou. Self-consistency improves chain of thought reasoning in language models. arXiv preprint arXiv:2203.11171, 2022. 2, 3, 7", + "[54] Xingyao Wang, Yangyi Chen, Lifan Yuan, Yizhe Zhang, Yunzhu Li, Hao Peng, and Heng Ji. Executable code actions elicit better lIm agents. arXiv preprint arXiv:2402.01030, 2024. 3, 7", + "[55] Xingyao Wang, Boxuan Li, Yufan Song, Frank F Xu, Xiangru Tang, Mingchen Zhuge, Jiayi Pan, Yueqi Song, Bowen Li, Jaskirat Singh, et al. Opendevin: An open platform for ai software developers as generalist agents. arXiv preprint arXiv:2407.16741, 2024. 3", + "[56] Zhiruo Wang, Shuyan Zhou, Daniel Fried, and Graham Neubig. Execution-based evaluation for open-domain code generation. arXiv preprint arXiv:2212.10481, 2022. 3", + "[57] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. In NeurIPS, 2022. 2, 3, 7", + "[58] Qingyun Wu, Gagan Bansal, Jieyu Zhang, Yiran Wu, Shaokun Zhang, Erkang Zhu, Beibin Li, Li Jiang, Xiaoyun Zhang, and Chi Wang. Autogen: Enabling next-gen llm applications via multi-agent conversation framework. arXiv preprint arXiv:2308.08155, 2023. 3", + "[59] Zhiyong Wu, Chengcheng Han, Zichen Ding, Zhenmin Weng, Zhoumianze Liu, Shunyu Yao, Tao Yu, and Lingpeng" + ], + "bbox": [ + 516, + 90, + 903, + 901 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "24623", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Kong. Os-copilot: Towards generalist computer agents with self-improvement. arXiv preprint arXiv:2402.07456, 2024. 3", + "[60] Zhiheng Xi, Wenxiang Chen, Xin Guo, Wei He, Yiwen Ding, Boyang Hong, Ming Zhang, Junzhe Wang, Senjie Jin, Enyu Zhou, et al. The rise and potential of large language model based agents: A survey. arXiv preprint arXiv:2309.07864, 2023. 3", + "[61] Zhiheng Xi, Yiwen Ding, Wenxiang Chen, Boyang Hong, Honglin Guo, Junzhe Wang, Dingwen Yang, Chenyang Liao, Xin Guo, Wei He, et al. Agentgym: Evolving large language model-based agents across diverse environments. arXiv preprint arXiv:2406.04151, 2024. 3", + "[62] Tianbao Xie, Fan Zhou, Zhoujun Cheng, Peng Shi, Luoxuan Weng, Yitao Liu, Toh Jing Hua, Junning Zhao, Qian Liu, Che Liu, et al. Openagents: An open platform for language agents in the wild. arXiv preprint arXiv:2310.10634, 2023. 3", + "[63] Tianbao Xie, Danyang Zhang, Jixuan Chen, Xiaochuan Li, Siheng Zhao, Ruisheng Cao, Toh Jing Hua, Zhoujun Cheng, Dongchan Shin, Fangyu Lei, et al. Osworld: Benchmarking multimodal agents for open-ended tasks in real computer environments. arXiv preprint arXiv:2404.07972, 2024. 3", + "[64] Fangzhi Xu, Qiushi Sun, Kanzhi Cheng, Jun Liu, Yu Qiao, and Zhiyong Wu. Interactive evolution: A neural-symbolic self-training framework for large language models. arXiv preprint arXiv:2406.11736, 2024. 3", + "[65] Shuyuan Xu, Zelong Li, Kai Mei, and Yongfeng Zhang. Core: LIm as interpreter for natural language programming, pseudo-code programming, and flow programming of ai agents. arXiv preprint arXiv:2405.06907, 2024. 2", + "[66] John Yang, Carlos E Jimenez, Alexander Wettig, Kilian Lieret, Shunyu Yao, Karthik Narasimhan, and Ofir Press. Swe-agent: Agent-computer interfaces enable automated software engineering. arXiv preprint arXiv:2405.15793, 2024.3", + "[67] John Yang, Akshara Prabhakar, Karthik Narasimhan, and Shunyu Yao. Intercode: Standardizing and benchmarking interactive coding with execution feedback. In NeurIPS, 2024. 3", + "[68] Zhao Yang, Jiaxuan Liu, Yucheng Han, Xin Chen, Zebiao Huang, Bin Fu, and Gang Yu. Appagent: Multimodal agents as smartphone users. arXiv preprint arXiv:2312.13771, 2023.3", + "[69] Shunyu Yao, Howard Chen, John Yang, and Karthik Narasimhan. Webshop: Towards scalable real-world web interaction with grounded language agents. In NeurIPS, 2022. 3", + "[70] Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, and Yuan Cao. React: Synergizing reasoning and acting in language models. arXiv preprint arXiv:2210.03629, 2022. 3", + "[71] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023. 3", + "[72] Pengcheng Yin, Wen-Ding Li, Kefan Xiao, Abhishek Rao, Yeming Wen, Kensen Shi, Joshua Howland, Paige Bailey," + ], + "bbox": [ + 91, + 90, + 482, + 901 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Michele Catasta, Henryk Michalewski, et al. Natural language to code generation in interactive data science notebooks. arXiv preprint arXiv:2212.09248, 2022. 3", + "[73] Zhengqing Yuan, Ruoxi Chen, Zhaoxu Li, Haolong Jia, Lifang He, Chi Wang, and Lichao Sun. Mora: Enabling generalist video generation via a multi-agent framework. arXiv preprint arXiv:2403.13248, 2024. 3", + "[74] Matei Zaharia, Omar Khattab, Lingjiao Chen, Jared Quincy Davis, Heather Miller, Chris Potts, James Zou, Michael Carbin, Jonathan Frankle, Naveen Rao, and Ali Ghodsi. The shift from models to compound ai systems. https://bair.berkeley.edu/blog/2024/02/18/compound-ai-systems/, 2024.1,3", + "[75] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In CVPR, 2023. 3", + "[76] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Imposing consistent light. https://github.com/lllyasviel/IC-Light, 2024.3", + "[77] Wenqi Zhang, Yongliang Shen, Weiming Lu, and Yueting Zhuang. Data-copilot: Bridging billions of data and humans with autonomous workflow. arXiv preprint arXiv:2306.07209, 2023. 3, 5", + "[78] Zeyu Zhang, Xiaohe Bo, Chen Ma, Rui Li, Xu Chen, Quanyu Dai, Jieming Zhu, Zhenhua Dong, and Ji-Rong Wen. A survey on the memory mechanism of large language model based agents. arXiv preprint arXiv:2404.13501, 2024. 3", + "[79] Wanjun Zhong, Lianghong Guo, Qiqi Gao, He Ye, and Yanlin Wang. Memorybank: Enhancing large language models with long-term memory. In AAAI, 2024. 3", + "[80] Shuyan Zhou, Frank F Xu, Hao Zhu, Xuhui Zhou, Robert Lo, Abishek Sridhar, Xianyi Cheng, Tianyue Ou, Yonatan Bisk, Daniel Fried, et al. Webarena: A realistic web environment for building autonomous agents. arXiv preprint arXiv:2307.13854, 2023. 3", + "[81] Wangchunshu Zhou, Yuchen Eleanor Jiang, Long Li, Jia-long Wu, Tiannan Wang, Shi Qiu, Jintian Zhang, Jing Chen, Ruipu Wu, Shuai Wang, et al. Agents: An open-source framework for autonomous language agents. arXiv preprint arXiv:2309.07870, 2023. 3", + "[82] Wangchunshu Zhou, Yixin Ou, Shengwei Ding, Long Li, Jialong Wu, Tiannan Wang, Jiamin Chen, Shuai Wang, Xiaohua Xu, Ningyu Zhang, et al. Symbolic learning enables self-evolving agents. arXiv preprint arXiv:2406.18532, 2024. 3", + "[83] Mingchen Zhuge, Wenyi Wang, Louis Kirsch, Francesco Faccio, Dmitrii Khizbullin, and Jürgen Schmidhuber. Gptswarm: Language agents as estimizable graphs. In ICML, 2024. 3" + ], + "bbox": [ + 516, + 90, + 903, + 768 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "24624", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/ComfyBench_ Benchmarking LLM-based Agents in ComfyUI for Autonomously Designing Collaborative AI Systems/9eeb1c75-c4c6-45dd-9896-e14c13fb270d_model.json b/2025/ComfyBench_ Benchmarking LLM-based Agents in ComfyUI for Autonomously Designing Collaborative AI Systems/9eeb1c75-c4c6-45dd-9896-e14c13fb270d_model.json new file mode 100644 index 0000000000000000000000000000000000000000..a093aac32da2c552b58e9efb80b3eba30a3f7b6e --- /dev/null +++ b/2025/ComfyBench_ Benchmarking LLM-based Agents in ComfyUI for Autonomously Designing Collaborative AI Systems/9eeb1c75-c4c6-45dd-9896-e14c13fb270d_model.json @@ -0,0 +1,2840 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.096, + 0.13, + 0.903, + 0.177 + ], + "angle": 0, + "content": "ComfyBench: Benchmarking LLM-based Agents in ComfyUI for Autonomous Designing Collaborative AI Systems" + }, + { + "type": "text", + "bbox": [ + 0.108, + 0.202, + 0.89, + 0.259 + ], + "angle": 0, + "content": "Xiangyuan Xue\\(^{1,2*}\\) Zeyu Lu\\(^{1,2}\\) Di Huang\\(^{3}\\) Zidong Wang\\(^{1,4}\\) Wanli Ouyang\\(^{1,4}\\) Lei Bai\\(^{1\\boxtimes}\\) \n\\(^{1}\\)Shanghai Artificial Intelligence Laboratory \n\\(^{2}\\)Shanghai Jiao Tong University \n\\(^{3}\\)The University of Sydney \n\\(^{4}\\)The Chinese University of Hong Kong" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.291, + 0.327, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.324, + 0.486, + 0.749 + ], + "angle": 0, + "content": "Much previous AI research has focused on developing monolithic models to maximize their intelligence, with the primary goal of enhancing performance on specific tasks. In contrast, this work attempts to study using LLM-based agents to design collaborative AI systems autonomously. To explore this problem, we first introduce ComfyBench to evaluate agents' ability to design collaborative AI systems in ComfyUI. ComfyBench is a comprehensive benchmark comprising 200 diverse tasks covering various instruction-following generation challenges, along with detailed annotations for 3,205 nodes and 20 workflows. Based on ComfyBench, we further develop ComfyAgent, a novel framework that empowers LLM-based agents to autonomously design collaborative AI systems by generating workflows. ComfyAgent is based on two core concepts. First, it represents workflows with code, which can be reversibly converted into workflows and executed as collaborative systems by the interpreter. Second, it constructs a multi-agent system that cooperates to learn from existing workflows and generate new workflows for a given task. While experimental results demonstrate that ComfyAgent achieves a comparable resolve rate to o1-preview and significantly surpasses other agents on ComfyBench, ComfyAgent has resolved only \\(15\\%\\) of creative tasks. LLM-based agents still have a long way to go in autonomously designing collaborative AI systems. Progress with ComfyBench is paving the way for more intelligent and autonomous collaborative AI systems. Our code is available at: https://github.com/xyxQwQ/ComfyBench." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.777, + 0.222, + 0.793 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.803, + 0.484, + 0.866 + ], + "angle": 0, + "content": "The recent evolution of AI is defined by the growing importance of collaborative AI systems, which integrate multiple models and tools to work as a whole collaborative system [74]. The success of ChatGPT [37] presents the pos" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.293, + 0.908, + 0.429 + ], + "angle": 0, + "content": "sibility of integrating a wide range of tasks, such as web browsing [35], image generation [43], and code execution, into a single conversational agent. Unlike traditional AI models that function as single entities, collaborative AI systems integrate multiple AI components, each contributing unique capabilities to solve complex problems. This shift towards integration has become crucial for achieving state-of-the-art results, as it leverages the combined strengths of diverse AI functionalities within a unified framework." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.43, + 0.909, + 0.627 + ], + "angle": 0, + "content": "Unfortunately, the design space and optimization of collaborative systems often require significant human expertise [31]. For instance, AlphaCode 2 [2] utilizes an intricate AI system to achieve expert-level performance in competitive programming, which involves fine-tuned LLMs for coding and scoring, expert models for clustering, and modules for code execution and filtering. These components collaborate in a precise manner carefully designed by human experts to achieve state-of-the-art performance, which is not achievable by any single component or their casual combination. This brings us to a pivotal question: Can we develop an agent, akin to a human expert, capable of autonomously designing collaborative AI systems?" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.628, + 0.909, + 0.719 + ], + "angle": 0, + "content": "In this work, we explore this question in ComfyUI [13], a trending and open-source platform that supports various generative tools and models in terms of nodes. Within ComfyUI, users can flexibly construct and connect nodes to build the workflow of the collaborative AI system, which is capable of producing highly customized artwork." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.911, + 0.903 + ], + "angle": 0, + "content": "To evaluate agents's ability to design collaborative AI systems in ComfyUI, we introduce ComfyBench, a comprehensive benchmark shown in Figure 1, which comprises 200 diverse tasks covering various instruction-following generation challenges, along with detailed annotations for 3,205 nodes and 20 workflows. We categorize tasks into three levels of difficulty: vanilla, complex, and creative. In each task, agents are required to create workflows that describe collaborative AI systems within ComfyUI, which can then be executed to produce the desired outcomes specified in the task instructions. We assess the performance of agents using two evaluation metrics: pass rate, showing the" + }, + { + "type": "page_footnote", + "bbox": [ + 0.09, + 0.875, + 0.483, + 0.901 + ], + "angle": 0, + "content": "*This work was done during his internship at Shanghai Artificial Intelligence Laboratory. Corresponding author, baisanshi@gmail.com." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "24614" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.094, + 0.089, + 0.208, + 0.105 + ], + "angle": 0, + "content": "(a) ComfyBench" + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.107, + 0.905, + 0.255 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.095, + 0.261, + 0.207, + 0.276 + ], + "angle": 0, + "content": "(b) ComfyAgent" + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.276, + 0.907, + 0.361 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.372, + 0.907, + 0.456 + ], + "angle": 0, + "content": "Figure 1. (a) ComfyBench is a comprehensive benchmark to evaluate agents's ability to design collaborative AI systems in ComfyUI. Given the task instruction, agents are required to learn from documents and create workflows to describe collaborative AI systems. The performance is measured by pass rate and resolve rate, reflecting whether the workflow can be correctly executed and whether the task requirements are realized. (b) ComfyAgent builds collaborative AI systems in ComfyUI by generating workflows. The workflows are converted into equivalent code so that LLMs can better understand them. ComfyAgent can learn from existing workflows and autonomously design new ones. The generated workflows can be interpreted as collaborative AI systems to complete given tasks." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.483, + 0.483, + 0.513 + ], + "angle": 0, + "content": "success of workflow execution, and resovle rate, reflecting the effectiveness in meeting task requirements." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.518, + 0.483, + 0.685 + ], + "angle": 0, + "content": "Based on ComfyBench, we find such tasks present great challenges for agents. First, the native JSON representation of workflows is inadequate for LLM-based agents to fully comprehend the intrinsic logic and dependencies within the systems [65]. Second, properly organizing the pipeline and utilizing AI-related modules require substantial expert knowledge and complex cognitive processes [21]. Traditional methods, such as Chain-of-Thought (CoT) [57] or Retrieval-Augmented Generation (RAG) [29], suffer from context limitation and hallucination problems, thus failing to generate complex workflows reliably." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.69, + 0.483, + 0.901 + ], + "angle": 0, + "content": "To tackle the aforementioned problems, we introduce ComfyAgent, a novel framework shown in Figure 1, which empowers LLM-based agents to autonomously design collaborative AI systems by generating workflows. The core concepts behind ComfyAgent are two-fold. First, it represents workflows with code, which can be reversibly converted into workflows and executed as collaborative systems by the interpreter. Second, it constructs a multi-agent system [19] which cooperates to learn from existing workflows and generate new workflows for a given task. The core of the multi-agent system is PlanAgent, responsible for the global planning of workflows by task instruction. After collecting document descriptions of all components and code examples of collected workflows, RetrievalAgent au" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.483, + 0.907, + 0.558 + ], + "angle": 0, + "content": "tomatically retrieves and learns key information about the given task from code. CombineAgent and AdaptAgent are responsible for code modification to construct a workflow that can fulfill the specified task. Upon task completion, PlanAgent finishes the task to form the final workflow." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.56, + 0.909, + 0.697 + ], + "angle": 0, + "content": "We evaluate common baseline agents [9, 29, 53, 57], as well as our proposed ComfyAgent, on ComfyBench. The experimental results indicate that ComfyAgent achieves comparable pass rate and resolve rate to o1-preview [39] and significantly surpasses other agents on ComfyBench. However, ComfyAgent has resolved only \\(15\\%\\) of creative tasks, which is still far from satisfactory. We believe that LLM-based agents still have a long way to go for autonomously designing collaborative AI systems." + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.7, + 0.859, + 0.713 + ], + "angle": 0, + "content": "Our contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.717, + 0.905, + 0.776 + ], + "angle": 0, + "content": "- We introduce ComfyBench, the first-of-its-kind comprehensive benchmark for the development of agents capable of designing and executing a wide range of collaborative AI systems in ComfyUI." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.778, + 0.905, + 0.837 + ], + "angle": 0, + "content": "- We propose ComfyAgent, which represents workflows with code and constructs a novel multi-agent framework to empower LLM-based agents in autonomously designing collaborative AI systems." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.838, + 0.905, + 0.898 + ], + "angle": 0, + "content": "- Experimental results demonstrate that with our intricate multi-agent design, ComfyAgent achieves a comparable pass rate and resolve rate to o1-preview and significantly surpasses other agents on ComfyBench." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.717, + 0.905, + 0.898 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24615" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.09, + 0.234, + 0.106 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.115, + 0.276, + 0.132 + ], + "angle": 0, + "content": "2.1. LLM-based Agents" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.139, + 0.483, + 0.41 + ], + "angle": 0, + "content": "LLM-based agents leverage LLMs to interact with external tools and solve real-world problems [52, 60]. Agents can exploit the powerful capabilities of LLMs in cognition and creativity, adapting to various novel tasks without relying on human intervention [30, 34]. Besides, agents have access to a wide range of external tools, making up for the natural limitation of processing only natural language [12, 77]. Researchers have been devoted to enhancing the capabilities of LLM-based agents. Many prompt-based methods [7, 33, 47, 53, 54, 57, 70] are proven effective for improving agents' planning and reasoning. Memory mechanism [78, 79] and RAG [16, 29] further extend agents' working context and knowledge base. Reinforcement learning [6, 23, 41], serving as a post-training paradigm, allowing even stronger enhancement and alignment. These various methods hasten the advent of advanced theories for agents, such as self-evolution [61, 64, 82] and unified architecture [48, 81]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.412, + 0.483, + 0.534 + ], + "angle": 0, + "content": "Based on well-designed module architectures and interaction mechanisms, LLM-based agents can be applied to solve a wide range of complex tasks, such as web navigation [14, 27, 35, 69, 80], interface operation [10, 59, 62, 63, 68], code generation [55, 56, 66, 72], and other tasks in specified domains [17, 22, 46, 50, 67]. These tasks, previously considered to require expert knowledge and human intervention, can now be accomplished by agents reliably." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.545, + 0.325, + 0.56 + ], + "angle": 0, + "content": "2.2. Collaborative AI Systems" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.568, + 0.483, + 0.734 + ], + "angle": 0, + "content": "Zaharia et al. [74] points out that state-of-the-art results are increasingly obtained by collaborative AI systems instead of monolithic models, which has been verified by many previous works [2, 29, 36, 49, 51]. Yuan et al. [73] even designs a multi-agent framework that incorporates several advanced visual agents to replicate generalist video generation demonstrated by Sora [8]. The advantages of collaborative AI systems are prominent. Dynamic systems make it possible to fetch external knowledge with components such as retrievers. Besides, extra modules such as verifier make the results controllable and trustworthy." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Compared to monolithic models, collaborative AI systems have posed new challenges in design and optimization. Novel paradigms are emerging to tackle these challenges, of which the most popular is composition frameworks, which build applications out of calls to models and other components [11, 20, 28, 32, 58]. DSPy [24] automatically optimizes the prompt instructions, few-shot examples and other parameters in the pipeline to maximize the end-to-end performance. GPTSwarm [83] supports building agents in the form of graphs and optimizing the parameters and connections as an entire system." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.09, + 0.649, + 0.107 + ], + "angle": 0, + "content": "3. ComfyBench" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.115, + 0.907, + 0.206 + ], + "angle": 0, + "content": "We propose ComfyBench, a new benchmark for evaluating the capability of LLM-based agents to autonomously design collaborative AI systems within the ComfyUI [13] platform. Given the task instruction, ComfyBench requires agents to generate workflows to describe collaborative AI systems in ComfyUI, producing the desired results after execution." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.214, + 0.696, + 0.23 + ], + "angle": 0, + "content": "3.1. ComfyUI Platform" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.236, + 0.907, + 0.417 + ], + "angle": 0, + "content": "Overview. ComfyUI is an open-source platform for designing and executing generative pipelines, where users can design their workflows by creating and connecting a series of nodes. Each node stands for a module, which can be a model, a tool, or a widget. On the user side, the workflows are visualized as directed acyclic graphs (DAGs), so that users can operate them intuitively. On the server side, however, the workflows are saved in a JSON object, which describes the nodes and connectivity in a structured manner. ComfyUI executes the JSON format workflows at the granularity of nodes, automatically loading models, conducting inference, passing variables, and returning results to users." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.418, + 0.909, + 0.691 + ], + "angle": 0, + "content": "Features. ComfyUI provides native support for common models and tools related to Stable Diffusion [44] [44], including but not limited to VAEs [25], CLIPs [42], LoRAs [45], IP-Adapters [71], and ControlNets [75]. With the joint efforts of the open-source community, hundreds of ComfyUI extensions are developed to support various latest models and tools, such as InsightFace [18], IC-Light [76], and Segment Anything [26], which bring about immense potential for realizing flexible and controllable generation tasks. By designing intricate workflows, users can produce fantastic artwork that cannot be achieved by common tools. Limitations. Despite the fancy effects that ComfyUI can achieve, the workflows are always manually crafted, which can be quite expensive. On the one hand, expert-level skills are required to figure out the appropriate nodes and parameters that can produce the desired results. On the other hand, workflows are designed to solve a specific task, while creating new workflows can be laborious and time-consuming." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.698, + 0.715, + 0.713 + ], + "angle": 0, + "content": "3.2. Benchmark Contents" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.907, + 0.901 + ], + "angle": 0, + "content": "ComfyBench collects annotations for 3205 nodes and 20 workflows and construct 200 tasks as is shown in Figure 2. More examples can be found in the supplementary material. Node document provides complete documentation for the 3205 nodes registered in ComfyUI until September 2024, serving as an important manual for agents to utilize the nodes properly. The source code is cloned from Salt AI Documentation [1], which is then cleaned and arranged into documentation, describing the usage of each node and their inputs and outputs. This structured documentation serves as a solid reference resource for understanding the complex node ecosystem within the ComfyUI framework." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.521, + 0.957 + ], + "angle": 0, + "content": "24616" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.089, + 0.307, + 0.325 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.312, + 0.089, + 0.482, + 0.325 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.338, + 0.483, + 0.381 + ], + "angle": 0, + "content": "Figure 2. ComfyBench provides annotations for 3205 nodes and 20 workflows, together with 200 task instructions categorized into three difficulty levels: vanilla, complex, and creative." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.396, + 0.483, + 0.501 + ], + "angle": 0, + "content": "Curriculum workflow is a set of 20 tutorial workflows in ComfyUI, which introduces some common nodes and routines to build workflows for solving basic tasks. All the workflows are verified feasible with common nodes and models in ComfyUI and correctly annotated with their purposes and principles. The curriculum workflow allows agents to master useful skills in a human-like manner." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.502, + 0.484, + 0.593 + ], + "angle": 0, + "content": "Task instruction serves as the core part of the benchmark, containing 200 task instructions. Each instruction defines a specific task, which should be solved by designing and executing a ComfyUI workflow. To reflect the generalization ability of agents, the tasks are manually categorized into three difficulty levels: vanilla, complex, and creative." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.594, + 0.483, + 0.639 + ], + "angle": 0, + "content": "- Vanilla includes 100 tasks, which can be solved by learning from a single curriculum workflow, involving minor modifications on several nodes and their parameters." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.64, + 0.483, + 0.699 + ], + "angle": 0, + "content": "- Complex includes 60 tasks, which necessitates learning from more than one curriculum workflow, involving functional combinations of multiple pipelines and minor adjustments on nodes and parameters." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.7, + 0.483, + 0.76 + ], + "angle": 0, + "content": "- Creative includes 40 tasks, which cannot be solved by directly imitating one or more curriculum workflows, requiring agents to understand the underlying principles and apply the learned skills in a novel way." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.594, + 0.483, + 0.76 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.761, + 0.484, + 0.806 + ], + "angle": 0, + "content": "Besides, we provide an auxiliary dataset as the input for some of the tasks. If an image or video is involved in the task, it will be specified in the instruction." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.818, + 0.275, + 0.833 + ], + "angle": 0, + "content": "3.3. Evaluation Metrics" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Different from the common benchmarks for image or video generation, we are evaluating the generated workflows, so traditional metrics are not applicable. We divide the evaluation into two progressive stages. The first stage is to check" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.168 + ], + "angle": 0, + "content": "whether the generated workflows are syntactically and semantically correct, and the second stage is to check whether the generated workflows can produce the expected results in the task instructions. We design two metrics: pass rate and resolve rate, corresponding to the two stages respectively." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.168, + 0.905, + 0.258 + ], + "angle": 0, + "content": "Pass rate measures the ratio of tasks where generated workflows are syntactically and semantically correct. In practice, each workflow is sent to the server of ComfyUI, which automatically conducts validation and execution. A task will be marked as passed only if the server finishes the execution process and returns a success message." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.259, + 0.907, + 0.485 + ], + "angle": 0, + "content": "Resolve rate measures the ratio of tasks where generated workflows can produce the expected results in the task instruction. In other words, we need to automatically determine the consistency between the produced images or videos and the task requirements. In practice, we apply Visual Language Models (VLMs), where different prompt templates are written according to the input and output modalities. Specifically, we adopt GPT-4o for evaluation, as it demonstrates strong capabilities for understanding language and analyzing visual content. For images, we directly encode them into the prompt. For videos, we uniformly sample up to 10 frames before encoding, which compresses the context length within a reasonable range. A task will be marked as resolved only if the GPT-4o model confirms that all the requirements are satisfied." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.496, + 0.696, + 0.511 + ], + "angle": 0, + "content": "3.4. Human Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.518, + 0.906, + 0.609 + ], + "angle": 0, + "content": "The computation of the resolve rate involves using VLMs to determine the consistency between the produced images or videos and the task requirements, which unavoidably involves subjectivity. To verify the reliability of VLM-based evaluation, we conduct a human evaluation to analyze the agreement between human evaluators and GPT-4o." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.61, + 0.906, + 0.762 + ], + "angle": 0, + "content": "We randomly select 70 tasks that are either completed by ComfyAgent proposed in Section 4 or by a human expert. Human evaluators and the latest GPT-4o model are asked to provide their judgments respectively. We compute the average scores for each question under a sample size of 560 and conduct a correlation analysis. Results are presented in Table 1, where the statistics indicate a strong agreement between human evaluators and GPT-4o, which verifies the reliability of our proposed VLM-based evaluation. More details can be found in the supplementary material." + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.775, + 0.905, + 0.803 + ], + "angle": 0, + "content": "Table 1. Correlation analysis between the average scores given by human evaluators and GPT-4o indicates a strong agreement." + }, + { + "type": "table", + "bbox": [ + 0.558, + 0.814, + 0.864, + 0.892 + ], + "angle": 0, + "content": "
MetricStatisticP-value
Kendall's τ0.6455.427 × 10-11
Pearson's r0.7641.482 × 10-14
Spearman's ρ0.7373.672 × 10-13
" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24617" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.089, + 0.907, + 0.257 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.267, + 0.908, + 0.337 + ], + "angle": 0, + "content": "Figure 3. The architecture of the ComfyAgent framework. Multiple agents cooperate to design workflows in a step-by-step manner. Given the task instruction, the planner initializes the memory and produces a plan. For each step, the planner updates the plan and forms an action. Different actions, including combine, adapt, and retrieve, are then handled by corresponding agents. After combination or adaptation, refine action will be conducted to ensure the correctness. All the agents can interact with the memory, which consists of history, reference, and workspace. Once the task is deemed completed, the planner will finish the procedure and save the workflow." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.362, + 0.226, + 0.379 + ], + "angle": 0, + "content": "4. ComfyAgent" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.387, + 0.325, + 0.403 + ], + "angle": 0, + "content": "4.1. Workflow Representation" + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.423, + 0.481, + 0.662 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.674, + 0.483, + 0.703 + ], + "angle": 0, + "content": "Figure 4. Examples of four common formats to represent workflows: flow graph, JSON, element list, and code." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Collaborative AI systems coordinate multiple interacting components, such as AI models and external tools, using workflows. These workflows are essentially high-level diagrams that outline the operational logic. They can be naturally represented as directed acyclic graphs (DAGs) consisting of two main elements: vertices and directed edges. Vertices represent information processing components with input and output interfaces, while directed edges represent information flow between components. Specifically, the incoming edges of a vertex represent the external information it receives, while the outgoing edges signify the information it outputs, with multiple edges illustrating various" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.363, + 0.905, + 0.409 + ], + "angle": 0, + "content": "information paths. This abstraction allows efficient construction and modification of collaborative AI systems, as demonstrated by platforms like ComfyUI [13]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.41, + 0.906, + 0.576 + ], + "angle": 0, + "content": "There are four common formats to represent workflows: flow graph, JSON, element list, and code, as shown in Figure 4. Flow graph [13] provides intuitive visualization for humans but is unsuitable for LLM or VLM processing. JSON, typically used in [77], provides a structured representation but is unwieldy for complex workflows due to redundant information and LLMs' context limitations. Element list, typically used in [31], is closer to natural language and provides a more compact representation, but it lacks explicit topological relationships, hindering LLMs from correctly processing complex workflows." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.578, + 0.906, + 0.713 + ], + "angle": 0, + "content": "In contrast, code emerges as the most effective representation, offering various advantages including Turing completeness, rich semantic information, and natural compatibility with LLMs' code generation capabilities. We implement code representation using a restricted subset of Python-like syntax, involving basic operations and control structures while excluding advanced features. We also conduct an ablation study in Section 5.4 to demonstrate the superiority of code representation over other representations." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.726, + 0.741, + 0.741 + ], + "angle": 0, + "content": "4.2. Multi-Agent Framework" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.906, + 0.84 + ], + "angle": 0, + "content": "Through code representation, we can effectively prompt LLMs to generate workflows for the expected collaborative AI systems. However, with the complexity of workflows increasing, traditional agents suffer from context limitation and hallucination problems, failing to generate workflows correctly and stably, which is proved in Section 5.3." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.841, + 0.906, + 0.901 + ], + "angle": 0, + "content": "To achieve reliable generation of ComfyUI workflows, we propose ComfyAgent, a multi-agent framework to design workflows in a step-by-step manner. As demonstrated in Figure 3, ComfyAgent consists of three independent" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "24618" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.908, + 0.119 + ], + "angle": 0, + "content": "Table 2. Evaluation results of all the agents on ComfyBench. The average task-wise token consumption, the pass rate and resolve rate of every task category denoted by %P and %R, and the overall result on ComfyBench, are reported. The best results are highlighted in bold." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.13, + 0.907, + 0.356 + ], + "angle": 0, + "content": "
AgentCostVanillaComplexCreativeTotal
#Tokens (k)%P%R%P%R%P%R%P%R
GPT-4o + Zero-shot0.500.00.00.00.00.00.00.00.0
GPT-4o + Few-shot2.5532.027.016.78.37.50.022.516.0
GPT-4o + CoT2.5044.029.011.78.312.50.028.017.0
GPT-4o + CoT-SC11.0345.034.011.75.015.00.029.018.5
Claude-3.5-Sonnet + RAG5.7327.013.023.06.77.50.022.08.5
Llama-3.1-70B + RAG5.0458.032.023.010.015.05.039.020.0
GPT-4o + RAG4.9762.041.045.021.740.07.552.023.0
o1-mini + RAG6.0532.016.021.78.312.57.525.012.0
o1-preview + RAG9.4670.046.048.323.330.012.555.532.5
Llama-3.1-70B + ComfyAgent24.2263.035.026.718.320.05.043.524.0
GPT-4o + ComfyAgent14.9667.046.048.321.740.015.056.032.5
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.38, + 0.365, + 0.394 + ], + "angle": 0, + "content": "modules: Memory, Planner, and Actions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.396, + 0.483, + 0.488 + ], + "angle": 0, + "content": "Memory stores the recent state of ComfyAgent, which is built on psychological theories [4]. Specifically, it refers to the working memory [5], reflecting the current circumstances, namely the recent history behaviors, results from intermediate steps, internal reasoning, and external knowledge. We formulate them into three parts." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.488, + 0.483, + 0.518 + ], + "angle": 0, + "content": "- History maintains recent plans and actions of Planner, enabling action review for subsequent planning." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.518, + 0.483, + 0.548 + ], + "angle": 0, + "content": "- Reference stores information retrieved from the knowledge base, and can be updated through active retrieval." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.549, + 0.483, + 0.594 + ], + "angle": 0, + "content": "- **Workspace contains the current workflow together with its natural language annotation, which maintains a complete record of the current working status.**" + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.488, + 0.483, + 0.594 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.595, + 0.483, + 0.64 + ], + "angle": 0, + "content": "Planner, which can be specified as PlanAgent, serves as the core of ComfyAgent, providing the global scheme to design and modify workflows under the task instruction." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.642, + 0.483, + 0.687 + ], + "angle": 0, + "content": "- At the beginning of the task, PlanAgent selects an existing workflow to initialize the memory and produces a thorough multi-step plan based on the task instruction." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.687, + 0.483, + 0.732 + ], + "angle": 0, + "content": "- For each step, PlanAgent produces a high-level plan, together with an action based on the current memory, which will then be processed by other agents." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.733, + 0.483, + 0.777 + ], + "angle": 0, + "content": "- For each step, PlanAgent evaluates the completion status of the task. Once the task is deemed completed, PlanAgent will finish the procedure and save the workflow." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.642, + 0.483, + 0.777 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.779, + 0.483, + 0.824 + ], + "angle": 0, + "content": "Actions represent different actions that can be selected by PlanAgent. All the actions are meant to modify the current memory. In ComfyAgent, we have three actions." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.826, + 0.483, + 0.871 + ], + "angle": 0, + "content": "- Combine is processed by CombineAgent, which aims to combine the current workflow with another workflow from references specified by PlanAgent." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.483, + 0.902 + ], + "angle": 0, + "content": "- Adapt is processed by AdaptAgent, which is responsible for adapting the details (e.g. parameters) of the current" + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.826, + 0.483, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.527, + 0.38, + 0.868, + 0.395 + ], + "angle": 0, + "content": "workflow based on the prompt given by PlanAgent." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.395, + 0.905, + 0.44 + ], + "angle": 0, + "content": "- Retrieve is processed by RetrieveAgent, which follows the prompt given by PlanAgent to retrieve relevant information and update references accordingly." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.441, + 0.905, + 0.486 + ], + "angle": 0, + "content": "After combination or adaptation, the updated workflow will be checked and refined by RefineAgent before entering the workspace, which avoids the accumulation of errors." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.487, + 0.906, + 0.578 + ], + "angle": 0, + "content": "After the action is processed, ComfyAgent will enter a new step, where PlanAgent updates the existing plan and forms a new action. In such a way, ComfyAgent gradually generates the workflow. When the procedure is finished, the code representation will be converted into the standard format, which exactly describes the collaborative AI system." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.594, + 0.646, + 0.612 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.62, + 0.671, + 0.636 + ], + "angle": 0, + "content": "5.1. Baseline Agents" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.643, + 0.906, + 0.719 + ], + "angle": 0, + "content": "ComfyBench places strong requirements on the learning and generalization ability of agents, necessitating LLMs with powerful reasoning capabilities and sufficient context length. Therefore, we conduct experiments on the advanced models, involving both commercial and open-source LLMs." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.72, + 0.905, + 0.765 + ], + "angle": 0, + "content": "- Llama-3.1 (llama-3.1-70b-instruct) [15] is an LLM developed by Meta, achieving state-of-the-art performance in the open-source community." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.765, + 0.905, + 0.811 + ], + "angle": 0, + "content": "- Claude-3.5 (claude-3.5-sonnet-20240620) [3] is a commercial LLM developed by Anthropic, with a reputation for its creativity and interaction quality." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.811, + 0.905, + 0.87 + ], + "angle": 0, + "content": "- GPT-4o (gpt-4o-2024-08-06) [38] is currently the mainstream model provided by OpenAI and most widely used by the community. It still maintains state-of-the-art performance in leaderboards of various domains." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.905, + 0.902 + ], + "angle": 0, + "content": "- o1-mini (o1-mini-2024-09-12) and o1-preview (o1-preview-2024-09-12) [39] are the latest mod" + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.72, + 0.905, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24619" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.104, + 0.092, + 0.483, + 0.122 + ], + "angle": 0, + "content": "els released by OpenAI, which focus on complex reasoning tasks such as mathematics and programming." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.123, + 0.483, + 0.168 + ], + "angle": 0, + "content": "We adopt five methods that have proved to be universally effective by the community and can be conveniently adapted to solve the tasks in ComfyBench." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.169, + 0.483, + 0.198 + ], + "angle": 0, + "content": "- Zero-shot Learning directly feeds LLMs with the task instruction to conduct inference." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.2, + 0.483, + 0.245 + ], + "angle": 0, + "content": "- Few-shot Learning [9] provides a set of demonstrations in the prompt, which utilizes the in-context learning ability of LLMs and facilitates the correct generation." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.245, + 0.483, + 0.29 + ], + "angle": 0, + "content": "- Chain-of-Thought (CoT) [57] is improved based on few-shot learning, which instructs the agent to articulate the reasoning process before providing the final answer." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.29, + 0.483, + 0.334 + ], + "angle": 0, + "content": "- CoT with Self-consistency (CoT-SC) [53] is an improved version of CoT, which ensembles parallel trajectories and then selects the most consistent answer." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.335, + 0.483, + 0.38 + ], + "angle": 0, + "content": "- Retrieval-Augmented Generation (RAG) [29] learns from the retrieved demonstrations, which efficiently utilizes the most relevant information." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.169, + 0.483, + 0.38 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.382, + 0.484, + 0.457 + ], + "angle": 0, + "content": "Considering the limited experimental budget, we evaluate the methods mainly on GPT-4o. ComfyAgent is also evaluated on Llama-3.1 to verify its performance on the open-source model. RAG is evaluated on all the models to provide a horizontal comparison of their capabilities." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.469, + 0.308, + 0.485 + ], + "angle": 0, + "content": "5.2. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.492, + 0.483, + 0.719 + ], + "angle": 0, + "content": "ComfyUI Settings. The experiments are conducted on a fixed version of ComfyUI with necessary extensions. A set of common models is installed to make the workflows runnable and the results reproducible. In ComfyUI, the workflows are natively represented in JSON format. We implement a parser to extract the entire DAG described in the JSON object, which is then converted into a sequence of Python-like function calls in topological order. Another parser is implemented to restore this conversion. This ensures that workflows can be effectively generated by agents through code representation while remaining fully compatible with the ComfyUI execution environment. Note that the node documents are not fed into the agents due to context limitations, even though this technique would potentially improve their performance." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.72, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Agent Parameters. We follow CodeAct [54] to format the interaction with LLMs. The parameters of the agents are manually set to exploit their largest potential within a limited context length. For few-shot learning, CoT, and CoT-SC, the number of demonstrations is consistently set to 3, but CoT-SC will ensemble 3 trajectories in parallel to form the final answer. For RAG and ComfyAgent, we adopt an advanced embedding model (text-embedding-3-large) [40] released by OpenAI to encode the texts for retrieval, and the number of retrieved demonstrations is set to 5. For ComfyAgent, the modules are implemented as different roles played by the" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.09, + 0.907, + 0.133 + ], + "angle": 0, + "content": "Table 3. Ablation results of RAG on GPT-4o with different workflow representations on ComfyBench. The pass rate and resolve rate are reported, denoted by %P and %R respectively." + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.143, + 0.905, + 0.248 + ], + "angle": 0, + "content": "
RepresentationVanillaComplexCreative
%P%R%P%R%P%R
Flow Graph------
JSON52.033.025.08.317.57.5
Element List51.030.023.35.020.010.0
Code62.041.045.015.030.07.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.263, + 0.907, + 0.32 + ], + "angle": 0, + "content": "Table 4. Ablation results of ComfyAgent on GPT-4o with different architectures on ComfyBench. Each variant is implemented by removing one agent from the multi-agent framework. The pass rate and resolve rate are reported, denoted by %P and %R respectively." + }, + { + "type": "table", + "bbox": [ + 0.524, + 0.33, + 0.899, + 0.45 + ], + "angle": 0, + "content": "
ArchitectureVanillaComplexCreative
%P%R%P%R%P%R
ComfyAgent67.046.048.321.740.015.0
w/o Combine69.043.015.05.035.015.0
w/o Adapt52.026.020.010.017.55.0
w/o Retrieve27.016.020.05.07.57.5
w/o Refine65.042.043.321.735.012.5
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.475, + 0.907, + 0.507 + ], + "angle": 0, + "content": "same LLM. We provide the prompt implementation and an example trajectory in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.516, + 0.695, + 0.53 + ], + "angle": 0, + "content": "5.3. Evaluation Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.538, + 0.907, + 0.629 + ], + "angle": 0, + "content": "We report the evaluation results of different agents on ComfyBench in Table 2. All the agents follow the parameter settings in Section 5.2 and adopt code representation in Section 4.1. We also present some examples generated by ComfyAgent on ComfyBench in Table 5. More examples can be found in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.63, + 0.906, + 0.825 + ], + "angle": 0, + "content": "In terms of methods, zero-shot learning fails due to a lack of domain knowledge. CoT achieves a better performance than few-shot learning, and CoT-SC shows a slight improvement over CoT. However, these agents fail to achieve comparable performance with retrieval-based agents, because they cannot fully exploit the external knowledge provided by ComfyBench. ComfyAgent outperforms all the baselines in every task category, achieving an overall pass rate of \\(56\\%\\) and resolve rate of \\(32.5\\%\\), which indicates that the multi-agent framework applied in ComfyAgent is capable of alleviating the hallucination problem and improving the generation quality and stability. This improvement is consistent on both open-source and commercial models." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.907, + 0.9 + ], + "angle": 0, + "content": "In terms of models, Claude-3.5 and o1-mini show poor performance under RAG, which is mainly caused by the hallucination problem, where models modify the nodes or parameters without verifying their validity. In contrast, Llama-3.1 and GPT-4o show less hallucination and achieve" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24620" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.908, + 0.119 + ], + "angle": 0, + "content": "Table 5. Some examples generated by ComfyAgent on ComfyBench. Given the task instruction, ComfyAgent designs a workflow to describe the collaborative AI system in ComfyUI, which is then executed to generate the result required by the instruction." + }, + { + "type": "title", + "bbox": [ + 0.177, + 0.136, + 0.284, + 0.148 + ], + "angle": 0, + "content": "Task Instruction" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.154, + 0.351, + 0.209 + ], + "angle": 0, + "content": "You are given an image of a male celebrity. Transform the man in the image into a beautiful woman with ponytail hair while preserving her facial identity." + }, + { + "type": "image_caption", + "bbox": [ + 0.145, + 0.216, + 0.184, + 0.23 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image", + "bbox": [ + 0.104, + 0.232, + 0.227, + 0.328 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.273, + 0.216, + 0.315, + 0.229 + ], + "angle": 0, + "content": "Result" + }, + { + "type": "image", + "bbox": [ + 0.231, + 0.232, + 0.357, + 0.328 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.446, + 0.136, + 0.553, + 0.148 + ], + "angle": 0, + "content": "Task Instruction" + }, + { + "type": "text", + "bbox": [ + 0.371, + 0.154, + 0.62, + 0.209 + ], + "angle": 0, + "content": "You are given an image of a standing cat. Replace the background with a scene of a cozy living room while keeping the lighting and shadows consistent." + }, + { + "type": "image_caption", + "bbox": [ + 0.415, + 0.216, + 0.453, + 0.23 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image", + "bbox": [ + 0.374, + 0.232, + 0.498, + 0.328 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.542, + 0.216, + 0.585, + 0.229 + ], + "angle": 0, + "content": "Result" + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.232, + 0.626, + 0.328 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.716, + 0.136, + 0.822, + 0.148 + ], + "angle": 0, + "content": "Task Instruction" + }, + { + "type": "text", + "bbox": [ + 0.641, + 0.154, + 0.89, + 0.209 + ], + "angle": 0, + "content": "You are given an image of a red bridge with a person standing on it. Remove the person from the image while maintaining the original appearance of the bridge." + }, + { + "type": "image_caption", + "bbox": [ + 0.684, + 0.216, + 0.722, + 0.23 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image", + "bbox": [ + 0.643, + 0.232, + 0.767, + 0.328 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.811, + 0.216, + 0.855, + 0.23 + ], + "angle": 0, + "content": "Result" + }, + { + "type": "image", + "bbox": [ + 0.771, + 0.232, + 0.895, + 0.328 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.361, + 0.483, + 0.392 + ], + "angle": 0, + "content": "better performance, and o1-preview achieves an outstanding performance comparable to ComfyAgent on GPT-4o." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.407, + 0.255, + 0.421 + ], + "angle": 0, + "content": "5.4. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.431, + 0.484, + 0.627 + ], + "angle": 0, + "content": "Representation. We conduct an ablation study to verify the advantages of representing workflows with code introduced in Section 4.1. We implement three variants of RAG on GPT-4o, where the workflows are respectively represented in JSON, element list, and code. The variants are evaluated on ComfyBench with the same parameter settings as in Section 5.3. Flow graph is not implemented because it is not understandable by LLM-based agents. We report the ablation results in Table 3, where RAG with code representation shows superior performance over other representations, which verifies our claim. RAG with JSON representation is expected to perform even worse without the unique feature of GPT-4o to generate structured JSON objects." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.63, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Architecture. We conduct another ablation to demonstrate the rationality of the multi-agent framework introduced in Section 4.2, where four variants of ComfyAgent on GPT-4o are implemented, each removing one agent from the original framework. We report the ablation results in Table 4. We find that removing any of the four agents will lead to a performance drop. Removing CombineAgent alleviates hallucination in the vanilla category, but leads to failure in the complex category, demonstrating the necessity of integrating multiple demonstrations. Removing AdaptAgent or RetrieveAgent leads to a significant performance drop in all the categories, which underscores the critical role of detail modification and knowledge utilization. Removing RefineAgent leads to a minor but noticeable performance drop in all the categories, indicating its effectiveness in mitigating error accumulation. Therefore, we can conclude that our meticulously-designed multi-agent framework contributes to the superior performance of ComfyAgent." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.36, + 0.634, + 0.376 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.387, + 0.907, + 0.735 + ], + "angle": 0, + "content": "In this work, we investigate the potential of LLM-based agents to autonomously design collaborative AI systems. Building upon the open-source ComfyUI [13] platform, we present ComfyBench, an autonomous workflow design environment serving as both a benchmark for evaluating the capabilities of agents in designing collaborative systems and a means to extend the applicability of ComfyUI in real-world, language-instruction-based generative applications. We further develop ComfyAgent, which achieves a comparable resolve rate to o1-preview [39] and significantly surpasses other agents on ComfyBench. We point out two key insights underpinning ComfyAgent. First, representing workflows through code is efficient, which provides rich semantics for agents to better understand complex systems. Second, the multi-agent framework of ComfyAgent highlights the substantial benefits of task decomposition and collaborative generation. However, the top-performing agent, ComfyAgent, has achieved a resolve rate of only \\(15\\%\\) for creative tasks on ComfyBench. LLM-based agents still have a long way to go in designing collaborative AI systems. We hope that ComfyBench and ComfyAgent can serve as valuable resources in advancing the development of intelligent and autonomous collaborative AI systems." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.736, + 0.907, + 0.857 + ], + "angle": 0, + "content": "Limitations. Well-designed fine-tuning methods could potentially elevate agents to an expert level on ComfyBench, but they are quite expensive and thus not involved in our work. In addition, ComfyAgent operates in an open-loop manner. An alternative approach would involve using VLMs as evaluators to provide feedback [21], enabling a closed-loop system where agents can refine their workflows based on evaluations, thus leading to better results." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.858, + 0.907, + 0.888 + ], + "angle": 0, + "content": "Acknowledgments. This work is supported by the Shanghai Municipal Science and Technology Major Project." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "24621" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.091, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.116, + 0.48, + 0.142 + ], + "angle": 0, + "content": "[1] Salt AI. Salt node documentation. https://docs. getsalt.ai/md/, 2024.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.145, + 0.483, + 0.199 + ], + "angle": 0, + "content": "[2] Google DeepMind AlphaCode Team. Alphacode 2 technical report. https://storage.googleapis.com/deepmind-media/AlphaCode2/AlphaCode2_Tech_Report.pdf, 2024.1,3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.201, + 0.482, + 0.24 + ], + "angle": 0, + "content": "[3] Anthropic. Claude 3.5 sonnet. https://www.anthropic.com/news/claude-3-5-sonnet, 2024.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.243, + 0.483, + 0.282 + ], + "angle": 0, + "content": "[4] Richard C Atkinson. Human memory: A proposed system and its control processes. The psychology of learning and motivation, 1968. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.285, + 0.433, + 0.298 + ], + "angle": 0, + "content": "[5] Alan Baddeley. Working memory. Science, 1992. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.3, + 0.482, + 0.368 + ], + "angle": 0, + "content": "[6] Yuntao Bai, Andy Jones, Kamal Ndousse, Amanda Askell, Anna Chen, Nova DasSarma, Dawn Drain, Stanislav Fort, Deep Ganguli, Tom Henighan, et al. Training a helpful and harmless assistant with reinforcement learning from human feedback. arXiv preprint arXiv:2204.05862, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.37, + 0.482, + 0.437 + ], + "angle": 0, + "content": "[7] Maciej Besta, Nils Blach, Ales Kubicek, Robert Gerstenberger, Michal Podstawski, Lukas Gianinazzi, Joanna Gajda, Tomasz Lehmann, Hubert Niewiadomski, Piotr Nczyk, et al. Graph of thoughts: Solving elaborate problems with large language models. In AAAI, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.44, + 0.482, + 0.534 + ], + "angle": 0, + "content": "[8] Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Li Jing, David Schnurr, Joe Taylor, Troy Luhman, Eric Luhman, Clarence Ng, Ricky Wang, and Aditya Ramesh. Video generation models as world simulators. https://openai.com/research/video-generation-models-as-world-simulators, 2024.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.537, + 0.482, + 0.592 + ], + "angle": 0, + "content": "[9] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. In NeurIPS, 2020. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.594, + 0.482, + 0.674 + ], + "angle": 0, + "content": "[10] Ruisheng Cao, Fangyu Lei, Haoyuan Wu, Jixuan Chen, Yeqiao Fu, Hongcheng Gao, Xinzhuang Xiong, Hanchong Zhang, Yuchen Mao, Wenjing Hu, et al. Spider2-v: How far are multimodal agents from automating data science and engineering workflows? arXiv preprint arXiv:2407.10956, 2024.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.677, + 0.482, + 0.746 + ], + "angle": 0, + "content": "[11] Weize Chen, Yusheng Su, Jingwei Zuo, Cheng Yang, Chenfei Yuan, Chen Qian, Chi-Min Chan, Yujia Qin, Yaxi Lu, Ruobing Xie, et al. Agentverse: Facilitating multi-agent collaboration and exploring emergent behaviors in agents. arXiv preprint arXiv:2308.10848, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.748, + 0.482, + 0.802 + ], + "angle": 0, + "content": "[12] Zehui Chen, Kuikun Liu, Qiuchen Wang, Jiangning Liu, Wenwei Zhang, Kai Chen, and Feng Zhao. Mindsearch: Mimicking human minds elicits deep ai searcher. arXiv preprint arXiv:2407.20183, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.804, + 0.482, + 0.83 + ], + "angle": 0, + "content": "[13] comfyanonymous. Comfyui. https://github.com/comfyanonymous/ComfyUI, 2023. 1, 3, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.832, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[14] Xiang Deng, Yu Gu, Boyuan Zheng, Shijie Chen, Sam Stevens, Boshi Wang, Huan Sun, and Yu Su. Mind2web: Towards a generalist agent for the web. In NeurIPS, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.874, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[15] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.116, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.094, + 0.905, + 0.133 + ], + "angle": 0, + "content": "Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.136, + 0.905, + 0.189 + ], + "angle": 0, + "content": "[16] Yunfan Gao, Yun Xiong, Xinyu Gao, Kangxiang Jia, Jinliu Pan, Yuxi Bi, Yi Dai, Jiawei Sun, and Haofen Wang. Retrieval-augmented generation for large language models: A survey. arXiv preprint arXiv:2312.10997, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.191, + 0.905, + 0.232 + ], + "angle": 0, + "content": "[17] Yingqiang Ge, Wenyue Hua, Kai Mei, Juntao Tan, Shuyuan Xu, Zelong Li, Yongfeng Zhang, et al. Openagi: When llm meets domain experts. In NeurlPS, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.234, + 0.905, + 0.287 + ], + "angle": 0, + "content": "[18] Jia Guo, Jiankang Deng, Xiang An, Jack Yu, and Baris Gecer. Insightface: 2d and 3d face analysis project. https://github.com/deepinsight/insightface, 2023.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.289, + 0.905, + 0.356 + ], + "angle": 0, + "content": "[19] Taicheng Guo, Xiuying Chen, Yaqi Wang, Ruidi Chang, Shichao Pei, Nitesh V Chawla, Olaf Wiest, and Xiangliang Zhang. Large language model based multi-agents: A survey of progress and challenges. arXiv preprint arXiv:2402.01680, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.358, + 0.905, + 0.426 + ], + "angle": 0, + "content": "[20] Sirui Hong, Xiawu Zheng, Jonathan Chen, Yuheng Cheng, Jinlin Wang, Ceyao Zhang, Zili Wang, Steven Ka Shing Yau, Zijuan Lin, Liyang Zhou, et al. Metagpt: Meta programming for multi-agent collaborative framework. arXiv preprint arXiv:2308.00352, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.428, + 0.905, + 0.467 + ], + "angle": 0, + "content": "[21] Shengran Hu, Cong Lu, and Jeff Clune. Automated design of agentic systems. arXiv preprint arXiv:2408.08435, 2024. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.47, + 0.904, + 0.51 + ], + "angle": 0, + "content": "[22] Ian Huang, Guandao Yang, and Leonidas Guibas. Blender alchemy: Editing 3d graphics with vision-language models. arXiv preprint arXiv:2404.17672, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.512, + 0.905, + 0.552 + ], + "angle": 0, + "content": "[23] Timo Kaufmann, Paul Weng, Viktor Bengs, and Eyke Hüllermeier. A survey of reinforcement learning from human feedback. arXiv preprint arXiv:2312.14925, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.554, + 0.905, + 0.634 + ], + "angle": 0, + "content": "[24] Omar Khattab, Arnav Singhvi, Paridhi Maheshwari, Zhiyuan Zhang, Keshav Santhanam, Sri Vardhamanan, Saiful Haq, Ashutosh Sharma, Thomas T Joshi, Hanna Moazam, et al. Dspy: Compiling declarative language model calls into self-improving pipelines. arXiv preprint arXiv:2310.03714, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.637, + 0.905, + 0.664 + ], + "angle": 0, + "content": "[25] Diederik P Kingma. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.665, + 0.905, + 0.733 + ], + "angle": 0, + "content": "[26] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dólar, and Ross Girshick. Segment anything. arXiv preprint arXiv:2304.02643, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.735, + 0.905, + 0.803 + ], + "angle": 0, + "content": "[27] Jing Yu Koh, Robert Lo, Lawrence Jang, Vikram Duvvur, Ming Chong Lim, Po-Yu Huang, Graham Neubig, Shuyan Zhou, Ruslan Salakhutdinov, and Daniel Fried. Visualwebarena: Evaluating multimodal agents on realistic visual web tasks. arXiv preprint arXiv:2401.13649, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.804, + 0.905, + 0.831 + ], + "angle": 0, + "content": "[28] LangChain. Applications that can reason. powered by langchain. https://www.langchain.com/, 2023.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.833, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[29] Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Kuttler, Mike Lewis, Wen-tau Yih, Tim Rocktäschel, et al. Retrieval-augmented generation for knowledge-intensive nlp tasks. In NeurIPS, 2020. 2, 3, 7" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.094, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24622" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[30] Guohao Li, Hasan Hammoud, Hani Itani, Dmitrii Khizbullin, and Bernard Ghanem. Camel: Communicative agents for\" mind\" exploration of large language model society. In NeurIPS, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.218 + ], + "angle": 0, + "content": "[31] Zelong Li, Shuyuan Xu, Kai Mei, Wenyue Hua, Balaji Rama, Om Raheja, Hao Wang, He Zhu, and Yongfeng Zhang. Autoflow: Automated workflow generation for large language model agents. arXiv preprint arXiv:2407.12821, 2024. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.22, + 0.482, + 0.288 + ], + "angle": 0, + "content": "[32] Tian Liang, Zhiwei He, Wenxiang Jiao, Xing Wang, Yan Wang, Rui Wang, Yujiu Yang, Shuming Shi, and Zhaopeng Tu. Encouraging divergent thinking in large language models through multi-agent debate. arXiv preprint arXiv:2305.19118, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.291, + 0.482, + 0.346 + ], + "angle": 0, + "content": "[33] Aman Madaan, Niket Tandon, Prakhar Gupta, Skyler Hallinan, Luyu Gao, Sarah Wiegreffe, Uri Alon, Nouha Dziri, Shrimai Prabhumoye, Yiming Yang, et al. Self-refine: Iterative refinement with self-feedback. In NeurIPS, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.348, + 0.482, + 0.389 + ], + "angle": 0, + "content": "[34] Kai Mei, Zelong Li, Shuyuan Xu, Ruosong Ye, Yingqiang Ge, and Yongfeng Zhang. LIm agent operating system. arXiv preprint arXiv:2403.16971, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.391, + 0.482, + 0.459 + ], + "angle": 0, + "content": "[35] Reiichiro Nakano, Jacob Hilton, Suchir Balaji, Jeff Wu, Long Ouyang, Christina Kim, Christopher Hesse, Shantanu Jain, Vineet Kosaraju, William Saunders, et al. Webgpt: Browser-assisted question-answering with human feedback. arXiv preprint arXiv:2112.09332, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.461, + 0.482, + 0.53 + ], + "angle": 0, + "content": "[36] Harsha Nori, Yin Tat Lee, Sheng Zhang, Dean Carignan, Richard Edgar, Nicolo Fusi, Nicholas King, Jonathan Larson, Yuanzhi Li, Weishung Liu, et al. Can generalist foundation models outcompete special-purpose tuning? case study in medicine. arXiv preprint arXiv:2311.16452, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.532, + 0.482, + 0.559 + ], + "angle": 0, + "content": "[37] OpenAI. Chatgpt. https://openai.com/chatgpt/overview/, 2024.1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.562, + 0.482, + 0.588 + ], + "angle": 0, + "content": "[38] OpenAI. Hello gpt-4o. https://openai.com/index/hello-gpt-4o/, 2024.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.591, + 0.482, + 0.631 + ], + "angle": 0, + "content": "[39] OpenAI. Introducing openai ol-preview. https://openai.com/index/introducing-openai-ol-preview/, 2024.2,6,8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.633, + 0.482, + 0.674 + ], + "angle": 0, + "content": "[40] OpenAI. New embedding models and api updates. https://openai.com/index/new-embedding-models-and-api-updates/, 2024.7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.676, + 0.482, + 0.744 + ], + "angle": 0, + "content": "[41] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. In NeurIPS, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.746, + 0.482, + 0.814 + ], + "angle": 0, + "content": "[42] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.817, + 0.482, + 0.858 + ], + "angle": 0, + "content": "[43] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In ICML, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.86, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[44] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.134 + ], + "angle": 0, + "content": "[45] Simo Ryu. Low-rank adaptation for fast text-to-image diffusion fine-tuning. https://github.com/cloneofsimo/lora, 2023.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[46] Yongliang Shen, Kaitao Song, Xu Tan, Dongsheng Li, Weiming Lu, and Yueting Zhuang. Hugginggpt: Solving ai tasks with chatgpt and its friends in hugging face. In NeurIPS, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.193, + 0.905, + 0.247 + ], + "angle": 0, + "content": "[47] Chan Hee Song, Jiaman Wu, Clayton Washington, Brian M Sadler, Wei-Lun Chao, and Yu Su. Llm-planner: Few-shot grounded planning for embodied agents with large language models. In CVPR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.249, + 0.905, + 0.291 + ], + "angle": 0, + "content": "[48] Theodore R Sumers, Shunyu Yao, Karthik Narasimhan, and Thomas L Griffiths. Cognitive architectures for language agents. arXiv preprint arXiv:2309.02427, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.293, + 0.905, + 0.36 + ], + "angle": 0, + "content": "[49] Gemini Team, Rohan Anil, Sebastian Borgeaud, Yonghui Wu, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, et al. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.363, + 0.905, + 0.404 + ], + "angle": 0, + "content": "[50] Amitayush Thakur, Yeming Wen, and Swarat Chaudhuri. A language-agent approach to formal theorem-proving. arXiv preprint arXiv:2310.04353, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.406, + 0.905, + 0.447 + ], + "angle": 0, + "content": "[51] Trieu H Trinh, Yuhuai Wu, Quoc V Le, He He, and Thang Luong. Solving olympiad geometry without human demonstrations. Nature, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.449, + 0.905, + 0.504 + ], + "angle": 0, + "content": "[52] Lei Wang, Chen Ma, Xueyang Feng, Zeyu Zhang, Hao Yang, Jingsen Zhang, Zhiyuan Chen, Jiakai Tang, Xu Chen, Yankai Lin, et al. A survey on large language model based autonomous agents. Frontiers of Computer Science, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.506, + 0.905, + 0.573 + ], + "angle": 0, + "content": "[53] Xuezhi Wang, Jason Wei, Dale Schuurmans, Quoc Le, Ed Chi, Sharan Narang, Aakanksha Chowdhery, and Denny Zhou. Self-consistency improves chain of thought reasoning in language models. arXiv preprint arXiv:2203.11171, 2022. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.576, + 0.905, + 0.63 + ], + "angle": 0, + "content": "[54] Xingyao Wang, Yangyi Chen, Lifan Yuan, Yizhe Zhang, Yunzhu Li, Hao Peng, and Heng Ji. Executable code actions elicit better lIm agents. arXiv preprint arXiv:2402.01030, 2024. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.633, + 0.905, + 0.701 + ], + "angle": 0, + "content": "[55] Xingyao Wang, Boxuan Li, Yufan Song, Frank F Xu, Xiangru Tang, Mingchen Zhuge, Jiayi Pan, Yueqi Song, Bowen Li, Jaskirat Singh, et al. Opendevin: An open platform for ai software developers as generalist agents. arXiv preprint arXiv:2407.16741, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.704, + 0.905, + 0.744 + ], + "angle": 0, + "content": "[56] Zhiruo Wang, Shuyan Zhou, Daniel Fried, and Graham Neubig. Execution-based evaluation for open-domain code generation. arXiv preprint arXiv:2212.10481, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.746, + 0.905, + 0.801 + ], + "angle": 0, + "content": "[57] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. In NeurIPS, 2022. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.803, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[58] Qingyun Wu, Gagan Bansal, Jieyu Zhang, Yiran Wu, Shaokun Zhang, Erkang Zhu, Beibin Li, Li Jiang, Xiaoyun Zhang, and Chi Wang. Autogen: Enabling next-gen llm applications via multi-agent conversation framework. arXiv preprint arXiv:2308.08155, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.905, + 0.902 + ], + "angle": 0, + "content": "[59] Zhiyong Wu, Chengcheng Han, Zichen Ding, Zhenmin Weng, Zhoumianze Liu, Shunyu Yao, Tao Yu, and Lingpeng" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24623" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.133 + ], + "angle": 0, + "content": "Kong. Os-copilot: Towards generalist computer agents with self-improvement. arXiv preprint arXiv:2402.07456, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.134, + 0.483, + 0.203 + ], + "angle": 0, + "content": "[60] Zhiheng Xi, Wenxiang Chen, Xin Guo, Wei He, Yiwen Ding, Boyang Hong, Ming Zhang, Junzhe Wang, Senjie Jin, Enyu Zhou, et al. The rise and potential of large language model based agents: A survey. arXiv preprint arXiv:2309.07864, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.204, + 0.483, + 0.273 + ], + "angle": 0, + "content": "[61] Zhiheng Xi, Yiwen Ding, Wenxiang Chen, Boyang Hong, Honglin Guo, Junzhe Wang, Dingwen Yang, Chenyang Liao, Xin Guo, Wei He, et al. Agentgym: Evolving large language model-based agents across diverse environments. arXiv preprint arXiv:2406.04151, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.274, + 0.482, + 0.342 + ], + "angle": 0, + "content": "[62] Tianbao Xie, Fan Zhou, Zhoujun Cheng, Peng Shi, Luoxuan Weng, Yitao Liu, Toh Jing Hua, Junning Zhao, Qian Liu, Che Liu, et al. Openagents: An open platform for language agents in the wild. arXiv preprint arXiv:2310.10634, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.343, + 0.483, + 0.412 + ], + "angle": 0, + "content": "[63] Tianbao Xie, Danyang Zhang, Jixuan Chen, Xiaochuan Li, Siheng Zhao, Ruisheng Cao, Toh Jing Hua, Zhoujun Cheng, Dongchan Shin, Fangyu Lei, et al. Osworld: Benchmarking multimodal agents for open-ended tasks in real computer environments. arXiv preprint arXiv:2404.07972, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.413, + 0.483, + 0.468 + ], + "angle": 0, + "content": "[64] Fangzhi Xu, Qiushi Sun, Kanzhi Cheng, Jun Liu, Yu Qiao, and Zhiyong Wu. Interactive evolution: A neural-symbolic self-training framework for large language models. arXiv preprint arXiv:2406.11736, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.469, + 0.483, + 0.524 + ], + "angle": 0, + "content": "[65] Shuyuan Xu, Zelong Li, Kai Mei, and Yongfeng Zhang. Core: LIm as interpreter for natural language programming, pseudo-code programming, and flow programming of ai agents. arXiv preprint arXiv:2405.06907, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.525, + 0.483, + 0.592 + ], + "angle": 0, + "content": "[66] John Yang, Carlos E Jimenez, Alexander Wettig, Kilian Lieret, Shunyu Yao, Karthik Narasimhan, and Ofir Press. Swe-agent: Agent-computer interfaces enable automated software engineering. arXiv preprint arXiv:2405.15793, 2024.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.594, + 0.483, + 0.648 + ], + "angle": 0, + "content": "[67] John Yang, Akshara Prabhakar, Karthik Narasimhan, and Shunyu Yao. Intercode: Standardizing and benchmarking interactive coding with execution feedback. In NeurIPS, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.65, + 0.483, + 0.704 + ], + "angle": 0, + "content": "[68] Zhao Yang, Jiaxuan Liu, Yucheng Han, Xin Chen, Zebiao Huang, Bin Fu, and Gang Yu. Appagent: Multimodal agents as smartphone users. arXiv preprint arXiv:2312.13771, 2023.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.705, + 0.483, + 0.76 + ], + "angle": 0, + "content": "[69] Shunyu Yao, Howard Chen, John Yang, and Karthik Narasimhan. Webshop: Towards scalable real-world web interaction with grounded language agents. In NeurIPS, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.761, + 0.483, + 0.816 + ], + "angle": 0, + "content": "[70] Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, and Yuan Cao. React: Synergizing reasoning and acting in language models. arXiv preprint arXiv:2210.03629, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.817, + 0.483, + 0.871 + ], + "angle": 0, + "content": "[71] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.872, + 0.483, + 0.902 + ], + "angle": 0, + "content": "[72] Pengcheng Yin, Wen-Ding Li, Kefan Xiao, Abhishek Rao, Yeming Wen, Kensen Shi, Joshua Howland, Paige Bailey," + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.092, + 0.905, + 0.134 + ], + "angle": 0, + "content": "Michele Catasta, Henryk Michalewski, et al. Natural language to code generation in interactive data science notebooks. arXiv preprint arXiv:2212.09248, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[73] Zhengqing Yuan, Ruoxi Chen, Zhaoxu Li, Haolong Jia, Lifang He, Chi Wang, and Lichao Sun. Mora: Enabling generalist video generation via a multi-agent framework. arXiv preprint arXiv:2403.13248, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.192, + 0.905, + 0.275 + ], + "angle": 0, + "content": "[74] Matei Zaharia, Omar Khattab, Lingjiao Chen, Jared Quincy Davis, Heather Miller, Chris Potts, James Zou, Michael Carbin, Jonathan Frankle, Naveen Rao, and Ali Ghodsi. The shift from models to compound ai systems. https://bair.berkeley.edu/blog/2024/02/18/compound-ai-systems/, 2024.1,3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.277, + 0.905, + 0.317 + ], + "angle": 0, + "content": "[75] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In CVPR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.319, + 0.905, + 0.36 + ], + "angle": 0, + "content": "[76] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Imposing consistent light. https://github.com/lllyasviel/IC-Light, 2024.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.362, + 0.905, + 0.416 + ], + "angle": 0, + "content": "[77] Wenqi Zhang, Yongliang Shen, Weiming Lu, and Yueting Zhuang. Data-copilot: Bridging billions of data and humans with autonomous workflow. arXiv preprint arXiv:2306.07209, 2023. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.418, + 0.905, + 0.473 + ], + "angle": 0, + "content": "[78] Zeyu Zhang, Xiaohe Bo, Chen Ma, Rui Li, Xu Chen, Quanyu Dai, Jieming Zhu, Zhenhua Dong, and Ji-Rong Wen. A survey on the memory mechanism of large language model based agents. arXiv preprint arXiv:2404.13501, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.475, + 0.905, + 0.517 + ], + "angle": 0, + "content": "[79] Wanjun Zhong, Lianghong Guo, Qiqi Gao, He Ye, and Yanlin Wang. Memorybank: Enhancing large language models with long-term memory. In AAAI, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.518, + 0.905, + 0.586 + ], + "angle": 0, + "content": "[80] Shuyan Zhou, Frank F Xu, Hao Zhu, Xuhui Zhou, Robert Lo, Abishek Sridhar, Xianyi Cheng, Tianyue Ou, Yonatan Bisk, Daniel Fried, et al. Webarena: A realistic web environment for building autonomous agents. arXiv preprint arXiv:2307.13854, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.588, + 0.905, + 0.657 + ], + "angle": 0, + "content": "[81] Wangchunshu Zhou, Yuchen Eleanor Jiang, Long Li, Jia-long Wu, Tiannan Wang, Shi Qiu, Jintian Zhang, Jing Chen, Ruipu Wu, Shuai Wang, et al. Agents: An open-source framework for autonomous language agents. arXiv preprint arXiv:2309.07870, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.659, + 0.905, + 0.714 + ], + "angle": 0, + "content": "[82] Wangchunshu Zhou, Yixin Ou, Shengwei Ding, Long Li, Jialong Wu, Tiannan Wang, Jiamin Chen, Shuai Wang, Xiaohua Xu, Ningyu Zhang, et al. Symbolic learning enables self-evolving agents. arXiv preprint arXiv:2406.18532, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.715, + 0.905, + 0.77 + ], + "angle": 0, + "content": "[83] Mingchen Zhuge, Wenyi Wang, Louis Kirsch, Francesco Faccio, Dmitrii Khizbullin, and Jürgen Schmidhuber. Gptswarm: Language agents as estimizable graphs. In ICML, 2024. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.77 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "24624" + } + ] +] \ No newline at end of file diff --git a/2025/ComfyBench_ Benchmarking LLM-based Agents in ComfyUI for Autonomously Designing Collaborative AI Systems/9eeb1c75-c4c6-45dd-9896-e14c13fb270d_origin.pdf b/2025/ComfyBench_ Benchmarking LLM-based Agents in ComfyUI for Autonomously Designing Collaborative AI Systems/9eeb1c75-c4c6-45dd-9896-e14c13fb270d_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..376aff87dfbee974b615188adab65593cf815c52 --- /dev/null +++ b/2025/ComfyBench_ Benchmarking LLM-based Agents in ComfyUI for Autonomously Designing Collaborative AI Systems/9eeb1c75-c4c6-45dd-9896-e14c13fb270d_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0063cb4df63c1976a72a873dfbde71c80c3cbf3a2e9fc0e9d5305a15903ed0c2 +size 3920409 diff --git a/2025/ComfyBench_ Benchmarking LLM-based Agents in ComfyUI for Autonomously Designing Collaborative AI Systems/full.md b/2025/ComfyBench_ Benchmarking LLM-based Agents in ComfyUI for Autonomously Designing Collaborative AI Systems/full.md new file mode 100644 index 0000000000000000000000000000000000000000..abf82854f521e689464d6ee74ef84d85bbf1c24c --- /dev/null +++ b/2025/ComfyBench_ Benchmarking LLM-based Agents in ComfyUI for Autonomously Designing Collaborative AI Systems/full.md @@ -0,0 +1,352 @@ +# ComfyBench: Benchmarking LLM-based Agents in ComfyUI for Autonomous Designing Collaborative AI Systems + +Xiangyuan Xue $^{1,2*}$ Zeyu Lu $^{1,2}$ Di Huang $^{3}$ Zidong Wang $^{1,4}$ Wanli Ouyang $^{1,4}$ Lei Bai $^{1\boxtimes}$ $^{1}$ Shanghai Artificial Intelligence Laboratory + $^{2}$ Shanghai Jiao Tong University + $^{3}$ The University of Sydney + $^{4}$ The Chinese University of Hong Kong + +# Abstract + +Much previous AI research has focused on developing monolithic models to maximize their intelligence, with the primary goal of enhancing performance on specific tasks. In contrast, this work attempts to study using LLM-based agents to design collaborative AI systems autonomously. To explore this problem, we first introduce ComfyBench to evaluate agents' ability to design collaborative AI systems in ComfyUI. ComfyBench is a comprehensive benchmark comprising 200 diverse tasks covering various instruction-following generation challenges, along with detailed annotations for 3,205 nodes and 20 workflows. Based on ComfyBench, we further develop ComfyAgent, a novel framework that empowers LLM-based agents to autonomously design collaborative AI systems by generating workflows. ComfyAgent is based on two core concepts. First, it represents workflows with code, which can be reversibly converted into workflows and executed as collaborative systems by the interpreter. Second, it constructs a multi-agent system that cooperates to learn from existing workflows and generate new workflows for a given task. While experimental results demonstrate that ComfyAgent achieves a comparable resolve rate to o1-preview and significantly surpasses other agents on ComfyBench, ComfyAgent has resolved only $15\%$ of creative tasks. LLM-based agents still have a long way to go in autonomously designing collaborative AI systems. Progress with ComfyBench is paving the way for more intelligent and autonomous collaborative AI systems. Our code is available at: https://github.com/xyxQwQ/ComfyBench. + +# 1. Introduction + +The recent evolution of AI is defined by the growing importance of collaborative AI systems, which integrate multiple models and tools to work as a whole collaborative system [74]. The success of ChatGPT [37] presents the pos + +sibility of integrating a wide range of tasks, such as web browsing [35], image generation [43], and code execution, into a single conversational agent. Unlike traditional AI models that function as single entities, collaborative AI systems integrate multiple AI components, each contributing unique capabilities to solve complex problems. This shift towards integration has become crucial for achieving state-of-the-art results, as it leverages the combined strengths of diverse AI functionalities within a unified framework. + +Unfortunately, the design space and optimization of collaborative systems often require significant human expertise [31]. For instance, AlphaCode 2 [2] utilizes an intricate AI system to achieve expert-level performance in competitive programming, which involves fine-tuned LLMs for coding and scoring, expert models for clustering, and modules for code execution and filtering. These components collaborate in a precise manner carefully designed by human experts to achieve state-of-the-art performance, which is not achievable by any single component or their casual combination. This brings us to a pivotal question: Can we develop an agent, akin to a human expert, capable of autonomously designing collaborative AI systems? + +In this work, we explore this question in ComfyUI [13], a trending and open-source platform that supports various generative tools and models in terms of nodes. Within ComfyUI, users can flexibly construct and connect nodes to build the workflow of the collaborative AI system, which is capable of producing highly customized artwork. + +To evaluate agents's ability to design collaborative AI systems in ComfyUI, we introduce ComfyBench, a comprehensive benchmark shown in Figure 1, which comprises 200 diverse tasks covering various instruction-following generation challenges, along with detailed annotations for 3,205 nodes and 20 workflows. We categorize tasks into three levels of difficulty: vanilla, complex, and creative. In each task, agents are required to create workflows that describe collaborative AI systems within ComfyUI, which can then be executed to produce the desired outcomes specified in the task instructions. We assess the performance of agents using two evaluation metrics: pass rate, showing the + +![](images/306174134a52a2729f8f0f05a8100caea5e4cd467a0c3d888cebccad81ba374a.jpg) +(a) ComfyBench + +![](images/164f6d4b2a57e631b2ce3976504a198483319364e23deda3dbe4ff97aa8b596f.jpg) +(b) ComfyAgent +Figure 1. (a) ComfyBench is a comprehensive benchmark to evaluate agents's ability to design collaborative AI systems in ComfyUI. Given the task instruction, agents are required to learn from documents and create workflows to describe collaborative AI systems. The performance is measured by pass rate and resolve rate, reflecting whether the workflow can be correctly executed and whether the task requirements are realized. (b) ComfyAgent builds collaborative AI systems in ComfyUI by generating workflows. The workflows are converted into equivalent code so that LLMs can better understand them. ComfyAgent can learn from existing workflows and autonomously design new ones. The generated workflows can be interpreted as collaborative AI systems to complete given tasks. + +success of workflow execution, and resovle rate, reflecting the effectiveness in meeting task requirements. + +Based on ComfyBench, we find such tasks present great challenges for agents. First, the native JSON representation of workflows is inadequate for LLM-based agents to fully comprehend the intrinsic logic and dependencies within the systems [65]. Second, properly organizing the pipeline and utilizing AI-related modules require substantial expert knowledge and complex cognitive processes [21]. Traditional methods, such as Chain-of-Thought (CoT) [57] or Retrieval-Augmented Generation (RAG) [29], suffer from context limitation and hallucination problems, thus failing to generate complex workflows reliably. + +To tackle the aforementioned problems, we introduce ComfyAgent, a novel framework shown in Figure 1, which empowers LLM-based agents to autonomously design collaborative AI systems by generating workflows. The core concepts behind ComfyAgent are two-fold. First, it represents workflows with code, which can be reversibly converted into workflows and executed as collaborative systems by the interpreter. Second, it constructs a multi-agent system [19] which cooperates to learn from existing workflows and generate new workflows for a given task. The core of the multi-agent system is PlanAgent, responsible for the global planning of workflows by task instruction. After collecting document descriptions of all components and code examples of collected workflows, RetrievalAgent au + +tomatically retrieves and learns key information about the given task from code. CombineAgent and AdaptAgent are responsible for code modification to construct a workflow that can fulfill the specified task. Upon task completion, PlanAgent finishes the task to form the final workflow. + +We evaluate common baseline agents [9, 29, 53, 57], as well as our proposed ComfyAgent, on ComfyBench. The experimental results indicate that ComfyAgent achieves comparable pass rate and resolve rate to o1-preview [39] and significantly surpasses other agents on ComfyBench. However, ComfyAgent has resolved only $15\%$ of creative tasks, which is still far from satisfactory. We believe that LLM-based agents still have a long way to go for autonomously designing collaborative AI systems. + +Our contributions can be summarized as follows: + +- We introduce ComfyBench, the first-of-its-kind comprehensive benchmark for the development of agents capable of designing and executing a wide range of collaborative AI systems in ComfyUI. +- We propose ComfyAgent, which represents workflows with code and constructs a novel multi-agent framework to empower LLM-based agents in autonomously designing collaborative AI systems. +- Experimental results demonstrate that with our intricate multi-agent design, ComfyAgent achieves a comparable pass rate and resolve rate to o1-preview and significantly surpasses other agents on ComfyBench. + +# 2. Related Work + +# 2.1. LLM-based Agents + +LLM-based agents leverage LLMs to interact with external tools and solve real-world problems [52, 60]. Agents can exploit the powerful capabilities of LLMs in cognition and creativity, adapting to various novel tasks without relying on human intervention [30, 34]. Besides, agents have access to a wide range of external tools, making up for the natural limitation of processing only natural language [12, 77]. Researchers have been devoted to enhancing the capabilities of LLM-based agents. Many prompt-based methods [7, 33, 47, 53, 54, 57, 70] are proven effective for improving agents' planning and reasoning. Memory mechanism [78, 79] and RAG [16, 29] further extend agents' working context and knowledge base. Reinforcement learning [6, 23, 41], serving as a post-training paradigm, allowing even stronger enhancement and alignment. These various methods hasten the advent of advanced theories for agents, such as self-evolution [61, 64, 82] and unified architecture [48, 81]. + +Based on well-designed module architectures and interaction mechanisms, LLM-based agents can be applied to solve a wide range of complex tasks, such as web navigation [14, 27, 35, 69, 80], interface operation [10, 59, 62, 63, 68], code generation [55, 56, 66, 72], and other tasks in specified domains [17, 22, 46, 50, 67]. These tasks, previously considered to require expert knowledge and human intervention, can now be accomplished by agents reliably. + +# 2.2. Collaborative AI Systems + +Zaharia et al. [74] points out that state-of-the-art results are increasingly obtained by collaborative AI systems instead of monolithic models, which has been verified by many previous works [2, 29, 36, 49, 51]. Yuan et al. [73] even designs a multi-agent framework that incorporates several advanced visual agents to replicate generalist video generation demonstrated by Sora [8]. The advantages of collaborative AI systems are prominent. Dynamic systems make it possible to fetch external knowledge with components such as retrievers. Besides, extra modules such as verifier make the results controllable and trustworthy. + +Compared to monolithic models, collaborative AI systems have posed new challenges in design and optimization. Novel paradigms are emerging to tackle these challenges, of which the most popular is composition frameworks, which build applications out of calls to models and other components [11, 20, 28, 32, 58]. DSPy [24] automatically optimizes the prompt instructions, few-shot examples and other parameters in the pipeline to maximize the end-to-end performance. GPTSwarm [83] supports building agents in the form of graphs and optimizing the parameters and connections as an entire system. + +# 3. ComfyBench + +We propose ComfyBench, a new benchmark for evaluating the capability of LLM-based agents to autonomously design collaborative AI systems within the ComfyUI [13] platform. Given the task instruction, ComfyBench requires agents to generate workflows to describe collaborative AI systems in ComfyUI, producing the desired results after execution. + +# 3.1. ComfyUI Platform + +Overview. ComfyUI is an open-source platform for designing and executing generative pipelines, where users can design their workflows by creating and connecting a series of nodes. Each node stands for a module, which can be a model, a tool, or a widget. On the user side, the workflows are visualized as directed acyclic graphs (DAGs), so that users can operate them intuitively. On the server side, however, the workflows are saved in a JSON object, which describes the nodes and connectivity in a structured manner. ComfyUI executes the JSON format workflows at the granularity of nodes, automatically loading models, conducting inference, passing variables, and returning results to users. + +Features. ComfyUI provides native support for common models and tools related to Stable Diffusion [44] [44], including but not limited to VAEs [25], CLIPs [42], LoRAs [45], IP-Adapters [71], and ControlNets [75]. With the joint efforts of the open-source community, hundreds of ComfyUI extensions are developed to support various latest models and tools, such as InsightFace [18], IC-Light [76], and Segment Anything [26], which bring about immense potential for realizing flexible and controllable generation tasks. By designing intricate workflows, users can produce fantastic artwork that cannot be achieved by common tools. Limitations. Despite the fancy effects that ComfyUI can achieve, the workflows are always manually crafted, which can be quite expensive. On the one hand, expert-level skills are required to figure out the appropriate nodes and parameters that can produce the desired results. On the other hand, workflows are designed to solve a specific task, while creating new workflows can be laborious and time-consuming. + +# 3.2. Benchmark Contents + +ComfyBench collects annotations for 3205 nodes and 20 workflows and construct 200 tasks as is shown in Figure 2. More examples can be found in the supplementary material. Node document provides complete documentation for the 3205 nodes registered in ComfyUI until September 2024, serving as an important manual for agents to utilize the nodes properly. The source code is cloned from Salt AI Documentation [1], which is then cleaned and arranged into documentation, describing the usage of each node and their inputs and outputs. This structured documentation serves as a solid reference resource for understanding the complex node ecosystem within the ComfyUI framework. + +![](images/e936fb0a80897dd3fc39df4e872a7f0f0639304d6354931290ff1361e194c04a.jpg) +Figure 2. ComfyBench provides annotations for 3205 nodes and 20 workflows, together with 200 task instructions categorized into three difficulty levels: vanilla, complex, and creative. + +![](images/758c9e14cecc1cbc7c36afddb1d449e623eca026940e70b3083ba616f2ac9440.jpg) + +Curriculum workflow is a set of 20 tutorial workflows in ComfyUI, which introduces some common nodes and routines to build workflows for solving basic tasks. All the workflows are verified feasible with common nodes and models in ComfyUI and correctly annotated with their purposes and principles. The curriculum workflow allows agents to master useful skills in a human-like manner. + +Task instruction serves as the core part of the benchmark, containing 200 task instructions. Each instruction defines a specific task, which should be solved by designing and executing a ComfyUI workflow. To reflect the generalization ability of agents, the tasks are manually categorized into three difficulty levels: vanilla, complex, and creative. + +- Vanilla includes 100 tasks, which can be solved by learning from a single curriculum workflow, involving minor modifications on several nodes and their parameters. +- Complex includes 60 tasks, which necessitates learning from more than one curriculum workflow, involving functional combinations of multiple pipelines and minor adjustments on nodes and parameters. +- Creative includes 40 tasks, which cannot be solved by directly imitating one or more curriculum workflows, requiring agents to understand the underlying principles and apply the learned skills in a novel way. + +Besides, we provide an auxiliary dataset as the input for some of the tasks. If an image or video is involved in the task, it will be specified in the instruction. + +# 3.3. Evaluation Metrics + +Different from the common benchmarks for image or video generation, we are evaluating the generated workflows, so traditional metrics are not applicable. We divide the evaluation into two progressive stages. The first stage is to check + +whether the generated workflows are syntactically and semantically correct, and the second stage is to check whether the generated workflows can produce the expected results in the task instructions. We design two metrics: pass rate and resolve rate, corresponding to the two stages respectively. + +Pass rate measures the ratio of tasks where generated workflows are syntactically and semantically correct. In practice, each workflow is sent to the server of ComfyUI, which automatically conducts validation and execution. A task will be marked as passed only if the server finishes the execution process and returns a success message. + +Resolve rate measures the ratio of tasks where generated workflows can produce the expected results in the task instruction. In other words, we need to automatically determine the consistency between the produced images or videos and the task requirements. In practice, we apply Visual Language Models (VLMs), where different prompt templates are written according to the input and output modalities. Specifically, we adopt GPT-4o for evaluation, as it demonstrates strong capabilities for understanding language and analyzing visual content. For images, we directly encode them into the prompt. For videos, we uniformly sample up to 10 frames before encoding, which compresses the context length within a reasonable range. A task will be marked as resolved only if the GPT-4o model confirms that all the requirements are satisfied. + +# 3.4. Human Evaluation + +The computation of the resolve rate involves using VLMs to determine the consistency between the produced images or videos and the task requirements, which unavoidably involves subjectivity. To verify the reliability of VLM-based evaluation, we conduct a human evaluation to analyze the agreement between human evaluators and GPT-4o. + +We randomly select 70 tasks that are either completed by ComfyAgent proposed in Section 4 or by a human expert. Human evaluators and the latest GPT-4o model are asked to provide their judgments respectively. We compute the average scores for each question under a sample size of 560 and conduct a correlation analysis. Results are presented in Table 1, where the statistics indicate a strong agreement between human evaluators and GPT-4o, which verifies the reliability of our proposed VLM-based evaluation. More details can be found in the supplementary material. + +Table 1. Correlation analysis between the average scores given by human evaluators and GPT-4o indicates a strong agreement. + +
MetricStatisticP-value
Kendall's τ0.6455.427 × 10-11
Pearson's r0.7641.482 × 10-14
Spearman's ρ0.7373.672 × 10-13
+ +![](images/03ca6b663ec4ba2105a522171c52d968b5127a0306d822290786268bbf4b60b4.jpg) +Figure 3. The architecture of the ComfyAgent framework. Multiple agents cooperate to design workflows in a step-by-step manner. Given the task instruction, the planner initializes the memory and produces a plan. For each step, the planner updates the plan and forms an action. Different actions, including combine, adapt, and retrieve, are then handled by corresponding agents. After combination or adaptation, refine action will be conducted to ensure the correctness. All the agents can interact with the memory, which consists of history, reference, and workspace. Once the task is deemed completed, the planner will finish the procedure and save the workflow. + +# 4. ComfyAgent + +# 4.1. Workflow Representation + +![](images/340be82b2a1bbcc03409b3e02d0b87f2bf68485fb68ef616563355887aad67fb.jpg) +Figure 4. Examples of four common formats to represent workflows: flow graph, JSON, element list, and code. + +Collaborative AI systems coordinate multiple interacting components, such as AI models and external tools, using workflows. These workflows are essentially high-level diagrams that outline the operational logic. They can be naturally represented as directed acyclic graphs (DAGs) consisting of two main elements: vertices and directed edges. Vertices represent information processing components with input and output interfaces, while directed edges represent information flow between components. Specifically, the incoming edges of a vertex represent the external information it receives, while the outgoing edges signify the information it outputs, with multiple edges illustrating various + +information paths. This abstraction allows efficient construction and modification of collaborative AI systems, as demonstrated by platforms like ComfyUI [13]. + +There are four common formats to represent workflows: flow graph, JSON, element list, and code, as shown in Figure 4. Flow graph [13] provides intuitive visualization for humans but is unsuitable for LLM or VLM processing. JSON, typically used in [77], provides a structured representation but is unwieldy for complex workflows due to redundant information and LLMs' context limitations. Element list, typically used in [31], is closer to natural language and provides a more compact representation, but it lacks explicit topological relationships, hindering LLMs from correctly processing complex workflows. + +In contrast, code emerges as the most effective representation, offering various advantages including Turing completeness, rich semantic information, and natural compatibility with LLMs' code generation capabilities. We implement code representation using a restricted subset of Python-like syntax, involving basic operations and control structures while excluding advanced features. We also conduct an ablation study in Section 5.4 to demonstrate the superiority of code representation over other representations. + +# 4.2. Multi-Agent Framework + +Through code representation, we can effectively prompt LLMs to generate workflows for the expected collaborative AI systems. However, with the complexity of workflows increasing, traditional agents suffer from context limitation and hallucination problems, failing to generate workflows correctly and stably, which is proved in Section 5.3. + +To achieve reliable generation of ComfyUI workflows, we propose ComfyAgent, a multi-agent framework to design workflows in a step-by-step manner. As demonstrated in Figure 3, ComfyAgent consists of three independent + +Table 2. Evaluation results of all the agents on ComfyBench. The average task-wise token consumption, the pass rate and resolve rate of every task category denoted by %P and %R, and the overall result on ComfyBench, are reported. The best results are highlighted in bold. + +
AgentCostVanillaComplexCreativeTotal
#Tokens (k)%P%R%P%R%P%R%P%R
GPT-4o + Zero-shot0.500.00.00.00.00.00.00.00.0
GPT-4o + Few-shot2.5532.027.016.78.37.50.022.516.0
GPT-4o + CoT2.5044.029.011.78.312.50.028.017.0
GPT-4o + CoT-SC11.0345.034.011.75.015.00.029.018.5
Claude-3.5-Sonnet + RAG5.7327.013.023.06.77.50.022.08.5
Llama-3.1-70B + RAG5.0458.032.023.010.015.05.039.020.0
GPT-4o + RAG4.9762.041.045.021.740.07.552.023.0
o1-mini + RAG6.0532.016.021.78.312.57.525.012.0
o1-preview + RAG9.4670.046.048.323.330.012.555.532.5
Llama-3.1-70B + ComfyAgent24.2263.035.026.718.320.05.043.524.0
GPT-4o + ComfyAgent14.9667.046.048.321.740.015.056.032.5
+ +modules: Memory, Planner, and Actions. + +Memory stores the recent state of ComfyAgent, which is built on psychological theories [4]. Specifically, it refers to the working memory [5], reflecting the current circumstances, namely the recent history behaviors, results from intermediate steps, internal reasoning, and external knowledge. We formulate them into three parts. + +- History maintains recent plans and actions of Planner, enabling action review for subsequent planning. +- Reference stores information retrieved from the knowledge base, and can be updated through active retrieval. +- **Workspace contains the current workflow together with its natural language annotation, which maintains a complete record of the current working status.** + +Planner, which can be specified as PlanAgent, serves as the core of ComfyAgent, providing the global scheme to design and modify workflows under the task instruction. + +- At the beginning of the task, PlanAgent selects an existing workflow to initialize the memory and produces a thorough multi-step plan based on the task instruction. +- For each step, PlanAgent produces a high-level plan, together with an action based on the current memory, which will then be processed by other agents. +- For each step, PlanAgent evaluates the completion status of the task. Once the task is deemed completed, PlanAgent will finish the procedure and save the workflow. + +Actions represent different actions that can be selected by PlanAgent. All the actions are meant to modify the current memory. In ComfyAgent, we have three actions. + +- Combine is processed by CombineAgent, which aims to combine the current workflow with another workflow from references specified by PlanAgent. +- Adapt is processed by AdaptAgent, which is responsible for adapting the details (e.g. parameters) of the current + +workflow based on the prompt given by PlanAgent. + +- Retrieve is processed by RetrieveAgent, which follows the prompt given by PlanAgent to retrieve relevant information and update references accordingly. + +After combination or adaptation, the updated workflow will be checked and refined by RefineAgent before entering the workspace, which avoids the accumulation of errors. + +After the action is processed, ComfyAgent will enter a new step, where PlanAgent updates the existing plan and forms a new action. In such a way, ComfyAgent gradually generates the workflow. When the procedure is finished, the code representation will be converted into the standard format, which exactly describes the collaborative AI system. + +# 5. Experiments + +# 5.1. Baseline Agents + +ComfyBench places strong requirements on the learning and generalization ability of agents, necessitating LLMs with powerful reasoning capabilities and sufficient context length. Therefore, we conduct experiments on the advanced models, involving both commercial and open-source LLMs. + +- Llama-3.1 (llama-3.1-70b-instruct) [15] is an LLM developed by Meta, achieving state-of-the-art performance in the open-source community. +- Claude-3.5 (claude-3.5-sonnet-20240620) [3] is a commercial LLM developed by Anthropic, with a reputation for its creativity and interaction quality. +- GPT-4o (gpt-4o-2024-08-06) [38] is currently the mainstream model provided by OpenAI and most widely used by the community. It still maintains state-of-the-art performance in leaderboards of various domains. +- o1-mini (o1-mini-2024-09-12) and o1-preview (o1-preview-2024-09-12) [39] are the latest mod + +els released by OpenAI, which focus on complex reasoning tasks such as mathematics and programming. + +We adopt five methods that have proved to be universally effective by the community and can be conveniently adapted to solve the tasks in ComfyBench. + +- Zero-shot Learning directly feeds LLMs with the task instruction to conduct inference. +- Few-shot Learning [9] provides a set of demonstrations in the prompt, which utilizes the in-context learning ability of LLMs and facilitates the correct generation. +- Chain-of-Thought (CoT) [57] is improved based on few-shot learning, which instructs the agent to articulate the reasoning process before providing the final answer. +- CoT with Self-consistency (CoT-SC) [53] is an improved version of CoT, which ensembles parallel trajectories and then selects the most consistent answer. +- Retrieval-Augmented Generation (RAG) [29] learns from the retrieved demonstrations, which efficiently utilizes the most relevant information. + +Considering the limited experimental budget, we evaluate the methods mainly on GPT-4o. ComfyAgent is also evaluated on Llama-3.1 to verify its performance on the open-source model. RAG is evaluated on all the models to provide a horizontal comparison of their capabilities. + +# 5.2. Implementation Details + +ComfyUI Settings. The experiments are conducted on a fixed version of ComfyUI with necessary extensions. A set of common models is installed to make the workflows runnable and the results reproducible. In ComfyUI, the workflows are natively represented in JSON format. We implement a parser to extract the entire DAG described in the JSON object, which is then converted into a sequence of Python-like function calls in topological order. Another parser is implemented to restore this conversion. This ensures that workflows can be effectively generated by agents through code representation while remaining fully compatible with the ComfyUI execution environment. Note that the node documents are not fed into the agents due to context limitations, even though this technique would potentially improve their performance. + +Agent Parameters. We follow CodeAct [54] to format the interaction with LLMs. The parameters of the agents are manually set to exploit their largest potential within a limited context length. For few-shot learning, CoT, and CoT-SC, the number of demonstrations is consistently set to 3, but CoT-SC will ensemble 3 trajectories in parallel to form the final answer. For RAG and ComfyAgent, we adopt an advanced embedding model (text-embedding-3-large) [40] released by OpenAI to encode the texts for retrieval, and the number of retrieved demonstrations is set to 5. For ComfyAgent, the modules are implemented as different roles played by the + +Table 3. Ablation results of RAG on GPT-4o with different workflow representations on ComfyBench. The pass rate and resolve rate are reported, denoted by %P and %R respectively. + +
RepresentationVanillaComplexCreative
%P%R%P%R%P%R
Flow Graph------
JSON52.033.025.08.317.57.5
Element List51.030.023.35.020.010.0
Code62.041.045.015.030.07.5
+ +Table 4. Ablation results of ComfyAgent on GPT-4o with different architectures on ComfyBench. Each variant is implemented by removing one agent from the multi-agent framework. The pass rate and resolve rate are reported, denoted by %P and %R respectively. + +
ArchitectureVanillaComplexCreative
%P%R%P%R%P%R
ComfyAgent67.046.048.321.740.015.0
w/o Combine69.043.015.05.035.015.0
w/o Adapt52.026.020.010.017.55.0
w/o Retrieve27.016.020.05.07.57.5
w/o Refine65.042.043.321.735.012.5
+ +same LLM. We provide the prompt implementation and an example trajectory in the supplementary material. + +# 5.3. Evaluation Results + +We report the evaluation results of different agents on ComfyBench in Table 2. All the agents follow the parameter settings in Section 5.2 and adopt code representation in Section 4.1. We also present some examples generated by ComfyAgent on ComfyBench in Table 5. More examples can be found in the supplementary material. + +In terms of methods, zero-shot learning fails due to a lack of domain knowledge. CoT achieves a better performance than few-shot learning, and CoT-SC shows a slight improvement over CoT. However, these agents fail to achieve comparable performance with retrieval-based agents, because they cannot fully exploit the external knowledge provided by ComfyBench. ComfyAgent outperforms all the baselines in every task category, achieving an overall pass rate of $56\%$ and resolve rate of $32.5\%$ , which indicates that the multi-agent framework applied in ComfyAgent is capable of alleviating the hallucination problem and improving the generation quality and stability. This improvement is consistent on both open-source and commercial models. + +In terms of models, Claude-3.5 and o1-mini show poor performance under RAG, which is mainly caused by the hallucination problem, where models modify the nodes or parameters without verifying their validity. In contrast, Llama-3.1 and GPT-4o show less hallucination and achieve + +Table 5. Some examples generated by ComfyAgent on ComfyBench. Given the task instruction, ComfyAgent designs a workflow to describe the collaborative AI system in ComfyUI, which is then executed to generate the result required by the instruction. + +# Task Instruction + +You are given an image of a male celebrity. Transform the man in the image into a beautiful woman with ponytail hair while preserving her facial identity. + +![](images/9850924b40131c41758c93ce99995792158edeae95609ce6e8686df0f16391ce.jpg) +Input + +![](images/e5c87d670caae48f57e25f3c959eab525508c1c4f2922472c15c79247d0bd3d9.jpg) +Result + +# Task Instruction + +You are given an image of a standing cat. Replace the background with a scene of a cozy living room while keeping the lighting and shadows consistent. + +![](images/864ec1fe4c53ac904d443498c676c50f83e2fea4928a0a1b0717d3d9b669389a.jpg) +Input + +![](images/6766d351c96778144f5b08142df8f83fe284d52258f533f9ecc74f4f4eae225d.jpg) +Result + +# Task Instruction + +You are given an image of a red bridge with a person standing on it. Remove the person from the image while maintaining the original appearance of the bridge. + +![](images/7851733a1bc670c7d7e0c418aa23d2104fe11fd8035cb220e3ddb13a69d5a6e4.jpg) +Input + +![](images/9b3d6606dda2e2a66e4db56d8c3891be29311c94e680b0b935dd1e934ae44cf0.jpg) +Result + +better performance, and o1-preview achieves an outstanding performance comparable to ComfyAgent on GPT-4o. + +# 5.4. Ablation Studies + +Representation. We conduct an ablation study to verify the advantages of representing workflows with code introduced in Section 4.1. We implement three variants of RAG on GPT-4o, where the workflows are respectively represented in JSON, element list, and code. The variants are evaluated on ComfyBench with the same parameter settings as in Section 5.3. Flow graph is not implemented because it is not understandable by LLM-based agents. We report the ablation results in Table 3, where RAG with code representation shows superior performance over other representations, which verifies our claim. RAG with JSON representation is expected to perform even worse without the unique feature of GPT-4o to generate structured JSON objects. + +Architecture. We conduct another ablation to demonstrate the rationality of the multi-agent framework introduced in Section 4.2, where four variants of ComfyAgent on GPT-4o are implemented, each removing one agent from the original framework. We report the ablation results in Table 4. We find that removing any of the four agents will lead to a performance drop. Removing CombineAgent alleviates hallucination in the vanilla category, but leads to failure in the complex category, demonstrating the necessity of integrating multiple demonstrations. Removing AdaptAgent or RetrieveAgent leads to a significant performance drop in all the categories, which underscores the critical role of detail modification and knowledge utilization. Removing RefineAgent leads to a minor but noticeable performance drop in all the categories, indicating its effectiveness in mitigating error accumulation. Therefore, we can conclude that our meticulously-designed multi-agent framework contributes to the superior performance of ComfyAgent. + +# 6. Conclusion + +In this work, we investigate the potential of LLM-based agents to autonomously design collaborative AI systems. Building upon the open-source ComfyUI [13] platform, we present ComfyBench, an autonomous workflow design environment serving as both a benchmark for evaluating the capabilities of agents in designing collaborative systems and a means to extend the applicability of ComfyUI in real-world, language-instruction-based generative applications. We further develop ComfyAgent, which achieves a comparable resolve rate to o1-preview [39] and significantly surpasses other agents on ComfyBench. We point out two key insights underpinning ComfyAgent. First, representing workflows through code is efficient, which provides rich semantics for agents to better understand complex systems. Second, the multi-agent framework of ComfyAgent highlights the substantial benefits of task decomposition and collaborative generation. However, the top-performing agent, ComfyAgent, has achieved a resolve rate of only $15\%$ for creative tasks on ComfyBench. LLM-based agents still have a long way to go in designing collaborative AI systems. We hope that ComfyBench and ComfyAgent can serve as valuable resources in advancing the development of intelligent and autonomous collaborative AI systems. + +Limitations. Well-designed fine-tuning methods could potentially elevate agents to an expert level on ComfyBench, but they are quite expensive and thus not involved in our work. In addition, ComfyAgent operates in an open-loop manner. An alternative approach would involve using VLMs as evaluators to provide feedback [21], enabling a closed-loop system where agents can refine their workflows based on evaluations, thus leading to better results. + +Acknowledgments. This work is supported by the Shanghai Municipal Science and Technology Major Project. + +# References + +[1] Salt AI. Salt node documentation. https://docs. getsalt.ai/md/, 2024.3 +[2] Google DeepMind AlphaCode Team. Alphacode 2 technical report. https://storage.googleapis.com/deepmind-media/AlphaCode2/AlphaCode2_Tech_Report.pdf, 2024.1,3 +[3] Anthropic. Claude 3.5 sonnet. https://www.anthropic.com/news/claude-3-5-sonnet, 2024.6 +[4] Richard C Atkinson. Human memory: A proposed system and its control processes. The psychology of learning and motivation, 1968. 6 +[5] Alan Baddeley. Working memory. Science, 1992. 6 +[6] Yuntao Bai, Andy Jones, Kamal Ndousse, Amanda Askell, Anna Chen, Nova DasSarma, Dawn Drain, Stanislav Fort, Deep Ganguli, Tom Henighan, et al. Training a helpful and harmless assistant with reinforcement learning from human feedback. arXiv preprint arXiv:2204.05862, 2022. 3 +[7] Maciej Besta, Nils Blach, Ales Kubicek, Robert Gerstenberger, Michal Podstawski, Lukas Gianinazzi, Joanna Gajda, Tomasz Lehmann, Hubert Niewiadomski, Piotr Nczyk, et al. Graph of thoughts: Solving elaborate problems with large language models. In AAAI, 2024. 3 +[8] Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Li Jing, David Schnurr, Joe Taylor, Troy Luhman, Eric Luhman, Clarence Ng, Ricky Wang, and Aditya Ramesh. Video generation models as world simulators. https://openai.com/research/video-generation-models-as-world-simulators, 2024.3 +[9] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. In NeurIPS, 2020. 2, 7 +[10] Ruisheng Cao, Fangyu Lei, Haoyuan Wu, Jixuan Chen, Yeqiao Fu, Hongcheng Gao, Xinzhuang Xiong, Hanchong Zhang, Yuchen Mao, Wenjing Hu, et al. Spider2-v: How far are multimodal agents from automating data science and engineering workflows? arXiv preprint arXiv:2407.10956, 2024.3 +[11] Weize Chen, Yusheng Su, Jingwei Zuo, Cheng Yang, Chenfei Yuan, Chen Qian, Chi-Min Chan, Yujia Qin, Yaxi Lu, Ruobing Xie, et al. Agentverse: Facilitating multi-agent collaboration and exploring emergent behaviors in agents. arXiv preprint arXiv:2308.10848, 2023. 3 +[12] Zehui Chen, Kuikun Liu, Qiuchen Wang, Jiangning Liu, Wenwei Zhang, Kai Chen, and Feng Zhao. Mindsearch: Mimicking human minds elicits deep ai searcher. arXiv preprint arXiv:2407.20183, 2024. 3 +[13] comfyanonymous. Comfyui. https://github.com/comfyanonymous/ComfyUI, 2023. 1, 3, 5, 8 +[14] Xiang Deng, Yu Gu, Boyuan Zheng, Shijie Chen, Sam Stevens, Boshi Wang, Huan Sun, and Yu Su. Mind2web: Towards a generalist agent for the web. In NeurIPS, 2024. 3 +[15] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil + +Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 6 +[16] Yunfan Gao, Yun Xiong, Xinyu Gao, Kangxiang Jia, Jinliu Pan, Yuxi Bi, Yi Dai, Jiawei Sun, and Haofen Wang. Retrieval-augmented generation for large language models: A survey. arXiv preprint arXiv:2312.10997, 2023. 3 +[17] Yingqiang Ge, Wenyue Hua, Kai Mei, Juntao Tan, Shuyuan Xu, Zelong Li, Yongfeng Zhang, et al. Openagi: When llm meets domain experts. In NeurlPS, 2024. 3 +[18] Jia Guo, Jiankang Deng, Xiang An, Jack Yu, and Baris Gecer. Insightface: 2d and 3d face analysis project. https://github.com/deepinsight/insightface, 2023.3 +[19] Taicheng Guo, Xiuying Chen, Yaqi Wang, Ruidi Chang, Shichao Pei, Nitesh V Chawla, Olaf Wiest, and Xiangliang Zhang. Large language model based multi-agents: A survey of progress and challenges. arXiv preprint arXiv:2402.01680, 2024. 2 +[20] Sirui Hong, Xiawu Zheng, Jonathan Chen, Yuheng Cheng, Jinlin Wang, Ceyao Zhang, Zili Wang, Steven Ka Shing Yau, Zijuan Lin, Liyang Zhou, et al. Metagpt: Meta programming for multi-agent collaborative framework. arXiv preprint arXiv:2308.00352, 2023. 3 +[21] Shengran Hu, Cong Lu, and Jeff Clune. Automated design of agentic systems. arXiv preprint arXiv:2408.08435, 2024. 2, 8 +[22] Ian Huang, Guandao Yang, and Leonidas Guibas. Blender alchemy: Editing 3d graphics with vision-language models. arXiv preprint arXiv:2404.17672, 2024. 3 +[23] Timo Kaufmann, Paul Weng, Viktor Bengs, and Eyke Hüllermeier. A survey of reinforcement learning from human feedback. arXiv preprint arXiv:2312.14925, 2023. 3 +[24] Omar Khattab, Arnav Singhvi, Paridhi Maheshwari, Zhiyuan Zhang, Keshav Santhanam, Sri Vardhamanan, Saiful Haq, Ashutosh Sharma, Thomas T Joshi, Hanna Moazam, et al. Dspy: Compiling declarative language model calls into self-improving pipelines. arXiv preprint arXiv:2310.03714, 2023. 3 +[25] Diederik P Kingma. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 3 +[26] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dólar, and Ross Girshick. Segment anything. arXiv preprint arXiv:2304.02643, 2023. 3 +[27] Jing Yu Koh, Robert Lo, Lawrence Jang, Vikram Duvvur, Ming Chong Lim, Po-Yu Huang, Graham Neubig, Shuyan Zhou, Ruslan Salakhutdinov, and Daniel Fried. Visualwebarena: Evaluating multimodal agents on realistic visual web tasks. arXiv preprint arXiv:2401.13649, 2024. 3 +[28] LangChain. Applications that can reason. powered by langchain. https://www.langchain.com/, 2023.3 +[29] Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Kuttler, Mike Lewis, Wen-tau Yih, Tim Rocktäschel, et al. Retrieval-augmented generation for knowledge-intensive nlp tasks. In NeurIPS, 2020. 2, 3, 7 + +[30] Guohao Li, Hasan Hammoud, Hani Itani, Dmitrii Khizbullin, and Bernard Ghanem. Camel: Communicative agents for" mind" exploration of large language model society. In NeurIPS, 2023. 3 +[31] Zelong Li, Shuyuan Xu, Kai Mei, Wenyue Hua, Balaji Rama, Om Raheja, Hao Wang, He Zhu, and Yongfeng Zhang. Autoflow: Automated workflow generation for large language model agents. arXiv preprint arXiv:2407.12821, 2024. 1, 5 +[32] Tian Liang, Zhiwei He, Wenxiang Jiao, Xing Wang, Yan Wang, Rui Wang, Yujiu Yang, Shuming Shi, and Zhaopeng Tu. Encouraging divergent thinking in large language models through multi-agent debate. arXiv preprint arXiv:2305.19118, 2023. 3 +[33] Aman Madaan, Niket Tandon, Prakhar Gupta, Skyler Hallinan, Luyu Gao, Sarah Wiegreffe, Uri Alon, Nouha Dziri, Shrimai Prabhumoye, Yiming Yang, et al. Self-refine: Iterative refinement with self-feedback. In NeurIPS, 2024. 3 +[34] Kai Mei, Zelong Li, Shuyuan Xu, Ruosong Ye, Yingqiang Ge, and Yongfeng Zhang. LIm agent operating system. arXiv preprint arXiv:2403.16971, 2024. 3 +[35] Reiichiro Nakano, Jacob Hilton, Suchir Balaji, Jeff Wu, Long Ouyang, Christina Kim, Christopher Hesse, Shantanu Jain, Vineet Kosaraju, William Saunders, et al. Webgpt: Browser-assisted question-answering with human feedback. arXiv preprint arXiv:2112.09332, 2021. 1, 3 +[36] Harsha Nori, Yin Tat Lee, Sheng Zhang, Dean Carignan, Richard Edgar, Nicolo Fusi, Nicholas King, Jonathan Larson, Yuanzhi Li, Weishung Liu, et al. Can generalist foundation models outcompete special-purpose tuning? case study in medicine. arXiv preprint arXiv:2311.16452, 2023. 3 +[37] OpenAI. Chatgpt. https://openai.com/chatgpt/overview/, 2024.1 +[38] OpenAI. Hello gpt-4o. https://openai.com/index/hello-gpt-4o/, 2024.6 +[39] OpenAI. Introducing openai ol-preview. https://openai.com/index/introducing-openai-ol-preview/, 2024.2,6,8 +[40] OpenAI. New embedding models and api updates. https://openai.com/index/new-embedding-models-and-api-updates/, 2024.7 +[41] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. In NeurIPS, 2022. 3 +[42] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 3 +[43] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In ICML, 2021. 1 +[44] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 3 + +[45] Simo Ryu. Low-rank adaptation for fast text-to-image diffusion fine-tuning. https://github.com/cloneofsimo/lora, 2023.3 +[46] Yongliang Shen, Kaitao Song, Xu Tan, Dongsheng Li, Weiming Lu, and Yueting Zhuang. Hugginggpt: Solving ai tasks with chatgpt and its friends in hugging face. In NeurIPS, 2024. 3 +[47] Chan Hee Song, Jiaman Wu, Clayton Washington, Brian M Sadler, Wei-Lun Chao, and Yu Su. Llm-planner: Few-shot grounded planning for embodied agents with large language models. In CVPR, 2023. 3 +[48] Theodore R Sumers, Shunyu Yao, Karthik Narasimhan, and Thomas L Griffiths. Cognitive architectures for language agents. arXiv preprint arXiv:2309.02427, 2023. 3 +[49] Gemini Team, Rohan Anil, Sebastian Borgeaud, Yonghui Wu, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, et al. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805, 2023. 3 +[50] Amitayush Thakur, Yeming Wen, and Swarat Chaudhuri. A language-agent approach to formal theorem-proving. arXiv preprint arXiv:2310.04353, 2023. 3 +[51] Trieu H Trinh, Yuhuai Wu, Quoc V Le, He He, and Thang Luong. Solving olympiad geometry without human demonstrations. Nature, 2024. 3 +[52] Lei Wang, Chen Ma, Xueyang Feng, Zeyu Zhang, Hao Yang, Jingsen Zhang, Zhiyuan Chen, Jiakai Tang, Xu Chen, Yankai Lin, et al. A survey on large language model based autonomous agents. Frontiers of Computer Science, 2024. 3 +[53] Xuezhi Wang, Jason Wei, Dale Schuurmans, Quoc Le, Ed Chi, Sharan Narang, Aakanksha Chowdhery, and Denny Zhou. Self-consistency improves chain of thought reasoning in language models. arXiv preprint arXiv:2203.11171, 2022. 2, 3, 7 +[54] Xingyao Wang, Yangyi Chen, Lifan Yuan, Yizhe Zhang, Yunzhu Li, Hao Peng, and Heng Ji. Executable code actions elicit better lIm agents. arXiv preprint arXiv:2402.01030, 2024. 3, 7 +[55] Xingyao Wang, Boxuan Li, Yufan Song, Frank F Xu, Xiangru Tang, Mingchen Zhuge, Jiayi Pan, Yueqi Song, Bowen Li, Jaskirat Singh, et al. Opendevin: An open platform for ai software developers as generalist agents. arXiv preprint arXiv:2407.16741, 2024. 3 +[56] Zhiruo Wang, Shuyan Zhou, Daniel Fried, and Graham Neubig. Execution-based evaluation for open-domain code generation. arXiv preprint arXiv:2212.10481, 2022. 3 +[57] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. In NeurIPS, 2022. 2, 3, 7 +[58] Qingyun Wu, Gagan Bansal, Jieyu Zhang, Yiran Wu, Shaokun Zhang, Erkang Zhu, Beibin Li, Li Jiang, Xiaoyun Zhang, and Chi Wang. Autogen: Enabling next-gen llm applications via multi-agent conversation framework. arXiv preprint arXiv:2308.08155, 2023. 3 +[59] Zhiyong Wu, Chengcheng Han, Zichen Ding, Zhenmin Weng, Zhoumianze Liu, Shunyu Yao, Tao Yu, and Lingpeng + +Kong. Os-copilot: Towards generalist computer agents with self-improvement. arXiv preprint arXiv:2402.07456, 2024. 3 +[60] Zhiheng Xi, Wenxiang Chen, Xin Guo, Wei He, Yiwen Ding, Boyang Hong, Ming Zhang, Junzhe Wang, Senjie Jin, Enyu Zhou, et al. The rise and potential of large language model based agents: A survey. arXiv preprint arXiv:2309.07864, 2023. 3 +[61] Zhiheng Xi, Yiwen Ding, Wenxiang Chen, Boyang Hong, Honglin Guo, Junzhe Wang, Dingwen Yang, Chenyang Liao, Xin Guo, Wei He, et al. Agentgym: Evolving large language model-based agents across diverse environments. arXiv preprint arXiv:2406.04151, 2024. 3 +[62] Tianbao Xie, Fan Zhou, Zhoujun Cheng, Peng Shi, Luoxuan Weng, Yitao Liu, Toh Jing Hua, Junning Zhao, Qian Liu, Che Liu, et al. Openagents: An open platform for language agents in the wild. arXiv preprint arXiv:2310.10634, 2023. 3 +[63] Tianbao Xie, Danyang Zhang, Jixuan Chen, Xiaochuan Li, Siheng Zhao, Ruisheng Cao, Toh Jing Hua, Zhoujun Cheng, Dongchan Shin, Fangyu Lei, et al. Osworld: Benchmarking multimodal agents for open-ended tasks in real computer environments. arXiv preprint arXiv:2404.07972, 2024. 3 +[64] Fangzhi Xu, Qiushi Sun, Kanzhi Cheng, Jun Liu, Yu Qiao, and Zhiyong Wu. Interactive evolution: A neural-symbolic self-training framework for large language models. arXiv preprint arXiv:2406.11736, 2024. 3 +[65] Shuyuan Xu, Zelong Li, Kai Mei, and Yongfeng Zhang. Core: LIm as interpreter for natural language programming, pseudo-code programming, and flow programming of ai agents. arXiv preprint arXiv:2405.06907, 2024. 2 +[66] John Yang, Carlos E Jimenez, Alexander Wettig, Kilian Lieret, Shunyu Yao, Karthik Narasimhan, and Ofir Press. Swe-agent: Agent-computer interfaces enable automated software engineering. arXiv preprint arXiv:2405.15793, 2024.3 +[67] John Yang, Akshara Prabhakar, Karthik Narasimhan, and Shunyu Yao. Intercode: Standardizing and benchmarking interactive coding with execution feedback. In NeurIPS, 2024. 3 +[68] Zhao Yang, Jiaxuan Liu, Yucheng Han, Xin Chen, Zebiao Huang, Bin Fu, and Gang Yu. Appagent: Multimodal agents as smartphone users. arXiv preprint arXiv:2312.13771, 2023.3 +[69] Shunyu Yao, Howard Chen, John Yang, and Karthik Narasimhan. Webshop: Towards scalable real-world web interaction with grounded language agents. In NeurIPS, 2022. 3 +[70] Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, and Yuan Cao. React: Synergizing reasoning and acting in language models. arXiv preprint arXiv:2210.03629, 2022. 3 +[71] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023. 3 +[72] Pengcheng Yin, Wen-Ding Li, Kefan Xiao, Abhishek Rao, Yeming Wen, Kensen Shi, Joshua Howland, Paige Bailey, + +Michele Catasta, Henryk Michalewski, et al. Natural language to code generation in interactive data science notebooks. arXiv preprint arXiv:2212.09248, 2022. 3 +[73] Zhengqing Yuan, Ruoxi Chen, Zhaoxu Li, Haolong Jia, Lifang He, Chi Wang, and Lichao Sun. Mora: Enabling generalist video generation via a multi-agent framework. arXiv preprint arXiv:2403.13248, 2024. 3 +[74] Matei Zaharia, Omar Khattab, Lingjiao Chen, Jared Quincy Davis, Heather Miller, Chris Potts, James Zou, Michael Carbin, Jonathan Frankle, Naveen Rao, and Ali Ghodsi. The shift from models to compound ai systems. https://bair.berkeley.edu/blog/2024/02/18/compound-ai-systems/, 2024.1,3 +[75] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In CVPR, 2023. 3 +[76] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Imposing consistent light. https://github.com/lllyasviel/IC-Light, 2024.3 +[77] Wenqi Zhang, Yongliang Shen, Weiming Lu, and Yueting Zhuang. Data-copilot: Bridging billions of data and humans with autonomous workflow. arXiv preprint arXiv:2306.07209, 2023. 3, 5 +[78] Zeyu Zhang, Xiaohe Bo, Chen Ma, Rui Li, Xu Chen, Quanyu Dai, Jieming Zhu, Zhenhua Dong, and Ji-Rong Wen. A survey on the memory mechanism of large language model based agents. arXiv preprint arXiv:2404.13501, 2024. 3 +[79] Wanjun Zhong, Lianghong Guo, Qiqi Gao, He Ye, and Yanlin Wang. Memorybank: Enhancing large language models with long-term memory. In AAAI, 2024. 3 +[80] Shuyan Zhou, Frank F Xu, Hao Zhu, Xuhui Zhou, Robert Lo, Abishek Sridhar, Xianyi Cheng, Tianyue Ou, Yonatan Bisk, Daniel Fried, et al. Webarena: A realistic web environment for building autonomous agents. arXiv preprint arXiv:2307.13854, 2023. 3 +[81] Wangchunshu Zhou, Yuchen Eleanor Jiang, Long Li, Jia-long Wu, Tiannan Wang, Shi Qiu, Jintian Zhang, Jing Chen, Ruipu Wu, Shuai Wang, et al. Agents: An open-source framework for autonomous language agents. arXiv preprint arXiv:2309.07870, 2023. 3 +[82] Wangchunshu Zhou, Yixin Ou, Shengwei Ding, Long Li, Jialong Wu, Tiannan Wang, Jiamin Chen, Shuai Wang, Xiaohua Xu, Ningyu Zhang, et al. Symbolic learning enables self-evolving agents. arXiv preprint arXiv:2406.18532, 2024. 3 +[83] Mingchen Zhuge, Wenyi Wang, Louis Kirsch, Francesco Faccio, Dmitrii Khizbullin, and Jürgen Schmidhuber. Gptswarm: Language agents as estimizable graphs. In ICML, 2024. 3 \ No newline at end of file diff --git a/2025/ComfyBench_ Benchmarking LLM-based Agents in ComfyUI for Autonomously Designing Collaborative AI Systems/images.zip b/2025/ComfyBench_ Benchmarking LLM-based Agents in ComfyUI for Autonomously Designing Collaborative AI Systems/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..e75029f5ab11456882033cadcf28d17d2c5c0bc5 --- /dev/null +++ b/2025/ComfyBench_ Benchmarking LLM-based Agents in ComfyUI for Autonomously Designing Collaborative AI Systems/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:534ded3e51d7bcff641acf876df265a0b66e116a2e6e864ae84746cc7ec528d8 +size 484842 diff --git a/2025/ComfyBench_ Benchmarking LLM-based Agents in ComfyUI for Autonomously Designing Collaborative AI Systems/layout.json b/2025/ComfyBench_ Benchmarking LLM-based Agents in ComfyUI for Autonomously Designing Collaborative AI Systems/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..2e15fbc5bf87a8ca594c8483c049f17e42a38556 --- /dev/null +++ b/2025/ComfyBench_ Benchmarking LLM-based Agents in ComfyUI for Autonomously Designing Collaborative AI Systems/layout.json @@ -0,0 +1,8814 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 58, + 102, + 552, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 102, + 552, + 140 + ], + "spans": [ + { + "bbox": [ + 58, + 102, + 552, + 140 + ], + "type": "text", + "content": "ComfyBench: Benchmarking LLM-based Agents in ComfyUI for Autonomous Designing Collaborative AI Systems" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 66, + 159, + 544, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 159, + 544, + 205 + ], + "spans": [ + { + "bbox": [ + 66, + 159, + 544, + 205 + ], + "type": "text", + "content": "Xiangyuan Xue" + }, + { + "bbox": [ + 66, + 159, + 544, + 205 + ], + "type": "inline_equation", + "content": "^{1,2*}" + }, + { + "bbox": [ + 66, + 159, + 544, + 205 + ], + "type": "text", + "content": " Zeyu Lu" + }, + { + "bbox": [ + 66, + 159, + 544, + 205 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 66, + 159, + 544, + 205 + ], + "type": "text", + "content": " Di Huang" + }, + { + "bbox": [ + 66, + 159, + 544, + 205 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 66, + 159, + 544, + 205 + ], + "type": "text", + "content": " Zidong Wang" + }, + { + "bbox": [ + 66, + 159, + 544, + 205 + ], + "type": "inline_equation", + "content": "^{1,4}" + }, + { + "bbox": [ + 66, + 159, + 544, + 205 + ], + "type": "text", + "content": " Wanli Ouyang" + }, + { + "bbox": [ + 66, + 159, + 544, + 205 + ], + "type": "inline_equation", + "content": "^{1,4}" + }, + { + "bbox": [ + 66, + 159, + 544, + 205 + ], + "type": "text", + "content": " Lei Bai" + }, + { + "bbox": [ + 66, + 159, + 544, + 205 + ], + "type": "inline_equation", + "content": "^{1\\boxtimes}" + }, + { + "bbox": [ + 66, + 159, + 544, + 205 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 66, + 159, + 544, + 205 + ], + "type": "text", + "content": "Shanghai Artificial Intelligence Laboratory \n" + }, + { + "bbox": [ + 66, + 159, + 544, + 205 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 66, + 159, + 544, + 205 + ], + "type": "text", + "content": "Shanghai Jiao Tong University \n" + }, + { + "bbox": [ + 66, + 159, + 544, + 205 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 66, + 159, + 544, + 205 + ], + "type": "text", + "content": "The University of Sydney \n" + }, + { + "bbox": [ + 66, + 159, + 544, + 205 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 66, + 159, + 544, + 205 + ], + "type": "text", + "content": "The Chinese University of Hong Kong" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 151, + 230, + 200, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 230, + 200, + 243 + ], + "spans": [ + { + "bbox": [ + 151, + 230, + 200, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 256, + 297, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 256, + 297, + 593 + ], + "spans": [ + { + "bbox": [ + 54, + 256, + 297, + 593 + ], + "type": "text", + "content": "Much previous AI research has focused on developing monolithic models to maximize their intelligence, with the primary goal of enhancing performance on specific tasks. In contrast, this work attempts to study using LLM-based agents to design collaborative AI systems autonomously. To explore this problem, we first introduce ComfyBench to evaluate agents' ability to design collaborative AI systems in ComfyUI. ComfyBench is a comprehensive benchmark comprising 200 diverse tasks covering various instruction-following generation challenges, along with detailed annotations for 3,205 nodes and 20 workflows. Based on ComfyBench, we further develop ComfyAgent, a novel framework that empowers LLM-based agents to autonomously design collaborative AI systems by generating workflows. ComfyAgent is based on two core concepts. First, it represents workflows with code, which can be reversibly converted into workflows and executed as collaborative systems by the interpreter. Second, it constructs a multi-agent system that cooperates to learn from existing workflows and generate new workflows for a given task. While experimental results demonstrate that ComfyAgent achieves a comparable resolve rate to o1-preview and significantly surpasses other agents on ComfyBench, ComfyAgent has resolved only " + }, + { + "bbox": [ + 54, + 256, + 297, + 593 + ], + "type": "inline_equation", + "content": "15\\%" + }, + { + "bbox": [ + 54, + 256, + 297, + 593 + ], + "type": "text", + "content": " of creative tasks. LLM-based agents still have a long way to go in autonomously designing collaborative AI systems. Progress with ComfyBench is paving the way for more intelligent and autonomous collaborative AI systems. Our code is available at: https://github.com/xyxQwQ/ComfyBench." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 615, + 135, + 628 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 615, + 135, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 615, + 135, + 628 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 635, + 296, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 635, + 296, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 635, + 296, + 685 + ], + "type": "text", + "content": "The recent evolution of AI is defined by the growing importance of collaborative AI systems, which integrate multiple models and tools to work as a whole collaborative system [74]. The success of ChatGPT [37] presents the pos" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 232, + 555, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 232, + 555, + 339 + ], + "spans": [ + { + "bbox": [ + 313, + 232, + 555, + 339 + ], + "type": "text", + "content": "sibility of integrating a wide range of tasks, such as web browsing [35], image generation [43], and code execution, into a single conversational agent. Unlike traditional AI models that function as single entities, collaborative AI systems integrate multiple AI components, each contributing unique capabilities to solve complex problems. This shift towards integration has become crucial for achieving state-of-the-art results, as it leverages the combined strengths of diverse AI functionalities within a unified framework." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 340, + 556, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 340, + 556, + 496 + ], + "spans": [ + { + "bbox": [ + 313, + 340, + 556, + 496 + ], + "type": "text", + "content": "Unfortunately, the design space and optimization of collaborative systems often require significant human expertise [31]. For instance, AlphaCode 2 [2] utilizes an intricate AI system to achieve expert-level performance in competitive programming, which involves fine-tuned LLMs for coding and scoring, expert models for clustering, and modules for code execution and filtering. These components collaborate in a precise manner carefully designed by human experts to achieve state-of-the-art performance, which is not achievable by any single component or their casual combination. This brings us to a pivotal question: Can we develop an agent, akin to a human expert, capable of autonomously designing collaborative AI systems?" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 497, + 556, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 497, + 556, + 569 + ], + "spans": [ + { + "bbox": [ + 313, + 497, + 556, + 569 + ], + "type": "text", + "content": "In this work, we explore this question in ComfyUI [13], a trending and open-source platform that supports various generative tools and models in terms of nodes. Within ComfyUI, users can flexibly construct and connect nodes to build the workflow of the collaborative AI system, which is capable of producing highly customized artwork." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 570, + 557, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 557, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 557, + 715 + ], + "type": "text", + "content": "To evaluate agents's ability to design collaborative AI systems in ComfyUI, we introduce ComfyBench, a comprehensive benchmark shown in Figure 1, which comprises 200 diverse tasks covering various instruction-following generation challenges, along with detailed annotations for 3,205 nodes and 20 workflows. We categorize tasks into three levels of difficulty: vanilla, complex, and creative. In each task, agents are required to create workflows that describe collaborative AI systems within ComfyUI, which can then be executed to produce the desired outcomes specified in the task instructions. We assess the performance of agents using two evaluation metrics: pass rate, showing the" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "type": "text", + "content": "*This work was done during his internship at Shanghai Artificial Intelligence Laboratory. Corresponding author, baisanshi@gmail.com." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "24614" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 84, + 553, + 201 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 127, + 83 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 127, + 83 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 127, + 83 + ], + "type": "text", + "content": "(a) ComfyBench" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 58, + 84, + 553, + 201 + ], + "lines": [ + { + "bbox": [ + 58, + 84, + 553, + 201 + ], + "spans": [ + { + "bbox": [ + 58, + 84, + 553, + 201 + ], + "type": "image", + "image_path": "306174134a52a2729f8f0f05a8100caea5e4cd467a0c3d888cebccad81ba374a.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 57, + 218, + 555, + 285 + ], + "blocks": [ + { + "bbox": [ + 58, + 206, + 126, + 218 + ], + "lines": [ + { + "bbox": [ + 58, + 206, + 126, + 218 + ], + "spans": [ + { + "bbox": [ + 58, + 206, + 126, + 218 + ], + "type": "text", + "content": "(b) ComfyAgent" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 57, + 218, + 555, + 285 + ], + "lines": [ + { + "bbox": [ + 57, + 218, + 555, + 285 + ], + "spans": [ + { + "bbox": [ + 57, + 218, + 555, + 285 + ], + "type": "image", + "image_path": "164f6d4b2a57e631b2ce3976504a198483319364e23deda3dbe4ff97aa8b596f.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 294, + 555, + 361 + ], + "lines": [ + { + "bbox": [ + 55, + 294, + 555, + 361 + ], + "spans": [ + { + "bbox": [ + 55, + 294, + 555, + 361 + ], + "type": "text", + "content": "Figure 1. (a) ComfyBench is a comprehensive benchmark to evaluate agents's ability to design collaborative AI systems in ComfyUI. Given the task instruction, agents are required to learn from documents and create workflows to describe collaborative AI systems. The performance is measured by pass rate and resolve rate, reflecting whether the workflow can be correctly executed and whether the task requirements are realized. (b) ComfyAgent builds collaborative AI systems in ComfyUI by generating workflows. The workflows are converted into equivalent code so that LLMs can better understand them. ComfyAgent can learn from existing workflows and autonomously design new ones. The generated workflows can be interpreted as collaborative AI systems to complete given tasks." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 382, + 295, + 406 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 382, + 295, + 406 + ], + "spans": [ + { + "bbox": [ + 55, + 382, + 295, + 406 + ], + "type": "text", + "content": "success of workflow execution, and resovle rate, reflecting the effectiveness in meeting task requirements." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 410, + 295, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 410, + 295, + 542 + ], + "spans": [ + { + "bbox": [ + 54, + 410, + 295, + 542 + ], + "type": "text", + "content": "Based on ComfyBench, we find such tasks present great challenges for agents. First, the native JSON representation of workflows is inadequate for LLM-based agents to fully comprehend the intrinsic logic and dependencies within the systems [65]. Second, properly organizing the pipeline and utilizing AI-related modules require substantial expert knowledge and complex cognitive processes [21]. Traditional methods, such as Chain-of-Thought (CoT) [57] or Retrieval-Augmented Generation (RAG) [29], suffer from context limitation and hallucination problems, thus failing to generate complex workflows reliably." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 546, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 546, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 546, + 295, + 713 + ], + "type": "text", + "content": "To tackle the aforementioned problems, we introduce ComfyAgent, a novel framework shown in Figure 1, which empowers LLM-based agents to autonomously design collaborative AI systems by generating workflows. The core concepts behind ComfyAgent are two-fold. First, it represents workflows with code, which can be reversibly converted into workflows and executed as collaborative systems by the interpreter. Second, it constructs a multi-agent system [19] which cooperates to learn from existing workflows and generate new workflows for a given task. The core of the multi-agent system is PlanAgent, responsible for the global planning of workflows by task instruction. After collecting document descriptions of all components and code examples of collected workflows, RetrievalAgent au" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 382, + 555, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 382, + 555, + 441 + ], + "spans": [ + { + "bbox": [ + 313, + 382, + 555, + 441 + ], + "type": "text", + "content": "tomatically retrieves and learns key information about the given task from code. CombineAgent and AdaptAgent are responsible for code modification to construct a workflow that can fulfill the specified task. Upon task completion, PlanAgent finishes the task to form the final workflow." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 443, + 556, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 443, + 556, + 552 + ], + "spans": [ + { + "bbox": [ + 313, + 443, + 556, + 552 + ], + "type": "text", + "content": "We evaluate common baseline agents [9, 29, 53, 57], as well as our proposed ComfyAgent, on ComfyBench. The experimental results indicate that ComfyAgent achieves comparable pass rate and resolve rate to o1-preview [39] and significantly surpasses other agents on ComfyBench. However, ComfyAgent has resolved only " + }, + { + "bbox": [ + 313, + 443, + 556, + 552 + ], + "type": "inline_equation", + "content": "15\\%" + }, + { + "bbox": [ + 313, + 443, + 556, + 552 + ], + "type": "text", + "content": " of creative tasks, which is still far from satisfactory. We believe that LLM-based agents still have a long way to go for autonomously designing collaborative AI systems." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 325, + 554, + 525, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 554, + 525, + 564 + ], + "spans": [ + { + "bbox": [ + 325, + 554, + 525, + 564 + ], + "type": "text", + "content": "Our contributions can be summarized as follows:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 567, + 553, + 711 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 314, + 567, + 553, + 614 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 567, + 553, + 614 + ], + "spans": [ + { + "bbox": [ + 314, + 567, + 553, + 614 + ], + "type": "text", + "content": "- We introduce ComfyBench, the first-of-its-kind comprehensive benchmark for the development of agents capable of designing and executing a wide range of collaborative AI systems in ComfyUI." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 616, + 553, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 616, + 553, + 662 + ], + "spans": [ + { + "bbox": [ + 313, + 616, + 553, + 662 + ], + "type": "text", + "content": "- We propose ComfyAgent, which represents workflows with code and constructs a novel multi-agent framework to empower LLM-based agents in autonomously designing collaborative AI systems." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 663, + 553, + 711 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 663, + 553, + 711 + ], + "spans": [ + { + "bbox": [ + 313, + 663, + 553, + 711 + ], + "type": "text", + "content": "- Experimental results demonstrate that with our intricate multi-agent design, ComfyAgent achieves a comparable pass rate and resolve rate to o1-preview and significantly surpasses other agents on ComfyBench." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24615" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 143, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 143, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 143, + 83 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 168, + 104 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 168, + 104 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 168, + 104 + ], + "type": "text", + "content": "2.1. LLM-based Agents" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 110, + 295, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 110, + 295, + 324 + ], + "spans": [ + { + "bbox": [ + 55, + 110, + 295, + 324 + ], + "type": "text", + "content": "LLM-based agents leverage LLMs to interact with external tools and solve real-world problems [52, 60]. Agents can exploit the powerful capabilities of LLMs in cognition and creativity, adapting to various novel tasks without relying on human intervention [30, 34]. Besides, agents have access to a wide range of external tools, making up for the natural limitation of processing only natural language [12, 77]. Researchers have been devoted to enhancing the capabilities of LLM-based agents. Many prompt-based methods [7, 33, 47, 53, 54, 57, 70] are proven effective for improving agents' planning and reasoning. Memory mechanism [78, 79] and RAG [16, 29] further extend agents' working context and knowledge base. Reinforcement learning [6, 23, 41], serving as a post-training paradigm, allowing even stronger enhancement and alignment. These various methods hasten the advent of advanced theories for agents, such as self-evolution [61, 64, 82] and unified architecture [48, 81]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 326, + 295, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 326, + 295, + 422 + ], + "spans": [ + { + "bbox": [ + 55, + 326, + 295, + 422 + ], + "type": "text", + "content": "Based on well-designed module architectures and interaction mechanisms, LLM-based agents can be applied to solve a wide range of complex tasks, such as web navigation [14, 27, 35, 69, 80], interface operation [10, 59, 62, 63, 68], code generation [55, 56, 66, 72], and other tasks in specified domains [17, 22, 46, 50, 67]. These tasks, previously considered to require expert knowledge and human intervention, can now be accomplished by agents reliably." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 431, + 198, + 443 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 431, + 198, + 443 + ], + "spans": [ + { + "bbox": [ + 55, + 431, + 198, + 443 + ], + "type": "text", + "content": "2.2. Collaborative AI Systems" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 449, + 295, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 449, + 295, + 581 + ], + "spans": [ + { + "bbox": [ + 55, + 449, + 295, + 581 + ], + "type": "text", + "content": "Zaharia et al. [74] points out that state-of-the-art results are increasingly obtained by collaborative AI systems instead of monolithic models, which has been verified by many previous works [2, 29, 36, 49, 51]. Yuan et al. [73] even designs a multi-agent framework that incorporates several advanced visual agents to replicate generalist video generation demonstrated by Sora [8]. The advantages of collaborative AI systems are prominent. Dynamic systems make it possible to fetch external knowledge with components such as retrievers. Besides, extra modules such as verifier make the results controllable and trustworthy." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "type": "text", + "content": "Compared to monolithic models, collaborative AI systems have posed new challenges in design and optimization. Novel paradigms are emerging to tackle these challenges, of which the most popular is composition frameworks, which build applications out of calls to models and other components [11, 20, 28, 32, 58]. DSPy [24] automatically optimizes the prompt instructions, few-shot examples and other parameters in the pipeline to maximize the end-to-end performance. GPTSwarm [83] supports building agents in the form of graphs and optimizing the parameters and connections as an entire system." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 71, + 397, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 71, + 397, + 84 + ], + "spans": [ + { + "bbox": [ + 314, + 71, + 397, + 84 + ], + "type": "text", + "content": "3. ComfyBench" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 91, + 555, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 91, + 555, + 163 + ], + "spans": [ + { + "bbox": [ + 313, + 91, + 555, + 163 + ], + "type": "text", + "content": "We propose ComfyBench, a new benchmark for evaluating the capability of LLM-based agents to autonomously design collaborative AI systems within the ComfyUI [13] platform. Given the task instruction, ComfyBench requires agents to generate workflows to describe collaborative AI systems in ComfyUI, producing the desired results after execution." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 169, + 425, + 182 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 169, + 425, + 182 + ], + "spans": [ + { + "bbox": [ + 313, + 169, + 425, + 182 + ], + "type": "text", + "content": "3.1. ComfyUI Platform" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 186, + 555, + 330 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 186, + 555, + 330 + ], + "spans": [ + { + "bbox": [ + 313, + 186, + 555, + 330 + ], + "type": "text", + "content": "Overview. ComfyUI is an open-source platform for designing and executing generative pipelines, where users can design their workflows by creating and connecting a series of nodes. Each node stands for a module, which can be a model, a tool, or a widget. On the user side, the workflows are visualized as directed acyclic graphs (DAGs), so that users can operate them intuitively. On the server side, however, the workflows are saved in a JSON object, which describes the nodes and connectivity in a structured manner. ComfyUI executes the JSON format workflows at the granularity of nodes, automatically loading models, conducting inference, passing variables, and returning results to users." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 331, + 556, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 331, + 556, + 547 + ], + "spans": [ + { + "bbox": [ + 313, + 331, + 556, + 547 + ], + "type": "text", + "content": "Features. ComfyUI provides native support for common models and tools related to Stable Diffusion [44] [44], including but not limited to VAEs [25], CLIPs [42], LoRAs [45], IP-Adapters [71], and ControlNets [75]. With the joint efforts of the open-source community, hundreds of ComfyUI extensions are developed to support various latest models and tools, such as InsightFace [18], IC-Light [76], and Segment Anything [26], which bring about immense potential for realizing flexible and controllable generation tasks. By designing intricate workflows, users can produce fantastic artwork that cannot be achieved by common tools. Limitations. Despite the fancy effects that ComfyUI can achieve, the workflows are always manually crafted, which can be quite expensive. On the one hand, expert-level skills are required to figure out the appropriate nodes and parameters that can produce the desired results. On the other hand, workflows are designed to solve a specific task, while creating new workflows can be laborious and time-consuming." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 552, + 437, + 564 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 552, + 437, + 564 + ], + "spans": [ + { + "bbox": [ + 313, + 552, + 437, + 564 + ], + "type": "text", + "content": "3.2. Benchmark Contents" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "content": "ComfyBench collects annotations for 3205 nodes and 20 workflows and construct 200 tasks as is shown in Figure 2. More examples can be found in the supplementary material. Node document provides complete documentation for the 3205 nodes registered in ComfyUI until September 2024, serving as an important manual for agents to utilize the nodes properly. The source code is cloned from Salt AI Documentation [1], which is then cleaned and arranged into documentation, describing the usage of each node and their inputs and outputs. This structured documentation serves as a solid reference resource for understanding the complex node ecosystem within the ComfyUI framework." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "24616" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 70, + 187, + 257 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 187, + 257 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 187, + 257 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 187, + 257 + ], + "type": "image", + "image_path": "e936fb0a80897dd3fc39df4e872a7f0f0639304d6354931290ff1361e194c04a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 267, + 295, + 301 + ], + "lines": [ + { + "bbox": [ + 55, + 267, + 295, + 301 + ], + "spans": [ + { + "bbox": [ + 55, + 267, + 295, + 301 + ], + "type": "text", + "content": "Figure 2. ComfyBench provides annotations for 3205 nodes and 20 workflows, together with 200 task instructions categorized into three difficulty levels: vanilla, complex, and creative." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 190, + 70, + 294, + 257 + ], + "blocks": [ + { + "bbox": [ + 190, + 70, + 294, + 257 + ], + "lines": [ + { + "bbox": [ + 190, + 70, + 294, + 257 + ], + "spans": [ + { + "bbox": [ + 190, + 70, + 294, + 257 + ], + "type": "image", + "image_path": "758c9e14cecc1cbc7c36afddb1d449e623eca026940e70b3083ba616f2ac9440.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 313, + 295, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 313, + 295, + 396 + ], + "spans": [ + { + "bbox": [ + 54, + 313, + 295, + 396 + ], + "type": "text", + "content": "Curriculum workflow is a set of 20 tutorial workflows in ComfyUI, which introduces some common nodes and routines to build workflows for solving basic tasks. All the workflows are verified feasible with common nodes and models in ComfyUI and correctly annotated with their purposes and principles. The curriculum workflow allows agents to master useful skills in a human-like manner." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 397, + 296, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 397, + 296, + 469 + ], + "spans": [ + { + "bbox": [ + 55, + 397, + 296, + 469 + ], + "type": "text", + "content": "Task instruction serves as the core part of the benchmark, containing 200 task instructions. Each instruction defines a specific task, which should be solved by designing and executing a ComfyUI workflow. To reflect the generalization ability of agents, the tasks are manually categorized into three difficulty levels: vanilla, complex, and creative." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 470, + 295, + 601 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 55, + 470, + 295, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 470, + 295, + 506 + ], + "spans": [ + { + "bbox": [ + 55, + 470, + 295, + 506 + ], + "type": "text", + "content": "- Vanilla includes 100 tasks, which can be solved by learning from a single curriculum workflow, involving minor modifications on several nodes and their parameters." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 506, + 295, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 506, + 295, + 553 + ], + "spans": [ + { + "bbox": [ + 55, + 506, + 295, + 553 + ], + "type": "text", + "content": "- Complex includes 60 tasks, which necessitates learning from more than one curriculum workflow, involving functional combinations of multiple pipelines and minor adjustments on nodes and parameters." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 554, + 295, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 554, + 295, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 554, + 295, + 601 + ], + "type": "text", + "content": "- Creative includes 40 tasks, which cannot be solved by directly imitating one or more curriculum workflows, requiring agents to understand the underlying principles and apply the learned skills in a novel way." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 602, + 296, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 602, + 296, + 638 + ], + "spans": [ + { + "bbox": [ + 55, + 602, + 296, + 638 + ], + "type": "text", + "content": "Besides, we provide an auxiliary dataset as the input for some of the tasks. If an image or video is involved in the task, it will be specified in the instruction." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 647, + 168, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 647, + 168, + 659 + ], + "spans": [ + { + "bbox": [ + 55, + 647, + 168, + 659 + ], + "type": "text", + "content": "3.3. Evaluation Metrics" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "content": "Different from the common benchmarks for image or video generation, we are evaluating the generated workflows, so traditional metrics are not applicable. We divide the evaluation into two progressive stages. The first stage is to check" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "content": "whether the generated workflows are syntactically and semantically correct, and the second stage is to check whether the generated workflows can produce the expected results in the task instructions. We design two metrics: pass rate and resolve rate, corresponding to the two stages respectively." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 133, + 553, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 133, + 553, + 204 + ], + "spans": [ + { + "bbox": [ + 313, + 133, + 553, + 204 + ], + "type": "text", + "content": "Pass rate measures the ratio of tasks where generated workflows are syntactically and semantically correct. In practice, each workflow is sent to the server of ComfyUI, which automatically conducts validation and execution. A task will be marked as passed only if the server finishes the execution process and returns a success message." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 205, + 555, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 205, + 555, + 384 + ], + "spans": [ + { + "bbox": [ + 313, + 205, + 555, + 384 + ], + "type": "text", + "content": "Resolve rate measures the ratio of tasks where generated workflows can produce the expected results in the task instruction. In other words, we need to automatically determine the consistency between the produced images or videos and the task requirements. In practice, we apply Visual Language Models (VLMs), where different prompt templates are written according to the input and output modalities. Specifically, we adopt GPT-4o for evaluation, as it demonstrates strong capabilities for understanding language and analyzing visual content. For images, we directly encode them into the prompt. For videos, we uniformly sample up to 10 frames before encoding, which compresses the context length within a reasonable range. A task will be marked as resolved only if the GPT-4o model confirms that all the requirements are satisfied." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 392, + 425, + 404 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 392, + 425, + 404 + ], + "spans": [ + { + "bbox": [ + 313, + 392, + 425, + 404 + ], + "type": "text", + "content": "3.4. Human Evaluation" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 410, + 554, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 410, + 554, + 482 + ], + "spans": [ + { + "bbox": [ + 313, + 410, + 554, + 482 + ], + "type": "text", + "content": "The computation of the resolve rate involves using VLMs to determine the consistency between the produced images or videos and the task requirements, which unavoidably involves subjectivity. To verify the reliability of VLM-based evaluation, we conduct a human evaluation to analyze the agreement between human evaluators and GPT-4o." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 483, + 554, + 603 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 483, + 554, + 603 + ], + "spans": [ + { + "bbox": [ + 313, + 483, + 554, + 603 + ], + "type": "text", + "content": "We randomly select 70 tasks that are either completed by ComfyAgent proposed in Section 4 or by a human expert. Human evaluators and the latest GPT-4o model are asked to provide their judgments respectively. We compute the average scores for each question under a sample size of 560 and conduct a correlation analysis. Results are presented in Table 1, where the statistics indicate a strong agreement between human evaluators and GPT-4o, which verifies the reliability of our proposed VLM-based evaluation. More details can be found in the supplementary material." + } + ] + } + ], + "index": 17 + }, + { + "type": "table", + "bbox": [ + 341, + 644, + 528, + 706 + ], + "blocks": [ + { + "bbox": [ + 313, + 613, + 553, + 635 + ], + "lines": [ + { + "bbox": [ + 313, + 613, + 553, + 635 + ], + "spans": [ + { + "bbox": [ + 313, + 613, + 553, + 635 + ], + "type": "text", + "content": "Table 1. Correlation analysis between the average scores given by human evaluators and GPT-4o indicates a strong agreement." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 341, + 644, + 528, + 706 + ], + "lines": [ + { + "bbox": [ + 341, + 644, + 528, + 706 + ], + "spans": [ + { + "bbox": [ + 341, + 644, + 528, + 706 + ], + "type": "table", + "html": "
MetricStatisticP-value
Kendall's τ0.6455.427 × 10-11
Pearson's r0.7641.482 × 10-14
Spearman's ρ0.7373.672 × 10-13
", + "image_path": "a493af92df48a2aa7bd3ad3821a9b3842eda4ead0aaea2dd2ab82de9900059c7.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "table_body" + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24617" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 70, + 555, + 203 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 555, + 203 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 555, + 203 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 555, + 203 + ], + "type": "image", + "image_path": "03ca6b663ec4ba2105a522171c52d968b5127a0306d822290786268bbf4b60b4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 211, + 555, + 266 + ], + "lines": [ + { + "bbox": [ + 55, + 211, + 555, + 266 + ], + "spans": [ + { + "bbox": [ + 55, + 211, + 555, + 266 + ], + "type": "text", + "content": "Figure 3. The architecture of the ComfyAgent framework. Multiple agents cooperate to design workflows in a step-by-step manner. Given the task instruction, the planner initializes the memory and produces a plan. For each step, the planner updates the plan and forms an action. Different actions, including combine, adapt, and retrieve, are then handled by corresponding agents. After combination or adaptation, refine action will be conducted to ensure the correctness. All the agents can interact with the memory, which consists of history, reference, and workspace. Once the task is deemed completed, the planner will finish the procedure and save the workflow." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 286, + 138, + 300 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 286, + 138, + 300 + ], + "spans": [ + { + "bbox": [ + 55, + 286, + 138, + 300 + ], + "type": "text", + "content": "4. ComfyAgent" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 306, + 198, + 319 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 306, + 198, + 319 + ], + "spans": [ + { + "bbox": [ + 55, + 306, + 198, + 319 + ], + "type": "text", + "content": "4.1. Workflow Representation" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 57, + 335, + 294, + 524 + ], + "blocks": [ + { + "bbox": [ + 57, + 335, + 294, + 524 + ], + "lines": [ + { + "bbox": [ + 57, + 335, + 294, + 524 + ], + "spans": [ + { + "bbox": [ + 57, + 335, + 294, + 524 + ], + "type": "image", + "image_path": "340be82b2a1bbcc03409b3e02d0b87f2bf68485fb68ef616563355887aad67fb.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 533, + 295, + 556 + ], + "lines": [ + { + "bbox": [ + 55, + 533, + 295, + 556 + ], + "spans": [ + { + "bbox": [ + 55, + 533, + 295, + 556 + ], + "type": "text", + "content": "Figure 4. Examples of four common formats to represent workflows: flow graph, JSON, element list, and code." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "content": "Collaborative AI systems coordinate multiple interacting components, such as AI models and external tools, using workflows. These workflows are essentially high-level diagrams that outline the operational logic. They can be naturally represented as directed acyclic graphs (DAGs) consisting of two main elements: vertices and directed edges. Vertices represent information processing components with input and output interfaces, while directed edges represent information flow between components. Specifically, the incoming edges of a vertex represent the external information it receives, while the outgoing edges signify the information it outputs, with multiple edges illustrating various" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 287, + 553, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 287, + 553, + 323 + ], + "spans": [ + { + "bbox": [ + 313, + 287, + 553, + 323 + ], + "type": "text", + "content": "information paths. This abstraction allows efficient construction and modification of collaborative AI systems, as demonstrated by platforms like ComfyUI [13]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 324, + 554, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 324, + 554, + 456 + ], + "spans": [ + { + "bbox": [ + 313, + 324, + 554, + 456 + ], + "type": "text", + "content": "There are four common formats to represent workflows: flow graph, JSON, element list, and code, as shown in Figure 4. Flow graph [13] provides intuitive visualization for humans but is unsuitable for LLM or VLM processing. JSON, typically used in [77], provides a structured representation but is unwieldy for complex workflows due to redundant information and LLMs' context limitations. Element list, typically used in [31], is closer to natural language and provides a more compact representation, but it lacks explicit topological relationships, hindering LLMs from correctly processing complex workflows." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 457, + 554, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 457, + 554, + 564 + ], + "spans": [ + { + "bbox": [ + 313, + 457, + 554, + 564 + ], + "type": "text", + "content": "In contrast, code emerges as the most effective representation, offering various advantages including Turing completeness, rich semantic information, and natural compatibility with LLMs' code generation capabilities. We implement code representation using a restricted subset of Python-like syntax, involving basic operations and control structures while excluding advanced features. We also conduct an ablation study in Section 5.4 to demonstrate the superiority of code representation over other representations." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 574, + 453, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 574, + 453, + 586 + ], + "spans": [ + { + "bbox": [ + 313, + 574, + 453, + 586 + ], + "type": "text", + "content": "4.2. Multi-Agent Framework" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 594, + 554, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 554, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 554, + 665 + ], + "type": "text", + "content": "Through code representation, we can effectively prompt LLMs to generate workflows for the expected collaborative AI systems. However, with the complexity of workflows increasing, traditional agents suffer from context limitation and hallucination problems, failing to generate workflows correctly and stably, which is proved in Section 5.3." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 666, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 666, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 666, + 554, + 713 + ], + "type": "text", + "content": "To achieve reliable generation of ComfyUI workflows, we propose ComfyAgent, a multi-agent framework to design workflows in a step-by-step manner. As demonstrated in Figure 3, ComfyAgent consists of three independent" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "24618" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 102, + 555, + 281 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "text", + "content": "Table 2. Evaluation results of all the agents on ComfyBench. The average task-wise token consumption, the pass rate and resolve rate of every task category denoted by %P and %R, and the overall result on ComfyBench, are reported. The best results are highlighted in bold." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 102, + 555, + 281 + ], + "lines": [ + { + "bbox": [ + 56, + 102, + 555, + 281 + ], + "spans": [ + { + "bbox": [ + 56, + 102, + 555, + 281 + ], + "type": "table", + "html": "
AgentCostVanillaComplexCreativeTotal
#Tokens (k)%P%R%P%R%P%R%P%R
GPT-4o + Zero-shot0.500.00.00.00.00.00.00.00.0
GPT-4o + Few-shot2.5532.027.016.78.37.50.022.516.0
GPT-4o + CoT2.5044.029.011.78.312.50.028.017.0
GPT-4o + CoT-SC11.0345.034.011.75.015.00.029.018.5
Claude-3.5-Sonnet + RAG5.7327.013.023.06.77.50.022.08.5
Llama-3.1-70B + RAG5.0458.032.023.010.015.05.039.020.0
GPT-4o + RAG4.9762.041.045.021.740.07.552.023.0
o1-mini + RAG6.0532.016.021.78.312.57.525.012.0
o1-preview + RAG9.4670.046.048.323.330.012.555.532.5
Llama-3.1-70B + ComfyAgent24.2263.035.026.718.320.05.043.524.0
GPT-4o + ComfyAgent14.9667.046.048.321.740.015.056.032.5
", + "image_path": "3d91971c2a74b4f6cee354a1a71d868971f9d1de174d1a2c048aad04a6d01cb3.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 300, + 223, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 300, + 223, + 312 + ], + "spans": [ + { + "bbox": [ + 55, + 300, + 223, + 312 + ], + "type": "text", + "content": "modules: Memory, Planner, and Actions." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 313, + 295, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 313, + 295, + 386 + ], + "spans": [ + { + "bbox": [ + 55, + 313, + 295, + 386 + ], + "type": "text", + "content": "Memory stores the recent state of ComfyAgent, which is built on psychological theories [4]. Specifically, it refers to the working memory [5], reflecting the current circumstances, namely the recent history behaviors, results from intermediate steps, internal reasoning, and external knowledge. We formulate them into three parts." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 386, + 295, + 470 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 55, + 386, + 295, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 386, + 295, + 410 + ], + "spans": [ + { + "bbox": [ + 55, + 386, + 295, + 410 + ], + "type": "text", + "content": "- History maintains recent plans and actions of Planner, enabling action review for subsequent planning." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 410, + 295, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 410, + 295, + 434 + ], + "spans": [ + { + "bbox": [ + 55, + 410, + 295, + 434 + ], + "type": "text", + "content": "- Reference stores information retrieved from the knowledge base, and can be updated through active retrieval." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 434, + 295, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 434, + 295, + 470 + ], + "spans": [ + { + "bbox": [ + 55, + 434, + 295, + 470 + ], + "type": "text", + "content": "- **Workspace contains the current workflow together with its natural language annotation, which maintains a complete record of the current working status.**" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 471, + 295, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 471, + 295, + 506 + ], + "spans": [ + { + "bbox": [ + 55, + 471, + 295, + 506 + ], + "type": "text", + "content": "Planner, which can be specified as PlanAgent, serves as the core of ComfyAgent, providing the global scheme to design and modify workflows under the task instruction." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 508, + 295, + 615 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 55, + 508, + 295, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 508, + 295, + 544 + ], + "spans": [ + { + "bbox": [ + 55, + 508, + 295, + 544 + ], + "type": "text", + "content": "- At the beginning of the task, PlanAgent selects an existing workflow to initialize the memory and produces a thorough multi-step plan based on the task instruction." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 544, + 295, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 544, + 295, + 579 + ], + "spans": [ + { + "bbox": [ + 55, + 544, + 295, + 579 + ], + "type": "text", + "content": "- For each step, PlanAgent produces a high-level plan, together with an action based on the current memory, which will then be processed by other agents." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 580, + 295, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 580, + 295, + 615 + ], + "spans": [ + { + "bbox": [ + 55, + 580, + 295, + 615 + ], + "type": "text", + "content": "- For each step, PlanAgent evaluates the completion status of the task. Once the task is deemed completed, PlanAgent will finish the procedure and save the workflow." + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 616, + 295, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 616, + 295, + 652 + ], + "spans": [ + { + "bbox": [ + 55, + 616, + 295, + 652 + ], + "type": "text", + "content": "Actions represent different actions that can be selected by PlanAgent. All the actions are meant to modify the current memory. In ComfyAgent, we have three actions." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 654, + 295, + 714 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 55, + 654, + 295, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 295, + 689 + ], + "type": "text", + "content": "- Combine is processed by CombineAgent, which aims to combine the current workflow with another workflow from references specified by PlanAgent." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": "- Adapt is processed by AdaptAgent, which is responsible for adapting the details (e.g. parameters) of the current" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 322, + 300, + 531, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 300, + 531, + 312 + ], + "spans": [ + { + "bbox": [ + 322, + 300, + 531, + 312 + ], + "type": "text", + "content": "workflow based on the prompt given by PlanAgent." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 312, + 553, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 312, + 553, + 348 + ], + "spans": [ + { + "bbox": [ + 313, + 312, + 553, + 348 + ], + "type": "text", + "content": "- Retrieve is processed by RetrieveAgent, which follows the prompt given by PlanAgent to retrieve relevant information and update references accordingly." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 349, + 553, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 349, + 553, + 384 + ], + "spans": [ + { + "bbox": [ + 313, + 349, + 553, + 384 + ], + "type": "text", + "content": "After combination or adaptation, the updated workflow will be checked and refined by RefineAgent before entering the workspace, which avoids the accumulation of errors." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 385, + 554, + 457 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 385, + 554, + 457 + ], + "spans": [ + { + "bbox": [ + 313, + 385, + 554, + 457 + ], + "type": "text", + "content": "After the action is processed, ComfyAgent will enter a new step, where PlanAgent updates the existing plan and forms a new action. In such a way, ComfyAgent gradually generates the workflow. When the procedure is finished, the code representation will be converted into the standard format, which exactly describes the collaborative AI system." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 314, + 470, + 395, + 484 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 470, + 395, + 484 + ], + "spans": [ + { + "bbox": [ + 314, + 470, + 395, + 484 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 314, + 491, + 410, + 503 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 491, + 410, + 503 + ], + "spans": [ + { + "bbox": [ + 314, + 491, + 410, + 503 + ], + "type": "text", + "content": "5.1. Baseline Agents" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 509, + 554, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 509, + 554, + 569 + ], + "spans": [ + { + "bbox": [ + 313, + 509, + 554, + 569 + ], + "type": "text", + "content": "ComfyBench places strong requirements on the learning and generalization ability of agents, necessitating LLMs with powerful reasoning capabilities and sufficient context length. Therefore, we conduct experiments on the advanced models, involving both commercial and open-source LLMs." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 570, + 553, + 714 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 313, + 570, + 553, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 553, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 553, + 605 + ], + "type": "text", + "content": "- Llama-3.1 (llama-3.1-70b-instruct) [15] is an LLM developed by Meta, achieving state-of-the-art performance in the open-source community." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 605, + 553, + 642 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 553, + 642 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 553, + 642 + ], + "type": "text", + "content": "- Claude-3.5 (claude-3.5-sonnet-20240620) [3] is a commercial LLM developed by Anthropic, with a reputation for its creativity and interaction quality." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 642, + 553, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 553, + 689 + ], + "type": "text", + "content": "- GPT-4o (gpt-4o-2024-08-06) [38] is currently the mainstream model provided by OpenAI and most widely used by the community. It still maintains state-of-the-art performance in leaderboards of various domains." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "type": "text", + "content": "- o1-mini (o1-mini-2024-09-12) and o1-preview (o1-preview-2024-09-12) [39] are the latest mod" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24619" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 63, + 72, + 295, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 72, + 295, + 96 + ], + "spans": [ + { + "bbox": [ + 63, + 72, + 295, + 96 + ], + "type": "text", + "content": "els released by OpenAI, which focus on complex reasoning tasks such as mathematics and programming." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 97, + 295, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 97, + 295, + 133 + ], + "spans": [ + { + "bbox": [ + 55, + 97, + 295, + 133 + ], + "type": "text", + "content": "We adopt five methods that have proved to be universally effective by the community and can be conveniently adapted to solve the tasks in ComfyBench." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 133, + 295, + 300 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 55, + 133, + 295, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 133, + 295, + 156 + ], + "spans": [ + { + "bbox": [ + 55, + 133, + 295, + 156 + ], + "type": "text", + "content": "- Zero-shot Learning directly feeds LLMs with the task instruction to conduct inference." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 158, + 295, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 158, + 295, + 194 + ], + "spans": [ + { + "bbox": [ + 55, + 158, + 295, + 194 + ], + "type": "text", + "content": "- Few-shot Learning [9] provides a set of demonstrations in the prompt, which utilizes the in-context learning ability of LLMs and facilitates the correct generation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 194, + 295, + 229 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 194, + 295, + 229 + ], + "spans": [ + { + "bbox": [ + 55, + 194, + 295, + 229 + ], + "type": "text", + "content": "- Chain-of-Thought (CoT) [57] is improved based on few-shot learning, which instructs the agent to articulate the reasoning process before providing the final answer." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 229, + 295, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 229, + 295, + 264 + ], + "spans": [ + { + "bbox": [ + 55, + 229, + 295, + 264 + ], + "type": "text", + "content": "- CoT with Self-consistency (CoT-SC) [53] is an improved version of CoT, which ensembles parallel trajectories and then selects the most consistent answer." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 265, + 295, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 265, + 295, + 300 + ], + "spans": [ + { + "bbox": [ + 55, + 265, + 295, + 300 + ], + "type": "text", + "content": "- Retrieval-Augmented Generation (RAG) [29] learns from the retrieved demonstrations, which efficiently utilizes the most relevant information." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 302, + 296, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 302, + 296, + 361 + ], + "spans": [ + { + "bbox": [ + 55, + 302, + 296, + 361 + ], + "type": "text", + "content": "Considering the limited experimental budget, we evaluate the methods mainly on GPT-4o. ComfyAgent is also evaluated on Llama-3.1 to verify its performance on the open-source model. RAG is evaluated on all the models to provide a horizontal comparison of their capabilities." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 371, + 188, + 384 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 371, + 188, + 384 + ], + "spans": [ + { + "bbox": [ + 55, + 371, + 188, + 384 + ], + "type": "text", + "content": "5.2. Implementation Details" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 54, + 389, + 295, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 389, + 295, + 569 + ], + "spans": [ + { + "bbox": [ + 54, + 389, + 295, + 569 + ], + "type": "text", + "content": "ComfyUI Settings. The experiments are conducted on a fixed version of ComfyUI with necessary extensions. A set of common models is installed to make the workflows runnable and the results reproducible. In ComfyUI, the workflows are natively represented in JSON format. We implement a parser to extract the entire DAG described in the JSON object, which is then converted into a sequence of Python-like function calls in topological order. Another parser is implemented to restore this conversion. This ensures that workflows can be effectively generated by agents through code representation while remaining fully compatible with the ComfyUI execution environment. Note that the node documents are not fed into the agents due to context limitations, even though this technique would potentially improve their performance." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 54, + 570, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 570, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 570, + 295, + 714 + ], + "type": "text", + "content": "Agent Parameters. We follow CodeAct [54] to format the interaction with LLMs. The parameters of the agents are manually set to exploit their largest potential within a limited context length. For few-shot learning, CoT, and CoT-SC, the number of demonstrations is consistently set to 3, but CoT-SC will ensemble 3 trajectories in parallel to form the final answer. For RAG and ComfyAgent, we adopt an advanced embedding model (text-embedding-3-large) [40] released by OpenAI to encode the texts for retrieval, and the number of retrieved demonstrations is set to 5. For ComfyAgent, the modules are implemented as different roles played by the" + } + ] + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 315, + 113, + 553, + 196 + ], + "blocks": [ + { + "bbox": [ + 313, + 71, + 555, + 105 + ], + "lines": [ + { + "bbox": [ + 313, + 71, + 555, + 105 + ], + "spans": [ + { + "bbox": [ + 313, + 71, + 555, + 105 + ], + "type": "text", + "content": "Table 3. Ablation results of RAG on GPT-4o with different workflow representations on ComfyBench. The pass rate and resolve rate are reported, denoted by %P and %R respectively." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 113, + 553, + 196 + ], + "lines": [ + { + "bbox": [ + 315, + 113, + 553, + 196 + ], + "spans": [ + { + "bbox": [ + 315, + 113, + 553, + 196 + ], + "type": "table", + "html": "
RepresentationVanillaComplexCreative
%P%R%P%R%P%R
Flow Graph------
JSON52.033.025.08.317.57.5
Element List51.030.023.35.020.010.0
Code62.041.045.015.030.07.5
", + "image_path": "c874ec8ed6c149495edaf30f90f06e8ef484ec27c0c34f2cb964600f4c8e164a.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "type": "table", + "bbox": [ + 320, + 261, + 550, + 356 + ], + "blocks": [ + { + "bbox": [ + 313, + 208, + 555, + 253 + ], + "lines": [ + { + "bbox": [ + 313, + 208, + 555, + 253 + ], + "spans": [ + { + "bbox": [ + 313, + 208, + 555, + 253 + ], + "type": "text", + "content": "Table 4. Ablation results of ComfyAgent on GPT-4o with different architectures on ComfyBench. Each variant is implemented by removing one agent from the multi-agent framework. The pass rate and resolve rate are reported, denoted by %P and %R respectively." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 320, + 261, + 550, + 356 + ], + "lines": [ + { + "bbox": [ + 320, + 261, + 550, + 356 + ], + "spans": [ + { + "bbox": [ + 320, + 261, + 550, + 356 + ], + "type": "table", + "html": "
ArchitectureVanillaComplexCreative
%P%R%P%R%P%R
ComfyAgent67.046.048.321.740.015.0
w/o Combine69.043.015.05.035.015.0
w/o Adapt52.026.020.010.017.55.0
w/o Retrieve27.016.020.05.07.57.5
w/o Refine65.042.043.321.735.012.5
", + "image_path": "607da0b1567adb3d1f37bd70c7c5715c0d3eca0bb03de12a7d73a4a75b8fbaa1.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_body" + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 376, + 555, + 401 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 376, + 555, + 401 + ], + "spans": [ + { + "bbox": [ + 313, + 376, + 555, + 401 + ], + "type": "text", + "content": "same LLM. We provide the prompt implementation and an example trajectory in the supplementary material." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 408, + 425, + 419 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 408, + 425, + 419 + ], + "spans": [ + { + "bbox": [ + 313, + 408, + 425, + 419 + ], + "type": "text", + "content": "5.3. Evaluation Results" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 426, + 555, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 555, + 498 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 555, + 498 + ], + "type": "text", + "content": "We report the evaluation results of different agents on ComfyBench in Table 2. All the agents follow the parameter settings in Section 5.2 and adopt code representation in Section 4.1. We also present some examples generated by ComfyAgent on ComfyBench in Table 5. More examples can be found in the supplementary material." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 498, + 554, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 554, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 554, + 653 + ], + "type": "text", + "content": "In terms of methods, zero-shot learning fails due to a lack of domain knowledge. CoT achieves a better performance than few-shot learning, and CoT-SC shows a slight improvement over CoT. However, these agents fail to achieve comparable performance with retrieval-based agents, because they cannot fully exploit the external knowledge provided by ComfyBench. ComfyAgent outperforms all the baselines in every task category, achieving an overall pass rate of " + }, + { + "bbox": [ + 313, + 498, + 554, + 653 + ], + "type": "inline_equation", + "content": "56\\%" + }, + { + "bbox": [ + 313, + 498, + 554, + 653 + ], + "type": "text", + "content": " and resolve rate of " + }, + { + "bbox": [ + 313, + 498, + 554, + 653 + ], + "type": "inline_equation", + "content": "32.5\\%" + }, + { + "bbox": [ + 313, + 498, + 554, + 653 + ], + "type": "text", + "content": ", which indicates that the multi-agent framework applied in ComfyAgent is capable of alleviating the hallucination problem and improving the generation quality and stability. This improvement is consistent on both open-source and commercial models." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 654, + 555, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 555, + 712 + ], + "type": "text", + "content": "In terms of models, Claude-3.5 and o1-mini show poor performance under RAG, which is mainly caused by the hallucination problem, where models modify the nodes or parameters without verifying their validity. In contrast, Llama-3.1 and GPT-4o show less hallucination and achieve" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24620" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "text", + "content": "Table 5. Some examples generated by ComfyAgent on ComfyBench. Given the task instruction, ComfyAgent designs a workflow to describe the collaborative AI system in ComfyUI, which is then executed to generate the result required by the instruction." + } + ] + } + ], + "index": 0, + "type": "text" + }, + { + "bbox": [ + 108, + 107, + 173, + 117 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 107, + 173, + 117 + ], + "spans": [ + { + "bbox": [ + 108, + 107, + 173, + 117 + ], + "type": "text", + "content": "Task Instruction" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 121, + 214, + 165 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 121, + 214, + 165 + ], + "spans": [ + { + "bbox": [ + 62, + 121, + 214, + 165 + ], + "type": "text", + "content": "You are given an image of a male celebrity. Transform the man in the image into a beautiful woman with ponytail hair while preserving her facial identity." + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 63, + 183, + 138, + 259 + ], + "blocks": [ + { + "bbox": [ + 88, + 171, + 112, + 182 + ], + "lines": [ + { + "bbox": [ + 88, + 171, + 112, + 182 + ], + "spans": [ + { + "bbox": [ + 88, + 171, + 112, + 182 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 63, + 183, + 138, + 259 + ], + "lines": [ + { + "bbox": [ + 63, + 183, + 138, + 259 + ], + "spans": [ + { + "bbox": [ + 63, + 183, + 138, + 259 + ], + "type": "image", + "image_path": "9850924b40131c41758c93ce99995792158edeae95609ce6e8686df0f16391ce.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 141, + 183, + 218, + 259 + ], + "blocks": [ + { + "bbox": [ + 167, + 171, + 192, + 181 + ], + "lines": [ + { + "bbox": [ + 167, + 171, + 192, + 181 + ], + "spans": [ + { + "bbox": [ + 167, + 171, + 192, + 181 + ], + "type": "text", + "content": "Result" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 141, + 183, + 218, + 259 + ], + "lines": [ + { + "bbox": [ + 141, + 183, + 218, + 259 + ], + "spans": [ + { + "bbox": [ + 141, + 183, + 218, + 259 + ], + "type": "image", + "image_path": "e5c87d670caae48f57e25f3c959eab525508c1c4f2922472c15c79247d0bd3d9.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 272, + 107, + 338, + 117 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 272, + 107, + 338, + 117 + ], + "spans": [ + { + "bbox": [ + 272, + 107, + 338, + 117 + ], + "type": "text", + "content": "Task Instruction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 227, + 121, + 379, + 165 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 227, + 121, + 379, + 165 + ], + "spans": [ + { + "bbox": [ + 227, + 121, + 379, + 165 + ], + "type": "text", + "content": "You are given an image of a standing cat. Replace the background with a scene of a cozy living room while keeping the lighting and shadows consistent." + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 228, + 183, + 304, + 259 + ], + "blocks": [ + { + "bbox": [ + 253, + 171, + 277, + 182 + ], + "lines": [ + { + "bbox": [ + 253, + 171, + 277, + 182 + ], + "spans": [ + { + "bbox": [ + 253, + 171, + 277, + 182 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 228, + 183, + 304, + 259 + ], + "lines": [ + { + "bbox": [ + 228, + 183, + 304, + 259 + ], + "spans": [ + { + "bbox": [ + 228, + 183, + 304, + 259 + ], + "type": "image", + "image_path": "864ec1fe4c53ac904d443498c676c50f83e2fea4928a0a1b0717d3d9b669389a.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 307, + 183, + 383, + 259 + ], + "blocks": [ + { + "bbox": [ + 331, + 171, + 358, + 181 + ], + "lines": [ + { + "bbox": [ + 331, + 171, + 358, + 181 + ], + "spans": [ + { + "bbox": [ + 331, + 171, + 358, + 181 + ], + "type": "text", + "content": "Result" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 307, + 183, + 383, + 259 + ], + "lines": [ + { + "bbox": [ + 307, + 183, + 383, + 259 + ], + "spans": [ + { + "bbox": [ + 307, + 183, + 383, + 259 + ], + "type": "image", + "image_path": "6766d351c96778144f5b08142df8f83fe284d52258f533f9ecc74f4f4eae225d.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 438, + 107, + 503, + 117 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 438, + 107, + 503, + 117 + ], + "spans": [ + { + "bbox": [ + 438, + 107, + 503, + 117 + ], + "type": "text", + "content": "Task Instruction" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 392, + 121, + 544, + 165 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 392, + 121, + 544, + 165 + ], + "spans": [ + { + "bbox": [ + 392, + 121, + 544, + 165 + ], + "type": "text", + "content": "You are given an image of a red bridge with a person standing on it. Remove the person from the image while maintaining the original appearance of the bridge." + } + ] + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 393, + 183, + 469, + 259 + ], + "blocks": [ + { + "bbox": [ + 418, + 171, + 441, + 182 + ], + "lines": [ + { + "bbox": [ + 418, + 171, + 441, + 182 + ], + "spans": [ + { + "bbox": [ + 418, + 171, + 441, + 182 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 393, + 183, + 469, + 259 + ], + "lines": [ + { + "bbox": [ + 393, + 183, + 469, + 259 + ], + "spans": [ + { + "bbox": [ + 393, + 183, + 469, + 259 + ], + "type": "image", + "image_path": "7851733a1bc670c7d7e0c418aa23d2104fe11fd8035cb220e3ddb13a69d5a6e4.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 471, + 183, + 547, + 259 + ], + "blocks": [ + { + "bbox": [ + 496, + 171, + 523, + 182 + ], + "lines": [ + { + "bbox": [ + 496, + 171, + 523, + 182 + ], + "spans": [ + { + "bbox": [ + 496, + 171, + 523, + 182 + ], + "type": "text", + "content": "Result" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 471, + 183, + 547, + 259 + ], + "lines": [ + { + "bbox": [ + 471, + 183, + 547, + 259 + ], + "spans": [ + { + "bbox": [ + 471, + 183, + 547, + 259 + ], + "type": "image", + "image_path": "9b3d6606dda2e2a66e4db56d8c3891be29311c94e680b0b935dd1e934ae44cf0.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "bbox": [ + 55, + 285, + 295, + 310 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 285, + 295, + 310 + ], + "spans": [ + { + "bbox": [ + 55, + 285, + 295, + 310 + ], + "type": "text", + "content": "better performance, and o1-preview achieves an outstanding performance comparable to ComfyAgent on GPT-4o." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 55, + 322, + 156, + 333 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 322, + 156, + 333 + ], + "spans": [ + { + "bbox": [ + 55, + 322, + 156, + 333 + ], + "type": "text", + "content": "5.4. Ablation Studies" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 55, + 341, + 296, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 341, + 296, + 496 + ], + "spans": [ + { + "bbox": [ + 55, + 341, + 296, + 496 + ], + "type": "text", + "content": "Representation. We conduct an ablation study to verify the advantages of representing workflows with code introduced in Section 4.1. We implement three variants of RAG on GPT-4o, where the workflows are respectively represented in JSON, element list, and code. The variants are evaluated on ComfyBench with the same parameter settings as in Section 5.3. Flow graph is not implemented because it is not understandable by LLM-based agents. We report the ablation results in Table 3, where RAG with code representation shows superior performance over other representations, which verifies our claim. RAG with JSON representation is expected to perform even worse without the unique feature of GPT-4o to generate structured JSON objects." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 55, + 498, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 296, + 714 + ], + "type": "text", + "content": "Architecture. We conduct another ablation to demonstrate the rationality of the multi-agent framework introduced in Section 4.2, where four variants of ComfyAgent on GPT-4o are implemented, each removing one agent from the original framework. We report the ablation results in Table 4. We find that removing any of the four agents will lead to a performance drop. Removing CombineAgent alleviates hallucination in the vanilla category, but leads to failure in the complex category, demonstrating the necessity of integrating multiple demonstrations. Removing AdaptAgent or RetrieveAgent leads to a significant performance drop in all the categories, which underscores the critical role of detail modification and knowledge utilization. Removing RefineAgent leads to a minor but noticeable performance drop in all the categories, indicating its effectiveness in mitigating error accumulation. Therefore, we can conclude that our meticulously-designed multi-agent framework contributes to the superior performance of ComfyAgent." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 314, + 285, + 388, + 297 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 285, + 388, + 297 + ], + "spans": [ + { + "bbox": [ + 314, + 285, + 388, + 297 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 306, + 555, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 306, + 555, + 582 + ], + "spans": [ + { + "bbox": [ + 313, + 306, + 555, + 582 + ], + "type": "text", + "content": "In this work, we investigate the potential of LLM-based agents to autonomously design collaborative AI systems. Building upon the open-source ComfyUI [13] platform, we present ComfyBench, an autonomous workflow design environment serving as both a benchmark for evaluating the capabilities of agents in designing collaborative systems and a means to extend the applicability of ComfyUI in real-world, language-instruction-based generative applications. We further develop ComfyAgent, which achieves a comparable resolve rate to o1-preview [39] and significantly surpasses other agents on ComfyBench. We point out two key insights underpinning ComfyAgent. First, representing workflows through code is efficient, which provides rich semantics for agents to better understand complex systems. Second, the multi-agent framework of ComfyAgent highlights the substantial benefits of task decomposition and collaborative generation. However, the top-performing agent, ComfyAgent, has achieved a resolve rate of only " + }, + { + "bbox": [ + 313, + 306, + 555, + 582 + ], + "type": "inline_equation", + "content": "15\\%" + }, + { + "bbox": [ + 313, + 306, + 555, + 582 + ], + "type": "text", + "content": " for creative tasks on ComfyBench. LLM-based agents still have a long way to go in designing collaborative AI systems. We hope that ComfyBench and ComfyAgent can serve as valuable resources in advancing the development of intelligent and autonomous collaborative AI systems." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 582, + 555, + 678 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 555, + 678 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 555, + 678 + ], + "type": "text", + "content": "Limitations. Well-designed fine-tuning methods could potentially elevate agents to an expert level on ComfyBench, but they are quite expensive and thus not involved in our work. In addition, ComfyAgent operates in an open-loop manner. An alternative approach would involve using VLMs as evaluators to provide feedback [21], enabling a closed-loop system where agents can refine their workflows based on evaluations, thus leading to better results." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 679, + 555, + 703 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 679, + 555, + 703 + ], + "spans": [ + { + "bbox": [ + 313, + 679, + 555, + 703 + ], + "type": "text", + "content": "Acknowledgments. This work is supported by the Shanghai Municipal Science and Technology Major Project." + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "24621" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 61, + 91, + 293, + 112 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 293, + 112 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 293, + 112 + ], + "type": "text", + "content": "[1] Salt AI. Salt node documentation. https://docs. getsalt.ai/md/, 2024.3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 114, + 295, + 157 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 114, + 295, + 157 + ], + "spans": [ + { + "bbox": [ + 61, + 114, + 295, + 157 + ], + "type": "text", + "content": "[2] Google DeepMind AlphaCode Team. Alphacode 2 technical report. https://storage.googleapis.com/deepmind-media/AlphaCode2/AlphaCode2_Tech_Report.pdf, 2024.1,3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 159, + 294, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 159, + 294, + 190 + ], + "spans": [ + { + "bbox": [ + 62, + 159, + 294, + 190 + ], + "type": "text", + "content": "[3] Anthropic. Claude 3.5 sonnet. https://www.anthropic.com/news/claude-3-5-sonnet, 2024.6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 192, + 295, + 223 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 192, + 295, + 223 + ], + "spans": [ + { + "bbox": [ + 62, + 192, + 295, + 223 + ], + "type": "text", + "content": "[4] Richard C Atkinson. Human memory: A proposed system and its control processes. The psychology of learning and motivation, 1968. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 225, + 264, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 225, + 264, + 236 + ], + "spans": [ + { + "bbox": [ + 62, + 225, + 264, + 236 + ], + "type": "text", + "content": "[5] Alan Baddeley. Working memory. Science, 1992. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 237, + 294, + 291 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 237, + 294, + 291 + ], + "spans": [ + { + "bbox": [ + 62, + 237, + 294, + 291 + ], + "type": "text", + "content": "[6] Yuntao Bai, Andy Jones, Kamal Ndousse, Amanda Askell, Anna Chen, Nova DasSarma, Dawn Drain, Stanislav Fort, Deep Ganguli, Tom Henighan, et al. Training a helpful and harmless assistant with reinforcement learning from human feedback. arXiv preprint arXiv:2204.05862, 2022. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 293, + 294, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 293, + 294, + 346 + ], + "spans": [ + { + "bbox": [ + 62, + 293, + 294, + 346 + ], + "type": "text", + "content": "[7] Maciej Besta, Nils Blach, Ales Kubicek, Robert Gerstenberger, Michal Podstawski, Lukas Gianinazzi, Joanna Gajda, Tomasz Lehmann, Hubert Niewiadomski, Piotr Nczyk, et al. Graph of thoughts: Solving elaborate problems with large language models. In AAAI, 2024. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 348, + 294, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 348, + 294, + 422 + ], + "spans": [ + { + "bbox": [ + 62, + 348, + 294, + 422 + ], + "type": "text", + "content": "[8] Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Li Jing, David Schnurr, Joe Taylor, Troy Luhman, Eric Luhman, Clarence Ng, Ricky Wang, and Aditya Ramesh. Video generation models as world simulators. https://openai.com/research/video-generation-models-as-world-simulators, 2024.3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 425, + 294, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 425, + 294, + 468 + ], + "spans": [ + { + "bbox": [ + 62, + 425, + 294, + 468 + ], + "type": "text", + "content": "[9] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. In NeurIPS, 2020. 2, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 470, + 294, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 470, + 294, + 533 + ], + "spans": [ + { + "bbox": [ + 57, + 470, + 294, + 533 + ], + "type": "text", + "content": "[10] Ruisheng Cao, Fangyu Lei, Haoyuan Wu, Jixuan Chen, Yeqiao Fu, Hongcheng Gao, Xinzhuang Xiong, Hanchong Zhang, Yuchen Mao, Wenjing Hu, et al. Spider2-v: How far are multimodal agents from automating data science and engineering workflows? arXiv preprint arXiv:2407.10956, 2024.3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 536, + 294, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 536, + 294, + 590 + ], + "spans": [ + { + "bbox": [ + 57, + 536, + 294, + 590 + ], + "type": "text", + "content": "[11] Weize Chen, Yusheng Su, Jingwei Zuo, Cheng Yang, Chenfei Yuan, Chen Qian, Chi-Min Chan, Yujia Qin, Yaxi Lu, Ruobing Xie, et al. Agentverse: Facilitating multi-agent collaboration and exploring emergent behaviors in agents. arXiv preprint arXiv:2308.10848, 2023. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 592, + 294, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 592, + 294, + 635 + ], + "spans": [ + { + "bbox": [ + 57, + 592, + 294, + 635 + ], + "type": "text", + "content": "[12] Zehui Chen, Kuikun Liu, Qiuchen Wang, Jiangning Liu, Wenwei Zhang, Kai Chen, and Feng Zhao. Mindsearch: Mimicking human minds elicits deep ai searcher. arXiv preprint arXiv:2407.20183, 2024. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 636, + 294, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 636, + 294, + 657 + ], + "spans": [ + { + "bbox": [ + 57, + 636, + 294, + 657 + ], + "type": "text", + "content": "[13] comfyanonymous. Comfyui. https://github.com/comfyanonymous/ComfyUI, 2023. 1, 3, 5, 8" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 658, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 658, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 57, + 658, + 294, + 689 + ], + "type": "text", + "content": "[14] Xiang Deng, Yu Gu, Boyuan Zheng, Shijie Chen, Sam Stevens, Boshi Wang, Huan Sun, and Yu Su. Mind2web: Towards a generalist agent for the web. In NeurIPS, 2024. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 692, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 692, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 692, + 294, + 712 + ], + "type": "text", + "content": "[15] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 74, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 335, + 74, + 553, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 74, + 553, + 105 + ], + "spans": [ + { + "bbox": [ + 335, + 74, + 553, + 105 + ], + "type": "text", + "content": "Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 107, + 553, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 107, + 553, + 149 + ], + "spans": [ + { + "bbox": [ + 317, + 107, + 553, + 149 + ], + "type": "text", + "content": "[16] Yunfan Gao, Yun Xiong, Xinyu Gao, Kangxiang Jia, Jinliu Pan, Yuxi Bi, Yi Dai, Jiawei Sun, and Haofen Wang. Retrieval-augmented generation for large language models: A survey. arXiv preprint arXiv:2312.10997, 2023. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 151, + 553, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 151, + 553, + 183 + ], + "spans": [ + { + "bbox": [ + 317, + 151, + 553, + 183 + ], + "type": "text", + "content": "[17] Yingqiang Ge, Wenyue Hua, Kai Mei, Juntao Tan, Shuyuan Xu, Zelong Li, Yongfeng Zhang, et al. Openagi: When llm meets domain experts. In NeurlPS, 2024. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 185, + 553, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 185, + 553, + 227 + ], + "spans": [ + { + "bbox": [ + 317, + 185, + 553, + 227 + ], + "type": "text", + "content": "[18] Jia Guo, Jiankang Deng, Xiang An, Jack Yu, and Baris Gecer. Insightface: 2d and 3d face analysis project. https://github.com/deepinsight/insightface, 2023.3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 228, + 553, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 228, + 553, + 281 + ], + "spans": [ + { + "bbox": [ + 317, + 228, + 553, + 281 + ], + "type": "text", + "content": "[19] Taicheng Guo, Xiuying Chen, Yaqi Wang, Ruidi Chang, Shichao Pei, Nitesh V Chawla, Olaf Wiest, and Xiangliang Zhang. Large language model based multi-agents: A survey of progress and challenges. arXiv preprint arXiv:2402.01680, 2024. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 283, + 553, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 283, + 553, + 337 + ], + "spans": [ + { + "bbox": [ + 317, + 283, + 553, + 337 + ], + "type": "text", + "content": "[20] Sirui Hong, Xiawu Zheng, Jonathan Chen, Yuheng Cheng, Jinlin Wang, Ceyao Zhang, Zili Wang, Steven Ka Shing Yau, Zijuan Lin, Liyang Zhou, et al. Metagpt: Meta programming for multi-agent collaborative framework. arXiv preprint arXiv:2308.00352, 2023. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 338, + 553, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 338, + 553, + 369 + ], + "spans": [ + { + "bbox": [ + 317, + 338, + 553, + 369 + ], + "type": "text", + "content": "[21] Shengran Hu, Cong Lu, and Jeff Clune. Automated design of agentic systems. arXiv preprint arXiv:2408.08435, 2024. 2, 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 372, + 553, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 372, + 553, + 403 + ], + "spans": [ + { + "bbox": [ + 317, + 372, + 553, + 403 + ], + "type": "text", + "content": "[22] Ian Huang, Guandao Yang, and Leonidas Guibas. Blender alchemy: Editing 3d graphics with vision-language models. arXiv preprint arXiv:2404.17672, 2024. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 405, + 553, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 405, + 553, + 437 + ], + "spans": [ + { + "bbox": [ + 317, + 405, + 553, + 437 + ], + "type": "text", + "content": "[23] Timo Kaufmann, Paul Weng, Viktor Bengs, and Eyke Hüllermeier. A survey of reinforcement learning from human feedback. arXiv preprint arXiv:2312.14925, 2023. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 438, + 553, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 438, + 553, + 502 + ], + "spans": [ + { + "bbox": [ + 317, + 438, + 553, + 502 + ], + "type": "text", + "content": "[24] Omar Khattab, Arnav Singhvi, Paridhi Maheshwari, Zhiyuan Zhang, Keshav Santhanam, Sri Vardhamanan, Saiful Haq, Ashutosh Sharma, Thomas T Joshi, Hanna Moazam, et al. Dspy: Compiling declarative language model calls into self-improving pipelines. arXiv preprint arXiv:2310.03714, 2023. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 504, + 553, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 504, + 553, + 525 + ], + "spans": [ + { + "bbox": [ + 316, + 504, + 553, + 525 + ], + "type": "text", + "content": "[25] Diederik P Kingma. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 526, + 553, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 526, + 553, + 580 + ], + "spans": [ + { + "bbox": [ + 317, + 526, + 553, + 580 + ], + "type": "text", + "content": "[26] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dólar, and Ross Girshick. Segment anything. arXiv preprint arXiv:2304.02643, 2023. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 317, + 582, + 553, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 582, + 553, + 635 + ], + "spans": [ + { + "bbox": [ + 317, + 582, + 553, + 635 + ], + "type": "text", + "content": "[27] Jing Yu Koh, Robert Lo, Lawrence Jang, Vikram Duvvur, Ming Chong Lim, Po-Yu Huang, Graham Neubig, Shuyan Zhou, Ruslan Salakhutdinov, and Daniel Fried. Visualwebarena: Evaluating multimodal agents on realistic visual web tasks. arXiv preprint arXiv:2401.13649, 2024. 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 317, + 636, + 553, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 636, + 553, + 658 + ], + "spans": [ + { + "bbox": [ + 317, + 636, + 553, + 658 + ], + "type": "text", + "content": "[28] LangChain. Applications that can reason. powered by langchain. https://www.langchain.com/, 2023.3" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 317, + 659, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 659, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 317, + 659, + 553, + 712 + ], + "type": "text", + "content": "[29] Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Kuttler, Mike Lewis, Wen-tau Yih, Tim Rocktäschel, et al. Retrieval-augmented generation for knowledge-intensive nlp tasks. In NeurIPS, 2020. 2, 3, 7" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "24622" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "type": "text", + "content": "[30] Guohao Li, Hasan Hammoud, Hani Itani, Dmitrii Khizbullin, and Bernard Ghanem. Camel: Communicative agents for\" mind\" exploration of large language model society. In NeurIPS, 2023. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 172 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 172 + ], + "type": "text", + "content": "[31] Zelong Li, Shuyuan Xu, Kai Mei, Wenyue Hua, Balaji Rama, Om Raheja, Hao Wang, He Zhu, and Yongfeng Zhang. Autoflow: Automated workflow generation for large language model agents. arXiv preprint arXiv:2407.12821, 2024. 1, 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 174, + 294, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 174, + 294, + 228 + ], + "spans": [ + { + "bbox": [ + 56, + 174, + 294, + 228 + ], + "type": "text", + "content": "[32] Tian Liang, Zhiwei He, Wenxiang Jiao, Xing Wang, Yan Wang, Rui Wang, Yujiu Yang, Shuming Shi, and Zhaopeng Tu. Encouraging divergent thinking in large language models through multi-agent debate. arXiv preprint arXiv:2305.19118, 2023. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 230, + 294, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 230, + 294, + 274 + ], + "spans": [ + { + "bbox": [ + 56, + 230, + 294, + 274 + ], + "type": "text", + "content": "[33] Aman Madaan, Niket Tandon, Prakhar Gupta, Skyler Hallinan, Luyu Gao, Sarah Wiegreffe, Uri Alon, Nouha Dziri, Shrimai Prabhumoye, Yiming Yang, et al. Self-refine: Iterative refinement with self-feedback. In NeurIPS, 2024. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 275, + 294, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 275, + 294, + 308 + ], + "spans": [ + { + "bbox": [ + 56, + 275, + 294, + 308 + ], + "type": "text", + "content": "[34] Kai Mei, Zelong Li, Shuyuan Xu, Ruosong Ye, Yingqiang Ge, and Yongfeng Zhang. LIm agent operating system. arXiv preprint arXiv:2403.16971, 2024. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 309, + 294, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 309, + 294, + 363 + ], + "spans": [ + { + "bbox": [ + 56, + 309, + 294, + 363 + ], + "type": "text", + "content": "[35] Reiichiro Nakano, Jacob Hilton, Suchir Balaji, Jeff Wu, Long Ouyang, Christina Kim, Christopher Hesse, Shantanu Jain, Vineet Kosaraju, William Saunders, et al. Webgpt: Browser-assisted question-answering with human feedback. arXiv preprint arXiv:2112.09332, 2021. 1, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 365, + 294, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 365, + 294, + 419 + ], + "spans": [ + { + "bbox": [ + 56, + 365, + 294, + 419 + ], + "type": "text", + "content": "[36] Harsha Nori, Yin Tat Lee, Sheng Zhang, Dean Carignan, Richard Edgar, Nicolo Fusi, Nicholas King, Jonathan Larson, Yuanzhi Li, Weishung Liu, et al. Can generalist foundation models outcompete special-purpose tuning? case study in medicine. arXiv preprint arXiv:2311.16452, 2023. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 421, + 294, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 421, + 294, + 442 + ], + "spans": [ + { + "bbox": [ + 56, + 421, + 294, + 442 + ], + "type": "text", + "content": "[37] OpenAI. Chatgpt. https://openai.com/chatgpt/overview/, 2024.1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 445, + 294, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 445, + 294, + 465 + ], + "spans": [ + { + "bbox": [ + 56, + 445, + 294, + 465 + ], + "type": "text", + "content": "[38] OpenAI. Hello gpt-4o. https://openai.com/index/hello-gpt-4o/, 2024.6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 468, + 294, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 468, + 294, + 499 + ], + "spans": [ + { + "bbox": [ + 56, + 468, + 294, + 499 + ], + "type": "text", + "content": "[39] OpenAI. Introducing openai ol-preview. https://openai.com/index/introducing-openai-ol-preview/, 2024.2,6,8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 501, + 294, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 501, + 294, + 533 + ], + "spans": [ + { + "bbox": [ + 56, + 501, + 294, + 533 + ], + "type": "text", + "content": "[40] OpenAI. New embedding models and api updates. https://openai.com/index/new-embedding-models-and-api-updates/, 2024.7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 535, + 294, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 535, + 294, + 589 + ], + "spans": [ + { + "bbox": [ + 56, + 535, + 294, + 589 + ], + "type": "text", + "content": "[41] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. In NeurIPS, 2022. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 590, + 294, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 590, + 294, + 644 + ], + "spans": [ + { + "bbox": [ + 56, + 590, + 294, + 644 + ], + "type": "text", + "content": "[42] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 647, + 294, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 294, + 679 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 294, + 679 + ], + "type": "text", + "content": "[43] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In ICML, 2021. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 681, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 681, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 681, + 294, + 712 + ], + "type": "text", + "content": "[44] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 3" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 714 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 106 + ], + "type": "text", + "content": "[45] Simo Ryu. Low-rank adaptation for fast text-to-image diffusion fine-tuning. https://github.com/cloneofsimo/lora, 2023.3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 107, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 150 + ], + "type": "text", + "content": "[46] Yongliang Shen, Kaitao Song, Xu Tan, Dongsheng Li, Weiming Lu, and Yueting Zhuang. Hugginggpt: Solving ai tasks with chatgpt and its friends in hugging face. In NeurIPS, 2024. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 152, + 553, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 553, + 195 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 553, + 195 + ], + "type": "text", + "content": "[47] Chan Hee Song, Jiaman Wu, Clayton Washington, Brian M Sadler, Wei-Lun Chao, and Yu Su. Llm-planner: Few-shot grounded planning for embodied agents with large language models. In CVPR, 2023. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 197, + 553, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 197, + 553, + 230 + ], + "spans": [ + { + "bbox": [ + 316, + 197, + 553, + 230 + ], + "type": "text", + "content": "[48] Theodore R Sumers, Shunyu Yao, Karthik Narasimhan, and Thomas L Griffiths. Cognitive architectures for language agents. arXiv preprint arXiv:2309.02427, 2023. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 232, + 553, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 232, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 316, + 232, + 553, + 285 + ], + "type": "text", + "content": "[49] Gemini Team, Rohan Anil, Sebastian Borgeaud, Yonghui Wu, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, et al. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805, 2023. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 287, + 553, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 287, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 316, + 287, + 553, + 319 + ], + "type": "text", + "content": "[50] Amitayush Thakur, Yeming Wen, and Swarat Chaudhuri. A language-agent approach to formal theorem-proving. arXiv preprint arXiv:2310.04353, 2023. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 321, + 553, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 321, + 553, + 354 + ], + "spans": [ + { + "bbox": [ + 316, + 321, + 553, + 354 + ], + "type": "text", + "content": "[51] Trieu H Trinh, Yuhuai Wu, Quoc V Le, He He, and Thang Luong. Solving olympiad geometry without human demonstrations. Nature, 2024. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 355, + 553, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 355, + 553, + 399 + ], + "spans": [ + { + "bbox": [ + 316, + 355, + 553, + 399 + ], + "type": "text", + "content": "[52] Lei Wang, Chen Ma, Xueyang Feng, Zeyu Zhang, Hao Yang, Jingsen Zhang, Zhiyuan Chen, Jiakai Tang, Xu Chen, Yankai Lin, et al. A survey on large language model based autonomous agents. Frontiers of Computer Science, 2024. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 400, + 553, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 400, + 553, + 453 + ], + "spans": [ + { + "bbox": [ + 316, + 400, + 553, + 453 + ], + "type": "text", + "content": "[53] Xuezhi Wang, Jason Wei, Dale Schuurmans, Quoc Le, Ed Chi, Sharan Narang, Aakanksha Chowdhery, and Denny Zhou. Self-consistency improves chain of thought reasoning in language models. arXiv preprint arXiv:2203.11171, 2022. 2, 3, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 456, + 553, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 456, + 553, + 498 + ], + "spans": [ + { + "bbox": [ + 316, + 456, + 553, + 498 + ], + "type": "text", + "content": "[54] Xingyao Wang, Yangyi Chen, Lifan Yuan, Yizhe Zhang, Yunzhu Li, Hao Peng, and Heng Ji. Executable code actions elicit better lIm agents. arXiv preprint arXiv:2402.01030, 2024. 3, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 501, + 553, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 501, + 553, + 555 + ], + "spans": [ + { + "bbox": [ + 316, + 501, + 553, + 555 + ], + "type": "text", + "content": "[55] Xingyao Wang, Boxuan Li, Yufan Song, Frank F Xu, Xiangru Tang, Mingchen Zhuge, Jiayi Pan, Yueqi Song, Bowen Li, Jaskirat Singh, et al. Opendevin: An open platform for ai software developers as generalist agents. arXiv preprint arXiv:2407.16741, 2024. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 557, + 553, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 557, + 553, + 589 + ], + "spans": [ + { + "bbox": [ + 316, + 557, + 553, + 589 + ], + "type": "text", + "content": "[56] Zhiruo Wang, Shuyan Zhou, Daniel Fried, and Graham Neubig. Execution-based evaluation for open-domain code generation. arXiv preprint arXiv:2212.10481, 2022. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 590, + 553, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 590, + 553, + 634 + ], + "spans": [ + { + "bbox": [ + 316, + 590, + 553, + 634 + ], + "type": "text", + "content": "[57] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. In NeurIPS, 2022. 2, 3, 7" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 635, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 553, + 689 + ], + "type": "text", + "content": "[58] Qingyun Wu, Gagan Bansal, Jieyu Zhang, Yiran Wu, Shaokun Zhang, Erkang Zhu, Beibin Li, Li Jiang, Xiaoyun Zhang, and Chi Wang. Autogen: Enabling next-gen llm applications via multi-agent conversation framework. arXiv preprint arXiv:2308.08155, 2023. 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 692, + 553, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 553, + 714 + ], + "type": "text", + "content": "[59] Zhiyong Wu, Chengcheng Han, Zichen Ding, Zhenmin Weng, Zhoumianze Liu, Shunyu Yao, Tao Yu, and Lingpeng" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24623" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 105 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 105 + ], + "type": "text", + "content": "Kong. Os-copilot: Towards generalist computer agents with self-improvement. arXiv preprint arXiv:2402.07456, 2024. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 106, + 295, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 106, + 295, + 160 + ], + "spans": [ + { + "bbox": [ + 56, + 106, + 295, + 160 + ], + "type": "text", + "content": "[60] Zhiheng Xi, Wenxiang Chen, Xin Guo, Wei He, Yiwen Ding, Boyang Hong, Ming Zhang, Junzhe Wang, Senjie Jin, Enyu Zhou, et al. The rise and potential of large language model based agents: A survey. arXiv preprint arXiv:2309.07864, 2023. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 161, + 295, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 161, + 295, + 216 + ], + "spans": [ + { + "bbox": [ + 56, + 161, + 295, + 216 + ], + "type": "text", + "content": "[61] Zhiheng Xi, Yiwen Ding, Wenxiang Chen, Boyang Hong, Honglin Guo, Junzhe Wang, Dingwen Yang, Chenyang Liao, Xin Guo, Wei He, et al. Agentgym: Evolving large language model-based agents across diverse environments. arXiv preprint arXiv:2406.04151, 2024. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 217, + 294, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 217, + 294, + 270 + ], + "spans": [ + { + "bbox": [ + 56, + 217, + 294, + 270 + ], + "type": "text", + "content": "[62] Tianbao Xie, Fan Zhou, Zhoujun Cheng, Peng Shi, Luoxuan Weng, Yitao Liu, Toh Jing Hua, Junning Zhao, Qian Liu, Che Liu, et al. Openagents: An open platform for language agents in the wild. arXiv preprint arXiv:2310.10634, 2023. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 271, + 295, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 271, + 295, + 326 + ], + "spans": [ + { + "bbox": [ + 56, + 271, + 295, + 326 + ], + "type": "text", + "content": "[63] Tianbao Xie, Danyang Zhang, Jixuan Chen, Xiaochuan Li, Siheng Zhao, Ruisheng Cao, Toh Jing Hua, Zhoujun Cheng, Dongchan Shin, Fangyu Lei, et al. Osworld: Benchmarking multimodal agents for open-ended tasks in real computer environments. arXiv preprint arXiv:2404.07972, 2024. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 327, + 295, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 327, + 295, + 370 + ], + "spans": [ + { + "bbox": [ + 56, + 327, + 295, + 370 + ], + "type": "text", + "content": "[64] Fangzhi Xu, Qiushi Sun, Kanzhi Cheng, Jun Liu, Yu Qiao, and Zhiyong Wu. Interactive evolution: A neural-symbolic self-training framework for large language models. arXiv preprint arXiv:2406.11736, 2024. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 371, + 295, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 371, + 295, + 415 + ], + "spans": [ + { + "bbox": [ + 56, + 371, + 295, + 415 + ], + "type": "text", + "content": "[65] Shuyuan Xu, Zelong Li, Kai Mei, and Yongfeng Zhang. Core: LIm as interpreter for natural language programming, pseudo-code programming, and flow programming of ai agents. arXiv preprint arXiv:2405.06907, 2024. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 415, + 295, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 415, + 295, + 468 + ], + "spans": [ + { + "bbox": [ + 56, + 415, + 295, + 468 + ], + "type": "text", + "content": "[66] John Yang, Carlos E Jimenez, Alexander Wettig, Kilian Lieret, Shunyu Yao, Karthik Narasimhan, and Ofir Press. Swe-agent: Agent-computer interfaces enable automated software engineering. arXiv preprint arXiv:2405.15793, 2024.3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 470, + 295, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 470, + 295, + 513 + ], + "spans": [ + { + "bbox": [ + 56, + 470, + 295, + 513 + ], + "type": "text", + "content": "[67] John Yang, Akshara Prabhakar, Karthik Narasimhan, and Shunyu Yao. Intercode: Standardizing and benchmarking interactive coding with execution feedback. In NeurIPS, 2024. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 514, + 295, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 514, + 295, + 557 + ], + "spans": [ + { + "bbox": [ + 56, + 514, + 295, + 557 + ], + "type": "text", + "content": "[68] Zhao Yang, Jiaxuan Liu, Yucheng Han, Xin Chen, Zebiao Huang, Bin Fu, and Gang Yu. Appagent: Multimodal agents as smartphone users. arXiv preprint arXiv:2312.13771, 2023.3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 558, + 295, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 558, + 295, + 601 + ], + "spans": [ + { + "bbox": [ + 56, + 558, + 295, + 601 + ], + "type": "text", + "content": "[69] Shunyu Yao, Howard Chen, John Yang, and Karthik Narasimhan. Webshop: Towards scalable real-world web interaction with grounded language agents. In NeurIPS, 2022. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 602, + 295, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 602, + 295, + 646 + ], + "spans": [ + { + "bbox": [ + 56, + 602, + 295, + 646 + ], + "type": "text", + "content": "[70] Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, and Yuan Cao. React: Synergizing reasoning and acting in language models. arXiv preprint arXiv:2210.03629, 2022. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 647, + 295, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 295, + 689 + ], + "type": "text", + "content": "[71] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 690, + 295, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 690, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 690, + 295, + 714 + ], + "type": "text", + "content": "[72] Pengcheng Yin, Wen-Ding Li, Kefan Xiao, Abhishek Rao, Yeming Wen, Kensen Shi, Joshua Howland, Paige Bailey," + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 609 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 333, + 72, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 72, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 333, + 72, + 553, + 106 + ], + "type": "text", + "content": "Michele Catasta, Henryk Michalewski, et al. Natural language to code generation in interactive data science notebooks. arXiv preprint arXiv:2212.09248, 2022. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 107, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 150 + ], + "type": "text", + "content": "[73] Zhengqing Yuan, Ruoxi Chen, Zhaoxu Li, Haolong Jia, Lifang He, Chi Wang, and Lichao Sun. Mora: Enabling generalist video generation via a multi-agent framework. arXiv preprint arXiv:2403.13248, 2024. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 152, + 553, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 553, + 217 + ], + "type": "text", + "content": "[74] Matei Zaharia, Omar Khattab, Lingjiao Chen, Jared Quincy Davis, Heather Miller, Chris Potts, James Zou, Michael Carbin, Jonathan Frankle, Naveen Rao, and Ali Ghodsi. The shift from models to compound ai systems. https://bair.berkeley.edu/blog/2024/02/18/compound-ai-systems/, 2024.1,3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 219, + 553, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 219, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 316, + 219, + 553, + 251 + ], + "type": "text", + "content": "[75] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In CVPR, 2023. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 252, + 553, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 252, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 316, + 252, + 553, + 285 + ], + "type": "text", + "content": "[76] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Imposing consistent light. https://github.com/lllyasviel/IC-Light, 2024.3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 286, + 553, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 286, + 553, + 329 + ], + "spans": [ + { + "bbox": [ + 316, + 286, + 553, + 329 + ], + "type": "text", + "content": "[77] Wenqi Zhang, Yongliang Shen, Weiming Lu, and Yueting Zhuang. Data-copilot: Bridging billions of data and humans with autonomous workflow. arXiv preprint arXiv:2306.07209, 2023. 3, 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 331, + 553, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 331, + 553, + 374 + ], + "spans": [ + { + "bbox": [ + 316, + 331, + 553, + 374 + ], + "type": "text", + "content": "[78] Zeyu Zhang, Xiaohe Bo, Chen Ma, Rui Li, Xu Chen, Quanyu Dai, Jieming Zhu, Zhenhua Dong, and Ji-Rong Wen. A survey on the memory mechanism of large language model based agents. arXiv preprint arXiv:2404.13501, 2024. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 376, + 553, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 376, + 553, + 409 + ], + "spans": [ + { + "bbox": [ + 316, + 376, + 553, + 409 + ], + "type": "text", + "content": "[79] Wanjun Zhong, Lianghong Guo, Qiqi Gao, He Ye, and Yanlin Wang. Memorybank: Enhancing large language models with long-term memory. In AAAI, 2024. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 410, + 553, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 410, + 553, + 464 + ], + "spans": [ + { + "bbox": [ + 316, + 410, + 553, + 464 + ], + "type": "text", + "content": "[80] Shuyan Zhou, Frank F Xu, Hao Zhu, Xuhui Zhou, Robert Lo, Abishek Sridhar, Xianyi Cheng, Tianyue Ou, Yonatan Bisk, Daniel Fried, et al. Webarena: A realistic web environment for building autonomous agents. arXiv preprint arXiv:2307.13854, 2023. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 465, + 553, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 465, + 553, + 520 + ], + "spans": [ + { + "bbox": [ + 316, + 465, + 553, + 520 + ], + "type": "text", + "content": "[81] Wangchunshu Zhou, Yuchen Eleanor Jiang, Long Li, Jia-long Wu, Tiannan Wang, Shi Qiu, Jintian Zhang, Jing Chen, Ruipu Wu, Shuai Wang, et al. Agents: An open-source framework for autonomous language agents. arXiv preprint arXiv:2309.07870, 2023. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 521, + 553, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 521, + 553, + 565 + ], + "spans": [ + { + "bbox": [ + 316, + 521, + 553, + 565 + ], + "type": "text", + "content": "[82] Wangchunshu Zhou, Yixin Ou, Shengwei Ding, Long Li, Jialong Wu, Tiannan Wang, Jiamin Chen, Shuai Wang, Xiaohua Xu, Ningyu Zhang, et al. Symbolic learning enables self-evolving agents. arXiv preprint arXiv:2406.18532, 2024. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 566, + 553, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 566, + 553, + 609 + ], + "spans": [ + { + "bbox": [ + 316, + 566, + 553, + 609 + ], + "type": "text", + "content": "[83] Mingchen Zhuge, Wenyi Wang, Louis Kirsch, Francesco Faccio, Dmitrii Khizbullin, and Jürgen Schmidhuber. Gptswarm: Language agents as estimizable graphs. In ICML, 2024. 3" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "24624" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Common3D_ Self-Supervised Learning of 3D Morphable Models for Common Objects in Neural Feature Space/ecdc4e97-8d79-41e2-abfb-5c0cebf6aa2a_content_list.json b/2025/Common3D_ Self-Supervised Learning of 3D Morphable Models for Common Objects in Neural Feature Space/ecdc4e97-8d79-41e2-abfb-5c0cebf6aa2a_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..d555a2d96e14ca18a939e18fc433750d2413ef8f --- /dev/null +++ b/2025/Common3D_ Self-Supervised Learning of 3D Morphable Models for Common Objects in Neural Feature Space/ecdc4e97-8d79-41e2-abfb-5c0cebf6aa2a_content_list.json @@ -0,0 +1,1571 @@ +[ + { + "type": "text", + "text": "Common3D: Self-Supervised Learning of 3D Morphable Models for Common Objects in Neural Feature Space", + "text_level": 1, + "bbox": [ + 174, + 130, + 823, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Leonhard Sommer", + "bbox": [ + 130, + 202, + 292, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Olaf Dunkel", + "bbox": [ + 333, + 202, + 447, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Christian Theobalt 2", + "bbox": [ + 488, + 202, + 650, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Adam Kortylewski 1,2", + "bbox": [ + 691, + 202, + 864, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1University of Freiburg", + "bbox": [ + 243, + 237, + 431, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Max Planck Institute for Informatics", + "bbox": [ + 450, + 237, + 746, + 256 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/01a12ec6fb8d4c297aa800494cbbf4a7addfdc1f272cef0f4fcf3164415677bc.jpg", + "image_caption": [ + "Figure 1. Common3D learns category-specific 3D morphable models from few casually captured videos completely self-supervised, and can estimate the 3D object shape (visualized from two viewpoints), 2D-3D correspondences, and the 3D object pose via inverse rendering." + ], + "image_footnote": [], + "bbox": [ + 93, + 290, + 901, + 510 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 560, + 326, + 575 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3D morphable models (3DMMs) are a powerful tool to represent the possible shapes and appearances of an object category. Given a single test image, 3DMMs can be used to solve various tasks, such as predicting the 3D shape, pose, semantic correspondence, and instance segmentation of an object. Unfortunately, 3DMMs are only available for very few object categories that are of particular interest, like faces or human bodies, as they require a demanding 3D data acquisition and category-specific training process. In contrast, we introduce a new method, Common3D, that learns 3DMMs of common objects in a fully self-supervised manner from a collection of object-centric videos. For this purpose, our model represents objects as a learned 3D template mesh and a deformation field that is parameterized as an image-conditioned neural network. Different from prior works, Common3D represents the object appearance with neural features instead of RGB colors, which enables the learning of more generalizable representations through an abstraction from pixel intensities. Importantly, we train the appearance features using a contrastive", + "bbox": [ + 88, + 598, + 485, + 902 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "objective by exploiting the correspondences defined through the deformable template mesh. This leads to higher quality correspondence features compared to related works and a significantly improved model performance at estimating 3D object pose and semantic correspondence. Common3D is the first completely self-supervised method that can solve various vision tasks in a zero-shot manner. We release all code at github.com/GenIntel/common3d.", + "bbox": [ + 511, + 561, + 908, + 684 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 707, + 645, + 723 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Estimating the 3D shape and pose of an object from a single image requires prior knowledge about the possible shapes and appearances of an object category. Learning such prior models is a long-standing problem in computer vision, with 3D morphable models (3DMMs) [5] emerging as a very successful approach to address this problem. A popular application of 3DMMs is to perform single image 3D reconstruction via inverse rendering [1, 5, 55], also referred to as analysis-by-synthesis approach [77]. This enables solving under-constrained vision tasks, for example predicting an object's shape behind an occluder [11] or estimating the", + "bbox": [ + 511, + 734, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "6468", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3D object shape and pose from a single image. However, training 3DMMs requires sophisticated data acquisition setups [3, 21, 33], involving laser scanners or multi-view studio captures, and a significant manual effort in cleaning and registration of the data [5, 34, 84]. This approach is difficult to scale to more common everyday objects.", + "bbox": [ + 89, + 90, + 480, + 181 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "An alternative approach that emerged recently is to learn a 3D prior from 2D images only [58, 67, 74, 81]. However, learning such priors from in-the-wild images is a major challenge and typically requires significant category-specific design choices, with various solutions having been proposed for different object categories [23, 32, 68, 69, 81]. This makes it difficult to scale to the long tail of common objects that can appear in natural images.", + "bbox": [ + 88, + 184, + 480, + 305 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We introduce Common3D, a new way of integrating self-supervised learning, neural rendering, and 3D modeling to learn 3D object priors for common objects from a collection of object-centric videos. Once trained, the model can estimate 2D-3D correspondences, it can be leveraged for inverse rendering to estimate the 3D pose, and instance segmentation of common objects from a single image, see Fig. 1. At the core of our method is a category-specific 3D morphable model (3DMM) that represents the possible variations of different object instances. The 3DMM is represented using a learned prototypical 3D shape and a shape deformation field that is parameterized as an image-conditioned neural network. Training our model requires canonical camera poses, which we estimate unsupervised by adopting the method of [52].", + "bbox": [ + 91, + 306, + 482, + 532 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "One core challenge for learning 3DMMs of common objects is to be flexible enough to capture the highly variable shapes and topologies of various object categories, as well as the possible variations of instances within one category. We address this issue by using a hybrid volumetric mesh representation based on DMTet [50]. Hence, the shape is defined volumetrically in canonical space, but a mesh is extracted on the fly for morphing and rendering the model. This allows to retain the flexibility of implicit neural representations for capturing highly variable shapes and topologies [8, 36, 42], while also enabling the use of meshes for an efficient rendering and regularization during training.", + "bbox": [ + 91, + 536, + 482, + 717 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Another important challenge is to perform inverse rendering with 3D object priors, as it requires the optimization of a highly non-convex image-based reconstruction loss [46, 49, 55]. We address this problem in Common3D by representing the object appearance as a semantic feature field instead of RGB colors. Performing inverse rendering on the feature-level is easier because the features can directly encode 2D-3D correspondences. While related work [32, 69] directly uses DINO [41] as appearance features, recent studies showed that DINO has severe limitations at estimating semantic correspondence accurately [35, 78]. Therefore, we train an additional adapter module on top of", + "bbox": [ + 89, + 719, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "DINO by using the 3D object prior to further enhance the feature extractor at accurately estimating 2D-3D correspondences. This leads to higher quality semantic correspondence features and a largely enhanced model performance.", + "bbox": [ + 511, + 90, + 903, + 151 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To summarize, we make the following contributions: (1) We propose a new way of integrating self-supervised learning, neural rendering, and 3D modeling to enable the joint learning of 3D object priors for common objects from a collection of object-centric videos, (2) an effective mechanism for learning correspondence features by training a 3D object prior jointly with an additional adapter module on top of DINO, using a local contrastive objective that exploits the correspondences estimated by the 3D model, and (3) Common3D is the first completely self-supervised method that can estimate the shape, pose, and semantic correspondence of common objects from a single image.", + "bbox": [ + 511, + 152, + 903, + 332 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 345, + 653, + 361 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our model has two core components: An adapter module that enhances correspondence learning and a 3D morphable model that is learned without supervision. Therefore, we first provide an overview of correspondence learning methods and subsequently discuss related work on learning 3D morphable models from images and from synthetic data.", + "bbox": [ + 511, + 372, + 903, + 462 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Self-Supervised Correspondence Learning. Finding semantic correspondences across instances of an object category is a challenging computer vision task [37]. Self-supervised models like DINOv2 [41] and Stable Diffusion [47] were shown to have remarkable capabilities at estimating zero-shot semantic correspondences [79]. Despite these advances, unsupervised models are still not performing well on all categories. This can be accounted to ambiguous features of similar parts [35] or the lower performance for viewpoint changes [78]. Previous works have addressed this by mapping the 2D features to a coarse spherical representation of the object [35] or by applying heuristic augmentation methods to account for the viewpoint dependency [78]. Recent related work [51] showed that by leveraging 3D shape representations, the 2D-3D correspondence mapping can be improved significantly. However, their approach requires a detailed categorical mesh and is limited to a small selected set of animals. Similarly, [52] uses multiple views from object-centric videos to account for the viewpoint dependency. Further, they obtain a simple rigid 3D template and learn 2D-3D correspondences to estimate the 3D pose of an object. In contrast, our proposed Common3D learns an underlying deformable 3D shape representation self-supervised, which enables us to generalize to various common objects. Moreover, our approach of learning correspondence features by training a 3D object prior jointly with an additional adapter module on top of DINO enables high quality features to emerge.", + "bbox": [ + 511, + 463, + 903, + 885 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "6469", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/d191c29ecff0a63279bd5c07df407856295f3ae5528a9dbf43c7457267ae1d00.jpg", + "image_caption": [ + "Figure 2. Method Overview. At the core of our method is a category-level template with semantic features that is acquired using a neural SDF with Differentiable Marching Tetrahedra (DMTet) where features are attached using a feature field. The category-level template is morphed using an MLP that is conditioned on a latent code $l$ and rotated using the pose $\\pi$ . The pose $\\pi$ is estimated unsupervised at training time and predicted at inference time. DINOv2 serves as a feature encoder and its output is processed in two branches: One for estimating the latent code $l$ with a block of convolutional layers; and another branch serves as an adapter to enhance the DINO features for better correspondence learning. As a result, the features from the adapter are better suited for finding correspondences. For example, the tire features of DINO have a similar encoding and are, hence, ambiguous, whereas it is much more distinct after the adapter. We also visualize the training objectives: The appearance objective compares two surface probabilities: 1) The geometric probability $p(\\mathrm{v}_j|\\mathrm{v}_i,\\mathrm{V})$ , as defined via the Euclidean distances of the mesh vertices; and 2) the appearance probability $\\bar{p} (\\mathrm{f}_j|s_i,\\mathrm{F},\\beta)$ , acquired by comparing the 2D image features $s_i$ with the surface features $F$ . The remaining geometric objectives enforce the 3D shape and the projected mask to be consistent with the data acquired from a SFM pipeline." + ], + "image_footnote": [], + "bbox": [ + 94, + 88, + 558, + 309 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/85d242279d4038a3267766926e54de4c6624ef9af42258a3d3421ce2d55c3c0b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 568, + 88, + 903, + 309 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Learning 3D Morphable Models from Images. There was remarkable progress made in the self-supervised learning of category-specific 3D morphable models from images. Originally, approaches used camera viewpoint, 2D keypoints, and masks to predict the 3D shape [24, 59] and texture [22, 64]. Several works achieved to mitigate the requirement of camera viewpoint and 2D keypoint annotations by exploiting symmetries [12, 18, 31, 60], by adding multi-view viewpoint hypothesis [12], by requiring consistent self-supervised co-part segmentation [31] using [19], by cycle-consistent surface maps [27, 28, 60], or by temporal consistency for video clips [56, 57, 68]. Few works even achieve to not require mask annotations, [38, 58, 66], by being constrained to front views of faces [66], or coarse-to-fine learning of texture and shape [38]. However, most of these works focus on modeling the object in very fine detail, such that they can generate object instances down to the RGB pixel-level, making the inverse rendering process inherently difficult [46, 49, 55]. In contrast, our focus lies on self-supervised learning of a 3D morphable model that is particularly well suited for solving vision tasks through inverse rendering. Therefore, we represent the object appearance using neural features instead of RGB colors, which facilitates the inverse rendering process as the features learn to support the inverse rendering process by encoding 2D-3D correspondences.", + "bbox": [ + 91, + 505, + 483, + 898 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Another line of work learns 3D object priors without viewpoint annotations by replicating DINO features [20, 32, 53, 69, 75, 76]. In contrast to us, these works are restricted to animal categories. Further, they are not enhancing the image feature encoder for learning correspondences. The work of Zhang et al. is related to ours [80], as it also enhances the image feature encoder, but not in an inverse rendering setting. Further, they rely on a detailed template prior shape and focus on categories with few deformations and rotational symmetries, such as a bottle.", + "bbox": [ + 511, + 505, + 906, + 655 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "There exist also multiple works that focus on videospecific deformation model [9, 43, 73, 74]. In contrast to us, they have to re-train for each new instance and cannot find correspondences across instances.", + "bbox": [ + 511, + 657, + 906, + 717 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3D Morphable Models from Synthetic Data. Various work use synthetic data [7] to learn 3D representations. With dense annotation, in terms of a complete surface and annotated canonical frames, several works learn impressive 3D autoencoders [16], [42], and image-conditioned 3D reconstruction [65], [17]. Hereby, implicit 3D representations have outperformed explicit ones and were extended to learn category-specific 3D morphable models [10], [82], and [54]. In contrast, our model does not require synthetic data of several thousands instances per category, but only requires few casually captured object-centric videos. This provides novel 3D morphable models for various categories.", + "bbox": [ + 511, + 719, + 908, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "6470", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. A 3D Morphable Model of Neural Features", + "text_level": 1, + "bbox": [ + 89, + 89, + 475, + 107 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section we introduce Common3D, a method for self-supervised learning of 3D morphable models (3DMM) for common objects from only object-centric videos. One of the core challenges that we need to resolve is to find an approach that can deal with highly variable shapes and topologies within the same object category, while at the same time being general enough such that it works for as many categories as possible, hence requiring minimal category-specific heuristics. One key design choice is to represent the morphable shape model with a hybrid volumetric-mesh representation that is flexible, but also enables using mesh-based regularizers during training (Sec. 3.1). Our second key design choice is to represent the object appearance in terms of neural features and not RGB colors (Sec. 3.2), which enables the feature extractor to learn robust 2D-3D correspondences and, hence, facilitates the inverse rendering process. We discuss our training strategy and objectives in Sec. 3.3, and elaborate the inference mechanism in Sec. 3.4. Fig. 2 provides an overview of our method.", + "bbox": [ + 89, + 114, + 483, + 402 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Morphable Shape Representation", + "text_level": 1, + "bbox": [ + 89, + 416, + 385, + 431 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We represent 3D object priors with a category-level template that has neural features attached to its surface, together with instance-specific geometric deformations.", + "bbox": [ + 89, + 440, + 483, + 486 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Category-level Template. One core challenge in our work is to develop a flexible approach for learning the category-level template shape of objects, which generalizes across the highly variable shapes and topologies of common objects. While purely mesh-based representations can have advantages, they are difficult to optimize from scratch, leading to problems that often require ad-hoc heuristics such as re-meshing [13, 72]. To enable a more flexible way of optimizing the category-level 3D shape, we use a hybrid volumetric-mesh representation. This combines the advantages of an implicit and explicit 3D representation. In particular, the object shape is represented with an implicit model that is given as a signed distance field $\\phi_{\\mathrm{sdf}}$ which can adapt to complex geometries flexibly. It can be converted on the fly into an explicit mesh via the Differentiable Marching Tetrahedra (DMTet) algorithm [50]. Specifically, the signed distance field (SDF) is evaluated at the vertices of a tetrahedral grid and converted into a watertight mesh in a fast and differentiable manner. We denote the triangle mesh as $\\mathrm{M} = \\{\\mathrm{V},\\mathrm{E},\\mathrm{F}\\}$ with the vertices $\\mathrm{V} = \\{\\mathrm{v}_i\\in \\mathbb{R}^3\\}_{i = 1}^{|\\mathrm{V}|}$ , the edges $\\mathrm{E} = \\{(\\mathrm{v}_i,\\mathrm{v}_j)_e\\in \\mathbb{R}^3\\}_{e = 1}^{|\\mathrm{E}|}$ and the features $\\mathrm{F} = \\{\\mathrm{f}_i\\in \\mathbb{R}^D\\}_{i = 1}^{|F|}$ . Once converted to a mesh, the 3D object prior can be rendered into a 2D feature map by applying standard rasterization. The conversion of the SDF into a mesh enables us to use mesh-based regularizers when learning the deformable model, for example to enforce rigidity. Moreover, the vertices provide us", + "bbox": [ + 89, + 488, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "direct access to corresponding points among the deformed templates, which will enable us to learn features that represent 2D-3D correspondences effectively during training (see Sec. 3.2 and Sec. 3.3 for details).", + "bbox": [ + 511, + 90, + 905, + 150 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Instance-Level Deformation. We adopt an affine transformation field [82] to model the deformations of the category-level shape to different object instances. In contrast to [82] we transform the vertices of the category-level mesh instead of the signed distance field. This is more efficient, as re-extracting the mesh for each instance is not required. The function $\\phi_{a}(\\mathrm{v},\\mathrm{l}):\\mathbb{R}^{3}\\times \\mathrm{L}\\mapsto \\mathbb{R}^{3}$ maps each vertex $\\mathbf{v}$ from the category-level canonical space together with the image-specific latent code $\\mathbf{l}$ to a new position", + "bbox": [ + 511, + 152, + 906, + 287 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\phi_ {a} (\\mathrm {v}, 1) = \\alpha (\\mathrm {v}, 1) \\odot \\mathrm {v} + \\delta (\\mathrm {v}, 1), \\quad \\alpha , \\delta \\in \\mathbb {R} ^ {3}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 557, + 296, + 906, + 313 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "with the intermediate variables for scale $\\alpha$ and translation $\\delta$ . Hereby, $\\alpha$ and $\\delta$ are the outputs of an MLP conditioned on the vertex position $\\mathbf{v}$ and the latent code $\\mathbf{l}$ . We compute the image-specific latent code $\\mathbf{l}$ with an image feature encoder $\\psi_{1}$ that consists of a DINOv2 backbone and additional convolutional layers (see Fig. 2). We denote the deformed mesh $M_{def}(\\mathrm{I}) = \\{\\mathrm{V}_{def}(\\mathrm{I}),\\mathrm{E},\\mathcal{F},\\mathrm{F}\\}$ with the deformed vertices $V_{def}(\\mathrm{I}) = \\{\\phi_a(\\mathrm{v}_i,\\psi_1(\\mathrm{I}))\\in \\mathbb{R}^3\\}_{i = 1}^{|\\mathrm{V}|}$ and the image I.", + "bbox": [ + 511, + 321, + 905, + 448 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Appearance as Semantic Feature Field", + "text_level": 1, + "bbox": [ + 511, + 454, + 844, + 470 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Prior works in the context of 3DMMs traditionally represent the object appearance using RGB colors, which makes it inherently difficult to apply them for solving vision tasks through inverse rendering [46, 49, 55]. A core contribution of our work is to extend 3DMMs to facilitate the inverse rendering process. For this purpose, we represent the appearance using neural features instead of RGB colors. The key advantage of this approach is that the neural features can be invariant to image-specific appearance details and instead focus on encoding semantic correspondences that facilitate the inverse rendering process. Our approach is inspired from prior works [69] that learn deformable 3D models through inverse rendering on the feature-level. These works use DINO [41] features that allow estimating semantic correspondences reasonably well [6]. However, recent studies have shown that DINO features also have limitations, e.g., for differentiating similar object parts, or [78] or in terms of dense prediction [79]. Therefore, we choose to learn an adapter on top of the DINO backbone to further improve the image features for better capturing semantic correspondences. Specifically, we compute a feature map of semantic features as $S = \\Phi_w(\\tau(I)) \\in \\mathbb{R}^{D \\times H \\times W}$ , where $\\tau$ is a frozen DINOv2 [41] backbone and $\\Phi_w$ is the adapter with $w$ being the parameters of several convolutional layers (see Sec. 4.1 for the detailed architecture). The adapter is trained jointly with the shape model as described in Sec. 3.3.", + "bbox": [ + 511, + 477, + 906, + 869 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "On the 3DMM, we represent the object appearance as a set of feature vectors at every mesh vertex $\\mathbf{F} = \\{\\mathbf{f}_i \\in$", + "bbox": [ + 511, + 869, + 905, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "6471", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$\\mathbb{R}^D\\}_{i=1}^{|V|}$ . The features are generated by a feature field [40] $\\mathrm{f}_i = \\phi_{\\text{feat}}(\\mathrm{v}_i) : \\mathbb{R}^3 \\mapsto \\mathbb{R}^D$ that maps each vertex $\\mathrm{v}_i$ to a feature vector $\\mathrm{f}_i$ . Importantly, as the vertices of the deformable mesh encode corresponding points in different object instances, the adapter can learn more accurate 2D-3D correspondences compared to self-supervised methods that learn purely in 2D, see Fig. 2.", + "bbox": [ + 89, + 88, + 483, + 198 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Training Objectives", + "text_level": 1, + "bbox": [ + 89, + 205, + 279, + 222 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We train Common3D from a collection of videos of objects. Our training process jointly optimizes the category-level template shape, the deformation model, the image feature encoder adapter, and the semantic feature field. Initially, we obtain a 3D point cloud from every video using Structure-from-Motion [48]. We need to align the individual 3D reconstructions of object instances into a canonical coordinate frame to obtain a camera pose annotation $\\pi$ relative to the object that is coherent across videos. We achieve this by adopting the method proposed in [52] to get initial camera pose annotations in an unsupervised manner. In the following, we discuss the geometric objectives, the appearance objective, and finally, the combined objective used for training the model.", + "bbox": [ + 89, + 229, + 483, + 440 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Geometric Objectives. For optimizing the geometry, we formulate 2D reconstruction objectives using the mask and a 3D reconstruction objective using the 3D point cloud from each instance. Simultaneously, we regularize the category-level template's SDF and introduce mesh-based regularizers for highly rigid instance-level deformations.", + "bbox": [ + 89, + 441, + 483, + 530 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "First, we penalize the pixel-wise deviation $(u,v)$ of our rendered mask $\\bar{\\mathfrak{m}} (\\mathrm{M}_{def},\\mathrm{I},\\pi ,u,v)$ with the pose $\\pi$ from the object-centric mask m, provided by PointRend [26]. For this, we compute the mean squared error", + "bbox": [ + 89, + 531, + 483, + 592 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\mathrm {m}} \\left(\\mathrm {M} _ {d e f}, \\mathrm {I}, \\pi , \\mathrm {m}, u, v\\right) = \\\\ \\left| \\left| \\bar {\\mathrm {m}} \\left(\\mathrm {M} _ {d e f}, \\mathrm {I}, \\pi , u, v\\right) - \\mathrm {m} (u, v) \\right| \\right| ^ {2}. \\tag {2} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 173, + 603, + 480, + 638 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We also maximize the pixel-wise overlap with the mask's distance transform $\\mathrm{m_{dt}}(u,v)$ , as follows:", + "bbox": [ + 89, + 648, + 482, + 679 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\mathrm {m d t}} \\left(\\mathrm {M} _ {\\text {d e f}}, \\mathrm {I}, \\pi , \\mathrm {m} _ {\\mathrm {d t}}, u, v\\right) = \\tag {3} \\\\ - \\bar {\\mathrm {m}} (\\mathrm {M} _ {d e f}, \\mathrm {I}, \\pi , u, v) \\mathrm {m} _ {\\mathrm {d t}} (u, v). \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 176, + 689, + 480, + 724 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Hereby, the distance transform equals the distance to the silhouette for each pixel inside the mask, all pixels outside the mask are equal to zero. Note that this objective may only increase the object's volume, leading to a category-level template that encapsulates all objects like a visual hull [30]. Without this objective, and given the conflicting objectives of fitting a single category-level template to multiple different instances, the SDF might result in disconnected parts.", + "bbox": [ + 89, + 734, + 482, + 854 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To ensure that the deformed mesh reconstructs the shape of the individual object instances, we formulate a 3D chamfer distance based on the canonicalized 3D point cloud", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$\\mathcal{P}_{inst}$ , obtained from our object-centric video,", + "bbox": [ + 513, + 90, + 818, + 106 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {C D} \\left(\\mathrm {M} _ {d e f}, \\mathrm {I}, \\mathcal {P} _ {i n s t}\\right) = \\frac {1}{| \\mathrm {V} _ {d e f} | + | \\mathcal {P} _ {i n s t} |} \\\\ \\left(\\sum_ {\\mathrm {v} _ {i} \\in \\mathrm {V} _ {d e f}} | | \\mathrm {v} _ {i} - x _ {\\chi (\\mathrm {v} _ {i})} | | _ {2} + \\sum_ {x _ {i} \\in \\mathcal {P} _ {i n s t}} | | x _ {i} - \\mathrm {v} _ {\\chi (x _ {i})} | | _ {2}\\right), \\tag {4} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 521, + 118, + 903, + 212 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "with $\\chi : \\mathbb{R}^3 \\mapsto \\mathbb{R}^3$ returning the nearest neighbor of a 3D point to either the vertices $V_{\\text{def}}$ or the point cloud $\\mathcal{P}_{\\text{inst}}$ .", + "bbox": [ + 511, + 213, + 903, + 244 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Further, we optimize the category-level template using the Eikonal regularization as proposed in [15] to ensure the satisfaction of the Eikonal property of a signed distance function, which states that the norm of the gradient equals to one. This is required as DMTet only provides gradients via the extracted mesh close to the surface, which leaves the points more far away from the surface undefined. Therefore, we sample a set of points $\\mathcal{P}_{sdf}$ uniformly in the canonical space and around the extracted vertices, and average over each point $x\\in \\mathcal{P}_{sdf}$", + "bbox": [ + 511, + 244, + 906, + 396 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {s d f} (\\mathrm {M}, x) = \\left(\\left| \\right| \\nabla \\phi_ {s d f} (x) \\left| \\right| _ {2} - 1\\right) ^ {2}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 407, + 903, + 426 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We also penalize large deformations from the category-level template using the mean squared error", + "bbox": [ + 511, + 439, + 903, + 469 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {d e f} (\\mathrm {M}, \\mathrm {M} _ {d e f}, \\mathrm {I}) = \\frac {1}{| \\mathrm {V} |} \\sum_ {\\mathrm {v} \\in \\mathrm {V}} \\left\\| \\mathrm {v} - \\phi_ {a} (\\mathrm {v}, \\psi_ {1} (\\mathrm {I})) \\right\\| ^ {2}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 529, + 481, + 903, + 517 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To ensure that deformations of connected vertices are smooth, we adopt the point-pair regularization from [82] to only consider connected vertices", + "bbox": [ + 511, + 530, + 903, + 574 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {d e f - s m} (\\mathrm {M}, \\mathrm {M} _ {d e f}, \\mathrm {I}) = \\frac {1}{| \\mathrm {E} |} \\\\ \\sum_ {\\mathrm {v} _ {i}, \\mathrm {v} _ {j} \\in \\mathrm {E}} \\frac {\\left| \\left| \\left[ \\mathrm {v} _ {i} - \\phi_ {a} \\left(\\mathrm {v} _ {i} , \\psi_ {1} (\\mathrm {I})\\right) \\right] - \\left[ \\mathrm {v} _ {j} - \\phi_ {a} \\left(\\mathrm {v} _ {j} , \\psi_ {1} (\\mathrm {I})\\right) \\right] \\right| \\right| _ {2}}{\\left| \\left| \\mathrm {v} _ {i} - \\mathrm {v} _ {j} \\right| \\right| _ {2}}. \\tag {7} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 587, + 903, + 672 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Appearance Objective. In addition to the morphable shape representation, we also train the parameters of the semantic feature field $\\phi_{\\text{feat}}(\\mathbf{v}_i)$ and the parameters of the adapter module $\\Phi_w$ of the image feature encoder. In order to learn the semantic feature field and the feature encoder simultaneously, we cannot directly use a reconstruction objective, as this optimization collapses to the trivial solution of a constant feature for the whole image and the whole surface. We prevent a collapse to the trivial solution by instead optimizing an objective that encourages unique 2D-3D correspondences to emerge, which is inspired from related work on contrastive learning for keypoint detection [4]. In particular, we model for each image feature $s_i$ the 3D correspondence as surface probability $\\bar{\\mathfrak{p}}$ , which is computed by the similarity to all vertex features $F$ . Moreover,", + "bbox": [ + 511, + 674, + 906, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "6472", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "we add a background feature $\\beta$ , which enables our model to differentiate pixelwise between surface and background", + "bbox": [ + 89, + 90, + 482, + 121 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {p} \\left(f _ {j} \\mid s _ {i}, F, \\beta\\right) = \\frac {\\exp \\left(\\kappa \\left\\langle s _ {i} , f _ {j} \\right\\rangle\\right)}{\\sum_ {k} \\exp \\left(\\kappa \\left\\langle s _ {i} , f _ {k} \\right\\rangle\\right) + \\exp \\left(\\kappa \\left\\langle s _ {i} , \\beta \\right\\rangle\\right)}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 106, + 128, + 480, + 174 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\kappa$ is the temperature parameter. Similarly, the background probability $\\bar{\\mathfrak{p}} (\\beta |s_i,\\mathrm{F},\\beta)$ is defined for each image feature. In contrast to previous work, we use farthest point sampling as in [44] to retrieve a subset of vertices and features. To mitigate additional notation, we neglect this sampling in the following. As ground truth, we model the surface probability $\\mathbb{P}$ over the set of vertices as in [39, 62]. Therefore, the surface probability evaluated at the vertex $\\mathbf{v}_i$ is defined by the Euclidean distance to all other vertices", + "bbox": [ + 89, + 176, + 483, + 311 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {p} \\left(\\mathrm {v} _ {j} \\mid \\mathrm {v} _ {i}, \\mathrm {V}\\right) = \\frac {\\exp \\left(- \\frac {\\left| \\left| \\mathrm {v} _ {i} - \\mathrm {v} _ {j} \\right| \\right| ^ {2}}{2 \\sigma^ {2}}\\right)}{\\sum_ {k} \\exp \\left(\\frac {\\left| \\left| \\mathrm {v} _ {i} - \\mathrm {v} _ {k} \\right| \\right| ^ {2}}{2 \\sigma^ {2}}\\right)}. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 161, + 319, + 482, + 371 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Now we can render this surface probability using the camera pose $\\hat{\\mathrm{p}} (\\mathrm{v}_j|\\mathrm{M}_{def},\\pi ,u,v)$ . For pixels outside the surface we set the background probability $\\hat{\\mathrm{p}} (\\beta |\\mathrm{M}_{def},\\pi ,u,v)$ to 1. Jointly, we estimate these probabilities $\\bar{\\mathrm{p}} (\\mathrm{f}_j|\\Phi_w(\\mathrm{I},u,v),\\mathrm{F},\\beta)$ . Thereupon, we optimize the pixel-wise cross-entropy loss", + "bbox": [ + 89, + 378, + 483, + 470 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\text {a p p}} \\left(\\Phi_ {w}, \\mathrm {M} _ {\\text {d e f}}, \\mathrm {I})\\right) = \\\\ - \\sum_ {j} \\log (\\bar {\\mathrm {p}} (f _ {j})) \\hat {\\mathrm {p}} (v _ {j}) - \\log (\\bar {\\mathrm {p}} (\\beta)) \\hat {\\mathrm {p}} (\\beta) \\cdot \\tag {10} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 153, + 479, + 480, + 530 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Combined Objective. First, we only optimize the category-level template using only the geometric losses", + "bbox": [ + 89, + 539, + 482, + 570 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\lambda_ {C D} \\mathcal {L} _ {C D} + \\lambda_ {\\mathrm {m}} \\mathcal {L} _ {\\mathrm {m}} + \\lambda_ {\\mathrm {m d t}} \\mathcal {L} _ {\\mathrm {m d t}} + \\lambda_ {s d f} \\mathcal {L} _ {s d f}, \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 102, + 579, + 480, + 595 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "with the balancing weights $\\lambda_{i}$", + "bbox": [ + 89, + 604, + 292, + 619 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Second, we optimize the category-level feature field, the adapters, and the instance-level deformations using the weighted sum of geometric and appearance losses", + "bbox": [ + 89, + 619, + 482, + 666 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} = \\lambda_ {\\text {a p p}} \\mathcal {L} _ {\\text {a p p}} + \\lambda_ {C D} \\mathcal {L} _ {C D} + \\lambda_ {\\mathrm {m}} \\mathcal {L} _ {\\mathrm {m}} + \\lambda_ {\\mathrm {m d t}} \\mathcal {L} _ {\\mathrm {m d t}} \\tag {12} \\\\ + \\lambda_ {s d f} \\mathcal {L} _ {s d f} + \\lambda_ {d e f} \\mathcal {L} _ {d e f} + \\lambda_ {d e f - s m} \\mathcal {L} _ {d e f - s m}. \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 102, + 674, + 482, + 710 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.4. Inference via Inverse Rendering", + "text_level": 1, + "bbox": [ + 89, + 718, + 372, + 734 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "During inference, our main application is to infer the shape latent $l$ and the camera pose $\\pi$ . The shape latent can be directly predicted using the trained latent encoder $\\psi_{1}$ . While the pose could alternatively also be predicted with a feedforward neural network, other works have demonstrated that inverse rendering can significantly improve the results [63]. Therefore, we optimize the pixelwise surface probabilities during inference applying the reconstruction loss", + "bbox": [ + 89, + 741, + 483, + 863 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\max _ {\\pi} \\sum_ {u, v} \\bar {p} \\left(f _ {r} \\mid f _ {i}, F, \\beta\\right), \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 209, + 869, + 482, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "with the rendered image feature $\\mathrm{f}_r = \\Psi_r(\\mathrm{M}_{def},\\pi ,u,v)$ and the image feature extracted from the test image $\\mathrm{f}_i = \\Phi_w(\\tau (\\mathrm{I},u,v))$ . Note that we use three degrees of freedom for parametrizing the object rotation. Following [62], we initialize our pose using uniform sampling.", + "bbox": [ + 511, + 90, + 905, + 167 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 179, + 645, + 196 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We conduct extensive experiments and evaluate the effectiveness of Common3D on various downstream tasks for a variety of object categories. We use the same trained model for all tasks, demonstrating the universal utilization of combining geometry with correspondence-aware appearance in one single model. We describe our experimental setup in Sec. 4.1. Subsequently, we demonstrate that our model improves over prior self-supervised work in 3D object pose estimation (Sec. 4.2), in instance segmentation (Sec. 4.3), and in semantic correspondence estimation (Sec. 4.4).", + "bbox": [ + 511, + 204, + 906, + 354 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experimental Details", + "text_level": 1, + "bbox": [ + 511, + 364, + 712, + 378 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Training Dataset. To train Common3D, we use the object-centric videos provided by the Common Objects in 3D (CO3D) dataset [45]. For each category, the videos capture a large variety of instances. In total, it provides 19k videos across 50 categories, with each video containing approximately 100 to 200 frames. Each video is supplemented by point clouds and relative camera poses using Structure-from-Motion [48] and object masks using PointRend [26]. We choose the same videos as used in [52], where videos that were captured too narrow to or not focusing on the object were filtered out, and only videos were retained where the full object is visible. For each category, we use a maximum of 50 videos. Due to the filtering we use fewer videos for the following ten categories: 17 for \"remote\", 15 for \"mouse\", 16 for \"tv\", 7 for \"toilet\", 41 for \"toybus\", 28 for \"hairdryer\", 49 for \"couch\", and 23 for \"cellphone\". Before training, we center each object's coordinate frame and normalize its scale such that its point cloud ranges from -1 to 1 on the axis with the maximum range. We achieve adopt the method proposed in [52] to obtain canonical camera pose annotations across videos in an unsupervised manner.", + "bbox": [ + 511, + 387, + 906, + 704 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation on In-the-Wild Datasets. We evaluate our model at various tasks on in-the-wild data. In particular, we use three common datasets PASCAL3D+ [70], ObjectNet3D [71], and SPair-71k [37]. PASCAL3D+ and ObjectNet3D provide pose and segmentation annotations per image, which enables evaluation of 3D object pose estimation as well as instance segmentation. While PASCAL3D+ yields 12 rigid classes of PASCAL VOC 2012, ObjectNet3D covers, beyond that, 100 categories. We use the same 20 common categories across ObjectNet3D and CO3D as used in [52]. This even includes the domain gap of a toy bus in CO3D to a real one in PASCAL3D+ and ObjectNet3D. In total, we evaluate on 6233 images of PAS", + "bbox": [ + 511, + 704, + 908, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6473", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/905fb67d29b11a7f70c651914b39aced8d0534cec09f27fe369e0bd8e3a7c730.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Method∅∅∅∅∅○AVG
Sup.StarMap [83]83.294.475.479.868.885.882.5
VoGE [63]82.698.190.594.987.583.990.9
Unsup.ZSP [14]61.721.442.652.943.139.046.0
UOP3D [52]58.479.351.976.667.053.169.2
Ours67.282.967.880.978.151.175.3
", + "bbox": [ + 94, + 88, + 480, + 175 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/c44b669dc5b5e275e0471c5a6369dab56cb3b253dab39fded6d6f7571e5d7640.jpg", + "table_caption": [ + "Table 1. Evaluation of 3D object pose estimation on 7 categories of PASCAL3D+, we supplement the missing results in the appendix. The reported metric is the $30^{\\circ}$ accuracy. Note, that ZSP [14] requires also depth and runs an order of a magnitude slower." + ], + "table_footnote": [], + "table_body": "
草草草草草草草草草草草草草草
ZSP [14]23.150.860.346.436.833.021.7
UOP3D [52]18.062.198.154.652.238.214.1
Ours25.577.099.051.865.944.719.4
草草草草草草草草草草草草AVG
ZSP [14]46.860.528.841.625.856.337.9
UOP3D [52]26.953.344.754.415.539.653.2
Ours34.554.238.552.415.765.052.4
", + "bbox": [ + 94, + 257, + 480, + 377 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "CAL3D+ using the same subset as in [61], and on 12039 images of ObjectNet3D. We center all objects, as proposed in [83]. Further, we evaluate our feature adapter on the semantic correspondence task of SPair71k across 6 common categories, which we validate on 3898 image pairs.", + "bbox": [ + 88, + 470, + 482, + 546 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Implementation Details. As backbone we leverage DINOv2 [41] with the ViT-S model [2], for which we freeze all 21M parameters. To compare with prior work on semantic correspondence estimation Sec. 4.4, we additionally trained one ViT-B model. From an input resolution of $448 \\times 448$ , the backbone maps to a $32 \\times 32$ feature map. Our adapter, consisting of three ResNet blocks, upsamples the $32 \\times 32$ feature map to $64 \\times 64$ . For efficient differentiable rendering, we make use of the rasterizer proposed by [29], which achieves differentiability using anti-aliasing. For each category-specific morphable model we train, using Adam [25], for 20 epochs to optimize the template shape, and another 20 to optimize the appearance together with the deformation model. We use early stopping on a hold-out set comprising $10\\%$ of the videos. On a single NVIDIA GeForce RTX 2080 the training does not exceed 10 hours.", + "bbox": [ + 88, + 547, + 482, + 790 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. In-the-Wild 3D Pose Estimation", + "text_level": 1, + "bbox": [ + 89, + 801, + 372, + 816 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We evaluate our 3D object prior at single-image 3D pose estimation on in-the-wild data. We report the pose estimation results using the $30^{\\circ}$ accuracy in Tab. 1 for PASCAL3D+ and in Tab. 2 for ObjectNet3D. In Tab. 1, we also report 3D pose estimation results from two supervised", + "bbox": [ + 89, + 824, + 483, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/29ac7d5ec9db44eae4a9023491e51cb158301fd0e22bf644a5795ef29c2d1f25.jpg", + "table_caption": [ + "Table 2. Evaluation of 3D object pose estimation on 20 categories of ObjectNet3D. We supplement the missing results in the appendix. We report the $30^{\\circ}$ accuracy. Note, that ZSP [14] requires also depth input and runs an order of a magnitude slower." + ], + "table_footnote": [], + "table_body": "
201918171615AVG
UOP3D [52]26.066.579.142.371.763.673.0
Ours29.364.782.552.276.965.977.0
", + "bbox": [ + 516, + 88, + 903, + 138 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/1e42bdd88cd46d51108152778fc3022c9df50a496df7b22d1a2b92ff34a48b59.jpg", + "table_caption": [ + "Table 3. Evaluation of instance segmentation on 7 categories of PASCAL3D+. We report the IoU in percentage with the occlusion-free instance mask provided by the ground truth pose and mesh." + ], + "table_footnote": [], + "table_body": "
井井井井井井井
UOP3D [52]68.949.979.267.742.969.650.9
Ours69.454.182.868.153.366.853.6
井井井井井井井AVG
UOP3D [52]76.962.464.161.462.564.773.864.0
Ours83.170.662.771.765.366.778.367.0
", + "bbox": [ + 517, + 205, + 903, + 303 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. Evaluation of instance segmentation on 20 categories of ObjectNet3D, we supplement the missing results in the appendix. We report the IoU in percentage with the occlusion-free instance mask given by rendering the annotated mesh in the annotated pose.", + "bbox": [ + 511, + 311, + 906, + 369 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "methods [63, 83] that are trained from the PASCAL3D+ training set. Notably, all unsupervised methods are trained from CO3D data, as they require videos as input during training, and hence are evaluated in an out-of-domain setting. Our method improves over prior work on most of the categories except for categories with few shape variance across instances such as cellphone, microwave, and toaster. On average, the $30^{\\circ}$ accuracy improves significantly over prior work from $69.2\\%$ to $75.3\\%$ by $+6.3$ points on PASCAL3D+ and from $52.4\\%$ to $56.8\\%$ by $+4.4$ points on ObjectNet3D. We attribute this performance increase to the correspondence learning with a deformable shape, which facilitates the training and, therefore, generalizes better. This is illustrated in the qualitative examples in Fig. 3", + "bbox": [ + 511, + 396, + 906, + 609 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. In-the-Wild Instance Segmentation", + "text_level": 1, + "bbox": [ + 511, + 618, + 818, + 635 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We validate the deformation capabilities of Common3D using the instance segmentation masks of PASCAL3D+ in Tab. 3 and ObjectNet3D in Tab. 4. We compare our model with UOP3D [52], as it is the current state-of-the-art at self-supervised learning 3D object priors for common objects, and was also trained using CO3D videos. Common3D outperforms UOP3D on almost all categories and on average increases the IoU by 3.8 percentage points from $60.3\\%$ to $64.1\\%$ on PASCAL3D+ and by 3 percentage points, from $64\\%$ to $67\\%$ on ObjectNet3D. A qualitative comparison of our method against UOP3D is depicted in Fig. 3. It shows that Common3D achieves much more accurate highly accurate despite a large variability in the object instances.", + "bbox": [ + 511, + 642, + 906, + 840 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Semantic Correspondence", + "text_level": 1, + "bbox": [ + 511, + 849, + 751, + 866 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We demonstrate the significant benefit of our semantic feature adapter by evaluating its performance on semantic cor", + "bbox": [ + 511, + 869, + 905, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "6474", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e52efaa4cfbe7868ea5d56274a733ef4a8789e26c136429e4f8d75735f5575d5.jpg", + "image_caption": [ + "Figure 3. Qualitative results on the ObjectNet3D dataset. In the second row the results of our method are illustrated, in the third the results of UOP3D [52]. Notably, our method fits the object more accurately, resulting in improved 3D pose and segmentation accuracy." + ], + "image_footnote": [], + "bbox": [ + 94, + 89, + 903, + 303 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/2d69897f918ddfb8db1495429864ab6d0545a626a72bc304bfd85a3fb2533466.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
6073.367.246.366.042.259.4
SpherMaps [35]61.273.351.536.261.024.247.9
DINOv2 [41]62.052.366.244.260.731.954.0
Ours60.360.5
", + "bbox": [ + 93, + 368, + 480, + 438 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "respondence estimation in Tab. 5. We evaluate our model compared to the self-supervised DINOv2 model as well as SpherMaps [35], which require coarse viewpoint annotations. We follow the evaluation protocol of SpherMaps and find correspondences by weighting the image features similarities from the DINOv2 backbone and the feature similarities from our adapter. We weight the backbone's features with 0.8 and ours with 0.2. On average, we can largely increase the PCK@0.1 compared to the self-supervised baseline DINOv2 by $6.1\\%$ percentage points from $47.9\\%$ to $54\\%$ . For some categories, we can even achieve a competitive performance compared to the supervised baseline SpherMaps. We believe that one major challenge remains the domain gap between the training and test data.", + "bbox": [ + 88, + 530, + 482, + 742 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Ablation", + "text_level": 1, + "bbox": [ + 89, + 768, + 194, + 784 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In Tab. 6, we ablate our model design choices regarding the adapter, the appearance loss as well as the instance deformation model. We observe, that instance-level deformation has only a marginal relevance for the pose estimation, but proves essential for segmentation and semantic correspondence estimation. Furthermore, the adapter and the dense appearance loss are significantly improving our method.", + "bbox": [ + 89, + 794, + 482, + 902 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/e98e70821037e2217d35c0f88a3baa08b1ab99dfa42a0887d4abe643abc3fa74.jpg", + "table_caption": [ + "Table 5. Evaluation of the PCK@0.1 on 6 categories of SPair-71k [37]. Note, that SpherMaps [35] uses coarse viewpoint annotations, while DINOv2 and our approach do not require any." + ], + "table_footnote": [], + "table_body": "
30°10°IoUPCK@0.1
w/o Adapt.25.96.660.532.4
w/o Adapt., w Rec. Loss13.92.156.232.4
w/o Dense App. Loss56.712.258.845.3
w/o Inst. Deform.75.820.258.645.3
Ours75.321.064.147.8
", + "bbox": [ + 516, + 368, + 903, + 469 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 6. Ablation of our model design choices. We evaluate the 3D pose accuracy and instance segmentation on PASCAL3D+ using $30^{\\circ}$ , $10^{\\circ}$ accuracy and IoU. Further, the semantic correspondences are evaluated on SPair71-k using PCK@0.1. We observe that the adapter is crucial for our method, as well as the dense appearance loss. Further, our instance deformation model yields a minor disadvantage for the coarse $30^{\\circ}$ pose estimation, while improving the fine-grained $10^{\\circ}$ pose estimation, and proves essential for the instance segmentation and the semantic correspondences. Note that in this standard setting we use the ViT-S backbone, in contrast to the results reported in Tab. 5.", + "bbox": [ + 511, + 479, + 906, + 633 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 662, + 633, + 679 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We have introduced Common3D, a new model that can learn 3D morphable models from a collection of casually captured videos completely self-supervised. This model can, at test time, estimate the 3D object shape, 2D-3D correspondences, and the 3D object pose. Common3D connects two different subfields of computer vision, self-supervised representation learning and the learning of 3D deformable object priors. We believe that in the long term, this may unify many vision tasks into a common inference process, while also creating a better 3D spatial understanding in vision models.", + "bbox": [ + 511, + 686, + 906, + 851 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgment. This work was funded by the German Research Foundation (DFG) - 468670075, and 499552394 (SFB 1597 - Small Data).", + "bbox": [ + 511, + 854, + 908, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "6475", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Oswald Aldrian and William AP Smith. Inverse rendering of faces with a 3d morphable model. IEEE transactions on pattern analysis and machine intelligence, 35(5):1080-1093, 2012. 1", + "[2] Dosovitskiy Alexey. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv: 2010.11929, 2020.7", + "[3] Dragomir Anguelov, Praveen Srinivasan, Daphne Koller, Sebastian Thrun, Jim Rodgers, and James Davis. Scape: shape completion and animation of people. In ACM SIGGRAPH 2005 Papers, pages 408-416. 2005. 2", + "[4] Yutong Bai, Angtian Wang, Adam Kortylewski, and Alan Yuille. Coke: Contrastive learning for robust keypoint detection. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 65-74, 2023. 5", + "[5] V BLANZ and T VETTER. A morphable model for the synthesis of 3d faces. In Computer graphics proceedings, annual conference series, pages 187-194. Association for Computing Machinery SIGGRAPH, 1999. 1, 2", + "[6] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 4", + "[7] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 3", + "[8] Zhiqin Chen and Hao Zhang. Learning implicit fields for generative shape modeling. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5939-5948, 2019. 2", + "[9] Devikalyan Das, Christopher Wewer, Raza Yunus, Eddy Ilg, and Jan Eric Lenssen. Neural parametric gaussians for monocular non-rigid object reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10715-10725, 2024. 3", + "[10] Yu Deng, Jiaolong Yang, and Xin Tong. Deformed implicit field: Modeling 3d shapes with learned dense correspondence. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10286-10296, 2021. 3", + "[11] Bernhard Egger, Sandro Schonborn, Andreas Schneider, Adam Kortylewski, Andreas Morel-Forster, Clemens Blumer, and Thomas Vetter. Occlusion-aware 3d morphable models and an illumination prior for face image analysis. International Journal of Computer Vision, 126:1269-1287, 2018. 1", + "[12] Shubham Goel, Angjoo Kanazawa, and Jitendra Malik. Shape and viewpoint without keypoints. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XV 16, pages 88-104. Springer, 2020. 3", + "[13] Shubham Goel, Georgia Gkioxari, and Jitendra Malik. Differentiable stereopsis: Meshes from multiple views using" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "differentiable rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8635-8644, 2022. 4", + "[14] Walter Goodwin, Sagar Vaze, Ioannis Havoutis, and Ingmar Posner. Zero-shot category-level object pose estimation. In European Conference on Computer Vision, pages 516-532. Springer, 2022. 7", + "[15] Amos Gropp, Lior Yariv, Niv Haim, Matan Atzmon, and Yaron Lipman. Implicit geometric regularization for learning shapes. arXiv preprint arXiv:2002.10099, 2020. 5", + "[16] Thibault Groueix, Matthew Fisher, Vladimir G Kim, Bryan C Russell, and Mathieu Aubry. Atlasnet: a papiermâché approach to learning 3d surface generation (2018). arXiv preprint arXiv:1802.05384, 11, 1802. 3", + "[17] Paul Henderson, Vagia Tsiminaki, and Christoph H Lampert. Leveraging 2d data to learn textured 3d mesh generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7498-7507, 2020. 3", + "[18] Tao Hu, Liwei Wang, Xiaogang Xu, Shu Liu, and Jiaya Jia. Self-supervised 3d mesh reconstruction from single images. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6002-6011, 2021. 3", + "[19] Wei-Chih Hung, Varun Jampani, Sifei Liu, Pavlo Molchanov, Ming-Hsuan Yang, and Jan Kautz. Scops: Self-supervised co-part segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 869-878, 2019. 3", + "[20] Tomas Jakab, Ruining Li, Shangzhe Wu, Christian Rupprecht, and Andrea Vedaldi. Farm3d: Learning articulated 3d animals by distilling 2d diffusion. In 2024 International Conference on 3D Vision (3DV), pages 852-861. IEEE, 2024. 3", + "[21] Hanbyul Joo, Tomas Simon, and Yaser Sheikh. Total capture: A 3d deformation model for tracking faces, hands, and bodies. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8320-8329, 2018. 2", + "[22] Angjoo Kanazawa, Shubham Tulsiani, Alexei A Efros, and Jitendra Malik. Learning category-specific mesh reconstruction from image collections. In Proceedings of the European Conference on Computer Vision (ECCV), pages 371-386, 2018. 3", + "[23] Angjoo Kanazawa, Shubham Tulsiani, Alexei A Efros, and Jitendra Malik. Learning category-specific mesh reconstruction from image collections. In Proceedings of the European Conference on Computer Vision (ECCV), pages 371-386, 2018. 2", + "[24] Abhishek Kar, Shubham Tulsiani, Joao Carreira, and Jitendra Malik. Category-specific object reconstruction from a single image. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1966-1974, 2015. 3", + "[25] Diederik P Kingma. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 7", + "[26] Alexander Kirillov, Yuxin Wu, Kaiming He, and Ross Girshick. Pointrend: Image segmentation as rendering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9799-9808, 2020. 5, 6" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "6476", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[27] Filippos Kokkinos and Iasonas Kokkinos. To the point: Correspondence-driven monocular 3d category reconstruction. Advances in Neural Information Processing Systems, 34:7760-7772, 2021. 3", + "[28] Nilesh Kulkarni, Abhinav Gupta, David F Fouhey, and Shubham Tulsiani. Articulation-aware canonical surface mapping. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 452-461, 2020. 3", + "[29] Samuli Laine, Janne Hellsten, Tero Karras, Yeongho Seol, Jaakko Lehtinen, and Timo Aila. Modular primitives for high-performance differentiable rendering. ACM Transactions on Graphics, 39(6), 2020. 7", + "[30] Aldo Laurentini. The visual hull concept for silhouette-based image understanding. IEEE Transactions on pattern analysis and machine intelligence, 16(2):150-162, 1994. 5", + "[31] Xueting Li, Sifei Liu, Kihwan Kim, Shalini De Mello, Varun Jampani, Ming-Hsuan Yang, and Jan Kautz. Self-supervised single-view 3d reconstruction via semantic consistency. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIV 16, pages 677-693. Springer, 2020. 3", + "[32] Zizhang Li, Dor Litvak, Ruining Li, Yunzhi Zhang, Tomas Jakab, Christian Rupprecht, Shangzhe Wu, Andrea Vedaldi, and Jiajun Wu. Learning the 3d fauna of the web. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9752-9762, 2024. 2, 3", + "[33] Yebin Liu, Juergen Gall, Carsten Stoll, Qionghai Dai, Hans-Peter Seidel, and Christian Theobalt. Markerless motion capture of multiple characters using multiview image segmentation. IEEE transactions on pattern analysis and machine intelligence, 35(11):2720-2735, 2013. 2", + "[34] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J Black. Smpl: a skinned multiperson linear model. ACM Transactions on Graphics (TOG), 34(6):1-16, 2015. 2", + "[35] Octave Mariotti, Oisin Mac Aodha, and Hakan Bilen. Improving semantic correspondence with viewpoint-guided spherical maps. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19521-19530, 2024. 2, 8", + "[36] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4460-4470, 2019. 2", + "[37] Juhong Min, Jongmin Lee, Jean Ponce, and Minsu Cho. Spair-71k: A large-scale benchmark for semantic correspondence. arXiv preprint arXiv:1908.10543, 2019. 2, 6, 8", + "[38] Tom Monnier, Matthew Fisher, Alexei A Efros, and Mathieu Aubry. Share with thy neighbors: Single-view reconstruction by cross-instance consistency. In European Conference on Computer Vision, pages 285-303. Springer, 2022. 3", + "[39] Natalia Neverova, David Novotny, Marc Szafraniec, Vasil Khalidov, Patrick Labatut, and Andrea Vedaldi. Continuous surface embeddings. Advances in Neural Information Processing Systems, 33:17258-17270, 2020. 6" + ], + "bbox": [ + 91, + 92, + 480, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[40] Michael Oechsle, Lars Mescheder, Michael Niemeyer, Thilo Strauss, and Andreas Geiger. Texture fields: Learning texture representations in function space. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4531-4540, 2019. 5", + "[41] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 2, 4, 7, 8", + "[42] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 165-174, 2019. 2, 3", + "[43] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5865-5874, 2021. 3", + "[44] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems, 30, 2017. 6", + "[45] Jeremy Reizenstein, Roman Shapovalov, Philipp Henzler, Luca Sbordone, Patrick Labatut, and David Novotny. Common objects in 3d: Large-scale learning and evaluation of real-life 3d category reconstruction. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10901-10911, 2021. 6", + "[46] Lukasz Romaszko, Christopher KI Williams, Pol Moreno, and Pushmeet Kohli. Vision-as-inverse-graphics: Obtaining a rich 3d explanation of a scene from a single image. In Proceedings of the IEEE International Conference on Computer Vision Workshops, pages 851–859, 2017. 2, 3, 4", + "[47] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2", + "[48] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-Motion Revisited. In Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 5, 6", + "[49] Sandro Schonborn, Bernhard Egger, Andreas Morel-Forster, and Thomas Vetter. Markov chain monte carlo for automated face image analysis. International Journal of Computer Vision, 123:160–183, 2017. 2, 3, 4", + "[50] Tianchang Shen, Jun Gao, Kangxue Yin, Ming-Yu Liu, and Sanja Fidler. Deep marching tetrahedra: a hybrid representation for high-resolution 3d shape synthesis. Advances in Neural Information Processing Systems, 34:6087-6101, 2021. 2, 4", + "[51] Aleksandar Shtedritski, Christian Rupprecht, and Andrea Vedaldi. Shic: Shape-image correspondences with no key-point supervision. arXiv preprint arXiv:2407.18907, 2024. 2", + "[52] Leonhard Sommer, Artur Jesslen, Eddy Ilg, and Adam Kortylewski. Unsupervised learning of category-level 3d pose" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "6477", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "from object-centric videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22787-22796, 2024. 2, 5, 6, 7, 8", + "[53] Keqiang Sun, Dor Litvak, Yunzhi Zhang, Hongsheng Li, Ji-jun Wu, and Shangzhe Wu. Ponomation: Learning articulated 3d animal motions from unlabeled online videos. In European Conference on Computer Vision, pages 100-119. Springer, 2025. 3", + "[54] Shanlin Sun, Kun Han, Deying Kong, Hao Tang, Xiangyi Yan, and Xiaohui Xie. Topology-preserving shape reconstruction and registration via neural diffeomorphic flow. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20845-20855, 2022. 3", + "[55] Ayush Tewari, Michael Zollhofer, Hyeongwoo Kim, Pablo Garrido, Florian Bernard, Patrick Perez, and Christian Theobalt. Mofa: Model-based deep convolutional face autoencoder for unsupervised monocular reconstruction. In Proceedings of the IEEE international conference on computer vision workshops, pages 1274-1283, 2017. 1, 2, 3, 4", + "[56] Ayush Tewari, Florian Bernard, Pablo Garrido, Gaurav Bharaj, Mohamed Elgharib, Hans-Peter Seidel, Patrick Pérez, Michael Zollhofer, and Christian Theobalt. Fml: Face model learning from videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10812-10822, 2019. 3", + "[57] Ayush Tewari, Hans-Peter Seidel, Mohamed Elgharib, Christian Theobalt, et al. Learning complete 3d morphable face models from images and videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3361-3371, 2021. 3", + "[58] Ayush Tewari, Xingang Pan, Ohad Fried, Maneesh Agrawala, Christian Theobalt, et al. Disentangled3d: Learning a 3d generative model with disentangled geometry and appearance from monocular images. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1516-1525, 2022. 2, 3", + "[59] Shubham Tulsiani, Abhishek Kar, Joao Carreira, and Jitendra Malik. Learning category-specific deformable 3d models for object reconstruction. IEEE transactions on pattern analysis and machine intelligence, 39(4):719-731, 2016. 3", + "[60] Shubham Tulsiani, Nilesh Kulkarni, and Abhinav Gupta. Implicit mesh reconstruction from unannotated image collections. arXiv preprint arXiv:2007.08504, 2020. 3", + "[61] Angtian Wang, Yihong Sun, Adam Kortylewski, and Alan L Yuille. Robust object detection under occlusion with context-aware compositionalnets. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12645-12654, 2020. 7", + "[62] Angtian Wang, Adam Kortylewski, and Alan Yuille. Nemo: Neural mesh models of contrastive features for robust 3d pose estimation. arXiv preprint arXiv:2101.12378, 2021. 6", + "[63] Angtian Wang, Peng Wang, Jian Sun, Adam Kortylewski, and Alan Yuille. Voge: a differentiable volume renderer using gaussian ellipsoids for analysis-by-synthesis. arXiv preprint arXiv:2205.15401, 2022. 6, 7", + "[64] Haowen Wang, Zhipeng Fan, Zhen Zhao, Zhengping Che, Zhiyuan Xu, Dong Liu, Feifei Feng, Yakun Huang, Xiuquan" + ], + "bbox": [ + 91, + 90, + 482, + 901 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Qiao, and Jian Tang. Dtf-net: Category-level pose estimation and shape reconstruction via deformable template field. In Proceedings of the 31st ACM International Conference on Multimedia, pages 3676-3685, 2023. 3", + "[65] Nanyang Wang, Yinda Zhang, Zhuwen Li, Yanwei Fu, Wei Liu, and Yu-Gang Jiang. Pixel2mesh: Generating 3d mesh models from single rgb images. In Proceedings of the European conference on computer vision (ECCV), pages 52-67, 2018. 3", + "[66] Shangzhe Wu, Christian Rupprecht, and Andrea Vedaldi. Unsupervised learning of probably symmetric deformable 3d objects from images in the wild. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1-10, 2020. 3", + "[67] Shangzhe Wu, Christian Rupprecht, and Andrea Vedaldi. Unsupervised learning of probably symmetric deformable 3d objects from images in the wild. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1-10, 2020. 2", + "[68] Shangzhe Wu, Tomas Jakab, Christian Rupprecht, and Andrea Vedaldi. Dove: Learning deformable 3d objects by watching videos. International Journal of Computer Vision, 131(10):2623-2634, 2023. 2, 3", + "[69] Shangzhe Wu, Ruining Li, Tomas Jakab, Christian Rupprecht, and Andrea Vedaldi. Magicpony: Learning articulated 3d animals in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8792-8802, 2023. 2, 3, 4", + "[70] Yu Xiang, Roozbeh Mottaghi, and Silvio Savarese. Beyond Pascal: A benchmark for 3d object detection in the wild. In IEEE winter conference on applications of computer vision, pages 75-82. IEEE, 2014. 6", + "[71] Yu Xiang, Wonhui Kim, Wei Chen, Jingwei Ji, Christopher Choy, Hao Su, Roozbeh Mottaghi, Leonidas Guibas, and Silvio Savarese. Objectnet3d: A large scale database for 3d object recognition. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part VIII 14, pages 160-176. Springer, 2016. 6", + "[72] Gengshan Yang, Deqing Sun, Varun Jampani, Daniel Vlasic, Forrester Cole, Huiwen Chang, Deva Ramanan, William T Freeman, and Ce Liu. Lasr: Learning articulated shape reconstruction from a monocular video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15980-15989, 2021. 4", + "[73] Gengshan Yang, Deqing Sun, Varun Jampani, Daniel Vlasic, Forrester Cole, Ce Liu, and Deva Ramanan. Viser: Video-specific surface embeddings for articulated 3d shape reconstruction. Advances in Neural Information Processing Systems, 34:19326-19338, 2021. 3", + "[74] Gengshan Yang, Minh Vo, Natalia Neverova, Deva Ramanan, Andrea Vedaldi, and Hanbyul Joo. Banmo: Building animatable 3d neural models from many casual videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2863-2873, 2022. 2, 3", + "[75] Chun-Han Yao, Wei-Chih Hung, Yuanzhen Li, Michael Rubinstein, Ming-Hsuan Yang, and Varun Jampani. Lassie:" + ], + "bbox": [ + 516, + 92, + 906, + 901 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "6478", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Learning articulated shapes from sparse image ensemble via 3d part discovery. Advances in Neural Information Processing Systems, 35:15296-15308, 2022. 3", + "[76] Chun-Han Yao, Wei-Chih Hung, Yuanzhen Li, Michael Rubinstein, Ming-Hsuan Yang, and Varun Jampani. Hi-lassie: High-fidelity articulated shape and skeleton discovery from sparse image ensemble. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4853-4862, 2023. 3", + "[77] Alan Yuille and Daniel Kersten. Vision as bayesian inference: analysis by synthesis? Trends in cognitive sciences, 10(7):301-308, 2006. 1", + "[78] Junyi Zhang, Charles Herrmann, Junhwa Hur, Eric Chen, Varun Jampani, Deqing Sun, and Ming-Hsuan Yang. Telling left from right: Identifying geometry-aware semantic correspondence. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3076-3085, 2024. 2, 4", + "[79] Junyi Zhang, Charles Herrmann, Junhwa Hur, Luisa Polania Cabrera, Varun Jampani, Deqing Sun, and Ming-Hsuan Yang. A tale of two features: Stable diffusion complements dino for zero-shot semantic correspondence. Advances in Neural Information Processing Systems, 36, 2024. 2, 4", + "[80] Kaifeng Zhang, Yang Fu, Shubhankar Borse, Hong Cai, Fatih Porikli, and Xiaolong Wang. Self-supervised geometric correspondence for category-level 6d object pose estimation in the wild. arXiv preprint arXiv:2210.07199, 2022. 3", + "[81] Yunzhi Zhang, Shangzhe Wu, Noah Snavely, and Jiajun Wu. Seeing a rose in five thousand ways. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 962-971, 2023. 2", + "[82] Zerong Zheng, Tao Yu, Qionghai Dai, and Yebin Liu. Deep implicit templates for 3d shape representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1429-1439, 2021. 3, 4, 5", + "[83] Xingyi Zhou, Arjun Karpur, Linjie Luo, and Qixing Huang. Starmap for category-agnostic keypoint and viewpoint estimation. In Proceedings of the European Conference on Computer Vision (ECCV), pages 318–334, 2018. 7", + "[84] Silvia Zuffi, Angjoo Kanazawa, David W Jacobs, and Michael J Black. 3d menagerie: Modeling the 3d shape and pose of animals. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6365-6373, 2017. 2" + ], + "bbox": [ + 91, + 90, + 482, + 710 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "6479", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2025/Common3D_ Self-Supervised Learning of 3D Morphable Models for Common Objects in Neural Feature Space/ecdc4e97-8d79-41e2-abfb-5c0cebf6aa2a_model.json b/2025/Common3D_ Self-Supervised Learning of 3D Morphable Models for Common Objects in Neural Feature Space/ecdc4e97-8d79-41e2-abfb-5c0cebf6aa2a_model.json new file mode 100644 index 0000000000000000000000000000000000000000..192a4fc7e5e34d76fab856d2e133208386f39a3b --- /dev/null +++ b/2025/Common3D_ Self-Supervised Learning of 3D Morphable Models for Common Objects in Neural Feature Space/ecdc4e97-8d79-41e2-abfb-5c0cebf6aa2a_model.json @@ -0,0 +1,2468 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.175, + 0.131, + 0.825, + 0.177 + ], + "angle": 0, + "content": "Common3D: Self-Supervised Learning of 3D Morphable Models for Common Objects in Neural Feature Space" + }, + { + "type": "text", + "bbox": [ + 0.131, + 0.203, + 0.294, + 0.221 + ], + "angle": 0, + "content": "Leonhard Sommer" + }, + { + "type": "text", + "bbox": [ + 0.334, + 0.203, + 0.448, + 0.221 + ], + "angle": 0, + "content": "Olaf Dunkel" + }, + { + "type": "text", + "bbox": [ + 0.489, + 0.203, + 0.651, + 0.221 + ], + "angle": 0, + "content": "Christian Theobalt 2" + }, + { + "type": "text", + "bbox": [ + 0.692, + 0.203, + 0.865, + 0.222 + ], + "angle": 0, + "content": "Adam Kortylewski 1,2" + }, + { + "type": "text", + "bbox": [ + 0.245, + 0.238, + 0.432, + 0.258 + ], + "angle": 0, + "content": "1University of Freiburg" + }, + { + "type": "text", + "bbox": [ + 0.451, + 0.238, + 0.748, + 0.257 + ], + "angle": 0, + "content": "\\(^{2}\\)Max Planck Institute for Informatics" + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.291, + 0.902, + 0.511 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.52, + 0.908, + 0.551 + ], + "angle": 0, + "content": "Figure 1. Common3D learns category-specific 3D morphable models from few casually captured videos completely self-supervised, and can estimate the 3D object shape (visualized from two viewpoints), 2D-3D correspondences, and the 3D object pose via inverse rendering." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.561, + 0.327, + 0.577 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.599, + 0.486, + 0.903 + ], + "angle": 0, + "content": "3D morphable models (3DMMs) are a powerful tool to represent the possible shapes and appearances of an object category. Given a single test image, 3DMMs can be used to solve various tasks, such as predicting the 3D shape, pose, semantic correspondence, and instance segmentation of an object. Unfortunately, 3DMMs are only available for very few object categories that are of particular interest, like faces or human bodies, as they require a demanding 3D data acquisition and category-specific training process. In contrast, we introduce a new method, Common3D, that learns 3DMMs of common objects in a fully self-supervised manner from a collection of object-centric videos. For this purpose, our model represents objects as a learned 3D template mesh and a deformation field that is parameterized as an image-conditioned neural network. Different from prior works, Common3D represents the object appearance with neural features instead of RGB colors, which enables the learning of more generalizable representations through an abstraction from pixel intensities. Importantly, we train the appearance features using a contrastive" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.563, + 0.909, + 0.685 + ], + "angle": 0, + "content": "objective by exploiting the correspondences defined through the deformable template mesh. This leads to higher quality correspondence features compared to related works and a significantly improved model performance at estimating 3D object pose and semantic correspondence. Common3D is the first completely self-supervised method that can solve various vision tasks in a zero-shot manner. We release all code at github.com/GenIntel/common3d." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.708, + 0.646, + 0.724 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Estimating the 3D shape and pose of an object from a single image requires prior knowledge about the possible shapes and appearances of an object category. Learning such prior models is a long-standing problem in computer vision, with 3D morphable models (3DMMs) [5] emerging as a very successful approach to address this problem. A popular application of 3DMMs is to perform single image 3D reconstruction via inverse rendering [1, 5, 55], also referred to as analysis-by-synthesis approach [77]. This enables solving under-constrained vision tasks, for example predicting an object's shape behind an occluder [11] or estimating the" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6468" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.182 + ], + "angle": 0, + "content": "3D object shape and pose from a single image. However, training 3DMMs requires sophisticated data acquisition setups [3, 21, 33], involving laser scanners or multi-view studio captures, and a significant manual effort in cleaning and registration of the data [5, 34, 84]. This approach is difficult to scale to more common everyday objects." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.185, + 0.482, + 0.306 + ], + "angle": 0, + "content": "An alternative approach that emerged recently is to learn a 3D prior from 2D images only [58, 67, 74, 81]. However, learning such priors from in-the-wild images is a major challenge and typically requires significant category-specific design choices, with various solutions having been proposed for different object categories [23, 32, 68, 69, 81]. This makes it difficult to scale to the long tail of common objects that can appear in natural images." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.308, + 0.483, + 0.534 + ], + "angle": 0, + "content": "We introduce Common3D, a new way of integrating self-supervised learning, neural rendering, and 3D modeling to learn 3D object priors for common objects from a collection of object-centric videos. Once trained, the model can estimate 2D-3D correspondences, it can be leveraged for inverse rendering to estimate the 3D pose, and instance segmentation of common objects from a single image, see Fig. 1. At the core of our method is a category-specific 3D morphable model (3DMM) that represents the possible variations of different object instances. The 3DMM is represented using a learned prototypical 3D shape and a shape deformation field that is parameterized as an image-conditioned neural network. Training our model requires canonical camera poses, which we estimate unsupervised by adopting the method of [52]." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.537, + 0.483, + 0.718 + ], + "angle": 0, + "content": "One core challenge for learning 3DMMs of common objects is to be flexible enough to capture the highly variable shapes and topologies of various object categories, as well as the possible variations of instances within one category. We address this issue by using a hybrid volumetric mesh representation based on DMTet [50]. Hence, the shape is defined volumetrically in canonical space, but a mesh is extracted on the fly for morphing and rendering the model. This allows to retain the flexibility of implicit neural representations for capturing highly variable shapes and topologies [8, 36, 42], while also enabling the use of meshes for an efficient rendering and regularization during training." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Another important challenge is to perform inverse rendering with 3D object priors, as it requires the optimization of a highly non-convex image-based reconstruction loss [46, 49, 55]. We address this problem in Common3D by representing the object appearance as a semantic feature field instead of RGB colors. Performing inverse rendering on the feature-level is easier because the features can directly encode 2D-3D correspondences. While related work [32, 69] directly uses DINO [41] as appearance features, recent studies showed that DINO has severe limitations at estimating semantic correspondence accurately [35, 78]. Therefore, we train an additional adapter module on top of" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.152 + ], + "angle": 0, + "content": "DINO by using the 3D object prior to further enhance the feature extractor at accurately estimating 2D-3D correspondences. This leads to higher quality semantic correspondence features and a largely enhanced model performance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.153, + 0.905, + 0.333 + ], + "angle": 0, + "content": "To summarize, we make the following contributions: (1) We propose a new way of integrating self-supervised learning, neural rendering, and 3D modeling to enable the joint learning of 3D object priors for common objects from a collection of object-centric videos, (2) an effective mechanism for learning correspondence features by training a 3D object prior jointly with an additional adapter module on top of DINO, using a local contrastive objective that exploits the correspondences estimated by the 3D model, and (3) Common3D is the first completely self-supervised method that can estimate the shape, pose, and semantic correspondence of common objects from a single image." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.347, + 0.655, + 0.362 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.373, + 0.905, + 0.463 + ], + "angle": 0, + "content": "Our model has two core components: An adapter module that enhances correspondence learning and a 3D morphable model that is learned without supervision. Therefore, we first provide an overview of correspondence learning methods and subsequently discuss related work on learning 3D morphable models from images and from synthetic data." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.464, + 0.905, + 0.886 + ], + "angle": 0, + "content": "Self-Supervised Correspondence Learning. Finding semantic correspondences across instances of an object category is a challenging computer vision task [37]. Self-supervised models like DINOv2 [41] and Stable Diffusion [47] were shown to have remarkable capabilities at estimating zero-shot semantic correspondences [79]. Despite these advances, unsupervised models are still not performing well on all categories. This can be accounted to ambiguous features of similar parts [35] or the lower performance for viewpoint changes [78]. Previous works have addressed this by mapping the 2D features to a coarse spherical representation of the object [35] or by applying heuristic augmentation methods to account for the viewpoint dependency [78]. Recent related work [51] showed that by leveraging 3D shape representations, the 2D-3D correspondence mapping can be improved significantly. However, their approach requires a detailed categorical mesh and is limited to a small selected set of animals. Similarly, [52] uses multiple views from object-centric videos to account for the viewpoint dependency. Further, they obtain a simple rigid 3D template and learn 2D-3D correspondences to estimate the 3D pose of an object. In contrast, our proposed Common3D learns an underlying deformable 3D shape representation self-supervised, which enables us to generalize to various common objects. Moreover, our approach of learning correspondence features by training a 3D object prior jointly with an additional adapter module on top of DINO enables high quality features to emerge." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "6469" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.089, + 0.56, + 0.31 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.569, + 0.089, + 0.904, + 0.31 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.326, + 0.909, + 0.48 + ], + "angle": 0, + "content": "Figure 2. Method Overview. At the core of our method is a category-level template with semantic features that is acquired using a neural SDF with Differentiable Marching Tetrahedra (DMTet) where features are attached using a feature field. The category-level template is morphed using an MLP that is conditioned on a latent code \\( l \\) and rotated using the pose \\( \\pi \\). The pose \\( \\pi \\) is estimated unsupervised at training time and predicted at inference time. DINOv2 serves as a feature encoder and its output is processed in two branches: One for estimating the latent code \\( l \\) with a block of convolutional layers; and another branch serves as an adapter to enhance the DINO features for better correspondence learning. As a result, the features from the adapter are better suited for finding correspondences. For example, the tire features of DINO have a similar encoding and are, hence, ambiguous, whereas it is much more distinct after the adapter. We also visualize the training objectives: The appearance objective compares two surface probabilities: 1) The geometric probability \\( p(\\mathrm{v}_j|\\mathrm{v}_i,\\mathrm{V}) \\), as defined via the Euclidean distances of the mesh vertices; and 2) the appearance probability \\( \\bar{p} (\\mathrm{f}_j|s_i,\\mathrm{F},\\beta) \\), acquired by comparing the 2D image features \\( s_i \\) with the surface features \\( F \\). The remaining geometric objectives enforce the 3D shape and the projected mask to be consistent with the data acquired from a SFM pipeline." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.506, + 0.485, + 0.899 + ], + "angle": 0, + "content": "Learning 3D Morphable Models from Images. There was remarkable progress made in the self-supervised learning of category-specific 3D morphable models from images. Originally, approaches used camera viewpoint, 2D keypoints, and masks to predict the 3D shape [24, 59] and texture [22, 64]. Several works achieved to mitigate the requirement of camera viewpoint and 2D keypoint annotations by exploiting symmetries [12, 18, 31, 60], by adding multi-view viewpoint hypothesis [12], by requiring consistent self-supervised co-part segmentation [31] using [19], by cycle-consistent surface maps [27, 28, 60], or by temporal consistency for video clips [56, 57, 68]. Few works even achieve to not require mask annotations, [38, 58, 66], by being constrained to front views of faces [66], or coarse-to-fine learning of texture and shape [38]. However, most of these works focus on modeling the object in very fine detail, such that they can generate object instances down to the RGB pixel-level, making the inverse rendering process inherently difficult [46, 49, 55]. In contrast, our focus lies on self-supervised learning of a 3D morphable model that is particularly well suited for solving vision tasks through inverse rendering. Therefore, we represent the object appearance using neural features instead of RGB colors, which facilitates the inverse rendering process as the features learn to support the inverse rendering process by encoding 2D-3D correspondences." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.506, + 0.907, + 0.656 + ], + "angle": 0, + "content": "Another line of work learns 3D object priors without viewpoint annotations by replicating DINO features [20, 32, 53, 69, 75, 76]. In contrast to us, these works are restricted to animal categories. Further, they are not enhancing the image feature encoder for learning correspondences. The work of Zhang et al. is related to ours [80], as it also enhances the image feature encoder, but not in an inverse rendering setting. Further, they rely on a detailed template prior shape and focus on categories with few deformations and rotational symmetries, such as a bottle." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.658, + 0.907, + 0.718 + ], + "angle": 0, + "content": "There exist also multiple works that focus on videospecific deformation model [9, 43, 73, 74]. In contrast to us, they have to re-train for each new instance and cannot find correspondences across instances." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.909, + 0.903 + ], + "angle": 0, + "content": "3D Morphable Models from Synthetic Data. Various work use synthetic data [7] to learn 3D representations. With dense annotation, in terms of a complete surface and annotated canonical frames, several works learn impressive 3D autoencoders [16], [42], and image-conditioned 3D reconstruction [65], [17]. Hereby, implicit 3D representations have outperformed explicit ones and were extended to learn category-specific 3D morphable models [10], [82], and [54]. In contrast, our model does not require synthetic data of several thousands instances per category, but only requires few casually captured object-centric videos. This provides novel 3D morphable models for various categories." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "6470" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.09, + 0.476, + 0.108 + ], + "angle": 0, + "content": "3. A 3D Morphable Model of Neural Features" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.116, + 0.485, + 0.404 + ], + "angle": 0, + "content": "In this section we introduce Common3D, a method for self-supervised learning of 3D morphable models (3DMM) for common objects from only object-centric videos. One of the core challenges that we need to resolve is to find an approach that can deal with highly variable shapes and topologies within the same object category, while at the same time being general enough such that it works for as many categories as possible, hence requiring minimal category-specific heuristics. One key design choice is to represent the morphable shape model with a hybrid volumetric-mesh representation that is flexible, but also enables using mesh-based regularizers during training (Sec. 3.1). Our second key design choice is to represent the object appearance in terms of neural features and not RGB colors (Sec. 3.2), which enables the feature extractor to learn robust 2D-3D correspondences and, hence, facilitates the inverse rendering process. We discuss our training strategy and objectives in Sec. 3.3, and elaborate the inference mechanism in Sec. 3.4. Fig. 2 provides an overview of our method." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.417, + 0.386, + 0.433 + ], + "angle": 0, + "content": "3.1. Morphable Shape Representation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.441, + 0.484, + 0.487 + ], + "angle": 0, + "content": "We represent 3D object priors with a category-level template that has neural features attached to its surface, together with instance-specific geometric deformations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.489, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Category-level Template. One core challenge in our work is to develop a flexible approach for learning the category-level template shape of objects, which generalizes across the highly variable shapes and topologies of common objects. While purely mesh-based representations can have advantages, they are difficult to optimize from scratch, leading to problems that often require ad-hoc heuristics such as re-meshing [13, 72]. To enable a more flexible way of optimizing the category-level 3D shape, we use a hybrid volumetric-mesh representation. This combines the advantages of an implicit and explicit 3D representation. In particular, the object shape is represented with an implicit model that is given as a signed distance field \\(\\phi_{\\mathrm{sdf}}\\) which can adapt to complex geometries flexibly. It can be converted on the fly into an explicit mesh via the Differentiable Marching Tetrahedra (DMTet) algorithm [50]. Specifically, the signed distance field (SDF) is evaluated at the vertices of a tetrahedral grid and converted into a watertight mesh in a fast and differentiable manner. We denote the triangle mesh as \\(\\mathrm{M} = \\{\\mathrm{V},\\mathrm{E},\\mathrm{F}\\}\\) with the vertices \\(\\mathrm{V} = \\{\\mathrm{v}_i\\in \\mathbb{R}^3\\}_{i = 1}^{|\\mathrm{V}|}\\), the edges \\(\\mathrm{E} = \\{(\\mathrm{v}_i,\\mathrm{v}_j)_e\\in \\mathbb{R}^3\\}_{e = 1}^{|\\mathrm{E}|}\\) and the features \\(\\mathrm{F} = \\{\\mathrm{f}_i\\in \\mathbb{R}^D\\}_{i = 1}^{|F|}\\). Once converted to a mesh, the 3D object prior can be rendered into a 2D feature map by applying standard rasterization. The conversion of the SDF into a mesh enables us to use mesh-based regularizers when learning the deformable model, for example to enforce rigidity. Moreover, the vertices provide us" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.151 + ], + "angle": 0, + "content": "direct access to corresponding points among the deformed templates, which will enable us to learn features that represent 2D-3D correspondences effectively during training (see Sec. 3.2 and Sec. 3.3 for details)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.153, + 0.907, + 0.289 + ], + "angle": 0, + "content": "Instance-Level Deformation. We adopt an affine transformation field [82] to model the deformations of the category-level shape to different object instances. In contrast to [82] we transform the vertices of the category-level mesh instead of the signed distance field. This is more efficient, as re-extracting the mesh for each instance is not required. The function \\(\\phi_{a}(\\mathrm{v},\\mathrm{l}):\\mathbb{R}^{3}\\times \\mathrm{L}\\mapsto \\mathbb{R}^{3}\\) maps each vertex \\(\\mathbf{v}\\) from the category-level canonical space together with the image-specific latent code \\(\\mathbf{l}\\) to a new position" + }, + { + "type": "equation", + "bbox": [ + 0.558, + 0.297, + 0.907, + 0.314 + ], + "angle": 0, + "content": "\\[\n\\phi_ {a} (\\mathrm {v}, 1) = \\alpha (\\mathrm {v}, 1) \\odot \\mathrm {v} + \\delta (\\mathrm {v}, 1), \\quad \\alpha , \\delta \\in \\mathbb {R} ^ {3}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.323, + 0.906, + 0.449 + ], + "angle": 0, + "content": "with the intermediate variables for scale \\(\\alpha\\) and translation \\(\\delta\\). Hereby, \\(\\alpha\\) and \\(\\delta\\) are the outputs of an MLP conditioned on the vertex position \\(\\mathbf{v}\\) and the latent code \\(\\mathbf{l}\\). We compute the image-specific latent code \\(\\mathbf{l}\\) with an image feature encoder \\(\\psi_{1}\\) that consists of a DINOv2 backbone and additional convolutional layers (see Fig. 2). We denote the deformed mesh \\(M_{def}(\\mathrm{I}) = \\{\\mathrm{V}_{def}(\\mathrm{I}),\\mathrm{E},\\mathcal{F},\\mathrm{F}\\}\\) with the deformed vertices \\(V_{def}(\\mathrm{I}) = \\{\\phi_a(\\mathrm{v}_i,\\psi_1(\\mathrm{I}))\\in \\mathbb{R}^3\\}_{i = 1}^{|\\mathrm{V}|}\\) and the image I." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.455, + 0.845, + 0.471 + ], + "angle": 0, + "content": "3.2. Appearance as Semantic Feature Field" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.478, + 0.907, + 0.87 + ], + "angle": 0, + "content": "Prior works in the context of 3DMMs traditionally represent the object appearance using RGB colors, which makes it inherently difficult to apply them for solving vision tasks through inverse rendering [46, 49, 55]. A core contribution of our work is to extend 3DMMs to facilitate the inverse rendering process. For this purpose, we represent the appearance using neural features instead of RGB colors. The key advantage of this approach is that the neural features can be invariant to image-specific appearance details and instead focus on encoding semantic correspondences that facilitate the inverse rendering process. Our approach is inspired from prior works [69] that learn deformable 3D models through inverse rendering on the feature-level. These works use DINO [41] features that allow estimating semantic correspondences reasonably well [6]. However, recent studies have shown that DINO features also have limitations, e.g., for differentiating similar object parts, or [78] or in terms of dense prediction [79]. Therefore, we choose to learn an adapter on top of the DINO backbone to further improve the image features for better capturing semantic correspondences. Specifically, we compute a feature map of semantic features as \\( S = \\Phi_w(\\tau(I)) \\in \\mathbb{R}^{D \\times H \\times W} \\), where \\( \\tau \\) is a frozen DINOv2 [41] backbone and \\( \\Phi_w \\) is the adapter with \\( w \\) being the parameters of several convolutional layers (see Sec. 4.1 for the detailed architecture). The adapter is trained jointly with the shape model as described in Sec. 3.3." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.906, + 0.902 + ], + "angle": 0, + "content": "On the 3DMM, we represent the object appearance as a set of feature vectors at every mesh vertex \\( \\mathbf{F} = \\{\\mathbf{f}_i \\in \\)" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "6471" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.089, + 0.485, + 0.199 + ], + "angle": 0, + "content": "\\(\\mathbb{R}^D\\}_{i=1}^{|V|}\\). The features are generated by a feature field [40] \\(\\mathrm{f}_i = \\phi_{\\text{feat}}(\\mathrm{v}_i) : \\mathbb{R}^3 \\mapsto \\mathbb{R}^D\\) that maps each vertex \\(\\mathrm{v}_i\\) to a feature vector \\(\\mathrm{f}_i\\). Importantly, as the vertices of the deformable mesh encode corresponding points in different object instances, the adapter can learn more accurate 2D-3D correspondences compared to self-supervised methods that learn purely in 2D, see Fig. 2." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.207, + 0.281, + 0.223 + ], + "angle": 0, + "content": "3.3. Training Objectives" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.23, + 0.484, + 0.441 + ], + "angle": 0, + "content": "We train Common3D from a collection of videos of objects. Our training process jointly optimizes the category-level template shape, the deformation model, the image feature encoder adapter, and the semantic feature field. Initially, we obtain a 3D point cloud from every video using Structure-from-Motion [48]. We need to align the individual 3D reconstructions of object instances into a canonical coordinate frame to obtain a camera pose annotation \\(\\pi\\) relative to the object that is coherent across videos. We achieve this by adopting the method proposed in [52] to get initial camera pose annotations in an unsupervised manner. In the following, we discuss the geometric objectives, the appearance objective, and finally, the combined objective used for training the model." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.442, + 0.484, + 0.531 + ], + "angle": 0, + "content": "Geometric Objectives. For optimizing the geometry, we formulate 2D reconstruction objectives using the mask and a 3D reconstruction objective using the 3D point cloud from each instance. Simultaneously, we regularize the category-level template's SDF and introduce mesh-based regularizers for highly rigid instance-level deformations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.532, + 0.484, + 0.593 + ], + "angle": 0, + "content": "First, we penalize the pixel-wise deviation \\((u,v)\\) of our rendered mask \\(\\bar{\\mathfrak{m}} (\\mathrm{M}_{def},\\mathrm{I},\\pi ,u,v)\\) with the pose \\(\\pi\\) from the object-centric mask m, provided by PointRend [26]. For this, we compute the mean squared error" + }, + { + "type": "equation", + "bbox": [ + 0.174, + 0.604, + 0.482, + 0.64 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {\\mathrm {m}} \\left(\\mathrm {M} _ {d e f}, \\mathrm {I}, \\pi , \\mathrm {m}, u, v\\right) = \\\\ \\left| \\left| \\bar {\\mathrm {m}} \\left(\\mathrm {M} _ {d e f}, \\mathrm {I}, \\pi , u, v\\right) - \\mathrm {m} (u, v) \\right| \\right| ^ {2}. \\tag {2} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.65, + 0.483, + 0.68 + ], + "angle": 0, + "content": "We also maximize the pixel-wise overlap with the mask's distance transform \\(\\mathrm{m_{dt}}(u,v)\\), as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.177, + 0.69, + 0.482, + 0.725 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {\\mathrm {m d t}} \\left(\\mathrm {M} _ {\\text {d e f}}, \\mathrm {I}, \\pi , \\mathrm {m} _ {\\mathrm {d t}}, u, v\\right) = \\tag {3} \\\\ - \\bar {\\mathrm {m}} (\\mathrm {M} _ {d e f}, \\mathrm {I}, \\pi , u, v) \\mathrm {m} _ {\\mathrm {d t}} (u, v). \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.483, + 0.856 + ], + "angle": 0, + "content": "Hereby, the distance transform equals the distance to the silhouette for each pixel inside the mask, all pixels outside the mask are equal to zero. Note that this objective may only increase the object's volume, leading to a category-level template that encapsulates all objects like a visual hull [30]. Without this objective, and given the conflicting objectives of fitting a single category-level template to multiple different instances, the SDF might result in disconnected parts." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "To ensure that the deformed mesh reconstructs the shape of the individual object instances, we formulate a 3D chamfer distance based on the canonicalized 3D point cloud" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.092, + 0.82, + 0.107 + ], + "angle": 0, + "content": "\\(\\mathcal{P}_{inst}\\), obtained from our object-centric video," + }, + { + "type": "equation", + "bbox": [ + 0.522, + 0.119, + 0.905, + 0.213 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {C D} \\left(\\mathrm {M} _ {d e f}, \\mathrm {I}, \\mathcal {P} _ {i n s t}\\right) = \\frac {1}{| \\mathrm {V} _ {d e f} | + | \\mathcal {P} _ {i n s t} |} \\\\ \\left(\\sum_ {\\mathrm {v} _ {i} \\in \\mathrm {V} _ {d e f}} | | \\mathrm {v} _ {i} - x _ {\\chi (\\mathrm {v} _ {i})} | | _ {2} + \\sum_ {x _ {i} \\in \\mathcal {P} _ {i n s t}} | | x _ {i} - \\mathrm {v} _ {\\chi (x _ {i})} | | _ {2}\\right), \\tag {4} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.214, + 0.905, + 0.245 + ], + "angle": 0, + "content": "with \\(\\chi : \\mathbb{R}^3 \\mapsto \\mathbb{R}^3\\) returning the nearest neighbor of a 3D point to either the vertices \\(V_{\\text{def}}\\) or the point cloud \\(\\mathcal{P}_{\\text{inst}}\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.245, + 0.907, + 0.397 + ], + "angle": 0, + "content": "Further, we optimize the category-level template using the Eikonal regularization as proposed in [15] to ensure the satisfaction of the Eikonal property of a signed distance function, which states that the norm of the gradient equals to one. This is required as DMTet only provides gradients via the extracted mesh close to the surface, which leaves the points more far away from the surface undefined. Therefore, we sample a set of points \\(\\mathcal{P}_{sdf}\\) uniformly in the canonical space and around the extracted vertices, and average over each point \\(x\\in \\mathcal{P}_{sdf}\\)" + }, + { + "type": "equation", + "bbox": [ + 0.587, + 0.409, + 0.905, + 0.427 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {s d f} (\\mathrm {M}, x) = \\left(\\left| \\right| \\nabla \\phi_ {s d f} (x) \\left| \\right| _ {2} - 1\\right) ^ {2}. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.44, + 0.905, + 0.47 + ], + "angle": 0, + "content": "We also penalize large deformations from the category-level template using the mean squared error" + }, + { + "type": "equation", + "bbox": [ + 0.53, + 0.482, + 0.905, + 0.518 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {d e f} (\\mathrm {M}, \\mathrm {M} _ {d e f}, \\mathrm {I}) = \\frac {1}{| \\mathrm {V} |} \\sum_ {\\mathrm {v} \\in \\mathrm {V}} \\left\\| \\mathrm {v} - \\phi_ {a} (\\mathrm {v}, \\psi_ {1} (\\mathrm {I})) \\right\\| ^ {2}. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.531, + 0.905, + 0.575 + ], + "angle": 0, + "content": "To ensure that deformations of connected vertices are smooth, we adopt the point-pair regularization from [82] to only consider connected vertices" + }, + { + "type": "equation", + "bbox": [ + 0.521, + 0.588, + 0.905, + 0.674 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {d e f - s m} (\\mathrm {M}, \\mathrm {M} _ {d e f}, \\mathrm {I}) = \\frac {1}{| \\mathrm {E} |} \\\\ \\sum_ {\\mathrm {v} _ {i}, \\mathrm {v} _ {j} \\in \\mathrm {E}} \\frac {\\left| \\left| \\left[ \\mathrm {v} _ {i} - \\phi_ {a} \\left(\\mathrm {v} _ {i} , \\psi_ {1} (\\mathrm {I})\\right) \\right] - \\left[ \\mathrm {v} _ {j} - \\phi_ {a} \\left(\\mathrm {v} _ {j} , \\psi_ {1} (\\mathrm {I})\\right) \\right] \\right| \\right| _ {2}}{\\left| \\left| \\mathrm {v} _ {i} - \\mathrm {v} _ {j} \\right| \\right| _ {2}}. \\tag {7} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Appearance Objective. In addition to the morphable shape representation, we also train the parameters of the semantic feature field \\(\\phi_{\\text{feat}}(\\mathbf{v}_i)\\) and the parameters of the adapter module \\(\\Phi_w\\) of the image feature encoder. In order to learn the semantic feature field and the feature encoder simultaneously, we cannot directly use a reconstruction objective, as this optimization collapses to the trivial solution of a constant feature for the whole image and the whole surface. We prevent a collapse to the trivial solution by instead optimizing an objective that encourages unique 2D-3D correspondences to emerge, which is inspired from related work on contrastive learning for keypoint detection [4]. In particular, we model for each image feature \\(s_i\\) the 3D correspondence as surface probability \\(\\bar{\\mathfrak{p}}\\), which is computed by the similarity to all vertex features \\(F\\). Moreover," + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6472" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.122 + ], + "angle": 0, + "content": "we add a background feature \\(\\beta\\), which enables our model to differentiate pixelwise between surface and background" + }, + { + "type": "equation", + "bbox": [ + 0.107, + 0.129, + 0.482, + 0.175 + ], + "angle": 0, + "content": "\\[\n\\bar {p} \\left(f _ {j} \\mid s _ {i}, F, \\beta\\right) = \\frac {\\exp \\left(\\kappa \\left\\langle s _ {i} , f _ {j} \\right\\rangle\\right)}{\\sum_ {k} \\exp \\left(\\kappa \\left\\langle s _ {i} , f _ {k} \\right\\rangle\\right) + \\exp \\left(\\kappa \\left\\langle s _ {i} , \\beta \\right\\rangle\\right)}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.177, + 0.484, + 0.312 + ], + "angle": 0, + "content": "where \\(\\kappa\\) is the temperature parameter. Similarly, the background probability \\(\\bar{\\mathfrak{p}} (\\beta |s_i,\\mathrm{F},\\beta)\\) is defined for each image feature. In contrast to previous work, we use farthest point sampling as in [44] to retrieve a subset of vertices and features. To mitigate additional notation, we neglect this sampling in the following. As ground truth, we model the surface probability \\(\\mathbb{P}\\) over the set of vertices as in [39, 62]. Therefore, the surface probability evaluated at the vertex \\(\\mathbf{v}_i\\) is defined by the Euclidean distance to all other vertices" + }, + { + "type": "equation", + "bbox": [ + 0.162, + 0.32, + 0.483, + 0.372 + ], + "angle": 0, + "content": "\\[\n\\mathrm {p} \\left(\\mathrm {v} _ {j} \\mid \\mathrm {v} _ {i}, \\mathrm {V}\\right) = \\frac {\\exp \\left(- \\frac {\\left| \\left| \\mathrm {v} _ {i} - \\mathrm {v} _ {j} \\right| \\right| ^ {2}}{2 \\sigma^ {2}}\\right)}{\\sum_ {k} \\exp \\left(\\frac {\\left| \\left| \\mathrm {v} _ {i} - \\mathrm {v} _ {k} \\right| \\right| ^ {2}}{2 \\sigma^ {2}}\\right)}. \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.38, + 0.484, + 0.471 + ], + "angle": 0, + "content": "Now we can render this surface probability using the camera pose \\(\\hat{\\mathrm{p}} (\\mathrm{v}_j|\\mathrm{M}_{def},\\pi ,u,v)\\). For pixels outside the surface we set the background probability \\(\\hat{\\mathrm{p}} (\\beta |\\mathrm{M}_{def},\\pi ,u,v)\\) to 1. Jointly, we estimate these probabilities \\(\\bar{\\mathrm{p}} (\\mathrm{f}_j|\\Phi_w(\\mathrm{I},u,v),\\mathrm{F},\\beta)\\). Thereupon, we optimize the pixel-wise cross-entropy loss" + }, + { + "type": "equation", + "bbox": [ + 0.154, + 0.48, + 0.482, + 0.531 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {\\text {a p p}} \\left(\\Phi_ {w}, \\mathrm {M} _ {\\text {d e f}}, \\mathrm {I})\\right) = \\\\ - \\sum_ {j} \\log (\\bar {\\mathrm {p}} (f _ {j})) \\hat {\\mathrm {p}} (v _ {j}) - \\log (\\bar {\\mathrm {p}} (\\beta)) \\hat {\\mathrm {p}} (\\beta) \\cdot \\tag {10} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.54, + 0.483, + 0.571 + ], + "angle": 0, + "content": "Combined Objective. First, we only optimize the category-level template using only the geometric losses" + }, + { + "type": "equation", + "bbox": [ + 0.104, + 0.58, + 0.482, + 0.597 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\lambda_ {C D} \\mathcal {L} _ {C D} + \\lambda_ {\\mathrm {m}} \\mathcal {L} _ {\\mathrm {m}} + \\lambda_ {\\mathrm {m d t}} \\mathcal {L} _ {\\mathrm {m d t}} + \\lambda_ {s d f} \\mathcal {L} _ {s d f}, \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.606, + 0.294, + 0.62 + ], + "angle": 0, + "content": "with the balancing weights \\(\\lambda_{i}\\)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.621, + 0.483, + 0.667 + ], + "angle": 0, + "content": "Second, we optimize the category-level feature field, the adapters, and the instance-level deformations using the weighted sum of geometric and appearance losses" + }, + { + "type": "equation", + "bbox": [ + 0.104, + 0.675, + 0.483, + 0.711 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} = \\lambda_ {\\text {a p p}} \\mathcal {L} _ {\\text {a p p}} + \\lambda_ {C D} \\mathcal {L} _ {C D} + \\lambda_ {\\mathrm {m}} \\mathcal {L} _ {\\mathrm {m}} + \\lambda_ {\\mathrm {m d t}} \\mathcal {L} _ {\\mathrm {m d t}} \\tag {12} \\\\ + \\lambda_ {s d f} \\mathcal {L} _ {s d f} + \\lambda_ {d e f} \\mathcal {L} _ {d e f} + \\lambda_ {d e f - s m} \\mathcal {L} _ {d e f - s m}. \\\\ \\end{array}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.719, + 0.373, + 0.735 + ], + "angle": 0, + "content": "3.4. Inference via Inverse Rendering" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.742, + 0.484, + 0.864 + ], + "angle": 0, + "content": "During inference, our main application is to infer the shape latent \\( l \\) and the camera pose \\( \\pi \\). The shape latent can be directly predicted using the trained latent encoder \\( \\psi_{1} \\). While the pose could alternatively also be predicted with a feedforward neural network, other works have demonstrated that inverse rendering can significantly improve the results [63]. Therefore, we optimize the pixelwise surface probabilities during inference applying the reconstruction loss" + }, + { + "type": "equation", + "bbox": [ + 0.21, + 0.871, + 0.483, + 0.903 + ], + "angle": 0, + "content": "\\[\n\\max _ {\\pi} \\sum_ {u, v} \\bar {p} \\left(f _ {r} \\mid f _ {i}, F, \\beta\\right), \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.906, + 0.168 + ], + "angle": 0, + "content": "with the rendered image feature \\(\\mathrm{f}_r = \\Psi_r(\\mathrm{M}_{def},\\pi ,u,v)\\) and the image feature extracted from the test image \\(\\mathrm{f}_i = \\Phi_w(\\tau (\\mathrm{I},u,v))\\) . Note that we use three degrees of freedom for parametrizing the object rotation. Following [62], we initialize our pose using uniform sampling." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.18, + 0.646, + 0.197 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.205, + 0.907, + 0.356 + ], + "angle": 0, + "content": "We conduct extensive experiments and evaluate the effectiveness of Common3D on various downstream tasks for a variety of object categories. We use the same trained model for all tasks, demonstrating the universal utilization of combining geometry with correspondence-aware appearance in one single model. We describe our experimental setup in Sec. 4.1. Subsequently, we demonstrate that our model improves over prior self-supervised work in 3D object pose estimation (Sec. 4.2), in instance segmentation (Sec. 4.3), and in semantic correspondence estimation (Sec. 4.4)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.365, + 0.714, + 0.38 + ], + "angle": 0, + "content": "4.1. Experimental Details" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.388, + 0.907, + 0.705 + ], + "angle": 0, + "content": "Training Dataset. To train Common3D, we use the object-centric videos provided by the Common Objects in 3D (CO3D) dataset [45]. For each category, the videos capture a large variety of instances. In total, it provides 19k videos across 50 categories, with each video containing approximately 100 to 200 frames. Each video is supplemented by point clouds and relative camera poses using Structure-from-Motion [48] and object masks using PointRend [26]. We choose the same videos as used in [52], where videos that were captured too narrow to or not focusing on the object were filtered out, and only videos were retained where the full object is visible. For each category, we use a maximum of 50 videos. Due to the filtering we use fewer videos for the following ten categories: 17 for \"remote\", 15 for \"mouse\", 16 for \"tv\", 7 for \"toilet\", 41 for \"toybus\", 28 for \"hairdryer\", 49 for \"couch\", and 23 for \"cellphone\". Before training, we center each object's coordinate frame and normalize its scale such that its point cloud ranges from -1 to 1 on the axis with the maximum range. We achieve adopt the method proposed in [52] to obtain canonical camera pose annotations across videos in an unsupervised manner." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Evaluation on In-the-Wild Datasets. We evaluate our model at various tasks on in-the-wild data. In particular, we use three common datasets PASCAL3D+ [70], ObjectNet3D [71], and SPair-71k [37]. PASCAL3D+ and ObjectNet3D provide pose and segmentation annotations per image, which enables evaluation of 3D object pose estimation as well as instance segmentation. While PASCAL3D+ yields 12 rigid classes of PASCAL VOC 2012, ObjectNet3D covers, beyond that, 100 categories. We use the same 20 common categories across ObjectNet3D and CO3D as used in [52]. This even includes the domain gap of a toy bus in CO3D to a real one in PASCAL3D+ and ObjectNet3D. In total, we evaluate on 6233 images of PAS" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6473" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.089, + 0.481, + 0.176 + ], + "angle": 0, + "content": "
Method∅∅∅∅∅○AVG
Sup.StarMap [83]83.294.475.479.868.885.882.5
VoGE [63]82.698.190.594.987.583.990.9
Unsup.ZSP [14]61.721.442.652.943.139.046.0
UOP3D [52]58.479.351.976.667.053.169.2
Ours67.282.967.880.978.151.175.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.186, + 0.483, + 0.243 + ], + "angle": 0, + "content": "Table 1. Evaluation of 3D object pose estimation on 7 categories of PASCAL3D+, we supplement the missing results in the appendix. The reported metric is the \\(30^{\\circ}\\) accuracy. Note, that ZSP [14] requires also depth and runs an order of a magnitude slower." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.258, + 0.481, + 0.378 + ], + "angle": 0, + "content": "
草草草草草草草草草草草草草草
ZSP [14]23.150.860.346.436.833.021.7
UOP3D [52]18.062.198.154.652.238.214.1
Ours25.577.099.051.865.944.719.4
草草草草草草草草草草草草AVG
ZSP [14]46.860.528.841.625.856.337.9
UOP3D [52]26.953.344.754.415.539.653.2
Ours34.554.238.552.415.765.052.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.388, + 0.483, + 0.445 + ], + "angle": 0, + "content": "Table 2. Evaluation of 3D object pose estimation on 20 categories of ObjectNet3D. We supplement the missing results in the appendix. We report the \\(30^{\\circ}\\) accuracy. Note, that ZSP [14] requires also depth input and runs an order of a magnitude slower." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.471, + 0.483, + 0.547 + ], + "angle": 0, + "content": "CAL3D+ using the same subset as in [61], and on 12039 images of ObjectNet3D. We center all objects, as proposed in [83]. Further, we evaluate our feature adapter on the semantic correspondence task of SPair71k across 6 common categories, which we validate on 3898 image pairs." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.548, + 0.483, + 0.791 + ], + "angle": 0, + "content": "Implementation Details. As backbone we leverage DINOv2 [41] with the ViT-S model [2], for which we freeze all 21M parameters. To compare with prior work on semantic correspondence estimation Sec. 4.4, we additionally trained one ViT-B model. From an input resolution of \\(448 \\times 448\\), the backbone maps to a \\(32 \\times 32\\) feature map. Our adapter, consisting of three ResNet blocks, upsamples the \\(32 \\times 32\\) feature map to \\(64 \\times 64\\). For efficient differentiable rendering, we make use of the rasterizer proposed by [29], which achieves differentiability using anti-aliasing. For each category-specific morphable model we train, using Adam [25], for 20 epochs to optimize the template shape, and another 20 to optimize the appearance together with the deformation model. We use early stopping on a hold-out set comprising \\(10\\%\\) of the videos. On a single NVIDIA GeForce RTX 2080 the training does not exceed 10 hours." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.802, + 0.373, + 0.817 + ], + "angle": 0, + "content": "4.2. In-the-Wild 3D Pose Estimation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.484, + 0.903 + ], + "angle": 0, + "content": "We evaluate our 3D object prior at single-image 3D pose estimation on in-the-wild data. We report the pose estimation results using the \\(30^{\\circ}\\) accuracy in Tab. 1 for PASCAL3D+ and in Tab. 2 for ObjectNet3D. In Tab. 1, we also report 3D pose estimation results from two supervised" + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.089, + 0.905, + 0.139 + ], + "angle": 0, + "content": "
201918171615AVG
UOP3D [52]26.066.579.142.371.763.673.0
Ours29.364.782.552.276.965.977.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.149, + 0.907, + 0.192 + ], + "angle": 0, + "content": "Table 3. Evaluation of instance segmentation on 7 categories of PASCAL3D+. We report the IoU in percentage with the occlusion-free instance mask provided by the ground truth pose and mesh." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.206, + 0.905, + 0.304 + ], + "angle": 0, + "content": "
井井井井井井井
UOP3D [52]68.949.979.267.742.969.650.9
Ours69.454.182.868.153.366.853.6
井井井井井井井AVG
UOP3D [52]76.962.464.161.462.564.773.864.0
Ours83.170.662.771.765.366.778.367.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.313, + 0.907, + 0.371 + ], + "angle": 0, + "content": "Table 4. Evaluation of instance segmentation on 20 categories of ObjectNet3D, we supplement the missing results in the appendix. We report the IoU in percentage with the occlusion-free instance mask given by rendering the annotated mesh in the annotated pose." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.397, + 0.907, + 0.61 + ], + "angle": 0, + "content": "methods [63, 83] that are trained from the PASCAL3D+ training set. Notably, all unsupervised methods are trained from CO3D data, as they require videos as input during training, and hence are evaluated in an out-of-domain setting. Our method improves over prior work on most of the categories except for categories with few shape variance across instances such as cellphone, microwave, and toaster. On average, the \\(30^{\\circ}\\) accuracy improves significantly over prior work from \\(69.2\\%\\) to \\(75.3\\%\\) by \\(+6.3\\) points on PASCAL3D+ and from \\(52.4\\%\\) to \\(56.8\\%\\) by \\(+4.4\\) points on ObjectNet3D. We attribute this performance increase to the correspondence learning with a deformable shape, which facilitates the training and, therefore, generalizes better. This is illustrated in the qualitative examples in Fig. 3" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.619, + 0.82, + 0.636 + ], + "angle": 0, + "content": "4.3. In-the-Wild Instance Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.643, + 0.907, + 0.841 + ], + "angle": 0, + "content": "We validate the deformation capabilities of Common3D using the instance segmentation masks of PASCAL3D+ in Tab. 3 and ObjectNet3D in Tab. 4. We compare our model with UOP3D [52], as it is the current state-of-the-art at self-supervised learning 3D object priors for common objects, and was also trained using CO3D videos. Common3D outperforms UOP3D on almost all categories and on average increases the IoU by 3.8 percentage points from \\(60.3\\%\\) to \\(64.1\\%\\) on PASCAL3D+ and by 3 percentage points, from \\(64\\%\\) to \\(67\\%\\) on ObjectNet3D. A qualitative comparison of our method against UOP3D is depicted in Fig. 3. It shows that Common3D achieves much more accurate highly accurate despite a large variability in the object instances." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.851, + 0.753, + 0.867 + ], + "angle": 0, + "content": "4.4. Semantic Correspondence" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.906, + 0.902 + ], + "angle": 0, + "content": "We demonstrate the significant benefit of our semantic feature adapter by evaluating its performance on semantic cor" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6474" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.096, + 0.09, + 0.904, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.317, + 0.908, + 0.347 + ], + "angle": 0, + "content": "Figure 3. Qualitative results on the ObjectNet3D dataset. In the second row the results of our method are illustrated, in the third the results of UOP3D [52]. Notably, our method fits the object more accurately, resulting in improved 3D pose and segmentation accuracy." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.369, + 0.481, + 0.439 + ], + "angle": 0, + "content": "
6073.367.246.366.042.259.4
SpherMaps [35]61.273.351.536.261.024.247.9
DINOv2 [41]62.052.366.244.260.731.954.0
Ours60.360.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.45, + 0.483, + 0.493 + ], + "angle": 0, + "content": "Table 5. Evaluation of the PCK@0.1 on 6 categories of SPair-71k [37]. Note, that SpherMaps [35] uses coarse viewpoint annotations, while DINOv2 and our approach do not require any." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.531, + 0.483, + 0.743 + ], + "angle": 0, + "content": "respondence estimation in Tab. 5. We evaluate our model compared to the self-supervised DINOv2 model as well as SpherMaps [35], which require coarse viewpoint annotations. We follow the evaluation protocol of SpherMaps and find correspondences by weighting the image features similarities from the DINOv2 backbone and the feature similarities from our adapter. We weight the backbone's features with 0.8 and ours with 0.2. On average, we can largely increase the PCK@0.1 compared to the self-supervised baseline DINOv2 by \\(6.1\\%\\) percentage points from \\(47.9\\%\\) to \\(54\\%\\). For some categories, we can even achieve a competitive performance compared to the supervised baseline SpherMaps. We believe that one major challenge remains the domain gap between the training and test data." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.77, + 0.196, + 0.785 + ], + "angle": 0, + "content": "4.5. Ablation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.483, + 0.903 + ], + "angle": 0, + "content": "In Tab. 6, we ablate our model design choices regarding the adapter, the appearance loss as well as the instance deformation model. We observe, that instance-level deformation has only a marginal relevance for the pose estimation, but proves essential for segmentation and semantic correspondence estimation. Furthermore, the adapter and the dense appearance loss are significantly improving our method." + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.369, + 0.905, + 0.47 + ], + "angle": 0, + "content": "
30°10°IoUPCK@0.1
w/o Adapt.25.96.660.532.4
w/o Adapt., w Rec. Loss13.92.156.232.4
w/o Dense App. Loss56.712.258.845.3
w/o Inst. Deform.75.820.258.645.3
Ours75.321.064.147.8
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.481, + 0.907, + 0.634 + ], + "angle": 0, + "content": "Table 6. Ablation of our model design choices. We evaluate the 3D pose accuracy and instance segmentation on PASCAL3D+ using \\(30^{\\circ}\\), \\(10^{\\circ}\\) accuracy and IoU. Further, the semantic correspondences are evaluated on SPair71-k using PCK@0.1. We observe that the adapter is crucial for our method, as well as the dense appearance loss. Further, our instance deformation model yields a minor disadvantage for the coarse \\(30^{\\circ}\\) pose estimation, while improving the fine-grained \\(10^{\\circ}\\) pose estimation, and proves essential for the instance segmentation and the semantic correspondences. Note that in this standard setting we use the ViT-S backbone, in contrast to the results reported in Tab. 5." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.664, + 0.634, + 0.68 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.687, + 0.907, + 0.852 + ], + "angle": 0, + "content": "We have introduced Common3D, a new model that can learn 3D morphable models from a collection of casually captured videos completely self-supervised. This model can, at test time, estimate the 3D object shape, 2D-3D correspondences, and the 3D object pose. Common3D connects two different subfields of computer vision, self-supervised representation learning and the learning of 3D deformable object priors. We believe that in the long term, this may unify many vision tasks into a common inference process, while also creating a better 3D spatial understanding in vision models." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Acknowledgment. This work was funded by the German Research Foundation (DFG) - 468670075, and 499552394 (SFB 1597 - Small Data)." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6475" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.116, + 0.484, + 0.169 + ], + "angle": 0, + "content": "[1] Oswald Aldrian and William AP Smith. Inverse rendering of faces with a 3d morphable model. IEEE transactions on pattern analysis and machine intelligence, 35(5):1080-1093, 2012. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.172, + 0.484, + 0.211 + ], + "angle": 0, + "content": "[2] Dosovitskiy Alexey. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv: 2010.11929, 2020.7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.215, + 0.484, + 0.269 + ], + "angle": 0, + "content": "[3] Dragomir Anguelov, Praveen Srinivasan, Daphne Koller, Sebastian Thrun, Jim Rodgers, and James Davis. Scape: shape completion and animation of people. In ACM SIGGRAPH 2005 Papers, pages 408-416. 2005. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.271, + 0.483, + 0.325 + ], + "angle": 0, + "content": "[4] Yutong Bai, Angtian Wang, Adam Kortylewski, and Alan Yuille. Coke: Contrastive learning for robust keypoint detection. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 65-74, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.327, + 0.483, + 0.381 + ], + "angle": 0, + "content": "[5] V BLANZ and T VETTER. A morphable model for the synthesis of 3d faces. In Computer graphics proceedings, annual conference series, pages 187-194. Association for Computing Machinery SIGGRAPH, 1999. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.383, + 0.483, + 0.451 + ], + "angle": 0, + "content": "[6] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.453, + 0.483, + 0.52 + ], + "angle": 0, + "content": "[7] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.523, + 0.483, + 0.578 + ], + "angle": 0, + "content": "[8] Zhiqin Chen and Hao Zhang. Learning implicit fields for generative shape modeling. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5939-5948, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.581, + 0.483, + 0.648 + ], + "angle": 0, + "content": "[9] Devikalyan Das, Christopher Wewer, Raza Yunus, Eddy Ilg, and Jan Eric Lenssen. Neural parametric gaussians for monocular non-rigid object reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10715-10725, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.65, + 0.483, + 0.716 + ], + "angle": 0, + "content": "[10] Yu Deng, Jiaolong Yang, and Xin Tong. Deformed implicit field: Modeling 3d shapes with learned dense correspondence. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10286-10296, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.72, + 0.483, + 0.8 + ], + "angle": 0, + "content": "[11] Bernhard Egger, Sandro Schonborn, Andreas Schneider, Adam Kortylewski, Andreas Morel-Forster, Clemens Blumer, and Thomas Vetter. Occlusion-aware 3d morphable models and an illumination prior for face image analysis. International Journal of Computer Vision, 126:1269-1287, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.804, + 0.483, + 0.871 + ], + "angle": 0, + "content": "[12] Shubham Goel, Angjoo Kanazawa, and Jitendra Malik. Shape and viewpoint without keypoints. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XV 16, pages 88-104. Springer, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.874, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[13] Shubham Goel, Georgia Gkioxari, and Jitendra Malik. Differentiable stereopsis: Meshes from multiple views using" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.116, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.548, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "differentiable rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8635-8644, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.136, + 0.905, + 0.191 + ], + "angle": 0, + "content": "[14] Walter Goodwin, Sagar Vaze, Ioannis Havoutis, and Ingmar Posner. Zero-shot category-level object pose estimation. In European Conference on Computer Vision, pages 516-532. Springer, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.193, + 0.905, + 0.234 + ], + "angle": 0, + "content": "[15] Amos Gropp, Lior Yariv, Niv Haim, Matan Atzmon, and Yaron Lipman. Implicit geometric regularization for learning shapes. arXiv preprint arXiv:2002.10099, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.236, + 0.905, + 0.291 + ], + "angle": 0, + "content": "[16] Thibault Groueix, Matthew Fisher, Vladimir G Kim, Bryan C Russell, and Mathieu Aubry. Atlasnet: a papiermâché approach to learning 3d surface generation (2018). arXiv preprint arXiv:1802.05384, 11, 1802. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.293, + 0.905, + 0.347 + ], + "angle": 0, + "content": "[17] Paul Henderson, Vagia Tsiminaki, and Christoph H Lampert. Leveraging 2d data to learn textured 3d mesh generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7498-7507, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.35, + 0.905, + 0.404 + ], + "angle": 0, + "content": "[18] Tao Hu, Liwei Wang, Xiaogang Xu, Shu Liu, and Jiaya Jia. Self-supervised 3d mesh reconstruction from single images. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6002-6011, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.407, + 0.907, + 0.475 + ], + "angle": 0, + "content": "[19] Wei-Chih Hung, Varun Jampani, Sifei Liu, Pavlo Molchanov, Ming-Hsuan Yang, and Jan Kautz. Scops: Self-supervised co-part segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 869-878, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.477, + 0.905, + 0.544 + ], + "angle": 0, + "content": "[20] Tomas Jakab, Ruining Li, Shangzhe Wu, Christian Rupprecht, and Andrea Vedaldi. Farm3d: Learning articulated 3d animals by distilling 2d diffusion. In 2024 International Conference on 3D Vision (3DV), pages 852-861. IEEE, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.548, + 0.905, + 0.603 + ], + "angle": 0, + "content": "[21] Hanbyul Joo, Tomas Simon, and Yaser Sheikh. Total capture: A 3d deformation model for tracking faces, hands, and bodies. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8320-8329, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.605, + 0.905, + 0.672 + ], + "angle": 0, + "content": "[22] Angjoo Kanazawa, Shubham Tulsiani, Alexei A Efros, and Jitendra Malik. Learning category-specific mesh reconstruction from image collections. In Proceedings of the European Conference on Computer Vision (ECCV), pages 371-386, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.676, + 0.905, + 0.743 + ], + "angle": 0, + "content": "[23] Angjoo Kanazawa, Shubham Tulsiani, Alexei A Efros, and Jitendra Malik. Learning category-specific mesh reconstruction from image collections. In Proceedings of the European Conference on Computer Vision (ECCV), pages 371-386, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.746, + 0.905, + 0.813 + ], + "angle": 0, + "content": "[24] Abhishek Kar, Shubham Tulsiani, Joao Carreira, and Jitendra Malik. Category-specific object reconstruction from a single image. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1966-1974, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.817, + 0.905, + 0.843 + ], + "angle": 0, + "content": "[25] Diederik P Kingma. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.846, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[26] Alexander Kirillov, Yuxin Wu, Kaiming He, and Ross Girshick. Pointrend: Image segmentation as rendering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9799-9808, 2020. 5, 6" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.956 + ], + "angle": 0, + "content": "6476" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[27] Filippos Kokkinos and Iasonas Kokkinos. To the point: Correspondence-driven monocular 3d category reconstruction. Advances in Neural Information Processing Systems, 34:7760-7772, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.15, + 0.482, + 0.218 + ], + "angle": 0, + "content": "[28] Nilesh Kulkarni, Abhinav Gupta, David F Fouhey, and Shubham Tulsiani. Articulation-aware canonical surface mapping. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 452-461, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.221, + 0.482, + 0.276 + ], + "angle": 0, + "content": "[29] Samuli Laine, Janne Hellsten, Tero Karras, Yeongho Seol, Jaakko Lehtinen, and Timo Aila. Modular primitives for high-performance differentiable rendering. ACM Transactions on Graphics, 39(6), 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.278, + 0.482, + 0.319 + ], + "angle": 0, + "content": "[30] Aldo Laurentini. The visual hull concept for silhouette-based image understanding. IEEE Transactions on pattern analysis and machine intelligence, 16(2):150-162, 1994. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.321, + 0.482, + 0.404 + ], + "angle": 0, + "content": "[31] Xueting Li, Sifei Liu, Kihwan Kim, Shalini De Mello, Varun Jampani, Ming-Hsuan Yang, and Jan Kautz. Self-supervised single-view 3d reconstruction via semantic consistency. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIV 16, pages 677-693. Springer, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.406, + 0.482, + 0.475 + ], + "angle": 0, + "content": "[32] Zizhang Li, Dor Litvak, Ruining Li, Yunzhi Zhang, Tomas Jakab, Christian Rupprecht, Shangzhe Wu, Andrea Vedaldi, and Jiajun Wu. Learning the 3d fauna of the web. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9752-9762, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.477, + 0.482, + 0.546 + ], + "angle": 0, + "content": "[33] Yebin Liu, Juergen Gall, Carsten Stoll, Qionghai Dai, Hans-Peter Seidel, and Christian Theobalt. Markerless motion capture of multiple characters using multiview image segmentation. IEEE transactions on pattern analysis and machine intelligence, 35(11):2720-2735, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.548, + 0.482, + 0.602 + ], + "angle": 0, + "content": "[34] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J Black. Smpl: a skinned multiperson linear model. ACM Transactions on Graphics (TOG), 34(6):1-16, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.604, + 0.482, + 0.672 + ], + "angle": 0, + "content": "[35] Octave Mariotti, Oisin Mac Aodha, and Hakan Bilen. Improving semantic correspondence with viewpoint-guided spherical maps. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19521-19530, 2024. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.675, + 0.482, + 0.743 + ], + "angle": 0, + "content": "[36] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4460-4470, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.746, + 0.482, + 0.787 + ], + "angle": 0, + "content": "[37] Juhong Min, Jongmin Lee, Jean Ponce, and Minsu Cho. Spair-71k: A large-scale benchmark for semantic correspondence. arXiv preprint arXiv:1908.10543, 2019. 2, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.789, + 0.482, + 0.844 + ], + "angle": 0, + "content": "[38] Tom Monnier, Matthew Fisher, Alexei A Efros, and Mathieu Aubry. Share with thy neighbors: Single-view reconstruction by cross-instance consistency. In European Conference on Computer Vision, pages 285-303. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[39] Natalia Neverova, David Novotny, Marc Szafraniec, Vasil Khalidov, Patrick Labatut, and Andrea Vedaldi. Continuous surface embeddings. Advances in Neural Information Processing Systems, 33:17258-17270, 2020. 6" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.161 + ], + "angle": 0, + "content": "[40] Michael Oechsle, Lars Mescheder, Michael Niemeyer, Thilo Strauss, and Andreas Geiger. Texture fields: Learning texture representations in function space. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4531-4540, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.163, + 0.905, + 0.232 + ], + "angle": 0, + "content": "[41] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 2, 4, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.233, + 0.905, + 0.302 + ], + "angle": 0, + "content": "[42] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 165-174, 2019. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.303, + 0.905, + 0.37 + ], + "angle": 0, + "content": "[43] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5865-5874, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.372, + 0.905, + 0.427 + ], + "angle": 0, + "content": "[44] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems, 30, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.428, + 0.905, + 0.51 + ], + "angle": 0, + "content": "[45] Jeremy Reizenstein, Roman Shapovalov, Philipp Henzler, Luca Sbordone, Patrick Labatut, and David Novotny. Common objects in 3d: Large-scale learning and evaluation of real-life 3d category reconstruction. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10901-10911, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.511, + 0.905, + 0.58 + ], + "angle": 0, + "content": "[46] Lukasz Romaszko, Christopher KI Williams, Pol Moreno, and Pushmeet Kohli. Vision-as-inverse-graphics: Obtaining a rich 3d explanation of a scene from a single image. In Proceedings of the IEEE International Conference on Computer Vision Workshops, pages 851–859, 2017. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.581, + 0.905, + 0.649 + ], + "angle": 0, + "content": "[47] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.65, + 0.905, + 0.692 + ], + "angle": 0, + "content": "[48] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-Motion Revisited. In Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.693, + 0.905, + 0.747 + ], + "angle": 0, + "content": "[49] Sandro Schonborn, Bernhard Egger, Andreas Morel-Forster, and Thomas Vetter. Markov chain monte carlo for automated face image analysis. International Journal of Computer Vision, 123:160–183, 2017. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.748, + 0.905, + 0.816 + ], + "angle": 0, + "content": "[50] Tianchang Shen, Jun Gao, Kangxue Yin, Ming-Yu Liu, and Sanja Fidler. Deep marching tetrahedra: a hybrid representation for high-resolution 3d shape synthesis. Advances in Neural Information Processing Systems, 34:6087-6101, 2021. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.818, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[51] Aleksandar Shtedritski, Christian Rupprecht, and Andrea Vedaldi. Shic: Shape-image correspondences with no key-point supervision. arXiv preprint arXiv:2407.18907, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[52] Leonhard Sommer, Artur Jesslen, Eddy Ilg, and Adam Kortylewski. Unsupervised learning of category-level 3d pose" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "6477" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.126, + 0.092, + 0.483, + 0.134 + ], + "angle": 0, + "content": "from object-centric videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22787-22796, 2024. 2, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.483, + 0.205 + ], + "angle": 0, + "content": "[53] Keqiang Sun, Dor Litvak, Yunzhi Zhang, Hongsheng Li, Ji-jun Wu, and Shangzhe Wu. Ponomation: Learning articulated 3d animal motions from unlabeled online videos. In European Conference on Computer Vision, pages 100-119. Springer, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.207, + 0.482, + 0.276 + ], + "angle": 0, + "content": "[54] Shanlin Sun, Kun Han, Deying Kong, Hao Tang, Xiangyi Yan, and Xiaohui Xie. Topology-preserving shape reconstruction and registration via neural diffeomorphic flow. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20845-20855, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.278, + 0.482, + 0.36 + ], + "angle": 0, + "content": "[55] Ayush Tewari, Michael Zollhofer, Hyeongwoo Kim, Pablo Garrido, Florian Bernard, Patrick Perez, and Christian Theobalt. Mofa: Model-based deep convolutional face autoencoder for unsupervised monocular reconstruction. In Proceedings of the IEEE international conference on computer vision workshops, pages 1274-1283, 2017. 1, 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.362, + 0.482, + 0.445 + ], + "angle": 0, + "content": "[56] Ayush Tewari, Florian Bernard, Pablo Garrido, Gaurav Bharaj, Mohamed Elgharib, Hans-Peter Seidel, Patrick Pérez, Michael Zollhofer, and Christian Theobalt. Fml: Face model learning from videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10812-10822, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.447, + 0.483, + 0.516 + ], + "angle": 0, + "content": "[57] Ayush Tewari, Hans-Peter Seidel, Mohamed Elgharib, Christian Theobalt, et al. Learning complete 3d morphable face models from images and videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3361-3371, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.518, + 0.483, + 0.6 + ], + "angle": 0, + "content": "[58] Ayush Tewari, Xingang Pan, Ohad Fried, Maneesh Agrawala, Christian Theobalt, et al. Disentangled3d: Learning a 3d generative model with disentangled geometry and appearance from monocular images. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1516-1525, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.602, + 0.482, + 0.657 + ], + "angle": 0, + "content": "[59] Shubham Tulsiani, Abhishek Kar, Joao Carreira, and Jitendra Malik. Learning category-specific deformable 3d models for object reconstruction. IEEE transactions on pattern analysis and machine intelligence, 39(4):719-731, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.66, + 0.482, + 0.701 + ], + "angle": 0, + "content": "[60] Shubham Tulsiani, Nilesh Kulkarni, and Abhinav Gupta. Implicit mesh reconstruction from unannotated image collections. arXiv preprint arXiv:2007.08504, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.703, + 0.482, + 0.771 + ], + "angle": 0, + "content": "[61] Angtian Wang, Yihong Sun, Adam Kortylewski, and Alan L Yuille. Robust object detection under occlusion with context-aware compositionalnets. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12645-12654, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.773, + 0.482, + 0.814 + ], + "angle": 0, + "content": "[62] Angtian Wang, Adam Kortylewski, and Alan Yuille. Nemo: Neural mesh models of contrastive features for robust 3d pose estimation. arXiv preprint arXiv:2101.12378, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.817, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[63] Angtian Wang, Peng Wang, Jian Sun, Adam Kortylewski, and Alan Yuille. Voge: a differentiable volume renderer using gaussian ellipsoids for analysis-by-synthesis. arXiv preprint arXiv:2205.15401, 2022. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.902 + ], + "angle": 0, + "content": "[64] Haowen Wang, Zhipeng Fan, Zhen Zhao, Zhengping Che, Zhiyuan Xu, Dong Liu, Feifei Feng, Yakun Huang, Xiuquan" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "Qiao, and Jian Tang. Dtf-net: Category-level pose estimation and shape reconstruction via deformable template field. In Proceedings of the 31st ACM International Conference on Multimedia, pages 3676-3685, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.149, + 0.905, + 0.216 + ], + "angle": 0, + "content": "[65] Nanyang Wang, Yinda Zhang, Zhuwen Li, Yanwei Fu, Wei Liu, and Yu-Gang Jiang. Pixel2mesh: Generating 3d mesh models from single rgb images. In Proceedings of the European conference on computer vision (ECCV), pages 52-67, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.219, + 0.907, + 0.287 + ], + "angle": 0, + "content": "[66] Shangzhe Wu, Christian Rupprecht, and Andrea Vedaldi. Unsupervised learning of probably symmetric deformable 3d objects from images in the wild. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1-10, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.289, + 0.907, + 0.357 + ], + "angle": 0, + "content": "[67] Shangzhe Wu, Christian Rupprecht, and Andrea Vedaldi. Unsupervised learning of probably symmetric deformable 3d objects from images in the wild. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1-10, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.359, + 0.905, + 0.412 + ], + "angle": 0, + "content": "[68] Shangzhe Wu, Tomas Jakab, Christian Rupprecht, and Andrea Vedaldi. Dove: Learning deformable 3d objects by watching videos. International Journal of Computer Vision, 131(10):2623-2634, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.414, + 0.905, + 0.482 + ], + "angle": 0, + "content": "[69] Shangzhe Wu, Ruining Li, Tomas Jakab, Christian Rupprecht, and Andrea Vedaldi. Magicpony: Learning articulated 3d animals in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8792-8802, 2023. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.484, + 0.905, + 0.538 + ], + "angle": 0, + "content": "[70] Yu Xiang, Roozbeh Mottaghi, and Silvio Savarese. Beyond Pascal: A benchmark for 3d object detection in the wild. In IEEE winter conference on applications of computer vision, pages 75-82. IEEE, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.54, + 0.905, + 0.635 + ], + "angle": 0, + "content": "[71] Yu Xiang, Wonhui Kim, Wei Chen, Jingwei Ji, Christopher Choy, Hao Su, Roozbeh Mottaghi, Leonidas Guibas, and Silvio Savarese. Objectnet3d: A large scale database for 3d object recognition. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part VIII 14, pages 160-176. Springer, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.637, + 0.905, + 0.719 + ], + "angle": 0, + "content": "[72] Gengshan Yang, Deqing Sun, Varun Jampani, Daniel Vlasic, Forrester Cole, Huiwen Chang, Deva Ramanan, William T Freeman, and Ce Liu. Lasr: Learning articulated shape reconstruction from a monocular video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15980-15989, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.721, + 0.905, + 0.788 + ], + "angle": 0, + "content": "[73] Gengshan Yang, Deqing Sun, Varun Jampani, Daniel Vlasic, Forrester Cole, Ce Liu, and Deva Ramanan. Viser: Video-specific surface embeddings for articulated 3d shape reconstruction. Advances in Neural Information Processing Systems, 34:19326-19338, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.79, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[74] Gengshan Yang, Minh Vo, Natalia Neverova, Deva Ramanan, Andrea Vedaldi, and Hanbyul Joo. Banmo: Building animatable 3d neural models from many casual videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2863-2873, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.902 + ], + "angle": 0, + "content": "[75] Chun-Han Yao, Wei-Chih Hung, Yuanzhen Li, Michael Rubinstein, Ming-Hsuan Yang, and Varun Jampani. Lassie:" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "6478" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.134 + ], + "angle": 0, + "content": "Learning articulated shapes from sparse image ensemble via 3d part discovery. Advances in Neural Information Processing Systems, 35:15296-15308, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.483, + 0.219 + ], + "angle": 0, + "content": "[76] Chun-Han Yao, Wei-Chih Hung, Yuanzhen Li, Michael Rubinstein, Ming-Hsuan Yang, and Varun Jampani. Hi-lassie: High-fidelity articulated shape and skeleton discovery from sparse image ensemble. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4853-4862, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.221, + 0.482, + 0.261 + ], + "angle": 0, + "content": "[77] Alan Yuille and Daniel Kersten. Vision as bayesian inference: analysis by synthesis? Trends in cognitive sciences, 10(7):301-308, 2006. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.263, + 0.482, + 0.344 + ], + "angle": 0, + "content": "[78] Junyi Zhang, Charles Herrmann, Junhwa Hur, Eric Chen, Varun Jampani, Deqing Sun, and Ming-Hsuan Yang. Telling left from right: Identifying geometry-aware semantic correspondence. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3076-3085, 2024. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.347, + 0.482, + 0.415 + ], + "angle": 0, + "content": "[79] Junyi Zhang, Charles Herrmann, Junhwa Hur, Luisa Polania Cabrera, Varun Jampani, Deqing Sun, and Ming-Hsuan Yang. A tale of two features: Stable diffusion complements dino for zero-shot semantic correspondence. Advances in Neural Information Processing Systems, 36, 2024. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.418, + 0.482, + 0.471 + ], + "angle": 0, + "content": "[80] Kaifeng Zhang, Yang Fu, Shubhankar Borse, Hong Cai, Fatih Porikli, and Xiaolong Wang. Self-supervised geometric correspondence for category-level 6d object pose estimation in the wild. arXiv preprint arXiv:2210.07199, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.474, + 0.483, + 0.529 + ], + "angle": 0, + "content": "[81] Yunzhi Zhang, Shangzhe Wu, Noah Snavely, and Jiajun Wu. Seeing a rose in five thousand ways. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 962-971, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.531, + 0.483, + 0.586 + ], + "angle": 0, + "content": "[82] Zerong Zheng, Tao Yu, Qionghai Dai, and Yebin Liu. Deep implicit templates for 3d shape representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1429-1439, 2021. 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.588, + 0.482, + 0.642 + ], + "angle": 0, + "content": "[83] Xingyi Zhou, Arjun Karpur, Linjie Luo, and Qixing Huang. Starmap for category-agnostic keypoint and viewpoint estimation. In Proceedings of the European Conference on Computer Vision (ECCV), pages 318–334, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.644, + 0.482, + 0.711 + ], + "angle": 0, + "content": "[84] Silvia Zuffi, Angjoo Kanazawa, David W Jacobs, and Michael J Black. 3d menagerie: Modeling the 3d shape and pose of animals. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6365-6373, 2017. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.711 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "6479" + } + ] +] \ No newline at end of file diff --git a/2025/Common3D_ Self-Supervised Learning of 3D Morphable Models for Common Objects in Neural Feature Space/ecdc4e97-8d79-41e2-abfb-5c0cebf6aa2a_origin.pdf b/2025/Common3D_ Self-Supervised Learning of 3D Morphable Models for Common Objects in Neural Feature Space/ecdc4e97-8d79-41e2-abfb-5c0cebf6aa2a_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..4a569d0a2cc9e64a43c2ebafdb64fdd373307bf8 --- /dev/null +++ b/2025/Common3D_ Self-Supervised Learning of 3D Morphable Models for Common Objects in Neural Feature Space/ecdc4e97-8d79-41e2-abfb-5c0cebf6aa2a_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:63a41e457bba8aa04c9e9fda70e06ab8c2a6ecedf4c520eef5009b5fd488186d +size 8430939 diff --git a/2025/Common3D_ Self-Supervised Learning of 3D Morphable Models for Common Objects in Neural Feature Space/full.md b/2025/Common3D_ Self-Supervised Learning of 3D Morphable Models for Common Objects in Neural Feature Space/full.md new file mode 100644 index 0000000000000000000000000000000000000000..ca43d90a266ed378758d7fb1d0a1882a66f29146 --- /dev/null +++ b/2025/Common3D_ Self-Supervised Learning of 3D Morphable Models for Common Objects in Neural Feature Space/full.md @@ -0,0 +1,343 @@ +# Common3D: Self-Supervised Learning of 3D Morphable Models for Common Objects in Neural Feature Space + +Leonhard Sommer + +Olaf Dunkel + +Christian Theobalt 2 + +Adam Kortylewski 1,2 + +1University of Freiburg + +$^{2}$ Max Planck Institute for Informatics + +![](images/01a12ec6fb8d4c297aa800494cbbf4a7addfdc1f272cef0f4fcf3164415677bc.jpg) +Figure 1. Common3D learns category-specific 3D morphable models from few casually captured videos completely self-supervised, and can estimate the 3D object shape (visualized from two viewpoints), 2D-3D correspondences, and the 3D object pose via inverse rendering. + +# Abstract + +3D morphable models (3DMMs) are a powerful tool to represent the possible shapes and appearances of an object category. Given a single test image, 3DMMs can be used to solve various tasks, such as predicting the 3D shape, pose, semantic correspondence, and instance segmentation of an object. Unfortunately, 3DMMs are only available for very few object categories that are of particular interest, like faces or human bodies, as they require a demanding 3D data acquisition and category-specific training process. In contrast, we introduce a new method, Common3D, that learns 3DMMs of common objects in a fully self-supervised manner from a collection of object-centric videos. For this purpose, our model represents objects as a learned 3D template mesh and a deformation field that is parameterized as an image-conditioned neural network. Different from prior works, Common3D represents the object appearance with neural features instead of RGB colors, which enables the learning of more generalizable representations through an abstraction from pixel intensities. Importantly, we train the appearance features using a contrastive + +objective by exploiting the correspondences defined through the deformable template mesh. This leads to higher quality correspondence features compared to related works and a significantly improved model performance at estimating 3D object pose and semantic correspondence. Common3D is the first completely self-supervised method that can solve various vision tasks in a zero-shot manner. We release all code at github.com/GenIntel/common3d. + +# 1. Introduction + +Estimating the 3D shape and pose of an object from a single image requires prior knowledge about the possible shapes and appearances of an object category. Learning such prior models is a long-standing problem in computer vision, with 3D morphable models (3DMMs) [5] emerging as a very successful approach to address this problem. A popular application of 3DMMs is to perform single image 3D reconstruction via inverse rendering [1, 5, 55], also referred to as analysis-by-synthesis approach [77]. This enables solving under-constrained vision tasks, for example predicting an object's shape behind an occluder [11] or estimating the + +3D object shape and pose from a single image. However, training 3DMMs requires sophisticated data acquisition setups [3, 21, 33], involving laser scanners or multi-view studio captures, and a significant manual effort in cleaning and registration of the data [5, 34, 84]. This approach is difficult to scale to more common everyday objects. + +An alternative approach that emerged recently is to learn a 3D prior from 2D images only [58, 67, 74, 81]. However, learning such priors from in-the-wild images is a major challenge and typically requires significant category-specific design choices, with various solutions having been proposed for different object categories [23, 32, 68, 69, 81]. This makes it difficult to scale to the long tail of common objects that can appear in natural images. + +We introduce Common3D, a new way of integrating self-supervised learning, neural rendering, and 3D modeling to learn 3D object priors for common objects from a collection of object-centric videos. Once trained, the model can estimate 2D-3D correspondences, it can be leveraged for inverse rendering to estimate the 3D pose, and instance segmentation of common objects from a single image, see Fig. 1. At the core of our method is a category-specific 3D morphable model (3DMM) that represents the possible variations of different object instances. The 3DMM is represented using a learned prototypical 3D shape and a shape deformation field that is parameterized as an image-conditioned neural network. Training our model requires canonical camera poses, which we estimate unsupervised by adopting the method of [52]. + +One core challenge for learning 3DMMs of common objects is to be flexible enough to capture the highly variable shapes and topologies of various object categories, as well as the possible variations of instances within one category. We address this issue by using a hybrid volumetric mesh representation based on DMTet [50]. Hence, the shape is defined volumetrically in canonical space, but a mesh is extracted on the fly for morphing and rendering the model. This allows to retain the flexibility of implicit neural representations for capturing highly variable shapes and topologies [8, 36, 42], while also enabling the use of meshes for an efficient rendering and regularization during training. + +Another important challenge is to perform inverse rendering with 3D object priors, as it requires the optimization of a highly non-convex image-based reconstruction loss [46, 49, 55]. We address this problem in Common3D by representing the object appearance as a semantic feature field instead of RGB colors. Performing inverse rendering on the feature-level is easier because the features can directly encode 2D-3D correspondences. While related work [32, 69] directly uses DINO [41] as appearance features, recent studies showed that DINO has severe limitations at estimating semantic correspondence accurately [35, 78]. Therefore, we train an additional adapter module on top of + +DINO by using the 3D object prior to further enhance the feature extractor at accurately estimating 2D-3D correspondences. This leads to higher quality semantic correspondence features and a largely enhanced model performance. + +To summarize, we make the following contributions: (1) We propose a new way of integrating self-supervised learning, neural rendering, and 3D modeling to enable the joint learning of 3D object priors for common objects from a collection of object-centric videos, (2) an effective mechanism for learning correspondence features by training a 3D object prior jointly with an additional adapter module on top of DINO, using a local contrastive objective that exploits the correspondences estimated by the 3D model, and (3) Common3D is the first completely self-supervised method that can estimate the shape, pose, and semantic correspondence of common objects from a single image. + +# 2. Related Work + +Our model has two core components: An adapter module that enhances correspondence learning and a 3D morphable model that is learned without supervision. Therefore, we first provide an overview of correspondence learning methods and subsequently discuss related work on learning 3D morphable models from images and from synthetic data. + +Self-Supervised Correspondence Learning. Finding semantic correspondences across instances of an object category is a challenging computer vision task [37]. Self-supervised models like DINOv2 [41] and Stable Diffusion [47] were shown to have remarkable capabilities at estimating zero-shot semantic correspondences [79]. Despite these advances, unsupervised models are still not performing well on all categories. This can be accounted to ambiguous features of similar parts [35] or the lower performance for viewpoint changes [78]. Previous works have addressed this by mapping the 2D features to a coarse spherical representation of the object [35] or by applying heuristic augmentation methods to account for the viewpoint dependency [78]. Recent related work [51] showed that by leveraging 3D shape representations, the 2D-3D correspondence mapping can be improved significantly. However, their approach requires a detailed categorical mesh and is limited to a small selected set of animals. Similarly, [52] uses multiple views from object-centric videos to account for the viewpoint dependency. Further, they obtain a simple rigid 3D template and learn 2D-3D correspondences to estimate the 3D pose of an object. In contrast, our proposed Common3D learns an underlying deformable 3D shape representation self-supervised, which enables us to generalize to various common objects. Moreover, our approach of learning correspondence features by training a 3D object prior jointly with an additional adapter module on top of DINO enables high quality features to emerge. + +![](images/d191c29ecff0a63279bd5c07df407856295f3ae5528a9dbf43c7457267ae1d00.jpg) +Figure 2. Method Overview. At the core of our method is a category-level template with semantic features that is acquired using a neural SDF with Differentiable Marching Tetrahedra (DMTet) where features are attached using a feature field. The category-level template is morphed using an MLP that is conditioned on a latent code $l$ and rotated using the pose $\pi$ . The pose $\pi$ is estimated unsupervised at training time and predicted at inference time. DINOv2 serves as a feature encoder and its output is processed in two branches: One for estimating the latent code $l$ with a block of convolutional layers; and another branch serves as an adapter to enhance the DINO features for better correspondence learning. As a result, the features from the adapter are better suited for finding correspondences. For example, the tire features of DINO have a similar encoding and are, hence, ambiguous, whereas it is much more distinct after the adapter. We also visualize the training objectives: The appearance objective compares two surface probabilities: 1) The geometric probability $p(\mathrm{v}_j|\mathrm{v}_i,\mathrm{V})$ , as defined via the Euclidean distances of the mesh vertices; and 2) the appearance probability $\bar{p} (\mathrm{f}_j|s_i,\mathrm{F},\beta)$ , acquired by comparing the 2D image features $s_i$ with the surface features $F$ . The remaining geometric objectives enforce the 3D shape and the projected mask to be consistent with the data acquired from a SFM pipeline. + +![](images/85d242279d4038a3267766926e54de4c6624ef9af42258a3d3421ce2d55c3c0b.jpg) + +Learning 3D Morphable Models from Images. There was remarkable progress made in the self-supervised learning of category-specific 3D morphable models from images. Originally, approaches used camera viewpoint, 2D keypoints, and masks to predict the 3D shape [24, 59] and texture [22, 64]. Several works achieved to mitigate the requirement of camera viewpoint and 2D keypoint annotations by exploiting symmetries [12, 18, 31, 60], by adding multi-view viewpoint hypothesis [12], by requiring consistent self-supervised co-part segmentation [31] using [19], by cycle-consistent surface maps [27, 28, 60], or by temporal consistency for video clips [56, 57, 68]. Few works even achieve to not require mask annotations, [38, 58, 66], by being constrained to front views of faces [66], or coarse-to-fine learning of texture and shape [38]. However, most of these works focus on modeling the object in very fine detail, such that they can generate object instances down to the RGB pixel-level, making the inverse rendering process inherently difficult [46, 49, 55]. In contrast, our focus lies on self-supervised learning of a 3D morphable model that is particularly well suited for solving vision tasks through inverse rendering. Therefore, we represent the object appearance using neural features instead of RGB colors, which facilitates the inverse rendering process as the features learn to support the inverse rendering process by encoding 2D-3D correspondences. + +Another line of work learns 3D object priors without viewpoint annotations by replicating DINO features [20, 32, 53, 69, 75, 76]. In contrast to us, these works are restricted to animal categories. Further, they are not enhancing the image feature encoder for learning correspondences. The work of Zhang et al. is related to ours [80], as it also enhances the image feature encoder, but not in an inverse rendering setting. Further, they rely on a detailed template prior shape and focus on categories with few deformations and rotational symmetries, such as a bottle. + +There exist also multiple works that focus on videospecific deformation model [9, 43, 73, 74]. In contrast to us, they have to re-train for each new instance and cannot find correspondences across instances. + +3D Morphable Models from Synthetic Data. Various work use synthetic data [7] to learn 3D representations. With dense annotation, in terms of a complete surface and annotated canonical frames, several works learn impressive 3D autoencoders [16], [42], and image-conditioned 3D reconstruction [65], [17]. Hereby, implicit 3D representations have outperformed explicit ones and were extended to learn category-specific 3D morphable models [10], [82], and [54]. In contrast, our model does not require synthetic data of several thousands instances per category, but only requires few casually captured object-centric videos. This provides novel 3D morphable models for various categories. + +# 3. A 3D Morphable Model of Neural Features + +In this section we introduce Common3D, a method for self-supervised learning of 3D morphable models (3DMM) for common objects from only object-centric videos. One of the core challenges that we need to resolve is to find an approach that can deal with highly variable shapes and topologies within the same object category, while at the same time being general enough such that it works for as many categories as possible, hence requiring minimal category-specific heuristics. One key design choice is to represent the morphable shape model with a hybrid volumetric-mesh representation that is flexible, but also enables using mesh-based regularizers during training (Sec. 3.1). Our second key design choice is to represent the object appearance in terms of neural features and not RGB colors (Sec. 3.2), which enables the feature extractor to learn robust 2D-3D correspondences and, hence, facilitates the inverse rendering process. We discuss our training strategy and objectives in Sec. 3.3, and elaborate the inference mechanism in Sec. 3.4. Fig. 2 provides an overview of our method. + +# 3.1. Morphable Shape Representation + +We represent 3D object priors with a category-level template that has neural features attached to its surface, together with instance-specific geometric deformations. + +Category-level Template. One core challenge in our work is to develop a flexible approach for learning the category-level template shape of objects, which generalizes across the highly variable shapes and topologies of common objects. While purely mesh-based representations can have advantages, they are difficult to optimize from scratch, leading to problems that often require ad-hoc heuristics such as re-meshing [13, 72]. To enable a more flexible way of optimizing the category-level 3D shape, we use a hybrid volumetric-mesh representation. This combines the advantages of an implicit and explicit 3D representation. In particular, the object shape is represented with an implicit model that is given as a signed distance field $\phi_{\mathrm{sdf}}$ which can adapt to complex geometries flexibly. It can be converted on the fly into an explicit mesh via the Differentiable Marching Tetrahedra (DMTet) algorithm [50]. Specifically, the signed distance field (SDF) is evaluated at the vertices of a tetrahedral grid and converted into a watertight mesh in a fast and differentiable manner. We denote the triangle mesh as $\mathrm{M} = \{\mathrm{V},\mathrm{E},\mathrm{F}\}$ with the vertices $\mathrm{V} = \{\mathrm{v}_i\in \mathbb{R}^3\}_{i = 1}^{|\mathrm{V}|}$ , the edges $\mathrm{E} = \{(\mathrm{v}_i,\mathrm{v}_j)_e\in \mathbb{R}^3\}_{e = 1}^{|\mathrm{E}|}$ and the features $\mathrm{F} = \{\mathrm{f}_i\in \mathbb{R}^D\}_{i = 1}^{|F|}$ . Once converted to a mesh, the 3D object prior can be rendered into a 2D feature map by applying standard rasterization. The conversion of the SDF into a mesh enables us to use mesh-based regularizers when learning the deformable model, for example to enforce rigidity. Moreover, the vertices provide us + +direct access to corresponding points among the deformed templates, which will enable us to learn features that represent 2D-3D correspondences effectively during training (see Sec. 3.2 and Sec. 3.3 for details). + +Instance-Level Deformation. We adopt an affine transformation field [82] to model the deformations of the category-level shape to different object instances. In contrast to [82] we transform the vertices of the category-level mesh instead of the signed distance field. This is more efficient, as re-extracting the mesh for each instance is not required. The function $\phi_{a}(\mathrm{v},\mathrm{l}):\mathbb{R}^{3}\times \mathrm{L}\mapsto \mathbb{R}^{3}$ maps each vertex $\mathbf{v}$ from the category-level canonical space together with the image-specific latent code $\mathbf{l}$ to a new position + +$$ +\phi_ {a} (\mathrm {v}, 1) = \alpha (\mathrm {v}, 1) \odot \mathrm {v} + \delta (\mathrm {v}, 1), \quad \alpha , \delta \in \mathbb {R} ^ {3}, \tag {1} +$$ + +with the intermediate variables for scale $\alpha$ and translation $\delta$ . Hereby, $\alpha$ and $\delta$ are the outputs of an MLP conditioned on the vertex position $\mathbf{v}$ and the latent code $\mathbf{l}$ . We compute the image-specific latent code $\mathbf{l}$ with an image feature encoder $\psi_{1}$ that consists of a DINOv2 backbone and additional convolutional layers (see Fig. 2). We denote the deformed mesh $M_{def}(\mathrm{I}) = \{\mathrm{V}_{def}(\mathrm{I}),\mathrm{E},\mathcal{F},\mathrm{F}\}$ with the deformed vertices $V_{def}(\mathrm{I}) = \{\phi_a(\mathrm{v}_i,\psi_1(\mathrm{I}))\in \mathbb{R}^3\}_{i = 1}^{|\mathrm{V}|}$ and the image I. + +# 3.2. Appearance as Semantic Feature Field + +Prior works in the context of 3DMMs traditionally represent the object appearance using RGB colors, which makes it inherently difficult to apply them for solving vision tasks through inverse rendering [46, 49, 55]. A core contribution of our work is to extend 3DMMs to facilitate the inverse rendering process. For this purpose, we represent the appearance using neural features instead of RGB colors. The key advantage of this approach is that the neural features can be invariant to image-specific appearance details and instead focus on encoding semantic correspondences that facilitate the inverse rendering process. Our approach is inspired from prior works [69] that learn deformable 3D models through inverse rendering on the feature-level. These works use DINO [41] features that allow estimating semantic correspondences reasonably well [6]. However, recent studies have shown that DINO features also have limitations, e.g., for differentiating similar object parts, or [78] or in terms of dense prediction [79]. Therefore, we choose to learn an adapter on top of the DINO backbone to further improve the image features for better capturing semantic correspondences. Specifically, we compute a feature map of semantic features as $S = \Phi_w(\tau(I)) \in \mathbb{R}^{D \times H \times W}$ , where $\tau$ is a frozen DINOv2 [41] backbone and $\Phi_w$ is the adapter with $w$ being the parameters of several convolutional layers (see Sec. 4.1 for the detailed architecture). The adapter is trained jointly with the shape model as described in Sec. 3.3. + +On the 3DMM, we represent the object appearance as a set of feature vectors at every mesh vertex $\mathbf{F} = \{\mathbf{f}_i \in$ + +$\mathbb{R}^D\}_{i=1}^{|V|}$ . The features are generated by a feature field [40] $\mathrm{f}_i = \phi_{\text{feat}}(\mathrm{v}_i) : \mathbb{R}^3 \mapsto \mathbb{R}^D$ that maps each vertex $\mathrm{v}_i$ to a feature vector $\mathrm{f}_i$ . Importantly, as the vertices of the deformable mesh encode corresponding points in different object instances, the adapter can learn more accurate 2D-3D correspondences compared to self-supervised methods that learn purely in 2D, see Fig. 2. + +# 3.3. Training Objectives + +We train Common3D from a collection of videos of objects. Our training process jointly optimizes the category-level template shape, the deformation model, the image feature encoder adapter, and the semantic feature field. Initially, we obtain a 3D point cloud from every video using Structure-from-Motion [48]. We need to align the individual 3D reconstructions of object instances into a canonical coordinate frame to obtain a camera pose annotation $\pi$ relative to the object that is coherent across videos. We achieve this by adopting the method proposed in [52] to get initial camera pose annotations in an unsupervised manner. In the following, we discuss the geometric objectives, the appearance objective, and finally, the combined objective used for training the model. + +Geometric Objectives. For optimizing the geometry, we formulate 2D reconstruction objectives using the mask and a 3D reconstruction objective using the 3D point cloud from each instance. Simultaneously, we regularize the category-level template's SDF and introduce mesh-based regularizers for highly rigid instance-level deformations. + +First, we penalize the pixel-wise deviation $(u,v)$ of our rendered mask $\bar{\mathfrak{m}} (\mathrm{M}_{def},\mathrm{I},\pi ,u,v)$ with the pose $\pi$ from the object-centric mask m, provided by PointRend [26]. For this, we compute the mean squared error + +$$ +\begin{array}{l} \mathcal {L} _ {\mathrm {m}} \left(\mathrm {M} _ {d e f}, \mathrm {I}, \pi , \mathrm {m}, u, v\right) = \\ \left| \left| \bar {\mathrm {m}} \left(\mathrm {M} _ {d e f}, \mathrm {I}, \pi , u, v\right) - \mathrm {m} (u, v) \right| \right| ^ {2}. \tag {2} \\ \end{array} +$$ + +We also maximize the pixel-wise overlap with the mask's distance transform $\mathrm{m_{dt}}(u,v)$ , as follows: + +$$ +\begin{array}{l} \mathcal {L} _ {\mathrm {m d t}} \left(\mathrm {M} _ {\text {d e f}}, \mathrm {I}, \pi , \mathrm {m} _ {\mathrm {d t}}, u, v\right) = \tag {3} \\ - \bar {\mathrm {m}} (\mathrm {M} _ {d e f}, \mathrm {I}, \pi , u, v) \mathrm {m} _ {\mathrm {d t}} (u, v). \\ \end{array} +$$ + +Hereby, the distance transform equals the distance to the silhouette for each pixel inside the mask, all pixels outside the mask are equal to zero. Note that this objective may only increase the object's volume, leading to a category-level template that encapsulates all objects like a visual hull [30]. Without this objective, and given the conflicting objectives of fitting a single category-level template to multiple different instances, the SDF might result in disconnected parts. + +To ensure that the deformed mesh reconstructs the shape of the individual object instances, we formulate a 3D chamfer distance based on the canonicalized 3D point cloud + +$\mathcal{P}_{inst}$ , obtained from our object-centric video, + +$$ +\begin{array}{l} \mathcal {L} _ {C D} \left(\mathrm {M} _ {d e f}, \mathrm {I}, \mathcal {P} _ {i n s t}\right) = \frac {1}{| \mathrm {V} _ {d e f} | + | \mathcal {P} _ {i n s t} |} \\ \left(\sum_ {\mathrm {v} _ {i} \in \mathrm {V} _ {d e f}} | | \mathrm {v} _ {i} - x _ {\chi (\mathrm {v} _ {i})} | | _ {2} + \sum_ {x _ {i} \in \mathcal {P} _ {i n s t}} | | x _ {i} - \mathrm {v} _ {\chi (x _ {i})} | | _ {2}\right), \tag {4} \\ \end{array} +$$ + +with $\chi : \mathbb{R}^3 \mapsto \mathbb{R}^3$ returning the nearest neighbor of a 3D point to either the vertices $V_{\text{def}}$ or the point cloud $\mathcal{P}_{\text{inst}}$ . + +Further, we optimize the category-level template using the Eikonal regularization as proposed in [15] to ensure the satisfaction of the Eikonal property of a signed distance function, which states that the norm of the gradient equals to one. This is required as DMTet only provides gradients via the extracted mesh close to the surface, which leaves the points more far away from the surface undefined. Therefore, we sample a set of points $\mathcal{P}_{sdf}$ uniformly in the canonical space and around the extracted vertices, and average over each point $x\in \mathcal{P}_{sdf}$ + +$$ +\mathcal {L} _ {s d f} (\mathrm {M}, x) = \left(\left| \right| \nabla \phi_ {s d f} (x) \left| \right| _ {2} - 1\right) ^ {2}. \tag {5} +$$ + +We also penalize large deformations from the category-level template using the mean squared error + +$$ +\mathcal {L} _ {d e f} (\mathrm {M}, \mathrm {M} _ {d e f}, \mathrm {I}) = \frac {1}{| \mathrm {V} |} \sum_ {\mathrm {v} \in \mathrm {V}} \left\| \mathrm {v} - \phi_ {a} (\mathrm {v}, \psi_ {1} (\mathrm {I})) \right\| ^ {2}. \tag {6} +$$ + +To ensure that deformations of connected vertices are smooth, we adopt the point-pair regularization from [82] to only consider connected vertices + +$$ +\begin{array}{l} \mathcal {L} _ {d e f - s m} (\mathrm {M}, \mathrm {M} _ {d e f}, \mathrm {I}) = \frac {1}{| \mathrm {E} |} \\ \sum_ {\mathrm {v} _ {i}, \mathrm {v} _ {j} \in \mathrm {E}} \frac {\left| \left| \left[ \mathrm {v} _ {i} - \phi_ {a} \left(\mathrm {v} _ {i} , \psi_ {1} (\mathrm {I})\right) \right] - \left[ \mathrm {v} _ {j} - \phi_ {a} \left(\mathrm {v} _ {j} , \psi_ {1} (\mathrm {I})\right) \right] \right| \right| _ {2}}{\left| \left| \mathrm {v} _ {i} - \mathrm {v} _ {j} \right| \right| _ {2}}. \tag {7} \\ \end{array} +$$ + +Appearance Objective. In addition to the morphable shape representation, we also train the parameters of the semantic feature field $\phi_{\text{feat}}(\mathbf{v}_i)$ and the parameters of the adapter module $\Phi_w$ of the image feature encoder. In order to learn the semantic feature field and the feature encoder simultaneously, we cannot directly use a reconstruction objective, as this optimization collapses to the trivial solution of a constant feature for the whole image and the whole surface. We prevent a collapse to the trivial solution by instead optimizing an objective that encourages unique 2D-3D correspondences to emerge, which is inspired from related work on contrastive learning for keypoint detection [4]. In particular, we model for each image feature $s_i$ the 3D correspondence as surface probability $\bar{\mathfrak{p}}$ , which is computed by the similarity to all vertex features $F$ . Moreover, + +we add a background feature $\beta$ , which enables our model to differentiate pixelwise between surface and background + +$$ +\bar {p} \left(f _ {j} \mid s _ {i}, F, \beta\right) = \frac {\exp \left(\kappa \left\langle s _ {i} , f _ {j} \right\rangle\right)}{\sum_ {k} \exp \left(\kappa \left\langle s _ {i} , f _ {k} \right\rangle\right) + \exp \left(\kappa \left\langle s _ {i} , \beta \right\rangle\right)}, \tag {8} +$$ + +where $\kappa$ is the temperature parameter. Similarly, the background probability $\bar{\mathfrak{p}} (\beta |s_i,\mathrm{F},\beta)$ is defined for each image feature. In contrast to previous work, we use farthest point sampling as in [44] to retrieve a subset of vertices and features. To mitigate additional notation, we neglect this sampling in the following. As ground truth, we model the surface probability $\mathbb{P}$ over the set of vertices as in [39, 62]. Therefore, the surface probability evaluated at the vertex $\mathbf{v}_i$ is defined by the Euclidean distance to all other vertices + +$$ +\mathrm {p} \left(\mathrm {v} _ {j} \mid \mathrm {v} _ {i}, \mathrm {V}\right) = \frac {\exp \left(- \frac {\left| \left| \mathrm {v} _ {i} - \mathrm {v} _ {j} \right| \right| ^ {2}}{2 \sigma^ {2}}\right)}{\sum_ {k} \exp \left(\frac {\left| \left| \mathrm {v} _ {i} - \mathrm {v} _ {k} \right| \right| ^ {2}}{2 \sigma^ {2}}\right)}. \tag {9} +$$ + +Now we can render this surface probability using the camera pose $\hat{\mathrm{p}} (\mathrm{v}_j|\mathrm{M}_{def},\pi ,u,v)$ . For pixels outside the surface we set the background probability $\hat{\mathrm{p}} (\beta |\mathrm{M}_{def},\pi ,u,v)$ to 1. Jointly, we estimate these probabilities $\bar{\mathrm{p}} (\mathrm{f}_j|\Phi_w(\mathrm{I},u,v),\mathrm{F},\beta)$ . Thereupon, we optimize the pixel-wise cross-entropy loss + +$$ +\begin{array}{l} \mathcal {L} _ {\text {a p p}} \left(\Phi_ {w}, \mathrm {M} _ {\text {d e f}}, \mathrm {I})\right) = \\ - \sum_ {j} \log (\bar {\mathrm {p}} (f _ {j})) \hat {\mathrm {p}} (v _ {j}) - \log (\bar {\mathrm {p}} (\beta)) \hat {\mathrm {p}} (\beta) \cdot \tag {10} \\ \end{array} +$$ + +Combined Objective. First, we only optimize the category-level template using only the geometric losses + +$$ +\mathcal {L} = \lambda_ {C D} \mathcal {L} _ {C D} + \lambda_ {\mathrm {m}} \mathcal {L} _ {\mathrm {m}} + \lambda_ {\mathrm {m d t}} \mathcal {L} _ {\mathrm {m d t}} + \lambda_ {s d f} \mathcal {L} _ {s d f}, \tag {11} +$$ + +with the balancing weights $\lambda_{i}$ + +Second, we optimize the category-level feature field, the adapters, and the instance-level deformations using the weighted sum of geometric and appearance losses + +$$ +\begin{array}{l} \mathcal {L} = \lambda_ {\text {a p p}} \mathcal {L} _ {\text {a p p}} + \lambda_ {C D} \mathcal {L} _ {C D} + \lambda_ {\mathrm {m}} \mathcal {L} _ {\mathrm {m}} + \lambda_ {\mathrm {m d t}} \mathcal {L} _ {\mathrm {m d t}} \tag {12} \\ + \lambda_ {s d f} \mathcal {L} _ {s d f} + \lambda_ {d e f} \mathcal {L} _ {d e f} + \lambda_ {d e f - s m} \mathcal {L} _ {d e f - s m}. \\ \end{array} +$$ + +# 3.4. Inference via Inverse Rendering + +During inference, our main application is to infer the shape latent $l$ and the camera pose $\pi$ . The shape latent can be directly predicted using the trained latent encoder $\psi_{1}$ . While the pose could alternatively also be predicted with a feedforward neural network, other works have demonstrated that inverse rendering can significantly improve the results [63]. Therefore, we optimize the pixelwise surface probabilities during inference applying the reconstruction loss + +$$ +\max _ {\pi} \sum_ {u, v} \bar {p} \left(f _ {r} \mid f _ {i}, F, \beta\right), \tag {13} +$$ + +with the rendered image feature $\mathrm{f}_r = \Psi_r(\mathrm{M}_{def},\pi ,u,v)$ and the image feature extracted from the test image $\mathrm{f}_i = \Phi_w(\tau (\mathrm{I},u,v))$ . Note that we use three degrees of freedom for parametrizing the object rotation. Following [62], we initialize our pose using uniform sampling. + +# 4. Experiments + +We conduct extensive experiments and evaluate the effectiveness of Common3D on various downstream tasks for a variety of object categories. We use the same trained model for all tasks, demonstrating the universal utilization of combining geometry with correspondence-aware appearance in one single model. We describe our experimental setup in Sec. 4.1. Subsequently, we demonstrate that our model improves over prior self-supervised work in 3D object pose estimation (Sec. 4.2), in instance segmentation (Sec. 4.3), and in semantic correspondence estimation (Sec. 4.4). + +# 4.1. Experimental Details + +Training Dataset. To train Common3D, we use the object-centric videos provided by the Common Objects in 3D (CO3D) dataset [45]. For each category, the videos capture a large variety of instances. In total, it provides 19k videos across 50 categories, with each video containing approximately 100 to 200 frames. Each video is supplemented by point clouds and relative camera poses using Structure-from-Motion [48] and object masks using PointRend [26]. We choose the same videos as used in [52], where videos that were captured too narrow to or not focusing on the object were filtered out, and only videos were retained where the full object is visible. For each category, we use a maximum of 50 videos. Due to the filtering we use fewer videos for the following ten categories: 17 for "remote", 15 for "mouse", 16 for "tv", 7 for "toilet", 41 for "toybus", 28 for "hairdryer", 49 for "couch", and 23 for "cellphone". Before training, we center each object's coordinate frame and normalize its scale such that its point cloud ranges from -1 to 1 on the axis with the maximum range. We achieve adopt the method proposed in [52] to obtain canonical camera pose annotations across videos in an unsupervised manner. + +Evaluation on In-the-Wild Datasets. We evaluate our model at various tasks on in-the-wild data. In particular, we use three common datasets PASCAL3D+ [70], ObjectNet3D [71], and SPair-71k [37]. PASCAL3D+ and ObjectNet3D provide pose and segmentation annotations per image, which enables evaluation of 3D object pose estimation as well as instance segmentation. While PASCAL3D+ yields 12 rigid classes of PASCAL VOC 2012, ObjectNet3D covers, beyond that, 100 categories. We use the same 20 common categories across ObjectNet3D and CO3D as used in [52]. This even includes the domain gap of a toy bus in CO3D to a real one in PASCAL3D+ and ObjectNet3D. In total, we evaluate on 6233 images of PAS + +
Method∅∅∅∅∅○AVG
Sup.StarMap [83]83.294.475.479.868.885.882.5
VoGE [63]82.698.190.594.987.583.990.9
Unsup.ZSP [14]61.721.442.652.943.139.046.0
UOP3D [52]58.479.351.976.667.053.169.2
Ours67.282.967.880.978.151.175.3
+ +Table 1. Evaluation of 3D object pose estimation on 7 categories of PASCAL3D+, we supplement the missing results in the appendix. The reported metric is the $30^{\circ}$ accuracy. Note, that ZSP [14] requires also depth and runs an order of a magnitude slower. + +
草草草草草草草草草草草草草草
ZSP [14]23.150.860.346.436.833.021.7
UOP3D [52]18.062.198.154.652.238.214.1
Ours25.577.099.051.865.944.719.4
草草草草草草草草草草草草AVG
ZSP [14]46.860.528.841.625.856.337.9
UOP3D [52]26.953.344.754.415.539.653.2
Ours34.554.238.552.415.765.052.4
+ +CAL3D+ using the same subset as in [61], and on 12039 images of ObjectNet3D. We center all objects, as proposed in [83]. Further, we evaluate our feature adapter on the semantic correspondence task of SPair71k across 6 common categories, which we validate on 3898 image pairs. + +Implementation Details. As backbone we leverage DINOv2 [41] with the ViT-S model [2], for which we freeze all 21M parameters. To compare with prior work on semantic correspondence estimation Sec. 4.4, we additionally trained one ViT-B model. From an input resolution of $448 \times 448$ , the backbone maps to a $32 \times 32$ feature map. Our adapter, consisting of three ResNet blocks, upsamples the $32 \times 32$ feature map to $64 \times 64$ . For efficient differentiable rendering, we make use of the rasterizer proposed by [29], which achieves differentiability using anti-aliasing. For each category-specific morphable model we train, using Adam [25], for 20 epochs to optimize the template shape, and another 20 to optimize the appearance together with the deformation model. We use early stopping on a hold-out set comprising $10\%$ of the videos. On a single NVIDIA GeForce RTX 2080 the training does not exceed 10 hours. + +# 4.2. In-the-Wild 3D Pose Estimation + +We evaluate our 3D object prior at single-image 3D pose estimation on in-the-wild data. We report the pose estimation results using the $30^{\circ}$ accuracy in Tab. 1 for PASCAL3D+ and in Tab. 2 for ObjectNet3D. In Tab. 1, we also report 3D pose estimation results from two supervised + +Table 2. Evaluation of 3D object pose estimation on 20 categories of ObjectNet3D. We supplement the missing results in the appendix. We report the $30^{\circ}$ accuracy. Note, that ZSP [14] requires also depth input and runs an order of a magnitude slower. + +
201918171615AVG
UOP3D [52]26.066.579.142.371.763.673.0
Ours29.364.782.552.276.965.977.0
+ +Table 3. Evaluation of instance segmentation on 7 categories of PASCAL3D+. We report the IoU in percentage with the occlusion-free instance mask provided by the ground truth pose and mesh. + +
井井井井井井井
UOP3D [52]68.949.979.267.742.969.650.9
Ours69.454.182.868.153.366.853.6
井井井井井井井AVG
UOP3D [52]76.962.464.161.462.564.773.864.0
Ours83.170.662.771.765.366.778.367.0
+ +Table 4. Evaluation of instance segmentation on 20 categories of ObjectNet3D, we supplement the missing results in the appendix. We report the IoU in percentage with the occlusion-free instance mask given by rendering the annotated mesh in the annotated pose. + +methods [63, 83] that are trained from the PASCAL3D+ training set. Notably, all unsupervised methods are trained from CO3D data, as they require videos as input during training, and hence are evaluated in an out-of-domain setting. Our method improves over prior work on most of the categories except for categories with few shape variance across instances such as cellphone, microwave, and toaster. On average, the $30^{\circ}$ accuracy improves significantly over prior work from $69.2\%$ to $75.3\%$ by $+6.3$ points on PASCAL3D+ and from $52.4\%$ to $56.8\%$ by $+4.4$ points on ObjectNet3D. We attribute this performance increase to the correspondence learning with a deformable shape, which facilitates the training and, therefore, generalizes better. This is illustrated in the qualitative examples in Fig. 3 + +# 4.3. In-the-Wild Instance Segmentation + +We validate the deformation capabilities of Common3D using the instance segmentation masks of PASCAL3D+ in Tab. 3 and ObjectNet3D in Tab. 4. We compare our model with UOP3D [52], as it is the current state-of-the-art at self-supervised learning 3D object priors for common objects, and was also trained using CO3D videos. Common3D outperforms UOP3D on almost all categories and on average increases the IoU by 3.8 percentage points from $60.3\%$ to $64.1\%$ on PASCAL3D+ and by 3 percentage points, from $64\%$ to $67\%$ on ObjectNet3D. A qualitative comparison of our method against UOP3D is depicted in Fig. 3. It shows that Common3D achieves much more accurate highly accurate despite a large variability in the object instances. + +# 4.4. Semantic Correspondence + +We demonstrate the significant benefit of our semantic feature adapter by evaluating its performance on semantic cor + +![](images/e52efaa4cfbe7868ea5d56274a733ef4a8789e26c136429e4f8d75735f5575d5.jpg) +Figure 3. Qualitative results on the ObjectNet3D dataset. In the second row the results of our method are illustrated, in the third the results of UOP3D [52]. Notably, our method fits the object more accurately, resulting in improved 3D pose and segmentation accuracy. + +
6073.367.246.366.042.259.4
SpherMaps [35]61.273.351.536.261.024.247.9
DINOv2 [41]62.052.366.244.260.731.954.0
Ours60.360.5
+ +respondence estimation in Tab. 5. We evaluate our model compared to the self-supervised DINOv2 model as well as SpherMaps [35], which require coarse viewpoint annotations. We follow the evaluation protocol of SpherMaps and find correspondences by weighting the image features similarities from the DINOv2 backbone and the feature similarities from our adapter. We weight the backbone's features with 0.8 and ours with 0.2. On average, we can largely increase the PCK@0.1 compared to the self-supervised baseline DINOv2 by $6.1\%$ percentage points from $47.9\%$ to $54\%$ . For some categories, we can even achieve a competitive performance compared to the supervised baseline SpherMaps. We believe that one major challenge remains the domain gap between the training and test data. + +# 4.5. Ablation + +In Tab. 6, we ablate our model design choices regarding the adapter, the appearance loss as well as the instance deformation model. We observe, that instance-level deformation has only a marginal relevance for the pose estimation, but proves essential for segmentation and semantic correspondence estimation. Furthermore, the adapter and the dense appearance loss are significantly improving our method. + +Table 5. Evaluation of the PCK@0.1 on 6 categories of SPair-71k [37]. Note, that SpherMaps [35] uses coarse viewpoint annotations, while DINOv2 and our approach do not require any. + +
30°10°IoUPCK@0.1
w/o Adapt.25.96.660.532.4
w/o Adapt., w Rec. Loss13.92.156.232.4
w/o Dense App. Loss56.712.258.845.3
w/o Inst. Deform.75.820.258.645.3
Ours75.321.064.147.8
+ +Table 6. Ablation of our model design choices. We evaluate the 3D pose accuracy and instance segmentation on PASCAL3D+ using $30^{\circ}$ , $10^{\circ}$ accuracy and IoU. Further, the semantic correspondences are evaluated on SPair71-k using PCK@0.1. We observe that the adapter is crucial for our method, as well as the dense appearance loss. Further, our instance deformation model yields a minor disadvantage for the coarse $30^{\circ}$ pose estimation, while improving the fine-grained $10^{\circ}$ pose estimation, and proves essential for the instance segmentation and the semantic correspondences. Note that in this standard setting we use the ViT-S backbone, in contrast to the results reported in Tab. 5. + +# 5. Conclusion + +We have introduced Common3D, a new model that can learn 3D morphable models from a collection of casually captured videos completely self-supervised. This model can, at test time, estimate the 3D object shape, 2D-3D correspondences, and the 3D object pose. Common3D connects two different subfields of computer vision, self-supervised representation learning and the learning of 3D deformable object priors. We believe that in the long term, this may unify many vision tasks into a common inference process, while also creating a better 3D spatial understanding in vision models. + +Acknowledgment. This work was funded by the German Research Foundation (DFG) - 468670075, and 499552394 (SFB 1597 - Small Data). + +# References + +[1] Oswald Aldrian and William AP Smith. Inverse rendering of faces with a 3d morphable model. IEEE transactions on pattern analysis and machine intelligence, 35(5):1080-1093, 2012. 1 +[2] Dosovitskiy Alexey. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv: 2010.11929, 2020.7 +[3] Dragomir Anguelov, Praveen Srinivasan, Daphne Koller, Sebastian Thrun, Jim Rodgers, and James Davis. Scape: shape completion and animation of people. In ACM SIGGRAPH 2005 Papers, pages 408-416. 2005. 2 +[4] Yutong Bai, Angtian Wang, Adam Kortylewski, and Alan Yuille. Coke: Contrastive learning for robust keypoint detection. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 65-74, 2023. 5 +[5] V BLANZ and T VETTER. A morphable model for the synthesis of 3d faces. In Computer graphics proceedings, annual conference series, pages 187-194. Association for Computing Machinery SIGGRAPH, 1999. 1, 2 +[6] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 4 +[7] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 3 +[8] Zhiqin Chen and Hao Zhang. Learning implicit fields for generative shape modeling. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5939-5948, 2019. 2 +[9] Devikalyan Das, Christopher Wewer, Raza Yunus, Eddy Ilg, and Jan Eric Lenssen. Neural parametric gaussians for monocular non-rigid object reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10715-10725, 2024. 3 +[10] Yu Deng, Jiaolong Yang, and Xin Tong. Deformed implicit field: Modeling 3d shapes with learned dense correspondence. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10286-10296, 2021. 3 +[11] Bernhard Egger, Sandro Schonborn, Andreas Schneider, Adam Kortylewski, Andreas Morel-Forster, Clemens Blumer, and Thomas Vetter. Occlusion-aware 3d morphable models and an illumination prior for face image analysis. International Journal of Computer Vision, 126:1269-1287, 2018. 1 +[12] Shubham Goel, Angjoo Kanazawa, and Jitendra Malik. Shape and viewpoint without keypoints. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XV 16, pages 88-104. Springer, 2020. 3 +[13] Shubham Goel, Georgia Gkioxari, and Jitendra Malik. Differentiable stereopsis: Meshes from multiple views using + +differentiable rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8635-8644, 2022. 4 +[14] Walter Goodwin, Sagar Vaze, Ioannis Havoutis, and Ingmar Posner. Zero-shot category-level object pose estimation. In European Conference on Computer Vision, pages 516-532. Springer, 2022. 7 +[15] Amos Gropp, Lior Yariv, Niv Haim, Matan Atzmon, and Yaron Lipman. Implicit geometric regularization for learning shapes. arXiv preprint arXiv:2002.10099, 2020. 5 +[16] Thibault Groueix, Matthew Fisher, Vladimir G Kim, Bryan C Russell, and Mathieu Aubry. Atlasnet: a papiermâché approach to learning 3d surface generation (2018). arXiv preprint arXiv:1802.05384, 11, 1802. 3 +[17] Paul Henderson, Vagia Tsiminaki, and Christoph H Lampert. Leveraging 2d data to learn textured 3d mesh generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7498-7507, 2020. 3 +[18] Tao Hu, Liwei Wang, Xiaogang Xu, Shu Liu, and Jiaya Jia. Self-supervised 3d mesh reconstruction from single images. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6002-6011, 2021. 3 +[19] Wei-Chih Hung, Varun Jampani, Sifei Liu, Pavlo Molchanov, Ming-Hsuan Yang, and Jan Kautz. Scops: Self-supervised co-part segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 869-878, 2019. 3 +[20] Tomas Jakab, Ruining Li, Shangzhe Wu, Christian Rupprecht, and Andrea Vedaldi. Farm3d: Learning articulated 3d animals by distilling 2d diffusion. In 2024 International Conference on 3D Vision (3DV), pages 852-861. IEEE, 2024. 3 +[21] Hanbyul Joo, Tomas Simon, and Yaser Sheikh. Total capture: A 3d deformation model for tracking faces, hands, and bodies. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8320-8329, 2018. 2 +[22] Angjoo Kanazawa, Shubham Tulsiani, Alexei A Efros, and Jitendra Malik. Learning category-specific mesh reconstruction from image collections. In Proceedings of the European Conference on Computer Vision (ECCV), pages 371-386, 2018. 3 +[23] Angjoo Kanazawa, Shubham Tulsiani, Alexei A Efros, and Jitendra Malik. Learning category-specific mesh reconstruction from image collections. In Proceedings of the European Conference on Computer Vision (ECCV), pages 371-386, 2018. 2 +[24] Abhishek Kar, Shubham Tulsiani, Joao Carreira, and Jitendra Malik. Category-specific object reconstruction from a single image. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1966-1974, 2015. 3 +[25] Diederik P Kingma. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 7 +[26] Alexander Kirillov, Yuxin Wu, Kaiming He, and Ross Girshick. Pointrend: Image segmentation as rendering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9799-9808, 2020. 5, 6 + +[27] Filippos Kokkinos and Iasonas Kokkinos. To the point: Correspondence-driven monocular 3d category reconstruction. Advances in Neural Information Processing Systems, 34:7760-7772, 2021. 3 +[28] Nilesh Kulkarni, Abhinav Gupta, David F Fouhey, and Shubham Tulsiani. Articulation-aware canonical surface mapping. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 452-461, 2020. 3 +[29] Samuli Laine, Janne Hellsten, Tero Karras, Yeongho Seol, Jaakko Lehtinen, and Timo Aila. Modular primitives for high-performance differentiable rendering. ACM Transactions on Graphics, 39(6), 2020. 7 +[30] Aldo Laurentini. The visual hull concept for silhouette-based image understanding. IEEE Transactions on pattern analysis and machine intelligence, 16(2):150-162, 1994. 5 +[31] Xueting Li, Sifei Liu, Kihwan Kim, Shalini De Mello, Varun Jampani, Ming-Hsuan Yang, and Jan Kautz. Self-supervised single-view 3d reconstruction via semantic consistency. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIV 16, pages 677-693. Springer, 2020. 3 +[32] Zizhang Li, Dor Litvak, Ruining Li, Yunzhi Zhang, Tomas Jakab, Christian Rupprecht, Shangzhe Wu, Andrea Vedaldi, and Jiajun Wu. Learning the 3d fauna of the web. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9752-9762, 2024. 2, 3 +[33] Yebin Liu, Juergen Gall, Carsten Stoll, Qionghai Dai, Hans-Peter Seidel, and Christian Theobalt. Markerless motion capture of multiple characters using multiview image segmentation. IEEE transactions on pattern analysis and machine intelligence, 35(11):2720-2735, 2013. 2 +[34] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J Black. Smpl: a skinned multiperson linear model. ACM Transactions on Graphics (TOG), 34(6):1-16, 2015. 2 +[35] Octave Mariotti, Oisin Mac Aodha, and Hakan Bilen. Improving semantic correspondence with viewpoint-guided spherical maps. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19521-19530, 2024. 2, 8 +[36] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4460-4470, 2019. 2 +[37] Juhong Min, Jongmin Lee, Jean Ponce, and Minsu Cho. Spair-71k: A large-scale benchmark for semantic correspondence. arXiv preprint arXiv:1908.10543, 2019. 2, 6, 8 +[38] Tom Monnier, Matthew Fisher, Alexei A Efros, and Mathieu Aubry. Share with thy neighbors: Single-view reconstruction by cross-instance consistency. In European Conference on Computer Vision, pages 285-303. Springer, 2022. 3 +[39] Natalia Neverova, David Novotny, Marc Szafraniec, Vasil Khalidov, Patrick Labatut, and Andrea Vedaldi. Continuous surface embeddings. Advances in Neural Information Processing Systems, 33:17258-17270, 2020. 6 + +[40] Michael Oechsle, Lars Mescheder, Michael Niemeyer, Thilo Strauss, and Andreas Geiger. Texture fields: Learning texture representations in function space. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4531-4540, 2019. 5 +[41] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 2, 4, 7, 8 +[42] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 165-174, 2019. 2, 3 +[43] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5865-5874, 2021. 3 +[44] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems, 30, 2017. 6 +[45] Jeremy Reizenstein, Roman Shapovalov, Philipp Henzler, Luca Sbordone, Patrick Labatut, and David Novotny. Common objects in 3d: Large-scale learning and evaluation of real-life 3d category reconstruction. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10901-10911, 2021. 6 +[46] Lukasz Romaszko, Christopher KI Williams, Pol Moreno, and Pushmeet Kohli. Vision-as-inverse-graphics: Obtaining a rich 3d explanation of a scene from a single image. In Proceedings of the IEEE International Conference on Computer Vision Workshops, pages 851–859, 2017. 2, 3, 4 +[47] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2 +[48] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-Motion Revisited. In Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 5, 6 +[49] Sandro Schonborn, Bernhard Egger, Andreas Morel-Forster, and Thomas Vetter. Markov chain monte carlo for automated face image analysis. International Journal of Computer Vision, 123:160–183, 2017. 2, 3, 4 +[50] Tianchang Shen, Jun Gao, Kangxue Yin, Ming-Yu Liu, and Sanja Fidler. Deep marching tetrahedra: a hybrid representation for high-resolution 3d shape synthesis. Advances in Neural Information Processing Systems, 34:6087-6101, 2021. 2, 4 +[51] Aleksandar Shtedritski, Christian Rupprecht, and Andrea Vedaldi. Shic: Shape-image correspondences with no key-point supervision. arXiv preprint arXiv:2407.18907, 2024. 2 +[52] Leonhard Sommer, Artur Jesslen, Eddy Ilg, and Adam Kortylewski. Unsupervised learning of category-level 3d pose + +from object-centric videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22787-22796, 2024. 2, 5, 6, 7, 8 +[53] Keqiang Sun, Dor Litvak, Yunzhi Zhang, Hongsheng Li, Ji-jun Wu, and Shangzhe Wu. Ponomation: Learning articulated 3d animal motions from unlabeled online videos. In European Conference on Computer Vision, pages 100-119. Springer, 2025. 3 +[54] Shanlin Sun, Kun Han, Deying Kong, Hao Tang, Xiangyi Yan, and Xiaohui Xie. Topology-preserving shape reconstruction and registration via neural diffeomorphic flow. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20845-20855, 2022. 3 +[55] Ayush Tewari, Michael Zollhofer, Hyeongwoo Kim, Pablo Garrido, Florian Bernard, Patrick Perez, and Christian Theobalt. Mofa: Model-based deep convolutional face autoencoder for unsupervised monocular reconstruction. In Proceedings of the IEEE international conference on computer vision workshops, pages 1274-1283, 2017. 1, 2, 3, 4 +[56] Ayush Tewari, Florian Bernard, Pablo Garrido, Gaurav Bharaj, Mohamed Elgharib, Hans-Peter Seidel, Patrick Pérez, Michael Zollhofer, and Christian Theobalt. Fml: Face model learning from videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10812-10822, 2019. 3 +[57] Ayush Tewari, Hans-Peter Seidel, Mohamed Elgharib, Christian Theobalt, et al. Learning complete 3d morphable face models from images and videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3361-3371, 2021. 3 +[58] Ayush Tewari, Xingang Pan, Ohad Fried, Maneesh Agrawala, Christian Theobalt, et al. Disentangled3d: Learning a 3d generative model with disentangled geometry and appearance from monocular images. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1516-1525, 2022. 2, 3 +[59] Shubham Tulsiani, Abhishek Kar, Joao Carreira, and Jitendra Malik. Learning category-specific deformable 3d models for object reconstruction. IEEE transactions on pattern analysis and machine intelligence, 39(4):719-731, 2016. 3 +[60] Shubham Tulsiani, Nilesh Kulkarni, and Abhinav Gupta. Implicit mesh reconstruction from unannotated image collections. arXiv preprint arXiv:2007.08504, 2020. 3 +[61] Angtian Wang, Yihong Sun, Adam Kortylewski, and Alan L Yuille. Robust object detection under occlusion with context-aware compositionalnets. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12645-12654, 2020. 7 +[62] Angtian Wang, Adam Kortylewski, and Alan Yuille. Nemo: Neural mesh models of contrastive features for robust 3d pose estimation. arXiv preprint arXiv:2101.12378, 2021. 6 +[63] Angtian Wang, Peng Wang, Jian Sun, Adam Kortylewski, and Alan Yuille. Voge: a differentiable volume renderer using gaussian ellipsoids for analysis-by-synthesis. arXiv preprint arXiv:2205.15401, 2022. 6, 7 +[64] Haowen Wang, Zhipeng Fan, Zhen Zhao, Zhengping Che, Zhiyuan Xu, Dong Liu, Feifei Feng, Yakun Huang, Xiuquan + +Qiao, and Jian Tang. Dtf-net: Category-level pose estimation and shape reconstruction via deformable template field. In Proceedings of the 31st ACM International Conference on Multimedia, pages 3676-3685, 2023. 3 +[65] Nanyang Wang, Yinda Zhang, Zhuwen Li, Yanwei Fu, Wei Liu, and Yu-Gang Jiang. Pixel2mesh: Generating 3d mesh models from single rgb images. In Proceedings of the European conference on computer vision (ECCV), pages 52-67, 2018. 3 +[66] Shangzhe Wu, Christian Rupprecht, and Andrea Vedaldi. Unsupervised learning of probably symmetric deformable 3d objects from images in the wild. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1-10, 2020. 3 +[67] Shangzhe Wu, Christian Rupprecht, and Andrea Vedaldi. Unsupervised learning of probably symmetric deformable 3d objects from images in the wild. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1-10, 2020. 2 +[68] Shangzhe Wu, Tomas Jakab, Christian Rupprecht, and Andrea Vedaldi. Dove: Learning deformable 3d objects by watching videos. International Journal of Computer Vision, 131(10):2623-2634, 2023. 2, 3 +[69] Shangzhe Wu, Ruining Li, Tomas Jakab, Christian Rupprecht, and Andrea Vedaldi. Magicpony: Learning articulated 3d animals in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8792-8802, 2023. 2, 3, 4 +[70] Yu Xiang, Roozbeh Mottaghi, and Silvio Savarese. Beyond Pascal: A benchmark for 3d object detection in the wild. In IEEE winter conference on applications of computer vision, pages 75-82. IEEE, 2014. 6 +[71] Yu Xiang, Wonhui Kim, Wei Chen, Jingwei Ji, Christopher Choy, Hao Su, Roozbeh Mottaghi, Leonidas Guibas, and Silvio Savarese. Objectnet3d: A large scale database for 3d object recognition. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part VIII 14, pages 160-176. Springer, 2016. 6 +[72] Gengshan Yang, Deqing Sun, Varun Jampani, Daniel Vlasic, Forrester Cole, Huiwen Chang, Deva Ramanan, William T Freeman, and Ce Liu. Lasr: Learning articulated shape reconstruction from a monocular video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15980-15989, 2021. 4 +[73] Gengshan Yang, Deqing Sun, Varun Jampani, Daniel Vlasic, Forrester Cole, Ce Liu, and Deva Ramanan. Viser: Video-specific surface embeddings for articulated 3d shape reconstruction. Advances in Neural Information Processing Systems, 34:19326-19338, 2021. 3 +[74] Gengshan Yang, Minh Vo, Natalia Neverova, Deva Ramanan, Andrea Vedaldi, and Hanbyul Joo. Banmo: Building animatable 3d neural models from many casual videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2863-2873, 2022. 2, 3 +[75] Chun-Han Yao, Wei-Chih Hung, Yuanzhen Li, Michael Rubinstein, Ming-Hsuan Yang, and Varun Jampani. Lassie: + +Learning articulated shapes from sparse image ensemble via 3d part discovery. Advances in Neural Information Processing Systems, 35:15296-15308, 2022. 3 +[76] Chun-Han Yao, Wei-Chih Hung, Yuanzhen Li, Michael Rubinstein, Ming-Hsuan Yang, and Varun Jampani. Hi-lassie: High-fidelity articulated shape and skeleton discovery from sparse image ensemble. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4853-4862, 2023. 3 +[77] Alan Yuille and Daniel Kersten. Vision as bayesian inference: analysis by synthesis? Trends in cognitive sciences, 10(7):301-308, 2006. 1 +[78] Junyi Zhang, Charles Herrmann, Junhwa Hur, Eric Chen, Varun Jampani, Deqing Sun, and Ming-Hsuan Yang. Telling left from right: Identifying geometry-aware semantic correspondence. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3076-3085, 2024. 2, 4 +[79] Junyi Zhang, Charles Herrmann, Junhwa Hur, Luisa Polania Cabrera, Varun Jampani, Deqing Sun, and Ming-Hsuan Yang. A tale of two features: Stable diffusion complements dino for zero-shot semantic correspondence. Advances in Neural Information Processing Systems, 36, 2024. 2, 4 +[80] Kaifeng Zhang, Yang Fu, Shubhankar Borse, Hong Cai, Fatih Porikli, and Xiaolong Wang. Self-supervised geometric correspondence for category-level 6d object pose estimation in the wild. arXiv preprint arXiv:2210.07199, 2022. 3 +[81] Yunzhi Zhang, Shangzhe Wu, Noah Snavely, and Jiajun Wu. Seeing a rose in five thousand ways. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 962-971, 2023. 2 +[82] Zerong Zheng, Tao Yu, Qionghai Dai, and Yebin Liu. Deep implicit templates for 3d shape representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1429-1439, 2021. 3, 4, 5 +[83] Xingyi Zhou, Arjun Karpur, Linjie Luo, and Qixing Huang. Starmap for category-agnostic keypoint and viewpoint estimation. In Proceedings of the European Conference on Computer Vision (ECCV), pages 318–334, 2018. 7 +[84] Silvia Zuffi, Angjoo Kanazawa, David W Jacobs, and Michael J Black. 3d menagerie: Modeling the 3d shape and pose of animals. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6365-6373, 2017. 2 \ No newline at end of file diff --git a/2025/Common3D_ Self-Supervised Learning of 3D Morphable Models for Common Objects in Neural Feature Space/images.zip b/2025/Common3D_ Self-Supervised Learning of 3D Morphable Models for Common Objects in Neural Feature Space/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..eea1eac2bf7a82fc3665417da7e35d09312af5f8 --- /dev/null +++ b/2025/Common3D_ Self-Supervised Learning of 3D Morphable Models for Common Objects in Neural Feature Space/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bcd22618a0b7147629a304cf167c990507d96b3d71321b72a90ce3174ac76641 +size 485599 diff --git a/2025/Common3D_ Self-Supervised Learning of 3D Morphable Models for Common Objects in Neural Feature Space/layout.json b/2025/Common3D_ Self-Supervised Learning of 3D Morphable Models for Common Objects in Neural Feature Space/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..4589110e2c4f3705d0302764dbf88262dcb4308e --- /dev/null +++ b/2025/Common3D_ Self-Supervised Learning of 3D Morphable Models for Common Objects in Neural Feature Space/layout.json @@ -0,0 +1,9250 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 107, + 103, + 504, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 103, + 504, + 140 + ], + "spans": [ + { + "bbox": [ + 107, + 103, + 504, + 140 + ], + "type": "text", + "content": "Common3D: Self-Supervised Learning of 3D Morphable Models for Common Objects in Neural Feature Space" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 80, + 160, + 179, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 160, + 179, + 175 + ], + "spans": [ + { + "bbox": [ + 80, + 160, + 179, + 175 + ], + "type": "text", + "content": "Leonhard Sommer" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 204, + 160, + 274, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 204, + 160, + 274, + 175 + ], + "spans": [ + { + "bbox": [ + 204, + 160, + 274, + 175 + ], + "type": "text", + "content": "Olaf Dunkel" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 299, + 160, + 398, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 160, + 398, + 175 + ], + "spans": [ + { + "bbox": [ + 299, + 160, + 398, + 175 + ], + "type": "text", + "content": "Christian Theobalt 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 423, + 160, + 529, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 423, + 160, + 529, + 175 + ], + "spans": [ + { + "bbox": [ + 423, + 160, + 529, + 175 + ], + "type": "text", + "content": "Adam Kortylewski 1,2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 149, + 188, + 264, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 188, + 264, + 204 + ], + "spans": [ + { + "bbox": [ + 149, + 188, + 264, + 204 + ], + "type": "text", + "content": "1University of Freiburg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 276, + 188, + 457, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 276, + 188, + 457, + 203 + ], + "spans": [ + { + "bbox": [ + 276, + 188, + 457, + 203 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 276, + 188, + 457, + 203 + ], + "type": "text", + "content": "Max Planck Institute for Informatics" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 57, + 230, + 552, + 404 + ], + "blocks": [ + { + "bbox": [ + 57, + 230, + 552, + 404 + ], + "lines": [ + { + "bbox": [ + 57, + 230, + 552, + 404 + ], + "spans": [ + { + "bbox": [ + 57, + 230, + 552, + 404 + ], + "type": "image", + "image_path": "01a12ec6fb8d4c297aa800494cbbf4a7addfdc1f272cef0f4fcf3164415677bc.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 411, + 555, + 436 + ], + "lines": [ + { + "bbox": [ + 55, + 411, + 555, + 436 + ], + "spans": [ + { + "bbox": [ + 55, + 411, + 555, + 436 + ], + "type": "text", + "content": "Figure 1. Common3D learns category-specific 3D morphable models from few casually captured videos completely self-supervised, and can estimate the 3D object shape (visualized from two viewpoints), 2D-3D correspondences, and the 3D object pose via inverse rendering." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 151, + 444, + 200, + 456 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 444, + 200, + 456 + ], + "spans": [ + { + "bbox": [ + 151, + 444, + 200, + 456 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 54, + 474, + 297, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 474, + 297, + 715 + ], + "spans": [ + { + "bbox": [ + 54, + 474, + 297, + 715 + ], + "type": "text", + "content": "3D morphable models (3DMMs) are a powerful tool to represent the possible shapes and appearances of an object category. Given a single test image, 3DMMs can be used to solve various tasks, such as predicting the 3D shape, pose, semantic correspondence, and instance segmentation of an object. Unfortunately, 3DMMs are only available for very few object categories that are of particular interest, like faces or human bodies, as they require a demanding 3D data acquisition and category-specific training process. In contrast, we introduce a new method, Common3D, that learns 3DMMs of common objects in a fully self-supervised manner from a collection of object-centric videos. For this purpose, our model represents objects as a learned 3D template mesh and a deformation field that is parameterized as an image-conditioned neural network. Different from prior works, Common3D represents the object appearance with neural features instead of RGB colors, which enables the learning of more generalizable representations through an abstraction from pixel intensities. Importantly, we train the appearance features using a contrastive" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 445, + 556, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 445, + 556, + 542 + ], + "spans": [ + { + "bbox": [ + 313, + 445, + 556, + 542 + ], + "type": "text", + "content": "objective by exploiting the correspondences defined through the deformable template mesh. This leads to higher quality correspondence features compared to related works and a significantly improved model performance at estimating 3D object pose and semantic correspondence. Common3D is the first completely self-supervised method that can solve various vision tasks in a zero-shot manner. We release all code at github.com/GenIntel/common3d." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 314, + 560, + 395, + 573 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 560, + 395, + 573 + ], + "spans": [ + { + "bbox": [ + 314, + 560, + 395, + 573 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "text", + "content": "Estimating the 3D shape and pose of an object from a single image requires prior knowledge about the possible shapes and appearances of an object category. Learning such prior models is a long-standing problem in computer vision, with 3D morphable models (3DMMs) [5] emerging as a very successful approach to address this problem. A popular application of 3DMMs is to perform single image 3D reconstruction via inverse rendering [1, 5, 55], also referred to as analysis-by-synthesis approach [77]. This enables solving under-constrained vision tasks, for example predicting an object's shape behind an occluder [11] or estimating the" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6468" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "content": "3D object shape and pose from a single image. However, training 3DMMs requires sophisticated data acquisition setups [3, 21, 33], involving laser scanners or multi-view studio captures, and a significant manual effort in cleaning and registration of the data [5, 34, 84]. This approach is difficult to scale to more common everyday objects." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 146, + 294, + 242 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 146, + 294, + 242 + ], + "spans": [ + { + "bbox": [ + 54, + 146, + 294, + 242 + ], + "type": "text", + "content": "An alternative approach that emerged recently is to learn a 3D prior from 2D images only [58, 67, 74, 81]. However, learning such priors from in-the-wild images is a major challenge and typically requires significant category-specific design choices, with various solutions having been proposed for different object categories [23, 32, 68, 69, 81]. This makes it difficult to scale to the long tail of common objects that can appear in natural images." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 243, + 295, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 243, + 295, + 422 + ], + "spans": [ + { + "bbox": [ + 56, + 243, + 295, + 422 + ], + "type": "text", + "content": "We introduce Common3D, a new way of integrating self-supervised learning, neural rendering, and 3D modeling to learn 3D object priors for common objects from a collection of object-centric videos. Once trained, the model can estimate 2D-3D correspondences, it can be leveraged for inverse rendering to estimate the 3D pose, and instance segmentation of common objects from a single image, see Fig. 1. At the core of our method is a category-specific 3D morphable model (3DMM) that represents the possible variations of different object instances. The 3DMM is represented using a learned prototypical 3D shape and a shape deformation field that is parameterized as an image-conditioned neural network. Training our model requires canonical camera poses, which we estimate unsupervised by adopting the method of [52]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 425, + 295, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 425, + 295, + 568 + ], + "spans": [ + { + "bbox": [ + 56, + 425, + 295, + 568 + ], + "type": "text", + "content": "One core challenge for learning 3DMMs of common objects is to be flexible enough to capture the highly variable shapes and topologies of various object categories, as well as the possible variations of instances within one category. We address this issue by using a hybrid volumetric mesh representation based on DMTet [50]. Hence, the shape is defined volumetrically in canonical space, but a mesh is extracted on the fly for morphing and rendering the model. This allows to retain the flexibility of implicit neural representations for capturing highly variable shapes and topologies [8, 36, 42], while also enabling the use of meshes for an efficient rendering and regularization during training." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 570, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 295, + 713 + ], + "type": "text", + "content": "Another important challenge is to perform inverse rendering with 3D object priors, as it requires the optimization of a highly non-convex image-based reconstruction loss [46, 49, 55]. We address this problem in Common3D by representing the object appearance as a semantic feature field instead of RGB colors. Performing inverse rendering on the feature-level is easier because the features can directly encode 2D-3D correspondences. While related work [32, 69] directly uses DINO [41] as appearance features, recent studies showed that DINO has severe limitations at estimating semantic correspondence accurately [35, 78]. Therefore, we train an additional adapter module on top of" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "content": "DINO by using the 3D object prior to further enhance the feature extractor at accurately estimating 2D-3D correspondences. This leads to higher quality semantic correspondence features and a largely enhanced model performance." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 121, + 553, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 121, + 553, + 263 + ], + "spans": [ + { + "bbox": [ + 313, + 121, + 553, + 263 + ], + "type": "text", + "content": "To summarize, we make the following contributions: (1) We propose a new way of integrating self-supervised learning, neural rendering, and 3D modeling to enable the joint learning of 3D object priors for common objects from a collection of object-centric videos, (2) an effective mechanism for learning correspondence features by training a 3D object prior jointly with an additional adapter module on top of DINO, using a local contrastive objective that exploits the correspondences estimated by the 3D model, and (3) Common3D is the first completely self-supervised method that can estimate the shape, pose, and semantic correspondence of common objects from a single image." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 274, + 400, + 286 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 274, + 400, + 286 + ], + "spans": [ + { + "bbox": [ + 314, + 274, + 400, + 286 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 295, + 553, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 295, + 553, + 366 + ], + "spans": [ + { + "bbox": [ + 313, + 295, + 553, + 366 + ], + "type": "text", + "content": "Our model has two core components: An adapter module that enhances correspondence learning and a 3D morphable model that is learned without supervision. Therefore, we first provide an overview of correspondence learning methods and subsequently discuss related work on learning 3D morphable models from images and from synthetic data." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 367, + 553, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 367, + 553, + 701 + ], + "spans": [ + { + "bbox": [ + 313, + 367, + 553, + 701 + ], + "type": "text", + "content": "Self-Supervised Correspondence Learning. Finding semantic correspondences across instances of an object category is a challenging computer vision task [37]. Self-supervised models like DINOv2 [41] and Stable Diffusion [47] were shown to have remarkable capabilities at estimating zero-shot semantic correspondences [79]. Despite these advances, unsupervised models are still not performing well on all categories. This can be accounted to ambiguous features of similar parts [35] or the lower performance for viewpoint changes [78]. Previous works have addressed this by mapping the 2D features to a coarse spherical representation of the object [35] or by applying heuristic augmentation methods to account for the viewpoint dependency [78]. Recent related work [51] showed that by leveraging 3D shape representations, the 2D-3D correspondence mapping can be improved significantly. However, their approach requires a detailed categorical mesh and is limited to a small selected set of animals. Similarly, [52] uses multiple views from object-centric videos to account for the viewpoint dependency. Further, they obtain a simple rigid 3D template and learn 2D-3D correspondences to estimate the 3D pose of an object. In contrast, our proposed Common3D learns an underlying deformable 3D shape representation self-supervised, which enables us to generalize to various common objects. Moreover, our approach of learning correspondence features by training a 3D object prior jointly with an additional adapter module on top of DINO enables high quality features to emerge." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6469" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 70, + 342, + 245 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 342, + 245 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 342, + 245 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 342, + 245 + ], + "type": "image", + "image_path": "d191c29ecff0a63279bd5c07df407856295f3ae5528a9dbf43c7457267ae1d00.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 258, + 556, + 380 + ], + "lines": [ + { + "bbox": [ + 55, + 258, + 556, + 380 + ], + "spans": [ + { + "bbox": [ + 55, + 258, + 556, + 380 + ], + "type": "text", + "content": "Figure 2. Method Overview. At the core of our method is a category-level template with semantic features that is acquired using a neural SDF with Differentiable Marching Tetrahedra (DMTet) where features are attached using a feature field. The category-level template is morphed using an MLP that is conditioned on a latent code " + }, + { + "bbox": [ + 55, + 258, + 556, + 380 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 55, + 258, + 556, + 380 + ], + "type": "text", + "content": " and rotated using the pose " + }, + { + "bbox": [ + 55, + 258, + 556, + 380 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 55, + 258, + 556, + 380 + ], + "type": "text", + "content": ". The pose " + }, + { + "bbox": [ + 55, + 258, + 556, + 380 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 55, + 258, + 556, + 380 + ], + "type": "text", + "content": " is estimated unsupervised at training time and predicted at inference time. DINOv2 serves as a feature encoder and its output is processed in two branches: One for estimating the latent code " + }, + { + "bbox": [ + 55, + 258, + 556, + 380 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 55, + 258, + 556, + 380 + ], + "type": "text", + "content": " with a block of convolutional layers; and another branch serves as an adapter to enhance the DINO features for better correspondence learning. As a result, the features from the adapter are better suited for finding correspondences. For example, the tire features of DINO have a similar encoding and are, hence, ambiguous, whereas it is much more distinct after the adapter. We also visualize the training objectives: The appearance objective compares two surface probabilities: 1) The geometric probability " + }, + { + "bbox": [ + 55, + 258, + 556, + 380 + ], + "type": "inline_equation", + "content": "p(\\mathrm{v}_j|\\mathrm{v}_i,\\mathrm{V})" + }, + { + "bbox": [ + 55, + 258, + 556, + 380 + ], + "type": "text", + "content": ", as defined via the Euclidean distances of the mesh vertices; and 2) the appearance probability " + }, + { + "bbox": [ + 55, + 258, + 556, + 380 + ], + "type": "inline_equation", + "content": "\\bar{p} (\\mathrm{f}_j|s_i,\\mathrm{F},\\beta)" + }, + { + "bbox": [ + 55, + 258, + 556, + 380 + ], + "type": "text", + "content": ", acquired by comparing the 2D image features " + }, + { + "bbox": [ + 55, + 258, + 556, + 380 + ], + "type": "inline_equation", + "content": "s_i" + }, + { + "bbox": [ + 55, + 258, + 556, + 380 + ], + "type": "text", + "content": " with the surface features " + }, + { + "bbox": [ + 55, + 258, + 556, + 380 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 55, + 258, + 556, + 380 + ], + "type": "text", + "content": ". The remaining geometric objectives enforce the 3D shape and the projected mask to be consistent with the data acquired from a SFM pipeline." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 348, + 70, + 553, + 245 + ], + "blocks": [ + { + "bbox": [ + 348, + 70, + 553, + 245 + ], + "lines": [ + { + "bbox": [ + 348, + 70, + 553, + 245 + ], + "spans": [ + { + "bbox": [ + 348, + 70, + 553, + 245 + ], + "type": "image", + "image_path": "85d242279d4038a3267766926e54de4c6624ef9af42258a3d3421ce2d55c3c0b.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 400, + 296, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 400, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 400, + 296, + 712 + ], + "type": "text", + "content": "Learning 3D Morphable Models from Images. There was remarkable progress made in the self-supervised learning of category-specific 3D morphable models from images. Originally, approaches used camera viewpoint, 2D keypoints, and masks to predict the 3D shape [24, 59] and texture [22, 64]. Several works achieved to mitigate the requirement of camera viewpoint and 2D keypoint annotations by exploiting symmetries [12, 18, 31, 60], by adding multi-view viewpoint hypothesis [12], by requiring consistent self-supervised co-part segmentation [31] using [19], by cycle-consistent surface maps [27, 28, 60], or by temporal consistency for video clips [56, 57, 68]. Few works even achieve to not require mask annotations, [38, 58, 66], by being constrained to front views of faces [66], or coarse-to-fine learning of texture and shape [38]. However, most of these works focus on modeling the object in very fine detail, such that they can generate object instances down to the RGB pixel-level, making the inverse rendering process inherently difficult [46, 49, 55]. In contrast, our focus lies on self-supervised learning of a 3D morphable model that is particularly well suited for solving vision tasks through inverse rendering. Therefore, we represent the object appearance using neural features instead of RGB colors, which facilitates the inverse rendering process as the features learn to support the inverse rendering process by encoding 2D-3D correspondences." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 400, + 555, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 400, + 555, + 519 + ], + "spans": [ + { + "bbox": [ + 313, + 400, + 555, + 519 + ], + "type": "text", + "content": "Another line of work learns 3D object priors without viewpoint annotations by replicating DINO features [20, 32, 53, 69, 75, 76]. In contrast to us, these works are restricted to animal categories. Further, they are not enhancing the image feature encoder for learning correspondences. The work of Zhang et al. is related to ours [80], as it also enhances the image feature encoder, but not in an inverse rendering setting. Further, they rely on a detailed template prior shape and focus on categories with few deformations and rotational symmetries, such as a bottle." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 521, + 555, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 555, + 568 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 555, + 568 + ], + "type": "text", + "content": "There exist also multiple works that focus on videospecific deformation model [9, 43, 73, 74]. In contrast to us, they have to re-train for each new instance and cannot find correspondences across instances." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "text", + "content": "3D Morphable Models from Synthetic Data. Various work use synthetic data [7] to learn 3D representations. With dense annotation, in terms of a complete surface and annotated canonical frames, several works learn impressive 3D autoencoders [16], [42], and image-conditioned 3D reconstruction [65], [17]. Hereby, implicit 3D representations have outperformed explicit ones and were extended to learn category-specific 3D morphable models [10], [82], and [54]. In contrast, our model does not require synthetic data of several thousands instances per category, but only requires few casually captured object-centric videos. This provides novel 3D morphable models for various categories." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6470" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 71, + 291, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 71, + 291, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 291, + 85 + ], + "type": "text", + "content": "3. A 3D Morphable Model of Neural Features" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 319 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 319 + ], + "type": "text", + "content": "In this section we introduce Common3D, a method for self-supervised learning of 3D morphable models (3DMM) for common objects from only object-centric videos. One of the core challenges that we need to resolve is to find an approach that can deal with highly variable shapes and topologies within the same object category, while at the same time being general enough such that it works for as many categories as possible, hence requiring minimal category-specific heuristics. One key design choice is to represent the morphable shape model with a hybrid volumetric-mesh representation that is flexible, but also enables using mesh-based regularizers during training (Sec. 3.1). Our second key design choice is to represent the object appearance in terms of neural features and not RGB colors (Sec. 3.2), which enables the feature extractor to learn robust 2D-3D correspondences and, hence, facilitates the inverse rendering process. We discuss our training strategy and objectives in Sec. 3.3, and elaborate the inference mechanism in Sec. 3.4. Fig. 2 provides an overview of our method." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 330, + 236, + 342 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 330, + 236, + 342 + ], + "spans": [ + { + "bbox": [ + 55, + 330, + 236, + 342 + ], + "type": "text", + "content": "3.1. Morphable Shape Representation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 349, + 296, + 385 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 349, + 296, + 385 + ], + "spans": [ + { + "bbox": [ + 55, + 349, + 296, + 385 + ], + "type": "text", + "content": "We represent 3D object priors with a category-level template that has neural features attached to its surface, together with instance-specific geometric deformations." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 387, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 387, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 387, + 296, + 714 + ], + "type": "text", + "content": "Category-level Template. One core challenge in our work is to develop a flexible approach for learning the category-level template shape of objects, which generalizes across the highly variable shapes and topologies of common objects. While purely mesh-based representations can have advantages, they are difficult to optimize from scratch, leading to problems that often require ad-hoc heuristics such as re-meshing [13, 72]. To enable a more flexible way of optimizing the category-level 3D shape, we use a hybrid volumetric-mesh representation. This combines the advantages of an implicit and explicit 3D representation. In particular, the object shape is represented with an implicit model that is given as a signed distance field " + }, + { + "bbox": [ + 55, + 387, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\phi_{\\mathrm{sdf}}" + }, + { + "bbox": [ + 55, + 387, + 296, + 714 + ], + "type": "text", + "content": " which can adapt to complex geometries flexibly. It can be converted on the fly into an explicit mesh via the Differentiable Marching Tetrahedra (DMTet) algorithm [50]. Specifically, the signed distance field (SDF) is evaluated at the vertices of a tetrahedral grid and converted into a watertight mesh in a fast and differentiable manner. We denote the triangle mesh as " + }, + { + "bbox": [ + 55, + 387, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{M} = \\{\\mathrm{V},\\mathrm{E},\\mathrm{F}\\}" + }, + { + "bbox": [ + 55, + 387, + 296, + 714 + ], + "type": "text", + "content": " with the vertices " + }, + { + "bbox": [ + 55, + 387, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{V} = \\{\\mathrm{v}_i\\in \\mathbb{R}^3\\}_{i = 1}^{|\\mathrm{V}|}" + }, + { + "bbox": [ + 55, + 387, + 296, + 714 + ], + "type": "text", + "content": ", the edges " + }, + { + "bbox": [ + 55, + 387, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{E} = \\{(\\mathrm{v}_i,\\mathrm{v}_j)_e\\in \\mathbb{R}^3\\}_{e = 1}^{|\\mathrm{E}|}" + }, + { + "bbox": [ + 55, + 387, + 296, + 714 + ], + "type": "text", + "content": " and the features " + }, + { + "bbox": [ + 55, + 387, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{F} = \\{\\mathrm{f}_i\\in \\mathbb{R}^D\\}_{i = 1}^{|F|}" + }, + { + "bbox": [ + 55, + 387, + 296, + 714 + ], + "type": "text", + "content": ". Once converted to a mesh, the 3D object prior can be rendered into a 2D feature map by applying standard rasterization. The conversion of the SDF into a mesh enables us to use mesh-based regularizers when learning the deformable model, for example to enforce rigidity. Moreover, the vertices provide us" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 554, + 119 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 119 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 119 + ], + "type": "text", + "content": "direct access to corresponding points among the deformed templates, which will enable us to learn features that represent 2D-3D correspondences effectively during training (see Sec. 3.2 and Sec. 3.3 for details)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 121, + 555, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 121, + 555, + 228 + ], + "spans": [ + { + "bbox": [ + 313, + 121, + 555, + 228 + ], + "type": "text", + "content": "Instance-Level Deformation. We adopt an affine transformation field [82] to model the deformations of the category-level shape to different object instances. In contrast to [82] we transform the vertices of the category-level mesh instead of the signed distance field. This is more efficient, as re-extracting the mesh for each instance is not required. The function " + }, + { + "bbox": [ + 313, + 121, + 555, + 228 + ], + "type": "inline_equation", + "content": "\\phi_{a}(\\mathrm{v},\\mathrm{l}):\\mathbb{R}^{3}\\times \\mathrm{L}\\mapsto \\mathbb{R}^{3}" + }, + { + "bbox": [ + 313, + 121, + 555, + 228 + ], + "type": "text", + "content": " maps each vertex " + }, + { + "bbox": [ + 313, + 121, + 555, + 228 + ], + "type": "inline_equation", + "content": "\\mathbf{v}" + }, + { + "bbox": [ + 313, + 121, + 555, + 228 + ], + "type": "text", + "content": " from the category-level canonical space together with the image-specific latent code " + }, + { + "bbox": [ + 313, + 121, + 555, + 228 + ], + "type": "inline_equation", + "content": "\\mathbf{l}" + }, + { + "bbox": [ + 313, + 121, + 555, + 228 + ], + "type": "text", + "content": " to a new position" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 341, + 235, + 555, + 248 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 341, + 235, + 555, + 248 + ], + "spans": [ + { + "bbox": [ + 341, + 235, + 555, + 248 + ], + "type": "interline_equation", + "content": "\\phi_ {a} (\\mathrm {v}, 1) = \\alpha (\\mathrm {v}, 1) \\odot \\mathrm {v} + \\delta (\\mathrm {v}, 1), \\quad \\alpha , \\delta \\in \\mathbb {R} ^ {3}, \\tag {1}", + "image_path": "d470082df9e268076b1390ceb49e2fb5f8c5e669f2da63f4fb06d7303b85f0f2.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 255, + 554, + 355 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 255, + 554, + 355 + ], + "spans": [ + { + "bbox": [ + 313, + 255, + 554, + 355 + ], + "type": "text", + "content": "with the intermediate variables for scale " + }, + { + "bbox": [ + 313, + 255, + 554, + 355 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 255, + 554, + 355 + ], + "type": "text", + "content": " and translation " + }, + { + "bbox": [ + 313, + 255, + 554, + 355 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 313, + 255, + 554, + 355 + ], + "type": "text", + "content": ". Hereby, " + }, + { + "bbox": [ + 313, + 255, + 554, + 355 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 255, + 554, + 355 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 255, + 554, + 355 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 313, + 255, + 554, + 355 + ], + "type": "text", + "content": " are the outputs of an MLP conditioned on the vertex position " + }, + { + "bbox": [ + 313, + 255, + 554, + 355 + ], + "type": "inline_equation", + "content": "\\mathbf{v}" + }, + { + "bbox": [ + 313, + 255, + 554, + 355 + ], + "type": "text", + "content": " and the latent code " + }, + { + "bbox": [ + 313, + 255, + 554, + 355 + ], + "type": "inline_equation", + "content": "\\mathbf{l}" + }, + { + "bbox": [ + 313, + 255, + 554, + 355 + ], + "type": "text", + "content": ". We compute the image-specific latent code " + }, + { + "bbox": [ + 313, + 255, + 554, + 355 + ], + "type": "inline_equation", + "content": "\\mathbf{l}" + }, + { + "bbox": [ + 313, + 255, + 554, + 355 + ], + "type": "text", + "content": " with an image feature encoder " + }, + { + "bbox": [ + 313, + 255, + 554, + 355 + ], + "type": "inline_equation", + "content": "\\psi_{1}" + }, + { + "bbox": [ + 313, + 255, + 554, + 355 + ], + "type": "text", + "content": " that consists of a DINOv2 backbone and additional convolutional layers (see Fig. 2). We denote the deformed mesh " + }, + { + "bbox": [ + 313, + 255, + 554, + 355 + ], + "type": "inline_equation", + "content": "M_{def}(\\mathrm{I}) = \\{\\mathrm{V}_{def}(\\mathrm{I}),\\mathrm{E},\\mathcal{F},\\mathrm{F}\\}" + }, + { + "bbox": [ + 313, + 255, + 554, + 355 + ], + "type": "text", + "content": " with the deformed vertices " + }, + { + "bbox": [ + 313, + 255, + 554, + 355 + ], + "type": "inline_equation", + "content": "V_{def}(\\mathrm{I}) = \\{\\phi_a(\\mathrm{v}_i,\\psi_1(\\mathrm{I}))\\in \\mathbb{R}^3\\}_{i = 1}^{|\\mathrm{V}|}" + }, + { + "bbox": [ + 313, + 255, + 554, + 355 + ], + "type": "text", + "content": " and the image I." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 360, + 517, + 373 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 360, + 517, + 373 + ], + "spans": [ + { + "bbox": [ + 313, + 360, + 517, + 373 + ], + "type": "text", + "content": "3.2. Appearance as Semantic Feature Field" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 378, + 555, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 378, + 555, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 378, + 555, + 689 + ], + "type": "text", + "content": "Prior works in the context of 3DMMs traditionally represent the object appearance using RGB colors, which makes it inherently difficult to apply them for solving vision tasks through inverse rendering [46, 49, 55]. A core contribution of our work is to extend 3DMMs to facilitate the inverse rendering process. For this purpose, we represent the appearance using neural features instead of RGB colors. The key advantage of this approach is that the neural features can be invariant to image-specific appearance details and instead focus on encoding semantic correspondences that facilitate the inverse rendering process. Our approach is inspired from prior works [69] that learn deformable 3D models through inverse rendering on the feature-level. These works use DINO [41] features that allow estimating semantic correspondences reasonably well [6]. However, recent studies have shown that DINO features also have limitations, e.g., for differentiating similar object parts, or [78] or in terms of dense prediction [79]. Therefore, we choose to learn an adapter on top of the DINO backbone to further improve the image features for better capturing semantic correspondences. Specifically, we compute a feature map of semantic features as " + }, + { + "bbox": [ + 313, + 378, + 555, + 689 + ], + "type": "inline_equation", + "content": "S = \\Phi_w(\\tau(I)) \\in \\mathbb{R}^{D \\times H \\times W}" + }, + { + "bbox": [ + 313, + 378, + 555, + 689 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 378, + 555, + 689 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 378, + 555, + 689 + ], + "type": "text", + "content": " is a frozen DINOv2 [41] backbone and " + }, + { + "bbox": [ + 313, + 378, + 555, + 689 + ], + "type": "inline_equation", + "content": "\\Phi_w" + }, + { + "bbox": [ + 313, + 378, + 555, + 689 + ], + "type": "text", + "content": " is the adapter with " + }, + { + "bbox": [ + 313, + 378, + 555, + 689 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 313, + 378, + 555, + 689 + ], + "type": "text", + "content": " being the parameters of several convolutional layers (see Sec. 4.1 for the detailed architecture). The adapter is trained jointly with the shape model as described in Sec. 3.3." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "content": "On the 3DMM, we represent the object appearance as a set of feature vectors at every mesh vertex " + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{F} = \\{\\mathbf{f}_i \\in" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "6471" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 70, + 296, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 70, + 296, + 157 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 296, + 157 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^D\\}_{i=1}^{|V|}" + }, + { + "bbox": [ + 55, + 70, + 296, + 157 + ], + "type": "text", + "content": ". The features are generated by a feature field [40] " + }, + { + "bbox": [ + 55, + 70, + 296, + 157 + ], + "type": "inline_equation", + "content": "\\mathrm{f}_i = \\phi_{\\text{feat}}(\\mathrm{v}_i) : \\mathbb{R}^3 \\mapsto \\mathbb{R}^D" + }, + { + "bbox": [ + 55, + 70, + 296, + 157 + ], + "type": "text", + "content": " that maps each vertex " + }, + { + "bbox": [ + 55, + 70, + 296, + 157 + ], + "type": "inline_equation", + "content": "\\mathrm{v}_i" + }, + { + "bbox": [ + 55, + 70, + 296, + 157 + ], + "type": "text", + "content": " to a feature vector " + }, + { + "bbox": [ + 55, + 70, + 296, + 157 + ], + "type": "inline_equation", + "content": "\\mathrm{f}_i" + }, + { + "bbox": [ + 55, + 70, + 296, + 157 + ], + "type": "text", + "content": ". Importantly, as the vertices of the deformable mesh encode corresponding points in different object instances, the adapter can learn more accurate 2D-3D correspondences compared to self-supervised methods that learn purely in 2D, see Fig. 2." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 163, + 171, + 176 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 163, + 171, + 176 + ], + "spans": [ + { + "bbox": [ + 55, + 163, + 171, + 176 + ], + "type": "text", + "content": "3.3. Training Objectives" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 182, + 296, + 349 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 182, + 296, + 349 + ], + "spans": [ + { + "bbox": [ + 55, + 182, + 296, + 349 + ], + "type": "text", + "content": "We train Common3D from a collection of videos of objects. Our training process jointly optimizes the category-level template shape, the deformation model, the image feature encoder adapter, and the semantic feature field. Initially, we obtain a 3D point cloud from every video using Structure-from-Motion [48]. We need to align the individual 3D reconstructions of object instances into a canonical coordinate frame to obtain a camera pose annotation " + }, + { + "bbox": [ + 55, + 182, + 296, + 349 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 55, + 182, + 296, + 349 + ], + "type": "text", + "content": " relative to the object that is coherent across videos. We achieve this by adopting the method proposed in [52] to get initial camera pose annotations in an unsupervised manner. In the following, we discuss the geometric objectives, the appearance objective, and finally, the combined objective used for training the model." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 350, + 296, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 350, + 296, + 420 + ], + "spans": [ + { + "bbox": [ + 55, + 350, + 296, + 420 + ], + "type": "text", + "content": "Geometric Objectives. For optimizing the geometry, we formulate 2D reconstruction objectives using the mask and a 3D reconstruction objective using the 3D point cloud from each instance. Simultaneously, we regularize the category-level template's SDF and introduce mesh-based regularizers for highly rigid instance-level deformations." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 421, + 296, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 421, + 296, + 469 + ], + "spans": [ + { + "bbox": [ + 55, + 421, + 296, + 469 + ], + "type": "text", + "content": "First, we penalize the pixel-wise deviation " + }, + { + "bbox": [ + 55, + 421, + 296, + 469 + ], + "type": "inline_equation", + "content": "(u,v)" + }, + { + "bbox": [ + 55, + 421, + 296, + 469 + ], + "type": "text", + "content": " of our rendered mask " + }, + { + "bbox": [ + 55, + 421, + 296, + 469 + ], + "type": "inline_equation", + "content": "\\bar{\\mathfrak{m}} (\\mathrm{M}_{def},\\mathrm{I},\\pi ,u,v)" + }, + { + "bbox": [ + 55, + 421, + 296, + 469 + ], + "type": "text", + "content": " with the pose " + }, + { + "bbox": [ + 55, + 421, + 296, + 469 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 55, + 421, + 296, + 469 + ], + "type": "text", + "content": " from the object-centric mask m, provided by PointRend [26]. For this, we compute the mean squared error" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 478, + 294, + 506 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 106, + 478, + 294, + 506 + ], + "spans": [ + { + "bbox": [ + 106, + 478, + 294, + 506 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {\\mathrm {m}} \\left(\\mathrm {M} _ {d e f}, \\mathrm {I}, \\pi , \\mathrm {m}, u, v\\right) = \\\\ \\left| \\left| \\bar {\\mathrm {m}} \\left(\\mathrm {M} _ {d e f}, \\mathrm {I}, \\pi , u, v\\right) - \\mathrm {m} (u, v) \\right| \\right| ^ {2}. \\tag {2} \\\\ \\end{array}", + "image_path": "6f960b333872a391d1a8ab1b81458475604523891d9873aebfd40f4e556524b6.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 514, + 295, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 514, + 295, + 538 + ], + "spans": [ + { + "bbox": [ + 55, + 514, + 295, + 538 + ], + "type": "text", + "content": "We also maximize the pixel-wise overlap with the mask's distance transform " + }, + { + "bbox": [ + 55, + 514, + 295, + 538 + ], + "type": "inline_equation", + "content": "\\mathrm{m_{dt}}(u,v)" + }, + { + "bbox": [ + 55, + 514, + 295, + 538 + ], + "type": "text", + "content": ", as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 108, + 546, + 294, + 574 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 546, + 294, + 574 + ], + "spans": [ + { + "bbox": [ + 108, + 546, + 294, + 574 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {\\mathrm {m d t}} \\left(\\mathrm {M} _ {\\text {d e f}}, \\mathrm {I}, \\pi , \\mathrm {m} _ {\\mathrm {d t}}, u, v\\right) = \\tag {3} \\\\ - \\bar {\\mathrm {m}} (\\mathrm {M} _ {d e f}, \\mathrm {I}, \\pi , u, v) \\mathrm {m} _ {\\mathrm {d t}} (u, v). \\\\ \\end{array}", + "image_path": "acead5238b4a6d8c02abc2e579e65eca2f49887204890f630f0132b7beb82a98.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 582, + 295, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 295, + 677 + ], + "type": "text", + "content": "Hereby, the distance transform equals the distance to the silhouette for each pixel inside the mask, all pixels outside the mask are equal to zero. Note that this objective may only increase the object's volume, leading to a category-level template that encapsulates all objects like a visual hull [30]. Without this objective, and given the conflicting objectives of fitting a single category-level template to multiple different instances, the SDF might result in disconnected parts." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "To ensure that the deformed mesh reconstructs the shape of the individual object instances, we formulate a 3D chamfer distance based on the canonicalized 3D point cloud" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 72, + 501, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 501, + 84 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 501, + 84 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{inst}" + }, + { + "bbox": [ + 314, + 72, + 501, + 84 + ], + "type": "text", + "content": ", obtained from our object-centric video," + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 319, + 94, + 553, + 168 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 94, + 553, + 168 + ], + "spans": [ + { + "bbox": [ + 319, + 94, + 553, + 168 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {C D} \\left(\\mathrm {M} _ {d e f}, \\mathrm {I}, \\mathcal {P} _ {i n s t}\\right) = \\frac {1}{| \\mathrm {V} _ {d e f} | + | \\mathcal {P} _ {i n s t} |} \\\\ \\left(\\sum_ {\\mathrm {v} _ {i} \\in \\mathrm {V} _ {d e f}} | | \\mathrm {v} _ {i} - x _ {\\chi (\\mathrm {v} _ {i})} | | _ {2} + \\sum_ {x _ {i} \\in \\mathcal {P} _ {i n s t}} | | x _ {i} - \\mathrm {v} _ {\\chi (x _ {i})} | | _ {2}\\right), \\tag {4} \\\\ \\end{array}", + "image_path": "cc786ff4b35425355630e7c32de841277c583f788603602ca9ae09778421be00.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 169, + 553, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 169, + 553, + 194 + ], + "spans": [ + { + "bbox": [ + 313, + 169, + 553, + 194 + ], + "type": "text", + "content": "with " + }, + { + "bbox": [ + 313, + 169, + 553, + 194 + ], + "type": "inline_equation", + "content": "\\chi : \\mathbb{R}^3 \\mapsto \\mathbb{R}^3" + }, + { + "bbox": [ + 313, + 169, + 553, + 194 + ], + "type": "text", + "content": " returning the nearest neighbor of a 3D point to either the vertices " + }, + { + "bbox": [ + 313, + 169, + 553, + 194 + ], + "type": "inline_equation", + "content": "V_{\\text{def}}" + }, + { + "bbox": [ + 313, + 169, + 553, + 194 + ], + "type": "text", + "content": " or the point cloud " + }, + { + "bbox": [ + 313, + 169, + 553, + 194 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{\\text{inst}}" + }, + { + "bbox": [ + 313, + 169, + 553, + 194 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 194, + 555, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 194, + 555, + 314 + ], + "spans": [ + { + "bbox": [ + 313, + 194, + 555, + 314 + ], + "type": "text", + "content": "Further, we optimize the category-level template using the Eikonal regularization as proposed in [15] to ensure the satisfaction of the Eikonal property of a signed distance function, which states that the norm of the gradient equals to one. This is required as DMTet only provides gradients via the extracted mesh close to the surface, which leaves the points more far away from the surface undefined. Therefore, we sample a set of points " + }, + { + "bbox": [ + 313, + 194, + 555, + 314 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{sdf}" + }, + { + "bbox": [ + 313, + 194, + 555, + 314 + ], + "type": "text", + "content": " uniformly in the canonical space and around the extracted vertices, and average over each point " + }, + { + "bbox": [ + 313, + 194, + 555, + 314 + ], + "type": "inline_equation", + "content": "x\\in \\mathcal{P}_{sdf}" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 359, + 323, + 553, + 338 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 323, + 553, + 338 + ], + "spans": [ + { + "bbox": [ + 359, + 323, + 553, + 338 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {s d f} (\\mathrm {M}, x) = \\left(\\left| \\right| \\nabla \\phi_ {s d f} (x) \\left| \\right| _ {2} - 1\\right) ^ {2}. \\tag {5}", + "image_path": "b9995dc312483700b993654020fdcec561fe1a2b7bc9fcf973330f4f664a7012.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 348, + 553, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 348, + 553, + 372 + ], + "spans": [ + { + "bbox": [ + 313, + 348, + 553, + 372 + ], + "type": "text", + "content": "We also penalize large deformations from the category-level template using the mean squared error" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 324, + 381, + 553, + 410 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 381, + 553, + 410 + ], + "spans": [ + { + "bbox": [ + 324, + 381, + 553, + 410 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {d e f} (\\mathrm {M}, \\mathrm {M} _ {d e f}, \\mathrm {I}) = \\frac {1}{| \\mathrm {V} |} \\sum_ {\\mathrm {v} \\in \\mathrm {V}} \\left\\| \\mathrm {v} - \\phi_ {a} (\\mathrm {v}, \\psi_ {1} (\\mathrm {I})) \\right\\| ^ {2}. \\tag {6}", + "image_path": "9545ec802b1adbb4c4ec87ea724037ace1dce7cf2ce94d6512eaaa7981324694.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 420, + 553, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 420, + 553, + 455 + ], + "spans": [ + { + "bbox": [ + 313, + 420, + 553, + 455 + ], + "type": "text", + "content": "To ensure that deformations of connected vertices are smooth, we adopt the point-pair regularization from [82] to only consider connected vertices" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 318, + 465, + 553, + 533 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 465, + 553, + 533 + ], + "spans": [ + { + "bbox": [ + 318, + 465, + 553, + 533 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {d e f - s m} (\\mathrm {M}, \\mathrm {M} _ {d e f}, \\mathrm {I}) = \\frac {1}{| \\mathrm {E} |} \\\\ \\sum_ {\\mathrm {v} _ {i}, \\mathrm {v} _ {j} \\in \\mathrm {E}} \\frac {\\left| \\left| \\left[ \\mathrm {v} _ {i} - \\phi_ {a} \\left(\\mathrm {v} _ {i} , \\psi_ {1} (\\mathrm {I})\\right) \\right] - \\left[ \\mathrm {v} _ {j} - \\phi_ {a} \\left(\\mathrm {v} _ {j} , \\psi_ {1} (\\mathrm {I})\\right) \\right] \\right| \\right| _ {2}}{\\left| \\left| \\mathrm {v} _ {i} - \\mathrm {v} _ {j} \\right| \\right| _ {2}}. \\tag {7} \\\\ \\end{array}", + "image_path": "feb290952dd3ef9c538ca52ecb0833dae452dcf5c277a5c389872a683b5c23af.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "text", + "content": "Appearance Objective. In addition to the morphable shape representation, we also train the parameters of the semantic feature field " + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\phi_{\\text{feat}}(\\mathbf{v}_i)" + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "text", + "content": " and the parameters of the adapter module " + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\Phi_w" + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "text", + "content": " of the image feature encoder. In order to learn the semantic feature field and the feature encoder simultaneously, we cannot directly use a reconstruction objective, as this optimization collapses to the trivial solution of a constant feature for the whole image and the whole surface. We prevent a collapse to the trivial solution by instead optimizing an objective that encourages unique 2D-3D correspondences to emerge, which is inspired from related work on contrastive learning for keypoint detection [4]. In particular, we model for each image feature " + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "inline_equation", + "content": "s_i" + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "text", + "content": " the 3D correspondence as surface probability " + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\bar{\\mathfrak{p}}" + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "text", + "content": ", which is computed by the similarity to all vertex features " + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "text", + "content": ". Moreover," + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6472" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "type": "text", + "content": "we add a background feature " + }, + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "type": "text", + "content": ", which enables our model to differentiate pixelwise between surface and background" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 65, + 102, + 294, + 138 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 102, + 294, + 138 + ], + "spans": [ + { + "bbox": [ + 65, + 102, + 294, + 138 + ], + "type": "interline_equation", + "content": "\\bar {p} \\left(f _ {j} \\mid s _ {i}, F, \\beta\\right) = \\frac {\\exp \\left(\\kappa \\left\\langle s _ {i} , f _ {j} \\right\\rangle\\right)}{\\sum_ {k} \\exp \\left(\\kappa \\left\\langle s _ {i} , f _ {k} \\right\\rangle\\right) + \\exp \\left(\\kappa \\left\\langle s _ {i} , \\beta \\right\\rangle\\right)}, \\tag {8}", + "image_path": "519dd784ec41fbed77a75a1e421bf6f44ecca45d3538234549d579f240646b03.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 140, + 296, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 140, + 296, + 247 + ], + "spans": [ + { + "bbox": [ + 55, + 140, + 296, + 247 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 140, + 296, + 247 + ], + "type": "inline_equation", + "content": "\\kappa" + }, + { + "bbox": [ + 55, + 140, + 296, + 247 + ], + "type": "text", + "content": " is the temperature parameter. Similarly, the background probability " + }, + { + "bbox": [ + 55, + 140, + 296, + 247 + ], + "type": "inline_equation", + "content": "\\bar{\\mathfrak{p}} (\\beta |s_i,\\mathrm{F},\\beta)" + }, + { + "bbox": [ + 55, + 140, + 296, + 247 + ], + "type": "text", + "content": " is defined for each image feature. In contrast to previous work, we use farthest point sampling as in [44] to retrieve a subset of vertices and features. To mitigate additional notation, we neglect this sampling in the following. As ground truth, we model the surface probability " + }, + { + "bbox": [ + 55, + 140, + 296, + 247 + ], + "type": "inline_equation", + "content": "\\mathbb{P}" + }, + { + "bbox": [ + 55, + 140, + 296, + 247 + ], + "type": "text", + "content": " over the set of vertices as in [39, 62]. Therefore, the surface probability evaluated at the vertex " + }, + { + "bbox": [ + 55, + 140, + 296, + 247 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_i" + }, + { + "bbox": [ + 55, + 140, + 296, + 247 + ], + "type": "text", + "content": " is defined by the Euclidean distance to all other vertices" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 99, + 253, + 295, + 294 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 253, + 295, + 294 + ], + "spans": [ + { + "bbox": [ + 99, + 253, + 295, + 294 + ], + "type": "interline_equation", + "content": "\\mathrm {p} \\left(\\mathrm {v} _ {j} \\mid \\mathrm {v} _ {i}, \\mathrm {V}\\right) = \\frac {\\exp \\left(- \\frac {\\left| \\left| \\mathrm {v} _ {i} - \\mathrm {v} _ {j} \\right| \\right| ^ {2}}{2 \\sigma^ {2}}\\right)}{\\sum_ {k} \\exp \\left(\\frac {\\left| \\left| \\mathrm {v} _ {i} - \\mathrm {v} _ {k} \\right| \\right| ^ {2}}{2 \\sigma^ {2}}\\right)}. \\tag {9}", + "image_path": "ac12a699a00ba0c4e7f30e3b13b073ffad975a854e40ca5bdf862ca594156693.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 300, + 296, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 300, + 296, + 373 + ], + "spans": [ + { + "bbox": [ + 55, + 300, + 296, + 373 + ], + "type": "text", + "content": "Now we can render this surface probability using the camera pose " + }, + { + "bbox": [ + 55, + 300, + 296, + 373 + ], + "type": "inline_equation", + "content": "\\hat{\\mathrm{p}} (\\mathrm{v}_j|\\mathrm{M}_{def},\\pi ,u,v)" + }, + { + "bbox": [ + 55, + 300, + 296, + 373 + ], + "type": "text", + "content": ". For pixels outside the surface we set the background probability " + }, + { + "bbox": [ + 55, + 300, + 296, + 373 + ], + "type": "inline_equation", + "content": "\\hat{\\mathrm{p}} (\\beta |\\mathrm{M}_{def},\\pi ,u,v)" + }, + { + "bbox": [ + 55, + 300, + 296, + 373 + ], + "type": "text", + "content": " to 1. Jointly, we estimate these probabilities " + }, + { + "bbox": [ + 55, + 300, + 296, + 373 + ], + "type": "inline_equation", + "content": "\\bar{\\mathrm{p}} (\\mathrm{f}_j|\\Phi_w(\\mathrm{I},u,v),\\mathrm{F},\\beta)" + }, + { + "bbox": [ + 55, + 300, + 296, + 373 + ], + "type": "text", + "content": ". Thereupon, we optimize the pixel-wise cross-entropy loss" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 94, + 380, + 294, + 420 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 380, + 294, + 420 + ], + "spans": [ + { + "bbox": [ + 94, + 380, + 294, + 420 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {\\text {a p p}} \\left(\\Phi_ {w}, \\mathrm {M} _ {\\text {d e f}}, \\mathrm {I})\\right) = \\\\ - \\sum_ {j} \\log (\\bar {\\mathrm {p}} (f _ {j})) \\hat {\\mathrm {p}} (v _ {j}) - \\log (\\bar {\\mathrm {p}} (\\beta)) \\hat {\\mathrm {p}} (\\beta) \\cdot \\tag {10} \\\\ \\end{array}", + "image_path": "3dcaebee48b3d075feac9c71fa62ccd90a76e678274cc8b62978448c0761de5a.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 427, + 295, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 427, + 295, + 452 + ], + "spans": [ + { + "bbox": [ + 55, + 427, + 295, + 452 + ], + "type": "text", + "content": "Combined Objective. First, we only optimize the category-level template using only the geometric losses" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 63, + 459, + 294, + 472 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 459, + 294, + 472 + ], + "spans": [ + { + "bbox": [ + 63, + 459, + 294, + 472 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\lambda_ {C D} \\mathcal {L} _ {C D} + \\lambda_ {\\mathrm {m}} \\mathcal {L} _ {\\mathrm {m}} + \\lambda_ {\\mathrm {m d t}} \\mathcal {L} _ {\\mathrm {m d t}} + \\lambda_ {s d f} \\mathcal {L} _ {s d f}, \\tag {11}", + "image_path": "eb6b51b72cc63a4de33946ff7a3bc8e0502738b90f732af2db10c58f7eebc629.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 479, + 179, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 479, + 179, + 491 + ], + "spans": [ + { + "bbox": [ + 55, + 479, + 179, + 491 + ], + "type": "text", + "content": "with the balancing weights " + }, + { + "bbox": [ + 55, + 479, + 179, + 491 + ], + "type": "inline_equation", + "content": "\\lambda_{i}" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 491, + 295, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 491, + 295, + 528 + ], + "spans": [ + { + "bbox": [ + 55, + 491, + 295, + 528 + ], + "type": "text", + "content": "Second, we optimize the category-level feature field, the adapters, and the instance-level deformations using the weighted sum of geometric and appearance losses" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 63, + 534, + 295, + 563 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 534, + 295, + 563 + ], + "spans": [ + { + "bbox": [ + 63, + 534, + 295, + 563 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} = \\lambda_ {\\text {a p p}} \\mathcal {L} _ {\\text {a p p}} + \\lambda_ {C D} \\mathcal {L} _ {C D} + \\lambda_ {\\mathrm {m}} \\mathcal {L} _ {\\mathrm {m}} + \\lambda_ {\\mathrm {m d t}} \\mathcal {L} _ {\\mathrm {m d t}} \\tag {12} \\\\ + \\lambda_ {s d f} \\mathcal {L} _ {s d f} + \\lambda_ {d e f} \\mathcal {L} _ {d e f} + \\lambda_ {d e f - s m} \\mathcal {L} _ {d e f - s m}. \\\\ \\end{array}", + "image_path": "80389cc2477abb983e59a1dd8138a796328035f1474ff1d84e9473e9f0fab9bc.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 569, + 228, + 582 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 569, + 228, + 582 + ], + "spans": [ + { + "bbox": [ + 55, + 569, + 228, + 582 + ], + "type": "text", + "content": "3.4. Inference via Inverse Rendering" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 587, + 296, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 587, + 296, + 684 + ], + "spans": [ + { + "bbox": [ + 55, + 587, + 296, + 684 + ], + "type": "text", + "content": "During inference, our main application is to infer the shape latent " + }, + { + "bbox": [ + 55, + 587, + 296, + 684 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 55, + 587, + 296, + 684 + ], + "type": "text", + "content": " and the camera pose " + }, + { + "bbox": [ + 55, + 587, + 296, + 684 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 55, + 587, + 296, + 684 + ], + "type": "text", + "content": ". The shape latent can be directly predicted using the trained latent encoder " + }, + { + "bbox": [ + 55, + 587, + 296, + 684 + ], + "type": "inline_equation", + "content": "\\psi_{1}" + }, + { + "bbox": [ + 55, + 587, + 296, + 684 + ], + "type": "text", + "content": ". While the pose could alternatively also be predicted with a feedforward neural network, other works have demonstrated that inverse rendering can significantly improve the results [63]. Therefore, we optimize the pixelwise surface probabilities during inference applying the reconstruction loss" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 128, + 689, + 295, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 128, + 689, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 128, + 689, + 295, + 715 + ], + "type": "interline_equation", + "content": "\\max _ {\\pi} \\sum_ {u, v} \\bar {p} \\left(f _ {r} \\mid f _ {i}, F, \\beta\\right), \\tag {13}", + "image_path": "62de5724a098fc2e0db9898d7d6411b983b48645e497d0c6e5bbe0a2f0ae1cdf.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "text", + "content": "with the rendered image feature " + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "inline_equation", + "content": "\\mathrm{f}_r = \\Psi_r(\\mathrm{M}_{def},\\pi ,u,v)" + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "text", + "content": " and the image feature extracted from the test image " + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "inline_equation", + "content": "\\mathrm{f}_i = \\Phi_w(\\tau (\\mathrm{I},u,v))" + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "text", + "content": " . Note that we use three degrees of freedom for parametrizing the object rotation. Following [62], we initialize our pose using uniform sampling." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 142, + 395, + 156 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 142, + 395, + 156 + ], + "spans": [ + { + "bbox": [ + 313, + 142, + 395, + 156 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 162, + 555, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 162, + 555, + 281 + ], + "spans": [ + { + "bbox": [ + 313, + 162, + 555, + 281 + ], + "type": "text", + "content": "We conduct extensive experiments and evaluate the effectiveness of Common3D on various downstream tasks for a variety of object categories. We use the same trained model for all tasks, demonstrating the universal utilization of combining geometry with correspondence-aware appearance in one single model. We describe our experimental setup in Sec. 4.1. Subsequently, we demonstrate that our model improves over prior self-supervised work in 3D object pose estimation (Sec. 4.2), in instance segmentation (Sec. 4.3), and in semantic correspondence estimation (Sec. 4.4)." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 289, + 436, + 300 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 289, + 436, + 300 + ], + "spans": [ + { + "bbox": [ + 313, + 289, + 436, + 300 + ], + "type": "text", + "content": "4.1. Experimental Details" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 307, + 555, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 307, + 555, + 558 + ], + "spans": [ + { + "bbox": [ + 313, + 307, + 555, + 558 + ], + "type": "text", + "content": "Training Dataset. To train Common3D, we use the object-centric videos provided by the Common Objects in 3D (CO3D) dataset [45]. For each category, the videos capture a large variety of instances. In total, it provides 19k videos across 50 categories, with each video containing approximately 100 to 200 frames. Each video is supplemented by point clouds and relative camera poses using Structure-from-Motion [48] and object masks using PointRend [26]. We choose the same videos as used in [52], where videos that were captured too narrow to or not focusing on the object were filtered out, and only videos were retained where the full object is visible. For each category, we use a maximum of 50 videos. Due to the filtering we use fewer videos for the following ten categories: 17 for \"remote\", 15 for \"mouse\", 16 for \"tv\", 7 for \"toilet\", 41 for \"toybus\", 28 for \"hairdryer\", 49 for \"couch\", and 23 for \"cellphone\". Before training, we center each object's coordinate frame and normalize its scale such that its point cloud ranges from -1 to 1 on the axis with the maximum range. We achieve adopt the method proposed in [52] to obtain canonical camera pose annotations across videos in an unsupervised manner." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 558, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 556, + 714 + ], + "type": "text", + "content": "Evaluation on In-the-Wild Datasets. We evaluate our model at various tasks on in-the-wild data. In particular, we use three common datasets PASCAL3D+ [70], ObjectNet3D [71], and SPair-71k [37]. PASCAL3D+ and ObjectNet3D provide pose and segmentation annotations per image, which enables evaluation of 3D object pose estimation as well as instance segmentation. While PASCAL3D+ yields 12 rigid classes of PASCAL VOC 2012, ObjectNet3D covers, beyond that, 100 categories. We use the same 20 common categories across ObjectNet3D and CO3D as used in [52]. This even includes the domain gap of a toy bus in CO3D to a real one in PASCAL3D+ and ObjectNet3D. In total, we evaluate on 6233 images of PAS" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6473" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 70, + 294, + 139 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 294, + 139 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 294, + 139 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 294, + 139 + ], + "type": "table", + "html": "
Method∅∅∅∅∅○AVG
Sup.StarMap [83]83.294.475.479.868.885.882.5
VoGE [63]82.698.190.594.987.583.990.9
Unsup.ZSP [14]61.721.442.652.943.139.046.0
UOP3D [52]58.479.351.976.667.053.169.2
Ours67.282.967.880.978.151.175.3
", + "image_path": "905fb67d29b11a7f70c651914b39aced8d0534cec09f27fe369e0bd8e3a7c730.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 58, + 204, + 294, + 299 + ], + "blocks": [ + { + "bbox": [ + 55, + 147, + 295, + 192 + ], + "lines": [ + { + "bbox": [ + 55, + 147, + 295, + 192 + ], + "spans": [ + { + "bbox": [ + 55, + 147, + 295, + 192 + ], + "type": "text", + "content": "Table 1. Evaluation of 3D object pose estimation on 7 categories of PASCAL3D+, we supplement the missing results in the appendix. The reported metric is the " + }, + { + "bbox": [ + 55, + 147, + 295, + 192 + ], + "type": "inline_equation", + "content": "30^{\\circ}" + }, + { + "bbox": [ + 55, + 147, + 295, + 192 + ], + "type": "text", + "content": " accuracy. Note, that ZSP [14] requires also depth and runs an order of a magnitude slower." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 204, + 294, + 299 + ], + "lines": [ + { + "bbox": [ + 58, + 204, + 294, + 299 + ], + "spans": [ + { + "bbox": [ + 58, + 204, + 294, + 299 + ], + "type": "table", + "html": "
草草草草草草草草草草草草草草
ZSP [14]23.150.860.346.436.833.021.7
UOP3D [52]18.062.198.154.652.238.214.1
Ours25.577.099.051.865.944.719.4
草草草草草草草草草草草草AVG
ZSP [14]46.860.528.841.625.856.337.9
UOP3D [52]26.953.344.754.415.539.653.2
Ours34.554.238.552.415.765.052.4
", + "image_path": "c44b669dc5b5e275e0471c5a6369dab56cb3b253dab39fded6d6f7571e5d7640.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 373, + 295, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 373, + 295, + 433 + ], + "spans": [ + { + "bbox": [ + 54, + 373, + 295, + 433 + ], + "type": "text", + "content": "CAL3D+ using the same subset as in [61], and on 12039 images of ObjectNet3D. We center all objects, as proposed in [83]. Further, we evaluate our feature adapter on the semantic correspondence task of SPair71k across 6 common categories, which we validate on 3898 image pairs." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 434, + 295, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 434, + 295, + 626 + ], + "spans": [ + { + "bbox": [ + 54, + 434, + 295, + 626 + ], + "type": "text", + "content": "Implementation Details. As backbone we leverage DINOv2 [41] with the ViT-S model [2], for which we freeze all 21M parameters. To compare with prior work on semantic correspondence estimation Sec. 4.4, we additionally trained one ViT-B model. From an input resolution of " + }, + { + "bbox": [ + 54, + 434, + 295, + 626 + ], + "type": "inline_equation", + "content": "448 \\times 448" + }, + { + "bbox": [ + 54, + 434, + 295, + 626 + ], + "type": "text", + "content": ", the backbone maps to a " + }, + { + "bbox": [ + 54, + 434, + 295, + 626 + ], + "type": "inline_equation", + "content": "32 \\times 32" + }, + { + "bbox": [ + 54, + 434, + 295, + 626 + ], + "type": "text", + "content": " feature map. Our adapter, consisting of three ResNet blocks, upsamples the " + }, + { + "bbox": [ + 54, + 434, + 295, + 626 + ], + "type": "inline_equation", + "content": "32 \\times 32" + }, + { + "bbox": [ + 54, + 434, + 295, + 626 + ], + "type": "text", + "content": " feature map to " + }, + { + "bbox": [ + 54, + 434, + 295, + 626 + ], + "type": "inline_equation", + "content": "64 \\times 64" + }, + { + "bbox": [ + 54, + 434, + 295, + 626 + ], + "type": "text", + "content": ". For efficient differentiable rendering, we make use of the rasterizer proposed by [29], which achieves differentiability using anti-aliasing. For each category-specific morphable model we train, using Adam [25], for 20 epochs to optimize the template shape, and another 20 to optimize the appearance together with the deformation model. We use early stopping on a hold-out set comprising " + }, + { + "bbox": [ + 54, + 434, + 295, + 626 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 54, + 434, + 295, + 626 + ], + "type": "text", + "content": " of the videos. On a single NVIDIA GeForce RTX 2080 the training does not exceed 10 hours." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 635, + 228, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 635, + 228, + 647 + ], + "spans": [ + { + "bbox": [ + 55, + 635, + 228, + 647 + ], + "type": "text", + "content": "4.2. In-the-Wild 3D Pose Estimation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "content": "We evaluate our 3D object prior at single-image 3D pose estimation on in-the-wild data. We report the pose estimation results using the " + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "inline_equation", + "content": "30^{\\circ}" + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "content": " accuracy in Tab. 1 for PASCAL3D+ and in Tab. 2 for ObjectNet3D. In Tab. 1, we also report 3D pose estimation results from two supervised" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 316, + 70, + 553, + 110 + ], + "blocks": [ + { + "bbox": [ + 55, + 307, + 295, + 352 + ], + "lines": [ + { + "bbox": [ + 55, + 307, + 295, + 352 + ], + "spans": [ + { + "bbox": [ + 55, + 307, + 295, + 352 + ], + "type": "text", + "content": "Table 2. Evaluation of 3D object pose estimation on 20 categories of ObjectNet3D. We supplement the missing results in the appendix. We report the " + }, + { + "bbox": [ + 55, + 307, + 295, + 352 + ], + "type": "inline_equation", + "content": "30^{\\circ}" + }, + { + "bbox": [ + 55, + 307, + 295, + 352 + ], + "type": "text", + "content": " accuracy. Note, that ZSP [14] requires also depth input and runs an order of a magnitude slower." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 316, + 70, + 553, + 110 + ], + "lines": [ + { + "bbox": [ + 316, + 70, + 553, + 110 + ], + "spans": [ + { + "bbox": [ + 316, + 70, + 553, + 110 + ], + "type": "table", + "html": "
201918171615AVG
UOP3D [52]26.066.579.142.371.763.673.0
Ours29.364.782.552.276.965.977.0
", + "image_path": "29ac7d5ec9db44eae4a9023491e51cb158301fd0e22bf644a5795ef29c2d1f25.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 317, + 163, + 553, + 240 + ], + "blocks": [ + { + "bbox": [ + 313, + 118, + 555, + 152 + ], + "lines": [ + { + "bbox": [ + 313, + 118, + 555, + 152 + ], + "spans": [ + { + "bbox": [ + 313, + 118, + 555, + 152 + ], + "type": "text", + "content": "Table 3. Evaluation of instance segmentation on 7 categories of PASCAL3D+. We report the IoU in percentage with the occlusion-free instance mask provided by the ground truth pose and mesh." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 163, + 553, + 240 + ], + "lines": [ + { + "bbox": [ + 317, + 163, + 553, + 240 + ], + "spans": [ + { + "bbox": [ + 317, + 163, + 553, + 240 + ], + "type": "table", + "html": "
井井井井井井井
UOP3D [52]68.949.979.267.742.969.650.9
Ours69.454.182.868.153.366.853.6
井井井井井井井AVG
UOP3D [52]76.962.464.161.462.564.773.864.0
Ours83.170.662.771.765.366.778.367.0
", + "image_path": "1e42bdd88cd46d51108152778fc3022c9df50a496df7b22d1a2b92ff34a48b59.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 247, + 555, + 293 + ], + "lines": [ + { + "bbox": [ + 313, + 247, + 555, + 293 + ], + "spans": [ + { + "bbox": [ + 313, + 247, + 555, + 293 + ], + "type": "text", + "content": "Table 4. Evaluation of instance segmentation on 20 categories of ObjectNet3D, we supplement the missing results in the appendix. We report the IoU in percentage with the occlusion-free instance mask given by rendering the annotated mesh in the annotated pose." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 314, + 555, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 314, + 555, + 483 + ], + "spans": [ + { + "bbox": [ + 313, + 314, + 555, + 483 + ], + "type": "text", + "content": "methods [63, 83] that are trained from the PASCAL3D+ training set. Notably, all unsupervised methods are trained from CO3D data, as they require videos as input during training, and hence are evaluated in an out-of-domain setting. Our method improves over prior work on most of the categories except for categories with few shape variance across instances such as cellphone, microwave, and toaster. On average, the " + }, + { + "bbox": [ + 313, + 314, + 555, + 483 + ], + "type": "inline_equation", + "content": "30^{\\circ}" + }, + { + "bbox": [ + 313, + 314, + 555, + 483 + ], + "type": "text", + "content": " accuracy improves significantly over prior work from " + }, + { + "bbox": [ + 313, + 314, + 555, + 483 + ], + "type": "inline_equation", + "content": "69.2\\%" + }, + { + "bbox": [ + 313, + 314, + 555, + 483 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 314, + 555, + 483 + ], + "type": "inline_equation", + "content": "75.3\\%" + }, + { + "bbox": [ + 313, + 314, + 555, + 483 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 313, + 314, + 555, + 483 + ], + "type": "inline_equation", + "content": "+6.3" + }, + { + "bbox": [ + 313, + 314, + 555, + 483 + ], + "type": "text", + "content": " points on PASCAL3D+ and from " + }, + { + "bbox": [ + 313, + 314, + 555, + 483 + ], + "type": "inline_equation", + "content": "52.4\\%" + }, + { + "bbox": [ + 313, + 314, + 555, + 483 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 314, + 555, + 483 + ], + "type": "inline_equation", + "content": "56.8\\%" + }, + { + "bbox": [ + 313, + 314, + 555, + 483 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 313, + 314, + 555, + 483 + ], + "type": "inline_equation", + "content": "+4.4" + }, + { + "bbox": [ + 313, + 314, + 555, + 483 + ], + "type": "text", + "content": " points on ObjectNet3D. We attribute this performance increase to the correspondence learning with a deformable shape, which facilitates the training and, therefore, generalizes better. This is illustrated in the qualitative examples in Fig. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 490, + 501, + 503 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 490, + 501, + 503 + ], + "spans": [ + { + "bbox": [ + 313, + 490, + 501, + 503 + ], + "type": "text", + "content": "4.3. In-the-Wild Instance Segmentation" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 509, + 555, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 509, + 555, + 666 + ], + "spans": [ + { + "bbox": [ + 313, + 509, + 555, + 666 + ], + "type": "text", + "content": "We validate the deformation capabilities of Common3D using the instance segmentation masks of PASCAL3D+ in Tab. 3 and ObjectNet3D in Tab. 4. We compare our model with UOP3D [52], as it is the current state-of-the-art at self-supervised learning 3D object priors for common objects, and was also trained using CO3D videos. Common3D outperforms UOP3D on almost all categories and on average increases the IoU by 3.8 percentage points from " + }, + { + "bbox": [ + 313, + 509, + 555, + 666 + ], + "type": "inline_equation", + "content": "60.3\\%" + }, + { + "bbox": [ + 313, + 509, + 555, + 666 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 509, + 555, + 666 + ], + "type": "inline_equation", + "content": "64.1\\%" + }, + { + "bbox": [ + 313, + 509, + 555, + 666 + ], + "type": "text", + "content": " on PASCAL3D+ and by 3 percentage points, from " + }, + { + "bbox": [ + 313, + 509, + 555, + 666 + ], + "type": "inline_equation", + "content": "64\\%" + }, + { + "bbox": [ + 313, + 509, + 555, + 666 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 509, + 555, + 666 + ], + "type": "inline_equation", + "content": "67\\%" + }, + { + "bbox": [ + 313, + 509, + 555, + 666 + ], + "type": "text", + "content": " on ObjectNet3D. A qualitative comparison of our method against UOP3D is depicted in Fig. 3. It shows that Common3D achieves much more accurate highly accurate despite a large variability in the object instances." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 673, + 460, + 686 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 673, + 460, + 686 + ], + "spans": [ + { + "bbox": [ + 313, + 673, + 460, + 686 + ], + "type": "text", + "content": "4.4. Semantic Correspondence" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "content": "We demonstrate the significant benefit of our semantic feature adapter by evaluating its performance on semantic cor" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6474" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 71, + 553, + 240 + ], + "blocks": [ + { + "bbox": [ + 58, + 71, + 553, + 240 + ], + "lines": [ + { + "bbox": [ + 58, + 71, + 553, + 240 + ], + "spans": [ + { + "bbox": [ + 58, + 71, + 553, + 240 + ], + "type": "image", + "image_path": "e52efaa4cfbe7868ea5d56274a733ef4a8789e26c136429e4f8d75735f5575d5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 251, + 555, + 274 + ], + "lines": [ + { + "bbox": [ + 55, + 251, + 555, + 274 + ], + "spans": [ + { + "bbox": [ + 55, + 251, + 555, + 274 + ], + "type": "text", + "content": "Figure 3. Qualitative results on the ObjectNet3D dataset. In the second row the results of our method are illustrated, in the third the results of UOP3D [52]. Notably, our method fits the object more accurately, resulting in improved 3D pose and segmentation accuracy." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 57, + 292, + 294, + 347 + ], + "blocks": [ + { + "bbox": [ + 57, + 292, + 294, + 347 + ], + "lines": [ + { + "bbox": [ + 57, + 292, + 294, + 347 + ], + "spans": [ + { + "bbox": [ + 57, + 292, + 294, + 347 + ], + "type": "table", + "html": "
6073.367.246.366.042.259.4
SpherMaps [35]61.273.351.536.261.024.247.9
DINOv2 [41]62.052.366.244.260.731.954.0
Ours60.360.5
", + "image_path": "2d69897f918ddfb8db1495429864ab6d0545a626a72bc304bfd85a3fb2533466.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 420, + 295, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 420, + 295, + 588 + ], + "spans": [ + { + "bbox": [ + 54, + 420, + 295, + 588 + ], + "type": "text", + "content": "respondence estimation in Tab. 5. We evaluate our model compared to the self-supervised DINOv2 model as well as SpherMaps [35], which require coarse viewpoint annotations. We follow the evaluation protocol of SpherMaps and find correspondences by weighting the image features similarities from the DINOv2 backbone and the feature similarities from our adapter. We weight the backbone's features with 0.8 and ours with 0.2. On average, we can largely increase the PCK@0.1 compared to the self-supervised baseline DINOv2 by " + }, + { + "bbox": [ + 54, + 420, + 295, + 588 + ], + "type": "inline_equation", + "content": "6.1\\%" + }, + { + "bbox": [ + 54, + 420, + 295, + 588 + ], + "type": "text", + "content": " percentage points from " + }, + { + "bbox": [ + 54, + 420, + 295, + 588 + ], + "type": "inline_equation", + "content": "47.9\\%" + }, + { + "bbox": [ + 54, + 420, + 295, + 588 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 54, + 420, + 295, + 588 + ], + "type": "inline_equation", + "content": "54\\%" + }, + { + "bbox": [ + 54, + 420, + 295, + 588 + ], + "type": "text", + "content": ". For some categories, we can even achieve a competitive performance compared to the supervised baseline SpherMaps. We believe that one major challenge remains the domain gap between the training and test data." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 609, + 119, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 609, + 119, + 621 + ], + "spans": [ + { + "bbox": [ + 55, + 609, + 119, + 621 + ], + "type": "text", + "content": "4.5. Ablation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 629, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 295, + 715 + ], + "type": "text", + "content": "In Tab. 6, we ablate our model design choices regarding the adapter, the appearance loss as well as the instance deformation model. We observe, that instance-level deformation has only a marginal relevance for the pose estimation, but proves essential for segmentation and semantic correspondence estimation. Furthermore, the adapter and the dense appearance loss are significantly improving our method." + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 316, + 292, + 553, + 372 + ], + "blocks": [ + { + "bbox": [ + 55, + 356, + 295, + 390 + ], + "lines": [ + { + "bbox": [ + 55, + 356, + 295, + 390 + ], + "spans": [ + { + "bbox": [ + 55, + 356, + 295, + 390 + ], + "type": "text", + "content": "Table 5. Evaluation of the PCK@0.1 on 6 categories of SPair-71k [37]. Note, that SpherMaps [35] uses coarse viewpoint annotations, while DINOv2 and our approach do not require any." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 316, + 292, + 553, + 372 + ], + "lines": [ + { + "bbox": [ + 316, + 292, + 553, + 372 + ], + "spans": [ + { + "bbox": [ + 316, + 292, + 553, + 372 + ], + "type": "table", + "html": "
30°10°IoUPCK@0.1
w/o Adapt.25.96.660.532.4
w/o Adapt., w Rec. Loss13.92.156.232.4
w/o Dense App. Loss56.712.258.845.3
w/o Inst. Deform.75.820.258.645.3
Ours75.321.064.147.8
", + "image_path": "e98e70821037e2217d35c0f88a3baa08b1ab99dfa42a0887d4abe643abc3fa74.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 380, + 555, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 380, + 555, + 502 + ], + "spans": [ + { + "bbox": [ + 313, + 380, + 555, + 502 + ], + "type": "text", + "content": "Table 6. Ablation of our model design choices. We evaluate the 3D pose accuracy and instance segmentation on PASCAL3D+ using " + }, + { + "bbox": [ + 313, + 380, + 555, + 502 + ], + "type": "inline_equation", + "content": "30^{\\circ}" + }, + { + "bbox": [ + 313, + 380, + 555, + 502 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 380, + 555, + 502 + ], + "type": "inline_equation", + "content": "10^{\\circ}" + }, + { + "bbox": [ + 313, + 380, + 555, + 502 + ], + "type": "text", + "content": " accuracy and IoU. Further, the semantic correspondences are evaluated on SPair71-k using PCK@0.1. We observe that the adapter is crucial for our method, as well as the dense appearance loss. Further, our instance deformation model yields a minor disadvantage for the coarse " + }, + { + "bbox": [ + 313, + 380, + 555, + 502 + ], + "type": "inline_equation", + "content": "30^{\\circ}" + }, + { + "bbox": [ + 313, + 380, + 555, + 502 + ], + "type": "text", + "content": " pose estimation, while improving the fine-grained " + }, + { + "bbox": [ + 313, + 380, + 555, + 502 + ], + "type": "inline_equation", + "content": "10^{\\circ}" + }, + { + "bbox": [ + 313, + 380, + 555, + 502 + ], + "type": "text", + "content": " pose estimation, and proves essential for the instance segmentation and the semantic correspondences. Note that in this standard setting we use the ViT-S backbone, in contrast to the results reported in Tab. 5." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 525, + 388, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 525, + 388, + 538 + ], + "spans": [ + { + "bbox": [ + 313, + 525, + 388, + 538 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 544, + 555, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 544, + 555, + 674 + ], + "spans": [ + { + "bbox": [ + 313, + 544, + 555, + 674 + ], + "type": "text", + "content": "We have introduced Common3D, a new model that can learn 3D morphable models from a collection of casually captured videos completely self-supervised. This model can, at test time, estimate the 3D object shape, 2D-3D correspondences, and the 3D object pose. Common3D connects two different subfields of computer vision, self-supervised representation learning and the learning of 3D deformable object priors. We believe that in the long term, this may unify many vision tasks into a common inference process, while also creating a better 3D spatial understanding in vision models." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 677, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 556, + 713 + ], + "type": "text", + "content": "Acknowledgment. This work was funded by the German Research Foundation (DFG) - 468670075, and 499552394 (SFB 1597 - Small Data)." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6475" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 133 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 133 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 133 + ], + "type": "text", + "content": "[1] Oswald Aldrian and William AP Smith. Inverse rendering of faces with a 3d morphable model. IEEE transactions on pattern analysis and machine intelligence, 35(5):1080-1093, 2012. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 136, + 296, + 167 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 136, + 296, + 167 + ], + "spans": [ + { + "bbox": [ + 61, + 136, + 296, + 167 + ], + "type": "text", + "content": "[2] Dosovitskiy Alexey. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv: 2010.11929, 2020.7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 170, + 296, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 170, + 296, + 213 + ], + "spans": [ + { + "bbox": [ + 62, + 170, + 296, + 213 + ], + "type": "text", + "content": "[3] Dragomir Anguelov, Praveen Srinivasan, Daphne Koller, Sebastian Thrun, Jim Rodgers, and James Davis. Scape: shape completion and animation of people. In ACM SIGGRAPH 2005 Papers, pages 408-416. 2005. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 214, + 295, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 214, + 295, + 257 + ], + "spans": [ + { + "bbox": [ + 62, + 214, + 295, + 257 + ], + "type": "text", + "content": "[4] Yutong Bai, Angtian Wang, Adam Kortylewski, and Alan Yuille. Coke: Contrastive learning for robust keypoint detection. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 65-74, 2023. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 258, + 295, + 301 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 258, + 295, + 301 + ], + "spans": [ + { + "bbox": [ + 62, + 258, + 295, + 301 + ], + "type": "text", + "content": "[5] V BLANZ and T VETTER. A morphable model for the synthesis of 3d faces. In Computer graphics proceedings, annual conference series, pages 187-194. Association for Computing Machinery SIGGRAPH, 1999. 1, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 303, + 295, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 303, + 295, + 357 + ], + "spans": [ + { + "bbox": [ + 62, + 303, + 295, + 357 + ], + "type": "text", + "content": "[6] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 358, + 295, + 411 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 358, + 295, + 411 + ], + "spans": [ + { + "bbox": [ + 62, + 358, + 295, + 411 + ], + "type": "text", + "content": "[7] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 414, + 295, + 457 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 414, + 295, + 457 + ], + "spans": [ + { + "bbox": [ + 62, + 414, + 295, + 457 + ], + "type": "text", + "content": "[8] Zhiqin Chen and Hao Zhang. Learning implicit fields for generative shape modeling. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5939-5948, 2019. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 460, + 295, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 460, + 295, + 513 + ], + "spans": [ + { + "bbox": [ + 62, + 460, + 295, + 513 + ], + "type": "text", + "content": "[9] Devikalyan Das, Christopher Wewer, Raza Yunus, Eddy Ilg, and Jan Eric Lenssen. Neural parametric gaussians for monocular non-rigid object reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10715-10725, 2024. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 514, + 295, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 514, + 295, + 567 + ], + "spans": [ + { + "bbox": [ + 57, + 514, + 295, + 567 + ], + "type": "text", + "content": "[10] Yu Deng, Jiaolong Yang, and Xin Tong. Deformed implicit field: Modeling 3d shapes with learned dense correspondence. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10286-10296, 2021. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 570, + 295, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 570, + 295, + 633 + ], + "spans": [ + { + "bbox": [ + 57, + 570, + 295, + 633 + ], + "type": "text", + "content": "[11] Bernhard Egger, Sandro Schonborn, Andreas Schneider, Adam Kortylewski, Andreas Morel-Forster, Clemens Blumer, and Thomas Vetter. Occlusion-aware 3d morphable models and an illumination prior for face image analysis. International Journal of Computer Vision, 126:1269-1287, 2018. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 636, + 295, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 636, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 57, + 636, + 295, + 689 + ], + "type": "text", + "content": "[12] Shubham Goel, Angjoo Kanazawa, and Jitendra Malik. Shape and viewpoint without keypoints. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XV 16, pages 88-104. Springer, 2020. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 692, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 692, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 692, + 295, + 713 + ], + "type": "text", + "content": "[13] Shubham Goel, Georgia Gkioxari, and Jitendra Malik. Differentiable stereopsis: Meshes from multiple views using" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 106 + ], + "type": "text", + "content": "differentiable rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8635-8644, 2022. 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 107, + 553, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 107, + 553, + 151 + ], + "spans": [ + { + "bbox": [ + 317, + 107, + 553, + 151 + ], + "type": "text", + "content": "[14] Walter Goodwin, Sagar Vaze, Ioannis Havoutis, and Ingmar Posner. Zero-shot category-level object pose estimation. In European Conference on Computer Vision, pages 516-532. Springer, 2022. 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 152, + 553, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 152, + 553, + 185 + ], + "spans": [ + { + "bbox": [ + 317, + 152, + 553, + 185 + ], + "type": "text", + "content": "[15] Amos Gropp, Lior Yariv, Niv Haim, Matan Atzmon, and Yaron Lipman. Implicit geometric regularization for learning shapes. arXiv preprint arXiv:2002.10099, 2020. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 186, + 553, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 186, + 553, + 230 + ], + "spans": [ + { + "bbox": [ + 317, + 186, + 553, + 230 + ], + "type": "text", + "content": "[16] Thibault Groueix, Matthew Fisher, Vladimir G Kim, Bryan C Russell, and Mathieu Aubry. Atlasnet: a papiermâché approach to learning 3d surface generation (2018). arXiv preprint arXiv:1802.05384, 11, 1802. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 232, + 553, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 232, + 553, + 274 + ], + "spans": [ + { + "bbox": [ + 317, + 232, + 553, + 274 + ], + "type": "text", + "content": "[17] Paul Henderson, Vagia Tsiminaki, and Christoph H Lampert. Leveraging 2d data to learn textured 3d mesh generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7498-7507, 2020. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 277, + 553, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 277, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 317, + 277, + 553, + 319 + ], + "type": "text", + "content": "[18] Tao Hu, Liwei Wang, Xiaogang Xu, Shu Liu, and Jiaya Jia. Self-supervised 3d mesh reconstruction from single images. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6002-6011, 2021. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 322, + 555, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 322, + 555, + 376 + ], + "spans": [ + { + "bbox": [ + 317, + 322, + 555, + 376 + ], + "type": "text", + "content": "[19] Wei-Chih Hung, Varun Jampani, Sifei Liu, Pavlo Molchanov, Ming-Hsuan Yang, and Jan Kautz. Scops: Self-supervised co-part segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 869-878, 2019. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 377, + 553, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 377, + 553, + 430 + ], + "spans": [ + { + "bbox": [ + 317, + 377, + 553, + 430 + ], + "type": "text", + "content": "[20] Tomas Jakab, Ruining Li, Shangzhe Wu, Christian Rupprecht, and Andrea Vedaldi. Farm3d: Learning articulated 3d animals by distilling 2d diffusion. In 2024 International Conference on 3D Vision (3DV), pages 852-861. IEEE, 2024. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 434, + 553, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 434, + 553, + 477 + ], + "spans": [ + { + "bbox": [ + 317, + 434, + 553, + 477 + ], + "type": "text", + "content": "[21] Hanbyul Joo, Tomas Simon, and Yaser Sheikh. Total capture: A 3d deformation model for tracking faces, hands, and bodies. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8320-8329, 2018. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 479, + 553, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 479, + 553, + 532 + ], + "spans": [ + { + "bbox": [ + 317, + 479, + 553, + 532 + ], + "type": "text", + "content": "[22] Angjoo Kanazawa, Shubham Tulsiani, Alexei A Efros, and Jitendra Malik. Learning category-specific mesh reconstruction from image collections. In Proceedings of the European Conference on Computer Vision (ECCV), pages 371-386, 2018. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 535, + 553, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 535, + 553, + 588 + ], + "spans": [ + { + "bbox": [ + 317, + 535, + 553, + 588 + ], + "type": "text", + "content": "[23] Angjoo Kanazawa, Shubham Tulsiani, Alexei A Efros, and Jitendra Malik. Learning category-specific mesh reconstruction from image collections. In Proceedings of the European Conference on Computer Vision (ECCV), pages 371-386, 2018. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 590, + 553, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 590, + 553, + 643 + ], + "spans": [ + { + "bbox": [ + 317, + 590, + 553, + 643 + ], + "type": "text", + "content": "[24] Abhishek Kar, Shubham Tulsiani, Joao Carreira, and Jitendra Malik. Category-specific object reconstruction from a single image. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1966-1974, 2015. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 647, + 553, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 667 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 667 + ], + "type": "text", + "content": "[25] Diederik P Kingma. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 7" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 670, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 670, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 670, + 553, + 713 + ], + "type": "text", + "content": "[26] Alexander Kirillov, Yuxin Wu, Kaiming He, and Ross Girshick. Pointrend: Image segmentation as rendering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9799-9808, 2020. 5, 6" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "6476" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "text", + "content": "[27] Filippos Kokkinos and Iasonas Kokkinos. To the point: Correspondence-driven monocular 3d category reconstruction. Advances in Neural Information Processing Systems, 34:7760-7772, 2021. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 172 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 172 + ], + "type": "text", + "content": "[28] Nilesh Kulkarni, Abhinav Gupta, David F Fouhey, and Shubham Tulsiani. Articulation-aware canonical surface mapping. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 452-461, 2020. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 175, + 294, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 294, + 218 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 294, + 218 + ], + "type": "text", + "content": "[29] Samuli Laine, Janne Hellsten, Tero Karras, Yeongho Seol, Jaakko Lehtinen, and Timo Aila. Modular primitives for high-performance differentiable rendering. ACM Transactions on Graphics, 39(6), 2020. 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 220, + 294, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 220, + 294, + 252 + ], + "spans": [ + { + "bbox": [ + 56, + 220, + 294, + 252 + ], + "type": "text", + "content": "[30] Aldo Laurentini. The visual hull concept for silhouette-based image understanding. IEEE Transactions on pattern analysis and machine intelligence, 16(2):150-162, 1994. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 254, + 294, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 254, + 294, + 319 + ], + "spans": [ + { + "bbox": [ + 56, + 254, + 294, + 319 + ], + "type": "text", + "content": "[31] Xueting Li, Sifei Liu, Kihwan Kim, Shalini De Mello, Varun Jampani, Ming-Hsuan Yang, and Jan Kautz. Self-supervised single-view 3d reconstruction via semantic consistency. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIV 16, pages 677-693. Springer, 2020. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 321, + 294, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 321, + 294, + 376 + ], + "spans": [ + { + "bbox": [ + 56, + 321, + 294, + 376 + ], + "type": "text", + "content": "[32] Zizhang Li, Dor Litvak, Ruining Li, Yunzhi Zhang, Tomas Jakab, Christian Rupprecht, Shangzhe Wu, Andrea Vedaldi, and Jiajun Wu. Learning the 3d fauna of the web. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9752-9762, 2024. 2, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 377, + 294, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 377, + 294, + 432 + ], + "spans": [ + { + "bbox": [ + 56, + 377, + 294, + 432 + ], + "type": "text", + "content": "[33] Yebin Liu, Juergen Gall, Carsten Stoll, Qionghai Dai, Hans-Peter Seidel, and Christian Theobalt. Markerless motion capture of multiple characters using multiview image segmentation. IEEE transactions on pattern analysis and machine intelligence, 35(11):2720-2735, 2013. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 434, + 294, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 434, + 294, + 476 + ], + "spans": [ + { + "bbox": [ + 56, + 434, + 294, + 476 + ], + "type": "text", + "content": "[34] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J Black. Smpl: a skinned multiperson linear model. ACM Transactions on Graphics (TOG), 34(6):1-16, 2015. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 478, + 294, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 478, + 294, + 532 + ], + "spans": [ + { + "bbox": [ + 56, + 478, + 294, + 532 + ], + "type": "text", + "content": "[35] Octave Mariotti, Oisin Mac Aodha, and Hakan Bilen. Improving semantic correspondence with viewpoint-guided spherical maps. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19521-19530, 2024. 2, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 534, + 294, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 534, + 294, + 588 + ], + "spans": [ + { + "bbox": [ + 56, + 534, + 294, + 588 + ], + "type": "text", + "content": "[36] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4460-4470, 2019. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 590, + 294, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 590, + 294, + 623 + ], + "spans": [ + { + "bbox": [ + 56, + 590, + 294, + 623 + ], + "type": "text", + "content": "[37] Juhong Min, Jongmin Lee, Jean Ponce, and Minsu Cho. Spair-71k: A large-scale benchmark for semantic correspondence. arXiv preprint arXiv:1908.10543, 2019. 2, 6, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 624, + 294, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 294, + 668 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 294, + 668 + ], + "type": "text", + "content": "[38] Tom Monnier, Matthew Fisher, Alexei A Efros, and Mathieu Aubry. Share with thy neighbors: Single-view reconstruction by cross-instance consistency. In European Conference on Computer Vision, pages 285-303. Springer, 2022. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 670, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 294, + 713 + ], + "type": "text", + "content": "[39] Natalia Neverova, David Novotny, Marc Szafraniec, Vasil Khalidov, Patrick Labatut, and Andrea Vedaldi. Continuous surface embeddings. Advances in Neural Information Processing Systems, 33:17258-17270, 2020. 6" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "type": "text", + "content": "[40] Michael Oechsle, Lars Mescheder, Michael Niemeyer, Thilo Strauss, and Andreas Geiger. Texture fields: Learning texture representations in function space. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4531-4540, 2019. 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 129, + 553, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 553, + 183 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 553, + 183 + ], + "type": "text", + "content": "[41] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 2, 4, 7, 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 184, + 553, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 184, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 316, + 184, + 553, + 239 + ], + "type": "text", + "content": "[42] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 165-174, 2019. 2, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 239, + 553, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 239, + 553, + 293 + ], + "spans": [ + { + "bbox": [ + 316, + 239, + 553, + 293 + ], + "type": "text", + "content": "[43] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5865-5874, 2021. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 294, + 553, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 294, + 553, + 338 + ], + "spans": [ + { + "bbox": [ + 316, + 294, + 553, + 338 + ], + "type": "text", + "content": "[44] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems, 30, 2017. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 338, + 553, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 338, + 553, + 403 + ], + "spans": [ + { + "bbox": [ + 316, + 338, + 553, + 403 + ], + "type": "text", + "content": "[45] Jeremy Reizenstein, Roman Shapovalov, Philipp Henzler, Luca Sbordone, Patrick Labatut, and David Novotny. Common objects in 3d: Large-scale learning and evaluation of real-life 3d category reconstruction. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10901-10911, 2021. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 404, + 553, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 404, + 553, + 459 + ], + "spans": [ + { + "bbox": [ + 316, + 404, + 553, + 459 + ], + "type": "text", + "content": "[46] Lukasz Romaszko, Christopher KI Williams, Pol Moreno, and Pushmeet Kohli. Vision-as-inverse-graphics: Obtaining a rich 3d explanation of a scene from a single image. In Proceedings of the IEEE International Conference on Computer Vision Workshops, pages 851–859, 2017. 2, 3, 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 460, + 553, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 460, + 553, + 514 + ], + "spans": [ + { + "bbox": [ + 316, + 460, + 553, + 514 + ], + "type": "text", + "content": "[47] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 514, + 553, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 514, + 553, + 548 + ], + "spans": [ + { + "bbox": [ + 316, + 514, + 553, + 548 + ], + "type": "text", + "content": "[48] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-Motion Revisited. In Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 5, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 548, + 553, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 548, + 553, + 591 + ], + "spans": [ + { + "bbox": [ + 316, + 548, + 553, + 591 + ], + "type": "text", + "content": "[49] Sandro Schonborn, Bernhard Egger, Andreas Morel-Forster, and Thomas Vetter. Markov chain monte carlo for automated face image analysis. International Journal of Computer Vision, 123:160–183, 2017. 2, 3, 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 592, + 553, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 592, + 553, + 646 + ], + "spans": [ + { + "bbox": [ + 316, + 592, + 553, + 646 + ], + "type": "text", + "content": "[50] Tianchang Shen, Jun Gao, Kangxue Yin, Ming-Yu Liu, and Sanja Fidler. Deep marching tetrahedra: a hybrid representation for high-resolution 3d shape synthesis. Advances in Neural Information Processing Systems, 34:6087-6101, 2021. 2, 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "type": "text", + "content": "[51] Aleksandar Shtedritski, Christian Rupprecht, and Andrea Vedaldi. Shic: Shape-image correspondences with no key-point supervision. arXiv preprint arXiv:2407.18907, 2024. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "text", + "content": "[52] Leonhard Sommer, Artur Jesslen, Eddy Ilg, and Adam Kortylewski. Unsupervised learning of category-level 3d pose" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6477" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 77, + 72, + 295, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 72, + 295, + 106 + ], + "spans": [ + { + "bbox": [ + 77, + 72, + 295, + 106 + ], + "type": "text", + "content": "from object-centric videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22787-22796, 2024. 2, 5, 6, 7, 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 295, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 295, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 295, + 162 + ], + "type": "text", + "content": "[53] Keqiang Sun, Dor Litvak, Yunzhi Zhang, Hongsheng Li, Ji-jun Wu, and Shangzhe Wu. Ponomation: Learning articulated 3d animal motions from unlabeled online videos. In European Conference on Computer Vision, pages 100-119. Springer, 2025. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 294, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 294, + 218 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 294, + 218 + ], + "type": "text", + "content": "[54] Shanlin Sun, Kun Han, Deying Kong, Hao Tang, Xiangyi Yan, and Xiaohui Xie. Topology-preserving shape reconstruction and registration via neural diffeomorphic flow. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20845-20855, 2022. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 220, + 294, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 220, + 294, + 285 + ], + "spans": [ + { + "bbox": [ + 56, + 220, + 294, + 285 + ], + "type": "text", + "content": "[55] Ayush Tewari, Michael Zollhofer, Hyeongwoo Kim, Pablo Garrido, Florian Bernard, Patrick Perez, and Christian Theobalt. Mofa: Model-based deep convolutional face autoencoder for unsupervised monocular reconstruction. In Proceedings of the IEEE international conference on computer vision workshops, pages 1274-1283, 2017. 1, 2, 3, 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 286, + 294, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 286, + 294, + 352 + ], + "spans": [ + { + "bbox": [ + 56, + 286, + 294, + 352 + ], + "type": "text", + "content": "[56] Ayush Tewari, Florian Bernard, Pablo Garrido, Gaurav Bharaj, Mohamed Elgharib, Hans-Peter Seidel, Patrick Pérez, Michael Zollhofer, and Christian Theobalt. Fml: Face model learning from videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10812-10822, 2019. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 354, + 295, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 354, + 295, + 408 + ], + "spans": [ + { + "bbox": [ + 56, + 354, + 295, + 408 + ], + "type": "text", + "content": "[57] Ayush Tewari, Hans-Peter Seidel, Mohamed Elgharib, Christian Theobalt, et al. Learning complete 3d morphable face models from images and videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3361-3371, 2021. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 410, + 295, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 410, + 295, + 475 + ], + "spans": [ + { + "bbox": [ + 56, + 410, + 295, + 475 + ], + "type": "text", + "content": "[58] Ayush Tewari, Xingang Pan, Ohad Fried, Maneesh Agrawala, Christian Theobalt, et al. Disentangled3d: Learning a 3d generative model with disentangled geometry and appearance from monocular images. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1516-1525, 2022. 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 476, + 294, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 476, + 294, + 520 + ], + "spans": [ + { + "bbox": [ + 56, + 476, + 294, + 520 + ], + "type": "text", + "content": "[59] Shubham Tulsiani, Abhishek Kar, Joao Carreira, and Jitendra Malik. Learning category-specific deformable 3d models for object reconstruction. IEEE transactions on pattern analysis and machine intelligence, 39(4):719-731, 2016. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 522, + 294, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 522, + 294, + 555 + ], + "spans": [ + { + "bbox": [ + 56, + 522, + 294, + 555 + ], + "type": "text", + "content": "[60] Shubham Tulsiani, Nilesh Kulkarni, and Abhinav Gupta. Implicit mesh reconstruction from unannotated image collections. arXiv preprint arXiv:2007.08504, 2020. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 556, + 294, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 556, + 294, + 610 + ], + "spans": [ + { + "bbox": [ + 56, + 556, + 294, + 610 + ], + "type": "text", + "content": "[61] Angtian Wang, Yihong Sun, Adam Kortylewski, and Alan L Yuille. Robust object detection under occlusion with context-aware compositionalnets. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12645-12654, 2020. 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 612, + 294, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 612, + 294, + 644 + ], + "spans": [ + { + "bbox": [ + 56, + 612, + 294, + 644 + ], + "type": "text", + "content": "[62] Angtian Wang, Adam Kortylewski, and Alan Yuille. Nemo: Neural mesh models of contrastive features for robust 3d pose estimation. arXiv preprint arXiv:2101.12378, 2021. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 647, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 294, + 689 + ], + "type": "text", + "content": "[63] Angtian Wang, Peng Wang, Jian Sun, Adam Kortylewski, and Alan Yuille. Voge: a differentiable volume renderer using gaussian ellipsoids for analysis-by-synthesis. arXiv preprint arXiv:2205.15401, 2022. 6, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 691, + 294, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 714 + ], + "type": "text", + "content": "[64] Haowen Wang, Zhipeng Fan, Zhen Zhao, Zhengping Che, Zhiyuan Xu, Dong Liu, Feifei Feng, Yakun Huang, Xiuquan" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 714 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 116 + ], + "type": "text", + "content": "Qiao, and Jian Tang. Dtf-net: Category-level pose estimation and shape reconstruction via deformable template field. In Proceedings of the 31st ACM International Conference on Multimedia, pages 3676-3685, 2023. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 118, + 553, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 553, + 171 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 553, + 171 + ], + "type": "text", + "content": "[65] Nanyang Wang, Yinda Zhang, Zhuwen Li, Yanwei Fu, Wei Liu, and Yu-Gang Jiang. Pixel2mesh: Generating 3d mesh models from single rgb images. In Proceedings of the European conference on computer vision (ECCV), pages 52-67, 2018. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 173, + 555, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 173, + 555, + 227 + ], + "spans": [ + { + "bbox": [ + 316, + 173, + 555, + 227 + ], + "type": "text", + "content": "[66] Shangzhe Wu, Christian Rupprecht, and Andrea Vedaldi. Unsupervised learning of probably symmetric deformable 3d objects from images in the wild. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1-10, 2020. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 228, + 555, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 228, + 555, + 282 + ], + "spans": [ + { + "bbox": [ + 316, + 228, + 555, + 282 + ], + "type": "text", + "content": "[67] Shangzhe Wu, Christian Rupprecht, and Andrea Vedaldi. Unsupervised learning of probably symmetric deformable 3d objects from images in the wild. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1-10, 2020. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 284, + 553, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 284, + 553, + 326 + ], + "spans": [ + { + "bbox": [ + 316, + 284, + 553, + 326 + ], + "type": "text", + "content": "[68] Shangzhe Wu, Tomas Jakab, Christian Rupprecht, and Andrea Vedaldi. Dove: Learning deformable 3d objects by watching videos. International Journal of Computer Vision, 131(10):2623-2634, 2023. 2, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 327, + 553, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 327, + 553, + 381 + ], + "spans": [ + { + "bbox": [ + 316, + 327, + 553, + 381 + ], + "type": "text", + "content": "[69] Shangzhe Wu, Ruining Li, Tomas Jakab, Christian Rupprecht, and Andrea Vedaldi. Magicpony: Learning articulated 3d animals in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8792-8802, 2023. 2, 3, 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 383, + 553, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 383, + 553, + 426 + ], + "spans": [ + { + "bbox": [ + 316, + 383, + 553, + 426 + ], + "type": "text", + "content": "[70] Yu Xiang, Roozbeh Mottaghi, and Silvio Savarese. Beyond Pascal: A benchmark for 3d object detection in the wild. In IEEE winter conference on applications of computer vision, pages 75-82. IEEE, 2014. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 427, + 553, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 427, + 553, + 502 + ], + "spans": [ + { + "bbox": [ + 316, + 427, + 553, + 502 + ], + "type": "text", + "content": "[71] Yu Xiang, Wonhui Kim, Wei Chen, Jingwei Ji, Christopher Choy, Hao Su, Roozbeh Mottaghi, Leonidas Guibas, and Silvio Savarese. Objectnet3d: A large scale database for 3d object recognition. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part VIII 14, pages 160-176. Springer, 2016. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 504, + 553, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 504, + 553, + 569 + ], + "spans": [ + { + "bbox": [ + 316, + 504, + 553, + 569 + ], + "type": "text", + "content": "[72] Gengshan Yang, Deqing Sun, Varun Jampani, Daniel Vlasic, Forrester Cole, Huiwen Chang, Deva Ramanan, William T Freeman, and Ce Liu. Lasr: Learning articulated shape reconstruction from a monocular video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15980-15989, 2021. 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 571, + 553, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 571, + 553, + 624 + ], + "spans": [ + { + "bbox": [ + 316, + 571, + 553, + 624 + ], + "type": "text", + "content": "[73] Gengshan Yang, Deqing Sun, Varun Jampani, Daniel Vlasic, Forrester Cole, Ce Liu, and Deva Ramanan. Viser: Video-specific surface embeddings for articulated 3d shape reconstruction. Advances in Neural Information Processing Systems, 34:19326-19338, 2021. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 625, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 625, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 625, + 553, + 689 + ], + "type": "text", + "content": "[74] Gengshan Yang, Minh Vo, Natalia Neverova, Deva Ramanan, Andrea Vedaldi, and Hanbyul Joo. Banmo: Building animatable 3d neural models from many casual videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2863-2873, 2022. 2, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 691, + 553, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 714 + ], + "type": "text", + "content": "[75] Chun-Han Yao, Wei-Chih Hung, Yuanzhen Li, Michael Rubinstein, Ming-Hsuan Yang, and Varun Jampani. Lassie:" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6478" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 563 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 106 + ], + "type": "text", + "content": "Learning articulated shapes from sparse image ensemble via 3d part discovery. Advances in Neural Information Processing Systems, 35:15296-15308, 2022. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 295, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 295, + 173 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 295, + 173 + ], + "type": "text", + "content": "[76] Chun-Han Yao, Wei-Chih Hung, Yuanzhen Li, Michael Rubinstein, Ming-Hsuan Yang, and Varun Jampani. Hi-lassie: High-fidelity articulated shape and skeleton discovery from sparse image ensemble. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4853-4862, 2023. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 175, + 294, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 294, + 206 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 294, + 206 + ], + "type": "text", + "content": "[77] Alan Yuille and Daniel Kersten. Vision as bayesian inference: analysis by synthesis? Trends in cognitive sciences, 10(7):301-308, 2006. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 208, + 294, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 208, + 294, + 272 + ], + "spans": [ + { + "bbox": [ + 56, + 208, + 294, + 272 + ], + "type": "text", + "content": "[78] Junyi Zhang, Charles Herrmann, Junhwa Hur, Eric Chen, Varun Jampani, Deqing Sun, and Ming-Hsuan Yang. Telling left from right: Identifying geometry-aware semantic correspondence. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3076-3085, 2024. 2, 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 274, + 294, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 274, + 294, + 328 + ], + "spans": [ + { + "bbox": [ + 56, + 274, + 294, + 328 + ], + "type": "text", + "content": "[79] Junyi Zhang, Charles Herrmann, Junhwa Hur, Luisa Polania Cabrera, Varun Jampani, Deqing Sun, and Ming-Hsuan Yang. A tale of two features: Stable diffusion complements dino for zero-shot semantic correspondence. Advances in Neural Information Processing Systems, 36, 2024. 2, 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 331, + 294, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 331, + 294, + 373 + ], + "spans": [ + { + "bbox": [ + 56, + 331, + 294, + 373 + ], + "type": "text", + "content": "[80] Kaifeng Zhang, Yang Fu, Shubhankar Borse, Hong Cai, Fatih Porikli, and Xiaolong Wang. Self-supervised geometric correspondence for category-level 6d object pose estimation in the wild. arXiv preprint arXiv:2210.07199, 2022. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 375, + 295, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 375, + 295, + 418 + ], + "spans": [ + { + "bbox": [ + 56, + 375, + 295, + 418 + ], + "type": "text", + "content": "[81] Yunzhi Zhang, Shangzhe Wu, Noah Snavely, and Jiajun Wu. Seeing a rose in five thousand ways. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 962-971, 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 420, + 295, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 420, + 295, + 464 + ], + "spans": [ + { + "bbox": [ + 56, + 420, + 295, + 464 + ], + "type": "text", + "content": "[82] Zerong Zheng, Tao Yu, Qionghai Dai, and Yebin Liu. Deep implicit templates for 3d shape representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1429-1439, 2021. 3, 4, 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 465, + 294, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 465, + 294, + 508 + ], + "spans": [ + { + "bbox": [ + 56, + 465, + 294, + 508 + ], + "type": "text", + "content": "[83] Xingyi Zhou, Arjun Karpur, Linjie Luo, and Qixing Huang. Starmap for category-agnostic keypoint and viewpoint estimation. In Proceedings of the European Conference on Computer Vision (ECCV), pages 318–334, 2018. 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 510, + 294, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 510, + 294, + 563 + ], + "spans": [ + { + "bbox": [ + 56, + 510, + 294, + 563 + ], + "type": "text", + "content": "[84] Silvia Zuffi, Angjoo Kanazawa, David W Jacobs, and Michael J Black. 3d menagerie: Modeling the 3d shape and pose of animals. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6365-6373, 2017. 2" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6479" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Commonsense Video Question Answering through Video-Grounded Entailment Tree Reasoning/e1080574-1050-4c6e-8509-597d8efaff00_content_list.json b/2025/Commonsense Video Question Answering through Video-Grounded Entailment Tree Reasoning/e1080574-1050-4c6e-8509-597d8efaff00_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..7f5e92b2dafc92745908abd53303ed31e7876741 --- /dev/null +++ b/2025/Commonsense Video Question Answering through Video-Grounded Entailment Tree Reasoning/e1080574-1050-4c6e-8509-597d8efaff00_content_list.json @@ -0,0 +1,1293 @@ +[ + { + "type": "text", + "text": "Commonsense Video Question Answering through Video-Grounded Entailment Tree Reasoning", + "text_level": 1, + "bbox": [ + 156, + 130, + 841, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Huabin Liu $^{1,3*}$ , Filip Ilievski $^{2}$ , Cees G. M. Snoek $^{3}$ $^{1}$ Shanghai Jiao Tong University $^{2}$ Vrije Universiteit Amsterdam $^{3}$ University of Amsterdam huabinliu@sjtu.edu.cn, f.ilievski@vu.nl, c.g.m.snoek@uva.nl", + "bbox": [ + 116, + 200, + 880, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 291, + 326, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "This paper proposes the first video-grounded entailment tree reasoning method for commonsense video question answering (VQA). Despite the remarkable progress of large visual-language models (VLMs), there are growing concerns that they learn spurious correlations between videos and likely answers, reinforced by their black-box nature and remaining benchmarking biases. Our method explicitly grounds VQA tasks to video fragments in four steps: entailment tree construction, video-language entailment verification, tree reasoning, and dynamic tree expansion. A vital benefit of the method is its generalizability to current video- and image-based VLMs across reasoning types. To support fair evaluation, we devise a de-biasing procedure based on large-language models that rewrites VQA benchmark answer sets to enforce model reasoning. Systematic experiments on existing and de-biased benchmarks highlight the impact of our method components across benchmarks, VLMs, and reasoning types.", + "bbox": [ + 89, + 323, + 483, + 597 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 627, + 220, + 642 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "This paper proposes a video-grounded reasoning method for commonsense video question answering (VQA). VQA has a long tradition in computer vision [11, 16, 25, 26], with remarkable recent progress obtained through video-and image-language models [10, 12, 14, 17, 33] (throughout this paper collectively referred to as vision-language models, or VLMs). Yet, there are growing concerns that their improved performance is based on learning shortcut associations between videos and likely answers, as opposed to reasoning [27]. Such concerns are reinforced by the black-box nature of these models [12, 14], which prohibits a deeper understanding of their decision-making process.", + "bbox": [ + 89, + 647, + 482, + 828 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We are inspired by recent work in natural language processing, where entailment trees have emerged as a mechanism to explicitly analyze answer candidates, using LLMs", + "bbox": [ + 89, + 829, + 483, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d1fff7e0f875c0e1b0d4fdbcbc480d3ce9de73267e40e372e9b2108e926a526f.jpg", + "image_caption": [ + "Figure 1. Given a video questioning answering task, our framework performs explicit reasoning over an entailment tree, where answer options are transformed into statements. These statements are then recursively decomposed and verified based on videotegued evidence relevant to the question." + ], + "image_footnote": [], + "bbox": [ + 517, + 290, + 906, + 558 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "to recursively decompose a candidate into hypotheses and natural language inference formalisms to evaluate the hypotheses [4]. Entailment trees provide an explicit reasoning chain that explains the model's decision-making process and enables verification of each step, thus addressing concerns about shortcut learning. Recently, Sanders et al. [19] have devised a mechanism to apply entailment trees to videos. However, their work assumes video transcripts are explicitly provided to evaluate answers, thus avoiding the complexity of grounding hypotheses into video content. In this paper, we propose the first video-grounded entailment tree reasoning method for commonsense VQA.", + "bbox": [ + 509, + 657, + 906, + 838 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Our method explicitly grounds VQA tasks to video fragments in four steps: (i) entailment tree construction, (ii) video-language entailment verification, (iii) tree reasoning, and (iv) dynamic tree expansion. As shown in Fig. 1, given", + "bbox": [ + 511, + 839, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Work was done as a visiting student at University of Amsterdam", + "bbox": [ + 114, + 886, + 465, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "3262", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "a video and a multiple-choice question, we generate a statement for each answer candidate that acts as a first-level hypothesis. We decompose each statement iteratively, aiming to produce sub-statements that can be confidently verified in the video. The video is itself decomposed into partitions, consisting of sets of frames. Verifying each statement is then a matter of aligning it to a video partition. A vital benefit of the method is its generalizability to current video and image-based VLMs across reasoning types, including temporal and causal. To demonstrate its video reasoning ability, we develop an answer-set de-bias procedure supported by an LLM that ensures that VQA benchmarks [11, 26] are adequate for reasoning in videos without relying on spurious correlations. Our experiments show that our video-grounded entailment tree method consistently improves video- and image-based baselines on both the existing and de-biased benchmarks. Moreover, it performs on par with, and often better than, state-of-the-art video-based VLMs while leveraging $257 \\times$ fewer parameters. Further ablations show that the method benefits from considering both textual and video information and that its performance is especially strong on causal and temporal questions.", + "bbox": [ + 93, + 90, + 480, + 422 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 94, + 441, + 225, + 455 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Video question answering. Recent research has shown that while video-based VLMs can achieve state-of-the-art performance, their answers are sensitive to object size, positioning, and speed [1, 28]. Moreover, when answering temporal and spatial questions, VLMs rely on textual biases to \"guess\" answers rather than performing genuine understanding and reasoning over visual-text information [32].", + "bbox": [ + 94, + 460, + 480, + 566 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To improve the robustness and interpretability of VLMs, one line of research enriches VLMs with visual grounding functionality during QA, which enables VLMs to localize relevant video moments [18, 27, 30] or key frames [15, 22] to support answers. However, while these methods localize visual evidence, the process by which VLMs use it to deduce answers remains opaque. Another approach leverages external LLMs as reasoners or agents to enhance interpretability in textual modality. For instance, LLoVi [31] converts VQA to a text-based QA task via video captioning, then prompts an LLM to provide answers. Similarly, VideoAgent [21] uses an LLM to recursively determine if the current frames can answer the given question based on their textual descriptions. However, these methods heavily rely on the reasoning capabilities of LLMs. Like VLMs, the LLM reasoning process remains a black box, and hallucinations are common. Recently, TV-trees [19] attempted to perform explicit reasoning over both visual and textual modalities using a neuro-symbolic system. However, their work assumes video transcripts are explicitly provided to evaluate answers, thus avoiding the complexity of grounding hypotheses into video content. Instead, we contribute a", + "bbox": [ + 93, + 569, + 480, + 898 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "general framework for explicit reasoning in commonsense VQA, fueled by a grounding component that aligns question components with video fragments.", + "bbox": [ + 517, + 92, + 903, + 136 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Beyond methodology, some works focus on providing fair and comprehensive VLM evaluations in VQA tasks by creating new benchmarks [3, 6, 9]. These benchmarks contain videos with diverse scenarios and durations, with carefully crafted questions and options designed to prevent textual shortcuts that VLMs might exploit. Video-specific questions (e.g., compositional action reasoning) [1], which require insights beyond textual associations, are included to test commonsense reasoning in VLMs. Addressing concerns of remaining biases in such benchmarks [6], we contribute an LLM-based answer-set de-biasing procedure to ensure that VQA benchmarks [11, 26] are adequate to evaluate reasoning in videos rather than spurious correlations.", + "bbox": [ + 517, + 138, + 903, + 333 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Systematic language reasoning. As LLMs demonstrate great potential in reasoning, there has been considerable interest in using LLMs to generate systematic explanations to support their answers. The series of Chain-of-Thought prompting [2, 23, 29] encourages LLMs to think step-by-step to perform explicit multi-hop reasoning, providing free-form reasoning steps before arriving at an answer. However, such implicit explanations are not grounded in external knowledge or evidence, which may lead to unverifiable and unfaithful reasoning. Since the development of EntailmentBank [4], research has increasingly focused on constructing explanation trees [20, 24] and graphs [8], encouraging models to generate step-wise entailment proofs of a statement using a set of supporting facts. Entailer [20] introduced this systematic explanation framework into language-based multiple-choice QA, performing explicit reasoning by generating entailment trees grounded in the model's internal beliefs. REFLEX [8] extends the entailment tree to form a belief graph for QA models, aiming to address consistency issues by intervening in the intermediate reasoning steps. Instead of grounding facts in predefined rules or model beliefs, NELLIE [24] adopts Prolog-based inference engines and external natural language corpora to build entailment trees as explainable reasoning for multiple-choice QA tasks. While such techniques for natural language processing inspire our framework, we generalize entailment trees to VQA, contributing a novel grounding method that aligns entailment trees with video fragments.", + "bbox": [ + 517, + 335, + 903, + 757 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Video-grounded entailment tree reasoning", + "text_level": 1, + "bbox": [ + 516, + 775, + 888, + 791 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This paper devises a novel explainable framework for grounded commonsense VQA. It derives the answers through systematic reasoning over video-text information with entailment trees. Specifically, in the entailment tree (Fig. 2a), each candidate answer is decomposed into statements that entail the answer, explaining why each answer could be plausible. These statements are then grounded", + "bbox": [ + 517, + 795, + 903, + 898 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "3263", + "bbox": [ + 483, + 946, + 513, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "in relevant visual evidence from the video to prove or refute them (Fig. 2b). While entailment trees in natural language processing are constructed based on a model's internal knowledge or corpora [8, 20, 24], we ground entailment trees into video fragments. Finally, backtracking through the entailment tree leads to systematic reasoning over the statements (Fig. 3). Thus, answers can be deduced by a systematic structure with explicit reasoning paths and explanations rather than relying on opaque, black-box models.", + "bbox": [ + 89, + 90, + 480, + 227 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Entailment tree construction", + "text_level": 1, + "bbox": [ + 89, + 238, + 346, + 252 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Initial statement generation. Given a question and its answer candidates, we first convert each question-answer pair into a declarative sentence that preserves the semantic meaning of the original QA pair. As a result, an $N$ -way multiple-choice QA problem produces a set of statements, denoted as $\\mathcal{D} = d_{1},\\ldots ,d_{N}$ . For example, the two-way question \"What did the boy in white do after he first took the balloon? (A) resting on a chair (B) carries it toward the hula hoop\" is transformed into: $\\mathcal{D}:\\{d_1 =$ \"The boy in white resting on a chair after first taking the balloon.\", $d_{2} =$ \"The boy in white carries it toward the hula hoop after first taking the balloon.\" $\\}$ . Thus, selecting the best answer equals identifying the correct statement for a given video.", + "bbox": [ + 89, + 253, + 482, + 449 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recursive statement decomposition. For each initial statement in $\\mathcal{D}$ , we generate two sub-statements as proofs that support the statement: Statement $\\Leftarrow$ Sub-statement1, Sub-statement2. The statement is True if and only if both its sub-statements are proved to be True, i.e., the sub-statements entail the statement. Proving the original statements is thus translated into proving two simpler sub-statements. This procedure is recursive: the sub-statements can be further decomposed into further sub-statements that entail them. Therefore, to construct an entailment tree, we recursively decompose these sub-statements as new statements in the next tree layer until reaching the maximum depth or meeting the stop criterion. Fig. 2(a) presents an example of entailment tree generation.", + "bbox": [ + 89, + 450, + 482, + 662 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We leverage LLM prompting for both the initial statement generation and the statement decomposition, as these are linguistic tasks (see implementation details).", + "bbox": [ + 89, + 662, + 482, + 708 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Video-language entailment verification", + "text_level": 1, + "bbox": [ + 89, + 718, + 423, + 734 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given the entailment tree, the framework then verifies language statements based on the grounded video content as evidence. Specifically, each statement in the entailment tree must be proven or refuted by analyzing the video. A straightforward solution is to encode the whole video to collect information that can be used to verify the statement. However, the critical visual evidence that accurately verifies a statement tends to exist in a local moment instead of the whole video. Therefore, we develop a novel video grounding that guides the verification process to the moments with relevant visual evidence.", + "bbox": [ + 89, + 734, + 482, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Question-aware video captioning. Given a video, we convert its visual information into detailed textual information. Specifically, we input video frames into a VLM-based captioner $\\mathrm{Cap}(\\cdot)$ to obtain a caption $c_{i} = \\mathrm{Cap}(f_{i})$ for each frame. However, captioning frames individually can overlook essential details or introduce irrelevant information for VQA. In commonsense VQA, questions often focus on specific facts already observed in the video. For example, a typical temporal reasoning question is \"What happened before/after Event-A?\" where Event-A refers to a fact statement about an event in the video. The fact referenced by the question can be leveraged to guide video understanding. To this end, we first extract the anchor fact indicated by the question and provide it to $\\mathrm{Cap}(\\cdot)$ as prior knowledge, encouraging the generation of relevant captions. Moreover, captions from all previous frames are also provided for each current frame to ensure $\\mathrm{Cap}(\\cdot)$ captures the temporal context from the past. This process is formulated as:", + "bbox": [ + 511, + 90, + 903, + 362 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nc _ {i} = \\operatorname {C a p} \\left(f _ {i} \\mid F, \\left(c _ {1}, \\dots , c _ {i - 1}\\right)\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 596, + 376, + 903, + 393 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $F$ indicates the fact statement.", + "bbox": [ + 511, + 405, + 756, + 419 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Video evidence grounding. For commonsense VQA, depending on how the question reasons around the fact statement, the necessary evidence for answers can be gathered from specific video moments. For instance, in the case of temporal reasoning (e.g., before or after questions), the answer should be inferred from moments occurring either before or after the time of the relevant fact. Following this intuition, we design a two-step evidence-grounding strategy to localize the critical moments for answering.", + "bbox": [ + 511, + 420, + 903, + 556 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "First, given the frame-wise captions, we retrieve a keyframe deemed most relevant to the fact statement, which we refer to as the anchor frame. A straightforward retrieval approach would involve comparing each $c_i$ with the fact description using specific metrics to identify the anchor frame. However, we enhance retrieval accuracy by adopting a structured semantic retrieval strategy. Specifically, the textual descriptions of each frame and fact statement are converted into structured triplets. These triplets capture the attributes and relationships of objects in each frame through structured semantics. As shown in Fig. 2(a), rather than directly comparing raw textual descriptions of frames and fact statements, we use these triplets for retrieval. Inspired by the success of using LLMs for retrieval tasks, we prompt an LLM to conduct anchor frame retrieval using the triplets of the fact statement as the query. The LLM then identifies and returns the most relevant frame ID, i.e., its timestamp.", + "bbox": [ + 511, + 556, + 905, + 814 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nt _ {\\text {a n c h o r}} = \\operatorname {R t v} \\left(c _ {i}, F\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 637, + 825, + 903, + 843 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $t_{\\mathrm{anchor}}$ is the time stamp of the anchor frame, $\\mathrm{Rtv}(\\cdot)$ denotes the retrieval process. Second, we determine the final moment where we should look centered on the $t_{\\mathrm{anchor}}$ , to", + "bbox": [ + 511, + 854, + 903, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3264", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/9f3359d2c101bb0a8e86b760c57a1259094279cacb3e42086b24c0b83e63a8be.jpg", + "image_caption": [ + "Figure 2. Overview of our framework. (a) The generation of the entailment tree, where statements are recursively decomposed until the tree reaches its max depth or meets the stop criterion. (b) The process of video-language entailment verification: the input video is first converted into textual descriptions. Each caption is then parsed into structured semantics. Given the fact statement as a query, we retrieve the anchor frame. Then, based on the temporal or causal navigation indicated by questions, the visual evidence moment can be grounded." + ], + "image_footnote": [], + "bbox": [ + 94, + 88, + 903, + 433 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "incorporate the temporal relations present in the question. Therefore, based on the anchor frame, the navigation for the moment is selected from \"look ahead, look behind, look around\" by considering the question:", + "bbox": [ + 89, + 516, + 483, + 578 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {M} = \\operatorname {G n d} \\left(t _ {\\text {a n c h o r}} \\mid \\mathcal {Q}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 210, + 590, + 480, + 609 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\operatorname{Gnd}(\\cdot)$ is the grounding process and $\\mathcal{M}$ denotes the grounded continuous interval in the video. Then, frames are resampled from the video within $M$ and used as visual evidence proving or refuting entailment tree statements.", + "bbox": [ + 89, + 622, + 483, + 681 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Visual-text statement prover. Given grounded visual evidence $\\mathcal{M}$ of the video, statements are estimated to be true or false. Specifically, we employ a VLM as the statement prover, denoted as $\\mathrm{Prv}(\\cdot)$ , to evaluate each statement within the tree by probing VLM's internal belief on this statement. Each statement will be transformed into a binary QA task, with possible options being True or False. We then directly probe the $\\mathrm{Prv}(\\cdot)$ with the binary QA prompt and use the next token prediction probabilities of the words to elicit the model's belief. We normalize the prediction logits of the two options to get the confidence score of that statement. The above process is formulated as:", + "bbox": [ + 89, + 684, + 483, + 864 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ns _ {d} = \\Pr \\operatorname {r v} (\\mathcal {M}, h), s \\in [ 0, 1 ], \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 187, + 880, + 482, + 897 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/d3b28f2f6252dbcf4a3b151041a2110b13482ff2e7af4b3cef3f8eba367f972a.jpg", + "image_caption": [ + "Figure 3. Illustration of dynamic tree generation and backtrace. In Step-3, when the proof score of the left statement calculated from its child nodes is less than its direct score $(0.63 < 0.8)$ , its decomposition is pruned and stops." + ], + "image_footnote": [], + "bbox": [ + 513, + 512, + 903, + 609 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{M}$ is the grounded moment and $h$ denotes the statement that needs to be verified.", + "bbox": [ + 511, + 690, + 906, + 720 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Dynamic entailment tree expansion", + "text_level": 1, + "bbox": [ + 511, + 733, + 821, + 748 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "So far, we have performed statement decomposition recursively to construct an entailment tree with pre-defined depth. However, not all statements need to be verified recursively, especially those easily determined to be true or false by VLMs. Moreover, as the depth increases, some statement sentences are atomic and directly verifiable. Thus, to improve the efficiency of the reasoning process, we further adopt a strategy to expand the entailment tree dynamically. Specifically, each statement $d$ is tied with two confidence scores provided by the $\\operatorname{Prv}(\\cdot)$ :", + "bbox": [ + 511, + 750, + 906, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "3265", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/de93e1108f78b4b924bb625a222dfe4a9d1eb3d32f5f2397f848d19e9d5719f3.jpg", + "image_caption": [ + "Figure 4. Illustration of commonsense bias in video question answering. The example is selected from the NExT-QA dataset." + ], + "image_footnote": [], + "bbox": [ + 96, + 88, + 480, + 349 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "(1) The direct score $s_d$ , which indicates the belief of $\\operatorname{Prv}(\\cdot)$ model in $d$ .", + "(2) The proof score $s_p$ , denoting how confidently the model can prove $d$ , is calculated by multiplying the scores of its direct sub-statements." + ], + "bbox": [ + 78, + 412, + 482, + 486 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For a statement $d$ , the goal of decomposition is to establish a more reliable and convincing proof path than merely evaluating whether $d$ is true by VLMs. If the decomposition-based reasoning can prove $d$ with higher confidence than its direct score, the overall confidence for statement $d$ should increase. Otherwise, the decomposition should be disregarded. Thus, in the dynamic tree expansion, if decomposition does not enhance a statement's score, it is pruned, and that statement node becomes a leaf in the entailment tree.", + "bbox": [ + 89, + 488, + 483, + 622 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Fig. 3 presents a toy example. This criterion ensures that only beneficial decompositions are retained, significantly enhancing the tree reasoning process's efficiency.", + "bbox": [ + 89, + 625, + 483, + 670 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Reasoning over the entailment tree", + "text_level": 1, + "bbox": [ + 89, + 681, + 393, + 696 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Finally, we perform a backtrace through the entailment tree to calculate the confidence score of each top statement. Specifically, the final score for each statement is produced by comparing its direct score $s_d$ and proof score $s_p$ , i.e., $s = \\max(s_d, s_p)$ during backtrace (as shown in Fig. 3). The overall framework selects the answer corresponding to the statement at the top layer with top-scoring proof.", + "bbox": [ + 89, + 696, + 482, + 804 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. De-biasing commonsense VQA answer sets", + "text_level": 1, + "bbox": [ + 89, + 820, + 473, + 837 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To demonstrate the reasoning ability of video-grounded entailment trees, it is essential to evaluate using commonsense VQA benchmarks that enforce model reasoning. Recent work [13, 18, 27] has provided evidence that shortcuts", + "bbox": [ + 89, + 840, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/e3a2bf99076e33621d676eb4bb7041017cbde2dfe28a52b3a509b12f8f37a998.jpg", + "image_caption": [ + "Figure 5. Prompt used for rewriting answers on NExT-QA." + ], + "image_footnote": [], + "bbox": [ + 524, + 94, + 895, + 280 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "are present in VQA datasets which enables VLMs to solve these tasks based on textual associations rather than video-grounded reasoning. While VQA benchmarks increasingly focus on commonsense reasoning skills, such as temporal (e.g. after, before) or causal (how, why, what if) relationships in video content, reasoning shortcuts affect the validity of their evaluation. This is illustrated in Fig. 4 (top), where the correct answer (D) is much more relevant to the question and also aligns best with real-world expectations. Consequently, a VLM (VideoLLaVA [14] used in this example) can answer this question correctly by leveraging such associations and without analyzing the video content. Meanwhile, replacing the answer set distractors with other commonsensical answer candidates, as illustrated in Fig. 4 (bottom), makes this task challenging for VLMs. Here, a VLM switches its answer incorrectly to option (C), which confirms the impact of commonsense associations and the lack of grounded reasoning by these models.", + "bbox": [ + 511, + 356, + 906, + 628 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To this end, we devise a de-biasing procedure that mitigates reasoning shortcuts in commonsense VQA answer sets. Our de-biasing procedure transforms multiple-choice VQA benchmarks (e.g., NExT-QA) by rewriting their answer distractors while keeping their question and ground-truth answer intact. We prompt an LLM (LLaMA-3) to implement the rewriting procedure for each original QA set. Fig. 5 shows the detailed prompt we used for LLaMA-3 on NExT-QA dataset. This procedure ensures that (1) the answers cannot be easily derived from the QA set associations and (2) the answer remains consistent with the original QA pair. Thus, our procedure enables the scalable construction of de-biased QA sets by leveraging the commonsense associations in LLMs. The next section, focusing on experimental evaluation, analyzes the application of de-biasing to various datasets and its impact on the performance of VLMs, with and without entailment tree reasoning.", + "bbox": [ + 511, + 643, + 908, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "3266", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 89, + 90, + 222, + 106 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Experimental setup", + "text_level": 1, + "bbox": [ + 89, + 109, + 277, + 125 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. We test our framework on three VQA benchmarks: (1) NExT-QA [26], a VQA benchmark for causal and temporal reasoning. (2) IntentQA [11], which focuses on video intent reasoning from both causal and temporal aspects. (3) Video-MME [6], which is a recently proposed comprehensive evaluation benchmark for video analysis; we use its \"short-term\" split (video length $< 2$ mins) and 4 question types (temporal, spatial, action, and object reasoning) highly related to commonsense reasoning are selected.", + "bbox": [ + 89, + 126, + 480, + 261 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation. We report model performances on our rewritten test set for each dataset and its original test set. We evaluate our framework on all datasets under the multiple-choice QA setting, using a standard accuracy metric.", + "bbox": [ + 89, + 263, + 480, + 324 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines. Our baselines represent three categories:", + "bbox": [ + 89, + 325, + 431, + 339 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Video-based VLMs: Video-based VLMs are widely used for VQA tasks, so we include Video-LLaVA [14], VideoChat2 [12], and VideoLLaMA [33]. To test the effectiveness of our framework, we integrate these VLMs by replacing our $\\mathsf{Prv}(\\cdot)$ with specific VLM models.", + "- Image-based VLMs: We include BLIP-2 [10] and LLaVA-1.5 [17] as Image-LLM baselines.", + "- State-of-the-art VQA approaches: Recent works in VQA, such as VideoTree [22], VideoAgent [21], and LLoVi [31], are included as strong baselines." + ], + "bbox": [ + 89, + 340, + 480, + 491 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation details. Our entire framework is training-and human annotation-free. We use LLaMA-3-8B [5] to handle basic functionalities, including (1) converting original QA into declarative statements, (2) statement decomposition, (3) structured semantic extraction and retrieval, and (4) guiding evidence grounding. Detailed prompts for each functionality are provided in the supplementary material. For frame-wise captioning across all datasets, we use LLaVA-1.5 [17] as our default captioner. When comparing with state-of-the-art VQA methods (e.g., VideoAgent), we follow their setup by replacing the captioner with the stronger CogAgent [7] model for fair comparison. When integrating our framework with VLMs, the VLM itself serves as the Prover, $\\mathsf{Prv}(\\cdot)$ . For video-LLMs, frames are uniformly sampled from the grounded video moment to meet their input requirements (VideoChat2: 16 frames; VideoLLaVA & VideoLLaMA: 8 frames). For image-language models like LLaVA, we sample 8 frames from the grounded moment and process them individually; the final confidence score is obtained by averaging scores across frames. During dynamic tree generation, we also set a max depth of 5 for the overall entailment tree to improve the efficiency.", + "bbox": [ + 89, + 493, + 482, + 825 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Main results", + "text_level": 1, + "bbox": [ + 89, + 838, + 223, + 852 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Benefit for image and video-based VLMs. Tab. 1 summarizes the results of our method compared to baselines. Integrating entailment tree reasoning brings consistent im", + "bbox": [ + 89, + 854, + 480, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "provement across the video- and image-based VLMs for all datasets (1-4% on average). This finding includes the recently proposed benchmark VideoMME, which poses much more challenging videos and questions. The benefits are particularly regular for temporal reasoning (improvement in 14 out of 15 cases), which illustrates how our explicit reasoning process enhances temporal commonsense QA. Image-based VLMs, which initially lack temporal modeling capabilities, perform poorly when directly applied to video QA tasks. However, our framework provides a significant performance boost for these models up to $8\\%$ for LLAVA-1.5. By reasoning over multiple sub-problems rather than tackling the entire complex question simultaneously, our reasoning method makes the task more manageable for both video- and image-based VLMs.", + "bbox": [ + 511, + 90, + 903, + 316 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Results on de-biased QA sets. As shown in Tab. 2, all models experience considerable performance drops on the de-biased set of the same VQA dataset, which aligns with our observation that current VLMs often rely on textual bias in commonsense reasoning tasks. Notably, video-based VLMs show an $8\\% -10\\%$ decrease in the de-biased set even though the question and the correct answer remain unchanged. In contrast, our proposed framework, which derives answers through an explicit reasoning process based on specific visual evidence, demonstrates much greater robustness on the de-biased set. The improvement brought by our framework on the de-biased set is even higher than on the original test sets. In turn, our framework compensates for the performance loss of the VLMs on the de-biased set. This analysis underscores our framework's potential to mitigate textual bias in commonsense reasoning. Furthermore, the performance differences between the original and de-biased QA sets highlight VQA benchmarks' limitations in evaluating VLMs' true reasoning abilities.", + "bbox": [ + 511, + 316, + 903, + 603 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Comparison with state-of-the-art. Tab. 3 compares our framework's results on the de-biased sets to state-of-the-art VQA approaches. The table shows that, next to the consistent benefit our framework provides to various VLMs, it is also competitive with state-of-the-art VQA methods. When applying an advanced captioner and reasoner that aligns with VideoAgent and VideoTree, our framework yields new state-of-the-art results in some cases. In particular, our framework performs best on temporal reasoning questions for all three benchmarks and outperforms all methods on the IntentQA dataset. Importantly, our method reaches such competitive performance despite using $257 \\times$ fewer parameters for its reasoning compared to state-of-the-art methods.", + "bbox": [ + 511, + 604, + 903, + 800 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.3. Ablation studies", + "text_level": 1, + "bbox": [ + 511, + 809, + 673, + 823 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Ablation experiments are conducted using VideoLLaVA's baseline with our framework. Results are reported on the test set of the NExT-QA dataset.", + "bbox": [ + 511, + 825, + 903, + 869 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Impact of LLMs for statement decomposition. Entailment generation in our framework relies on prompting an", + "bbox": [ + 511, + 869, + 903, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "3267", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/b3d2e3be807e912e7d73a85364ee50e63248ed1c6858badd95a5968c92474e60.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelNExT-QAIntentQAVideoMMEAvg
TemporalCausalTemporalCausalTemporalSpatialActionObject
Image-based VLMsBLIP-2 [10]38.336.143.848.625.426.924.228.634.0
+Ours45.341.848.952.530.527.427.728.837.9
LLaVA-1.5 [17]37.840.745.850.031.133.627.729.837.1
+Ours45.647.948.454.736.736.831.430.741.5
Video-based VLMsVideoChat2 [12]56.962.160.463.250.352.449.550.155.6
+Ours57.861.662.363.852.851.551.350.056.4
VideoLLaVA [14]56.060.453.960.747.644.346.249.752.3
+Ours58.362.757.661.848.346.149.850.354.2
VideoLLaMA [33]55.460.255.156.744.847.244.749.351.7
+Ours58.160.454.558.947.447.848.649.153.1
", + "bbox": [ + 91, + 88, + 906, + 263 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/8ce1c9037fa7df55b32339d610051b204d6c291308b6df4a6c4e159d97f27d78.jpg", + "table_caption": [ + "Table 1. Impact on image and video-based VLMs on the original NExT-QA, IntentQA, and VideoMME test sets. Our framework increases accuracy of all video- and image-based VLMs by $1 - 4\\%$ on average across all data partitions. Temporal and action partitions benefit most." + ], + "table_footnote": [], + "table_body": "
ModelBLIP-2+OursLLaVA+OursVideo Chat2+OursVideo LLaVA+OursVideo LLaMA+Ours
NExT-QAOriginal37.243.639.346.859.559.758.260.557.859.3
Rewritten33.539.834.844.945.449.051.155.441.447.0
Intent-QAOriginal46.250.747.951.661.863.157.359.755.956.7
Rewritten38.245.542.748.652.655.750.554.746.350.0
Avg38.844.941.248.054.856.954.357.650.453.3
", + "bbox": [ + 91, + 310, + 906, + 422 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Results on de-biased QA sets. Video-based VLMs show significant decreases in the rewritten de-biased set. In contrast, our framework demonstrates much greater robustness on the rewritten set.", + "bbox": [ + 89, + 426, + 906, + 455 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "external LLM to recursively decompose statements (cf. Sec. 3.1), which is crucial in guiding reasoning paths. Consequently, we tested various LLMs for entailment tree generation, including open-source models (LLama-3 and Mistral) of different sizes and proprietary LLMs (GPT-4 and Gemini-1.5). The results are summarized in Tab. 4. As expected, the proprietary model GPT-4, known for its strong step-by-step reasoning capabilities, delivers the best performance across all settings. Scaling up LLaMA-3 to 70B offers improvements over the 8B model, though with a notable increase in inference time. As the overall performance difference between models is within $1\\%$ , we select LLaMA-3-8B as the default for integrating our framework into VLMs due to its free availability and efficiency.", + "bbox": [ + 88, + 481, + 482, + 694 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Ablation on grounding components. Next, we test the effectiveness of each component in our grounding module (Sec. 3.2). The results, summarized in Tab. 5, indicate that both fact-conditional captioning and structure-guided retrieval enhance overall performance by improving grounding accuracy. However, using only structure-guided retrieval results in a slight performance drop, possibly because $\\mathsf{Cap}(\\cdot)$ introduces irrelevant semantic information that doesn't align with the question's focus, and the structured representation can make identifying anchor frames more challenging. In contrast, fact-conditional captioning alone yields substantial improvement, demonstrating that this straightforward approach can yield an effective", + "bbox": [ + 88, + 704, + 482, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "and more controllable textual description for videos by conditioning on prior knowledge or relevant facts.", + "bbox": [ + 511, + 481, + 903, + 512 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Impact of length of video frames. We further ablate the impact of input video frame length in our framework to determine the optimal number of frame-wise captions to generate per video. The results, summarized in Tab. 6, show that ideal performance is achieved only when sampling a sufficient number of frames (at least 16 for NExTQA). When fewer frames are used (e.g., 4 or 8), key anchor frames may be missed, reducing the accuracy of grounded visual evidence. Additionally, while increasing the frame count to 32 yields the best performance, it also increases the calls required for $\\mathrm{Cap}(\\cdot)$ to generate frame-wise captions. Balancing efficiency with performance gains, we set 24 frames as the default in our implementation.", + "bbox": [ + 511, + 517, + 906, + 714 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effectiveness of evidence grounding. Our method grounds relevant video fragments to support statements in the entailment tree (Sec. 3.2). To validate its effectiveness, we compare it with two other variations as sources of visual evidence: (1) without evidence grounding, using the full video as evidence, and (2) upper-bound results: manually annotated temporal boundaries provided in the NExT-GQA dataset, indicating where the QA models should focus when producing correct answers. The results are shown in Tab. 7. Compared to the baseline, our video-grounded method provides consistent improvements across the original and debiased sets. The improvement is more apparent in the de", + "bbox": [ + 511, + 719, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "3268", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/f570a25d65df90b4c64b191ff41b84de1be208a666c90923ec94f82e7852fc1b.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodModel ReasonerNExT-QA*IntentQA*VideoMME
TemporalCausalTemporalCausalTemporalSpatialActionObject
VideoAgent [21]GPT-4 (1.8T)58.266.660.461.0----
VideoTree [22]GPT-4 (1.8T)60.266.456.760.155.754.354.252.6
LLoVi [31]GPT-4 (1.8T)53.160.858.661.852.255.351.850.8
OursVideoLLAVA (7B)60.865.961.062.655.953.854.050.8
", + "bbox": [ + 91, + 88, + 906, + 180 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/0c3ccde910e7dbe8814485cc03c34975697d2d70e851901636368b9addca4541.jpg", + "table_caption": [ + "Table 3. Comparison with state-of-the-art. Results for NExT-QA and IntentQA are reported under the de-biased set (the results on the original sets are similar; we provide them in the Appendix). The 'Reasoner' in these approaches is similar to the \"Prover\" in our framework. The captioner for all methods is CogAgent [7]. Despite other methods relying on much stronger reasoning models, our approach yields competitive performance (four state-of-the-art results) and high parameter efficiency ( $257 \\times$ fewer than GPT-4 reasoners)." + ], + "table_footnote": [], + "table_body": "
Model classLLMNExT-QA
OriginalRewritten
Open-sourceMistral-7B60.055.6
LLaMA-3-8B60.555.4
LLaMA-3-70B61.355.9
ProprietaryGemini-1.5-Pro61.155.2
GPT-461.656.1
", + "bbox": [ + 93, + 263, + 480, + 375 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/b4a18976df1527aebb95f866e5602612294cee2040bc0f8a0790348def79dc8b.jpg", + "table_caption": [ + "Table 4. Impact of LLMs for statement decomposition. The open-source and proprietary models are ordered ascendingly by size. Larger models, especially GPT-4, are best at decomposition, but the smaller models (e.g., LlaMa-3-8B) come close." + ], + "table_footnote": [], + "table_body": "
ComponentsNExT-QA
Fact-conditioned captioningStructure-based retrievalOriginalRewritten
56.249.6
✓59.553.3
✓58.452.7
✓✓60.555.4
", + "bbox": [ + 94, + 452, + 482, + 556 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/5feb288937de3b56dc9bf07f5d2f9b49c3d82851a6a64d195c69ced4547b6ad6.jpg", + "table_caption": [ + "Table 5. Ablation on grounding components, showing that both fact-conditional captioning and structure-guided retrieval enhance overall performance by improving grounding accuracy." + ], + "table_footnote": [], + "table_body": "
Acc\n(NExT-QA)Frame number
48162432
Original57.759.360.560.761.0
Rewritten51.953.455.455.455.7
", + "bbox": [ + 116, + 619, + 459, + 686 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "biased set, where the answer options are more semantically similar and require more precise, discriminative visual evidence. Using the ground-truth fragment can further boost our approach, suggesting that enhancing grounding accuracy could further improve our framework.", + "bbox": [ + 89, + 747, + 482, + 823 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effectiveness of dynamic tree expansion. The depth of the entailment tree determines the granularity of reasoning (Sec. 3.3). This ablation analyzes how tree depth impacts overall performance and compares a fixed-depth approach with our dynamic tree generation strategy. Increases", + "bbox": [ + 89, + 824, + 483, + 902 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/25cb4bf3eb3ffdb08bed047d963fe981531b8ab29a467f87ccf9fdc4a845ef0f.jpg", + "table_caption": [ + "Table 6. Impact of video frame amount. Strong performance requires a sufficiently high frame number (over 16 for NExT-QA)." + ], + "table_footnote": [], + "table_body": "
Video fragmentFullGrounded (ours)GT
NExT-QAOriginal58.360.561.8
Rewritten51.755.456.9
", + "bbox": [ + 514, + 263, + 905, + 314 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/2b25ba2ccfbd2864f810ede221197abc2ceaae122bbfe00fa48b990d8b38b3be.jpg", + "table_caption": [ + "Table 7. Effectiveness of evidence grounding. Our videotegrounded method yields clear improvement over using the full video. More precise grounding can further enhance our accuracy." + ], + "table_footnote": [], + "table_body": "
StrategyStatic (Depth=)Dynamic
2345
NExT-QAOriginal58.859.260.260.360.5
Rewritten52.053.455.655.355.4
", + "bbox": [ + 514, + 375, + 905, + 431 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 8. Effectiveness of dynamic tree expansion. It yields superior accuracy while increasing reasoning efficiency.", + "bbox": [ + 511, + 436, + 905, + 465 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ing the depth of reasoning yields significant improvements, as complex, long statements are broken down into concise sub-statements that VLMs can understand more effectively. However, extending reasoning beyond the 4th layer offers diminishing returns; for NExT-QA, the original statements' complexity constrains the task, and some 5th-layer sub-statements become overly simplistic and less effective for reasoning. This finding highlights the necessity of our dynamic strategy. Applying the dynamic tree expansion strategy, we can see that the performance outperforms the fixed-depth paradigm. In the meantime, the dynamic strategy increases the reasoning efficiency over the entailment tree, more details about efficiency comparison can be found in our supplementary material.", + "bbox": [ + 511, + 489, + 906, + 702 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 715, + 633, + 729 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper proposed the first video-grounded entailment tree framework for VQA. Moreover, we also contributed a de-biasing procedure to avoid spurious correlations during evaluation and applied it to enhance representative benchmarks. Extensive experiments with five video- and image-based VLMs demonstrate consistent benefits of our method on these benchmarks. Besides, our proposed framework performs on par with state-of-the-art video reasoning methods despite using $257 \\times$ fewer parameters. While de-biasing hurts VLM accuracy, our framework regains the accuracy losses and is competitive with state-of-the-art VQA methods.", + "bbox": [ + 511, + 733, + 906, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "3269", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Piyush Bagad, Makarand Tapaswi, and Cees GM Snoek. Test of time: Instilling video-language models with a sense of time. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2503-2516, 2023. 2", + "[2] Maciej Besta, Nils Blach, Ales Kubicek, Robert Gerstenberger, Michal Podstawski, Lukas Gianinazzi, Joanna Gajda, Tomasz Lehmann, Hubert Niewiadomski, Piotr Nczyk, et al. Graph of thoughts: Solving elaborate problems with large language models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 17682-17690, 2024. 2", + "[3] Tieyuan Chen, Huabin Liu, Tianyao He, Yihang Chen, Chaofan Gan, Xiao Ma, Cheng Zhong, Yang Zhang, Yingxue Wang, Hui Lin, et al. Mecd: Unlocking multi-event causal discovery in video reasoning. arXiv preprint arXiv:2409.17647, 2024. 2", + "[4] Bhavana Dalvi, Peter Jansen, Oyvind Tafjord, Zhengnan Xie, Hannah Smith, Leighanna Pipatanangkura, and Peter Clark. Explaining answers with entailment trees. arXiv preprint arXiv:2104.08661, 2021. 1, 2", + "[5] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024.6", + "[6] Chaoyou Fu, Yuhan Dai, Yongdong Luo, Lei Li, Shuhuai Ren, Renrui Zhang, Zihan Wang, Chenyu Zhou, Yunhang Shen, Mengdan Zhang, et al. Video-mme: The first-ever comprehensive evaluation benchmark of multi-modal llms in video analysis. arXiv preprint arXiv:2405.21075, 2024. 2, 6", + "[7] Wenyi Hong, Weihan Wang, Qingsong Lv, Jiazheng Xu, Wenmeng Yu, Junhui Ji, Yan Wang, Zihan Wang, Yuxiao Dong, Ming Ding, et al. Cogagent: A visual language model for gui agents. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14281-14290, 2024. 6, 8", + "[8] Nora Kassner, Oyvind Tafjord, Ashish Sabharwal, Kyle Richardson, Hinrich Schuetze, and Peter Clark. Language models with rationality. arXiv preprint arXiv:2305.14250, 2023. 2, 3", + "[9] Muhammad Uzair Khattak, Muhammad Ferjad Naeem, Jameel Hassan, Muzammal Naseer, Federico Tombari, Fahad Shahbaz Khan, and Salman Khan. Complex video reasoning and robustness evaluation suite for video-lmms. arXiv preprint arXiv:2405.03690, 2024. 2", + "[10] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730–19742. PMLR, 2023. 1, 6, 7", + "[11] Jiapeng Li, Ping Wei, Wenjuan Han, and Lifeng Fan. Intentqa: Context-aware video intent reasoning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11963-11974, 2023. 1, 2, 6", + "[12] Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Yi Liu, Zun Wang, Jilan Xu, Guo Chen, Ping Luo, et al." + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Mvbench: A comprehensive multi-modal video understanding benchmark. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22195-22206, 2024. 1, 6, 7", + "[13] Yicong Li, Xiang Wang, Junbin Xiao, Wei Ji, and Tat-Seng Chua. Invariant grounding for video question answering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2928-2937, 2022. 5", + "[14] Bin Lin, Yang Ye, Bin Zhu, Jiaxi Cui, Munan Ning, Peng Jin, and Li Yuan. Video-llava: Learning united visual representation by alignment before projection. arXiv preprint arXiv:2311.10122, 2023. 1, 5, 6, 7", + "[15] Huabin Liu, Weixian Lv, John See, and Weiyao Lin. Task-adaptive spatial-temporal video sampler for few-shot action recognition. In Proceedings of the 30th ACM International Conference on Multimedia, pages 6230-6240, 2022. 2", + "[16] Huabin Liu, Weiyao Lin, Tieyuan Chen, Yuxi Li, Shuyuan Li, and John See. Few-shot action recognition via intra- and inter-video information maximization. arXiv preprint arXiv:2305.06114, 2023. 1", + "[17] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26296-26306, 2024. 1, 6, 7", + "[18] Huabin Liu, Xiao Ma, Cheng Zhong, Yang Zhang, and Weiyao Lin. Timecraft: Navigate weakly-supervised temporal grounded video question answering via bi-directional reasoning. In European Conference on Computer Vision, pages 92-107. Springer, 2025. 2, 5", + "[19] Kate Sanders, Nathaniel Weir, and Benjamin Van Durme. Tv-trees: Multimodal entailment trees for neuro-symbolic video reasoning. arXiv preprint arXiv:2402.19467, 2024. 1, 2", + "[20] Oyvind Tafjord, Bhavana Dalvi Mishra, and Peter Clark. Entailer: Answering questions with faithful and truthful chains of reasoning. arXiv preprint arXiv:2210.12217, 2022. 2, 3", + "[21] Xiaohan Wang, Yuhui Zhang, Orr Zohar, and Serena Yeung-Levy. Videoagent: Long-form video understanding with large language model as agent. arXiv preprint arXiv:2403.10517, 2024. 2, 6, 8", + "[22] Ziyang Wang, Shoubin Yu, Elias Stengel-Eskin, Jaehong Yoon, Feng Cheng, Gedas Bertasius, and Mohit Bansal. Videotree: Adaptive tree-based video representation for llm reasoning on long videos. arXiv preprint arXiv:2405.19209, 2024. 2, 6, 8", + "[23] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824-24837, 2022. 2", + "[24] Nathaniel Weir and Benjamin Van Durme. Dynamic generation of grounded logical explanations in a neuro-symbolic expert system. arXiv preprint arXiv:2209.07662, 2022. 2, 3", + "[25] Hao Wu, Huabin Liu, Yu Qiao, and Xiao Sun. Dibs: Enhancing dense video captioning with unlabeled videos via pseudo boundary enrichment and online refinement. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18699-18708, 2024. 1" + ], + "bbox": [ + 516, + 92, + 905, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "3270", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[26] Junbin Xiao, Xindi Shang, Angela Yao, and Tat-Seng Chua. Next-qa: Next phase of question-answering to explaining temporal actions. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9777–9786, 2021. 1, 2, 6", + "[27] Junbin Xiao, Angela Yao, Yicong Li, and Tat-Seng Chua. Can i trust your answer? visually grounded video question answering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13204-13214, 2024. 1, 2, 5", + "[28] Yu Yang, Besmira Nushi, Hamid Palangi, and Baharan Mirzasoleiman. Mitigating spurious correlations in multimodal models during fine-tuning. In International Conference on Machine Learning, pages 39365-39379. PMLR, 2023. 2", + "[29] Shunyu Yao, Dian Yu, Jeffrey Zhao, Izhak Shafran, Tom Griffiths, Yuan Cao, and Karthik Narasimhan. Tree of thoughts: Deliberate problem solving with large language models. Advances in Neural Information Processing Systems, 36, 2024. 2", + "[30] Shoubin Yu, Jaemin Cho, Prateek Yadav, and Mohit Bansal. Self-chained image-language model for video localization and question answering. Advances in Neural Information Processing Systems, 36, 2024. 2", + "[31] Ce Zhang, Taixi Lu, Md Mohaiminul Islam, Ziyang Wang, Shoubin Yu, Mohit Bansal, and Gedas Bertasius. A simple llm framework for long-range video question-answering. arXiv preprint arXiv:2312.17235, 2023. 2, 6, 8", + "[32] Gengyuan Zhang, Yurui Zhang, Kerui Zhang, and Volker Tresp. Can vision-language models be a good guesser? exploring vlms for times and location reasoning. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 636-645, 2024. 2", + "[33] Hang Zhang, Xin Li, and Lidong Bing. Video-llama: An instruction-tuned audio-visual language model for video understanding. arXiv preprint arXiv:2306.02858, 2023. 1, 6, 7" + ], + "bbox": [ + 91, + 92, + 480, + 612 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "3271", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Commonsense Video Question Answering through Video-Grounded Entailment Tree Reasoning/e1080574-1050-4c6e-8509-597d8efaff00_model.json b/2025/Commonsense Video Question Answering through Video-Grounded Entailment Tree Reasoning/e1080574-1050-4c6e-8509-597d8efaff00_model.json new file mode 100644 index 0000000000000000000000000000000000000000..72b35dd74fb12aa9a72df401e78a35cbd0027542 --- /dev/null +++ b/2025/Commonsense Video Question Answering through Video-Grounded Entailment Tree Reasoning/e1080574-1050-4c6e-8509-597d8efaff00_model.json @@ -0,0 +1,1738 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.158, + 0.131, + 0.842, + 0.177 + ], + "angle": 0, + "content": "Commonsense Video Question Answering through Video-Grounded Entailment Tree Reasoning" + }, + { + "type": "text", + "bbox": [ + 0.117, + 0.202, + 0.882, + 0.257 + ], + "angle": 0, + "content": "Huabin Liu\\(^{1,3*}\\), Filip Ilievski\\(^{2}\\), Cees G. M. Snoek\\(^{3}\\) \n\\(^{1}\\) Shanghai Jiao Tong University \\(^{2}\\) Vrije Universiteit Amsterdam \\(^{3}\\) University of Amsterdam huabinliu@sjtu.edu.cn, f.ilievski@vu.nl, c.g.m.snoek@uva.nl" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.292, + 0.327, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.324, + 0.485, + 0.598 + ], + "angle": 0, + "content": "This paper proposes the first video-grounded entailment tree reasoning method for commonsense video question answering (VQA). Despite the remarkable progress of large visual-language models (VLMs), there are growing concerns that they learn spurious correlations between videos and likely answers, reinforced by their black-box nature and remaining benchmarking biases. Our method explicitly grounds VQA tasks to video fragments in four steps: entailment tree construction, video-language entailment verification, tree reasoning, and dynamic tree expansion. A vital benefit of the method is its generalizability to current video- and image-based VLMs across reasoning types. To support fair evaluation, we devise a de-biasing procedure based on large-language models that rewrites VQA benchmark answer sets to enforce model reasoning. Systematic experiments on existing and de-biased benchmarks highlight the impact of our method components across benchmarks, VLMs, and reasoning types." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.628, + 0.222, + 0.643 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.648, + 0.483, + 0.829 + ], + "angle": 0, + "content": "This paper proposes a video-grounded reasoning method for commonsense video question answering (VQA). VQA has a long tradition in computer vision [11, 16, 25, 26], with remarkable recent progress obtained through video-and image-language models [10, 12, 14, 17, 33] (throughout this paper collectively referred to as vision-language models, or VLMs). Yet, there are growing concerns that their improved performance is based on learning shortcut associations between videos and likely answers, as opposed to reasoning [27]. Such concerns are reinforced by the black-box nature of these models [12, 14], which prohibits a deeper understanding of their decision-making process." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.83, + 0.484, + 0.876 + ], + "angle": 0, + "content": "We are inspired by recent work in natural language processing, where entailment trees have emerged as a mechanism to explicitly analyze answer candidates, using LLMs" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.291, + 0.907, + 0.559 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.567, + 0.907, + 0.637 + ], + "angle": 0, + "content": "Figure 1. Given a video questioning answering task, our framework performs explicit reasoning over an entailment tree, where answer options are transformed into statements. These statements are then recursively decomposed and verified based on videotegued evidence relevant to the question." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.659, + 0.907, + 0.839 + ], + "angle": 0, + "content": "to recursively decompose a candidate into hypotheses and natural language inference formalisms to evaluate the hypotheses [4]. Entailment trees provide an explicit reasoning chain that explains the model's decision-making process and enables verification of each step, thus addressing concerns about shortcut learning. Recently, Sanders et al. [19] have devised a mechanism to apply entailment trees to videos. However, their work assumes video transcripts are explicitly provided to evaluate answers, thus avoiding the complexity of grounding hypotheses into video content. In this paper, we propose the first video-grounded entailment tree reasoning method for commonsense VQA." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Our method explicitly grounds VQA tasks to video fragments in four steps: (i) entailment tree construction, (ii) video-language entailment verification, (iii) tree reasoning, and (iv) dynamic tree expansion. As shown in Fig. 1, given" + }, + { + "type": "page_footnote", + "bbox": [ + 0.115, + 0.887, + 0.466, + 0.901 + ], + "angle": 0, + "content": "* Work was done as a visiting student at University of Amsterdam" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3262" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.094, + 0.092, + 0.482, + 0.424 + ], + "angle": 0, + "content": "a video and a multiple-choice question, we generate a statement for each answer candidate that acts as a first-level hypothesis. We decompose each statement iteratively, aiming to produce sub-statements that can be confidently verified in the video. The video is itself decomposed into partitions, consisting of sets of frames. Verifying each statement is then a matter of aligning it to a video partition. A vital benefit of the method is its generalizability to current video and image-based VLMs across reasoning types, including temporal and causal. To demonstrate its video reasoning ability, we develop an answer-set de-bias procedure supported by an LLM that ensures that VQA benchmarks [11, 26] are adequate for reasoning in videos without relying on spurious correlations. Our experiments show that our video-grounded entailment tree method consistently improves video- and image-based baselines on both the existing and de-biased benchmarks. Moreover, it performs on par with, and often better than, state-of-the-art video-based VLMs while leveraging \\(257 \\times\\) fewer parameters. Further ablations show that the method benefits from considering both textual and video information and that its performance is especially strong on causal and temporal questions." + }, + { + "type": "title", + "bbox": [ + 0.095, + 0.442, + 0.227, + 0.456 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "text", + "bbox": [ + 0.095, + 0.462, + 0.482, + 0.567 + ], + "angle": 0, + "content": "Video question answering. Recent research has shown that while video-based VLMs can achieve state-of-the-art performance, their answers are sensitive to object size, positioning, and speed [1, 28]. Moreover, when answering temporal and spatial questions, VLMs rely on textual biases to \"guess\" answers rather than performing genuine understanding and reasoning over visual-text information [32]." + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.57, + 0.482, + 0.9 + ], + "angle": 0, + "content": "To improve the robustness and interpretability of VLMs, one line of research enriches VLMs with visual grounding functionality during QA, which enables VLMs to localize relevant video moments [18, 27, 30] or key frames [15, 22] to support answers. However, while these methods localize visual evidence, the process by which VLMs use it to deduce answers remains opaque. Another approach leverages external LLMs as reasoners or agents to enhance interpretability in textual modality. For instance, LLoVi [31] converts VQA to a text-based QA task via video captioning, then prompts an LLM to provide answers. Similarly, VideoAgent [21] uses an LLM to recursively determine if the current frames can answer the given question based on their textual descriptions. However, these methods heavily rely on the reasoning capabilities of LLMs. Like VLMs, the LLM reasoning process remains a black box, and hallucinations are common. Recently, TV-trees [19] attempted to perform explicit reasoning over both visual and textual modalities using a neuro-symbolic system. However, their work assumes video transcripts are explicitly provided to evaluate answers, thus avoiding the complexity of grounding hypotheses into video content. Instead, we contribute a" + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.093, + 0.904, + 0.137 + ], + "angle": 0, + "content": "general framework for explicit reasoning in commonsense VQA, fueled by a grounding component that aligns question components with video fragments." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.139, + 0.905, + 0.334 + ], + "angle": 0, + "content": "Beyond methodology, some works focus on providing fair and comprehensive VLM evaluations in VQA tasks by creating new benchmarks [3, 6, 9]. These benchmarks contain videos with diverse scenarios and durations, with carefully crafted questions and options designed to prevent textual shortcuts that VLMs might exploit. Video-specific questions (e.g., compositional action reasoning) [1], which require insights beyond textual associations, are included to test commonsense reasoning in VLMs. Addressing concerns of remaining biases in such benchmarks [6], we contribute an LLM-based answer-set de-biasing procedure to ensure that VQA benchmarks [11, 26] are adequate to evaluate reasoning in videos rather than spurious correlations." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.336, + 0.905, + 0.758 + ], + "angle": 0, + "content": "Systematic language reasoning. As LLMs demonstrate great potential in reasoning, there has been considerable interest in using LLMs to generate systematic explanations to support their answers. The series of Chain-of-Thought prompting [2, 23, 29] encourages LLMs to think step-by-step to perform explicit multi-hop reasoning, providing free-form reasoning steps before arriving at an answer. However, such implicit explanations are not grounded in external knowledge or evidence, which may lead to unverifiable and unfaithful reasoning. Since the development of EntailmentBank [4], research has increasingly focused on constructing explanation trees [20, 24] and graphs [8], encouraging models to generate step-wise entailment proofs of a statement using a set of supporting facts. Entailer [20] introduced this systematic explanation framework into language-based multiple-choice QA, performing explicit reasoning by generating entailment trees grounded in the model's internal beliefs. REFLEX [8] extends the entailment tree to form a belief graph for QA models, aiming to address consistency issues by intervening in the intermediate reasoning steps. Instead of grounding facts in predefined rules or model beliefs, NELLIE [24] adopts Prolog-based inference engines and external natural language corpora to build entailment trees as explainable reasoning for multiple-choice QA tasks. While such techniques for natural language processing inspire our framework, we generalize entailment trees to VQA, contributing a novel grounding method that aligns entailment trees with video fragments." + }, + { + "type": "title", + "bbox": [ + 0.517, + 0.776, + 0.89, + 0.792 + ], + "angle": 0, + "content": "3. Video-grounded entailment tree reasoning" + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.796, + 0.904, + 0.9 + ], + "angle": 0, + "content": "This paper devises a novel explainable framework for grounded commonsense VQA. It derives the answers through systematic reasoning over video-text information with entailment trees. Specifically, in the entailment tree (Fig. 2a), each candidate answer is decomposed into statements that entail the answer, explaining why each answer could be plausible. These statements are then grounded" + }, + { + "type": "page_number", + "bbox": [ + 0.485, + 0.947, + 0.514, + 0.957 + ], + "angle": 0, + "content": "3263" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.228 + ], + "angle": 0, + "content": "in relevant visual evidence from the video to prove or refute them (Fig. 2b). While entailment trees in natural language processing are constructed based on a model's internal knowledge or corpora [8, 20, 24], we ground entailment trees into video fragments. Finally, backtracking through the entailment tree leads to systematic reasoning over the statements (Fig. 3). Thus, answers can be deduced by a systematic structure with explicit reasoning paths and explanations rather than relying on opaque, black-box models." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.239, + 0.348, + 0.253 + ], + "angle": 0, + "content": "3.1. Entailment tree construction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.255, + 0.483, + 0.45 + ], + "angle": 0, + "content": "Initial statement generation. Given a question and its answer candidates, we first convert each question-answer pair into a declarative sentence that preserves the semantic meaning of the original QA pair. As a result, an \\(N\\)-way multiple-choice QA problem produces a set of statements, denoted as \\(\\mathcal{D} = d_{1},\\ldots ,d_{N}\\). For example, the two-way question \"What did the boy in white do after he first took the balloon? (A) resting on a chair (B) carries it toward the hula hoop\" is transformed into: \\(\\mathcal{D}:\\{d_1 =\\) \"The boy in white resting on a chair after first taking the balloon.\", \\(d_{2} =\\) \"The boy in white carries it toward the hula hoop after first taking the balloon.\" \\(\\}\\). Thus, selecting the best answer equals identifying the correct statement for a given video." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.452, + 0.483, + 0.663 + ], + "angle": 0, + "content": "Recursive statement decomposition. For each initial statement in \\(\\mathcal{D}\\), we generate two sub-statements as proofs that support the statement: Statement \\(\\Leftarrow\\) Sub-statement1, Sub-statement2. The statement is True if and only if both its sub-statements are proved to be True, i.e., the sub-statements entail the statement. Proving the original statements is thus translated into proving two simpler sub-statements. This procedure is recursive: the sub-statements can be further decomposed into further sub-statements that entail them. Therefore, to construct an entailment tree, we recursively decompose these sub-statements as new statements in the next tree layer until reaching the maximum depth or meeting the stop criterion. Fig. 2(a) presents an example of entailment tree generation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.663, + 0.483, + 0.709 + ], + "angle": 0, + "content": "We leverage LLM prompting for both the initial statement generation and the statement decomposition, as these are linguistic tasks (see implementation details)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.719, + 0.424, + 0.735 + ], + "angle": 0, + "content": "3.2. Video-language entailment verification" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Given the entailment tree, the framework then verifies language statements based on the grounded video content as evidence. Specifically, each statement in the entailment tree must be proven or refuted by analyzing the video. A straightforward solution is to encode the whole video to collect information that can be used to verify the statement. However, the critical visual evidence that accurately verifies a statement tends to exist in a local moment instead of the whole video. Therefore, we develop a novel video grounding that guides the verification process to the moments with relevant visual evidence." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.363 + ], + "angle": 0, + "content": "Question-aware video captioning. Given a video, we convert its visual information into detailed textual information. Specifically, we input video frames into a VLM-based captioner \\(\\mathrm{Cap}(\\cdot)\\) to obtain a caption \\(c_{i} = \\mathrm{Cap}(f_{i})\\) for each frame. However, captioning frames individually can overlook essential details or introduce irrelevant information for VQA. In commonsense VQA, questions often focus on specific facts already observed in the video. For example, a typical temporal reasoning question is \"What happened before/after Event-A?\" where Event-A refers to a fact statement about an event in the video. The fact referenced by the question can be leveraged to guide video understanding. To this end, we first extract the anchor fact indicated by the question and provide it to \\(\\mathrm{Cap}(\\cdot)\\) as prior knowledge, encouraging the generation of relevant captions. Moreover, captions from all previous frames are also provided for each current frame to ensure \\(\\mathrm{Cap}(\\cdot)\\) captures the temporal context from the past. This process is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.597, + 0.377, + 0.905, + 0.394 + ], + "angle": 0, + "content": "\\[\nc _ {i} = \\operatorname {C a p} \\left(f _ {i} \\mid F, \\left(c _ {1}, \\dots , c _ {i - 1}\\right)\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.406, + 0.758, + 0.42 + ], + "angle": 0, + "content": "where \\(F\\) indicates the fact statement." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.421, + 0.905, + 0.557 + ], + "angle": 0, + "content": "Video evidence grounding. For commonsense VQA, depending on how the question reasons around the fact statement, the necessary evidence for answers can be gathered from specific video moments. For instance, in the case of temporal reasoning (e.g., before or after questions), the answer should be inferred from moments occurring either before or after the time of the relevant fact. Following this intuition, we design a two-step evidence-grounding strategy to localize the critical moments for answering." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.558, + 0.906, + 0.815 + ], + "angle": 0, + "content": "First, given the frame-wise captions, we retrieve a keyframe deemed most relevant to the fact statement, which we refer to as the anchor frame. A straightforward retrieval approach would involve comparing each \\( c_i \\) with the fact description using specific metrics to identify the anchor frame. However, we enhance retrieval accuracy by adopting a structured semantic retrieval strategy. Specifically, the textual descriptions of each frame and fact statement are converted into structured triplets. These triplets capture the attributes and relationships of objects in each frame through structured semantics. As shown in Fig. 2(a), rather than directly comparing raw textual descriptions of frames and fact statements, we use these triplets for retrieval. Inspired by the success of using LLMs for retrieval tasks, we prompt an LLM to conduct anchor frame retrieval using the triplets of the fact statement as the query. The LLM then identifies and returns the most relevant frame ID, i.e., its timestamp." + }, + { + "type": "equation", + "bbox": [ + 0.638, + 0.827, + 0.905, + 0.844 + ], + "angle": 0, + "content": "\\[\nt _ {\\text {a n c h o r}} = \\operatorname {R t v} \\left(c _ {i}, F\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.856, + 0.905, + 0.901 + ], + "angle": 0, + "content": "where \\( t_{\\mathrm{anchor}} \\) is the time stamp of the anchor frame, \\( \\mathrm{Rtv}(\\cdot) \\) denotes the retrieval process. Second, we determine the final moment where we should look centered on the \\( t_{\\mathrm{anchor}} \\), to" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3264" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.089, + 0.905, + 0.434 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.44, + 0.908, + 0.498 + ], + "angle": 0, + "content": "Figure 2. Overview of our framework. (a) The generation of the entailment tree, where statements are recursively decomposed until the tree reaches its max depth or meets the stop criterion. (b) The process of video-language entailment verification: the input video is first converted into textual descriptions. Each caption is then parsed into structured semantics. Given the fact statement as a query, we retrieve the anchor frame. Then, based on the temporal or causal navigation indicated by questions, the visual evidence moment can be grounded." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.517, + 0.484, + 0.579 + ], + "angle": 0, + "content": "incorporate the temporal relations present in the question. Therefore, based on the anchor frame, the navigation for the moment is selected from \"look ahead, look behind, look around\" by considering the question:" + }, + { + "type": "equation", + "bbox": [ + 0.212, + 0.592, + 0.482, + 0.61 + ], + "angle": 0, + "content": "\\[\n\\mathcal {M} = \\operatorname {G n d} \\left(t _ {\\text {a n c h o r}} \\mid \\mathcal {Q}\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.623, + 0.484, + 0.683 + ], + "angle": 0, + "content": "where \\(\\operatorname{Gnd}(\\cdot)\\) is the grounding process and \\(\\mathcal{M}\\) denotes the grounded continuous interval in the video. Then, frames are resampled from the video within \\(M\\) and used as visual evidence proving or refuting entailment tree statements." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.685, + 0.484, + 0.866 + ], + "angle": 0, + "content": "Visual-text statement prover. Given grounded visual evidence \\(\\mathcal{M}\\) of the video, statements are estimated to be true or false. Specifically, we employ a VLM as the statement prover, denoted as \\(\\mathrm{Prv}(\\cdot)\\), to evaluate each statement within the tree by probing VLM's internal belief on this statement. Each statement will be transformed into a binary QA task, with possible options being True or False. We then directly probe the \\(\\mathrm{Prv}(\\cdot)\\) with the binary QA prompt and use the next token prediction probabilities of the words to elicit the model's belief. We normalize the prediction logits of the two options to get the confidence score of that statement. The above process is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.188, + 0.881, + 0.483, + 0.898 + ], + "angle": 0, + "content": "\\[\ns _ {d} = \\Pr \\operatorname {r v} (\\mathcal {M}, h), s \\in [ 0, 1 ], \\tag {4}\n\\]" + }, + { + "type": "image", + "bbox": [ + 0.514, + 0.513, + 0.905, + 0.611 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.62, + 0.907, + 0.677 + ], + "angle": 0, + "content": "Figure 3. Illustration of dynamic tree generation and backtrace. In Step-3, when the proof score of the left statement calculated from its child nodes is less than its direct score \\((0.63 < 0.8)\\), its decomposition is pruned and stops." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.691, + 0.907, + 0.721 + ], + "angle": 0, + "content": "where \\(\\mathcal{M}\\) is the grounded moment and \\(h\\) denotes the statement that needs to be verified." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.734, + 0.822, + 0.749 + ], + "angle": 0, + "content": "3.3. Dynamic entailment tree expansion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.907, + 0.902 + ], + "angle": 0, + "content": "So far, we have performed statement decomposition recursively to construct an entailment tree with pre-defined depth. However, not all statements need to be verified recursively, especially those easily determined to be true or false by VLMs. Moreover, as the depth increases, some statement sentences are atomic and directly verifiable. Thus, to improve the efficiency of the reasoning process, we further adopt a strategy to expand the entailment tree dynamically. Specifically, each statement \\( d \\) is tied with two confidence scores provided by the \\( \\operatorname{Prv}(\\cdot) \\):" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3265" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.089, + 0.482, + 0.351 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.366, + 0.483, + 0.395 + ], + "angle": 0, + "content": "Figure 4. Illustration of commonsense bias in video question answering. The example is selected from the NExT-QA dataset." + }, + { + "type": "text", + "bbox": [ + 0.079, + 0.413, + 0.483, + 0.441 + ], + "angle": 0, + "content": "(1) The direct score \\( s_d \\), which indicates the belief of \\( \\operatorname{Prv}(\\cdot) \\) model in \\( d \\)." + }, + { + "type": "text", + "bbox": [ + 0.079, + 0.443, + 0.483, + 0.487 + ], + "angle": 0, + "content": "(2) The proof score \\( s_p \\), denoting how confidently the model can prove \\( d \\), is calculated by multiplying the scores of its direct sub-statements." + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.413, + 0.483, + 0.487 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.489, + 0.484, + 0.623 + ], + "angle": 0, + "content": "For a statement \\( d \\), the goal of decomposition is to establish a more reliable and convincing proof path than merely evaluating whether \\( d \\) is true by VLMs. If the decomposition-based reasoning can prove \\( d \\) with higher confidence than its direct score, the overall confidence for statement \\( d \\) should increase. Otherwise, the decomposition should be disregarded. Thus, in the dynamic tree expansion, if decomposition does not enhance a statement's score, it is pruned, and that statement node becomes a leaf in the entailment tree." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.625, + 0.484, + 0.671 + ], + "angle": 0, + "content": "Fig. 3 presents a toy example. This criterion ensures that only beneficial decompositions are retained, significantly enhancing the tree reasoning process's efficiency." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.682, + 0.395, + 0.697 + ], + "angle": 0, + "content": "3.4. Reasoning over the entailment tree" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.698, + 0.483, + 0.805 + ], + "angle": 0, + "content": "Finally, we perform a backtrace through the entailment tree to calculate the confidence score of each top statement. Specifically, the final score for each statement is produced by comparing its direct score \\( s_d \\) and proof score \\( s_p \\), i.e., \\( s = \\max(s_d, s_p) \\) during backtrace (as shown in Fig. 3). The overall framework selects the answer corresponding to the statement at the top layer with top-scoring proof." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.821, + 0.474, + 0.838 + ], + "angle": 0, + "content": "4. De-biasing commonsense VQA answer sets" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.841, + 0.484, + 0.901 + ], + "angle": 0, + "content": "To demonstrate the reasoning ability of video-grounded entailment trees, it is essential to evaluate using commonsense VQA benchmarks that enforce model reasoning. Recent work [13, 18, 27] has provided evidence that shortcuts" + }, + { + "type": "image", + "bbox": [ + 0.526, + 0.095, + 0.897, + 0.281 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.534, + 0.295, + 0.885, + 0.31 + ], + "angle": 0, + "content": "Figure 5. Prompt used for rewriting answers on NExT-QA." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.357, + 0.907, + 0.63 + ], + "angle": 0, + "content": "are present in VQA datasets which enables VLMs to solve these tasks based on textual associations rather than video-grounded reasoning. While VQA benchmarks increasingly focus on commonsense reasoning skills, such as temporal (e.g. after, before) or causal (how, why, what if) relationships in video content, reasoning shortcuts affect the validity of their evaluation. This is illustrated in Fig. 4 (top), where the correct answer (D) is much more relevant to the question and also aligns best with real-world expectations. Consequently, a VLM (VideoLLaVA [14] used in this example) can answer this question correctly by leveraging such associations and without analyzing the video content. Meanwhile, replacing the answer set distractors with other commonsensical answer candidates, as illustrated in Fig. 4 (bottom), makes this task challenging for VLMs. Here, a VLM switches its answer incorrectly to option (C), which confirms the impact of commonsense associations and the lack of grounded reasoning by these models." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.645, + 0.909, + 0.901 + ], + "angle": 0, + "content": "To this end, we devise a de-biasing procedure that mitigates reasoning shortcuts in commonsense VQA answer sets. Our de-biasing procedure transforms multiple-choice VQA benchmarks (e.g., NExT-QA) by rewriting their answer distractors while keeping their question and ground-truth answer intact. We prompt an LLM (LLaMA-3) to implement the rewriting procedure for each original QA set. Fig. 5 shows the detailed prompt we used for LLaMA-3 on NExT-QA dataset. This procedure ensures that (1) the answers cannot be easily derived from the QA set associations and (2) the answer remains consistent with the original QA pair. Thus, our procedure enables the scalable construction of de-biased QA sets by leveraging the commonsense associations in LLMs. The next section, focusing on experimental evaluation, analyzes the application of de-biasing to various datasets and its impact on the performance of VLMs, with and without entailment tree reasoning." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3266" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.223, + 0.107 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.11, + 0.278, + 0.126 + ], + "angle": 0, + "content": "5.1. Experimental setup" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.127, + 0.482, + 0.262 + ], + "angle": 0, + "content": "Datasets. We test our framework on three VQA benchmarks: (1) NExT-QA [26], a VQA benchmark for causal and temporal reasoning. (2) IntentQA [11], which focuses on video intent reasoning from both causal and temporal aspects. (3) Video-MME [6], which is a recently proposed comprehensive evaluation benchmark for video analysis; we use its \"short-term\" split (video length \\(< 2\\) mins) and 4 question types (temporal, spatial, action, and object reasoning) highly related to commonsense reasoning are selected." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.264, + 0.482, + 0.325 + ], + "angle": 0, + "content": "Evaluation. We report model performances on our rewritten test set for each dataset and its original test set. We evaluate our framework on all datasets under the multiple-choice QA setting, using a standard accuracy metric." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.326, + 0.433, + 0.34 + ], + "angle": 0, + "content": "Baselines. Our baselines represent three categories:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.342, + 0.481, + 0.416 + ], + "angle": 0, + "content": "- Video-based VLMs: Video-based VLMs are widely used for VQA tasks, so we include Video-LLaVA [14], VideoChat2 [12], and VideoLLaMA [33]. To test the effectiveness of our framework, we integrate these VLMs by replacing our \\(\\mathsf{Prv}(\\cdot)\\) with specific VLM models." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.418, + 0.481, + 0.447 + ], + "angle": 0, + "content": "- Image-based VLMs: We include BLIP-2 [10] and LLaVA-1.5 [17] as Image-LLM baselines." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.448, + 0.481, + 0.492 + ], + "angle": 0, + "content": "- State-of-the-art VQA approaches: Recent works in VQA, such as VideoTree [22], VideoAgent [21], and LLoVi [31], are included as strong baselines." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.342, + 0.481, + 0.492 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.494, + 0.483, + 0.827 + ], + "angle": 0, + "content": "Implementation details. Our entire framework is training-and human annotation-free. We use LLaMA-3-8B [5] to handle basic functionalities, including (1) converting original QA into declarative statements, (2) statement decomposition, (3) structured semantic extraction and retrieval, and (4) guiding evidence grounding. Detailed prompts for each functionality are provided in the supplementary material. For frame-wise captioning across all datasets, we use LLaVA-1.5 [17] as our default captioner. When comparing with state-of-the-art VQA methods (e.g., VideoAgent), we follow their setup by replacing the captioner with the stronger CogAgent [7] model for fair comparison. When integrating our framework with VLMs, the VLM itself serves as the Prover, \\(\\mathsf{Prv}(\\cdot)\\). For video-LLMs, frames are uniformly sampled from the grounded video moment to meet their input requirements (VideoChat2: 16 frames; VideoLLaVA & VideoLLaMA: 8 frames). For image-language models like LLaVA, we sample 8 frames from the grounded moment and process them individually; the final confidence score is obtained by averaging scores across frames. During dynamic tree generation, we also set a max depth of 5 for the overall entailment tree to improve the efficiency." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.839, + 0.224, + 0.853 + ], + "angle": 0, + "content": "5.2. Main results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.482, + 0.901 + ], + "angle": 0, + "content": "Benefit for image and video-based VLMs. Tab. 1 summarizes the results of our method compared to baselines. Integrating entailment tree reasoning brings consistent im" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.317 + ], + "angle": 0, + "content": "provement across the video- and image-based VLMs for all datasets (1-4% on average). This finding includes the recently proposed benchmark VideoMME, which poses much more challenging videos and questions. The benefits are particularly regular for temporal reasoning (improvement in 14 out of 15 cases), which illustrates how our explicit reasoning process enhances temporal commonsense QA. Image-based VLMs, which initially lack temporal modeling capabilities, perform poorly when directly applied to video QA tasks. However, our framework provides a significant performance boost for these models up to \\(8\\%\\) for LLAVA-1.5. By reasoning over multiple sub-problems rather than tackling the entire complex question simultaneously, our reasoning method makes the task more manageable for both video- and image-based VLMs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.318, + 0.905, + 0.604 + ], + "angle": 0, + "content": "Results on de-biased QA sets. As shown in Tab. 2, all models experience considerable performance drops on the de-biased set of the same VQA dataset, which aligns with our observation that current VLMs often rely on textual bias in commonsense reasoning tasks. Notably, video-based VLMs show an \\(8\\% -10\\%\\) decrease in the de-biased set even though the question and the correct answer remain unchanged. In contrast, our proposed framework, which derives answers through an explicit reasoning process based on specific visual evidence, demonstrates much greater robustness on the de-biased set. The improvement brought by our framework on the de-biased set is even higher than on the original test sets. In turn, our framework compensates for the performance loss of the VLMs on the de-biased set. This analysis underscores our framework's potential to mitigate textual bias in commonsense reasoning. Furthermore, the performance differences between the original and de-biased QA sets highlight VQA benchmarks' limitations in evaluating VLMs' true reasoning abilities." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.606, + 0.905, + 0.801 + ], + "angle": 0, + "content": "Comparison with state-of-the-art. Tab. 3 compares our framework's results on the de-biased sets to state-of-the-art VQA approaches. The table shows that, next to the consistent benefit our framework provides to various VLMs, it is also competitive with state-of-the-art VQA methods. When applying an advanced captioner and reasoner that aligns with VideoAgent and VideoTree, our framework yields new state-of-the-art results in some cases. In particular, our framework performs best on temporal reasoning questions for all three benchmarks and outperforms all methods on the IntentQA dataset. Importantly, our method reaches such competitive performance despite using \\(257 \\times\\) fewer parameters for its reasoning compared to state-of-the-art methods." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.81, + 0.674, + 0.824 + ], + "angle": 0, + "content": "5.3. Ablation studies" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.904, + 0.87 + ], + "angle": 0, + "content": "Ablation experiments are conducted using VideoLLaVA's baseline with our framework. Results are reported on the test set of the NExT-QA dataset." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Impact of LLMs for statement decomposition. Entailment generation in our framework relies on prompting an" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3267" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.093, + 0.089, + 0.907, + 0.265 + ], + "angle": 0, + "content": "
ModelNExT-QAIntentQAVideoMMEAvg
TemporalCausalTemporalCausalTemporalSpatialActionObject
Image-based VLMsBLIP-2 [10]38.336.143.848.625.426.924.228.634.0
+Ours45.341.848.952.530.527.427.728.837.9
LLaVA-1.5 [17]37.840.745.850.031.133.627.729.837.1
+Ours45.647.948.454.736.736.831.430.741.5
Video-based VLMsVideoChat2 [12]56.962.160.463.250.352.449.550.155.6
+Ours57.861.662.363.852.851.551.350.056.4
VideoLLaVA [14]56.060.453.960.747.644.346.249.752.3
+Ours58.362.757.661.848.346.149.850.354.2
VideoLLaMA [33]55.460.255.156.744.847.244.749.351.7
+Ours58.160.454.558.947.447.848.649.153.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.269, + 0.907, + 0.298 + ], + "angle": 0, + "content": "Table 1. Impact on image and video-based VLMs on the original NExT-QA, IntentQA, and VideoMME test sets. Our framework increases accuracy of all video- and image-based VLMs by \\(1 - 4\\%\\) on average across all data partitions. Temporal and action partitions benefit most." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.311, + 0.907, + 0.423 + ], + "angle": 0, + "content": "
ModelBLIP-2+OursLLaVA+OursVideo Chat2+OursVideo LLaVA+OursVideo LLaMA+Ours
NExT-QAOriginal37.243.639.346.859.559.758.260.557.859.3
Rewritten33.539.834.844.945.449.051.155.441.447.0
Intent-QAOriginal46.250.747.951.661.863.157.359.755.956.7
Rewritten38.245.542.748.652.655.750.554.746.350.0
Avg38.844.941.248.054.856.954.357.650.453.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.428, + 0.907, + 0.456 + ], + "angle": 0, + "content": "Table 2. Results on de-biased QA sets. Video-based VLMs show significant decreases in the rewritten de-biased set. In contrast, our framework demonstrates much greater robustness on the rewritten set." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.482, + 0.483, + 0.695 + ], + "angle": 0, + "content": "external LLM to recursively decompose statements (cf. Sec. 3.1), which is crucial in guiding reasoning paths. Consequently, we tested various LLMs for entailment tree generation, including open-source models (LLama-3 and Mistral) of different sizes and proprietary LLMs (GPT-4 and Gemini-1.5). The results are summarized in Tab. 4. As expected, the proprietary model GPT-4, known for its strong step-by-step reasoning capabilities, delivers the best performance across all settings. Scaling up LLaMA-3 to 70B offers improvements over the 8B model, though with a notable increase in inference time. As the overall performance difference between models is within \\(1\\%\\), we select LLaMA-3-8B as the default for integrating our framework into VLMs due to its free availability and efficiency." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.705, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Ablation on grounding components. Next, we test the effectiveness of each component in our grounding module (Sec. 3.2). The results, summarized in Tab. 5, indicate that both fact-conditional captioning and structure-guided retrieval enhance overall performance by improving grounding accuracy. However, using only structure-guided retrieval results in a slight performance drop, possibly because \\(\\mathsf{Cap}(\\cdot)\\) introduces irrelevant semantic information that doesn't align with the question's focus, and the structured representation can make identifying anchor frames more challenging. In contrast, fact-conditional captioning alone yields substantial improvement, demonstrating that this straightforward approach can yield an effective" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.482, + 0.905, + 0.513 + ], + "angle": 0, + "content": "and more controllable textual description for videos by conditioning on prior knowledge or relevant facts." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.518, + 0.907, + 0.715 + ], + "angle": 0, + "content": "Impact of length of video frames. We further ablate the impact of input video frame length in our framework to determine the optimal number of frame-wise captions to generate per video. The results, summarized in Tab. 6, show that ideal performance is achieved only when sampling a sufficient number of frames (at least 16 for NExTQA). When fewer frames are used (e.g., 4 or 8), key anchor frames may be missed, reducing the accuracy of grounded visual evidence. Additionally, while increasing the frame count to 32 yields the best performance, it also increases the calls required for \\(\\mathrm{Cap}(\\cdot)\\) to generate frame-wise captions. Balancing efficiency with performance gains, we set 24 frames as the default in our implementation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Effectiveness of evidence grounding. Our method grounds relevant video fragments to support statements in the entailment tree (Sec. 3.2). To validate its effectiveness, we compare it with two other variations as sources of visual evidence: (1) without evidence grounding, using the full video as evidence, and (2) upper-bound results: manually annotated temporal boundaries provided in the NExT-GQA dataset, indicating where the QA models should focus when producing correct answers. The results are shown in Tab. 7. Compared to the baseline, our video-grounded method provides consistent improvements across the original and debiased sets. The improvement is more apparent in the de" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3268" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.093, + 0.089, + 0.907, + 0.181 + ], + "angle": 0, + "content": "
MethodModel ReasonerNExT-QA*IntentQA*VideoMME
TemporalCausalTemporalCausalTemporalSpatialActionObject
VideoAgent [21]GPT-4 (1.8T)58.266.660.461.0----
VideoTree [22]GPT-4 (1.8T)60.266.456.760.155.754.354.252.6
LLoVi [31]GPT-4 (1.8T)53.160.858.661.852.255.351.850.8
OursVideoLLAVA (7B)60.865.961.062.655.953.854.050.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.185, + 0.907, + 0.242 + ], + "angle": 0, + "content": "Table 3. Comparison with state-of-the-art. Results for NExT-QA and IntentQA are reported under the de-biased set (the results on the original sets are similar; we provide them in the Appendix). The 'Reasoner' in these approaches is similar to the \"Prover\" in our framework. The captioner for all methods is CogAgent [7]. Despite other methods relying on much stronger reasoning models, our approach yields competitive performance (four state-of-the-art results) and high parameter efficiency (\\(257 \\times\\) fewer than GPT-4 reasoners)." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.265, + 0.481, + 0.376 + ], + "angle": 0, + "content": "
Model classLLMNExT-QA
OriginalRewritten
Open-sourceMistral-7B60.055.6
LLaMA-3-8B60.555.4
LLaMA-3-70B61.355.9
ProprietaryGemini-1.5-Pro61.155.2
GPT-461.656.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.38, + 0.483, + 0.437 + ], + "angle": 0, + "content": "Table 4. Impact of LLMs for statement decomposition. The open-source and proprietary models are ordered ascendingly by size. Larger models, especially GPT-4, are best at decomposition, but the smaller models (e.g., LlaMa-3-8B) come close." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.453, + 0.483, + 0.557 + ], + "angle": 0, + "content": "
ComponentsNExT-QA
Fact-conditioned captioningStructure-based retrievalOriginalRewritten
56.249.6
✓59.553.3
✓58.452.7
✓✓60.555.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.562, + 0.483, + 0.605 + ], + "angle": 0, + "content": "Table 5. Ablation on grounding components, showing that both fact-conditional captioning and structure-guided retrieval enhance overall performance by improving grounding accuracy." + }, + { + "type": "table", + "bbox": [ + 0.117, + 0.621, + 0.46, + 0.687 + ], + "angle": 0, + "content": "
Acc\n(NExT-QA)Frame number
48162432
Original57.759.360.560.761.0
Rewritten51.953.455.455.455.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.69, + 0.483, + 0.719 + ], + "angle": 0, + "content": "Table 6. Impact of video frame amount. Strong performance requires a sufficiently high frame number (over 16 for NExT-QA)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.748, + 0.483, + 0.824 + ], + "angle": 0, + "content": "biased set, where the answer options are more semantically similar and require more precise, discriminative visual evidence. Using the ground-truth fragment can further boost our approach, suggesting that enhancing grounding accuracy could further improve our framework." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Effectiveness of dynamic tree expansion. The depth of the entailment tree determines the granularity of reasoning (Sec. 3.3). This ablation analyzes how tree depth impacts overall performance and compares a fixed-depth approach with our dynamic tree generation strategy. Increases" + }, + { + "type": "table", + "bbox": [ + 0.515, + 0.265, + 0.906, + 0.315 + ], + "angle": 0, + "content": "
Video fragmentFullGrounded (ours)GT
NExT-QAOriginal58.360.561.8
Rewritten51.755.456.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.32, + 0.907, + 0.363 + ], + "angle": 0, + "content": "Table 7. Effectiveness of evidence grounding. Our videotegrounded method yields clear improvement over using the full video. More precise grounding can further enhance our accuracy." + }, + { + "type": "table", + "bbox": [ + 0.515, + 0.375, + 0.906, + 0.433 + ], + "angle": 0, + "content": "
StrategyStatic (Depth=)Dynamic
2345
NExT-QAOriginal58.859.260.260.360.5
Rewritten52.053.455.655.355.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.437, + 0.906, + 0.467 + ], + "angle": 0, + "content": "Table 8. Effectiveness of dynamic tree expansion. It yields superior accuracy while increasing reasoning efficiency." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.491, + 0.907, + 0.703 + ], + "angle": 0, + "content": "ing the depth of reasoning yields significant improvements, as complex, long statements are broken down into concise sub-statements that VLMs can understand more effectively. However, extending reasoning beyond the 4th layer offers diminishing returns; for NExT-QA, the original statements' complexity constrains the task, and some 5th-layer sub-statements become overly simplistic and less effective for reasoning. This finding highlights the necessity of our dynamic strategy. Applying the dynamic tree expansion strategy, we can see that the performance outperforms the fixed-depth paradigm. In the meantime, the dynamic strategy increases the reasoning efficiency over the entailment tree, more details about efficiency comparison can be found in our supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.716, + 0.634, + 0.731 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.734, + 0.907, + 0.9 + ], + "angle": 0, + "content": "This paper proposed the first video-grounded entailment tree framework for VQA. Moreover, we also contributed a de-biasing procedure to avoid spurious correlations during evaluation and applied it to enhance representative benchmarks. Extensive experiments with five video- and image-based VLMs demonstrate consistent benefits of our method on these benchmarks. Besides, our proposed framework performs on par with state-of-the-art video reasoning methods despite using \\(257 \\times\\) fewer parameters. While de-biasing hurts VLM accuracy, our framework regains the accuracy losses and is competitive with state-of-the-art VQA methods." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3269" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.183 + ], + "angle": 0, + "content": "[1] Piyush Bagad, Makarand Tapaswi, and Cees GM Snoek. Test of time: Instilling video-language models with a sense of time. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2503-2516, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.186, + 0.484, + 0.269 + ], + "angle": 0, + "content": "[2] Maciej Besta, Nils Blach, Ales Kubicek, Robert Gerstenberger, Michal Podstawski, Lukas Gianinazzi, Joanna Gajda, Tomasz Lehmann, Hubert Niewiadomski, Piotr Nczyk, et al. Graph of thoughts: Solving elaborate problems with large language models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 17682-17690, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.271, + 0.483, + 0.338 + ], + "angle": 0, + "content": "[3] Tieyuan Chen, Huabin Liu, Tianyao He, Yihang Chen, Chaofan Gan, Xiao Ma, Cheng Zhong, Yang Zhang, Yingxue Wang, Hui Lin, et al. Mecd: Unlocking multi-event causal discovery in video reasoning. arXiv preprint arXiv:2409.17647, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.341, + 0.483, + 0.394 + ], + "angle": 0, + "content": "[4] Bhavana Dalvi, Peter Jansen, Oyvind Tafjord, Zhengnan Xie, Hannah Smith, Leighanna Pipatanangkura, and Peter Clark. Explaining answers with entailment trees. arXiv preprint arXiv:2104.08661, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.397, + 0.483, + 0.464 + ], + "angle": 0, + "content": "[5] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.467, + 0.483, + 0.535 + ], + "angle": 0, + "content": "[6] Chaoyou Fu, Yuhan Dai, Yongdong Luo, Lei Li, Shuhuai Ren, Renrui Zhang, Zihan Wang, Chenyu Zhou, Yunhang Shen, Mengdan Zhang, et al. Video-mme: The first-ever comprehensive evaluation benchmark of multi-modal llms in video analysis. arXiv preprint arXiv:2405.21075, 2024. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.537, + 0.483, + 0.619 + ], + "angle": 0, + "content": "[7] Wenyi Hong, Weihan Wang, Qingsong Lv, Jiazheng Xu, Wenmeng Yu, Junhui Ji, Yan Wang, Zihan Wang, Yuxiao Dong, Ming Ding, et al. Cogagent: A visual language model for gui agents. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14281-14290, 2024. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.621, + 0.483, + 0.675 + ], + "angle": 0, + "content": "[8] Nora Kassner, Oyvind Tafjord, Ashish Sabharwal, Kyle Richardson, Hinrich Schuetze, and Peter Clark. Language models with rationality. arXiv preprint arXiv:2305.14250, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.677, + 0.483, + 0.746 + ], + "angle": 0, + "content": "[9] Muhammad Uzair Khattak, Muhammad Ferjad Naeem, Jameel Hassan, Muzammal Naseer, Federico Tombari, Fahad Shahbaz Khan, and Salman Khan. Complex video reasoning and robustness evaluation suite for video-lmms. arXiv preprint arXiv:2405.03690, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.747, + 0.483, + 0.815 + ], + "angle": 0, + "content": "[10] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730–19742. PMLR, 2023. 1, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.818, + 0.483, + 0.872 + ], + "angle": 0, + "content": "[11] Jiapeng Li, Ping Wei, Wenjuan Han, and Lifeng Fan. Intentqa: Context-aware video intent reasoning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11963-11974, 2023. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.874, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[12] Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Yi Liu, Zun Wang, Jilan Xu, Guo Chen, Ping Luo, et al." + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.906, + 0.147 + ], + "angle": 0, + "content": "Mvbench: A comprehensive multi-modal video understanding benchmark. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22195-22206, 2024. 1, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.149, + 0.906, + 0.204 + ], + "angle": 0, + "content": "[13] Yicong Li, Xiang Wang, Junbin Xiao, Wei Ji, and Tat-Seng Chua. Invariant grounding for video question answering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2928-2937, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.205, + 0.906, + 0.259 + ], + "angle": 0, + "content": "[14] Bin Lin, Yang Ye, Bin Zhu, Jiaxi Cui, Munan Ning, Peng Jin, and Li Yuan. Video-llava: Learning united visual representation by alignment before projection. arXiv preprint arXiv:2311.10122, 2023. 1, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.261, + 0.906, + 0.315 + ], + "angle": 0, + "content": "[15] Huabin Liu, Weixian Lv, John See, and Weiyao Lin. Task-adaptive spatial-temporal video sampler for few-shot action recognition. In Proceedings of the 30th ACM International Conference on Multimedia, pages 6230-6240, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.317, + 0.906, + 0.37 + ], + "angle": 0, + "content": "[16] Huabin Liu, Weiyao Lin, Tieyuan Chen, Yuxi Li, Shuyuan Li, and John See. Few-shot action recognition via intra- and inter-video information maximization. arXiv preprint arXiv:2305.06114, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.372, + 0.906, + 0.427 + ], + "angle": 0, + "content": "[17] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26296-26306, 2024. 1, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.429, + 0.906, + 0.496 + ], + "angle": 0, + "content": "[18] Huabin Liu, Xiao Ma, Cheng Zhong, Yang Zhang, and Weiyao Lin. Timecraft: Navigate weakly-supervised temporal grounded video question answering via bi-directional reasoning. In European Conference on Computer Vision, pages 92-107. Springer, 2025. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.498, + 0.906, + 0.55 + ], + "angle": 0, + "content": "[19] Kate Sanders, Nathaniel Weir, and Benjamin Van Durme. Tv-trees: Multimodal entailment trees for neuro-symbolic video reasoning. arXiv preprint arXiv:2402.19467, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.553, + 0.906, + 0.595 + ], + "angle": 0, + "content": "[20] Oyvind Tafjord, Bhavana Dalvi Mishra, and Peter Clark. Entailer: Answering questions with faithful and truthful chains of reasoning. arXiv preprint arXiv:2210.12217, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.597, + 0.906, + 0.65 + ], + "angle": 0, + "content": "[21] Xiaohan Wang, Yuhui Zhang, Orr Zohar, and Serena Yeung-Levy. Videoagent: Long-form video understanding with large language model as agent. arXiv preprint arXiv:2403.10517, 2024. 2, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.652, + 0.906, + 0.718 + ], + "angle": 0, + "content": "[22] Ziyang Wang, Shoubin Yu, Elias Stengel-Eskin, Jaehong Yoon, Feng Cheng, Gedas Bertasius, and Mohit Bansal. Videotree: Adaptive tree-based video representation for llm reasoning on long videos. arXiv preprint arXiv:2405.19209, 2024. 2, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.721, + 0.906, + 0.789 + ], + "angle": 0, + "content": "[23] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824-24837, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.79, + 0.906, + 0.831 + ], + "angle": 0, + "content": "[24] Nathaniel Weir and Benjamin Van Durme. Dynamic generation of grounded logical explanations in a neuro-symbolic expert system. arXiv preprint arXiv:2209.07662, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.833, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[25] Hao Wu, Huabin Liu, Yu Qiao, and Xiao Sun. Dibs: Enhancing dense video captioning with unlabeled videos via pseudo boundary enrichment and online refinement. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18699-18708, 2024. 1" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "3270" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.161 + ], + "angle": 0, + "content": "[26] Junbin Xiao, Xindi Shang, Angela Yao, and Tat-Seng Chua. Next-qa: Next phase of question-answering to explaining temporal actions. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9777–9786, 2021. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.164, + 0.482, + 0.232 + ], + "angle": 0, + "content": "[27] Junbin Xiao, Angela Yao, Yicong Li, and Tat-Seng Chua. Can i trust your answer? visually grounded video question answering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13204-13214, 2024. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.234, + 0.482, + 0.302 + ], + "angle": 0, + "content": "[28] Yu Yang, Besmira Nushi, Hamid Palangi, and Baharan Mirzasoleiman. Mitigating spurious correlations in multimodal models during fine-tuning. In International Conference on Machine Learning, pages 39365-39379. PMLR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.304, + 0.482, + 0.373 + ], + "angle": 0, + "content": "[29] Shunyu Yao, Dian Yu, Jeffrey Zhao, Izhak Shafran, Tom Griffiths, Yuan Cao, and Karthik Narasimhan. Tree of thoughts: Deliberate problem solving with large language models. Advances in Neural Information Processing Systems, 36, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.375, + 0.482, + 0.431 + ], + "angle": 0, + "content": "[30] Shoubin Yu, Jaemin Cho, Prateek Yadav, and Mohit Bansal. Self-chained image-language model for video localization and question answering. Advances in Neural Information Processing Systems, 36, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.432, + 0.482, + 0.487 + ], + "angle": 0, + "content": "[31] Ce Zhang, Taixi Lu, Md Mohaiminul Islam, Ziyang Wang, Shoubin Yu, Mohit Bansal, and Gedas Bertasius. A simple llm framework for long-range video question-answering. arXiv preprint arXiv:2312.17235, 2023. 2, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.488, + 0.482, + 0.557 + ], + "angle": 0, + "content": "[32] Gengyuan Zhang, Yurui Zhang, Kerui Zhang, and Volker Tresp. Can vision-language models be a good guesser? exploring vlms for times and location reasoning. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 636-645, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.559, + 0.482, + 0.613 + ], + "angle": 0, + "content": "[33] Hang Zhang, Xin Li, and Lidong Bing. Video-llama: An instruction-tuned audio-visual language model for video understanding. arXiv preprint arXiv:2306.02858, 2023. 1, 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.613 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.515, + 0.956 + ], + "angle": 0, + "content": "3271" + } + ] +] \ No newline at end of file diff --git a/2025/Commonsense Video Question Answering through Video-Grounded Entailment Tree Reasoning/e1080574-1050-4c6e-8509-597d8efaff00_origin.pdf b/2025/Commonsense Video Question Answering through Video-Grounded Entailment Tree Reasoning/e1080574-1050-4c6e-8509-597d8efaff00_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..e0f8327053b23e87b2942727052742d37d66d4c3 --- /dev/null +++ b/2025/Commonsense Video Question Answering through Video-Grounded Entailment Tree Reasoning/e1080574-1050-4c6e-8509-597d8efaff00_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3d9d12af5aa15823cf98d60d97a1e45d326e3e83dfd525db5499feced16831e2 +size 4752699 diff --git a/2025/Commonsense Video Question Answering through Video-Grounded Entailment Tree Reasoning/full.md b/2025/Commonsense Video Question Answering through Video-Grounded Entailment Tree Reasoning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..e8665d021835c10bddf34ff6eb9842cb1a7c4e6d --- /dev/null +++ b/2025/Commonsense Video Question Answering through Video-Grounded Entailment Tree Reasoning/full.md @@ -0,0 +1,245 @@ +# Commonsense Video Question Answering through Video-Grounded Entailment Tree Reasoning + +Huabin Liu $^{1,3*}$ , Filip Ilievski $^{2}$ , Cees G. M. Snoek $^{3}$ $^{1}$ Shanghai Jiao Tong University $^{2}$ Vrije Universiteit Amsterdam $^{3}$ University of Amsterdam huabinliu@sjtu.edu.cn, f.ilievski@vu.nl, c.g.m.snoek@uva.nl + +# Abstract + +This paper proposes the first video-grounded entailment tree reasoning method for commonsense video question answering (VQA). Despite the remarkable progress of large visual-language models (VLMs), there are growing concerns that they learn spurious correlations between videos and likely answers, reinforced by their black-box nature and remaining benchmarking biases. Our method explicitly grounds VQA tasks to video fragments in four steps: entailment tree construction, video-language entailment verification, tree reasoning, and dynamic tree expansion. A vital benefit of the method is its generalizability to current video- and image-based VLMs across reasoning types. To support fair evaluation, we devise a de-biasing procedure based on large-language models that rewrites VQA benchmark answer sets to enforce model reasoning. Systematic experiments on existing and de-biased benchmarks highlight the impact of our method components across benchmarks, VLMs, and reasoning types. + +# 1. Introduction + +This paper proposes a video-grounded reasoning method for commonsense video question answering (VQA). VQA has a long tradition in computer vision [11, 16, 25, 26], with remarkable recent progress obtained through video-and image-language models [10, 12, 14, 17, 33] (throughout this paper collectively referred to as vision-language models, or VLMs). Yet, there are growing concerns that their improved performance is based on learning shortcut associations between videos and likely answers, as opposed to reasoning [27]. Such concerns are reinforced by the black-box nature of these models [12, 14], which prohibits a deeper understanding of their decision-making process. + +We are inspired by recent work in natural language processing, where entailment trees have emerged as a mechanism to explicitly analyze answer candidates, using LLMs + +![](images/d1fff7e0f875c0e1b0d4fdbcbc480d3ce9de73267e40e372e9b2108e926a526f.jpg) +Figure 1. Given a video questioning answering task, our framework performs explicit reasoning over an entailment tree, where answer options are transformed into statements. These statements are then recursively decomposed and verified based on videotegued evidence relevant to the question. + +to recursively decompose a candidate into hypotheses and natural language inference formalisms to evaluate the hypotheses [4]. Entailment trees provide an explicit reasoning chain that explains the model's decision-making process and enables verification of each step, thus addressing concerns about shortcut learning. Recently, Sanders et al. [19] have devised a mechanism to apply entailment trees to videos. However, their work assumes video transcripts are explicitly provided to evaluate answers, thus avoiding the complexity of grounding hypotheses into video content. In this paper, we propose the first video-grounded entailment tree reasoning method for commonsense VQA. + +Our method explicitly grounds VQA tasks to video fragments in four steps: (i) entailment tree construction, (ii) video-language entailment verification, (iii) tree reasoning, and (iv) dynamic tree expansion. As shown in Fig. 1, given + +a video and a multiple-choice question, we generate a statement for each answer candidate that acts as a first-level hypothesis. We decompose each statement iteratively, aiming to produce sub-statements that can be confidently verified in the video. The video is itself decomposed into partitions, consisting of sets of frames. Verifying each statement is then a matter of aligning it to a video partition. A vital benefit of the method is its generalizability to current video and image-based VLMs across reasoning types, including temporal and causal. To demonstrate its video reasoning ability, we develop an answer-set de-bias procedure supported by an LLM that ensures that VQA benchmarks [11, 26] are adequate for reasoning in videos without relying on spurious correlations. Our experiments show that our video-grounded entailment tree method consistently improves video- and image-based baselines on both the existing and de-biased benchmarks. Moreover, it performs on par with, and often better than, state-of-the-art video-based VLMs while leveraging $257 \times$ fewer parameters. Further ablations show that the method benefits from considering both textual and video information and that its performance is especially strong on causal and temporal questions. + +# 2. Related work + +Video question answering. Recent research has shown that while video-based VLMs can achieve state-of-the-art performance, their answers are sensitive to object size, positioning, and speed [1, 28]. Moreover, when answering temporal and spatial questions, VLMs rely on textual biases to "guess" answers rather than performing genuine understanding and reasoning over visual-text information [32]. + +To improve the robustness and interpretability of VLMs, one line of research enriches VLMs with visual grounding functionality during QA, which enables VLMs to localize relevant video moments [18, 27, 30] or key frames [15, 22] to support answers. However, while these methods localize visual evidence, the process by which VLMs use it to deduce answers remains opaque. Another approach leverages external LLMs as reasoners or agents to enhance interpretability in textual modality. For instance, LLoVi [31] converts VQA to a text-based QA task via video captioning, then prompts an LLM to provide answers. Similarly, VideoAgent [21] uses an LLM to recursively determine if the current frames can answer the given question based on their textual descriptions. However, these methods heavily rely on the reasoning capabilities of LLMs. Like VLMs, the LLM reasoning process remains a black box, and hallucinations are common. Recently, TV-trees [19] attempted to perform explicit reasoning over both visual and textual modalities using a neuro-symbolic system. However, their work assumes video transcripts are explicitly provided to evaluate answers, thus avoiding the complexity of grounding hypotheses into video content. Instead, we contribute a + +general framework for explicit reasoning in commonsense VQA, fueled by a grounding component that aligns question components with video fragments. + +Beyond methodology, some works focus on providing fair and comprehensive VLM evaluations in VQA tasks by creating new benchmarks [3, 6, 9]. These benchmarks contain videos with diverse scenarios and durations, with carefully crafted questions and options designed to prevent textual shortcuts that VLMs might exploit. Video-specific questions (e.g., compositional action reasoning) [1], which require insights beyond textual associations, are included to test commonsense reasoning in VLMs. Addressing concerns of remaining biases in such benchmarks [6], we contribute an LLM-based answer-set de-biasing procedure to ensure that VQA benchmarks [11, 26] are adequate to evaluate reasoning in videos rather than spurious correlations. + +Systematic language reasoning. As LLMs demonstrate great potential in reasoning, there has been considerable interest in using LLMs to generate systematic explanations to support their answers. The series of Chain-of-Thought prompting [2, 23, 29] encourages LLMs to think step-by-step to perform explicit multi-hop reasoning, providing free-form reasoning steps before arriving at an answer. However, such implicit explanations are not grounded in external knowledge or evidence, which may lead to unverifiable and unfaithful reasoning. Since the development of EntailmentBank [4], research has increasingly focused on constructing explanation trees [20, 24] and graphs [8], encouraging models to generate step-wise entailment proofs of a statement using a set of supporting facts. Entailer [20] introduced this systematic explanation framework into language-based multiple-choice QA, performing explicit reasoning by generating entailment trees grounded in the model's internal beliefs. REFLEX [8] extends the entailment tree to form a belief graph for QA models, aiming to address consistency issues by intervening in the intermediate reasoning steps. Instead of grounding facts in predefined rules or model beliefs, NELLIE [24] adopts Prolog-based inference engines and external natural language corpora to build entailment trees as explainable reasoning for multiple-choice QA tasks. While such techniques for natural language processing inspire our framework, we generalize entailment trees to VQA, contributing a novel grounding method that aligns entailment trees with video fragments. + +# 3. Video-grounded entailment tree reasoning + +This paper devises a novel explainable framework for grounded commonsense VQA. It derives the answers through systematic reasoning over video-text information with entailment trees. Specifically, in the entailment tree (Fig. 2a), each candidate answer is decomposed into statements that entail the answer, explaining why each answer could be plausible. These statements are then grounded + +in relevant visual evidence from the video to prove or refute them (Fig. 2b). While entailment trees in natural language processing are constructed based on a model's internal knowledge or corpora [8, 20, 24], we ground entailment trees into video fragments. Finally, backtracking through the entailment tree leads to systematic reasoning over the statements (Fig. 3). Thus, answers can be deduced by a systematic structure with explicit reasoning paths and explanations rather than relying on opaque, black-box models. + +# 3.1. Entailment tree construction + +Initial statement generation. Given a question and its answer candidates, we first convert each question-answer pair into a declarative sentence that preserves the semantic meaning of the original QA pair. As a result, an $N$ -way multiple-choice QA problem produces a set of statements, denoted as $\mathcal{D} = d_{1},\ldots ,d_{N}$ . For example, the two-way question "What did the boy in white do after he first took the balloon? (A) resting on a chair (B) carries it toward the hula hoop" is transformed into: $\mathcal{D}:\{d_1 =$ "The boy in white resting on a chair after first taking the balloon.", $d_{2} =$ "The boy in white carries it toward the hula hoop after first taking the balloon." $\}$ . Thus, selecting the best answer equals identifying the correct statement for a given video. + +Recursive statement decomposition. For each initial statement in $\mathcal{D}$ , we generate two sub-statements as proofs that support the statement: Statement $\Leftarrow$ Sub-statement1, Sub-statement2. The statement is True if and only if both its sub-statements are proved to be True, i.e., the sub-statements entail the statement. Proving the original statements is thus translated into proving two simpler sub-statements. This procedure is recursive: the sub-statements can be further decomposed into further sub-statements that entail them. Therefore, to construct an entailment tree, we recursively decompose these sub-statements as new statements in the next tree layer until reaching the maximum depth or meeting the stop criterion. Fig. 2(a) presents an example of entailment tree generation. + +We leverage LLM prompting for both the initial statement generation and the statement decomposition, as these are linguistic tasks (see implementation details). + +# 3.2. Video-language entailment verification + +Given the entailment tree, the framework then verifies language statements based on the grounded video content as evidence. Specifically, each statement in the entailment tree must be proven or refuted by analyzing the video. A straightforward solution is to encode the whole video to collect information that can be used to verify the statement. However, the critical visual evidence that accurately verifies a statement tends to exist in a local moment instead of the whole video. Therefore, we develop a novel video grounding that guides the verification process to the moments with relevant visual evidence. + +Question-aware video captioning. Given a video, we convert its visual information into detailed textual information. Specifically, we input video frames into a VLM-based captioner $\mathrm{Cap}(\cdot)$ to obtain a caption $c_{i} = \mathrm{Cap}(f_{i})$ for each frame. However, captioning frames individually can overlook essential details or introduce irrelevant information for VQA. In commonsense VQA, questions often focus on specific facts already observed in the video. For example, a typical temporal reasoning question is "What happened before/after Event-A?" where Event-A refers to a fact statement about an event in the video. The fact referenced by the question can be leveraged to guide video understanding. To this end, we first extract the anchor fact indicated by the question and provide it to $\mathrm{Cap}(\cdot)$ as prior knowledge, encouraging the generation of relevant captions. Moreover, captions from all previous frames are also provided for each current frame to ensure $\mathrm{Cap}(\cdot)$ captures the temporal context from the past. This process is formulated as: + +$$ +c _ {i} = \operatorname {C a p} \left(f _ {i} \mid F, \left(c _ {1}, \dots , c _ {i - 1}\right)\right), \tag {1} +$$ + +where $F$ indicates the fact statement. + +Video evidence grounding. For commonsense VQA, depending on how the question reasons around the fact statement, the necessary evidence for answers can be gathered from specific video moments. For instance, in the case of temporal reasoning (e.g., before or after questions), the answer should be inferred from moments occurring either before or after the time of the relevant fact. Following this intuition, we design a two-step evidence-grounding strategy to localize the critical moments for answering. + +First, given the frame-wise captions, we retrieve a keyframe deemed most relevant to the fact statement, which we refer to as the anchor frame. A straightforward retrieval approach would involve comparing each $c_i$ with the fact description using specific metrics to identify the anchor frame. However, we enhance retrieval accuracy by adopting a structured semantic retrieval strategy. Specifically, the textual descriptions of each frame and fact statement are converted into structured triplets. These triplets capture the attributes and relationships of objects in each frame through structured semantics. As shown in Fig. 2(a), rather than directly comparing raw textual descriptions of frames and fact statements, we use these triplets for retrieval. Inspired by the success of using LLMs for retrieval tasks, we prompt an LLM to conduct anchor frame retrieval using the triplets of the fact statement as the query. The LLM then identifies and returns the most relevant frame ID, i.e., its timestamp. + +$$ +t _ {\text {a n c h o r}} = \operatorname {R t v} \left(c _ {i}, F\right), \tag {2} +$$ + +where $t_{\mathrm{anchor}}$ is the time stamp of the anchor frame, $\mathrm{Rtv}(\cdot)$ denotes the retrieval process. Second, we determine the final moment where we should look centered on the $t_{\mathrm{anchor}}$ , to + +![](images/9f3359d2c101bb0a8e86b760c57a1259094279cacb3e42086b24c0b83e63a8be.jpg) +Figure 2. Overview of our framework. (a) The generation of the entailment tree, where statements are recursively decomposed until the tree reaches its max depth or meets the stop criterion. (b) The process of video-language entailment verification: the input video is first converted into textual descriptions. Each caption is then parsed into structured semantics. Given the fact statement as a query, we retrieve the anchor frame. Then, based on the temporal or causal navigation indicated by questions, the visual evidence moment can be grounded. + +incorporate the temporal relations present in the question. Therefore, based on the anchor frame, the navigation for the moment is selected from "look ahead, look behind, look around" by considering the question: + +$$ +\mathcal {M} = \operatorname {G n d} \left(t _ {\text {a n c h o r}} \mid \mathcal {Q}\right), \tag {3} +$$ + +where $\operatorname{Gnd}(\cdot)$ is the grounding process and $\mathcal{M}$ denotes the grounded continuous interval in the video. Then, frames are resampled from the video within $M$ and used as visual evidence proving or refuting entailment tree statements. + +Visual-text statement prover. Given grounded visual evidence $\mathcal{M}$ of the video, statements are estimated to be true or false. Specifically, we employ a VLM as the statement prover, denoted as $\mathrm{Prv}(\cdot)$ , to evaluate each statement within the tree by probing VLM's internal belief on this statement. Each statement will be transformed into a binary QA task, with possible options being True or False. We then directly probe the $\mathrm{Prv}(\cdot)$ with the binary QA prompt and use the next token prediction probabilities of the words to elicit the model's belief. We normalize the prediction logits of the two options to get the confidence score of that statement. The above process is formulated as: + +$$ +s _ {d} = \Pr \operatorname {r v} (\mathcal {M}, h), s \in [ 0, 1 ], \tag {4} +$$ + +![](images/d3b28f2f6252dbcf4a3b151041a2110b13482ff2e7af4b3cef3f8eba367f972a.jpg) +Figure 3. Illustration of dynamic tree generation and backtrace. In Step-3, when the proof score of the left statement calculated from its child nodes is less than its direct score $(0.63 < 0.8)$ , its decomposition is pruned and stops. + +where $\mathcal{M}$ is the grounded moment and $h$ denotes the statement that needs to be verified. + +# 3.3. Dynamic entailment tree expansion + +So far, we have performed statement decomposition recursively to construct an entailment tree with pre-defined depth. However, not all statements need to be verified recursively, especially those easily determined to be true or false by VLMs. Moreover, as the depth increases, some statement sentences are atomic and directly verifiable. Thus, to improve the efficiency of the reasoning process, we further adopt a strategy to expand the entailment tree dynamically. Specifically, each statement $d$ is tied with two confidence scores provided by the $\operatorname{Prv}(\cdot)$ : + +![](images/de93e1108f78b4b924bb625a222dfe4a9d1eb3d32f5f2397f848d19e9d5719f3.jpg) +Figure 4. Illustration of commonsense bias in video question answering. The example is selected from the NExT-QA dataset. + +(1) The direct score $s_d$ , which indicates the belief of $\operatorname{Prv}(\cdot)$ model in $d$ . +(2) The proof score $s_p$ , denoting how confidently the model can prove $d$ , is calculated by multiplying the scores of its direct sub-statements. + +For a statement $d$ , the goal of decomposition is to establish a more reliable and convincing proof path than merely evaluating whether $d$ is true by VLMs. If the decomposition-based reasoning can prove $d$ with higher confidence than its direct score, the overall confidence for statement $d$ should increase. Otherwise, the decomposition should be disregarded. Thus, in the dynamic tree expansion, if decomposition does not enhance a statement's score, it is pruned, and that statement node becomes a leaf in the entailment tree. + +Fig. 3 presents a toy example. This criterion ensures that only beneficial decompositions are retained, significantly enhancing the tree reasoning process's efficiency. + +# 3.4. Reasoning over the entailment tree + +Finally, we perform a backtrace through the entailment tree to calculate the confidence score of each top statement. Specifically, the final score for each statement is produced by comparing its direct score $s_d$ and proof score $s_p$ , i.e., $s = \max(s_d, s_p)$ during backtrace (as shown in Fig. 3). The overall framework selects the answer corresponding to the statement at the top layer with top-scoring proof. + +# 4. De-biasing commonsense VQA answer sets + +To demonstrate the reasoning ability of video-grounded entailment trees, it is essential to evaluate using commonsense VQA benchmarks that enforce model reasoning. Recent work [13, 18, 27] has provided evidence that shortcuts + +![](images/e3a2bf99076e33621d676eb4bb7041017cbde2dfe28a52b3a509b12f8f37a998.jpg) +Figure 5. Prompt used for rewriting answers on NExT-QA. + +are present in VQA datasets which enables VLMs to solve these tasks based on textual associations rather than video-grounded reasoning. While VQA benchmarks increasingly focus on commonsense reasoning skills, such as temporal (e.g. after, before) or causal (how, why, what if) relationships in video content, reasoning shortcuts affect the validity of their evaluation. This is illustrated in Fig. 4 (top), where the correct answer (D) is much more relevant to the question and also aligns best with real-world expectations. Consequently, a VLM (VideoLLaVA [14] used in this example) can answer this question correctly by leveraging such associations and without analyzing the video content. Meanwhile, replacing the answer set distractors with other commonsensical answer candidates, as illustrated in Fig. 4 (bottom), makes this task challenging for VLMs. Here, a VLM switches its answer incorrectly to option (C), which confirms the impact of commonsense associations and the lack of grounded reasoning by these models. + +To this end, we devise a de-biasing procedure that mitigates reasoning shortcuts in commonsense VQA answer sets. Our de-biasing procedure transforms multiple-choice VQA benchmarks (e.g., NExT-QA) by rewriting their answer distractors while keeping their question and ground-truth answer intact. We prompt an LLM (LLaMA-3) to implement the rewriting procedure for each original QA set. Fig. 5 shows the detailed prompt we used for LLaMA-3 on NExT-QA dataset. This procedure ensures that (1) the answers cannot be easily derived from the QA set associations and (2) the answer remains consistent with the original QA pair. Thus, our procedure enables the scalable construction of de-biased QA sets by leveraging the commonsense associations in LLMs. The next section, focusing on experimental evaluation, analyzes the application of de-biasing to various datasets and its impact on the performance of VLMs, with and without entailment tree reasoning. + +# 5. Experiments + +# 5.1. Experimental setup + +Datasets. We test our framework on three VQA benchmarks: (1) NExT-QA [26], a VQA benchmark for causal and temporal reasoning. (2) IntentQA [11], which focuses on video intent reasoning from both causal and temporal aspects. (3) Video-MME [6], which is a recently proposed comprehensive evaluation benchmark for video analysis; we use its "short-term" split (video length $< 2$ mins) and 4 question types (temporal, spatial, action, and object reasoning) highly related to commonsense reasoning are selected. + +Evaluation. We report model performances on our rewritten test set for each dataset and its original test set. We evaluate our framework on all datasets under the multiple-choice QA setting, using a standard accuracy metric. + +Baselines. Our baselines represent three categories: + +- Video-based VLMs: Video-based VLMs are widely used for VQA tasks, so we include Video-LLaVA [14], VideoChat2 [12], and VideoLLaMA [33]. To test the effectiveness of our framework, we integrate these VLMs by replacing our $\mathsf{Prv}(\cdot)$ with specific VLM models. +- Image-based VLMs: We include BLIP-2 [10] and LLaVA-1.5 [17] as Image-LLM baselines. +- State-of-the-art VQA approaches: Recent works in VQA, such as VideoTree [22], VideoAgent [21], and LLoVi [31], are included as strong baselines. + +Implementation details. Our entire framework is training-and human annotation-free. We use LLaMA-3-8B [5] to handle basic functionalities, including (1) converting original QA into declarative statements, (2) statement decomposition, (3) structured semantic extraction and retrieval, and (4) guiding evidence grounding. Detailed prompts for each functionality are provided in the supplementary material. For frame-wise captioning across all datasets, we use LLaVA-1.5 [17] as our default captioner. When comparing with state-of-the-art VQA methods (e.g., VideoAgent), we follow their setup by replacing the captioner with the stronger CogAgent [7] model for fair comparison. When integrating our framework with VLMs, the VLM itself serves as the Prover, $\mathsf{Prv}(\cdot)$ . For video-LLMs, frames are uniformly sampled from the grounded video moment to meet their input requirements (VideoChat2: 16 frames; VideoLLaVA & VideoLLaMA: 8 frames). For image-language models like LLaVA, we sample 8 frames from the grounded moment and process them individually; the final confidence score is obtained by averaging scores across frames. During dynamic tree generation, we also set a max depth of 5 for the overall entailment tree to improve the efficiency. + +# 5.2. Main results + +Benefit for image and video-based VLMs. Tab. 1 summarizes the results of our method compared to baselines. Integrating entailment tree reasoning brings consistent im + +provement across the video- and image-based VLMs for all datasets (1-4% on average). This finding includes the recently proposed benchmark VideoMME, which poses much more challenging videos and questions. The benefits are particularly regular for temporal reasoning (improvement in 14 out of 15 cases), which illustrates how our explicit reasoning process enhances temporal commonsense QA. Image-based VLMs, which initially lack temporal modeling capabilities, perform poorly when directly applied to video QA tasks. However, our framework provides a significant performance boost for these models up to $8\%$ for LLAVA-1.5. By reasoning over multiple sub-problems rather than tackling the entire complex question simultaneously, our reasoning method makes the task more manageable for both video- and image-based VLMs. + +Results on de-biased QA sets. As shown in Tab. 2, all models experience considerable performance drops on the de-biased set of the same VQA dataset, which aligns with our observation that current VLMs often rely on textual bias in commonsense reasoning tasks. Notably, video-based VLMs show an $8\% -10\%$ decrease in the de-biased set even though the question and the correct answer remain unchanged. In contrast, our proposed framework, which derives answers through an explicit reasoning process based on specific visual evidence, demonstrates much greater robustness on the de-biased set. The improvement brought by our framework on the de-biased set is even higher than on the original test sets. In turn, our framework compensates for the performance loss of the VLMs on the de-biased set. This analysis underscores our framework's potential to mitigate textual bias in commonsense reasoning. Furthermore, the performance differences between the original and de-biased QA sets highlight VQA benchmarks' limitations in evaluating VLMs' true reasoning abilities. + +Comparison with state-of-the-art. Tab. 3 compares our framework's results on the de-biased sets to state-of-the-art VQA approaches. The table shows that, next to the consistent benefit our framework provides to various VLMs, it is also competitive with state-of-the-art VQA methods. When applying an advanced captioner and reasoner that aligns with VideoAgent and VideoTree, our framework yields new state-of-the-art results in some cases. In particular, our framework performs best on temporal reasoning questions for all three benchmarks and outperforms all methods on the IntentQA dataset. Importantly, our method reaches such competitive performance despite using $257 \times$ fewer parameters for its reasoning compared to state-of-the-art methods. + +# 5.3. Ablation studies + +Ablation experiments are conducted using VideoLLaVA's baseline with our framework. Results are reported on the test set of the NExT-QA dataset. + +Impact of LLMs for statement decomposition. Entailment generation in our framework relies on prompting an + +
ModelNExT-QAIntentQAVideoMMEAvg
TemporalCausalTemporalCausalTemporalSpatialActionObject
Image-based VLMsBLIP-2 [10]38.336.143.848.625.426.924.228.634.0
+Ours45.341.848.952.530.527.427.728.837.9
LLaVA-1.5 [17]37.840.745.850.031.133.627.729.837.1
+Ours45.647.948.454.736.736.831.430.741.5
Video-based VLMsVideoChat2 [12]56.962.160.463.250.352.449.550.155.6
+Ours57.861.662.363.852.851.551.350.056.4
VideoLLaVA [14]56.060.453.960.747.644.346.249.752.3
+Ours58.362.757.661.848.346.149.850.354.2
VideoLLaMA [33]55.460.255.156.744.847.244.749.351.7
+Ours58.160.454.558.947.447.848.649.153.1
+ +Table 1. Impact on image and video-based VLMs on the original NExT-QA, IntentQA, and VideoMME test sets. Our framework increases accuracy of all video- and image-based VLMs by $1 - 4\%$ on average across all data partitions. Temporal and action partitions benefit most. + +
ModelBLIP-2+OursLLaVA+OursVideo Chat2+OursVideo LLaVA+OursVideo LLaMA+Ours
NExT-QAOriginal37.243.639.346.859.559.758.260.557.859.3
Rewritten33.539.834.844.945.449.051.155.441.447.0
Intent-QAOriginal46.250.747.951.661.863.157.359.755.956.7
Rewritten38.245.542.748.652.655.750.554.746.350.0
Avg38.844.941.248.054.856.954.357.650.453.3
+ +Table 2. Results on de-biased QA sets. Video-based VLMs show significant decreases in the rewritten de-biased set. In contrast, our framework demonstrates much greater robustness on the rewritten set. + +external LLM to recursively decompose statements (cf. Sec. 3.1), which is crucial in guiding reasoning paths. Consequently, we tested various LLMs for entailment tree generation, including open-source models (LLama-3 and Mistral) of different sizes and proprietary LLMs (GPT-4 and Gemini-1.5). The results are summarized in Tab. 4. As expected, the proprietary model GPT-4, known for its strong step-by-step reasoning capabilities, delivers the best performance across all settings. Scaling up LLaMA-3 to 70B offers improvements over the 8B model, though with a notable increase in inference time. As the overall performance difference between models is within $1\%$ , we select LLaMA-3-8B as the default for integrating our framework into VLMs due to its free availability and efficiency. + +Ablation on grounding components. Next, we test the effectiveness of each component in our grounding module (Sec. 3.2). The results, summarized in Tab. 5, indicate that both fact-conditional captioning and structure-guided retrieval enhance overall performance by improving grounding accuracy. However, using only structure-guided retrieval results in a slight performance drop, possibly because $\mathsf{Cap}(\cdot)$ introduces irrelevant semantic information that doesn't align with the question's focus, and the structured representation can make identifying anchor frames more challenging. In contrast, fact-conditional captioning alone yields substantial improvement, demonstrating that this straightforward approach can yield an effective + +and more controllable textual description for videos by conditioning on prior knowledge or relevant facts. + +Impact of length of video frames. We further ablate the impact of input video frame length in our framework to determine the optimal number of frame-wise captions to generate per video. The results, summarized in Tab. 6, show that ideal performance is achieved only when sampling a sufficient number of frames (at least 16 for NExTQA). When fewer frames are used (e.g., 4 or 8), key anchor frames may be missed, reducing the accuracy of grounded visual evidence. Additionally, while increasing the frame count to 32 yields the best performance, it also increases the calls required for $\mathrm{Cap}(\cdot)$ to generate frame-wise captions. Balancing efficiency with performance gains, we set 24 frames as the default in our implementation. + +Effectiveness of evidence grounding. Our method grounds relevant video fragments to support statements in the entailment tree (Sec. 3.2). To validate its effectiveness, we compare it with two other variations as sources of visual evidence: (1) without evidence grounding, using the full video as evidence, and (2) upper-bound results: manually annotated temporal boundaries provided in the NExT-GQA dataset, indicating where the QA models should focus when producing correct answers. The results are shown in Tab. 7. Compared to the baseline, our video-grounded method provides consistent improvements across the original and debiased sets. The improvement is more apparent in the de + +
MethodModel ReasonerNExT-QA*IntentQA*VideoMME
TemporalCausalTemporalCausalTemporalSpatialActionObject
VideoAgent [21]GPT-4 (1.8T)58.266.660.461.0----
VideoTree [22]GPT-4 (1.8T)60.266.456.760.155.754.354.252.6
LLoVi [31]GPT-4 (1.8T)53.160.858.661.852.255.351.850.8
OursVideoLLAVA (7B)60.865.961.062.655.953.854.050.8
+ +Table 3. Comparison with state-of-the-art. Results for NExT-QA and IntentQA are reported under the de-biased set (the results on the original sets are similar; we provide them in the Appendix). The 'Reasoner' in these approaches is similar to the "Prover" in our framework. The captioner for all methods is CogAgent [7]. Despite other methods relying on much stronger reasoning models, our approach yields competitive performance (four state-of-the-art results) and high parameter efficiency ( $257 \times$ fewer than GPT-4 reasoners). + +
Model classLLMNExT-QA
OriginalRewritten
Open-sourceMistral-7B60.055.6
LLaMA-3-8B60.555.4
LLaMA-3-70B61.355.9
ProprietaryGemini-1.5-Pro61.155.2
GPT-461.656.1
+ +Table 4. Impact of LLMs for statement decomposition. The open-source and proprietary models are ordered ascendingly by size. Larger models, especially GPT-4, are best at decomposition, but the smaller models (e.g., LlaMa-3-8B) come close. + +
ComponentsNExT-QA
Fact-conditioned captioningStructure-based retrievalOriginalRewritten
56.249.6
✓59.553.3
✓58.452.7
✓✓60.555.4
+ +Table 5. Ablation on grounding components, showing that both fact-conditional captioning and structure-guided retrieval enhance overall performance by improving grounding accuracy. + +
Acc +(NExT-QA)Frame number
48162432
Original57.759.360.560.761.0
Rewritten51.953.455.455.455.7
+ +biased set, where the answer options are more semantically similar and require more precise, discriminative visual evidence. Using the ground-truth fragment can further boost our approach, suggesting that enhancing grounding accuracy could further improve our framework. + +Effectiveness of dynamic tree expansion. The depth of the entailment tree determines the granularity of reasoning (Sec. 3.3). This ablation analyzes how tree depth impacts overall performance and compares a fixed-depth approach with our dynamic tree generation strategy. Increases + +Table 6. Impact of video frame amount. Strong performance requires a sufficiently high frame number (over 16 for NExT-QA). + +
Video fragmentFullGrounded (ours)GT
NExT-QAOriginal58.360.561.8
Rewritten51.755.456.9
+ +Table 7. Effectiveness of evidence grounding. Our videotegrounded method yields clear improvement over using the full video. More precise grounding can further enhance our accuracy. + +
StrategyStatic (Depth=)Dynamic
2345
NExT-QAOriginal58.859.260.260.360.5
Rewritten52.053.455.655.355.4
+ +Table 8. Effectiveness of dynamic tree expansion. It yields superior accuracy while increasing reasoning efficiency. + +ing the depth of reasoning yields significant improvements, as complex, long statements are broken down into concise sub-statements that VLMs can understand more effectively. However, extending reasoning beyond the 4th layer offers diminishing returns; for NExT-QA, the original statements' complexity constrains the task, and some 5th-layer sub-statements become overly simplistic and less effective for reasoning. This finding highlights the necessity of our dynamic strategy. Applying the dynamic tree expansion strategy, we can see that the performance outperforms the fixed-depth paradigm. In the meantime, the dynamic strategy increases the reasoning efficiency over the entailment tree, more details about efficiency comparison can be found in our supplementary material. + +# 6. Conclusion + +This paper proposed the first video-grounded entailment tree framework for VQA. Moreover, we also contributed a de-biasing procedure to avoid spurious correlations during evaluation and applied it to enhance representative benchmarks. Extensive experiments with five video- and image-based VLMs demonstrate consistent benefits of our method on these benchmarks. Besides, our proposed framework performs on par with state-of-the-art video reasoning methods despite using $257 \times$ fewer parameters. While de-biasing hurts VLM accuracy, our framework regains the accuracy losses and is competitive with state-of-the-art VQA methods. + +# References + +[1] Piyush Bagad, Makarand Tapaswi, and Cees GM Snoek. Test of time: Instilling video-language models with a sense of time. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2503-2516, 2023. 2 +[2] Maciej Besta, Nils Blach, Ales Kubicek, Robert Gerstenberger, Michal Podstawski, Lukas Gianinazzi, Joanna Gajda, Tomasz Lehmann, Hubert Niewiadomski, Piotr Nczyk, et al. Graph of thoughts: Solving elaborate problems with large language models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 17682-17690, 2024. 2 +[3] Tieyuan Chen, Huabin Liu, Tianyao He, Yihang Chen, Chaofan Gan, Xiao Ma, Cheng Zhong, Yang Zhang, Yingxue Wang, Hui Lin, et al. Mecd: Unlocking multi-event causal discovery in video reasoning. arXiv preprint arXiv:2409.17647, 2024. 2 +[4] Bhavana Dalvi, Peter Jansen, Oyvind Tafjord, Zhengnan Xie, Hannah Smith, Leighanna Pipatanangkura, and Peter Clark. Explaining answers with entailment trees. arXiv preprint arXiv:2104.08661, 2021. 1, 2 +[5] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024.6 +[6] Chaoyou Fu, Yuhan Dai, Yongdong Luo, Lei Li, Shuhuai Ren, Renrui Zhang, Zihan Wang, Chenyu Zhou, Yunhang Shen, Mengdan Zhang, et al. Video-mme: The first-ever comprehensive evaluation benchmark of multi-modal llms in video analysis. arXiv preprint arXiv:2405.21075, 2024. 2, 6 +[7] Wenyi Hong, Weihan Wang, Qingsong Lv, Jiazheng Xu, Wenmeng Yu, Junhui Ji, Yan Wang, Zihan Wang, Yuxiao Dong, Ming Ding, et al. Cogagent: A visual language model for gui agents. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14281-14290, 2024. 6, 8 +[8] Nora Kassner, Oyvind Tafjord, Ashish Sabharwal, Kyle Richardson, Hinrich Schuetze, and Peter Clark. Language models with rationality. arXiv preprint arXiv:2305.14250, 2023. 2, 3 +[9] Muhammad Uzair Khattak, Muhammad Ferjad Naeem, Jameel Hassan, Muzammal Naseer, Federico Tombari, Fahad Shahbaz Khan, and Salman Khan. Complex video reasoning and robustness evaluation suite for video-lmms. arXiv preprint arXiv:2405.03690, 2024. 2 +[10] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730–19742. PMLR, 2023. 1, 6, 7 +[11] Jiapeng Li, Ping Wei, Wenjuan Han, and Lifeng Fan. Intentqa: Context-aware video intent reasoning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11963-11974, 2023. 1, 2, 6 +[12] Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Yi Liu, Zun Wang, Jilan Xu, Guo Chen, Ping Luo, et al. + +Mvbench: A comprehensive multi-modal video understanding benchmark. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22195-22206, 2024. 1, 6, 7 +[13] Yicong Li, Xiang Wang, Junbin Xiao, Wei Ji, and Tat-Seng Chua. Invariant grounding for video question answering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2928-2937, 2022. 5 +[14] Bin Lin, Yang Ye, Bin Zhu, Jiaxi Cui, Munan Ning, Peng Jin, and Li Yuan. Video-llava: Learning united visual representation by alignment before projection. arXiv preprint arXiv:2311.10122, 2023. 1, 5, 6, 7 +[15] Huabin Liu, Weixian Lv, John See, and Weiyao Lin. Task-adaptive spatial-temporal video sampler for few-shot action recognition. In Proceedings of the 30th ACM International Conference on Multimedia, pages 6230-6240, 2022. 2 +[16] Huabin Liu, Weiyao Lin, Tieyuan Chen, Yuxi Li, Shuyuan Li, and John See. Few-shot action recognition via intra- and inter-video information maximization. arXiv preprint arXiv:2305.06114, 2023. 1 +[17] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26296-26306, 2024. 1, 6, 7 +[18] Huabin Liu, Xiao Ma, Cheng Zhong, Yang Zhang, and Weiyao Lin. Timecraft: Navigate weakly-supervised temporal grounded video question answering via bi-directional reasoning. In European Conference on Computer Vision, pages 92-107. Springer, 2025. 2, 5 +[19] Kate Sanders, Nathaniel Weir, and Benjamin Van Durme. Tv-trees: Multimodal entailment trees for neuro-symbolic video reasoning. arXiv preprint arXiv:2402.19467, 2024. 1, 2 +[20] Oyvind Tafjord, Bhavana Dalvi Mishra, and Peter Clark. Entailer: Answering questions with faithful and truthful chains of reasoning. arXiv preprint arXiv:2210.12217, 2022. 2, 3 +[21] Xiaohan Wang, Yuhui Zhang, Orr Zohar, and Serena Yeung-Levy. Videoagent: Long-form video understanding with large language model as agent. arXiv preprint arXiv:2403.10517, 2024. 2, 6, 8 +[22] Ziyang Wang, Shoubin Yu, Elias Stengel-Eskin, Jaehong Yoon, Feng Cheng, Gedas Bertasius, and Mohit Bansal. Videotree: Adaptive tree-based video representation for llm reasoning on long videos. arXiv preprint arXiv:2405.19209, 2024. 2, 6, 8 +[23] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824-24837, 2022. 2 +[24] Nathaniel Weir and Benjamin Van Durme. Dynamic generation of grounded logical explanations in a neuro-symbolic expert system. arXiv preprint arXiv:2209.07662, 2022. 2, 3 +[25] Hao Wu, Huabin Liu, Yu Qiao, and Xiao Sun. Dibs: Enhancing dense video captioning with unlabeled videos via pseudo boundary enrichment and online refinement. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18699-18708, 2024. 1 + +[26] Junbin Xiao, Xindi Shang, Angela Yao, and Tat-Seng Chua. Next-qa: Next phase of question-answering to explaining temporal actions. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9777–9786, 2021. 1, 2, 6 +[27] Junbin Xiao, Angela Yao, Yicong Li, and Tat-Seng Chua. Can i trust your answer? visually grounded video question answering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13204-13214, 2024. 1, 2, 5 +[28] Yu Yang, Besmira Nushi, Hamid Palangi, and Baharan Mirzasoleiman. Mitigating spurious correlations in multimodal models during fine-tuning. In International Conference on Machine Learning, pages 39365-39379. PMLR, 2023. 2 +[29] Shunyu Yao, Dian Yu, Jeffrey Zhao, Izhak Shafran, Tom Griffiths, Yuan Cao, and Karthik Narasimhan. Tree of thoughts: Deliberate problem solving with large language models. Advances in Neural Information Processing Systems, 36, 2024. 2 +[30] Shoubin Yu, Jaemin Cho, Prateek Yadav, and Mohit Bansal. Self-chained image-language model for video localization and question answering. Advances in Neural Information Processing Systems, 36, 2024. 2 +[31] Ce Zhang, Taixi Lu, Md Mohaiminul Islam, Ziyang Wang, Shoubin Yu, Mohit Bansal, and Gedas Bertasius. A simple llm framework for long-range video question-answering. arXiv preprint arXiv:2312.17235, 2023. 2, 6, 8 +[32] Gengyuan Zhang, Yurui Zhang, Kerui Zhang, and Volker Tresp. Can vision-language models be a good guesser? exploring vlms for times and location reasoning. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 636-645, 2024. 2 +[33] Hang Zhang, Xin Li, and Lidong Bing. Video-llama: An instruction-tuned audio-visual language model for video understanding. arXiv preprint arXiv:2306.02858, 2023. 1, 6, 7 \ No newline at end of file diff --git a/2025/Commonsense Video Question Answering through Video-Grounded Entailment Tree Reasoning/images.zip b/2025/Commonsense Video Question Answering through Video-Grounded Entailment Tree Reasoning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..4f10653f66d949e88ba8426c875d7c5f1f94c698 --- /dev/null +++ b/2025/Commonsense Video Question Answering through Video-Grounded Entailment Tree Reasoning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:481fba62f22e557c03702597f672195a483cd719dcdd0e071936127c34bcb3b7 +size 657324 diff --git a/2025/Commonsense Video Question Answering through Video-Grounded Entailment Tree Reasoning/layout.json b/2025/Commonsense Video Question Answering through Video-Grounded Entailment Tree Reasoning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..d634482c45c2c3f2aae025e65d23d8eb08cf36c5 --- /dev/null +++ b/2025/Commonsense Video Question Answering through Video-Grounded Entailment Tree Reasoning/layout.json @@ -0,0 +1,6487 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 96, + 103, + 515, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 103, + 515, + 140 + ], + "spans": [ + { + "bbox": [ + 96, + 103, + 515, + 140 + ], + "type": "text", + "content": "Commonsense Video Question Answering through Video-Grounded Entailment Tree Reasoning" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 71, + 159, + 539, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 159, + 539, + 203 + ], + "spans": [ + { + "bbox": [ + 71, + 159, + 539, + 203 + ], + "type": "text", + "content": "Huabin Liu" + }, + { + "bbox": [ + 71, + 159, + 539, + 203 + ], + "type": "inline_equation", + "content": "^{1,3*}" + }, + { + "bbox": [ + 71, + 159, + 539, + 203 + ], + "type": "text", + "content": ", Filip Ilievski" + }, + { + "bbox": [ + 71, + 159, + 539, + 203 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 71, + 159, + 539, + 203 + ], + "type": "text", + "content": ", Cees G. M. Snoek" + }, + { + "bbox": [ + 71, + 159, + 539, + 203 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 71, + 159, + 539, + 203 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 71, + 159, + 539, + 203 + ], + "type": "text", + "content": " Shanghai Jiao Tong University " + }, + { + "bbox": [ + 71, + 159, + 539, + 203 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 71, + 159, + 539, + 203 + ], + "type": "text", + "content": " Vrije Universiteit Amsterdam " + }, + { + "bbox": [ + 71, + 159, + 539, + 203 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 71, + 159, + 539, + 203 + ], + "type": "text", + "content": " University of Amsterdam huabinliu@sjtu.edu.cn, f.ilievski@vu.nl, c.g.m.snoek@uva.nl" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "spans": [ + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 256, + 296, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 256, + 296, + 473 + ], + "spans": [ + { + "bbox": [ + 55, + 256, + 296, + 473 + ], + "type": "text", + "content": "This paper proposes the first video-grounded entailment tree reasoning method for commonsense video question answering (VQA). Despite the remarkable progress of large visual-language models (VLMs), there are growing concerns that they learn spurious correlations between videos and likely answers, reinforced by their black-box nature and remaining benchmarking biases. Our method explicitly grounds VQA tasks to video fragments in four steps: entailment tree construction, video-language entailment verification, tree reasoning, and dynamic tree expansion. A vital benefit of the method is its generalizability to current video- and image-based VLMs across reasoning types. To support fair evaluation, we devise a de-biasing procedure based on large-language models that rewrites VQA benchmark answer sets to enforce model reasoning. Systematic experiments on existing and de-biased benchmarks highlight the impact of our method components across benchmarks, VLMs, and reasoning types." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 497, + 135, + 509 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 497, + 135, + 509 + ], + "spans": [ + { + "bbox": [ + 55, + 497, + 135, + 509 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 513, + 295, + 656 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 513, + 295, + 656 + ], + "spans": [ + { + "bbox": [ + 55, + 513, + 295, + 656 + ], + "type": "text", + "content": "This paper proposes a video-grounded reasoning method for commonsense video question answering (VQA). VQA has a long tradition in computer vision [11, 16, 25, 26], with remarkable recent progress obtained through video-and image-language models [10, 12, 14, 17, 33] (throughout this paper collectively referred to as vision-language models, or VLMs). Yet, there are growing concerns that their improved performance is based on learning shortcut associations between videos and likely answers, as opposed to reasoning [27]. Such concerns are reinforced by the black-box nature of these models [12, 14], which prohibits a deeper understanding of their decision-making process." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 657, + 296, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 657, + 296, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 657, + 296, + 693 + ], + "type": "text", + "content": "We are inspired by recent work in natural language processing, where entailment trees have emerged as a mechanism to explicitly analyze answer candidates, using LLMs" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 317, + 230, + 555, + 442 + ], + "blocks": [ + { + "bbox": [ + 317, + 230, + 555, + 442 + ], + "lines": [ + { + "bbox": [ + 317, + 230, + 555, + 442 + ], + "spans": [ + { + "bbox": [ + 317, + 230, + 555, + 442 + ], + "type": "image", + "image_path": "d1fff7e0f875c0e1b0d4fdbcbc480d3ce9de73267e40e372e9b2108e926a526f.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 449, + 555, + 504 + ], + "lines": [ + { + "bbox": [ + 313, + 449, + 555, + 504 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 555, + 504 + ], + "type": "text", + "content": "Figure 1. Given a video questioning answering task, our framework performs explicit reasoning over an entailment tree, where answer options are transformed into statements. These statements are then recursively decomposed and verified based on videotegued evidence relevant to the question." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 312, + 521, + 555, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 521, + 555, + 664 + ], + "spans": [ + { + "bbox": [ + 312, + 521, + 555, + 664 + ], + "type": "text", + "content": "to recursively decompose a candidate into hypotheses and natural language inference formalisms to evaluate the hypotheses [4]. Entailment trees provide an explicit reasoning chain that explains the model's decision-making process and enables verification of each step, thus addressing concerns about shortcut learning. Recently, Sanders et al. [19] have devised a mechanism to apply entailment trees to videos. However, their work assumes video transcripts are explicitly provided to evaluate answers, thus avoiding the complexity of grounding hypotheses into video content. In this paper, we propose the first video-grounded entailment tree reasoning method for commonsense VQA." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "content": "Our method explicitly grounds VQA tasks to video fragments in four steps: (i) entailment tree construction, (ii) video-language entailment verification, (iii) tree reasoning, and (iv) dynamic tree expansion. As shown in Fig. 1, given" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 70, + 702, + 285, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 702, + 285, + 713 + ], + "spans": [ + { + "bbox": [ + 70, + 702, + 285, + 713 + ], + "type": "text", + "content": "* Work was done as a visiting student at University of Amsterdam" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3262" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 57, + 72, + 294, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 72, + 294, + 335 + ], + "spans": [ + { + "bbox": [ + 57, + 72, + 294, + 335 + ], + "type": "text", + "content": "a video and a multiple-choice question, we generate a statement for each answer candidate that acts as a first-level hypothesis. We decompose each statement iteratively, aiming to produce sub-statements that can be confidently verified in the video. The video is itself decomposed into partitions, consisting of sets of frames. Verifying each statement is then a matter of aligning it to a video partition. A vital benefit of the method is its generalizability to current video and image-based VLMs across reasoning types, including temporal and causal. To demonstrate its video reasoning ability, we develop an answer-set de-bias procedure supported by an LLM that ensures that VQA benchmarks [11, 26] are adequate for reasoning in videos without relying on spurious correlations. Our experiments show that our video-grounded entailment tree method consistently improves video- and image-based baselines on both the existing and de-biased benchmarks. Moreover, it performs on par with, and often better than, state-of-the-art video-based VLMs while leveraging " + }, + { + "bbox": [ + 57, + 72, + 294, + 335 + ], + "type": "inline_equation", + "content": "257 \\times" + }, + { + "bbox": [ + 57, + 72, + 294, + 335 + ], + "type": "text", + "content": " fewer parameters. Further ablations show that the method benefits from considering both textual and video information and that its performance is especially strong on causal and temporal questions." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 58, + 350, + 138, + 361 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 350, + 138, + 361 + ], + "spans": [ + { + "bbox": [ + 58, + 350, + 138, + 361 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 365, + 294, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 365, + 294, + 449 + ], + "spans": [ + { + "bbox": [ + 58, + 365, + 294, + 449 + ], + "type": "text", + "content": "Video question answering. Recent research has shown that while video-based VLMs can achieve state-of-the-art performance, their answers are sensitive to object size, positioning, and speed [1, 28]. Moreover, when answering temporal and spatial questions, VLMs rely on textual biases to \"guess\" answers rather than performing genuine understanding and reasoning over visual-text information [32]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 451, + 294, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 451, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 451, + 294, + 712 + ], + "type": "text", + "content": "To improve the robustness and interpretability of VLMs, one line of research enriches VLMs with visual grounding functionality during QA, which enables VLMs to localize relevant video moments [18, 27, 30] or key frames [15, 22] to support answers. However, while these methods localize visual evidence, the process by which VLMs use it to deduce answers remains opaque. Another approach leverages external LLMs as reasoners or agents to enhance interpretability in textual modality. For instance, LLoVi [31] converts VQA to a text-based QA task via video captioning, then prompts an LLM to provide answers. Similarly, VideoAgent [21] uses an LLM to recursively determine if the current frames can answer the given question based on their textual descriptions. However, these methods heavily rely on the reasoning capabilities of LLMs. Like VLMs, the LLM reasoning process remains a black box, and hallucinations are common. Recently, TV-trees [19] attempted to perform explicit reasoning over both visual and textual modalities using a neuro-symbolic system. However, their work assumes video transcripts are explicitly provided to evaluate answers, thus avoiding the complexity of grounding hypotheses into video content. Instead, we contribute a" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 317, + 73, + 553, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 73, + 553, + 108 + ], + "spans": [ + { + "bbox": [ + 317, + 73, + 553, + 108 + ], + "type": "text", + "content": "general framework for explicit reasoning in commonsense VQA, fueled by a grounding component that aligns question components with video fragments." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 317, + 110, + 553, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 110, + 553, + 264 + ], + "spans": [ + { + "bbox": [ + 317, + 110, + 553, + 264 + ], + "type": "text", + "content": "Beyond methodology, some works focus on providing fair and comprehensive VLM evaluations in VQA tasks by creating new benchmarks [3, 6, 9]. These benchmarks contain videos with diverse scenarios and durations, with carefully crafted questions and options designed to prevent textual shortcuts that VLMs might exploit. Video-specific questions (e.g., compositional action reasoning) [1], which require insights beyond textual associations, are included to test commonsense reasoning in VLMs. Addressing concerns of remaining biases in such benchmarks [6], we contribute an LLM-based answer-set de-biasing procedure to ensure that VQA benchmarks [11, 26] are adequate to evaluate reasoning in videos rather than spurious correlations." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 317, + 266, + 553, + 600 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 266, + 553, + 600 + ], + "spans": [ + { + "bbox": [ + 317, + 266, + 553, + 600 + ], + "type": "text", + "content": "Systematic language reasoning. As LLMs demonstrate great potential in reasoning, there has been considerable interest in using LLMs to generate systematic explanations to support their answers. The series of Chain-of-Thought prompting [2, 23, 29] encourages LLMs to think step-by-step to perform explicit multi-hop reasoning, providing free-form reasoning steps before arriving at an answer. However, such implicit explanations are not grounded in external knowledge or evidence, which may lead to unverifiable and unfaithful reasoning. Since the development of EntailmentBank [4], research has increasingly focused on constructing explanation trees [20, 24] and graphs [8], encouraging models to generate step-wise entailment proofs of a statement using a set of supporting facts. Entailer [20] introduced this systematic explanation framework into language-based multiple-choice QA, performing explicit reasoning by generating entailment trees grounded in the model's internal beliefs. REFLEX [8] extends the entailment tree to form a belief graph for QA models, aiming to address consistency issues by intervening in the intermediate reasoning steps. Instead of grounding facts in predefined rules or model beliefs, NELLIE [24] adopts Prolog-based inference engines and external natural language corpora to build entailment trees as explainable reasoning for multiple-choice QA tasks. While such techniques for natural language processing inspire our framework, we generalize entailment trees to VQA, contributing a novel grounding method that aligns entailment trees with video fragments." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 316, + 614, + 544, + 627 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 614, + 544, + 627 + ], + "spans": [ + { + "bbox": [ + 316, + 614, + 544, + 627 + ], + "type": "text", + "content": "3. Video-grounded entailment tree reasoning" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 317, + 630, + 553, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 630, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 317, + 630, + 553, + 712 + ], + "type": "text", + "content": "This paper devises a novel explainable framework for grounded commonsense VQA. It derives the answers through systematic reasoning over video-text information with entailment trees. Specifically, in the entailment tree (Fig. 2a), each candidate answer is decomposed into statements that entail the answer, explaining why each answer could be plausible. These statements are then grounded" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 750, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 750, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 750, + 314, + 757 + ], + "type": "text", + "content": "3263" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 180 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 180 + ], + "type": "text", + "content": "in relevant visual evidence from the video to prove or refute them (Fig. 2b). While entailment trees in natural language processing are constructed based on a model's internal knowledge or corpora [8, 20, 24], we ground entailment trees into video fragments. Finally, backtracking through the entailment tree leads to systematic reasoning over the statements (Fig. 3). Thus, answers can be deduced by a systematic structure with explicit reasoning paths and explanations rather than relying on opaque, black-box models." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 189, + 212, + 200 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 189, + 212, + 200 + ], + "spans": [ + { + "bbox": [ + 55, + 189, + 212, + 200 + ], + "type": "text", + "content": "3.1. Entailment tree construction" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 201, + 295, + 356 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 201, + 295, + 356 + ], + "spans": [ + { + "bbox": [ + 55, + 201, + 295, + 356 + ], + "type": "text", + "content": "Initial statement generation. Given a question and its answer candidates, we first convert each question-answer pair into a declarative sentence that preserves the semantic meaning of the original QA pair. As a result, an " + }, + { + "bbox": [ + 55, + 201, + 295, + 356 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 201, + 295, + 356 + ], + "type": "text", + "content": "-way multiple-choice QA problem produces a set of statements, denoted as " + }, + { + "bbox": [ + 55, + 201, + 295, + 356 + ], + "type": "inline_equation", + "content": "\\mathcal{D} = d_{1},\\ldots ,d_{N}" + }, + { + "bbox": [ + 55, + 201, + 295, + 356 + ], + "type": "text", + "content": ". For example, the two-way question \"What did the boy in white do after he first took the balloon? (A) resting on a chair (B) carries it toward the hula hoop\" is transformed into: " + }, + { + "bbox": [ + 55, + 201, + 295, + 356 + ], + "type": "inline_equation", + "content": "\\mathcal{D}:\\{d_1 =" + }, + { + "bbox": [ + 55, + 201, + 295, + 356 + ], + "type": "text", + "content": " \"The boy in white resting on a chair after first taking the balloon.\", " + }, + { + "bbox": [ + 55, + 201, + 295, + 356 + ], + "type": "inline_equation", + "content": "d_{2} =" + }, + { + "bbox": [ + 55, + 201, + 295, + 356 + ], + "type": "text", + "content": " \"The boy in white carries it toward the hula hoop after first taking the balloon.\" " + }, + { + "bbox": [ + 55, + 201, + 295, + 356 + ], + "type": "inline_equation", + "content": "\\}" + }, + { + "bbox": [ + 55, + 201, + 295, + 356 + ], + "type": "text", + "content": ". Thus, selecting the best answer equals identifying the correct statement for a given video." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 357, + 295, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 357, + 295, + 525 + ], + "spans": [ + { + "bbox": [ + 55, + 357, + 295, + 525 + ], + "type": "text", + "content": "Recursive statement decomposition. For each initial statement in " + }, + { + "bbox": [ + 55, + 357, + 295, + 525 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 55, + 357, + 295, + 525 + ], + "type": "text", + "content": ", we generate two sub-statements as proofs that support the statement: Statement " + }, + { + "bbox": [ + 55, + 357, + 295, + 525 + ], + "type": "inline_equation", + "content": "\\Leftarrow" + }, + { + "bbox": [ + 55, + 357, + 295, + 525 + ], + "type": "text", + "content": " Sub-statement1, Sub-statement2. The statement is True if and only if both its sub-statements are proved to be True, i.e., the sub-statements entail the statement. Proving the original statements is thus translated into proving two simpler sub-statements. This procedure is recursive: the sub-statements can be further decomposed into further sub-statements that entail them. Therefore, to construct an entailment tree, we recursively decompose these sub-statements as new statements in the next tree layer until reaching the maximum depth or meeting the stop criterion. Fig. 2(a) presents an example of entailment tree generation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 525, + 295, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 525, + 295, + 561 + ], + "spans": [ + { + "bbox": [ + 55, + 525, + 295, + 561 + ], + "type": "text", + "content": "We leverage LLM prompting for both the initial statement generation and the statement decomposition, as these are linguistic tasks (see implementation details)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 569, + 259, + 582 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 569, + 259, + 582 + ], + "spans": [ + { + "bbox": [ + 55, + 569, + 259, + 582 + ], + "type": "text", + "content": "3.2. Video-language entailment verification" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "text", + "content": "Given the entailment tree, the framework then verifies language statements based on the grounded video content as evidence. Specifically, each statement in the entailment tree must be proven or refuted by analyzing the video. A straightforward solution is to encode the whole video to collect information that can be used to verify the statement. However, the critical visual evidence that accurately verifies a statement tends to exist in a local moment instead of the whole video. Therefore, we develop a novel video grounding that guides the verification process to the moments with relevant visual evidence." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 553, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 287 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 287 + ], + "type": "text", + "content": "Question-aware video captioning. Given a video, we convert its visual information into detailed textual information. Specifically, we input video frames into a VLM-based captioner " + }, + { + "bbox": [ + 313, + 72, + 553, + 287 + ], + "type": "inline_equation", + "content": "\\mathrm{Cap}(\\cdot)" + }, + { + "bbox": [ + 313, + 72, + 553, + 287 + ], + "type": "text", + "content": " to obtain a caption " + }, + { + "bbox": [ + 313, + 72, + 553, + 287 + ], + "type": "inline_equation", + "content": "c_{i} = \\mathrm{Cap}(f_{i})" + }, + { + "bbox": [ + 313, + 72, + 553, + 287 + ], + "type": "text", + "content": " for each frame. However, captioning frames individually can overlook essential details or introduce irrelevant information for VQA. In commonsense VQA, questions often focus on specific facts already observed in the video. For example, a typical temporal reasoning question is \"What happened before/after Event-A?\" where Event-A refers to a fact statement about an event in the video. The fact referenced by the question can be leveraged to guide video understanding. To this end, we first extract the anchor fact indicated by the question and provide it to " + }, + { + "bbox": [ + 313, + 72, + 553, + 287 + ], + "type": "inline_equation", + "content": "\\mathrm{Cap}(\\cdot)" + }, + { + "bbox": [ + 313, + 72, + 553, + 287 + ], + "type": "text", + "content": " as prior knowledge, encouraging the generation of relevant captions. Moreover, captions from all previous frames are also provided for each current frame to ensure " + }, + { + "bbox": [ + 313, + 72, + 553, + 287 + ], + "type": "inline_equation", + "content": "\\mathrm{Cap}(\\cdot)" + }, + { + "bbox": [ + 313, + 72, + 553, + 287 + ], + "type": "text", + "content": " captures the temporal context from the past. This process is formulated as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 365, + 298, + 553, + 312 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 298, + 553, + 312 + ], + "spans": [ + { + "bbox": [ + 365, + 298, + 553, + 312 + ], + "type": "interline_equation", + "content": "c _ {i} = \\operatorname {C a p} \\left(f _ {i} \\mid F, \\left(c _ {1}, \\dots , c _ {i - 1}\\right)\\right), \\tag {1}", + "image_path": "60bd8e7b4d7e2c94a0b04edb3865f8107d2d472c561e14c0fa77100d2b684fe4.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 321, + 463, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 321, + 463, + 332 + ], + "spans": [ + { + "bbox": [ + 313, + 321, + 463, + 332 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 321, + 463, + 332 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 313, + 321, + 463, + 332 + ], + "type": "text", + "content": " indicates the fact statement." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 333, + 553, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 333, + 553, + 441 + ], + "spans": [ + { + "bbox": [ + 313, + 333, + 553, + 441 + ], + "type": "text", + "content": "Video evidence grounding. For commonsense VQA, depending on how the question reasons around the fact statement, the necessary evidence for answers can be gathered from specific video moments. For instance, in the case of temporal reasoning (e.g., before or after questions), the answer should be inferred from moments occurring either before or after the time of the relevant fact. Following this intuition, we design a two-step evidence-grounding strategy to localize the critical moments for answering." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 441, + 554, + 645 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 441, + 554, + 645 + ], + "spans": [ + { + "bbox": [ + 313, + 441, + 554, + 645 + ], + "type": "text", + "content": "First, given the frame-wise captions, we retrieve a keyframe deemed most relevant to the fact statement, which we refer to as the anchor frame. A straightforward retrieval approach would involve comparing each " + }, + { + "bbox": [ + 313, + 441, + 554, + 645 + ], + "type": "inline_equation", + "content": "c_i" + }, + { + "bbox": [ + 313, + 441, + 554, + 645 + ], + "type": "text", + "content": " with the fact description using specific metrics to identify the anchor frame. However, we enhance retrieval accuracy by adopting a structured semantic retrieval strategy. Specifically, the textual descriptions of each frame and fact statement are converted into structured triplets. These triplets capture the attributes and relationships of objects in each frame through structured semantics. As shown in Fig. 2(a), rather than directly comparing raw textual descriptions of frames and fact statements, we use these triplets for retrieval. Inspired by the success of using LLMs for retrieval tasks, we prompt an LLM to conduct anchor frame retrieval using the triplets of the fact statement as the query. The LLM then identifies and returns the most relevant frame ID, i.e., its timestamp." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 390, + 654, + 553, + 668 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 390, + 654, + 553, + 668 + ], + "spans": [ + { + "bbox": [ + 390, + 654, + 553, + 668 + ], + "type": "interline_equation", + "content": "t _ {\\text {a n c h o r}} = \\operatorname {R t v} \\left(c _ {i}, F\\right), \\tag {2}", + "image_path": "300ae2815069814cc1b12575498d92d3ee1386b1b75253d89ebcb7137c986826.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "inline_equation", + "content": "t_{\\mathrm{anchor}}" + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "content": " is the time stamp of the anchor frame, " + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{Rtv}(\\cdot)" + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "content": " denotes the retrieval process. Second, we determine the final moment where we should look centered on the " + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "inline_equation", + "content": "t_{\\mathrm{anchor}}" + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "content": ", to" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3264" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 70, + 553, + 343 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 553, + 343 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 553, + 343 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 553, + 343 + ], + "type": "image", + "image_path": "9f3359d2c101bb0a8e86b760c57a1259094279cacb3e42086b24c0b83e63a8be.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 348, + 555, + 394 + ], + "lines": [ + { + "bbox": [ + 54, + 348, + 555, + 394 + ], + "spans": [ + { + "bbox": [ + 54, + 348, + 555, + 394 + ], + "type": "text", + "content": "Figure 2. Overview of our framework. (a) The generation of the entailment tree, where statements are recursively decomposed until the tree reaches its max depth or meets the stop criterion. (b) The process of video-language entailment verification: the input video is first converted into textual descriptions. Each caption is then parsed into structured semantics. Given the fact statement as a query, we retrieve the anchor frame. Then, based on the temporal or causal navigation indicated by questions, the visual evidence moment can be grounded." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 409, + 296, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 409, + 296, + 458 + ], + "spans": [ + { + "bbox": [ + 55, + 409, + 296, + 458 + ], + "type": "text", + "content": "incorporate the temporal relations present in the question. Therefore, based on the anchor frame, the navigation for the moment is selected from \"look ahead, look behind, look around\" by considering the question:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 129, + 468, + 294, + 483 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 468, + 294, + 483 + ], + "spans": [ + { + "bbox": [ + 129, + 468, + 294, + 483 + ], + "type": "interline_equation", + "content": "\\mathcal {M} = \\operatorname {G n d} \\left(t _ {\\text {a n c h o r}} \\mid \\mathcal {Q}\\right), \\tag {3}", + "image_path": "38614fffdea105fcd5e5f4cf6e01e5f79f0c395066e0e257362dad61ed79c12e.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 493, + 296, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 493, + 296, + 540 + ], + "spans": [ + { + "bbox": [ + 55, + 493, + 296, + 540 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 493, + 296, + 540 + ], + "type": "inline_equation", + "content": "\\operatorname{Gnd}(\\cdot)" + }, + { + "bbox": [ + 55, + 493, + 296, + 540 + ], + "type": "text", + "content": " is the grounding process and " + }, + { + "bbox": [ + 55, + 493, + 296, + 540 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 55, + 493, + 296, + 540 + ], + "type": "text", + "content": " denotes the grounded continuous interval in the video. Then, frames are resampled from the video within " + }, + { + "bbox": [ + 55, + 493, + 296, + 540 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 493, + 296, + 540 + ], + "type": "text", + "content": " and used as visual evidence proving or refuting entailment tree statements." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 542, + 296, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 542, + 296, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 542, + 296, + 685 + ], + "type": "text", + "content": "Visual-text statement prover. Given grounded visual evidence " + }, + { + "bbox": [ + 55, + 542, + 296, + 685 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 55, + 542, + 296, + 685 + ], + "type": "text", + "content": " of the video, statements are estimated to be true or false. Specifically, we employ a VLM as the statement prover, denoted as " + }, + { + "bbox": [ + 55, + 542, + 296, + 685 + ], + "type": "inline_equation", + "content": "\\mathrm{Prv}(\\cdot)" + }, + { + "bbox": [ + 55, + 542, + 296, + 685 + ], + "type": "text", + "content": ", to evaluate each statement within the tree by probing VLM's internal belief on this statement. Each statement will be transformed into a binary QA task, with possible options being True or False. We then directly probe the " + }, + { + "bbox": [ + 55, + 542, + 296, + 685 + ], + "type": "inline_equation", + "content": "\\mathrm{Prv}(\\cdot)" + }, + { + "bbox": [ + 55, + 542, + 296, + 685 + ], + "type": "text", + "content": " with the binary QA prompt and use the next token prediction probabilities of the words to elicit the model's belief. We normalize the prediction logits of the two options to get the confidence score of that statement. The above process is formulated as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 115, + 697, + 295, + 711 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 697, + 295, + 711 + ], + "spans": [ + { + "bbox": [ + 115, + 697, + 295, + 711 + ], + "type": "interline_equation", + "content": "s _ {d} = \\Pr \\operatorname {r v} (\\mathcal {M}, h), s \\in [ 0, 1 ], \\tag {4}", + "image_path": "ed37f8ff062c02012d1905ac0f8a3b5356aa3a3eac30e2a68fc41678404bfe29.jpg" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 314, + 406, + 553, + 483 + ], + "blocks": [ + { + "bbox": [ + 314, + 406, + 553, + 483 + ], + "lines": [ + { + "bbox": [ + 314, + 406, + 553, + 483 + ], + "spans": [ + { + "bbox": [ + 314, + 406, + 553, + 483 + ], + "type": "image", + "image_path": "d3b28f2f6252dbcf4a3b151041a2110b13482ff2e7af4b3cef3f8eba367f972a.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 491, + 555, + 536 + ], + "lines": [ + { + "bbox": [ + 313, + 491, + 555, + 536 + ], + "spans": [ + { + "bbox": [ + 313, + 491, + 555, + 536 + ], + "type": "text", + "content": "Figure 3. Illustration of dynamic tree generation and backtrace. In Step-3, when the proof score of the left statement calculated from its child nodes is less than its direct score " + }, + { + "bbox": [ + 313, + 491, + 555, + 536 + ], + "type": "inline_equation", + "content": "(0.63 < 0.8)" + }, + { + "bbox": [ + 313, + 491, + 555, + 536 + ], + "type": "text", + "content": ", its decomposition is pruned and stops." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 547, + 555, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 547, + 555, + 571 + ], + "spans": [ + { + "bbox": [ + 313, + 547, + 555, + 571 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 547, + 555, + 571 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 313, + 547, + 555, + 571 + ], + "type": "text", + "content": " is the grounded moment and " + }, + { + "bbox": [ + 313, + 547, + 555, + 571 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 313, + 547, + 555, + 571 + ], + "type": "text", + "content": " denotes the statement that needs to be verified." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 581, + 503, + 593 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 581, + 503, + 593 + ], + "spans": [ + { + "bbox": [ + 313, + 581, + 503, + 593 + ], + "type": "text", + "content": "3.3. Dynamic entailment tree expansion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "content": "So far, we have performed statement decomposition recursively to construct an entailment tree with pre-defined depth. However, not all statements need to be verified recursively, especially those easily determined to be true or false by VLMs. Moreover, as the depth increases, some statement sentences are atomic and directly verifiable. Thus, to improve the efficiency of the reasoning process, we further adopt a strategy to expand the entailment tree dynamically. Specifically, each statement " + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "content": " is tied with two confidence scores provided by the " + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\operatorname{Prv}(\\cdot)" + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3265" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 70, + 294, + 277 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 294, + 277 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 294, + 277 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 294, + 277 + ], + "type": "image", + "image_path": "de93e1108f78b4b924bb625a222dfe4a9d1eb3d32f5f2397f848d19e9d5719f3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 289, + 295, + 312 + ], + "lines": [ + { + "bbox": [ + 55, + 289, + 295, + 312 + ], + "spans": [ + { + "bbox": [ + 55, + 289, + 295, + 312 + ], + "type": "text", + "content": "Figure 4. Illustration of commonsense bias in video question answering. The example is selected from the NExT-QA dataset." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 327, + 295, + 385 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 48, + 327, + 295, + 349 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 327, + 295, + 349 + ], + "spans": [ + { + "bbox": [ + 48, + 327, + 295, + 349 + ], + "type": "text", + "content": "(1) The direct score " + }, + { + "bbox": [ + 48, + 327, + 295, + 349 + ], + "type": "inline_equation", + "content": "s_d" + }, + { + "bbox": [ + 48, + 327, + 295, + 349 + ], + "type": "text", + "content": ", which indicates the belief of " + }, + { + "bbox": [ + 48, + 327, + 295, + 349 + ], + "type": "inline_equation", + "content": "\\operatorname{Prv}(\\cdot)" + }, + { + "bbox": [ + 48, + 327, + 295, + 349 + ], + "type": "text", + "content": " model in " + }, + { + "bbox": [ + 48, + 327, + 295, + 349 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 48, + 327, + 295, + 349 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 350, + 295, + 385 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 350, + 295, + 385 + ], + "spans": [ + { + "bbox": [ + 48, + 350, + 295, + 385 + ], + "type": "text", + "content": "(2) The proof score " + }, + { + "bbox": [ + 48, + 350, + 295, + 385 + ], + "type": "inline_equation", + "content": "s_p" + }, + { + "bbox": [ + 48, + 350, + 295, + 385 + ], + "type": "text", + "content": ", denoting how confidently the model can prove " + }, + { + "bbox": [ + 48, + 350, + 295, + 385 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 48, + 350, + 295, + 385 + ], + "type": "text", + "content": ", is calculated by multiplying the scores of its direct sub-statements." + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 387, + 296, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 387, + 296, + 493 + ], + "spans": [ + { + "bbox": [ + 55, + 387, + 296, + 493 + ], + "type": "text", + "content": "For a statement " + }, + { + "bbox": [ + 55, + 387, + 296, + 493 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 55, + 387, + 296, + 493 + ], + "type": "text", + "content": ", the goal of decomposition is to establish a more reliable and convincing proof path than merely evaluating whether " + }, + { + "bbox": [ + 55, + 387, + 296, + 493 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 55, + 387, + 296, + 493 + ], + "type": "text", + "content": " is true by VLMs. If the decomposition-based reasoning can prove " + }, + { + "bbox": [ + 55, + 387, + 296, + 493 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 55, + 387, + 296, + 493 + ], + "type": "text", + "content": " with higher confidence than its direct score, the overall confidence for statement " + }, + { + "bbox": [ + 55, + 387, + 296, + 493 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 55, + 387, + 296, + 493 + ], + "type": "text", + "content": " should increase. Otherwise, the decomposition should be disregarded. Thus, in the dynamic tree expansion, if decomposition does not enhance a statement's score, it is pruned, and that statement node becomes a leaf in the entailment tree." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 495, + 296, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 495, + 296, + 531 + ], + "spans": [ + { + "bbox": [ + 55, + 495, + 296, + 531 + ], + "type": "text", + "content": "Fig. 3 presents a toy example. This criterion ensures that only beneficial decompositions are retained, significantly enhancing the tree reasoning process's efficiency." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 540, + 241, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 540, + 241, + 552 + ], + "spans": [ + { + "bbox": [ + 55, + 540, + 241, + 552 + ], + "type": "text", + "content": "3.4. Reasoning over the entailment tree" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 552, + 295, + 637 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 552, + 295, + 637 + ], + "spans": [ + { + "bbox": [ + 55, + 552, + 295, + 637 + ], + "type": "text", + "content": "Finally, we perform a backtrace through the entailment tree to calculate the confidence score of each top statement. Specifically, the final score for each statement is produced by comparing its direct score " + }, + { + "bbox": [ + 55, + 552, + 295, + 637 + ], + "type": "inline_equation", + "content": "s_d" + }, + { + "bbox": [ + 55, + 552, + 295, + 637 + ], + "type": "text", + "content": " and proof score " + }, + { + "bbox": [ + 55, + 552, + 295, + 637 + ], + "type": "inline_equation", + "content": "s_p" + }, + { + "bbox": [ + 55, + 552, + 295, + 637 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 55, + 552, + 295, + 637 + ], + "type": "inline_equation", + "content": "s = \\max(s_d, s_p)" + }, + { + "bbox": [ + 55, + 552, + 295, + 637 + ], + "type": "text", + "content": " during backtrace (as shown in Fig. 3). The overall framework selects the answer corresponding to the statement at the top layer with top-scoring proof." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 650, + 290, + 663 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 650, + 290, + 663 + ], + "spans": [ + { + "bbox": [ + 55, + 650, + 290, + 663 + ], + "type": "text", + "content": "4. De-biasing commonsense VQA answer sets" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 666, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 666, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 666, + 296, + 713 + ], + "type": "text", + "content": "To demonstrate the reasoning ability of video-grounded entailment trees, it is essential to evaluate using commonsense VQA benchmarks that enforce model reasoning. Recent work [13, 18, 27] has provided evidence that shortcuts" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 321, + 75, + 548, + 222 + ], + "blocks": [ + { + "bbox": [ + 321, + 75, + 548, + 222 + ], + "lines": [ + { + "bbox": [ + 321, + 75, + 548, + 222 + ], + "spans": [ + { + "bbox": [ + 321, + 75, + 548, + 222 + ], + "type": "image", + "image_path": "e3a2bf99076e33621d676eb4bb7041017cbde2dfe28a52b3a509b12f8f37a998.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 326, + 233, + 541, + 245 + ], + "lines": [ + { + "bbox": [ + 326, + 233, + 541, + 245 + ], + "spans": [ + { + "bbox": [ + 326, + 233, + 541, + 245 + ], + "type": "text", + "content": "Figure 5. Prompt used for rewriting answers on NExT-QA." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 282, + 555, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 282, + 555, + 498 + ], + "spans": [ + { + "bbox": [ + 313, + 282, + 555, + 498 + ], + "type": "text", + "content": "are present in VQA datasets which enables VLMs to solve these tasks based on textual associations rather than video-grounded reasoning. While VQA benchmarks increasingly focus on commonsense reasoning skills, such as temporal (e.g. after, before) or causal (how, why, what if) relationships in video content, reasoning shortcuts affect the validity of their evaluation. This is illustrated in Fig. 4 (top), where the correct answer (D) is much more relevant to the question and also aligns best with real-world expectations. Consequently, a VLM (VideoLLaVA [14] used in this example) can answer this question correctly by leveraging such associations and without analyzing the video content. Meanwhile, replacing the answer set distractors with other commonsensical answer candidates, as illustrated in Fig. 4 (bottom), makes this task challenging for VLMs. Here, a VLM switches its answer incorrectly to option (C), which confirms the impact of commonsense associations and the lack of grounded reasoning by these models." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 510, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 556, + 713 + ], + "type": "text", + "content": "To this end, we devise a de-biasing procedure that mitigates reasoning shortcuts in commonsense VQA answer sets. Our de-biasing procedure transforms multiple-choice VQA benchmarks (e.g., NExT-QA) by rewriting their answer distractors while keeping their question and ground-truth answer intact. We prompt an LLM (LLaMA-3) to implement the rewriting procedure for each original QA set. Fig. 5 shows the detailed prompt we used for LLaMA-3 on NExT-QA dataset. This procedure ensures that (1) the answers cannot be easily derived from the QA set associations and (2) the answer remains consistent with the original QA pair. Thus, our procedure enables the scalable construction of de-biased QA sets by leveraging the commonsense associations in LLMs. The next section, focusing on experimental evaluation, analyzes the application of de-biasing to various datasets and its impact on the performance of VLMs, with and without entailment tree reasoning." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3266" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 136, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 136, + 84 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 136, + 84 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 87, + 170, + 99 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 87, + 170, + 99 + ], + "spans": [ + { + "bbox": [ + 55, + 87, + 170, + 99 + ], + "type": "text", + "content": "5.1. Experimental setup" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 100, + 294, + 207 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 100, + 294, + 207 + ], + "spans": [ + { + "bbox": [ + 55, + 100, + 294, + 207 + ], + "type": "text", + "content": "Datasets. We test our framework on three VQA benchmarks: (1) NExT-QA [26], a VQA benchmark for causal and temporal reasoning. (2) IntentQA [11], which focuses on video intent reasoning from both causal and temporal aspects. (3) Video-MME [6], which is a recently proposed comprehensive evaluation benchmark for video analysis; we use its \"short-term\" split (video length " + }, + { + "bbox": [ + 55, + 100, + 294, + 207 + ], + "type": "inline_equation", + "content": "< 2" + }, + { + "bbox": [ + 55, + 100, + 294, + 207 + ], + "type": "text", + "content": " mins) and 4 question types (temporal, spatial, action, and object reasoning) highly related to commonsense reasoning are selected." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 209, + 294, + 257 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 209, + 294, + 257 + ], + "spans": [ + { + "bbox": [ + 55, + 209, + 294, + 257 + ], + "type": "text", + "content": "Evaluation. We report model performances on our rewritten test set for each dataset and its original test set. We evaluate our framework on all datasets under the multiple-choice QA setting, using a standard accuracy metric." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 258, + 264, + 269 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 258, + 264, + 269 + ], + "spans": [ + { + "bbox": [ + 55, + 258, + 264, + 269 + ], + "type": "text", + "content": "Baselines. Our baselines represent three categories:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 270, + 294, + 389 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 55, + 270, + 294, + 329 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 270, + 294, + 329 + ], + "spans": [ + { + "bbox": [ + 55, + 270, + 294, + 329 + ], + "type": "text", + "content": "- Video-based VLMs: Video-based VLMs are widely used for VQA tasks, so we include Video-LLaVA [14], VideoChat2 [12], and VideoLLaMA [33]. To test the effectiveness of our framework, we integrate these VLMs by replacing our " + }, + { + "bbox": [ + 55, + 270, + 294, + 329 + ], + "type": "inline_equation", + "content": "\\mathsf{Prv}(\\cdot)" + }, + { + "bbox": [ + 55, + 270, + 294, + 329 + ], + "type": "text", + "content": " with specific VLM models." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 331, + 294, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 331, + 294, + 354 + ], + "spans": [ + { + "bbox": [ + 55, + 331, + 294, + 354 + ], + "type": "text", + "content": "- Image-based VLMs: We include BLIP-2 [10] and LLaVA-1.5 [17] as Image-LLM baselines." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 354, + 294, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 354, + 294, + 389 + ], + "spans": [ + { + "bbox": [ + 55, + 354, + 294, + 389 + ], + "type": "text", + "content": "- State-of-the-art VQA approaches: Recent works in VQA, such as VideoTree [22], VideoAgent [21], and LLoVi [31], are included as strong baselines." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 391, + 295, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 391, + 295, + 654 + ], + "spans": [ + { + "bbox": [ + 55, + 391, + 295, + 654 + ], + "type": "text", + "content": "Implementation details. Our entire framework is training-and human annotation-free. We use LLaMA-3-8B [5] to handle basic functionalities, including (1) converting original QA into declarative statements, (2) statement decomposition, (3) structured semantic extraction and retrieval, and (4) guiding evidence grounding. Detailed prompts for each functionality are provided in the supplementary material. For frame-wise captioning across all datasets, we use LLaVA-1.5 [17] as our default captioner. When comparing with state-of-the-art VQA methods (e.g., VideoAgent), we follow their setup by replacing the captioner with the stronger CogAgent [7] model for fair comparison. When integrating our framework with VLMs, the VLM itself serves as the Prover, " + }, + { + "bbox": [ + 55, + 391, + 295, + 654 + ], + "type": "inline_equation", + "content": "\\mathsf{Prv}(\\cdot)" + }, + { + "bbox": [ + 55, + 391, + 295, + 654 + ], + "type": "text", + "content": ". For video-LLMs, frames are uniformly sampled from the grounded video moment to meet their input requirements (VideoChat2: 16 frames; VideoLLaVA & VideoLLaMA: 8 frames). For image-language models like LLaVA, we sample 8 frames from the grounded moment and process them individually; the final confidence score is obtained by averaging scores across frames. During dynamic tree generation, we also set a max depth of 5 for the overall entailment tree to improve the efficiency." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 664, + 137, + 675 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 664, + 137, + 675 + ], + "spans": [ + { + "bbox": [ + 55, + 664, + 137, + 675 + ], + "type": "text", + "content": "5.2. Main results" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 677, + 294, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 294, + 713 + ], + "type": "text", + "content": "Benefit for image and video-based VLMs. Tab. 1 summarizes the results of our method compared to baselines. Integrating entailment tree reasoning brings consistent im" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 72, + 553, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 251 + ], + "type": "text", + "content": "provement across the video- and image-based VLMs for all datasets (1-4% on average). This finding includes the recently proposed benchmark VideoMME, which poses much more challenging videos and questions. The benefits are particularly regular for temporal reasoning (improvement in 14 out of 15 cases), which illustrates how our explicit reasoning process enhances temporal commonsense QA. Image-based VLMs, which initially lack temporal modeling capabilities, perform poorly when directly applied to video QA tasks. However, our framework provides a significant performance boost for these models up to " + }, + { + "bbox": [ + 313, + 72, + 553, + 251 + ], + "type": "inline_equation", + "content": "8\\%" + }, + { + "bbox": [ + 313, + 72, + 553, + 251 + ], + "type": "text", + "content": " for LLAVA-1.5. By reasoning over multiple sub-problems rather than tackling the entire complex question simultaneously, our reasoning method makes the task more manageable for both video- and image-based VLMs." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 251, + 553, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 251, + 553, + 478 + ], + "spans": [ + { + "bbox": [ + 313, + 251, + 553, + 478 + ], + "type": "text", + "content": "Results on de-biased QA sets. As shown in Tab. 2, all models experience considerable performance drops on the de-biased set of the same VQA dataset, which aligns with our observation that current VLMs often rely on textual bias in commonsense reasoning tasks. Notably, video-based VLMs show an " + }, + { + "bbox": [ + 313, + 251, + 553, + 478 + ], + "type": "inline_equation", + "content": "8\\% -10\\%" + }, + { + "bbox": [ + 313, + 251, + 553, + 478 + ], + "type": "text", + "content": " decrease in the de-biased set even though the question and the correct answer remain unchanged. In contrast, our proposed framework, which derives answers through an explicit reasoning process based on specific visual evidence, demonstrates much greater robustness on the de-biased set. The improvement brought by our framework on the de-biased set is even higher than on the original test sets. In turn, our framework compensates for the performance loss of the VLMs on the de-biased set. This analysis underscores our framework's potential to mitigate textual bias in commonsense reasoning. Furthermore, the performance differences between the original and de-biased QA sets highlight VQA benchmarks' limitations in evaluating VLMs' true reasoning abilities." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 479, + 553, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 479, + 553, + 634 + ], + "spans": [ + { + "bbox": [ + 313, + 479, + 553, + 634 + ], + "type": "text", + "content": "Comparison with state-of-the-art. Tab. 3 compares our framework's results on the de-biased sets to state-of-the-art VQA approaches. The table shows that, next to the consistent benefit our framework provides to various VLMs, it is also competitive with state-of-the-art VQA methods. When applying an advanced captioner and reasoner that aligns with VideoAgent and VideoTree, our framework yields new state-of-the-art results in some cases. In particular, our framework performs best on temporal reasoning questions for all three benchmarks and outperforms all methods on the IntentQA dataset. Importantly, our method reaches such competitive performance despite using " + }, + { + "bbox": [ + 313, + 479, + 553, + 634 + ], + "type": "inline_equation", + "content": "257 \\times" + }, + { + "bbox": [ + 313, + 479, + 553, + 634 + ], + "type": "text", + "content": " fewer parameters for its reasoning compared to state-of-the-art methods." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 641, + 412, + 652 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 412, + 652 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 412, + 652 + ], + "type": "text", + "content": "5.3. Ablation studies" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 654, + 553, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 553, + 689 + ], + "type": "text", + "content": "Ablation experiments are conducted using VideoLLaVA's baseline with our framework. Results are reported on the test set of the NExT-QA dataset." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "content": "Impact of LLMs for statement decomposition. Entailment generation in our framework relies on prompting an" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3267" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 70, + 555, + 209 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 555, + 209 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 555, + 209 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 555, + 209 + ], + "type": "table", + "html": "
ModelNExT-QAIntentQAVideoMMEAvg
TemporalCausalTemporalCausalTemporalSpatialActionObject
Image-based VLMsBLIP-2 [10]38.336.143.848.625.426.924.228.634.0
+Ours45.341.848.952.530.527.427.728.837.9
LLaVA-1.5 [17]37.840.745.850.031.133.627.729.837.1
+Ours45.647.948.454.736.736.831.430.741.5
Video-based VLMsVideoChat2 [12]56.962.160.463.250.352.449.550.155.6
+Ours57.861.662.363.852.851.551.350.056.4
VideoLLaVA [14]56.060.453.960.747.644.346.249.752.3
+Ours58.362.757.661.848.346.149.850.354.2
VideoLLaMA [33]55.460.255.156.744.847.244.749.351.7
+Ours58.160.454.558.947.447.848.649.153.1
", + "image_path": "b3d2e3be807e912e7d73a85364ee50e63248ed1c6858badd95a5968c92474e60.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 56, + 246, + 555, + 335 + ], + "blocks": [ + { + "bbox": [ + 55, + 213, + 555, + 236 + ], + "lines": [ + { + "bbox": [ + 55, + 213, + 555, + 236 + ], + "spans": [ + { + "bbox": [ + 55, + 213, + 555, + 236 + ], + "type": "text", + "content": "Table 1. Impact on image and video-based VLMs on the original NExT-QA, IntentQA, and VideoMME test sets. Our framework increases accuracy of all video- and image-based VLMs by " + }, + { + "bbox": [ + 55, + 213, + 555, + 236 + ], + "type": "inline_equation", + "content": "1 - 4\\%" + }, + { + "bbox": [ + 55, + 213, + 555, + 236 + ], + "type": "text", + "content": " on average across all data partitions. Temporal and action partitions benefit most." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 246, + 555, + 335 + ], + "lines": [ + { + "bbox": [ + 56, + 246, + 555, + 335 + ], + "spans": [ + { + "bbox": [ + 56, + 246, + 555, + 335 + ], + "type": "table", + "html": "
ModelBLIP-2+OursLLaVA+OursVideo Chat2+OursVideo LLaVA+OursVideo LLaMA+Ours
NExT-QAOriginal37.243.639.346.859.559.758.260.557.859.3
Rewritten33.539.834.844.945.449.051.155.441.447.0
Intent-QAOriginal46.250.747.951.661.863.157.359.755.956.7
Rewritten38.245.542.748.652.655.750.554.746.350.0
Avg38.844.941.248.054.856.954.357.650.453.3
", + "image_path": "8ce1c9037fa7df55b32339d610051b204d6c291308b6df4a6c4e159d97f27d78.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 338, + 555, + 361 + ], + "lines": [ + { + "bbox": [ + 55, + 338, + 555, + 361 + ], + "spans": [ + { + "bbox": [ + 55, + 338, + 555, + 361 + ], + "type": "text", + "content": "Table 2. Results on de-biased QA sets. Video-based VLMs show significant decreases in the rewritten de-biased set. In contrast, our framework demonstrates much greater robustness on the rewritten set." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 381, + 295, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 381, + 295, + 550 + ], + "spans": [ + { + "bbox": [ + 54, + 381, + 295, + 550 + ], + "type": "text", + "content": "external LLM to recursively decompose statements (cf. Sec. 3.1), which is crucial in guiding reasoning paths. Consequently, we tested various LLMs for entailment tree generation, including open-source models (LLama-3 and Mistral) of different sizes and proprietary LLMs (GPT-4 and Gemini-1.5). The results are summarized in Tab. 4. As expected, the proprietary model GPT-4, known for its strong step-by-step reasoning capabilities, delivers the best performance across all settings. Scaling up LLaMA-3 to 70B offers improvements over the 8B model, though with a notable increase in inference time. As the overall performance difference between models is within " + }, + { + "bbox": [ + 54, + 381, + 295, + 550 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 54, + 381, + 295, + 550 + ], + "type": "text", + "content": ", we select LLaMA-3-8B as the default for integrating our framework into VLMs due to its free availability and efficiency." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 558, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 558, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 558, + 295, + 714 + ], + "type": "text", + "content": "Ablation on grounding components. Next, we test the effectiveness of each component in our grounding module (Sec. 3.2). The results, summarized in Tab. 5, indicate that both fact-conditional captioning and structure-guided retrieval enhance overall performance by improving grounding accuracy. However, using only structure-guided retrieval results in a slight performance drop, possibly because " + }, + { + "bbox": [ + 54, + 558, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\mathsf{Cap}(\\cdot)" + }, + { + "bbox": [ + 54, + 558, + 295, + 714 + ], + "type": "text", + "content": " introduces irrelevant semantic information that doesn't align with the question's focus, and the structured representation can make identifying anchor frames more challenging. In contrast, fact-conditional captioning alone yields substantial improvement, demonstrating that this straightforward approach can yield an effective" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 381, + 553, + 406 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 381, + 553, + 406 + ], + "spans": [ + { + "bbox": [ + 313, + 381, + 553, + 406 + ], + "type": "text", + "content": "and more controllable textual description for videos by conditioning on prior knowledge or relevant facts." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 410, + 555, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 410, + 555, + 566 + ], + "spans": [ + { + "bbox": [ + 313, + 410, + 555, + 566 + ], + "type": "text", + "content": "Impact of length of video frames. We further ablate the impact of input video frame length in our framework to determine the optimal number of frame-wise captions to generate per video. The results, summarized in Tab. 6, show that ideal performance is achieved only when sampling a sufficient number of frames (at least 16 for NExTQA). When fewer frames are used (e.g., 4 or 8), key anchor frames may be missed, reducing the accuracy of grounded visual evidence. Additionally, while increasing the frame count to 32 yields the best performance, it also increases the calls required for " + }, + { + "bbox": [ + 313, + 410, + 555, + 566 + ], + "type": "inline_equation", + "content": "\\mathrm{Cap}(\\cdot)" + }, + { + "bbox": [ + 313, + 410, + 555, + 566 + ], + "type": "text", + "content": " to generate frame-wise captions. Balancing efficiency with performance gains, we set 24 frames as the default in our implementation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "text", + "content": "Effectiveness of evidence grounding. Our method grounds relevant video fragments to support statements in the entailment tree (Sec. 3.2). To validate its effectiveness, we compare it with two other variations as sources of visual evidence: (1) without evidence grounding, using the full video as evidence, and (2) upper-bound results: manually annotated temporal boundaries provided in the NExT-GQA dataset, indicating where the QA models should focus when producing correct answers. The results are shown in Tab. 7. Compared to the baseline, our video-grounded method provides consistent improvements across the original and debiased sets. The improvement is more apparent in the de" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3268" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 70, + 555, + 143 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 555, + 143 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 555, + 143 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 555, + 143 + ], + "type": "table", + "html": "
MethodModel ReasonerNExT-QA*IntentQA*VideoMME
TemporalCausalTemporalCausalTemporalSpatialActionObject
VideoAgent [21]GPT-4 (1.8T)58.266.660.461.0----
VideoTree [22]GPT-4 (1.8T)60.266.456.760.155.754.354.252.6
LLoVi [31]GPT-4 (1.8T)53.160.858.661.852.255.351.850.8
OursVideoLLAVA (7B)60.865.961.062.655.953.854.050.8
", + "image_path": "f570a25d65df90b4c64b191ff41b84de1be208a666c90923ec94f82e7852fc1b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 57, + 209, + 294, + 297 + ], + "blocks": [ + { + "bbox": [ + 55, + 146, + 555, + 191 + ], + "lines": [ + { + "bbox": [ + 55, + 146, + 555, + 191 + ], + "spans": [ + { + "bbox": [ + 55, + 146, + 555, + 191 + ], + "type": "text", + "content": "Table 3. Comparison with state-of-the-art. Results for NExT-QA and IntentQA are reported under the de-biased set (the results on the original sets are similar; we provide them in the Appendix). The 'Reasoner' in these approaches is similar to the \"Prover\" in our framework. The captioner for all methods is CogAgent [7]. Despite other methods relying on much stronger reasoning models, our approach yields competitive performance (four state-of-the-art results) and high parameter efficiency (" + }, + { + "bbox": [ + 55, + 146, + 555, + 191 + ], + "type": "inline_equation", + "content": "257 \\times" + }, + { + "bbox": [ + 55, + 146, + 555, + 191 + ], + "type": "text", + "content": " fewer than GPT-4 reasoners)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 209, + 294, + 297 + ], + "lines": [ + { + "bbox": [ + 57, + 209, + 294, + 297 + ], + "spans": [ + { + "bbox": [ + 57, + 209, + 294, + 297 + ], + "type": "table", + "html": "
Model classLLMNExT-QA
OriginalRewritten
Open-sourceMistral-7B60.055.6
LLaMA-3-8B60.555.4
LLaMA-3-70B61.355.9
ProprietaryGemini-1.5-Pro61.155.2
GPT-461.656.1
", + "image_path": "0c3ccde910e7dbe8814485cc03c34975697d2d70e851901636368b9addca4541.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 58, + 358, + 295, + 441 + ], + "blocks": [ + { + "bbox": [ + 55, + 300, + 295, + 346 + ], + "lines": [ + { + "bbox": [ + 55, + 300, + 295, + 346 + ], + "spans": [ + { + "bbox": [ + 55, + 300, + 295, + 346 + ], + "type": "text", + "content": "Table 4. Impact of LLMs for statement decomposition. The open-source and proprietary models are ordered ascendingly by size. Larger models, especially GPT-4, are best at decomposition, but the smaller models (e.g., LlaMa-3-8B) come close." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 358, + 295, + 441 + ], + "lines": [ + { + "bbox": [ + 58, + 358, + 295, + 441 + ], + "spans": [ + { + "bbox": [ + 58, + 358, + 295, + 441 + ], + "type": "table", + "html": "
ComponentsNExT-QA
Fact-conditioned captioningStructure-based retrievalOriginalRewritten
56.249.6
✓59.553.3
✓58.452.7
✓✓60.555.4
", + "image_path": "b4a18976df1527aebb95f866e5602612294cee2040bc0f8a0790348def79dc8b.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 71, + 491, + 281, + 544 + ], + "blocks": [ + { + "bbox": [ + 55, + 445, + 295, + 479 + ], + "lines": [ + { + "bbox": [ + 55, + 445, + 295, + 479 + ], + "spans": [ + { + "bbox": [ + 55, + 445, + 295, + 479 + ], + "type": "text", + "content": "Table 5. Ablation on grounding components, showing that both fact-conditional captioning and structure-guided retrieval enhance overall performance by improving grounding accuracy." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 71, + 491, + 281, + 544 + ], + "lines": [ + { + "bbox": [ + 71, + 491, + 281, + 544 + ], + "spans": [ + { + "bbox": [ + 71, + 491, + 281, + 544 + ], + "type": "table", + "html": "
Acc\n(NExT-QA)Frame number
48162432
Original57.759.360.560.761.0
Rewritten51.953.455.455.455.7
", + "image_path": "5feb288937de3b56dc9bf07f5d2f9b49c3d82851a6a64d195c69ced4547b6ad6.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 592, + 295, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 592, + 295, + 652 + ], + "spans": [ + { + "bbox": [ + 55, + 592, + 295, + 652 + ], + "type": "text", + "content": "biased set, where the answer options are more semantically similar and require more precise, discriminative visual evidence. Using the ground-truth fragment can further boost our approach, suggesting that enhancing grounding accuracy could further improve our framework." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "content": "Effectiveness of dynamic tree expansion. The depth of the entailment tree determines the granularity of reasoning (Sec. 3.3). This ablation analyzes how tree depth impacts overall performance and compares a fixed-depth approach with our dynamic tree generation strategy. Increases" + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 315, + 209, + 554, + 249 + ], + "blocks": [ + { + "bbox": [ + 55, + 546, + 295, + 569 + ], + "lines": [ + { + "bbox": [ + 55, + 546, + 295, + 569 + ], + "spans": [ + { + "bbox": [ + 55, + 546, + 295, + 569 + ], + "type": "text", + "content": "Table 6. Impact of video frame amount. Strong performance requires a sufficiently high frame number (over 16 for NExT-QA)." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 209, + 554, + 249 + ], + "lines": [ + { + "bbox": [ + 315, + 209, + 554, + 249 + ], + "spans": [ + { + "bbox": [ + 315, + 209, + 554, + 249 + ], + "type": "table", + "html": "
Video fragmentFullGrounded (ours)GT
NExT-QAOriginal58.360.561.8
Rewritten51.755.456.9
", + "image_path": "25cb4bf3eb3ffdb08bed047d963fe981531b8ab29a467f87ccf9fdc4a845ef0f.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 315, + 297, + 554, + 342 + ], + "blocks": [ + { + "bbox": [ + 313, + 253, + 555, + 287 + ], + "lines": [ + { + "bbox": [ + 313, + 253, + 555, + 287 + ], + "spans": [ + { + "bbox": [ + 313, + 253, + 555, + 287 + ], + "type": "text", + "content": "Table 7. Effectiveness of evidence grounding. Our videotegrounded method yields clear improvement over using the full video. More precise grounding can further enhance our accuracy." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 297, + 554, + 342 + ], + "lines": [ + { + "bbox": [ + 315, + 297, + 554, + 342 + ], + "spans": [ + { + "bbox": [ + 315, + 297, + 554, + 342 + ], + "type": "table", + "html": "
StrategyStatic (Depth=)Dynamic
2345
NExT-QAOriginal58.859.260.260.360.5
Rewritten52.053.455.655.355.4
", + "image_path": "2b25ba2ccfbd2864f810ede221197abc2ceaae122bbfe00fa48b990d8b38b3be.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 346, + 554, + 369 + ], + "lines": [ + { + "bbox": [ + 313, + 346, + 554, + 369 + ], + "spans": [ + { + "bbox": [ + 313, + 346, + 554, + 369 + ], + "type": "text", + "content": "Table 8. Effectiveness of dynamic tree expansion. It yields superior accuracy while increasing reasoning efficiency." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 388, + 555, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 388, + 555, + 556 + ], + "spans": [ + { + "bbox": [ + 313, + 388, + 555, + 556 + ], + "type": "text", + "content": "ing the depth of reasoning yields significant improvements, as complex, long statements are broken down into concise sub-statements that VLMs can understand more effectively. However, extending reasoning beyond the 4th layer offers diminishing returns; for NExT-QA, the original statements' complexity constrains the task, and some 5th-layer sub-statements become overly simplistic and less effective for reasoning. This finding highlights the necessity of our dynamic strategy. Applying the dynamic tree expansion strategy, we can see that the performance outperforms the fixed-depth paradigm. In the meantime, the dynamic strategy increases the reasoning efficiency over the entailment tree, more details about efficiency comparison can be found in our supplementary material." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 567, + 388, + 578 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 567, + 388, + 578 + ], + "spans": [ + { + "bbox": [ + 313, + 567, + 388, + 578 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 581, + 555, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 581, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 581, + 555, + 712 + ], + "type": "text", + "content": "This paper proposed the first video-grounded entailment tree framework for VQA. Moreover, we also contributed a de-biasing procedure to avoid spurious correlations during evaluation and applied it to enhance representative benchmarks. Extensive experiments with five video- and image-based VLMs demonstrate consistent benefits of our method on these benchmarks. Besides, our proposed framework performs on par with state-of-the-art video reasoning methods despite using " + }, + { + "bbox": [ + 313, + 581, + 555, + 712 + ], + "type": "inline_equation", + "content": "257 \\times" + }, + { + "bbox": [ + 313, + 581, + 555, + 712 + ], + "type": "text", + "content": " fewer parameters. While de-biasing hurts VLM accuracy, our framework regains the accuracy losses and is competitive with state-of-the-art VQA methods." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3269" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 144 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 144 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 144 + ], + "type": "text", + "content": "[1] Piyush Bagad, Makarand Tapaswi, and Cees GM Snoek. Test of time: Instilling video-language models with a sense of time. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2503-2516, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 147, + 296, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 147, + 296, + 213 + ], + "spans": [ + { + "bbox": [ + 61, + 147, + 296, + 213 + ], + "type": "text", + "content": "[2] Maciej Besta, Nils Blach, Ales Kubicek, Robert Gerstenberger, Michal Podstawski, Lukas Gianinazzi, Joanna Gajda, Tomasz Lehmann, Hubert Niewiadomski, Piotr Nczyk, et al. Graph of thoughts: Solving elaborate problems with large language models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 17682-17690, 2024. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 214, + 295, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 214, + 295, + 267 + ], + "spans": [ + { + "bbox": [ + 62, + 214, + 295, + 267 + ], + "type": "text", + "content": "[3] Tieyuan Chen, Huabin Liu, Tianyao He, Yihang Chen, Chaofan Gan, Xiao Ma, Cheng Zhong, Yang Zhang, Yingxue Wang, Hui Lin, et al. Mecd: Unlocking multi-event causal discovery in video reasoning. arXiv preprint arXiv:2409.17647, 2024. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 270, + 295, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 270, + 295, + 312 + ], + "spans": [ + { + "bbox": [ + 62, + 270, + 295, + 312 + ], + "type": "text", + "content": "[4] Bhavana Dalvi, Peter Jansen, Oyvind Tafjord, Zhengnan Xie, Hannah Smith, Leighanna Pipatanangkura, and Peter Clark. Explaining answers with entailment trees. arXiv preprint arXiv:2104.08661, 2021. 1, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 314, + 295, + 367 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 314, + 295, + 367 + ], + "spans": [ + { + "bbox": [ + 62, + 314, + 295, + 367 + ], + "type": "text", + "content": "[5] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024.6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 369, + 295, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 369, + 295, + 423 + ], + "spans": [ + { + "bbox": [ + 62, + 369, + 295, + 423 + ], + "type": "text", + "content": "[6] Chaoyou Fu, Yuhan Dai, Yongdong Luo, Lei Li, Shuhuai Ren, Renrui Zhang, Zihan Wang, Chenyu Zhou, Yunhang Shen, Mengdan Zhang, et al. Video-mme: The first-ever comprehensive evaluation benchmark of multi-modal llms in video analysis. arXiv preprint arXiv:2405.21075, 2024. 2, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 425, + 295, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 425, + 295, + 490 + ], + "spans": [ + { + "bbox": [ + 62, + 425, + 295, + 490 + ], + "type": "text", + "content": "[7] Wenyi Hong, Weihan Wang, Qingsong Lv, Jiazheng Xu, Wenmeng Yu, Junhui Ji, Yan Wang, Zihan Wang, Yuxiao Dong, Ming Ding, et al. Cogagent: A visual language model for gui agents. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14281-14290, 2024. 6, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 491, + 295, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 491, + 295, + 534 + ], + "spans": [ + { + "bbox": [ + 62, + 491, + 295, + 534 + ], + "type": "text", + "content": "[8] Nora Kassner, Oyvind Tafjord, Ashish Sabharwal, Kyle Richardson, Hinrich Schuetze, and Peter Clark. Language models with rationality. arXiv preprint arXiv:2305.14250, 2023. 2, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 536, + 295, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 536, + 295, + 590 + ], + "spans": [ + { + "bbox": [ + 62, + 536, + 295, + 590 + ], + "type": "text", + "content": "[9] Muhammad Uzair Khattak, Muhammad Ferjad Naeem, Jameel Hassan, Muzammal Naseer, Federico Tombari, Fahad Shahbaz Khan, and Salman Khan. Complex video reasoning and robustness evaluation suite for video-lmms. arXiv preprint arXiv:2405.03690, 2024. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 591, + 295, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 591, + 295, + 645 + ], + "spans": [ + { + "bbox": [ + 57, + 591, + 295, + 645 + ], + "type": "text", + "content": "[10] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730–19742. PMLR, 2023. 1, 6, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 647, + 295, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 295, + 690 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 295, + 690 + ], + "type": "text", + "content": "[11] Jiapeng Li, Ping Wei, Wenjuan Han, and Lifeng Fan. Intentqa: Context-aware video intent reasoning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11963-11974, 2023. 1, 2, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 692, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 692, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 692, + 295, + 713 + ], + "type": "text", + "content": "[12] Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Yi Liu, Zun Wang, Jilan Xu, Guo Chen, Ping Luo, et al." + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 554, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 333, + 73, + 554, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 554, + 116 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 554, + 116 + ], + "type": "text", + "content": "Mvbench: A comprehensive multi-modal video understanding benchmark. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22195-22206, 2024. 1, 6, 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 118, + 554, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 554, + 161 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 554, + 161 + ], + "type": "text", + "content": "[13] Yicong Li, Xiang Wang, Junbin Xiao, Wei Ji, and Tat-Seng Chua. Invariant grounding for video question answering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2928-2937, 2022. 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 162, + 554, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 162, + 554, + 205 + ], + "spans": [ + { + "bbox": [ + 317, + 162, + 554, + 205 + ], + "type": "text", + "content": "[14] Bin Lin, Yang Ye, Bin Zhu, Jiaxi Cui, Munan Ning, Peng Jin, and Li Yuan. Video-llava: Learning united visual representation by alignment before projection. arXiv preprint arXiv:2311.10122, 2023. 1, 5, 6, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 206, + 554, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 206, + 554, + 249 + ], + "spans": [ + { + "bbox": [ + 316, + 206, + 554, + 249 + ], + "type": "text", + "content": "[15] Huabin Liu, Weixian Lv, John See, and Weiyao Lin. Task-adaptive spatial-temporal video sampler for few-shot action recognition. In Proceedings of the 30th ACM International Conference on Multimedia, pages 6230-6240, 2022. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 251, + 554, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 251, + 554, + 293 + ], + "spans": [ + { + "bbox": [ + 317, + 251, + 554, + 293 + ], + "type": "text", + "content": "[16] Huabin Liu, Weiyao Lin, Tieyuan Chen, Yuxi Li, Shuyuan Li, and John See. Few-shot action recognition via intra- and inter-video information maximization. arXiv preprint arXiv:2305.06114, 2023. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 294, + 554, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 294, + 554, + 338 + ], + "spans": [ + { + "bbox": [ + 317, + 294, + 554, + 338 + ], + "type": "text", + "content": "[17] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26296-26306, 2024. 1, 6, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 339, + 554, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 339, + 554, + 392 + ], + "spans": [ + { + "bbox": [ + 317, + 339, + 554, + 392 + ], + "type": "text", + "content": "[18] Huabin Liu, Xiao Ma, Cheng Zhong, Yang Zhang, and Weiyao Lin. Timecraft: Navigate weakly-supervised temporal grounded video question answering via bi-directional reasoning. In European Conference on Computer Vision, pages 92-107. Springer, 2025. 2, 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 394, + 554, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 394, + 554, + 435 + ], + "spans": [ + { + "bbox": [ + 317, + 394, + 554, + 435 + ], + "type": "text", + "content": "[19] Kate Sanders, Nathaniel Weir, and Benjamin Van Durme. Tv-trees: Multimodal entailment trees for neuro-symbolic video reasoning. arXiv preprint arXiv:2402.19467, 2024. 1, 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 437, + 554, + 471 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 437, + 554, + 471 + ], + "spans": [ + { + "bbox": [ + 316, + 437, + 554, + 471 + ], + "type": "text", + "content": "[20] Oyvind Tafjord, Bhavana Dalvi Mishra, and Peter Clark. Entailer: Answering questions with faithful and truthful chains of reasoning. arXiv preprint arXiv:2210.12217, 2022. 2, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 472, + 554, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 472, + 554, + 514 + ], + "spans": [ + { + "bbox": [ + 317, + 472, + 554, + 514 + ], + "type": "text", + "content": "[21] Xiaohan Wang, Yuhui Zhang, Orr Zohar, and Serena Yeung-Levy. Videoagent: Long-form video understanding with large language model as agent. arXiv preprint arXiv:2403.10517, 2024. 2, 6, 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 516, + 554, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 516, + 554, + 568 + ], + "spans": [ + { + "bbox": [ + 317, + 516, + 554, + 568 + ], + "type": "text", + "content": "[22] Ziyang Wang, Shoubin Yu, Elias Stengel-Eskin, Jaehong Yoon, Feng Cheng, Gedas Bertasius, and Mohit Bansal. Videotree: Adaptive tree-based video representation for llm reasoning on long videos. arXiv preprint arXiv:2405.19209, 2024. 2, 6, 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 571, + 554, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 571, + 554, + 624 + ], + "spans": [ + { + "bbox": [ + 317, + 571, + 554, + 624 + ], + "type": "text", + "content": "[23] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824-24837, 2022. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 625, + 554, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 625, + 554, + 658 + ], + "spans": [ + { + "bbox": [ + 316, + 625, + 554, + 658 + ], + "type": "text", + "content": "[24] Nathaniel Weir and Benjamin Van Durme. Dynamic generation of grounded logical explanations in a neuro-symbolic expert system. arXiv preprint arXiv:2209.07662, 2022. 2, 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 659, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 659, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 659, + 554, + 713 + ], + "type": "text", + "content": "[25] Hao Wu, Huabin Liu, Yu Qiao, and Xiao Sun. Dibs: Enhancing dense video captioning with unlabeled videos via pseudo boundary enrichment and online refinement. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18699-18708, 2024. 1" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "3270" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 485 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 127 + ], + "type": "text", + "content": "[26] Junbin Xiao, Xindi Shang, Angela Yao, and Tat-Seng Chua. Next-qa: Next phase of question-answering to explaining temporal actions. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9777–9786, 2021. 1, 2, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 294, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 294, + 183 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 294, + 183 + ], + "type": "text", + "content": "[27] Junbin Xiao, Angela Yao, Yicong Li, and Tat-Seng Chua. Can i trust your answer? visually grounded video question answering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13204-13214, 2024. 1, 2, 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 185, + 294, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 185, + 294, + 239 + ], + "spans": [ + { + "bbox": [ + 56, + 185, + 294, + 239 + ], + "type": "text", + "content": "[28] Yu Yang, Besmira Nushi, Hamid Palangi, and Baharan Mirzasoleiman. Mitigating spurious correlations in multimodal models during fine-tuning. In International Conference on Machine Learning, pages 39365-39379. PMLR, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 240, + 294, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 240, + 294, + 295 + ], + "spans": [ + { + "bbox": [ + 56, + 240, + 294, + 295 + ], + "type": "text", + "content": "[29] Shunyu Yao, Dian Yu, Jeffrey Zhao, Izhak Shafran, Tom Griffiths, Yuan Cao, and Karthik Narasimhan. Tree of thoughts: Deliberate problem solving with large language models. Advances in Neural Information Processing Systems, 36, 2024. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 297, + 294, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 297, + 294, + 341 + ], + "spans": [ + { + "bbox": [ + 56, + 297, + 294, + 341 + ], + "type": "text", + "content": "[30] Shoubin Yu, Jaemin Cho, Prateek Yadav, and Mohit Bansal. Self-chained image-language model for video localization and question answering. Advances in Neural Information Processing Systems, 36, 2024. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 342, + 294, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 342, + 294, + 385 + ], + "spans": [ + { + "bbox": [ + 56, + 342, + 294, + 385 + ], + "type": "text", + "content": "[31] Ce Zhang, Taixi Lu, Md Mohaiminul Islam, Ziyang Wang, Shoubin Yu, Mohit Bansal, and Gedas Bertasius. A simple llm framework for long-range video question-answering. arXiv preprint arXiv:2312.17235, 2023. 2, 6, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 386, + 294, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 386, + 294, + 441 + ], + "spans": [ + { + "bbox": [ + 56, + 386, + 294, + 441 + ], + "type": "text", + "content": "[32] Gengyuan Zhang, Yurui Zhang, Kerui Zhang, and Volker Tresp. Can vision-language models be a good guesser? exploring vlms for times and location reasoning. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 636-645, 2024. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 442, + 294, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 442, + 294, + 485 + ], + "spans": [ + { + "bbox": [ + 56, + 442, + 294, + 485 + ], + "type": "text", + "content": "[33] Hang Zhang, Xin Li, and Lidong Bing. Video-llama: An instruction-tuned audio-visual language model for video understanding. arXiv preprint arXiv:2306.02858, 2023. 1, 6, 7" + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3271" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Community Forensics_ Using Thousands of Generators to Train Fake Image Detectors/f9d7ae16-50eb-4d82-828a-d6dd35d593a9_content_list.json b/2025/Community Forensics_ Using Thousands of Generators to Train Fake Image Detectors/f9d7ae16-50eb-4d82-828a-d6dd35d593a9_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..d5079ec7ae46dbb543a8c99ea802b69dae8de656 --- /dev/null +++ b/2025/Community Forensics_ Using Thousands of Generators to Train Fake Image Detectors/f9d7ae16-50eb-4d82-828a-d6dd35d593a9_content_list.json @@ -0,0 +1,2002 @@ +[ + { + "type": "text", + "text": "Community Forensics: Using Thousands of Generators to Train Fake Image Detectors", + "text_level": 1, + "bbox": [ + 176, + 130, + 818, + 175 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jeongsoo Park Andrew Owens University of Michigan", + "bbox": [ + 336, + 203, + 658, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 273, + 326, + 289 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "One of the key challenges of detecting AI-generated images is spotting images that have been created by previously unseen generative models. We argue that the limited diversity of the training data is a major obstacle to addressing this problem, and we propose a new dataset that is significantly larger and more diverse than prior works. As part of creating this dataset, we systematically download thousands of text-to-image latent diffusion models and sample images from them. We also collect images from dozens of popular open source and commercial models. The resulting dataset contains 2.7M images that have been sampled from 4803 different models. These images collectively capture a wide range of scene content, generator architectures, and image processing settings. Using this dataset, we study the generalization abilities of fake image detectors. Our experiments suggest that detection performance improves as the number of models in the training set increases, even when these models have similar architectures. We also find that increasing the diversity of the models improves detection performance, and that our trained detectors generalize better than those trained on other datasets. The dataset can be found in", + "bbox": [ + 88, + 306, + 485, + 623 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "https://jespark.net/projects/2024/community_forensics", + "bbox": [ + 89, + 625, + 455, + 638 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 666, + 220, + 681 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Our ability to automatically generate realistic images is quickly outpacing our ability to detect them, potentially leading to a state of affairs in which neither humans nor machines can reliably tell real from fake. While the field of image forensics has been developing methods to address this problem, existing fake image detectors still struggle with generalization. These methods often excel at detecting images from generators that were present in their training sets, but fail when given images sampled from unseen models [87, 101, 102, 125].", + "bbox": [ + 89, + 688, + 482, + 838 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "A core challenge is dealing with the large amounts of variation between models. Each generator has a potentially unique combination of the architecture, loss function, and training distribution. Even seemingly minor differences in", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/653df4c788aa9892c3e27775054108cb9797cf24bedc786da1a1ff5df83470f9.jpg", + "image_caption": [ + "Figure 1. Performance vs. model diversity. We use images sampled from different numbers of open source latent diffusion models in our Community Forensics dataset to train fake image detectors (shown in Fig. 2a). We fix the total number of images and only vary the number of models. The detector's performance increases across all generator types as we train from more models, even though these added models are entirely latent diffusion. This improvement is largest for test images from out-of-distribution generators, such as pixel-based diffusion models or GANs." + ], + "image_footnote": [], + "bbox": [ + 526, + 276, + 890, + 481 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "low-level image processing details, such as the ways that training images are resized or compressed, can strongly influence detection accuracy [125]. As a result of these model-specific idiosyncrasies, a generator's images may evade detection, even when images from architecturally similar models exist in the training set. This issue has been exacerbated by the thousands of open source models that are now available online, many of which extend pretrained base models in complex ways.", + "bbox": [ + 511, + 628, + 908, + 763 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We hypothesize that the lack of diversity in training datasets is a major source of these shortcomings. Although today's datasets often contain millions of fake images, they come from a relatively small number of generators. As a result, this data fails to capture many sources of variation that one might encounter in the wild. These limitations also make it challenging to accurately benchmark performance, since it is easy for cues that work well on one set of generators to fail on others.", + "bbox": [ + 511, + 763, + 908, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "8245", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d1f5072360d364ca47d054567c71f9c5839673cf0b8b2d63a1aa62a5e415580c.jpg", + "image_caption": [ + "Dreamshaper v6" + ], + "image_footnote": [], + "bbox": [ + 181, + 90, + 250, + 152 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/6fe9aa0173c01026b84b59881178e13aa1eb95ccb09bf03e14da648a3e1b7ce9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 261, + 90, + 326, + 152 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/f18b5a9ce3439f4e81c3684d1eb232b9139713186c8cf0fbac687c95099c022a.jpg", + "image_caption": [ + "Plat diffusion" + ], + "image_footnote": [], + "bbox": [ + 341, + 90, + 411, + 152 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/5fce98152588acb7455cc5f2fa8bc91cfea7bb6956a8c43e61e19ce49c1c93c8.jpg", + "image_caption": [ + "SoteMix" + ], + "image_footnote": [], + "bbox": [ + 421, + 90, + 493, + 152 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "", + "bbox": [ + 421, + 97, + 493, + 152 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/d1ddd487777710b7b38dcd33a26da7bfbc50be43bf1bcb1148ba2820d91ca332.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 421, + 160, + 493, + 214 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/51c233be4c773bc38d072851e63c7202864285b451288867fda884ec0b8f9606.jpg", + "image_caption": [ + "Technoalbum" + ], + "image_footnote": [], + "bbox": [ + 421, + 220, + 493, + 276 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Nightvisionxl 0791", + "bbox": [ + 341, + 276, + 413, + 284 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/2ebea408eef862ea73fa5d22b6093a1f7f0d775f83e9386a3ef0bae503afa177.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 341, + 220, + 413, + 276 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Nightvisionxl 0791", + "bbox": [ + 341, + 276, + 413, + 284 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Albedobase x1 v21", + "bbox": [ + 421, + 276, + 493, + 284 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/4326e416ed416ee6148740a80fae62fc0e38844e93d1775e1ce55d502cd1365c.jpg", + "image_caption": [ + "Nightvisionxl 0791" + ], + "image_footnote": [], + "bbox": [ + 341, + 286, + 413, + 337 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/3ba0c90a0452fa25ad27528355032421adeb240c5c1ccdd172d3d1d27736ef00.jpg", + "image_caption": [ + "Enasswhidaly", + "(a) Systematically collected diffusion models (4763 models)", + "Figure 2. The Community Forensics dataset. Our dataset contains images sampled from three types of generative models. (a) We systematically download open-source latent diffusion models from a model-sharing community [34, 123]. (b) We select popular open source generators with a variety of architectures and training procedures. (c) We sample from both closed and open state-of-the-art commercial models. We present example images and their corresponding model names." + ], + "image_footnote": [], + "bbox": [ + 341, + 345, + 413, + 398 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Albedobase x1 v21", + "bbox": [ + 421, + 276, + 493, + 284 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/c0ea5d2a9ad257f237e85f37379d6059719c942552d2227be09afbb897f5d1a4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 421, + 286, + 493, + 337 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/bcd1b05f51e19ae663ab794f3f071b619c39102a82daab31320e9ca1bebf17e0.jpg", + "image_caption": [ + "Msart 541" + ], + "image_footnote": [], + "bbox": [ + 421, + 345, + 493, + 398 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/c1f0500db9b788ec7a905b477645e3d86b575ff187a21068ba19fec9ad579572.jpg", + "image_caption": [ + "91", + "First avarfar" + ], + "image_footnote": [], + "bbox": [ + 504, + 90, + 573, + 152 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/914cb7b7cd514259ed7b6e25c49761a1b827a8fc3a1aa3885ae3d7d40d418705.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 160, + 573, + 214 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/fa4a628fee64f9e2493ca5a29c7054ea0efa00da4dbce0c165ebf526e61f7b2a.jpg", + "image_caption": [ + "AnalogMadness real", + "DucHaitenAnimated" + ], + "image_footnote": [], + "bbox": [ + 504, + 222, + 573, + 276 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/b9218cfe77d7f7d2a088456a4948bad36fa0cf4db97cbdfc86138418761ef07d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 286, + 573, + 337 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/e08a81e17a58d551f93f4cd9eb9bf9ece3a23e23c149da6b6c1e1f665a7316db.jpg", + "image_caption": [ + "A can thr" + ], + "image_footnote": [], + "bbox": [ + 504, + 345, + 573, + 398 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/64d1bc6de9df69c3b6489c266458244477b6c5581016a61fc98459b86bef746a.jpg", + "image_caption": [ + "DeepFloyd", + "StyleGANXL" + ], + "image_footnote": [], + "bbox": [ + 596, + 114, + 666, + 167 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/52f32b7d1105a788025b0416e2c1cf562640a87f9ffd3bd5258f42ec65f18f67.jpg", + "image_caption": [ + "VQDiffusion" + ], + "image_footnote": [], + "bbox": [ + 673, + 114, + 745, + 167 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/e7b8aeb2ec6a6ce68a2797374b002d91d0ea47aa5cb3f6258777ff4b876b365b.jpg", + "image_caption": [ + "GuidedDiffusion", + "(b) Manually chosen open source models (19 models)" + ], + "image_footnote": [], + "bbox": [ + 673, + 175, + 743, + 228 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/da4130d5200db2a94b4074279315b43279828e5da2a2bf30dcad9390b1ce0eb9.jpg", + "image_caption": [ + "StyleGAN2" + ], + "image_footnote": [], + "bbox": [ + 751, + 114, + 821, + 167 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/d7f0d9b425fba9ca111eb271725ffab93f8510480c37387aea1ad9086ba63d4b.jpg", + "image_caption": [ + "Glide" + ], + "image_footnote": [], + "bbox": [ + 751, + 175, + 821, + 228 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/fc175043f62aa72dbea529adae15d4b9ab30ec1ed9675b72e654e175288e9064.jpg", + "image_caption": [ + "DiT" + ], + "image_footnote": [], + "bbox": [ + 830, + 114, + 898, + 167 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/410392cc44f32b435cf1546850125a2658df3053f13c091bc89b1f4bb159ff87.jpg", + "image_caption": [ + "CIPS" + ], + "image_footnote": [], + "bbox": [ + 830, + 175, + 898, + 228 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/11b1f9cf60b78ad52a794a446d7ac4f29f65ef9e149766a7093a8363e338a34e.jpg", + "image_caption": [ + "Dalle2" + ], + "image_footnote": [], + "bbox": [ + 596, + 273, + 665, + 325 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/475dbce8cf8f39c96856c26a14e28cb4c347ad99ebcf60710ed8a06066669514.jpg", + "image_caption": [ + "FLUX dev", + "(c) Commercial models (11 models)" + ], + "image_footnote": [], + "bbox": [ + 596, + 334, + 665, + 387 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/ad7c7fedf172621ef758bba396770ceb5b4b15d01d989f8384f8b4329781a5c3.jpg", + "image_caption": [ + "MidjourneyV6:" + ], + "image_footnote": [], + "bbox": [ + 674, + 273, + 743, + 325 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/ef5a72b79b986288854273ed87a7dec6f25c0515ffaa242cdc9adb4f839165da.jpg", + "image_caption": [ + "Firefly Image2" + ], + "image_footnote": [], + "bbox": [ + 671, + 335, + 743, + 387 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/aa991346df578eeb395ee35990d987960343854f61dc320668f545a73e0cd742.jpg", + "image_caption": [ + "MidjourneyV5 2" + ], + "image_footnote": [], + "bbox": [ + 751, + 273, + 821, + 325 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/bbccfa2ca88d61d08de5831fc11c6f7463100e834f77a4a353170e05ff30e38c.jpg", + "image_caption": [ + "Firefly Image3" + ], + "image_footnote": [], + "bbox": [ + 751, + 334, + 821, + 387 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/0b8ea5970c68a34b9f7ad0ed8a746394eaf18029cfac47be0388eefe418a9888.jpg", + "image_caption": [ + "Dalle3" + ], + "image_footnote": [], + "bbox": [ + 830, + 273, + 898, + 327 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/bf07435f7a74b9f29f483914d3cb78a88d28f7bbddd0c65b23e2c9f1c281c58c.jpg", + "image_caption": [ + "IdeogramV1" + ], + "image_footnote": [], + "bbox": [ + 830, + 334, + 898, + 387 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address these problems, we propose Community Forensics, a dataset that is significantly more diverse and comprehensive than those in prior works (Fig. 2). Our dataset contains images generated by: (a) thousands of systematically downloaded open-source latent diffusion models, (b) hand-selected open source models with various architectures, and (c) state-of-the-art commercial models.", + "bbox": [ + 89, + 489, + 483, + 594 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To acquire large numbers of models, we sample images from thousands of text-to-image diffusion models hosted on a popular model-sharing website, Hugging Face [34]. We exploit the fact that these models use a common programming library [123] and thus can be sampled in a standardized way. A large fraction of them are extensions of Stable Diffusion [100], but collectively capture a variety of common model variations, such as in the architecture, image processing, and image content. We also sample images from many other open source models, including GANs [37], autoregressive models [39], and consistency models [79, 114]. To ensure sufficient diversity in the image content of the generated images, we sample generators using captions sourced from a mix of several existing datasets whenever possible.", + "bbox": [ + 89, + 595, + 483, + 806 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our dataset contains 4803 distinct models, roughly $34 \\times$ more than the most extensive previous dataset that sample images from generative models [5, 7, 15, 32, 43, 87, 125, 133], and covers a variety of recent model designs (Fig. 2).", + "bbox": [ + 89, + 809, + 483, + 869 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We use this dataset to study generalization in the generated image detection problem. Our experiments confirm", + "bbox": [ + 89, + 869, + 483, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "that increasing the diversity of data improves generalization, consistent with previous findings in other areas of computer vision [48, 107, 108]. In contrast to other work, our dataset aims to increase diversity via number of generators, an underexplored axis of diversity. Through experiments, we find:", + "bbox": [ + 511, + 489, + 906, + 565 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Classifiers trained on our dataset obtain strong performance, both on our newly proposed evaluations and on multiple previously-proposed benchmarks.", + "- Adding more generative models improves generalization. Fig. 1 demonstrates the performance of fake image detection when trained on samples from varying numbers of diffusion models. Notably, the performance improves as more models are added, even across different architectures.", + "- Including diverse generative model architectures significantly improves results, since classifiers do not fully generalize between generator architectures.", + "- Standard classifiers perform well. In contrast to observations from recent work, we find that end-to-end training of classifiers based on CNNs or ViTs generalizes well, with qualitatively similar to that of other recognition problems." + ], + "bbox": [ + 513, + 566, + 906, + 792 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 513, + 804, + 650, + 819 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Datasets for detecting generated images. A number of datasets have been proposed for specifically detecting \"deepfake\" images containing manipulated faces [25, 63, 67, 71, 101, 102, 134]. Rather than focusing on face manipulation, we address creating general-purpose meth", + "bbox": [ + 511, + 825, + 906, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "8246", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ods that can detect images that have been directly produced by generative models. Wang et al. [125] proposed a widely-used dataset of CNN-generated images, mixing images from GANs [8, 12, 58, 59, 90, 132] with other models [9, 10, 18, 70, 102]. This work showed that forensics models generalize between generative models, providing motivation for training on large datasets of diverse generators. However, their classifier was trained on images from a single GAN and was highly sensitive to data augmentation parameters, and more recent work shows that it does not generalize to newer models [14, 87]. Ojha et al. [87] introduced a dataset of recent diffusion models and found that training a linear classifier on CLIP features [96] extracted from ProGAN-generated images performed well. Cozzolino et al. [15] extend this work by studying the performance of CLIP-based detectors on various generative models and datasets. Epstein et al. [32] simulated detecting fake images in an online way by training a detector up to a certain year and testing it on generators released after that year. Zhu et al. [133] collected 1.4M generated images from 8 different generators. These datasets, however, only consider a handful of models (less than 20 each), limiting the generalization of their detectors. Asnani et al. proposed RED116 [5], collecting 116K images from 116 generative models for predicting the model hyperparameters of a given generated image. Guo et al. [43] extends RED116 to 140 generative models with 140K images total. We improve upon these works by collecting much more diverse generative models to improve the performance and generalization of the detector. In concurrent work, Hong et al. [49] acquires user-created images from Midjourney and CivitAI. This strategy is complementary to ours: while it aims to collect in-the-wild fake images, its distribution is centered on images that users share, and the models are not necessarily identifiable, making it challenging to rigorously analyze the dataset's contents and to interpret experiments conducted on it.", + "bbox": [ + 89, + 90, + 485, + 636 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Fingerprint-based image forensics methods. Classic work on image forensics relied on methods based on image statistics [95] and physical constraints [56], rather than learning. A number of datasets have been created for detecting images that have been manipulated using traditional methods, such as with photo editors [21, 26, 53, 64, 84]. Recent works focus on detecting synthetic images by inspecting the generator fingerprints. Zhang et al. [131] and Marra et al. [80] proposed identifying the spatial fingerprints left by the generator to detect synthetic images. Others focus on spectral anomalies to detect synthetic images. Durall et al. [29] and Dzanic et al. [30] identified that CNN-generated images fail to reproduce certain spectral properties of real images. Corvi et al. [14] study the frequency fingerprints of the generated images and analyzes the cross-architecture generalization of the detector. Bammey [7] uses high-frequency artifacts to detect generated images. However, these approaches may be", + "bbox": [ + 89, + 643, + 486, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "brittle since the artifacts they rely on can be eliminated by post-processing [15]. We instead approach this problem in a data-driven manner, scaling the number of models, images, and architectures. Recent work has created ensembles of fake image classifiers [50]. In parallel, researchers have detected text generated by language models using supervised learning and heuristics [6, 35, 55, 65, 82, 103, 113, 121], which closely resemble those in visual forensics. However, no existing techniques that we are aware of aim to collect comprehensive datasets of community-created generators.", + "bbox": [ + 511, + 90, + 908, + 243 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Out-of-distribution generalization. Our work is related to the out-of-distribution recognition problem as it involves generalizing to unseen generators and image processing pipelines. A variety of approaches have been proposed for this problem, based on likelihood ratios [72, 98, 128], self-supervision [47, 83, 109, 124], internal model statistics [46, 104], temperature scaling [3, 73], and via energy-based models [28, 31, 75]. Works by Schuhmann et al. [108] and Hendrycks et al. [48] show that diverse training data and data augmentation are important to improving the robustness to out-of-distribution samples. Our results are in line with these conclusions, as we find that a diverse set of generative models and stronger augmentations improve generalization.", + "bbox": [ + 511, + 247, + 910, + 445 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. The Community Forensics Dataset", + "text_level": 1, + "bbox": [ + 511, + 457, + 821, + 474 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To support our goal of studying generalization in generated image detection, we collect a dataset of images sampled from a wide range of models (Fig. 2). Our dataset consists of: (a) a large and systematically collected set of \"in-the-wild\" text-to-image latent diffusion models obtained from a model-sharing website, (b) hand-selected models from other open source architectures, and (c) closed and open state-of-the-art commercial models. We also pair these generated images with real images from other datasets. For all images in our dataset, we preserve the original image format where possible, without any additional compression or resizing. This is to mitigate potential bias and performance degradation in out-of-distribution settings due to unwanted artifacts [40, 42, 97]. Our dataset contains significantly more models than previous works (Tab. 1) and spans a wider range of architectures, processing pipelines, and semantic content.", + "bbox": [ + 511, + 479, + 908, + 720 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Systematically collecting generative models", + "text_level": 1, + "bbox": [ + 511, + 726, + 880, + 742 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We perform our systematic collection using publicly available, open source models that use the Hugging Face diffusers library [34, 123] because: 1) it is a popular library for creating text-to-image models and is widely used by hobbyists, 2) thousands of such models are publicly indexed, and 3) it provides a standard interface by which we can sample images. We process the models in the order of popularity, as indicated by the number of downloads. Our", + "bbox": [ + 511, + 744, + 908, + 867 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "1We use \"open source\" to describe models with public weights and code, even if they may be closed in some aspect (e.g., private training data).", + "bbox": [ + 511, + 875, + 906, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "8247", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/5412ad9454e799c6beca1aee8f12485a99c281b83568c9ef69c3ebdea8131176.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetModelsImagesArchitecturesTraining setup
Wang et al. [125]11362KGAN, Perceptual, Deepfake, ...ProGAN [58] vs. LSUN [129]
Ojha et al. [87]4*10K*GAN, Perceptual, Diffusion, ...ProGAN [58] vs. LSUN [129]
Epstein et al. [32]14570KDiffusionDiffusion vs. LAION [108]
Cozzolino et al. [15]1826KDiffusionLDM [100] vs. MS-COCO [74]
Synthbuster [7]910KDiffusionDiffusion vs. Dresden [36]
GenImage [133]81.4MDiffusion, GANDiffusion, GAN vs. ImageNet [23]
RED116 [5]116116KGAN, VAE, Autoregressive, ...Many vs. Many
RED140 [43]140140KDiffusion, GAN, VAE, ...Many vs. Many
Ours48032.7MDiffusion, GAN, Autoregressive, ...Many vs. Many
", + "bbox": [ + 91, + 88, + 903, + 250 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Table 1. Comparison with existing forensics datasets. We compare the size of the dataset with existing datasets containing identifiable generative models. We only count the number of generated images. Our dataset contains significantly more generative models than prior works. $\\ast$ : Only counting the unique evaluation set by Ojha et al. [87] as their dataset is based on Wang et al. [125].", + "bbox": [ + 89, + 258, + 906, + 303 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "pipeline downloads each model and extracts relevant hyperparameters (e.g., number of diffusion steps and guidance scale), sampling pipeline configurations, and metadata, from both the model cards on the model-sharing webpage and the metadata retrieved by the diffusers library [34, 123]. We sample images using a distribution of text prompts obtained from various real datasets (Sec. 3.3). Since experiments suggest diminishing returns for repeatedly sampling from any given model, we sample a few hundred images from each one. We sample 4763 models with approximately 403 images each, for a total of $1.9\\mathrm{M}$ images from this process.", + "bbox": [ + 88, + 313, + 480, + 479 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "While the lack of documentation in each model and the scale of data collection make it challenging to exactly characterize the model designs in this set, they appear to be entirely (or almost entirely) based on latent diffusion. More specifically, we categorize models as being based on latent diffusion if they perform a denoising process on a latent representation. Based on this criterion and the self-reported tags, all models in our systematically collected set appear to be based on latent diffusion. While pixel-based diffusion models also use the diffusers library (e.g., DeepFloyd [22]), they were incompatible with our automated generation pipeline. We record such incompatible models and manually sample them to either construct an out-of-distribution test set (Sec. 3.4), or as manually-chosen models for training data (Sec. 3.2).", + "bbox": [ + 89, + 481, + 482, + 691 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We show examples of the sampled images in Fig. 2. In the supplement, we provide examples of models and information from their project pages. These models generate a variety of images, with various types of semantic content and preprocessing. For example, a large fraction of these models adapt variations of a popular pretrained latent diffusion model, Stable Diffusion [100], to different downstream applications, and use a number of adaptation strategies (e.g., using LoRA [51]). We provide the model metadata with each image to enable other possible forensics applications. We discuss these in the supplement and provide information about image and model licenses.", + "bbox": [ + 89, + 694, + 482, + 875 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Collecting images from other architectures", + "text_level": 1, + "bbox": [ + 511, + 313, + 877, + 329 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Images from manually chosen models. To ensure that our dataset contains a broader range of models, we manually select 19 models from public repositories and sample an average of 40,738 images per model, resulting in 774K images total. We note that this number is itself on par with (or more than) many prior datasets with identifiable generative models [7, 15, 32, 87, 125, 133]. We include several GANs (e.g., StyleGANs [60-62, 106], BigGAN [8], StyleSwin [130], GigaGAN [57], ProGAN [58], ProjectedGAN [105], GANformer [52], SAN [115], and CIPS [4]), pixel-based diffusion models (e.g., GLIDE [85], ADM [24], and DeepFloyd [22]), latent diffusion models (e.g., VQ-Diffusion [41], Diffusion Transformers [92], and Latent Flow Matching [20]), and an autoregressive model (Taming Transformers [33]).", + "bbox": [ + 511, + 332, + 906, + 544 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Images from commercial models. We sample 15K images from 11 commercial models using LAION-based captions to evaluate the generalization to state-of-the-art models with typically unknown architectures: DALL-E 2, 3 [88, 89], Ideogram V1, V2 [2], Midjourney V5, V6 [81], Firefly Image 2, 3 [1], FLUX.1-dev, Schnell [68], and Imagen 3 [38].", + "bbox": [ + 511, + 551, + 908, + 645 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Collecting real images", + "text_level": 1, + "bbox": [ + 511, + 651, + 720, + 667 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To help study how real images influence forensics models, we source real images from a variety of existing datasets: LAION [107], ImageNet [23], COCO [74], FFHQ [59], CelebA [76], MetFaces [60], AFHQ [13], Forchheim [44], IMD2020 [86], Landscapes HQ [112], and VISION [111].3", + "bbox": [ + 511, + 671, + 908, + 750 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Curating the evaluation set", + "text_level": 1, + "bbox": [ + 511, + 755, + 758, + 771 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We construct our evaluation set using the incompatible models from our automated sampling pipeline, commercial models (Sec. 3.2), and manually collected open source models. The evaluation set comprises 26K images sampled from 21 models not included in the training set.", + "bbox": [ + 511, + 775, + 908, + 852 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "2We note that this definition includes latent consistency models [78].", + "bbox": [ + 107, + 886, + 470, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "3Following common convention, we refer to these images as real images, even though they may be synthetic (e.g., graphic design). More precisely, our goal is to distinguish \"AI-generated\" images from the originals.", + "bbox": [ + 511, + 862, + 906, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "8248", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This includes our commercial models set and an additional 11K images from 10 models: Deci Diffusion V2 [120], GALIP [119], KandinskyV2.2 [110], Kvikontent [66], LCM-LoRA-SDv1.5, LCM-LoRA-SDXL, LCM-LoRA-SSD1B [79], Stable Cascade [93], DF-GAN [118], and HDiT [16], sampled using RAISE [19], ImageNet [23], FFHQ [59], and COCO [74]-based captions.", + "bbox": [ + 88, + 90, + 483, + 196 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The generated images are paired with the source real data that are used to prompt the generators. However, since some of the real datasets do not have appropriate licenses for redistribution (e.g., LAION [107, 108]), we created a public version of our evaluation set by pairing the generated images with openly licensed COCO [74] and FFHQ [59] which allow redistribution for non-commercial purposes. The public version of our evaluation set will serve as an easily reproducible and shareable evaluation set that will complement our default set. We will refer to our default set as the comprehensive evaluation set. We will also release the instructions to reconstruct our comprehensive set. However, note that it may not be possible to exactly reconstruct this set in the future due to link rot.", + "bbox": [ + 91, + 198, + 485, + 407 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Generating images", + "text_level": 1, + "bbox": [ + 89, + 414, + 272, + 430 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Unconditional models are sampled until we reach the desired number of images. For class conditional models, we sample an equal number of images per class. Text-conditional models are sampled using captions obtained from real images (Sec. 3.3). We either use captions that are already present in the dataset or use BLIP [69] to generate them. Some models such as GigaGAN [57] and HDiT [16] do not provide a pretrained model, so we instead use their pre-generated images. Generated images are saved in PNG format to avoid compression artifacts. However, Firefly [1] generated images are saved in JPEG format as their web UI only allows downloading in JPEG.", + "bbox": [ + 89, + 433, + 485, + 614 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 621, + 223, + 638 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We use our dataset to conduct a study of generalization in visual forensics, asking a number of questions: (1) How well do forensics models trained on our dataset generalize to unseen models? (2) Does adding more models improve detection performance? (3) How does diversity of the training data affect performance? (4) What architectures and data augmentation schemes are most successful?", + "bbox": [ + 89, + 642, + 483, + 750 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Training image forensics models", + "text_level": 1, + "bbox": [ + 89, + 753, + 375, + 770 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We train binary classifiers that detect generated images using our dataset to study the generalization in image forensics. We construct our training set of 5.4M images by pairing 2.7M generated images with 2.7M real images.", + "bbox": [ + 89, + 773, + 483, + 835 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training and evaluation setup. We evaluate the models trained on our dataset and compare them with prior works [43, 87, 125, 133]. As RED140 [43] already contains RED116 [5], we do not compare with RED116 dataset.", + "bbox": [ + 89, + 839, + 485, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Following prior works [87, 125], we use the threshold-independent mean average precision (mAP) and accuracy (Acc.) as our evaluation metrics. We compute the mAP and accuracy by averaging the results of each generative model. We use five evaluation sets: Wang et al. [125], Ojha et al. [87], Synthbuster [7], GenImage [133], and our evaluation set. All evaluation sets apart from GenImage [133] evaluate out-of-distribution performance for all classifiers. GenImage [133] evaluation set, however, contains the same set of generators used in training, and is an in-distribution evaluation set for their classifiers. Concretely, the evaluation set by Wang et al. [125] and Ojha et al. [87, 125] contains models such as CRN [10], CycleGAN [132], DALL-E [88], DeepFake [25], IMLE [70], SAN [18], StarGAN [12], and SITD [9] which are unseen by both their and our classifiers. Synthbuster [7] evaluation set is comprised of RAISE [19]-based synthetic images of DALL-E [88, 89], Firefly [1], Glide [85], Midjourney [81], and Stable Diffusion [94, 100], and is mostly out of distribution for all classifiers. GenImage [133] evaluation set is a validation split of their training set; they use an identical set of models to train their classifier: ADM [24], BigGAN [8], Glide [85], Midjourney [81], Stable Diffusion [100], VQ-Diffusion [41], and Wukong [127].", + "bbox": [ + 511, + 90, + 908, + 439 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Model architecture. Building on prior works which mainly used CLIP-ViT [27, 54, 96] and ResNet-50 [45], we consider ViT [27] and ConvNeXt [77] pretrained models for our classifiers. We use a plain ViT-S backbone [27] pretrained on CLIP objective [54, 96] using LAION-2B [108], ImageNet 21K, and ImageNet 1K datasets [23]. We also experiment with a ConvNeXt-S model [77] pretrained on ImageNet 21K and ImageNet 1K datasets [23]. We replace the classification head with a linear layer with sigmoid activation that outputs the probability of the image being generated. Unlike prior works [15, 87] that freeze the CLIP-ViT backbone, we train the backbone end-to-end. The models are obtained through timm [11, 126] library on Hugging Face. We experiment with two input resolutions, $224^2$ and $384^2$ , to evaluate the impact of the input resolution on the detector's performance. We denote the detector with $384^2$ input resolution as High res. We implement the models using PyTorch [91]. The hyperparameters are detailed in the supplement.", + "bbox": [ + 511, + 450, + 908, + 723 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Data augmentation. Prior work considered augmentations that were designed to simulate postprocessing, such as flipping, cropping, Gaussian blur, and JPEG recompression to train their detectors [7, 15, 87, 125]. We propose an augmentation scheme that extends this approach and compare it with previously proposed augmentation methods. We expand the set of augmentations to handle additional transformations that can occur in the wild, such as padding, resizing, rotation, and shear, and integrate them into a framework that can apply complex sequences of transformations. We introduce a modified version of RandAugment [17] that applies a", + "bbox": [ + 511, + 734, + 910, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "8249", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/6436d7f66a12015bf915de151eb2fa7d6abc03563f100037b3d21b4e3c0236cc.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelEvaluation Set (mAP)Evaluation Set (Acc)
Wang et al. [125]Ojha et al. [87]SB [7]GenImage [133]OursMeanWang et al. [125]Ojha et al. [87]SB [7]GenImage [133]OursMean
Wang et al. [125]0.8970.6960.5160.6420.5370.6000.6480.7140.5270.5080.5330.5130.5170.552
Ojha et al. [87]0.9390.9570.6200.7970.5920.6560.7600.7910.8210.5320.6410.5400.5480.646
GenImage [133]0.9290.9840.8130.9990.9120.9680.9340.7950.9660.7190.9900.8180.8860.862
RED140 [43] - High res.0.9000.9540.7650.9270.7640.8610.8620.6940.7800.5580.6740.5620.5650.639
Ours0.9640.9910.9040.9900.9710.9770.9660.8730.9500.8180.9460.8610.8880.889
Ours - High res.0.9670.9960.9740.9980.9870.9940.9860.9010.9700.9080.9570.8920.9120.923
", + "bbox": [ + 98, + 88, + 906, + 220 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Generalization of AI-generated image detectors across benchmarks. We evaluate the classifiers trained on our dataset on several benchmarks, including our own. We also evaluate several previously released classifiers. Our Comprehensive set (abbreviated as Comp.) pairs the generated images with original real data; the Public set pairs them with openly licensed COCO [74] and FFHQ [59] for license-compliant redistribution of the evaluation set (Sec. 3.4). We use plain CLIP-ViT-S [27, 54, 96] architecture with $224^{2}$ and $384^{2}$ (High res.) input resolutions, Wang et al. [125] and GenImage [133] use ResNet-50 [45] with $224^{2}$ input resolution, and Ojha et al. [87] uses CLIP-ViT-L with $224^{2}$ input resolution as the backbone. Our classifiers show robust performance across all evaluation sets, outperforming all baselines in out-of-distribution settings ([7, 87, 125] and Ours) and nearly matches GenImage [133] on its in-distribution evaluation set. $\\spadesuit$ : RED140 [43] is trained following our training procedure with $384^{2}$ input resolution as they do not provide a trained classifier.", + "bbox": [ + 89, + 224, + 906, + 335 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/962707b7f8c50d48ce8f78f437e2c69577554671c1e94e0602d66f8aeaec0aeb.jpg", + "image_caption": [ + "Figure 3. Performance across generator types. We evaluate the classifier performance across five generator types - latent diffusion, pixel diffusion, commercial models, GANs, and other architecture (Stable Cascade [93]), and report the mean. We additionally evaluate three forensics methods, Aerblade [99], LGrad [116], and NPR [117]. Our classifiers show robust performance across all generator types, whereas prior works struggle to generalize." + ], + "image_footnote": [], + "bbox": [ + 91, + 345, + 480, + 435 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "randomly-ordered sequence of augmentations to the images. Specifically, our modified RandAugment samples a random number $n$ between 0 and $n_{\\mathrm{max}}$ for each augmentation type. Then, it applies the augmentations in random order until $n$ augmentations are applied for each augmentation type to the image. We use various augmentations, including in-memory JPEG compression, random resizing with random interpolation methods, cropping, flipping, rotation, translation, shear, padding, and cutout.", + "bbox": [ + 89, + 554, + 483, + 690 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Generalization across benchmarks", + "text_level": 1, + "bbox": [ + 89, + 699, + 392, + 713 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We first evaluate how well classifiers trained on our dataset perform across benchmarks. In Table 2, we observe that our models outperform the prior works [43, 87, 125, 133] on all evaluation sets except GenImage. This is expected since the GenImage evaluation set is a validation split of their training set; all of the generators are already seen by their classifier. On all other unseen benchmarks, our classifiers outperform all prior works. Notably, our classifiers achieve very high performance (0.987 mAP and $89.2\\%$ accuracy) on our out-of-distribution, comprehensive evaluation set, with a significant margin over prior works. This gap in performance can be traced to our training data which in", + "bbox": [ + 89, + 719, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b2050c6c5265689e96c9dc3268ab08e1cb231abb9c6ddef646fe10e377ee3cc7.jpg", + "image_caption": [ + "Figure 4. Performance with increasing number of images. We train a classifier with varying numbers of images from two sets: 1000 randomly chosen models and 10 popular (highly downloaded) models in the systematically collected subset. We report the mean and standard error bands for each data point across 4 randomly sampled subsets. The classifier trained from 1000 random models outperforms 10 popular models in all cases. Notably, the accuracy gap is wider than that of mAP, which may suggest that having a diverse set of models improves accuracy threshold calibration." + ], + "image_footnote": [], + "bbox": [ + 516, + 345, + 709, + 457 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/71f687f50ab6ed2107f86a86ca9738dc0eeb779d3afd5d7e956d3ee90cb40b2a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 710, + 345, + 901, + 458 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "corporates a substantially richer variety of generators than existing works. Consequently, our classifiers demonstrate robust generalization to out-of-distribution data, where prior works often struggle.", + "bbox": [ + 511, + 602, + 905, + 662 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To better illustrate the generalization of the classifiers, in Figure 3, we split our comprehensive evaluation set into five subsets and evaluate them: latent diffusion, pixel diffusion, commercial models, GANs, and other architecture type (Stable Cascade [93]). Furthermore, we evaluate three additional forensics methods: Aerblade [99], LGrad [116], and NPR [117]. Our classifiers show strong performance across all generator types, unlike prior works which struggle to generalize. For the following experiments, we use our best-performing model (High res.) unless stated otherwise.", + "bbox": [ + 511, + 662, + 906, + 814 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Impact of model diversity", + "text_level": 1, + "bbox": [ + 511, + 820, + 746, + 835 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Next, we examine the impact of the number of models in training data. We train classifiers with images sampled from 3 to 3333 generators and evaluate them (Fig. 1). To ensure that the gains are not due to simply sampling qualita", + "bbox": [ + 511, + 839, + 906, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "8250", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/de3a685a1acf23bf0081bb9d302fdcdecba91a2e02a740ea8299ba0c369f2745.jpg", + "image_caption": [ + "(a) Impact of generator type diversity" + ], + "image_footnote": [], + "bbox": [ + 94, + 87, + 285, + 189 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/86a32a5591270b141562b2d967f9c61148be4d1ec7231ff9414b918c18f904f7.jpg", + "image_caption": [ + "(b) Classifier backbone comparison", + "Figure 5. (a) Performance and model diversity. We compare detection performance for commercial models using classifiers trained on different subsets of the dataset: the systematically collected latent diffusion models, the manually chosen models containing diverse generator types, and both. As diversity increases, so does performance. (b) Classifier backbone comparison. We compare the architectures across datasets: ours, GenImage [133], and Wang et al. [125]. Performance is similar between architectures." + ], + "image_footnote": [], + "bbox": [ + 290, + 87, + 480, + 190 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "tively different architectures, we only use our systematically collected latent diffusion models. For each data point, we sample 10 random subsets of models with 100K training images each and report the mean and standard error bands. We use an extended evaluation set that includes non-latent diffusion generators from our training set, which allows us to comprehensively assess the generalization capability of the classifiers trained exclusively on latent diffusion models. We find that the performance steadily increases with the number of models. However, the performance begins to flatten out beyond 1000 models, suggesting diminishing returns. Interestingly, the performance also improves on out-of-distribution architectures such as GANs and pixel-based diffusion models, even though the classifier is only trained on latent diffusion models.", + "bbox": [ + 88, + 340, + 482, + 565 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In Figure 4, we vary the number of images from two sets: 1000 randomly chosen models and 10 popular models (as denoted by their number of downloads) downloaded from our systematically collected diffusion models. While the results show that the performance improves with more training images, it begins to plateau at approximately $27\\mathrm{K}$ images. Moreover, the classifier trained on 1000 models outperforms the 10 models in all cases, indicating that model diversity is important for strong performance. We also note that the accuracy gap is noticeably wider than that of mAP, which may suggest that model diversity is crucial in calibrating the accuracy thresholds of the classifiers.", + "bbox": [ + 89, + 566, + 482, + 748 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our experiments show that the performance improvements from increasing the number of models may plateau when they are limited to a single generator type (Fig. 1). In Figure 5a, we show that the diversity of the generator type also plays a major role in generalization. We train classifiers on three different sets of training data: our systematically collected set, manually chosen set, and a full set consisting of both subsets. The systematic set comprises entirely of latent diffusion models, and the manual set contains numerous generator types, including GANs, latent and pixel-based dif", + "bbox": [ + 89, + 750, + 485, + 901 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0af5c5242036ada06279f525959cfa606d6065e317152240bfc22a526c1dbddc.jpg", + "image_caption": [ + "(a) Impact of frozen backbones", + "Figure 6. (a) Evaluating frozen backbones. Freezing the pretrained backbone, a common practice in prior works [15, 87], consistently decreases the performance. (b) Analyzing source and generated data alignment. We evaluate how the pairing of the real datasets affects performance. 'R' denotes the real dataset used in training, and 'F' indicates the source dataset used to obtain the captions for prompting the generators. The results suggest that pairing the source data (i.e., real data used to prompt the generators) with the generated images is not essential for performance." + ], + "image_footnote": [], + "bbox": [ + 517, + 87, + 705, + 190 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c408e5535218ed15163009d68d13d62ad24b504eb97ca65653f5bf9f8f90ab5b.jpg", + "image_caption": [ + "(b) Semantic alignment analysis" + ], + "image_footnote": [], + "bbox": [ + 714, + 88, + 903, + 190 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "fusion, and autoregressive models (Sec. 3.2). The classifier trained on the manual set with more diverse generator types shows similar performance compared to the one trained on the systematic set, despite containing fewer than half the number of images (774K vs. 1.9M). Additionally, we find that the two sets are complementary; the performance is further improved when we train using both sets.", + "bbox": [ + 511, + 352, + 906, + 459 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Analysis of design choices", + "text_level": 1, + "bbox": [ + 511, + 465, + 746, + 481 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We examine the impact of various design choices, including some suggested in earlier works. In particular, we investigate the choice of backbone models, freezing the backbone, semantic alignment between the real and generated data, and robustness to transformations.", + "bbox": [ + 511, + 484, + 908, + 560 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Classifier backbone comparison. We compare the performance of the classifier trained using CLIP-ViT [27, 54, 96] and ConvNeXt [77] backbones following our training procedure in Figure 5b. We examine three datasets: ours, GenImage [133], and Wang et al. [125]. We observe similar performance between architectures across all datasets.", + "bbox": [ + 511, + 568, + 908, + 659 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Frozen backbone. Prior works [15, 87] suggested using a frozen CLIP-ViT backbone for training the classifiers. We study this by training the classifiers with both frozen and unfrozen pretrained backbones, using CLIP-ViT [27, 54, 96] and ConvNeXt [77]. As shown in Fig. 6a, freezing the backbone consistently leads to worse results, indicating that end-to-end training is crucial for high performance.", + "bbox": [ + 511, + 666, + 908, + 772 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Semantic alignment. Existing works often pair the generated images with the source dataset (i.e., the real dataset used to prompt or generate the images) arguing that misaligned data can introduce bias [7, 15, 87, 125]. We test this practice in Fig. 6b by examining the performance with both semantically aligned and misaligned real datasets. Specifically, we consider two real datasets: one comprised exclusively from LAION [107] and another combining ImageNet [23],", + "bbox": [ + 511, + 779, + 910, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "8251", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8a307ef1c5bdc4bcf944ceeb4dc3742fe7c1ec364fe5b71bb387b42936583893.jpg", + "image_caption": [ + "Figure 7. Robustness to various transformations. Our classifiers display robust performance across transformations. Other works generally show more sensitivity to these factors." + ], + "image_footnote": [], + "bbox": [ + 91, + 89, + 480, + 297 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "MS-COCO [74], LandscapesHQ [112], Forchheim [44], VISION [111], and IMD2020 [86]. We sample our systematically collected latent diffusion models using these two sets and categorize the generated images by their source dataset. The performance differences of all pairs were marginal, suggesting that strict alignment may not be as critical with sufficiently large data.", + "bbox": [ + 89, + 368, + 483, + 474 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Robustness to transformations. Figure 7 illustrates the robustness of the classifiers against various transformations. Following prior works [87, 125], we test robustness to JPEG compression and Gaussian blur. Additionally, we examine robustness to rotation, resizing, padding, and shear, as they commonly occur in real-world scenarios. For padding, we randomly pad the width or height of the image with a given fraction and scale it back to the original size. Similarly, in resize, we randomly upsample or downsample the height and width of the image by a given fraction and resize it back to the original size (e.g., if fraction is 0.3, resize the height and width to $0.7 \\times$ or $1.3 \\times$ and then scale it back to the original size). The results demonstrate that our models are more robust to transformations than existing models. Specifically, GenImage [133] is notably more sensitive to Gaussian blur, JPEG compression, and resizing artifacts; classifier by Ojha et al. [87] displays sensitivity to shear transforms, and the one by Wang et al. [125] performs poorly overall.", + "bbox": [ + 89, + 507, + 483, + 780 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Other forensics applications", + "text_level": 1, + "bbox": [ + 89, + 800, + 341, + 816 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our dataset may enable further forensics studies that can take advantage of our diverse array of generators. To illustrate this, in the supplement, we identify the type of generator used to synthesize a given image by using $k$ -nearest neighbor in the feature space of our classifier.", + "bbox": [ + 89, + 824, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 89, + 633, + 104 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we studied the problem of generalizing to unseen generative models in synthetic image detection. We proposed a new dataset, Community Forensics, which contains 4803 models and 2.7M images collected from various public sources. We studied the impact of model diversity and demonstrated that it plays a crucial role in enhancing data diversity and generalization performance. We trained classifiers on our dataset, studied their ability to generalize in various settings, and evaluated previously proposed models and training practices.", + "bbox": [ + 511, + 112, + 906, + 263 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We do not intend for our dataset to be used to train classifiers that are directly used in the wild. Detecting in-the-wild synthetic images remains a challenging open problem, and detection errors can have severe consequences (e.g., falsely accusing an author of creating fake images or allowing misinformation to be certified as real). We hope that our work will serve as a stepping stone for future research in this area by providing tools and insights for studying generalization and data collection strategies.", + "bbox": [ + 511, + 265, + 908, + 402 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations. While our dataset is diverse, a large portion of the data is diffusion-based, especially models based on Stable Diffusion [100]. Despite this, Figure 1 demonstrates that adding more diffusion models to the training data improves the generalization capability of the classifiers across other generator types. Furthermore, Figure 5a illustrates improved classifier performance when combining the diffusion-only systematic set with a more diverse manual set. These results suggest that despite their similarities, each diffusion model may uniquely contribute through its semantic content, image processing pipelines, or architectural variations (Sec. 3.1). Future work may explore collecting more diverse models, including GANs, VQ-VAEs [122], and autoregressive models. We also note that the generative models sourced from the community may contain inappropriate content. While in many contexts it is important to detect such images, these generated images may require further scrutiny before being used in other downstream applications. Finally, although our experiments suggest that our forensics classifiers generalize to unseen models better than those of previous work, their error rates are still too high for many critical applications.", + "bbox": [ + 511, + 417, + 908, + 734 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements. We thank the creators of the many open source models that we used to collect the Community Forensics dataset. We thank Chenhao Zheng, Cameron Johnson, Matthias Kirchner, Daniel Geng, Ziyang Chen, Ayush Shrivastava, Yiming Dou, Chao Feng, Xuanchen Lu, Zihao Wei, Zixuan Pan, Inbum Park, Rohit Banerjee, and Ang Cao for the valuable discussions and feedback. This research was developed with funding from the Defense Advanced Research Projects Agency (DARPA) under Contract No. HR001120C0123.", + "bbox": [ + 511, + 750, + 910, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8252", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Adobe. Firefly. https://www.adobe.com/products/firefly, 2023.4, 5", + "[2] Ideogram AI. Ideogram. https://ideogram.ai, 2024.4", + "[3] Rushil Anirudh and Jayaraman J Thiagarajan. Out of distribution detection via neural network anchoring. In Asian Conference on Machine Learning, pages 32-47. PMLR, 2023. 3", + "[4] Ivan Anokhin, Kirill Demochkin, Taras Khakhulin, Gleb Sterkin, Victor Lempitsky, and Denis Korzhenkov. Image generators with conditionally-independent pixel synthesis. arXiv preprint arXiv:2011.13775, 2020. 4", + "[5] Vishal Asnani, Xi Yin, Tal Hassner, and Xiaoming Liu. Reverse engineering of generative models: Inferring model hyperparameters from generated images. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(12):15477-15493, 2023. 2, 3, 4, 5", + "[6] Anton Bakhtin, Sam Gross, Myle Ott, Yuntian Deng, Marc'Aurelio Ranzato, and Arthur Szlam. Real or fake? learning to discriminate machine from human generated text. arXiv preprint arXiv:1906.03351, 2019. 3", + "[7] Quentin Bammey. Synthbuster: Towards detection of diffusion model generated images. IEEE Open Journal of Signal Processing, 2023. 2, 3, 4, 5, 6, 7", + "[8] Andrew Brock, Jeff Donahue, and Karen Simonyan. Large scale GAN training for high fidelity natural image synthesis. In International Conference on Learning Representations, 2019. 3, 4, 5", + "[9] Chen Chen, Qifeng Chen, Jia Xu, and Vladlen Koltun. Learning to see in the dark. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3291-3300, 2018. 3, 5", + "[10] Qifeng Chen and Vladlen Koltun. Photographic image synthesis with cascaded refinement networks. In Proceedings of the IEEE international conference on computer vision, pages 1511-1520, 2017. 3, 5", + "[11] Mehdi Cherti, Romain Beaumont, Ross Wightman, Mitchell Wortsman, Gabriel Ilharco, Cade Gordon, Christoph Schuhmann, Ludwig Schmidt, and Jenia Jitsev. Reproducible scaling laws for contrastive language-image learning. arXiv preprint arXiv:2212.07143, 2022.5", + "[12] Yunjey Choi, Minje Choi, Munyoung Kim, Jung-Woo Ha, Sunghun Kim, and Jaegul Choo. Stargan: Unified generative adversarial networks for multi-domain image-to-image translation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2018. 3, 5", + "[13] Yunjey Choi, Youngjung Uh, Jaejun Yoo, and Jung-Woo Ha. Stargan v2: Diverse image synthesis for multiple domains. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2020. 4", + "[14] Riccardo Corvi, Davide Cozzolino, Giada Zingarini, Giovanni Poggi, Koki Nagano, and Luisa Verdoliva. On the detection of synthetic images generated by diffusion models. In ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1-5. IEEE, 2023. 3" + ], + "bbox": [ + 99, + 114, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Davide Cozzolino, Giovanni Poggi, Riccardo Corvi, Matthias Nießner, and Luisa Verdoliva. Raising the bar of ai-generated image detection with clip. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4356-4366, 2024. 2, 3, 4, 5, 7", + "[16] Katherine Crowson, Stefan Andreas Baumann, Alex Birch, Tanishq Mathew Abraham, Daniel Z Kaplan, and Enrico Shippole. Scalable high-resolution pixel-space image synthesis with hourglass diffusion transformers. In Proceedings of the 41st International Conference on Machine Learning, pages 9550-9575. PMLR, 2024. 5", + "[17] Ekin Dogus Cubuk, Barret Zoph, Jon Shlens, and Quoc Le. Randaugment: Practical automated data augmentation with a reduced search space. In Advances in Neural Information Processing Systems, pages 18613-18624. Curran Associates, Inc., 2020. 5", + "[18] Tao Dai, Jianrui Cai, Yongbing Zhang, Shu-Tao Xia, and Lei Zhang. Second-order attention network for single image super-resolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11065-11074, 2019. 3, 5", + "[19] Duc-Tien Dang-Nguyen, Cecilia Pasquini, Valentina Conotter, and Giulia Boato. Raise: A raw images dataset for digital image forensics. In Proceedings of the 6th ACM multimedia systems conference, pages 219-224, 2015. 5", + "[20] Quan Dao, Hao Phung, Binh Nguyen, and Anh Tran. Flow matching in latent space. arXiv preprint arXiv:2307.08698, 2023.4", + "[21] Tiago José De Carvalho, Christian Riess, Eli Angelopoulou, Helio Pedrini, and Anderson de Rezende Rocha. Exposing digital image forgeries by illumination color classification. IEEE Transactions on Information Forensics and Security, 8(7):1182-1194, 2013. 3", + "[22] DeepFloyd. Deepfloyd. https://huggingface.co/DeepFloyd/IF-I-L-v1.0, 2024.4", + "[23] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 4, 5, 7", + "[24] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 4, 5", + "[25] Brian Dolhansky, Joanna Bitton, Ben Pflaum, Jikuo Lu, Russ Howes, Menglin Wang, and Cristian Canton Ferrer. The deepfake detection challenge (dfdc) dataset. arXiv preprint arXiv:2006.07397, 2020. 2, 5", + "[26] Jing Dong, Wei Wang, and Tieniu Tan. Casia image tampering detection evaluation database. In 2013 IEEE China summit and international conference on signal and information processing, pages 422-426. IEEE, 2013. 3", + "[27] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. ICLR, 2021. 5, 6, 7" + ], + "bbox": [ + 524, + 92, + 906, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "8253", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[28] Yilun Du, Shuang Li, Joshua Tenenbaum, and Igor Mordatch. Improved contrastive divergence training of energy-based models. In International Conference on Machine Learning, pages 2837-2848. PMLR, 2021. 3", + "[29] Ricard Durall, Margret Keuper, and Janis Keuper. Watch your up-convolution: Cnn based generative deep neural networks are failing to reproduce spectral distributions. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7890–7899, 2020. 3", + "[30] Tarik Dzanic, Karan Shah, and Freddie Witherden. Fourier spectrum discrepancies in deep network generated images. Advances in neural information processing systems, 33:3022-3032, 2020. 3", + "[31] Sven Elflein, Bertrand Charpentier, Daniel Zügner, and Stephan Gunnemann. On out-of-distribution detection with energy-based models, 2021. 3", + "[32] David C Epstein, Ishan Jain, Oliver Wang, and Richard Zhang. Online detection of ai-generated images. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 382-392, 2023. 2, 3, 4", + "[33] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12873-12883, 2021. 4", + "[34] Hugging Face. Hugging face diffusers library. https://huggingface.co/models?library=diffusers, accessed on June 05, 2022, 2022. 2, 3, 4", + "[35] Sebastian Gehrmann, Hendrik Strobelt, and Alexander M Rush. Gltr: Statistical detection and visualization of generated text. arXiv preprint arXiv:1906.04043, 2019. 3", + "[36] Thomas Gloe and Rainer Böhme. The'dresden image database' for benchmarking digital image forensics. In Proceedings of the 2010 ACM symposium on applied computing, pages 1584-1590, 2010. 4", + "[37] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in Neural Information Processing Systems, 27, 2014. 2", + "[38] Google. Imagen 3. https://deepmind.google/technologies/imagen-3, 2024.4", + "[39] Karol Gregor, Ivo Danihelka, Andriy Mnih, Charles Blundell, and Daan Wierstra. Deep autoregressive networks. In International Conference on Machine Learning, pages 1242-1250. PMLR, 2014. 2", + "[40] Patrick Grommelt, Louis Weiss, Franz-Josef Pfreundt, and Janis Keuper. Fake or.jpeg? revealing common biases in generated image detection datasets. arXiv preprint arXiv:2403.17608, 2024. 3", + "[41] Shuyang Gu, Dong Chen, Jianmin Bao, Fang Wen, Bo Zhang, Dongdong Chen, Lu Yuan, and Baining Guo. Vector quantized diffusion model for text-to-image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10696-10706, 2022. 4, 5", + "[42] Fabrizio Guillaro, Giada Zingarini, Ben Usman, Avneesh Sud, Davide Cozzolino, and Luisa Verdoliva. A bias-free training paradigm for more general ai-generated image detection, 2024. 3" + ], + "bbox": [ + 99, + 92, + 485, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[43] Xiao Guo, Vishal Asnani, Sijja Liu, and Xiaoming Liu. Tracing hyperparameter dependencies for model parsing via learnable graph pooling network. Advances in Neural Information Processing Systems, 37:116899-116932, 2024. 2, 3, 4, 5, 6", + "[44] Benjamin Hadwiger and Christian Riess. The forchheim image database for camera identification in the wild. In Pattern Recognition. ICPR International Workshops and Challenges: Virtual Event, January 10–15, 2021, Proceedings, Part VI, pages 500–515. Springer, 2021. 4, 8", + "[45] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 5, 6", + "[46] Dan Hendrycks and Kevin Gimpel. A baseline for detecting misclassified and out-of-distribution examples in neural networks. In International Conference on Learning Representations, 2016. 3", + "[47] Dan Hendrycks, Mantas Mazeika, Saurav Kadavath, and Dawn Song. Using self-supervised learning can improve model robustness and uncertainty. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2019. 3", + "[48] Dan Hendrycks, Steven Basart, Norman Mu, Saurav Kadavath, Frank Wang, Evan Dorundo, Rahul Desai, Tyler Zhu, Samyak Parajuli, Mike Guo, Dawn Song, Jacob Steinhardt, and Justin Gilmer. The many faces of robustness: A critical analysis of out-of-distribution generalization. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 8340-8349, 2021. 2, 3", + "[49] Yan Hong and Jianfu Zhang. Wildfake: A large-scale challenging dataset for ai-generated images detection. arXiv preprint arXiv:2402.11843, 2024. 3", + "[50] Shuwei Hou, Yan Ju, Chengzhe Sun, Shan Jia, Lipeng Ke, Riky Zhou, Anita Nikolich, and Siwei Lyu. Deepfake-meter v2.0: An open platform for deepfake detection. arXiv preprint arXiv:2404.13146, 2024. 3", + "[51] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 4", + "[52] Drew A Hudson and Larry Zitnick. Generative adversarial transformers. In International conference on machine learning, pages 4487-4499. PMLR, 2021. 4", + "[53] Minyoung Huh, Andrew Liu, Andrew Owens, and Alexei A Efros. Fighting fake news: Image splice detection via learned self-consistency. In Proceedings of the European conference on computer vision (ECCV), 2018. 3", + "[54] Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. Openclip, 2021. If you use this software, please cite it as below. 5, 6, 7", + "[55] Ganesh Jawahar, Muhammad Abdul-Mageed, and Laks VS Lakshmanan. Automatic detection of machine generated text: A critical survey. arXiv preprint arXiv:2011.01314, 2020. 3" + ], + "bbox": [ + 522, + 92, + 906, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "8254", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[56] Micah K Johnson and Hany Farid. Exposing digital forgeries in complex lighting environments. IEEE Transactions on Information Forensics and Security, 2(3):450-461, 2007. 3", + "[57] Minguk Kang, Jun-Yan Zhu, Richard Zhang, Jaesik Park, Eli Shechtman, Sylvain Paris, and Taesung Park. Scaling up gans for text-to-image synthesis. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 4, 5", + "[58] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of gans for improved quality, stability, and variation. In International Conference on Learning Representations, 2018. 3, 4", + "[59] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 3, 4, 5, 6", + "[60] Tero Karras, Miika Aittala, Janne Hellsten, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Training generative adversarial networks with limited data. Advances in neural information processing systems, 33:12104-12114, 2020. 4", + "[61] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8110-8119, 2020.", + "[62] Tero Karras, Miika Aittala, Samuli Laine, Erik Härkonen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Alias-free generative adversarial networks. Advances in neural information processing systems, 34:852-863, 2021. 4", + "[63] Hasam Khalid, Shahroz Tariq, Minha Kim, and Simon Woo. Fakeavceleb: A novel audio-video multimodal deepfake dataset. In Proceedings of the Neural Information Processing Systems Track on Datasets and Benchmarks, 2021. 2", + "[64] P. Korus and J. Huang. Multi-scale analysis strategies in prnu-based tampering localization. IEEE Trans. on Information Forensics & Security, 2017. 3", + "[65] Kalpesh Krishna, Yixiao Song, Marzena Karpinska, John Wieting, and Mohit Iyyer. Paraphrasing evades detectors of ai-generated text, but retrieval is an effective defense. Advances in Neural Information Processing Systems, 36, 2024. 3", + "[66] Kvikontent. Kvikontent-midjourney v6. https://huggingface.co/Kvikontent/midjourney-v6, 2023.5", + "[67] Patrick Kwon, Jaeseong You, Gyuhyeon Nam, Sungwoo Park, and Gyeongsu Chae. Kodf: A large-scale korean deepfake detection dataset. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021. 2", + "[68] Black Forst Labs. Flux. https://blackforestlabs.ai, 2024.4", + "[69] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, 2022. 5", + "[70] Ke Li, Tianhao Zhang, and Jitendra Malik. Diverse image synthesis from semantic layouts via conditional imle. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4220-4229, 2019. 3, 5" + ], + "bbox": [ + 99, + 90, + 485, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[71] Yuezun Li, Xin Yang, Pu Sun, Honggang Qi, and Siwei Lyu. Celeb-df: A large-scale challenging dataset for deepfake forensics. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2020. 2", + "[72] Yewen Li, Chaojie Wang, Xiaobo Xia, Tongliang Liu, Bo An, et al. Out-of-distribution detection with an adaptive likelihood ratio on informative hierarchical vae. Advances in Neural Information Processing Systems, 35:7383-7396, 2022. 3", + "[73] Shiyu Liang, Yixuan Li, and R Srikant. Enhancing the reliability of out-of-distribution image detection in neural networks. In International Conference on Learning Representations, 2018. 3", + "[74] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 4, 5, 6, 8", + "[75] Weitang Liu, Xiaoyun Wang, John Owens, and Yixuan Li. Energy-based out-of-distribution detection. In Advances in Neural Information Processing Systems, pages 21464-21475. Curran Associates, Inc., 2020. 3", + "[76] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaou Tang. Deep learning face attributes in the wild. In Proceedings of International Conference on Computer Vision (ICCV), 2015. 4", + "[77] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11976-11986, 2022. 5, 7", + "[78] Simian Luo, Yiqin Tan, Longbo Huang, Jian Li, and Hang Zhao. Latent consistency models: Synthesizing high-resolution images with few-step inference. arXiv preprint arXiv:2310.04378, 2023. 4", + "[79] Simian Luo, Yiqin Tan, Suraj Patil, Daniel Gu, Patrick von Platen, Apolinário Passos, Longbo Huang, Jian Li, and Hang Zhao. Lcm-lora: A universal stable-diffusion acceleration module. arXiv preprint arXiv:2311.05556, 2023. 2, 5", + "[80] Francesco Marra, Diego Gragnaniello, Luisa Verdoliva, and Giovanni Poggi. Do gans leave artificial fingerprints? In 2019 IEEE conference on multimedia information processing and retrieval (MIPR), pages 506-511. IEEE, 2019. 3", + "[81] Inc. Midjourney. Midjourney. https://www.midjourney.com/home, 2022. 4, 5", + "[82] Eric Mitchell, Yoonho Lee, Alexander Khazatsky, Christopher D Manning, and Chelsea Finn. Detectgpt: Zero-shot machine-generated text detection using probability curvature. In International Conference on Machine Learning, pages 24950-24962. PMLR, 2023. 3", + "[83] Sina Mohseni, Mandar Pitale, JBS Yadawa, and Zhangyang Wang. Self-supervised learning for generalizable out-of-distribution detection. Proceedings of the AAAI Conference on Artificial Intelligence, 34(04):5216-5223, 2020. 3", + "[84] Tian-Tsong Ng, Shih-Fu Chang, and Q Sun. A data set of authentic and spliced image blocks. Columbia University, ADVENT Technical Report, 4, 2004. 3" + ], + "bbox": [ + 522, + 92, + 906, + 900 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "8255", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[85] Alexander Quinn Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. In International Conference on Machine Learning, pages 16784-16804. PMLR, 2022. 4, 5", + "[86] Adam Novozamsky, Babak Mahdian, and Stanislav Saic. Imd2020: A large-scale annotated dataset tailored for detecting manipulated images. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision Workshops, pages 71-80, 2020. 4, 8", + "[87] Utkarsh Ojha, Yuheng Li, and Yong Jae Lee. Towards universal fake image detectors that generalize across generative models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 24480-24489, 2023. 1, 2, 3, 4, 5, 6, 7, 8", + "[88] OpenAI. Dall-e 2. https://openai.com/index/dall-e-2, 2022. 4, 5", + "[89] OpenAI. Dall-e 3. https://openai.com/index/dall-e-3, 2023.4, 5", + "[90] Taesung Park, Ming-Yu Liu, Ting-Chun Wang, and JunYan Zhu. Semantic image synthesis with spatially-adaptive normalization. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2019. 3", + "[91] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019. 5", + "[92] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4195-4205, 2023. 4", + "[93] Pablo Pernias, Dominic Rampas, Mats Leon Richter, Christopher Pal, and Marc Aubreville. Würstchen: An efficient architecture for large-scale text-to-image diffusion models. In The Twelfth International Conference on Learning Representations, 2023. 5, 6", + "[94] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. In The Twelfth International Conference on Learning Representations, 2023. 5", + "[95] Alin C Popescu and Hany Farid. Exposing digital forgeries by detecting traces of resampling. IEEE Transactions on signal processing, 2005. 3", + "[96] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 3, 5, 6, 7", + "[97] Anirudh Sundara Rajan, Utkarsh Ojha, Jedidiah Schoesser, and Yong Jae Lee. Aligned datasets improve detection of latent diffusion-generated images, 2025. 3" + ], + "bbox": [ + 99, + 92, + 485, + 900 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[98] Jie Ren, Peter J Liu, Emily Fertig, Jasper Snoek, Ryan Poplin, Mark Depristo, Joshua Dillon, and Balaji Lakshminarayanan. Likelihood ratios for out-of-distribution detection. Advances in neural information processing systems, 32, 2019. 3", + "[99] Jonas Ricker, Denis Lukovnikov, and Asja Fischer. Aeroblade: Training-free detection of latent diffusion images using autoencoder reconstruction error. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9130-9140, 2024. 6", + "[100] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, 2022. 2, 4, 5, 8", + "[101] Andreas Rössler, Davide Cozzolino, Luisa Verdoliva, Christian Riess, Justus Thies, and Matthias Nießner. Faceforensics: A large-scale video dataset for forgery detection in human faces. arXiv preprint arXiv:1803.09179, 2018. 1, 2", + "[102] Andreas Rossler, Davide Cozzolino, Luisa Verdoliva, Christian Riess, Justus Thies, and Matthias Nießner. Faceforensics++: Learning to detect manipulated facial images. In Proceedings of the IEEE/CVF international conference on computer vision, 2019. 1, 2, 3", + "[103] Vinu Sankar Sadasivan, Aounon Kumar, Sriram Balasubramanian, Wenxiao Wang, and Soheil Feizi. Can ai-generated text be reliably detected? arXiv preprint arXiv:2303.11156, 2023. 3", + "[104] Chandramouli Shama Sastry and Sageev Oore. Detecting out-of-distribution examples with Gram matrices. In Proceedings of the 37th International Conference on Machine Learning, pages 8491-8501. PMLR, 2020. 3", + "[105] Axel Sauer, Kashyap Chitta, Jens Müller, and Andreas Geiger. Projected gans converge faster. In Advances in Neural Information Processing Systems (NeurIPS), 2021. 4", + "[106] Axel Sauer, Katja Schwarz, and Andreas Geiger. Styleganxl: Scaling stylegan to large diverse datasets. In ACM SIGGRAPH 2022 conference proceedings, pages 1-10, 2022. 4", + "[107] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021. 2, 4, 5, 7", + "[108] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade W Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortman, Patrick Schramowski, Srivatsa R Kundurthy, Katherine Crowson, Ludwig Schmidt, Robert Kaczmarczyk, and Jenia Jitsev. LAION-5b: An open large-scale dataset for training next generation image-text models. In Thirty-sixth Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2022. 2, 3, 4, 5", + "[109] Vikash Sehwag, Mung Chiang, and Prateek Mittal. Ssd: A unified framework for self-supervised outlier detection. In International Conference on Learning Representations, 2020. 3" + ], + "bbox": [ + 516, + 92, + 906, + 898 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "8256", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[110] Arseniy Shakhmatov, Anton Razzhigaev, Aleksandr Nikolich, Vladimir Arkhipkin, Igor Pavlov, Andrey Kuznetsov, and Denis Dimitrov. Kandinsky 2.2. https://github.com/ai-forever/Kandinsky-2, 2023.5", + "[111] Dasara Shullani, Marco Fontani, Massimo Iuliani, Omar Al Shaya, and Alessandro Piva. Vision: a video and image dataset for source identification. EURASIP Journal on Information Security, 2017:1-16, 2017. 4, 8", + "[112] Ivan Skorokhodov, Grigorii Sotnikov, and Mohamed Elhoseiny. Aligning latent and image spaces to connect the unconnectable. arXiv preprint arXiv:2104.06954, 2021. 4, 8", + "[113] Irene Solaiman, Miles Brundage, Jack Clark, Amanda Askell, Ariel Herbert-Voss, Jeff Wu, Alec Radford, Gretchen Krueger, Jong Wook Kim, Sarah Kreps, et al. Release strategies and the social impacts of language models. arXiv preprint arXiv:1908.09203, 2019. 3", + "[114] Yang Song, Prafulla Dhariwal, Mark Chen, and Ilya Sutskever. Consistency models. In Proceedings of the 40th International Conference on Machine Learning, pages 32211-32252, 2023. 2", + "[115] Yuhta Takida, Masaaki Imaizumi, Takashi Shibuya, Chieh-Hsin Lai, Toshimitsu Uesaka, Naoki Murata, and Yuki Mitsufuji. SAN: Inducing metrizability of GAN with discriminative normalized linear layer. In The Twelfth International Conference on Learning Representations, 2024. 4", + "[116] Chuangchuang Tan, Yao Zhao, Shikui Wei, Guanghua Gu, and Yunchao Wei. Learning on gradients: Generalized artifacts representation for gan-generated images detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12105-12114, 2023. 6", + "[117] Chuangchuang Tan, Yao Zhao, Shikui Wei, Guanghua Gu, Ping Liu, and Yunchao Wei. Rethinking the up-sampling operations in cnn-based generative network for generalizable deepfake detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 28130-28139, 2024. 6", + "[118] Ming Tao, Hao Tang, Fei Wu, Xiao-Yuan Jing, Bing-Kun Bao, and Changsheng Xu. Df-gan: A simple and effective baseline for text-to-image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16515-16525, 2022. 5", + "[119] Ming Tao, Bing-Kun Bao, Hao Tang, and Changsheng Xu. Galip: Generative adversarial clips for text-to-image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14214-14223, 2023. 5", + "[120] DeciAI Research Team. Decidiffusion 2.0, 2024. 5", + "[121] Adaku Uchendu, Thai Le, Kai Shu, and Dongwon Lee. Authorship attribution for neural text generation. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 8384–8395, 2020. 3", + "[122] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017. 8" + ], + "bbox": [ + 91, + 90, + 485, + 900 + ], + "page_idx": 12 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[123] Patrick von Platen, Suraj Patil, Anton Lozhkov, Pedro Cuenca, Nathan Lambert, Kashif Rasul, Mishig Davaadorj, Dhruv Nair, Sayak Paul, William Berman, Yiyi Xu, Steven Liu, and Thomas Wolf. Diffusers: State-of-the-art diffusion models. https://github.com/huggingface/ diffusers, 2022. 2, 3, 4", + "[124] Apoorv Vyas, Nataraj Jammalamadaka, Xia Zhu, Dipankar Das, Bharat Kaul, and Theodore L. Willke. Out-of-distribution detection using an ensemble of self-supervised leave-out classifiers. In Proceedings of the European Conference on Computer Vision (ECCV), 2018. 3", + "[125] Sheng-Yu Wang, Oliver Wang, Richard Zhang, Andrew Owens, and Alexei A. Efros. Cnn-generated images are surprisingly easy to spot... for now. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 1, 2, 3, 4, 5, 6, 7, 8", + "[126] Ross Wightman. Pytorch image models. https://github.com/huggingface/pytorch-image-models, 2019.5", + "[127] Wukong. https://xihe.mindspore.cn/modelzoo/wukong, 2022.5", + "[128] Zhisheng Xiao, Qing Yan, and Yali Amit. Likelihood regret: An out-of-distribution detection score for variational autoencoder. In Advances in Neural Information Processing Systems, pages 20685-20696. Curran Associates, Inc., 2020. 3", + "[129] Fisher Yu, Yinda Zhang, Shuran Song, Ari Seff, and Jianxiong Xiao. Lsun: Construction of a large-scale image dataset using deep learning with humans in the loop. arXiv preprint arXiv:1506.03365, 2015. 4", + "[130] Bowen Zhang, Shuyang Gu, Bo Zhang, Jianmin Bao, Dong Chen, Fang Wen, Yong Wang, and Baining Guo. Styleswin: Transformer-based gan for high-resolution image generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11304-11314, 2022. 4", + "[131] Xu Zhang, Svebor Karaman, and Shih-Fu Chang. Detecting and simulating artifacts in gan fake images. In 2019 IEEE international workshop on information forensics and security (WIFS), pages 1-6. IEEE, 2019. 3", + "[132] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In Computer Vision (ICCV), 2017 IEEE International Conference on, 2017. 3, 5", + "[133] Mingjian Zhu, Hanting Chen, Qiangyu YAN, Xudong Huang, Guanyu Lin, Wei Li, Zhijun Tu, Hailin Hu, Jie Hu, and Yunhe Wang. Genimage: A million-scale benchmark for detecting ai-generated image. In Advances in Neural Information Processing Systems Dataset and Benchmarks Track, pages 77771-77782, 2023. 2, 3, 4, 5, 6, 7, 8", + "[134] Bojia Zi, Minghao Chang, Jingjing Chen, Xingjun Ma, and Yu-Gang Jiang. Wilddeepfake: A challenging real-world dataset for deepfake detection. In Proceedings of the 28th ACM international conference on multimedia, 2020. 2" + ], + "bbox": [ + 516, + 90, + 906, + 825 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "8257", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 12 + } +] \ No newline at end of file diff --git a/2025/Community Forensics_ Using Thousands of Generators to Train Fake Image Detectors/f9d7ae16-50eb-4d82-828a-d6dd35d593a9_model.json b/2025/Community Forensics_ Using Thousands of Generators to Train Fake Image Detectors/f9d7ae16-50eb-4d82-828a-d6dd35d593a9_model.json new file mode 100644 index 0000000000000000000000000000000000000000..f5c023e679a9a72ba84992d947d38ccba1a895dc --- /dev/null +++ b/2025/Community Forensics_ Using Thousands of Generators to Train Fake Image Detectors/f9d7ae16-50eb-4d82-828a-d6dd35d593a9_model.json @@ -0,0 +1,3669 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.178, + 0.131, + 0.82, + 0.176 + ], + "angle": 0, + "content": "Community Forensics: Using Thousands of Generators to Train Fake Image Detectors" + }, + { + "type": "text", + "bbox": [ + 0.337, + 0.204, + 0.66, + 0.24 + ], + "angle": 0, + "content": "Jeongsoo Park Andrew Owens University of Michigan" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.274, + 0.327, + 0.29 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.307, + 0.486, + 0.624 + ], + "angle": 0, + "content": "One of the key challenges of detecting AI-generated images is spotting images that have been created by previously unseen generative models. We argue that the limited diversity of the training data is a major obstacle to addressing this problem, and we propose a new dataset that is significantly larger and more diverse than prior works. As part of creating this dataset, we systematically download thousands of text-to-image latent diffusion models and sample images from them. We also collect images from dozens of popular open source and commercial models. The resulting dataset contains 2.7M images that have been sampled from 4803 different models. These images collectively capture a wide range of scene content, generator architectures, and image processing settings. Using this dataset, we study the generalization abilities of fake image detectors. Our experiments suggest that detection performance improves as the number of models in the training set increases, even when these models have similar architectures. We also find that increasing the diversity of the models improves detection performance, and that our trained detectors generalize better than those trained on other datasets. The dataset can be found in" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.626, + 0.456, + 0.639 + ], + "angle": 0, + "content": "https://jespark.net/projects/2024/community_forensics" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.667, + 0.222, + 0.683 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.483, + 0.839 + ], + "angle": 0, + "content": "Our ability to automatically generate realistic images is quickly outpacing our ability to detect them, potentially leading to a state of affairs in which neither humans nor machines can reliably tell real from fake. While the field of image forensics has been developing methods to address this problem, existing fake image detectors still struggle with generalization. These methods often excel at detecting images from generators that were present in their training sets, but fail when given images sampled from unseen models [87, 101, 102, 125]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "A core challenge is dealing with the large amounts of variation between models. Each generator has a potentially unique combination of the architecture, loss function, and training distribution. Even seemingly minor differences in" + }, + { + "type": "image", + "bbox": [ + 0.527, + 0.277, + 0.891, + 0.482 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.492, + 0.908, + 0.617 + ], + "angle": 0, + "content": "Figure 1. Performance vs. model diversity. We use images sampled from different numbers of open source latent diffusion models in our Community Forensics dataset to train fake image detectors (shown in Fig. 2a). We fix the total number of images and only vary the number of models. The detector's performance increases across all generator types as we train from more models, even though these added models are entirely latent diffusion. This improvement is largest for test images from out-of-distribution generators, such as pixel-based diffusion models or GANs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.629, + 0.909, + 0.765 + ], + "angle": 0, + "content": "low-level image processing details, such as the ways that training images are resized or compressed, can strongly influence detection accuracy [125]. As a result of these model-specific idiosyncrasies, a generator's images may evade detection, even when images from architecturally similar models exist in the training set. This issue has been exacerbated by the thousands of open source models that are now available online, many of which extend pretrained base models in complex ways." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.909, + 0.901 + ], + "angle": 0, + "content": "We hypothesize that the lack of diversity in training datasets is a major source of these shortcomings. Although today's datasets often contain millions of fake images, they come from a relatively small number of generators. As a result, this data fails to capture many sources of variation that one might encounter in the wild. These limitations also make it challenging to accurately benchmark performance, since it is easy for cues that work well on one set of generators to fail on others." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8245" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.1, + 0.091, + 0.166, + 0.153 + ], + "angle": 0, + "content": "Dreamshaper v6" + }, + { + "type": "image", + "bbox": [ + 0.182, + 0.091, + 0.251, + 0.153 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.263, + 0.091, + 0.328, + 0.153 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.342, + 0.091, + 0.413, + 0.153 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.422, + 0.091, + 0.495, + 0.153 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.536, + 0.091, + 0.545, + 0.098 + ], + "angle": 0, + "content": "91" + }, + { + "type": "image", + "bbox": [ + 0.423, + 0.098, + 0.495, + 0.153 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.443, + 0.154, + 0.474, + 0.161 + ], + "angle": 0, + "content": "SoteMix" + }, + { + "type": "image", + "bbox": [ + 0.423, + 0.161, + 0.495, + 0.215 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.437, + 0.216, + 0.482, + 0.222 + ], + "angle": 0, + "content": "Technoalbum" + }, + { + "type": "image", + "bbox": [ + 0.423, + 0.222, + 0.495, + 0.277 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.343, + 0.154, + 0.408, + 0.161 + ], + "angle": 0, + "content": "Plat diffusion" + }, + { + "type": "image_caption", + "bbox": [ + 0.343, + 0.277, + 0.414, + 0.285 + ], + "angle": 0, + "content": "Nightvisionxl 0791" + }, + { + "type": "image", + "bbox": [ + 0.343, + 0.222, + 0.414, + 0.277 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.343, + 0.28, + 0.414, + 0.287 + ], + "angle": 0, + "content": "Nightvisionxl 0791" + }, + { + "type": "image_caption", + "bbox": [ + 0.343, + 0.277, + 0.414, + 0.285 + ], + "angle": 0, + "content": "Nightvisionxl 0791" + }, + { + "type": "image_caption", + "bbox": [ + 0.423, + 0.277, + 0.495, + 0.285 + ], + "angle": 0, + "content": "Albedobase x1 v21" + }, + { + "type": "image", + "bbox": [ + 0.343, + 0.287, + 0.414, + 0.338 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.352, + 0.339, + 0.404, + 0.346 + ], + "angle": 0, + "content": "Enasswhidaly" + }, + { + "type": "image", + "bbox": [ + 0.343, + 0.346, + 0.414, + 0.4 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.423, + 0.277, + 0.495, + 0.285 + ], + "angle": 0, + "content": "Albedobase x1 v21" + }, + { + "type": "image", + "bbox": [ + 0.423, + 0.287, + 0.495, + 0.338 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.44, + 0.339, + 0.478, + 0.346 + ], + "angle": 0, + "content": "Msart 541" + }, + { + "type": "image", + "bbox": [ + 0.423, + 0.347, + 0.495, + 0.4 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.091, + 0.575, + 0.153 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.154, + 0.567, + 0.161 + ], + "angle": 0, + "content": "First avarfar" + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.161, + 0.575, + 0.215 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.504, + 0.216, + 0.576, + 0.223 + ], + "angle": 0, + "content": "AnalogMadness real" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.223, + 0.575, + 0.277 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.505, + 0.277, + 0.575, + 0.285 + ], + "angle": 0, + "content": "DucHaitenAnimated" + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.287, + 0.575, + 0.338 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.521, + 0.339, + 0.559, + 0.346 + ], + "angle": 0, + "content": "A can thr" + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.347, + 0.575, + 0.4 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.612, + 0.109, + 0.649, + 0.116 + ], + "angle": 0, + "content": "DeepFloyd" + }, + { + "type": "image", + "bbox": [ + 0.597, + 0.116, + 0.667, + 0.168 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.609, + 0.17, + 0.651, + 0.176 + ], + "angle": 0, + "content": "StyleGANXL" + }, + { + "type": "image_caption", + "bbox": [ + 0.675, + 0.109, + 0.731, + 0.116 + ], + "angle": 0, + "content": "VQDiffusion" + }, + { + "type": "image", + "bbox": [ + 0.674, + 0.116, + 0.746, + 0.168 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.678, + 0.17, + 0.74, + 0.176 + ], + "angle": 0, + "content": "GuidedDiffusion" + }, + { + "type": "image", + "bbox": [ + 0.674, + 0.176, + 0.744, + 0.229 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.769, + 0.109, + 0.806, + 0.116 + ], + "angle": 0, + "content": "StyleGAN2" + }, + { + "type": "image", + "bbox": [ + 0.752, + 0.116, + 0.822, + 0.168 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.777, + 0.17, + 0.797, + 0.176 + ], + "angle": 0, + "content": "Glide" + }, + { + "type": "image", + "bbox": [ + 0.752, + 0.176, + 0.822, + 0.229 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.859, + 0.109, + 0.871, + 0.116 + ], + "angle": 0, + "content": "DiT" + }, + { + "type": "image", + "bbox": [ + 0.831, + 0.116, + 0.9, + 0.169 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.857, + 0.17, + 0.874, + 0.176 + ], + "angle": 0, + "content": "CIPS" + }, + { + "type": "image", + "bbox": [ + 0.831, + 0.176, + 0.9, + 0.229 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.606, + 0.233, + 0.889, + 0.245 + ], + "angle": 0, + "content": "(b) Manually chosen open source models (19 models)" + }, + { + "type": "image_caption", + "bbox": [ + 0.619, + 0.267, + 0.643, + 0.274 + ], + "angle": 0, + "content": "Dalle2" + }, + { + "type": "image", + "bbox": [ + 0.597, + 0.274, + 0.666, + 0.327 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.615, + 0.328, + 0.648, + 0.335 + ], + "angle": 0, + "content": "FLUX dev" + }, + { + "type": "image", + "bbox": [ + 0.597, + 0.335, + 0.666, + 0.388 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.677, + 0.268, + 0.734, + 0.275 + ], + "angle": 0, + "content": "MidjourneyV6:" + }, + { + "type": "image", + "bbox": [ + 0.675, + 0.275, + 0.744, + 0.327 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.68, + 0.328, + 0.737, + 0.336 + ], + "angle": 0, + "content": "Firefly Image2" + }, + { + "type": "image", + "bbox": [ + 0.673, + 0.336, + 0.744, + 0.388 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.757, + 0.268, + 0.816, + 0.275 + ], + "angle": 0, + "content": "MidjourneyV5 2" + }, + { + "type": "image", + "bbox": [ + 0.753, + 0.275, + 0.822, + 0.327 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.757, + 0.328, + 0.815, + 0.335 + ], + "angle": 0, + "content": "Firefly Image3" + }, + { + "type": "image", + "bbox": [ + 0.753, + 0.335, + 0.822, + 0.388 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.853, + 0.268, + 0.878, + 0.275 + ], + "angle": 0, + "content": "Dalle3" + }, + { + "type": "image", + "bbox": [ + 0.831, + 0.275, + 0.9, + 0.328 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.845, + 0.329, + 0.885, + 0.335 + ], + "angle": 0, + "content": "IdeogramV1" + }, + { + "type": "image", + "bbox": [ + 0.831, + 0.335, + 0.9, + 0.388 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.179, + 0.406, + 0.496, + 0.419 + ], + "angle": 0, + "content": "(a) Systematically collected diffusion models (4763 models)" + }, + { + "type": "image_caption", + "bbox": [ + 0.652, + 0.392, + 0.843, + 0.403 + ], + "angle": 0, + "content": "(c) Commercial models (11 models)" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.423, + 0.907, + 0.48 + ], + "angle": 0, + "content": "Figure 2. The Community Forensics dataset. Our dataset contains images sampled from three types of generative models. (a) We systematically download open-source latent diffusion models from a model-sharing community [34, 123]. (b) We select popular open source generators with a variety of architectures and training procedures. (c) We sample from both closed and open state-of-the-art commercial models. We present example images and their corresponding model names." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.49, + 0.484, + 0.595 + ], + "angle": 0, + "content": "To address these problems, we propose Community Forensics, a dataset that is significantly more diverse and comprehensive than those in prior works (Fig. 2). Our dataset contains images generated by: (a) thousands of systematically downloaded open-source latent diffusion models, (b) hand-selected open source models with various architectures, and (c) state-of-the-art commercial models." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.597, + 0.484, + 0.808 + ], + "angle": 0, + "content": "To acquire large numbers of models, we sample images from thousands of text-to-image diffusion models hosted on a popular model-sharing website, Hugging Face [34]. We exploit the fact that these models use a common programming library [123] and thus can be sampled in a standardized way. A large fraction of them are extensions of Stable Diffusion [100], but collectively capture a variety of common model variations, such as in the architecture, image processing, and image content. We also sample images from many other open source models, including GANs [37], autoregressive models [39], and consistency models [79, 114]. To ensure sufficient diversity in the image content of the generated images, we sample generators using captions sourced from a mix of several existing datasets whenever possible." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.484, + 0.87 + ], + "angle": 0, + "content": "Our dataset contains 4803 distinct models, roughly \\(34 \\times\\) more than the most extensive previous dataset that sample images from generative models [5, 7, 15, 32, 43, 87, 125, 133], and covers a variety of recent model designs (Fig. 2)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.901 + ], + "angle": 0, + "content": "We use this dataset to study generalization in the generated image detection problem. Our experiments confirm" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.49, + 0.907, + 0.566 + ], + "angle": 0, + "content": "that increasing the diversity of data improves generalization, consistent with previous findings in other areas of computer vision [48, 107, 108]. In contrast to other work, our dataset aims to increase diversity via number of generators, an underexplored axis of diversity. Through experiments, we find:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.568, + 0.907, + 0.613 + ], + "angle": 0, + "content": "- Classifiers trained on our dataset obtain strong performance, both on our newly proposed evaluations and on multiple previously-proposed benchmarks." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.614, + 0.907, + 0.687 + ], + "angle": 0, + "content": "- Adding more generative models improves generalization. Fig. 1 demonstrates the performance of fake image detection when trained on samples from varying numbers of diffusion models. Notably, the performance improves as more models are added, even across different architectures." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.689, + 0.907, + 0.733 + ], + "angle": 0, + "content": "- Including diverse generative model architectures significantly improves results, since classifiers do not fully generalize between generator architectures." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.734, + 0.907, + 0.794 + ], + "angle": 0, + "content": "- Standard classifiers perform well. In contrast to observations from recent work, we find that end-to-end training of classifiers based on CNNs or ViTs generalizes well, with qualitatively similar to that of other recognition problems." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.568, + 0.907, + 0.794 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.805, + 0.651, + 0.82 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Datasets for detecting generated images. A number of datasets have been proposed for specifically detecting \"deepfake\" images containing manipulated faces [25, 63, 67, 71, 101, 102, 134]. Rather than focusing on face manipulation, we address creating general-purpose meth" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8246" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.091, + 0.486, + 0.637 + ], + "angle": 0, + "content": "ods that can detect images that have been directly produced by generative models. Wang et al. [125] proposed a widely-used dataset of CNN-generated images, mixing images from GANs [8, 12, 58, 59, 90, 132] with other models [9, 10, 18, 70, 102]. This work showed that forensics models generalize between generative models, providing motivation for training on large datasets of diverse generators. However, their classifier was trained on images from a single GAN and was highly sensitive to data augmentation parameters, and more recent work shows that it does not generalize to newer models [14, 87]. Ojha et al. [87] introduced a dataset of recent diffusion models and found that training a linear classifier on CLIP features [96] extracted from ProGAN-generated images performed well. Cozzolino et al. [15] extend this work by studying the performance of CLIP-based detectors on various generative models and datasets. Epstein et al. [32] simulated detecting fake images in an online way by training a detector up to a certain year and testing it on generators released after that year. Zhu et al. [133] collected 1.4M generated images from 8 different generators. These datasets, however, only consider a handful of models (less than 20 each), limiting the generalization of their detectors. Asnani et al. proposed RED116 [5], collecting 116K images from 116 generative models for predicting the model hyperparameters of a given generated image. Guo et al. [43] extends RED116 to 140 generative models with 140K images total. We improve upon these works by collecting much more diverse generative models to improve the performance and generalization of the detector. In concurrent work, Hong et al. [49] acquires user-created images from Midjourney and CivitAI. This strategy is complementary to ours: while it aims to collect in-the-wild fake images, its distribution is centered on images that users share, and the models are not necessarily identifiable, making it challenging to rigorously analyze the dataset's contents and to interpret experiments conducted on it." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.645, + 0.487, + 0.902 + ], + "angle": 0, + "content": "Fingerprint-based image forensics methods. Classic work on image forensics relied on methods based on image statistics [95] and physical constraints [56], rather than learning. A number of datasets have been created for detecting images that have been manipulated using traditional methods, such as with photo editors [21, 26, 53, 64, 84]. Recent works focus on detecting synthetic images by inspecting the generator fingerprints. Zhang et al. [131] and Marra et al. [80] proposed identifying the spatial fingerprints left by the generator to detect synthetic images. Others focus on spectral anomalies to detect synthetic images. Durall et al. [29] and Dzanic et al. [30] identified that CNN-generated images fail to reproduce certain spectral properties of real images. Corvi et al. [14] study the frequency fingerprints of the generated images and analyzes the cross-architecture generalization of the detector. Bammey [7] uses high-frequency artifacts to detect generated images. However, these approaches may be" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.909, + 0.244 + ], + "angle": 0, + "content": "brittle since the artifacts they rely on can be eliminated by post-processing [15]. We instead approach this problem in a data-driven manner, scaling the number of models, images, and architectures. Recent work has created ensembles of fake image classifiers [50]. In parallel, researchers have detected text generated by language models using supervised learning and heuristics [6, 35, 55, 65, 82, 103, 113, 121], which closely resemble those in visual forensics. However, no existing techniques that we are aware of aim to collect comprehensive datasets of community-created generators." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.248, + 0.911, + 0.446 + ], + "angle": 0, + "content": "Out-of-distribution generalization. Our work is related to the out-of-distribution recognition problem as it involves generalizing to unseen generators and image processing pipelines. A variety of approaches have been proposed for this problem, based on likelihood ratios [72, 98, 128], self-supervision [47, 83, 109, 124], internal model statistics [46, 104], temperature scaling [3, 73], and via energy-based models [28, 31, 75]. Works by Schuhmann et al. [108] and Hendrycks et al. [48] show that diverse training data and data augmentation are important to improving the robustness to out-of-distribution samples. Our results are in line with these conclusions, as we find that a diverse set of generative models and stronger augmentations improve generalization." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.458, + 0.822, + 0.475 + ], + "angle": 0, + "content": "3. The Community Forensics Dataset" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.48, + 0.909, + 0.722 + ], + "angle": 0, + "content": "To support our goal of studying generalization in generated image detection, we collect a dataset of images sampled from a wide range of models (Fig. 2). Our dataset consists of: (a) a large and systematically collected set of \"in-the-wild\" text-to-image latent diffusion models obtained from a model-sharing website, (b) hand-selected models from other open source architectures, and (c) closed and open state-of-the-art commercial models. We also pair these generated images with real images from other datasets. For all images in our dataset, we preserve the original image format where possible, without any additional compression or resizing. This is to mitigate potential bias and performance degradation in out-of-distribution settings due to unwanted artifacts [40, 42, 97]. Our dataset contains significantly more models than previous works (Tab. 1) and spans a wider range of architectures, processing pipelines, and semantic content." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.727, + 0.882, + 0.743 + ], + "angle": 0, + "content": "3.1. Systematically collecting generative models" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.746, + 0.909, + 0.868 + ], + "angle": 0, + "content": "We perform our systematic collection using publicly available, open source models that use the Hugging Face diffusers library [34, 123] because: 1) it is a popular library for creating text-to-image models and is widely used by hobbyists, 2) thousands of such models are publicly indexed, and 3) it provides a standard interface by which we can sample images. We process the models in the order of popularity, as indicated by the number of downloads. Our" + }, + { + "type": "page_footnote", + "bbox": [ + 0.513, + 0.876, + 0.908, + 0.902 + ], + "angle": 0, + "content": "1We use \"open source\" to describe models with public weights and code, even if they may be closed in some aspect (e.g., private training data)." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8247" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.092, + 0.089, + 0.905, + 0.25 + ], + "angle": 0, + "content": "
DatasetModelsImagesArchitecturesTraining setup
Wang et al. [125]11362KGAN, Perceptual, Deepfake, ...ProGAN [58] vs. LSUN [129]
Ojha et al. [87]4*10K*GAN, Perceptual, Diffusion, ...ProGAN [58] vs. LSUN [129]
Epstein et al. [32]14570KDiffusionDiffusion vs. LAION [108]
Cozzolino et al. [15]1826KDiffusionLDM [100] vs. MS-COCO [74]
Synthbuster [7]910KDiffusionDiffusion vs. Dresden [36]
GenImage [133]81.4MDiffusion, GANDiffusion, GAN vs. ImageNet [23]
RED116 [5]116116KGAN, VAE, Autoregressive, ...Many vs. Many
RED140 [43]140140KDiffusion, GAN, VAE, ...Many vs. Many
Ours48032.7MDiffusion, GAN, Autoregressive, ...Many vs. Many
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.259, + 0.907, + 0.304 + ], + "angle": 0, + "content": "Table 1. Comparison with existing forensics datasets. We compare the size of the dataset with existing datasets containing identifiable generative models. We only count the number of generated images. Our dataset contains significantly more generative models than prior works. \\(\\ast\\): Only counting the unique evaluation set by Ojha et al. [87] as their dataset is based on Wang et al. [125]." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.314, + 0.482, + 0.48 + ], + "angle": 0, + "content": "pipeline downloads each model and extracts relevant hyperparameters (e.g., number of diffusion steps and guidance scale), sampling pipeline configurations, and metadata, from both the model cards on the model-sharing webpage and the metadata retrieved by the diffusers library [34, 123]. We sample images using a distribution of text prompts obtained from various real datasets (Sec. 3.3). Since experiments suggest diminishing returns for repeatedly sampling from any given model, we sample a few hundred images from each one. We sample 4763 models with approximately 403 images each, for a total of \\(1.9\\mathrm{M}\\) images from this process." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.482, + 0.483, + 0.693 + ], + "angle": 0, + "content": "While the lack of documentation in each model and the scale of data collection make it challenging to exactly characterize the model designs in this set, they appear to be entirely (or almost entirely) based on latent diffusion. More specifically, we categorize models as being based on latent diffusion if they perform a denoising process on a latent representation. Based on this criterion and the self-reported tags, all models in our systematically collected set appear to be based on latent diffusion. While pixel-based diffusion models also use the diffusers library (e.g., DeepFloyd [22]), they were incompatible with our automated generation pipeline. We record such incompatible models and manually sample them to either construct an out-of-distribution test set (Sec. 3.4), or as manually-chosen models for training data (Sec. 3.2)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.695, + 0.483, + 0.875 + ], + "angle": 0, + "content": "We show examples of the sampled images in Fig. 2. In the supplement, we provide examples of models and information from their project pages. These models generate a variety of images, with various types of semantic content and preprocessing. For example, a large fraction of these models adapt variations of a popular pretrained latent diffusion model, Stable Diffusion [100], to different downstream applications, and use a number of adaptation strategies (e.g., using LoRA [51]). We provide the model metadata with each image to enable other possible forensics applications. We discuss these in the supplement and provide information about image and model licenses." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.314, + 0.878, + 0.33 + ], + "angle": 0, + "content": "3.2. Collecting images from other architectures" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.333, + 0.908, + 0.545 + ], + "angle": 0, + "content": "Images from manually chosen models. To ensure that our dataset contains a broader range of models, we manually select 19 models from public repositories and sample an average of 40,738 images per model, resulting in 774K images total. We note that this number is itself on par with (or more than) many prior datasets with identifiable generative models [7, 15, 32, 87, 125, 133]. We include several GANs (e.g., StyleGANs [60-62, 106], BigGAN [8], StyleSwin [130], GigaGAN [57], ProGAN [58], ProjectedGAN [105], GANformer [52], SAN [115], and CIPS [4]), pixel-based diffusion models (e.g., GLIDE [85], ADM [24], and DeepFloyd [22]), latent diffusion models (e.g., VQ-Diffusion [41], Diffusion Transformers [92], and Latent Flow Matching [20]), and an autoregressive model (Taming Transformers [33])." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.553, + 0.909, + 0.646 + ], + "angle": 0, + "content": "Images from commercial models. We sample 15K images from 11 commercial models using LAION-based captions to evaluate the generalization to state-of-the-art models with typically unknown architectures: DALL-E 2, 3 [88, 89], Ideogram V1, V2 [2], Midjourney V5, V6 [81], Firefly Image 2, 3 [1], FLUX.1-dev, Schnell [68], and Imagen 3 [38]." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.652, + 0.722, + 0.669 + ], + "angle": 0, + "content": "3.3. Collecting real images" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.672, + 0.909, + 0.75 + ], + "angle": 0, + "content": "To help study how real images influence forensics models, we source real images from a variety of existing datasets: LAION [107], ImageNet [23], COCO [74], FFHQ [59], CelebA [76], MetFaces [60], AFHQ [13], Forchheim [44], IMD2020 [86], Landscapes HQ [112], and VISION [111].3" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.756, + 0.759, + 0.772 + ], + "angle": 0, + "content": "3.4. Curating the evaluation set" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.776, + 0.909, + 0.853 + ], + "angle": 0, + "content": "We construct our evaluation set using the incompatible models from our automated sampling pipeline, commercial models (Sec. 3.2), and manually collected open source models. The evaluation set comprises 26K images sampled from 21 models not included in the training set." + }, + { + "type": "page_footnote", + "bbox": [ + 0.108, + 0.887, + 0.472, + 0.902 + ], + "angle": 0, + "content": "2We note that this definition includes latent consistency models [78]." + }, + { + "type": "page_footnote", + "bbox": [ + 0.513, + 0.863, + 0.908, + 0.903 + ], + "angle": 0, + "content": "3Following common convention, we refer to these images as real images, even though they may be synthetic (e.g., graphic design). More precisely, our goal is to distinguish \"AI-generated\" images from the originals." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8248" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.089, + 0.092, + 0.485, + 0.198 + ], + "angle": 0, + "content": "This includes our commercial models set and an additional 11K images from 10 models: Deci Diffusion V2 [120], GALIP [119], KandinskyV2.2 [110], Kvikontent [66], LCM-LoRA-SDv1.5, LCM-LoRA-SDXL, LCM-LoRA-SSD1B [79], Stable Cascade [93], DF-GAN [118], and HDiT [16], sampled using RAISE [19], ImageNet [23], FFHQ [59], and COCO [74]-based captions." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.199, + 0.486, + 0.408 + ], + "angle": 0, + "content": "The generated images are paired with the source real data that are used to prompt the generators. However, since some of the real datasets do not have appropriate licenses for redistribution (e.g., LAION [107, 108]), we created a public version of our evaluation set by pairing the generated images with openly licensed COCO [74] and FFHQ [59] which allow redistribution for non-commercial purposes. The public version of our evaluation set will serve as an easily reproducible and shareable evaluation set that will complement our default set. We will refer to our default set as the comprehensive evaluation set. We will also release the instructions to reconstruct our comprehensive set. However, note that it may not be possible to exactly reconstruct this set in the future due to link rot." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.415, + 0.273, + 0.431 + ], + "angle": 0, + "content": "3.5. Generating images" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.434, + 0.486, + 0.616 + ], + "angle": 0, + "content": "Unconditional models are sampled until we reach the desired number of images. For class conditional models, we sample an equal number of images per class. Text-conditional models are sampled using captions obtained from real images (Sec. 3.3). We either use captions that are already present in the dataset or use BLIP [69] to generate them. Some models such as GigaGAN [57] and HDiT [16] do not provide a pretrained model, so we instead use their pre-generated images. Generated images are saved in PNG format to avoid compression artifacts. However, Firefly [1] generated images are saved in JPEG format as their web UI only allows downloading in JPEG." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.622, + 0.225, + 0.639 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.643, + 0.485, + 0.75 + ], + "angle": 0, + "content": "We use our dataset to conduct a study of generalization in visual forensics, asking a number of questions: (1) How well do forensics models trained on our dataset generalize to unseen models? (2) Does adding more models improve detection performance? (3) How does diversity of the training data affect performance? (4) What architectures and data augmentation schemes are most successful?" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.755, + 0.376, + 0.771 + ], + "angle": 0, + "content": "4.1. Training image forensics models" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.775, + 0.485, + 0.837 + ], + "angle": 0, + "content": "We train binary classifiers that detect generated images using our dataset to study the generalization in image forensics. We construct our training set of 5.4M images by pairing 2.7M generated images with 2.7M real images." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.486, + 0.902 + ], + "angle": 0, + "content": "Training and evaluation setup. We evaluate the models trained on our dataset and compare them with prior works [43, 87, 125, 133]. As RED140 [43] already contains RED116 [5], we do not compare with RED116 dataset." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.909, + 0.44 + ], + "angle": 0, + "content": "Following prior works [87, 125], we use the threshold-independent mean average precision (mAP) and accuracy (Acc.) as our evaluation metrics. We compute the mAP and accuracy by averaging the results of each generative model. We use five evaluation sets: Wang et al. [125], Ojha et al. [87], Synthbuster [7], GenImage [133], and our evaluation set. All evaluation sets apart from GenImage [133] evaluate out-of-distribution performance for all classifiers. GenImage [133] evaluation set, however, contains the same set of generators used in training, and is an in-distribution evaluation set for their classifiers. Concretely, the evaluation set by Wang et al. [125] and Ojha et al. [87, 125] contains models such as CRN [10], CycleGAN [132], DALL-E [88], DeepFake [25], IMLE [70], SAN [18], StarGAN [12], and SITD [9] which are unseen by both their and our classifiers. Synthbuster [7] evaluation set is comprised of RAISE [19]-based synthetic images of DALL-E [88, 89], Firefly [1], Glide [85], Midjourney [81], and Stable Diffusion [94, 100], and is mostly out of distribution for all classifiers. GenImage [133] evaluation set is a validation split of their training set; they use an identical set of models to train their classifier: ADM [24], BigGAN [8], Glide [85], Midjourney [81], Stable Diffusion [100], VQ-Diffusion [41], and Wukong [127]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.451, + 0.909, + 0.724 + ], + "angle": 0, + "content": "Model architecture. Building on prior works which mainly used CLIP-ViT [27, 54, 96] and ResNet-50 [45], we consider ViT [27] and ConvNeXt [77] pretrained models for our classifiers. We use a plain ViT-S backbone [27] pretrained on CLIP objective [54, 96] using LAION-2B [108], ImageNet 21K, and ImageNet 1K datasets [23]. We also experiment with a ConvNeXt-S model [77] pretrained on ImageNet 21K and ImageNet 1K datasets [23]. We replace the classification head with a linear layer with sigmoid activation that outputs the probability of the image being generated. Unlike prior works [15, 87] that freeze the CLIP-ViT backbone, we train the backbone end-to-end. The models are obtained through timm [11, 126] library on Hugging Face. We experiment with two input resolutions, \\(224^2\\) and \\(384^2\\), to evaluate the impact of the input resolution on the detector's performance. We denote the detector with \\(384^2\\) input resolution as High res. We implement the models using PyTorch [91]. The hyperparameters are detailed in the supplement." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.911, + 0.903 + ], + "angle": 0, + "content": "Data augmentation. Prior work considered augmentations that were designed to simulate postprocessing, such as flipping, cropping, Gaussian blur, and JPEG recompression to train their detectors [7, 15, 87, 125]. We propose an augmentation scheme that extends this approach and compare it with previously proposed augmentation methods. We expand the set of augmentations to handle additional transformations that can occur in the wild, such as padding, resizing, rotation, and shear, and integrate them into a framework that can apply complex sequences of transformations. We introduce a modified version of RandAugment [17] that applies a" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8249" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.099, + 0.089, + 0.907, + 0.221 + ], + "angle": 0, + "content": "
ModelEvaluation Set (mAP)Evaluation Set (Acc)
Wang et al. [125]Ojha et al. [87]SB [7]GenImage [133]OursMeanWang et al. [125]Ojha et al. [87]SB [7]GenImage [133]OursMean
Wang et al. [125]0.8970.6960.5160.6420.5370.6000.6480.7140.5270.5080.5330.5130.5170.552
Ojha et al. [87]0.9390.9570.6200.7970.5920.6560.7600.7910.8210.5320.6410.5400.5480.646
GenImage [133]0.9290.9840.8130.9990.9120.9680.9340.7950.9660.7190.9900.8180.8860.862
RED140 [43] - High res.0.9000.9540.7650.9270.7640.8610.8620.6940.7800.5580.6740.5620.5650.639
Ours0.9640.9910.9040.9900.9710.9770.9660.8730.9500.8180.9460.8610.8880.889
Ours - High res.0.9670.9960.9740.9980.9870.9940.9860.9010.9700.9080.9570.8920.9120.923
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.226, + 0.907, + 0.337 + ], + "angle": 0, + "content": "Table 2. Generalization of AI-generated image detectors across benchmarks. We evaluate the classifiers trained on our dataset on several benchmarks, including our own. We also evaluate several previously released classifiers. Our Comprehensive set (abbreviated as Comp.) pairs the generated images with original real data; the Public set pairs them with openly licensed COCO [74] and FFHQ [59] for license-compliant redistribution of the evaluation set (Sec. 3.4). We use plain CLIP-ViT-S [27, 54, 96] architecture with \\(224^{2}\\) and \\(384^{2}\\) (High res.) input resolutions, Wang et al. [125] and GenImage [133] use ResNet-50 [45] with \\(224^{2}\\) input resolution, and Ojha et al. [87] uses CLIP-ViT-L with \\(224^{2}\\) input resolution as the backbone. Our classifiers show robust performance across all evaluation sets, outperforming all baselines in out-of-distribution settings ([7, 87, 125] and Ours) and nearly matches GenImage [133] on its in-distribution evaluation set. \\(\\spadesuit\\): RED140 [43] is trained following our training procedure with \\(384^{2}\\) input resolution as they do not provide a trained classifier." + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.347, + 0.482, + 0.436 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.444, + 0.484, + 0.542 + ], + "angle": 0, + "content": "Figure 3. Performance across generator types. We evaluate the classifier performance across five generator types - latent diffusion, pixel diffusion, commercial models, GANs, and other architecture (Stable Cascade [93]), and report the mean. We additionally evaluate three forensics methods, Aerblade [99], LGrad [116], and NPR [117]. Our classifiers show robust performance across all generator types, whereas prior works struggle to generalize." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.555, + 0.484, + 0.691 + ], + "angle": 0, + "content": "randomly-ordered sequence of augmentations to the images. Specifically, our modified RandAugment samples a random number \\( n \\) between 0 and \\( n_{\\mathrm{max}} \\) for each augmentation type. Then, it applies the augmentations in random order until \\( n \\) augmentations are applied for each augmentation type to the image. We use various augmentations, including in-memory JPEG compression, random resizing with random interpolation methods, cropping, flipping, rotation, translation, shear, padding, and cutout." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.7, + 0.393, + 0.714 + ], + "angle": 0, + "content": "4.2. Generalization across benchmarks" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.484, + 0.901 + ], + "angle": 0, + "content": "We first evaluate how well classifiers trained on our dataset perform across benchmarks. In Table 2, we observe that our models outperform the prior works [43, 87, 125, 133] on all evaluation sets except GenImage. This is expected since the GenImage evaluation set is a validation split of their training set; all of the generators are already seen by their classifier. On all other unseen benchmarks, our classifiers outperform all prior works. Notably, our classifiers achieve very high performance (0.987 mAP and \\(89.2\\%\\) accuracy) on our out-of-distribution, comprehensive evaluation set, with a significant margin over prior works. This gap in performance can be traced to our training data which in" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.347, + 0.71, + 0.458 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.712, + 0.347, + 0.903, + 0.459 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.466, + 0.907, + 0.591 + ], + "angle": 0, + "content": "Figure 4. Performance with increasing number of images. We train a classifier with varying numbers of images from two sets: 1000 randomly chosen models and 10 popular (highly downloaded) models in the systematically collected subset. We report the mean and standard error bands for each data point across 4 randomly sampled subsets. The classifier trained from 1000 random models outperforms 10 popular models in all cases. Notably, the accuracy gap is wider than that of mAP, which may suggest that having a diverse set of models improves accuracy threshold calibration." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.603, + 0.906, + 0.663 + ], + "angle": 0, + "content": "corporates a substantially richer variety of generators than existing works. Consequently, our classifiers demonstrate robust generalization to out-of-distribution data, where prior works often struggle." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.664, + 0.907, + 0.815 + ], + "angle": 0, + "content": "To better illustrate the generalization of the classifiers, in Figure 3, we split our comprehensive evaluation set into five subsets and evaluate them: latent diffusion, pixel diffusion, commercial models, GANs, and other architecture type (Stable Cascade [93]). Furthermore, we evaluate three additional forensics methods: Aerblade [99], LGrad [116], and NPR [117]. Our classifiers show strong performance across all generator types, unlike prior works which struggle to generalize. For the following experiments, we use our best-performing model (High res.) unless stated otherwise." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.821, + 0.747, + 0.837 + ], + "angle": 0, + "content": "4.3. Impact of model diversity" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Next, we examine the impact of the number of models in training data. We train classifiers with images sampled from 3 to 3333 generators and evaluate them (Fig. 1). To ensure that the gains are not due to simply sampling qualita" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8250" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.088, + 0.287, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.092, + 0.194, + 0.285, + 0.207 + ], + "angle": 0, + "content": "(a) Impact of generator type diversity" + }, + { + "type": "image", + "bbox": [ + 0.291, + 0.088, + 0.482, + 0.191 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.291, + 0.194, + 0.48, + 0.207 + ], + "angle": 0, + "content": "(b) Classifier backbone comparison" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.215, + 0.485, + 0.326 + ], + "angle": 0, + "content": "Figure 5. (a) Performance and model diversity. We compare detection performance for commercial models using classifiers trained on different subsets of the dataset: the systematically collected latent diffusion models, the manually chosen models containing diverse generator types, and both. As diversity increases, so does performance. (b) Classifier backbone comparison. We compare the architectures across datasets: ours, GenImage [133], and Wang et al. [125]. Performance is similar between architectures." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.341, + 0.483, + 0.566 + ], + "angle": 0, + "content": "tively different architectures, we only use our systematically collected latent diffusion models. For each data point, we sample 10 random subsets of models with 100K training images each and report the mean and standard error bands. We use an extended evaluation set that includes non-latent diffusion generators from our training set, which allows us to comprehensively assess the generalization capability of the classifiers trained exclusively on latent diffusion models. We find that the performance steadily increases with the number of models. However, the performance begins to flatten out beyond 1000 models, suggesting diminishing returns. Interestingly, the performance also improves on out-of-distribution architectures such as GANs and pixel-based diffusion models, even though the classifier is only trained on latent diffusion models." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.568, + 0.483, + 0.749 + ], + "angle": 0, + "content": "In Figure 4, we vary the number of images from two sets: 1000 randomly chosen models and 10 popular models (as denoted by their number of downloads) downloaded from our systematically collected diffusion models. While the results show that the performance improves with more training images, it begins to plateau at approximately \\(27\\mathrm{K}\\) images. Moreover, the classifier trained on 1000 models outperforms the 10 models in all cases, indicating that model diversity is important for strong performance. We also note that the accuracy gap is noticeably wider than that of mAP, which may suggest that model diversity is crucial in calibrating the accuracy thresholds of the classifiers." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.486, + 0.902 + ], + "angle": 0, + "content": "Our experiments show that the performance improvements from increasing the number of models may plateau when they are limited to a single generator type (Fig. 1). In Figure 5a, we show that the diversity of the generator type also plays a major role in generalization. We train classifiers on three different sets of training data: our systematically collected set, manually chosen set, and a full set consisting of both subsets. The systematic set comprises entirely of latent diffusion models, and the manual set contains numerous generator types, including GANs, latent and pixel-based dif" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.088, + 0.707, + 0.191 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.529, + 0.194, + 0.696, + 0.207 + ], + "angle": 0, + "content": "(a) Impact of frozen backbones" + }, + { + "type": "image", + "bbox": [ + 0.715, + 0.089, + 0.904, + 0.191 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.724, + 0.194, + 0.895, + 0.207 + ], + "angle": 0, + "content": "(b) Semantic alignment analysis" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.215, + 0.908, + 0.342 + ], + "angle": 0, + "content": "Figure 6. (a) Evaluating frozen backbones. Freezing the pretrained backbone, a common practice in prior works [15, 87], consistently decreases the performance. (b) Analyzing source and generated data alignment. We evaluate how the pairing of the real datasets affects performance. 'R' denotes the real dataset used in training, and 'F' indicates the source dataset used to obtain the captions for prompting the generators. The results suggest that pairing the source data (i.e., real data used to prompt the generators) with the generated images is not essential for performance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.353, + 0.907, + 0.46 + ], + "angle": 0, + "content": "fusion, and autoregressive models (Sec. 3.2). The classifier trained on the manual set with more diverse generator types shows similar performance compared to the one trained on the systematic set, despite containing fewer than half the number of images (774K vs. 1.9M). Additionally, we find that the two sets are complementary; the performance is further improved when we train using both sets." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.467, + 0.747, + 0.482 + ], + "angle": 0, + "content": "4.4. Analysis of design choices" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.486, + 0.909, + 0.561 + ], + "angle": 0, + "content": "We examine the impact of various design choices, including some suggested in earlier works. In particular, we investigate the choice of backbone models, freezing the backbone, semantic alignment between the real and generated data, and robustness to transformations." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.569, + 0.909, + 0.66 + ], + "angle": 0, + "content": "Classifier backbone comparison. We compare the performance of the classifier trained using CLIP-ViT [27, 54, 96] and ConvNeXt [77] backbones following our training procedure in Figure 5b. We examine three datasets: ours, GenImage [133], and Wang et al. [125]. We observe similar performance between architectures across all datasets." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.667, + 0.909, + 0.773 + ], + "angle": 0, + "content": "Frozen backbone. Prior works [15, 87] suggested using a frozen CLIP-ViT backbone for training the classifiers. We study this by training the classifiers with both frozen and unfrozen pretrained backbones, using CLIP-ViT [27, 54, 96] and ConvNeXt [77]. As shown in Fig. 6a, freezing the backbone consistently leads to worse results, indicating that end-to-end training is crucial for high performance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.911, + 0.902 + ], + "angle": 0, + "content": "Semantic alignment. Existing works often pair the generated images with the source dataset (i.e., the real dataset used to prompt or generate the images) arguing that misaligned data can introduce bias [7, 15, 87, 125]. We test this practice in Fig. 6b by examining the performance with both semantically aligned and misaligned real datasets. Specifically, we consider two real datasets: one comprised exclusively from LAION [107] and another combining ImageNet [23]," + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "8251" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.09, + 0.482, + 0.299 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.305, + 0.483, + 0.348 + ], + "angle": 0, + "content": "Figure 7. Robustness to various transformations. Our classifiers display robust performance across transformations. Other works generally show more sensitivity to these factors." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.369, + 0.484, + 0.475 + ], + "angle": 0, + "content": "MS-COCO [74], LandscapesHQ [112], Forchheim [44], VISION [111], and IMD2020 [86]. We sample our systematically collected latent diffusion models using these two sets and categorize the generated images by their source dataset. The performance differences of all pairs were marginal, suggesting that strict alignment may not be as critical with sufficiently large data." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.508, + 0.484, + 0.781 + ], + "angle": 0, + "content": "Robustness to transformations. Figure 7 illustrates the robustness of the classifiers against various transformations. Following prior works [87, 125], we test robustness to JPEG compression and Gaussian blur. Additionally, we examine robustness to rotation, resizing, padding, and shear, as they commonly occur in real-world scenarios. For padding, we randomly pad the width or height of the image with a given fraction and scale it back to the original size. Similarly, in resize, we randomly upsample or downsample the height and width of the image by a given fraction and resize it back to the original size (e.g., if fraction is 0.3, resize the height and width to \\(0.7 \\times\\) or \\(1.3 \\times\\) and then scale it back to the original size). The results demonstrate that our models are more robust to transformations than existing models. Specifically, GenImage [133] is notably more sensitive to Gaussian blur, JPEG compression, and resizing artifacts; classifier by Ojha et al. [87] displays sensitivity to shear transforms, and the one by Wang et al. [125] performs poorly overall." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.801, + 0.342, + 0.817 + ], + "angle": 0, + "content": "4.5. Other forensics applications" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Our dataset may enable further forensics studies that can take advantage of our diverse array of generators. To illustrate this, in the supplement, we identify the type of generator used to synthesize a given image by using \\( k \\)-nearest neighbor in the feature space of our classifier." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.09, + 0.634, + 0.105 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.113, + 0.907, + 0.265 + ], + "angle": 0, + "content": "In this paper, we studied the problem of generalizing to unseen generative models in synthetic image detection. We proposed a new dataset, Community Forensics, which contains 4803 models and 2.7M images collected from various public sources. We studied the impact of model diversity and demonstrated that it plays a crucial role in enhancing data diversity and generalization performance. We trained classifiers on our dataset, studied their ability to generalize in various settings, and evaluated previously proposed models and training practices." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.266, + 0.909, + 0.403 + ], + "angle": 0, + "content": "We do not intend for our dataset to be used to train classifiers that are directly used in the wild. Detecting in-the-wild synthetic images remains a challenging open problem, and detection errors can have severe consequences (e.g., falsely accusing an author of creating fake images or allowing misinformation to be certified as real). We hope that our work will serve as a stepping stone for future research in this area by providing tools and insights for studying generalization and data collection strategies." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.418, + 0.909, + 0.735 + ], + "angle": 0, + "content": "Limitations. While our dataset is diverse, a large portion of the data is diffusion-based, especially models based on Stable Diffusion [100]. Despite this, Figure 1 demonstrates that adding more diffusion models to the training data improves the generalization capability of the classifiers across other generator types. Furthermore, Figure 5a illustrates improved classifier performance when combining the diffusion-only systematic set with a more diverse manual set. These results suggest that despite their similarities, each diffusion model may uniquely contribute through its semantic content, image processing pipelines, or architectural variations (Sec. 3.1). Future work may explore collecting more diverse models, including GANs, VQ-VAEs [122], and autoregressive models. We also note that the generative models sourced from the community may contain inappropriate content. While in many contexts it is important to detect such images, these generated images may require further scrutiny before being used in other downstream applications. Finally, although our experiments suggest that our forensics classifiers generalize to unseen models better than those of previous work, their error rates are still too high for many critical applications." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.911, + 0.9 + ], + "angle": 0, + "content": "Acknowledgements. We thank the creators of the many open source models that we used to collect the Community Forensics dataset. We thank Chenhao Zheng, Cameron Johnson, Matthias Kirchner, Daniel Geng, Ziyang Chen, Ayush Shrivastava, Yiming Dou, Chao Feng, Xuanchen Lu, Zihao Wei, Zixuan Pan, Inbum Park, Rohit Banerjee, and Ang Cao for the valuable discussions and feedback. This research was developed with funding from the Defense Advanced Research Projects Agency (DARPA) under Contract No. HR001120C0123." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8252" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.115, + 0.484, + 0.143 + ], + "angle": 0, + "content": "[1] Adobe. Firefly. https://www.adobe.com/products/firefly, 2023.4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.145, + 0.484, + 0.159 + ], + "angle": 0, + "content": "[2] Ideogram AI. Ideogram. https://ideogram.ai, 2024.4" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.161, + 0.484, + 0.214 + ], + "angle": 0, + "content": "[3] Rushil Anirudh and Jayaraman J Thiagarajan. Out of distribution detection via neural network anchoring. In Asian Conference on Machine Learning, pages 32-47. PMLR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.218, + 0.484, + 0.273 + ], + "angle": 0, + "content": "[4] Ivan Anokhin, Kirill Demochkin, Taras Khakhulin, Gleb Sterkin, Victor Lempitsky, and Denis Korzhenkov. Image generators with conditionally-independent pixel synthesis. arXiv preprint arXiv:2011.13775, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.275, + 0.484, + 0.343 + ], + "angle": 0, + "content": "[5] Vishal Asnani, Xi Yin, Tal Hassner, and Xiaoming Liu. Reverse engineering of generative models: Inferring model hyperparameters from generated images. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(12):15477-15493, 2023. 2, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.346, + 0.484, + 0.401 + ], + "angle": 0, + "content": "[6] Anton Bakhtin, Sam Gross, Myle Ott, Yuntian Deng, Marc'Aurelio Ranzato, and Arthur Szlam. Real or fake? learning to discriminate machine from human generated text. arXiv preprint arXiv:1906.03351, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.403, + 0.484, + 0.445 + ], + "angle": 0, + "content": "[7] Quentin Bammey. Synthbuster: Towards detection of diffusion model generated images. IEEE Open Journal of Signal Processing, 2023. 2, 3, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.447, + 0.484, + 0.5 + ], + "angle": 0, + "content": "[8] Andrew Brock, Jeff Donahue, and Karen Simonyan. Large scale GAN training for high fidelity natural image synthesis. In International Conference on Learning Representations, 2019. 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.504, + 0.484, + 0.559 + ], + "angle": 0, + "content": "[9] Chen Chen, Qifeng Chen, Jia Xu, and Vladlen Koltun. Learning to see in the dark. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3291-3300, 2018. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.562, + 0.484, + 0.616 + ], + "angle": 0, + "content": "[10] Qifeng Chen and Vladlen Koltun. Photographic image synthesis with cascaded refinement networks. In Proceedings of the IEEE international conference on computer vision, pages 1511-1520, 2017. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.619, + 0.484, + 0.687 + ], + "angle": 0, + "content": "[11] Mehdi Cherti, Romain Beaumont, Ross Wightman, Mitchell Wortsman, Gabriel Ilharco, Cade Gordon, Christoph Schuhmann, Ludwig Schmidt, and Jenia Jitsev. Reproducible scaling laws for contrastive language-image learning. arXiv preprint arXiv:2212.07143, 2022.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.69, + 0.484, + 0.758 + ], + "angle": 0, + "content": "[12] Yunjey Choi, Minje Choi, Munyoung Kim, Jung-Woo Ha, Sunghun Kim, and Jaegul Choo. Stargan: Unified generative adversarial networks for multi-domain image-to-image translation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2018. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.761, + 0.484, + 0.815 + ], + "angle": 0, + "content": "[13] Yunjey Choi, Youngjung Uh, Jaejun Yoo, and Jung-Woo Ha. Stargan v2: Diverse image synthesis for multiple domains. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.818, + 0.484, + 0.899 + ], + "angle": 0, + "content": "[14] Riccardo Corvi, Davide Cozzolino, Giada Zingarini, Giovanni Poggi, Koki Nagano, and Luisa Verdoliva. On the detection of synthetic images generated by diffusion models. In ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1-5. IEEE, 2023. 3" + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.093, + 0.908, + 0.162 + ], + "angle": 0, + "content": "[15] Davide Cozzolino, Giovanni Poggi, Riccardo Corvi, Matthias Nießner, and Luisa Verdoliva. Raising the bar of ai-generated image detection with clip. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4356-4366, 2024. 2, 3, 4, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.165, + 0.908, + 0.247 + ], + "angle": 0, + "content": "[16] Katherine Crowson, Stefan Andreas Baumann, Alex Birch, Tanishq Mathew Abraham, Daniel Z Kaplan, and Enrico Shippole. Scalable high-resolution pixel-space image synthesis with hourglass diffusion transformers. In Proceedings of the 41st International Conference on Machine Learning, pages 9550-9575. PMLR, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.251, + 0.908, + 0.318 + ], + "angle": 0, + "content": "[17] Ekin Dogus Cubuk, Barret Zoph, Jon Shlens, and Quoc Le. Randaugment: Practical automated data augmentation with a reduced search space. In Advances in Neural Information Processing Systems, pages 18613-18624. Curran Associates, Inc., 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.323, + 0.908, + 0.39 + ], + "angle": 0, + "content": "[18] Tao Dai, Jianrui Cai, Yongbing Zhang, Shu-Tao Xia, and Lei Zhang. Second-order attention network for single image super-resolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11065-11074, 2019. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.394, + 0.908, + 0.45 + ], + "angle": 0, + "content": "[19] Duc-Tien Dang-Nguyen, Cecilia Pasquini, Valentina Conotter, and Giulia Boato. Raise: A raw images dataset for digital image forensics. In Proceedings of the 6th ACM multimedia systems conference, pages 219-224, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.453, + 0.908, + 0.492 + ], + "angle": 0, + "content": "[20] Quan Dao, Hao Phung, Binh Nguyen, and Anh Tran. Flow matching in latent space. arXiv preprint arXiv:2307.08698, 2023.4" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.497, + 0.908, + 0.565 + ], + "angle": 0, + "content": "[21] Tiago José De Carvalho, Christian Riess, Eli Angelopoulou, Helio Pedrini, and Anderson de Rezende Rocha. Exposing digital image forgeries by illumination color classification. IEEE Transactions on Information Forensics and Security, 8(7):1182-1194, 2013. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.569, + 0.907, + 0.595 + ], + "angle": 0, + "content": "[22] DeepFloyd. Deepfloyd. https://huggingface.co/DeepFloyd/IF-I-L-v1.0, 2024.4" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.6, + 0.908, + 0.654 + ], + "angle": 0, + "content": "[23] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 4, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.657, + 0.908, + 0.699 + ], + "angle": 0, + "content": "[24] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.702, + 0.908, + 0.756 + ], + "angle": 0, + "content": "[25] Brian Dolhansky, Joanna Bitton, Ben Pflaum, Jikuo Lu, Russ Howes, Menglin Wang, and Cristian Canton Ferrer. The deepfake detection challenge (dfdc) dataset. arXiv preprint arXiv:2006.07397, 2020. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.76, + 0.908, + 0.815 + ], + "angle": 0, + "content": "[26] Jing Dong, Wei Wang, and Tieniu Tan. Casia image tampering detection evaluation database. In 2013 IEEE China summit and international conference on signal and information processing, pages 422-426. IEEE, 2013. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.818, + 0.908, + 0.899 + ], + "angle": 0, + "content": "[27] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. ICLR, 2021. 5, 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.525, + 0.093, + 0.908, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8253" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.093, + 0.486, + 0.147 + ], + "angle": 0, + "content": "[28] Yilun Du, Shuang Li, Joshua Tenenbaum, and Igor Mordatch. Improved contrastive divergence training of energy-based models. In International Conference on Machine Learning, pages 2837-2848. PMLR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.149, + 0.485, + 0.218 + ], + "angle": 0, + "content": "[29] Ricard Durall, Margret Keuper, and Janis Keuper. Watch your up-convolution: Cnn based generative deep neural networks are failing to reproduce spectral distributions. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7890–7899, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.219, + 0.486, + 0.272 + ], + "angle": 0, + "content": "[30] Tarik Dzanic, Karan Shah, and Freddie Witherden. Fourier spectrum discrepancies in deep network generated images. Advances in neural information processing systems, 33:3022-3032, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.274, + 0.483, + 0.315 + ], + "angle": 0, + "content": "[31] Sven Elflein, Bertrand Charpentier, Daniel Zügner, and Stephan Gunnemann. On out-of-distribution detection with energy-based models, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.316, + 0.484, + 0.372 + ], + "angle": 0, + "content": "[32] David C Epstein, Ishan Jain, Oliver Wang, and Richard Zhang. Online detection of ai-generated images. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 382-392, 2023. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.372, + 0.484, + 0.427 + ], + "angle": 0, + "content": "[33] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12873-12883, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.428, + 0.484, + 0.468 + ], + "angle": 0, + "content": "[34] Hugging Face. Hugging face diffusers library. https://huggingface.co/models?library=diffusers, accessed on June 05, 2022, 2022. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.47, + 0.484, + 0.511 + ], + "angle": 0, + "content": "[35] Sebastian Gehrmann, Hendrik Strobelt, and Alexander M Rush. Gltr: Statistical detection and visualization of generated text. arXiv preprint arXiv:1906.04043, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.512, + 0.484, + 0.566 + ], + "angle": 0, + "content": "[36] Thomas Gloe and Rainer Böhme. The'dresden image database' for benchmarking digital image forensics. In Proceedings of the 2010 ACM symposium on applied computing, pages 1584-1590, 2010. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.568, + 0.482, + 0.622 + ], + "angle": 0, + "content": "[37] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in Neural Information Processing Systems, 27, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.623, + 0.483, + 0.651 + ], + "angle": 0, + "content": "[38] Google. Imagen 3. https://deepmind.google/technologies/imagen-3, 2024.4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.652, + 0.484, + 0.705 + ], + "angle": 0, + "content": "[39] Karol Gregor, Ivo Danihelka, Andriy Mnih, Charles Blundell, and Daan Wierstra. Deep autoregressive networks. In International Conference on Machine Learning, pages 1242-1250. PMLR, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.707, + 0.484, + 0.761 + ], + "angle": 0, + "content": "[40] Patrick Grommelt, Louis Weiss, Franz-Josef Pfreundt, and Janis Keuper. Fake or.jpeg? revealing common biases in generated image detection datasets. arXiv preprint arXiv:2403.17608, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.763, + 0.484, + 0.843 + ], + "angle": 0, + "content": "[41] Shuyang Gu, Dong Chen, Jianmin Bao, Fang Wen, Bo Zhang, Dongdong Chen, Lu Yuan, and Baining Guo. Vector quantized diffusion model for text-to-image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10696-10706, 2022. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.846, + 0.484, + 0.9 + ], + "angle": 0, + "content": "[42] Fabrizio Guillaro, Giada Zingarini, Ben Usman, Avneesh Sud, Davide Cozzolino, and Luisa Verdoliva. A bias-free training paradigm for more general ai-generated image detection, 2024. 3" + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.093, + 0.486, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.093, + 0.908, + 0.161 + ], + "angle": 0, + "content": "[43] Xiao Guo, Vishal Asnani, Sijja Liu, and Xiaoming Liu. Tracing hyperparameter dependencies for model parsing via learnable graph pooling network. Advances in Neural Information Processing Systems, 37:116899-116932, 2024. 2, 3, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.163, + 0.908, + 0.232 + ], + "angle": 0, + "content": "[44] Benjamin Hadwiger and Christian Riess. The forchheim image database for camera identification in the wild. In Pattern Recognition. ICPR International Workshops and Challenges: Virtual Event, January 10–15, 2021, Proceedings, Part VI, pages 500–515. Springer, 2021. 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.233, + 0.908, + 0.288 + ], + "angle": 0, + "content": "[45] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.289, + 0.908, + 0.343 + ], + "angle": 0, + "content": "[46] Dan Hendrycks and Kevin Gimpel. A baseline for detecting misclassified and out-of-distribution examples in neural networks. In International Conference on Learning Representations, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.344, + 0.907, + 0.412 + ], + "angle": 0, + "content": "[47] Dan Hendrycks, Mantas Mazeika, Saurav Kadavath, and Dawn Song. Using self-supervised learning can improve model robustness and uncertainty. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.414, + 0.908, + 0.511 + ], + "angle": 0, + "content": "[48] Dan Hendrycks, Steven Basart, Norman Mu, Saurav Kadavath, Frank Wang, Evan Dorundo, Rahul Desai, Tyler Zhu, Samyak Parajuli, Mike Guo, Dawn Song, Jacob Steinhardt, and Justin Gilmer. The many faces of robustness: A critical analysis of out-of-distribution generalization. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 8340-8349, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.511, + 0.908, + 0.553 + ], + "angle": 0, + "content": "[49] Yan Hong and Jianfu Zhang. Wildfake: A large-scale challenging dataset for ai-generated images detection. arXiv preprint arXiv:2402.11843, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.553, + 0.908, + 0.608 + ], + "angle": 0, + "content": "[50] Shuwei Hou, Yan Ju, Chengzhe Sun, Shan Jia, Lipeng Ke, Riky Zhou, Anita Nikolich, and Siwei Lyu. Deepfake-meter v2.0: An open platform for deepfake detection. arXiv preprint arXiv:2404.13146, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.609, + 0.908, + 0.664 + ], + "angle": 0, + "content": "[51] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.665, + 0.907, + 0.707 + ], + "angle": 0, + "content": "[52] Drew A Hudson and Larry Zitnick. Generative adversarial transformers. In International conference on machine learning, pages 4487-4499. PMLR, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.707, + 0.907, + 0.762 + ], + "angle": 0, + "content": "[53] Minyoung Huh, Andrew Liu, Andrew Owens, and Alexei A Efros. Fighting fake news: Image splice detection via learned self-consistency. In Proceedings of the European conference on computer vision (ECCV), 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.763, + 0.908, + 0.845 + ], + "angle": 0, + "content": "[54] Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. Openclip, 2021. If you use this software, please cite it as below. 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.846, + 0.908, + 0.9 + ], + "angle": 0, + "content": "[55] Ganesh Jawahar, Muhammad Abdul-Mageed, and Laks VS Lakshmanan. Automatic detection of machine generated text: A critical survey. arXiv preprint arXiv:2011.01314, 2020. 3" + }, + { + "type": "list", + "bbox": [ + 0.524, + 0.093, + 0.908, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8254" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.092, + 0.486, + 0.134 + ], + "angle": 0, + "content": "[56] Micah K Johnson and Hany Farid. Exposing digital forgeries in complex lighting environments. IEEE Transactions on Information Forensics and Security, 2(3):450-461, 2007. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.134, + 0.485, + 0.204 + ], + "angle": 0, + "content": "[57] Minguk Kang, Jun-Yan Zhu, Richard Zhang, Jaesik Park, Eli Shechtman, Sylvain Paris, and Taesung Park. Scaling up gans for text-to-image synthesis. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.205, + 0.484, + 0.259 + ], + "angle": 0, + "content": "[58] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of gans for improved quality, stability, and variation. In International Conference on Learning Representations, 2018. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.26, + 0.484, + 0.327 + ], + "angle": 0, + "content": "[59] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 3, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.33, + 0.484, + 0.385 + ], + "angle": 0, + "content": "[60] Tero Karras, Miika Aittala, Janne Hellsten, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Training generative adversarial networks with limited data. Advances in neural information processing systems, 33:12104-12114, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.386, + 0.484, + 0.454 + ], + "angle": 0, + "content": "[61] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8110-8119, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.455, + 0.484, + 0.51 + ], + "angle": 0, + "content": "[62] Tero Karras, Miika Aittala, Samuli Laine, Erik Härkonen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Alias-free generative adversarial networks. Advances in neural information processing systems, 34:852-863, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.511, + 0.484, + 0.566 + ], + "angle": 0, + "content": "[63] Hasam Khalid, Shahroz Tariq, Minha Kim, and Simon Woo. Fakeavceleb: A novel audio-video multimodal deepfake dataset. In Proceedings of the Neural Information Processing Systems Track on Datasets and Benchmarks, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.567, + 0.484, + 0.608 + ], + "angle": 0, + "content": "[64] P. Korus and J. Huang. Multi-scale analysis strategies in prnu-based tampering localization. IEEE Trans. on Information Forensics & Security, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.609, + 0.484, + 0.676 + ], + "angle": 0, + "content": "[65] Kalpesh Krishna, Yixiao Song, Marzena Karpinska, John Wieting, and Mohit Iyyer. Paraphrasing evades detectors of ai-generated text, but retrieval is an effective defense. Advances in Neural Information Processing Systems, 36, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.678, + 0.484, + 0.706 + ], + "angle": 0, + "content": "[66] Kvikontent. Kvikontent-midjourney v6. https://huggingface.co/Kvikontent/midjourney-v6, 2023.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.707, + 0.484, + 0.761 + ], + "angle": 0, + "content": "[67] Patrick Kwon, Jaeseong You, Gyuhyeon Nam, Sungwoo Park, and Gyeongsu Chae. Kodf: A large-scale korean deepfake detection dataset. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.762, + 0.484, + 0.788 + ], + "angle": 0, + "content": "[68] Black Forst Labs. Flux. https://blackforestlabs.ai, 2024.4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.79, + 0.484, + 0.844 + ], + "angle": 0, + "content": "[69] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.846, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[70] Ke Li, Tianhao Zhang, and Jitendra Malik. Diverse image synthesis from semantic layouts via conditional imle. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4220-4229, 2019. 3, 5" + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.092, + 0.486, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.093, + 0.907, + 0.147 + ], + "angle": 0, + "content": "[71] Yuezun Li, Xin Yang, Pu Sun, Honggang Qi, and Siwei Lyu. Celeb-df: A large-scale challenging dataset for deepfake forensics. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.149, + 0.907, + 0.216 + ], + "angle": 0, + "content": "[72] Yewen Li, Chaojie Wang, Xiaobo Xia, Tongliang Liu, Bo An, et al. Out-of-distribution detection with an adaptive likelihood ratio on informative hierarchical vae. Advances in Neural Information Processing Systems, 35:7383-7396, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.218, + 0.907, + 0.272 + ], + "angle": 0, + "content": "[73] Shiyu Liang, Yixuan Li, and R Srikant. Enhancing the reliability of out-of-distribution image detection in neural networks. In International Conference on Learning Representations, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.274, + 0.907, + 0.357 + ], + "angle": 0, + "content": "[74] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 4, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.358, + 0.907, + 0.412 + ], + "angle": 0, + "content": "[75] Weitang Liu, Xiaoyun Wang, John Owens, and Yixuan Li. Energy-based out-of-distribution detection. In Advances in Neural Information Processing Systems, pages 21464-21475. Curran Associates, Inc., 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.414, + 0.907, + 0.466 + ], + "angle": 0, + "content": "[76] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaou Tang. Deep learning face attributes in the wild. In Proceedings of International Conference on Computer Vision (ICCV), 2015. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.469, + 0.907, + 0.537 + ], + "angle": 0, + "content": "[77] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11976-11986, 2022. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.539, + 0.907, + 0.593 + ], + "angle": 0, + "content": "[78] Simian Luo, Yiqin Tan, Longbo Huang, Jian Li, and Hang Zhao. Latent consistency models: Synthesizing high-resolution images with few-step inference. arXiv preprint arXiv:2310.04378, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.595, + 0.906, + 0.649 + ], + "angle": 0, + "content": "[79] Simian Luo, Yiqin Tan, Suraj Patil, Daniel Gu, Patrick von Platen, Apolinário Passos, Longbo Huang, Jian Li, and Hang Zhao. Lcm-lora: A universal stable-diffusion acceleration module. arXiv preprint arXiv:2311.05556, 2023. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.651, + 0.907, + 0.705 + ], + "angle": 0, + "content": "[80] Francesco Marra, Diego Gragnaniello, Luisa Verdoliva, and Giovanni Poggi. Do gans leave artificial fingerprints? In 2019 IEEE conference on multimedia information processing and retrieval (MIPR), pages 506-511. IEEE, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.707, + 0.907, + 0.732 + ], + "angle": 0, + "content": "[81] Inc. Midjourney. Midjourney. https://www.midjourney.com/home, 2022. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.734, + 0.907, + 0.802 + ], + "angle": 0, + "content": "[82] Eric Mitchell, Yoonho Lee, Alexander Khazatsky, Christopher D Manning, and Chelsea Finn. Detectgpt: Zero-shot machine-generated text detection using probability curvature. In International Conference on Machine Learning, pages 24950-24962. PMLR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.804, + 0.907, + 0.858 + ], + "angle": 0, + "content": "[83] Sina Mohseni, Mandar Pitale, JBS Yadawa, and Zhangyang Wang. Self-supervised learning for generalizable out-of-distribution detection. Proceedings of the AAAI Conference on Artificial Intelligence, 34(04):5216-5223, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.86, + 0.907, + 0.901 + ], + "angle": 0, + "content": "[84] Tian-Tsong Ng, Shih-Fu Chang, and Q Sun. A data set of authentic and spliced image blocks. Columbia University, ADVENT Technical Report, 4, 2004. 3" + }, + { + "type": "list", + "bbox": [ + 0.523, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8255" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.093, + 0.486, + 0.176 + ], + "angle": 0, + "content": "[85] Alexander Quinn Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. In International Conference on Machine Learning, pages 16784-16804. PMLR, 2022. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.178, + 0.485, + 0.247 + ], + "angle": 0, + "content": "[86] Adam Novozamsky, Babak Mahdian, and Stanislav Saic. Imd2020: A large-scale annotated dataset tailored for detecting manipulated images. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision Workshops, pages 71-80, 2020. 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.249, + 0.484, + 0.317 + ], + "angle": 0, + "content": "[87] Utkarsh Ojha, Yuheng Li, and Yong Jae Lee. Towards universal fake image detectors that generalize across generative models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 24480-24489, 2023. 1, 2, 3, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.319, + 0.485, + 0.346 + ], + "angle": 0, + "content": "[88] OpenAI. Dall-e 2. https://openai.com/index/dall-e-2, 2022. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.349, + 0.484, + 0.375 + ], + "angle": 0, + "content": "[89] OpenAI. Dall-e 3. https://openai.com/index/dall-e-3, 2023.4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.378, + 0.484, + 0.433 + ], + "angle": 0, + "content": "[90] Taesung Park, Ming-Yu Liu, Ting-Chun Wang, and JunYan Zhu. Semantic image synthesis with spatially-adaptive normalization. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.435, + 0.484, + 0.516 + ], + "angle": 0, + "content": "[91] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.519, + 0.484, + 0.573 + ], + "angle": 0, + "content": "[92] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4195-4205, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.576, + 0.484, + 0.645 + ], + "angle": 0, + "content": "[93] Pablo Pernias, Dominic Rampas, Mats Leon Richter, Christopher Pal, and Marc Aubreville. Würstchen: An efficient architecture for large-scale text-to-image diffusion models. In The Twelfth International Conference on Learning Representations, 2023. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.647, + 0.484, + 0.728 + ], + "angle": 0, + "content": "[94] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. In The Twelfth International Conference on Learning Representations, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.732, + 0.482, + 0.773 + ], + "angle": 0, + "content": "[95] Alin C Popescu and Hany Farid. Exposing digital forgeries by detecting traces of resampling. IEEE Transactions on signal processing, 2005. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.775, + 0.484, + 0.857 + ], + "angle": 0, + "content": "[96] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.859, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[97] Anirudh Sundara Rajan, Utkarsh Ojha, Jedidiah Schoesser, and Yong Jae Lee. Aligned datasets improve detection of latent diffusion-generated images, 2025. 3" + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.093, + 0.486, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.093, + 0.907, + 0.161 + ], + "angle": 0, + "content": "[98] Jie Ren, Peter J Liu, Emily Fertig, Jasper Snoek, Ryan Poplin, Mark Depristo, Joshua Dillon, and Balaji Lakshminarayanan. Likelihood ratios for out-of-distribution detection. Advances in neural information processing systems, 32, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.164, + 0.907, + 0.233 + ], + "angle": 0, + "content": "[99] Jonas Ricker, Denis Lukovnikov, and Asja Fischer. Aeroblade: Training-free detection of latent diffusion images using autoencoder reconstruction error. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9130-9140, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.235, + 0.907, + 0.304 + ], + "angle": 0, + "content": "[100] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, 2022. 2, 4, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.305, + 0.907, + 0.361 + ], + "angle": 0, + "content": "[101] Andreas Rössler, Davide Cozzolino, Luisa Verdoliva, Christian Riess, Justus Thies, and Matthias Nießner. Faceforensics: A large-scale video dataset for forgery detection in human faces. arXiv preprint arXiv:1803.09179, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.363, + 0.907, + 0.431 + ], + "angle": 0, + "content": "[102] Andreas Rossler, Davide Cozzolino, Luisa Verdoliva, Christian Riess, Justus Thies, and Matthias Nießner. Faceforensics++: Learning to detect manipulated facial images. In Proceedings of the IEEE/CVF international conference on computer vision, 2019. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.434, + 0.907, + 0.488 + ], + "angle": 0, + "content": "[103] Vinu Sankar Sadasivan, Aounon Kumar, Sriram Balasubramanian, Wenxiao Wang, and Soheil Feizi. Can ai-generated text be reliably detected? arXiv preprint arXiv:2303.11156, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.49, + 0.907, + 0.546 + ], + "angle": 0, + "content": "[104] Chandramouli Shama Sastry and Sageev Oore. Detecting out-of-distribution examples with Gram matrices. In Proceedings of the 37th International Conference on Machine Learning, pages 8491-8501. PMLR, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.548, + 0.907, + 0.59 + ], + "angle": 0, + "content": "[105] Axel Sauer, Kashyap Chitta, Jens Müller, and Andreas Geiger. Projected gans converge faster. In Advances in Neural Information Processing Systems (NeurIPS), 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.591, + 0.907, + 0.645 + ], + "angle": 0, + "content": "[106] Axel Sauer, Katja Schwarz, and Andreas Geiger. Styleganxl: Scaling stylegan to large diverse datasets. In ACM SIGGRAPH 2022 conference proceedings, pages 1-10, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.648, + 0.907, + 0.717 + ], + "angle": 0, + "content": "[107] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021. 2, 4, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.719, + 0.907, + 0.843 + ], + "angle": 0, + "content": "[108] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade W Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortman, Patrick Schramowski, Srivatsa R Kundurthy, Katherine Crowson, Ludwig Schmidt, Robert Kaczmarczyk, and Jenia Jitsev. LAION-5b: An open large-scale dataset for training next generation image-text models. In Thirty-sixth Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2022. 2, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.907, + 0.899 + ], + "angle": 0, + "content": "[109] Vikash Sehwag, Mung Chiang, and Prateek Mittal. Ssd: A unified framework for self-supervised outlier detection. In International Conference on Learning Representations, 2020. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8256" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.148 + ], + "angle": 0, + "content": "[110] Arseniy Shakhmatov, Anton Razzhigaev, Aleksandr Nikolich, Vladimir Arkhipkin, Igor Pavlov, Andrey Kuznetsov, and Denis Dimitrov. Kandinsky 2.2. https://github.com/ai-forever/Kandinsky-2, 2023.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.485, + 0.205 + ], + "angle": 0, + "content": "[111] Dasara Shullani, Marco Fontani, Massimo Iuliani, Omar Al Shaya, and Alessandro Piva. Vision: a video and image dataset for source identification. EURASIP Journal on Information Security, 2017:1-16, 2017. 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.207, + 0.484, + 0.261 + ], + "angle": 0, + "content": "[112] Ivan Skorokhodov, Grigorii Sotnikov, and Mohamed Elhoseiny. Aligning latent and image spaces to connect the unconnectable. arXiv preprint arXiv:2104.06954, 2021. 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.264, + 0.484, + 0.332 + ], + "angle": 0, + "content": "[113] Irene Solaiman, Miles Brundage, Jack Clark, Amanda Askell, Ariel Herbert-Voss, Jeff Wu, Alec Radford, Gretchen Krueger, Jong Wook Kim, Sarah Kreps, et al. Release strategies and the social impacts of language models. arXiv preprint arXiv:1908.09203, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.334, + 0.484, + 0.389 + ], + "angle": 0, + "content": "[114] Yang Song, Prafulla Dhariwal, Mark Chen, and Ilya Sutskever. Consistency models. In Proceedings of the 40th International Conference on Machine Learning, pages 32211-32252, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.391, + 0.484, + 0.461 + ], + "angle": 0, + "content": "[115] Yuhta Takida, Masaaki Imaizumi, Takashi Shibuya, Chieh-Hsin Lai, Toshimitsu Uesaka, Naoki Murata, and Yuki Mitsufuji. SAN: Inducing metrizability of GAN with discriminative normalized linear layer. In The Twelfth International Conference on Learning Representations, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.462, + 0.484, + 0.544 + ], + "angle": 0, + "content": "[116] Chuangchuang Tan, Yao Zhao, Shikui Wei, Guanghua Gu, and Yunchao Wei. Learning on gradients: Generalized artifacts representation for gan-generated images detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12105-12114, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.547, + 0.484, + 0.63 + ], + "angle": 0, + "content": "[117] Chuangchuang Tan, Yao Zhao, Shikui Wei, Guanghua Gu, Ping Liu, and Yunchao Wei. Rethinking the up-sampling operations in cnn-based generative network for generalizable deepfake detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 28130-28139, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.632, + 0.484, + 0.701 + ], + "angle": 0, + "content": "[118] Ming Tao, Hao Tang, Fei Wu, Xiao-Yuan Jing, Bing-Kun Bao, and Changsheng Xu. Df-gan: A simple and effective baseline for text-to-image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16515-16525, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.702, + 0.484, + 0.77 + ], + "angle": 0, + "content": "[119] Ming Tao, Bing-Kun Bao, Hao Tang, and Changsheng Xu. Galip: Generative adversarial clips for text-to-image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14214-14223, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.772, + 0.439, + 0.786 + ], + "angle": 0, + "content": "[120] DeciAI Research Team. Decidiffusion 2.0, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.788, + 0.484, + 0.856 + ], + "angle": 0, + "content": "[121] Adaku Uchendu, Thai Le, Kai Shu, and Dongwon Lee. Authorship attribution for neural text generation. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 8384–8395, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.859, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[122] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017. 8" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.908, + 0.176 + ], + "angle": 0, + "content": "[123] Patrick von Platen, Suraj Patil, Anton Lozhkov, Pedro Cuenca, Nathan Lambert, Kashif Rasul, Mishig Davaadorj, Dhruv Nair, Sayak Paul, William Berman, Yiyi Xu, Steven Liu, and Thomas Wolf. Diffusers: State-of-the-art diffusion models. https://github.com/huggingface/ diffusers, 2022. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.178, + 0.908, + 0.247 + ], + "angle": 0, + "content": "[124] Apoorv Vyas, Nataraj Jammalamadaka, Xia Zhu, Dipankar Das, Bharat Kaul, and Theodore L. Willke. Out-of-distribution detection using an ensemble of self-supervised leave-out classifiers. In Proceedings of the European Conference on Computer Vision (ECCV), 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.248, + 0.907, + 0.317 + ], + "angle": 0, + "content": "[125] Sheng-Yu Wang, Oliver Wang, Richard Zhang, Andrew Owens, and Alexei A. Efros. Cnn-generated images are surprisingly easy to spot... for now. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 1, 2, 3, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.318, + 0.908, + 0.346 + ], + "angle": 0, + "content": "[126] Ross Wightman. Pytorch image models. https://github.com/huggingface/pytorch-image-models, 2019.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.347, + 0.907, + 0.374 + ], + "angle": 0, + "content": "[127] Wukong. https://xihe.mindspore.cn/modelzoo/wukong, 2022.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.375, + 0.908, + 0.444 + ], + "angle": 0, + "content": "[128] Zhisheng Xiao, Qing Yan, and Yali Amit. Likelihood regret: An out-of-distribution detection score for variational autoencoder. In Advances in Neural Information Processing Systems, pages 20685-20696. Curran Associates, Inc., 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.446, + 0.908, + 0.502 + ], + "angle": 0, + "content": "[129] Fisher Yu, Yinda Zhang, Shuran Song, Ari Seff, and Jianxiong Xiao. Lsun: Construction of a large-scale image dataset using deep learning with humans in the loop. arXiv preprint arXiv:1506.03365, 2015. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.503, + 0.908, + 0.572 + ], + "angle": 0, + "content": "[130] Bowen Zhang, Shuyang Gu, Bo Zhang, Jianmin Bao, Dong Chen, Fang Wen, Yong Wang, and Baining Guo. Styleswin: Transformer-based gan for high-resolution image generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11304-11314, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.573, + 0.908, + 0.629 + ], + "angle": 0, + "content": "[131] Xu Zhang, Svebor Karaman, and Shih-Fu Chang. Detecting and simulating artifacts in gan fake images. In 2019 IEEE international workshop on information forensics and security (WIFS), pages 1-6. IEEE, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.63, + 0.908, + 0.686 + ], + "angle": 0, + "content": "[132] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In Computer Vision (ICCV), 2017 IEEE International Conference on, 2017. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.687, + 0.908, + 0.769 + ], + "angle": 0, + "content": "[133] Mingjian Zhu, Hanting Chen, Qiangyu YAN, Xudong Huang, Guanyu Lin, Wei Li, Zhijun Tu, Hailin Hu, Jie Hu, and Yunhe Wang. Genimage: A million-scale benchmark for detecting ai-generated image. In Advances in Neural Information Processing Systems Dataset and Benchmarks Track, pages 77771-77782, 2023. 2, 3, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.77, + 0.907, + 0.827 + ], + "angle": 0, + "content": "[134] Bojia Zi, Minghao Chang, Jingjing Chen, Xingjun Ma, and Yu-Gang Jiang. Wilddeepfake: A challenging real-world dataset for deepfake detection. In Proceedings of the 28th ACM international conference on multimedia, 2020. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.908, + 0.827 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8257" + } + ] +] \ No newline at end of file diff --git a/2025/Community Forensics_ Using Thousands of Generators to Train Fake Image Detectors/f9d7ae16-50eb-4d82-828a-d6dd35d593a9_origin.pdf b/2025/Community Forensics_ Using Thousands of Generators to Train Fake Image Detectors/f9d7ae16-50eb-4d82-828a-d6dd35d593a9_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..0231cb8f1cd0a387d87ca7358024843480ca922a --- /dev/null +++ b/2025/Community Forensics_ Using Thousands of Generators to Train Fake Image Detectors/f9d7ae16-50eb-4d82-828a-d6dd35d593a9_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:99bb0de0be752a489ad1472d126bb406d14ed2d0527fc3da832eb7a97f159368 +size 3237595 diff --git a/2025/Community Forensics_ Using Thousands of Generators to Train Fake Image Detectors/full.md b/2025/Community Forensics_ Using Thousands of Generators to Train Fake Image Detectors/full.md new file mode 100644 index 0000000000000000000000000000000000000000..c35a60614f373a5baf97569d5ea72654a6191ea9 --- /dev/null +++ b/2025/Community Forensics_ Using Thousands of Generators to Train Fake Image Detectors/full.md @@ -0,0 +1,432 @@ +# Community Forensics: Using Thousands of Generators to Train Fake Image Detectors + +Jeongsoo Park Andrew Owens University of Michigan + +# Abstract + +One of the key challenges of detecting AI-generated images is spotting images that have been created by previously unseen generative models. We argue that the limited diversity of the training data is a major obstacle to addressing this problem, and we propose a new dataset that is significantly larger and more diverse than prior works. As part of creating this dataset, we systematically download thousands of text-to-image latent diffusion models and sample images from them. We also collect images from dozens of popular open source and commercial models. The resulting dataset contains 2.7M images that have been sampled from 4803 different models. These images collectively capture a wide range of scene content, generator architectures, and image processing settings. Using this dataset, we study the generalization abilities of fake image detectors. Our experiments suggest that detection performance improves as the number of models in the training set increases, even when these models have similar architectures. We also find that increasing the diversity of the models improves detection performance, and that our trained detectors generalize better than those trained on other datasets. The dataset can be found in + +https://jespark.net/projects/2024/community_forensics + +# 1. Introduction + +Our ability to automatically generate realistic images is quickly outpacing our ability to detect them, potentially leading to a state of affairs in which neither humans nor machines can reliably tell real from fake. While the field of image forensics has been developing methods to address this problem, existing fake image detectors still struggle with generalization. These methods often excel at detecting images from generators that were present in their training sets, but fail when given images sampled from unseen models [87, 101, 102, 125]. + +A core challenge is dealing with the large amounts of variation between models. Each generator has a potentially unique combination of the architecture, loss function, and training distribution. Even seemingly minor differences in + +![](images/653df4c788aa9892c3e27775054108cb9797cf24bedc786da1a1ff5df83470f9.jpg) +Figure 1. Performance vs. model diversity. We use images sampled from different numbers of open source latent diffusion models in our Community Forensics dataset to train fake image detectors (shown in Fig. 2a). We fix the total number of images and only vary the number of models. The detector's performance increases across all generator types as we train from more models, even though these added models are entirely latent diffusion. This improvement is largest for test images from out-of-distribution generators, such as pixel-based diffusion models or GANs. + +low-level image processing details, such as the ways that training images are resized or compressed, can strongly influence detection accuracy [125]. As a result of these model-specific idiosyncrasies, a generator's images may evade detection, even when images from architecturally similar models exist in the training set. This issue has been exacerbated by the thousands of open source models that are now available online, many of which extend pretrained base models in complex ways. + +We hypothesize that the lack of diversity in training datasets is a major source of these shortcomings. Although today's datasets often contain millions of fake images, they come from a relatively small number of generators. As a result, this data fails to capture many sources of variation that one might encounter in the wild. These limitations also make it challenging to accurately benchmark performance, since it is easy for cues that work well on one set of generators to fail on others. + +![](images/d1f5072360d364ca47d054567c71f9c5839673cf0b8b2d63a1aa62a5e415580c.jpg) +Dreamshaper v6 + +![](images/6fe9aa0173c01026b84b59881178e13aa1eb95ccb09bf03e14da648a3e1b7ce9.jpg) + +![](images/f18b5a9ce3439f4e81c3684d1eb232b9139713186c8cf0fbac687c95099c022a.jpg) +Plat diffusion + +![](images/5fce98152588acb7455cc5f2fa8bc91cfea7bb6956a8c43e61e19ce49c1c93c8.jpg) +SoteMix + +![](images/d1ddd487777710b7b38dcd33a26da7bfbc50be43bf1bcb1148ba2820d91ca332.jpg) + +![](images/51c233be4c773bc38d072851e63c7202864285b451288867fda884ec0b8f9606.jpg) +Technoalbum + +Nightvisionxl 0791 + +![](images/2ebea408eef862ea73fa5d22b6093a1f7f0d775f83e9386a3ef0bae503afa177.jpg) + +Nightvisionxl 0791 + +Albedobase x1 v21 + +![](images/4326e416ed416ee6148740a80fae62fc0e38844e93d1775e1ce55d502cd1365c.jpg) +Nightvisionxl 0791 + +![](images/3ba0c90a0452fa25ad27528355032421adeb240c5c1ccdd172d3d1d27736ef00.jpg) +Enasswhidaly +(a) Systematically collected diffusion models (4763 models) +Figure 2. The Community Forensics dataset. Our dataset contains images sampled from three types of generative models. (a) We systematically download open-source latent diffusion models from a model-sharing community [34, 123]. (b) We select popular open source generators with a variety of architectures and training procedures. (c) We sample from both closed and open state-of-the-art commercial models. We present example images and their corresponding model names. + +Albedobase x1 v21 + +![](images/c0ea5d2a9ad257f237e85f37379d6059719c942552d2227be09afbb897f5d1a4.jpg) + +![](images/bcd1b05f51e19ae663ab794f3f071b619c39102a82daab31320e9ca1bebf17e0.jpg) +Msart 541 + +![](images/c1f0500db9b788ec7a905b477645e3d86b575ff187a21068ba19fec9ad579572.jpg) +91 +First avarfar + +![](images/914cb7b7cd514259ed7b6e25c49761a1b827a8fc3a1aa3885ae3d7d40d418705.jpg) + +![](images/fa4a628fee64f9e2493ca5a29c7054ea0efa00da4dbce0c165ebf526e61f7b2a.jpg) +AnalogMadness real +DucHaitenAnimated + +![](images/b9218cfe77d7f7d2a088456a4948bad36fa0cf4db97cbdfc86138418761ef07d.jpg) + +![](images/e08a81e17a58d551f93f4cd9eb9bf9ece3a23e23c149da6b6c1e1f665a7316db.jpg) +A can thr + +![](images/64d1bc6de9df69c3b6489c266458244477b6c5581016a61fc98459b86bef746a.jpg) +DeepFloyd +StyleGANXL + +![](images/52f32b7d1105a788025b0416e2c1cf562640a87f9ffd3bd5258f42ec65f18f67.jpg) +VQDiffusion + +![](images/e7b8aeb2ec6a6ce68a2797374b002d91d0ea47aa5cb3f6258777ff4b876b365b.jpg) +GuidedDiffusion +(b) Manually chosen open source models (19 models) + +![](images/da4130d5200db2a94b4074279315b43279828e5da2a2bf30dcad9390b1ce0eb9.jpg) +StyleGAN2 + +![](images/d7f0d9b425fba9ca111eb271725ffab93f8510480c37387aea1ad9086ba63d4b.jpg) +Glide + +![](images/fc175043f62aa72dbea529adae15d4b9ab30ec1ed9675b72e654e175288e9064.jpg) +DiT + +![](images/410392cc44f32b435cf1546850125a2658df3053f13c091bc89b1f4bb159ff87.jpg) +CIPS + +![](images/11b1f9cf60b78ad52a794a446d7ac4f29f65ef9e149766a7093a8363e338a34e.jpg) +Dalle2 + +![](images/475dbce8cf8f39c96856c26a14e28cb4c347ad99ebcf60710ed8a06066669514.jpg) +FLUX dev +(c) Commercial models (11 models) + +![](images/ad7c7fedf172621ef758bba396770ceb5b4b15d01d989f8384f8b4329781a5c3.jpg) +MidjourneyV6: + +![](images/ef5a72b79b986288854273ed87a7dec6f25c0515ffaa242cdc9adb4f839165da.jpg) +Firefly Image2 + +![](images/aa991346df578eeb395ee35990d987960343854f61dc320668f545a73e0cd742.jpg) +MidjourneyV5 2 + +![](images/bbccfa2ca88d61d08de5831fc11c6f7463100e834f77a4a353170e05ff30e38c.jpg) +Firefly Image3 + +![](images/0b8ea5970c68a34b9f7ad0ed8a746394eaf18029cfac47be0388eefe418a9888.jpg) +Dalle3 + +![](images/bf07435f7a74b9f29f483914d3cb78a88d28f7bbddd0c65b23e2c9f1c281c58c.jpg) +IdeogramV1 + +To address these problems, we propose Community Forensics, a dataset that is significantly more diverse and comprehensive than those in prior works (Fig. 2). Our dataset contains images generated by: (a) thousands of systematically downloaded open-source latent diffusion models, (b) hand-selected open source models with various architectures, and (c) state-of-the-art commercial models. + +To acquire large numbers of models, we sample images from thousands of text-to-image diffusion models hosted on a popular model-sharing website, Hugging Face [34]. We exploit the fact that these models use a common programming library [123] and thus can be sampled in a standardized way. A large fraction of them are extensions of Stable Diffusion [100], but collectively capture a variety of common model variations, such as in the architecture, image processing, and image content. We also sample images from many other open source models, including GANs [37], autoregressive models [39], and consistency models [79, 114]. To ensure sufficient diversity in the image content of the generated images, we sample generators using captions sourced from a mix of several existing datasets whenever possible. + +Our dataset contains 4803 distinct models, roughly $34 \times$ more than the most extensive previous dataset that sample images from generative models [5, 7, 15, 32, 43, 87, 125, 133], and covers a variety of recent model designs (Fig. 2). + +We use this dataset to study generalization in the generated image detection problem. Our experiments confirm + +that increasing the diversity of data improves generalization, consistent with previous findings in other areas of computer vision [48, 107, 108]. In contrast to other work, our dataset aims to increase diversity via number of generators, an underexplored axis of diversity. Through experiments, we find: + +- Classifiers trained on our dataset obtain strong performance, both on our newly proposed evaluations and on multiple previously-proposed benchmarks. +- Adding more generative models improves generalization. Fig. 1 demonstrates the performance of fake image detection when trained on samples from varying numbers of diffusion models. Notably, the performance improves as more models are added, even across different architectures. +- Including diverse generative model architectures significantly improves results, since classifiers do not fully generalize between generator architectures. +- Standard classifiers perform well. In contrast to observations from recent work, we find that end-to-end training of classifiers based on CNNs or ViTs generalizes well, with qualitatively similar to that of other recognition problems. + +# 2. Related work + +Datasets for detecting generated images. A number of datasets have been proposed for specifically detecting "deepfake" images containing manipulated faces [25, 63, 67, 71, 101, 102, 134]. Rather than focusing on face manipulation, we address creating general-purpose meth + +ods that can detect images that have been directly produced by generative models. Wang et al. [125] proposed a widely-used dataset of CNN-generated images, mixing images from GANs [8, 12, 58, 59, 90, 132] with other models [9, 10, 18, 70, 102]. This work showed that forensics models generalize between generative models, providing motivation for training on large datasets of diverse generators. However, their classifier was trained on images from a single GAN and was highly sensitive to data augmentation parameters, and more recent work shows that it does not generalize to newer models [14, 87]. Ojha et al. [87] introduced a dataset of recent diffusion models and found that training a linear classifier on CLIP features [96] extracted from ProGAN-generated images performed well. Cozzolino et al. [15] extend this work by studying the performance of CLIP-based detectors on various generative models and datasets. Epstein et al. [32] simulated detecting fake images in an online way by training a detector up to a certain year and testing it on generators released after that year. Zhu et al. [133] collected 1.4M generated images from 8 different generators. These datasets, however, only consider a handful of models (less than 20 each), limiting the generalization of their detectors. Asnani et al. proposed RED116 [5], collecting 116K images from 116 generative models for predicting the model hyperparameters of a given generated image. Guo et al. [43] extends RED116 to 140 generative models with 140K images total. We improve upon these works by collecting much more diverse generative models to improve the performance and generalization of the detector. In concurrent work, Hong et al. [49] acquires user-created images from Midjourney and CivitAI. This strategy is complementary to ours: while it aims to collect in-the-wild fake images, its distribution is centered on images that users share, and the models are not necessarily identifiable, making it challenging to rigorously analyze the dataset's contents and to interpret experiments conducted on it. + +Fingerprint-based image forensics methods. Classic work on image forensics relied on methods based on image statistics [95] and physical constraints [56], rather than learning. A number of datasets have been created for detecting images that have been manipulated using traditional methods, such as with photo editors [21, 26, 53, 64, 84]. Recent works focus on detecting synthetic images by inspecting the generator fingerprints. Zhang et al. [131] and Marra et al. [80] proposed identifying the spatial fingerprints left by the generator to detect synthetic images. Others focus on spectral anomalies to detect synthetic images. Durall et al. [29] and Dzanic et al. [30] identified that CNN-generated images fail to reproduce certain spectral properties of real images. Corvi et al. [14] study the frequency fingerprints of the generated images and analyzes the cross-architecture generalization of the detector. Bammey [7] uses high-frequency artifacts to detect generated images. However, these approaches may be + +brittle since the artifacts they rely on can be eliminated by post-processing [15]. We instead approach this problem in a data-driven manner, scaling the number of models, images, and architectures. Recent work has created ensembles of fake image classifiers [50]. In parallel, researchers have detected text generated by language models using supervised learning and heuristics [6, 35, 55, 65, 82, 103, 113, 121], which closely resemble those in visual forensics. However, no existing techniques that we are aware of aim to collect comprehensive datasets of community-created generators. + +Out-of-distribution generalization. Our work is related to the out-of-distribution recognition problem as it involves generalizing to unseen generators and image processing pipelines. A variety of approaches have been proposed for this problem, based on likelihood ratios [72, 98, 128], self-supervision [47, 83, 109, 124], internal model statistics [46, 104], temperature scaling [3, 73], and via energy-based models [28, 31, 75]. Works by Schuhmann et al. [108] and Hendrycks et al. [48] show that diverse training data and data augmentation are important to improving the robustness to out-of-distribution samples. Our results are in line with these conclusions, as we find that a diverse set of generative models and stronger augmentations improve generalization. + +# 3. The Community Forensics Dataset + +To support our goal of studying generalization in generated image detection, we collect a dataset of images sampled from a wide range of models (Fig. 2). Our dataset consists of: (a) a large and systematically collected set of "in-the-wild" text-to-image latent diffusion models obtained from a model-sharing website, (b) hand-selected models from other open source architectures, and (c) closed and open state-of-the-art commercial models. We also pair these generated images with real images from other datasets. For all images in our dataset, we preserve the original image format where possible, without any additional compression or resizing. This is to mitigate potential bias and performance degradation in out-of-distribution settings due to unwanted artifacts [40, 42, 97]. Our dataset contains significantly more models than previous works (Tab. 1) and spans a wider range of architectures, processing pipelines, and semantic content. + +# 3.1. Systematically collecting generative models + +We perform our systematic collection using publicly available, open source models that use the Hugging Face diffusers library [34, 123] because: 1) it is a popular library for creating text-to-image models and is widely used by hobbyists, 2) thousands of such models are publicly indexed, and 3) it provides a standard interface by which we can sample images. We process the models in the order of popularity, as indicated by the number of downloads. Our + +
DatasetModelsImagesArchitecturesTraining setup
Wang et al. [125]11362KGAN, Perceptual, Deepfake, ...ProGAN [58] vs. LSUN [129]
Ojha et al. [87]4*10K*GAN, Perceptual, Diffusion, ...ProGAN [58] vs. LSUN [129]
Epstein et al. [32]14570KDiffusionDiffusion vs. LAION [108]
Cozzolino et al. [15]1826KDiffusionLDM [100] vs. MS-COCO [74]
Synthbuster [7]910KDiffusionDiffusion vs. Dresden [36]
GenImage [133]81.4MDiffusion, GANDiffusion, GAN vs. ImageNet [23]
RED116 [5]116116KGAN, VAE, Autoregressive, ...Many vs. Many
RED140 [43]140140KDiffusion, GAN, VAE, ...Many vs. Many
Ours48032.7MDiffusion, GAN, Autoregressive, ...Many vs. Many
+ +Table 1. Comparison with existing forensics datasets. We compare the size of the dataset with existing datasets containing identifiable generative models. We only count the number of generated images. Our dataset contains significantly more generative models than prior works. $\ast$ : Only counting the unique evaluation set by Ojha et al. [87] as their dataset is based on Wang et al. [125]. + +pipeline downloads each model and extracts relevant hyperparameters (e.g., number of diffusion steps and guidance scale), sampling pipeline configurations, and metadata, from both the model cards on the model-sharing webpage and the metadata retrieved by the diffusers library [34, 123]. We sample images using a distribution of text prompts obtained from various real datasets (Sec. 3.3). Since experiments suggest diminishing returns for repeatedly sampling from any given model, we sample a few hundred images from each one. We sample 4763 models with approximately 403 images each, for a total of $1.9\mathrm{M}$ images from this process. + +While the lack of documentation in each model and the scale of data collection make it challenging to exactly characterize the model designs in this set, they appear to be entirely (or almost entirely) based on latent diffusion. More specifically, we categorize models as being based on latent diffusion if they perform a denoising process on a latent representation. Based on this criterion and the self-reported tags, all models in our systematically collected set appear to be based on latent diffusion. While pixel-based diffusion models also use the diffusers library (e.g., DeepFloyd [22]), they were incompatible with our automated generation pipeline. We record such incompatible models and manually sample them to either construct an out-of-distribution test set (Sec. 3.4), or as manually-chosen models for training data (Sec. 3.2). + +We show examples of the sampled images in Fig. 2. In the supplement, we provide examples of models and information from their project pages. These models generate a variety of images, with various types of semantic content and preprocessing. For example, a large fraction of these models adapt variations of a popular pretrained latent diffusion model, Stable Diffusion [100], to different downstream applications, and use a number of adaptation strategies (e.g., using LoRA [51]). We provide the model metadata with each image to enable other possible forensics applications. We discuss these in the supplement and provide information about image and model licenses. + +# 3.2. Collecting images from other architectures + +Images from manually chosen models. To ensure that our dataset contains a broader range of models, we manually select 19 models from public repositories and sample an average of 40,738 images per model, resulting in 774K images total. We note that this number is itself on par with (or more than) many prior datasets with identifiable generative models [7, 15, 32, 87, 125, 133]. We include several GANs (e.g., StyleGANs [60-62, 106], BigGAN [8], StyleSwin [130], GigaGAN [57], ProGAN [58], ProjectedGAN [105], GANformer [52], SAN [115], and CIPS [4]), pixel-based diffusion models (e.g., GLIDE [85], ADM [24], and DeepFloyd [22]), latent diffusion models (e.g., VQ-Diffusion [41], Diffusion Transformers [92], and Latent Flow Matching [20]), and an autoregressive model (Taming Transformers [33]). + +Images from commercial models. We sample 15K images from 11 commercial models using LAION-based captions to evaluate the generalization to state-of-the-art models with typically unknown architectures: DALL-E 2, 3 [88, 89], Ideogram V1, V2 [2], Midjourney V5, V6 [81], Firefly Image 2, 3 [1], FLUX.1-dev, Schnell [68], and Imagen 3 [38]. + +# 3.3. Collecting real images + +To help study how real images influence forensics models, we source real images from a variety of existing datasets: LAION [107], ImageNet [23], COCO [74], FFHQ [59], CelebA [76], MetFaces [60], AFHQ [13], Forchheim [44], IMD2020 [86], Landscapes HQ [112], and VISION [111].3 + +# 3.4. Curating the evaluation set + +We construct our evaluation set using the incompatible models from our automated sampling pipeline, commercial models (Sec. 3.2), and manually collected open source models. The evaluation set comprises 26K images sampled from 21 models not included in the training set. + +This includes our commercial models set and an additional 11K images from 10 models: Deci Diffusion V2 [120], GALIP [119], KandinskyV2.2 [110], Kvikontent [66], LCM-LoRA-SDv1.5, LCM-LoRA-SDXL, LCM-LoRA-SSD1B [79], Stable Cascade [93], DF-GAN [118], and HDiT [16], sampled using RAISE [19], ImageNet [23], FFHQ [59], and COCO [74]-based captions. + +The generated images are paired with the source real data that are used to prompt the generators. However, since some of the real datasets do not have appropriate licenses for redistribution (e.g., LAION [107, 108]), we created a public version of our evaluation set by pairing the generated images with openly licensed COCO [74] and FFHQ [59] which allow redistribution for non-commercial purposes. The public version of our evaluation set will serve as an easily reproducible and shareable evaluation set that will complement our default set. We will refer to our default set as the comprehensive evaluation set. We will also release the instructions to reconstruct our comprehensive set. However, note that it may not be possible to exactly reconstruct this set in the future due to link rot. + +# 3.5. Generating images + +Unconditional models are sampled until we reach the desired number of images. For class conditional models, we sample an equal number of images per class. Text-conditional models are sampled using captions obtained from real images (Sec. 3.3). We either use captions that are already present in the dataset or use BLIP [69] to generate them. Some models such as GigaGAN [57] and HDiT [16] do not provide a pretrained model, so we instead use their pre-generated images. Generated images are saved in PNG format to avoid compression artifacts. However, Firefly [1] generated images are saved in JPEG format as their web UI only allows downloading in JPEG. + +# 4. Experiments + +We use our dataset to conduct a study of generalization in visual forensics, asking a number of questions: (1) How well do forensics models trained on our dataset generalize to unseen models? (2) Does adding more models improve detection performance? (3) How does diversity of the training data affect performance? (4) What architectures and data augmentation schemes are most successful? + +# 4.1. Training image forensics models + +We train binary classifiers that detect generated images using our dataset to study the generalization in image forensics. We construct our training set of 5.4M images by pairing 2.7M generated images with 2.7M real images. + +Training and evaluation setup. We evaluate the models trained on our dataset and compare them with prior works [43, 87, 125, 133]. As RED140 [43] already contains RED116 [5], we do not compare with RED116 dataset. + +Following prior works [87, 125], we use the threshold-independent mean average precision (mAP) and accuracy (Acc.) as our evaluation metrics. We compute the mAP and accuracy by averaging the results of each generative model. We use five evaluation sets: Wang et al. [125], Ojha et al. [87], Synthbuster [7], GenImage [133], and our evaluation set. All evaluation sets apart from GenImage [133] evaluate out-of-distribution performance for all classifiers. GenImage [133] evaluation set, however, contains the same set of generators used in training, and is an in-distribution evaluation set for their classifiers. Concretely, the evaluation set by Wang et al. [125] and Ojha et al. [87, 125] contains models such as CRN [10], CycleGAN [132], DALL-E [88], DeepFake [25], IMLE [70], SAN [18], StarGAN [12], and SITD [9] which are unseen by both their and our classifiers. Synthbuster [7] evaluation set is comprised of RAISE [19]-based synthetic images of DALL-E [88, 89], Firefly [1], Glide [85], Midjourney [81], and Stable Diffusion [94, 100], and is mostly out of distribution for all classifiers. GenImage [133] evaluation set is a validation split of their training set; they use an identical set of models to train their classifier: ADM [24], BigGAN [8], Glide [85], Midjourney [81], Stable Diffusion [100], VQ-Diffusion [41], and Wukong [127]. + +Model architecture. Building on prior works which mainly used CLIP-ViT [27, 54, 96] and ResNet-50 [45], we consider ViT [27] and ConvNeXt [77] pretrained models for our classifiers. We use a plain ViT-S backbone [27] pretrained on CLIP objective [54, 96] using LAION-2B [108], ImageNet 21K, and ImageNet 1K datasets [23]. We also experiment with a ConvNeXt-S model [77] pretrained on ImageNet 21K and ImageNet 1K datasets [23]. We replace the classification head with a linear layer with sigmoid activation that outputs the probability of the image being generated. Unlike prior works [15, 87] that freeze the CLIP-ViT backbone, we train the backbone end-to-end. The models are obtained through timm [11, 126] library on Hugging Face. We experiment with two input resolutions, $224^2$ and $384^2$ , to evaluate the impact of the input resolution on the detector's performance. We denote the detector with $384^2$ input resolution as High res. We implement the models using PyTorch [91]. The hyperparameters are detailed in the supplement. + +Data augmentation. Prior work considered augmentations that were designed to simulate postprocessing, such as flipping, cropping, Gaussian blur, and JPEG recompression to train their detectors [7, 15, 87, 125]. We propose an augmentation scheme that extends this approach and compare it with previously proposed augmentation methods. We expand the set of augmentations to handle additional transformations that can occur in the wild, such as padding, resizing, rotation, and shear, and integrate them into a framework that can apply complex sequences of transformations. We introduce a modified version of RandAugment [17] that applies a + +
ModelEvaluation Set (mAP)Evaluation Set (Acc)
Wang et al. [125]Ojha et al. [87]SB [7]GenImage [133]OursMeanWang et al. [125]Ojha et al. [87]SB [7]GenImage [133]OursMean
Wang et al. [125]0.8970.6960.5160.6420.5370.6000.6480.7140.5270.5080.5330.5130.5170.552
Ojha et al. [87]0.9390.9570.6200.7970.5920.6560.7600.7910.8210.5320.6410.5400.5480.646
GenImage [133]0.9290.9840.8130.9990.9120.9680.9340.7950.9660.7190.9900.8180.8860.862
RED140 [43] - High res.0.9000.9540.7650.9270.7640.8610.8620.6940.7800.5580.6740.5620.5650.639
Ours0.9640.9910.9040.9900.9710.9770.9660.8730.9500.8180.9460.8610.8880.889
Ours - High res.0.9670.9960.9740.9980.9870.9940.9860.9010.9700.9080.9570.8920.9120.923
+ +Table 2. Generalization of AI-generated image detectors across benchmarks. We evaluate the classifiers trained on our dataset on several benchmarks, including our own. We also evaluate several previously released classifiers. Our Comprehensive set (abbreviated as Comp.) pairs the generated images with original real data; the Public set pairs them with openly licensed COCO [74] and FFHQ [59] for license-compliant redistribution of the evaluation set (Sec. 3.4). We use plain CLIP-ViT-S [27, 54, 96] architecture with $224^{2}$ and $384^{2}$ (High res.) input resolutions, Wang et al. [125] and GenImage [133] use ResNet-50 [45] with $224^{2}$ input resolution, and Ojha et al. [87] uses CLIP-ViT-L with $224^{2}$ input resolution as the backbone. Our classifiers show robust performance across all evaluation sets, outperforming all baselines in out-of-distribution settings ([7, 87, 125] and Ours) and nearly matches GenImage [133] on its in-distribution evaluation set. $\spadesuit$ : RED140 [43] is trained following our training procedure with $384^{2}$ input resolution as they do not provide a trained classifier. + +![](images/962707b7f8c50d48ce8f78f437e2c69577554671c1e94e0602d66f8aeaec0aeb.jpg) +Figure 3. Performance across generator types. We evaluate the classifier performance across five generator types - latent diffusion, pixel diffusion, commercial models, GANs, and other architecture (Stable Cascade [93]), and report the mean. We additionally evaluate three forensics methods, Aerblade [99], LGrad [116], and NPR [117]. Our classifiers show robust performance across all generator types, whereas prior works struggle to generalize. + +randomly-ordered sequence of augmentations to the images. Specifically, our modified RandAugment samples a random number $n$ between 0 and $n_{\mathrm{max}}$ for each augmentation type. Then, it applies the augmentations in random order until $n$ augmentations are applied for each augmentation type to the image. We use various augmentations, including in-memory JPEG compression, random resizing with random interpolation methods, cropping, flipping, rotation, translation, shear, padding, and cutout. + +# 4.2. Generalization across benchmarks + +We first evaluate how well classifiers trained on our dataset perform across benchmarks. In Table 2, we observe that our models outperform the prior works [43, 87, 125, 133] on all evaluation sets except GenImage. This is expected since the GenImage evaluation set is a validation split of their training set; all of the generators are already seen by their classifier. On all other unseen benchmarks, our classifiers outperform all prior works. Notably, our classifiers achieve very high performance (0.987 mAP and $89.2\%$ accuracy) on our out-of-distribution, comprehensive evaluation set, with a significant margin over prior works. This gap in performance can be traced to our training data which in + +![](images/b2050c6c5265689e96c9dc3268ab08e1cb231abb9c6ddef646fe10e377ee3cc7.jpg) +Figure 4. Performance with increasing number of images. We train a classifier with varying numbers of images from two sets: 1000 randomly chosen models and 10 popular (highly downloaded) models in the systematically collected subset. We report the mean and standard error bands for each data point across 4 randomly sampled subsets. The classifier trained from 1000 random models outperforms 10 popular models in all cases. Notably, the accuracy gap is wider than that of mAP, which may suggest that having a diverse set of models improves accuracy threshold calibration. + +![](images/71f687f50ab6ed2107f86a86ca9738dc0eeb779d3afd5d7e956d3ee90cb40b2a.jpg) + +corporates a substantially richer variety of generators than existing works. Consequently, our classifiers demonstrate robust generalization to out-of-distribution data, where prior works often struggle. + +To better illustrate the generalization of the classifiers, in Figure 3, we split our comprehensive evaluation set into five subsets and evaluate them: latent diffusion, pixel diffusion, commercial models, GANs, and other architecture type (Stable Cascade [93]). Furthermore, we evaluate three additional forensics methods: Aerblade [99], LGrad [116], and NPR [117]. Our classifiers show strong performance across all generator types, unlike prior works which struggle to generalize. For the following experiments, we use our best-performing model (High res.) unless stated otherwise. + +# 4.3. Impact of model diversity + +Next, we examine the impact of the number of models in training data. We train classifiers with images sampled from 3 to 3333 generators and evaluate them (Fig. 1). To ensure that the gains are not due to simply sampling qualita + +![](images/de3a685a1acf23bf0081bb9d302fdcdecba91a2e02a740ea8299ba0c369f2745.jpg) +(a) Impact of generator type diversity + +![](images/86a32a5591270b141562b2d967f9c61148be4d1ec7231ff9414b918c18f904f7.jpg) +(b) Classifier backbone comparison +Figure 5. (a) Performance and model diversity. We compare detection performance for commercial models using classifiers trained on different subsets of the dataset: the systematically collected latent diffusion models, the manually chosen models containing diverse generator types, and both. As diversity increases, so does performance. (b) Classifier backbone comparison. We compare the architectures across datasets: ours, GenImage [133], and Wang et al. [125]. Performance is similar between architectures. + +tively different architectures, we only use our systematically collected latent diffusion models. For each data point, we sample 10 random subsets of models with 100K training images each and report the mean and standard error bands. We use an extended evaluation set that includes non-latent diffusion generators from our training set, which allows us to comprehensively assess the generalization capability of the classifiers trained exclusively on latent diffusion models. We find that the performance steadily increases with the number of models. However, the performance begins to flatten out beyond 1000 models, suggesting diminishing returns. Interestingly, the performance also improves on out-of-distribution architectures such as GANs and pixel-based diffusion models, even though the classifier is only trained on latent diffusion models. + +In Figure 4, we vary the number of images from two sets: 1000 randomly chosen models and 10 popular models (as denoted by their number of downloads) downloaded from our systematically collected diffusion models. While the results show that the performance improves with more training images, it begins to plateau at approximately $27\mathrm{K}$ images. Moreover, the classifier trained on 1000 models outperforms the 10 models in all cases, indicating that model diversity is important for strong performance. We also note that the accuracy gap is noticeably wider than that of mAP, which may suggest that model diversity is crucial in calibrating the accuracy thresholds of the classifiers. + +Our experiments show that the performance improvements from increasing the number of models may plateau when they are limited to a single generator type (Fig. 1). In Figure 5a, we show that the diversity of the generator type also plays a major role in generalization. We train classifiers on three different sets of training data: our systematically collected set, manually chosen set, and a full set consisting of both subsets. The systematic set comprises entirely of latent diffusion models, and the manual set contains numerous generator types, including GANs, latent and pixel-based dif + +![](images/0af5c5242036ada06279f525959cfa606d6065e317152240bfc22a526c1dbddc.jpg) +(a) Impact of frozen backbones +Figure 6. (a) Evaluating frozen backbones. Freezing the pretrained backbone, a common practice in prior works [15, 87], consistently decreases the performance. (b) Analyzing source and generated data alignment. We evaluate how the pairing of the real datasets affects performance. 'R' denotes the real dataset used in training, and 'F' indicates the source dataset used to obtain the captions for prompting the generators. The results suggest that pairing the source data (i.e., real data used to prompt the generators) with the generated images is not essential for performance. + +![](images/c408e5535218ed15163009d68d13d62ad24b504eb97ca65653f5bf9f8f90ab5b.jpg) +(b) Semantic alignment analysis + +fusion, and autoregressive models (Sec. 3.2). The classifier trained on the manual set with more diverse generator types shows similar performance compared to the one trained on the systematic set, despite containing fewer than half the number of images (774K vs. 1.9M). Additionally, we find that the two sets are complementary; the performance is further improved when we train using both sets. + +# 4.4. Analysis of design choices + +We examine the impact of various design choices, including some suggested in earlier works. In particular, we investigate the choice of backbone models, freezing the backbone, semantic alignment between the real and generated data, and robustness to transformations. + +Classifier backbone comparison. We compare the performance of the classifier trained using CLIP-ViT [27, 54, 96] and ConvNeXt [77] backbones following our training procedure in Figure 5b. We examine three datasets: ours, GenImage [133], and Wang et al. [125]. We observe similar performance between architectures across all datasets. + +Frozen backbone. Prior works [15, 87] suggested using a frozen CLIP-ViT backbone for training the classifiers. We study this by training the classifiers with both frozen and unfrozen pretrained backbones, using CLIP-ViT [27, 54, 96] and ConvNeXt [77]. As shown in Fig. 6a, freezing the backbone consistently leads to worse results, indicating that end-to-end training is crucial for high performance. + +Semantic alignment. Existing works often pair the generated images with the source dataset (i.e., the real dataset used to prompt or generate the images) arguing that misaligned data can introduce bias [7, 15, 87, 125]. We test this practice in Fig. 6b by examining the performance with both semantically aligned and misaligned real datasets. Specifically, we consider two real datasets: one comprised exclusively from LAION [107] and another combining ImageNet [23], + +![](images/8a307ef1c5bdc4bcf944ceeb4dc3742fe7c1ec364fe5b71bb387b42936583893.jpg) +Figure 7. Robustness to various transformations. Our classifiers display robust performance across transformations. Other works generally show more sensitivity to these factors. + +MS-COCO [74], LandscapesHQ [112], Forchheim [44], VISION [111], and IMD2020 [86]. We sample our systematically collected latent diffusion models using these two sets and categorize the generated images by their source dataset. The performance differences of all pairs were marginal, suggesting that strict alignment may not be as critical with sufficiently large data. + +Robustness to transformations. Figure 7 illustrates the robustness of the classifiers against various transformations. Following prior works [87, 125], we test robustness to JPEG compression and Gaussian blur. Additionally, we examine robustness to rotation, resizing, padding, and shear, as they commonly occur in real-world scenarios. For padding, we randomly pad the width or height of the image with a given fraction and scale it back to the original size. Similarly, in resize, we randomly upsample or downsample the height and width of the image by a given fraction and resize it back to the original size (e.g., if fraction is 0.3, resize the height and width to $0.7 \times$ or $1.3 \times$ and then scale it back to the original size). The results demonstrate that our models are more robust to transformations than existing models. Specifically, GenImage [133] is notably more sensitive to Gaussian blur, JPEG compression, and resizing artifacts; classifier by Ojha et al. [87] displays sensitivity to shear transforms, and the one by Wang et al. [125] performs poorly overall. + +# 4.5. Other forensics applications + +Our dataset may enable further forensics studies that can take advantage of our diverse array of generators. To illustrate this, in the supplement, we identify the type of generator used to synthesize a given image by using $k$ -nearest neighbor in the feature space of our classifier. + +# 5. Conclusion + +In this paper, we studied the problem of generalizing to unseen generative models in synthetic image detection. We proposed a new dataset, Community Forensics, which contains 4803 models and 2.7M images collected from various public sources. We studied the impact of model diversity and demonstrated that it plays a crucial role in enhancing data diversity and generalization performance. We trained classifiers on our dataset, studied their ability to generalize in various settings, and evaluated previously proposed models and training practices. + +We do not intend for our dataset to be used to train classifiers that are directly used in the wild. Detecting in-the-wild synthetic images remains a challenging open problem, and detection errors can have severe consequences (e.g., falsely accusing an author of creating fake images or allowing misinformation to be certified as real). We hope that our work will serve as a stepping stone for future research in this area by providing tools and insights for studying generalization and data collection strategies. + +Limitations. While our dataset is diverse, a large portion of the data is diffusion-based, especially models based on Stable Diffusion [100]. Despite this, Figure 1 demonstrates that adding more diffusion models to the training data improves the generalization capability of the classifiers across other generator types. Furthermore, Figure 5a illustrates improved classifier performance when combining the diffusion-only systematic set with a more diverse manual set. These results suggest that despite their similarities, each diffusion model may uniquely contribute through its semantic content, image processing pipelines, or architectural variations (Sec. 3.1). Future work may explore collecting more diverse models, including GANs, VQ-VAEs [122], and autoregressive models. We also note that the generative models sourced from the community may contain inappropriate content. While in many contexts it is important to detect such images, these generated images may require further scrutiny before being used in other downstream applications. Finally, although our experiments suggest that our forensics classifiers generalize to unseen models better than those of previous work, their error rates are still too high for many critical applications. + +Acknowledgements. We thank the creators of the many open source models that we used to collect the Community Forensics dataset. We thank Chenhao Zheng, Cameron Johnson, Matthias Kirchner, Daniel Geng, Ziyang Chen, Ayush Shrivastava, Yiming Dou, Chao Feng, Xuanchen Lu, Zihao Wei, Zixuan Pan, Inbum Park, Rohit Banerjee, and Ang Cao for the valuable discussions and feedback. This research was developed with funding from the Defense Advanced Research Projects Agency (DARPA) under Contract No. HR001120C0123. + +# References + +[1] Adobe. Firefly. https://www.adobe.com/products/firefly, 2023.4, 5 +[2] Ideogram AI. Ideogram. https://ideogram.ai, 2024.4 +[3] Rushil Anirudh and Jayaraman J Thiagarajan. Out of distribution detection via neural network anchoring. In Asian Conference on Machine Learning, pages 32-47. PMLR, 2023. 3 +[4] Ivan Anokhin, Kirill Demochkin, Taras Khakhulin, Gleb Sterkin, Victor Lempitsky, and Denis Korzhenkov. Image generators with conditionally-independent pixel synthesis. arXiv preprint arXiv:2011.13775, 2020. 4 +[5] Vishal Asnani, Xi Yin, Tal Hassner, and Xiaoming Liu. Reverse engineering of generative models: Inferring model hyperparameters from generated images. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(12):15477-15493, 2023. 2, 3, 4, 5 +[6] Anton Bakhtin, Sam Gross, Myle Ott, Yuntian Deng, Marc'Aurelio Ranzato, and Arthur Szlam. Real or fake? learning to discriminate machine from human generated text. arXiv preprint arXiv:1906.03351, 2019. 3 +[7] Quentin Bammey. Synthbuster: Towards detection of diffusion model generated images. IEEE Open Journal of Signal Processing, 2023. 2, 3, 4, 5, 6, 7 +[8] Andrew Brock, Jeff Donahue, and Karen Simonyan. Large scale GAN training for high fidelity natural image synthesis. In International Conference on Learning Representations, 2019. 3, 4, 5 +[9] Chen Chen, Qifeng Chen, Jia Xu, and Vladlen Koltun. Learning to see in the dark. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3291-3300, 2018. 3, 5 +[10] Qifeng Chen and Vladlen Koltun. Photographic image synthesis with cascaded refinement networks. In Proceedings of the IEEE international conference on computer vision, pages 1511-1520, 2017. 3, 5 +[11] Mehdi Cherti, Romain Beaumont, Ross Wightman, Mitchell Wortsman, Gabriel Ilharco, Cade Gordon, Christoph Schuhmann, Ludwig Schmidt, and Jenia Jitsev. Reproducible scaling laws for contrastive language-image learning. arXiv preprint arXiv:2212.07143, 2022.5 +[12] Yunjey Choi, Minje Choi, Munyoung Kim, Jung-Woo Ha, Sunghun Kim, and Jaegul Choo. Stargan: Unified generative adversarial networks for multi-domain image-to-image translation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2018. 3, 5 +[13] Yunjey Choi, Youngjung Uh, Jaejun Yoo, and Jung-Woo Ha. Stargan v2: Diverse image synthesis for multiple domains. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2020. 4 +[14] Riccardo Corvi, Davide Cozzolino, Giada Zingarini, Giovanni Poggi, Koki Nagano, and Luisa Verdoliva. On the detection of synthetic images generated by diffusion models. In ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1-5. IEEE, 2023. 3 + +[15] Davide Cozzolino, Giovanni Poggi, Riccardo Corvi, Matthias Nießner, and Luisa Verdoliva. Raising the bar of ai-generated image detection with clip. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4356-4366, 2024. 2, 3, 4, 5, 7 +[16] Katherine Crowson, Stefan Andreas Baumann, Alex Birch, Tanishq Mathew Abraham, Daniel Z Kaplan, and Enrico Shippole. Scalable high-resolution pixel-space image synthesis with hourglass diffusion transformers. In Proceedings of the 41st International Conference on Machine Learning, pages 9550-9575. PMLR, 2024. 5 +[17] Ekin Dogus Cubuk, Barret Zoph, Jon Shlens, and Quoc Le. Randaugment: Practical automated data augmentation with a reduced search space. In Advances in Neural Information Processing Systems, pages 18613-18624. Curran Associates, Inc., 2020. 5 +[18] Tao Dai, Jianrui Cai, Yongbing Zhang, Shu-Tao Xia, and Lei Zhang. Second-order attention network for single image super-resolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11065-11074, 2019. 3, 5 +[19] Duc-Tien Dang-Nguyen, Cecilia Pasquini, Valentina Conotter, and Giulia Boato. Raise: A raw images dataset for digital image forensics. In Proceedings of the 6th ACM multimedia systems conference, pages 219-224, 2015. 5 +[20] Quan Dao, Hao Phung, Binh Nguyen, and Anh Tran. Flow matching in latent space. arXiv preprint arXiv:2307.08698, 2023.4 +[21] Tiago José De Carvalho, Christian Riess, Eli Angelopoulou, Helio Pedrini, and Anderson de Rezende Rocha. Exposing digital image forgeries by illumination color classification. IEEE Transactions on Information Forensics and Security, 8(7):1182-1194, 2013. 3 +[22] DeepFloyd. Deepfloyd. https://huggingface.co/DeepFloyd/IF-I-L-v1.0, 2024.4 +[23] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 4, 5, 7 +[24] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 4, 5 +[25] Brian Dolhansky, Joanna Bitton, Ben Pflaum, Jikuo Lu, Russ Howes, Menglin Wang, and Cristian Canton Ferrer. The deepfake detection challenge (dfdc) dataset. arXiv preprint arXiv:2006.07397, 2020. 2, 5 +[26] Jing Dong, Wei Wang, and Tieniu Tan. Casia image tampering detection evaluation database. In 2013 IEEE China summit and international conference on signal and information processing, pages 422-426. IEEE, 2013. 3 +[27] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. ICLR, 2021. 5, 6, 7 + +[28] Yilun Du, Shuang Li, Joshua Tenenbaum, and Igor Mordatch. Improved contrastive divergence training of energy-based models. In International Conference on Machine Learning, pages 2837-2848. PMLR, 2021. 3 +[29] Ricard Durall, Margret Keuper, and Janis Keuper. Watch your up-convolution: Cnn based generative deep neural networks are failing to reproduce spectral distributions. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7890–7899, 2020. 3 +[30] Tarik Dzanic, Karan Shah, and Freddie Witherden. Fourier spectrum discrepancies in deep network generated images. Advances in neural information processing systems, 33:3022-3032, 2020. 3 +[31] Sven Elflein, Bertrand Charpentier, Daniel Zügner, and Stephan Gunnemann. On out-of-distribution detection with energy-based models, 2021. 3 +[32] David C Epstein, Ishan Jain, Oliver Wang, and Richard Zhang. Online detection of ai-generated images. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 382-392, 2023. 2, 3, 4 +[33] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12873-12883, 2021. 4 +[34] Hugging Face. Hugging face diffusers library. https://huggingface.co/models?library=diffusers, accessed on June 05, 2022, 2022. 2, 3, 4 +[35] Sebastian Gehrmann, Hendrik Strobelt, and Alexander M Rush. Gltr: Statistical detection and visualization of generated text. arXiv preprint arXiv:1906.04043, 2019. 3 +[36] Thomas Gloe and Rainer Böhme. The'dresden image database' for benchmarking digital image forensics. In Proceedings of the 2010 ACM symposium on applied computing, pages 1584-1590, 2010. 4 +[37] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in Neural Information Processing Systems, 27, 2014. 2 +[38] Google. Imagen 3. https://deepmind.google/technologies/imagen-3, 2024.4 +[39] Karol Gregor, Ivo Danihelka, Andriy Mnih, Charles Blundell, and Daan Wierstra. Deep autoregressive networks. In International Conference on Machine Learning, pages 1242-1250. PMLR, 2014. 2 +[40] Patrick Grommelt, Louis Weiss, Franz-Josef Pfreundt, and Janis Keuper. Fake or.jpeg? revealing common biases in generated image detection datasets. arXiv preprint arXiv:2403.17608, 2024. 3 +[41] Shuyang Gu, Dong Chen, Jianmin Bao, Fang Wen, Bo Zhang, Dongdong Chen, Lu Yuan, and Baining Guo. Vector quantized diffusion model for text-to-image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10696-10706, 2022. 4, 5 +[42] Fabrizio Guillaro, Giada Zingarini, Ben Usman, Avneesh Sud, Davide Cozzolino, and Luisa Verdoliva. A bias-free training paradigm for more general ai-generated image detection, 2024. 3 + +[43] Xiao Guo, Vishal Asnani, Sijja Liu, and Xiaoming Liu. Tracing hyperparameter dependencies for model parsing via learnable graph pooling network. Advances in Neural Information Processing Systems, 37:116899-116932, 2024. 2, 3, 4, 5, 6 +[44] Benjamin Hadwiger and Christian Riess. The forchheim image database for camera identification in the wild. In Pattern Recognition. ICPR International Workshops and Challenges: Virtual Event, January 10–15, 2021, Proceedings, Part VI, pages 500–515. Springer, 2021. 4, 8 +[45] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 5, 6 +[46] Dan Hendrycks and Kevin Gimpel. A baseline for detecting misclassified and out-of-distribution examples in neural networks. In International Conference on Learning Representations, 2016. 3 +[47] Dan Hendrycks, Mantas Mazeika, Saurav Kadavath, and Dawn Song. Using self-supervised learning can improve model robustness and uncertainty. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2019. 3 +[48] Dan Hendrycks, Steven Basart, Norman Mu, Saurav Kadavath, Frank Wang, Evan Dorundo, Rahul Desai, Tyler Zhu, Samyak Parajuli, Mike Guo, Dawn Song, Jacob Steinhardt, and Justin Gilmer. The many faces of robustness: A critical analysis of out-of-distribution generalization. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 8340-8349, 2021. 2, 3 +[49] Yan Hong and Jianfu Zhang. Wildfake: A large-scale challenging dataset for ai-generated images detection. arXiv preprint arXiv:2402.11843, 2024. 3 +[50] Shuwei Hou, Yan Ju, Chengzhe Sun, Shan Jia, Lipeng Ke, Riky Zhou, Anita Nikolich, and Siwei Lyu. Deepfake-meter v2.0: An open platform for deepfake detection. arXiv preprint arXiv:2404.13146, 2024. 3 +[51] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 4 +[52] Drew A Hudson and Larry Zitnick. Generative adversarial transformers. In International conference on machine learning, pages 4487-4499. PMLR, 2021. 4 +[53] Minyoung Huh, Andrew Liu, Andrew Owens, and Alexei A Efros. Fighting fake news: Image splice detection via learned self-consistency. In Proceedings of the European conference on computer vision (ECCV), 2018. 3 +[54] Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. Openclip, 2021. If you use this software, please cite it as below. 5, 6, 7 +[55] Ganesh Jawahar, Muhammad Abdul-Mageed, and Laks VS Lakshmanan. Automatic detection of machine generated text: A critical survey. arXiv preprint arXiv:2011.01314, 2020. 3 + +[56] Micah K Johnson and Hany Farid. Exposing digital forgeries in complex lighting environments. IEEE Transactions on Information Forensics and Security, 2(3):450-461, 2007. 3 +[57] Minguk Kang, Jun-Yan Zhu, Richard Zhang, Jaesik Park, Eli Shechtman, Sylvain Paris, and Taesung Park. Scaling up gans for text-to-image synthesis. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 4, 5 +[58] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of gans for improved quality, stability, and variation. In International Conference on Learning Representations, 2018. 3, 4 +[59] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 3, 4, 5, 6 +[60] Tero Karras, Miika Aittala, Janne Hellsten, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Training generative adversarial networks with limited data. Advances in neural information processing systems, 33:12104-12114, 2020. 4 +[61] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8110-8119, 2020. +[62] Tero Karras, Miika Aittala, Samuli Laine, Erik Härkonen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Alias-free generative adversarial networks. Advances in neural information processing systems, 34:852-863, 2021. 4 +[63] Hasam Khalid, Shahroz Tariq, Minha Kim, and Simon Woo. Fakeavceleb: A novel audio-video multimodal deepfake dataset. In Proceedings of the Neural Information Processing Systems Track on Datasets and Benchmarks, 2021. 2 +[64] P. Korus and J. Huang. Multi-scale analysis strategies in prnu-based tampering localization. IEEE Trans. on Information Forensics & Security, 2017. 3 +[65] Kalpesh Krishna, Yixiao Song, Marzena Karpinska, John Wieting, and Mohit Iyyer. Paraphrasing evades detectors of ai-generated text, but retrieval is an effective defense. Advances in Neural Information Processing Systems, 36, 2024. 3 +[66] Kvikontent. Kvikontent-midjourney v6. https://huggingface.co/Kvikontent/midjourney-v6, 2023.5 +[67] Patrick Kwon, Jaeseong You, Gyuhyeon Nam, Sungwoo Park, and Gyeongsu Chae. Kodf: A large-scale korean deepfake detection dataset. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021. 2 +[68] Black Forst Labs. Flux. https://blackforestlabs.ai, 2024.4 +[69] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, 2022. 5 +[70] Ke Li, Tianhao Zhang, and Jitendra Malik. Diverse image synthesis from semantic layouts via conditional imle. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4220-4229, 2019. 3, 5 + +[71] Yuezun Li, Xin Yang, Pu Sun, Honggang Qi, and Siwei Lyu. Celeb-df: A large-scale challenging dataset for deepfake forensics. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2020. 2 +[72] Yewen Li, Chaojie Wang, Xiaobo Xia, Tongliang Liu, Bo An, et al. Out-of-distribution detection with an adaptive likelihood ratio on informative hierarchical vae. Advances in Neural Information Processing Systems, 35:7383-7396, 2022. 3 +[73] Shiyu Liang, Yixuan Li, and R Srikant. Enhancing the reliability of out-of-distribution image detection in neural networks. In International Conference on Learning Representations, 2018. 3 +[74] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 4, 5, 6, 8 +[75] Weitang Liu, Xiaoyun Wang, John Owens, and Yixuan Li. Energy-based out-of-distribution detection. In Advances in Neural Information Processing Systems, pages 21464-21475. Curran Associates, Inc., 2020. 3 +[76] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaou Tang. Deep learning face attributes in the wild. In Proceedings of International Conference on Computer Vision (ICCV), 2015. 4 +[77] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11976-11986, 2022. 5, 7 +[78] Simian Luo, Yiqin Tan, Longbo Huang, Jian Li, and Hang Zhao. Latent consistency models: Synthesizing high-resolution images with few-step inference. arXiv preprint arXiv:2310.04378, 2023. 4 +[79] Simian Luo, Yiqin Tan, Suraj Patil, Daniel Gu, Patrick von Platen, Apolinário Passos, Longbo Huang, Jian Li, and Hang Zhao. Lcm-lora: A universal stable-diffusion acceleration module. arXiv preprint arXiv:2311.05556, 2023. 2, 5 +[80] Francesco Marra, Diego Gragnaniello, Luisa Verdoliva, and Giovanni Poggi. Do gans leave artificial fingerprints? In 2019 IEEE conference on multimedia information processing and retrieval (MIPR), pages 506-511. IEEE, 2019. 3 +[81] Inc. Midjourney. Midjourney. https://www.midjourney.com/home, 2022. 4, 5 +[82] Eric Mitchell, Yoonho Lee, Alexander Khazatsky, Christopher D Manning, and Chelsea Finn. Detectgpt: Zero-shot machine-generated text detection using probability curvature. In International Conference on Machine Learning, pages 24950-24962. PMLR, 2023. 3 +[83] Sina Mohseni, Mandar Pitale, JBS Yadawa, and Zhangyang Wang. Self-supervised learning for generalizable out-of-distribution detection. Proceedings of the AAAI Conference on Artificial Intelligence, 34(04):5216-5223, 2020. 3 +[84] Tian-Tsong Ng, Shih-Fu Chang, and Q Sun. A data set of authentic and spliced image blocks. Columbia University, ADVENT Technical Report, 4, 2004. 3 + +[85] Alexander Quinn Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. In International Conference on Machine Learning, pages 16784-16804. PMLR, 2022. 4, 5 +[86] Adam Novozamsky, Babak Mahdian, and Stanislav Saic. Imd2020: A large-scale annotated dataset tailored for detecting manipulated images. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision Workshops, pages 71-80, 2020. 4, 8 +[87] Utkarsh Ojha, Yuheng Li, and Yong Jae Lee. Towards universal fake image detectors that generalize across generative models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 24480-24489, 2023. 1, 2, 3, 4, 5, 6, 7, 8 +[88] OpenAI. Dall-e 2. https://openai.com/index/dall-e-2, 2022. 4, 5 +[89] OpenAI. Dall-e 3. https://openai.com/index/dall-e-3, 2023.4, 5 +[90] Taesung Park, Ming-Yu Liu, Ting-Chun Wang, and JunYan Zhu. Semantic image synthesis with spatially-adaptive normalization. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2019. 3 +[91] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019. 5 +[92] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4195-4205, 2023. 4 +[93] Pablo Pernias, Dominic Rampas, Mats Leon Richter, Christopher Pal, and Marc Aubreville. Würstchen: An efficient architecture for large-scale text-to-image diffusion models. In The Twelfth International Conference on Learning Representations, 2023. 5, 6 +[94] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. In The Twelfth International Conference on Learning Representations, 2023. 5 +[95] Alin C Popescu and Hany Farid. Exposing digital forgeries by detecting traces of resampling. IEEE Transactions on signal processing, 2005. 3 +[96] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 3, 5, 6, 7 +[97] Anirudh Sundara Rajan, Utkarsh Ojha, Jedidiah Schoesser, and Yong Jae Lee. Aligned datasets improve detection of latent diffusion-generated images, 2025. 3 + +[98] Jie Ren, Peter J Liu, Emily Fertig, Jasper Snoek, Ryan Poplin, Mark Depristo, Joshua Dillon, and Balaji Lakshminarayanan. Likelihood ratios for out-of-distribution detection. Advances in neural information processing systems, 32, 2019. 3 +[99] Jonas Ricker, Denis Lukovnikov, and Asja Fischer. Aeroblade: Training-free detection of latent diffusion images using autoencoder reconstruction error. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9130-9140, 2024. 6 +[100] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, 2022. 2, 4, 5, 8 +[101] Andreas Rössler, Davide Cozzolino, Luisa Verdoliva, Christian Riess, Justus Thies, and Matthias Nießner. Faceforensics: A large-scale video dataset for forgery detection in human faces. arXiv preprint arXiv:1803.09179, 2018. 1, 2 +[102] Andreas Rossler, Davide Cozzolino, Luisa Verdoliva, Christian Riess, Justus Thies, and Matthias Nießner. Faceforensics++: Learning to detect manipulated facial images. In Proceedings of the IEEE/CVF international conference on computer vision, 2019. 1, 2, 3 +[103] Vinu Sankar Sadasivan, Aounon Kumar, Sriram Balasubramanian, Wenxiao Wang, and Soheil Feizi. Can ai-generated text be reliably detected? arXiv preprint arXiv:2303.11156, 2023. 3 +[104] Chandramouli Shama Sastry and Sageev Oore. Detecting out-of-distribution examples with Gram matrices. In Proceedings of the 37th International Conference on Machine Learning, pages 8491-8501. PMLR, 2020. 3 +[105] Axel Sauer, Kashyap Chitta, Jens Müller, and Andreas Geiger. Projected gans converge faster. In Advances in Neural Information Processing Systems (NeurIPS), 2021. 4 +[106] Axel Sauer, Katja Schwarz, and Andreas Geiger. Styleganxl: Scaling stylegan to large diverse datasets. In ACM SIGGRAPH 2022 conference proceedings, pages 1-10, 2022. 4 +[107] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021. 2, 4, 5, 7 +[108] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade W Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortman, Patrick Schramowski, Srivatsa R Kundurthy, Katherine Crowson, Ludwig Schmidt, Robert Kaczmarczyk, and Jenia Jitsev. LAION-5b: An open large-scale dataset for training next generation image-text models. In Thirty-sixth Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2022. 2, 3, 4, 5 +[109] Vikash Sehwag, Mung Chiang, and Prateek Mittal. Ssd: A unified framework for self-supervised outlier detection. In International Conference on Learning Representations, 2020. 3 + +[110] Arseniy Shakhmatov, Anton Razzhigaev, Aleksandr Nikolich, Vladimir Arkhipkin, Igor Pavlov, Andrey Kuznetsov, and Denis Dimitrov. Kandinsky 2.2. https://github.com/ai-forever/Kandinsky-2, 2023.5 +[111] Dasara Shullani, Marco Fontani, Massimo Iuliani, Omar Al Shaya, and Alessandro Piva. Vision: a video and image dataset for source identification. EURASIP Journal on Information Security, 2017:1-16, 2017. 4, 8 +[112] Ivan Skorokhodov, Grigorii Sotnikov, and Mohamed Elhoseiny. Aligning latent and image spaces to connect the unconnectable. arXiv preprint arXiv:2104.06954, 2021. 4, 8 +[113] Irene Solaiman, Miles Brundage, Jack Clark, Amanda Askell, Ariel Herbert-Voss, Jeff Wu, Alec Radford, Gretchen Krueger, Jong Wook Kim, Sarah Kreps, et al. Release strategies and the social impacts of language models. arXiv preprint arXiv:1908.09203, 2019. 3 +[114] Yang Song, Prafulla Dhariwal, Mark Chen, and Ilya Sutskever. Consistency models. In Proceedings of the 40th International Conference on Machine Learning, pages 32211-32252, 2023. 2 +[115] Yuhta Takida, Masaaki Imaizumi, Takashi Shibuya, Chieh-Hsin Lai, Toshimitsu Uesaka, Naoki Murata, and Yuki Mitsufuji. SAN: Inducing metrizability of GAN with discriminative normalized linear layer. In The Twelfth International Conference on Learning Representations, 2024. 4 +[116] Chuangchuang Tan, Yao Zhao, Shikui Wei, Guanghua Gu, and Yunchao Wei. Learning on gradients: Generalized artifacts representation for gan-generated images detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12105-12114, 2023. 6 +[117] Chuangchuang Tan, Yao Zhao, Shikui Wei, Guanghua Gu, Ping Liu, and Yunchao Wei. Rethinking the up-sampling operations in cnn-based generative network for generalizable deepfake detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 28130-28139, 2024. 6 +[118] Ming Tao, Hao Tang, Fei Wu, Xiao-Yuan Jing, Bing-Kun Bao, and Changsheng Xu. Df-gan: A simple and effective baseline for text-to-image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16515-16525, 2022. 5 +[119] Ming Tao, Bing-Kun Bao, Hao Tang, and Changsheng Xu. Galip: Generative adversarial clips for text-to-image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14214-14223, 2023. 5 +[120] DeciAI Research Team. Decidiffusion 2.0, 2024. 5 +[121] Adaku Uchendu, Thai Le, Kai Shu, and Dongwon Lee. Authorship attribution for neural text generation. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 8384–8395, 2020. 3 +[122] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017. 8 + +[123] Patrick von Platen, Suraj Patil, Anton Lozhkov, Pedro Cuenca, Nathan Lambert, Kashif Rasul, Mishig Davaadorj, Dhruv Nair, Sayak Paul, William Berman, Yiyi Xu, Steven Liu, and Thomas Wolf. Diffusers: State-of-the-art diffusion models. https://github.com/huggingface/ diffusers, 2022. 2, 3, 4 +[124] Apoorv Vyas, Nataraj Jammalamadaka, Xia Zhu, Dipankar Das, Bharat Kaul, and Theodore L. Willke. Out-of-distribution detection using an ensemble of self-supervised leave-out classifiers. In Proceedings of the European Conference on Computer Vision (ECCV), 2018. 3 +[125] Sheng-Yu Wang, Oliver Wang, Richard Zhang, Andrew Owens, and Alexei A. Efros. Cnn-generated images are surprisingly easy to spot... for now. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 1, 2, 3, 4, 5, 6, 7, 8 +[126] Ross Wightman. Pytorch image models. https://github.com/huggingface/pytorch-image-models, 2019.5 +[127] Wukong. https://xihe.mindspore.cn/modelzoo/wukong, 2022.5 +[128] Zhisheng Xiao, Qing Yan, and Yali Amit. Likelihood regret: An out-of-distribution detection score for variational autoencoder. In Advances in Neural Information Processing Systems, pages 20685-20696. Curran Associates, Inc., 2020. 3 +[129] Fisher Yu, Yinda Zhang, Shuran Song, Ari Seff, and Jianxiong Xiao. Lsun: Construction of a large-scale image dataset using deep learning with humans in the loop. arXiv preprint arXiv:1506.03365, 2015. 4 +[130] Bowen Zhang, Shuyang Gu, Bo Zhang, Jianmin Bao, Dong Chen, Fang Wen, Yong Wang, and Baining Guo. Styleswin: Transformer-based gan for high-resolution image generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11304-11314, 2022. 4 +[131] Xu Zhang, Svebor Karaman, and Shih-Fu Chang. Detecting and simulating artifacts in gan fake images. In 2019 IEEE international workshop on information forensics and security (WIFS), pages 1-6. IEEE, 2019. 3 +[132] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In Computer Vision (ICCV), 2017 IEEE International Conference on, 2017. 3, 5 +[133] Mingjian Zhu, Hanting Chen, Qiangyu YAN, Xudong Huang, Guanyu Lin, Wei Li, Zhijun Tu, Hailin Hu, Jie Hu, and Yunhe Wang. Genimage: A million-scale benchmark for detecting ai-generated image. In Advances in Neural Information Processing Systems Dataset and Benchmarks Track, pages 77771-77782, 2023. 2, 3, 4, 5, 6, 7, 8 +[134] Bojia Zi, Minghao Chang, Jingjing Chen, Xingjun Ma, and Yu-Gang Jiang. Wilddeepfake: A challenging real-world dataset for deepfake detection. In Proceedings of the 28th ACM international conference on multimedia, 2020. 2 \ No newline at end of file diff --git a/2025/Community Forensics_ Using Thousands of Generators to Train Fake Image Detectors/images.zip b/2025/Community Forensics_ Using Thousands of Generators to Train Fake Image Detectors/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..e5f5597b88342a220cdf805b8ad7d17837cf6c38 --- /dev/null +++ b/2025/Community Forensics_ Using Thousands of Generators to Train Fake Image Detectors/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c39ac47bfb5f650c220b7cf5873cd9d08edeafb51a450cbbab2eba4df293b32c +size 449178 diff --git a/2025/Community Forensics_ Using Thousands of Generators to Train Fake Image Detectors/layout.json b/2025/Community Forensics_ Using Thousands of Generators to Train Fake Image Detectors/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..52a41b1e739e29f341e946a6032d0900f9eaecb2 --- /dev/null +++ b/2025/Community Forensics_ Using Thousands of Generators to Train Fake Image Detectors/layout.json @@ -0,0 +1,11769 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 108, + 103, + 501, + 139 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 103, + 501, + 139 + ], + "spans": [ + { + "bbox": [ + 108, + 103, + 501, + 139 + ], + "type": "text", + "content": "Community Forensics: Using Thousands of Generators to Train Fake Image Detectors" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 206, + 161, + 403, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 206, + 161, + 403, + 190 + ], + "spans": [ + { + "bbox": [ + 206, + 161, + 403, + 190 + ], + "type": "text", + "content": "Jeongsoo Park Andrew Owens University of Michigan" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 151, + 217, + 200, + 229 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 217, + 200, + 229 + ], + "spans": [ + { + "bbox": [ + 151, + 217, + 200, + 229 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 243, + 297, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 243, + 297, + 494 + ], + "spans": [ + { + "bbox": [ + 54, + 243, + 297, + 494 + ], + "type": "text", + "content": "One of the key challenges of detecting AI-generated images is spotting images that have been created by previously unseen generative models. We argue that the limited diversity of the training data is a major obstacle to addressing this problem, and we propose a new dataset that is significantly larger and more diverse than prior works. As part of creating this dataset, we systematically download thousands of text-to-image latent diffusion models and sample images from them. We also collect images from dozens of popular open source and commercial models. The resulting dataset contains 2.7M images that have been sampled from 4803 different models. These images collectively capture a wide range of scene content, generator architectures, and image processing settings. Using this dataset, we study the generalization abilities of fake image detectors. Our experiments suggest that detection performance improves as the number of models in the training set increases, even when these models have similar architectures. We also find that increasing the diversity of the models improves detection performance, and that our trained detectors generalize better than those trained on other datasets. The dataset can be found in" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 495, + 279, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 495, + 279, + 506 + ], + "spans": [ + { + "bbox": [ + 55, + 495, + 279, + 506 + ], + "type": "text", + "content": "https://jespark.net/projects/2024/community_forensics" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 528, + 135, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 528, + 135, + 540 + ], + "spans": [ + { + "bbox": [ + 55, + 528, + 135, + 540 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 545, + 295, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 295, + 664 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 295, + 664 + ], + "type": "text", + "content": "Our ability to automatically generate realistic images is quickly outpacing our ability to detect them, potentially leading to a state of affairs in which neither humans nor machines can reliably tell real from fake. While the field of image forensics has been developing methods to address this problem, existing fake image detectors still struggle with generalization. These methods often excel at detecting images from generators that were present in their training sets, but fail when given images sampled from unseen models [87, 101, 102, 125]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "A core challenge is dealing with the large amounts of variation between models. Each generator has a potentially unique combination of the architecture, loss function, and training distribution. Even seemingly minor differences in" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 322, + 219, + 545, + 381 + ], + "blocks": [ + { + "bbox": [ + 322, + 219, + 545, + 381 + ], + "lines": [ + { + "bbox": [ + 322, + 219, + 545, + 381 + ], + "spans": [ + { + "bbox": [ + 322, + 219, + 545, + 381 + ], + "type": "image", + "image_path": "653df4c788aa9892c3e27775054108cb9797cf24bedc786da1a1ff5df83470f9.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 389, + 555, + 488 + ], + "lines": [ + { + "bbox": [ + 313, + 389, + 555, + 488 + ], + "spans": [ + { + "bbox": [ + 313, + 389, + 555, + 488 + ], + "type": "text", + "content": "Figure 1. Performance vs. model diversity. We use images sampled from different numbers of open source latent diffusion models in our Community Forensics dataset to train fake image detectors (shown in Fig. 2a). We fix the total number of images and only vary the number of models. The detector's performance increases across all generator types as we train from more models, even though these added models are entirely latent diffusion. This improvement is largest for test images from out-of-distribution generators, such as pixel-based diffusion models or GANs." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 498, + 556, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 556, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 556, + 605 + ], + "type": "text", + "content": "low-level image processing details, such as the ways that training images are resized or compressed, can strongly influence detection accuracy [125]. As a result of these model-specific idiosyncrasies, a generator's images may evade detection, even when images from architecturally similar models exist in the training set. This issue has been exacerbated by the thousands of open source models that are now available online, many of which extend pretrained base models in complex ways." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 605, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 556, + 713 + ], + "type": "text", + "content": "We hypothesize that the lack of diversity in training datasets is a major source of these shortcomings. Although today's datasets often contain millions of fake images, they come from a relatively small number of generators. As a result, this data fails to capture many sources of variation that one might encounter in the wild. These limitations also make it challenging to accurately benchmark performance, since it is easy for cues that work well on one set of generators to fail on others." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8245" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 111, + 72, + 153, + 121 + ], + "blocks": [ + { + "bbox": [ + 61, + 72, + 101, + 121 + ], + "lines": [ + { + "bbox": [ + 61, + 72, + 101, + 121 + ], + "spans": [ + { + "bbox": [ + 61, + 72, + 101, + 121 + ], + "type": "text", + "content": "Dreamshaper v6" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 111, + 72, + 153, + 121 + ], + "lines": [ + { + "bbox": [ + 111, + 72, + 153, + 121 + ], + "spans": [ + { + "bbox": [ + 111, + 72, + 153, + 121 + ], + "type": "image", + "image_path": "d1f5072360d364ca47d054567c71f9c5839673cf0b8b2d63a1aa62a5e415580c.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 160, + 72, + 200, + 121 + ], + "blocks": [ + { + "bbox": [ + 160, + 72, + 200, + 121 + ], + "lines": [ + { + "bbox": [ + 160, + 72, + 200, + 121 + ], + "spans": [ + { + "bbox": [ + 160, + 72, + 200, + 121 + ], + "type": "image", + "image_path": "6fe9aa0173c01026b84b59881178e13aa1eb95ccb09bf03e14da648a3e1b7ce9.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 209, + 72, + 252, + 121 + ], + "blocks": [ + { + "bbox": [ + 209, + 72, + 252, + 121 + ], + "lines": [ + { + "bbox": [ + 209, + 72, + 252, + 121 + ], + "spans": [ + { + "bbox": [ + 209, + 72, + 252, + 121 + ], + "type": "image", + "image_path": "f18b5a9ce3439f4e81c3684d1eb232b9139713186c8cf0fbac687c95099c022a.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 209, + 121, + 249, + 127 + ], + "lines": [ + { + "bbox": [ + 209, + 121, + 249, + 127 + ], + "spans": [ + { + "bbox": [ + 209, + 121, + 249, + 127 + ], + "type": "text", + "content": "Plat diffusion" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 258, + 72, + 302, + 121 + ], + "blocks": [ + { + "bbox": [ + 258, + 72, + 302, + 121 + ], + "lines": [ + { + "bbox": [ + 258, + 72, + 302, + 121 + ], + "spans": [ + { + "bbox": [ + 258, + 72, + 302, + 121 + ], + "type": "image", + "image_path": "5fce98152588acb7455cc5f2fa8bc91cfea7bb6956a8c43e61e19ce49c1c93c8.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 271, + 121, + 290, + 127 + ], + "lines": [ + { + "bbox": [ + 271, + 121, + 290, + 127 + ], + "spans": [ + { + "bbox": [ + 271, + 121, + 290, + 127 + ], + "type": "text", + "content": "SoteMix" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 258, + 77, + 302, + 121 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 258, + 77, + 302, + 121 + ], + "spans": [ + { + "bbox": [ + 258, + 77, + 302, + 121 + ], + "type": "image", + "image_path": "4af300b328e789d19d90fe53f2ec05ca8a6fadae813d13cc44f1221ac2348325.jpg" + } + ] + } + ], + "index": 6, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 258, + 127, + 302, + 170 + ], + "blocks": [ + { + "bbox": [ + 258, + 127, + 302, + 170 + ], + "lines": [ + { + "bbox": [ + 258, + 127, + 302, + 170 + ], + "spans": [ + { + "bbox": [ + 258, + 127, + 302, + 170 + ], + "type": "image", + "image_path": "d1ddd487777710b7b38dcd33a26da7bfbc50be43bf1bcb1148ba2820d91ca332.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 258, + 175, + 302, + 219 + ], + "blocks": [ + { + "bbox": [ + 267, + 171, + 294, + 175 + ], + "lines": [ + { + "bbox": [ + 267, + 171, + 294, + 175 + ], + "spans": [ + { + "bbox": [ + 267, + 171, + 294, + 175 + ], + "type": "text", + "content": "Technoalbum" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 258, + 175, + 302, + 219 + ], + "lines": [ + { + "bbox": [ + 258, + 175, + 302, + 219 + ], + "spans": [ + { + "bbox": [ + 258, + 175, + 302, + 219 + ], + "type": "image", + "image_path": "51c233be4c773bc38d072851e63c7202864285b451288867fda884ec0b8f9606.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 209, + 219, + 253, + 225 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 209, + 219, + 253, + 225 + ], + "spans": [ + { + "bbox": [ + 209, + 219, + 253, + 225 + ], + "type": "text", + "content": "Nightvisionxl 0791" + } + ] + } + ], + "index": 12, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 209, + 175, + 253, + 219 + ], + "blocks": [ + { + "bbox": [ + 209, + 175, + 253, + 219 + ], + "lines": [ + { + "bbox": [ + 209, + 175, + 253, + 219 + ], + "spans": [ + { + "bbox": [ + 209, + 175, + 253, + 219 + ], + "type": "image", + "image_path": "2ebea408eef862ea73fa5d22b6093a1f7f0d775f83e9386a3ef0bae503afa177.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 209, + 219, + 253, + 225 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 209, + 219, + 253, + 225 + ], + "spans": [ + { + "bbox": [ + 209, + 219, + 253, + 225 + ], + "type": "text", + "content": "Nightvisionxl 0791" + } + ] + } + ], + "index": 15, + "type": "text" + }, + { + "bbox": [ + 258, + 219, + 302, + 225 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 258, + 219, + 302, + 225 + ], + "spans": [ + { + "bbox": [ + 258, + 219, + 302, + 225 + ], + "type": "text", + "content": "Albedobase x1 v21" + } + ] + } + ], + "index": 16, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 209, + 227, + 253, + 267 + ], + "blocks": [ + { + "bbox": [ + 209, + 221, + 253, + 227 + ], + "lines": [ + { + "bbox": [ + 209, + 221, + 253, + 227 + ], + "spans": [ + { + "bbox": [ + 209, + 221, + 253, + 227 + ], + "type": "text", + "content": "Nightvisionxl 0791" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 209, + 227, + 253, + 267 + ], + "lines": [ + { + "bbox": [ + 209, + 227, + 253, + 267 + ], + "spans": [ + { + "bbox": [ + 209, + 227, + 253, + 267 + ], + "type": "image", + "image_path": "4326e416ed416ee6148740a80fae62fc0e38844e93d1775e1ce55d502cd1365c.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 209, + 274, + 253, + 316 + ], + "blocks": [ + { + "bbox": [ + 215, + 268, + 247, + 274 + ], + "lines": [ + { + "bbox": [ + 215, + 268, + 247, + 274 + ], + "spans": [ + { + "bbox": [ + 215, + 268, + 247, + 274 + ], + "type": "text", + "content": "Enasswhidaly" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 209, + 274, + 253, + 316 + ], + "lines": [ + { + "bbox": [ + 209, + 274, + 253, + 316 + ], + "spans": [ + { + "bbox": [ + 209, + 274, + 253, + 316 + ], + "type": "image", + "image_path": "3ba0c90a0452fa25ad27528355032421adeb240c5c1ccdd172d3d1d27736ef00.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 109, + 321, + 303, + 331 + ], + "lines": [ + { + "bbox": [ + 109, + 321, + 303, + 331 + ], + "spans": [ + { + "bbox": [ + 109, + 321, + 303, + 331 + ], + "type": "text", + "content": "(a) Systematically collected diffusion models (4763 models)" + } + ] + } + ], + "index": 65, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 335, + 555, + 380 + ], + "lines": [ + { + "bbox": [ + 55, + 335, + 555, + 380 + ], + "spans": [ + { + "bbox": [ + 55, + 335, + 555, + 380 + ], + "type": "text", + "content": "Figure 2. The Community Forensics dataset. Our dataset contains images sampled from three types of generative models. (a) We systematically download open-source latent diffusion models from a model-sharing community [34, 123]. (b) We select popular open source generators with a variety of architectures and training procedures. (c) We sample from both closed and open state-of-the-art commercial models. We present example images and their corresponding model names." + } + ] + } + ], + "index": 67, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "bbox": [ + 258, + 219, + 302, + 225 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 258, + 219, + 302, + 225 + ], + "spans": [ + { + "bbox": [ + 258, + 219, + 302, + 225 + ], + "type": "text", + "content": "Albedobase x1 v21" + } + ] + } + ], + "index": 20, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 258, + 227, + 302, + 267 + ], + "blocks": [ + { + "bbox": [ + 258, + 227, + 302, + 267 + ], + "lines": [ + { + "bbox": [ + 258, + 227, + 302, + 267 + ], + "spans": [ + { + "bbox": [ + 258, + 227, + 302, + 267 + ], + "type": "image", + "image_path": "c0ea5d2a9ad257f237e85f37379d6059719c942552d2227be09afbb897f5d1a4.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 258, + 274, + 302, + 316 + ], + "blocks": [ + { + "bbox": [ + 269, + 268, + 292, + 274 + ], + "lines": [ + { + "bbox": [ + 269, + 268, + 292, + 274 + ], + "spans": [ + { + "bbox": [ + 269, + 268, + 292, + 274 + ], + "type": "text", + "content": "Msart 541" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 258, + 274, + 302, + 316 + ], + "lines": [ + { + "bbox": [ + 258, + 274, + 302, + 316 + ], + "spans": [ + { + "bbox": [ + 258, + 274, + 302, + 316 + ], + "type": "image", + "image_path": "bcd1b05f51e19ae663ab794f3f071b619c39102a82daab31320e9ca1bebf17e0.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 309, + 72, + 351, + 121 + ], + "blocks": [ + { + "bbox": [ + 328, + 72, + 333, + 77 + ], + "lines": [ + { + "bbox": [ + 328, + 72, + 333, + 77 + ], + "spans": [ + { + "bbox": [ + 328, + 72, + 333, + 77 + ], + "type": "text", + "content": "91" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 309, + 72, + 351, + 121 + ], + "lines": [ + { + "bbox": [ + 309, + 72, + 351, + 121 + ], + "spans": [ + { + "bbox": [ + 309, + 72, + 351, + 121 + ], + "type": "image", + "image_path": "c1f0500db9b788ec7a905b477645e3d86b575ff187a21068ba19fec9ad579572.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 121, + 347, + 127 + ], + "lines": [ + { + "bbox": [ + 313, + 121, + 347, + 127 + ], + "spans": [ + { + "bbox": [ + 313, + 121, + 347, + 127 + ], + "type": "text", + "content": "First avarfar" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 309, + 127, + 351, + 170 + ], + "blocks": [ + { + "bbox": [ + 309, + 127, + 351, + 170 + ], + "lines": [ + { + "bbox": [ + 309, + 127, + 351, + 170 + ], + "spans": [ + { + "bbox": [ + 309, + 127, + 351, + 170 + ], + "type": "image", + "image_path": "914cb7b7cd514259ed7b6e25c49761a1b827a8fc3a1aa3885ae3d7d40d418705.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 309, + 176, + 351, + 219 + ], + "blocks": [ + { + "bbox": [ + 308, + 171, + 352, + 176 + ], + "lines": [ + { + "bbox": [ + 308, + 171, + 352, + 176 + ], + "spans": [ + { + "bbox": [ + 308, + 171, + 352, + 176 + ], + "type": "text", + "content": "AnalogMadness real" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 309, + 176, + 351, + 219 + ], + "lines": [ + { + "bbox": [ + 309, + 176, + 351, + 219 + ], + "spans": [ + { + "bbox": [ + 309, + 176, + 351, + 219 + ], + "type": "image", + "image_path": "fa4a628fee64f9e2493ca5a29c7054ea0efa00da4dbce0c165ebf526e61f7b2a.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 309, + 219, + 351, + 225 + ], + "lines": [ + { + "bbox": [ + 309, + 219, + 351, + 225 + ], + "spans": [ + { + "bbox": [ + 309, + 219, + 351, + 225 + ], + "type": "text", + "content": "DucHaitenAnimated" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 309, + 227, + 351, + 267 + ], + "blocks": [ + { + "bbox": [ + 309, + 227, + 351, + 267 + ], + "lines": [ + { + "bbox": [ + 309, + 227, + 351, + 267 + ], + "spans": [ + { + "bbox": [ + 309, + 227, + 351, + 267 + ], + "type": "image", + "image_path": "b9218cfe77d7f7d2a088456a4948bad36fa0cf4db97cbdfc86138418761ef07d.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 309, + 274, + 351, + 316 + ], + "blocks": [ + { + "bbox": [ + 318, + 268, + 342, + 274 + ], + "lines": [ + { + "bbox": [ + 318, + 268, + 342, + 274 + ], + "spans": [ + { + "bbox": [ + 318, + 268, + 342, + 274 + ], + "type": "text", + "content": "A can thr" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 309, + 274, + 351, + 316 + ], + "lines": [ + { + "bbox": [ + 309, + 274, + 351, + 316 + ], + "spans": [ + { + "bbox": [ + 309, + 274, + 351, + 316 + ], + "type": "image", + "image_path": "e08a81e17a58d551f93f4cd9eb9bf9ece3a23e23c149da6b6c1e1f665a7316db.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 365, + 91, + 408, + 133 + ], + "blocks": [ + { + "bbox": [ + 374, + 86, + 397, + 91 + ], + "lines": [ + { + "bbox": [ + 374, + 86, + 397, + 91 + ], + "spans": [ + { + "bbox": [ + 374, + 86, + 397, + 91 + ], + "type": "text", + "content": "DeepFloyd" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 365, + 91, + 408, + 133 + ], + "lines": [ + { + "bbox": [ + 365, + 91, + 408, + 133 + ], + "spans": [ + { + "bbox": [ + 365, + 91, + 408, + 133 + ], + "type": "image", + "image_path": "64d1bc6de9df69c3b6489c266458244477b6c5581016a61fc98459b86bef746a.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 372, + 134, + 398, + 139 + ], + "lines": [ + { + "bbox": [ + 372, + 134, + 398, + 139 + ], + "spans": [ + { + "bbox": [ + 372, + 134, + 398, + 139 + ], + "type": "text", + "content": "StyleGANXL" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_caption" + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 412, + 91, + 456, + 133 + ], + "blocks": [ + { + "bbox": [ + 413, + 86, + 447, + 91 + ], + "lines": [ + { + "bbox": [ + 413, + 86, + 447, + 91 + ], + "spans": [ + { + "bbox": [ + 413, + 86, + 447, + 91 + ], + "type": "text", + "content": "VQDiffusion" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 412, + 91, + 456, + 133 + ], + "lines": [ + { + "bbox": [ + 412, + 91, + 456, + 133 + ], + "spans": [ + { + "bbox": [ + 412, + 91, + 456, + 133 + ], + "type": "image", + "image_path": "52f32b7d1105a788025b0416e2c1cf562640a87f9ffd3bd5258f42ec65f18f67.jpg" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_body" + } + ], + "index": 37 + }, + { + "type": "image", + "bbox": [ + 412, + 139, + 455, + 181 + ], + "blocks": [ + { + "bbox": [ + 414, + 134, + 452, + 139 + ], + "lines": [ + { + "bbox": [ + 414, + 134, + 452, + 139 + ], + "spans": [ + { + "bbox": [ + 414, + 134, + 452, + 139 + ], + "type": "text", + "content": "GuidedDiffusion" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 412, + 139, + 455, + 181 + ], + "lines": [ + { + "bbox": [ + 412, + 139, + 455, + 181 + ], + "spans": [ + { + "bbox": [ + 412, + 139, + 455, + 181 + ], + "type": "image", + "image_path": "e7b8aeb2ec6a6ce68a2797374b002d91d0ea47aa5cb3f6258777ff4b876b365b.jpg" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 370, + 184, + 544, + 194 + ], + "lines": [ + { + "bbox": [ + 370, + 184, + 544, + 194 + ], + "spans": [ + { + "bbox": [ + 370, + 184, + 544, + 194 + ], + "type": "text", + "content": "(b) Manually chosen open source models (19 models)" + } + ] + } + ], + "index": 48, + "angle": 0, + "type": "image_caption" + } + ], + "index": 39 + }, + { + "type": "image", + "bbox": [ + 460, + 91, + 503, + 133 + ], + "blocks": [ + { + "bbox": [ + 470, + 86, + 493, + 91 + ], + "lines": [ + { + "bbox": [ + 470, + 86, + 493, + 91 + ], + "spans": [ + { + "bbox": [ + 470, + 86, + 493, + 91 + ], + "type": "text", + "content": "StyleGAN2" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 460, + 91, + 503, + 133 + ], + "lines": [ + { + "bbox": [ + 460, + 91, + 503, + 133 + ], + "spans": [ + { + "bbox": [ + 460, + 91, + 503, + 133 + ], + "type": "image", + "image_path": "da4130d5200db2a94b4074279315b43279828e5da2a2bf30dcad9390b1ce0eb9.jpg" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_body" + } + ], + "index": 41 + }, + { + "type": "image", + "bbox": [ + 460, + 139, + 503, + 181 + ], + "blocks": [ + { + "bbox": [ + 475, + 134, + 487, + 139 + ], + "lines": [ + { + "bbox": [ + 475, + 134, + 487, + 139 + ], + "spans": [ + { + "bbox": [ + 475, + 134, + 487, + 139 + ], + "type": "text", + "content": "Glide" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 460, + 139, + 503, + 181 + ], + "lines": [ + { + "bbox": [ + 460, + 139, + 503, + 181 + ], + "spans": [ + { + "bbox": [ + 460, + 139, + 503, + 181 + ], + "type": "image", + "image_path": "d7f0d9b425fba9ca111eb271725ffab93f8510480c37387aea1ad9086ba63d4b.jpg" + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_body" + } + ], + "index": 43 + }, + { + "type": "image", + "bbox": [ + 508, + 91, + 550, + 133 + ], + "blocks": [ + { + "bbox": [ + 525, + 86, + 533, + 91 + ], + "lines": [ + { + "bbox": [ + 525, + 86, + 533, + 91 + ], + "spans": [ + { + "bbox": [ + 525, + 86, + 533, + 91 + ], + "type": "text", + "content": "DiT" + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 508, + 91, + 550, + 133 + ], + "lines": [ + { + "bbox": [ + 508, + 91, + 550, + 133 + ], + "spans": [ + { + "bbox": [ + 508, + 91, + 550, + 133 + ], + "type": "image", + "image_path": "fc175043f62aa72dbea529adae15d4b9ab30ec1ed9675b72e654e175288e9064.jpg" + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "image_body" + } + ], + "index": 45 + }, + { + "type": "image", + "bbox": [ + 508, + 139, + 550, + 181 + ], + "blocks": [ + { + "bbox": [ + 524, + 134, + 534, + 139 + ], + "lines": [ + { + "bbox": [ + 524, + 134, + 534, + 139 + ], + "spans": [ + { + "bbox": [ + 524, + 134, + 534, + 139 + ], + "type": "text", + "content": "CIPS" + } + ] + } + ], + "index": 46, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 508, + 139, + 550, + 181 + ], + "lines": [ + { + "bbox": [ + 508, + 139, + 550, + 181 + ], + "spans": [ + { + "bbox": [ + 508, + 139, + 550, + 181 + ], + "type": "image", + "image_path": "410392cc44f32b435cf1546850125a2658df3053f13c091bc89b1f4bb159ff87.jpg" + } + ] + } + ], + "index": 47, + "angle": 0, + "type": "image_body" + } + ], + "index": 47 + }, + { + "type": "image", + "bbox": [ + 365, + 217, + 407, + 258 + ], + "blocks": [ + { + "bbox": [ + 378, + 211, + 393, + 217 + ], + "lines": [ + { + "bbox": [ + 378, + 211, + 393, + 217 + ], + "spans": [ + { + "bbox": [ + 378, + 211, + 393, + 217 + ], + "type": "text", + "content": "Dalle2" + } + ] + } + ], + "index": 49, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 365, + 217, + 407, + 258 + ], + "lines": [ + { + "bbox": [ + 365, + 217, + 407, + 258 + ], + "spans": [ + { + "bbox": [ + 365, + 217, + 407, + 258 + ], + "type": "image", + "image_path": "11b1f9cf60b78ad52a794a446d7ac4f29f65ef9e149766a7093a8363e338a34e.jpg" + } + ] + } + ], + "index": 50, + "angle": 0, + "type": "image_body" + } + ], + "index": 50 + }, + { + "type": "image", + "bbox": [ + 365, + 265, + 407, + 307 + ], + "blocks": [ + { + "bbox": [ + 376, + 259, + 396, + 265 + ], + "lines": [ + { + "bbox": [ + 376, + 259, + 396, + 265 + ], + "spans": [ + { + "bbox": [ + 376, + 259, + 396, + 265 + ], + "type": "text", + "content": "FLUX dev" + } + ] + } + ], + "index": 51, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 365, + 265, + 407, + 307 + ], + "lines": [ + { + "bbox": [ + 365, + 265, + 407, + 307 + ], + "spans": [ + { + "bbox": [ + 365, + 265, + 407, + 307 + ], + "type": "image", + "image_path": "475dbce8cf8f39c96856c26a14e28cb4c347ad99ebcf60710ed8a06066669514.jpg" + } + ] + } + ], + "index": 52, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 399, + 310, + 515, + 319 + ], + "lines": [ + { + "bbox": [ + 399, + 310, + 515, + 319 + ], + "spans": [ + { + "bbox": [ + 399, + 310, + 515, + 319 + ], + "type": "text", + "content": "(c) Commercial models (11 models)" + } + ] + } + ], + "index": 66, + "angle": 0, + "type": "image_caption" + } + ], + "index": 52 + }, + { + "type": "image", + "bbox": [ + 413, + 217, + 455, + 258 + ], + "blocks": [ + { + "bbox": [ + 414, + 212, + 449, + 217 + ], + "lines": [ + { + "bbox": [ + 414, + 212, + 449, + 217 + ], + "spans": [ + { + "bbox": [ + 414, + 212, + 449, + 217 + ], + "type": "text", + "content": "MidjourneyV6:" + } + ] + } + ], + "index": 53, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 413, + 217, + 455, + 258 + ], + "lines": [ + { + "bbox": [ + 413, + 217, + 455, + 258 + ], + "spans": [ + { + "bbox": [ + 413, + 217, + 455, + 258 + ], + "type": "image", + "image_path": "ad7c7fedf172621ef758bba396770ceb5b4b15d01d989f8384f8b4329781a5c3.jpg" + } + ] + } + ], + "index": 54, + "angle": 0, + "type": "image_body" + } + ], + "index": 54 + }, + { + "type": "image", + "bbox": [ + 411, + 266, + 455, + 307 + ], + "blocks": [ + { + "bbox": [ + 416, + 259, + 451, + 266 + ], + "lines": [ + { + "bbox": [ + 416, + 259, + 451, + 266 + ], + "spans": [ + { + "bbox": [ + 416, + 259, + 451, + 266 + ], + "type": "text", + "content": "Firefly Image2" + } + ] + } + ], + "index": 55, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 411, + 266, + 455, + 307 + ], + "lines": [ + { + "bbox": [ + 411, + 266, + 455, + 307 + ], + "spans": [ + { + "bbox": [ + 411, + 266, + 455, + 307 + ], + "type": "image", + "image_path": "ef5a72b79b986288854273ed87a7dec6f25c0515ffaa242cdc9adb4f839165da.jpg" + } + ] + } + ], + "index": 56, + "angle": 0, + "type": "image_body" + } + ], + "index": 56 + }, + { + "type": "image", + "bbox": [ + 460, + 217, + 503, + 258 + ], + "blocks": [ + { + "bbox": [ + 463, + 212, + 499, + 217 + ], + "lines": [ + { + "bbox": [ + 463, + 212, + 499, + 217 + ], + "spans": [ + { + "bbox": [ + 463, + 212, + 499, + 217 + ], + "type": "text", + "content": "MidjourneyV5 2" + } + ] + } + ], + "index": 57, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 460, + 217, + 503, + 258 + ], + "lines": [ + { + "bbox": [ + 460, + 217, + 503, + 258 + ], + "spans": [ + { + "bbox": [ + 460, + 217, + 503, + 258 + ], + "type": "image", + "image_path": "aa991346df578eeb395ee35990d987960343854f61dc320668f545a73e0cd742.jpg" + } + ] + } + ], + "index": 58, + "angle": 0, + "type": "image_body" + } + ], + "index": 58 + }, + { + "type": "image", + "bbox": [ + 460, + 265, + 503, + 307 + ], + "blocks": [ + { + "bbox": [ + 463, + 259, + 498, + 265 + ], + "lines": [ + { + "bbox": [ + 463, + 259, + 498, + 265 + ], + "spans": [ + { + "bbox": [ + 463, + 259, + 498, + 265 + ], + "type": "text", + "content": "Firefly Image3" + } + ] + } + ], + "index": 59, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 460, + 265, + 503, + 307 + ], + "lines": [ + { + "bbox": [ + 460, + 265, + 503, + 307 + ], + "spans": [ + { + "bbox": [ + 460, + 265, + 503, + 307 + ], + "type": "image", + "image_path": "bbccfa2ca88d61d08de5831fc11c6f7463100e834f77a4a353170e05ff30e38c.jpg" + } + ] + } + ], + "index": 60, + "angle": 0, + "type": "image_body" + } + ], + "index": 60 + }, + { + "type": "image", + "bbox": [ + 508, + 217, + 550, + 259 + ], + "blocks": [ + { + "bbox": [ + 522, + 212, + 537, + 217 + ], + "lines": [ + { + "bbox": [ + 522, + 212, + 537, + 217 + ], + "spans": [ + { + "bbox": [ + 522, + 212, + 537, + 217 + ], + "type": "text", + "content": "Dalle3" + } + ] + } + ], + "index": 61, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 508, + 217, + 550, + 259 + ], + "lines": [ + { + "bbox": [ + 508, + 217, + 550, + 259 + ], + "spans": [ + { + "bbox": [ + 508, + 217, + 550, + 259 + ], + "type": "image", + "image_path": "0b8ea5970c68a34b9f7ad0ed8a746394eaf18029cfac47be0388eefe418a9888.jpg" + } + ] + } + ], + "index": 62, + "angle": 0, + "type": "image_body" + } + ], + "index": 62 + }, + { + "type": "image", + "bbox": [ + 508, + 265, + 550, + 307 + ], + "blocks": [ + { + "bbox": [ + 517, + 260, + 541, + 265 + ], + "lines": [ + { + "bbox": [ + 517, + 260, + 541, + 265 + ], + "spans": [ + { + "bbox": [ + 517, + 260, + 541, + 265 + ], + "type": "text", + "content": "IdeogramV1" + } + ] + } + ], + "index": 63, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 508, + 265, + 550, + 307 + ], + "lines": [ + { + "bbox": [ + 508, + 265, + 550, + 307 + ], + "spans": [ + { + "bbox": [ + 508, + 265, + 550, + 307 + ], + "type": "image", + "image_path": "bf07435f7a74b9f29f483914d3cb78a88d28f7bbddd0c65b23e2c9f1c281c58c.jpg" + } + ] + } + ], + "index": 64, + "angle": 0, + "type": "image_body" + } + ], + "index": 64 + }, + { + "bbox": [ + 55, + 388, + 296, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 388, + 296, + 471 + ], + "spans": [ + { + "bbox": [ + 55, + 388, + 296, + 471 + ], + "type": "text", + "content": "To address these problems, we propose Community Forensics, a dataset that is significantly more diverse and comprehensive than those in prior works (Fig. 2). Our dataset contains images generated by: (a) thousands of systematically downloaded open-source latent diffusion models, (b) hand-selected open source models with various architectures, and (c) state-of-the-art commercial models." + } + ] + } + ], + "index": 68 + }, + { + "bbox": [ + 55, + 472, + 296, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 472, + 296, + 639 + ], + "spans": [ + { + "bbox": [ + 55, + 472, + 296, + 639 + ], + "type": "text", + "content": "To acquire large numbers of models, we sample images from thousands of text-to-image diffusion models hosted on a popular model-sharing website, Hugging Face [34]. We exploit the fact that these models use a common programming library [123] and thus can be sampled in a standardized way. A large fraction of them are extensions of Stable Diffusion [100], but collectively capture a variety of common model variations, such as in the architecture, image processing, and image content. We also sample images from many other open source models, including GANs [37], autoregressive models [39], and consistency models [79, 114]. To ensure sufficient diversity in the image content of the generated images, we sample generators using captions sourced from a mix of several existing datasets whenever possible." + } + ] + } + ], + "index": 69 + }, + { + "bbox": [ + 55, + 641, + 296, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 296, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 296, + 689 + ], + "type": "text", + "content": "Our dataset contains 4803 distinct models, roughly " + }, + { + "bbox": [ + 55, + 641, + 296, + 689 + ], + "type": "inline_equation", + "content": "34 \\times" + }, + { + "bbox": [ + 55, + 641, + 296, + 689 + ], + "type": "text", + "content": " more than the most extensive previous dataset that sample images from generative models [5, 7, 15, 32, 43, 87, 125, 133], and covers a variety of recent model designs (Fig. 2)." + } + ] + } + ], + "index": 70 + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": "We use this dataset to study generalization in the generated image detection problem. Our experiments confirm" + } + ] + } + ], + "index": 71 + }, + { + "bbox": [ + 313, + 388, + 555, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 388, + 555, + 448 + ], + "spans": [ + { + "bbox": [ + 313, + 388, + 555, + 448 + ], + "type": "text", + "content": "that increasing the diversity of data improves generalization, consistent with previous findings in other areas of computer vision [48, 107, 108]. In contrast to other work, our dataset aims to increase diversity via number of generators, an underexplored axis of diversity. Through experiments, we find:" + } + ] + } + ], + "index": 72 + }, + { + "bbox": [ + 314, + 449, + 555, + 628 + ], + "type": "list", + "angle": 0, + "index": 77, + "blocks": [ + { + "bbox": [ + 314, + 449, + 555, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 449, + 555, + 485 + ], + "spans": [ + { + "bbox": [ + 314, + 449, + 555, + 485 + ], + "type": "text", + "content": "- Classifiers trained on our dataset obtain strong performance, both on our newly proposed evaluations and on multiple previously-proposed benchmarks." + } + ] + } + ], + "index": 73 + }, + { + "bbox": [ + 314, + 486, + 555, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 486, + 555, + 544 + ], + "spans": [ + { + "bbox": [ + 314, + 486, + 555, + 544 + ], + "type": "text", + "content": "- Adding more generative models improves generalization. Fig. 1 demonstrates the performance of fake image detection when trained on samples from varying numbers of diffusion models. Notably, the performance improves as more models are added, even across different architectures." + } + ] + } + ], + "index": 74 + }, + { + "bbox": [ + 314, + 545, + 555, + 580 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 545, + 555, + 580 + ], + "spans": [ + { + "bbox": [ + 314, + 545, + 555, + 580 + ], + "type": "text", + "content": "- Including diverse generative model architectures significantly improves results, since classifiers do not fully generalize between generator architectures." + } + ] + } + ], + "index": 75 + }, + { + "bbox": [ + 314, + 581, + 555, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 581, + 555, + 628 + ], + "spans": [ + { + "bbox": [ + 314, + 581, + 555, + 628 + ], + "type": "text", + "content": "- Standard classifiers perform well. In contrast to observations from recent work, we find that end-to-end training of classifiers based on CNNs or ViTs generalizes well, with qualitatively similar to that of other recognition problems." + } + ] + } + ], + "index": 76 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 637, + 398, + 649 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 637, + 398, + 649 + ], + "spans": [ + { + "bbox": [ + 314, + 637, + 398, + 649 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 78 + }, + { + "bbox": [ + 313, + 654, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 555, + 713 + ], + "type": "text", + "content": "Datasets for detecting generated images. A number of datasets have been proposed for specifically detecting \"deepfake\" images containing manipulated faces [25, 63, 67, 71, 101, 102, 134]. Rather than focusing on face manipulation, we address creating general-purpose meth" + } + ] + } + ], + "index": 79 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8246" + } + ] + } + ], + "index": 80 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 504 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 504 + ], + "type": "text", + "content": "ods that can detect images that have been directly produced by generative models. Wang et al. [125] proposed a widely-used dataset of CNN-generated images, mixing images from GANs [8, 12, 58, 59, 90, 132] with other models [9, 10, 18, 70, 102]. This work showed that forensics models generalize between generative models, providing motivation for training on large datasets of diverse generators. However, their classifier was trained on images from a single GAN and was highly sensitive to data augmentation parameters, and more recent work shows that it does not generalize to newer models [14, 87]. Ojha et al. [87] introduced a dataset of recent diffusion models and found that training a linear classifier on CLIP features [96] extracted from ProGAN-generated images performed well. Cozzolino et al. [15] extend this work by studying the performance of CLIP-based detectors on various generative models and datasets. Epstein et al. [32] simulated detecting fake images in an online way by training a detector up to a certain year and testing it on generators released after that year. Zhu et al. [133] collected 1.4M generated images from 8 different generators. These datasets, however, only consider a handful of models (less than 20 each), limiting the generalization of their detectors. Asnani et al. proposed RED116 [5], collecting 116K images from 116 generative models for predicting the model hyperparameters of a given generated image. Guo et al. [43] extends RED116 to 140 generative models with 140K images total. We improve upon these works by collecting much more diverse generative models to improve the performance and generalization of the detector. In concurrent work, Hong et al. [49] acquires user-created images from Midjourney and CivitAI. This strategy is complementary to ours: while it aims to collect in-the-wild fake images, its distribution is centered on images that users share, and the models are not necessarily identifiable, making it challenging to rigorously analyze the dataset's contents and to interpret experiments conducted on it." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 510, + 298, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 298, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 298, + 714 + ], + "type": "text", + "content": "Fingerprint-based image forensics methods. Classic work on image forensics relied on methods based on image statistics [95] and physical constraints [56], rather than learning. A number of datasets have been created for detecting images that have been manipulated using traditional methods, such as with photo editors [21, 26, 53, 64, 84]. Recent works focus on detecting synthetic images by inspecting the generator fingerprints. Zhang et al. [131] and Marra et al. [80] proposed identifying the spatial fingerprints left by the generator to detect synthetic images. Others focus on spectral anomalies to detect synthetic images. Durall et al. [29] and Dzanic et al. [30] identified that CNN-generated images fail to reproduce certain spectral properties of real images. Corvi et al. [14] study the frequency fingerprints of the generated images and analyzes the cross-architecture generalization of the detector. Bammey [7] uses high-frequency artifacts to detect generated images. However, these approaches may be" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 313, + 72, + 556, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 556, + 193 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 556, + 193 + ], + "type": "text", + "content": "brittle since the artifacts they rely on can be eliminated by post-processing [15]. We instead approach this problem in a data-driven manner, scaling the number of models, images, and architectures. Recent work has created ensembles of fake image classifiers [50]. In parallel, researchers have detected text generated by language models using supervised learning and heuristics [6, 35, 55, 65, 82, 103, 113, 121], which closely resemble those in visual forensics. However, no existing techniques that we are aware of aim to collect comprehensive datasets of community-created generators." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 313, + 196, + 557, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 196, + 557, + 353 + ], + "spans": [ + { + "bbox": [ + 313, + 196, + 557, + 353 + ], + "type": "text", + "content": "Out-of-distribution generalization. Our work is related to the out-of-distribution recognition problem as it involves generalizing to unseen generators and image processing pipelines. A variety of approaches have been proposed for this problem, based on likelihood ratios [72, 98, 128], self-supervision [47, 83, 109, 124], internal model statistics [46, 104], temperature scaling [3, 73], and via energy-based models [28, 31, 75]. Works by Schuhmann et al. [108] and Hendrycks et al. [48] show that diverse training data and data augmentation are important to improving the robustness to out-of-distribution samples. Our results are in line with these conclusions, as we find that a diverse set of generative models and stronger augmentations improve generalization." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 362, + 503, + 376 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 362, + 503, + 376 + ], + "spans": [ + { + "bbox": [ + 313, + 362, + 503, + 376 + ], + "type": "text", + "content": "3. The Community Forensics Dataset" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 380, + 556, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 380, + 556, + 571 + ], + "spans": [ + { + "bbox": [ + 313, + 380, + 556, + 571 + ], + "type": "text", + "content": "To support our goal of studying generalization in generated image detection, we collect a dataset of images sampled from a wide range of models (Fig. 2). Our dataset consists of: (a) a large and systematically collected set of \"in-the-wild\" text-to-image latent diffusion models obtained from a model-sharing website, (b) hand-selected models from other open source architectures, and (c) closed and open state-of-the-art commercial models. We also pair these generated images with real images from other datasets. For all images in our dataset, we preserve the original image format where possible, without any additional compression or resizing. This is to mitigate potential bias and performance degradation in out-of-distribution settings due to unwanted artifacts [40, 42, 97]. Our dataset contains significantly more models than previous works (Tab. 1) and spans a wider range of architectures, processing pipelines, and semantic content." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 575, + 539, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 575, + 539, + 588 + ], + "spans": [ + { + "bbox": [ + 313, + 575, + 539, + 588 + ], + "type": "text", + "content": "3.1. Systematically collecting generative models" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 590, + 556, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 590, + 556, + 687 + ], + "spans": [ + { + "bbox": [ + 313, + 590, + 556, + 687 + ], + "type": "text", + "content": "We perform our systematic collection using publicly available, open source models that use the Hugging Face diffusers library [34, 123] because: 1) it is a popular library for creating text-to-image models and is widely used by hobbyists, 2) thousands of such models are publicly indexed, and 3) it provides a standard interface by which we can sample images. We process the models in the order of popularity, as indicated by the number of downloads. Our" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 313, + 693, + 555, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 693, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 693, + 555, + 714 + ], + "type": "text", + "content": "1We use \"open source\" to describe models with public weights and code, even if they may be closed in some aspect (e.g., private training data)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8247" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 70, + 553, + 198 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 553, + 198 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 553, + 198 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 553, + 198 + ], + "type": "table", + "html": "
DatasetModelsImagesArchitecturesTraining setup
Wang et al. [125]11362KGAN, Perceptual, Deepfake, ...ProGAN [58] vs. LSUN [129]
Ojha et al. [87]4*10K*GAN, Perceptual, Diffusion, ...ProGAN [58] vs. LSUN [129]
Epstein et al. [32]14570KDiffusionDiffusion vs. LAION [108]
Cozzolino et al. [15]1826KDiffusionLDM [100] vs. MS-COCO [74]
Synthbuster [7]910KDiffusionDiffusion vs. Dresden [36]
GenImage [133]81.4MDiffusion, GANDiffusion, GAN vs. ImageNet [23]
RED116 [5]116116KGAN, VAE, Autoregressive, ...Many vs. Many
RED140 [43]140140KDiffusion, GAN, VAE, ...Many vs. Many
Ours48032.7MDiffusion, GAN, Autoregressive, ...Many vs. Many
", + "image_path": "5412ad9454e799c6beca1aee8f12485a99c281b83568c9ef69c3ebdea8131176.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 205, + 555, + 240 + ], + "lines": [ + { + "bbox": [ + 55, + 205, + 555, + 240 + ], + "spans": [ + { + "bbox": [ + 55, + 205, + 555, + 240 + ], + "type": "text", + "content": "Table 1. Comparison with existing forensics datasets. We compare the size of the dataset with existing datasets containing identifiable generative models. We only count the number of generated images. Our dataset contains significantly more generative models than prior works. " + }, + { + "bbox": [ + 55, + 205, + 555, + 240 + ], + "type": "inline_equation", + "content": "\\ast" + }, + { + "bbox": [ + 55, + 205, + 555, + 240 + ], + "type": "text", + "content": ": Only counting the unique evaluation set by Ojha et al. [87] as their dataset is based on Wang et al. [125]." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 248, + 294, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 248, + 294, + 380 + ], + "spans": [ + { + "bbox": [ + 54, + 248, + 294, + 380 + ], + "type": "text", + "content": "pipeline downloads each model and extracts relevant hyperparameters (e.g., number of diffusion steps and guidance scale), sampling pipeline configurations, and metadata, from both the model cards on the model-sharing webpage and the metadata retrieved by the diffusers library [34, 123]. We sample images using a distribution of text prompts obtained from various real datasets (Sec. 3.3). Since experiments suggest diminishing returns for repeatedly sampling from any given model, we sample a few hundred images from each one. We sample 4763 models with approximately 403 images each, for a total of " + }, + { + "bbox": [ + 54, + 248, + 294, + 380 + ], + "type": "inline_equation", + "content": "1.9\\mathrm{M}" + }, + { + "bbox": [ + 54, + 248, + 294, + 380 + ], + "type": "text", + "content": " images from this process." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 381, + 295, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 381, + 295, + 548 + ], + "spans": [ + { + "bbox": [ + 55, + 381, + 295, + 548 + ], + "type": "text", + "content": "While the lack of documentation in each model and the scale of data collection make it challenging to exactly characterize the model designs in this set, they appear to be entirely (or almost entirely) based on latent diffusion. More specifically, we categorize models as being based on latent diffusion if they perform a denoising process on a latent representation. Based on this criterion and the self-reported tags, all models in our systematically collected set appear to be based on latent diffusion. While pixel-based diffusion models also use the diffusers library (e.g., DeepFloyd [22]), they were incompatible with our automated generation pipeline. We record such incompatible models and manually sample them to either construct an out-of-distribution test set (Sec. 3.4), or as manually-chosen models for training data (Sec. 3.2)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 550, + 295, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 550, + 295, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 550, + 295, + 693 + ], + "type": "text", + "content": "We show examples of the sampled images in Fig. 2. In the supplement, we provide examples of models and information from their project pages. These models generate a variety of images, with various types of semantic content and preprocessing. For example, a large fraction of these models adapt variations of a popular pretrained latent diffusion model, Stable Diffusion [100], to different downstream applications, and use a number of adaptation strategies (e.g., using LoRA [51]). We provide the model metadata with each image to enable other possible forensics applications. We discuss these in the supplement and provide information about image and model licenses." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 248, + 537, + 261 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 248, + 537, + 261 + ], + "spans": [ + { + "bbox": [ + 313, + 248, + 537, + 261 + ], + "type": "text", + "content": "3.2. Collecting images from other architectures" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 263, + 555, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 263, + 555, + 431 + ], + "spans": [ + { + "bbox": [ + 313, + 263, + 555, + 431 + ], + "type": "text", + "content": "Images from manually chosen models. To ensure that our dataset contains a broader range of models, we manually select 19 models from public repositories and sample an average of 40,738 images per model, resulting in 774K images total. We note that this number is itself on par with (or more than) many prior datasets with identifiable generative models [7, 15, 32, 87, 125, 133]. We include several GANs (e.g., StyleGANs [60-62, 106], BigGAN [8], StyleSwin [130], GigaGAN [57], ProGAN [58], ProjectedGAN [105], GANformer [52], SAN [115], and CIPS [4]), pixel-based diffusion models (e.g., GLIDE [85], ADM [24], and DeepFloyd [22]), latent diffusion models (e.g., VQ-Diffusion [41], Diffusion Transformers [92], and Latent Flow Matching [20]), and an autoregressive model (Taming Transformers [33])." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 437, + 556, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 556, + 511 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 556, + 511 + ], + "type": "text", + "content": "Images from commercial models. We sample 15K images from 11 commercial models using LAION-based captions to evaluate the generalization to state-of-the-art models with typically unknown architectures: DALL-E 2, 3 [88, 89], Ideogram V1, V2 [2], Midjourney V5, V6 [81], Firefly Image 2, 3 [1], FLUX.1-dev, Schnell [68], and Imagen 3 [38]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 516, + 441, + 529 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 516, + 441, + 529 + ], + "spans": [ + { + "bbox": [ + 313, + 516, + 441, + 529 + ], + "type": "text", + "content": "3.3. Collecting real images" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 532, + 556, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 532, + 556, + 594 + ], + "spans": [ + { + "bbox": [ + 313, + 532, + 556, + 594 + ], + "type": "text", + "content": "To help study how real images influence forensics models, we source real images from a variety of existing datasets: LAION [107], ImageNet [23], COCO [74], FFHQ [59], CelebA [76], MetFaces [60], AFHQ [13], Forchheim [44], IMD2020 [86], Landscapes HQ [112], and VISION [111].3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 598, + 464, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 598, + 464, + 611 + ], + "spans": [ + { + "bbox": [ + 313, + 598, + 464, + 611 + ], + "type": "text", + "content": "3.4. Curating the evaluation set" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 614, + 556, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 614, + 556, + 675 + ], + "spans": [ + { + "bbox": [ + 313, + 614, + 556, + 675 + ], + "type": "text", + "content": "We construct our evaluation set using the incompatible models from our automated sampling pipeline, commercial models (Sec. 3.2), and manually collected open source models. The evaluation set comprises 26K images sampled from 21 models not included in the training set." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 66, + 702, + 288, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 702, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 66, + 702, + 288, + 714 + ], + "type": "text", + "content": "2We note that this definition includes latent consistency models [78]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 683, + 555, + 715 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 683, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 683, + 555, + 715 + ], + "type": "text", + "content": "3Following common convention, we refer to these images as real images, even though they may be synthetic (e.g., graphic design). More precisely, our goal is to distinguish \"AI-generated\" images from the originals." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8248" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 54, + 72, + 296, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 72, + 296, + 156 + ], + "spans": [ + { + "bbox": [ + 54, + 72, + 296, + 156 + ], + "type": "text", + "content": "This includes our commercial models set and an additional 11K images from 10 models: Deci Diffusion V2 [120], GALIP [119], KandinskyV2.2 [110], Kvikontent [66], LCM-LoRA-SDv1.5, LCM-LoRA-SDXL, LCM-LoRA-SSD1B [79], Stable Cascade [93], DF-GAN [118], and HDiT [16], sampled using RAISE [19], ImageNet [23], FFHQ [59], and COCO [74]-based captions." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 157, + 297, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 157, + 297, + 323 + ], + "spans": [ + { + "bbox": [ + 56, + 157, + 297, + 323 + ], + "type": "text", + "content": "The generated images are paired with the source real data that are used to prompt the generators. However, since some of the real datasets do not have appropriate licenses for redistribution (e.g., LAION [107, 108]), we created a public version of our evaluation set by pairing the generated images with openly licensed COCO [74] and FFHQ [59] which allow redistribution for non-commercial purposes. The public version of our evaluation set will serve as an easily reproducible and shareable evaluation set that will complement our default set. We will refer to our default set as the comprehensive evaluation set. We will also release the instructions to reconstruct our comprehensive set. However, note that it may not be possible to exactly reconstruct this set in the future due to link rot." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 328, + 167, + 341 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 328, + 167, + 341 + ], + "spans": [ + { + "bbox": [ + 55, + 328, + 167, + 341 + ], + "type": "text", + "content": "3.5. Generating images" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 343, + 297, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 343, + 297, + 487 + ], + "spans": [ + { + "bbox": [ + 55, + 343, + 297, + 487 + ], + "type": "text", + "content": "Unconditional models are sampled until we reach the desired number of images. For class conditional models, we sample an equal number of images per class. Text-conditional models are sampled using captions obtained from real images (Sec. 3.3). We either use captions that are already present in the dataset or use BLIP [69] to generate them. Some models such as GigaGAN [57] and HDiT [16] do not provide a pretrained model, so we instead use their pre-generated images. Generated images are saved in PNG format to avoid compression artifacts. However, Firefly [1] generated images are saved in JPEG format as their web UI only allows downloading in JPEG." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 492, + 137, + 506 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 492, + 137, + 506 + ], + "spans": [ + { + "bbox": [ + 55, + 492, + 137, + 506 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 509, + 296, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 509, + 296, + 594 + ], + "spans": [ + { + "bbox": [ + 55, + 509, + 296, + 594 + ], + "type": "text", + "content": "We use our dataset to conduct a study of generalization in visual forensics, asking a number of questions: (1) How well do forensics models trained on our dataset generalize to unseen models? (2) Does adding more models improve detection performance? (3) How does diversity of the training data affect performance? (4) What architectures and data augmentation schemes are most successful?" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 597, + 230, + 610 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 597, + 230, + 610 + ], + "spans": [ + { + "bbox": [ + 55, + 597, + 230, + 610 + ], + "type": "text", + "content": "4.1. Training image forensics models" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 613, + 296, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 613, + 296, + 662 + ], + "spans": [ + { + "bbox": [ + 55, + 613, + 296, + 662 + ], + "type": "text", + "content": "We train binary classifiers that detect generated images using our dataset to study the generalization in image forensics. We construct our training set of 5.4M images by pairing 2.7M generated images with 2.7M real images." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 665, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 297, + 714 + ], + "type": "text", + "content": "Training and evaluation setup. We evaluate the models trained on our dataset and compare them with prior works [43, 87, 125, 133]. As RED140 [43] already contains RED116 [5], we do not compare with RED116 dataset." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 556, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 556, + 348 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 556, + 348 + ], + "type": "text", + "content": "Following prior works [87, 125], we use the threshold-independent mean average precision (mAP) and accuracy (Acc.) as our evaluation metrics. We compute the mAP and accuracy by averaging the results of each generative model. We use five evaluation sets: Wang et al. [125], Ojha et al. [87], Synthbuster [7], GenImage [133], and our evaluation set. All evaluation sets apart from GenImage [133] evaluate out-of-distribution performance for all classifiers. GenImage [133] evaluation set, however, contains the same set of generators used in training, and is an in-distribution evaluation set for their classifiers. Concretely, the evaluation set by Wang et al. [125] and Ojha et al. [87, 125] contains models such as CRN [10], CycleGAN [132], DALL-E [88], DeepFake [25], IMLE [70], SAN [18], StarGAN [12], and SITD [9] which are unseen by both their and our classifiers. Synthbuster [7] evaluation set is comprised of RAISE [19]-based synthetic images of DALL-E [88, 89], Firefly [1], Glide [85], Midjourney [81], and Stable Diffusion [94, 100], and is mostly out of distribution for all classifiers. GenImage [133] evaluation set is a validation split of their training set; they use an identical set of models to train their classifier: ADM [24], BigGAN [8], Glide [85], Midjourney [81], Stable Diffusion [100], VQ-Diffusion [41], and Wukong [127]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 357, + 556, + 573 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 357, + 556, + 573 + ], + "spans": [ + { + "bbox": [ + 313, + 357, + 556, + 573 + ], + "type": "text", + "content": "Model architecture. Building on prior works which mainly used CLIP-ViT [27, 54, 96] and ResNet-50 [45], we consider ViT [27] and ConvNeXt [77] pretrained models for our classifiers. We use a plain ViT-S backbone [27] pretrained on CLIP objective [54, 96] using LAION-2B [108], ImageNet 21K, and ImageNet 1K datasets [23]. We also experiment with a ConvNeXt-S model [77] pretrained on ImageNet 21K and ImageNet 1K datasets [23]. We replace the classification head with a linear layer with sigmoid activation that outputs the probability of the image being generated. Unlike prior works [15, 87] that freeze the CLIP-ViT backbone, we train the backbone end-to-end. The models are obtained through timm [11, 126] library on Hugging Face. We experiment with two input resolutions, " + }, + { + "bbox": [ + 313, + 357, + 556, + 573 + ], + "type": "inline_equation", + "content": "224^2" + }, + { + "bbox": [ + 313, + 357, + 556, + 573 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 357, + 556, + 573 + ], + "type": "inline_equation", + "content": "384^2" + }, + { + "bbox": [ + 313, + 357, + 556, + 573 + ], + "type": "text", + "content": ", to evaluate the impact of the input resolution on the detector's performance. We denote the detector with " + }, + { + "bbox": [ + 313, + 357, + 556, + 573 + ], + "type": "inline_equation", + "content": "384^2" + }, + { + "bbox": [ + 313, + 357, + 556, + 573 + ], + "type": "text", + "content": " input resolution as High res. We implement the models using PyTorch [91]. The hyperparameters are detailed in the supplement." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 582, + 557, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 557, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 557, + 715 + ], + "type": "text", + "content": "Data augmentation. Prior work considered augmentations that were designed to simulate postprocessing, such as flipping, cropping, Gaussian blur, and JPEG recompression to train their detectors [7, 15, 87, 125]. We propose an augmentation scheme that extends this approach and compare it with previously proposed augmentation methods. We expand the set of augmentations to handle additional transformations that can occur in the wild, such as padding, resizing, rotation, and shear, and integrate them into a framework that can apply complex sequences of transformations. We introduce a modified version of RandAugment [17] that applies a" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8249" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 60, + 70, + 555, + 175 + ], + "blocks": [ + { + "bbox": [ + 60, + 70, + 555, + 175 + ], + "lines": [ + { + "bbox": [ + 60, + 70, + 555, + 175 + ], + "spans": [ + { + "bbox": [ + 60, + 70, + 555, + 175 + ], + "type": "table", + "html": "
ModelEvaluation Set (mAP)Evaluation Set (Acc)
Wang et al. [125]Ojha et al. [87]SB [7]GenImage [133]OursMeanWang et al. [125]Ojha et al. [87]SB [7]GenImage [133]OursMean
Wang et al. [125]0.8970.6960.5160.6420.5370.6000.6480.7140.5270.5080.5330.5130.5170.552
Ojha et al. [87]0.9390.9570.6200.7970.5920.6560.7600.7910.8210.5320.6410.5400.5480.646
GenImage [133]0.9290.9840.8130.9990.9120.9680.9340.7950.9660.7190.9900.8180.8860.862
RED140 [43] - High res.0.9000.9540.7650.9270.7640.8610.8620.6940.7800.5580.6740.5620.5650.639
Ours0.9640.9910.9040.9900.9710.9770.9660.8730.9500.8180.9460.8610.8880.889
Ours - High res.0.9670.9960.9740.9980.9870.9940.9860.9010.9700.9080.9570.8920.9120.923
", + "image_path": "6436d7f66a12015bf915de151eb2fa7d6abc03563f100037b3d21b4e3c0236cc.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 178, + 555, + 266 + ], + "lines": [ + { + "bbox": [ + 55, + 178, + 555, + 266 + ], + "spans": [ + { + "bbox": [ + 55, + 178, + 555, + 266 + ], + "type": "text", + "content": "Table 2. Generalization of AI-generated image detectors across benchmarks. We evaluate the classifiers trained on our dataset on several benchmarks, including our own. We also evaluate several previously released classifiers. Our Comprehensive set (abbreviated as Comp.) pairs the generated images with original real data; the Public set pairs them with openly licensed COCO [74] and FFHQ [59] for license-compliant redistribution of the evaluation set (Sec. 3.4). We use plain CLIP-ViT-S [27, 54, 96] architecture with " + }, + { + "bbox": [ + 55, + 178, + 555, + 266 + ], + "type": "inline_equation", + "content": "224^{2}" + }, + { + "bbox": [ + 55, + 178, + 555, + 266 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 178, + 555, + 266 + ], + "type": "inline_equation", + "content": "384^{2}" + }, + { + "bbox": [ + 55, + 178, + 555, + 266 + ], + "type": "text", + "content": " (High res.) input resolutions, Wang et al. [125] and GenImage [133] use ResNet-50 [45] with " + }, + { + "bbox": [ + 55, + 178, + 555, + 266 + ], + "type": "inline_equation", + "content": "224^{2}" + }, + { + "bbox": [ + 55, + 178, + 555, + 266 + ], + "type": "text", + "content": " input resolution, and Ojha et al. [87] uses CLIP-ViT-L with " + }, + { + "bbox": [ + 55, + 178, + 555, + 266 + ], + "type": "inline_equation", + "content": "224^{2}" + }, + { + "bbox": [ + 55, + 178, + 555, + 266 + ], + "type": "text", + "content": " input resolution as the backbone. Our classifiers show robust performance across all evaluation sets, outperforming all baselines in out-of-distribution settings ([7, 87, 125] and Ours) and nearly matches GenImage [133] on its in-distribution evaluation set. " + }, + { + "bbox": [ + 55, + 178, + 555, + 266 + ], + "type": "inline_equation", + "content": "\\spadesuit" + }, + { + "bbox": [ + 55, + 178, + 555, + 266 + ], + "type": "text", + "content": ": RED140 [43] is trained following our training procedure with " + }, + { + "bbox": [ + 55, + 178, + 555, + 266 + ], + "type": "inline_equation", + "content": "384^{2}" + }, + { + "bbox": [ + 55, + 178, + 555, + 266 + ], + "type": "text", + "content": " input resolution as they do not provide a trained classifier." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 56, + 274, + 294, + 345 + ], + "blocks": [ + { + "bbox": [ + 56, + 274, + 294, + 345 + ], + "lines": [ + { + "bbox": [ + 56, + 274, + 294, + 345 + ], + "spans": [ + { + "bbox": [ + 56, + 274, + 294, + 345 + ], + "type": "image", + "image_path": "962707b7f8c50d48ce8f78f437e2c69577554671c1e94e0602d66f8aeaec0aeb.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 351, + 296, + 429 + ], + "lines": [ + { + "bbox": [ + 55, + 351, + 296, + 429 + ], + "spans": [ + { + "bbox": [ + 55, + 351, + 296, + 429 + ], + "type": "text", + "content": "Figure 3. Performance across generator types. We evaluate the classifier performance across five generator types - latent diffusion, pixel diffusion, commercial models, GANs, and other architecture (Stable Cascade [93]), and report the mean. We additionally evaluate three forensics methods, Aerblade [99], LGrad [116], and NPR [117]. Our classifiers show robust performance across all generator types, whereas prior works struggle to generalize." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 439, + 296, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 439, + 296, + 547 + ], + "spans": [ + { + "bbox": [ + 55, + 439, + 296, + 547 + ], + "type": "text", + "content": "randomly-ordered sequence of augmentations to the images. Specifically, our modified RandAugment samples a random number " + }, + { + "bbox": [ + 55, + 439, + 296, + 547 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 439, + 296, + 547 + ], + "type": "text", + "content": " between 0 and " + }, + { + "bbox": [ + 55, + 439, + 296, + 547 + ], + "type": "inline_equation", + "content": "n_{\\mathrm{max}}" + }, + { + "bbox": [ + 55, + 439, + 296, + 547 + ], + "type": "text", + "content": " for each augmentation type. Then, it applies the augmentations in random order until " + }, + { + "bbox": [ + 55, + 439, + 296, + 547 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 439, + 296, + 547 + ], + "type": "text", + "content": " augmentations are applied for each augmentation type to the image. We use various augmentations, including in-memory JPEG compression, random resizing with random interpolation methods, cropping, flipping, rotation, translation, shear, padding, and cutout." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 554, + 240, + 565 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 554, + 240, + 565 + ], + "spans": [ + { + "bbox": [ + 55, + 554, + 240, + 565 + ], + "type": "text", + "content": "4.2. Generalization across benchmarks" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "content": "We first evaluate how well classifiers trained on our dataset perform across benchmarks. In Table 2, we observe that our models outperform the prior works [43, 87, 125, 133] on all evaluation sets except GenImage. This is expected since the GenImage evaluation set is a validation split of their training set; all of the generators are already seen by their classifier. On all other unseen benchmarks, our classifiers outperform all prior works. Notably, our classifiers achieve very high performance (0.987 mAP and " + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "inline_equation", + "content": "89.2\\%" + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "content": " accuracy) on our out-of-distribution, comprehensive evaluation set, with a significant margin over prior works. This gap in performance can be traced to our training data which in" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 316, + 274, + 434, + 362 + ], + "blocks": [ + { + "bbox": [ + 316, + 274, + 434, + 362 + ], + "lines": [ + { + "bbox": [ + 316, + 274, + 434, + 362 + ], + "spans": [ + { + "bbox": [ + 316, + 274, + 434, + 362 + ], + "type": "image", + "image_path": "b2050c6c5265689e96c9dc3268ab08e1cb231abb9c6ddef646fe10e377ee3cc7.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 369, + 555, + 468 + ], + "lines": [ + { + "bbox": [ + 313, + 369, + 555, + 468 + ], + "spans": [ + { + "bbox": [ + 313, + 369, + 555, + 468 + ], + "type": "text", + "content": "Figure 4. Performance with increasing number of images. We train a classifier with varying numbers of images from two sets: 1000 randomly chosen models and 10 popular (highly downloaded) models in the systematically collected subset. We report the mean and standard error bands for each data point across 4 randomly sampled subsets. The classifier trained from 1000 random models outperforms 10 popular models in all cases. Notably, the accuracy gap is wider than that of mAP, which may suggest that having a diverse set of models improves accuracy threshold calibration." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 435, + 274, + 552, + 363 + ], + "blocks": [ + { + "bbox": [ + 435, + 274, + 552, + 363 + ], + "lines": [ + { + "bbox": [ + 435, + 274, + 552, + 363 + ], + "spans": [ + { + "bbox": [ + 435, + 274, + 552, + 363 + ], + "type": "image", + "image_path": "71f687f50ab6ed2107f86a86ca9738dc0eeb779d3afd5d7e956d3ee90cb40b2a.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 477, + 554, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 477, + 554, + 525 + ], + "spans": [ + { + "bbox": [ + 313, + 477, + 554, + 525 + ], + "type": "text", + "content": "corporates a substantially richer variety of generators than existing works. Consequently, our classifiers demonstrate robust generalization to out-of-distribution data, where prior works often struggle." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 525, + 555, + 645 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 525, + 555, + 645 + ], + "spans": [ + { + "bbox": [ + 313, + 525, + 555, + 645 + ], + "type": "text", + "content": "To better illustrate the generalization of the classifiers, in Figure 3, we split our comprehensive evaluation set into five subsets and evaluate them: latent diffusion, pixel diffusion, commercial models, GANs, and other architecture type (Stable Cascade [93]). Furthermore, we evaluate three additional forensics methods: Aerblade [99], LGrad [116], and NPR [117]. Our classifiers show strong performance across all generator types, unlike prior works which struggle to generalize. For the following experiments, we use our best-performing model (High res.) unless stated otherwise." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 650, + 457, + 662 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 650, + 457, + 662 + ], + "spans": [ + { + "bbox": [ + 313, + 650, + 457, + 662 + ], + "type": "text", + "content": "4.3. Impact of model diversity" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "type": "text", + "content": "Next, we examine the impact of the number of models in training data. We train classifiers with images sampled from 3 to 3333 generators and evaluate them (Fig. 1). To ensure that the gains are not due to simply sampling qualita" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8250" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 69, + 175, + 150 + ], + "blocks": [ + { + "bbox": [ + 58, + 69, + 175, + 150 + ], + "lines": [ + { + "bbox": [ + 58, + 69, + 175, + 150 + ], + "spans": [ + { + "bbox": [ + 58, + 69, + 175, + 150 + ], + "type": "image", + "image_path": "de3a685a1acf23bf0081bb9d302fdcdecba91a2e02a740ea8299ba0c369f2745.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 56, + 153, + 174, + 163 + ], + "lines": [ + { + "bbox": [ + 56, + 153, + 174, + 163 + ], + "spans": [ + { + "bbox": [ + 56, + 153, + 174, + 163 + ], + "type": "text", + "content": "(a) Impact of generator type diversity" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 178, + 69, + 294, + 151 + ], + "blocks": [ + { + "bbox": [ + 178, + 69, + 294, + 151 + ], + "lines": [ + { + "bbox": [ + 178, + 69, + 294, + 151 + ], + "spans": [ + { + "bbox": [ + 178, + 69, + 294, + 151 + ], + "type": "image", + "image_path": "86a32a5591270b141562b2d967f9c61148be4d1ec7231ff9414b918c18f904f7.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 178, + 153, + 293, + 163 + ], + "lines": [ + { + "bbox": [ + 178, + 153, + 293, + 163 + ], + "spans": [ + { + "bbox": [ + 178, + 153, + 293, + 163 + ], + "type": "text", + "content": "(b) Classifier backbone comparison" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 170, + 296, + 258 + ], + "lines": [ + { + "bbox": [ + 55, + 170, + 296, + 258 + ], + "spans": [ + { + "bbox": [ + 55, + 170, + 296, + 258 + ], + "type": "text", + "content": "Figure 5. (a) Performance and model diversity. We compare detection performance for commercial models using classifiers trained on different subsets of the dataset: the systematically collected latent diffusion models, the manually chosen models containing diverse generator types, and both. As diversity increases, so does performance. (b) Classifier backbone comparison. We compare the architectures across datasets: ours, GenImage [133], and Wang et al. [125]. Performance is similar between architectures." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 270, + 295, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 270, + 295, + 448 + ], + "spans": [ + { + "bbox": [ + 54, + 270, + 295, + 448 + ], + "type": "text", + "content": "tively different architectures, we only use our systematically collected latent diffusion models. For each data point, we sample 10 random subsets of models with 100K training images each and report the mean and standard error bands. We use an extended evaluation set that includes non-latent diffusion generators from our training set, which allows us to comprehensively assess the generalization capability of the classifiers trained exclusively on latent diffusion models. We find that the performance steadily increases with the number of models. However, the performance begins to flatten out beyond 1000 models, suggesting diminishing returns. Interestingly, the performance also improves on out-of-distribution architectures such as GANs and pixel-based diffusion models, even though the classifier is only trained on latent diffusion models." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 449, + 295, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 449, + 295, + 593 + ], + "spans": [ + { + "bbox": [ + 55, + 449, + 295, + 593 + ], + "type": "text", + "content": "In Figure 4, we vary the number of images from two sets: 1000 randomly chosen models and 10 popular models (as denoted by their number of downloads) downloaded from our systematically collected diffusion models. While the results show that the performance improves with more training images, it begins to plateau at approximately " + }, + { + "bbox": [ + 55, + 449, + 295, + 593 + ], + "type": "inline_equation", + "content": "27\\mathrm{K}" + }, + { + "bbox": [ + 55, + 449, + 295, + 593 + ], + "type": "text", + "content": " images. Moreover, the classifier trained on 1000 models outperforms the 10 models in all cases, indicating that model diversity is important for strong performance. We also note that the accuracy gap is noticeably wider than that of mAP, which may suggest that model diversity is crucial in calibrating the accuracy thresholds of the classifiers." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 594, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 297, + 714 + ], + "type": "text", + "content": "Our experiments show that the performance improvements from increasing the number of models may plateau when they are limited to a single generator type (Fig. 1). In Figure 5a, we show that the diversity of the generator type also plays a major role in generalization. We train classifiers on three different sets of training data: our systematically collected set, manually chosen set, and a full set consisting of both subsets. The systematic set comprises entirely of latent diffusion models, and the manual set contains numerous generator types, including GANs, latent and pixel-based dif" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 317, + 69, + 432, + 151 + ], + "blocks": [ + { + "bbox": [ + 317, + 69, + 432, + 151 + ], + "lines": [ + { + "bbox": [ + 317, + 69, + 432, + 151 + ], + "spans": [ + { + "bbox": [ + 317, + 69, + 432, + 151 + ], + "type": "image", + "image_path": "0af5c5242036ada06279f525959cfa606d6065e317152240bfc22a526c1dbddc.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 323, + 153, + 425, + 163 + ], + "lines": [ + { + "bbox": [ + 323, + 153, + 425, + 163 + ], + "spans": [ + { + "bbox": [ + 323, + 153, + 425, + 163 + ], + "type": "text", + "content": "(a) Impact of frozen backbones" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 170, + 555, + 270 + ], + "lines": [ + { + "bbox": [ + 313, + 170, + 555, + 270 + ], + "spans": [ + { + "bbox": [ + 313, + 170, + 555, + 270 + ], + "type": "text", + "content": "Figure 6. (a) Evaluating frozen backbones. Freezing the pretrained backbone, a common practice in prior works [15, 87], consistently decreases the performance. (b) Analyzing source and generated data alignment. We evaluate how the pairing of the real datasets affects performance. 'R' denotes the real dataset used in training, and 'F' indicates the source dataset used to obtain the captions for prompting the generators. The results suggest that pairing the source data (i.e., real data used to prompt the generators) with the generated images is not essential for performance." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 437, + 70, + 553, + 151 + ], + "blocks": [ + { + "bbox": [ + 437, + 70, + 553, + 151 + ], + "lines": [ + { + "bbox": [ + 437, + 70, + 553, + 151 + ], + "spans": [ + { + "bbox": [ + 437, + 70, + 553, + 151 + ], + "type": "image", + "image_path": "c408e5535218ed15163009d68d13d62ad24b504eb97ca65653f5bf9f8f90ab5b.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 443, + 153, + 547, + 163 + ], + "lines": [ + { + "bbox": [ + 443, + 153, + 547, + 163 + ], + "spans": [ + { + "bbox": [ + 443, + 153, + 547, + 163 + ], + "type": "text", + "content": "(b) Semantic alignment analysis" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 279, + 555, + 364 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 279, + 555, + 364 + ], + "spans": [ + { + "bbox": [ + 313, + 279, + 555, + 364 + ], + "type": "text", + "content": "fusion, and autoregressive models (Sec. 3.2). The classifier trained on the manual set with more diverse generator types shows similar performance compared to the one trained on the systematic set, despite containing fewer than half the number of images (774K vs. 1.9M). Additionally, we find that the two sets are complementary; the performance is further improved when we train using both sets." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 369, + 457, + 381 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 369, + 457, + 381 + ], + "spans": [ + { + "bbox": [ + 313, + 369, + 457, + 381 + ], + "type": "text", + "content": "4.4. Analysis of design choices" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 384, + 556, + 444 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 384, + 556, + 444 + ], + "spans": [ + { + "bbox": [ + 313, + 384, + 556, + 444 + ], + "type": "text", + "content": "We examine the impact of various design choices, including some suggested in earlier works. In particular, we investigate the choice of backbone models, freezing the backbone, semantic alignment between the real and generated data, and robustness to transformations." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 450, + 556, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 450, + 556, + 522 + ], + "spans": [ + { + "bbox": [ + 313, + 450, + 556, + 522 + ], + "type": "text", + "content": "Classifier backbone comparison. We compare the performance of the classifier trained using CLIP-ViT [27, 54, 96] and ConvNeXt [77] backbones following our training procedure in Figure 5b. We examine three datasets: ours, GenImage [133], and Wang et al. [125]. We observe similar performance between architectures across all datasets." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 528, + 556, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 528, + 556, + 612 + ], + "spans": [ + { + "bbox": [ + 313, + 528, + 556, + 612 + ], + "type": "text", + "content": "Frozen backbone. Prior works [15, 87] suggested using a frozen CLIP-ViT backbone for training the classifiers. We study this by training the classifiers with both frozen and unfrozen pretrained backbones, using CLIP-ViT [27, 54, 96] and ConvNeXt [77]. As shown in Fig. 6a, freezing the backbone consistently leads to worse results, indicating that end-to-end training is crucial for high performance." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 617, + 557, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 557, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 557, + 714 + ], + "type": "text", + "content": "Semantic alignment. Existing works often pair the generated images with the source dataset (i.e., the real dataset used to prompt or generate the images) arguing that misaligned data can introduce bias [7, 15, 87, 125]. We test this practice in Fig. 6b by examining the performance with both semantically aligned and misaligned real datasets. Specifically, we consider two real datasets: one comprised exclusively from LAION [107] and another combining ImageNet [23]," + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "8251" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 71, + 294, + 236 + ], + "blocks": [ + { + "bbox": [ + 56, + 71, + 294, + 236 + ], + "lines": [ + { + "bbox": [ + 56, + 71, + 294, + 236 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 294, + 236 + ], + "type": "image", + "image_path": "8a307ef1c5bdc4bcf944ceeb4dc3742fe7c1ec364fe5b71bb387b42936583893.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 241, + 295, + 275 + ], + "lines": [ + { + "bbox": [ + 55, + 241, + 295, + 275 + ], + "spans": [ + { + "bbox": [ + 55, + 241, + 295, + 275 + ], + "type": "text", + "content": "Figure 7. Robustness to various transformations. Our classifiers display robust performance across transformations. Other works generally show more sensitivity to these factors." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 292, + 296, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 292, + 296, + 376 + ], + "spans": [ + { + "bbox": [ + 55, + 292, + 296, + 376 + ], + "type": "text", + "content": "MS-COCO [74], LandscapesHQ [112], Forchheim [44], VISION [111], and IMD2020 [86]. We sample our systematically collected latent diffusion models using these two sets and categorize the generated images by their source dataset. The performance differences of all pairs were marginal, suggesting that strict alignment may not be as critical with sufficiently large data." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 402, + 296, + 618 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 402, + 296, + 618 + ], + "spans": [ + { + "bbox": [ + 55, + 402, + 296, + 618 + ], + "type": "text", + "content": "Robustness to transformations. Figure 7 illustrates the robustness of the classifiers against various transformations. Following prior works [87, 125], we test robustness to JPEG compression and Gaussian blur. Additionally, we examine robustness to rotation, resizing, padding, and shear, as they commonly occur in real-world scenarios. For padding, we randomly pad the width or height of the image with a given fraction and scale it back to the original size. Similarly, in resize, we randomly upsample or downsample the height and width of the image by a given fraction and resize it back to the original size (e.g., if fraction is 0.3, resize the height and width to " + }, + { + "bbox": [ + 55, + 402, + 296, + 618 + ], + "type": "inline_equation", + "content": "0.7 \\times" + }, + { + "bbox": [ + 55, + 402, + 296, + 618 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 55, + 402, + 296, + 618 + ], + "type": "inline_equation", + "content": "1.3 \\times" + }, + { + "bbox": [ + 55, + 402, + 296, + 618 + ], + "type": "text", + "content": " and then scale it back to the original size). The results demonstrate that our models are more robust to transformations than existing models. Specifically, GenImage [133] is notably more sensitive to Gaussian blur, JPEG compression, and resizing artifacts; classifier by Ojha et al. [87] displays sensitivity to shear transforms, and the one by Wang et al. [125] performs poorly overall." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 634, + 209, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 634, + 209, + 647 + ], + "spans": [ + { + "bbox": [ + 55, + 634, + 209, + 647 + ], + "type": "text", + "content": "4.5. Other forensics applications" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "type": "text", + "content": "Our dataset may enable further forensics studies that can take advantage of our diverse array of generators. To illustrate this, in the supplement, we identify the type of generator used to synthesize a given image by using " + }, + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "type": "text", + "content": "-nearest neighbor in the feature space of our classifier." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 71, + 388, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 71, + 388, + 83 + ], + "spans": [ + { + "bbox": [ + 314, + 71, + 388, + 83 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 89, + 555, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 89, + 555, + 209 + ], + "spans": [ + { + "bbox": [ + 313, + 89, + 555, + 209 + ], + "type": "text", + "content": "In this paper, we studied the problem of generalizing to unseen generative models in synthetic image detection. We proposed a new dataset, Community Forensics, which contains 4803 models and 2.7M images collected from various public sources. We studied the impact of model diversity and demonstrated that it plays a crucial role in enhancing data diversity and generalization performance. We trained classifiers on our dataset, studied their ability to generalize in various settings, and evaluated previously proposed models and training practices." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 210, + 556, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 210, + 556, + 319 + ], + "spans": [ + { + "bbox": [ + 313, + 210, + 556, + 319 + ], + "type": "text", + "content": "We do not intend for our dataset to be used to train classifiers that are directly used in the wild. Detecting in-the-wild synthetic images remains a challenging open problem, and detection errors can have severe consequences (e.g., falsely accusing an author of creating fake images or allowing misinformation to be certified as real). We hope that our work will serve as a stepping stone for future research in this area by providing tools and insights for studying generalization and data collection strategies." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 331, + 556, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 331, + 556, + 582 + ], + "spans": [ + { + "bbox": [ + 313, + 331, + 556, + 582 + ], + "type": "text", + "content": "Limitations. While our dataset is diverse, a large portion of the data is diffusion-based, especially models based on Stable Diffusion [100]. Despite this, Figure 1 demonstrates that adding more diffusion models to the training data improves the generalization capability of the classifiers across other generator types. Furthermore, Figure 5a illustrates improved classifier performance when combining the diffusion-only systematic set with a more diverse manual set. These results suggest that despite their similarities, each diffusion model may uniquely contribute through its semantic content, image processing pipelines, or architectural variations (Sec. 3.1). Future work may explore collecting more diverse models, including GANs, VQ-VAEs [122], and autoregressive models. We also note that the generative models sourced from the community may contain inappropriate content. While in many contexts it is important to detect such images, these generated images may require further scrutiny before being used in other downstream applications. Finally, although our experiments suggest that our forensics classifiers generalize to unseen models better than those of previous work, their error rates are still too high for many critical applications." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 594, + 557, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 557, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 557, + 712 + ], + "type": "text", + "content": "Acknowledgements. We thank the creators of the many open source models that we used to collect the Community Forensics dataset. We thank Chenhao Zheng, Cameron Johnson, Matthias Kirchner, Daniel Geng, Ziyang Chen, Ayush Shrivastava, Yiming Dou, Chao Feng, Xuanchen Lu, Zihao Wei, Zixuan Pan, Inbum Park, Rohit Banerjee, and Ang Cao for the valuable discussions and feedback. This research was developed with funding from the Defense Advanced Research Projects Agency (DARPA) under Contract No. HR001120C0123." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8252" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 91, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 66, + 91, + 296, + 113 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 91, + 296, + 113 + ], + "spans": [ + { + "bbox": [ + 66, + 91, + 296, + 113 + ], + "type": "text", + "content": "[1] Adobe. Firefly. https://www.adobe.com/products/firefly, 2023.4, 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 114, + 296, + 125 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 114, + 296, + 125 + ], + "spans": [ + { + "bbox": [ + 66, + 114, + 296, + 125 + ], + "type": "text", + "content": "[2] Ideogram AI. Ideogram. https://ideogram.ai, 2024.4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 66, + 127, + 296, + 169 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 127, + 296, + 169 + ], + "spans": [ + { + "bbox": [ + 66, + 127, + 296, + 169 + ], + "type": "text", + "content": "[3] Rushil Anirudh and Jayaraman J Thiagarajan. Out of distribution detection via neural network anchoring. In Asian Conference on Machine Learning, pages 32-47. PMLR, 2023. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 66, + 172, + 296, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 172, + 296, + 216 + ], + "spans": [ + { + "bbox": [ + 66, + 172, + 296, + 216 + ], + "type": "text", + "content": "[4] Ivan Anokhin, Kirill Demochkin, Taras Khakhulin, Gleb Sterkin, Victor Lempitsky, and Denis Korzhenkov. Image generators with conditionally-independent pixel synthesis. arXiv preprint arXiv:2011.13775, 2020. 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 66, + 217, + 296, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 217, + 296, + 271 + ], + "spans": [ + { + "bbox": [ + 66, + 217, + 296, + 271 + ], + "type": "text", + "content": "[5] Vishal Asnani, Xi Yin, Tal Hassner, and Xiaoming Liu. Reverse engineering of generative models: Inferring model hyperparameters from generated images. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(12):15477-15493, 2023. 2, 3, 4, 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 66, + 274, + 296, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 274, + 296, + 317 + ], + "spans": [ + { + "bbox": [ + 66, + 274, + 296, + 317 + ], + "type": "text", + "content": "[6] Anton Bakhtin, Sam Gross, Myle Ott, Yuntian Deng, Marc'Aurelio Ranzato, and Arthur Szlam. Real or fake? learning to discriminate machine from human generated text. arXiv preprint arXiv:1906.03351, 2019. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 66, + 319, + 296, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 319, + 296, + 352 + ], + "spans": [ + { + "bbox": [ + 66, + 319, + 296, + 352 + ], + "type": "text", + "content": "[7] Quentin Bammey. Synthbuster: Towards detection of diffusion model generated images. IEEE Open Journal of Signal Processing, 2023. 2, 3, 4, 5, 6, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 66, + 354, + 296, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 354, + 296, + 396 + ], + "spans": [ + { + "bbox": [ + 66, + 354, + 296, + 396 + ], + "type": "text", + "content": "[8] Andrew Brock, Jeff Donahue, and Karen Simonyan. Large scale GAN training for high fidelity natural image synthesis. In International Conference on Learning Representations, 2019. 3, 4, 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 66, + 399, + 296, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 399, + 296, + 442 + ], + "spans": [ + { + "bbox": [ + 66, + 399, + 296, + 442 + ], + "type": "text", + "content": "[9] Chen Chen, Qifeng Chen, Jia Xu, and Vladlen Koltun. Learning to see in the dark. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3291-3300, 2018. 3, 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 61, + 445, + 296, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 445, + 296, + 487 + ], + "spans": [ + { + "bbox": [ + 61, + 445, + 296, + 487 + ], + "type": "text", + "content": "[10] Qifeng Chen and Vladlen Koltun. Photographic image synthesis with cascaded refinement networks. In Proceedings of the IEEE international conference on computer vision, pages 1511-1520, 2017. 3, 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 61, + 490, + 296, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 490, + 296, + 544 + ], + "spans": [ + { + "bbox": [ + 61, + 490, + 296, + 544 + ], + "type": "text", + "content": "[11] Mehdi Cherti, Romain Beaumont, Ross Wightman, Mitchell Wortsman, Gabriel Ilharco, Cade Gordon, Christoph Schuhmann, Ludwig Schmidt, and Jenia Jitsev. Reproducible scaling laws for contrastive language-image learning. arXiv preprint arXiv:2212.07143, 2022.5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 61, + 546, + 296, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 546, + 296, + 600 + ], + "spans": [ + { + "bbox": [ + 61, + 546, + 296, + 600 + ], + "type": "text", + "content": "[12] Yunjey Choi, Minje Choi, Munyoung Kim, Jung-Woo Ha, Sunghun Kim, and Jaegul Choo. Stargan: Unified generative adversarial networks for multi-domain image-to-image translation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2018. 3, 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 61, + 602, + 296, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 602, + 296, + 645 + ], + "spans": [ + { + "bbox": [ + 61, + 602, + 296, + 645 + ], + "type": "text", + "content": "[13] Yunjey Choi, Youngjung Uh, Jaejun Yoo, and Jung-Woo Ha. Stargan v2: Diverse image synthesis for multiple domains. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2020. 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 61, + 647, + 296, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 647, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 61, + 647, + 296, + 712 + ], + "type": "text", + "content": "[14] Riccardo Corvi, Davide Cozzolino, Giada Zingarini, Giovanni Poggi, Koki Nagano, and Luisa Verdoliva. On the detection of synthetic images generated by diffusion models. In ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1-5. IEEE, 2023. 3" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 321, + 73, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 321, + 73, + 555, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 73, + 555, + 128 + ], + "spans": [ + { + "bbox": [ + 321, + 73, + 555, + 128 + ], + "type": "text", + "content": "[15] Davide Cozzolino, Giovanni Poggi, Riccardo Corvi, Matthias Nießner, and Luisa Verdoliva. Raising the bar of ai-generated image detection with clip. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4356-4366, 2024. 2, 3, 4, 5, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 321, + 130, + 555, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 130, + 555, + 195 + ], + "spans": [ + { + "bbox": [ + 321, + 130, + 555, + 195 + ], + "type": "text", + "content": "[16] Katherine Crowson, Stefan Andreas Baumann, Alex Birch, Tanishq Mathew Abraham, Daniel Z Kaplan, and Enrico Shippole. Scalable high-resolution pixel-space image synthesis with hourglass diffusion transformers. In Proceedings of the 41st International Conference on Machine Learning, pages 9550-9575. PMLR, 2024. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 321, + 198, + 555, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 198, + 555, + 251 + ], + "spans": [ + { + "bbox": [ + 321, + 198, + 555, + 251 + ], + "type": "text", + "content": "[17] Ekin Dogus Cubuk, Barret Zoph, Jon Shlens, and Quoc Le. Randaugment: Practical automated data augmentation with a reduced search space. In Advances in Neural Information Processing Systems, pages 18613-18624. Curran Associates, Inc., 2020. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 321, + 255, + 555, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 255, + 555, + 308 + ], + "spans": [ + { + "bbox": [ + 321, + 255, + 555, + 308 + ], + "type": "text", + "content": "[18] Tao Dai, Jianrui Cai, Yongbing Zhang, Shu-Tao Xia, and Lei Zhang. Second-order attention network for single image super-resolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11065-11074, 2019. 3, 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 321, + 312, + 555, + 356 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 312, + 555, + 356 + ], + "spans": [ + { + "bbox": [ + 321, + 312, + 555, + 356 + ], + "type": "text", + "content": "[19] Duc-Tien Dang-Nguyen, Cecilia Pasquini, Valentina Conotter, and Giulia Boato. Raise: A raw images dataset for digital image forensics. In Proceedings of the 6th ACM multimedia systems conference, pages 219-224, 2015. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 321, + 358, + 555, + 389 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 358, + 555, + 389 + ], + "spans": [ + { + "bbox": [ + 321, + 358, + 555, + 389 + ], + "type": "text", + "content": "[20] Quan Dao, Hao Phung, Binh Nguyen, and Anh Tran. Flow matching in latent space. arXiv preprint arXiv:2307.08698, 2023.4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 321, + 393, + 555, + 447 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 393, + 555, + 447 + ], + "spans": [ + { + "bbox": [ + 321, + 393, + 555, + 447 + ], + "type": "text", + "content": "[21] Tiago José De Carvalho, Christian Riess, Eli Angelopoulou, Helio Pedrini, and Anderson de Rezende Rocha. Exposing digital image forgeries by illumination color classification. IEEE Transactions on Information Forensics and Security, 8(7):1182-1194, 2013. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 321, + 450, + 555, + 471 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 450, + 555, + 471 + ], + "spans": [ + { + "bbox": [ + 321, + 450, + 555, + 471 + ], + "type": "text", + "content": "[22] DeepFloyd. Deepfloyd. https://huggingface.co/DeepFloyd/IF-I-L-v1.0, 2024.4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 321, + 475, + 555, + 517 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 475, + 555, + 517 + ], + "spans": [ + { + "bbox": [ + 321, + 475, + 555, + 517 + ], + "type": "text", + "content": "[23] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 4, 5, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 321, + 520, + 555, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 520, + 555, + 553 + ], + "spans": [ + { + "bbox": [ + 321, + 520, + 555, + 553 + ], + "type": "text", + "content": "[24] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 4, 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 321, + 555, + 555, + 598 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 555, + 555, + 598 + ], + "spans": [ + { + "bbox": [ + 321, + 555, + 555, + 598 + ], + "type": "text", + "content": "[25] Brian Dolhansky, Joanna Bitton, Ben Pflaum, Jikuo Lu, Russ Howes, Menglin Wang, and Cristian Canton Ferrer. The deepfake detection challenge (dfdc) dataset. arXiv preprint arXiv:2006.07397, 2020. 2, 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 321, + 601, + 555, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 601, + 555, + 645 + ], + "spans": [ + { + "bbox": [ + 321, + 601, + 555, + 645 + ], + "type": "text", + "content": "[26] Jing Dong, Wei Wang, and Tieniu Tan. Casia image tampering detection evaluation database. In 2013 IEEE China summit and international conference on signal and information processing, pages 422-426. IEEE, 2013. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 321, + 647, + 555, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 647, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 321, + 647, + 555, + 712 + ], + "type": "text", + "content": "[27] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. ICLR, 2021. 5, 6, 7" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8253" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 61, + 73, + 297, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 73, + 297, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 73, + 297, + 116 + ], + "spans": [ + { + "bbox": [ + 61, + 73, + 297, + 116 + ], + "type": "text", + "content": "[28] Yilun Du, Shuang Li, Joshua Tenenbaum, and Igor Mordatch. Improved contrastive divergence training of energy-based models. In International Conference on Machine Learning, pages 2837-2848. PMLR, 2021. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 118, + 296, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 118, + 296, + 172 + ], + "spans": [ + { + "bbox": [ + 61, + 118, + 296, + 172 + ], + "type": "text", + "content": "[29] Ricard Durall, Margret Keuper, and Janis Keuper. Watch your up-convolution: Cnn based generative deep neural networks are failing to reproduce spectral distributions. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7890–7899, 2020. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 173, + 297, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 173, + 297, + 215 + ], + "spans": [ + { + "bbox": [ + 62, + 173, + 297, + 215 + ], + "type": "text", + "content": "[30] Tarik Dzanic, Karan Shah, and Freddie Witherden. Fourier spectrum discrepancies in deep network generated images. Advances in neural information processing systems, 33:3022-3032, 2020. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 217, + 295, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 217, + 295, + 249 + ], + "spans": [ + { + "bbox": [ + 62, + 217, + 295, + 249 + ], + "type": "text", + "content": "[31] Sven Elflein, Bertrand Charpentier, Daniel Zügner, and Stephan Gunnemann. On out-of-distribution detection with energy-based models, 2021. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 250, + 296, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 250, + 296, + 294 + ], + "spans": [ + { + "bbox": [ + 62, + 250, + 296, + 294 + ], + "type": "text", + "content": "[32] David C Epstein, Ishan Jain, Oliver Wang, and Richard Zhang. Online detection of ai-generated images. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 382-392, 2023. 2, 3, 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 294, + 296, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 294, + 296, + 338 + ], + "spans": [ + { + "bbox": [ + 62, + 294, + 296, + 338 + ], + "type": "text", + "content": "[33] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12873-12883, 2021. 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 338, + 296, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 338, + 296, + 370 + ], + "spans": [ + { + "bbox": [ + 62, + 338, + 296, + 370 + ], + "type": "text", + "content": "[34] Hugging Face. Hugging face diffusers library. https://huggingface.co/models?library=diffusers, accessed on June 05, 2022, 2022. 2, 3, 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 372, + 296, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 372, + 296, + 404 + ], + "spans": [ + { + "bbox": [ + 62, + 372, + 296, + 404 + ], + "type": "text", + "content": "[35] Sebastian Gehrmann, Hendrik Strobelt, and Alexander M Rush. Gltr: Statistical detection and visualization of generated text. arXiv preprint arXiv:1906.04043, 2019. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 405, + 296, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 405, + 296, + 448 + ], + "spans": [ + { + "bbox": [ + 62, + 405, + 296, + 448 + ], + "type": "text", + "content": "[36] Thomas Gloe and Rainer Böhme. The'dresden image database' for benchmarking digital image forensics. In Proceedings of the 2010 ACM symposium on applied computing, pages 1584-1590, 2010. 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 449, + 294, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 449, + 294, + 492 + ], + "spans": [ + { + "bbox": [ + 62, + 449, + 294, + 492 + ], + "type": "text", + "content": "[37] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in Neural Information Processing Systems, 27, 2014. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 493, + 295, + 515 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 493, + 295, + 515 + ], + "spans": [ + { + "bbox": [ + 62, + 493, + 295, + 515 + ], + "type": "text", + "content": "[38] Google. Imagen 3. https://deepmind.google/technologies/imagen-3, 2024.4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 516, + 296, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 516, + 296, + 558 + ], + "spans": [ + { + "bbox": [ + 62, + 516, + 296, + 558 + ], + "type": "text", + "content": "[39] Karol Gregor, Ivo Danihelka, Andriy Mnih, Charles Blundell, and Daan Wierstra. Deep autoregressive networks. In International Conference on Machine Learning, pages 1242-1250. PMLR, 2014. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 62, + 559, + 296, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 559, + 296, + 602 + ], + "spans": [ + { + "bbox": [ + 62, + 559, + 296, + 602 + ], + "type": "text", + "content": "[40] Patrick Grommelt, Louis Weiss, Franz-Josef Pfreundt, and Janis Keuper. Fake or.jpeg? revealing common biases in generated image detection datasets. arXiv preprint arXiv:2403.17608, 2024. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 62, + 604, + 296, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 604, + 296, + 667 + ], + "spans": [ + { + "bbox": [ + 62, + 604, + 296, + 667 + ], + "type": "text", + "content": "[41] Shuyang Gu, Dong Chen, Jianmin Bao, Fang Wen, Bo Zhang, Dongdong Chen, Lu Yuan, and Baining Guo. Vector quantized diffusion model for text-to-image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10696-10706, 2022. 4, 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 62, + 670, + 296, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 670, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 62, + 670, + 296, + 712 + ], + "type": "text", + "content": "[42] Fabrizio Guillaro, Giada Zingarini, Ben Usman, Avneesh Sud, Davide Cozzolino, and Luisa Verdoliva. A bias-free training paradigm for more general ai-generated image detection, 2024. 3" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 320, + 73, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 320, + 73, + 555, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 73, + 555, + 127 + ], + "spans": [ + { + "bbox": [ + 320, + 73, + 555, + 127 + ], + "type": "text", + "content": "[43] Xiao Guo, Vishal Asnani, Sijja Liu, and Xiaoming Liu. Tracing hyperparameter dependencies for model parsing via learnable graph pooling network. Advances in Neural Information Processing Systems, 37:116899-116932, 2024. 2, 3, 4, 5, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 320, + 129, + 555, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 129, + 555, + 183 + ], + "spans": [ + { + "bbox": [ + 320, + 129, + 555, + 183 + ], + "type": "text", + "content": "[44] Benjamin Hadwiger and Christian Riess. The forchheim image database for camera identification in the wild. In Pattern Recognition. ICPR International Workshops and Challenges: Virtual Event, January 10–15, 2021, Proceedings, Part VI, pages 500–515. Springer, 2021. 4, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 320, + 184, + 555, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 184, + 555, + 228 + ], + "spans": [ + { + "bbox": [ + 320, + 184, + 555, + 228 + ], + "type": "text", + "content": "[45] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 5, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 320, + 228, + 555, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 228, + 555, + 271 + ], + "spans": [ + { + "bbox": [ + 320, + 228, + 555, + 271 + ], + "type": "text", + "content": "[46] Dan Hendrycks and Kevin Gimpel. A baseline for detecting misclassified and out-of-distribution examples in neural networks. In International Conference on Learning Representations, 2016. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 320, + 272, + 555, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 272, + 555, + 326 + ], + "spans": [ + { + "bbox": [ + 320, + 272, + 555, + 326 + ], + "type": "text", + "content": "[47] Dan Hendrycks, Mantas Mazeika, Saurav Kadavath, and Dawn Song. Using self-supervised learning can improve model robustness and uncertainty. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2019. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 320, + 327, + 555, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 327, + 555, + 404 + ], + "spans": [ + { + "bbox": [ + 320, + 327, + 555, + 404 + ], + "type": "text", + "content": "[48] Dan Hendrycks, Steven Basart, Norman Mu, Saurav Kadavath, Frank Wang, Evan Dorundo, Rahul Desai, Tyler Zhu, Samyak Parajuli, Mike Guo, Dawn Song, Jacob Steinhardt, and Justin Gilmer. The many faces of robustness: A critical analysis of out-of-distribution generalization. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 8340-8349, 2021. 2, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 320, + 404, + 555, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 404, + 555, + 437 + ], + "spans": [ + { + "bbox": [ + 320, + 404, + 555, + 437 + ], + "type": "text", + "content": "[49] Yan Hong and Jianfu Zhang. Wildfake: A large-scale challenging dataset for ai-generated images detection. arXiv preprint arXiv:2402.11843, 2024. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 320, + 437, + 555, + 481 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 437, + 555, + 481 + ], + "spans": [ + { + "bbox": [ + 320, + 437, + 555, + 481 + ], + "type": "text", + "content": "[50] Shuwei Hou, Yan Ju, Chengzhe Sun, Shan Jia, Lipeng Ke, Riky Zhou, Anita Nikolich, and Siwei Lyu. Deepfake-meter v2.0: An open platform for deepfake detection. arXiv preprint arXiv:2404.13146, 2024. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 320, + 482, + 555, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 482, + 555, + 525 + ], + "spans": [ + { + "bbox": [ + 320, + 482, + 555, + 525 + ], + "type": "text", + "content": "[51] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 320, + 526, + 555, + 559 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 526, + 555, + 559 + ], + "spans": [ + { + "bbox": [ + 320, + 526, + 555, + 559 + ], + "type": "text", + "content": "[52] Drew A Hudson and Larry Zitnick. Generative adversarial transformers. In International conference on machine learning, pages 4487-4499. PMLR, 2021. 4" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 320, + 559, + 555, + 603 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 559, + 555, + 603 + ], + "spans": [ + { + "bbox": [ + 320, + 559, + 555, + 603 + ], + "type": "text", + "content": "[53] Minyoung Huh, Andrew Liu, Andrew Owens, and Alexei A Efros. Fighting fake news: Image splice detection via learned self-consistency. In Proceedings of the European conference on computer vision (ECCV), 2018. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 320, + 604, + 555, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 604, + 555, + 669 + ], + "spans": [ + { + "bbox": [ + 320, + 604, + 555, + 669 + ], + "type": "text", + "content": "[54] Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. Openclip, 2021. If you use this software, please cite it as below. 5, 6, 7" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 320, + 670, + 555, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 670, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 320, + 670, + 555, + 712 + ], + "type": "text", + "content": "[55] Ganesh Jawahar, Muhammad Abdul-Mageed, and Laks VS Lakshmanan. Automatic detection of machine generated text: A critical survey. arXiv preprint arXiv:2011.01314, 2020. 3" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8254" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 61, + 72, + 297, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 72, + 297, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 72, + 297, + 106 + ], + "spans": [ + { + "bbox": [ + 61, + 72, + 297, + 106 + ], + "type": "text", + "content": "[56] Micah K Johnson and Hany Farid. Exposing digital forgeries in complex lighting environments. IEEE Transactions on Information Forensics and Security, 2(3):450-461, 2007. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 106, + 296, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 106, + 296, + 161 + ], + "spans": [ + { + "bbox": [ + 61, + 106, + 296, + 161 + ], + "type": "text", + "content": "[57] Minguk Kang, Jun-Yan Zhu, Richard Zhang, Jaesik Park, Eli Shechtman, Sylvain Paris, and Taesung Park. Scaling up gans for text-to-image synthesis. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 4, 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 162, + 296, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 162, + 296, + 205 + ], + "spans": [ + { + "bbox": [ + 62, + 162, + 296, + 205 + ], + "type": "text", + "content": "[58] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of gans for improved quality, stability, and variation. In International Conference on Learning Representations, 2018. 3, 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 205, + 296, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 205, + 296, + 258 + ], + "spans": [ + { + "bbox": [ + 62, + 205, + 296, + 258 + ], + "type": "text", + "content": "[59] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 3, 4, 5, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 261, + 296, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 261, + 296, + 304 + ], + "spans": [ + { + "bbox": [ + 62, + 261, + 296, + 304 + ], + "type": "text", + "content": "[60] Tero Karras, Miika Aittala, Janne Hellsten, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Training generative adversarial networks with limited data. Advances in neural information processing systems, 33:12104-12114, 2020. 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 305, + 296, + 359 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 305, + 296, + 359 + ], + "spans": [ + { + "bbox": [ + 62, + 305, + 296, + 359 + ], + "type": "text", + "content": "[61] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8110-8119, 2020." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 360, + 296, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 360, + 296, + 403 + ], + "spans": [ + { + "bbox": [ + 62, + 360, + 296, + 403 + ], + "type": "text", + "content": "[62] Tero Karras, Miika Aittala, Samuli Laine, Erik Härkonen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Alias-free generative adversarial networks. Advances in neural information processing systems, 34:852-863, 2021. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 404, + 296, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 404, + 296, + 448 + ], + "spans": [ + { + "bbox": [ + 62, + 404, + 296, + 448 + ], + "type": "text", + "content": "[63] Hasam Khalid, Shahroz Tariq, Minha Kim, and Simon Woo. Fakeavceleb: A novel audio-video multimodal deepfake dataset. In Proceedings of the Neural Information Processing Systems Track on Datasets and Benchmarks, 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 449, + 296, + 481 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 449, + 296, + 481 + ], + "spans": [ + { + "bbox": [ + 62, + 449, + 296, + 481 + ], + "type": "text", + "content": "[64] P. Korus and J. Huang. Multi-scale analysis strategies in prnu-based tampering localization. IEEE Trans. on Information Forensics & Security, 2017. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 482, + 296, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 482, + 296, + 535 + ], + "spans": [ + { + "bbox": [ + 62, + 482, + 296, + 535 + ], + "type": "text", + "content": "[65] Kalpesh Krishna, Yixiao Song, Marzena Karpinska, John Wieting, and Mohit Iyyer. Paraphrasing evades detectors of ai-generated text, but retrieval is an effective defense. Advances in Neural Information Processing Systems, 36, 2024. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 536, + 296, + 559 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 536, + 296, + 559 + ], + "spans": [ + { + "bbox": [ + 62, + 536, + 296, + 559 + ], + "type": "text", + "content": "[66] Kvikontent. Kvikontent-midjourney v6. https://huggingface.co/Kvikontent/midjourney-v6, 2023.5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 559, + 296, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 559, + 296, + 602 + ], + "spans": [ + { + "bbox": [ + 62, + 559, + 296, + 602 + ], + "type": "text", + "content": "[67] Patrick Kwon, Jaeseong You, Gyuhyeon Nam, Sungwoo Park, and Gyeongsu Chae. Kodf: A large-scale korean deepfake detection dataset. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 62, + 603, + 296, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 603, + 296, + 624 + ], + "spans": [ + { + "bbox": [ + 62, + 603, + 296, + 624 + ], + "type": "text", + "content": "[68] Black Forst Labs. Flux. https://blackforestlabs.ai, 2024.4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 62, + 625, + 296, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 625, + 296, + 668 + ], + "spans": [ + { + "bbox": [ + 62, + 625, + 296, + 668 + ], + "type": "text", + "content": "[69] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, 2022. 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 62, + 670, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 670, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 62, + 670, + 296, + 713 + ], + "type": "text", + "content": "[70] Ke Li, Tianhao Zhang, and Jitendra Malik. Diverse image synthesis from semantic layouts via conditional imle. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4220-4229, 2019. 3, 5" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 320, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 320, + 73, + 555, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 73, + 555, + 116 + ], + "spans": [ + { + "bbox": [ + 320, + 73, + 555, + 116 + ], + "type": "text", + "content": "[71] Yuezun Li, Xin Yang, Pu Sun, Honggang Qi, and Siwei Lyu. Celeb-df: A large-scale challenging dataset for deepfake forensics. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2020. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 320, + 118, + 555, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 118, + 555, + 171 + ], + "spans": [ + { + "bbox": [ + 320, + 118, + 555, + 171 + ], + "type": "text", + "content": "[72] Yewen Li, Chaojie Wang, Xiaobo Xia, Tongliang Liu, Bo An, et al. Out-of-distribution detection with an adaptive likelihood ratio on informative hierarchical vae. Advances in Neural Information Processing Systems, 35:7383-7396, 2022. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 320, + 172, + 555, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 172, + 555, + 215 + ], + "spans": [ + { + "bbox": [ + 320, + 172, + 555, + 215 + ], + "type": "text", + "content": "[73] Shiyu Liang, Yixuan Li, and R Srikant. Enhancing the reliability of out-of-distribution image detection in neural networks. In International Conference on Learning Representations, 2018. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 320, + 217, + 555, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 217, + 555, + 282 + ], + "spans": [ + { + "bbox": [ + 320, + 217, + 555, + 282 + ], + "type": "text", + "content": "[74] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 4, 5, 6, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 320, + 283, + 555, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 283, + 555, + 326 + ], + "spans": [ + { + "bbox": [ + 320, + 283, + 555, + 326 + ], + "type": "text", + "content": "[75] Weitang Liu, Xiaoyun Wang, John Owens, and Yixuan Li. Energy-based out-of-distribution detection. In Advances in Neural Information Processing Systems, pages 21464-21475. Curran Associates, Inc., 2020. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 320, + 327, + 555, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 327, + 555, + 369 + ], + "spans": [ + { + "bbox": [ + 320, + 327, + 555, + 369 + ], + "type": "text", + "content": "[76] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaou Tang. Deep learning face attributes in the wild. In Proceedings of International Conference on Computer Vision (ICCV), 2015. 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 320, + 371, + 555, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 371, + 555, + 425 + ], + "spans": [ + { + "bbox": [ + 320, + 371, + 555, + 425 + ], + "type": "text", + "content": "[77] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11976-11986, 2022. 5, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 320, + 426, + 555, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 426, + 555, + 469 + ], + "spans": [ + { + "bbox": [ + 320, + 426, + 555, + 469 + ], + "type": "text", + "content": "[78] Simian Luo, Yiqin Tan, Longbo Huang, Jian Li, and Hang Zhao. Latent consistency models: Synthesizing high-resolution images with few-step inference. arXiv preprint arXiv:2310.04378, 2023. 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 320, + 471, + 554, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 471, + 554, + 514 + ], + "spans": [ + { + "bbox": [ + 320, + 471, + 554, + 514 + ], + "type": "text", + "content": "[79] Simian Luo, Yiqin Tan, Suraj Patil, Daniel Gu, Patrick von Platen, Apolinário Passos, Longbo Huang, Jian Li, and Hang Zhao. Lcm-lora: A universal stable-diffusion acceleration module. arXiv preprint arXiv:2311.05556, 2023. 2, 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 320, + 515, + 555, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 515, + 555, + 558 + ], + "spans": [ + { + "bbox": [ + 320, + 515, + 555, + 558 + ], + "type": "text", + "content": "[80] Francesco Marra, Diego Gragnaniello, Luisa Verdoliva, and Giovanni Poggi. Do gans leave artificial fingerprints? In 2019 IEEE conference on multimedia information processing and retrieval (MIPR), pages 506-511. IEEE, 2019. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 320, + 559, + 555, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 559, + 555, + 579 + ], + "spans": [ + { + "bbox": [ + 320, + 559, + 555, + 579 + ], + "type": "text", + "content": "[81] Inc. Midjourney. Midjourney. https://www.midjourney.com/home, 2022. 4, 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 320, + 581, + 555, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 581, + 555, + 635 + ], + "spans": [ + { + "bbox": [ + 320, + 581, + 555, + 635 + ], + "type": "text", + "content": "[82] Eric Mitchell, Yoonho Lee, Alexander Khazatsky, Christopher D Manning, and Chelsea Finn. Detectgpt: Zero-shot machine-generated text detection using probability curvature. In International Conference on Machine Learning, pages 24950-24962. PMLR, 2023. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 320, + 636, + 555, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 636, + 555, + 679 + ], + "spans": [ + { + "bbox": [ + 320, + 636, + 555, + 679 + ], + "type": "text", + "content": "[83] Sina Mohseni, Mandar Pitale, JBS Yadawa, and Zhangyang Wang. Self-supervised learning for generalizable out-of-distribution detection. Proceedings of the AAAI Conference on Artificial Intelligence, 34(04):5216-5223, 2020. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 320, + 681, + 555, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 681, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 320, + 681, + 555, + 713 + ], + "type": "text", + "content": "[84] Tian-Tsong Ng, Shih-Fu Chang, and Q Sun. A data set of authentic and spliced image blocks. Columbia University, ADVENT Technical Report, 4, 2004. 3" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8255" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 61, + 73, + 297, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 73, + 297, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 73, + 297, + 139 + ], + "spans": [ + { + "bbox": [ + 61, + 73, + 297, + 139 + ], + "type": "text", + "content": "[85] Alexander Quinn Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. In International Conference on Machine Learning, pages 16784-16804. PMLR, 2022. 4, 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 140, + 296, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 140, + 296, + 195 + ], + "spans": [ + { + "bbox": [ + 61, + 140, + 296, + 195 + ], + "type": "text", + "content": "[86] Adam Novozamsky, Babak Mahdian, and Stanislav Saic. Imd2020: A large-scale annotated dataset tailored for detecting manipulated images. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision Workshops, pages 71-80, 2020. 4, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 197, + 296, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 197, + 296, + 251 + ], + "spans": [ + { + "bbox": [ + 62, + 197, + 296, + 251 + ], + "type": "text", + "content": "[87] Utkarsh Ojha, Yuheng Li, and Yong Jae Lee. Towards universal fake image detectors that generalize across generative models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 24480-24489, 2023. 1, 2, 3, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 252, + 296, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 252, + 296, + 274 + ], + "spans": [ + { + "bbox": [ + 62, + 252, + 296, + 274 + ], + "type": "text", + "content": "[88] OpenAI. Dall-e 2. https://openai.com/index/dall-e-2, 2022. 4, 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 276, + 296, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 276, + 296, + 297 + ], + "spans": [ + { + "bbox": [ + 62, + 276, + 296, + 297 + ], + "type": "text", + "content": "[89] OpenAI. Dall-e 3. https://openai.com/index/dall-e-3, 2023.4, 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 299, + 296, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 299, + 296, + 342 + ], + "spans": [ + { + "bbox": [ + 62, + 299, + 296, + 342 + ], + "type": "text", + "content": "[90] Taesung Park, Ming-Yu Liu, Ting-Chun Wang, and JunYan Zhu. Semantic image synthesis with spatially-adaptive normalization. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2019. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 344, + 296, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 344, + 296, + 408 + ], + "spans": [ + { + "bbox": [ + 62, + 344, + 296, + 408 + ], + "type": "text", + "content": "[91] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 411, + 296, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 411, + 296, + 453 + ], + "spans": [ + { + "bbox": [ + 62, + 411, + 296, + 453 + ], + "type": "text", + "content": "[92] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4195-4205, 2023. 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 456, + 296, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 456, + 296, + 510 + ], + "spans": [ + { + "bbox": [ + 62, + 456, + 296, + 510 + ], + "type": "text", + "content": "[93] Pablo Pernias, Dominic Rampas, Mats Leon Richter, Christopher Pal, and Marc Aubreville. Würstchen: An efficient architecture for large-scale text-to-image diffusion models. In The Twelfth International Conference on Learning Representations, 2023. 5, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 512, + 296, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 512, + 296, + 576 + ], + "spans": [ + { + "bbox": [ + 62, + 512, + 296, + 576 + ], + "type": "text", + "content": "[94] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. In The Twelfth International Conference on Learning Representations, 2023. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 579, + 294, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 579, + 294, + 612 + ], + "spans": [ + { + "bbox": [ + 62, + 579, + 294, + 612 + ], + "type": "text", + "content": "[95] Alin C Popescu and Hany Farid. Exposing digital forgeries by detecting traces of resampling. IEEE Transactions on signal processing, 2005. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 613, + 296, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 613, + 296, + 678 + ], + "spans": [ + { + "bbox": [ + 62, + 613, + 296, + 678 + ], + "type": "text", + "content": "[96] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 3, 5, 6, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 62, + 680, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 680, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 62, + 680, + 296, + 713 + ], + "type": "text", + "content": "[97] Anirudh Sundara Rajan, Utkarsh Ojha, Jedidiah Schoesser, and Yong Jae Lee. Aligned datasets improve detection of latent diffusion-generated images, 2025. 3" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 320, + 73, + 555, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 73, + 555, + 127 + ], + "spans": [ + { + "bbox": [ + 320, + 73, + 555, + 127 + ], + "type": "text", + "content": "[98] Jie Ren, Peter J Liu, Emily Fertig, Jasper Snoek, Ryan Poplin, Mark Depristo, Joshua Dillon, and Balaji Lakshminarayanan. Likelihood ratios for out-of-distribution detection. Advances in neural information processing systems, 32, 2019. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 320, + 129, + 555, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 129, + 555, + 184 + ], + "spans": [ + { + "bbox": [ + 320, + 129, + 555, + 184 + ], + "type": "text", + "content": "[99] Jonas Ricker, Denis Lukovnikov, and Asja Fischer. Aeroblade: Training-free detection of latent diffusion images using autoencoder reconstruction error. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9130-9140, 2024. 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 186, + 555, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 186, + 555, + 240 + ], + "spans": [ + { + "bbox": [ + 316, + 186, + 555, + 240 + ], + "type": "text", + "content": "[100] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, 2022. 2, 4, 5, 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 241, + 555, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 241, + 555, + 285 + ], + "spans": [ + { + "bbox": [ + 316, + 241, + 555, + 285 + ], + "type": "text", + "content": "[101] Andreas Rössler, Davide Cozzolino, Luisa Verdoliva, Christian Riess, Justus Thies, and Matthias Nießner. Faceforensics: A large-scale video dataset for forgery detection in human faces. arXiv preprint arXiv:1803.09179, 2018. 1, 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 287, + 555, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 287, + 555, + 341 + ], + "spans": [ + { + "bbox": [ + 316, + 287, + 555, + 341 + ], + "type": "text", + "content": "[102] Andreas Rossler, Davide Cozzolino, Luisa Verdoliva, Christian Riess, Justus Thies, and Matthias Nießner. Faceforensics++: Learning to detect manipulated facial images. In Proceedings of the IEEE/CVF international conference on computer vision, 2019. 1, 2, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 343, + 555, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 343, + 555, + 386 + ], + "spans": [ + { + "bbox": [ + 316, + 343, + 555, + 386 + ], + "type": "text", + "content": "[103] Vinu Sankar Sadasivan, Aounon Kumar, Sriram Balasubramanian, Wenxiao Wang, and Soheil Feizi. Can ai-generated text be reliably detected? arXiv preprint arXiv:2303.11156, 2023. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 388, + 555, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 388, + 555, + 432 + ], + "spans": [ + { + "bbox": [ + 316, + 388, + 555, + 432 + ], + "type": "text", + "content": "[104] Chandramouli Shama Sastry and Sageev Oore. Detecting out-of-distribution examples with Gram matrices. In Proceedings of the 37th International Conference on Machine Learning, pages 8491-8501. PMLR, 2020. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 434, + 555, + 467 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 434, + 555, + 467 + ], + "spans": [ + { + "bbox": [ + 316, + 434, + 555, + 467 + ], + "type": "text", + "content": "[105] Axel Sauer, Kashyap Chitta, Jens Müller, and Andreas Geiger. Projected gans converge faster. In Advances in Neural Information Processing Systems (NeurIPS), 2021. 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 468, + 555, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 468, + 555, + 510 + ], + "spans": [ + { + "bbox": [ + 316, + 468, + 555, + 510 + ], + "type": "text", + "content": "[106] Axel Sauer, Katja Schwarz, and Andreas Geiger. Styleganxl: Scaling stylegan to large diverse datasets. In ACM SIGGRAPH 2022 conference proceedings, pages 1-10, 2022. 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 513, + 555, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 513, + 555, + 567 + ], + "spans": [ + { + "bbox": [ + 316, + 513, + 555, + 567 + ], + "type": "text", + "content": "[107] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021. 2, 4, 5, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 569, + 555, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 569, + 555, + 667 + ], + "spans": [ + { + "bbox": [ + 316, + 569, + 555, + 667 + ], + "type": "text", + "content": "[108] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade W Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortman, Patrick Schramowski, Srivatsa R Kundurthy, Katherine Crowson, Ludwig Schmidt, Robert Kaczmarczyk, and Jenia Jitsev. LAION-5b: An open large-scale dataset for training next generation image-text models. In Thirty-sixth Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2022. 2, 3, 4, 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 670, + 555, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 555, + 712 + ], + "type": "text", + "content": "[109] Vikash Sehwag, Mung Chiang, and Prateek Mittal. Ssd: A unified framework for self-supervised outlier detection. In International Conference on Learning Representations, 2020. 3" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "8256" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 297, + 117 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 297, + 117 + ], + "type": "text", + "content": "[110] Arseniy Shakhmatov, Anton Razzhigaev, Aleksandr Nikolich, Vladimir Arkhipkin, Igor Pavlov, Andrey Kuznetsov, and Denis Dimitrov. Kandinsky 2.2. https://github.com/ai-forever/Kandinsky-2, 2023.5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 296, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 296, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 296, + 162 + ], + "type": "text", + "content": "[111] Dasara Shullani, Marco Fontani, Massimo Iuliani, Omar Al Shaya, and Alessandro Piva. Vision: a video and image dataset for source identification. EURASIP Journal on Information Security, 2017:1-16, 2017. 4, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 296, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 296, + 206 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 296, + 206 + ], + "type": "text", + "content": "[112] Ivan Skorokhodov, Grigorii Sotnikov, and Mohamed Elhoseiny. Aligning latent and image spaces to connect the unconnectable. arXiv preprint arXiv:2104.06954, 2021. 4, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 209, + 296, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 209, + 296, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 209, + 296, + 262 + ], + "type": "text", + "content": "[113] Irene Solaiman, Miles Brundage, Jack Clark, Amanda Askell, Ariel Herbert-Voss, Jeff Wu, Alec Radford, Gretchen Krueger, Jong Wook Kim, Sarah Kreps, et al. Release strategies and the social impacts of language models. arXiv preprint arXiv:1908.09203, 2019. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 264, + 296, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 264, + 296, + 308 + ], + "spans": [ + { + "bbox": [ + 56, + 264, + 296, + 308 + ], + "type": "text", + "content": "[114] Yang Song, Prafulla Dhariwal, Mark Chen, and Ilya Sutskever. Consistency models. In Proceedings of the 40th International Conference on Machine Learning, pages 32211-32252, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 309, + 296, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 309, + 296, + 365 + ], + "spans": [ + { + "bbox": [ + 56, + 309, + 296, + 365 + ], + "type": "text", + "content": "[115] Yuhta Takida, Masaaki Imaizumi, Takashi Shibuya, Chieh-Hsin Lai, Toshimitsu Uesaka, Naoki Murata, and Yuki Mitsufuji. SAN: Inducing metrizability of GAN with discriminative normalized linear layer. In The Twelfth International Conference on Learning Representations, 2024. 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 365, + 296, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 365, + 296, + 430 + ], + "spans": [ + { + "bbox": [ + 56, + 365, + 296, + 430 + ], + "type": "text", + "content": "[116] Chuangchuang Tan, Yao Zhao, Shikui Wei, Guanghua Gu, and Yunchao Wei. Learning on gradients: Generalized artifacts representation for gan-generated images detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12105-12114, 2023. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 433, + 296, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 433, + 296, + 498 + ], + "spans": [ + { + "bbox": [ + 56, + 433, + 296, + 498 + ], + "type": "text", + "content": "[117] Chuangchuang Tan, Yao Zhao, Shikui Wei, Guanghua Gu, Ping Liu, and Yunchao Wei. Rethinking the up-sampling operations in cnn-based generative network for generalizable deepfake detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 28130-28139, 2024. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 500, + 296, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 500, + 296, + 555 + ], + "spans": [ + { + "bbox": [ + 56, + 500, + 296, + 555 + ], + "type": "text", + "content": "[118] Ming Tao, Hao Tang, Fei Wu, Xiao-Yuan Jing, Bing-Kun Bao, and Changsheng Xu. Df-gan: A simple and effective baseline for text-to-image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16515-16525, 2022. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 555, + 296, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 555, + 296, + 609 + ], + "spans": [ + { + "bbox": [ + 56, + 555, + 296, + 609 + ], + "type": "text", + "content": "[119] Ming Tao, Bing-Kun Bao, Hao Tang, and Changsheng Xu. Galip: Generative adversarial clips for text-to-image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14214-14223, 2023. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 611, + 268, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 611, + 268, + 622 + ], + "spans": [ + { + "bbox": [ + 56, + 611, + 268, + 622 + ], + "type": "text", + "content": "[120] DeciAI Research Team. Decidiffusion 2.0, 2024. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 624, + 296, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 296, + 677 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 296, + 677 + ], + "type": "text", + "content": "[121] Adaku Uchendu, Thai Le, Kai Shu, and Dongwon Lee. Authorship attribution for neural text generation. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 8384–8395, 2020. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 680, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 680, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 680, + 296, + 713 + ], + "type": "text", + "content": "[122] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017. 8" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 555, + 654 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 316, + 72, + 555, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 555, + 139 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 555, + 139 + ], + "type": "text", + "content": "[123] Patrick von Platen, Suraj Patil, Anton Lozhkov, Pedro Cuenca, Nathan Lambert, Kashif Rasul, Mishig Davaadorj, Dhruv Nair, Sayak Paul, William Berman, Yiyi Xu, Steven Liu, and Thomas Wolf. Diffusers: State-of-the-art diffusion models. https://github.com/huggingface/ diffusers, 2022. 2, 3, 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 140, + 555, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 140, + 555, + 195 + ], + "spans": [ + { + "bbox": [ + 316, + 140, + 555, + 195 + ], + "type": "text", + "content": "[124] Apoorv Vyas, Nataraj Jammalamadaka, Xia Zhu, Dipankar Das, Bharat Kaul, and Theodore L. Willke. Out-of-distribution detection using an ensemble of self-supervised leave-out classifiers. In Proceedings of the European Conference on Computer Vision (ECCV), 2018. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 196, + 555, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 196, + 555, + 251 + ], + "spans": [ + { + "bbox": [ + 316, + 196, + 555, + 251 + ], + "type": "text", + "content": "[125] Sheng-Yu Wang, Oliver Wang, Richard Zhang, Andrew Owens, and Alexei A. Efros. Cnn-generated images are surprisingly easy to spot... for now. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 1, 2, 3, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 251, + 555, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 251, + 555, + 274 + ], + "spans": [ + { + "bbox": [ + 316, + 251, + 555, + 274 + ], + "type": "text", + "content": "[126] Ross Wightman. Pytorch image models. https://github.com/huggingface/pytorch-image-models, 2019.5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 274, + 555, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 274, + 555, + 296 + ], + "spans": [ + { + "bbox": [ + 316, + 274, + 555, + 296 + ], + "type": "text", + "content": "[127] Wukong. https://xihe.mindspore.cn/modelzoo/wukong, 2022.5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 297, + 555, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 297, + 555, + 351 + ], + "spans": [ + { + "bbox": [ + 316, + 297, + 555, + 351 + ], + "type": "text", + "content": "[128] Zhisheng Xiao, Qing Yan, and Yali Amit. Likelihood regret: An out-of-distribution detection score for variational autoencoder. In Advances in Neural Information Processing Systems, pages 20685-20696. Curran Associates, Inc., 2020. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 353, + 555, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 353, + 555, + 397 + ], + "spans": [ + { + "bbox": [ + 316, + 353, + 555, + 397 + ], + "type": "text", + "content": "[129] Fisher Yu, Yinda Zhang, Shuran Song, Ari Seff, and Jianxiong Xiao. Lsun: Construction of a large-scale image dataset using deep learning with humans in the loop. arXiv preprint arXiv:1506.03365, 2015. 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 398, + 555, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 398, + 555, + 453 + ], + "spans": [ + { + "bbox": [ + 316, + 398, + 555, + 453 + ], + "type": "text", + "content": "[130] Bowen Zhang, Shuyang Gu, Bo Zhang, Jianmin Bao, Dong Chen, Fang Wen, Yong Wang, and Baining Guo. Styleswin: Transformer-based gan for high-resolution image generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11304-11314, 2022. 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 453, + 555, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 453, + 555, + 498 + ], + "spans": [ + { + "bbox": [ + 316, + 453, + 555, + 498 + ], + "type": "text", + "content": "[131] Xu Zhang, Svebor Karaman, and Shih-Fu Chang. Detecting and simulating artifacts in gan fake images. In 2019 IEEE international workshop on information forensics and security (WIFS), pages 1-6. IEEE, 2019. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 498, + 555, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 498, + 555, + 543 + ], + "spans": [ + { + "bbox": [ + 316, + 498, + 555, + 543 + ], + "type": "text", + "content": "[132] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In Computer Vision (ICCV), 2017 IEEE International Conference on, 2017. 3, 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 544, + 555, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 544, + 555, + 609 + ], + "spans": [ + { + "bbox": [ + 316, + 544, + 555, + 609 + ], + "type": "text", + "content": "[133] Mingjian Zhu, Hanting Chen, Qiangyu YAN, Xudong Huang, Guanyu Lin, Wei Li, Zhijun Tu, Hailin Hu, Jie Hu, and Yunhe Wang. Genimage: A million-scale benchmark for detecting ai-generated image. In Advances in Neural Information Processing Systems Dataset and Benchmarks Track, pages 77771-77782, 2023. 2, 3, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 609, + 555, + 654 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 609, + 555, + 654 + ], + "spans": [ + { + "bbox": [ + 316, + 609, + 555, + 654 + ], + "type": "text", + "content": "[134] Bojia Zi, Minghao Chang, Jingjing Chen, Xingjun Ma, and Yu-Gang Jiang. Wilddeepfake: A challenging real-world dataset for deepfake detection. In Proceedings of the 28th ACM international conference on multimedia, 2020. 2" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8257" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/CompGS_ Unleashing 2D Compositionality for Compositional Text-to-3D via Dynamically Optimizing 3D Gaussians/d9350216-461c-42c5-bc6e-3da77f71b98b_content_list.json b/2025/CompGS_ Unleashing 2D Compositionality for Compositional Text-to-3D via Dynamically Optimizing 3D Gaussians/d9350216-461c-42c5-bc6e-3da77f71b98b_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..a85f280942219c63cd00fa8f604a974f96d769b7 --- /dev/null +++ b/2025/CompGS_ Unleashing 2D Compositionality for Compositional Text-to-3D via Dynamically Optimizing 3D Gaussians/d9350216-461c-42c5-bc6e-3da77f71b98b_content_list.json @@ -0,0 +1,1636 @@ +[ + { + "type": "text", + "text": "COMPGS: Unleashing 2D Compositionality for Compositional Text-to-3D via Dynamically Optimizing 3D Gaussians", + "text_level": 1, + "bbox": [ + 109, + 128, + 888, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Chongjian Ge $^{1,2,3}$ Chenfeng Xu $^{1}$ Yuanfeng Ji $^{2}$ Chensheng Peng $^{1}$", + "bbox": [ + 236, + 202, + 756, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Masayoshi Tomizuka1 Ping Luo2 Mingyu Ding1,4* Varun Jampani5* Wei Zhan1", + "bbox": [ + 181, + 220, + 818, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 University of California, Berkeley 2 The University of Hong Kong", + "bbox": [ + 222, + 239, + 776, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3 Adobe Research 4 UNC-Chapel Hill 5 Stability AI", + "bbox": [ + 272, + 256, + 723, + 273 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/711f275c50c8f04f3791954f81ce0e166e265f31101abe14eb16497beee1ebcb.jpg", + "image_caption": [ + "Single 3D Generation", + "a beautiful butterfly" + ], + "image_footnote": [], + "bbox": [ + 130, + 316, + 284, + 388 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e314c9b51b5258fc6be1831c3b5f283d3208699c1cee517117fa3e913c794d06.jpg", + "image_caption": [ + "an owl" + ], + "image_footnote": [], + "bbox": [ + 142, + 412, + 287, + 481 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2245fd20debe10497379649f31b346c997201c81f643591b5d83994ccfecc0ef.jpg", + "image_caption": [ + "a bird is drinking water from a cup" + ], + "image_footnote": [], + "bbox": [ + 313, + 315, + 491, + 383 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7c26be72f1f1f944b7c0183724152f786963d9d3f432c274900ae5bf9953cfef.jpg", + "image_caption": [ + "a footballer is kicking a soccer ball", + "an owl perches on a branch" + ], + "image_footnote": [], + "bbox": [ + 315, + 414, + 483, + 484 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6c558797819cb60bc2f573e896a806940ddcd16c8f3e32decd7185f64458568e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 514, + 323, + 558, + 381 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2fcc4315a30b8be42f6dade02e4b9ec2909888df3c2ea19e4b5c0c5a2ac51347.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 573, + 332, + 661, + 378 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/483d3b1f11e17f8e19223f57a2277944ccab76b0e53319b149b3241003904ebe.jpg", + "image_caption": [ + "an owl perches on a branch near a pinecone" + ], + "image_footnote": [], + "bbox": [ + 504, + 417, + 653, + 482 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e521b80e6d8e6bf74a34430ac8f46111c0932774919ce2a117c687c84c446bd0.jpg", + "image_caption": [ + "Compositional 3D Generation" + ], + "image_footnote": [], + "bbox": [ + 676, + 315, + 771, + 386 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4e1d03f14646c0d0fdd9ffb11b46b21e9e05f812111ab677c229aa5262ef53dc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 727, + 316, + 807, + 367 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a96d1900beff7b3043bbd55d8c13dc57870d25a7e40c2b74ea64a5c06812c4b7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 818, + 315, + 861, + 386 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/fb91e0ee73ba362d49758d690786580adec6a94fcb0d76d47c313c895167fa11.jpg", + "image_caption": [ + "a photographer is capturing abeautiful butterfly with camera" + ], + "image_footnote": [], + "bbox": [ + 679, + 412, + 761, + 482 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3c5d98a57054db4df6a264de847be3dd3c554ed184f362b9a72dc2cacb8ec524.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 779, + 415, + 849, + 483 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a2be1b7abbcb20740667e2dc4186db6b131370839f3f176e4fb601e1d8316c0c.jpg", + "image_caption": [ + "an owl perches on a branch near a pinecone, with a rat below the branch", + "Figure 1. Illustration of compositional 3D Generation and COMPGS. All the contents are generated by COMPGS. Top row: COMPGS is capable of generating either a single object (e.g., a butterfly) or generating compositional objects with reasonable interactions (e.g., the rightmost figure in the top row). Middle row: Beyond text-to-3D generation, COMPGS can be easily extend to 3D editing by progressively adding objects. The colored texts (e.g., 'a branch', 'a pinecone', 'a rat' in the rightmost figure) denote the added part compared to its previous asset. Bottom row: COMPGS achieves compositional text-to-3D by transferring 2D compositionality to initialize 3D Gaussians. COMPGS is further trained with dynamic SDS optimization to produce plausible results." + ], + "image_footnote": [], + "bbox": [ + 127, + 523, + 869, + 635 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 724, + 326, + 739 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent breakthroughs in text-guided image generation have significantly advanced the field of 3D generation. While generating a single high-quality 3D object is now feasible, generating multiple objects with reasonable interactions within a 3D space, a.k.a. compositional 3D generation, presents substantial challenges. This paper introduces COMPGS, a novel generative framework that employs 3D Gaussian Splatting (GS) for efficient, compositional text-to-3D content generation. To achieve this goal, two core designs are proposed:", + "bbox": [ + 89, + 763, + 485, + 901 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "(1) 3D Gaussians Initialization with 2D compositionality: We transfer the well-established 2D compositionality to initialize the Gaussian parameters on an entity-by-entity basis, ensuring both consistent 3D priors for each entity and reasonable interactions among multiple entities; (2) Dynamic Optimization: We propose a dynamic strategy to optimize 3D Gaussians using Score Distillation Sampling (SDS) loss. COMPGS first automatically decomposes 3D Gaussians into distinct entity parts, enabling optimization at both the entity and composition levels. Additionally, COMPGS optimizes across objects of varying scales by dynamically adjusting the", + "bbox": [ + 511, + 724, + 908, + 893 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation.", + "bbox": [ + 236, + 0, + 810, + 18 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "Except for this watermark, it is identical to the accepted version;", + "bbox": [ + 323, + 16, + 723, + 31 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 292, + 31, + 754, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "18509", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "spatial parameters of each entity, enhancing the generation of fine-grained details, particularly in smaller entities. Qualitative comparisons and quantitative evaluations on $\\mathrm{T}^3$ Bench demonstrate the effectiveness of COMPGS in generating compositional 3D objects with superior image quality and semantic alignment over existing methods. COMPGS can also be easily extended to progressively adding 3D objects, facilitating complex scene generation. We hope COMPGS will provide new insights to the compositional 3D generation.", + "bbox": [ + 88, + 90, + 486, + 229 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 253, + 225, + 271 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3D content creation is essential to the modern media industry, yet it has traditionally been labor-intensive and necessitated professional expertise. Typically, designing a single 3D object takes several hours for an experienced designer, and creating complex scenes with multiple 3D objects (as shown in Fig. 1) requires even more effort. Inspired by the recent success of diffusion models in the text-to-image generation [28, 51, 56, 61], much research has focused on exploring 2D diffusion models for text-to-3D generation. Previous work can be divided into two main methodologies: (1) Feed-forward generation [29, 39, 73], which entails training generalizable diffusion models on 3D data; and (2) Optimization-based generation [10, 41, 47, 52, 68, 71], which utilizes the pretrained 2D diffusion guidance to optimize 3D representations via Score Distillation Sampling (SDS) [52].", + "bbox": [ + 91, + 280, + 485, + 521 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "While existing work has demonstrated the feasibility of generating single 3D objects, they often struggle to produce compositional 3D content with multiple objects and complex interactions. For example, (1) Feed-forward generation methods struggle in generalizing to complex textual descriptions since the amount of 3D training data is extremely limited [17, 18], and most of the data contains only one object; (2) Optimization-based generation methods face significant challenges with current 2D diffusion guidance when optimizing compositional 3D objects. Typically, using 2D diffusion guidance to optimize a single object is feasible, as it's easy to incorporate various attributes into a single object. However, in compositional 3D generation, 2D guidance struggles to accurately compose multiple objects with different attributes and relationships into a coherent scene [30]. For example, given the prompt 'a blue bench on the left of a green car,' distinguishing different attributes within the implicit 2D diffusion priors is challenging, resulting in misaligning attributes to different objects or generating incorrect spatial relationships. Thus, optimization-based compositional generation with standard 2D guidance often causes problems like 3D inconsistencies, multi-faced objects, semantic drift, etc [26, 58].", + "bbox": [ + 91, + 522, + 483, + 868 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we propose COMPGS, a generative system based on 3D Gaussian Splatting (GS) [33] for compositional", + "bbox": [ + 89, + 869, + 483, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "text-to-3D generation. To achieve this, we introduce two core designs:", + "bbox": [ + 511, + 90, + 906, + 121 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3D Gaussians Initialization with 2D Compositionality. Unlike the implicit representation, i.e., NeRF [48], COMPGS uses 3D Gaussians as the representation, which helps to achieve feasible parameter initialization with a coarse 3D shape [31, 77]. As shown in the bottom row of Fig. 1, we first apply a text-to-image model [3, 7, 55] to generate an image that accurately captures the compositionality of multiple objects. Then the image is segmented into different sub-objects (a.k.a. entities) according to the entity information in the given prompt. These segmented entities are processed through an image-to-3D model [29, 67] to obtain coarse 3D shapes, which are used to roughly initialize the Gaussian parameters in 3D space, thereby transferring the 2D compositionality to 3D representations.", + "bbox": [ + 511, + 125, + 908, + 335 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Dynamic Optimization. Current 2D guidance [41, 47, 52] struggles with optimizing multiple 3D objects simultaneously in a scene, often leading to 3D inconsistencies and semantic shifts. To address these issues, we introduce a dynamic optimization strategy based on Score Distillation Sampling (SDS) loss, consisting of two key components: (1) COMPGS dynamically divides the training process to optimize different parts of 3D Gaussians. Specifically, it alternates between optimizing a single object (entity-level optimization) and the entire scene (composition-level optimization). This is achieved by labeling and filtering the Gaussian parameters for inference, and updating them through masking gradients. (2) COMPGS dynamically trains the entity-level Gaussians within a normalized 3D space, crucial for compositional 3D generation where object sizes vary. This is particularly challenging when dealing with very small objects, as optimizing Gaussian parameters in a limited 3D space may not adequately capture detailed textures. To mitigate this, we first scale the subspace of each entity to a predefined, standardized volume. Following each training iteration within these standardized volumes, we rescale the Gaussian parameters back to their original sizes. This method dynamically maintains volume consistency throughout the training process, thereby facilitating the capture of fine details within smaller objects.", + "bbox": [ + 511, + 339, + 908, + 715 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "With these designs, COMPGS is capable of generating high-quality compositional 3D objects (as shown in the first row of Fig. 1) and progressive 3D adding (as shown in the second row of Fig. 1). We demonstrate the effectiveness of COMPGS both qualitatively and quantitatively. Qualitative comparisons through our user study indicate that COMPGS offers superior image quality and semantic alignment compared to existing models (e.g., Fantasia3D [10], ProlificDreamer [71], VP3D [11], etc.). Besides, COMPGS's performance on $\\mathrm{T}^3$ Bench [26] quantitatively highlights its advantages in both semantic control and high-fidelity generation. We hope COMPGS can provide valuable insights into", + "bbox": [ + 511, + 719, + 910, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "18510", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "the research of compositional 3D generation.", + "bbox": [ + 91, + 90, + 387, + 106 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 119, + 232, + 133 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Multi-modality 3D Generation. Multi-modality 3D generation can basically categorized into the feed-forward system and optimization-based system. The former one is trained end-to-end on multi-view dataset [5, 17] for zero-shot generation, typically based on 3D representations, including NeRF [29, 67], 3D Gaussians [66, 73], tri-planes [59, 69] and feature grids [32]. Despite the excellent performance achieved, the single-object training data [5, 17, 18, 45] limits their generative abilities in scenarios involving multiple objects. Optimization-based 3D systems lift 2D diffusion priors [7, 8, 51, 56, 57] for 3D generation, and typically train 3D representations on a prompt-by-prompt basis. For example, DreamFusion [52] introduces Score Distillation Sampling (SDS) loss to transfer 2D diffusion models into the 3D domain. Magic3D [53] employs a coarse-to-fine scheme to improve both efficiency and effectiveness, while Fantasy3D [10] decouples the modelling of geometry and appearance. Despite the advancements of these methods addressing various challenges, they also usually struggle with compositional 3D generation. $\\mathrm{T}^3$ bench's examination of ten prominent optimization-based models [10, 41, 47, 52, 58, 63, 68, 71, 77] revealed frequent issues in compositional generation. The implicit 2D diffusion guidance used in these methods [10, 52, 58, 71] often fails to maintain 3D consistency across different views, leading to significant discrepancies in compositional 3D scenes.", + "bbox": [ + 89, + 143, + 485, + 535 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Compositional Generation. Compositional generation involves creating content involving multiple objects with logical interactions. These interactions include, but are not limited to, concept relations [43], attribute association with colors [6, 22, 50], and spatial relationships between objects [9, 72]. There is considerable focus on various aspects of compositional 2D generation, such as learning from human feedback [20, 37, 75, 80], enhancing image captions [3, 7, 8, 16], designing effective networks [44], and etc. T2I-CompBench [30] further proposed a comprehensive benchmark and boosted the compositionality of text-to-image models. Compared to 2D compositional generation, its 3D counterpart is under-explored due to 3D geometry and appearance complexities [1, 12, 25, 40, 64, 74, 76]. For example, Set-the-Scene [15] and Scenewiz3d [79] propose to adopt object layouts to generate compositional scenes. However, they are mainly based on the implicit NeRF representations, so the features between objects cannot be well decoupled, resulting in a relatively blurred rendering effect. VP3D [11] takes a different strategy to employ visual features for compositional generation. However, its visual features are only used as implicit supervision, which does not adequately ensure 3D consistency or address the Janus problems [58]. Besides, LucidDreaming [70] requires", + "bbox": [ + 89, + 537, + 486, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "human-annotated layout priors for local optimization, which is labour-intensive and inaccurate; while GraphDreamer [23] optimizes all interactive objects from scratch, which is time-consuming and difficult to optimize. The most concurrent works, GALA3D [81], propose optimizing the spatial relationships of well-trained Gaussians for compositional generation. Though it achieves plausible spatial interactions, it is inefficient in generating complex mutual interactions among objects. More comparisons are in the appendix. We propose COMPGS to address the above challenges, emphasizing efficient optimization and the generation of complex interactions.", + "bbox": [ + 511, + 90, + 908, + 272 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 511, + 279, + 648, + 296 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we revisit 3D Gaussian Splitting [33] and diffusion priors for 3D generation. We then provide an overview of COMPGS, which includes initializing 3D Gaussians to incorporate compositionality priors and dynamic SDS optimization to generate high-fidelity, 3D consistent objects with complex interactions. The notations used in this section are also detailed in Appendix A.1 for clarity.", + "bbox": [ + 511, + 304, + 908, + 411 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminaries", + "text_level": 1, + "bbox": [ + 511, + 419, + 653, + 434 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3D Gaussian Splatting (GS). GS [33] has recently revolutionized novel-view synthesis of objects/scenes by its real-time rendering. Specifically, GS represents the 3D objects/scenes by $N$ explicit anisotropic Gaussians with center positions $\\mu_{i}$ , covariances $\\Sigma_{i}$ , opacities $\\alpha_{i}$ and colors $c_{i}$ , where $i\\in N$ . The color $\\mathcal{C}(\\mathbf{p})$ of image pixel $\\mathbf{p}$ can be calculated through point-based volume rendering [35, 36] by integrating the color and density of the 3D Gaussians intersected by a ray, as follows:", + "bbox": [ + 511, + 440, + 908, + 566 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {C} (\\mathbf {p}) = \\sum_ {i = 1} ^ {N} c _ {i} \\sigma_ {i} \\prod_ {j = 1} ^ {i - 1} (1 - \\sigma_ {j}), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 558, + 575, + 906, + 614 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\sigma_ {i} = \\alpha_ {i} \\exp \\left[ - \\frac {1}{2} \\left(\\mathbf {p} - \\hat {\\boldsymbol {\\mu}} _ {i}\\right) ^ {\\top} \\hat {\\Sigma} _ {i} ^ {- 1} \\left(\\mathbf {p} - \\hat {\\boldsymbol {\\mu}} _ {i}\\right) \\right], \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 616, + 906, + 648 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\hat{\\mu}_i$ and $\\hat{\\Sigma}_i$ denote the projected center positions and covariances of the 2D Gaussians, transformed from 3D space to the 2D camera's image plane. The image plane can be segmented into tiles during rendering for parallel processing. Unlike implicit representations, such as NeRF [2, 48, 49], GS offers two key advantages for compositional 3D generation: (1) 3D Gaussians facilitate localized rendering, allowing for the rendering of object $A$ in one specific sub-space and object $B$ in another, thus enabling global compositional generation on a divide-and-conquer basis; (2) 3D Gaussians enable direct parameter initialization, simplifying the integration of compositional priors at the initialization stage.", + "bbox": [ + 511, + 657, + 908, + 839 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Diffusion Priors for 3D Generation. Diffusion-based generative models (DMs) [19, 60, 62] have been widely utilized to provide implicit priors for 3D object generation via score distillation sampling (SDS) [52]. Specifically, given a 3D", + "bbox": [ + 511, + 840, + 908, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "18511", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/14db37253ead1a77c3d0cef047e8e1ea58295e51bbb6dbc7bf6bd904574900bd.jpg", + "image_caption": [ + "Figure 2. Overall pipeline of COMPGS. Given a compositional prompt $V$ , we first use an LLM to decompose it into entity-level prompts $\\{v_l\\}$ , guiding the segmentation of each entity from the compositional image generated by T2I models. The segmented images initialize entity-level 3D Gaussians via image-to-3D models [29, 67]. COMPGS employs a dynamic optimization strategy, alternating between composition-level optimization of $\\theta$ and entity-level optimization of $\\{\\theta_l\\}$ . For entity-level optimization, COMPGS dynamically maintains volume consistency to refine the details of each objects, particularly the small one." + ], + "image_footnote": [], + "bbox": [ + 112, + 107, + 883, + 314 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "model whose parameters are $\\theta$ , a differentiable rendering process $g$ , the rendered images could be obtained via $\\mathrm{x} = g(\\theta)$ . To ensure the rendered images resemble those generated by the DM $\\phi$ , SDS first formulates the sampled noise $\\hat{\\epsilon}_{\\phi}(\\mathrm{z}_t;v,t)$ with the noisy image $\\mathrm{z}_t$ , text embedding $v$ , and noise level $t$ . By comparing the difference between the added Gaussian noise $\\epsilon$ and the predicted noise $\\hat{\\epsilon}_{\\phi}$ , SDS constructs gradients that could be back-propagated to update $\\theta$ via:", + "bbox": [ + 89, + 395, + 483, + 516 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\nabla_ {\\theta} \\mathcal {L} _ {\\mathrm {S D S}} (\\phi , \\mathrm {x} = g (\\theta)) \\triangleq \\mathbb {E} _ {t, \\epsilon} \\left[ w (t) \\left(\\hat {\\epsilon} _ {\\phi} \\left(\\mathbf {z} _ {t}, v, t\\right) - \\epsilon\\right) \\frac {\\partial \\mathrm {x}}{\\partial \\theta} \\right], \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 102, + 539, + 483, + 580 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $w(t)$ is a weighting function, and the Classifier-free guidance [27] typically amplifies the text conditioning $v$ .", + "bbox": [ + 89, + 580, + 482, + 612 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.COMPGS", + "text_level": 1, + "bbox": [ + 89, + 632, + 194, + 646 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We propose COMPGS for efficiently transferring the 2D compositionality to facilitate compositional 3D generation with 3D Gaussians. The overall framework of COMPGS is shown in Fig. 2. We first generate a well-composed image from the given complex prompt. After extracting the subobject (also named entity) information within the prompt by LLM, we utilize the entity prompt to segment the composed image into different parts. Each part will be adopted to initialize a specific space of Gaussians. During the training stage, we propose a dynamic SDS optimization strategy. This strategy first automatically decomposes the training to optimize either entity-level Gaussians or composition-level Gaussians. Then, it employs a volume-adaptive strategy to dynamically optimize varying-sized entities within a consistent 3D space. We detail the initialization process and dynamic SDS optimization in the following sections.", + "bbox": [ + 88, + 657, + 485, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.1. 3D Gaussians Init with 2D Compositionality", + "text_level": 1, + "bbox": [ + 511, + 393, + 864, + 410 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Unlike the implicit NeRF [48] representations, explicit 3D Gaussians can be easily initialized with 3D shapes and colors, which facilitates introducing rough 3D priors to ensure 3D consistency [77]. Although existing text-to-3D or image-to-3D models can generate a single 3D object, they struggle with compositional 3D generation as mentioned in Sec. 2, which hinders the creation of compositional 3D priors needed for initialization. Considering this, we propose initializing the compositional Gaussians on an entity-by-entity basis.", + "bbox": [ + 511, + 415, + 906, + 565 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As shown in Fig. 2 (left), given a complex prompt $V$ , (e.g., 'an owl perches on a branch near a pinecone'), we first adopt a 2D diffusion model [3] to generate a composed image $I$ that faithfully captures the compositionality of multiple objects. We extract each entity information in prompt $V$ by prompting LLM to obtain $L$ different entity-level prompts (i.e., $L = 3$ for $v_{1}$ 'an owl', $v_{2}$ 'a branch', and $v_{3}$ 'a pinecone'). These prompts are adopted in a text-guided segmentation model [34] to decompose image $I$ into various parts $\\{I_l\\}$ . Until now, each segmented image has only one entity, facilitating using existing image-based 3D generation models [67] to predict a rough triangle mesh $m_{l}(l \\in L)$ of the corresponding 3D entity $l$ . To initialize 3D Gaussians $\\theta_{l}$ , we index $N$ points from each mesh; the center positions $\\mu_i^l \\in \\mathbb{R}^3 (i \\in N, l \\in L)$ are the centers of each vertex of $m_{l}$ , and the texture colors $c_i^l \\in \\mathbb{R}^3 (i \\in N, l \\in L)$ are queried from each vertex of $m_{l}$ . During the image-to-3D process, we did not perform any cropping or padding on the image $I_{l}$ , ensuring that the spatial positions of each mesh $m_{l}$ correspond to their 2D spatial positions. Additionally, we have marked the 3D layouts of each entity asset with 3D bounding boxes bbox $l$ for subsequent optimization. The", + "bbox": [ + 509, + 568, + 908, + 902 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "$V = \\text{an owl perches on a branch near a pinecone}^{\\prime}$ ( $v_{1} = \\text{an owl}; v_{2} = \\text{a branch}; v_{3} = \\text{a pinecone}^{\\prime}$ )", + "bbox": [ + 116, + 90, + 723, + 104 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "18512", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "bounding box coordinates are determined by the outer-most center positions of the entity Gaussian.", + "bbox": [ + 89, + 90, + 485, + 122 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2.2. Dynamic SDS Optimization", + "text_level": 1, + "bbox": [ + 89, + 131, + 331, + 147 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As mentioned in Sec. 2, existing SDS methods [10, 52, 58, 71] struggle in optimizing compositional 3D scenes due to the challenges of using implicit diffusion priors to maintain multi-objects consistency and interactions. To address this, we introduce dynamic SDS optimization to enable existing SDS losses to remain effective in compositional generation. Dynamic SDS optimization is performed through the following two procedures.", + "bbox": [ + 89, + 152, + 485, + 272 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Automatically Decomposing Optimization to Different Levels. We adopt the Decomposed Optimization (DO) strategy that divides the entire training into $L + 1$ stages, including $L$ stages for entity-level optimization to ensure 3D consistency of each $L$ entities, and one stage for composition-level optimization to refine inter-entity interactions. In each training iteration, we randomly render either (1) an entity Gaussian $\\theta_{l}(l \\in L)$ to obtain the entity image $\\mathrm{x}_l = g(\\theta_l)$ ( $\\theta_{l} \\in \\mathrm{bbox}_{l}$ ), or (2) the composed Gaussians $\\theta$ to obtain the composed image $\\mathrm{x} = g(\\theta)$ . Here $g$ denotes the Gaussian Splitting rendering [33] described in Eq. 1.", + "bbox": [ + 89, + 275, + 483, + 441 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To update Gaussian parameters, we apply Eq. 3 for gradients backpropagation. Specifically, for entity-level optimization, we substitute the text embedding $v$ by $v_{l}$ and parameters $\\theta$ by $\\theta_{l}$ in Eq. 3 to optimize entity Gaussian. We adopt MV-Dream [58] as $\\phi$ , which provides multi-view diffusion priors that can better maintain 3D consistency. To ensure that other entity's parameters remain unchanged, we mask the gradients that are not within the corresponding 3D bounding boxes $\\mathrm{bbox}_l$ when updating $\\theta_{l}$ . For composition-level optimization, we integrate both 2D [56] and 3D diffusion priors [58] to jointly optimize the overall Gaussian parameters $\\theta$ . The rationale is that 2D priors promote geometry exploration [53], which enhances the generation of plausible interactions between different entities.", + "bbox": [ + 89, + 443, + 485, + 654 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Volume-adaptively Optimizing Entity-level Gaussians. Compared to the single object generation, compositional 3D generation using SDS presents additional challenges. Specifically, when objects vary in size, optimizing directly in the original 3D space often results in suboptimal generation for smaller objects, as shown the 'pinecone' in Fig. 2. To mitigate this issue, we propose a Volume-adaptive Optimization (VAO) strategy that enhances optimization across objects of varying sizes by dynamically standardizing each entity's space to a standardized volumetric scale. This approach is especially beneficial for improving the generation of fine-grained texture details in smaller entities.", + "bbox": [ + 89, + 656, + 485, + 838 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To scale the 3D bounding box of an entity Gaussian $\\theta_{l}$ with a standardized volumetric space bbox $^{std}$ , we first determine the necessary transformations for the centre positions $\\mu$ . Specifically, we calculate the shift parameters $\\beta$ , and", + "bbox": [ + 89, + 839, + 485, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "scale parameters $\\lambda$ , as follows:", + "bbox": [ + 513, + 90, + 720, + 106 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\beta = \\operatorname {M e a n} (\\mathrm {b b o x} _ {l}); \\quad \\lambda = \\mathrm {b b o x} _ {s t d} / \\mathrm {b b o x} _ {l}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 573, + 113, + 906, + 128 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where Mean computes the center coordinates of the given bounding box, and $\\mathrm{bbox}_{std}$ is the target standardized space. Then, we zoom-in the entity Gaussian on its center positions by $\\hat{\\mu} = \\lambda (\\mu -\\beta)$ , where $\\hat{\\mu}$ denotes the transformed center positions of entity Gaussian. We substitute all $\\mu$ to $\\hat{\\mu}$ in Eq. 1 for SDS optimization. After each training iteration, we perform zooming-back via $\\mu = \\hat{\\mu} /\\lambda +\\beta$ . The proposed transformation ensures dynamically conducting SDS optimization on each entity Gaussian at a standardized and consistent scale. Experiments in Sec. 4.3 demonstrate its effectiveness in generating high-quality 3D assets.", + "bbox": [ + 511, + 135, + 906, + 303 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Progressive Adding Objects with COMPGS", + "text_level": 1, + "bbox": [ + 511, + 306, + 874, + 321 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The dynamic SDS optimization enables COMPGS to explicitly control the generation of different parts in 3D scene. This capability can be utilized for progressive increasing 3D objects. Specifically, given well-trained 3D Gaussians $\\theta$ , we begin by rendering an image from the front view, denoted as $\\mathbf{x} = g(\\theta)$ . Unlike previous works that directly add 3D objects [13], we adopt MagicBrush [78] to modify the 2D rendered image $\\mathbf{x}$ . The edited 2D image can then be utilized to initialize new Gaussians $\\hat{\\theta}$ in the original 3D space. We further train the added 3D Gaussians to finally introduce new objects into the compositional scenes. As shown in Fig. 1, given well-trained 3D Gaussians (e.g., 'an owl'), we progressively modify the 2D images to add new objects, such as the 'branch' and 'pinecone.' We then train the corresponding 3D Gaussians via dynamic SDS optimization to incorporate these new elements into the 3D scenes.", + "bbox": [ + 511, + 328, + 908, + 569 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 582, + 648, + 599 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Settings", + "text_level": 1, + "bbox": [ + 511, + 606, + 720, + 623 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details. We implemented COMPGS using ThreeStudio [24]. We use DALL-E [3] to generate the well-established 2D compositionality, LangSAM [46] to segment the entity image, and TripoSR [67] to generate a preliminary 3D prior (i.e., mesh) from each entity image, respectively. During the training, we employed MVDream [58] to optimize entity-level Gaussians, and both MVDream [58] and stabilityai/stablediffusion-2-1-base [57] to optimize composition-level Gaussians. The guidance scale for diffusion models was set as 50. Timestamps were uniformly selected from 0.02 to 0.55 for the first 1,000 iterations and then adjusted to a range from 0.02 to 0.15 for subsequent iterations. We initialized the Gaussian points at $N = 10k$ , and progressively increased the density to a maximum of $1000k$ points. We optimized the overall compositional scenes through 10k iterations to achieve optimal results. However, it is worth noting that we empirically found that training with 5k iterations already produces", + "bbox": [ + 511, + 628, + 908, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "18513", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "plausible 3D assets. The learning rates are detailed in Appendix A.3. We set the camera parameters, including radius, azimuth, elevation and FoV (field of view) to be the same as [58]. Experiments were conducted on NVIDIA A100 GPUs(40G). The prompt used for decomposing are provided in Appendix A.11. Our code will be released.", + "bbox": [ + 89, + 90, + 480, + 181 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metrics. In addition to user studies and qualitative comparisons, we utilize $\\mathrm{T}^3$ Bench [26] to render 300 prompts in order to evaluate compositional 3D generation on two criteria: (1) the visual quality of the 3D objects and (2) the alignment between the 3D objects and the input prompt. For quality evaluation, we captured multi-focal and multiview images and sent them to text-image scoring models, CLIP [54], to obtain an average quality score of the generated 3D scene. Regarding the textual alignment, we first followed [45] to perform 3D captioning via BLIP [38] and GPT4, and then computed the recall of the original prompt within the generated caption via ROUGE-L [42].", + "bbox": [ + 89, + 183, + 482, + 363 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Performance Comparisons and Analysis.", + "text_level": 1, + "bbox": [ + 89, + 377, + 437, + 393 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Qualitative Evaluation. In Fig. 3, we present qualitative comparisons of compositional 3D generation. Our model, COMPGS, is compared with several open-source methods, including DreamFusion [52], Magic3D [53], Latent-NeRF [47], Fantasia3D [10], SJC [68], and ProlificDreamer [71], as well as methods specifically designed to handle intricate text prompts, VP3D [11] in Fig. 3. The results demonstrate that COMPGS generates compositional 3D objects with superior quality, greater consistency, and more plausible interactions. DreamFusion, for example, fails to generate reasonable multi-objects from compositional prompts, revealing limitations in SDS loss for optimizing multiple objects simultaneously. Magic3D and Latent-NeRF, though guided by a 3D mesh prior, struggle with generating complex and compositional 3D geometry. Despite employing advanced SDS loss, Fantasia3D, SJC, and ProlificDreamer only improve the appearance of the content but do not fundamentally address multi-object generation. Notably, VP3D employs features from compositional images and human feedback to guide compositional 3D generation. However, it often produces unexpected artifacts (e.g., redundant dots in the top two rows of Fig. 3) or less plausible interactions (e.g., strange spatial relationships in the fourth row). This is likely due to the insufficient utilization of 2D compositionality, as VP3D adopts the implicit image features to guide optimization. In contrast, COMPGS generates high-quality and composed content that strictly aligns with the given complex prompts, demonstrating its effectiveness over others. We provide more visual comparisons with both the close-source and open-source methods (including GALA3D [81] in Fig. 9, GraphDreamer [23], DreamGaussian [65] in Fig. 10, etc.) in Appendix A.5 as well as the attached video.", + "bbox": [ + 89, + 400, + 482, + 883 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We also demonstrate COMPGS's capabilities in generat", + "bbox": [ + 109, + 885, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/ef6b73e27fa21b844e02d6f4cd563a97b34d0fa9493c66cfdae747aa1d9435fb.jpg", + "table_caption": [ + "Table 1. Quantitative comparisons on $\\mathrm{T}^3$ Bench [26] and user studies show COMPGS outperforms feed-forward, optimization-based, and compositional generation models." + ], + "table_footnote": [], + "table_body": "
MethodT3Bench (Multiple Objects)User Study Ranking Score↑
Quality↑Alignment↑Average↑
OpenLRM [29]15.225.520.4-
TripoSR [67]16.728.622.7-
DreamFusion [52]17.314.816.14.54
SJC [68]17.75.811.73.08
LatentNeRF [47]21.719.520.64.14
Fantasia3D [10]22.714.318.52.43
ProlificDreamer [71]45.725.835.83.56
Magic3D [41]26.624.825.74.30
Set-the-Scene [15]20.829.925.4-
VP3D [11]49.131.540.36.71
COMPGS (ours)54.2 (+5.1)37.9(+6.4)46.1(+5.8)7.23
", + "bbox": [ + 517, + 135, + 903, + 292 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ing more diverse contents in Fig. 4. COMPGS accurately generates both simple spatial relationships (shown in the first row), and more complex interactions, where objects should react to each other (shown in the second row). Besides, COMPGS is not limited to generating just two objects, it can generate multiple objects by expanding the number of entity-level Gaussians. For example, in the third row, COMPGS generates scenes with three and four entities, each with high-quality 3D shapes and textures.", + "bbox": [ + 511, + 305, + 906, + 441 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We conduct a user study for further evaluation. We randomly selected 15 prompts in $\\mathrm{T}^3$ Bench, and collected the 3D assets generated by different models. These collected 3D assets were then distributed to individuals for ranking the models based on (1) 3D visual quality and (3) text-3D alignment. We average the ranking of different models as their scores. Results in Tab. 1 show that representative optimization-based models received relatively low average scores, highlighting their limitations in compositional generation. VP3D [11] outperforms its predecessors due to the incorporation of image features but still ranks lower than COMPGS, which further validates our effectiveness.", + "bbox": [ + 511, + 444, + 908, + 625 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Quantitative Evaluation. In Tab. 1, we benchmark the representative models on $\\mathrm{T}^3$ Bench, focusing on the multi-objects track. Results indicate COMPGS achieves superior performance in both quality and textual alignment. Compared with feed-forward methods in the first block, COMPGS significantly improves the generation quality by a large margin. Though COMPGS is initialized from the 3D priors obtained from [67], COMPGS still enhances text-3D alignment, likely due to the dynamic optimization on composition-level Gaussians. Furthermore, compared with the optimization-based methods, COMPGS shows significant advantages in both two metrics. Even when compared to methods specifically designed for compositional generation, e.g., VP3D [11] and Set-the-Scene [15], COMPGS demonstrates clear improvements, indicating the effectiveness of our designs. Note that COMPGS is not limited to compositional generation; the results in the track of single-object generation are included in Appendix A.5.", + "bbox": [ + 511, + 628, + 908, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "18514", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "'A florist is making a bouquet with fresh flowers'", + "bbox": [ + 354, + 87, + 643, + 101 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "'An artist is painting on a blank canvas'", + "bbox": [ + 377, + 189, + 617, + 202 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Hot popcorn jump out from the red striped popcorn maker", + "bbox": [ + 321, + 291, + 674, + 305 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "'A half-eaten sandwich sits next to a lukewarm thermos'", + "bbox": [ + 331, + 393, + 663, + 406 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "DreamFusion", + "bbox": [ + 91, + 492, + 158, + 503 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Magic3D", + "bbox": [ + 212, + 492, + 259, + 503 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "LatentNeRF", + "bbox": [ + 297, + 492, + 357, + 503 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Fantasia3D", + "bbox": [ + 401, + 492, + 457, + 503 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "SJC", + "bbox": [ + 526, + 492, + 549, + 502 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ProlificDreamer", + "bbox": [ + 607, + 492, + 684, + 502 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "VP3D", + "bbox": [ + 743, + 492, + 774, + 502 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "COMPGS (ours)", + "bbox": [ + 823, + 492, + 901, + 503 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0034243e454e8469075c79dde2ee3045ae2aec9a840090edfa6944b627f7a1ab.jpg", + "image_caption": [ + "Figure 3. Qualitative comparisons between COMPGS and other text-to-3D models on $\\mathrm{T}^3$ Bench (multiple objects track). Compared to others, COMPGS is better at generating highly-composed, high-quality 3D contents that strictly align with the given texts. Watch the animations by clicking them (Not all PDF readers support playing animations. Best viewed in Acrobat/Foxit Reader)." + ], + "image_footnote": [], + "bbox": [ + 96, + 553, + 483, + 797 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/fc33d919392e937f5707680aacf60f7db395bb1c9db7c8339123e54e93255f4a.jpg", + "image_caption": [ + "Figure 4. More generated samples by COMPGS. Four views are shown." + ], + "image_footnote": [], + "bbox": [ + 511, + 555, + 883, + 797 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Runtime Comparisons. Though COMPGS is trained with 10k steps, we observed that training the model for 5k iterations already produces high-quality content with minimal loss of texture details. We show runtime comparisons with other models in Appendix A.4. In fact, compositional", + "bbox": [ + 89, + 821, + 485, + 898 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3D generation (left) requires a longer training time than single-object generation (right) due to its complexity. Compositional 3D involves optimizing individual objects and ensuring the consistency of compositional scenes, making them more intricate. The runtime is generally relative to", + "bbox": [ + 511, + 821, + 908, + 898 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "18515", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7ed8b65f4d2728aec34d12a6988f4a028e98d31d371fa07d88445558537990e5.jpg", + "image_caption": [ + "Figure 5. 3D Editing examples of COMPGS. More examples could be found in Appendix A.10." + ], + "image_footnote": [], + "bbox": [ + 112, + 59, + 472, + 234 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7b4b21f09381cde132f41dc3ee7e5bb78a06d9bd541f3a94a5236c24ddafa9b2.jpg", + "image_caption": [ + "Figure 6. Visual results of the ablation studies on key designs." + ], + "image_footnote": [], + "bbox": [ + 107, + 268, + 465, + 436 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "the number of objects involved. Compared to open-source compositional 3D methods such as Set-the-Scene [15], Progressive3D [14], and GraphDraemer [23], our proposed COMPGS is more efficient in training. For instance, given the prompt \"a parrot talks beside a perch and two bowls,\" Progressive3D takes approximately 250 minutes for 3D generation, while Set-the-Scene requires around 110 minutes. Since many compositional scene generation methods are not open-sourced, we also present the training steps listed in the papers for straightforward comparisons. In addition, CompGS demonstrates comparable and even superior efficiency over other methods in text-to-single object generation. For example, methods such as Magic3D [41] and Fantastic3D [10] require over five hours to optimize a single object. In contrast, CompGS can achieve the same task in approximately 30 minutes. Details are in Appendix A.4.", + "bbox": [ + 89, + 470, + 485, + 712 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Extended Applications - Progressive 3D Adding. COMPGS allows progressive 3D adding for compositional scenes. We present 3D adding examples in Fig. 5. These examples demonstrate that by transferring edited 2D compositionality and employing dynamic SDS optimization, COMPGS can progressively incorporate new 3D entities into the original 3D scenes. For example, COMPGS can generate the 'chair', 'panda', 'hat' and 'plant' step by step as shown in Fig. 5.", + "bbox": [ + 89, + 724, + 483, + 859 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations. Detailed limitations and future explorations are provided in Appendix A.12 and Appendix A.13.", + "bbox": [ + 89, + 871, + 482, + 902 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/30b5a6525d2ac6ea7a84e1171a6a791afbf6819c9bcdee28c789a050c7fa167c.jpg", + "table_caption": [ + "Table 2. Ablation Studies on ${\\mathrm{T}}^{3}$ Bench [26]." + ], + "table_footnote": [], + "table_body": "
ComponentQualityAlignmentAverage
Full Setting53.838.045.9
- w/o GS Init.22.818.720.8
- w/o DO Strategy46.835.241.0
- w/o VAO Strategy50.836.443.6
", + "bbox": [ + 552, + 88, + 864, + 172 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 513, + 184, + 666, + 200 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We conduct ablation studies to validate the effectiveness of three key designs in COMPGS: Gaussian parameters initialization, decomposed optimization (DO), and volume-adaptive optimization (VAO). We randomly choose 20 prompts from $\\mathrm{T}^3$ Bench for the quantitative evaluation. Meanwhile, we use Fig. 6 to visualize the effect of each component.", + "bbox": [ + 511, + 207, + 906, + 313 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Initialization. Instead of initializing Gaussian parameters using 2D compositionality, we conduct random initialization within a predefined 3D bounding box for each entity Gaussian. A significant decrease in the quality and alignment metric is observed in Tab. 2. Besides, the visualization in Fig. 6 confirms that training from random initialization with predefined layouts may result in low-quality textures (e.g., owl's face) and missing entities (e.g., the branch).", + "bbox": [ + 511, + 314, + 906, + 435 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Decomposed Optimization. We analyze the decomposed training by discarding the entity-level optimization. Results in Tab. 2 and Fig. 6 both indicate that removing decomposed training leads to low-fidelity generation. For example, in Fig. 6, optimizing composition-level Gaussians results in a relatively blurred generation on the owl's fur and the pinecone, validating that optimization on each single entity is necessary to effectively leverage 2D diffusion priors to guide 3D generation.", + "bbox": [ + 511, + 436, + 908, + 571 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Volume-adaptive Optimization. To evaluate the effectiveness of volume-adaptive optimization, we choose to optimize the entity Gaussians in their original 3D space. We observed that optimizing different entities within varying sizes of 3D space slightly decreased in quality and alignment, which is noticeable for smaller objects in the compositional scene. For instance, the 'pinecone' and 'branch' trained without volume-adaptive optimization exhibited fewer fine-grained details, shown as the green leaves on the branch and the shadow on the pinecone in Fig. 6.", + "bbox": [ + 511, + 573, + 908, + 724 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 738, + 633, + 753 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we introduce COMPGS, a user-friendly, optimization-based framework, that achieves compositional text-to-3D generation utilizing Gaussian Splitting. Our three core designs, including initializing 3D Gaussians with 2D compositionality, decomposed optimization for either entity-level or composition-level Gaussians, and volume-adaptive optimization, enable COMPGS to generate complex interactions between multiple 3D entities. We hope this work encourages further research in compositional 3D generation.", + "bbox": [ + 511, + 763, + 908, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "18516", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgment The work was supported by Berkeley DeepDrive. The authors thank Google Cloud Platform for providing the computational resources. The work was done during Chongjian Ge's visit to University of California, Berkeley.", + "bbox": [ + 89, + 90, + 485, + 167 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 184, + 187, + 200 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Haotian Bai, Yuanhuiyi Lyu, Lutao Jiang, Sijia Li, Haonan Lu, Xiaodong Lin, and Lin Wang. Componerf: Text-guided multi-object compositional nerf with editable 3d scene layout. arXiv preprint arXiv:2303.13843, 2023.", + "[2] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In IEEE/CVF International Conference on Computer Vision, 2021.", + "[3] James Betker, Gabriel Goh, Li Jing, Tim Brooks, Jianfeng Wang, Linjie Li, Long Ouyang, Juntang Zhuang, Joyce Lee, Yufei Guo, et al. Improving image generation with better captions. Computer Science. https://cdn.openaia.com/papers/dall-e-3.pdf, 2023.", + "[4] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023.", + "[5] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015.", + "[6] Hila Chefer, Yuval Alaluf, Yael Vinker, Lior Wolf, and Daniel Cohen-Or. Attend-and-excite: Attention-based semantic guidance for text-to-image diffusion models. ACM Transactions on Graphics (TOG), 2023.", + "[7] Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, et al. Pixart-alpha: Fast training of diffusion transformer for photorealistic text-to-image synthesis. 2023.", + "[8] Junsong Chen, Chongjian Ge, Enze Xie, Yue Wu, Lewei Yao, Xiaozhe Ren, Zhongdao Wang, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart- $\\backslash$ sigma: Weak-to-strong training of diffusion transformer for 4k text-to-image generation. arXiv preprint arXiv:2403.04692, 2024.", + "[9] Minghao Chen, Iro Laina, and Andrea Vedaldi. Training-free layout control with cross-attention guidance. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, 2024.", + "[10] Rui Chen, Yongwei Chen, Ningxin Jiao, and Kui Jia. Fantasia3d: Disentangling geometry and appearance for high-quality text-to-3d content creation. In IEEE/CVF International Conference on Computer Vision, 2023.", + "[11] Yang Chen, Yingwei Pan, Haibo Yang, Ting Yao, and Tao Mei. Vp3d: Unleashing 2d visual prompt for text-to-3d generation. arXiv preprint arXiv:2403.17001, 2024." + ], + "bbox": [ + 93, + 209, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[12] Yongwei Chen, Tengfei Wang, Tong Wu, Xingang Pan, Kui Jia, and Ziwei Liu. Comboverse: Compositional 3d assets creation using spatially-aware diffusion guidance. arXiv preprint arXiv:2403.12409, 2024.", + "[13] Kai Cheng, Xiaoxiao Long, Kaizhi Yang, Yao Yao, Wei Yin, Yuexin Ma, Wenping Wang, and Xuejin Chen. Gaussianpro: 3d gaussian splatting with progressive propagation. arXiv preprint arXiv:2402.14650, 2024.", + "[14] Xinhua Cheng, Tianyu Yang, Jianan Wang, Yu Li, Lei Zhang, Jian Zhang, and Li Yuan. Progressive3d: Progressively local editing for text-to-3d content creation with complex semantic prompts. arXiv preprint arXiv:2310.11784, 2023.", + "[15] Dana Cohen-Bar, Elad Richardson, Gal Metzer, Raja Giryes, and Daniel Cohen-Or. Set-the-scene: Global-local training for generating controllable nerf scenes. In IEEE/CVF International Conference on Computer Vision, 2023.", + "[16] Xiaoliang Dai, Ji Hou, Chih-Yao Ma, Sam Tsai, Jialiang Wang, Rui Wang, Peizhao Zhang, Simon Vandenhende, Xiaofang Wang, Abhimanyu Dubey, et al. Emu: Enhancing image generation models using photogenic needles in a haystack. arXiv preprint arXiv:2309.15807, 2023.", + "[17] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objaverse: A universe of annotated 3d objects. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023.", + "[18] Matt Deitke, Ruoshi Liu, Matthew Wallingford, Huong Ngo, Oscar Michel, Aditya Kusupati, Alan Fan, Christian Laforte, Vikram Voleti, Samir Yitzhak Gadre, et al. Objverse-xl: A universe of $10\\mathrm{m} + 3\\mathrm{d}$ objects. Advances in Neural Information Processing Systems, 2024.", + "[19] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in Neural Information Processing Systems, 2021.", + "[20] Hanze Dong, Wei Xiong, Deepanshu Goyal, Yihan Zhang, Winnie Chow, Rui Pan, Shizhe Diao, Jipeng Zhang, Kashun Shum, and Tong Zhang. Raft: Reward ranked finetuning for generative foundation model alignment. arXiv preprint arXiv:2304.06767, 2023.", + "[21] Danny Driess, Zhiao Huang, Yunzhu Li, Russ Tedrake, and Marc Toussaint. Learning multi-object dynamics with compositional neural radiance fields. In Conference on robot learning, pages 1755-1768. PMLR, 2023.", + "[22] Weixi Feng, Xuehai He, Tsu-Jui Fu, Varun Jampani, Arjun Akula, Pradyumna Narayana, Sugato Basu, Xin Eric Wang, and William Yang Wang. Training-free structured diffusion guidance for compositional text-to-image synthesis. arXiv preprint arXiv:2212.05032, 2022.", + "[23] Gege Gao, Weiyang Liu, Anpei Chen, Andreas Geiger, and Bernhard Scholkopf. Graphdreamer: Compositional 3d scene synthesis from scene graphs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21295-21304, 2024.", + "[24] Yuan-Chen Guo, Ying-Tian Liu, Ruizhi Shao, Christian Laforte, Vikram Voleti, Guan Luo, Chia-Hao Chen, Zixin Zou, Chen Wang, Yan-Pei Cao, and Song-Hai Zhang. threestudio: A unified framework for 3d content generation." + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "18517", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "https://github.com/threestudio-project/threestudio, 2023.", + "[25] Haonan Han, Rui Yang, Huan Liao, Jiankai Xing, Zunnan Xu, Xiaoming Yu, Junwei Zha, Xiu Li, and Wanhua Li. Reparo: Compositional 3d assets generation with differentiable 3d layout alignment. arXiv preprint arXiv:2405.18525, 2024.", + "[26] Yuze He, Yushi Bai, Matthieu Lin, Wang Zhao, Yubin Hu, Jenny Sheng, Ran Yi, Juanzi Li, and Yong-Jin Liu. T3bench: Benchmarking current progress in text-to-3d generation. arXiv preprint arXiv:2310.02977, 2023.", + "[27] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. Advances in Neural Information Processing Systems, 2022.", + "[28] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 2020.", + "[29] Yicong Hong, Kai Zhang, Jiuxiang Gu, Sai Bi, Yang Zhou, Difan Liu, Feng Liu, Kalyan Sunkavalli, Trung Bui, and Hao Tan. Lrm: Large reconstruction model for single image to 3d. arXiv preprint arXiv:2311.04400, 2023.", + "[30] Kaiyi Huang, Kaiyue Sun, Enze Xie, Zhenguo Li, and Xihui Liu. T2i-compbench: A comprehensive benchmark for open-world compositional text-to-image generation. Advances in Neural Information Processing Systems, 2023.", + "[31] Heewoo Jun and Alex Nichol. Shap-e: Generating conditional 3d implicit functions. arXiv preprint arXiv:2305.02463, 2023.", + "[32] Animesh Karnewar, Andrea Vedaldi, David Novotny, and Niloy J Mitra. Holodiffusion: Training a 3d diffusion model using 2d images. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023.", + "[33] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splattering for real-time radiance field rendering. ACM Transactions on Graphics, 2023.", + "[34] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In IEEE/CVF International Conference on Computer Vision, 2023.", + "[35] Georgios Kopanas, Julien Philip, Thomas Leimkuhler, and George Drettakis. Point-based neural rendering with per-view optimization. In Computer Graphics Forum, 2021.", + "[36] Georgios Kopanas, Thomas Leimkuhler, Gilles Rainer, Clément Jambon, and George Drettakis. Neural point catauastics for novel-view synthesis of reflections. ACM Transactions on Graphics (TOG), 2022.", + "[37] Kimin Lee, Hao Liu, Moonkyung Ryu, Olivia Watkins, Yuqing Du, Craig Boutilier, Pieter Abbeel, Mohammad Ghavamzadeh, and Shixiang Shane Gu. Aligning text-to-image models using human feedback. arXiv preprint arXiv:2302.12192, 2023.", + "[38] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International Conference on Machine Learning, 2022.", + "[39] Jiahao Li, Hao Tan, Kai Zhang, Zexiang Xu, Fujun Luan, Yinghao Xu, Yicong Hong, Kalyan Sunkavalli, Greg Shakhnarovich, and Sai Bi. Instant3d: Fast text-to-3d with" + ], + "bbox": [ + 91, + 92, + 483, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "sparse-view generation and large reconstruction model. arXiv preprint arXiv:2311.06214, 2023.", + "[40] Runjia Li, Junlin Han, Luke Melas-Kyriazi, Chunyi Sun, Zhaochong An, Zhongrui Gui, Shuyang Sun, Philip Torr, and Tomas Jakab. Dreambeast: Distilling 3d fantastical animals with part-aware knowledge transfer. arXiv preprint arXiv:2409.08271, 2024.", + "[41] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023.", + "[42] Chin-Yew Lin. Rouge: A package for automatic evaluation of summaries. In Text summarization branches out, 2004.", + "[43] Nan Liu, Shuang Li, Yilun Du, Antonio Torralba, and Joshua B Tenenbaum. Compositional visual generation with composable diffusion models. In European Conference on Computer Vision, 2022.", + "[44] Nan Liu, Shuang Li, Yilun Du, Antonio Torralba, and Joshua B Tenenbaum. Compositional visual generation with composable diffusion models. In European Conference on Computer Vision, 2022.", + "[45] Tiange Luo, Chris Rockwell, Honglak Lee, and Justin Johnson. Scalable 3d captioning with pretrained models. Advances in Neural Information Processing Systems, 2024.", + "[46] Luca Medeiros. Language segment-anything: Sam with text prompt. https://github.com/luca-medeiros/ lang-segment-anything, 2024. Accessed: 2024-05-21.", + "[47] Gal Metzer, Elad Richardson, Or Patashnik, Raja Giryes, and Daniel Cohen-Or. Latent-nerf for shape-guided generation of 3d shapes and textures. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023.", + "[48] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 2021.", + "[49] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM transactions on graphics (TOG), 2022.", + "[50] Dong Huk Park, Samaneh Azadi, Xihui Liu, Trevor Darrell, and Anna Rohrbach. Benchmark for compositional text-to-image synthesis. In Advances in Neural Information Processing Systems, 2021.", + "[51] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023.", + "[52] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022.", + "[53] Guocheng Qian, Jinjie Mai, Abdullah Hamdi, Jian Ren, Aliaksandr Siarohin, Bing Li, Hsin-Ying Lee, Ivan Skorokhodov, Peter Wonka, Sergey Tulyakov, et al. Magic123: One image to high-quality 3d object generation using both 2d and 3d diffusion priors. arXiv preprint arXiv:2306.17843, 2023." + ], + "bbox": [ + 516, + 92, + 906, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "18518", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[54] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, 2021.", + "[55] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning, 2021.", + "[56] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022.", + "[57] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022.", + "[58] Yichun Shi, Peng Wang, Jianglong Ye, Mai Long, Kejie Li, and Xiao Yang. Mvdream: Multi-view diffusion for 3d generation. arXiv preprint arXiv:2308.16512, 2023.", + "[59] J Ryan Shue, Eric Ryan Chan, Ryan Po, Zachary Ankner, Jiajun Wu, and Gordon Wetzstein. 3d neural field generation using triplane diffusion. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023.", + "[60] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International Conference on Machine Learning, 2015.", + "[61] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020.", + "[62] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020.", + "[63] Jingxiang Sun, Bo Zhang, Ruizhi Shao, Lizhen Wang, Wen Liu, Zhenda Xie, and Yebin Liu. Dreamcraft3d: Hierarchical 3d generation with bootstrapped diffusion prior. arXiv preprint arXiv:2310.16818, 2023.", + "[64] Shaorong Sun, Shuchao Pang, Yazhou Yao, and Xiaoshui Huang. Comogen: A controllable text-to-3d multi-object generation framework. arXiv preprint arXiv:2409.00590, 2024.", + "[65] Jiaxiang Tang, Jiawei Ren, Hang Zhou, Ziwei Liu, and Gang Zeng. Dreamgaussian: Generative gaussian splatting for efficient 3d content creation. arXiv preprint arXiv:2309.16653, 2023.", + "[66] Jiaxiang Tang, Zhaoxi Chen, Xiaokang Chen, Tengfei Wang, Gang Zeng, and Ziwei Liu. Lgm: Large multi-view gaussian model for high-resolution 3d content creation. arXiv preprint arXiv:2402.05054, 2024.", + "[67] Dmitry Tochilkin, David Pankratz, Zexiang Liu, Zixuan Huang, Adam Letts, Yangguang Li, Ding Liang, Christian Laforte, Varun Jampani, and Yan-Pei Cao. Triposr: Fast 3d object reconstruction from a single image. arXiv preprint arXiv:2403.02151, 2024." + ], + "bbox": [ + 91, + 90, + 485, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[68] Haochen Wang, Xiaodan Du, Jiahao Li, Raymond A Yeh, and Greg Shakhnarovich. Score jacobian chaining: Lifting pretrained 2d diffusion models for 3d generation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023.", + "[69] Tengfei Wang, Bo Zhang, Ting Zhang, Shuyang Gu, Jianmin Bao, Tadas Baltrusaitis, Jingjing Shen, Dong Chen, Fang Wen, Qifeng Chen, et al. Rodin: A generative model for sculpting 3d digital avatars using diffusion. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023.", + "[70] Zhaoning Wang, Ming Li, and Chen Chen. Luciddreaming: Controllable object-centric 3d generation. arXiv preprint arXiv:2312.00588, 2023.", + "[71] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. Advances in Neural Information Processing Systems, 2024.", + "[72] Qiucheng Wu, Yujuan Liu, Handong Zhao, Trung Bui, Zhe Lin, Yang Zhang, and Shiyu Chang. Harnessing the spatial-temporal attention of diffusion models for high-fidelity text-to-image synthesis. In IEEE/CVF International Conference on Computer Vision, 2023.", + "[73] Yinghao Xu, Zifan Shi, Wang Yifan, Hansheng Chen, Ceyuan Yang, Sida Peng, Yujun Shen, and Gordon Wetzstein. Grm: Large gaussian reconstruction model for efficient 3d reconstruction and generation. arXiv preprint arXiv:2403.14621, 2024.", + "[74] Han Yan, Yang Li, Zhennan Wu, Shenzhen Chen, Weixuan Sun, Taizhang Shang, Weizhe Liu, Tian Chen, Xiaqiang Dai, Chao Ma, et al. Frankenstein: Generating semantic-compositional 3d scenes in one tri-plane. 2024.", + "[75] Ling Yang, Zhaochen Yu, Chenlin Meng, Minkai Xu, Stefano Ermon, and Bin Cui. Mastering text-to-image diffusion: Recognition, planning, and generating with multimodal llms. arXiv preprint arXiv:2401.11708, 2024.", + "[76] Ling Yang, Zixiang Zhang, Junlin Han, Bohan Zeng, Runjia Li, Philip Torr, and Wentao Zhang. Semantic score distillation sampling for compositional text-to-3d generation. arXiv preprint arXiv:2410.09009, 2024.", + "[77] Taoran Yi, Jiemin Fang, Junjie Wang, Guanjun Wu, Lingxi Xie, Xiaopeng Zhang, Wenyu Liu, Qi Tian, and Xinggang Wang. Gaussiandreamer: Fast generation from text to 3d gaussians by bridging 2d and 3d diffusion models. arXiv preprint arXiv, 2023.", + "[78] Kai Zhang, Lingbo Mo, Wenhu Chen, Huan Sun, and Yu Su. Magicbrush: A manually annotated dataset for instruction-guided image editing. Advances in Neural Information Processing Systems, 2024.", + "[79] Qihang Zhang, Chaoyang Wang, Aliaksandr Siarohin, Peiye Zhuang, Yinghao Xu, Ceyuan Yang, Dahua Lin, Bolei Zhou, Sergey Tulyakov, and Hsin-Ying Lee. Towards text-guided 3d scene composition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6829-6838, 2024.", + "[80] Shu Zhang, Xinyi Yang, Yihao Feng, Can Qin, Chia-Chih Chen, Ning Yu, Zeyuan Chen, Huan Wang, Silvio Savarese," + ], + "bbox": [ + 516, + 90, + 906, + 901 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "18519", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Stefano Ermon, et al. Hive: Harnessing human feedback for instructional visual editing. arXiv preprint arXiv:2303.09618, 2023.", + "[81] Xiaoyu Zhou, Xingjian Ran, Yajiao Xiong, Jinlin He, Zhiwei Lin, Yongtao Wang, Deqing Sun, and Ming-Hsuan Yang. Gala3d: Towards text-to-3d complex scene generation via layout-guided generative gaussian splatting. arXiv preprint arXiv:2402.07207, 2024." + ], + "bbox": [ + 91, + 90, + 483, + 204 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "18520", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2025/CompGS_ Unleashing 2D Compositionality for Compositional Text-to-3D via Dynamically Optimizing 3D Gaussians/d9350216-461c-42c5-bc6e-3da77f71b98b_model.json b/2025/CompGS_ Unleashing 2D Compositionality for Compositional Text-to-3D via Dynamically Optimizing 3D Gaussians/d9350216-461c-42c5-bc6e-3da77f71b98b_model.json new file mode 100644 index 0000000000000000000000000000000000000000..65a1821891debdd37d9ab524113e3e99fc40abbf --- /dev/null +++ b/2025/CompGS_ Unleashing 2D Compositionality for Compositional Text-to-3D via Dynamically Optimizing 3D Gaussians/d9350216-461c-42c5-bc6e-3da77f71b98b_model.json @@ -0,0 +1,2589 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.02 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation." + }, + { + "type": "header", + "bbox": [ + 0.325, + 0.017, + 0.725, + 0.032 + ], + "angle": 0, + "content": "Except for this watermark, it is identical to the accepted version;" + }, + { + "type": "header", + "bbox": [ + 0.294, + 0.032, + 0.756, + 0.047 + ], + "angle": 0, + "content": "the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.111, + 0.13, + 0.89, + 0.177 + ], + "angle": 0, + "content": "COMPGS: Unleashing 2D Compositionality for Compositional Text-to-3D via Dynamically Optimizing 3D Gaussians" + }, + { + "type": "text", + "bbox": [ + 0.238, + 0.203, + 0.758, + 0.222 + ], + "angle": 0, + "content": "Chongjian Ge\\(^{1,2,3}\\) Chenfeng Xu\\(^{1}\\) Yuanfeng Ji\\(^{2}\\) Chensheng Peng\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.182, + 0.222, + 0.82, + 0.24 + ], + "angle": 0, + "content": "Masayoshi Tomizuka1 Ping Luo2 Mingyu Ding1,4* Varun Jampani5* Wei Zhan1" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.24, + 0.777, + 0.258 + ], + "angle": 0, + "content": "1 University of California, Berkeley 2 The University of Hong Kong" + }, + { + "type": "text", + "bbox": [ + 0.274, + 0.257, + 0.725, + 0.275 + ], + "angle": 0, + "content": "3 Adobe Research 4 UNC-Chapel Hill 5 Stability AI" + }, + { + "type": "image_caption", + "bbox": [ + 0.144, + 0.298, + 0.277, + 0.312 + ], + "angle": 0, + "content": "Single 3D Generation" + }, + { + "type": "image", + "bbox": [ + 0.132, + 0.318, + 0.285, + 0.389 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.145, + 0.389, + 0.274, + 0.402 + ], + "angle": 0, + "content": "a beautiful butterfly" + }, + { + "type": "image", + "bbox": [ + 0.143, + 0.413, + 0.289, + 0.482 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.186, + 0.489, + 0.228, + 0.499 + ], + "angle": 0, + "content": "an owl" + }, + { + "type": "image_caption", + "bbox": [ + 0.495, + 0.298, + 0.677, + 0.312 + ], + "angle": 0, + "content": "Compositional 3D Generation" + }, + { + "type": "image", + "bbox": [ + 0.315, + 0.316, + 0.493, + 0.384 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.313, + 0.388, + 0.454, + 0.412 + ], + "angle": 0, + "content": "a footballer is kicking a soccer ball" + }, + { + "type": "image", + "bbox": [ + 0.316, + 0.415, + 0.484, + 0.485 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.331, + 0.489, + 0.452, + 0.512 + ], + "angle": 0, + "content": "an owl perches on a branch" + }, + { + "type": "image", + "bbox": [ + 0.515, + 0.324, + 0.56, + 0.382 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.575, + 0.333, + 0.663, + 0.379 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.492, + 0.388, + 0.635, + 0.413 + ], + "angle": 0, + "content": "a bird is drinking water from a cup" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.418, + 0.655, + 0.483 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.497, + 0.489, + 0.633, + 0.512 + ], + "angle": 0, + "content": "an owl perches on a branch near a pinecone" + }, + { + "type": "image", + "bbox": [ + 0.677, + 0.316, + 0.772, + 0.387 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.728, + 0.318, + 0.808, + 0.368 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.819, + 0.316, + 0.862, + 0.387 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.659, + 0.389, + 0.854, + 0.413 + ], + "angle": 0, + "content": "a photographer is capturing abeautiful butterfly with camera" + }, + { + "type": "image", + "bbox": [ + 0.681, + 0.414, + 0.762, + 0.483 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.781, + 0.416, + 0.851, + 0.484 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.653, + 0.488, + 0.859, + 0.524 + ], + "angle": 0, + "content": "an owl perches on a branch near a pinecone, with a rat below the branch" + }, + { + "type": "image", + "bbox": [ + 0.129, + 0.525, + 0.87, + 0.636 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.637, + 0.907, + 0.72 + ], + "angle": 0, + "content": "Figure 1. Illustration of compositional 3D Generation and COMPGS. All the contents are generated by COMPGS. Top row: COMPGS is capable of generating either a single object (e.g., a butterfly) or generating compositional objects with reasonable interactions (e.g., the rightmost figure in the top row). Middle row: Beyond text-to-3D generation, COMPGS can be easily extend to 3D editing by progressively adding objects. The colored texts (e.g., 'a branch', 'a pinecone', 'a rat' in the rightmost figure) denote the added part compared to its previous asset. Bottom row: COMPGS achieves compositional text-to-3D by transferring 2D compositionality to initialize 3D Gaussians. COMPGS is further trained with dynamic SDS optimization to produce plausible results." + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.725, + 0.327, + 0.74 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.486, + 0.902 + ], + "angle": 0, + "content": "Recent breakthroughs in text-guided image generation have significantly advanced the field of 3D generation. While generating a single high-quality 3D object is now feasible, generating multiple objects with reasonable interactions within a 3D space, a.k.a. compositional 3D generation, presents substantial challenges. This paper introduces COMPGS, a novel generative framework that employs 3D Gaussian Splatting (GS) for efficient, compositional text-to-3D content generation. To achieve this goal, two core designs are proposed:" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.726, + 0.909, + 0.894 + ], + "angle": 0, + "content": "(1) 3D Gaussians Initialization with 2D compositionality: We transfer the well-established 2D compositionality to initialize the Gaussian parameters on an entity-by-entity basis, ensuring both consistent 3D priors for each entity and reasonable interactions among multiple entities; (2) Dynamic Optimization: We propose a dynamic strategy to optimize 3D Gaussians using Score Distillation Sampling (SDS) loss. COMPGS first automatically decomposes 3D Gaussians into distinct entity parts, enabling optimization at both the entity and composition levels. Additionally, COMPGS optimizes across objects of varying scales by dynamically adjusting the" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18509" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.089, + 0.092, + 0.488, + 0.23 + ], + "angle": 0, + "content": "spatial parameters of each entity, enhancing the generation of fine-grained details, particularly in smaller entities. Qualitative comparisons and quantitative evaluations on \\(\\mathrm{T}^3\\) Bench demonstrate the effectiveness of COMPGS in generating compositional 3D objects with superior image quality and semantic alignment over existing methods. COMPGS can also be easily extended to progressively adding 3D objects, facilitating complex scene generation. We hope COMPGS will provide new insights to the compositional 3D generation." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.255, + 0.226, + 0.272 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.281, + 0.486, + 0.522 + ], + "angle": 0, + "content": "3D content creation is essential to the modern media industry, yet it has traditionally been labor-intensive and necessitated professional expertise. Typically, designing a single 3D object takes several hours for an experienced designer, and creating complex scenes with multiple 3D objects (as shown in Fig. 1) requires even more effort. Inspired by the recent success of diffusion models in the text-to-image generation [28, 51, 56, 61], much research has focused on exploring 2D diffusion models for text-to-3D generation. Previous work can be divided into two main methodologies: (1) Feed-forward generation [29, 39, 73], which entails training generalizable diffusion models on 3D data; and (2) Optimization-based generation [10, 41, 47, 52, 68, 71], which utilizes the pretrained 2D diffusion guidance to optimize 3D representations via Score Distillation Sampling (SDS) [52]." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.523, + 0.485, + 0.869 + ], + "angle": 0, + "content": "While existing work has demonstrated the feasibility of generating single 3D objects, they often struggle to produce compositional 3D content with multiple objects and complex interactions. For example, (1) Feed-forward generation methods struggle in generalizing to complex textual descriptions since the amount of 3D training data is extremely limited [17, 18], and most of the data contains only one object; (2) Optimization-based generation methods face significant challenges with current 2D diffusion guidance when optimizing compositional 3D objects. Typically, using 2D diffusion guidance to optimize a single object is feasible, as it's easy to incorporate various attributes into a single object. However, in compositional 3D generation, 2D guidance struggles to accurately compose multiple objects with different attributes and relationships into a coherent scene [30]. For example, given the prompt 'a blue bench on the left of a green car,' distinguishing different attributes within the implicit 2D diffusion priors is challenging, resulting in misaligning attributes to different objects or generating incorrect spatial relationships. Thus, optimization-based compositional generation with standard 2D guidance often causes problems like 3D inconsistencies, multi-faced objects, semantic drift, etc [26, 58]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.485, + 0.903 + ], + "angle": 0, + "content": "In this work, we propose COMPGS, a generative system based on 3D Gaussian Splatting (GS) [33] for compositional" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.122 + ], + "angle": 0, + "content": "text-to-3D generation. To achieve this, we introduce two core designs:" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.125, + 0.909, + 0.337 + ], + "angle": 0, + "content": "3D Gaussians Initialization with 2D Compositionality. Unlike the implicit representation, i.e., NeRF [48], COMPGS uses 3D Gaussians as the representation, which helps to achieve feasible parameter initialization with a coarse 3D shape [31, 77]. As shown in the bottom row of Fig. 1, we first apply a text-to-image model [3, 7, 55] to generate an image that accurately captures the compositionality of multiple objects. Then the image is segmented into different sub-objects (a.k.a. entities) according to the entity information in the given prompt. These segmented entities are processed through an image-to-3D model [29, 67] to obtain coarse 3D shapes, which are used to roughly initialize the Gaussian parameters in 3D space, thereby transferring the 2D compositionality to 3D representations." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.34, + 0.909, + 0.717 + ], + "angle": 0, + "content": "Dynamic Optimization. Current 2D guidance [41, 47, 52] struggles with optimizing multiple 3D objects simultaneously in a scene, often leading to 3D inconsistencies and semantic shifts. To address these issues, we introduce a dynamic optimization strategy based on Score Distillation Sampling (SDS) loss, consisting of two key components: (1) COMPGS dynamically divides the training process to optimize different parts of 3D Gaussians. Specifically, it alternates between optimizing a single object (entity-level optimization) and the entire scene (composition-level optimization). This is achieved by labeling and filtering the Gaussian parameters for inference, and updating them through masking gradients. (2) COMPGS dynamically trains the entity-level Gaussians within a normalized 3D space, crucial for compositional 3D generation where object sizes vary. This is particularly challenging when dealing with very small objects, as optimizing Gaussian parameters in a limited 3D space may not adequately capture detailed textures. To mitigate this, we first scale the subspace of each entity to a predefined, standardized volume. Following each training iteration within these standardized volumes, we rescale the Gaussian parameters back to their original sizes. This method dynamically maintains volume consistency throughout the training process, thereby facilitating the capture of fine details within smaller objects." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.911, + 0.902 + ], + "angle": 0, + "content": "With these designs, COMPGS is capable of generating high-quality compositional 3D objects (as shown in the first row of Fig. 1) and progressive 3D adding (as shown in the second row of Fig. 1). We demonstrate the effectiveness of COMPGS both qualitatively and quantitatively. Qualitative comparisons through our user study indicate that COMPGS offers superior image quality and semantic alignment compared to existing models (e.g., Fantasia3D [10], ProlificDreamer [71], VP3D [11], etc.). Besides, COMPGS's performance on \\(\\mathrm{T}^3\\) Bench [26] quantitatively highlights its advantages in both semantic control and high-fidelity generation. We hope COMPGS can provide valuable insights into" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "18510" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.092, + 0.092, + 0.388, + 0.107 + ], + "angle": 0, + "content": "the research of compositional 3D generation." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.12, + 0.233, + 0.135 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.144, + 0.486, + 0.536 + ], + "angle": 0, + "content": "Multi-modality 3D Generation. Multi-modality 3D generation can basically categorized into the feed-forward system and optimization-based system. The former one is trained end-to-end on multi-view dataset [5, 17] for zero-shot generation, typically based on 3D representations, including NeRF [29, 67], 3D Gaussians [66, 73], tri-planes [59, 69] and feature grids [32]. Despite the excellent performance achieved, the single-object training data [5, 17, 18, 45] limits their generative abilities in scenarios involving multiple objects. Optimization-based 3D systems lift 2D diffusion priors [7, 8, 51, 56, 57] for 3D generation, and typically train 3D representations on a prompt-by-prompt basis. For example, DreamFusion [52] introduces Score Distillation Sampling (SDS) loss to transfer 2D diffusion models into the 3D domain. Magic3D [53] employs a coarse-to-fine scheme to improve both efficiency and effectiveness, while Fantasy3D [10] decouples the modelling of geometry and appearance. Despite the advancements of these methods addressing various challenges, they also usually struggle with compositional 3D generation. \\(\\mathrm{T}^3\\) bench's examination of ten prominent optimization-based models [10, 41, 47, 52, 58, 63, 68, 71, 77] revealed frequent issues in compositional generation. The implicit 2D diffusion guidance used in these methods [10, 52, 58, 71] often fails to maintain 3D consistency across different views, leading to significant discrepancies in compositional 3D scenes." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.539, + 0.487, + 0.902 + ], + "angle": 0, + "content": "Compositional Generation. Compositional generation involves creating content involving multiple objects with logical interactions. These interactions include, but are not limited to, concept relations [43], attribute association with colors [6, 22, 50], and spatial relationships between objects [9, 72]. There is considerable focus on various aspects of compositional 2D generation, such as learning from human feedback [20, 37, 75, 80], enhancing image captions [3, 7, 8, 16], designing effective networks [44], and etc. T2I-CompBench [30] further proposed a comprehensive benchmark and boosted the compositionality of text-to-image models. Compared to 2D compositional generation, its 3D counterpart is under-explored due to 3D geometry and appearance complexities [1, 12, 25, 40, 64, 74, 76]. For example, Set-the-Scene [15] and Scenewiz3d [79] propose to adopt object layouts to generate compositional scenes. However, they are mainly based on the implicit NeRF representations, so the features between objects cannot be well decoupled, resulting in a relatively blurred rendering effect. VP3D [11] takes a different strategy to employ visual features for compositional generation. However, its visual features are only used as implicit supervision, which does not adequately ensure 3D consistency or address the Janus problems [58]. Besides, LucidDreaming [70] requires" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.909, + 0.273 + ], + "angle": 0, + "content": "human-annotated layout priors for local optimization, which is labour-intensive and inaccurate; while GraphDreamer [23] optimizes all interactive objects from scratch, which is time-consuming and difficult to optimize. The most concurrent works, GALA3D [81], propose optimizing the spatial relationships of well-trained Gaussians for compositional generation. Though it achieves plausible spatial interactions, it is inefficient in generating complex mutual interactions among objects. More comparisons are in the appendix. We propose COMPGS to address the above challenges, emphasizing efficient optimization and the generation of complex interactions." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.28, + 0.649, + 0.297 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.305, + 0.909, + 0.412 + ], + "angle": 0, + "content": "In this section, we revisit 3D Gaussian Splitting [33] and diffusion priors for 3D generation. We then provide an overview of COMPGS, which includes initializing 3D Gaussians to incorporate compositionality priors and dynamic SDS optimization to generate high-fidelity, 3D consistent objects with complex interactions. The notations used in this section are also detailed in Appendix A.1 for clarity." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.42, + 0.655, + 0.435 + ], + "angle": 0, + "content": "3.1. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.441, + 0.909, + 0.568 + ], + "angle": 0, + "content": "3D Gaussian Splatting (GS). GS [33] has recently revolutionized novel-view synthesis of objects/scenes by its real-time rendering. Specifically, GS represents the 3D objects/scenes by \\(N\\) explicit anisotropic Gaussians with center positions \\(\\mu_{i}\\), covariances \\(\\Sigma_{i}\\), opacities \\(\\alpha_{i}\\) and colors \\(c_{i}\\), where \\(i\\in N\\). The color \\(\\mathcal{C}(\\mathbf{p})\\) of image pixel \\(\\mathbf{p}\\) can be calculated through point-based volume rendering [35, 36] by integrating the color and density of the 3D Gaussians intersected by a ray, as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.56, + 0.577, + 0.907, + 0.615 + ], + "angle": 0, + "content": "\\[\n\\mathcal {C} (\\mathbf {p}) = \\sum_ {i = 1} ^ {N} c _ {i} \\sigma_ {i} \\prod_ {j = 1} ^ {i - 1} (1 - \\sigma_ {j}), \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.577, + 0.617, + 0.907, + 0.649 + ], + "angle": 0, + "content": "\\[\n\\sigma_ {i} = \\alpha_ {i} \\exp \\left[ - \\frac {1}{2} \\left(\\mathbf {p} - \\hat {\\boldsymbol {\\mu}} _ {i}\\right) ^ {\\top} \\hat {\\Sigma} _ {i} ^ {- 1} \\left(\\mathbf {p} - \\hat {\\boldsymbol {\\mu}} _ {i}\\right) \\right], \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.659, + 0.909, + 0.84 + ], + "angle": 0, + "content": "where \\(\\hat{\\mu}_i\\) and \\(\\hat{\\Sigma}_i\\) denote the projected center positions and covariances of the 2D Gaussians, transformed from 3D space to the 2D camera's image plane. The image plane can be segmented into tiles during rendering for parallel processing. Unlike implicit representations, such as NeRF [2, 48, 49], GS offers two key advantages for compositional 3D generation: (1) 3D Gaussians facilitate localized rendering, allowing for the rendering of object \\(A\\) in one specific sub-space and object \\(B\\) in another, thus enabling global compositional generation on a divide-and-conquer basis; (2) 3D Gaussians enable direct parameter initialization, simplifying the integration of compositional priors at the initialization stage." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.841, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Diffusion Priors for 3D Generation. Diffusion-based generative models (DMs) [19, 60, 62] have been widely utilized to provide implicit priors for 3D object generation via score distillation sampling (SDS) [52]. Specifically, given a 3D" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "18511" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.117, + 0.091, + 0.724, + 0.105 + ], + "angle": 0, + "content": "\\(V = \\text{an owl perches on a branch near a pinecone}^{\\prime}\\) (\\(v_{1} = \\text{an owl}; v_{2} = \\text{a branch}; v_{3} = \\text{a pinecone}^{\\prime}\\))" + }, + { + "type": "image", + "bbox": [ + 0.114, + 0.108, + 0.884, + 0.315 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.316, + 0.907, + 0.382 + ], + "angle": 0, + "content": "Figure 2. Overall pipeline of COMPGS. Given a compositional prompt \\( V \\), we first use an LLM to decompose it into entity-level prompts \\( \\{v_l\\} \\), guiding the segmentation of each entity from the compositional image generated by T2I models. The segmented images initialize entity-level 3D Gaussians via image-to-3D models [29, 67]. COMPGS employs a dynamic optimization strategy, alternating between composition-level optimization of \\( \\theta \\) and entity-level optimization of \\( \\{\\theta_l\\} \\). For entity-level optimization, COMPGS dynamically maintains volume consistency to refine the details of each objects, particularly the small one." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.396, + 0.484, + 0.517 + ], + "angle": 0, + "content": "model whose parameters are \\(\\theta\\), a differentiable rendering process \\(g\\), the rendered images could be obtained via \\(\\mathrm{x} = g(\\theta)\\). To ensure the rendered images resemble those generated by the DM \\(\\phi\\), SDS first formulates the sampled noise \\(\\hat{\\epsilon}_{\\phi}(\\mathrm{z}_t;v,t)\\) with the noisy image \\(\\mathrm{z}_t\\), text embedding \\(v\\), and noise level \\(t\\). By comparing the difference between the added Gaussian noise \\(\\epsilon\\) and the predicted noise \\(\\hat{\\epsilon}_{\\phi}\\), SDS constructs gradients that could be back-propagated to update \\(\\theta\\) via:" + }, + { + "type": "equation", + "bbox": [ + 0.104, + 0.54, + 0.484, + 0.581 + ], + "angle": 0, + "content": "\\[\n\\nabla_ {\\theta} \\mathcal {L} _ {\\mathrm {S D S}} (\\phi , \\mathrm {x} = g (\\theta)) \\triangleq \\mathbb {E} _ {t, \\epsilon} \\left[ w (t) \\left(\\hat {\\epsilon} _ {\\phi} \\left(\\mathbf {z} _ {t}, v, t\\right) - \\epsilon\\right) \\frac {\\partial \\mathrm {x}}{\\partial \\theta} \\right], \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.582, + 0.483, + 0.613 + ], + "angle": 0, + "content": "where \\( w(t) \\) is a weighting function, and the Classifier-free guidance [27] typically amplifies the text conditioning \\( v \\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.633, + 0.196, + 0.647 + ], + "angle": 0, + "content": "3.2.COMPGS" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.659, + 0.486, + 0.902 + ], + "angle": 0, + "content": "We propose COMPGS for efficiently transferring the 2D compositionality to facilitate compositional 3D generation with 3D Gaussians. The overall framework of COMPGS is shown in Fig. 2. We first generate a well-composed image from the given complex prompt. After extracting the subobject (also named entity) information within the prompt by LLM, we utilize the entity prompt to segment the composed image into different parts. Each part will be adopted to initialize a specific space of Gaussians. During the training stage, we propose a dynamic SDS optimization strategy. This strategy first automatically decomposes the training to optimize either entity-level Gaussians or composition-level Gaussians. Then, it employs a volume-adaptive strategy to dynamically optimize varying-sized entities within a consistent 3D space. We detail the initialization process and dynamic SDS optimization in the following sections." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.395, + 0.866, + 0.411 + ], + "angle": 0, + "content": "3.2.1. 3D Gaussians Init with 2D Compositionality" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.416, + 0.907, + 0.566 + ], + "angle": 0, + "content": "Unlike the implicit NeRF [48] representations, explicit 3D Gaussians can be easily initialized with 3D shapes and colors, which facilitates introducing rough 3D priors to ensure 3D consistency [77]. Although existing text-to-3D or image-to-3D models can generate a single 3D object, they struggle with compositional 3D generation as mentioned in Sec. 2, which hinders the creation of compositional 3D priors needed for initialization. Considering this, we propose initializing the compositional Gaussians on an entity-by-entity basis." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.569, + 0.909, + 0.903 + ], + "angle": 0, + "content": "As shown in Fig. 2 (left), given a complex prompt \\( V \\), (e.g., 'an owl perches on a branch near a pinecone'), we first adopt a 2D diffusion model [3] to generate a composed image \\( I \\) that faithfully captures the compositionality of multiple objects. We extract each entity information in prompt \\( V \\) by prompting LLM to obtain \\( L \\) different entity-level prompts (i.e., \\( L = 3 \\) for \\( v_{1} \\) 'an owl', \\( v_{2} \\) 'a branch', and \\( v_{3} \\) 'a pinecone'). These prompts are adopted in a text-guided segmentation model [34] to decompose image \\( I \\) into various parts \\( \\{I_l\\} \\). Until now, each segmented image has only one entity, facilitating using existing image-based 3D generation models [67] to predict a rough triangle mesh \\( m_{l}(l \\in L) \\) of the corresponding 3D entity \\( l \\). To initialize 3D Gaussians \\( \\theta_{l} \\), we index \\( N \\) points from each mesh; the center positions \\( \\mu_i^l \\in \\mathbb{R}^3 (i \\in N, l \\in L) \\) are the centers of each vertex of \\( m_{l} \\), and the texture colors \\( c_i^l \\in \\mathbb{R}^3 (i \\in N, l \\in L) \\) are queried from each vertex of \\( m_{l} \\). During the image-to-3D process, we did not perform any cropping or padding on the image \\( I_{l} \\), ensuring that the spatial positions of each mesh \\( m_{l} \\) correspond to their 2D spatial positions. Additionally, we have marked the 3D layouts of each entity asset with 3D bounding boxes bbox \\( l \\) for subsequent optimization. The" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18512" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.123 + ], + "angle": 0, + "content": "bounding box coordinates are determined by the outer-most center positions of the entity Gaussian." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.132, + 0.332, + 0.148 + ], + "angle": 0, + "content": "3.2.2. Dynamic SDS Optimization" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.153, + 0.486, + 0.273 + ], + "angle": 0, + "content": "As mentioned in Sec. 2, existing SDS methods [10, 52, 58, 71] struggle in optimizing compositional 3D scenes due to the challenges of using implicit diffusion priors to maintain multi-objects consistency and interactions. To address this, we introduce dynamic SDS optimization to enable existing SDS losses to remain effective in compositional generation. Dynamic SDS optimization is performed through the following two procedures." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.276, + 0.485, + 0.442 + ], + "angle": 0, + "content": "Automatically Decomposing Optimization to Different Levels. We adopt the Decomposed Optimization (DO) strategy that divides the entire training into \\( L + 1 \\) stages, including \\( L \\) stages for entity-level optimization to ensure 3D consistency of each \\( L \\) entities, and one stage for composition-level optimization to refine inter-entity interactions. In each training iteration, we randomly render either (1) an entity Gaussian \\( \\theta_{l}(l \\in L) \\) to obtain the entity image \\( \\mathrm{x}_l = g(\\theta_l) \\) (\\( \\theta_{l} \\in \\mathrm{bbox}_{l} \\)), or (2) the composed Gaussians \\( \\theta \\) to obtain the composed image \\( \\mathrm{x} = g(\\theta) \\). Here \\( g \\) denotes the Gaussian Splitting rendering [33] described in Eq. 1." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.444, + 0.486, + 0.655 + ], + "angle": 0, + "content": "To update Gaussian parameters, we apply Eq. 3 for gradients backpropagation. Specifically, for entity-level optimization, we substitute the text embedding \\( v \\) by \\( v_{l} \\) and parameters \\( \\theta \\) by \\( \\theta_{l} \\) in Eq. 3 to optimize entity Gaussian. We adopt MV-Dream [58] as \\( \\phi \\), which provides multi-view diffusion priors that can better maintain 3D consistency. To ensure that other entity's parameters remain unchanged, we mask the gradients that are not within the corresponding 3D bounding boxes \\( \\mathrm{bbox}_l \\) when updating \\( \\theta_{l} \\). For composition-level optimization, we integrate both 2D [56] and 3D diffusion priors [58] to jointly optimize the overall Gaussian parameters \\( \\theta \\). The rationale is that 2D priors promote geometry exploration [53], which enhances the generation of plausible interactions between different entities." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.657, + 0.486, + 0.839 + ], + "angle": 0, + "content": "Volume-adaptively Optimizing Entity-level Gaussians. Compared to the single object generation, compositional 3D generation using SDS presents additional challenges. Specifically, when objects vary in size, optimizing directly in the original 3D space often results in suboptimal generation for smaller objects, as shown the 'pinecone' in Fig. 2. To mitigate this issue, we propose a Volume-adaptive Optimization (VAO) strategy that enhances optimization across objects of varying sizes by dynamically standardizing each entity's space to a standardized volumetric scale. This approach is especially beneficial for improving the generation of fine-grained texture details in smaller entities." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.486, + 0.902 + ], + "angle": 0, + "content": "To scale the 3D bounding box of an entity Gaussian \\(\\theta_{l}\\) with a standardized volumetric space bbox \\(^{std}\\), we first determine the necessary transformations for the centre positions \\(\\mu\\). Specifically, we calculate the shift parameters \\(\\beta\\), and" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.092, + 0.722, + 0.107 + ], + "angle": 0, + "content": "scale parameters \\(\\lambda\\), as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.575, + 0.114, + 0.907, + 0.129 + ], + "angle": 0, + "content": "\\[\n\\beta = \\operatorname {M e a n} (\\mathrm {b b o x} _ {l}); \\quad \\lambda = \\mathrm {b b o x} _ {s t d} / \\mathrm {b b o x} _ {l}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.136, + 0.907, + 0.304 + ], + "angle": 0, + "content": "where Mean computes the center coordinates of the given bounding box, and \\(\\mathrm{bbox}_{std}\\) is the target standardized space. Then, we zoom-in the entity Gaussian on its center positions by \\(\\hat{\\mu} = \\lambda (\\mu -\\beta)\\), where \\(\\hat{\\mu}\\) denotes the transformed center positions of entity Gaussian. We substitute all \\(\\mu\\) to \\(\\hat{\\mu}\\) in Eq. 1 for SDS optimization. After each training iteration, we perform zooming-back via \\(\\mu = \\hat{\\mu} /\\lambda +\\beta\\). The proposed transformation ensures dynamically conducting SDS optimization on each entity Gaussian at a standardized and consistent scale. Experiments in Sec. 4.3 demonstrate its effectiveness in generating high-quality 3D assets." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.307, + 0.875, + 0.323 + ], + "angle": 0, + "content": "3.3. Progressive Adding Objects with COMPGS" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.329, + 0.909, + 0.57 + ], + "angle": 0, + "content": "The dynamic SDS optimization enables COMPGS to explicitly control the generation of different parts in 3D scene. This capability can be utilized for progressive increasing 3D objects. Specifically, given well-trained 3D Gaussians \\(\\theta\\), we begin by rendering an image from the front view, denoted as \\(\\mathbf{x} = g(\\theta)\\). Unlike previous works that directly add 3D objects [13], we adopt MagicBrush [78] to modify the 2D rendered image \\(\\mathbf{x}\\). The edited 2D image can then be utilized to initialize new Gaussians \\(\\hat{\\theta}\\) in the original 3D space. We further train the added 3D Gaussians to finally introduce new objects into the compositional scenes. As shown in Fig. 1, given well-trained 3D Gaussians (e.g., 'an owl'), we progressively modify the 2D images to add new objects, such as the 'branch' and 'pinecone.' We then train the corresponding 3D Gaussians via dynamic SDS optimization to incorporate these new elements into the 3D scenes." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.583, + 0.649, + 0.6 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.607, + 0.722, + 0.624 + ], + "angle": 0, + "content": "4.1. Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.629, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Implementation Details. We implemented COMPGS using ThreeStudio [24]. We use DALL-E [3] to generate the well-established 2D compositionality, LangSAM [46] to segment the entity image, and TripoSR [67] to generate a preliminary 3D prior (i.e., mesh) from each entity image, respectively. During the training, we employed MVDream [58] to optimize entity-level Gaussians, and both MVDream [58] and stabilityai/stablediffusion-2-1-base [57] to optimize composition-level Gaussians. The guidance scale for diffusion models was set as 50. Timestamps were uniformly selected from 0.02 to 0.55 for the first 1,000 iterations and then adjusted to a range from 0.02 to 0.15 for subsequent iterations. We initialized the Gaussian points at \\( N = 10k \\), and progressively increased the density to a maximum of \\( 1000k \\) points. We optimized the overall compositional scenes through 10k iterations to achieve optimal results. However, it is worth noting that we empirically found that training with 5k iterations already produces" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18513" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.182 + ], + "angle": 0, + "content": "plausible 3D assets. The learning rates are detailed in Appendix A.3. We set the camera parameters, including radius, azimuth, elevation and FoV (field of view) to be the same as [58]. Experiments were conducted on NVIDIA A100 GPUs(40G). The prompt used for decomposing are provided in Appendix A.11. Our code will be released." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.184, + 0.483, + 0.364 + ], + "angle": 0, + "content": "Evaluation Metrics. In addition to user studies and qualitative comparisons, we utilize \\(\\mathrm{T}^3\\) Bench [26] to render 300 prompts in order to evaluate compositional 3D generation on two criteria: (1) the visual quality of the 3D objects and (2) the alignment between the 3D objects and the input prompt. For quality evaluation, we captured multi-focal and multiview images and sent them to text-image scoring models, CLIP [54], to obtain an average quality score of the generated 3D scene. Regarding the textual alignment, we first followed [45] to perform 3D captioning via BLIP [38] and GPT4, and then computed the recall of the original prompt within the generated caption via ROUGE-L [42]." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.378, + 0.439, + 0.394 + ], + "angle": 0, + "content": "4.2. Performance Comparisons and Analysis." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.401, + 0.483, + 0.884 + ], + "angle": 0, + "content": "Qualitative Evaluation. In Fig. 3, we present qualitative comparisons of compositional 3D generation. Our model, COMPGS, is compared with several open-source methods, including DreamFusion [52], Magic3D [53], Latent-NeRF [47], Fantasia3D [10], SJC [68], and ProlificDreamer [71], as well as methods specifically designed to handle intricate text prompts, VP3D [11] in Fig. 3. The results demonstrate that COMPGS generates compositional 3D objects with superior quality, greater consistency, and more plausible interactions. DreamFusion, for example, fails to generate reasonable multi-objects from compositional prompts, revealing limitations in SDS loss for optimizing multiple objects simultaneously. Magic3D and Latent-NeRF, though guided by a 3D mesh prior, struggle with generating complex and compositional 3D geometry. Despite employing advanced SDS loss, Fantasia3D, SJC, and ProlificDreamer only improve the appearance of the content but do not fundamentally address multi-object generation. Notably, VP3D employs features from compositional images and human feedback to guide compositional 3D generation. However, it often produces unexpected artifacts (e.g., redundant dots in the top two rows of Fig. 3) or less plausible interactions (e.g., strange spatial relationships in the fourth row). This is likely due to the insufficient utilization of 2D compositionality, as VP3D adopts the implicit image features to guide optimization. In contrast, COMPGS generates high-quality and composed content that strictly aligns with the given complex prompts, demonstrating its effectiveness over others. We provide more visual comparisons with both the close-source and open-source methods (including GALA3D [81] in Fig. 9, GraphDreamer [23], DreamGaussian [65] in Fig. 10, etc.) in Appendix A.5 as well as the attached video." + }, + { + "type": "text", + "bbox": [ + 0.111, + 0.886, + 0.483, + 0.901 + ], + "angle": 0, + "content": "We also demonstrate COMPGS's capabilities in generat" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.09, + 0.907, + 0.131 + ], + "angle": 0, + "content": "Table 1. Quantitative comparisons on \\(\\mathrm{T}^3\\) Bench [26] and user studies show COMPGS outperforms feed-forward, optimization-based, and compositional generation models." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.136, + 0.905, + 0.294 + ], + "angle": 0, + "content": "
MethodT3Bench (Multiple Objects)User Study Ranking Score↑
Quality↑Alignment↑Average↑
OpenLRM [29]15.225.520.4-
TripoSR [67]16.728.622.7-
DreamFusion [52]17.314.816.14.54
SJC [68]17.75.811.73.08
LatentNeRF [47]21.719.520.64.14
Fantasia3D [10]22.714.318.52.43
ProlificDreamer [71]45.725.835.83.56
Magic3D [41]26.624.825.74.30
Set-the-Scene [15]20.829.925.4-
VP3D [11]49.131.540.36.71
COMPGS (ours)54.2 (+5.1)37.9(+6.4)46.1(+5.8)7.23
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.306, + 0.907, + 0.442 + ], + "angle": 0, + "content": "ing more diverse contents in Fig. 4. COMPGS accurately generates both simple spatial relationships (shown in the first row), and more complex interactions, where objects should react to each other (shown in the second row). Besides, COMPGS is not limited to generating just two objects, it can generate multiple objects by expanding the number of entity-level Gaussians. For example, in the third row, COMPGS generates scenes with three and four entities, each with high-quality 3D shapes and textures." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.445, + 0.909, + 0.626 + ], + "angle": 0, + "content": "We conduct a user study for further evaluation. We randomly selected 15 prompts in \\(\\mathrm{T}^3\\) Bench, and collected the 3D assets generated by different models. These collected 3D assets were then distributed to individuals for ranking the models based on (1) 3D visual quality and (3) text-3D alignment. We average the ranking of different models as their scores. Results in Tab. 1 show that representative optimization-based models received relatively low average scores, highlighting their limitations in compositional generation. VP3D [11] outperforms its predecessors due to the incorporation of image features but still ranks lower than COMPGS, which further validates our effectiveness." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.629, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Quantitative Evaluation. In Tab. 1, we benchmark the representative models on \\(\\mathrm{T}^3\\) Bench, focusing on the multi-objects track. Results indicate COMPGS achieves superior performance in both quality and textual alignment. Compared with feed-forward methods in the first block, COMPGS significantly improves the generation quality by a large margin. Though COMPGS is initialized from the 3D priors obtained from [67], COMPGS still enhances text-3D alignment, likely due to the dynamic optimization on composition-level Gaussians. Furthermore, compared with the optimization-based methods, COMPGS shows significant advantages in both two metrics. Even when compared to methods specifically designed for compositional generation, e.g., VP3D [11] and Set-the-Scene [15], COMPGS demonstrates clear improvements, indicating the effectiveness of our designs. Note that COMPGS is not limited to compositional generation; the results in the track of single-object generation are included in Appendix A.5." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.956 + ], + "angle": 0, + "content": "18514" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.355, + 0.088, + 0.644, + 0.102 + ], + "angle": 0, + "content": "'A florist is making a bouquet with fresh flowers'" + }, + { + "type": "text", + "bbox": [ + 0.379, + 0.19, + 0.619, + 0.203 + ], + "angle": 0, + "content": "'An artist is painting on a blank canvas'" + }, + { + "type": "text", + "bbox": [ + 0.322, + 0.292, + 0.676, + 0.306 + ], + "angle": 0, + "content": "Hot popcorn jump out from the red striped popcorn maker" + }, + { + "type": "text", + "bbox": [ + 0.333, + 0.394, + 0.665, + 0.407 + ], + "angle": 0, + "content": "'A half-eaten sandwich sits next to a lukewarm thermos'" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.493, + 0.159, + 0.504 + ], + "angle": 0, + "content": "DreamFusion" + }, + { + "type": "text", + "bbox": [ + 0.213, + 0.493, + 0.26, + 0.504 + ], + "angle": 0, + "content": "Magic3D" + }, + { + "type": "text", + "bbox": [ + 0.298, + 0.493, + 0.358, + 0.504 + ], + "angle": 0, + "content": "LatentNeRF" + }, + { + "type": "text", + "bbox": [ + 0.403, + 0.493, + 0.458, + 0.504 + ], + "angle": 0, + "content": "Fantasia3D" + }, + { + "type": "text", + "bbox": [ + 0.527, + 0.493, + 0.55, + 0.503 + ], + "angle": 0, + "content": "SJC" + }, + { + "type": "text", + "bbox": [ + 0.608, + 0.493, + 0.685, + 0.503 + ], + "angle": 0, + "content": "ProlificDreamer" + }, + { + "type": "text", + "bbox": [ + 0.744, + 0.493, + 0.776, + 0.503 + ], + "angle": 0, + "content": "VP3D" + }, + { + "type": "text", + "bbox": [ + 0.824, + 0.493, + 0.902, + 0.504 + ], + "angle": 0, + "content": "COMPGS (ours)" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.511, + 0.907, + 0.553 + ], + "angle": 0, + "content": "Figure 3. Qualitative comparisons between COMPGS and other text-to-3D models on \\(\\mathrm{T}^3\\)Bench (multiple objects track). Compared to others, COMPGS is better at generating highly-composed, high-quality 3D contents that strictly align with the given texts. Watch the animations by clicking them (Not all PDF readers support playing animations. Best viewed in Acrobat/Foxit Reader)." + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.554, + 0.485, + 0.799 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.556, + 0.884, + 0.799 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.278, + 0.803, + 0.719, + 0.817 + ], + "angle": 0, + "content": "Figure 4. More generated samples by COMPGS. Four views are shown." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.822, + 0.486, + 0.899 + ], + "angle": 0, + "content": "Runtime Comparisons. Though COMPGS is trained with 10k steps, we observed that training the model for 5k iterations already produces high-quality content with minimal loss of texture details. We show runtime comparisons with other models in Appendix A.4. In fact, compositional" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.822, + 0.909, + 0.899 + ], + "angle": 0, + "content": "3D generation (left) requires a longer training time than single-object generation (right) due to its complexity. Compositional 3D involves optimizing individual objects and ensuring the consistency of compositional scenes, making them more intricate. The runtime is generally relative to" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18515" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.113, + 0.06, + 0.473, + 0.236 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.102, + 0.237, + 0.472, + 0.265 + ], + "angle": 0, + "content": "Figure 5. 3D Editing examples of COMPGS. More examples could be found in Appendix A.10." + }, + { + "type": "image", + "bbox": [ + 0.109, + 0.269, + 0.466, + 0.437 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.1, + 0.443, + 0.469, + 0.457 + ], + "angle": 0, + "content": "Figure 6. Visual results of the ablation studies on key designs." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.471, + 0.486, + 0.713 + ], + "angle": 0, + "content": "the number of objects involved. Compared to open-source compositional 3D methods such as Set-the-Scene [15], Progressive3D [14], and GraphDraemer [23], our proposed COMPGS is more efficient in training. For instance, given the prompt \"a parrot talks beside a perch and two bowls,\" Progressive3D takes approximately 250 minutes for 3D generation, while Set-the-Scene requires around 110 minutes. Since many compositional scene generation methods are not open-sourced, we also present the training steps listed in the papers for straightforward comparisons. In addition, CompGS demonstrates comparable and even superior efficiency over other methods in text-to-single object generation. For example, methods such as Magic3D [41] and Fantastic3D [10] require over five hours to optimize a single object. In contrast, CompGS can achieve the same task in approximately 30 minutes. Details are in Appendix A.4." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.725, + 0.485, + 0.86 + ], + "angle": 0, + "content": "Extended Applications - Progressive 3D Adding. COMPGS allows progressive 3D adding for compositional scenes. We present 3D adding examples in Fig. 5. These examples demonstrate that by transferring edited 2D compositionality and employing dynamic SDS optimization, COMPGS can progressively incorporate new 3D entities into the original 3D scenes. For example, COMPGS can generate the 'chair', 'panda', 'hat' and 'plant' step by step as shown in Fig. 5." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.872, + 0.483, + 0.903 + ], + "angle": 0, + "content": "Limitations. Detailed limitations and future explorations are provided in Appendix A.12 and Appendix A.13." + }, + { + "type": "table_caption", + "bbox": [ + 0.578, + 0.075, + 0.84, + 0.088 + ], + "angle": 0, + "content": "Table 2. Ablation Studies on \\( {\\mathrm{T}}^{3} \\) Bench [26]." + }, + { + "type": "table", + "bbox": [ + 0.553, + 0.089, + 0.865, + 0.173 + ], + "angle": 0, + "content": "
ComponentQualityAlignmentAverage
Full Setting53.838.045.9
- w/o GS Init.22.818.720.8
- w/o DO Strategy46.835.241.0
- w/o VAO Strategy50.836.443.6
" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.185, + 0.667, + 0.202 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.208, + 0.908, + 0.314 + ], + "angle": 0, + "content": "We conduct ablation studies to validate the effectiveness of three key designs in COMPGS: Gaussian parameters initialization, decomposed optimization (DO), and volume-adaptive optimization (VAO). We randomly choose 20 prompts from \\(\\mathrm{T}^3\\)Bench for the quantitative evaluation. Meanwhile, we use Fig. 6 to visualize the effect of each component." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.315, + 0.907, + 0.436 + ], + "angle": 0, + "content": "Initialization. Instead of initializing Gaussian parameters using 2D compositionality, we conduct random initialization within a predefined 3D bounding box for each entity Gaussian. A significant decrease in the quality and alignment metric is observed in Tab. 2. Besides, the visualization in Fig. 6 confirms that training from random initialization with predefined layouts may result in low-quality textures (e.g., owl's face) and missing entities (e.g., the branch)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.437, + 0.909, + 0.573 + ], + "angle": 0, + "content": "Decomposed Optimization. We analyze the decomposed training by discarding the entity-level optimization. Results in Tab. 2 and Fig. 6 both indicate that removing decomposed training leads to low-fidelity generation. For example, in Fig. 6, optimizing composition-level Gaussians results in a relatively blurred generation on the owl's fur and the pinecone, validating that optimization on each single entity is necessary to effectively leverage 2D diffusion priors to guide 3D generation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.574, + 0.909, + 0.726 + ], + "angle": 0, + "content": "Volume-adaptive Optimization. To evaluate the effectiveness of volume-adaptive optimization, we choose to optimize the entity Gaussians in their original 3D space. We observed that optimizing different entities within varying sizes of 3D space slightly decreased in quality and alignment, which is noticeable for smaller objects in the compositional scene. For instance, the 'pinecone' and 'branch' trained without volume-adaptive optimization exhibited fewer fine-grained details, shown as the green leaves on the branch and the shadow on the pinecone in Fig. 6." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.739, + 0.634, + 0.754 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.909, + 0.901 + ], + "angle": 0, + "content": "In this paper, we introduce COMPGS, a user-friendly, optimization-based framework, that achieves compositional text-to-3D generation utilizing Gaussian Splitting. Our three core designs, including initializing 3D Gaussians with 2D compositionality, decomposed optimization for either entity-level or composition-level Gaussians, and volume-adaptive optimization, enable COMPGS to generate complex interactions between multiple 3D entities. We hope this work encourages further research in compositional 3D generation." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18516" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.168 + ], + "angle": 0, + "content": "Acknowledgment The work was supported by Berkeley DeepDrive. The authors thank Google Cloud Platform for providing the computational resources. The work was done during Chongjian Ge's visit to University of California, Berkeley." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.185, + 0.188, + 0.201 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.21, + 0.484, + 0.265 + ], + "angle": 0, + "content": "[1] Haotian Bai, Yuanhuiyi Lyu, Lutao Jiang, Sijia Li, Haonan Lu, Xiaodong Lin, and Lin Wang. Componerf: Text-guided multi-object compositional nerf with editable 3d scene layout. arXiv preprint arXiv:2303.13843, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.268, + 0.484, + 0.337 + ], + "angle": 0, + "content": "[2] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In IEEE/CVF International Conference on Computer Vision, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.34, + 0.484, + 0.409 + ], + "angle": 0, + "content": "[3] James Betker, Gabriel Goh, Li Jing, Tim Brooks, Jianfeng Wang, Linjie Li, Long Ouyang, Juntang Zhuang, Joyce Lee, Yufei Guo, et al. Improving image generation with better captions. Computer Science. https://cdn.openaia.com/papers/dall-e-3.pdf, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.412, + 0.484, + 0.481 + ], + "angle": 0, + "content": "[4] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.484, + 0.484, + 0.552 + ], + "angle": 0, + "content": "[5] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.556, + 0.484, + 0.611 + ], + "angle": 0, + "content": "[6] Hila Chefer, Yuval Alaluf, Yael Vinker, Lior Wolf, and Daniel Cohen-Or. Attend-and-excite: Attention-based semantic guidance for text-to-image diffusion models. ACM Transactions on Graphics (TOG), 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.614, + 0.484, + 0.669 + ], + "angle": 0, + "content": "[7] Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, et al. Pixart-alpha: Fast training of diffusion transformer for photorealistic text-to-image synthesis. 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.672, + 0.484, + 0.741 + ], + "angle": 0, + "content": "[8] Junsong Chen, Chongjian Ge, Enze Xie, Yue Wu, Lewei Yao, Xiaozhe Ren, Zhongdao Wang, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart- \\(\\backslash\\) sigma: Weak-to-strong training of diffusion transformer for 4k text-to-image generation. arXiv preprint arXiv:2403.04692, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.744, + 0.484, + 0.799 + ], + "angle": 0, + "content": "[9] Minghao Chen, Iro Laina, and Andrea Vedaldi. Training-free layout control with cross-attention guidance. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.802, + 0.484, + 0.857 + ], + "angle": 0, + "content": "[10] Rui Chen, Yongwei Chen, Ningxin Jiao, and Kui Jia. Fantasia3d: Disentangling geometry and appearance for high-quality text-to-3d content creation. In IEEE/CVF International Conference on Computer Vision, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.859, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[11] Yang Chen, Yingwei Pan, Haibo Yang, Ting Yao, and Tao Mei. Vp3d: Unleashing 2d visual prompt for text-to-3d generation. arXiv preprint arXiv:2403.17001, 2024." + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.21, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.147 + ], + "angle": 0, + "content": "[12] Yongwei Chen, Tengfei Wang, Tong Wu, Xingang Pan, Kui Jia, and Ziwei Liu. Comboverse: Compositional 3d assets creation using spatially-aware diffusion guidance. arXiv preprint arXiv:2403.12409, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.15, + 0.907, + 0.205 + ], + "angle": 0, + "content": "[13] Kai Cheng, Xiaoxiao Long, Kaizhi Yang, Yao Yao, Wei Yin, Yuexin Ma, Wenping Wang, and Xuejin Chen. Gaussianpro: 3d gaussian splatting with progressive propagation. arXiv preprint arXiv:2402.14650, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.208, + 0.907, + 0.262 + ], + "angle": 0, + "content": "[14] Xinhua Cheng, Tianyu Yang, Jianan Wang, Yu Li, Lei Zhang, Jian Zhang, and Li Yuan. Progressive3d: Progressively local editing for text-to-3d content creation with complex semantic prompts. arXiv preprint arXiv:2310.11784, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.264, + 0.907, + 0.319 + ], + "angle": 0, + "content": "[15] Dana Cohen-Bar, Elad Richardson, Gal Metzer, Raja Giryes, and Daniel Cohen-Or. Set-the-scene: Global-local training for generating controllable nerf scenes. In IEEE/CVF International Conference on Computer Vision, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.321, + 0.907, + 0.389 + ], + "angle": 0, + "content": "[16] Xiaoliang Dai, Ji Hou, Chih-Yao Ma, Sam Tsai, Jialiang Wang, Rui Wang, Peizhao Zhang, Simon Vandenhende, Xiaofang Wang, Abhimanyu Dubey, et al. Emu: Enhancing image generation models using photogenic needles in a haystack. arXiv preprint arXiv:2309.15807, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.392, + 0.907, + 0.461 + ], + "angle": 0, + "content": "[17] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objaverse: A universe of annotated 3d objects. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.463, + 0.907, + 0.531 + ], + "angle": 0, + "content": "[18] Matt Deitke, Ruoshi Liu, Matthew Wallingford, Huong Ngo, Oscar Michel, Aditya Kusupati, Alan Fan, Christian Laforte, Vikram Voleti, Samir Yitzhak Gadre, et al. Objverse-xl: A universe of \\(10\\mathrm{m} + 3\\mathrm{d}\\) objects. Advances in Neural Information Processing Systems, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.533, + 0.906, + 0.575 + ], + "angle": 0, + "content": "[19] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in Neural Information Processing Systems, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.577, + 0.907, + 0.645 + ], + "angle": 0, + "content": "[20] Hanze Dong, Wei Xiong, Deepanshu Goyal, Yihan Zhang, Winnie Chow, Rui Pan, Shizhe Diao, Jipeng Zhang, Kashun Shum, and Tong Zhang. Raft: Reward ranked finetuning for generative foundation model alignment. arXiv preprint arXiv:2304.06767, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.648, + 0.907, + 0.702 + ], + "angle": 0, + "content": "[21] Danny Driess, Zhiao Huang, Yunzhu Li, Russ Tedrake, and Marc Toussaint. Learning multi-object dynamics with compositional neural radiance fields. In Conference on robot learning, pages 1755-1768. PMLR, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.704, + 0.907, + 0.773 + ], + "angle": 0, + "content": "[22] Weixi Feng, Xuehai He, Tsu-Jui Fu, Varun Jampani, Arjun Akula, Pradyumna Narayana, Sugato Basu, Xin Eric Wang, and William Yang Wang. Training-free structured diffusion guidance for compositional text-to-image synthesis. arXiv preprint arXiv:2212.05032, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.776, + 0.907, + 0.844 + ], + "angle": 0, + "content": "[23] Gege Gao, Weiyang Liu, Anpei Chen, Andreas Geiger, and Bernhard Scholkopf. Graphdreamer: Compositional 3d scene synthesis from scene graphs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21295-21304, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.907, + 0.901 + ], + "angle": 0, + "content": "[24] Yuan-Chen Guo, Ying-Tian Liu, Ruizhi Shao, Christian Laforte, Vikram Voleti, Guan Luo, Chia-Hao Chen, Zixin Zou, Chen Wang, Yan-Pei Cao, and Song-Hai Zhang. threestudio: A unified framework for 3d content generation." + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "18517" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.093, + 0.484, + 0.12 + ], + "angle": 0, + "content": "https://github.com/threestudio-project/threestudio, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.484, + 0.177 + ], + "angle": 0, + "content": "[25] Haonan Han, Rui Yang, Huan Liao, Jiankai Xing, Zunnan Xu, Xiaoming Yu, Junwei Zha, Xiu Li, and Wanhua Li. Reparo: Compositional 3d assets generation with differentiable 3d layout alignment. arXiv preprint arXiv:2405.18525, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.179, + 0.484, + 0.233 + ], + "angle": 0, + "content": "[26] Yuze He, Yushi Bai, Matthieu Lin, Wang Zhao, Yubin Hu, Jenny Sheng, Ran Yi, Juanzi Li, and Yong-Jin Liu. T3bench: Benchmarking current progress in text-to-3d generation. arXiv preprint arXiv:2310.02977, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.235, + 0.484, + 0.275 + ], + "angle": 0, + "content": "[27] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. Advances in Neural Information Processing Systems, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.278, + 0.484, + 0.319 + ], + "angle": 0, + "content": "[28] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.321, + 0.484, + 0.375 + ], + "angle": 0, + "content": "[29] Yicong Hong, Kai Zhang, Jiuxiang Gu, Sai Bi, Yang Zhou, Difan Liu, Feng Liu, Kalyan Sunkavalli, Trung Bui, and Hao Tan. Lrm: Large reconstruction model for single image to 3d. arXiv preprint arXiv:2311.04400, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.378, + 0.484, + 0.433 + ], + "angle": 0, + "content": "[30] Kaiyi Huang, Kaiyue Sun, Enze Xie, Zhenguo Li, and Xihui Liu. T2i-compbench: A comprehensive benchmark for open-world compositional text-to-image generation. Advances in Neural Information Processing Systems, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.435, + 0.484, + 0.462 + ], + "angle": 0, + "content": "[31] Heewoo Jun and Alex Nichol. Shap-e: Generating conditional 3d implicit functions. arXiv preprint arXiv:2305.02463, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.464, + 0.484, + 0.517 + ], + "angle": 0, + "content": "[32] Animesh Karnewar, Andrea Vedaldi, David Novotny, and Niloy J Mitra. Holodiffusion: Training a 3d diffusion model using 2d images. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.519, + 0.484, + 0.56 + ], + "angle": 0, + "content": "[33] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splattering for real-time radiance field rendering. ACM Transactions on Graphics, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.563, + 0.484, + 0.63 + ], + "angle": 0, + "content": "[34] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In IEEE/CVF International Conference on Computer Vision, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.633, + 0.484, + 0.674 + ], + "angle": 0, + "content": "[35] Georgios Kopanas, Julien Philip, Thomas Leimkuhler, and George Drettakis. Point-based neural rendering with per-view optimization. In Computer Graphics Forum, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.676, + 0.484, + 0.731 + ], + "angle": 0, + "content": "[36] Georgios Kopanas, Thomas Leimkuhler, Gilles Rainer, Clément Jambon, and George Drettakis. Neural point catauastics for novel-view synthesis of reflections. ACM Transactions on Graphics (TOG), 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.733, + 0.484, + 0.8 + ], + "angle": 0, + "content": "[37] Kimin Lee, Hao Liu, Moonkyung Ryu, Olivia Watkins, Yuqing Du, Craig Boutilier, Pieter Abbeel, Mohammad Ghavamzadeh, and Shixiang Shane Gu. Aligning text-to-image models using human feedback. arXiv preprint arXiv:2302.12192, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.803, + 0.484, + 0.858 + ], + "angle": 0, + "content": "[38] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International Conference on Machine Learning, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.86, + 0.484, + 0.9 + ], + "angle": 0, + "content": "[39] Jiahao Li, Hao Tan, Kai Zhang, Zexiang Xu, Fujun Luan, Yinghao Xu, Yicong Hong, Kalyan Sunkavalli, Greg Shakhnarovich, and Sai Bi. Instant3d: Fast text-to-3d with" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "sparse-view generation and large reconstruction model. arXiv preprint arXiv:2311.06214, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.907, + 0.19 + ], + "angle": 0, + "content": "[40] Runjia Li, Junlin Han, Luke Melas-Kyriazi, Chunyi Sun, Zhaochong An, Zhongrui Gui, Shuyang Sun, Philip Torr, and Tomas Jakab. Dreambeast: Distilling 3d fantastical animals with part-aware knowledge transfer. arXiv preprint arXiv:2409.08271, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.193, + 0.907, + 0.262 + ], + "angle": 0, + "content": "[41] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.264, + 0.905, + 0.29 + ], + "angle": 0, + "content": "[42] Chin-Yew Lin. Rouge: A package for automatic evaluation of summaries. In Text summarization branches out, 2004." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.292, + 0.905, + 0.347 + ], + "angle": 0, + "content": "[43] Nan Liu, Shuang Li, Yilun Du, Antonio Torralba, and Joshua B Tenenbaum. Compositional visual generation with composable diffusion models. In European Conference on Computer Vision, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.349, + 0.905, + 0.404 + ], + "angle": 0, + "content": "[44] Nan Liu, Shuang Li, Yilun Du, Antonio Torralba, and Joshua B Tenenbaum. Compositional visual generation with composable diffusion models. In European Conference on Computer Vision, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.406, + 0.905, + 0.447 + ], + "angle": 0, + "content": "[45] Tiange Luo, Chris Rockwell, Honglak Lee, and Justin Johnson. Scalable 3d captioning with pretrained models. Advances in Neural Information Processing Systems, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.449, + 0.907, + 0.503 + ], + "angle": 0, + "content": "[46] Luca Medeiros. Language segment-anything: Sam with text prompt. https://github.com/luca-medeiros/ lang-segment-anything, 2024. Accessed: 2024-05-21." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.505, + 0.905, + 0.56 + ], + "angle": 0, + "content": "[47] Gal Metzer, Elad Richardson, Or Patashnik, Raja Giryes, and Daniel Cohen-Or. Latent-nerf for shape-guided generation of 3d shapes and textures. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.563, + 0.907, + 0.617 + ], + "angle": 0, + "content": "[48] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.619, + 0.907, + 0.672 + ], + "angle": 0, + "content": "[49] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM transactions on graphics (TOG), 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.675, + 0.907, + 0.731 + ], + "angle": 0, + "content": "[50] Dong Huk Park, Samaneh Azadi, Xihui Liu, Trevor Darrell, and Anna Rohrbach. Benchmark for compositional text-to-image synthesis. In Advances in Neural Information Processing Systems, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.733, + 0.907, + 0.787 + ], + "angle": 0, + "content": "[51] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.789, + 0.907, + 0.83 + ], + "angle": 0, + "content": "[52] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.832, + 0.907, + 0.9 + ], + "angle": 0, + "content": "[53] Guocheng Qian, Jinjie Mai, Abdullah Hamdi, Jian Ren, Aliaksandr Siarohin, Bing Li, Hsin-Ying Lee, Ivan Skorokhodov, Peter Wonka, Sergey Tulyakov, et al. Magic123: One image to high-quality 3d object generation using both 2d and 3d diffusion priors. arXiv preprint arXiv:2306.17843, 2023." + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18518" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.162 + ], + "angle": 0, + "content": "[54] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.165, + 0.485, + 0.22 + ], + "angle": 0, + "content": "[55] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.223, + 0.484, + 0.277 + ], + "angle": 0, + "content": "[56] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.281, + 0.484, + 0.336 + ], + "angle": 0, + "content": "[57] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.339, + 0.484, + 0.38 + ], + "angle": 0, + "content": "[58] Yichun Shi, Peng Wang, Jianglong Ye, Mai Long, Kejie Li, and Xiao Yang. Mvdream: Multi-view diffusion for 3d generation. arXiv preprint arXiv:2308.16512, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.383, + 0.484, + 0.437 + ], + "angle": 0, + "content": "[59] J Ryan Shue, Eric Ryan Chan, Ryan Po, Zachary Ankner, Jiajun Wu, and Gordon Wetzstein. 3d neural field generation using triplane diffusion. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.441, + 0.484, + 0.495 + ], + "angle": 0, + "content": "[60] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International Conference on Machine Learning, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.499, + 0.484, + 0.538 + ], + "angle": 0, + "content": "[61] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.542, + 0.484, + 0.598 + ], + "angle": 0, + "content": "[62] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.601, + 0.484, + 0.655 + ], + "angle": 0, + "content": "[63] Jingxiang Sun, Bo Zhang, Ruizhi Shao, Lizhen Wang, Wen Liu, Zhenda Xie, and Yebin Liu. Dreamcraft3d: Hierarchical 3d generation with bootstrapped diffusion prior. arXiv preprint arXiv:2310.16818, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.658, + 0.484, + 0.712 + ], + "angle": 0, + "content": "[64] Shaorong Sun, Shuchao Pang, Yazhou Yao, and Xiaoshui Huang. Comogen: A controllable text-to-3d multi-object generation framework. arXiv preprint arXiv:2409.00590, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.716, + 0.484, + 0.77 + ], + "angle": 0, + "content": "[65] Jiaxiang Tang, Jiawei Ren, Hang Zhou, Ziwei Liu, and Gang Zeng. Dreamgaussian: Generative gaussian splatting for efficient 3d content creation. arXiv preprint arXiv:2309.16653, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.774, + 0.484, + 0.828 + ], + "angle": 0, + "content": "[66] Jiaxiang Tang, Zhaoxi Chen, Xiaokang Chen, Tengfei Wang, Gang Zeng, and Ziwei Liu. Lgm: Large multi-view gaussian model for high-resolution 3d content creation. arXiv preprint arXiv:2402.05054, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.484, + 0.9 + ], + "angle": 0, + "content": "[67] Dmitry Tochilkin, David Pankratz, Zexiang Liu, Zixuan Huang, Adam Letts, Yangguang Li, Ding Liang, Christian Laforte, Varun Jampani, and Yan-Pei Cao. Triposr: Fast 3d object reconstruction from a single image. arXiv preprint arXiv:2403.02151, 2024." + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.908, + 0.16 + ], + "angle": 0, + "content": "[68] Haochen Wang, Xiaodan Du, Jiahao Li, Raymond A Yeh, and Greg Shakhnarovich. Score jacobian chaining: Lifting pretrained 2d diffusion models for 3d generation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.164, + 0.907, + 0.233 + ], + "angle": 0, + "content": "[69] Tengfei Wang, Bo Zhang, Ting Zhang, Shuyang Gu, Jianmin Bao, Tadas Baltrusaitis, Jingjing Shen, Dong Chen, Fang Wen, Qifeng Chen, et al. Rodin: A generative model for sculpting 3d digital avatars using diffusion. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.235, + 0.907, + 0.275 + ], + "angle": 0, + "content": "[70] Zhaoning Wang, Ming Li, and Chen Chen. Luciddreaming: Controllable object-centric 3d generation. arXiv preprint arXiv:2312.00588, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.278, + 0.907, + 0.345 + ], + "angle": 0, + "content": "[71] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. Advances in Neural Information Processing Systems, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.349, + 0.907, + 0.417 + ], + "angle": 0, + "content": "[72] Qiucheng Wu, Yujuan Liu, Handong Zhao, Trung Bui, Zhe Lin, Yang Zhang, and Shiyu Chang. Harnessing the spatial-temporal attention of diffusion models for high-fidelity text-to-image synthesis. In IEEE/CVF International Conference on Computer Vision, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.42, + 0.908, + 0.487 + ], + "angle": 0, + "content": "[73] Yinghao Xu, Zifan Shi, Wang Yifan, Hansheng Chen, Ceyuan Yang, Sida Peng, Yujun Shen, and Gordon Wetzstein. Grm: Large gaussian reconstruction model for efficient 3d reconstruction and generation. arXiv preprint arXiv:2403.14621, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.49, + 0.908, + 0.545 + ], + "angle": 0, + "content": "[74] Han Yan, Yang Li, Zhennan Wu, Shenzhen Chen, Weixuan Sun, Taizhang Shang, Weizhe Liu, Tian Chen, Xiaqiang Dai, Chao Ma, et al. Frankenstein: Generating semantic-compositional 3d scenes in one tri-plane. 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.548, + 0.908, + 0.602 + ], + "angle": 0, + "content": "[75] Ling Yang, Zhaochen Yu, Chenlin Meng, Minkai Xu, Stefano Ermon, and Bin Cui. Mastering text-to-image diffusion: Recognition, planning, and generating with multimodal llms. arXiv preprint arXiv:2401.11708, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.604, + 0.908, + 0.659 + ], + "angle": 0, + "content": "[76] Ling Yang, Zixiang Zhang, Junlin Han, Bohan Zeng, Runjia Li, Philip Torr, and Wentao Zhang. Semantic score distillation sampling for compositional text-to-3d generation. arXiv preprint arXiv:2410.09009, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.661, + 0.907, + 0.731 + ], + "angle": 0, + "content": "[77] Taoran Yi, Jiemin Fang, Junjie Wang, Guanjun Wu, Lingxi Xie, Xiaopeng Zhang, Wenyu Liu, Qi Tian, and Xinggang Wang. Gaussiandreamer: Fast generation from text to 3d gaussians by bridging 2d and 3d diffusion models. arXiv preprint arXiv, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.733, + 0.908, + 0.787 + ], + "angle": 0, + "content": "[78] Kai Zhang, Lingbo Mo, Wenhu Chen, Huan Sun, and Yu Su. Magicbrush: A manually annotated dataset for instruction-guided image editing. Advances in Neural Information Processing Systems, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.789, + 0.908, + 0.872 + ], + "angle": 0, + "content": "[79] Qihang Zhang, Chaoyang Wang, Aliaksandr Siarohin, Peiye Zhuang, Yinghao Xu, Ceyuan Yang, Dahua Lin, Bolei Zhou, Sergey Tulyakov, and Hsin-Ying Lee. Towards text-guided 3d scene composition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6829-6838, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.907, + 0.902 + ], + "angle": 0, + "content": "[80] Shu Zhang, Xinyi Yang, Yihao Feng, Can Qin, Chia-Chih Chen, Ning Yu, Zeyuan Chen, Huan Wang, Silvio Savarese," + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.908, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18519" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.124, + 0.092, + 0.484, + 0.134 + ], + "angle": 0, + "content": "Stefano Ermon, et al. Hive: Harnessing human feedback for instructional visual editing. arXiv preprint arXiv:2303.09618, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.484, + 0.205 + ], + "angle": 0, + "content": "[81] Xiaoyu Zhou, Xingjian Ran, Yajiao Xiong, Jinlin He, Zhiwei Lin, Yongtao Wang, Deqing Sun, and Ming-Hsuan Yang. Gala3d: Towards text-to-3d complex scene generation via layout-guided generative gaussian splatting. arXiv preprint arXiv:2402.07207, 2024." + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.205 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.956 + ], + "angle": 0, + "content": "18520" + } + ] +] \ No newline at end of file diff --git a/2025/CompGS_ Unleashing 2D Compositionality for Compositional Text-to-3D via Dynamically Optimizing 3D Gaussians/d9350216-461c-42c5-bc6e-3da77f71b98b_origin.pdf b/2025/CompGS_ Unleashing 2D Compositionality for Compositional Text-to-3D via Dynamically Optimizing 3D Gaussians/d9350216-461c-42c5-bc6e-3da77f71b98b_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..16f42826d6536b6297c50df27c0fe2aab4bc7df8 --- /dev/null +++ b/2025/CompGS_ Unleashing 2D Compositionality for Compositional Text-to-3D via Dynamically Optimizing 3D Gaussians/d9350216-461c-42c5-bc6e-3da77f71b98b_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2abb9abca170c0c87108ca272c607216f2be0519df7eb2482bf27aec3e3bc9cf +size 3440721 diff --git a/2025/CompGS_ Unleashing 2D Compositionality for Compositional Text-to-3D via Dynamically Optimizing 3D Gaussians/full.md b/2025/CompGS_ Unleashing 2D Compositionality for Compositional Text-to-3D via Dynamically Optimizing 3D Gaussians/full.md new file mode 100644 index 0000000000000000000000000000000000000000..b16033dd94fa4a824e5eec6957beee73e805554e --- /dev/null +++ b/2025/CompGS_ Unleashing 2D Compositionality for Compositional Text-to-3D via Dynamically Optimizing 3D Gaussians/full.md @@ -0,0 +1,332 @@ +# COMPGS: Unleashing 2D Compositionality for Compositional Text-to-3D via Dynamically Optimizing 3D Gaussians + +Chongjian Ge $^{1,2,3}$ Chenfeng Xu $^{1}$ Yuanfeng Ji $^{2}$ Chensheng Peng $^{1}$ + +Masayoshi Tomizuka1 Ping Luo2 Mingyu Ding1,4* Varun Jampani5* Wei Zhan1 + +1 University of California, Berkeley 2 The University of Hong Kong + +3 Adobe Research 4 UNC-Chapel Hill 5 Stability AI + +![](images/711f275c50c8f04f3791954f81ce0e166e265f31101abe14eb16497beee1ebcb.jpg) +Single 3D Generation +a beautiful butterfly + +![](images/e314c9b51b5258fc6be1831c3b5f283d3208699c1cee517117fa3e913c794d06.jpg) +an owl + +![](images/2245fd20debe10497379649f31b346c997201c81f643591b5d83994ccfecc0ef.jpg) +a bird is drinking water from a cup + +![](images/7c26be72f1f1f944b7c0183724152f786963d9d3f432c274900ae5bf9953cfef.jpg) +a footballer is kicking a soccer ball +an owl perches on a branch + +![](images/6c558797819cb60bc2f573e896a806940ddcd16c8f3e32decd7185f64458568e.jpg) + +![](images/2fcc4315a30b8be42f6dade02e4b9ec2909888df3c2ea19e4b5c0c5a2ac51347.jpg) + +![](images/483d3b1f11e17f8e19223f57a2277944ccab76b0e53319b149b3241003904ebe.jpg) +an owl perches on a branch near a pinecone + +![](images/e521b80e6d8e6bf74a34430ac8f46111c0932774919ce2a117c687c84c446bd0.jpg) +Compositional 3D Generation + +![](images/4e1d03f14646c0d0fdd9ffb11b46b21e9e05f812111ab677c229aa5262ef53dc.jpg) + +![](images/a96d1900beff7b3043bbd55d8c13dc57870d25a7e40c2b74ea64a5c06812c4b7.jpg) + +![](images/fb91e0ee73ba362d49758d690786580adec6a94fcb0d76d47c313c895167fa11.jpg) +a photographer is capturing abeautiful butterfly with camera + +![](images/3c5d98a57054db4df6a264de847be3dd3c554ed184f362b9a72dc2cacb8ec524.jpg) + +![](images/a2be1b7abbcb20740667e2dc4186db6b131370839f3f176e4fb601e1d8316c0c.jpg) +an owl perches on a branch near a pinecone, with a rat below the branch +Figure 1. Illustration of compositional 3D Generation and COMPGS. All the contents are generated by COMPGS. Top row: COMPGS is capable of generating either a single object (e.g., a butterfly) or generating compositional objects with reasonable interactions (e.g., the rightmost figure in the top row). Middle row: Beyond text-to-3D generation, COMPGS can be easily extend to 3D editing by progressively adding objects. The colored texts (e.g., 'a branch', 'a pinecone', 'a rat' in the rightmost figure) denote the added part compared to its previous asset. Bottom row: COMPGS achieves compositional text-to-3D by transferring 2D compositionality to initialize 3D Gaussians. COMPGS is further trained with dynamic SDS optimization to produce plausible results. + +# Abstract + +Recent breakthroughs in text-guided image generation have significantly advanced the field of 3D generation. While generating a single high-quality 3D object is now feasible, generating multiple objects with reasonable interactions within a 3D space, a.k.a. compositional 3D generation, presents substantial challenges. This paper introduces COMPGS, a novel generative framework that employs 3D Gaussian Splatting (GS) for efficient, compositional text-to-3D content generation. To achieve this goal, two core designs are proposed: + +(1) 3D Gaussians Initialization with 2D compositionality: We transfer the well-established 2D compositionality to initialize the Gaussian parameters on an entity-by-entity basis, ensuring both consistent 3D priors for each entity and reasonable interactions among multiple entities; (2) Dynamic Optimization: We propose a dynamic strategy to optimize 3D Gaussians using Score Distillation Sampling (SDS) loss. COMPGS first automatically decomposes 3D Gaussians into distinct entity parts, enabling optimization at both the entity and composition levels. Additionally, COMPGS optimizes across objects of varying scales by dynamically adjusting the + +spatial parameters of each entity, enhancing the generation of fine-grained details, particularly in smaller entities. Qualitative comparisons and quantitative evaluations on $\mathrm{T}^3$ Bench demonstrate the effectiveness of COMPGS in generating compositional 3D objects with superior image quality and semantic alignment over existing methods. COMPGS can also be easily extended to progressively adding 3D objects, facilitating complex scene generation. We hope COMPGS will provide new insights to the compositional 3D generation. + +# 1. Introduction + +3D content creation is essential to the modern media industry, yet it has traditionally been labor-intensive and necessitated professional expertise. Typically, designing a single 3D object takes several hours for an experienced designer, and creating complex scenes with multiple 3D objects (as shown in Fig. 1) requires even more effort. Inspired by the recent success of diffusion models in the text-to-image generation [28, 51, 56, 61], much research has focused on exploring 2D diffusion models for text-to-3D generation. Previous work can be divided into two main methodologies: (1) Feed-forward generation [29, 39, 73], which entails training generalizable diffusion models on 3D data; and (2) Optimization-based generation [10, 41, 47, 52, 68, 71], which utilizes the pretrained 2D diffusion guidance to optimize 3D representations via Score Distillation Sampling (SDS) [52]. + +While existing work has demonstrated the feasibility of generating single 3D objects, they often struggle to produce compositional 3D content with multiple objects and complex interactions. For example, (1) Feed-forward generation methods struggle in generalizing to complex textual descriptions since the amount of 3D training data is extremely limited [17, 18], and most of the data contains only one object; (2) Optimization-based generation methods face significant challenges with current 2D diffusion guidance when optimizing compositional 3D objects. Typically, using 2D diffusion guidance to optimize a single object is feasible, as it's easy to incorporate various attributes into a single object. However, in compositional 3D generation, 2D guidance struggles to accurately compose multiple objects with different attributes and relationships into a coherent scene [30]. For example, given the prompt 'a blue bench on the left of a green car,' distinguishing different attributes within the implicit 2D diffusion priors is challenging, resulting in misaligning attributes to different objects or generating incorrect spatial relationships. Thus, optimization-based compositional generation with standard 2D guidance often causes problems like 3D inconsistencies, multi-faced objects, semantic drift, etc [26, 58]. + +In this work, we propose COMPGS, a generative system based on 3D Gaussian Splatting (GS) [33] for compositional + +text-to-3D generation. To achieve this, we introduce two core designs: + +3D Gaussians Initialization with 2D Compositionality. Unlike the implicit representation, i.e., NeRF [48], COMPGS uses 3D Gaussians as the representation, which helps to achieve feasible parameter initialization with a coarse 3D shape [31, 77]. As shown in the bottom row of Fig. 1, we first apply a text-to-image model [3, 7, 55] to generate an image that accurately captures the compositionality of multiple objects. Then the image is segmented into different sub-objects (a.k.a. entities) according to the entity information in the given prompt. These segmented entities are processed through an image-to-3D model [29, 67] to obtain coarse 3D shapes, which are used to roughly initialize the Gaussian parameters in 3D space, thereby transferring the 2D compositionality to 3D representations. + +Dynamic Optimization. Current 2D guidance [41, 47, 52] struggles with optimizing multiple 3D objects simultaneously in a scene, often leading to 3D inconsistencies and semantic shifts. To address these issues, we introduce a dynamic optimization strategy based on Score Distillation Sampling (SDS) loss, consisting of two key components: (1) COMPGS dynamically divides the training process to optimize different parts of 3D Gaussians. Specifically, it alternates between optimizing a single object (entity-level optimization) and the entire scene (composition-level optimization). This is achieved by labeling and filtering the Gaussian parameters for inference, and updating them through masking gradients. (2) COMPGS dynamically trains the entity-level Gaussians within a normalized 3D space, crucial for compositional 3D generation where object sizes vary. This is particularly challenging when dealing with very small objects, as optimizing Gaussian parameters in a limited 3D space may not adequately capture detailed textures. To mitigate this, we first scale the subspace of each entity to a predefined, standardized volume. Following each training iteration within these standardized volumes, we rescale the Gaussian parameters back to their original sizes. This method dynamically maintains volume consistency throughout the training process, thereby facilitating the capture of fine details within smaller objects. + +With these designs, COMPGS is capable of generating high-quality compositional 3D objects (as shown in the first row of Fig. 1) and progressive 3D adding (as shown in the second row of Fig. 1). We demonstrate the effectiveness of COMPGS both qualitatively and quantitatively. Qualitative comparisons through our user study indicate that COMPGS offers superior image quality and semantic alignment compared to existing models (e.g., Fantasia3D [10], ProlificDreamer [71], VP3D [11], etc.). Besides, COMPGS's performance on $\mathrm{T}^3$ Bench [26] quantitatively highlights its advantages in both semantic control and high-fidelity generation. We hope COMPGS can provide valuable insights into + +the research of compositional 3D generation. + +# 2. Related Work + +Multi-modality 3D Generation. Multi-modality 3D generation can basically categorized into the feed-forward system and optimization-based system. The former one is trained end-to-end on multi-view dataset [5, 17] for zero-shot generation, typically based on 3D representations, including NeRF [29, 67], 3D Gaussians [66, 73], tri-planes [59, 69] and feature grids [32]. Despite the excellent performance achieved, the single-object training data [5, 17, 18, 45] limits their generative abilities in scenarios involving multiple objects. Optimization-based 3D systems lift 2D diffusion priors [7, 8, 51, 56, 57] for 3D generation, and typically train 3D representations on a prompt-by-prompt basis. For example, DreamFusion [52] introduces Score Distillation Sampling (SDS) loss to transfer 2D diffusion models into the 3D domain. Magic3D [53] employs a coarse-to-fine scheme to improve both efficiency and effectiveness, while Fantasy3D [10] decouples the modelling of geometry and appearance. Despite the advancements of these methods addressing various challenges, they also usually struggle with compositional 3D generation. $\mathrm{T}^3$ bench's examination of ten prominent optimization-based models [10, 41, 47, 52, 58, 63, 68, 71, 77] revealed frequent issues in compositional generation. The implicit 2D diffusion guidance used in these methods [10, 52, 58, 71] often fails to maintain 3D consistency across different views, leading to significant discrepancies in compositional 3D scenes. + +Compositional Generation. Compositional generation involves creating content involving multiple objects with logical interactions. These interactions include, but are not limited to, concept relations [43], attribute association with colors [6, 22, 50], and spatial relationships between objects [9, 72]. There is considerable focus on various aspects of compositional 2D generation, such as learning from human feedback [20, 37, 75, 80], enhancing image captions [3, 7, 8, 16], designing effective networks [44], and etc. T2I-CompBench [30] further proposed a comprehensive benchmark and boosted the compositionality of text-to-image models. Compared to 2D compositional generation, its 3D counterpart is under-explored due to 3D geometry and appearance complexities [1, 12, 25, 40, 64, 74, 76]. For example, Set-the-Scene [15] and Scenewiz3d [79] propose to adopt object layouts to generate compositional scenes. However, they are mainly based on the implicit NeRF representations, so the features between objects cannot be well decoupled, resulting in a relatively blurred rendering effect. VP3D [11] takes a different strategy to employ visual features for compositional generation. However, its visual features are only used as implicit supervision, which does not adequately ensure 3D consistency or address the Janus problems [58]. Besides, LucidDreaming [70] requires + +human-annotated layout priors for local optimization, which is labour-intensive and inaccurate; while GraphDreamer [23] optimizes all interactive objects from scratch, which is time-consuming and difficult to optimize. The most concurrent works, GALA3D [81], propose optimizing the spatial relationships of well-trained Gaussians for compositional generation. Though it achieves plausible spatial interactions, it is inefficient in generating complex mutual interactions among objects. More comparisons are in the appendix. We propose COMPGS to address the above challenges, emphasizing efficient optimization and the generation of complex interactions. + +# 3. Methodology + +In this section, we revisit 3D Gaussian Splitting [33] and diffusion priors for 3D generation. We then provide an overview of COMPGS, which includes initializing 3D Gaussians to incorporate compositionality priors and dynamic SDS optimization to generate high-fidelity, 3D consistent objects with complex interactions. The notations used in this section are also detailed in Appendix A.1 for clarity. + +# 3.1. Preliminaries + +3D Gaussian Splatting (GS). GS [33] has recently revolutionized novel-view synthesis of objects/scenes by its real-time rendering. Specifically, GS represents the 3D objects/scenes by $N$ explicit anisotropic Gaussians with center positions $\mu_{i}$ , covariances $\Sigma_{i}$ , opacities $\alpha_{i}$ and colors $c_{i}$ , where $i\in N$ . The color $\mathcal{C}(\mathbf{p})$ of image pixel $\mathbf{p}$ can be calculated through point-based volume rendering [35, 36] by integrating the color and density of the 3D Gaussians intersected by a ray, as follows: + +$$ +\mathcal {C} (\mathbf {p}) = \sum_ {i = 1} ^ {N} c _ {i} \sigma_ {i} \prod_ {j = 1} ^ {i - 1} (1 - \sigma_ {j}), \tag {1} +$$ + +$$ +\sigma_ {i} = \alpha_ {i} \exp \left[ - \frac {1}{2} \left(\mathbf {p} - \hat {\boldsymbol {\mu}} _ {i}\right) ^ {\top} \hat {\Sigma} _ {i} ^ {- 1} \left(\mathbf {p} - \hat {\boldsymbol {\mu}} _ {i}\right) \right], \tag {2} +$$ + +where $\hat{\mu}_i$ and $\hat{\Sigma}_i$ denote the projected center positions and covariances of the 2D Gaussians, transformed from 3D space to the 2D camera's image plane. The image plane can be segmented into tiles during rendering for parallel processing. Unlike implicit representations, such as NeRF [2, 48, 49], GS offers two key advantages for compositional 3D generation: (1) 3D Gaussians facilitate localized rendering, allowing for the rendering of object $A$ in one specific sub-space and object $B$ in another, thus enabling global compositional generation on a divide-and-conquer basis; (2) 3D Gaussians enable direct parameter initialization, simplifying the integration of compositional priors at the initialization stage. + +Diffusion Priors for 3D Generation. Diffusion-based generative models (DMs) [19, 60, 62] have been widely utilized to provide implicit priors for 3D object generation via score distillation sampling (SDS) [52]. Specifically, given a 3D + +![](images/14db37253ead1a77c3d0cef047e8e1ea58295e51bbb6dbc7bf6bd904574900bd.jpg) +Figure 2. Overall pipeline of COMPGS. Given a compositional prompt $V$ , we first use an LLM to decompose it into entity-level prompts $\{v_l\}$ , guiding the segmentation of each entity from the compositional image generated by T2I models. The segmented images initialize entity-level 3D Gaussians via image-to-3D models [29, 67]. COMPGS employs a dynamic optimization strategy, alternating between composition-level optimization of $\theta$ and entity-level optimization of $\{\theta_l\}$ . For entity-level optimization, COMPGS dynamically maintains volume consistency to refine the details of each objects, particularly the small one. + +model whose parameters are $\theta$ , a differentiable rendering process $g$ , the rendered images could be obtained via $\mathrm{x} = g(\theta)$ . To ensure the rendered images resemble those generated by the DM $\phi$ , SDS first formulates the sampled noise $\hat{\epsilon}_{\phi}(\mathrm{z}_t;v,t)$ with the noisy image $\mathrm{z}_t$ , text embedding $v$ , and noise level $t$ . By comparing the difference between the added Gaussian noise $\epsilon$ and the predicted noise $\hat{\epsilon}_{\phi}$ , SDS constructs gradients that could be back-propagated to update $\theta$ via: + +$$ +\nabla_ {\theta} \mathcal {L} _ {\mathrm {S D S}} (\phi , \mathrm {x} = g (\theta)) \triangleq \mathbb {E} _ {t, \epsilon} \left[ w (t) \left(\hat {\epsilon} _ {\phi} \left(\mathbf {z} _ {t}, v, t\right) - \epsilon\right) \frac {\partial \mathrm {x}}{\partial \theta} \right], \tag {3} +$$ + +where $w(t)$ is a weighting function, and the Classifier-free guidance [27] typically amplifies the text conditioning $v$ . + +# 3.2.COMPGS + +We propose COMPGS for efficiently transferring the 2D compositionality to facilitate compositional 3D generation with 3D Gaussians. The overall framework of COMPGS is shown in Fig. 2. We first generate a well-composed image from the given complex prompt. After extracting the subobject (also named entity) information within the prompt by LLM, we utilize the entity prompt to segment the composed image into different parts. Each part will be adopted to initialize a specific space of Gaussians. During the training stage, we propose a dynamic SDS optimization strategy. This strategy first automatically decomposes the training to optimize either entity-level Gaussians or composition-level Gaussians. Then, it employs a volume-adaptive strategy to dynamically optimize varying-sized entities within a consistent 3D space. We detail the initialization process and dynamic SDS optimization in the following sections. + +# 3.2.1. 3D Gaussians Init with 2D Compositionality + +Unlike the implicit NeRF [48] representations, explicit 3D Gaussians can be easily initialized with 3D shapes and colors, which facilitates introducing rough 3D priors to ensure 3D consistency [77]. Although existing text-to-3D or image-to-3D models can generate a single 3D object, they struggle with compositional 3D generation as mentioned in Sec. 2, which hinders the creation of compositional 3D priors needed for initialization. Considering this, we propose initializing the compositional Gaussians on an entity-by-entity basis. + +As shown in Fig. 2 (left), given a complex prompt $V$ , (e.g., 'an owl perches on a branch near a pinecone'), we first adopt a 2D diffusion model [3] to generate a composed image $I$ that faithfully captures the compositionality of multiple objects. We extract each entity information in prompt $V$ by prompting LLM to obtain $L$ different entity-level prompts (i.e., $L = 3$ for $v_{1}$ 'an owl', $v_{2}$ 'a branch', and $v_{3}$ 'a pinecone'). These prompts are adopted in a text-guided segmentation model [34] to decompose image $I$ into various parts $\{I_l\}$ . Until now, each segmented image has only one entity, facilitating using existing image-based 3D generation models [67] to predict a rough triangle mesh $m_{l}(l \in L)$ of the corresponding 3D entity $l$ . To initialize 3D Gaussians $\theta_{l}$ , we index $N$ points from each mesh; the center positions $\mu_i^l \in \mathbb{R}^3 (i \in N, l \in L)$ are the centers of each vertex of $m_{l}$ , and the texture colors $c_i^l \in \mathbb{R}^3 (i \in N, l \in L)$ are queried from each vertex of $m_{l}$ . During the image-to-3D process, we did not perform any cropping or padding on the image $I_{l}$ , ensuring that the spatial positions of each mesh $m_{l}$ correspond to their 2D spatial positions. Additionally, we have marked the 3D layouts of each entity asset with 3D bounding boxes bbox $l$ for subsequent optimization. The + +bounding box coordinates are determined by the outer-most center positions of the entity Gaussian. + +# 3.2.2. Dynamic SDS Optimization + +As mentioned in Sec. 2, existing SDS methods [10, 52, 58, 71] struggle in optimizing compositional 3D scenes due to the challenges of using implicit diffusion priors to maintain multi-objects consistency and interactions. To address this, we introduce dynamic SDS optimization to enable existing SDS losses to remain effective in compositional generation. Dynamic SDS optimization is performed through the following two procedures. + +Automatically Decomposing Optimization to Different Levels. We adopt the Decomposed Optimization (DO) strategy that divides the entire training into $L + 1$ stages, including $L$ stages for entity-level optimization to ensure 3D consistency of each $L$ entities, and one stage for composition-level optimization to refine inter-entity interactions. In each training iteration, we randomly render either (1) an entity Gaussian $\theta_{l}(l \in L)$ to obtain the entity image $\mathrm{x}_l = g(\theta_l)$ ( $\theta_{l} \in \mathrm{bbox}_{l}$ ), or (2) the composed Gaussians $\theta$ to obtain the composed image $\mathrm{x} = g(\theta)$ . Here $g$ denotes the Gaussian Splitting rendering [33] described in Eq. 1. + +To update Gaussian parameters, we apply Eq. 3 for gradients backpropagation. Specifically, for entity-level optimization, we substitute the text embedding $v$ by $v_{l}$ and parameters $\theta$ by $\theta_{l}$ in Eq. 3 to optimize entity Gaussian. We adopt MV-Dream [58] as $\phi$ , which provides multi-view diffusion priors that can better maintain 3D consistency. To ensure that other entity's parameters remain unchanged, we mask the gradients that are not within the corresponding 3D bounding boxes $\mathrm{bbox}_l$ when updating $\theta_{l}$ . For composition-level optimization, we integrate both 2D [56] and 3D diffusion priors [58] to jointly optimize the overall Gaussian parameters $\theta$ . The rationale is that 2D priors promote geometry exploration [53], which enhances the generation of plausible interactions between different entities. + +Volume-adaptively Optimizing Entity-level Gaussians. Compared to the single object generation, compositional 3D generation using SDS presents additional challenges. Specifically, when objects vary in size, optimizing directly in the original 3D space often results in suboptimal generation for smaller objects, as shown the 'pinecone' in Fig. 2. To mitigate this issue, we propose a Volume-adaptive Optimization (VAO) strategy that enhances optimization across objects of varying sizes by dynamically standardizing each entity's space to a standardized volumetric scale. This approach is especially beneficial for improving the generation of fine-grained texture details in smaller entities. + +To scale the 3D bounding box of an entity Gaussian $\theta_{l}$ with a standardized volumetric space bbox $^{std}$ , we first determine the necessary transformations for the centre positions $\mu$ . Specifically, we calculate the shift parameters $\beta$ , and + +scale parameters $\lambda$ , as follows: + +$$ +\beta = \operatorname {M e a n} (\mathrm {b b o x} _ {l}); \quad \lambda = \mathrm {b b o x} _ {s t d} / \mathrm {b b o x} _ {l}, \tag {4} +$$ + +where Mean computes the center coordinates of the given bounding box, and $\mathrm{bbox}_{std}$ is the target standardized space. Then, we zoom-in the entity Gaussian on its center positions by $\hat{\mu} = \lambda (\mu -\beta)$ , where $\hat{\mu}$ denotes the transformed center positions of entity Gaussian. We substitute all $\mu$ to $\hat{\mu}$ in Eq. 1 for SDS optimization. After each training iteration, we perform zooming-back via $\mu = \hat{\mu} /\lambda +\beta$ . The proposed transformation ensures dynamically conducting SDS optimization on each entity Gaussian at a standardized and consistent scale. Experiments in Sec. 4.3 demonstrate its effectiveness in generating high-quality 3D assets. + +# 3.3. Progressive Adding Objects with COMPGS + +The dynamic SDS optimization enables COMPGS to explicitly control the generation of different parts in 3D scene. This capability can be utilized for progressive increasing 3D objects. Specifically, given well-trained 3D Gaussians $\theta$ , we begin by rendering an image from the front view, denoted as $\mathbf{x} = g(\theta)$ . Unlike previous works that directly add 3D objects [13], we adopt MagicBrush [78] to modify the 2D rendered image $\mathbf{x}$ . The edited 2D image can then be utilized to initialize new Gaussians $\hat{\theta}$ in the original 3D space. We further train the added 3D Gaussians to finally introduce new objects into the compositional scenes. As shown in Fig. 1, given well-trained 3D Gaussians (e.g., 'an owl'), we progressively modify the 2D images to add new objects, such as the 'branch' and 'pinecone.' We then train the corresponding 3D Gaussians via dynamic SDS optimization to incorporate these new elements into the 3D scenes. + +# 4. Experiments + +# 4.1. Experimental Settings + +Implementation Details. We implemented COMPGS using ThreeStudio [24]. We use DALL-E [3] to generate the well-established 2D compositionality, LangSAM [46] to segment the entity image, and TripoSR [67] to generate a preliminary 3D prior (i.e., mesh) from each entity image, respectively. During the training, we employed MVDream [58] to optimize entity-level Gaussians, and both MVDream [58] and stabilityai/stablediffusion-2-1-base [57] to optimize composition-level Gaussians. The guidance scale for diffusion models was set as 50. Timestamps were uniformly selected from 0.02 to 0.55 for the first 1,000 iterations and then adjusted to a range from 0.02 to 0.15 for subsequent iterations. We initialized the Gaussian points at $N = 10k$ , and progressively increased the density to a maximum of $1000k$ points. We optimized the overall compositional scenes through 10k iterations to achieve optimal results. However, it is worth noting that we empirically found that training with 5k iterations already produces + +plausible 3D assets. The learning rates are detailed in Appendix A.3. We set the camera parameters, including radius, azimuth, elevation and FoV (field of view) to be the same as [58]. Experiments were conducted on NVIDIA A100 GPUs(40G). The prompt used for decomposing are provided in Appendix A.11. Our code will be released. + +Evaluation Metrics. In addition to user studies and qualitative comparisons, we utilize $\mathrm{T}^3$ Bench [26] to render 300 prompts in order to evaluate compositional 3D generation on two criteria: (1) the visual quality of the 3D objects and (2) the alignment between the 3D objects and the input prompt. For quality evaluation, we captured multi-focal and multiview images and sent them to text-image scoring models, CLIP [54], to obtain an average quality score of the generated 3D scene. Regarding the textual alignment, we first followed [45] to perform 3D captioning via BLIP [38] and GPT4, and then computed the recall of the original prompt within the generated caption via ROUGE-L [42]. + +# 4.2. Performance Comparisons and Analysis. + +Qualitative Evaluation. In Fig. 3, we present qualitative comparisons of compositional 3D generation. Our model, COMPGS, is compared with several open-source methods, including DreamFusion [52], Magic3D [53], Latent-NeRF [47], Fantasia3D [10], SJC [68], and ProlificDreamer [71], as well as methods specifically designed to handle intricate text prompts, VP3D [11] in Fig. 3. The results demonstrate that COMPGS generates compositional 3D objects with superior quality, greater consistency, and more plausible interactions. DreamFusion, for example, fails to generate reasonable multi-objects from compositional prompts, revealing limitations in SDS loss for optimizing multiple objects simultaneously. Magic3D and Latent-NeRF, though guided by a 3D mesh prior, struggle with generating complex and compositional 3D geometry. Despite employing advanced SDS loss, Fantasia3D, SJC, and ProlificDreamer only improve the appearance of the content but do not fundamentally address multi-object generation. Notably, VP3D employs features from compositional images and human feedback to guide compositional 3D generation. However, it often produces unexpected artifacts (e.g., redundant dots in the top two rows of Fig. 3) or less plausible interactions (e.g., strange spatial relationships in the fourth row). This is likely due to the insufficient utilization of 2D compositionality, as VP3D adopts the implicit image features to guide optimization. In contrast, COMPGS generates high-quality and composed content that strictly aligns with the given complex prompts, demonstrating its effectiveness over others. We provide more visual comparisons with both the close-source and open-source methods (including GALA3D [81] in Fig. 9, GraphDreamer [23], DreamGaussian [65] in Fig. 10, etc.) in Appendix A.5 as well as the attached video. + +We also demonstrate COMPGS's capabilities in generat + +Table 1. Quantitative comparisons on $\mathrm{T}^3$ Bench [26] and user studies show COMPGS outperforms feed-forward, optimization-based, and compositional generation models. + +
MethodT3Bench (Multiple Objects)User Study Ranking Score↑
Quality↑Alignment↑Average↑
OpenLRM [29]15.225.520.4-
TripoSR [67]16.728.622.7-
DreamFusion [52]17.314.816.14.54
SJC [68]17.75.811.73.08
LatentNeRF [47]21.719.520.64.14
Fantasia3D [10]22.714.318.52.43
ProlificDreamer [71]45.725.835.83.56
Magic3D [41]26.624.825.74.30
Set-the-Scene [15]20.829.925.4-
VP3D [11]49.131.540.36.71
COMPGS (ours)54.2 (+5.1)37.9(+6.4)46.1(+5.8)7.23
+ +ing more diverse contents in Fig. 4. COMPGS accurately generates both simple spatial relationships (shown in the first row), and more complex interactions, where objects should react to each other (shown in the second row). Besides, COMPGS is not limited to generating just two objects, it can generate multiple objects by expanding the number of entity-level Gaussians. For example, in the third row, COMPGS generates scenes with three and four entities, each with high-quality 3D shapes and textures. + +We conduct a user study for further evaluation. We randomly selected 15 prompts in $\mathrm{T}^3$ Bench, and collected the 3D assets generated by different models. These collected 3D assets were then distributed to individuals for ranking the models based on (1) 3D visual quality and (3) text-3D alignment. We average the ranking of different models as their scores. Results in Tab. 1 show that representative optimization-based models received relatively low average scores, highlighting their limitations in compositional generation. VP3D [11] outperforms its predecessors due to the incorporation of image features but still ranks lower than COMPGS, which further validates our effectiveness. + +Quantitative Evaluation. In Tab. 1, we benchmark the representative models on $\mathrm{T}^3$ Bench, focusing on the multi-objects track. Results indicate COMPGS achieves superior performance in both quality and textual alignment. Compared with feed-forward methods in the first block, COMPGS significantly improves the generation quality by a large margin. Though COMPGS is initialized from the 3D priors obtained from [67], COMPGS still enhances text-3D alignment, likely due to the dynamic optimization on composition-level Gaussians. Furthermore, compared with the optimization-based methods, COMPGS shows significant advantages in both two metrics. Even when compared to methods specifically designed for compositional generation, e.g., VP3D [11] and Set-the-Scene [15], COMPGS demonstrates clear improvements, indicating the effectiveness of our designs. Note that COMPGS is not limited to compositional generation; the results in the track of single-object generation are included in Appendix A.5. + +'A florist is making a bouquet with fresh flowers' + +'An artist is painting on a blank canvas' + +Hot popcorn jump out from the red striped popcorn maker + +'A half-eaten sandwich sits next to a lukewarm thermos' + +DreamFusion + +Magic3D + +LatentNeRF + +Fantasia3D + +SJC + +ProlificDreamer + +VP3D + +COMPGS (ours) + +![](images/0034243e454e8469075c79dde2ee3045ae2aec9a840090edfa6944b627f7a1ab.jpg) +Figure 3. Qualitative comparisons between COMPGS and other text-to-3D models on $\mathrm{T}^3$ Bench (multiple objects track). Compared to others, COMPGS is better at generating highly-composed, high-quality 3D contents that strictly align with the given texts. Watch the animations by clicking them (Not all PDF readers support playing animations. Best viewed in Acrobat/Foxit Reader). + +![](images/fc33d919392e937f5707680aacf60f7db395bb1c9db7c8339123e54e93255f4a.jpg) +Figure 4. More generated samples by COMPGS. Four views are shown. + +Runtime Comparisons. Though COMPGS is trained with 10k steps, we observed that training the model for 5k iterations already produces high-quality content with minimal loss of texture details. We show runtime comparisons with other models in Appendix A.4. In fact, compositional + +3D generation (left) requires a longer training time than single-object generation (right) due to its complexity. Compositional 3D involves optimizing individual objects and ensuring the consistency of compositional scenes, making them more intricate. The runtime is generally relative to + +![](images/7ed8b65f4d2728aec34d12a6988f4a028e98d31d371fa07d88445558537990e5.jpg) +Figure 5. 3D Editing examples of COMPGS. More examples could be found in Appendix A.10. + +![](images/7b4b21f09381cde132f41dc3ee7e5bb78a06d9bd541f3a94a5236c24ddafa9b2.jpg) +Figure 6. Visual results of the ablation studies on key designs. + +the number of objects involved. Compared to open-source compositional 3D methods such as Set-the-Scene [15], Progressive3D [14], and GraphDraemer [23], our proposed COMPGS is more efficient in training. For instance, given the prompt "a parrot talks beside a perch and two bowls," Progressive3D takes approximately 250 minutes for 3D generation, while Set-the-Scene requires around 110 minutes. Since many compositional scene generation methods are not open-sourced, we also present the training steps listed in the papers for straightforward comparisons. In addition, CompGS demonstrates comparable and even superior efficiency over other methods in text-to-single object generation. For example, methods such as Magic3D [41] and Fantastic3D [10] require over five hours to optimize a single object. In contrast, CompGS can achieve the same task in approximately 30 minutes. Details are in Appendix A.4. + +Extended Applications - Progressive 3D Adding. COMPGS allows progressive 3D adding for compositional scenes. We present 3D adding examples in Fig. 5. These examples demonstrate that by transferring edited 2D compositionality and employing dynamic SDS optimization, COMPGS can progressively incorporate new 3D entities into the original 3D scenes. For example, COMPGS can generate the 'chair', 'panda', 'hat' and 'plant' step by step as shown in Fig. 5. + +Limitations. Detailed limitations and future explorations are provided in Appendix A.12 and Appendix A.13. + +Table 2. Ablation Studies on ${\mathrm{T}}^{3}$ Bench [26]. + +
ComponentQualityAlignmentAverage
Full Setting53.838.045.9
- w/o GS Init.22.818.720.8
- w/o DO Strategy46.835.241.0
- w/o VAO Strategy50.836.443.6
+ +# 4.3. Ablation Study + +We conduct ablation studies to validate the effectiveness of three key designs in COMPGS: Gaussian parameters initialization, decomposed optimization (DO), and volume-adaptive optimization (VAO). We randomly choose 20 prompts from $\mathrm{T}^3$ Bench for the quantitative evaluation. Meanwhile, we use Fig. 6 to visualize the effect of each component. + +Initialization. Instead of initializing Gaussian parameters using 2D compositionality, we conduct random initialization within a predefined 3D bounding box for each entity Gaussian. A significant decrease in the quality and alignment metric is observed in Tab. 2. Besides, the visualization in Fig. 6 confirms that training from random initialization with predefined layouts may result in low-quality textures (e.g., owl's face) and missing entities (e.g., the branch). + +Decomposed Optimization. We analyze the decomposed training by discarding the entity-level optimization. Results in Tab. 2 and Fig. 6 both indicate that removing decomposed training leads to low-fidelity generation. For example, in Fig. 6, optimizing composition-level Gaussians results in a relatively blurred generation on the owl's fur and the pinecone, validating that optimization on each single entity is necessary to effectively leverage 2D diffusion priors to guide 3D generation. + +Volume-adaptive Optimization. To evaluate the effectiveness of volume-adaptive optimization, we choose to optimize the entity Gaussians in their original 3D space. We observed that optimizing different entities within varying sizes of 3D space slightly decreased in quality and alignment, which is noticeable for smaller objects in the compositional scene. For instance, the 'pinecone' and 'branch' trained without volume-adaptive optimization exhibited fewer fine-grained details, shown as the green leaves on the branch and the shadow on the pinecone in Fig. 6. + +# 5. Conclusion + +In this paper, we introduce COMPGS, a user-friendly, optimization-based framework, that achieves compositional text-to-3D generation utilizing Gaussian Splitting. Our three core designs, including initializing 3D Gaussians with 2D compositionality, decomposed optimization for either entity-level or composition-level Gaussians, and volume-adaptive optimization, enable COMPGS to generate complex interactions between multiple 3D entities. We hope this work encourages further research in compositional 3D generation. + +Acknowledgment The work was supported by Berkeley DeepDrive. The authors thank Google Cloud Platform for providing the computational resources. The work was done during Chongjian Ge's visit to University of California, Berkeley. + +# References + +[1] Haotian Bai, Yuanhuiyi Lyu, Lutao Jiang, Sijia Li, Haonan Lu, Xiaodong Lin, and Lin Wang. Componerf: Text-guided multi-object compositional nerf with editable 3d scene layout. arXiv preprint arXiv:2303.13843, 2023. +[2] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In IEEE/CVF International Conference on Computer Vision, 2021. +[3] James Betker, Gabriel Goh, Li Jing, Tim Brooks, Jianfeng Wang, Linjie Li, Long Ouyang, Juntang Zhuang, Joyce Lee, Yufei Guo, et al. Improving image generation with better captions. Computer Science. https://cdn.openaia.com/papers/dall-e-3.pdf, 2023. +[4] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. +[5] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. +[6] Hila Chefer, Yuval Alaluf, Yael Vinker, Lior Wolf, and Daniel Cohen-Or. Attend-and-excite: Attention-based semantic guidance for text-to-image diffusion models. ACM Transactions on Graphics (TOG), 2023. +[7] Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, et al. Pixart-alpha: Fast training of diffusion transformer for photorealistic text-to-image synthesis. 2023. +[8] Junsong Chen, Chongjian Ge, Enze Xie, Yue Wu, Lewei Yao, Xiaozhe Ren, Zhongdao Wang, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart- $\backslash$ sigma: Weak-to-strong training of diffusion transformer for 4k text-to-image generation. arXiv preprint arXiv:2403.04692, 2024. +[9] Minghao Chen, Iro Laina, and Andrea Vedaldi. Training-free layout control with cross-attention guidance. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, 2024. +[10] Rui Chen, Yongwei Chen, Ningxin Jiao, and Kui Jia. Fantasia3d: Disentangling geometry and appearance for high-quality text-to-3d content creation. In IEEE/CVF International Conference on Computer Vision, 2023. +[11] Yang Chen, Yingwei Pan, Haibo Yang, Ting Yao, and Tao Mei. Vp3d: Unleashing 2d visual prompt for text-to-3d generation. arXiv preprint arXiv:2403.17001, 2024. + +[12] Yongwei Chen, Tengfei Wang, Tong Wu, Xingang Pan, Kui Jia, and Ziwei Liu. Comboverse: Compositional 3d assets creation using spatially-aware diffusion guidance. arXiv preprint arXiv:2403.12409, 2024. +[13] Kai Cheng, Xiaoxiao Long, Kaizhi Yang, Yao Yao, Wei Yin, Yuexin Ma, Wenping Wang, and Xuejin Chen. Gaussianpro: 3d gaussian splatting with progressive propagation. arXiv preprint arXiv:2402.14650, 2024. +[14] Xinhua Cheng, Tianyu Yang, Jianan Wang, Yu Li, Lei Zhang, Jian Zhang, and Li Yuan. Progressive3d: Progressively local editing for text-to-3d content creation with complex semantic prompts. arXiv preprint arXiv:2310.11784, 2023. +[15] Dana Cohen-Bar, Elad Richardson, Gal Metzer, Raja Giryes, and Daniel Cohen-Or. Set-the-scene: Global-local training for generating controllable nerf scenes. In IEEE/CVF International Conference on Computer Vision, 2023. +[16] Xiaoliang Dai, Ji Hou, Chih-Yao Ma, Sam Tsai, Jialiang Wang, Rui Wang, Peizhao Zhang, Simon Vandenhende, Xiaofang Wang, Abhimanyu Dubey, et al. Emu: Enhancing image generation models using photogenic needles in a haystack. arXiv preprint arXiv:2309.15807, 2023. +[17] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objaverse: A universe of annotated 3d objects. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. +[18] Matt Deitke, Ruoshi Liu, Matthew Wallingford, Huong Ngo, Oscar Michel, Aditya Kusupati, Alan Fan, Christian Laforte, Vikram Voleti, Samir Yitzhak Gadre, et al. Objverse-xl: A universe of $10\mathrm{m} + 3\mathrm{d}$ objects. Advances in Neural Information Processing Systems, 2024. +[19] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in Neural Information Processing Systems, 2021. +[20] Hanze Dong, Wei Xiong, Deepanshu Goyal, Yihan Zhang, Winnie Chow, Rui Pan, Shizhe Diao, Jipeng Zhang, Kashun Shum, and Tong Zhang. Raft: Reward ranked finetuning for generative foundation model alignment. arXiv preprint arXiv:2304.06767, 2023. +[21] Danny Driess, Zhiao Huang, Yunzhu Li, Russ Tedrake, and Marc Toussaint. Learning multi-object dynamics with compositional neural radiance fields. In Conference on robot learning, pages 1755-1768. PMLR, 2023. +[22] Weixi Feng, Xuehai He, Tsu-Jui Fu, Varun Jampani, Arjun Akula, Pradyumna Narayana, Sugato Basu, Xin Eric Wang, and William Yang Wang. Training-free structured diffusion guidance for compositional text-to-image synthesis. arXiv preprint arXiv:2212.05032, 2022. +[23] Gege Gao, Weiyang Liu, Anpei Chen, Andreas Geiger, and Bernhard Scholkopf. Graphdreamer: Compositional 3d scene synthesis from scene graphs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21295-21304, 2024. +[24] Yuan-Chen Guo, Ying-Tian Liu, Ruizhi Shao, Christian Laforte, Vikram Voleti, Guan Luo, Chia-Hao Chen, Zixin Zou, Chen Wang, Yan-Pei Cao, and Song-Hai Zhang. threestudio: A unified framework for 3d content generation. + +https://github.com/threestudio-project/threestudio, 2023. +[25] Haonan Han, Rui Yang, Huan Liao, Jiankai Xing, Zunnan Xu, Xiaoming Yu, Junwei Zha, Xiu Li, and Wanhua Li. Reparo: Compositional 3d assets generation with differentiable 3d layout alignment. arXiv preprint arXiv:2405.18525, 2024. +[26] Yuze He, Yushi Bai, Matthieu Lin, Wang Zhao, Yubin Hu, Jenny Sheng, Ran Yi, Juanzi Li, and Yong-Jin Liu. T3bench: Benchmarking current progress in text-to-3d generation. arXiv preprint arXiv:2310.02977, 2023. +[27] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. Advances in Neural Information Processing Systems, 2022. +[28] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 2020. +[29] Yicong Hong, Kai Zhang, Jiuxiang Gu, Sai Bi, Yang Zhou, Difan Liu, Feng Liu, Kalyan Sunkavalli, Trung Bui, and Hao Tan. Lrm: Large reconstruction model for single image to 3d. arXiv preprint arXiv:2311.04400, 2023. +[30] Kaiyi Huang, Kaiyue Sun, Enze Xie, Zhenguo Li, and Xihui Liu. T2i-compbench: A comprehensive benchmark for open-world compositional text-to-image generation. Advances in Neural Information Processing Systems, 2023. +[31] Heewoo Jun and Alex Nichol. Shap-e: Generating conditional 3d implicit functions. arXiv preprint arXiv:2305.02463, 2023. +[32] Animesh Karnewar, Andrea Vedaldi, David Novotny, and Niloy J Mitra. Holodiffusion: Training a 3d diffusion model using 2d images. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. +[33] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splattering for real-time radiance field rendering. ACM Transactions on Graphics, 2023. +[34] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In IEEE/CVF International Conference on Computer Vision, 2023. +[35] Georgios Kopanas, Julien Philip, Thomas Leimkuhler, and George Drettakis. Point-based neural rendering with per-view optimization. In Computer Graphics Forum, 2021. +[36] Georgios Kopanas, Thomas Leimkuhler, Gilles Rainer, Clément Jambon, and George Drettakis. Neural point catauastics for novel-view synthesis of reflections. ACM Transactions on Graphics (TOG), 2022. +[37] Kimin Lee, Hao Liu, Moonkyung Ryu, Olivia Watkins, Yuqing Du, Craig Boutilier, Pieter Abbeel, Mohammad Ghavamzadeh, and Shixiang Shane Gu. Aligning text-to-image models using human feedback. arXiv preprint arXiv:2302.12192, 2023. +[38] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International Conference on Machine Learning, 2022. +[39] Jiahao Li, Hao Tan, Kai Zhang, Zexiang Xu, Fujun Luan, Yinghao Xu, Yicong Hong, Kalyan Sunkavalli, Greg Shakhnarovich, and Sai Bi. Instant3d: Fast text-to-3d with + +sparse-view generation and large reconstruction model. arXiv preprint arXiv:2311.06214, 2023. +[40] Runjia Li, Junlin Han, Luke Melas-Kyriazi, Chunyi Sun, Zhaochong An, Zhongrui Gui, Shuyang Sun, Philip Torr, and Tomas Jakab. Dreambeast: Distilling 3d fantastical animals with part-aware knowledge transfer. arXiv preprint arXiv:2409.08271, 2024. +[41] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. +[42] Chin-Yew Lin. Rouge: A package for automatic evaluation of summaries. In Text summarization branches out, 2004. +[43] Nan Liu, Shuang Li, Yilun Du, Antonio Torralba, and Joshua B Tenenbaum. Compositional visual generation with composable diffusion models. In European Conference on Computer Vision, 2022. +[44] Nan Liu, Shuang Li, Yilun Du, Antonio Torralba, and Joshua B Tenenbaum. Compositional visual generation with composable diffusion models. In European Conference on Computer Vision, 2022. +[45] Tiange Luo, Chris Rockwell, Honglak Lee, and Justin Johnson. Scalable 3d captioning with pretrained models. Advances in Neural Information Processing Systems, 2024. +[46] Luca Medeiros. Language segment-anything: Sam with text prompt. https://github.com/luca-medeiros/ lang-segment-anything, 2024. Accessed: 2024-05-21. +[47] Gal Metzer, Elad Richardson, Or Patashnik, Raja Giryes, and Daniel Cohen-Or. Latent-nerf for shape-guided generation of 3d shapes and textures. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. +[48] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 2021. +[49] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM transactions on graphics (TOG), 2022. +[50] Dong Huk Park, Samaneh Azadi, Xihui Liu, Trevor Darrell, and Anna Rohrbach. Benchmark for compositional text-to-image synthesis. In Advances in Neural Information Processing Systems, 2021. +[51] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. +[52] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. +[53] Guocheng Qian, Jinjie Mai, Abdullah Hamdi, Jian Ren, Aliaksandr Siarohin, Bing Li, Hsin-Ying Lee, Ivan Skorokhodov, Peter Wonka, Sergey Tulyakov, et al. Magic123: One image to high-quality 3d object generation using both 2d and 3d diffusion priors. arXiv preprint arXiv:2306.17843, 2023. + +[54] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, 2021. +[55] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning, 2021. +[56] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. +[57] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. +[58] Yichun Shi, Peng Wang, Jianglong Ye, Mai Long, Kejie Li, and Xiao Yang. Mvdream: Multi-view diffusion for 3d generation. arXiv preprint arXiv:2308.16512, 2023. +[59] J Ryan Shue, Eric Ryan Chan, Ryan Po, Zachary Ankner, Jiajun Wu, and Gordon Wetzstein. 3d neural field generation using triplane diffusion. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. +[60] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International Conference on Machine Learning, 2015. +[61] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. +[62] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. +[63] Jingxiang Sun, Bo Zhang, Ruizhi Shao, Lizhen Wang, Wen Liu, Zhenda Xie, and Yebin Liu. Dreamcraft3d: Hierarchical 3d generation with bootstrapped diffusion prior. arXiv preprint arXiv:2310.16818, 2023. +[64] Shaorong Sun, Shuchao Pang, Yazhou Yao, and Xiaoshui Huang. Comogen: A controllable text-to-3d multi-object generation framework. arXiv preprint arXiv:2409.00590, 2024. +[65] Jiaxiang Tang, Jiawei Ren, Hang Zhou, Ziwei Liu, and Gang Zeng. Dreamgaussian: Generative gaussian splatting for efficient 3d content creation. arXiv preprint arXiv:2309.16653, 2023. +[66] Jiaxiang Tang, Zhaoxi Chen, Xiaokang Chen, Tengfei Wang, Gang Zeng, and Ziwei Liu. Lgm: Large multi-view gaussian model for high-resolution 3d content creation. arXiv preprint arXiv:2402.05054, 2024. +[67] Dmitry Tochilkin, David Pankratz, Zexiang Liu, Zixuan Huang, Adam Letts, Yangguang Li, Ding Liang, Christian Laforte, Varun Jampani, and Yan-Pei Cao. Triposr: Fast 3d object reconstruction from a single image. arXiv preprint arXiv:2403.02151, 2024. + +[68] Haochen Wang, Xiaodan Du, Jiahao Li, Raymond A Yeh, and Greg Shakhnarovich. Score jacobian chaining: Lifting pretrained 2d diffusion models for 3d generation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. +[69] Tengfei Wang, Bo Zhang, Ting Zhang, Shuyang Gu, Jianmin Bao, Tadas Baltrusaitis, Jingjing Shen, Dong Chen, Fang Wen, Qifeng Chen, et al. Rodin: A generative model for sculpting 3d digital avatars using diffusion. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. +[70] Zhaoning Wang, Ming Li, and Chen Chen. Luciddreaming: Controllable object-centric 3d generation. arXiv preprint arXiv:2312.00588, 2023. +[71] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. Advances in Neural Information Processing Systems, 2024. +[72] Qiucheng Wu, Yujuan Liu, Handong Zhao, Trung Bui, Zhe Lin, Yang Zhang, and Shiyu Chang. Harnessing the spatial-temporal attention of diffusion models for high-fidelity text-to-image synthesis. In IEEE/CVF International Conference on Computer Vision, 2023. +[73] Yinghao Xu, Zifan Shi, Wang Yifan, Hansheng Chen, Ceyuan Yang, Sida Peng, Yujun Shen, and Gordon Wetzstein. Grm: Large gaussian reconstruction model for efficient 3d reconstruction and generation. arXiv preprint arXiv:2403.14621, 2024. +[74] Han Yan, Yang Li, Zhennan Wu, Shenzhen Chen, Weixuan Sun, Taizhang Shang, Weizhe Liu, Tian Chen, Xiaqiang Dai, Chao Ma, et al. Frankenstein: Generating semantic-compositional 3d scenes in one tri-plane. 2024. +[75] Ling Yang, Zhaochen Yu, Chenlin Meng, Minkai Xu, Stefano Ermon, and Bin Cui. Mastering text-to-image diffusion: Recognition, planning, and generating with multimodal llms. arXiv preprint arXiv:2401.11708, 2024. +[76] Ling Yang, Zixiang Zhang, Junlin Han, Bohan Zeng, Runjia Li, Philip Torr, and Wentao Zhang. Semantic score distillation sampling for compositional text-to-3d generation. arXiv preprint arXiv:2410.09009, 2024. +[77] Taoran Yi, Jiemin Fang, Junjie Wang, Guanjun Wu, Lingxi Xie, Xiaopeng Zhang, Wenyu Liu, Qi Tian, and Xinggang Wang. Gaussiandreamer: Fast generation from text to 3d gaussians by bridging 2d and 3d diffusion models. arXiv preprint arXiv, 2023. +[78] Kai Zhang, Lingbo Mo, Wenhu Chen, Huan Sun, and Yu Su. Magicbrush: A manually annotated dataset for instruction-guided image editing. Advances in Neural Information Processing Systems, 2024. +[79] Qihang Zhang, Chaoyang Wang, Aliaksandr Siarohin, Peiye Zhuang, Yinghao Xu, Ceyuan Yang, Dahua Lin, Bolei Zhou, Sergey Tulyakov, and Hsin-Ying Lee. Towards text-guided 3d scene composition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6829-6838, 2024. +[80] Shu Zhang, Xinyi Yang, Yihao Feng, Can Qin, Chia-Chih Chen, Ning Yu, Zeyuan Chen, Huan Wang, Silvio Savarese, + +Stefano Ermon, et al. Hive: Harnessing human feedback for instructional visual editing. arXiv preprint arXiv:2303.09618, 2023. +[81] Xiaoyu Zhou, Xingjian Ran, Yajiao Xiong, Jinlin He, Zhiwei Lin, Yongtao Wang, Deqing Sun, and Ming-Hsuan Yang. Gala3d: Towards text-to-3d complex scene generation via layout-guided generative gaussian splatting. arXiv preprint arXiv:2402.07207, 2024. \ No newline at end of file diff --git a/2025/CompGS_ Unleashing 2D Compositionality for Compositional Text-to-3D via Dynamically Optimizing 3D Gaussians/images.zip b/2025/CompGS_ Unleashing 2D Compositionality for Compositional Text-to-3D via Dynamically Optimizing 3D Gaussians/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..e1234a9267fe8af8992d9a9a4d0f2e6804074b52 --- /dev/null +++ b/2025/CompGS_ Unleashing 2D Compositionality for Compositional Text-to-3D via Dynamically Optimizing 3D Gaussians/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:871ce5424265e59989a142edb1505bbf7bd5926e911023d7003cb6cedddb7b84 +size 518430 diff --git a/2025/CompGS_ Unleashing 2D Compositionality for Compositional Text-to-3D via Dynamically Optimizing 3D Gaussians/layout.json b/2025/CompGS_ Unleashing 2D Compositionality for Compositional Text-to-3D via Dynamically Optimizing 3D Gaussians/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..eb50fb829f5c64e8acc4fdc3e3381389c082592a --- /dev/null +++ b/2025/CompGS_ Unleashing 2D Compositionality for Compositional Text-to-3D via Dynamically Optimizing 3D Gaussians/layout.json @@ -0,0 +1,9960 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 67, + 102, + 544, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 102, + 544, + 140 + ], + "spans": [ + { + "bbox": [ + 67, + 102, + 544, + 140 + ], + "type": "text", + "content": "COMPGS: Unleashing 2D Compositionality for Compositional Text-to-3D via Dynamically Optimizing 3D Gaussians" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 145, + 160, + 463, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 160, + 463, + 175 + ], + "spans": [ + { + "bbox": [ + 145, + 160, + 463, + 175 + ], + "type": "text", + "content": "Chongjian Ge" + }, + { + "bbox": [ + 145, + 160, + 463, + 175 + ], + "type": "inline_equation", + "content": "^{1,2,3}" + }, + { + "bbox": [ + 145, + 160, + 463, + 175 + ], + "type": "text", + "content": " Chenfeng Xu" + }, + { + "bbox": [ + 145, + 160, + 463, + 175 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 145, + 160, + 463, + 175 + ], + "type": "text", + "content": " Yuanfeng Ji" + }, + { + "bbox": [ + 145, + 160, + 463, + 175 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 145, + 160, + 463, + 175 + ], + "type": "text", + "content": " Chensheng Peng" + }, + { + "bbox": [ + 145, + 160, + 463, + 175 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 111, + 175, + 501, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 175, + 501, + 190 + ], + "spans": [ + { + "bbox": [ + 111, + 175, + 501, + 190 + ], + "type": "text", + "content": "Masayoshi Tomizuka1 Ping Luo2 Mingyu Ding1,4* Varun Jampani5* Wei Zhan1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 136, + 190, + 475, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 190, + 475, + 204 + ], + "spans": [ + { + "bbox": [ + 136, + 190, + 475, + 204 + ], + "type": "text", + "content": "1 University of California, Berkeley 2 The University of Hong Kong" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 167, + 203, + 443, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 167, + 203, + 443, + 217 + ], + "spans": [ + { + "bbox": [ + 167, + 203, + 443, + 217 + ], + "type": "text", + "content": "3 Adobe Research 4 UNC-Chapel Hill 5 Stability AI" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 80, + 251, + 174, + 308 + ], + "blocks": [ + { + "bbox": [ + 88, + 236, + 169, + 247 + ], + "lines": [ + { + "bbox": [ + 88, + 236, + 169, + 247 + ], + "spans": [ + { + "bbox": [ + 88, + 236, + 169, + 247 + ], + "type": "text", + "content": "Single 3D Generation" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 80, + 251, + 174, + 308 + ], + "lines": [ + { + "bbox": [ + 80, + 251, + 174, + 308 + ], + "spans": [ + { + "bbox": [ + 80, + 251, + 174, + 308 + ], + "type": "image", + "image_path": "711f275c50c8f04f3791954f81ce0e166e265f31101abe14eb16497beee1ebcb.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 88, + 308, + 167, + 318 + ], + "lines": [ + { + "bbox": [ + 88, + 308, + 167, + 318 + ], + "spans": [ + { + "bbox": [ + 88, + 308, + 167, + 318 + ], + "type": "text", + "content": "a beautiful butterfly" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 87, + 327, + 176, + 381 + ], + "blocks": [ + { + "bbox": [ + 87, + 327, + 176, + 381 + ], + "lines": [ + { + "bbox": [ + 87, + 327, + 176, + 381 + ], + "spans": [ + { + "bbox": [ + 87, + 327, + 176, + 381 + ], + "type": "image", + "image_path": "e314c9b51b5258fc6be1831c3b5f283d3208699c1cee517117fa3e913c794d06.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 113, + 387, + 139, + 395 + ], + "lines": [ + { + "bbox": [ + 113, + 387, + 139, + 395 + ], + "spans": [ + { + "bbox": [ + 113, + 387, + 139, + 395 + ], + "type": "text", + "content": "an owl" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 192, + 250, + 301, + 304 + ], + "blocks": [ + { + "bbox": [ + 192, + 250, + 301, + 304 + ], + "lines": [ + { + "bbox": [ + 192, + 250, + 301, + 304 + ], + "spans": [ + { + "bbox": [ + 192, + 250, + 301, + 304 + ], + "type": "image", + "image_path": "2245fd20debe10497379649f31b346c997201c81f643591b5d83994ccfecc0ef.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 301, + 307, + 388, + 327 + ], + "lines": [ + { + "bbox": [ + 301, + 307, + 388, + 327 + ], + "spans": [ + { + "bbox": [ + 301, + 307, + 388, + 327 + ], + "type": "text", + "content": "a bird is drinking water from a cup" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 193, + 328, + 296, + 384 + ], + "blocks": [ + { + "bbox": [ + 191, + 307, + 277, + 326 + ], + "lines": [ + { + "bbox": [ + 191, + 307, + 277, + 326 + ], + "spans": [ + { + "bbox": [ + 191, + 307, + 277, + 326 + ], + "type": "text", + "content": "a footballer is kicking a soccer ball" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 193, + 328, + 296, + 384 + ], + "lines": [ + { + "bbox": [ + 193, + 328, + 296, + 384 + ], + "spans": [ + { + "bbox": [ + 193, + 328, + 296, + 384 + ], + "type": "image", + "image_path": "7c26be72f1f1f944b7c0183724152f786963d9d3f432c274900ae5bf9953cfef.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 202, + 387, + 276, + 405 + ], + "lines": [ + { + "bbox": [ + 202, + 387, + 276, + 405 + ], + "spans": [ + { + "bbox": [ + 202, + 387, + 276, + 405 + ], + "type": "text", + "content": "an owl perches on a branch" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 315, + 256, + 342, + 302 + ], + "blocks": [ + { + "bbox": [ + 315, + 256, + 342, + 302 + ], + "lines": [ + { + "bbox": [ + 315, + 256, + 342, + 302 + ], + "spans": [ + { + "bbox": [ + 315, + 256, + 342, + 302 + ], + "type": "image", + "image_path": "6c558797819cb60bc2f573e896a806940ddcd16c8f3e32decd7185f64458568e.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 351, + 263, + 405, + 300 + ], + "blocks": [ + { + "bbox": [ + 351, + 263, + 405, + 300 + ], + "lines": [ + { + "bbox": [ + 351, + 263, + 405, + 300 + ], + "spans": [ + { + "bbox": [ + 351, + 263, + 405, + 300 + ], + "type": "image", + "image_path": "2fcc4315a30b8be42f6dade02e4b9ec2909888df3c2ea19e4b5c0c5a2ac51347.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 309, + 331, + 400, + 382 + ], + "blocks": [ + { + "bbox": [ + 309, + 331, + 400, + 382 + ], + "lines": [ + { + "bbox": [ + 309, + 331, + 400, + 382 + ], + "spans": [ + { + "bbox": [ + 309, + 331, + 400, + 382 + ], + "type": "image", + "image_path": "483d3b1f11e17f8e19223f57a2277944ccab76b0e53319b149b3241003904ebe.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 387, + 387, + 405 + ], + "lines": [ + { + "bbox": [ + 304, + 387, + 387, + 405 + ], + "spans": [ + { + "bbox": [ + 304, + 387, + 387, + 405 + ], + "type": "text", + "content": "an owl perches on a branch near a pinecone" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 414, + 250, + 472, + 306 + ], + "blocks": [ + { + "bbox": [ + 302, + 236, + 414, + 247 + ], + "lines": [ + { + "bbox": [ + 302, + 236, + 414, + 247 + ], + "spans": [ + { + "bbox": [ + 302, + 236, + 414, + 247 + ], + "type": "text", + "content": "Compositional 3D Generation" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 414, + 250, + 472, + 306 + ], + "lines": [ + { + "bbox": [ + 414, + 250, + 472, + 306 + ], + "spans": [ + { + "bbox": [ + 414, + 250, + 472, + 306 + ], + "type": "image", + "image_path": "e521b80e6d8e6bf74a34430ac8f46111c0932774919ce2a117c687c84c446bd0.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 445, + 251, + 494, + 291 + ], + "blocks": [ + { + "bbox": [ + 445, + 251, + 494, + 291 + ], + "lines": [ + { + "bbox": [ + 445, + 251, + 494, + 291 + ], + "spans": [ + { + "bbox": [ + 445, + 251, + 494, + 291 + ], + "type": "image", + "image_path": "4e1d03f14646c0d0fdd9ffb11b46b21e9e05f812111ab677c229aa5262ef53dc.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 501, + 250, + 527, + 306 + ], + "blocks": [ + { + "bbox": [ + 501, + 250, + 527, + 306 + ], + "lines": [ + { + "bbox": [ + 501, + 250, + 527, + 306 + ], + "spans": [ + { + "bbox": [ + 501, + 250, + 527, + 306 + ], + "type": "image", + "image_path": "a96d1900beff7b3043bbd55d8c13dc57870d25a7e40c2b74ea64a5c06812c4b7.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 416, + 327, + 466, + 382 + ], + "blocks": [ + { + "bbox": [ + 403, + 308, + 522, + 327 + ], + "lines": [ + { + "bbox": [ + 403, + 308, + 522, + 327 + ], + "spans": [ + { + "bbox": [ + 403, + 308, + 522, + 327 + ], + "type": "text", + "content": "a photographer is capturing abeautiful butterfly with camera" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 416, + 327, + 466, + 382 + ], + "lines": [ + { + "bbox": [ + 416, + 327, + 466, + 382 + ], + "spans": [ + { + "bbox": [ + 416, + 327, + 466, + 382 + ], + "type": "image", + "image_path": "fb91e0ee73ba362d49758d690786580adec6a94fcb0d76d47c313c895167fa11.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 477, + 329, + 520, + 383 + ], + "blocks": [ + { + "bbox": [ + 477, + 329, + 520, + 383 + ], + "lines": [ + { + "bbox": [ + 477, + 329, + 520, + 383 + ], + "spans": [ + { + "bbox": [ + 477, + 329, + 520, + 383 + ], + "type": "image", + "image_path": "3c5d98a57054db4df6a264de847be3dd3c554ed184f362b9a72dc2cacb8ec524.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 78, + 415, + 532, + 503 + ], + "blocks": [ + { + "bbox": [ + 399, + 386, + 525, + 415 + ], + "lines": [ + { + "bbox": [ + 399, + 386, + 525, + 415 + ], + "spans": [ + { + "bbox": [ + 399, + 386, + 525, + 415 + ], + "type": "text", + "content": "an owl perches on a branch near a pinecone, with a rat below the branch" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 78, + 415, + 532, + 503 + ], + "lines": [ + { + "bbox": [ + 78, + 415, + 532, + 503 + ], + "spans": [ + { + "bbox": [ + 78, + 415, + 532, + 503 + ], + "type": "image", + "image_path": "a2be1b7abbcb20740667e2dc4186db6b131370839f3f176e4fb601e1d8316c0c.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 504, + 555, + 570 + ], + "lines": [ + { + "bbox": [ + 55, + 504, + 555, + 570 + ], + "spans": [ + { + "bbox": [ + 55, + 504, + 555, + 570 + ], + "type": "text", + "content": "Figure 1. Illustration of compositional 3D Generation and COMPGS. All the contents are generated by COMPGS. Top row: COMPGS is capable of generating either a single object (e.g., a butterfly) or generating compositional objects with reasonable interactions (e.g., the rightmost figure in the top row). Middle row: Beyond text-to-3D generation, COMPGS can be easily extend to 3D editing by progressively adding objects. The colored texts (e.g., 'a branch', 'a pinecone', 'a rat' in the rightmost figure) denote the added part compared to its previous asset. Bottom row: COMPGS achieves compositional text-to-3D by transferring 2D compositionality to initialize 3D Gaussians. COMPGS is further trained with dynamic SDS optimization to produce plausible results." + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_caption" + } + ], + "index": 31 + }, + { + "bbox": [ + 152, + 574, + 200, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 574, + 200, + 586 + ], + "spans": [ + { + "bbox": [ + 152, + 574, + 200, + 586 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 55, + 605, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 297, + 714 + ], + "type": "text", + "content": "Recent breakthroughs in text-guided image generation have significantly advanced the field of 3D generation. While generating a single high-quality 3D object is now feasible, generating multiple objects with reasonable interactions within a 3D space, a.k.a. compositional 3D generation, presents substantial challenges. This paper introduces COMPGS, a novel generative framework that employs 3D Gaussian Splatting (GS) for efficient, compositional text-to-3D content generation. To achieve this goal, two core designs are proposed:" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 313, + 574, + 556, + 708 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 574, + 556, + 708 + ], + "spans": [ + { + "bbox": [ + 313, + 574, + 556, + 708 + ], + "type": "text", + "content": "(1) 3D Gaussians Initialization with 2D compositionality: We transfer the well-established 2D compositionality to initialize the Gaussian parameters on an entity-by-entity basis, ensuring both consistent 3D priors for each entity and reasonable interactions among multiple entities; (2) Dynamic Optimization: We propose a dynamic strategy to optimize 3D Gaussians using Score Distillation Sampling (SDS) loss. COMPGS first automatically decomposes 3D Gaussians into distinct entity parts, enabling optimization at both the entity and composition levels. Additionally, COMPGS optimizes across objects of varying scales by dynamically adjusting the" + } + ] + } + ], + "index": 35 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 15 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 15 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 15 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 198, + 13, + 443, + 25 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 13, + 443, + 25 + ], + "spans": [ + { + "bbox": [ + 198, + 13, + 443, + 25 + ], + "type": "text", + "content": "Except for this watermark, it is identical to the accepted version;" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 179, + 25, + 462, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 25, + 462, + 37 + ], + "spans": [ + { + "bbox": [ + 179, + 25, + 462, + 37 + ], + "type": "text", + "content": "the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18509" + } + ] + } + ], + "index": 36 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 54, + 72, + 298, + 182 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 72, + 298, + 182 + ], + "spans": [ + { + "bbox": [ + 54, + 72, + 298, + 182 + ], + "type": "text", + "content": "spatial parameters of each entity, enhancing the generation of fine-grained details, particularly in smaller entities. Qualitative comparisons and quantitative evaluations on " + }, + { + "bbox": [ + 54, + 72, + 298, + 182 + ], + "type": "inline_equation", + "content": "\\mathrm{T}^3" + }, + { + "bbox": [ + 54, + 72, + 298, + 182 + ], + "type": "text", + "content": " Bench demonstrate the effectiveness of COMPGS in generating compositional 3D objects with superior image quality and semantic alignment over existing methods. COMPGS can also be easily extended to progressively adding 3D objects, facilitating complex scene generation. We hope COMPGS will provide new insights to the compositional 3D generation." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 201, + 138, + 215 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 201, + 138, + 215 + ], + "spans": [ + { + "bbox": [ + 55, + 201, + 138, + 215 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 222, + 297, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 222, + 297, + 413 + ], + "spans": [ + { + "bbox": [ + 56, + 222, + 297, + 413 + ], + "type": "text", + "content": "3D content creation is essential to the modern media industry, yet it has traditionally been labor-intensive and necessitated professional expertise. Typically, designing a single 3D object takes several hours for an experienced designer, and creating complex scenes with multiple 3D objects (as shown in Fig. 1) requires even more effort. Inspired by the recent success of diffusion models in the text-to-image generation [28, 51, 56, 61], much research has focused on exploring 2D diffusion models for text-to-3D generation. Previous work can be divided into two main methodologies: (1) Feed-forward generation [29, 39, 73], which entails training generalizable diffusion models on 3D data; and (2) Optimization-based generation [10, 41, 47, 52, 68, 71], which utilizes the pretrained 2D diffusion guidance to optimize 3D representations via Score Distillation Sampling (SDS) [52]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 414, + 296, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 414, + 296, + 688 + ], + "spans": [ + { + "bbox": [ + 56, + 414, + 296, + 688 + ], + "type": "text", + "content": "While existing work has demonstrated the feasibility of generating single 3D objects, they often struggle to produce compositional 3D content with multiple objects and complex interactions. For example, (1) Feed-forward generation methods struggle in generalizing to complex textual descriptions since the amount of 3D training data is extremely limited [17, 18], and most of the data contains only one object; (2) Optimization-based generation methods face significant challenges with current 2D diffusion guidance when optimizing compositional 3D objects. Typically, using 2D diffusion guidance to optimize a single object is feasible, as it's easy to incorporate various attributes into a single object. However, in compositional 3D generation, 2D guidance struggles to accurately compose multiple objects with different attributes and relationships into a coherent scene [30]. For example, given the prompt 'a blue bench on the left of a green car,' distinguishing different attributes within the implicit 2D diffusion priors is challenging, resulting in misaligning attributes to different objects or generating incorrect spatial relationships. Thus, optimization-based compositional generation with standard 2D guidance often causes problems like 3D inconsistencies, multi-faced objects, semantic drift, etc [26, 58]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": "In this work, we propose COMPGS, a generative system based on 3D Gaussian Splatting (GS) [33] for compositional" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "type": "text", + "content": "text-to-3D generation. To achieve this, we introduce two core designs:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 99, + 556, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 99, + 556, + 266 + ], + "spans": [ + { + "bbox": [ + 313, + 99, + 556, + 266 + ], + "type": "text", + "content": "3D Gaussians Initialization with 2D Compositionality. Unlike the implicit representation, i.e., NeRF [48], COMPGS uses 3D Gaussians as the representation, which helps to achieve feasible parameter initialization with a coarse 3D shape [31, 77]. As shown in the bottom row of Fig. 1, we first apply a text-to-image model [3, 7, 55] to generate an image that accurately captures the compositionality of multiple objects. Then the image is segmented into different sub-objects (a.k.a. entities) according to the entity information in the given prompt. These segmented entities are processed through an image-to-3D model [29, 67] to obtain coarse 3D shapes, which are used to roughly initialize the Gaussian parameters in 3D space, thereby transferring the 2D compositionality to 3D representations." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 269, + 556, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 269, + 556, + 567 + ], + "spans": [ + { + "bbox": [ + 313, + 269, + 556, + 567 + ], + "type": "text", + "content": "Dynamic Optimization. Current 2D guidance [41, 47, 52] struggles with optimizing multiple 3D objects simultaneously in a scene, often leading to 3D inconsistencies and semantic shifts. To address these issues, we introduce a dynamic optimization strategy based on Score Distillation Sampling (SDS) loss, consisting of two key components: (1) COMPGS dynamically divides the training process to optimize different parts of 3D Gaussians. Specifically, it alternates between optimizing a single object (entity-level optimization) and the entire scene (composition-level optimization). This is achieved by labeling and filtering the Gaussian parameters for inference, and updating them through masking gradients. (2) COMPGS dynamically trains the entity-level Gaussians within a normalized 3D space, crucial for compositional 3D generation where object sizes vary. This is particularly challenging when dealing with very small objects, as optimizing Gaussian parameters in a limited 3D space may not adequately capture detailed textures. To mitigate this, we first scale the subspace of each entity to a predefined, standardized volume. Following each training iteration within these standardized volumes, we rescale the Gaussian parameters back to their original sizes. This method dynamically maintains volume consistency throughout the training process, thereby facilitating the capture of fine details within smaller objects." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 570, + 557, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 557, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 557, + 714 + ], + "type": "text", + "content": "With these designs, COMPGS is capable of generating high-quality compositional 3D objects (as shown in the first row of Fig. 1) and progressive 3D adding (as shown in the second row of Fig. 1). We demonstrate the effectiveness of COMPGS both qualitatively and quantitatively. Qualitative comparisons through our user study indicate that COMPGS offers superior image quality and semantic alignment compared to existing models (e.g., Fantasia3D [10], ProlificDreamer [71], VP3D [11], etc.). Besides, COMPGS's performance on " + }, + { + "bbox": [ + 313, + 570, + 557, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{T}^3" + }, + { + "bbox": [ + 313, + 570, + 557, + 714 + ], + "type": "text", + "content": " Bench [26] quantitatively highlights its advantages in both semantic control and high-fidelity generation. We hope COMPGS can provide valuable insights into" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "18510" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 237, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 237, + 84 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 237, + 84 + ], + "type": "text", + "content": "the research of compositional 3D generation." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 95, + 142, + 106 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 95, + 142, + 106 + ], + "spans": [ + { + "bbox": [ + 55, + 95, + 142, + 106 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 114, + 297, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 114, + 297, + 424 + ], + "spans": [ + { + "bbox": [ + 55, + 114, + 297, + 424 + ], + "type": "text", + "content": "Multi-modality 3D Generation. Multi-modality 3D generation can basically categorized into the feed-forward system and optimization-based system. The former one is trained end-to-end on multi-view dataset [5, 17] for zero-shot generation, typically based on 3D representations, including NeRF [29, 67], 3D Gaussians [66, 73], tri-planes [59, 69] and feature grids [32]. Despite the excellent performance achieved, the single-object training data [5, 17, 18, 45] limits their generative abilities in scenarios involving multiple objects. Optimization-based 3D systems lift 2D diffusion priors [7, 8, 51, 56, 57] for 3D generation, and typically train 3D representations on a prompt-by-prompt basis. For example, DreamFusion [52] introduces Score Distillation Sampling (SDS) loss to transfer 2D diffusion models into the 3D domain. Magic3D [53] employs a coarse-to-fine scheme to improve both efficiency and effectiveness, while Fantasy3D [10] decouples the modelling of geometry and appearance. Despite the advancements of these methods addressing various challenges, they also usually struggle with compositional 3D generation. " + }, + { + "bbox": [ + 55, + 114, + 297, + 424 + ], + "type": "inline_equation", + "content": "\\mathrm{T}^3" + }, + { + "bbox": [ + 55, + 114, + 297, + 424 + ], + "type": "text", + "content": " bench's examination of ten prominent optimization-based models [10, 41, 47, 52, 58, 63, 68, 71, 77] revealed frequent issues in compositional generation. The implicit 2D diffusion guidance used in these methods [10, 52, 58, 71] often fails to maintain 3D consistency across different views, leading to significant discrepancies in compositional 3D scenes." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 426, + 298, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 426, + 298, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 426, + 298, + 714 + ], + "type": "text", + "content": "Compositional Generation. Compositional generation involves creating content involving multiple objects with logical interactions. These interactions include, but are not limited to, concept relations [43], attribute association with colors [6, 22, 50], and spatial relationships between objects [9, 72]. There is considerable focus on various aspects of compositional 2D generation, such as learning from human feedback [20, 37, 75, 80], enhancing image captions [3, 7, 8, 16], designing effective networks [44], and etc. T2I-CompBench [30] further proposed a comprehensive benchmark and boosted the compositionality of text-to-image models. Compared to 2D compositional generation, its 3D counterpart is under-explored due to 3D geometry and appearance complexities [1, 12, 25, 40, 64, 74, 76]. For example, Set-the-Scene [15] and Scenewiz3d [79] propose to adopt object layouts to generate compositional scenes. However, they are mainly based on the implicit NeRF representations, so the features between objects cannot be well decoupled, resulting in a relatively blurred rendering effect. VP3D [11] takes a different strategy to employ visual features for compositional generation. However, its visual features are only used as implicit supervision, which does not adequately ensure 3D consistency or address the Janus problems [58]. Besides, LucidDreaming [70] requires" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 556, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 556, + 216 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 556, + 216 + ], + "type": "text", + "content": "human-annotated layout priors for local optimization, which is labour-intensive and inaccurate; while GraphDreamer [23] optimizes all interactive objects from scratch, which is time-consuming and difficult to optimize. The most concurrent works, GALA3D [81], propose optimizing the spatial relationships of well-trained Gaussians for compositional generation. Though it achieves plausible spatial interactions, it is inefficient in generating complex mutual interactions among objects. More comparisons are in the appendix. We propose COMPGS to address the above challenges, emphasizing efficient optimization and the generation of complex interactions." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 221, + 397, + 235 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 221, + 397, + 235 + ], + "spans": [ + { + "bbox": [ + 313, + 221, + 397, + 235 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 241, + 556, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 241, + 556, + 326 + ], + "spans": [ + { + "bbox": [ + 313, + 241, + 556, + 326 + ], + "type": "text", + "content": "In this section, we revisit 3D Gaussian Splitting [33] and diffusion priors for 3D generation. We then provide an overview of COMPGS, which includes initializing 3D Gaussians to incorporate compositionality priors and dynamic SDS optimization to generate high-fidelity, 3D consistent objects with complex interactions. The notations used in this section are also detailed in Appendix A.1 for clarity." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 332, + 400, + 344 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 332, + 400, + 344 + ], + "spans": [ + { + "bbox": [ + 313, + 332, + 400, + 344 + ], + "type": "text", + "content": "3.1. Preliminaries" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 349, + 556, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 349, + 556, + 449 + ], + "spans": [ + { + "bbox": [ + 313, + 349, + 556, + 449 + ], + "type": "text", + "content": "3D Gaussian Splatting (GS). GS [33] has recently revolutionized novel-view synthesis of objects/scenes by its real-time rendering. Specifically, GS represents the 3D objects/scenes by " + }, + { + "bbox": [ + 313, + 349, + 556, + 449 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 349, + 556, + 449 + ], + "type": "text", + "content": " explicit anisotropic Gaussians with center positions " + }, + { + "bbox": [ + 313, + 349, + 556, + 449 + ], + "type": "inline_equation", + "content": "\\mu_{i}" + }, + { + "bbox": [ + 313, + 349, + 556, + 449 + ], + "type": "text", + "content": ", covariances " + }, + { + "bbox": [ + 313, + 349, + 556, + 449 + ], + "type": "inline_equation", + "content": "\\Sigma_{i}" + }, + { + "bbox": [ + 313, + 349, + 556, + 449 + ], + "type": "text", + "content": ", opacities " + }, + { + "bbox": [ + 313, + 349, + 556, + 449 + ], + "type": "inline_equation", + "content": "\\alpha_{i}" + }, + { + "bbox": [ + 313, + 349, + 556, + 449 + ], + "type": "text", + "content": " and colors " + }, + { + "bbox": [ + 313, + 349, + 556, + 449 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 313, + 349, + 556, + 449 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 349, + 556, + 449 + ], + "type": "inline_equation", + "content": "i\\in N" + }, + { + "bbox": [ + 313, + 349, + 556, + 449 + ], + "type": "text", + "content": ". The color " + }, + { + "bbox": [ + 313, + 349, + 556, + 449 + ], + "type": "inline_equation", + "content": "\\mathcal{C}(\\mathbf{p})" + }, + { + "bbox": [ + 313, + 349, + 556, + 449 + ], + "type": "text", + "content": " of image pixel " + }, + { + "bbox": [ + 313, + 349, + 556, + 449 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 313, + 349, + 556, + 449 + ], + "type": "text", + "content": " can be calculated through point-based volume rendering [35, 36] by integrating the color and density of the 3D Gaussians intersected by a ray, as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 342, + 456, + 555, + 487 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 342, + 456, + 555, + 487 + ], + "spans": [ + { + "bbox": [ + 342, + 456, + 555, + 487 + ], + "type": "interline_equation", + "content": "\\mathcal {C} (\\mathbf {p}) = \\sum_ {i = 1} ^ {N} c _ {i} \\sigma_ {i} \\prod_ {j = 1} ^ {i - 1} (1 - \\sigma_ {j}), \\tag {1}", + "image_path": "bcde5b4f32edac86b726f1ae6ddebd63a1b9be5b69d9a3263973f45d8fe652a1.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 353, + 488, + 555, + 514 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 488, + 555, + 514 + ], + "spans": [ + { + "bbox": [ + 353, + 488, + 555, + 514 + ], + "type": "interline_equation", + "content": "\\sigma_ {i} = \\alpha_ {i} \\exp \\left[ - \\frac {1}{2} \\left(\\mathbf {p} - \\hat {\\boldsymbol {\\mu}} _ {i}\\right) ^ {\\top} \\hat {\\Sigma} _ {i} ^ {- 1} \\left(\\mathbf {p} - \\hat {\\boldsymbol {\\mu}} _ {i}\\right) \\right], \\tag {2}", + "image_path": "ef86fb7d6f047bd0b1aa9bef804a14cf71f47b0b55e568c22aa5237ec0cc1863.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 521, + 556, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 556, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 556, + 665 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 521, + 556, + 665 + ], + "type": "inline_equation", + "content": "\\hat{\\mu}_i" + }, + { + "bbox": [ + 313, + 521, + 556, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 521, + 556, + 665 + ], + "type": "inline_equation", + "content": "\\hat{\\Sigma}_i" + }, + { + "bbox": [ + 313, + 521, + 556, + 665 + ], + "type": "text", + "content": " denote the projected center positions and covariances of the 2D Gaussians, transformed from 3D space to the 2D camera's image plane. The image plane can be segmented into tiles during rendering for parallel processing. Unlike implicit representations, such as NeRF [2, 48, 49], GS offers two key advantages for compositional 3D generation: (1) 3D Gaussians facilitate localized rendering, allowing for the rendering of object " + }, + { + "bbox": [ + 313, + 521, + 556, + 665 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 313, + 521, + 556, + 665 + ], + "type": "text", + "content": " in one specific sub-space and object " + }, + { + "bbox": [ + 313, + 521, + 556, + 665 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 313, + 521, + 556, + 665 + ], + "type": "text", + "content": " in another, thus enabling global compositional generation on a divide-and-conquer basis; (2) 3D Gaussians enable direct parameter initialization, simplifying the integration of compositional priors at the initialization stage." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 666, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 666, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 666, + 556, + 714 + ], + "type": "text", + "content": "Diffusion Priors for 3D Generation. Diffusion-based generative models (DMs) [19, 60, 62] have been widely utilized to provide implicit priors for 3D object generation via score distillation sampling (SDS) [52]. Specifically, given a 3D" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18511" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 69, + 85, + 541, + 249 + ], + "blocks": [ + { + "bbox": [ + 69, + 85, + 541, + 249 + ], + "lines": [ + { + "bbox": [ + 69, + 85, + 541, + 249 + ], + "spans": [ + { + "bbox": [ + 69, + 85, + 541, + 249 + ], + "type": "image", + "image_path": "14db37253ead1a77c3d0cef047e8e1ea58295e51bbb6dbc7bf6bd904574900bd.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 250, + 555, + 302 + ], + "lines": [ + { + "bbox": [ + 55, + 250, + 555, + 302 + ], + "spans": [ + { + "bbox": [ + 55, + 250, + 555, + 302 + ], + "type": "text", + "content": "Figure 2. Overall pipeline of COMPGS. Given a compositional prompt " + }, + { + "bbox": [ + 55, + 250, + 555, + 302 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 55, + 250, + 555, + 302 + ], + "type": "text", + "content": ", we first use an LLM to decompose it into entity-level prompts " + }, + { + "bbox": [ + 55, + 250, + 555, + 302 + ], + "type": "inline_equation", + "content": "\\{v_l\\}" + }, + { + "bbox": [ + 55, + 250, + 555, + 302 + ], + "type": "text", + "content": ", guiding the segmentation of each entity from the compositional image generated by T2I models. The segmented images initialize entity-level 3D Gaussians via image-to-3D models [29, 67]. COMPGS employs a dynamic optimization strategy, alternating between composition-level optimization of " + }, + { + "bbox": [ + 55, + 250, + 555, + 302 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 55, + 250, + 555, + 302 + ], + "type": "text", + "content": " and entity-level optimization of " + }, + { + "bbox": [ + 55, + 250, + 555, + 302 + ], + "type": "inline_equation", + "content": "\\{\\theta_l\\}" + }, + { + "bbox": [ + 55, + 250, + 555, + 302 + ], + "type": "text", + "content": ". For entity-level optimization, COMPGS dynamically maintains volume consistency to refine the details of each objects, particularly the small one." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 313, + 296, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 313, + 296, + 409 + ], + "spans": [ + { + "bbox": [ + 55, + 313, + 296, + 409 + ], + "type": "text", + "content": "model whose parameters are " + }, + { + "bbox": [ + 55, + 313, + 296, + 409 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 55, + 313, + 296, + 409 + ], + "type": "text", + "content": ", a differentiable rendering process " + }, + { + "bbox": [ + 55, + 313, + 296, + 409 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 55, + 313, + 296, + 409 + ], + "type": "text", + "content": ", the rendered images could be obtained via " + }, + { + "bbox": [ + 55, + 313, + 296, + 409 + ], + "type": "inline_equation", + "content": "\\mathrm{x} = g(\\theta)" + }, + { + "bbox": [ + 55, + 313, + 296, + 409 + ], + "type": "text", + "content": ". To ensure the rendered images resemble those generated by the DM " + }, + { + "bbox": [ + 55, + 313, + 296, + 409 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 55, + 313, + 296, + 409 + ], + "type": "text", + "content": ", SDS first formulates the sampled noise " + }, + { + "bbox": [ + 55, + 313, + 296, + 409 + ], + "type": "inline_equation", + "content": "\\hat{\\epsilon}_{\\phi}(\\mathrm{z}_t;v,t)" + }, + { + "bbox": [ + 55, + 313, + 296, + 409 + ], + "type": "text", + "content": " with the noisy image " + }, + { + "bbox": [ + 55, + 313, + 296, + 409 + ], + "type": "inline_equation", + "content": "\\mathrm{z}_t" + }, + { + "bbox": [ + 55, + 313, + 296, + 409 + ], + "type": "text", + "content": ", text embedding " + }, + { + "bbox": [ + 55, + 313, + 296, + 409 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 55, + 313, + 296, + 409 + ], + "type": "text", + "content": ", and noise level " + }, + { + "bbox": [ + 55, + 313, + 296, + 409 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 313, + 296, + 409 + ], + "type": "text", + "content": ". By comparing the difference between the added Gaussian noise " + }, + { + "bbox": [ + 55, + 313, + 296, + 409 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 55, + 313, + 296, + 409 + ], + "type": "text", + "content": " and the predicted noise " + }, + { + "bbox": [ + 55, + 313, + 296, + 409 + ], + "type": "inline_equation", + "content": "\\hat{\\epsilon}_{\\phi}" + }, + { + "bbox": [ + 55, + 313, + 296, + 409 + ], + "type": "text", + "content": ", SDS constructs gradients that could be back-propagated to update " + }, + { + "bbox": [ + 55, + 313, + 296, + 409 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 55, + 313, + 296, + 409 + ], + "type": "text", + "content": " via:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 63, + 427, + 296, + 460 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 427, + 296, + 460 + ], + "spans": [ + { + "bbox": [ + 63, + 427, + 296, + 460 + ], + "type": "interline_equation", + "content": "\\nabla_ {\\theta} \\mathcal {L} _ {\\mathrm {S D S}} (\\phi , \\mathrm {x} = g (\\theta)) \\triangleq \\mathbb {E} _ {t, \\epsilon} \\left[ w (t) \\left(\\hat {\\epsilon} _ {\\phi} \\left(\\mathbf {z} _ {t}, v, t\\right) - \\epsilon\\right) \\frac {\\partial \\mathrm {x}}{\\partial \\theta} \\right], \\tag {3}", + "image_path": "f1a757a215e64ad7f0a1ff97517bd08adecd21a8f6a4190481506095d2a61dd2.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 460, + 295, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 460, + 295, + 485 + ], + "spans": [ + { + "bbox": [ + 55, + 460, + 295, + 485 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 460, + 295, + 485 + ], + "type": "inline_equation", + "content": "w(t)" + }, + { + "bbox": [ + 55, + 460, + 295, + 485 + ], + "type": "text", + "content": " is a weighting function, and the Classifier-free guidance [27] typically amplifies the text conditioning " + }, + { + "bbox": [ + 55, + 460, + 295, + 485 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 55, + 460, + 295, + 485 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 501, + 119, + 512 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 501, + 119, + 512 + ], + "spans": [ + { + "bbox": [ + 55, + 501, + 119, + 512 + ], + "type": "text", + "content": "3.2.COMPGS" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 521, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 521, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 521, + 297, + 714 + ], + "type": "text", + "content": "We propose COMPGS for efficiently transferring the 2D compositionality to facilitate compositional 3D generation with 3D Gaussians. The overall framework of COMPGS is shown in Fig. 2. We first generate a well-composed image from the given complex prompt. After extracting the subobject (also named entity) information within the prompt by LLM, we utilize the entity prompt to segment the composed image into different parts. Each part will be adopted to initialize a specific space of Gaussians. During the training stage, we propose a dynamic SDS optimization strategy. This strategy first automatically decomposes the training to optimize either entity-level Gaussians or composition-level Gaussians. Then, it employs a volume-adaptive strategy to dynamically optimize varying-sized entities within a consistent 3D space. We detail the initialization process and dynamic SDS optimization in the following sections." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 312, + 529, + 325 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 312, + 529, + 325 + ], + "spans": [ + { + "bbox": [ + 313, + 312, + 529, + 325 + ], + "type": "text", + "content": "3.2.1. 3D Gaussians Init with 2D Compositionality" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 329, + 555, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 329, + 555, + 448 + ], + "spans": [ + { + "bbox": [ + 313, + 329, + 555, + 448 + ], + "type": "text", + "content": "Unlike the implicit NeRF [48] representations, explicit 3D Gaussians can be easily initialized with 3D shapes and colors, which facilitates introducing rough 3D priors to ensure 3D consistency [77]. Although existing text-to-3D or image-to-3D models can generate a single 3D object, they struggle with compositional 3D generation as mentioned in Sec. 2, which hinders the creation of compositional 3D priors needed for initialization. Considering this, we propose initializing the compositional Gaussians on an entity-by-entity basis." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 312, + 450, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 450, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 312, + 450, + 556, + 715 + ], + "type": "text", + "content": "As shown in Fig. 2 (left), given a complex prompt " + }, + { + "bbox": [ + 312, + 450, + 556, + 715 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 312, + 450, + 556, + 715 + ], + "type": "text", + "content": ", (e.g., 'an owl perches on a branch near a pinecone'), we first adopt a 2D diffusion model [3] to generate a composed image " + }, + { + "bbox": [ + 312, + 450, + 556, + 715 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 312, + 450, + 556, + 715 + ], + "type": "text", + "content": " that faithfully captures the compositionality of multiple objects. We extract each entity information in prompt " + }, + { + "bbox": [ + 312, + 450, + 556, + 715 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 312, + 450, + 556, + 715 + ], + "type": "text", + "content": " by prompting LLM to obtain " + }, + { + "bbox": [ + 312, + 450, + 556, + 715 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 312, + 450, + 556, + 715 + ], + "type": "text", + "content": " different entity-level prompts (i.e., " + }, + { + "bbox": [ + 312, + 450, + 556, + 715 + ], + "type": "inline_equation", + "content": "L = 3" + }, + { + "bbox": [ + 312, + 450, + 556, + 715 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 312, + 450, + 556, + 715 + ], + "type": "inline_equation", + "content": "v_{1}" + }, + { + "bbox": [ + 312, + 450, + 556, + 715 + ], + "type": "text", + "content": " 'an owl', " + }, + { + "bbox": [ + 312, + 450, + 556, + 715 + ], + "type": "inline_equation", + "content": "v_{2}" + }, + { + "bbox": [ + 312, + 450, + 556, + 715 + ], + "type": "text", + "content": " 'a branch', and " + }, + { + "bbox": [ + 312, + 450, + 556, + 715 + ], + "type": "inline_equation", + "content": "v_{3}" + }, + { + "bbox": [ + 312, + 450, + 556, + 715 + ], + "type": "text", + "content": " 'a pinecone'). These prompts are adopted in a text-guided segmentation model [34] to decompose image " + }, + { + "bbox": [ + 312, + 450, + 556, + 715 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 312, + 450, + 556, + 715 + ], + "type": "text", + "content": " into various parts " + }, + { + "bbox": [ + 312, + 450, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\{I_l\\}" + }, + { + "bbox": [ + 312, + 450, + 556, + 715 + ], + "type": "text", + "content": ". Until now, each segmented image has only one entity, facilitating using existing image-based 3D generation models [67] to predict a rough triangle mesh " + }, + { + "bbox": [ + 312, + 450, + 556, + 715 + ], + "type": "inline_equation", + "content": "m_{l}(l \\in L)" + }, + { + "bbox": [ + 312, + 450, + 556, + 715 + ], + "type": "text", + "content": " of the corresponding 3D entity " + }, + { + "bbox": [ + 312, + 450, + 556, + 715 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 312, + 450, + 556, + 715 + ], + "type": "text", + "content": ". To initialize 3D Gaussians " + }, + { + "bbox": [ + 312, + 450, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\theta_{l}" + }, + { + "bbox": [ + 312, + 450, + 556, + 715 + ], + "type": "text", + "content": ", we index " + }, + { + "bbox": [ + 312, + 450, + 556, + 715 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 312, + 450, + 556, + 715 + ], + "type": "text", + "content": " points from each mesh; the center positions " + }, + { + "bbox": [ + 312, + 450, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mu_i^l \\in \\mathbb{R}^3 (i \\in N, l \\in L)" + }, + { + "bbox": [ + 312, + 450, + 556, + 715 + ], + "type": "text", + "content": " are the centers of each vertex of " + }, + { + "bbox": [ + 312, + 450, + 556, + 715 + ], + "type": "inline_equation", + "content": "m_{l}" + }, + { + "bbox": [ + 312, + 450, + 556, + 715 + ], + "type": "text", + "content": ", and the texture colors " + }, + { + "bbox": [ + 312, + 450, + 556, + 715 + ], + "type": "inline_equation", + "content": "c_i^l \\in \\mathbb{R}^3 (i \\in N, l \\in L)" + }, + { + "bbox": [ + 312, + 450, + 556, + 715 + ], + "type": "text", + "content": " are queried from each vertex of " + }, + { + "bbox": [ + 312, + 450, + 556, + 715 + ], + "type": "inline_equation", + "content": "m_{l}" + }, + { + "bbox": [ + 312, + 450, + 556, + 715 + ], + "type": "text", + "content": ". During the image-to-3D process, we did not perform any cropping or padding on the image " + }, + { + "bbox": [ + 312, + 450, + 556, + 715 + ], + "type": "inline_equation", + "content": "I_{l}" + }, + { + "bbox": [ + 312, + 450, + 556, + 715 + ], + "type": "text", + "content": ", ensuring that the spatial positions of each mesh " + }, + { + "bbox": [ + 312, + 450, + 556, + 715 + ], + "type": "inline_equation", + "content": "m_{l}" + }, + { + "bbox": [ + 312, + 450, + 556, + 715 + ], + "type": "text", + "content": " correspond to their 2D spatial positions. Additionally, we have marked the 3D layouts of each entity asset with 3D bounding boxes bbox " + }, + { + "bbox": [ + 312, + 450, + 556, + 715 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 312, + 450, + 556, + 715 + ], + "type": "text", + "content": " for subsequent optimization. The" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 71, + 72, + 443, + 83 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 72, + 443, + 83 + ], + "spans": [ + { + "bbox": [ + 71, + 72, + 443, + 83 + ], + "type": "inline_equation", + "content": "V = \\text{an owl perches on a branch near a pinecone}^{\\prime}" + }, + { + "bbox": [ + 71, + 72, + 443, + 83 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 71, + 72, + 443, + 83 + ], + "type": "inline_equation", + "content": "v_{1} = \\text{an owl}; v_{2} = \\text{a branch}; v_{3} = \\text{a pinecone}^{\\prime}" + }, + { + "bbox": [ + 71, + 72, + 443, + 83 + ], + "type": "text", + "content": ")" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18512" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 97 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 97 + ], + "type": "text", + "content": "bounding box coordinates are determined by the outer-most center positions of the entity Gaussian." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 104, + 203, + 117 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 104, + 203, + 117 + ], + "spans": [ + { + "bbox": [ + 55, + 104, + 203, + 117 + ], + "type": "text", + "content": "3.2.2. Dynamic SDS Optimization" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 121, + 297, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 121, + 297, + 216 + ], + "spans": [ + { + "bbox": [ + 55, + 121, + 297, + 216 + ], + "type": "text", + "content": "As mentioned in Sec. 2, existing SDS methods [10, 52, 58, 71] struggle in optimizing compositional 3D scenes due to the challenges of using implicit diffusion priors to maintain multi-objects consistency and interactions. To address this, we introduce dynamic SDS optimization to enable existing SDS losses to remain effective in compositional generation. Dynamic SDS optimization is performed through the following two procedures." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 218, + 296, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 218, + 296, + 350 + ], + "spans": [ + { + "bbox": [ + 55, + 218, + 296, + 350 + ], + "type": "text", + "content": "Automatically Decomposing Optimization to Different Levels. We adopt the Decomposed Optimization (DO) strategy that divides the entire training into " + }, + { + "bbox": [ + 55, + 218, + 296, + 350 + ], + "type": "inline_equation", + "content": "L + 1" + }, + { + "bbox": [ + 55, + 218, + 296, + 350 + ], + "type": "text", + "content": " stages, including " + }, + { + "bbox": [ + 55, + 218, + 296, + 350 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 55, + 218, + 296, + 350 + ], + "type": "text", + "content": " stages for entity-level optimization to ensure 3D consistency of each " + }, + { + "bbox": [ + 55, + 218, + 296, + 350 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 55, + 218, + 296, + 350 + ], + "type": "text", + "content": " entities, and one stage for composition-level optimization to refine inter-entity interactions. In each training iteration, we randomly render either (1) an entity Gaussian " + }, + { + "bbox": [ + 55, + 218, + 296, + 350 + ], + "type": "inline_equation", + "content": "\\theta_{l}(l \\in L)" + }, + { + "bbox": [ + 55, + 218, + 296, + 350 + ], + "type": "text", + "content": " to obtain the entity image " + }, + { + "bbox": [ + 55, + 218, + 296, + 350 + ], + "type": "inline_equation", + "content": "\\mathrm{x}_l = g(\\theta_l)" + }, + { + "bbox": [ + 55, + 218, + 296, + 350 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 55, + 218, + 296, + 350 + ], + "type": "inline_equation", + "content": "\\theta_{l} \\in \\mathrm{bbox}_{l}" + }, + { + "bbox": [ + 55, + 218, + 296, + 350 + ], + "type": "text", + "content": "), or (2) the composed Gaussians " + }, + { + "bbox": [ + 55, + 218, + 296, + 350 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 55, + 218, + 296, + 350 + ], + "type": "text", + "content": " to obtain the composed image " + }, + { + "bbox": [ + 55, + 218, + 296, + 350 + ], + "type": "inline_equation", + "content": "\\mathrm{x} = g(\\theta)" + }, + { + "bbox": [ + 55, + 218, + 296, + 350 + ], + "type": "text", + "content": ". Here " + }, + { + "bbox": [ + 55, + 218, + 296, + 350 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 55, + 218, + 296, + 350 + ], + "type": "text", + "content": " denotes the Gaussian Splitting rendering [33] described in Eq. 1." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 351, + 297, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 351, + 297, + 518 + ], + "spans": [ + { + "bbox": [ + 55, + 351, + 297, + 518 + ], + "type": "text", + "content": "To update Gaussian parameters, we apply Eq. 3 for gradients backpropagation. Specifically, for entity-level optimization, we substitute the text embedding " + }, + { + "bbox": [ + 55, + 351, + 297, + 518 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 55, + 351, + 297, + 518 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 55, + 351, + 297, + 518 + ], + "type": "inline_equation", + "content": "v_{l}" + }, + { + "bbox": [ + 55, + 351, + 297, + 518 + ], + "type": "text", + "content": " and parameters " + }, + { + "bbox": [ + 55, + 351, + 297, + 518 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 55, + 351, + 297, + 518 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 55, + 351, + 297, + 518 + ], + "type": "inline_equation", + "content": "\\theta_{l}" + }, + { + "bbox": [ + 55, + 351, + 297, + 518 + ], + "type": "text", + "content": " in Eq. 3 to optimize entity Gaussian. We adopt MV-Dream [58] as " + }, + { + "bbox": [ + 55, + 351, + 297, + 518 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 55, + 351, + 297, + 518 + ], + "type": "text", + "content": ", which provides multi-view diffusion priors that can better maintain 3D consistency. To ensure that other entity's parameters remain unchanged, we mask the gradients that are not within the corresponding 3D bounding boxes " + }, + { + "bbox": [ + 55, + 351, + 297, + 518 + ], + "type": "inline_equation", + "content": "\\mathrm{bbox}_l" + }, + { + "bbox": [ + 55, + 351, + 297, + 518 + ], + "type": "text", + "content": " when updating " + }, + { + "bbox": [ + 55, + 351, + 297, + 518 + ], + "type": "inline_equation", + "content": "\\theta_{l}" + }, + { + "bbox": [ + 55, + 351, + 297, + 518 + ], + "type": "text", + "content": ". For composition-level optimization, we integrate both 2D [56] and 3D diffusion priors [58] to jointly optimize the overall Gaussian parameters " + }, + { + "bbox": [ + 55, + 351, + 297, + 518 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 55, + 351, + 297, + 518 + ], + "type": "text", + "content": ". The rationale is that 2D priors promote geometry exploration [53], which enhances the generation of plausible interactions between different entities." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 520, + 297, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 520, + 297, + 664 + ], + "spans": [ + { + "bbox": [ + 55, + 520, + 297, + 664 + ], + "type": "text", + "content": "Volume-adaptively Optimizing Entity-level Gaussians. Compared to the single object generation, compositional 3D generation using SDS presents additional challenges. Specifically, when objects vary in size, optimizing directly in the original 3D space often results in suboptimal generation for smaller objects, as shown the 'pinecone' in Fig. 2. To mitigate this issue, we propose a Volume-adaptive Optimization (VAO) strategy that enhances optimization across objects of varying sizes by dynamically standardizing each entity's space to a standardized volumetric scale. This approach is especially beneficial for improving the generation of fine-grained texture details in smaller entities." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 665, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 297, + 714 + ], + "type": "text", + "content": "To scale the 3D bounding box of an entity Gaussian " + }, + { + "bbox": [ + 55, + 665, + 297, + 714 + ], + "type": "inline_equation", + "content": "\\theta_{l}" + }, + { + "bbox": [ + 55, + 665, + 297, + 714 + ], + "type": "text", + "content": " with a standardized volumetric space bbox " + }, + { + "bbox": [ + 55, + 665, + 297, + 714 + ], + "type": "inline_equation", + "content": "^{std}" + }, + { + "bbox": [ + 55, + 665, + 297, + 714 + ], + "type": "text", + "content": ", we first determine the necessary transformations for the centre positions " + }, + { + "bbox": [ + 55, + 665, + 297, + 714 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 55, + 665, + 297, + 714 + ], + "type": "text", + "content": ". Specifically, we calculate the shift parameters " + }, + { + "bbox": [ + 55, + 665, + 297, + 714 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 55, + 665, + 297, + 714 + ], + "type": "text", + "content": ", and" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 72, + 441, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 441, + 84 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 441, + 84 + ], + "type": "text", + "content": "scale parameters " + }, + { + "bbox": [ + 314, + 72, + 441, + 84 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 314, + 72, + 441, + 84 + ], + "type": "text", + "content": ", as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 351, + 90, + 555, + 102 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 90, + 555, + 102 + ], + "spans": [ + { + "bbox": [ + 351, + 90, + 555, + 102 + ], + "type": "interline_equation", + "content": "\\beta = \\operatorname {M e a n} (\\mathrm {b b o x} _ {l}); \\quad \\lambda = \\mathrm {b b o x} _ {s t d} / \\mathrm {b b o x} _ {l}, \\tag {4}", + "image_path": "cfb1937cb88e510330e1a5a5f257611acf2a114d162157583dc38c4314c2004c.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 107, + 555, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 107, + 555, + 240 + ], + "spans": [ + { + "bbox": [ + 313, + 107, + 555, + 240 + ], + "type": "text", + "content": "where Mean computes the center coordinates of the given bounding box, and " + }, + { + "bbox": [ + 313, + 107, + 555, + 240 + ], + "type": "inline_equation", + "content": "\\mathrm{bbox}_{std}" + }, + { + "bbox": [ + 313, + 107, + 555, + 240 + ], + "type": "text", + "content": " is the target standardized space. Then, we zoom-in the entity Gaussian on its center positions by " + }, + { + "bbox": [ + 313, + 107, + 555, + 240 + ], + "type": "inline_equation", + "content": "\\hat{\\mu} = \\lambda (\\mu -\\beta)" + }, + { + "bbox": [ + 313, + 107, + 555, + 240 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 107, + 555, + 240 + ], + "type": "inline_equation", + "content": "\\hat{\\mu}" + }, + { + "bbox": [ + 313, + 107, + 555, + 240 + ], + "type": "text", + "content": " denotes the transformed center positions of entity Gaussian. We substitute all " + }, + { + "bbox": [ + 313, + 107, + 555, + 240 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 313, + 107, + 555, + 240 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 107, + 555, + 240 + ], + "type": "inline_equation", + "content": "\\hat{\\mu}" + }, + { + "bbox": [ + 313, + 107, + 555, + 240 + ], + "type": "text", + "content": " in Eq. 1 for SDS optimization. After each training iteration, we perform zooming-back via " + }, + { + "bbox": [ + 313, + 107, + 555, + 240 + ], + "type": "inline_equation", + "content": "\\mu = \\hat{\\mu} /\\lambda +\\beta" + }, + { + "bbox": [ + 313, + 107, + 555, + 240 + ], + "type": "text", + "content": ". The proposed transformation ensures dynamically conducting SDS optimization on each entity Gaussian at a standardized and consistent scale. Experiments in Sec. 4.3 demonstrate its effectiveness in generating high-quality 3D assets." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 243, + 535, + 255 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 243, + 535, + 255 + ], + "spans": [ + { + "bbox": [ + 313, + 243, + 535, + 255 + ], + "type": "text", + "content": "3.3. Progressive Adding Objects with COMPGS" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 260, + 556, + 451 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 260, + 556, + 451 + ], + "spans": [ + { + "bbox": [ + 313, + 260, + 556, + 451 + ], + "type": "text", + "content": "The dynamic SDS optimization enables COMPGS to explicitly control the generation of different parts in 3D scene. This capability can be utilized for progressive increasing 3D objects. Specifically, given well-trained 3D Gaussians " + }, + { + "bbox": [ + 313, + 260, + 556, + 451 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 260, + 556, + 451 + ], + "type": "text", + "content": ", we begin by rendering an image from the front view, denoted as " + }, + { + "bbox": [ + 313, + 260, + 556, + 451 + ], + "type": "inline_equation", + "content": "\\mathbf{x} = g(\\theta)" + }, + { + "bbox": [ + 313, + 260, + 556, + 451 + ], + "type": "text", + "content": ". Unlike previous works that directly add 3D objects [13], we adopt MagicBrush [78] to modify the 2D rendered image " + }, + { + "bbox": [ + 313, + 260, + 556, + 451 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 313, + 260, + 556, + 451 + ], + "type": "text", + "content": ". The edited 2D image can then be utilized to initialize new Gaussians " + }, + { + "bbox": [ + 313, + 260, + 556, + 451 + ], + "type": "inline_equation", + "content": "\\hat{\\theta}" + }, + { + "bbox": [ + 313, + 260, + 556, + 451 + ], + "type": "text", + "content": " in the original 3D space. We further train the added 3D Gaussians to finally introduce new objects into the compositional scenes. As shown in Fig. 1, given well-trained 3D Gaussians (e.g., 'an owl'), we progressively modify the 2D images to add new objects, such as the 'branch' and 'pinecone.' We then train the corresponding 3D Gaussians via dynamic SDS optimization to incorporate these new elements into the 3D scenes." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 461, + 397, + 475 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 461, + 397, + 475 + ], + "spans": [ + { + "bbox": [ + 313, + 461, + 397, + 475 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 480, + 441, + 494 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 480, + 441, + 494 + ], + "spans": [ + { + "bbox": [ + 313, + 480, + 441, + 494 + ], + "type": "text", + "content": "4.1. Experimental Settings" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 498, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 556, + 715 + ], + "type": "text", + "content": "Implementation Details. We implemented COMPGS using ThreeStudio [24]. We use DALL-E [3] to generate the well-established 2D compositionality, LangSAM [46] to segment the entity image, and TripoSR [67] to generate a preliminary 3D prior (i.e., mesh) from each entity image, respectively. During the training, we employed MVDream [58] to optimize entity-level Gaussians, and both MVDream [58] and stabilityai/stablediffusion-2-1-base [57] to optimize composition-level Gaussians. The guidance scale for diffusion models was set as 50. Timestamps were uniformly selected from 0.02 to 0.55 for the first 1,000 iterations and then adjusted to a range from 0.02 to 0.15 for subsequent iterations. We initialized the Gaussian points at " + }, + { + "bbox": [ + 313, + 498, + 556, + 715 + ], + "type": "inline_equation", + "content": "N = 10k" + }, + { + "bbox": [ + 313, + 498, + 556, + 715 + ], + "type": "text", + "content": ", and progressively increased the density to a maximum of " + }, + { + "bbox": [ + 313, + 498, + 556, + 715 + ], + "type": "inline_equation", + "content": "1000k" + }, + { + "bbox": [ + 313, + 498, + 556, + 715 + ], + "type": "text", + "content": " points. We optimized the overall compositional scenes through 10k iterations to achieve optimal results. However, it is worth noting that we empirically found that training with 5k iterations already produces" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18513" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "content": "plausible 3D assets. The learning rates are detailed in Appendix A.3. We set the camera parameters, including radius, azimuth, elevation and FoV (field of view) to be the same as [58]. Experiments were conducted on NVIDIA A100 GPUs(40G). The prompt used for decomposing are provided in Appendix A.11. Our code will be released." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 145, + 295, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 145, + 295, + 288 + ], + "spans": [ + { + "bbox": [ + 55, + 145, + 295, + 288 + ], + "type": "text", + "content": "Evaluation Metrics. In addition to user studies and qualitative comparisons, we utilize " + }, + { + "bbox": [ + 55, + 145, + 295, + 288 + ], + "type": "inline_equation", + "content": "\\mathrm{T}^3" + }, + { + "bbox": [ + 55, + 145, + 295, + 288 + ], + "type": "text", + "content": " Bench [26] to render 300 prompts in order to evaluate compositional 3D generation on two criteria: (1) the visual quality of the 3D objects and (2) the alignment between the 3D objects and the input prompt. For quality evaluation, we captured multi-focal and multiview images and sent them to text-image scoring models, CLIP [54], to obtain an average quality score of the generated 3D scene. Regarding the textual alignment, we first followed [45] to perform 3D captioning via BLIP [38] and GPT4, and then computed the recall of the original prompt within the generated caption via ROUGE-L [42]." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 299, + 268, + 312 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 299, + 268, + 312 + ], + "spans": [ + { + "bbox": [ + 55, + 299, + 268, + 312 + ], + "type": "text", + "content": "4.2. Performance Comparisons and Analysis." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 317, + 295, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 317, + 295, + 700 + ], + "spans": [ + { + "bbox": [ + 55, + 317, + 295, + 700 + ], + "type": "text", + "content": "Qualitative Evaluation. In Fig. 3, we present qualitative comparisons of compositional 3D generation. Our model, COMPGS, is compared with several open-source methods, including DreamFusion [52], Magic3D [53], Latent-NeRF [47], Fantasia3D [10], SJC [68], and ProlificDreamer [71], as well as methods specifically designed to handle intricate text prompts, VP3D [11] in Fig. 3. The results demonstrate that COMPGS generates compositional 3D objects with superior quality, greater consistency, and more plausible interactions. DreamFusion, for example, fails to generate reasonable multi-objects from compositional prompts, revealing limitations in SDS loss for optimizing multiple objects simultaneously. Magic3D and Latent-NeRF, though guided by a 3D mesh prior, struggle with generating complex and compositional 3D geometry. Despite employing advanced SDS loss, Fantasia3D, SJC, and ProlificDreamer only improve the appearance of the content but do not fundamentally address multi-object generation. Notably, VP3D employs features from compositional images and human feedback to guide compositional 3D generation. However, it often produces unexpected artifacts (e.g., redundant dots in the top two rows of Fig. 3) or less plausible interactions (e.g., strange spatial relationships in the fourth row). This is likely due to the insufficient utilization of 2D compositionality, as VP3D adopts the implicit image features to guide optimization. In contrast, COMPGS generates high-quality and composed content that strictly aligns with the given complex prompts, demonstrating its effectiveness over others. We provide more visual comparisons with both the close-source and open-source methods (including GALA3D [81] in Fig. 9, GraphDreamer [23], DreamGaussian [65] in Fig. 10, etc.) in Appendix A.5 as well as the attached video." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "type": "text", + "content": "We also demonstrate COMPGS's capabilities in generat" + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 317, + 107, + 553, + 232 + ], + "blocks": [ + { + "bbox": [ + 313, + 71, + 555, + 103 + ], + "lines": [ + { + "bbox": [ + 313, + 71, + 555, + 103 + ], + "spans": [ + { + "bbox": [ + 313, + 71, + 555, + 103 + ], + "type": "text", + "content": "Table 1. Quantitative comparisons on " + }, + { + "bbox": [ + 313, + 71, + 555, + 103 + ], + "type": "inline_equation", + "content": "\\mathrm{T}^3" + }, + { + "bbox": [ + 313, + 71, + 555, + 103 + ], + "type": "text", + "content": " Bench [26] and user studies show COMPGS outperforms feed-forward, optimization-based, and compositional generation models." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 107, + 553, + 232 + ], + "lines": [ + { + "bbox": [ + 317, + 107, + 553, + 232 + ], + "spans": [ + { + "bbox": [ + 317, + 107, + 553, + 232 + ], + "type": "table", + "html": "
MethodT3Bench (Multiple Objects)User Study Ranking Score↑
Quality↑Alignment↑Average↑
OpenLRM [29]15.225.520.4-
TripoSR [67]16.728.622.7-
DreamFusion [52]17.314.816.14.54
SJC [68]17.75.811.73.08
LatentNeRF [47]21.719.520.64.14
Fantasia3D [10]22.714.318.52.43
ProlificDreamer [71]45.725.835.83.56
Magic3D [41]26.624.825.74.30
Set-the-Scene [15]20.829.925.4-
VP3D [11]49.131.540.36.71
COMPGS (ours)54.2 (+5.1)37.9(+6.4)46.1(+5.8)7.23
", + "image_path": "ef6b73e27fa21b844e02d6f4cd563a97b34d0fa9493c66cfdae747aa1d9435fb.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 242, + 555, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 242, + 555, + 350 + ], + "spans": [ + { + "bbox": [ + 313, + 242, + 555, + 350 + ], + "type": "text", + "content": "ing more diverse contents in Fig. 4. COMPGS accurately generates both simple spatial relationships (shown in the first row), and more complex interactions, where objects should react to each other (shown in the second row). Besides, COMPGS is not limited to generating just two objects, it can generate multiple objects by expanding the number of entity-level Gaussians. For example, in the third row, COMPGS generates scenes with three and four entities, each with high-quality 3D shapes and textures." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 352, + 556, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 352, + 556, + 495 + ], + "spans": [ + { + "bbox": [ + 313, + 352, + 556, + 495 + ], + "type": "text", + "content": "We conduct a user study for further evaluation. We randomly selected 15 prompts in " + }, + { + "bbox": [ + 313, + 352, + 556, + 495 + ], + "type": "inline_equation", + "content": "\\mathrm{T}^3" + }, + { + "bbox": [ + 313, + 352, + 556, + 495 + ], + "type": "text", + "content": " Bench, and collected the 3D assets generated by different models. These collected 3D assets were then distributed to individuals for ranking the models based on (1) 3D visual quality and (3) text-3D alignment. We average the ranking of different models as their scores. Results in Tab. 1 show that representative optimization-based models received relatively low average scores, highlighting their limitations in compositional generation. VP3D [11] outperforms its predecessors due to the incorporation of image features but still ranks lower than COMPGS, which further validates our effectiveness." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 498, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 556, + 713 + ], + "type": "text", + "content": "Quantitative Evaluation. In Tab. 1, we benchmark the representative models on " + }, + { + "bbox": [ + 313, + 498, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{T}^3" + }, + { + "bbox": [ + 313, + 498, + 556, + 713 + ], + "type": "text", + "content": " Bench, focusing on the multi-objects track. Results indicate COMPGS achieves superior performance in both quality and textual alignment. Compared with feed-forward methods in the first block, COMPGS significantly improves the generation quality by a large margin. Though COMPGS is initialized from the 3D priors obtained from [67], COMPGS still enhances text-3D alignment, likely due to the dynamic optimization on composition-level Gaussians. Furthermore, compared with the optimization-based methods, COMPGS shows significant advantages in both two metrics. Even when compared to methods specifically designed for compositional generation, e.g., VP3D [11] and Set-the-Scene [15], COMPGS demonstrates clear improvements, indicating the effectiveness of our designs. Note that COMPGS is not limited to compositional generation; the results in the track of single-object generation are included in Appendix A.5." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18514" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 217, + 69, + 394, + 80 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 217, + 69, + 394, + 80 + ], + "spans": [ + { + "bbox": [ + 217, + 69, + 394, + 80 + ], + "type": "text", + "content": "'A florist is making a bouquet with fresh flowers'" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 231, + 150, + 378, + 160 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 231, + 150, + 378, + 160 + ], + "spans": [ + { + "bbox": [ + 231, + 150, + 378, + 160 + ], + "type": "text", + "content": "'An artist is painting on a blank canvas'" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 197, + 231, + 413, + 242 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 197, + 231, + 413, + 242 + ], + "spans": [ + { + "bbox": [ + 197, + 231, + 413, + 242 + ], + "type": "text", + "content": "Hot popcorn jump out from the red striped popcorn maker" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 203, + 312, + 406, + 322 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 203, + 312, + 406, + 322 + ], + "spans": [ + { + "bbox": [ + 203, + 312, + 406, + 322 + ], + "type": "text", + "content": "'A half-eaten sandwich sits next to a lukewarm thermos'" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 390, + 97, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 390, + 97, + 399 + ], + "spans": [ + { + "bbox": [ + 56, + 390, + 97, + 399 + ], + "type": "text", + "content": "DreamFusion" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 390, + 159, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 390, + 159, + 399 + ], + "spans": [ + { + "bbox": [ + 130, + 390, + 159, + 399 + ], + "type": "text", + "content": "Magic3D" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 182, + 390, + 219, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 182, + 390, + 219, + 399 + ], + "spans": [ + { + "bbox": [ + 182, + 390, + 219, + 399 + ], + "type": "text", + "content": "LatentNeRF" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 246, + 390, + 280, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 246, + 390, + 280, + 399 + ], + "spans": [ + { + "bbox": [ + 246, + 390, + 280, + 399 + ], + "type": "text", + "content": "Fantasia3D" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 322, + 390, + 336, + 398 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 390, + 336, + 398 + ], + "spans": [ + { + "bbox": [ + 322, + 390, + 336, + 398 + ], + "type": "text", + "content": "SJC" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 372, + 390, + 419, + 398 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 390, + 419, + 398 + ], + "spans": [ + { + "bbox": [ + 372, + 390, + 419, + 398 + ], + "type": "text", + "content": "ProlificDreamer" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 455, + 390, + 474, + 398 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 455, + 390, + 474, + 398 + ], + "spans": [ + { + "bbox": [ + 455, + 390, + 474, + 398 + ], + "type": "text", + "content": "VP3D" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 504, + 390, + 552, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 504, + 390, + 552, + 399 + ], + "spans": [ + { + "bbox": [ + 504, + 390, + 552, + 399 + ], + "type": "text", + "content": "COMPGS (ours)" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 59, + 438, + 296, + 632 + ], + "blocks": [ + { + "bbox": [ + 55, + 404, + 555, + 437 + ], + "lines": [ + { + "bbox": [ + 55, + 404, + 555, + 437 + ], + "spans": [ + { + "bbox": [ + 55, + 404, + 555, + 437 + ], + "type": "text", + "content": "Figure 3. Qualitative comparisons between COMPGS and other text-to-3D models on " + }, + { + "bbox": [ + 55, + 404, + 555, + 437 + ], + "type": "inline_equation", + "content": "\\mathrm{T}^3" + }, + { + "bbox": [ + 55, + 404, + 555, + 437 + ], + "type": "text", + "content": "Bench (multiple objects track). Compared to others, COMPGS is better at generating highly-composed, high-quality 3D contents that strictly align with the given texts. Watch the animations by clicking them (Not all PDF readers support playing animations. Best viewed in Acrobat/Foxit Reader)." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 59, + 438, + 296, + 632 + ], + "lines": [ + { + "bbox": [ + 59, + 438, + 296, + 632 + ], + "spans": [ + { + "bbox": [ + 59, + 438, + 296, + 632 + ], + "type": "image", + "image_path": "0034243e454e8469075c79dde2ee3045ae2aec9a840090edfa6944b627f7a1ab.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 313, + 440, + 541, + 632 + ], + "blocks": [ + { + "bbox": [ + 313, + 440, + 541, + 632 + ], + "lines": [ + { + "bbox": [ + 313, + 440, + 541, + 632 + ], + "spans": [ + { + "bbox": [ + 313, + 440, + 541, + 632 + ], + "type": "image", + "image_path": "fc33d919392e937f5707680aacf60f7db395bb1c9db7c8339123e54e93255f4a.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 170, + 635, + 440, + 647 + ], + "lines": [ + { + "bbox": [ + 170, + 635, + 440, + 647 + ], + "spans": [ + { + "bbox": [ + 170, + 635, + 440, + 647 + ], + "type": "text", + "content": "Figure 4. More generated samples by COMPGS. Four views are shown." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 651, + 297, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 651, + 297, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 651, + 297, + 712 + ], + "type": "text", + "content": "Runtime Comparisons. Though COMPGS is trained with 10k steps, we observed that training the model for 5k iterations already produces high-quality content with minimal loss of texture details. We show runtime comparisons with other models in Appendix A.4. In fact, compositional" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 651, + 556, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 651, + 556, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 651, + 556, + 712 + ], + "type": "text", + "content": "3D generation (left) requires a longer training time than single-object generation (right) due to its complexity. Compositional 3D involves optimizing individual objects and ensuring the consistency of compositional scenes, making them more intricate. The runtime is generally relative to" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18515" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 69, + 47, + 289, + 186 + ], + "blocks": [ + { + "bbox": [ + 69, + 47, + 289, + 186 + ], + "lines": [ + { + "bbox": [ + 69, + 47, + 289, + 186 + ], + "spans": [ + { + "bbox": [ + 69, + 47, + 289, + 186 + ], + "type": "image", + "image_path": "7ed8b65f4d2728aec34d12a6988f4a028e98d31d371fa07d88445558537990e5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 62, + 187, + 288, + 209 + ], + "lines": [ + { + "bbox": [ + 62, + 187, + 288, + 209 + ], + "spans": [ + { + "bbox": [ + 62, + 187, + 288, + 209 + ], + "type": "text", + "content": "Figure 5. 3D Editing examples of COMPGS. More examples could be found in Appendix A.10." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 66, + 213, + 285, + 346 + ], + "blocks": [ + { + "bbox": [ + 66, + 213, + 285, + 346 + ], + "lines": [ + { + "bbox": [ + 66, + 213, + 285, + 346 + ], + "spans": [ + { + "bbox": [ + 66, + 213, + 285, + 346 + ], + "type": "image", + "image_path": "7b4b21f09381cde132f41dc3ee7e5bb78a06d9bd541f3a94a5236c24ddafa9b2.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 61, + 350, + 287, + 361 + ], + "lines": [ + { + "bbox": [ + 61, + 350, + 287, + 361 + ], + "spans": [ + { + "bbox": [ + 61, + 350, + 287, + 361 + ], + "type": "text", + "content": "Figure 6. Visual results of the ablation studies on key designs." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 373, + 297, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 373, + 297, + 564 + ], + "spans": [ + { + "bbox": [ + 55, + 373, + 297, + 564 + ], + "type": "text", + "content": "the number of objects involved. Compared to open-source compositional 3D methods such as Set-the-Scene [15], Progressive3D [14], and GraphDraemer [23], our proposed COMPGS is more efficient in training. For instance, given the prompt \"a parrot talks beside a perch and two bowls,\" Progressive3D takes approximately 250 minutes for 3D generation, while Set-the-Scene requires around 110 minutes. Since many compositional scene generation methods are not open-sourced, we also present the training steps listed in the papers for straightforward comparisons. In addition, CompGS demonstrates comparable and even superior efficiency over other methods in text-to-single object generation. For example, methods such as Magic3D [41] and Fantastic3D [10] require over five hours to optimize a single object. In contrast, CompGS can achieve the same task in approximately 30 minutes. Details are in Appendix A.4." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 574, + 296, + 681 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 574, + 296, + 681 + ], + "spans": [ + { + "bbox": [ + 55, + 574, + 296, + 681 + ], + "type": "text", + "content": "Extended Applications - Progressive 3D Adding. COMPGS allows progressive 3D adding for compositional scenes. We present 3D adding examples in Fig. 5. These examples demonstrate that by transferring edited 2D compositionality and employing dynamic SDS optimization, COMPGS can progressively incorporate new 3D entities into the original 3D scenes. For example, COMPGS can generate the 'chair', 'panda', 'hat' and 'plant' step by step as shown in Fig. 5." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 690, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 690, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 690, + 295, + 715 + ], + "type": "text", + "content": "Limitations. Detailed limitations and future explorations are provided in Appendix A.12 and Appendix A.13." + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 338, + 70, + 529, + 137 + ], + "blocks": [ + { + "bbox": [ + 353, + 59, + 514, + 69 + ], + "lines": [ + { + "bbox": [ + 353, + 59, + 514, + 69 + ], + "spans": [ + { + "bbox": [ + 353, + 59, + 514, + 69 + ], + "type": "text", + "content": "Table 2. Ablation Studies on " + }, + { + "bbox": [ + 353, + 59, + 514, + 69 + ], + "type": "inline_equation", + "content": "{\\mathrm{T}}^{3}" + }, + { + "bbox": [ + 353, + 59, + 514, + 69 + ], + "type": "text", + "content": " Bench [26]." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 338, + 70, + 529, + 137 + ], + "lines": [ + { + "bbox": [ + 338, + 70, + 529, + 137 + ], + "spans": [ + { + "bbox": [ + 338, + 70, + 529, + 137 + ], + "type": "table", + "html": "
ComponentQualityAlignmentAverage
Full Setting53.838.045.9
- w/o GS Init.22.818.720.8
- w/o DO Strategy46.835.241.0
- w/o VAO Strategy50.836.443.6
", + "image_path": "30b5a6525d2ac6ea7a84e1171a6a791afbf6819c9bcdee28c789a050c7fa167c.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 146, + 408, + 159 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 146, + 408, + 159 + ], + "spans": [ + { + "bbox": [ + 314, + 146, + 408, + 159 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 164, + 555, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 164, + 555, + 248 + ], + "spans": [ + { + "bbox": [ + 313, + 164, + 555, + 248 + ], + "type": "text", + "content": "We conduct ablation studies to validate the effectiveness of three key designs in COMPGS: Gaussian parameters initialization, decomposed optimization (DO), and volume-adaptive optimization (VAO). We randomly choose 20 prompts from " + }, + { + "bbox": [ + 313, + 164, + 555, + 248 + ], + "type": "inline_equation", + "content": "\\mathrm{T}^3" + }, + { + "bbox": [ + 313, + 164, + 555, + 248 + ], + "type": "text", + "content": "Bench for the quantitative evaluation. Meanwhile, we use Fig. 6 to visualize the effect of each component." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 249, + 555, + 345 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 249, + 555, + 345 + ], + "spans": [ + { + "bbox": [ + 313, + 249, + 555, + 345 + ], + "type": "text", + "content": "Initialization. Instead of initializing Gaussian parameters using 2D compositionality, we conduct random initialization within a predefined 3D bounding box for each entity Gaussian. A significant decrease in the quality and alignment metric is observed in Tab. 2. Besides, the visualization in Fig. 6 confirms that training from random initialization with predefined layouts may result in low-quality textures (e.g., owl's face) and missing entities (e.g., the branch)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 346, + 556, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 346, + 556, + 453 + ], + "spans": [ + { + "bbox": [ + 313, + 346, + 556, + 453 + ], + "type": "text", + "content": "Decomposed Optimization. We analyze the decomposed training by discarding the entity-level optimization. Results in Tab. 2 and Fig. 6 both indicate that removing decomposed training leads to low-fidelity generation. For example, in Fig. 6, optimizing composition-level Gaussians results in a relatively blurred generation on the owl's fur and the pinecone, validating that optimization on each single entity is necessary to effectively leverage 2D diffusion priors to guide 3D generation." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 454, + 556, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 454, + 556, + 574 + ], + "spans": [ + { + "bbox": [ + 313, + 454, + 556, + 574 + ], + "type": "text", + "content": "Volume-adaptive Optimization. To evaluate the effectiveness of volume-adaptive optimization, we choose to optimize the entity Gaussians in their original 3D space. We observed that optimizing different entities within varying sizes of 3D space slightly decreased in quality and alignment, which is noticeable for smaller objects in the compositional scene. For instance, the 'pinecone' and 'branch' trained without volume-adaptive optimization exhibited fewer fine-grained details, shown as the green leaves on the branch and the shadow on the pinecone in Fig. 6." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 314, + 585, + 388, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 585, + 388, + 597 + ], + "spans": [ + { + "bbox": [ + 314, + 585, + 388, + 597 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 605, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 556, + 713 + ], + "type": "text", + "content": "In this paper, we introduce COMPGS, a user-friendly, optimization-based framework, that achieves compositional text-to-3D generation utilizing Gaussian Splitting. Our three core designs, including initializing 3D Gaussians with 2D compositionality, decomposed optimization for either entity-level or composition-level Gaussians, and volume-adaptive optimization, enable COMPGS to generate complex interactions between multiple 3D entities. We hope this work encourages further research in compositional 3D generation." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18516" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 133 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 133 + ], + "type": "text", + "content": "Acknowledgment The work was supported by Berkeley DeepDrive. The authors thank Google Cloud Platform for providing the computational resources. The work was done during Chongjian Ge's visit to University of California, Berkeley." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 146, + 115, + 159 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 146, + 115, + 159 + ], + "spans": [ + { + "bbox": [ + 56, + 146, + 115, + 159 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 166, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 166, + 296, + 209 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 166, + 296, + 209 + ], + "spans": [ + { + "bbox": [ + 61, + 166, + 296, + 209 + ], + "type": "text", + "content": "[1] Haotian Bai, Yuanhuiyi Lyu, Lutao Jiang, Sijia Li, Haonan Lu, Xiaodong Lin, and Lin Wang. Componerf: Text-guided multi-object compositional nerf with editable 3d scene layout. arXiv preprint arXiv:2303.13843, 2023." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 212, + 296, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 212, + 296, + 266 + ], + "spans": [ + { + "bbox": [ + 61, + 212, + 296, + 266 + ], + "type": "text", + "content": "[2] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In IEEE/CVF International Conference on Computer Vision, 2021." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 269, + 296, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 269, + 296, + 323 + ], + "spans": [ + { + "bbox": [ + 61, + 269, + 296, + 323 + ], + "type": "text", + "content": "[3] James Betker, Gabriel Goh, Li Jing, Tim Brooks, Jianfeng Wang, Linjie Li, Long Ouyang, Juntang Zhuang, Joyce Lee, Yufei Guo, et al. Improving image generation with better captions. Computer Science. https://cdn.openaia.com/papers/dall-e-3.pdf, 2023." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 326, + 296, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 326, + 296, + 380 + ], + "spans": [ + { + "bbox": [ + 61, + 326, + 296, + 380 + ], + "type": "text", + "content": "[4] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 383, + 296, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 383, + 296, + 437 + ], + "spans": [ + { + "bbox": [ + 62, + 383, + 296, + 437 + ], + "type": "text", + "content": "[5] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 440, + 296, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 440, + 296, + 483 + ], + "spans": [ + { + "bbox": [ + 62, + 440, + 296, + 483 + ], + "type": "text", + "content": "[6] Hila Chefer, Yuval Alaluf, Yael Vinker, Lior Wolf, and Daniel Cohen-Or. Attend-and-excite: Attention-based semantic guidance for text-to-image diffusion models. ACM Transactions on Graphics (TOG), 2023." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 486, + 296, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 486, + 296, + 529 + ], + "spans": [ + { + "bbox": [ + 62, + 486, + 296, + 529 + ], + "type": "text", + "content": "[7] Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, et al. Pixart-alpha: Fast training of diffusion transformer for photorealistic text-to-image synthesis. 2023." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 532, + 296, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 532, + 296, + 586 + ], + "spans": [ + { + "bbox": [ + 62, + 532, + 296, + 586 + ], + "type": "text", + "content": "[8] Junsong Chen, Chongjian Ge, Enze Xie, Yue Wu, Lewei Yao, Xiaozhe Ren, Zhongdao Wang, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart- " + }, + { + "bbox": [ + 62, + 532, + 296, + 586 + ], + "type": "inline_equation", + "content": "\\backslash" + }, + { + "bbox": [ + 62, + 532, + 296, + 586 + ], + "type": "text", + "content": " sigma: Weak-to-strong training of diffusion transformer for 4k text-to-image generation. arXiv preprint arXiv:2403.04692, 2024." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 589, + 296, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 589, + 296, + 632 + ], + "spans": [ + { + "bbox": [ + 62, + 589, + 296, + 632 + ], + "type": "text", + "content": "[9] Minghao Chen, Iro Laina, and Andrea Vedaldi. Training-free layout control with cross-attention guidance. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, 2024." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 635, + 296, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 635, + 296, + 678 + ], + "spans": [ + { + "bbox": [ + 57, + 635, + 296, + 678 + ], + "type": "text", + "content": "[10] Rui Chen, Yongwei Chen, Ningxin Jiao, and Kui Jia. Fantasia3d: Disentangling geometry and appearance for high-quality text-to-3d content creation. In IEEE/CVF International Conference on Computer Vision, 2023." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 680, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 680, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 680, + 296, + 713 + ], + "type": "text", + "content": "[11] Yang Chen, Yingwei Pan, Haibo Yang, Ting Yao, and Tao Mei. Vp3d: Unleashing 2d visual prompt for text-to-3d generation. arXiv preprint arXiv:2403.17001, 2024." + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 316, + 73, + 555, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 555, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 555, + 116 + ], + "type": "text", + "content": "[12] Yongwei Chen, Tengfei Wang, Tong Wu, Xingang Pan, Kui Jia, and Ziwei Liu. Comboverse: Compositional 3d assets creation using spatially-aware diffusion guidance. arXiv preprint arXiv:2403.12409, 2024." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 118, + 555, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 118, + 555, + 162 + ], + "spans": [ + { + "bbox": [ + 317, + 118, + 555, + 162 + ], + "type": "text", + "content": "[13] Kai Cheng, Xiaoxiao Long, Kaizhi Yang, Yao Yao, Wei Yin, Yuexin Ma, Wenping Wang, and Xuejin Chen. Gaussianpro: 3d gaussian splatting with progressive propagation. arXiv preprint arXiv:2402.14650, 2024." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 164, + 555, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 164, + 555, + 207 + ], + "spans": [ + { + "bbox": [ + 317, + 164, + 555, + 207 + ], + "type": "text", + "content": "[14] Xinhua Cheng, Tianyu Yang, Jianan Wang, Yu Li, Lei Zhang, Jian Zhang, and Li Yuan. Progressive3d: Progressively local editing for text-to-3d content creation with complex semantic prompts. arXiv preprint arXiv:2310.11784, 2023." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 209, + 555, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 209, + 555, + 252 + ], + "spans": [ + { + "bbox": [ + 316, + 209, + 555, + 252 + ], + "type": "text", + "content": "[15] Dana Cohen-Bar, Elad Richardson, Gal Metzer, Raja Giryes, and Daniel Cohen-Or. Set-the-scene: Global-local training for generating controllable nerf scenes. In IEEE/CVF International Conference on Computer Vision, 2023." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 254, + 555, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 254, + 555, + 308 + ], + "spans": [ + { + "bbox": [ + 317, + 254, + 555, + 308 + ], + "type": "text", + "content": "[16] Xiaoliang Dai, Ji Hou, Chih-Yao Ma, Sam Tsai, Jialiang Wang, Rui Wang, Peizhao Zhang, Simon Vandenhende, Xiaofang Wang, Abhimanyu Dubey, et al. Emu: Enhancing image generation models using photogenic needles in a haystack. arXiv preprint arXiv:2309.15807, 2023." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 310, + 555, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 310, + 555, + 365 + ], + "spans": [ + { + "bbox": [ + 317, + 310, + 555, + 365 + ], + "type": "text", + "content": "[17] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objaverse: A universe of annotated 3d objects. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 366, + 555, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 366, + 555, + 420 + ], + "spans": [ + { + "bbox": [ + 317, + 366, + 555, + 420 + ], + "type": "text", + "content": "[18] Matt Deitke, Ruoshi Liu, Matthew Wallingford, Huong Ngo, Oscar Michel, Aditya Kusupati, Alan Fan, Christian Laforte, Vikram Voleti, Samir Yitzhak Gadre, et al. Objverse-xl: A universe of " + }, + { + "bbox": [ + 317, + 366, + 555, + 420 + ], + "type": "inline_equation", + "content": "10\\mathrm{m} + 3\\mathrm{d}" + }, + { + "bbox": [ + 317, + 366, + 555, + 420 + ], + "type": "text", + "content": " objects. Advances in Neural Information Processing Systems, 2024." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 422, + 554, + 455 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 422, + 554, + 455 + ], + "spans": [ + { + "bbox": [ + 316, + 422, + 554, + 455 + ], + "type": "text", + "content": "[19] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in Neural Information Processing Systems, 2021." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 456, + 555, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 456, + 555, + 510 + ], + "spans": [ + { + "bbox": [ + 317, + 456, + 555, + 510 + ], + "type": "text", + "content": "[20] Hanze Dong, Wei Xiong, Deepanshu Goyal, Yihan Zhang, Winnie Chow, Rui Pan, Shizhe Diao, Jipeng Zhang, Kashun Shum, and Tong Zhang. Raft: Reward ranked finetuning for generative foundation model alignment. arXiv preprint arXiv:2304.06767, 2023." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 513, + 555, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 513, + 555, + 555 + ], + "spans": [ + { + "bbox": [ + 317, + 513, + 555, + 555 + ], + "type": "text", + "content": "[21] Danny Driess, Zhiao Huang, Yunzhu Li, Russ Tedrake, and Marc Toussaint. Learning multi-object dynamics with compositional neural radiance fields. In Conference on robot learning, pages 1755-1768. PMLR, 2023." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 557, + 555, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 557, + 555, + 612 + ], + "spans": [ + { + "bbox": [ + 317, + 557, + 555, + 612 + ], + "type": "text", + "content": "[22] Weixi Feng, Xuehai He, Tsu-Jui Fu, Varun Jampani, Arjun Akula, Pradyumna Narayana, Sugato Basu, Xin Eric Wang, and William Yang Wang. Training-free structured diffusion guidance for compositional text-to-image synthesis. arXiv preprint arXiv:2212.05032, 2022." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 614, + 555, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 614, + 555, + 668 + ], + "spans": [ + { + "bbox": [ + 317, + 614, + 555, + 668 + ], + "type": "text", + "content": "[23] Gege Gao, Weiyang Liu, Anpei Chen, Andreas Geiger, and Bernhard Scholkopf. Graphdreamer: Compositional 3d scene synthesis from scene graphs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21295-21304, 2024." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 670, + 555, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 555, + 713 + ], + "type": "text", + "content": "[24] Yuan-Chen Guo, Ying-Tian Liu, Ruizhi Shao, Christian Laforte, Vikram Voleti, Guan Luo, Chia-Hao Chen, Zixin Zou, Chen Wang, Yan-Pei Cao, and Song-Hai Zhang. threestudio: A unified framework for 3d content generation." + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18517" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 76, + 73, + 296, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 73, + 296, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 73, + 296, + 95 + ], + "type": "text", + "content": "https://github.com/threestudio-project/threestudio, 2023." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 296, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 296, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 296, + 140 + ], + "type": "text", + "content": "[25] Haonan Han, Rui Yang, Huan Liao, Jiankai Xing, Zunnan Xu, Xiaoming Yu, Junwei Zha, Xiu Li, and Wanhua Li. Reparo: Compositional 3d assets generation with differentiable 3d layout alignment. arXiv preprint arXiv:2405.18525, 2024." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 141, + 296, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 141, + 296, + 184 + ], + "spans": [ + { + "bbox": [ + 56, + 141, + 296, + 184 + ], + "type": "text", + "content": "[26] Yuze He, Yushi Bai, Matthieu Lin, Wang Zhao, Yubin Hu, Jenny Sheng, Ran Yi, Juanzi Li, and Yong-Jin Liu. T3bench: Benchmarking current progress in text-to-3d generation. arXiv preprint arXiv:2310.02977, 2023." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 186, + 296, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 186, + 296, + 217 + ], + "spans": [ + { + "bbox": [ + 56, + 186, + 296, + 217 + ], + "type": "text", + "content": "[27] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. Advances in Neural Information Processing Systems, 2022." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 220, + 296, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 220, + 296, + 252 + ], + "spans": [ + { + "bbox": [ + 56, + 220, + 296, + 252 + ], + "type": "text", + "content": "[28] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 2020." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 254, + 296, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 254, + 296, + 297 + ], + "spans": [ + { + "bbox": [ + 56, + 254, + 296, + 297 + ], + "type": "text", + "content": "[29] Yicong Hong, Kai Zhang, Jiuxiang Gu, Sai Bi, Yang Zhou, Difan Liu, Feng Liu, Kalyan Sunkavalli, Trung Bui, and Hao Tan. Lrm: Large reconstruction model for single image to 3d. arXiv preprint arXiv:2311.04400, 2023." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 299, + 296, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 299, + 296, + 342 + ], + "spans": [ + { + "bbox": [ + 56, + 299, + 296, + 342 + ], + "type": "text", + "content": "[30] Kaiyi Huang, Kaiyue Sun, Enze Xie, Zhenguo Li, and Xihui Liu. T2i-compbench: A comprehensive benchmark for open-world compositional text-to-image generation. Advances in Neural Information Processing Systems, 2023." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 344, + 296, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 344, + 296, + 365 + ], + "spans": [ + { + "bbox": [ + 56, + 344, + 296, + 365 + ], + "type": "text", + "content": "[31] Heewoo Jun and Alex Nichol. Shap-e: Generating conditional 3d implicit functions. arXiv preprint arXiv:2305.02463, 2023." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 367, + 296, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 367, + 296, + 409 + ], + "spans": [ + { + "bbox": [ + 56, + 367, + 296, + 409 + ], + "type": "text", + "content": "[32] Animesh Karnewar, Andrea Vedaldi, David Novotny, and Niloy J Mitra. Holodiffusion: Training a 3d diffusion model using 2d images. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 411, + 296, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 411, + 296, + 443 + ], + "spans": [ + { + "bbox": [ + 56, + 411, + 296, + 443 + ], + "type": "text", + "content": "[33] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splattering for real-time radiance field rendering. ACM Transactions on Graphics, 2023." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 445, + 296, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 445, + 296, + 498 + ], + "spans": [ + { + "bbox": [ + 56, + 445, + 296, + 498 + ], + "type": "text", + "content": "[34] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In IEEE/CVF International Conference on Computer Vision, 2023." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 501, + 296, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 501, + 296, + 533 + ], + "spans": [ + { + "bbox": [ + 56, + 501, + 296, + 533 + ], + "type": "text", + "content": "[35] Georgios Kopanas, Julien Philip, Thomas Leimkuhler, and George Drettakis. Point-based neural rendering with per-view optimization. In Computer Graphics Forum, 2021." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 535, + 296, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 535, + 296, + 578 + ], + "spans": [ + { + "bbox": [ + 56, + 535, + 296, + 578 + ], + "type": "text", + "content": "[36] Georgios Kopanas, Thomas Leimkuhler, Gilles Rainer, Clément Jambon, and George Drettakis. Neural point catauastics for novel-view synthesis of reflections. ACM Transactions on Graphics (TOG), 2022." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 580, + 296, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 580, + 296, + 633 + ], + "spans": [ + { + "bbox": [ + 56, + 580, + 296, + 633 + ], + "type": "text", + "content": "[37] Kimin Lee, Hao Liu, Moonkyung Ryu, Olivia Watkins, Yuqing Du, Craig Boutilier, Pieter Abbeel, Mohammad Ghavamzadeh, and Shixiang Shane Gu. Aligning text-to-image models using human feedback. arXiv preprint arXiv:2302.12192, 2023." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 635, + 296, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 296, + 679 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 296, + 679 + ], + "type": "text", + "content": "[38] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International Conference on Machine Learning, 2022." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 56, + 681, + 296, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 681, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 681, + 296, + 712 + ], + "type": "text", + "content": "[39] Jiahao Li, Hao Tan, Kai Zhang, Zexiang Xu, Fujun Luan, Yinghao Xu, Yicong Hong, Kalyan Sunkavalli, Greg Shakhnarovich, and Sai Bi. Instant3d: Fast text-to-3d with" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "sparse-view generation and large reconstruction model. arXiv preprint arXiv:2311.06214, 2023." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 96, + 555, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 555, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 555, + 150 + ], + "type": "text", + "content": "[40] Runjia Li, Junlin Han, Luke Melas-Kyriazi, Chunyi Sun, Zhaochong An, Zhongrui Gui, Shuyang Sun, Philip Torr, and Tomas Jakab. Dreambeast: Distilling 3d fantastical animals with part-aware knowledge transfer. arXiv preprint arXiv:2409.08271, 2024." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 152, + 555, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 555, + 207 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 555, + 207 + ], + "type": "text", + "content": "[41] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 209, + 553, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 209, + 553, + 229 + ], + "spans": [ + { + "bbox": [ + 316, + 209, + 553, + 229 + ], + "type": "text", + "content": "[42] Chin-Yew Lin. Rouge: A package for automatic evaluation of summaries. In Text summarization branches out, 2004." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 231, + 553, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 231, + 553, + 274 + ], + "spans": [ + { + "bbox": [ + 316, + 231, + 553, + 274 + ], + "type": "text", + "content": "[43] Nan Liu, Shuang Li, Yilun Du, Antonio Torralba, and Joshua B Tenenbaum. Compositional visual generation with composable diffusion models. In European Conference on Computer Vision, 2022." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 276, + 553, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 276, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 316, + 276, + 553, + 319 + ], + "type": "text", + "content": "[44] Nan Liu, Shuang Li, Yilun Du, Antonio Torralba, and Joshua B Tenenbaum. Compositional visual generation with composable diffusion models. In European Conference on Computer Vision, 2022." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 321, + 553, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 321, + 553, + 354 + ], + "spans": [ + { + "bbox": [ + 316, + 321, + 553, + 354 + ], + "type": "text", + "content": "[45] Tiange Luo, Chris Rockwell, Honglak Lee, and Justin Johnson. Scalable 3d captioning with pretrained models. Advances in Neural Information Processing Systems, 2024." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 355, + 555, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 355, + 555, + 398 + ], + "spans": [ + { + "bbox": [ + 316, + 355, + 555, + 398 + ], + "type": "text", + "content": "[46] Luca Medeiros. Language segment-anything: Sam with text prompt. https://github.com/luca-medeiros/ lang-segment-anything, 2024. Accessed: 2024-05-21." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 399, + 553, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 399, + 553, + 443 + ], + "spans": [ + { + "bbox": [ + 316, + 399, + 553, + 443 + ], + "type": "text", + "content": "[47] Gal Metzer, Elad Richardson, Or Patashnik, Raja Giryes, and Daniel Cohen-Or. Latent-nerf for shape-guided generation of 3d shapes and textures. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 445, + 555, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 445, + 555, + 488 + ], + "spans": [ + { + "bbox": [ + 316, + 445, + 555, + 488 + ], + "type": "text", + "content": "[48] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 2021." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 490, + 555, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 490, + 555, + 532 + ], + "spans": [ + { + "bbox": [ + 316, + 490, + 555, + 532 + ], + "type": "text", + "content": "[49] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM transactions on graphics (TOG), 2022." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 534, + 555, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 534, + 555, + 578 + ], + "spans": [ + { + "bbox": [ + 316, + 534, + 555, + 578 + ], + "type": "text", + "content": "[50] Dong Huk Park, Samaneh Azadi, Xihui Liu, Trevor Darrell, and Anna Rohrbach. Benchmark for compositional text-to-image synthesis. In Advances in Neural Information Processing Systems, 2021." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 580, + 555, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 580, + 555, + 623 + ], + "spans": [ + { + "bbox": [ + 316, + 580, + 555, + 623 + ], + "type": "text", + "content": "[51] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 624, + 555, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 555, + 657 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 555, + 657 + ], + "type": "text", + "content": "[52] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 658, + 555, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 555, + 712 + ], + "type": "text", + "content": "[53] Guocheng Qian, Jinjie Mai, Abdullah Hamdi, Jian Ren, Aliaksandr Siarohin, Bing Li, Hsin-Ying Lee, Ivan Skorokhodov, Peter Wonka, Sergey Tulyakov, et al. Magic123: One image to high-quality 3d object generation using both 2d and 3d diffusion priors. arXiv preprint arXiv:2306.17843, 2023." + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18518" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 297, + 128 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 297, + 128 + ], + "type": "text", + "content": "[54] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, 2021." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 130, + 296, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 130, + 296, + 174 + ], + "spans": [ + { + "bbox": [ + 56, + 130, + 296, + 174 + ], + "type": "text", + "content": "[55] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning, 2021." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 176, + 296, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 176, + 296, + 219 + ], + "spans": [ + { + "bbox": [ + 56, + 176, + 296, + 219 + ], + "type": "text", + "content": "[56] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 222, + 296, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 222, + 296, + 266 + ], + "spans": [ + { + "bbox": [ + 56, + 222, + 296, + 266 + ], + "type": "text", + "content": "[57] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 268, + 296, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 268, + 296, + 300 + ], + "spans": [ + { + "bbox": [ + 56, + 268, + 296, + 300 + ], + "type": "text", + "content": "[58] Yichun Shi, Peng Wang, Jianglong Ye, Mai Long, Kejie Li, and Xiao Yang. Mvdream: Multi-view diffusion for 3d generation. arXiv preprint arXiv:2308.16512, 2023." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 303, + 296, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 303, + 296, + 346 + ], + "spans": [ + { + "bbox": [ + 56, + 303, + 296, + 346 + ], + "type": "text", + "content": "[59] J Ryan Shue, Eric Ryan Chan, Ryan Po, Zachary Ankner, Jiajun Wu, and Gordon Wetzstein. 3d neural field generation using triplane diffusion. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 349, + 296, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 349, + 296, + 392 + ], + "spans": [ + { + "bbox": [ + 56, + 349, + 296, + 392 + ], + "type": "text", + "content": "[60] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International Conference on Machine Learning, 2015." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 395, + 296, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 395, + 296, + 426 + ], + "spans": [ + { + "bbox": [ + 56, + 395, + 296, + 426 + ], + "type": "text", + "content": "[61] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 429, + 296, + 473 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 429, + 296, + 473 + ], + "spans": [ + { + "bbox": [ + 56, + 429, + 296, + 473 + ], + "type": "text", + "content": "[62] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 475, + 296, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 475, + 296, + 518 + ], + "spans": [ + { + "bbox": [ + 56, + 475, + 296, + 518 + ], + "type": "text", + "content": "[63] Jingxiang Sun, Bo Zhang, Ruizhi Shao, Lizhen Wang, Wen Liu, Zhenda Xie, and Yebin Liu. Dreamcraft3d: Hierarchical 3d generation with bootstrapped diffusion prior. arXiv preprint arXiv:2310.16818, 2023." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 521, + 296, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 521, + 296, + 563 + ], + "spans": [ + { + "bbox": [ + 56, + 521, + 296, + 563 + ], + "type": "text", + "content": "[64] Shaorong Sun, Shuchao Pang, Yazhou Yao, and Xiaoshui Huang. Comogen: A controllable text-to-3d multi-object generation framework. arXiv preprint arXiv:2409.00590, 2024." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 567, + 296, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 567, + 296, + 609 + ], + "spans": [ + { + "bbox": [ + 56, + 567, + 296, + 609 + ], + "type": "text", + "content": "[65] Jiaxiang Tang, Jiawei Ren, Hang Zhou, Ziwei Liu, and Gang Zeng. Dreamgaussian: Generative gaussian splatting for efficient 3d content creation. arXiv preprint arXiv:2309.16653, 2023." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 613, + 296, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 613, + 296, + 655 + ], + "spans": [ + { + "bbox": [ + 56, + 613, + 296, + 655 + ], + "type": "text", + "content": "[66] Jiaxiang Tang, Zhaoxi Chen, Xiaokang Chen, Tengfei Wang, Gang Zeng, and Ziwei Liu. Lgm: Large multi-view gaussian model for high-resolution 3d content creation. arXiv preprint arXiv:2402.05054, 2024." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 658, + 296, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 296, + 712 + ], + "type": "text", + "content": "[67] Dmitry Tochilkin, David Pankratz, Zexiang Liu, Zixuan Huang, Adam Letts, Yangguang Li, Ding Liang, Christian Laforte, Varun Jampani, and Yan-Pei Cao. Triposr: Fast 3d object reconstruction from a single image. arXiv preprint arXiv:2403.02151, 2024." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 555, + 714 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 316, + 72, + 555, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 555, + 126 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 555, + 126 + ], + "type": "text", + "content": "[68] Haochen Wang, Xiaodan Du, Jiahao Li, Raymond A Yeh, and Greg Shakhnarovich. Score jacobian chaining: Lifting pretrained 2d diffusion models for 3d generation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 129, + 555, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 555, + 184 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 555, + 184 + ], + "type": "text", + "content": "[69] Tengfei Wang, Bo Zhang, Ting Zhang, Shuyang Gu, Jianmin Bao, Tadas Baltrusaitis, Jingjing Shen, Dong Chen, Fang Wen, Qifeng Chen, et al. Rodin: A generative model for sculpting 3d digital avatars using diffusion. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 186, + 555, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 186, + 555, + 217 + ], + "spans": [ + { + "bbox": [ + 316, + 186, + 555, + 217 + ], + "type": "text", + "content": "[70] Zhaoning Wang, Ming Li, and Chen Chen. Luciddreaming: Controllable object-centric 3d generation. arXiv preprint arXiv:2312.00588, 2023." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 220, + 555, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 220, + 555, + 273 + ], + "spans": [ + { + "bbox": [ + 316, + 220, + 555, + 273 + ], + "type": "text", + "content": "[71] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. Advances in Neural Information Processing Systems, 2024." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 276, + 555, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 276, + 555, + 330 + ], + "spans": [ + { + "bbox": [ + 316, + 276, + 555, + 330 + ], + "type": "text", + "content": "[72] Qiucheng Wu, Yujuan Liu, Handong Zhao, Trung Bui, Zhe Lin, Yang Zhang, and Shiyu Chang. Harnessing the spatial-temporal attention of diffusion models for high-fidelity text-to-image synthesis. In IEEE/CVF International Conference on Computer Vision, 2023." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 332, + 555, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 332, + 555, + 385 + ], + "spans": [ + { + "bbox": [ + 316, + 332, + 555, + 385 + ], + "type": "text", + "content": "[73] Yinghao Xu, Zifan Shi, Wang Yifan, Hansheng Chen, Ceyuan Yang, Sida Peng, Yujun Shen, and Gordon Wetzstein. Grm: Large gaussian reconstruction model for efficient 3d reconstruction and generation. arXiv preprint arXiv:2403.14621, 2024." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 388, + 555, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 388, + 555, + 431 + ], + "spans": [ + { + "bbox": [ + 316, + 388, + 555, + 431 + ], + "type": "text", + "content": "[74] Han Yan, Yang Li, Zhennan Wu, Shenzhen Chen, Weixuan Sun, Taizhang Shang, Weizhe Liu, Tian Chen, Xiaqiang Dai, Chao Ma, et al. Frankenstein: Generating semantic-compositional 3d scenes in one tri-plane. 2024." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 434, + 555, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 434, + 555, + 476 + ], + "spans": [ + { + "bbox": [ + 316, + 434, + 555, + 476 + ], + "type": "text", + "content": "[75] Ling Yang, Zhaochen Yu, Chenlin Meng, Minkai Xu, Stefano Ermon, and Bin Cui. Mastering text-to-image diffusion: Recognition, planning, and generating with multimodal llms. arXiv preprint arXiv:2401.11708, 2024." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 478, + 555, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 478, + 555, + 521 + ], + "spans": [ + { + "bbox": [ + 316, + 478, + 555, + 521 + ], + "type": "text", + "content": "[76] Ling Yang, Zixiang Zhang, Junlin Han, Bohan Zeng, Runjia Li, Philip Torr, and Wentao Zhang. Semantic score distillation sampling for compositional text-to-3d generation. arXiv preprint arXiv:2410.09009, 2024." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 523, + 555, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 523, + 555, + 578 + ], + "spans": [ + { + "bbox": [ + 316, + 523, + 555, + 578 + ], + "type": "text", + "content": "[77] Taoran Yi, Jiemin Fang, Junjie Wang, Guanjun Wu, Lingxi Xie, Xiaopeng Zhang, Wenyu Liu, Qi Tian, and Xinggang Wang. Gaussiandreamer: Fast generation from text to 3d gaussians by bridging 2d and 3d diffusion models. arXiv preprint arXiv, 2023." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 580, + 555, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 580, + 555, + 623 + ], + "spans": [ + { + "bbox": [ + 316, + 580, + 555, + 623 + ], + "type": "text", + "content": "[78] Kai Zhang, Lingbo Mo, Wenhu Chen, Huan Sun, and Yu Su. Magicbrush: A manually annotated dataset for instruction-guided image editing. Advances in Neural Information Processing Systems, 2024." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 624, + 555, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 555, + 690 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 555, + 690 + ], + "type": "text", + "content": "[79] Qihang Zhang, Chaoyang Wang, Aliaksandr Siarohin, Peiye Zhuang, Yinghao Xu, Ceyuan Yang, Dahua Lin, Bolei Zhou, Sergey Tulyakov, and Hsin-Ying Lee. Towards text-guided 3d scene composition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6829-6838, 2024." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 692, + 555, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 555, + 714 + ], + "type": "text", + "content": "[80] Shu Zhang, Xinyi Yang, Yihao Feng, Can Qin, Chia-Chih Chen, Ning Yu, Zeyuan Chen, Huan Wang, Silvio Savarese," + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18519" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 162 + ], + "type": "list", + "angle": 0, + "index": 2, + "blocks": [ + { + "bbox": [ + 75, + 72, + 296, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 72, + 296, + 106 + ], + "spans": [ + { + "bbox": [ + 75, + 72, + 296, + 106 + ], + "type": "text", + "content": "Stefano Ermon, et al. Hive: Harnessing human feedback for instructional visual editing. arXiv preprint arXiv:2303.09618, 2023." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 296, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 296, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 296, + 162 + ], + "type": "text", + "content": "[81] Xiaoyu Zhou, Xingjian Ran, Yajiao Xiong, Jinlin He, Zhiwei Lin, Yongtao Wang, Deqing Sun, and Ming-Hsuan Yang. Gala3d: Towards text-to-3d complex scene generation via layout-guided generative gaussian splatting. arXiv preprint arXiv:2402.07207, 2024." + } + ] + } + ], + "index": 1 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "18520" + } + ] + } + ], + "index": 3 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Compass Control_ Multi Object Orientation Control for Text-to-Image Generation/9022e55b-ce42-4998-ae5e-db01a18e1b74_content_list.json b/2025/Compass Control_ Multi Object Orientation Control for Text-to-Image Generation/9022e55b-ce42-4998-ae5e-db01a18e1b74_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..19beabdaca8d15a4def649bcaf9e80fa142c59f8 --- /dev/null +++ b/2025/Compass Control_ Multi Object Orientation Control for Text-to-Image Generation/9022e55b-ce42-4998-ae5e-db01a18e1b74_content_list.json @@ -0,0 +1,1663 @@ +[ + { + "type": "text", + "text": "Compass Control: Multi Object Orientation Control for Text-to-Image Generation", + "text_level": 1, + "bbox": [ + 142, + 130, + 854, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Rishubh Parihar $^{1*}$", + "bbox": [ + 169, + 202, + 316, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Vaibhav Agrawal $^{2*†}$ $^{1}$ IISc Bangalore", + "bbox": [ + 330, + 202, + 488, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Sachidanand VS1 \n2IIT Hyderabad", + "bbox": [ + 503, + 202, + 643, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "R. Venkatesh Babu", + "bbox": [ + 665, + 202, + 825, + 220 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c814fe204320a3daaa9b2ef57781c1e019d52279a27c6bbc0ba8a35b3a2af27f.jpg", + "image_caption": [ + "a)", + "Stable Diffusion", + "A photo of a SUV facing left and a sedan facing right on a road." + ], + "image_footnote": [], + "bbox": [ + 171, + 289, + 297, + 383 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/52ee9b5bcc110321c82b00948eb5a9810421f953f230a1db5da6aec14ab65e82.jpg", + "image_caption": [ + "Compass Control", + "'A photo of a [ ] SUV and a [ sedan on a road..." + ], + "image_footnote": [], + "bbox": [ + 302, + 289, + 428, + 383 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/aa137149fd5a43de9c9084bb41432ebd9ae6bffd5fde6166eba88d1693e22540.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 268, + 524, + 286 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/67e13276661e762abca383cc8adc443b2f3e24b9a72cd7059887a9aa63976ae7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 632, + 268, + 653, + 285 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d4a82691815ff9a9092aa2ac781024049d10d1565a7f95da32838b4a1d6944fa.jpg", + "image_caption": [ + "'A photo of a mother walking with a pram on a snowy street, festive Christmas lights, beautiful winter evening scene'" + ], + "image_footnote": [], + "bbox": [ + 446, + 287, + 578, + 388 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0f790c5db6fab1fa2d98b01c27e0b0ead46303d7faf58e190b4fbc75264e0cd7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 580, + 287, + 709, + 388 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/cdcc98c2742b63c71ddcd7eed64023733992c98bff86a155e6f32eb3ea6d5210.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 764, + 270, + 790, + 286 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3a19bb20023145ea8b6ef1c0c35bace1fd0d3434972a0707782ff0ac1cecfc56.jpg", + "image_caption": [ + "A photo of a sofa and a teddy in a sunlit greenhouse filled with tropical plants" + ], + "image_footnote": [], + "bbox": [ + 151, + 414, + 397, + 527 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/96261237efb6e4d8087995463867ab156adbed812a08fd84969e9d98673f6e1d.jpg", + "image_caption": [ + "A photo of a sparrow and a sparrow and a hen among fallen autumn leaves carpeting the ground, soft afternoon glow around" + ], + "image_footnote": [], + "bbox": [ + 397, + 414, + 643, + 527 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0630f7c2043dd9cb060dad580bb0106f0d32322ac232802ea474de41636eb969.jpg", + "image_caption": [ + "'A photo of a sedan and a sedan and a camel and a camel and a camel in a vast desert with dust rising in the background'" + ], + "image_footnote": [], + "bbox": [ + 643, + 412, + 841, + 527 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e861c8201aa0333269d5d33ffa5f153e83aa05312459040436f78d4426900274.jpg", + "image_caption": [ + "Personalization with 3D Orientation Control", + "A photo of $V^{*}$ car in front of the leaning tower of Pisa in Italy", + "Figure 1. We present Compass Control, a method to generate multi-object scenes with orientation control from text-to-image diffusion models. Given a text prompt and an orientation of each object (shown as frustum, the colored face is the forward direction), our method generates scenes that align with both the prompt and specified orientations. Additionally, with a few $(\\approx 10)$ unposed images of a new object, our model is personalized to generate the object in target orientations." + ], + "image_footnote": [], + "bbox": [ + 151, + 575, + 841, + 684 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 784, + 326, + 800 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Existing approaches for controlling text-to-image diffusion models, while powerful, do not allow for explicit 3D object-centric control, such as precise control of object orienta", + "bbox": [ + 89, + 816, + 483, + 863 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tion. In this work, we address the problem of multi-object orientation control in text-to-image diffusion models. This enables the generation of diverse multi-object scenes with precise orientation control for each object. The key idea is to condition the diffusion model with a set of orientation-aware compass tokens, one for each object, along with text tokens. A light-weight encoder network predicts these com", + "bbox": [ + 511, + 786, + 906, + 892 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*equal contribution.", + "bbox": [ + 107, + 875, + 215, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$\\dagger$ work done during an internship at VAL, IISc", + "bbox": [ + 109, + 887, + 352, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "2791", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "pass tokens taking object orientation as the input. The model is trained on a synthetic dataset of procedurally generated scenes, each containing one or two 3D assets on a plain background. However, direct training this framework results in poor orientation control as well as leads to entanglement among objects. To mitigate this, we intervene in the generation process and constrain the cross-attention maps of each compass token to its corresponding object regions. The trained model is able to achieve precise orientation control for a) complex objects not seen during training and b) multi-object scenes with more than two objects, indicating strong generalization capabilities. Further, when combined with personalization methods, our method precisely controls the orientation of the new object in diverse contexts. Our method achieves state-of-the-art orientation control and text alignment quantified with extensive evaluations and a user study. Project website", + "bbox": [ + 89, + 90, + 480, + 347 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 354, + 222, + 369 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Imagine a visual artist aiming to create a scene featuring two cars facing each other. They prompt a text-to-image model with - 'A photo of a sedan facing right and an SUV facing left'. However, as shown in Fig. 1 (a), the generated image may not always accurately capture the intended object orientations. Moreover, relying on text prompts to control object orientation (e.g., 'facing right') is imprecise and requires iterative prompting. Can we design an alternate interface for text-to-image models that accepts target orientation angle as input along with the text prompts? Such an interface will allow for precise orientation control for each object, eliminating the need for iterative prompt adjustments and streamlining the creative process.", + "bbox": [ + 88, + 372, + 482, + 568 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Several works have been proposed to achieve finer control in text-to-image models such as changing object appearance, scene layouts or image style [5, 8, 11, 21, 22, 43, 53, 55, 63, 66]. While effective for controlling 2D attributes of the image, these methods fail to accurately control 3D attributes. More recently, several methods have been proposed to control 3D properties in text-to-image models, such as camera viewpoint [6, 31], scene lighting [13], or scene layout using 3D bounding boxes [4]. However, these approaches either require dense 3D information, like multi-view images or accurate 3D bounding boxes, or limited to simple single-object scenes (Tab. 1). In this work, we present a novel interface to condition text-to-image diffusion models to generate multi-object scenes with precise 3D orientation control, without the need for multi-view images or 3D bounding boxes.", + "bbox": [ + 89, + 569, + 482, + 810 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Text-to-image (T2I) diffusion models enable the generation of objects with specific attributes via text prompts (e.g., 'a red car'). Motivated by this, we encode the object orientation as an additional attribute in the text embedding space of the T2I model. Specifically, we introduce a special token, dubbed as compass token (c) along with each object token", + "bbox": [ + 89, + 810, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "table", + "img_path": "images/94bef5a692f6713c347e41069d4f78be1ac10ec18c02c3f40496e1b32cbb5510.jpg", + "table_caption": [ + "Table 1. Comparison with Related works. Our approach uniquely allows for object-centric orientation control and generalize to novel categories without any explicit 3D representation." + ], + "table_footnote": [], + "table_body": "
MethodCD-360 [31]LooseControl [4]Cont-3D-Words [13]ViewNeTI [6]Ours
InputCam Pose3D boxesCam PoseCam PoseOrientation
3D conditioningExplicitExplicitImplicitImplicitImplicit
Novel classesX✓X✓✓
Multiple Object Control✓✓XX✓
", + "bbox": [ + 514, + 136, + 906, + 196 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "in the prompt (e.g., 'A photo of $\\mathbf{c_1}$ SUV and $\\mathbf{c_2}$ sedan on a road'). Each compass token is predicted by a lightweight encoder model taking the prescribed orientation angle as input. This formulation preserves the original interface of the base T2I model and enables precise object-centric orientation control in multi-object scenes. We train the encoder model and fine-tune the denoising U-Net with LoRA [24] on a synthetic dataset of scenes containing one or two 3D assets placed in diverse layouts on an empty floor.", + "bbox": [ + 511, + 202, + 903, + 338 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We discover that directly injecting compass tokens with the prompt tokens leads to poor orientation control, as the compass token attends to irrelevant image regions, limiting its influence on its corresponding object (Fig. 4(a)). To address this, we propose Coupled Attention Localization (CALL) mechanism, where we constrain the cross-attention maps of the compass token and its corresponding object token within a 2D bounding box. This results in a tight coupling between the two tokens, enabling the compass token to precisely control the orientation for its corresponding object. Additionally, for multi-object scenes, this results in an appropriate binding between each compass token and its corresponding object token, leading to disentangled orientation control of individual objects (Fig. 4(b)).", + "bbox": [ + 511, + 338, + 903, + 550 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The proposed approach achieves precise orientation control for unseen objects (e.g., pram) and can generalize to scenes with more than two objects, despite being trained on one and two object scenes only (see Fig. 1). Further, given a few unposed images of a real object, we can personalize the model to control the orientation of the new object. We evaluate our method against several baselines, achieving superior performance both quantitatively and in a user study. In summary, our primary contributions are:", + "bbox": [ + 511, + 551, + 905, + 686 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. Compass Control - A method for conditioning text-to-image diffusion models on object orientation, enabling precise orientation control for individual objects in multi-object scene generation.", + "2. Coupled Attention Localization - A mechanism to restrict the influence of the input object orientation to the corresponding object, ensuring effective object-centric orientation control and object disentanglement.", + "3. Strong generalization of Compass Control for precise orientation control to unseen objects and complex multi-object scenes, though trained on simple synthetic scenes.", + "4. Personalization of Compass Control - Given a few unposed images of a real object, our method can perform orientation control of the new object in diverse contexts." + ], + "bbox": [ + 514, + 690, + 903, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2792", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 89, + 90, + 228, + 104 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Controlled generation in T2I models. Several works have been proposed to achieve fine-grained control in text-to-image diffusion models [47, 49, 51]. Recent works resort to manipulation of the text embeddings [20, 27, 42, 57, 67], or attention maps in the diffusion U-Net [1, 2, 8, 10, 15, 22, 29, 40, 43, 48, 55, 56] for controlling the generated image. Additional encoder models can be trained to condition the T2I models on a new modality such as depth, bounding boxes, or object identity [32, 63, 66]. Another set of works personalizes the diffusion model given with a few subject images [30, 50] enabling generation of the learned subject in different backgrounds. Recent work on guiding diffusion models [19, 35, 41] allows inference time control over the scene contents, allowing the control of object location, appearance, shape, and skeleton pose. However, these controls are limited to 2D.", + "bbox": [ + 89, + 109, + 483, + 349 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3D-aware image editing. Recent works leverage the rich generation capability of the T2I model to perform 3D aware editing [40, 52, 59]. Specifically, they use the input scene depth as an additional input and use it to warp the internal features of the diffusion models. This enables zero-shot geometric 3D edits such as translating or rotating an object. However, these methods are limited to the editing of a single object. Another line of work uses multi-view input images and trains an implicit 3D representation such as a radiance field in the diffusion feature space [31, 44] to perform 3D consistent editing. More recently, few works leverage 3D Gaussian splat representation along with T2I models to perform scene editing [12, 36]. However, the above methods require scene-specific training and require multi-view images of accurate depth maps as input. Another direction explores large-scale training of diffusion model on a specific dataset [25, 37, 61] allowing for 3D scene editing. However, these methods fail to generalize to in-the-wild real-world scenes outside the distribution of training datasets.", + "bbox": [ + 89, + 356, + 483, + 642 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3D control in generation. Earlier works train generative models from scratch with explicit scene controls such as 3D blobs [58], or radiance fields of individual objects [3, 26, 28, 38, 39, 62] to control the generated scene. Recent works have shown the existence of 3D properties in text-to-image diffusion models [6, 16-18, 65]. This has fuelled the research that leverages this knowledge from T2I models for 3D generation. A set of works lifts this knowledge to 3D by distilling from pre-trained T2I models [33, 46, 54, 60, 64, 68]. Another set of recent works [4, 6, 13] leverage this underlying 3D knowledge for controlling 3D properties of the generated scene with an additional conditioning mechanism. ViewNeTI [6] learns a 3D view token to control the camera view for the task of novel-view synthesis from a dataset of multi-view images. ViewNeTI is natively designed for novel view synthesis but is unable to generate objects in diverse contexts. A continuous word", + "bbox": [ + 89, + 643, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/8d77d907befbebfa7ecda7d17f7608f7593fad56482dedf4aade4aaa5e9a06e3.jpg", + "image_caption": [ + "Figure 2. Synthetic data generation. We curate 10 diverse 3D assets, and render them in diverse layouts and orientations in Blender [14]. The rendered scenes are augmented with realistic generations from Canny [7] ControlNet [66]. The final dataset consists of one and two object scenes." + ], + "image_footnote": [], + "bbox": [ + 522, + 95, + 898, + 268 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "representation is trained in [13] for 3D scene properties such as orientation and lighting on a single 3D mesh. However, both these approaches are limited to simple scenes and control only the global view angle. The closest to our work is [4] that conditions the T2I model on loose depth maps using ControlNet [66]. Loose depth maps are created using 3D object boxes and scene boundaries to enable 3D object-centric control. However, this approach relies on precise 3D boxes, making it cumbersome at inference. In contrast, our method only requires coarse 2D bounding boxes and orientation angles, offering a more user-friendly solution.", + "bbox": [ + 511, + 356, + 906, + 523 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 511, + 531, + 604, + 545 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Text-to-Image Diffusion Models. Diffusion models, when directly applied in the pixel space, are computationally expensive due to their iterative nature. To mitigate this, latent diffusion models [49] apply the diffusion process in the smaller resolution latent space of a pretrained autoencoder. Further, the generation can be conditioned on text by injecting text features into the diffusion U-Net with additional cross-attention layers. The cross-attention maps allow for precise control during generation [19, 23].", + "bbox": [ + 511, + 549, + 906, + 686 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Dataset", + "text_level": 1, + "bbox": [ + 511, + 696, + 609, + 709 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Synthetic scene generation: We curated a list of 10 diverse 3D assets from the web: ostrich, helicopter, shoe, jeep, teddy bear, lion, sedan, horse, motorbike and sofa to aid in model generalization (see Fig. 2). Our dataset consists of 1000 one-object and 7900 two-object scenes, rendered in Blender [14]. For each image, we save the 2D bounding boxes and the orientations of the objects. The objects are placed at varied locations and in varied orientations to increase layout diversity.", + "bbox": [ + 511, + 714, + 905, + 851 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Augmentations: Directly training on this dataset results in overfitting to the plain background and the black floor, as shown in ablations. To address this, we augment the", + "bbox": [ + 511, + 854, + 905, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "2793", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/ccc2c8387f2daaaab6064526d1a3df41c68dfc877b90e2930ff6a7e73c763c89.jpg", + "image_caption": [ + "Figure 3. Compass Control. Given an orientation angle $\\theta_{j}$ , we project it to a compass token with a lightweight encoder model. The compass tokens are interleaved with the text tokens (as shown in the figure) and passed through the text encoder. The outputs of the text encoder are used to condition the denoising process in the U-Net. We train $\\mathcal{P}$ and also fine-tune the U-Net using LoRA [24]." + ], + "image_footnote": [], + "bbox": [ + 124, + 89, + 447, + 345 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "dataset using ControlNet [66] to place objects in diverse backgrounds while retaining known orientations. For each rendered image, we extract its Canny [7] edge map to condition ControlNet to generate the objects in diverse contexts (for e.g., 'in a garden...', 'near a lake...', etc). This approach preserves object orientations while altering their appearance. We manually filter out the inconsistent augmented images. Further details on the dataset creation process can be found in the Suppl. Sec.H.", + "bbox": [ + 89, + 439, + 483, + 575 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Orientation convention: In this paper, we parameterize orientation with a single angle $\\theta$ , rotation around the up axis in the world coordinate system (pointing towards the sky). We define $\\theta = 0$ as a reference when the object faces exactly towards the right (e.g., sedan in Fig. 2). We parameterize with a single orientation angle, as most of our objects are land objects, and only this rotation axis results in plausible object orientations. However, our method is not limited to a single orientation angle, and we present results in Suppl. Sec.B for conditioning on three orientation angles.", + "bbox": [ + 89, + 580, + 483, + 731 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Compass Control", + "text_level": 1, + "bbox": [ + 89, + 744, + 261, + 760 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given a text prompt $\\mathcal{T}$ consisting of $N$ object names $\\{o_1,o_2,\\dots,o_N\\}$ (e.g., 'A photo of a jeep and a sedan and a horse in a garden') and their corresponding 3D orientation angles $\\{\\theta_{1},\\theta_{2},\\ldots ,\\theta_{N}\\}$ we introduce a set of compass tokens $\\{\\mathbf{c}_1,\\mathbf{c}_2,\\dots,\\mathbf{c}_{\\mathbf{N}}\\}$ to control the orientations of the respective objects. A compass token $\\mathbf{c}_{\\mathbf{n}}$ is an embedding in the input space of the text encoder. It is predicted by a lightweight MLP encoder network $\\mathcal{P}$ (see Fig. 3), which takes as input the object orientation angle $\\theta_{k}$ . The compass tokens are", + "bbox": [ + 89, + 763, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "prepended before their corresponding object tokens (e.g., 'A photo of a c1 jeep and a c2 sedan and a c3 horse in a garden'), and passed through the text encoder. The outputs of the text encoder are used to condition the denoising U-Net.", + "bbox": [ + 511, + 90, + 906, + 151 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "However, directly training the above framework on the dataset from Sec. 3.1 fails to learn accurate orientation control. We hypothesize that this is because the added compass tokens are unrestricted and can attend to irrelevant image regions, limiting their influence on their corresponding objects. This is evident in the cross-attention maps for the compass tokens, which are indeed diffused in other image regions (see Fig. 4 (a)). Furthermore, this becomes a severe issue in multi-object scenes, where a single compass token can attend to multiple objects, resulting in the entanglement between different compass tokens(see Fig. 4 (b)). Existing works have shown that cross-attention maps closely control the image layouts [23] in the generated image. Motivated by this, we design a cross-attention localization approach.", + "bbox": [ + 511, + 152, + 908, + 364 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Coupled Attention Localization (CALL)", + "text_level": 1, + "bbox": [ + 511, + 371, + 857, + 386 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our key idea is to constrain the cross-attention maps for both the compass token and the corresponding object token inside a given 2D bounding box. This enables tight association between the object and the compass tokens. Additionally, it enables explicit control over the object location during generation. Specifically, during training, we use the saved object bounding boxes $\\{b_1,\\dots b_N\\}$ to compute a set of loose square bounding boxes $\\{b_1^l,\\dots b_N^l\\}$ . The side length $a_{n}$ of the loose box $b_{n}^{l}$ is computed as $a_{n} = \\lambda * \\max(h_{b_{n}},w_{b_{n}})$ , where $h_{b_n}$ and $w_{b_n}$ are the height and width of the object box $b_{n}$ and $\\lambda > 1$ is a padding factor controlling the looseness of the box. Next, we compute a binary mask $m_{n}$ from the loose bounding box $b_{n}^{l}$ , such that m is 0 inside the box $b_{n}^{l}$ and $-\\infty$ outside. We use it to mask out the cross attentions $(\\Psi)$ of the object token $o_{n}$ and compass token $\\mathbf{c}_{\\mathbf{n}}$ as follows:", + "bbox": [ + 511, + 388, + 908, + 630 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\Psi (\\mathbf {c} _ {\\mathbf {n}}) = \\operatorname {s o f t m a x} \\left(m + \\frac {Q K (\\mathbf {c} _ {\\mathbf {n}}) ^ {T}}{\\sqrt {d _ {K}}}\\right)\n$$\n", + "text_format": "latex", + "bbox": [ + 584, + 631, + 830, + 667 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\Psi \\left(\\mathbf {o} _ {\\mathbf {n}}\\right) = \\operatorname {s o f t m a x} \\left(m + \\frac {Q K \\left(o _ {n}\\right) ^ {T}}{\\sqrt {d _ {K}}}\\right)\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 665, + 828, + 702 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where the query feature $Q$ comes from the U-Net features, and the key features $K(\\mathbf{c_n})$ and $K(\\mathbf{o_n})$ come from the respective tokens. This masking operation is performed at all the diffusion timesteps and cross-attention layers. We find that using a loose mask is highly effective, providing greater flexibility during the generation process. This attention localization mechanism is dubbed as Coupled Attention Localization, or CALL for short. Adding CALL mechanism during training and inference has key advantages for composing multi-object scenes: a) Appropriate binding of each compass token $\\mathbf{c_n}$ with its corresponding object token $o_n$ leads to disentangled orientation control of individual objects. b) Constraining the cross-attention for the object to", + "bbox": [ + 511, + 704, + 908, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "2794", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/a639508dea2ee3aa028e0955a5025f3333c966d483b0152aeb67358d45c26dfe.jpg", + "image_caption": [ + "a)", + "Figure 4. Binding the compass tokens: We visualize the averaged cross attention of the compass token(s) when training with CALL (shown on the left) and without it (shown on the right). CALL localizes the influence of the compass token at the correct regions, which (a) improves orientation control (b) disentangles orientations in multi-object scenes. In (b), $\\mathbf{c}_1$ and $\\mathbf{c}_2$ are compass tokens for car and motorbike, respectively." + ], + "image_footnote": [], + "bbox": [ + 106, + 92, + 467, + 290 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "kens $o_n$ to non-overlapping bounding boxes results in disentanglement between the objects themselves (a known issue in T2I models [10]), enabling strong generalization to complex scenes with multiple objects.", + "bbox": [ + 89, + 415, + 483, + 476 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training. We train our encoder model $\\mathcal{P}$ and fine-tune the denoising U-Net with LoRA [24] on the synthetic dataset from Sec. 3.1. The LoRA training is extremely parameter efficient and preserves the behavior of the base T2I model. We use the proposed CALL mechanism for effective learning of orientation control. However, for the effective working of CALL, the object must be generated within the loose bounding box, as the compass token's influence is restricted to this region. To this end, we first train on simple single-object scenes, to first learn the bounding box adherence for the generated object, and then train on a mix of both single and two-object scenes thereafter. We contrast this two-staged training procedure with the single-stage training at an intermediate training iteration in Fig. 5. The generated objects adhere to the bounding box better in the two-staged training compared to single-stage training. This leads to effective learning of orientation control, as we have shown in ablative experiments (Sec. 4).", + "bbox": [ + 89, + 487, + 483, + 758 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Inference. During the inference phase, Compass Control expects the text prompt containing the objects, desired orientations, and optional coarse 2D bounding boxes as input. Using loose bounding boxes during training offers a significant advantage here, as we can even spawn non-overlapping boxes heuristically, as shown in the Suppl. Sec.F. We use the text tokens and the compass tokens together to condition the diffusion model.", + "bbox": [ + 89, + 770, + 483, + 891 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/b4fd45a83a5fc1c7452d1ac451c68ae47f06c2c3f5969820f905bc2e4da161b2.jpg", + "image_caption": [ + "w/o staged training at iteration i" + ], + "image_footnote": [], + "bbox": [ + 517, + 108, + 612, + 180 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/4b19886fe092ccf748d04a39a49a0bd8b21fddfd4f73e585cac93d29041aecc8.jpg", + "image_caption": [ + "w staged training at iteration i" + ], + "image_footnote": [], + "bbox": [ + 614, + 107, + 709, + 180 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/ce7dc394ce4370c8fd7e6c950648b92a0f692d339212e83e28ca2862c58ea199.jpg", + "image_caption": [ + "w/o staged training at iteration i", + "Figure 5. Staged training results in improved adherence of objects to the bounding boxes, leading to orientation learning." + ], + "image_footnote": [], + "bbox": [ + 710, + 107, + 805, + 181 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/19b63436009884d7dbe2f3dd23680ef79d010a533dc7f4b9e22e22fb849b6066.jpg", + "image_caption": [ + "w staged trainingat iteration i" + ], + "image_footnote": [], + "bbox": [ + 807, + 107, + 901, + 181 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Personalization", + "text_level": 1, + "bbox": [ + 513, + 214, + 668, + 228 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The design of Compass Control as a conditioning mechanism preserves the original capabilities of T2I models, such as personalization. Given a few unposed images of an object $(\\approx 10)$ , we apply Dreambooth [50] with LoRA and associate a special token $\\hat{u}$ for the input object using Compass Control's fine-tuned UNet. During inference, we can generate the object in desired orientation $\\theta$ with simple prompts; e.g., 'A photo of a $\\mathbf{c}(\\theta)$ $\\hat{u}$ car on the beach,' where $\\mathbf{c}(\\theta)$ is the compass token.", + "bbox": [ + 511, + 233, + 906, + 369 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 513, + 378, + 643, + 395 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experiment setup", + "text_level": 1, + "bbox": [ + 513, + 400, + 686, + 416 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Dataset. We use the synthetic dataset from Sec. 3.1 consisting of 8900 rendered images and 6010 augmented images consisting of one and two object scenes for training our model. For quantitative evaluation, we construct a test set of 10 scene prompts having one/two object names generated from ChatGPT [9]. We use a mix of seen objects (horse, jeep, sedan, sofa, teddy, lion) and unseen objects (boat, dolphin, ship, SUV, tractor) in the prompts. For each object and prompt combination, we use a set of 10 randomly sampled orientations. The list of input text prompts and orientations is given in the Supp. Sec.L.", + "bbox": [ + 511, + 419, + 905, + 585 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details. We use Stable Diffusion v2.1 [49] as our base T2I model and use LoRA rank 4 for fine-tuning its UNet. We present additional results on Stable Diffusion-XL [45] in Suppl. Sec. D. Our encoder model $\\mathcal{P}$ is a lightweight MLP: three linear layers with ReLU. We train our model for 25,000 steps with a batch size of 4 with AdamW optimizer and a fixed learning rate of $10^{-4}$ . We keep the bounding box padding $\\lambda = 1.2$ for CALL. The training takes 24 hours on a single A6000 GPU.", + "bbox": [ + 511, + 589, + 905, + 726 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation Metrics. We evaluate for a) Text Alignment using CLIP similarly; b) %Object Generation - we evaluate the presence of the intended object using Grounding-DINO [34] and threshold on the objectness score for each object in the prompt. c) Angular Error - to evaluate the orientation consistency, we compute the Angular error between the input orientation angle $\\theta$ and the orientation of the generated object using a pretrained orientation predictor. Further details about the orientation predictor and the implementation details of the metrics are in the Suppl Sec.I.", + "bbox": [ + 511, + 729, + 906, + 881 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Baselines. As no prior method tackles our same task, we", + "bbox": [ + 511, + 885, + 905, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "2795", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/16d98db56fdb42175d98be89bbfde2c0e4bad477190b7eb3a07ee22b6b5af359.jpg", + "image_caption": [ + "\"A photo of a sedan spinning on a lazy susan in a mall\"" + ], + "image_footnote": [], + "bbox": [ + 155, + 90, + 383, + 195 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/466f125f22b75be25fec15dcc3c5168f5033660976109c54e685b642e681d63d.jpg", + "image_caption": [ + "'A photo of a pirate ship sailing in a sea with dark clouds near a port'" + ], + "image_footnote": [], + "bbox": [ + 385, + 90, + 612, + 195 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/6adbadff4797831c8c4ecd920d9ec1294f5d864629358df96ca94fc2ca50894a.jpg", + "image_caption": [ + "'A photo of a helicopter flying over a magnificent city, the sky is cloudy at sunset'" + ], + "image_footnote": [], + "bbox": [ + 612, + 90, + 841, + 195 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/86949ffae888969e88bff03d9734f5f35acb62becbeb1607710436824ec28372.jpg", + "image_caption": [ + "A photo of a jeep and a sedan in front of a serene waterfall with trees scattered around the region, and stones scattered in the water" + ], + "image_footnote": [], + "bbox": [ + 155, + 217, + 383, + 321 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/728605eb13ba2d178deeb06acb54b6608bd7a0027934c7b4e97d24cfe3377ce7.jpg", + "image_caption": [ + "A photo of a boat and a yacht floating in a river, featuring a lot of greenery, the sun is setting in the background" + ], + "image_footnote": [], + "bbox": [ + 385, + 217, + 612, + 321 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/068b79fafe5baea9de5179375ea056940852b4a9ed558b62e43d02baf66dc8a9.jpg", + "image_caption": [ + "'A man and a woman talking to each other sitting in a garden'" + ], + "image_footnote": [], + "bbox": [ + 614, + 217, + 841, + 321 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b00f6b28095813cb09ba4cc410b5b96e69d5892d894cdda34421156a0c5421bc.jpg", + "image_caption": [ + "A photo of a dog and a dog and a pram in a rustic living room", + "A photo of a sedan and a sedan and a SUV moving down a modern highway with a lush green forest on the sides at sunset", + "'A photo of a bicycle and a bicycle on a wooden bridge, a boat and a boat on the shiny waters on a sunny afternoon", + "'A photo of a shoe and a shoe and a teddy and a sofa and a teddy in a lavish living room with soft lighting and wooden flooring'", + "A photo of a SUV and a SUV and a SUV and a sedan and a sedan and a sedan at a city intersection at night, featuring tall buildings", + "Figure 6. Main Results. Compass Control generates complex scenes aligned with the text prompts and the orientations (shown as frustum, the colored face is forward direction). It generalizes well to unseen object categories - pirate ship, boat, yacht, bicycle, pram, dog, and even human. Further, it generates high-quality compositions of several objects, despite being trained only on one and two object scenes." + ], + "image_footnote": [], + "bbox": [ + 156, + 351, + 836, + 483 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "compare against methods that allow for either camera pose control or 3D object pose control in text-to-image models: a) Continuous 3D Words (Cont-3D-Words) [13]: Following their exact setup, we train a 3D word for controlling the object orientation on renderings of a single 3D asset - Sedan and its ControlNet augmentations. b) ViewNeTI [6]: In contrast to Cont-3D-Words, ViewNeTI allows for training on multiple 3D assets; for fair evaluation, we train ViewNeTI on our training dataset and condition the T2I on object orientation instead of 3D camera pose. Notably, both of these methods are limited to a global view control. Hence, we evaluate on only single-object scenes. c) LooseControl [4], allows for multi-object control by conditioning on loose depth maps (Fig. 7) formed by 3D object boxes. We use template 3D bounding boxes for each test object and place them in the scene, with random orientation and location (similar to Sec. 3.1). For a fair comparison, we use the 2D boxes corresponding to the 3D boxes in our outputs. Notably, LooseControl does not take exact orientation as input as a 180 flipped box also has the same loose depth;", + "bbox": [ + 89, + 585, + 483, + 888 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "we consider this while computing the Angular error. Further, it requires the user to provide accurate 3D bounding boxes during inference, which is cumbersome, whereas our method requires only loose 2D boxes. Additional baseline details are in Suppl. Sec.J.", + "bbox": [ + 511, + 585, + 906, + 660 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Main Results", + "text_level": 1, + "bbox": [ + 511, + 669, + 653, + 684 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Qualitative results. We present our method's results in Fig. 6. Our method is able to generate complex multi-object scenes with precise orientation control of individual objects, even though it is trained with single and two object scenes. Further, it is able to generalize well to challenging unseen objects such as humans and prams. Interestingly, there was no water-based subject in the training dataset, yet our method can achieve precise orientation control for a ship, yacht, and boat. These strong generalization capabilities of Compass Control can be attributed to effective attention constraining with CALL and diversity in the 3D assets used for training. The conditioning mechanism of Compass Control is generalized, and we present results for jointly controlling all three orientation angles, camera", + "bbox": [ + 509, + 688, + 906, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "2796", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/9212754278149d3b1978daa1833367cce4c428f8bb0913b94bdc55b13cfc7834.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Single objectText Align. ↓% Obj. Generated ↑Angular Err. ↓
ViewNeTI [6]22.120.9200.596
Cont-3D-words [13]29.880.7320.509
LooseControl [4]31.600.6560.385
Ours32.980.9680.198
Multiple objectText Align. ↓% Obj. Generated ↑Angular Err. ↓
LooseControl [4]31.730.7780.372
Ours33.930.9640.215
", + "bbox": [ + 99, + 80, + 470, + 189 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Quantitative comparison. We compute Text Alignment, using CLIP, % of correct subject generation and Angular error between the predicted and input orientations.", + "bbox": [ + 89, + 193, + 482, + 236 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "elevation, and object scale in Suppl. Sec.B & C.", + "bbox": [ + 89, + 248, + 410, + 262 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Baseline comparison. We compare our method to all three baselines on single-object scenes and additionally include multi-object scene comparison with LooseControl in Fig. 7 and Tab. 2. Cont-3D-words morphs the generated objects into a sedan shape seen during training and generates washed-out backgrounds. This results in poor text alignment and a lower percentage of intended object generation. ViewNeTI can generate better object shapes in a given orientation; however, it overfits to the black backgrounds seen during training, leading to poor text alignment. This is primarily because ViewNeTI does not accept ControlNet augmentations in its original form as it is designed for novel view synthesis. LooseControl generates realistic single-subject scenes following the given text prompt. However, in some cases, the object orientation is not followed (e.g., sofa, teddy). For a multi-object generation (Fig. 7b), LooseControl either misses the object during generation (horse in the second column and cars in the last column) or distorts the object shapes. LooseControl distorts the object to a box like appearance (columns 3 and 4) as it learns LoRA over the original depth-conditioned ControlNet, which follows the depth input closely, resulting in an inferior object generation score. Further, though we adjust for a 180-degree flip in computing Angular error for LooseControl, our method achieves significantly lower Angular error, demonstrating a strong orientation adherence.", + "bbox": [ + 89, + 263, + 483, + 656 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. User study", + "text_level": 1, + "bbox": [ + 89, + 665, + 210, + 681 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We conducted a user study with 57 participants to compare all methods on text alignment, object quality, and orientation consistency. Users rated 90 image pairs for single-object scenes and 30 for multi-object scenes, choosing the better image in each pair, sampled from our method and baselines. In total, we obtained 5130 ratings for single-object and 1710 for two-object scenes. Results (Fig. 8) show users preferred our method overall, with LooseControl scoring well in text alignment for single objects but falling short across all metrics in multi-object scenes.", + "bbox": [ + 89, + 686, + 483, + 839 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Personalization", + "text_level": 1, + "bbox": [ + 89, + 848, + 246, + 862 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We present personalization results in Fig. 9. With only 10 unposed images of an object, our method can gener", + "bbox": [ + 89, + 869, + 483, + 902 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/51728fef24abdbe76c2b183c50b23e3a170f4bf96768ae09fe5750e3c8c9885e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 519, + 92, + 903, + 378 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/df2781b37d444df1c0ae2fe625895a43f6de79f264bd379fcb89e8695f363758.jpg", + "image_caption": [ + "Figure 7. Qualitative Comparison. We compare our method against three baselines. Cont-3D-words [13] does not generate the intended object whereas View-NeTI [6] generates objects in plain backgrounds. LooseControl [4] generates realistic compositions but does not follow the input orientation well. In contrast, our method aligns with the input text prompt and follows the input orientation, while generating realistic scenes." + ], + "image_footnote": [], + "bbox": [ + 519, + 381, + 903, + 575 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/896492711fe23ef86613d8624248e230d2614f17aef752fe60d0a16e1feb20c7.jpg", + "image_caption": [ + "Figure 8. User study. We compare all methods on the three image metrics. Each bar indicates the fraction of people that preferred our result (gray) vs the baseline (other color)." + ], + "image_footnote": [], + "bbox": [ + 526, + 680, + 903, + 806 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ate the object with precise orientation control in various contexts. Furthermore, we can jointly optimize Dream", + "bbox": [ + 511, + 869, + 906, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "2797", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "booth [50] LoRA weights for two objects (e.g., a chair and a teddy bear), enabling multi-object personalization with object-centric orientation control. We compare our method with CD-360 [31] and achieve comparable performance. Unlike CD-360 [31], which requires $\\approx 100$ object images with camera pose, we require only a few unposed images, making it more convenient and user-friendly.", + "bbox": [ + 89, + 90, + 483, + 196 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c0debfd42436386ef4388565d187f9463232cb558d84d837b29ae960340f2c35.jpg", + "image_caption": [ + "Figure 9. Personalization. Given a few $(\\approx 10)$ unposed images of an object, our method can personalize the diffusion models and allow for orientation control of the new object. Notably, our method can also generate scenes with two personalized objects with precise orientation control. Additionally, we compare our method with CustomDiffusion-360 [31] that uses $\\approx 100$ posed images." + ], + "image_footnote": [], + "bbox": [ + 96, + 212, + 478, + 494 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Ablations", + "text_level": 1, + "bbox": [ + 89, + 599, + 202, + 612 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We present the results of the ablation study in Fig. 10 on generated scenes with 1, 3, and 5 objects. We focus on three key design choices and generate scenes with a variable number of objects:", + "bbox": [ + 89, + 617, + 483, + 678 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Staged training: Training Compass Control in a single stage results in poor adherence to the bounding boxes. This especially affects complex multi-object layouts, where some objects tend to leak outside their box and suppress the generation of neighboring objects. This results in generating a lesser number of objects in the scene.", + "bbox": [ + 89, + 681, + 483, + 772 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "CALL is crucial for accurate orientation control. Without it the compass token attends to irrelevant image regions, resulting in poor orientation control. Further, without CALL the object tokens entangle with each other during generation (a known issue in T2I models [10]). This results in generating a lesser number of objects in the scene.", + "bbox": [ + 89, + 775, + 483, + 867 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Augmentations: Without the ControlNet augmentations, the model overfits the training backgrounds, resulting in", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/56e2ac254e87aba415b3bf01ce1a0dc51a0b9512e0fa69f5924fb75239c335bc.jpg", + "image_caption": [ + "Figure 10. Ablation studies. We show the impact of several design choices of our approach. See Sec. 4.5 for details." + ], + "image_footnote": [], + "bbox": [ + 519, + 93, + 901, + 358 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "black backgrounds. Hence, ControlNet augmentation is necessary to generate objects in diverse contexts.", + "bbox": [ + 511, + 400, + 905, + 430 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion and Discussion", + "text_level": 1, + "bbox": [ + 511, + 436, + 761, + 452 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations. Our method struggles to control the orientation of objects that are occluded or have significant overlap.", + "bbox": [ + 511, + 459, + 905, + 489 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In these scenarios, the model either fails to generate one of the objects or mixes the attributes between objects (Fig. 11). Further, our single-angle orientation representation is too simplistic for modeling complex non-rigid objects such as humans.", + "bbox": [ + 511, + 489, + 691, + 638 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8c6868b60fe5e8347718177f95b2e181fd16618ad7437bbcfde826792fc3e003.jpg", + "image_caption": [ + "Figure 11. Failure Cases" + ], + "image_footnote": [], + "bbox": [ + 718, + 489, + 895, + 626 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we focus on presenting a generalized framework to condition text-to-image models on 3D controls that can be adapted easily for other representations.", + "bbox": [ + 511, + 641, + 905, + 685 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Conclusion. In this work, we propose a method to condition pre-trained text-to-image models with 3D object orientation while preserving its rich image-generation capabilities. We train the conditioning module on a small synthetic dataset via staged training and involving attention regularization. These modifications enable strong generalization of the model, allowing for precise orientation control for unseen object categories and individual objects in a multi-object scene. Further, it can be seamlessly integrated with personalization methods to achieve orientation control of personalized objects. This work is a testament that text-to-image diffusion models innately have some form of 3D understanding, and interesting 3D controls can be obtained with appropriate conditionings.", + "bbox": [ + 511, + 689, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "2798", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements. We thank Srinjay Sarkar, Abhijnya Bhat, and Tejan Karmali for thoroughly reviewing the manuscript. This work is supported by Meesho and PMRF by the Government of India.", + "bbox": [ + 89, + 90, + 485, + 148 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 176, + 187, + 193 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Omri Avrahami, Rinon Gal, Gal Chechik, Ohad Fried, Dani Lischinski, Arash Vahdat, and Weili Nie. Diffuhaul: A training-free method for object dragging in images. arXiv preprint arXiv:2406.01594, 2024. 3", + "[2] Omer Bar-Tal, Lior Yariv, Yaron Lipman, and Tali Dekel. Multidiffusion: Fusing diffusion paths for controlled image generation. 2023. 3", + "[3] Miguel Angel Bautista, Pengsheng Guo, Samira Abnar, Walter Talbott, Alexander Toshev, Zhuoyuan Chen, Laurent Dinh, Shuangfei Zhai, Hanlin Goh, Daniel Ulbricht, et al. Gaudi: A neural architect for immersive 3d scene generation. Advances in Neural Information Processing Systems, 35:25102-25116, 2022. 3", + "[4] Shariq Farooq Bhat, Niloy Mitra, and Peter Wonka. Loosecontrol: Lifting controlnet for generalized depth conditioning. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 2, 3, 6, 7", + "[5] Manuel Brack, Felix Friedrich, Katharia Kornmeier, Linoy Tsaban, Patrick Schramowski, Kristian Kersting, and Apolinário Passos. Ledits++: Limitless image editing using text-to-image models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8861-8870, 2024. 2", + "[6] James Burgess, Kuan-Chieh Wang, and Serena Yeung. Viewpoint textual inversion: Unleashing novel view synthesis with pretrained 2d diffusion models. arXiv preprint arXiv:2309.07986, 2023. 2, 3, 6, 7", + "[7] John Canny. A computational approach to edge detection. IEEE Transactions on pattern analysis and machine intelligence, (6):679-698, 1986. 3, 4", + "[8] Mingdeng Cao, Xintao Wang, Zhongang Qi, Ying Shan, Xiaohu Qie, and Yinqiang Zheng. Masactrl: Tuning-free mutual self-attention control for consistent image synthesis and editing. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22560-22570, 2023. 2, 3", + "[9] Chatgpt. Chatgpt. In https://chat.openai.com-chat, 2022.5", + "[10] Hila Chefer, Yuval Alaluf, Yael Vinker, Lior Wolf, and Daniel Cohen-Or. Attend-and-excite: Attention-based semantic guidance for text-to-image diffusion models. ACM Transactions on Graphics (TOG), 42(4):1-10, 2023. 3, 5, 8", + "[11] Xi Chen, Lianghua Huang, Yu Liu, Yujun Shen, Deli Zhao, and Hengshuang Zhao. Anydoor: Zero-shot object-level image customization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6593-6602, 2024. 2", + "[12] Yiwen Chen, Zilong Chen, Chi Zhang, Feng Wang, Xiaofeng Yang, Yikai Wang, Zhongang Cai, Lei Yang, Huaping Liu, and Guosheng Lin. Gaussianeditor: Swift and control" + ], + "bbox": [ + 93, + 202, + 482, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "able 3d editing with gaussian splatting. corr abs/2311.14521 (2023), 2023. 3", + "[13] Ta-Ying Cheng, Matheus Gadelha, Thibault Groueix, Matthew Fisher, Radomir Mech, Andrew Markham, and Niki Trigoni. Learning continuous 3d words for text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6753-6762, 2024. 2, 3, 6, 7", + "[14] Blender Online Community. Blender - a 3D modelling and rendering package. Blender Foundation, Stichting Blender Foundation, Amsterdam, 2018. 3", + "[15] Omer Dahary, Or Patashnik, Kfir Aberman, and Daniel Cohen-Or. Be yourself: Bounded attention for multi-subject text-to-image generation. arXiv preprint arXiv:2403.16990, 2(5), 2024. 3", + "[16] Ankit Dhiman, Manan Shah, Rishubh Parihar, Yash Bhalgat, Lokesh R Boregowda, and R Venkatesh Babu. Reflecting reality: Enabling diffusion models to produce faithful mirror reflections. arXiv preprint arXiv:2409.14677, 2024. 3", + "[17] Xiaodan Du, Nicholas Kolkin, Greg Shakhnarovich, and Anand Bhattachad. Generative models: What do they know? do they know things? let's find out! arXiv preprint arXiv:2311.17137, 2023.", + "[18] Mohamed El Banani, Amit Raj, Kevis-Kokitsi Maninis, Abhishek Kar, Yuanzhen Li, Michael Rubinstein, Deqing Sun, Leonidas Guibas, Justin Johnson, and Varun Jampani. Probing the 3d awareness of visual foundation models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21795-21806, 2024. 3", + "[19] Dave Epstein, Allan Jabri, Ben Poole, Alexei Efros, and Aleksander Holynski. Diffusion self-guidance for controllable image generation. Advances in Neural Information Processing Systems, 36:16222-16239, 2023. 3", + "[20] Songwei Ge, Taesung Park, Jun-Yan Zhu, and Jia-Bin Huang. Expressive text-to-image generation with rich text. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7545-7556, 2023. 3", + "[21] Daniel Geng and Andrew Owens. Motion guidance: Diffusion-based image editing with differentiable motion estimators. In The Twelfth International Conference on Learning Representations. 2", + "[22] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 2, 3", + "[23] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 3, 4", + "[24] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 2, 4, 5", + "[25] Allan Jabri, Sjoerd van Steenkiste, Emiel Hoogeboom, Mehdi SM Sajjadi, and Thomas Kipf. Dorsal: Diffusion for object-centric representations of scenes et. al. arXiv preprint arXiv:2306.08068, 2023. 3" + ], + "bbox": [ + 517, + 92, + 903, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "2799", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[26] Kunal Kathare, Ankit Dhiman, K Vikas Gowda, Siddharth Aravindan, Shubham Monga, Basavaraja Shanthappa Vandrotti, and Lokesh R Boregowda. Instructive3d: Editing large reconstruction models with text instructions. In Proceedings of the Winter Conference on Applications of Computer Vision (WACV), pages 3246-3256, 2025. 3", + "[27] Bahjat Kawar, Shiran Zada, Oran Lang, Omer Tov, Huiwen Chang, Tali Dekel, Inbar Mosseri, and Michal Irani. Imagc: Text-based real image editing with diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6007-6017, 2023. 3", + "[28] Hyunsu Kim, Gayoung Lee, Yunjay Choi, Jin-Hwa Kim, and Jun-Yan Zhu. 3d-aware blending with generative nerfs. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22906-22918, 2023. 3", + "[29] Yunji Kim, Jiyoung Lee, Jin-Hwa Kim, Jung-Woo Ha, and Jun-Yan Zhu. Dense text-to-image generation with attention modulation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7701-7711, 2023. 3", + "[30] Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization of text-to-image diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1931-1941, 2023. 3", + "[31] Nupur Kumari, Grace Su, Richard Zhang, Taesung Park, Eli Shechtman, and Jun-Yan Zhu. Customizing text-to-image diffusion with camera viewpoint control. arXiv preprint arXiv:2404.12333, 2024. 2, 3, 8", + "[32] Yuheng Li, Haotian Liu, Qingyang Wu, Fangzhou Mu, Jianwei Yang, Jianfeng Gao, Chunyuan Li, and Yong Jae Lee. Gligen: Open-set grounded text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22511-22521, 2023. 3", + "[33] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 300–309, 2023. 3", + "[34] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023. 5", + "[35] Grace Luo, Trevor Darrell, Oliver Wang, Dan B Goldman, and Aleksander Holynski. Readout guidance: Learning control from diffusion features. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8217-8227, 2024. 3", + "[36] Guan Luo, Tian-Xing Xu, Ying-Tian Liu, Xiao-Xiong Fan, Fang-Lue Zhang, and Song-Hai Zhang. 3d gaussian editing with a single image. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 6627-6636, 2024. 3", + "[37] Oscar Michel, Anand Bhattachad, Eli VanderBilt, Ranjay Krishna, Aniruddha Kembhavi, and Tanmay Gupta. Object 3dit: Language-guided 3d-aware image editing. Advances in Neural Information Processing Systems, 36, 2024. 3" + ], + "bbox": [ + 91, + 90, + 480, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[38] Thu Nguyen-Phuoc, Chuan Li, Lucas Theis, Christian Richardt, and Yong-Liang Yang. Hologan: Unsupervised learning of 3d representations from natural images. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7588–7597, 2019. 3", + "[39] Michael Niemeyer and Andreas Geiger. Giraffe: Representing scenes as compositional generative neural feature fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11453-11464, 2021. 3", + "[40] Karran Pandey, Paul Guerrero, Matheus Gadelha, Yannick Hold-Geoffroy, Karan Singh, and Niloy J Mitra. Diffusion handles enabling 3d edits for diffusion models by lifting activations to 3d. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7695–7704, 2024. 3", + "[41] Rishubh Parihar, Abhijnya Bhat, Abhipsa Basu, Saswat Mallick, Jogendra Nath Kundu, and R Venkatesh Babu. Balancing act: distribution-guided debiasing in diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6668-6678, 2024. 3", + "[42] Rishubh Parihar, VS Sachidanand, Sabariswaran Mani, Tejan Karmali, and R Venkatesh Babu. Precisecontrol: Enhancing text-to-image diffusion models with fine-grained attribute control. In European Conference on Computer Vision, pages 469-487. Springer, 2024. 3", + "[43] Or Patashnik, Daniel Garibi, Idan Azuri, Hadar Averbuch-Elor, and Daniel Cohen-Or. Localizing object-level shape variations with text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 23051–23061, 2023. 2, 3", + "[44] Or Patashnik, Rinon Gal, Daniel Cohen-Or, Jun-Yan Zhu, and Fernando De la Torre. Consolidating attention features for multi-view image editing. arXiv preprint arXiv:2402.14792, 2024. 3", + "[45] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis, 2023. 5", + "[46] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 3", + "[47] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 3", + "[48] Harsh Rangwani, Aishwarya Agarwal, Kuldeep Kulkarni, R Venkatesh Babu, and Srikrishna Karanam. Crafting parts for expressive object composition. arXiv preprint arXiv:2406.10197, 2024.3", + "[49] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 3, 5", + "[50] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine" + ], + "bbox": [ + 516, + 90, + 903, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "2800", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 22500-22510, 2023. 3, 5, 8", + "[51] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural information processing systems, 35:36479-36494, 2022. 3", + "[52] Rahul Sajnani, Jeroen Vanbaar, Jie Min, Kapil Katyal, and Srinath Sridhar. Geodiffuser: Geometry-based image editing with diffusion models. arXiv preprint arXiv:2404.14403, 2024.3", + "[53] Yizhi Song, Zhifei Zhang, Zhe Lin, Scott Cohen, Brian Price, Jianming Zhang, Soo Ye Kim, and Daniel Aliaga. Object-stitch: Object compositing with diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18310-18319, 2023. 2", + "[54] Alexander Vilesov, Pradyumna Chari, and Achuta Kadambi. Cg3d: Compositional generation for text-to-3d via gaussian splatting. arXiv preprint arXiv:2311.17907, 2023. 3", + "[55] Andrey Voynov, Kfir Aberman, and Daniel Cohen-Or. Sketch-guided text-to-image diffusion models. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-11, 2023. 2, 3", + "[56] Andrey Voynov, Kfir Aberman, and Daniel Cohen-Or. Sketch-guided text-to-image diffusion models. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1–11, 2023. 3", + "[57] Andrey Voynov, Qinghao Chu, Daniel Cohen-Or, and Kfir Aberman. $\\mathfrak{p}+$ : Extended textual conditioning in text-to-image generation. arXiv preprint arXiv:2303.09522, 2023. 3", + "[58] Qian Wang, Yiqun Wang, Michael Birsak, and Peter Wonka. Blobgan-3d: A spatially-disentangled 3d-aware generative model for indoor scenes. arXiv preprint arXiv:2303.14706, 2023. 3", + "[59] Ruicheng Wang, Jianfeng Xiang, Jiaolong Yang, and Xin Tong. Diffusion models are geometry critics: Single image 3d editing using pre-trained diffusion priors. arXiv preprint arXiv:2403.11503, 2024. 3", + "[60] Zhaoning Wang, Ming Li, and Chen Chen. Luciddreaming: Controllable object-centric 3d generation. arXiv preprint arXiv:2312.00588, 2023. 3", + "[61] Ziyi Wu, Yulia Rubanova, Rishabh Kabra, Drew A Hudson, Igor Gilitschenski, Yusuf Aytar, Sjoerd van Steenkiste, Kelsey R Allen, and Thomas Kipf. Neural assets: 3d-aware multi-object scene synthesis with image diffusion models. arXiv preprint arXiv:2406.09292, 2024. 3", + "[62] Yang Xue, Yuheng Li, Krishna Kumar Singh, and Yong Jae Lee. Giraffe hd: A high-resolution 3d-aware generative model. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18440-18449, 2022. 3", + "[63] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to" + ], + "bbox": [ + 91, + 92, + 483, + 902 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "image diffusion models. arXiv preprint arXiv:2308.06721, 2023.2,3", + "[64] Hong-Xing Yu, Haoyi Duan, Charles Herrmann, William T Freeman, and Jiajun Wu. Wonderworld: Interactive 3d scene generation from a single image. arXiv preprint arXiv:2406.09394, 2024. 3", + "[65] Guanqi Zhan, Chuanxia Zheng, Weidi Xie, and Andrew Zisserman. What does stable diffusion know about the 3d scene? arXiv preprint arXiv:2310.06836, 2023. 3", + "[66] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 2, 3, 4", + "[67] Yuxin Zhang, Weiming Dong, Fan Tang, Nisha Huang, Haibin Huang, Chongyang Ma, Tong-Yee Lee, Oliver Deussen, and Changsheng Xu. Prospect: Prompt spectrum for attribute-aware personalization of diffusion models. ACM Transactions on Graphics (TOG), 42(6):1-14, 2023. 3", + "[68] Shijie Zhou, Zhiwen Fan, Dejia Xu, Haoran Chang, Pradyumna Chari, Tejas Bharadwaj, Suya You, Zhangyang Wang, and Achuta Kadambi. Dreamscene360: Unconstrained text-to-3d scene generation with panoramic gaussian splatting. In European Conference on Computer Vision, pages 324–342. Springer, 2025. 3" + ], + "bbox": [ + 516, + 92, + 903, + 431 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "2801", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Compass Control_ Multi Object Orientation Control for Text-to-Image Generation/9022e55b-ce42-4998-ae5e-db01a18e1b74_model.json b/2025/Compass Control_ Multi Object Orientation Control for Text-to-Image Generation/9022e55b-ce42-4998-ae5e-db01a18e1b74_model.json new file mode 100644 index 0000000000000000000000000000000000000000..7c9c4016e33dfae6ac3d866c9d7c05877e7ab71f --- /dev/null +++ b/2025/Compass Control_ Multi Object Orientation Control for Text-to-Image Generation/9022e55b-ce42-4998-ae5e-db01a18e1b74_model.json @@ -0,0 +1,2708 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.143, + 0.131, + 0.856, + 0.175 + ], + "angle": 0, + "content": "Compass Control: Multi Object Orientation Control for Text-to-Image Generation" + }, + { + "type": "text", + "bbox": [ + 0.17, + 0.203, + 0.317, + 0.221 + ], + "angle": 0, + "content": "Rishubh Parihar\\(^{1*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.331, + 0.203, + 0.49, + 0.241 + ], + "angle": 0, + "content": "Vaibhav Agrawal\\(^{2*†}\\) \n\\(^{1}\\)IISc Bangalore" + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.203, + 0.645, + 0.24 + ], + "angle": 0, + "content": "Sachidanand VS1 \n2IIT Hyderabad" + }, + { + "type": "text", + "bbox": [ + 0.666, + 0.203, + 0.826, + 0.221 + ], + "angle": 0, + "content": "R. Venkatesh Babu" + }, + { + "type": "image_caption", + "bbox": [ + 0.162, + 0.277, + 0.175, + 0.287 + ], + "angle": 0, + "content": "a)" + }, + { + "type": "image_caption", + "bbox": [ + 0.195, + 0.277, + 0.281, + 0.288 + ], + "angle": 0, + "content": "Stable Diffusion" + }, + { + "type": "image", + "bbox": [ + 0.172, + 0.29, + 0.299, + 0.385 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.167, + 0.386, + 0.302, + 0.404 + ], + "angle": 0, + "content": "A photo of a SUV facing left and a sedan facing right on a road." + }, + { + "type": "image_caption", + "bbox": [ + 0.318, + 0.277, + 0.411, + 0.288 + ], + "angle": 0, + "content": "Compass Control" + }, + { + "type": "image", + "bbox": [ + 0.303, + 0.29, + 0.429, + 0.385 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.306, + 0.386, + 0.424, + 0.406 + ], + "angle": 0, + "content": "'A photo of a [ ] SUV and a [ sedan on a road..." + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.27, + 0.525, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.633, + 0.27, + 0.655, + 0.286 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.447, + 0.288, + 0.58, + 0.39 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.581, + 0.289, + 0.71, + 0.39 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.765, + 0.271, + 0.791, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.475, + 0.392, + 0.792, + 0.411 + ], + "angle": 0, + "content": "'A photo of a mother walking with a pram on a snowy street, festive Christmas lights, beautiful winter evening scene'" + }, + { + "type": "image", + "bbox": [ + 0.152, + 0.415, + 0.398, + 0.529 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.194, + 0.531, + 0.358, + 0.55 + ], + "angle": 0, + "content": "A photo of a sofa and a teddy in a sunlit greenhouse filled with tropical plants" + }, + { + "type": "image", + "bbox": [ + 0.398, + 0.415, + 0.645, + 0.529 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.411, + 0.53, + 0.633, + 0.558 + ], + "angle": 0, + "content": "A photo of a sparrow and a sparrow and a hen among fallen autumn leaves carpeting the ground, soft afternoon glow around" + }, + { + "type": "image", + "bbox": [ + 0.645, + 0.414, + 0.842, + 0.529 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.649, + 0.531, + 0.833, + 0.559 + ], + "angle": 0, + "content": "'A photo of a sedan and a sedan and a camel and a camel and a camel in a vast desert with dust rising in the background'" + }, + { + "type": "image_caption", + "bbox": [ + 0.387, + 0.564, + 0.665, + 0.575 + ], + "angle": 0, + "content": "Personalization with 3D Orientation Control" + }, + { + "type": "image", + "bbox": [ + 0.153, + 0.577, + 0.842, + 0.685 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.416, + 0.685, + 0.665, + 0.695 + ], + "angle": 0, + "content": "A photo of \\(V^{*}\\) car in front of the leaning tower of Pisa in Italy" + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.699, + 0.908, + 0.755 + ], + "angle": 0, + "content": "Figure 1. We present Compass Control, a method to generate multi-object scenes with orientation control from text-to-image diffusion models. Given a text prompt and an orientation of each object (shown as frustum, the colored face is the forward direction), our method generates scenes that align with both the prompt and specified orientations. Additionally, with a few \\((\\approx 10)\\) unposed images of a new object, our model is personalized to generate the object in target orientations." + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.785, + 0.327, + 0.801 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.818, + 0.484, + 0.864 + ], + "angle": 0, + "content": "Existing approaches for controlling text-to-image diffusion models, while powerful, do not allow for explicit 3D object-centric control, such as precise control of object orienta" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.787, + 0.907, + 0.893 + ], + "angle": 0, + "content": "tion. In this work, we address the problem of multi-object orientation control in text-to-image diffusion models. This enables the generation of diverse multi-object scenes with precise orientation control for each object. The key idea is to condition the diffusion model with a set of orientation-aware compass tokens, one for each object, along with text tokens. A light-weight encoder network predicts these com" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.876, + 0.217, + 0.888 + ], + "angle": 0, + "content": "*equal contribution." + }, + { + "type": "page_footnote", + "bbox": [ + 0.111, + 0.888, + 0.353, + 0.9 + ], + "angle": 0, + "content": "\\(\\dagger\\) work done during an internship at VAL, IISc" + }, + { + "type": "list", + "bbox": [ + 0.109, + 0.876, + 0.353, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "2791" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.348 + ], + "angle": 0, + "content": "pass tokens taking object orientation as the input. The model is trained on a synthetic dataset of procedurally generated scenes, each containing one or two 3D assets on a plain background. However, direct training this framework results in poor orientation control as well as leads to entanglement among objects. To mitigate this, we intervene in the generation process and constrain the cross-attention maps of each compass token to its corresponding object regions. The trained model is able to achieve precise orientation control for a) complex objects not seen during training and b) multi-object scenes with more than two objects, indicating strong generalization capabilities. Further, when combined with personalization methods, our method precisely controls the orientation of the new object in diverse contexts. Our method achieves state-of-the-art orientation control and text alignment quantified with extensive evaluations and a user study. Project website" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.355, + 0.223, + 0.37 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.373, + 0.483, + 0.569 + ], + "angle": 0, + "content": "Imagine a visual artist aiming to create a scene featuring two cars facing each other. They prompt a text-to-image model with - 'A photo of a sedan facing right and an SUV facing left'. However, as shown in Fig. 1 (a), the generated image may not always accurately capture the intended object orientations. Moreover, relying on text prompts to control object orientation (e.g., 'facing right') is imprecise and requires iterative prompting. Can we design an alternate interface for text-to-image models that accepts target orientation angle as input along with the text prompts? Such an interface will allow for precise orientation control for each object, eliminating the need for iterative prompt adjustments and streamlining the creative process." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.57, + 0.483, + 0.811 + ], + "angle": 0, + "content": "Several works have been proposed to achieve finer control in text-to-image models such as changing object appearance, scene layouts or image style [5, 8, 11, 21, 22, 43, 53, 55, 63, 66]. While effective for controlling 2D attributes of the image, these methods fail to accurately control 3D attributes. More recently, several methods have been proposed to control 3D properties in text-to-image models, such as camera viewpoint [6, 31], scene lighting [13], or scene layout using 3D bounding boxes [4]. However, these approaches either require dense 3D information, like multi-view images or accurate 3D bounding boxes, or limited to simple single-object scenes (Tab. 1). In this work, we present a novel interface to condition text-to-image diffusion models to generate multi-object scenes with precise 3D orientation control, without the need for multi-view images or 3D bounding boxes." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Text-to-image (T2I) diffusion models enable the generation of objects with specific attributes via text prompts (e.g., 'a red car'). Motivated by this, we encode the object orientation as an additional attribute in the text embedding space of the T2I model. Specifically, we introduce a special token, dubbed as compass token (c) along with each object token" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.09, + 0.905, + 0.132 + ], + "angle": 0, + "content": "Table 1. Comparison with Related works. Our approach uniquely allows for object-centric orientation control and generalize to novel categories without any explicit 3D representation." + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.137, + 0.907, + 0.197 + ], + "angle": 0, + "content": "
MethodCD-360 [31]LooseControl [4]Cont-3D-Words [13]ViewNeTI [6]Ours
InputCam Pose3D boxesCam PoseCam PoseOrientation
3D conditioningExplicitExplicitImplicitImplicitImplicit
Novel classesX✓X✓✓
Multiple Object Control✓✓XX✓
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.203, + 0.905, + 0.339 + ], + "angle": 0, + "content": "in the prompt (e.g., 'A photo of \\(\\mathbf{c_1}\\) SUV and \\(\\mathbf{c_2}\\) sedan on a road'). Each compass token is predicted by a lightweight encoder model taking the prescribed orientation angle as input. This formulation preserves the original interface of the base T2I model and enables precise object-centric orientation control in multi-object scenes. We train the encoder model and fine-tune the denoising U-Net with LoRA [24] on a synthetic dataset of scenes containing one or two 3D assets placed in diverse layouts on an empty floor." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.339, + 0.905, + 0.551 + ], + "angle": 0, + "content": "We discover that directly injecting compass tokens with the prompt tokens leads to poor orientation control, as the compass token attends to irrelevant image regions, limiting its influence on its corresponding object (Fig. 4(a)). To address this, we propose Coupled Attention Localization (CALL) mechanism, where we constrain the cross-attention maps of the compass token and its corresponding object token within a 2D bounding box. This results in a tight coupling between the two tokens, enabling the compass token to precisely control the orientation for its corresponding object. Additionally, for multi-object scenes, this results in an appropriate binding between each compass token and its corresponding object token, leading to disentangled orientation control of individual objects (Fig. 4(b))." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.552, + 0.906, + 0.687 + ], + "angle": 0, + "content": "The proposed approach achieves precise orientation control for unseen objects (e.g., pram) and can generalize to scenes with more than two objects, despite being trained on one and two object scenes only (see Fig. 1). Further, given a few unposed images of a real object, we can personalize the model to control the orientation of the new object. We evaluate our method against several baselines, achieving superior performance both quantitatively and in a user study. In summary, our primary contributions are:" + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.691, + 0.905, + 0.751 + ], + "angle": 0, + "content": "1. Compass Control - A method for conditioning text-to-image diffusion models on object orientation, enabling precise orientation control for individual objects in multi-object scene generation." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.751, + 0.905, + 0.811 + ], + "angle": 0, + "content": "2. Coupled Attention Localization - A mechanism to restrict the influence of the input object orientation to the corresponding object, ensuring effective object-centric orientation control and object disentanglement." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.812, + 0.905, + 0.856 + ], + "angle": 0, + "content": "3. Strong generalization of Compass Control for precise orientation control to unseen objects and complex multi-object scenes, though trained on simple synthetic scenes." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.857, + 0.905, + 0.902 + ], + "angle": 0, + "content": "4. Personalization of Compass Control - Given a few unposed images of a real object, our method can perform orientation control of the new object in diverse contexts." + }, + { + "type": "list", + "bbox": [ + 0.515, + 0.691, + 0.905, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2792" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.229, + 0.105 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.11, + 0.484, + 0.351 + ], + "angle": 0, + "content": "Controlled generation in T2I models. Several works have been proposed to achieve fine-grained control in text-to-image diffusion models [47, 49, 51]. Recent works resort to manipulation of the text embeddings [20, 27, 42, 57, 67], or attention maps in the diffusion U-Net [1, 2, 8, 10, 15, 22, 29, 40, 43, 48, 55, 56] for controlling the generated image. Additional encoder models can be trained to condition the T2I models on a new modality such as depth, bounding boxes, or object identity [32, 63, 66]. Another set of works personalizes the diffusion model given with a few subject images [30, 50] enabling generation of the learned subject in different backgrounds. Recent work on guiding diffusion models [19, 35, 41] allows inference time control over the scene contents, allowing the control of object location, appearance, shape, and skeleton pose. However, these controls are limited to 2D." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.357, + 0.484, + 0.643 + ], + "angle": 0, + "content": "3D-aware image editing. Recent works leverage the rich generation capability of the T2I model to perform 3D aware editing [40, 52, 59]. Specifically, they use the input scene depth as an additional input and use it to warp the internal features of the diffusion models. This enables zero-shot geometric 3D edits such as translating or rotating an object. However, these methods are limited to the editing of a single object. Another line of work uses multi-view input images and trains an implicit 3D representation such as a radiance field in the diffusion feature space [31, 44] to perform 3D consistent editing. More recently, few works leverage 3D Gaussian splat representation along with T2I models to perform scene editing [12, 36]. However, the above methods require scene-specific training and require multi-view images of accurate depth maps as input. Another direction explores large-scale training of diffusion model on a specific dataset [25, 37, 61] allowing for 3D scene editing. However, these methods fail to generalize to in-the-wild real-world scenes outside the distribution of training datasets." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.645, + 0.484, + 0.901 + ], + "angle": 0, + "content": "3D control in generation. Earlier works train generative models from scratch with explicit scene controls such as 3D blobs [58], or radiance fields of individual objects [3, 26, 28, 38, 39, 62] to control the generated scene. Recent works have shown the existence of 3D properties in text-to-image diffusion models [6, 16-18, 65]. This has fuelled the research that leverages this knowledge from T2I models for 3D generation. A set of works lifts this knowledge to 3D by distilling from pre-trained T2I models [33, 46, 54, 60, 64, 68]. Another set of recent works [4, 6, 13] leverage this underlying 3D knowledge for controlling 3D properties of the generated scene with an additional conditioning mechanism. ViewNeTI [6] learns a 3D view token to control the camera view for the task of novel-view synthesis from a dataset of multi-view images. ViewNeTI is natively designed for novel view synthesis but is unable to generate objects in diverse contexts. A continuous word" + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.096, + 0.9, + 0.269 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.275, + 0.907, + 0.344 + ], + "angle": 0, + "content": "Figure 2. Synthetic data generation. We curate 10 diverse 3D assets, and render them in diverse layouts and orientations in Blender [14]. The rendered scenes are augmented with realistic generations from Canny [7] ControlNet [66]. The final dataset consists of one and two object scenes." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.357, + 0.907, + 0.525 + ], + "angle": 0, + "content": "representation is trained in [13] for 3D scene properties such as orientation and lighting on a single 3D mesh. However, both these approaches are limited to simple scenes and control only the global view angle. The closest to our work is [4] that conditions the T2I model on loose depth maps using ControlNet [66]. Loose depth maps are created using 3D object boxes and scene boundaries to enable 3D object-centric control. However, this approach relies on precise 3D boxes, making it cumbersome at inference. In contrast, our method only requires coarse 2D bounding boxes and orientation angles, offering a more user-friendly solution." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.532, + 0.606, + 0.546 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.55, + 0.907, + 0.687 + ], + "angle": 0, + "content": "Text-to-Image Diffusion Models. Diffusion models, when directly applied in the pixel space, are computationally expensive due to their iterative nature. To mitigate this, latent diffusion models [49] apply the diffusion process in the smaller resolution latent space of a pretrained autoencoder. Further, the generation can be conditioned on text by injecting text features into the diffusion U-Net with additional cross-attention layers. The cross-attention maps allow for precise control during generation [19, 23]." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.697, + 0.61, + 0.71 + ], + "angle": 0, + "content": "3.1. Dataset" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.715, + 0.906, + 0.852 + ], + "angle": 0, + "content": "Synthetic scene generation: We curated a list of 10 diverse 3D assets from the web: ostrich, helicopter, shoe, jeep, teddy bear, lion, sedan, horse, motorbike and sofa to aid in model generalization (see Fig. 2). Our dataset consists of 1000 one-object and 7900 two-object scenes, rendered in Blender [14]. For each image, we save the 2D bounding boxes and the orientations of the objects. The objects are placed at varied locations and in varied orientations to increase layout diversity." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Augmentations: Directly training on this dataset results in overfitting to the plain background and the black floor, as shown in ablations. To address this, we augment the" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2793" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.125, + 0.09, + 0.449, + 0.346 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.349, + 0.485, + 0.432 + ], + "angle": 0, + "content": "Figure 3. Compass Control. Given an orientation angle \\(\\theta_{j}\\), we project it to a compass token with a lightweight encoder model. The compass tokens are interleaved with the text tokens (as shown in the figure) and passed through the text encoder. The outputs of the text encoder are used to condition the denoising process in the U-Net. We train \\(\\mathcal{P}\\) and also fine-tune the U-Net using LoRA [24]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.44, + 0.484, + 0.576 + ], + "angle": 0, + "content": "dataset using ControlNet [66] to place objects in diverse backgrounds while retaining known orientations. For each rendered image, we extract its Canny [7] edge map to condition ControlNet to generate the objects in diverse contexts (for e.g., 'in a garden...', 'near a lake...', etc). This approach preserves object orientations while altering their appearance. We manually filter out the inconsistent augmented images. Further details on the dataset creation process can be found in the Suppl. Sec.H." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.581, + 0.484, + 0.732 + ], + "angle": 0, + "content": "Orientation convention: In this paper, we parameterize orientation with a single angle \\(\\theta\\), rotation around the up axis in the world coordinate system (pointing towards the sky). We define \\(\\theta = 0\\) as a reference when the object faces exactly towards the right (e.g., sedan in Fig. 2). We parameterize with a single orientation angle, as most of our objects are land objects, and only this rotation axis results in plausible object orientations. However, our method is not limited to a single orientation angle, and we present results in Suppl. Sec.B for conditioning on three orientation angles." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.745, + 0.262, + 0.761 + ], + "angle": 0, + "content": "3.2. Compass Control" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Given a text prompt \\(\\mathcal{T}\\) consisting of \\(N\\) object names \\(\\{o_1,o_2,\\dots,o_N\\}\\) (e.g., 'A photo of a jeep and a sedan and a horse in a garden') and their corresponding 3D orientation angles \\(\\{\\theta_{1},\\theta_{2},\\ldots ,\\theta_{N}\\}\\) we introduce a set of compass tokens \\(\\{\\mathbf{c}_1,\\mathbf{c}_2,\\dots,\\mathbf{c}_{\\mathbf{N}}\\}\\) to control the orientations of the respective objects. A compass token \\(\\mathbf{c}_{\\mathbf{n}}\\) is an embedding in the input space of the text encoder. It is predicted by a lightweight MLP encoder network \\(\\mathcal{P}\\) (see Fig. 3), which takes as input the object orientation angle \\(\\theta_{k}\\). The compass tokens are" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.152 + ], + "angle": 0, + "content": "prepended before their corresponding object tokens (e.g., 'A photo of a c1 jeep and a c2 sedan and a c3 horse in a garden'), and passed through the text encoder. The outputs of the text encoder are used to condition the denoising U-Net." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.153, + 0.909, + 0.365 + ], + "angle": 0, + "content": "However, directly training the above framework on the dataset from Sec. 3.1 fails to learn accurate orientation control. We hypothesize that this is because the added compass tokens are unrestricted and can attend to irrelevant image regions, limiting their influence on their corresponding objects. This is evident in the cross-attention maps for the compass tokens, which are indeed diffused in other image regions (see Fig. 4 (a)). Furthermore, this becomes a severe issue in multi-object scenes, where a single compass token can attend to multiple objects, resulting in the entanglement between different compass tokens(see Fig. 4 (b)). Existing works have shown that cross-attention maps closely control the image layouts [23] in the generated image. Motivated by this, we design a cross-attention localization approach." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.372, + 0.859, + 0.387 + ], + "angle": 0, + "content": "3.3. Coupled Attention Localization (CALL)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.39, + 0.909, + 0.631 + ], + "angle": 0, + "content": "Our key idea is to constrain the cross-attention maps for both the compass token and the corresponding object token inside a given 2D bounding box. This enables tight association between the object and the compass tokens. Additionally, it enables explicit control over the object location during generation. Specifically, during training, we use the saved object bounding boxes \\(\\{b_1,\\dots b_N\\}\\) to compute a set of loose square bounding boxes \\(\\{b_1^l,\\dots b_N^l\\}\\). The side length \\(a_{n}\\) of the loose box \\(b_{n}^{l}\\) is computed as \\(a_{n} = \\lambda * \\max(h_{b_{n}},w_{b_{n}})\\), where \\(h_{b_n}\\) and \\(w_{b_n}\\) are the height and width of the object box \\(b_{n}\\) and \\(\\lambda > 1\\) is a padding factor controlling the looseness of the box. Next, we compute a binary mask \\(m_{n}\\) from the loose bounding box \\(b_{n}^{l}\\), such that m is 0 inside the box \\(b_{n}^{l}\\) and \\(-\\infty\\) outside. We use it to mask out the cross attentions \\((\\Psi)\\) of the object token \\(o_{n}\\) and compass token \\(\\mathbf{c}_{\\mathbf{n}}\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.586, + 0.632, + 0.831, + 0.668 + ], + "angle": 0, + "content": "\\[\n\\Psi (\\mathbf {c} _ {\\mathbf {n}}) = \\operatorname {s o f t m a x} \\left(m + \\frac {Q K (\\mathbf {c} _ {\\mathbf {n}}) ^ {T}}{\\sqrt {d _ {K}}}\\right)\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.588, + 0.666, + 0.83, + 0.703 + ], + "angle": 0, + "content": "\\[\n\\Psi \\left(\\mathbf {o} _ {\\mathbf {n}}\\right) = \\operatorname {s o f t m a x} \\left(m + \\frac {Q K \\left(o _ {n}\\right) ^ {T}}{\\sqrt {d _ {K}}}\\right)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.909, + 0.901 + ], + "angle": 0, + "content": "where the query feature \\( Q \\) comes from the U-Net features, and the key features \\( K(\\mathbf{c_n}) \\) and \\( K(\\mathbf{o_n}) \\) come from the respective tokens. This masking operation is performed at all the diffusion timesteps and cross-attention layers. We find that using a loose mask is highly effective, providing greater flexibility during the generation process. This attention localization mechanism is dubbed as Coupled Attention Localization, or CALL for short. Adding CALL mechanism during training and inference has key advantages for composing multi-object scenes: a) Appropriate binding of each compass token \\( \\mathbf{c_n} \\) with its corresponding object token \\( o_n \\) leads to disentangled orientation control of individual objects. b) Constraining the cross-attention for the object to" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2794" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.105, + 0.093, + 0.12, + 0.104 + ], + "angle": 0, + "content": "a)" + }, + { + "type": "image", + "bbox": [ + 0.107, + 0.093, + 0.468, + 0.291 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.296, + 0.484, + 0.395 + ], + "angle": 0, + "content": "Figure 4. Binding the compass tokens: We visualize the averaged cross attention of the compass token(s) when training with CALL (shown on the left) and without it (shown on the right). CALL localizes the influence of the compass token at the correct regions, which (a) improves orientation control (b) disentangles orientations in multi-object scenes. In (b), \\(\\mathbf{c}_1\\) and \\(\\mathbf{c}_2\\) are compass tokens for car and motorbike, respectively." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.416, + 0.484, + 0.477 + ], + "angle": 0, + "content": "kens \\( o_n \\) to non-overlapping bounding boxes results in disentanglement between the objects themselves (a known issue in T2I models [10]), enabling strong generalization to complex scenes with multiple objects." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.488, + 0.484, + 0.76 + ], + "angle": 0, + "content": "Training. We train our encoder model \\(\\mathcal{P}\\) and fine-tune the denoising U-Net with LoRA [24] on the synthetic dataset from Sec. 3.1. The LoRA training is extremely parameter efficient and preserves the behavior of the base T2I model. We use the proposed CALL mechanism for effective learning of orientation control. However, for the effective working of CALL, the object must be generated within the loose bounding box, as the compass token's influence is restricted to this region. To this end, we first train on simple single-object scenes, to first learn the bounding box adherence for the generated object, and then train on a mix of both single and two-object scenes thereafter. We contrast this two-staged training procedure with the single-stage training at an intermediate training iteration in Fig. 5. The generated objects adhere to the bounding box better in the two-staged training compared to single-stage training. This leads to effective learning of orientation control, as we have shown in ablative experiments (Sec. 4)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.771, + 0.484, + 0.892 + ], + "angle": 0, + "content": "Inference. During the inference phase, Compass Control expects the text prompt containing the objects, desired orientations, and optional coarse 2D bounding boxes as input. Using loose bounding boxes during training offers a significant advantage here, as we can even spawn non-overlapping boxes heuristically, as shown in the Suppl. Sec.F. We use the text tokens and the compass tokens together to condition the diffusion model." + }, + { + "type": "image_caption", + "bbox": [ + 0.52, + 0.09, + 0.61, + 0.108 + ], + "angle": 0, + "content": "w/o staged training at iteration i" + }, + { + "type": "image_caption", + "bbox": [ + 0.617, + 0.09, + 0.703, + 0.108 + ], + "angle": 0, + "content": "w staged training at iteration i" + }, + { + "type": "image_caption", + "bbox": [ + 0.715, + 0.09, + 0.804, + 0.107 + ], + "angle": 0, + "content": "w/o staged training at iteration i" + }, + { + "type": "image_caption", + "bbox": [ + 0.814, + 0.09, + 0.903, + 0.182 + ], + "angle": 0, + "content": "w staged trainingat iteration i" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.109, + 0.614, + 0.181 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.615, + 0.108, + 0.71, + 0.181 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.712, + 0.108, + 0.807, + 0.182 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.808, + 0.108, + 0.903, + 0.182 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.184, + 0.905, + 0.213 + ], + "angle": 0, + "content": "Figure 5. Staged training results in improved adherence of objects to the bounding boxes, leading to orientation learning." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.215, + 0.669, + 0.229 + ], + "angle": 0, + "content": "3.4. Personalization" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.234, + 0.907, + 0.37 + ], + "angle": 0, + "content": "The design of Compass Control as a conditioning mechanism preserves the original capabilities of T2I models, such as personalization. Given a few unposed images of an object \\((\\approx 10)\\), we apply Dreambooth [50] with LoRA and associate a special token \\(\\hat{u}\\) for the input object using Compass Control's fine-tuned UNet. During inference, we can generate the object in desired orientation \\(\\theta\\) with simple prompts; e.g., 'A photo of a \\(\\mathbf{c}(\\theta)\\) \\(\\hat{u}\\) car on the beach,' where \\(\\mathbf{c}(\\theta)\\) is the compass token." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.379, + 0.645, + 0.396 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.401, + 0.687, + 0.417 + ], + "angle": 0, + "content": "4.1. Experiment setup" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.42, + 0.906, + 0.587 + ], + "angle": 0, + "content": "Dataset. We use the synthetic dataset from Sec. 3.1 consisting of 8900 rendered images and 6010 augmented images consisting of one and two object scenes for training our model. For quantitative evaluation, we construct a test set of 10 scene prompts having one/two object names generated from ChatGPT [9]. We use a mix of seen objects (horse, jeep, sedan, sofa, teddy, lion) and unseen objects (boat, dolphin, ship, SUV, tractor) in the prompts. For each object and prompt combination, we use a set of 10 randomly sampled orientations. The list of input text prompts and orientations is given in the Supp. Sec.L." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.59, + 0.906, + 0.727 + ], + "angle": 0, + "content": "Implementation Details. We use Stable Diffusion v2.1 [49] as our base T2I model and use LoRA rank 4 for fine-tuning its UNet. We present additional results on Stable Diffusion-XL [45] in Suppl. Sec. D. Our encoder model \\(\\mathcal{P}\\) is a lightweight MLP: three linear layers with ReLU. We train our model for 25,000 steps with a batch size of 4 with AdamW optimizer and a fixed learning rate of \\(10^{-4}\\). We keep the bounding box padding \\(\\lambda = 1.2\\) for CALL. The training takes 24 hours on a single A6000 GPU." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.73, + 0.907, + 0.882 + ], + "angle": 0, + "content": "Evaluation Metrics. We evaluate for a) Text Alignment using CLIP similarly; b) %Object Generation - we evaluate the presence of the intended object using Grounding-DINO [34] and threshold on the objectness score for each object in the prompt. c) Angular Error - to evaluate the orientation consistency, we compute the Angular error between the input orientation angle \\(\\theta\\) and the orientation of the generated object using a pretrained orientation predictor. Further details about the orientation predictor and the implementation details of the metrics are in the Suppl Sec.I." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.886, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Baselines. As no prior method tackles our same task, we" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2795" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.156, + 0.092, + 0.385, + 0.196 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.168, + 0.195, + 0.374, + 0.204 + ], + "angle": 0, + "content": "\"A photo of a sedan spinning on a lazy susan in a mall\"" + }, + { + "type": "image", + "bbox": [ + 0.387, + 0.092, + 0.613, + 0.196 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.404, + 0.196, + 0.594, + 0.213 + ], + "angle": 0, + "content": "'A photo of a pirate ship sailing in a sea with dark clouds near a port'" + }, + { + "type": "image", + "bbox": [ + 0.613, + 0.092, + 0.842, + 0.196 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.624, + 0.196, + 0.829, + 0.213 + ], + "angle": 0, + "content": "'A photo of a helicopter flying over a magnificent city, the sky is cloudy at sunset'" + }, + { + "type": "image", + "bbox": [ + 0.156, + 0.218, + 0.385, + 0.322 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.171, + 0.321, + 0.375, + 0.347 + ], + "angle": 0, + "content": "A photo of a jeep and a sedan in front of a serene waterfall with trees scattered around the region, and stones scattered in the water" + }, + { + "type": "image", + "bbox": [ + 0.386, + 0.218, + 0.613, + 0.322 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.408, + 0.321, + 0.594, + 0.348 + ], + "angle": 0, + "content": "A photo of a boat and a yacht floating in a river, featuring a lot of greenery, the sun is setting in the background" + }, + { + "type": "image", + "bbox": [ + 0.615, + 0.218, + 0.842, + 0.322 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.639, + 0.321, + 0.802, + 0.339 + ], + "angle": 0, + "content": "'A man and a woman talking to each other sitting in a garden'" + }, + { + "type": "image", + "bbox": [ + 0.158, + 0.352, + 0.837, + 0.484 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.171, + 0.487, + 0.281, + 0.514 + ], + "angle": 0, + "content": "A photo of a dog and a dog and a pram in a rustic living room" + }, + { + "type": "image_caption", + "bbox": [ + 0.295, + 0.486, + 0.422, + 0.529 + ], + "angle": 0, + "content": "A photo of a sedan and a sedan and a SUV moving down a modern highway with a lush green forest on the sides at sunset" + }, + { + "type": "image_caption", + "bbox": [ + 0.437, + 0.485, + 0.556, + 0.519 + ], + "angle": 0, + "content": "'A photo of a bicycle and a bicycle on a wooden bridge, a boat and a boat on the shiny waters on a sunny afternoon" + }, + { + "type": "image_caption", + "bbox": [ + 0.575, + 0.485, + 0.691, + 0.529 + ], + "angle": 0, + "content": "'A photo of a shoe and a shoe and a teddy and a sofa and a teddy in a lavish living room with soft lighting and wooden flooring'" + }, + { + "type": "image_caption", + "bbox": [ + 0.709, + 0.486, + 0.829, + 0.53 + ], + "angle": 0, + "content": "A photo of a SUV and a SUV and a SUV and a sedan and a sedan and a sedan at a city intersection at night, featuring tall buildings" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.532, + 0.907, + 0.575 + ], + "angle": 0, + "content": "Figure 6. Main Results. Compass Control generates complex scenes aligned with the text prompts and the orientations (shown as frustum, the colored face is forward direction). It generalizes well to unseen object categories - pirate ship, boat, yacht, bicycle, pram, dog, and even human. Further, it generates high-quality compositions of several objects, despite being trained only on one and two object scenes." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.586, + 0.485, + 0.889 + ], + "angle": 0, + "content": "compare against methods that allow for either camera pose control or 3D object pose control in text-to-image models: a) Continuous 3D Words (Cont-3D-Words) [13]: Following their exact setup, we train a 3D word for controlling the object orientation on renderings of a single 3D asset - Sedan and its ControlNet augmentations. b) ViewNeTI [6]: In contrast to Cont-3D-Words, ViewNeTI allows for training on multiple 3D assets; for fair evaluation, we train ViewNeTI on our training dataset and condition the T2I on object orientation instead of 3D camera pose. Notably, both of these methods are limited to a global view control. Hence, we evaluate on only single-object scenes. c) LooseControl [4], allows for multi-object control by conditioning on loose depth maps (Fig. 7) formed by 3D object boxes. We use template 3D bounding boxes for each test object and place them in the scene, with random orientation and location (similar to Sec. 3.1). For a fair comparison, we use the 2D boxes corresponding to the 3D boxes in our outputs. Notably, LooseControl does not take exact orientation as input as a 180 flipped box also has the same loose depth;" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.586, + 0.907, + 0.661 + ], + "angle": 0, + "content": "we consider this while computing the Angular error. Further, it requires the user to provide accurate 3D bounding boxes during inference, which is cumbersome, whereas our method requires only loose 2D boxes. Additional baseline details are in Suppl. Sec.J." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.67, + 0.655, + 0.685 + ], + "angle": 0, + "content": "4.2. Main Results" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.689, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Qualitative results. We present our method's results in Fig. 6. Our method is able to generate complex multi-object scenes with precise orientation control of individual objects, even though it is trained with single and two object scenes. Further, it is able to generalize well to challenging unseen objects such as humans and prams. Interestingly, there was no water-based subject in the training dataset, yet our method can achieve precise orientation control for a ship, yacht, and boat. These strong generalization capabilities of Compass Control can be attributed to effective attention constraining with CALL and diversity in the 3D assets used for training. The conditioning mechanism of Compass Control is generalized, and we present results for jointly controlling all three orientation angles, camera" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2796" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.1, + 0.081, + 0.472, + 0.19 + ], + "angle": 0, + "content": "
Single objectText Align. ↓% Obj. Generated ↑Angular Err. ↓
ViewNeTI [6]22.120.9200.596
Cont-3D-words [13]29.880.7320.509
LooseControl [4]31.600.6560.385
Ours32.980.9680.198
Multiple objectText Align. ↓% Obj. Generated ↑Angular Err. ↓
LooseControl [4]31.730.7780.372
Ours33.930.9640.215
" + }, + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.194, + 0.483, + 0.237 + ], + "angle": 0, + "content": "Table 2. Quantitative comparison. We compute Text Alignment, using CLIP, % of correct subject generation and Angular error between the predicted and input orientations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.249, + 0.411, + 0.263 + ], + "angle": 0, + "content": "elevation, and object scale in Suppl. Sec.B & C." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.264, + 0.485, + 0.657 + ], + "angle": 0, + "content": "Baseline comparison. We compare our method to all three baselines on single-object scenes and additionally include multi-object scene comparison with LooseControl in Fig. 7 and Tab. 2. Cont-3D-words morphs the generated objects into a sedan shape seen during training and generates washed-out backgrounds. This results in poor text alignment and a lower percentage of intended object generation. ViewNeTI can generate better object shapes in a given orientation; however, it overfits to the black backgrounds seen during training, leading to poor text alignment. This is primarily because ViewNeTI does not accept ControlNet augmentations in its original form as it is designed for novel view synthesis. LooseControl generates realistic single-subject scenes following the given text prompt. However, in some cases, the object orientation is not followed (e.g., sofa, teddy). For a multi-object generation (Fig. 7b), LooseControl either misses the object during generation (horse in the second column and cars in the last column) or distorts the object shapes. LooseControl distorts the object to a box like appearance (columns 3 and 4) as it learns LoRA over the original depth-conditioned ControlNet, which follows the depth input closely, resulting in an inferior object generation score. Further, though we adjust for a 180-degree flip in computing Angular error for LooseControl, our method achieves significantly lower Angular error, demonstrating a strong orientation adherence." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.666, + 0.212, + 0.683 + ], + "angle": 0, + "content": "4.3. User study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.688, + 0.484, + 0.84 + ], + "angle": 0, + "content": "We conducted a user study with 57 participants to compare all methods on text alignment, object quality, and orientation consistency. Users rated 90 image pairs for single-object scenes and 30 for multi-object scenes, choosing the better image in each pair, sampled from our method and baselines. In total, we obtained 5130 ratings for single-object and 1710 for two-object scenes. Results (Fig. 8) show users preferred our method overall, with LooseControl scoring well in text alignment for single objects but falling short across all metrics in multi-object scenes." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.849, + 0.247, + 0.863 + ], + "angle": 0, + "content": "4.4. Personalization" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.484, + 0.903 + ], + "angle": 0, + "content": "We present personalization results in Fig. 9. With only 10 unposed images of an object, our method can gener" + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.093, + 0.905, + 0.379 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.382, + 0.904, + 0.576 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.581, + 0.907, + 0.678 + ], + "angle": 0, + "content": "Figure 7. Qualitative Comparison. We compare our method against three baselines. Cont-3D-words [13] does not generate the intended object whereas View-NeTI [6] generates objects in plain backgrounds. LooseControl [4] generates realistic compositions but does not follow the input orientation well. In contrast, our method aligns with the input text prompt and follows the input orientation, while generating realistic scenes." + }, + { + "type": "image", + "bbox": [ + 0.527, + 0.681, + 0.905, + 0.807 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.815, + 0.907, + 0.858 + ], + "angle": 0, + "content": "Figure 8. User study. We compare all methods on the three image metrics. Each bar indicates the fraction of people that preferred our result (gray) vs the baseline (other color)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.907, + 0.903 + ], + "angle": 0, + "content": "ate the object with precise orientation control in various contexts. Furthermore, we can jointly optimize Dream" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2797" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.198 + ], + "angle": 0, + "content": "booth [50] LoRA weights for two objects (e.g., a chair and a teddy bear), enabling multi-object personalization with object-centric orientation control. We compare our method with CD-360 [31] and achieve comparable performance. Unlike CD-360 [31], which requires \\(\\approx 100\\) object images with camera pose, we require only a few unposed images, making it more convenient and user-friendly." + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.213, + 0.479, + 0.496 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.497, + 0.484, + 0.581 + ], + "angle": 0, + "content": "Figure 9. Personalization. Given a few \\((\\approx 10)\\) unposed images of an object, our method can personalize the diffusion models and allow for orientation control of the new object. Notably, our method can also generate scenes with two personalized objects with precise orientation control. Additionally, we compare our method with CustomDiffusion-360 [31] that uses \\(\\approx 100\\) posed images." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.6, + 0.203, + 0.613 + ], + "angle": 0, + "content": "4.5. Ablations" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.618, + 0.484, + 0.679 + ], + "angle": 0, + "content": "We present the results of the ablation study in Fig. 10 on generated scenes with 1, 3, and 5 objects. We focus on three key design choices and generate scenes with a variable number of objects:" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.682, + 0.484, + 0.773 + ], + "angle": 0, + "content": "Staged training: Training Compass Control in a single stage results in poor adherence to the bounding boxes. This especially affects complex multi-object layouts, where some objects tend to leak outside their box and suppress the generation of neighboring objects. This results in generating a lesser number of objects in the scene." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.776, + 0.484, + 0.868 + ], + "angle": 0, + "content": "CALL is crucial for accurate orientation control. Without it the compass token attends to irrelevant image regions, resulting in poor orientation control. Further, without CALL the object tokens entangle with each other during generation (a known issue in T2I models [10]). This results in generating a lesser number of objects in the scene." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Augmentations: Without the ControlNet augmentations, the model overfits the training backgrounds, resulting in" + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.094, + 0.903, + 0.359 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.361, + 0.905, + 0.388 + ], + "angle": 0, + "content": "Figure 10. Ablation studies. We show the impact of several design choices of our approach. See Sec. 4.5 for details." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.401, + 0.906, + 0.431 + ], + "angle": 0, + "content": "black backgrounds. Hence, ControlNet augmentation is necessary to generate objects in diverse contexts." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.438, + 0.762, + 0.453 + ], + "angle": 0, + "content": "5. Conclusion and Discussion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.46, + 0.906, + 0.49 + ], + "angle": 0, + "content": "Limitations. Our method struggles to control the orientation of objects that are occluded or have significant overlap." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.491, + 0.692, + 0.64 + ], + "angle": 0, + "content": "In these scenarios, the model either fails to generate one of the objects or mixes the attributes between objects (Fig. 11). Further, our single-angle orientation representation is too simplistic for modeling complex non-rigid objects such as humans." + }, + { + "type": "image", + "bbox": [ + 0.719, + 0.491, + 0.897, + 0.627 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.741, + 0.627, + 0.887, + 0.638 + ], + "angle": 0, + "content": "Figure 11. Failure Cases" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.642, + 0.906, + 0.686 + ], + "angle": 0, + "content": "In this work, we focus on presenting a generalized framework to condition text-to-image models on 3D controls that can be adapted easily for other representations." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.69, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Conclusion. In this work, we propose a method to condition pre-trained text-to-image models with 3D object orientation while preserving its rich image-generation capabilities. We train the conditioning module on a small synthetic dataset via staged training and involving attention regularization. These modifications enable strong generalization of the model, allowing for precise orientation control for unseen object categories and individual objects in a multi-object scene. Further, it can be seamlessly integrated with personalization methods to achieve orientation control of personalized objects. This work is a testament that text-to-image diffusion models innately have some form of 3D understanding, and interesting 3D controls can be obtained with appropriate conditionings." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2798" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.15 + ], + "angle": 0, + "content": "Acknowledgements. We thank Srinjay Sarkar, Abhijnya Bhat, and Tejan Karmali for thoroughly reviewing the manuscript. This work is supported by Meesho and PMRF by the Government of India." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.178, + 0.188, + 0.194 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.203, + 0.482, + 0.258 + ], + "angle": 0, + "content": "[1] Omri Avrahami, Rinon Gal, Gal Chechik, Ohad Fried, Dani Lischinski, Arash Vahdat, and Weili Nie. Diffuhaul: A training-free method for object dragging in images. arXiv preprint arXiv:2406.01594, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.261, + 0.482, + 0.3 + ], + "angle": 0, + "content": "[2] Omer Bar-Tal, Lior Yariv, Yaron Lipman, and Tali Dekel. Multidiffusion: Fusing diffusion paths for controlled image generation. 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.304, + 0.482, + 0.385 + ], + "angle": 0, + "content": "[3] Miguel Angel Bautista, Pengsheng Guo, Samira Abnar, Walter Talbott, Alexander Toshev, Zhuoyuan Chen, Laurent Dinh, Shuangfei Zhai, Hanlin Goh, Daniel Ulbricht, et al. Gaudi: A neural architect for immersive 3d scene generation. Advances in Neural Information Processing Systems, 35:25102-25116, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.389, + 0.482, + 0.442 + ], + "angle": 0, + "content": "[4] Shariq Farooq Bhat, Niloy Mitra, and Peter Wonka. Loosecontrol: Lifting controlnet for generalized depth conditioning. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.446, + 0.482, + 0.527 + ], + "angle": 0, + "content": "[5] Manuel Brack, Felix Friedrich, Katharia Kornmeier, Linoy Tsaban, Patrick Schramowski, Kristian Kersting, and Apolinário Passos. Ledits++: Limitless image editing using text-to-image models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8861-8870, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.531, + 0.482, + 0.585 + ], + "angle": 0, + "content": "[6] James Burgess, Kuan-Chieh Wang, and Serena Yeung. Viewpoint textual inversion: Unleashing novel view synthesis with pretrained 2d diffusion models. arXiv preprint arXiv:2309.07986, 2023. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.588, + 0.482, + 0.628 + ], + "angle": 0, + "content": "[7] John Canny. A computational approach to edge detection. IEEE Transactions on pattern analysis and machine intelligence, (6):679-698, 1986. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.632, + 0.482, + 0.711 + ], + "angle": 0, + "content": "[8] Mingdeng Cao, Xintao Wang, Zhongang Qi, Ying Shan, Xiaohu Qie, and Yinqiang Zheng. Masactrl: Tuning-free mutual self-attention control for consistent image synthesis and editing. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22560-22570, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.716, + 0.476, + 0.73 + ], + "angle": 0, + "content": "[9] Chatgpt. Chatgpt. In https://chat.openai.com-chat, 2022.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.732, + 0.483, + 0.786 + ], + "angle": 0, + "content": "[10] Hila Chefer, Yuval Alaluf, Yael Vinker, Lior Wolf, and Daniel Cohen-Or. Attend-and-excite: Attention-based semantic guidance for text-to-image diffusion models. ACM Transactions on Graphics (TOG), 42(4):1-10, 2023. 3, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.789, + 0.483, + 0.856 + ], + "angle": 0, + "content": "[11] Xi Chen, Lianghua Huang, Yu Liu, Yujun Shen, Deli Zhao, and Hengshuang Zhao. Anydoor: Zero-shot object-level image customization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6593-6602, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.86, + 0.483, + 0.9 + ], + "angle": 0, + "content": "[12] Yiwen Chen, Zilong Chen, Chi Zhang, Feng Wang, Xiaofeng Yang, Yikai Wang, Zhongang Cai, Lei Yang, Huaping Liu, and Guosheng Lin. Gaussianeditor: Swift and control" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.203, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "able 3d editing with gaussian splatting. corr abs/2311.14521 (2023), 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.123, + 0.905, + 0.204 + ], + "angle": 0, + "content": "[13] Ta-Ying Cheng, Matheus Gadelha, Thibault Groueix, Matthew Fisher, Radomir Mech, Andrew Markham, and Niki Trigoni. Learning continuous 3d words for text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6753-6762, 2024. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.208, + 0.905, + 0.247 + ], + "angle": 0, + "content": "[14] Blender Online Community. Blender - a 3D modelling and rendering package. Blender Foundation, Stichting Blender Foundation, Amsterdam, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.25, + 0.904, + 0.304 + ], + "angle": 0, + "content": "[15] Omer Dahary, Or Patashnik, Kfir Aberman, and Daniel Cohen-Or. Be yourself: Bounded attention for multi-subject text-to-image generation. arXiv preprint arXiv:2403.16990, 2(5), 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.307, + 0.904, + 0.361 + ], + "angle": 0, + "content": "[16] Ankit Dhiman, Manan Shah, Rishubh Parihar, Yash Bhalgat, Lokesh R Boregowda, and R Venkatesh Babu. Reflecting reality: Enabling diffusion models to produce faithful mirror reflections. arXiv preprint arXiv:2409.14677, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.364, + 0.905, + 0.417 + ], + "angle": 0, + "content": "[17] Xiaodan Du, Nicholas Kolkin, Greg Shakhnarovich, and Anand Bhattachad. Generative models: What do they know? do they know things? let's find out! arXiv preprint arXiv:2311.17137, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.421, + 0.905, + 0.503 + ], + "angle": 0, + "content": "[18] Mohamed El Banani, Amit Raj, Kevis-Kokitsi Maninis, Abhishek Kar, Yuanzhen Li, Michael Rubinstein, Deqing Sun, Leonidas Guibas, Justin Johnson, and Varun Jampani. Probing the 3d awareness of visual foundation models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21795-21806, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.505, + 0.905, + 0.56 + ], + "angle": 0, + "content": "[19] Dave Epstein, Allan Jabri, Ben Poole, Alexei Efros, and Aleksander Holynski. Diffusion self-guidance for controllable image generation. Advances in Neural Information Processing Systems, 36:16222-16239, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.562, + 0.905, + 0.617 + ], + "angle": 0, + "content": "[20] Songwei Ge, Taesung Park, Jun-Yan Zhu, and Jia-Bin Huang. Expressive text-to-image generation with rich text. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7545-7556, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.62, + 0.905, + 0.673 + ], + "angle": 0, + "content": "[21] Daniel Geng and Andrew Owens. Motion guidance: Diffusion-based image editing with differentiable motion estimators. In The Twelfth International Conference on Learning Representations. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.676, + 0.905, + 0.73 + ], + "angle": 0, + "content": "[22] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.733, + 0.905, + 0.786 + ], + "angle": 0, + "content": "[23] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.789, + 0.905, + 0.844 + ], + "angle": 0, + "content": "[24] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.847, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[25] Allan Jabri, Sjoerd van Steenkiste, Emiel Hoogeboom, Mehdi SM Sajjadi, and Thomas Kipf. Dorsal: Diffusion for object-centric representations of scenes et. al. arXiv preprint arXiv:2306.08068, 2023. 3" + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.093, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2799" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.176 + ], + "angle": 0, + "content": "[26] Kunal Kathare, Ankit Dhiman, K Vikas Gowda, Siddharth Aravindan, Shubham Monga, Basavaraja Shanthappa Vandrotti, and Lokesh R Boregowda. Instructive3d: Editing large reconstruction models with text instructions. In Proceedings of the Winter Conference on Applications of Computer Vision (WACV), pages 3246-3256, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.177, + 0.482, + 0.245 + ], + "angle": 0, + "content": "[27] Bahjat Kawar, Shiran Zada, Oran Lang, Omer Tov, Huiwen Chang, Tali Dekel, Inbar Mosseri, and Michal Irani. Imagc: Text-based real image editing with diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6007-6017, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.247, + 0.482, + 0.302 + ], + "angle": 0, + "content": "[28] Hyunsu Kim, Gayoung Lee, Yunjay Choi, Jin-Hwa Kim, and Jun-Yan Zhu. 3d-aware blending with generative nerfs. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22906-22918, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.303, + 0.482, + 0.357 + ], + "angle": 0, + "content": "[29] Yunji Kim, Jiyoung Lee, Jin-Hwa Kim, Jung-Woo Ha, and Jun-Yan Zhu. Dense text-to-image generation with attention modulation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7701-7711, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.359, + 0.482, + 0.426 + ], + "angle": 0, + "content": "[30] Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization of text-to-image diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1931-1941, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.428, + 0.482, + 0.482 + ], + "angle": 0, + "content": "[31] Nupur Kumari, Grace Su, Richard Zhang, Taesung Park, Eli Shechtman, and Jun-Yan Zhu. Customizing text-to-image diffusion with camera viewpoint control. arXiv preprint arXiv:2404.12333, 2024. 2, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.484, + 0.482, + 0.552 + ], + "angle": 0, + "content": "[32] Yuheng Li, Haotian Liu, Qingyang Wu, Fangzhou Mu, Jianwei Yang, Jianfeng Gao, Chunyuan Li, and Yong Jae Lee. Gligen: Open-set grounded text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22511-22521, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.553, + 0.482, + 0.636 + ], + "angle": 0, + "content": "[33] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 300–309, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.637, + 0.482, + 0.704 + ], + "angle": 0, + "content": "[34] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.707, + 0.482, + 0.775 + ], + "angle": 0, + "content": "[35] Grace Luo, Trevor Darrell, Oliver Wang, Dan B Goldman, and Aleksander Holynski. Readout guidance: Learning control from diffusion features. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8217-8227, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.776, + 0.482, + 0.843 + ], + "angle": 0, + "content": "[36] Guan Luo, Tian-Xing Xu, Ying-Tian Liu, Xiao-Xiong Fan, Fang-Lue Zhang, and Song-Hai Zhang. 3d gaussian editing with a single image. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 6627-6636, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.845, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[37] Oscar Michel, Anand Bhattachad, Eli VanderBilt, Ranjay Krishna, Aniruddha Kembhavi, and Tanmay Gupta. Object 3dit: Language-guided 3d-aware image editing. Advances in Neural Information Processing Systems, 36, 2024. 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.161 + ], + "angle": 0, + "content": "[38] Thu Nguyen-Phuoc, Chuan Li, Lucas Theis, Christian Richardt, and Yong-Liang Yang. Hologan: Unsupervised learning of 3d representations from natural images. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7588–7597, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.163, + 0.905, + 0.23 + ], + "angle": 0, + "content": "[39] Michael Niemeyer and Andreas Geiger. Giraffe: Representing scenes as compositional generative neural feature fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11453-11464, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.232, + 0.905, + 0.313 + ], + "angle": 0, + "content": "[40] Karran Pandey, Paul Guerrero, Matheus Gadelha, Yannick Hold-Geoffroy, Karan Singh, and Niloy J Mitra. Diffusion handles enabling 3d edits for diffusion models by lifting activations to 3d. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7695–7704, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.316, + 0.905, + 0.396 + ], + "angle": 0, + "content": "[41] Rishubh Parihar, Abhijnya Bhat, Abhipsa Basu, Saswat Mallick, Jogendra Nath Kundu, and R Venkatesh Babu. Balancing act: distribution-guided debiasing in diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6668-6678, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.399, + 0.905, + 0.468 + ], + "angle": 0, + "content": "[42] Rishubh Parihar, VS Sachidanand, Sabariswaran Mani, Tejan Karmali, and R Venkatesh Babu. Precisecontrol: Enhancing text-to-image diffusion models with fine-grained attribute control. In European Conference on Computer Vision, pages 469-487. Springer, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.47, + 0.905, + 0.537 + ], + "angle": 0, + "content": "[43] Or Patashnik, Daniel Garibi, Idan Azuri, Hadar Averbuch-Elor, and Daniel Cohen-Or. Localizing object-level shape variations with text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 23051–23061, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.539, + 0.905, + 0.592 + ], + "angle": 0, + "content": "[44] Or Patashnik, Rinon Gal, Daniel Cohen-Or, Jun-Yan Zhu, and Fernando De la Torre. Consolidating attention features for multi-view image editing. arXiv preprint arXiv:2402.14792, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.595, + 0.905, + 0.649 + ], + "angle": 0, + "content": "[45] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.651, + 0.905, + 0.691 + ], + "angle": 0, + "content": "[46] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.693, + 0.905, + 0.746 + ], + "angle": 0, + "content": "[47] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.748, + 0.905, + 0.802 + ], + "angle": 0, + "content": "[48] Harsh Rangwani, Aishwarya Agarwal, Kuldeep Kulkarni, R Venkatesh Babu, and Srikrishna Karanam. Crafting parts for expressive object composition. arXiv preprint arXiv:2406.10197, 2024.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.804, + 0.905, + 0.872 + ], + "angle": 0, + "content": "[49] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[50] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "2800" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.093, + 0.484, + 0.147 + ], + "angle": 0, + "content": "tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 22500-22510, 2023. 3, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.483, + 0.234 + ], + "angle": 0, + "content": "[51] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural information processing systems, 35:36479-36494, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.234, + 0.483, + 0.288 + ], + "angle": 0, + "content": "[52] Rahul Sajnani, Jeroen Vanbaar, Jie Min, Kapil Katyal, and Srinath Sridhar. Geodiffuser: Geometry-based image editing with diffusion models. arXiv preprint arXiv:2404.14403, 2024.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.291, + 0.483, + 0.36 + ], + "angle": 0, + "content": "[53] Yizhi Song, Zhifei Zhang, Zhe Lin, Scott Cohen, Brian Price, Jianming Zhang, Soo Ye Kim, and Daniel Aliaga. Object-stitch: Object compositing with diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18310-18319, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.361, + 0.483, + 0.404 + ], + "angle": 0, + "content": "[54] Alexander Vilesov, Pradyumna Chari, and Achuta Kadambi. Cg3d: Compositional generation for text-to-3d via gaussian splatting. arXiv preprint arXiv:2311.17907, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.405, + 0.483, + 0.459 + ], + "angle": 0, + "content": "[55] Andrey Voynov, Kfir Aberman, and Daniel Cohen-Or. Sketch-guided text-to-image diffusion models. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-11, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.461, + 0.483, + 0.516 + ], + "angle": 0, + "content": "[56] Andrey Voynov, Kfir Aberman, and Daniel Cohen-Or. Sketch-guided text-to-image diffusion models. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1–11, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.518, + 0.483, + 0.572 + ], + "angle": 0, + "content": "[57] Andrey Voynov, Qinghao Chu, Daniel Cohen-Or, and Kfir Aberman. \\(\\mathfrak{p}+\\): Extended textual conditioning in text-to-image generation. arXiv preprint arXiv:2303.09522, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.575, + 0.483, + 0.63 + ], + "angle": 0, + "content": "[58] Qian Wang, Yiqun Wang, Michael Birsak, and Peter Wonka. Blobgan-3d: A spatially-disentangled 3d-aware generative model for indoor scenes. arXiv preprint arXiv:2303.14706, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.632, + 0.483, + 0.687 + ], + "angle": 0, + "content": "[59] Ruicheng Wang, Jianfeng Xiang, Jiaolong Yang, and Xin Tong. Diffusion models are geometry critics: Single image 3d editing using pre-trained diffusion priors. arXiv preprint arXiv:2403.11503, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.689, + 0.483, + 0.731 + ], + "angle": 0, + "content": "[60] Zhaoning Wang, Ming Li, and Chen Chen. Luciddreaming: Controllable object-centric 3d generation. arXiv preprint arXiv:2312.00588, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.732, + 0.483, + 0.801 + ], + "angle": 0, + "content": "[61] Ziyi Wu, Yulia Rubanova, Rishabh Kabra, Drew A Hudson, Igor Gilitschenski, Yusuf Aytar, Sjoerd van Steenkiste, Kelsey R Allen, and Thomas Kipf. Neural assets: 3d-aware multi-object scene synthesis with image diffusion models. arXiv preprint arXiv:2406.09292, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.802, + 0.483, + 0.871 + ], + "angle": 0, + "content": "[62] Yang Xue, Yuheng Li, Krishna Kumar Singh, and Yong Jae Lee. Giraffe hd: A high-resolution 3d-aware generative model. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18440-18449, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.872, + 0.483, + 0.903 + ], + "angle": 0, + "content": "[63] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.484, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "image diffusion models. arXiv preprint arXiv:2308.06721, 2023.2,3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.178 + ], + "angle": 0, + "content": "[64] Hong-Xing Yu, Haoyi Duan, Charles Herrmann, William T Freeman, and Jiajun Wu. Wonderworld: Interactive 3d scene generation from a single image. arXiv preprint arXiv:2406.09394, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.905, + 0.22 + ], + "angle": 0, + "content": "[65] Guanqi Zhan, Chuanxia Zheng, Weidi Xie, and Andrew Zisserman. What does stable diffusion know about the 3d scene? arXiv preprint arXiv:2310.06836, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.222, + 0.905, + 0.277 + ], + "angle": 0, + "content": "[66] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.278, + 0.905, + 0.348 + ], + "angle": 0, + "content": "[67] Yuxin Zhang, Weiming Dong, Fan Tang, Nisha Huang, Haibin Huang, Chongyang Ma, Tong-Yee Lee, Oliver Deussen, and Changsheng Xu. Prospect: Prompt spectrum for attribute-aware personalization of diffusion models. ACM Transactions on Graphics (TOG), 42(6):1-14, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.349, + 0.905, + 0.432 + ], + "angle": 0, + "content": "[68] Shijie Zhou, Zhiwen Fan, Dejia Xu, Haoran Chang, Pradyumna Chari, Tejas Bharadwaj, Suya You, Zhangyang Wang, and Achuta Kadambi. Dreamscene360: Unconstrained text-to-3d scene generation with panoramic gaussian splatting. In European Conference on Computer Vision, pages 324–342. Springer, 2025. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.432 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "2801" + } + ] +] \ No newline at end of file diff --git a/2025/Compass Control_ Multi Object Orientation Control for Text-to-Image Generation/9022e55b-ce42-4998-ae5e-db01a18e1b74_origin.pdf b/2025/Compass Control_ Multi Object Orientation Control for Text-to-Image Generation/9022e55b-ce42-4998-ae5e-db01a18e1b74_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..2c8da3fb2e47d17e3edd45757525d3178e60550f --- /dev/null +++ b/2025/Compass Control_ Multi Object Orientation Control for Text-to-Image Generation/9022e55b-ce42-4998-ae5e-db01a18e1b74_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fe58061dc387e02ef0d639a06e58dcf8fbaf9539c04e9918faaba107d729d460 +size 9715579 diff --git a/2025/Compass Control_ Multi Object Orientation Control for Text-to-Image Generation/full.md b/2025/Compass Control_ Multi Object Orientation Control for Text-to-Image Generation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..a280abd40cefdec48185f5782f290832e788e56c --- /dev/null +++ b/2025/Compass Control_ Multi Object Orientation Control for Text-to-Image Generation/full.md @@ -0,0 +1,340 @@ +# Compass Control: Multi Object Orientation Control for Text-to-Image Generation + +Rishubh Parihar $^{1*}$ + +Vaibhav Agrawal $^{2*†}$ $^{1}$ IISc Bangalore + +Sachidanand VS1 +2IIT Hyderabad + +R. Venkatesh Babu + +![](images/c814fe204320a3daaa9b2ef57781c1e019d52279a27c6bbc0ba8a35b3a2af27f.jpg) +a) +Stable Diffusion +A photo of a SUV facing left and a sedan facing right on a road. + +![](images/52ee9b5bcc110321c82b00948eb5a9810421f953f230a1db5da6aec14ab65e82.jpg) +Compass Control +'A photo of a [ ] SUV and a [ sedan on a road... + +![](images/aa137149fd5a43de9c9084bb41432ebd9ae6bffd5fde6166eba88d1693e22540.jpg) + +![](images/67e13276661e762abca383cc8adc443b2f3e24b9a72cd7059887a9aa63976ae7.jpg) + +![](images/d4a82691815ff9a9092aa2ac781024049d10d1565a7f95da32838b4a1d6944fa.jpg) +'A photo of a mother walking with a pram on a snowy street, festive Christmas lights, beautiful winter evening scene' + +![](images/0f790c5db6fab1fa2d98b01c27e0b0ead46303d7faf58e190b4fbc75264e0cd7.jpg) + +![](images/cdcc98c2742b63c71ddcd7eed64023733992c98bff86a155e6f32eb3ea6d5210.jpg) + +![](images/3a19bb20023145ea8b6ef1c0c35bace1fd0d3434972a0707782ff0ac1cecfc56.jpg) +A photo of a sofa and a teddy in a sunlit greenhouse filled with tropical plants + +![](images/96261237efb6e4d8087995463867ab156adbed812a08fd84969e9d98673f6e1d.jpg) +A photo of a sparrow and a sparrow and a hen among fallen autumn leaves carpeting the ground, soft afternoon glow around + +![](images/0630f7c2043dd9cb060dad580bb0106f0d32322ac232802ea474de41636eb969.jpg) +'A photo of a sedan and a sedan and a camel and a camel and a camel in a vast desert with dust rising in the background' + +![](images/e861c8201aa0333269d5d33ffa5f153e83aa05312459040436f78d4426900274.jpg) +Personalization with 3D Orientation Control +A photo of $V^{*}$ car in front of the leaning tower of Pisa in Italy +Figure 1. We present Compass Control, a method to generate multi-object scenes with orientation control from text-to-image diffusion models. Given a text prompt and an orientation of each object (shown as frustum, the colored face is the forward direction), our method generates scenes that align with both the prompt and specified orientations. Additionally, with a few $(\approx 10)$ unposed images of a new object, our model is personalized to generate the object in target orientations. + +# Abstract + +Existing approaches for controlling text-to-image diffusion models, while powerful, do not allow for explicit 3D object-centric control, such as precise control of object orienta + +tion. In this work, we address the problem of multi-object orientation control in text-to-image diffusion models. This enables the generation of diverse multi-object scenes with precise orientation control for each object. The key idea is to condition the diffusion model with a set of orientation-aware compass tokens, one for each object, along with text tokens. A light-weight encoder network predicts these com + +pass tokens taking object orientation as the input. The model is trained on a synthetic dataset of procedurally generated scenes, each containing one or two 3D assets on a plain background. However, direct training this framework results in poor orientation control as well as leads to entanglement among objects. To mitigate this, we intervene in the generation process and constrain the cross-attention maps of each compass token to its corresponding object regions. The trained model is able to achieve precise orientation control for a) complex objects not seen during training and b) multi-object scenes with more than two objects, indicating strong generalization capabilities. Further, when combined with personalization methods, our method precisely controls the orientation of the new object in diverse contexts. Our method achieves state-of-the-art orientation control and text alignment quantified with extensive evaluations and a user study. Project website + +# 1. Introduction + +Imagine a visual artist aiming to create a scene featuring two cars facing each other. They prompt a text-to-image model with - 'A photo of a sedan facing right and an SUV facing left'. However, as shown in Fig. 1 (a), the generated image may not always accurately capture the intended object orientations. Moreover, relying on text prompts to control object orientation (e.g., 'facing right') is imprecise and requires iterative prompting. Can we design an alternate interface for text-to-image models that accepts target orientation angle as input along with the text prompts? Such an interface will allow for precise orientation control for each object, eliminating the need for iterative prompt adjustments and streamlining the creative process. + +Several works have been proposed to achieve finer control in text-to-image models such as changing object appearance, scene layouts or image style [5, 8, 11, 21, 22, 43, 53, 55, 63, 66]. While effective for controlling 2D attributes of the image, these methods fail to accurately control 3D attributes. More recently, several methods have been proposed to control 3D properties in text-to-image models, such as camera viewpoint [6, 31], scene lighting [13], or scene layout using 3D bounding boxes [4]. However, these approaches either require dense 3D information, like multi-view images or accurate 3D bounding boxes, or limited to simple single-object scenes (Tab. 1). In this work, we present a novel interface to condition text-to-image diffusion models to generate multi-object scenes with precise 3D orientation control, without the need for multi-view images or 3D bounding boxes. + +Text-to-image (T2I) diffusion models enable the generation of objects with specific attributes via text prompts (e.g., 'a red car'). Motivated by this, we encode the object orientation as an additional attribute in the text embedding space of the T2I model. Specifically, we introduce a special token, dubbed as compass token (c) along with each object token + +Table 1. Comparison with Related works. Our approach uniquely allows for object-centric orientation control and generalize to novel categories without any explicit 3D representation. + +
MethodCD-360 [31]LooseControl [4]Cont-3D-Words [13]ViewNeTI [6]Ours
InputCam Pose3D boxesCam PoseCam PoseOrientation
3D conditioningExplicitExplicitImplicitImplicitImplicit
Novel classesX✓X✓✓
Multiple Object Control✓✓XX✓
+ +in the prompt (e.g., 'A photo of $\mathbf{c_1}$ SUV and $\mathbf{c_2}$ sedan on a road'). Each compass token is predicted by a lightweight encoder model taking the prescribed orientation angle as input. This formulation preserves the original interface of the base T2I model and enables precise object-centric orientation control in multi-object scenes. We train the encoder model and fine-tune the denoising U-Net with LoRA [24] on a synthetic dataset of scenes containing one or two 3D assets placed in diverse layouts on an empty floor. + +We discover that directly injecting compass tokens with the prompt tokens leads to poor orientation control, as the compass token attends to irrelevant image regions, limiting its influence on its corresponding object (Fig. 4(a)). To address this, we propose Coupled Attention Localization (CALL) mechanism, where we constrain the cross-attention maps of the compass token and its corresponding object token within a 2D bounding box. This results in a tight coupling between the two tokens, enabling the compass token to precisely control the orientation for its corresponding object. Additionally, for multi-object scenes, this results in an appropriate binding between each compass token and its corresponding object token, leading to disentangled orientation control of individual objects (Fig. 4(b)). + +The proposed approach achieves precise orientation control for unseen objects (e.g., pram) and can generalize to scenes with more than two objects, despite being trained on one and two object scenes only (see Fig. 1). Further, given a few unposed images of a real object, we can personalize the model to control the orientation of the new object. We evaluate our method against several baselines, achieving superior performance both quantitatively and in a user study. In summary, our primary contributions are: + +1. Compass Control - A method for conditioning text-to-image diffusion models on object orientation, enabling precise orientation control for individual objects in multi-object scene generation. +2. Coupled Attention Localization - A mechanism to restrict the influence of the input object orientation to the corresponding object, ensuring effective object-centric orientation control and object disentanglement. +3. Strong generalization of Compass Control for precise orientation control to unseen objects and complex multi-object scenes, though trained on simple synthetic scenes. +4. Personalization of Compass Control - Given a few unposed images of a real object, our method can perform orientation control of the new object in diverse contexts. + +# 2. Related work + +Controlled generation in T2I models. Several works have been proposed to achieve fine-grained control in text-to-image diffusion models [47, 49, 51]. Recent works resort to manipulation of the text embeddings [20, 27, 42, 57, 67], or attention maps in the diffusion U-Net [1, 2, 8, 10, 15, 22, 29, 40, 43, 48, 55, 56] for controlling the generated image. Additional encoder models can be trained to condition the T2I models on a new modality such as depth, bounding boxes, or object identity [32, 63, 66]. Another set of works personalizes the diffusion model given with a few subject images [30, 50] enabling generation of the learned subject in different backgrounds. Recent work on guiding diffusion models [19, 35, 41] allows inference time control over the scene contents, allowing the control of object location, appearance, shape, and skeleton pose. However, these controls are limited to 2D. + +3D-aware image editing. Recent works leverage the rich generation capability of the T2I model to perform 3D aware editing [40, 52, 59]. Specifically, they use the input scene depth as an additional input and use it to warp the internal features of the diffusion models. This enables zero-shot geometric 3D edits such as translating or rotating an object. However, these methods are limited to the editing of a single object. Another line of work uses multi-view input images and trains an implicit 3D representation such as a radiance field in the diffusion feature space [31, 44] to perform 3D consistent editing. More recently, few works leverage 3D Gaussian splat representation along with T2I models to perform scene editing [12, 36]. However, the above methods require scene-specific training and require multi-view images of accurate depth maps as input. Another direction explores large-scale training of diffusion model on a specific dataset [25, 37, 61] allowing for 3D scene editing. However, these methods fail to generalize to in-the-wild real-world scenes outside the distribution of training datasets. + +3D control in generation. Earlier works train generative models from scratch with explicit scene controls such as 3D blobs [58], or radiance fields of individual objects [3, 26, 28, 38, 39, 62] to control the generated scene. Recent works have shown the existence of 3D properties in text-to-image diffusion models [6, 16-18, 65]. This has fuelled the research that leverages this knowledge from T2I models for 3D generation. A set of works lifts this knowledge to 3D by distilling from pre-trained T2I models [33, 46, 54, 60, 64, 68]. Another set of recent works [4, 6, 13] leverage this underlying 3D knowledge for controlling 3D properties of the generated scene with an additional conditioning mechanism. ViewNeTI [6] learns a 3D view token to control the camera view for the task of novel-view synthesis from a dataset of multi-view images. ViewNeTI is natively designed for novel view synthesis but is unable to generate objects in diverse contexts. A continuous word + +![](images/8d77d907befbebfa7ecda7d17f7608f7593fad56482dedf4aade4aaa5e9a06e3.jpg) +Figure 2. Synthetic data generation. We curate 10 diverse 3D assets, and render them in diverse layouts and orientations in Blender [14]. The rendered scenes are augmented with realistic generations from Canny [7] ControlNet [66]. The final dataset consists of one and two object scenes. + +representation is trained in [13] for 3D scene properties such as orientation and lighting on a single 3D mesh. However, both these approaches are limited to simple scenes and control only the global view angle. The closest to our work is [4] that conditions the T2I model on loose depth maps using ControlNet [66]. Loose depth maps are created using 3D object boxes and scene boundaries to enable 3D object-centric control. However, this approach relies on precise 3D boxes, making it cumbersome at inference. In contrast, our method only requires coarse 2D bounding boxes and orientation angles, offering a more user-friendly solution. + +# 3. Method + +Text-to-Image Diffusion Models. Diffusion models, when directly applied in the pixel space, are computationally expensive due to their iterative nature. To mitigate this, latent diffusion models [49] apply the diffusion process in the smaller resolution latent space of a pretrained autoencoder. Further, the generation can be conditioned on text by injecting text features into the diffusion U-Net with additional cross-attention layers. The cross-attention maps allow for precise control during generation [19, 23]. + +# 3.1. Dataset + +Synthetic scene generation: We curated a list of 10 diverse 3D assets from the web: ostrich, helicopter, shoe, jeep, teddy bear, lion, sedan, horse, motorbike and sofa to aid in model generalization (see Fig. 2). Our dataset consists of 1000 one-object and 7900 two-object scenes, rendered in Blender [14]. For each image, we save the 2D bounding boxes and the orientations of the objects. The objects are placed at varied locations and in varied orientations to increase layout diversity. + +Augmentations: Directly training on this dataset results in overfitting to the plain background and the black floor, as shown in ablations. To address this, we augment the + +![](images/ccc2c8387f2daaaab6064526d1a3df41c68dfc877b90e2930ff6a7e73c763c89.jpg) +Figure 3. Compass Control. Given an orientation angle $\theta_{j}$ , we project it to a compass token with a lightweight encoder model. The compass tokens are interleaved with the text tokens (as shown in the figure) and passed through the text encoder. The outputs of the text encoder are used to condition the denoising process in the U-Net. We train $\mathcal{P}$ and also fine-tune the U-Net using LoRA [24]. + +dataset using ControlNet [66] to place objects in diverse backgrounds while retaining known orientations. For each rendered image, we extract its Canny [7] edge map to condition ControlNet to generate the objects in diverse contexts (for e.g., 'in a garden...', 'near a lake...', etc). This approach preserves object orientations while altering their appearance. We manually filter out the inconsistent augmented images. Further details on the dataset creation process can be found in the Suppl. Sec.H. + +Orientation convention: In this paper, we parameterize orientation with a single angle $\theta$ , rotation around the up axis in the world coordinate system (pointing towards the sky). We define $\theta = 0$ as a reference when the object faces exactly towards the right (e.g., sedan in Fig. 2). We parameterize with a single orientation angle, as most of our objects are land objects, and only this rotation axis results in plausible object orientations. However, our method is not limited to a single orientation angle, and we present results in Suppl. Sec.B for conditioning on three orientation angles. + +# 3.2. Compass Control + +Given a text prompt $\mathcal{T}$ consisting of $N$ object names $\{o_1,o_2,\dots,o_N\}$ (e.g., 'A photo of a jeep and a sedan and a horse in a garden') and their corresponding 3D orientation angles $\{\theta_{1},\theta_{2},\ldots ,\theta_{N}\}$ we introduce a set of compass tokens $\{\mathbf{c}_1,\mathbf{c}_2,\dots,\mathbf{c}_{\mathbf{N}}\}$ to control the orientations of the respective objects. A compass token $\mathbf{c}_{\mathbf{n}}$ is an embedding in the input space of the text encoder. It is predicted by a lightweight MLP encoder network $\mathcal{P}$ (see Fig. 3), which takes as input the object orientation angle $\theta_{k}$ . The compass tokens are + +prepended before their corresponding object tokens (e.g., 'A photo of a c1 jeep and a c2 sedan and a c3 horse in a garden'), and passed through the text encoder. The outputs of the text encoder are used to condition the denoising U-Net. + +However, directly training the above framework on the dataset from Sec. 3.1 fails to learn accurate orientation control. We hypothesize that this is because the added compass tokens are unrestricted and can attend to irrelevant image regions, limiting their influence on their corresponding objects. This is evident in the cross-attention maps for the compass tokens, which are indeed diffused in other image regions (see Fig. 4 (a)). Furthermore, this becomes a severe issue in multi-object scenes, where a single compass token can attend to multiple objects, resulting in the entanglement between different compass tokens(see Fig. 4 (b)). Existing works have shown that cross-attention maps closely control the image layouts [23] in the generated image. Motivated by this, we design a cross-attention localization approach. + +# 3.3. Coupled Attention Localization (CALL) + +Our key idea is to constrain the cross-attention maps for both the compass token and the corresponding object token inside a given 2D bounding box. This enables tight association between the object and the compass tokens. Additionally, it enables explicit control over the object location during generation. Specifically, during training, we use the saved object bounding boxes $\{b_1,\dots b_N\}$ to compute a set of loose square bounding boxes $\{b_1^l,\dots b_N^l\}$ . The side length $a_{n}$ of the loose box $b_{n}^{l}$ is computed as $a_{n} = \lambda * \max(h_{b_{n}},w_{b_{n}})$ , where $h_{b_n}$ and $w_{b_n}$ are the height and width of the object box $b_{n}$ and $\lambda > 1$ is a padding factor controlling the looseness of the box. Next, we compute a binary mask $m_{n}$ from the loose bounding box $b_{n}^{l}$ , such that m is 0 inside the box $b_{n}^{l}$ and $-\infty$ outside. We use it to mask out the cross attentions $(\Psi)$ of the object token $o_{n}$ and compass token $\mathbf{c}_{\mathbf{n}}$ as follows: + +$$ +\Psi (\mathbf {c} _ {\mathbf {n}}) = \operatorname {s o f t m a x} \left(m + \frac {Q K (\mathbf {c} _ {\mathbf {n}}) ^ {T}}{\sqrt {d _ {K}}}\right) +$$ + +$$ +\Psi \left(\mathbf {o} _ {\mathbf {n}}\right) = \operatorname {s o f t m a x} \left(m + \frac {Q K \left(o _ {n}\right) ^ {T}}{\sqrt {d _ {K}}}\right) +$$ + +where the query feature $Q$ comes from the U-Net features, and the key features $K(\mathbf{c_n})$ and $K(\mathbf{o_n})$ come from the respective tokens. This masking operation is performed at all the diffusion timesteps and cross-attention layers. We find that using a loose mask is highly effective, providing greater flexibility during the generation process. This attention localization mechanism is dubbed as Coupled Attention Localization, or CALL for short. Adding CALL mechanism during training and inference has key advantages for composing multi-object scenes: a) Appropriate binding of each compass token $\mathbf{c_n}$ with its corresponding object token $o_n$ leads to disentangled orientation control of individual objects. b) Constraining the cross-attention for the object to + +![](images/a639508dea2ee3aa028e0955a5025f3333c966d483b0152aeb67358d45c26dfe.jpg) +a) +Figure 4. Binding the compass tokens: We visualize the averaged cross attention of the compass token(s) when training with CALL (shown on the left) and without it (shown on the right). CALL localizes the influence of the compass token at the correct regions, which (a) improves orientation control (b) disentangles orientations in multi-object scenes. In (b), $\mathbf{c}_1$ and $\mathbf{c}_2$ are compass tokens for car and motorbike, respectively. + +kens $o_n$ to non-overlapping bounding boxes results in disentanglement between the objects themselves (a known issue in T2I models [10]), enabling strong generalization to complex scenes with multiple objects. + +Training. We train our encoder model $\mathcal{P}$ and fine-tune the denoising U-Net with LoRA [24] on the synthetic dataset from Sec. 3.1. The LoRA training is extremely parameter efficient and preserves the behavior of the base T2I model. We use the proposed CALL mechanism for effective learning of orientation control. However, for the effective working of CALL, the object must be generated within the loose bounding box, as the compass token's influence is restricted to this region. To this end, we first train on simple single-object scenes, to first learn the bounding box adherence for the generated object, and then train on a mix of both single and two-object scenes thereafter. We contrast this two-staged training procedure with the single-stage training at an intermediate training iteration in Fig. 5. The generated objects adhere to the bounding box better in the two-staged training compared to single-stage training. This leads to effective learning of orientation control, as we have shown in ablative experiments (Sec. 4). + +Inference. During the inference phase, Compass Control expects the text prompt containing the objects, desired orientations, and optional coarse 2D bounding boxes as input. Using loose bounding boxes during training offers a significant advantage here, as we can even spawn non-overlapping boxes heuristically, as shown in the Suppl. Sec.F. We use the text tokens and the compass tokens together to condition the diffusion model. + +![](images/b4fd45a83a5fc1c7452d1ac451c68ae47f06c2c3f5969820f905bc2e4da161b2.jpg) +w/o staged training at iteration i + +![](images/4b19886fe092ccf748d04a39a49a0bd8b21fddfd4f73e585cac93d29041aecc8.jpg) +w staged training at iteration i + +![](images/ce7dc394ce4370c8fd7e6c950648b92a0f692d339212e83e28ca2862c58ea199.jpg) +w/o staged training at iteration i +Figure 5. Staged training results in improved adherence of objects to the bounding boxes, leading to orientation learning. + +![](images/19b63436009884d7dbe2f3dd23680ef79d010a533dc7f4b9e22e22fb849b6066.jpg) +w staged trainingat iteration i + +# 3.4. Personalization + +The design of Compass Control as a conditioning mechanism preserves the original capabilities of T2I models, such as personalization. Given a few unposed images of an object $(\approx 10)$ , we apply Dreambooth [50] with LoRA and associate a special token $\hat{u}$ for the input object using Compass Control's fine-tuned UNet. During inference, we can generate the object in desired orientation $\theta$ with simple prompts; e.g., 'A photo of a $\mathbf{c}(\theta)$ $\hat{u}$ car on the beach,' where $\mathbf{c}(\theta)$ is the compass token. + +# 4. Experiments + +# 4.1. Experiment setup + +Dataset. We use the synthetic dataset from Sec. 3.1 consisting of 8900 rendered images and 6010 augmented images consisting of one and two object scenes for training our model. For quantitative evaluation, we construct a test set of 10 scene prompts having one/two object names generated from ChatGPT [9]. We use a mix of seen objects (horse, jeep, sedan, sofa, teddy, lion) and unseen objects (boat, dolphin, ship, SUV, tractor) in the prompts. For each object and prompt combination, we use a set of 10 randomly sampled orientations. The list of input text prompts and orientations is given in the Supp. Sec.L. + +Implementation Details. We use Stable Diffusion v2.1 [49] as our base T2I model and use LoRA rank 4 for fine-tuning its UNet. We present additional results on Stable Diffusion-XL [45] in Suppl. Sec. D. Our encoder model $\mathcal{P}$ is a lightweight MLP: three linear layers with ReLU. We train our model for 25,000 steps with a batch size of 4 with AdamW optimizer and a fixed learning rate of $10^{-4}$ . We keep the bounding box padding $\lambda = 1.2$ for CALL. The training takes 24 hours on a single A6000 GPU. + +Evaluation Metrics. We evaluate for a) Text Alignment using CLIP similarly; b) %Object Generation - we evaluate the presence of the intended object using Grounding-DINO [34] and threshold on the objectness score for each object in the prompt. c) Angular Error - to evaluate the orientation consistency, we compute the Angular error between the input orientation angle $\theta$ and the orientation of the generated object using a pretrained orientation predictor. Further details about the orientation predictor and the implementation details of the metrics are in the Suppl Sec.I. + +Baselines. As no prior method tackles our same task, we + +![](images/16d98db56fdb42175d98be89bbfde2c0e4bad477190b7eb3a07ee22b6b5af359.jpg) +"A photo of a sedan spinning on a lazy susan in a mall" + +![](images/466f125f22b75be25fec15dcc3c5168f5033660976109c54e685b642e681d63d.jpg) +'A photo of a pirate ship sailing in a sea with dark clouds near a port' + +![](images/6adbadff4797831c8c4ecd920d9ec1294f5d864629358df96ca94fc2ca50894a.jpg) +'A photo of a helicopter flying over a magnificent city, the sky is cloudy at sunset' + +![](images/86949ffae888969e88bff03d9734f5f35acb62becbeb1607710436824ec28372.jpg) +A photo of a jeep and a sedan in front of a serene waterfall with trees scattered around the region, and stones scattered in the water + +![](images/728605eb13ba2d178deeb06acb54b6608bd7a0027934c7b4e97d24cfe3377ce7.jpg) +A photo of a boat and a yacht floating in a river, featuring a lot of greenery, the sun is setting in the background + +![](images/068b79fafe5baea9de5179375ea056940852b4a9ed558b62e43d02baf66dc8a9.jpg) +'A man and a woman talking to each other sitting in a garden' + +![](images/b00f6b28095813cb09ba4cc410b5b96e69d5892d894cdda34421156a0c5421bc.jpg) +A photo of a dog and a dog and a pram in a rustic living room +A photo of a sedan and a sedan and a SUV moving down a modern highway with a lush green forest on the sides at sunset +'A photo of a bicycle and a bicycle on a wooden bridge, a boat and a boat on the shiny waters on a sunny afternoon +'A photo of a shoe and a shoe and a teddy and a sofa and a teddy in a lavish living room with soft lighting and wooden flooring' +A photo of a SUV and a SUV and a SUV and a sedan and a sedan and a sedan at a city intersection at night, featuring tall buildings +Figure 6. Main Results. Compass Control generates complex scenes aligned with the text prompts and the orientations (shown as frustum, the colored face is forward direction). It generalizes well to unseen object categories - pirate ship, boat, yacht, bicycle, pram, dog, and even human. Further, it generates high-quality compositions of several objects, despite being trained only on one and two object scenes. + +compare against methods that allow for either camera pose control or 3D object pose control in text-to-image models: a) Continuous 3D Words (Cont-3D-Words) [13]: Following their exact setup, we train a 3D word for controlling the object orientation on renderings of a single 3D asset - Sedan and its ControlNet augmentations. b) ViewNeTI [6]: In contrast to Cont-3D-Words, ViewNeTI allows for training on multiple 3D assets; for fair evaluation, we train ViewNeTI on our training dataset and condition the T2I on object orientation instead of 3D camera pose. Notably, both of these methods are limited to a global view control. Hence, we evaluate on only single-object scenes. c) LooseControl [4], allows for multi-object control by conditioning on loose depth maps (Fig. 7) formed by 3D object boxes. We use template 3D bounding boxes for each test object and place them in the scene, with random orientation and location (similar to Sec. 3.1). For a fair comparison, we use the 2D boxes corresponding to the 3D boxes in our outputs. Notably, LooseControl does not take exact orientation as input as a 180 flipped box also has the same loose depth; + +we consider this while computing the Angular error. Further, it requires the user to provide accurate 3D bounding boxes during inference, which is cumbersome, whereas our method requires only loose 2D boxes. Additional baseline details are in Suppl. Sec.J. + +# 4.2. Main Results + +Qualitative results. We present our method's results in Fig. 6. Our method is able to generate complex multi-object scenes with precise orientation control of individual objects, even though it is trained with single and two object scenes. Further, it is able to generalize well to challenging unseen objects such as humans and prams. Interestingly, there was no water-based subject in the training dataset, yet our method can achieve precise orientation control for a ship, yacht, and boat. These strong generalization capabilities of Compass Control can be attributed to effective attention constraining with CALL and diversity in the 3D assets used for training. The conditioning mechanism of Compass Control is generalized, and we present results for jointly controlling all three orientation angles, camera + +
Single objectText Align. ↓% Obj. Generated ↑Angular Err. ↓
ViewNeTI [6]22.120.9200.596
Cont-3D-words [13]29.880.7320.509
LooseControl [4]31.600.6560.385
Ours32.980.9680.198
Multiple objectText Align. ↓% Obj. Generated ↑Angular Err. ↓
LooseControl [4]31.730.7780.372
Ours33.930.9640.215
+ +Table 2. Quantitative comparison. We compute Text Alignment, using CLIP, % of correct subject generation and Angular error between the predicted and input orientations. + +elevation, and object scale in Suppl. Sec.B & C. + +Baseline comparison. We compare our method to all three baselines on single-object scenes and additionally include multi-object scene comparison with LooseControl in Fig. 7 and Tab. 2. Cont-3D-words morphs the generated objects into a sedan shape seen during training and generates washed-out backgrounds. This results in poor text alignment and a lower percentage of intended object generation. ViewNeTI can generate better object shapes in a given orientation; however, it overfits to the black backgrounds seen during training, leading to poor text alignment. This is primarily because ViewNeTI does not accept ControlNet augmentations in its original form as it is designed for novel view synthesis. LooseControl generates realistic single-subject scenes following the given text prompt. However, in some cases, the object orientation is not followed (e.g., sofa, teddy). For a multi-object generation (Fig. 7b), LooseControl either misses the object during generation (horse in the second column and cars in the last column) or distorts the object shapes. LooseControl distorts the object to a box like appearance (columns 3 and 4) as it learns LoRA over the original depth-conditioned ControlNet, which follows the depth input closely, resulting in an inferior object generation score. Further, though we adjust for a 180-degree flip in computing Angular error for LooseControl, our method achieves significantly lower Angular error, demonstrating a strong orientation adherence. + +# 4.3. User study + +We conducted a user study with 57 participants to compare all methods on text alignment, object quality, and orientation consistency. Users rated 90 image pairs for single-object scenes and 30 for multi-object scenes, choosing the better image in each pair, sampled from our method and baselines. In total, we obtained 5130 ratings for single-object and 1710 for two-object scenes. Results (Fig. 8) show users preferred our method overall, with LooseControl scoring well in text alignment for single objects but falling short across all metrics in multi-object scenes. + +# 4.4. Personalization + +We present personalization results in Fig. 9. With only 10 unposed images of an object, our method can gener + +![](images/51728fef24abdbe76c2b183c50b23e3a170f4bf96768ae09fe5750e3c8c9885e.jpg) + +![](images/df2781b37d444df1c0ae2fe625895a43f6de79f264bd379fcb89e8695f363758.jpg) +Figure 7. Qualitative Comparison. We compare our method against three baselines. Cont-3D-words [13] does not generate the intended object whereas View-NeTI [6] generates objects in plain backgrounds. LooseControl [4] generates realistic compositions but does not follow the input orientation well. In contrast, our method aligns with the input text prompt and follows the input orientation, while generating realistic scenes. + +![](images/896492711fe23ef86613d8624248e230d2614f17aef752fe60d0a16e1feb20c7.jpg) +Figure 8. User study. We compare all methods on the three image metrics. Each bar indicates the fraction of people that preferred our result (gray) vs the baseline (other color). + +ate the object with precise orientation control in various contexts. Furthermore, we can jointly optimize Dream + +booth [50] LoRA weights for two objects (e.g., a chair and a teddy bear), enabling multi-object personalization with object-centric orientation control. We compare our method with CD-360 [31] and achieve comparable performance. Unlike CD-360 [31], which requires $\approx 100$ object images with camera pose, we require only a few unposed images, making it more convenient and user-friendly. + +![](images/c0debfd42436386ef4388565d187f9463232cb558d84d837b29ae960340f2c35.jpg) +Figure 9. Personalization. Given a few $(\approx 10)$ unposed images of an object, our method can personalize the diffusion models and allow for orientation control of the new object. Notably, our method can also generate scenes with two personalized objects with precise orientation control. Additionally, we compare our method with CustomDiffusion-360 [31] that uses $\approx 100$ posed images. + +# 4.5. Ablations + +We present the results of the ablation study in Fig. 10 on generated scenes with 1, 3, and 5 objects. We focus on three key design choices and generate scenes with a variable number of objects: + +Staged training: Training Compass Control in a single stage results in poor adherence to the bounding boxes. This especially affects complex multi-object layouts, where some objects tend to leak outside their box and suppress the generation of neighboring objects. This results in generating a lesser number of objects in the scene. + +CALL is crucial for accurate orientation control. Without it the compass token attends to irrelevant image regions, resulting in poor orientation control. Further, without CALL the object tokens entangle with each other during generation (a known issue in T2I models [10]). This results in generating a lesser number of objects in the scene. + +Augmentations: Without the ControlNet augmentations, the model overfits the training backgrounds, resulting in + +![](images/56e2ac254e87aba415b3bf01ce1a0dc51a0b9512e0fa69f5924fb75239c335bc.jpg) +Figure 10. Ablation studies. We show the impact of several design choices of our approach. See Sec. 4.5 for details. + +black backgrounds. Hence, ControlNet augmentation is necessary to generate objects in diverse contexts. + +# 5. Conclusion and Discussion + +Limitations. Our method struggles to control the orientation of objects that are occluded or have significant overlap. + +In these scenarios, the model either fails to generate one of the objects or mixes the attributes between objects (Fig. 11). Further, our single-angle orientation representation is too simplistic for modeling complex non-rigid objects such as humans. + +![](images/8c6868b60fe5e8347718177f95b2e181fd16618ad7437bbcfde826792fc3e003.jpg) +Figure 11. Failure Cases + +In this work, we focus on presenting a generalized framework to condition text-to-image models on 3D controls that can be adapted easily for other representations. + +Conclusion. In this work, we propose a method to condition pre-trained text-to-image models with 3D object orientation while preserving its rich image-generation capabilities. We train the conditioning module on a small synthetic dataset via staged training and involving attention regularization. These modifications enable strong generalization of the model, allowing for precise orientation control for unseen object categories and individual objects in a multi-object scene. Further, it can be seamlessly integrated with personalization methods to achieve orientation control of personalized objects. This work is a testament that text-to-image diffusion models innately have some form of 3D understanding, and interesting 3D controls can be obtained with appropriate conditionings. + +Acknowledgements. We thank Srinjay Sarkar, Abhijnya Bhat, and Tejan Karmali for thoroughly reviewing the manuscript. This work is supported by Meesho and PMRF by the Government of India. + +# References + +[1] Omri Avrahami, Rinon Gal, Gal Chechik, Ohad Fried, Dani Lischinski, Arash Vahdat, and Weili Nie. Diffuhaul: A training-free method for object dragging in images. arXiv preprint arXiv:2406.01594, 2024. 3 +[2] Omer Bar-Tal, Lior Yariv, Yaron Lipman, and Tali Dekel. Multidiffusion: Fusing diffusion paths for controlled image generation. 2023. 3 +[3] Miguel Angel Bautista, Pengsheng Guo, Samira Abnar, Walter Talbott, Alexander Toshev, Zhuoyuan Chen, Laurent Dinh, Shuangfei Zhai, Hanlin Goh, Daniel Ulbricht, et al. Gaudi: A neural architect for immersive 3d scene generation. Advances in Neural Information Processing Systems, 35:25102-25116, 2022. 3 +[4] Shariq Farooq Bhat, Niloy Mitra, and Peter Wonka. Loosecontrol: Lifting controlnet for generalized depth conditioning. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 2, 3, 6, 7 +[5] Manuel Brack, Felix Friedrich, Katharia Kornmeier, Linoy Tsaban, Patrick Schramowski, Kristian Kersting, and Apolinário Passos. Ledits++: Limitless image editing using text-to-image models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8861-8870, 2024. 2 +[6] James Burgess, Kuan-Chieh Wang, and Serena Yeung. Viewpoint textual inversion: Unleashing novel view synthesis with pretrained 2d diffusion models. arXiv preprint arXiv:2309.07986, 2023. 2, 3, 6, 7 +[7] John Canny. A computational approach to edge detection. IEEE Transactions on pattern analysis and machine intelligence, (6):679-698, 1986. 3, 4 +[8] Mingdeng Cao, Xintao Wang, Zhongang Qi, Ying Shan, Xiaohu Qie, and Yinqiang Zheng. Masactrl: Tuning-free mutual self-attention control for consistent image synthesis and editing. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22560-22570, 2023. 2, 3 +[9] Chatgpt. Chatgpt. In https://chat.openai.com-chat, 2022.5 +[10] Hila Chefer, Yuval Alaluf, Yael Vinker, Lior Wolf, and Daniel Cohen-Or. Attend-and-excite: Attention-based semantic guidance for text-to-image diffusion models. ACM Transactions on Graphics (TOG), 42(4):1-10, 2023. 3, 5, 8 +[11] Xi Chen, Lianghua Huang, Yu Liu, Yujun Shen, Deli Zhao, and Hengshuang Zhao. Anydoor: Zero-shot object-level image customization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6593-6602, 2024. 2 +[12] Yiwen Chen, Zilong Chen, Chi Zhang, Feng Wang, Xiaofeng Yang, Yikai Wang, Zhongang Cai, Lei Yang, Huaping Liu, and Guosheng Lin. Gaussianeditor: Swift and control + +able 3d editing with gaussian splatting. corr abs/2311.14521 (2023), 2023. 3 +[13] Ta-Ying Cheng, Matheus Gadelha, Thibault Groueix, Matthew Fisher, Radomir Mech, Andrew Markham, and Niki Trigoni. Learning continuous 3d words for text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6753-6762, 2024. 2, 3, 6, 7 +[14] Blender Online Community. Blender - a 3D modelling and rendering package. Blender Foundation, Stichting Blender Foundation, Amsterdam, 2018. 3 +[15] Omer Dahary, Or Patashnik, Kfir Aberman, and Daniel Cohen-Or. Be yourself: Bounded attention for multi-subject text-to-image generation. arXiv preprint arXiv:2403.16990, 2(5), 2024. 3 +[16] Ankit Dhiman, Manan Shah, Rishubh Parihar, Yash Bhalgat, Lokesh R Boregowda, and R Venkatesh Babu. Reflecting reality: Enabling diffusion models to produce faithful mirror reflections. arXiv preprint arXiv:2409.14677, 2024. 3 +[17] Xiaodan Du, Nicholas Kolkin, Greg Shakhnarovich, and Anand Bhattachad. Generative models: What do they know? do they know things? let's find out! arXiv preprint arXiv:2311.17137, 2023. +[18] Mohamed El Banani, Amit Raj, Kevis-Kokitsi Maninis, Abhishek Kar, Yuanzhen Li, Michael Rubinstein, Deqing Sun, Leonidas Guibas, Justin Johnson, and Varun Jampani. Probing the 3d awareness of visual foundation models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21795-21806, 2024. 3 +[19] Dave Epstein, Allan Jabri, Ben Poole, Alexei Efros, and Aleksander Holynski. Diffusion self-guidance for controllable image generation. Advances in Neural Information Processing Systems, 36:16222-16239, 2023. 3 +[20] Songwei Ge, Taesung Park, Jun-Yan Zhu, and Jia-Bin Huang. Expressive text-to-image generation with rich text. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7545-7556, 2023. 3 +[21] Daniel Geng and Andrew Owens. Motion guidance: Diffusion-based image editing with differentiable motion estimators. In The Twelfth International Conference on Learning Representations. 2 +[22] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 2, 3 +[23] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 3, 4 +[24] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 2, 4, 5 +[25] Allan Jabri, Sjoerd van Steenkiste, Emiel Hoogeboom, Mehdi SM Sajjadi, and Thomas Kipf. Dorsal: Diffusion for object-centric representations of scenes et. al. arXiv preprint arXiv:2306.08068, 2023. 3 + +[26] Kunal Kathare, Ankit Dhiman, K Vikas Gowda, Siddharth Aravindan, Shubham Monga, Basavaraja Shanthappa Vandrotti, and Lokesh R Boregowda. Instructive3d: Editing large reconstruction models with text instructions. In Proceedings of the Winter Conference on Applications of Computer Vision (WACV), pages 3246-3256, 2025. 3 +[27] Bahjat Kawar, Shiran Zada, Oran Lang, Omer Tov, Huiwen Chang, Tali Dekel, Inbar Mosseri, and Michal Irani. Imagc: Text-based real image editing with diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6007-6017, 2023. 3 +[28] Hyunsu Kim, Gayoung Lee, Yunjay Choi, Jin-Hwa Kim, and Jun-Yan Zhu. 3d-aware blending with generative nerfs. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22906-22918, 2023. 3 +[29] Yunji Kim, Jiyoung Lee, Jin-Hwa Kim, Jung-Woo Ha, and Jun-Yan Zhu. Dense text-to-image generation with attention modulation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7701-7711, 2023. 3 +[30] Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization of text-to-image diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1931-1941, 2023. 3 +[31] Nupur Kumari, Grace Su, Richard Zhang, Taesung Park, Eli Shechtman, and Jun-Yan Zhu. Customizing text-to-image diffusion with camera viewpoint control. arXiv preprint arXiv:2404.12333, 2024. 2, 3, 8 +[32] Yuheng Li, Haotian Liu, Qingyang Wu, Fangzhou Mu, Jianwei Yang, Jianfeng Gao, Chunyuan Li, and Yong Jae Lee. Gligen: Open-set grounded text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22511-22521, 2023. 3 +[33] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 300–309, 2023. 3 +[34] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023. 5 +[35] Grace Luo, Trevor Darrell, Oliver Wang, Dan B Goldman, and Aleksander Holynski. Readout guidance: Learning control from diffusion features. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8217-8227, 2024. 3 +[36] Guan Luo, Tian-Xing Xu, Ying-Tian Liu, Xiao-Xiong Fan, Fang-Lue Zhang, and Song-Hai Zhang. 3d gaussian editing with a single image. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 6627-6636, 2024. 3 +[37] Oscar Michel, Anand Bhattachad, Eli VanderBilt, Ranjay Krishna, Aniruddha Kembhavi, and Tanmay Gupta. Object 3dit: Language-guided 3d-aware image editing. Advances in Neural Information Processing Systems, 36, 2024. 3 + +[38] Thu Nguyen-Phuoc, Chuan Li, Lucas Theis, Christian Richardt, and Yong-Liang Yang. Hologan: Unsupervised learning of 3d representations from natural images. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7588–7597, 2019. 3 +[39] Michael Niemeyer and Andreas Geiger. Giraffe: Representing scenes as compositional generative neural feature fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11453-11464, 2021. 3 +[40] Karran Pandey, Paul Guerrero, Matheus Gadelha, Yannick Hold-Geoffroy, Karan Singh, and Niloy J Mitra. Diffusion handles enabling 3d edits for diffusion models by lifting activations to 3d. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7695–7704, 2024. 3 +[41] Rishubh Parihar, Abhijnya Bhat, Abhipsa Basu, Saswat Mallick, Jogendra Nath Kundu, and R Venkatesh Babu. Balancing act: distribution-guided debiasing in diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6668-6678, 2024. 3 +[42] Rishubh Parihar, VS Sachidanand, Sabariswaran Mani, Tejan Karmali, and R Venkatesh Babu. Precisecontrol: Enhancing text-to-image diffusion models with fine-grained attribute control. In European Conference on Computer Vision, pages 469-487. Springer, 2024. 3 +[43] Or Patashnik, Daniel Garibi, Idan Azuri, Hadar Averbuch-Elor, and Daniel Cohen-Or. Localizing object-level shape variations with text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 23051–23061, 2023. 2, 3 +[44] Or Patashnik, Rinon Gal, Daniel Cohen-Or, Jun-Yan Zhu, and Fernando De la Torre. Consolidating attention features for multi-view image editing. arXiv preprint arXiv:2402.14792, 2024. 3 +[45] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis, 2023. 5 +[46] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 3 +[47] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 3 +[48] Harsh Rangwani, Aishwarya Agarwal, Kuldeep Kulkarni, R Venkatesh Babu, and Srikrishna Karanam. Crafting parts for expressive object composition. arXiv preprint arXiv:2406.10197, 2024.3 +[49] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 3, 5 +[50] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine + +tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 22500-22510, 2023. 3, 5, 8 +[51] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural information processing systems, 35:36479-36494, 2022. 3 +[52] Rahul Sajnani, Jeroen Vanbaar, Jie Min, Kapil Katyal, and Srinath Sridhar. Geodiffuser: Geometry-based image editing with diffusion models. arXiv preprint arXiv:2404.14403, 2024.3 +[53] Yizhi Song, Zhifei Zhang, Zhe Lin, Scott Cohen, Brian Price, Jianming Zhang, Soo Ye Kim, and Daniel Aliaga. Object-stitch: Object compositing with diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18310-18319, 2023. 2 +[54] Alexander Vilesov, Pradyumna Chari, and Achuta Kadambi. Cg3d: Compositional generation for text-to-3d via gaussian splatting. arXiv preprint arXiv:2311.17907, 2023. 3 +[55] Andrey Voynov, Kfir Aberman, and Daniel Cohen-Or. Sketch-guided text-to-image diffusion models. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-11, 2023. 2, 3 +[56] Andrey Voynov, Kfir Aberman, and Daniel Cohen-Or. Sketch-guided text-to-image diffusion models. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1–11, 2023. 3 +[57] Andrey Voynov, Qinghao Chu, Daniel Cohen-Or, and Kfir Aberman. $\mathfrak{p}+$ : Extended textual conditioning in text-to-image generation. arXiv preprint arXiv:2303.09522, 2023. 3 +[58] Qian Wang, Yiqun Wang, Michael Birsak, and Peter Wonka. Blobgan-3d: A spatially-disentangled 3d-aware generative model for indoor scenes. arXiv preprint arXiv:2303.14706, 2023. 3 +[59] Ruicheng Wang, Jianfeng Xiang, Jiaolong Yang, and Xin Tong. Diffusion models are geometry critics: Single image 3d editing using pre-trained diffusion priors. arXiv preprint arXiv:2403.11503, 2024. 3 +[60] Zhaoning Wang, Ming Li, and Chen Chen. Luciddreaming: Controllable object-centric 3d generation. arXiv preprint arXiv:2312.00588, 2023. 3 +[61] Ziyi Wu, Yulia Rubanova, Rishabh Kabra, Drew A Hudson, Igor Gilitschenski, Yusuf Aytar, Sjoerd van Steenkiste, Kelsey R Allen, and Thomas Kipf. Neural assets: 3d-aware multi-object scene synthesis with image diffusion models. arXiv preprint arXiv:2406.09292, 2024. 3 +[62] Yang Xue, Yuheng Li, Krishna Kumar Singh, and Yong Jae Lee. Giraffe hd: A high-resolution 3d-aware generative model. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18440-18449, 2022. 3 +[63] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to + +image diffusion models. arXiv preprint arXiv:2308.06721, 2023.2,3 +[64] Hong-Xing Yu, Haoyi Duan, Charles Herrmann, William T Freeman, and Jiajun Wu. Wonderworld: Interactive 3d scene generation from a single image. arXiv preprint arXiv:2406.09394, 2024. 3 +[65] Guanqi Zhan, Chuanxia Zheng, Weidi Xie, and Andrew Zisserman. What does stable diffusion know about the 3d scene? arXiv preprint arXiv:2310.06836, 2023. 3 +[66] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 2, 3, 4 +[67] Yuxin Zhang, Weiming Dong, Fan Tang, Nisha Huang, Haibin Huang, Chongyang Ma, Tong-Yee Lee, Oliver Deussen, and Changsheng Xu. Prospect: Prompt spectrum for attribute-aware personalization of diffusion models. ACM Transactions on Graphics (TOG), 42(6):1-14, 2023. 3 +[68] Shijie Zhou, Zhiwen Fan, Dejia Xu, Haoran Chang, Pradyumna Chari, Tejas Bharadwaj, Suya You, Zhangyang Wang, and Achuta Kadambi. Dreamscene360: Unconstrained text-to-3d scene generation with panoramic gaussian splatting. In European Conference on Computer Vision, pages 324–342. Springer, 2025. 3 \ No newline at end of file diff --git a/2025/Compass Control_ Multi Object Orientation Control for Text-to-Image Generation/images.zip b/2025/Compass Control_ Multi Object Orientation Control for Text-to-Image Generation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..1256260ad4b67d6cf04751350ca423cd148f659b --- /dev/null +++ b/2025/Compass Control_ Multi Object Orientation Control for Text-to-Image Generation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:69439b230bb4842b11a1b12a6871d0fee8c8a16855f1f2e500173a5baa1b3fed +size 922545 diff --git a/2025/Compass Control_ Multi Object Orientation Control for Text-to-Image Generation/layout.json b/2025/Compass Control_ Multi Object Orientation Control for Text-to-Image Generation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..7cdd4d240757687bda6af4bee526ed5ff27741a5 --- /dev/null +++ b/2025/Compass Control_ Multi Object Orientation Control for Text-to-Image Generation/layout.json @@ -0,0 +1,9553 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 87, + 103, + 523, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 87, + 103, + 523, + 138 + ], + "spans": [ + { + "bbox": [ + 87, + 103, + 523, + 138 + ], + "type": "text", + "content": "Compass Control: Multi Object Orientation Control for Text-to-Image Generation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 104, + 160, + 194, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 160, + 194, + 175 + ], + "spans": [ + { + "bbox": [ + 104, + 160, + 194, + 175 + ], + "type": "text", + "content": "Rishubh Parihar" + }, + { + "bbox": [ + 104, + 160, + 194, + 175 + ], + "type": "inline_equation", + "content": "^{1*}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 202, + 160, + 299, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 202, + 160, + 299, + 190 + ], + "spans": [ + { + "bbox": [ + 202, + 160, + 299, + 190 + ], + "type": "text", + "content": "Vaibhav Agrawal" + }, + { + "bbox": [ + 202, + 160, + 299, + 190 + ], + "type": "inline_equation", + "content": "^{2*†}" + }, + { + "bbox": [ + 202, + 160, + 299, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 202, + 160, + 299, + 190 + ], + "type": "text", + "content": "IISc Bangalore" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 308, + 160, + 394, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 160, + 394, + 190 + ], + "spans": [ + { + "bbox": [ + 308, + 160, + 394, + 190 + ], + "type": "text", + "content": "Sachidanand VS1 \n2IIT Hyderabad" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 407, + 160, + 505, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 407, + 160, + 505, + 175 + ], + "spans": [ + { + "bbox": [ + 407, + 160, + 505, + 175 + ], + "type": "text", + "content": "R. Venkatesh Babu" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 105, + 229, + 182, + 304 + ], + "blocks": [ + { + "bbox": [ + 99, + 219, + 107, + 227 + ], + "lines": [ + { + "bbox": [ + 99, + 219, + 107, + 227 + ], + "spans": [ + { + "bbox": [ + 99, + 219, + 107, + 227 + ], + "type": "text", + "content": "a)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 119, + 219, + 171, + 228 + ], + "lines": [ + { + "bbox": [ + 119, + 219, + 171, + 228 + ], + "spans": [ + { + "bbox": [ + 119, + 219, + 171, + 228 + ], + "type": "text", + "content": "Stable Diffusion" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 105, + 229, + 182, + 304 + ], + "lines": [ + { + "bbox": [ + 105, + 229, + 182, + 304 + ], + "spans": [ + { + "bbox": [ + 105, + 229, + 182, + 304 + ], + "type": "image", + "image_path": "c814fe204320a3daaa9b2ef57781c1e019d52279a27c6bbc0ba8a35b3a2af27f.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 102, + 305, + 184, + 319 + ], + "lines": [ + { + "bbox": [ + 102, + 305, + 184, + 319 + ], + "spans": [ + { + "bbox": [ + 102, + 305, + 184, + 319 + ], + "type": "text", + "content": "A photo of a SUV facing left and a sedan facing right on a road." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 185, + 229, + 262, + 304 + ], + "blocks": [ + { + "bbox": [ + 194, + 219, + 251, + 228 + ], + "lines": [ + { + "bbox": [ + 194, + 219, + 251, + 228 + ], + "spans": [ + { + "bbox": [ + 194, + 219, + 251, + 228 + ], + "type": "text", + "content": "Compass Control" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 185, + 229, + 262, + 304 + ], + "lines": [ + { + "bbox": [ + 185, + 229, + 262, + 304 + ], + "spans": [ + { + "bbox": [ + 185, + 229, + 262, + 304 + ], + "type": "image", + "image_path": "52ee9b5bcc110321c82b00948eb5a9810421f953f230a1db5da6aec14ab65e82.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 187, + 305, + 259, + 321 + ], + "lines": [ + { + "bbox": [ + 187, + 305, + 259, + 321 + ], + "spans": [ + { + "bbox": [ + 187, + 305, + 259, + 321 + ], + "type": "text", + "content": "'A photo of a [ ] SUV and a [ sedan on a road..." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 306, + 213, + 321, + 227 + ], + "blocks": [ + { + "bbox": [ + 306, + 213, + 321, + 227 + ], + "lines": [ + { + "bbox": [ + 306, + 213, + 321, + 227 + ], + "spans": [ + { + "bbox": [ + 306, + 213, + 321, + 227 + ], + "type": "image", + "image_path": "aa137149fd5a43de9c9084bb41432ebd9ae6bffd5fde6166eba88d1693e22540.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 387, + 213, + 400, + 226 + ], + "blocks": [ + { + "bbox": [ + 387, + 213, + 400, + 226 + ], + "lines": [ + { + "bbox": [ + 387, + 213, + 400, + 226 + ], + "spans": [ + { + "bbox": [ + 387, + 213, + 400, + 226 + ], + "type": "image", + "image_path": "67e13276661e762abca383cc8adc443b2f3e24b9a72cd7059887a9aa63976ae7.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 273, + 228, + 354, + 308 + ], + "blocks": [ + { + "bbox": [ + 273, + 228, + 354, + 308 + ], + "lines": [ + { + "bbox": [ + 273, + 228, + 354, + 308 + ], + "spans": [ + { + "bbox": [ + 273, + 228, + 354, + 308 + ], + "type": "image", + "image_path": "d4a82691815ff9a9092aa2ac781024049d10d1565a7f95da32838b4a1d6944fa.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 290, + 310, + 484, + 325 + ], + "lines": [ + { + "bbox": [ + 290, + 310, + 484, + 325 + ], + "spans": [ + { + "bbox": [ + 290, + 310, + 484, + 325 + ], + "type": "text", + "content": "'A photo of a mother walking with a pram on a snowy street, festive Christmas lights, beautiful winter evening scene'" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 355, + 228, + 434, + 308 + ], + "blocks": [ + { + "bbox": [ + 355, + 228, + 434, + 308 + ], + "lines": [ + { + "bbox": [ + 355, + 228, + 434, + 308 + ], + "spans": [ + { + "bbox": [ + 355, + 228, + 434, + 308 + ], + "type": "image", + "image_path": "0f790c5db6fab1fa2d98b01c27e0b0ead46303d7faf58e190b4fbc75264e0cd7.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 468, + 214, + 484, + 227 + ], + "blocks": [ + { + "bbox": [ + 468, + 214, + 484, + 227 + ], + "lines": [ + { + "bbox": [ + 468, + 214, + 484, + 227 + ], + "spans": [ + { + "bbox": [ + 468, + 214, + 484, + 227 + ], + "type": "image", + "image_path": "cdcc98c2742b63c71ddcd7eed64023733992c98bff86a155e6f32eb3ea6d5210.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 93, + 328, + 243, + 418 + ], + "blocks": [ + { + "bbox": [ + 93, + 328, + 243, + 418 + ], + "lines": [ + { + "bbox": [ + 93, + 328, + 243, + 418 + ], + "spans": [ + { + "bbox": [ + 93, + 328, + 243, + 418 + ], + "type": "image", + "image_path": "3a19bb20023145ea8b6ef1c0c35bace1fd0d3434972a0707782ff0ac1cecfc56.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 118, + 420, + 219, + 435 + ], + "lines": [ + { + "bbox": [ + 118, + 420, + 219, + 435 + ], + "spans": [ + { + "bbox": [ + 118, + 420, + 219, + 435 + ], + "type": "text", + "content": "A photo of a sofa and a teddy in a sunlit greenhouse filled with tropical plants" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 243, + 328, + 394, + 418 + ], + "blocks": [ + { + "bbox": [ + 243, + 328, + 394, + 418 + ], + "lines": [ + { + "bbox": [ + 243, + 328, + 394, + 418 + ], + "spans": [ + { + "bbox": [ + 243, + 328, + 394, + 418 + ], + "type": "image", + "image_path": "96261237efb6e4d8087995463867ab156adbed812a08fd84969e9d98673f6e1d.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 251, + 419, + 387, + 441 + ], + "lines": [ + { + "bbox": [ + 251, + 419, + 387, + 441 + ], + "spans": [ + { + "bbox": [ + 251, + 419, + 387, + 441 + ], + "type": "text", + "content": "A photo of a sparrow and a sparrow and a hen among fallen autumn leaves carpeting the ground, soft afternoon glow around" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 394, + 327, + 515, + 418 + ], + "blocks": [ + { + "bbox": [ + 394, + 327, + 515, + 418 + ], + "lines": [ + { + "bbox": [ + 394, + 327, + 515, + 418 + ], + "spans": [ + { + "bbox": [ + 394, + 327, + 515, + 418 + ], + "type": "image", + "image_path": "0630f7c2043dd9cb060dad580bb0106f0d32322ac232802ea474de41636eb969.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 397, + 420, + 509, + 442 + ], + "lines": [ + { + "bbox": [ + 397, + 420, + 509, + 442 + ], + "spans": [ + { + "bbox": [ + 397, + 420, + 509, + 442 + ], + "type": "text", + "content": "'A photo of a sedan and a sedan and a camel and a camel and a camel in a vast desert with dust rising in the background'" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 93, + 456, + 515, + 542 + ], + "blocks": [ + { + "bbox": [ + 236, + 446, + 406, + 455 + ], + "lines": [ + { + "bbox": [ + 236, + 446, + 406, + 455 + ], + "spans": [ + { + "bbox": [ + 236, + 446, + 406, + 455 + ], + "type": "text", + "content": "Personalization with 3D Orientation Control" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 93, + 456, + 515, + 542 + ], + "lines": [ + { + "bbox": [ + 93, + 456, + 515, + 542 + ], + "spans": [ + { + "bbox": [ + 93, + 456, + 515, + 542 + ], + "type": "image", + "image_path": "e861c8201aa0333269d5d33ffa5f153e83aa05312459040436f78d4426900274.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 254, + 542, + 406, + 550 + ], + "lines": [ + { + "bbox": [ + 254, + 542, + 406, + 550 + ], + "spans": [ + { + "bbox": [ + 254, + 542, + 406, + 550 + ], + "type": "text", + "content": "A photo of " + }, + { + "bbox": [ + 254, + 542, + 406, + 550 + ], + "type": "inline_equation", + "content": "V^{*}" + }, + { + "bbox": [ + 254, + 542, + 406, + 550 + ], + "type": "text", + "content": " car in front of the leaning tower of Pisa in Italy" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 54, + 553, + 555, + 597 + ], + "lines": [ + { + "bbox": [ + 54, + 553, + 555, + 597 + ], + "spans": [ + { + "bbox": [ + 54, + 553, + 555, + 597 + ], + "type": "text", + "content": "Figure 1. We present Compass Control, a method to generate multi-object scenes with orientation control from text-to-image diffusion models. Given a text prompt and an orientation of each object (shown as frustum, the colored face is the forward direction), our method generates scenes that align with both the prompt and specified orientations. Additionally, with a few " + }, + { + "bbox": [ + 54, + 553, + 555, + 597 + ], + "type": "inline_equation", + "content": "(\\approx 10)" + }, + { + "bbox": [ + 54, + 553, + 555, + 597 + ], + "type": "text", + "content": " unposed images of a new object, our model is personalized to generate the object in target orientations." + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 27 + }, + { + "bbox": [ + 152, + 621, + 200, + 634 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 621, + 200, + 634 + ], + "spans": [ + { + "bbox": [ + 152, + 621, + 200, + 634 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 55, + 647, + 296, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 647, + 296, + 684 + ], + "spans": [ + { + "bbox": [ + 55, + 647, + 296, + 684 + ], + "type": "text", + "content": "Existing approaches for controlling text-to-image diffusion models, while powerful, do not allow for explicit 3D object-centric control, such as precise control of object orienta" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 313, + 623, + 555, + 707 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 623, + 555, + 707 + ], + "spans": [ + { + "bbox": [ + 313, + 623, + 555, + 707 + ], + "type": "text", + "content": "tion. In this work, we address the problem of multi-object orientation control in text-to-image diffusion models. This enables the generation of diverse multi-object scenes with precise orientation control for each object. The key idea is to condition the diffusion model with a set of orientation-aware compass tokens, one for each object, along with text tokens. A light-weight encoder network predicts these com" + } + ] + } + ], + "index": 32 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 693, + 132, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 693, + 132, + 703 + ], + "spans": [ + { + "bbox": [ + 66, + 693, + 132, + 703 + ], + "type": "text", + "content": "*equal contribution." + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 67, + 703, + 216, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 216, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 216, + 712 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 67, + 703, + 216, + 712 + ], + "type": "text", + "content": " work done during an internship at VAL, IISc" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "2791" + } + ] + } + ], + "index": 36 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 275 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 275 + ], + "type": "text", + "content": "pass tokens taking object orientation as the input. The model is trained on a synthetic dataset of procedurally generated scenes, each containing one or two 3D assets on a plain background. However, direct training this framework results in poor orientation control as well as leads to entanglement among objects. To mitigate this, we intervene in the generation process and constrain the cross-attention maps of each compass token to its corresponding object regions. The trained model is able to achieve precise orientation control for a) complex objects not seen during training and b) multi-object scenes with more than two objects, indicating strong generalization capabilities. Further, when combined with personalization methods, our method precisely controls the orientation of the new object in diverse contexts. Our method achieves state-of-the-art orientation control and text alignment quantified with extensive evaluations and a user study. Project website" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 281, + 136, + 293 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 281, + 136, + 293 + ], + "spans": [ + { + "bbox": [ + 55, + 281, + 136, + 293 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 295, + 295, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 295, + 295, + 450 + ], + "spans": [ + { + "bbox": [ + 54, + 295, + 295, + 450 + ], + "type": "text", + "content": "Imagine a visual artist aiming to create a scene featuring two cars facing each other. They prompt a text-to-image model with - 'A photo of a sedan facing right and an SUV facing left'. However, as shown in Fig. 1 (a), the generated image may not always accurately capture the intended object orientations. Moreover, relying on text prompts to control object orientation (e.g., 'facing right') is imprecise and requires iterative prompting. Can we design an alternate interface for text-to-image models that accepts target orientation angle as input along with the text prompts? Such an interface will allow for precise orientation control for each object, eliminating the need for iterative prompt adjustments and streamlining the creative process." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 451, + 295, + 642 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 451, + 295, + 642 + ], + "spans": [ + { + "bbox": [ + 55, + 451, + 295, + 642 + ], + "type": "text", + "content": "Several works have been proposed to achieve finer control in text-to-image models such as changing object appearance, scene layouts or image style [5, 8, 11, 21, 22, 43, 53, 55, 63, 66]. While effective for controlling 2D attributes of the image, these methods fail to accurately control 3D attributes. More recently, several methods have been proposed to control 3D properties in text-to-image models, such as camera viewpoint [6, 31], scene lighting [13], or scene layout using 3D bounding boxes [4]. However, these approaches either require dense 3D information, like multi-view images or accurate 3D bounding boxes, or limited to simple single-object scenes (Tab. 1). In this work, we present a novel interface to condition text-to-image diffusion models to generate multi-object scenes with precise 3D orientation control, without the need for multi-view images or 3D bounding boxes." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "type": "text", + "content": "Text-to-image (T2I) diffusion models enable the generation of objects with specific attributes via text prompts (e.g., 'a red car'). Motivated by this, we encode the object orientation as an additional attribute in the text embedding space of the T2I model. Specifically, we introduce a special token, dubbed as compass token (c) along with each object token" + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 315, + 108, + 555, + 156 + ], + "blocks": [ + { + "bbox": [ + 313, + 71, + 553, + 104 + ], + "lines": [ + { + "bbox": [ + 313, + 71, + 553, + 104 + ], + "spans": [ + { + "bbox": [ + 313, + 71, + 553, + 104 + ], + "type": "text", + "content": "Table 1. Comparison with Related works. Our approach uniquely allows for object-centric orientation control and generalize to novel categories without any explicit 3D representation." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 108, + 555, + 156 + ], + "lines": [ + { + "bbox": [ + 315, + 108, + 555, + 156 + ], + "spans": [ + { + "bbox": [ + 315, + 108, + 555, + 156 + ], + "type": "table", + "html": "
MethodCD-360 [31]LooseControl [4]Cont-3D-Words [13]ViewNeTI [6]Ours
InputCam Pose3D boxesCam PoseCam PoseOrientation
3D conditioningExplicitExplicitImplicitImplicitImplicit
Novel classesX✓X✓✓
Multiple Object Control✓✓XX✓
", + "image_path": "94bef5a692f6713c347e41069d4f78be1ac10ec18c02c3f40496e1b32cbb5510.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 160, + 553, + 268 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 160, + 553, + 268 + ], + "spans": [ + { + "bbox": [ + 313, + 160, + 553, + 268 + ], + "type": "text", + "content": "in the prompt (e.g., 'A photo of " + }, + { + "bbox": [ + 313, + 160, + 553, + 268 + ], + "type": "inline_equation", + "content": "\\mathbf{c_1}" + }, + { + "bbox": [ + 313, + 160, + 553, + 268 + ], + "type": "text", + "content": " SUV and " + }, + { + "bbox": [ + 313, + 160, + 553, + 268 + ], + "type": "inline_equation", + "content": "\\mathbf{c_2}" + }, + { + "bbox": [ + 313, + 160, + 553, + 268 + ], + "type": "text", + "content": " sedan on a road'). Each compass token is predicted by a lightweight encoder model taking the prescribed orientation angle as input. This formulation preserves the original interface of the base T2I model and enables precise object-centric orientation control in multi-object scenes. We train the encoder model and fine-tune the denoising U-Net with LoRA [24] on a synthetic dataset of scenes containing one or two 3D assets placed in diverse layouts on an empty floor." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 268, + 553, + 436 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 268, + 553, + 436 + ], + "spans": [ + { + "bbox": [ + 313, + 268, + 553, + 436 + ], + "type": "text", + "content": "We discover that directly injecting compass tokens with the prompt tokens leads to poor orientation control, as the compass token attends to irrelevant image regions, limiting its influence on its corresponding object (Fig. 4(a)). To address this, we propose Coupled Attention Localization (CALL) mechanism, where we constrain the cross-attention maps of the compass token and its corresponding object token within a 2D bounding box. This results in a tight coupling between the two tokens, enabling the compass token to precisely control the orientation for its corresponding object. Additionally, for multi-object scenes, this results in an appropriate binding between each compass token and its corresponding object token, leading to disentangled orientation control of individual objects (Fig. 4(b))." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 437, + 554, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 554, + 544 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 554, + 544 + ], + "type": "text", + "content": "The proposed approach achieves precise orientation control for unseen objects (e.g., pram) and can generalize to scenes with more than two objects, despite being trained on one and two object scenes only (see Fig. 1). Further, given a few unposed images of a real object, we can personalize the model to control the orientation of the new object. We evaluate our method against several baselines, achieving superior performance both quantitatively and in a user study. In summary, our primary contributions are:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 315, + 547, + 553, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 315, + 547, + 553, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 547, + 553, + 594 + ], + "spans": [ + { + "bbox": [ + 315, + 547, + 553, + 594 + ], + "type": "text", + "content": "1. Compass Control - A method for conditioning text-to-image diffusion models on object orientation, enabling precise orientation control for individual objects in multi-object scene generation." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 315, + 594, + 553, + 642 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 594, + 553, + 642 + ], + "spans": [ + { + "bbox": [ + 315, + 594, + 553, + 642 + ], + "type": "text", + "content": "2. Coupled Attention Localization - A mechanism to restrict the influence of the input object orientation to the corresponding object, ensuring effective object-centric orientation control and object disentanglement." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 315, + 643, + 553, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 643, + 553, + 677 + ], + "spans": [ + { + "bbox": [ + 315, + 643, + 553, + 677 + ], + "type": "text", + "content": "3. Strong generalization of Compass Control for precise orientation control to unseen objects and complex multi-object scenes, though trained on simple synthetic scenes." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 315, + 678, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 678, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 315, + 678, + 553, + 714 + ], + "type": "text", + "content": "4. Personalization of Compass Control - Given a few unposed images of a real object, our method can perform orientation control of the new object in diverse contexts." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2792" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 140, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 140, + 83 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 140, + 83 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 87, + 296, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 87, + 296, + 277 + ], + "spans": [ + { + "bbox": [ + 55, + 87, + 296, + 277 + ], + "type": "text", + "content": "Controlled generation in T2I models. Several works have been proposed to achieve fine-grained control in text-to-image diffusion models [47, 49, 51]. Recent works resort to manipulation of the text embeddings [20, 27, 42, 57, 67], or attention maps in the diffusion U-Net [1, 2, 8, 10, 15, 22, 29, 40, 43, 48, 55, 56] for controlling the generated image. Additional encoder models can be trained to condition the T2I models on a new modality such as depth, bounding boxes, or object identity [32, 63, 66]. Another set of works personalizes the diffusion model given with a few subject images [30, 50] enabling generation of the learned subject in different backgrounds. Recent work on guiding diffusion models [19, 35, 41] allows inference time control over the scene contents, allowing the control of object location, appearance, shape, and skeleton pose. However, these controls are limited to 2D." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 282, + 296, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 282, + 296, + 509 + ], + "spans": [ + { + "bbox": [ + 55, + 282, + 296, + 509 + ], + "type": "text", + "content": "3D-aware image editing. Recent works leverage the rich generation capability of the T2I model to perform 3D aware editing [40, 52, 59]. Specifically, they use the input scene depth as an additional input and use it to warp the internal features of the diffusion models. This enables zero-shot geometric 3D edits such as translating or rotating an object. However, these methods are limited to the editing of a single object. Another line of work uses multi-view input images and trains an implicit 3D representation such as a radiance field in the diffusion feature space [31, 44] to perform 3D consistent editing. More recently, few works leverage 3D Gaussian splat representation along with T2I models to perform scene editing [12, 36]. However, the above methods require scene-specific training and require multi-view images of accurate depth maps as input. Another direction explores large-scale training of diffusion model on a specific dataset [25, 37, 61] allowing for 3D scene editing. However, these methods fail to generalize to in-the-wild real-world scenes outside the distribution of training datasets." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 510, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 296, + 713 + ], + "type": "text", + "content": "3D control in generation. Earlier works train generative models from scratch with explicit scene controls such as 3D blobs [58], or radiance fields of individual objects [3, 26, 28, 38, 39, 62] to control the generated scene. Recent works have shown the existence of 3D properties in text-to-image diffusion models [6, 16-18, 65]. This has fuelled the research that leverages this knowledge from T2I models for 3D generation. A set of works lifts this knowledge to 3D by distilling from pre-trained T2I models [33, 46, 54, 60, 64, 68]. Another set of recent works [4, 6, 13] leverage this underlying 3D knowledge for controlling 3D properties of the generated scene with an additional conditioning mechanism. ViewNeTI [6] learns a 3D view token to control the camera view for the task of novel-view synthesis from a dataset of multi-view images. ViewNeTI is natively designed for novel view synthesis but is unable to generate objects in diverse contexts. A continuous word" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 320, + 76, + 550, + 213 + ], + "blocks": [ + { + "bbox": [ + 320, + 76, + 550, + 213 + ], + "lines": [ + { + "bbox": [ + 320, + 76, + 550, + 213 + ], + "spans": [ + { + "bbox": [ + 320, + 76, + 550, + 213 + ], + "type": "image", + "image_path": "8d77d907befbebfa7ecda7d17f7608f7593fad56482dedf4aade4aaa5e9a06e3.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 217, + 555, + 272 + ], + "lines": [ + { + "bbox": [ + 313, + 217, + 555, + 272 + ], + "spans": [ + { + "bbox": [ + 313, + 217, + 555, + 272 + ], + "type": "text", + "content": "Figure 2. Synthetic data generation. We curate 10 diverse 3D assets, and render them in diverse layouts and orientations in Blender [14]. The rendered scenes are augmented with realistic generations from Canny [7] ControlNet [66]. The final dataset consists of one and two object scenes." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 282, + 555, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 282, + 555, + 415 + ], + "spans": [ + { + "bbox": [ + 313, + 282, + 555, + 415 + ], + "type": "text", + "content": "representation is trained in [13] for 3D scene properties such as orientation and lighting on a single 3D mesh. However, both these approaches are limited to simple scenes and control only the global view angle. The closest to our work is [4] that conditions the T2I model on loose depth maps using ControlNet [66]. Loose depth maps are created using 3D object boxes and scene boundaries to enable 3D object-centric control. However, this approach relies on precise 3D boxes, making it cumbersome at inference. In contrast, our method only requires coarse 2D bounding boxes and orientation angles, offering a more user-friendly solution." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 421, + 370, + 432 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 421, + 370, + 432 + ], + "spans": [ + { + "bbox": [ + 313, + 421, + 370, + 432 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 435, + 555, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 435, + 555, + 544 + ], + "spans": [ + { + "bbox": [ + 313, + 435, + 555, + 544 + ], + "type": "text", + "content": "Text-to-Image Diffusion Models. Diffusion models, when directly applied in the pixel space, are computationally expensive due to their iterative nature. To mitigate this, latent diffusion models [49] apply the diffusion process in the smaller resolution latent space of a pretrained autoencoder. Further, the generation can be conditioned on text by injecting text features into the diffusion U-Net with additional cross-attention layers. The cross-attention maps allow for precise control during generation [19, 23]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 552, + 373, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 552, + 373, + 562 + ], + "spans": [ + { + "bbox": [ + 313, + 552, + 373, + 562 + ], + "type": "text", + "content": "3.1. Dataset" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 566, + 554, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 566, + 554, + 674 + ], + "spans": [ + { + "bbox": [ + 313, + 566, + 554, + 674 + ], + "type": "text", + "content": "Synthetic scene generation: We curated a list of 10 diverse 3D assets from the web: ostrich, helicopter, shoe, jeep, teddy bear, lion, sedan, horse, motorbike and sofa to aid in model generalization (see Fig. 2). Our dataset consists of 1000 one-object and 7900 two-object scenes, rendered in Blender [14]. For each image, we save the 2D bounding boxes and the orientations of the objects. The objects are placed at varied locations and in varied orientations to increase layout diversity." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "content": "Augmentations: Directly training on this dataset results in overfitting to the plain background and the black floor, as shown in ablations. To address this, we augment the" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2793" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 76, + 71, + 274, + 274 + ], + "blocks": [ + { + "bbox": [ + 76, + 71, + 274, + 274 + ], + "lines": [ + { + "bbox": [ + 76, + 71, + 274, + 274 + ], + "spans": [ + { + "bbox": [ + 76, + 71, + 274, + 274 + ], + "type": "image", + "image_path": "ccc2c8387f2daaaab6064526d1a3df41c68dfc877b90e2930ff6a7e73c763c89.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 276, + 296, + 342 + ], + "lines": [ + { + "bbox": [ + 55, + 276, + 296, + 342 + ], + "spans": [ + { + "bbox": [ + 55, + 276, + 296, + 342 + ], + "type": "text", + "content": "Figure 3. Compass Control. Given an orientation angle " + }, + { + "bbox": [ + 55, + 276, + 296, + 342 + ], + "type": "inline_equation", + "content": "\\theta_{j}" + }, + { + "bbox": [ + 55, + 276, + 296, + 342 + ], + "type": "text", + "content": ", we project it to a compass token with a lightweight encoder model. The compass tokens are interleaved with the text tokens (as shown in the figure) and passed through the text encoder. The outputs of the text encoder are used to condition the denoising process in the U-Net. We train " + }, + { + "bbox": [ + 55, + 276, + 296, + 342 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 55, + 276, + 296, + 342 + ], + "type": "text", + "content": " and also fine-tune the U-Net using LoRA [24]." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 348, + 296, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 348, + 296, + 456 + ], + "spans": [ + { + "bbox": [ + 55, + 348, + 296, + 456 + ], + "type": "text", + "content": "dataset using ControlNet [66] to place objects in diverse backgrounds while retaining known orientations. For each rendered image, we extract its Canny [7] edge map to condition ControlNet to generate the objects in diverse contexts (for e.g., 'in a garden...', 'near a lake...', etc). This approach preserves object orientations while altering their appearance. We manually filter out the inconsistent augmented images. Further details on the dataset creation process can be found in the Suppl. Sec.H." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 460, + 296, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 460, + 296, + 579 + ], + "spans": [ + { + "bbox": [ + 55, + 460, + 296, + 579 + ], + "type": "text", + "content": "Orientation convention: In this paper, we parameterize orientation with a single angle " + }, + { + "bbox": [ + 55, + 460, + 296, + 579 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 55, + 460, + 296, + 579 + ], + "type": "text", + "content": ", rotation around the up axis in the world coordinate system (pointing towards the sky). We define " + }, + { + "bbox": [ + 55, + 460, + 296, + 579 + ], + "type": "inline_equation", + "content": "\\theta = 0" + }, + { + "bbox": [ + 55, + 460, + 296, + 579 + ], + "type": "text", + "content": " as a reference when the object faces exactly towards the right (e.g., sedan in Fig. 2). We parameterize with a single orientation angle, as most of our objects are land objects, and only this rotation axis results in plausible object orientations. However, our method is not limited to a single orientation angle, and we present results in Suppl. Sec.B for conditioning on three orientation angles." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 590, + 160, + 602 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 590, + 160, + 602 + ], + "spans": [ + { + "bbox": [ + 55, + 590, + 160, + 602 + ], + "type": "text", + "content": "3.2. Compass Control" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "content": "Given a text prompt " + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "content": " consisting of " + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "content": " object names " + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\{o_1,o_2,\\dots,o_N\\}" + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "content": " (e.g., 'A photo of a jeep and a sedan and a horse in a garden') and their corresponding 3D orientation angles " + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\{\\theta_{1},\\theta_{2},\\ldots ,\\theta_{N}\\}" + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "content": " we introduce a set of compass tokens " + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{c}_1,\\mathbf{c}_2,\\dots,\\mathbf{c}_{\\mathbf{N}}\\}" + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "content": " to control the orientations of the respective objects. A compass token " + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_{\\mathbf{n}}" + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "content": " is an embedding in the input space of the text encoder. It is predicted by a lightweight MLP encoder network " + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "content": " (see Fig. 3), which takes as input the object orientation angle " + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\theta_{k}" + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "content": ". The compass tokens are" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 555, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 120 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 120 + ], + "type": "text", + "content": "prepended before their corresponding object tokens (e.g., 'A photo of a c1 jeep and a c2 sedan and a c3 horse in a garden'), and passed through the text encoder. The outputs of the text encoder are used to condition the denoising U-Net." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 121, + 556, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 121, + 556, + 289 + ], + "spans": [ + { + "bbox": [ + 313, + 121, + 556, + 289 + ], + "type": "text", + "content": "However, directly training the above framework on the dataset from Sec. 3.1 fails to learn accurate orientation control. We hypothesize that this is because the added compass tokens are unrestricted and can attend to irrelevant image regions, limiting their influence on their corresponding objects. This is evident in the cross-attention maps for the compass tokens, which are indeed diffused in other image regions (see Fig. 4 (a)). Furthermore, this becomes a severe issue in multi-object scenes, where a single compass token can attend to multiple objects, resulting in the entanglement between different compass tokens(see Fig. 4 (b)). Existing works have shown that cross-attention maps closely control the image layouts [23] in the generated image. Motivated by this, we design a cross-attention localization approach." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 294, + 525, + 306 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 294, + 525, + 306 + ], + "spans": [ + { + "bbox": [ + 313, + 294, + 525, + 306 + ], + "type": "text", + "content": "3.3. Coupled Attention Localization (CALL)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 308, + 556, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 308, + 556, + 499 + ], + "spans": [ + { + "bbox": [ + 313, + 308, + 556, + 499 + ], + "type": "text", + "content": "Our key idea is to constrain the cross-attention maps for both the compass token and the corresponding object token inside a given 2D bounding box. This enables tight association between the object and the compass tokens. Additionally, it enables explicit control over the object location during generation. Specifically, during training, we use the saved object bounding boxes " + }, + { + "bbox": [ + 313, + 308, + 556, + 499 + ], + "type": "inline_equation", + "content": "\\{b_1,\\dots b_N\\}" + }, + { + "bbox": [ + 313, + 308, + 556, + 499 + ], + "type": "text", + "content": " to compute a set of loose square bounding boxes " + }, + { + "bbox": [ + 313, + 308, + 556, + 499 + ], + "type": "inline_equation", + "content": "\\{b_1^l,\\dots b_N^l\\}" + }, + { + "bbox": [ + 313, + 308, + 556, + 499 + ], + "type": "text", + "content": ". The side length " + }, + { + "bbox": [ + 313, + 308, + 556, + 499 + ], + "type": "inline_equation", + "content": "a_{n}" + }, + { + "bbox": [ + 313, + 308, + 556, + 499 + ], + "type": "text", + "content": " of the loose box " + }, + { + "bbox": [ + 313, + 308, + 556, + 499 + ], + "type": "inline_equation", + "content": "b_{n}^{l}" + }, + { + "bbox": [ + 313, + 308, + 556, + 499 + ], + "type": "text", + "content": " is computed as " + }, + { + "bbox": [ + 313, + 308, + 556, + 499 + ], + "type": "inline_equation", + "content": "a_{n} = \\lambda * \\max(h_{b_{n}},w_{b_{n}})" + }, + { + "bbox": [ + 313, + 308, + 556, + 499 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 308, + 556, + 499 + ], + "type": "inline_equation", + "content": "h_{b_n}" + }, + { + "bbox": [ + 313, + 308, + 556, + 499 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 308, + 556, + 499 + ], + "type": "inline_equation", + "content": "w_{b_n}" + }, + { + "bbox": [ + 313, + 308, + 556, + 499 + ], + "type": "text", + "content": " are the height and width of the object box " + }, + { + "bbox": [ + 313, + 308, + 556, + 499 + ], + "type": "inline_equation", + "content": "b_{n}" + }, + { + "bbox": [ + 313, + 308, + 556, + 499 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 308, + 556, + 499 + ], + "type": "inline_equation", + "content": "\\lambda > 1" + }, + { + "bbox": [ + 313, + 308, + 556, + 499 + ], + "type": "text", + "content": " is a padding factor controlling the looseness of the box. Next, we compute a binary mask " + }, + { + "bbox": [ + 313, + 308, + 556, + 499 + ], + "type": "inline_equation", + "content": "m_{n}" + }, + { + "bbox": [ + 313, + 308, + 556, + 499 + ], + "type": "text", + "content": " from the loose bounding box " + }, + { + "bbox": [ + 313, + 308, + 556, + 499 + ], + "type": "inline_equation", + "content": "b_{n}^{l}" + }, + { + "bbox": [ + 313, + 308, + 556, + 499 + ], + "type": "text", + "content": ", such that m is 0 inside the box " + }, + { + "bbox": [ + 313, + 308, + 556, + 499 + ], + "type": "inline_equation", + "content": "b_{n}^{l}" + }, + { + "bbox": [ + 313, + 308, + 556, + 499 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 308, + 556, + 499 + ], + "type": "inline_equation", + "content": "-\\infty" + }, + { + "bbox": [ + 313, + 308, + 556, + 499 + ], + "type": "text", + "content": " outside. We use it to mask out the cross attentions " + }, + { + "bbox": [ + 313, + 308, + 556, + 499 + ], + "type": "inline_equation", + "content": "(\\Psi)" + }, + { + "bbox": [ + 313, + 308, + 556, + 499 + ], + "type": "text", + "content": " of the object token " + }, + { + "bbox": [ + 313, + 308, + 556, + 499 + ], + "type": "inline_equation", + "content": "o_{n}" + }, + { + "bbox": [ + 313, + 308, + 556, + 499 + ], + "type": "text", + "content": " and compass token " + }, + { + "bbox": [ + 313, + 308, + 556, + 499 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_{\\mathbf{n}}" + }, + { + "bbox": [ + 313, + 308, + 556, + 499 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 358, + 500, + 508, + 529 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 500, + 508, + 529 + ], + "spans": [ + { + "bbox": [ + 358, + 500, + 508, + 529 + ], + "type": "interline_equation", + "content": "\\Psi (\\mathbf {c} _ {\\mathbf {n}}) = \\operatorname {s o f t m a x} \\left(m + \\frac {Q K (\\mathbf {c} _ {\\mathbf {n}}) ^ {T}}{\\sqrt {d _ {K}}}\\right)", + "image_path": "b5641de493f5d2c720949807308f44563ed9ebeb81f971629a80c0258ad9db3e.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 359, + 527, + 507, + 556 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 527, + 507, + 556 + ], + "spans": [ + { + "bbox": [ + 359, + 527, + 507, + 556 + ], + "type": "interline_equation", + "content": "\\Psi \\left(\\mathbf {o} _ {\\mathbf {n}}\\right) = \\operatorname {s o f t m a x} \\left(m + \\frac {Q K \\left(o _ {n}\\right) ^ {T}}{\\sqrt {d _ {K}}}\\right)", + "image_path": "b56eb08cba417f5d5a8f04538bd97690ad18c01ad646a54f56914d017120f293.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 558, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 556, + 713 + ], + "type": "text", + "content": "where the query feature " + }, + { + "bbox": [ + 313, + 558, + 556, + 713 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 313, + 558, + 556, + 713 + ], + "type": "text", + "content": " comes from the U-Net features, and the key features " + }, + { + "bbox": [ + 313, + 558, + 556, + 713 + ], + "type": "inline_equation", + "content": "K(\\mathbf{c_n})" + }, + { + "bbox": [ + 313, + 558, + 556, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 558, + 556, + 713 + ], + "type": "inline_equation", + "content": "K(\\mathbf{o_n})" + }, + { + "bbox": [ + 313, + 558, + 556, + 713 + ], + "type": "text", + "content": " come from the respective tokens. This masking operation is performed at all the diffusion timesteps and cross-attention layers. We find that using a loose mask is highly effective, providing greater flexibility during the generation process. This attention localization mechanism is dubbed as Coupled Attention Localization, or CALL for short. Adding CALL mechanism during training and inference has key advantages for composing multi-object scenes: a) Appropriate binding of each compass token " + }, + { + "bbox": [ + 313, + 558, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{c_n}" + }, + { + "bbox": [ + 313, + 558, + 556, + 713 + ], + "type": "text", + "content": " with its corresponding object token " + }, + { + "bbox": [ + 313, + 558, + 556, + 713 + ], + "type": "inline_equation", + "content": "o_n" + }, + { + "bbox": [ + 313, + 558, + 556, + 713 + ], + "type": "text", + "content": " leads to disentangled orientation control of individual objects. b) Constraining the cross-attention for the object to" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2794" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 65, + 73, + 286, + 230 + ], + "blocks": [ + { + "bbox": [ + 64, + 73, + 73, + 82 + ], + "lines": [ + { + "bbox": [ + 64, + 73, + 73, + 82 + ], + "spans": [ + { + "bbox": [ + 64, + 73, + 73, + 82 + ], + "type": "text", + "content": "a)" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 65, + 73, + 286, + 230 + ], + "lines": [ + { + "bbox": [ + 65, + 73, + 286, + 230 + ], + "spans": [ + { + "bbox": [ + 65, + 73, + 286, + 230 + ], + "type": "image", + "image_path": "a639508dea2ee3aa028e0955a5025f3333c966d483b0152aeb67358d45c26dfe.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 234, + 296, + 312 + ], + "lines": [ + { + "bbox": [ + 55, + 234, + 296, + 312 + ], + "spans": [ + { + "bbox": [ + 55, + 234, + 296, + 312 + ], + "type": "text", + "content": "Figure 4. Binding the compass tokens: We visualize the averaged cross attention of the compass token(s) when training with CALL (shown on the left) and without it (shown on the right). CALL localizes the influence of the compass token at the correct regions, which (a) improves orientation control (b) disentangles orientations in multi-object scenes. In (b), " + }, + { + "bbox": [ + 55, + 234, + 296, + 312 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_1" + }, + { + "bbox": [ + 55, + 234, + 296, + 312 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 234, + 296, + 312 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_2" + }, + { + "bbox": [ + 55, + 234, + 296, + 312 + ], + "type": "text", + "content": " are compass tokens for car and motorbike, respectively." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 329, + 296, + 377 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 329, + 296, + 377 + ], + "spans": [ + { + "bbox": [ + 55, + 329, + 296, + 377 + ], + "type": "text", + "content": "kens " + }, + { + "bbox": [ + 55, + 329, + 296, + 377 + ], + "type": "inline_equation", + "content": "o_n" + }, + { + "bbox": [ + 55, + 329, + 296, + 377 + ], + "type": "text", + "content": " to non-overlapping bounding boxes results in disentanglement between the objects themselves (a known issue in T2I models [10]), enabling strong generalization to complex scenes with multiple objects." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 386, + 296, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 386, + 296, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 386, + 296, + 601 + ], + "type": "text", + "content": "Training. We train our encoder model " + }, + { + "bbox": [ + 55, + 386, + 296, + 601 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 55, + 386, + 296, + 601 + ], + "type": "text", + "content": " and fine-tune the denoising U-Net with LoRA [24] on the synthetic dataset from Sec. 3.1. The LoRA training is extremely parameter efficient and preserves the behavior of the base T2I model. We use the proposed CALL mechanism for effective learning of orientation control. However, for the effective working of CALL, the object must be generated within the loose bounding box, as the compass token's influence is restricted to this region. To this end, we first train on simple single-object scenes, to first learn the bounding box adherence for the generated object, and then train on a mix of both single and two-object scenes thereafter. We contrast this two-staged training procedure with the single-stage training at an intermediate training iteration in Fig. 5. The generated objects adhere to the bounding box better in the two-staged training compared to single-stage training. This leads to effective learning of orientation control, as we have shown in ablative experiments (Sec. 4)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 610, + 296, + 706 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 610, + 296, + 706 + ], + "spans": [ + { + "bbox": [ + 55, + 610, + 296, + 706 + ], + "type": "text", + "content": "Inference. During the inference phase, Compass Control expects the text prompt containing the objects, desired orientations, and optional coarse 2D bounding boxes as input. Using loose bounding boxes during training offers a significant advantage here, as we can even spawn non-overlapping boxes heuristically, as shown in the Suppl. Sec.F. We use the text tokens and the compass tokens together to condition the diffusion model." + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 317, + 86, + 375, + 143 + ], + "blocks": [ + { + "bbox": [ + 318, + 71, + 373, + 85 + ], + "lines": [ + { + "bbox": [ + 318, + 71, + 373, + 85 + ], + "spans": [ + { + "bbox": [ + 318, + 71, + 373, + 85 + ], + "type": "text", + "content": "w/o staged training at iteration i" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 317, + 86, + 375, + 143 + ], + "lines": [ + { + "bbox": [ + 317, + 86, + 375, + 143 + ], + "spans": [ + { + "bbox": [ + 317, + 86, + 375, + 143 + ], + "type": "image", + "image_path": "b4fd45a83a5fc1c7452d1ac451c68ae47f06c2c3f5969820f905bc2e4da161b2.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 376, + 85, + 434, + 143 + ], + "blocks": [ + { + "bbox": [ + 377, + 71, + 430, + 85 + ], + "lines": [ + { + "bbox": [ + 377, + 71, + 430, + 85 + ], + "spans": [ + { + "bbox": [ + 377, + 71, + 430, + 85 + ], + "type": "text", + "content": "w staged training at iteration i" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 376, + 85, + 434, + 143 + ], + "lines": [ + { + "bbox": [ + 376, + 85, + 434, + 143 + ], + "spans": [ + { + "bbox": [ + 376, + 85, + 434, + 143 + ], + "type": "image", + "image_path": "4b19886fe092ccf748d04a39a49a0bd8b21fddfd4f73e585cac93d29041aecc8.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 435, + 85, + 493, + 144 + ], + "blocks": [ + { + "bbox": [ + 437, + 71, + 492, + 84 + ], + "lines": [ + { + "bbox": [ + 437, + 71, + 492, + 84 + ], + "spans": [ + { + "bbox": [ + 437, + 71, + 492, + 84 + ], + "type": "text", + "content": "w/o staged training at iteration i" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 435, + 85, + 493, + 144 + ], + "lines": [ + { + "bbox": [ + 435, + 85, + 493, + 144 + ], + "spans": [ + { + "bbox": [ + 435, + 85, + 493, + 144 + ], + "type": "image", + "image_path": "ce7dc394ce4370c8fd7e6c950648b92a0f692d339212e83e28ca2862c58ea199.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 145, + 553, + 168 + ], + "lines": [ + { + "bbox": [ + 313, + 145, + 553, + 168 + ], + "spans": [ + { + "bbox": [ + 313, + 145, + 553, + 168 + ], + "type": "text", + "content": "Figure 5. Staged training results in improved adherence of objects to the bounding boxes, leading to orientation learning." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 494, + 85, + 552, + 144 + ], + "blocks": [ + { + "bbox": [ + 498, + 71, + 552, + 144 + ], + "lines": [ + { + "bbox": [ + 498, + 71, + 552, + 144 + ], + "spans": [ + { + "bbox": [ + 498, + 71, + 552, + 144 + ], + "type": "text", + "content": "w staged trainingat iteration i" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 494, + 85, + 552, + 144 + ], + "lines": [ + { + "bbox": [ + 494, + 85, + 552, + 144 + ], + "spans": [ + { + "bbox": [ + 494, + 85, + 552, + 144 + ], + "type": "image", + "image_path": "19b63436009884d7dbe2f3dd23680ef79d010a533dc7f4b9e22e22fb849b6066.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 314, + 170, + 409, + 181 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 170, + 409, + 181 + ], + "spans": [ + { + "bbox": [ + 314, + 170, + 409, + 181 + ], + "type": "text", + "content": "3.4. Personalization" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 185, + 555, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 185, + 555, + 293 + ], + "spans": [ + { + "bbox": [ + 313, + 185, + 555, + 293 + ], + "type": "text", + "content": "The design of Compass Control as a conditioning mechanism preserves the original capabilities of T2I models, such as personalization. Given a few unposed images of an object " + }, + { + "bbox": [ + 313, + 185, + 555, + 293 + ], + "type": "inline_equation", + "content": "(\\approx 10)" + }, + { + "bbox": [ + 313, + 185, + 555, + 293 + ], + "type": "text", + "content": ", we apply Dreambooth [50] with LoRA and associate a special token " + }, + { + "bbox": [ + 313, + 185, + 555, + 293 + ], + "type": "inline_equation", + "content": "\\hat{u}" + }, + { + "bbox": [ + 313, + 185, + 555, + 293 + ], + "type": "text", + "content": " for the input object using Compass Control's fine-tuned UNet. During inference, we can generate the object in desired orientation " + }, + { + "bbox": [ + 313, + 185, + 555, + 293 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 185, + 555, + 293 + ], + "type": "text", + "content": " with simple prompts; e.g., 'A photo of a " + }, + { + "bbox": [ + 313, + 185, + 555, + 293 + ], + "type": "inline_equation", + "content": "\\mathbf{c}(\\theta)" + }, + { + "bbox": [ + 313, + 185, + 555, + 293 + ], + "type": "inline_equation", + "content": "\\hat{u}" + }, + { + "bbox": [ + 313, + 185, + 555, + 293 + ], + "type": "text", + "content": " car on the beach,' where " + }, + { + "bbox": [ + 313, + 185, + 555, + 293 + ], + "type": "inline_equation", + "content": "\\mathbf{c}(\\theta)" + }, + { + "bbox": [ + 313, + 185, + 555, + 293 + ], + "type": "text", + "content": " is the compass token." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 314, + 300, + 394, + 313 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 300, + 394, + 313 + ], + "spans": [ + { + "bbox": [ + 314, + 300, + 394, + 313 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 314, + 317, + 420, + 330 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 317, + 420, + 330 + ], + "spans": [ + { + "bbox": [ + 314, + 317, + 420, + 330 + ], + "type": "text", + "content": "4.1. Experiment setup" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 332, + 554, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 332, + 554, + 464 + ], + "spans": [ + { + "bbox": [ + 313, + 332, + 554, + 464 + ], + "type": "text", + "content": "Dataset. We use the synthetic dataset from Sec. 3.1 consisting of 8900 rendered images and 6010 augmented images consisting of one and two object scenes for training our model. For quantitative evaluation, we construct a test set of 10 scene prompts having one/two object names generated from ChatGPT [9]. We use a mix of seen objects (horse, jeep, sedan, sofa, teddy, lion) and unseen objects (boat, dolphin, ship, SUV, tractor) in the prompts. For each object and prompt combination, we use a set of 10 randomly sampled orientations. The list of input text prompts and orientations is given in the Supp. Sec.L." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 467, + 554, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 467, + 554, + 575 + ], + "spans": [ + { + "bbox": [ + 313, + 467, + 554, + 575 + ], + "type": "text", + "content": "Implementation Details. We use Stable Diffusion v2.1 [49] as our base T2I model and use LoRA rank 4 for fine-tuning its UNet. We present additional results on Stable Diffusion-XL [45] in Suppl. Sec. D. Our encoder model " + }, + { + "bbox": [ + 313, + 467, + 554, + 575 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 313, + 467, + 554, + 575 + ], + "type": "text", + "content": " is a lightweight MLP: three linear layers with ReLU. We train our model for 25,000 steps with a batch size of 4 with AdamW optimizer and a fixed learning rate of " + }, + { + "bbox": [ + 313, + 467, + 554, + 575 + ], + "type": "inline_equation", + "content": "10^{-4}" + }, + { + "bbox": [ + 313, + 467, + 554, + 575 + ], + "type": "text", + "content": ". We keep the bounding box padding " + }, + { + "bbox": [ + 313, + 467, + 554, + 575 + ], + "type": "inline_equation", + "content": "\\lambda = 1.2" + }, + { + "bbox": [ + 313, + 467, + 554, + 575 + ], + "type": "text", + "content": " for CALL. The training takes 24 hours on a single A6000 GPU." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 578, + 555, + 698 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 578, + 555, + 698 + ], + "spans": [ + { + "bbox": [ + 313, + 578, + 555, + 698 + ], + "type": "text", + "content": "Evaluation Metrics. We evaluate for a) Text Alignment using CLIP similarly; b) %Object Generation - we evaluate the presence of the intended object using Grounding-DINO [34] and threshold on the objectness score for each object in the prompt. c) Angular Error - to evaluate the orientation consistency, we compute the Angular error between the input orientation angle " + }, + { + "bbox": [ + 313, + 578, + 555, + 698 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 578, + 555, + 698 + ], + "type": "text", + "content": " and the orientation of the generated object using a pretrained orientation predictor. Further details about the orientation predictor and the implementation details of the metrics are in the Suppl Sec.I." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 701, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 701, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 701, + 554, + 713 + ], + "type": "text", + "content": "Baselines. As no prior method tackles our same task, we" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2795" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 95, + 72, + 235, + 155 + ], + "blocks": [ + { + "bbox": [ + 95, + 72, + 235, + 155 + ], + "lines": [ + { + "bbox": [ + 95, + 72, + 235, + 155 + ], + "spans": [ + { + "bbox": [ + 95, + 72, + 235, + 155 + ], + "type": "image", + "image_path": "16d98db56fdb42175d98be89bbfde2c0e4bad477190b7eb3a07ee22b6b5af359.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 102, + 154, + 228, + 161 + ], + "lines": [ + { + "bbox": [ + 102, + 154, + 228, + 161 + ], + "spans": [ + { + "bbox": [ + 102, + 154, + 228, + 161 + ], + "type": "text", + "content": "\"A photo of a sedan spinning on a lazy susan in a mall\"" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 236, + 72, + 375, + 155 + ], + "blocks": [ + { + "bbox": [ + 236, + 72, + 375, + 155 + ], + "lines": [ + { + "bbox": [ + 236, + 72, + 375, + 155 + ], + "spans": [ + { + "bbox": [ + 236, + 72, + 375, + 155 + ], + "type": "image", + "image_path": "466f125f22b75be25fec15dcc3c5168f5033660976109c54e685b642e681d63d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 247, + 155, + 363, + 168 + ], + "lines": [ + { + "bbox": [ + 247, + 155, + 363, + 168 + ], + "spans": [ + { + "bbox": [ + 247, + 155, + 363, + 168 + ], + "type": "text", + "content": "'A photo of a pirate ship sailing in a sea with dark clouds near a port'" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 375, + 72, + 515, + 155 + ], + "blocks": [ + { + "bbox": [ + 375, + 72, + 515, + 155 + ], + "lines": [ + { + "bbox": [ + 375, + 72, + 515, + 155 + ], + "spans": [ + { + "bbox": [ + 375, + 72, + 515, + 155 + ], + "type": "image", + "image_path": "6adbadff4797831c8c4ecd920d9ec1294f5d864629358df96ca94fc2ca50894a.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 381, + 155, + 507, + 168 + ], + "lines": [ + { + "bbox": [ + 381, + 155, + 507, + 168 + ], + "spans": [ + { + "bbox": [ + 381, + 155, + 507, + 168 + ], + "type": "text", + "content": "'A photo of a helicopter flying over a magnificent city, the sky is cloudy at sunset'" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 95, + 172, + 235, + 255 + ], + "blocks": [ + { + "bbox": [ + 95, + 172, + 235, + 255 + ], + "lines": [ + { + "bbox": [ + 95, + 172, + 235, + 255 + ], + "spans": [ + { + "bbox": [ + 95, + 172, + 235, + 255 + ], + "type": "image", + "image_path": "86949ffae888969e88bff03d9734f5f35acb62becbeb1607710436824ec28372.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 104, + 254, + 229, + 274 + ], + "lines": [ + { + "bbox": [ + 104, + 254, + 229, + 274 + ], + "spans": [ + { + "bbox": [ + 104, + 254, + 229, + 274 + ], + "type": "text", + "content": "A photo of a jeep and a sedan in front of a serene waterfall with trees scattered around the region, and stones scattered in the water" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 236, + 172, + 375, + 255 + ], + "blocks": [ + { + "bbox": [ + 236, + 172, + 375, + 255 + ], + "lines": [ + { + "bbox": [ + 236, + 172, + 375, + 255 + ], + "spans": [ + { + "bbox": [ + 236, + 172, + 375, + 255 + ], + "type": "image", + "image_path": "728605eb13ba2d178deeb06acb54b6608bd7a0027934c7b4e97d24cfe3377ce7.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 249, + 254, + 363, + 275 + ], + "lines": [ + { + "bbox": [ + 249, + 254, + 363, + 275 + ], + "spans": [ + { + "bbox": [ + 249, + 254, + 363, + 275 + ], + "type": "text", + "content": "A photo of a boat and a yacht floating in a river, featuring a lot of greenery, the sun is setting in the background" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 376, + 172, + 515, + 255 + ], + "blocks": [ + { + "bbox": [ + 376, + 172, + 515, + 255 + ], + "lines": [ + { + "bbox": [ + 376, + 172, + 515, + 255 + ], + "spans": [ + { + "bbox": [ + 376, + 172, + 515, + 255 + ], + "type": "image", + "image_path": "068b79fafe5baea9de5179375ea056940852b4a9ed558b62e43d02baf66dc8a9.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 391, + 254, + 490, + 268 + ], + "lines": [ + { + "bbox": [ + 391, + 254, + 490, + 268 + ], + "spans": [ + { + "bbox": [ + 391, + 254, + 490, + 268 + ], + "type": "text", + "content": "'A man and a woman talking to each other sitting in a garden'" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 96, + 278, + 512, + 383 + ], + "blocks": [ + { + "bbox": [ + 96, + 278, + 512, + 383 + ], + "lines": [ + { + "bbox": [ + 96, + 278, + 512, + 383 + ], + "spans": [ + { + "bbox": [ + 96, + 278, + 512, + 383 + ], + "type": "image", + "image_path": "b00f6b28095813cb09ba4cc410b5b96e69d5892d894cdda34421156a0c5421bc.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 104, + 385, + 171, + 407 + ], + "lines": [ + { + "bbox": [ + 104, + 385, + 171, + 407 + ], + "spans": [ + { + "bbox": [ + 104, + 385, + 171, + 407 + ], + "type": "text", + "content": "A photo of a dog and a dog and a pram in a rustic living room" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 180, + 384, + 258, + 418 + ], + "lines": [ + { + "bbox": [ + 180, + 384, + 258, + 418 + ], + "spans": [ + { + "bbox": [ + 180, + 384, + 258, + 418 + ], + "type": "text", + "content": "A photo of a sedan and a sedan and a SUV moving down a modern highway with a lush green forest on the sides at sunset" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 267, + 384, + 340, + 411 + ], + "lines": [ + { + "bbox": [ + 267, + 384, + 340, + 411 + ], + "spans": [ + { + "bbox": [ + 267, + 384, + 340, + 411 + ], + "type": "text", + "content": "'A photo of a bicycle and a bicycle on a wooden bridge, a boat and a boat on the shiny waters on a sunny afternoon" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 351, + 384, + 422, + 418 + ], + "lines": [ + { + "bbox": [ + 351, + 384, + 422, + 418 + ], + "spans": [ + { + "bbox": [ + 351, + 384, + 422, + 418 + ], + "type": "text", + "content": "'A photo of a shoe and a shoe and a teddy and a sofa and a teddy in a lavish living room with soft lighting and wooden flooring'" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 433, + 384, + 507, + 419 + ], + "lines": [ + { + "bbox": [ + 433, + 384, + 507, + 419 + ], + "spans": [ + { + "bbox": [ + 433, + 384, + 507, + 419 + ], + "type": "text", + "content": "A photo of a SUV and a SUV and a SUV and a sedan and a sedan and a sedan at a city intersection at night, featuring tall buildings" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 421, + 555, + 455 + ], + "lines": [ + { + "bbox": [ + 55, + 421, + 555, + 455 + ], + "spans": [ + { + "bbox": [ + 55, + 421, + 555, + 455 + ], + "type": "text", + "content": "Figure 6. Main Results. Compass Control generates complex scenes aligned with the text prompts and the orientations (shown as frustum, the colored face is forward direction). It generalizes well to unseen object categories - pirate ship, boat, yacht, bicycle, pram, dog, and even human. Further, it generates high-quality compositions of several objects, despite being trained only on one and two object scenes." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 464, + 296, + 704 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 464, + 296, + 704 + ], + "spans": [ + { + "bbox": [ + 55, + 464, + 296, + 704 + ], + "type": "text", + "content": "compare against methods that allow for either camera pose control or 3D object pose control in text-to-image models: a) Continuous 3D Words (Cont-3D-Words) [13]: Following their exact setup, we train a 3D word for controlling the object orientation on renderings of a single 3D asset - Sedan and its ControlNet augmentations. b) ViewNeTI [6]: In contrast to Cont-3D-Words, ViewNeTI allows for training on multiple 3D assets; for fair evaluation, we train ViewNeTI on our training dataset and condition the T2I on object orientation instead of 3D camera pose. Notably, both of these methods are limited to a global view control. Hence, we evaluate on only single-object scenes. c) LooseControl [4], allows for multi-object control by conditioning on loose depth maps (Fig. 7) formed by 3D object boxes. We use template 3D bounding boxes for each test object and place them in the scene, with random orientation and location (similar to Sec. 3.1). For a fair comparison, we use the 2D boxes corresponding to the 3D boxes in our outputs. Notably, LooseControl does not take exact orientation as input as a 180 flipped box also has the same loose depth;" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 464, + 555, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 464, + 555, + 523 + ], + "spans": [ + { + "bbox": [ + 313, + 464, + 555, + 523 + ], + "type": "text", + "content": "we consider this while computing the Angular error. Further, it requires the user to provide accurate 3D bounding boxes during inference, which is cumbersome, whereas our method requires only loose 2D boxes. Additional baseline details are in Suppl. Sec.J." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 530, + 400, + 542 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 530, + 400, + 542 + ], + "spans": [ + { + "bbox": [ + 313, + 530, + 400, + 542 + ], + "type": "text", + "content": "4.2. Main Results" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 312, + 545, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 545, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 312, + 545, + 555, + 715 + ], + "type": "text", + "content": "Qualitative results. We present our method's results in Fig. 6. Our method is able to generate complex multi-object scenes with precise orientation control of individual objects, even though it is trained with single and two object scenes. Further, it is able to generalize well to challenging unseen objects such as humans and prams. Interestingly, there was no water-based subject in the training dataset, yet our method can achieve precise orientation control for a ship, yacht, and boat. These strong generalization capabilities of Compass Control can be attributed to effective attention constraining with CALL and diversity in the 3D assets used for training. The conditioning mechanism of Compass Control is generalized, and we present results for jointly controlling all three orientation angles, camera" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "2796" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 61, + 64, + 288, + 150 + ], + "blocks": [ + { + "bbox": [ + 61, + 64, + 288, + 150 + ], + "lines": [ + { + "bbox": [ + 61, + 64, + 288, + 150 + ], + "spans": [ + { + "bbox": [ + 61, + 64, + 288, + 150 + ], + "type": "table", + "html": "
Single objectText Align. ↓% Obj. Generated ↑Angular Err. ↓
ViewNeTI [6]22.120.9200.596
Cont-3D-words [13]29.880.7320.509
LooseControl [4]31.600.6560.385
Ours32.980.9680.198
Multiple objectText Align. ↓% Obj. Generated ↑Angular Err. ↓
LooseControl [4]31.730.7780.372
Ours33.930.9640.215
", + "image_path": "9212754278149d3b1978daa1833367cce4c428f8bb0913b94bdc55b13cfc7834.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 153, + 295, + 187 + ], + "lines": [ + { + "bbox": [ + 55, + 153, + 295, + 187 + ], + "spans": [ + { + "bbox": [ + 55, + 153, + 295, + 187 + ], + "type": "text", + "content": "Table 2. Quantitative comparison. We compute Text Alignment, using CLIP, % of correct subject generation and Angular error between the predicted and input orientations." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 197, + 251, + 208 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 197, + 251, + 208 + ], + "spans": [ + { + "bbox": [ + 55, + 197, + 251, + 208 + ], + "type": "text", + "content": "elevation, and object scale in Suppl. Sec.B & C." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 209, + 296, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 209, + 296, + 520 + ], + "spans": [ + { + "bbox": [ + 55, + 209, + 296, + 520 + ], + "type": "text", + "content": "Baseline comparison. We compare our method to all three baselines on single-object scenes and additionally include multi-object scene comparison with LooseControl in Fig. 7 and Tab. 2. Cont-3D-words morphs the generated objects into a sedan shape seen during training and generates washed-out backgrounds. This results in poor text alignment and a lower percentage of intended object generation. ViewNeTI can generate better object shapes in a given orientation; however, it overfits to the black backgrounds seen during training, leading to poor text alignment. This is primarily because ViewNeTI does not accept ControlNet augmentations in its original form as it is designed for novel view synthesis. LooseControl generates realistic single-subject scenes following the given text prompt. However, in some cases, the object orientation is not followed (e.g., sofa, teddy). For a multi-object generation (Fig. 7b), LooseControl either misses the object during generation (horse in the second column and cars in the last column) or distorts the object shapes. LooseControl distorts the object to a box like appearance (columns 3 and 4) as it learns LoRA over the original depth-conditioned ControlNet, which follows the depth input closely, resulting in an inferior object generation score. Further, though we adjust for a 180-degree flip in computing Angular error for LooseControl, our method achieves significantly lower Angular error, demonstrating a strong orientation adherence." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 527, + 129, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 527, + 129, + 540 + ], + "spans": [ + { + "bbox": [ + 55, + 527, + 129, + 540 + ], + "type": "text", + "content": "4.3. User study" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 544, + 296, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 544, + 296, + 665 + ], + "spans": [ + { + "bbox": [ + 55, + 544, + 296, + 665 + ], + "type": "text", + "content": "We conducted a user study with 57 participants to compare all methods on text alignment, object quality, and orientation consistency. Users rated 90 image pairs for single-object scenes and 30 for multi-object scenes, choosing the better image in each pair, sampled from our method and baselines. In total, we obtained 5130 ratings for single-object and 1710 for two-object scenes. Results (Fig. 8) show users preferred our method overall, with LooseControl scoring well in text alignment for single objects but falling short across all metrics in multi-object scenes." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 672, + 151, + 683 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 672, + 151, + 683 + ], + "spans": [ + { + "bbox": [ + 55, + 672, + 151, + 683 + ], + "type": "text", + "content": "4.4. Personalization" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": "We present personalization results in Fig. 9. With only 10 unposed images of an object, our method can gener" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 318, + 73, + 553, + 300 + ], + "blocks": [ + { + "bbox": [ + 318, + 73, + 553, + 300 + ], + "lines": [ + { + "bbox": [ + 318, + 73, + 553, + 300 + ], + "spans": [ + { + "bbox": [ + 318, + 73, + 553, + 300 + ], + "type": "image", + "image_path": "51728fef24abdbe76c2b183c50b23e3a170f4bf96768ae09fe5750e3c8c9885e.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 318, + 302, + 553, + 456 + ], + "blocks": [ + { + "bbox": [ + 318, + 302, + 553, + 456 + ], + "lines": [ + { + "bbox": [ + 318, + 302, + 553, + 456 + ], + "spans": [ + { + "bbox": [ + 318, + 302, + 553, + 456 + ], + "type": "image", + "image_path": "df2781b37d444df1c0ae2fe625895a43f6de79f264bd379fcb89e8695f363758.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 460, + 555, + 536 + ], + "lines": [ + { + "bbox": [ + 313, + 460, + 555, + 536 + ], + "spans": [ + { + "bbox": [ + 313, + 460, + 555, + 536 + ], + "type": "text", + "content": "Figure 7. Qualitative Comparison. We compare our method against three baselines. Cont-3D-words [13] does not generate the intended object whereas View-NeTI [6] generates objects in plain backgrounds. LooseControl [4] generates realistic compositions but does not follow the input orientation well. In contrast, our method aligns with the input text prompt and follows the input orientation, while generating realistic scenes." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 322, + 539, + 553, + 639 + ], + "blocks": [ + { + "bbox": [ + 322, + 539, + 553, + 639 + ], + "lines": [ + { + "bbox": [ + 322, + 539, + 553, + 639 + ], + "spans": [ + { + "bbox": [ + 322, + 539, + 553, + 639 + ], + "type": "image", + "image_path": "896492711fe23ef86613d8624248e230d2614f17aef752fe60d0a16e1feb20c7.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 645, + 555, + 679 + ], + "lines": [ + { + "bbox": [ + 313, + 645, + 555, + 679 + ], + "spans": [ + { + "bbox": [ + 313, + 645, + 555, + 679 + ], + "type": "text", + "content": "Figure 8. User study. We compare all methods on the three image metrics. Each bar indicates the fraction of people that preferred our result (gray) vs the baseline (other color)." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "content": "ate the object with precise orientation control in various contexts. Furthermore, we can jointly optimize Dream" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2797" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": "booth [50] LoRA weights for two objects (e.g., a chair and a teddy bear), enabling multi-object personalization with object-centric orientation control. We compare our method with CD-360 [31] and achieve comparable performance. Unlike CD-360 [31], which requires " + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "inline_equation", + "content": "\\approx 100" + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": " object images with camera pose, we require only a few unposed images, making it more convenient and user-friendly." + } + ] + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 59, + 168, + 293, + 392 + ], + "blocks": [ + { + "bbox": [ + 59, + 168, + 293, + 392 + ], + "lines": [ + { + "bbox": [ + 59, + 168, + 293, + 392 + ], + "spans": [ + { + "bbox": [ + 59, + 168, + 293, + 392 + ], + "type": "image", + "image_path": "c0debfd42436386ef4388565d187f9463232cb558d84d837b29ae960340f2c35.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 393, + 296, + 460 + ], + "lines": [ + { + "bbox": [ + 55, + 393, + 296, + 460 + ], + "spans": [ + { + "bbox": [ + 55, + 393, + 296, + 460 + ], + "type": "text", + "content": "Figure 9. Personalization. Given a few " + }, + { + "bbox": [ + 55, + 393, + 296, + 460 + ], + "type": "inline_equation", + "content": "(\\approx 10)" + }, + { + "bbox": [ + 55, + 393, + 296, + 460 + ], + "type": "text", + "content": " unposed images of an object, our method can personalize the diffusion models and allow for orientation control of the new object. Notably, our method can also generate scenes with two personalized objects with precise orientation control. Additionally, we compare our method with CustomDiffusion-360 [31] that uses " + }, + { + "bbox": [ + 55, + 393, + 296, + 460 + ], + "type": "inline_equation", + "content": "\\approx 100" + }, + { + "bbox": [ + 55, + 393, + 296, + 460 + ], + "type": "text", + "content": " posed images." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 475, + 124, + 485 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 475, + 124, + 485 + ], + "spans": [ + { + "bbox": [ + 55, + 475, + 124, + 485 + ], + "type": "text", + "content": "4.5. Ablations" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 489, + 296, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 489, + 296, + 537 + ], + "spans": [ + { + "bbox": [ + 55, + 489, + 296, + 537 + ], + "type": "text", + "content": "We present the results of the ablation study in Fig. 10 on generated scenes with 1, 3, and 5 objects. We focus on three key design choices and generate scenes with a variable number of objects:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 540, + 296, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 540, + 296, + 612 + ], + "spans": [ + { + "bbox": [ + 55, + 540, + 296, + 612 + ], + "type": "text", + "content": "Staged training: Training Compass Control in a single stage results in poor adherence to the bounding boxes. This especially affects complex multi-object layouts, where some objects tend to leak outside their box and suppress the generation of neighboring objects. This results in generating a lesser number of objects in the scene." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 614, + 296, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 614, + 296, + 687 + ], + "spans": [ + { + "bbox": [ + 55, + 614, + 296, + 687 + ], + "type": "text", + "content": "CALL is crucial for accurate orientation control. Without it the compass token attends to irrelevant image regions, resulting in poor orientation control. Further, without CALL the object tokens entangle with each other during generation (a known issue in T2I models [10]). This results in generating a lesser number of objects in the scene." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "Augmentations: Without the ControlNet augmentations, the model overfits the training backgrounds, resulting in" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 318, + 74, + 552, + 284 + ], + "blocks": [ + { + "bbox": [ + 318, + 74, + 552, + 284 + ], + "lines": [ + { + "bbox": [ + 318, + 74, + 552, + 284 + ], + "spans": [ + { + "bbox": [ + 318, + 74, + 552, + 284 + ], + "type": "image", + "image_path": "56e2ac254e87aba415b3bf01ce1a0dc51a0b9512e0fa69f5924fb75239c335bc.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 285, + 553, + 307 + ], + "lines": [ + { + "bbox": [ + 313, + 285, + 553, + 307 + ], + "spans": [ + { + "bbox": [ + 313, + 285, + 553, + 307 + ], + "type": "text", + "content": "Figure 10. Ablation studies. We show the impact of several design choices of our approach. See Sec. 4.5 for details." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 317, + 554, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 317, + 554, + 341 + ], + "spans": [ + { + "bbox": [ + 313, + 317, + 554, + 341 + ], + "type": "text", + "content": "black backgrounds. Hence, ControlNet augmentation is necessary to generate objects in diverse contexts." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 346, + 466, + 358 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 346, + 466, + 358 + ], + "spans": [ + { + "bbox": [ + 313, + 346, + 466, + 358 + ], + "type": "text", + "content": "5. Conclusion and Discussion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 364, + 554, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 364, + 554, + 388 + ], + "spans": [ + { + "bbox": [ + 313, + 364, + 554, + 388 + ], + "type": "text", + "content": "Limitations. Our method struggles to control the orientation of objects that are occluded or have significant overlap." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 388, + 423, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 388, + 423, + 506 + ], + "spans": [ + { + "bbox": [ + 313, + 388, + 423, + 506 + ], + "type": "text", + "content": "In these scenarios, the model either fails to generate one of the objects or mixes the attributes between objects (Fig. 11). Further, our single-angle orientation representation is too simplistic for modeling complex non-rigid objects such as humans." + } + ] + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 440, + 388, + 548, + 496 + ], + "blocks": [ + { + "bbox": [ + 440, + 388, + 548, + 496 + ], + "lines": [ + { + "bbox": [ + 440, + 388, + 548, + 496 + ], + "spans": [ + { + "bbox": [ + 440, + 388, + 548, + 496 + ], + "type": "image", + "image_path": "8c6868b60fe5e8347718177f95b2e181fd16618ad7437bbcfde826792fc3e003.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 453, + 496, + 542, + 505 + ], + "lines": [ + { + "bbox": [ + 453, + 496, + 542, + 505 + ], + "spans": [ + { + "bbox": [ + 453, + 496, + 542, + 505 + ], + "type": "text", + "content": "Figure 11. Failure Cases" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 508, + 554, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 508, + 554, + 543 + ], + "spans": [ + { + "bbox": [ + 313, + 508, + 554, + 543 + ], + "type": "text", + "content": "In this work, we focus on presenting a generalized framework to condition text-to-image models on 3D controls that can be adapted easily for other representations." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 546, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 546, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 546, + 555, + 713 + ], + "type": "text", + "content": "Conclusion. In this work, we propose a method to condition pre-trained text-to-image models with 3D object orientation while preserving its rich image-generation capabilities. We train the conditioning module on a small synthetic dataset via staged training and involving attention regularization. These modifications enable strong generalization of the model, allowing for precise orientation control for unseen object categories and individual objects in a multi-object scene. Further, it can be seamlessly integrated with personalization methods to achieve orientation control of personalized objects. This work is a testament that text-to-image diffusion models innately have some form of 3D understanding, and interesting 3D controls can be obtained with appropriate conditionings." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "2798" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 118 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 118 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 118 + ], + "type": "text", + "content": "Acknowledgements. We thank Srinjay Sarkar, Abhijnya Bhat, and Tejan Karmali for thoroughly reviewing the manuscript. This work is supported by Meesho and PMRF by the Government of India." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 140, + 115, + 153 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 140, + 115, + 153 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 115, + 153 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 160, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 160, + 294, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 160, + 294, + 204 + ], + "spans": [ + { + "bbox": [ + 61, + 160, + 294, + 204 + ], + "type": "text", + "content": "[1] Omri Avrahami, Rinon Gal, Gal Chechik, Ohad Fried, Dani Lischinski, Arash Vahdat, and Weili Nie. Diffuhaul: A training-free method for object dragging in images. arXiv preprint arXiv:2406.01594, 2024. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 206, + 294, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 206, + 294, + 237 + ], + "spans": [ + { + "bbox": [ + 62, + 206, + 294, + 237 + ], + "type": "text", + "content": "[2] Omer Bar-Tal, Lior Yariv, Yaron Lipman, and Tali Dekel. Multidiffusion: Fusing diffusion paths for controlled image generation. 2023. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 240, + 294, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 240, + 294, + 304 + ], + "spans": [ + { + "bbox": [ + 62, + 240, + 294, + 304 + ], + "type": "text", + "content": "[3] Miguel Angel Bautista, Pengsheng Guo, Samira Abnar, Walter Talbott, Alexander Toshev, Zhuoyuan Chen, Laurent Dinh, Shuangfei Zhai, Hanlin Goh, Daniel Ulbricht, et al. Gaudi: A neural architect for immersive 3d scene generation. Advances in Neural Information Processing Systems, 35:25102-25116, 2022. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 308, + 294, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 308, + 294, + 350 + ], + "spans": [ + { + "bbox": [ + 62, + 308, + 294, + 350 + ], + "type": "text", + "content": "[4] Shariq Farooq Bhat, Niloy Mitra, and Peter Wonka. Loosecontrol: Lifting controlnet for generalized depth conditioning. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 2, 3, 6, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 353, + 294, + 417 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 353, + 294, + 417 + ], + "spans": [ + { + "bbox": [ + 62, + 353, + 294, + 417 + ], + "type": "text", + "content": "[5] Manuel Brack, Felix Friedrich, Katharia Kornmeier, Linoy Tsaban, Patrick Schramowski, Kristian Kersting, and Apolinário Passos. Ledits++: Limitless image editing using text-to-image models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8861-8870, 2024. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 420, + 294, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 420, + 294, + 463 + ], + "spans": [ + { + "bbox": [ + 62, + 420, + 294, + 463 + ], + "type": "text", + "content": "[6] James Burgess, Kuan-Chieh Wang, and Serena Yeung. Viewpoint textual inversion: Unleashing novel view synthesis with pretrained 2d diffusion models. arXiv preprint arXiv:2309.07986, 2023. 2, 3, 6, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 465, + 294, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 465, + 294, + 497 + ], + "spans": [ + { + "bbox": [ + 62, + 465, + 294, + 497 + ], + "type": "text", + "content": "[7] John Canny. A computational approach to edge detection. IEEE Transactions on pattern analysis and machine intelligence, (6):679-698, 1986. 3, 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 500, + 294, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 500, + 294, + 563 + ], + "spans": [ + { + "bbox": [ + 62, + 500, + 294, + 563 + ], + "type": "text", + "content": "[8] Mingdeng Cao, Xintao Wang, Zhongang Qi, Ying Shan, Xiaohu Qie, and Yinqiang Zheng. Masactrl: Tuning-free mutual self-attention control for consistent image synthesis and editing. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22560-22570, 2023. 2, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 567, + 291, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 567, + 291, + 578 + ], + "spans": [ + { + "bbox": [ + 62, + 567, + 291, + 578 + ], + "type": "text", + "content": "[9] Chatgpt. Chatgpt. In https://chat.openai.com-chat, 2022.5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 579, + 295, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 579, + 295, + 622 + ], + "spans": [ + { + "bbox": [ + 57, + 579, + 295, + 622 + ], + "type": "text", + "content": "[10] Hila Chefer, Yuval Alaluf, Yael Vinker, Lior Wolf, and Daniel Cohen-Or. Attend-and-excite: Attention-based semantic guidance for text-to-image diffusion models. ACM Transactions on Graphics (TOG), 42(4):1-10, 2023. 3, 5, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 624, + 295, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 624, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 57, + 624, + 295, + 677 + ], + "type": "text", + "content": "[11] Xi Chen, Lianghua Huang, Yu Liu, Yujun Shen, Deli Zhao, and Hengshuang Zhao. Anydoor: Zero-shot object-level image customization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6593-6602, 2024. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 681, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 681, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 681, + 295, + 712 + ], + "type": "text", + "content": "[12] Yiwen Chen, Zilong Chen, Chi Zhang, Feng Wang, Xiaofeng Yang, Yikai Wang, Zhongang Cai, Lei Yang, Huaping Liu, and Guosheng Lin. Gaussianeditor: Swift and control" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 317, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "text", + "content": "able 3d editing with gaussian splatting. corr abs/2311.14521 (2023), 2023. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 97, + 553, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 97, + 553, + 161 + ], + "spans": [ + { + "bbox": [ + 317, + 97, + 553, + 161 + ], + "type": "text", + "content": "[13] Ta-Ying Cheng, Matheus Gadelha, Thibault Groueix, Matthew Fisher, Radomir Mech, Andrew Markham, and Niki Trigoni. Learning continuous 3d words for text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6753-6762, 2024. 2, 3, 6, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 164, + 553, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 164, + 553, + 195 + ], + "spans": [ + { + "bbox": [ + 317, + 164, + 553, + 195 + ], + "type": "text", + "content": "[14] Blender Online Community. Blender - a 3D modelling and rendering package. Blender Foundation, Stichting Blender Foundation, Amsterdam, 2018. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 198, + 553, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 198, + 553, + 240 + ], + "spans": [ + { + "bbox": [ + 317, + 198, + 553, + 240 + ], + "type": "text", + "content": "[15] Omer Dahary, Or Patashnik, Kfir Aberman, and Daniel Cohen-Or. Be yourself: Bounded attention for multi-subject text-to-image generation. arXiv preprint arXiv:2403.16990, 2(5), 2024. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 243, + 553, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 243, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 317, + 243, + 553, + 285 + ], + "type": "text", + "content": "[16] Ankit Dhiman, Manan Shah, Rishubh Parihar, Yash Bhalgat, Lokesh R Boregowda, and R Venkatesh Babu. Reflecting reality: Enabling diffusion models to produce faithful mirror reflections. arXiv preprint arXiv:2409.14677, 2024. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 288, + 553, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 288, + 553, + 330 + ], + "spans": [ + { + "bbox": [ + 317, + 288, + 553, + 330 + ], + "type": "text", + "content": "[17] Xiaodan Du, Nicholas Kolkin, Greg Shakhnarovich, and Anand Bhattachad. Generative models: What do they know? do they know things? let's find out! arXiv preprint arXiv:2311.17137, 2023." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 333, + 553, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 333, + 553, + 398 + ], + "spans": [ + { + "bbox": [ + 317, + 333, + 553, + 398 + ], + "type": "text", + "content": "[18] Mohamed El Banani, Amit Raj, Kevis-Kokitsi Maninis, Abhishek Kar, Yuanzhen Li, Michael Rubinstein, Deqing Sun, Leonidas Guibas, Justin Johnson, and Varun Jampani. Probing the 3d awareness of visual foundation models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21795-21806, 2024. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 399, + 553, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 399, + 553, + 443 + ], + "spans": [ + { + "bbox": [ + 317, + 399, + 553, + 443 + ], + "type": "text", + "content": "[19] Dave Epstein, Allan Jabri, Ben Poole, Alexei Efros, and Aleksander Holynski. Diffusion self-guidance for controllable image generation. Advances in Neural Information Processing Systems, 36:16222-16239, 2023. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 445, + 553, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 445, + 553, + 488 + ], + "spans": [ + { + "bbox": [ + 317, + 445, + 553, + 488 + ], + "type": "text", + "content": "[20] Songwei Ge, Taesung Park, Jun-Yan Zhu, and Jia-Bin Huang. Expressive text-to-image generation with rich text. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7545-7556, 2023. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 491, + 553, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 491, + 553, + 533 + ], + "spans": [ + { + "bbox": [ + 317, + 491, + 553, + 533 + ], + "type": "text", + "content": "[21] Daniel Geng and Andrew Owens. Motion guidance: Diffusion-based image editing with differentiable motion estimators. In The Twelfth International Conference on Learning Representations. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 535, + 553, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 535, + 553, + 578 + ], + "spans": [ + { + "bbox": [ + 317, + 535, + 553, + 578 + ], + "type": "text", + "content": "[22] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 2, 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 580, + 553, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 580, + 553, + 622 + ], + "spans": [ + { + "bbox": [ + 317, + 580, + 553, + 622 + ], + "type": "text", + "content": "[23] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 3, 4" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 624, + 553, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 624, + 553, + 668 + ], + "spans": [ + { + "bbox": [ + 317, + 624, + 553, + 668 + ], + "type": "text", + "content": "[24] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 2, 4, 5" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 670, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 670, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 317, + 670, + 553, + 712 + ], + "type": "text", + "content": "[25] Allan Jabri, Sjoerd van Steenkiste, Emiel Hoogeboom, Mehdi SM Sajjadi, and Thomas Kipf. Dorsal: Diffusion for object-centric representations of scenes et. al. arXiv preprint arXiv:2306.08068, 2023. 3" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "2799" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 139 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 139 + ], + "type": "text", + "content": "[26] Kunal Kathare, Ankit Dhiman, K Vikas Gowda, Siddharth Aravindan, Shubham Monga, Basavaraja Shanthappa Vandrotti, and Lokesh R Boregowda. Instructive3d: Editing large reconstruction models with text instructions. In Proceedings of the Winter Conference on Applications of Computer Vision (WACV), pages 3246-3256, 2025. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 140, + 294, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 140, + 294, + 194 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 294, + 194 + ], + "type": "text", + "content": "[27] Bahjat Kawar, Shiran Zada, Oran Lang, Omer Tov, Huiwen Chang, Tali Dekel, Inbar Mosseri, and Michal Irani. Imagc: Text-based real image editing with diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6007-6017, 2023. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 195, + 294, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 195, + 294, + 239 + ], + "spans": [ + { + "bbox": [ + 56, + 195, + 294, + 239 + ], + "type": "text", + "content": "[28] Hyunsu Kim, Gayoung Lee, Yunjay Choi, Jin-Hwa Kim, and Jun-Yan Zhu. 3d-aware blending with generative nerfs. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22906-22918, 2023. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 239, + 294, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 239, + 294, + 282 + ], + "spans": [ + { + "bbox": [ + 56, + 239, + 294, + 282 + ], + "type": "text", + "content": "[29] Yunji Kim, Jiyoung Lee, Jin-Hwa Kim, Jung-Woo Ha, and Jun-Yan Zhu. Dense text-to-image generation with attention modulation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7701-7711, 2023. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 284, + 294, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 284, + 294, + 337 + ], + "spans": [ + { + "bbox": [ + 56, + 284, + 294, + 337 + ], + "type": "text", + "content": "[30] Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization of text-to-image diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1931-1941, 2023. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 338, + 294, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 338, + 294, + 381 + ], + "spans": [ + { + "bbox": [ + 56, + 338, + 294, + 381 + ], + "type": "text", + "content": "[31] Nupur Kumari, Grace Su, Richard Zhang, Taesung Park, Eli Shechtman, and Jun-Yan Zhu. Customizing text-to-image diffusion with camera viewpoint control. arXiv preprint arXiv:2404.12333, 2024. 2, 3, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 383, + 294, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 383, + 294, + 437 + ], + "spans": [ + { + "bbox": [ + 56, + 383, + 294, + 437 + ], + "type": "text", + "content": "[32] Yuheng Li, Haotian Liu, Qingyang Wu, Fangzhou Mu, Jianwei Yang, Jianfeng Gao, Chunyuan Li, and Yong Jae Lee. Gligen: Open-set grounded text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22511-22521, 2023. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 437, + 294, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 437, + 294, + 503 + ], + "spans": [ + { + "bbox": [ + 56, + 437, + 294, + 503 + ], + "type": "text", + "content": "[33] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 300–309, 2023. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 504, + 294, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 504, + 294, + 557 + ], + "spans": [ + { + "bbox": [ + 56, + 504, + 294, + 557 + ], + "type": "text", + "content": "[34] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 559, + 294, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 559, + 294, + 613 + ], + "spans": [ + { + "bbox": [ + 56, + 559, + 294, + 613 + ], + "type": "text", + "content": "[35] Grace Luo, Trevor Darrell, Oliver Wang, Dan B Goldman, and Aleksander Holynski. Readout guidance: Learning control from diffusion features. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8217-8227, 2024. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 614, + 294, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 614, + 294, + 667 + ], + "spans": [ + { + "bbox": [ + 56, + 614, + 294, + 667 + ], + "type": "text", + "content": "[36] Guan Luo, Tian-Xing Xu, Ying-Tian Liu, Xiao-Xiong Fan, Fang-Lue Zhang, and Song-Hai Zhang. 3d gaussian editing with a single image. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 6627-6636, 2024. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 669, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 669, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 669, + 294, + 713 + ], + "type": "text", + "content": "[37] Oscar Michel, Anand Bhattachad, Eli VanderBilt, Ranjay Krishna, Aniruddha Kembhavi, and Tanmay Gupta. Object 3dit: Language-guided 3d-aware image editing. Advances in Neural Information Processing Systems, 36, 2024. 3" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 127 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 127 + ], + "type": "text", + "content": "[38] Thu Nguyen-Phuoc, Chuan Li, Lucas Theis, Christian Richardt, and Yong-Liang Yang. Hologan: Unsupervised learning of 3d representations from natural images. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7588–7597, 2019. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 129, + 553, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 553, + 182 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 553, + 182 + ], + "type": "text", + "content": "[39] Michael Niemeyer and Andreas Geiger. Giraffe: Representing scenes as compositional generative neural feature fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11453-11464, 2021. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 183, + 553, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 183, + 553, + 247 + ], + "spans": [ + { + "bbox": [ + 316, + 183, + 553, + 247 + ], + "type": "text", + "content": "[40] Karran Pandey, Paul Guerrero, Matheus Gadelha, Yannick Hold-Geoffroy, Karan Singh, and Niloy J Mitra. Diffusion handles enabling 3d edits for diffusion models by lifting activations to 3d. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7695–7704, 2024. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 250, + 553, + 313 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 250, + 553, + 313 + ], + "spans": [ + { + "bbox": [ + 316, + 250, + 553, + 313 + ], + "type": "text", + "content": "[41] Rishubh Parihar, Abhijnya Bhat, Abhipsa Basu, Saswat Mallick, Jogendra Nath Kundu, and R Venkatesh Babu. Balancing act: distribution-guided debiasing in diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6668-6678, 2024. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 316, + 553, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 316, + 553, + 370 + ], + "spans": [ + { + "bbox": [ + 316, + 316, + 553, + 370 + ], + "type": "text", + "content": "[42] Rishubh Parihar, VS Sachidanand, Sabariswaran Mani, Tejan Karmali, and R Venkatesh Babu. Precisecontrol: Enhancing text-to-image diffusion models with fine-grained attribute control. In European Conference on Computer Vision, pages 469-487. Springer, 2024. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 372, + 553, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 372, + 553, + 425 + ], + "spans": [ + { + "bbox": [ + 316, + 372, + 553, + 425 + ], + "type": "text", + "content": "[43] Or Patashnik, Daniel Garibi, Idan Azuri, Hadar Averbuch-Elor, and Daniel Cohen-Or. Localizing object-level shape variations with text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 23051–23061, 2023. 2, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 426, + 553, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 426, + 553, + 468 + ], + "spans": [ + { + "bbox": [ + 316, + 426, + 553, + 468 + ], + "type": "text", + "content": "[44] Or Patashnik, Rinon Gal, Daniel Cohen-Or, Jun-Yan Zhu, and Fernando De la Torre. Consolidating attention features for multi-view image editing. arXiv preprint arXiv:2402.14792, 2024. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 471, + 553, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 471, + 553, + 514 + ], + "spans": [ + { + "bbox": [ + 316, + 471, + 553, + 514 + ], + "type": "text", + "content": "[45] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis, 2023. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 515, + 553, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 515, + 553, + 547 + ], + "spans": [ + { + "bbox": [ + 316, + 515, + 553, + 547 + ], + "type": "text", + "content": "[46] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 548, + 553, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 548, + 553, + 590 + ], + "spans": [ + { + "bbox": [ + 316, + 548, + 553, + 590 + ], + "type": "text", + "content": "[47] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 592, + 553, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 592, + 553, + 635 + ], + "spans": [ + { + "bbox": [ + 316, + 592, + 553, + 635 + ], + "type": "text", + "content": "[48] Harsh Rangwani, Aishwarya Agarwal, Kuldeep Kulkarni, R Venkatesh Babu, and Srikrishna Karanam. Crafting parts for expressive object composition. arXiv preprint arXiv:2406.10197, 2024.3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 636, + 553, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 636, + 553, + 690 + ], + "spans": [ + { + "bbox": [ + 316, + 636, + 553, + 690 + ], + "type": "text", + "content": "[49] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 3, 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 692, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 553, + 712 + ], + "type": "text", + "content": "[50] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2800" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 296, + 715 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 76, + 73, + 296, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 73, + 296, + 116 + ], + "spans": [ + { + "bbox": [ + 76, + 73, + 296, + 116 + ], + "type": "text", + "content": "tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 22500-22510, 2023. 3, 5, 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 295, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 295, + 185 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 295, + 185 + ], + "type": "text", + "content": "[51] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural information processing systems, 35:36479-36494, 2022. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 185, + 295, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 185, + 295, + 228 + ], + "spans": [ + { + "bbox": [ + 56, + 185, + 295, + 228 + ], + "type": "text", + "content": "[52] Rahul Sajnani, Jeroen Vanbaar, Jie Min, Kapil Katyal, and Srinath Sridhar. Geodiffuser: Geometry-based image editing with diffusion models. arXiv preprint arXiv:2404.14403, 2024.3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 230, + 295, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 230, + 295, + 285 + ], + "spans": [ + { + "bbox": [ + 56, + 230, + 295, + 285 + ], + "type": "text", + "content": "[53] Yizhi Song, Zhifei Zhang, Zhe Lin, Scott Cohen, Brian Price, Jianming Zhang, Soo Ye Kim, and Daniel Aliaga. Object-stitch: Object compositing with diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18310-18319, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 285, + 295, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 285, + 295, + 319 + ], + "spans": [ + { + "bbox": [ + 56, + 285, + 295, + 319 + ], + "type": "text", + "content": "[54] Alexander Vilesov, Pradyumna Chari, and Achuta Kadambi. Cg3d: Compositional generation for text-to-3d via gaussian splatting. arXiv preprint arXiv:2311.17907, 2023. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 320, + 295, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 320, + 295, + 363 + ], + "spans": [ + { + "bbox": [ + 56, + 320, + 295, + 363 + ], + "type": "text", + "content": "[55] Andrey Voynov, Kfir Aberman, and Daniel Cohen-Or. Sketch-guided text-to-image diffusion models. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-11, 2023. 2, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 365, + 295, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 365, + 295, + 408 + ], + "spans": [ + { + "bbox": [ + 56, + 365, + 295, + 408 + ], + "type": "text", + "content": "[56] Andrey Voynov, Kfir Aberman, and Daniel Cohen-Or. Sketch-guided text-to-image diffusion models. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1–11, 2023. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 410, + 295, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 410, + 295, + 453 + ], + "spans": [ + { + "bbox": [ + 56, + 410, + 295, + 453 + ], + "type": "text", + "content": "[57] Andrey Voynov, Qinghao Chu, Daniel Cohen-Or, and Kfir Aberman. " + }, + { + "bbox": [ + 56, + 410, + 295, + 453 + ], + "type": "inline_equation", + "content": "\\mathfrak{p}+" + }, + { + "bbox": [ + 56, + 410, + 295, + 453 + ], + "type": "text", + "content": ": Extended textual conditioning in text-to-image generation. arXiv preprint arXiv:2303.09522, 2023. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 455, + 295, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 455, + 295, + 498 + ], + "spans": [ + { + "bbox": [ + 56, + 455, + 295, + 498 + ], + "type": "text", + "content": "[58] Qian Wang, Yiqun Wang, Michael Birsak, and Peter Wonka. Blobgan-3d: A spatially-disentangled 3d-aware generative model for indoor scenes. arXiv preprint arXiv:2303.14706, 2023. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 500, + 295, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 500, + 295, + 544 + ], + "spans": [ + { + "bbox": [ + 56, + 500, + 295, + 544 + ], + "type": "text", + "content": "[59] Ruicheng Wang, Jianfeng Xiang, Jiaolong Yang, and Xin Tong. Diffusion models are geometry critics: Single image 3d editing using pre-trained diffusion priors. arXiv preprint arXiv:2403.11503, 2024. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 545, + 295, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 545, + 295, + 578 + ], + "spans": [ + { + "bbox": [ + 56, + 545, + 295, + 578 + ], + "type": "text", + "content": "[60] Zhaoning Wang, Ming Li, and Chen Chen. Luciddreaming: Controllable object-centric 3d generation. arXiv preprint arXiv:2312.00588, 2023. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 579, + 295, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 579, + 295, + 634 + ], + "spans": [ + { + "bbox": [ + 56, + 579, + 295, + 634 + ], + "type": "text", + "content": "[61] Ziyi Wu, Yulia Rubanova, Rishabh Kabra, Drew A Hudson, Igor Gilitschenski, Yusuf Aytar, Sjoerd van Steenkiste, Kelsey R Allen, and Thomas Kipf. Neural assets: 3d-aware multi-object scene synthesis with image diffusion models. arXiv preprint arXiv:2406.09292, 2024. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 635, + 295, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 295, + 689 + ], + "type": "text", + "content": "[62] Yang Xue, Yuheng Li, Krishna Kumar Singh, and Yong Jae Lee. Giraffe hd: A high-resolution 3d-aware generative model. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18440-18449, 2022. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 690, + 295, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 690, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 56, + 690, + 295, + 715 + ], + "type": "text", + "content": "[63] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 342 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "image diffusion models. arXiv preprint arXiv:2308.06721, 2023.2,3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "type": "text", + "content": "[64] Hong-Xing Yu, Haoyi Duan, Charles Herrmann, William T Freeman, and Jiajun Wu. Wonderworld: Interactive 3d scene generation from a single image. arXiv preprint arXiv:2406.09394, 2024. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 141, + 553, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 553, + 174 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 553, + 174 + ], + "type": "text", + "content": "[65] Guanqi Zhan, Chuanxia Zheng, Weidi Xie, and Andrew Zisserman. What does stable diffusion know about the 3d scene? arXiv preprint arXiv:2310.06836, 2023. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 175, + 553, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 553, + 219 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 553, + 219 + ], + "type": "text", + "content": "[66] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 2, 3, 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 220, + 553, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 220, + 553, + 275 + ], + "spans": [ + { + "bbox": [ + 316, + 220, + 553, + 275 + ], + "type": "text", + "content": "[67] Yuxin Zhang, Weiming Dong, Fan Tang, Nisha Huang, Haibin Huang, Chongyang Ma, Tong-Yee Lee, Oliver Deussen, and Changsheng Xu. Prospect: Prompt spectrum for attribute-aware personalization of diffusion models. ACM Transactions on Graphics (TOG), 42(6):1-14, 2023. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 276, + 553, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 276, + 553, + 342 + ], + "spans": [ + { + "bbox": [ + 316, + 276, + 553, + 342 + ], + "type": "text", + "content": "[68] Shijie Zhou, Zhiwen Fan, Dejia Xu, Haoran Chang, Pradyumna Chari, Tejas Bharadwaj, Suya You, Zhangyang Wang, and Achuta Kadambi. Dreamscene360: Unconstrained text-to-3d scene generation with panoramic gaussian splatting. In European Conference on Computer Vision, pages 324–342. Springer, 2025. 3" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "2801" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Complementary Advantages_ Exploiting Cross-Field Frequency Correlation for NIR-Assisted Image Denoising/c9ee5d0d-57af-4b5c-aacf-8ca5596906ca_content_list.json b/2025/Complementary Advantages_ Exploiting Cross-Field Frequency Correlation for NIR-Assisted Image Denoising/c9ee5d0d-57af-4b5c-aacf-8ca5596906ca_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..d59ded4916c20ef7b04c2050fafd2bed4478e532 --- /dev/null +++ b/2025/Complementary Advantages_ Exploiting Cross-Field Frequency Correlation for NIR-Assisted Image Denoising/c9ee5d0d-57af-4b5c-aacf-8ca5596906ca_content_list.json @@ -0,0 +1,1604 @@ +[ + { + "type": "text", + "text": "Complementary Advantages: Exploiting Cross-Field Frequency Correlation for NIR-Assisted Image Denoising", + "text_level": 1, + "bbox": [ + 98, + 128, + 898, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yuchen Wang $^{1}$ Hongyuan Wang $^{1}$ Lizhi Wang $^{2,3*}$", + "bbox": [ + 297, + 202, + 764, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xin Wang $^{1}$ Lin Zhu $^{1}$ Wanxuan Lu $^{4}$ Hua Huang $^{2,3}$", + "bbox": [ + 276, + 220, + 785, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 School of Computer Science and Technology, Beijing Institute of Technology", + "bbox": [ + 217, + 239, + 844, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ School of Artificial Intelligence, Beijing Normal University", + "bbox": [ + 287, + 257, + 772, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3 Engineering Research Center of Intelligent Technology and Educational Application, Ministry of Education", + "bbox": [ + 99, + 273, + 960, + 292 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{4}$ Aerospace Information Research Institute, Chinese Academy of Sciences", + "bbox": [ + 235, + 292, + 826, + 309 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 344, + 326, + 359 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Existing single-image denoising algorithms often struggle to restore details when dealing with complex noisy images. The introduction of near-infrared (NIR) images offers new possibilities for RGB image denoising. However, due to the inconsistency between NIR and RGB images, the existing works still struggle to balance the contributions of two fields in the process of image fusion. In response to this, in this paper, we develop a cross-field Frequency Correlation Exploiting Network (FCENet) for NIR-assisted image denoising. We first propose the frequency correlation prior based on an in-depth statistical frequency analysis of NIR-RGB image pairs. The prior reveals the complementary correlation of NIR and RGB images in the frequency domain. Leveraging frequency correlation prior, we then establish a frequency learning framework composed of Frequency Dynamic Selection Mechanism (FDSM) and Frequency Exhaustive Fusion Mechanism (FEFM). FDSM dynamically selects complementary information from NIR and RGB images in the frequency domain, and FEFM strengthens the control of common and differential features during the fusion process of NIR and RGB features. Extensive experiments on simulated and real data validate that the proposed method outperforms other state-of-the-art methods. The code will be released at https://github.com/yuchenwang815/FCENet.", + "bbox": [ + 86, + 376, + 485, + 755 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 786, + 220, + 801 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Due to the inherent physical limitations of digital imaging devices, the captured images are often affected by various types of noise, which have a considerable impact on applications such as 24-hour surveillance, autonomous driving,", + "bbox": [ + 89, + 811, + 483, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/cd04717a2e57779998dce53b0b02024e66048ef79e423846adb28afdd725e4c7.jpg", + "image_caption": [ + "Figure 1. Comparisons of PSNR, FLOPs, and Parameters on the DVD dataset [18] are presented. Green circles represent the single-image denoising algorithms, and blue circles represent the NIR-assisted image denoising algorithms. The circle radius represents the number of parameters. The proposed method (red circles) achieves the superior performance while maintaining efficiency." + ], + "image_footnote": [], + "bbox": [ + 519, + 347, + 901, + 566 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "and smartphone photography. Although the development of deep learning techniques [24, 36] has brought forth a variety of image denoising methods [1, 3, 13, 28, 47, 51], existing algorithms still find it difficult to deal with complex noisy images, especially in low-light environment. Restoring clean detailed information from noisy images that lack substantial information is highly ill-posed. Therefore, it is necessary to introduce supplementary information from new fields. Fortunately, NIR images with high Signal to Noise Ratio (SNR) can be captured at a low cost and are suitable for everyday photography. Thus, NIR-assisted image denoising [26, 37, 46] has become a promising solution for image denoising.", + "bbox": [ + 509, + 672, + 906, + 868 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "However, it is challenging to effectively leverage information from different fields to achieve noise reduction. Due", + "bbox": [ + 511, + 869, + 906, + 898 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding Author: Lizhi Wang (wanglizhi@bnu.edu.cn)", + "bbox": [ + 107, + 886, + 437, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "12679", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b49cca02186ed0919b0d221791bed6ad52902dfa423cbcdb42edcb991e259751.jpg", + "image_caption": [ + "Figure 2. Visual comparisons on the challenging noisy RGB-NIR image pairs. Pay attention to the differences in structural and color information between two fields. Our method produces a better denoising result with clear details and fewer artifacts." + ], + "image_footnote": [], + "bbox": [ + 93, + 89, + 482, + 194 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "to the influence of the inherent reflective spectra of objects, there are inconsistencies between NIR and RGB images. As shown in Fig. 2, firstly, there is a significant difference in color and brightness information between two fields. Secondly, some structural texture of RGB image disappears in NIR image, and additional artifacts may appear. The color and structural inconsistencies between NIR and RGB images pose a significant challenge for the task of NIR-assisted image denoising.", + "bbox": [ + 88, + 263, + 482, + 400 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In existing researches, some methods attempt to directly integrate inconsistent features from two fields [31, 32, 43, 44]. However, these methods indiscriminately fuse crossfield information, relying on the fitting capabilities of neural networks to learn the fusion of NIR and RGB images under spatial inconsistency, which often leads to suboptimal denoising results and lower computational efficiency. Some methods reinforce common features from NIR and RGB images and diminish the weight of inconsistent features during the fusion [11, 18, 19, 46]. These image fusion schemes typically only enhance the common information in NIR and RGB during fusion, but some differential information such as the sharp textures in NIR and the color information in RGB is discarded.", + "bbox": [ + 89, + 401, + 482, + 611 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we develop a cross-field Frequency Correlation Exploiting Network (FCENet) for NIR-assisted image denoising. We conduct a series of analyses on the characteristics of NIR and RGB images in the frequency domain and obtain a key observation: in the same scene, the similarity between noisy RGB images and clean RGB (noise-free RGB) images decreases from low to high frequency, whereas the similarity between NIR images and clean RGB images increases from low to high frequency. This significant cross-field frequency correlation prior redefines crossfield fusion as a process of frequency domain adaptive selection, motivating us to subsequently establish detailed frequency domain learning to precisely control the restoration of texture details and colors.", + "bbox": [ + 89, + 613, + 482, + 823 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Specifically, the frequency domain fusion network consists of two mechanisms: Frequency Dynamic Selection Mechanism (FDSM) and Frequency Exhaustive Fusion Mechanism (FEFM). In order to comprehensively extract useful features from inconsistent images, the FDSM is for-", + "bbox": [ + 89, + 825, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "mulated based on dynamic filters under the guidance of frequency domain prior. The FDSM efficiently screens complementary features from NIR and RGB in the frequency domain, providing high quality material for subsequent image fusion. Then, the FEFM is formulated to ensure thorough frequency domain fusion. The FEFM integrates local and long-range correlations across fields to fully extract the common feature from cross-field information, and employs a differential cross-attention approach to complement some key differential high frequency features in NIR features.", + "bbox": [ + 511, + 90, + 903, + 241 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The proposed FCENet can utilize complementary information from cross-field images. Extensive experiments on simulated and real data validate that our method outperforms other state-of-the-art methods in terms of image quality and computational efficiency. The main contributions of this work are summarized as follows:", + "bbox": [ + 511, + 242, + 906, + 332 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We establish the frequency correlation prior between NIR and RGB images through a series of frequency domain analyses of two fields, which prompts the integration of complementary features in the frequency domain.", + "- We propose an efficient frequency dynamic selection mechanism based on the frequency correlation prior to identify which useful low-frequency and high frequency information should be retained in NIR and RGB features.", + "- We develop an exhaustive frequency domain fusion mechanism that models the local similarity and long-range correlation in the frequency domain of NIR and RGB images, and supplements some critical high frequency differential features." + ], + "bbox": [ + 513, + 334, + 903, + 531 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 546, + 653, + 561 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Single Image Denoising", + "text_level": 1, + "bbox": [ + 511, + 571, + 730, + 588 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Image denoising [12, 29] is a classic task in low-level vision. In recent years, deep learning has greatly improved the accuracy and efficiency of denoising [2, 7, 14, 22, 23, 35]. DnCNN [50] uses a simple network with batch normalization and residual learning, which outperforms traditional denoising methods. MPRNet [47] designs a multi-stage restoration framework, and Restormer [48] proposes an efficient transformer framework for high-resolution images. Xformer [49] proposes a transformer framework that combines spatial self-attention and channel self-attention. However, despite their strong denoising capabilities, single-image denoising still inevitably loses a lot of spatial information, leading to overly smooth edge textures in the denoised images.", + "bbox": [ + 511, + 594, + 906, + 806 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. NIR-Assisted Image Restoration", + "text_level": 1, + "bbox": [ + 511, + 816, + 797, + 834 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To achieve better denoising effects, researchers have attempted to use NIR images to assist in RGB image denoising [8, 26, 38, 46]. Compared with single-image recovery, NIR images can help restore the details of degraded im", + "bbox": [ + 511, + 839, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "12680", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ages [37, 43]. In early research, Yan et al. [46] proposes a scale map from the perspective of gradient inconsistency to find universal usable edges and smooth transitions in NIR images. With the proliferation of deep learning, CUNet [11] uses a convolutional sparse coding model to extract common features from cross-field images. FGDNet [31] and SANet [32] fuse the features of two fields in the frequency domain. MNNet [44] proposes an observation model that takes into account the modality gap between the target and guiding images. DVN [18] integrates structural inconsistency priors into deep features, specifically strengthening the common deep features in NIR and RGB images. NAID [43] proposes Selective Fusion Module (SFM), which can be plugged into advanced image restoration networks. However, under extreme noisy conditions, existing NIR-assisted denoising methods fail to balance color accuracy, detail retention, and artifact suppression.", + "bbox": [ + 89, + 90, + 483, + 349 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Frequency Domain Image Restoration", + "text_level": 1, + "bbox": [ + 89, + 356, + 419, + 372 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Frequency analysis in digital signal processing emphasizes the frequency properties of signals [33, 45]. Recently, frequency domain analysis has shown great potential in improving the performance of various learning-based image restoration frameworks [17, 42, 53]. In [21], An efficient estimation of scaled dot-product attention based on a frequency domain self-attention solver (FSAS) is proposed. HDNet [16] proposes a high-resolution dual-domain learning network. SFNet [10] uses filters to dynamically decouple feature mappings into different frequency components and adaptively extracts useful features. In this paper, we emphasize the importance of frequency domain information for the research on fusion under inconsistency and design a comprehensive frequency domain fusion scheme for this task to ensure the utilization of complementary information from NIR and RGB images.", + "bbox": [ + 89, + 378, + 483, + 619 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 632, + 181, + 648 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we first introduce the cross-field frequency correlation prior of NIR and RGB images. Based on the cross-field frequency correlation prior, a detailed frequency domain scheme is formulated, including the Frequency Dynamic Selection Mechanism (FDSM) and the Frequency Exhaustive Fusion Mechanism (FEFM) for thorough frequency exploitation.", + "bbox": [ + 89, + 657, + 483, + 765 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Cross-Field Frequency Correlation Prior", + "text_level": 1, + "bbox": [ + 89, + 772, + 441, + 789 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To address the fusion of NIR and RGB images under spatial inconsistency, it is imperative to distinguish which information in NIR and RGB images is useful and which is redundant or even erroneous for denoising. We randomly select 30 pairs of NIR images, noisy RGB images, and clean RGB images from the existing RGB-NIR dataset and perform fourier transform on them. To facilitate the study of the", + "bbox": [ + 89, + 794, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/096f74f99473b4f900382ea287a60b9aef5d7e855526fdb576d89c7bf4322dd3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 522, + 90, + 707, + 215 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/1abe5b305419a3f6cb834643f540880b69dc4499cac145398b035d0757f87b68.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 707, + 90, + 895, + 215 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/0dbbf62a8d3d8cf87517605b8790a7f54bfe8587b0a7801c430547f3026b957c.jpg", + "image_caption": [ + "Figure 3. The analysis of NIR and RGB frequency correlation. (a) The frequency correlation between noisy RGB and clean RGB, with the horizontal axis indicating the cutoff frequency of the high-pass filter. Different colored curves represent different scenarios. (b) The frequency correlation between NIR and clean RGB. (c) The visualization results of images output by a fixed frequency filter, with the first row showing high frequency (HF) part and the second row showing low frequency (LF) part." + ], + "image_footnote": [], + "bbox": [ + 522, + 218, + 653, + 339 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/4af54a2411b9aa6fd12ed836ff2212bb0e02b2889fb2388cd6ba9f4cdd159c59.jpg", + "image_caption": [ + "(c)" + ], + "image_footnote": [], + "bbox": [ + 656, + 218, + 771, + 339 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/07037fa6455ef9225e2af54ebc34984d995d8d629f0641df51950c43c02d529e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 774, + 218, + 890, + 339 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "correlation between NIR and RGB in the frequency domain, we apply a high-pass filter to the aforementioned three kinds of frequency domain images and convert the filtered results back to the spatial domain. We separately computed the Structural Similarity (SSIM) [40] between NIR/noisy RGB images and clean RGB images at different cutoff frequencies as a measure of frequency correlation $S$ , as shown in Fig. 3a and Fig. 3b. The process can be described as", + "bbox": [ + 511, + 474, + 906, + 597 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nS = \\operatorname {S S I M} \\left(\\mathcal {F} ^ {- 1} \\left(\\mathrm {H} \\left(F _ {\\text {t a r g e t}}\\right)\\right), \\mathcal {F} ^ {- 1} \\left(\\mathrm {H} \\left(F _ {\\text {g t}}\\right)\\right)\\right) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 555, + 603, + 906, + 622 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathcal{F}^{-1}(\\cdot)$ denote 2D-IDFT operations, $F_{target}$ and $F_{gt}$ represent NIR or noisy RGB and clean RGB after 2D-DFT transformation. $\\mathrm{H}(\\cdot)$ represents the filtering of the input image with a high-pass filter. The analysis results indicate that, in the main energy frequency, the image similarity between the input noisy RGB and clean RGB images decreases progressively as the frequency increases, while the image similarity between the input NIR image and clean RGB image increases from low to high frequency. Fig. 3c also shows that for noisy RGB images, their high frequency information is mainly composed of noise and lacks detail information, while the low frequency information more faithfully reflects the color in the original image, and NIR images are just the opposite. These cross-field Frequency correlation priors indicate that it is simple and efficient to distinguish useful information from NIR and RGB images in the frequency domain, which inspires us to propose a comprehensive frequency domain learning framework.", + "bbox": [ + 511, + 627, + 908, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "12681", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/5866955e7c4a4770d9b89dadcfc5461df22c2aca17b6c21291642ac9b6cbe000.jpg", + "image_caption": [ + "Noisy RGB (Input)" + ], + "image_footnote": [], + "bbox": [ + 96, + 90, + 251, + 160 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/d920fb41101c44be5427913d7679d3f21b9d747a8dd7e9c2785e9acc8d0aeef4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 321, + 92, + 475, + 159 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/3f21648d2b5ed62c367a1019c5f30fd90f41e5ca6c6d539c52fb454fed0eb311.jpg", + "image_caption": [ + "Clean NIR (Input)", + "Figure 4. The overall architecture of the proposed cross-field Frequency Correlation Exploiting Network (FCENet). In the first stage, pre-denoising is performed. In the second stage, NIR image is introduced for NIR-assisted image denoising." + ], + "image_footnote": [], + "bbox": [ + 94, + 175, + 478, + 406 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Overview Framework", + "text_level": 1, + "bbox": [ + 89, + 483, + 295, + 498 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The proposed frequency domain fusion framework, as depicted in Fig. 4, includes two stages for the gradual restoration of images. Both stages incorporate an encoder-decoder architecture based on the U-Net [30], featuring two downsampling layers and two upsampling layers. Skip connections are employed between the encoder and decoder features. A Supervised Attention Module (SAM) [47] is utilized to connect features from the first stage to the second stage for further processing.", + "bbox": [ + 89, + 505, + 483, + 641 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the first stage, the input noisy RGB image undergoes preliminary denoising to reduce the modality gap with the NIR image. In the second stage, we perform multi-scale fusion of the encoded features of the NIR and RGB images. Specifically, at each scale of the encoder, NIR encoded features and RGB pre-denoise features are fed into the frequency domain fusion modules based on FDSM and FEFM, as shown in Fig. 5.", + "bbox": [ + 89, + 642, + 483, + 762 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Frequency Dynamic Selection Mechanism", + "text_level": 1, + "bbox": [ + 89, + 772, + 447, + 787 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The frequency correlation prior indicates that extracting cross-field complementary information from the frequency domain is efficient. For feature extraction, traditional convolutional kernels share weights across the entire feature map, lacking flexibility to address the numerous inconsistencies between NIR and RGB images, which affects the efficiency of extraction. We introduce frequency domain", + "bbox": [ + 89, + 794, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "dynamic convolution to adapt to specific inputs and effectively extract complementary features.", + "bbox": [ + 511, + 90, + 903, + 119 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Specifically, given the NIR features $\\mathbf{N} \\in \\mathbb{R}^{C \\times H \\times W}$ and pre-denoised features $\\mathbf{R} \\in \\mathbb{R}^{C \\times H \\times W}$ , we concatenate them after layernorm and undergo convolution and the activate function to obtain a coarsely fused feature map. After that, we separate the fused feature map to obtain the aggregated features $\\mathbf{A}_{\\mathbf{R}}$ and $\\mathbf{A}_{\\mathbf{N}}$ for NIR and RGB individually. These aggregated features are used to generate dynamic filtering kernels to selectively extract useful features from the NIR and RGB input features.", + "bbox": [ + 511, + 119, + 903, + 258 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For dynamic filters, some works [27, 52] that generate deep separable and spatially variant dynamic filtering kernels are usually computationally intensive and time-consuming, and require a large amount of data. We employ a method similar to [6, 34] by generating dynamic weights to dynamically aggregate multiple parallel convolutional kernels related to the input, and performing this process in the frequency domain. Specifically, the aggregated features $\\mathbf{A}_{\\mathbf{R}}$ and $\\mathbf{A}_{\\mathbf{N}}$ are used to generate the weights for $k$ learnable frequency domain filters. For each feature channel, these filters are linearly combined to form a frequency domain global filter, which is used to select useful frequency domain information from the input NIR and RGB features. Finally, $\\mathbf{F}_{\\mathbf{R}}$ , $\\mathbf{F}_{\\mathbf{N}}$ denoting the filtered useful features of NIR and RGB can be represented as", + "bbox": [ + 511, + 260, + 906, + 486 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {D F} _ {\\mathbf {I}} = \\operatorname {S o f t m a x} (\\operatorname {M L P} (\\operatorname {P o o l i n g} (\\mathbf {A} _ {\\mathbf {I}}))) \\cdot \\mathbf {G} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 562, + 501, + 903, + 517 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {F} _ {\\mathbf {I}} = \\mathcal {F} ^ {- 1} (\\mathcal {F} (\\mathbf {I}) \\odot \\mathbf {D F} _ {\\mathbf {I}}) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 622, + 549, + 903, + 566 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{DF}_{\\mathbf{I}}$ is frequency domain dynamic filter, and $\\mathbf{I}$ represents $\\mathbf{R}$ or $\\mathbf{N}$ . $\\mathbf{G} \\in \\mathbb{R}^{M \\times k}$ represents the combination of $k$ learnable frequency domain filtering kernels, and $\\mathbf{M}$ is the spatial dimension of the convolutional kernel. $\\mathcal{F}(\\cdot)$ and $\\mathcal{F}^{-1}(\\cdot)$ denote the 2D-DFT and 2D-IDFT operations, respectively.", + "bbox": [ + 511, + 577, + 905, + 667 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Frequency Exhaustive Fusion Mechanism", + "text_level": 1, + "bbox": [ + 511, + 681, + 870, + 696 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "After the NIR and RGB feature maps pass through the FDSM, the useful features are retained. Next, to fuse the two inconsistent fields, we further integrate their information in the frequency domain and propose the FEFM, with this process driven by prior frequency correlations. Specifically, the proposed FEFM includes two key components: Common Feature Reinforcement (CFR) mechanism and Differential Feature Reinforcement (DFR) mechanism. CFR comprehensively explores the correlation between NIR and RGB features in the frequency domain to strengthen the modeling of common features, while DFR employs differential modeling to enhance key differential high frequency features.", + "bbox": [ + 511, + 703, + 906, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "12682", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/ed8870cfdcef13a312ceaab3eafb76c2359f11206d819cd593cbb66bfd5bb935.jpg", + "image_caption": [ + "Figure 5. The architecture of the proposed Frequency Dynamic Selection Mechanism (FDSM) and Frequency Exhaustive Fusion Mechanism (FEFM) for thorough frequency exploitation." + ], + "image_footnote": [], + "bbox": [ + 98, + 90, + 504, + 352 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/a3813b66fc45b9c7286ce0915b5ead4f9be1099372b027b1b80c5a523c6ca42a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 509, + 90, + 901, + 352 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Common Feature Reinforcement Mechanism. The convolution theorem establishes that the correlation or convolution of two signals in the spatial domain is equal to their element-wise product in the frequency domain. Leveraging this property, previous work [21] integrate the frequency domain into the self-attention mechanism, simplifying matrix multiplication into lightweight element-wise dot product operations. Although the dot product is efficient, it only calculates the point-wise correlation coefficients in the frequency domain, missing the long-range correlations in the frequency domain. To more comprehensively model the common features of NIR and RGB, we apply a pointwise convolution and a depthwise convolution to the input NIR and RGB feature maps $\\mathbf{F}_{\\mathbf{R}}$ , $\\mathbf{F}_{\\mathbf{N}}$ , to obtain encoded features $\\mathbf{Q} = W_{d}^{Q}W_{p}^{Q}\\mathbf{F}_{\\mathbf{R}}$ , $\\mathbf{K} = W_{d}^{K}W_{p}^{K}\\mathbf{F}_{\\mathbf{N}}$ , $\\mathbf{V} = W_{d}^{V}W_{p}^{V}\\mathbf{F}_{\\mathbf{N}}$ . To facilitate the calculation of frequency domain correlations, we apply 2D-DFT to the input features to produce frequency cubes $\\mathbf{F}_{\\mathbf{Q}}$ and $\\mathbf{F}_{\\mathbf{K}}$ . We simultaneously model the long-range correlation and point-wise correlation of NIR and RGB features in the frequency domain, resulting in a correlation map that integrates local similarity and long-range channel correlation within the frequency domain. The output of CFR can be represented as", + "bbox": [ + 91, + 400, + 483, + 750 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {F} _ {\\mathbf {C F R}} = \\left(\\mathbf {F} _ {\\mathbf {Q}} \\odot \\mathbf {F} _ {\\mathbf {K}}\\right) \\cdot \\mathrm {S o f t m a x} (\\mathbf {F} _ {\\mathbf {Q}} \\cdot \\mathbf {F} _ {\\mathbf {K}} ^ {\\top} / \\alpha) \\quad (4)\n$$\n", + "text_format": "latex", + "bbox": [ + 130, + 753, + 482, + 773 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\alpha$ is a learnable scaling parameter.", + "bbox": [ + 89, + 780, + 362, + 795 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Differential Feature Reinforcement Mechanism. Traditional cross-attention [15, 25, 39] structures are typically used to capture common features but do not effectively utilize differential information. Although this approach is correct in general image fusion frameworks, in this paper, the features of the two fields are considered useful after passing through the frequency domain selection mechanism. By", + "bbox": [ + 89, + 795, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "observing the feature maps (Fig. 9), we find that the high frequency texture information in the NIR features is clear compared to the RGB feature maps. The common features obtained by weighted fusion through cross-attention remain somewhat blurry, which greatly affects the denoising effect. These differential high frequency features on NIR are crucial for image restoration. We adopt a differential cross-attention to complement the missing differential information, and the output of CFR can be represented as", + "bbox": [ + 511, + 400, + 906, + 537 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {F} _ {\\mathbf {D F R}} = \\left(\\mathbf {V} - \\lambda \\mathbf {V} \\odot \\mathcal {F} ^ {- 1} \\left(\\mathbf {F} _ {\\mathbf {C F R}}\\right)\\right) \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 581, + 553, + 906, + 571 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda$ is a learnable scalar that dynamically controls the weights of two feature maps. Although DFR effectively models the differential high frequency information, the fused features lack some common features and background information. To further refine the fused features, we perform a dot product between the encoded features Q and the output of CFR, and add the result to the output of DFR as the final fused result.", + "bbox": [ + 511, + 588, + 906, + 708 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Loss Function", + "text_level": 1, + "bbox": [ + 511, + 723, + 660, + 738 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We adopt the Charbonnier loss [5], and adding the frequency domain loss [9]. The overall loss function is represented as", + "bbox": [ + 511, + 747, + 905, + 792 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} = \\mathcal {L} _ {\\text {c h a r b o n n i e r}} \\left(\\mathbf {X} _ {1}, \\mathbf {T}\\right) + \\mathcal {L} _ {\\text {c h a r b o n n i e r}} \\left(\\mathbf {X} _ {2}, \\mathbf {T}\\right) \\tag {6} \\\\ + \\alpha \\cdot \\mathcal {L} _ {\\text {c h a r b o n n i e r}} (\\mathcal {F} (\\mathbf {X} _ {2}), \\mathcal {F} (\\mathbf {T})) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 531, + 804, + 903, + 840 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{X}_1$ and $\\mathbf{X}_2$ are the outputs of pre-denoising subnetwork and whole network, respectively. $\\mathbf{T}$ is ground truth and $\\alpha$ is set as 0.1 [9].", + "bbox": [ + 511, + 854, + 906, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "12683", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/d505563a74781eb99f16582db7a79b4a840e189841ce46afbc36714df2e51c76.jpg", + "image_caption": [ + "Figure 6. The qualitative comparison among our FCENet and the state-of-the-art methods on the noisy RGB-NIR pairs from DVD[18] with different noise levels. Our method achieves better results in detail recovery, artifact removal, and color restoration." + ], + "image_footnote": [], + "bbox": [ + 94, + 77, + 903, + 349 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/ff78c319ea4e0658d0fc0a052c62c1119dc0be6dbb00c03bd4a73cec721a40bc.jpg", + "table_caption": [ + "Table 1. Comparison of different methods on the DVD test set with noise levels $\\sigma = 2,4,6$ . The best and second-best results are highlighted in boldface and underlined, respectively." + ], + "table_footnote": [], + "table_body": "
Methodsσ = 2σ = 4σ = 6Complexity
PSNR↑SSIM↑PSNR↑SSIM↑PSNR↑SSIM↑FLOPs(G)Params(M)
Uformer [41]31.340.94929.120.92727.580.9089.6020.63
Restormer [48]31.300.94929.200.92827.760.91036.0626.13
MPRNet [47]31.780.95129.370.92927.840.910133.5115.74
Xformer [49]31.410.95029.300.92827.430.90839.3425.22
CUNet [11]28.760.92326.810.89825.640.8756.720.44
SANet [32]30.130.93827.840.91726.410.90125.015.38
MNNet [44]30.140.94428.260.92326.750.90610.970.76
DVN [18]31.500.95529.620.94028.260.92748.306.97
NIR-Restormer [43]31.810.95729.810.94228.310.92845.3130.26
FCENet-L (Ours)31.870.96029.910.94628.360.9345.711.16
FCENet (Ours)32.430.96330.370.95028.780.93919.524.44
", + "bbox": [ + 91, + 422, + 906, + 654 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 664, + 223, + 681 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we evaluate the proposed method on public datasets and compare it with other NIR-assisted image denoising methods.", + "bbox": [ + 89, + 689, + 482, + 736 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experimental Settings", + "text_level": 1, + "bbox": [ + 89, + 742, + 297, + 758 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. We evaluate the NIR-assisted image denoising method on the DVD [18] and IVRG [4] datasets. The DVD dataset consists of 307 high-resolution NIR-RGB image pairs, which are converted into 5000 image pairs of size $3 \\times 256 \\times 256$ for training and 1000 image pairs of size $3 \\times 256 \\times 256$ for testing. The IVRG dataset contains NIR and RGB image pairs from 9 different scenes. For each scene, as in [44], experiments randomly select 30/8/15 pairs for training/validation/testing.", + "bbox": [ + 89, + 763, + 483, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Parameter settings. The proposed network is implemented with Pytorch and trained using the Adam optimizer [20] $(\\beta_{1} = 0.9, \\beta_{2} = 0.999)$ on patches of size $128 \\times 128$ with a batch size of 16. For the DVD dataset, we simulate low-light conditions by randomly reducing the mean of the original images as in [18] and add Gaussian-Poisson mixed noise with a range of 1 to 16. The model is trained for 80 epochs. The initial learning rate is $2 \\times 10^{-4}$ , and a cosine annealing strategy is used to gradually reduce the learning rate to $1 \\times 10^{-6}$ . For the IVRG dataset, following the settings in [44], it is cropped into 16,374 training patches, and then additive Gaussian noise with random noise levels is used to simulate the target images, considering three noise levels in testing: $\\sigma = 25$ , 50, and 75. The model is trained for 100 epochs with an initial learning rate of $5 \\times 10^{-4}$ .", + "bbox": [ + 511, + 666, + 906, + 893 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "12684", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/40efdddf4da078317dab4769d3ee5e455ac254c04ffdfa881969bf3f48163ffd.jpg", + "image_caption": [ + "Real noisy RGB" + ], + "image_footnote": [], + "bbox": [ + 96, + 93, + 197, + 170 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f6011350f2ffe760c900d560fc9e4ebb58e0e6b4b103774a34699cac3b691679.jpg", + "image_caption": [ + "NIR" + ], + "image_footnote": [], + "bbox": [ + 197, + 93, + 297, + 171 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ef188faa8bc5f6c4996a255ad18e4769a0c232fc5cab0b083b406464413bacbf.jpg", + "image_caption": [ + "Xformer" + ], + "image_footnote": [], + "bbox": [ + 297, + 93, + 398, + 171 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/bcf826294ca471d329803b0a14ae106dcc66b786b76b4369ad0d492855228561.jpg", + "image_caption": [ + "SANet" + ], + "image_footnote": [], + "bbox": [ + 400, + 93, + 498, + 171 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/97b3fd47acfa9f240401dc6ce6cb0b38ea5313e1047e3244dca4a6c6472e687d.jpg", + "image_caption": [ + "MNNet" + ], + "image_footnote": [], + "bbox": [ + 500, + 93, + 599, + 171 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/456535714bfa1a68771516b2fa7ca084b3afe05ce4180a22b900ba1591981231.jpg", + "image_caption": [ + "DVN" + ], + "image_footnote": [], + "bbox": [ + 602, + 93, + 700, + 171 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6e8981fd2e5ac2e36e1c2c94d0646e6912b6166772e72e329a5db301ffd41821.jpg", + "image_caption": [ + "NIR-Restormer" + ], + "image_footnote": [], + "bbox": [ + 700, + 93, + 802, + 171 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2f7e27f46a8dca2591f5672c4b3cdb886d653f922f4f36685ec3d42ae93eabdd.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 803, + 93, + 903, + 171 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/7d9a2944bfe35f59276accc93230167e45dec510f217568b35a868434d9a7bf4.jpg", + "table_caption": [ + "Table 2. Comparison of different methods on the IVRG test set with noise levels $\\sigma = 25, 50, 75$ . The best and second-best results are highlighted in boldface and underlined, respectively." + ], + "table_footnote": [], + "table_body": "
Methodsσ = 25σ = 50σ = 75
PSNR↑SSIM↑PSNR↑SSIM↑PSNR↑SSIM↑
Restormer [48]31.720.87627.670.78824.500.722
Uformer [41]31.970.87827.830.79324.580.726
MPRNet [47]31.890.87827.800.79424.570.729
Xformer [49]31.980.87927.860.79424.610.731
CUNet [11]31.370.87627.780.81224.640.764
SANet [32]31.540.88528.130.82524.810.785
MNNet [44]32.590.90428.660.85125.250.808
DVN [18]32.760.90628.730.85225.280.809
NIR-Restormer [43]32.860.90628.770.85325.300.811
FCENet-L (Ours)32.850.90928.800.85725.330.816
FCENet (Ours)33.260.91529.080.86625.480.826
", + "bbox": [ + 91, + 273, + 483, + 474 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Experimental Results", + "text_level": 1, + "bbox": [ + 89, + 484, + 292, + 500 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In our experimental evaluation of our method, we compare our results with nine models, including five state-of-the-art NIR-assisted RGB denoising methods: CUNet [11], SANet [32], MNNet [44], DVN [18], and NIR-Restormer [43], and four state-of-the-art single-image denoising methods: MPRNet [47], Restormer [48], Uformer [41], and Xformer [49]. To further demonstrate the superiority of our method, we additionally provide the denoising results of a light version (-L) by changing the initial feature map channel count from 64 to 36, which has only $1.16\\mathrm{M}$ parameters. All compared methods are trained on the same training set as ours.", + "bbox": [ + 88, + 508, + 482, + 688 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Evaluations on the DVN dataset. We first evaluate our method on the DVD dataset, with quantitative results shown in Table 1. It can be observed that the proposed FCENet, and even FCENet-L, outperform all other single-image denoising methods and NIR-assisted denoising methods in terms of PSNR and SSIM with moderate computational and memory costs. The qualitative comparison in Fig. 6 clearly illustrates that our method has advantages over other state-of-the-art methods in noise removal, color preservation, and detail recovery. For single-image denoising methods like Restormer, although they have achieved advanced denoising effects in low-noise scenarios, they still lack effective means to address detail texture loss in high-noise situations. For guided denoising methods, MNNet is unable to effec", + "bbox": [ + 89, + 689, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d17f45f53a63c92a062f1cf5ffcf65f1835c6e2eeb48f9a28fe466e3cfe22f51.jpg", + "image_caption": [ + "Figure 7. Visual comparisons on real-world RGB/NIR image pairs. Our method demonstrates better visual results.", + "Figure 8. Frequency domain dynamic filters on NIR and RGB side at different depths. The first row represents the frequency domain filters of RGB side, and the second row represents the filters of NIR side." + ], + "image_footnote": [], + "bbox": [ + 517, + 222, + 898, + 321 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "tively handle the inconsistencies between NIR and RGB images. DVN strengthens consistent high frequency textures but loses critical differential high frequency information in NIR and sacrifices the authenticity of low-frequency color. NIR-Restormer has good denoising effects but lacks effective utilization of inconsistent information, leading to over-smoothing and color distortion. Our method recovers detailed texture information while more faithfully preserving low-frequency color information in RGB images. This proves that the proposed method can efficiently utilize the complementary information of NIR and RGB images in the fusion through a series of frequency domain learning.", + "bbox": [ + 511, + 402, + 906, + 583 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "**Evaluations on the IVRG dataset.** We evaluate our method on the IVRG dataset. Following the previous method, we consider three noise levels in the test, including $\\sigma = 25, 50$ , and 75. Quantitative results are shown in Table 2. It can be observed that compared to the state-of-the-art single-image denoising algorithms, our method significantly improves the performance of image denoising, demonstrating the effectiveness of NIR images. Compared to existing NIR-assisted denoising methods, our method also has significant advantages.", + "bbox": [ + 511, + 584, + 905, + 733 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Evaluations on real-world experiment. To further evaluate the performance of our model in real noisy environments, we conduct qualitative experiments on NIR-RGB data pairs in real low-light conditions. As shown in Fig. 7, the introduction of NIR images brings more detailed textures but also a large number of inconsistencies. SANet fails to handle inconsistencies, resulting in a significant amount of artifacts, while MNNet, DVN and NIR-Restormer are unable to thoroughly denoise. In contrast, our method can eliminate inconsistencies while introducing more detailed information.", + "bbox": [ + 511, + 734, + 906, + 898 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "12685", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/415b7f9d523bce9b955e5a264060101cacda6dc3c9949db7ee0cde29a37e8073.jpg", + "table_caption": [ + "Table 3. Ablation study of various components of our method." + ], + "table_footnote": [], + "table_body": "
BaselineFDSMFEFMPSNR↑SSIM↑
(a)✓29.130.936
(b)✓✓29.640.944
(c)✓✓29.960.947
(d)✓✓✓30.370.950
", + "bbox": [ + 98, + 107, + 473, + 199 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/2f19b5329320bf2696df6bb31fa99b6d96670a5f53f660374a4c19af601895d8.jpg", + "table_caption": [ + "Table 4. Ablation on the branches of FDSM." + ], + "table_footnote": [], + "table_body": "
MethodsPSNR↑SSIM↑
w/o FDSM29.960.947
Spatial Dynamic Selection30.110.948
Frequency Dynamic Selection30.370.950
", + "bbox": [ + 109, + 227, + 460, + 303 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/40325c736a1b6bb8a477ad6ea7395112b997e4e4075423009a5f9ece8a125358.jpg", + "table_caption": [ + "Table 5. Ablation on the branches of FEFM." + ], + "table_footnote": [], + "table_body": "
MethodsPSNR↑SSIM↑
Sum29.640.944
Cross-attention29.850.946
DFR30.150.948
DFR + CFR30.370.950
", + "bbox": [ + 153, + 330, + 418, + 422 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 440, + 243, + 455 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We conduct ablation studies with FCENet on the DVD dataset under $\\sigma = 4$ . We compare two mechanisms proposed in this paper with a two-stage U-Net baseline, with the results shown in Table 3. In the U-Net baseline, crossfield images are fused through addition. By deploying FDSM and FEFM, the model achieves gains of $0.51\\mathrm{dB}$ and $0.83\\mathrm{dB}$ , respectively. Finally, combining all contributions significantly improves the denoising level compared to the baseline.", + "bbox": [ + 89, + 462, + 482, + 597 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of FDSM. Using FDSM allows for better extraction of complementary information from NIR and RGB while handling inconsistent regions. In Fig. 8, we show the frequency domain dynamic filter kernels in FDSM. It can be observed that the filter kernels processing NIR feature maps regularly enhance the learning of mid-to-high frequency information, while those processing RGB feature maps mainly learn low frequency information from the input images. This is consistent with the conclusions drawn from our frequency domain correlation priors. Additionally, we conduct experiments using spatial dynamic convolution in this module. The results in Table 4 ultimately prove the effectiveness of extracting relevant features from the frequency domain in NIR and RGB images and the stability of dynamic convolution in extracting inconsistent features.", + "bbox": [ + 89, + 598, + 482, + 824 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of FEFM. Traditional cross-attention fusion modules only strengthen common features but neglect differential features. FEFM strengthens the learning of common features and enhances the learning of some high frequency differential information in NIR feature maps. Fig. 9 shows", + "bbox": [ + 89, + 824, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/21b075348c642adcbe9eb117e4a3c2668d4b456f4430f0feee086f1a0f69c0a9.jpg", + "image_caption": [ + "Figure 9. Feature maps obtained from two fields. The first column represents the input feature maps of NIR and RGB, the second column represents the output feature of cross-attention, and the third column is the output feature of FEFM." + ], + "image_footnote": [], + "bbox": [ + 517, + 90, + 903, + 229 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "the input NIR and RGB feature maps after frequency domain selection, the fused feature maps after cross-attention fusion, and the fused feature maps after FEFM. Compared to the NIR feature maps, the input RGB feature maps still lack high frequency features. It can be seen that the proposed FEFM strengthens common features while supplementing some differential texture information from NIR. We compare two mechanisms in FEFM, CFR and DFR, with traditional cross-attention fusion modules and direct feature summation. The results in Table 5 prove effectiveness of our method in strengthening common features in frequency domain and modeling differential features.", + "bbox": [ + 511, + 309, + 906, + 491 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 508, + 633, + 523 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, a cross-field frequency correlation exploiting network for NIR-assisted image denoising is proposed. Specifically, we first conduct a series of analyses on the characteristics of NIR and RGB images in frequency domain and obtain frequency correlation prior. Based on the prior, the frequency dynamic selection mechanism dynamically extracts complementary information from cross-field images in the frequency domain, and the frequency exhaustive fusion mechanism is proposed to enhance common features and differential high frequency features during fusion process. Extensive experiments demonstrate that the proposed method is significantly better than other algorithms. It is worth mentioning that existing methods may fail on significantly misaligned cross-field images. In future work, we plan to explore advanced methods such as feature point matching or incorporating deformable convolutional kernels to handle significantly misaligned cross-field images.", + "bbox": [ + 511, + 537, + 906, + 792 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 511, + 813, + 671, + 830 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work is supported by the National Natural Science Foundation of China under Grants 62322204, 62131003, and the Key Laboratory of Target Cognition and Application Technology.", + "bbox": [ + 511, + 839, + 905, + 901 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "12686", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 90, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Abdelrahman Abdelhamed, Mahmoud Affifi, Radu Timofte, and Michael S Brown. Ntire 2020 challenge on real image denoising: Dataset, methods and results. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 496-497, 2020. 1", + "[2] Saeed Anwar and Nick Barnes. Real image denoising with feature attention. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3155-3164, 2019. 2", + "[3] Joshua Batson and Loic Royer. Noise2self: Blind denoising by self-supervision. In International Conference on Machine Learning, pages 524-533, 2019. 1", + "[4] Matthew Brown and Sabine Susstrunk. Multi-spectral sift for scene category recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 177-184, 2011. 6", + "[5] Pierre Charbonnier, Laure Blanc-Feraud, Gilles Aubert, and Michel Barlaud. Two deterministic half-quadratic regularization algorithms for computed imaging. In Proceedings of the International Conference on Image Processing, pages 168-172, 1994. 5", + "[6] Yinpeng Chen, Xiyang Dai, Mengchen Liu, Dongdong Chen, Lu Yuan, and Zicheng Liu. Dynamic convolution: Attention over convolution kernels. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11030-11039, 2020. 4", + "[7] Shen Cheng, Yuzhi Wang, Haibin Huang, Donghao Liu, Haoqiang Fan, and Shuaicheng Liu. Nbnet: Noise basis learning for image denoising with subspace projection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4896-4906, 2021. 2", + "[8] Yuxiao Cheng, Runzhao Yang, Zhihong Zhang, Jinli Suo, and Qionghai Dai. A mutually boosting dual sensor computational camera for high quality dark videography. Information Fusion, 93:429-440, 2023. 2", + "[9] Sung-Jin Cho, Seo-Won Ji, Jun-Pyo Hong, Seung-Won Jung, and Sung-Jea Ko. Rethinking coarse-to-fine approach in single image deblurring. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4641-4650, 2021. 5", + "[10] Yuning Cui, Yi Tao, Zhenshan Bing, Wenqi Ren, Xinwei Gao, Xiaochun Cao, Kai Huang, and Alois Knoll. Selective frequency network for image restoration. In International Conference on Learning Representations, 2023. 3", + "[11] Xin Deng and Pier Luigi Dragotti. Deep convolutional neural network for multi-modal image restoration and fusion. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(10):3333-3348, 2020. 2, 3, 6, 7", + "[12] Michael Elad and Michal Aharon. Image denoising via sparse and redundant representations over learned dictionaries. IEEE Transactions on Image processing, 15(12):3736-3745, 2006. 2", + "[13] Hansen Feng, Lizhi Wang, Yuzhi Wang, and Hua Huang. Learnability enhancement for low-light raw denoising: Where paired real data meets noise modeling. In Proceed-" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ings of the ACM International Conference on Multimedia, pages 1436-1444, 2022. 1", + "[14] Hansen Feng, Lizhi Wang, Yuzhi Wang, Haoqiang Fan, and Hua Huang. Learnability enhancement for low-light raw image denoising: A data perspective. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(1):370-387, 2024. 2", + "[15] Yuanshen Guan, Ruikang Xu, Mingde Yao, Lizhi Wang, and Zhiwei Xiong. Mutual-guided dynamic network for image fusion. In Proceedings of the ACM International Conference on Multimedia, pages 1779-1788, 2023. 5", + "[16] Xiaowan Hu, Yuanhao Cai, Jing Lin, Haoqian Wang, Xin Yuan, Yulun Zhang, Radu Timofte, and Luc Van Gool. Hdnet: High-resolution dual-domain learning for spectral compressive imaging. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17542-17551, 2022. 3", + "[17] Liming Jiang, Bo Dai, Wayne Wu, and Chen Change Loy. Focal frequency loss for image reconstruction and synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13919-13929, 2021. 3", + "[18] Shuangping Jin, Bingbing Yu, Minhao Jing, Yi Zhou, Jiajun Liang, and Renhe Ji. Darkvisionnet: Low-light imaging via rgb-nir fusion with deep inconsistency prior. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1104-1112, 2022. 1, 2, 3, 6, 7", + "[19] Beomjun Kim, Jean Ponce, and Bumsub Ham. Deformable kernel networks for joint image filtering. International Journal of Computer Vision, 129(2):579-600, 2021. 2", + "[20] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6", + "[21] Lingshun Kong, Jiangxin Dong, Jianjun Ge, Mingqiang Li, and Jinshan Pan. Efficient frequency domain-based transformers for high-quality image deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5886-5895, 2023. 3, 5", + "[22] Tong Li, Hansen Feng, Lizhi Wang, Zhiwei Xiong, and Hua Huang. Stimulating the diffusion model for image denoising via adaptive embedding and ensembling. In IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 8240-8257, 2024. 2", + "[23] Tong Li, Lizhi Wang, Zhiyuan Xu, Lin Zhu, Wanxuan Lu, and Hua Huang. Positive2negative: Breaking the information-lossy barrier in self-supervised single image denoising. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2025. 2", + "[24] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Swinir: Image restoration using swim transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1833-1844, 2021. 1", + "[25] Hezheng Lin, Xing Cheng, Xiangyu Wu, and Dong Shen. Cat: Cross attention in vision transformer. In Proceedings of the IEEE International Conference on Multimedia and Expo, pages 1-6, 2022. 5" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "12687", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[26] Feifan Lv, Yinqiang Zheng, Yicheng Li, and Feng Lu. An integrated enhancement solution for 24-hour colorful imaging. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 11725-11732, 2020. 1, 2", + "[27] Ruijun Ma, Shuyi Li, Bob Zhang, and Zhengming Li. Generative adaptive convolutions for real-world noisy image denoising. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1935-1943, 2022. 4", + "[28] Youssef Mansour and Reinhard Heckel. Zero-shot noise2noise: Efficient image denoising without any data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14018-14027, 2023. 1", + "[29] Javier Portilla, Vasily Strela, Martin J Wainwright, and Eero P Simoncelli. Image denoising using scale mixtures of gaussians in the wavelet domain. IEEE Transactions on Image Processing, 12(11):1338-1351, 2003. 2", + "[30] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention, pages 234-241, 2015. 4", + "[31] Zehua Sheng, Xiongwei Liu, Si-Yuan Cao, Hui-Liang Shen, and Huaqi Zhang. Frequency-domain deep guided image denoising. IEEE Transactions on Multimedia, 25:6767-6781, 2022. 2, 3", + "[32] Zehua Sheng, Zhu Yu, Xiongwei Liu, Si-Yuan Cao, Yuqi Liu, Hui-Liang Shen, and Huaqi Zhang. Structure aggregation for cross-spectral stereo image guided denoising. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13997-14006, 2023. 2, 3, 6, 7", + "[33] Matthew Tancik, Pratul Srinivasan, Ben Mildenhall, Sara Fridovich-Keil, Nithin Raghavan, Utkarsh Singhal, Ravi Ramamoorthi, Jonathan Barron, and Ren Ng. Fourier features let networks learn high frequency functions in low dimensional domains. Advances in Neural Information Processing Systems, 33:7537-7547, 2020. 3", + "[34] Yuki Tatsunami and Masato Taki. Fft-based dynamic token mixer for vision. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 15328-15336, 2024. 4", + "[35] Chunwei Tian, Lunke Fei, Wenxian Zheng, Yong Xu, Wangmeng Zuo, and Chia-Wen Lin. Deep learning on image denoising: An overview. *Neural Networks*, 131:251–275, 2020. 2", + "[36] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in Neural Information Processing Systems, 30, 2017. 1", + "[37] Renjie Wan, Boxin Shi, Wenhan Yang, Bihan Wen, LingYu Duan, and Alex C. Kot. Purifying low-light images via near-infrared enlightened image. IEEE Transactions on Multimedia, 25:8006-8019, 2023. 1, 3", + "[38] Jian Wang, Tianfan Xue, Jonathan T Barron, and Jiawen Chen. Stereoscopic dark flash for low-light photography. In IEEE International Conference on Computational Photography, pages 1-10, 2019. 2", + "[39] Xin Wang, Lizhi Wang, Xiangtian Ma, Maoqing Zhang, Lin Zhu, and Hua Huang. In2set: Intra-inter similarity exploiting transformer for dual-camera compressive hyperspectral" + ], + "bbox": [ + 91, + 92, + 480, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "imaging. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24881-24891, 2024. 5", + "[40] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing, 13(4):600-612, 2004. 3", + "[41] Zhendong Wang, Xiaodong Cun, Jianmin Bao, Wengang Zhou, Jianzhuang Liu, and Houqiang Li. Uformer: A general u-shaped transformer for image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17683-17693, 2022. 6, 7", + "[42] Yunxuan Wei, Shuhang Gu, Yawei Li, Radu Timofte, Longcun Jin, and Hengjie Song. Unsupervised real-world image super resolution via domain-distance aware training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13385-13394, 2021. 3", + "[43] Rongjian Xu, Zhilu Zhang, Renlong Wu, and Wangmeng Zuo. Nir-assisted image denoising: A selective fusion approach and a real-world benchmark dataset. arXiv preprint arXiv:2404.08514, 2024. 2, 3, 6, 7", + "[44] Shuang Xu, Jiangshe Zhang, Jialin Wang, Kai Sun, Chunxia Zhang, Junmin Liu, and Junying Hu. A model-driven network for guided image denoising. Information Fusion, 85: 60-71, 2022. 2, 3, 6, 7", + "[45] Zhi-Qin John Xu, Yaoyu Zhang, Tao Luo, Yanyang Xiao, and Zheng Ma. Frequency principle: Fourier analysis sheds light on deep neural networks. arXiv preprint arXiv:1901.06523, 2019. 3", + "[46] Qiong Yan, Xiaoyong Shen, Li Xu, Shaojie Zhuo, Xiaopeng Zhang, Liang Shen, and Jiaya Jia. Cross-field joint image restoration via scale map. In Proceedings of the IEEE International Conference on Computer Vision, pages 1537-1544, 2013. 1, 2, 3", + "[47] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, Ming-Hsuan Yang, and Ling Shao. Multi-stage progressive image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14821-14831, 2021. 1, 2, 4, 6, 7", + "[48] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5728-5739, 2022. 2, 6, 7", + "[49] Jiale Zhang, Yulun Zhang, Jinjin Gu, Jiahua Dong, Linghe Kong, and Xiaokang Yang. Xformer: Hybrid x-shaped transformer for image denoising. arXiv preprint arXiv:2303.06440, 2023. 2, 6, 7", + "[50] Kai Zhang, Wangmeng Zuo, Yunjin Chen, Deyu Meng, and Lei Zhang. Beyond a gaussian denoiser: Residual learning of deep cnn for image denoising. IEEE Transactions on Image Processing, 26(7):3142-3155, 2017. 2", + "[51] Kai Zhang, Wangmeng Zuo, and Lei Zhang. Ffdnet: Toward a fast and flexible solution for cnn-based image denoising. IEEE Transactions on Image Processing, 27(9):4608-4622, 2018. 1" + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "12688", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[52] Jingkai Zhou, Varun Jampani, Zhixiong Pi, Qiong Liu, and Ming-Hsuan Yang. Decoupled dynamic filter networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6647-6656, 2021. 4", + "[53] Wenbin Zou, Mingchao Jiang, Yunchen Zhang, Liang Chen, Zhiyong Lu, and Yi Wu. Sdwnet: A straight dilated network with wavelet transformation for image deblurring. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1895–1904, 2021. 3" + ], + "bbox": [ + 91, + 92, + 480, + 218 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "12689", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Complementary Advantages_ Exploiting Cross-Field Frequency Correlation for NIR-Assisted Image Denoising/c9ee5d0d-57af-4b5c-aacf-8ca5596906ca_model.json b/2025/Complementary Advantages_ Exploiting Cross-Field Frequency Correlation for NIR-Assisted Image Denoising/c9ee5d0d-57af-4b5c-aacf-8ca5596906ca_model.json new file mode 100644 index 0000000000000000000000000000000000000000..b9c61fc656658f972f4d2275da6436f50b05bb68 --- /dev/null +++ b/2025/Complementary Advantages_ Exploiting Cross-Field Frequency Correlation for NIR-Assisted Image Denoising/c9ee5d0d-57af-4b5c-aacf-8ca5596906ca_model.json @@ -0,0 +1,2334 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.099, + 0.13, + 0.9, + 0.177 + ], + "angle": 0, + "content": "Complementary Advantages: Exploiting Cross-Field Frequency Correlation for NIR-Assisted Image Denoising" + }, + { + "type": "text", + "bbox": [ + 0.299, + 0.203, + 0.765, + 0.222 + ], + "angle": 0, + "content": "Yuchen Wang\\(^{1}\\) Hongyuan Wang\\(^{1}\\) Lizhi Wang\\(^{2,3*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.277, + 0.222, + 0.786, + 0.24 + ], + "angle": 0, + "content": "Xin Wang\\(^{1}\\) Lin Zhu\\(^{1}\\) Wanxuan Lu\\(^{4}\\) Hua Huang\\(^{2,3}\\)" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.24, + 0.845, + 0.258 + ], + "angle": 0, + "content": "1 School of Computer Science and Technology, Beijing Institute of Technology" + }, + { + "type": "text", + "bbox": [ + 0.288, + 0.258, + 0.774, + 0.275 + ], + "angle": 0, + "content": "\\(^{2}\\) School of Artificial Intelligence, Beijing Normal University" + }, + { + "type": "text", + "bbox": [ + 0.1, + 0.275, + 0.962, + 0.293 + ], + "angle": 0, + "content": "3 Engineering Research Center of Intelligent Technology and Educational Application, Ministry of Education" + }, + { + "type": "text", + "bbox": [ + 0.236, + 0.293, + 0.827, + 0.31 + ], + "angle": 0, + "content": "\\(^{4}\\) Aerospace Information Research Institute, Chinese Academy of Sciences" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.345, + 0.327, + 0.361 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.377, + 0.486, + 0.756 + ], + "angle": 0, + "content": "Existing single-image denoising algorithms often struggle to restore details when dealing with complex noisy images. The introduction of near-infrared (NIR) images offers new possibilities for RGB image denoising. However, due to the inconsistency between NIR and RGB images, the existing works still struggle to balance the contributions of two fields in the process of image fusion. In response to this, in this paper, we develop a cross-field Frequency Correlation Exploiting Network (FCENet) for NIR-assisted image denoising. We first propose the frequency correlation prior based on an in-depth statistical frequency analysis of NIR-RGB image pairs. The prior reveals the complementary correlation of NIR and RGB images in the frequency domain. Leveraging frequency correlation prior, we then establish a frequency learning framework composed of Frequency Dynamic Selection Mechanism (FDSM) and Frequency Exhaustive Fusion Mechanism (FEFM). FDSM dynamically selects complementary information from NIR and RGB images in the frequency domain, and FEFM strengthens the control of common and differential features during the fusion process of NIR and RGB features. Extensive experiments on simulated and real data validate that the proposed method outperforms other state-of-the-art methods. The code will be released at https://github.com/yuchenwang815/FCENet." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.787, + 0.222, + 0.803 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.813, + 0.484, + 0.875 + ], + "angle": 0, + "content": "Due to the inherent physical limitations of digital imaging devices, the captured images are often affected by various types of noise, which have a considerable impact on applications such as 24-hour surveillance, autonomous driving," + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.348, + 0.902, + 0.568 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.574, + 0.907, + 0.658 + ], + "angle": 0, + "content": "Figure 1. Comparisons of PSNR, FLOPs, and Parameters on the DVD dataset [18] are presented. Green circles represent the single-image denoising algorithms, and blue circles represent the NIR-assisted image denoising algorithms. The circle radius represents the number of parameters. The proposed method (red circles) achieves the superior performance while maintaining efficiency." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.673, + 0.907, + 0.869 + ], + "angle": 0, + "content": "and smartphone photography. Although the development of deep learning techniques [24, 36] has brought forth a variety of image denoising methods [1, 3, 13, 28, 47, 51], existing algorithms still find it difficult to deal with complex noisy images, especially in low-light environment. Restoring clean detailed information from noisy images that lack substantial information is highly ill-posed. Therefore, it is necessary to introduce supplementary information from new fields. Fortunately, NIR images with high Signal to Noise Ratio (SNR) can be captured at a low cost and are suitable for everyday photography. Thus, NIR-assisted image denoising [26, 37, 46] has become a promising solution for image denoising." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.907, + 0.9 + ], + "angle": 0, + "content": "However, it is challenging to effectively leverage information from different fields to achieve noise reduction. Due" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.887, + 0.438, + 0.901 + ], + "angle": 0, + "content": "*Corresponding Author: Lizhi Wang (wanglizhi@bnu.edu.cn)" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12679" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.09, + 0.483, + 0.195 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.199, + 0.483, + 0.255 + ], + "angle": 0, + "content": "Figure 2. Visual comparisons on the challenging noisy RGB-NIR image pairs. Pay attention to the differences in structural and color information between two fields. Our method produces a better denoising result with clear details and fewer artifacts." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.265, + 0.483, + 0.401 + ], + "angle": 0, + "content": "to the influence of the inherent reflective spectra of objects, there are inconsistencies between NIR and RGB images. As shown in Fig. 2, firstly, there is a significant difference in color and brightness information between two fields. Secondly, some structural texture of RGB image disappears in NIR image, and additional artifacts may appear. The color and structural inconsistencies between NIR and RGB images pose a significant challenge for the task of NIR-assisted image denoising." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.402, + 0.483, + 0.612 + ], + "angle": 0, + "content": "In existing researches, some methods attempt to directly integrate inconsistent features from two fields [31, 32, 43, 44]. However, these methods indiscriminately fuse crossfield information, relying on the fitting capabilities of neural networks to learn the fusion of NIR and RGB images under spatial inconsistency, which often leads to suboptimal denoising results and lower computational efficiency. Some methods reinforce common features from NIR and RGB images and diminish the weight of inconsistent features during the fusion [11, 18, 19, 46]. These image fusion schemes typically only enhance the common information in NIR and RGB during fusion, but some differential information such as the sharp textures in NIR and the color information in RGB is discarded." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.614, + 0.483, + 0.824 + ], + "angle": 0, + "content": "In this paper, we develop a cross-field Frequency Correlation Exploiting Network (FCENet) for NIR-assisted image denoising. We conduct a series of analyses on the characteristics of NIR and RGB images in the frequency domain and obtain a key observation: in the same scene, the similarity between noisy RGB images and clean RGB (noise-free RGB) images decreases from low to high frequency, whereas the similarity between NIR images and clean RGB images increases from low to high frequency. This significant cross-field frequency correlation prior redefines crossfield fusion as a process of frequency domain adaptive selection, motivating us to subsequently establish detailed frequency domain learning to precisely control the restoration of texture details and colors." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Specifically, the frequency domain fusion network consists of two mechanisms: Frequency Dynamic Selection Mechanism (FDSM) and Frequency Exhaustive Fusion Mechanism (FEFM). In order to comprehensively extract useful features from inconsistent images, the FDSM is for-" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.242 + ], + "angle": 0, + "content": "mulated based on dynamic filters under the guidance of frequency domain prior. The FDSM efficiently screens complementary features from NIR and RGB in the frequency domain, providing high quality material for subsequent image fusion. Then, the FEFM is formulated to ensure thorough frequency domain fusion. The FEFM integrates local and long-range correlations across fields to fully extract the common feature from cross-field information, and employs a differential cross-attention approach to complement some key differential high frequency features in NIR features." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.243, + 0.907, + 0.333 + ], + "angle": 0, + "content": "The proposed FCENet can utilize complementary information from cross-field images. Extensive experiments on simulated and real data validate that our method outperforms other state-of-the-art methods in terms of image quality and computational efficiency. The main contributions of this work are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.335, + 0.905, + 0.395 + ], + "angle": 0, + "content": "- We establish the frequency correlation prior between NIR and RGB images through a series of frequency domain analyses of two fields, which prompts the integration of complementary features in the frequency domain." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.396, + 0.905, + 0.455 + ], + "angle": 0, + "content": "- We propose an efficient frequency dynamic selection mechanism based on the frequency correlation prior to identify which useful low-frequency and high frequency information should be retained in NIR and RGB features." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.456, + 0.905, + 0.532 + ], + "angle": 0, + "content": "- We develop an exhaustive frequency domain fusion mechanism that models the local similarity and long-range correlation in the frequency domain of NIR and RGB images, and supplements some critical high frequency differential features." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.335, + 0.905, + 0.532 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.547, + 0.655, + 0.563 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.572, + 0.731, + 0.589 + ], + "angle": 0, + "content": "2.1. Single Image Denoising" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.595, + 0.907, + 0.807 + ], + "angle": 0, + "content": "Image denoising [12, 29] is a classic task in low-level vision. In recent years, deep learning has greatly improved the accuracy and efficiency of denoising [2, 7, 14, 22, 23, 35]. DnCNN [50] uses a simple network with batch normalization and residual learning, which outperforms traditional denoising methods. MPRNet [47] designs a multi-stage restoration framework, and Restormer [48] proposes an efficient transformer framework for high-resolution images. Xformer [49] proposes a transformer framework that combines spatial self-attention and channel self-attention. However, despite their strong denoising capabilities, single-image denoising still inevitably loses a lot of spatial information, leading to overly smooth edge textures in the denoised images." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.818, + 0.799, + 0.835 + ], + "angle": 0, + "content": "2.2. NIR-Assisted Image Restoration" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.905, + 0.901 + ], + "angle": 0, + "content": "To achieve better denoising effects, researchers have attempted to use NIR images to assist in RGB image denoising [8, 26, 38, 46]. Compared with single-image recovery, NIR images can help restore the details of degraded im" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12680" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.091, + 0.485, + 0.35 + ], + "angle": 0, + "content": "ages [37, 43]. In early research, Yan et al. [46] proposes a scale map from the perspective of gradient inconsistency to find universal usable edges and smooth transitions in NIR images. With the proliferation of deep learning, CUNet [11] uses a convolutional sparse coding model to extract common features from cross-field images. FGDNet [31] and SANet [32] fuse the features of two fields in the frequency domain. MNNet [44] proposes an observation model that takes into account the modality gap between the target and guiding images. DVN [18] integrates structural inconsistency priors into deep features, specifically strengthening the common deep features in NIR and RGB images. NAID [43] proposes Selective Fusion Module (SFM), which can be plugged into advanced image restoration networks. However, under extreme noisy conditions, existing NIR-assisted denoising methods fail to balance color accuracy, detail retention, and artifact suppression." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.357, + 0.421, + 0.373 + ], + "angle": 0, + "content": "2.3. Frequency Domain Image Restoration" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.379, + 0.484, + 0.621 + ], + "angle": 0, + "content": "Frequency analysis in digital signal processing emphasizes the frequency properties of signals [33, 45]. Recently, frequency domain analysis has shown great potential in improving the performance of various learning-based image restoration frameworks [17, 42, 53]. In [21], An efficient estimation of scaled dot-product attention based on a frequency domain self-attention solver (FSAS) is proposed. HDNet [16] proposes a high-resolution dual-domain learning network. SFNet [10] uses filters to dynamically decouple feature mappings into different frequency components and adaptively extracts useful features. In this paper, we emphasize the importance of frequency domain information for the research on fusion under inconsistency and design a comprehensive frequency domain fusion scheme for this task to ensure the utilization of complementary information from NIR and RGB images." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.633, + 0.182, + 0.649 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.659, + 0.484, + 0.766 + ], + "angle": 0, + "content": "In this section, we first introduce the cross-field frequency correlation prior of NIR and RGB images. Based on the cross-field frequency correlation prior, a detailed frequency domain scheme is formulated, including the Frequency Dynamic Selection Mechanism (FDSM) and the Frequency Exhaustive Fusion Mechanism (FEFM) for thorough frequency exploitation." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.773, + 0.442, + 0.79 + ], + "angle": 0, + "content": "3.1. Cross-Field Frequency Correlation Prior" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.484, + 0.901 + ], + "angle": 0, + "content": "To address the fusion of NIR and RGB images under spatial inconsistency, it is imperative to distinguish which information in NIR and RGB images is useful and which is redundant or even erroneous for denoising. We randomly select 30 pairs of NIR images, noisy RGB images, and clean RGB images from the existing RGB-NIR dataset and perform fourier transform on them. To facilitate the study of the" + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.091, + 0.709, + 0.217 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.709, + 0.091, + 0.897, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.524, + 0.219, + 0.655, + 0.34 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.658, + 0.219, + 0.772, + 0.34 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.775, + 0.219, + 0.892, + 0.34 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.699, + 0.34, + 0.717, + 0.35 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.352, + 0.907, + 0.465 + ], + "angle": 0, + "content": "Figure 3. The analysis of NIR and RGB frequency correlation. (a) The frequency correlation between noisy RGB and clean RGB, with the horizontal axis indicating the cutoff frequency of the high-pass filter. Different colored curves represent different scenarios. (b) The frequency correlation between NIR and clean RGB. (c) The visualization results of images output by a fixed frequency filter, with the first row showing high frequency (HF) part and the second row showing low frequency (LF) part." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.476, + 0.907, + 0.598 + ], + "angle": 0, + "content": "correlation between NIR and RGB in the frequency domain, we apply a high-pass filter to the aforementioned three kinds of frequency domain images and convert the filtered results back to the spatial domain. We separately computed the Structural Similarity (SSIM) [40] between NIR/noisy RGB images and clean RGB images at different cutoff frequencies as a measure of frequency correlation \\( S \\), as shown in Fig. 3a and Fig. 3b. The process can be described as" + }, + { + "type": "equation", + "bbox": [ + 0.557, + 0.604, + 0.907, + 0.623 + ], + "angle": 0, + "content": "\\[\nS = \\operatorname {S S I M} \\left(\\mathcal {F} ^ {- 1} \\left(\\mathrm {H} \\left(F _ {\\text {t a r g e t}}\\right)\\right), \\mathcal {F} ^ {- 1} \\left(\\mathrm {H} \\left(F _ {\\text {g t}}\\right)\\right)\\right) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.628, + 0.909, + 0.902 + ], + "angle": 0, + "content": "where \\(\\mathcal{F}^{-1}(\\cdot)\\) denote 2D-IDFT operations, \\(F_{target}\\) and \\(F_{gt}\\) represent NIR or noisy RGB and clean RGB after 2D-DFT transformation. \\(\\mathrm{H}(\\cdot)\\) represents the filtering of the input image with a high-pass filter. The analysis results indicate that, in the main energy frequency, the image similarity between the input noisy RGB and clean RGB images decreases progressively as the frequency increases, while the image similarity between the input NIR image and clean RGB image increases from low to high frequency. Fig. 3c also shows that for noisy RGB images, their high frequency information is mainly composed of noise and lacks detail information, while the low frequency information more faithfully reflects the color in the original image, and NIR images are just the opposite. These cross-field Frequency correlation priors indicate that it is simple and efficient to distinguish useful information from NIR and RGB images in the frequency domain, which inspires us to propose a comprehensive frequency domain learning framework." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "12681" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.092, + 0.253, + 0.161 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.119, + 0.163, + 0.232, + 0.175 + ], + "angle": 0, + "content": "Noisy RGB (Input)" + }, + { + "type": "image", + "bbox": [ + 0.323, + 0.093, + 0.477, + 0.16 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.345, + 0.163, + 0.454, + 0.175 + ], + "angle": 0, + "content": "Clean NIR (Input)" + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.176, + 0.48, + 0.407 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.412, + 0.483, + 0.469 + ], + "angle": 0, + "content": "Figure 4. The overall architecture of the proposed cross-field Frequency Correlation Exploiting Network (FCENet). In the first stage, pre-denoising is performed. In the second stage, NIR image is introduced for NIR-assisted image denoising." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.484, + 0.296, + 0.499 + ], + "angle": 0, + "content": "3.2. Overview Framework" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.506, + 0.484, + 0.642 + ], + "angle": 0, + "content": "The proposed frequency domain fusion framework, as depicted in Fig. 4, includes two stages for the gradual restoration of images. Both stages incorporate an encoder-decoder architecture based on the U-Net [30], featuring two downsampling layers and two upsampling layers. Skip connections are employed between the encoder and decoder features. A Supervised Attention Module (SAM) [47] is utilized to connect features from the first stage to the second stage for further processing." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.643, + 0.484, + 0.763 + ], + "angle": 0, + "content": "In the first stage, the input noisy RGB image undergoes preliminary denoising to reduce the modality gap with the NIR image. In the second stage, we perform multi-scale fusion of the encoded features of the NIR and RGB images. Specifically, at each scale of the encoder, NIR encoded features and RGB pre-denoise features are fed into the frequency domain fusion modules based on FDSM and FEFM, as shown in Fig. 5." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.773, + 0.449, + 0.789 + ], + "angle": 0, + "content": "3.3. Frequency Dynamic Selection Mechanism" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.484, + 0.901 + ], + "angle": 0, + "content": "The frequency correlation prior indicates that extracting cross-field complementary information from the frequency domain is efficient. For feature extraction, traditional convolutional kernels share weights across the entire feature map, lacking flexibility to address the numerous inconsistencies between NIR and RGB images, which affects the efficiency of extraction. We introduce frequency domain" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.121 + ], + "angle": 0, + "content": "dynamic convolution to adapt to specific inputs and effectively extract complementary features." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.121, + 0.905, + 0.259 + ], + "angle": 0, + "content": "Specifically, given the NIR features \\(\\mathbf{N} \\in \\mathbb{R}^{C \\times H \\times W}\\) and pre-denoised features \\(\\mathbf{R} \\in \\mathbb{R}^{C \\times H \\times W}\\), we concatenate them after layernorm and undergo convolution and the activate function to obtain a coarsely fused feature map. After that, we separate the fused feature map to obtain the aggregated features \\(\\mathbf{A}_{\\mathbf{R}}\\) and \\(\\mathbf{A}_{\\mathbf{N}}\\) for NIR and RGB individually. These aggregated features are used to generate dynamic filtering kernels to selectively extract useful features from the NIR and RGB input features." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.261, + 0.907, + 0.487 + ], + "angle": 0, + "content": "For dynamic filters, some works [27, 52] that generate deep separable and spatially variant dynamic filtering kernels are usually computationally intensive and time-consuming, and require a large amount of data. We employ a method similar to [6, 34] by generating dynamic weights to dynamically aggregate multiple parallel convolutional kernels related to the input, and performing this process in the frequency domain. Specifically, the aggregated features \\(\\mathbf{A}_{\\mathbf{R}}\\) and \\(\\mathbf{A}_{\\mathbf{N}}\\) are used to generate the weights for \\(k\\) learnable frequency domain filters. For each feature channel, these filters are linearly combined to form a frequency domain global filter, which is used to select useful frequency domain information from the input NIR and RGB features. Finally, \\(\\mathbf{F}_{\\mathbf{R}}\\), \\(\\mathbf{F}_{\\mathbf{N}}\\) denoting the filtered useful features of NIR and RGB can be represented as" + }, + { + "type": "equation", + "bbox": [ + 0.563, + 0.502, + 0.905, + 0.518 + ], + "angle": 0, + "content": "\\[\n\\mathbf {D F} _ {\\mathbf {I}} = \\operatorname {S o f t m a x} (\\operatorname {M L P} (\\operatorname {P o o l i n g} (\\mathbf {A} _ {\\mathbf {I}}))) \\cdot \\mathbf {G} \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.623, + 0.55, + 0.905, + 0.567 + ], + "angle": 0, + "content": "\\[\n\\mathbf {F} _ {\\mathbf {I}} = \\mathcal {F} ^ {- 1} (\\mathcal {F} (\\mathbf {I}) \\odot \\mathbf {D F} _ {\\mathbf {I}}) \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.578, + 0.906, + 0.669 + ], + "angle": 0, + "content": "where \\(\\mathbf{DF}_{\\mathbf{I}}\\) is frequency domain dynamic filter, and \\(\\mathbf{I}\\) represents \\(\\mathbf{R}\\) or \\(\\mathbf{N}\\). \\(\\mathbf{G} \\in \\mathbb{R}^{M \\times k}\\) represents the combination of \\(k\\) learnable frequency domain filtering kernels, and \\(\\mathbf{M}\\) is the spatial dimension of the convolutional kernel. \\(\\mathcal{F}(\\cdot)\\) and \\(\\mathcal{F}^{-1}(\\cdot)\\) denote the 2D-DFT and 2D-IDFT operations, respectively." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.682, + 0.871, + 0.698 + ], + "angle": 0, + "content": "3.4. Frequency Exhaustive Fusion Mechanism" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.704, + 0.907, + 0.901 + ], + "angle": 0, + "content": "After the NIR and RGB feature maps pass through the FDSM, the useful features are retained. Next, to fuse the two inconsistent fields, we further integrate their information in the frequency domain and propose the FEFM, with this process driven by prior frequency correlations. Specifically, the proposed FEFM includes two key components: Common Feature Reinforcement (CFR) mechanism and Differential Feature Reinforcement (DFR) mechanism. CFR comprehensively explores the correlation between NIR and RGB features in the frequency domain to strengthen the modeling of common features, while DFR employs differential modeling to enhance key differential high frequency features." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12682" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.092, + 0.506, + 0.353 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.092, + 0.902, + 0.353 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.361, + 0.908, + 0.391 + ], + "angle": 0, + "content": "Figure 5. The architecture of the proposed Frequency Dynamic Selection Mechanism (FDSM) and Frequency Exhaustive Fusion Mechanism (FEFM) for thorough frequency exploitation." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.401, + 0.485, + 0.75 + ], + "angle": 0, + "content": "Common Feature Reinforcement Mechanism. The convolution theorem establishes that the correlation or convolution of two signals in the spatial domain is equal to their element-wise product in the frequency domain. Leveraging this property, previous work [21] integrate the frequency domain into the self-attention mechanism, simplifying matrix multiplication into lightweight element-wise dot product operations. Although the dot product is efficient, it only calculates the point-wise correlation coefficients in the frequency domain, missing the long-range correlations in the frequency domain. To more comprehensively model the common features of NIR and RGB, we apply a pointwise convolution and a depthwise convolution to the input NIR and RGB feature maps \\(\\mathbf{F}_{\\mathbf{R}}\\), \\(\\mathbf{F}_{\\mathbf{N}}\\), to obtain encoded features \\(\\mathbf{Q} = W_{d}^{Q}W_{p}^{Q}\\mathbf{F}_{\\mathbf{R}}\\), \\(\\mathbf{K} = W_{d}^{K}W_{p}^{K}\\mathbf{F}_{\\mathbf{N}}\\), \\(\\mathbf{V} = W_{d}^{V}W_{p}^{V}\\mathbf{F}_{\\mathbf{N}}\\). To facilitate the calculation of frequency domain correlations, we apply 2D-DFT to the input features to produce frequency cubes \\(\\mathbf{F}_{\\mathbf{Q}}\\) and \\(\\mathbf{F}_{\\mathbf{K}}\\). We simultaneously model the long-range correlation and point-wise correlation of NIR and RGB features in the frequency domain, resulting in a correlation map that integrates local similarity and long-range channel correlation within the frequency domain. The output of CFR can be represented as" + }, + { + "type": "equation", + "bbox": [ + 0.131, + 0.755, + 0.483, + 0.775 + ], + "angle": 0, + "content": "\\[\n\\mathbf {F} _ {\\mathbf {C F R}} = \\left(\\mathbf {F} _ {\\mathbf {Q}} \\odot \\mathbf {F} _ {\\mathbf {K}}\\right) \\cdot \\mathrm {S o f t m a x} (\\mathbf {F} _ {\\mathbf {Q}} \\cdot \\mathbf {F} _ {\\mathbf {K}} ^ {\\top} / \\alpha) \\quad (4)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.781, + 0.364, + 0.796 + ], + "angle": 0, + "content": "where \\(\\alpha\\) is a learnable scaling parameter." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Differential Feature Reinforcement Mechanism. Traditional cross-attention [15, 25, 39] structures are typically used to capture common features but do not effectively utilize differential information. Although this approach is correct in general image fusion frameworks, in this paper, the features of the two fields are considered useful after passing through the frequency domain selection mechanism. By" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.401, + 0.907, + 0.538 + ], + "angle": 0, + "content": "observing the feature maps (Fig. 9), we find that the high frequency texture information in the NIR features is clear compared to the RGB feature maps. The common features obtained by weighted fusion through cross-attention remain somewhat blurry, which greatly affects the denoising effect. These differential high frequency features on NIR are crucial for image restoration. We adopt a differential cross-attention to complement the missing differential information, and the output of CFR can be represented as" + }, + { + "type": "equation", + "bbox": [ + 0.583, + 0.554, + 0.907, + 0.573 + ], + "angle": 0, + "content": "\\[\n\\mathbf {F} _ {\\mathbf {D F R}} = \\left(\\mathbf {V} - \\lambda \\mathbf {V} \\odot \\mathcal {F} ^ {- 1} \\left(\\mathbf {F} _ {\\mathbf {C F R}}\\right)\\right) \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.589, + 0.907, + 0.709 + ], + "angle": 0, + "content": "where \\(\\lambda\\) is a learnable scalar that dynamically controls the weights of two feature maps. Although DFR effectively models the differential high frequency information, the fused features lack some common features and background information. To further refine the fused features, we perform a dot product between the encoded features Q and the output of CFR, and add the result to the output of DFR as the final fused result." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.724, + 0.661, + 0.739 + ], + "angle": 0, + "content": "3.5. Loss Function" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.748, + 0.906, + 0.793 + ], + "angle": 0, + "content": "We adopt the Charbonnier loss [5], and adding the frequency domain loss [9]. The overall loss function is represented as" + }, + { + "type": "equation", + "bbox": [ + 0.532, + 0.805, + 0.905, + 0.841 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} = \\mathcal {L} _ {\\text {c h a r b o n n i e r}} \\left(\\mathbf {X} _ {1}, \\mathbf {T}\\right) + \\mathcal {L} _ {\\text {c h a r b o n n i e r}} \\left(\\mathbf {X} _ {2}, \\mathbf {T}\\right) \\tag {6} \\\\ + \\alpha \\cdot \\mathcal {L} _ {\\text {c h a r b o n n i e r}} (\\mathcal {F} (\\mathbf {X} _ {2}), \\mathcal {F} (\\mathbf {T})) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.907, + 0.901 + ], + "angle": 0, + "content": "where \\(\\mathbf{X}_1\\) and \\(\\mathbf{X}_2\\) are the outputs of pre-denoising subnetwork and whole network, respectively. \\(\\mathbf{T}\\) is ground truth and \\(\\alpha\\) is set as 0.1 [9]." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12683" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.078, + 0.905, + 0.35 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.353, + 0.907, + 0.381 + ], + "angle": 0, + "content": "Figure 6. The qualitative comparison among our FCENet and the state-of-the-art methods on the noisy RGB-NIR pairs from DVD[18] with different noise levels. Our method achieves better results in detail recovery, artifact removal, and color restoration." + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.392, + 0.907, + 0.42 + ], + "angle": 0, + "content": "Table 1. Comparison of different methods on the DVD test set with noise levels \\( \\sigma = 2,4,6 \\) . The best and second-best results are highlighted in boldface and underlined, respectively." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.424, + 0.907, + 0.655 + ], + "angle": 0, + "content": "
Methodsσ = 2σ = 4σ = 6Complexity
PSNR↑SSIM↑PSNR↑SSIM↑PSNR↑SSIM↑FLOPs(G)Params(M)
Uformer [41]31.340.94929.120.92727.580.9089.6020.63
Restormer [48]31.300.94929.200.92827.760.91036.0626.13
MPRNet [47]31.780.95129.370.92927.840.910133.5115.74
Xformer [49]31.410.95029.300.92827.430.90839.3425.22
CUNet [11]28.760.92326.810.89825.640.8756.720.44
SANet [32]30.130.93827.840.91726.410.90125.015.38
MNNet [44]30.140.94428.260.92326.750.90610.970.76
DVN [18]31.500.95529.620.94028.260.92748.306.97
NIR-Restormer [43]31.810.95729.810.94228.310.92845.3130.26
FCENet-L (Ours)31.870.96029.910.94628.360.9345.711.16
FCENet (Ours)32.430.96330.370.95028.780.93919.524.44
" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.665, + 0.224, + 0.682 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.69, + 0.483, + 0.737 + ], + "angle": 0, + "content": "In this section, we evaluate the proposed method on public datasets and compare it with other NIR-assisted image denoising methods." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.743, + 0.299, + 0.759 + ], + "angle": 0, + "content": "4.1. Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Datasets. We evaluate the NIR-assisted image denoising method on the DVD [18] and IVRG [4] datasets. The DVD dataset consists of 307 high-resolution NIR-RGB image pairs, which are converted into 5000 image pairs of size \\(3 \\times 256 \\times 256\\) for training and 1000 image pairs of size \\(3 \\times 256 \\times 256\\) for testing. The IVRG dataset contains NIR and RGB image pairs from 9 different scenes. For each scene, as in [44], experiments randomly select 30/8/15 pairs for training/validation/testing." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.667, + 0.907, + 0.895 + ], + "angle": 0, + "content": "Parameter settings. The proposed network is implemented with Pytorch and trained using the Adam optimizer [20] \\((\\beta_{1} = 0.9, \\beta_{2} = 0.999)\\) on patches of size \\(128 \\times 128\\) with a batch size of 16. For the DVD dataset, we simulate low-light conditions by randomly reducing the mean of the original images as in [18] and add Gaussian-Poisson mixed noise with a range of 1 to 16. The model is trained for 80 epochs. The initial learning rate is \\(2 \\times 10^{-4}\\), and a cosine annealing strategy is used to gradually reduce the learning rate to \\(1 \\times 10^{-6}\\). For the IVRG dataset, following the settings in [44], it is cropped into 16,374 training patches, and then additive Gaussian noise with random noise levels is used to simulate the target images, considering three noise levels in testing: \\(\\sigma = 25\\), 50, and 75. The model is trained for 100 epochs with an initial learning rate of \\(5 \\times 10^{-4}\\)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12684" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.094, + 0.198, + 0.171 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.103, + 0.174, + 0.195, + 0.187 + ], + "angle": 0, + "content": "Real noisy RGB" + }, + { + "type": "image", + "bbox": [ + 0.199, + 0.094, + 0.298, + 0.172 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.234, + 0.176, + 0.259, + 0.186 + ], + "angle": 0, + "content": "NIR" + }, + { + "type": "image", + "bbox": [ + 0.299, + 0.094, + 0.399, + 0.172 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.325, + 0.176, + 0.375, + 0.187 + ], + "angle": 0, + "content": "Xformer" + }, + { + "type": "image", + "bbox": [ + 0.401, + 0.094, + 0.499, + 0.172 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.427, + 0.176, + 0.467, + 0.187 + ], + "angle": 0, + "content": "SANet" + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.094, + 0.601, + 0.172 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.529, + 0.176, + 0.573, + 0.187 + ], + "angle": 0, + "content": "MNNet" + }, + { + "type": "image", + "bbox": [ + 0.603, + 0.094, + 0.701, + 0.172 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.638, + 0.176, + 0.669, + 0.187 + ], + "angle": 0, + "content": "DVN" + }, + { + "type": "image", + "bbox": [ + 0.702, + 0.094, + 0.803, + 0.172 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.71, + 0.176, + 0.797, + 0.187 + ], + "angle": 0, + "content": "NIR-Restormer" + }, + { + "type": "image", + "bbox": [ + 0.805, + 0.094, + 0.904, + 0.172 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.837, + 0.176, + 0.869, + 0.187 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.16, + 0.194, + 0.836, + 0.208 + ], + "angle": 0, + "content": "Figure 7. Visual comparisons on real-world RGB/NIR image pairs. Our method demonstrates better visual results." + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.222, + 0.483, + 0.264 + ], + "angle": 0, + "content": "Table 2. Comparison of different methods on the IVRG test set with noise levels \\(\\sigma = 25, 50, 75\\). The best and second-best results are highlighted in boldface and underlined, respectively." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.275, + 0.485, + 0.475 + ], + "angle": 0, + "content": "
Methodsσ = 25σ = 50σ = 75
PSNR↑SSIM↑PSNR↑SSIM↑PSNR↑SSIM↑
Restormer [48]31.720.87627.670.78824.500.722
Uformer [41]31.970.87827.830.79324.580.726
MPRNet [47]31.890.87827.800.79424.570.729
Xformer [49]31.980.87927.860.79424.610.731
CUNet [11]31.370.87627.780.81224.640.764
SANet [32]31.540.88528.130.82524.810.785
MNNet [44]32.590.90428.660.85125.250.808
DVN [18]32.760.90628.730.85225.280.809
NIR-Restormer [43]32.860.90628.770.85325.300.811
FCENet-L (Ours)32.850.90928.800.85725.330.816
FCENet (Ours)33.260.91529.080.86625.480.826
" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.485, + 0.294, + 0.5 + ], + "angle": 0, + "content": "4.2. Experimental Results" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.509, + 0.483, + 0.689 + ], + "angle": 0, + "content": "In our experimental evaluation of our method, we compare our results with nine models, including five state-of-the-art NIR-assisted RGB denoising methods: CUNet [11], SANet [32], MNNet [44], DVN [18], and NIR-Restormer [43], and four state-of-the-art single-image denoising methods: MPRNet [47], Restormer [48], Uformer [41], and Xformer [49]. To further demonstrate the superiority of our method, we additionally provide the denoising results of a light version (-L) by changing the initial feature map channel count from 64 to 36, which has only \\(1.16\\mathrm{M}\\) parameters. All compared methods are trained on the same training set as ours." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.69, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Evaluations on the DVN dataset. We first evaluate our method on the DVD dataset, with quantitative results shown in Table 1. It can be observed that the proposed FCENet, and even FCENet-L, outperform all other single-image denoising methods and NIR-assisted denoising methods in terms of PSNR and SSIM with moderate computational and memory costs. The qualitative comparison in Fig. 6 clearly illustrates that our method has advantages over other state-of-the-art methods in noise removal, color preservation, and detail recovery. For single-image denoising methods like Restormer, although they have achieved advanced denoising effects in low-noise scenarios, they still lack effective means to address detail texture loss in high-noise situations. For guided denoising methods, MNNet is unable to effec" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.223, + 0.9, + 0.322 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.334, + 0.907, + 0.389 + ], + "angle": 0, + "content": "Figure 8. Frequency domain dynamic filters on NIR and RGB side at different depths. The first row represents the frequency domain filters of RGB side, and the second row represents the filters of NIR side." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.403, + 0.907, + 0.584 + ], + "angle": 0, + "content": "tively handle the inconsistencies between NIR and RGB images. DVN strengthens consistent high frequency textures but loses critical differential high frequency information in NIR and sacrifices the authenticity of low-frequency color. NIR-Restormer has good denoising effects but lacks effective utilization of inconsistent information, leading to over-smoothing and color distortion. Our method recovers detailed texture information while more faithfully preserving low-frequency color information in RGB images. This proves that the proposed method can efficiently utilize the complementary information of NIR and RGB images in the fusion through a series of frequency domain learning." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.585, + 0.906, + 0.734 + ], + "angle": 0, + "content": "**Evaluations on the IVRG dataset.** We evaluate our method on the IVRG dataset. Following the previous method, we consider three noise levels in the test, including \\(\\sigma = 25, 50\\), and 75. Quantitative results are shown in Table 2. It can be observed that compared to the state-of-the-art single-image denoising algorithms, our method significantly improves the performance of image denoising, demonstrating the effectiveness of NIR images. Compared to existing NIR-assisted denoising methods, our method also has significant advantages." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.907, + 0.9 + ], + "angle": 0, + "content": "Evaluations on real-world experiment. To further evaluate the performance of our model in real noisy environments, we conduct qualitative experiments on NIR-RGB data pairs in real low-light conditions. As shown in Fig. 7, the introduction of NIR images brings more detailed textures but also a large number of inconsistencies. SANet fails to handle inconsistencies, resulting in a significant amount of artifacts, while MNNet, DVN and NIR-Restormer are unable to thoroughly denoise. In contrast, our method can eliminate inconsistencies while introducing more detailed information." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "12685" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.101, + 0.09, + 0.472, + 0.104 + ], + "angle": 0, + "content": "Table 3. Ablation study of various components of our method." + }, + { + "type": "table", + "bbox": [ + 0.099, + 0.108, + 0.474, + 0.2 + ], + "angle": 0, + "content": "
BaselineFDSMFEFMPSNR↑SSIM↑
(a)✓29.130.936
(b)✓✓29.640.944
(c)✓✓29.960.947
(d)✓✓✓30.370.950
" + }, + { + "type": "table_caption", + "bbox": [ + 0.154, + 0.211, + 0.419, + 0.224 + ], + "angle": 0, + "content": "Table 4. Ablation on the branches of FDSM." + }, + { + "type": "table", + "bbox": [ + 0.11, + 0.228, + 0.462, + 0.304 + ], + "angle": 0, + "content": "
MethodsPSNR↑SSIM↑
w/o FDSM29.960.947
Spatial Dynamic Selection30.110.948
Frequency Dynamic Selection30.370.950
" + }, + { + "type": "table_caption", + "bbox": [ + 0.154, + 0.314, + 0.419, + 0.327 + ], + "angle": 0, + "content": "Table 5. Ablation on the branches of FEFM." + }, + { + "type": "table", + "bbox": [ + 0.154, + 0.332, + 0.419, + 0.424 + ], + "angle": 0, + "content": "
MethodsPSNR↑SSIM↑
Sum29.640.944
Cross-attention29.850.946
DFR30.150.948
DFR + CFR30.370.950
" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.441, + 0.245, + 0.457 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.463, + 0.483, + 0.598 + ], + "angle": 0, + "content": "We conduct ablation studies with FCENet on the DVD dataset under \\(\\sigma = 4\\). We compare two mechanisms proposed in this paper with a two-stage U-Net baseline, with the results shown in Table 3. In the U-Net baseline, crossfield images are fused through addition. By deploying FDSM and FEFM, the model achieves gains of \\(0.51\\mathrm{dB}\\) and \\(0.83\\mathrm{dB}\\), respectively. Finally, combining all contributions significantly improves the denoising level compared to the baseline." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.599, + 0.483, + 0.825 + ], + "angle": 0, + "content": "Effect of FDSM. Using FDSM allows for better extraction of complementary information from NIR and RGB while handling inconsistent regions. In Fig. 8, we show the frequency domain dynamic filter kernels in FDSM. It can be observed that the filter kernels processing NIR feature maps regularly enhance the learning of mid-to-high frequency information, while those processing RGB feature maps mainly learn low frequency information from the input images. This is consistent with the conclusions drawn from our frequency domain correlation priors. Additionally, we conduct experiments using spatial dynamic convolution in this module. The results in Table 4 ultimately prove the effectiveness of extracting relevant features from the frequency domain in NIR and RGB images and the stability of dynamic convolution in extracting inconsistent features." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Effect of FEFM. Traditional cross-attention fusion modules only strengthen common features but neglect differential features. FEFM strengthens the learning of common features and enhances the learning of some high frequency differential information in NIR feature maps. Fig. 9 shows" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.091, + 0.905, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.237, + 0.907, + 0.293 + ], + "angle": 0, + "content": "Figure 9. Feature maps obtained from two fields. The first column represents the input feature maps of NIR and RGB, the second column represents the output feature of cross-attention, and the third column is the output feature of FEFM." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.31, + 0.907, + 0.492 + ], + "angle": 0, + "content": "the input NIR and RGB feature maps after frequency domain selection, the fused feature maps after cross-attention fusion, and the fused feature maps after FEFM. Compared to the NIR feature maps, the input RGB feature maps still lack high frequency features. It can be seen that the proposed FEFM strengthens common features while supplementing some differential texture information from NIR. We compare two mechanisms in FEFM, CFR and DFR, with traditional cross-attention fusion modules and direct feature summation. The results in Table 5 prove effectiveness of our method in strengthening common features in frequency domain and modeling differential features." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.51, + 0.634, + 0.525 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.538, + 0.907, + 0.794 + ], + "angle": 0, + "content": "In this paper, a cross-field frequency correlation exploiting network for NIR-assisted image denoising is proposed. Specifically, we first conduct a series of analyses on the characteristics of NIR and RGB images in frequency domain and obtain frequency correlation prior. Based on the prior, the frequency dynamic selection mechanism dynamically extracts complementary information from cross-field images in the frequency domain, and the frequency exhaustive fusion mechanism is proposed to enhance common features and differential high frequency features during fusion process. Extensive experiments demonstrate that the proposed method is significantly better than other algorithms. It is worth mentioning that existing methods may fail on significantly misaligned cross-field images. In future work, we plan to explore advanced methods such as feature point matching or incorporating deformable convolutional kernels to handle significantly misaligned cross-field images." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.814, + 0.673, + 0.831 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.906, + 0.902 + ], + "angle": 0, + "content": "This work is supported by the National Natural Science Foundation of China under Grants 62322204, 62131003, and the Key Laboratory of Target Cognition and Application Technology." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "12686" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.091, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.184 + ], + "angle": 0, + "content": "[1] Abdelrahman Abdelhamed, Mahmoud Affifi, Radu Timofte, and Michael S Brown. Ntire 2020 challenge on real image denoising: Dataset, methods and results. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 496-497, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.187, + 0.484, + 0.241 + ], + "angle": 0, + "content": "[2] Saeed Anwar and Nick Barnes. Real image denoising with feature attention. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3155-3164, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.245, + 0.483, + 0.285 + ], + "angle": 0, + "content": "[3] Joshua Batson and Loic Royer. Noise2self: Blind denoising by self-supervision. In International Conference on Machine Learning, pages 524-533, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.288, + 0.483, + 0.342 + ], + "angle": 0, + "content": "[4] Matthew Brown and Sabine Susstrunk. Multi-spectral sift for scene category recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 177-184, 2011. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.345, + 0.483, + 0.413 + ], + "angle": 0, + "content": "[5] Pierre Charbonnier, Laure Blanc-Feraud, Gilles Aubert, and Michel Barlaud. Two deterministic half-quadratic regularization algorithms for computed imaging. In Proceedings of the International Conference on Image Processing, pages 168-172, 1994. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.417, + 0.483, + 0.485 + ], + "angle": 0, + "content": "[6] Yinpeng Chen, Xiyang Dai, Mengchen Liu, Dongdong Chen, Lu Yuan, and Zicheng Liu. Dynamic convolution: Attention over convolution kernels. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11030-11039, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.488, + 0.483, + 0.557 + ], + "angle": 0, + "content": "[7] Shen Cheng, Yuzhi Wang, Haibin Huang, Donghao Liu, Haoqiang Fan, and Shuaicheng Liu. Nbnet: Noise basis learning for image denoising with subspace projection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4896-4906, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.559, + 0.483, + 0.613 + ], + "angle": 0, + "content": "[8] Yuxiao Cheng, Runzhao Yang, Zhihong Zhang, Jinli Suo, and Qionghai Dai. A mutually boosting dual sensor computational camera for high quality dark videography. Information Fusion, 93:429-440, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.617, + 0.483, + 0.683 + ], + "angle": 0, + "content": "[9] Sung-Jin Cho, Seo-Won Ji, Jun-Pyo Hong, Seung-Won Jung, and Sung-Jea Ko. Rethinking coarse-to-fine approach in single image deblurring. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4641-4650, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.687, + 0.483, + 0.743 + ], + "angle": 0, + "content": "[10] Yuning Cui, Yi Tao, Zhenshan Bing, Wenqi Ren, Xinwei Gao, Xiaochun Cao, Kai Huang, and Alois Knoll. Selective frequency network for image restoration. In International Conference on Learning Representations, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.746, + 0.483, + 0.799 + ], + "angle": 0, + "content": "[11] Xin Deng and Pier Luigi Dragotti. Deep convolutional neural network for multi-modal image restoration and fusion. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(10):3333-3348, 2020. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.803, + 0.483, + 0.856 + ], + "angle": 0, + "content": "[12] Michael Elad and Michal Aharon. Image denoising via sparse and redundant representations over learned dictionaries. IEEE Transactions on Image processing, 15(12):3736-3745, 2006. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.86, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[13] Hansen Feng, Lizhi Wang, Yuzhi Wang, and Hua Huang. Learnability enhancement for low-light raw denoising: Where paired real data meets noise modeling. In Proceed-" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.548, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "ings of the ACM International Conference on Multimedia, pages 1436-1444, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.124, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[14] Hansen Feng, Lizhi Wang, Yuzhi Wang, Haoqiang Fan, and Hua Huang. Learnability enhancement for low-light raw image denoising: A data perspective. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(1):370-387, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.195, + 0.905, + 0.25 + ], + "angle": 0, + "content": "[15] Yuanshen Guan, Ruikang Xu, Mingde Yao, Lizhi Wang, and Zhiwei Xiong. Mutual-guided dynamic network for image fusion. In Proceedings of the ACM International Conference on Multimedia, pages 1779-1788, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.254, + 0.905, + 0.335 + ], + "angle": 0, + "content": "[16] Xiaowan Hu, Yuanhao Cai, Jing Lin, Haoqian Wang, Xin Yuan, Yulun Zhang, Radu Timofte, and Luc Van Gool. Hdnet: High-resolution dual-domain learning for spectral compressive imaging. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17542-17551, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.339, + 0.905, + 0.395 + ], + "angle": 0, + "content": "[17] Liming Jiang, Bo Dai, Wayne Wu, and Chen Change Loy. Focal frequency loss for image reconstruction and synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13919-13929, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.398, + 0.905, + 0.466 + ], + "angle": 0, + "content": "[18] Shuangping Jin, Bingbing Yu, Minhao Jing, Yi Zhou, Jiajun Liang, and Renhe Ji. Darkvisionnet: Low-light imaging via rgb-nir fusion with deep inconsistency prior. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1104-1112, 2022. 1, 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.47, + 0.904, + 0.511 + ], + "angle": 0, + "content": "[19] Beomjun Kim, Jean Ponce, and Bumsub Ham. Deformable kernel networks for joint image filtering. International Journal of Computer Vision, 129(2):579-600, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.514, + 0.905, + 0.553 + ], + "angle": 0, + "content": "[20] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.557, + 0.905, + 0.627 + ], + "angle": 0, + "content": "[21] Lingshun Kong, Jiangxin Dong, Jianjun Ge, Mingqiang Li, and Jinshan Pan. Efficient frequency domain-based transformers for high-quality image deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5886-5895, 2023. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.63, + 0.905, + 0.698 + ], + "angle": 0, + "content": "[22] Tong Li, Hansen Feng, Lizhi Wang, Zhiwei Xiong, and Hua Huang. Stimulating the diffusion model for image denoising via adaptive embedding and ensembling. In IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 8240-8257, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.702, + 0.905, + 0.771 + ], + "angle": 0, + "content": "[23] Tong Li, Lizhi Wang, Zhiyuan Xu, Lin Zhu, Wanxuan Lu, and Hua Huang. Positive2negative: Breaking the information-lossy barrier in self-supervised single image denoising. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.774, + 0.905, + 0.841 + ], + "angle": 0, + "content": "[24] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Swinir: Image restoration using swim transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1833-1844, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.907, + 0.901 + ], + "angle": 0, + "content": "[25] Hezheng Lin, Xing Cheng, Xiangyu Wu, and Dong Shen. Cat: Cross attention in vision transformer. In Proceedings of the IEEE International Conference on Multimedia and Expo, pages 1-6, 2022. 5" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12687" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[26] Feifan Lv, Yinqiang Zheng, Yicheng Li, and Feng Lu. An integrated enhancement solution for 24-hour colorful imaging. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 11725-11732, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.203 + ], + "angle": 0, + "content": "[27] Ruijun Ma, Shuyi Li, Bob Zhang, and Zhengming Li. Generative adaptive convolutions for real-world noisy image denoising. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1935-1943, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.205, + 0.482, + 0.259 + ], + "angle": 0, + "content": "[28] Youssef Mansour and Reinhard Heckel. Zero-shot noise2noise: Efficient image denoising without any data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14018-14027, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.261, + 0.482, + 0.315 + ], + "angle": 0, + "content": "[29] Javier Portilla, Vasily Strela, Martin J Wainwright, and Eero P Simoncelli. Image denoising using scale mixtures of gaussians in the wavelet domain. IEEE Transactions on Image Processing, 12(11):1338-1351, 2003. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.317, + 0.482, + 0.371 + ], + "angle": 0, + "content": "[30] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention, pages 234-241, 2015. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.373, + 0.482, + 0.425 + ], + "angle": 0, + "content": "[31] Zehua Sheng, Xiongwei Liu, Si-Yuan Cao, Hui-Liang Shen, and Huaqi Zhang. Frequency-domain deep guided image denoising. IEEE Transactions on Multimedia, 25:6767-6781, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.428, + 0.482, + 0.509 + ], + "angle": 0, + "content": "[32] Zehua Sheng, Zhu Yu, Xiongwei Liu, Si-Yuan Cao, Yuqi Liu, Hui-Liang Shen, and Huaqi Zhang. Structure aggregation for cross-spectral stereo image guided denoising. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13997-14006, 2023. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.511, + 0.482, + 0.594 + ], + "angle": 0, + "content": "[33] Matthew Tancik, Pratul Srinivasan, Ben Mildenhall, Sara Fridovich-Keil, Nithin Raghavan, Utkarsh Singhal, Ravi Ramamoorthi, Jonathan Barron, and Ren Ng. Fourier features let networks learn high frequency functions in low dimensional domains. Advances in Neural Information Processing Systems, 33:7537-7547, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.595, + 0.482, + 0.636 + ], + "angle": 0, + "content": "[34] Yuki Tatsunami and Masato Taki. Fft-based dynamic token mixer for vision. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 15328-15336, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.637, + 0.482, + 0.69 + ], + "angle": 0, + "content": "[35] Chunwei Tian, Lunke Fei, Wenxian Zheng, Yong Xu, Wangmeng Zuo, and Chia-Wen Lin. Deep learning on image denoising: An overview. *Neural Networks*, 131:251–275, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.692, + 0.482, + 0.747 + ], + "angle": 0, + "content": "[36] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in Neural Information Processing Systems, 30, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.749, + 0.482, + 0.802 + ], + "angle": 0, + "content": "[37] Renjie Wan, Boxin Shi, Wenhan Yang, Bihan Wen, LingYu Duan, and Alex C. Kot. Purifying low-light images via near-infrared enlightened image. IEEE Transactions on Multimedia, 25:8006-8019, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.804, + 0.482, + 0.858 + ], + "angle": 0, + "content": "[38] Jian Wang, Tianfan Xue, Jonathan T Barron, and Jiawen Chen. Stereoscopic dark flash for low-light photography. In IEEE International Conference on Computational Photography, pages 1-10, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.86, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[39] Xin Wang, Lizhi Wang, Xiangtian Ma, Maoqing Zhang, Lin Zhu, and Hua Huang. In2set: Intra-inter similarity exploiting transformer for dual-camera compressive hyperspectral" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.133 + ], + "angle": 0, + "content": "imaging. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24881-24891, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.135, + 0.905, + 0.189 + ], + "angle": 0, + "content": "[40] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing, 13(4):600-612, 2004. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.191, + 0.905, + 0.259 + ], + "angle": 0, + "content": "[41] Zhendong Wang, Xiaodong Cun, Jianmin Bao, Wengang Zhou, Jianzhuang Liu, and Houqiang Li. Uformer: A general u-shaped transformer for image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17683-17693, 2022. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.261, + 0.905, + 0.328 + ], + "angle": 0, + "content": "[42] Yunxuan Wei, Shuhang Gu, Yawei Li, Radu Timofte, Longcun Jin, and Hengjie Song. Unsupervised real-world image super resolution via domain-distance aware training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13385-13394, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.33, + 0.905, + 0.384 + ], + "angle": 0, + "content": "[43] Rongjian Xu, Zhilu Zhang, Renlong Wu, and Wangmeng Zuo. Nir-assisted image denoising: A selective fusion approach and a real-world benchmark dataset. arXiv preprint arXiv:2404.08514, 2024. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.386, + 0.905, + 0.44 + ], + "angle": 0, + "content": "[44] Shuang Xu, Jiangshe Zhang, Jialin Wang, Kai Sun, Chunxia Zhang, Junmin Liu, and Junying Hu. A model-driven network for guided image denoising. Information Fusion, 85: 60-71, 2022. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.442, + 0.905, + 0.495 + ], + "angle": 0, + "content": "[45] Zhi-Qin John Xu, Yaoyu Zhang, Tao Luo, Yanyang Xiao, and Zheng Ma. Frequency principle: Fourier analysis sheds light on deep neural networks. arXiv preprint arXiv:1901.06523, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.498, + 0.905, + 0.565 + ], + "angle": 0, + "content": "[46] Qiong Yan, Xiaoyong Shen, Li Xu, Shaojie Zhuo, Xiaopeng Zhang, Liang Shen, and Jiaya Jia. Cross-field joint image restoration via scale map. In Proceedings of the IEEE International Conference on Computer Vision, pages 1537-1544, 2013. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.568, + 0.905, + 0.649 + ], + "angle": 0, + "content": "[47] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, Ming-Hsuan Yang, and Ling Shao. Multi-stage progressive image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14821-14831, 2021. 1, 2, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.651, + 0.905, + 0.733 + ], + "angle": 0, + "content": "[48] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5728-5739, 2022. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.735, + 0.905, + 0.788 + ], + "angle": 0, + "content": "[49] Jiale Zhang, Yulun Zhang, Jinjin Gu, Jiahua Dong, Linghe Kong, and Xiaokang Yang. Xformer: Hybrid x-shaped transformer for image denoising. arXiv preprint arXiv:2303.06440, 2023. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.79, + 0.905, + 0.844 + ], + "angle": 0, + "content": "[50] Kai Zhang, Wangmeng Zuo, Yunjin Chen, Deyu Meng, and Lei Zhang. Beyond a gaussian denoiser: Residual learning of deep cnn for image denoising. IEEE Transactions on Image Processing, 26(7):3142-3155, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.905, + 0.899 + ], + "angle": 0, + "content": "[51] Kai Zhang, Wangmeng Zuo, and Lei Zhang. Ffdnet: Toward a fast and flexible solution for cnn-based image denoising. IEEE Transactions on Image Processing, 27(9):4608-4622, 2018. 1" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12688" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[52] Jingkai Zhou, Varun Jampani, Zhixiong Pi, Qiong Liu, and Ming-Hsuan Yang. Decoupled dynamic filter networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6647-6656, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.219 + ], + "angle": 0, + "content": "[53] Wenbin Zou, Mingchao Jiang, Yunchen Zhang, Liang Chen, Zhiyong Lu, and Yi Wu. Sdwnet: A straight dilated network with wavelet transformation for image deblurring. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1895–1904, 2021. 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.956 + ], + "angle": 0, + "content": "12689" + } + ] +] \ No newline at end of file diff --git a/2025/Complementary Advantages_ Exploiting Cross-Field Frequency Correlation for NIR-Assisted Image Denoising/c9ee5d0d-57af-4b5c-aacf-8ca5596906ca_origin.pdf b/2025/Complementary Advantages_ Exploiting Cross-Field Frequency Correlation for NIR-Assisted Image Denoising/c9ee5d0d-57af-4b5c-aacf-8ca5596906ca_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..22c8cd045d11b63de5513d273ac6c8060898945f --- /dev/null +++ b/2025/Complementary Advantages_ Exploiting Cross-Field Frequency Correlation for NIR-Assisted Image Denoising/c9ee5d0d-57af-4b5c-aacf-8ca5596906ca_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:770bdda03b3bbae21096e006917ba3dcc787e63e6fdcdecfea2175ee515b4c4c +size 4085828 diff --git a/2025/Complementary Advantages_ Exploiting Cross-Field Frequency Correlation for NIR-Assisted Image Denoising/full.md b/2025/Complementary Advantages_ Exploiting Cross-Field Frequency Correlation for NIR-Assisted Image Denoising/full.md new file mode 100644 index 0000000000000000000000000000000000000000..8a122c6ba77bdb9ea914aa1af8655afc4982fbb0 --- /dev/null +++ b/2025/Complementary Advantages_ Exploiting Cross-Field Frequency Correlation for NIR-Assisted Image Denoising/full.md @@ -0,0 +1,319 @@ +# Complementary Advantages: Exploiting Cross-Field Frequency Correlation for NIR-Assisted Image Denoising + +Yuchen Wang $^{1}$ Hongyuan Wang $^{1}$ Lizhi Wang $^{2,3*}$ + +Xin Wang $^{1}$ Lin Zhu $^{1}$ Wanxuan Lu $^{4}$ Hua Huang $^{2,3}$ + +1 School of Computer Science and Technology, Beijing Institute of Technology + +$^{2}$ School of Artificial Intelligence, Beijing Normal University + +3 Engineering Research Center of Intelligent Technology and Educational Application, Ministry of Education + +$^{4}$ Aerospace Information Research Institute, Chinese Academy of Sciences + +# Abstract + +Existing single-image denoising algorithms often struggle to restore details when dealing with complex noisy images. The introduction of near-infrared (NIR) images offers new possibilities for RGB image denoising. However, due to the inconsistency between NIR and RGB images, the existing works still struggle to balance the contributions of two fields in the process of image fusion. In response to this, in this paper, we develop a cross-field Frequency Correlation Exploiting Network (FCENet) for NIR-assisted image denoising. We first propose the frequency correlation prior based on an in-depth statistical frequency analysis of NIR-RGB image pairs. The prior reveals the complementary correlation of NIR and RGB images in the frequency domain. Leveraging frequency correlation prior, we then establish a frequency learning framework composed of Frequency Dynamic Selection Mechanism (FDSM) and Frequency Exhaustive Fusion Mechanism (FEFM). FDSM dynamically selects complementary information from NIR and RGB images in the frequency domain, and FEFM strengthens the control of common and differential features during the fusion process of NIR and RGB features. Extensive experiments on simulated and real data validate that the proposed method outperforms other state-of-the-art methods. The code will be released at https://github.com/yuchenwang815/FCENet. + +# 1. Introduction + +Due to the inherent physical limitations of digital imaging devices, the captured images are often affected by various types of noise, which have a considerable impact on applications such as 24-hour surveillance, autonomous driving, + +![](images/cd04717a2e57779998dce53b0b02024e66048ef79e423846adb28afdd725e4c7.jpg) +Figure 1. Comparisons of PSNR, FLOPs, and Parameters on the DVD dataset [18] are presented. Green circles represent the single-image denoising algorithms, and blue circles represent the NIR-assisted image denoising algorithms. The circle radius represents the number of parameters. The proposed method (red circles) achieves the superior performance while maintaining efficiency. + +and smartphone photography. Although the development of deep learning techniques [24, 36] has brought forth a variety of image denoising methods [1, 3, 13, 28, 47, 51], existing algorithms still find it difficult to deal with complex noisy images, especially in low-light environment. Restoring clean detailed information from noisy images that lack substantial information is highly ill-posed. Therefore, it is necessary to introduce supplementary information from new fields. Fortunately, NIR images with high Signal to Noise Ratio (SNR) can be captured at a low cost and are suitable for everyday photography. Thus, NIR-assisted image denoising [26, 37, 46] has become a promising solution for image denoising. + +However, it is challenging to effectively leverage information from different fields to achieve noise reduction. Due + +![](images/b49cca02186ed0919b0d221791bed6ad52902dfa423cbcdb42edcb991e259751.jpg) +Figure 2. Visual comparisons on the challenging noisy RGB-NIR image pairs. Pay attention to the differences in structural and color information between two fields. Our method produces a better denoising result with clear details and fewer artifacts. + +to the influence of the inherent reflective spectra of objects, there are inconsistencies between NIR and RGB images. As shown in Fig. 2, firstly, there is a significant difference in color and brightness information between two fields. Secondly, some structural texture of RGB image disappears in NIR image, and additional artifacts may appear. The color and structural inconsistencies between NIR and RGB images pose a significant challenge for the task of NIR-assisted image denoising. + +In existing researches, some methods attempt to directly integrate inconsistent features from two fields [31, 32, 43, 44]. However, these methods indiscriminately fuse crossfield information, relying on the fitting capabilities of neural networks to learn the fusion of NIR and RGB images under spatial inconsistency, which often leads to suboptimal denoising results and lower computational efficiency. Some methods reinforce common features from NIR and RGB images and diminish the weight of inconsistent features during the fusion [11, 18, 19, 46]. These image fusion schemes typically only enhance the common information in NIR and RGB during fusion, but some differential information such as the sharp textures in NIR and the color information in RGB is discarded. + +In this paper, we develop a cross-field Frequency Correlation Exploiting Network (FCENet) for NIR-assisted image denoising. We conduct a series of analyses on the characteristics of NIR and RGB images in the frequency domain and obtain a key observation: in the same scene, the similarity between noisy RGB images and clean RGB (noise-free RGB) images decreases from low to high frequency, whereas the similarity between NIR images and clean RGB images increases from low to high frequency. This significant cross-field frequency correlation prior redefines crossfield fusion as a process of frequency domain adaptive selection, motivating us to subsequently establish detailed frequency domain learning to precisely control the restoration of texture details and colors. + +Specifically, the frequency domain fusion network consists of two mechanisms: Frequency Dynamic Selection Mechanism (FDSM) and Frequency Exhaustive Fusion Mechanism (FEFM). In order to comprehensively extract useful features from inconsistent images, the FDSM is for- + +mulated based on dynamic filters under the guidance of frequency domain prior. The FDSM efficiently screens complementary features from NIR and RGB in the frequency domain, providing high quality material for subsequent image fusion. Then, the FEFM is formulated to ensure thorough frequency domain fusion. The FEFM integrates local and long-range correlations across fields to fully extract the common feature from cross-field information, and employs a differential cross-attention approach to complement some key differential high frequency features in NIR features. + +The proposed FCENet can utilize complementary information from cross-field images. Extensive experiments on simulated and real data validate that our method outperforms other state-of-the-art methods in terms of image quality and computational efficiency. The main contributions of this work are summarized as follows: + +- We establish the frequency correlation prior between NIR and RGB images through a series of frequency domain analyses of two fields, which prompts the integration of complementary features in the frequency domain. +- We propose an efficient frequency dynamic selection mechanism based on the frequency correlation prior to identify which useful low-frequency and high frequency information should be retained in NIR and RGB features. +- We develop an exhaustive frequency domain fusion mechanism that models the local similarity and long-range correlation in the frequency domain of NIR and RGB images, and supplements some critical high frequency differential features. + +# 2. Related Work + +# 2.1. Single Image Denoising + +Image denoising [12, 29] is a classic task in low-level vision. In recent years, deep learning has greatly improved the accuracy and efficiency of denoising [2, 7, 14, 22, 23, 35]. DnCNN [50] uses a simple network with batch normalization and residual learning, which outperforms traditional denoising methods. MPRNet [47] designs a multi-stage restoration framework, and Restormer [48] proposes an efficient transformer framework for high-resolution images. Xformer [49] proposes a transformer framework that combines spatial self-attention and channel self-attention. However, despite their strong denoising capabilities, single-image denoising still inevitably loses a lot of spatial information, leading to overly smooth edge textures in the denoised images. + +# 2.2. NIR-Assisted Image Restoration + +To achieve better denoising effects, researchers have attempted to use NIR images to assist in RGB image denoising [8, 26, 38, 46]. Compared with single-image recovery, NIR images can help restore the details of degraded im + +ages [37, 43]. In early research, Yan et al. [46] proposes a scale map from the perspective of gradient inconsistency to find universal usable edges and smooth transitions in NIR images. With the proliferation of deep learning, CUNet [11] uses a convolutional sparse coding model to extract common features from cross-field images. FGDNet [31] and SANet [32] fuse the features of two fields in the frequency domain. MNNet [44] proposes an observation model that takes into account the modality gap between the target and guiding images. DVN [18] integrates structural inconsistency priors into deep features, specifically strengthening the common deep features in NIR and RGB images. NAID [43] proposes Selective Fusion Module (SFM), which can be plugged into advanced image restoration networks. However, under extreme noisy conditions, existing NIR-assisted denoising methods fail to balance color accuracy, detail retention, and artifact suppression. + +# 2.3. Frequency Domain Image Restoration + +Frequency analysis in digital signal processing emphasizes the frequency properties of signals [33, 45]. Recently, frequency domain analysis has shown great potential in improving the performance of various learning-based image restoration frameworks [17, 42, 53]. In [21], An efficient estimation of scaled dot-product attention based on a frequency domain self-attention solver (FSAS) is proposed. HDNet [16] proposes a high-resolution dual-domain learning network. SFNet [10] uses filters to dynamically decouple feature mappings into different frequency components and adaptively extracts useful features. In this paper, we emphasize the importance of frequency domain information for the research on fusion under inconsistency and design a comprehensive frequency domain fusion scheme for this task to ensure the utilization of complementary information from NIR and RGB images. + +# 3. Method + +In this section, we first introduce the cross-field frequency correlation prior of NIR and RGB images. Based on the cross-field frequency correlation prior, a detailed frequency domain scheme is formulated, including the Frequency Dynamic Selection Mechanism (FDSM) and the Frequency Exhaustive Fusion Mechanism (FEFM) for thorough frequency exploitation. + +# 3.1. Cross-Field Frequency Correlation Prior + +To address the fusion of NIR and RGB images under spatial inconsistency, it is imperative to distinguish which information in NIR and RGB images is useful and which is redundant or even erroneous for denoising. We randomly select 30 pairs of NIR images, noisy RGB images, and clean RGB images from the existing RGB-NIR dataset and perform fourier transform on them. To facilitate the study of the + +![](images/096f74f99473b4f900382ea287a60b9aef5d7e855526fdb576d89c7bf4322dd3.jpg) + +![](images/1abe5b305419a3f6cb834643f540880b69dc4499cac145398b035d0757f87b68.jpg) + +![](images/0dbbf62a8d3d8cf87517605b8790a7f54bfe8587b0a7801c430547f3026b957c.jpg) +Figure 3. The analysis of NIR and RGB frequency correlation. (a) The frequency correlation between noisy RGB and clean RGB, with the horizontal axis indicating the cutoff frequency of the high-pass filter. Different colored curves represent different scenarios. (b) The frequency correlation between NIR and clean RGB. (c) The visualization results of images output by a fixed frequency filter, with the first row showing high frequency (HF) part and the second row showing low frequency (LF) part. + +![](images/4af54a2411b9aa6fd12ed836ff2212bb0e02b2889fb2388cd6ba9f4cdd159c59.jpg) +(c) + +![](images/07037fa6455ef9225e2af54ebc34984d995d8d629f0641df51950c43c02d529e.jpg) + +correlation between NIR and RGB in the frequency domain, we apply a high-pass filter to the aforementioned three kinds of frequency domain images and convert the filtered results back to the spatial domain. We separately computed the Structural Similarity (SSIM) [40] between NIR/noisy RGB images and clean RGB images at different cutoff frequencies as a measure of frequency correlation $S$ , as shown in Fig. 3a and Fig. 3b. The process can be described as + +$$ +S = \operatorname {S S I M} \left(\mathcal {F} ^ {- 1} \left(\mathrm {H} \left(F _ {\text {t a r g e t}}\right)\right), \mathcal {F} ^ {- 1} \left(\mathrm {H} \left(F _ {\text {g t}}\right)\right)\right) \tag {1} +$$ + +where $\mathcal{F}^{-1}(\cdot)$ denote 2D-IDFT operations, $F_{target}$ and $F_{gt}$ represent NIR or noisy RGB and clean RGB after 2D-DFT transformation. $\mathrm{H}(\cdot)$ represents the filtering of the input image with a high-pass filter. The analysis results indicate that, in the main energy frequency, the image similarity between the input noisy RGB and clean RGB images decreases progressively as the frequency increases, while the image similarity between the input NIR image and clean RGB image increases from low to high frequency. Fig. 3c also shows that for noisy RGB images, their high frequency information is mainly composed of noise and lacks detail information, while the low frequency information more faithfully reflects the color in the original image, and NIR images are just the opposite. These cross-field Frequency correlation priors indicate that it is simple and efficient to distinguish useful information from NIR and RGB images in the frequency domain, which inspires us to propose a comprehensive frequency domain learning framework. + +![](images/5866955e7c4a4770d9b89dadcfc5461df22c2aca17b6c21291642ac9b6cbe000.jpg) +Noisy RGB (Input) + +![](images/d920fb41101c44be5427913d7679d3f21b9d747a8dd7e9c2785e9acc8d0aeef4.jpg) + +![](images/3f21648d2b5ed62c367a1019c5f30fd90f41e5ca6c6d539c52fb454fed0eb311.jpg) +Clean NIR (Input) +Figure 4. The overall architecture of the proposed cross-field Frequency Correlation Exploiting Network (FCENet). In the first stage, pre-denoising is performed. In the second stage, NIR image is introduced for NIR-assisted image denoising. + +# 3.2. Overview Framework + +The proposed frequency domain fusion framework, as depicted in Fig. 4, includes two stages for the gradual restoration of images. Both stages incorporate an encoder-decoder architecture based on the U-Net [30], featuring two downsampling layers and two upsampling layers. Skip connections are employed between the encoder and decoder features. A Supervised Attention Module (SAM) [47] is utilized to connect features from the first stage to the second stage for further processing. + +In the first stage, the input noisy RGB image undergoes preliminary denoising to reduce the modality gap with the NIR image. In the second stage, we perform multi-scale fusion of the encoded features of the NIR and RGB images. Specifically, at each scale of the encoder, NIR encoded features and RGB pre-denoise features are fed into the frequency domain fusion modules based on FDSM and FEFM, as shown in Fig. 5. + +# 3.3. Frequency Dynamic Selection Mechanism + +The frequency correlation prior indicates that extracting cross-field complementary information from the frequency domain is efficient. For feature extraction, traditional convolutional kernels share weights across the entire feature map, lacking flexibility to address the numerous inconsistencies between NIR and RGB images, which affects the efficiency of extraction. We introduce frequency domain + +dynamic convolution to adapt to specific inputs and effectively extract complementary features. + +Specifically, given the NIR features $\mathbf{N} \in \mathbb{R}^{C \times H \times W}$ and pre-denoised features $\mathbf{R} \in \mathbb{R}^{C \times H \times W}$ , we concatenate them after layernorm and undergo convolution and the activate function to obtain a coarsely fused feature map. After that, we separate the fused feature map to obtain the aggregated features $\mathbf{A}_{\mathbf{R}}$ and $\mathbf{A}_{\mathbf{N}}$ for NIR and RGB individually. These aggregated features are used to generate dynamic filtering kernels to selectively extract useful features from the NIR and RGB input features. + +For dynamic filters, some works [27, 52] that generate deep separable and spatially variant dynamic filtering kernels are usually computationally intensive and time-consuming, and require a large amount of data. We employ a method similar to [6, 34] by generating dynamic weights to dynamically aggregate multiple parallel convolutional kernels related to the input, and performing this process in the frequency domain. Specifically, the aggregated features $\mathbf{A}_{\mathbf{R}}$ and $\mathbf{A}_{\mathbf{N}}$ are used to generate the weights for $k$ learnable frequency domain filters. For each feature channel, these filters are linearly combined to form a frequency domain global filter, which is used to select useful frequency domain information from the input NIR and RGB features. Finally, $\mathbf{F}_{\mathbf{R}}$ , $\mathbf{F}_{\mathbf{N}}$ denoting the filtered useful features of NIR and RGB can be represented as + +$$ +\mathbf {D F} _ {\mathbf {I}} = \operatorname {S o f t m a x} (\operatorname {M L P} (\operatorname {P o o l i n g} (\mathbf {A} _ {\mathbf {I}}))) \cdot \mathbf {G} \tag {2} +$$ + +$$ +\mathbf {F} _ {\mathbf {I}} = \mathcal {F} ^ {- 1} (\mathcal {F} (\mathbf {I}) \odot \mathbf {D F} _ {\mathbf {I}}) \tag {3} +$$ + +where $\mathbf{DF}_{\mathbf{I}}$ is frequency domain dynamic filter, and $\mathbf{I}$ represents $\mathbf{R}$ or $\mathbf{N}$ . $\mathbf{G} \in \mathbb{R}^{M \times k}$ represents the combination of $k$ learnable frequency domain filtering kernels, and $\mathbf{M}$ is the spatial dimension of the convolutional kernel. $\mathcal{F}(\cdot)$ and $\mathcal{F}^{-1}(\cdot)$ denote the 2D-DFT and 2D-IDFT operations, respectively. + +# 3.4. Frequency Exhaustive Fusion Mechanism + +After the NIR and RGB feature maps pass through the FDSM, the useful features are retained. Next, to fuse the two inconsistent fields, we further integrate their information in the frequency domain and propose the FEFM, with this process driven by prior frequency correlations. Specifically, the proposed FEFM includes two key components: Common Feature Reinforcement (CFR) mechanism and Differential Feature Reinforcement (DFR) mechanism. CFR comprehensively explores the correlation between NIR and RGB features in the frequency domain to strengthen the modeling of common features, while DFR employs differential modeling to enhance key differential high frequency features. + +![](images/ed8870cfdcef13a312ceaab3eafb76c2359f11206d819cd593cbb66bfd5bb935.jpg) +Figure 5. The architecture of the proposed Frequency Dynamic Selection Mechanism (FDSM) and Frequency Exhaustive Fusion Mechanism (FEFM) for thorough frequency exploitation. + +![](images/a3813b66fc45b9c7286ce0915b5ead4f9be1099372b027b1b80c5a523c6ca42a.jpg) + +Common Feature Reinforcement Mechanism. The convolution theorem establishes that the correlation or convolution of two signals in the spatial domain is equal to their element-wise product in the frequency domain. Leveraging this property, previous work [21] integrate the frequency domain into the self-attention mechanism, simplifying matrix multiplication into lightweight element-wise dot product operations. Although the dot product is efficient, it only calculates the point-wise correlation coefficients in the frequency domain, missing the long-range correlations in the frequency domain. To more comprehensively model the common features of NIR and RGB, we apply a pointwise convolution and a depthwise convolution to the input NIR and RGB feature maps $\mathbf{F}_{\mathbf{R}}$ , $\mathbf{F}_{\mathbf{N}}$ , to obtain encoded features $\mathbf{Q} = W_{d}^{Q}W_{p}^{Q}\mathbf{F}_{\mathbf{R}}$ , $\mathbf{K} = W_{d}^{K}W_{p}^{K}\mathbf{F}_{\mathbf{N}}$ , $\mathbf{V} = W_{d}^{V}W_{p}^{V}\mathbf{F}_{\mathbf{N}}$ . To facilitate the calculation of frequency domain correlations, we apply 2D-DFT to the input features to produce frequency cubes $\mathbf{F}_{\mathbf{Q}}$ and $\mathbf{F}_{\mathbf{K}}$ . We simultaneously model the long-range correlation and point-wise correlation of NIR and RGB features in the frequency domain, resulting in a correlation map that integrates local similarity and long-range channel correlation within the frequency domain. The output of CFR can be represented as + +$$ +\mathbf {F} _ {\mathbf {C F R}} = \left(\mathbf {F} _ {\mathbf {Q}} \odot \mathbf {F} _ {\mathbf {K}}\right) \cdot \mathrm {S o f t m a x} (\mathbf {F} _ {\mathbf {Q}} \cdot \mathbf {F} _ {\mathbf {K}} ^ {\top} / \alpha) \quad (4) +$$ + +where $\alpha$ is a learnable scaling parameter. + +Differential Feature Reinforcement Mechanism. Traditional cross-attention [15, 25, 39] structures are typically used to capture common features but do not effectively utilize differential information. Although this approach is correct in general image fusion frameworks, in this paper, the features of the two fields are considered useful after passing through the frequency domain selection mechanism. By + +observing the feature maps (Fig. 9), we find that the high frequency texture information in the NIR features is clear compared to the RGB feature maps. The common features obtained by weighted fusion through cross-attention remain somewhat blurry, which greatly affects the denoising effect. These differential high frequency features on NIR are crucial for image restoration. We adopt a differential cross-attention to complement the missing differential information, and the output of CFR can be represented as + +$$ +\mathbf {F} _ {\mathbf {D F R}} = \left(\mathbf {V} - \lambda \mathbf {V} \odot \mathcal {F} ^ {- 1} \left(\mathbf {F} _ {\mathbf {C F R}}\right)\right) \tag {5} +$$ + +where $\lambda$ is a learnable scalar that dynamically controls the weights of two feature maps. Although DFR effectively models the differential high frequency information, the fused features lack some common features and background information. To further refine the fused features, we perform a dot product between the encoded features Q and the output of CFR, and add the result to the output of DFR as the final fused result. + +# 3.5. Loss Function + +We adopt the Charbonnier loss [5], and adding the frequency domain loss [9]. The overall loss function is represented as + +$$ +\begin{array}{l} \mathcal {L} = \mathcal {L} _ {\text {c h a r b o n n i e r}} \left(\mathbf {X} _ {1}, \mathbf {T}\right) + \mathcal {L} _ {\text {c h a r b o n n i e r}} \left(\mathbf {X} _ {2}, \mathbf {T}\right) \tag {6} \\ + \alpha \cdot \mathcal {L} _ {\text {c h a r b o n n i e r}} (\mathcal {F} (\mathbf {X} _ {2}), \mathcal {F} (\mathbf {T})) \\ \end{array} +$$ + +where $\mathbf{X}_1$ and $\mathbf{X}_2$ are the outputs of pre-denoising subnetwork and whole network, respectively. $\mathbf{T}$ is ground truth and $\alpha$ is set as 0.1 [9]. + +![](images/d505563a74781eb99f16582db7a79b4a840e189841ce46afbc36714df2e51c76.jpg) +Figure 6. The qualitative comparison among our FCENet and the state-of-the-art methods on the noisy RGB-NIR pairs from DVD[18] with different noise levels. Our method achieves better results in detail recovery, artifact removal, and color restoration. + +Table 1. Comparison of different methods on the DVD test set with noise levels $\sigma = 2,4,6$ . The best and second-best results are highlighted in boldface and underlined, respectively. + +
Methodsσ = 2σ = 4σ = 6Complexity
PSNR↑SSIM↑PSNR↑SSIM↑PSNR↑SSIM↑FLOPs(G)Params(M)
Uformer [41]31.340.94929.120.92727.580.9089.6020.63
Restormer [48]31.300.94929.200.92827.760.91036.0626.13
MPRNet [47]31.780.95129.370.92927.840.910133.5115.74
Xformer [49]31.410.95029.300.92827.430.90839.3425.22
CUNet [11]28.760.92326.810.89825.640.8756.720.44
SANet [32]30.130.93827.840.91726.410.90125.015.38
MNNet [44]30.140.94428.260.92326.750.90610.970.76
DVN [18]31.500.95529.620.94028.260.92748.306.97
NIR-Restormer [43]31.810.95729.810.94228.310.92845.3130.26
FCENet-L (Ours)31.870.96029.910.94628.360.9345.711.16
FCENet (Ours)32.430.96330.370.95028.780.93919.524.44
+ +# 4. Experiments + +In this section, we evaluate the proposed method on public datasets and compare it with other NIR-assisted image denoising methods. + +# 4.1. Experimental Settings + +Datasets. We evaluate the NIR-assisted image denoising method on the DVD [18] and IVRG [4] datasets. The DVD dataset consists of 307 high-resolution NIR-RGB image pairs, which are converted into 5000 image pairs of size $3 \times 256 \times 256$ for training and 1000 image pairs of size $3 \times 256 \times 256$ for testing. The IVRG dataset contains NIR and RGB image pairs from 9 different scenes. For each scene, as in [44], experiments randomly select 30/8/15 pairs for training/validation/testing. + +Parameter settings. The proposed network is implemented with Pytorch and trained using the Adam optimizer [20] $(\beta_{1} = 0.9, \beta_{2} = 0.999)$ on patches of size $128 \times 128$ with a batch size of 16. For the DVD dataset, we simulate low-light conditions by randomly reducing the mean of the original images as in [18] and add Gaussian-Poisson mixed noise with a range of 1 to 16. The model is trained for 80 epochs. The initial learning rate is $2 \times 10^{-4}$ , and a cosine annealing strategy is used to gradually reduce the learning rate to $1 \times 10^{-6}$ . For the IVRG dataset, following the settings in [44], it is cropped into 16,374 training patches, and then additive Gaussian noise with random noise levels is used to simulate the target images, considering three noise levels in testing: $\sigma = 25$ , 50, and 75. The model is trained for 100 epochs with an initial learning rate of $5 \times 10^{-4}$ . + +![](images/40efdddf4da078317dab4769d3ee5e455ac254c04ffdfa881969bf3f48163ffd.jpg) +Real noisy RGB + +![](images/f6011350f2ffe760c900d560fc9e4ebb58e0e6b4b103774a34699cac3b691679.jpg) +NIR + +![](images/ef188faa8bc5f6c4996a255ad18e4769a0c232fc5cab0b083b406464413bacbf.jpg) +Xformer + +![](images/bcf826294ca471d329803b0a14ae106dcc66b786b76b4369ad0d492855228561.jpg) +SANet + +![](images/97b3fd47acfa9f240401dc6ce6cb0b38ea5313e1047e3244dca4a6c6472e687d.jpg) +MNNet + +![](images/456535714bfa1a68771516b2fa7ca084b3afe05ce4180a22b900ba1591981231.jpg) +DVN + +![](images/6e8981fd2e5ac2e36e1c2c94d0646e6912b6166772e72e329a5db301ffd41821.jpg) +NIR-Restormer + +![](images/2f7e27f46a8dca2591f5672c4b3cdb886d653f922f4f36685ec3d42ae93eabdd.jpg) +Ours + +Table 2. Comparison of different methods on the IVRG test set with noise levels $\sigma = 25, 50, 75$ . The best and second-best results are highlighted in boldface and underlined, respectively. + +
Methodsσ = 25σ = 50σ = 75
PSNR↑SSIM↑PSNR↑SSIM↑PSNR↑SSIM↑
Restormer [48]31.720.87627.670.78824.500.722
Uformer [41]31.970.87827.830.79324.580.726
MPRNet [47]31.890.87827.800.79424.570.729
Xformer [49]31.980.87927.860.79424.610.731
CUNet [11]31.370.87627.780.81224.640.764
SANet [32]31.540.88528.130.82524.810.785
MNNet [44]32.590.90428.660.85125.250.808
DVN [18]32.760.90628.730.85225.280.809
NIR-Restormer [43]32.860.90628.770.85325.300.811
FCENet-L (Ours)32.850.90928.800.85725.330.816
FCENet (Ours)33.260.91529.080.86625.480.826
+ +# 4.2. Experimental Results + +In our experimental evaluation of our method, we compare our results with nine models, including five state-of-the-art NIR-assisted RGB denoising methods: CUNet [11], SANet [32], MNNet [44], DVN [18], and NIR-Restormer [43], and four state-of-the-art single-image denoising methods: MPRNet [47], Restormer [48], Uformer [41], and Xformer [49]. To further demonstrate the superiority of our method, we additionally provide the denoising results of a light version (-L) by changing the initial feature map channel count from 64 to 36, which has only $1.16\mathrm{M}$ parameters. All compared methods are trained on the same training set as ours. + +Evaluations on the DVN dataset. We first evaluate our method on the DVD dataset, with quantitative results shown in Table 1. It can be observed that the proposed FCENet, and even FCENet-L, outperform all other single-image denoising methods and NIR-assisted denoising methods in terms of PSNR and SSIM with moderate computational and memory costs. The qualitative comparison in Fig. 6 clearly illustrates that our method has advantages over other state-of-the-art methods in noise removal, color preservation, and detail recovery. For single-image denoising methods like Restormer, although they have achieved advanced denoising effects in low-noise scenarios, they still lack effective means to address detail texture loss in high-noise situations. For guided denoising methods, MNNet is unable to effec + +![](images/d17f45f53a63c92a062f1cf5ffcf65f1835c6e2eeb48f9a28fe466e3cfe22f51.jpg) +Figure 7. Visual comparisons on real-world RGB/NIR image pairs. Our method demonstrates better visual results. +Figure 8. Frequency domain dynamic filters on NIR and RGB side at different depths. The first row represents the frequency domain filters of RGB side, and the second row represents the filters of NIR side. + +tively handle the inconsistencies between NIR and RGB images. DVN strengthens consistent high frequency textures but loses critical differential high frequency information in NIR and sacrifices the authenticity of low-frequency color. NIR-Restormer has good denoising effects but lacks effective utilization of inconsistent information, leading to over-smoothing and color distortion. Our method recovers detailed texture information while more faithfully preserving low-frequency color information in RGB images. This proves that the proposed method can efficiently utilize the complementary information of NIR and RGB images in the fusion through a series of frequency domain learning. + +**Evaluations on the IVRG dataset.** We evaluate our method on the IVRG dataset. Following the previous method, we consider three noise levels in the test, including $\sigma = 25, 50$ , and 75. Quantitative results are shown in Table 2. It can be observed that compared to the state-of-the-art single-image denoising algorithms, our method significantly improves the performance of image denoising, demonstrating the effectiveness of NIR images. Compared to existing NIR-assisted denoising methods, our method also has significant advantages. + +Evaluations on real-world experiment. To further evaluate the performance of our model in real noisy environments, we conduct qualitative experiments on NIR-RGB data pairs in real low-light conditions. As shown in Fig. 7, the introduction of NIR images brings more detailed textures but also a large number of inconsistencies. SANet fails to handle inconsistencies, resulting in a significant amount of artifacts, while MNNet, DVN and NIR-Restormer are unable to thoroughly denoise. In contrast, our method can eliminate inconsistencies while introducing more detailed information. + +Table 3. Ablation study of various components of our method. + +
BaselineFDSMFEFMPSNR↑SSIM↑
(a)✓29.130.936
(b)✓✓29.640.944
(c)✓✓29.960.947
(d)✓✓✓30.370.950
+ +Table 4. Ablation on the branches of FDSM. + +
MethodsPSNR↑SSIM↑
w/o FDSM29.960.947
Spatial Dynamic Selection30.110.948
Frequency Dynamic Selection30.370.950
+ +Table 5. Ablation on the branches of FEFM. + +
MethodsPSNR↑SSIM↑
Sum29.640.944
Cross-attention29.850.946
DFR30.150.948
DFR + CFR30.370.950
+ +# 4.3. Ablation Study + +We conduct ablation studies with FCENet on the DVD dataset under $\sigma = 4$ . We compare two mechanisms proposed in this paper with a two-stage U-Net baseline, with the results shown in Table 3. In the U-Net baseline, crossfield images are fused through addition. By deploying FDSM and FEFM, the model achieves gains of $0.51\mathrm{dB}$ and $0.83\mathrm{dB}$ , respectively. Finally, combining all contributions significantly improves the denoising level compared to the baseline. + +Effect of FDSM. Using FDSM allows for better extraction of complementary information from NIR and RGB while handling inconsistent regions. In Fig. 8, we show the frequency domain dynamic filter kernels in FDSM. It can be observed that the filter kernels processing NIR feature maps regularly enhance the learning of mid-to-high frequency information, while those processing RGB feature maps mainly learn low frequency information from the input images. This is consistent with the conclusions drawn from our frequency domain correlation priors. Additionally, we conduct experiments using spatial dynamic convolution in this module. The results in Table 4 ultimately prove the effectiveness of extracting relevant features from the frequency domain in NIR and RGB images and the stability of dynamic convolution in extracting inconsistent features. + +Effect of FEFM. Traditional cross-attention fusion modules only strengthen common features but neglect differential features. FEFM strengthens the learning of common features and enhances the learning of some high frequency differential information in NIR feature maps. Fig. 9 shows + +![](images/21b075348c642adcbe9eb117e4a3c2668d4b456f4430f0feee086f1a0f69c0a9.jpg) +Figure 9. Feature maps obtained from two fields. The first column represents the input feature maps of NIR and RGB, the second column represents the output feature of cross-attention, and the third column is the output feature of FEFM. + +the input NIR and RGB feature maps after frequency domain selection, the fused feature maps after cross-attention fusion, and the fused feature maps after FEFM. Compared to the NIR feature maps, the input RGB feature maps still lack high frequency features. It can be seen that the proposed FEFM strengthens common features while supplementing some differential texture information from NIR. We compare two mechanisms in FEFM, CFR and DFR, with traditional cross-attention fusion modules and direct feature summation. The results in Table 5 prove effectiveness of our method in strengthening common features in frequency domain and modeling differential features. + +# 5. Conclusion + +In this paper, a cross-field frequency correlation exploiting network for NIR-assisted image denoising is proposed. Specifically, we first conduct a series of analyses on the characteristics of NIR and RGB images in frequency domain and obtain frequency correlation prior. Based on the prior, the frequency dynamic selection mechanism dynamically extracts complementary information from cross-field images in the frequency domain, and the frequency exhaustive fusion mechanism is proposed to enhance common features and differential high frequency features during fusion process. Extensive experiments demonstrate that the proposed method is significantly better than other algorithms. It is worth mentioning that existing methods may fail on significantly misaligned cross-field images. In future work, we plan to explore advanced methods such as feature point matching or incorporating deformable convolutional kernels to handle significantly misaligned cross-field images. + +# Acknowledgments + +This work is supported by the National Natural Science Foundation of China under Grants 62322204, 62131003, and the Key Laboratory of Target Cognition and Application Technology. + +# References + +[1] Abdelrahman Abdelhamed, Mahmoud Affifi, Radu Timofte, and Michael S Brown. Ntire 2020 challenge on real image denoising: Dataset, methods and results. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 496-497, 2020. 1 +[2] Saeed Anwar and Nick Barnes. Real image denoising with feature attention. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3155-3164, 2019. 2 +[3] Joshua Batson and Loic Royer. Noise2self: Blind denoising by self-supervision. In International Conference on Machine Learning, pages 524-533, 2019. 1 +[4] Matthew Brown and Sabine Susstrunk. Multi-spectral sift for scene category recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 177-184, 2011. 6 +[5] Pierre Charbonnier, Laure Blanc-Feraud, Gilles Aubert, and Michel Barlaud. Two deterministic half-quadratic regularization algorithms for computed imaging. In Proceedings of the International Conference on Image Processing, pages 168-172, 1994. 5 +[6] Yinpeng Chen, Xiyang Dai, Mengchen Liu, Dongdong Chen, Lu Yuan, and Zicheng Liu. Dynamic convolution: Attention over convolution kernels. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11030-11039, 2020. 4 +[7] Shen Cheng, Yuzhi Wang, Haibin Huang, Donghao Liu, Haoqiang Fan, and Shuaicheng Liu. Nbnet: Noise basis learning for image denoising with subspace projection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4896-4906, 2021. 2 +[8] Yuxiao Cheng, Runzhao Yang, Zhihong Zhang, Jinli Suo, and Qionghai Dai. A mutually boosting dual sensor computational camera for high quality dark videography. Information Fusion, 93:429-440, 2023. 2 +[9] Sung-Jin Cho, Seo-Won Ji, Jun-Pyo Hong, Seung-Won Jung, and Sung-Jea Ko. Rethinking coarse-to-fine approach in single image deblurring. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4641-4650, 2021. 5 +[10] Yuning Cui, Yi Tao, Zhenshan Bing, Wenqi Ren, Xinwei Gao, Xiaochun Cao, Kai Huang, and Alois Knoll. Selective frequency network for image restoration. In International Conference on Learning Representations, 2023. 3 +[11] Xin Deng and Pier Luigi Dragotti. Deep convolutional neural network for multi-modal image restoration and fusion. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(10):3333-3348, 2020. 2, 3, 6, 7 +[12] Michael Elad and Michal Aharon. Image denoising via sparse and redundant representations over learned dictionaries. IEEE Transactions on Image processing, 15(12):3736-3745, 2006. 2 +[13] Hansen Feng, Lizhi Wang, Yuzhi Wang, and Hua Huang. Learnability enhancement for low-light raw denoising: Where paired real data meets noise modeling. In Proceed- + +ings of the ACM International Conference on Multimedia, pages 1436-1444, 2022. 1 +[14] Hansen Feng, Lizhi Wang, Yuzhi Wang, Haoqiang Fan, and Hua Huang. Learnability enhancement for low-light raw image denoising: A data perspective. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(1):370-387, 2024. 2 +[15] Yuanshen Guan, Ruikang Xu, Mingde Yao, Lizhi Wang, and Zhiwei Xiong. Mutual-guided dynamic network for image fusion. In Proceedings of the ACM International Conference on Multimedia, pages 1779-1788, 2023. 5 +[16] Xiaowan Hu, Yuanhao Cai, Jing Lin, Haoqian Wang, Xin Yuan, Yulun Zhang, Radu Timofte, and Luc Van Gool. Hdnet: High-resolution dual-domain learning for spectral compressive imaging. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17542-17551, 2022. 3 +[17] Liming Jiang, Bo Dai, Wayne Wu, and Chen Change Loy. Focal frequency loss for image reconstruction and synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13919-13929, 2021. 3 +[18] Shuangping Jin, Bingbing Yu, Minhao Jing, Yi Zhou, Jiajun Liang, and Renhe Ji. Darkvisionnet: Low-light imaging via rgb-nir fusion with deep inconsistency prior. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1104-1112, 2022. 1, 2, 3, 6, 7 +[19] Beomjun Kim, Jean Ponce, and Bumsub Ham. Deformable kernel networks for joint image filtering. International Journal of Computer Vision, 129(2):579-600, 2021. 2 +[20] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6 +[21] Lingshun Kong, Jiangxin Dong, Jianjun Ge, Mingqiang Li, and Jinshan Pan. Efficient frequency domain-based transformers for high-quality image deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5886-5895, 2023. 3, 5 +[22] Tong Li, Hansen Feng, Lizhi Wang, Zhiwei Xiong, and Hua Huang. Stimulating the diffusion model for image denoising via adaptive embedding and ensembling. In IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 8240-8257, 2024. 2 +[23] Tong Li, Lizhi Wang, Zhiyuan Xu, Lin Zhu, Wanxuan Lu, and Hua Huang. Positive2negative: Breaking the information-lossy barrier in self-supervised single image denoising. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2025. 2 +[24] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Swinir: Image restoration using swim transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1833-1844, 2021. 1 +[25] Hezheng Lin, Xing Cheng, Xiangyu Wu, and Dong Shen. Cat: Cross attention in vision transformer. In Proceedings of the IEEE International Conference on Multimedia and Expo, pages 1-6, 2022. 5 + +[26] Feifan Lv, Yinqiang Zheng, Yicheng Li, and Feng Lu. An integrated enhancement solution for 24-hour colorful imaging. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 11725-11732, 2020. 1, 2 +[27] Ruijun Ma, Shuyi Li, Bob Zhang, and Zhengming Li. Generative adaptive convolutions for real-world noisy image denoising. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1935-1943, 2022. 4 +[28] Youssef Mansour and Reinhard Heckel. Zero-shot noise2noise: Efficient image denoising without any data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14018-14027, 2023. 1 +[29] Javier Portilla, Vasily Strela, Martin J Wainwright, and Eero P Simoncelli. Image denoising using scale mixtures of gaussians in the wavelet domain. IEEE Transactions on Image Processing, 12(11):1338-1351, 2003. 2 +[30] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention, pages 234-241, 2015. 4 +[31] Zehua Sheng, Xiongwei Liu, Si-Yuan Cao, Hui-Liang Shen, and Huaqi Zhang. Frequency-domain deep guided image denoising. IEEE Transactions on Multimedia, 25:6767-6781, 2022. 2, 3 +[32] Zehua Sheng, Zhu Yu, Xiongwei Liu, Si-Yuan Cao, Yuqi Liu, Hui-Liang Shen, and Huaqi Zhang. Structure aggregation for cross-spectral stereo image guided denoising. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13997-14006, 2023. 2, 3, 6, 7 +[33] Matthew Tancik, Pratul Srinivasan, Ben Mildenhall, Sara Fridovich-Keil, Nithin Raghavan, Utkarsh Singhal, Ravi Ramamoorthi, Jonathan Barron, and Ren Ng. Fourier features let networks learn high frequency functions in low dimensional domains. Advances in Neural Information Processing Systems, 33:7537-7547, 2020. 3 +[34] Yuki Tatsunami and Masato Taki. Fft-based dynamic token mixer for vision. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 15328-15336, 2024. 4 +[35] Chunwei Tian, Lunke Fei, Wenxian Zheng, Yong Xu, Wangmeng Zuo, and Chia-Wen Lin. Deep learning on image denoising: An overview. *Neural Networks*, 131:251–275, 2020. 2 +[36] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in Neural Information Processing Systems, 30, 2017. 1 +[37] Renjie Wan, Boxin Shi, Wenhan Yang, Bihan Wen, LingYu Duan, and Alex C. Kot. Purifying low-light images via near-infrared enlightened image. IEEE Transactions on Multimedia, 25:8006-8019, 2023. 1, 3 +[38] Jian Wang, Tianfan Xue, Jonathan T Barron, and Jiawen Chen. Stereoscopic dark flash for low-light photography. In IEEE International Conference on Computational Photography, pages 1-10, 2019. 2 +[39] Xin Wang, Lizhi Wang, Xiangtian Ma, Maoqing Zhang, Lin Zhu, and Hua Huang. In2set: Intra-inter similarity exploiting transformer for dual-camera compressive hyperspectral + +imaging. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24881-24891, 2024. 5 +[40] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing, 13(4):600-612, 2004. 3 +[41] Zhendong Wang, Xiaodong Cun, Jianmin Bao, Wengang Zhou, Jianzhuang Liu, and Houqiang Li. Uformer: A general u-shaped transformer for image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17683-17693, 2022. 6, 7 +[42] Yunxuan Wei, Shuhang Gu, Yawei Li, Radu Timofte, Longcun Jin, and Hengjie Song. Unsupervised real-world image super resolution via domain-distance aware training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13385-13394, 2021. 3 +[43] Rongjian Xu, Zhilu Zhang, Renlong Wu, and Wangmeng Zuo. Nir-assisted image denoising: A selective fusion approach and a real-world benchmark dataset. arXiv preprint arXiv:2404.08514, 2024. 2, 3, 6, 7 +[44] Shuang Xu, Jiangshe Zhang, Jialin Wang, Kai Sun, Chunxia Zhang, Junmin Liu, and Junying Hu. A model-driven network for guided image denoising. Information Fusion, 85: 60-71, 2022. 2, 3, 6, 7 +[45] Zhi-Qin John Xu, Yaoyu Zhang, Tao Luo, Yanyang Xiao, and Zheng Ma. Frequency principle: Fourier analysis sheds light on deep neural networks. arXiv preprint arXiv:1901.06523, 2019. 3 +[46] Qiong Yan, Xiaoyong Shen, Li Xu, Shaojie Zhuo, Xiaopeng Zhang, Liang Shen, and Jiaya Jia. Cross-field joint image restoration via scale map. In Proceedings of the IEEE International Conference on Computer Vision, pages 1537-1544, 2013. 1, 2, 3 +[47] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, Ming-Hsuan Yang, and Ling Shao. Multi-stage progressive image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14821-14831, 2021. 1, 2, 4, 6, 7 +[48] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5728-5739, 2022. 2, 6, 7 +[49] Jiale Zhang, Yulun Zhang, Jinjin Gu, Jiahua Dong, Linghe Kong, and Xiaokang Yang. Xformer: Hybrid x-shaped transformer for image denoising. arXiv preprint arXiv:2303.06440, 2023. 2, 6, 7 +[50] Kai Zhang, Wangmeng Zuo, Yunjin Chen, Deyu Meng, and Lei Zhang. Beyond a gaussian denoiser: Residual learning of deep cnn for image denoising. IEEE Transactions on Image Processing, 26(7):3142-3155, 2017. 2 +[51] Kai Zhang, Wangmeng Zuo, and Lei Zhang. Ffdnet: Toward a fast and flexible solution for cnn-based image denoising. IEEE Transactions on Image Processing, 27(9):4608-4622, 2018. 1 + +[52] Jingkai Zhou, Varun Jampani, Zhixiong Pi, Qiong Liu, and Ming-Hsuan Yang. Decoupled dynamic filter networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6647-6656, 2021. 4 +[53] Wenbin Zou, Mingchao Jiang, Yunchen Zhang, Liang Chen, Zhiyong Lu, and Yi Wu. Sdwnet: A straight dilated network with wavelet transformation for image deblurring. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1895–1904, 2021. 3 \ No newline at end of file diff --git a/2025/Complementary Advantages_ Exploiting Cross-Field Frequency Correlation for NIR-Assisted Image Denoising/images.zip b/2025/Complementary Advantages_ Exploiting Cross-Field Frequency Correlation for NIR-Assisted Image Denoising/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..052197dac9024eedddd595769b428b5c23e32032 --- /dev/null +++ b/2025/Complementary Advantages_ Exploiting Cross-Field Frequency Correlation for NIR-Assisted Image Denoising/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:19e4380066cbfc285a60d2d1170aad0034234839b0de898d8c5501c116a69f3a +size 725039 diff --git a/2025/Complementary Advantages_ Exploiting Cross-Field Frequency Correlation for NIR-Assisted Image Denoising/layout.json b/2025/Complementary Advantages_ Exploiting Cross-Field Frequency Correlation for NIR-Assisted Image Denoising/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..b65e11ff13791cf6ec2ad9a1f588de88b676caf6 --- /dev/null +++ b/2025/Complementary Advantages_ Exploiting Cross-Field Frequency Correlation for NIR-Assisted Image Denoising/layout.json @@ -0,0 +1,8450 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 60, + 102, + 550, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 102, + 550, + 140 + ], + "spans": [ + { + "bbox": [ + 60, + 102, + 550, + 140 + ], + "type": "text", + "content": "Complementary Advantages: Exploiting Cross-Field Frequency Correlation for NIR-Assisted Image Denoising" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 182, + 160, + 468, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 182, + 160, + 468, + 175 + ], + "spans": [ + { + "bbox": [ + 182, + 160, + 468, + 175 + ], + "type": "text", + "content": "Yuchen Wang" + }, + { + "bbox": [ + 182, + 160, + 468, + 175 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 182, + 160, + 468, + 175 + ], + "type": "text", + "content": " Hongyuan Wang" + }, + { + "bbox": [ + 182, + 160, + 468, + 175 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 182, + 160, + 468, + 175 + ], + "type": "text", + "content": " Lizhi Wang" + }, + { + "bbox": [ + 182, + 160, + 468, + 175 + ], + "type": "inline_equation", + "content": "^{2,3*}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 169, + 175, + 481, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 169, + 175, + 481, + 190 + ], + "spans": [ + { + "bbox": [ + 169, + 175, + 481, + 190 + ], + "type": "text", + "content": "Xin Wang" + }, + { + "bbox": [ + 169, + 175, + 481, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 169, + 175, + 481, + 190 + ], + "type": "text", + "content": " Lin Zhu" + }, + { + "bbox": [ + 169, + 175, + 481, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 169, + 175, + 481, + 190 + ], + "type": "text", + "content": " Wanxuan Lu" + }, + { + "bbox": [ + 169, + 175, + 481, + 190 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 169, + 175, + 481, + 190 + ], + "type": "text", + "content": " Hua Huang" + }, + { + "bbox": [ + 169, + 175, + 481, + 190 + ], + "type": "inline_equation", + "content": "^{2,3}" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 190, + 517, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 190, + 517, + 204 + ], + "spans": [ + { + "bbox": [ + 133, + 190, + 517, + 204 + ], + "type": "text", + "content": "1 School of Computer Science and Technology, Beijing Institute of Technology" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 176, + 204, + 473, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 176, + 204, + 473, + 217 + ], + "spans": [ + { + "bbox": [ + 176, + 204, + 473, + 217 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 176, + 204, + 473, + 217 + ], + "type": "text", + "content": " School of Artificial Intelligence, Beijing Normal University" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 61, + 217, + 588, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 217, + 588, + 232 + ], + "spans": [ + { + "bbox": [ + 61, + 217, + 588, + 232 + ], + "type": "text", + "content": "3 Engineering Research Center of Intelligent Technology and Educational Application, Ministry of Education" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 144, + 232, + 506, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 144, + 232, + 506, + 245 + ], + "spans": [ + { + "bbox": [ + 144, + 232, + 506, + 245 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 144, + 232, + 506, + 245 + ], + "type": "text", + "content": " Aerospace Information Research Institute, Chinese Academy of Sciences" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 151, + 273, + 200, + 285 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 273, + 200, + 285 + ], + "spans": [ + { + "bbox": [ + 151, + 273, + 200, + 285 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 298, + 297, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 298, + 297, + 598 + ], + "spans": [ + { + "bbox": [ + 53, + 298, + 297, + 598 + ], + "type": "text", + "content": "Existing single-image denoising algorithms often struggle to restore details when dealing with complex noisy images. The introduction of near-infrared (NIR) images offers new possibilities for RGB image denoising. However, due to the inconsistency between NIR and RGB images, the existing works still struggle to balance the contributions of two fields in the process of image fusion. In response to this, in this paper, we develop a cross-field Frequency Correlation Exploiting Network (FCENet) for NIR-assisted image denoising. We first propose the frequency correlation prior based on an in-depth statistical frequency analysis of NIR-RGB image pairs. The prior reveals the complementary correlation of NIR and RGB images in the frequency domain. Leveraging frequency correlation prior, we then establish a frequency learning framework composed of Frequency Dynamic Selection Mechanism (FDSM) and Frequency Exhaustive Fusion Mechanism (FEFM). FDSM dynamically selects complementary information from NIR and RGB images in the frequency domain, and FEFM strengthens the control of common and differential features during the fusion process of NIR and RGB features. Extensive experiments on simulated and real data validate that the proposed method outperforms other state-of-the-art methods. The code will be released at https://github.com/yuchenwang815/FCENet." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 623, + 135, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 623, + 135, + 635 + ], + "spans": [ + { + "bbox": [ + 56, + 623, + 135, + 635 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 643, + 296, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 643, + 296, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 643, + 296, + 693 + ], + "type": "text", + "content": "Due to the inherent physical limitations of digital imaging devices, the captured images are often affected by various types of noise, which have a considerable impact on applications such as 24-hour surveillance, autonomous driving," + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 318, + 275, + 552, + 449 + ], + "blocks": [ + { + "bbox": [ + 318, + 275, + 552, + 449 + ], + "lines": [ + { + "bbox": [ + 318, + 275, + 552, + 449 + ], + "spans": [ + { + "bbox": [ + 318, + 275, + 552, + 449 + ], + "type": "image", + "image_path": "cd04717a2e57779998dce53b0b02024e66048ef79e423846adb28afdd725e4c7.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 454, + 555, + 521 + ], + "lines": [ + { + "bbox": [ + 313, + 454, + 555, + 521 + ], + "spans": [ + { + "bbox": [ + 313, + 454, + 555, + 521 + ], + "type": "text", + "content": "Figure 1. Comparisons of PSNR, FLOPs, and Parameters on the DVD dataset [18] are presented. Green circles represent the single-image denoising algorithms, and blue circles represent the NIR-assisted image denoising algorithms. The circle radius represents the number of parameters. The proposed method (red circles) achieves the superior performance while maintaining efficiency." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 312, + 533, + 555, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 533, + 555, + 688 + ], + "spans": [ + { + "bbox": [ + 312, + 533, + 555, + 688 + ], + "type": "text", + "content": "and smartphone photography. Although the development of deep learning techniques [24, 36] has brought forth a variety of image denoising methods [1, 3, 13, 28, 47, 51], existing algorithms still find it difficult to deal with complex noisy images, especially in low-light environment. Restoring clean detailed information from noisy images that lack substantial information is highly ill-posed. Therefore, it is necessary to introduce supplementary information from new fields. Fortunately, NIR images with high Signal to Noise Ratio (SNR) can be captured at a low cost and are suitable for everyday photography. Thus, NIR-assisted image denoising [26, 37, 46] has become a promising solution for image denoising." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 689, + 555, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 712 + ], + "type": "text", + "content": "However, it is challenging to effectively leverage information from different fields to achieve noise reduction. Due" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 702, + 268, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 702, + 268, + 713 + ], + "spans": [ + { + "bbox": [ + 66, + 702, + 268, + 713 + ], + "type": "text", + "content": "*Corresponding Author: Lizhi Wang (wanglizhi@bnu.edu.cn)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12679" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 71, + 295, + 154 + ], + "blocks": [ + { + "bbox": [ + 57, + 71, + 295, + 154 + ], + "lines": [ + { + "bbox": [ + 57, + 71, + 295, + 154 + ], + "spans": [ + { + "bbox": [ + 57, + 71, + 295, + 154 + ], + "type": "image", + "image_path": "b49cca02186ed0919b0d221791bed6ad52902dfa423cbcdb42edcb991e259751.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 157, + 295, + 201 + ], + "lines": [ + { + "bbox": [ + 55, + 157, + 295, + 201 + ], + "spans": [ + { + "bbox": [ + 55, + 157, + 295, + 201 + ], + "type": "text", + "content": "Figure 2. Visual comparisons on the challenging noisy RGB-NIR image pairs. Pay attention to the differences in structural and color information between two fields. Our method produces a better denoising result with clear details and fewer artifacts." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 209, + 295, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 209, + 295, + 317 + ], + "spans": [ + { + "bbox": [ + 54, + 209, + 295, + 317 + ], + "type": "text", + "content": "to the influence of the inherent reflective spectra of objects, there are inconsistencies between NIR and RGB images. As shown in Fig. 2, firstly, there is a significant difference in color and brightness information between two fields. Secondly, some structural texture of RGB image disappears in NIR image, and additional artifacts may appear. The color and structural inconsistencies between NIR and RGB images pose a significant challenge for the task of NIR-assisted image denoising." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 318, + 295, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 318, + 295, + 484 + ], + "spans": [ + { + "bbox": [ + 55, + 318, + 295, + 484 + ], + "type": "text", + "content": "In existing researches, some methods attempt to directly integrate inconsistent features from two fields [31, 32, 43, 44]. However, these methods indiscriminately fuse crossfield information, relying on the fitting capabilities of neural networks to learn the fusion of NIR and RGB images under spatial inconsistency, which often leads to suboptimal denoising results and lower computational efficiency. Some methods reinforce common features from NIR and RGB images and diminish the weight of inconsistent features during the fusion [11, 18, 19, 46]. These image fusion schemes typically only enhance the common information in NIR and RGB during fusion, but some differential information such as the sharp textures in NIR and the color information in RGB is discarded." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 486, + 295, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 486, + 295, + 652 + ], + "spans": [ + { + "bbox": [ + 55, + 486, + 295, + 652 + ], + "type": "text", + "content": "In this paper, we develop a cross-field Frequency Correlation Exploiting Network (FCENet) for NIR-assisted image denoising. We conduct a series of analyses on the characteristics of NIR and RGB images in the frequency domain and obtain a key observation: in the same scene, the similarity between noisy RGB images and clean RGB (noise-free RGB) images decreases from low to high frequency, whereas the similarity between NIR images and clean RGB images increases from low to high frequency. This significant cross-field frequency correlation prior redefines crossfield fusion as a process of frequency domain adaptive selection, motivating us to subsequently establish detailed frequency domain learning to precisely control the restoration of texture details and colors." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 654, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 295, + 713 + ], + "type": "text", + "content": "Specifically, the frequency domain fusion network consists of two mechanisms: Frequency Dynamic Selection Mechanism (FDSM) and Frequency Exhaustive Fusion Mechanism (FEFM). In order to comprehensively extract useful features from inconsistent images, the FDSM is for-" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 553, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 191 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 191 + ], + "type": "text", + "content": "mulated based on dynamic filters under the guidance of frequency domain prior. The FDSM efficiently screens complementary features from NIR and RGB in the frequency domain, providing high quality material for subsequent image fusion. Then, the FEFM is formulated to ensure thorough frequency domain fusion. The FEFM integrates local and long-range correlations across fields to fully extract the common feature from cross-field information, and employs a differential cross-attention approach to complement some key differential high frequency features in NIR features." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 192, + 555, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 192, + 555, + 263 + ], + "spans": [ + { + "bbox": [ + 313, + 192, + 555, + 263 + ], + "type": "text", + "content": "The proposed FCENet can utilize complementary information from cross-field images. Extensive experiments on simulated and real data validate that our method outperforms other state-of-the-art methods in terms of image quality and computational efficiency. The main contributions of this work are summarized as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 265, + 553, + 421 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 314, + 265, + 553, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 265, + 553, + 312 + ], + "spans": [ + { + "bbox": [ + 314, + 265, + 553, + 312 + ], + "type": "text", + "content": "- We establish the frequency correlation prior between NIR and RGB images through a series of frequency domain analyses of two fields, which prompts the integration of complementary features in the frequency domain." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 315, + 313, + 553, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 313, + 553, + 360 + ], + "spans": [ + { + "bbox": [ + 315, + 313, + 553, + 360 + ], + "type": "text", + "content": "- We propose an efficient frequency dynamic selection mechanism based on the frequency correlation prior to identify which useful low-frequency and high frequency information should be retained in NIR and RGB features." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 315, + 361, + 553, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 361, + 553, + 421 + ], + "spans": [ + { + "bbox": [ + 315, + 361, + 553, + 421 + ], + "type": "text", + "content": "- We develop an exhaustive frequency domain fusion mechanism that models the local similarity and long-range correlation in the frequency domain of NIR and RGB images, and supplements some critical high frequency differential features." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 433, + 400, + 445 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 433, + 400, + 445 + ], + "spans": [ + { + "bbox": [ + 314, + 433, + 400, + 445 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 453, + 447, + 466 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 453, + 447, + 466 + ], + "spans": [ + { + "bbox": [ + 313, + 453, + 447, + 466 + ], + "type": "text", + "content": "2.1. Single Image Denoising" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 471, + 555, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 471, + 555, + 639 + ], + "spans": [ + { + "bbox": [ + 313, + 471, + 555, + 639 + ], + "type": "text", + "content": "Image denoising [12, 29] is a classic task in low-level vision. In recent years, deep learning has greatly improved the accuracy and efficiency of denoising [2, 7, 14, 22, 23, 35]. DnCNN [50] uses a simple network with batch normalization and residual learning, which outperforms traditional denoising methods. MPRNet [47] designs a multi-stage restoration framework, and Restormer [48] proposes an efficient transformer framework for high-resolution images. Xformer [49] proposes a transformer framework that combines spatial self-attention and channel self-attention. However, despite their strong denoising capabilities, single-image denoising still inevitably loses a lot of spatial information, leading to overly smooth edge textures in the denoised images." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 647, + 488, + 661 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 647, + 488, + 661 + ], + "spans": [ + { + "bbox": [ + 313, + 647, + 488, + 661 + ], + "type": "text", + "content": "2.2. NIR-Assisted Image Restoration" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "type": "text", + "content": "To achieve better denoising effects, researchers have attempted to use NIR images to assist in RGB image denoising [8, 26, 38, 46]. Compared with single-image recovery, NIR images can help restore the details of degraded im" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12680" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 277 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 277 + ], + "type": "text", + "content": "ages [37, 43]. In early research, Yan et al. [46] proposes a scale map from the perspective of gradient inconsistency to find universal usable edges and smooth transitions in NIR images. With the proliferation of deep learning, CUNet [11] uses a convolutional sparse coding model to extract common features from cross-field images. FGDNet [31] and SANet [32] fuse the features of two fields in the frequency domain. MNNet [44] proposes an observation model that takes into account the modality gap between the target and guiding images. DVN [18] integrates structural inconsistency priors into deep features, specifically strengthening the common deep features in NIR and RGB images. NAID [43] proposes Selective Fusion Module (SFM), which can be plugged into advanced image restoration networks. However, under extreme noisy conditions, existing NIR-assisted denoising methods fail to balance color accuracy, detail retention, and artifact suppression." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 282, + 257, + 295 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 282, + 257, + 295 + ], + "spans": [ + { + "bbox": [ + 55, + 282, + 257, + 295 + ], + "type": "text", + "content": "2.3. Frequency Domain Image Restoration" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 300, + 296, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 300, + 296, + 491 + ], + "spans": [ + { + "bbox": [ + 55, + 300, + 296, + 491 + ], + "type": "text", + "content": "Frequency analysis in digital signal processing emphasizes the frequency properties of signals [33, 45]. Recently, frequency domain analysis has shown great potential in improving the performance of various learning-based image restoration frameworks [17, 42, 53]. In [21], An efficient estimation of scaled dot-product attention based on a frequency domain self-attention solver (FSAS) is proposed. HDNet [16] proposes a high-resolution dual-domain learning network. SFNet [10] uses filters to dynamically decouple feature mappings into different frequency components and adaptively extracts useful features. In this paper, we emphasize the importance of frequency domain information for the research on fusion under inconsistency and design a comprehensive frequency domain fusion scheme for this task to ensure the utilization of complementary information from NIR and RGB images." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 501, + 111, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 501, + 111, + 514 + ], + "spans": [ + { + "bbox": [ + 55, + 501, + 111, + 514 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 521, + 296, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 296, + 606 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 296, + 606 + ], + "type": "text", + "content": "In this section, we first introduce the cross-field frequency correlation prior of NIR and RGB images. Based on the cross-field frequency correlation prior, a detailed frequency domain scheme is formulated, including the Frequency Dynamic Selection Mechanism (FDSM) and the Frequency Exhaustive Fusion Mechanism (FEFM) for thorough frequency exploitation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 612, + 270, + 625 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 612, + 270, + 625 + ], + "spans": [ + { + "bbox": [ + 55, + 612, + 270, + 625 + ], + "type": "text", + "content": "3.1. Cross-Field Frequency Correlation Prior" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "content": "To address the fusion of NIR and RGB images under spatial inconsistency, it is imperative to distinguish which information in NIR and RGB images is useful and which is redundant or even erroneous for denoising. We randomly select 30 pairs of NIR images, noisy RGB images, and clean RGB images from the existing RGB-NIR dataset and perform fourier transform on them. To facilitate the study of the" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 320, + 72, + 433, + 171 + ], + "blocks": [ + { + "bbox": [ + 320, + 72, + 433, + 171 + ], + "lines": [ + { + "bbox": [ + 320, + 72, + 433, + 171 + ], + "spans": [ + { + "bbox": [ + 320, + 72, + 433, + 171 + ], + "type": "image", + "image_path": "096f74f99473b4f900382ea287a60b9aef5d7e855526fdb576d89c7bf4322dd3.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 433, + 72, + 548, + 171 + ], + "blocks": [ + { + "bbox": [ + 433, + 72, + 548, + 171 + ], + "lines": [ + { + "bbox": [ + 433, + 72, + 548, + 171 + ], + "spans": [ + { + "bbox": [ + 433, + 72, + 548, + 171 + ], + "type": "image", + "image_path": "1abe5b305419a3f6cb834643f540880b69dc4499cac145398b035d0757f87b68.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 320, + 173, + 400, + 269 + ], + "blocks": [ + { + "bbox": [ + 320, + 173, + 400, + 269 + ], + "lines": [ + { + "bbox": [ + 320, + 173, + 400, + 269 + ], + "spans": [ + { + "bbox": [ + 320, + 173, + 400, + 269 + ], + "type": "image", + "image_path": "0dbbf62a8d3d8cf87517605b8790a7f54bfe8587b0a7801c430547f3026b957c.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 278, + 555, + 368 + ], + "lines": [ + { + "bbox": [ + 313, + 278, + 555, + 368 + ], + "spans": [ + { + "bbox": [ + 313, + 278, + 555, + 368 + ], + "type": "text", + "content": "Figure 3. The analysis of NIR and RGB frequency correlation. (a) The frequency correlation between noisy RGB and clean RGB, with the horizontal axis indicating the cutoff frequency of the high-pass filter. Different colored curves represent different scenarios. (b) The frequency correlation between NIR and clean RGB. (c) The visualization results of images output by a fixed frequency filter, with the first row showing high frequency (HF) part and the second row showing low frequency (LF) part." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 402, + 173, + 472, + 269 + ], + "blocks": [ + { + "bbox": [ + 402, + 173, + 472, + 269 + ], + "lines": [ + { + "bbox": [ + 402, + 173, + 472, + 269 + ], + "spans": [ + { + "bbox": [ + 402, + 173, + 472, + 269 + ], + "type": "image", + "image_path": "4af54a2411b9aa6fd12ed836ff2212bb0e02b2889fb2388cd6ba9f4cdd159c59.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 427, + 269, + 438, + 277 + ], + "lines": [ + { + "bbox": [ + 427, + 269, + 438, + 277 + ], + "spans": [ + { + "bbox": [ + 427, + 269, + 438, + 277 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 474, + 173, + 545, + 269 + ], + "blocks": [ + { + "bbox": [ + 474, + 173, + 545, + 269 + ], + "lines": [ + { + "bbox": [ + 474, + 173, + 545, + 269 + ], + "spans": [ + { + "bbox": [ + 474, + 173, + 545, + 269 + ], + "type": "image", + "image_path": "07037fa6455ef9225e2af54ebc34984d995d8d629f0641df51950c43c02d529e.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 376, + 555, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 376, + 555, + 473 + ], + "spans": [ + { + "bbox": [ + 313, + 376, + 555, + 473 + ], + "type": "text", + "content": "correlation between NIR and RGB in the frequency domain, we apply a high-pass filter to the aforementioned three kinds of frequency domain images and convert the filtered results back to the spatial domain. We separately computed the Structural Similarity (SSIM) [40] between NIR/noisy RGB images and clean RGB images at different cutoff frequencies as a measure of frequency correlation " + }, + { + "bbox": [ + 313, + 376, + 555, + 473 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 313, + 376, + 555, + 473 + ], + "type": "text", + "content": ", as shown in Fig. 3a and Fig. 3b. The process can be described as" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 340, + 478, + 555, + 493 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 340, + 478, + 555, + 493 + ], + "spans": [ + { + "bbox": [ + 340, + 478, + 555, + 493 + ], + "type": "interline_equation", + "content": "S = \\operatorname {S S I M} \\left(\\mathcal {F} ^ {- 1} \\left(\\mathrm {H} \\left(F _ {\\text {t a r g e t}}\\right)\\right), \\mathcal {F} ^ {- 1} \\left(\\mathrm {H} \\left(F _ {\\text {g t}}\\right)\\right)\\right) \\tag {1}", + "image_path": "426c1d8aad79c813432774ccf93e0c5528f041b7f011d9c9d3811a200436e46e.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 497, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 497, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 497, + 556, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 497, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{F}^{-1}(\\cdot)" + }, + { + "bbox": [ + 313, + 497, + 556, + 714 + ], + "type": "text", + "content": " denote 2D-IDFT operations, " + }, + { + "bbox": [ + 313, + 497, + 556, + 714 + ], + "type": "inline_equation", + "content": "F_{target}" + }, + { + "bbox": [ + 313, + 497, + 556, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 497, + 556, + 714 + ], + "type": "inline_equation", + "content": "F_{gt}" + }, + { + "bbox": [ + 313, + 497, + 556, + 714 + ], + "type": "text", + "content": " represent NIR or noisy RGB and clean RGB after 2D-DFT transformation. " + }, + { + "bbox": [ + 313, + 497, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{H}(\\cdot)" + }, + { + "bbox": [ + 313, + 497, + 556, + 714 + ], + "type": "text", + "content": " represents the filtering of the input image with a high-pass filter. The analysis results indicate that, in the main energy frequency, the image similarity between the input noisy RGB and clean RGB images decreases progressively as the frequency increases, while the image similarity between the input NIR image and clean RGB image increases from low to high frequency. Fig. 3c also shows that for noisy RGB images, their high frequency information is mainly composed of noise and lacks detail information, while the low frequency information more faithfully reflects the color in the original image, and NIR images are just the opposite. These cross-field Frequency correlation priors indicate that it is simple and efficient to distinguish useful information from NIR and RGB images in the frequency domain, which inspires us to propose a comprehensive frequency domain learning framework." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12681" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 72, + 154, + 127 + ], + "blocks": [ + { + "bbox": [ + 59, + 72, + 154, + 127 + ], + "lines": [ + { + "bbox": [ + 59, + 72, + 154, + 127 + ], + "spans": [ + { + "bbox": [ + 59, + 72, + 154, + 127 + ], + "type": "image", + "image_path": "5866955e7c4a4770d9b89dadcfc5461df22c2aca17b6c21291642ac9b6cbe000.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 72, + 129, + 141, + 138 + ], + "lines": [ + { + "bbox": [ + 72, + 129, + 141, + 138 + ], + "spans": [ + { + "bbox": [ + 72, + 129, + 141, + 138 + ], + "type": "text", + "content": "Noisy RGB (Input)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 197, + 73, + 291, + 126 + ], + "blocks": [ + { + "bbox": [ + 197, + 73, + 291, + 126 + ], + "lines": [ + { + "bbox": [ + 197, + 73, + 291, + 126 + ], + "spans": [ + { + "bbox": [ + 197, + 73, + 291, + 126 + ], + "type": "image", + "image_path": "d920fb41101c44be5427913d7679d3f21b9d747a8dd7e9c2785e9acc8d0aeef4.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 58, + 139, + 293, + 322 + ], + "blocks": [ + { + "bbox": [ + 211, + 129, + 277, + 138 + ], + "lines": [ + { + "bbox": [ + 211, + 129, + 277, + 138 + ], + "spans": [ + { + "bbox": [ + 211, + 129, + 277, + 138 + ], + "type": "text", + "content": "Clean NIR (Input)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 58, + 139, + 293, + 322 + ], + "lines": [ + { + "bbox": [ + 58, + 139, + 293, + 322 + ], + "spans": [ + { + "bbox": [ + 58, + 139, + 293, + 322 + ], + "type": "image", + "image_path": "3f21648d2b5ed62c367a1019c5f30fd90f41e5ca6c6d539c52fb454fed0eb311.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 326, + 295, + 371 + ], + "lines": [ + { + "bbox": [ + 55, + 326, + 295, + 371 + ], + "spans": [ + { + "bbox": [ + 55, + 326, + 295, + 371 + ], + "type": "text", + "content": "Figure 4. The overall architecture of the proposed cross-field Frequency Correlation Exploiting Network (FCENet). In the first stage, pre-denoising is performed. In the second stage, NIR image is introduced for NIR-assisted image denoising." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 383, + 181, + 395 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 383, + 181, + 395 + ], + "spans": [ + { + "bbox": [ + 55, + 383, + 181, + 395 + ], + "type": "text", + "content": "3.2. Overview Framework" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 400, + 296, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 400, + 296, + 508 + ], + "spans": [ + { + "bbox": [ + 55, + 400, + 296, + 508 + ], + "type": "text", + "content": "The proposed frequency domain fusion framework, as depicted in Fig. 4, includes two stages for the gradual restoration of images. Both stages incorporate an encoder-decoder architecture based on the U-Net [30], featuring two downsampling layers and two upsampling layers. Skip connections are employed between the encoder and decoder features. A Supervised Attention Module (SAM) [47] is utilized to connect features from the first stage to the second stage for further processing." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 509, + 296, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 509, + 296, + 604 + ], + "spans": [ + { + "bbox": [ + 55, + 509, + 296, + 604 + ], + "type": "text", + "content": "In the first stage, the input noisy RGB image undergoes preliminary denoising to reduce the modality gap with the NIR image. In the second stage, we perform multi-scale fusion of the encoded features of the NIR and RGB images. Specifically, at each scale of the encoder, NIR encoded features and RGB pre-denoise features are fed into the frequency domain fusion modules based on FDSM and FEFM, as shown in Fig. 5." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 612, + 274, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 612, + 274, + 624 + ], + "spans": [ + { + "bbox": [ + 55, + 612, + 274, + 624 + ], + "type": "text", + "content": "3.3. Frequency Dynamic Selection Mechanism" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "content": "The frequency correlation prior indicates that extracting cross-field complementary information from the frequency domain is efficient. For feature extraction, traditional convolutional kernels share weights across the entire feature map, lacking flexibility to address the numerous inconsistencies between NIR and RGB images, which affects the efficiency of extraction. We introduce frequency domain" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 72, + 553, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 95 + ], + "type": "text", + "content": "dynamic convolution to adapt to specific inputs and effectively extract complementary features." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 95, + 553, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 95, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 313, + 95, + 553, + 205 + ], + "type": "text", + "content": "Specifically, given the NIR features " + }, + { + "bbox": [ + 313, + 95, + 553, + 205 + ], + "type": "inline_equation", + "content": "\\mathbf{N} \\in \\mathbb{R}^{C \\times H \\times W}" + }, + { + "bbox": [ + 313, + 95, + 553, + 205 + ], + "type": "text", + "content": " and pre-denoised features " + }, + { + "bbox": [ + 313, + 95, + 553, + 205 + ], + "type": "inline_equation", + "content": "\\mathbf{R} \\in \\mathbb{R}^{C \\times H \\times W}" + }, + { + "bbox": [ + 313, + 95, + 553, + 205 + ], + "type": "text", + "content": ", we concatenate them after layernorm and undergo convolution and the activate function to obtain a coarsely fused feature map. After that, we separate the fused feature map to obtain the aggregated features " + }, + { + "bbox": [ + 313, + 95, + 553, + 205 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_{\\mathbf{R}}" + }, + { + "bbox": [ + 313, + 95, + 553, + 205 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 95, + 553, + 205 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_{\\mathbf{N}}" + }, + { + "bbox": [ + 313, + 95, + 553, + 205 + ], + "type": "text", + "content": " for NIR and RGB individually. These aggregated features are used to generate dynamic filtering kernels to selectively extract useful features from the NIR and RGB input features." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 206, + 555, + 385 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 206, + 555, + 385 + ], + "spans": [ + { + "bbox": [ + 313, + 206, + 555, + 385 + ], + "type": "text", + "content": "For dynamic filters, some works [27, 52] that generate deep separable and spatially variant dynamic filtering kernels are usually computationally intensive and time-consuming, and require a large amount of data. We employ a method similar to [6, 34] by generating dynamic weights to dynamically aggregate multiple parallel convolutional kernels related to the input, and performing this process in the frequency domain. Specifically, the aggregated features " + }, + { + "bbox": [ + 313, + 206, + 555, + 385 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_{\\mathbf{R}}" + }, + { + "bbox": [ + 313, + 206, + 555, + 385 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 206, + 555, + 385 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_{\\mathbf{N}}" + }, + { + "bbox": [ + 313, + 206, + 555, + 385 + ], + "type": "text", + "content": " are used to generate the weights for " + }, + { + "bbox": [ + 313, + 206, + 555, + 385 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 206, + 555, + 385 + ], + "type": "text", + "content": " learnable frequency domain filters. For each feature channel, these filters are linearly combined to form a frequency domain global filter, which is used to select useful frequency domain information from the input NIR and RGB features. Finally, " + }, + { + "bbox": [ + 313, + 206, + 555, + 385 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{\\mathbf{R}}" + }, + { + "bbox": [ + 313, + 206, + 555, + 385 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 206, + 555, + 385 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{\\mathbf{N}}" + }, + { + "bbox": [ + 313, + 206, + 555, + 385 + ], + "type": "text", + "content": " denoting the filtered useful features of NIR and RGB can be represented as" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 344, + 397, + 553, + 410 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 344, + 397, + 553, + 410 + ], + "spans": [ + { + "bbox": [ + 344, + 397, + 553, + 410 + ], + "type": "interline_equation", + "content": "\\mathbf {D F} _ {\\mathbf {I}} = \\operatorname {S o f t m a x} (\\operatorname {M L P} (\\operatorname {P o o l i n g} (\\mathbf {A} _ {\\mathbf {I}}))) \\cdot \\mathbf {G} \\tag {2}", + "image_path": "eaa79fc9e72006d81675f5da0d73d233d9c2be779785f18ca067d37807ceb3b7.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 381, + 435, + 553, + 449 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 381, + 435, + 553, + 449 + ], + "spans": [ + { + "bbox": [ + 381, + 435, + 553, + 449 + ], + "type": "interline_equation", + "content": "\\mathbf {F} _ {\\mathbf {I}} = \\mathcal {F} ^ {- 1} (\\mathcal {F} (\\mathbf {I}) \\odot \\mathbf {D F} _ {\\mathbf {I}}) \\tag {3}", + "image_path": "bfffa05073b65ba0090647428c1415ca90a83ab5a4c380eb687eecb858b8312d.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 457, + 554, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 457, + 554, + 529 + ], + "spans": [ + { + "bbox": [ + 313, + 457, + 554, + 529 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 457, + 554, + 529 + ], + "type": "inline_equation", + "content": "\\mathbf{DF}_{\\mathbf{I}}" + }, + { + "bbox": [ + 313, + 457, + 554, + 529 + ], + "type": "text", + "content": " is frequency domain dynamic filter, and " + }, + { + "bbox": [ + 313, + 457, + 554, + 529 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 313, + 457, + 554, + 529 + ], + "type": "text", + "content": " represents " + }, + { + "bbox": [ + 313, + 457, + 554, + 529 + ], + "type": "inline_equation", + "content": "\\mathbf{R}" + }, + { + "bbox": [ + 313, + 457, + 554, + 529 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 313, + 457, + 554, + 529 + ], + "type": "inline_equation", + "content": "\\mathbf{N}" + }, + { + "bbox": [ + 313, + 457, + 554, + 529 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 313, + 457, + 554, + 529 + ], + "type": "inline_equation", + "content": "\\mathbf{G} \\in \\mathbb{R}^{M \\times k}" + }, + { + "bbox": [ + 313, + 457, + 554, + 529 + ], + "type": "text", + "content": " represents the combination of " + }, + { + "bbox": [ + 313, + 457, + 554, + 529 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 457, + 554, + 529 + ], + "type": "text", + "content": " learnable frequency domain filtering kernels, and " + }, + { + "bbox": [ + 313, + 457, + 554, + 529 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 313, + 457, + 554, + 529 + ], + "type": "text", + "content": " is the spatial dimension of the convolutional kernel. " + }, + { + "bbox": [ + 313, + 457, + 554, + 529 + ], + "type": "inline_equation", + "content": "\\mathcal{F}(\\cdot)" + }, + { + "bbox": [ + 313, + 457, + 554, + 529 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 457, + 554, + 529 + ], + "type": "inline_equation", + "content": "\\mathcal{F}^{-1}(\\cdot)" + }, + { + "bbox": [ + 313, + 457, + 554, + 529 + ], + "type": "text", + "content": " denote the 2D-DFT and 2D-IDFT operations, respectively." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 540, + 533, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 540, + 533, + 552 + ], + "spans": [ + { + "bbox": [ + 313, + 540, + 533, + 552 + ], + "type": "text", + "content": "3.4. Frequency Exhaustive Fusion Mechanism" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 557, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 557, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 557, + 555, + 713 + ], + "type": "text", + "content": "After the NIR and RGB feature maps pass through the FDSM, the useful features are retained. Next, to fuse the two inconsistent fields, we further integrate their information in the frequency domain and propose the FEFM, with this process driven by prior frequency correlations. Specifically, the proposed FEFM includes two key components: Common Feature Reinforcement (CFR) mechanism and Differential Feature Reinforcement (DFR) mechanism. CFR comprehensively explores the correlation between NIR and RGB features in the frequency domain to strengthen the modeling of common features, while DFR employs differential modeling to enhance key differential high frequency features." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12682" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 72, + 309, + 279 + ], + "blocks": [ + { + "bbox": [ + 60, + 72, + 309, + 279 + ], + "lines": [ + { + "bbox": [ + 60, + 72, + 309, + 279 + ], + "spans": [ + { + "bbox": [ + 60, + 72, + 309, + 279 + ], + "type": "image", + "image_path": "ed8870cfdcef13a312ceaab3eafb76c2359f11206d819cd593cbb66bfd5bb935.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 285, + 555, + 309 + ], + "lines": [ + { + "bbox": [ + 55, + 285, + 555, + 309 + ], + "spans": [ + { + "bbox": [ + 55, + 285, + 555, + 309 + ], + "type": "text", + "content": "Figure 5. The architecture of the proposed Frequency Dynamic Selection Mechanism (FDSM) and Frequency Exhaustive Fusion Mechanism (FEFM) for thorough frequency exploitation." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 312, + 72, + 552, + 279 + ], + "blocks": [ + { + "bbox": [ + 312, + 72, + 552, + 279 + ], + "lines": [ + { + "bbox": [ + 312, + 72, + 552, + 279 + ], + "spans": [ + { + "bbox": [ + 312, + 72, + 552, + 279 + ], + "type": "image", + "image_path": "a3813b66fc45b9c7286ce0915b5ead4f9be1099372b027b1b80c5a523c6ca42a.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 317, + 296, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 317, + 296, + 594 + ], + "spans": [ + { + "bbox": [ + 56, + 317, + 296, + 594 + ], + "type": "text", + "content": "Common Feature Reinforcement Mechanism. The convolution theorem establishes that the correlation or convolution of two signals in the spatial domain is equal to their element-wise product in the frequency domain. Leveraging this property, previous work [21] integrate the frequency domain into the self-attention mechanism, simplifying matrix multiplication into lightweight element-wise dot product operations. Although the dot product is efficient, it only calculates the point-wise correlation coefficients in the frequency domain, missing the long-range correlations in the frequency domain. To more comprehensively model the common features of NIR and RGB, we apply a pointwise convolution and a depthwise convolution to the input NIR and RGB feature maps " + }, + { + "bbox": [ + 56, + 317, + 296, + 594 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{\\mathbf{R}}" + }, + { + "bbox": [ + 56, + 317, + 296, + 594 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 56, + 317, + 296, + 594 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{\\mathbf{N}}" + }, + { + "bbox": [ + 56, + 317, + 296, + 594 + ], + "type": "text", + "content": ", to obtain encoded features " + }, + { + "bbox": [ + 56, + 317, + 296, + 594 + ], + "type": "inline_equation", + "content": "\\mathbf{Q} = W_{d}^{Q}W_{p}^{Q}\\mathbf{F}_{\\mathbf{R}}" + }, + { + "bbox": [ + 56, + 317, + 296, + 594 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 56, + 317, + 296, + 594 + ], + "type": "inline_equation", + "content": "\\mathbf{K} = W_{d}^{K}W_{p}^{K}\\mathbf{F}_{\\mathbf{N}}" + }, + { + "bbox": [ + 56, + 317, + 296, + 594 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 56, + 317, + 296, + 594 + ], + "type": "inline_equation", + "content": "\\mathbf{V} = W_{d}^{V}W_{p}^{V}\\mathbf{F}_{\\mathbf{N}}" + }, + { + "bbox": [ + 56, + 317, + 296, + 594 + ], + "type": "text", + "content": ". To facilitate the calculation of frequency domain correlations, we apply 2D-DFT to the input features to produce frequency cubes " + }, + { + "bbox": [ + 56, + 317, + 296, + 594 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{\\mathbf{Q}}" + }, + { + "bbox": [ + 56, + 317, + 296, + 594 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 56, + 317, + 296, + 594 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{\\mathbf{K}}" + }, + { + "bbox": [ + 56, + 317, + 296, + 594 + ], + "type": "text", + "content": ". We simultaneously model the long-range correlation and point-wise correlation of NIR and RGB features in the frequency domain, resulting in a correlation map that integrates local similarity and long-range channel correlation within the frequency domain. The output of CFR can be represented as" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 80, + 597, + 295, + 613 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 597, + 295, + 613 + ], + "spans": [ + { + "bbox": [ + 80, + 597, + 295, + 613 + ], + "type": "interline_equation", + "content": "\\mathbf {F} _ {\\mathbf {C F R}} = \\left(\\mathbf {F} _ {\\mathbf {Q}} \\odot \\mathbf {F} _ {\\mathbf {K}}\\right) \\cdot \\mathrm {S o f t m a x} (\\mathbf {F} _ {\\mathbf {Q}} \\cdot \\mathbf {F} _ {\\mathbf {K}} ^ {\\top} / \\alpha) \\quad (4)", + "image_path": "c25fc9c1cfa042706892dee0901785a08efdd3b95e4ba621135c587eb236ac95.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 618, + 222, + 630 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 222, + 630 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 222, + 630 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 618, + 222, + 630 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 618, + 222, + 630 + ], + "type": "text", + "content": " is a learnable scaling parameter." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "text", + "content": "Differential Feature Reinforcement Mechanism. Traditional cross-attention [15, 25, 39] structures are typically used to capture common features but do not effectively utilize differential information. Although this approach is correct in general image fusion frameworks, in this paper, the features of the two fields are considered useful after passing through the frequency domain selection mechanism. By" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 317, + 555, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 317, + 555, + 426 + ], + "spans": [ + { + "bbox": [ + 313, + 317, + 555, + 426 + ], + "type": "text", + "content": "observing the feature maps (Fig. 9), we find that the high frequency texture information in the NIR features is clear compared to the RGB feature maps. The common features obtained by weighted fusion through cross-attention remain somewhat blurry, which greatly affects the denoising effect. These differential high frequency features on NIR are crucial for image restoration. We adopt a differential cross-attention to complement the missing differential information, and the output of CFR can be represented as" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 356, + 438, + 555, + 453 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 356, + 438, + 555, + 453 + ], + "spans": [ + { + "bbox": [ + 356, + 438, + 555, + 453 + ], + "type": "interline_equation", + "content": "\\mathbf {F} _ {\\mathbf {D F R}} = \\left(\\mathbf {V} - \\lambda \\mathbf {V} \\odot \\mathcal {F} ^ {- 1} \\left(\\mathbf {F} _ {\\mathbf {C F R}}\\right)\\right) \\tag {5}", + "image_path": "af33146f5a270bad3a17d448fb7ea19981e528726b7e9f6f55a01bf3955a6488.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 466, + 555, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 466, + 555, + 561 + ], + "spans": [ + { + "bbox": [ + 313, + 466, + 555, + 561 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 466, + 555, + 561 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 466, + 555, + 561 + ], + "type": "text", + "content": " is a learnable scalar that dynamically controls the weights of two feature maps. Although DFR effectively models the differential high frequency information, the fused features lack some common features and background information. To further refine the fused features, we perform a dot product between the encoded features Q and the output of CFR, and add the result to the output of DFR as the final fused result." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 573, + 404, + 585 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 573, + 404, + 585 + ], + "spans": [ + { + "bbox": [ + 313, + 573, + 404, + 585 + ], + "type": "text", + "content": "3.5. Loss Function" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 592, + 554, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 592, + 554, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 592, + 554, + 628 + ], + "type": "text", + "content": "We adopt the Charbonnier loss [5], and adding the frequency domain loss [9]. The overall loss function is represented as" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 325, + 637, + 553, + 666 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 637, + 553, + 666 + ], + "spans": [ + { + "bbox": [ + 325, + 637, + 553, + 666 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} = \\mathcal {L} _ {\\text {c h a r b o n n i e r}} \\left(\\mathbf {X} _ {1}, \\mathbf {T}\\right) + \\mathcal {L} _ {\\text {c h a r b o n n i e r}} \\left(\\mathbf {X} _ {2}, \\mathbf {T}\\right) \\tag {6} \\\\ + \\alpha \\cdot \\mathcal {L} _ {\\text {c h a r b o n n i e r}} (\\mathcal {F} (\\mathbf {X} _ {2}), \\mathcal {F} (\\mathbf {T})) \\\\ \\end{array}", + "image_path": "0a90e69d02b479dae8dfd0e8002f44a1370acdabd9476b9b8844d283f7522167.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_1" + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_2" + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "content": " are the outputs of pre-denoising subnetwork and whole network, respectively. " + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{T}" + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "content": " is ground truth and " + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "content": " is set as 0.1 [9]." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12683" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 61, + 553, + 277 + ], + "blocks": [ + { + "bbox": [ + 58, + 61, + 553, + 277 + ], + "lines": [ + { + "bbox": [ + 58, + 61, + 553, + 277 + ], + "spans": [ + { + "bbox": [ + 58, + 61, + 553, + 277 + ], + "type": "image", + "image_path": "d505563a74781eb99f16582db7a79b4a840e189841ce46afbc36714df2e51c76.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 279, + 555, + 301 + ], + "lines": [ + { + "bbox": [ + 55, + 279, + 555, + 301 + ], + "spans": [ + { + "bbox": [ + 55, + 279, + 555, + 301 + ], + "type": "text", + "content": "Figure 6. The qualitative comparison among our FCENet and the state-of-the-art methods on the noisy RGB-NIR pairs from DVD[18] with different noise levels. Our method achieves better results in detail recovery, artifact removal, and color restoration." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 56, + 335, + 555, + 518 + ], + "blocks": [ + { + "bbox": [ + 55, + 310, + 555, + 332 + ], + "lines": [ + { + "bbox": [ + 55, + 310, + 555, + 332 + ], + "spans": [ + { + "bbox": [ + 55, + 310, + 555, + 332 + ], + "type": "text", + "content": "Table 1. Comparison of different methods on the DVD test set with noise levels " + }, + { + "bbox": [ + 55, + 310, + 555, + 332 + ], + "type": "inline_equation", + "content": "\\sigma = 2,4,6" + }, + { + "bbox": [ + 55, + 310, + 555, + 332 + ], + "type": "text", + "content": " . The best and second-best results are highlighted in boldface and underlined, respectively." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 335, + 555, + 518 + ], + "lines": [ + { + "bbox": [ + 56, + 335, + 555, + 518 + ], + "spans": [ + { + "bbox": [ + 56, + 335, + 555, + 518 + ], + "type": "table", + "html": "
Methodsσ = 2σ = 4σ = 6Complexity
PSNR↑SSIM↑PSNR↑SSIM↑PSNR↑SSIM↑FLOPs(G)Params(M)
Uformer [41]31.340.94929.120.92727.580.9089.6020.63
Restormer [48]31.300.94929.200.92827.760.91036.0626.13
MPRNet [47]31.780.95129.370.92927.840.910133.5115.74
Xformer [49]31.410.95029.300.92827.430.90839.3425.22
CUNet [11]28.760.92326.810.89825.640.8756.720.44
SANet [32]30.130.93827.840.91726.410.90125.015.38
MNNet [44]30.140.94428.260.92326.750.90610.970.76
DVN [18]31.500.95529.620.94028.260.92748.306.97
NIR-Restormer [43]31.810.95729.810.94228.310.92845.3130.26
FCENet-L (Ours)31.870.96029.910.94628.360.9345.711.16
FCENet (Ours)32.430.96330.370.95028.780.93919.524.44
", + "image_path": "ff78c319ea4e0658d0fc0a052c62c1119dc0be6dbb00c03bd4a73cec721a40bc.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 526, + 137, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 526, + 137, + 540 + ], + "spans": [ + { + "bbox": [ + 55, + 526, + 137, + 540 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 546, + 295, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 546, + 295, + 583 + ], + "spans": [ + { + "bbox": [ + 55, + 546, + 295, + 583 + ], + "type": "text", + "content": "In this section, we evaluate the proposed method on public datasets and compare it with other NIR-assisted image denoising methods." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 588, + 182, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 588, + 182, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 588, + 182, + 601 + ], + "type": "text", + "content": "4.1. Experimental Settings" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "text", + "content": "Datasets. We evaluate the NIR-assisted image denoising method on the DVD [18] and IVRG [4] datasets. The DVD dataset consists of 307 high-resolution NIR-RGB image pairs, which are converted into 5000 image pairs of size " + }, + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "inline_equation", + "content": "3 \\times 256 \\times 256" + }, + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "text", + "content": " for training and 1000 image pairs of size " + }, + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "inline_equation", + "content": "3 \\times 256 \\times 256" + }, + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "text", + "content": " for testing. The IVRG dataset contains NIR and RGB image pairs from 9 different scenes. For each scene, as in [44], experiments randomly select 30/8/15 pairs for training/validation/testing." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 528, + 555, + 708 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 528, + 555, + 708 + ], + "spans": [ + { + "bbox": [ + 313, + 528, + 555, + 708 + ], + "type": "text", + "content": "Parameter settings. The proposed network is implemented with Pytorch and trained using the Adam optimizer [20] " + }, + { + "bbox": [ + 313, + 528, + 555, + 708 + ], + "type": "inline_equation", + "content": "(\\beta_{1} = 0.9, \\beta_{2} = 0.999)" + }, + { + "bbox": [ + 313, + 528, + 555, + 708 + ], + "type": "text", + "content": " on patches of size " + }, + { + "bbox": [ + 313, + 528, + 555, + 708 + ], + "type": "inline_equation", + "content": "128 \\times 128" + }, + { + "bbox": [ + 313, + 528, + 555, + 708 + ], + "type": "text", + "content": " with a batch size of 16. For the DVD dataset, we simulate low-light conditions by randomly reducing the mean of the original images as in [18] and add Gaussian-Poisson mixed noise with a range of 1 to 16. The model is trained for 80 epochs. The initial learning rate is " + }, + { + "bbox": [ + 313, + 528, + 555, + 708 + ], + "type": "inline_equation", + "content": "2 \\times 10^{-4}" + }, + { + "bbox": [ + 313, + 528, + 555, + 708 + ], + "type": "text", + "content": ", and a cosine annealing strategy is used to gradually reduce the learning rate to " + }, + { + "bbox": [ + 313, + 528, + 555, + 708 + ], + "type": "inline_equation", + "content": "1 \\times 10^{-6}" + }, + { + "bbox": [ + 313, + 528, + 555, + 708 + ], + "type": "text", + "content": ". For the IVRG dataset, following the settings in [44], it is cropped into 16,374 training patches, and then additive Gaussian noise with random noise levels is used to simulate the target images, considering three noise levels in testing: " + }, + { + "bbox": [ + 313, + 528, + 555, + 708 + ], + "type": "inline_equation", + "content": "\\sigma = 25" + }, + { + "bbox": [ + 313, + 528, + 555, + 708 + ], + "type": "text", + "content": ", 50, and 75. The model is trained for 100 epochs with an initial learning rate of " + }, + { + "bbox": [ + 313, + 528, + 555, + 708 + ], + "type": "inline_equation", + "content": "5 \\times 10^{-4}" + }, + { + "bbox": [ + 313, + 528, + 555, + 708 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "12684" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 74, + 121, + 135 + ], + "blocks": [ + { + "bbox": [ + 59, + 74, + 121, + 135 + ], + "lines": [ + { + "bbox": [ + 59, + 74, + 121, + 135 + ], + "spans": [ + { + "bbox": [ + 59, + 74, + 121, + 135 + ], + "type": "image", + "image_path": "40efdddf4da078317dab4769d3ee5e455ac254c04ffdfa881969bf3f48163ffd.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 63, + 137, + 119, + 148 + ], + "lines": [ + { + "bbox": [ + 63, + 137, + 119, + 148 + ], + "spans": [ + { + "bbox": [ + 63, + 137, + 119, + 148 + ], + "type": "text", + "content": "Real noisy RGB" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 121, + 74, + 182, + 136 + ], + "blocks": [ + { + "bbox": [ + 121, + 74, + 182, + 136 + ], + "lines": [ + { + "bbox": [ + 121, + 74, + 182, + 136 + ], + "spans": [ + { + "bbox": [ + 121, + 74, + 182, + 136 + ], + "type": "image", + "image_path": "f6011350f2ffe760c900d560fc9e4ebb58e0e6b4b103774a34699cac3b691679.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 143, + 139, + 158, + 147 + ], + "lines": [ + { + "bbox": [ + 143, + 139, + 158, + 147 + ], + "spans": [ + { + "bbox": [ + 143, + 139, + 158, + 147 + ], + "type": "text", + "content": "NIR" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 182, + 74, + 244, + 136 + ], + "blocks": [ + { + "bbox": [ + 182, + 74, + 244, + 136 + ], + "lines": [ + { + "bbox": [ + 182, + 74, + 244, + 136 + ], + "spans": [ + { + "bbox": [ + 182, + 74, + 244, + 136 + ], + "type": "image", + "image_path": "ef188faa8bc5f6c4996a255ad18e4769a0c232fc5cab0b083b406464413bacbf.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 198, + 139, + 229, + 148 + ], + "lines": [ + { + "bbox": [ + 198, + 139, + 229, + 148 + ], + "spans": [ + { + "bbox": [ + 198, + 139, + 229, + 148 + ], + "type": "text", + "content": "Xformer" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 245, + 74, + 305, + 136 + ], + "blocks": [ + { + "bbox": [ + 245, + 74, + 305, + 136 + ], + "lines": [ + { + "bbox": [ + 245, + 74, + 305, + 136 + ], + "spans": [ + { + "bbox": [ + 245, + 74, + 305, + 136 + ], + "type": "image", + "image_path": "bcf826294ca471d329803b0a14ae106dcc66b786b76b4369ad0d492855228561.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 261, + 139, + 285, + 148 + ], + "lines": [ + { + "bbox": [ + 261, + 139, + 285, + 148 + ], + "spans": [ + { + "bbox": [ + 261, + 139, + 285, + 148 + ], + "type": "text", + "content": "SANet" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 306, + 74, + 367, + 136 + ], + "blocks": [ + { + "bbox": [ + 306, + 74, + 367, + 136 + ], + "lines": [ + { + "bbox": [ + 306, + 74, + 367, + 136 + ], + "spans": [ + { + "bbox": [ + 306, + 74, + 367, + 136 + ], + "type": "image", + "image_path": "97b3fd47acfa9f240401dc6ce6cb0b38ea5313e1047e3244dca4a6c6472e687d.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 323, + 139, + 350, + 148 + ], + "lines": [ + { + "bbox": [ + 323, + 139, + 350, + 148 + ], + "spans": [ + { + "bbox": [ + 323, + 139, + 350, + 148 + ], + "type": "text", + "content": "MNNet" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 369, + 74, + 429, + 136 + ], + "blocks": [ + { + "bbox": [ + 369, + 74, + 429, + 136 + ], + "lines": [ + { + "bbox": [ + 369, + 74, + 429, + 136 + ], + "spans": [ + { + "bbox": [ + 369, + 74, + 429, + 136 + ], + "type": "image", + "image_path": "456535714bfa1a68771516b2fa7ca084b3afe05ce4180a22b900ba1591981231.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 390, + 139, + 409, + 148 + ], + "lines": [ + { + "bbox": [ + 390, + 139, + 409, + 148 + ], + "spans": [ + { + "bbox": [ + 390, + 139, + 409, + 148 + ], + "type": "text", + "content": "DVN" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 429, + 74, + 491, + 136 + ], + "blocks": [ + { + "bbox": [ + 429, + 74, + 491, + 136 + ], + "lines": [ + { + "bbox": [ + 429, + 74, + 491, + 136 + ], + "spans": [ + { + "bbox": [ + 429, + 74, + 491, + 136 + ], + "type": "image", + "image_path": "6e8981fd2e5ac2e36e1c2c94d0646e6912b6166772e72e329a5db301ffd41821.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 434, + 139, + 487, + 148 + ], + "lines": [ + { + "bbox": [ + 434, + 139, + 487, + 148 + ], + "spans": [ + { + "bbox": [ + 434, + 139, + 487, + 148 + ], + "type": "text", + "content": "NIR-Restormer" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 492, + 74, + 553, + 136 + ], + "blocks": [ + { + "bbox": [ + 492, + 74, + 553, + 136 + ], + "lines": [ + { + "bbox": [ + 492, + 74, + 553, + 136 + ], + "spans": [ + { + "bbox": [ + 492, + 74, + 553, + 136 + ], + "type": "image", + "image_path": "2f7e27f46a8dca2591f5672c4b3cdb886d653f922f4f36685ec3d42ae93eabdd.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 512, + 139, + 531, + 148 + ], + "lines": [ + { + "bbox": [ + 512, + 139, + 531, + 148 + ], + "spans": [ + { + "bbox": [ + 512, + 139, + 531, + 148 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "table", + "bbox": [ + 56, + 217, + 296, + 376 + ], + "blocks": [ + { + "bbox": [ + 55, + 175, + 295, + 209 + ], + "lines": [ + { + "bbox": [ + 55, + 175, + 295, + 209 + ], + "spans": [ + { + "bbox": [ + 55, + 175, + 295, + 209 + ], + "type": "text", + "content": "Table 2. Comparison of different methods on the IVRG test set with noise levels " + }, + { + "bbox": [ + 55, + 175, + 295, + 209 + ], + "type": "inline_equation", + "content": "\\sigma = 25, 50, 75" + }, + { + "bbox": [ + 55, + 175, + 295, + 209 + ], + "type": "text", + "content": ". The best and second-best results are highlighted in boldface and underlined, respectively." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 217, + 296, + 376 + ], + "lines": [ + { + "bbox": [ + 56, + 217, + 296, + 376 + ], + "spans": [ + { + "bbox": [ + 56, + 217, + 296, + 376 + ], + "type": "table", + "html": "
Methodsσ = 25σ = 50σ = 75
PSNR↑SSIM↑PSNR↑SSIM↑PSNR↑SSIM↑
Restormer [48]31.720.87627.670.78824.500.722
Uformer [41]31.970.87827.830.79324.580.726
MPRNet [47]31.890.87827.800.79424.570.729
Xformer [49]31.980.87927.860.79424.610.731
CUNet [11]31.370.87627.780.81224.640.764
SANet [32]31.540.88528.130.82524.810.785
MNNet [44]32.590.90428.660.85125.250.808
DVN [18]32.760.90628.730.85225.280.809
NIR-Restormer [43]32.860.90628.770.85325.300.811
FCENet-L (Ours)32.850.90928.800.85725.330.816
FCENet (Ours)33.260.91529.080.86625.480.826
", + "image_path": "7d9a2944bfe35f59276accc93230167e45dec510f217568b35a868434d9a7bf4.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "table_body" + } + ], + "index": 18 + }, + { + "bbox": [ + 55, + 384, + 179, + 396 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 384, + 179, + 396 + ], + "spans": [ + { + "bbox": [ + 55, + 384, + 179, + 396 + ], + "type": "text", + "content": "4.2. Experimental Results" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 54, + 403, + 295, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 403, + 295, + 545 + ], + "spans": [ + { + "bbox": [ + 54, + 403, + 295, + 545 + ], + "type": "text", + "content": "In our experimental evaluation of our method, we compare our results with nine models, including five state-of-the-art NIR-assisted RGB denoising methods: CUNet [11], SANet [32], MNNet [44], DVN [18], and NIR-Restormer [43], and four state-of-the-art single-image denoising methods: MPRNet [47], Restormer [48], Uformer [41], and Xformer [49]. To further demonstrate the superiority of our method, we additionally provide the denoising results of a light version (-L) by changing the initial feature map channel count from 64 to 36, which has only " + }, + { + "bbox": [ + 54, + 403, + 295, + 545 + ], + "type": "inline_equation", + "content": "1.16\\mathrm{M}" + }, + { + "bbox": [ + 54, + 403, + 295, + 545 + ], + "type": "text", + "content": " parameters. All compared methods are trained on the same training set as ours." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 55, + 546, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 546, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 546, + 295, + 713 + ], + "type": "text", + "content": "Evaluations on the DVN dataset. We first evaluate our method on the DVD dataset, with quantitative results shown in Table 1. It can be observed that the proposed FCENet, and even FCENet-L, outperform all other single-image denoising methods and NIR-assisted denoising methods in terms of PSNR and SSIM with moderate computational and memory costs. The qualitative comparison in Fig. 6 clearly illustrates that our method has advantages over other state-of-the-art methods in noise removal, color preservation, and detail recovery. For single-image denoising methods like Restormer, although they have achieved advanced denoising effects in low-noise scenarios, they still lack effective means to address detail texture loss in high-noise situations. For guided denoising methods, MNNet is unable to effec" + } + ] + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 317, + 176, + 550, + 255 + ], + "blocks": [ + { + "bbox": [ + 97, + 153, + 511, + 164 + ], + "lines": [ + { + "bbox": [ + 97, + 153, + 511, + 164 + ], + "spans": [ + { + "bbox": [ + 97, + 153, + 511, + 164 + ], + "type": "text", + "content": "Figure 7. Visual comparisons on real-world RGB/NIR image pairs. Our method demonstrates better visual results." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 317, + 176, + 550, + 255 + ], + "lines": [ + { + "bbox": [ + 317, + 176, + 550, + 255 + ], + "spans": [ + { + "bbox": [ + 317, + 176, + 550, + 255 + ], + "type": "image", + "image_path": "d17f45f53a63c92a062f1cf5ffcf65f1835c6e2eeb48f9a28fe466e3cfe22f51.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 264, + 555, + 308 + ], + "lines": [ + { + "bbox": [ + 313, + 264, + 555, + 308 + ], + "spans": [ + { + "bbox": [ + 313, + 264, + 555, + 308 + ], + "type": "text", + "content": "Figure 8. Frequency domain dynamic filters on NIR and RGB side at different depths. The first row represents the frequency domain filters of RGB side, and the second row represents the filters of NIR side." + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 319, + 555, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 319, + 555, + 462 + ], + "spans": [ + { + "bbox": [ + 313, + 319, + 555, + 462 + ], + "type": "text", + "content": "tively handle the inconsistencies between NIR and RGB images. DVN strengthens consistent high frequency textures but loses critical differential high frequency information in NIR and sacrifices the authenticity of low-frequency color. NIR-Restormer has good denoising effects but lacks effective utilization of inconsistent information, leading to over-smoothing and color distortion. Our method recovers detailed texture information while more faithfully preserving low-frequency color information in RGB images. This proves that the proposed method can efficiently utilize the complementary information of NIR and RGB images in the fusion through a series of frequency domain learning." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 463, + 554, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 463, + 554, + 581 + ], + "spans": [ + { + "bbox": [ + 313, + 463, + 554, + 581 + ], + "type": "text", + "content": "**Evaluations on the IVRG dataset.** We evaluate our method on the IVRG dataset. Following the previous method, we consider three noise levels in the test, including " + }, + { + "bbox": [ + 313, + 463, + 554, + 581 + ], + "type": "inline_equation", + "content": "\\sigma = 25, 50" + }, + { + "bbox": [ + 313, + 463, + 554, + 581 + ], + "type": "text", + "content": ", and 75. Quantitative results are shown in Table 2. It can be observed that compared to the state-of-the-art single-image denoising algorithms, our method significantly improves the performance of image denoising, demonstrating the effectiveness of NIR images. Compared to existing NIR-assisted denoising methods, our method also has significant advantages." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 582, + 555, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 555, + 712 + ], + "type": "text", + "content": "Evaluations on real-world experiment. To further evaluate the performance of our model in real noisy environments, we conduct qualitative experiments on NIR-RGB data pairs in real low-light conditions. As shown in Fig. 7, the introduction of NIR images brings more detailed textures but also a large number of inconsistencies. SANet fails to handle inconsistencies, resulting in a significant amount of artifacts, while MNNet, DVN and NIR-Restormer are unable to thoroughly denoise. In contrast, our method can eliminate inconsistencies while introducing more detailed information." + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "12685" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 60, + 85, + 290, + 158 + ], + "blocks": [ + { + "bbox": [ + 61, + 71, + 288, + 82 + ], + "lines": [ + { + "bbox": [ + 61, + 71, + 288, + 82 + ], + "spans": [ + { + "bbox": [ + 61, + 71, + 288, + 82 + ], + "type": "text", + "content": "Table 3. Ablation study of various components of our method." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 60, + 85, + 290, + 158 + ], + "lines": [ + { + "bbox": [ + 60, + 85, + 290, + 158 + ], + "spans": [ + { + "bbox": [ + 60, + 85, + 290, + 158 + ], + "type": "table", + "html": "
BaselineFDSMFEFMPSNR↑SSIM↑
(a)✓29.130.936
(b)✓✓29.640.944
(c)✓✓29.960.947
(d)✓✓✓30.370.950
", + "image_path": "415b7f9d523bce9b955e5a264060101cacda6dc3c9949db7ee0cde29a37e8073.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 67, + 180, + 282, + 240 + ], + "blocks": [ + { + "bbox": [ + 94, + 167, + 256, + 177 + ], + "lines": [ + { + "bbox": [ + 94, + 167, + 256, + 177 + ], + "spans": [ + { + "bbox": [ + 94, + 167, + 256, + 177 + ], + "type": "text", + "content": "Table 4. Ablation on the branches of FDSM." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 67, + 180, + 282, + 240 + ], + "lines": [ + { + "bbox": [ + 67, + 180, + 282, + 240 + ], + "spans": [ + { + "bbox": [ + 67, + 180, + 282, + 240 + ], + "type": "table", + "html": "
MethodsPSNR↑SSIM↑
w/o FDSM29.960.947
Spatial Dynamic Selection30.110.948
Frequency Dynamic Selection30.370.950
", + "image_path": "2f19b5329320bf2696df6bb31fa99b6d96670a5f53f660374a4c19af601895d8.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 94, + 262, + 256, + 335 + ], + "blocks": [ + { + "bbox": [ + 94, + 248, + 256, + 258 + ], + "lines": [ + { + "bbox": [ + 94, + 248, + 256, + 258 + ], + "spans": [ + { + "bbox": [ + 94, + 248, + 256, + 258 + ], + "type": "text", + "content": "Table 5. Ablation on the branches of FEFM." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 94, + 262, + 256, + 335 + ], + "lines": [ + { + "bbox": [ + 94, + 262, + 256, + 335 + ], + "spans": [ + { + "bbox": [ + 94, + 262, + 256, + 335 + ], + "type": "table", + "html": "
MethodsPSNR↑SSIM↑
Sum29.640.944
Cross-attention29.850.946
DFR30.150.948
DFR + CFR30.370.950
", + "image_path": "40325c736a1b6bb8a477ad6ea7395112b997e4e4075423009a5f9ece8a125358.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 349, + 149, + 361 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 349, + 149, + 361 + ], + "spans": [ + { + "bbox": [ + 55, + 349, + 149, + 361 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 366, + 295, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 366, + 295, + 473 + ], + "spans": [ + { + "bbox": [ + 55, + 366, + 295, + 473 + ], + "type": "text", + "content": "We conduct ablation studies with FCENet on the DVD dataset under " + }, + { + "bbox": [ + 55, + 366, + 295, + 473 + ], + "type": "inline_equation", + "content": "\\sigma = 4" + }, + { + "bbox": [ + 55, + 366, + 295, + 473 + ], + "type": "text", + "content": ". We compare two mechanisms proposed in this paper with a two-stage U-Net baseline, with the results shown in Table 3. In the U-Net baseline, crossfield images are fused through addition. By deploying FDSM and FEFM, the model achieves gains of " + }, + { + "bbox": [ + 55, + 366, + 295, + 473 + ], + "type": "inline_equation", + "content": "0.51\\mathrm{dB}" + }, + { + "bbox": [ + 55, + 366, + 295, + 473 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 366, + 295, + 473 + ], + "type": "inline_equation", + "content": "0.83\\mathrm{dB}" + }, + { + "bbox": [ + 55, + 366, + 295, + 473 + ], + "type": "text", + "content": ", respectively. Finally, combining all contributions significantly improves the denoising level compared to the baseline." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 474, + 295, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 474, + 295, + 653 + ], + "spans": [ + { + "bbox": [ + 55, + 474, + 295, + 653 + ], + "type": "text", + "content": "Effect of FDSM. Using FDSM allows for better extraction of complementary information from NIR and RGB while handling inconsistent regions. In Fig. 8, we show the frequency domain dynamic filter kernels in FDSM. It can be observed that the filter kernels processing NIR feature maps regularly enhance the learning of mid-to-high frequency information, while those processing RGB feature maps mainly learn low frequency information from the input images. This is consistent with the conclusions drawn from our frequency domain correlation priors. Additionally, we conduct experiments using spatial dynamic convolution in this module. The results in Table 4 ultimately prove the effectiveness of extracting relevant features from the frequency domain in NIR and RGB images and the stability of dynamic convolution in extracting inconsistent features." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 653, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 295, + 713 + ], + "type": "text", + "content": "Effect of FEFM. Traditional cross-attention fusion modules only strengthen common features but neglect differential features. FEFM strengthens the learning of common features and enhances the learning of some high frequency differential information in NIR feature maps. Fig. 9 shows" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 317, + 72, + 553, + 182 + ], + "blocks": [ + { + "bbox": [ + 317, + 72, + 553, + 182 + ], + "lines": [ + { + "bbox": [ + 317, + 72, + 553, + 182 + ], + "spans": [ + { + "bbox": [ + 317, + 72, + 553, + 182 + ], + "type": "image", + "image_path": "21b075348c642adcbe9eb117e4a3c2668d4b456f4430f0feee086f1a0f69c0a9.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 187, + 555, + 232 + ], + "lines": [ + { + "bbox": [ + 313, + 187, + 555, + 232 + ], + "spans": [ + { + "bbox": [ + 313, + 187, + 555, + 232 + ], + "type": "text", + "content": "Figure 9. Feature maps obtained from two fields. The first column represents the input feature maps of NIR and RGB, the second column represents the output feature of cross-attention, and the third column is the output feature of FEFM." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 245, + 555, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 245, + 555, + 389 + ], + "spans": [ + { + "bbox": [ + 313, + 245, + 555, + 389 + ], + "type": "text", + "content": "the input NIR and RGB feature maps after frequency domain selection, the fused feature maps after cross-attention fusion, and the fused feature maps after FEFM. Compared to the NIR feature maps, the input RGB feature maps still lack high frequency features. It can be seen that the proposed FEFM strengthens common features while supplementing some differential texture information from NIR. We compare two mechanisms in FEFM, CFR and DFR, with traditional cross-attention fusion modules and direct feature summation. The results in Table 5 prove effectiveness of our method in strengthening common features in frequency domain and modeling differential features." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 403, + 388, + 415 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 403, + 388, + 415 + ], + "spans": [ + { + "bbox": [ + 313, + 403, + 388, + 415 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 426, + 555, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 555, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 555, + 628 + ], + "type": "text", + "content": "In this paper, a cross-field frequency correlation exploiting network for NIR-assisted image denoising is proposed. Specifically, we first conduct a series of analyses on the characteristics of NIR and RGB images in frequency domain and obtain frequency correlation prior. Based on the prior, the frequency dynamic selection mechanism dynamically extracts complementary information from cross-field images in the frequency domain, and the frequency exhaustive fusion mechanism is proposed to enhance common features and differential high frequency features during fusion process. Extensive experiments demonstrate that the proposed method is significantly better than other algorithms. It is worth mentioning that existing methods may fail on significantly misaligned cross-field images. In future work, we plan to explore advanced methods such as feature point matching or incorporating deformable convolutional kernels to handle significantly misaligned cross-field images." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 644, + 411, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 644, + 411, + 658 + ], + "spans": [ + { + "bbox": [ + 313, + 644, + 411, + 658 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "type": "text", + "content": "This work is supported by the National Natural Science Foundation of China under Grants 62322204, 62131003, and the Key Laboratory of Target Cognition and Application Technology." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "12686" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "text", + "content": "[1] Abdelrahman Abdelhamed, Mahmoud Affifi, Radu Timofte, and Michael S Brown. Ntire 2020 challenge on real image denoising: Dataset, methods and results. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 496-497, 2020. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 148, + 296, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 148, + 296, + 190 + ], + "spans": [ + { + "bbox": [ + 61, + 148, + 296, + 190 + ], + "type": "text", + "content": "[2] Saeed Anwar and Nick Barnes. Real image denoising with feature attention. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3155-3164, 2019. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 194, + 295, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 194, + 295, + 225 + ], + "spans": [ + { + "bbox": [ + 62, + 194, + 295, + 225 + ], + "type": "text", + "content": "[3] Joshua Batson and Loic Royer. Noise2self: Blind denoising by self-supervision. In International Conference on Machine Learning, pages 524-533, 2019. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 228, + 295, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 228, + 295, + 270 + ], + "spans": [ + { + "bbox": [ + 62, + 228, + 295, + 270 + ], + "type": "text", + "content": "[4] Matthew Brown and Sabine Susstrunk. Multi-spectral sift for scene category recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 177-184, 2011. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 273, + 295, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 273, + 295, + 327 + ], + "spans": [ + { + "bbox": [ + 62, + 273, + 295, + 327 + ], + "type": "text", + "content": "[5] Pierre Charbonnier, Laure Blanc-Feraud, Gilles Aubert, and Michel Barlaud. Two deterministic half-quadratic regularization algorithms for computed imaging. In Proceedings of the International Conference on Image Processing, pages 168-172, 1994. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 330, + 295, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 330, + 295, + 384 + ], + "spans": [ + { + "bbox": [ + 62, + 330, + 295, + 384 + ], + "type": "text", + "content": "[6] Yinpeng Chen, Xiyang Dai, Mengchen Liu, Dongdong Chen, Lu Yuan, and Zicheng Liu. Dynamic convolution: Attention over convolution kernels. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11030-11039, 2020. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 386, + 295, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 386, + 295, + 441 + ], + "spans": [ + { + "bbox": [ + 62, + 386, + 295, + 441 + ], + "type": "text", + "content": "[7] Shen Cheng, Yuzhi Wang, Haibin Huang, Donghao Liu, Haoqiang Fan, and Shuaicheng Liu. Nbnet: Noise basis learning for image denoising with subspace projection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4896-4906, 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 442, + 295, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 442, + 295, + 485 + ], + "spans": [ + { + "bbox": [ + 62, + 442, + 295, + 485 + ], + "type": "text", + "content": "[8] Yuxiao Cheng, Runzhao Yang, Zhihong Zhang, Jinli Suo, and Qionghai Dai. A mutually boosting dual sensor computational camera for high quality dark videography. Information Fusion, 93:429-440, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 488, + 295, + 540 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 488, + 295, + 540 + ], + "spans": [ + { + "bbox": [ + 62, + 488, + 295, + 540 + ], + "type": "text", + "content": "[9] Sung-Jin Cho, Seo-Won Ji, Jun-Pyo Hong, Seung-Won Jung, and Sung-Jea Ko. Rethinking coarse-to-fine approach in single image deblurring. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4641-4650, 2021. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 544, + 295, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 544, + 295, + 588 + ], + "spans": [ + { + "bbox": [ + 57, + 544, + 295, + 588 + ], + "type": "text", + "content": "[10] Yuning Cui, Yi Tao, Zhenshan Bing, Wenqi Ren, Xinwei Gao, Xiaochun Cao, Kai Huang, and Alois Knoll. Selective frequency network for image restoration. In International Conference on Learning Representations, 2023. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 590, + 295, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 590, + 295, + 632 + ], + "spans": [ + { + "bbox": [ + 57, + 590, + 295, + 632 + ], + "type": "text", + "content": "[11] Xin Deng and Pier Luigi Dragotti. Deep convolutional neural network for multi-modal image restoration and fusion. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(10):3333-3348, 2020. 2, 3, 6, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 635, + 295, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 635, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 57, + 635, + 295, + 677 + ], + "type": "text", + "content": "[12] Michael Elad and Michal Aharon. Image denoising via sparse and redundant representations over learned dictionaries. IEEE Transactions on Image processing, 15(12):3736-3745, 2006. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 681, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 681, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 681, + 295, + 713 + ], + "type": "text", + "content": "[13] Hansen Feng, Lizhi Wang, Yuzhi Wang, and Hua Huang. Learnability enhancement for low-light raw denoising: Where paired real data meets noise modeling. In Proceed-" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "text", + "content": "ings of the ACM International Conference on Multimedia, pages 1436-1444, 2022. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 98, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 98, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 98, + 553, + 150 + ], + "type": "text", + "content": "[14] Hansen Feng, Lizhi Wang, Yuzhi Wang, Haoqiang Fan, and Hua Huang. Learnability enhancement for low-light raw image denoising: A data perspective. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(1):370-387, 2024. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 154, + 553, + 198 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 154, + 553, + 198 + ], + "spans": [ + { + "bbox": [ + 316, + 154, + 553, + 198 + ], + "type": "text", + "content": "[15] Yuanshen Guan, Ruikang Xu, Mingde Yao, Lizhi Wang, and Zhiwei Xiong. Mutual-guided dynamic network for image fusion. In Proceedings of the ACM International Conference on Multimedia, pages 1779-1788, 2023. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 201, + 553, + 265 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 201, + 553, + 265 + ], + "spans": [ + { + "bbox": [ + 317, + 201, + 553, + 265 + ], + "type": "text", + "content": "[16] Xiaowan Hu, Yuanhao Cai, Jing Lin, Haoqian Wang, Xin Yuan, Yulun Zhang, Radu Timofte, and Luc Van Gool. Hdnet: High-resolution dual-domain learning for spectral compressive imaging. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17542-17551, 2022. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 268, + 553, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 268, + 553, + 312 + ], + "spans": [ + { + "bbox": [ + 317, + 268, + 553, + 312 + ], + "type": "text", + "content": "[17] Liming Jiang, Bo Dai, Wayne Wu, and Chen Change Loy. Focal frequency loss for image reconstruction and synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13919-13929, 2021. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 315, + 553, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 315, + 553, + 369 + ], + "spans": [ + { + "bbox": [ + 317, + 315, + 553, + 369 + ], + "type": "text", + "content": "[18] Shuangping Jin, Bingbing Yu, Minhao Jing, Yi Zhou, Jiajun Liang, and Renhe Ji. Darkvisionnet: Low-light imaging via rgb-nir fusion with deep inconsistency prior. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1104-1112, 2022. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 372, + 553, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 372, + 553, + 404 + ], + "spans": [ + { + "bbox": [ + 316, + 372, + 553, + 404 + ], + "type": "text", + "content": "[19] Beomjun Kim, Jean Ponce, and Bumsub Ham. Deformable kernel networks for joint image filtering. International Journal of Computer Vision, 129(2):579-600, 2021. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 407, + 553, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 407, + 553, + 437 + ], + "spans": [ + { + "bbox": [ + 317, + 407, + 553, + 437 + ], + "type": "text", + "content": "[20] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 441, + 553, + 496 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 441, + 553, + 496 + ], + "spans": [ + { + "bbox": [ + 317, + 441, + 553, + 496 + ], + "type": "text", + "content": "[21] Lingshun Kong, Jiangxin Dong, Jianjun Ge, Mingqiang Li, and Jinshan Pan. Efficient frequency domain-based transformers for high-quality image deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5886-5895, 2023. 3, 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 498, + 553, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 498, + 553, + 552 + ], + "spans": [ + { + "bbox": [ + 317, + 498, + 553, + 552 + ], + "type": "text", + "content": "[22] Tong Li, Hansen Feng, Lizhi Wang, Zhiwei Xiong, and Hua Huang. Stimulating the diffusion model for image denoising via adaptive embedding and ensembling. In IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 8240-8257, 2024. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 555, + 553, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 555, + 553, + 610 + ], + "spans": [ + { + "bbox": [ + 317, + 555, + 553, + 610 + ], + "type": "text", + "content": "[23] Tong Li, Lizhi Wang, Zhiyuan Xu, Lin Zhu, Wanxuan Lu, and Hua Huang. Positive2negative: Breaking the information-lossy barrier in self-supervised single image denoising. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2025. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 613, + 553, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 613, + 553, + 666 + ], + "spans": [ + { + "bbox": [ + 317, + 613, + 553, + 666 + ], + "type": "text", + "content": "[24] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Swinir: Image restoration using swim transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1833-1844, 2021. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 670, + 555, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 555, + 713 + ], + "type": "text", + "content": "[25] Hezheng Lin, Xing Cheng, Xiangyu Wu, and Dong Shen. Cat: Cross attention in vision transformer. In Proceedings of the IEEE International Conference on Multimedia and Expo, pages 1-6, 2022. 5" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "12687" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "text", + "content": "[26] Feifan Lv, Yinqiang Zheng, Yicheng Li, and Feng Lu. An integrated enhancement solution for 24-hour colorful imaging. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 11725-11732, 2020. 1, 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 160 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 160 + ], + "type": "text", + "content": "[27] Ruijun Ma, Shuyi Li, Bob Zhang, and Zhengming Li. Generative adaptive convolutions for real-world noisy image denoising. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1935-1943, 2022. 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 162, + 294, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 162, + 294, + 205 + ], + "spans": [ + { + "bbox": [ + 56, + 162, + 294, + 205 + ], + "type": "text", + "content": "[28] Youssef Mansour and Reinhard Heckel. Zero-shot noise2noise: Efficient image denoising without any data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14018-14027, 2023. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 206, + 294, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 206, + 294, + 249 + ], + "spans": [ + { + "bbox": [ + 56, + 206, + 294, + 249 + ], + "type": "text", + "content": "[29] Javier Portilla, Vasily Strela, Martin J Wainwright, and Eero P Simoncelli. Image denoising using scale mixtures of gaussians in the wavelet domain. IEEE Transactions on Image Processing, 12(11):1338-1351, 2003. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 251, + 294, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 251, + 294, + 293 + ], + "spans": [ + { + "bbox": [ + 56, + 251, + 294, + 293 + ], + "type": "text", + "content": "[30] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention, pages 234-241, 2015. 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 295, + 294, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 295, + 294, + 336 + ], + "spans": [ + { + "bbox": [ + 56, + 295, + 294, + 336 + ], + "type": "text", + "content": "[31] Zehua Sheng, Xiongwei Liu, Si-Yuan Cao, Hui-Liang Shen, and Huaqi Zhang. Frequency-domain deep guided image denoising. IEEE Transactions on Multimedia, 25:6767-6781, 2022. 2, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 338, + 294, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 338, + 294, + 403 + ], + "spans": [ + { + "bbox": [ + 56, + 338, + 294, + 403 + ], + "type": "text", + "content": "[32] Zehua Sheng, Zhu Yu, Xiongwei Liu, Si-Yuan Cao, Yuqi Liu, Hui-Liang Shen, and Huaqi Zhang. Structure aggregation for cross-spectral stereo image guided denoising. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13997-14006, 2023. 2, 3, 6, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 404, + 294, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 404, + 294, + 470 + ], + "spans": [ + { + "bbox": [ + 56, + 404, + 294, + 470 + ], + "type": "text", + "content": "[33] Matthew Tancik, Pratul Srinivasan, Ben Mildenhall, Sara Fridovich-Keil, Nithin Raghavan, Utkarsh Singhal, Ravi Ramamoorthi, Jonathan Barron, and Ren Ng. Fourier features let networks learn high frequency functions in low dimensional domains. Advances in Neural Information Processing Systems, 33:7537-7547, 2020. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 471, + 294, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 471, + 294, + 503 + ], + "spans": [ + { + "bbox": [ + 56, + 471, + 294, + 503 + ], + "type": "text", + "content": "[34] Yuki Tatsunami and Masato Taki. Fft-based dynamic token mixer for vision. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 15328-15336, 2024. 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 504, + 294, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 504, + 294, + 546 + ], + "spans": [ + { + "bbox": [ + 56, + 504, + 294, + 546 + ], + "type": "text", + "content": "[35] Chunwei Tian, Lunke Fei, Wenxian Zheng, Yong Xu, Wangmeng Zuo, and Chia-Wen Lin. Deep learning on image denoising: An overview. *Neural Networks*, 131:251–275, 2020. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 548, + 294, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 548, + 294, + 591 + ], + "spans": [ + { + "bbox": [ + 56, + 548, + 294, + 591 + ], + "type": "text", + "content": "[36] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in Neural Information Processing Systems, 30, 2017. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 593, + 294, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 593, + 294, + 635 + ], + "spans": [ + { + "bbox": [ + 56, + 593, + 294, + 635 + ], + "type": "text", + "content": "[37] Renjie Wan, Boxin Shi, Wenhan Yang, Bihan Wen, LingYu Duan, and Alex C. Kot. Purifying low-light images via near-infrared enlightened image. IEEE Transactions on Multimedia, 25:8006-8019, 2023. 1, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 636, + 294, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 636, + 294, + 679 + ], + "spans": [ + { + "bbox": [ + 56, + 636, + 294, + 679 + ], + "type": "text", + "content": "[38] Jian Wang, Tianfan Xue, Jonathan T Barron, and Jiawen Chen. Stereoscopic dark flash for low-light photography. In IEEE International Conference on Computational Photography, pages 1-10, 2019. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 681, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 681, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 681, + 294, + 713 + ], + "type": "text", + "content": "[39] Xin Wang, Lizhi Wang, Xiangtian Ma, Maoqing Zhang, Lin Zhu, and Hua Huang. In2set: Intra-inter similarity exploiting transformer for dual-camera compressive hyperspectral" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 105 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 105 + ], + "type": "text", + "content": "imaging. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24881-24891, 2024. 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 106, + 553, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 106, + 553, + 149 + ], + "spans": [ + { + "bbox": [ + 316, + 106, + 553, + 149 + ], + "type": "text", + "content": "[40] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing, 13(4):600-612, 2004. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 151, + 553, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 151, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 316, + 151, + 553, + 205 + ], + "type": "text", + "content": "[41] Zhendong Wang, Xiaodong Cun, Jianmin Bao, Wengang Zhou, Jianzhuang Liu, and Houqiang Li. Uformer: A general u-shaped transformer for image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17683-17693, 2022. 6, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 206, + 553, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 206, + 553, + 259 + ], + "spans": [ + { + "bbox": [ + 316, + 206, + 553, + 259 + ], + "type": "text", + "content": "[42] Yunxuan Wei, Shuhang Gu, Yawei Li, Radu Timofte, Longcun Jin, and Hengjie Song. Unsupervised real-world image super resolution via domain-distance aware training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13385-13394, 2021. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 261, + 553, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 261, + 553, + 304 + ], + "spans": [ + { + "bbox": [ + 316, + 261, + 553, + 304 + ], + "type": "text", + "content": "[43] Rongjian Xu, Zhilu Zhang, Renlong Wu, and Wangmeng Zuo. Nir-assisted image denoising: A selective fusion approach and a real-world benchmark dataset. arXiv preprint arXiv:2404.08514, 2024. 2, 3, 6, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 305, + 553, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 305, + 553, + 348 + ], + "spans": [ + { + "bbox": [ + 316, + 305, + 553, + 348 + ], + "type": "text", + "content": "[44] Shuang Xu, Jiangshe Zhang, Jialin Wang, Kai Sun, Chunxia Zhang, Junmin Liu, and Junying Hu. A model-driven network for guided image denoising. Information Fusion, 85: 60-71, 2022. 2, 3, 6, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 350, + 553, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 350, + 553, + 392 + ], + "spans": [ + { + "bbox": [ + 316, + 350, + 553, + 392 + ], + "type": "text", + "content": "[45] Zhi-Qin John Xu, Yaoyu Zhang, Tao Luo, Yanyang Xiao, and Zheng Ma. Frequency principle: Fourier analysis sheds light on deep neural networks. arXiv preprint arXiv:1901.06523, 2019. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 394, + 553, + 447 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 394, + 553, + 447 + ], + "spans": [ + { + "bbox": [ + 316, + 394, + 553, + 447 + ], + "type": "text", + "content": "[46] Qiong Yan, Xiaoyong Shen, Li Xu, Shaojie Zhuo, Xiaopeng Zhang, Liang Shen, and Jiaya Jia. Cross-field joint image restoration via scale map. In Proceedings of the IEEE International Conference on Computer Vision, pages 1537-1544, 2013. 1, 2, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 449, + 553, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 449, + 553, + 514 + ], + "spans": [ + { + "bbox": [ + 316, + 449, + 553, + 514 + ], + "type": "text", + "content": "[47] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, Ming-Hsuan Yang, and Ling Shao. Multi-stage progressive image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14821-14831, 2021. 1, 2, 4, 6, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 515, + 553, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 515, + 553, + 580 + ], + "spans": [ + { + "bbox": [ + 316, + 515, + 553, + 580 + ], + "type": "text", + "content": "[48] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5728-5739, 2022. 2, 6, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 582, + 553, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 582, + 553, + 624 + ], + "spans": [ + { + "bbox": [ + 316, + 582, + 553, + 624 + ], + "type": "text", + "content": "[49] Jiale Zhang, Yulun Zhang, Jinjin Gu, Jiahua Dong, Linghe Kong, and Xiaokang Yang. Xformer: Hybrid x-shaped transformer for image denoising. arXiv preprint arXiv:2303.06440, 2023. 2, 6, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 625, + 553, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 625, + 553, + 668 + ], + "spans": [ + { + "bbox": [ + 316, + 625, + 553, + 668 + ], + "type": "text", + "content": "[50] Kai Zhang, Wangmeng Zuo, Yunjin Chen, Deyu Meng, and Lei Zhang. Beyond a gaussian denoiser: Residual learning of deep cnn for image denoising. IEEE Transactions on Image Processing, 26(7):3142-3155, 2017. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "text", + "content": "[51] Kai Zhang, Wangmeng Zuo, and Lei Zhang. Ffdnet: Toward a fast and flexible solution for cnn-based image denoising. IEEE Transactions on Image Processing, 27(9):4608-4622, 2018. 1" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "12688" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 173 + ], + "type": "list", + "angle": 0, + "index": 2, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "text", + "content": "[52] Jingkai Zhou, Varun Jampani, Zhixiong Pi, Qiong Liu, and Ming-Hsuan Yang. Decoupled dynamic filter networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6647-6656, 2021. 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 173 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 173 + ], + "type": "text", + "content": "[53] Wenbin Zou, Mingchao Jiang, Yunchen Zhang, Liang Chen, Zhiyong Lu, and Yi Wu. Sdwnet: A straight dilated network with wavelet transformation for image deblurring. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1895–1904, 2021. 3" + } + ] + } + ], + "index": 1 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "12689" + } + ] + } + ], + "index": 3 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Completion as Enhancement_ A Degradation-Aware Selective Image Guided Network for Depth Completion/5b8a9777-b777-4fb5-9ba2-85ebeb5a648a_content_list.json b/2025/Completion as Enhancement_ A Degradation-Aware Selective Image Guided Network for Depth Completion/5b8a9777-b777-4fb5-9ba2-85ebeb5a648a_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..45ab9dceb4974318d18896ca4c26daf177e23a32 --- /dev/null +++ b/2025/Completion as Enhancement_ A Degradation-Aware Selective Image Guided Network for Depth Completion/5b8a9777-b777-4fb5-9ba2-85ebeb5a648a_content_list.json @@ -0,0 +1,1854 @@ +[ + { + "type": "text", + "text": "Completion as Enhancement: A Degradation-Aware Selective Image Guided Network for Depth Completion", + "text_level": 1, + "bbox": [ + 106, + 130, + 890, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zhiqiang Yan Zhengxue Wang Kun Wang Jun Li\\* Jian Yang\\*", + "bbox": [ + 223, + 203, + 779, + 222 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "PCA Lab†, Nanjing University of Science and Technology", + "bbox": [ + 267, + 222, + 728, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{yanzq,zxwang,kunwang,junli,csjyang}@njust.edu.cn", + "bbox": [ + 279, + 241, + 712, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 291, + 326, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we introduce the Selective Image Guided Network (SigNet), a novel degradation-aware framework that transforms depth completion into depth enhancement for the first time. Moving beyond direct completion using convolutional neural networks (CNNs), SigNet initially densifies sparse depth data through non-CNN densification tools to obtain coarse yet dense depth. This approach eliminates the mismatch and ambiguity caused by direct convolution over irregularly sampled sparse data. Subsequently, SigNet redefines completion as enhancement, establishing a self-supervised degradation bridge between the coarse depth and the targeted dense depth for effective RGB-D fusion. To achieve this, SigNet leverages the implicit degradation to adaptively select high-frequency components (e.g., edges) of RGB data to compensate for the coarse depth. This degradation is further integrated into a multi-modal conditional Mamba, dynamically generating the state parameters to enable efficient global high-frequency information interaction. We conduct extensive experiments on the NYUv2, DIML, SUN RGBD, and TOFDC datasets, demonstrating the state-of-the-art (SOTA) performance of SigNet.", + "bbox": [ + 88, + 321, + 485, + 641 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 666, + 220, + 681 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Depth completion [27, 38, 48, 50] aims to predict a dense depth map from sparse and noisy depth measurements, often using the corresponding RGB image for auxiliary guidance. The resulting dense depth map can benefit various downstream applications such as scene reconstruction [36, 39, 47, 68] and self-driving [22, 45, 51, 63, 65, 67].", + "bbox": [ + 89, + 691, + 482, + 782 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The key challenges of this task lie in the sparse depth input data [21, 38, 40, 49, 50]. We summarize these challenges as mismatch and ambiguity. As shown in Fig. 1, the mismatch is caused by the irregularly sampled sparse", + "bbox": [ + 89, + 782, + 483, + 844 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4319659526abef301e7b370531fede098225e97d949d8e8117db046c20b330a4.jpg", + "image_caption": [ + "Figure 1. Illustration of our main concept. It redefines depth completion as depth enhancement. We first transform sparse depth input via non-CNN densification tools [14, 15], yielding coarse but dense depth. Then we predict the precise and dense depth from the coarse depth by leveraging degradation assumption [43, 59, 62]." + ], + "image_footnote": [], + "bbox": [ + 517, + 287, + 906, + 526 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "depth with varying point distributions, while convolutions are spatially invariant with fixed kernel shapes. Moreover, it is ambiguous for convolutions to distinguish between valid and invalid depth measurements since unknown pixels in the sparse depth are recognized as zero.", + "bbox": [ + 511, + 638, + 906, + 715 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To address these challenges, several excellent methods [21, 38, 40] have been proposed. For instance, SparseCOnvs [40] proposes sparsity invariant CNNs that explicitly model the location of missing data. To avoid directly convolving on sparse depth, Liu et al. [21] employs kernels learned by differentiable kernel regression layers to interpolate sparse data. BPNet [38] develops bilateral propagation, where the coefficients are generated based on both radiometric difference and spatial distance. These data-driven schemes are robust and impressive. However, they still cannot fully avoid the application of convolution on incomplete depth data, even though the depth data has become denser.", + "bbox": [ + 511, + 719, + 908, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding authors", + "bbox": [ + 107, + 851, + 240, + 863 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$\\dagger$ PCA Lab, Key Lab of Intelligent Perception and Systems for High-Dimensional Information of Ministry of Education, School of Computer Science and Engineering, Nanjing University of Science and Technology.", + "bbox": [ + 91, + 863, + 482, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "26943", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, our objective is to develop a novel solution that not only completely resolves the issues of mismatch and ambiguity but also provides a fresh perspective on revisiting the traditional depth completion task.", + "bbox": [ + 89, + 90, + 480, + 151 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To this end, we propose a novel framework called the Selective Image Guided Network (SigNet). As shown in Fig. 1, SigNet redefines depth completion as depth enhancement1. It first leverages non-CNN densification tools [14, 15] to obtain coarse but dense depth. This coarse depth is then enhanced by a CNN-based subnetwork. Inspired by previous approaches [59, 62], we assume that there exists unknown degradation $\\mathbf{h}$ between the coarse depth $\\mathbf{Z}$ and the target depth $\\mathbf{Y}$ , described as", + "bbox": [ + 89, + 152, + 480, + 287 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {Z} = \\mathbf {h} \\mathbf {Y} + \\mathbf {n}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 236, + 294, + 480, + 309 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "where $n$ denotes additive noise. Therefore, we can establish a connection between the source depth and the target depth through the degradation. A fundamental observation is that degradation typically occurs near edges, suggesting that this phenomenon can be leveraged to enhance RGB-D fusion. This is particularly relevant given that the coarse depth often lacks significant edge details, while the RGB image contains abundant high-frequency components. Specifically, SigNet introduces a Degradation-Aware Decomposition and Fusion (DADF) module. DADF begins by decomposing the implicit degradation $h$ into multiple patches and estimating their coefficients in the discrete cosine domain. Each patch is then used to adaptively select the most representative high-frequency information from the RGB image. To achieve effective RGB-D aggregation, DADF incorporates a conditional Mamba to encode long-range correlations. Unlike the vanilla Vision Mamba [25], which calculates state parameters from the input, DADF uses the degradation patches as additional conditions to generate the state parameters. This approach enables further global high-frequency information interaction.", + "bbox": [ + 91, + 316, + 482, + 632 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our contributions are as follows:", + "bbox": [ + 109, + 633, + 410, + 647 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce a novel framework called SigNet, which redefines depth completion as depth enhancement for the first time. By adopting a degradation-aware approach, SigNet not only offers a new perspective to the field but also fully eliminates the traditional challenges such as mismatch and ambiguity.", + "- We propose a Degradation-Aware Decomposition and Fusion (DADF) module. Utilizing degradation decomposition, DADF adaptively selects high-frequency components of RGB data to compensate for coarse depth. Furthermore, DADF achieves effective RGB-D fusion by incorporating conditional Mamba, where the state parameters are generated from degradation patches." + ], + "bbox": [ + 89, + 648, + 482, + 844 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- We evaluate SigNet through comprehensive experiments, demonstrating its superiority over previous SOTAs.", + "bbox": [ + 513, + 90, + 903, + 121 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 133, + 653, + 148 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Depth Completion. Depth completion can be broadly categorized into depth-only and image-guided categories.", + "bbox": [ + 511, + 159, + 903, + 188 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Depth-only methods [9, 12, 14, 28, 40, 41, 44, 50] obtain dense depth directly from a single sparse depth input. For example, IP-Basic [14] designs reasonable combinations of basic image processing operations to densify sparse depth without using CNNs. Ma et al. [28] employ a UNet-like architecture to progressively densify the sparse depth input. FusionNet [41] designs a lightweight completion network that aggregates both global and local information with confidence adjustment. Moreover, Lu et al. [26] introduce a color image branch as auxiliary supervision during training.", + "bbox": [ + 509, + 189, + 903, + 340 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Despite advancements in single-modal methods, their performance is often limited without supplementary reference information. To address this challenge, numerous image-guided techniques have been proposed [11, 16, 19, 22, 37, 38, 45, 50, 53, 61, 66]. For example, GuideNet [37] and RigNet [48] utilize color images to generate adaptive filtering kernels for depth feature extraction. GFormer [32] and CFormer [61] integrate convolution and transformer mechanisms to capture both local and global representations. LRRU [44] introduces a dynamic kernel scope that transitions from large to small to capture varying dependencies. FuseNet [3] and PointDC [56] incorporate LiDAR point cloud branches to model 3D geometries. BEV@DC [66] uses bird's-eye view to build a point-voxel architecture. Additionally, DepthPrompting [30] embeds priors of large depth models for better sensor-agnostic generalization.", + "bbox": [ + 511, + 340, + 903, + 580 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To further advance completion models, spatial propagation networks (SPNs) have been introduced. The initial SPN [23] is designed to learn a pairwise similarity matrix. CSPN [5] applies this technique to depth completion by conducting recursive convolutions with fixed local neighborhood kernels, while $\\mathrm{CSPN} + +$ [7] adapts by learning variable kernel sizes. PENet [11] expands the receptive fields using dilated convolutions. Furthermore, NLSPN [29] integrates non-local neighbors via deformable convolutions. DySPN [20] generates dynamic non-linear neighbors using an attention mechanism. Most recently, 3D SPNs [6] have been developed to encode 3D geometry. For example, GraphC-SPN [24] employs additional geometric constraints to regulate 3D propagation. TPVD [50] introduces a geometric SPN that constructs the affinity in 2D tri-perspective view spaces and their combined 3D projection space.", + "bbox": [ + 511, + 582, + 903, + 823 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Degradation Learning. Recently, color image restoration has been significantly advanced by improvements in degradation representations [43, 55, 59]. For example, Zhang et al. [59] construct a sophisticated yet practical degradation model to simulate real-world image degradation for", + "bbox": [ + 511, + 825, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "1Specifically depth super-resolution (DSR). Mainstream DSR methods restore target depth directly from upsampled depth. Consequently, the processes and the end goals of depth completion and DSR are identical, i.e., predicting clear depth from the same-resolution coarse depth.", + "bbox": [ + 89, + 849, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "26944", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/38425302d36afe1dcc84c4999e36e59711a1aa5fed916640e10ca19e005df0ea.jpg", + "image_caption": [ + "Figure 2. Pipeline of SigNet. The sparse depth data is initially filled to create a coarse depth map. We then utilize the degradation assumption in Eq. (1) to establish a connection between this coarse map and the target prediction. The color image features and degradation representations are employed to guide the multi-modal fusion through our proposed DADF module, as depicted in Fig. 3." + ], + "image_footnote": [], + "bbox": [ + 155, + 88, + 843, + 380 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "blind image super-resolution. Similarly, Zhou et al. [64] propose to estimate degradation for each local image region using a degradation-adaptive regression filter for blind super-resolution. Liang et al. [18] present a degradation-adaptive model that specifies the network parameters by estimating the degradation. Furthermore, several image restoration methods [17, 57, 58] have been proposed to simultaneously address multiple forms of degradation. For example, Zhang et al. [57] develop a multi-degradation image restoration network that incrementally learns degradation representations via clustering, without relying on prior knowledge of the degradation information. Zhang et al. [58] build an ingredient-oriented degradation reformulation network to model the relationship among different images. Additionally, Yao et al. [54] introduce a neural degradation representation learning method that captures latent degradation features, adaptively decomposing various types of degradation to enable comprehensive restoration.", + "bbox": [ + 88, + 450, + 485, + 723 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 738, + 183, + 753 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Overview. Fig. 2 shows our SigNet pipeline that consists of the depth branch, degradation branch, and image branch. In the depth branch, the sparse depth $\\mathbf{S}$ is first filled via nonCNN densification tools [14, 15], obtaining a coarse depth map $\\mathbf{Z}$ . According to Eq. (1), we employ denoising operators to weaken the negative impact of the noise $\\mathbf{n}$ , yielding $\\hat{\\mathbf{Z}}$ . Then, we use residual groups [60] to extract features of the denoised depth $\\hat{\\mathbf{Z}}$ and the color image $\\mathbf{I}$ in the image branch. In the degradation branch, the implicit degradation", + "bbox": [ + 89, + 763, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "$\\mathbf{D}$ is generated from a degradation network, where convolutions and residual groups are used. We thus utilize the image feature and degradation $\\mathbf{D}$ to facilitate RGB-D fusion via our proposed Degradation-Aware Decomposition and Fusion (DADF). In summary, our SigNet redefines the depth completion as depth enhancement through the degradation assumption in Eq. (1), making the model free from the challenges of mismatch and ambiguity.", + "bbox": [ + 511, + 450, + 906, + 571 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Task Switching", + "text_level": 1, + "bbox": [ + 511, + 578, + 668, + 593 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Degradation learning is widely applied in low-level image restoration [43, 55, 59], yielding impressive results. These tasks exhibit two key characteristics: 1) both input and target data are typically dense, and 2) the input data generally have a lower resolution than the target data. Actually, the low-resolution input data are typically upsampled using interpolation methods, such as bicubic interpolation. Consequently, the degradation matrix usually consists of a convolution whose kernel is derived from degradation representations and a downsampling function. In contrast, the depth completion task takes sparse depth data as input, and the resolutions of the input and target data are the same. Thus, we reframe depth completion as depth enhancement by drawing an analogy to image restoration.", + "bbox": [ + 511, + 599, + 906, + 811 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As shown in Fig. 2, we first fill the sparse depth $\\mathbf{S}$ using non-CNN densification tools [14, 15], denoted as $f_{ds}(\\cdot)$ , resulting in a coarse yet dense depth $\\mathbf{Z}$ , formulated as:", + "bbox": [ + 511, + 811, + 906, + 857 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {Z} = f _ {d s} (\\mathbf {S}). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 665, + 863, + 903, + 878 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As discussed in Sec. 1, this approach eliminates the mis", + "bbox": [ + 511, + 885, + 903, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "26945", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/7c0d8234f69e1a252db2271431e620e0e785ffc4e9e5184fa9352b8cc68fa270.jpg", + "image_caption": [ + "Figure 3. Overview of our DADF. $\\odot$ and CM refer to elementwise multiplication and the conditional Mamba, respectively." + ], + "image_footnote": [], + "bbox": [ + 91, + 85, + 483, + 265 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "match and ambiguity caused by direct convolution over irregularly sampled sparse data. Subsequently, we employ a degradation network to estimate the implicit degradation $\\mathbf{D}$ . The degradation kernel $\\mathbf{K}$ is produced from $\\mathbf{D}$ using convolutions. Consequently, given the coarse depth $\\mathbf{Z}$ and target depth $\\mathbf{Y}$ , Eq. (1) can be rewritten based on the task switching as follows:", + "bbox": [ + 89, + 327, + 483, + 434 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {Z} = \\mathbf {K} (\\mathbf {D}; \\vartheta) \\otimes \\mathbf {Y} + \\mathbf {n}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 202, + 443, + 482, + 460 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\otimes$ indicates convolution operation, $\\vartheta$ is the network parameter. Since the resolutions are the same, our degradation equation does not need the downsampling function.", + "bbox": [ + 89, + 470, + 483, + 516 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Degradation-Aware Decomposition and Fusion", + "text_level": 1, + "bbox": [ + 89, + 523, + 483, + 540 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As we know, degradation typically occurs near edges, suggesting that this cue can be leveraged to enhance RGB-D fusion, especially considering that the coarse depth often lacks significant edge details, while the RGB image contains abundant high-frequency components. Thus, we propose DADF that is shown in Fig. 3. We denote the image and depth features as $\\mathbf{F}_{img}$ and $\\mathbf{F}_{dep}$ , obtaining:", + "bbox": [ + 89, + 546, + 483, + 652 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {F} _ {i m g} = f _ {e} (\\mathbf {I}), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 215, + 662, + 482, + 679 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {F} _ {d e p} = f _ {e} \\left(f _ {d n} (\\mathbf {Z})\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 217, + 681, + 482, + 699 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $f_{e}(\\cdot)$ and $f_{dn}(\\cdot)$ refer to the feature extraction network and denoising function, respectively. As presented in Fig. 3, DADF takes as input the image feature $\\mathbf{F}_{img}$ , depth feature $\\mathbf{F}_{dep}$ , and degradation $\\mathbf{F}_D$ .", + "bbox": [ + 89, + 708, + 483, + 768 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Decomposition. To extract the most representative components of $\\mathbf{F}_D$ , we first apply the discrete cosine transform (DCT) to convert it into the frequency domain. Next, it adaptively learns a spectrum mask to enhance informative frequency bands while suppressing irrelevant ones. Finally, we reconstruct the spatial representation using the inverse DCT. This process is formulated as:", + "bbox": [ + 89, + 768, + 483, + 875 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{\\mathbf {D} _ {i} \\mid i = 1, 2, 3 \\right\\} = f _ {d c t} (\\mathbf {D}). \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 191, + 883, + 482, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Algorithm 1: Conditional SSM Block", + "text_level": 1, + "bbox": [ + 522, + 94, + 779, + 109 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Input: $\\mathbf{F}_{rgbd}$ , $\\{\\mathbf{D}_i|i = 1,2,3\\} :(\\mathrm{HW},\\mathrm{C})$", + "bbox": [ + 534, + 112, + 792, + 126 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Output: $\\hat{\\mathbf{F}}_{rgbd}:(\\mathrm{HW},\\mathrm{C})$", + "bbox": [ + 534, + 126, + 696, + 140 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1 A : (C, N) $\\leftarrow$ ParameterA", + "2 $\\mathbf{B}:(\\mathrm{HW},\\mathrm{N})\\gets \\mathbf{Linear}_B(f_{gru}(\\mathbf{F}_{rgbd},\\{\\mathbf{D}_i|i = 1,2,3\\}))$", + "3 $\\mathbf{C}:(\\mathrm{HW},\\mathrm{N})\\gets \\mathbf{Linear}_C(f_{gru}(\\mathbf{F}_{rgbd},\\{\\mathbf{D}_i|i = 1,2,3\\}))$", + "4 $\\Delta$ : (HW, C) $\\leftarrow$ $\\log(1 + \\exp(\\text{Linear}_\\Delta(f_{gru}(\\mathbf{F}_{rgbd}, \\{\\mathbf{D}_i | i = 1, 2, 3\\})) + \\text{Parameter}_\\Delta))$", + "5 $\\overline{\\mathbf{A}}: (\\mathrm{HW}, \\mathrm{C}, \\mathrm{N}) \\leftarrow \\exp(\\Delta \\otimes \\mathbf{A})$", + "6 $\\overline{\\mathbf{B}}:\\mathrm{(HW,C,N)}\\gets \\pmb {\\Delta}\\otimes \\mathbf{B}$", + "7 $\\hat{\\mathbf{F}}_{rgbd} \\gets \\mathrm{SSM}(\\overline{\\mathbf{A}},\\overline{\\mathbf{B}},\\mathbf{C})(\\mathbf{F}_{rgbd})$", + "8 return $\\hat{\\mathbf{F}}_{rgbd}$" + ], + "bbox": [ + 521, + 141, + 834, + 315 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We apply concatenation, convolution, pooling, and Sigmoid operations, collectively denoted as $f_{m}(\\cdot)$ , to generate a mask $m$ . This mask is then multiplied with the combined frequency components to produce an updated mask $\\hat{m}$ , enabling the adaptive selection of the representative image feature $\\hat{\\mathbf{F}}_{img}$ . The process is defined as:", + "bbox": [ + 511, + 349, + 906, + 441 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nm = f _ {m} \\left(\\mathbf {F} _ {\\text {i m g}}, \\mathbf {D}\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 593, + 453, + 903, + 469 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {m} = f _ {i d c t} \\left(m \\cdot \\left(f _ {a c} \\left(\\sum_ {i = 1} ^ {3} \\mathbf {D} _ {i}\\right)\\right)\\right), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 594, + 473, + 903, + 500 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {F}} _ {i m g} = \\hat {m} \\cdot \\mathbf {F} _ {i m g}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 594, + 501, + 903, + 520 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $f_{idct}$ refers to inverse DCT and $f_{ac}(\\cdot)$ represents the combination function of addition and convolutions.", + "bbox": [ + 511, + 532, + 903, + 561 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Fusion. As a result, we further leverage the decomposed degradation $\\mathbf{D}_i$ to enable effective RGB-D fusion. It is well-known that Mamba [10] is good at handling long-range correlations. Inspired by recent Vision Mamba [25], we present a variant, i.e., Conditional Mamba (CM). As shown in Fig. 3, the selected image feature $\\hat{\\mathbf{F}}_{img}$ and raw depth feature $\\mathbf{F}_{dep}$ are concatenated into CM, where the decomposed degradation $\\mathbf{D}_i$ serves as the condition:", + "bbox": [ + 511, + 561, + 906, + 684 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {F} _ {r g b d} = f _ {c c} \\left(\\tilde {\\mathbf {F}} _ {i m g}, \\mathbf {F} _ {d e p}\\right), \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 695, + 903, + 713 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {F}} _ {r g b d} = f _ {c m} \\left(\\mathbf {F} _ {r g b d}, \\left\\{\\mathbf {D} _ {i} \\mid i = 1, 2, 3 \\right\\}\\right), \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 578, + 717, + 903, + 734 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $f_{cc}$ denotes the combination function of concatenation and convolutions, $f_{cm}$ is the Condition Mamba function. As reported in Alg. 1, the state parameters $\\mathbf{B}$ and $\\mathbf{C}$ are calculated based on both the RGB-D input and degradation using gated recurrent convolution units.", + "bbox": [ + 511, + 747, + 906, + 821 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Loss Function", + "text_level": 1, + "bbox": [ + 511, + 832, + 658, + 847 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As shown in our pipeline, the total loss function $\\mathcal{L}_t$ consists of a reconstruction loss $\\mathcal{L}_r$ and a self-supervised degradation learning loss $\\mathcal{L}_d$ . Let $\\hat{\\mathbf{Y}}$ and $\\mathbf{Y}$ denote the ground truth", + "bbox": [ + 511, + 854, + 906, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "26946", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/e5f92265e359e440d8787dfe0df627f93e9a03e5bea2a8467bde02142b45aaba.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetMethodParams. (M) ↓RMSE (mm) ↓REL ↓δ1(%) ↑δ2(%) ↑δ3(%) ↑Venue
NYUv2Bilateral [1]-5320.13285.193.595.9CVPR 2005
S2D [28]26.12300.05494.597.398.9ICRA 2019
CSPN [4]17.41730.02096.398.699.5ECCV 2018
DfuseNet [34]-1560.01698.899.799.9ITSC 2019
DM-LRN [33]-2050.01498.899.699.9IROS 2021
RDF-GAN [42]-1390.01398.799.699.9CVPR 2022
AGG-Net [2]129.1920.01499.499.9100.0ICCV 2023
BPNet [38]89.9880.01199.699.9100.0CVPR 2024
SigNet (ours)3.3830.01299.799.9100.0CVPR 2025
DIMLBilateral [1]-6360.18983.088.892.4CVPR 2005
CSPN [4]17.41620.03396.198.799.6ECCV 2018
DfuseNet [34]-1430.02398.499.499.9ITSC 2019
DM-LRN [33]-1490.01599.099.699.9IROS 2021
AGG-Net [2]129.1780.01199.699.9100.0ICCV 2023
BPNet [38]89.9680.01199.499.8100.0CVPR 2024
SigNet (ours)3.3550.00699.799.9100.0CVPR 2025
SUN RGBDS2D [28]26.13290.07493.997.098.1ICRA 2019
CSPN [4]17.42950.13795.697.598.4ECCV 2018
DLiDAR [31]53.42790.06196.998.098.4CVPR 2019
DM-LRN [33]-2670.06397.698.298.7IROS 2021
RDF-GAN [42]-2550.05996.998.499.0CVPR 2022
AGG-Net [2]129.11520.03898.599.099.4ICCV 2023
BPNet [38]89.91240.02598.999.499.8CVPR 2024
SigNet (ours)3.31160.02299.099.499.8CVPR 2025
TOFDCCSPN [4]17.42240.04294.595.396.5ECCV 2018
GuideNet [37]73.51460.03097.698.999.5TIP 2020
PENet [11]131.52410.04394.695.395.5ICRA 2021
NLSPN [29]25.81740.02996.497.998.9ECCV 2020
CFounder [61]83.51130.02999.199.699.9CVPR 2023
RigNet [48]65.21330.02597.699.199.7ECCV 2022
GraphCSPN [24]26.42530.05292.096.998.7ECCV 2022
PointDC [56]25.11090.02198.599.299.6ICCV 2023
TPVD [50]31.2920.01499.199.699.9CVPR 2024
SigNet (ours)3.3870.01299.299.799.9CVPR 2025
", + "bbox": [ + 125, + 87, + 870, + 588 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1. Quantitative comparison on four depth completion datasets. The best and suboptimal results are highlighted. Note that most of the results on NYUv2, DIML, and SUN RGBD are borrowed directly from AGG-Net [2], while those of TOFDC are from TPVD [50].", + "bbox": [ + 89, + 593, + 906, + 621 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "depth and the predicted depth, respectively. The reconstruction loss $\\mathcal{L}_r$ is defined as:", + "bbox": [ + 89, + 647, + 482, + 678 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {r} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} | | \\hat {\\mathbf {Y}} _ {i} - \\mathbf {Y} _ {i} | | _ {1}, \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 194, + 686, + 482, + 728 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $N$ represents the number of training samples, and $\\| \\cdot \\| _1$ denotes the $L_{1}$ norm. Subsequently, we introduce the degradation loss $\\mathcal{L}_d$ to facilitate the degradation learning process through:", + "bbox": [ + 89, + 734, + 483, + 797 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {d} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\left| \\left| \\mathbf {K} \\otimes \\mathbf {Y} - \\mathbf {Z} \\right| \\right| _ {1}, \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 184, + 805, + 482, + 845 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{K}$ refers to the degradation kernel learned from the implicit degradation representation $\\mathbf{D}$ . $\\mathbf{Z}$ corresponds to the coarse depth. $\\otimes$ denotes the convolution operation, where", + "bbox": [ + 89, + 854, + 483, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$\\mathbf{K}$ serves as the convolution kernel. As a result, the total training loss is formulated as:", + "bbox": [ + 511, + 647, + 906, + 676 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {t} = \\mathcal {L} _ {r} + \\lambda \\mathcal {L} _ {d}, \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 651, + 686, + 906, + 704 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda$ is a hyper-parameter, empirically set to 0.1. By the constraint of the auxiliary loss $\\mathcal{L}_d$ , our model can build additional connections between the input and output, thus predicting more accurate depth results.", + "bbox": [ + 511, + 710, + 906, + 772 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiment", + "text_level": 1, + "bbox": [ + 511, + 784, + 638, + 801 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We evaluate our SigNet method on four popular datasets, including NYUv2, DIML, SUN RGBD, and TOFDC benchmarks. For NYUv2, DIML, and SUN RGBD, we maintain the settings used in AGG-Net [2], while for TOFDC, we follow the settings of TPVD [50]. Refer to our supplementary material for additional details on the experimental setups.", + "bbox": [ + 511, + 809, + 906, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "26947", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/09e2b28f93e8421502ee794244836a54580f10db81a9fb4a6e4e5ab74e6bf353.jpg", + "image_caption": [ + "(a) RGB" + ], + "image_footnote": [], + "bbox": [ + 138, + 89, + 282, + 258 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/6f96e3f1244f41c55d4c841468fdd4b7e96c91d675d80575c6bda5e03ad1c750.jpg", + "image_caption": [ + "(b) NLSPN" + ], + "image_footnote": [], + "bbox": [ + 282, + 89, + 426, + 258 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f464526efacec3d8e3a64f967d67d4939157e57e04901401248b286c18d11759.jpg", + "image_caption": [ + "(c) RigNet" + ], + "image_footnote": [], + "bbox": [ + 428, + 90, + 571, + 258 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a1f283232dc99d1999c7033b1638338683da706908622868c1d3ff483f8b3e76.jpg", + "image_caption": [ + "(d) SigNet" + ], + "image_footnote": [], + "bbox": [ + 571, + 90, + 714, + 258 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/2655180d87c800e552ac88a50546c20277dba44f1befcc3b4f702de33425ca42.jpg", + "image_caption": [ + "(e) GT" + ], + "image_footnote": [], + "bbox": [ + 715, + 90, + 857, + 258 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/8bfff01cb3e7aac297369b169baabfb607cf9a47496a4543b3c4386f4b3ca229.jpg", + "image_caption": [ + "Figure 4. Visual comparison with SOTA methods on TOFDC dataset, including NLSPN [29], RigNet [48], and our SigNet.", + "(a) RGB" + ], + "image_footnote": [], + "bbox": [ + 138, + 305, + 282, + 434 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/fcf0d73704f99531662243d392dd0e0ec29df5431bdf265d2e110fd6a1ff2e7e.jpg", + "image_caption": [ + "(b)CSPN" + ], + "image_footnote": [], + "bbox": [ + 282, + 306, + 426, + 434 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c1719cb69f606f6d26cba0fb05765d7ca0ba26ee7b10e2ecfcd3c95dd06288b3.jpg", + "image_caption": [ + "(c) AGG-Net" + ], + "image_footnote": [], + "bbox": [ + 426, + 306, + 571, + 434 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/7ec5dd05e3aadc8c97a32bf2624a6465f77147c52b528cbbf3af93088595db63.jpg", + "image_caption": [ + "(d) SigNet" + ], + "image_footnote": [], + "bbox": [ + 571, + 306, + 714, + 434 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/0cf62c4d48a96b8ec265c3ad71ad81d45f673267f400c22e3e6c685196c80f35.jpg", + "image_caption": [ + "(e) GT" + ], + "image_footnote": [], + "bbox": [ + 715, + 306, + 857, + 434 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/35ad799d658fde72b175e426b03d3a612be886f8910aa84baeb3b930251f8405.jpg", + "image_caption": [ + "Figure 5. Visual comparison with SOTA approaches on DIML dataset, including CSPN [4], AGG-Net [2], and our SigNet.", + "RGB", + "Figure 6. Visual examples of filled coarse depth maps and degradation representations on the DIML dataset." + ], + "image_footnote": [], + "bbox": [ + 138, + 481, + 256, + 593 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/015771f92579e5009873de3f94b2cd7c44c8fbdbddade9c11e149e6c742f09e0.jpg", + "image_caption": [ + "Coarse depth" + ], + "image_footnote": [], + "bbox": [ + 258, + 481, + 375, + 593 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/7ab8df37afd80c775888b1080b3f672b48440efb273378472f51c967366030a3.jpg", + "image_caption": [ + "Degradation" + ], + "image_footnote": [], + "bbox": [ + 377, + 481, + 496, + 593 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ddfaa1a958b53be0a288ed96f85c9c1b3242c29b066e904f890b87245fcae813.jpg", + "image_caption": [ + "RGB" + ], + "image_footnote": [], + "bbox": [ + 500, + 481, + 617, + 593 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/afd2c3b49df17bf5ed6643dd14e73146d16f8e572bc8d9491ea2f73202135ca0.jpg", + "image_caption": [ + "Coarse depth" + ], + "image_footnote": [], + "bbox": [ + 619, + 481, + 736, + 593 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/6b65f37cd96e75aebc7d20ba2a1e2f01ba8e7a367b5d65d3dcf7e9fc3d47ebc3.jpg", + "image_caption": [ + "Degradation" + ], + "image_footnote": [], + "bbox": [ + 738, + 481, + 857, + 593 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Comparison with SOTA", + "text_level": 1, + "bbox": [ + 89, + 654, + 310, + 671 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To assess the effectiveness of our SigNet, we benchmark it against the traditional bilateral filtering method [1] and some recent SOTA deep learning approaches such as [2, 28, 29, 34, 38]. The error and accuracy results for the competing models are sourced from their original publications or obtained by retraining them following the default settings.", + "bbox": [ + 89, + 676, + 483, + 768 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1.1. Performance Analysis", + "text_level": 1, + "bbox": [ + 89, + 775, + 290, + 790 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "On NYUv2. As a widely used benchmark dataset, NYUv2 is often employed to evaluate the effectiveness of models. From the first row of Table 1, we can observe that SigNet achieves competitive performance on the NYUv2 dataset, with an RMSE that is $5mm$ lower than the second-best method. Additionally, although our method lags behind BPNet [38] by only 0.001 in the REL metric, SigNet signifi", + "bbox": [ + 89, + 794, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "cantly reduces the number of parameters by $96\\%$ .", + "bbox": [ + 511, + 655, + 841, + 671 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "On DIML. DIML is collected from both indoor and outdoor scenes, suffering from severe edge shadows and irregular holes. Our method is implemented using the same experimental settings as previous approaches [2, 33]. As shown in the second row of Table 1, SigNet demonstrates the best performance across all evaluation metrics. For example, compared to the suboptimal model, our approach decreases the RMSE by $19\\%$ and the REL by $45\\%$ .", + "bbox": [ + 511, + 672, + 906, + 792 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "On SUN RGBD. The large-scale SUN RGB-D dataset is leveraged to evaluate the generalization capability of our SigNet across different sensors and scenes. Table 1 (third row) reveals that our method surpasses previous state-of-the-art approaches across all metrics. Notably, SigNet achieves a $8mm$ lower RMSE than the latest BPNet [38] and a $36mm$ lower RMSE compared to AGG-Net [2].", + "bbox": [ + 511, + 794, + 906, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "26948", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/e78f307dde527af33a161b1487a26e624940b01c8b35bb439d305f79b817c6d3.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Density1%5%35%65%95%
TPVD [50]156.6139.7112.297.392.8
SigNet144.6121.599.490.184.3
Improv. ↑12.018.212.87.28.5
", + "bbox": [ + 106, + 88, + 467, + 162 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/fb3a9a02651eeca5f3b0e2aeabcc78be927334df8256767d18eedc3ec1a38bb2.jpg", + "image_caption": [ + "Figure 7. Complexity comparison on the TOFDC dataset, where larger circle areas indicate higher inference time." + ], + "image_footnote": [], + "bbox": [ + 102, + 191, + 472, + 406 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "On TOFDC. TOFDC encompasses extensive data captured under various lighting conditions and in open spaces. It is utilized to further assess the performance of the proposed model. The quantitative results in the last row of Table 1 indicate that our method achieves state-of-the-art performance. Compared to the second-best TPVD [50], SigNet reduces the RMSE by $5mm$ and the REL by 0.002, while also decreasing the number of parameters by $27.9M$ .", + "bbox": [ + 89, + 460, + 482, + 583 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.1.2. Visualization Analysis", + "text_level": 1, + "bbox": [ + 89, + 589, + 290, + 604 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Figs. 4 and 5 showcase visual results on TOFDC and DIML, highlighting SigNet's ability to recover precise depth structures. Compared to previous models, our method yields sharper predictions, such as the leaves in Fig. 4 and the chair in Fig. 5, achieving closer alignment with the ground truth. These results affirm the effectiveness of our approach in enhancing depth completion. Additionally, Fig. 6 presents some degradation examples, further demonstrating the ability of SigNet to model and address such challenges.", + "bbox": [ + 89, + 608, + 482, + 744 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.1.3. Number of Valid Points", + "text_level": 1, + "bbox": [ + 89, + 751, + 302, + 765 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Tab. 2 lists a comparison of different valid points on DIML. It indicates that our SigNet consistently outperforms TPVD, achieving an average improvement of $9.7\\%$ across the five densities. These results highlight the generalization capability of our method across varying valid points.", + "bbox": [ + 89, + 768, + 482, + 845 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.1.4. Complexity Analysis", + "text_level": 1, + "bbox": [ + 89, + 851, + 282, + 866 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Fig. 7 shows that our method achieves an excellent balance among RMSE, parameters, and inference time. For exam-", + "bbox": [ + 89, + 869, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/80c5b7a042ec265a149a53143f820c504173d2b1b1ca499bdfd78f0aaa85845b.jpg", + "table_caption": [ + "Table 2. Density comparison of TPVD and SigNet on TOFDC." + ], + "table_footnote": [], + "table_body": "
SigNetDeg. BridgeRGB-D FusionRMSE (mm)
DensifyDenoiseConc.Att.DADF-DDADF-F
i108
ii✓✓95
iii✓✓✓92
iv✓✓✓74
v✓✓✓68
v✓✓✓✓55
", + "bbox": [ + 517, + 88, + 903, + 200 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Ablation study of our SigNet on the DIML dataset. Conc.: concat, Att.: attention, DADF-D: degradation decomposition in DADF module, DADF-F: conditional Mamba fusion.", + "bbox": [ + 511, + 208, + 903, + 250 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ple, compared to the fewer parameter CSPN [4], our SigNet not only reduces parameters by $81\\%$ and inference time by $82\\%$ , but also decreases RMSE by $61\\%$ . Additionally, our method surpasses the latest TPVD [50] with a $5\\%$ lower RMSE and $78\\%$ lower FPS, while reducing parameters by $89\\%$ and inference time by $78\\%$ .", + "bbox": [ + 511, + 275, + 906, + 367 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Ablation Study", + "text_level": 1, + "bbox": [ + 511, + 377, + 666, + 393 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "SigNet Designs. Table 3 presents the ablation results on the DIML test split. The iteration of DADF is set to five. SigNet-i serves as the baseline, where both the degradation bridge and DADF module are removed, and the RGB-D features are fused using addition. Building on this, SigNet-ii redefines depth completion as depth enhancement through densification and degradation, concatenating the RGB and degradation features, resulting in a $13\\mathrm{mm}$ improvement. This highlights the superiority of our task transformation approach. Additionally, SigNet-iii incorporates a Gaussian denoising operator to further reduce error. To better utilize the degradation, SigNet-iv applies a Sigmoid function and element-wise multiplication to obtain an attention map, enabling adaptive selection of RGB features. Consequently, performance improves significantly from $92\\mathrm{mm}$ to $74\\mathrm{mm}$ . Furthermore, SigNet-v employs degradation-aware decomposition, reducing the error by an additional $6\\mathrm{mm}$ . The results of SigNet-iii, iv, and v demonstrate the effectiveness of our adaptive feature selection strategy. Notably, in SigNet-iii, iv, and v, the selected RGB and depth features are encoded through concatenation and convolutions. Finally, with the introduction of our conditional Mamba, SigNet-vi surpasses SigNet-v by $13\\mathrm{mm}$ in RMSE, representing a significant improvement. Overall, each component positively impacts the baseline.", + "bbox": [ + 511, + 398, + 906, + 777 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Mamba Fusion and Iteration of DADF. Fig. 8 (a) illustrates the ablation study of Mamba fusion on the DIML dataset. For rapid validation, we employ DADF only once. The baseline ViM directly encodes the combined RGB-D features using the vanilla Vision Mamba [25]. When we apply our Conditional Mamba with initial degradation and RGB-D features as input, the error decreases by $8\\mathrm{mm}$ . Additionally, the introduction of degradation decomposition", + "bbox": [ + 511, + 779, + 908, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "26949", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8ca0db0eb846cd175d65914b9fbdb356cc3241ac68fdaff946507be3ff0f360e.jpg", + "image_caption": [ + "Figure 8. Ablation study of our DADF on the DIML dataset." + ], + "image_footnote": [], + "bbox": [ + 99, + 88, + 473, + 222 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/61b61ea52479da242e2640ed726c906b7414b5b982e1050219904e4cc010cb48.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Path number357911
RMSE (mm)55.054.654.554.154.0
", + "bbox": [ + 104, + 258, + 467, + 306 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. Ablation on the number of decomposition path in DADF.", + "bbox": [ + 91, + 315, + 480, + 329 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "using DCT further enhances performance. We also observe that the results improve as the number of DCT components increases, highlighting the effectiveness of our gradual DCT decomposition design. Fig. 8 (b) presents the ablation study of our DADF with different iterations on the DIML dataset. We find that performance improves with an increasing number of iterations. For an optimal balance between efficiency and effectiveness, we set the final iteration count to five. Moreover, as shown in Fig.3, the degradation $\\mathbf{D}$ can be decomposed into $n$ components using DCT. As indicated in Tab.4, the error gradually decreases as $n$ increases. However, considering the computational complexity, we select the top-3 most representative components to focus on the high-frequency components of $\\mathbf{F}_{img}$ and update the Mamba condition accordingly.", + "bbox": [ + 89, + 356, + 483, + 583 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Failure Case", + "text_level": 1, + "bbox": [ + 89, + 590, + 225, + 606 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our method is primarily evaluated in indoor scenes and some simple outdoor scenes with closer distances. By constraining dense predictions with coarse depth maps, SigNet can effectively predict their degradation representations. However, as depicted in Fig. 9, SigNet fails to estimate reasonable degradation on the outdoor KITTI benchmark [40]. We identify two main reasons for this:", + "bbox": [ + 89, + 613, + 482, + 717 + ], + "page_idx": 7 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1) The ground truth depth maps of KITTI are very sparse, with only about $30\\%$ valid pixels. This sparsity leads to inconsistent depth edges, resulting in a relatively weak constraint. For instance, the degradation in Fig. 9 even introduces some initial LiDAR lines.", + "2) The depth distance of KITTI data is much larger, which significantly weakens edge details. As observed, the degradation in Fig. 9 can only learn very coarse and blurry representations near edges at short range, while at long range, it learns incorrect representations." + ], + "bbox": [ + 89, + 719, + 482, + 869 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "One potential solution is to provide the model with supervisory signals regarding dense contents and precise", + "bbox": [ + 89, + 869, + 482, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6e2213f0f81a15f97da16e501338e98fef2ad3febb47a980a081f121657ae439.jpg", + "image_caption": [ + "Figure 9. Visual results of failure cases on KITTI benchmark [40]. DA v2 denotes the large depth model DepthAnything v2 [52]." + ], + "image_footnote": [], + "bbox": [ + 514, + 88, + 903, + 388 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "edges. It is well-known that large depth models can produce dense depth with impressive sharp edges. For example, we employ DepthAnything v2 [52] to generate the corresponding dense depth. We discover that the edges of cars and even tree leaves are much clearer. Although the depth scale does not match that of KITTI, we can generate a map that offers constraints near edges via gradient or boundary extraction techniques [46]. Certainly, the results of DepthAnything v2 [52] or DepthAnything v2 itself can also be utilized to facilitate depth completion, warranting further exploration.", + "bbox": [ + 511, + 446, + 906, + 599 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 609, + 633, + 626 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we redefined depth completion as depth enhancement for the first time with our novel framework, SigNet. SigNet leveraged non-CNN densification tools to transform sparse depth data into coarse yet dense depth, fully eliminating the mismatch and ambiguity caused by direct convolution over irregularly sampled sparse data. More importantly, by establishing a self-supervised degradation bridge, SigNet effectively fused RGB-D data. It utilized degradation decomposition to dynamically select high-frequency components of RGB and build a conditional Mamba for the depth enhancement. Extensive experiments validated the superiority of SigNet, highlighting its potential for depth related applications.", + "bbox": [ + 511, + 636, + 906, + 833 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 511, + 845, + 681, + 861 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work was supported by the National Natural Science Fund of China (U24A20330, 62361166670 and 62072242).", + "bbox": [ + 511, + 869, + 905, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "26950", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Antoni Buades, Bartomeu Coll, and J-M Morel. A non-local algorithm for image denoising. In CVPR, pages 60-65, 2005. 5, 6", + "[2] Dongyue Chen, Tingxuan Huang, Zhimin Song, Shizhuo Deng, and Tong Jia. Agg-net: Attention guided gated convolutional network for depth image completion. In ICCV, pages 8853-8862, 2023. 5, 6, 2", + "[3] Yun Chen, Bin Yang, Ming Liang, and Raquel Urtasun. Learning joint 2d-3d representations for depth completion. In ICCV, pages 10023–10032, 2019. 2", + "[4] Xinjing Cheng, Peng Wang, and Ruigang Yang. Learning depth with convolutional spatial propagation network. In ECCV, pages 103-119, 2018. 5, 6, 7", + "[5] Xinjing Cheng, Peng Wang, and Ruigang Yang. Depth estimation via affinity learned with convolutional spatial propagation network. In ECCV, pages 103-119, 2018. 2", + "[6] Xinjing Cheng, Peng Wang, and Ruigang Yang. Learning depth with convolutional spatial propagation network. IEEE Transactions on Pattern Analysis and Machine Intelligence, 42(10):2361-2379, 2019. 2", + "[7] Xinjing Cheng, Peng Wang, Chenye Guan, and Ruigang Yang. Cspn++: Learning context and resource aware convolutional spatial propagation networks for depth completion. In AAAI, pages 10615-10622, 2020. 2", + "[8] Jaehoon Cho, Dongbo Min, Youngjung Kim, and Kwanghoon Sohn. A large rgb-d dataset for semi-supervised monocular depth estimation. arXiv preprint arXiv:1904.10230, 2019. 2", + "[9] Abdelrahman Eldisokey, Michael Felsberg, and Fahad Shahbaz Khan. Confidence propagation through cnns for guided sparse depth regression. IEEE Transactions on Pattern Analysis and Machine Intelligence, 42(10):2423-2436, 2020. 2", + "[10] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023. 4", + "[11] Mu Hu, Shuling Wang, Bin Li, Shiyu Ning, Li Fan, and Xiaojin Gong. Penet: Towards precise and efficient image guided depth completion. In ICRA, 2021. 2, 5", + "[12] Maximilian Jaritz, Raoul De Charette, Emilie Wirbel, Xavier Perrotton, and Fawzi Nashashibi. Sparse and dense data with cnns: Depth completion and semantic segmentation. In 3DV, pages 52-60, 2018. 2", + "[13] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In Computer Ence, 2014. 2", + "[14] Jason Ku, Ali Harakeh, and Steven L Waslander. In defense of classical image processing: Fast depth completion on the cpu. In CRV, pages 16-22, 2018. 1, 2, 3", + "[15] Anat Levin, Dani Lischinski, and Yair Weiss. Colorization using optimization. In SIGGRAPH, pages 689-694, 2004. 1, 2, 3", + "[16] Ang Li, Zejian Yuan, Yonggen Ling, Wanchao Chi, Chong Zhang, et al. A multi-scale guided cascade hourglass network for depth completion. In WACV, pages 32-40, 2020. 2" + ], + "bbox": [ + 93, + 114, + 482, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[17] Boyun Li, Xiao Liu, Peng Hu, Zhongqin Wu, Jiancheng Lv, and Xi Peng. All-in-one image restoration for unknown corruption. In CVPR, pages 17452-17462, 2022. 3", + "[18] Jie Liang, Hui Zeng, and Lei Zhang. Efficient and degradation-adaptive network for real-world image superresolution. In ECCV, pages 574-591, 2022. 3", + "[19] Yuankai Lin, Tao Cheng, Qi Zhong, Wending Zhou, and Hua Yang. Dynamic spatial propagation network for depth completion. In AAAI, pages 1638-1646, 2022. 2", + "[20] Yuankai Lin, Hua Yang, Tao Cheng, Wending Zhou, and Zhouping Yin. Dyspn: Learning dynamic affinity for image-guided depth completion. IEEE Transactions on Circuits and Systems for Video Technology, pages 1-1, 2023. 2", + "[21] Lina Liu, Yiyi Liao, Yue Wang, Andreas Geiger, and Yong Liu. Learning steering kernels for guided depth completion. IEEE Transactions on Image Processing, 30:2850-2861, 2021. 1", + "[22] Lina Liu, Xibin Song, Xiaoyang Lyu, Junwei Diao, Mengmeng Wang, Yong Liu, and Liangjun Zhang. Fcfr-net: Feature fusion based coarse-to-fine residual learning for depth completion. In AAAI, pages 2136-2144, 2021. 1, 2", + "[23] Sifei Liu, Shalini De Mello, Jinwei Gu, Guangyu Zhong, Ming-Hsuan Yang, and Jan Kautz. Learning affinity via spatial propagation networks. In NeurIPS, 2017. 2", + "[24] Xin Liu, Xiaofei Shao, Bo Wang, Yali Li, and Shengjin Wang. Graphcspn: Geometry-aware depth completion via dynamic gcns. In ECCV, pages 90-107. Springer, 2022. 2, 5", + "[25] Yue Liu, Yunjie Tian, Yuzhong Zhao, Hongtian Yu, Lingxi Xie, Yaowei Wang, Qixiang Ye, and Yunfan Liu. Vmamba: Visual state space model. arXiv preprint arXiv:2401.10166, 2024. 2, 4, 7", + "[26] Kaiyue Lu, Nick Barnes, Saeed Anwar, and Liang Zheng. From depth what can you see? depth completion via auxiliary image reconstruction. In CVPR, pages 11306-11315, 2020. 2", + "[27] Fangchang Ma and Sertac Karaman. Sparse-to-dense: Depth prediction from sparse depth samples and a single image. In ICRA, pages 4796-4803. IEEE, 2018. 1", + "[28] Fangchang Ma, Guilherme Venturelli Cavalheiro, and Sertac Karaman. Self-supervised sparse-to-dense: Self-supervised depth completion from lidar and monocular camera. In ICRA, 2019. 2, 5, 6", + "[29] Jinsun Park, Kyungdon Joo, Zhe Hu, Chi-Kuei Liu, and In So Kweon. Non-local spatial propagation network for depth completion. In ECCV, 2020. 2, 5, 6", + "[30] Jin-Hwi Park, Chanhwi Jeong, Junoh Lee, and Hae-Gon Jeon. Depth prompting for sensor-agnostic depth estimation. In CVPR, pages 9859–9869, 2024. 2", + "[31] Jiaxiong Qiu, Zhaopeng Cui, Yinda Zhang, Xingdi Zhang, Shuaicheng Liu, Bing Zeng, and Marc Pollefeys. Deep surface normal guided depth prediction for outdoor scene from sparse lidar data and single color image. In CVPR, pages 3313-3322, 2019. 5, 2", + "[32] Kyeongha Rho, Jinsung Ha, and Youngjung Kim. Guideformer: Transformers for image guided depth completion. In CVPR, pages 6250-6259, 2022. 2" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "26951", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[33] Dmitry Senushkin, Mikhail Romanov, Ilia Belikov, Nikolay Patakin, and Anton Konushin. Decoder modulation for indoor depth completion. In IROS, pages 2181-2188. IEEE, 2021. 5, 6", + "[34] Shreyas S Shivakumar, Ty Nguyen, Ian D Miller, Steven W Chen, Vijay Kumar, and Camillo J Taylor. Dfusenet: Deep fusion of rgb and sparse depth information for image guided dense depth completion. In ITSC, pages 13-20. IEEE, 2019. 5, 6", + "[35] Nathan Silberman, Derek Hoiem, Pushmeet Kohli, and Rob Fergus. Indoor segmentation and support inference from rgbd images. In ECCV, pages 746-760. Springer, 2012. 2", + "[36] Shuran Song, Samuel P Lichtenberg, and Jianxiong Xiao. Sun rgb-d: A rgb-d scene understanding benchmark suite. In CVPR, pages 567-576, 2015. 1, 2", + "[37] Jie Tang, Fei-Peng Tian, Wei Feng, Jian Li, and Ping Tan. Learning guided convolutional network for depth completion. IEEE Transactions on Image Processing, 30:1116-1129, 2020. 2, 5", + "[38] Jie Tang, Fei-Peng Tian, Boshi An, Jian Li, and Ping Tan. Bilateral propagation network for depth completion. In CVPR, pages 9763-9772, 2024. 1, 2, 5, 6", + "[39] Diego Thomas and Akihiro Sugimoto. A flexible scene representation for 3d reconstruction using an rgb-d camera. In ICCV, pages 2800-2807, 2013. 1", + "[40] Jonas Uhrig, Nick Schneider, Lukas Schneider, Uwe Franke, Thomas Brox, and Andreas Geiger. Sparsity invariant cnns. In 3DV, pages 11–20, 2017. 1, 2, 8", + "[41] Wouter Van Gansbeke, Davy Neven, Bert De Brabandere, and Luc Van Gool. Sparse and noisy lidar completion with rgb guidance and uncertainty. In MVA, pages 1-6, 2019. 2", + "[42] Haowen Wang, Mingyuan Wang, Zhengping Che, Zhiyuan Xu, Xiuquan Qiao, Mengshi Qi, Feifei Feng, and Jian Tang. RGB-depth fusion gan for indoor depth completion. In CVPR, pages 6209-6218, 2022. 5", + "[43] Longguang Wang, Yingqian Wang, Xiaoyu Dong, Qingyu Xu, Jungang Yang, Wei An, and Yulan Guo. Unsupervised degradation representation learning for blind superresolution. In CVPR, pages 10581-10590, 2021. 1, 2, 3", + "[44] Yufei Wang, Bo Li, Ge Zhang, Qi Liu, Tao Gao, and Yuchao Dai. Lrru: Long-short range recurrent updating networks for depth completion. In ICCV, pages 9422-9432, 2023. 2", + "[45] Yufei Wang, Ge Zhang, Shaoqian Wang, Bo Li, Qi Liu, Le Hui, and Yuchao Dai. Improving depth completion via depth feature upsampling. In CVPR, pages 21104-21113, 2024. 1, 2", + "[46] Zhengxue Wang, Zhiqiang Yan, and Jian Yang. Sgnet: Structure guided network via gradient-frequency awareness for depth map super-resolution. In AAAI, pages 5823-5831, 2024. 8", + "[47] Zhiqiang Yan, Xiang Li, Kun Wang, Zhenyu Zhang, Jun Li, and Jian Yang. Multi-modal masked pre-training for monocular panoramic depth completion. In ECCV, pages 378-395, 2022. 1", + "[48] Zhiqiang Yan, Kun Wang, Xiang Li, Zhenyu Zhang, Jun Li, and Jian Yang. Rignet: Repetitive image guided network for depth completion. In ECCV, pages 214-230, 2022. 1, 2, 5, 6" + ], + "bbox": [ + 91, + 90, + 480, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[49] Zhiqiang Yan, Kun Wang, Xiang Li, Zhenyu Zhang, Jun Li, and Jian Yang. Desnet: Decomposed scale-consistent network for unsupervised depth completion. In AAAI, pages 3109-3117, 2023. 1", + "[50] Zhiqiang Yan, Yuankai Lin, Kun Wang, Yupeng Zheng, Yufei Wang, Zhenyu Zhang, Jun Li, and Jian Yang. Tri-perspective view decomposition for geometry-aware depth completion. In CVPR, pages 4874-4884, 2024. 1, 2, 5, 7", + "[51] Zhiqiang Yan, Yupeng Zheng, Deng-Ping Fan, Xiang Li, Jun Li, and Jian Yang. Learnable differencing center for nighttime depth perception. Visual Intelligence, 2(1):15, 2024. 1", + "[52] Lihe Yang, Bingyi Kang, Zilong Huang, Zhen Zhao, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything v2. arXiv preprint arXiv:2406.09414, 2024. 8", + "[53] Yanchao Yang, Alex Wong, and Stefano Soatto. Dense depth posterior (ddp) from single image and sparse range. In CVPR, pages 3353-3362, 2020. 2", + "[54] Mingde Yao, Ruikang Xu, Yuanshen Guan, Jie Huang, and Zhiwei Xiong. Neural degradation representation learning for all-in-one image restoration. IEEE Transactions on Image Processing, 2024. 3", + "[55] Guanghao Yin, Wei Wang, Zehuan Yuan, Wei Ji, Dongdong Yu, Shouqian Sun, Tat-Seng Chua, and Changhu Wang. Conditional hyper-network for blind super-resolution with multiple degradations. IEEE Transactions on Image Processing, 31:3949-3960, 2022. 2, 3", + "[56] Zhu Yu, Zehua Sheng, Zili Zhou, Lun Luo, Si-Yuan Cao, Hong Gu, Huaqi Zhang, and Hui-Liang Shen. Aggregating feature point cloud for depth completion. In ICCV, pages 8732-8743, 2023. 2, 5", + "[57] Cheng Zhang, Yu Zhu, Qingsen Yan, Jinqiu Sun, and Yanning Zhang. All-in-one multi-degradation image restoration network via hierarchical degradation representation. In ACM MM, pages 2285-2293, 2023. 3", + "[58] Jinghao Zhang, Jie Huang, Mingde Yao, Zizheng Yang, Hu Yu, Man Zhou, and Feng Zhao. Ingredient-oriented multi-degradation learning for image restoration. In CVPR, pages 5825-5835, 2023. 3", + "[59] Kai Zhang, Jingyun Liang, Luc Van Gool, and Radu Timofte. Designing a practical degradation model for deep blind image super-resolution. In ICCV, pages 4791-4800, 2021. 1, 2, 3", + "[60] Yulun Zhang, Kunpeng Li, Kai Li, Lichen Wang, Bineng Zhong, and Yun Fu. Image super-resolution using very deep residual channel attention networks. In ECCV, pages 286-301, 2018. 3", + "[61] Youmin Zhang, Xianda Guo, Matteo Poggi, Zheng Zhu, Guan Huang, and Stefano Mattoccia. Completionformer: Depth completion with convolutions and vision transformers. In CVPR, pages 18527-18536, 2023. 2, 5", + "[62] Zhiwei Zhong, Xianming Liu, Junjun Jiang, Debin Zhao, and Xiangyang Ji. Guided depth map super-resolution: A survey. ACM Computing Surveys, 55(14s):1-36, 2023. 1, 2", + "[63] Hanyu Zhou, Yi Chang, Wending Yan, and Luxin Yan. Unsupervised cumulative domain adaptation for foggy scene optical flow. In CVPR, pages 9569-9578, 2023. 1" + ], + "bbox": [ + 514, + 90, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "26952", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[64] Hongyang Zhou, Xiaobin Zhu, Jianqing Zhu, Zheng Han, Shi-Xue Zhang, Jingyan Qin, and Xu-Cheng Yin. Learning correction filter via degradation-adaptive regression for blind single image super-resolution. In ICCV, pages 12365-12375, 2023. 3", + "[65] Hanyu Zhou, Yi Chang, Zhiwei Shi, Wending Yan, Gang Chen, Yonghong Tian, and Luxin Yan. Adverse weather optical flow: Cumulative homogeneous-heterogeneous adaptation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 1", + "[66] Wending Zhou, Xu Yan, Yinghong Liao, Yuankai Lin, Jin Huang, Gangming Zhao, Shuguang Cui, and Zhen Li. Bev@ dc: Bird's-eye view assisted training for depth completion. In CVPR, pages 9233-9242, 2023. 2", + "[67] Ziyue Zhu, Qiang Meng, Xiao Wang, Ke Wang, Liujiang Yan, and Jian Yang. Curricular object manipulation in lidar-based object detection. In CVPR, pages 1125-1135, 2023. 1", + "[68] Ziyue Zhu, Shenlong Wang, Jin Xie, Jiang-Jiang Liu, Jingdong Wang, and Jian Yang. Voxelsplat: Dynamic gaussian splatting as an effective loss for occupancy and flow prediction. In CVPR, 2025. 1" + ], + "bbox": [ + 91, + 92, + 480, + 401 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "26953", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Completion as Enhancement_ A Degradation-Aware Selective Image Guided Network for Depth Completion/5b8a9777-b777-4fb5-9ba2-85ebeb5a648a_model.json b/2025/Completion as Enhancement_ A Degradation-Aware Selective Image Guided Network for Depth Completion/5b8a9777-b777-4fb5-9ba2-85ebeb5a648a_model.json new file mode 100644 index 0000000000000000000000000000000000000000..e7a2cb5adb384e7cd21507a0097e5c5d555ca260 --- /dev/null +++ b/2025/Completion as Enhancement_ A Degradation-Aware Selective Image Guided Network for Depth Completion/5b8a9777-b777-4fb5-9ba2-85ebeb5a648a_model.json @@ -0,0 +1,2818 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.107, + 0.131, + 0.892, + 0.177 + ], + "angle": 0, + "content": "Completion as Enhancement: A Degradation-Aware Selective Image Guided Network for Depth Completion" + }, + { + "type": "text", + "bbox": [ + 0.224, + 0.204, + 0.78, + 0.223 + ], + "angle": 0, + "content": "Zhiqiang Yan Zhengxue Wang Kun Wang Jun Li\\* Jian Yang\\*" + }, + { + "type": "text", + "bbox": [ + 0.269, + 0.223, + 0.729, + 0.24 + ], + "angle": 0, + "content": "PCA Lab†, Nanjing University of Science and Technology" + }, + { + "type": "text", + "bbox": [ + 0.28, + 0.242, + 0.714, + 0.257 + ], + "angle": 0, + "content": "{yanzq,zxwang,kunwang,junli,csjyang}@njust.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.292, + 0.327, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.323, + 0.486, + 0.642 + ], + "angle": 0, + "content": "In this paper, we introduce the Selective Image Guided Network (SigNet), a novel degradation-aware framework that transforms depth completion into depth enhancement for the first time. Moving beyond direct completion using convolutional neural networks (CNNs), SigNet initially densifies sparse depth data through non-CNN densification tools to obtain coarse yet dense depth. This approach eliminates the mismatch and ambiguity caused by direct convolution over irregularly sampled sparse data. Subsequently, SigNet redefines completion as enhancement, establishing a self-supervised degradation bridge between the coarse depth and the targeted dense depth for effective RGB-D fusion. To achieve this, SigNet leverages the implicit degradation to adaptively select high-frequency components (e.g., edges) of RGB data to compensate for the coarse depth. This degradation is further integrated into a multi-modal conditional Mamba, dynamically generating the state parameters to enable efficient global high-frequency information interaction. We conduct extensive experiments on the NYUv2, DIML, SUN RGBD, and TOFDC datasets, demonstrating the state-of-the-art (SOTA) performance of SigNet." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.667, + 0.222, + 0.683 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.693, + 0.483, + 0.783 + ], + "angle": 0, + "content": "Depth completion [27, 38, 48, 50] aims to predict a dense depth map from sparse and noisy depth measurements, often using the corresponding RGB image for auxiliary guidance. The resulting dense depth map can benefit various downstream applications such as scene reconstruction [36, 39, 47, 68] and self-driving [22, 45, 51, 63, 65, 67]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.784, + 0.484, + 0.845 + ], + "angle": 0, + "content": "The key challenges of this task lie in the sparse depth input data [21, 38, 40, 49, 50]. We summarize these challenges as mismatch and ambiguity. As shown in Fig. 1, the mismatch is caused by the irregularly sampled sparse" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.288, + 0.907, + 0.527 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.536, + 0.907, + 0.607 + ], + "angle": 0, + "content": "Figure 1. Illustration of our main concept. It redefines depth completion as depth enhancement. We first transform sparse depth input via non-CNN densification tools [14, 15], yielding coarse but dense depth. Then we predict the precise and dense depth from the coarse depth by leveraging degradation assumption [43, 59, 62]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.64, + 0.907, + 0.716 + ], + "angle": 0, + "content": "depth with varying point distributions, while convolutions are spatially invariant with fixed kernel shapes. Moreover, it is ambiguous for convolutions to distinguish between valid and invalid depth measurements since unknown pixels in the sparse depth are recognized as zero." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.909, + 0.902 + ], + "angle": 0, + "content": "To address these challenges, several excellent methods [21, 38, 40] have been proposed. For instance, SparseCOnvs [40] proposes sparsity invariant CNNs that explicitly model the location of missing data. To avoid directly convolving on sparse depth, Liu et al. [21] employs kernels learned by differentiable kernel regression layers to interpolate sparse data. BPNet [38] develops bilateral propagation, where the coefficients are generated based on both radiometric difference and spatial distance. These data-driven schemes are robust and impressive. However, they still cannot fully avoid the application of convolution on incomplete depth data, even though the depth data has become denser." + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.852, + 0.241, + 0.864 + ], + "angle": 0, + "content": "*Corresponding authors" + }, + { + "type": "page_footnote", + "bbox": [ + 0.093, + 0.864, + 0.483, + 0.901 + ], + "angle": 0, + "content": "\\(\\dagger\\)PCA Lab, Key Lab of Intelligent Perception and Systems for High-Dimensional Information of Ministry of Education, School of Computer Science and Engineering, Nanjing University of Science and Technology." + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.852, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26943" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.152 + ], + "angle": 0, + "content": "In this paper, our objective is to develop a novel solution that not only completely resolves the issues of mismatch and ambiguity but also provides a fresh perspective on revisiting the traditional depth completion task." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.153, + 0.482, + 0.288 + ], + "angle": 0, + "content": "To this end, we propose a novel framework called the Selective Image Guided Network (SigNet). As shown in Fig. 1, SigNet redefines depth completion as depth enhancement1. It first leverages non-CNN densification tools [14, 15] to obtain coarse but dense depth. This coarse depth is then enhanced by a CNN-based subnetwork. Inspired by previous approaches [59, 62], we assume that there exists unknown degradation \\( \\mathbf{h} \\) between the coarse depth \\( \\mathbf{Z} \\) and the target depth \\( \\mathbf{Y} \\), described as" + }, + { + "type": "equation", + "bbox": [ + 0.237, + 0.295, + 0.482, + 0.31 + ], + "angle": 0, + "content": "\\[\n\\mathbf {Z} = \\mathbf {h} \\mathbf {Y} + \\mathbf {n}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.317, + 0.483, + 0.633 + ], + "angle": 0, + "content": "where \\( n \\) denotes additive noise. Therefore, we can establish a connection between the source depth and the target depth through the degradation. A fundamental observation is that degradation typically occurs near edges, suggesting that this phenomenon can be leveraged to enhance RGB-D fusion. This is particularly relevant given that the coarse depth often lacks significant edge details, while the RGB image contains abundant high-frequency components. Specifically, SigNet introduces a Degradation-Aware Decomposition and Fusion (DADF) module. DADF begins by decomposing the implicit degradation \\( h \\) into multiple patches and estimating their coefficients in the discrete cosine domain. Each patch is then used to adaptively select the most representative high-frequency information from the RGB image. To achieve effective RGB-D aggregation, DADF incorporates a conditional Mamba to encode long-range correlations. Unlike the vanilla Vision Mamba [25], which calculates state parameters from the input, DADF uses the degradation patches as additional conditions to generate the state parameters. This approach enables further global high-frequency information interaction." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.635, + 0.411, + 0.648 + ], + "angle": 0, + "content": "In summary, our contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.649, + 0.483, + 0.739 + ], + "angle": 0, + "content": "- We introduce a novel framework called SigNet, which redefines depth completion as depth enhancement for the first time. By adopting a degradation-aware approach, SigNet not only offers a new perspective to the field but also fully eliminates the traditional challenges such as mismatch and ambiguity." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.74, + 0.483, + 0.845 + ], + "angle": 0, + "content": "- We propose a Degradation-Aware Decomposition and Fusion (DADF) module. Utilizing degradation decomposition, DADF adaptively selects high-frequency components of RGB data to compensate for coarse depth. Furthermore, DADF achieves effective RGB-D fusion by incorporating conditional Mamba, where the state parameters are generated from degradation patches." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.649, + 0.483, + 0.845 + ], + "angle": 0, + "content": null + }, + { + "type": "page_footnote", + "bbox": [ + 0.09, + 0.851, + 0.483, + 0.901 + ], + "angle": 0, + "content": "1Specifically depth super-resolution (DSR). Mainstream DSR methods restore target depth directly from upsampled depth. Consequently, the processes and the end goals of depth completion and DSR are identical, i.e., predicting clear depth from the same-resolution coarse depth." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.092, + 0.905, + 0.122 + ], + "angle": 0, + "content": "- We evaluate SigNet through comprehensive experiments, demonstrating its superiority over previous SOTAs." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.134, + 0.655, + 0.15 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.16, + 0.905, + 0.189 + ], + "angle": 0, + "content": "Depth Completion. Depth completion can be broadly categorized into depth-only and image-guided categories." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.19, + 0.905, + 0.341 + ], + "angle": 0, + "content": "Depth-only methods [9, 12, 14, 28, 40, 41, 44, 50] obtain dense depth directly from a single sparse depth input. For example, IP-Basic [14] designs reasonable combinations of basic image processing operations to densify sparse depth without using CNNs. Ma et al. [28] employ a UNet-like architecture to progressively densify the sparse depth input. FusionNet [41] designs a lightweight completion network that aggregates both global and local information with confidence adjustment. Moreover, Lu et al. [26] introduce a color image branch as auxiliary supervision during training." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.342, + 0.905, + 0.582 + ], + "angle": 0, + "content": "Despite advancements in single-modal methods, their performance is often limited without supplementary reference information. To address this challenge, numerous image-guided techniques have been proposed [11, 16, 19, 22, 37, 38, 45, 50, 53, 61, 66]. For example, GuideNet [37] and RigNet [48] utilize color images to generate adaptive filtering kernels for depth feature extraction. GFormer [32] and CFormer [61] integrate convolution and transformer mechanisms to capture both local and global representations. LRRU [44] introduces a dynamic kernel scope that transitions from large to small to capture varying dependencies. FuseNet [3] and PointDC [56] incorporate LiDAR point cloud branches to model 3D geometries. BEV@DC [66] uses bird's-eye view to build a point-voxel architecture. Additionally, DepthPrompting [30] embeds priors of large depth models for better sensor-agnostic generalization." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.583, + 0.905, + 0.824 + ], + "angle": 0, + "content": "To further advance completion models, spatial propagation networks (SPNs) have been introduced. The initial SPN [23] is designed to learn a pairwise similarity matrix. CSPN [5] applies this technique to depth completion by conducting recursive convolutions with fixed local neighborhood kernels, while \\(\\mathrm{CSPN} + +\\) [7] adapts by learning variable kernel sizes. PENet [11] expands the receptive fields using dilated convolutions. Furthermore, NLSPN [29] integrates non-local neighbors via deformable convolutions. DySPN [20] generates dynamic non-linear neighbors using an attention mechanism. Most recently, 3D SPNs [6] have been developed to encode 3D geometry. For example, GraphC-SPN [24] employs additional geometric constraints to regulate 3D propagation. TPVD [50] introduces a geometric SPN that constructs the affinity in 2D tri-perspective view spaces and their combined 3D projection space." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Degradation Learning. Recently, color image restoration has been significantly advanced by improvements in degradation representations [43, 55, 59]. For example, Zhang et al. [59] construct a sophisticated yet practical degradation model to simulate real-world image degradation for" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26944" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.156, + 0.089, + 0.844, + 0.381 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.387, + 0.908, + 0.43 + ], + "angle": 0, + "content": "Figure 2. Pipeline of SigNet. The sparse depth data is initially filled to create a coarse depth map. We then utilize the degradation assumption in Eq. (1) to establish a connection between this coarse map and the target prediction. The color image features and degradation representations are employed to guide the multi-modal fusion through our proposed DADF module, as depicted in Fig. 3." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.451, + 0.486, + 0.724 + ], + "angle": 0, + "content": "blind image super-resolution. Similarly, Zhou et al. [64] propose to estimate degradation for each local image region using a degradation-adaptive regression filter for blind super-resolution. Liang et al. [18] present a degradation-adaptive model that specifies the network parameters by estimating the degradation. Furthermore, several image restoration methods [17, 57, 58] have been proposed to simultaneously address multiple forms of degradation. For example, Zhang et al. [57] develop a multi-degradation image restoration network that incrementally learns degradation representations via clustering, without relying on prior knowledge of the degradation information. Zhang et al. [58] build an ingredient-oriented degradation reformulation network to model the relationship among different images. Additionally, Yao et al. [54] introduce a neural degradation representation learning method that captures latent degradation features, adaptively decomposing various types of degradation to enable comprehensive restoration." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.739, + 0.184, + 0.755 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.485, + 0.902 + ], + "angle": 0, + "content": "Overview. Fig. 2 shows our SigNet pipeline that consists of the depth branch, degradation branch, and image branch. In the depth branch, the sparse depth \\(\\mathbf{S}\\) is first filled via nonCNN densification tools [14, 15], obtaining a coarse depth map \\(\\mathbf{Z}\\). According to Eq. (1), we employ denoising operators to weaken the negative impact of the noise \\(\\mathbf{n}\\), yielding \\(\\hat{\\mathbf{Z}}\\). Then, we use residual groups [60] to extract features of the denoised depth \\(\\hat{\\mathbf{Z}}\\) and the color image \\(\\mathbf{I}\\) in the image branch. In the degradation branch, the implicit degradation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.451, + 0.907, + 0.572 + ], + "angle": 0, + "content": "\\(\\mathbf{D}\\) is generated from a degradation network, where convolutions and residual groups are used. We thus utilize the image feature and degradation \\(\\mathbf{D}\\) to facilitate RGB-D fusion via our proposed Degradation-Aware Decomposition and Fusion (DADF). In summary, our SigNet redefines the depth completion as depth enhancement through the degradation assumption in Eq. (1), making the model free from the challenges of mismatch and ambiguity." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.579, + 0.669, + 0.594 + ], + "angle": 0, + "content": "3.1. Task Switching" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.6, + 0.907, + 0.812 + ], + "angle": 0, + "content": "Degradation learning is widely applied in low-level image restoration [43, 55, 59], yielding impressive results. These tasks exhibit two key characteristics: 1) both input and target data are typically dense, and 2) the input data generally have a lower resolution than the target data. Actually, the low-resolution input data are typically upsampled using interpolation methods, such as bicubic interpolation. Consequently, the degradation matrix usually consists of a convolution whose kernel is derived from degradation representations and a downsampling function. In contrast, the depth completion task takes sparse depth data as input, and the resolutions of the input and target data are the same. Thus, we reframe depth completion as depth enhancement by drawing an analogy to image restoration." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.812, + 0.907, + 0.858 + ], + "angle": 0, + "content": "As shown in Fig. 2, we first fill the sparse depth \\(\\mathbf{S}\\) using non-CNN densification tools [14, 15], denoted as \\(f_{ds}(\\cdot)\\), resulting in a coarse yet dense depth \\(\\mathbf{Z}\\), formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.666, + 0.864, + 0.905, + 0.88 + ], + "angle": 0, + "content": "\\[\n\\mathbf {Z} = f _ {d s} (\\mathbf {S}). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.886, + 0.905, + 0.902 + ], + "angle": 0, + "content": "As discussed in Sec. 1, this approach eliminates the mis" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26945" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.087, + 0.485, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.274, + 0.483, + 0.304 + ], + "angle": 0, + "content": "Figure 3. Overview of our DADF. \\(\\odot\\) and CM refer to elementwise multiplication and the conditional Mamba, respectively." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.328, + 0.484, + 0.435 + ], + "angle": 0, + "content": "match and ambiguity caused by direct convolution over irregularly sampled sparse data. Subsequently, we employ a degradation network to estimate the implicit degradation \\(\\mathbf{D}\\). The degradation kernel \\(\\mathbf{K}\\) is produced from \\(\\mathbf{D}\\) using convolutions. Consequently, given the coarse depth \\(\\mathbf{Z}\\) and target depth \\(\\mathbf{Y}\\), Eq. (1) can be rewritten based on the task switching as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.203, + 0.444, + 0.483, + 0.461 + ], + "angle": 0, + "content": "\\[\n\\mathbf {Z} = \\mathbf {K} (\\mathbf {D}; \\vartheta) \\otimes \\mathbf {Y} + \\mathbf {n}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.471, + 0.484, + 0.517 + ], + "angle": 0, + "content": "where \\(\\otimes\\) indicates convolution operation, \\(\\vartheta\\) is the network parameter. Since the resolutions are the same, our degradation equation does not need the downsampling function." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.525, + 0.484, + 0.541 + ], + "angle": 0, + "content": "3.2. Degradation-Aware Decomposition and Fusion" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.547, + 0.484, + 0.654 + ], + "angle": 0, + "content": "As we know, degradation typically occurs near edges, suggesting that this cue can be leveraged to enhance RGB-D fusion, especially considering that the coarse depth often lacks significant edge details, while the RGB image contains abundant high-frequency components. Thus, we propose DADF that is shown in Fig. 3. We denote the image and depth features as \\(\\mathbf{F}_{img}\\) and \\(\\mathbf{F}_{dep}\\), obtaining:" + }, + { + "type": "equation", + "bbox": [ + 0.217, + 0.663, + 0.483, + 0.68 + ], + "angle": 0, + "content": "\\[\n\\mathbf {F} _ {i m g} = f _ {e} (\\mathbf {I}), \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.218, + 0.682, + 0.483, + 0.7 + ], + "angle": 0, + "content": "\\[\n\\mathbf {F} _ {d e p} = f _ {e} \\left(f _ {d n} (\\mathbf {Z})\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.709, + 0.484, + 0.77 + ], + "angle": 0, + "content": "where \\( f_{e}(\\cdot) \\) and \\( f_{dn}(\\cdot) \\) refer to the feature extraction network and denoising function, respectively. As presented in Fig. 3, DADF takes as input the image feature \\( \\mathbf{F}_{img} \\), depth feature \\( \\mathbf{F}_{dep} \\), and degradation \\( \\mathbf{F}_D \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.77, + 0.484, + 0.875 + ], + "angle": 0, + "content": "Decomposition. To extract the most representative components of \\(\\mathbf{F}_D\\), we first apply the discrete cosine transform (DCT) to convert it into the frequency domain. Next, it adaptively learns a spectrum mask to enhance informative frequency bands while suppressing irrelevant ones. Finally, we reconstruct the spatial representation using the inverse DCT. This process is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.192, + 0.885, + 0.483, + 0.903 + ], + "angle": 0, + "content": "\\[\n\\left\\{\\mathbf {D} _ {i} \\mid i = 1, 2, 3 \\right\\} = f _ {d c t} (\\mathbf {D}). \\tag {6}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.524, + 0.095, + 0.78, + 0.11 + ], + "angle": 0, + "content": "Algorithm 1: Conditional SSM Block" + }, + { + "type": "text", + "bbox": [ + 0.535, + 0.113, + 0.793, + 0.127 + ], + "angle": 0, + "content": "Input: \\(\\mathbf{F}_{rgbd}\\), \\(\\{\\mathbf{D}_i|i = 1,2,3\\} :(\\mathrm{HW},\\mathrm{C})\\)" + }, + { + "type": "text", + "bbox": [ + 0.535, + 0.127, + 0.697, + 0.141 + ], + "angle": 0, + "content": "Output: \\(\\hat{\\mathbf{F}}_{rgbd}:(\\mathrm{HW},\\mathrm{C})\\)" + }, + { + "type": "text", + "bbox": [ + 0.522, + 0.142, + 0.716, + 0.154 + ], + "angle": 0, + "content": "1 A : (C, N) \\(\\leftarrow\\) ParameterA" + }, + { + "type": "text", + "bbox": [ + 0.522, + 0.155, + 0.835, + 0.182 + ], + "angle": 0, + "content": "2 \\(\\mathbf{B}:(\\mathrm{HW},\\mathrm{N})\\gets \\mathbf{Linear}_B(f_{gru}(\\mathbf{F}_{rgbd},\\{\\mathbf{D}_i|i = 1,2,3\\}))\\)" + }, + { + "type": "text", + "bbox": [ + 0.522, + 0.182, + 0.835, + 0.21 + ], + "angle": 0, + "content": "3 \\(\\mathbf{C}:(\\mathrm{HW},\\mathrm{N})\\gets \\mathbf{Linear}_C(f_{gru}(\\mathbf{F}_{rgbd},\\{\\mathbf{D}_i|i = 1,2,3\\}))\\)" + }, + { + "type": "text", + "bbox": [ + 0.522, + 0.21, + 0.819, + 0.252 + ], + "angle": 0, + "content": "4 \\(\\Delta\\) : (HW, C) \\(\\leftarrow\\) \\(\\log(1 + \\exp(\\text{Linear}_\\Delta(f_{gru}(\\mathbf{F}_{rgbd}, \\{\\mathbf{D}_i | i = 1, 2, 3\\})) + \\text{Parameter}_\\Delta))\\)" + }, + { + "type": "text", + "bbox": [ + 0.522, + 0.253, + 0.742, + 0.268 + ], + "angle": 0, + "content": "5 \\(\\overline{\\mathbf{A}}: (\\mathrm{HW}, \\mathrm{C}, \\mathrm{N}) \\leftarrow \\exp(\\Delta \\otimes \\mathbf{A})\\)" + }, + { + "type": "text", + "bbox": [ + 0.522, + 0.268, + 0.706, + 0.282 + ], + "angle": 0, + "content": "6 \\(\\overline{\\mathbf{B}}:\\mathrm{(HW,C,N)}\\gets \\pmb {\\Delta}\\otimes \\mathbf{B}\\)" + }, + { + "type": "text", + "bbox": [ + 0.522, + 0.284, + 0.737, + 0.3 + ], + "angle": 0, + "content": "7 \\(\\hat{\\mathbf{F}}_{rgbd} \\gets \\mathrm{SSM}(\\overline{\\mathbf{A}},\\overline{\\mathbf{B}},\\mathbf{C})(\\mathbf{F}_{rgbd})\\)" + }, + { + "type": "text", + "bbox": [ + 0.522, + 0.3, + 0.617, + 0.316 + ], + "angle": 0, + "content": "8 return \\(\\hat{\\mathbf{F}}_{rgbd}\\)" + }, + { + "type": "list", + "bbox": [ + 0.522, + 0.142, + 0.835, + 0.316 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.35, + 0.907, + 0.442 + ], + "angle": 0, + "content": "We apply concatenation, convolution, pooling, and Sigmoid operations, collectively denoted as \\( f_{m}(\\cdot) \\), to generate a mask \\( m \\). This mask is then multiplied with the combined frequency components to produce an updated mask \\( \\hat{m} \\), enabling the adaptive selection of the representative image feature \\( \\hat{\\mathbf{F}}_{img} \\). The process is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.594, + 0.454, + 0.905, + 0.47 + ], + "angle": 0, + "content": "\\[\nm = f _ {m} \\left(\\mathbf {F} _ {\\text {i m g}}, \\mathbf {D}\\right), \\tag {7}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.595, + 0.474, + 0.905, + 0.5 + ], + "angle": 0, + "content": "\\[\n\\hat {m} = f _ {i d c t} \\left(m \\cdot \\left(f _ {a c} \\left(\\sum_ {i = 1} ^ {3} \\mathbf {D} _ {i}\\right)\\right)\\right), \\tag {8}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.595, + 0.502, + 0.905, + 0.521 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {F}} _ {i m g} = \\hat {m} \\cdot \\mathbf {F} _ {i m g}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.533, + 0.905, + 0.562 + ], + "angle": 0, + "content": "where \\( f_{idct} \\) refers to inverse DCT and \\( f_{ac}(\\cdot) \\) represents the combination function of addition and convolutions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.563, + 0.907, + 0.685 + ], + "angle": 0, + "content": "Fusion. As a result, we further leverage the decomposed degradation \\(\\mathbf{D}_i\\) to enable effective RGB-D fusion. It is well-known that Mamba [10] is good at handling long-range correlations. Inspired by recent Vision Mamba [25], we present a variant, i.e., Conditional Mamba (CM). As shown in Fig. 3, the selected image feature \\(\\hat{\\mathbf{F}}_{img}\\) and raw depth feature \\(\\mathbf{F}_{dep}\\) are concatenated into CM, where the decomposed degradation \\(\\mathbf{D}_i\\) serves as the condition:" + }, + { + "type": "equation", + "bbox": [ + 0.578, + 0.696, + 0.905, + 0.714 + ], + "angle": 0, + "content": "\\[\n\\mathbf {F} _ {r g b d} = f _ {c c} \\left(\\tilde {\\mathbf {F}} _ {i m g}, \\mathbf {F} _ {d e p}\\right), \\tag {10}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.579, + 0.718, + 0.905, + 0.735 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {F}} _ {r g b d} = f _ {c m} \\left(\\mathbf {F} _ {r g b d}, \\left\\{\\mathbf {D} _ {i} \\mid i = 1, 2, 3 \\right\\}\\right), \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.748, + 0.907, + 0.823 + ], + "angle": 0, + "content": "where \\( f_{cc} \\) denotes the combination function of concatenation and convolutions, \\( f_{cm} \\) is the Condition Mamba function. As reported in Alg. 1, the state parameters \\( \\mathbf{B} \\) and \\( \\mathbf{C} \\) are calculated based on both the RGB-D input and degradation using gated recurrent convolution units." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.833, + 0.66, + 0.848 + ], + "angle": 0, + "content": "3.3. Loss Function" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.907, + 0.902 + ], + "angle": 0, + "content": "As shown in our pipeline, the total loss function \\(\\mathcal{L}_t\\) consists of a reconstruction loss \\(\\mathcal{L}_r\\) and a self-supervised degradation learning loss \\(\\mathcal{L}_d\\). Let \\(\\hat{\\mathbf{Y}}\\) and \\(\\mathbf{Y}\\) denote the ground truth" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "26946" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.126, + 0.088, + 0.872, + 0.589 + ], + "angle": 0, + "content": "
DatasetMethodParams. (M) ↓RMSE (mm) ↓REL ↓δ1(%) ↑δ2(%) ↑δ3(%) ↑Venue
NYUv2Bilateral [1]-5320.13285.193.595.9CVPR 2005
S2D [28]26.12300.05494.597.398.9ICRA 2019
CSPN [4]17.41730.02096.398.699.5ECCV 2018
DfuseNet [34]-1560.01698.899.799.9ITSC 2019
DM-LRN [33]-2050.01498.899.699.9IROS 2021
RDF-GAN [42]-1390.01398.799.699.9CVPR 2022
AGG-Net [2]129.1920.01499.499.9100.0ICCV 2023
BPNet [38]89.9880.01199.699.9100.0CVPR 2024
SigNet (ours)3.3830.01299.799.9100.0CVPR 2025
DIMLBilateral [1]-6360.18983.088.892.4CVPR 2005
CSPN [4]17.41620.03396.198.799.6ECCV 2018
DfuseNet [34]-1430.02398.499.499.9ITSC 2019
DM-LRN [33]-1490.01599.099.699.9IROS 2021
AGG-Net [2]129.1780.01199.699.9100.0ICCV 2023
BPNet [38]89.9680.01199.499.8100.0CVPR 2024
SigNet (ours)3.3550.00699.799.9100.0CVPR 2025
SUN RGBDS2D [28]26.13290.07493.997.098.1ICRA 2019
CSPN [4]17.42950.13795.697.598.4ECCV 2018
DLiDAR [31]53.42790.06196.998.098.4CVPR 2019
DM-LRN [33]-2670.06397.698.298.7IROS 2021
RDF-GAN [42]-2550.05996.998.499.0CVPR 2022
AGG-Net [2]129.11520.03898.599.099.4ICCV 2023
BPNet [38]89.91240.02598.999.499.8CVPR 2024
SigNet (ours)3.31160.02299.099.499.8CVPR 2025
TOFDCCSPN [4]17.42240.04294.595.396.5ECCV 2018
GuideNet [37]73.51460.03097.698.999.5TIP 2020
PENet [11]131.52410.04394.695.395.5ICRA 2021
NLSPN [29]25.81740.02996.497.998.9ECCV 2020
CFounder [61]83.51130.02999.199.699.9CVPR 2023
RigNet [48]65.21330.02597.699.199.7ECCV 2022
GraphCSPN [24]26.42530.05292.096.998.7ECCV 2022
PointDC [56]25.11090.02198.599.299.6ICCV 2023
TPVD [50]31.2920.01499.199.699.9CVPR 2024
SigNet (ours)3.3870.01299.299.799.9CVPR 2025
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.594, + 0.908, + 0.622 + ], + "angle": 0, + "content": "Table 1. Quantitative comparison on four depth completion datasets. The best and suboptimal results are highlighted. Note that most of the results on NYUv2, DIML, and SUN RGBD are borrowed directly from AGG-Net [2], while those of TOFDC are from TPVD [50]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.648, + 0.483, + 0.679 + ], + "angle": 0, + "content": "depth and the predicted depth, respectively. The reconstruction loss \\(\\mathcal{L}_r\\) is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.196, + 0.687, + 0.483, + 0.729 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {r} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} | | \\hat {\\mathbf {Y}} _ {i} - \\mathbf {Y} _ {i} | | _ {1}, \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.736, + 0.484, + 0.798 + ], + "angle": 0, + "content": "where \\(N\\) represents the number of training samples, and \\(\\| \\cdot \\| _1\\) denotes the \\(L_{1}\\) norm. Subsequently, we introduce the degradation loss \\(\\mathcal{L}_d\\) to facilitate the degradation learning process through:" + }, + { + "type": "equation", + "bbox": [ + 0.185, + 0.806, + 0.483, + 0.847 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {d} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\left| \\left| \\mathbf {K} \\otimes \\mathbf {Y} - \\mathbf {Z} \\right| \\right| _ {1}, \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.903 + ], + "angle": 0, + "content": "where \\(\\mathbf{K}\\) refers to the degradation kernel learned from the implicit degradation representation \\(\\mathbf{D}\\). \\(\\mathbf{Z}\\) corresponds to the coarse depth. \\(\\otimes\\) denotes the convolution operation, where" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.648, + 0.907, + 0.678 + ], + "angle": 0, + "content": "\\(\\mathbf{K}\\) serves as the convolution kernel. As a result, the total training loss is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.652, + 0.688, + 0.907, + 0.705 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {t} = \\mathcal {L} _ {r} + \\lambda \\mathcal {L} _ {d}, \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.712, + 0.907, + 0.773 + ], + "angle": 0, + "content": "where \\(\\lambda\\) is a hyper-parameter, empirically set to 0.1. By the constraint of the auxiliary loss \\(\\mathcal{L}_d\\), our model can build additional connections between the input and output, thus predicting more accurate depth results." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.785, + 0.64, + 0.803 + ], + "angle": 0, + "content": "4. Experiment" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.907, + 0.903 + ], + "angle": 0, + "content": "We evaluate our SigNet method on four popular datasets, including NYUv2, DIML, SUN RGBD, and TOFDC benchmarks. For NYUv2, DIML, and SUN RGBD, we maintain the settings used in AGG-Net [2], while for TOFDC, we follow the settings of TPVD [50]. Refer to our supplementary material for additional details on the experimental setups." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "26947" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.139, + 0.09, + 0.283, + 0.259 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.186, + 0.262, + 0.235, + 0.274 + ], + "angle": 0, + "content": "(a) RGB" + }, + { + "type": "image", + "bbox": [ + 0.284, + 0.09, + 0.427, + 0.259 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.324, + 0.262, + 0.386, + 0.274 + ], + "angle": 0, + "content": "(b) NLSPN" + }, + { + "type": "image", + "bbox": [ + 0.429, + 0.091, + 0.572, + 0.259 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.47, + 0.262, + 0.529, + 0.275 + ], + "angle": 0, + "content": "(c) RigNet" + }, + { + "type": "image", + "bbox": [ + 0.573, + 0.091, + 0.715, + 0.259 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.615, + 0.262, + 0.672, + 0.275 + ], + "angle": 0, + "content": "(d) SigNet" + }, + { + "type": "image", + "bbox": [ + 0.716, + 0.091, + 0.859, + 0.259 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.768, + 0.262, + 0.806, + 0.274 + ], + "angle": 0, + "content": "(e) GT" + }, + { + "type": "image_caption", + "bbox": [ + 0.134, + 0.281, + 0.861, + 0.296 + ], + "angle": 0, + "content": "Figure 4. Visual comparison with SOTA methods on TOFDC dataset, including NLSPN [29], RigNet [48], and our SigNet." + }, + { + "type": "image", + "bbox": [ + 0.139, + 0.306, + 0.283, + 0.435 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.187, + 0.439, + 0.235, + 0.451 + ], + "angle": 0, + "content": "(a) RGB" + }, + { + "type": "image", + "bbox": [ + 0.284, + 0.307, + 0.427, + 0.435 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.328, + 0.439, + 0.383, + 0.451 + ], + "angle": 0, + "content": "(b)CSPN" + }, + { + "type": "image", + "bbox": [ + 0.428, + 0.307, + 0.572, + 0.435 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.464, + 0.439, + 0.535, + 0.451 + ], + "angle": 0, + "content": "(c) AGG-Net" + }, + { + "type": "image", + "bbox": [ + 0.573, + 0.307, + 0.715, + 0.435 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.614, + 0.439, + 0.672, + 0.451 + ], + "angle": 0, + "content": "(d) SigNet" + }, + { + "type": "image", + "bbox": [ + 0.716, + 0.307, + 0.859, + 0.435 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.768, + 0.439, + 0.805, + 0.451 + ], + "angle": 0, + "content": "(e) GT" + }, + { + "type": "image_caption", + "bbox": [ + 0.136, + 0.458, + 0.859, + 0.473 + ], + "angle": 0, + "content": "Figure 5. Visual comparison with SOTA approaches on DIML dataset, including CSPN [4], AGG-Net [2], and our SigNet." + }, + { + "type": "image", + "bbox": [ + 0.139, + 0.482, + 0.257, + 0.594 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.183, + 0.598, + 0.214, + 0.609 + ], + "angle": 0, + "content": "RGB" + }, + { + "type": "image", + "bbox": [ + 0.259, + 0.482, + 0.377, + 0.594 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.28, + 0.598, + 0.355, + 0.61 + ], + "angle": 0, + "content": "Coarse depth" + }, + { + "type": "image", + "bbox": [ + 0.378, + 0.482, + 0.497, + 0.594 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.402, + 0.598, + 0.472, + 0.61 + ], + "angle": 0, + "content": "Degradation" + }, + { + "type": "image", + "bbox": [ + 0.5, + 0.482, + 0.619, + 0.594 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.545, + 0.598, + 0.574, + 0.609 + ], + "angle": 0, + "content": "RGB" + }, + { + "type": "image", + "bbox": [ + 0.62, + 0.482, + 0.738, + 0.594 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.641, + 0.598, + 0.715, + 0.61 + ], + "angle": 0, + "content": "Coarse depth" + }, + { + "type": "image", + "bbox": [ + 0.739, + 0.482, + 0.859, + 0.594 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.763, + 0.598, + 0.833, + 0.61 + ], + "angle": 0, + "content": "Degradation" + }, + { + "type": "image_caption", + "bbox": [ + 0.178, + 0.619, + 0.817, + 0.633 + ], + "angle": 0, + "content": "Figure 6. Visual examples of filled coarse depth maps and degradation representations on the DIML dataset." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.655, + 0.312, + 0.672 + ], + "angle": 0, + "content": "4.1. Comparison with SOTA" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.677, + 0.484, + 0.769 + ], + "angle": 0, + "content": "To assess the effectiveness of our SigNet, we benchmark it against the traditional bilateral filtering method [1] and some recent SOTA deep learning approaches such as [2, 28, 29, 34, 38]. The error and accuracy results for the competing models are sourced from their original publications or obtained by retraining them following the default settings." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.776, + 0.291, + 0.791 + ], + "angle": 0, + "content": "4.1.1. Performance Analysis" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.484, + 0.901 + ], + "angle": 0, + "content": "On NYUv2. As a widely used benchmark dataset, NYUv2 is often employed to evaluate the effectiveness of models. From the first row of Table 1, we can observe that SigNet achieves competitive performance on the NYUv2 dataset, with an RMSE that is \\(5mm\\) lower than the second-best method. Additionally, although our method lags behind BPNet [38] by only 0.001 in the REL metric, SigNet signifi" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.656, + 0.842, + 0.672 + ], + "angle": 0, + "content": "cantly reduces the number of parameters by \\(96\\%\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.673, + 0.907, + 0.793 + ], + "angle": 0, + "content": "On DIML. DIML is collected from both indoor and outdoor scenes, suffering from severe edge shadows and irregular holes. Our method is implemented using the same experimental settings as previous approaches [2, 33]. As shown in the second row of Table 1, SigNet demonstrates the best performance across all evaluation metrics. For example, compared to the suboptimal model, our approach decreases the RMSE by \\(19\\%\\) and the REL by \\(45\\%\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.907, + 0.901 + ], + "angle": 0, + "content": "On SUN RGBD. The large-scale SUN RGB-D dataset is leveraged to evaluate the generalization capability of our SigNet across different sensors and scenes. Table 1 (third row) reveals that our method surpasses previous state-of-the-art approaches across all metrics. Notably, SigNet achieves a \\(8mm\\) lower RMSE than the latest BPNet [38] and a \\(36mm\\) lower RMSE compared to AGG-Net [2]." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26948" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.107, + 0.089, + 0.468, + 0.164 + ], + "angle": 0, + "content": "
Density1%5%35%65%95%
TPVD [50]156.6139.7112.297.392.8
SigNet144.6121.599.490.184.3
Improv. ↑12.018.212.87.28.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.101, + 0.17, + 0.473, + 0.184 + ], + "angle": 0, + "content": "Table 2. Density comparison of TPVD and SigNet on TOFDC." + }, + { + "type": "image", + "bbox": [ + 0.104, + 0.193, + 0.473, + 0.407 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.414, + 0.483, + 0.442 + ], + "angle": 0, + "content": "Figure 7. Complexity comparison on the TOFDC dataset, where larger circle areas indicate higher inference time." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.462, + 0.483, + 0.584 + ], + "angle": 0, + "content": "On TOFDC. TOFDC encompasses extensive data captured under various lighting conditions and in open spaces. It is utilized to further assess the performance of the proposed model. The quantitative results in the last row of Table 1 indicate that our method achieves state-of-the-art performance. Compared to the second-best TPVD [50], SigNet reduces the RMSE by \\(5mm\\) and the REL by 0.002, while also decreasing the number of parameters by \\(27.9M\\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.59, + 0.292, + 0.605 + ], + "angle": 0, + "content": "4.1.2. Visualization Analysis" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.609, + 0.483, + 0.745 + ], + "angle": 0, + "content": "Figs. 4 and 5 showcase visual results on TOFDC and DIML, highlighting SigNet's ability to recover precise depth structures. Compared to previous models, our method yields sharper predictions, such as the leaves in Fig. 4 and the chair in Fig. 5, achieving closer alignment with the ground truth. These results affirm the effectiveness of our approach in enhancing depth completion. Additionally, Fig. 6 presents some degradation examples, further demonstrating the ability of SigNet to model and address such challenges." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.752, + 0.303, + 0.766 + ], + "angle": 0, + "content": "4.1.3. Number of Valid Points" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.77, + 0.483, + 0.847 + ], + "angle": 0, + "content": "Tab. 2 lists a comparison of different valid points on DIML. It indicates that our SigNet consistently outperforms TPVD, achieving an average improvement of \\(9.7\\%\\) across the five densities. These results highlight the generalization capability of our method across varying valid points." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.852, + 0.283, + 0.867 + ], + "angle": 0, + "content": "4.1.4. Complexity Analysis" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Fig. 7 shows that our method achieves an excellent balance among RMSE, parameters, and inference time. For exam-" + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.089, + 0.905, + 0.201 + ], + "angle": 0, + "content": "
SigNetDeg. BridgeRGB-D FusionRMSE (mm)
DensifyDenoiseConc.Att.DADF-DDADF-F
i108
ii✓✓95
iii✓✓✓92
iv✓✓✓74
v✓✓✓68
v✓✓✓✓55
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.209, + 0.905, + 0.251 + ], + "angle": 0, + "content": "Table 3. Ablation study of our SigNet on the DIML dataset. Conc.: concat, Att.: attention, DADF-D: degradation decomposition in DADF module, DADF-F: conditional Mamba fusion." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.276, + 0.907, + 0.368 + ], + "angle": 0, + "content": "ple, compared to the fewer parameter CSPN [4], our SigNet not only reduces parameters by \\(81\\%\\) and inference time by \\(82\\%\\), but also decreases RMSE by \\(61\\%\\). Additionally, our method surpasses the latest TPVD [50] with a \\(5\\%\\) lower RMSE and \\(78\\%\\) lower FPS, while reducing parameters by \\(89\\%\\) and inference time by \\(78\\%\\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.378, + 0.667, + 0.394 + ], + "angle": 0, + "content": "4.2. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.4, + 0.907, + 0.779 + ], + "angle": 0, + "content": "SigNet Designs. Table 3 presents the ablation results on the DIML test split. The iteration of DADF is set to five. SigNet-i serves as the baseline, where both the degradation bridge and DADF module are removed, and the RGB-D features are fused using addition. Building on this, SigNet-ii redefines depth completion as depth enhancement through densification and degradation, concatenating the RGB and degradation features, resulting in a \\(13\\mathrm{mm}\\) improvement. This highlights the superiority of our task transformation approach. Additionally, SigNet-iii incorporates a Gaussian denoising operator to further reduce error. To better utilize the degradation, SigNet-iv applies a Sigmoid function and element-wise multiplication to obtain an attention map, enabling adaptive selection of RGB features. Consequently, performance improves significantly from \\(92\\mathrm{mm}\\) to \\(74\\mathrm{mm}\\). Furthermore, SigNet-v employs degradation-aware decomposition, reducing the error by an additional \\(6\\mathrm{mm}\\). The results of SigNet-iii, iv, and v demonstrate the effectiveness of our adaptive feature selection strategy. Notably, in SigNet-iii, iv, and v, the selected RGB and depth features are encoded through concatenation and convolutions. Finally, with the introduction of our conditional Mamba, SigNet-vi surpasses SigNet-v by \\(13\\mathrm{mm}\\) in RMSE, representing a significant improvement. Overall, each component positively impacts the baseline." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Mamba Fusion and Iteration of DADF. Fig. 8 (a) illustrates the ablation study of Mamba fusion on the DIML dataset. For rapid validation, we employ DADF only once. The baseline ViM directly encodes the combined RGB-D features using the vanilla Vision Mamba [25]. When we apply our Conditional Mamba with initial degradation and RGB-D features as input, the error decreases by \\(8\\mathrm{mm}\\). Additionally, the introduction of degradation decomposition" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26949" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.101, + 0.089, + 0.475, + 0.223 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.107, + 0.231, + 0.466, + 0.245 + ], + "angle": 0, + "content": "Figure 8. Ablation study of our DADF on the DIML dataset." + }, + { + "type": "table", + "bbox": [ + 0.106, + 0.26, + 0.468, + 0.307 + ], + "angle": 0, + "content": "
Path number357911
RMSE (mm)55.054.654.554.154.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.092, + 0.316, + 0.482, + 0.33 + ], + "angle": 0, + "content": "Table 4. Ablation on the number of decomposition path in DADF." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.357, + 0.484, + 0.584 + ], + "angle": 0, + "content": "using DCT further enhances performance. We also observe that the results improve as the number of DCT components increases, highlighting the effectiveness of our gradual DCT decomposition design. Fig. 8 (b) presents the ablation study of our DADF with different iterations on the DIML dataset. We find that performance improves with an increasing number of iterations. For an optimal balance between efficiency and effectiveness, we set the final iteration count to five. Moreover, as shown in Fig.3, the degradation \\(\\mathbf{D}\\) can be decomposed into \\(n\\) components using DCT. As indicated in Tab.4, the error gradually decreases as \\(n\\) increases. However, considering the computational complexity, we select the top-3 most representative components to focus on the high-frequency components of \\(\\mathbf{F}_{img}\\) and update the Mamba condition accordingly." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.592, + 0.226, + 0.607 + ], + "angle": 0, + "content": "4.3. Failure Case" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.614, + 0.483, + 0.718 + ], + "angle": 0, + "content": "Our method is primarily evaluated in indoor scenes and some simple outdoor scenes with closer distances. By constraining dense predictions with coarse depth maps, SigNet can effectively predict their degradation representations. However, as depicted in Fig. 9, SigNet fails to estimate reasonable degradation on the outdoor KITTI benchmark [40]. We identify two main reasons for this:" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.483, + 0.795 + ], + "angle": 0, + "content": "1) The ground truth depth maps of KITTI are very sparse, with only about \\(30\\%\\) valid pixels. This sparsity leads to inconsistent depth edges, resulting in a relatively weak constraint. For instance, the degradation in Fig. 9 even introduces some initial LiDAR lines." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.483, + 0.871 + ], + "angle": 0, + "content": "2) The depth distance of KITTI data is much larger, which significantly weakens edge details. As observed, the degradation in Fig. 9 can only learn very coarse and blurry representations near edges at short range, while at long range, it learns incorrect representations." + }, + { + "type": "list", + "bbox": [ + 0.09, + 0.72, + 0.483, + 0.871 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.483, + 0.902 + ], + "angle": 0, + "content": "One potential solution is to provide the model with supervisory signals regarding dense contents and precise" + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.089, + 0.905, + 0.389 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.397, + 0.905, + 0.426 + ], + "angle": 0, + "content": "Figure 9. Visual results of failure cases on KITTI benchmark [40]. DA v2 denotes the large depth model DepthAnything v2 [52]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.447, + 0.907, + 0.6 + ], + "angle": 0, + "content": "edges. It is well-known that large depth models can produce dense depth with impressive sharp edges. For example, we employ DepthAnything v2 [52] to generate the corresponding dense depth. We discover that the edges of cars and even tree leaves are much clearer. Although the depth scale does not match that of KITTI, we can generate a map that offers constraints near edges via gradient or boundary extraction techniques [46]. Certainly, the results of DepthAnything v2 [52] or DepthAnything v2 itself can also be utilized to facilitate depth completion, warranting further exploration." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.611, + 0.634, + 0.627 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.637, + 0.907, + 0.834 + ], + "angle": 0, + "content": "In this paper, we redefined depth completion as depth enhancement for the first time with our novel framework, SigNet. SigNet leveraged non-CNN densification tools to transform sparse depth data into coarse yet dense depth, fully eliminating the mismatch and ambiguity caused by direct convolution over irregularly sampled sparse data. More importantly, by establishing a self-supervised degradation bridge, SigNet effectively fused RGB-D data. It utilized degradation decomposition to dynamically select high-frequency components of RGB and build a conditional Mamba for the depth enhancement. Extensive experiments validated the superiority of SigNet, highlighting its potential for depth related applications." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.846, + 0.682, + 0.862 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.906, + 0.901 + ], + "angle": 0, + "content": "This work was supported by the National Natural Science Fund of China (U24A20330, 62361166670 and 62072242)." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26950" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.482, + 0.156 + ], + "angle": 0, + "content": "[1] Antoni Buades, Bartomeu Coll, and J-M Morel. A non-local algorithm for image denoising. In CVPR, pages 60-65, 2005. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.159, + 0.483, + 0.215 + ], + "angle": 0, + "content": "[2] Dongyue Chen, Tingxuan Huang, Zhimin Song, Shizhuo Deng, and Tong Jia. Agg-net: Attention guided gated convolutional network for depth image completion. In ICCV, pages 8853-8862, 2023. 5, 6, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.218, + 0.482, + 0.259 + ], + "angle": 0, + "content": "[3] Yun Chen, Bin Yang, Ming Liang, and Raquel Urtasun. Learning joint 2d-3d representations for depth completion. In ICCV, pages 10023–10032, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.262, + 0.482, + 0.303 + ], + "angle": 0, + "content": "[4] Xinjing Cheng, Peng Wang, and Ruigang Yang. Learning depth with convolutional spatial propagation network. In ECCV, pages 103-119, 2018. 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.307, + 0.482, + 0.348 + ], + "angle": 0, + "content": "[5] Xinjing Cheng, Peng Wang, and Ruigang Yang. Depth estimation via affinity learned with convolutional spatial propagation network. In ECCV, pages 103-119, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.35, + 0.482, + 0.404 + ], + "angle": 0, + "content": "[6] Xinjing Cheng, Peng Wang, and Ruigang Yang. Learning depth with convolutional spatial propagation network. IEEE Transactions on Pattern Analysis and Machine Intelligence, 42(10):2361-2379, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.408, + 0.482, + 0.462 + ], + "angle": 0, + "content": "[7] Xinjing Cheng, Peng Wang, Chenye Guan, and Ruigang Yang. Cspn++: Learning context and resource aware convolutional spatial propagation networks for depth completion. In AAAI, pages 10615-10622, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.466, + 0.482, + 0.52 + ], + "angle": 0, + "content": "[8] Jaehoon Cho, Dongbo Min, Youngjung Kim, and Kwanghoon Sohn. A large rgb-d dataset for semi-supervised monocular depth estimation. arXiv preprint arXiv:1904.10230, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.524, + 0.482, + 0.579 + ], + "angle": 0, + "content": "[9] Abdelrahman Eldisokey, Michael Felsberg, and Fahad Shahbaz Khan. Confidence propagation through cnns for guided sparse depth regression. IEEE Transactions on Pattern Analysis and Machine Intelligence, 42(10):2423-2436, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.582, + 0.482, + 0.622 + ], + "angle": 0, + "content": "[10] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.626, + 0.482, + 0.667 + ], + "angle": 0, + "content": "[11] Mu Hu, Shuling Wang, Bin Li, Shiyu Ning, Li Fan, and Xiaojin Gong. Penet: Towards precise and efficient image guided depth completion. In ICRA, 2021. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.67, + 0.482, + 0.724 + ], + "angle": 0, + "content": "[12] Maximilian Jaritz, Raoul De Charette, Emilie Wirbel, Xavier Perrotton, and Fawzi Nashashibi. Sparse and dense data with cnns: Depth completion and semantic segmentation. In 3DV, pages 52-60, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.728, + 0.482, + 0.755 + ], + "angle": 0, + "content": "[13] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In Computer Ence, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.758, + 0.482, + 0.799 + ], + "angle": 0, + "content": "[14] Jason Ku, Ali Harakeh, and Steven L Waslander. In defense of classical image processing: Fast depth completion on the cpu. In CRV, pages 16-22, 2018. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.802, + 0.482, + 0.842 + ], + "angle": 0, + "content": "[15] Anat Levin, Dani Lischinski, and Yair Weiss. Colorization using optimization. In SIGGRAPH, pages 689-694, 2004. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.482, + 0.899 + ], + "angle": 0, + "content": "[16] Ang Li, Zejian Yuan, Yonggen Ling, Wanchao Chi, Chong Zhang, et al. A multi-scale guided cascade hourglass network for depth completion. In WACV, pages 32-40, 2020. 2" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.483, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.135 + ], + "angle": 0, + "content": "[17] Boyun Li, Xiao Liu, Peng Hu, Zhongqin Wu, Jiancheng Lv, and Xi Peng. All-in-one image restoration for unknown corruption. In CVPR, pages 17452-17462, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.137, + 0.905, + 0.178 + ], + "angle": 0, + "content": "[18] Jie Liang, Hui Zeng, and Lei Zhang. Efficient and degradation-adaptive network for real-world image superresolution. In ECCV, pages 574-591, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.181, + 0.905, + 0.222 + ], + "angle": 0, + "content": "[19] Yuankai Lin, Tao Cheng, Qi Zhong, Wending Zhou, and Hua Yang. Dynamic spatial propagation network for depth completion. In AAAI, pages 1638-1646, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.224, + 0.905, + 0.279 + ], + "angle": 0, + "content": "[20] Yuankai Lin, Hua Yang, Tao Cheng, Wending Zhou, and Zhouping Yin. Dyspn: Learning dynamic affinity for image-guided depth completion. IEEE Transactions on Circuits and Systems for Video Technology, pages 1-1, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.282, + 0.905, + 0.336 + ], + "angle": 0, + "content": "[21] Lina Liu, Yiyi Liao, Yue Wang, Andreas Geiger, and Yong Liu. Learning steering kernels for guided depth completion. IEEE Transactions on Image Processing, 30:2850-2861, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.34, + 0.905, + 0.394 + ], + "angle": 0, + "content": "[22] Lina Liu, Xibin Song, Xiaoyang Lyu, Junwei Diao, Mengmeng Wang, Yong Liu, and Liangjun Zhang. Fcfr-net: Feature fusion based coarse-to-fine residual learning for depth completion. In AAAI, pages 2136-2144, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.397, + 0.905, + 0.438 + ], + "angle": 0, + "content": "[23] Sifei Liu, Shalini De Mello, Jinwei Gu, Guangyu Zhong, Ming-Hsuan Yang, and Jan Kautz. Learning affinity via spatial propagation networks. In NeurIPS, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.441, + 0.905, + 0.482 + ], + "angle": 0, + "content": "[24] Xin Liu, Xiaofei Shao, Bo Wang, Yali Li, and Shengjin Wang. Graphcspn: Geometry-aware depth completion via dynamic gcns. In ECCV, pages 90-107. Springer, 2022. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.485, + 0.905, + 0.538 + ], + "angle": 0, + "content": "[25] Yue Liu, Yunjie Tian, Yuzhong Zhao, Hongtian Yu, Lingxi Xie, Yaowei Wang, Qixiang Ye, and Yunfan Liu. Vmamba: Visual state space model. arXiv preprint arXiv:2401.10166, 2024. 2, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.542, + 0.905, + 0.596 + ], + "angle": 0, + "content": "[26] Kaiyue Lu, Nick Barnes, Saeed Anwar, and Liang Zheng. From depth what can you see? depth completion via auxiliary image reconstruction. In CVPR, pages 11306-11315, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.6, + 0.905, + 0.641 + ], + "angle": 0, + "content": "[27] Fangchang Ma and Sertac Karaman. Sparse-to-dense: Depth prediction from sparse depth samples and a single image. In ICRA, pages 4796-4803. IEEE, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.643, + 0.905, + 0.697 + ], + "angle": 0, + "content": "[28] Fangchang Ma, Guilherme Venturelli Cavalheiro, and Sertac Karaman. Self-supervised sparse-to-dense: Self-supervised depth completion from lidar and monocular camera. In ICRA, 2019. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.701, + 0.905, + 0.742 + ], + "angle": 0, + "content": "[29] Jinsun Park, Kyungdon Joo, Zhe Hu, Chi-Kuei Liu, and In So Kweon. Non-local spatial propagation network for depth completion. In ECCV, 2020. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.745, + 0.905, + 0.786 + ], + "angle": 0, + "content": "[30] Jin-Hwi Park, Chanhwi Jeong, Junoh Lee, and Hae-Gon Jeon. Depth prompting for sensor-agnostic depth estimation. In CVPR, pages 9859–9869, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.789, + 0.905, + 0.857 + ], + "angle": 0, + "content": "[31] Jiaxiong Qiu, Zhaopeng Cui, Yinda Zhang, Xingdi Zhang, Shuaicheng Liu, Bing Zeng, and Marc Pollefeys. Deep surface normal guided depth prediction for outdoor scene from sparse lidar data and single color image. In CVPR, pages 3313-3322, 2019. 5, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[32] Kyeongha Rho, Jinsung Ha, and Youngjung Kim. Guideformer: Transformers for image guided depth completion. In CVPR, pages 6250-6259, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "26951" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.146 + ], + "angle": 0, + "content": "[33] Dmitry Senushkin, Mikhail Romanov, Ilia Belikov, Nikolay Patakin, and Anton Konushin. Decoder modulation for indoor depth completion. In IROS, pages 2181-2188. IEEE, 2021. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.218 + ], + "angle": 0, + "content": "[34] Shreyas S Shivakumar, Ty Nguyen, Ian D Miller, Steven W Chen, Vijay Kumar, and Camillo J Taylor. Dfusenet: Deep fusion of rgb and sparse depth information for image guided dense depth completion. In ITSC, pages 13-20. IEEE, 2019. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.22, + 0.482, + 0.262 + ], + "angle": 0, + "content": "[35] Nathan Silberman, Derek Hoiem, Pushmeet Kohli, and Rob Fergus. Indoor segmentation and support inference from rgbd images. In ECCV, pages 746-760. Springer, 2012. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.263, + 0.482, + 0.304 + ], + "angle": 0, + "content": "[36] Shuran Song, Samuel P Lichtenberg, and Jianxiong Xiao. Sun rgb-d: A rgb-d scene understanding benchmark suite. In CVPR, pages 567-576, 2015. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.305, + 0.482, + 0.359 + ], + "angle": 0, + "content": "[37] Jie Tang, Fei-Peng Tian, Wei Feng, Jian Li, and Ping Tan. Learning guided convolutional network for depth completion. IEEE Transactions on Image Processing, 30:1116-1129, 2020. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.362, + 0.482, + 0.403 + ], + "angle": 0, + "content": "[38] Jie Tang, Fei-Peng Tian, Boshi An, Jian Li, and Ping Tan. Bilateral propagation network for depth completion. In CVPR, pages 9763-9772, 2024. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.405, + 0.482, + 0.446 + ], + "angle": 0, + "content": "[39] Diego Thomas and Akihiro Sugimoto. A flexible scene representation for 3d reconstruction using an rgb-d camera. In ICCV, pages 2800-2807, 2013. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.448, + 0.482, + 0.489 + ], + "angle": 0, + "content": "[40] Jonas Uhrig, Nick Schneider, Lukas Schneider, Uwe Franke, Thomas Brox, and Andreas Geiger. Sparsity invariant cnns. In 3DV, pages 11–20, 2017. 1, 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.491, + 0.482, + 0.533 + ], + "angle": 0, + "content": "[41] Wouter Van Gansbeke, Davy Neven, Bert De Brabandere, and Luc Van Gool. Sparse and noisy lidar completion with rgb guidance and uncertainty. In MVA, pages 1-6, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.534, + 0.482, + 0.588 + ], + "angle": 0, + "content": "[42] Haowen Wang, Mingyuan Wang, Zhengping Che, Zhiyuan Xu, Xiuquan Qiao, Mengshi Qi, Feifei Feng, and Jian Tang. RGB-depth fusion gan for indoor depth completion. In CVPR, pages 6209-6218, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.59, + 0.482, + 0.645 + ], + "angle": 0, + "content": "[43] Longguang Wang, Yingqian Wang, Xiaoyu Dong, Qingyu Xu, Jungang Yang, Wei An, and Yulan Guo. Unsupervised degradation representation learning for blind superresolution. In CVPR, pages 10581-10590, 2021. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.647, + 0.482, + 0.688 + ], + "angle": 0, + "content": "[44] Yufei Wang, Bo Li, Ge Zhang, Qi Liu, Tao Gao, and Yuchao Dai. Lrru: Long-short range recurrent updating networks for depth completion. In ICCV, pages 9422-9432, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.689, + 0.482, + 0.743 + ], + "angle": 0, + "content": "[45] Yufei Wang, Ge Zhang, Shaoqian Wang, Bo Li, Qi Liu, Le Hui, and Yuchao Dai. Improving depth completion via depth feature upsampling. In CVPR, pages 21104-21113, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.746, + 0.482, + 0.8 + ], + "angle": 0, + "content": "[46] Zhengxue Wang, Zhiqiang Yan, and Jian Yang. Sgnet: Structure guided network via gradient-frequency awareness for depth map super-resolution. In AAAI, pages 5823-5831, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.803, + 0.482, + 0.856 + ], + "angle": 0, + "content": "[47] Zhiqiang Yan, Xiang Li, Kun Wang, Zhenyu Zhang, Jun Li, and Jian Yang. Multi-modal masked pre-training for monocular panoramic depth completion. In ECCV, pages 378-395, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.859, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[48] Zhiqiang Yan, Kun Wang, Xiang Li, Zhenyu Zhang, Jun Li, and Jian Yang. Rignet: Repetitive image guided network for depth completion. In ECCV, pages 214-230, 2022. 1, 2, 5, 6" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.092, + 0.905, + 0.146 + ], + "angle": 0, + "content": "[49] Zhiqiang Yan, Kun Wang, Xiang Li, Zhenyu Zhang, Jun Li, and Jian Yang. Desnet: Decomposed scale-consistent network for unsupervised depth completion. In AAAI, pages 3109-3117, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.15, + 0.905, + 0.207 + ], + "angle": 0, + "content": "[50] Zhiqiang Yan, Yuankai Lin, Kun Wang, Yupeng Zheng, Yufei Wang, Zhenyu Zhang, Jun Li, and Jian Yang. Tri-perspective view decomposition for geometry-aware depth completion. In CVPR, pages 4874-4884, 2024. 1, 2, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.209, + 0.905, + 0.25 + ], + "angle": 0, + "content": "[51] Zhiqiang Yan, Yupeng Zheng, Deng-Ping Fan, Xiang Li, Jun Li, and Jian Yang. Learnable differencing center for nighttime depth perception. Visual Intelligence, 2(1):15, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.252, + 0.905, + 0.293 + ], + "angle": 0, + "content": "[52] Lihe Yang, Bingyi Kang, Zilong Huang, Zhen Zhao, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything v2. arXiv preprint arXiv:2406.09414, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.296, + 0.905, + 0.338 + ], + "angle": 0, + "content": "[53] Yanchao Yang, Alex Wong, and Stefano Soatto. Dense depth posterior (ddp) from single image and sparse range. In CVPR, pages 3353-3362, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.34, + 0.905, + 0.395 + ], + "angle": 0, + "content": "[54] Mingde Yao, Ruikang Xu, Yuanshen Guan, Jie Huang, and Zhiwei Xiong. Neural degradation representation learning for all-in-one image restoration. IEEE Transactions on Image Processing, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.398, + 0.905, + 0.466 + ], + "angle": 0, + "content": "[55] Guanghao Yin, Wei Wang, Zehuan Yuan, Wei Ji, Dongdong Yu, Shouqian Sun, Tat-Seng Chua, and Changhu Wang. Conditional hyper-network for blind super-resolution with multiple degradations. IEEE Transactions on Image Processing, 31:3949-3960, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.469, + 0.905, + 0.523 + ], + "angle": 0, + "content": "[56] Zhu Yu, Zehua Sheng, Zili Zhou, Lun Luo, Si-Yuan Cao, Hong Gu, Huaqi Zhang, and Hui-Liang Shen. Aggregating feature point cloud for depth completion. In ICCV, pages 8732-8743, 2023. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.527, + 0.905, + 0.581 + ], + "angle": 0, + "content": "[57] Cheng Zhang, Yu Zhu, Qingsen Yan, Jinqiu Sun, and Yanning Zhang. All-in-one multi-degradation image restoration network via hierarchical degradation representation. In ACM MM, pages 2285-2293, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.584, + 0.905, + 0.639 + ], + "angle": 0, + "content": "[58] Jinghao Zhang, Jie Huang, Mingde Yao, Zizheng Yang, Hu Yu, Man Zhou, and Feng Zhao. Ingredient-oriented multi-degradation learning for image restoration. In CVPR, pages 5825-5835, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.642, + 0.905, + 0.696 + ], + "angle": 0, + "content": "[59] Kai Zhang, Jingyun Liang, Luc Van Gool, and Radu Timofte. Designing a practical degradation model for deep blind image super-resolution. In ICCV, pages 4791-4800, 2021. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.7, + 0.905, + 0.754 + ], + "angle": 0, + "content": "[60] Yulun Zhang, Kunpeng Li, Kai Li, Lichen Wang, Bineng Zhong, and Yun Fu. Image super-resolution using very deep residual channel attention networks. In ECCV, pages 286-301, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.758, + 0.905, + 0.813 + ], + "angle": 0, + "content": "[61] Youmin Zhang, Xianda Guo, Matteo Poggi, Zheng Zhu, Guan Huang, and Stefano Mattoccia. Completionformer: Depth completion with convolutions and vision transformers. In CVPR, pages 18527-18536, 2023. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.816, + 0.905, + 0.856 + ], + "angle": 0, + "content": "[62] Zhiwei Zhong, Xianming Liu, Junjun Jiang, Debin Zhao, and Xiangyang Ji. Guided depth map super-resolution: A survey. ACM Computing Surveys, 55(14s):1-36, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.859, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[63] Hanyu Zhou, Yi Chang, Wending Yan, and Luxin Yan. Unsupervised cumulative domain adaptation for foggy scene optical flow. In CVPR, pages 9569-9578, 2023. 1" + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.092, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "26952" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.16 + ], + "angle": 0, + "content": "[64] Hongyang Zhou, Xiaobin Zhu, Jianqing Zhu, Zheng Han, Shi-Xue Zhang, Jingyan Qin, and Xu-Cheng Yin. Learning correction filter via degradation-adaptive regression for blind single image super-resolution. In ICCV, pages 12365-12375, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.164, + 0.482, + 0.232 + ], + "angle": 0, + "content": "[65] Hanyu Zhou, Yi Chang, Zhiwei Shi, Wending Yan, Gang Chen, Yonghong Tian, and Luxin Yan. Adverse weather optical flow: Cumulative homogeneous-heterogeneous adaptation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.234, + 0.482, + 0.289 + ], + "angle": 0, + "content": "[66] Wending Zhou, Xu Yan, Yinghong Liao, Yuankai Lin, Jin Huang, Gangming Zhao, Shuguang Cui, and Zhen Li. Bev@ dc: Bird's-eye view assisted training for depth completion. In CVPR, pages 9233-9242, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.291, + 0.482, + 0.344 + ], + "angle": 0, + "content": "[67] Ziyue Zhu, Qiang Meng, Xiao Wang, Ke Wang, Liujiang Yan, and Jian Yang. Curricular object manipulation in lidar-based object detection. In CVPR, pages 1125-1135, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.347, + 0.482, + 0.402 + ], + "angle": 0, + "content": "[68] Ziyue Zhu, Shenlong Wang, Jin Xie, Jiang-Jiang Liu, Jingdong Wang, and Jian Yang. Voxelsplat: Dynamic gaussian splatting as an effective loss for occupancy and flow prediction. In CVPR, 2025. 1" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.402 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.956 + ], + "angle": 0, + "content": "26953" + } + ] +] \ No newline at end of file diff --git a/2025/Completion as Enhancement_ A Degradation-Aware Selective Image Guided Network for Depth Completion/5b8a9777-b777-4fb5-9ba2-85ebeb5a648a_origin.pdf b/2025/Completion as Enhancement_ A Degradation-Aware Selective Image Guided Network for Depth Completion/5b8a9777-b777-4fb5-9ba2-85ebeb5a648a_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..4692170e9bb984277c351eef8cdc6287015415c9 --- /dev/null +++ b/2025/Completion as Enhancement_ A Degradation-Aware Selective Image Guided Network for Depth Completion/5b8a9777-b777-4fb5-9ba2-85ebeb5a648a_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:895be4d9b7fd917ad1261f72c524f9a3dd9dce5e1c7ae1ad7bb00cd8014a4902 +size 2956656 diff --git a/2025/Completion as Enhancement_ A Degradation-Aware Selective Image Guided Network for Depth Completion/full.md b/2025/Completion as Enhancement_ A Degradation-Aware Selective Image Guided Network for Depth Completion/full.md new file mode 100644 index 0000000000000000000000000000000000000000..ed974bddc3f359d8cac1608eb106d8f86bd7dfac --- /dev/null +++ b/2025/Completion as Enhancement_ A Degradation-Aware Selective Image Guided Network for Depth Completion/full.md @@ -0,0 +1,388 @@ +# Completion as Enhancement: A Degradation-Aware Selective Image Guided Network for Depth Completion + +Zhiqiang Yan Zhengxue Wang Kun Wang Jun Li\* Jian Yang\* + +PCA Lab†, Nanjing University of Science and Technology + +{yanzq,zxwang,kunwang,junli,csjyang}@njust.edu.cn + +# Abstract + +In this paper, we introduce the Selective Image Guided Network (SigNet), a novel degradation-aware framework that transforms depth completion into depth enhancement for the first time. Moving beyond direct completion using convolutional neural networks (CNNs), SigNet initially densifies sparse depth data through non-CNN densification tools to obtain coarse yet dense depth. This approach eliminates the mismatch and ambiguity caused by direct convolution over irregularly sampled sparse data. Subsequently, SigNet redefines completion as enhancement, establishing a self-supervised degradation bridge between the coarse depth and the targeted dense depth for effective RGB-D fusion. To achieve this, SigNet leverages the implicit degradation to adaptively select high-frequency components (e.g., edges) of RGB data to compensate for the coarse depth. This degradation is further integrated into a multi-modal conditional Mamba, dynamically generating the state parameters to enable efficient global high-frequency information interaction. We conduct extensive experiments on the NYUv2, DIML, SUN RGBD, and TOFDC datasets, demonstrating the state-of-the-art (SOTA) performance of SigNet. + +# 1. Introduction + +Depth completion [27, 38, 48, 50] aims to predict a dense depth map from sparse and noisy depth measurements, often using the corresponding RGB image for auxiliary guidance. The resulting dense depth map can benefit various downstream applications such as scene reconstruction [36, 39, 47, 68] and self-driving [22, 45, 51, 63, 65, 67]. + +The key challenges of this task lie in the sparse depth input data [21, 38, 40, 49, 50]. We summarize these challenges as mismatch and ambiguity. As shown in Fig. 1, the mismatch is caused by the irregularly sampled sparse + +![](images/4319659526abef301e7b370531fede098225e97d949d8e8117db046c20b330a4.jpg) +Figure 1. Illustration of our main concept. It redefines depth completion as depth enhancement. We first transform sparse depth input via non-CNN densification tools [14, 15], yielding coarse but dense depth. Then we predict the precise and dense depth from the coarse depth by leveraging degradation assumption [43, 59, 62]. + +depth with varying point distributions, while convolutions are spatially invariant with fixed kernel shapes. Moreover, it is ambiguous for convolutions to distinguish between valid and invalid depth measurements since unknown pixels in the sparse depth are recognized as zero. + +To address these challenges, several excellent methods [21, 38, 40] have been proposed. For instance, SparseCOnvs [40] proposes sparsity invariant CNNs that explicitly model the location of missing data. To avoid directly convolving on sparse depth, Liu et al. [21] employs kernels learned by differentiable kernel regression layers to interpolate sparse data. BPNet [38] develops bilateral propagation, where the coefficients are generated based on both radiometric difference and spatial distance. These data-driven schemes are robust and impressive. However, they still cannot fully avoid the application of convolution on incomplete depth data, even though the depth data has become denser. + +In this paper, our objective is to develop a novel solution that not only completely resolves the issues of mismatch and ambiguity but also provides a fresh perspective on revisiting the traditional depth completion task. + +To this end, we propose a novel framework called the Selective Image Guided Network (SigNet). As shown in Fig. 1, SigNet redefines depth completion as depth enhancement1. It first leverages non-CNN densification tools [14, 15] to obtain coarse but dense depth. This coarse depth is then enhanced by a CNN-based subnetwork. Inspired by previous approaches [59, 62], we assume that there exists unknown degradation $\mathbf{h}$ between the coarse depth $\mathbf{Z}$ and the target depth $\mathbf{Y}$ , described as + +$$ +\mathbf {Z} = \mathbf {h} \mathbf {Y} + \mathbf {n}, \tag {1} +$$ + +where $n$ denotes additive noise. Therefore, we can establish a connection between the source depth and the target depth through the degradation. A fundamental observation is that degradation typically occurs near edges, suggesting that this phenomenon can be leveraged to enhance RGB-D fusion. This is particularly relevant given that the coarse depth often lacks significant edge details, while the RGB image contains abundant high-frequency components. Specifically, SigNet introduces a Degradation-Aware Decomposition and Fusion (DADF) module. DADF begins by decomposing the implicit degradation $h$ into multiple patches and estimating their coefficients in the discrete cosine domain. Each patch is then used to adaptively select the most representative high-frequency information from the RGB image. To achieve effective RGB-D aggregation, DADF incorporates a conditional Mamba to encode long-range correlations. Unlike the vanilla Vision Mamba [25], which calculates state parameters from the input, DADF uses the degradation patches as additional conditions to generate the state parameters. This approach enables further global high-frequency information interaction. + +In summary, our contributions are as follows: + +- We introduce a novel framework called SigNet, which redefines depth completion as depth enhancement for the first time. By adopting a degradation-aware approach, SigNet not only offers a new perspective to the field but also fully eliminates the traditional challenges such as mismatch and ambiguity. +- We propose a Degradation-Aware Decomposition and Fusion (DADF) module. Utilizing degradation decomposition, DADF adaptively selects high-frequency components of RGB data to compensate for coarse depth. Furthermore, DADF achieves effective RGB-D fusion by incorporating conditional Mamba, where the state parameters are generated from degradation patches. + +- We evaluate SigNet through comprehensive experiments, demonstrating its superiority over previous SOTAs. + +# 2. Related Work + +Depth Completion. Depth completion can be broadly categorized into depth-only and image-guided categories. + +Depth-only methods [9, 12, 14, 28, 40, 41, 44, 50] obtain dense depth directly from a single sparse depth input. For example, IP-Basic [14] designs reasonable combinations of basic image processing operations to densify sparse depth without using CNNs. Ma et al. [28] employ a UNet-like architecture to progressively densify the sparse depth input. FusionNet [41] designs a lightweight completion network that aggregates both global and local information with confidence adjustment. Moreover, Lu et al. [26] introduce a color image branch as auxiliary supervision during training. + +Despite advancements in single-modal methods, their performance is often limited without supplementary reference information. To address this challenge, numerous image-guided techniques have been proposed [11, 16, 19, 22, 37, 38, 45, 50, 53, 61, 66]. For example, GuideNet [37] and RigNet [48] utilize color images to generate adaptive filtering kernels for depth feature extraction. GFormer [32] and CFormer [61] integrate convolution and transformer mechanisms to capture both local and global representations. LRRU [44] introduces a dynamic kernel scope that transitions from large to small to capture varying dependencies. FuseNet [3] and PointDC [56] incorporate LiDAR point cloud branches to model 3D geometries. BEV@DC [66] uses bird's-eye view to build a point-voxel architecture. Additionally, DepthPrompting [30] embeds priors of large depth models for better sensor-agnostic generalization. + +To further advance completion models, spatial propagation networks (SPNs) have been introduced. The initial SPN [23] is designed to learn a pairwise similarity matrix. CSPN [5] applies this technique to depth completion by conducting recursive convolutions with fixed local neighborhood kernels, while $\mathrm{CSPN} + +$ [7] adapts by learning variable kernel sizes. PENet [11] expands the receptive fields using dilated convolutions. Furthermore, NLSPN [29] integrates non-local neighbors via deformable convolutions. DySPN [20] generates dynamic non-linear neighbors using an attention mechanism. Most recently, 3D SPNs [6] have been developed to encode 3D geometry. For example, GraphC-SPN [24] employs additional geometric constraints to regulate 3D propagation. TPVD [50] introduces a geometric SPN that constructs the affinity in 2D tri-perspective view spaces and their combined 3D projection space. + +Degradation Learning. Recently, color image restoration has been significantly advanced by improvements in degradation representations [43, 55, 59]. For example, Zhang et al. [59] construct a sophisticated yet practical degradation model to simulate real-world image degradation for + +![](images/38425302d36afe1dcc84c4999e36e59711a1aa5fed916640e10ca19e005df0ea.jpg) +Figure 2. Pipeline of SigNet. The sparse depth data is initially filled to create a coarse depth map. We then utilize the degradation assumption in Eq. (1) to establish a connection between this coarse map and the target prediction. The color image features and degradation representations are employed to guide the multi-modal fusion through our proposed DADF module, as depicted in Fig. 3. + +blind image super-resolution. Similarly, Zhou et al. [64] propose to estimate degradation for each local image region using a degradation-adaptive regression filter for blind super-resolution. Liang et al. [18] present a degradation-adaptive model that specifies the network parameters by estimating the degradation. Furthermore, several image restoration methods [17, 57, 58] have been proposed to simultaneously address multiple forms of degradation. For example, Zhang et al. [57] develop a multi-degradation image restoration network that incrementally learns degradation representations via clustering, without relying on prior knowledge of the degradation information. Zhang et al. [58] build an ingredient-oriented degradation reformulation network to model the relationship among different images. Additionally, Yao et al. [54] introduce a neural degradation representation learning method that captures latent degradation features, adaptively decomposing various types of degradation to enable comprehensive restoration. + +# 3. Method + +Overview. Fig. 2 shows our SigNet pipeline that consists of the depth branch, degradation branch, and image branch. In the depth branch, the sparse depth $\mathbf{S}$ is first filled via nonCNN densification tools [14, 15], obtaining a coarse depth map $\mathbf{Z}$ . According to Eq. (1), we employ denoising operators to weaken the negative impact of the noise $\mathbf{n}$ , yielding $\hat{\mathbf{Z}}$ . Then, we use residual groups [60] to extract features of the denoised depth $\hat{\mathbf{Z}}$ and the color image $\mathbf{I}$ in the image branch. In the degradation branch, the implicit degradation + +$\mathbf{D}$ is generated from a degradation network, where convolutions and residual groups are used. We thus utilize the image feature and degradation $\mathbf{D}$ to facilitate RGB-D fusion via our proposed Degradation-Aware Decomposition and Fusion (DADF). In summary, our SigNet redefines the depth completion as depth enhancement through the degradation assumption in Eq. (1), making the model free from the challenges of mismatch and ambiguity. + +# 3.1. Task Switching + +Degradation learning is widely applied in low-level image restoration [43, 55, 59], yielding impressive results. These tasks exhibit two key characteristics: 1) both input and target data are typically dense, and 2) the input data generally have a lower resolution than the target data. Actually, the low-resolution input data are typically upsampled using interpolation methods, such as bicubic interpolation. Consequently, the degradation matrix usually consists of a convolution whose kernel is derived from degradation representations and a downsampling function. In contrast, the depth completion task takes sparse depth data as input, and the resolutions of the input and target data are the same. Thus, we reframe depth completion as depth enhancement by drawing an analogy to image restoration. + +As shown in Fig. 2, we first fill the sparse depth $\mathbf{S}$ using non-CNN densification tools [14, 15], denoted as $f_{ds}(\cdot)$ , resulting in a coarse yet dense depth $\mathbf{Z}$ , formulated as: + +$$ +\mathbf {Z} = f _ {d s} (\mathbf {S}). \tag {2} +$$ + +As discussed in Sec. 1, this approach eliminates the mis + +![](images/7c0d8234f69e1a252db2271431e620e0e785ffc4e9e5184fa9352b8cc68fa270.jpg) +Figure 3. Overview of our DADF. $\odot$ and CM refer to elementwise multiplication and the conditional Mamba, respectively. + +match and ambiguity caused by direct convolution over irregularly sampled sparse data. Subsequently, we employ a degradation network to estimate the implicit degradation $\mathbf{D}$ . The degradation kernel $\mathbf{K}$ is produced from $\mathbf{D}$ using convolutions. Consequently, given the coarse depth $\mathbf{Z}$ and target depth $\mathbf{Y}$ , Eq. (1) can be rewritten based on the task switching as follows: + +$$ +\mathbf {Z} = \mathbf {K} (\mathbf {D}; \vartheta) \otimes \mathbf {Y} + \mathbf {n}, \tag {3} +$$ + +where $\otimes$ indicates convolution operation, $\vartheta$ is the network parameter. Since the resolutions are the same, our degradation equation does not need the downsampling function. + +# 3.2. Degradation-Aware Decomposition and Fusion + +As we know, degradation typically occurs near edges, suggesting that this cue can be leveraged to enhance RGB-D fusion, especially considering that the coarse depth often lacks significant edge details, while the RGB image contains abundant high-frequency components. Thus, we propose DADF that is shown in Fig. 3. We denote the image and depth features as $\mathbf{F}_{img}$ and $\mathbf{F}_{dep}$ , obtaining: + +$$ +\mathbf {F} _ {i m g} = f _ {e} (\mathbf {I}), \tag {4} +$$ + +$$ +\mathbf {F} _ {d e p} = f _ {e} \left(f _ {d n} (\mathbf {Z})\right), \tag {5} +$$ + +where $f_{e}(\cdot)$ and $f_{dn}(\cdot)$ refer to the feature extraction network and denoising function, respectively. As presented in Fig. 3, DADF takes as input the image feature $\mathbf{F}_{img}$ , depth feature $\mathbf{F}_{dep}$ , and degradation $\mathbf{F}_D$ . + +Decomposition. To extract the most representative components of $\mathbf{F}_D$ , we first apply the discrete cosine transform (DCT) to convert it into the frequency domain. Next, it adaptively learns a spectrum mask to enhance informative frequency bands while suppressing irrelevant ones. Finally, we reconstruct the spatial representation using the inverse DCT. This process is formulated as: + +$$ +\left\{\mathbf {D} _ {i} \mid i = 1, 2, 3 \right\} = f _ {d c t} (\mathbf {D}). \tag {6} +$$ + +# Algorithm 1: Conditional SSM Block + +Input: $\mathbf{F}_{rgbd}$ , $\{\mathbf{D}_i|i = 1,2,3\} :(\mathrm{HW},\mathrm{C})$ + +Output: $\hat{\mathbf{F}}_{rgbd}:(\mathrm{HW},\mathrm{C})$ + +1 A : (C, N) $\leftarrow$ ParameterA +2 $\mathbf{B}:(\mathrm{HW},\mathrm{N})\gets \mathbf{Linear}_B(f_{gru}(\mathbf{F}_{rgbd},\{\mathbf{D}_i|i = 1,2,3\}))$ +3 $\mathbf{C}:(\mathrm{HW},\mathrm{N})\gets \mathbf{Linear}_C(f_{gru}(\mathbf{F}_{rgbd},\{\mathbf{D}_i|i = 1,2,3\}))$ +4 $\Delta$ : (HW, C) $\leftarrow$ $\log(1 + \exp(\text{Linear}_\Delta(f_{gru}(\mathbf{F}_{rgbd}, \{\mathbf{D}_i | i = 1, 2, 3\})) + \text{Parameter}_\Delta))$ +5 $\overline{\mathbf{A}}: (\mathrm{HW}, \mathrm{C}, \mathrm{N}) \leftarrow \exp(\Delta \otimes \mathbf{A})$ +6 $\overline{\mathbf{B}}:\mathrm{(HW,C,N)}\gets \pmb {\Delta}\otimes \mathbf{B}$ +7 $\hat{\mathbf{F}}_{rgbd} \gets \mathrm{SSM}(\overline{\mathbf{A}},\overline{\mathbf{B}},\mathbf{C})(\mathbf{F}_{rgbd})$ +8 return $\hat{\mathbf{F}}_{rgbd}$ + +We apply concatenation, convolution, pooling, and Sigmoid operations, collectively denoted as $f_{m}(\cdot)$ , to generate a mask $m$ . This mask is then multiplied with the combined frequency components to produce an updated mask $\hat{m}$ , enabling the adaptive selection of the representative image feature $\hat{\mathbf{F}}_{img}$ . The process is defined as: + +$$ +m = f _ {m} \left(\mathbf {F} _ {\text {i m g}}, \mathbf {D}\right), \tag {7} +$$ + +$$ +\hat {m} = f _ {i d c t} \left(m \cdot \left(f _ {a c} \left(\sum_ {i = 1} ^ {3} \mathbf {D} _ {i}\right)\right)\right), \tag {8} +$$ + +$$ +\hat {\mathbf {F}} _ {i m g} = \hat {m} \cdot \mathbf {F} _ {i m g}, \tag {9} +$$ + +where $f_{idct}$ refers to inverse DCT and $f_{ac}(\cdot)$ represents the combination function of addition and convolutions. + +Fusion. As a result, we further leverage the decomposed degradation $\mathbf{D}_i$ to enable effective RGB-D fusion. It is well-known that Mamba [10] is good at handling long-range correlations. Inspired by recent Vision Mamba [25], we present a variant, i.e., Conditional Mamba (CM). As shown in Fig. 3, the selected image feature $\hat{\mathbf{F}}_{img}$ and raw depth feature $\mathbf{F}_{dep}$ are concatenated into CM, where the decomposed degradation $\mathbf{D}_i$ serves as the condition: + +$$ +\mathbf {F} _ {r g b d} = f _ {c c} \left(\tilde {\mathbf {F}} _ {i m g}, \mathbf {F} _ {d e p}\right), \tag {10} +$$ + +$$ +\hat {\mathbf {F}} _ {r g b d} = f _ {c m} \left(\mathbf {F} _ {r g b d}, \left\{\mathbf {D} _ {i} \mid i = 1, 2, 3 \right\}\right), \tag {11} +$$ + +where $f_{cc}$ denotes the combination function of concatenation and convolutions, $f_{cm}$ is the Condition Mamba function. As reported in Alg. 1, the state parameters $\mathbf{B}$ and $\mathbf{C}$ are calculated based on both the RGB-D input and degradation using gated recurrent convolution units. + +# 3.3. Loss Function + +As shown in our pipeline, the total loss function $\mathcal{L}_t$ consists of a reconstruction loss $\mathcal{L}_r$ and a self-supervised degradation learning loss $\mathcal{L}_d$ . Let $\hat{\mathbf{Y}}$ and $\mathbf{Y}$ denote the ground truth + +
DatasetMethodParams. (M) ↓RMSE (mm) ↓REL ↓δ1(%) ↑δ2(%) ↑δ3(%) ↑Venue
NYUv2Bilateral [1]-5320.13285.193.595.9CVPR 2005
S2D [28]26.12300.05494.597.398.9ICRA 2019
CSPN [4]17.41730.02096.398.699.5ECCV 2018
DfuseNet [34]-1560.01698.899.799.9ITSC 2019
DM-LRN [33]-2050.01498.899.699.9IROS 2021
RDF-GAN [42]-1390.01398.799.699.9CVPR 2022
AGG-Net [2]129.1920.01499.499.9100.0ICCV 2023
BPNet [38]89.9880.01199.699.9100.0CVPR 2024
SigNet (ours)3.3830.01299.799.9100.0CVPR 2025
DIMLBilateral [1]-6360.18983.088.892.4CVPR 2005
CSPN [4]17.41620.03396.198.799.6ECCV 2018
DfuseNet [34]-1430.02398.499.499.9ITSC 2019
DM-LRN [33]-1490.01599.099.699.9IROS 2021
AGG-Net [2]129.1780.01199.699.9100.0ICCV 2023
BPNet [38]89.9680.01199.499.8100.0CVPR 2024
SigNet (ours)3.3550.00699.799.9100.0CVPR 2025
SUN RGBDS2D [28]26.13290.07493.997.098.1ICRA 2019
CSPN [4]17.42950.13795.697.598.4ECCV 2018
DLiDAR [31]53.42790.06196.998.098.4CVPR 2019
DM-LRN [33]-2670.06397.698.298.7IROS 2021
RDF-GAN [42]-2550.05996.998.499.0CVPR 2022
AGG-Net [2]129.11520.03898.599.099.4ICCV 2023
BPNet [38]89.91240.02598.999.499.8CVPR 2024
SigNet (ours)3.31160.02299.099.499.8CVPR 2025
TOFDCCSPN [4]17.42240.04294.595.396.5ECCV 2018
GuideNet [37]73.51460.03097.698.999.5TIP 2020
PENet [11]131.52410.04394.695.395.5ICRA 2021
NLSPN [29]25.81740.02996.497.998.9ECCV 2020
CFounder [61]83.51130.02999.199.699.9CVPR 2023
RigNet [48]65.21330.02597.699.199.7ECCV 2022
GraphCSPN [24]26.42530.05292.096.998.7ECCV 2022
PointDC [56]25.11090.02198.599.299.6ICCV 2023
TPVD [50]31.2920.01499.199.699.9CVPR 2024
SigNet (ours)3.3870.01299.299.799.9CVPR 2025
+ +Table 1. Quantitative comparison on four depth completion datasets. The best and suboptimal results are highlighted. Note that most of the results on NYUv2, DIML, and SUN RGBD are borrowed directly from AGG-Net [2], while those of TOFDC are from TPVD [50]. + +depth and the predicted depth, respectively. The reconstruction loss $\mathcal{L}_r$ is defined as: + +$$ +\mathcal {L} _ {r} = \frac {1}{N} \sum_ {i = 1} ^ {N} | | \hat {\mathbf {Y}} _ {i} - \mathbf {Y} _ {i} | | _ {1}, \tag {12} +$$ + +where $N$ represents the number of training samples, and $\| \cdot \| _1$ denotes the $L_{1}$ norm. Subsequently, we introduce the degradation loss $\mathcal{L}_d$ to facilitate the degradation learning process through: + +$$ +\mathcal {L} _ {d} = \frac {1}{N} \sum_ {i = 1} ^ {N} \left| \left| \mathbf {K} \otimes \mathbf {Y} - \mathbf {Z} \right| \right| _ {1}, \tag {13} +$$ + +where $\mathbf{K}$ refers to the degradation kernel learned from the implicit degradation representation $\mathbf{D}$ . $\mathbf{Z}$ corresponds to the coarse depth. $\otimes$ denotes the convolution operation, where + +$\mathbf{K}$ serves as the convolution kernel. As a result, the total training loss is formulated as: + +$$ +\mathcal {L} _ {t} = \mathcal {L} _ {r} + \lambda \mathcal {L} _ {d}, \tag {14} +$$ + +where $\lambda$ is a hyper-parameter, empirically set to 0.1. By the constraint of the auxiliary loss $\mathcal{L}_d$ , our model can build additional connections between the input and output, thus predicting more accurate depth results. + +# 4. Experiment + +We evaluate our SigNet method on four popular datasets, including NYUv2, DIML, SUN RGBD, and TOFDC benchmarks. For NYUv2, DIML, and SUN RGBD, we maintain the settings used in AGG-Net [2], while for TOFDC, we follow the settings of TPVD [50]. Refer to our supplementary material for additional details on the experimental setups. + +![](images/09e2b28f93e8421502ee794244836a54580f10db81a9fb4a6e4e5ab74e6bf353.jpg) +(a) RGB + +![](images/6f96e3f1244f41c55d4c841468fdd4b7e96c91d675d80575c6bda5e03ad1c750.jpg) +(b) NLSPN + +![](images/f464526efacec3d8e3a64f967d67d4939157e57e04901401248b286c18d11759.jpg) +(c) RigNet + +![](images/a1f283232dc99d1999c7033b1638338683da706908622868c1d3ff483f8b3e76.jpg) +(d) SigNet + +![](images/2655180d87c800e552ac88a50546c20277dba44f1befcc3b4f702de33425ca42.jpg) +(e) GT + +![](images/8bfff01cb3e7aac297369b169baabfb607cf9a47496a4543b3c4386f4b3ca229.jpg) +Figure 4. Visual comparison with SOTA methods on TOFDC dataset, including NLSPN [29], RigNet [48], and our SigNet. +(a) RGB + +![](images/fcf0d73704f99531662243d392dd0e0ec29df5431bdf265d2e110fd6a1ff2e7e.jpg) +(b)CSPN + +![](images/c1719cb69f606f6d26cba0fb05765d7ca0ba26ee7b10e2ecfcd3c95dd06288b3.jpg) +(c) AGG-Net + +![](images/7ec5dd05e3aadc8c97a32bf2624a6465f77147c52b528cbbf3af93088595db63.jpg) +(d) SigNet + +![](images/0cf62c4d48a96b8ec265c3ad71ad81d45f673267f400c22e3e6c685196c80f35.jpg) +(e) GT + +![](images/35ad799d658fde72b175e426b03d3a612be886f8910aa84baeb3b930251f8405.jpg) +Figure 5. Visual comparison with SOTA approaches on DIML dataset, including CSPN [4], AGG-Net [2], and our SigNet. +RGB +Figure 6. Visual examples of filled coarse depth maps and degradation representations on the DIML dataset. + +![](images/015771f92579e5009873de3f94b2cd7c44c8fbdbddade9c11e149e6c742f09e0.jpg) +Coarse depth + +![](images/7ab8df37afd80c775888b1080b3f672b48440efb273378472f51c967366030a3.jpg) +Degradation + +![](images/ddfaa1a958b53be0a288ed96f85c9c1b3242c29b066e904f890b87245fcae813.jpg) +RGB + +![](images/afd2c3b49df17bf5ed6643dd14e73146d16f8e572bc8d9491ea2f73202135ca0.jpg) +Coarse depth + +![](images/6b65f37cd96e75aebc7d20ba2a1e2f01ba8e7a367b5d65d3dcf7e9fc3d47ebc3.jpg) +Degradation + +# 4.1. Comparison with SOTA + +To assess the effectiveness of our SigNet, we benchmark it against the traditional bilateral filtering method [1] and some recent SOTA deep learning approaches such as [2, 28, 29, 34, 38]. The error and accuracy results for the competing models are sourced from their original publications or obtained by retraining them following the default settings. + +# 4.1.1. Performance Analysis + +On NYUv2. As a widely used benchmark dataset, NYUv2 is often employed to evaluate the effectiveness of models. From the first row of Table 1, we can observe that SigNet achieves competitive performance on the NYUv2 dataset, with an RMSE that is $5mm$ lower than the second-best method. Additionally, although our method lags behind BPNet [38] by only 0.001 in the REL metric, SigNet signifi + +cantly reduces the number of parameters by $96\%$ . + +On DIML. DIML is collected from both indoor and outdoor scenes, suffering from severe edge shadows and irregular holes. Our method is implemented using the same experimental settings as previous approaches [2, 33]. As shown in the second row of Table 1, SigNet demonstrates the best performance across all evaluation metrics. For example, compared to the suboptimal model, our approach decreases the RMSE by $19\%$ and the REL by $45\%$ . + +On SUN RGBD. The large-scale SUN RGB-D dataset is leveraged to evaluate the generalization capability of our SigNet across different sensors and scenes. Table 1 (third row) reveals that our method surpasses previous state-of-the-art approaches across all metrics. Notably, SigNet achieves a $8mm$ lower RMSE than the latest BPNet [38] and a $36mm$ lower RMSE compared to AGG-Net [2]. + +
Density1%5%35%65%95%
TPVD [50]156.6139.7112.297.392.8
SigNet144.6121.599.490.184.3
Improv. ↑12.018.212.87.28.5
+ +![](images/fb3a9a02651eeca5f3b0e2aeabcc78be927334df8256767d18eedc3ec1a38bb2.jpg) +Figure 7. Complexity comparison on the TOFDC dataset, where larger circle areas indicate higher inference time. + +On TOFDC. TOFDC encompasses extensive data captured under various lighting conditions and in open spaces. It is utilized to further assess the performance of the proposed model. The quantitative results in the last row of Table 1 indicate that our method achieves state-of-the-art performance. Compared to the second-best TPVD [50], SigNet reduces the RMSE by $5mm$ and the REL by 0.002, while also decreasing the number of parameters by $27.9M$ . + +# 4.1.2. Visualization Analysis + +Figs. 4 and 5 showcase visual results on TOFDC and DIML, highlighting SigNet's ability to recover precise depth structures. Compared to previous models, our method yields sharper predictions, such as the leaves in Fig. 4 and the chair in Fig. 5, achieving closer alignment with the ground truth. These results affirm the effectiveness of our approach in enhancing depth completion. Additionally, Fig. 6 presents some degradation examples, further demonstrating the ability of SigNet to model and address such challenges. + +# 4.1.3. Number of Valid Points + +Tab. 2 lists a comparison of different valid points on DIML. It indicates that our SigNet consistently outperforms TPVD, achieving an average improvement of $9.7\%$ across the five densities. These results highlight the generalization capability of our method across varying valid points. + +# 4.1.4. Complexity Analysis + +Fig. 7 shows that our method achieves an excellent balance among RMSE, parameters, and inference time. For exam- + +Table 2. Density comparison of TPVD and SigNet on TOFDC. + +
SigNetDeg. BridgeRGB-D FusionRMSE (mm)
DensifyDenoiseConc.Att.DADF-DDADF-F
i108
ii✓✓95
iii✓✓✓92
iv✓✓✓74
v✓✓✓68
v✓✓✓✓55
+ +Table 3. Ablation study of our SigNet on the DIML dataset. Conc.: concat, Att.: attention, DADF-D: degradation decomposition in DADF module, DADF-F: conditional Mamba fusion. + +ple, compared to the fewer parameter CSPN [4], our SigNet not only reduces parameters by $81\%$ and inference time by $82\%$ , but also decreases RMSE by $61\%$ . Additionally, our method surpasses the latest TPVD [50] with a $5\%$ lower RMSE and $78\%$ lower FPS, while reducing parameters by $89\%$ and inference time by $78\%$ . + +# 4.2. Ablation Study + +SigNet Designs. Table 3 presents the ablation results on the DIML test split. The iteration of DADF is set to five. SigNet-i serves as the baseline, where both the degradation bridge and DADF module are removed, and the RGB-D features are fused using addition. Building on this, SigNet-ii redefines depth completion as depth enhancement through densification and degradation, concatenating the RGB and degradation features, resulting in a $13\mathrm{mm}$ improvement. This highlights the superiority of our task transformation approach. Additionally, SigNet-iii incorporates a Gaussian denoising operator to further reduce error. To better utilize the degradation, SigNet-iv applies a Sigmoid function and element-wise multiplication to obtain an attention map, enabling adaptive selection of RGB features. Consequently, performance improves significantly from $92\mathrm{mm}$ to $74\mathrm{mm}$ . Furthermore, SigNet-v employs degradation-aware decomposition, reducing the error by an additional $6\mathrm{mm}$ . The results of SigNet-iii, iv, and v demonstrate the effectiveness of our adaptive feature selection strategy. Notably, in SigNet-iii, iv, and v, the selected RGB and depth features are encoded through concatenation and convolutions. Finally, with the introduction of our conditional Mamba, SigNet-vi surpasses SigNet-v by $13\mathrm{mm}$ in RMSE, representing a significant improvement. Overall, each component positively impacts the baseline. + +Mamba Fusion and Iteration of DADF. Fig. 8 (a) illustrates the ablation study of Mamba fusion on the DIML dataset. For rapid validation, we employ DADF only once. The baseline ViM directly encodes the combined RGB-D features using the vanilla Vision Mamba [25]. When we apply our Conditional Mamba with initial degradation and RGB-D features as input, the error decreases by $8\mathrm{mm}$ . Additionally, the introduction of degradation decomposition + +![](images/8ca0db0eb846cd175d65914b9fbdb356cc3241ac68fdaff946507be3ff0f360e.jpg) +Figure 8. Ablation study of our DADF on the DIML dataset. + +
Path number357911
RMSE (mm)55.054.654.554.154.0
+ +Table 4. Ablation on the number of decomposition path in DADF. + +using DCT further enhances performance. We also observe that the results improve as the number of DCT components increases, highlighting the effectiveness of our gradual DCT decomposition design. Fig. 8 (b) presents the ablation study of our DADF with different iterations on the DIML dataset. We find that performance improves with an increasing number of iterations. For an optimal balance between efficiency and effectiveness, we set the final iteration count to five. Moreover, as shown in Fig.3, the degradation $\mathbf{D}$ can be decomposed into $n$ components using DCT. As indicated in Tab.4, the error gradually decreases as $n$ increases. However, considering the computational complexity, we select the top-3 most representative components to focus on the high-frequency components of $\mathbf{F}_{img}$ and update the Mamba condition accordingly. + +# 4.3. Failure Case + +Our method is primarily evaluated in indoor scenes and some simple outdoor scenes with closer distances. By constraining dense predictions with coarse depth maps, SigNet can effectively predict their degradation representations. However, as depicted in Fig. 9, SigNet fails to estimate reasonable degradation on the outdoor KITTI benchmark [40]. We identify two main reasons for this: + +1) The ground truth depth maps of KITTI are very sparse, with only about $30\%$ valid pixels. This sparsity leads to inconsistent depth edges, resulting in a relatively weak constraint. For instance, the degradation in Fig. 9 even introduces some initial LiDAR lines. +2) The depth distance of KITTI data is much larger, which significantly weakens edge details. As observed, the degradation in Fig. 9 can only learn very coarse and blurry representations near edges at short range, while at long range, it learns incorrect representations. + +One potential solution is to provide the model with supervisory signals regarding dense contents and precise + +![](images/6e2213f0f81a15f97da16e501338e98fef2ad3febb47a980a081f121657ae439.jpg) +Figure 9. Visual results of failure cases on KITTI benchmark [40]. DA v2 denotes the large depth model DepthAnything v2 [52]. + +edges. It is well-known that large depth models can produce dense depth with impressive sharp edges. For example, we employ DepthAnything v2 [52] to generate the corresponding dense depth. We discover that the edges of cars and even tree leaves are much clearer. Although the depth scale does not match that of KITTI, we can generate a map that offers constraints near edges via gradient or boundary extraction techniques [46]. Certainly, the results of DepthAnything v2 [52] or DepthAnything v2 itself can also be utilized to facilitate depth completion, warranting further exploration. + +# 5. Conclusion + +In this paper, we redefined depth completion as depth enhancement for the first time with our novel framework, SigNet. SigNet leveraged non-CNN densification tools to transform sparse depth data into coarse yet dense depth, fully eliminating the mismatch and ambiguity caused by direct convolution over irregularly sampled sparse data. More importantly, by establishing a self-supervised degradation bridge, SigNet effectively fused RGB-D data. It utilized degradation decomposition to dynamically select high-frequency components of RGB and build a conditional Mamba for the depth enhancement. Extensive experiments validated the superiority of SigNet, highlighting its potential for depth related applications. + +# Acknowledgements + +This work was supported by the National Natural Science Fund of China (U24A20330, 62361166670 and 62072242). + +# References + +[1] Antoni Buades, Bartomeu Coll, and J-M Morel. A non-local algorithm for image denoising. In CVPR, pages 60-65, 2005. 5, 6 +[2] Dongyue Chen, Tingxuan Huang, Zhimin Song, Shizhuo Deng, and Tong Jia. Agg-net: Attention guided gated convolutional network for depth image completion. In ICCV, pages 8853-8862, 2023. 5, 6, 2 +[3] Yun Chen, Bin Yang, Ming Liang, and Raquel Urtasun. Learning joint 2d-3d representations for depth completion. In ICCV, pages 10023–10032, 2019. 2 +[4] Xinjing Cheng, Peng Wang, and Ruigang Yang. Learning depth with convolutional spatial propagation network. In ECCV, pages 103-119, 2018. 5, 6, 7 +[5] Xinjing Cheng, Peng Wang, and Ruigang Yang. Depth estimation via affinity learned with convolutional spatial propagation network. In ECCV, pages 103-119, 2018. 2 +[6] Xinjing Cheng, Peng Wang, and Ruigang Yang. Learning depth with convolutional spatial propagation network. IEEE Transactions on Pattern Analysis and Machine Intelligence, 42(10):2361-2379, 2019. 2 +[7] Xinjing Cheng, Peng Wang, Chenye Guan, and Ruigang Yang. Cspn++: Learning context and resource aware convolutional spatial propagation networks for depth completion. In AAAI, pages 10615-10622, 2020. 2 +[8] Jaehoon Cho, Dongbo Min, Youngjung Kim, and Kwanghoon Sohn. A large rgb-d dataset for semi-supervised monocular depth estimation. arXiv preprint arXiv:1904.10230, 2019. 2 +[9] Abdelrahman Eldisokey, Michael Felsberg, and Fahad Shahbaz Khan. Confidence propagation through cnns for guided sparse depth regression. IEEE Transactions on Pattern Analysis and Machine Intelligence, 42(10):2423-2436, 2020. 2 +[10] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023. 4 +[11] Mu Hu, Shuling Wang, Bin Li, Shiyu Ning, Li Fan, and Xiaojin Gong. Penet: Towards precise and efficient image guided depth completion. In ICRA, 2021. 2, 5 +[12] Maximilian Jaritz, Raoul De Charette, Emilie Wirbel, Xavier Perrotton, and Fawzi Nashashibi. Sparse and dense data with cnns: Depth completion and semantic segmentation. In 3DV, pages 52-60, 2018. 2 +[13] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In Computer Ence, 2014. 2 +[14] Jason Ku, Ali Harakeh, and Steven L Waslander. In defense of classical image processing: Fast depth completion on the cpu. In CRV, pages 16-22, 2018. 1, 2, 3 +[15] Anat Levin, Dani Lischinski, and Yair Weiss. Colorization using optimization. In SIGGRAPH, pages 689-694, 2004. 1, 2, 3 +[16] Ang Li, Zejian Yuan, Yonggen Ling, Wanchao Chi, Chong Zhang, et al. A multi-scale guided cascade hourglass network for depth completion. In WACV, pages 32-40, 2020. 2 + +[17] Boyun Li, Xiao Liu, Peng Hu, Zhongqin Wu, Jiancheng Lv, and Xi Peng. All-in-one image restoration for unknown corruption. In CVPR, pages 17452-17462, 2022. 3 +[18] Jie Liang, Hui Zeng, and Lei Zhang. Efficient and degradation-adaptive network for real-world image superresolution. In ECCV, pages 574-591, 2022. 3 +[19] Yuankai Lin, Tao Cheng, Qi Zhong, Wending Zhou, and Hua Yang. Dynamic spatial propagation network for depth completion. In AAAI, pages 1638-1646, 2022. 2 +[20] Yuankai Lin, Hua Yang, Tao Cheng, Wending Zhou, and Zhouping Yin. Dyspn: Learning dynamic affinity for image-guided depth completion. IEEE Transactions on Circuits and Systems for Video Technology, pages 1-1, 2023. 2 +[21] Lina Liu, Yiyi Liao, Yue Wang, Andreas Geiger, and Yong Liu. Learning steering kernels for guided depth completion. IEEE Transactions on Image Processing, 30:2850-2861, 2021. 1 +[22] Lina Liu, Xibin Song, Xiaoyang Lyu, Junwei Diao, Mengmeng Wang, Yong Liu, and Liangjun Zhang. Fcfr-net: Feature fusion based coarse-to-fine residual learning for depth completion. In AAAI, pages 2136-2144, 2021. 1, 2 +[23] Sifei Liu, Shalini De Mello, Jinwei Gu, Guangyu Zhong, Ming-Hsuan Yang, and Jan Kautz. Learning affinity via spatial propagation networks. In NeurIPS, 2017. 2 +[24] Xin Liu, Xiaofei Shao, Bo Wang, Yali Li, and Shengjin Wang. Graphcspn: Geometry-aware depth completion via dynamic gcns. In ECCV, pages 90-107. Springer, 2022. 2, 5 +[25] Yue Liu, Yunjie Tian, Yuzhong Zhao, Hongtian Yu, Lingxi Xie, Yaowei Wang, Qixiang Ye, and Yunfan Liu. Vmamba: Visual state space model. arXiv preprint arXiv:2401.10166, 2024. 2, 4, 7 +[26] Kaiyue Lu, Nick Barnes, Saeed Anwar, and Liang Zheng. From depth what can you see? depth completion via auxiliary image reconstruction. In CVPR, pages 11306-11315, 2020. 2 +[27] Fangchang Ma and Sertac Karaman. Sparse-to-dense: Depth prediction from sparse depth samples and a single image. In ICRA, pages 4796-4803. IEEE, 2018. 1 +[28] Fangchang Ma, Guilherme Venturelli Cavalheiro, and Sertac Karaman. Self-supervised sparse-to-dense: Self-supervised depth completion from lidar and monocular camera. In ICRA, 2019. 2, 5, 6 +[29] Jinsun Park, Kyungdon Joo, Zhe Hu, Chi-Kuei Liu, and In So Kweon. Non-local spatial propagation network for depth completion. In ECCV, 2020. 2, 5, 6 +[30] Jin-Hwi Park, Chanhwi Jeong, Junoh Lee, and Hae-Gon Jeon. Depth prompting for sensor-agnostic depth estimation. In CVPR, pages 9859–9869, 2024. 2 +[31] Jiaxiong Qiu, Zhaopeng Cui, Yinda Zhang, Xingdi Zhang, Shuaicheng Liu, Bing Zeng, and Marc Pollefeys. Deep surface normal guided depth prediction for outdoor scene from sparse lidar data and single color image. In CVPR, pages 3313-3322, 2019. 5, 2 +[32] Kyeongha Rho, Jinsung Ha, and Youngjung Kim. Guideformer: Transformers for image guided depth completion. In CVPR, pages 6250-6259, 2022. 2 + +[33] Dmitry Senushkin, Mikhail Romanov, Ilia Belikov, Nikolay Patakin, and Anton Konushin. Decoder modulation for indoor depth completion. In IROS, pages 2181-2188. IEEE, 2021. 5, 6 +[34] Shreyas S Shivakumar, Ty Nguyen, Ian D Miller, Steven W Chen, Vijay Kumar, and Camillo J Taylor. Dfusenet: Deep fusion of rgb and sparse depth information for image guided dense depth completion. In ITSC, pages 13-20. IEEE, 2019. 5, 6 +[35] Nathan Silberman, Derek Hoiem, Pushmeet Kohli, and Rob Fergus. Indoor segmentation and support inference from rgbd images. In ECCV, pages 746-760. Springer, 2012. 2 +[36] Shuran Song, Samuel P Lichtenberg, and Jianxiong Xiao. Sun rgb-d: A rgb-d scene understanding benchmark suite. In CVPR, pages 567-576, 2015. 1, 2 +[37] Jie Tang, Fei-Peng Tian, Wei Feng, Jian Li, and Ping Tan. Learning guided convolutional network for depth completion. IEEE Transactions on Image Processing, 30:1116-1129, 2020. 2, 5 +[38] Jie Tang, Fei-Peng Tian, Boshi An, Jian Li, and Ping Tan. Bilateral propagation network for depth completion. In CVPR, pages 9763-9772, 2024. 1, 2, 5, 6 +[39] Diego Thomas and Akihiro Sugimoto. A flexible scene representation for 3d reconstruction using an rgb-d camera. In ICCV, pages 2800-2807, 2013. 1 +[40] Jonas Uhrig, Nick Schneider, Lukas Schneider, Uwe Franke, Thomas Brox, and Andreas Geiger. Sparsity invariant cnns. In 3DV, pages 11–20, 2017. 1, 2, 8 +[41] Wouter Van Gansbeke, Davy Neven, Bert De Brabandere, and Luc Van Gool. Sparse and noisy lidar completion with rgb guidance and uncertainty. In MVA, pages 1-6, 2019. 2 +[42] Haowen Wang, Mingyuan Wang, Zhengping Che, Zhiyuan Xu, Xiuquan Qiao, Mengshi Qi, Feifei Feng, and Jian Tang. RGB-depth fusion gan for indoor depth completion. In CVPR, pages 6209-6218, 2022. 5 +[43] Longguang Wang, Yingqian Wang, Xiaoyu Dong, Qingyu Xu, Jungang Yang, Wei An, and Yulan Guo. Unsupervised degradation representation learning for blind superresolution. In CVPR, pages 10581-10590, 2021. 1, 2, 3 +[44] Yufei Wang, Bo Li, Ge Zhang, Qi Liu, Tao Gao, and Yuchao Dai. Lrru: Long-short range recurrent updating networks for depth completion. In ICCV, pages 9422-9432, 2023. 2 +[45] Yufei Wang, Ge Zhang, Shaoqian Wang, Bo Li, Qi Liu, Le Hui, and Yuchao Dai. Improving depth completion via depth feature upsampling. In CVPR, pages 21104-21113, 2024. 1, 2 +[46] Zhengxue Wang, Zhiqiang Yan, and Jian Yang. Sgnet: Structure guided network via gradient-frequency awareness for depth map super-resolution. In AAAI, pages 5823-5831, 2024. 8 +[47] Zhiqiang Yan, Xiang Li, Kun Wang, Zhenyu Zhang, Jun Li, and Jian Yang. Multi-modal masked pre-training for monocular panoramic depth completion. In ECCV, pages 378-395, 2022. 1 +[48] Zhiqiang Yan, Kun Wang, Xiang Li, Zhenyu Zhang, Jun Li, and Jian Yang. Rignet: Repetitive image guided network for depth completion. In ECCV, pages 214-230, 2022. 1, 2, 5, 6 + +[49] Zhiqiang Yan, Kun Wang, Xiang Li, Zhenyu Zhang, Jun Li, and Jian Yang. Desnet: Decomposed scale-consistent network for unsupervised depth completion. In AAAI, pages 3109-3117, 2023. 1 +[50] Zhiqiang Yan, Yuankai Lin, Kun Wang, Yupeng Zheng, Yufei Wang, Zhenyu Zhang, Jun Li, and Jian Yang. Tri-perspective view decomposition for geometry-aware depth completion. In CVPR, pages 4874-4884, 2024. 1, 2, 5, 7 +[51] Zhiqiang Yan, Yupeng Zheng, Deng-Ping Fan, Xiang Li, Jun Li, and Jian Yang. Learnable differencing center for nighttime depth perception. Visual Intelligence, 2(1):15, 2024. 1 +[52] Lihe Yang, Bingyi Kang, Zilong Huang, Zhen Zhao, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything v2. arXiv preprint arXiv:2406.09414, 2024. 8 +[53] Yanchao Yang, Alex Wong, and Stefano Soatto. Dense depth posterior (ddp) from single image and sparse range. In CVPR, pages 3353-3362, 2020. 2 +[54] Mingde Yao, Ruikang Xu, Yuanshen Guan, Jie Huang, and Zhiwei Xiong. Neural degradation representation learning for all-in-one image restoration. IEEE Transactions on Image Processing, 2024. 3 +[55] Guanghao Yin, Wei Wang, Zehuan Yuan, Wei Ji, Dongdong Yu, Shouqian Sun, Tat-Seng Chua, and Changhu Wang. Conditional hyper-network for blind super-resolution with multiple degradations. IEEE Transactions on Image Processing, 31:3949-3960, 2022. 2, 3 +[56] Zhu Yu, Zehua Sheng, Zili Zhou, Lun Luo, Si-Yuan Cao, Hong Gu, Huaqi Zhang, and Hui-Liang Shen. Aggregating feature point cloud for depth completion. In ICCV, pages 8732-8743, 2023. 2, 5 +[57] Cheng Zhang, Yu Zhu, Qingsen Yan, Jinqiu Sun, and Yanning Zhang. All-in-one multi-degradation image restoration network via hierarchical degradation representation. In ACM MM, pages 2285-2293, 2023. 3 +[58] Jinghao Zhang, Jie Huang, Mingde Yao, Zizheng Yang, Hu Yu, Man Zhou, and Feng Zhao. Ingredient-oriented multi-degradation learning for image restoration. In CVPR, pages 5825-5835, 2023. 3 +[59] Kai Zhang, Jingyun Liang, Luc Van Gool, and Radu Timofte. Designing a practical degradation model for deep blind image super-resolution. In ICCV, pages 4791-4800, 2021. 1, 2, 3 +[60] Yulun Zhang, Kunpeng Li, Kai Li, Lichen Wang, Bineng Zhong, and Yun Fu. Image super-resolution using very deep residual channel attention networks. In ECCV, pages 286-301, 2018. 3 +[61] Youmin Zhang, Xianda Guo, Matteo Poggi, Zheng Zhu, Guan Huang, and Stefano Mattoccia. Completionformer: Depth completion with convolutions and vision transformers. In CVPR, pages 18527-18536, 2023. 2, 5 +[62] Zhiwei Zhong, Xianming Liu, Junjun Jiang, Debin Zhao, and Xiangyang Ji. Guided depth map super-resolution: A survey. ACM Computing Surveys, 55(14s):1-36, 2023. 1, 2 +[63] Hanyu Zhou, Yi Chang, Wending Yan, and Luxin Yan. Unsupervised cumulative domain adaptation for foggy scene optical flow. In CVPR, pages 9569-9578, 2023. 1 + +[64] Hongyang Zhou, Xiaobin Zhu, Jianqing Zhu, Zheng Han, Shi-Xue Zhang, Jingyan Qin, and Xu-Cheng Yin. Learning correction filter via degradation-adaptive regression for blind single image super-resolution. In ICCV, pages 12365-12375, 2023. 3 +[65] Hanyu Zhou, Yi Chang, Zhiwei Shi, Wending Yan, Gang Chen, Yonghong Tian, and Luxin Yan. Adverse weather optical flow: Cumulative homogeneous-heterogeneous adaptation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 1 +[66] Wending Zhou, Xu Yan, Yinghong Liao, Yuankai Lin, Jin Huang, Gangming Zhao, Shuguang Cui, and Zhen Li. Bev@ dc: Bird's-eye view assisted training for depth completion. In CVPR, pages 9233-9242, 2023. 2 +[67] Ziyue Zhu, Qiang Meng, Xiao Wang, Ke Wang, Liujiang Yan, and Jian Yang. Curricular object manipulation in lidar-based object detection. In CVPR, pages 1125-1135, 2023. 1 +[68] Ziyue Zhu, Shenlong Wang, Jin Xie, Jiang-Jiang Liu, Jingdong Wang, and Jian Yang. Voxelsplat: Dynamic gaussian splatting as an effective loss for occupancy and flow prediction. In CVPR, 2025. 1 \ No newline at end of file diff --git a/2025/Completion as Enhancement_ A Degradation-Aware Selective Image Guided Network for Depth Completion/images.zip b/2025/Completion as Enhancement_ A Degradation-Aware Selective Image Guided Network for Depth Completion/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..c82f2ad1cdb20fbd548580108e10e2902944c72d --- /dev/null +++ b/2025/Completion as Enhancement_ A Degradation-Aware Selective Image Guided Network for Depth Completion/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d7b21772ec19367e4300521cb48d2a254e95d4b549dc3e1901b18081475f6eb3 +size 721261 diff --git a/2025/Completion as Enhancement_ A Degradation-Aware Selective Image Guided Network for Depth Completion/layout.json b/2025/Completion as Enhancement_ A Degradation-Aware Selective Image Guided Network for Depth Completion/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..9878997b996933f784bce0d64026fcfcb5ef9d98 --- /dev/null +++ b/2025/Completion as Enhancement_ A Degradation-Aware Selective Image Guided Network for Depth Completion/layout.json @@ -0,0 +1,10624 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 65, + 103, + 545, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 103, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 65, + 103, + 545, + 140 + ], + "type": "text", + "content": "Completion as Enhancement: A Degradation-Aware Selective Image Guided Network for Depth Completion" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 137, + 161, + 477, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 161, + 477, + 176 + ], + "spans": [ + { + "bbox": [ + 137, + 161, + 477, + 176 + ], + "type": "text", + "content": "Zhiqiang Yan Zhengxue Wang Kun Wang Jun Li\\* Jian Yang\\*" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 164, + 176, + 446, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 164, + 176, + 446, + 190 + ], + "spans": [ + { + "bbox": [ + 164, + 176, + 446, + 190 + ], + "type": "text", + "content": "PCA Lab†, Nanjing University of Science and Technology" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 171, + 191, + 436, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 171, + 191, + 436, + 203 + ], + "spans": [ + { + "bbox": [ + 171, + 191, + 436, + 203 + ], + "type": "text", + "content": "{yanzq,zxwang,kunwang,junli,csjyang}@njust.edu.cn" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "spans": [ + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 255, + 297, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 255, + 297, + 508 + ], + "spans": [ + { + "bbox": [ + 54, + 255, + 297, + 508 + ], + "type": "text", + "content": "In this paper, we introduce the Selective Image Guided Network (SigNet), a novel degradation-aware framework that transforms depth completion into depth enhancement for the first time. Moving beyond direct completion using convolutional neural networks (CNNs), SigNet initially densifies sparse depth data through non-CNN densification tools to obtain coarse yet dense depth. This approach eliminates the mismatch and ambiguity caused by direct convolution over irregularly sampled sparse data. Subsequently, SigNet redefines completion as enhancement, establishing a self-supervised degradation bridge between the coarse depth and the targeted dense depth for effective RGB-D fusion. To achieve this, SigNet leverages the implicit degradation to adaptively select high-frequency components (e.g., edges) of RGB data to compensate for the coarse depth. This degradation is further integrated into a multi-modal conditional Mamba, dynamically generating the state parameters to enable efficient global high-frequency information interaction. We conduct extensive experiments on the NYUv2, DIML, SUN RGBD, and TOFDC datasets, demonstrating the state-of-the-art (SOTA) performance of SigNet." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 528, + 135, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 528, + 135, + 540 + ], + "spans": [ + { + "bbox": [ + 55, + 528, + 135, + 540 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 548, + 295, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 548, + 295, + 620 + ], + "spans": [ + { + "bbox": [ + 55, + 548, + 295, + 620 + ], + "type": "text", + "content": "Depth completion [27, 38, 48, 50] aims to predict a dense depth map from sparse and noisy depth measurements, often using the corresponding RGB image for auxiliary guidance. The resulting dense depth map can benefit various downstream applications such as scene reconstruction [36, 39, 47, 68] and self-driving [22, 45, 51, 63, 65, 67]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 620, + 296, + 669 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 620, + 296, + 669 + ], + "spans": [ + { + "bbox": [ + 55, + 620, + 296, + 669 + ], + "type": "text", + "content": "The key challenges of this task lie in the sparse depth input data [21, 38, 40, 49, 50]. We summarize these challenges as mismatch and ambiguity. As shown in Fig. 1, the mismatch is caused by the irregularly sampled sparse" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 317, + 228, + 555, + 417 + ], + "blocks": [ + { + "bbox": [ + 317, + 228, + 555, + 417 + ], + "lines": [ + { + "bbox": [ + 317, + 228, + 555, + 417 + ], + "spans": [ + { + "bbox": [ + 317, + 228, + 555, + 417 + ], + "type": "image", + "image_path": "4319659526abef301e7b370531fede098225e97d949d8e8117db046c20b330a4.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 424, + 555, + 480 + ], + "lines": [ + { + "bbox": [ + 313, + 424, + 555, + 480 + ], + "spans": [ + { + "bbox": [ + 313, + 424, + 555, + 480 + ], + "type": "text", + "content": "Figure 1. Illustration of our main concept. It redefines depth completion as depth enhancement. We first transform sparse depth input via non-CNN densification tools [14, 15], yielding coarse but dense depth. Then we predict the precise and dense depth from the coarse depth by leveraging degradation assumption [43, 59, 62]." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 506, + 555, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 506, + 555, + 567 + ], + "spans": [ + { + "bbox": [ + 313, + 506, + 555, + 567 + ], + "type": "text", + "content": "depth with varying point distributions, while convolutions are spatially invariant with fixed kernel shapes. Moreover, it is ambiguous for convolutions to distinguish between valid and invalid depth measurements since unknown pixels in the sparse depth are recognized as zero." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "type": "text", + "content": "To address these challenges, several excellent methods [21, 38, 40] have been proposed. For instance, SparseCOnvs [40] proposes sparsity invariant CNNs that explicitly model the location of missing data. To avoid directly convolving on sparse depth, Liu et al. [21] employs kernels learned by differentiable kernel regression layers to interpolate sparse data. BPNet [38] develops bilateral propagation, where the coefficients are generated based on both radiometric difference and spatial distance. These data-driven schemes are robust and impressive. However, they still cannot fully avoid the application of convolution on incomplete depth data, even though the depth data has become denser." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 674, + 147, + 684 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 674, + 147, + 684 + ], + "spans": [ + { + "bbox": [ + 66, + 674, + 147, + 684 + ], + "type": "text", + "content": "*Corresponding authors" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 56, + 684, + 295, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 684, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 684, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 56, + 684, + 295, + 713 + ], + "type": "text", + "content": "PCA Lab, Key Lab of Intelligent Perception and Systems for High-Dimensional Information of Ministry of Education, School of Computer Science and Engineering, Nanjing University of Science and Technology." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26943" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "content": "In this paper, our objective is to develop a novel solution that not only completely resolves the issues of mismatch and ambiguity but also provides a fresh perspective on revisiting the traditional depth completion task." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 121, + 294, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 121, + 294, + 228 + ], + "spans": [ + { + "bbox": [ + 55, + 121, + 294, + 228 + ], + "type": "text", + "content": "To this end, we propose a novel framework called the Selective Image Guided Network (SigNet). As shown in Fig. 1, SigNet redefines depth completion as depth enhancement1. It first leverages non-CNN densification tools [14, 15] to obtain coarse but dense depth. This coarse depth is then enhanced by a CNN-based subnetwork. Inspired by previous approaches [59, 62], we assume that there exists unknown degradation " + }, + { + "bbox": [ + 55, + 121, + 294, + 228 + ], + "type": "inline_equation", + "content": "\\mathbf{h}" + }, + { + "bbox": [ + 55, + 121, + 294, + 228 + ], + "type": "text", + "content": " between the coarse depth " + }, + { + "bbox": [ + 55, + 121, + 294, + 228 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}" + }, + { + "bbox": [ + 55, + 121, + 294, + 228 + ], + "type": "text", + "content": " and the target depth " + }, + { + "bbox": [ + 55, + 121, + 294, + 228 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}" + }, + { + "bbox": [ + 55, + 121, + 294, + 228 + ], + "type": "text", + "content": ", described as" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 145, + 233, + 294, + 245 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 233, + 294, + 245 + ], + "spans": [ + { + "bbox": [ + 145, + 233, + 294, + 245 + ], + "type": "interline_equation", + "content": "\\mathbf {Z} = \\mathbf {h} \\mathbf {Y} + \\mathbf {n}, \\tag {1}", + "image_path": "15fab58835718701936df2815728e7e075feefe732ec1f0637ef7501a29eb7bd.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 251, + 295, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 251, + 295, + 501 + ], + "spans": [ + { + "bbox": [ + 56, + 251, + 295, + 501 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 56, + 251, + 295, + 501 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 56, + 251, + 295, + 501 + ], + "type": "text", + "content": " denotes additive noise. Therefore, we can establish a connection between the source depth and the target depth through the degradation. A fundamental observation is that degradation typically occurs near edges, suggesting that this phenomenon can be leveraged to enhance RGB-D fusion. This is particularly relevant given that the coarse depth often lacks significant edge details, while the RGB image contains abundant high-frequency components. Specifically, SigNet introduces a Degradation-Aware Decomposition and Fusion (DADF) module. DADF begins by decomposing the implicit degradation " + }, + { + "bbox": [ + 56, + 251, + 295, + 501 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 56, + 251, + 295, + 501 + ], + "type": "text", + "content": " into multiple patches and estimating their coefficients in the discrete cosine domain. Each patch is then used to adaptively select the most representative high-frequency information from the RGB image. To achieve effective RGB-D aggregation, DADF incorporates a conditional Mamba to encode long-range correlations. Unlike the vanilla Vision Mamba [25], which calculates state parameters from the input, DADF uses the degradation patches as additional conditions to generate the state parameters. This approach enables further global high-frequency information interaction." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 502, + 251, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 502, + 251, + 513 + ], + "spans": [ + { + "bbox": [ + 67, + 502, + 251, + 513 + ], + "type": "text", + "content": "In summary, our contributions are as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 514, + 295, + 669 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 56, + 514, + 295, + 585 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 514, + 295, + 585 + ], + "spans": [ + { + "bbox": [ + 56, + 514, + 295, + 585 + ], + "type": "text", + "content": "- We introduce a novel framework called SigNet, which redefines depth completion as depth enhancement for the first time. By adopting a degradation-aware approach, SigNet not only offers a new perspective to the field but also fully eliminates the traditional challenges such as mismatch and ambiguity." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 586, + 295, + 669 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 586, + 295, + 669 + ], + "spans": [ + { + "bbox": [ + 55, + 586, + 295, + 669 + ], + "type": "text", + "content": "- We propose a Degradation-Aware Decomposition and Fusion (DADF) module. Utilizing degradation decomposition, DADF adaptively selects high-frequency components of RGB data to compensate for coarse depth. Furthermore, DADF achieves effective RGB-D fusion by incorporating conditional Mamba, where the state parameters are generated from degradation patches." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 72, + 553, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 553, + 96 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 553, + 96 + ], + "type": "text", + "content": "- We evaluate SigNet through comprehensive experiments, demonstrating its superiority over previous SOTAs." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 106, + 400, + 118 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 106, + 400, + 118 + ], + "spans": [ + { + "bbox": [ + 314, + 106, + 400, + 118 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 126, + 553, + 149 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 126, + 553, + 149 + ], + "spans": [ + { + "bbox": [ + 313, + 126, + 553, + 149 + ], + "type": "text", + "content": "Depth Completion. Depth completion can be broadly categorized into depth-only and image-guided categories." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 312, + 150, + 553, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 150, + 553, + 270 + ], + "spans": [ + { + "bbox": [ + 312, + 150, + 553, + 270 + ], + "type": "text", + "content": "Depth-only methods [9, 12, 14, 28, 40, 41, 44, 50] obtain dense depth directly from a single sparse depth input. For example, IP-Basic [14] designs reasonable combinations of basic image processing operations to densify sparse depth without using CNNs. Ma et al. [28] employ a UNet-like architecture to progressively densify the sparse depth input. FusionNet [41] designs a lightweight completion network that aggregates both global and local information with confidence adjustment. Moreover, Lu et al. [26] introduce a color image branch as auxiliary supervision during training." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 270, + 553, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 270, + 553, + 460 + ], + "spans": [ + { + "bbox": [ + 313, + 270, + 553, + 460 + ], + "type": "text", + "content": "Despite advancements in single-modal methods, their performance is often limited without supplementary reference information. To address this challenge, numerous image-guided techniques have been proposed [11, 16, 19, 22, 37, 38, 45, 50, 53, 61, 66]. For example, GuideNet [37] and RigNet [48] utilize color images to generate adaptive filtering kernels for depth feature extraction. GFormer [32] and CFormer [61] integrate convolution and transformer mechanisms to capture both local and global representations. LRRU [44] introduces a dynamic kernel scope that transitions from large to small to capture varying dependencies. FuseNet [3] and PointDC [56] incorporate LiDAR point cloud branches to model 3D geometries. BEV@DC [66] uses bird's-eye view to build a point-voxel architecture. Additionally, DepthPrompting [30] embeds priors of large depth models for better sensor-agnostic generalization." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 461, + 553, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 461, + 553, + 652 + ], + "spans": [ + { + "bbox": [ + 313, + 461, + 553, + 652 + ], + "type": "text", + "content": "To further advance completion models, spatial propagation networks (SPNs) have been introduced. The initial SPN [23] is designed to learn a pairwise similarity matrix. CSPN [5] applies this technique to depth completion by conducting recursive convolutions with fixed local neighborhood kernels, while " + }, + { + "bbox": [ + 313, + 461, + 553, + 652 + ], + "type": "inline_equation", + "content": "\\mathrm{CSPN} + +" + }, + { + "bbox": [ + 313, + 461, + 553, + 652 + ], + "type": "text", + "content": " [7] adapts by learning variable kernel sizes. PENet [11] expands the receptive fields using dilated convolutions. Furthermore, NLSPN [29] integrates non-local neighbors via deformable convolutions. DySPN [20] generates dynamic non-linear neighbors using an attention mechanism. Most recently, 3D SPNs [6] have been developed to encode 3D geometry. For example, GraphC-SPN [24] employs additional geometric constraints to regulate 3D propagation. TPVD [50] introduces a geometric SPN that constructs the affinity in 2D tri-perspective view spaces and their combined 3D projection space." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "text", + "content": "Degradation Learning. Recently, color image restoration has been significantly advanced by improvements in degradation representations [43, 55, 59]. For example, Zhang et al. [59] construct a sophisticated yet practical degradation model to simulate real-world image degradation for" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 55, + 673, + 295, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 673, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 673, + 295, + 713 + ], + "type": "text", + "content": "1Specifically depth super-resolution (DSR). Mainstream DSR methods restore target depth directly from upsampled depth. Consequently, the processes and the end goals of depth completion and DSR are identical, i.e., predicting clear depth from the same-resolution coarse depth." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26944" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 95, + 70, + 516, + 301 + ], + "blocks": [ + { + "bbox": [ + 95, + 70, + 516, + 301 + ], + "lines": [ + { + "bbox": [ + 95, + 70, + 516, + 301 + ], + "spans": [ + { + "bbox": [ + 95, + 70, + 516, + 301 + ], + "type": "image", + "image_path": "38425302d36afe1dcc84c4999e36e59711a1aa5fed916640e10ca19e005df0ea.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 306, + 555, + 340 + ], + "lines": [ + { + "bbox": [ + 54, + 306, + 555, + 340 + ], + "spans": [ + { + "bbox": [ + 54, + 306, + 555, + 340 + ], + "type": "text", + "content": "Figure 2. Pipeline of SigNet. The sparse depth data is initially filled to create a coarse depth map. We then utilize the degradation assumption in Eq. (1) to establish a connection between this coarse map and the target prediction. The color image features and degradation representations are employed to guide the multi-modal fusion through our proposed DADF module, as depicted in Fig. 3." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 357, + 297, + 573 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 357, + 297, + 573 + ], + "spans": [ + { + "bbox": [ + 54, + 357, + 297, + 573 + ], + "type": "text", + "content": "blind image super-resolution. Similarly, Zhou et al. [64] propose to estimate degradation for each local image region using a degradation-adaptive regression filter for blind super-resolution. Liang et al. [18] present a degradation-adaptive model that specifies the network parameters by estimating the degradation. Furthermore, several image restoration methods [17, 57, 58] have been proposed to simultaneously address multiple forms of degradation. For example, Zhang et al. [57] develop a multi-degradation image restoration network that incrementally learns degradation representations via clustering, without relying on prior knowledge of the degradation information. Zhang et al. [58] build an ingredient-oriented degradation reformulation network to model the relationship among different images. Additionally, Yao et al. [54] introduce a neural degradation representation learning method that captures latent degradation features, adaptively decomposing various types of degradation to enable comprehensive restoration." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 585, + 112, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 585, + 112, + 597 + ], + "spans": [ + { + "bbox": [ + 55, + 585, + 112, + 597 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "text", + "content": "Overview. Fig. 2 shows our SigNet pipeline that consists of the depth branch, degradation branch, and image branch. In the depth branch, the sparse depth " + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{S}" + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "text", + "content": " is first filled via nonCNN densification tools [14, 15], obtaining a coarse depth map " + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}" + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "text", + "content": ". According to Eq. (1), we employ denoising operators to weaken the negative impact of the noise " + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{n}" + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "text", + "content": ", yielding " + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{Z}}" + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "text", + "content": ". Then, we use residual groups [60] to extract features of the denoised depth " + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{Z}}" + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "text", + "content": " and the color image " + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "text", + "content": " in the image branch. In the degradation branch, the implicit degradation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 357, + 555, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 357, + 555, + 453 + ], + "spans": [ + { + "bbox": [ + 313, + 357, + 555, + 453 + ], + "type": "inline_equation", + "content": "\\mathbf{D}" + }, + { + "bbox": [ + 313, + 357, + 555, + 453 + ], + "type": "text", + "content": " is generated from a degradation network, where convolutions and residual groups are used. We thus utilize the image feature and degradation " + }, + { + "bbox": [ + 313, + 357, + 555, + 453 + ], + "type": "inline_equation", + "content": "\\mathbf{D}" + }, + { + "bbox": [ + 313, + 357, + 555, + 453 + ], + "type": "text", + "content": " to facilitate RGB-D fusion via our proposed Degradation-Aware Decomposition and Fusion (DADF). In summary, our SigNet redefines the depth completion as depth enhancement through the degradation assumption in Eq. (1), making the model free from the challenges of mismatch and ambiguity." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 458, + 409, + 470 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 458, + 409, + 470 + ], + "spans": [ + { + "bbox": [ + 313, + 458, + 409, + 470 + ], + "type": "text", + "content": "3.1. Task Switching" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 475, + 555, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 475, + 555, + 643 + ], + "spans": [ + { + "bbox": [ + 313, + 475, + 555, + 643 + ], + "type": "text", + "content": "Degradation learning is widely applied in low-level image restoration [43, 55, 59], yielding impressive results. These tasks exhibit two key characteristics: 1) both input and target data are typically dense, and 2) the input data generally have a lower resolution than the target data. Actually, the low-resolution input data are typically upsampled using interpolation methods, such as bicubic interpolation. Consequently, the degradation matrix usually consists of a convolution whose kernel is derived from degradation representations and a downsampling function. In contrast, the depth completion task takes sparse depth data as input, and the resolutions of the input and target data are the same. Thus, we reframe depth completion as depth enhancement by drawing an analogy to image restoration." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 643, + 555, + 679 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 643, + 555, + 679 + ], + "spans": [ + { + "bbox": [ + 313, + 643, + 555, + 679 + ], + "type": "text", + "content": "As shown in Fig. 2, we first fill the sparse depth " + }, + { + "bbox": [ + 313, + 643, + 555, + 679 + ], + "type": "inline_equation", + "content": "\\mathbf{S}" + }, + { + "bbox": [ + 313, + 643, + 555, + 679 + ], + "type": "text", + "content": " using non-CNN densification tools [14, 15], denoted as " + }, + { + "bbox": [ + 313, + 643, + 555, + 679 + ], + "type": "inline_equation", + "content": "f_{ds}(\\cdot)" + }, + { + "bbox": [ + 313, + 643, + 555, + 679 + ], + "type": "text", + "content": ", resulting in a coarse yet dense depth " + }, + { + "bbox": [ + 313, + 643, + 555, + 679 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}" + }, + { + "bbox": [ + 313, + 643, + 555, + 679 + ], + "type": "text", + "content": ", formulated as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 407, + 684, + 553, + 696 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 407, + 684, + 553, + 696 + ], + "spans": [ + { + "bbox": [ + 407, + 684, + 553, + 696 + ], + "type": "interline_equation", + "content": "\\mathbf {Z} = f _ {d s} (\\mathbf {S}). \\tag {2}", + "image_path": "f9562edb5d214752e9a0564ccd1485be9a3b0cbab5a0f9b89a63d31670e8041e.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 701, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 701, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 701, + 553, + 714 + ], + "type": "text", + "content": "As discussed in Sec. 1, this approach eliminates the mis" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26945" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 68, + 296, + 210 + ], + "blocks": [ + { + "bbox": [ + 56, + 68, + 296, + 210 + ], + "lines": [ + { + "bbox": [ + 56, + 68, + 296, + 210 + ], + "spans": [ + { + "bbox": [ + 56, + 68, + 296, + 210 + ], + "type": "image", + "image_path": "7c0d8234f69e1a252db2271431e620e0e785ffc4e9e5184fa9352b8cc68fa270.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 217, + 295, + 240 + ], + "lines": [ + { + "bbox": [ + 55, + 217, + 295, + 240 + ], + "spans": [ + { + "bbox": [ + 55, + 217, + 295, + 240 + ], + "type": "text", + "content": "Figure 3. Overview of our DADF. " + }, + { + "bbox": [ + 55, + 217, + 295, + 240 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 55, + 217, + 295, + 240 + ], + "type": "text", + "content": " and CM refer to elementwise multiplication and the conditional Mamba, respectively." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 259, + 296, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 259, + 296, + 344 + ], + "spans": [ + { + "bbox": [ + 55, + 259, + 296, + 344 + ], + "type": "text", + "content": "match and ambiguity caused by direct convolution over irregularly sampled sparse data. Subsequently, we employ a degradation network to estimate the implicit degradation " + }, + { + "bbox": [ + 55, + 259, + 296, + 344 + ], + "type": "inline_equation", + "content": "\\mathbf{D}" + }, + { + "bbox": [ + 55, + 259, + 296, + 344 + ], + "type": "text", + "content": ". The degradation kernel " + }, + { + "bbox": [ + 55, + 259, + 296, + 344 + ], + "type": "inline_equation", + "content": "\\mathbf{K}" + }, + { + "bbox": [ + 55, + 259, + 296, + 344 + ], + "type": "text", + "content": " is produced from " + }, + { + "bbox": [ + 55, + 259, + 296, + 344 + ], + "type": "inline_equation", + "content": "\\mathbf{D}" + }, + { + "bbox": [ + 55, + 259, + 296, + 344 + ], + "type": "text", + "content": " using convolutions. Consequently, given the coarse depth " + }, + { + "bbox": [ + 55, + 259, + 296, + 344 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}" + }, + { + "bbox": [ + 55, + 259, + 296, + 344 + ], + "type": "text", + "content": " and target depth " + }, + { + "bbox": [ + 55, + 259, + 296, + 344 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}" + }, + { + "bbox": [ + 55, + 259, + 296, + 344 + ], + "type": "text", + "content": ", Eq. (1) can be rewritten based on the task switching as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 124, + 351, + 295, + 365 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 124, + 351, + 295, + 365 + ], + "spans": [ + { + "bbox": [ + 124, + 351, + 295, + 365 + ], + "type": "interline_equation", + "content": "\\mathbf {Z} = \\mathbf {K} (\\mathbf {D}; \\vartheta) \\otimes \\mathbf {Y} + \\mathbf {n}, \\tag {3}", + "image_path": "8116721511732c4a1a4236dc7dca5cecdc79ebf9e84a41bf368b2cf3e8894822.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 373, + 296, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 373, + 296, + 409 + ], + "spans": [ + { + "bbox": [ + 55, + 373, + 296, + 409 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 373, + 296, + 409 + ], + "type": "inline_equation", + "content": "\\otimes" + }, + { + "bbox": [ + 55, + 373, + 296, + 409 + ], + "type": "text", + "content": " indicates convolution operation, " + }, + { + "bbox": [ + 55, + 373, + 296, + 409 + ], + "type": "inline_equation", + "content": "\\vartheta" + }, + { + "bbox": [ + 55, + 373, + 296, + 409 + ], + "type": "text", + "content": " is the network parameter. Since the resolutions are the same, our degradation equation does not need the downsampling function." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 415, + 296, + 428 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 415, + 296, + 428 + ], + "spans": [ + { + "bbox": [ + 55, + 415, + 296, + 428 + ], + "type": "text", + "content": "3.2. Degradation-Aware Decomposition and Fusion" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 433, + 296, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 433, + 296, + 517 + ], + "spans": [ + { + "bbox": [ + 55, + 433, + 296, + 517 + ], + "type": "text", + "content": "As we know, degradation typically occurs near edges, suggesting that this cue can be leveraged to enhance RGB-D fusion, especially considering that the coarse depth often lacks significant edge details, while the RGB image contains abundant high-frequency components. Thus, we propose DADF that is shown in Fig. 3. We denote the image and depth features as " + }, + { + "bbox": [ + 55, + 433, + 296, + 517 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{img}" + }, + { + "bbox": [ + 55, + 433, + 296, + 517 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 433, + 296, + 517 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{dep}" + }, + { + "bbox": [ + 55, + 433, + 296, + 517 + ], + "type": "text", + "content": ", obtaining:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 525, + 295, + 538 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 525, + 295, + 538 + ], + "spans": [ + { + "bbox": [ + 132, + 525, + 295, + 538 + ], + "type": "interline_equation", + "content": "\\mathbf {F} _ {i m g} = f _ {e} (\\mathbf {I}), \\tag {4}", + "image_path": "21b51a907700c21b051d1f4fbc850a69094e3f51a9c18b36e55d02c5f72e3eb3.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 133, + 540, + 295, + 554 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 540, + 295, + 554 + ], + "spans": [ + { + "bbox": [ + 133, + 540, + 295, + 554 + ], + "type": "interline_equation", + "content": "\\mathbf {F} _ {d e p} = f _ {e} \\left(f _ {d n} (\\mathbf {Z})\\right), \\tag {5}", + "image_path": "c9f07f67a206aabd5c05ac3a3a065434a969681356770cde917ae5fa4401d5f4.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 561, + 296, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 561, + 296, + 609 + ], + "spans": [ + { + "bbox": [ + 55, + 561, + 296, + 609 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 561, + 296, + 609 + ], + "type": "inline_equation", + "content": "f_{e}(\\cdot)" + }, + { + "bbox": [ + 55, + 561, + 296, + 609 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 561, + 296, + 609 + ], + "type": "inline_equation", + "content": "f_{dn}(\\cdot)" + }, + { + "bbox": [ + 55, + 561, + 296, + 609 + ], + "type": "text", + "content": " refer to the feature extraction network and denoising function, respectively. As presented in Fig. 3, DADF takes as input the image feature " + }, + { + "bbox": [ + 55, + 561, + 296, + 609 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{img}" + }, + { + "bbox": [ + 55, + 561, + 296, + 609 + ], + "type": "text", + "content": ", depth feature " + }, + { + "bbox": [ + 55, + 561, + 296, + 609 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{dep}" + }, + { + "bbox": [ + 55, + 561, + 296, + 609 + ], + "type": "text", + "content": ", and degradation " + }, + { + "bbox": [ + 55, + 561, + 296, + 609 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_D" + }, + { + "bbox": [ + 55, + 561, + 296, + 609 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 609, + 296, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 609, + 296, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 609, + 296, + 693 + ], + "type": "text", + "content": "Decomposition. To extract the most representative components of " + }, + { + "bbox": [ + 55, + 609, + 296, + 693 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_D" + }, + { + "bbox": [ + 55, + 609, + 296, + 693 + ], + "type": "text", + "content": ", we first apply the discrete cosine transform (DCT) to convert it into the frequency domain. Next, it adaptively learns a spectrum mask to enhance informative frequency bands while suppressing irrelevant ones. Finally, we reconstruct the spatial representation using the inverse DCT. This process is formulated as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 117, + 700, + 295, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 117, + 700, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 117, + 700, + 295, + 715 + ], + "type": "interline_equation", + "content": "\\left\\{\\mathbf {D} _ {i} \\mid i = 1, 2, 3 \\right\\} = f _ {d c t} (\\mathbf {D}). \\tag {6}", + "image_path": "ff3a2c3e6c60d7830e0c4784b233b6a32b196147b3677af00e363be6c95ae581.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 320, + 75, + 477, + 87 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 75, + 477, + 87 + ], + "spans": [ + { + "bbox": [ + 320, + 75, + 477, + 87 + ], + "type": "text", + "content": "Algorithm 1: Conditional SSM Block" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 327, + 89, + 485, + 100 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 89, + 485, + 100 + ], + "spans": [ + { + "bbox": [ + 327, + 89, + 485, + 100 + ], + "type": "text", + "content": "Input: " + }, + { + "bbox": [ + 327, + 89, + 485, + 100 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{rgbd}" + }, + { + "bbox": [ + 327, + 89, + 485, + 100 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 327, + 89, + 485, + 100 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{D}_i|i = 1,2,3\\} :(\\mathrm{HW},\\mathrm{C})" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 327, + 100, + 426, + 111 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 100, + 426, + 111 + ], + "spans": [ + { + "bbox": [ + 327, + 100, + 426, + 111 + ], + "type": "text", + "content": "Output: " + }, + { + "bbox": [ + 327, + 100, + 426, + 111 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{F}}_{rgbd}:(\\mathrm{HW},\\mathrm{C})" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 319, + 112, + 511, + 250 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 319, + 112, + 438, + 121 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 112, + 438, + 121 + ], + "spans": [ + { + "bbox": [ + 319, + 112, + 438, + 121 + ], + "type": "text", + "content": "1 A : (C, N) " + }, + { + "bbox": [ + 319, + 112, + 438, + 121 + ], + "type": "inline_equation", + "content": "\\leftarrow" + }, + { + "bbox": [ + 319, + 112, + 438, + 121 + ], + "type": "text", + "content": " ParameterA" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 319, + 122, + 511, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 122, + 511, + 144 + ], + "spans": [ + { + "bbox": [ + 319, + 122, + 511, + 144 + ], + "type": "text", + "content": "2 " + }, + { + "bbox": [ + 319, + 122, + 511, + 144 + ], + "type": "inline_equation", + "content": "\\mathbf{B}:(\\mathrm{HW},\\mathrm{N})\\gets \\mathbf{Linear}_B(f_{gru}(\\mathbf{F}_{rgbd},\\{\\mathbf{D}_i|i = 1,2,3\\}))" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 319, + 144, + 511, + 166 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 144, + 511, + 166 + ], + "spans": [ + { + "bbox": [ + 319, + 144, + 511, + 166 + ], + "type": "text", + "content": "3 " + }, + { + "bbox": [ + 319, + 144, + 511, + 166 + ], + "type": "inline_equation", + "content": "\\mathbf{C}:(\\mathrm{HW},\\mathrm{N})\\gets \\mathbf{Linear}_C(f_{gru}(\\mathbf{F}_{rgbd},\\{\\mathbf{D}_i|i = 1,2,3\\}))" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 319, + 166, + 501, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 166, + 501, + 199 + ], + "spans": [ + { + "bbox": [ + 319, + 166, + 501, + 199 + ], + "type": "text", + "content": "4 " + }, + { + "bbox": [ + 319, + 166, + 501, + 199 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 319, + 166, + 501, + 199 + ], + "type": "text", + "content": " : (HW, C) " + }, + { + "bbox": [ + 319, + 166, + 501, + 199 + ], + "type": "inline_equation", + "content": "\\leftarrow" + }, + { + "bbox": [ + 319, + 166, + 501, + 199 + ], + "type": "inline_equation", + "content": "\\log(1 + \\exp(\\text{Linear}_\\Delta(f_{gru}(\\mathbf{F}_{rgbd}, \\{\\mathbf{D}_i | i = 1, 2, 3\\})) + \\text{Parameter}_\\Delta))" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 319, + 200, + 454, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 200, + 454, + 212 + ], + "spans": [ + { + "bbox": [ + 319, + 200, + 454, + 212 + ], + "type": "text", + "content": "5 " + }, + { + "bbox": [ + 319, + 200, + 454, + 212 + ], + "type": "inline_equation", + "content": "\\overline{\\mathbf{A}}: (\\mathrm{HW}, \\mathrm{C}, \\mathrm{N}) \\leftarrow \\exp(\\Delta \\otimes \\mathbf{A})" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 319, + 212, + 432, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 212, + 432, + 223 + ], + "spans": [ + { + "bbox": [ + 319, + 212, + 432, + 223 + ], + "type": "text", + "content": "6 " + }, + { + "bbox": [ + 319, + 212, + 432, + 223 + ], + "type": "inline_equation", + "content": "\\overline{\\mathbf{B}}:\\mathrm{(HW,C,N)}\\gets \\pmb {\\Delta}\\otimes \\mathbf{B}" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 319, + 224, + 451, + 237 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 224, + 451, + 237 + ], + "spans": [ + { + "bbox": [ + 319, + 224, + 451, + 237 + ], + "type": "text", + "content": "7 " + }, + { + "bbox": [ + 319, + 224, + 451, + 237 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{F}}_{rgbd} \\gets \\mathrm{SSM}(\\overline{\\mathbf{A}},\\overline{\\mathbf{B}},\\mathbf{C})(\\mathbf{F}_{rgbd})" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 319, + 237, + 377, + 250 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 237, + 377, + 250 + ], + "spans": [ + { + "bbox": [ + 319, + 237, + 377, + 250 + ], + "type": "text", + "content": "8 return " + }, + { + "bbox": [ + 319, + 237, + 377, + 250 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{F}}_{rgbd}" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 277, + 555, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 277, + 555, + 350 + ], + "spans": [ + { + "bbox": [ + 313, + 277, + 555, + 350 + ], + "type": "text", + "content": "We apply concatenation, convolution, pooling, and Sigmoid operations, collectively denoted as " + }, + { + "bbox": [ + 313, + 277, + 555, + 350 + ], + "type": "inline_equation", + "content": "f_{m}(\\cdot)" + }, + { + "bbox": [ + 313, + 277, + 555, + 350 + ], + "type": "text", + "content": ", to generate a mask " + }, + { + "bbox": [ + 313, + 277, + 555, + 350 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 277, + 555, + 350 + ], + "type": "text", + "content": ". This mask is then multiplied with the combined frequency components to produce an updated mask " + }, + { + "bbox": [ + 313, + 277, + 555, + 350 + ], + "type": "inline_equation", + "content": "\\hat{m}" + }, + { + "bbox": [ + 313, + 277, + 555, + 350 + ], + "type": "text", + "content": ", enabling the adaptive selection of the representative image feature " + }, + { + "bbox": [ + 313, + 277, + 555, + 350 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{F}}_{img}" + }, + { + "bbox": [ + 313, + 277, + 555, + 350 + ], + "type": "text", + "content": ". The process is defined as:" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 363, + 359, + 553, + 372 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 363, + 359, + 553, + 372 + ], + "spans": [ + { + "bbox": [ + 363, + 359, + 553, + 372 + ], + "type": "interline_equation", + "content": "m = f _ {m} \\left(\\mathbf {F} _ {\\text {i m g}}, \\mathbf {D}\\right), \\tag {7}", + "image_path": "fd8a804e2b9c0ef532bac7e1ef5aed3fa7b6c7714001f9fcdca01c1170ad82ea.jpg" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 364, + 375, + 553, + 396 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 364, + 375, + 553, + 396 + ], + "spans": [ + { + "bbox": [ + 364, + 375, + 553, + 396 + ], + "type": "interline_equation", + "content": "\\hat {m} = f _ {i d c t} \\left(m \\cdot \\left(f _ {a c} \\left(\\sum_ {i = 1} ^ {3} \\mathbf {D} _ {i}\\right)\\right)\\right), \\tag {8}", + "image_path": "21e4bab81a5d6a3149d4db5cc0849fef09fb40e647623226be7539ac8512773e.jpg" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 364, + 397, + 553, + 412 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 364, + 397, + 553, + 412 + ], + "spans": [ + { + "bbox": [ + 364, + 397, + 553, + 412 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {F}} _ {i m g} = \\hat {m} \\cdot \\mathbf {F} _ {i m g}, \\tag {9}", + "image_path": "2793a1a88b142c6441396b030ece54a2a6c71aece27d8a0983cb9e161cc69b23.jpg" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 313, + 422, + 553, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 422, + 553, + 445 + ], + "spans": [ + { + "bbox": [ + 313, + 422, + 553, + 445 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 422, + 553, + 445 + ], + "type": "inline_equation", + "content": "f_{idct}" + }, + { + "bbox": [ + 313, + 422, + 553, + 445 + ], + "type": "text", + "content": " refers to inverse DCT and " + }, + { + "bbox": [ + 313, + 422, + 553, + 445 + ], + "type": "inline_equation", + "content": "f_{ac}(\\cdot)" + }, + { + "bbox": [ + 313, + 422, + 553, + 445 + ], + "type": "text", + "content": " represents the combination function of addition and convolutions." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 313, + 445, + 555, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 445, + 555, + 542 + ], + "spans": [ + { + "bbox": [ + 313, + 445, + 555, + 542 + ], + "type": "text", + "content": "Fusion. As a result, we further leverage the decomposed degradation " + }, + { + "bbox": [ + 313, + 445, + 555, + 542 + ], + "type": "inline_equation", + "content": "\\mathbf{D}_i" + }, + { + "bbox": [ + 313, + 445, + 555, + 542 + ], + "type": "text", + "content": " to enable effective RGB-D fusion. It is well-known that Mamba [10] is good at handling long-range correlations. Inspired by recent Vision Mamba [25], we present a variant, i.e., Conditional Mamba (CM). As shown in Fig. 3, the selected image feature " + }, + { + "bbox": [ + 313, + 445, + 555, + 542 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{F}}_{img}" + }, + { + "bbox": [ + 313, + 445, + 555, + 542 + ], + "type": "text", + "content": " and raw depth feature " + }, + { + "bbox": [ + 313, + 445, + 555, + 542 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{dep}" + }, + { + "bbox": [ + 313, + 445, + 555, + 542 + ], + "type": "text", + "content": " are concatenated into CM, where the decomposed degradation " + }, + { + "bbox": [ + 313, + 445, + 555, + 542 + ], + "type": "inline_equation", + "content": "\\mathbf{D}_i" + }, + { + "bbox": [ + 313, + 445, + 555, + 542 + ], + "type": "text", + "content": " serves as the condition:" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 353, + 551, + 553, + 565 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 551, + 553, + 565 + ], + "spans": [ + { + "bbox": [ + 353, + 551, + 553, + 565 + ], + "type": "interline_equation", + "content": "\\mathbf {F} _ {r g b d} = f _ {c c} \\left(\\tilde {\\mathbf {F}} _ {i m g}, \\mathbf {F} _ {d e p}\\right), \\tag {10}", + "image_path": "7eb95f0ed4038d7acbc9cb6dd094a618dcb6ca375f2d62b78751f358783dbbcf.jpg" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 354, + 568, + 553, + 582 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 568, + 553, + 582 + ], + "spans": [ + { + "bbox": [ + 354, + 568, + 553, + 582 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {F}} _ {r g b d} = f _ {c m} \\left(\\mathbf {F} _ {r g b d}, \\left\\{\\mathbf {D} _ {i} \\mid i = 1, 2, 3 \\right\\}\\right), \\tag {11}", + "image_path": "3f6440818de0005c46a52c4fc100a23709e49a583be54236d0442af99b9ce63b.jpg" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 313, + 592, + 555, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 592, + 555, + 651 + ], + "spans": [ + { + "bbox": [ + 313, + 592, + 555, + 651 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 592, + 555, + 651 + ], + "type": "inline_equation", + "content": "f_{cc}" + }, + { + "bbox": [ + 313, + 592, + 555, + 651 + ], + "type": "text", + "content": " denotes the combination function of concatenation and convolutions, " + }, + { + "bbox": [ + 313, + 592, + 555, + 651 + ], + "type": "inline_equation", + "content": "f_{cm}" + }, + { + "bbox": [ + 313, + 592, + 555, + 651 + ], + "type": "text", + "content": " is the Condition Mamba function. As reported in Alg. 1, the state parameters " + }, + { + "bbox": [ + 313, + 592, + 555, + 651 + ], + "type": "inline_equation", + "content": "\\mathbf{B}" + }, + { + "bbox": [ + 313, + 592, + 555, + 651 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 592, + 555, + 651 + ], + "type": "inline_equation", + "content": "\\mathbf{C}" + }, + { + "bbox": [ + 313, + 592, + 555, + 651 + ], + "type": "text", + "content": " are calculated based on both the RGB-D input and degradation using gated recurrent convolution units." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 313, + 659, + 403, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 659, + 403, + 671 + ], + "spans": [ + { + "bbox": [ + 313, + 659, + 403, + 671 + ], + "type": "text", + "content": "3.3. Loss Function" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "type": "text", + "content": "As shown in our pipeline, the total loss function " + }, + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_t" + }, + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "type": "text", + "content": " consists of a reconstruction loss " + }, + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_r" + }, + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "type": "text", + "content": " and a self-supervised degradation learning loss " + }, + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_d" + }, + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "type": "text", + "content": ". Let " + }, + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{Y}}" + }, + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}" + }, + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "type": "text", + "content": " denote the ground truth" + } + ] + } + ], + "index": 34 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "26946" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 77, + 69, + 533, + 466 + ], + "blocks": [ + { + "bbox": [ + 77, + 69, + 533, + 466 + ], + "lines": [ + { + "bbox": [ + 77, + 69, + 533, + 466 + ], + "spans": [ + { + "bbox": [ + 77, + 69, + 533, + 466 + ], + "type": "table", + "html": "
DatasetMethodParams. (M) ↓RMSE (mm) ↓REL ↓δ1(%) ↑δ2(%) ↑δ3(%) ↑Venue
NYUv2Bilateral [1]-5320.13285.193.595.9CVPR 2005
S2D [28]26.12300.05494.597.398.9ICRA 2019
CSPN [4]17.41730.02096.398.699.5ECCV 2018
DfuseNet [34]-1560.01698.899.799.9ITSC 2019
DM-LRN [33]-2050.01498.899.699.9IROS 2021
RDF-GAN [42]-1390.01398.799.699.9CVPR 2022
AGG-Net [2]129.1920.01499.499.9100.0ICCV 2023
BPNet [38]89.9880.01199.699.9100.0CVPR 2024
SigNet (ours)3.3830.01299.799.9100.0CVPR 2025
DIMLBilateral [1]-6360.18983.088.892.4CVPR 2005
CSPN [4]17.41620.03396.198.799.6ECCV 2018
DfuseNet [34]-1430.02398.499.499.9ITSC 2019
DM-LRN [33]-1490.01599.099.699.9IROS 2021
AGG-Net [2]129.1780.01199.699.9100.0ICCV 2023
BPNet [38]89.9680.01199.499.8100.0CVPR 2024
SigNet (ours)3.3550.00699.799.9100.0CVPR 2025
SUN RGBDS2D [28]26.13290.07493.997.098.1ICRA 2019
CSPN [4]17.42950.13795.697.598.4ECCV 2018
DLiDAR [31]53.42790.06196.998.098.4CVPR 2019
DM-LRN [33]-2670.06397.698.298.7IROS 2021
RDF-GAN [42]-2550.05996.998.499.0CVPR 2022
AGG-Net [2]129.11520.03898.599.099.4ICCV 2023
BPNet [38]89.91240.02598.999.499.8CVPR 2024
SigNet (ours)3.31160.02299.099.499.8CVPR 2025
TOFDCCSPN [4]17.42240.04294.595.396.5ECCV 2018
GuideNet [37]73.51460.03097.698.999.5TIP 2020
PENet [11]131.52410.04394.695.395.5ICRA 2021
NLSPN [29]25.81740.02996.497.998.9ECCV 2020
CFounder [61]83.51130.02999.199.699.9CVPR 2023
RigNet [48]65.21330.02597.699.199.7ECCV 2022
GraphCSPN [24]26.42530.05292.096.998.7ECCV 2022
PointDC [56]25.11090.02198.599.299.6ICCV 2023
TPVD [50]31.2920.01499.199.699.9CVPR 2024
SigNet (ours)3.3870.01299.299.799.9CVPR 2025
", + "image_path": "e5f92265e359e440d8787dfe0df627f93e9a03e5bea2a8467bde02142b45aaba.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 470, + 555, + 492 + ], + "lines": [ + { + "bbox": [ + 55, + 470, + 555, + 492 + ], + "spans": [ + { + "bbox": [ + 55, + 470, + 555, + 492 + ], + "type": "text", + "content": "Table 1. Quantitative comparison on four depth completion datasets. The best and suboptimal results are highlighted. Note that most of the results on NYUv2, DIML, and SUN RGBD are borrowed directly from AGG-Net [2], while those of TOFDC are from TPVD [50]." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 513, + 295, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 513, + 295, + 537 + ], + "spans": [ + { + "bbox": [ + 55, + 513, + 295, + 537 + ], + "type": "text", + "content": "depth and the predicted depth, respectively. The reconstruction loss " + }, + { + "bbox": [ + 55, + 513, + 295, + 537 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_r" + }, + { + "bbox": [ + 55, + 513, + 295, + 537 + ], + "type": "text", + "content": " is defined as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 119, + 544, + 295, + 577 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 119, + 544, + 295, + 577 + ], + "spans": [ + { + "bbox": [ + 119, + 544, + 295, + 577 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {r} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} | | \\hat {\\mathbf {Y}} _ {i} - \\mathbf {Y} _ {i} | | _ {1}, \\tag {12}", + "image_path": "8c8d72b95c58b92a0b0f9723cdf1f94669cbfd2e4921aad622541d524045a3cd.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 582, + 296, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 296, + 632 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 296, + 632 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 582, + 296, + 632 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 582, + 296, + 632 + ], + "type": "text", + "content": " represents the number of training samples, and " + }, + { + "bbox": [ + 55, + 582, + 296, + 632 + ], + "type": "inline_equation", + "content": "\\| \\cdot \\| _1" + }, + { + "bbox": [ + 55, + 582, + 296, + 632 + ], + "type": "text", + "content": " denotes the " + }, + { + "bbox": [ + 55, + 582, + 296, + 632 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 55, + 582, + 296, + 632 + ], + "type": "text", + "content": " norm. Subsequently, we introduce the degradation loss " + }, + { + "bbox": [ + 55, + 582, + 296, + 632 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_d" + }, + { + "bbox": [ + 55, + 582, + 296, + 632 + ], + "type": "text", + "content": " to facilitate the degradation learning process through:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 113, + 638, + 295, + 670 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 638, + 295, + 670 + ], + "spans": [ + { + "bbox": [ + 113, + 638, + 295, + 670 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {d} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\left| \\left| \\mathbf {K} \\otimes \\mathbf {Y} - \\mathbf {Z} \\right| \\right| _ {1}, \\tag {13}", + "image_path": "57bf056f6b8183c1d7c78f698a2004cf36cfb8178925f6d09a530be597eb5016.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{K}" + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": " refers to the degradation kernel learned from the implicit degradation representation " + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{D}" + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}" + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": " corresponds to the coarse depth. " + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\otimes" + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": " denotes the convolution operation, where" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 513, + 555, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 513, + 555, + 536 + ], + "spans": [ + { + "bbox": [ + 313, + 513, + 555, + 536 + ], + "type": "inline_equation", + "content": "\\mathbf{K}" + }, + { + "bbox": [ + 313, + 513, + 555, + 536 + ], + "type": "text", + "content": " serves as the convolution kernel. As a result, the total training loss is formulated as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 399, + 544, + 555, + 558 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 399, + 544, + 555, + 558 + ], + "spans": [ + { + "bbox": [ + 399, + 544, + 555, + 558 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {t} = \\mathcal {L} _ {r} + \\lambda \\mathcal {L} _ {d}, \\tag {14}", + "image_path": "5a67ef7185f30a7ce884dd1e8ae164283ff2857739edf99e6756458be871ae65.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 563, + 555, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 563, + 555, + 612 + ], + "spans": [ + { + "bbox": [ + 313, + 563, + 555, + 612 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 563, + 555, + 612 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 563, + 555, + 612 + ], + "type": "text", + "content": " is a hyper-parameter, empirically set to 0.1. By the constraint of the auxiliary loss " + }, + { + "bbox": [ + 313, + 563, + 555, + 612 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_d" + }, + { + "bbox": [ + 313, + 563, + 555, + 612 + ], + "type": "text", + "content": ", our model can build additional connections between the input and output, thus predicting more accurate depth results." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 621, + 391, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 621, + 391, + 635 + ], + "spans": [ + { + "bbox": [ + 313, + 621, + 391, + 635 + ], + "type": "text", + "content": "4. Experiment" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 641, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 555, + 715 + ], + "type": "text", + "content": "We evaluate our SigNet method on four popular datasets, including NYUv2, DIML, SUN RGBD, and TOFDC benchmarks. For NYUv2, DIML, and SUN RGBD, we maintain the settings used in AGG-Net [2], while for TOFDC, we follow the settings of TPVD [50]. Refer to our supplementary material for additional details on the experimental setups." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "26947" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 85, + 71, + 173, + 205 + ], + "blocks": [ + { + "bbox": [ + 85, + 71, + 173, + 205 + ], + "lines": [ + { + "bbox": [ + 85, + 71, + 173, + 205 + ], + "spans": [ + { + "bbox": [ + 85, + 71, + 173, + 205 + ], + "type": "image", + "image_path": "09e2b28f93e8421502ee794244836a54580f10db81a9fb4a6e4e5ab74e6bf353.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 113, + 207, + 143, + 217 + ], + "lines": [ + { + "bbox": [ + 113, + 207, + 143, + 217 + ], + "spans": [ + { + "bbox": [ + 113, + 207, + 143, + 217 + ], + "type": "text", + "content": "(a) RGB" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 173, + 71, + 261, + 205 + ], + "blocks": [ + { + "bbox": [ + 173, + 71, + 261, + 205 + ], + "lines": [ + { + "bbox": [ + 173, + 71, + 261, + 205 + ], + "spans": [ + { + "bbox": [ + 173, + 71, + 261, + 205 + ], + "type": "image", + "image_path": "6f96e3f1244f41c55d4c841468fdd4b7e96c91d675d80575c6bda5e03ad1c750.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 198, + 207, + 236, + 217 + ], + "lines": [ + { + "bbox": [ + 198, + 207, + 236, + 217 + ], + "spans": [ + { + "bbox": [ + 198, + 207, + 236, + 217 + ], + "type": "text", + "content": "(b) NLSPN" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 262, + 72, + 350, + 205 + ], + "blocks": [ + { + "bbox": [ + 262, + 72, + 350, + 205 + ], + "lines": [ + { + "bbox": [ + 262, + 72, + 350, + 205 + ], + "spans": [ + { + "bbox": [ + 262, + 72, + 350, + 205 + ], + "type": "image", + "image_path": "f464526efacec3d8e3a64f967d67d4939157e57e04901401248b286c18d11759.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 287, + 207, + 323, + 217 + ], + "lines": [ + { + "bbox": [ + 287, + 207, + 323, + 217 + ], + "spans": [ + { + "bbox": [ + 287, + 207, + 323, + 217 + ], + "type": "text", + "content": "(c) RigNet" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 350, + 72, + 437, + 205 + ], + "blocks": [ + { + "bbox": [ + 350, + 72, + 437, + 205 + ], + "lines": [ + { + "bbox": [ + 350, + 72, + 437, + 205 + ], + "spans": [ + { + "bbox": [ + 350, + 72, + 437, + 205 + ], + "type": "image", + "image_path": "a1f283232dc99d1999c7033b1638338683da706908622868c1d3ff483f8b3e76.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 376, + 207, + 411, + 217 + ], + "lines": [ + { + "bbox": [ + 376, + 207, + 411, + 217 + ], + "spans": [ + { + "bbox": [ + 376, + 207, + 411, + 217 + ], + "type": "text", + "content": "(d) SigNet" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 438, + 72, + 525, + 205 + ], + "blocks": [ + { + "bbox": [ + 438, + 72, + 525, + 205 + ], + "lines": [ + { + "bbox": [ + 438, + 72, + 525, + 205 + ], + "spans": [ + { + "bbox": [ + 438, + 72, + 525, + 205 + ], + "type": "image", + "image_path": "2655180d87c800e552ac88a50546c20277dba44f1befcc3b4f702de33425ca42.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 470, + 207, + 493, + 217 + ], + "lines": [ + { + "bbox": [ + 470, + 207, + 493, + 217 + ], + "spans": [ + { + "bbox": [ + 470, + 207, + 493, + 217 + ], + "type": "text", + "content": "(e) GT" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 85, + 242, + 173, + 344 + ], + "blocks": [ + { + "bbox": [ + 82, + 222, + 526, + 234 + ], + "lines": [ + { + "bbox": [ + 82, + 222, + 526, + 234 + ], + "spans": [ + { + "bbox": [ + 82, + 222, + 526, + 234 + ], + "type": "text", + "content": "Figure 4. Visual comparison with SOTA methods on TOFDC dataset, including NLSPN [29], RigNet [48], and our SigNet." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 85, + 242, + 173, + 344 + ], + "lines": [ + { + "bbox": [ + 85, + 242, + 173, + 344 + ], + "spans": [ + { + "bbox": [ + 85, + 242, + 173, + 344 + ], + "type": "image", + "image_path": "8bfff01cb3e7aac297369b169baabfb607cf9a47496a4543b3c4386f4b3ca229.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 114, + 347, + 143, + 357 + ], + "lines": [ + { + "bbox": [ + 114, + 347, + 143, + 357 + ], + "spans": [ + { + "bbox": [ + 114, + 347, + 143, + 357 + ], + "type": "text", + "content": "(a) RGB" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 173, + 243, + 261, + 344 + ], + "blocks": [ + { + "bbox": [ + 173, + 243, + 261, + 344 + ], + "lines": [ + { + "bbox": [ + 173, + 243, + 261, + 344 + ], + "spans": [ + { + "bbox": [ + 173, + 243, + 261, + 344 + ], + "type": "image", + "image_path": "fcf0d73704f99531662243d392dd0e0ec29df5431bdf265d2e110fd6a1ff2e7e.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 200, + 347, + 234, + 357 + ], + "lines": [ + { + "bbox": [ + 200, + 347, + 234, + 357 + ], + "spans": [ + { + "bbox": [ + 200, + 347, + 234, + 357 + ], + "type": "text", + "content": "(b)CSPN" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 261, + 243, + 350, + 344 + ], + "blocks": [ + { + "bbox": [ + 261, + 243, + 350, + 344 + ], + "lines": [ + { + "bbox": [ + 261, + 243, + 350, + 344 + ], + "spans": [ + { + "bbox": [ + 261, + 243, + 350, + 344 + ], + "type": "image", + "image_path": "c1719cb69f606f6d26cba0fb05765d7ca0ba26ee7b10e2ecfcd3c95dd06288b3.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 283, + 347, + 327, + 357 + ], + "lines": [ + { + "bbox": [ + 283, + 347, + 327, + 357 + ], + "spans": [ + { + "bbox": [ + 283, + 347, + 327, + 357 + ], + "type": "text", + "content": "(c) AGG-Net" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 350, + 243, + 437, + 344 + ], + "blocks": [ + { + "bbox": [ + 350, + 243, + 437, + 344 + ], + "lines": [ + { + "bbox": [ + 350, + 243, + 437, + 344 + ], + "spans": [ + { + "bbox": [ + 350, + 243, + 437, + 344 + ], + "type": "image", + "image_path": "7ec5dd05e3aadc8c97a32bf2624a6465f77147c52b528cbbf3af93088595db63.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 375, + 347, + 411, + 357 + ], + "lines": [ + { + "bbox": [ + 375, + 347, + 411, + 357 + ], + "spans": [ + { + "bbox": [ + 375, + 347, + 411, + 357 + ], + "type": "text", + "content": "(d) SigNet" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 438, + 243, + 525, + 344 + ], + "blocks": [ + { + "bbox": [ + 438, + 243, + 525, + 344 + ], + "lines": [ + { + "bbox": [ + 438, + 243, + 525, + 344 + ], + "spans": [ + { + "bbox": [ + 438, + 243, + 525, + 344 + ], + "type": "image", + "image_path": "0cf62c4d48a96b8ec265c3ad71ad81d45f673267f400c22e3e6c685196c80f35.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 470, + 347, + 492, + 357 + ], + "lines": [ + { + "bbox": [ + 470, + 347, + 492, + 357 + ], + "spans": [ + { + "bbox": [ + 470, + 347, + 492, + 357 + ], + "type": "text", + "content": "(e) GT" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 85, + 381, + 157, + 470 + ], + "blocks": [ + { + "bbox": [ + 83, + 362, + 525, + 374 + ], + "lines": [ + { + "bbox": [ + 83, + 362, + 525, + 374 + ], + "spans": [ + { + "bbox": [ + 83, + 362, + 525, + 374 + ], + "type": "text", + "content": "Figure 5. Visual comparison with SOTA approaches on DIML dataset, including CSPN [4], AGG-Net [2], and our SigNet." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 85, + 381, + 157, + 470 + ], + "lines": [ + { + "bbox": [ + 85, + 381, + 157, + 470 + ], + "spans": [ + { + "bbox": [ + 85, + 381, + 157, + 470 + ], + "type": "image", + "image_path": "35ad799d658fde72b175e426b03d3a612be886f8910aa84baeb3b930251f8405.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 111, + 473, + 130, + 482 + ], + "lines": [ + { + "bbox": [ + 111, + 473, + 130, + 482 + ], + "spans": [ + { + "bbox": [ + 111, + 473, + 130, + 482 + ], + "type": "text", + "content": "RGB" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 108, + 490, + 500, + 501 + ], + "lines": [ + { + "bbox": [ + 108, + 490, + 500, + 501 + ], + "spans": [ + { + "bbox": [ + 108, + 490, + 500, + 501 + ], + "type": "text", + "content": "Figure 6. Visual examples of filled coarse depth maps and degradation representations on the DIML dataset." + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 158, + 381, + 230, + 470 + ], + "blocks": [ + { + "bbox": [ + 158, + 381, + 230, + 470 + ], + "lines": [ + { + "bbox": [ + 158, + 381, + 230, + 470 + ], + "spans": [ + { + "bbox": [ + 158, + 381, + 230, + 470 + ], + "type": "image", + "image_path": "015771f92579e5009873de3f94b2cd7c44c8fbdbddade9c11e149e6c742f09e0.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 171, + 473, + 217, + 483 + ], + "lines": [ + { + "bbox": [ + 171, + 473, + 217, + 483 + ], + "spans": [ + { + "bbox": [ + 171, + 473, + 217, + 483 + ], + "type": "text", + "content": "Coarse depth" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 231, + 381, + 304, + 470 + ], + "blocks": [ + { + "bbox": [ + 231, + 381, + 304, + 470 + ], + "lines": [ + { + "bbox": [ + 231, + 381, + 304, + 470 + ], + "spans": [ + { + "bbox": [ + 231, + 381, + 304, + 470 + ], + "type": "image", + "image_path": "7ab8df37afd80c775888b1080b3f672b48440efb273378472f51c967366030a3.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 246, + 473, + 288, + 483 + ], + "lines": [ + { + "bbox": [ + 246, + 473, + 288, + 483 + ], + "spans": [ + { + "bbox": [ + 246, + 473, + 288, + 483 + ], + "type": "text", + "content": "Degradation" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 306, + 381, + 378, + 470 + ], + "blocks": [ + { + "bbox": [ + 306, + 381, + 378, + 470 + ], + "lines": [ + { + "bbox": [ + 306, + 381, + 378, + 470 + ], + "spans": [ + { + "bbox": [ + 306, + 381, + 378, + 470 + ], + "type": "image", + "image_path": "ddfaa1a958b53be0a288ed96f85c9c1b3242c29b066e904f890b87245fcae813.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 333, + 473, + 351, + 482 + ], + "lines": [ + { + "bbox": [ + 333, + 473, + 351, + 482 + ], + "spans": [ + { + "bbox": [ + 333, + 473, + 351, + 482 + ], + "type": "text", + "content": "RGB" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 379, + 381, + 451, + 470 + ], + "blocks": [ + { + "bbox": [ + 379, + 381, + 451, + 470 + ], + "lines": [ + { + "bbox": [ + 379, + 381, + 451, + 470 + ], + "spans": [ + { + "bbox": [ + 379, + 381, + 451, + 470 + ], + "type": "image", + "image_path": "afd2c3b49df17bf5ed6643dd14e73146d16f8e572bc8d9491ea2f73202135ca0.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 392, + 473, + 437, + 483 + ], + "lines": [ + { + "bbox": [ + 392, + 473, + 437, + 483 + ], + "spans": [ + { + "bbox": [ + 392, + 473, + 437, + 483 + ], + "type": "text", + "content": "Coarse depth" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 452, + 381, + 525, + 470 + ], + "blocks": [ + { + "bbox": [ + 452, + 381, + 525, + 470 + ], + "lines": [ + { + "bbox": [ + 452, + 381, + 525, + 470 + ], + "spans": [ + { + "bbox": [ + 452, + 381, + 525, + 470 + ], + "type": "image", + "image_path": "6b65f37cd96e75aebc7d20ba2a1e2f01ba8e7a367b5d65d3dcf7e9fc3d47ebc3.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 466, + 473, + 509, + 483 + ], + "lines": [ + { + "bbox": [ + 466, + 473, + 509, + 483 + ], + "spans": [ + { + "bbox": [ + 466, + 473, + 509, + 483 + ], + "type": "text", + "content": "Degradation" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + } + ], + "index": 32 + }, + { + "bbox": [ + 55, + 518, + 190, + 532 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 518, + 190, + 532 + ], + "spans": [ + { + "bbox": [ + 55, + 518, + 190, + 532 + ], + "type": "text", + "content": "4.1. Comparison with SOTA" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 55, + 536, + 296, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 536, + 296, + 609 + ], + "spans": [ + { + "bbox": [ + 55, + 536, + 296, + 609 + ], + "type": "text", + "content": "To assess the effectiveness of our SigNet, we benchmark it against the traditional bilateral filtering method [1] and some recent SOTA deep learning approaches such as [2, 28, 29, 34, 38]. The error and accuracy results for the competing models are sourced from their original publications or obtained by retraining them following the default settings." + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 55, + 614, + 178, + 626 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 614, + 178, + 626 + ], + "spans": [ + { + "bbox": [ + 55, + 614, + 178, + 626 + ], + "type": "text", + "content": "4.1.1. Performance Analysis" + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "content": "On NYUv2. As a widely used benchmark dataset, NYUv2 is often employed to evaluate the effectiveness of models. From the first row of Table 1, we can observe that SigNet achieves competitive performance on the NYUv2 dataset, with an RMSE that is " + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "inline_equation", + "content": "5mm" + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "content": " lower than the second-best method. Additionally, although our method lags behind BPNet [38] by only 0.001 in the REL metric, SigNet signifi" + } + ] + } + ], + "index": 38 + }, + { + "bbox": [ + 313, + 519, + 515, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 519, + 515, + 532 + ], + "spans": [ + { + "bbox": [ + 313, + 519, + 515, + 532 + ], + "type": "text", + "content": "cantly reduces the number of parameters by " + }, + { + "bbox": [ + 313, + 519, + 515, + 532 + ], + "type": "inline_equation", + "content": "96\\%" + }, + { + "bbox": [ + 313, + 519, + 515, + 532 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 39 + }, + { + "bbox": [ + 313, + 533, + 555, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 555, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 555, + 628 + ], + "type": "text", + "content": "On DIML. DIML is collected from both indoor and outdoor scenes, suffering from severe edge shadows and irregular holes. Our method is implemented using the same experimental settings as previous approaches [2, 33]. As shown in the second row of Table 1, SigNet demonstrates the best performance across all evaluation metrics. For example, compared to the suboptimal model, our approach decreases the RMSE by " + }, + { + "bbox": [ + 313, + 533, + 555, + 628 + ], + "type": "inline_equation", + "content": "19\\%" + }, + { + "bbox": [ + 313, + 533, + 555, + 628 + ], + "type": "text", + "content": " and the REL by " + }, + { + "bbox": [ + 313, + 533, + 555, + 628 + ], + "type": "inline_equation", + "content": "45\\%" + }, + { + "bbox": [ + 313, + 533, + 555, + 628 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 40 + }, + { + "bbox": [ + 313, + 629, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 555, + 713 + ], + "type": "text", + "content": "On SUN RGBD. The large-scale SUN RGB-D dataset is leveraged to evaluate the generalization capability of our SigNet across different sensors and scenes. Table 1 (third row) reveals that our method surpasses previous state-of-the-art approaches across all metrics. Notably, SigNet achieves a " + }, + { + "bbox": [ + 313, + 629, + 555, + 713 + ], + "type": "inline_equation", + "content": "8mm" + }, + { + "bbox": [ + 313, + 629, + 555, + 713 + ], + "type": "text", + "content": " lower RMSE than the latest BPNet [38] and a " + }, + { + "bbox": [ + 313, + 629, + 555, + 713 + ], + "type": "inline_equation", + "content": "36mm" + }, + { + "bbox": [ + 313, + 629, + 555, + 713 + ], + "type": "text", + "content": " lower RMSE compared to AGG-Net [2]." + } + ] + } + ], + "index": 41 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26948" + } + ] + } + ], + "index": 42 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 65, + 70, + 286, + 129 + ], + "blocks": [ + { + "bbox": [ + 65, + 70, + 286, + 129 + ], + "lines": [ + { + "bbox": [ + 65, + 70, + 286, + 129 + ], + "spans": [ + { + "bbox": [ + 65, + 70, + 286, + 129 + ], + "type": "table", + "html": "
Density1%5%35%65%95%
TPVD [50]156.6139.7112.297.392.8
SigNet144.6121.599.490.184.3
Improv. ↑12.018.212.87.28.5
", + "image_path": "e78f307dde527af33a161b1487a26e624940b01c8b35bb439d305f79b817c6d3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 63, + 152, + 289, + 322 + ], + "blocks": [ + { + "bbox": [ + 63, + 152, + 289, + 322 + ], + "lines": [ + { + "bbox": [ + 63, + 152, + 289, + 322 + ], + "spans": [ + { + "bbox": [ + 63, + 152, + 289, + 322 + ], + "type": "image", + "image_path": "fb3a9a02651eeca5f3b0e2aeabcc78be927334df8256767d18eedc3ec1a38bb2.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 327, + 295, + 350 + ], + "lines": [ + { + "bbox": [ + 55, + 327, + 295, + 350 + ], + "spans": [ + { + "bbox": [ + 55, + 327, + 295, + 350 + ], + "type": "text", + "content": "Figure 7. Complexity comparison on the TOFDC dataset, where larger circle areas indicate higher inference time." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 365, + 295, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 365, + 295, + 462 + ], + "spans": [ + { + "bbox": [ + 55, + 365, + 295, + 462 + ], + "type": "text", + "content": "On TOFDC. TOFDC encompasses extensive data captured under various lighting conditions and in open spaces. It is utilized to further assess the performance of the proposed model. The quantitative results in the last row of Table 1 indicate that our method achieves state-of-the-art performance. Compared to the second-best TPVD [50], SigNet reduces the RMSE by " + }, + { + "bbox": [ + 55, + 365, + 295, + 462 + ], + "type": "inline_equation", + "content": "5mm" + }, + { + "bbox": [ + 55, + 365, + 295, + 462 + ], + "type": "text", + "content": " and the REL by 0.002, while also decreasing the number of parameters by " + }, + { + "bbox": [ + 55, + 365, + 295, + 462 + ], + "type": "inline_equation", + "content": "27.9M" + }, + { + "bbox": [ + 55, + 365, + 295, + 462 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 467, + 178, + 479 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 467, + 178, + 479 + ], + "spans": [ + { + "bbox": [ + 55, + 467, + 178, + 479 + ], + "type": "text", + "content": "4.1.2. Visualization Analysis" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 482, + 295, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 482, + 295, + 590 + ], + "spans": [ + { + "bbox": [ + 55, + 482, + 295, + 590 + ], + "type": "text", + "content": "Figs. 4 and 5 showcase visual results on TOFDC and DIML, highlighting SigNet's ability to recover precise depth structures. Compared to previous models, our method yields sharper predictions, such as the leaves in Fig. 4 and the chair in Fig. 5, achieving closer alignment with the ground truth. These results affirm the effectiveness of our approach in enhancing depth completion. Additionally, Fig. 6 presents some degradation examples, further demonstrating the ability of SigNet to model and address such challenges." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 595, + 185, + 606 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 595, + 185, + 606 + ], + "spans": [ + { + "bbox": [ + 55, + 595, + 185, + 606 + ], + "type": "text", + "content": "4.1.3. Number of Valid Points" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 609, + 295, + 670 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 609, + 295, + 670 + ], + "spans": [ + { + "bbox": [ + 55, + 609, + 295, + 670 + ], + "type": "text", + "content": "Tab. 2 lists a comparison of different valid points on DIML. It indicates that our SigNet consistently outperforms TPVD, achieving an average improvement of " + }, + { + "bbox": [ + 55, + 609, + 295, + 670 + ], + "type": "inline_equation", + "content": "9.7\\%" + }, + { + "bbox": [ + 55, + 609, + 295, + 670 + ], + "type": "text", + "content": " across the five densities. These results highlight the generalization capability of our method across varying valid points." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 674, + 173, + 686 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 674, + 173, + 686 + ], + "spans": [ + { + "bbox": [ + 55, + 674, + 173, + 686 + ], + "type": "text", + "content": "4.1.4. Complexity Analysis" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "content": "Fig. 7 shows that our method achieves an excellent balance among RMSE, parameters, and inference time. For exam-" + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 317, + 70, + 553, + 159 + ], + "blocks": [ + { + "bbox": [ + 61, + 134, + 289, + 145 + ], + "lines": [ + { + "bbox": [ + 61, + 134, + 289, + 145 + ], + "spans": [ + { + "bbox": [ + 61, + 134, + 289, + 145 + ], + "type": "text", + "content": "Table 2. Density comparison of TPVD and SigNet on TOFDC." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 70, + 553, + 159 + ], + "lines": [ + { + "bbox": [ + 317, + 70, + 553, + 159 + ], + "spans": [ + { + "bbox": [ + 317, + 70, + 553, + 159 + ], + "type": "table", + "html": "
SigNetDeg. BridgeRGB-D FusionRMSE (mm)
DensifyDenoiseConc.Att.DADF-DDADF-F
i108
ii✓✓95
iii✓✓✓92
iv✓✓✓74
v✓✓✓68
v✓✓✓✓55
", + "image_path": "80c5b7a042ec265a149a53143f820c504173d2b1b1ca499bdfd78f0aaa85845b.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 165, + 553, + 198 + ], + "lines": [ + { + "bbox": [ + 313, + 165, + 553, + 198 + ], + "spans": [ + { + "bbox": [ + 313, + 165, + 553, + 198 + ], + "type": "text", + "content": "Table 3. Ablation study of our SigNet on the DIML dataset. Conc.: concat, Att.: attention, DADF-D: degradation decomposition in DADF module, DADF-F: conditional Mamba fusion." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 218, + 555, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 218, + 555, + 291 + ], + "spans": [ + { + "bbox": [ + 313, + 218, + 555, + 291 + ], + "type": "text", + "content": "ple, compared to the fewer parameter CSPN [4], our SigNet not only reduces parameters by " + }, + { + "bbox": [ + 313, + 218, + 555, + 291 + ], + "type": "inline_equation", + "content": "81\\%" + }, + { + "bbox": [ + 313, + 218, + 555, + 291 + ], + "type": "text", + "content": " and inference time by " + }, + { + "bbox": [ + 313, + 218, + 555, + 291 + ], + "type": "inline_equation", + "content": "82\\%" + }, + { + "bbox": [ + 313, + 218, + 555, + 291 + ], + "type": "text", + "content": ", but also decreases RMSE by " + }, + { + "bbox": [ + 313, + 218, + 555, + 291 + ], + "type": "inline_equation", + "content": "61\\%" + }, + { + "bbox": [ + 313, + 218, + 555, + 291 + ], + "type": "text", + "content": ". Additionally, our method surpasses the latest TPVD [50] with a " + }, + { + "bbox": [ + 313, + 218, + 555, + 291 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 313, + 218, + 555, + 291 + ], + "type": "text", + "content": " lower RMSE and " + }, + { + "bbox": [ + 313, + 218, + 555, + 291 + ], + "type": "inline_equation", + "content": "78\\%" + }, + { + "bbox": [ + 313, + 218, + 555, + 291 + ], + "type": "text", + "content": " lower FPS, while reducing parameters by " + }, + { + "bbox": [ + 313, + 218, + 555, + 291 + ], + "type": "inline_equation", + "content": "89\\%" + }, + { + "bbox": [ + 313, + 218, + 555, + 291 + ], + "type": "text", + "content": " and inference time by " + }, + { + "bbox": [ + 313, + 218, + 555, + 291 + ], + "type": "inline_equation", + "content": "78\\%" + }, + { + "bbox": [ + 313, + 218, + 555, + 291 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 299, + 408, + 312 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 299, + 408, + 312 + ], + "spans": [ + { + "bbox": [ + 313, + 299, + 408, + 312 + ], + "type": "text", + "content": "4.2. Ablation Study" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 316, + 555, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 316, + 555, + 616 + ], + "spans": [ + { + "bbox": [ + 313, + 316, + 555, + 616 + ], + "type": "text", + "content": "SigNet Designs. Table 3 presents the ablation results on the DIML test split. The iteration of DADF is set to five. SigNet-i serves as the baseline, where both the degradation bridge and DADF module are removed, and the RGB-D features are fused using addition. Building on this, SigNet-ii redefines depth completion as depth enhancement through densification and degradation, concatenating the RGB and degradation features, resulting in a " + }, + { + "bbox": [ + 313, + 316, + 555, + 616 + ], + "type": "inline_equation", + "content": "13\\mathrm{mm}" + }, + { + "bbox": [ + 313, + 316, + 555, + 616 + ], + "type": "text", + "content": " improvement. This highlights the superiority of our task transformation approach. Additionally, SigNet-iii incorporates a Gaussian denoising operator to further reduce error. To better utilize the degradation, SigNet-iv applies a Sigmoid function and element-wise multiplication to obtain an attention map, enabling adaptive selection of RGB features. Consequently, performance improves significantly from " + }, + { + "bbox": [ + 313, + 316, + 555, + 616 + ], + "type": "inline_equation", + "content": "92\\mathrm{mm}" + }, + { + "bbox": [ + 313, + 316, + 555, + 616 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 316, + 555, + 616 + ], + "type": "inline_equation", + "content": "74\\mathrm{mm}" + }, + { + "bbox": [ + 313, + 316, + 555, + 616 + ], + "type": "text", + "content": ". Furthermore, SigNet-v employs degradation-aware decomposition, reducing the error by an additional " + }, + { + "bbox": [ + 313, + 316, + 555, + 616 + ], + "type": "inline_equation", + "content": "6\\mathrm{mm}" + }, + { + "bbox": [ + 313, + 316, + 555, + 616 + ], + "type": "text", + "content": ". The results of SigNet-iii, iv, and v demonstrate the effectiveness of our adaptive feature selection strategy. Notably, in SigNet-iii, iv, and v, the selected RGB and depth features are encoded through concatenation and convolutions. Finally, with the introduction of our conditional Mamba, SigNet-vi surpasses SigNet-v by " + }, + { + "bbox": [ + 313, + 316, + 555, + 616 + ], + "type": "inline_equation", + "content": "13\\mathrm{mm}" + }, + { + "bbox": [ + 313, + 316, + 555, + 616 + ], + "type": "text", + "content": " in RMSE, representing a significant improvement. Overall, each component positively impacts the baseline." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "type": "text", + "content": "Mamba Fusion and Iteration of DADF. Fig. 8 (a) illustrates the ablation study of Mamba fusion on the DIML dataset. For rapid validation, we employ DADF only once. The baseline ViM directly encodes the combined RGB-D features using the vanilla Vision Mamba [25]. When we apply our Conditional Mamba with initial degradation and RGB-D features as input, the error decreases by " + }, + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "type": "inline_equation", + "content": "8\\mathrm{mm}" + }, + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "type": "text", + "content": ". Additionally, the introduction of degradation decomposition" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26949" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 70, + 290, + 176 + ], + "blocks": [ + { + "bbox": [ + 61, + 70, + 290, + 176 + ], + "lines": [ + { + "bbox": [ + 61, + 70, + 290, + 176 + ], + "spans": [ + { + "bbox": [ + 61, + 70, + 290, + 176 + ], + "type": "image", + "image_path": "8ca0db0eb846cd175d65914b9fbdb356cc3241ac68fdaff946507be3ff0f360e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 65, + 182, + 285, + 194 + ], + "lines": [ + { + "bbox": [ + 65, + 182, + 285, + 194 + ], + "spans": [ + { + "bbox": [ + 65, + 182, + 285, + 194 + ], + "type": "text", + "content": "Figure 8. Ablation study of our DADF on the DIML dataset." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 64, + 205, + 286, + 243 + ], + "blocks": [ + { + "bbox": [ + 64, + 205, + 286, + 243 + ], + "lines": [ + { + "bbox": [ + 64, + 205, + 286, + 243 + ], + "spans": [ + { + "bbox": [ + 64, + 205, + 286, + 243 + ], + "type": "table", + "html": "
Path number357911
RMSE (mm)55.054.654.554.154.0
", + "image_path": "61b61ea52479da242e2640ed726c906b7414b5b982e1050219904e4cc010cb48.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 250, + 294, + 261 + ], + "lines": [ + { + "bbox": [ + 56, + 250, + 294, + 261 + ], + "spans": [ + { + "bbox": [ + 56, + 250, + 294, + 261 + ], + "type": "text", + "content": "Table 4. Ablation on the number of decomposition path in DADF." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 282, + 296, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 282, + 296, + 462 + ], + "spans": [ + { + "bbox": [ + 55, + 282, + 296, + 462 + ], + "type": "text", + "content": "using DCT further enhances performance. We also observe that the results improve as the number of DCT components increases, highlighting the effectiveness of our gradual DCT decomposition design. Fig. 8 (b) presents the ablation study of our DADF with different iterations on the DIML dataset. We find that performance improves with an increasing number of iterations. For an optimal balance between efficiency and effectiveness, we set the final iteration count to five. Moreover, as shown in Fig.3, the degradation " + }, + { + "bbox": [ + 55, + 282, + 296, + 462 + ], + "type": "inline_equation", + "content": "\\mathbf{D}" + }, + { + "bbox": [ + 55, + 282, + 296, + 462 + ], + "type": "text", + "content": " can be decomposed into " + }, + { + "bbox": [ + 55, + 282, + 296, + 462 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 282, + 296, + 462 + ], + "type": "text", + "content": " components using DCT. As indicated in Tab.4, the error gradually decreases as " + }, + { + "bbox": [ + 55, + 282, + 296, + 462 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 282, + 296, + 462 + ], + "type": "text", + "content": " increases. However, considering the computational complexity, we select the top-3 most representative components to focus on the high-frequency components of " + }, + { + "bbox": [ + 55, + 282, + 296, + 462 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{img}" + }, + { + "bbox": [ + 55, + 282, + 296, + 462 + ], + "type": "text", + "content": " and update the Mamba condition accordingly." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 468, + 138, + 480 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 468, + 138, + 480 + ], + "spans": [ + { + "bbox": [ + 55, + 468, + 138, + 480 + ], + "type": "text", + "content": "4.3. Failure Case" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 486, + 295, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 486, + 295, + 568 + ], + "spans": [ + { + "bbox": [ + 55, + 486, + 295, + 568 + ], + "type": "text", + "content": "Our method is primarily evaluated in indoor scenes and some simple outdoor scenes with closer distances. By constraining dense predictions with coarse depth maps, SigNet can effectively predict their degradation representations. However, as depicted in Fig. 9, SigNet fails to estimate reasonable degradation on the outdoor KITTI benchmark [40]. We identify two main reasons for this:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 570, + 295, + 689 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 55, + 570, + 295, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 295, + 629 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 295, + 629 + ], + "type": "text", + "content": "1) The ground truth depth maps of KITTI are very sparse, with only about " + }, + { + "bbox": [ + 55, + 570, + 295, + 629 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 55, + 570, + 295, + 629 + ], + "type": "text", + "content": " valid pixels. This sparsity leads to inconsistent depth edges, resulting in a relatively weak constraint. For instance, the degradation in Fig. 9 even introduces some initial LiDAR lines." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 630, + 295, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 295, + 689 + ], + "type": "text", + "content": "2) The depth distance of KITTI data is much larger, which significantly weakens edge details. As observed, the degradation in Fig. 9 can only learn very coarse and blurry representations near edges at short range, while at long range, it learns incorrect representations." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": "One potential solution is to provide the model with supervisory signals regarding dense contents and precise" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 315, + 70, + 553, + 308 + ], + "blocks": [ + { + "bbox": [ + 315, + 70, + 553, + 308 + ], + "lines": [ + { + "bbox": [ + 315, + 70, + 553, + 308 + ], + "spans": [ + { + "bbox": [ + 315, + 70, + 553, + 308 + ], + "type": "image", + "image_path": "6e2213f0f81a15f97da16e501338e98fef2ad3febb47a980a081f121657ae439.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 314, + 553, + 337 + ], + "lines": [ + { + "bbox": [ + 313, + 314, + 553, + 337 + ], + "spans": [ + { + "bbox": [ + 313, + 314, + 553, + 337 + ], + "type": "text", + "content": "Figure 9. Visual results of failure cases on KITTI benchmark [40]. DA v2 denotes the large depth model DepthAnything v2 [52]." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 354, + 555, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 354, + 555, + 475 + ], + "spans": [ + { + "bbox": [ + 313, + 354, + 555, + 475 + ], + "type": "text", + "content": "edges. It is well-known that large depth models can produce dense depth with impressive sharp edges. For example, we employ DepthAnything v2 [52] to generate the corresponding dense depth. We discover that the edges of cars and even tree leaves are much clearer. Although the depth scale does not match that of KITTI, we can generate a map that offers constraints near edges via gradient or boundary extraction techniques [46]. Certainly, the results of DepthAnything v2 [52] or DepthAnything v2 itself can also be utilized to facilitate depth completion, warranting further exploration." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 483, + 388, + 496 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 483, + 388, + 496 + ], + "spans": [ + { + "bbox": [ + 313, + 483, + 388, + 496 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 504, + 555, + 660 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 504, + 555, + 660 + ], + "spans": [ + { + "bbox": [ + 313, + 504, + 555, + 660 + ], + "type": "text", + "content": "In this paper, we redefined depth completion as depth enhancement for the first time with our novel framework, SigNet. SigNet leveraged non-CNN densification tools to transform sparse depth data into coarse yet dense depth, fully eliminating the mismatch and ambiguity caused by direct convolution over irregularly sampled sparse data. More importantly, by establishing a self-supervised degradation bridge, SigNet effectively fused RGB-D data. It utilized degradation decomposition to dynamically select high-frequency components of RGB and build a conditional Mamba for the depth enhancement. Extensive experiments validated the superiority of SigNet, highlighting its potential for depth related applications." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 670, + 417, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 670, + 417, + 682 + ], + "spans": [ + { + "bbox": [ + 313, + 670, + 417, + 682 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "content": "This work was supported by the National Natural Science Fund of China (U24A20330, 62361166670 and 62072242)." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26950" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 61, + 91, + 294, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 294, + 123 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 294, + 123 + ], + "type": "text", + "content": "[1] Antoni Buades, Bartomeu Coll, and J-M Morel. A non-local algorithm for image denoising. In CVPR, pages 60-65, 2005. 5, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 125, + 295, + 170 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 125, + 295, + 170 + ], + "spans": [ + { + "bbox": [ + 61, + 125, + 295, + 170 + ], + "type": "text", + "content": "[2] Dongyue Chen, Tingxuan Huang, Zhimin Song, Shizhuo Deng, and Tong Jia. Agg-net: Attention guided gated convolutional network for depth image completion. In ICCV, pages 8853-8862, 2023. 5, 6, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 172, + 294, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 172, + 294, + 205 + ], + "spans": [ + { + "bbox": [ + 62, + 172, + 294, + 205 + ], + "type": "text", + "content": "[3] Yun Chen, Bin Yang, Ming Liang, and Raquel Urtasun. Learning joint 2d-3d representations for depth completion. In ICCV, pages 10023–10032, 2019. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 207, + 294, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 207, + 294, + 239 + ], + "spans": [ + { + "bbox": [ + 62, + 207, + 294, + 239 + ], + "type": "text", + "content": "[4] Xinjing Cheng, Peng Wang, and Ruigang Yang. Learning depth with convolutional spatial propagation network. In ECCV, pages 103-119, 2018. 5, 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 243, + 294, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 243, + 294, + 275 + ], + "spans": [ + { + "bbox": [ + 62, + 243, + 294, + 275 + ], + "type": "text", + "content": "[5] Xinjing Cheng, Peng Wang, and Ruigang Yang. Depth estimation via affinity learned with convolutional spatial propagation network. In ECCV, pages 103-119, 2018. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 277, + 294, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 277, + 294, + 319 + ], + "spans": [ + { + "bbox": [ + 62, + 277, + 294, + 319 + ], + "type": "text", + "content": "[6] Xinjing Cheng, Peng Wang, and Ruigang Yang. Learning depth with convolutional spatial propagation network. IEEE Transactions on Pattern Analysis and Machine Intelligence, 42(10):2361-2379, 2019. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 323, + 294, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 323, + 294, + 365 + ], + "spans": [ + { + "bbox": [ + 62, + 323, + 294, + 365 + ], + "type": "text", + "content": "[7] Xinjing Cheng, Peng Wang, Chenye Guan, and Ruigang Yang. Cspn++: Learning context and resource aware convolutional spatial propagation networks for depth completion. In AAAI, pages 10615-10622, 2020. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 369, + 294, + 411 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 369, + 294, + 411 + ], + "spans": [ + { + "bbox": [ + 62, + 369, + 294, + 411 + ], + "type": "text", + "content": "[8] Jaehoon Cho, Dongbo Min, Youngjung Kim, and Kwanghoon Sohn. A large rgb-d dataset for semi-supervised monocular depth estimation. arXiv preprint arXiv:1904.10230, 2019. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 415, + 294, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 415, + 294, + 458 + ], + "spans": [ + { + "bbox": [ + 62, + 415, + 294, + 458 + ], + "type": "text", + "content": "[9] Abdelrahman Eldisokey, Michael Felsberg, and Fahad Shahbaz Khan. Confidence propagation through cnns for guided sparse depth regression. IEEE Transactions on Pattern Analysis and Machine Intelligence, 42(10):2423-2436, 2020. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 460, + 294, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 460, + 294, + 492 + ], + "spans": [ + { + "bbox": [ + 57, + 460, + 294, + 492 + ], + "type": "text", + "content": "[10] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023. 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 495, + 294, + 528 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 495, + 294, + 528 + ], + "spans": [ + { + "bbox": [ + 57, + 495, + 294, + 528 + ], + "type": "text", + "content": "[11] Mu Hu, Shuling Wang, Bin Li, Shiyu Ning, Li Fan, and Xiaojin Gong. Penet: Towards precise and efficient image guided depth completion. In ICRA, 2021. 2, 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 530, + 294, + 573 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 530, + 294, + 573 + ], + "spans": [ + { + "bbox": [ + 57, + 530, + 294, + 573 + ], + "type": "text", + "content": "[12] Maximilian Jaritz, Raoul De Charette, Emilie Wirbel, Xavier Perrotton, and Fawzi Nashashibi. Sparse and dense data with cnns: Depth completion and semantic segmentation. In 3DV, pages 52-60, 2018. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 576, + 294, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 576, + 294, + 597 + ], + "spans": [ + { + "bbox": [ + 57, + 576, + 294, + 597 + ], + "type": "text", + "content": "[13] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In Computer Ence, 2014. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 600, + 294, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 600, + 294, + 632 + ], + "spans": [ + { + "bbox": [ + 57, + 600, + 294, + 632 + ], + "type": "text", + "content": "[14] Jason Ku, Ali Harakeh, and Steven L Waslander. In defense of classical image processing: Fast depth completion on the cpu. In CRV, pages 16-22, 2018. 1, 2, 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 635, + 294, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 635, + 294, + 666 + ], + "spans": [ + { + "bbox": [ + 57, + 635, + 294, + 666 + ], + "type": "text", + "content": "[15] Anat Levin, Dani Lischinski, and Yair Weiss. Colorization using optimization. In SIGGRAPH, pages 689-694, 2004. 1, 2, 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 57, + 670, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 294, + 712 + ], + "type": "text", + "content": "[16] Ang Li, Zejian Yuan, Yonggen Ling, Wanchao Chi, Chong Zhang, et al. A multi-scale guided cascade hourglass network for depth completion. In WACV, pages 32-40, 2020. 2" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 34, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "text", + "content": "[17] Boyun Li, Xiao Liu, Peng Hu, Zhongqin Wu, Jiancheng Lv, and Xi Peng. All-in-one image restoration for unknown corruption. In CVPR, pages 17452-17462, 2022. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 108, + 553, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 108, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 317, + 108, + 553, + 140 + ], + "type": "text", + "content": "[18] Jie Liang, Hui Zeng, and Lei Zhang. Efficient and degradation-adaptive network for real-world image superresolution. In ECCV, pages 574-591, 2022. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 143, + 553, + 175 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 143, + 553, + 175 + ], + "spans": [ + { + "bbox": [ + 317, + 143, + 553, + 175 + ], + "type": "text", + "content": "[19] Yuankai Lin, Tao Cheng, Qi Zhong, Wending Zhou, and Hua Yang. Dynamic spatial propagation network for depth completion. In AAAI, pages 1638-1646, 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 177, + 553, + 220 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 177, + 553, + 220 + ], + "spans": [ + { + "bbox": [ + 317, + 177, + 553, + 220 + ], + "type": "text", + "content": "[20] Yuankai Lin, Hua Yang, Tao Cheng, Wending Zhou, and Zhouping Yin. Dyspn: Learning dynamic affinity for image-guided depth completion. IEEE Transactions on Circuits and Systems for Video Technology, pages 1-1, 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 223, + 553, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 223, + 553, + 266 + ], + "spans": [ + { + "bbox": [ + 316, + 223, + 553, + 266 + ], + "type": "text", + "content": "[21] Lina Liu, Yiyi Liao, Yue Wang, Andreas Geiger, and Yong Liu. Learning steering kernels for guided depth completion. IEEE Transactions on Image Processing, 30:2850-2861, 2021. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 269, + 553, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 269, + 553, + 312 + ], + "spans": [ + { + "bbox": [ + 316, + 269, + 553, + 312 + ], + "type": "text", + "content": "[22] Lina Liu, Xibin Song, Xiaoyang Lyu, Junwei Diao, Mengmeng Wang, Yong Liu, and Liangjun Zhang. Fcfr-net: Feature fusion based coarse-to-fine residual learning for depth completion. In AAAI, pages 2136-2144, 2021. 1, 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 314, + 553, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 314, + 553, + 346 + ], + "spans": [ + { + "bbox": [ + 316, + 314, + 553, + 346 + ], + "type": "text", + "content": "[23] Sifei Liu, Shalini De Mello, Jinwei Gu, Guangyu Zhong, Ming-Hsuan Yang, and Jan Kautz. Learning affinity via spatial propagation networks. In NeurIPS, 2017. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 349, + 553, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 349, + 553, + 381 + ], + "spans": [ + { + "bbox": [ + 316, + 349, + 553, + 381 + ], + "type": "text", + "content": "[24] Xin Liu, Xiaofei Shao, Bo Wang, Yali Li, and Shengjin Wang. Graphcspn: Geometry-aware depth completion via dynamic gcns. In ECCV, pages 90-107. Springer, 2022. 2, 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 384, + 553, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 384, + 553, + 426 + ], + "spans": [ + { + "bbox": [ + 317, + 384, + 553, + 426 + ], + "type": "text", + "content": "[25] Yue Liu, Yunjie Tian, Yuzhong Zhao, Hongtian Yu, Lingxi Xie, Yaowei Wang, Qixiang Ye, and Yunfan Liu. Vmamba: Visual state space model. arXiv preprint arXiv:2401.10166, 2024. 2, 4, 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 429, + 553, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 429, + 553, + 472 + ], + "spans": [ + { + "bbox": [ + 317, + 429, + 553, + 472 + ], + "type": "text", + "content": "[26] Kaiyue Lu, Nick Barnes, Saeed Anwar, and Liang Zheng. From depth what can you see? depth completion via auxiliary image reconstruction. In CVPR, pages 11306-11315, 2020. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 475, + 553, + 507 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 475, + 553, + 507 + ], + "spans": [ + { + "bbox": [ + 317, + 475, + 553, + 507 + ], + "type": "text", + "content": "[27] Fangchang Ma and Sertac Karaman. Sparse-to-dense: Depth prediction from sparse depth samples and a single image. In ICRA, pages 4796-4803. IEEE, 2018. 1" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 317, + 509, + 553, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 509, + 553, + 552 + ], + "spans": [ + { + "bbox": [ + 317, + 509, + 553, + 552 + ], + "type": "text", + "content": "[28] Fangchang Ma, Guilherme Venturelli Cavalheiro, and Sertac Karaman. Self-supervised sparse-to-dense: Self-supervised depth completion from lidar and monocular camera. In ICRA, 2019. 2, 5, 6" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 317, + 555, + 553, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 555, + 553, + 587 + ], + "spans": [ + { + "bbox": [ + 317, + 555, + 553, + 587 + ], + "type": "text", + "content": "[29] Jinsun Park, Kyungdon Joo, Zhe Hu, Chi-Kuei Liu, and In So Kweon. Non-local spatial propagation network for depth completion. In ECCV, 2020. 2, 5, 6" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 317, + 590, + 553, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 590, + 553, + 622 + ], + "spans": [ + { + "bbox": [ + 317, + 590, + 553, + 622 + ], + "type": "text", + "content": "[30] Jin-Hwi Park, Chanhwi Jeong, Junoh Lee, and Hae-Gon Jeon. Depth prompting for sensor-agnostic depth estimation. In CVPR, pages 9859–9869, 2024. 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 317, + 624, + 553, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 624, + 553, + 678 + ], + "spans": [ + { + "bbox": [ + 317, + 624, + 553, + 678 + ], + "type": "text", + "content": "[31] Jiaxiong Qiu, Zhaopeng Cui, Yinda Zhang, Xingdi Zhang, Shuaicheng Liu, Bing Zeng, and Marc Pollefeys. Deep surface normal guided depth prediction for outdoor scene from sparse lidar data and single color image. In CVPR, pages 3313-3322, 2019. 5, 2" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "text", + "content": "[32] Kyeongha Rho, Jinsung Ha, and Youngjung Kim. Guideformer: Transformers for image guided depth completion. In CVPR, pages 6250-6259, 2022. 2" + } + ] + } + ], + "index": 33 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "26951" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 115 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 115 + ], + "type": "text", + "content": "[33] Dmitry Senushkin, Mikhail Romanov, Ilia Belikov, Nikolay Patakin, and Anton Konushin. Decoder modulation for indoor depth completion. In IROS, pages 2181-2188. IEEE, 2021. 5, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 172 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 172 + ], + "type": "text", + "content": "[34] Shreyas S Shivakumar, Ty Nguyen, Ian D Miller, Steven W Chen, Vijay Kumar, and Camillo J Taylor. Dfusenet: Deep fusion of rgb and sparse depth information for image guided dense depth completion. In ITSC, pages 13-20. IEEE, 2019. 5, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 174, + 294, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 174, + 294, + 207 + ], + "spans": [ + { + "bbox": [ + 56, + 174, + 294, + 207 + ], + "type": "text", + "content": "[35] Nathan Silberman, Derek Hoiem, Pushmeet Kohli, and Rob Fergus. Indoor segmentation and support inference from rgbd images. In ECCV, pages 746-760. Springer, 2012. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 208, + 294, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 208, + 294, + 240 + ], + "spans": [ + { + "bbox": [ + 56, + 208, + 294, + 240 + ], + "type": "text", + "content": "[36] Shuran Song, Samuel P Lichtenberg, and Jianxiong Xiao. Sun rgb-d: A rgb-d scene understanding benchmark suite. In CVPR, pages 567-576, 2015. 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 241, + 294, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 241, + 294, + 284 + ], + "spans": [ + { + "bbox": [ + 56, + 241, + 294, + 284 + ], + "type": "text", + "content": "[37] Jie Tang, Fei-Peng Tian, Wei Feng, Jian Li, and Ping Tan. Learning guided convolutional network for depth completion. IEEE Transactions on Image Processing, 30:1116-1129, 2020. 2, 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 286, + 294, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 286, + 294, + 319 + ], + "spans": [ + { + "bbox": [ + 56, + 286, + 294, + 319 + ], + "type": "text", + "content": "[38] Jie Tang, Fei-Peng Tian, Boshi An, Jian Li, and Ping Tan. Bilateral propagation network for depth completion. In CVPR, pages 9763-9772, 2024. 1, 2, 5, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 320, + 294, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 320, + 294, + 353 + ], + "spans": [ + { + "bbox": [ + 56, + 320, + 294, + 353 + ], + "type": "text", + "content": "[39] Diego Thomas and Akihiro Sugimoto. A flexible scene representation for 3d reconstruction using an rgb-d camera. In ICCV, pages 2800-2807, 2013. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 354, + 294, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 354, + 294, + 387 + ], + "spans": [ + { + "bbox": [ + 56, + 354, + 294, + 387 + ], + "type": "text", + "content": "[40] Jonas Uhrig, Nick Schneider, Lukas Schneider, Uwe Franke, Thomas Brox, and Andreas Geiger. Sparsity invariant cnns. In 3DV, pages 11–20, 2017. 1, 2, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 388, + 294, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 388, + 294, + 422 + ], + "spans": [ + { + "bbox": [ + 56, + 388, + 294, + 422 + ], + "type": "text", + "content": "[41] Wouter Van Gansbeke, Davy Neven, Bert De Brabandere, and Luc Van Gool. Sparse and noisy lidar completion with rgb guidance and uncertainty. In MVA, pages 1-6, 2019. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 422, + 294, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 422, + 294, + 465 + ], + "spans": [ + { + "bbox": [ + 56, + 422, + 294, + 465 + ], + "type": "text", + "content": "[42] Haowen Wang, Mingyuan Wang, Zhengping Che, Zhiyuan Xu, Xiuquan Qiao, Mengshi Qi, Feifei Feng, and Jian Tang. RGB-depth fusion gan for indoor depth completion. In CVPR, pages 6209-6218, 2022. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 467, + 294, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 467, + 294, + 510 + ], + "spans": [ + { + "bbox": [ + 56, + 467, + 294, + 510 + ], + "type": "text", + "content": "[43] Longguang Wang, Yingqian Wang, Xiaoyu Dong, Qingyu Xu, Jungang Yang, Wei An, and Yulan Guo. Unsupervised degradation representation learning for blind superresolution. In CVPR, pages 10581-10590, 2021. 1, 2, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 512, + 294, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 512, + 294, + 544 + ], + "spans": [ + { + "bbox": [ + 56, + 512, + 294, + 544 + ], + "type": "text", + "content": "[44] Yufei Wang, Bo Li, Ge Zhang, Qi Liu, Tao Gao, and Yuchao Dai. Lrru: Long-short range recurrent updating networks for depth completion. In ICCV, pages 9422-9432, 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 545, + 294, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 545, + 294, + 588 + ], + "spans": [ + { + "bbox": [ + 56, + 545, + 294, + 588 + ], + "type": "text", + "content": "[45] Yufei Wang, Ge Zhang, Shaoqian Wang, Bo Li, Qi Liu, Le Hui, and Yuchao Dai. Improving depth completion via depth feature upsampling. In CVPR, pages 21104-21113, 2024. 1, 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 590, + 294, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 590, + 294, + 633 + ], + "spans": [ + { + "bbox": [ + 56, + 590, + 294, + 633 + ], + "type": "text", + "content": "[46] Zhengxue Wang, Zhiqiang Yan, and Jian Yang. Sgnet: Structure guided network via gradient-frequency awareness for depth map super-resolution. In AAAI, pages 5823-5831, 2024. 8" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 635, + 294, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 294, + 677 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 294, + 677 + ], + "type": "text", + "content": "[47] Zhiqiang Yan, Xiang Li, Kun Wang, Zhenyu Zhang, Jun Li, and Jian Yang. Multi-modal masked pre-training for monocular panoramic depth completion. In ECCV, pages 378-395, 2022. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 56, + 680, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 680, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 680, + 294, + 713 + ], + "type": "text", + "content": "[48] Zhiqiang Yan, Kun Wang, Xiang Li, Zhenyu Zhang, Jun Li, and Jian Yang. Rignet: Repetitive image guided network for depth completion. In ECCV, pages 214-230, 2022. 1, 2, 5, 6" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 72, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 315, + 72, + 553, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 72, + 553, + 115 + ], + "spans": [ + { + "bbox": [ + 315, + 72, + 553, + 115 + ], + "type": "text", + "content": "[49] Zhiqiang Yan, Kun Wang, Xiang Li, Zhenyu Zhang, Jun Li, and Jian Yang. Desnet: Decomposed scale-consistent network for unsupervised depth completion. In AAAI, pages 3109-3117, 2023. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 315, + 118, + 553, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 118, + 553, + 163 + ], + "spans": [ + { + "bbox": [ + 315, + 118, + 553, + 163 + ], + "type": "text", + "content": "[50] Zhiqiang Yan, Yuankai Lin, Kun Wang, Yupeng Zheng, Yufei Wang, Zhenyu Zhang, Jun Li, and Jian Yang. Tri-perspective view decomposition for geometry-aware depth completion. In CVPR, pages 4874-4884, 2024. 1, 2, 5, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 315, + 165, + 553, + 198 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 165, + 553, + 198 + ], + "spans": [ + { + "bbox": [ + 315, + 165, + 553, + 198 + ], + "type": "text", + "content": "[51] Zhiqiang Yan, Yupeng Zheng, Deng-Ping Fan, Xiang Li, Jun Li, and Jian Yang. Learnable differencing center for nighttime depth perception. Visual Intelligence, 2(1):15, 2024. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 315, + 199, + 553, + 232 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 199, + 553, + 232 + ], + "spans": [ + { + "bbox": [ + 315, + 199, + 553, + 232 + ], + "type": "text", + "content": "[52] Lihe Yang, Bingyi Kang, Zilong Huang, Zhen Zhao, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything v2. arXiv preprint arXiv:2406.09414, 2024. 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 315, + 234, + 553, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 234, + 553, + 267 + ], + "spans": [ + { + "bbox": [ + 315, + 234, + 553, + 267 + ], + "type": "text", + "content": "[53] Yanchao Yang, Alex Wong, and Stefano Soatto. Dense depth posterior (ddp) from single image and sparse range. In CVPR, pages 3353-3362, 2020. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 315, + 269, + 553, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 269, + 553, + 312 + ], + "spans": [ + { + "bbox": [ + 315, + 269, + 553, + 312 + ], + "type": "text", + "content": "[54] Mingde Yao, Ruikang Xu, Yuanshen Guan, Jie Huang, and Zhiwei Xiong. Neural degradation representation learning for all-in-one image restoration. IEEE Transactions on Image Processing, 2024. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 315, + 315, + 553, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 315, + 553, + 369 + ], + "spans": [ + { + "bbox": [ + 315, + 315, + 553, + 369 + ], + "type": "text", + "content": "[55] Guanghao Yin, Wei Wang, Zehuan Yuan, Wei Ji, Dongdong Yu, Shouqian Sun, Tat-Seng Chua, and Changhu Wang. Conditional hyper-network for blind super-resolution with multiple degradations. IEEE Transactions on Image Processing, 31:3949-3960, 2022. 2, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 315, + 371, + 553, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 371, + 553, + 414 + ], + "spans": [ + { + "bbox": [ + 315, + 371, + 553, + 414 + ], + "type": "text", + "content": "[56] Zhu Yu, Zehua Sheng, Zili Zhou, Lun Luo, Si-Yuan Cao, Hong Gu, Huaqi Zhang, and Hui-Liang Shen. Aggregating feature point cloud for depth completion. In ICCV, pages 8732-8743, 2023. 2, 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 315, + 417, + 553, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 417, + 553, + 460 + ], + "spans": [ + { + "bbox": [ + 315, + 417, + 553, + 460 + ], + "type": "text", + "content": "[57] Cheng Zhang, Yu Zhu, Qingsen Yan, Jinqiu Sun, and Yanning Zhang. All-in-one multi-degradation image restoration network via hierarchical degradation representation. In ACM MM, pages 2285-2293, 2023. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 315, + 462, + 553, + 506 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 462, + 553, + 506 + ], + "spans": [ + { + "bbox": [ + 315, + 462, + 553, + 506 + ], + "type": "text", + "content": "[58] Jinghao Zhang, Jie Huang, Mingde Yao, Zizheng Yang, Hu Yu, Man Zhou, and Feng Zhao. Ingredient-oriented multi-degradation learning for image restoration. In CVPR, pages 5825-5835, 2023. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 315, + 508, + 553, + 551 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 508, + 553, + 551 + ], + "spans": [ + { + "bbox": [ + 315, + 508, + 553, + 551 + ], + "type": "text", + "content": "[59] Kai Zhang, Jingyun Liang, Luc Van Gool, and Radu Timofte. Designing a practical degradation model for deep blind image super-resolution. In ICCV, pages 4791-4800, 2021. 1, 2, 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 315, + 554, + 553, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 554, + 553, + 597 + ], + "spans": [ + { + "bbox": [ + 315, + 554, + 553, + 597 + ], + "type": "text", + "content": "[60] Yulun Zhang, Kunpeng Li, Kai Li, Lichen Wang, Bineng Zhong, and Yun Fu. Image super-resolution using very deep residual channel attention networks. In ECCV, pages 286-301, 2018. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 315, + 600, + 553, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 600, + 553, + 643 + ], + "spans": [ + { + "bbox": [ + 315, + 600, + 553, + 643 + ], + "type": "text", + "content": "[61] Youmin Zhang, Xianda Guo, Matteo Poggi, Zheng Zhu, Guan Huang, and Stefano Mattoccia. Completionformer: Depth completion with convolutions and vision transformers. In CVPR, pages 18527-18536, 2023. 2, 5" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 315, + 646, + 553, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 646, + 553, + 677 + ], + "spans": [ + { + "bbox": [ + 315, + 646, + 553, + 677 + ], + "type": "text", + "content": "[62] Zhiwei Zhong, Xianming Liu, Junjun Jiang, Debin Zhao, and Xiangyang Ji. Guided depth map super-resolution: A survey. ACM Computing Surveys, 55(14s):1-36, 2023. 1, 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 315, + 680, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 680, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 315, + 680, + 553, + 713 + ], + "type": "text", + "content": "[63] Hanyu Zhou, Yi Chang, Wending Yan, and Luxin Yan. Unsupervised cumulative domain adaptation for foggy scene optical flow. In CVPR, pages 9569-9578, 2023. 1" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "26952" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 318 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 126 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 126 + ], + "type": "text", + "content": "[64] Hongyang Zhou, Xiaobin Zhu, Jianqing Zhu, Zheng Han, Shi-Xue Zhang, Jingyan Qin, and Xu-Cheng Yin. Learning correction filter via degradation-adaptive regression for blind single image super-resolution. In ICCV, pages 12365-12375, 2023. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 294, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 294, + 183 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 294, + 183 + ], + "type": "text", + "content": "[65] Hanyu Zhou, Yi Chang, Zhiwei Shi, Wending Yan, Gang Chen, Yonghong Tian, and Luxin Yan. Adverse weather optical flow: Cumulative homogeneous-heterogeneous adaptation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 185, + 294, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 185, + 294, + 228 + ], + "spans": [ + { + "bbox": [ + 56, + 185, + 294, + 228 + ], + "type": "text", + "content": "[66] Wending Zhou, Xu Yan, Yinghong Liao, Yuankai Lin, Jin Huang, Gangming Zhao, Shuguang Cui, and Zhen Li. Bev@ dc: Bird's-eye view assisted training for depth completion. In CVPR, pages 9233-9242, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 230, + 294, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 230, + 294, + 272 + ], + "spans": [ + { + "bbox": [ + 56, + 230, + 294, + 272 + ], + "type": "text", + "content": "[67] Ziyue Zhu, Qiang Meng, Xiao Wang, Ke Wang, Liujiang Yan, and Jian Yang. Curricular object manipulation in lidar-based object detection. In CVPR, pages 1125-1135, 2023. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 274, + 294, + 318 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 274, + 294, + 318 + ], + "spans": [ + { + "bbox": [ + 56, + 274, + 294, + 318 + ], + "type": "text", + "content": "[68] Ziyue Zhu, Shenlong Wang, Jin Xie, Jiang-Jiang Liu, Jingdong Wang, and Jian Yang. Voxelsplat: Dynamic gaussian splatting as an effective loss for occupancy and flow prediction. In CVPR, 2025. 1" + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "26953" + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Complexity Experts are Task-Discriminative Learners for Any Image Restoration/811da5ba-4d1d-41b9-ac9b-e649712870d3_content_list.json b/2025/Complexity Experts are Task-Discriminative Learners for Any Image Restoration/811da5ba-4d1d-41b9-ac9b-e649712870d3_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..8b9a52de1e1f05591061789429ab9bf2d4c34ae8 --- /dev/null +++ b/2025/Complexity Experts are Task-Discriminative Learners for Any Image Restoration/811da5ba-4d1d-41b9-ac9b-e649712870d3_content_list.json @@ -0,0 +1,1391 @@ +[ + { + "type": "text", + "text": "Complexity Experts are Task-Discriminative Learners for Any Image Restoration", + "text_level": 1, + "bbox": [ + 91, + 130, + 906, + 154 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Eduard Zamfir1 Zongwei Wu1* Nancy Mehta1 Yuedong Tan1 Danda Pani Paudel2 Yulun Zhang3 Radu Timofte1", + "bbox": [ + 230, + 179, + 764, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1Computer Vision Lab, CAIDAS & IFI, University of Würzburg", + "bbox": [ + 305, + 218, + 692, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2INSAIT, Sofia University \"St. Kliment Ohridski\" 3Shanghai Jiao Tong University", + "bbox": [ + 243, + 234, + 751, + 252 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 286, + 326, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent advancements in all-in-one image restoration models have revolutionized the ability to address diverse degradations through a unified framework. However, parameters tied to specific tasks often remain inactive for other tasks, making mixture-of-experts (MoE) architectures a natural extension. Despite this, MoEs often show inconsistent behavior, with some experts unexpectedly generalizing across tasks while others struggle within their intended scope. This hinders leveraging MoEs' computational benefits by bypassing irrelevant experts during inference. We attribute this undesired behavior to the uniform and rigid architecture of traditional MoEs. To address this, we introduce \"complexity experts\" - flexible expert blocks with varying computational complexity and receptive fields. A key challenge is assigning tasks to each expert, as degradation complexity is unknown in advance. Thus, we execute tasks with a simple bias toward lower complexity. To our surprise, this preference effectively drives task-specific allocation, assigning tasks to experts with the appropriate complexity. Extensive experiments validate our approach, demonstrating the ability to bypass irrelevant experts during inference while maintaining superior performance. The proposed MoCE-IR model outperforms state-of-the-art methods, affirming its efficiency and practical applicability. The source code and models are publicly available at eduardzamfir.github.io/MoCE-IR/", + "bbox": [ + 88, + 316, + 485, + 696 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 717, + 220, + 732 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Image restoration [47, 67, 68] is a fundamental problem in computer vision, dealing with reconstructing high-quality images from deteriorated observations. Adverse conditions such as noise [22, 67], haze [5, 38], or rain [7, 59] significantly impact the practical utility of images in downstream applications across various domains, including autonomous navigation [7, 49] or augmented reality [10, 15, 45]. Deep learning-based approaches have remarkably advanced this field, particularly for task-specific image restoration prob", + "bbox": [ + 89, + 742, + 483, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9ca792966f6d0aa5f464e7d8e451aceaf073605373aab8e6a655da80f6d992b6.jpg", + "image_caption": [ + "(a) Dense Model" + ], + "image_footnote": [], + "bbox": [ + 516, + 285, + 689, + 377 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/69324a597e6cf8a2befc7729dad123701ab939bdfff45e37303cf1462c07513c.jpg", + "image_caption": [ + "(b) MoE" + ], + "image_footnote": [], + "bbox": [ + 709, + 285, + 901, + 377 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9d9df16144abeaa5ca33f19b160da4030b5c8dc685c5800b445a29032c84e6d8.jpg", + "image_caption": [ + "(c) Complexity Experts (ours)" + ], + "image_footnote": [], + "bbox": [ + 516, + 407, + 694, + 500 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b46b5d1a1584d2df93afcce3f9e2ab92f068210dd752173968389d617f0a7802.jpg", + "image_caption": [ + "Figure 1. Motivation. (a) Dense all-in-one restoration methods [24, 57] often inefficiently allocate parameters when handling multiple degradation types. (b) While recent Mixture-of-Experts (MoE) approaches [61, 62] address this through sparse computation, their rigid routing mechanisms uniformly distribute inputs across experts without considering the natural relationships between degradations. (c) To overcome these limitations, we introduce Complexity Experts - adaptive processing blocks with size-varying computational units. Our framework dynamically allocates model capacity using a spring-inspired force mechanism that continuously guides routing decisions toward simpler experts when possible, with the force proportional to the complexity of the input degradation. While initially designed for computational efficiency, this approach naturally emerges as a task-discriminative learning framework, assigning degradations to the most suitable experts. This makes it particularly effective for all-in-one restoration methods, where both task-specific processing and cross-degradation knowledge sharing are crucial." + ], + "image_footnote": [], + "bbox": [ + 709, + 407, + 905, + 500 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "lems [6, 8, 16, 28, 52, 65]. Recent all-in-onerestoration models [9, 12, 13, 24, 27, 36, 49, 66], however, demonstrate the possibility of handling multiple degradation types within a single model, offering more practical solutions compared to traditional task-specific approaches without the extensive re-training need for each new degradation type. Notable works employ visual [25, 36, 50] or language-based prompt-", + "bbox": [ + 511, + 794, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author", + "bbox": [ + 109, + 887, + 235, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "12753", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ing [1, 9, 30], contrastive learning [24, 66], and diffusion-based models [1, 30].", + "bbox": [ + 89, + 90, + 485, + 121 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Despite their success, we observe that the aforementioned models often suffer from inefficiencies, as parameters tied to specific degradations often remain inactive or underutilized when addressing unrelated tasks [57]. This naturally suggests the use of mixture-of-experts (MoE) architectures for task-specific processing. Yet, current MoE-based approaches [29, 57, 62] typically incorporate routing mechanisms based on language [57] or degradation priors [29, 62], leading to imbalanced optimization, where some experts generalize well, while others struggle with their intended tasks. This inconsistency limits the potential computational benefits of bypassing task-irrelevant experts during inference.", + "bbox": [ + 88, + 123, + 485, + 305 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We attribute this undesired behavior to two limitations in current approaches. First, the uniform and rigid architecture of existing MoE models fails to account for the varying complexity requirements across different restoration tasks. For instance, motion blur demands localized processing with strong spatial awareness, while haze removal requires broader contextual understanding [4, 35], necessitating adaptive processing aligned to the task requirements. Second, the challenge of appropriately routing tasks to experts is complicated by the unknown complexity of each degradation type a priori. Typically, MoE models [37, 39, 43, 70] aim to balance expert utilization, preventing the model from collapsing into a single-expert dependency.", + "bbox": [ + 89, + 306, + 485, + 503 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we introduce a novel mixture-of-complexity-experts (MoCE) framework for all-in-one image restoration that directly addresses these limitations. Our key innovation lies in designing expert blocks with increasing computational complexity and receptive fields, allowing the model to adaptively match processing capacity with task requirements.", + "bbox": [ + 89, + 506, + 483, + 597 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address the routing challenge, we introduce a complexity-aware allocation mechanism that preferentially directs tasks to lower-complexity experts. This approach implements a complexity-proportional bias term, analogous to a mechanical spring force, to guide expert selection. Surprisingly, this straightforward approach results in effective task-specific allocation, naturally directing inputs to experts with appropriate complexity levels. Our MoCE framework offers two significant advantages: it ensures efficient inference by selectively bypassing irrelevant experts, thereby reducing computational overhead, while simultaneously maintaining high-quality restoration performance across diverse tasks. Through extensive experiments across multiple image restoration tasks, we demonstrate that our approach not only preserves restoration quality but surpasses recent state-of-the-art methods. These results validate both the effectiveness of our complexity-based expert design and the practicality of our routing strategy for real-world applications. Our approach establishes a new benchmark for all-in-one image restoration, enhancing both efficiency and fidelity.", + "bbox": [ + 89, + 599, + 485, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The principal contributions of this work are threefold:", + "bbox": [ + 532, + 90, + 888, + 106 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose MoCE-IR, achieving SoTA all-in-one image restoration performance with improved efficiency compared to prior works.", + "- Our novel MoE layer selectively activates complexity experts based on input requirements, unifying task-specific and holistic learning in a single architecture.", + "- We develop a complexity-aware routing mechanism that balances restoration quality with computational efficiency by adaptive expert allocation." + ], + "bbox": [ + 513, + 108, + 906, + 244 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 513, + 263, + 663, + 279 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Task Specific Image Restoration. Reconstructing the clean image from its degraded counterpart is a highly ill-posed problem, however, a great body of work have addressed image restoration from a data-driven learning perspective, achieving tremendous results compared to prior hand-crafted methods [22, 26, 47, 52, 65, 67, 68]. Most proposed solutions build on convolutional [6, 47, 67, 68] or Transformer-based architectures [8, 28, 52, 65] addressing single degradation tasks, such as denoising [8, 67, 68], dehazing [41, 42, 55] or deraining [18, 40]. Contrary to CNN-based networks, Transformer offer strong modeling capabilities for capturing global dependencies, which makes them outstanding image restorers [28, 65, 69]. Self-attention's quadratic complexity w.r.t the image size poses a challenge for resource-constrained applications. Conversely, convolutions offer fast and efficient processing with limited global context, but they scale more effectively with increasing input size. Recently, FFTformer [21] introduced an efficient approximation of query and key interactions in the Fourier domain, significantly reducing computational demands compared to window-based attention methods for image deblurring. Building on this, we incorporate specialized experts into a Transformer-based architecture [65] to efficiently capture both shared and distinct contextual information, addressing the diverse demands of image restoration tasks.", + "bbox": [ + 511, + 291, + 908, + 667 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "All-in-One Image Restoration. Restoring degraded images often requires multiple models, complicating practical implementation, especially when images exhibit a combination of various degradations. An emerging field known as all-in-one image restoration is advancing in low-level computer vision, utilizing a single deep blind restoration model to tackle multiple degradation types simultaneously [7, 19, 36, 49, 64, 66]. The seminal work, AirNet [24] achieves blind all-in-one image restoration using contrastive learning to extract degradation representations, which guide the restoration process. Similarly, IDR [66] employs a meta-learning-based two-stage approach to model degradation through underlying physical principles. Prompt-based learning [25, 36, 50] has also gained traction, with [36] introducing tunable prompts to encode degradation-specific in", + "bbox": [ + 511, + 674, + 910, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "12754", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/d89bfebbc3610f17c2faf0b54f62894d3210c7910aa4df55795ffce9cf632e7e.jpg", + "image_caption": [ + "(a) Architecture overview" + ], + "image_footnote": [], + "bbox": [ + 98, + 88, + 361, + 306 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/f3471c152ca5a81a0179dd375026f12638ffa2b774dd409c7c7378e6a64d89b0.jpg", + "image_caption": [ + "(b) Transformer blocks." + ], + "image_footnote": [], + "bbox": [ + 398, + 90, + 599, + 299 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/a849b69f4a55e7e65565adb2930570db71fcfea3615872b33bde60103a703f26.jpg", + "image_caption": [ + "(c) Mixture-of-complexity-experts", + "Figure 2. Proposed MoCE-IR framework. Despite recent advances in MoE-based image restoration [57, 62, 63], inconsistent expert behavior—where some experts over-generalize while others underperform—limits their computational efficiency. We address this through complexity experts: flexible blocks with varying computational capacity and receptive fields. Our MoCE-IR employs an asymmetric encoder-decoder architecture where each decoder block contains a mixture-of-complexity-experts layer for adaptive capacity routing." + ], + "image_footnote": [], + "bbox": [ + 633, + 90, + 903, + 301 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "formation, albeit with a high parameter count. With recent advancements in multimodal integration, such as combining images and language, there has been growing interest in language-guided image restoration. Notable approaches include DA-CLIP [30], InstructIR [9], and UniProcessor [12]. However, these methods often depend on high-quality text prompts and typically require LLMs, limiting their applicability on resource-constrained devices. In contrast, our contribution lies in a novel, parameter-efficient approach that focuses on selective activation, effectively addressing diverse degradations while remaining lightweight and adaptable.", + "bbox": [ + 88, + 396, + 483, + 561 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Dynamic Networks. Dynamic networks have evolved from basic conditional computation [3] to sophisticated Mixture-of-Experts (MoE) architectures that expand model capacity while maintaining efficient inference costs [37, 43, 46]. This approach has proven particularly effective in transformer-based NLP and high-level vision tasks. In image restoration, Path-Restore [60] introduced content-aware patch routing with difficulty-regulated rewards, while recent all-in-one models [29, 57, 62] leverage various priors for expert routing. Our work addresses the challenge of adaptively routing corrupted images to appropriate computational units within a generalist restoration framework. By dynamically aligning image characteristics with computational requirements, we achieve a unified model in both parameter and architectural space, advancing beyond previous limitations.", + "bbox": [ + 89, + 568, + 483, + 792 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 813, + 181, + 828 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we outline the core principles of our All-in-One image restoration method. As shown in Fig. 2a, our pipeline features a U-shaped architecture [44, 65] with an asymmetric encoder-decoder design. A $3 \\times 3$ convolution", + "bbox": [ + 89, + 839, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "first extracts shallow features from the degraded input, which then pass through 4 levels of encoding and decoding stages. While both stages utilize Transformer blocks [36, 65], our novel MoCE layers are exclusively integrated into the decoder blocks (Fig. 2b). To further enhance the decoder's feature enrichment capabilities, we introduce high-frequency guidance via Sobel-filtered global feature vector to promote the frequency awareness of the gating functions. Finally, a global residual connection links the shallow features to the output of the refinement stage, refining the crucial high-frequency details before producing the restored image.", + "bbox": [ + 511, + 396, + 906, + 561 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Mixture-of-Complexity-Experts", + "text_level": 1, + "bbox": [ + 511, + 571, + 794, + 588 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To overcome the limitations of uniform MoE architectures in image restoration, we propose an efficient input-adaptive model for all-in-one image restoration that dynamically adjusts processing capacity to match task requirements.", + "bbox": [ + 511, + 594, + 906, + 654 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Technically, at the core of our framework is a specialized MoCE layer, consisting of $n$ complexity experts $\\mathbf{E}$ and a single shared expert $\\mathbf{S}$ . As shown in Fig. 2b, our experts are designed with increasing computational complexity and progressively larger receptive fields, in contrast to the uniform architecture of traditional MoE designs [37, 43, 46]. The interaction between these components is orchestrated through a two-level gating mechanism: first projecting routed tokens into expert-specific embeddings, then combining specialized and agnostic features through modulation, and finally merging the outputs via cross-attention. This enables our model to capture both degradation-specific features and interdegradation relationships, while maintaining a bias toward computationally efficient processing paths.", + "bbox": [ + 511, + 655, + 908, + 867 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Expert Design. Each complexity expert block as shown in Fig. 2c is designed based on two fundamental principles:", + "bbox": [ + 511, + 869, + 906, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "12755", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ensuring computational efficiency and capturing hierarchical spatial features, which are critical for restoring fine details and addressing diverse degradation patterns. Given the growing computational demands associated with increasing expert numbers and capacities, it is crucial to prioritize an overall efficient expert design. To this end, we implement a nested expert structure, progressively reducing channel dimensionality $r$ within each subsequent expert to control the computational overhead. Simultaneously, we increase the receptive field, i.e. window partition size $w$ , within each expert to adaptively balance the localized and global processing, tailored to the specific requirements of the input degradation. Thus, our complexity experts are constructed by scaling the design along two key axes: window partition size and channel dimension.", + "bbox": [ + 89, + 90, + 483, + 316 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "More specifically, each expert $\\mathbf{E}$ projects its input tokens to the embedding $\\mathbf{x}_E\\in \\mathbb{R}^{H\\times W\\times r}$ , where $r = C / 2^i$ for $i\\in \\{i,\\dots,n\\}$ , to highlight the most relevant information along the channel dimension. Following this, a window-based self-attention (WSA) mechanism captures spatial information effectively. To optimize performance, we employ an FFT-based approximation [21] for efficient matrix multiplication between queries $Q$ and keys $K$ in the Fourier domain. The refined features are then transformed back into the spatial domain and re-projected to their original dimensionality. Lastly, the output of the expert block $\\hat{\\mathbf{x}}_E^i$ is modulated through element-wise multiplication with the holistic features from the shared expert, S. As shared expert, we employ the transposed self-attention (T-SA) module [65] in the channel dimension, while the complexity experts operate spatially. More concretely, the modulated expert features $\\mathbf{y}_{\\mathbf{E}}^{\\mathrm{i}}$ are obtained as following:", + "bbox": [ + 89, + 318, + 483, + 575 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {y} _ {\\mathbf {E}} ^ {i} = \\hat {\\mathbf {x}} _ {E} ^ {i} \\odot \\mathbf {S} (\\mathbf {x}) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 151, + 585, + 483, + 604 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\text {w i t h} \\hat {\\mathbf {x}} _ {E} ^ {i} = \\mathbf {E} ^ {i} \\left(\\mathbf {x} _ {E} ^ {i}\\right) = \\operatorname {F F T - W S A} _ {\\mathrm {W} _ {i}} \\left(\\mathbf {W} _ {C \\rightarrow r _ {i}} ^ {i} \\mathbf {x} _ {E} ^ {i}\\right) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 119, + 607, + 483, + 626 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {a n d} \\mathbf {S} (\\mathbf {x}) = \\mathrm {T} - \\operatorname {S A} (\\mathbf {x}) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 112, + 628, + 483, + 643 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where linear layers for projection are denoted as $\\mathbf{W}$ , $\\odot$ denotes element-wise multiplication and $i$ is the index of the current complexity expert $\\mathbf{E}^i$ with window size $w_i$ and embedding dimensionality $r_i$ . For example, the most lightweight expert is characterized by the smallest embedding dimension $r_1$ and window size $w_1$ , while the dimensions and window partitions for the other experts can increase linearly. Throughout our network, the capacity of the shared expert remains constant, with $r = C$ .", + "bbox": [ + 89, + 656, + 483, + 791 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Complexity-aware Routing", + "text_level": 1, + "bbox": [ + 89, + 801, + 336, + 819 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Drawing inspiration from sparse MoE [43, 46], we integrate linear layers within each decoder block to enable a routing mechanism, associating input features $\\mathbf{x} \\in \\mathbb{R}^{H \\times W \\times C}$ with their corresponding specialized complexity experts $\\mathbf{E}$ . However, in the context of image restoration, a key challenge", + "bbox": [ + 89, + 825, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "arises in achieving a scale-invariant tokenization of input images, ensuring consistency across varying resolutions and scales. While previous MoE approaches [37, 43, 46] employ a token-based routing, we opt for an image-level routing strategy instead, where we select experts for entire input image. The routing function $g(\\mathbf{x})$ orchestrates the allocation of input samples based on the required compute to corresponding complexity experts $\\mathbf{E}^i$ , indexed by $i \\in 1, \\dots, n$ with $n$ representing the total number of experts. Within $g(\\mathbf{x})$ , we select the top- $k$ elements of the softmax distribution associating input tokens to experts, setting all other elements to zero. In practice, we use $k = 1$ while $\\epsilon$ is sampled independently $\\epsilon \\sim \\mathcal{N}(0, 1/n^2)$ entry-wise enabling noisy top-1 routing:", + "bbox": [ + 511, + 90, + 906, + 303 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ng (\\mathbf {x}) = \\operatorname {t o p} _ {k} \\left(\\operatorname {S o f t m a x} \\left(\\mathbf {W} \\mathbf {x} + \\epsilon\\right)\\right) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 596, + 316, + 906, + 334 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "During training, we incorporate an auxiliary loss $\\mathcal{L}_{\\mathrm{aux}}$ inspired by the load balancing loss proposed by Riquelme et al. [43]. We complement the importance term to favor experts based on their computational complexity, allowing for more effective balancing aligned with expert capacity. As in [43], the balanced expert utilization is encouraged through an importance loss, where the importance of each expert $\\mathbf{E}^i$ for a batch of images is defined as the sum of the routing weights associated with the $i$ -th expert across the entire batch. Additionally, we introduce a complexity bias $\\mathbf{b}$ by calculating the number of learnable parameters $p_i$ in each expert $\\mathbf{E}^i$ and normalizing these values with respect to the largest parameter count of the experts, $p_{\\mathrm{max}}$ :", + "bbox": [ + 511, + 347, + 908, + 544 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {I m p} _ {i} (\\mathbf {x}) = \\left(\\sum_ {x \\in \\mathbf {x}} \\operatorname {S o f t m a x} (W x) _ {i}\\right) * \\mathbf {b}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 555, + 906, + 597 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\text {w i t h} \\mathbf {b} = \\left[ \\frac {p _ {1}}{p _ {\\max }}, \\frac {p _ {2}}{p _ {\\max }}, \\dots , \\frac {p _ {n}}{p _ {\\max }} \\right], \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 599, + 906, + 617 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $W$ is the layer-specific weight for the router $g(\\mathbf{x})$ . This approach assigns progressively lower weights to experts with lower complexity. For intuitive understanding, this mechanism parallels a mechanical spring system, where the restoring force is proportional to both displacement and the spring constant. In our context, the expert's parameter count corresponds to displacement, while the normalization factor serves as the material-specific constant, collectively determining the magnitude of complexity bias. Without this, the routers lack a basis for selecting specific experts, leading to randomized routing and thus suboptimal model capacity utilization. Lastly, the complexity-aware importance Imp is used to compute the total auxiliary loss $\\mathcal{L}_{\\mathrm{aux}}$ defined as:", + "bbox": [ + 511, + 628, + 908, + 827 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {a u x}} (\\mathbf {x}) = ^ {1 / 2} \\operatorname {C V} \\left(\\operatorname {I m p} (\\mathbf {x})\\right) ^ {2} + ^ {1 / 2} \\operatorname {C V} \\left(\\operatorname {L o a d} (\\mathbf {x})\\right) ^ {2}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 531, + 839, + 906, + 857 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where CV denotes the coefficient of variation [43]. The effect of the auxiliary loss is shown in Fig. 5 and Tab. 5a.", + "bbox": [ + 511, + 869, + 906, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "12756", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/f5fb81bcffd931c7467eae82bb00a904fa500ddd9f2d38c8be58ee9fac0d79ff.jpg", + "table_caption": [ + "Table 1. Comparison to state-of-the-art on three degradations. PSNR (dB, ↑) and SSIM (↑) are reported on the full RGB images. Our method sets a new state-of-the-art on average across all benchmarks on two model scales, showcasing its scalability while being more efficient than prior work. ‘-’ represents unreported results and ‘*’ indicates VLM-based approaches. The best performances are highlighted." + ], + "table_footnote": [], + "table_body": "
MethodParams.DehazingDerainingDenoisingAverage
SOTSRain100LBSD68σ=15BSD68σ=25BSD68σ=50
LightBRDNet [48]-23.23.89527.42.89532.26.89829.76.83626.34.69327.80.843
LPNet [14]-20.84.82824.88.78426.47.77824.77.74821.26.55223.64.738
FDGAN [11]-24.71.92929.89.93330.25.91028.81.86826.43.77628.02.883
DL [13]2M26.92.93132.62.93133.05.91430.41.86126.90.74029.98.876
AirNet [24]9M27.94.96234.90.96733.92.93331.26.88828.00.79731.20.910
MoCE-IR-S (ours)11M30.94.97938.22.98334.08.93331.42.88828.16.79832.57.916
HeavyMPRNet [64]16M25.28.95533.57.95433.54.92730.89.88027.56.77930.17.899
PromptIR [36]36M30.58.97436.37.97233.98.93331.31.88828.06.79932.06.913
Gridformer [51]34M30.37.97037.15.97233.93.93131.37.88728.11.80132.19.912
Art-PromptIR [54]33M30.83.97937.94.98234.06.93431.42.89128.14.80132.49.917
DA-CLIP* [30]125M29.46.96336.28.96830.02.82124.86.58522.29.476--
UniProcessor* [12]1002M31.66.97938.17.98234.08.93531.42.89128.17.80332.70.918
MoCE-IR (ours)25M31.34.97938.57.98434.11.93231.45.88828.18.80032.73.917
", + "bbox": [ + 91, + 133, + 903, + 311 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/8927f4c4bcbe4df6683b717c7148e9f56291212ac166a7250533e535cbb44c7e.jpg", + "table_caption": [ + "Table 2. Comparison to state-of-the-art on five degradations. PSNR (dB, ↑) and SSIM (↑) are reported on the full RGB images with (*) denoting general image restorers, others are specialized all-in-one approaches. Under more challenging degradations, our MoCE-IR framework outperforms prior work across two model scales. The best performances are highlighted." + ], + "table_footnote": [], + "table_body": "
MethodParams.DehazingDerainingDenoisingDeblurringLow-LightAverage
SOTSRain100LBSD68σ=25GoProLOLv1
LightSwinIR* [26]1M21.50.89130.78.92330.59.86824.52.77317.81.72325.04.835
DL [13]2M20.54.82621.96.76223.09.74519.86.67219.83.71221.05.743
TAPE [27]1M22.16.86129.67.90430.18.85524.47.76318.97.62125.09.801
AirNet [24]9M21.04.88432.98.95130.91.88224.35.78118.18.73525.49.847
MoCE-IR-S (ours)11M31.33.97837.21.97831.25.88428.90.87721.68.85130.08.913
HeavyNAFNet* [6]17M25.23.93935.56.96731.02.88326.53.80820.49.80927.76.881
DGUNet* [33]17M24.78.94036.62.97131.10.88327.25.83721.87.82328.32.891
Restormer* [65]26M24.09.92734.81.96231.49.88427.22.82920.41.80627.60.881
MambaIR [16]27M25.81.94436.55.97131.41.88428.61.87522.49.83228.97.901
Transweather [49]38M21.32.88529.43.90529.00.84125.12.75721.21.79225.22.836
IDR [66]15M25.24.94335.63.96531.60.88727.87.84621.34.82628.34.893
Gridformer [51]34M26.79.95136.61.97131.45.88529.22.88422.59.83129.33.904
InstructIR-5D [9]17M27.10.95636.84.97331.40.87329.40.88623.00.83629.55.908
MoCE-IR (ours)25M30.48.97438.04.98231.34.88730.05.89923.00.85230.58.919
", + "bbox": [ + 91, + 358, + 903, + 547 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 561, + 223, + 579 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We conduct experiments by strictly following previous works in general image restoration [36, 66] under two different settings: (i) All-in-One and (ii) Composited degradations. In the all-in-one setting, a unified model is trained across multiple degradation types, where we consider three and five distinct degradations. In the Composited degradations setting, a single model is trained to handle both isolated degradation occurrences as well as combinations of up to three different types of degradations.", + "bbox": [ + 86, + 590, + 482, + 728 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details. Our MoCE-IR framework is end-to-end trainable requiring no multi-stage optimization of any components. The architecture features an asymmetrical, four-level encoder-decoder design. The encoder consists of four levels with varying numbers of transformer blocks, arranged as $[4,6,6,8]$ from the top level to the lowest. The decoder includes three levels, containing $[2,4,4]$ transformer blocks, respectively. In each MoCE layer, we use $n = 4$ nested experts, where the dimensionality of the expert embedding is given by $R = C / 2^i$ for $i\\in \\{i,\\dots,n\\}$ . The network's initial embedding dimension is set to $C = 32$ , doubling at each", + "bbox": [ + 88, + 734, + 485, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "subsequent level while the spatial resolution is halved. We follow the training configuration of prior work [36], and train our models for 120 epochs with a batch size of 32. We optimize the $L_{1}$ loss in the RGB and Fourier domain using the Adam [20] optimizer $(\\beta_{1} = 0.9, \\beta_{2} = 0.999)$ with an initial learning rate of $2 \\times 10^{-4}$ and cosine decay. During training, we utilize crops of size $128^{2}$ with horizontal and vertical flips as augmentations.", + "bbox": [ + 511, + 564, + 906, + 684 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. For all-in-one, we follow existing work [24, 36] and include following datasets: For image denoising, we combine the BSD400 [2] and WED [31] datasets, adding Gaussian noise at levels $\\sigma \\in [15, 25, 50]$ to create noisy images. Testing is conducted on the BSD68 [32]. For deraining, we use Rain100L [56]. The dehazing task utilises the SOTS [23] dataset. For deblurring and low-light enhancement, we employ the GoPro [34] and the LOL-v1 [53] dataset, respectively. To develop a unified model for all tasks, we merge these datasets in a three (AIO-3) or five (AIO-5) degradation setting, and train for 120 epochs and directly evaluate them across different tasks. For the composite degradation setting, we use the CDD11 dataset [17] and train our method for 200 epochs, keeping all other settings unchanged.", + "bbox": [ + 511, + 688, + 908, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "12757", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/26e1aa0b976f3626d752e08df45eba2f6b226148c023cabcbba4f3dc9a923da6.jpg", + "table_caption": [ + "Table 3. Comparison to state-of-the-art on composited degradations. PSNR (dB, ↑) and SSIM (↑) are reported on the full RGB images with (*) denoting general image restorers, others are specialized all-in-one approaches. Our MoCE-IR method consistently outperforms even larger models, with favorable results in composited degradation scenarios." + ], + "table_footnote": [], + "table_body": "
MethodParams.CDD11-SingleCDD11-DoubleCDD11-TripleAvg.
Low (L)Haze (H)Rain (R)Snow (S)L+HL+RL+SH+RH+SL+H+RL+H+S
AirNet [24]9M24.83.77824.21.95126.55.89126.79.91923.23.77922.82.71023.29.72322.21.86823.29.90121.80.70822.24.72523.75.814
PromptIR [36]36M26.32.80526.10.96931.56.94631.53.96024.49.78925.05.77124.51.76124.54.92423.70.92523.74.75223.33.74725.90.850
WGWSNet [71]26M24.39.77427.90.98233.15.96434.43.97324.27.80025.06.77224.60.76527.23.95527.65.96023.90.77223.97.77126.96.863
WeatherDiff [72]83M23.58.76321.99.90424.85.88524.80.88821.83.75622.69.73022.12.70721.25.86821.99.86821.23.71621.04.69822.49.799
OneRestore [17]6M26.48.82632.52.99033.40.96434.31.97325.79.82225.58.79925.19.78929.99.95730.21.96424.78.78824.90.79128.47.878
MoCE-IR-S (ours)11M27.26.82432.66.99034.31.97035.91.98026.24.81726.25.80026.04.79329.93.96430.19.97025.41.78925.39.79029.05.881
DehazingInputAirNetPromptIROursGT
Deraining
Denoising
Input CropsError Maps (darker means better)GT Crops
", + "bbox": [ + 89, + 132, + 906, + 628 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Figure 3. Visual results. We compare MoCE-IR-S to AirNet [24], and PromptIR [36] in the all-in-one setting with three degradations. MoCE-IR-S effectively removes haze and rain streaks while preserving image sharpness, achieving high-quality restoration. An error heatmap is provided, with color transitioning from black to white to indicate increasing pixel-wise error.", + "bbox": [ + 89, + 633, + 906, + 678 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Comparison to State-of-the-Art Methods", + "text_level": 1, + "bbox": [ + 89, + 684, + 439, + 700 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "All-in-One: Three Degradations. We compare our all-in-one restoration approach with specialized restoration methods, including both image-only models, such as MPR-Net [64], AirNet [24], PromptIR [36], and Gridformer [51], and vision-language models like DA-CLIP [30], InstructIR [9], and UniProcessor [12]. All methods were trained simultaneously on three types of degradation: haze, rain, and Gaussian noise. The results in Tab. 1 show that our proposed model, MoCE-IR, consistently outperforms previous works, both in light and heavy model scales. In the light category, MoCE-IR achieves the highest scores across all benchmarks with an average improvement in PSNR of", + "bbox": [ + 88, + 719, + 485, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "1.37 dB, surpassing other models like AirNet [24]. In the heavy scale, MoCE-IR competes closely with UniProcessor [12], achieving an average PSNR of 32.73 dB with $98\\%$ fewer parameters, validating our efficiency and effectiveness.", + "bbox": [ + 511, + 685, + 906, + 746 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "All-in-One: Five Degradations. Based upon [24, 66], we expand the three-degradation setting to include deblurring and low-light image enhancement, further validating our method's effectiveness. As shown in Tab. 2, our approach excels by learning specialized experts with input-adaptive complexity and maximizing synergies across degradations. Our framework outperforms previous methods on the lightest model scale, with an average PSNR improvement of 4.31 dB over AirNet [24]. Even compared to larger networks, our lightest model surpasses the VLM-based InstructIR [9] by", + "bbox": [ + 511, + 750, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "12758", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4054f08cddc89f8dc03a0e68ef767cca245144c4b5b439a6ef161def7dbc8169.jpg", + "image_caption": [ + "Figure 4. Complexity-efficiency tradeoff. Visualization of PSNR and parameter counts of proposed method compared to prior work. Proposed MoCE-IR surpasses prior methods, achieving SoTA results in all-in-one image restoration with enhanced efficiency." + ], + "image_footnote": [], + "bbox": [ + 96, + 90, + 480, + 220 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/be61f764f8f35449e2e76073f400a4a8868f18fd70cc69e520321ed5337e051b.jpg", + "table_caption": [ + "Table 4. Computational demands. GFLOPS are computed on an input image of size $224 \\times 224$ using a NVIDIA RTX 4090 GPU." + ], + "table_footnote": [], + "table_body": "
MethodParams.MemoryFLOPSRuntime
AirNet [24]8.93M4829M238G42.17 ± 0.23
PromptIR [36]35.59M9830M132G41.28 ± 0.43
IDR [66]15.34M4905M98G-
MoCE-IR (ours)25.35M5887M80.59 ± 5.21G23.36 ± 2.47
MoCE-IR-S (ours)11.47M4228M36.93 ± 2.32G22.15 ± 2.59
", + "bbox": [ + 91, + 316, + 483, + 393 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "0.33 dB on average, and on a comparable model scale, our framework extends this advantage by an additional 0.50 dB.", + "bbox": [ + 89, + 414, + 483, + 445 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "All-in-One: Composed Degradations. To simulate more realistic degradation scenarios, recent work [17] has trained all-in-one restorers not only on a mix of different degradations but also on more challenging cases where multiple degradation types are combined within a single image. We extend the previous AIO settings by including rain, haze, snow, and low illumination as individual degradation types, along with various composite degradation scenarios, resulting in a total of eleven distinct restoration settings. As shown in Tab. 3, our method outperforms previous all-in-one restorers by an average improvement of $0.64\\mathrm{dB}$ over OneRestore [17], while being among the smallest models in size. This further validates our effectiveness with adaptive modeling capacity based on input complexity.", + "bbox": [ + 88, + 453, + 485, + 667 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Visual Results. We show visual results obtained under the three degradation settings in Fig. 3. In certain demanding hazy scenarios, both AirNet [24] and PromptIR [36] reveal constraints in completely eradicating haziness, resulting in noticeable color intensity discrepancies, whereas our approach ensures precise color reconstruction. Furthermore, in challenging rainy scenes, these popular approaches continue to exhibit notable remnants of rain, which our method adeptly eliminates, distinguishing itself from other approaches. Additionally, our method also generates clear and sharp denoised outputs. We further provide the error maps for visual differences where our method contains less noticeable errors. Coupled with the quantitative comparisons, these findings underscore the effectiveness of our method. Please refer to the supplementary for more visuals.", + "bbox": [ + 89, + 674, + 485, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 5. Architecture analysis. (a) We investigate scaling the expert embedding $r$ and the bias normalization for MoCE-IR-S. (b) We compare MoCE to several baselines [29, 43, 58], demonstrating that our complexity experts consistently outperform them. Notably, complexity-biased routing further boosts performance, surpassing standard load balancing (LB) [43]. PSNR (dB, ↑) and SSIM (↑) are reported on full RGB images.", + "bbox": [ + 511, + 88, + 906, + 186 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/f0aa766f6b63e5170f535a8b98b9f0ba537e0b168935c8cdcbf8caabd7a7f6e4.jpg", + "table_caption": [ + "(a) Ablation on the expert scaling." + ], + "table_footnote": [], + "table_body": "
Embedding r scalingBias normalizationAvg. AIO-3
NestedpMin32.44.913
ExponentialpMax32.47.914
NestedpMax32.57.916
(b) Ablation on the complexity-aware routing.
MethodSOTSRain100LBSD6815-50Avg.
MoCE + Complexity Bias (ours)30.94.97938.22.98331.22.873
Baseline (MoE + LB [45])30.28.97737.77.98131.14.871
w/ MoCE + LB [45]30.13.97838.24.98331.19.872
w/ AMIR [58]29.97.97537.90.98131.17.872
w/ WM-MoE [29]30.10.97636.72.97631.05.869
", + "bbox": [ + 514, + 202, + 903, + 358 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Ablation Studies", + "text_level": 1, + "bbox": [ + 511, + 368, + 679, + 383 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We conduct detailed studies on the components within our method, highlighting the emergence of task-discriminative learners. All experiments are conducted in the all-in-one setting with three degradations. Please refer to the supplementary for additional ablations.", + "bbox": [ + 511, + 391, + 908, + 467 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "**Complexity-Efficiency Trade-off.** Our MoCE-IR framework addresses the critical challenge of scalability in vision tasks by optimizing both reconstruction quality and computational efficiency. We visualize in Fig. 4 the progressive improvements in reconstruction fidelity of our framework on two complexity scales. Furthermore, Tab. 4 validates that our approach significantly outperforms SoTA methods in terms of memory utilization. Unlike traditional MoE architectures with fixed computational demands, our complexity experts enable dynamic FLOPS allocation. For instance, our lightweight model averages 36.9 GFLOPS, while the heavyweight model averages 80.59 GFLOPS. This adaptive computation, combined with lower memory consumption and parameter counts, gives MoCE-IR a substantial advantage over contemporary methods.", + "bbox": [ + 511, + 472, + 908, + 699 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Expert Scaling. As shown in Tab. 5a, we systematically evaluate different scaling strategies regarding the embedding dimensionality $r$ of each complexity expert. Our investigation compares exponential and nested parameter scaling, where the window partition is always fixed and follows a $2^{(2 + i)}$ progression. Further, we investigate the bias normalization, whether normalizing to the lowest or highest expert capacity. Overall, the nested parameter scaling emerges as the optimal choice, providing sufficient expert channel diversity while maintaining reasonable computational costs. Normalizing router weights by the highest expert capacity, rather than the minimum, proves crucial for performance as minimum capacity normalization forces router weights", + "bbox": [ + 511, + 704, + 910, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "12759", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/de8b9b36e22e38c7ca1958a064c3aee7684308a39347bd4dbc5d731bad47f8db.jpg", + "image_caption": [ + "Figure 5. Routing visualization for the AIO-3 setting. (a) While load balancing [43] ensures uniform expert utilization, it neglects shared task dependencies and task-specific characteristics, limiting restoration quality. (b)-(d) Complexity-aware routing fosters task discrimination by directing complex degradations to experts with broader contextual understanding and vice versa. This allows some experts to generalize across tasks while others specialize, enhancing adaptability. We visualize the average decisions made by each router for dehazing, deraining and denoising. The y-axis indicates increasing expert complexity." + ], + "image_footnote": [], + "bbox": [ + 106, + 90, + 313, + 200 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5625323f40b572039dc3c40b879f2066643f57312f94319dfb78d6e15058b691.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 318, + 90, + 504, + 200 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6183b41c6ff9a9faa9bac3b25e2d0a586dfd507f0e5491de087fe9d94342db88.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 509, + 90, + 699, + 200 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/de6dd3673d508523700cdc658412cf2af71d9236496cd0646803af38d99699e6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 704, + 90, + 892, + 200 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/a578ba8e6e69589f5c08684298fe5ad42fe8233ec52e5b28b3099ccca554bb05.jpg", + "table_caption": [ + "Table 6. Expert generalization. We report PSNR for MoCE-IR-S with expert assignments fixed post-training. $\\mathcal{E}$ denotes expert." + ], + "table_footnote": [], + "table_body": "
DegradationRoutingLearned ChoiceManual Choice
\\( \\mathcal{E}_1 \\)\\( \\mathcal{E}_2 \\)\\( \\mathcal{E}_3 \\)\\( \\mathcal{E}_4 \\)
Rain NoiseLoad Balance [45]Not Applicable28.2733.3328.2433.3728.2533.3728.2733.40
Rain NoiseComplexity Bias (ours)\\( \\mathcal{E}_1 \\)30.4530.2129.7324.65
\\( \\mathcal{E}_4 \\)33.9233.9333.9334.00
", + "bbox": [ + 91, + 311, + 483, + 393 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "of higher capacity experts to escalate uncontrollably, thus excluding them from the gating mechanism.", + "bbox": [ + 89, + 405, + 483, + 436 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Impact of Complexity-aware Routing. Building on the findings from Tab. 5a, we adopt nested expert scaling and evaluate our complexity-aware routing strategy against conventional load balancing [43], which enforces uniform expert usage. As shown in Tab. 5b, our MoCE-IR framework achieves superior performance by preferentially routing to lower-complexity experts, activating higher-complexity experts only when their added capacity provides clear benefits. This efficient model utilization not only boosts image restoration performance, consistently surpassing related adaptive restoration methods [29, 58], but also enhances out-of-distribution generalization as MoCE-IR effectively handles degradations that with significant deviation from the training data; see the supplementary for more evaluations.", + "bbox": [ + 88, + 440, + 483, + 652 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Routing Analysis. Fig. 5 illustrates the average expert selection per layer for different degradations: haze, rain, and noise. Higher-complexity experts are favored for tasks requiring broader context, like haze and noise, while lighter experts are chosen for rain streak removal. Notably, some experts handle multiple tasks, while others remain task-specific. Our complexity-conditioned routing promotes task-discriminative learning by directing inputs to relevant experts, preventing the collapse into a single-expert dependency, ensuring efficient and diverse operation.", + "bbox": [ + 89, + 657, + 483, + 808 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The baseline MoE [43] approach, lacking complexity awareness and scaled experts, uniformly distributes degraded samples across all experts. Our results show that this uniform allocation is suboptimal for all-in-one restoration, as it fails to leverage task-specific commonalities and distinctions effectively. Tab. 6 highlights the limitations of standard", + "bbox": [ + 89, + 810, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "MoEs [43], where fixed post-training expert assignments result in uniform, task-agnostic performance, offering only moderate cross-task generalization. In contrast, MoCEs effectively generalize across degradations while exhibiting task-discriminative behavior, routing inputs to the most suitable experts for optimal performance.", + "bbox": [ + 511, + 282, + 906, + 375 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion, Limitations and Discussion", + "text_level": 1, + "bbox": [ + 511, + 388, + 867, + 404 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper presents MoCE-IR, an efficient all-in-one image restoration model leveraging complexity experts, specialized units with varying computational demands and receptive fields. By aligning degradation tasks with suitable resources, MoCE-IR overcomes limitations of prior methods. Its built-in bias toward simpler paths enhances task-discriminative learning, enabling swift inference by bypassing irrelevant experts without sacrificing restoration quality.", + "bbox": [ + 511, + 412, + 906, + 532 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Extensive experiments across diverse degradation settings, including more demanding scenarios with multiple composited degradations, demonstrate our computational efficiency through selective expert utilization and consistent improvement over state-of-the-art methods.", + "bbox": [ + 511, + 534, + 906, + 609 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "While MoCE-IR shows substantial improvements, opportunities remain. The current image-level routing imposes scalability constraints. Future work could explore SoftMoE [37] architectures for fine-grained spatial routing or early-exiting approaches [39] to further enhance processing speeds and input adaptability. Extending to synthetic-to-real adaptation, especially for video, represents a crucial next step. Performance optimization through mixed-precision computation across experts could further enhance speed and efficiency for resource-limited deployments.", + "bbox": [ + 511, + 609, + 908, + 762 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements. The authors sincerely thank the reviewers and all members of the program committee for their tremendous efforts and incisive feedback. This research was supported by the Alexander von Humboldt Foundation, the Ministry of Education and Science of Bulgaria (support for INSAIT, under the Bulgarian National Roadmap for Research Infrastructure), the Shanghai Municipal Science and Technology Major Project (2021SHZDZX0102), and the Fundamental Research Funds for the Central Universities.", + "bbox": [ + 511, + 763, + 908, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "12760", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Yuang Ai, Huaibo Huang, Xiaogiang Zhou, Jiexiang Wang, and Ran He. Multimodal prompt perceiver: Empower adaptiveness generalizability and fidelity for all-in-one image restoration. In CVPR, 2024. 2", + "[2] Pablo Arbelaez, Michael Maire, Charless Fowlkes, and Jitendra Malik. Contour detection and hierarchical image segmentation. TPAMI, 33(5):898-916, 2010. 5", + "[3] Yoshua Bengio, Nicholas Léonard, and Aaron Courville. Estimating or propagating gradients through stochastic neurons for conditional computation. arXiv preprint arXiv:1308.3432, 2013. 3", + "[4] Dana Berman and Shai Avidan. Non-local image dehazing. In CVPR, 2016. 2", + "[5] Bolun Cai, Xiangmin Xu, Kui Jia, Chunmei Qing, and Dacheng Tao. Dehazenet: An end-to-end system for single image haze removal. TIP, 25(11):5187-5198, 2016. 1", + "[6] Liangyu Chen, Xiaojie Chu, Xiangyu Zhang, and Jian Sun. Simple baselines for image restoration. In ECCV, 2022. 1, 2, 5", + "[7] Yu-Wei Chen and Soo-Chang Pei. Always clear days: Degradation type and severity aware all-in-one adverse weather removal. arXiv preprint arXiv:2310.18293, 2023. 1, 2", + "[8] Zheng Chen, Yulun Zhang, Jinjin Gu, Linghe Kong, Xin Yuan, et al. Cross aggregation transformer for image restoration. NeurIPS, 2022. 1, 2", + "[9] Marcos V Conde, Gregor Geigle, and Radu Timofte. Instructir: High-quality image restoration following human instructions. In ECCV, 2024. 1, 2, 3, 5, 6", + "[10] Xin Dang, Hai Wang, Jie Ren, and Le Chen. An application performance optimization model of mobile augmented reality based on hd restoration. In CBD, pages 201-206. IEEE, 2020. 1", + "[11] Yu Dong, Yihao Liu, He Zhang, Shifeng Chen, and Yu Qiao. Fd-gan: Generative adversarial networks with fusion-discriminator for single image dehazing. In AAAI, 2020. 5", + "[12] Huiyu Duan, Xiongkuo Min, Sijing Wu, Wei Shen, and Guangtao Zhai. Uniprocessor: A text-induced unified low-level image processor. In ECCV, 2024. 1, 3, 5, 6", + "[13] Qingnan Fan, Dongdong Chen, Lu Yuan, Gang Hua, Nenghai Yu, and Baoquan Chen. A general decoupled learning framework for parameterized image operators. TPAMI, 43(1): 33-47, 2019. 1, 5", + "[14] Hongyun Gao, Xin Tao, Xiaoyong Shen, and Jiaya Jia. Dynamic scene deblurring with parameter selective sharing and nested skip connections. In CVPR, 2019. 5", + "[15] Florin Girbacia, Silviu Butnariu, Alex Petre Orman, and Cristian Cezar Postelnicu. Virtual restoration of deteriorated religious heritage objects using augmented reality technologies. European Journal of Science and Theology, 9(2):223-231, 2013. 1", + "[16] Hang Guo, Jinmin Li, Tao Dai, Zhihao Ouyang, Xudong Ren, and Shu-Tao Xia. Mambair: A simple baseline for image restoration with state-space model. In ECCV, 2024. 1, 5", + "[17] Yu Guo, Yuan Gao, Yuxu Lu, Ryan Wen Liu, and Shengfeng He. Onerestore: A universal restoration framework for composite degradation. In ECCV, 2024. 5, 6, 7" + ], + "bbox": [ + 93, + 114, + 485, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[18] Kui Jiang, Zhongyuan Wang, Peng Yi, Chen Chen, Baojin Huang, Yimin Luo, Jiayi Ma, and Junjun Jiang. Multi-scale progressive fusion network for single image deraining. In CVPR, 2020. 2", + "[19] Yitong Jiang, Zhaoyang Zhang, Tianfan Xue, and Jinwei Gu. Autodir: Automatic all-in-one image restoration with latent diffusion. arXiv preprint arXiv:2310.10123, 2023. 2", + "[20] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv, 2017. 5", + "[21] Lingshun Kong, Jiangxin Dong, Jianjun Ge, Mingqiang Li, and Jinshan Pan. Efficient frequency domain-based transformers for high-quality image deblurring. In CVPR, 2023. 2, 4", + "[22] Jaakko Lehtinen, Jacob Munkberg, Jon Hasselgren, Samuli Laine, Tero Karras, Miika Aittala, and Timo Aila. Noise2noise: Learning image restoration without clean data. arXiv preprint arXiv:1803.04189, 2018. 1, 2", + "[23] Boyi Li, Wenqi Ren, Dengpan Fu, Dacheng Tao, Dan Feng, Wenjun Zeng, and Zhangyang Wang. Benchmarking single-image dehazing and beyond. TIP, 28(1):492-505, 2018. 5", + "[24] Boyun Li, Xiao Liu, Peng Hu, Zhongqin Wu, Jiancheng Lv, and Xi Peng. All-In-One Image Restoration for Unknown Corruption. In CVPR, 2022. 1, 2, 5, 6, 7", + "[25] Zilong Li, Yiming Lei, Chenglong Ma, Junping Zhang, and Hongming Shan. Prompt-in-prompt learning for universal image restoration. arXiv preprint arXiv:2312.05038, 2023. 1, 2", + "[26] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Swinir: Image restoration using swim transformer. In ICCV, 2021. 2, 5", + "[27] Lin Liu, Lingxi Xie, Xiaopeng Zhang, Shanxin Yuan, Xiangyu Chen, Wengang Zhou, Houqiang Li, and Qi Tian. Tape: Task-agnostic prior embedding for image restoration. In ECCV, 2022. 1, 5", + "[28] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, 2021. 1, 2", + "[29] Yulin Luo, Rui Zhao, Xiaobao Wei, Jinwei Chen, Yijie Lu, Shenghao Xie, Tianyu Wang, Ruiqin Xiong, Ming Lu, and Shanghang Zhang. Wm-moe: Weather-aware multi-scale mixture-of-experts for blind adverse weather removal. arXiv preprint arXiv:2303.13739, 2023. 2, 3, 7, 8", + "[30] Ziwei Luo, Fredrik K Gustafsson, Zheng Zhao, Jens Sjolund, and Thomas B Schon. Controlling vision-language models for multi-task image restoration. In ICLR, 2024. 2, 3, 5, 6", + "[31] Kede Ma, Zhengfang Duanmu, Qingbo Wu, Zhou Wang, Hongwei Yong, Hongliang Li, and Lei Zhang. Waterloo exploration database: New challenges for image quality assessment models. TIP, 26(2):1004-1016, 2016. 5", + "[32] David Martin, Charless Fowlkes, Doron Tal, and Jitendra Malik. A database of human segmented natural images and its application to evaluating segmentation algorithms and measuring ecological statistics. In ICCV, 2001. 5", + "[33] Chong Mou, Qian Wang, and Jian Zhang. Deep generalized unfolding networks for image restoration. In CVPR, 2022. 5" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "12761", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[34] Seungjun Nah, Tae Hyun Kim, and Kyoung Mu Lee. Deep multi-scale convolutional neural network for dynamic scene deblurring. In CVPR, 2017. 5", + "[35] Jinshan Pan, Deqing Sun, Hans Pfister, and Ming-Hsuan Yang. Blind image deblurring using dark channel prior. In CVPR, 2016. 2", + "[36] Vaishnav Potlapalli, Syed Waqas Zamir, Salman Khan, and Fahad Khan. Prompt: Prompting for all-in-one image restoration. In NeurIPS, 2023. 1, 2, 3, 5, 6, 7", + "[37] Joan Puigcerver, Carlos Riquelme Ruiz, Basil Mustafa, and Neil Houlsby. From sparse to soft mixtures of experts. In ICLR, 2024. 2, 3, 4, 8", + "[38] Yanyun Qu, Yizi Chen, Jingying Huang, and Yuan Xie. Enhanced pix2pix dehazing network. In CVPR, 2019. 1", + "[39] David Raposo, Sam Ritter, Blake Richards, Timothy Lillicrap, Peter Conway Humphreys, and Adam Santoro. Mixture-of-depths: Dynamically allocating compute in transformer-based language models, 2024. 2, 8", + "[40] Dongwei Ren, Wangmeng Zuo, Qinghua Hu, Pengfei Zhu, and Deyu Meng. Progressive image deraining networks: A better and simpler baseline. In CVPR, 2019. 2", + "[41] Wenqi Ren, Lin Ma, Jiawei Zhang, Jinshan Pan, Xiaochun Cao, Wei Liu, and Ming-Hsuan Yang. Gated fusion network for single image dehazing. In CVPR, 2018. 2", + "[42] Wenqi Ren, Jinshan Pan, Hua Zhang, Xiaochun Cao, and Ming-Hsuan Yang. Single image dehazing via multi-scale convolutional neural networks with holistic edges. In ICCV, 2020. 2", + "[43] Carlos Riquelme, Joan Puigcerver, Basil Mustafa, Maxim Neumann, Rodolphe Jenatton, André Susano Pinto, Daniel Keysers, and Neil Houlsby. Scaling vision with sparse mixture of experts. *NeurIPS*, 2021. 2, 3, 4, 7, 8", + "[44] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-Net: Convolutional networks for biomedical image segmentation. In MICCAI, pages 234–241. Springer, 2015. 3", + "[45] Giovanni Saggio, Davide Borra, et al. Augmented reality for restoration/reconstruction of artefacts with artistic or historical value. In Augmented reality: some emerging application areas, pages 59-86. InTech Publications, 2011. 1", + "[46] Noam Shazeer, Azalia Mirhoseini, Krzysztof Maziarz, Andy Davis, Quoc Le, Geoffrey Hinton, and Jeff Dean. Outrageously large neural networks: The sparsely-gated mixture-of-experts layer, 2017. 3, 4", + "[47] Ying Tai, Jian Yang, Xiaoming Liu, and Chunyan Xu. Memnet: A persistent memory network for image restoration. In ICCV, 2017. 1, 2", + "[48] Chunwei Tian, Yong Xu, and Wangmeng Zuo. Image denoising using deep cnn with batch renormalization. NN, 2020. 5", + "[49] Jeya Maria Jose Valanarasu, Rajeev Yasarla, and Vishal M Patel. Transweather: Transformer-based restoration of images degraded by adverse weather conditions. In CVPR, 2022. 1, 2, 5", + "[50] Cong Wang, Jinshan Pan, Wei Wang, Jiangxin Dong, Mengzhu Wang, Yakun Ju, and Junyang Chen. Promptstorer: A prompting image restoration method with degradation perception. In NeurIPS, 2023. 1, 2" + ], + "bbox": [ + 91, + 90, + 483, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[51] Tao Wang, Kaihao Zhang, Ziqian Shao, Wenhan Luo, Bjorn Stenger, Tong Lu, Tae-Kyun Kim, Wei Liu, and Hongdong Li. Gridformer: Residual dense transformer with grid structure for image restoration in adverse weather conditions. ICCV, 2024. 5, 6", + "[52] Zhendong Wang, Xiaodong Cun, Jianmin Bao, Wengang Zhou, Jianzhuang Liu, and Houqiang Li. Uformer: A general u-shaped transformer for image restoration. In CVPR, 2022. 1, 2", + "[53] Chen Wei, Wenjing Wang, Wenhan Yang, and Jiaying Liu. Deep retina decomposition for low-light enhancement. arXiv preprint arXiv:1808.04560, 2018. 5", + "[54] Gang Wu, Junjun Jiang, Kui Jiang, and Xianming Liu. Harmony in diversity: Improving all-in-one image restoration via multi-task collaboration. In ACMMM, 2024. 5", + "[55] Haiyan Wu, Yanyun Qu, Shaohui Lin, Jian Zhou, Ruizhi Qiao, Zhizhong Zhang, Yuan Xie, and Lizhuang Ma. Contrastive learning for compact single image dehazing. In CVPR, 2021. 2", + "[56] Fuzhi Yang, Huan Yang, Jianlong Fu, Hongtao Lu, and Baining Guo. Learning texture transformer network for image super-resolution. In CVPR, 2020. 5", + "[57] Hao Yang, Liyuan Pan, Yan Yang, and Wei Liang. Language-driven all-in-one adverse weather removal. In CVPR, 2024. 1, 2, 3", + "[58] Zhiwen Yang, Haowei Chen, Ziniu Qian, Yang Yi, Hui Zhang, Dan Zhao, Bingzheng Wei, and Yan Xu. All-in-one medical image restoration via task-adaptive routing. In MICCAI, 2024. 7, 8", + "[59] Rajeev Yasarla and Vishal M Patel. Uncertainty guided multiscale residual learning-using a cycle spinning cnn for single image de-raining. In CVPR, 2019. 1", + "[60] Ke Yu, Xintao Wang, Chao Dong, Xiaou Tang, and Chen Change Loy. Path-restore: Learning network path selection for image restoration. TPAMI, 2022. 3", + "[61] Xiaoyan Yu, Shen Zhou, Huafeng Li, and Liehuang Zhu. Multi-expert adaptive selection: Task-balancing for all-in-one image restoration. arXiv preprint arXiv:2407.19139, 2024. 1", + "[62] Eduard Zamfir, Zongwei Wu, Nancy Mehta, Danda Pani Paudel, Yulun Zhang, and Radu Timofte. Efficient degradation-aware any image restoration. arXiv preprint arXiv:2405.15475, 2024. 1, 2, 3", + "[63] Eduard Zamfir, Zongwei Wu, Nancy Mehta, Yulun Zhang, and Radu Timofte. See more details: Efficient image superresolution by experts mining. In ICML. PMLR, 2024. 3", + "[64] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, Ming-Hsuan Yang, and Ling Shao. Multi-stage progressive image restoration. In CVPR, 2021. 2, 5, 6", + "[65] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. In CVPR, 2022. 1, 2, 3, 4, 5", + "[66] Jinghao Zhang, Jie Huang, Mingde Yao, Zizheng Yang, Hu Yu, Man Zhou, and Feng Zhao. Ingredient-oriented multidegradation learning for image restoration. In CVPR, 2023. 1, 2, 5, 6, 7" + ], + "bbox": [ + 516, + 90, + 906, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "12762", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[67] Kai Zhang, Wangmeng Zuo, Shuhang Gu, and Lei Zhang. Learning deep cnn denoiser prior for image restoration. In CVPR, 2017. 1, 2", + "[68] Yulun Zhang, Kunpeng Li, Kai Li, Bineng Zhong, and Yun Fu. Residual non-local attention networks for image restoration. arXiv preprint arXiv:1903.10082, 2019. 1, 2", + "[69] Haiyu Zhao, Yuanbiao Gou, Boyun Li, Dezhong Peng, Jiancheng Lv, and Xi Peng. Comprehensive and delicate: An efficient transformer for image restoration. In CVPR, 2023. 2", + "[70] Yanqi Zhou, Tao Lei, Hanxiao Liu, Nan Du, Yanping Huang, Vincent Zhao, Andrew M Dai, Quoc V Le, James Laudon, et al. Mixture-of-experts with expert choice routing. NeurIPS, 2022. 2", + "[71] Yurui Zhu, Tianyu Wang, Xueyang Fu, Xuanyu Yang, Xin Guo, Jifeng Dai, Yu Qiao, and Xiaowei Hu. Learning weather-general and weather-specific features for image restoration under multiple adverse weather conditions. In CVPR, 2023. 6", + "[72] Ozan Özdenizci and Robert Legenstein. Restoring vision in adverse weather conditions with patch-based denoising diffusion models. TPAMI, 2023. 6" + ], + "bbox": [ + 91, + 90, + 485, + 388 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "12763", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Complexity Experts are Task-Discriminative Learners for Any Image Restoration/811da5ba-4d1d-41b9-ac9b-e649712870d3_model.json b/2025/Complexity Experts are Task-Discriminative Learners for Any Image Restoration/811da5ba-4d1d-41b9-ac9b-e649712870d3_model.json new file mode 100644 index 0000000000000000000000000000000000000000..2ad9541661296e84a598da322a3029ba6806bdc0 --- /dev/null +++ b/2025/Complexity Experts are Task-Discriminative Learners for Any Image Restoration/811da5ba-4d1d-41b9-ac9b-e649712870d3_model.json @@ -0,0 +1,2235 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.131, + 0.908, + 0.155 + ], + "angle": 0, + "content": "Complexity Experts are Task-Discriminative Learners for Any Image Restoration" + }, + { + "type": "text", + "bbox": [ + 0.232, + 0.18, + 0.766, + 0.218 + ], + "angle": 0, + "content": "Eduard Zamfir1 Zongwei Wu1* Nancy Mehta1 Yuedong Tan1 Danda Pani Paudel2 Yulun Zhang3 Radu Timofte1" + }, + { + "type": "text", + "bbox": [ + 0.307, + 0.219, + 0.693, + 0.235 + ], + "angle": 0, + "content": "1Computer Vision Lab, CAIDAS & IFI, University of Würzburg" + }, + { + "type": "text", + "bbox": [ + 0.245, + 0.235, + 0.753, + 0.253 + ], + "angle": 0, + "content": "2INSAIT, Sofia University \"St. Kliment Ohridski\" 3Shanghai Jiao Tong University" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.287, + 0.327, + 0.303 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.318, + 0.486, + 0.697 + ], + "angle": 0, + "content": "Recent advancements in all-in-one image restoration models have revolutionized the ability to address diverse degradations through a unified framework. However, parameters tied to specific tasks often remain inactive for other tasks, making mixture-of-experts (MoE) architectures a natural extension. Despite this, MoEs often show inconsistent behavior, with some experts unexpectedly generalizing across tasks while others struggle within their intended scope. This hinders leveraging MoEs' computational benefits by bypassing irrelevant experts during inference. We attribute this undesired behavior to the uniform and rigid architecture of traditional MoEs. To address this, we introduce \"complexity experts\" - flexible expert blocks with varying computational complexity and receptive fields. A key challenge is assigning tasks to each expert, as degradation complexity is unknown in advance. Thus, we execute tasks with a simple bias toward lower complexity. To our surprise, this preference effectively drives task-specific allocation, assigning tasks to experts with the appropriate complexity. Extensive experiments validate our approach, demonstrating the ability to bypass irrelevant experts during inference while maintaining superior performance. The proposed MoCE-IR model outperforms state-of-the-art methods, affirming its efficiency and practical applicability. The source code and models are publicly available at eduardzamfir.github.io/MoCE-IR/" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.718, + 0.222, + 0.733 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.743, + 0.485, + 0.88 + ], + "angle": 0, + "content": "Image restoration [47, 67, 68] is a fundamental problem in computer vision, dealing with reconstructing high-quality images from deteriorated observations. Adverse conditions such as noise [22, 67], haze [5, 38], or rain [7, 59] significantly impact the practical utility of images in downstream applications across various domains, including autonomous navigation [7, 49] or augmented reality [10, 15, 45]. Deep learning-based approaches have remarkably advanced this field, particularly for task-specific image restoration prob" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.286, + 0.691, + 0.378 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.551, + 0.388, + 0.649, + 0.401 + ], + "angle": 0, + "content": "(a) Dense Model" + }, + { + "type": "image", + "bbox": [ + 0.71, + 0.286, + 0.902, + 0.378 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.783, + 0.388, + 0.835, + 0.401 + ], + "angle": 0, + "content": "(b) MoE" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.409, + 0.695, + 0.501 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.625, + 0.51, + 0.796, + 0.524 + ], + "angle": 0, + "content": "(c) Complexity Experts (ours)" + }, + { + "type": "image", + "bbox": [ + 0.71, + 0.408, + 0.906, + 0.501 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.531, + 0.909, + 0.781 + ], + "angle": 0, + "content": "Figure 1. Motivation. (a) Dense all-in-one restoration methods [24, 57] often inefficiently allocate parameters when handling multiple degradation types. (b) While recent Mixture-of-Experts (MoE) approaches [61, 62] address this through sparse computation, their rigid routing mechanisms uniformly distribute inputs across experts without considering the natural relationships between degradations. (c) To overcome these limitations, we introduce Complexity Experts - adaptive processing blocks with size-varying computational units. Our framework dynamically allocates model capacity using a spring-inspired force mechanism that continuously guides routing decisions toward simpler experts when possible, with the force proportional to the complexity of the input degradation. While initially designed for computational efficiency, this approach naturally emerges as a task-discriminative learning framework, assigning degradations to the most suitable experts. This makes it particularly effective for all-in-one restoration methods, where both task-specific processing and cross-degradation knowledge sharing are crucial." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.909, + 0.903 + ], + "angle": 0, + "content": "lems [6, 8, 16, 28, 52, 65]. Recent all-in-onerestoration models [9, 12, 13, 24, 27, 36, 49, 66], however, demonstrate the possibility of handling multiple degradation types within a single model, offering more practical solutions compared to traditional task-specific approaches without the extensive re-training need for each new degradation type. Notable works employ visual [25, 36, 50] or language-based prompt-" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.236, + 0.901 + ], + "angle": 0, + "content": "*Corresponding author" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12753" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.122 + ], + "angle": 0, + "content": "ing [1, 9, 30], contrastive learning [24, 66], and diffusion-based models [1, 30]." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.124, + 0.486, + 0.306 + ], + "angle": 0, + "content": "Despite their success, we observe that the aforementioned models often suffer from inefficiencies, as parameters tied to specific degradations often remain inactive or underutilized when addressing unrelated tasks [57]. This naturally suggests the use of mixture-of-experts (MoE) architectures for task-specific processing. Yet, current MoE-based approaches [29, 57, 62] typically incorporate routing mechanisms based on language [57] or degradation priors [29, 62], leading to imbalanced optimization, where some experts generalize well, while others struggle with their intended tasks. This inconsistency limits the potential computational benefits of bypassing task-irrelevant experts during inference." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.308, + 0.486, + 0.504 + ], + "angle": 0, + "content": "We attribute this undesired behavior to two limitations in current approaches. First, the uniform and rigid architecture of existing MoE models fails to account for the varying complexity requirements across different restoration tasks. For instance, motion blur demands localized processing with strong spatial awareness, while haze removal requires broader contextual understanding [4, 35], necessitating adaptive processing aligned to the task requirements. Second, the challenge of appropriately routing tasks to experts is complicated by the unknown complexity of each degradation type a priori. Typically, MoE models [37, 39, 43, 70] aim to balance expert utilization, preventing the model from collapsing into a single-expert dependency." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.507, + 0.485, + 0.598 + ], + "angle": 0, + "content": "In this work, we introduce a novel mixture-of-complexity-experts (MoCE) framework for all-in-one image restoration that directly addresses these limitations. Our key innovation lies in designing expert blocks with increasing computational complexity and receptive fields, allowing the model to adaptively match processing capacity with task requirements." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.6, + 0.486, + 0.901 + ], + "angle": 0, + "content": "To address the routing challenge, we introduce a complexity-aware allocation mechanism that preferentially directs tasks to lower-complexity experts. This approach implements a complexity-proportional bias term, analogous to a mechanical spring force, to guide expert selection. Surprisingly, this straightforward approach results in effective task-specific allocation, naturally directing inputs to experts with appropriate complexity levels. Our MoCE framework offers two significant advantages: it ensures efficient inference by selectively bypassing irrelevant experts, thereby reducing computational overhead, while simultaneously maintaining high-quality restoration performance across diverse tasks. Through extensive experiments across multiple image restoration tasks, we demonstrate that our approach not only preserves restoration quality but surpasses recent state-of-the-art methods. These results validate both the effectiveness of our complexity-based expert design and the practicality of our routing strategy for real-world applications. Our approach establishes a new benchmark for all-in-one image restoration, enhancing both efficiency and fidelity." + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.092, + 0.89, + 0.107 + ], + "angle": 0, + "content": "The principal contributions of this work are threefold:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.109, + 0.907, + 0.154 + ], + "angle": 0, + "content": "- We propose MoCE-IR, achieving SoTA all-in-one image restoration performance with improved efficiency compared to prior works." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.155, + 0.907, + 0.199 + ], + "angle": 0, + "content": "- Our novel MoE layer selectively activates complexity experts based on input requirements, unifying task-specific and holistic learning in a single architecture." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.2, + 0.906, + 0.245 + ], + "angle": 0, + "content": "- We develop a complexity-aware routing mechanism that balances restoration quality with computational efficiency by adaptive expert allocation." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.109, + 0.907, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.265, + 0.664, + 0.28 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.292, + 0.909, + 0.669 + ], + "angle": 0, + "content": "Task Specific Image Restoration. Reconstructing the clean image from its degraded counterpart is a highly ill-posed problem, however, a great body of work have addressed image restoration from a data-driven learning perspective, achieving tremendous results compared to prior hand-crafted methods [22, 26, 47, 52, 65, 67, 68]. Most proposed solutions build on convolutional [6, 47, 67, 68] or Transformer-based architectures [8, 28, 52, 65] addressing single degradation tasks, such as denoising [8, 67, 68], dehazing [41, 42, 55] or deraining [18, 40]. Contrary to CNN-based networks, Transformer offer strong modeling capabilities for capturing global dependencies, which makes them outstanding image restorers [28, 65, 69]. Self-attention's quadratic complexity w.r.t the image size poses a challenge for resource-constrained applications. Conversely, convolutions offer fast and efficient processing with limited global context, but they scale more effectively with increasing input size. Recently, FFTformer [21] introduced an efficient approximation of query and key interactions in the Fourier domain, significantly reducing computational demands compared to window-based attention methods for image deblurring. Building on this, we incorporate specialized experts into a Transformer-based architecture [65] to efficiently capture both shared and distinct contextual information, addressing the diverse demands of image restoration tasks." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.675, + 0.911, + 0.901 + ], + "angle": 0, + "content": "All-in-One Image Restoration. Restoring degraded images often requires multiple models, complicating practical implementation, especially when images exhibit a combination of various degradations. An emerging field known as all-in-one image restoration is advancing in low-level computer vision, utilizing a single deep blind restoration model to tackle multiple degradation types simultaneously [7, 19, 36, 49, 64, 66]. The seminal work, AirNet [24] achieves blind all-in-one image restoration using contrastive learning to extract degradation representations, which guide the restoration process. Similarly, IDR [66] employs a meta-learning-based two-stage approach to model degradation through underlying physical principles. Prompt-based learning [25, 36, 50] has also gained traction, with [36] introducing tunable prompts to encode degradation-specific in" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "12754" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.089, + 0.362, + 0.308 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.16, + 0.311, + 0.296, + 0.322 + ], + "angle": 0, + "content": "(a) Architecture overview" + }, + { + "type": "image", + "bbox": [ + 0.399, + 0.092, + 0.6, + 0.3 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.435, + 0.31, + 0.562, + 0.323 + ], + "angle": 0, + "content": "(b) Transformer blocks." + }, + { + "type": "image", + "bbox": [ + 0.635, + 0.092, + 0.905, + 0.302 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.68, + 0.31, + 0.86, + 0.323 + ], + "angle": 0, + "content": "(c) Mixture-of-complexity-experts" + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.332, + 0.907, + 0.388 + ], + "angle": 0, + "content": "Figure 2. Proposed MoCE-IR framework. Despite recent advances in MoE-based image restoration [57, 62, 63], inconsistent expert behavior—where some experts over-generalize while others underperform—limits their computational efficiency. We address this through complexity experts: flexible blocks with varying computational capacity and receptive fields. Our MoCE-IR employs an asymmetric encoder-decoder architecture where each decoder block contains a mixture-of-complexity-experts layer for adaptive capacity routing." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.397, + 0.484, + 0.563 + ], + "angle": 0, + "content": "formation, albeit with a high parameter count. With recent advancements in multimodal integration, such as combining images and language, there has been growing interest in language-guided image restoration. Notable approaches include DA-CLIP [30], InstructIR [9], and UniProcessor [12]. However, these methods often depend on high-quality text prompts and typically require LLMs, limiting their applicability on resource-constrained devices. In contrast, our contribution lies in a novel, parameter-efficient approach that focuses on selective activation, effectively addressing diverse degradations while remaining lightweight and adaptable." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.569, + 0.484, + 0.794 + ], + "angle": 0, + "content": "Dynamic Networks. Dynamic networks have evolved from basic conditional computation [3] to sophisticated Mixture-of-Experts (MoE) architectures that expand model capacity while maintaining efficient inference costs [37, 43, 46]. This approach has proven particularly effective in transformer-based NLP and high-level vision tasks. In image restoration, Path-Restore [60] introduced content-aware patch routing with difficulty-regulated rewards, while recent all-in-one models [29, 57, 62] leverage various priors for expert routing. Our work addresses the challenge of adaptively routing corrupted images to appropriate computational units within a generalist restoration framework. By dynamically aligning image characteristics with computational requirements, we achieve a unified model in both parameter and architectural space, advancing beyond previous limitations." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.814, + 0.182, + 0.829 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.901 + ], + "angle": 0, + "content": "In this section, we outline the core principles of our All-in-One image restoration method. As shown in Fig. 2a, our pipeline features a U-shaped architecture [44, 65] with an asymmetric encoder-decoder design. A \\(3 \\times 3\\) convolution" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.397, + 0.908, + 0.563 + ], + "angle": 0, + "content": "first extracts shallow features from the degraded input, which then pass through 4 levels of encoding and decoding stages. While both stages utilize Transformer blocks [36, 65], our novel MoCE layers are exclusively integrated into the decoder blocks (Fig. 2b). To further enhance the decoder's feature enrichment capabilities, we introduce high-frequency guidance via Sobel-filtered global feature vector to promote the frequency awareness of the gating functions. Finally, a global residual connection links the shallow features to the output of the refinement stage, refining the crucial high-frequency details before producing the restored image." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.573, + 0.795, + 0.589 + ], + "angle": 0, + "content": "3.1. Mixture-of-Complexity-Experts" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.595, + 0.907, + 0.655 + ], + "angle": 0, + "content": "To overcome the limitations of uniform MoE architectures in image restoration, we propose an efficient input-adaptive model for all-in-one image restoration that dynamically adjusts processing capacity to match task requirements." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.656, + 0.909, + 0.868 + ], + "angle": 0, + "content": "Technically, at the core of our framework is a specialized MoCE layer, consisting of \\( n \\) complexity experts \\( \\mathbf{E} \\) and a single shared expert \\( \\mathbf{S} \\). As shown in Fig. 2b, our experts are designed with increasing computational complexity and progressively larger receptive fields, in contrast to the uniform architecture of traditional MoE designs [37, 43, 46]. The interaction between these components is orchestrated through a two-level gating mechanism: first projecting routed tokens into expert-specific embeddings, then combining specialized and agnostic features through modulation, and finally merging the outputs via cross-attention. This enables our model to capture both degradation-specific features and interdegradation relationships, while maintaining a bias toward computationally efficient processing paths." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.908, + 0.902 + ], + "angle": 0, + "content": "Expert Design. Each complexity expert block as shown in Fig. 2c is designed based on two fundamental principles:" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "12755" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.484, + 0.317 + ], + "angle": 0, + "content": "ensuring computational efficiency and capturing hierarchical spatial features, which are critical for restoring fine details and addressing diverse degradation patterns. Given the growing computational demands associated with increasing expert numbers and capacities, it is crucial to prioritize an overall efficient expert design. To this end, we implement a nested expert structure, progressively reducing channel dimensionality \\( r \\) within each subsequent expert to control the computational overhead. Simultaneously, we increase the receptive field, i.e. window partition size \\( w \\), within each expert to adaptively balance the localized and global processing, tailored to the specific requirements of the input degradation. Thus, our complexity experts are constructed by scaling the design along two key axes: window partition size and channel dimension." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.319, + 0.484, + 0.576 + ], + "angle": 0, + "content": "More specifically, each expert \\(\\mathbf{E}\\) projects its input tokens to the embedding \\(\\mathbf{x}_E\\in \\mathbb{R}^{H\\times W\\times r}\\), where \\(r = C / 2^i\\) for \\(i\\in \\{i,\\dots,n\\}\\), to highlight the most relevant information along the channel dimension. Following this, a window-based self-attention (WSA) mechanism captures spatial information effectively. To optimize performance, we employ an FFT-based approximation [21] for efficient matrix multiplication between queries \\(Q\\) and keys \\(K\\) in the Fourier domain. The refined features are then transformed back into the spatial domain and re-projected to their original dimensionality. Lastly, the output of the expert block \\(\\hat{\\mathbf{x}}_E^i\\) is modulated through element-wise multiplication with the holistic features from the shared expert, S. As shared expert, we employ the transposed self-attention (T-SA) module [65] in the channel dimension, while the complexity experts operate spatially. More concretely, the modulated expert features \\(\\mathbf{y}_{\\mathbf{E}}^{\\mathrm{i}}\\) are obtained as following:" + }, + { + "type": "equation", + "bbox": [ + 0.152, + 0.587, + 0.484, + 0.605 + ], + "angle": 0, + "content": "\\[\n\\mathbf {y} _ {\\mathbf {E}} ^ {i} = \\hat {\\mathbf {x}} _ {E} ^ {i} \\odot \\mathbf {S} (\\mathbf {x}) \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.12, + 0.608, + 0.484, + 0.627 + ], + "angle": 0, + "content": "\\[\n\\text {w i t h} \\hat {\\mathbf {x}} _ {E} ^ {i} = \\mathbf {E} ^ {i} \\left(\\mathbf {x} _ {E} ^ {i}\\right) = \\operatorname {F F T - W S A} _ {\\mathrm {W} _ {i}} \\left(\\mathbf {W} _ {C \\rightarrow r _ {i}} ^ {i} \\mathbf {x} _ {E} ^ {i}\\right) \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.113, + 0.629, + 0.484, + 0.645 + ], + "angle": 0, + "content": "\\[\n\\operatorname {a n d} \\mathbf {S} (\\mathbf {x}) = \\mathrm {T} - \\operatorname {S A} (\\mathbf {x}) \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.657, + 0.484, + 0.792 + ], + "angle": 0, + "content": "where linear layers for projection are denoted as \\(\\mathbf{W}\\), \\(\\odot\\) denotes element-wise multiplication and \\(i\\) is the index of the current complexity expert \\(\\mathbf{E}^i\\) with window size \\(w_i\\) and embedding dimensionality \\(r_i\\). For example, the most lightweight expert is characterized by the smallest embedding dimension \\(r_1\\) and window size \\(w_1\\), while the dimensions and window partitions for the other experts can increase linearly. Throughout our network, the capacity of the shared expert remains constant, with \\(r = C\\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.803, + 0.337, + 0.82 + ], + "angle": 0, + "content": "3.2. Complexity-aware Routing" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Drawing inspiration from sparse MoE [43, 46], we integrate linear layers within each decoder block to enable a routing mechanism, associating input features \\(\\mathbf{x} \\in \\mathbb{R}^{H \\times W \\times C}\\) with their corresponding specialized complexity experts \\(\\mathbf{E}\\). However, in the context of image restoration, a key challenge" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.908, + 0.304 + ], + "angle": 0, + "content": "arises in achieving a scale-invariant tokenization of input images, ensuring consistency across varying resolutions and scales. While previous MoE approaches [37, 43, 46] employ a token-based routing, we opt for an image-level routing strategy instead, where we select experts for entire input image. The routing function \\( g(\\mathbf{x}) \\) orchestrates the allocation of input samples based on the required compute to corresponding complexity experts \\( \\mathbf{E}^i \\), indexed by \\( i \\in 1, \\dots, n \\) with \\( n \\) representing the total number of experts. Within \\( g(\\mathbf{x}) \\), we select the top- \\( k \\) elements of the softmax distribution associating input tokens to experts, setting all other elements to zero. In practice, we use \\( k = 1 \\) while \\( \\epsilon \\) is sampled independently \\( \\epsilon \\sim \\mathcal{N}(0, 1/n^2) \\) entry-wise enabling noisy top-1 routing:" + }, + { + "type": "equation", + "bbox": [ + 0.598, + 0.318, + 0.907, + 0.335 + ], + "angle": 0, + "content": "\\[\ng (\\mathbf {x}) = \\operatorname {t o p} _ {k} \\left(\\operatorname {S o f t m a x} \\left(\\mathbf {W} \\mathbf {x} + \\epsilon\\right)\\right) \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.348, + 0.909, + 0.545 + ], + "angle": 0, + "content": "During training, we incorporate an auxiliary loss \\(\\mathcal{L}_{\\mathrm{aux}}\\) inspired by the load balancing loss proposed by Riquelme et al. [43]. We complement the importance term to favor experts based on their computational complexity, allowing for more effective balancing aligned with expert capacity. As in [43], the balanced expert utilization is encouraged through an importance loss, where the importance of each expert \\(\\mathbf{E}^i\\) for a batch of images is defined as the sum of the routing weights associated with the \\(i\\)-th expert across the entire batch. Additionally, we introduce a complexity bias \\(\\mathbf{b}\\) by calculating the number of learnable parameters \\(p_i\\) in each expert \\(\\mathbf{E}^i\\) and normalizing these values with respect to the largest parameter count of the experts, \\(p_{\\mathrm{max}}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.577, + 0.556, + 0.907, + 0.598 + ], + "angle": 0, + "content": "\\[\n\\operatorname {I m p} _ {i} (\\mathbf {x}) = \\left(\\sum_ {x \\in \\mathbf {x}} \\operatorname {S o f t m a x} (W x) _ {i}\\right) * \\mathbf {b}, \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.588, + 0.6, + 0.907, + 0.618 + ], + "angle": 0, + "content": "\\[\n\\text {w i t h} \\mathbf {b} = \\left[ \\frac {p _ {1}}{p _ {\\max }}, \\frac {p _ {2}}{p _ {\\max }}, \\dots , \\frac {p _ {n}}{p _ {\\max }} \\right], \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.63, + 0.909, + 0.828 + ], + "angle": 0, + "content": "where \\(W\\) is the layer-specific weight for the router \\(g(\\mathbf{x})\\). This approach assigns progressively lower weights to experts with lower complexity. For intuitive understanding, this mechanism parallels a mechanical spring system, where the restoring force is proportional to both displacement and the spring constant. In our context, the expert's parameter count corresponds to displacement, while the normalization factor serves as the material-specific constant, collectively determining the magnitude of complexity bias. Without this, the routers lack a basis for selecting specific experts, leading to randomized routing and thus suboptimal model capacity utilization. Lastly, the complexity-aware importance Imp is used to compute the total auxiliary loss \\(\\mathcal{L}_{\\mathrm{aux}}\\) defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.532, + 0.84, + 0.907, + 0.858 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {a u x}} (\\mathbf {x}) = ^ {1 / 2} \\operatorname {C V} \\left(\\operatorname {I m p} (\\mathbf {x})\\right) ^ {2} + ^ {1 / 2} \\operatorname {C V} \\left(\\operatorname {L o a d} (\\mathbf {x})\\right) ^ {2}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.908, + 0.901 + ], + "angle": 0, + "content": "where CV denotes the coefficient of variation [43]. The effect of the auxiliary loss is shown in Fig. 5 and Tab. 5a." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "12756" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.907, + 0.133 + ], + "angle": 0, + "content": "Table 1. Comparison to state-of-the-art on three degradations. PSNR (dB, ↑) and SSIM (↑) are reported on the full RGB images. Our method sets a new state-of-the-art on average across all benchmarks on two model scales, showcasing its scalability while being more efficient than prior work. ‘-’ represents unreported results and ‘*’ indicates VLM-based approaches. The best performances are highlighted." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.134, + 0.905, + 0.312 + ], + "angle": 0, + "content": "
MethodParams.DehazingDerainingDenoisingAverage
SOTSRain100LBSD68σ=15BSD68σ=25BSD68σ=50
LightBRDNet [48]-23.23.89527.42.89532.26.89829.76.83626.34.69327.80.843
LPNet [14]-20.84.82824.88.78426.47.77824.77.74821.26.55223.64.738
FDGAN [11]-24.71.92929.89.93330.25.91028.81.86826.43.77628.02.883
DL [13]2M26.92.93132.62.93133.05.91430.41.86126.90.74029.98.876
AirNet [24]9M27.94.96234.90.96733.92.93331.26.88828.00.79731.20.910
MoCE-IR-S (ours)11M30.94.97938.22.98334.08.93331.42.88828.16.79832.57.916
HeavyMPRNet [64]16M25.28.95533.57.95433.54.92730.89.88027.56.77930.17.899
PromptIR [36]36M30.58.97436.37.97233.98.93331.31.88828.06.79932.06.913
Gridformer [51]34M30.37.97037.15.97233.93.93131.37.88728.11.80132.19.912
Art-PromptIR [54]33M30.83.97937.94.98234.06.93431.42.89128.14.80132.49.917
DA-CLIP* [30]125M29.46.96336.28.96830.02.82124.86.58522.29.476--
UniProcessor* [12]1002M31.66.97938.17.98234.08.93531.42.89128.17.80332.70.918
MoCE-IR (ours)25M31.34.97938.57.98434.11.93231.45.88828.18.80032.73.917
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.316, + 0.907, + 0.358 + ], + "angle": 0, + "content": "Table 2. Comparison to state-of-the-art on five degradations. PSNR (dB, ↑) and SSIM (↑) are reported on the full RGB images with (*) denoting general image restorers, others are specialized all-in-one approaches. Under more challenging degradations, our MoCE-IR framework outperforms prior work across two model scales. The best performances are highlighted." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.359, + 0.905, + 0.549 + ], + "angle": 0, + "content": "
MethodParams.DehazingDerainingDenoisingDeblurringLow-LightAverage
SOTSRain100LBSD68σ=25GoProLOLv1
LightSwinIR* [26]1M21.50.89130.78.92330.59.86824.52.77317.81.72325.04.835
DL [13]2M20.54.82621.96.76223.09.74519.86.67219.83.71221.05.743
TAPE [27]1M22.16.86129.67.90430.18.85524.47.76318.97.62125.09.801
AirNet [24]9M21.04.88432.98.95130.91.88224.35.78118.18.73525.49.847
MoCE-IR-S (ours)11M31.33.97837.21.97831.25.88428.90.87721.68.85130.08.913
HeavyNAFNet* [6]17M25.23.93935.56.96731.02.88326.53.80820.49.80927.76.881
DGUNet* [33]17M24.78.94036.62.97131.10.88327.25.83721.87.82328.32.891
Restormer* [65]26M24.09.92734.81.96231.49.88427.22.82920.41.80627.60.881
MambaIR [16]27M25.81.94436.55.97131.41.88428.61.87522.49.83228.97.901
Transweather [49]38M21.32.88529.43.90529.00.84125.12.75721.21.79225.22.836
IDR [66]15M25.24.94335.63.96531.60.88727.87.84621.34.82628.34.893
Gridformer [51]34M26.79.95136.61.97131.45.88529.22.88422.59.83129.33.904
InstructIR-5D [9]17M27.10.95636.84.97331.40.87329.40.88623.00.83629.55.908
MoCE-IR (ours)25M30.48.97438.04.98231.34.88730.05.89923.00.85230.58.919
" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.563, + 0.224, + 0.58 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.592, + 0.483, + 0.729 + ], + "angle": 0, + "content": "We conduct experiments by strictly following previous works in general image restoration [36, 66] under two different settings: (i) All-in-One and (ii) Composited degradations. In the all-in-one setting, a unified model is trained across multiple degradation types, where we consider three and five distinct degradations. In the Composited degradations setting, a single model is trained to handle both isolated degradation occurrences as well as combinations of up to three different types of degradations." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.735, + 0.486, + 0.902 + ], + "angle": 0, + "content": "Implementation Details. Our MoCE-IR framework is end-to-end trainable requiring no multi-stage optimization of any components. The architecture features an asymmetrical, four-level encoder-decoder design. The encoder consists of four levels with varying numbers of transformer blocks, arranged as \\([4,6,6,8]\\) from the top level to the lowest. The decoder includes three levels, containing \\([2,4,4]\\) transformer blocks, respectively. In each MoCE layer, we use \\(n = 4\\) nested experts, where the dimensionality of the expert embedding is given by \\(R = C / 2^i\\) for \\(i\\in \\{i,\\dots,n\\}\\). The network's initial embedding dimension is set to \\(C = 32\\), doubling at each" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.565, + 0.907, + 0.685 + ], + "angle": 0, + "content": "subsequent level while the spatial resolution is halved. We follow the training configuration of prior work [36], and train our models for 120 epochs with a batch size of 32. We optimize the \\( L_{1} \\) loss in the RGB and Fourier domain using the Adam [20] optimizer \\( (\\beta_{1} = 0.9, \\beta_{2} = 0.999) \\) with an initial learning rate of \\( 2 \\times 10^{-4} \\) and cosine decay. During training, we utilize crops of size \\( 128^{2} \\) with horizontal and vertical flips as augmentations." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Datasets. For all-in-one, we follow existing work [24, 36] and include following datasets: For image denoising, we combine the BSD400 [2] and WED [31] datasets, adding Gaussian noise at levels \\(\\sigma \\in [15, 25, 50]\\) to create noisy images. Testing is conducted on the BSD68 [32]. For deraining, we use Rain100L [56]. The dehazing task utilises the SOTS [23] dataset. For deblurring and low-light enhancement, we employ the GoPro [34] and the LOL-v1 [53] dataset, respectively. To develop a unified model for all tasks, we merge these datasets in a three (AIO-3) or five (AIO-5) degradation setting, and train for 120 epochs and directly evaluate them across different tasks. For the composite degradation setting, we use the CDD11 dataset [17] and train our method for 200 epochs, keeping all other settings unchanged." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12757" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.09, + 0.907, + 0.131 + ], + "angle": 0, + "content": "Table 3. Comparison to state-of-the-art on composited degradations. PSNR (dB, ↑) and SSIM (↑) are reported on the full RGB images with (*) denoting general image restorers, others are specialized all-in-one approaches. Our MoCE-IR method consistently outperforms even larger models, with favorable results in composited degradation scenarios." + }, + { + "type": "table", + "bbox": [ + 0.09, + 0.133, + 0.907, + 0.63 + ], + "angle": 0, + "content": "
MethodParams.CDD11-SingleCDD11-DoubleCDD11-TripleAvg.
Low (L)Haze (H)Rain (R)Snow (S)L+HL+RL+SH+RH+SL+H+RL+H+S
AirNet [24]9M24.83.77824.21.95126.55.89126.79.91923.23.77922.82.71023.29.72322.21.86823.29.90121.80.70822.24.72523.75.814
PromptIR [36]36M26.32.80526.10.96931.56.94631.53.96024.49.78925.05.77124.51.76124.54.92423.70.92523.74.75223.33.74725.90.850
WGWSNet [71]26M24.39.77427.90.98233.15.96434.43.97324.27.80025.06.77224.60.76527.23.95527.65.96023.90.77223.97.77126.96.863
WeatherDiff [72]83M23.58.76321.99.90424.85.88524.80.88821.83.75622.69.73022.12.70721.25.86821.99.86821.23.71621.04.69822.49.799
OneRestore [17]6M26.48.82632.52.99033.40.96434.31.97325.79.82225.58.79925.19.78929.99.95730.21.96424.78.78824.90.79128.47.878
MoCE-IR-S (ours)11M27.26.82432.66.99034.31.97035.91.98026.24.81726.25.80026.04.79329.93.96430.19.97025.41.78925.39.79029.05.881
DehazingInputAirNetPromptIROursGT
Deraining
Denoising
Input CropsError Maps (darker means better)GT Crops
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.635, + 0.907, + 0.679 + ], + "angle": 0, + "content": "Figure 3. Visual results. We compare MoCE-IR-S to AirNet [24], and PromptIR [36] in the all-in-one setting with three degradations. MoCE-IR-S effectively removes haze and rain streaks while preserving image sharpness, achieving high-quality restoration. An error heatmap is provided, with color transitioning from black to white to indicate increasing pixel-wise error." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.685, + 0.441, + 0.701 + ], + "angle": 0, + "content": "4.1. Comparison to State-of-the-Art Methods" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.72, + 0.486, + 0.902 + ], + "angle": 0, + "content": "All-in-One: Three Degradations. We compare our all-in-one restoration approach with specialized restoration methods, including both image-only models, such as MPR-Net [64], AirNet [24], PromptIR [36], and Gridformer [51], and vision-language models like DA-CLIP [30], InstructIR [9], and UniProcessor [12]. All methods were trained simultaneously on three types of degradation: haze, rain, and Gaussian noise. The results in Tab. 1 show that our proposed model, MoCE-IR, consistently outperforms previous works, both in light and heavy model scales. In the light category, MoCE-IR achieves the highest scores across all benchmarks with an average improvement in PSNR of" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.686, + 0.907, + 0.747 + ], + "angle": 0, + "content": "1.37 dB, surpassing other models like AirNet [24]. In the heavy scale, MoCE-IR competes closely with UniProcessor [12], achieving an average PSNR of 32.73 dB with \\(98\\%\\) fewer parameters, validating our efficiency and effectiveness." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.909, + 0.903 + ], + "angle": 0, + "content": "All-in-One: Five Degradations. Based upon [24, 66], we expand the three-degradation setting to include deblurring and low-light image enhancement, further validating our method's effectiveness. As shown in Tab. 2, our approach excels by learning specialized experts with input-adaptive complexity and maximizing synergies across degradations. Our framework outperforms previous methods on the lightest model scale, with an average PSNR improvement of 4.31 dB over AirNet [24]. Even compared to larger networks, our lightest model surpasses the VLM-based InstructIR [9] by" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12758" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.098, + 0.091, + 0.482, + 0.221 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.228, + 0.485, + 0.286 + ], + "angle": 0, + "content": "Figure 4. Complexity-efficiency tradeoff. Visualization of PSNR and parameter counts of proposed method compared to prior work. Proposed MoCE-IR surpasses prior methods, achieving SoTA results in all-in-one image restoration with enhanced efficiency." + }, + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.289, + 0.484, + 0.318 + ], + "angle": 0, + "content": "Table 4. Computational demands. GFLOPS are computed on an input image of size \\(224 \\times 224\\) using a NVIDIA RTX 4090 GPU." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.318, + 0.484, + 0.395 + ], + "angle": 0, + "content": "
MethodParams.MemoryFLOPSRuntime
AirNet [24]8.93M4829M238G42.17 ± 0.23
PromptIR [36]35.59M9830M132G41.28 ± 0.43
IDR [66]15.34M4905M98G-
MoCE-IR (ours)25.35M5887M80.59 ± 5.21G23.36 ± 2.47
MoCE-IR-S (ours)11.47M4228M36.93 ± 2.32G22.15 ± 2.59
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.415, + 0.484, + 0.446 + ], + "angle": 0, + "content": "0.33 dB on average, and on a comparable model scale, our framework extends this advantage by an additional 0.50 dB." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.454, + 0.486, + 0.668 + ], + "angle": 0, + "content": "All-in-One: Composed Degradations. To simulate more realistic degradation scenarios, recent work [17] has trained all-in-one restorers not only on a mix of different degradations but also on more challenging cases where multiple degradation types are combined within a single image. We extend the previous AIO settings by including rain, haze, snow, and low illumination as individual degradation types, along with various composite degradation scenarios, resulting in a total of eleven distinct restoration settings. As shown in Tab. 3, our method outperforms previous all-in-one restorers by an average improvement of \\(0.64\\mathrm{dB}\\) over OneRestore [17], while being among the smallest models in size. This further validates our effectiveness with adaptive modeling capacity based on input complexity." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.675, + 0.486, + 0.902 + ], + "angle": 0, + "content": "Visual Results. We show visual results obtained under the three degradation settings in Fig. 3. In certain demanding hazy scenarios, both AirNet [24] and PromptIR [36] reveal constraints in completely eradicating haziness, resulting in noticeable color intensity discrepancies, whereas our approach ensures precise color reconstruction. Furthermore, in challenging rainy scenes, these popular approaches continue to exhibit notable remnants of rain, which our method adeptly eliminates, distinguishing itself from other approaches. Additionally, our method also generates clear and sharp denoised outputs. We further provide the error maps for visual differences where our method contains less noticeable errors. Coupled with the quantitative comparisons, these findings underscore the effectiveness of our method. Please refer to the supplementary for more visuals." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.089, + 0.907, + 0.187 + ], + "angle": 0, + "content": "Table 5. Architecture analysis. (a) We investigate scaling the expert embedding \\( r \\) and the bias normalization for MoCE-IR-S. (b) We compare MoCE to several baselines [29, 43, 58], demonstrating that our complexity experts consistently outperform them. Notably, complexity-biased routing further boosts performance, surpassing standard load balancing (LB) [43]. PSNR (dB, ↑) and SSIM (↑) are reported on full RGB images." + }, + { + "type": "table_caption", + "bbox": [ + 0.618, + 0.189, + 0.8, + 0.201 + ], + "angle": 0, + "content": "(a) Ablation on the expert scaling." + }, + { + "type": "table", + "bbox": [ + 0.515, + 0.203, + 0.905, + 0.359 + ], + "angle": 0, + "content": "
Embedding r scalingBias normalizationAvg. AIO-3
NestedpMin32.44.913
ExponentialpMax32.47.914
NestedpMax32.57.916
(b) Ablation on the complexity-aware routing.
MethodSOTSRain100LBSD6815-50Avg.
MoCE + Complexity Bias (ours)30.94.97938.22.98331.22.873
Baseline (MoE + LB [45])30.28.97737.77.98131.14.871
w/ MoCE + LB [45]30.13.97838.24.98331.19.872
w/ AMIR [58]29.97.97537.90.98131.17.872
w/ WM-MoE [29]30.10.97636.72.97631.05.869
" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.369, + 0.68, + 0.385 + ], + "angle": 0, + "content": "4.2. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.392, + 0.909, + 0.468 + ], + "angle": 0, + "content": "We conduct detailed studies on the components within our method, highlighting the emergence of task-discriminative learners. All experiments are conducted in the all-in-one setting with three degradations. Please refer to the supplementary for additional ablations." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.473, + 0.909, + 0.7 + ], + "angle": 0, + "content": "**Complexity-Efficiency Trade-off.** Our MoCE-IR framework addresses the critical challenge of scalability in vision tasks by optimizing both reconstruction quality and computational efficiency. We visualize in Fig. 4 the progressive improvements in reconstruction fidelity of our framework on two complexity scales. Furthermore, Tab. 4 validates that our approach significantly outperforms SoTA methods in terms of memory utilization. Unlike traditional MoE architectures with fixed computational demands, our complexity experts enable dynamic FLOPS allocation. For instance, our lightweight model averages 36.9 GFLOPS, while the heavyweight model averages 80.59 GFLOPS. This adaptive computation, combined with lower memory consumption and parameter counts, gives MoCE-IR a substantial advantage over contemporary methods." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.911, + 0.903 + ], + "angle": 0, + "content": "Expert Scaling. As shown in Tab. 5a, we systematically evaluate different scaling strategies regarding the embedding dimensionality \\( r \\) of each complexity expert. Our investigation compares exponential and nested parameter scaling, where the window partition is always fixed and follows a \\( 2^{(2 + i)} \\) progression. Further, we investigate the bias normalization, whether normalizing to the lowest or highest expert capacity. Overall, the nested parameter scaling emerges as the optimal choice, providing sufficient expert channel diversity while maintaining reasonable computational costs. Normalizing router weights by the highest expert capacity, rather than the minimum, proves crucial for performance as minimum capacity normalization forces router weights" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "12759" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.107, + 0.092, + 0.315, + 0.202 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.319, + 0.092, + 0.506, + 0.202 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.092, + 0.7, + 0.202 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.705, + 0.092, + 0.893, + 0.202 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.205, + 0.907, + 0.277 + ], + "angle": 0, + "content": "Figure 5. Routing visualization for the AIO-3 setting. (a) While load balancing [43] ensures uniform expert utilization, it neglects shared task dependencies and task-specific characteristics, limiting restoration quality. (b)-(d) Complexity-aware routing fosters task discrimination by directing complex degradations to experts with broader contextual understanding and vice versa. This allows some experts to generalize across tasks while others specialize, enhancing adaptability. We visualize the average decisions made by each router for dehazing, deraining and denoising. The y-axis indicates increasing expert complexity." + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.282, + 0.483, + 0.31 + ], + "angle": 0, + "content": "Table 6. Expert generalization. We report PSNR for MoCE-IR-S with expert assignments fixed post-training. \\(\\mathcal{E}\\) denotes expert." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.312, + 0.485, + 0.395 + ], + "angle": 0, + "content": "
DegradationRoutingLearned ChoiceManual Choice
\\( \\mathcal{E}_1 \\)\\( \\mathcal{E}_2 \\)\\( \\mathcal{E}_3 \\)\\( \\mathcal{E}_4 \\)
Rain NoiseLoad Balance [45]Not Applicable28.2733.3328.2433.3728.2533.3728.2733.40
Rain NoiseComplexity Bias (ours)\\( \\mathcal{E}_1 \\)30.4530.2129.7324.65
\\( \\mathcal{E}_4 \\)33.9233.9333.9334.00
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.406, + 0.484, + 0.437 + ], + "angle": 0, + "content": "of higher capacity experts to escalate uncontrollably, thus excluding them from the gating mechanism." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.441, + 0.484, + 0.653 + ], + "angle": 0, + "content": "Impact of Complexity-aware Routing. Building on the findings from Tab. 5a, we adopt nested expert scaling and evaluate our complexity-aware routing strategy against conventional load balancing [43], which enforces uniform expert usage. As shown in Tab. 5b, our MoCE-IR framework achieves superior performance by preferentially routing to lower-complexity experts, activating higher-complexity experts only when their added capacity provides clear benefits. This efficient model utilization not only boosts image restoration performance, consistently surpassing related adaptive restoration methods [29, 58], but also enhances out-of-distribution generalization as MoCE-IR effectively handles degradations that with significant deviation from the training data; see the supplementary for more evaluations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.658, + 0.484, + 0.809 + ], + "angle": 0, + "content": "Routing Analysis. Fig. 5 illustrates the average expert selection per layer for different degradations: haze, rain, and noise. Higher-complexity experts are favored for tasks requiring broader context, like haze and noise, while lighter experts are chosen for rain streak removal. Notably, some experts handle multiple tasks, while others remain task-specific. Our complexity-conditioned routing promotes task-discriminative learning by directing inputs to relevant experts, preventing the collapse into a single-expert dependency, ensuring efficient and diverse operation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.484, + 0.901 + ], + "angle": 0, + "content": "The baseline MoE [43] approach, lacking complexity awareness and scaled experts, uniformly distributes degraded samples across all experts. Our results show that this uniform allocation is suboptimal for all-in-one restoration, as it fails to leverage task-specific commonalities and distinctions effectively. Tab. 6 highlights the limitations of standard" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.284, + 0.907, + 0.376 + ], + "angle": 0, + "content": "MoEs [43], where fixed post-training expert assignments result in uniform, task-agnostic performance, offering only moderate cross-task generalization. In contrast, MoCEs effectively generalize across degradations while exhibiting task-discriminative behavior, routing inputs to the most suitable experts for optimal performance." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.389, + 0.869, + 0.405 + ], + "angle": 0, + "content": "5. Conclusion, Limitations and Discussion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.414, + 0.907, + 0.534 + ], + "angle": 0, + "content": "This paper presents MoCE-IR, an efficient all-in-one image restoration model leveraging complexity experts, specialized units with varying computational demands and receptive fields. By aligning degradation tasks with suitable resources, MoCE-IR overcomes limitations of prior methods. Its built-in bias toward simpler paths enhances task-discriminative learning, enabling swift inference by bypassing irrelevant experts without sacrificing restoration quality." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.535, + 0.907, + 0.61 + ], + "angle": 0, + "content": "Extensive experiments across diverse degradation settings, including more demanding scenarios with multiple composited degradations, demonstrate our computational efficiency through selective expert utilization and consistent improvement over state-of-the-art methods." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.61, + 0.909, + 0.763 + ], + "angle": 0, + "content": "While MoCE-IR shows substantial improvements, opportunities remain. The current image-level routing imposes scalability constraints. Future work could explore SoftMoE [37] architectures for fine-grained spatial routing or early-exiting approaches [39] to further enhance processing speeds and input adaptability. Extending to synthetic-to-real adaptation, especially for video, represents a crucial next step. Performance optimization through mixed-precision computation across experts could further enhance speed and efficiency for resource-limited deployments." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Acknowledgements. The authors sincerely thank the reviewers and all members of the program committee for their tremendous efforts and incisive feedback. This research was supported by the Alexander von Humboldt Foundation, the Ministry of Education and Science of Bulgaria (support for INSAIT, under the Bulgarian National Roadmap for Research Infrastructure), the Shanghai Municipal Science and Technology Major Project (2021SHZDZX0102), and the Fundamental Research Funds for the Central Universities." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "12760" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.486, + 0.17 + ], + "angle": 0, + "content": "[1] Yuang Ai, Huaibo Huang, Xiaogiang Zhou, Jiexiang Wang, and Ran He. Multimodal prompt perceiver: Empower adaptiveness generalizability and fidelity for all-in-one image restoration. In CVPR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.171, + 0.485, + 0.212 + ], + "angle": 0, + "content": "[2] Pablo Arbelaez, Michael Maire, Charless Fowlkes, and Jitendra Malik. Contour detection and hierarchical image segmentation. TPAMI, 33(5):898-916, 2010. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.213, + 0.484, + 0.268 + ], + "angle": 0, + "content": "[3] Yoshua Bengio, Nicholas Léonard, and Aaron Courville. Estimating or propagating gradients through stochastic neurons for conditional computation. arXiv preprint arXiv:1308.3432, 2013. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.269, + 0.484, + 0.297 + ], + "angle": 0, + "content": "[4] Dana Berman and Shai Avidan. Non-local image dehazing. In CVPR, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.298, + 0.484, + 0.34 + ], + "angle": 0, + "content": "[5] Bolun Cai, Xiangmin Xu, Kui Jia, Chunmei Qing, and Dacheng Tao. Dehazenet: An end-to-end system for single image haze removal. TIP, 25(11):5187-5198, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.341, + 0.484, + 0.38 + ], + "angle": 0, + "content": "[6] Liangyu Chen, Xiaojie Chu, Xiangyu Zhang, and Jian Sun. Simple baselines for image restoration. In ECCV, 2022. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.382, + 0.484, + 0.424 + ], + "angle": 0, + "content": "[7] Yu-Wei Chen and Soo-Chang Pei. Always clear days: Degradation type and severity aware all-in-one adverse weather removal. arXiv preprint arXiv:2310.18293, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.425, + 0.484, + 0.465 + ], + "angle": 0, + "content": "[8] Zheng Chen, Yulun Zhang, Jinjin Gu, Linghe Kong, Xin Yuan, et al. Cross aggregation transformer for image restoration. NeurIPS, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.467, + 0.484, + 0.508 + ], + "angle": 0, + "content": "[9] Marcos V Conde, Gregor Geigle, and Radu Timofte. Instructir: High-quality image restoration following human instructions. In ECCV, 2024. 1, 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.509, + 0.484, + 0.562 + ], + "angle": 0, + "content": "[10] Xin Dang, Hai Wang, Jie Ren, and Le Chen. An application performance optimization model of mobile augmented reality based on hd restoration. In CBD, pages 201-206. IEEE, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.565, + 0.484, + 0.607 + ], + "angle": 0, + "content": "[11] Yu Dong, Yihao Liu, He Zhang, Shifeng Chen, and Yu Qiao. Fd-gan: Generative adversarial networks with fusion-discriminator for single image dehazing. In AAAI, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.608, + 0.484, + 0.648 + ], + "angle": 0, + "content": "[12] Huiyu Duan, Xiongkuo Min, Sijing Wu, Wei Shen, and Guangtao Zhai. Uniprocessor: A text-induced unified low-level image processor. In ECCV, 2024. 1, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.65, + 0.484, + 0.704 + ], + "angle": 0, + "content": "[13] Qingnan Fan, Dongdong Chen, Lu Yuan, Gang Hua, Nenghai Yu, and Baoquan Chen. A general decoupled learning framework for parameterized image operators. TPAMI, 43(1): 33-47, 2019. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.705, + 0.484, + 0.747 + ], + "angle": 0, + "content": "[14] Hongyun Gao, Xin Tao, Xiaoyong Shen, and Jiaya Jia. Dynamic scene deblurring with parameter selective sharing and nested skip connections. In CVPR, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.748, + 0.484, + 0.815 + ], + "angle": 0, + "content": "[15] Florin Girbacia, Silviu Butnariu, Alex Petre Orman, and Cristian Cezar Postelnicu. Virtual restoration of deteriorated religious heritage objects using augmented reality technologies. European Journal of Science and Theology, 9(2):223-231, 2013. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.817, + 0.484, + 0.859 + ], + "angle": 0, + "content": "[16] Hang Guo, Jinmin Li, Tao Dai, Zhihao Ouyang, Xudong Ren, and Shu-Tao Xia. Mambair: A simple baseline for image restoration with state-space model. In ECCV, 2024. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.86, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[17] Yu Guo, Yuan Gao, Yuxu Lu, Ryan Wen Liu, and Shengfeng He. Onerestore: A universal restoration framework for composite degradation. In ECCV, 2024. 5, 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.486, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.147 + ], + "angle": 0, + "content": "[18] Kui Jiang, Zhongyuan Wang, Peng Yi, Chen Chen, Baojin Huang, Yimin Luo, Jiayi Ma, and Junjun Jiang. Multi-scale progressive fusion network for single image deraining. In CVPR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.15, + 0.906, + 0.192 + ], + "angle": 0, + "content": "[19] Yitong Jiang, Zhaoyang Zhang, Tianfan Xue, and Jinwei Gu. Autodir: Automatic all-in-one image restoration with latent diffusion. arXiv preprint arXiv:2310.10123, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.194, + 0.906, + 0.222 + ], + "angle": 0, + "content": "[20] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.224, + 0.908, + 0.278 + ], + "angle": 0, + "content": "[21] Lingshun Kong, Jiangxin Dong, Jianjun Ge, Mingqiang Li, and Jinshan Pan. Efficient frequency domain-based transformers for high-quality image deblurring. In CVPR, 2023. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.281, + 0.908, + 0.337 + ], + "angle": 0, + "content": "[22] Jaakko Lehtinen, Jacob Munkberg, Jon Hasselgren, Samuli Laine, Tero Karras, Miika Aittala, and Timo Aila. Noise2noise: Learning image restoration without clean data. arXiv preprint arXiv:1803.04189, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.339, + 0.908, + 0.381 + ], + "angle": 0, + "content": "[23] Boyi Li, Wenqi Ren, Dengpan Fu, Dacheng Tao, Dan Feng, Wenjun Zeng, and Zhangyang Wang. Benchmarking single-image dehazing and beyond. TIP, 28(1):492-505, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.383, + 0.906, + 0.424 + ], + "angle": 0, + "content": "[24] Boyun Li, Xiao Liu, Peng Hu, Zhongqin Wu, Jiancheng Lv, and Xi Peng. All-In-One Image Restoration for Unknown Corruption. In CVPR, 2022. 1, 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.426, + 0.906, + 0.48 + ], + "angle": 0, + "content": "[25] Zilong Li, Yiming Lei, Chenglong Ma, Junping Zhang, and Hongming Shan. Prompt-in-prompt learning for universal image restoration. arXiv preprint arXiv:2312.05038, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.484, + 0.906, + 0.525 + ], + "angle": 0, + "content": "[26] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Swinir: Image restoration using swim transformer. In ICCV, 2021. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.528, + 0.908, + 0.582 + ], + "angle": 0, + "content": "[27] Lin Liu, Lingxi Xie, Xiaopeng Zhang, Shanxin Yuan, Xiangyu Chen, Wengang Zhou, Houqiang Li, and Qi Tian. Tape: Task-agnostic prior embedding for image restoration. In ECCV, 2022. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.585, + 0.906, + 0.64 + ], + "angle": 0, + "content": "[28] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.643, + 0.906, + 0.713 + ], + "angle": 0, + "content": "[29] Yulin Luo, Rui Zhao, Xiaobao Wei, Jinwei Chen, Yijie Lu, Shenghao Xie, Tianyu Wang, Ruiqin Xiong, Ming Lu, and Shanghang Zhang. Wm-moe: Weather-aware multi-scale mixture-of-experts for blind adverse weather removal. arXiv preprint arXiv:2303.13739, 2023. 2, 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.714, + 0.906, + 0.756 + ], + "angle": 0, + "content": "[30] Ziwei Luo, Fredrik K Gustafsson, Zheng Zhao, Jens Sjolund, and Thomas B Schon. Controlling vision-language models for multi-task image restoration. In ICLR, 2024. 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.758, + 0.906, + 0.813 + ], + "angle": 0, + "content": "[31] Kede Ma, Zhengfang Duanmu, Qingbo Wu, Zhou Wang, Hongwei Yong, Hongliang Li, and Lei Zhang. Waterloo exploration database: New challenges for image quality assessment models. TIP, 26(2):1004-1016, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.816, + 0.906, + 0.871 + ], + "angle": 0, + "content": "[32] David Martin, Charless Fowlkes, Doron Tal, and Jitendra Malik. A database of human segmented natural images and its application to evaluating segmentation algorithms and measuring ecological statistics. In ICCV, 2001. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[33] Chong Mou, Qian Wang, and Jian Zhang. Deep generalized unfolding networks for image restoration. In CVPR, 2022. 5" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "12761" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.134 + ], + "angle": 0, + "content": "[34] Seungjun Nah, Tae Hyun Kim, and Kyoung Mu Lee. Deep multi-scale convolutional neural network for dynamic scene deblurring. In CVPR, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.484, + 0.176 + ], + "angle": 0, + "content": "[35] Jinshan Pan, Deqing Sun, Hans Pfister, and Ming-Hsuan Yang. Blind image deblurring using dark channel prior. In CVPR, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.179, + 0.482, + 0.22 + ], + "angle": 0, + "content": "[36] Vaishnav Potlapalli, Syed Waqas Zamir, Salman Khan, and Fahad Khan. Prompt: Prompting for all-in-one image restoration. In NeurIPS, 2023. 1, 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.222, + 0.482, + 0.262 + ], + "angle": 0, + "content": "[37] Joan Puigcerver, Carlos Riquelme Ruiz, Basil Mustafa, and Neil Houlsby. From sparse to soft mixtures of experts. In ICLR, 2024. 2, 3, 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.264, + 0.484, + 0.292 + ], + "angle": 0, + "content": "[38] Yanyun Qu, Yizi Chen, Jingying Huang, and Yuan Xie. Enhanced pix2pix dehazing network. In CVPR, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.294, + 0.484, + 0.347 + ], + "angle": 0, + "content": "[39] David Raposo, Sam Ritter, Blake Richards, Timothy Lillicrap, Peter Conway Humphreys, and Adam Santoro. Mixture-of-depths: Dynamically allocating compute in transformer-based language models, 2024. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.35, + 0.484, + 0.39 + ], + "angle": 0, + "content": "[40] Dongwei Ren, Wangmeng Zuo, Qinghua Hu, Pengfei Zhu, and Deyu Meng. Progressive image deraining networks: A better and simpler baseline. In CVPR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.392, + 0.482, + 0.433 + ], + "angle": 0, + "content": "[41] Wenqi Ren, Lin Ma, Jiawei Zhang, Jinshan Pan, Xiaochun Cao, Wei Liu, and Ming-Hsuan Yang. Gated fusion network for single image dehazing. In CVPR, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.435, + 0.484, + 0.489 + ], + "angle": 0, + "content": "[42] Wenqi Ren, Jinshan Pan, Hua Zhang, Xiaochun Cao, and Ming-Hsuan Yang. Single image dehazing via multi-scale convolutional neural networks with holistic edges. In ICCV, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.491, + 0.482, + 0.546 + ], + "angle": 0, + "content": "[43] Carlos Riquelme, Joan Puigcerver, Basil Mustafa, Maxim Neumann, Rodolphe Jenatton, André Susano Pinto, Daniel Keysers, and Neil Houlsby. Scaling vision with sparse mixture of experts. *NeurIPS*, 2021. 2, 3, 4, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.548, + 0.484, + 0.589 + ], + "angle": 0, + "content": "[44] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-Net: Convolutional networks for biomedical image segmentation. In MICCAI, pages 234–241. Springer, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.591, + 0.482, + 0.646 + ], + "angle": 0, + "content": "[45] Giovanni Saggio, Davide Borra, et al. Augmented reality for restoration/reconstruction of artefacts with artistic or historical value. In Augmented reality: some emerging application areas, pages 59-86. InTech Publications, 2011. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.647, + 0.484, + 0.702 + ], + "angle": 0, + "content": "[46] Noam Shazeer, Azalia Mirhoseini, Krzysztof Maziarz, Andy Davis, Quoc Le, Geoffrey Hinton, and Jeff Dean. Outrageously large neural networks: The sparsely-gated mixture-of-experts layer, 2017. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.704, + 0.484, + 0.744 + ], + "angle": 0, + "content": "[47] Ying Tai, Jian Yang, Xiaoming Liu, and Chunyan Xu. Memnet: A persistent memory network for image restoration. In ICCV, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.747, + 0.484, + 0.787 + ], + "angle": 0, + "content": "[48] Chunwei Tian, Yong Xu, and Wangmeng Zuo. Image denoising using deep cnn with batch renormalization. NN, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.789, + 0.484, + 0.843 + ], + "angle": 0, + "content": "[49] Jeya Maria Jose Valanarasu, Rajeev Yasarla, and Vishal M Patel. Transweather: Transformer-based restoration of images degraded by adverse weather conditions. In CVPR, 2022. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[50] Cong Wang, Jinshan Pan, Wei Wang, Jiangxin Dong, Mengzhu Wang, Yakun Ju, and Junyang Chen. Promptstorer: A prompting image restoration method with degradation perception. In NeurIPS, 2023. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.907, + 0.161 + ], + "angle": 0, + "content": "[51] Tao Wang, Kaihao Zhang, Ziqian Shao, Wenhan Luo, Bjorn Stenger, Tong Lu, Tae-Kyun Kim, Wei Liu, and Hongdong Li. Gridformer: Residual dense transformer with grid structure for image restoration in adverse weather conditions. ICCV, 2024. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.164, + 0.907, + 0.218 + ], + "angle": 0, + "content": "[52] Zhendong Wang, Xiaodong Cun, Jianmin Bao, Wengang Zhou, Jianzhuang Liu, and Houqiang Li. Uformer: A general u-shaped transformer for image restoration. In CVPR, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.221, + 0.907, + 0.261 + ], + "angle": 0, + "content": "[53] Chen Wei, Wenjing Wang, Wenhan Yang, and Jiaying Liu. Deep retina decomposition for low-light enhancement. arXiv preprint arXiv:1808.04560, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.263, + 0.907, + 0.303 + ], + "angle": 0, + "content": "[54] Gang Wu, Junjun Jiang, Kui Jiang, and Xianming Liu. Harmony in diversity: Improving all-in-one image restoration via multi-task collaboration. In ACMMM, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.306, + 0.907, + 0.358 + ], + "angle": 0, + "content": "[55] Haiyan Wu, Yanyun Qu, Shaohui Lin, Jian Zhou, Ruizhi Qiao, Zhizhong Zhang, Yuan Xie, and Lizhuang Ma. Contrastive learning for compact single image dehazing. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.362, + 0.907, + 0.403 + ], + "angle": 0, + "content": "[56] Fuzhi Yang, Huan Yang, Jianlong Fu, Hongtao Lu, and Baining Guo. Learning texture transformer network for image super-resolution. In CVPR, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.406, + 0.907, + 0.446 + ], + "angle": 0, + "content": "[57] Hao Yang, Liyuan Pan, Yan Yang, and Wei Liang. Language-driven all-in-one adverse weather removal. In CVPR, 2024. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.449, + 0.907, + 0.502 + ], + "angle": 0, + "content": "[58] Zhiwen Yang, Haowei Chen, Ziniu Qian, Yang Yi, Hui Zhang, Dan Zhao, Bingzheng Wei, and Yan Xu. All-in-one medical image restoration via task-adaptive routing. In MICCAI, 2024. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.504, + 0.907, + 0.546 + ], + "angle": 0, + "content": "[59] Rajeev Yasarla and Vishal M Patel. Uncertainty guided multiscale residual learning-using a cycle spinning cnn for single image de-raining. In CVPR, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.548, + 0.907, + 0.588 + ], + "angle": 0, + "content": "[60] Ke Yu, Xintao Wang, Chao Dong, Xiaou Tang, and Chen Change Loy. Path-restore: Learning network path selection for image restoration. TPAMI, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.591, + 0.907, + 0.631 + ], + "angle": 0, + "content": "[61] Xiaoyan Yu, Shen Zhou, Huafeng Li, and Liehuang Zhu. Multi-expert adaptive selection: Task-balancing for all-in-one image restoration. arXiv preprint arXiv:2407.19139, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.634, + 0.907, + 0.687 + ], + "angle": 0, + "content": "[62] Eduard Zamfir, Zongwei Wu, Nancy Mehta, Danda Pani Paudel, Yulun Zhang, and Radu Timofte. Efficient degradation-aware any image restoration. arXiv preprint arXiv:2405.15475, 2024. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.689, + 0.907, + 0.731 + ], + "angle": 0, + "content": "[63] Eduard Zamfir, Zongwei Wu, Nancy Mehta, Yulun Zhang, and Radu Timofte. See more details: Efficient image superresolution by experts mining. In ICML. PMLR, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.733, + 0.907, + 0.787 + ], + "angle": 0, + "content": "[64] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, Ming-Hsuan Yang, and Ling Shao. Multi-stage progressive image restoration. In CVPR, 2021. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.789, + 0.907, + 0.843 + ], + "angle": 0, + "content": "[65] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. In CVPR, 2022. 1, 2, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.907, + 0.9 + ], + "angle": 0, + "content": "[66] Jinghao Zhang, Jie Huang, Mingde Yao, Zizheng Yang, Hu Yu, Man Zhou, and Feng Zhao. Ingredient-oriented multidegradation learning for image restoration. In CVPR, 2023. 1, 2, 5, 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.907, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "12762" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.134 + ], + "angle": 0, + "content": "[67] Kai Zhang, Wangmeng Zuo, Shuhang Gu, and Lei Zhang. Learning deep cnn denoiser prior for image restoration. In CVPR, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.485, + 0.178 + ], + "angle": 0, + "content": "[68] Yulun Zhang, Kunpeng Li, Kai Li, Bineng Zhong, and Yun Fu. Residual non-local attention networks for image restoration. arXiv preprint arXiv:1903.10082, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.179, + 0.485, + 0.233 + ], + "angle": 0, + "content": "[69] Haiyu Zhao, Yuanbiao Gou, Boyun Li, Dezhong Peng, Jiancheng Lv, and Xi Peng. Comprehensive and delicate: An efficient transformer for image restoration. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.235, + 0.485, + 0.289 + ], + "angle": 0, + "content": "[70] Yanqi Zhou, Tao Lei, Hanxiao Liu, Nan Du, Yanping Huang, Vincent Zhao, Andrew M Dai, Quoc V Le, James Laudon, et al. Mixture-of-experts with expert choice routing. NeurIPS, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.291, + 0.485, + 0.347 + ], + "angle": 0, + "content": "[71] Yurui Zhu, Tianyu Wang, Xueyang Fu, Xuanyu Yang, Xin Guo, Jifeng Dai, Yu Qiao, and Xiaowei Hu. Learning weather-general and weather-specific features for image restoration under multiple adverse weather conditions. In CVPR, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.349, + 0.485, + 0.389 + ], + "angle": 0, + "content": "[72] Ozan Özdenizci and Robert Legenstein. Restoring vision in adverse weather conditions with patch-based denoising diffusion models. TPAMI, 2023. 6" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.389 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "12763" + } + ] +] \ No newline at end of file diff --git a/2025/Complexity Experts are Task-Discriminative Learners for Any Image Restoration/811da5ba-4d1d-41b9-ac9b-e649712870d3_origin.pdf b/2025/Complexity Experts are Task-Discriminative Learners for Any Image Restoration/811da5ba-4d1d-41b9-ac9b-e649712870d3_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..81e903502a6d82885d7a2df8dc8f2966f98deb69 --- /dev/null +++ b/2025/Complexity Experts are Task-Discriminative Learners for Any Image Restoration/811da5ba-4d1d-41b9-ac9b-e649712870d3_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ed6e7c31ed9544858da65744ce488d6333e0eb311ce7f1511b714cfca2a137b6 +size 2595223 diff --git a/2025/Complexity Experts are Task-Discriminative Learners for Any Image Restoration/full.md b/2025/Complexity Experts are Task-Discriminative Learners for Any Image Restoration/full.md new file mode 100644 index 0000000000000000000000000000000000000000..af40af1ef90bf435f33a367d31b95569ada31c02 --- /dev/null +++ b/2025/Complexity Experts are Task-Discriminative Learners for Any Image Restoration/full.md @@ -0,0 +1,296 @@ +# Complexity Experts are Task-Discriminative Learners for Any Image Restoration + +Eduard Zamfir1 Zongwei Wu1* Nancy Mehta1 Yuedong Tan1 Danda Pani Paudel2 Yulun Zhang3 Radu Timofte1 + +1Computer Vision Lab, CAIDAS & IFI, University of Würzburg + +2INSAIT, Sofia University "St. Kliment Ohridski" 3Shanghai Jiao Tong University + +# Abstract + +Recent advancements in all-in-one image restoration models have revolutionized the ability to address diverse degradations through a unified framework. However, parameters tied to specific tasks often remain inactive for other tasks, making mixture-of-experts (MoE) architectures a natural extension. Despite this, MoEs often show inconsistent behavior, with some experts unexpectedly generalizing across tasks while others struggle within their intended scope. This hinders leveraging MoEs' computational benefits by bypassing irrelevant experts during inference. We attribute this undesired behavior to the uniform and rigid architecture of traditional MoEs. To address this, we introduce "complexity experts" - flexible expert blocks with varying computational complexity and receptive fields. A key challenge is assigning tasks to each expert, as degradation complexity is unknown in advance. Thus, we execute tasks with a simple bias toward lower complexity. To our surprise, this preference effectively drives task-specific allocation, assigning tasks to experts with the appropriate complexity. Extensive experiments validate our approach, demonstrating the ability to bypass irrelevant experts during inference while maintaining superior performance. The proposed MoCE-IR model outperforms state-of-the-art methods, affirming its efficiency and practical applicability. The source code and models are publicly available at eduardzamfir.github.io/MoCE-IR/ + +# 1. Introduction + +Image restoration [47, 67, 68] is a fundamental problem in computer vision, dealing with reconstructing high-quality images from deteriorated observations. Adverse conditions such as noise [22, 67], haze [5, 38], or rain [7, 59] significantly impact the practical utility of images in downstream applications across various domains, including autonomous navigation [7, 49] or augmented reality [10, 15, 45]. Deep learning-based approaches have remarkably advanced this field, particularly for task-specific image restoration prob + +![](images/9ca792966f6d0aa5f464e7d8e451aceaf073605373aab8e6a655da80f6d992b6.jpg) +(a) Dense Model + +![](images/69324a597e6cf8a2befc7729dad123701ab939bdfff45e37303cf1462c07513c.jpg) +(b) MoE + +![](images/9d9df16144abeaa5ca33f19b160da4030b5c8dc685c5800b445a29032c84e6d8.jpg) +(c) Complexity Experts (ours) + +![](images/b46b5d1a1584d2df93afcce3f9e2ab92f068210dd752173968389d617f0a7802.jpg) +Figure 1. Motivation. (a) Dense all-in-one restoration methods [24, 57] often inefficiently allocate parameters when handling multiple degradation types. (b) While recent Mixture-of-Experts (MoE) approaches [61, 62] address this through sparse computation, their rigid routing mechanisms uniformly distribute inputs across experts without considering the natural relationships between degradations. (c) To overcome these limitations, we introduce Complexity Experts - adaptive processing blocks with size-varying computational units. Our framework dynamically allocates model capacity using a spring-inspired force mechanism that continuously guides routing decisions toward simpler experts when possible, with the force proportional to the complexity of the input degradation. While initially designed for computational efficiency, this approach naturally emerges as a task-discriminative learning framework, assigning degradations to the most suitable experts. This makes it particularly effective for all-in-one restoration methods, where both task-specific processing and cross-degradation knowledge sharing are crucial. + +lems [6, 8, 16, 28, 52, 65]. Recent all-in-onerestoration models [9, 12, 13, 24, 27, 36, 49, 66], however, demonstrate the possibility of handling multiple degradation types within a single model, offering more practical solutions compared to traditional task-specific approaches without the extensive re-training need for each new degradation type. Notable works employ visual [25, 36, 50] or language-based prompt- + +ing [1, 9, 30], contrastive learning [24, 66], and diffusion-based models [1, 30]. + +Despite their success, we observe that the aforementioned models often suffer from inefficiencies, as parameters tied to specific degradations often remain inactive or underutilized when addressing unrelated tasks [57]. This naturally suggests the use of mixture-of-experts (MoE) architectures for task-specific processing. Yet, current MoE-based approaches [29, 57, 62] typically incorporate routing mechanisms based on language [57] or degradation priors [29, 62], leading to imbalanced optimization, where some experts generalize well, while others struggle with their intended tasks. This inconsistency limits the potential computational benefits of bypassing task-irrelevant experts during inference. + +We attribute this undesired behavior to two limitations in current approaches. First, the uniform and rigid architecture of existing MoE models fails to account for the varying complexity requirements across different restoration tasks. For instance, motion blur demands localized processing with strong spatial awareness, while haze removal requires broader contextual understanding [4, 35], necessitating adaptive processing aligned to the task requirements. Second, the challenge of appropriately routing tasks to experts is complicated by the unknown complexity of each degradation type a priori. Typically, MoE models [37, 39, 43, 70] aim to balance expert utilization, preventing the model from collapsing into a single-expert dependency. + +In this work, we introduce a novel mixture-of-complexity-experts (MoCE) framework for all-in-one image restoration that directly addresses these limitations. Our key innovation lies in designing expert blocks with increasing computational complexity and receptive fields, allowing the model to adaptively match processing capacity with task requirements. + +To address the routing challenge, we introduce a complexity-aware allocation mechanism that preferentially directs tasks to lower-complexity experts. This approach implements a complexity-proportional bias term, analogous to a mechanical spring force, to guide expert selection. Surprisingly, this straightforward approach results in effective task-specific allocation, naturally directing inputs to experts with appropriate complexity levels. Our MoCE framework offers two significant advantages: it ensures efficient inference by selectively bypassing irrelevant experts, thereby reducing computational overhead, while simultaneously maintaining high-quality restoration performance across diverse tasks. Through extensive experiments across multiple image restoration tasks, we demonstrate that our approach not only preserves restoration quality but surpasses recent state-of-the-art methods. These results validate both the effectiveness of our complexity-based expert design and the practicality of our routing strategy for real-world applications. Our approach establishes a new benchmark for all-in-one image restoration, enhancing both efficiency and fidelity. + +The principal contributions of this work are threefold: + +- We propose MoCE-IR, achieving SoTA all-in-one image restoration performance with improved efficiency compared to prior works. +- Our novel MoE layer selectively activates complexity experts based on input requirements, unifying task-specific and holistic learning in a single architecture. +- We develop a complexity-aware routing mechanism that balances restoration quality with computational efficiency by adaptive expert allocation. + +# 2. Related Works + +Task Specific Image Restoration. Reconstructing the clean image from its degraded counterpart is a highly ill-posed problem, however, a great body of work have addressed image restoration from a data-driven learning perspective, achieving tremendous results compared to prior hand-crafted methods [22, 26, 47, 52, 65, 67, 68]. Most proposed solutions build on convolutional [6, 47, 67, 68] or Transformer-based architectures [8, 28, 52, 65] addressing single degradation tasks, such as denoising [8, 67, 68], dehazing [41, 42, 55] or deraining [18, 40]. Contrary to CNN-based networks, Transformer offer strong modeling capabilities for capturing global dependencies, which makes them outstanding image restorers [28, 65, 69]. Self-attention's quadratic complexity w.r.t the image size poses a challenge for resource-constrained applications. Conversely, convolutions offer fast and efficient processing with limited global context, but they scale more effectively with increasing input size. Recently, FFTformer [21] introduced an efficient approximation of query and key interactions in the Fourier domain, significantly reducing computational demands compared to window-based attention methods for image deblurring. Building on this, we incorporate specialized experts into a Transformer-based architecture [65] to efficiently capture both shared and distinct contextual information, addressing the diverse demands of image restoration tasks. + +All-in-One Image Restoration. Restoring degraded images often requires multiple models, complicating practical implementation, especially when images exhibit a combination of various degradations. An emerging field known as all-in-one image restoration is advancing in low-level computer vision, utilizing a single deep blind restoration model to tackle multiple degradation types simultaneously [7, 19, 36, 49, 64, 66]. The seminal work, AirNet [24] achieves blind all-in-one image restoration using contrastive learning to extract degradation representations, which guide the restoration process. Similarly, IDR [66] employs a meta-learning-based two-stage approach to model degradation through underlying physical principles. Prompt-based learning [25, 36, 50] has also gained traction, with [36] introducing tunable prompts to encode degradation-specific in + +![](images/d89bfebbc3610f17c2faf0b54f62894d3210c7910aa4df55795ffce9cf632e7e.jpg) +(a) Architecture overview + +![](images/f3471c152ca5a81a0179dd375026f12638ffa2b774dd409c7c7378e6a64d89b0.jpg) +(b) Transformer blocks. + +![](images/a849b69f4a55e7e65565adb2930570db71fcfea3615872b33bde60103a703f26.jpg) +(c) Mixture-of-complexity-experts +Figure 2. Proposed MoCE-IR framework. Despite recent advances in MoE-based image restoration [57, 62, 63], inconsistent expert behavior—where some experts over-generalize while others underperform—limits their computational efficiency. We address this through complexity experts: flexible blocks with varying computational capacity and receptive fields. Our MoCE-IR employs an asymmetric encoder-decoder architecture where each decoder block contains a mixture-of-complexity-experts layer for adaptive capacity routing. + +formation, albeit with a high parameter count. With recent advancements in multimodal integration, such as combining images and language, there has been growing interest in language-guided image restoration. Notable approaches include DA-CLIP [30], InstructIR [9], and UniProcessor [12]. However, these methods often depend on high-quality text prompts and typically require LLMs, limiting their applicability on resource-constrained devices. In contrast, our contribution lies in a novel, parameter-efficient approach that focuses on selective activation, effectively addressing diverse degradations while remaining lightweight and adaptable. + +Dynamic Networks. Dynamic networks have evolved from basic conditional computation [3] to sophisticated Mixture-of-Experts (MoE) architectures that expand model capacity while maintaining efficient inference costs [37, 43, 46]. This approach has proven particularly effective in transformer-based NLP and high-level vision tasks. In image restoration, Path-Restore [60] introduced content-aware patch routing with difficulty-regulated rewards, while recent all-in-one models [29, 57, 62] leverage various priors for expert routing. Our work addresses the challenge of adaptively routing corrupted images to appropriate computational units within a generalist restoration framework. By dynamically aligning image characteristics with computational requirements, we achieve a unified model in both parameter and architectural space, advancing beyond previous limitations. + +# 3. Method + +In this section, we outline the core principles of our All-in-One image restoration method. As shown in Fig. 2a, our pipeline features a U-shaped architecture [44, 65] with an asymmetric encoder-decoder design. A $3 \times 3$ convolution + +first extracts shallow features from the degraded input, which then pass through 4 levels of encoding and decoding stages. While both stages utilize Transformer blocks [36, 65], our novel MoCE layers are exclusively integrated into the decoder blocks (Fig. 2b). To further enhance the decoder's feature enrichment capabilities, we introduce high-frequency guidance via Sobel-filtered global feature vector to promote the frequency awareness of the gating functions. Finally, a global residual connection links the shallow features to the output of the refinement stage, refining the crucial high-frequency details before producing the restored image. + +# 3.1. Mixture-of-Complexity-Experts + +To overcome the limitations of uniform MoE architectures in image restoration, we propose an efficient input-adaptive model for all-in-one image restoration that dynamically adjusts processing capacity to match task requirements. + +Technically, at the core of our framework is a specialized MoCE layer, consisting of $n$ complexity experts $\mathbf{E}$ and a single shared expert $\mathbf{S}$ . As shown in Fig. 2b, our experts are designed with increasing computational complexity and progressively larger receptive fields, in contrast to the uniform architecture of traditional MoE designs [37, 43, 46]. The interaction between these components is orchestrated through a two-level gating mechanism: first projecting routed tokens into expert-specific embeddings, then combining specialized and agnostic features through modulation, and finally merging the outputs via cross-attention. This enables our model to capture both degradation-specific features and interdegradation relationships, while maintaining a bias toward computationally efficient processing paths. + +Expert Design. Each complexity expert block as shown in Fig. 2c is designed based on two fundamental principles: + +ensuring computational efficiency and capturing hierarchical spatial features, which are critical for restoring fine details and addressing diverse degradation patterns. Given the growing computational demands associated with increasing expert numbers and capacities, it is crucial to prioritize an overall efficient expert design. To this end, we implement a nested expert structure, progressively reducing channel dimensionality $r$ within each subsequent expert to control the computational overhead. Simultaneously, we increase the receptive field, i.e. window partition size $w$ , within each expert to adaptively balance the localized and global processing, tailored to the specific requirements of the input degradation. Thus, our complexity experts are constructed by scaling the design along two key axes: window partition size and channel dimension. + +More specifically, each expert $\mathbf{E}$ projects its input tokens to the embedding $\mathbf{x}_E\in \mathbb{R}^{H\times W\times r}$ , where $r = C / 2^i$ for $i\in \{i,\dots,n\}$ , to highlight the most relevant information along the channel dimension. Following this, a window-based self-attention (WSA) mechanism captures spatial information effectively. To optimize performance, we employ an FFT-based approximation [21] for efficient matrix multiplication between queries $Q$ and keys $K$ in the Fourier domain. The refined features are then transformed back into the spatial domain and re-projected to their original dimensionality. Lastly, the output of the expert block $\hat{\mathbf{x}}_E^i$ is modulated through element-wise multiplication with the holistic features from the shared expert, S. As shared expert, we employ the transposed self-attention (T-SA) module [65] in the channel dimension, while the complexity experts operate spatially. More concretely, the modulated expert features $\mathbf{y}_{\mathbf{E}}^{\mathrm{i}}$ are obtained as following: + +$$ +\mathbf {y} _ {\mathbf {E}} ^ {i} = \hat {\mathbf {x}} _ {E} ^ {i} \odot \mathbf {S} (\mathbf {x}) \tag {1} +$$ + +$$ +\text {w i t h} \hat {\mathbf {x}} _ {E} ^ {i} = \mathbf {E} ^ {i} \left(\mathbf {x} _ {E} ^ {i}\right) = \operatorname {F F T - W S A} _ {\mathrm {W} _ {i}} \left(\mathbf {W} _ {C \rightarrow r _ {i}} ^ {i} \mathbf {x} _ {E} ^ {i}\right) \tag {2} +$$ + +$$ +\operatorname {a n d} \mathbf {S} (\mathbf {x}) = \mathrm {T} - \operatorname {S A} (\mathbf {x}) \tag {3} +$$ + +where linear layers for projection are denoted as $\mathbf{W}$ , $\odot$ denotes element-wise multiplication and $i$ is the index of the current complexity expert $\mathbf{E}^i$ with window size $w_i$ and embedding dimensionality $r_i$ . For example, the most lightweight expert is characterized by the smallest embedding dimension $r_1$ and window size $w_1$ , while the dimensions and window partitions for the other experts can increase linearly. Throughout our network, the capacity of the shared expert remains constant, with $r = C$ . + +# 3.2. Complexity-aware Routing + +Drawing inspiration from sparse MoE [43, 46], we integrate linear layers within each decoder block to enable a routing mechanism, associating input features $\mathbf{x} \in \mathbb{R}^{H \times W \times C}$ with their corresponding specialized complexity experts $\mathbf{E}$ . However, in the context of image restoration, a key challenge + +arises in achieving a scale-invariant tokenization of input images, ensuring consistency across varying resolutions and scales. While previous MoE approaches [37, 43, 46] employ a token-based routing, we opt for an image-level routing strategy instead, where we select experts for entire input image. The routing function $g(\mathbf{x})$ orchestrates the allocation of input samples based on the required compute to corresponding complexity experts $\mathbf{E}^i$ , indexed by $i \in 1, \dots, n$ with $n$ representing the total number of experts. Within $g(\mathbf{x})$ , we select the top- $k$ elements of the softmax distribution associating input tokens to experts, setting all other elements to zero. In practice, we use $k = 1$ while $\epsilon$ is sampled independently $\epsilon \sim \mathcal{N}(0, 1/n^2)$ entry-wise enabling noisy top-1 routing: + +$$ +g (\mathbf {x}) = \operatorname {t o p} _ {k} \left(\operatorname {S o f t m a x} \left(\mathbf {W} \mathbf {x} + \epsilon\right)\right) \tag {4} +$$ + +During training, we incorporate an auxiliary loss $\mathcal{L}_{\mathrm{aux}}$ inspired by the load balancing loss proposed by Riquelme et al. [43]. We complement the importance term to favor experts based on their computational complexity, allowing for more effective balancing aligned with expert capacity. As in [43], the balanced expert utilization is encouraged through an importance loss, where the importance of each expert $\mathbf{E}^i$ for a batch of images is defined as the sum of the routing weights associated with the $i$ -th expert across the entire batch. Additionally, we introduce a complexity bias $\mathbf{b}$ by calculating the number of learnable parameters $p_i$ in each expert $\mathbf{E}^i$ and normalizing these values with respect to the largest parameter count of the experts, $p_{\mathrm{max}}$ : + +$$ +\operatorname {I m p} _ {i} (\mathbf {x}) = \left(\sum_ {x \in \mathbf {x}} \operatorname {S o f t m a x} (W x) _ {i}\right) * \mathbf {b}, \tag {5} +$$ + +$$ +\text {w i t h} \mathbf {b} = \left[ \frac {p _ {1}}{p _ {\max }}, \frac {p _ {2}}{p _ {\max }}, \dots , \frac {p _ {n}}{p _ {\max }} \right], \tag {6} +$$ + +where $W$ is the layer-specific weight for the router $g(\mathbf{x})$ . This approach assigns progressively lower weights to experts with lower complexity. For intuitive understanding, this mechanism parallels a mechanical spring system, where the restoring force is proportional to both displacement and the spring constant. In our context, the expert's parameter count corresponds to displacement, while the normalization factor serves as the material-specific constant, collectively determining the magnitude of complexity bias. Without this, the routers lack a basis for selecting specific experts, leading to randomized routing and thus suboptimal model capacity utilization. Lastly, the complexity-aware importance Imp is used to compute the total auxiliary loss $\mathcal{L}_{\mathrm{aux}}$ defined as: + +$$ +\mathcal {L} _ {\text {a u x}} (\mathbf {x}) = ^ {1 / 2} \operatorname {C V} \left(\operatorname {I m p} (\mathbf {x})\right) ^ {2} + ^ {1 / 2} \operatorname {C V} \left(\operatorname {L o a d} (\mathbf {x})\right) ^ {2}, \tag {7} +$$ + +where CV denotes the coefficient of variation [43]. The effect of the auxiliary loss is shown in Fig. 5 and Tab. 5a. + +Table 1. Comparison to state-of-the-art on three degradations. PSNR (dB, ↑) and SSIM (↑) are reported on the full RGB images. Our method sets a new state-of-the-art on average across all benchmarks on two model scales, showcasing its scalability while being more efficient than prior work. ‘-’ represents unreported results and ‘*’ indicates VLM-based approaches. The best performances are highlighted. + +
MethodParams.DehazingDerainingDenoisingAverage
SOTSRain100LBSD68σ=15BSD68σ=25BSD68σ=50
LightBRDNet [48]-23.23.89527.42.89532.26.89829.76.83626.34.69327.80.843
LPNet [14]-20.84.82824.88.78426.47.77824.77.74821.26.55223.64.738
FDGAN [11]-24.71.92929.89.93330.25.91028.81.86826.43.77628.02.883
DL [13]2M26.92.93132.62.93133.05.91430.41.86126.90.74029.98.876
AirNet [24]9M27.94.96234.90.96733.92.93331.26.88828.00.79731.20.910
MoCE-IR-S (ours)11M30.94.97938.22.98334.08.93331.42.88828.16.79832.57.916
HeavyMPRNet [64]16M25.28.95533.57.95433.54.92730.89.88027.56.77930.17.899
PromptIR [36]36M30.58.97436.37.97233.98.93331.31.88828.06.79932.06.913
Gridformer [51]34M30.37.97037.15.97233.93.93131.37.88728.11.80132.19.912
Art-PromptIR [54]33M30.83.97937.94.98234.06.93431.42.89128.14.80132.49.917
DA-CLIP* [30]125M29.46.96336.28.96830.02.82124.86.58522.29.476--
UniProcessor* [12]1002M31.66.97938.17.98234.08.93531.42.89128.17.80332.70.918
MoCE-IR (ours)25M31.34.97938.57.98434.11.93231.45.88828.18.80032.73.917
+ +Table 2. Comparison to state-of-the-art on five degradations. PSNR (dB, ↑) and SSIM (↑) are reported on the full RGB images with (*) denoting general image restorers, others are specialized all-in-one approaches. Under more challenging degradations, our MoCE-IR framework outperforms prior work across two model scales. The best performances are highlighted. + +
MethodParams.DehazingDerainingDenoisingDeblurringLow-LightAverage
SOTSRain100LBSD68σ=25GoProLOLv1
LightSwinIR* [26]1M21.50.89130.78.92330.59.86824.52.77317.81.72325.04.835
DL [13]2M20.54.82621.96.76223.09.74519.86.67219.83.71221.05.743
TAPE [27]1M22.16.86129.67.90430.18.85524.47.76318.97.62125.09.801
AirNet [24]9M21.04.88432.98.95130.91.88224.35.78118.18.73525.49.847
MoCE-IR-S (ours)11M31.33.97837.21.97831.25.88428.90.87721.68.85130.08.913
HeavyNAFNet* [6]17M25.23.93935.56.96731.02.88326.53.80820.49.80927.76.881
DGUNet* [33]17M24.78.94036.62.97131.10.88327.25.83721.87.82328.32.891
Restormer* [65]26M24.09.92734.81.96231.49.88427.22.82920.41.80627.60.881
MambaIR [16]27M25.81.94436.55.97131.41.88428.61.87522.49.83228.97.901
Transweather [49]38M21.32.88529.43.90529.00.84125.12.75721.21.79225.22.836
IDR [66]15M25.24.94335.63.96531.60.88727.87.84621.34.82628.34.893
Gridformer [51]34M26.79.95136.61.97131.45.88529.22.88422.59.83129.33.904
InstructIR-5D [9]17M27.10.95636.84.97331.40.87329.40.88623.00.83629.55.908
MoCE-IR (ours)25M30.48.97438.04.98231.34.88730.05.89923.00.85230.58.919
+ +# 4. Experiments + +We conduct experiments by strictly following previous works in general image restoration [36, 66] under two different settings: (i) All-in-One and (ii) Composited degradations. In the all-in-one setting, a unified model is trained across multiple degradation types, where we consider three and five distinct degradations. In the Composited degradations setting, a single model is trained to handle both isolated degradation occurrences as well as combinations of up to three different types of degradations. + +Implementation Details. Our MoCE-IR framework is end-to-end trainable requiring no multi-stage optimization of any components. The architecture features an asymmetrical, four-level encoder-decoder design. The encoder consists of four levels with varying numbers of transformer blocks, arranged as $[4,6,6,8]$ from the top level to the lowest. The decoder includes three levels, containing $[2,4,4]$ transformer blocks, respectively. In each MoCE layer, we use $n = 4$ nested experts, where the dimensionality of the expert embedding is given by $R = C / 2^i$ for $i\in \{i,\dots,n\}$ . The network's initial embedding dimension is set to $C = 32$ , doubling at each + +subsequent level while the spatial resolution is halved. We follow the training configuration of prior work [36], and train our models for 120 epochs with a batch size of 32. We optimize the $L_{1}$ loss in the RGB and Fourier domain using the Adam [20] optimizer $(\beta_{1} = 0.9, \beta_{2} = 0.999)$ with an initial learning rate of $2 \times 10^{-4}$ and cosine decay. During training, we utilize crops of size $128^{2}$ with horizontal and vertical flips as augmentations. + +Datasets. For all-in-one, we follow existing work [24, 36] and include following datasets: For image denoising, we combine the BSD400 [2] and WED [31] datasets, adding Gaussian noise at levels $\sigma \in [15, 25, 50]$ to create noisy images. Testing is conducted on the BSD68 [32]. For deraining, we use Rain100L [56]. The dehazing task utilises the SOTS [23] dataset. For deblurring and low-light enhancement, we employ the GoPro [34] and the LOL-v1 [53] dataset, respectively. To develop a unified model for all tasks, we merge these datasets in a three (AIO-3) or five (AIO-5) degradation setting, and train for 120 epochs and directly evaluate them across different tasks. For the composite degradation setting, we use the CDD11 dataset [17] and train our method for 200 epochs, keeping all other settings unchanged. + +Table 3. Comparison to state-of-the-art on composited degradations. PSNR (dB, ↑) and SSIM (↑) are reported on the full RGB images with (*) denoting general image restorers, others are specialized all-in-one approaches. Our MoCE-IR method consistently outperforms even larger models, with favorable results in composited degradation scenarios. + +
MethodParams.CDD11-SingleCDD11-DoubleCDD11-TripleAvg.
Low (L)Haze (H)Rain (R)Snow (S)L+HL+RL+SH+RH+SL+H+RL+H+S
AirNet [24]9M24.83.77824.21.95126.55.89126.79.91923.23.77922.82.71023.29.72322.21.86823.29.90121.80.70822.24.72523.75.814
PromptIR [36]36M26.32.80526.10.96931.56.94631.53.96024.49.78925.05.77124.51.76124.54.92423.70.92523.74.75223.33.74725.90.850
WGWSNet [71]26M24.39.77427.90.98233.15.96434.43.97324.27.80025.06.77224.60.76527.23.95527.65.96023.90.77223.97.77126.96.863
WeatherDiff [72]83M23.58.76321.99.90424.85.88524.80.88821.83.75622.69.73022.12.70721.25.86821.99.86821.23.71621.04.69822.49.799
OneRestore [17]6M26.48.82632.52.99033.40.96434.31.97325.79.82225.58.79925.19.78929.99.95730.21.96424.78.78824.90.79128.47.878
MoCE-IR-S (ours)11M27.26.82432.66.99034.31.97035.91.98026.24.81726.25.80026.04.79329.93.96430.19.97025.41.78925.39.79029.05.881
DehazingInputAirNetPromptIROursGT
Deraining
Denoising
Input CropsError Maps (darker means better)GT Crops
+ +Figure 3. Visual results. We compare MoCE-IR-S to AirNet [24], and PromptIR [36] in the all-in-one setting with three degradations. MoCE-IR-S effectively removes haze and rain streaks while preserving image sharpness, achieving high-quality restoration. An error heatmap is provided, with color transitioning from black to white to indicate increasing pixel-wise error. + +# 4.1. Comparison to State-of-the-Art Methods + +All-in-One: Three Degradations. We compare our all-in-one restoration approach with specialized restoration methods, including both image-only models, such as MPR-Net [64], AirNet [24], PromptIR [36], and Gridformer [51], and vision-language models like DA-CLIP [30], InstructIR [9], and UniProcessor [12]. All methods were trained simultaneously on three types of degradation: haze, rain, and Gaussian noise. The results in Tab. 1 show that our proposed model, MoCE-IR, consistently outperforms previous works, both in light and heavy model scales. In the light category, MoCE-IR achieves the highest scores across all benchmarks with an average improvement in PSNR of + +1.37 dB, surpassing other models like AirNet [24]. In the heavy scale, MoCE-IR competes closely with UniProcessor [12], achieving an average PSNR of 32.73 dB with $98\%$ fewer parameters, validating our efficiency and effectiveness. + +All-in-One: Five Degradations. Based upon [24, 66], we expand the three-degradation setting to include deblurring and low-light image enhancement, further validating our method's effectiveness. As shown in Tab. 2, our approach excels by learning specialized experts with input-adaptive complexity and maximizing synergies across degradations. Our framework outperforms previous methods on the lightest model scale, with an average PSNR improvement of 4.31 dB over AirNet [24]. Even compared to larger networks, our lightest model surpasses the VLM-based InstructIR [9] by + +![](images/4054f08cddc89f8dc03a0e68ef767cca245144c4b5b439a6ef161def7dbc8169.jpg) +Figure 4. Complexity-efficiency tradeoff. Visualization of PSNR and parameter counts of proposed method compared to prior work. Proposed MoCE-IR surpasses prior methods, achieving SoTA results in all-in-one image restoration with enhanced efficiency. + +Table 4. Computational demands. GFLOPS are computed on an input image of size $224 \times 224$ using a NVIDIA RTX 4090 GPU. + +
MethodParams.MemoryFLOPSRuntime
AirNet [24]8.93M4829M238G42.17 ± 0.23
PromptIR [36]35.59M9830M132G41.28 ± 0.43
IDR [66]15.34M4905M98G-
MoCE-IR (ours)25.35M5887M80.59 ± 5.21G23.36 ± 2.47
MoCE-IR-S (ours)11.47M4228M36.93 ± 2.32G22.15 ± 2.59
+ +0.33 dB on average, and on a comparable model scale, our framework extends this advantage by an additional 0.50 dB. + +All-in-One: Composed Degradations. To simulate more realistic degradation scenarios, recent work [17] has trained all-in-one restorers not only on a mix of different degradations but also on more challenging cases where multiple degradation types are combined within a single image. We extend the previous AIO settings by including rain, haze, snow, and low illumination as individual degradation types, along with various composite degradation scenarios, resulting in a total of eleven distinct restoration settings. As shown in Tab. 3, our method outperforms previous all-in-one restorers by an average improvement of $0.64\mathrm{dB}$ over OneRestore [17], while being among the smallest models in size. This further validates our effectiveness with adaptive modeling capacity based on input complexity. + +Visual Results. We show visual results obtained under the three degradation settings in Fig. 3. In certain demanding hazy scenarios, both AirNet [24] and PromptIR [36] reveal constraints in completely eradicating haziness, resulting in noticeable color intensity discrepancies, whereas our approach ensures precise color reconstruction. Furthermore, in challenging rainy scenes, these popular approaches continue to exhibit notable remnants of rain, which our method adeptly eliminates, distinguishing itself from other approaches. Additionally, our method also generates clear and sharp denoised outputs. We further provide the error maps for visual differences where our method contains less noticeable errors. Coupled with the quantitative comparisons, these findings underscore the effectiveness of our method. Please refer to the supplementary for more visuals. + +Table 5. Architecture analysis. (a) We investigate scaling the expert embedding $r$ and the bias normalization for MoCE-IR-S. (b) We compare MoCE to several baselines [29, 43, 58], demonstrating that our complexity experts consistently outperform them. Notably, complexity-biased routing further boosts performance, surpassing standard load balancing (LB) [43]. PSNR (dB, ↑) and SSIM (↑) are reported on full RGB images. + +(a) Ablation on the expert scaling. + +
Embedding r scalingBias normalizationAvg. AIO-3
NestedpMin32.44.913
ExponentialpMax32.47.914
NestedpMax32.57.916
(b) Ablation on the complexity-aware routing.
MethodSOTSRain100LBSD6815-50Avg.
MoCE + Complexity Bias (ours)30.94.97938.22.98331.22.873
Baseline (MoE + LB [45])30.28.97737.77.98131.14.871
w/ MoCE + LB [45]30.13.97838.24.98331.19.872
w/ AMIR [58]29.97.97537.90.98131.17.872
w/ WM-MoE [29]30.10.97636.72.97631.05.869
+ +# 4.2. Ablation Studies + +We conduct detailed studies on the components within our method, highlighting the emergence of task-discriminative learners. All experiments are conducted in the all-in-one setting with three degradations. Please refer to the supplementary for additional ablations. + +**Complexity-Efficiency Trade-off.** Our MoCE-IR framework addresses the critical challenge of scalability in vision tasks by optimizing both reconstruction quality and computational efficiency. We visualize in Fig. 4 the progressive improvements in reconstruction fidelity of our framework on two complexity scales. Furthermore, Tab. 4 validates that our approach significantly outperforms SoTA methods in terms of memory utilization. Unlike traditional MoE architectures with fixed computational demands, our complexity experts enable dynamic FLOPS allocation. For instance, our lightweight model averages 36.9 GFLOPS, while the heavyweight model averages 80.59 GFLOPS. This adaptive computation, combined with lower memory consumption and parameter counts, gives MoCE-IR a substantial advantage over contemporary methods. + +Expert Scaling. As shown in Tab. 5a, we systematically evaluate different scaling strategies regarding the embedding dimensionality $r$ of each complexity expert. Our investigation compares exponential and nested parameter scaling, where the window partition is always fixed and follows a $2^{(2 + i)}$ progression. Further, we investigate the bias normalization, whether normalizing to the lowest or highest expert capacity. Overall, the nested parameter scaling emerges as the optimal choice, providing sufficient expert channel diversity while maintaining reasonable computational costs. Normalizing router weights by the highest expert capacity, rather than the minimum, proves crucial for performance as minimum capacity normalization forces router weights + +![](images/de8b9b36e22e38c7ca1958a064c3aee7684308a39347bd4dbc5d731bad47f8db.jpg) +Figure 5. Routing visualization for the AIO-3 setting. (a) While load balancing [43] ensures uniform expert utilization, it neglects shared task dependencies and task-specific characteristics, limiting restoration quality. (b)-(d) Complexity-aware routing fosters task discrimination by directing complex degradations to experts with broader contextual understanding and vice versa. This allows some experts to generalize across tasks while others specialize, enhancing adaptability. We visualize the average decisions made by each router for dehazing, deraining and denoising. The y-axis indicates increasing expert complexity. + +![](images/5625323f40b572039dc3c40b879f2066643f57312f94319dfb78d6e15058b691.jpg) + +![](images/6183b41c6ff9a9faa9bac3b25e2d0a586dfd507f0e5491de087fe9d94342db88.jpg) + +![](images/de6dd3673d508523700cdc658412cf2af71d9236496cd0646803af38d99699e6.jpg) + +Table 6. Expert generalization. We report PSNR for MoCE-IR-S with expert assignments fixed post-training. $\mathcal{E}$ denotes expert. + +
DegradationRoutingLearned ChoiceManual Choice
\( \mathcal{E}_1 \)\( \mathcal{E}_2 \)\( \mathcal{E}_3 \)\( \mathcal{E}_4 \)
Rain NoiseLoad Balance [45]Not Applicable28.2733.3328.2433.3728.2533.3728.2733.40
Rain NoiseComplexity Bias (ours)\( \mathcal{E}_1 \)30.4530.2129.7324.65
\( \mathcal{E}_4 \)33.9233.9333.9334.00
+ +of higher capacity experts to escalate uncontrollably, thus excluding them from the gating mechanism. + +Impact of Complexity-aware Routing. Building on the findings from Tab. 5a, we adopt nested expert scaling and evaluate our complexity-aware routing strategy against conventional load balancing [43], which enforces uniform expert usage. As shown in Tab. 5b, our MoCE-IR framework achieves superior performance by preferentially routing to lower-complexity experts, activating higher-complexity experts only when their added capacity provides clear benefits. This efficient model utilization not only boosts image restoration performance, consistently surpassing related adaptive restoration methods [29, 58], but also enhances out-of-distribution generalization as MoCE-IR effectively handles degradations that with significant deviation from the training data; see the supplementary for more evaluations. + +Routing Analysis. Fig. 5 illustrates the average expert selection per layer for different degradations: haze, rain, and noise. Higher-complexity experts are favored for tasks requiring broader context, like haze and noise, while lighter experts are chosen for rain streak removal. Notably, some experts handle multiple tasks, while others remain task-specific. Our complexity-conditioned routing promotes task-discriminative learning by directing inputs to relevant experts, preventing the collapse into a single-expert dependency, ensuring efficient and diverse operation. + +The baseline MoE [43] approach, lacking complexity awareness and scaled experts, uniformly distributes degraded samples across all experts. Our results show that this uniform allocation is suboptimal for all-in-one restoration, as it fails to leverage task-specific commonalities and distinctions effectively. Tab. 6 highlights the limitations of standard + +MoEs [43], where fixed post-training expert assignments result in uniform, task-agnostic performance, offering only moderate cross-task generalization. In contrast, MoCEs effectively generalize across degradations while exhibiting task-discriminative behavior, routing inputs to the most suitable experts for optimal performance. + +# 5. Conclusion, Limitations and Discussion + +This paper presents MoCE-IR, an efficient all-in-one image restoration model leveraging complexity experts, specialized units with varying computational demands and receptive fields. By aligning degradation tasks with suitable resources, MoCE-IR overcomes limitations of prior methods. Its built-in bias toward simpler paths enhances task-discriminative learning, enabling swift inference by bypassing irrelevant experts without sacrificing restoration quality. + +Extensive experiments across diverse degradation settings, including more demanding scenarios with multiple composited degradations, demonstrate our computational efficiency through selective expert utilization and consistent improvement over state-of-the-art methods. + +While MoCE-IR shows substantial improvements, opportunities remain. The current image-level routing imposes scalability constraints. Future work could explore SoftMoE [37] architectures for fine-grained spatial routing or early-exiting approaches [39] to further enhance processing speeds and input adaptability. Extending to synthetic-to-real adaptation, especially for video, represents a crucial next step. Performance optimization through mixed-precision computation across experts could further enhance speed and efficiency for resource-limited deployments. + +Acknowledgements. The authors sincerely thank the reviewers and all members of the program committee for their tremendous efforts and incisive feedback. This research was supported by the Alexander von Humboldt Foundation, the Ministry of Education and Science of Bulgaria (support for INSAIT, under the Bulgarian National Roadmap for Research Infrastructure), the Shanghai Municipal Science and Technology Major Project (2021SHZDZX0102), and the Fundamental Research Funds for the Central Universities. + +# References + +[1] Yuang Ai, Huaibo Huang, Xiaogiang Zhou, Jiexiang Wang, and Ran He. Multimodal prompt perceiver: Empower adaptiveness generalizability and fidelity for all-in-one image restoration. In CVPR, 2024. 2 +[2] Pablo Arbelaez, Michael Maire, Charless Fowlkes, and Jitendra Malik. Contour detection and hierarchical image segmentation. TPAMI, 33(5):898-916, 2010. 5 +[3] Yoshua Bengio, Nicholas Léonard, and Aaron Courville. Estimating or propagating gradients through stochastic neurons for conditional computation. arXiv preprint arXiv:1308.3432, 2013. 3 +[4] Dana Berman and Shai Avidan. Non-local image dehazing. In CVPR, 2016. 2 +[5] Bolun Cai, Xiangmin Xu, Kui Jia, Chunmei Qing, and Dacheng Tao. Dehazenet: An end-to-end system for single image haze removal. TIP, 25(11):5187-5198, 2016. 1 +[6] Liangyu Chen, Xiaojie Chu, Xiangyu Zhang, and Jian Sun. Simple baselines for image restoration. In ECCV, 2022. 1, 2, 5 +[7] Yu-Wei Chen and Soo-Chang Pei. Always clear days: Degradation type and severity aware all-in-one adverse weather removal. arXiv preprint arXiv:2310.18293, 2023. 1, 2 +[8] Zheng Chen, Yulun Zhang, Jinjin Gu, Linghe Kong, Xin Yuan, et al. Cross aggregation transformer for image restoration. NeurIPS, 2022. 1, 2 +[9] Marcos V Conde, Gregor Geigle, and Radu Timofte. Instructir: High-quality image restoration following human instructions. In ECCV, 2024. 1, 2, 3, 5, 6 +[10] Xin Dang, Hai Wang, Jie Ren, and Le Chen. An application performance optimization model of mobile augmented reality based on hd restoration. In CBD, pages 201-206. IEEE, 2020. 1 +[11] Yu Dong, Yihao Liu, He Zhang, Shifeng Chen, and Yu Qiao. Fd-gan: Generative adversarial networks with fusion-discriminator for single image dehazing. In AAAI, 2020. 5 +[12] Huiyu Duan, Xiongkuo Min, Sijing Wu, Wei Shen, and Guangtao Zhai. Uniprocessor: A text-induced unified low-level image processor. In ECCV, 2024. 1, 3, 5, 6 +[13] Qingnan Fan, Dongdong Chen, Lu Yuan, Gang Hua, Nenghai Yu, and Baoquan Chen. A general decoupled learning framework for parameterized image operators. TPAMI, 43(1): 33-47, 2019. 1, 5 +[14] Hongyun Gao, Xin Tao, Xiaoyong Shen, and Jiaya Jia. Dynamic scene deblurring with parameter selective sharing and nested skip connections. In CVPR, 2019. 5 +[15] Florin Girbacia, Silviu Butnariu, Alex Petre Orman, and Cristian Cezar Postelnicu. Virtual restoration of deteriorated religious heritage objects using augmented reality technologies. European Journal of Science and Theology, 9(2):223-231, 2013. 1 +[16] Hang Guo, Jinmin Li, Tao Dai, Zhihao Ouyang, Xudong Ren, and Shu-Tao Xia. Mambair: A simple baseline for image restoration with state-space model. In ECCV, 2024. 1, 5 +[17] Yu Guo, Yuan Gao, Yuxu Lu, Ryan Wen Liu, and Shengfeng He. Onerestore: A universal restoration framework for composite degradation. In ECCV, 2024. 5, 6, 7 + +[18] Kui Jiang, Zhongyuan Wang, Peng Yi, Chen Chen, Baojin Huang, Yimin Luo, Jiayi Ma, and Junjun Jiang. Multi-scale progressive fusion network for single image deraining. In CVPR, 2020. 2 +[19] Yitong Jiang, Zhaoyang Zhang, Tianfan Xue, and Jinwei Gu. Autodir: Automatic all-in-one image restoration with latent diffusion. arXiv preprint arXiv:2310.10123, 2023. 2 +[20] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv, 2017. 5 +[21] Lingshun Kong, Jiangxin Dong, Jianjun Ge, Mingqiang Li, and Jinshan Pan. Efficient frequency domain-based transformers for high-quality image deblurring. In CVPR, 2023. 2, 4 +[22] Jaakko Lehtinen, Jacob Munkberg, Jon Hasselgren, Samuli Laine, Tero Karras, Miika Aittala, and Timo Aila. Noise2noise: Learning image restoration without clean data. arXiv preprint arXiv:1803.04189, 2018. 1, 2 +[23] Boyi Li, Wenqi Ren, Dengpan Fu, Dacheng Tao, Dan Feng, Wenjun Zeng, and Zhangyang Wang. Benchmarking single-image dehazing and beyond. TIP, 28(1):492-505, 2018. 5 +[24] Boyun Li, Xiao Liu, Peng Hu, Zhongqin Wu, Jiancheng Lv, and Xi Peng. All-In-One Image Restoration for Unknown Corruption. In CVPR, 2022. 1, 2, 5, 6, 7 +[25] Zilong Li, Yiming Lei, Chenglong Ma, Junping Zhang, and Hongming Shan. Prompt-in-prompt learning for universal image restoration. arXiv preprint arXiv:2312.05038, 2023. 1, 2 +[26] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Swinir: Image restoration using swim transformer. In ICCV, 2021. 2, 5 +[27] Lin Liu, Lingxi Xie, Xiaopeng Zhang, Shanxin Yuan, Xiangyu Chen, Wengang Zhou, Houqiang Li, and Qi Tian. Tape: Task-agnostic prior embedding for image restoration. In ECCV, 2022. 1, 5 +[28] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, 2021. 1, 2 +[29] Yulin Luo, Rui Zhao, Xiaobao Wei, Jinwei Chen, Yijie Lu, Shenghao Xie, Tianyu Wang, Ruiqin Xiong, Ming Lu, and Shanghang Zhang. Wm-moe: Weather-aware multi-scale mixture-of-experts for blind adverse weather removal. arXiv preprint arXiv:2303.13739, 2023. 2, 3, 7, 8 +[30] Ziwei Luo, Fredrik K Gustafsson, Zheng Zhao, Jens Sjolund, and Thomas B Schon. Controlling vision-language models for multi-task image restoration. In ICLR, 2024. 2, 3, 5, 6 +[31] Kede Ma, Zhengfang Duanmu, Qingbo Wu, Zhou Wang, Hongwei Yong, Hongliang Li, and Lei Zhang. Waterloo exploration database: New challenges for image quality assessment models. TIP, 26(2):1004-1016, 2016. 5 +[32] David Martin, Charless Fowlkes, Doron Tal, and Jitendra Malik. A database of human segmented natural images and its application to evaluating segmentation algorithms and measuring ecological statistics. In ICCV, 2001. 5 +[33] Chong Mou, Qian Wang, and Jian Zhang. Deep generalized unfolding networks for image restoration. In CVPR, 2022. 5 + +[34] Seungjun Nah, Tae Hyun Kim, and Kyoung Mu Lee. Deep multi-scale convolutional neural network for dynamic scene deblurring. In CVPR, 2017. 5 +[35] Jinshan Pan, Deqing Sun, Hans Pfister, and Ming-Hsuan Yang. Blind image deblurring using dark channel prior. In CVPR, 2016. 2 +[36] Vaishnav Potlapalli, Syed Waqas Zamir, Salman Khan, and Fahad Khan. Prompt: Prompting for all-in-one image restoration. In NeurIPS, 2023. 1, 2, 3, 5, 6, 7 +[37] Joan Puigcerver, Carlos Riquelme Ruiz, Basil Mustafa, and Neil Houlsby. From sparse to soft mixtures of experts. In ICLR, 2024. 2, 3, 4, 8 +[38] Yanyun Qu, Yizi Chen, Jingying Huang, and Yuan Xie. Enhanced pix2pix dehazing network. In CVPR, 2019. 1 +[39] David Raposo, Sam Ritter, Blake Richards, Timothy Lillicrap, Peter Conway Humphreys, and Adam Santoro. Mixture-of-depths: Dynamically allocating compute in transformer-based language models, 2024. 2, 8 +[40] Dongwei Ren, Wangmeng Zuo, Qinghua Hu, Pengfei Zhu, and Deyu Meng. Progressive image deraining networks: A better and simpler baseline. In CVPR, 2019. 2 +[41] Wenqi Ren, Lin Ma, Jiawei Zhang, Jinshan Pan, Xiaochun Cao, Wei Liu, and Ming-Hsuan Yang. Gated fusion network for single image dehazing. In CVPR, 2018. 2 +[42] Wenqi Ren, Jinshan Pan, Hua Zhang, Xiaochun Cao, and Ming-Hsuan Yang. Single image dehazing via multi-scale convolutional neural networks with holistic edges. In ICCV, 2020. 2 +[43] Carlos Riquelme, Joan Puigcerver, Basil Mustafa, Maxim Neumann, Rodolphe Jenatton, André Susano Pinto, Daniel Keysers, and Neil Houlsby. Scaling vision with sparse mixture of experts. *NeurIPS*, 2021. 2, 3, 4, 7, 8 +[44] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-Net: Convolutional networks for biomedical image segmentation. In MICCAI, pages 234–241. Springer, 2015. 3 +[45] Giovanni Saggio, Davide Borra, et al. Augmented reality for restoration/reconstruction of artefacts with artistic or historical value. In Augmented reality: some emerging application areas, pages 59-86. InTech Publications, 2011. 1 +[46] Noam Shazeer, Azalia Mirhoseini, Krzysztof Maziarz, Andy Davis, Quoc Le, Geoffrey Hinton, and Jeff Dean. Outrageously large neural networks: The sparsely-gated mixture-of-experts layer, 2017. 3, 4 +[47] Ying Tai, Jian Yang, Xiaoming Liu, and Chunyan Xu. Memnet: A persistent memory network for image restoration. In ICCV, 2017. 1, 2 +[48] Chunwei Tian, Yong Xu, and Wangmeng Zuo. Image denoising using deep cnn with batch renormalization. NN, 2020. 5 +[49] Jeya Maria Jose Valanarasu, Rajeev Yasarla, and Vishal M Patel. Transweather: Transformer-based restoration of images degraded by adverse weather conditions. In CVPR, 2022. 1, 2, 5 +[50] Cong Wang, Jinshan Pan, Wei Wang, Jiangxin Dong, Mengzhu Wang, Yakun Ju, and Junyang Chen. Promptstorer: A prompting image restoration method with degradation perception. In NeurIPS, 2023. 1, 2 + +[51] Tao Wang, Kaihao Zhang, Ziqian Shao, Wenhan Luo, Bjorn Stenger, Tong Lu, Tae-Kyun Kim, Wei Liu, and Hongdong Li. Gridformer: Residual dense transformer with grid structure for image restoration in adverse weather conditions. ICCV, 2024. 5, 6 +[52] Zhendong Wang, Xiaodong Cun, Jianmin Bao, Wengang Zhou, Jianzhuang Liu, and Houqiang Li. Uformer: A general u-shaped transformer for image restoration. In CVPR, 2022. 1, 2 +[53] Chen Wei, Wenjing Wang, Wenhan Yang, and Jiaying Liu. Deep retina decomposition for low-light enhancement. arXiv preprint arXiv:1808.04560, 2018. 5 +[54] Gang Wu, Junjun Jiang, Kui Jiang, and Xianming Liu. Harmony in diversity: Improving all-in-one image restoration via multi-task collaboration. In ACMMM, 2024. 5 +[55] Haiyan Wu, Yanyun Qu, Shaohui Lin, Jian Zhou, Ruizhi Qiao, Zhizhong Zhang, Yuan Xie, and Lizhuang Ma. Contrastive learning for compact single image dehazing. In CVPR, 2021. 2 +[56] Fuzhi Yang, Huan Yang, Jianlong Fu, Hongtao Lu, and Baining Guo. Learning texture transformer network for image super-resolution. In CVPR, 2020. 5 +[57] Hao Yang, Liyuan Pan, Yan Yang, and Wei Liang. Language-driven all-in-one adverse weather removal. In CVPR, 2024. 1, 2, 3 +[58] Zhiwen Yang, Haowei Chen, Ziniu Qian, Yang Yi, Hui Zhang, Dan Zhao, Bingzheng Wei, and Yan Xu. All-in-one medical image restoration via task-adaptive routing. In MICCAI, 2024. 7, 8 +[59] Rajeev Yasarla and Vishal M Patel. Uncertainty guided multiscale residual learning-using a cycle spinning cnn for single image de-raining. In CVPR, 2019. 1 +[60] Ke Yu, Xintao Wang, Chao Dong, Xiaou Tang, and Chen Change Loy. Path-restore: Learning network path selection for image restoration. TPAMI, 2022. 3 +[61] Xiaoyan Yu, Shen Zhou, Huafeng Li, and Liehuang Zhu. Multi-expert adaptive selection: Task-balancing for all-in-one image restoration. arXiv preprint arXiv:2407.19139, 2024. 1 +[62] Eduard Zamfir, Zongwei Wu, Nancy Mehta, Danda Pani Paudel, Yulun Zhang, and Radu Timofte. Efficient degradation-aware any image restoration. arXiv preprint arXiv:2405.15475, 2024. 1, 2, 3 +[63] Eduard Zamfir, Zongwei Wu, Nancy Mehta, Yulun Zhang, and Radu Timofte. See more details: Efficient image superresolution by experts mining. In ICML. PMLR, 2024. 3 +[64] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, Ming-Hsuan Yang, and Ling Shao. Multi-stage progressive image restoration. In CVPR, 2021. 2, 5, 6 +[65] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. In CVPR, 2022. 1, 2, 3, 4, 5 +[66] Jinghao Zhang, Jie Huang, Mingde Yao, Zizheng Yang, Hu Yu, Man Zhou, and Feng Zhao. Ingredient-oriented multidegradation learning for image restoration. In CVPR, 2023. 1, 2, 5, 6, 7 + +[67] Kai Zhang, Wangmeng Zuo, Shuhang Gu, and Lei Zhang. Learning deep cnn denoiser prior for image restoration. In CVPR, 2017. 1, 2 +[68] Yulun Zhang, Kunpeng Li, Kai Li, Bineng Zhong, and Yun Fu. Residual non-local attention networks for image restoration. arXiv preprint arXiv:1903.10082, 2019. 1, 2 +[69] Haiyu Zhao, Yuanbiao Gou, Boyun Li, Dezhong Peng, Jiancheng Lv, and Xi Peng. Comprehensive and delicate: An efficient transformer for image restoration. In CVPR, 2023. 2 +[70] Yanqi Zhou, Tao Lei, Hanxiao Liu, Nan Du, Yanping Huang, Vincent Zhao, Andrew M Dai, Quoc V Le, James Laudon, et al. Mixture-of-experts with expert choice routing. NeurIPS, 2022. 2 +[71] Yurui Zhu, Tianyu Wang, Xueyang Fu, Xuanyu Yang, Xin Guo, Jifeng Dai, Yu Qiao, and Xiaowei Hu. Learning weather-general and weather-specific features for image restoration under multiple adverse weather conditions. In CVPR, 2023. 6 +[72] Ozan Özdenizci and Robert Legenstein. Restoring vision in adverse weather conditions with patch-based denoising diffusion models. TPAMI, 2023. 6 \ No newline at end of file diff --git a/2025/Complexity Experts are Task-Discriminative Learners for Any Image Restoration/images.zip b/2025/Complexity Experts are Task-Discriminative Learners for Any Image Restoration/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..7ac9acc63a6b7b0807e95e755f5650632a340f3e --- /dev/null +++ b/2025/Complexity Experts are Task-Discriminative Learners for Any Image Restoration/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:be6f76840c0553ed19c7dc0d70412984672cfd8b60bc220b6f3367f2b3cd7e5b +size 807217 diff --git a/2025/Complexity Experts are Task-Discriminative Learners for Any Image Restoration/layout.json b/2025/Complexity Experts are Task-Discriminative Learners for Any Image Restoration/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..bc80d297256149fa5b78410aabe3bbbdd1bced4a --- /dev/null +++ b/2025/Complexity Experts are Task-Discriminative Learners for Any Image Restoration/layout.json @@ -0,0 +1,8115 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 56, + 103, + 555, + 122 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 103, + 555, + 122 + ], + "spans": [ + { + "bbox": [ + 56, + 103, + 555, + 122 + ], + "type": "text", + "content": "Complexity Experts are Task-Discriminative Learners for Any Image Restoration" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 141, + 142, + 468, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 142, + 468, + 172 + ], + "spans": [ + { + "bbox": [ + 141, + 142, + 468, + 172 + ], + "type": "text", + "content": "Eduard Zamfir1 Zongwei Wu1* Nancy Mehta1 Yuedong Tan1 Danda Pani Paudel2 Yulun Zhang3 Radu Timofte1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 187, + 173, + 424, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 187, + 173, + 424, + 186 + ], + "spans": [ + { + "bbox": [ + 187, + 173, + 424, + 186 + ], + "type": "text", + "content": "1Computer Vision Lab, CAIDAS & IFI, University of Würzburg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 149, + 186, + 460, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 186, + 460, + 200 + ], + "spans": [ + { + "bbox": [ + 149, + 186, + 460, + 200 + ], + "type": "text", + "content": "2INSAIT, Sofia University \"St. Kliment Ohridski\" 3Shanghai Jiao Tong University" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "spans": [ + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 251, + 297, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 251, + 297, + 552 + ], + "spans": [ + { + "bbox": [ + 54, + 251, + 297, + 552 + ], + "type": "text", + "content": "Recent advancements in all-in-one image restoration models have revolutionized the ability to address diverse degradations through a unified framework. However, parameters tied to specific tasks often remain inactive for other tasks, making mixture-of-experts (MoE) architectures a natural extension. Despite this, MoEs often show inconsistent behavior, with some experts unexpectedly generalizing across tasks while others struggle within their intended scope. This hinders leveraging MoEs' computational benefits by bypassing irrelevant experts during inference. We attribute this undesired behavior to the uniform and rigid architecture of traditional MoEs. To address this, we introduce \"complexity experts\" - flexible expert blocks with varying computational complexity and receptive fields. A key challenge is assigning tasks to each expert, as degradation complexity is unknown in advance. Thus, we execute tasks with a simple bias toward lower complexity. To our surprise, this preference effectively drives task-specific allocation, assigning tasks to experts with the appropriate complexity. Extensive experiments validate our approach, demonstrating the ability to bypass irrelevant experts during inference while maintaining superior performance. The proposed MoCE-IR model outperforms state-of-the-art methods, affirming its efficiency and practical applicability. The source code and models are publicly available at eduardzamfir.github.io/MoCE-IR/" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 568, + 135, + 580 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 568, + 135, + 580 + ], + "spans": [ + { + "bbox": [ + 55, + 568, + 135, + 580 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 588, + 296, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 588, + 296, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 588, + 296, + 696 + ], + "type": "text", + "content": "Image restoration [47, 67, 68] is a fundamental problem in computer vision, dealing with reconstructing high-quality images from deteriorated observations. Adverse conditions such as noise [22, 67], haze [5, 38], or rain [7, 59] significantly impact the practical utility of images in downstream applications across various domains, including autonomous navigation [7, 49] or augmented reality [10, 15, 45]. Deep learning-based approaches have remarkably advanced this field, particularly for task-specific image restoration prob" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 316, + 226, + 422, + 299 + ], + "blocks": [ + { + "bbox": [ + 316, + 226, + 422, + 299 + ], + "lines": [ + { + "bbox": [ + 316, + 226, + 422, + 299 + ], + "spans": [ + { + "bbox": [ + 316, + 226, + 422, + 299 + ], + "type": "image", + "image_path": "9ca792966f6d0aa5f464e7d8e451aceaf073605373aab8e6a655da80f6d992b6.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 337, + 307, + 397, + 317 + ], + "lines": [ + { + "bbox": [ + 337, + 307, + 397, + 317 + ], + "spans": [ + { + "bbox": [ + 337, + 307, + 397, + 317 + ], + "type": "text", + "content": "(a) Dense Model" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 434, + 226, + 552, + 299 + ], + "blocks": [ + { + "bbox": [ + 434, + 226, + 552, + 299 + ], + "lines": [ + { + "bbox": [ + 434, + 226, + 552, + 299 + ], + "spans": [ + { + "bbox": [ + 434, + 226, + 552, + 299 + ], + "type": "image", + "image_path": "69324a597e6cf8a2befc7729dad123701ab939bdfff45e37303cf1462c07513c.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 479, + 307, + 511, + 317 + ], + "lines": [ + { + "bbox": [ + 479, + 307, + 511, + 317 + ], + "spans": [ + { + "bbox": [ + 479, + 307, + 511, + 317 + ], + "type": "text", + "content": "(b) MoE" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 316, + 323, + 425, + 396 + ], + "blocks": [ + { + "bbox": [ + 316, + 323, + 425, + 396 + ], + "lines": [ + { + "bbox": [ + 316, + 323, + 425, + 396 + ], + "spans": [ + { + "bbox": [ + 316, + 323, + 425, + 396 + ], + "type": "image", + "image_path": "9d9df16144abeaa5ca33f19b160da4030b5c8dc685c5800b445a29032c84e6d8.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 382, + 403, + 487, + 415 + ], + "lines": [ + { + "bbox": [ + 382, + 403, + 487, + 415 + ], + "spans": [ + { + "bbox": [ + 382, + 403, + 487, + 415 + ], + "type": "text", + "content": "(c) Complexity Experts (ours)" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 434, + 323, + 554, + 396 + ], + "blocks": [ + { + "bbox": [ + 434, + 323, + 554, + 396 + ], + "lines": [ + { + "bbox": [ + 434, + 323, + 554, + 396 + ], + "spans": [ + { + "bbox": [ + 434, + 323, + 554, + 396 + ], + "type": "image", + "image_path": "b46b5d1a1584d2df93afcce3f9e2ab92f068210dd752173968389d617f0a7802.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 420, + 556, + 618 + ], + "lines": [ + { + "bbox": [ + 313, + 420, + 556, + 618 + ], + "spans": [ + { + "bbox": [ + 313, + 420, + 556, + 618 + ], + "type": "text", + "content": "Figure 1. Motivation. (a) Dense all-in-one restoration methods [24, 57] often inefficiently allocate parameters when handling multiple degradation types. (b) While recent Mixture-of-Experts (MoE) approaches [61, 62] address this through sparse computation, their rigid routing mechanisms uniformly distribute inputs across experts without considering the natural relationships between degradations. (c) To overcome these limitations, we introduce Complexity Experts - adaptive processing blocks with size-varying computational units. Our framework dynamically allocates model capacity using a spring-inspired force mechanism that continuously guides routing decisions toward simpler experts when possible, with the force proportional to the complexity of the input degradation. While initially designed for computational efficiency, this approach naturally emerges as a task-discriminative learning framework, assigning degradations to the most suitable experts. This makes it particularly effective for all-in-one restoration methods, where both task-specific processing and cross-degradation knowledge sharing are crucial." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "content": "lems [6, 8, 16, 28, 52, 65]. Recent all-in-onerestoration models [9, 12, 13, 24, 27, 36, 49, 66], however, demonstrate the possibility of handling multiple degradation types within a single model, offering more practical solutions compared to traditional task-specific approaches without the extensive re-training need for each new degradation type. Notable works employ visual [25, 36, 50] or language-based prompt-" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 703, + 144, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 144, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 144, + 713 + ], + "type": "text", + "content": "*Corresponding author" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12753" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 96 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 96 + ], + "type": "text", + "content": "ing [1, 9, 30], contrastive learning [24, 66], and diffusion-based models [1, 30]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 98, + 297, + 242 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 98, + 297, + 242 + ], + "spans": [ + { + "bbox": [ + 54, + 98, + 297, + 242 + ], + "type": "text", + "content": "Despite their success, we observe that the aforementioned models often suffer from inefficiencies, as parameters tied to specific degradations often remain inactive or underutilized when addressing unrelated tasks [57]. This naturally suggests the use of mixture-of-experts (MoE) architectures for task-specific processing. Yet, current MoE-based approaches [29, 57, 62] typically incorporate routing mechanisms based on language [57] or degradation priors [29, 62], leading to imbalanced optimization, where some experts generalize well, while others struggle with their intended tasks. This inconsistency limits the potential computational benefits of bypassing task-irrelevant experts during inference." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 243, + 297, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 243, + 297, + 399 + ], + "spans": [ + { + "bbox": [ + 55, + 243, + 297, + 399 + ], + "type": "text", + "content": "We attribute this undesired behavior to two limitations in current approaches. First, the uniform and rigid architecture of existing MoE models fails to account for the varying complexity requirements across different restoration tasks. For instance, motion blur demands localized processing with strong spatial awareness, while haze removal requires broader contextual understanding [4, 35], necessitating adaptive processing aligned to the task requirements. Second, the challenge of appropriately routing tasks to experts is complicated by the unknown complexity of each degradation type a priori. Typically, MoE models [37, 39, 43, 70] aim to balance expert utilization, preventing the model from collapsing into a single-expert dependency." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 401, + 296, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 401, + 296, + 473 + ], + "spans": [ + { + "bbox": [ + 55, + 401, + 296, + 473 + ], + "type": "text", + "content": "In this work, we introduce a novel mixture-of-complexity-experts (MoCE) framework for all-in-one image restoration that directly addresses these limitations. Our key innovation lies in designing expert blocks with increasing computational complexity and receptive fields, allowing the model to adaptively match processing capacity with task requirements." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 475, + 297, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 475, + 297, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 475, + 297, + 713 + ], + "type": "text", + "content": "To address the routing challenge, we introduce a complexity-aware allocation mechanism that preferentially directs tasks to lower-complexity experts. This approach implements a complexity-proportional bias term, analogous to a mechanical spring force, to guide expert selection. Surprisingly, this straightforward approach results in effective task-specific allocation, naturally directing inputs to experts with appropriate complexity levels. Our MoCE framework offers two significant advantages: it ensures efficient inference by selectively bypassing irrelevant experts, thereby reducing computational overhead, while simultaneously maintaining high-quality restoration performance across diverse tasks. Through extensive experiments across multiple image restoration tasks, we demonstrate that our approach not only preserves restoration quality but surpasses recent state-of-the-art methods. These results validate both the effectiveness of our complexity-based expert design and the practicality of our routing strategy for real-world applications. Our approach establishes a new benchmark for all-in-one image restoration, enhancing both efficiency and fidelity." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 326, + 72, + 544, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 72, + 544, + 84 + ], + "spans": [ + { + "bbox": [ + 326, + 72, + 544, + 84 + ], + "type": "text", + "content": "The principal contributions of this work are threefold:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 86, + 555, + 194 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 314, + 86, + 555, + 121 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 86, + 555, + 121 + ], + "spans": [ + { + "bbox": [ + 314, + 86, + 555, + 121 + ], + "type": "text", + "content": "- We propose MoCE-IR, achieving SoTA all-in-one image restoration performance with improved efficiency compared to prior works." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 122, + 555, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 122, + 555, + 157 + ], + "spans": [ + { + "bbox": [ + 314, + 122, + 555, + 157 + ], + "type": "text", + "content": "- Our novel MoE layer selectively activates complexity experts based on input requirements, unifying task-specific and holistic learning in a single architecture." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 158, + 554, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 158, + 554, + 194 + ], + "spans": [ + { + "bbox": [ + 314, + 158, + 554, + 194 + ], + "type": "text", + "content": "- We develop a complexity-aware routing mechanism that balances restoration quality with computational efficiency by adaptive expert allocation." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 209, + 406, + 221 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 209, + 406, + 221 + ], + "spans": [ + { + "bbox": [ + 314, + 209, + 406, + 221 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 231, + 556, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 231, + 556, + 529 + ], + "spans": [ + { + "bbox": [ + 313, + 231, + 556, + 529 + ], + "type": "text", + "content": "Task Specific Image Restoration. Reconstructing the clean image from its degraded counterpart is a highly ill-posed problem, however, a great body of work have addressed image restoration from a data-driven learning perspective, achieving tremendous results compared to prior hand-crafted methods [22, 26, 47, 52, 65, 67, 68]. Most proposed solutions build on convolutional [6, 47, 67, 68] or Transformer-based architectures [8, 28, 52, 65] addressing single degradation tasks, such as denoising [8, 67, 68], dehazing [41, 42, 55] or deraining [18, 40]. Contrary to CNN-based networks, Transformer offer strong modeling capabilities for capturing global dependencies, which makes them outstanding image restorers [28, 65, 69]. Self-attention's quadratic complexity w.r.t the image size poses a challenge for resource-constrained applications. Conversely, convolutions offer fast and efficient processing with limited global context, but they scale more effectively with increasing input size. Recently, FFTformer [21] introduced an efficient approximation of query and key interactions in the Fourier domain, significantly reducing computational demands compared to window-based attention methods for image deblurring. Building on this, we incorporate specialized experts into a Transformer-based architecture [65] to efficiently capture both shared and distinct contextual information, addressing the diverse demands of image restoration tasks." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 534, + 557, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 557, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 557, + 713 + ], + "type": "text", + "content": "All-in-One Image Restoration. Restoring degraded images often requires multiple models, complicating practical implementation, especially when images exhibit a combination of various degradations. An emerging field known as all-in-one image restoration is advancing in low-level computer vision, utilizing a single deep blind restoration model to tackle multiple degradation types simultaneously [7, 19, 36, 49, 64, 66]. The seminal work, AirNet [24] achieves blind all-in-one image restoration using contrastive learning to extract degradation representations, which guide the restoration process. Similarly, IDR [66] employs a meta-learning-based two-stage approach to model degradation through underlying physical principles. Prompt-based learning [25, 36, 50] has also gained traction, with [36] introducing tunable prompts to encode degradation-specific in" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "12754" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 70, + 221, + 243 + ], + "blocks": [ + { + "bbox": [ + 60, + 70, + 221, + 243 + ], + "lines": [ + { + "bbox": [ + 60, + 70, + 221, + 243 + ], + "spans": [ + { + "bbox": [ + 60, + 70, + 221, + 243 + ], + "type": "image", + "image_path": "d89bfebbc3610f17c2faf0b54f62894d3210c7910aa4df55795ffce9cf632e7e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 97, + 246, + 181, + 255 + ], + "lines": [ + { + "bbox": [ + 97, + 246, + 181, + 255 + ], + "spans": [ + { + "bbox": [ + 97, + 246, + 181, + 255 + ], + "type": "text", + "content": "(a) Architecture overview" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 244, + 72, + 367, + 237 + ], + "blocks": [ + { + "bbox": [ + 244, + 72, + 367, + 237 + ], + "lines": [ + { + "bbox": [ + 244, + 72, + 367, + 237 + ], + "spans": [ + { + "bbox": [ + 244, + 72, + 367, + 237 + ], + "type": "image", + "image_path": "f3471c152ca5a81a0179dd375026f12638ffa2b774dd409c7c7378e6a64d89b0.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 266, + 245, + 343, + 255 + ], + "lines": [ + { + "bbox": [ + 266, + 245, + 343, + 255 + ], + "spans": [ + { + "bbox": [ + 266, + 245, + 343, + 255 + ], + "type": "text", + "content": "(b) Transformer blocks." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 388, + 72, + 553, + 239 + ], + "blocks": [ + { + "bbox": [ + 388, + 72, + 553, + 239 + ], + "lines": [ + { + "bbox": [ + 388, + 72, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 388, + 72, + 553, + 239 + ], + "type": "image", + "image_path": "a849b69f4a55e7e65565adb2930570db71fcfea3615872b33bde60103a703f26.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 416, + 245, + 526, + 255 + ], + "lines": [ + { + "bbox": [ + 416, + 245, + 526, + 255 + ], + "spans": [ + { + "bbox": [ + 416, + 245, + 526, + 255 + ], + "type": "text", + "content": "(c) Mixture-of-complexity-experts" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 54, + 262, + 555, + 307 + ], + "lines": [ + { + "bbox": [ + 54, + 262, + 555, + 307 + ], + "spans": [ + { + "bbox": [ + 54, + 262, + 555, + 307 + ], + "type": "text", + "content": "Figure 2. Proposed MoCE-IR framework. Despite recent advances in MoE-based image restoration [57, 62, 63], inconsistent expert behavior—where some experts over-generalize while others underperform—limits their computational efficiency. We address this through complexity experts: flexible blocks with varying computational capacity and receptive fields. Our MoCE-IR employs an asymmetric encoder-decoder architecture where each decoder block contains a mixture-of-complexity-experts layer for adaptive capacity routing." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 314, + 296, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 314, + 296, + 445 + ], + "spans": [ + { + "bbox": [ + 54, + 314, + 296, + 445 + ], + "type": "text", + "content": "formation, albeit with a high parameter count. With recent advancements in multimodal integration, such as combining images and language, there has been growing interest in language-guided image restoration. Notable approaches include DA-CLIP [30], InstructIR [9], and UniProcessor [12]. However, these methods often depend on high-quality text prompts and typically require LLMs, limiting their applicability on resource-constrained devices. In contrast, our contribution lies in a novel, parameter-efficient approach that focuses on selective activation, effectively addressing diverse degradations while remaining lightweight and adaptable." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 450, + 296, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 450, + 296, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 450, + 296, + 628 + ], + "type": "text", + "content": "Dynamic Networks. Dynamic networks have evolved from basic conditional computation [3] to sophisticated Mixture-of-Experts (MoE) architectures that expand model capacity while maintaining efficient inference costs [37, 43, 46]. This approach has proven particularly effective in transformer-based NLP and high-level vision tasks. In image restoration, Path-Restore [60] introduced content-aware patch routing with difficulty-regulated rewards, while recent all-in-one models [29, 57, 62] leverage various priors for expert routing. Our work addresses the challenge of adaptively routing corrupted images to appropriate computational units within a generalist restoration framework. By dynamically aligning image characteristics with computational requirements, we achieve a unified model in both parameter and architectural space, advancing beyond previous limitations." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 644, + 111, + 656 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 644, + 111, + 656 + ], + "spans": [ + { + "bbox": [ + 55, + 644, + 111, + 656 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "content": "In this section, we outline the core principles of our All-in-One image restoration method. As shown in Fig. 2a, our pipeline features a U-shaped architecture [44, 65] with an asymmetric encoder-decoder design. A " + }, + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "content": " convolution" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 314, + 555, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 314, + 555, + 445 + ], + "spans": [ + { + "bbox": [ + 313, + 314, + 555, + 445 + ], + "type": "text", + "content": "first extracts shallow features from the degraded input, which then pass through 4 levels of encoding and decoding stages. While both stages utilize Transformer blocks [36, 65], our novel MoCE layers are exclusively integrated into the decoder blocks (Fig. 2b). To further enhance the decoder's feature enrichment capabilities, we introduce high-frequency guidance via Sobel-filtered global feature vector to promote the frequency awareness of the gating functions. Finally, a global residual connection links the shallow features to the output of the refinement stage, refining the crucial high-frequency details before producing the restored image." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 453, + 486, + 466 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 453, + 486, + 466 + ], + "spans": [ + { + "bbox": [ + 313, + 453, + 486, + 466 + ], + "type": "text", + "content": "3.1. Mixture-of-Complexity-Experts" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 471, + 555, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 471, + 555, + 518 + ], + "spans": [ + { + "bbox": [ + 313, + 471, + 555, + 518 + ], + "type": "text", + "content": "To overcome the limitations of uniform MoE architectures in image restoration, we propose an efficient input-adaptive model for all-in-one image restoration that dynamically adjusts processing capacity to match task requirements." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 519, + 556, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 519, + 556, + 687 + ], + "spans": [ + { + "bbox": [ + 313, + 519, + 556, + 687 + ], + "type": "text", + "content": "Technically, at the core of our framework is a specialized MoCE layer, consisting of " + }, + { + "bbox": [ + 313, + 519, + 556, + 687 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 519, + 556, + 687 + ], + "type": "text", + "content": " complexity experts " + }, + { + "bbox": [ + 313, + 519, + 556, + 687 + ], + "type": "inline_equation", + "content": "\\mathbf{E}" + }, + { + "bbox": [ + 313, + 519, + 556, + 687 + ], + "type": "text", + "content": " and a single shared expert " + }, + { + "bbox": [ + 313, + 519, + 556, + 687 + ], + "type": "inline_equation", + "content": "\\mathbf{S}" + }, + { + "bbox": [ + 313, + 519, + 556, + 687 + ], + "type": "text", + "content": ". As shown in Fig. 2b, our experts are designed with increasing computational complexity and progressively larger receptive fields, in contrast to the uniform architecture of traditional MoE designs [37, 43, 46]. The interaction between these components is orchestrated through a two-level gating mechanism: first projecting routed tokens into expert-specific embeddings, then combining specialized and agnostic features through modulation, and finally merging the outputs via cross-attention. This enables our model to capture both degradation-specific features and interdegradation relationships, while maintaining a bias toward computationally efficient processing paths." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": "Expert Design. Each complexity expert block as shown in Fig. 2c is designed based on two fundamental principles:" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12755" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 251 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 251 + ], + "type": "text", + "content": "ensuring computational efficiency and capturing hierarchical spatial features, which are critical for restoring fine details and addressing diverse degradation patterns. Given the growing computational demands associated with increasing expert numbers and capacities, it is crucial to prioritize an overall efficient expert design. To this end, we implement a nested expert structure, progressively reducing channel dimensionality " + }, + { + "bbox": [ + 55, + 72, + 296, + 251 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 55, + 72, + 296, + 251 + ], + "type": "text", + "content": " within each subsequent expert to control the computational overhead. Simultaneously, we increase the receptive field, i.e. window partition size " + }, + { + "bbox": [ + 55, + 72, + 296, + 251 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 55, + 72, + 296, + 251 + ], + "type": "text", + "content": ", within each expert to adaptively balance the localized and global processing, tailored to the specific requirements of the input degradation. Thus, our complexity experts are constructed by scaling the design along two key axes: window partition size and channel dimension." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 252, + 296, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 252, + 296, + 456 + ], + "spans": [ + { + "bbox": [ + 55, + 252, + 296, + 456 + ], + "type": "text", + "content": "More specifically, each expert " + }, + { + "bbox": [ + 55, + 252, + 296, + 456 + ], + "type": "inline_equation", + "content": "\\mathbf{E}" + }, + { + "bbox": [ + 55, + 252, + 296, + 456 + ], + "type": "text", + "content": " projects its input tokens to the embedding " + }, + { + "bbox": [ + 55, + 252, + 296, + 456 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_E\\in \\mathbb{R}^{H\\times W\\times r}" + }, + { + "bbox": [ + 55, + 252, + 296, + 456 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 252, + 296, + 456 + ], + "type": "inline_equation", + "content": "r = C / 2^i" + }, + { + "bbox": [ + 55, + 252, + 296, + 456 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 55, + 252, + 296, + 456 + ], + "type": "inline_equation", + "content": "i\\in \\{i,\\dots,n\\}" + }, + { + "bbox": [ + 55, + 252, + 296, + 456 + ], + "type": "text", + "content": ", to highlight the most relevant information along the channel dimension. Following this, a window-based self-attention (WSA) mechanism captures spatial information effectively. To optimize performance, we employ an FFT-based approximation [21] for efficient matrix multiplication between queries " + }, + { + "bbox": [ + 55, + 252, + 296, + 456 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 55, + 252, + 296, + 456 + ], + "type": "text", + "content": " and keys " + }, + { + "bbox": [ + 55, + 252, + 296, + 456 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 252, + 296, + 456 + ], + "type": "text", + "content": " in the Fourier domain. The refined features are then transformed back into the spatial domain and re-projected to their original dimensionality. Lastly, the output of the expert block " + }, + { + "bbox": [ + 55, + 252, + 296, + 456 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}_E^i" + }, + { + "bbox": [ + 55, + 252, + 296, + 456 + ], + "type": "text", + "content": " is modulated through element-wise multiplication with the holistic features from the shared expert, S. As shared expert, we employ the transposed self-attention (T-SA) module [65] in the channel dimension, while the complexity experts operate spatially. More concretely, the modulated expert features " + }, + { + "bbox": [ + 55, + 252, + 296, + 456 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_{\\mathbf{E}}^{\\mathrm{i}}" + }, + { + "bbox": [ + 55, + 252, + 296, + 456 + ], + "type": "text", + "content": " are obtained as following:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 93, + 464, + 296, + 479 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 464, + 296, + 479 + ], + "spans": [ + { + "bbox": [ + 93, + 464, + 296, + 479 + ], + "type": "interline_equation", + "content": "\\mathbf {y} _ {\\mathbf {E}} ^ {i} = \\hat {\\mathbf {x}} _ {E} ^ {i} \\odot \\mathbf {S} (\\mathbf {x}) \\tag {1}", + "image_path": "b05d3d519349f9581f41a9c54f903581c7d3ff5a776cdb99a2fec816829d8270.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 73, + 481, + 296, + 496 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 481, + 296, + 496 + ], + "spans": [ + { + "bbox": [ + 73, + 481, + 296, + 496 + ], + "type": "interline_equation", + "content": "\\text {w i t h} \\hat {\\mathbf {x}} _ {E} ^ {i} = \\mathbf {E} ^ {i} \\left(\\mathbf {x} _ {E} ^ {i}\\right) = \\operatorname {F F T - W S A} _ {\\mathrm {W} _ {i}} \\left(\\mathbf {W} _ {C \\rightarrow r _ {i}} ^ {i} \\mathbf {x} _ {E} ^ {i}\\right) \\tag {2}", + "image_path": "885fe8f50474bf9ace0a24e93f7f383fee398436658fcd6f4385002365324707.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 69, + 498, + 296, + 510 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 498, + 296, + 510 + ], + "spans": [ + { + "bbox": [ + 69, + 498, + 296, + 510 + ], + "type": "interline_equation", + "content": "\\operatorname {a n d} \\mathbf {S} (\\mathbf {x}) = \\mathrm {T} - \\operatorname {S A} (\\mathbf {x}) \\tag {3}", + "image_path": "19cc86c581bce2a089d867590f8da304037d2e7f3f4431455fddfa41d16b0a6f.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 520, + 296, + 627 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 520, + 296, + 627 + ], + "spans": [ + { + "bbox": [ + 55, + 520, + 296, + 627 + ], + "type": "text", + "content": "where linear layers for projection are denoted as " + }, + { + "bbox": [ + 55, + 520, + 296, + 627 + ], + "type": "inline_equation", + "content": "\\mathbf{W}" + }, + { + "bbox": [ + 55, + 520, + 296, + 627 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 520, + 296, + 627 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 55, + 520, + 296, + 627 + ], + "type": "text", + "content": " denotes element-wise multiplication and " + }, + { + "bbox": [ + 55, + 520, + 296, + 627 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 520, + 296, + 627 + ], + "type": "text", + "content": " is the index of the current complexity expert " + }, + { + "bbox": [ + 55, + 520, + 296, + 627 + ], + "type": "inline_equation", + "content": "\\mathbf{E}^i" + }, + { + "bbox": [ + 55, + 520, + 296, + 627 + ], + "type": "text", + "content": " with window size " + }, + { + "bbox": [ + 55, + 520, + 296, + 627 + ], + "type": "inline_equation", + "content": "w_i" + }, + { + "bbox": [ + 55, + 520, + 296, + 627 + ], + "type": "text", + "content": " and embedding dimensionality " + }, + { + "bbox": [ + 55, + 520, + 296, + 627 + ], + "type": "inline_equation", + "content": "r_i" + }, + { + "bbox": [ + 55, + 520, + 296, + 627 + ], + "type": "text", + "content": ". For example, the most lightweight expert is characterized by the smallest embedding dimension " + }, + { + "bbox": [ + 55, + 520, + 296, + 627 + ], + "type": "inline_equation", + "content": "r_1" + }, + { + "bbox": [ + 55, + 520, + 296, + 627 + ], + "type": "text", + "content": " and window size " + }, + { + "bbox": [ + 55, + 520, + 296, + 627 + ], + "type": "inline_equation", + "content": "w_1" + }, + { + "bbox": [ + 55, + 520, + 296, + 627 + ], + "type": "text", + "content": ", while the dimensions and window partitions for the other experts can increase linearly. Throughout our network, the capacity of the shared expert remains constant, with " + }, + { + "bbox": [ + 55, + 520, + 296, + 627 + ], + "type": "inline_equation", + "content": "r = C" + }, + { + "bbox": [ + 55, + 520, + 296, + 627 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 635, + 206, + 649 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 635, + 206, + 649 + ], + "spans": [ + { + "bbox": [ + 55, + 635, + 206, + 649 + ], + "type": "text", + "content": "3.2. Complexity-aware Routing" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "content": "Drawing inspiration from sparse MoE [43, 46], we integrate linear layers within each decoder block to enable a routing mechanism, associating input features " + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{x} \\in \\mathbb{R}^{H \\times W \\times C}" + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "content": " with their corresponding specialized complexity experts " + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{E}" + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "content": ". However, in the context of image restoration, a key challenge" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "type": "text", + "content": "arises in achieving a scale-invariant tokenization of input images, ensuring consistency across varying resolutions and scales. While previous MoE approaches [37, 43, 46] employ a token-based routing, we opt for an image-level routing strategy instead, where we select experts for entire input image. The routing function " + }, + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "type": "inline_equation", + "content": "g(\\mathbf{x})" + }, + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "type": "text", + "content": " orchestrates the allocation of input samples based on the required compute to corresponding complexity experts " + }, + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "type": "inline_equation", + "content": "\\mathbf{E}^i" + }, + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "type": "text", + "content": ", indexed by " + }, + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "type": "inline_equation", + "content": "i \\in 1, \\dots, n" + }, + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "type": "text", + "content": " representing the total number of experts. Within " + }, + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "type": "inline_equation", + "content": "g(\\mathbf{x})" + }, + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "type": "text", + "content": ", we select the top- " + }, + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "type": "text", + "content": " elements of the softmax distribution associating input tokens to experts, setting all other elements to zero. In practice, we use " + }, + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "type": "inline_equation", + "content": "k = 1" + }, + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "type": "text", + "content": " while " + }, + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "type": "text", + "content": " is sampled independently " + }, + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "type": "inline_equation", + "content": "\\epsilon \\sim \\mathcal{N}(0, 1/n^2)" + }, + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "type": "text", + "content": " entry-wise enabling noisy top-1 routing:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 365, + 251, + 555, + 265 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 251, + 555, + 265 + ], + "spans": [ + { + "bbox": [ + 365, + 251, + 555, + 265 + ], + "type": "interline_equation", + "content": "g (\\mathbf {x}) = \\operatorname {t o p} _ {k} \\left(\\operatorname {S o f t m a x} \\left(\\mathbf {W} \\mathbf {x} + \\epsilon\\right)\\right) \\tag {4}", + "image_path": "1020f518b3d6baf2c2dcd34c18aca0f2791251d938d1556f8bbed1e9e6954e0f.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 275, + 556, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 275, + 556, + 431 + ], + "spans": [ + { + "bbox": [ + 313, + 275, + 556, + 431 + ], + "type": "text", + "content": "During training, we incorporate an auxiliary loss " + }, + { + "bbox": [ + 313, + 275, + 556, + 431 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{aux}}" + }, + { + "bbox": [ + 313, + 275, + 556, + 431 + ], + "type": "text", + "content": " inspired by the load balancing loss proposed by Riquelme et al. [43]. We complement the importance term to favor experts based on their computational complexity, allowing for more effective balancing aligned with expert capacity. As in [43], the balanced expert utilization is encouraged through an importance loss, where the importance of each expert " + }, + { + "bbox": [ + 313, + 275, + 556, + 431 + ], + "type": "inline_equation", + "content": "\\mathbf{E}^i" + }, + { + "bbox": [ + 313, + 275, + 556, + 431 + ], + "type": "text", + "content": " for a batch of images is defined as the sum of the routing weights associated with the " + }, + { + "bbox": [ + 313, + 275, + 556, + 431 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 275, + 556, + 431 + ], + "type": "text", + "content": "-th expert across the entire batch. Additionally, we introduce a complexity bias " + }, + { + "bbox": [ + 313, + 275, + 556, + 431 + ], + "type": "inline_equation", + "content": "\\mathbf{b}" + }, + { + "bbox": [ + 313, + 275, + 556, + 431 + ], + "type": "text", + "content": " by calculating the number of learnable parameters " + }, + { + "bbox": [ + 313, + 275, + 556, + 431 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 313, + 275, + 556, + 431 + ], + "type": "text", + "content": " in each expert " + }, + { + "bbox": [ + 313, + 275, + 556, + 431 + ], + "type": "inline_equation", + "content": "\\mathbf{E}^i" + }, + { + "bbox": [ + 313, + 275, + 556, + 431 + ], + "type": "text", + "content": " and normalizing these values with respect to the largest parameter count of the experts, " + }, + { + "bbox": [ + 313, + 275, + 556, + 431 + ], + "type": "inline_equation", + "content": "p_{\\mathrm{max}}" + }, + { + "bbox": [ + 313, + 275, + 556, + 431 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 353, + 440, + 555, + 473 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 440, + 555, + 473 + ], + "spans": [ + { + "bbox": [ + 353, + 440, + 555, + 473 + ], + "type": "interline_equation", + "content": "\\operatorname {I m p} _ {i} (\\mathbf {x}) = \\left(\\sum_ {x \\in \\mathbf {x}} \\operatorname {S o f t m a x} (W x) _ {i}\\right) * \\mathbf {b}, \\tag {5}", + "image_path": "8e0fb99dc920600eef5bda968bc5ebd3dcea87e1038f5fdd0887801385a26cf7.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 359, + 475, + 555, + 489 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 475, + 555, + 489 + ], + "spans": [ + { + "bbox": [ + 359, + 475, + 555, + 489 + ], + "type": "interline_equation", + "content": "\\text {w i t h} \\mathbf {b} = \\left[ \\frac {p _ {1}}{p _ {\\max }}, \\frac {p _ {2}}{p _ {\\max }}, \\dots , \\frac {p _ {n}}{p _ {\\max }} \\right], \\tag {6}", + "image_path": "83c87bd7a434fbefbabf8fd9c7849015d17886e567f6b42393fcca50451db86a.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 498, + 556, + 655 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 556, + 655 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 556, + 655 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 498, + 556, + 655 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 313, + 498, + 556, + 655 + ], + "type": "text", + "content": " is the layer-specific weight for the router " + }, + { + "bbox": [ + 313, + 498, + 556, + 655 + ], + "type": "inline_equation", + "content": "g(\\mathbf{x})" + }, + { + "bbox": [ + 313, + 498, + 556, + 655 + ], + "type": "text", + "content": ". This approach assigns progressively lower weights to experts with lower complexity. For intuitive understanding, this mechanism parallels a mechanical spring system, where the restoring force is proportional to both displacement and the spring constant. In our context, the expert's parameter count corresponds to displacement, while the normalization factor serves as the material-specific constant, collectively determining the magnitude of complexity bias. Without this, the routers lack a basis for selecting specific experts, leading to randomized routing and thus suboptimal model capacity utilization. Lastly, the complexity-aware importance Imp is used to compute the total auxiliary loss " + }, + { + "bbox": [ + 313, + 498, + 556, + 655 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{aux}}" + }, + { + "bbox": [ + 313, + 498, + 556, + 655 + ], + "type": "text", + "content": " defined as:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 325, + 665, + 555, + 679 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 665, + 555, + 679 + ], + "spans": [ + { + "bbox": [ + 325, + 665, + 555, + 679 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {a u x}} (\\mathbf {x}) = ^ {1 / 2} \\operatorname {C V} \\left(\\operatorname {I m p} (\\mathbf {x})\\right) ^ {2} + ^ {1 / 2} \\operatorname {C V} \\left(\\operatorname {L o a d} (\\mathbf {x})\\right) ^ {2}, \\tag {7}", + "image_path": "c6e2020cbc673c00ccceca8880fb1591087c8ff4b31873346a172c6f02a4839a.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "content": "where CV denotes the coefficient of variation [43]. The effect of the auxiliary loss is shown in Fig. 5 and Tab. 5a." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "12756" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 106, + 553, + 247 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 105 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 105 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 105 + ], + "type": "text", + "content": "Table 1. Comparison to state-of-the-art on three degradations. PSNR (dB, ↑) and SSIM (↑) are reported on the full RGB images. Our method sets a new state-of-the-art on average across all benchmarks on two model scales, showcasing its scalability while being more efficient than prior work. ‘-’ represents unreported results and ‘*’ indicates VLM-based approaches. The best performances are highlighted." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 106, + 553, + 247 + ], + "lines": [ + { + "bbox": [ + 56, + 106, + 553, + 247 + ], + "spans": [ + { + "bbox": [ + 56, + 106, + 553, + 247 + ], + "type": "table", + "html": "
MethodParams.DehazingDerainingDenoisingAverage
SOTSRain100LBSD68σ=15BSD68σ=25BSD68σ=50
LightBRDNet [48]-23.23.89527.42.89532.26.89829.76.83626.34.69327.80.843
LPNet [14]-20.84.82824.88.78426.47.77824.77.74821.26.55223.64.738
FDGAN [11]-24.71.92929.89.93330.25.91028.81.86826.43.77628.02.883
DL [13]2M26.92.93132.62.93133.05.91430.41.86126.90.74029.98.876
AirNet [24]9M27.94.96234.90.96733.92.93331.26.88828.00.79731.20.910
MoCE-IR-S (ours)11M30.94.97938.22.98334.08.93331.42.88828.16.79832.57.916
HeavyMPRNet [64]16M25.28.95533.57.95433.54.92730.89.88027.56.77930.17.899
PromptIR [36]36M30.58.97436.37.97233.98.93331.31.88828.06.79932.06.913
Gridformer [51]34M30.37.97037.15.97233.93.93131.37.88728.11.80132.19.912
Art-PromptIR [54]33M30.83.97937.94.98234.06.93431.42.89128.14.80132.49.917
DA-CLIP* [30]125M29.46.96336.28.96830.02.82124.86.58522.29.476--
UniProcessor* [12]1002M31.66.97938.17.98234.08.93531.42.89128.17.80332.70.918
MoCE-IR (ours)25M31.34.97938.57.98434.11.93231.45.88828.18.80032.73.917
", + "image_path": "f5fb81bcffd931c7467eae82bb00a904fa500ddd9f2d38c8be58ee9fac0d79ff.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 56, + 284, + 553, + 434 + ], + "blocks": [ + { + "bbox": [ + 55, + 250, + 555, + 283 + ], + "lines": [ + { + "bbox": [ + 55, + 250, + 555, + 283 + ], + "spans": [ + { + "bbox": [ + 55, + 250, + 555, + 283 + ], + "type": "text", + "content": "Table 2. Comparison to state-of-the-art on five degradations. PSNR (dB, ↑) and SSIM (↑) are reported on the full RGB images with (*) denoting general image restorers, others are specialized all-in-one approaches. Under more challenging degradations, our MoCE-IR framework outperforms prior work across two model scales. The best performances are highlighted." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 284, + 553, + 434 + ], + "lines": [ + { + "bbox": [ + 56, + 284, + 553, + 434 + ], + "spans": [ + { + "bbox": [ + 56, + 284, + 553, + 434 + ], + "type": "table", + "html": "
MethodParams.DehazingDerainingDenoisingDeblurringLow-LightAverage
SOTSRain100LBSD68σ=25GoProLOLv1
LightSwinIR* [26]1M21.50.89130.78.92330.59.86824.52.77317.81.72325.04.835
DL [13]2M20.54.82621.96.76223.09.74519.86.67219.83.71221.05.743
TAPE [27]1M22.16.86129.67.90430.18.85524.47.76318.97.62125.09.801
AirNet [24]9M21.04.88432.98.95130.91.88224.35.78118.18.73525.49.847
MoCE-IR-S (ours)11M31.33.97837.21.97831.25.88428.90.87721.68.85130.08.913
HeavyNAFNet* [6]17M25.23.93935.56.96731.02.88326.53.80820.49.80927.76.881
DGUNet* [33]17M24.78.94036.62.97131.10.88327.25.83721.87.82328.32.891
Restormer* [65]26M24.09.92734.81.96231.49.88427.22.82920.41.80627.60.881
MambaIR [16]27M25.81.94436.55.97131.41.88428.61.87522.49.83228.97.901
Transweather [49]38M21.32.88529.43.90529.00.84125.12.75721.21.79225.22.836
IDR [66]15M25.24.94335.63.96531.60.88727.87.84621.34.82628.34.893
Gridformer [51]34M26.79.95136.61.97131.45.88529.22.88422.59.83129.33.904
InstructIR-5D [9]17M27.10.95636.84.97331.40.87329.40.88623.00.83629.55.908
MoCE-IR (ours)25M30.48.97438.04.98231.34.88730.05.89923.00.85230.58.919
", + "image_path": "8927f4c4bcbe4df6683b717c7148e9f56291212ac166a7250533e535cbb44c7e.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 445, + 137, + 459 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 445, + 137, + 459 + ], + "spans": [ + { + "bbox": [ + 55, + 445, + 137, + 459 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 468, + 295, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 468, + 295, + 577 + ], + "spans": [ + { + "bbox": [ + 53, + 468, + 295, + 577 + ], + "type": "text", + "content": "We conduct experiments by strictly following previous works in general image restoration [36, 66] under two different settings: (i) All-in-One and (ii) Composited degradations. In the all-in-one setting, a unified model is trained across multiple degradation types, where we consider three and five distinct degradations. In the Composited degradations setting, a single model is trained to handle both isolated degradation occurrences as well as combinations of up to three different types of degradations." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 582, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 582, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 582, + 297, + 714 + ], + "type": "text", + "content": "Implementation Details. Our MoCE-IR framework is end-to-end trainable requiring no multi-stage optimization of any components. The architecture features an asymmetrical, four-level encoder-decoder design. The encoder consists of four levels with varying numbers of transformer blocks, arranged as " + }, + { + "bbox": [ + 54, + 582, + 297, + 714 + ], + "type": "inline_equation", + "content": "[4,6,6,8]" + }, + { + "bbox": [ + 54, + 582, + 297, + 714 + ], + "type": "text", + "content": " from the top level to the lowest. The decoder includes three levels, containing " + }, + { + "bbox": [ + 54, + 582, + 297, + 714 + ], + "type": "inline_equation", + "content": "[2,4,4]" + }, + { + "bbox": [ + 54, + 582, + 297, + 714 + ], + "type": "text", + "content": " transformer blocks, respectively. In each MoCE layer, we use " + }, + { + "bbox": [ + 54, + 582, + 297, + 714 + ], + "type": "inline_equation", + "content": "n = 4" + }, + { + "bbox": [ + 54, + 582, + 297, + 714 + ], + "type": "text", + "content": " nested experts, where the dimensionality of the expert embedding is given by " + }, + { + "bbox": [ + 54, + 582, + 297, + 714 + ], + "type": "inline_equation", + "content": "R = C / 2^i" + }, + { + "bbox": [ + 54, + 582, + 297, + 714 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 54, + 582, + 297, + 714 + ], + "type": "inline_equation", + "content": "i\\in \\{i,\\dots,n\\}" + }, + { + "bbox": [ + 54, + 582, + 297, + 714 + ], + "type": "text", + "content": ". The network's initial embedding dimension is set to " + }, + { + "bbox": [ + 54, + 582, + 297, + 714 + ], + "type": "inline_equation", + "content": "C = 32" + }, + { + "bbox": [ + 54, + 582, + 297, + 714 + ], + "type": "text", + "content": ", doubling at each" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 447, + 555, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 447, + 555, + 542 + ], + "spans": [ + { + "bbox": [ + 313, + 447, + 555, + 542 + ], + "type": "text", + "content": "subsequent level while the spatial resolution is halved. We follow the training configuration of prior work [36], and train our models for 120 epochs with a batch size of 32. We optimize the " + }, + { + "bbox": [ + 313, + 447, + 555, + 542 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 313, + 447, + 555, + 542 + ], + "type": "text", + "content": " loss in the RGB and Fourier domain using the Adam [20] optimizer " + }, + { + "bbox": [ + 313, + 447, + 555, + 542 + ], + "type": "inline_equation", + "content": "(\\beta_{1} = 0.9, \\beta_{2} = 0.999)" + }, + { + "bbox": [ + 313, + 447, + 555, + 542 + ], + "type": "text", + "content": " with an initial learning rate of " + }, + { + "bbox": [ + 313, + 447, + 555, + 542 + ], + "type": "inline_equation", + "content": "2 \\times 10^{-4}" + }, + { + "bbox": [ + 313, + 447, + 555, + 542 + ], + "type": "text", + "content": " and cosine decay. During training, we utilize crops of size " + }, + { + "bbox": [ + 313, + 447, + 555, + 542 + ], + "type": "inline_equation", + "content": "128^{2}" + }, + { + "bbox": [ + 313, + 447, + 555, + 542 + ], + "type": "text", + "content": " with horizontal and vertical flips as augmentations." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 545, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 556, + 715 + ], + "type": "text", + "content": "Datasets. For all-in-one, we follow existing work [24, 36] and include following datasets: For image denoising, we combine the BSD400 [2] and WED [31] datasets, adding Gaussian noise at levels " + }, + { + "bbox": [ + 313, + 545, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\sigma \\in [15, 25, 50]" + }, + { + "bbox": [ + 313, + 545, + 556, + 715 + ], + "type": "text", + "content": " to create noisy images. Testing is conducted on the BSD68 [32]. For deraining, we use Rain100L [56]. The dehazing task utilises the SOTS [23] dataset. For deblurring and low-light enhancement, we employ the GoPro [34] and the LOL-v1 [53] dataset, respectively. To develop a unified model for all tasks, we merge these datasets in a three (AIO-3) or five (AIO-5) degradation setting, and train for 120 epochs and directly evaluate them across different tasks. For the composite degradation setting, we use the CDD11 dataset [17] and train our method for 200 epochs, keeping all other settings unchanged." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12757" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 55, + 105, + 555, + 498 + ], + "blocks": [ + { + "bbox": [ + 55, + 71, + 555, + 103 + ], + "lines": [ + { + "bbox": [ + 55, + 71, + 555, + 103 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 555, + 103 + ], + "type": "text", + "content": "Table 3. Comparison to state-of-the-art on composited degradations. PSNR (dB, ↑) and SSIM (↑) are reported on the full RGB images with (*) denoting general image restorers, others are specialized all-in-one approaches. Our MoCE-IR method consistently outperforms even larger models, with favorable results in composited degradation scenarios." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 55, + 105, + 555, + 498 + ], + "lines": [ + { + "bbox": [ + 55, + 105, + 555, + 498 + ], + "spans": [ + { + "bbox": [ + 55, + 105, + 555, + 498 + ], + "type": "table", + "html": "
MethodParams.CDD11-SingleCDD11-DoubleCDD11-TripleAvg.
Low (L)Haze (H)Rain (R)Snow (S)L+HL+RL+SH+RH+SL+H+RL+H+S
AirNet [24]9M24.83.77824.21.95126.55.89126.79.91923.23.77922.82.71023.29.72322.21.86823.29.90121.80.70822.24.72523.75.814
PromptIR [36]36M26.32.80526.10.96931.56.94631.53.96024.49.78925.05.77124.51.76124.54.92423.70.92523.74.75223.33.74725.90.850
WGWSNet [71]26M24.39.77427.90.98233.15.96434.43.97324.27.80025.06.77224.60.76527.23.95527.65.96023.90.77223.97.77126.96.863
WeatherDiff [72]83M23.58.76321.99.90424.85.88524.80.88821.83.75622.69.73022.12.70721.25.86821.99.86821.23.71621.04.69822.49.799
OneRestore [17]6M26.48.82632.52.99033.40.96434.31.97325.79.82225.58.79925.19.78929.99.95730.21.96424.78.78824.90.79128.47.878
MoCE-IR-S (ours)11M27.26.82432.66.99034.31.97035.91.98026.24.81726.25.80026.04.79329.93.96430.19.97025.41.78925.39.79029.05.881
DehazingInputAirNetPromptIROursGT
Deraining
Denoising
Input CropsError Maps (darker means better)GT Crops
", + "image_path": "26e1aa0b976f3626d752e08df45eba2f6b226148c023cabcbba4f3dc9a923da6.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 502, + 555, + 537 + ], + "lines": [ + { + "bbox": [ + 55, + 502, + 555, + 537 + ], + "spans": [ + { + "bbox": [ + 55, + 502, + 555, + 537 + ], + "type": "text", + "content": "Figure 3. Visual results. We compare MoCE-IR-S to AirNet [24], and PromptIR [36] in the all-in-one setting with three degradations. MoCE-IR-S effectively removes haze and rain streaks while preserving image sharpness, achieving high-quality restoration. An error heatmap is provided, with color transitioning from black to white to indicate increasing pixel-wise error." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 542, + 269, + 555 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 542, + 269, + 555 + ], + "spans": [ + { + "bbox": [ + 55, + 542, + 269, + 555 + ], + "type": "text", + "content": "4.1. Comparison to State-of-the-Art Methods" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 570, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 570, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 570, + 297, + 714 + ], + "type": "text", + "content": "All-in-One: Three Degradations. We compare our all-in-one restoration approach with specialized restoration methods, including both image-only models, such as MPR-Net [64], AirNet [24], PromptIR [36], and Gridformer [51], and vision-language models like DA-CLIP [30], InstructIR [9], and UniProcessor [12]. All methods were trained simultaneously on three types of degradation: haze, rain, and Gaussian noise. The results in Tab. 1 show that our proposed model, MoCE-IR, consistently outperforms previous works, both in light and heavy model scales. In the light category, MoCE-IR achieves the highest scores across all benchmarks with an average improvement in PSNR of" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 543, + 555, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 543, + 555, + 591 + ], + "spans": [ + { + "bbox": [ + 313, + 543, + 555, + 591 + ], + "type": "text", + "content": "1.37 dB, surpassing other models like AirNet [24]. In the heavy scale, MoCE-IR competes closely with UniProcessor [12], achieving an average PSNR of 32.73 dB with " + }, + { + "bbox": [ + 313, + 543, + 555, + 591 + ], + "type": "inline_equation", + "content": "98\\%" + }, + { + "bbox": [ + 313, + 543, + 555, + 591 + ], + "type": "text", + "content": " fewer parameters, validating our efficiency and effectiveness." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "content": "All-in-One: Five Degradations. Based upon [24, 66], we expand the three-degradation setting to include deblurring and low-light image enhancement, further validating our method's effectiveness. As shown in Tab. 2, our approach excels by learning specialized experts with input-adaptive complexity and maximizing synergies across degradations. Our framework outperforms previous methods on the lightest model scale, with an average PSNR improvement of 4.31 dB over AirNet [24]. Even compared to larger networks, our lightest model surpasses the VLM-based InstructIR [9] by" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12758" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 72, + 294, + 175 + ], + "blocks": [ + { + "bbox": [ + 59, + 72, + 294, + 175 + ], + "lines": [ + { + "bbox": [ + 59, + 72, + 294, + 175 + ], + "spans": [ + { + "bbox": [ + 59, + 72, + 294, + 175 + ], + "type": "image", + "image_path": "4054f08cddc89f8dc03a0e68ef767cca245144c4b5b439a6ef161def7dbc8169.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 180, + 296, + 226 + ], + "lines": [ + { + "bbox": [ + 55, + 180, + 296, + 226 + ], + "spans": [ + { + "bbox": [ + 55, + 180, + 296, + 226 + ], + "type": "text", + "content": "Figure 4. Complexity-efficiency tradeoff. Visualization of PSNR and parameter counts of proposed method compared to prior work. Proposed MoCE-IR surpasses prior methods, achieving SoTA results in all-in-one image restoration with enhanced efficiency." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 56, + 251, + 296, + 312 + ], + "blocks": [ + { + "bbox": [ + 55, + 228, + 296, + 251 + ], + "lines": [ + { + "bbox": [ + 55, + 228, + 296, + 251 + ], + "spans": [ + { + "bbox": [ + 55, + 228, + 296, + 251 + ], + "type": "text", + "content": "Table 4. Computational demands. GFLOPS are computed on an input image of size " + }, + { + "bbox": [ + 55, + 228, + 296, + 251 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 55, + 228, + 296, + 251 + ], + "type": "text", + "content": " using a NVIDIA RTX 4090 GPU." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 251, + 296, + 312 + ], + "lines": [ + { + "bbox": [ + 56, + 251, + 296, + 312 + ], + "spans": [ + { + "bbox": [ + 56, + 251, + 296, + 312 + ], + "type": "table", + "html": "
MethodParams.MemoryFLOPSRuntime
AirNet [24]8.93M4829M238G42.17 ± 0.23
PromptIR [36]35.59M9830M132G41.28 ± 0.43
IDR [66]15.34M4905M98G-
MoCE-IR (ours)25.35M5887M80.59 ± 5.21G23.36 ± 2.47
MoCE-IR-S (ours)11.47M4228M36.93 ± 2.32G22.15 ± 2.59
", + "image_path": "be61f764f8f35449e2e76073f400a4a8868f18fd70cc69e520321ed5337e051b.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 328, + 296, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 328, + 296, + 353 + ], + "spans": [ + { + "bbox": [ + 55, + 328, + 296, + 353 + ], + "type": "text", + "content": "0.33 dB on average, and on a comparable model scale, our framework extends this advantage by an additional 0.50 dB." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 359, + 297, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 359, + 297, + 529 + ], + "spans": [ + { + "bbox": [ + 54, + 359, + 297, + 529 + ], + "type": "text", + "content": "All-in-One: Composed Degradations. To simulate more realistic degradation scenarios, recent work [17] has trained all-in-one restorers not only on a mix of different degradations but also on more challenging cases where multiple degradation types are combined within a single image. We extend the previous AIO settings by including rain, haze, snow, and low illumination as individual degradation types, along with various composite degradation scenarios, resulting in a total of eleven distinct restoration settings. As shown in Tab. 3, our method outperforms previous all-in-one restorers by an average improvement of " + }, + { + "bbox": [ + 54, + 359, + 297, + 529 + ], + "type": "inline_equation", + "content": "0.64\\mathrm{dB}" + }, + { + "bbox": [ + 54, + 359, + 297, + 529 + ], + "type": "text", + "content": " over OneRestore [17], while being among the smallest models in size. This further validates our effectiveness with adaptive modeling capacity based on input complexity." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 534, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 297, + 714 + ], + "type": "text", + "content": "Visual Results. We show visual results obtained under the three degradation settings in Fig. 3. In certain demanding hazy scenarios, both AirNet [24] and PromptIR [36] reveal constraints in completely eradicating haziness, resulting in noticeable color intensity discrepancies, whereas our approach ensures precise color reconstruction. Furthermore, in challenging rainy scenes, these popular approaches continue to exhibit notable remnants of rain, which our method adeptly eliminates, distinguishing itself from other approaches. Additionally, our method also generates clear and sharp denoised outputs. We further provide the error maps for visual differences where our method contains less noticeable errors. Coupled with the quantitative comparisons, these findings underscore the effectiveness of our method. Please refer to the supplementary for more visuals." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 70, + 555, + 148 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 70, + 555, + 148 + ], + "spans": [ + { + "bbox": [ + 313, + 70, + 555, + 148 + ], + "type": "text", + "content": "Table 5. Architecture analysis. (a) We investigate scaling the expert embedding " + }, + { + "bbox": [ + 313, + 70, + 555, + 148 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 313, + 70, + 555, + 148 + ], + "type": "text", + "content": " and the bias normalization for MoCE-IR-S. (b) We compare MoCE to several baselines [29, 43, 58], demonstrating that our complexity experts consistently outperform them. Notably, complexity-biased routing further boosts performance, surpassing standard load balancing (LB) [43]. PSNR (dB, ↑) and SSIM (↑) are reported on full RGB images." + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 315, + 160, + 553, + 284 + ], + "blocks": [ + { + "bbox": [ + 378, + 149, + 489, + 159 + ], + "lines": [ + { + "bbox": [ + 378, + 149, + 489, + 159 + ], + "spans": [ + { + "bbox": [ + 378, + 149, + 489, + 159 + ], + "type": "text", + "content": "(a) Ablation on the expert scaling." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 160, + 553, + 284 + ], + "lines": [ + { + "bbox": [ + 315, + 160, + 553, + 284 + ], + "spans": [ + { + "bbox": [ + 315, + 160, + 553, + 284 + ], + "type": "table", + "html": "
Embedding r scalingBias normalizationAvg. AIO-3
NestedpMin32.44.913
ExponentialpMax32.47.914
NestedpMax32.57.916
(b) Ablation on the complexity-aware routing.
MethodSOTSRain100LBSD6815-50Avg.
MoCE + Complexity Bias (ours)30.94.97938.22.98331.22.873
Baseline (MoE + LB [45])30.28.97737.77.98131.14.871
w/ MoCE + LB [45]30.13.97838.24.98331.19.872
w/ AMIR [58]29.97.97537.90.98131.17.872
w/ WM-MoE [29]30.10.97636.72.97631.05.869
", + "image_path": "f0aa766f6b63e5170f535a8b98b9f0ba537e0b168935c8cdcbf8caabd7a7f6e4.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 292, + 416, + 304 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 292, + 416, + 304 + ], + "spans": [ + { + "bbox": [ + 313, + 292, + 416, + 304 + ], + "type": "text", + "content": "4.2. Ablation Studies" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 310, + 556, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 310, + 556, + 370 + ], + "spans": [ + { + "bbox": [ + 313, + 310, + 556, + 370 + ], + "type": "text", + "content": "We conduct detailed studies on the components within our method, highlighting the emergence of task-discriminative learners. All experiments are conducted in the all-in-one setting with three degradations. Please refer to the supplementary for additional ablations." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 374, + 556, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 374, + 556, + 554 + ], + "spans": [ + { + "bbox": [ + 313, + 374, + 556, + 554 + ], + "type": "text", + "content": "**Complexity-Efficiency Trade-off.** Our MoCE-IR framework addresses the critical challenge of scalability in vision tasks by optimizing both reconstruction quality and computational efficiency. We visualize in Fig. 4 the progressive improvements in reconstruction fidelity of our framework on two complexity scales. Furthermore, Tab. 4 validates that our approach significantly outperforms SoTA methods in terms of memory utilization. Unlike traditional MoE architectures with fixed computational demands, our complexity experts enable dynamic FLOPS allocation. For instance, our lightweight model averages 36.9 GFLOPS, while the heavyweight model averages 80.59 GFLOPS. This adaptive computation, combined with lower memory consumption and parameter counts, gives MoCE-IR a substantial advantage over contemporary methods." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 558, + 557, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 557, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 557, + 715 + ], + "type": "text", + "content": "Expert Scaling. As shown in Tab. 5a, we systematically evaluate different scaling strategies regarding the embedding dimensionality " + }, + { + "bbox": [ + 313, + 558, + 557, + 715 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 313, + 558, + 557, + 715 + ], + "type": "text", + "content": " of each complexity expert. Our investigation compares exponential and nested parameter scaling, where the window partition is always fixed and follows a " + }, + { + "bbox": [ + 313, + 558, + 557, + 715 + ], + "type": "inline_equation", + "content": "2^{(2 + i)}" + }, + { + "bbox": [ + 313, + 558, + 557, + 715 + ], + "type": "text", + "content": " progression. Further, we investigate the bias normalization, whether normalizing to the lowest or highest expert capacity. Overall, the nested parameter scaling emerges as the optimal choice, providing sufficient expert channel diversity while maintaining reasonable computational costs. Normalizing router weights by the highest expert capacity, rather than the minimum, proves crucial for performance as minimum capacity normalization forces router weights" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "12759" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 65, + 72, + 192, + 159 + ], + "blocks": [ + { + "bbox": [ + 65, + 72, + 192, + 159 + ], + "lines": [ + { + "bbox": [ + 65, + 72, + 192, + 159 + ], + "spans": [ + { + "bbox": [ + 65, + 72, + 192, + 159 + ], + "type": "image", + "image_path": "de8b9b36e22e38c7ca1958a064c3aee7684308a39347bd4dbc5d731bad47f8db.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 162, + 555, + 219 + ], + "lines": [ + { + "bbox": [ + 55, + 162, + 555, + 219 + ], + "spans": [ + { + "bbox": [ + 55, + 162, + 555, + 219 + ], + "type": "text", + "content": "Figure 5. Routing visualization for the AIO-3 setting. (a) While load balancing [43] ensures uniform expert utilization, it neglects shared task dependencies and task-specific characteristics, limiting restoration quality. (b)-(d) Complexity-aware routing fosters task discrimination by directing complex degradations to experts with broader contextual understanding and vice versa. This allows some experts to generalize across tasks while others specialize, enhancing adaptability. We visualize the average decisions made by each router for dehazing, deraining and denoising. The y-axis indicates increasing expert complexity." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 195, + 72, + 309, + 159 + ], + "blocks": [ + { + "bbox": [ + 195, + 72, + 309, + 159 + ], + "lines": [ + { + "bbox": [ + 195, + 72, + 309, + 159 + ], + "spans": [ + { + "bbox": [ + 195, + 72, + 309, + 159 + ], + "type": "image", + "image_path": "5625323f40b572039dc3c40b879f2066643f57312f94319dfb78d6e15058b691.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 312, + 72, + 428, + 159 + ], + "blocks": [ + { + "bbox": [ + 312, + 72, + 428, + 159 + ], + "lines": [ + { + "bbox": [ + 312, + 72, + 428, + 159 + ], + "spans": [ + { + "bbox": [ + 312, + 72, + 428, + 159 + ], + "type": "image", + "image_path": "6183b41c6ff9a9faa9bac3b25e2d0a586dfd507f0e5491de087fe9d94342db88.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 431, + 72, + 546, + 159 + ], + "blocks": [ + { + "bbox": [ + 431, + 72, + 546, + 159 + ], + "lines": [ + { + "bbox": [ + 431, + 72, + 546, + 159 + ], + "spans": [ + { + "bbox": [ + 431, + 72, + 546, + 159 + ], + "type": "image", + "image_path": "de6dd3673d508523700cdc658412cf2af71d9236496cd0646803af38d99699e6.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 56, + 247, + 296, + 312 + ], + "blocks": [ + { + "bbox": [ + 55, + 223, + 295, + 245 + ], + "lines": [ + { + "bbox": [ + 55, + 223, + 295, + 245 + ], + "spans": [ + { + "bbox": [ + 55, + 223, + 295, + 245 + ], + "type": "text", + "content": "Table 6. Expert generalization. We report PSNR for MoCE-IR-S with expert assignments fixed post-training. " + }, + { + "bbox": [ + 55, + 223, + 295, + 245 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 55, + 223, + 295, + 245 + ], + "type": "text", + "content": " denotes expert." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 247, + 296, + 312 + ], + "lines": [ + { + "bbox": [ + 56, + 247, + 296, + 312 + ], + "spans": [ + { + "bbox": [ + 56, + 247, + 296, + 312 + ], + "type": "table", + "html": "
DegradationRoutingLearned ChoiceManual Choice
\\( \\mathcal{E}_1 \\)\\( \\mathcal{E}_2 \\)\\( \\mathcal{E}_3 \\)\\( \\mathcal{E}_4 \\)
Rain NoiseLoad Balance [45]Not Applicable28.2733.3328.2433.3728.2533.3728.2733.40
Rain NoiseComplexity Bias (ours)\\( \\mathcal{E}_1 \\)30.4530.2129.7324.65
\\( \\mathcal{E}_4 \\)33.9233.9333.9334.00
", + "image_path": "a578ba8e6e69589f5c08684298fe5ad42fe8233ec52e5b28b3099ccca554bb05.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 321, + 296, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 321, + 296, + 346 + ], + "spans": [ + { + "bbox": [ + 55, + 321, + 296, + 346 + ], + "type": "text", + "content": "of higher capacity experts to escalate uncontrollably, thus excluding them from the gating mechanism." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 349, + 296, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 349, + 296, + 517 + ], + "spans": [ + { + "bbox": [ + 54, + 349, + 296, + 517 + ], + "type": "text", + "content": "Impact of Complexity-aware Routing. Building on the findings from Tab. 5a, we adopt nested expert scaling and evaluate our complexity-aware routing strategy against conventional load balancing [43], which enforces uniform expert usage. As shown in Tab. 5b, our MoCE-IR framework achieves superior performance by preferentially routing to lower-complexity experts, activating higher-complexity experts only when their added capacity provides clear benefits. This efficient model utilization not only boosts image restoration performance, consistently surpassing related adaptive restoration methods [29, 58], but also enhances out-of-distribution generalization as MoCE-IR effectively handles degradations that with significant deviation from the training data; see the supplementary for more evaluations." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 521, + 296, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 296, + 640 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 296, + 640 + ], + "type": "text", + "content": "Routing Analysis. Fig. 5 illustrates the average expert selection per layer for different degradations: haze, rain, and noise. Higher-complexity experts are favored for tasks requiring broader context, like haze and noise, while lighter experts are chosen for rain streak removal. Notably, some experts handle multiple tasks, while others remain task-specific. Our complexity-conditioned routing promotes task-discriminative learning by directing inputs to relevant experts, preventing the collapse into a single-expert dependency, ensuring efficient and diverse operation." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "type": "text", + "content": "The baseline MoE [43] approach, lacking complexity awareness and scaled experts, uniformly distributes degraded samples across all experts. Our results show that this uniform allocation is suboptimal for all-in-one restoration, as it fails to leverage task-specific commonalities and distinctions effectively. Tab. 6 highlights the limitations of standard" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 224, + 555, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 224, + 555, + 297 + ], + "spans": [ + { + "bbox": [ + 313, + 224, + 555, + 297 + ], + "type": "text", + "content": "MoEs [43], where fixed post-training expert assignments result in uniform, task-agnostic performance, offering only moderate cross-task generalization. In contrast, MoCEs effectively generalize across degradations while exhibiting task-discriminative behavior, routing inputs to the most suitable experts for optimal performance." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 308, + 531, + 320 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 308, + 531, + 320 + ], + "spans": [ + { + "bbox": [ + 313, + 308, + 531, + 320 + ], + "type": "text", + "content": "5. Conclusion, Limitations and Discussion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 327, + 555, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 327, + 555, + 422 + ], + "spans": [ + { + "bbox": [ + 313, + 327, + 555, + 422 + ], + "type": "text", + "content": "This paper presents MoCE-IR, an efficient all-in-one image restoration model leveraging complexity experts, specialized units with varying computational demands and receptive fields. By aligning degradation tasks with suitable resources, MoCE-IR overcomes limitations of prior methods. Its built-in bias toward simpler paths enhances task-discriminative learning, enabling swift inference by bypassing irrelevant experts without sacrificing restoration quality." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 423, + 555, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 423, + 555, + 483 + ], + "spans": [ + { + "bbox": [ + 313, + 423, + 555, + 483 + ], + "type": "text", + "content": "Extensive experiments across diverse degradation settings, including more demanding scenarios with multiple composited degradations, demonstrate our computational efficiency through selective expert utilization and consistent improvement over state-of-the-art methods." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 483, + 556, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 483, + 556, + 604 + ], + "spans": [ + { + "bbox": [ + 313, + 483, + 556, + 604 + ], + "type": "text", + "content": "While MoCE-IR shows substantial improvements, opportunities remain. The current image-level routing imposes scalability constraints. Future work could explore SoftMoE [37] architectures for fine-grained spatial routing or early-exiting approaches [39] to further enhance processing speeds and input adaptability. Extending to synthetic-to-real adaptation, especially for video, represents a crucial next step. Performance optimization through mixed-precision computation across experts could further enhance speed and efficiency for resource-limited deployments." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 605, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 556, + 713 + ], + "type": "text", + "content": "Acknowledgements. The authors sincerely thank the reviewers and all members of the program committee for their tremendous efforts and incisive feedback. This research was supported by the Alexander von Humboldt Foundation, the Ministry of Education and Science of Bulgaria (support for INSAIT, under the Bulgarian National Roadmap for Research Infrastructure), the Shanghai Municipal Science and Technology Major Project (2021SHZDZX0102), and the Fundamental Research Funds for the Central Universities." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "12760" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 297, + 713 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 61, + 91, + 297, + 134 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 297, + 134 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 297, + 134 + ], + "type": "text", + "content": "[1] Yuang Ai, Huaibo Huang, Xiaogiang Zhou, Jiexiang Wang, and Ran He. Multimodal prompt perceiver: Empower adaptiveness generalizability and fidelity for all-in-one image restoration. In CVPR, 2024. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 135, + 296, + 167 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 135, + 296, + 167 + ], + "spans": [ + { + "bbox": [ + 61, + 135, + 296, + 167 + ], + "type": "text", + "content": "[2] Pablo Arbelaez, Michael Maire, Charless Fowlkes, and Jitendra Malik. Contour detection and hierarchical image segmentation. TPAMI, 33(5):898-916, 2010. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 168, + 296, + 212 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 168, + 296, + 212 + ], + "spans": [ + { + "bbox": [ + 62, + 168, + 296, + 212 + ], + "type": "text", + "content": "[3] Yoshua Bengio, Nicholas Léonard, and Aaron Courville. Estimating or propagating gradients through stochastic neurons for conditional computation. arXiv preprint arXiv:1308.3432, 2013. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 213, + 296, + 235 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 213, + 296, + 235 + ], + "spans": [ + { + "bbox": [ + 62, + 213, + 296, + 235 + ], + "type": "text", + "content": "[4] Dana Berman and Shai Avidan. Non-local image dehazing. In CVPR, 2016. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 236, + 296, + 269 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 236, + 296, + 269 + ], + "spans": [ + { + "bbox": [ + 62, + 236, + 296, + 269 + ], + "type": "text", + "content": "[5] Bolun Cai, Xiangmin Xu, Kui Jia, Chunmei Qing, and Dacheng Tao. Dehazenet: An end-to-end system for single image haze removal. TIP, 25(11):5187-5198, 2016. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 270, + 296, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 270, + 296, + 300 + ], + "spans": [ + { + "bbox": [ + 62, + 270, + 296, + 300 + ], + "type": "text", + "content": "[6] Liangyu Chen, Xiaojie Chu, Xiangyu Zhang, and Jian Sun. Simple baselines for image restoration. In ECCV, 2022. 1, 2, 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 302, + 296, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 302, + 296, + 335 + ], + "spans": [ + { + "bbox": [ + 62, + 302, + 296, + 335 + ], + "type": "text", + "content": "[7] Yu-Wei Chen and Soo-Chang Pei. Always clear days: Degradation type and severity aware all-in-one adverse weather removal. arXiv preprint arXiv:2310.18293, 2023. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 336, + 296, + 368 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 336, + 296, + 368 + ], + "spans": [ + { + "bbox": [ + 62, + 336, + 296, + 368 + ], + "type": "text", + "content": "[8] Zheng Chen, Yulun Zhang, Jinjin Gu, Linghe Kong, Xin Yuan, et al. Cross aggregation transformer for image restoration. NeurIPS, 2022. 1, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 369, + 296, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 369, + 296, + 402 + ], + "spans": [ + { + "bbox": [ + 62, + 369, + 296, + 402 + ], + "type": "text", + "content": "[9] Marcos V Conde, Gregor Geigle, and Radu Timofte. Instructir: High-quality image restoration following human instructions. In ECCV, 2024. 1, 2, 3, 5, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 403, + 296, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 403, + 296, + 445 + ], + "spans": [ + { + "bbox": [ + 57, + 403, + 296, + 445 + ], + "type": "text", + "content": "[10] Xin Dang, Hai Wang, Jie Ren, and Le Chen. An application performance optimization model of mobile augmented reality based on hd restoration. In CBD, pages 201-206. IEEE, 2020. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 447, + 296, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 447, + 296, + 480 + ], + "spans": [ + { + "bbox": [ + 57, + 447, + 296, + 480 + ], + "type": "text", + "content": "[11] Yu Dong, Yihao Liu, He Zhang, Shifeng Chen, and Yu Qiao. Fd-gan: Generative adversarial networks with fusion-discriminator for single image dehazing. In AAAI, 2020. 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 481, + 296, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 481, + 296, + 513 + ], + "spans": [ + { + "bbox": [ + 57, + 481, + 296, + 513 + ], + "type": "text", + "content": "[12] Huiyu Duan, Xiongkuo Min, Sijing Wu, Wei Shen, and Guangtao Zhai. Uniprocessor: A text-induced unified low-level image processor. In ECCV, 2024. 1, 3, 5, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 514, + 296, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 514, + 296, + 557 + ], + "spans": [ + { + "bbox": [ + 57, + 514, + 296, + 557 + ], + "type": "text", + "content": "[13] Qingnan Fan, Dongdong Chen, Lu Yuan, Gang Hua, Nenghai Yu, and Baoquan Chen. A general decoupled learning framework for parameterized image operators. TPAMI, 43(1): 33-47, 2019. 1, 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 558, + 296, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 558, + 296, + 591 + ], + "spans": [ + { + "bbox": [ + 57, + 558, + 296, + 591 + ], + "type": "text", + "content": "[14] Hongyun Gao, Xin Tao, Xiaoyong Shen, and Jiaya Jia. Dynamic scene deblurring with parameter selective sharing and nested skip connections. In CVPR, 2019. 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 592, + 296, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 592, + 296, + 645 + ], + "spans": [ + { + "bbox": [ + 57, + 592, + 296, + 645 + ], + "type": "text", + "content": "[15] Florin Girbacia, Silviu Butnariu, Alex Petre Orman, and Cristian Cezar Postelnicu. Virtual restoration of deteriorated religious heritage objects using augmented reality technologies. European Journal of Science and Theology, 9(2):223-231, 2013. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 57, + 647, + 296, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 296, + 680 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 296, + 680 + ], + "type": "text", + "content": "[16] Hang Guo, Jinmin Li, Tao Dai, Zhihao Ouyang, Xudong Ren, and Shu-Tao Xia. Mambair: A simple baseline for image restoration with state-space model. In ECCV, 2024. 1, 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 57, + 681, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 681, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 681, + 296, + 713 + ], + "type": "text", + "content": "[17] Yu Guo, Yuan Gao, Yuxu Lu, Ryan Wen Liu, and Shengfeng He. Onerestore: A universal restoration framework for composite degradation. In ECCV, 2024. 5, 6, 7" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 35, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 116 + ], + "type": "text", + "content": "[18] Kui Jiang, Zhongyuan Wang, Peng Yi, Chen Chen, Baojin Huang, Yimin Luo, Jiayi Ma, and Junjun Jiang. Multi-scale progressive fusion network for single image deraining. In CVPR, 2020. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 118, + 554, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 118, + 554, + 152 + ], + "spans": [ + { + "bbox": [ + 317, + 118, + 554, + 152 + ], + "type": "text", + "content": "[19] Yitong Jiang, Zhaoyang Zhang, Tianfan Xue, and Jinwei Gu. Autodir: Automatic all-in-one image restoration with latent diffusion. arXiv preprint arXiv:2310.10123, 2023. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 153, + 554, + 175 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 153, + 554, + 175 + ], + "spans": [ + { + "bbox": [ + 316, + 153, + 554, + 175 + ], + "type": "text", + "content": "[20] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv, 2017. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 177, + 555, + 220 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 177, + 555, + 220 + ], + "spans": [ + { + "bbox": [ + 316, + 177, + 555, + 220 + ], + "type": "text", + "content": "[21] Lingshun Kong, Jiangxin Dong, Jianjun Ge, Mingqiang Li, and Jinshan Pan. Efficient frequency domain-based transformers for high-quality image deblurring. In CVPR, 2023. 2, 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 222, + 555, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 222, + 555, + 266 + ], + "spans": [ + { + "bbox": [ + 316, + 222, + 555, + 266 + ], + "type": "text", + "content": "[22] Jaakko Lehtinen, Jacob Munkberg, Jon Hasselgren, Samuli Laine, Tero Karras, Miika Aittala, and Timo Aila. Noise2noise: Learning image restoration without clean data. arXiv preprint arXiv:1803.04189, 2018. 1, 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 268, + 555, + 301 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 268, + 555, + 301 + ], + "spans": [ + { + "bbox": [ + 316, + 268, + 555, + 301 + ], + "type": "text", + "content": "[23] Boyi Li, Wenqi Ren, Dengpan Fu, Dacheng Tao, Dan Feng, Wenjun Zeng, and Zhangyang Wang. Benchmarking single-image dehazing and beyond. TIP, 28(1):492-505, 2018. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 303, + 554, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 303, + 554, + 335 + ], + "spans": [ + { + "bbox": [ + 316, + 303, + 554, + 335 + ], + "type": "text", + "content": "[24] Boyun Li, Xiao Liu, Peng Hu, Zhongqin Wu, Jiancheng Lv, and Xi Peng. All-In-One Image Restoration for Unknown Corruption. In CVPR, 2022. 1, 2, 5, 6, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 337, + 554, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 337, + 554, + 380 + ], + "spans": [ + { + "bbox": [ + 316, + 337, + 554, + 380 + ], + "type": "text", + "content": "[25] Zilong Li, Yiming Lei, Chenglong Ma, Junping Zhang, and Hongming Shan. Prompt-in-prompt learning for universal image restoration. arXiv preprint arXiv:2312.05038, 2023. 1, 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 383, + 554, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 383, + 554, + 415 + ], + "spans": [ + { + "bbox": [ + 316, + 383, + 554, + 415 + ], + "type": "text", + "content": "[26] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Swinir: Image restoration using swim transformer. In ICCV, 2021. 2, 5" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 418, + 555, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 418, + 555, + 460 + ], + "spans": [ + { + "bbox": [ + 316, + 418, + 555, + 460 + ], + "type": "text", + "content": "[27] Lin Liu, Lingxi Xie, Xiaopeng Zhang, Shanxin Yuan, Xiangyu Chen, Wengang Zhou, Houqiang Li, and Qi Tian. Tape: Task-agnostic prior embedding for image restoration. In ECCV, 2022. 1, 5" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 317, + 463, + 554, + 506 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 463, + 554, + 506 + ], + "spans": [ + { + "bbox": [ + 317, + 463, + 554, + 506 + ], + "type": "text", + "content": "[28] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, 2021. 1, 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 317, + 509, + 554, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 509, + 554, + 564 + ], + "spans": [ + { + "bbox": [ + 317, + 509, + 554, + 564 + ], + "type": "text", + "content": "[29] Yulin Luo, Rui Zhao, Xiaobao Wei, Jinwei Chen, Yijie Lu, Shenghao Xie, Tianyu Wang, Ruiqin Xiong, Ming Lu, and Shanghang Zhang. Wm-moe: Weather-aware multi-scale mixture-of-experts for blind adverse weather removal. arXiv preprint arXiv:2303.13739, 2023. 2, 3, 7, 8" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 565, + 554, + 598 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 565, + 554, + 598 + ], + "spans": [ + { + "bbox": [ + 316, + 565, + 554, + 598 + ], + "type": "text", + "content": "[30] Ziwei Luo, Fredrik K Gustafsson, Zheng Zhao, Jens Sjolund, and Thomas B Schon. Controlling vision-language models for multi-task image restoration. In ICLR, 2024. 2, 3, 5, 6" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 316, + 600, + 554, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 600, + 554, + 643 + ], + "spans": [ + { + "bbox": [ + 316, + 600, + 554, + 643 + ], + "type": "text", + "content": "[31] Kede Ma, Zhengfang Duanmu, Qingbo Wu, Zhou Wang, Hongwei Yong, Hongliang Li, and Lei Zhang. Waterloo exploration database: New challenges for image quality assessment models. TIP, 26(2):1004-1016, 2016. 5" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 316, + 646, + 554, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 646, + 554, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 646, + 554, + 689 + ], + "type": "text", + "content": "[32] David Martin, Charless Fowlkes, Doron Tal, and Jitendra Malik. A database of human segmented natural images and its application to evaluating segmentation algorithms and measuring ecological statistics. In ICCV, 2001. 5" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 316, + 691, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 554, + 713 + ], + "type": "text", + "content": "[33] Chong Mou, Qian Wang, and Jian Zhang. Deep generalized unfolding networks for image restoration. In CVPR, 2022. 5" + } + ] + } + ], + "index": 34 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12761" + } + ] + } + ], + "index": 36 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "type": "text", + "content": "[34] Seungjun Nah, Tae Hyun Kim, and Kyoung Mu Lee. Deep multi-scale convolutional neural network for dynamic scene deblurring. In CVPR, 2017. 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 296, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 296, + 139 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 296, + 139 + ], + "type": "text", + "content": "[35] Jinshan Pan, Deqing Sun, Hans Pfister, and Ming-Hsuan Yang. Blind image deblurring using dark channel prior. In CVPR, 2016. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 141, + 294, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 141, + 294, + 174 + ], + "spans": [ + { + "bbox": [ + 56, + 141, + 294, + 174 + ], + "type": "text", + "content": "[36] Vaishnav Potlapalli, Syed Waqas Zamir, Salman Khan, and Fahad Khan. Prompt: Prompting for all-in-one image restoration. In NeurIPS, 2023. 1, 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 175, + 294, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 294, + 207 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 294, + 207 + ], + "type": "text", + "content": "[37] Joan Puigcerver, Carlos Riquelme Ruiz, Basil Mustafa, and Neil Houlsby. From sparse to soft mixtures of experts. In ICLR, 2024. 2, 3, 4, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 57, + 209, + 296, + 231 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 209, + 296, + 231 + ], + "spans": [ + { + "bbox": [ + 57, + 209, + 296, + 231 + ], + "type": "text", + "content": "[38] Yanyun Qu, Yizi Chen, Jingying Huang, and Yuan Xie. Enhanced pix2pix dehazing network. In CVPR, 2019. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 57, + 232, + 296, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 232, + 296, + 274 + ], + "spans": [ + { + "bbox": [ + 57, + 232, + 296, + 274 + ], + "type": "text", + "content": "[39] David Raposo, Sam Ritter, Blake Richards, Timothy Lillicrap, Peter Conway Humphreys, and Adam Santoro. Mixture-of-depths: Dynamically allocating compute in transformer-based language models, 2024. 2, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 277, + 296, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 277, + 296, + 308 + ], + "spans": [ + { + "bbox": [ + 56, + 277, + 296, + 308 + ], + "type": "text", + "content": "[40] Dongwei Ren, Wangmeng Zuo, Qinghua Hu, Pengfei Zhu, and Deyu Meng. Progressive image deraining networks: A better and simpler baseline. In CVPR, 2019. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 57, + 310, + 294, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 310, + 294, + 342 + ], + "spans": [ + { + "bbox": [ + 57, + 310, + 294, + 342 + ], + "type": "text", + "content": "[41] Wenqi Ren, Lin Ma, Jiawei Zhang, Jinshan Pan, Xiaochun Cao, Wei Liu, and Ming-Hsuan Yang. Gated fusion network for single image dehazing. In CVPR, 2018. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 344, + 296, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 344, + 296, + 387 + ], + "spans": [ + { + "bbox": [ + 56, + 344, + 296, + 387 + ], + "type": "text", + "content": "[42] Wenqi Ren, Jinshan Pan, Hua Zhang, Xiaochun Cao, and Ming-Hsuan Yang. Single image dehazing via multi-scale convolutional neural networks with holistic edges. In ICCV, 2020. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 388, + 294, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 388, + 294, + 432 + ], + "spans": [ + { + "bbox": [ + 56, + 388, + 294, + 432 + ], + "type": "text", + "content": "[43] Carlos Riquelme, Joan Puigcerver, Basil Mustafa, Maxim Neumann, Rodolphe Jenatton, André Susano Pinto, Daniel Keysers, and Neil Houlsby. Scaling vision with sparse mixture of experts. *NeurIPS*, 2021. 2, 3, 4, 7, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 434, + 296, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 434, + 296, + 466 + ], + "spans": [ + { + "bbox": [ + 56, + 434, + 296, + 466 + ], + "type": "text", + "content": "[44] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-Net: Convolutional networks for biomedical image segmentation. In MICCAI, pages 234–241. Springer, 2015. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 468, + 294, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 468, + 294, + 511 + ], + "spans": [ + { + "bbox": [ + 56, + 468, + 294, + 511 + ], + "type": "text", + "content": "[45] Giovanni Saggio, Davide Borra, et al. Augmented reality for restoration/reconstruction of artefacts with artistic or historical value. In Augmented reality: some emerging application areas, pages 59-86. InTech Publications, 2011. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 512, + 296, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 512, + 296, + 555 + ], + "spans": [ + { + "bbox": [ + 56, + 512, + 296, + 555 + ], + "type": "text", + "content": "[46] Noam Shazeer, Azalia Mirhoseini, Krzysztof Maziarz, Andy Davis, Quoc Le, Geoffrey Hinton, and Jeff Dean. Outrageously large neural networks: The sparsely-gated mixture-of-experts layer, 2017. 3, 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 557, + 296, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 557, + 296, + 589 + ], + "spans": [ + { + "bbox": [ + 56, + 557, + 296, + 589 + ], + "type": "text", + "content": "[47] Ying Tai, Jian Yang, Xiaoming Liu, and Chunyan Xu. Memnet: A persistent memory network for image restoration. In ICCV, 2017. 1, 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 591, + 296, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 591, + 296, + 623 + ], + "spans": [ + { + "bbox": [ + 56, + 591, + 296, + 623 + ], + "type": "text", + "content": "[48] Chunwei Tian, Yong Xu, and Wangmeng Zuo. Image denoising using deep cnn with batch renormalization. NN, 2020. 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 56, + 624, + 296, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 296, + 667 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 296, + 667 + ], + "type": "text", + "content": "[49] Jeya Maria Jose Valanarasu, Rajeev Yasarla, and Vishal M Patel. Transweather: Transformer-based restoration of images degraded by adverse weather conditions. In CVPR, 2022. 1, 2, 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 56, + 670, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 296, + 713 + ], + "type": "text", + "content": "[50] Cong Wang, Jinshan Pan, Wei Wang, Jiangxin Dong, Mengzhu Wang, Yakun Ju, and Junyang Chen. Promptstorer: A prompting image restoration method with degradation perception. In NeurIPS, 2023. 1, 2" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 34, + "blocks": [ + { + "bbox": [ + 316, + 72, + 555, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 555, + 127 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 555, + 127 + ], + "type": "text", + "content": "[51] Tao Wang, Kaihao Zhang, Ziqian Shao, Wenhan Luo, Bjorn Stenger, Tong Lu, Tae-Kyun Kim, Wei Liu, and Hongdong Li. Gridformer: Residual dense transformer with grid structure for image restoration in adverse weather conditions. ICCV, 2024. 5, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 129, + 555, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 555, + 172 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 555, + 172 + ], + "type": "text", + "content": "[52] Zhendong Wang, Xiaodong Cun, Jianmin Bao, Wengang Zhou, Jianzhuang Liu, and Houqiang Li. Uformer: A general u-shaped transformer for image restoration. In CVPR, 2022. 1, 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 175, + 555, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 555, + 206 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 555, + 206 + ], + "type": "text", + "content": "[53] Chen Wei, Wenjing Wang, Wenhan Yang, and Jiaying Liu. Deep retina decomposition for low-light enhancement. arXiv preprint arXiv:1808.04560, 2018. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 208, + 555, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 208, + 555, + 239 + ], + "spans": [ + { + "bbox": [ + 316, + 208, + 555, + 239 + ], + "type": "text", + "content": "[54] Gang Wu, Junjun Jiang, Kui Jiang, and Xianming Liu. Harmony in diversity: Improving all-in-one image restoration via multi-task collaboration. In ACMMM, 2024. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 242, + 555, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 242, + 555, + 283 + ], + "spans": [ + { + "bbox": [ + 316, + 242, + 555, + 283 + ], + "type": "text", + "content": "[55] Haiyan Wu, Yanyun Qu, Shaohui Lin, Jian Zhou, Ruizhi Qiao, Zhizhong Zhang, Yuan Xie, and Lizhuang Ma. Contrastive learning for compact single image dehazing. In CVPR, 2021. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 286, + 555, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 286, + 555, + 319 + ], + "spans": [ + { + "bbox": [ + 316, + 286, + 555, + 319 + ], + "type": "text", + "content": "[56] Fuzhi Yang, Huan Yang, Jianlong Fu, Hongtao Lu, and Baining Guo. Learning texture transformer network for image super-resolution. In CVPR, 2020. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 321, + 555, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 321, + 555, + 353 + ], + "spans": [ + { + "bbox": [ + 316, + 321, + 555, + 353 + ], + "type": "text", + "content": "[57] Hao Yang, Liyuan Pan, Yan Yang, and Wei Liang. Language-driven all-in-one adverse weather removal. In CVPR, 2024. 1, 2, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 355, + 555, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 355, + 555, + 397 + ], + "spans": [ + { + "bbox": [ + 316, + 355, + 555, + 397 + ], + "type": "text", + "content": "[58] Zhiwen Yang, Haowei Chen, Ziniu Qian, Yang Yi, Hui Zhang, Dan Zhao, Bingzheng Wei, and Yan Xu. All-in-one medical image restoration via task-adaptive routing. In MICCAI, 2024. 7, 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 399, + 555, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 399, + 555, + 432 + ], + "spans": [ + { + "bbox": [ + 316, + 399, + 555, + 432 + ], + "type": "text", + "content": "[59] Rajeev Yasarla and Vishal M Patel. Uncertainty guided multiscale residual learning-using a cycle spinning cnn for single image de-raining. In CVPR, 2019. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 434, + 555, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 434, + 555, + 465 + ], + "spans": [ + { + "bbox": [ + 316, + 434, + 555, + 465 + ], + "type": "text", + "content": "[60] Ke Yu, Xintao Wang, Chao Dong, Xiaou Tang, and Chen Change Loy. Path-restore: Learning network path selection for image restoration. TPAMI, 2022. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 468, + 555, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 468, + 555, + 499 + ], + "spans": [ + { + "bbox": [ + 316, + 468, + 555, + 499 + ], + "type": "text", + "content": "[61] Xiaoyan Yu, Shen Zhou, Huafeng Li, and Liehuang Zhu. Multi-expert adaptive selection: Task-balancing for all-in-one image restoration. arXiv preprint arXiv:2407.19139, 2024. 1" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 502, + 555, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 502, + 555, + 544 + ], + "spans": [ + { + "bbox": [ + 316, + 502, + 555, + 544 + ], + "type": "text", + "content": "[62] Eduard Zamfir, Zongwei Wu, Nancy Mehta, Danda Pani Paudel, Yulun Zhang, and Radu Timofte. Efficient degradation-aware any image restoration. arXiv preprint arXiv:2405.15475, 2024. 1, 2, 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 545, + 555, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 545, + 555, + 578 + ], + "spans": [ + { + "bbox": [ + 316, + 545, + 555, + 578 + ], + "type": "text", + "content": "[63] Eduard Zamfir, Zongwei Wu, Nancy Mehta, Yulun Zhang, and Radu Timofte. See more details: Efficient image superresolution by experts mining. In ICML. PMLR, 2024. 3" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 580, + 555, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 580, + 555, + 623 + ], + "spans": [ + { + "bbox": [ + 316, + 580, + 555, + 623 + ], + "type": "text", + "content": "[64] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, Ming-Hsuan Yang, and Ling Shao. Multi-stage progressive image restoration. In CVPR, 2021. 2, 5, 6" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 316, + 624, + 555, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 555, + 667 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 555, + 667 + ], + "type": "text", + "content": "[65] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. In CVPR, 2022. 1, 2, 3, 4, 5" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 316, + 670, + 555, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 555, + 712 + ], + "type": "text", + "content": "[66] Jinghao Zhang, Jie Huang, Mingde Yao, Zizheng Yang, Hu Yu, Man Zhou, and Feng Zhao. Ingredient-oriented multidegradation learning for image restoration. In CVPR, 2023. 1, 2, 5, 6, 7" + } + ] + } + ], + "index": 33 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "12762" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 308 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 297, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 297, + 106 + ], + "type": "text", + "content": "[67] Kai Zhang, Wangmeng Zuo, Shuhang Gu, and Lei Zhang. Learning deep cnn denoiser prior for image restoration. In CVPR, 2017. 1, 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 296, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 296, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 296, + 140 + ], + "type": "text", + "content": "[68] Yulun Zhang, Kunpeng Li, Kai Li, Bineng Zhong, and Yun Fu. Residual non-local attention networks for image restoration. arXiv preprint arXiv:1903.10082, 2019. 1, 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 141, + 296, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 141, + 296, + 184 + ], + "spans": [ + { + "bbox": [ + 56, + 141, + 296, + 184 + ], + "type": "text", + "content": "[69] Haiyu Zhao, Yuanbiao Gou, Boyun Li, Dezhong Peng, Jiancheng Lv, and Xi Peng. Comprehensive and delicate: An efficient transformer for image restoration. In CVPR, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 186, + 296, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 186, + 296, + 228 + ], + "spans": [ + { + "bbox": [ + 56, + 186, + 296, + 228 + ], + "type": "text", + "content": "[70] Yanqi Zhou, Tao Lei, Hanxiao Liu, Nan Du, Yanping Huang, Vincent Zhao, Andrew M Dai, Quoc V Le, James Laudon, et al. Mixture-of-experts with expert choice routing. NeurIPS, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 230, + 296, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 230, + 296, + 274 + ], + "spans": [ + { + "bbox": [ + 56, + 230, + 296, + 274 + ], + "type": "text", + "content": "[71] Yurui Zhu, Tianyu Wang, Xueyang Fu, Xuanyu Yang, Xin Guo, Jifeng Dai, Yu Qiao, and Xiaowei Hu. Learning weather-general and weather-specific features for image restoration under multiple adverse weather conditions. In CVPR, 2023. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 276, + 296, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 276, + 296, + 308 + ], + "spans": [ + { + "bbox": [ + 56, + 276, + 296, + 308 + ], + "type": "text", + "content": "[72] Ozan Özdenizci and Robert Legenstein. Restoring vision in adverse weather conditions with patch-based denoising diffusion models. TPAMI, 2023. 6" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "12763" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Composing Parts for Expressive Object Generation/edecee1f-cf3a-417c-a3d7-2ea937e0cfb0_content_list.json b/2025/Composing Parts for Expressive Object Generation/edecee1f-cf3a-417c-a3d7-2ea937e0cfb0_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..81bf0cac8a599e3f542c1dff4ce96dbf486aa694 --- /dev/null +++ b/2025/Composing Parts for Expressive Object Generation/edecee1f-cf3a-417c-a3d7-2ea937e0cfb0_content_list.json @@ -0,0 +1,1642 @@ +[ + { + "type": "text", + "text": "Composing Parts for Expressive Object Generation", + "text_level": 1, + "bbox": [ + 238, + 130, + 758, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Harsh Rangwani1 Aishwarya Agarwal1 Kuldeep Kulkarni1 \nR. Venkatesh Babu2 Srikrishna Karanam1", + "bbox": [ + 246, + 179, + 746, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Adobe Research $^{2}$ Indian Institute of Science", + "bbox": [ + 305, + 232, + 687, + 250 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/cd5432a7005939731e497150583b3fbe95cd4d73c160d0332a360f8501af1de3.jpg", + "image_caption": [ + "a full body portrait photo of a person wearing a dress with long hair and glasses PartComposer: hair, glasses, dress" + ], + "image_footnote": [], + "bbox": [ + 98, + 280, + 197, + 363 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e13f311949598154339404d5dcacbcf7dc93972ffa2719c075d283bd73a8e611.jpg", + "image_caption": [ + "StableDiffusion" + ], + "image_footnote": [], + "bbox": [ + 204, + 287, + 297, + 363 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7f99108ebc84183ebe8d3deb2488a76ac3a098da2737fe469299f73fb5a2c284.jpg", + "image_caption": [ + "Rich Text Gen" + ], + "image_footnote": [], + "bbox": [ + 299, + 287, + 380, + 363 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/18f0cda5ade8ba47f96b71f4301b491d6bc71c20e15a36443a15116cb2e47b61.jpg", + "image_caption": [ + "PartComposer (Ours)" + ], + "image_footnote": [], + "bbox": [ + 379, + 287, + 495, + 363 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0aa271579bb01c6ababa8f61735064cd8927fc60625d7f9b97b3f346c69f5ac1.jpg", + "image_caption": [ + "Base Generation" + ], + "image_footnote": [], + "bbox": [ + 503, + 289, + 601, + 363 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5ba46fa145d4726e1122e747ab86aa294a2b1863d4c1a61be9ea580832cd46d2.jpg", + "image_caption": [ + "StableDiffusion", + "A photo of a yellow warbler" + ], + "image_footnote": [], + "bbox": [ + 607, + 287, + 797, + 363 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6c79eac0584f1dd146085c96e7f41970869375aad684081edae561f97669c963.jpg", + "image_caption": [ + "PartComposer (Ours)" + ], + "image_footnote": [], + "bbox": [ + 797, + 289, + 900, + 363 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/18ed7dca323cecb7b8369f55cc64530b182b28400d69c3bb952ee3198dbace26.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 96, + 383, + 197, + 465 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4b5f3d7461775cb04de5390b6eca162e4399a251c32c0fc57178d5729b4e0200.jpg", + "image_caption": [ + "Rich Text Gen", + "a photo of a white swan, 8k, full hd \nrtComposer: crown → a crown of a peacock" + ], + "image_footnote": [], + "bbox": [ + 197, + 383, + 297, + 465 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/08bf5885f58861b409b88bafcaf7286cbf6888ddbf86ad06ae8de1d4e33855de.jpg", + "image_caption": [ + "A young woman sits at a table in a beautiful, lush garden, painting by claude monet PartComposer: dress $\\rightarrow$ a dress in Uhuie-style", + "Figure 1. Base Generation (left) Comparison with Methods for Generations with Parts Details (right). PartComposer allows the generation of object images with specified attributes (color, style etc.) of parts for the chosen object in the base text prompt. StableDiffusion and Rich-Text [18] methods with part details either ignore the part instructions or generate inconsistent objects." + ], + "image_footnote": [], + "bbox": [ + 500, + 383, + 599, + 467 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d3cb014fdd70930f1cf40349861e3b4002971da70967ffc7a4a9306771d27268.jpg", + "image_caption": [ + "PartComposer: beak $\\rightarrow$ a hens beak" + ], + "image_footnote": [], + "bbox": [ + 606, + 383, + 900, + 467 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 550, + 326, + 566 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Image composition and generation are processes where the artists need control over various parts of the generated images. However, the current state-of-the-art generation models, like Stable Diffusion, cannot handle fine-grained part-level attributes in the text prompts. Specifically, when additional attribute details are added to the base text prompt, these text-to-image models either generate an image vastly different from the image generated from the base prompt or ignore the attribute details. To mitigate these issues, we introduce PartComposer, a training-free method that enables image generation based on fine-grained part-level attributes specified for objects in the base text prompt. This allows more control for artists and enables novel object compositions by combining distinctive object parts. PartComposer first localizes object parts by denoising the object region from a specific diffusion process. This enables each part token to be localized to the right region. After obtaining part masks, we run a localized diffusion process in each part region based on fine-grained part attributes and combine them to produce the final image. All stages of PartComposer are based on repurposing a pre-trained diffusion model, which", + "bbox": [ + 86, + 583, + 485, + 901 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "enables it to generalize across domains. We demonstrate the effectiveness of part-level control provided by PartComposer through qualitative visual examples and quantitative comparisons with contemporary baselines.", + "bbox": [ + 511, + 551, + 908, + 613 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 645, + 643, + 660 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Image generation with large generative diffusion models like StableDiffusion [40], DALLE [38], etc., has become prevalent due to their superior quality and extensive world knowledge. These models are trained on large image-caption datasets and are trained to generate images based on a given text prompt (description). As image composition and creation are creative processes where the artists need control over various parts of the image being generated. However, adding additional details for controlling part appearance in the text prompt either changes the generated image entirely or ignores the part instructions [9] (Fig. 1).", + "bbox": [ + 509, + 672, + 906, + 838 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Various works aim to provide improved spatial controls to image generation, as they allow image generation conditioned on segmentation masks [5], edge maps [55], bounding boxes [10, 27] etc. Popular methods such as Control-", + "bbox": [ + 511, + 839, + 908, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "13209", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Net [55], GLIGEN [27], etc. require specifying training of these conditional modules, which allows for the controllability of these large generative models. Further, there has been some development of training-free approaches [9, 10] which enable controlled generation of objects by modulating the internal cross-attention activations in the diffusion process. This demonstrates that these pre-trained generative models contain information about the spatial parts of the image, and modulating them effectively can lead to image compositions", + "bbox": [ + 89, + 90, + 480, + 227 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, despite several attempts [6, 55], the semantic controllability of the image generation is still restricted to specifying details at the object level. The object level details can be restrictive, as often creative designers synthesize object parts (e.g. shirt, trouser etc.) and then compose them [23]. Further, variations in semantic parts are often very distinctive and are uniquely used to identify objects [12, 30]. For example, we usually identify species of birds by their unique beak. Due to this, semantic understanding and recognition of parts have been widely studied as a topic in computer vision [12, 13, 21, 31]. Hence, providing controllability for image synthesis at the object semantic part level can enable a large variety of image compositions [23]. Towards this goal, we introduce PartComposer, which provides users with an interface through which they can select the object in the scene and provide a semantic part-level description with fine-grained details for the object generation.", + "bbox": [ + 91, + 229, + 485, + 487 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In PartComposer, we develop a scheme to extract part-level localization masks from the Diffusion model. We introduce a parallel part diffusion process that generates masks for the object parts. The core idea of the approach is that by forcing the part diffusion model to specifically denoise only the object region in the image, it is possible to understand the locations of various parts in the object. After denoising just the object through the part model, the information inside the part diffusion model present in attention maps can be used to generate the masks for various object parts. In the following Part Generation stage, PartComposer utilizes the part masks and attributes of each part the user provides. We enable users to provide a highly expressive specification of parts by using a Rich-Text [18] interface, which allows the specifying attributes, such as style, color, etc., for each part. For the final image generation, taking inspiration from recent studies like Rich-Text Generation [18], Multi-Diffusion [5] etc., we compose the various object parts, by running parallel masked diffusion process for each part while combining them periodically into the image. This combination enables harmonious composition of parts, and also masking ensures only local modifications to the regions corresponding to each part of the object specified. In the PartComposer method, we only use the pre-trained StableDiffusion model, making it a generalizable and training-free approach.", + "bbox": [ + 91, + 491, + 483, + 867 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We extensively test the proposed PartComposer approach for the zero-shot object part segmentation, which is a chal", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "lenging setup in computer vision. We evaluate the part segmentation approach on DeepFashion [29] and CUB200 [50] datasets, where our method significantly outperforms the baseline StableDiffusion (SD). Further, to evaluate the PartComposer image composition abilities, we also provide quantitative results along with a user preference study, where our method significantly outperforms the baselines in generating images consistent with the described parts (Fig. 1). We summarize the core contributions of our paper below:", + "bbox": [ + 511, + 90, + 906, + 226 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. We introduce PartComposer, a training-free method that enables the generation of object images by using provided fine-grained details for the parts of the object. For example, while generating a bird image we can specify a detailed description of its beak.", + "2. In PartComposer, we introduce a novel Part-Diffusion process, which localizes and provides masks for parts of a base object generated by the Diffusion model (Fig. 3). To localize object parts, we introduce a novel segmentation scheme that uses the attention maps of the base diffusion and part diffusion process to obtain accurate masks for localized parts.", + "3. In PartComposer, after localization, we enable the generation of parts from the pre-trained diffusion model based on the Rich-Text description for the parts provided by the user. PartComposer then composes the image to harmoniously blend all object parts with the background by combining localized diffusion paths (Fig. 2)." + ], + "bbox": [ + 514, + 227, + 906, + 500 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 513, + 512, + 661, + 527 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Text-to-Image Generative Models. The text-to-image models synthesize images by following a textual description provided as a prompt. These models have recently become mainstream due to their superior image generation quality and significant knowledge base. This is an outcome of the availability of large-scale image caption datasets [42] and highly parallelized GPU clusters. Almost all kinds of generative models, such as GANs [22], Autoregressive [41, 53], and Diffusion models [38, 40], have shown significant improvements in quality with training on these large image caption datasets. Among these models, the StableDiffusion (SD) [40] models, based on denoising diffusion in latent space are popular due to their open-source nature, which we also utilize for experiments in our work.", + "bbox": [ + 511, + 537, + 908, + 748 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Text-to-Image Models for Downstream Tasks. Generative models, in general, have been useful for various downstream tasks, particularly ones based on per-pixel prediction like Segmentation [1, 11], Depth Estimation [7], etc. As layers near the image generation output have features that capture the pixel-pixel relation [7]. With the large-scale text-to-image generative models, these models often perform very competitively [46, 52] to discriminative methods, on tasks like segmentation. However, one commonality among most of these segmentation methods is that they operate at the", + "bbox": [ + 511, + 750, + 908, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "13210", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/9e8bec10eaa019538a476686de36837aa7edd59fc6536a23d9a6b5097712088d.jpg", + "image_caption": [ + "Figure 2. PartComposer takes the input of a base prompt, the selected object o, and a Rich-Text description ( i.e., footnote, colors, etc.) of the parts. The Part Diffusion process generates the object masks for the specified parts. Then, PartComposer runs a parallel Region Diffusion process to generate attributes of the specified parts, guided by base generation's intermediate outputs." + ], + "image_footnote": [], + "bbox": [ + 133, + 103, + 861, + 321 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "granularity of object or instance level. Text-to-image models here are at an advantage as the usual image captions describe the scene at an object level. In this work, we take a step further in exploring the object part-level knowledge of these text-to-image diffusion models.", + "bbox": [ + 88, + 372, + 480, + 446 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Part Discovery and Segmentation. Part Discovery and Attribution were an integral part of computer vision pipelines classically, as these approaches were robust to viewpoint variations [24, 31, 48]. In deep learning, the unsupervised (self-supervised) approaches for part discovery like SCOPS [21] and Unsup-Parts [12] became popular as they generalize to object parts across categories. In this work, we go one step ahead and operate in a zero-shot unsupervised part segmentation setting, generating part masks from T2I models.", + "bbox": [ + 88, + 449, + 482, + 584 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Controllable Image Generation. Masks, bounding boxes, edge maps, depth maps, etc., have been explored to control the generations of text-to-image diffusion models [6, 14, 27, 35, 36, 55] in addition to text. Further, various other approaches [2-4, 9, 34, 45, 49, 51] achieve control of image semantics through modulating diffusion models. Despite this, the text-to-image generation control at the object part level is under-explored; a recent work [33] tries to do it in a controlled supervised setup using part-masks. Contrary to that, in our work PartComposer we explore a generalized training-free zero-shot setting.", + "bbox": [ + 88, + 585, + 482, + 751 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 768, + 181, + 782 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we introduce PartComposer, our method to synthesize objects based on the description of parts of the objects. In PartComposer, we ask the user to specify a base prompt and the token for the object for which it wants to synthesize parts. Then, we provide the user a rich-text [18] editor (Fig. 2) to specify the parts and their description. PartComposer involves two diffusion steps, a) Part Localiza", + "bbox": [ + 88, + 794, + 482, + 898 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "tion: In the early diffusion stage, we get a mask for the object we want to divide into parts. Then, we perform denoising in later stage from a U-Net condition on parts to fill the masked region of the object, during which it learns to denoise different object parts. Due to this, the attention maps for various parts highlight the correct part region, which we use to extract the part mask. The infilling process is the major contribution of the PartComposer method (Fig. 3). b)", + "bbox": [ + 511, + 372, + 906, + 493 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Part Generation: For generating parts, we combine region-specific diffusion processes for various parts by iteratively merging them inspired by MultiDiffusion [5], Rich-Text Generation [18] etc. However, till now, most of these works have combined the diffusion process for generating objects; for the first time, we have demonstrated its effectiveness in generating object parts. We provide an overview of the PartComposer pipeline in Fig. 2.", + "bbox": [ + 511, + 494, + 908, + 614 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Problem Setup", + "text_level": 1, + "bbox": [ + 511, + 622, + 663, + 638 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We consider each span of tokens $\\mathbf{p}_i$ as indicative of describing one part of the object, with attributes $\\mathbf{a}_i$ describing its overall appearance. Our design choices are based on the rich-text generation [18], and we allow the user to specify the following type of attributes for part generation (Fig. 2): Part Description (i.e. footnote). It is an important attribute that specifies the part of an object; for example, for the part $(\\mathbf{p}_i)$ 'crown' of a bird, we can specify the attribute $(\\mathbf{a}_i^p = \\text{'a crown of a peacock')$ . This helps specify novel part descriptions, which can lead to artistic novel compositions.", + "bbox": [ + 511, + 643, + 908, + 794 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Font Color. It helps us specify the exact RGB values for the color attribute $\\mathbf{a}_i^c$ we want for the object part $\\mathbf{p}_i$ . The exact value of RGB allows fine-grained control over the color of the desired part, whereby just specifying specific colors like 'brick red' leads to ignorance by Stable Diffusion [18].", + "bbox": [ + 511, + 795, + 903, + 869 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Font Style. The font style is indicative of the specific artistic style $\\mathbf{a}_i^s$ like 'of Claude Monet' and 'of Van Gogh' when", + "bbox": [ + 511, + 869, + 903, + 898 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "13211", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/1c5ae7a02e4d99bd276ccf4fef6f5746b69ba170adfdbf53fd9061871a3d1c37.jpg", + "image_caption": [ + "Figure 3. Localization. We obtain the mask $(\\mathcal{M}_o)$ for object in middle of diffusion process. We then denoise object in masked region, using parts $\\mathbf{p}_i$ conditioned U-Net. Due to part denoising, the attention maps of $\\mathbf{p}_i$ can now localize parts as masks $\\mathbf{M}_{\\mathbf{p}_i}$ ." + ], + "image_footnote": [], + "bbox": [ + 93, + 88, + 903, + 315 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "synthesizing images of paintings. This instructs the model to generate a part following a specific style for the given prompt and blend it with other parts of the object (Sec. 5).", + "bbox": [ + 89, + 364, + 482, + 411 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Font Size. The font size controls the relative size of each part in a generation [19]. We use the $\\mathbf{a}_i^w$ to denote size.", + "bbox": [ + 89, + 411, + 482, + 441 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Part Localization", + "text_level": 1, + "bbox": [ + 89, + 450, + 261, + 464 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For part localization, we first run the base stable diffusion model for the given text prompt and extract the token map $\\mathcal{M}_o$ for the object $\\mathbf{o}$ specified by the user, by following the technique of clustering the self-attention maps [18, 37]. We obtain the mask using this clustering after denoising between the initial step $(T)$ until a threshold time step $T_{th}$ . After obtaining the binary object mask $\\mathcal{M}_o$ using attention masks [18, 37], we run two parallel diffusion processes where one contains input from the base prompt and the other contains input from the part prompts $\\hat{\\mathbf{p}} = [\\mathbf{p}_1,\\mathbf{p}_2,\\dots,\\mathbf{p}_n]$ , a token $\\mathbf{p}_i$ for each of the parts specified in the part text prompt. We denote the $\\hat{\\mathbf{b}}$ to denote the base text prompt and $\\hat{\\mathbf{p}}$ to denote the part prompt. For the time $t\\leq T_{th}$ (by default, we use $T_{th}\\approx T / 2$ ), we denoise the U-Net in the object region by denoising it with both the combination of the Part Prompt output and Base Prompt output. Due to this, the Part-Based U-Net gets the information regarding various parts in the object (Fig. 3). We now mathematically define the output noise $\\epsilon_t$ for the diffusion process with Part-Diffusion below:", + "bbox": [ + 88, + 470, + 485, + 758 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\epsilon_ {t} = \\alpha \\mathcal {M} _ {o} \\odot D (x _ {t}, \\hat {\\mathbf {p}}, t) + (1 - \\alpha \\mathcal {M} _ {o}) \\odot D (x _ {t}, \\hat {\\mathbf {b}}, t) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 98, + 782, + 483, + 800 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "here, the $\\alpha$ is the hyper-parameter controlling the strength of the part prompt diffusion output, and $D$ is the output of the pre-trained U-Net of a text-to-image diffusion model. The above denoising process is followed for t steps until the last step to produce a base image corresponding to the given prompt $\\hat{b}$ . Keeping a high $T_{th}$ and low $\\alpha$ , makes minimal", + "bbox": [ + 89, + 810, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "changes in output as if the denoising diffusion was done with original prompt $\\hat{b}$ (See Suppl. Sec.). With this part denoising of object, we obtain attention masks, from which we extract the localization information of part $\\mathbf{p}_t$ . We now describe the process of obtaining the part masks.", + "bbox": [ + 511, + 364, + 906, + 441 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Token Maps for Parts. We first take the part tokens $\\hat{\\mathbf{p}}$ , which are a concatenation of the part names (i.e. 'beak crown wings'), which may not make a meaningful text prompt. Hence, we initialize text embeddings for all these tokens $\\mathbf{p}_i$ by passing the meaningful text prompt having the following template: \"A photo of $\\mathbf{p}_i$ of a o\" where $\\mathbf{p}_i$ is the object part name and o is the object name. This serves two purposes: first, it makes the text embedding meaningful, and second, it introduces some invariance from the order of part specification in part prompt. These embeddings are then passed as text embeddings to the Part U-Net for denoising. After running the denoising process, we aggregate the self-attention maps across multi-heads and time steps (from 32 × 32 resolution) for both the base and part U-Net diffusion branches, taking inspiration from works that demonstrate that attention can localize objects [9, 18, 30, 44]. We then perform spectral clustering on these attention maps to form k segmentation maps $\\hat{\\mathbf{M}}$ ( $32 \\times 32$ ), based on pixel-pixel similarity. To attribute these K-segments to the part specified by the user, we aggregate the cross attention of Part U-Net diffusion process. For each token $\\mathbf{p}_j$ , we obtain the cross-attention score as follows: $\\hat{\\mathbf{m}}_j = \\frac{c_j}{\\sum_k c_k}$ , where $c_j$ is the cross attention score for each token. We proceed by aggregating the attention heads to obtain the average cross-attention scores and resizing them to $32 \\times 32$ , obtaining $\\hat{\\mathbf{m}}$ . We remove the start of text ( $\\langle \\mathrm{sot} \\rangle$ ) token for cross-attention and re-normalize it [9]. In other works [18, 37], as the tokens correspond to the objects being generated in the image, it's sure that token maps will be meaningful. However, this is not true for parts, as Part U-Net might not localize some", + "bbox": [ + 511, + 446, + 908, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "13212", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "parts. For determining if the part is localized, we propose to look at the max value of cross attention map spatially across the pixels; if we find that the following condition is met, the part is localized:", + "bbox": [ + 89, + 90, + 483, + 151 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL (j) = \\mathbb {1} \\left\\{\\max \\left(\\hat {\\mathbf {m}} _ {j}\\right) \\geq (1 - \\delta) \\frac {1}{K} \\right\\}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 163, + 157, + 483, + 188 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here $\\delta$ is a hyperparameter, and $K$ is the number of parts. Is This condition is robust in finding the localized parts (See Sec. 5 for ablation). For the parts $\\mathbf{p}_i$ , which are localized, we normalize the cross-attention map:", + "bbox": [ + 89, + 194, + 483, + 255 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {m}} _ {j} = L (j) \\frac {\\hat {\\mathbf {m}} _ {j} - \\min (\\hat {\\mathbf {m}} _ {j})}{\\max (\\hat {\\mathbf {m}} _ {j}) - \\min (\\hat {\\mathbf {m}} _ {j})} + (1 - L (j)) \\hat {\\mathbf {m}} _ {j}.\n$$\n", + "text_format": "latex", + "bbox": [ + 107, + 261, + 465, + 296 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We follow a dot-product-based protocol to assign each K cluster in the self-attention masks to a part, unlike the average attention protocol in the previous works [18, 37]. We find that dot product of normalized cross attention scores $\\hat{\\mathbf{m}}_j$ for each token with self-attention masks $\\hat{\\mathbf{M}}_j$ works better, as the attention maps in Part U-Net are noisy. Still, they are often correct for the regions that are localized in one specific area only (Suppl. Fig. 14). Hence, the dot product protocol favors those maps that are only localized in some areas of the image and don't have high attention values across all parts of the object. After obtaining the dot product scores for each part, we assign $\\hat{\\mathbf{M}}_j$ to $\\mathbf{p}_i$ with the highest scores. The mask for the part $\\mathbf{p}_i$ is finally given as the following:", + "bbox": [ + 89, + 301, + 483, + 500 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {M} _ {\\mathbf {p} _ {i}} = \\left\\{\\cup_ {j} \\hat {\\mathbf {M}} _ {j} \\mid \\operatorname {a r g m a x} _ {i} \\hat {\\mathbf {M}} _ {j} \\cdot \\hat {\\mathbf {m}} _ {i} = i \\text {a n d} \\hat {\\mathbf {M}} _ {j} \\cdot \\hat {\\mathbf {m}} _ {i} \\geq \\epsilon , \\right\\} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 521, + 483, + 553 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here $\\epsilon$ is the hyperparameter which controls the minimum similarity required between the attention mask and the part token. We combine the additional attention masks unassigned to any token, and name them as the background (other) token $\\mathbf{M}_b$ .", + "bbox": [ + 89, + 554, + 483, + 628 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Part Generation", + "text_level": 1, + "bbox": [ + 89, + 637, + 253, + 651 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We follow a similar protocol for the generation of part segments as in Rich-Text Generation. We tailor the rich text generation to compose the part regions of the object in the image in place of the original scene composition. We describe the part generation protocol below briefly and refer readers to Rich-Text Gen [18] for more details. For each part $\\mathbf{p}_i$ we run a region diffusion process, which runs in parallel for all the parts. We then combine the region diffusion processes to obtain the final noise prediction $\\epsilon_t$ as the masked $\\mathbf{M}_{\\mathbf{p}_i}$ sum of the denoiser outputs:", + "bbox": [ + 89, + 659, + 485, + 811 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\epsilon_ {t} = \\sum_ {i} \\mathbf {M} _ {\\mathbf {p} _ {i}} \\epsilon_ {t, \\mathbf {p} _ {i}} = \\sum_ {i} \\mathbf {M} _ {\\mathbf {p} _ {i}} \\odot D (x _ {t}, f (\\mathbf {p} _ {i}, \\mathbf {a} _ {i}), t) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 102, + 830, + 483, + 862 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $D$ is the pre-trained U-Net model, and $f(\\mathbf{p}_i,\\mathbf{a}_i)$ is the text description of the part $\\mathbf{p}_i$ constructed using the following", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/00ab72b86e894bcbc9ceccaba55020732970aac576a0369b3d46c107e7da795c.jpg", + "image_caption": [ + "Figure 4. Segmentation masks for the parts that are localized by Part Diffusion for DeepFashion (above) and CUB200 (below)." + ], + "image_footnote": [], + "bbox": [ + 519, + 89, + 898, + 176 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "process using the part tokens $\\mathbf{p}_i$ and attributes $\\mathbf{a}_i$ . Initially the text $f(\\mathbf{p}_i, \\mathbf{a}_i) = \\mathbf{p}_i$ , is set to part token itself. In case the part description (i.e. footnote) is available we set the $f(\\mathbf{p}_i, \\mathbf{a}_i) = \\mathbf{a}_i^p$ , further if the style attribute is available we do $f(\\mathbf{p}_i, \\mathbf{a}_i) = f(\\mathbf{p}_i, \\mathbf{a}_i) + \\text{in style of} + \\mathbf{a}_i^s$ . In case the color attribute is also specified, the nearest named color $\\hat{\\mathbf{a}}_i^c$ (e.g. red) for the specific RBG color $\\mathbf{a}_i^c$ is found, and $f(\\mathbf{p}_i, \\mathbf{a}_i) = \\hat{\\mathbf{a}}_i^c + ' ' + f(\\mathbf{p}_i, \\mathbf{a}_i)$ . The string $f(\\mathbf{p}_i, \\mathbf{a}_i)$ is the text input for the Diffusion to generate the part $\\mathbf{p}_i$ . We use the base prompt $\\hat{\\mathbf{b}}$ as $f(\\mathbf{p}_i, \\mathbf{a}_i)$ for the background masked region $\\mathbf{M}_b$ . Combining different diffusion outputs at every time $t$ helps generate a harmonious image after blending the defined parts of the object.", + "bbox": [ + 511, + 222, + 906, + 417 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Following Rich-Text [18], we also utilize the gradient guidance [16, 20] by taking the gradient of MSE loss between the estimated original image and the color value $\\mathbf{a}_i^c$ specified by the user. The gradient guidance helps generate the exact RGB color for part $\\mathbf{p}_i$ , which is impossible with just text guidance [18]. Further, as we use the base text framework of Rich-Text, we also can specify font size attribute to $\\mathbf{a}_i^f$ to control the relative importance of each part in the object image.", + "bbox": [ + 511, + 419, + 908, + 553 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Preservation of Other Parts. As we only intend to modify the object o from the original prompt, we also use the Self-Injection techniques from Plug and Play [47] to maintain the overall structure of the background from base prompt generation. Further, to ensure that our diffusion trajectory follows the same path as the base, the background region is also blended with base noise generation outputs.", + "bbox": [ + 511, + 554, + 908, + 660 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nx _ {t} = \\mathbf {M} _ {b} \\odot x _ {t} ^ {\\text {b a s e}} + (1 - \\mathbf {M} _ {b}) \\odot x _ {t} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 667, + 906, + 685 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In our case, the attention maps out of the object $\\mathbf{Mask}\\mathcal{M}_o$ and all the object parts not assigned to any part token comprise the background $\\mathbf{M}_b$ , and we start this blending process at $t = T_{blend}$ . We find this to be very useful in preserving the structure of the other parts of the image and just generating the described parts in the localized region. We provide an overview of the complete process in Fig. 2.", + "bbox": [ + 511, + 691, + 908, + 797 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experimental Analysis", + "text_level": 1, + "bbox": [ + 511, + 808, + 728, + 825 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Evaluation of Part Localization", + "text_level": 1, + "bbox": [ + 511, + 833, + 790, + 848 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We first evaluate the part localization module, which is based on the novel idea of denoising only the object region with the part-based diffusion outputs.", + "bbox": [ + 511, + 854, + 906, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "13213", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/0b0273349e42dfc9939d9ab4c28a7100bf544b2d6885296e80d72a613cb96e96.jpg", + "table_caption": [ + "Table 1. Comparison to Prior Works for the unsupervised part segmentation task. We follow Unsup-Part [12] for evaluation protocols and baseline results for $(\\mathbf{K} = 4)$ parts. We report clustering-based NMI and ARI metrics, which are higher for better segmentation outputs." + ], + "table_footnote": [], + "table_body": "
DeepFashion DatasetCUB200 Dataset
MethodFG-NMIFG-ARINMIARIFG-NMIFG-ARINMIARI
Unsupervised Learning
SCOPS [21]30.727.656.681.439.117.924.47.1
Unsup-Parts [12]44.846.668.190.646.021.043.519.6
DFF [11]----32.414.325.912.4
Unsupervised Zero-Shot
Rich-Text [18]16.05.248.358.73.10.33.10.3
StableDiffusion12.03.540.670.98.00.63.30.6
PartComposer(Ours)24.718.048.073.420.59.218.57.7
", + "bbox": [ + 166, + 121, + 831, + 316 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details. We use the StableDiffusion (SD) version 2.1 for our experimentation purposes. We use the DDIM Scheduler [43] with 50 steps to generate results for SD2.1 to evaluate the part localization process. As ground truth is not available for the part masks of the generated images, we use test sets of the commonly used DeepFashion [29] and CUB-200 [50] datasets for evaluation. These datasets are the standard datasets used for the evaluation of the unsupervised part segmentation approaches. The DeepFashion dataset contains images along with their part segmentation masks, divided into 14 categories of labels. The CUB200 dataset contains the key point annotations for the 14 key points specified for the bird categories. We provide further details in Suppl. Section.", + "bbox": [ + 88, + 321, + 486, + 532 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines and Problem Setting. As we operate in the setting of Zero-shot (i.e. no training) Unsupervised Part Segmentation, there are no previous works that report results in such a challenging setting. Hence, we provide results for the unsupervised learning approaches to facilitate comparison. We provide results for the SCOPS [21], which utilizes the internal features of the VGG model to train a model based on self-supervised loss functions to predict parts robustly across categories. The other stronger self-supervised baseline is Unsup-Parts [12], which uses contrastive loss functions to train a network based on equivariance and other vision properties to cluster the object regions into semantic parts. In addition to this, we also report results for the DFF [11] as they also operate in the same setting. We want to highlight that these unsupervised approaches require either training a neural network or performing clustering on the complete training data to segment parts. In contrast, our approach is training free and operates in a zero-shot fashion. Hence, the performance of zero-shot approaches is not directly comparable to unsupervised methods.", + "bbox": [ + 91, + 537, + 486, + 839 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Zero Shot Unsupervised Part Segmentation. As there is no benchmark to evaluate the part localization for generated images, we use the existing dataset of DeepFashion and CUB200 to obtain our results. To obtain segmentation for", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "each image, we first invert the image into the diffusion latent space using Null-Text Inversion [32] method (see Sec. 5 for ablation). We use the BLIP-V2 [25] captioner provided in Diffusers to obtain approximate text prompts for inversion. After providing the desired image and prompt, the Null-Text inversion provides us with inverted latent and unconditioned time embeddings to reconstruct the image of the dataset. We then construct a StableDiffusion baseline in which, in addition to the prompts, we append the list of part tokens $\\mathbf{p}$ to the prompt. We then use the segmentation algorithm described above in Sec. 3.2 to extract token maps corresponding to the part tokens $\\mathbf{p}_i$ . We also evaluate a Rich-Text [18] baseline segmentation algorithm for the same. For the proposed part-denoising approach based on parallel diffusion in PartComposer, we first get the inverted latents and embeddings. Then, we generate a base diffusion process to reconstruct the image and use Part Diffusion to fill the parts of the image. To very fairly compare the StableDiffusion performance with part-denoising, we keep all things the same except the part diffusion process to get masks. Further across these baselines, we use low classifier guidance to ensure proper reconstruction of the dataset images (See Suppl. for details).", + "bbox": [ + 511, + 321, + 908, + 655 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Protocol. We use the standard experimental protocol as used by earlier works [12] to facilitate the right point of comparison. We want to point out that in our work, the part name (e.g., beak, etc.) is associated with the localized mask, but the unsupervised approaches produce $(\\mathrm{K} = 4)$ parts without any part names. Hence, to make a fair comparison, we create 4 clusters of parts based on their locality, and finally generate segmentation masks with a maximum of 4 clusters. We provide the exact mapping of the part names to cluster labels in the Suppl. Section. We compare the specifically designed metrics [12] of NMI (Normalized Mutual Information) and ARI (Adjusted Rand Index) of the predicted cluster labels with part masks in the case of DeepFashion and key points in case of the CUB dataset. We report the metrics along with their foreground variants (FG-NMI and FG-ARI) for a holistic comparison of performance.", + "bbox": [ + 511, + 657, + 910, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "13214", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/8ce660be2b47af28118151ccd3f21a9815fa29a6634aa78c3e5c9f3e78410dfa.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 114, + 89, + 267, + 186 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/190c79cd63844621e26fc29025dab3fbdd375c615e10579360a13c7945077746.jpg", + "image_caption": [ + "Figure 5. Qualitative Comparison for PartComposer. (above) We provide base prompt 'a photo of a flamingo' and part prompt as 'beak - a pelicans beak'. (below) We generate a photo of a man, with part prompt as 'black jacket and blue jeans'. We find that other baselines either ignore instruction or change the entire composition. On the contrary, PartComposer can correctly localize parts and generate details." + ], + "image_footnote": [], + "bbox": [ + 114, + 188, + 267, + 284 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b76912a9f8ff072d94f0c8e42f6de4f760ff818a4488135154424211f29d200a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 277, + 90, + 428, + 181 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "", + "bbox": [ + 277, + 188, + 429, + 277 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "PartCraft (Ours)", + "bbox": [ + 313, + 277, + 395, + 284 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b3d287cc8aeb7867fe07c9a4427a8e66c9ea52d84d9ddd1742bba5178393a21e.jpg", + "image_caption": [ + "StableDiffusion" + ], + "image_footnote": [], + "bbox": [ + 428, + 90, + 578, + 180 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a4a04e6b8fe4eb13aed038c7e1f22c36d4c5ef6eb396580b499242f2800abe0d.jpg", + "image_caption": [ + "InstructPix2Pix" + ], + "image_footnote": [], + "bbox": [ + 578, + 90, + 730, + 180 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0af86d5f4d4db4124a76497fa148205cae060aebd7e979085eb352b94bd8d3de.jpg", + "image_caption": [ + "Rich Text Gen" + ], + "image_footnote": [], + "bbox": [ + 728, + 90, + 879, + 180 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "", + "bbox": [ + 277, + 188, + 429, + 277 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "PartCraft (Ours)", + "bbox": [ + 313, + 277, + 395, + 284 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b1e8a37755145c7331dfacbe8bcf9ffa4c66b0305a1b7e56c5fa86b8ff174ab3.jpg", + "image_caption": [ + "StableDiffusion" + ], + "image_footnote": [], + "bbox": [ + 429, + 188, + 578, + 277 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e78ad06de85adcbee3536b5f973a50d316cf659e4f6044ea31a982e8d677b149.jpg", + "image_caption": [ + "InstructPix2Pix" + ], + "image_footnote": [], + "bbox": [ + 578, + 188, + 728, + 277 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b61359ccfa2c89ebe844abc616efb258d0b87056856bd038e04cfd9a05089aa0.jpg", + "image_caption": [ + "Rich Text Gen" + ], + "image_footnote": [], + "bbox": [ + 728, + 188, + 879, + 277 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results. We tabulate the results for all the baselines in Table 1. Our approach, PartComposer, significantly improves over the baselines by $\\geq 5$ points in NMI and ARI across most metrics in the Unsupervised Zero-Shot setting. Further, in some cases our methods, NMI and ARI, are near the unsupervised learning approaches. This demonstrates that non-trivial part masks can be generated by harnessing the power of the pre-trained diffusion models, and they can serve as a good initial prior for learning part segmentation approaches by using them as base pre-trained models. We provide qualitative results for PartComposer, where we observe that PartComposer can associate the right part with the correct label (Fig 4). On the contrary, in the StableDiffusion baseline, the parts often get assigned to the wrong part tokens, which is the cause of degraded performance (Suppl. Fig. 16). We also observe that in the case of our algorithm, if the part gets localized based on the max condition defined in the segmentation algorithm, the part mask found is usually in the right region (See Fig. 4 and Table 2). This demonstrates the effectiveness of our part segmentation procedure, which often sides with not localizing objects rather than performing arbitrary assignments.", + "bbox": [ + 91, + 342, + 483, + 674 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. PartComposer Image Generation Evaluation", + "text_level": 1, + "bbox": [ + 89, + 681, + 470, + 698 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We now evaluate the object composition ability of the proposed approach, PartComposer, compared to the baselines. In this section, we use the base generation model as a StableDiffusion (SD) 1.5 model to do a fair comparison with the Rich-Text baseline. We used the same generation setting of a PNDM [28] scheduler with 41 steps and suggested classifier guidance 8.5. To perform a fair comparison, we keep all the parameters same as that of the Rich-Text [18] baseline.", + "bbox": [ + 89, + 704, + 483, + 824 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Baselines. For the task of generating the object image based on part-level details of the specified object, we use the strong baselines of Rich-Text Generation and InstructPix2Pix [8]. The Rich-Text Generation method generates the base image itself, whereas the InstructPix2Pix method requires us to", + "bbox": [ + 89, + 825, + 483, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "generate the base image. We also evaluate the standard StableDiffusion baseline in which we add all the part details in the text prompt to generate the desired image. We compare all these baselines with our proposed method, PartComposer, while ensuring that the base parameters and seed are the same for the base image generation. A recent work [33] regarding part generation operates in a fully supervised setup using part masks and cannot be used to create novel parts based on text instruction, making it incomparable. We defer specific details for the baselines to the Suppl. Section.", + "bbox": [ + 511, + 342, + 906, + 493 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Visual Comparison. We provide a visual comparison for the baselines for the a) bird image distinctive part generation and b) human image part generations. Across both cases (Fig. 5), our proposed method, PartComposer, only modifies the desired beak region and replaces it with the iconic pelican beak. The StableDiffusion baseline modifies the image completely while ignoring instructions (Supp. Fig. 12). In the Rich-Text generation, the full region corresponding to the bird gets modified instead of only the specified prompt in base generation. For the editing-based InstructPix2Pix method, we observe that it modifies the base image at a global level and cannot localize modification to the desired part region (Fig. 5). We further compare and find that recent models like SD3.5, Inpainting [17] and recent SotA editing methods [15, 26, 47, 54] can still not follow exact part instructions as seen in Supp. Fig. 8, 9 and 10. In contrast, observe that PartComposer-generated images are novel aesthetic compositions that follow the part-level instructions.", + "bbox": [ + 511, + 500, + 908, + 772 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Quantitative Evaluation. As the part-specific generation is fine-grained, we perform the both automatic and user-study evaluations. We performed a user study by inviting 50 participants (28 responded with a full survey). We evaluate the model by asking questions and evaluating metrics on three orthogonal aspects i.e., a) Localization of part generation in comparison to the base image (through LPIPS [56]), b) Text-Consistency of generated parts by measuring CLIP", + "bbox": [ + 511, + 780, + 908, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "13215", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/94368306424c8d25c29f0a7b17b893f031f03430a82874d82f08b48be672d12d.jpg", + "table_caption": [ + "Table 2. Ablation Analysis of PartComposer for Part Localization, showing performance improvement with each component." + ], + "table_footnote": [], + "table_body": "
PartComposer (Ablations)
MethodFG-NMIFG-ARI
PartComposer35.411.0
w/o Null-Text Inversion23.15.2
w/o Max Localization21.32.8
w/o Dot Product Localization23.75.0
w/o Independent Text31.28.9
PartComposer (Clustering)
PartComposer (K = 9)35.411.0
PartComposer (K = 4)20.42.8
PartComposer (K = 14)35.210.3
", + "bbox": [ + 91, + 114, + 460, + 316 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c7feb96f76df24e61954e428d9fcccf9414cd4b7cc0d2dd95e06b98b5ef2a6dc.jpg", + "image_caption": [ + "Figure 6. PartComposer generalizes over domains as shown by Claude Monet's painting style (above), where we specify the dress part of women to follow styles like Van Gogh and Ukiyo-e. We perform similar modifications to Woodie's image." + ], + "image_footnote": [], + "bbox": [ + 94, + 329, + 184, + 492 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9ba083cb1b50e0235a03d773b19f8abc2c86af8756b184a6179b5e24e63ac170.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 196, + 329, + 287, + 492 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/98a4c4e120dbae11deb0e00f5cbab1350b8b9e297287218afe3a75b67614aaf1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 292, + 329, + 383, + 491 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2a89e87e5a288a024d70157572874c1490e479d558bc4da45bdc22ac58433b55.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 390, + 329, + 480, + 491 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "similarity of the part detailed text prompt with the generated image and c) Aesthetic Quality of the generated image through LAION-5B [42] CLIP Aesthetic scorer. In total, we collect about 3.5k opinions. In Fig. 7, we summarize the results in which we observe that PartComposer is significantly preferred over the other baselines in terms of localized and consistent Part Generation. The automated evaluation results are provided in the table below, which also follow a similar trend as in the user study, demonstrating the effectiveness of PartComposer in a localized part generation while ensuring the aesthetics score is similar to that of the base model.", + "bbox": [ + 89, + 571, + 483, + 736 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Analysis and Discussion", + "text_level": 1, + "bbox": [ + 89, + 753, + 316, + 768 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Part Diffusion Based Localization. We ablate the components we have introduced in the PartDiffusion process of PartComposer (Sec. 3.2). We provide an analysis of the effect of using a) null-text inversion, b) max-based localization c) usage of dot product-based protocol in the part assignment, and d) independent text embeddings. We have used a subset of CUB-200 images to perform all evaluations, which is kept fixed across ablations. We tabulate the ablations in", + "bbox": [ + 89, + 779, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/850689bb6bce080f702d0358c762bd3ce0aca0b801107d68eb43f6972bb75596.jpg", + "image_caption": [ + "Figure 7. User Study and Quantitative Results for part-based image generation baselines." + ], + "image_footnote": [], + "bbox": [ + 521, + 119, + 908, + 339 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/a04a5a598777fc64d126427c0550dfe94bf47b6bca83ee1c177fed057a0f6c4e.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
PartComposerStableDiffusionInstructPix2PixRich-Text
LPIPS0.1680.4670.1890.243
CLIP0.2010.1830.1930.187
Aesthetic5.665.685.635.65
", + "bbox": [ + 514, + 345, + 903, + 393 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 2. We observe that all the components introduced in PartComposer contribute significantly to the performance.", + "bbox": [ + 511, + 409, + 906, + 439 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Generalization of PartComposer. In Fig. 6, above a painting in Claude Monet style generated by the base StableDiffusion method. We then use PartComposer to specify the dress part of the women's token, to dark green color in Monet Style, orange color dress in Van Gogh Style (middle), and green color dress in Ukiyo-e style. In Fig. 6, we provide results for the 'Woody' character from Toy Story. These results show that PartComposer can generate shirt, pants, and dress color variations, leading to aesthetic image combinations in synthetic and natural (Supp. Fig. 13) domains. This zero-shot generalization across domains demonstrates the creative activities that can be enabled with PartComposer.", + "bbox": [ + 511, + 444, + 908, + 627 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 657, + 633, + 672 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work we introduce PartComposer, a method to generate object images with fine-grained attribute details specified at the part level, using an expressive Rich-Text interface. The PartComposer method introduces a novel part diffusion process, which is responsible for denoising objects using the part features, and then utilizes a region-specific diffusion process to generate part details and compose the final image. PartComposer serves as initial work enabling rich-text-based training-free part-level control for SD models.", + "bbox": [ + 511, + 689, + 908, + 824 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations. We find that part generation is bottlenecked with the part understanding. If the part is localized correctly, the text-to-image model can generate specified details in PartComposer. Hence, improving the part-level localization of these models is a good direction for future works.", + "bbox": [ + 511, + 825, + 908, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "13216", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Rameen Abdal, Peihao Zhu, Niloy J Mitra, and Peter Wonka. Labels4free: Unsupervised segmentation using stylegan. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13970-13979, 2021. 2", + "[2] Aishwarya Agarwal, Srikrishna Karanam, K J Joseph, Apoorv Saxena, Koustava Goswami, and Balaji Vasan Srinivasan. Astar: Test-time attention segregation and retention for text-to-image synthesis, 2023. 3", + "[3] Aishwarya Agarwal, Srikrishna Karanam, and Balaji Vasan Srinivasan. Training-free color-style disentanglement for constrained text-to-image synthesis. arXiv preprint arXiv:2409.02429, 2024.", + "[4] Yuval Alaluf, Daniel Garibi, Or Patashnik, Hadar Averbuch-Elor, and Daniel Cohen-Or. Cross-image attention for zero-shot appearance transfer. arXiv preprint arXiv:2311.03335, 2023. 3", + "[5] Omer Bar-Tal, Lior Yariv, Yaron Lipman, and Tali Dekel. Multidiffusion: Fusing diffusion paths for controlled image generation. 2023. 1, 2, 3", + "[6] Shariq Farooq Bhat, Niloy J. Mitra, and Peter Wonka. Loosecontrol: Lifting controlnet for generalized depth conditioning, 2023. 2, 3", + "[7] Anand Bhattachad, Daniel McKee, Derek Hoiem, and David Forsyth. Stylegan knows normal, depth, albedo, and more. Advances in Neural Information Processing Systems, 36, 2024. 2", + "[8] Tim Brooks, Aleksander Holynski, and Alexei A. Efros. Instructpix2pix: Learning to follow image editing instructions. In CVPR, 2023. 7, 6", + "[9] Hila Chefer, Yuval Alaluf, Yael Vinker, Lior Wolf, and Daniel Cohen-Or. Attend-and-excite: Attention-based semantic guidance for text-to-image diffusion models. ACM Transactions on Graphics (TOG), 42(4):1-10, 2023. 1, 2, 3, 4", + "[10] Minghao Chen, Iro Laina, and Andrea Vedaldi. Training-free layout control with cross-attention guidance. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 5343–5353, 2024. 1, 2", + "[11] Jaehoon Choi, Taekyung Kim, and Changick Kim. Self-ensembling with gan-based data augmentation for domain adaptation in semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6830–6840, 2019. 2, 6", + "[12] Subhabrata Choudhury, Iro Laina, Christian Rupprecht, and Andrea Vedaldi. Unsupervised part discovery from contrastive reconstruction. Advances in Neural Information Processing Systems, 34:28104-28118, 2021. 2, 3, 6", + "[13] Edo Collins, Radhakrishna Achanta, and Sabine Susstrunk. Deep feature factorization for concept discovery. In Proceedings of the European Conference on Computer Vision (ECCV), pages 336-352, 2018. 2", + "[14] Omer Dahary, Or Patashnik, Kfir Aberman, and Daniel Cohen-Or. Be yourself: Bounded attention for multi-subject text-to-image generation. arXiv preprint arXiv:2403.16990, 2(5), 2024. 3", + "[15] Gilad Deutch, Rinon Gal, Daniel Garibi, Or Patashnik, and" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Daniel Cohen-Or. Turboedit: Text-based image editing using few-step diffusion models, 2024. 7, 1", + "[16] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 5", + "[17] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In *Forty-first International Conference on Machine Learning*, 2024. 7, 1", + "[18] Songwei Ge, Taesung Park, Jun-Yan Zhu, and Jia-Bin Huang. Expressive text-to-image generation with rich text. In IEEE International Conference on Computer Vision (ICCV), 2023. 1, 2, 3, 4, 5, 6, 7, 8", + "[19] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 4", + "[20] Jonathan Ho, William Chan, Chitwan Sahara, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 5", + "[21] Wei-Chih Hung, Varun Jampani, Sifei Liu, Pavlo Molchanov, Ming-Hsuan Yang, and Jan Kautz. Scops: Self-supervised co-part segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 869-878, 2019. 2, 3, 6", + "[22] Minguk Kang, Jun-Yan Zhu, Richard Zhang, Jaesik Park, Eli Shechtman, Sylvain Paris, and Taesung Park. Scaling up gans for text-to-image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10124-10134, 2023. 2", + "[23] Hyung-Kwon Ko, Gwanmo Park, Hyeon Jeon, Jaemin Jo, Juho Kim, and Jinwook Seo. Large-scale text-to-image generation models for visual artists' creative works. In Proceedings of the 28th International Conference on Intelligent User Interfaces, pages 919-933, 2023. 2", + "[24] Svetlana Lazebnik, Cordelia Schmid, and Jean Ponce. Semi-local affine parts for object recognition. In *British Machine Vision Conference*, BMVC 2004, Kingston, UK, September 7-9, 2004. Proceedings, pages 1-10. BMVA Press, 2004. 3", + "[25] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip2: Bootstrapping language-image pre-training with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597, 2023. 6", + "[26] Shanglin Li, Bohan Zeng, Yutang Feng, Sicheng Gao, Xiuhui Liu, Jiaming Liu, Lin Li, Xu Tang, Yao Hu, Jianzhuang Liu, et al. Zone: Zero-shot instruction-guided local editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6254-6263, 2024. 7, 1", + "[27] Yuheng Li, Haotian Liu, Qingyang Wu, Fangzhou Mu, Jianwei Yang, Jianfeng Gao, Chunyuan Li, and Yong Jae Lee. Gligen: Open-set grounded text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22511-22521, 2023. 1, 2, 3", + "[28] Luping Liu, Yi Ren, Zhijie Lin, and Zhou Zhao. Pseudo numerical methods for diffusion models on manifolds. In" + ], + "bbox": [ + 517, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "13217", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "International Conference on Learning Representations, 2022. 7", + "[29] Ziwei Liu, Ping Luo, Shi Qiu, Xiaogang Wang, and Xiaou Tang. Deepfashion: Powering robust clothes recognition and retrieval with rich annotations. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 2, 6", + "[30] Wan-Duo Kurt Ma, JP Lewis, W Bastiaan Kleijn, and Thomas Leung. Directed diffusion: Direct control of object placement through attention guidance. arXiv preprint arXiv:2302.13153, 2023. 2, 4", + "[31] Subhransu Maji, Esa Rahtu, Juho Kannala, Matthew Blaschko, and Andrea Vedaldi. Fine-grained visual classification of aircraft. arXiv preprint arXiv:1306.5151, 2013. 2, 3", + "[32] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6038–6047, 2023. 6, 5", + "[33] Kam Woh Ng, Xiatian Zhu, Yi-Zhe Song, and Tao Xiang. Partcraft: Crafting creative objects by parts. arXiv preprint arXiv:2407.04604, 2024. 3, 7, 6", + "[34] Rishubh Parihar, Abhijnya Bhat, Abhipsa Basu, Saswat Mallick, Jogendra Nath Kundu, and R Venkatesh Babu. Balancing act: distribution-guided debiasing in diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6668-6678, 2024. 3", + "[35] Rishubh Parihar, Harsh Gupta, Sachidanand VS, and R Venkatesh Babu. Text2place: Affordance-aware text guided human placement. In European Conference on Computer Vision, pages 57-77. Springer, 2024. 3", + "[36] Rishubh Parihar, VS Sachidanand, Sabariswaran Mani, Tejan Karmali, and R Venkatesh Babu. Precisecontrol: Enhancing text-to-image diffusion models with fine-grained attribute control. In European Conference on Computer Vision, pages 469-487. Springer, 2024. 3", + "[37] Or Patashnik, Daniel Garibi, Idan Azuri, Hadar Averbuch-Elor, and Daniel Cohen-Or. Localizing object-level shape variations with text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 4, 5", + "[38] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. 1, 2", + "[39] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, et al. Sam 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714, 2024. 1", + "[40] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 2" + ], + "bbox": [ + 91, + 90, + 483, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[41] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural information processing systems, 35:36479-36494, 2022. 2", + "[42] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278-25294, 2022. 2, 8", + "[43] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2105.05233, 2021. 6, 5", + "[44] Raphael Tang, Linqing Liu, Akshit Pandey, Zhiying Jiang, Gefei Yang, Karun Kumar, Pontus Stenetorp, Jimmy Lin, and Ferhan Ture. What the daam: Interpreting stable diffusion using cross attention. arXiv preprint arXiv:2210.04885, 2022. 4", + "[45] Yoad Tewel, Omri Kaduri, Rinon Gal, Yoni Kasten, Lior Wolf, Gal Chechik, and Yuval Atzmon. Training-free consistent text-to-image generation. arXiv preprint arXiv:2402.03286, 2024.3", + "[46] Junjiao Tian, Lavisha Aggarwal, Andrea Colaco, Zsolt Kira, and Mar Gonzalez-Franco. Diffuse, attend, and segment: Unsupervised zero-shot segmentation using stable diffusion. arXiv preprint arXiv:2308.12469, 2023. 2", + "[47] Narek Tumanyan, Michal Geyer, Shai Bagon, and Tali Dekel. Plug-and-play diffusion features for text-driven image-to-image translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1921-1930, 2023. 5, 7, 1", + "[48] Andrea Vedaldi, Siddharth Mahendran, Stavros Tsogkas, Subhransu Maji, Ross Girshick, Juho Kannala, Esa Rahtu, Iasonas Kokkinos, Matthew B. Blaschko, David Weiss, Ben Taskar, Karen Simonyan, Naomi Saphra, and Sammy Mohamed. Understanding objects in detail with fine-grained attributes. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2014. 3", + "[49] Andrey Voynov, Qinghao Chu, Daniel Cohen-Or, and Kfir Aberman. $p+$ : Extended textual conditioning in text-to-image generation. arXiv preprint arXiv:2303.09522, 2023. 3", + "[50] C. Wah, S. Branson, P. Welinder, P. Perona, and S. Belongie. Caltech-ucsd birds-200-2011 dataset. Technical Report CNSTR-2011-001, California Institute of Technology, 2011. 2, 6", + "[51] Zirui Wang, Zhizhou Sha, Zheng Ding, Yilin Wang, and Zhuowen Tu. Tokencompose: Grounding diffusion with token-level supervision. arXiv preprint arXiv:2312.03626, 2023.3", + "[52] Jiarui Xu, Sifei Liu, Arash Vahdat, Wonmin Byeon, Xiaolong Wang, and Shalini De Mello. Open-Vocabulary Panoptic Segmentation with Text-to-Image Diffusion Models. arXiv preprint arXiv:2303.04803, 2023. 2", + "[53] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei" + ], + "bbox": [ + 516, + 92, + 906, + 902 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "13218", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Yang, Burcu Karagol Ayan, et al. Scaling autoregressive models for content-rich text-to-image generation. arXiv preprint arXiv:2206.10789, 2(3):5, 2022. 2", + "[54] Kai Zhang, Lingbo Mo, Wenhu Chen, Huan Sun, and Yu Su. Magicbrush: A manually annotated dataset for instruction-guided image editing. Advances in Neural Information Processing Systems, 36:31428-31449, 2023. 7, 1", + "[55] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 1, 2, 3", + "[56] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 7" + ], + "bbox": [ + 91, + 90, + 483, + 316 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "13219", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Composing Parts for Expressive Object Generation/edecee1f-cf3a-417c-a3d7-2ea937e0cfb0_model.json b/2025/Composing Parts for Expressive Object Generation/edecee1f-cf3a-417c-a3d7-2ea937e0cfb0_model.json new file mode 100644 index 0000000000000000000000000000000000000000..8162e7841e48459aa7fa5e67dadaa21d368e6b1d --- /dev/null +++ b/2025/Composing Parts for Expressive Object Generation/edecee1f-cf3a-417c-a3d7-2ea937e0cfb0_model.json @@ -0,0 +1,2433 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.24, + 0.131, + 0.759, + 0.154 + ], + "angle": 0, + "content": "Composing Parts for Expressive Object Generation" + }, + { + "type": "text", + "bbox": [ + 0.247, + 0.18, + 0.748, + 0.216 + ], + "angle": 0, + "content": "Harsh Rangwani1 Aishwarya Agarwal1 Kuldeep Kulkarni1 \nR. Venkatesh Babu2 Srikrishna Karanam1" + }, + { + "type": "text", + "bbox": [ + 0.306, + 0.233, + 0.688, + 0.251 + ], + "angle": 0, + "content": "\\(^{1}\\)Adobe Research \\(^{2}\\) Indian Institute of Science" + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.281, + 0.199, + 0.364 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.108, + 0.365, + 0.489, + 0.382 + ], + "angle": 0, + "content": "a full body portrait photo of a person wearing a dress with long hair and glasses PartComposer: hair, glasses, dress" + }, + { + "type": "image_caption", + "bbox": [ + 0.204, + 0.281, + 0.28, + 0.288 + ], + "angle": 0, + "content": "StableDiffusion" + }, + { + "type": "image", + "bbox": [ + 0.205, + 0.289, + 0.299, + 0.364 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.326, + 0.281, + 0.374, + 0.288 + ], + "angle": 0, + "content": "Rich Text Gen" + }, + { + "type": "image", + "bbox": [ + 0.3, + 0.289, + 0.382, + 0.364 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.411, + 0.281, + 0.49, + 0.288 + ], + "angle": 0, + "content": "PartComposer (Ours)" + }, + { + "type": "image", + "bbox": [ + 0.38, + 0.289, + 0.496, + 0.364 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.524, + 0.282, + 0.581, + 0.288 + ], + "angle": 0, + "content": "Base Generation" + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.29, + 0.602, + 0.364 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.633, + 0.282, + 0.685, + 0.288 + ], + "angle": 0, + "content": "StableDiffusion" + }, + { + "type": "image", + "bbox": [ + 0.609, + 0.289, + 0.799, + 0.364 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.816, + 0.281, + 0.892, + 0.288 + ], + "angle": 0, + "content": "PartComposer (Ours)" + }, + { + "type": "image", + "bbox": [ + 0.799, + 0.29, + 0.901, + 0.364 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.635, + 0.365, + 0.767, + 0.373 + ], + "angle": 0, + "content": "A photo of a yellow warbler" + }, + { + "type": "image_caption", + "bbox": [ + 0.617, + 0.373, + 0.787, + 0.381 + ], + "angle": 0, + "content": "PartComposer: beak \\(\\rightarrow\\) a hens beak" + }, + { + "type": "image", + "bbox": [ + 0.098, + 0.385, + 0.198, + 0.467 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.199, + 0.385, + 0.299, + 0.467 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.299, + 0.385, + 0.377, + 0.392 + ], + "angle": 0, + "content": "Rich Text Gen" + }, + { + "type": "image_caption", + "bbox": [ + 0.206, + 0.468, + 0.406, + 0.485 + ], + "angle": 0, + "content": "a photo of a white swan, 8k, full hd \nrtComposer: crown → a crown of a peacock" + }, + { + "type": "image", + "bbox": [ + 0.5, + 0.385, + 0.601, + 0.468 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.607, + 0.385, + 0.901, + 0.468 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.504, + 0.468, + 0.9, + 0.485 + ], + "angle": 0, + "content": "A young woman sits at a table in a beautiful, lush garden, painting by claude monet PartComposer: dress \\(\\rightarrow\\) a dress in Uhuie-style" + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.495, + 0.907, + 0.539 + ], + "angle": 0, + "content": "Figure 1. Base Generation (left) Comparison with Methods for Generations with Parts Details (right). PartComposer allows the generation of object images with specified attributes (color, style etc.) of parts for the chosen object in the base text prompt. StableDiffusion and Rich-Text [18] methods with part details either ignore the part instructions or generate inconsistent objects." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.551, + 0.327, + 0.567 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.584, + 0.486, + 0.902 + ], + "angle": 0, + "content": "Image composition and generation are processes where the artists need control over various parts of the generated images. However, the current state-of-the-art generation models, like Stable Diffusion, cannot handle fine-grained part-level attributes in the text prompts. Specifically, when additional attribute details are added to the base text prompt, these text-to-image models either generate an image vastly different from the image generated from the base prompt or ignore the attribute details. To mitigate these issues, we introduce PartComposer, a training-free method that enables image generation based on fine-grained part-level attributes specified for objects in the base text prompt. This allows more control for artists and enables novel object compositions by combining distinctive object parts. PartComposer first localizes object parts by denoising the object region from a specific diffusion process. This enables each part token to be localized to the right region. After obtaining part masks, we run a localized diffusion process in each part region based on fine-grained part attributes and combine them to produce the final image. All stages of PartComposer are based on repurposing a pre-trained diffusion model, which" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.553, + 0.909, + 0.614 + ], + "angle": 0, + "content": "enables it to generalize across domains. We demonstrate the effectiveness of part-level control provided by PartComposer through qualitative visual examples and quantitative comparisons with contemporary baselines." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.646, + 0.645, + 0.661 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.673, + 0.907, + 0.839 + ], + "angle": 0, + "content": "Image generation with large generative diffusion models like StableDiffusion [40], DALLE [38], etc., has become prevalent due to their superior quality and extensive world knowledge. These models are trained on large image-caption datasets and are trained to generate images based on a given text prompt (description). As image composition and creation are creative processes where the artists need control over various parts of the image being generated. However, adding additional details for controlling part appearance in the text prompt either changes the generated image entirely or ignores the part instructions [9] (Fig. 1)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Various works aim to provide improved spatial controls to image generation, as they allow image generation conditioned on segmentation masks [5], edge maps [55], bounding boxes [10, 27] etc. Popular methods such as Control-" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13209" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.228 + ], + "angle": 0, + "content": "Net [55], GLIGEN [27], etc. require specifying training of these conditional modules, which allows for the controllability of these large generative models. Further, there has been some development of training-free approaches [9, 10] which enable controlled generation of objects by modulating the internal cross-attention activations in the diffusion process. This demonstrates that these pre-trained generative models contain information about the spatial parts of the image, and modulating them effectively can lead to image compositions" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.231, + 0.486, + 0.488 + ], + "angle": 0, + "content": "However, despite several attempts [6, 55], the semantic controllability of the image generation is still restricted to specifying details at the object level. The object level details can be restrictive, as often creative designers synthesize object parts (e.g. shirt, trouser etc.) and then compose them [23]. Further, variations in semantic parts are often very distinctive and are uniquely used to identify objects [12, 30]. For example, we usually identify species of birds by their unique beak. Due to this, semantic understanding and recognition of parts have been widely studied as a topic in computer vision [12, 13, 21, 31]. Hence, providing controllability for image synthesis at the object semantic part level can enable a large variety of image compositions [23]. Towards this goal, we introduce PartComposer, which provides users with an interface through which they can select the object in the scene and provide a semantic part-level description with fine-grained details for the object generation." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.492, + 0.485, + 0.868 + ], + "angle": 0, + "content": "In PartComposer, we develop a scheme to extract part-level localization masks from the Diffusion model. We introduce a parallel part diffusion process that generates masks for the object parts. The core idea of the approach is that by forcing the part diffusion model to specifically denoise only the object region in the image, it is possible to understand the locations of various parts in the object. After denoising just the object through the part model, the information inside the part diffusion model present in attention maps can be used to generate the masks for various object parts. In the following Part Generation stage, PartComposer utilizes the part masks and attributes of each part the user provides. We enable users to provide a highly expressive specification of parts by using a Rich-Text [18] interface, which allows the specifying attributes, such as style, color, etc., for each part. For the final image generation, taking inspiration from recent studies like Rich-Text Generation [18], Multi-Diffusion [5] etc., we compose the various object parts, by running parallel masked diffusion process for each part while combining them periodically into the image. This combination enables harmonious composition of parts, and also masking ensures only local modifications to the regions corresponding to each part of the object specified. In the PartComposer method, we only use the pre-trained StableDiffusion model, making it a generalizable and training-free approach." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.485, + 0.902 + ], + "angle": 0, + "content": "We extensively test the proposed PartComposer approach for the zero-shot object part segmentation, which is a chal" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.908, + 0.227 + ], + "angle": 0, + "content": "lenging setup in computer vision. We evaluate the part segmentation approach on DeepFashion [29] and CUB200 [50] datasets, where our method significantly outperforms the baseline StableDiffusion (SD). Further, to evaluate the PartComposer image composition abilities, we also provide quantitative results along with a user preference study, where our method significantly outperforms the baselines in generating images consistent with the described parts (Fig. 1). We summarize the core contributions of our paper below:" + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.228, + 0.907, + 0.302 + ], + "angle": 0, + "content": "1. We introduce PartComposer, a training-free method that enables the generation of object images by using provided fine-grained details for the parts of the object. For example, while generating a bird image we can specify a detailed description of its beak." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.303, + 0.907, + 0.408 + ], + "angle": 0, + "content": "2. In PartComposer, we introduce a novel Part-Diffusion process, which localizes and provides masks for parts of a base object generated by the Diffusion model (Fig. 3). To localize object parts, we introduce a novel segmentation scheme that uses the attention maps of the base diffusion and part diffusion process to obtain accurate masks for localized parts." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.409, + 0.907, + 0.5 + ], + "angle": 0, + "content": "3. In PartComposer, after localization, we enable the generation of parts from the pre-trained diffusion model based on the Rich-Text description for the parts provided by the user. PartComposer then composes the image to harmoniously blend all object parts with the background by combining localized diffusion paths (Fig. 2)." + }, + { + "type": "list", + "bbox": [ + 0.515, + 0.228, + 0.907, + 0.5 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.513, + 0.663, + 0.529 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.539, + 0.909, + 0.749 + ], + "angle": 0, + "content": "Text-to-Image Generative Models. The text-to-image models synthesize images by following a textual description provided as a prompt. These models have recently become mainstream due to their superior image generation quality and significant knowledge base. This is an outcome of the availability of large-scale image caption datasets [42] and highly parallelized GPU clusters. Almost all kinds of generative models, such as GANs [22], Autoregressive [41, 53], and Diffusion models [38, 40], have shown significant improvements in quality with training on these large image caption datasets. Among these models, the StableDiffusion (SD) [40] models, based on denoising diffusion in latent space are popular due to their open-source nature, which we also utilize for experiments in our work." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Text-to-Image Models for Downstream Tasks. Generative models, in general, have been useful for various downstream tasks, particularly ones based on per-pixel prediction like Segmentation [1, 11], Depth Estimation [7], etc. As layers near the image generation output have features that capture the pixel-pixel relation [7]. With the large-scale text-to-image generative models, these models often perform very competitively [46, 52] to discriminative methods, on tasks like segmentation. However, one commonality among most of these segmentation methods is that they operate at the" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13210" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.134, + 0.104, + 0.862, + 0.322 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.323, + 0.908, + 0.365 + ], + "angle": 0, + "content": "Figure 2. PartComposer takes the input of a base prompt, the selected object o, and a Rich-Text description ( i.e., footnote, colors, etc.) of the parts. The Part Diffusion process generates the object masks for the specified parts. Then, PartComposer runs a parallel Region Diffusion process to generate attributes of the specified parts, guided by base generation's intermediate outputs." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.373, + 0.482, + 0.448 + ], + "angle": 0, + "content": "granularity of object or instance level. Text-to-image models here are at an advantage as the usual image captions describe the scene at an object level. In this work, we take a step further in exploring the object part-level knowledge of these text-to-image diffusion models." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.45, + 0.483, + 0.585 + ], + "angle": 0, + "content": "Part Discovery and Segmentation. Part Discovery and Attribution were an integral part of computer vision pipelines classically, as these approaches were robust to viewpoint variations [24, 31, 48]. In deep learning, the unsupervised (self-supervised) approaches for part discovery like SCOPS [21] and Unsup-Parts [12] became popular as they generalize to object parts across categories. In this work, we go one step ahead and operate in a zero-shot unsupervised part segmentation setting, generating part masks from T2I models." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.587, + 0.483, + 0.752 + ], + "angle": 0, + "content": "Controllable Image Generation. Masks, bounding boxes, edge maps, depth maps, etc., have been explored to control the generations of text-to-image diffusion models [6, 14, 27, 35, 36, 55] in addition to text. Further, various other approaches [2-4, 9, 34, 45, 49, 51] achieve control of image semantics through modulating diffusion models. Despite this, the text-to-image generation control at the object part level is under-explored; a recent work [33] tries to do it in a controlled supervised setup using part-masks. Contrary to that, in our work PartComposer we explore a generalized training-free zero-shot setting." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.769, + 0.182, + 0.784 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.795, + 0.483, + 0.9 + ], + "angle": 0, + "content": "In this section, we introduce PartComposer, our method to synthesize objects based on the description of parts of the objects. In PartComposer, we ask the user to specify a base prompt and the token for the object for which it wants to synthesize parts. Then, we provide the user a rich-text [18] editor (Fig. 2) to specify the parts and their description. PartComposer involves two diffusion steps, a) Part Localiza" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.373, + 0.908, + 0.494 + ], + "angle": 0, + "content": "tion: In the early diffusion stage, we get a mask for the object we want to divide into parts. Then, we perform denoising in later stage from a U-Net condition on parts to fill the masked region of the object, during which it learns to denoise different object parts. Due to this, the attention maps for various parts highlight the correct part region, which we use to extract the part mask. The infilling process is the major contribution of the PartComposer method (Fig. 3). b)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.495, + 0.909, + 0.616 + ], + "angle": 0, + "content": "Part Generation: For generating parts, we combine region-specific diffusion processes for various parts by iteratively merging them inspired by MultiDiffusion [5], Rich-Text Generation [18] etc. However, till now, most of these works have combined the diffusion process for generating objects; for the first time, we have demonstrated its effectiveness in generating object parts. We provide an overview of the PartComposer pipeline in Fig. 2." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.623, + 0.665, + 0.639 + ], + "angle": 0, + "content": "3.1. Problem Setup" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.645, + 0.909, + 0.795 + ], + "angle": 0, + "content": "We consider each span of tokens \\(\\mathbf{p}_i\\) as indicative of describing one part of the object, with attributes \\(\\mathbf{a}_i\\) describing its overall appearance. Our design choices are based on the rich-text generation [18], and we allow the user to specify the following type of attributes for part generation (Fig. 2): Part Description (i.e. footnote). It is an important attribute that specifies the part of an object; for example, for the part \\((\\mathbf{p}_i)\\) 'crown' of a bird, we can specify the attribute \\((\\mathbf{a}_i^p = \\text{'a crown of a peacock')\\). This helps specify novel part descriptions, which can lead to artistic novel compositions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.905, + 0.87 + ], + "angle": 0, + "content": "Font Color. It helps us specify the exact RGB values for the color attribute \\(\\mathbf{a}_i^c\\) we want for the object part \\(\\mathbf{p}_i\\). The exact value of RGB allows fine-grained control over the color of the desired part, whereby just specifying specific colors like 'brick red' leads to ignorance by Stable Diffusion [18]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.905, + 0.9 + ], + "angle": 0, + "content": "Font Style. The font style is indicative of the specific artistic style \\( \\mathbf{a}_i^s \\) like 'of Claude Monet' and 'of Van Gogh' when" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "13211" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.089, + 0.905, + 0.316 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.329, + 0.908, + 0.361 + ], + "angle": 0, + "content": "Figure 3. Localization. We obtain the mask \\((\\mathcal{M}_o)\\) for object in middle of diffusion process. We then denoise object in masked region, using parts \\(\\mathbf{p}_i\\) conditioned U-Net. Due to part denoising, the attention maps of \\(\\mathbf{p}_i\\) can now localize parts as masks \\(\\mathbf{M}_{\\mathbf{p}_i}\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.366, + 0.483, + 0.412 + ], + "angle": 0, + "content": "synthesizing images of paintings. This instructs the model to generate a part following a specific style for the given prompt and blend it with other parts of the object (Sec. 5)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.412, + 0.483, + 0.443 + ], + "angle": 0, + "content": "Font Size. The font size controls the relative size of each part in a generation [19]. We use the \\(\\mathbf{a}_i^w\\) to denote size." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.451, + 0.262, + 0.465 + ], + "angle": 0, + "content": "3.2. Part Localization" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.472, + 0.486, + 0.76 + ], + "angle": 0, + "content": "For part localization, we first run the base stable diffusion model for the given text prompt and extract the token map \\(\\mathcal{M}_o\\) for the object \\(\\mathbf{o}\\) specified by the user, by following the technique of clustering the self-attention maps [18, 37]. We obtain the mask using this clustering after denoising between the initial step \\((T)\\) until a threshold time step \\(T_{th}\\). After obtaining the binary object mask \\(\\mathcal{M}_o\\) using attention masks [18, 37], we run two parallel diffusion processes where one contains input from the base prompt and the other contains input from the part prompts \\(\\hat{\\mathbf{p}} = [\\mathbf{p}_1,\\mathbf{p}_2,\\dots,\\mathbf{p}_n]\\), a token \\(\\mathbf{p}_i\\) for each of the parts specified in the part text prompt. We denote the \\(\\hat{\\mathbf{b}}\\) to denote the base text prompt and \\(\\hat{\\mathbf{p}}\\) to denote the part prompt. For the time \\(t\\leq T_{th}\\) (by default, we use \\(T_{th}\\approx T / 2\\)), we denoise the U-Net in the object region by denoising it with both the combination of the Part Prompt output and Base Prompt output. Due to this, the Part-Based U-Net gets the information regarding various parts in the object (Fig. 3). We now mathematically define the output noise \\(\\epsilon_t\\) for the diffusion process with Part-Diffusion below:" + }, + { + "type": "equation", + "bbox": [ + 0.099, + 0.784, + 0.484, + 0.801 + ], + "angle": 0, + "content": "\\[\n\\epsilon_ {t} = \\alpha \\mathcal {M} _ {o} \\odot D (x _ {t}, \\hat {\\mathbf {p}}, t) + (1 - \\alpha \\mathcal {M} _ {o}) \\odot D (x _ {t}, \\hat {\\mathbf {b}}, t) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.484, + 0.902 + ], + "angle": 0, + "content": "here, the \\(\\alpha\\) is the hyper-parameter controlling the strength of the part prompt diffusion output, and \\(D\\) is the output of the pre-trained U-Net of a text-to-image diffusion model. The above denoising process is followed for t steps until the last step to produce a base image corresponding to the given prompt \\(\\hat{b}\\). Keeping a high \\(T_{th}\\) and low \\(\\alpha\\), makes minimal" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.366, + 0.907, + 0.442 + ], + "angle": 0, + "content": "changes in output as if the denoising diffusion was done with original prompt \\(\\hat{b}\\) (See Suppl. Sec.). With this part denoising of object, we obtain attention masks, from which we extract the localization information of part \\(\\mathbf{p}_t\\). We now describe the process of obtaining the part masks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.447, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Token Maps for Parts. We first take the part tokens \\(\\hat{\\mathbf{p}}\\), which are a concatenation of the part names (i.e. 'beak crown wings'), which may not make a meaningful text prompt. Hence, we initialize text embeddings for all these tokens \\(\\mathbf{p}_i\\) by passing the meaningful text prompt having the following template: \"A photo of \\(\\mathbf{p}_i\\) of a o\" where \\(\\mathbf{p}_i\\) is the object part name and o is the object name. This serves two purposes: first, it makes the text embedding meaningful, and second, it introduces some invariance from the order of part specification in part prompt. These embeddings are then passed as text embeddings to the Part U-Net for denoising. After running the denoising process, we aggregate the self-attention maps across multi-heads and time steps (from 32 × 32 resolution) for both the base and part U-Net diffusion branches, taking inspiration from works that demonstrate that attention can localize objects [9, 18, 30, 44]. We then perform spectral clustering on these attention maps to form k segmentation maps \\(\\hat{\\mathbf{M}}\\) (\\(32 \\times 32\\)), based on pixel-pixel similarity. To attribute these K-segments to the part specified by the user, we aggregate the cross attention of Part U-Net diffusion process. For each token \\(\\mathbf{p}_j\\), we obtain the cross-attention score as follows: \\(\\hat{\\mathbf{m}}_j = \\frac{c_j}{\\sum_k c_k}\\), where \\(c_j\\) is the cross attention score for each token. We proceed by aggregating the attention heads to obtain the average cross-attention scores and resizing them to \\(32 \\times 32\\), obtaining \\(\\hat{\\mathbf{m}}\\). We remove the start of text (\\(\\langle \\mathrm{sot} \\rangle\\)) token for cross-attention and re-normalize it [9]. In other works [18, 37], as the tokens correspond to the objects being generated in the image, it's sure that token maps will be meaningful. However, this is not true for parts, as Part U-Net might not localize some" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13212" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.152 + ], + "angle": 0, + "content": "parts. For determining if the part is localized, we propose to look at the max value of cross attention map spatially across the pixels; if we find that the following condition is met, the part is localized:" + }, + { + "type": "equation", + "bbox": [ + 0.165, + 0.159, + 0.484, + 0.189 + ], + "angle": 0, + "content": "\\[\nL (j) = \\mathbb {1} \\left\\{\\max \\left(\\hat {\\mathbf {m}} _ {j}\\right) \\geq (1 - \\delta) \\frac {1}{K} \\right\\}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.195, + 0.485, + 0.256 + ], + "angle": 0, + "content": "Here \\(\\delta\\) is a hyperparameter, and \\(K\\) is the number of parts. Is This condition is robust in finding the localized parts (See Sec. 5 for ablation). For the parts \\(\\mathbf{p}_i\\), which are localized, we normalize the cross-attention map:" + }, + { + "type": "equation", + "bbox": [ + 0.108, + 0.262, + 0.466, + 0.297 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {m}} _ {j} = L (j) \\frac {\\hat {\\mathbf {m}} _ {j} - \\min (\\hat {\\mathbf {m}} _ {j})}{\\max (\\hat {\\mathbf {m}} _ {j}) - \\min (\\hat {\\mathbf {m}} _ {j})} + (1 - L (j)) \\hat {\\mathbf {m}} _ {j}.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.302, + 0.485, + 0.501 + ], + "angle": 0, + "content": "We follow a dot-product-based protocol to assign each K cluster in the self-attention masks to a part, unlike the average attention protocol in the previous works [18, 37]. We find that dot product of normalized cross attention scores \\(\\hat{\\mathbf{m}}_j\\) for each token with self-attention masks \\(\\hat{\\mathbf{M}}_j\\) works better, as the attention maps in Part U-Net are noisy. Still, they are often correct for the regions that are localized in one specific area only (Suppl. Fig. 14). Hence, the dot product protocol favors those maps that are only localized in some areas of the image and don't have high attention values across all parts of the object. After obtaining the dot product scores for each part, we assign \\(\\hat{\\mathbf{M}}_j\\) to \\(\\mathbf{p}_i\\) with the highest scores. The mask for the part \\(\\mathbf{p}_i\\) is finally given as the following:" + }, + { + "type": "equation", + "bbox": [ + 0.09, + 0.522, + 0.484, + 0.554 + ], + "angle": 0, + "content": "\\[\n\\mathbf {M} _ {\\mathbf {p} _ {i}} = \\left\\{\\cup_ {j} \\hat {\\mathbf {M}} _ {j} \\mid \\operatorname {a r g m a x} _ {i} \\hat {\\mathbf {M}} _ {j} \\cdot \\hat {\\mathbf {m}} _ {i} = i \\text {a n d} \\hat {\\mathbf {M}} _ {j} \\cdot \\hat {\\mathbf {m}} _ {i} \\geq \\epsilon , \\right\\} \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.555, + 0.485, + 0.63 + ], + "angle": 0, + "content": "Here \\(\\epsilon\\) is the hyperparameter which controls the minimum similarity required between the attention mask and the part token. We combine the additional attention masks unassigned to any token, and name them as the background (other) token \\(\\mathbf{M}_b\\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.638, + 0.254, + 0.652 + ], + "angle": 0, + "content": "3.3. Part Generation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.66, + 0.486, + 0.812 + ], + "angle": 0, + "content": "We follow a similar protocol for the generation of part segments as in Rich-Text Generation. We tailor the rich text generation to compose the part regions of the object in the image in place of the original scene composition. We describe the part generation protocol below briefly and refer readers to Rich-Text Gen [18] for more details. For each part \\(\\mathbf{p}_i\\) we run a region diffusion process, which runs in parallel for all the parts. We then combine the region diffusion processes to obtain the final noise prediction \\(\\epsilon_t\\) as the masked \\(\\mathbf{M}_{\\mathbf{p}_i}\\) sum of the denoiser outputs:" + }, + { + "type": "equation", + "bbox": [ + 0.103, + 0.832, + 0.484, + 0.863 + ], + "angle": 0, + "content": "\\[\n\\epsilon_ {t} = \\sum_ {i} \\mathbf {M} _ {\\mathbf {p} _ {i}} \\epsilon_ {t, \\mathbf {p} _ {i}} = \\sum_ {i} \\mathbf {M} _ {\\mathbf {p} _ {i}} \\odot D (x _ {t}, f (\\mathbf {p} _ {i}, \\mathbf {a} _ {i}), t) \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "where \\(D\\) is the pre-trained U-Net model, and \\(f(\\mathbf{p}_i,\\mathbf{a}_i)\\) is the text description of the part \\(\\mathbf{p}_i\\) constructed using the following" + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.09, + 0.899, + 0.177 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.19, + 0.907, + 0.218 + ], + "angle": 0, + "content": "Figure 4. Segmentation masks for the parts that are localized by Part Diffusion for DeepFashion (above) and CUB200 (below)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.223, + 0.907, + 0.419 + ], + "angle": 0, + "content": "process using the part tokens \\(\\mathbf{p}_i\\) and attributes \\(\\mathbf{a}_i\\). Initially the text \\(f(\\mathbf{p}_i, \\mathbf{a}_i) = \\mathbf{p}_i\\), is set to part token itself. In case the part description (i.e. footnote) is available we set the \\(f(\\mathbf{p}_i, \\mathbf{a}_i) = \\mathbf{a}_i^p\\), further if the style attribute is available we do \\(f(\\mathbf{p}_i, \\mathbf{a}_i) = f(\\mathbf{p}_i, \\mathbf{a}_i) + \\text{in style of} + \\mathbf{a}_i^s\\). In case the color attribute is also specified, the nearest named color \\(\\hat{\\mathbf{a}}_i^c\\) (e.g. red) for the specific RBG color \\(\\mathbf{a}_i^c\\) is found, and \\(f(\\mathbf{p}_i, \\mathbf{a}_i) = \\hat{\\mathbf{a}}_i^c + ' ' + f(\\mathbf{p}_i, \\mathbf{a}_i)\\). The string \\(f(\\mathbf{p}_i, \\mathbf{a}_i)\\) is the text input for the Diffusion to generate the part \\(\\mathbf{p}_i\\). We use the base prompt \\(\\hat{\\mathbf{b}}\\) as \\(f(\\mathbf{p}_i, \\mathbf{a}_i)\\) for the background masked region \\(\\mathbf{M}_b\\). Combining different diffusion outputs at every time \\(t\\) helps generate a harmonious image after blending the defined parts of the object." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.42, + 0.909, + 0.554 + ], + "angle": 0, + "content": "Following Rich-Text [18], we also utilize the gradient guidance [16, 20] by taking the gradient of MSE loss between the estimated original image and the color value \\(\\mathbf{a}_i^c\\) specified by the user. The gradient guidance helps generate the exact RGB color for part \\(\\mathbf{p}_i\\), which is impossible with just text guidance [18]. Further, as we use the base text framework of Rich-Text, we also can specify font size attribute to \\(\\mathbf{a}_i^f\\) to control the relative importance of each part in the object image." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.555, + 0.909, + 0.661 + ], + "angle": 0, + "content": "Preservation of Other Parts. As we only intend to modify the object o from the original prompt, we also use the Self-Injection techniques from Plug and Play [47] to maintain the overall structure of the background from base prompt generation. Further, to ensure that our diffusion trajectory follows the same path as the base, the background region is also blended with base noise generation outputs." + }, + { + "type": "equation", + "bbox": [ + 0.59, + 0.668, + 0.907, + 0.686 + ], + "angle": 0, + "content": "\\[\nx _ {t} = \\mathbf {M} _ {b} \\odot x _ {t} ^ {\\text {b a s e}} + (1 - \\mathbf {M} _ {b}) \\odot x _ {t} \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.692, + 0.909, + 0.798 + ], + "angle": 0, + "content": "In our case, the attention maps out of the object \\(\\mathbf{Mask}\\mathcal{M}_o\\) and all the object parts not assigned to any part token comprise the background \\(\\mathbf{M}_b\\), and we start this blending process at \\(t = T_{blend}\\). We find this to be very useful in preserving the structure of the other parts of the image and just generating the described parts in the localized region. We provide an overview of the complete process in Fig. 2." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.809, + 0.73, + 0.827 + ], + "angle": 0, + "content": "4. Experimental Analysis" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.834, + 0.792, + 0.849 + ], + "angle": 0, + "content": "4.1. Evaluation of Part Localization" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.908, + 0.901 + ], + "angle": 0, + "content": "We first evaluate the part localization module, which is based on the novel idea of denoising only the object region with the part-based diffusion outputs." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13213" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.908, + 0.119 + ], + "angle": 0, + "content": "Table 1. Comparison to Prior Works for the unsupervised part segmentation task. We follow Unsup-Part [12] for evaluation protocols and baseline results for \\((\\mathbf{K} = 4)\\) parts. We report clustering-based NMI and ARI metrics, which are higher for better segmentation outputs." + }, + { + "type": "table", + "bbox": [ + 0.168, + 0.122, + 0.833, + 0.317 + ], + "angle": 0, + "content": "
DeepFashion DatasetCUB200 Dataset
MethodFG-NMIFG-ARINMIARIFG-NMIFG-ARINMIARI
Unsupervised Learning
SCOPS [21]30.727.656.681.439.117.924.47.1
Unsup-Parts [12]44.846.668.190.646.021.043.519.6
DFF [11]----32.414.325.912.4
Unsupervised Zero-Shot
Rich-Text [18]16.05.248.358.73.10.33.10.3
StableDiffusion12.03.540.670.98.00.63.30.6
PartComposer(Ours)24.718.048.073.420.59.218.57.7
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.322, + 0.487, + 0.534 + ], + "angle": 0, + "content": "Implementation Details. We use the StableDiffusion (SD) version 2.1 for our experimentation purposes. We use the DDIM Scheduler [43] with 50 steps to generate results for SD2.1 to evaluate the part localization process. As ground truth is not available for the part masks of the generated images, we use test sets of the commonly used DeepFashion [29] and CUB-200 [50] datasets for evaluation. These datasets are the standard datasets used for the evaluation of the unsupervised part segmentation approaches. The DeepFashion dataset contains images along with their part segmentation masks, divided into 14 categories of labels. The CUB200 dataset contains the key point annotations for the 14 key points specified for the bird categories. We provide further details in Suppl. Section." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.538, + 0.487, + 0.84 + ], + "angle": 0, + "content": "Baselines and Problem Setting. As we operate in the setting of Zero-shot (i.e. no training) Unsupervised Part Segmentation, there are no previous works that report results in such a challenging setting. Hence, we provide results for the unsupervised learning approaches to facilitate comparison. We provide results for the SCOPS [21], which utilizes the internal features of the VGG model to train a model based on self-supervised loss functions to predict parts robustly across categories. The other stronger self-supervised baseline is Unsup-Parts [12], which uses contrastive loss functions to train a network based on equivariance and other vision properties to cluster the object regions into semantic parts. In addition to this, we also report results for the DFF [11] as they also operate in the same setting. We want to highlight that these unsupervised approaches require either training a neural network or performing clustering on the complete training data to segment parts. In contrast, our approach is training free and operates in a zero-shot fashion. Hence, the performance of zero-shot approaches is not directly comparable to unsupervised methods." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.485, + 0.902 + ], + "angle": 0, + "content": "Zero Shot Unsupervised Part Segmentation. As there is no benchmark to evaluate the part localization for generated images, we use the existing dataset of DeepFashion and CUB200 to obtain our results. To obtain segmentation for" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.322, + 0.909, + 0.656 + ], + "angle": 0, + "content": "each image, we first invert the image into the diffusion latent space using Null-Text Inversion [32] method (see Sec. 5 for ablation). We use the BLIP-V2 [25] captioner provided in Diffusers to obtain approximate text prompts for inversion. After providing the desired image and prompt, the Null-Text inversion provides us with inverted latent and unconditioned time embeddings to reconstruct the image of the dataset. We then construct a StableDiffusion baseline in which, in addition to the prompts, we append the list of part tokens \\(\\mathbf{p}\\) to the prompt. We then use the segmentation algorithm described above in Sec. 3.2 to extract token maps corresponding to the part tokens \\(\\mathbf{p}_i\\). We also evaluate a Rich-Text [18] baseline segmentation algorithm for the same. For the proposed part-denoising approach based on parallel diffusion in PartComposer, we first get the inverted latents and embeddings. Then, we generate a base diffusion process to reconstruct the image and use Part Diffusion to fill the parts of the image. To very fairly compare the StableDiffusion performance with part-denoising, we keep all things the same except the part diffusion process to get masks. Further across these baselines, we use low classifier guidance to ensure proper reconstruction of the dataset images (See Suppl. for details)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.659, + 0.911, + 0.902 + ], + "angle": 0, + "content": "Evaluation Protocol. We use the standard experimental protocol as used by earlier works [12] to facilitate the right point of comparison. We want to point out that in our work, the part name (e.g., beak, etc.) is associated with the localized mask, but the unsupervised approaches produce \\((\\mathrm{K} = 4)\\) parts without any part names. Hence, to make a fair comparison, we create 4 clusters of parts based on their locality, and finally generate segmentation masks with a maximum of 4 clusters. We provide the exact mapping of the part names to cluster labels in the Suppl. Section. We compare the specifically designed metrics [12] of NMI (Normalized Mutual Information) and ARI (Adjusted Rand Index) of the predicted cluster labels with part masks in the case of DeepFashion and key points in case of the CUB dataset. We report the metrics along with their foreground variants (FG-NMI and FG-ARI) for a holistic comparison of performance." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13214" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.115, + 0.09, + 0.268, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.115, + 0.189, + 0.268, + 0.285 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.279, + 0.091, + 0.429, + 0.183 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.279, + 0.189, + 0.43, + 0.279 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.314, + 0.278, + 0.397, + 0.285 + ], + "angle": 0, + "content": "PartCraft (Ours)" + }, + { + "type": "image", + "bbox": [ + 0.429, + 0.091, + 0.58, + 0.181 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.465, + 0.181, + 0.545, + 0.188 + ], + "angle": 0, + "content": "StableDiffusion" + }, + { + "type": "image", + "bbox": [ + 0.58, + 0.092, + 0.732, + 0.181 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.616, + 0.181, + 0.694, + 0.188 + ], + "angle": 0, + "content": "InstructPix2Pix" + }, + { + "type": "image", + "bbox": [ + 0.73, + 0.092, + 0.88, + 0.181 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.767, + 0.181, + 0.841, + 0.188 + ], + "angle": 0, + "content": "Rich Text Gen" + }, + { + "type": "image", + "bbox": [ + 0.279, + 0.189, + 0.43, + 0.278 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.314, + 0.278, + 0.397, + 0.285 + ], + "angle": 0, + "content": "PartCraft (Ours)" + }, + { + "type": "image", + "bbox": [ + 0.43, + 0.189, + 0.58, + 0.278 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.465, + 0.278, + 0.544, + 0.285 + ], + "angle": 0, + "content": "StableDiffusion" + }, + { + "type": "image", + "bbox": [ + 0.58, + 0.189, + 0.73, + 0.278 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.617, + 0.278, + 0.693, + 0.285 + ], + "angle": 0, + "content": "InstructPix2Pix" + }, + { + "type": "image", + "bbox": [ + 0.73, + 0.189, + 0.88, + 0.279 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.768, + 0.278, + 0.84, + 0.285 + ], + "angle": 0, + "content": "Rich Text Gen" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.292, + 0.908, + 0.335 + ], + "angle": 0, + "content": "Figure 5. Qualitative Comparison for PartComposer. (above) We provide base prompt 'a photo of a flamingo' and part prompt as 'beak - a pelicans beak'. (below) We generate a photo of a man, with part prompt as 'black jacket and blue jeans'. We find that other baselines either ignore instruction or change the entire composition. On the contrary, PartComposer can correctly localize parts and generate details." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.343, + 0.485, + 0.675 + ], + "angle": 0, + "content": "Results. We tabulate the results for all the baselines in Table 1. Our approach, PartComposer, significantly improves over the baselines by \\(\\geq 5\\) points in NMI and ARI across most metrics in the Unsupervised Zero-Shot setting. Further, in some cases our methods, NMI and ARI, are near the unsupervised learning approaches. This demonstrates that non-trivial part masks can be generated by harnessing the power of the pre-trained diffusion models, and they can serve as a good initial prior for learning part segmentation approaches by using them as base pre-trained models. We provide qualitative results for PartComposer, where we observe that PartComposer can associate the right part with the correct label (Fig 4). On the contrary, in the StableDiffusion baseline, the parts often get assigned to the wrong part tokens, which is the cause of degraded performance (Suppl. Fig. 16). We also observe that in the case of our algorithm, if the part gets localized based on the max condition defined in the segmentation algorithm, the part mask found is usually in the right region (See Fig. 4 and Table 2). This demonstrates the effectiveness of our part segmentation procedure, which often sides with not localizing objects rather than performing arbitrary assignments." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.683, + 0.472, + 0.699 + ], + "angle": 0, + "content": "4.2. PartComposer Image Generation Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.484, + 0.825 + ], + "angle": 0, + "content": "We now evaluate the object composition ability of the proposed approach, PartComposer, compared to the baselines. In this section, we use the base generation model as a StableDiffusion (SD) 1.5 model to do a fair comparison with the Rich-Text baseline. We used the same generation setting of a PNDM [28] scheduler with 41 steps and suggested classifier guidance 8.5. To perform a fair comparison, we keep all the parameters same as that of the Rich-Text [18] baseline." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Baselines. For the task of generating the object image based on part-level details of the specified object, we use the strong baselines of Rich-Text Generation and InstructPix2Pix [8]. The Rich-Text Generation method generates the base image itself, whereas the InstructPix2Pix method requires us to" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.343, + 0.907, + 0.494 + ], + "angle": 0, + "content": "generate the base image. We also evaluate the standard StableDiffusion baseline in which we add all the part details in the text prompt to generate the desired image. We compare all these baselines with our proposed method, PartComposer, while ensuring that the base parameters and seed are the same for the base image generation. A recent work [33] regarding part generation operates in a fully supervised setup using part masks and cannot be used to create novel parts based on text instruction, making it incomparable. We defer specific details for the baselines to the Suppl. Section." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.501, + 0.909, + 0.773 + ], + "angle": 0, + "content": "Visual Comparison. We provide a visual comparison for the baselines for the a) bird image distinctive part generation and b) human image part generations. Across both cases (Fig. 5), our proposed method, PartComposer, only modifies the desired beak region and replaces it with the iconic pelican beak. The StableDiffusion baseline modifies the image completely while ignoring instructions (Supp. Fig. 12). In the Rich-Text generation, the full region corresponding to the bird gets modified instead of only the specified prompt in base generation. For the editing-based InstructPix2Pix method, we observe that it modifies the base image at a global level and cannot localize modification to the desired part region (Fig. 5). We further compare and find that recent models like SD3.5, Inpainting [17] and recent SotA editing methods [15, 26, 47, 54] can still not follow exact part instructions as seen in Supp. Fig. 8, 9 and 10. In contrast, observe that PartComposer-generated images are novel aesthetic compositions that follow the part-level instructions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.781, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Quantitative Evaluation. As the part-specific generation is fine-grained, we perform the both automatic and user-study evaluations. We performed a user study by inviting 50 participants (28 responded with a full survey). We evaluate the model by asking questions and evaluating metrics on three orthogonal aspects i.e., a) Localization of part generation in comparison to the base image (through LPIPS [56]), b) Text-Consistency of generated parts by measuring CLIP" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13215" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.083, + 0.486, + 0.111 + ], + "angle": 0, + "content": "Table 2. Ablation Analysis of PartComposer for Part Localization, showing performance improvement with each component." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.115, + 0.462, + 0.317 + ], + "angle": 0, + "content": "
PartComposer (Ablations)
MethodFG-NMIFG-ARI
PartComposer35.411.0
w/o Null-Text Inversion23.15.2
w/o Max Localization21.32.8
w/o Dot Product Localization23.75.0
w/o Independent Text31.28.9
PartComposer (Clustering)
PartComposer (K = 9)35.411.0
PartComposer (K = 4)20.42.8
PartComposer (K = 14)35.210.3
" + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.33, + 0.185, + 0.493 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.197, + 0.33, + 0.289, + 0.493 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.293, + 0.33, + 0.385, + 0.492 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.391, + 0.33, + 0.482, + 0.492 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.505, + 0.484, + 0.561 + ], + "angle": 0, + "content": "Figure 6. PartComposer generalizes over domains as shown by Claude Monet's painting style (above), where we specify the dress part of women to follow styles like Van Gogh and Ukiyo-e. We perform similar modifications to Woodie's image." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.572, + 0.484, + 0.737 + ], + "angle": 0, + "content": "similarity of the part detailed text prompt with the generated image and c) Aesthetic Quality of the generated image through LAION-5B [42] CLIP Aesthetic scorer. In total, we collect about 3.5k opinions. In Fig. 7, we summarize the results in which we observe that PartComposer is significantly preferred over the other baselines in terms of localized and consistent Part Generation. The automated evaluation results are provided in the table below, which also follow a similar trend as in the user study, demonstrating the effectiveness of PartComposer in a localized part generation while ensuring the aesthetics score is similar to that of the base model." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.754, + 0.317, + 0.77 + ], + "angle": 0, + "content": "5. Analysis and Discussion" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.485, + 0.901 + ], + "angle": 0, + "content": "Part Diffusion Based Localization. We ablate the components we have introduced in the PartDiffusion process of PartComposer (Sec. 3.2). We provide an analysis of the effect of using a) null-text inversion, b) max-based localization c) usage of dot product-based protocol in the part assignment, and d) independent text embeddings. We have used a subset of CUB-200 images to perform all evaluations, which is kept fixed across ablations. We tabulate the ablations in" + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.09, + 0.907, + 0.117 + ], + "angle": 0, + "content": "Figure 7. User Study and Quantitative Results for part-based image generation baselines." + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.121, + 0.909, + 0.34 + ], + "angle": 0, + "content": null + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.346, + 0.905, + 0.394 + ], + "angle": 0, + "content": "
PartComposerStableDiffusionInstructPix2PixRich-Text
LPIPS0.1680.4670.1890.243
CLIP0.2010.1830.1930.187
Aesthetic5.665.685.635.65
" + }, + { + "type": "table_caption", + "bbox": [ + 0.512, + 0.41, + 0.907, + 0.44 + ], + "angle": 0, + "content": "Table 2. We observe that all the components introduced in PartComposer contribute significantly to the performance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.445, + 0.909, + 0.628 + ], + "angle": 0, + "content": "Generalization of PartComposer. In Fig. 6, above a painting in Claude Monet style generated by the base StableDiffusion method. We then use PartComposer to specify the dress part of the women's token, to dark green color in Monet Style, orange color dress in Van Gogh Style (middle), and green color dress in Ukiyo-e style. In Fig. 6, we provide results for the 'Woody' character from Toy Story. These results show that PartComposer can generate shirt, pants, and dress color variations, leading to aesthetic image combinations in synthetic and natural (Supp. Fig. 13) domains. This zero-shot generalization across domains demonstrates the creative activities that can be enabled with PartComposer." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.659, + 0.634, + 0.674 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.69, + 0.909, + 0.825 + ], + "angle": 0, + "content": "In this work we introduce PartComposer, a method to generate object images with fine-grained attribute details specified at the part level, using an expressive Rich-Text interface. The PartComposer method introduces a novel part diffusion process, which is responsible for denoising objects using the part features, and then utilizes a region-specific diffusion process to generate part details and compose the final image. PartComposer serves as initial work enabling rich-text-based training-free part-level control for SD models." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.909, + 0.9 + ], + "angle": 0, + "content": "Limitations. We find that part generation is bottlenecked with the part understanding. If the part is localized correctly, the text-to-image model can generate specified details in PartComposer. Hence, improving the part-level localization of these models is a good direction for future works." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13216" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.171 + ], + "angle": 0, + "content": "[1] Rameen Abdal, Peihao Zhu, Niloy J Mitra, and Peter Wonka. Labels4free: Unsupervised segmentation using stylegan. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13970-13979, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.173, + 0.484, + 0.227 + ], + "angle": 0, + "content": "[2] Aishwarya Agarwal, Srikrishna Karanam, K J Joseph, Apoorv Saxena, Koustava Goswami, and Balaji Vasan Srinivasan. Astar: Test-time attention segregation and retention for text-to-image synthesis, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.23, + 0.483, + 0.284 + ], + "angle": 0, + "content": "[3] Aishwarya Agarwal, Srikrishna Karanam, and Balaji Vasan Srinivasan. Training-free color-style disentanglement for constrained text-to-image synthesis. arXiv preprint arXiv:2409.02429, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.287, + 0.484, + 0.342 + ], + "angle": 0, + "content": "[4] Yuval Alaluf, Daniel Garibi, Or Patashnik, Hadar Averbuch-Elor, and Daniel Cohen-Or. Cross-image attention for zero-shot appearance transfer. arXiv preprint arXiv:2311.03335, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.344, + 0.484, + 0.385 + ], + "angle": 0, + "content": "[5] Omer Bar-Tal, Lior Yariv, Yaron Lipman, and Tali Dekel. Multidiffusion: Fusing diffusion paths for controlled image generation. 2023. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.387, + 0.484, + 0.428 + ], + "angle": 0, + "content": "[6] Shariq Farooq Bhat, Niloy J. Mitra, and Peter Wonka. Loosecontrol: Lifting controlnet for generalized depth conditioning, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.431, + 0.484, + 0.485 + ], + "angle": 0, + "content": "[7] Anand Bhattachad, Daniel McKee, Derek Hoiem, and David Forsyth. Stylegan knows normal, depth, albedo, and more. Advances in Neural Information Processing Systems, 36, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.488, + 0.484, + 0.528 + ], + "angle": 0, + "content": "[8] Tim Brooks, Aleksander Holynski, and Alexei A. Efros. Instructpix2pix: Learning to follow image editing instructions. In CVPR, 2023. 7, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.531, + 0.484, + 0.586 + ], + "angle": 0, + "content": "[9] Hila Chefer, Yuval Alaluf, Yael Vinker, Lior Wolf, and Daniel Cohen-Or. Attend-and-excite: Attention-based semantic guidance for text-to-image diffusion models. ACM Transactions on Graphics (TOG), 42(4):1-10, 2023. 1, 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.588, + 0.484, + 0.643 + ], + "angle": 0, + "content": "[10] Minghao Chen, Iro Laina, and Andrea Vedaldi. Training-free layout control with cross-attention guidance. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 5343–5353, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.645, + 0.484, + 0.714 + ], + "angle": 0, + "content": "[11] Jaehoon Choi, Taekyung Kim, and Changick Kim. Self-ensembling with gan-based data augmentation for domain adaptation in semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6830–6840, 2019. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.716, + 0.484, + 0.771 + ], + "angle": 0, + "content": "[12] Subhabrata Choudhury, Iro Laina, Christian Rupprecht, and Andrea Vedaldi. Unsupervised part discovery from contrastive reconstruction. Advances in Neural Information Processing Systems, 34:28104-28118, 2021. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.773, + 0.484, + 0.828 + ], + "angle": 0, + "content": "[13] Edo Collins, Radhakrishna Achanta, and Sabine Susstrunk. Deep feature factorization for concept discovery. In Proceedings of the European Conference on Computer Vision (ECCV), pages 336-352, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.83, + 0.484, + 0.885 + ], + "angle": 0, + "content": "[14] Omer Dahary, Or Patashnik, Kfir Aberman, and Daniel Cohen-Or. Be yourself: Bounded attention for multi-subject text-to-image generation. arXiv preprint arXiv:2403.16990, 2(5), 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.887, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[15] Gilad Deutch, Rinon Gal, Daniel Garibi, Or Patashnik, and" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "Daniel Cohen-Or. Turboedit: Text-based image editing using few-step diffusion models, 2024. 7, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.122, + 0.905, + 0.162 + ], + "angle": 0, + "content": "[16] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.164, + 0.907, + 0.233 + ], + "angle": 0, + "content": "[17] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In *Forty-first International Conference on Machine Learning*, 2024. 7, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.234, + 0.907, + 0.287 + ], + "angle": 0, + "content": "[18] Songwei Ge, Taesung Park, Jun-Yan Zhu, and Jia-Bin Huang. Expressive text-to-image generation with rich text. In IEEE International Conference on Computer Vision (ICCV), 2023. 1, 2, 3, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.289, + 0.907, + 0.343 + ], + "angle": 0, + "content": "[19] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.344, + 0.907, + 0.413 + ], + "angle": 0, + "content": "[20] Jonathan Ho, William Chan, Chitwan Sahara, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.414, + 0.907, + 0.482 + ], + "angle": 0, + "content": "[21] Wei-Chih Hung, Varun Jampani, Sifei Liu, Pavlo Molchanov, Ming-Hsuan Yang, and Jan Kautz. Scops: Self-supervised co-part segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 869-878, 2019. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.484, + 0.907, + 0.552 + ], + "angle": 0, + "content": "[22] Minguk Kang, Jun-Yan Zhu, Richard Zhang, Jaesik Park, Eli Shechtman, Sylvain Paris, and Taesung Park. Scaling up gans for text-to-image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10124-10134, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.554, + 0.907, + 0.622 + ], + "angle": 0, + "content": "[23] Hyung-Kwon Ko, Gwanmo Park, Hyeon Jeon, Jaemin Jo, Juho Kim, and Jinwook Seo. Large-scale text-to-image generation models for visual artists' creative works. In Proceedings of the 28th International Conference on Intelligent User Interfaces, pages 919-933, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.623, + 0.907, + 0.678 + ], + "angle": 0, + "content": "[24] Svetlana Lazebnik, Cordelia Schmid, and Jean Ponce. Semi-local affine parts for object recognition. In *British Machine Vision Conference*, BMVC 2004, Kingston, UK, September 7-9, 2004. Proceedings, pages 1-10. BMVA Press, 2004. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.679, + 0.907, + 0.734 + ], + "angle": 0, + "content": "[25] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip2: Bootstrapping language-image pre-training with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.735, + 0.907, + 0.803 + ], + "angle": 0, + "content": "[26] Shanglin Li, Bohan Zeng, Yutang Feng, Sicheng Gao, Xiuhui Liu, Jiaming Liu, Lin Li, Xu Tang, Yao Hu, Jianzhuang Liu, et al. Zone: Zero-shot instruction-guided local editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6254-6263, 2024. 7, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.804, + 0.907, + 0.873 + ], + "angle": 0, + "content": "[27] Yuheng Li, Haotian Liu, Qingyang Wu, Fangzhou Mu, Jianwei Yang, Jianfeng Gao, Chunyuan Li, and Yong Jae Lee. Gligen: Open-set grounded text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22511-22521, 2023. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.874, + 0.907, + 0.901 + ], + "angle": 0, + "content": "[28] Luping Liu, Yi Ren, Zhijie Lin, and Zhou Zhao. Pseudo numerical methods for diffusion models on manifolds. In" + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13217" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.484, + 0.119 + ], + "angle": 0, + "content": "International Conference on Learning Representations, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.484, + 0.191 + ], + "angle": 0, + "content": "[29] Ziwei Liu, Ping Luo, Shi Qiu, Xiaogang Wang, and Xiaou Tang. Deepfashion: Powering robust clothes recognition and retrieval with rich annotations. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.193, + 0.484, + 0.247 + ], + "angle": 0, + "content": "[30] Wan-Duo Kurt Ma, JP Lewis, W Bastiaan Kleijn, and Thomas Leung. Directed diffusion: Direct control of object placement through attention guidance. arXiv preprint arXiv:2302.13153, 2023. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.25, + 0.484, + 0.305 + ], + "angle": 0, + "content": "[31] Subhransu Maji, Esa Rahtu, Juho Kannala, Matthew Blaschko, and Andrea Vedaldi. Fine-grained visual classification of aircraft. arXiv preprint arXiv:1306.5151, 2013. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.307, + 0.484, + 0.376 + ], + "angle": 0, + "content": "[32] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6038–6047, 2023. 6, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.378, + 0.484, + 0.419 + ], + "angle": 0, + "content": "[33] Kam Woh Ng, Xiatian Zhu, Yi-Zhe Song, and Tao Xiang. Partcraft: Crafting creative objects by parts. arXiv preprint arXiv:2407.04604, 2024. 3, 7, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.421, + 0.484, + 0.49 + ], + "angle": 0, + "content": "[34] Rishubh Parihar, Abhijnya Bhat, Abhipsa Basu, Saswat Mallick, Jogendra Nath Kundu, and R Venkatesh Babu. Balancing act: distribution-guided debiasing in diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6668-6678, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.492, + 0.484, + 0.547 + ], + "angle": 0, + "content": "[35] Rishubh Parihar, Harsh Gupta, Sachidanand VS, and R Venkatesh Babu. Text2place: Affordance-aware text guided human placement. In European Conference on Computer Vision, pages 57-77. Springer, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.549, + 0.484, + 0.618 + ], + "angle": 0, + "content": "[36] Rishubh Parihar, VS Sachidanand, Sabariswaran Mani, Tejan Karmali, and R Venkatesh Babu. Precisecontrol: Enhancing text-to-image diffusion models with fine-grained attribute control. In European Conference on Computer Vision, pages 469-487. Springer, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.62, + 0.484, + 0.688 + ], + "angle": 0, + "content": "[37] Or Patashnik, Daniel Garibi, Idan Azuri, Hadar Averbuch-Elor, and Daniel Cohen-Or. Localizing object-level shape variations with text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.69, + 0.484, + 0.758 + ], + "angle": 0, + "content": "[38] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.761, + 0.484, + 0.829 + ], + "angle": 0, + "content": "[39] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, et al. Sam 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[40] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.176 + ], + "angle": 0, + "content": "[41] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural information processing systems, 35:36479-36494, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.907, + 0.26 + ], + "angle": 0, + "content": "[42] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278-25294, 2022. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.262, + 0.907, + 0.303 + ], + "angle": 0, + "content": "[43] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2105.05233, 2021. 6, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.305, + 0.907, + 0.373 + ], + "angle": 0, + "content": "[44] Raphael Tang, Linqing Liu, Akshit Pandey, Zhiying Jiang, Gefei Yang, Karun Kumar, Pontus Stenetorp, Jimmy Lin, and Ferhan Ture. What the daam: Interpreting stable diffusion using cross attention. arXiv preprint arXiv:2210.04885, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.376, + 0.907, + 0.431 + ], + "angle": 0, + "content": "[45] Yoad Tewel, Omri Kaduri, Rinon Gal, Yoni Kasten, Lior Wolf, Gal Chechik, and Yuval Atzmon. Training-free consistent text-to-image generation. arXiv preprint arXiv:2402.03286, 2024.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.433, + 0.907, + 0.489 + ], + "angle": 0, + "content": "[46] Junjiao Tian, Lavisha Aggarwal, Andrea Colaco, Zsolt Kira, and Mar Gonzalez-Franco. Diffuse, attend, and segment: Unsupervised zero-shot segmentation using stable diffusion. arXiv preprint arXiv:2308.12469, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.49, + 0.907, + 0.559 + ], + "angle": 0, + "content": "[47] Narek Tumanyan, Michal Geyer, Shai Bagon, and Tali Dekel. Plug-and-play diffusion features for text-driven image-to-image translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1921-1930, 2023. 5, 7, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.561, + 0.907, + 0.658 + ], + "angle": 0, + "content": "[48] Andrea Vedaldi, Siddharth Mahendran, Stavros Tsogkas, Subhransu Maji, Ross Girshick, Juho Kannala, Esa Rahtu, Iasonas Kokkinos, Matthew B. Blaschko, David Weiss, Ben Taskar, Karen Simonyan, Naomi Saphra, and Sammy Mohamed. Understanding objects in detail with fine-grained attributes. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.66, + 0.906, + 0.702 + ], + "angle": 0, + "content": "[49] Andrey Voynov, Qinghao Chu, Daniel Cohen-Or, and Kfir Aberman. \\(p+\\): Extended textual conditioning in text-to-image generation. arXiv preprint arXiv:2303.09522, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.703, + 0.907, + 0.757 + ], + "angle": 0, + "content": "[50] C. Wah, S. Branson, P. Welinder, P. Perona, and S. Belongie. Caltech-ucsd birds-200-2011 dataset. Technical Report CNSTR-2011-001, California Institute of Technology, 2011. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.759, + 0.907, + 0.814 + ], + "angle": 0, + "content": "[51] Zirui Wang, Zhizhou Sha, Zheng Ding, Yilin Wang, and Zhuowen Tu. Tokencompose: Grounding diffusion with token-level supervision. arXiv preprint arXiv:2312.03626, 2023.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.817, + 0.906, + 0.872 + ], + "angle": 0, + "content": "[52] Jiarui Xu, Sifei Liu, Arash Vahdat, Wonmin Byeon, Xiaolong Wang, and Shalini De Mello. Open-Vocabulary Panoptic Segmentation with Text-to-Image Diffusion Models. arXiv preprint arXiv:2303.04803, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.907, + 0.903 + ], + "angle": 0, + "content": "[53] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13218" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.124, + 0.092, + 0.484, + 0.134 + ], + "angle": 0, + "content": "Yang, Burcu Karagol Ayan, et al. Scaling autoregressive models for content-rich text-to-image generation. arXiv preprint arXiv:2206.10789, 2(3):5, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.484, + 0.19 + ], + "angle": 0, + "content": "[54] Kai Zhang, Lingbo Mo, Wenhu Chen, Huan Sun, and Yu Su. Magicbrush: A manually annotated dataset for instruction-guided image editing. Advances in Neural Information Processing Systems, 36:31428-31449, 2023. 7, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.192, + 0.482, + 0.248 + ], + "angle": 0, + "content": "[55] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.25, + 0.482, + 0.317 + ], + "angle": 0, + "content": "[56] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 7" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.317 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.956 + ], + "angle": 0, + "content": "13219" + } + ] +] \ No newline at end of file diff --git a/2025/Composing Parts for Expressive Object Generation/edecee1f-cf3a-417c-a3d7-2ea937e0cfb0_origin.pdf b/2025/Composing Parts for Expressive Object Generation/edecee1f-cf3a-417c-a3d7-2ea937e0cfb0_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..24fec867f44aeb4573c2659dc6ecf888197ddc5b --- /dev/null +++ b/2025/Composing Parts for Expressive Object Generation/edecee1f-cf3a-417c-a3d7-2ea937e0cfb0_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:82dcc908c3306115ec6294c1b817c59b1be179f94e8e9d740bac7197253129d9 +size 3032204 diff --git a/2025/Composing Parts for Expressive Object Generation/full.md b/2025/Composing Parts for Expressive Object Generation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..92140f6b4074a92400123113ea6e4e5b7075b6b0 --- /dev/null +++ b/2025/Composing Parts for Expressive Object Generation/full.md @@ -0,0 +1,328 @@ +# Composing Parts for Expressive Object Generation + +Harsh Rangwani1 Aishwarya Agarwal1 Kuldeep Kulkarni1 +R. Venkatesh Babu2 Srikrishna Karanam1 + +$^{1}$ Adobe Research $^{2}$ Indian Institute of Science + +![](images/cd5432a7005939731e497150583b3fbe95cd4d73c160d0332a360f8501af1de3.jpg) +a full body portrait photo of a person wearing a dress with long hair and glasses PartComposer: hair, glasses, dress + +![](images/e13f311949598154339404d5dcacbcf7dc93972ffa2719c075d283bd73a8e611.jpg) +StableDiffusion + +![](images/7f99108ebc84183ebe8d3deb2488a76ac3a098da2737fe469299f73fb5a2c284.jpg) +Rich Text Gen + +![](images/18f0cda5ade8ba47f96b71f4301b491d6bc71c20e15a36443a15116cb2e47b61.jpg) +PartComposer (Ours) + +![](images/0aa271579bb01c6ababa8f61735064cd8927fc60625d7f9b97b3f346c69f5ac1.jpg) +Base Generation + +![](images/5ba46fa145d4726e1122e747ab86aa294a2b1863d4c1a61be9ea580832cd46d2.jpg) +StableDiffusion +A photo of a yellow warbler + +![](images/6c79eac0584f1dd146085c96e7f41970869375aad684081edae561f97669c963.jpg) +PartComposer (Ours) + +![](images/18ed7dca323cecb7b8369f55cc64530b182b28400d69c3bb952ee3198dbace26.jpg) + +![](images/4b5f3d7461775cb04de5390b6eca162e4399a251c32c0fc57178d5729b4e0200.jpg) +Rich Text Gen +a photo of a white swan, 8k, full hd +rtComposer: crown → a crown of a peacock + +![](images/08bf5885f58861b409b88bafcaf7286cbf6888ddbf86ad06ae8de1d4e33855de.jpg) +A young woman sits at a table in a beautiful, lush garden, painting by claude monet PartComposer: dress $\rightarrow$ a dress in Uhuie-style +Figure 1. Base Generation (left) Comparison with Methods for Generations with Parts Details (right). PartComposer allows the generation of object images with specified attributes (color, style etc.) of parts for the chosen object in the base text prompt. StableDiffusion and Rich-Text [18] methods with part details either ignore the part instructions or generate inconsistent objects. + +![](images/d3cb014fdd70930f1cf40349861e3b4002971da70967ffc7a4a9306771d27268.jpg) +PartComposer: beak $\rightarrow$ a hens beak + +# Abstract + +Image composition and generation are processes where the artists need control over various parts of the generated images. However, the current state-of-the-art generation models, like Stable Diffusion, cannot handle fine-grained part-level attributes in the text prompts. Specifically, when additional attribute details are added to the base text prompt, these text-to-image models either generate an image vastly different from the image generated from the base prompt or ignore the attribute details. To mitigate these issues, we introduce PartComposer, a training-free method that enables image generation based on fine-grained part-level attributes specified for objects in the base text prompt. This allows more control for artists and enables novel object compositions by combining distinctive object parts. PartComposer first localizes object parts by denoising the object region from a specific diffusion process. This enables each part token to be localized to the right region. After obtaining part masks, we run a localized diffusion process in each part region based on fine-grained part attributes and combine them to produce the final image. All stages of PartComposer are based on repurposing a pre-trained diffusion model, which + +enables it to generalize across domains. We demonstrate the effectiveness of part-level control provided by PartComposer through qualitative visual examples and quantitative comparisons with contemporary baselines. + +# 1. Introduction + +Image generation with large generative diffusion models like StableDiffusion [40], DALLE [38], etc., has become prevalent due to their superior quality and extensive world knowledge. These models are trained on large image-caption datasets and are trained to generate images based on a given text prompt (description). As image composition and creation are creative processes where the artists need control over various parts of the image being generated. However, adding additional details for controlling part appearance in the text prompt either changes the generated image entirely or ignores the part instructions [9] (Fig. 1). + +Various works aim to provide improved spatial controls to image generation, as they allow image generation conditioned on segmentation masks [5], edge maps [55], bounding boxes [10, 27] etc. Popular methods such as Control- + +Net [55], GLIGEN [27], etc. require specifying training of these conditional modules, which allows for the controllability of these large generative models. Further, there has been some development of training-free approaches [9, 10] which enable controlled generation of objects by modulating the internal cross-attention activations in the diffusion process. This demonstrates that these pre-trained generative models contain information about the spatial parts of the image, and modulating them effectively can lead to image compositions + +However, despite several attempts [6, 55], the semantic controllability of the image generation is still restricted to specifying details at the object level. The object level details can be restrictive, as often creative designers synthesize object parts (e.g. shirt, trouser etc.) and then compose them [23]. Further, variations in semantic parts are often very distinctive and are uniquely used to identify objects [12, 30]. For example, we usually identify species of birds by their unique beak. Due to this, semantic understanding and recognition of parts have been widely studied as a topic in computer vision [12, 13, 21, 31]. Hence, providing controllability for image synthesis at the object semantic part level can enable a large variety of image compositions [23]. Towards this goal, we introduce PartComposer, which provides users with an interface through which they can select the object in the scene and provide a semantic part-level description with fine-grained details for the object generation. + +In PartComposer, we develop a scheme to extract part-level localization masks from the Diffusion model. We introduce a parallel part diffusion process that generates masks for the object parts. The core idea of the approach is that by forcing the part diffusion model to specifically denoise only the object region in the image, it is possible to understand the locations of various parts in the object. After denoising just the object through the part model, the information inside the part diffusion model present in attention maps can be used to generate the masks for various object parts. In the following Part Generation stage, PartComposer utilizes the part masks and attributes of each part the user provides. We enable users to provide a highly expressive specification of parts by using a Rich-Text [18] interface, which allows the specifying attributes, such as style, color, etc., for each part. For the final image generation, taking inspiration from recent studies like Rich-Text Generation [18], Multi-Diffusion [5] etc., we compose the various object parts, by running parallel masked diffusion process for each part while combining them periodically into the image. This combination enables harmonious composition of parts, and also masking ensures only local modifications to the regions corresponding to each part of the object specified. In the PartComposer method, we only use the pre-trained StableDiffusion model, making it a generalizable and training-free approach. + +We extensively test the proposed PartComposer approach for the zero-shot object part segmentation, which is a chal + +lenging setup in computer vision. We evaluate the part segmentation approach on DeepFashion [29] and CUB200 [50] datasets, where our method significantly outperforms the baseline StableDiffusion (SD). Further, to evaluate the PartComposer image composition abilities, we also provide quantitative results along with a user preference study, where our method significantly outperforms the baselines in generating images consistent with the described parts (Fig. 1). We summarize the core contributions of our paper below: + +1. We introduce PartComposer, a training-free method that enables the generation of object images by using provided fine-grained details for the parts of the object. For example, while generating a bird image we can specify a detailed description of its beak. +2. In PartComposer, we introduce a novel Part-Diffusion process, which localizes and provides masks for parts of a base object generated by the Diffusion model (Fig. 3). To localize object parts, we introduce a novel segmentation scheme that uses the attention maps of the base diffusion and part diffusion process to obtain accurate masks for localized parts. +3. In PartComposer, after localization, we enable the generation of parts from the pre-trained diffusion model based on the Rich-Text description for the parts provided by the user. PartComposer then composes the image to harmoniously blend all object parts with the background by combining localized diffusion paths (Fig. 2). + +# 2. Related Works + +Text-to-Image Generative Models. The text-to-image models synthesize images by following a textual description provided as a prompt. These models have recently become mainstream due to their superior image generation quality and significant knowledge base. This is an outcome of the availability of large-scale image caption datasets [42] and highly parallelized GPU clusters. Almost all kinds of generative models, such as GANs [22], Autoregressive [41, 53], and Diffusion models [38, 40], have shown significant improvements in quality with training on these large image caption datasets. Among these models, the StableDiffusion (SD) [40] models, based on denoising diffusion in latent space are popular due to their open-source nature, which we also utilize for experiments in our work. + +Text-to-Image Models for Downstream Tasks. Generative models, in general, have been useful for various downstream tasks, particularly ones based on per-pixel prediction like Segmentation [1, 11], Depth Estimation [7], etc. As layers near the image generation output have features that capture the pixel-pixel relation [7]. With the large-scale text-to-image generative models, these models often perform very competitively [46, 52] to discriminative methods, on tasks like segmentation. However, one commonality among most of these segmentation methods is that they operate at the + +![](images/9e8bec10eaa019538a476686de36837aa7edd59fc6536a23d9a6b5097712088d.jpg) +Figure 2. PartComposer takes the input of a base prompt, the selected object o, and a Rich-Text description ( i.e., footnote, colors, etc.) of the parts. The Part Diffusion process generates the object masks for the specified parts. Then, PartComposer runs a parallel Region Diffusion process to generate attributes of the specified parts, guided by base generation's intermediate outputs. + +granularity of object or instance level. Text-to-image models here are at an advantage as the usual image captions describe the scene at an object level. In this work, we take a step further in exploring the object part-level knowledge of these text-to-image diffusion models. + +Part Discovery and Segmentation. Part Discovery and Attribution were an integral part of computer vision pipelines classically, as these approaches were robust to viewpoint variations [24, 31, 48]. In deep learning, the unsupervised (self-supervised) approaches for part discovery like SCOPS [21] and Unsup-Parts [12] became popular as they generalize to object parts across categories. In this work, we go one step ahead and operate in a zero-shot unsupervised part segmentation setting, generating part masks from T2I models. + +Controllable Image Generation. Masks, bounding boxes, edge maps, depth maps, etc., have been explored to control the generations of text-to-image diffusion models [6, 14, 27, 35, 36, 55] in addition to text. Further, various other approaches [2-4, 9, 34, 45, 49, 51] achieve control of image semantics through modulating diffusion models. Despite this, the text-to-image generation control at the object part level is under-explored; a recent work [33] tries to do it in a controlled supervised setup using part-masks. Contrary to that, in our work PartComposer we explore a generalized training-free zero-shot setting. + +# 3. Method + +In this section, we introduce PartComposer, our method to synthesize objects based on the description of parts of the objects. In PartComposer, we ask the user to specify a base prompt and the token for the object for which it wants to synthesize parts. Then, we provide the user a rich-text [18] editor (Fig. 2) to specify the parts and their description. PartComposer involves two diffusion steps, a) Part Localiza + +tion: In the early diffusion stage, we get a mask for the object we want to divide into parts. Then, we perform denoising in later stage from a U-Net condition on parts to fill the masked region of the object, during which it learns to denoise different object parts. Due to this, the attention maps for various parts highlight the correct part region, which we use to extract the part mask. The infilling process is the major contribution of the PartComposer method (Fig. 3). b) + +Part Generation: For generating parts, we combine region-specific diffusion processes for various parts by iteratively merging them inspired by MultiDiffusion [5], Rich-Text Generation [18] etc. However, till now, most of these works have combined the diffusion process for generating objects; for the first time, we have demonstrated its effectiveness in generating object parts. We provide an overview of the PartComposer pipeline in Fig. 2. + +# 3.1. Problem Setup + +We consider each span of tokens $\mathbf{p}_i$ as indicative of describing one part of the object, with attributes $\mathbf{a}_i$ describing its overall appearance. Our design choices are based on the rich-text generation [18], and we allow the user to specify the following type of attributes for part generation (Fig. 2): Part Description (i.e. footnote). It is an important attribute that specifies the part of an object; for example, for the part $(\mathbf{p}_i)$ 'crown' of a bird, we can specify the attribute $(\mathbf{a}_i^p = \text{'a crown of a peacock')$ . This helps specify novel part descriptions, which can lead to artistic novel compositions. + +Font Color. It helps us specify the exact RGB values for the color attribute $\mathbf{a}_i^c$ we want for the object part $\mathbf{p}_i$ . The exact value of RGB allows fine-grained control over the color of the desired part, whereby just specifying specific colors like 'brick red' leads to ignorance by Stable Diffusion [18]. + +Font Style. The font style is indicative of the specific artistic style $\mathbf{a}_i^s$ like 'of Claude Monet' and 'of Van Gogh' when + +![](images/1c5ae7a02e4d99bd276ccf4fef6f5746b69ba170adfdbf53fd9061871a3d1c37.jpg) +Figure 3. Localization. We obtain the mask $(\mathcal{M}_o)$ for object in middle of diffusion process. We then denoise object in masked region, using parts $\mathbf{p}_i$ conditioned U-Net. Due to part denoising, the attention maps of $\mathbf{p}_i$ can now localize parts as masks $\mathbf{M}_{\mathbf{p}_i}$ . + +synthesizing images of paintings. This instructs the model to generate a part following a specific style for the given prompt and blend it with other parts of the object (Sec. 5). + +Font Size. The font size controls the relative size of each part in a generation [19]. We use the $\mathbf{a}_i^w$ to denote size. + +# 3.2. Part Localization + +For part localization, we first run the base stable diffusion model for the given text prompt and extract the token map $\mathcal{M}_o$ for the object $\mathbf{o}$ specified by the user, by following the technique of clustering the self-attention maps [18, 37]. We obtain the mask using this clustering after denoising between the initial step $(T)$ until a threshold time step $T_{th}$ . After obtaining the binary object mask $\mathcal{M}_o$ using attention masks [18, 37], we run two parallel diffusion processes where one contains input from the base prompt and the other contains input from the part prompts $\hat{\mathbf{p}} = [\mathbf{p}_1,\mathbf{p}_2,\dots,\mathbf{p}_n]$ , a token $\mathbf{p}_i$ for each of the parts specified in the part text prompt. We denote the $\hat{\mathbf{b}}$ to denote the base text prompt and $\hat{\mathbf{p}}$ to denote the part prompt. For the time $t\leq T_{th}$ (by default, we use $T_{th}\approx T / 2$ ), we denoise the U-Net in the object region by denoising it with both the combination of the Part Prompt output and Base Prompt output. Due to this, the Part-Based U-Net gets the information regarding various parts in the object (Fig. 3). We now mathematically define the output noise $\epsilon_t$ for the diffusion process with Part-Diffusion below: + +$$ +\epsilon_ {t} = \alpha \mathcal {M} _ {o} \odot D (x _ {t}, \hat {\mathbf {p}}, t) + (1 - \alpha \mathcal {M} _ {o}) \odot D (x _ {t}, \hat {\mathbf {b}}, t) \tag {1} +$$ + +here, the $\alpha$ is the hyper-parameter controlling the strength of the part prompt diffusion output, and $D$ is the output of the pre-trained U-Net of a text-to-image diffusion model. The above denoising process is followed for t steps until the last step to produce a base image corresponding to the given prompt $\hat{b}$ . Keeping a high $T_{th}$ and low $\alpha$ , makes minimal + +changes in output as if the denoising diffusion was done with original prompt $\hat{b}$ (See Suppl. Sec.). With this part denoising of object, we obtain attention masks, from which we extract the localization information of part $\mathbf{p}_t$ . We now describe the process of obtaining the part masks. + +Token Maps for Parts. We first take the part tokens $\hat{\mathbf{p}}$ , which are a concatenation of the part names (i.e. 'beak crown wings'), which may not make a meaningful text prompt. Hence, we initialize text embeddings for all these tokens $\mathbf{p}_i$ by passing the meaningful text prompt having the following template: "A photo of $\mathbf{p}_i$ of a o" where $\mathbf{p}_i$ is the object part name and o is the object name. This serves two purposes: first, it makes the text embedding meaningful, and second, it introduces some invariance from the order of part specification in part prompt. These embeddings are then passed as text embeddings to the Part U-Net for denoising. After running the denoising process, we aggregate the self-attention maps across multi-heads and time steps (from 32 × 32 resolution) for both the base and part U-Net diffusion branches, taking inspiration from works that demonstrate that attention can localize objects [9, 18, 30, 44]. We then perform spectral clustering on these attention maps to form k segmentation maps $\hat{\mathbf{M}}$ ( $32 \times 32$ ), based on pixel-pixel similarity. To attribute these K-segments to the part specified by the user, we aggregate the cross attention of Part U-Net diffusion process. For each token $\mathbf{p}_j$ , we obtain the cross-attention score as follows: $\hat{\mathbf{m}}_j = \frac{c_j}{\sum_k c_k}$ , where $c_j$ is the cross attention score for each token. We proceed by aggregating the attention heads to obtain the average cross-attention scores and resizing them to $32 \times 32$ , obtaining $\hat{\mathbf{m}}$ . We remove the start of text ( $\langle \mathrm{sot} \rangle$ ) token for cross-attention and re-normalize it [9]. In other works [18, 37], as the tokens correspond to the objects being generated in the image, it's sure that token maps will be meaningful. However, this is not true for parts, as Part U-Net might not localize some + +parts. For determining if the part is localized, we propose to look at the max value of cross attention map spatially across the pixels; if we find that the following condition is met, the part is localized: + +$$ +L (j) = \mathbb {1} \left\{\max \left(\hat {\mathbf {m}} _ {j}\right) \geq (1 - \delta) \frac {1}{K} \right\}. \tag {2} +$$ + +Here $\delta$ is a hyperparameter, and $K$ is the number of parts. Is This condition is robust in finding the localized parts (See Sec. 5 for ablation). For the parts $\mathbf{p}_i$ , which are localized, we normalize the cross-attention map: + +$$ +\hat {\mathbf {m}} _ {j} = L (j) \frac {\hat {\mathbf {m}} _ {j} - \min (\hat {\mathbf {m}} _ {j})}{\max (\hat {\mathbf {m}} _ {j}) - \min (\hat {\mathbf {m}} _ {j})} + (1 - L (j)) \hat {\mathbf {m}} _ {j}. +$$ + +We follow a dot-product-based protocol to assign each K cluster in the self-attention masks to a part, unlike the average attention protocol in the previous works [18, 37]. We find that dot product of normalized cross attention scores $\hat{\mathbf{m}}_j$ for each token with self-attention masks $\hat{\mathbf{M}}_j$ works better, as the attention maps in Part U-Net are noisy. Still, they are often correct for the regions that are localized in one specific area only (Suppl. Fig. 14). Hence, the dot product protocol favors those maps that are only localized in some areas of the image and don't have high attention values across all parts of the object. After obtaining the dot product scores for each part, we assign $\hat{\mathbf{M}}_j$ to $\mathbf{p}_i$ with the highest scores. The mask for the part $\mathbf{p}_i$ is finally given as the following: + +$$ +\mathbf {M} _ {\mathbf {p} _ {i}} = \left\{\cup_ {j} \hat {\mathbf {M}} _ {j} \mid \operatorname {a r g m a x} _ {i} \hat {\mathbf {M}} _ {j} \cdot \hat {\mathbf {m}} _ {i} = i \text {a n d} \hat {\mathbf {M}} _ {j} \cdot \hat {\mathbf {m}} _ {i} \geq \epsilon , \right\} \tag {3} +$$ + +Here $\epsilon$ is the hyperparameter which controls the minimum similarity required between the attention mask and the part token. We combine the additional attention masks unassigned to any token, and name them as the background (other) token $\mathbf{M}_b$ . + +# 3.3. Part Generation + +We follow a similar protocol for the generation of part segments as in Rich-Text Generation. We tailor the rich text generation to compose the part regions of the object in the image in place of the original scene composition. We describe the part generation protocol below briefly and refer readers to Rich-Text Gen [18] for more details. For each part $\mathbf{p}_i$ we run a region diffusion process, which runs in parallel for all the parts. We then combine the region diffusion processes to obtain the final noise prediction $\epsilon_t$ as the masked $\mathbf{M}_{\mathbf{p}_i}$ sum of the denoiser outputs: + +$$ +\epsilon_ {t} = \sum_ {i} \mathbf {M} _ {\mathbf {p} _ {i}} \epsilon_ {t, \mathbf {p} _ {i}} = \sum_ {i} \mathbf {M} _ {\mathbf {p} _ {i}} \odot D (x _ {t}, f (\mathbf {p} _ {i}, \mathbf {a} _ {i}), t) \tag {4} +$$ + +where $D$ is the pre-trained U-Net model, and $f(\mathbf{p}_i,\mathbf{a}_i)$ is the text description of the part $\mathbf{p}_i$ constructed using the following + +![](images/00ab72b86e894bcbc9ceccaba55020732970aac576a0369b3d46c107e7da795c.jpg) +Figure 4. Segmentation masks for the parts that are localized by Part Diffusion for DeepFashion (above) and CUB200 (below). + +process using the part tokens $\mathbf{p}_i$ and attributes $\mathbf{a}_i$ . Initially the text $f(\mathbf{p}_i, \mathbf{a}_i) = \mathbf{p}_i$ , is set to part token itself. In case the part description (i.e. footnote) is available we set the $f(\mathbf{p}_i, \mathbf{a}_i) = \mathbf{a}_i^p$ , further if the style attribute is available we do $f(\mathbf{p}_i, \mathbf{a}_i) = f(\mathbf{p}_i, \mathbf{a}_i) + \text{in style of} + \mathbf{a}_i^s$ . In case the color attribute is also specified, the nearest named color $\hat{\mathbf{a}}_i^c$ (e.g. red) for the specific RBG color $\mathbf{a}_i^c$ is found, and $f(\mathbf{p}_i, \mathbf{a}_i) = \hat{\mathbf{a}}_i^c + ' ' + f(\mathbf{p}_i, \mathbf{a}_i)$ . The string $f(\mathbf{p}_i, \mathbf{a}_i)$ is the text input for the Diffusion to generate the part $\mathbf{p}_i$ . We use the base prompt $\hat{\mathbf{b}}$ as $f(\mathbf{p}_i, \mathbf{a}_i)$ for the background masked region $\mathbf{M}_b$ . Combining different diffusion outputs at every time $t$ helps generate a harmonious image after blending the defined parts of the object. + +Following Rich-Text [18], we also utilize the gradient guidance [16, 20] by taking the gradient of MSE loss between the estimated original image and the color value $\mathbf{a}_i^c$ specified by the user. The gradient guidance helps generate the exact RGB color for part $\mathbf{p}_i$ , which is impossible with just text guidance [18]. Further, as we use the base text framework of Rich-Text, we also can specify font size attribute to $\mathbf{a}_i^f$ to control the relative importance of each part in the object image. + +Preservation of Other Parts. As we only intend to modify the object o from the original prompt, we also use the Self-Injection techniques from Plug and Play [47] to maintain the overall structure of the background from base prompt generation. Further, to ensure that our diffusion trajectory follows the same path as the base, the background region is also blended with base noise generation outputs. + +$$ +x _ {t} = \mathbf {M} _ {b} \odot x _ {t} ^ {\text {b a s e}} + (1 - \mathbf {M} _ {b}) \odot x _ {t} \tag {5} +$$ + +In our case, the attention maps out of the object $\mathbf{Mask}\mathcal{M}_o$ and all the object parts not assigned to any part token comprise the background $\mathbf{M}_b$ , and we start this blending process at $t = T_{blend}$ . We find this to be very useful in preserving the structure of the other parts of the image and just generating the described parts in the localized region. We provide an overview of the complete process in Fig. 2. + +# 4. Experimental Analysis + +# 4.1. Evaluation of Part Localization + +We first evaluate the part localization module, which is based on the novel idea of denoising only the object region with the part-based diffusion outputs. + +Table 1. Comparison to Prior Works for the unsupervised part segmentation task. We follow Unsup-Part [12] for evaluation protocols and baseline results for $(\mathbf{K} = 4)$ parts. We report clustering-based NMI and ARI metrics, which are higher for better segmentation outputs. + +
DeepFashion DatasetCUB200 Dataset
MethodFG-NMIFG-ARINMIARIFG-NMIFG-ARINMIARI
Unsupervised Learning
SCOPS [21]30.727.656.681.439.117.924.47.1
Unsup-Parts [12]44.846.668.190.646.021.043.519.6
DFF [11]----32.414.325.912.4
Unsupervised Zero-Shot
Rich-Text [18]16.05.248.358.73.10.33.10.3
StableDiffusion12.03.540.670.98.00.63.30.6
PartComposer(Ours)24.718.048.073.420.59.218.57.7
+ +Implementation Details. We use the StableDiffusion (SD) version 2.1 for our experimentation purposes. We use the DDIM Scheduler [43] with 50 steps to generate results for SD2.1 to evaluate the part localization process. As ground truth is not available for the part masks of the generated images, we use test sets of the commonly used DeepFashion [29] and CUB-200 [50] datasets for evaluation. These datasets are the standard datasets used for the evaluation of the unsupervised part segmentation approaches. The DeepFashion dataset contains images along with their part segmentation masks, divided into 14 categories of labels. The CUB200 dataset contains the key point annotations for the 14 key points specified for the bird categories. We provide further details in Suppl. Section. + +Baselines and Problem Setting. As we operate in the setting of Zero-shot (i.e. no training) Unsupervised Part Segmentation, there are no previous works that report results in such a challenging setting. Hence, we provide results for the unsupervised learning approaches to facilitate comparison. We provide results for the SCOPS [21], which utilizes the internal features of the VGG model to train a model based on self-supervised loss functions to predict parts robustly across categories. The other stronger self-supervised baseline is Unsup-Parts [12], which uses contrastive loss functions to train a network based on equivariance and other vision properties to cluster the object regions into semantic parts. In addition to this, we also report results for the DFF [11] as they also operate in the same setting. We want to highlight that these unsupervised approaches require either training a neural network or performing clustering on the complete training data to segment parts. In contrast, our approach is training free and operates in a zero-shot fashion. Hence, the performance of zero-shot approaches is not directly comparable to unsupervised methods. + +Zero Shot Unsupervised Part Segmentation. As there is no benchmark to evaluate the part localization for generated images, we use the existing dataset of DeepFashion and CUB200 to obtain our results. To obtain segmentation for + +each image, we first invert the image into the diffusion latent space using Null-Text Inversion [32] method (see Sec. 5 for ablation). We use the BLIP-V2 [25] captioner provided in Diffusers to obtain approximate text prompts for inversion. After providing the desired image and prompt, the Null-Text inversion provides us with inverted latent and unconditioned time embeddings to reconstruct the image of the dataset. We then construct a StableDiffusion baseline in which, in addition to the prompts, we append the list of part tokens $\mathbf{p}$ to the prompt. We then use the segmentation algorithm described above in Sec. 3.2 to extract token maps corresponding to the part tokens $\mathbf{p}_i$ . We also evaluate a Rich-Text [18] baseline segmentation algorithm for the same. For the proposed part-denoising approach based on parallel diffusion in PartComposer, we first get the inverted latents and embeddings. Then, we generate a base diffusion process to reconstruct the image and use Part Diffusion to fill the parts of the image. To very fairly compare the StableDiffusion performance with part-denoising, we keep all things the same except the part diffusion process to get masks. Further across these baselines, we use low classifier guidance to ensure proper reconstruction of the dataset images (See Suppl. for details). + +Evaluation Protocol. We use the standard experimental protocol as used by earlier works [12] to facilitate the right point of comparison. We want to point out that in our work, the part name (e.g., beak, etc.) is associated with the localized mask, but the unsupervised approaches produce $(\mathrm{K} = 4)$ parts without any part names. Hence, to make a fair comparison, we create 4 clusters of parts based on their locality, and finally generate segmentation masks with a maximum of 4 clusters. We provide the exact mapping of the part names to cluster labels in the Suppl. Section. We compare the specifically designed metrics [12] of NMI (Normalized Mutual Information) and ARI (Adjusted Rand Index) of the predicted cluster labels with part masks in the case of DeepFashion and key points in case of the CUB dataset. We report the metrics along with their foreground variants (FG-NMI and FG-ARI) for a holistic comparison of performance. + +![](images/8ce660be2b47af28118151ccd3f21a9815fa29a6634aa78c3e5c9f3e78410dfa.jpg) + +![](images/190c79cd63844621e26fc29025dab3fbdd375c615e10579360a13c7945077746.jpg) +Figure 5. Qualitative Comparison for PartComposer. (above) We provide base prompt 'a photo of a flamingo' and part prompt as 'beak - a pelicans beak'. (below) We generate a photo of a man, with part prompt as 'black jacket and blue jeans'. We find that other baselines either ignore instruction or change the entire composition. On the contrary, PartComposer can correctly localize parts and generate details. + +![](images/b76912a9f8ff072d94f0c8e42f6de4f760ff818a4488135154424211f29d200a.jpg) + +PartCraft (Ours) + +![](images/b3d287cc8aeb7867fe07c9a4427a8e66c9ea52d84d9ddd1742bba5178393a21e.jpg) +StableDiffusion + +![](images/a4a04e6b8fe4eb13aed038c7e1f22c36d4c5ef6eb396580b499242f2800abe0d.jpg) +InstructPix2Pix + +![](images/0af86d5f4d4db4124a76497fa148205cae060aebd7e979085eb352b94bd8d3de.jpg) +Rich Text Gen + +PartCraft (Ours) + +![](images/b1e8a37755145c7331dfacbe8bcf9ffa4c66b0305a1b7e56c5fa86b8ff174ab3.jpg) +StableDiffusion + +![](images/e78ad06de85adcbee3536b5f973a50d316cf659e4f6044ea31a982e8d677b149.jpg) +InstructPix2Pix + +![](images/b61359ccfa2c89ebe844abc616efb258d0b87056856bd038e04cfd9a05089aa0.jpg) +Rich Text Gen + +Results. We tabulate the results for all the baselines in Table 1. Our approach, PartComposer, significantly improves over the baselines by $\geq 5$ points in NMI and ARI across most metrics in the Unsupervised Zero-Shot setting. Further, in some cases our methods, NMI and ARI, are near the unsupervised learning approaches. This demonstrates that non-trivial part masks can be generated by harnessing the power of the pre-trained diffusion models, and they can serve as a good initial prior for learning part segmentation approaches by using them as base pre-trained models. We provide qualitative results for PartComposer, where we observe that PartComposer can associate the right part with the correct label (Fig 4). On the contrary, in the StableDiffusion baseline, the parts often get assigned to the wrong part tokens, which is the cause of degraded performance (Suppl. Fig. 16). We also observe that in the case of our algorithm, if the part gets localized based on the max condition defined in the segmentation algorithm, the part mask found is usually in the right region (See Fig. 4 and Table 2). This demonstrates the effectiveness of our part segmentation procedure, which often sides with not localizing objects rather than performing arbitrary assignments. + +# 4.2. PartComposer Image Generation Evaluation + +We now evaluate the object composition ability of the proposed approach, PartComposer, compared to the baselines. In this section, we use the base generation model as a StableDiffusion (SD) 1.5 model to do a fair comparison with the Rich-Text baseline. We used the same generation setting of a PNDM [28] scheduler with 41 steps and suggested classifier guidance 8.5. To perform a fair comparison, we keep all the parameters same as that of the Rich-Text [18] baseline. + +Baselines. For the task of generating the object image based on part-level details of the specified object, we use the strong baselines of Rich-Text Generation and InstructPix2Pix [8]. The Rich-Text Generation method generates the base image itself, whereas the InstructPix2Pix method requires us to + +generate the base image. We also evaluate the standard StableDiffusion baseline in which we add all the part details in the text prompt to generate the desired image. We compare all these baselines with our proposed method, PartComposer, while ensuring that the base parameters and seed are the same for the base image generation. A recent work [33] regarding part generation operates in a fully supervised setup using part masks and cannot be used to create novel parts based on text instruction, making it incomparable. We defer specific details for the baselines to the Suppl. Section. + +Visual Comparison. We provide a visual comparison for the baselines for the a) bird image distinctive part generation and b) human image part generations. Across both cases (Fig. 5), our proposed method, PartComposer, only modifies the desired beak region and replaces it with the iconic pelican beak. The StableDiffusion baseline modifies the image completely while ignoring instructions (Supp. Fig. 12). In the Rich-Text generation, the full region corresponding to the bird gets modified instead of only the specified prompt in base generation. For the editing-based InstructPix2Pix method, we observe that it modifies the base image at a global level and cannot localize modification to the desired part region (Fig. 5). We further compare and find that recent models like SD3.5, Inpainting [17] and recent SotA editing methods [15, 26, 47, 54] can still not follow exact part instructions as seen in Supp. Fig. 8, 9 and 10. In contrast, observe that PartComposer-generated images are novel aesthetic compositions that follow the part-level instructions. + +Quantitative Evaluation. As the part-specific generation is fine-grained, we perform the both automatic and user-study evaluations. We performed a user study by inviting 50 participants (28 responded with a full survey). We evaluate the model by asking questions and evaluating metrics on three orthogonal aspects i.e., a) Localization of part generation in comparison to the base image (through LPIPS [56]), b) Text-Consistency of generated parts by measuring CLIP + +Table 2. Ablation Analysis of PartComposer for Part Localization, showing performance improvement with each component. + +
PartComposer (Ablations)
MethodFG-NMIFG-ARI
PartComposer35.411.0
w/o Null-Text Inversion23.15.2
w/o Max Localization21.32.8
w/o Dot Product Localization23.75.0
w/o Independent Text31.28.9
PartComposer (Clustering)
PartComposer (K = 9)35.411.0
PartComposer (K = 4)20.42.8
PartComposer (K = 14)35.210.3
+ +![](images/c7feb96f76df24e61954e428d9fcccf9414cd4b7cc0d2dd95e06b98b5ef2a6dc.jpg) +Figure 6. PartComposer generalizes over domains as shown by Claude Monet's painting style (above), where we specify the dress part of women to follow styles like Van Gogh and Ukiyo-e. We perform similar modifications to Woodie's image. + +![](images/9ba083cb1b50e0235a03d773b19f8abc2c86af8756b184a6179b5e24e63ac170.jpg) + +![](images/98a4c4e120dbae11deb0e00f5cbab1350b8b9e297287218afe3a75b67614aaf1.jpg) + +![](images/2a89e87e5a288a024d70157572874c1490e479d558bc4da45bdc22ac58433b55.jpg) + +similarity of the part detailed text prompt with the generated image and c) Aesthetic Quality of the generated image through LAION-5B [42] CLIP Aesthetic scorer. In total, we collect about 3.5k opinions. In Fig. 7, we summarize the results in which we observe that PartComposer is significantly preferred over the other baselines in terms of localized and consistent Part Generation. The automated evaluation results are provided in the table below, which also follow a similar trend as in the user study, demonstrating the effectiveness of PartComposer in a localized part generation while ensuring the aesthetics score is similar to that of the base model. + +# 5. Analysis and Discussion + +Part Diffusion Based Localization. We ablate the components we have introduced in the PartDiffusion process of PartComposer (Sec. 3.2). We provide an analysis of the effect of using a) null-text inversion, b) max-based localization c) usage of dot product-based protocol in the part assignment, and d) independent text embeddings. We have used a subset of CUB-200 images to perform all evaluations, which is kept fixed across ablations. We tabulate the ablations in + +![](images/850689bb6bce080f702d0358c762bd3ce0aca0b801107d68eb43f6972bb75596.jpg) +Figure 7. User Study and Quantitative Results for part-based image generation baselines. + +
PartComposerStableDiffusionInstructPix2PixRich-Text
LPIPS0.1680.4670.1890.243
CLIP0.2010.1830.1930.187
Aesthetic5.665.685.635.65
+ +Table 2. We observe that all the components introduced in PartComposer contribute significantly to the performance. + +Generalization of PartComposer. In Fig. 6, above a painting in Claude Monet style generated by the base StableDiffusion method. We then use PartComposer to specify the dress part of the women's token, to dark green color in Monet Style, orange color dress in Van Gogh Style (middle), and green color dress in Ukiyo-e style. In Fig. 6, we provide results for the 'Woody' character from Toy Story. These results show that PartComposer can generate shirt, pants, and dress color variations, leading to aesthetic image combinations in synthetic and natural (Supp. Fig. 13) domains. This zero-shot generalization across domains demonstrates the creative activities that can be enabled with PartComposer. + +# 6. Conclusion + +In this work we introduce PartComposer, a method to generate object images with fine-grained attribute details specified at the part level, using an expressive Rich-Text interface. The PartComposer method introduces a novel part diffusion process, which is responsible for denoising objects using the part features, and then utilizes a region-specific diffusion process to generate part details and compose the final image. PartComposer serves as initial work enabling rich-text-based training-free part-level control for SD models. + +Limitations. We find that part generation is bottlenecked with the part understanding. If the part is localized correctly, the text-to-image model can generate specified details in PartComposer. Hence, improving the part-level localization of these models is a good direction for future works. + +# References + +[1] Rameen Abdal, Peihao Zhu, Niloy J Mitra, and Peter Wonka. Labels4free: Unsupervised segmentation using stylegan. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13970-13979, 2021. 2 +[2] Aishwarya Agarwal, Srikrishna Karanam, K J Joseph, Apoorv Saxena, Koustava Goswami, and Balaji Vasan Srinivasan. Astar: Test-time attention segregation and retention for text-to-image synthesis, 2023. 3 +[3] Aishwarya Agarwal, Srikrishna Karanam, and Balaji Vasan Srinivasan. Training-free color-style disentanglement for constrained text-to-image synthesis. arXiv preprint arXiv:2409.02429, 2024. +[4] Yuval Alaluf, Daniel Garibi, Or Patashnik, Hadar Averbuch-Elor, and Daniel Cohen-Or. Cross-image attention for zero-shot appearance transfer. arXiv preprint arXiv:2311.03335, 2023. 3 +[5] Omer Bar-Tal, Lior Yariv, Yaron Lipman, and Tali Dekel. Multidiffusion: Fusing diffusion paths for controlled image generation. 2023. 1, 2, 3 +[6] Shariq Farooq Bhat, Niloy J. Mitra, and Peter Wonka. Loosecontrol: Lifting controlnet for generalized depth conditioning, 2023. 2, 3 +[7] Anand Bhattachad, Daniel McKee, Derek Hoiem, and David Forsyth. Stylegan knows normal, depth, albedo, and more. Advances in Neural Information Processing Systems, 36, 2024. 2 +[8] Tim Brooks, Aleksander Holynski, and Alexei A. Efros. Instructpix2pix: Learning to follow image editing instructions. In CVPR, 2023. 7, 6 +[9] Hila Chefer, Yuval Alaluf, Yael Vinker, Lior Wolf, and Daniel Cohen-Or. Attend-and-excite: Attention-based semantic guidance for text-to-image diffusion models. ACM Transactions on Graphics (TOG), 42(4):1-10, 2023. 1, 2, 3, 4 +[10] Minghao Chen, Iro Laina, and Andrea Vedaldi. Training-free layout control with cross-attention guidance. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 5343–5353, 2024. 1, 2 +[11] Jaehoon Choi, Taekyung Kim, and Changick Kim. Self-ensembling with gan-based data augmentation for domain adaptation in semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6830–6840, 2019. 2, 6 +[12] Subhabrata Choudhury, Iro Laina, Christian Rupprecht, and Andrea Vedaldi. Unsupervised part discovery from contrastive reconstruction. Advances in Neural Information Processing Systems, 34:28104-28118, 2021. 2, 3, 6 +[13] Edo Collins, Radhakrishna Achanta, and Sabine Susstrunk. Deep feature factorization for concept discovery. In Proceedings of the European Conference on Computer Vision (ECCV), pages 336-352, 2018. 2 +[14] Omer Dahary, Or Patashnik, Kfir Aberman, and Daniel Cohen-Or. Be yourself: Bounded attention for multi-subject text-to-image generation. arXiv preprint arXiv:2403.16990, 2(5), 2024. 3 +[15] Gilad Deutch, Rinon Gal, Daniel Garibi, Or Patashnik, and + +Daniel Cohen-Or. Turboedit: Text-based image editing using few-step diffusion models, 2024. 7, 1 +[16] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 5 +[17] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In *Forty-first International Conference on Machine Learning*, 2024. 7, 1 +[18] Songwei Ge, Taesung Park, Jun-Yan Zhu, and Jia-Bin Huang. Expressive text-to-image generation with rich text. In IEEE International Conference on Computer Vision (ICCV), 2023. 1, 2, 3, 4, 5, 6, 7, 8 +[19] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 4 +[20] Jonathan Ho, William Chan, Chitwan Sahara, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 5 +[21] Wei-Chih Hung, Varun Jampani, Sifei Liu, Pavlo Molchanov, Ming-Hsuan Yang, and Jan Kautz. Scops: Self-supervised co-part segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 869-878, 2019. 2, 3, 6 +[22] Minguk Kang, Jun-Yan Zhu, Richard Zhang, Jaesik Park, Eli Shechtman, Sylvain Paris, and Taesung Park. Scaling up gans for text-to-image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10124-10134, 2023. 2 +[23] Hyung-Kwon Ko, Gwanmo Park, Hyeon Jeon, Jaemin Jo, Juho Kim, and Jinwook Seo. Large-scale text-to-image generation models for visual artists' creative works. In Proceedings of the 28th International Conference on Intelligent User Interfaces, pages 919-933, 2023. 2 +[24] Svetlana Lazebnik, Cordelia Schmid, and Jean Ponce. Semi-local affine parts for object recognition. In *British Machine Vision Conference*, BMVC 2004, Kingston, UK, September 7-9, 2004. Proceedings, pages 1-10. BMVA Press, 2004. 3 +[25] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip2: Bootstrapping language-image pre-training with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597, 2023. 6 +[26] Shanglin Li, Bohan Zeng, Yutang Feng, Sicheng Gao, Xiuhui Liu, Jiaming Liu, Lin Li, Xu Tang, Yao Hu, Jianzhuang Liu, et al. Zone: Zero-shot instruction-guided local editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6254-6263, 2024. 7, 1 +[27] Yuheng Li, Haotian Liu, Qingyang Wu, Fangzhou Mu, Jianwei Yang, Jianfeng Gao, Chunyuan Li, and Yong Jae Lee. Gligen: Open-set grounded text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22511-22521, 2023. 1, 2, 3 +[28] Luping Liu, Yi Ren, Zhijie Lin, and Zhou Zhao. Pseudo numerical methods for diffusion models on manifolds. In + +International Conference on Learning Representations, 2022. 7 +[29] Ziwei Liu, Ping Luo, Shi Qiu, Xiaogang Wang, and Xiaou Tang. Deepfashion: Powering robust clothes recognition and retrieval with rich annotations. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 2, 6 +[30] Wan-Duo Kurt Ma, JP Lewis, W Bastiaan Kleijn, and Thomas Leung. Directed diffusion: Direct control of object placement through attention guidance. arXiv preprint arXiv:2302.13153, 2023. 2, 4 +[31] Subhransu Maji, Esa Rahtu, Juho Kannala, Matthew Blaschko, and Andrea Vedaldi. Fine-grained visual classification of aircraft. arXiv preprint arXiv:1306.5151, 2013. 2, 3 +[32] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6038–6047, 2023. 6, 5 +[33] Kam Woh Ng, Xiatian Zhu, Yi-Zhe Song, and Tao Xiang. Partcraft: Crafting creative objects by parts. arXiv preprint arXiv:2407.04604, 2024. 3, 7, 6 +[34] Rishubh Parihar, Abhijnya Bhat, Abhipsa Basu, Saswat Mallick, Jogendra Nath Kundu, and R Venkatesh Babu. Balancing act: distribution-guided debiasing in diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6668-6678, 2024. 3 +[35] Rishubh Parihar, Harsh Gupta, Sachidanand VS, and R Venkatesh Babu. Text2place: Affordance-aware text guided human placement. In European Conference on Computer Vision, pages 57-77. Springer, 2024. 3 +[36] Rishubh Parihar, VS Sachidanand, Sabariswaran Mani, Tejan Karmali, and R Venkatesh Babu. Precisecontrol: Enhancing text-to-image diffusion models with fine-grained attribute control. In European Conference on Computer Vision, pages 469-487. Springer, 2024. 3 +[37] Or Patashnik, Daniel Garibi, Idan Azuri, Hadar Averbuch-Elor, and Daniel Cohen-Or. Localizing object-level shape variations with text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 4, 5 +[38] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. 1, 2 +[39] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, et al. Sam 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714, 2024. 1 +[40] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 2 + +[41] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural information processing systems, 35:36479-36494, 2022. 2 +[42] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278-25294, 2022. 2, 8 +[43] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2105.05233, 2021. 6, 5 +[44] Raphael Tang, Linqing Liu, Akshit Pandey, Zhiying Jiang, Gefei Yang, Karun Kumar, Pontus Stenetorp, Jimmy Lin, and Ferhan Ture. What the daam: Interpreting stable diffusion using cross attention. arXiv preprint arXiv:2210.04885, 2022. 4 +[45] Yoad Tewel, Omri Kaduri, Rinon Gal, Yoni Kasten, Lior Wolf, Gal Chechik, and Yuval Atzmon. Training-free consistent text-to-image generation. arXiv preprint arXiv:2402.03286, 2024.3 +[46] Junjiao Tian, Lavisha Aggarwal, Andrea Colaco, Zsolt Kira, and Mar Gonzalez-Franco. Diffuse, attend, and segment: Unsupervised zero-shot segmentation using stable diffusion. arXiv preprint arXiv:2308.12469, 2023. 2 +[47] Narek Tumanyan, Michal Geyer, Shai Bagon, and Tali Dekel. Plug-and-play diffusion features for text-driven image-to-image translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1921-1930, 2023. 5, 7, 1 +[48] Andrea Vedaldi, Siddharth Mahendran, Stavros Tsogkas, Subhransu Maji, Ross Girshick, Juho Kannala, Esa Rahtu, Iasonas Kokkinos, Matthew B. Blaschko, David Weiss, Ben Taskar, Karen Simonyan, Naomi Saphra, and Sammy Mohamed. Understanding objects in detail with fine-grained attributes. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2014. 3 +[49] Andrey Voynov, Qinghao Chu, Daniel Cohen-Or, and Kfir Aberman. $p+$ : Extended textual conditioning in text-to-image generation. arXiv preprint arXiv:2303.09522, 2023. 3 +[50] C. Wah, S. Branson, P. Welinder, P. Perona, and S. Belongie. Caltech-ucsd birds-200-2011 dataset. Technical Report CNSTR-2011-001, California Institute of Technology, 2011. 2, 6 +[51] Zirui Wang, Zhizhou Sha, Zheng Ding, Yilin Wang, and Zhuowen Tu. Tokencompose: Grounding diffusion with token-level supervision. arXiv preprint arXiv:2312.03626, 2023.3 +[52] Jiarui Xu, Sifei Liu, Arash Vahdat, Wonmin Byeon, Xiaolong Wang, and Shalini De Mello. Open-Vocabulary Panoptic Segmentation with Text-to-Image Diffusion Models. arXiv preprint arXiv:2303.04803, 2023. 2 +[53] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei + +Yang, Burcu Karagol Ayan, et al. Scaling autoregressive models for content-rich text-to-image generation. arXiv preprint arXiv:2206.10789, 2(3):5, 2022. 2 +[54] Kai Zhang, Lingbo Mo, Wenhu Chen, Huan Sun, and Yu Su. Magicbrush: A manually annotated dataset for instruction-guided image editing. Advances in Neural Information Processing Systems, 36:31428-31449, 2023. 7, 1 +[55] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 1, 2, 3 +[56] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 7 \ No newline at end of file diff --git a/2025/Composing Parts for Expressive Object Generation/images.zip b/2025/Composing Parts for Expressive Object Generation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..bf0d11175934901febfb2163f57468636f9b72c6 --- /dev/null +++ b/2025/Composing Parts for Expressive Object Generation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f7e1e4eb72e66ffa3b00738c293b2fd95b0effd75a487032b482159d9c20db32 +size 573166 diff --git a/2025/Composing Parts for Expressive Object Generation/layout.json b/2025/Composing Parts for Expressive Object Generation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..419769b574a8a5aacb1de8d4fa2e4b695effe33d --- /dev/null +++ b/2025/Composing Parts for Expressive Object Generation/layout.json @@ -0,0 +1,9391 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 146, + 103, + 464, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 103, + 464, + 121 + ], + "spans": [ + { + "bbox": [ + 146, + 103, + 464, + 121 + ], + "type": "text", + "content": "Composing Parts for Expressive Object Generation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 151, + 142, + 457, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 142, + 457, + 171 + ], + "spans": [ + { + "bbox": [ + 151, + 142, + 457, + 171 + ], + "type": "text", + "content": "Harsh Rangwani1 Aishwarya Agarwal1 Kuldeep Kulkarni1 \nR. Venkatesh Babu2 Srikrishna Karanam1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 187, + 184, + 421, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 187, + 184, + 421, + 198 + ], + "spans": [ + { + "bbox": [ + 187, + 184, + 421, + 198 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 187, + 184, + 421, + 198 + ], + "type": "text", + "content": "Adobe Research " + }, + { + "bbox": [ + 187, + 184, + 421, + 198 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 187, + 184, + 421, + 198 + ], + "type": "text", + "content": " Indian Institute of Science" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 60, + 222, + 121, + 288 + ], + "blocks": [ + { + "bbox": [ + 60, + 222, + 121, + 288 + ], + "lines": [ + { + "bbox": [ + 60, + 222, + 121, + 288 + ], + "spans": [ + { + "bbox": [ + 60, + 222, + 121, + 288 + ], + "type": "image", + "image_path": "cd5432a7005939731e497150583b3fbe95cd4d73c160d0332a360f8501af1de3.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 66, + 289, + 299, + 302 + ], + "lines": [ + { + "bbox": [ + 66, + 289, + 299, + 302 + ], + "spans": [ + { + "bbox": [ + 66, + 289, + 299, + 302 + ], + "type": "text", + "content": "a full body portrait photo of a person wearing a dress with long hair and glasses PartComposer: hair, glasses, dress" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 125, + 228, + 182, + 288 + ], + "blocks": [ + { + "bbox": [ + 124, + 222, + 171, + 228 + ], + "lines": [ + { + "bbox": [ + 124, + 222, + 171, + 228 + ], + "spans": [ + { + "bbox": [ + 124, + 222, + 171, + 228 + ], + "type": "text", + "content": "StableDiffusion" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 125, + 228, + 182, + 288 + ], + "lines": [ + { + "bbox": [ + 125, + 228, + 182, + 288 + ], + "spans": [ + { + "bbox": [ + 125, + 228, + 182, + 288 + ], + "type": "image", + "image_path": "e13f311949598154339404d5dcacbcf7dc93972ffa2719c075d283bd73a8e611.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 183, + 228, + 233, + 288 + ], + "blocks": [ + { + "bbox": [ + 199, + 222, + 228, + 228 + ], + "lines": [ + { + "bbox": [ + 199, + 222, + 228, + 228 + ], + "spans": [ + { + "bbox": [ + 199, + 222, + 228, + 228 + ], + "type": "text", + "content": "Rich Text Gen" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 183, + 228, + 233, + 288 + ], + "lines": [ + { + "bbox": [ + 183, + 228, + 233, + 288 + ], + "spans": [ + { + "bbox": [ + 183, + 228, + 233, + 288 + ], + "type": "image", + "image_path": "7f99108ebc84183ebe8d3deb2488a76ac3a098da2737fe469299f73fb5a2c284.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 232, + 228, + 303, + 288 + ], + "blocks": [ + { + "bbox": [ + 251, + 222, + 299, + 228 + ], + "lines": [ + { + "bbox": [ + 251, + 222, + 299, + 228 + ], + "spans": [ + { + "bbox": [ + 251, + 222, + 299, + 228 + ], + "type": "text", + "content": "PartComposer (Ours)" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 232, + 228, + 303, + 288 + ], + "lines": [ + { + "bbox": [ + 232, + 228, + 303, + 288 + ], + "spans": [ + { + "bbox": [ + 232, + 228, + 303, + 288 + ], + "type": "image", + "image_path": "18f0cda5ade8ba47f96b71f4301b491d6bc71c20e15a36443a15116cb2e47b61.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 308, + 229, + 368, + 288 + ], + "blocks": [ + { + "bbox": [ + 320, + 223, + 355, + 228 + ], + "lines": [ + { + "bbox": [ + 320, + 223, + 355, + 228 + ], + "spans": [ + { + "bbox": [ + 320, + 223, + 355, + 228 + ], + "type": "text", + "content": "Base Generation" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 308, + 229, + 368, + 288 + ], + "lines": [ + { + "bbox": [ + 308, + 229, + 368, + 288 + ], + "spans": [ + { + "bbox": [ + 308, + 229, + 368, + 288 + ], + "type": "image", + "image_path": "0aa271579bb01c6ababa8f61735064cd8927fc60625d7f9b97b3f346c69f5ac1.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 372, + 228, + 488, + 288 + ], + "blocks": [ + { + "bbox": [ + 387, + 223, + 419, + 228 + ], + "lines": [ + { + "bbox": [ + 387, + 223, + 419, + 228 + ], + "spans": [ + { + "bbox": [ + 387, + 223, + 419, + 228 + ], + "type": "text", + "content": "StableDiffusion" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 372, + 228, + 488, + 288 + ], + "lines": [ + { + "bbox": [ + 372, + 228, + 488, + 288 + ], + "spans": [ + { + "bbox": [ + 372, + 228, + 488, + 288 + ], + "type": "image", + "image_path": "5ba46fa145d4726e1122e747ab86aa294a2b1863d4c1a61be9ea580832cd46d2.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 388, + 289, + 469, + 295 + ], + "lines": [ + { + "bbox": [ + 388, + 289, + 469, + 295 + ], + "spans": [ + { + "bbox": [ + 388, + 289, + 469, + 295 + ], + "type": "text", + "content": "A photo of a yellow warbler" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 488, + 229, + 551, + 288 + ], + "blocks": [ + { + "bbox": [ + 499, + 222, + 545, + 228 + ], + "lines": [ + { + "bbox": [ + 499, + 222, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 499, + 222, + 545, + 228 + ], + "type": "text", + "content": "PartComposer (Ours)" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 488, + 229, + 551, + 288 + ], + "lines": [ + { + "bbox": [ + 488, + 229, + 551, + 288 + ], + "spans": [ + { + "bbox": [ + 488, + 229, + 551, + 288 + ], + "type": "image", + "image_path": "6c79eac0584f1dd146085c96e7f41970869375aad684081edae561f97669c963.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 59, + 304, + 121, + 369 + ], + "blocks": [ + { + "bbox": [ + 59, + 304, + 121, + 369 + ], + "lines": [ + { + "bbox": [ + 59, + 304, + 121, + 369 + ], + "spans": [ + { + "bbox": [ + 59, + 304, + 121, + 369 + ], + "type": "image", + "image_path": "18ed7dca323cecb7b8369f55cc64530b182b28400d69c3bb952ee3198dbace26.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 121, + 304, + 182, + 369 + ], + "blocks": [ + { + "bbox": [ + 121, + 304, + 182, + 369 + ], + "lines": [ + { + "bbox": [ + 121, + 304, + 182, + 369 + ], + "spans": [ + { + "bbox": [ + 121, + 304, + 182, + 369 + ], + "type": "image", + "image_path": "4b5f3d7461775cb04de5390b6eca162e4399a251c32c0fc57178d5729b4e0200.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 182, + 304, + 230, + 310 + ], + "lines": [ + { + "bbox": [ + 182, + 304, + 230, + 310 + ], + "spans": [ + { + "bbox": [ + 182, + 304, + 230, + 310 + ], + "type": "text", + "content": "Rich Text Gen" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 126, + 370, + 248, + 384 + ], + "lines": [ + { + "bbox": [ + 126, + 370, + 248, + 384 + ], + "spans": [ + { + "bbox": [ + 126, + 370, + 248, + 384 + ], + "type": "text", + "content": "a photo of a white swan, 8k, full hd \nrtComposer: crown → a crown of a peacock" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 306, + 304, + 367, + 370 + ], + "blocks": [ + { + "bbox": [ + 306, + 304, + 367, + 370 + ], + "lines": [ + { + "bbox": [ + 306, + 304, + 367, + 370 + ], + "spans": [ + { + "bbox": [ + 306, + 304, + 367, + 370 + ], + "type": "image", + "image_path": "08bf5885f58861b409b88bafcaf7286cbf6888ddbf86ad06ae8de1d4e33855de.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 308, + 370, + 550, + 384 + ], + "lines": [ + { + "bbox": [ + 308, + 370, + 550, + 384 + ], + "spans": [ + { + "bbox": [ + 308, + 370, + 550, + 384 + ], + "type": "text", + "content": "A young woman sits at a table in a beautiful, lush garden, painting by claude monet PartComposer: dress " + }, + { + "bbox": [ + 308, + 370, + 550, + 384 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 308, + 370, + 550, + 384 + ], + "type": "text", + "content": " a dress in Uhuie-style" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 54, + 392, + 555, + 426 + ], + "lines": [ + { + "bbox": [ + 54, + 392, + 555, + 426 + ], + "spans": [ + { + "bbox": [ + 54, + 392, + 555, + 426 + ], + "type": "text", + "content": "Figure 1. Base Generation (left) Comparison with Methods for Generations with Parts Details (right). PartComposer allows the generation of object images with specified attributes (color, style etc.) of parts for the chosen object in the base text prompt. StableDiffusion and Rich-Text [18] methods with part details either ignore the part instructions or generate inconsistent objects." + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 371, + 304, + 551, + 370 + ], + "blocks": [ + { + "bbox": [ + 377, + 295, + 481, + 301 + ], + "lines": [ + { + "bbox": [ + 377, + 295, + 481, + 301 + ], + "spans": [ + { + "bbox": [ + 377, + 295, + 481, + 301 + ], + "type": "text", + "content": "PartComposer: beak " + }, + { + "bbox": [ + 377, + 295, + 481, + 301 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 377, + 295, + 481, + 301 + ], + "type": "text", + "content": " a hens beak" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 371, + 304, + 551, + 370 + ], + "lines": [ + { + "bbox": [ + 371, + 304, + 551, + 370 + ], + "spans": [ + { + "bbox": [ + 371, + 304, + 551, + 370 + ], + "type": "image", + "image_path": "d3cb014fdd70930f1cf40349861e3b4002971da70967ffc7a4a9306771d27268.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "bbox": [ + 151, + 436, + 200, + 449 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 436, + 200, + 449 + ], + "spans": [ + { + "bbox": [ + 151, + 436, + 200, + 449 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 53, + 462, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 462, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 53, + 462, + 297, + 714 + ], + "type": "text", + "content": "Image composition and generation are processes where the artists need control over various parts of the generated images. However, the current state-of-the-art generation models, like Stable Diffusion, cannot handle fine-grained part-level attributes in the text prompts. Specifically, when additional attribute details are added to the base text prompt, these text-to-image models either generate an image vastly different from the image generated from the base prompt or ignore the attribute details. To mitigate these issues, we introduce PartComposer, a training-free method that enables image generation based on fine-grained part-level attributes specified for objects in the base text prompt. This allows more control for artists and enables novel object compositions by combining distinctive object parts. PartComposer first localizes object parts by denoising the object region from a specific diffusion process. This enables each part token to be localized to the right region. After obtaining part masks, we run a localized diffusion process in each part region based on fine-grained part attributes and combine them to produce the final image. All stages of PartComposer are based on repurposing a pre-trained diffusion model, which" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 313, + 437, + 556, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 556, + 486 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 556, + 486 + ], + "type": "text", + "content": "enables it to generalize across domains. We demonstrate the effectiveness of part-level control provided by PartComposer through qualitative visual examples and quantitative comparisons with contemporary baselines." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 314, + 511, + 394, + 523 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 511, + 394, + 523 + ], + "spans": [ + { + "bbox": [ + 314, + 511, + 394, + 523 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 312, + 533, + 555, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 533, + 555, + 664 + ], + "spans": [ + { + "bbox": [ + 312, + 533, + 555, + 664 + ], + "type": "text", + "content": "Image generation with large generative diffusion models like StableDiffusion [40], DALLE [38], etc., has become prevalent due to their superior quality and extensive world knowledge. These models are trained on large image-caption datasets and are trained to generate images based on a given text prompt (description). As image composition and creation are creative processes where the artists need control over various parts of the image being generated. However, adding additional details for controlling part appearance in the text prompt either changes the generated image entirely or ignores the part instructions [9] (Fig. 1)." + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 313, + 665, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 556, + 713 + ], + "type": "text", + "content": "Various works aim to provide improved spatial controls to image generation, as they allow image generation conditioned on segmentation masks [5], edge maps [55], bounding boxes [10, 27] etc. Popular methods such as Control-" + } + ] + } + ], + "index": 34 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13209" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 180 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 180 + ], + "type": "text", + "content": "Net [55], GLIGEN [27], etc. require specifying training of these conditional modules, which allows for the controllability of these large generative models. Further, there has been some development of training-free approaches [9, 10] which enable controlled generation of objects by modulating the internal cross-attention activations in the diffusion process. This demonstrates that these pre-trained generative models contain information about the spatial parts of the image, and modulating them effectively can lead to image compositions" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 182, + 297, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 182, + 297, + 386 + ], + "spans": [ + { + "bbox": [ + 56, + 182, + 297, + 386 + ], + "type": "text", + "content": "However, despite several attempts [6, 55], the semantic controllability of the image generation is still restricted to specifying details at the object level. The object level details can be restrictive, as often creative designers synthesize object parts (e.g. shirt, trouser etc.) and then compose them [23]. Further, variations in semantic parts are often very distinctive and are uniquely used to identify objects [12, 30]. For example, we usually identify species of birds by their unique beak. Due to this, semantic understanding and recognition of parts have been widely studied as a topic in computer vision [12, 13, 21, 31]. Hence, providing controllability for image synthesis at the object semantic part level can enable a large variety of image compositions [23]. Towards this goal, we introduce PartComposer, which provides users with an interface through which they can select the object in the scene and provide a semantic part-level description with fine-grained details for the object generation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 389, + 296, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 389, + 296, + 687 + ], + "spans": [ + { + "bbox": [ + 56, + 389, + 296, + 687 + ], + "type": "text", + "content": "In PartComposer, we develop a scheme to extract part-level localization masks from the Diffusion model. We introduce a parallel part diffusion process that generates masks for the object parts. The core idea of the approach is that by forcing the part diffusion model to specifically denoise only the object region in the image, it is possible to understand the locations of various parts in the object. After denoising just the object through the part model, the information inside the part diffusion model present in attention maps can be used to generate the masks for various object parts. In the following Part Generation stage, PartComposer utilizes the part masks and attributes of each part the user provides. We enable users to provide a highly expressive specification of parts by using a Rich-Text [18] interface, which allows the specifying attributes, such as style, color, etc., for each part. For the final image generation, taking inspiration from recent studies like Rich-Text Generation [18], Multi-Diffusion [5] etc., we compose the various object parts, by running parallel masked diffusion process for each part while combining them periodically into the image. This combination enables harmonious composition of parts, and also masking ensures only local modifications to the regions corresponding to each part of the object specified. In the PartComposer method, we only use the pre-trained StableDiffusion model, making it a generalizable and training-free approach." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "We extensively test the proposed PartComposer approach for the zero-shot object part segmentation, which is a chal" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 555, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 179 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 179 + ], + "type": "text", + "content": "lenging setup in computer vision. We evaluate the part segmentation approach on DeepFashion [29] and CUB200 [50] datasets, where our method significantly outperforms the baseline StableDiffusion (SD). Further, to evaluate the PartComposer image composition abilities, we also provide quantitative results along with a user preference study, where our method significantly outperforms the baselines in generating images consistent with the described parts (Fig. 1). We summarize the core contributions of our paper below:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 315, + 180, + 555, + 396 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 315, + 180, + 555, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 180, + 555, + 239 + ], + "spans": [ + { + "bbox": [ + 315, + 180, + 555, + 239 + ], + "type": "text", + "content": "1. We introduce PartComposer, a training-free method that enables the generation of object images by using provided fine-grained details for the parts of the object. For example, while generating a bird image we can specify a detailed description of its beak." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 315, + 239, + 555, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 239, + 555, + 323 + ], + "spans": [ + { + "bbox": [ + 315, + 239, + 555, + 323 + ], + "type": "text", + "content": "2. In PartComposer, we introduce a novel Part-Diffusion process, which localizes and provides masks for parts of a base object generated by the Diffusion model (Fig. 3). To localize object parts, we introduce a novel segmentation scheme that uses the attention maps of the base diffusion and part diffusion process to obtain accurate masks for localized parts." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 315, + 323, + 555, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 323, + 555, + 396 + ], + "spans": [ + { + "bbox": [ + 315, + 323, + 555, + 396 + ], + "type": "text", + "content": "3. In PartComposer, after localization, we enable the generation of parts from the pre-trained diffusion model based on the Rich-Text description for the parts provided by the user. PartComposer then composes the image to harmoniously blend all object parts with the background by combining localized diffusion paths (Fig. 2)." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 406, + 405, + 418 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 406, + 405, + 418 + ], + "spans": [ + { + "bbox": [ + 314, + 406, + 405, + 418 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 426, + 556, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 556, + 593 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 556, + 593 + ], + "type": "text", + "content": "Text-to-Image Generative Models. The text-to-image models synthesize images by following a textual description provided as a prompt. These models have recently become mainstream due to their superior image generation quality and significant knowledge base. This is an outcome of the availability of large-scale image caption datasets [42] and highly parallelized GPU clusters. Almost all kinds of generative models, such as GANs [22], Autoregressive [41, 53], and Diffusion models [38, 40], have shown significant improvements in quality with training on these large image caption datasets. Among these models, the StableDiffusion (SD) [40] models, based on denoising diffusion in latent space are popular due to their open-source nature, which we also utilize for experiments in our work." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 594, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 556, + 714 + ], + "type": "text", + "content": "Text-to-Image Models for Downstream Tasks. Generative models, in general, have been useful for various downstream tasks, particularly ones based on per-pixel prediction like Segmentation [1, 11], Depth Estimation [7], etc. As layers near the image generation output have features that capture the pixel-pixel relation [7]. With the large-scale text-to-image generative models, these models often perform very competitively [46, 52] to discriminative methods, on tasks like segmentation. However, one commonality among most of these segmentation methods is that they operate at the" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "13210" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 82, + 82, + 527, + 255 + ], + "blocks": [ + { + "bbox": [ + 82, + 82, + 527, + 255 + ], + "lines": [ + { + "bbox": [ + 82, + 82, + 527, + 255 + ], + "spans": [ + { + "bbox": [ + 82, + 82, + 527, + 255 + ], + "type": "image", + "image_path": "9e8bec10eaa019538a476686de36837aa7edd59fc6536a23d9a6b5097712088d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 255, + 555, + 289 + ], + "lines": [ + { + "bbox": [ + 54, + 255, + 555, + 289 + ], + "spans": [ + { + "bbox": [ + 54, + 255, + 555, + 289 + ], + "type": "text", + "content": "Figure 2. PartComposer takes the input of a base prompt, the selected object o, and a Rich-Text description ( i.e., footnote, colors, etc.) of the parts. The Part Diffusion process generates the object masks for the specified parts. Then, PartComposer runs a parallel Region Diffusion process to generate attributes of the specified parts, guided by base generation's intermediate outputs." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 295, + 294, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 295, + 294, + 354 + ], + "spans": [ + { + "bbox": [ + 54, + 295, + 294, + 354 + ], + "type": "text", + "content": "granularity of object or instance level. Text-to-image models here are at an advantage as the usual image captions describe the scene at an object level. In this work, we take a step further in exploring the object part-level knowledge of these text-to-image diffusion models." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 356, + 295, + 463 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 356, + 295, + 463 + ], + "spans": [ + { + "bbox": [ + 54, + 356, + 295, + 463 + ], + "type": "text", + "content": "Part Discovery and Segmentation. Part Discovery and Attribution were an integral part of computer vision pipelines classically, as these approaches were robust to viewpoint variations [24, 31, 48]. In deep learning, the unsupervised (self-supervised) approaches for part discovery like SCOPS [21] and Unsup-Parts [12] became popular as they generalize to object parts across categories. In this work, we go one step ahead and operate in a zero-shot unsupervised part segmentation setting, generating part masks from T2I models." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 464, + 295, + 595 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 464, + 295, + 595 + ], + "spans": [ + { + "bbox": [ + 54, + 464, + 295, + 595 + ], + "type": "text", + "content": "Controllable Image Generation. Masks, bounding boxes, edge maps, depth maps, etc., have been explored to control the generations of text-to-image diffusion models [6, 14, 27, 35, 36, 55] in addition to text. Further, various other approaches [2-4, 9, 34, 45, 49, 51] achieve control of image semantics through modulating diffusion models. Despite this, the text-to-image generation control at the object part level is under-explored; a recent work [33] tries to do it in a controlled supervised setup using part-masks. Contrary to that, in our work PartComposer we explore a generalized training-free zero-shot setting." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 609, + 111, + 620 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 609, + 111, + 620 + ], + "spans": [ + { + "bbox": [ + 55, + 609, + 111, + 620 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 629, + 295, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 629, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 54, + 629, + 295, + 712 + ], + "type": "text", + "content": "In this section, we introduce PartComposer, our method to synthesize objects based on the description of parts of the objects. In PartComposer, we ask the user to specify a base prompt and the token for the object for which it wants to synthesize parts. Then, we provide the user a rich-text [18] editor (Fig. 2) to specify the parts and their description. PartComposer involves two diffusion steps, a) Part Localiza" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 295, + 555, + 391 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 295, + 555, + 391 + ], + "spans": [ + { + "bbox": [ + 313, + 295, + 555, + 391 + ], + "type": "text", + "content": "tion: In the early diffusion stage, we get a mask for the object we want to divide into parts. Then, we perform denoising in later stage from a U-Net condition on parts to fill the masked region of the object, during which it learns to denoise different object parts. Due to this, the attention maps for various parts highlight the correct part region, which we use to extract the part mask. The infilling process is the major contribution of the PartComposer method (Fig. 3). b)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 392, + 556, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 392, + 556, + 487 + ], + "spans": [ + { + "bbox": [ + 313, + 392, + 556, + 487 + ], + "type": "text", + "content": "Part Generation: For generating parts, we combine region-specific diffusion processes for various parts by iteratively merging them inspired by MultiDiffusion [5], Rich-Text Generation [18] etc. However, till now, most of these works have combined the diffusion process for generating objects; for the first time, we have demonstrated its effectiveness in generating object parts. We provide an overview of the PartComposer pipeline in Fig. 2." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 493, + 406, + 506 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 493, + 406, + 506 + ], + "spans": [ + { + "bbox": [ + 313, + 493, + 406, + 506 + ], + "type": "text", + "content": "3.1. Problem Setup" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 510, + 556, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 556, + 629 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 556, + 629 + ], + "type": "text", + "content": "We consider each span of tokens " + }, + { + "bbox": [ + 313, + 510, + 556, + 629 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_i" + }, + { + "bbox": [ + 313, + 510, + 556, + 629 + ], + "type": "text", + "content": " as indicative of describing one part of the object, with attributes " + }, + { + "bbox": [ + 313, + 510, + 556, + 629 + ], + "type": "inline_equation", + "content": "\\mathbf{a}_i" + }, + { + "bbox": [ + 313, + 510, + 556, + 629 + ], + "type": "text", + "content": " describing its overall appearance. Our design choices are based on the rich-text generation [18], and we allow the user to specify the following type of attributes for part generation (Fig. 2): Part Description (i.e. footnote). It is an important attribute that specifies the part of an object; for example, for the part " + }, + { + "bbox": [ + 313, + 510, + 556, + 629 + ], + "type": "inline_equation", + "content": "(\\mathbf{p}_i)" + }, + { + "bbox": [ + 313, + 510, + 556, + 629 + ], + "type": "text", + "content": " 'crown' of a bird, we can specify the attribute " + }, + { + "bbox": [ + 313, + 510, + 556, + 629 + ], + "type": "inline_equation", + "content": "(\\mathbf{a}_i^p = \\text{'a crown of a peacock')" + }, + { + "bbox": [ + 313, + 510, + 556, + 629 + ], + "type": "text", + "content": ". This helps specify novel part descriptions, which can lead to artistic novel compositions." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 630, + 553, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 553, + 689 + ], + "type": "text", + "content": "Font Color. It helps us specify the exact RGB values for the color attribute " + }, + { + "bbox": [ + 313, + 630, + 553, + 689 + ], + "type": "inline_equation", + "content": "\\mathbf{a}_i^c" + }, + { + "bbox": [ + 313, + 630, + 553, + 689 + ], + "type": "text", + "content": " we want for the object part " + }, + { + "bbox": [ + 313, + 630, + 553, + 689 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_i" + }, + { + "bbox": [ + 313, + 630, + 553, + 689 + ], + "type": "text", + "content": ". The exact value of RGB allows fine-grained control over the color of the desired part, whereby just specifying specific colors like 'brick red' leads to ignorance by Stable Diffusion [18]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 689, + 553, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 553, + 712 + ], + "type": "text", + "content": "Font Style. The font style is indicative of the specific artistic style " + }, + { + "bbox": [ + 313, + 689, + 553, + 712 + ], + "type": "inline_equation", + "content": "\\mathbf{a}_i^s" + }, + { + "bbox": [ + 313, + 689, + 553, + 712 + ], + "type": "text", + "content": " like 'of Claude Monet' and 'of Van Gogh' when" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13211" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 70, + 553, + 250 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 553, + 250 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 553, + 250 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 553, + 250 + ], + "type": "image", + "image_path": "1c5ae7a02e4d99bd276ccf4fef6f5746b69ba170adfdbf53fd9061871a3d1c37.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 260, + 555, + 285 + ], + "lines": [ + { + "bbox": [ + 54, + 260, + 555, + 285 + ], + "spans": [ + { + "bbox": [ + 54, + 260, + 555, + 285 + ], + "type": "text", + "content": "Figure 3. Localization. We obtain the mask " + }, + { + "bbox": [ + 54, + 260, + 555, + 285 + ], + "type": "inline_equation", + "content": "(\\mathcal{M}_o)" + }, + { + "bbox": [ + 54, + 260, + 555, + 285 + ], + "type": "text", + "content": " for object in middle of diffusion process. We then denoise object in masked region, using parts " + }, + { + "bbox": [ + 54, + 260, + 555, + 285 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_i" + }, + { + "bbox": [ + 54, + 260, + 555, + 285 + ], + "type": "text", + "content": " conditioned U-Net. Due to part denoising, the attention maps of " + }, + { + "bbox": [ + 54, + 260, + 555, + 285 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_i" + }, + { + "bbox": [ + 54, + 260, + 555, + 285 + ], + "type": "text", + "content": " can now localize parts as masks " + }, + { + "bbox": [ + 54, + 260, + 555, + 285 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{\\mathbf{p}_i}" + }, + { + "bbox": [ + 54, + 260, + 555, + 285 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 289, + 295, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 289, + 295, + 326 + ], + "spans": [ + { + "bbox": [ + 55, + 289, + 295, + 326 + ], + "type": "text", + "content": "synthesizing images of paintings. This instructs the model to generate a part following a specific style for the given prompt and blend it with other parts of the object (Sec. 5)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 326, + 295, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 326, + 295, + 350 + ], + "spans": [ + { + "bbox": [ + 55, + 326, + 295, + 350 + ], + "type": "text", + "content": "Font Size. The font size controls the relative size of each part in a generation [19]. We use the " + }, + { + "bbox": [ + 55, + 326, + 295, + 350 + ], + "type": "inline_equation", + "content": "\\mathbf{a}_i^w" + }, + { + "bbox": [ + 55, + 326, + 295, + 350 + ], + "type": "text", + "content": " to denote size." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 357, + 160, + 368 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 357, + 160, + 368 + ], + "spans": [ + { + "bbox": [ + 55, + 357, + 160, + 368 + ], + "type": "text", + "content": "3.2. Part Localization" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 373, + 297, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 373, + 297, + 601 + ], + "spans": [ + { + "bbox": [ + 54, + 373, + 297, + 601 + ], + "type": "text", + "content": "For part localization, we first run the base stable diffusion model for the given text prompt and extract the token map " + }, + { + "bbox": [ + 54, + 373, + 297, + 601 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_o" + }, + { + "bbox": [ + 54, + 373, + 297, + 601 + ], + "type": "text", + "content": " for the object " + }, + { + "bbox": [ + 54, + 373, + 297, + 601 + ], + "type": "inline_equation", + "content": "\\mathbf{o}" + }, + { + "bbox": [ + 54, + 373, + 297, + 601 + ], + "type": "text", + "content": " specified by the user, by following the technique of clustering the self-attention maps [18, 37]. We obtain the mask using this clustering after denoising between the initial step " + }, + { + "bbox": [ + 54, + 373, + 297, + 601 + ], + "type": "inline_equation", + "content": "(T)" + }, + { + "bbox": [ + 54, + 373, + 297, + 601 + ], + "type": "text", + "content": " until a threshold time step " + }, + { + "bbox": [ + 54, + 373, + 297, + 601 + ], + "type": "inline_equation", + "content": "T_{th}" + }, + { + "bbox": [ + 54, + 373, + 297, + 601 + ], + "type": "text", + "content": ". After obtaining the binary object mask " + }, + { + "bbox": [ + 54, + 373, + 297, + 601 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_o" + }, + { + "bbox": [ + 54, + 373, + 297, + 601 + ], + "type": "text", + "content": " using attention masks [18, 37], we run two parallel diffusion processes where one contains input from the base prompt and the other contains input from the part prompts " + }, + { + "bbox": [ + 54, + 373, + 297, + 601 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{p}} = [\\mathbf{p}_1,\\mathbf{p}_2,\\dots,\\mathbf{p}_n]" + }, + { + "bbox": [ + 54, + 373, + 297, + 601 + ], + "type": "text", + "content": ", a token " + }, + { + "bbox": [ + 54, + 373, + 297, + 601 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_i" + }, + { + "bbox": [ + 54, + 373, + 297, + 601 + ], + "type": "text", + "content": " for each of the parts specified in the part text prompt. We denote the " + }, + { + "bbox": [ + 54, + 373, + 297, + 601 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{b}}" + }, + { + "bbox": [ + 54, + 373, + 297, + 601 + ], + "type": "text", + "content": " to denote the base text prompt and " + }, + { + "bbox": [ + 54, + 373, + 297, + 601 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{p}}" + }, + { + "bbox": [ + 54, + 373, + 297, + 601 + ], + "type": "text", + "content": " to denote the part prompt. For the time " + }, + { + "bbox": [ + 54, + 373, + 297, + 601 + ], + "type": "inline_equation", + "content": "t\\leq T_{th}" + }, + { + "bbox": [ + 54, + 373, + 297, + 601 + ], + "type": "text", + "content": " (by default, we use " + }, + { + "bbox": [ + 54, + 373, + 297, + 601 + ], + "type": "inline_equation", + "content": "T_{th}\\approx T / 2" + }, + { + "bbox": [ + 54, + 373, + 297, + 601 + ], + "type": "text", + "content": "), we denoise the U-Net in the object region by denoising it with both the combination of the Part Prompt output and Base Prompt output. Due to this, the Part-Based U-Net gets the information regarding various parts in the object (Fig. 3). We now mathematically define the output noise " + }, + { + "bbox": [ + 54, + 373, + 297, + 601 + ], + "type": "inline_equation", + "content": "\\epsilon_t" + }, + { + "bbox": [ + 54, + 373, + 297, + 601 + ], + "type": "text", + "content": " for the diffusion process with Part-Diffusion below:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 60, + 620, + 296, + 634 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 620, + 296, + 634 + ], + "spans": [ + { + "bbox": [ + 60, + 620, + 296, + 634 + ], + "type": "interline_equation", + "content": "\\epsilon_ {t} = \\alpha \\mathcal {M} _ {o} \\odot D (x _ {t}, \\hat {\\mathbf {p}}, t) + (1 - \\alpha \\mathcal {M} _ {o}) \\odot D (x _ {t}, \\hat {\\mathbf {b}}, t) \\tag {1}", + "image_path": "0b6db55990d4b0c738efcab072c2a58854899bf6d313257c46a414badab9d71e.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": "here, the " + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": " is the hyper-parameter controlling the strength of the part prompt diffusion output, and " + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": " is the output of the pre-trained U-Net of a text-to-image diffusion model. The above denoising process is followed for t steps until the last step to produce a base image corresponding to the given prompt " + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\hat{b}" + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": ". Keeping a high " + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "inline_equation", + "content": "T_{th}" + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": " and low " + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": ", makes minimal" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 289, + 555, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 289, + 555, + 350 + ], + "spans": [ + { + "bbox": [ + 313, + 289, + 555, + 350 + ], + "type": "text", + "content": "changes in output as if the denoising diffusion was done with original prompt " + }, + { + "bbox": [ + 313, + 289, + 555, + 350 + ], + "type": "inline_equation", + "content": "\\hat{b}" + }, + { + "bbox": [ + 313, + 289, + 555, + 350 + ], + "type": "text", + "content": " (See Suppl. Sec.). With this part denoising of object, we obtain attention masks, from which we extract the localization information of part " + }, + { + "bbox": [ + 313, + 289, + 555, + 350 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_t" + }, + { + "bbox": [ + 313, + 289, + 555, + 350 + ], + "type": "text", + "content": ". We now describe the process of obtaining the part masks." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 354, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 354, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 354, + 556, + 714 + ], + "type": "text", + "content": "Token Maps for Parts. We first take the part tokens " + }, + { + "bbox": [ + 313, + 354, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{p}}" + }, + { + "bbox": [ + 313, + 354, + 556, + 714 + ], + "type": "text", + "content": ", which are a concatenation of the part names (i.e. 'beak crown wings'), which may not make a meaningful text prompt. Hence, we initialize text embeddings for all these tokens " + }, + { + "bbox": [ + 313, + 354, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_i" + }, + { + "bbox": [ + 313, + 354, + 556, + 714 + ], + "type": "text", + "content": " by passing the meaningful text prompt having the following template: \"A photo of " + }, + { + "bbox": [ + 313, + 354, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_i" + }, + { + "bbox": [ + 313, + 354, + 556, + 714 + ], + "type": "text", + "content": " of a o\" where " + }, + { + "bbox": [ + 313, + 354, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_i" + }, + { + "bbox": [ + 313, + 354, + 556, + 714 + ], + "type": "text", + "content": " is the object part name and o is the object name. This serves two purposes: first, it makes the text embedding meaningful, and second, it introduces some invariance from the order of part specification in part prompt. These embeddings are then passed as text embeddings to the Part U-Net for denoising. After running the denoising process, we aggregate the self-attention maps across multi-heads and time steps (from 32 × 32 resolution) for both the base and part U-Net diffusion branches, taking inspiration from works that demonstrate that attention can localize objects [9, 18, 30, 44]. We then perform spectral clustering on these attention maps to form k segmentation maps " + }, + { + "bbox": [ + 313, + 354, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{M}}" + }, + { + "bbox": [ + 313, + 354, + 556, + 714 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 313, + 354, + 556, + 714 + ], + "type": "inline_equation", + "content": "32 \\times 32" + }, + { + "bbox": [ + 313, + 354, + 556, + 714 + ], + "type": "text", + "content": "), based on pixel-pixel similarity. To attribute these K-segments to the part specified by the user, we aggregate the cross attention of Part U-Net diffusion process. For each token " + }, + { + "bbox": [ + 313, + 354, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_j" + }, + { + "bbox": [ + 313, + 354, + 556, + 714 + ], + "type": "text", + "content": ", we obtain the cross-attention score as follows: " + }, + { + "bbox": [ + 313, + 354, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{m}}_j = \\frac{c_j}{\\sum_k c_k}" + }, + { + "bbox": [ + 313, + 354, + 556, + 714 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 354, + 556, + 714 + ], + "type": "inline_equation", + "content": "c_j" + }, + { + "bbox": [ + 313, + 354, + 556, + 714 + ], + "type": "text", + "content": " is the cross attention score for each token. We proceed by aggregating the attention heads to obtain the average cross-attention scores and resizing them to " + }, + { + "bbox": [ + 313, + 354, + 556, + 714 + ], + "type": "inline_equation", + "content": "32 \\times 32" + }, + { + "bbox": [ + 313, + 354, + 556, + 714 + ], + "type": "text", + "content": ", obtaining " + }, + { + "bbox": [ + 313, + 354, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{m}}" + }, + { + "bbox": [ + 313, + 354, + 556, + 714 + ], + "type": "text", + "content": ". We remove the start of text (" + }, + { + "bbox": [ + 313, + 354, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\langle \\mathrm{sot} \\rangle" + }, + { + "bbox": [ + 313, + 354, + 556, + 714 + ], + "type": "text", + "content": ") token for cross-attention and re-normalize it [9]. In other works [18, 37], as the tokens correspond to the objects being generated in the image, it's sure that token maps will be meaningful. However, this is not true for parts, as Part U-Net might not localize some" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13212" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "text", + "content": "parts. For determining if the part is localized, we propose to look at the max value of cross attention map spatially across the pixels; if we find that the following condition is met, the part is localized:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 100, + 125, + 296, + 149 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 125, + 296, + 149 + ], + "spans": [ + { + "bbox": [ + 100, + 125, + 296, + 149 + ], + "type": "interline_equation", + "content": "L (j) = \\mathbb {1} \\left\\{\\max \\left(\\hat {\\mathbf {m}} _ {j}\\right) \\geq (1 - \\delta) \\frac {1}{K} \\right\\}. \\tag {2}", + "image_path": "537b778afe411e49ebb2e24bdf98ddc91c847a93cf2f2cc532530b88874bacfa.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 154, + 296, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 154, + 296, + 202 + ], + "spans": [ + { + "bbox": [ + 55, + 154, + 296, + 202 + ], + "type": "text", + "content": "Here " + }, + { + "bbox": [ + 55, + 154, + 296, + 202 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 55, + 154, + 296, + 202 + ], + "type": "text", + "content": " is a hyperparameter, and " + }, + { + "bbox": [ + 55, + 154, + 296, + 202 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 154, + 296, + 202 + ], + "type": "text", + "content": " is the number of parts. Is This condition is robust in finding the localized parts (See Sec. 5 for ablation). For the parts " + }, + { + "bbox": [ + 55, + 154, + 296, + 202 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_i" + }, + { + "bbox": [ + 55, + 154, + 296, + 202 + ], + "type": "text", + "content": ", which are localized, we normalize the cross-attention map:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 66, + 207, + 285, + 235 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 207, + 285, + 235 + ], + "spans": [ + { + "bbox": [ + 66, + 207, + 285, + 235 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {m}} _ {j} = L (j) \\frac {\\hat {\\mathbf {m}} _ {j} - \\min (\\hat {\\mathbf {m}} _ {j})}{\\max (\\hat {\\mathbf {m}} _ {j}) - \\min (\\hat {\\mathbf {m}} _ {j})} + (1 - L (j)) \\hat {\\mathbf {m}} _ {j}.", + "image_path": "a2383e9a36cc8eb6a2d17b42389619b874b586f350896decba2208db8a9228f7.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 239, + 296, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 239, + 296, + 396 + ], + "spans": [ + { + "bbox": [ + 55, + 239, + 296, + 396 + ], + "type": "text", + "content": "We follow a dot-product-based protocol to assign each K cluster in the self-attention masks to a part, unlike the average attention protocol in the previous works [18, 37]. We find that dot product of normalized cross attention scores " + }, + { + "bbox": [ + 55, + 239, + 296, + 396 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{m}}_j" + }, + { + "bbox": [ + 55, + 239, + 296, + 396 + ], + "type": "text", + "content": " for each token with self-attention masks " + }, + { + "bbox": [ + 55, + 239, + 296, + 396 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{M}}_j" + }, + { + "bbox": [ + 55, + 239, + 296, + 396 + ], + "type": "text", + "content": " works better, as the attention maps in Part U-Net are noisy. Still, they are often correct for the regions that are localized in one specific area only (Suppl. Fig. 14). Hence, the dot product protocol favors those maps that are only localized in some areas of the image and don't have high attention values across all parts of the object. After obtaining the dot product scores for each part, we assign " + }, + { + "bbox": [ + 55, + 239, + 296, + 396 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{M}}_j" + }, + { + "bbox": [ + 55, + 239, + 296, + 396 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 239, + 296, + 396 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_i" + }, + { + "bbox": [ + 55, + 239, + 296, + 396 + ], + "type": "text", + "content": " with the highest scores. The mask for the part " + }, + { + "bbox": [ + 55, + 239, + 296, + 396 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_i" + }, + { + "bbox": [ + 55, + 239, + 296, + 396 + ], + "type": "text", + "content": " is finally given as the following:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 413, + 296, + 438 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 413, + 296, + 438 + ], + "spans": [ + { + "bbox": [ + 55, + 413, + 296, + 438 + ], + "type": "interline_equation", + "content": "\\mathbf {M} _ {\\mathbf {p} _ {i}} = \\left\\{\\cup_ {j} \\hat {\\mathbf {M}} _ {j} \\mid \\operatorname {a r g m a x} _ {i} \\hat {\\mathbf {M}} _ {j} \\cdot \\hat {\\mathbf {m}} _ {i} = i \\text {a n d} \\hat {\\mathbf {M}} _ {j} \\cdot \\hat {\\mathbf {m}} _ {i} \\geq \\epsilon , \\right\\} \\tag {3}", + "image_path": "885c3ebad938ee7e2482d3caa378be19b6b8c4013ca379f1cd5224390d170d33.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 439, + 296, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 439, + 296, + 498 + ], + "spans": [ + { + "bbox": [ + 55, + 439, + 296, + 498 + ], + "type": "text", + "content": "Here " + }, + { + "bbox": [ + 55, + 439, + 296, + 498 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 55, + 439, + 296, + 498 + ], + "type": "text", + "content": " is the hyperparameter which controls the minimum similarity required between the attention mask and the part token. We combine the additional attention masks unassigned to any token, and name them as the background (other) token " + }, + { + "bbox": [ + 55, + 439, + 296, + 498 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_b" + }, + { + "bbox": [ + 55, + 439, + 296, + 498 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 505, + 155, + 516 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 505, + 155, + 516 + ], + "spans": [ + { + "bbox": [ + 55, + 505, + 155, + 516 + ], + "type": "text", + "content": "3.3. Part Generation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 522, + 297, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 522, + 297, + 643 + ], + "spans": [ + { + "bbox": [ + 55, + 522, + 297, + 643 + ], + "type": "text", + "content": "We follow a similar protocol for the generation of part segments as in Rich-Text Generation. We tailor the rich text generation to compose the part regions of the object in the image in place of the original scene composition. We describe the part generation protocol below briefly and refer readers to Rich-Text Gen [18] for more details. For each part " + }, + { + "bbox": [ + 55, + 522, + 297, + 643 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_i" + }, + { + "bbox": [ + 55, + 522, + 297, + 643 + ], + "type": "text", + "content": " we run a region diffusion process, which runs in parallel for all the parts. We then combine the region diffusion processes to obtain the final noise prediction " + }, + { + "bbox": [ + 55, + 522, + 297, + 643 + ], + "type": "inline_equation", + "content": "\\epsilon_t" + }, + { + "bbox": [ + 55, + 522, + 297, + 643 + ], + "type": "text", + "content": " as the masked " + }, + { + "bbox": [ + 55, + 522, + 297, + 643 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{\\mathbf{p}_i}" + }, + { + "bbox": [ + 55, + 522, + 297, + 643 + ], + "type": "text", + "content": " sum of the denoiser outputs:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 63, + 658, + 296, + 683 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 658, + 296, + 683 + ], + "spans": [ + { + "bbox": [ + 63, + 658, + 296, + 683 + ], + "type": "interline_equation", + "content": "\\epsilon_ {t} = \\sum_ {i} \\mathbf {M} _ {\\mathbf {p} _ {i}} \\epsilon_ {t, \\mathbf {p} _ {i}} = \\sum_ {i} \\mathbf {M} _ {\\mathbf {p} _ {i}} \\odot D (x _ {t}, f (\\mathbf {p} _ {i}, \\mathbf {a} _ {i}), t) \\tag {4}", + "image_path": "b5605f3a38651d4e23da81fa5141718122313f5659c366867ada761308ed3585.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": " is the pre-trained U-Net model, and " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "f(\\mathbf{p}_i,\\mathbf{a}_i)" + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": " is the text description of the part " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_i" + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": " constructed using the following" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 318, + 71, + 550, + 140 + ], + "blocks": [ + { + "bbox": [ + 318, + 71, + 550, + 140 + ], + "lines": [ + { + "bbox": [ + 318, + 71, + 550, + 140 + ], + "spans": [ + { + "bbox": [ + 318, + 71, + 550, + 140 + ], + "type": "image", + "image_path": "00ab72b86e894bcbc9ceccaba55020732970aac576a0369b3d46c107e7da795c.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 150, + 555, + 172 + ], + "lines": [ + { + "bbox": [ + 313, + 150, + 555, + 172 + ], + "spans": [ + { + "bbox": [ + 313, + 150, + 555, + 172 + ], + "type": "text", + "content": "Figure 4. Segmentation masks for the parts that are localized by Part Diffusion for DeepFashion (above) and CUB200 (below)." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 176, + 555, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 176, + 555, + 331 + ], + "spans": [ + { + "bbox": [ + 313, + 176, + 555, + 331 + ], + "type": "text", + "content": "process using the part tokens " + }, + { + "bbox": [ + 313, + 176, + 555, + 331 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_i" + }, + { + "bbox": [ + 313, + 176, + 555, + 331 + ], + "type": "text", + "content": " and attributes " + }, + { + "bbox": [ + 313, + 176, + 555, + 331 + ], + "type": "inline_equation", + "content": "\\mathbf{a}_i" + }, + { + "bbox": [ + 313, + 176, + 555, + 331 + ], + "type": "text", + "content": ". Initially the text " + }, + { + "bbox": [ + 313, + 176, + 555, + 331 + ], + "type": "inline_equation", + "content": "f(\\mathbf{p}_i, \\mathbf{a}_i) = \\mathbf{p}_i" + }, + { + "bbox": [ + 313, + 176, + 555, + 331 + ], + "type": "text", + "content": ", is set to part token itself. In case the part description (i.e. footnote) is available we set the " + }, + { + "bbox": [ + 313, + 176, + 555, + 331 + ], + "type": "inline_equation", + "content": "f(\\mathbf{p}_i, \\mathbf{a}_i) = \\mathbf{a}_i^p" + }, + { + "bbox": [ + 313, + 176, + 555, + 331 + ], + "type": "text", + "content": ", further if the style attribute is available we do " + }, + { + "bbox": [ + 313, + 176, + 555, + 331 + ], + "type": "inline_equation", + "content": "f(\\mathbf{p}_i, \\mathbf{a}_i) = f(\\mathbf{p}_i, \\mathbf{a}_i) + \\text{in style of} + \\mathbf{a}_i^s" + }, + { + "bbox": [ + 313, + 176, + 555, + 331 + ], + "type": "text", + "content": ". In case the color attribute is also specified, the nearest named color " + }, + { + "bbox": [ + 313, + 176, + 555, + 331 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{a}}_i^c" + }, + { + "bbox": [ + 313, + 176, + 555, + 331 + ], + "type": "text", + "content": " (e.g. red) for the specific RBG color " + }, + { + "bbox": [ + 313, + 176, + 555, + 331 + ], + "type": "inline_equation", + "content": "\\mathbf{a}_i^c" + }, + { + "bbox": [ + 313, + 176, + 555, + 331 + ], + "type": "text", + "content": " is found, and " + }, + { + "bbox": [ + 313, + 176, + 555, + 331 + ], + "type": "inline_equation", + "content": "f(\\mathbf{p}_i, \\mathbf{a}_i) = \\hat{\\mathbf{a}}_i^c + ' ' + f(\\mathbf{p}_i, \\mathbf{a}_i)" + }, + { + "bbox": [ + 313, + 176, + 555, + 331 + ], + "type": "text", + "content": ". The string " + }, + { + "bbox": [ + 313, + 176, + 555, + 331 + ], + "type": "inline_equation", + "content": "f(\\mathbf{p}_i, \\mathbf{a}_i)" + }, + { + "bbox": [ + 313, + 176, + 555, + 331 + ], + "type": "text", + "content": " is the text input for the Diffusion to generate the part " + }, + { + "bbox": [ + 313, + 176, + 555, + 331 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_i" + }, + { + "bbox": [ + 313, + 176, + 555, + 331 + ], + "type": "text", + "content": ". We use the base prompt " + }, + { + "bbox": [ + 313, + 176, + 555, + 331 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{b}}" + }, + { + "bbox": [ + 313, + 176, + 555, + 331 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 313, + 176, + 555, + 331 + ], + "type": "inline_equation", + "content": "f(\\mathbf{p}_i, \\mathbf{a}_i)" + }, + { + "bbox": [ + 313, + 176, + 555, + 331 + ], + "type": "text", + "content": " for the background masked region " + }, + { + "bbox": [ + 313, + 176, + 555, + 331 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_b" + }, + { + "bbox": [ + 313, + 176, + 555, + 331 + ], + "type": "text", + "content": ". Combining different diffusion outputs at every time " + }, + { + "bbox": [ + 313, + 176, + 555, + 331 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 176, + 555, + 331 + ], + "type": "text", + "content": " helps generate a harmonious image after blending the defined parts of the object." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 332, + 556, + 438 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 332, + 556, + 438 + ], + "spans": [ + { + "bbox": [ + 313, + 332, + 556, + 438 + ], + "type": "text", + "content": "Following Rich-Text [18], we also utilize the gradient guidance [16, 20] by taking the gradient of MSE loss between the estimated original image and the color value " + }, + { + "bbox": [ + 313, + 332, + 556, + 438 + ], + "type": "inline_equation", + "content": "\\mathbf{a}_i^c" + }, + { + "bbox": [ + 313, + 332, + 556, + 438 + ], + "type": "text", + "content": " specified by the user. The gradient guidance helps generate the exact RGB color for part " + }, + { + "bbox": [ + 313, + 332, + 556, + 438 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_i" + }, + { + "bbox": [ + 313, + 332, + 556, + 438 + ], + "type": "text", + "content": ", which is impossible with just text guidance [18]. Further, as we use the base text framework of Rich-Text, we also can specify font size attribute to " + }, + { + "bbox": [ + 313, + 332, + 556, + 438 + ], + "type": "inline_equation", + "content": "\\mathbf{a}_i^f" + }, + { + "bbox": [ + 313, + 332, + 556, + 438 + ], + "type": "text", + "content": " to control the relative importance of each part in the object image." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 439, + 556, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 439, + 556, + 523 + ], + "spans": [ + { + "bbox": [ + 313, + 439, + 556, + 523 + ], + "type": "text", + "content": "Preservation of Other Parts. As we only intend to modify the object o from the original prompt, we also use the Self-Injection techniques from Plug and Play [47] to maintain the overall structure of the background from base prompt generation. Further, to ensure that our diffusion trajectory follows the same path as the base, the background region is also blended with base noise generation outputs." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 361, + 529, + 555, + 543 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 529, + 555, + 543 + ], + "spans": [ + { + "bbox": [ + 361, + 529, + 555, + 543 + ], + "type": "interline_equation", + "content": "x _ {t} = \\mathbf {M} _ {b} \\odot x _ {t} ^ {\\text {b a s e}} + (1 - \\mathbf {M} _ {b}) \\odot x _ {t} \\tag {5}", + "image_path": "609039dcec77284805b43174ba6c97ae061cd2053a7e758739126949a19c11a4.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 548, + 556, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 548, + 556, + 632 + ], + "spans": [ + { + "bbox": [ + 313, + 548, + 556, + 632 + ], + "type": "text", + "content": "In our case, the attention maps out of the object " + }, + { + "bbox": [ + 313, + 548, + 556, + 632 + ], + "type": "inline_equation", + "content": "\\mathbf{Mask}\\mathcal{M}_o" + }, + { + "bbox": [ + 313, + 548, + 556, + 632 + ], + "type": "text", + "content": " and all the object parts not assigned to any part token comprise the background " + }, + { + "bbox": [ + 313, + 548, + 556, + 632 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_b" + }, + { + "bbox": [ + 313, + 548, + 556, + 632 + ], + "type": "text", + "content": ", and we start this blending process at " + }, + { + "bbox": [ + 313, + 548, + 556, + 632 + ], + "type": "inline_equation", + "content": "t = T_{blend}" + }, + { + "bbox": [ + 313, + 548, + 556, + 632 + ], + "type": "text", + "content": ". We find this to be very useful in preserving the structure of the other parts of the image and just generating the described parts in the localized region. We provide an overview of the complete process in Fig. 2." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 640, + 446, + 654 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 640, + 446, + 654 + ], + "spans": [ + { + "bbox": [ + 313, + 640, + 446, + 654 + ], + "type": "text", + "content": "4. Experimental Analysis" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 660, + 484, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 660, + 484, + 672 + ], + "spans": [ + { + "bbox": [ + 313, + 660, + 484, + 672 + ], + "type": "text", + "content": "4.1. Evaluation of Part Localization" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "content": "We first evaluate the part localization module, which is based on the novel idea of denoising only the object region with the part-based diffusion outputs." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13213" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 102, + 96, + 509, + 251 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "text", + "content": "Table 1. Comparison to Prior Works for the unsupervised part segmentation task. We follow Unsup-Part [12] for evaluation protocols and baseline results for " + }, + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "inline_equation", + "content": "(\\mathbf{K} = 4)" + }, + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "text", + "content": " parts. We report clustering-based NMI and ARI metrics, which are higher for better segmentation outputs." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 102, + 96, + 509, + 251 + ], + "lines": [ + { + "bbox": [ + 102, + 96, + 509, + 251 + ], + "spans": [ + { + "bbox": [ + 102, + 96, + 509, + 251 + ], + "type": "table", + "html": "
DeepFashion DatasetCUB200 Dataset
MethodFG-NMIFG-ARINMIARIFG-NMIFG-ARINMIARI
Unsupervised Learning
SCOPS [21]30.727.656.681.439.117.924.47.1
Unsup-Parts [12]44.846.668.190.646.021.043.519.6
DFF [11]----32.414.325.912.4
Unsupervised Zero-Shot
Rich-Text [18]16.05.248.358.73.10.33.10.3
StableDiffusion12.03.540.670.98.00.63.30.6
PartComposer(Ours)24.718.048.073.420.59.218.57.7
", + "image_path": "0b0273349e42dfc9939d9ab4c28a7100bf544b2d6885296e80d72a613cb96e96.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 255, + 298, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 255, + 298, + 422 + ], + "spans": [ + { + "bbox": [ + 54, + 255, + 298, + 422 + ], + "type": "text", + "content": "Implementation Details. We use the StableDiffusion (SD) version 2.1 for our experimentation purposes. We use the DDIM Scheduler [43] with 50 steps to generate results for SD2.1 to evaluate the part localization process. As ground truth is not available for the part masks of the generated images, we use test sets of the commonly used DeepFashion [29] and CUB-200 [50] datasets for evaluation. These datasets are the standard datasets used for the evaluation of the unsupervised part segmentation approaches. The DeepFashion dataset contains images along with their part segmentation masks, divided into 14 categories of labels. The CUB200 dataset contains the key point annotations for the 14 key points specified for the bird categories. We provide further details in Suppl. Section." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 426, + 298, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 426, + 298, + 665 + ], + "spans": [ + { + "bbox": [ + 56, + 426, + 298, + 665 + ], + "type": "text", + "content": "Baselines and Problem Setting. As we operate in the setting of Zero-shot (i.e. no training) Unsupervised Part Segmentation, there are no previous works that report results in such a challenging setting. Hence, we provide results for the unsupervised learning approaches to facilitate comparison. We provide results for the SCOPS [21], which utilizes the internal features of the VGG model to train a model based on self-supervised loss functions to predict parts robustly across categories. The other stronger self-supervised baseline is Unsup-Parts [12], which uses contrastive loss functions to train a network based on equivariance and other vision properties to cluster the object regions into semantic parts. In addition to this, we also report results for the DFF [11] as they also operate in the same setting. We want to highlight that these unsupervised approaches require either training a neural network or performing clustering on the complete training data to segment parts. In contrast, our approach is training free and operates in a zero-shot fashion. Hence, the performance of zero-shot approaches is not directly comparable to unsupervised methods." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "Zero Shot Unsupervised Part Segmentation. As there is no benchmark to evaluate the part localization for generated images, we use the existing dataset of DeepFashion and CUB200 to obtain our results. To obtain segmentation for" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 255, + 556, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 255, + 556, + 519 + ], + "spans": [ + { + "bbox": [ + 313, + 255, + 556, + 519 + ], + "type": "text", + "content": "each image, we first invert the image into the diffusion latent space using Null-Text Inversion [32] method (see Sec. 5 for ablation). We use the BLIP-V2 [25] captioner provided in Diffusers to obtain approximate text prompts for inversion. After providing the desired image and prompt, the Null-Text inversion provides us with inverted latent and unconditioned time embeddings to reconstruct the image of the dataset. We then construct a StableDiffusion baseline in which, in addition to the prompts, we append the list of part tokens " + }, + { + "bbox": [ + 313, + 255, + 556, + 519 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 313, + 255, + 556, + 519 + ], + "type": "text", + "content": " to the prompt. We then use the segmentation algorithm described above in Sec. 3.2 to extract token maps corresponding to the part tokens " + }, + { + "bbox": [ + 313, + 255, + 556, + 519 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_i" + }, + { + "bbox": [ + 313, + 255, + 556, + 519 + ], + "type": "text", + "content": ". We also evaluate a Rich-Text [18] baseline segmentation algorithm for the same. For the proposed part-denoising approach based on parallel diffusion in PartComposer, we first get the inverted latents and embeddings. Then, we generate a base diffusion process to reconstruct the image and use Part Diffusion to fill the parts of the image. To very fairly compare the StableDiffusion performance with part-denoising, we keep all things the same except the part diffusion process to get masks. Further across these baselines, we use low classifier guidance to ensure proper reconstruction of the dataset images (See Suppl. for details)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 521, + 557, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 557, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 557, + 714 + ], + "type": "text", + "content": "Evaluation Protocol. We use the standard experimental protocol as used by earlier works [12] to facilitate the right point of comparison. We want to point out that in our work, the part name (e.g., beak, etc.) is associated with the localized mask, but the unsupervised approaches produce " + }, + { + "bbox": [ + 313, + 521, + 557, + 714 + ], + "type": "inline_equation", + "content": "(\\mathrm{K} = 4)" + }, + { + "bbox": [ + 313, + 521, + 557, + 714 + ], + "type": "text", + "content": " parts without any part names. Hence, to make a fair comparison, we create 4 clusters of parts based on their locality, and finally generate segmentation masks with a maximum of 4 clusters. We provide the exact mapping of the part names to cluster labels in the Suppl. Section. We compare the specifically designed metrics [12] of NMI (Normalized Mutual Information) and ARI (Adjusted Rand Index) of the predicted cluster labels with part masks in the case of DeepFashion and key points in case of the CUB dataset. We report the metrics along with their foreground variants (FG-NMI and FG-ARI) for a holistic comparison of performance." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "13214" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 70, + 71, + 164, + 148 + ], + "blocks": [ + { + "bbox": [ + 70, + 71, + 164, + 148 + ], + "lines": [ + { + "bbox": [ + 70, + 71, + 164, + 148 + ], + "spans": [ + { + "bbox": [ + 70, + 71, + 164, + 148 + ], + "type": "image", + "image_path": "8ce660be2b47af28118151ccd3f21a9815fa29a6634aa78c3e5c9f3e78410dfa.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 70, + 149, + 164, + 225 + ], + "blocks": [ + { + "bbox": [ + 70, + 149, + 164, + 225 + ], + "lines": [ + { + "bbox": [ + 70, + 149, + 164, + 225 + ], + "spans": [ + { + "bbox": [ + 70, + 149, + 164, + 225 + ], + "type": "image", + "image_path": "190c79cd63844621e26fc29025dab3fbdd375c615e10579360a13c7945077746.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 231, + 555, + 265 + ], + "lines": [ + { + "bbox": [ + 55, + 231, + 555, + 265 + ], + "spans": [ + { + "bbox": [ + 55, + 231, + 555, + 265 + ], + "type": "text", + "content": "Figure 5. Qualitative Comparison for PartComposer. (above) We provide base prompt 'a photo of a flamingo' and part prompt as 'beak - a pelicans beak'. (below) We generate a photo of a man, with part prompt as 'black jacket and blue jeans'. We find that other baselines either ignore instruction or change the entire composition. On the contrary, PartComposer can correctly localize parts and generate details." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 170, + 72, + 262, + 144 + ], + "blocks": [ + { + "bbox": [ + 170, + 72, + 262, + 144 + ], + "lines": [ + { + "bbox": [ + 170, + 72, + 262, + 144 + ], + "spans": [ + { + "bbox": [ + 170, + 72, + 262, + 144 + ], + "type": "image", + "image_path": "b76912a9f8ff072d94f0c8e42f6de4f760ff818a4488135154424211f29d200a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 170, + 149, + 263, + 220 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 170, + 149, + 263, + 220 + ], + "spans": [ + { + "bbox": [ + 170, + 149, + 263, + 220 + ], + "type": "image", + "image_path": "3b3265ecdc6d7fb0a937fbcdde20d1a42d93b6b88a208b74198a04d64c225431.jpg" + } + ] + } + ], + "index": 3, + "type": "text" + }, + { + "bbox": [ + 192, + 220, + 242, + 225 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 192, + 220, + 242, + 225 + ], + "spans": [ + { + "bbox": [ + 192, + 220, + 242, + 225 + ], + "type": "text", + "content": "PartCraft (Ours)" + } + ] + } + ], + "index": 4, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 262, + 72, + 354, + 143 + ], + "blocks": [ + { + "bbox": [ + 262, + 72, + 354, + 143 + ], + "lines": [ + { + "bbox": [ + 262, + 72, + 354, + 143 + ], + "spans": [ + { + "bbox": [ + 262, + 72, + 354, + 143 + ], + "type": "image", + "image_path": "b3d287cc8aeb7867fe07c9a4427a8e66c9ea52d84d9ddd1742bba5178393a21e.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 284, + 143, + 333, + 148 + ], + "lines": [ + { + "bbox": [ + 284, + 143, + 333, + 148 + ], + "spans": [ + { + "bbox": [ + 284, + 143, + 333, + 148 + ], + "type": "text", + "content": "StableDiffusion" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 354, + 72, + 447, + 143 + ], + "blocks": [ + { + "bbox": [ + 354, + 72, + 447, + 143 + ], + "lines": [ + { + "bbox": [ + 354, + 72, + 447, + 143 + ], + "spans": [ + { + "bbox": [ + 354, + 72, + 447, + 143 + ], + "type": "image", + "image_path": "a4a04e6b8fe4eb13aed038c7e1f22c36d4c5ef6eb396580b499242f2800abe0d.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 376, + 143, + 424, + 148 + ], + "lines": [ + { + "bbox": [ + 376, + 143, + 424, + 148 + ], + "spans": [ + { + "bbox": [ + 376, + 143, + 424, + 148 + ], + "type": "text", + "content": "InstructPix2Pix" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 446, + 72, + 538, + 143 + ], + "blocks": [ + { + "bbox": [ + 446, + 72, + 538, + 143 + ], + "lines": [ + { + "bbox": [ + 446, + 72, + 538, + 143 + ], + "spans": [ + { + "bbox": [ + 446, + 72, + 538, + 143 + ], + "type": "image", + "image_path": "0af86d5f4d4db4124a76497fa148205cae060aebd7e979085eb352b94bd8d3de.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 469, + 143, + 514, + 148 + ], + "lines": [ + { + "bbox": [ + 469, + 143, + 514, + 148 + ], + "spans": [ + { + "bbox": [ + 469, + 143, + 514, + 148 + ], + "type": "text", + "content": "Rich Text Gen" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 170, + 149, + 263, + 220 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 170, + 149, + 263, + 220 + ], + "spans": [ + { + "bbox": [ + 170, + 149, + 263, + 220 + ], + "type": "image", + "image_path": "3b3265ecdc6d7fb0a937fbcdde20d1a42d93b6b88a208b74198a04d64c225431.jpg" + } + ] + } + ], + "index": 11, + "type": "text" + }, + { + "bbox": [ + 192, + 220, + 242, + 225 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 192, + 220, + 242, + 225 + ], + "spans": [ + { + "bbox": [ + 192, + 220, + 242, + 225 + ], + "type": "text", + "content": "PartCraft (Ours)" + } + ] + } + ], + "index": 12, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 263, + 149, + 354, + 220 + ], + "blocks": [ + { + "bbox": [ + 263, + 149, + 354, + 220 + ], + "lines": [ + { + "bbox": [ + 263, + 149, + 354, + 220 + ], + "spans": [ + { + "bbox": [ + 263, + 149, + 354, + 220 + ], + "type": "image", + "image_path": "b1e8a37755145c7331dfacbe8bcf9ffa4c66b0305a1b7e56c5fa86b8ff174ab3.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 284, + 220, + 332, + 225 + ], + "lines": [ + { + "bbox": [ + 284, + 220, + 332, + 225 + ], + "spans": [ + { + "bbox": [ + 284, + 220, + 332, + 225 + ], + "type": "text", + "content": "StableDiffusion" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 354, + 149, + 446, + 220 + ], + "blocks": [ + { + "bbox": [ + 354, + 149, + 446, + 220 + ], + "lines": [ + { + "bbox": [ + 354, + 149, + 446, + 220 + ], + "spans": [ + { + "bbox": [ + 354, + 149, + 446, + 220 + ], + "type": "image", + "image_path": "e78ad06de85adcbee3536b5f973a50d316cf659e4f6044ea31a982e8d677b149.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 377, + 220, + 424, + 225 + ], + "lines": [ + { + "bbox": [ + 377, + 220, + 424, + 225 + ], + "spans": [ + { + "bbox": [ + 377, + 220, + 424, + 225 + ], + "type": "text", + "content": "InstructPix2Pix" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 446, + 149, + 538, + 220 + ], + "blocks": [ + { + "bbox": [ + 446, + 149, + 538, + 220 + ], + "lines": [ + { + "bbox": [ + 446, + 149, + 538, + 220 + ], + "spans": [ + { + "bbox": [ + 446, + 149, + 538, + 220 + ], + "type": "image", + "image_path": "b61359ccfa2c89ebe844abc616efb258d0b87056856bd038e04cfd9a05089aa0.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 470, + 220, + 514, + 225 + ], + "lines": [ + { + "bbox": [ + 470, + 220, + 514, + 225 + ], + "spans": [ + { + "bbox": [ + 470, + 220, + 514, + 225 + ], + "type": "text", + "content": "Rich Text Gen" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "bbox": [ + 56, + 271, + 296, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 271, + 296, + 534 + ], + "spans": [ + { + "bbox": [ + 56, + 271, + 296, + 534 + ], + "type": "text", + "content": "Results. We tabulate the results for all the baselines in Table 1. Our approach, PartComposer, significantly improves over the baselines by " + }, + { + "bbox": [ + 56, + 271, + 296, + 534 + ], + "type": "inline_equation", + "content": "\\geq 5" + }, + { + "bbox": [ + 56, + 271, + 296, + 534 + ], + "type": "text", + "content": " points in NMI and ARI across most metrics in the Unsupervised Zero-Shot setting. Further, in some cases our methods, NMI and ARI, are near the unsupervised learning approaches. This demonstrates that non-trivial part masks can be generated by harnessing the power of the pre-trained diffusion models, and they can serve as a good initial prior for learning part segmentation approaches by using them as base pre-trained models. We provide qualitative results for PartComposer, where we observe that PartComposer can associate the right part with the correct label (Fig 4). On the contrary, in the StableDiffusion baseline, the parts often get assigned to the wrong part tokens, which is the cause of degraded performance (Suppl. Fig. 16). We also observe that in the case of our algorithm, if the part gets localized based on the max condition defined in the segmentation algorithm, the part mask found is usually in the right region (See Fig. 4 and Table 2). This demonstrates the effectiveness of our part segmentation procedure, which often sides with not localizing objects rather than performing arbitrary assignments." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 55, + 540, + 288, + 553 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 540, + 288, + 553 + ], + "spans": [ + { + "bbox": [ + 55, + 540, + 288, + 553 + ], + "type": "text", + "content": "4.2. PartComposer Image Generation Evaluation" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 55, + 558, + 296, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 296, + 653 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 296, + 653 + ], + "type": "text", + "content": "We now evaluate the object composition ability of the proposed approach, PartComposer, compared to the baselines. In this section, we use the base generation model as a StableDiffusion (SD) 1.5 model to do a fair comparison with the Rich-Text baseline. We used the same generation setting of a PNDM [28] scheduler with 41 steps and suggested classifier guidance 8.5. To perform a fair comparison, we keep all the parameters same as that of the Rich-Text [18] baseline." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "content": "Baselines. For the task of generating the object image based on part-level details of the specified object, we use the strong baselines of Rich-Text Generation and InstructPix2Pix [8]. The Rich-Text Generation method generates the base image itself, whereas the InstructPix2Pix method requires us to" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 271, + 555, + 391 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 271, + 555, + 391 + ], + "spans": [ + { + "bbox": [ + 313, + 271, + 555, + 391 + ], + "type": "text", + "content": "generate the base image. We also evaluate the standard StableDiffusion baseline in which we add all the part details in the text prompt to generate the desired image. We compare all these baselines with our proposed method, PartComposer, while ensuring that the base parameters and seed are the same for the base image generation. A recent work [33] regarding part generation operates in a fully supervised setup using part masks and cannot be used to create novel parts based on text instruction, making it incomparable. We defer specific details for the baselines to the Suppl. Section." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 396, + 556, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 396, + 556, + 612 + ], + "spans": [ + { + "bbox": [ + 313, + 396, + 556, + 612 + ], + "type": "text", + "content": "Visual Comparison. We provide a visual comparison for the baselines for the a) bird image distinctive part generation and b) human image part generations. Across both cases (Fig. 5), our proposed method, PartComposer, only modifies the desired beak region and replaces it with the iconic pelican beak. The StableDiffusion baseline modifies the image completely while ignoring instructions (Supp. Fig. 12). In the Rich-Text generation, the full region corresponding to the bird gets modified instead of only the specified prompt in base generation. For the editing-based InstructPix2Pix method, we observe that it modifies the base image at a global level and cannot localize modification to the desired part region (Fig. 5). We further compare and find that recent models like SD3.5, Inpainting [17] and recent SotA editing methods [15, 26, 47, 54] can still not follow exact part instructions as seen in Supp. Fig. 8, 9 and 10. In contrast, observe that PartComposer-generated images are novel aesthetic compositions that follow the part-level instructions." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 618, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 618, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 618, + 556, + 714 + ], + "type": "text", + "content": "Quantitative Evaluation. As the part-specific generation is fine-grained, we perform the both automatic and user-study evaluations. We performed a user study by inviting 50 participants (28 responded with a full survey). We evaluate the model by asking questions and evaluating metrics on three orthogonal aspects i.e., a) Localization of part generation in comparison to the base image (through LPIPS [56]), b) Text-Consistency of generated parts by measuring CLIP" + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13215" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 91, + 282, + 251 + ], + "blocks": [ + { + "bbox": [ + 55, + 65, + 297, + 87 + ], + "lines": [ + { + "bbox": [ + 55, + 65, + 297, + 87 + ], + "spans": [ + { + "bbox": [ + 55, + 65, + 297, + 87 + ], + "type": "text", + "content": "Table 2. Ablation Analysis of PartComposer for Part Localization, showing performance improvement with each component." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 91, + 282, + 251 + ], + "lines": [ + { + "bbox": [ + 56, + 91, + 282, + 251 + ], + "spans": [ + { + "bbox": [ + 56, + 91, + 282, + 251 + ], + "type": "table", + "html": "
PartComposer (Ablations)
MethodFG-NMIFG-ARI
PartComposer35.411.0
w/o Null-Text Inversion23.15.2
w/o Max Localization21.32.8
w/o Dot Product Localization23.75.0
w/o Independent Text31.28.9
PartComposer (Clustering)
PartComposer (K = 9)35.411.0
PartComposer (K = 4)20.42.8
PartComposer (K = 14)35.210.3
", + "image_path": "94368306424c8d25c29f0a7b17b893f031f03430a82874d82f08b48be672d12d.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 58, + 261, + 113, + 390 + ], + "blocks": [ + { + "bbox": [ + 58, + 261, + 113, + 390 + ], + "lines": [ + { + "bbox": [ + 58, + 261, + 113, + 390 + ], + "spans": [ + { + "bbox": [ + 58, + 261, + 113, + 390 + ], + "type": "image", + "image_path": "c7feb96f76df24e61954e428d9fcccf9414cd4b7cc0d2dd95e06b98b5ef2a6dc.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 399, + 296, + 444 + ], + "lines": [ + { + "bbox": [ + 55, + 399, + 296, + 444 + ], + "spans": [ + { + "bbox": [ + 55, + 399, + 296, + 444 + ], + "type": "text", + "content": "Figure 6. PartComposer generalizes over domains as shown by Claude Monet's painting style (above), where we specify the dress part of women to follow styles like Van Gogh and Ukiyo-e. We perform similar modifications to Woodie's image." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 120, + 261, + 176, + 390 + ], + "blocks": [ + { + "bbox": [ + 120, + 261, + 176, + 390 + ], + "lines": [ + { + "bbox": [ + 120, + 261, + 176, + 390 + ], + "spans": [ + { + "bbox": [ + 120, + 261, + 176, + 390 + ], + "type": "image", + "image_path": "9ba083cb1b50e0235a03d773b19f8abc2c86af8756b184a6179b5e24e63ac170.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 179, + 261, + 235, + 389 + ], + "blocks": [ + { + "bbox": [ + 179, + 261, + 235, + 389 + ], + "lines": [ + { + "bbox": [ + 179, + 261, + 235, + 389 + ], + "spans": [ + { + "bbox": [ + 179, + 261, + 235, + 389 + ], + "type": "image", + "image_path": "98a4c4e120dbae11deb0e00f5cbab1350b8b9e297287218afe3a75b67614aaf1.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 239, + 261, + 294, + 389 + ], + "blocks": [ + { + "bbox": [ + 239, + 261, + 294, + 389 + ], + "lines": [ + { + "bbox": [ + 239, + 261, + 294, + 389 + ], + "spans": [ + { + "bbox": [ + 239, + 261, + 294, + 389 + ], + "type": "image", + "image_path": "2a89e87e5a288a024d70157572874c1490e479d558bc4da45bdc22ac58433b55.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 453, + 296, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 453, + 296, + 583 + ], + "spans": [ + { + "bbox": [ + 55, + 453, + 296, + 583 + ], + "type": "text", + "content": "similarity of the part detailed text prompt with the generated image and c) Aesthetic Quality of the generated image through LAION-5B [42] CLIP Aesthetic scorer. In total, we collect about 3.5k opinions. In Fig. 7, we summarize the results in which we observe that PartComposer is significantly preferred over the other baselines in terms of localized and consistent Part Generation. The automated evaluation results are provided in the table below, which also follow a similar trend as in the user study, demonstrating the effectiveness of PartComposer in a localized part generation while ensuring the aesthetics score is similar to that of the base model." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 597, + 194, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 597, + 194, + 609 + ], + "spans": [ + { + "bbox": [ + 55, + 597, + 194, + 609 + ], + "type": "text", + "content": "5. Analysis and Discussion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "text", + "content": "Part Diffusion Based Localization. We ablate the components we have introduced in the PartDiffusion process of PartComposer (Sec. 3.2). We provide an analysis of the effect of using a) null-text inversion, b) max-based localization c) usage of dot product-based protocol in the part assignment, and d) independent text embeddings. We have used a subset of CUB-200 images to perform all evaluations, which is kept fixed across ablations. We tabulate the ablations in" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 319, + 95, + 556, + 269 + ], + "blocks": [ + { + "bbox": [ + 313, + 71, + 555, + 92 + ], + "lines": [ + { + "bbox": [ + 313, + 71, + 555, + 92 + ], + "spans": [ + { + "bbox": [ + 313, + 71, + 555, + 92 + ], + "type": "text", + "content": "Figure 7. User Study and Quantitative Results for part-based image generation baselines." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 319, + 95, + 556, + 269 + ], + "lines": [ + { + "bbox": [ + 319, + 95, + 556, + 269 + ], + "spans": [ + { + "bbox": [ + 319, + 95, + 556, + 269 + ], + "type": "image", + "image_path": "850689bb6bce080f702d0358c762bd3ce0aca0b801107d68eb43f6972bb75596.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 315, + 274, + 553, + 312 + ], + "blocks": [ + { + "bbox": [ + 315, + 274, + 553, + 312 + ], + "lines": [ + { + "bbox": [ + 315, + 274, + 553, + 312 + ], + "spans": [ + { + "bbox": [ + 315, + 274, + 553, + 312 + ], + "type": "table", + "html": "
PartComposerStableDiffusionInstructPix2PixRich-Text
LPIPS0.1680.4670.1890.243
CLIP0.2010.1830.1930.187
Aesthetic5.665.685.635.65
", + "image_path": "a04a5a598777fc64d126427c0550dfe94bf47b6bca83ee1c177fed057a0f6c4e.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 324, + 555, + 348 + ], + "lines": [ + { + "bbox": [ + 313, + 324, + 555, + 348 + ], + "spans": [ + { + "bbox": [ + 313, + 324, + 555, + 348 + ], + "type": "text", + "content": "Table 2. We observe that all the components introduced in PartComposer contribute significantly to the performance." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 352, + 556, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 352, + 556, + 497 + ], + "spans": [ + { + "bbox": [ + 313, + 352, + 556, + 497 + ], + "type": "text", + "content": "Generalization of PartComposer. In Fig. 6, above a painting in Claude Monet style generated by the base StableDiffusion method. We then use PartComposer to specify the dress part of the women's token, to dark green color in Monet Style, orange color dress in Van Gogh Style (middle), and green color dress in Ukiyo-e style. In Fig. 6, we provide results for the 'Woody' character from Toy Story. These results show that PartComposer can generate shirt, pants, and dress color variations, leading to aesthetic image combinations in synthetic and natural (Supp. Fig. 13) domains. This zero-shot generalization across domains demonstrates the creative activities that can be enabled with PartComposer." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 521, + 388, + 533 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 388, + 533 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 388, + 533 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 546, + 556, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 546, + 556, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 546, + 556, + 653 + ], + "type": "text", + "content": "In this work we introduce PartComposer, a method to generate object images with fine-grained attribute details specified at the part level, using an expressive Rich-Text interface. The PartComposer method introduces a novel part diffusion process, which is responsible for denoising objects using the part features, and then utilizes a region-specific diffusion process to generate part details and compose the final image. PartComposer serves as initial work enabling rich-text-based training-free part-level control for SD models." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 654, + 556, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 556, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 556, + 712 + ], + "type": "text", + "content": "Limitations. We find that part generation is bottlenecked with the part understanding. If the part is localized correctly, the text-to-image model can generate specified details in PartComposer. Hence, improving the part-level localization of these models is a good direction for future works." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "13216" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "type": "text", + "content": "[1] Rameen Abdal, Peihao Zhu, Niloy J Mitra, and Peter Wonka. Labels4free: Unsupervised segmentation using stylegan. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13970-13979, 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 137, + 296, + 179 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 137, + 296, + 179 + ], + "spans": [ + { + "bbox": [ + 61, + 137, + 296, + 179 + ], + "type": "text", + "content": "[2] Aishwarya Agarwal, Srikrishna Karanam, K J Joseph, Apoorv Saxena, Koustava Goswami, and Balaji Vasan Srinivasan. Astar: Test-time attention segregation and retention for text-to-image synthesis, 2023. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 182, + 295, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 182, + 295, + 224 + ], + "spans": [ + { + "bbox": [ + 62, + 182, + 295, + 224 + ], + "type": "text", + "content": "[3] Aishwarya Agarwal, Srikrishna Karanam, and Balaji Vasan Srinivasan. Training-free color-style disentanglement for constrained text-to-image synthesis. arXiv preprint arXiv:2409.02429, 2024." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 227, + 296, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 227, + 296, + 270 + ], + "spans": [ + { + "bbox": [ + 62, + 227, + 296, + 270 + ], + "type": "text", + "content": "[4] Yuval Alaluf, Daniel Garibi, Or Patashnik, Hadar Averbuch-Elor, and Daniel Cohen-Or. Cross-image attention for zero-shot appearance transfer. arXiv preprint arXiv:2311.03335, 2023. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 272, + 296, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 272, + 296, + 304 + ], + "spans": [ + { + "bbox": [ + 62, + 272, + 296, + 304 + ], + "type": "text", + "content": "[5] Omer Bar-Tal, Lior Yariv, Yaron Lipman, and Tali Dekel. Multidiffusion: Fusing diffusion paths for controlled image generation. 2023. 1, 2, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 306, + 296, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 306, + 296, + 338 + ], + "spans": [ + { + "bbox": [ + 62, + 306, + 296, + 338 + ], + "type": "text", + "content": "[6] Shariq Farooq Bhat, Niloy J. Mitra, and Peter Wonka. Loosecontrol: Lifting controlnet for generalized depth conditioning, 2023. 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 341, + 296, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 341, + 296, + 384 + ], + "spans": [ + { + "bbox": [ + 62, + 341, + 296, + 384 + ], + "type": "text", + "content": "[7] Anand Bhattachad, Daniel McKee, Derek Hoiem, and David Forsyth. Stylegan knows normal, depth, albedo, and more. Advances in Neural Information Processing Systems, 36, 2024. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 386, + 296, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 386, + 296, + 418 + ], + "spans": [ + { + "bbox": [ + 62, + 386, + 296, + 418 + ], + "type": "text", + "content": "[8] Tim Brooks, Aleksander Holynski, and Alexei A. Efros. Instructpix2pix: Learning to follow image editing instructions. In CVPR, 2023. 7, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 420, + 296, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 420, + 296, + 464 + ], + "spans": [ + { + "bbox": [ + 62, + 420, + 296, + 464 + ], + "type": "text", + "content": "[9] Hila Chefer, Yuval Alaluf, Yael Vinker, Lior Wolf, and Daniel Cohen-Or. Attend-and-excite: Attention-based semantic guidance for text-to-image diffusion models. ACM Transactions on Graphics (TOG), 42(4):1-10, 2023. 1, 2, 3, 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 465, + 296, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 465, + 296, + 509 + ], + "spans": [ + { + "bbox": [ + 57, + 465, + 296, + 509 + ], + "type": "text", + "content": "[10] Minghao Chen, Iro Laina, and Andrea Vedaldi. Training-free layout control with cross-attention guidance. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 5343–5353, 2024. 1, 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 510, + 296, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 510, + 296, + 565 + ], + "spans": [ + { + "bbox": [ + 57, + 510, + 296, + 565 + ], + "type": "text", + "content": "[11] Jaehoon Choi, Taekyung Kim, and Changick Kim. Self-ensembling with gan-based data augmentation for domain adaptation in semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6830–6840, 2019. 2, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 567, + 296, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 567, + 296, + 610 + ], + "spans": [ + { + "bbox": [ + 57, + 567, + 296, + 610 + ], + "type": "text", + "content": "[12] Subhabrata Choudhury, Iro Laina, Christian Rupprecht, and Andrea Vedaldi. Unsupervised part discovery from contrastive reconstruction. Advances in Neural Information Processing Systems, 34:28104-28118, 2021. 2, 3, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 612, + 296, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 612, + 296, + 655 + ], + "spans": [ + { + "bbox": [ + 57, + 612, + 296, + 655 + ], + "type": "text", + "content": "[13] Edo Collins, Radhakrishna Achanta, and Sabine Susstrunk. Deep feature factorization for concept discovery. In Proceedings of the European Conference on Computer Vision (ECCV), pages 336-352, 2018. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 657, + 296, + 700 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 657, + 296, + 700 + ], + "spans": [ + { + "bbox": [ + 57, + 657, + 296, + 700 + ], + "type": "text", + "content": "[14] Omer Dahary, Or Patashnik, Kfir Aberman, and Daniel Cohen-Or. Be yourself: Bounded attention for multi-subject text-to-image generation. arXiv preprint arXiv:2403.16990, 2(5), 2024. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 702, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 702, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 702, + 296, + 713 + ], + "type": "text", + "content": "[15] Gilad Deutch, Rinon Gal, Daniel Garibi, Or Patashnik, and" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 317, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "text", + "content": "Daniel Cohen-Or. Turboedit: Text-based image editing using few-step diffusion models, 2024. 7, 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 96, + 553, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 96, + 553, + 128 + ], + "spans": [ + { + "bbox": [ + 317, + 96, + 553, + 128 + ], + "type": "text", + "content": "[16] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 129, + 555, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 129, + 555, + 184 + ], + "spans": [ + { + "bbox": [ + 317, + 129, + 555, + 184 + ], + "type": "text", + "content": "[17] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In *Forty-first International Conference on Machine Learning*, 2024. 7, 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 185, + 555, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 185, + 555, + 227 + ], + "spans": [ + { + "bbox": [ + 317, + 185, + 555, + 227 + ], + "type": "text", + "content": "[18] Songwei Ge, Taesung Park, Jun-Yan Zhu, and Jia-Bin Huang. Expressive text-to-image generation with rich text. In IEEE International Conference on Computer Vision (ICCV), 2023. 1, 2, 3, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 228, + 555, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 228, + 555, + 271 + ], + "spans": [ + { + "bbox": [ + 317, + 228, + 555, + 271 + ], + "type": "text", + "content": "[19] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 272, + 555, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 272, + 555, + 327 + ], + "spans": [ + { + "bbox": [ + 317, + 272, + 555, + 327 + ], + "type": "text", + "content": "[20] Jonathan Ho, William Chan, Chitwan Sahara, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 327, + 555, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 327, + 555, + 381 + ], + "spans": [ + { + "bbox": [ + 317, + 327, + 555, + 381 + ], + "type": "text", + "content": "[21] Wei-Chih Hung, Varun Jampani, Sifei Liu, Pavlo Molchanov, Ming-Hsuan Yang, and Jan Kautz. Scops: Self-supervised co-part segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 869-878, 2019. 2, 3, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 383, + 555, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 383, + 555, + 437 + ], + "spans": [ + { + "bbox": [ + 317, + 383, + 555, + 437 + ], + "type": "text", + "content": "[22] Minguk Kang, Jun-Yan Zhu, Richard Zhang, Jaesik Park, Eli Shechtman, Sylvain Paris, and Taesung Park. Scaling up gans for text-to-image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10124-10134, 2023. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 438, + 555, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 438, + 555, + 492 + ], + "spans": [ + { + "bbox": [ + 317, + 438, + 555, + 492 + ], + "type": "text", + "content": "[23] Hyung-Kwon Ko, Gwanmo Park, Hyeon Jeon, Jaemin Jo, Juho Kim, and Jinwook Seo. Large-scale text-to-image generation models for visual artists' creative works. In Proceedings of the 28th International Conference on Intelligent User Interfaces, pages 919-933, 2023. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 493, + 555, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 493, + 555, + 536 + ], + "spans": [ + { + "bbox": [ + 317, + 493, + 555, + 536 + ], + "type": "text", + "content": "[24] Svetlana Lazebnik, Cordelia Schmid, and Jean Ponce. Semi-local affine parts for object recognition. In *British Machine Vision Conference*, BMVC 2004, Kingston, UK, September 7-9, 2004. Proceedings, pages 1-10. BMVA Press, 2004. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 537, + 555, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 537, + 555, + 581 + ], + "spans": [ + { + "bbox": [ + 317, + 537, + 555, + 581 + ], + "type": "text", + "content": "[25] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip2: Bootstrapping language-image pre-training with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597, 2023. 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 582, + 555, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 582, + 555, + 635 + ], + "spans": [ + { + "bbox": [ + 317, + 582, + 555, + 635 + ], + "type": "text", + "content": "[26] Shanglin Li, Bohan Zeng, Yutang Feng, Sicheng Gao, Xiuhui Liu, Jiaming Liu, Lin Li, Xu Tang, Yao Hu, Jianzhuang Liu, et al. Zone: Zero-shot instruction-guided local editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6254-6263, 2024. 7, 1" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 317, + 636, + 555, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 636, + 555, + 691 + ], + "spans": [ + { + "bbox": [ + 317, + 636, + 555, + 691 + ], + "type": "text", + "content": "[27] Yuheng Li, Haotian Liu, Qingyang Wu, Fangzhou Mu, Jianwei Yang, Jianfeng Gao, Chunyuan Li, and Yong Jae Lee. Gligen: Open-set grounded text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22511-22521, 2023. 1, 2, 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 317, + 692, + 555, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 692, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 692, + 555, + 713 + ], + "type": "text", + "content": "[28] Luping Liu, Yi Ren, Zhijie Lin, and Zhou Zhao. Pseudo numerical methods for diffusion models on manifolds. In" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13217" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 76, + 72, + 296, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 296, + 94 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 296, + 94 + ], + "type": "text", + "content": "International Conference on Learning Representations, 2022. 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 296, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 296, + 151 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 296, + 151 + ], + "type": "text", + "content": "[29] Ziwei Liu, Ping Luo, Shi Qiu, Xiaogang Wang, and Xiaou Tang. Deepfashion: Powering robust clothes recognition and retrieval with rich annotations. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 2, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 296, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 296, + 195 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 296, + 195 + ], + "type": "text", + "content": "[30] Wan-Duo Kurt Ma, JP Lewis, W Bastiaan Kleijn, and Thomas Leung. Directed diffusion: Direct control of object placement through attention guidance. arXiv preprint arXiv:2302.13153, 2023. 2, 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 198, + 296, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 198, + 296, + 241 + ], + "spans": [ + { + "bbox": [ + 56, + 198, + 296, + 241 + ], + "type": "text", + "content": "[31] Subhransu Maji, Esa Rahtu, Juho Kannala, Matthew Blaschko, and Andrea Vedaldi. Fine-grained visual classification of aircraft. arXiv preprint arXiv:1306.5151, 2013. 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 243, + 296, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 243, + 296, + 297 + ], + "spans": [ + { + "bbox": [ + 56, + 243, + 296, + 297 + ], + "type": "text", + "content": "[32] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6038–6047, 2023. 6, 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 299, + 296, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 299, + 296, + 331 + ], + "spans": [ + { + "bbox": [ + 56, + 299, + 296, + 331 + ], + "type": "text", + "content": "[33] Kam Woh Ng, Xiatian Zhu, Yi-Zhe Song, and Tao Xiang. Partcraft: Crafting creative objects by parts. arXiv preprint arXiv:2407.04604, 2024. 3, 7, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 333, + 296, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 333, + 296, + 388 + ], + "spans": [ + { + "bbox": [ + 56, + 333, + 296, + 388 + ], + "type": "text", + "content": "[34] Rishubh Parihar, Abhijnya Bhat, Abhipsa Basu, Saswat Mallick, Jogendra Nath Kundu, and R Venkatesh Babu. Balancing act: distribution-guided debiasing in diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6668-6678, 2024. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 389, + 296, + 433 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 389, + 296, + 433 + ], + "spans": [ + { + "bbox": [ + 56, + 389, + 296, + 433 + ], + "type": "text", + "content": "[35] Rishubh Parihar, Harsh Gupta, Sachidanand VS, and R Venkatesh Babu. Text2place: Affordance-aware text guided human placement. In European Conference on Computer Vision, pages 57-77. Springer, 2024. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 434, + 296, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 434, + 296, + 489 + ], + "spans": [ + { + "bbox": [ + 56, + 434, + 296, + 489 + ], + "type": "text", + "content": "[36] Rishubh Parihar, VS Sachidanand, Sabariswaran Mani, Tejan Karmali, and R Venkatesh Babu. Precisecontrol: Enhancing text-to-image diffusion models with fine-grained attribute control. In European Conference on Computer Vision, pages 469-487. Springer, 2024. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 491, + 296, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 491, + 296, + 544 + ], + "spans": [ + { + "bbox": [ + 56, + 491, + 296, + 544 + ], + "type": "text", + "content": "[37] Or Patashnik, Daniel Garibi, Idan Azuri, Hadar Averbuch-Elor, and Daniel Cohen-Or. Localizing object-level shape variations with text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 4, 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 546, + 296, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 546, + 296, + 600 + ], + "spans": [ + { + "bbox": [ + 56, + 546, + 296, + 600 + ], + "type": "text", + "content": "[38] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. 1, 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 602, + 296, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 602, + 296, + 656 + ], + "spans": [ + { + "bbox": [ + 56, + 602, + 296, + 656 + ], + "type": "text", + "content": "[39] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, et al. Sam 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714, 2024. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 658, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 296, + 713 + ], + "type": "text", + "content": "[40] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 715 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 316, + 73, + 555, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 555, + 139 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 555, + 139 + ], + "type": "text", + "content": "[41] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural information processing systems, 35:36479-36494, 2022. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 141, + 555, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 555, + 205 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 555, + 205 + ], + "type": "text", + "content": "[42] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278-25294, 2022. 2, 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 207, + 555, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 207, + 555, + 239 + ], + "spans": [ + { + "bbox": [ + 316, + 207, + 555, + 239 + ], + "type": "text", + "content": "[43] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2105.05233, 2021. 6, 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 241, + 555, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 241, + 555, + 295 + ], + "spans": [ + { + "bbox": [ + 316, + 241, + 555, + 295 + ], + "type": "text", + "content": "[44] Raphael Tang, Linqing Liu, Akshit Pandey, Zhiying Jiang, Gefei Yang, Karun Kumar, Pontus Stenetorp, Jimmy Lin, and Ferhan Ture. What the daam: Interpreting stable diffusion using cross attention. arXiv preprint arXiv:2210.04885, 2022. 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 297, + 555, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 297, + 555, + 341 + ], + "spans": [ + { + "bbox": [ + 316, + 297, + 555, + 341 + ], + "type": "text", + "content": "[45] Yoad Tewel, Omri Kaduri, Rinon Gal, Yoni Kasten, Lior Wolf, Gal Chechik, and Yuval Atzmon. Training-free consistent text-to-image generation. arXiv preprint arXiv:2402.03286, 2024.3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 342, + 555, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 342, + 555, + 387 + ], + "spans": [ + { + "bbox": [ + 316, + 342, + 555, + 387 + ], + "type": "text", + "content": "[46] Junjiao Tian, Lavisha Aggarwal, Andrea Colaco, Zsolt Kira, and Mar Gonzalez-Franco. Diffuse, attend, and segment: Unsupervised zero-shot segmentation using stable diffusion. arXiv preprint arXiv:2308.12469, 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 388, + 555, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 388, + 555, + 442 + ], + "spans": [ + { + "bbox": [ + 316, + 388, + 555, + 442 + ], + "type": "text", + "content": "[47] Narek Tumanyan, Michal Geyer, Shai Bagon, and Tali Dekel. Plug-and-play diffusion features for text-driven image-to-image translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1921-1930, 2023. 5, 7, 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 444, + 555, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 444, + 555, + 521 + ], + "spans": [ + { + "bbox": [ + 316, + 444, + 555, + 521 + ], + "type": "text", + "content": "[48] Andrea Vedaldi, Siddharth Mahendran, Stavros Tsogkas, Subhransu Maji, Ross Girshick, Juho Kannala, Esa Rahtu, Iasonas Kokkinos, Matthew B. Blaschko, David Weiss, Ben Taskar, Karen Simonyan, Naomi Saphra, and Sammy Mohamed. Understanding objects in detail with fine-grained attributes. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2014. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 522, + 554, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 522, + 554, + 555 + ], + "spans": [ + { + "bbox": [ + 316, + 522, + 554, + 555 + ], + "type": "text", + "content": "[49] Andrey Voynov, Qinghao Chu, Daniel Cohen-Or, and Kfir Aberman. " + }, + { + "bbox": [ + 316, + 522, + 554, + 555 + ], + "type": "inline_equation", + "content": "p+" + }, + { + "bbox": [ + 316, + 522, + 554, + 555 + ], + "type": "text", + "content": ": Extended textual conditioning in text-to-image generation. arXiv preprint arXiv:2303.09522, 2023. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 556, + 555, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 556, + 555, + 599 + ], + "spans": [ + { + "bbox": [ + 316, + 556, + 555, + 599 + ], + "type": "text", + "content": "[50] C. Wah, S. Branson, P. Welinder, P. Perona, and S. Belongie. Caltech-ucsd birds-200-2011 dataset. Technical Report CNSTR-2011-001, California Institute of Technology, 2011. 2, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 601, + 555, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 601, + 555, + 644 + ], + "spans": [ + { + "bbox": [ + 316, + 601, + 555, + 644 + ], + "type": "text", + "content": "[51] Zirui Wang, Zhizhou Sha, Zheng Ding, Yilin Wang, and Zhuowen Tu. Tokencompose: Grounding diffusion with token-level supervision. arXiv preprint arXiv:2312.03626, 2023.3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 647, + 554, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 554, + 690 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 554, + 690 + ], + "type": "text", + "content": "[52] Jiarui Xu, Sifei Liu, Arash Vahdat, Wonmin Byeon, Xiaolong Wang, and Shalini De Mello. Open-Vocabulary Panoptic Segmentation with Text-to-Image Diffusion Models. arXiv preprint arXiv:2303.04803, 2023. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 692, + 555, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 555, + 715 + ], + "type": "text", + "content": "[53] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "13218" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 251 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 75, + 72, + 296, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 72, + 296, + 106 + ], + "spans": [ + { + "bbox": [ + 75, + 72, + 296, + 106 + ], + "type": "text", + "content": "Yang, Burcu Karagol Ayan, et al. Scaling autoregressive models for content-rich text-to-image generation. arXiv preprint arXiv:2206.10789, 2(3):5, 2022. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 296, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 296, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 296, + 150 + ], + "type": "text", + "content": "[54] Kai Zhang, Lingbo Mo, Wenhu Chen, Huan Sun, and Yu Su. Magicbrush: A manually annotated dataset for instruction-guided image editing. Advances in Neural Information Processing Systems, 36:31428-31449, 2023. 7, 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 294, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 294, + 196 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 294, + 196 + ], + "type": "text", + "content": "[55] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 1, 2, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 198, + 294, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 198, + 294, + 251 + ], + "spans": [ + { + "bbox": [ + 56, + 198, + 294, + 251 + ], + "type": "text", + "content": "[56] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 7" + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "13219" + } + ] + } + ], + "index": 5 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Compositional Caching for Training-free Open-vocabulary Attribute Detection/5997b1e0-b306-408f-91ce-2ab833ab2009_content_list.json b/2025/Compositional Caching for Training-free Open-vocabulary Attribute Detection/5997b1e0-b306-408f-91ce-2ab833ab2009_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..e20e2f0b88b9e3e10209dd4eee6cea22e23e24a5 --- /dev/null +++ b/2025/Compositional Caching for Training-free Open-vocabulary Attribute Detection/5997b1e0-b306-408f-91ce-2ab833ab2009_content_list.json @@ -0,0 +1,1626 @@ +[ + { + "type": "text", + "text": "Compositional Caching for Training-free Open-vocabulary Attribute Detection", + "text_level": 1, + "bbox": [ + 101, + 130, + 893, + 151 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Marco Garosi1 Alessandro Conti1 Gaowen Liu2 Elisa Ricci1,3 Massimiliano Mancini1 \n1University of Trento 2Cisco Research Fondazione Bruno Kessler \nhttps://comca-attributes. github.io", + "bbox": [ + 130, + 179, + 861, + 232 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 268, + 325, + 282 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Attribute detection is crucial for many computer vision tasks, as it enables systems to describe properties such as color, texture, and material. Current approaches often rely on labor-intensive annotation processes which are inherently limited: objects can be described at an arbitrary level of detail (e.g., color vs. color shades), leading to ambiguities when the annotators are not instructed carefully. Furthermore, they operate within a predefined set of attributes, reducing scalability and adaptability to unforeseen downstream applications. We present COMPOSITIONAL CACHING (COMCA), a training-free method for open-vocabulary attribute detection that overcomes these constraints. COMCA requires only the list of target attributes and objects as input, using them to populate an auxiliary cache of images by leveraging web-scale databases and Large Language Models to determine attribute-object compatibility. To account for the compositional nature of attributes, cache images receive soft attribute labels. Those are aggregated at inference time based on the similarity between the input and cache images, refining the predictions of underlying Vision-Language Models (VLMs). Importantly, our approach is model-agnostic, compatible with various VLMs. Experiments on public datasets demonstrate that COMCA significantly outperforms zero-shot and cache-based baselines, competing with recent training-based methods, proving that a carefully designed training-free approach can successfully address open-vocabulary attribute detection.", + "bbox": [ + 89, + 300, + 485, + 708 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 739, + 220, + 753 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Attributes shape our perception and interactions with the environment. Let us consider the example of Fig. 1. If our aim is to identify the animal, we may focus on its fur color, size, the presence of horns and the tail. Similarly, if we want to analyze its health status, we may focus on its age, fur patterns and shades, and its skin. This level of analysis requires models with fine-grained understanding of the visual input, which is usually obtained by training a model on datasets that annotate both objects and their corresponding", + "bbox": [ + 89, + 763, + 483, + 900 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/db59c2c80abbb2e7c48cbb7b17ea624025591c2b0ee627a3ce4681338528f924.jpg", + "image_caption": [ + "Figure 1. Example of attribute annotations. Attribute annotations are often: sparse, as they are not consistent across samples; incomplete, as not all attributes are annotated; ambiguous, as they can be subjective or miss a frame of reference. This makes open-vocabulary attribute detection challenging. Differently from previous works [5, 7], we do not rely on such annotations and propose COMCA, a training-free approach requiring no supervision." + ], + "image_footnote": [], + "bbox": [ + 517, + 268, + 903, + 415 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "attributes [38, 40]. However, collecting attribute annotations is both time-consuming and prone to ambiguities, especially when annotators lack clear instructions. Additionally, attribute annotations are often tailored to specific datasets and are ambiguous. For instance, in Fig. 1, one may list \"brown\" as a color for the left goat, but others may argue that it is a red-ish shade of brown. A simple solution is to provide annotators with a predefined list of attributes; however, (i) the list may not cover all possible annotations for a property (e.g., all colors) and (ii) it limits models trained on it to the given list itself, making them unable to recognize new attributes.", + "bbox": [ + 511, + 536, + 908, + 702 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent approaches overcome some of the challenges by developing open-vocabulary attribute detection models [5, 7, 16] that detect attributes specified in natural language at inference time. While these approaches generalize beyond the set of seen attributes and objects, their performance is still biased toward the specific dataset they are trained on. Indeed, training-based approaches may show significant degradation in performance in cross-dataset experiments (see Fig. 4).", + "bbox": [ + 511, + 703, + 908, + 824 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this work, we overcome potentially ambiguous and/or limited attribute annotations by introducing COMPOSITIONAL CACHING (COMCA), a training-free open-vocabulary attribute detection model. COMCA employs a Vision-Language Model (VLM) [41] and refines its", + "bbox": [ + 511, + 825, + 908, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "15098", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "predictions by utilizing an auxiliary cache that contains images together with automatically estimated attribute labels.", + "bbox": [ + 89, + 90, + 485, + 121 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Caching for refining CLIP predictions has been introduced in the context of image classification [57]. It requires storing images, either generated or retrieved [48, 57], for each of the classes of interests, together with a corresponding one-hot categorical label. The similarity of the input image with each of the examples is used to weight the labels in the cache and refine the model's prediction. While this approach is effective for object classification, it is prohibitively expensive in the context of attribute detection due to the much larger output space, comprising all possible combinations of attributes and objects. COMCA achieves scalability in caching by following two key compositional principles: (i) not all attributes are universally applicable to all objects, and (ii) every object inherently possesses multiple attributes.", + "bbox": [ + 88, + 125, + 485, + 335 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The first principle is implemented in two steps: attribute-object compatibility estimation and cache pair sampling. In the first phase, we estimate the probability of an attribute appearing with an object using the statistics of available databases (e.g., CC12M [6]) and Large Language Model (LLM) queries. Next, we populate the cache for a given attribute by sampling compatible objects from the estimated distribution and retrieving relevant images from a database. The second principle is implemented by self-labeling the cached elements. Initially, each sample is associated with the single attribute it was sampled for. Since this does not account for other attributes present in the image, we extend attribute labels through soft-labeling, i.e., comparing cache images to text embeddings of each attribute of interest. These scores modify a cache element's contribution to each target attribute, considering their implicit composition. We evaluate COMCA on two publicly available benchmarks, OVAD [5] and VAW [39]. COMCA demonstrates significant improvements over training-free baselines and is competitive with recent training-based approaches, while remaining unaffected by the biases of specific training sets. Notably, COMCA provides a consistent boost in performance regardless of the VLM pre-training and architecture.", + "bbox": [ + 91, + 339, + 486, + 686 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our contributions are: ① We propose COMCA, an approach specifically designed for training-free open-vocabulary attribute detection. COMCA exploits the compositional nature of objects and attributes to build an auxiliary cache that is used as an anchor to refine the model's predictions. ② We make caching scalable by extracting priors from databases and LLMs on which objects are associated with each attribute, using them to populate the cache. Soft-labeling further accounts for the multiplicity of attributes within an image. ③ We empirically show that COMCA outperforms all the cache-based baselines and achieves competitive performance against costly training-based methods, demonstrating the practical feasibility of training-free open vocabulary attribute detection.", + "bbox": [ + 89, + 689, + 485, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 513, + 90, + 651, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Attribute detection. Detecting visual attributes is a widely explored task in computer vision [4, 11-13], supporting various downstream applications such as zero-shot learning [1, 26, 27, 51], relative image comparison [8, 37, 46, 53], and compositional image retrieval [3, 19, 24, 44].", + "bbox": [ + 511, + 114, + 906, + 189 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recent works have introduced several benchmarks to promote and ease research in this area [18, 21, 25, 38-40, 49, 50, 53]. Despite these advances, annotating visual attributes remains a complex and challenging task. This is due to the wide variability in attributes across objects and contexts, as well as the difficulty of ensuring consistent annotations when multiple annotators are involved. Moreover, many datasets capture only a limited range of object attributes, limiting their usefulness for developing general-purpose models.", + "bbox": [ + 511, + 191, + 908, + 325 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Open-vocabulary attribute detection (OVAD) [5, 7] overcomes some of these limitations by detecting user-specified attributes, typically provided in natural language, even those that have not been observed during training. This task was first explored in [5], which introduced the OVAD benchmark and method. OVAD generates supervisory signals for attributes, nouns, and noun phrases by parsing captions and encoding them using CLIP text embeddings [41]. OvarNet [7] builds on this by fine-tuning CLIP with federated learning and additional supervision from external captions. LOWA [16] proposes a three-stage pipeline that aligns visual embeddings to objects, attributes, and free-form text. ArtVLM [63] uses a generative VLM to obtain attribute probabilities based on the input image. Although these methods operate in the open-vocabulary setting, they are training-based and require both a lot of annotated data and several computational resources (see Sec. 4). COMCA significantly differs from these approaches, as we explore for the first time a training-free, cache-based, strategy for attribute detection. Adapting Vision-Language Models. VLMs [28, 29, 31, 41, 54] bridge the gap between textual and visual data, showing remarkable zero-shot generalization abilities. Recent efforts have focused on improving their performance on specific downstream tasks through various strategies, broadly categorized into training-based and training-free methods.", + "bbox": [ + 511, + 327, + 908, + 704 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Training-based approaches typically involve tuning text prompts [22, 47, 52, 59, 62]. Some methods learn prompts directly [60] or predict them [59] from visual inputs to improve generalization. Other approaches train adapters to refine the model's predictions [15, 20]. While effective, these methods introduce additional parameters, which can negatively impact the model's zero-shot performance [59].", + "bbox": [ + 511, + 704, + 908, + 810 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Training-free methods take a different approach, such as incorporating additional knowledge into the prompts using natural language [32, 34, 43], or modifying predictions through attention layers [17] or normalization techniques [61]. Recent works also employ a cache of images and one-hot labels that represent the target classes [48, 57,", + "bbox": [ + 511, + 810, + 910, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "15099", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/ccb2c6f9ec37d0895662b61503a8abb082efad60b0167607ebd2378b1cad974d.jpg", + "image_caption": [ + "Figure 2. COMCA's cache construction. Given a list of attributes and objects, we compute their compatibility from a large database $D_r$ and with an LLM. The scores are merged and normalized to obtain the compatibility distribution, from which we sample cache entries and construct the cache. We enrich the latter with soft labels from the VLM-based similarity between cache images and attributes." + ], + "image_footnote": [], + "bbox": [ + 94, + 88, + 903, + 287 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "[64]. For example, TIP-Adapter [57] computes the similarity between input images and cached examples, using the associated labels for prediction. Similarly, SuS-X [48] creates a cache automatically, either by using a generative model [42] or retrieving from a large database [45], with distance-based scoring to weight the elements' contribution. While our work builds upon the same principles of TIP-Adapter and SuS-X, i.e., using a cache to refine predictions, it differs in four key aspects, as COMCA's cache: (i) is aware of attribute-object compatibility; (ii) accounts for attribute combinations; (iii) models the compositional nature of attributes and objects; (iv) scales efficiently with the number of attributes.", + "bbox": [ + 88, + 349, + 485, + 532 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Compositional Caching", + "text_level": 1, + "bbox": [ + 89, + 547, + 313, + 566 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we formally define the open-vocabulary attribute detection problem and describe caching as a tool for VLM adaptation (Sec. 3.1). Next, we describe how attribute-object compatibility informs the construction of scalable caches (Sec. 3.2). Finally, we detail our process for refining cache labels, taking into account multi-attribute compatibility (Sec. 3.3). Fig. 2 provides an overview of our approach.", + "bbox": [ + 88, + 575, + 485, + 683 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Background", + "text_level": 1, + "bbox": [ + 89, + 694, + 223, + 709 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Problem formulation. Given an image $x \\in \\mathcal{X}$ of an object and a list of attributes, our goal is to detect which attributes are present in the image. Let $\\mathcal{A}$ be the set of all existing attributes in natural language. At inference time, the model is provided with a set of $n$ target attributes $A = \\{a_{1},\\dots ,a_{n}\\}$ as input, where $A \\subset \\mathcal{A}$ . These are associated with $m$ target objects, which we denote as $O = \\{o_1,\\dots ,o_m\\}$ . Our goal is to build a function $f:\\mathcal{X}\\times \\mathcal{A}^n\\to [-1,1]^n$ that maps an image and a list of attributes to a binary vector indicating the presence (1) or absence (-1) of each attribute.", + "bbox": [ + 88, + 717, + 482, + 868 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We obtain $f$ without training, in contrast to previous works that learn $f$ using a dataset $D$ of image-label pairs", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "$(x, \\mathbf{a})$ with $\\mathbf{a} \\in [-1, 1]^n$ being the ground-truth binary vector1. We achieve this by implementing $f$ using a VLM combined with an auxiliary cache to refine its predictions.", + "bbox": [ + 511, + 349, + 906, + 395 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Adapting VLMs with caching. Following previous works [5, 7], we implement $f$ with a VLM. While we primarily consider CLIP [41] to describe our method and perform our evaluation, the proposed framework can integrate several other VLMs (see Tab. 2). The function $f$ has three elements: a visual encoder $f_{v}$ , a text encoder $f_{t}$ , and a prediction function. $f_{v}$ and $f_{t}$ map images in $\\mathcal{X}$ and natural language in $\\mathcal{T}$ , respectively, to a shared $d$ -dimensional multimodal space $\\mathcal{Z} \\in \\mathbb{R}^d$ , i.e., $f_{v}: \\mathcal{X} \\to \\mathcal{Z}$ and $f_{t}: \\mathcal{T} \\to \\mathcal{Z}$ . Next, the prediction function maps embeddings from the multimodal space to the final scores. A standard way for scoring an image $x$ against a class $c$ is by computing the cosine similarity between their embeddings: $^2$", + "bbox": [ + 511, + 396, + 908, + 592 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nf _ {\\mathrm {C L I P}} (x, c) = \\frac {f _ {v} (x) ^ {\\top} f _ {t} (c)}{| | f _ {v} (x) | | \\cdot | | f _ {t} (c) | |}. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 596, + 603, + 906, + 638 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "While effective, $f$ can be improved without any training by using an auxiliary cache that stores examples of the classes of interest [36, 48, 57]. Given a cache $\\mathcal{C} = \\{x_1^1, \\dots, x_C^K\\}$ that contains $K$ samples of $C$ classes, we can define the cache-based prediction for a class $c \\in C$ as:", + "bbox": [ + 511, + 646, + 908, + 723 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nf _ {\\mathcal {C}} (x, c) = \\sum_ {i = 1} ^ {K} \\eta_ {\\mathcal {C}} \\left(\\frac {f _ {v} (x) ^ {\\intercal} f _ {v} \\left(x _ {c} ^ {i}\\right)}{\\left| \\left| f _ {v} (x) \\right| \\right| \\cdot \\left| \\left| f _ {v} \\left(x _ {c} ^ {i}\\right) \\right| \\right|}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 563, + 733, + 906, + 773 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\eta_{\\mathcal{C}}(z) = \\exp (1 - \\beta (-z))$ is a normalization function regulated by the hyperparameter $\\beta$ [57].", + "bbox": [ + 511, + 786, + 906, + 816 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "1In some problem formulations, the dataset may contain (i) object labels, (ii) bounding boxes, (iii) an \"unknown\" value for some attributes. For simplicity, we omit the latter from the formulation and, following previous works [39], we consider the case where objects bounding boxes are given.", + "bbox": [ + 511, + 825, + 906, + 875 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "2For simplicity, we omit the presence of prompts such as \"a photo of\" when defining the text input.", + "bbox": [ + 514, + 876, + 903, + 898 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "15100", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The formulation in Eq. (2) has a main issue when applied to open-vocabulary attribute detection: the cache construction. On one hand, a cache considering all possible attribute labels is semantically complete but requires a prohibitive storage size of $K \\cdot n!$ images, e.g., $K \\cdot 117!$ for OVAD [5]. On the other hand, while storing $K$ positive examples per attribute has low memory requirements, it ignores the multi-attribute nature of images (e.g., shape, size, color, etc. can co-exist in the same sample), potentially leading to incorrect predictions (e.g., a high score because the color matches but the cache label refers to the shape). Moreover, attributes may appear differently across objects (e.g., wet car vs wet dog) and selecting a random subset of them may lead to a cache representative only of a few cases. In the following, we address these issues by focusing on the compatibility between objects and multiple attributes in the real world.", + "bbox": [ + 89, + 90, + 483, + 333 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Compatibility for scalable caching", + "text_level": 1, + "bbox": [ + 89, + 342, + 390, + 358 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Naively constructing the cache would lead to many problems, related both to labels' semantic and memory requirements (Sec. 3.1). In this section, we present a methodology that (i) has the same memory requirements as storing only $K$ examples per attribute and (ii) assigns multiple attribute labels per example, taking into account attribute-object compatibility. Scalable caching. To obtain a cache $\\mathcal{C}$ that is scalable with the number of attributes and accounts for object-specific information, we exploit the natural distributions of attributes across objects. In particular, let us define a function $\\varphi$ that, given an attribute $a \\in \\mathcal{A}$ and an object $o \\in \\mathcal{O}$ , returns their compatibility value as a scalar, i.e., $\\varphi : \\mathcal{A} \\times \\mathcal{O} \\to \\mathbb{R}$ .", + "bbox": [ + 89, + 364, + 483, + 545 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Considering an attribute $a$ and an object $o$ , the score $\\varphi(a, o)$ is proportional to the probability that $o$ exhibits $a$ . Thus, we can use the scores $\\Phi_O(a) = [\\varphi(a, o_1), \\dots, \\varphi(a, o_m)]$ as a proxy for the probability distribution of $a$ co-occurring with the elements of $O$ . We then assign each of the $K$ elements (i.e., shots) of the attribute $a$ to one object $o$ , sampling the object from $\\Phi_O(a)$ :", + "bbox": [ + 89, + 547, + 483, + 652 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {O} _ {a} ^ {K} = \\left\\{\\hat {o} _ {1} ^ {a}, \\dots , \\hat {o} _ {K} ^ {a} \\right\\}, \\text {w h e r e} \\hat {o} _ {i} ^ {a} \\sim \\tilde {\\Phi} _ {O} (a), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 138, + 661, + 483, + 681 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\tilde{\\Phi}_O(a)$ is the normalized version of $\\Phi_O(a)$ , i.e., $\\tilde{\\Phi}_O(a) = \\Phi_O(a) / \\sum \\Phi_O(a)$ . The set $\\hat{O}_a^K$ allows populating the cache with relevant objects and attributes. Let us denote a generic function mapping a textual query to an image as $\\mathrm{T2I}:\\mathcal{T}\\to \\mathcal{X}$ . The elements of the cache are:", + "bbox": [ + 89, + 691, + 483, + 767 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {C} _ {\\mathrm {H}} = \\left\\{x _ {a _ {1}} ^ {1}, \\dots , x _ {a _ {m}} ^ {K} \\right\\} (4) \\\\ = \\{\\mathrm {T 2 I} \\left(\\pi_ {g} \\circ a \\circ o _ {1} ^ {a}\\right), \\dots , \\mathrm {T 2 I} \\left(\\pi_ {g} \\circ a \\circ o _ {K} ^ {a}\\right) \\}, (4) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 116, + 776, + 483, + 813 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\pi_g$ is a contextual prompt for the image (e.g., \"a photo of\"). Following SuS-X [48], we implement T2I via retrieval. Specifically, we retrieve images from a database by scoring them against textual queries with the VLM (e.g., CLIP). While we focus on retrieval, COMCA", + "bbox": [ + 89, + 825, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "also supports the implementation of T2I via a text-to-image generative model, e.g., by feeding textual queries to Stable Diffusion [42]. We refer to the Supp. Mat. for more details. Estimating attribute-object compatibility. The key element that enables scalable caching is the function $\\varphi$ that estimates the viability of attribute-object compositions. We implement it by merging two scoring strategies.", + "bbox": [ + 511, + 90, + 906, + 196 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The first uses a generic large-scale database $D_r$ containing image-caption pairs. Using the captions $t \\in D_r$ , we can estimate the database score as the number of co-occurrences of attribute $a$ and object $o$ , i.e.:", + "bbox": [ + 511, + 196, + 906, + 257 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\varphi_ {\\mathrm {d b}} (a, o) = \\sum_ {(x, t) \\in D _ {r}} \\mathbb {1} (o \\subset t \\wedge a \\subset t), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 578, + 267, + 906, + 301 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbb{1}(z)$ is an indicator function whose value is 1 whenever the condition $z$ is true, and 0 otherwise.", + "bbox": [ + 511, + 313, + 906, + 342 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "While this process exploits real data, it may be biased on the specific captions and/or data collection process. For instance, captions rarely describe objects in their default state, such as a dry dog. Therefore, a second scoring strategy exploits the reasoning capabilities of an LLM, querying it to predict an attribute-object compatibility as:", + "bbox": [ + 511, + 343, + 905, + 434 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\varphi_ {\\mathrm {L L M}} (a, o) = \\mathrm {L L M} \\left(\\pi_ {c}, o, a\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 617, + 446, + 906, + 462 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\pi_c$ is a prompt specifying the compatibility task (see the Supp. Mat.). The objective is to assign a score on a specific range to the likelihood of the attribute-object pair.", + "bbox": [ + 511, + 473, + 905, + 518 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Taking into account that also LLMs have issues, such as hallucinations and biases [14, 33], we implement $\\varphi$ by combining database- and LLM-based scoring as:", + "bbox": [ + 511, + 518, + 905, + 564 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\varphi (a, o) = \\varphi_ {\\mathrm {d b}} (a, o) \\cdot \\varphi_ {\\mathrm {L L M}} (a, o), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 599, + 575, + 906, + 592 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "thus computing $\\Phi_O(a)$ and filling the cache accordingly.", + "bbox": [ + 511, + 603, + 885, + 619 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Compositional cache refinement", + "text_level": 1, + "bbox": [ + 511, + 627, + 795, + 643 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "After the process in Sec. 3.2, we have a cache $\\mathcal{C}_{\\mathbb{H}}$ populated with $K$ examples for each attribute $a \\in A$ . However, the cache assigns one hard attribute to each element, contrasting with the natural co-occurrence of attributes and leading to potentially wrong associations during inference. As a simple example, an image of a large and blue car obtained with the binding (large, car) will have a positive label only for large. As a consequence, if the test image shows a small blue car, the cache element will not contribute to the score of blue and may erroneously increase the score for the antonym large.", + "bbox": [ + 511, + 648, + 906, + 800 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Soft caching. To remedy to this issue, we refine the label associated to each cache element, exploiting the VLM capabilities to compute soft labels. Formally, the soft label $s_a$ for an attribute $a$ and a cache image $x_c$ is defined as:", + "bbox": [ + 511, + 801, + 906, + 859 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ns _ {a} \\left(x _ {c}\\right) = \\frac {f _ {t} (a) ^ {\\intercal} f _ {v} \\left(x _ {c}\\right)}{\\left| \\left| f _ {t} (a) \\right| \\right| \\cdot \\left| \\left| f _ {v} \\left(x _ {c}\\right) \\right| \\right|}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 604, + 869, + 906, + 905 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "15101", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where we compare the text embeddings of the attribute with the visual embeddings of the cache image. While these scores can directly serve as soft labels, we found them to exhibit low-variance due to the well-known modality-gap [30, 48]. We overcome this limitation and increase their relevance by defining normalized soft labels $\\bar{s}_a$ as:", + "bbox": [ + 89, + 90, + 483, + 181 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {s} _ {a} \\left(x _ {c}\\right) = \\frac {\\exp \\left(\\left(s _ {a} \\left(x _ {c}\\right) - \\mu_ {C}\\right) / \\sigma_ {C}\\right)}{\\sum_ {\\hat {a} \\in A} \\exp \\left(\\left(s _ {\\hat {a}} (x) - \\mu_ {C}\\right) / \\sigma_ {C}\\right)}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 142, + 191, + 483, + 227 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mu_{C}$ and $\\sigma_{C}$ are respectively the mean and standard deviation of the similarity between any attribute with any element of the cache. The first normalization expands the range of values of the similarity scores $s_a$ , while the softmax ensures that all samples have an equal contribution to the final prediction. In practice, we found the latter step to be important only on datasets with many attributes (e.g., VAW). We hypothesize this is due to differences in the output space cardinality, requiring different softmax scaling as in [58].", + "bbox": [ + 89, + 237, + 483, + 373 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "These scores are mixed with the original hard labels to obtain the final label $\\hat{s}_a(x_c) = (1 - \\alpha)\\cdot \\mathbb{1}(a = c) + \\alpha \\cdot \\overline{s}_a(x_c)$ where $\\alpha$ is a trade-off hyperparameter. With the soft labels $\\hat{s}_a$ , we can perform the cache-based prediction as:", + "bbox": [ + 89, + 375, + 483, + 434 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nf _ {\\mathcal {C}} ^ {\\mathrm {C O M C A}} (x, a) = \\sum_ {x _ {c}} ^ {\\mathcal {C}} \\eta_ {\\mathcal {C}} \\left(\\hat {s} _ {a} \\left(x _ {c}\\right) \\frac {f _ {v} (x) ^ {\\intercal} f _ {v} \\left(x _ {c}\\right)}{\\left\\| f _ {v} (x) \\right\\| \\cdot \\left\\| f _ {v} \\left(x _ {c}\\right) \\right\\|}\\right). \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 101, + 446, + 483, + 500 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In contrast to Eq. (2), here we consider each cache element $x_{c}$ as a potential contributor for attribute $a$ , even if $a \\neq c$ . Implicitly, this allows us to include an arbitrary number of (positive) attribute labels within each element of the cache.", + "bbox": [ + 89, + 501, + 483, + 561 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Final inference. Finally, the cache-based scoring defined in Eq. (10) is merged with the standard CLIP-based scores to obtain the final predictions:", + "bbox": [ + 89, + 561, + 483, + 606 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nf (x, a) = \\eta_ {A} \\left(\\lambda \\cdot f _ {\\mathcal {C}} ^ {\\mathrm {C O M C A}} (x, a) + f _ {\\mathcal {C}} ^ {\\mathrm {C L I P}} (x, a)\\right), \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 112, + 618, + 482, + 638 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda$ is an hyperparameter controlling the balance between the two and $\\eta_A(z) = \\mathrm{softmax}_A(z / \\max (z))$ , moving from simple scalar values to a distribution over the attributes space and leading to more stable scores. Note that Eq. (11) contains no element which needs training or labeled data for the specific attributes of interest. Given the target sets of attributes $A$ and objects $O$ , COMCA automatically re-generates the cache to update the inference procedure.", + "bbox": [ + 89, + 648, + 483, + 768 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 784, + 220, + 801 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. We evaluate COMCA on two datasets. OVAD [5] covers 80 classes and 117 attributes across 2,000 images from MS-COCO [38], for a total of 14,300 instances. Attributes are manually annotated as positive, negative, and unknown. VAW [39] has more concepts, i.e., 2,260 classes and 620 attributes across 10,392 images. VAW is more", + "bbox": [ + 89, + 809, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "sparse than OVAD, i.e., around 1.8 positive and 5.3 negative attributes per instance, and it is automatically annotated.", + "bbox": [ + 511, + 90, + 903, + 121 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Setting. Following prior works [5, 7], COMCA focuses on the box-given setting, where the bounding boxes of objects are given, as this allows us to assess its attribute detection capabilities more precisely. However, Tab. 3 also evaluates COMCA in the box-free setting by introducing an object detector. See the Supp. Mat. for more details on this setting.", + "bbox": [ + 511, + 122, + 906, + 212 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Metrics. We measure performance as mean Average Precision (mAP), and also report mAP scores on head, medium, and tail classes based on their frequency in the dataset.", + "bbox": [ + 511, + 213, + 906, + 258 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Baselines and competitors. We evaluate COMCA against training-based methods, i.e., OVAD [5], OvarNet [7], ArtVLM [63], and LOWA [16]. As we are the first to address the problem in a completely training-free fashion, we extend the baselines with several zero-shot methods [9, 28, 54-56], and adapt previous cache-based approaches to our task [48, 57]. For TIP-Adapter [57] and SuS-X [48] we construct a $K$ -shot cache, considering only attributes for sampling and disregarding their co-occurrences with objects. This follows their approach, which samples $K$ shots for each target class, i.e., attributes in our task. In addition, we test IAP [26, 27], which conditions predicted attributes on the inferred object category, using TIP-Adapter as the object classifier. We also consider a naive cache-based approach, i.e., \"image-based\", that uses CC12M [6] to build a cache specifically for each input image by retrieving the most similar samples. Since the retrieved images are unlabeled, we use our soft labels to annotate the cache. We construct all the proposed baselines on top of CLIP ViT-B/32 [9]. For additional details, please refer to the Supp. Mat.", + "bbox": [ + 511, + 260, + 908, + 561 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation details. We use CLIP ViT-B/32 [9] pretrained on LAION2B [45] as our primary VLM. Following TIP-Adapter [57], we set $\\lambda = 1.17$ in Eq. (11) and $\\beta = 1.0$ in Eq. (2). We set $\\alpha = 0.6$ to mix hard and soft labels on the validation set of VAW [39]. Lastly, we use CC12M [6] as the database in Eq. (5) and GPT 3.5 Turbo [35] as the LLM for Eq. (6), as we find it to perform better than GPT 4o-mini (see the Supp. Mat.). Unless otherwise stated, we use $K = 16$ shots per attribute, retrieving them from CC12M. We use the same hyperparameters for both datasets. For the box-free experiments, we use YOLOv11M [23] as the detector.", + "bbox": [ + 511, + 561, + 908, + 729 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Results", + "text_level": 1, + "bbox": [ + 513, + 739, + 607, + 753 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We present our comprehensive experimental evaluation comparing COMCA with state-of-the-art training- and cache-based methods. We also show its performance with different backbones and in the box-free setting. Moreover, we discuss qualitative results and limitations of other approaches.", + "bbox": [ + 511, + 763, + 906, + 839 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Comparison with the state of the art. Tab. 1 reports the results comparing zero-shot baselines, cache-based methods, and COMCA in the box-given setting. On both benchmarks, COMCA outperforms all the zero-shot and cache-based com", + "bbox": [ + 511, + 840, + 908, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "15102", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/8dee48bfbbe5b537f46c8c9eac02fa201ed9d51e683ab7ca8f722dcd58f0fb61.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodTrain sizeOVAD [5]VAW [39]
mAPHeadMediumTailmAPHeadMediumTail
Zero shotCLIP [9] RN5011.841.011.71.435.337.835.126.2
CLIP [9] ViT-B/3217.044.318.45.550.051.050.943.2
CLIP [9] ViT-L/1418.344.420.56.451.051.452.545.0
Cache-basedImage-based16.845.819.23.552.953.753.846.4
TIP-Adapter [57] + IAP [27]15.143.317.81.727.732.425.516.6
TIP-Adapter [57]16.744.419.73.157.557.559.451.2
SuS-X [48]20.248.924.64.530.233.828.820.3
ComCA27.454.334.69.058.158.259.951.7
Training basedOVAD [5]110k21.448.026.95.2----
OvarNet [7]190k28.658.635.59.568.5---
ArtVLM [63]N/A----71.975.072.159.4
LOWA [16]1.33M18.758.020.42.642.646.441.032.9
", + "bbox": [ + 91, + 88, + 915, + 335 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Comparison with state of the art. Green indicates COMCA. Bold indicates best among training-free methods. The symbol “-” indicates results for the competitors are not available on the original papers or the impossibility to run their method due to lack of public code and/or model weights. See the Supp. Mat. for an extended version of the table.", + "bbox": [ + 89, + 345, + 906, + 388 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "petitors, demonstrating the effectiveness of our specialized design of the cache. It scores an increment of $+10.6\\mathrm{mAP}$ on OVAD and $+5.2$ on VAW compared to the image-based baseline. Similarly, it surpasses TIP-Adapter [57] on both benchmarks by $+10.7\\mathrm{mAP}$ and $+0.6$ , respectively. SuS-X [48] increases performance on OVAD while decreasing it on VAW. Overall, COMCA has a gain of $+7.2\\mathrm{mAP}$ and $+27.9$ on the two datasets. Lastly, TIP-Adapter + IAP [27] demonstrates that naively using co-occurrences of attributes with objects [27] is insufficient to address this task. COMCA surpasses it by $+12.3\\mathrm{mAP}$ on OVAD and $+30.4$ on VAW.", + "bbox": [ + 88, + 396, + 480, + 561 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We also compare COMCA with training-based methods. Notably, COMCA outperforms LOWA [16] on both benchmarks, despite the latter having been trained on a large-scale dataset while our approach requires no training data.", + "bbox": [ + 89, + 564, + 482, + 625 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Experiments with multiple VLMs. To demonstrate that COMCA is model-agnostic, we test it on various backbones and VLMs [9, 28, 54-56] and report the results in Tab. 2. On both OVAD and VAW, COMCA boosts performance consistently for all the backbones up to $+11.5\\mathrm{mAP}$ on OVAD with CoCa ViT-L/14 [54] and $+15.8\\mathrm{mAP}$ on VAW with CLIP ResNet50 (RN50) [9]. On average, COMCA improves the performance of the original VLM by $+8.9$ on OVAD and $+7.0$ on VAW, showing COMCA has consistent advantages.", + "bbox": [ + 88, + 626, + 482, + 762 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Cross-datasets advantages. To demonstrate the generalization capabilities of our method, we report cross-dataset results in Fig. 4. We compare COMCA with OVAD when trained on various datasets (i.e., VAW [39], OVAD [5], MSCOCO [38]) and tested on the OVAD (left) and on the VAW (right) benchmarks. As shown in Fig. 4, performance significantly degrades when OVAD is tested on datasets different from the training one (yellow bars vs red bar), whereas our training-free COMCA achieves competitive performance", + "bbox": [ + 88, + 763, + 482, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "(green bars) with the same RN50 backbone. This reinforces our training-free method, which does not inherit biases from a training set, differently from training-based approaches.", + "bbox": [ + 511, + 396, + 903, + 441 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Box-free setting. In this work, we primarily focus on attribute recognition and therefore adopt the box-given setting. Here, we evaluate COMCA in the box-free setting to demonstrate that it achieves competitive performance in a less constrained scenario. Tab. 3 shows the results for zero-shot and cache-based methods, Specifically, we use CLIP ViT-B/32 [9] and YOLOv11M [23] as detector for all models. We note that all cache-based methods apart from SuS-X underperform the zero-shot baseline, with a decrease of up to $-3.1\\mathrm{mAP}$ for image-based. Only, SuS-X has a gain of $+2.8\\mathrm{mAP}$ w.r.t. CLIP ViT-B/32, with our method COMCA being superior and surpassing CLIP ViT-B/32 by $+9.0\\mathrm{mAP}$ . Notably, our model is competitive with the training-based ones, surpassing the performance of both OVAD $(+2.7\\mathrm{mAP})$ and LOWA $(+2.8\\mathrm{mAP})$ . We provide a more detailed analysis and an extended version of these results in the Supp. Mat.", + "bbox": [ + 511, + 444, + 905, + 686 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Qualitative results. Fig. 3, we show qualitative examples comparing training-based OVAD [5], CLIP ViT-B/32 [9], and COMCA with the same backbone. We consider the top-1 prediction for all methods. OVAD correctly identifies the bear's color but misclassifies the cat's color and wrongly predicts a \"tiled\" clothes pattern for two people. We speculate it likely predicts the same attribute for two objects (i.e., the $2^{nd}$ and the $6^{th}$ ) due to biases acquired during training. CLIP makes more errors, such as labeling the zebra as \"reflective\" and the airplane as \"tiny\", while misclassifying the baseball player as \"giant\". COMCA instead understands attributes better, for instance by accurately identifying the zebra's texture. These issues are less pronounced in COMCA thanks to our caching mechanism, which allows it to compare inputs", + "bbox": [ + 511, + 688, + 908, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "15103", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/fb350deee4ba76872502b745c25bb8fe5adf41460e93ed54b798acc6d4c9cdb9.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodOVAD [5]VAW [39]
BackboneCOMCAmAPΔmAPΔ
CLIP RN50 [9]11.835.3
✓19.2+7.451.1+15.8
CLIP ViT-B/32 [9]17.050.0
✓27.4+10.458.1+8.1
CLIP ViT-L/14 [9]18.351.0
✓24.8+6.561.0+10.0
SigLIP ViT-B/16 [56]13.752.5
✓22.1+8.459.8+7.3
CoCa ViT-B/32 [54]15.149.6
✓24.4+9.351.4+1.8
CoCa ViT-L/14 [54]14.449.8
✓25.9+11.554.5+4.7
BLIP [28]15.951.0
✓22.7+6.856.0+5.0
X-VLM [55]17.454.8
✓28.4+11.057.7+2.9
", + "bbox": [ + 94, + 88, + 480, + 316 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/37bdbd4f5a8456e96a664d9fbfe026e7f62fafb3aad4f8ab7405fb03964dffe7.jpg", + "table_caption": [ + "Table 2. Box-given results with different backbones. Green indicates COMCA applied on top of the backbone. $\\Delta$ indicates improvements w.r.t. the corresponding baseline. See the Supp. Mat. for an extended version of the table." + ], + "table_footnote": [], + "table_body": "
MethodOVAD [5]
mAPHeadMediumTail
Zero shotCLIP [9] RN5012.340.212.71.9
CLIP [9] ViT-B/3212.539.612.52.7
CLIP [9] ViT-L/1413.439.714.32.9
Cache-basedImage-based9.435.48.71.0
TIP [57] + IAP [27]9.836.99.21.0
TIP-Adapter [57]10.136.09.81.2
SuS-X [48]15.343.017.23.2
COMCA21.548.426.46.0
Training basedOVAD [5]18.847.722.04.6
OvarNet [7]27.256.833.68.9
LOWA [16]18.758.020.42.6
", + "bbox": [ + 511, + 88, + 898, + 300 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Results in the box-free setting. Green indicates COMCA. Bold indicates the best results among training-free methods. See the Supp. Mat. for an extended version of the table.", + "bbox": [ + 508, + 309, + 903, + 353 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/dca6ab54cfb7c11f58f092b69f981828ce8d75a4ff1d587ea69dbb145777b171.jpg", + "image_caption": [ + "OVAD CLIP COMCA", + "brown brown single, one, individual" + ], + "image_footnote": [], + "bbox": [ + 135, + 388, + 246, + 474 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7ce6350624792f5749ac16cee0b5d90dc83730a82509b80083a45b28ef00e985.jpg", + "image_caption": [ + "clothes pattern: tiled male neutral, calm, serious" + ], + "image_footnote": [], + "bbox": [ + 246, + 390, + 356, + 474 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/745fc7b5e48b84ea10d42653376b60c5c44208cd6b19e588c8fa0e21d346ae98.jpg", + "image_caption": [ + "plaid, tartan, checkered reflective striped, lined, pinstriped" + ], + "image_footnote": [], + "bbox": [ + 357, + 390, + 468, + 474 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/93e11421a085444e7676f52c6ae85268e6e4cd6181f7a437a51fd4735c0f961d.jpg", + "image_caption": [ + "large, giant, huge small, little, tiny smooth, sleek" + ], + "image_footnote": [], + "bbox": [ + 470, + 388, + 679, + 474 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1d7253325f036b30d81cf2d1cf6c400fa60d44663b32e2d82159fd575111c000.jpg", + "image_caption": [ + "yellow piece, cut small, little, tiny" + ], + "image_footnote": [], + "bbox": [ + 681, + 390, + 792, + 474 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c730c7329b0e3740b1aea6f8fc53ad1d4d8bc8f9de7b469c4355ea80974ed769.jpg", + "image_caption": [ + "clothes pattern: tiled big, large, giant, huge sitting, sit" + ], + "image_footnote": [], + "bbox": [ + 794, + 390, + 903, + 474 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6a23ffc3cb2bd9093dd8e46d13541458d3940461dac0717f96cd93cfa7ede7fc.jpg", + "image_caption": [ + "Figure 3. Qualitative results. Predictions of OVAD, CLIP and ComCA on some OVAD images. Green are correct ones, red are wrong.", + "Figure 4. Cross-dataset results. Results on OVAD (left) and VAW (right). Yellow indicates methods tested cross-domain, i.e., train and test set differ. Red indicates methods tested in-domain, and green our training-free method, which does not suffer from domain shift problems. All experiments use a RN50 backbone." + ], + "image_footnote": [], + "bbox": [ + 98, + 551, + 472, + 672 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "to self-labeled examples at test time.", + "bbox": [ + 89, + 763, + 333, + 777 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Ablation study", + "text_level": 1, + "bbox": [ + 89, + 787, + 240, + 803 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we analyze COMCA's components, studying how sampling strategies affect performance (Tab. 4), soft labeling's impact (Tab. 5 and Fig. 6), and the number of samples in the cache (Fig. 5). Additional experiments in the Supp. Mat. show the effect of (i) the hyperparameters $\\alpha$ and $\\lambda$ , (ii) constructing the cache with Stable Diffusion, and (iii)", + "bbox": [ + 89, + 810, + 483, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "using other LLMs for generative scores.", + "bbox": [ + 511, + 551, + 777, + 566 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Cache construction. In Tab. 4, we evaluate five strategies to construct the cache. The first, Random, samples $K$ objects for each attribute and creates a cache of $|A| \\times K$ examples, seeking only a match on the attribute. The second, Brute force, considers all possible combinations of attributes and objects, resulting in a large cache of $|O| \\times |A| \\times K$ samples. As VAW contains many attributes and objects, the Brute force cache would be too large and cannot be practically tested. As shown in Tab. 4, both approaches lead to very low results on OVAD. Compared to the zero-shot CLIP baseline, performance is similar (Random, -0.3 mAP) or even below (Brute force, -6.4 mAP). We hypothesize that Random ignores attribute-object relationships, while Brute force introduces many noisy and potentially detrimental samples, harming performance. On the contrary, Random improves by +7.5 on VAW. We hypothesize this is due to the nature of the dataset, which contains many attributes and objects.", + "bbox": [ + 509, + 566, + 906, + 824 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "When introducing priors on object-attributes compositions via a database (Eq. (5)) or an LLM (Eq. (6)), performance largely increases by up to $+4.1$ mAP on VAW w.r.t. the baseline for database scores. ComCA combines the two scores to obtain object priors, achieving the best perfor", + "bbox": [ + 511, + 825, + 908, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "15104", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/867d7eab45d6861eb7c42f75e372412fc798682c1239b3103caba68701c815de.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ConfigurationmAP
OVAD [5]VAW [39]
Zero-shot [9]17.050.0
Random16.757.5
Brute force10.6-
DB scores18.654.1
LLM scores21.155.4
DB + LLM scores26.455.2
", + "bbox": [ + 94, + 88, + 367, + 218 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/4ebc949e263ea7dd1c793f0e1ab87dcec9c28e9f14d1d14c006a9592589e9d9c.jpg", + "table_caption": [ + "Table 4. Ablation on cache construction. Green indicates our default configuration. Bold indicates the best results. Results are without soft labels. See the Supp. Mat. for an extended version of the table." + ], + "table_footnote": [], + "table_body": "
ConfigurationmAP
OVAD [5]VAW [39]
Zero-shot [9]17.050.0
One-hot26.455.2
Soft26.753.2
- Sharpening [2]21.527.7
- Softmax24.558.5
- Eq. (9)27.458.1
", + "bbox": [ + 380, + 87, + 653, + 222 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. Ablation on soft labels. Green indicates our default configuration. Bold indicates the best results. See the Supp. Mat. for an extended version of the table.", + "bbox": [ + 375, + 231, + 656, + 286 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3c696229124d5b3e2c71259862b246ecd9672b720a8d8f9fadda9fac9bc4ecce.jpg", + "image_caption": [ + "Figure 5. Ablation on number of samples per attribute. The purple line is mAP on OVAD, while the red line is mAP on VAW. Green indicates our default configuration." + ], + "image_footnote": [], + "bbox": [ + 673, + 95, + 893, + 214 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4997d68500826be54128ff262482948f7da2d213a30c33e0f5c98076a1e805dd.jpg", + "image_caption": [ + "Figure 6. Soft labels across cache priors. Performance on OVAD (left) and VAW (right), considering generative (LLM) and retrieval (CC3M, CC12M) scores. Yellow indicates one-hot labels, red soft labels and green COMCA, combining LLM and CC12M." + ], + "image_footnote": [], + "bbox": [ + 96, + 328, + 473, + 409 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "mance. These results confirm both (i) that attribute-object compatibility is important to build the cache, and (ii) that databases and LLMs provide complementary information.", + "bbox": [ + 89, + 500, + 482, + 545 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Soft labels. As discussed in Sec. 3.3, each image in the cache contains multiple attributes, even if we fill the cache considering only one of them. Here we analyze the soft labels obtained via Eq. (8), comparing them with the One-hot attribute labels of the initial cache, reporting the results in Tab. 5.", + "bbox": [ + 89, + 546, + 483, + 622 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The first strategy is to incorporate the soft scores as they are predicted (Soft in Tab. 5). This leads to results slightly above the One-hot baselines on OVAD $(+0.3\\mathrm{mAP})$ , while reducing the performance in VAW $(-2.0\\mathrm{mAP})$ . This is a direct consequence of allowing each sample to have a different level of relevance in the cache, something that becomes more prominent in VAW due to the larger number of attributes.", + "bbox": [ + 89, + 625, + 483, + 729 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To fix this issue, we can normalize the scores. We study three strategies: Sharpening, where the scores are normalized as in PAWS [2]; Softmax normalization over each sample; and the one we propose in Eq. (9), i.e., using the statistics of the cache-attributes similarity. Sharpening is not effective in this scenario, achieving the lowest performance across both settings. Softmax normalization leads to good results on VAW, e.g., +3.3 mAP w.r.t. One-hot, but it struggles on OVAD (24.5 mAP) due to the limited range of values.", + "bbox": [ + 89, + 732, + 482, + 868 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our strategy addresses both issues simultaneously, consistently improving metrics in both settings compared to", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "the One-hot strategy, scoring $27.4\\mathrm{mAP}$ on OVAD and 58.1 mAP on VAW. Note that these findings are general. Fig. 6 shows soft labels consistently enhance results for estimating attribute-object priors across all strategies and datasets. Improvements range from $+3.8\\mathrm{mAP}$ on VAW to $+5.4\\mathrm{mAP}$ on OVAD in the LLM-only setting.", + "bbox": [ + 511, + 327, + 906, + 417 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Number of samples per attribute. In Fig. 5, we measure how adding more samples (varying $K$ ) per attribute influences performance on the two datasets. We observe performance consistently increasing for OVAD while going from 1 to 16 samples, reaching a plateau when moving to $K = 24$ . On VAW, performance is stable for different values of $K$ due to the large number of attributes in the dataset (620), making the cache representative enough even when $K = 1$ . These results confirm the stability of COMCA.", + "bbox": [ + 511, + 419, + 908, + 555 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 571, + 633, + 587 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We introduced Compositional Caching for open-vocabulary attribute detection. COMCA uses a VLM and an auxiliary cache of images to solve this task in a training-free manner. Differently from previous works [48, 57], we construct the cache based on attribute-object compatibility derived from web-scale databases and LLMs. We also introduced a soft-labeling scheme to account for the co-occurrence of multiple attributes within an image. Combining these two components allows to model the natural co-occurrence of attributes in the real world, together and across objects, while making the cache construction scalable. Our experiments demonstrate that COMCA is a strong training-free approach. Since COMCA is training-free, it does not suffer from differences between training and test attributes, contrary to training-based methods. This also enables COMCA to be adapted to special domains (e.g., aerial images) more easily than training-based methods, which would require specific re-training. Future work will focus on devising methods to address attribute detection even in more challenging settings where objects or attributes are not given [10].", + "bbox": [ + 511, + 598, + 908, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "15105", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments. We acknowledge ISCRA for awarding this project access to the LEONARDO supercomputer, owned by the EuroHPC Joint Undertaking, hosted by CINECA (Italy). This work was sponsored by the MUR PNRR project FAIR - Future AI Research (PE00000013), funded by NextGeneration EU, and the EU projects AI4TRUST (No.101070190) and ELIAS (No.01120237).", + "bbox": [ + 89, + 90, + 485, + 196 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 210, + 187, + 224 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Zeynep Akata, Florent Perronnin, Zaid Harchaoui, and Cordelia Schmid. Label-embedding for attribute-based classification. In CVPR, 2013. 2", + "[2] Mahmoud Assran, Mathilde Caron, Ishan Misra, Piotr Bojanowski, Armand Joulin, Nicolas Ballas, and Michael Rabbat. Semi-supervised learning of visual features by nonparametrically predicting view assignments with support samples. In ICCV, 2021. 8", + "[3] Alberto Baldrati, Lorenzo Agnolucci, Marco Bertini, and Alberto Del Bimbo. Zero-shot composed image retrieval with textual inversion. In ICCV, 2023. 2", + "[4] Tamara L Berg, Alexander C Berg, and Jonathan Shih. Automatic attribute discovery and characterization from noisy web data. In ECCV, 2010. 2", + "[5] Maria A Bravo, Sudhanshu Mittal, Simon Ging, and Thomas Brox. Open-vocabulary attribute detection. In CVPR, 2023. 1, 2, 3, 4, 5, 6, 7, 8", + "[6] Soravit Changpinyo, Piyush Sharma, Nan Ding, and Radu Soricut. Conceptual 12m: Pushing web-scale image-text pre-training to recognize long-tail visual concepts. In CVPR, 2021. 2, 5", + "[7] Keyan Chen, Xiaolong Jiang, Yao Hu, Xu Tang, Yan Gao, Jianqi Chen, and Weidi Xie. Ovarnet: Towards open-vocabulary object attribute recognition. In CVPR, 2023. 1, 2, 3, 5, 6, 7", + "[8] Steven Chen and Kristen Grauman. Compare and contrast: Learning prominent visual differences. In CVPR, 2018. 2", + "[9] Mehdi Cherti, Romain Beaumont, Ross Wightman, Mitchell Wortsman, Gabriel Ilharco, Cade Gordon, Christoph Schuhmann, Ludwig Schmidt, and Jenia Jitsev. Reproducible scaling laws for contrastive language-image learning. In CVPR, 2023. 5, 6, 7, 8", + "[10] Alessandro Conti, Enrico Fini, Massimiliano Mancini, Paolo Rota, Yiming Wang, and Elisa Ricci. Vocabulary-free image classification. NeurIPS, 36, 2024. 8", + "[11] Ali Farhadi, Ian Endres, Derek Hoiem, and David Forsyth. Describing objects by their attributes. In CVPR, 2009. 2", + "[12] Ali Farhadi, Ian Endres, and Derek Hoiem. Attribute-centric recognition for cross-category generalization. In CVPR, 2010.", + "[13] Vittorio Ferrari and Andrew Zisserman. Learning visual attributes. NeurIPS, 2007. 2", + "[14] Isabel O Gallegos, Ryan A Rossi, Joe Barrow, Md Mehrab Tanjim, Sungchul Kim, Franck Dernoncourt, Tong Yu, Ruiyi Zhang, and Nesreen K Ahmed. Bias and fairness in large language models: A survey. arXiv preprint arXiv:2309.00770, 2023. 4" + ], + "bbox": [ + 93, + 234, + 485, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Peng Gao, Shijie Geng, Renrui Zhang, Teli Ma, Rongyao Fang, Yongfeng Zhang, Hongsheng Li, and Yu Qiao. Clip-adapter: Better vision-language models with feature adapters. IJCV, 132(2):581-595, 2024. 2", + "[16] Xiaoyuan Guo, Kezhen Chen, Jinmeng Rao, Yawen Zhang, Baochen Sun, and Jie Yang. Lowa: Localize objects in the wild with attributes. In NeurIPS-WS, 2023. 1, 2, 5, 6, 7", + "[17] Ziyu Guo, Renrui Zhang, Longtian Qiu, Xianzheng Ma, Xu-peng Miao, Xuming He, and Bin Cui. Calip: Zero-shot enhancement of clip with parameter-free attention. In AAAI, 2023. 2", + "[18] Xintong Han, Zuxuan Wu, Phoenix X Huang, Xiao Zhang, Menglong Zhu, Yuan Li, Yang Zhao, and Larry S Davis. Automatic spatially-aware fashion concept discovery. In ICCV, 2017. 2", + "[19] Mehrdad Hosseinzadeh and Yang Wang. Composed query image retrieval using locally bounded features. In CVPR, 2020. 2", + "[20] Edward J Hu, Phillip Wallis, Zeyuan Allen-Zhu, Yanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. In ICLR, 2021. 2", + "[21] Phillip Isola, Joseph J Lim, and Edward H Adelson. Discovering states and transformations in image collections. In CVPR, 2015. 2", + "[22] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In ECCV, 2022. 2", + "[23] Glenn Jocher and Jing Qiu. Ultralytics yolo11. 2024. 5, 6", + "[24] Shyamgopal Karthik, Karsten Roth, Massimiliano Mancini, and Zeynep Akata. Vision-by-language for training-free compositional image retrieval. In ICLR, 2024. 2", + "[25] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. IJCV, 123:32-73, 2017. 2", + "[26] Christoph H Lampert, Hannes Nickisch, and Stefan Harmeling. Learning to detect unseen object classes by between-class attribute transfer. In CVPR, 2009. 2, 5", + "[27] Christoph H Lampert, Hannes Nickisch, and Stefan Harmeling. Attribute-based classification for zero-shot visual object categorization. IEEE TPAMI, 36(3):453-465, 2013. 2, 5, 6, 7", + "[28] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, 2022. 2, 5, 6, 7", + "[29] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML, 2023. 2", + "[30] Victor Weixin Liang, Yuhui Zhang, Yongchan Kwon, Serena Yeung, and James Y Zou. Mind the gap: Understanding the modality gap in multi-modal contrastive representation learning. NeurIPS, 2022. 5", + "[31] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. NeurIPS, 2024. 2" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "15106", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[32] Sachit Menon and Carl Vondrick. Visual classification via description from large language models. In ICLR, 2023. 2", + "[33] Roberto Navigli, Simone Conia, and Björn Ross. Biases in large language models: origins, inventory, and discussion. ACM Journal of Data and Information Quality, 15(2):1-21, 2023. 4", + "[34] Zachary Novack, Julian McAuley, Zachary Chase Lipton, and Saurabh Garg. Chils: Zero-shot image classification with hierarchical label sets. In ICML, 2023. 2", + "[35] OpenAI. Introducing chatgpt. https://openai.com/blog/chatgpt, 2022.5", + "[36] Emin Orhan. A simple cache model for image recognition. NeurIPS, 2018. 3", + "[37] Devi Parikh and Kristen Grauman. Relative attributes. In ICCV, 2011. 2", + "[38] Genevieve Patterson and James Hays. Coco attributes: Attributes for people, animals, and objects. In ECCV, 2016. 1, 2, 5, 6", + "[39] Khoi Pham, Kushal Kafle, Zhe Lin, Zhihong Ding, Scott Cohen, Quan Tran, and Abhinav Shrivastava. Learning to predict visual attributes in the wild. In CVPR, 2021. 2, 3, 5, 6, 7, 8", + "[40] Khoi Pham, Kushal Kafle, Zhe Lin, Zhihong Ding, Scott Cohen, Quan Tran, and Abhinav Shrivastava. Improving closed and open-vocabulary attribute prediction using transformers. In ECCV, 2022. 1, 2", + "[41] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 1, 2, 3", + "[42] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 3, 4", + "[43] Karsten Roth, Jae Myung Kim, A Koepke, Oriol Vinyals, Cordelia Schmid, and Zeynep Akata. Waffling around for performance: Visual classification with random words and broad concepts. In ICCV, 2023. 2", + "[44] Kuniaki Saito, Kihyuk Sohn, Xiang Zhang, Chun-Liang Li, Chen-Yu Lee, Kate Saenko, and Tomas Pfister. Pic2word: Mapping pictures to words for zero-shot composed image retrieval. In CVPR, 2023. 2", + "[45] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. NeurIPS, 2022. 3, 5", + "[46] Behjat Siddique, Rogerio S Feris, and Larry S Davis. Image ranking and retrieval based on multi-attribute queries. In CVPR, 2011. 2", + "[47] Kihyuk Sohn, Huiwen Chang, José Lezama, Luisa Polania, Han Zhang, Yuan Hao, Irfan Essa, and Lu Jiang. Visual prompt tuning for generative transfer learning. In CVPR, 2023. 2", + "[48] Vishaal Udandarao, Ankush Gupta, and Samuel Albanie. Susx: Training-free name-only transfer of vision-language models. In ICCV, 2023. 2, 3, 4, 5, 6, 7, 8" + ], + "bbox": [ + 91, + 90, + 483, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[49] Peter Welinder, Steve Branson, Takeshi Mita, Catherine Wah, Florian Schroff, Serge Belongie, and Pietro Perona. Caltech-ucsd birds 200. 2010. 2", + "[50] Hui Wu, Yupeng Gao, Xiaoxiao Guo, Ziad Al-Halah, Steven Rennie, Kristen Grauman, and Rogerio Feris. Fashion iq: A new dataset towards retrieving images by natural language feedback. In CVPR, 2021. 2", + "[51] Yongqin Xian, Christoph H Lampert, Bernt Schiele, and Zeynep Akata. Zero-shot learning—a comprehensive evaluation of the good, the bad and the ugly. IEEE TPAMI, 41(9): 2251-2265, 2018. 2", + "[52] Hantao Yao, Rui Zhang, and Changsheng Xu. Visual-language prompt tuning with knowledge-guided context optimization. In CVPR, 2023. 2", + "[53] Aron Yu and Kristen Grauman. Fine-grained visual comparisons with local learning. In CVPR, 2014. 2", + "[54] Jiahui Yu, Zirui Wang, Vijay Vasudevan, Legg Yeung, Mojtaba Seyedhosseini, and Yonghui Wu. Coca: Contrastive captioners are image-text foundation models. TMLR, 2022. 2, 5, 6, 7", + "[55] Yan Zeng, Xinsong Zhang, and Hang Li. Multi-grained vision language pre-training: Aligning texts with visual concepts. ICML, 2022. 7", + "[56] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In ICCV, 2023. 5, 6, 7", + "[57] Renrui Zhang, Wei Zhang, Rongyao Fang, Peng Gao, Kunchang Li, Jifeng Dai, Yu Qiao, and Hongsheng Li. Tip-adapter: Training-free adaption of clip for few-shot classification. In ECCV, 2022. 2, 3, 5, 6, 7, 8", + "[58] Xiao Zhang, Rui Zhao, Yu Qiao, Xiaogang Wang, and Hongsheng Li. Adacos: Adaptively scaling cosine logits for effectively learning deep face representations. In CVPR, 2019. 5", + "[59] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Conditional prompt learning for vision-language models. In CVPR, 2022. 2", + "[60] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. IJCV, 130(9):2337-2348, 2022. 2", + "[61] Yifei Zhou, Juntao Ren, Fengyu Li, Ramin Zabih, and Ser Nam Lim. Test-time distribution normalization for contrastively learned visual-language models. NeurIPS, 2024. 2", + "[62] Beier Zhu, Yulei Niu, Yucheng Han, Yue Wu, and Hanwang Zhang. Prompt-aligned gradient for prompt tuning. In ICCV, 2023. 2", + "[63] William Y Zhu, Keren Ye, Junjie Ke, Jiahui Yu, Leonidas Guibas, Peyman Milanfar, and Feng Yang. Artvlm: Attribute recognition through vision-based prefix language modeling. ECCV, 2024. 2, 5, 6", + "[64] Xiangyang Zhu, Renrui Zhang, Bowei He, Aojun Zhou, Dong Wang, Bin Zhao, and Peng Gao. Not all features matter: Enhancing few-shot clip with adaptive prior refinement. In ICCV, 2023. 3" + ], + "bbox": [ + 514, + 90, + 906, + 869 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "15107", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Compositional Caching for Training-free Open-vocabulary Attribute Detection/5997b1e0-b306-408f-91ce-2ab833ab2009_model.json b/2025/Compositional Caching for Training-free Open-vocabulary Attribute Detection/5997b1e0-b306-408f-91ce-2ab833ab2009_model.json new file mode 100644 index 0000000000000000000000000000000000000000..cee13517ee71adf774e679084838ce4c5120b44d --- /dev/null +++ b/2025/Compositional Caching for Training-free Open-vocabulary Attribute Detection/5997b1e0-b306-408f-91ce-2ab833ab2009_model.json @@ -0,0 +1,2365 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.102, + 0.131, + 0.895, + 0.152 + ], + "angle": 0, + "content": "Compositional Caching for Training-free Open-vocabulary Attribute Detection" + }, + { + "type": "text", + "bbox": [ + 0.132, + 0.18, + 0.862, + 0.233 + ], + "angle": 0, + "content": "Marco Garosi1 Alessandro Conti1 Gaowen Liu2 Elisa Ricci1,3 Massimiliano Mancini1 \n1University of Trento 2Cisco Research Fondazione Bruno Kessler \nhttps://comca-attributes. github.io" + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.269, + 0.326, + 0.284 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.301, + 0.486, + 0.709 + ], + "angle": 0, + "content": "Attribute detection is crucial for many computer vision tasks, as it enables systems to describe properties such as color, texture, and material. Current approaches often rely on labor-intensive annotation processes which are inherently limited: objects can be described at an arbitrary level of detail (e.g., color vs. color shades), leading to ambiguities when the annotators are not instructed carefully. Furthermore, they operate within a predefined set of attributes, reducing scalability and adaptability to unforeseen downstream applications. We present COMPOSITIONAL CACHING (COMCA), a training-free method for open-vocabulary attribute detection that overcomes these constraints. COMCA requires only the list of target attributes and objects as input, using them to populate an auxiliary cache of images by leveraging web-scale databases and Large Language Models to determine attribute-object compatibility. To account for the compositional nature of attributes, cache images receive soft attribute labels. Those are aggregated at inference time based on the similarity between the input and cache images, refining the predictions of underlying Vision-Language Models (VLMs). Importantly, our approach is model-agnostic, compatible with various VLMs. Experiments on public datasets demonstrate that COMCA significantly outperforms zero-shot and cache-based baselines, competing with recent training-based methods, proving that a carefully designed training-free approach can successfully address open-vocabulary attribute detection." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.74, + 0.221, + 0.755 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Attributes shape our perception and interactions with the environment. Let us consider the example of Fig. 1. If our aim is to identify the animal, we may focus on its fur color, size, the presence of horns and the tail. Similarly, if we want to analyze its health status, we may focus on its age, fur patterns and shades, and its skin. This level of analysis requires models with fine-grained understanding of the visual input, which is usually obtained by training a model on datasets that annotate both objects and their corresponding" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.269, + 0.905, + 0.416 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.429, + 0.908, + 0.526 + ], + "angle": 0, + "content": "Figure 1. Example of attribute annotations. Attribute annotations are often: sparse, as they are not consistent across samples; incomplete, as not all attributes are annotated; ambiguous, as they can be subjective or miss a frame of reference. This makes open-vocabulary attribute detection challenging. Differently from previous works [5, 7], we do not rely on such annotations and propose COMCA, a training-free approach requiring no supervision." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.537, + 0.909, + 0.703 + ], + "angle": 0, + "content": "attributes [38, 40]. However, collecting attribute annotations is both time-consuming and prone to ambiguities, especially when annotators lack clear instructions. Additionally, attribute annotations are often tailored to specific datasets and are ambiguous. For instance, in Fig. 1, one may list \"brown\" as a color for the left goat, but others may argue that it is a red-ish shade of brown. A simple solution is to provide annotators with a predefined list of attributes; however, (i) the list may not cover all possible annotations for a property (e.g., all colors) and (ii) it limits models trained on it to the given list itself, making them unable to recognize new attributes." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.704, + 0.909, + 0.825 + ], + "angle": 0, + "content": "Recent approaches overcome some of the challenges by developing open-vocabulary attribute detection models [5, 7, 16] that detect attributes specified in natural language at inference time. While these approaches generalize beyond the set of seen attributes and objects, their performance is still biased toward the specific dataset they are trained on. Indeed, training-based approaches may show significant degradation in performance in cross-dataset experiments (see Fig. 4)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.909, + 0.901 + ], + "angle": 0, + "content": "In this work, we overcome potentially ambiguous and/or limited attribute annotations by introducing COMPOSITIONAL CACHING (COMCA), a training-free open-vocabulary attribute detection model. COMCA employs a Vision-Language Model (VLM) [41] and refines its" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "15098" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.122 + ], + "angle": 0, + "content": "predictions by utilizing an auxiliary cache that contains images together with automatically estimated attribute labels." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.125, + 0.486, + 0.337 + ], + "angle": 0, + "content": "Caching for refining CLIP predictions has been introduced in the context of image classification [57]. It requires storing images, either generated or retrieved [48, 57], for each of the classes of interests, together with a corresponding one-hot categorical label. The similarity of the input image with each of the examples is used to weight the labels in the cache and refine the model's prediction. While this approach is effective for object classification, it is prohibitively expensive in the context of attribute detection due to the much larger output space, comprising all possible combinations of attributes and objects. COMCA achieves scalability in caching by following two key compositional principles: (i) not all attributes are universally applicable to all objects, and (ii) every object inherently possesses multiple attributes." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.34, + 0.487, + 0.687 + ], + "angle": 0, + "content": "The first principle is implemented in two steps: attribute-object compatibility estimation and cache pair sampling. In the first phase, we estimate the probability of an attribute appearing with an object using the statistics of available databases (e.g., CC12M [6]) and Large Language Model (LLM) queries. Next, we populate the cache for a given attribute by sampling compatible objects from the estimated distribution and retrieving relevant images from a database. The second principle is implemented by self-labeling the cached elements. Initially, each sample is associated with the single attribute it was sampled for. Since this does not account for other attributes present in the image, we extend attribute labels through soft-labeling, i.e., comparing cache images to text embeddings of each attribute of interest. These scores modify a cache element's contribution to each target attribute, considering their implicit composition. We evaluate COMCA on two publicly available benchmarks, OVAD [5] and VAW [39]. COMCA demonstrates significant improvements over training-free baselines and is competitive with recent training-based approaches, while remaining unaffected by the biases of specific training sets. Notably, COMCA provides a consistent boost in performance regardless of the VLM pre-training and architecture." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.69, + 0.486, + 0.901 + ], + "angle": 0, + "content": "In summary, our contributions are: ① We propose COMCA, an approach specifically designed for training-free open-vocabulary attribute detection. COMCA exploits the compositional nature of objects and attributes to build an auxiliary cache that is used as an anchor to refine the model's predictions. ② We make caching scalable by extracting priors from databases and LLMs on which objects are associated with each attribute, using them to populate the cache. Soft-labeling further accounts for the multiplicity of attributes within an image. ③ We empirically show that COMCA outperforms all the cache-based baselines and achieves competitive performance against costly training-based methods, demonstrating the practical feasibility of training-free open vocabulary attribute detection." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.091, + 0.652, + 0.105 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.115, + 0.908, + 0.19 + ], + "angle": 0, + "content": "Attribute detection. Detecting visual attributes is a widely explored task in computer vision [4, 11-13], supporting various downstream applications such as zero-shot learning [1, 26, 27, 51], relative image comparison [8, 37, 46, 53], and compositional image retrieval [3, 19, 24, 44]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.192, + 0.909, + 0.327 + ], + "angle": 0, + "content": "Recent works have introduced several benchmarks to promote and ease research in this area [18, 21, 25, 38-40, 49, 50, 53]. Despite these advances, annotating visual attributes remains a complex and challenging task. This is due to the wide variability in attributes across objects and contexts, as well as the difficulty of ensuring consistent annotations when multiple annotators are involved. Moreover, many datasets capture only a limited range of object attributes, limiting their usefulness for developing general-purpose models." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.328, + 0.909, + 0.705 + ], + "angle": 0, + "content": "Open-vocabulary attribute detection (OVAD) [5, 7] overcomes some of these limitations by detecting user-specified attributes, typically provided in natural language, even those that have not been observed during training. This task was first explored in [5], which introduced the OVAD benchmark and method. OVAD generates supervisory signals for attributes, nouns, and noun phrases by parsing captions and encoding them using CLIP text embeddings [41]. OvarNet [7] builds on this by fine-tuning CLIP with federated learning and additional supervision from external captions. LOWA [16] proposes a three-stage pipeline that aligns visual embeddings to objects, attributes, and free-form text. ArtVLM [63] uses a generative VLM to obtain attribute probabilities based on the input image. Although these methods operate in the open-vocabulary setting, they are training-based and require both a lot of annotated data and several computational resources (see Sec. 4). COMCA significantly differs from these approaches, as we explore for the first time a training-free, cache-based, strategy for attribute detection. Adapting Vision-Language Models. VLMs [28, 29, 31, 41, 54] bridge the gap between textual and visual data, showing remarkable zero-shot generalization abilities. Recent efforts have focused on improving their performance on specific downstream tasks through various strategies, broadly categorized into training-based and training-free methods." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.909, + 0.811 + ], + "angle": 0, + "content": "Training-based approaches typically involve tuning text prompts [22, 47, 52, 59, 62]. Some methods learn prompts directly [60] or predict them [59] from visual inputs to improve generalization. Other approaches train adapters to refine the model's predictions [15, 20]. While effective, these methods introduce additional parameters, which can negatively impact the model's zero-shot performance [59]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.911, + 0.902 + ], + "angle": 0, + "content": "Training-free methods take a different approach, such as incorporating additional knowledge into the prompts using natural language [32, 34, 43], or modifying predictions through attention layers [17] or normalization techniques [61]. Recent works also employ a cache of images and one-hot labels that represent the target classes [48, 57," + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15099" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.089, + 0.905, + 0.288 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.3, + 0.905, + 0.342 + ], + "angle": 0, + "content": "Figure 2. COMCA's cache construction. Given a list of attributes and objects, we compute their compatibility from a large database \\( D_r \\) and with an LLM. The scores are merged and normalized to obtain the compatibility distribution, from which we sample cache entries and construct the cache. We enrich the latter with soft labels from the VLM-based similarity between cache images and attributes." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.351, + 0.486, + 0.534 + ], + "angle": 0, + "content": "[64]. For example, TIP-Adapter [57] computes the similarity between input images and cached examples, using the associated labels for prediction. Similarly, SuS-X [48] creates a cache automatically, either by using a generative model [42] or retrieving from a large database [45], with distance-based scoring to weight the elements' contribution. While our work builds upon the same principles of TIP-Adapter and SuS-X, i.e., using a cache to refine predictions, it differs in four key aspects, as COMCA's cache: (i) is aware of attribute-object compatibility; (ii) accounts for attribute combinations; (iii) models the compositional nature of attributes and objects; (iv) scales efficiently with the number of attributes." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.549, + 0.315, + 0.567 + ], + "angle": 0, + "content": "3. Compositional Caching" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.576, + 0.486, + 0.684 + ], + "angle": 0, + "content": "In this section, we formally define the open-vocabulary attribute detection problem and describe caching as a tool for VLM adaptation (Sec. 3.1). Next, we describe how attribute-object compatibility informs the construction of scalable caches (Sec. 3.2). Finally, we detail our process for refining cache labels, taking into account multi-attribute compatibility (Sec. 3.3). Fig. 2 provides an overview of our approach." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.695, + 0.224, + 0.71 + ], + "angle": 0, + "content": "3.1. Background" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.718, + 0.483, + 0.869 + ], + "angle": 0, + "content": "Problem formulation. Given an image \\( x \\in \\mathcal{X} \\) of an object and a list of attributes, our goal is to detect which attributes are present in the image. Let \\( \\mathcal{A} \\) be the set of all existing attributes in natural language. At inference time, the model is provided with a set of \\( n \\) target attributes \\( A = \\{a_{1},\\dots ,a_{n}\\} \\) as input, where \\( A \\subset \\mathcal{A} \\). These are associated with \\( m \\) target objects, which we denote as \\( O = \\{o_1,\\dots ,o_m\\} \\). Our goal is to build a function \\( f:\\mathcal{X}\\times \\mathcal{A}^n\\to [-1,1]^n \\) that maps an image and a list of attributes to a binary vector indicating the presence (1) or absence (-1) of each attribute." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "We obtain \\( f \\) without training, in contrast to previous works that learn \\( f \\) using a dataset \\( D \\) of image-label pairs" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.351, + 0.908, + 0.396 + ], + "angle": 0, + "content": "\\((x, \\mathbf{a})\\) with \\(\\mathbf{a} \\in [-1, 1]^n\\) being the ground-truth binary vector1. We achieve this by implementing \\(f\\) using a VLM combined with an auxiliary cache to refine its predictions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.397, + 0.909, + 0.593 + ], + "angle": 0, + "content": "Adapting VLMs with caching. Following previous works [5, 7], we implement \\( f \\) with a VLM. While we primarily consider CLIP [41] to describe our method and perform our evaluation, the proposed framework can integrate several other VLMs (see Tab. 2). The function \\( f \\) has three elements: a visual encoder \\( f_{v} \\), a text encoder \\( f_{t} \\), and a prediction function. \\( f_{v} \\) and \\( f_{t} \\) map images in \\( \\mathcal{X} \\) and natural language in \\( \\mathcal{T} \\), respectively, to a shared \\( d \\)-dimensional multimodal space \\( \\mathcal{Z} \\in \\mathbb{R}^d \\), i.e., \\( f_{v}: \\mathcal{X} \\to \\mathcal{Z} \\) and \\( f_{t}: \\mathcal{T} \\to \\mathcal{Z} \\). Next, the prediction function maps embeddings from the multimodal space to the final scores. A standard way for scoring an image \\( x \\) against a class \\( c \\) is by computing the cosine similarity between their embeddings:\\(^2\\)" + }, + { + "type": "equation", + "bbox": [ + 0.598, + 0.604, + 0.907, + 0.639 + ], + "angle": 0, + "content": "\\[\nf _ {\\mathrm {C L I P}} (x, c) = \\frac {f _ {v} (x) ^ {\\top} f _ {t} (c)}{| | f _ {v} (x) | | \\cdot | | f _ {t} (c) | |}. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.647, + 0.909, + 0.724 + ], + "angle": 0, + "content": "While effective, \\( f \\) can be improved without any training by using an auxiliary cache that stores examples of the classes of interest [36, 48, 57]. Given a cache \\( \\mathcal{C} = \\{x_1^1, \\dots, x_C^K\\} \\) that contains \\( K \\) samples of \\( C \\) classes, we can define the cache-based prediction for a class \\( c \\in C \\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.564, + 0.734, + 0.907, + 0.775 + ], + "angle": 0, + "content": "\\[\nf _ {\\mathcal {C}} (x, c) = \\sum_ {i = 1} ^ {K} \\eta_ {\\mathcal {C}} \\left(\\frac {f _ {v} (x) ^ {\\intercal} f _ {v} \\left(x _ {c} ^ {i}\\right)}{\\left| \\left| f _ {v} (x) \\right| \\right| \\cdot \\left| \\left| f _ {v} \\left(x _ {c} ^ {i}\\right) \\right| \\right|}\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.787, + 0.908, + 0.818 + ], + "angle": 0, + "content": "where \\(\\eta_{\\mathcal{C}}(z) = \\exp (1 - \\beta (-z))\\) is a normalization function regulated by the hyperparameter \\(\\beta\\) [57]." + }, + { + "type": "page_footnote", + "bbox": [ + 0.512, + 0.827, + 0.908, + 0.876 + ], + "angle": 0, + "content": "1In some problem formulations, the dataset may contain (i) object labels, (ii) bounding boxes, (iii) an \"unknown\" value for some attributes. For simplicity, we omit the latter from the formulation and, following previous works [39], we consider the case where objects bounding boxes are given." + }, + { + "type": "page_footnote", + "bbox": [ + 0.515, + 0.877, + 0.905, + 0.9 + ], + "angle": 0, + "content": "2For simplicity, we omit the presence of prompts such as \"a photo of\" when defining the text input." + }, + { + "type": "list", + "bbox": [ + 0.512, + 0.827, + 0.908, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15100" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.334 + ], + "angle": 0, + "content": "The formulation in Eq. (2) has a main issue when applied to open-vocabulary attribute detection: the cache construction. On one hand, a cache considering all possible attribute labels is semantically complete but requires a prohibitive storage size of \\( K \\cdot n! \\) images, e.g., \\( K \\cdot 117! \\) for OVAD [5]. On the other hand, while storing \\( K \\) positive examples per attribute has low memory requirements, it ignores the multi-attribute nature of images (e.g., shape, size, color, etc. can co-exist in the same sample), potentially leading to incorrect predictions (e.g., a high score because the color matches but the cache label refers to the shape). Moreover, attributes may appear differently across objects (e.g., wet car vs wet dog) and selecting a random subset of them may lead to a cache representative only of a few cases. In the following, we address these issues by focusing on the compatibility between objects and multiple attributes in the real world." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.343, + 0.391, + 0.359 + ], + "angle": 0, + "content": "3.2. Compatibility for scalable caching" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.365, + 0.484, + 0.546 + ], + "angle": 0, + "content": "Naively constructing the cache would lead to many problems, related both to labels' semantic and memory requirements (Sec. 3.1). In this section, we present a methodology that (i) has the same memory requirements as storing only \\( K \\) examples per attribute and (ii) assigns multiple attribute labels per example, taking into account attribute-object compatibility. Scalable caching. To obtain a cache \\( \\mathcal{C} \\) that is scalable with the number of attributes and accounts for object-specific information, we exploit the natural distributions of attributes across objects. In particular, let us define a function \\( \\varphi \\) that, given an attribute \\( a \\in \\mathcal{A} \\) and an object \\( o \\in \\mathcal{O} \\), returns their compatibility value as a scalar, i.e., \\( \\varphi : \\mathcal{A} \\times \\mathcal{O} \\to \\mathbb{R} \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.548, + 0.484, + 0.653 + ], + "angle": 0, + "content": "Considering an attribute \\( a \\) and an object \\( o \\), the score \\( \\varphi(a, o) \\) is proportional to the probability that \\( o \\) exhibits \\( a \\). Thus, we can use the scores \\( \\Phi_O(a) = [\\varphi(a, o_1), \\dots, \\varphi(a, o_m)] \\) as a proxy for the probability distribution of \\( a \\) co-occurring with the elements of \\( O \\). We then assign each of the \\( K \\) elements (i.e., shots) of the attribute \\( a \\) to one object \\( o \\), sampling the object from \\( \\Phi_O(a) \\):" + }, + { + "type": "equation", + "bbox": [ + 0.14, + 0.662, + 0.484, + 0.683 + ], + "angle": 0, + "content": "\\[\n\\hat {O} _ {a} ^ {K} = \\left\\{\\hat {o} _ {1} ^ {a}, \\dots , \\hat {o} _ {K} ^ {a} \\right\\}, \\text {w h e r e} \\hat {o} _ {i} ^ {a} \\sim \\tilde {\\Phi} _ {O} (a), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.693, + 0.484, + 0.768 + ], + "angle": 0, + "content": "where \\(\\tilde{\\Phi}_O(a)\\) is the normalized version of \\(\\Phi_O(a)\\), i.e., \\(\\tilde{\\Phi}_O(a) = \\Phi_O(a) / \\sum \\Phi_O(a)\\). The set \\(\\hat{O}_a^K\\) allows populating the cache with relevant objects and attributes. Let us denote a generic function mapping a textual query to an image as \\(\\mathrm{T2I}:\\mathcal{T}\\to \\mathcal{X}\\). The elements of the cache are:" + }, + { + "type": "equation", + "bbox": [ + 0.117, + 0.777, + 0.484, + 0.814 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {C} _ {\\mathrm {H}} = \\left\\{x _ {a _ {1}} ^ {1}, \\dots , x _ {a _ {m}} ^ {K} \\right\\} (4) \\\\ = \\{\\mathrm {T 2 I} \\left(\\pi_ {g} \\circ a \\circ o _ {1} ^ {a}\\right), \\dots , \\mathrm {T 2 I} \\left(\\pi_ {g} \\circ a \\circ o _ {K} ^ {a}\\right) \\}, (4) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.484, + 0.901 + ], + "angle": 0, + "content": "where \\(\\pi_g\\) is a contextual prompt for the image (e.g., \"a photo of\"). Following SuS-X [48], we implement T2I via retrieval. Specifically, we retrieve images from a database by scoring them against textual queries with the VLM (e.g., CLIP). While we focus on retrieval, COMCA" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.197 + ], + "angle": 0, + "content": "also supports the implementation of T2I via a text-to-image generative model, e.g., by feeding textual queries to Stable Diffusion [42]. We refer to the Supp. Mat. for more details. Estimating attribute-object compatibility. The key element that enables scalable caching is the function \\(\\varphi\\) that estimates the viability of attribute-object compositions. We implement it by merging two scoring strategies." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.198, + 0.907, + 0.258 + ], + "angle": 0, + "content": "The first uses a generic large-scale database \\( D_r \\) containing image-caption pairs. Using the captions \\( t \\in D_r \\), we can estimate the database score as the number of co-occurrences of attribute \\( a \\) and object \\( o \\), i.e.:" + }, + { + "type": "equation", + "bbox": [ + 0.58, + 0.268, + 0.907, + 0.303 + ], + "angle": 0, + "content": "\\[\n\\varphi_ {\\mathrm {d b}} (a, o) = \\sum_ {(x, t) \\in D _ {r}} \\mathbb {1} (o \\subset t \\wedge a \\subset t), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.314, + 0.907, + 0.343 + ], + "angle": 0, + "content": "where \\(\\mathbb{1}(z)\\) is an indicator function whose value is 1 whenever the condition \\(z\\) is true, and 0 otherwise." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.344, + 0.906, + 0.435 + ], + "angle": 0, + "content": "While this process exploits real data, it may be biased on the specific captions and/or data collection process. For instance, captions rarely describe objects in their default state, such as a dry dog. Therefore, a second scoring strategy exploits the reasoning capabilities of an LLM, querying it to predict an attribute-object compatibility as:" + }, + { + "type": "equation", + "bbox": [ + 0.618, + 0.447, + 0.907, + 0.463 + ], + "angle": 0, + "content": "\\[\n\\varphi_ {\\mathrm {L L M}} (a, o) = \\mathrm {L L M} \\left(\\pi_ {c}, o, a\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.474, + 0.906, + 0.519 + ], + "angle": 0, + "content": "where \\(\\pi_c\\) is a prompt specifying the compatibility task (see the Supp. Mat.). The objective is to assign a score on a specific range to the likelihood of the attribute-object pair." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.52, + 0.906, + 0.565 + ], + "angle": 0, + "content": "Taking into account that also LLMs have issues, such as hallucinations and biases [14, 33], we implement \\(\\varphi\\) by combining database- and LLM-based scoring as:" + }, + { + "type": "equation", + "bbox": [ + 0.601, + 0.576, + 0.907, + 0.593 + ], + "angle": 0, + "content": "\\[\n\\varphi (a, o) = \\varphi_ {\\mathrm {d b}} (a, o) \\cdot \\varphi_ {\\mathrm {L L M}} (a, o), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.604, + 0.887, + 0.62 + ], + "angle": 0, + "content": "thus computing \\(\\Phi_O(a)\\) and filling the cache accordingly." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.628, + 0.796, + 0.644 + ], + "angle": 0, + "content": "3.3. Compositional cache refinement" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.65, + 0.907, + 0.801 + ], + "angle": 0, + "content": "After the process in Sec. 3.2, we have a cache \\(\\mathcal{C}_{\\mathbb{H}}\\) populated with \\(K\\) examples for each attribute \\(a \\in A\\). However, the cache assigns one hard attribute to each element, contrasting with the natural co-occurrence of attributes and leading to potentially wrong associations during inference. As a simple example, an image of a large and blue car obtained with the binding (large, car) will have a positive label only for large. As a consequence, if the test image shows a small blue car, the cache element will not contribute to the score of blue and may erroneously increase the score for the antonym large." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.802, + 0.907, + 0.861 + ], + "angle": 0, + "content": "Soft caching. To remedy to this issue, we refine the label associated to each cache element, exploiting the VLM capabilities to compute soft labels. Formally, the soft label \\( s_a \\) for an attribute \\( a \\) and a cache image \\( x_c \\) is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.606, + 0.871, + 0.907, + 0.906 + ], + "angle": 0, + "content": "\\[\ns _ {a} \\left(x _ {c}\\right) = \\frac {f _ {t} (a) ^ {\\intercal} f _ {v} \\left(x _ {c}\\right)}{\\left| \\left| f _ {t} (a) \\right| \\right| \\cdot \\left| \\left| f _ {v} \\left(x _ {c}\\right) \\right| \\right|}, \\tag {8}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "15101" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.182 + ], + "angle": 0, + "content": "where we compare the text embeddings of the attribute with the visual embeddings of the cache image. While these scores can directly serve as soft labels, we found them to exhibit low-variance due to the well-known modality-gap [30, 48]. We overcome this limitation and increase their relevance by defining normalized soft labels \\(\\bar{s}_a\\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.143, + 0.193, + 0.484, + 0.228 + ], + "angle": 0, + "content": "\\[\n\\bar {s} _ {a} \\left(x _ {c}\\right) = \\frac {\\exp \\left(\\left(s _ {a} \\left(x _ {c}\\right) - \\mu_ {C}\\right) / \\sigma_ {C}\\right)}{\\sum_ {\\hat {a} \\in A} \\exp \\left(\\left(s _ {\\hat {a}} (x) - \\mu_ {C}\\right) / \\sigma_ {C}\\right)}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.238, + 0.485, + 0.374 + ], + "angle": 0, + "content": "where \\(\\mu_{C}\\) and \\(\\sigma_{C}\\) are respectively the mean and standard deviation of the similarity between any attribute with any element of the cache. The first normalization expands the range of values of the similarity scores \\(s_a\\), while the softmax ensures that all samples have an equal contribution to the final prediction. In practice, we found the latter step to be important only on datasets with many attributes (e.g., VAW). We hypothesize this is due to differences in the output space cardinality, requiring different softmax scaling as in [58]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.375, + 0.484, + 0.435 + ], + "angle": 0, + "content": "These scores are mixed with the original hard labels to obtain the final label \\(\\hat{s}_a(x_c) = (1 - \\alpha)\\cdot \\mathbb{1}(a = c) + \\alpha \\cdot \\overline{s}_a(x_c)\\) where \\(\\alpha\\) is a trade-off hyperparameter. With the soft labels \\(\\hat{s}_a\\), we can perform the cache-based prediction as:" + }, + { + "type": "equation", + "bbox": [ + 0.102, + 0.447, + 0.484, + 0.5 + ], + "angle": 0, + "content": "\\[\nf _ {\\mathcal {C}} ^ {\\mathrm {C O M C A}} (x, a) = \\sum_ {x _ {c}} ^ {\\mathcal {C}} \\eta_ {\\mathcal {C}} \\left(\\hat {s} _ {a} \\left(x _ {c}\\right) \\frac {f _ {v} (x) ^ {\\intercal} f _ {v} \\left(x _ {c}\\right)}{\\left\\| f _ {v} (x) \\right\\| \\cdot \\left\\| f _ {v} \\left(x _ {c}\\right) \\right\\|}\\right). \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.502, + 0.484, + 0.562 + ], + "angle": 0, + "content": "In contrast to Eq. (2), here we consider each cache element \\( x_{c} \\) as a potential contributor for attribute \\( a \\), even if \\( a \\neq c \\). Implicitly, this allows us to include an arbitrary number of (positive) attribute labels within each element of the cache." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.563, + 0.484, + 0.607 + ], + "angle": 0, + "content": "Final inference. Finally, the cache-based scoring defined in Eq. (10) is merged with the standard CLIP-based scores to obtain the final predictions:" + }, + { + "type": "equation", + "bbox": [ + 0.114, + 0.619, + 0.483, + 0.639 + ], + "angle": 0, + "content": "\\[\nf (x, a) = \\eta_ {A} \\left(\\lambda \\cdot f _ {\\mathcal {C}} ^ {\\mathrm {C O M C A}} (x, a) + f _ {\\mathcal {C}} ^ {\\mathrm {C L I P}} (x, a)\\right), \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.65, + 0.484, + 0.77 + ], + "angle": 0, + "content": "where \\(\\lambda\\) is an hyperparameter controlling the balance between the two and \\(\\eta_A(z) = \\mathrm{softmax}_A(z / \\max (z))\\), moving from simple scalar values to a distribution over the attributes space and leading to more stable scores. Note that Eq. (11) contains no element which needs training or labeled data for the specific attributes of interest. Given the target sets of attributes \\(A\\) and objects \\(O\\), COMCA automatically re-generates the cache to update the inference procedure." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.785, + 0.222, + 0.802 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Datasets. We evaluate COMCA on two datasets. OVAD [5] covers 80 classes and 117 attributes across 2,000 images from MS-COCO [38], for a total of 14,300 instances. Attributes are manually annotated as positive, negative, and unknown. VAW [39] has more concepts, i.e., 2,260 classes and 620 attributes across 10,392 images. VAW is more" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.122 + ], + "angle": 0, + "content": "sparse than OVAD, i.e., around 1.8 positive and 5.3 negative attributes per instance, and it is automatically annotated." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.123, + 0.907, + 0.213 + ], + "angle": 0, + "content": "Setting. Following prior works [5, 7], COMCA focuses on the box-given setting, where the bounding boxes of objects are given, as this allows us to assess its attribute detection capabilities more precisely. However, Tab. 3 also evaluates COMCA in the box-free setting by introducing an object detector. See the Supp. Mat. for more details on this setting." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.214, + 0.907, + 0.259 + ], + "angle": 0, + "content": "Metrics. We measure performance as mean Average Precision (mAP), and also report mAP scores on head, medium, and tail classes based on their frequency in the dataset." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.261, + 0.909, + 0.562 + ], + "angle": 0, + "content": "Baselines and competitors. We evaluate COMCA against training-based methods, i.e., OVAD [5], OvarNet [7], ArtVLM [63], and LOWA [16]. As we are the first to address the problem in a completely training-free fashion, we extend the baselines with several zero-shot methods [9, 28, 54-56], and adapt previous cache-based approaches to our task [48, 57]. For TIP-Adapter [57] and SuS-X [48] we construct a \\(K\\)-shot cache, considering only attributes for sampling and disregarding their co-occurrences with objects. This follows their approach, which samples \\(K\\) shots for each target class, i.e., attributes in our task. In addition, we test IAP [26, 27], which conditions predicted attributes on the inferred object category, using TIP-Adapter as the object classifier. We also consider a naive cache-based approach, i.e., \"image-based\", that uses CC12M [6] to build a cache specifically for each input image by retrieving the most similar samples. Since the retrieved images are unlabeled, we use our soft labels to annotate the cache. We construct all the proposed baselines on top of CLIP ViT-B/32 [9]. For additional details, please refer to the Supp. Mat." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.563, + 0.909, + 0.73 + ], + "angle": 0, + "content": "Implementation details. We use CLIP ViT-B/32 [9] pretrained on LAION2B [45] as our primary VLM. Following TIP-Adapter [57], we set \\(\\lambda = 1.17\\) in Eq. (11) and \\(\\beta = 1.0\\) in Eq. (2). We set \\(\\alpha = 0.6\\) to mix hard and soft labels on the validation set of VAW [39]. Lastly, we use CC12M [6] as the database in Eq. (5) and GPT 3.5 Turbo [35] as the LLM for Eq. (6), as we find it to perform better than GPT 4o-mini (see the Supp. Mat.). Unless otherwise stated, we use \\(K = 16\\) shots per attribute, retrieving them from CC12M. We use the same hyperparameters for both datasets. For the box-free experiments, we use YOLOv11M [23] as the detector." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.741, + 0.608, + 0.755 + ], + "angle": 0, + "content": "4.1. Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.764, + 0.907, + 0.84 + ], + "angle": 0, + "content": "We present our comprehensive experimental evaluation comparing COMCA with state-of-the-art training- and cache-based methods. We also show its performance with different backbones and in the box-free setting. Moreover, we discuss qualitative results and limitations of other approaches." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.841, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Comparison with the state of the art. Tab. 1 reports the results comparing zero-shot baselines, cache-based methods, and COMCA in the box-given setting. On both benchmarks, COMCA outperforms all the zero-shot and cache-based com" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15102" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.092, + 0.089, + 0.916, + 0.336 + ], + "angle": 0, + "content": "
MethodTrain sizeOVAD [5]VAW [39]
mAPHeadMediumTailmAPHeadMediumTail
Zero shotCLIP [9] RN5011.841.011.71.435.337.835.126.2
CLIP [9] ViT-B/3217.044.318.45.550.051.050.943.2
CLIP [9] ViT-L/1418.344.420.56.451.051.452.545.0
Cache-basedImage-based16.845.819.23.552.953.753.846.4
TIP-Adapter [57] + IAP [27]15.143.317.81.727.732.425.516.6
TIP-Adapter [57]16.744.419.73.157.557.559.451.2
SuS-X [48]20.248.924.64.530.233.828.820.3
ComCA27.454.334.69.058.158.259.951.7
Training basedOVAD [5]110k21.448.026.95.2----
OvarNet [7]190k28.658.635.59.568.5---
ArtVLM [63]N/A----71.975.072.159.4
LOWA [16]1.33M18.758.020.42.642.646.441.032.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.346, + 0.908, + 0.389 + ], + "angle": 0, + "content": "Table 1. Comparison with state of the art. Green indicates COMCA. Bold indicates best among training-free methods. The symbol “-” indicates results for the competitors are not available on the original papers or the impossibility to run their method due to lack of public code and/or model weights. See the Supp. Mat. for an extended version of the table." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.397, + 0.482, + 0.563 + ], + "angle": 0, + "content": "petitors, demonstrating the effectiveness of our specialized design of the cache. It scores an increment of \\(+10.6\\mathrm{mAP}\\) on OVAD and \\(+5.2\\) on VAW compared to the image-based baseline. Similarly, it surpasses TIP-Adapter [57] on both benchmarks by \\(+10.7\\mathrm{mAP}\\) and \\(+0.6\\), respectively. SuS-X [48] increases performance on OVAD while decreasing it on VAW. Overall, COMCA has a gain of \\(+7.2\\mathrm{mAP}\\) and \\(+27.9\\) on the two datasets. Lastly, TIP-Adapter + IAP [27] demonstrates that naively using co-occurrences of attributes with objects [27] is insufficient to address this task. COMCA surpasses it by \\(+12.3\\mathrm{mAP}\\) on OVAD and \\(+30.4\\) on VAW." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.565, + 0.483, + 0.626 + ], + "angle": 0, + "content": "We also compare COMCA with training-based methods. Notably, COMCA outperforms LOWA [16] on both benchmarks, despite the latter having been trained on a large-scale dataset while our approach requires no training data." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.627, + 0.483, + 0.763 + ], + "angle": 0, + "content": "Experiments with multiple VLMs. To demonstrate that COMCA is model-agnostic, we test it on various backbones and VLMs [9, 28, 54-56] and report the results in Tab. 2. On both OVAD and VAW, COMCA boosts performance consistently for all the backbones up to \\(+11.5\\mathrm{mAP}\\) on OVAD with CoCa ViT-L/14 [54] and \\(+15.8\\mathrm{mAP}\\) on VAW with CLIP ResNet50 (RN50) [9]. On average, COMCA improves the performance of the original VLM by \\(+8.9\\) on OVAD and \\(+7.0\\) on VAW, showing COMCA has consistent advantages." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.765, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Cross-datasets advantages. To demonstrate the generalization capabilities of our method, we report cross-dataset results in Fig. 4. We compare COMCA with OVAD when trained on various datasets (i.e., VAW [39], OVAD [5], MSCOCO [38]) and tested on the OVAD (left) and on the VAW (right) benchmarks. As shown in Fig. 4, performance significantly degrades when OVAD is tested on datasets different from the training one (yellow bars vs red bar), whereas our training-free COMCA achieves competitive performance" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.397, + 0.905, + 0.442 + ], + "angle": 0, + "content": "(green bars) with the same RN50 backbone. This reinforces our training-free method, which does not inherit biases from a training set, differently from training-based approaches." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.445, + 0.906, + 0.687 + ], + "angle": 0, + "content": "Box-free setting. In this work, we primarily focus on attribute recognition and therefore adopt the box-given setting. Here, we evaluate COMCA in the box-free setting to demonstrate that it achieves competitive performance in a less constrained scenario. Tab. 3 shows the results for zero-shot and cache-based methods, Specifically, we use CLIP ViT-B/32 [9] and YOLOv11M [23] as detector for all models. We note that all cache-based methods apart from SuS-X underperform the zero-shot baseline, with a decrease of up to \\(-3.1\\mathrm{mAP}\\) for image-based. Only, SuS-X has a gain of \\(+2.8\\mathrm{mAP}\\) w.r.t. CLIP ViT-B/32, with our method COMCA being superior and surpassing CLIP ViT-B/32 by \\(+9.0\\mathrm{mAP}\\). Notably, our model is competitive with the training-based ones, surpassing the performance of both OVAD \\((+2.7\\mathrm{mAP})\\) and LOWA \\((+2.8\\mathrm{mAP})\\). We provide a more detailed analysis and an extended version of these results in the Supp. Mat." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Qualitative results. Fig. 3, we show qualitative examples comparing training-based OVAD [5], CLIP ViT-B/32 [9], and COMCA with the same backbone. We consider the top-1 prediction for all methods. OVAD correctly identifies the bear's color but misclassifies the cat's color and wrongly predicts a \"tiled\" clothes pattern for two people. We speculate it likely predicts the same attribute for two objects (i.e., the \\(2^{nd}\\) and the \\(6^{th}\\)) due to biases acquired during training. CLIP makes more errors, such as labeling the zebra as \"reflective\" and the airplane as \"tiny\", while misclassifying the baseball player as \"giant\". COMCA instead understands attributes better, for instance by accurately identifying the zebra's texture. These issues are less pronounced in COMCA thanks to our caching mechanism, which allows it to compare inputs" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.956 + ], + "angle": 0, + "content": "15103" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.089, + 0.482, + 0.317 + ], + "angle": 0, + "content": "
MethodOVAD [5]VAW [39]
BackboneCOMCAmAPΔmAPΔ
CLIP RN50 [9]11.835.3
✓19.2+7.451.1+15.8
CLIP ViT-B/32 [9]17.050.0
✓27.4+10.458.1+8.1
CLIP ViT-L/14 [9]18.351.0
✓24.8+6.561.0+10.0
SigLIP ViT-B/16 [56]13.752.5
✓22.1+8.459.8+7.3
CoCa ViT-B/32 [54]15.149.6
✓24.4+9.351.4+1.8
CoCa ViT-L/14 [54]14.449.8
✓25.9+11.554.5+4.7
BLIP [28]15.951.0
✓22.7+6.856.0+5.0
X-VLM [55]17.454.8
✓28.4+11.057.7+2.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.328, + 0.484, + 0.384 + ], + "angle": 0, + "content": "Table 2. Box-given results with different backbones. Green indicates COMCA applied on top of the backbone. \\(\\Delta\\) indicates improvements w.r.t. the corresponding baseline. See the Supp. Mat. for an extended version of the table." + }, + { + "type": "table", + "bbox": [ + 0.513, + 0.089, + 0.9, + 0.301 + ], + "angle": 0, + "content": "
MethodOVAD [5]
mAPHeadMediumTail
Zero shotCLIP [9] RN5012.340.212.71.9
CLIP [9] ViT-B/3212.539.612.52.7
CLIP [9] ViT-L/1413.439.714.32.9
Cache-basedImage-based9.435.48.71.0
TIP [57] + IAP [27]9.836.99.21.0
TIP-Adapter [57]10.136.09.81.2
SuS-X [48]15.343.017.23.2
COMCA21.548.426.46.0
Training basedOVAD [5]18.847.722.04.6
OvarNet [7]27.256.833.68.9
LOWA [16]18.758.020.42.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.509, + 0.31, + 0.905, + 0.354 + ], + "angle": 0, + "content": "Table 3. Results in the box-free setting. Green indicates COMCA. Bold indicates the best results among training-free methods. See the Supp. Mat. for an extended version of the table." + }, + { + "type": "image", + "bbox": [ + 0.136, + 0.39, + 0.247, + 0.476 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.098, + 0.479, + 0.134, + 0.516 + ], + "angle": 0, + "content": "OVAD CLIP COMCA" + }, + { + "type": "image_caption", + "bbox": [ + 0.139, + 0.478, + 0.243, + 0.508 + ], + "angle": 0, + "content": "brown brown single, one, individual" + }, + { + "type": "image", + "bbox": [ + 0.248, + 0.391, + 0.357, + 0.476 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.254, + 0.479, + 0.352, + 0.508 + ], + "angle": 0, + "content": "clothes pattern: tiled male neutral, calm, serious" + }, + { + "type": "image", + "bbox": [ + 0.358, + 0.391, + 0.47, + 0.476 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.36, + 0.479, + 0.468, + 0.508 + ], + "angle": 0, + "content": "plaid, tartan, checkered reflective striped, lined, pinstriped" + }, + { + "type": "image", + "bbox": [ + 0.471, + 0.39, + 0.681, + 0.476 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.537, + 0.479, + 0.614, + 0.507 + ], + "angle": 0, + "content": "large, giant, huge small, little, tiny smooth, sleek" + }, + { + "type": "image", + "bbox": [ + 0.683, + 0.391, + 0.793, + 0.476 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.7, + 0.479, + 0.774, + 0.508 + ], + "angle": 0, + "content": "yellow piece, cut small, little, tiny" + }, + { + "type": "image", + "bbox": [ + 0.795, + 0.391, + 0.905, + 0.476 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.801, + 0.479, + 0.898, + 0.508 + ], + "angle": 0, + "content": "clothes pattern: tiled big, large, giant, huge sitting, sit" + }, + { + "type": "image_caption", + "bbox": [ + 0.091, + 0.521, + 0.908, + 0.536 + ], + "angle": 0, + "content": "Figure 3. Qualitative results. Predictions of OVAD, CLIP and ComCA on some OVAD images. Green are correct ones, red are wrong." + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.553, + 0.473, + 0.673 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.685, + 0.486, + 0.755 + ], + "angle": 0, + "content": "Figure 4. Cross-dataset results. Results on OVAD (left) and VAW (right). Yellow indicates methods tested cross-domain, i.e., train and test set differ. Red indicates methods tested in-domain, and green our training-free method, which does not suffer from domain shift problems. All experiments use a RN50 backbone." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.764, + 0.334, + 0.779 + ], + "angle": 0, + "content": "to self-labeled examples at test time." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.789, + 0.241, + 0.804 + ], + "angle": 0, + "content": "4.2. Ablation study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.484, + 0.901 + ], + "angle": 0, + "content": "In this section, we analyze COMCA's components, studying how sampling strategies affect performance (Tab. 4), soft labeling's impact (Tab. 5 and Fig. 6), and the number of samples in the cache (Fig. 5). Additional experiments in the Supp. Mat. show the effect of (i) the hyperparameters \\(\\alpha\\) and \\(\\lambda\\), (ii) constructing the cache with Stable Diffusion, and (iii)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.552, + 0.779, + 0.567 + ], + "angle": 0, + "content": "using other LLMs for generative scores." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.568, + 0.908, + 0.825 + ], + "angle": 0, + "content": "Cache construction. In Tab. 4, we evaluate five strategies to construct the cache. The first, Random, samples \\( K \\) objects for each attribute and creates a cache of \\( |A| \\times K \\) examples, seeking only a match on the attribute. The second, Brute force, considers all possible combinations of attributes and objects, resulting in a large cache of \\( |O| \\times |A| \\times K \\) samples. As VAW contains many attributes and objects, the Brute force cache would be too large and cannot be practically tested. As shown in Tab. 4, both approaches lead to very low results on OVAD. Compared to the zero-shot CLIP baseline, performance is similar (Random, -0.3 mAP) or even below (Brute force, -6.4 mAP). We hypothesize that Random ignores attribute-object relationships, while Brute force introduces many noisy and potentially detrimental samples, harming performance. On the contrary, Random improves by +7.5 on VAW. We hypothesize this is due to the nature of the dataset, which contains many attributes and objects." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.909, + 0.901 + ], + "angle": 0, + "content": "When introducing priors on object-attributes compositions via a database (Eq. (5)) or an LLM (Eq. (6)), performance largely increases by up to \\(+4.1\\) mAP on VAW w.r.t. the baseline for database scores. ComCA combines the two scores to obtain object priors, achieving the best perfor" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15104" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.089, + 0.368, + 0.219 + ], + "angle": 0, + "content": "
ConfigurationmAP
OVAD [5]VAW [39]
Zero-shot [9]17.050.0
Random16.757.5
Brute force10.6-
DB scores18.654.1
LLM scores21.155.4
DB + LLM scores26.455.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.227, + 0.37, + 0.296 + ], + "angle": 0, + "content": "Table 4. Ablation on cache construction. Green indicates our default configuration. Bold indicates the best results. Results are without soft labels. See the Supp. Mat. for an extended version of the table." + }, + { + "type": "table", + "bbox": [ + 0.382, + 0.088, + 0.655, + 0.223 + ], + "angle": 0, + "content": "
ConfigurationmAP
OVAD [5]VAW [39]
Zero-shot [9]17.050.0
One-hot26.455.2
Soft26.753.2
- Sharpening [2]21.527.7
- Softmax24.558.5
- Eq. (9)27.458.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.377, + 0.232, + 0.657, + 0.287 + ], + "angle": 0, + "content": "Table 5. Ablation on soft labels. Green indicates our default configuration. Bold indicates the best results. See the Supp. Mat. for an extended version of the table." + }, + { + "type": "image", + "bbox": [ + 0.674, + 0.096, + 0.895, + 0.215 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.665, + 0.232, + 0.908, + 0.302 + ], + "angle": 0, + "content": "Figure 5. Ablation on number of samples per attribute. The purple line is mAP on OVAD, while the red line is mAP on VAW. Green indicates our default configuration." + }, + { + "type": "image", + "bbox": [ + 0.098, + 0.329, + 0.475, + 0.41 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.424, + 0.483, + 0.48 + ], + "angle": 0, + "content": "Figure 6. Soft labels across cache priors. Performance on OVAD (left) and VAW (right), considering generative (LLM) and retrieval (CC3M, CC12M) scores. Yellow indicates one-hot labels, red soft labels and green COMCA, combining LLM and CC12M." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.5, + 0.483, + 0.546 + ], + "angle": 0, + "content": "mance. These results confirm both (i) that attribute-object compatibility is important to build the cache, and (ii) that databases and LLMs provide complementary information." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.547, + 0.484, + 0.623 + ], + "angle": 0, + "content": "Soft labels. As discussed in Sec. 3.3, each image in the cache contains multiple attributes, even if we fill the cache considering only one of them. Here we analyze the soft labels obtained via Eq. (8), comparing them with the One-hot attribute labels of the initial cache, reporting the results in Tab. 5." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.625, + 0.484, + 0.731 + ], + "angle": 0, + "content": "The first strategy is to incorporate the soft scores as they are predicted (Soft in Tab. 5). This leads to results slightly above the One-hot baselines on OVAD \\((+0.3\\mathrm{mAP})\\), while reducing the performance in VAW \\((-2.0\\mathrm{mAP})\\). This is a direct consequence of allowing each sample to have a different level of relevance in the cache, something that becomes more prominent in VAW due to the larger number of attributes." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.733, + 0.483, + 0.869 + ], + "angle": 0, + "content": "To fix this issue, we can normalize the scores. We study three strategies: Sharpening, where the scores are normalized as in PAWS [2]; Softmax normalization over each sample; and the one we propose in Eq. (9), i.e., using the statistics of the cache-attributes similarity. Sharpening is not effective in this scenario, achieving the lowest performance across both settings. Softmax normalization leads to good results on VAW, e.g., +3.3 mAP w.r.t. One-hot, but it struggles on OVAD (24.5 mAP) due to the limited range of values." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Our strategy addresses both issues simultaneously, consistently improving metrics in both settings compared to" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.328, + 0.907, + 0.419 + ], + "angle": 0, + "content": "the One-hot strategy, scoring \\(27.4\\mathrm{mAP}\\) on OVAD and 58.1 mAP on VAW. Note that these findings are general. Fig. 6 shows soft labels consistently enhance results for estimating attribute-object priors across all strategies and datasets. Improvements range from \\(+3.8\\mathrm{mAP}\\) on VAW to \\(+5.4\\mathrm{mAP}\\) on OVAD in the LLM-only setting." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.42, + 0.909, + 0.556 + ], + "angle": 0, + "content": "Number of samples per attribute. In Fig. 5, we measure how adding more samples (varying \\( K \\)) per attribute influences performance on the two datasets. We observe performance consistently increasing for OVAD while going from 1 to 16 samples, reaching a plateau when moving to \\( K = 24 \\). On VAW, performance is stable for different values of \\( K \\) due to the large number of attributes in the dataset (620), making the cache representative enough even when \\( K = 1 \\). These results confirm the stability of COMCA." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.573, + 0.634, + 0.588 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.599, + 0.909, + 0.903 + ], + "angle": 0, + "content": "We introduced Compositional Caching for open-vocabulary attribute detection. COMCA uses a VLM and an auxiliary cache of images to solve this task in a training-free manner. Differently from previous works [48, 57], we construct the cache based on attribute-object compatibility derived from web-scale databases and LLMs. We also introduced a soft-labeling scheme to account for the co-occurrence of multiple attributes within an image. Combining these two components allows to model the natural co-occurrence of attributes in the real world, together and across objects, while making the cache construction scalable. Our experiments demonstrate that COMCA is a strong training-free approach. Since COMCA is training-free, it does not suffer from differences between training and test attributes, contrary to training-based methods. This also enables COMCA to be adapted to special domains (e.g., aerial images) more easily than training-based methods, which would require specific re-training. Future work will focus on devising methods to address attribute detection even in more challenging settings where objects or attributes are not given [10]." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15105" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.198 + ], + "angle": 0, + "content": "Acknowledgments. We acknowledge ISCRA for awarding this project access to the LEONARDO supercomputer, owned by the EuroHPC Joint Undertaking, hosted by CINECA (Italy). This work was sponsored by the MUR PNRR project FAIR - Future AI Research (PE00000013), funded by NextGeneration EU, and the EU projects AI4TRUST (No.101070190) and ELIAS (No.01120237)." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.211, + 0.188, + 0.226 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.236, + 0.486, + 0.276 + ], + "angle": 0, + "content": "[1] Zeynep Akata, Florent Perronnin, Zaid Harchaoui, and Cordelia Schmid. Label-embedding for attribute-based classification. In CVPR, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.278, + 0.485, + 0.348 + ], + "angle": 0, + "content": "[2] Mahmoud Assran, Mathilde Caron, Ishan Misra, Piotr Bojanowski, Armand Joulin, Nicolas Ballas, and Michael Rabbat. Semi-supervised learning of visual features by nonparametrically predicting view assignments with support samples. In ICCV, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.35, + 0.483, + 0.39 + ], + "angle": 0, + "content": "[3] Alberto Baldrati, Lorenzo Agnolucci, Marco Bertini, and Alberto Del Bimbo. Zero-shot composed image retrieval with textual inversion. In ICCV, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.392, + 0.483, + 0.432 + ], + "angle": 0, + "content": "[4] Tamara L Berg, Alexander C Berg, and Jonathan Shih. Automatic attribute discovery and characterization from noisy web data. In ECCV, 2010. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.434, + 0.483, + 0.475 + ], + "angle": 0, + "content": "[5] Maria A Bravo, Sudhanshu Mittal, Simon Ging, and Thomas Brox. Open-vocabulary attribute detection. In CVPR, 2023. 1, 2, 3, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.477, + 0.483, + 0.532 + ], + "angle": 0, + "content": "[6] Soravit Changpinyo, Piyush Sharma, Nan Ding, and Radu Soricut. Conceptual 12m: Pushing web-scale image-text pre-training to recognize long-tail visual concepts. In CVPR, 2021. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.534, + 0.485, + 0.587 + ], + "angle": 0, + "content": "[7] Keyan Chen, Xiaolong Jiang, Yao Hu, Xu Tang, Yan Gao, Jianqi Chen, and Weidi Xie. Ovarnet: Towards open-vocabulary object attribute recognition. In CVPR, 2023. 1, 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.59, + 0.483, + 0.617 + ], + "angle": 0, + "content": "[8] Steven Chen and Kristen Grauman. Compare and contrast: Learning prominent visual differences. In CVPR, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.619, + 0.483, + 0.686 + ], + "angle": 0, + "content": "[9] Mehdi Cherti, Romain Beaumont, Ross Wightman, Mitchell Wortsman, Gabriel Ilharco, Cade Gordon, Christoph Schuhmann, Ludwig Schmidt, and Jenia Jitsev. Reproducible scaling laws for contrastive language-image learning. In CVPR, 2023. 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.689, + 0.483, + 0.73 + ], + "angle": 0, + "content": "[10] Alessandro Conti, Enrico Fini, Massimiliano Mancini, Paolo Rota, Yiming Wang, and Elisa Ricci. Vocabulary-free image classification. NeurIPS, 36, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.732, + 0.483, + 0.759 + ], + "angle": 0, + "content": "[11] Ali Farhadi, Ian Endres, Derek Hoiem, and David Forsyth. Describing objects by their attributes. In CVPR, 2009. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.761, + 0.483, + 0.801 + ], + "angle": 0, + "content": "[12] Ali Farhadi, Ian Endres, and Derek Hoiem. Attribute-centric recognition for cross-category generalization. In CVPR, 2010." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.804, + 0.483, + 0.83 + ], + "angle": 0, + "content": "[13] Vittorio Ferrari and Andrew Zisserman. Learning visual attributes. NeurIPS, 2007. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.832, + 0.483, + 0.899 + ], + "angle": 0, + "content": "[14] Isabel O Gallegos, Ryan A Rossi, Joe Barrow, Md Mehrab Tanjim, Sungchul Kim, Franck Dernoncourt, Tong Yu, Ruiyi Zhang, and Nesreen K Ahmed. Bias and fairness in large language models: A survey. arXiv preprint arXiv:2309.00770, 2023. 4" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.236, + 0.486, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.147 + ], + "angle": 0, + "content": "[15] Peng Gao, Shijie Geng, Renrui Zhang, Teli Ma, Rongyao Fang, Yongfeng Zhang, Hongsheng Li, and Yu Qiao. Clip-adapter: Better vision-language models with feature adapters. IJCV, 132(2):581-595, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.151, + 0.907, + 0.191 + ], + "angle": 0, + "content": "[16] Xiaoyuan Guo, Kezhen Chen, Jinmeng Rao, Yawen Zhang, Baochen Sun, and Jie Yang. Lowa: Localize objects in the wild with attributes. In NeurIPS-WS, 2023. 1, 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.194, + 0.908, + 0.247 + ], + "angle": 0, + "content": "[17] Ziyu Guo, Renrui Zhang, Longtian Qiu, Xianzheng Ma, Xu-peng Miao, Xuming He, and Bin Cui. Calip: Zero-shot enhancement of clip with parameter-free attention. In AAAI, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.251, + 0.908, + 0.305 + ], + "angle": 0, + "content": "[18] Xintong Han, Zuxuan Wu, Phoenix X Huang, Xiao Zhang, Menglong Zhu, Yuan Li, Yang Zhao, and Larry S Davis. Automatic spatially-aware fashion concept discovery. In ICCV, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.309, + 0.907, + 0.348 + ], + "angle": 0, + "content": "[19] Mehrdad Hosseinzadeh and Yang Wang. Composed query image retrieval using locally bounded features. In CVPR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.352, + 0.907, + 0.393 + ], + "angle": 0, + "content": "[20] Edward J Hu, Phillip Wallis, Zeyuan Allen-Zhu, Yanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. In ICLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.396, + 0.907, + 0.436 + ], + "angle": 0, + "content": "[21] Phillip Isola, Joseph J Lim, and Edward H Adelson. Discovering states and transformations in image collections. In CVPR, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.44, + 0.906, + 0.481 + ], + "angle": 0, + "content": "[22] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In ECCV, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.484, + 0.892, + 0.497 + ], + "angle": 0, + "content": "[23] Glenn Jocher and Jing Qiu. Ultralytics yolo11. 2024. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.5, + 0.908, + 0.54 + ], + "angle": 0, + "content": "[24] Shyamgopal Karthik, Karsten Roth, Massimiliano Mancini, and Zeynep Akata. Vision-by-language for training-free compositional image retrieval. In ICLR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.543, + 0.908, + 0.611 + ], + "angle": 0, + "content": "[25] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. IJCV, 123:32-73, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.614, + 0.907, + 0.654 + ], + "angle": 0, + "content": "[26] Christoph H Lampert, Hannes Nickisch, and Stefan Harmeling. Learning to detect unseen object classes by between-class attribute transfer. In CVPR, 2009. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.657, + 0.907, + 0.699 + ], + "angle": 0, + "content": "[27] Christoph H Lampert, Hannes Nickisch, and Stefan Harmeling. Attribute-based classification for zero-shot visual object categorization. IEEE TPAMI, 36(3):453-465, 2013. 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.702, + 0.908, + 0.755 + ], + "angle": 0, + "content": "[28] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, 2022. 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.759, + 0.908, + 0.813 + ], + "angle": 0, + "content": "[29] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.816, + 0.906, + 0.87 + ], + "angle": 0, + "content": "[30] Victor Weixin Liang, Yuhui Zhang, Yongchan Kwon, Serena Yeung, and James Y Zou. Mind the gap: Understanding the modality gap in multi-modal contrastive representation learning. NeurIPS, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.907, + 0.901 + ], + "angle": 0, + "content": "[31] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. NeurIPS, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15106" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.12 + ], + "angle": 0, + "content": "[32] Sachit Menon and Carl Vondrick. Visual classification via description from large language models. In ICLR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.484, + 0.176 + ], + "angle": 0, + "content": "[33] Roberto Navigli, Simone Conia, and Björn Ross. Biases in large language models: origins, inventory, and discussion. ACM Journal of Data and Information Quality, 15(2):1-21, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.179, + 0.482, + 0.219 + ], + "angle": 0, + "content": "[34] Zachary Novack, Julian McAuley, Zachary Chase Lipton, and Saurabh Garg. Chils: Zero-shot image classification with hierarchical label sets. In ICML, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.221, + 0.484, + 0.248 + ], + "angle": 0, + "content": "[35] OpenAI. Introducing chatgpt. https://openai.com/blog/chatgpt, 2022.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.25, + 0.484, + 0.276 + ], + "angle": 0, + "content": "[36] Emin Orhan. A simple cache model for image recognition. NeurIPS, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.279, + 0.482, + 0.305 + ], + "angle": 0, + "content": "[37] Devi Parikh and Kristen Grauman. Relative attributes. In ICCV, 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.308, + 0.484, + 0.348 + ], + "angle": 0, + "content": "[38] Genevieve Patterson and James Hays. Coco attributes: Attributes for people, animals, and objects. In ECCV, 2016. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.35, + 0.484, + 0.405 + ], + "angle": 0, + "content": "[39] Khoi Pham, Kushal Kafle, Zhe Lin, Zhihong Ding, Scott Cohen, Quan Tran, and Abhinav Shrivastava. Learning to predict visual attributes in the wild. In CVPR, 2021. 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.407, + 0.484, + 0.461 + ], + "angle": 0, + "content": "[40] Khoi Pham, Kushal Kafle, Zhe Lin, Zhihong Ding, Scott Cohen, Quan Tran, and Abhinav Shrivastava. Improving closed and open-vocabulary attribute prediction using transformers. In ECCV, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.463, + 0.484, + 0.532 + ], + "angle": 0, + "content": "[41] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.534, + 0.484, + 0.575 + ], + "angle": 0, + "content": "[42] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.577, + 0.484, + 0.631 + ], + "angle": 0, + "content": "[43] Karsten Roth, Jae Myung Kim, A Koepke, Oriol Vinyals, Cordelia Schmid, and Zeynep Akata. Waffling around for performance: Visual classification with random words and broad concepts. In ICCV, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.633, + 0.484, + 0.688 + ], + "angle": 0, + "content": "[44] Kuniaki Saito, Kihyuk Sohn, Xiang Zhang, Chun-Liang Li, Chen-Yu Lee, Kate Saenko, and Tomas Pfister. Pic2word: Mapping pictures to words for zero-shot composed image retrieval. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.69, + 0.484, + 0.759 + ], + "angle": 0, + "content": "[45] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. NeurIPS, 2022. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.761, + 0.482, + 0.801 + ], + "angle": 0, + "content": "[46] Behjat Siddique, Rogerio S Feris, and Larry S Davis. Image ranking and retrieval based on multi-attribute queries. In CVPR, 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.803, + 0.484, + 0.857 + ], + "angle": 0, + "content": "[47] Kihyuk Sohn, Huiwen Chang, José Lezama, Luisa Polania, Han Zhang, Yuan Hao, Irfan Essa, and Lu Jiang. Visual prompt tuning for generative transfer learning. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.859, + 0.484, + 0.9 + ], + "angle": 0, + "content": "[48] Vishaal Udandarao, Ankush Gupta, and Samuel Albanie. Susx: Training-free name-only transfer of vision-language models. In ICCV, 2023. 2, 3, 4, 5, 6, 7, 8" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.092, + 0.908, + 0.133 + ], + "angle": 0, + "content": "[49] Peter Welinder, Steve Branson, Takeshi Mita, Catherine Wah, Florian Schroff, Serge Belongie, and Pietro Perona. Caltech-ucsd birds 200. 2010. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.907, + 0.19 + ], + "angle": 0, + "content": "[50] Hui Wu, Yupeng Gao, Xiaoxiao Guo, Ziad Al-Halah, Steven Rennie, Kristen Grauman, and Rogerio Feris. Fashion iq: A new dataset towards retrieving images by natural language feedback. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.192, + 0.908, + 0.246 + ], + "angle": 0, + "content": "[51] Yongqin Xian, Christoph H Lampert, Bernt Schiele, and Zeynep Akata. Zero-shot learning—a comprehensive evaluation of the good, the bad and the ugly. IEEE TPAMI, 41(9): 2251-2265, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.249, + 0.908, + 0.289 + ], + "angle": 0, + "content": "[52] Hantao Yao, Rui Zhang, and Changsheng Xu. Visual-language prompt tuning with knowledge-guided context optimization. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.291, + 0.908, + 0.319 + ], + "angle": 0, + "content": "[53] Aron Yu and Kristen Grauman. Fine-grained visual comparisons with local learning. In CVPR, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.321, + 0.908, + 0.374 + ], + "angle": 0, + "content": "[54] Jiahui Yu, Zirui Wang, Vijay Vasudevan, Legg Yeung, Mojtaba Seyedhosseini, and Yonghui Wu. Coca: Contrastive captioners are image-text foundation models. TMLR, 2022. 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.377, + 0.907, + 0.417 + ], + "angle": 0, + "content": "[55] Yan Zeng, Xinsong Zhang, and Hang Li. Multi-grained vision language pre-training: Aligning texts with visual concepts. ICML, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.42, + 0.907, + 0.46 + ], + "angle": 0, + "content": "[56] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In ICCV, 2023. 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.462, + 0.908, + 0.517 + ], + "angle": 0, + "content": "[57] Renrui Zhang, Wei Zhang, Rongyao Fang, Peng Gao, Kunchang Li, Jifeng Dai, Yu Qiao, and Hongsheng Li. Tip-adapter: Training-free adaption of clip for few-shot classification. In ECCV, 2022. 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.519, + 0.908, + 0.572 + ], + "angle": 0, + "content": "[58] Xiao Zhang, Rui Zhao, Yu Qiao, Xiaogang Wang, and Hongsheng Li. Adacos: Adaptively scaling cosine logits for effectively learning deep face representations. In CVPR, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.576, + 0.907, + 0.616 + ], + "angle": 0, + "content": "[59] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Conditional prompt learning for vision-language models. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.618, + 0.907, + 0.659 + ], + "angle": 0, + "content": "[60] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. IJCV, 130(9):2337-2348, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.661, + 0.908, + 0.715 + ], + "angle": 0, + "content": "[61] Yifei Zhou, Juntao Ren, Fengyu Li, Ramin Zabih, and Ser Nam Lim. Test-time distribution normalization for contrastively learned visual-language models. NeurIPS, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.718, + 0.908, + 0.758 + ], + "angle": 0, + "content": "[62] Beier Zhu, Yulei Niu, Yucheng Han, Yue Wu, and Hanwang Zhang. Prompt-aligned gradient for prompt tuning. In ICCV, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.761, + 0.907, + 0.815 + ], + "angle": 0, + "content": "[63] William Y Zhu, Keren Ye, Junjie Ke, Jiahui Yu, Leonidas Guibas, Peyman Milanfar, and Feng Yang. Artvlm: Attribute recognition through vision-based prefix language modeling. ECCV, 2024. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.818, + 0.907, + 0.871 + ], + "angle": 0, + "content": "[64] Xiangyang Zhu, Renrui Zhang, Bowei He, Aojun Zhou, Dong Wang, Bin Zhao, and Peng Gao. Not all features matter: Enhancing few-shot clip with adaptive prior refinement. In ICCV, 2023. 3" + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.092, + 0.908, + 0.871 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15107" + } + ] +] \ No newline at end of file diff --git a/2025/Compositional Caching for Training-free Open-vocabulary Attribute Detection/5997b1e0-b306-408f-91ce-2ab833ab2009_origin.pdf b/2025/Compositional Caching for Training-free Open-vocabulary Attribute Detection/5997b1e0-b306-408f-91ce-2ab833ab2009_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..a6f0d0802a508ca4996a695aaa30227cd7f8fff3 --- /dev/null +++ b/2025/Compositional Caching for Training-free Open-vocabulary Attribute Detection/5997b1e0-b306-408f-91ce-2ab833ab2009_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b6d92c99640534c4863b3cd87110c5a1613b0f7f640666e7f286d7aefd5cf92f +size 2292097 diff --git a/2025/Compositional Caching for Training-free Open-vocabulary Attribute Detection/full.md b/2025/Compositional Caching for Training-free Open-vocabulary Attribute Detection/full.md new file mode 100644 index 0000000000000000000000000000000000000000..ed205906beeb392a7777c9d4ce7636f1431a0136 --- /dev/null +++ b/2025/Compositional Caching for Training-free Open-vocabulary Attribute Detection/full.md @@ -0,0 +1,338 @@ +# Compositional Caching for Training-free Open-vocabulary Attribute Detection + +Marco Garosi1 Alessandro Conti1 Gaowen Liu2 Elisa Ricci1,3 Massimiliano Mancini1 +1University of Trento 2Cisco Research Fondazione Bruno Kessler +https://comca-attributes. github.io + +# Abstract + +Attribute detection is crucial for many computer vision tasks, as it enables systems to describe properties such as color, texture, and material. Current approaches often rely on labor-intensive annotation processes which are inherently limited: objects can be described at an arbitrary level of detail (e.g., color vs. color shades), leading to ambiguities when the annotators are not instructed carefully. Furthermore, they operate within a predefined set of attributes, reducing scalability and adaptability to unforeseen downstream applications. We present COMPOSITIONAL CACHING (COMCA), a training-free method for open-vocabulary attribute detection that overcomes these constraints. COMCA requires only the list of target attributes and objects as input, using them to populate an auxiliary cache of images by leveraging web-scale databases and Large Language Models to determine attribute-object compatibility. To account for the compositional nature of attributes, cache images receive soft attribute labels. Those are aggregated at inference time based on the similarity between the input and cache images, refining the predictions of underlying Vision-Language Models (VLMs). Importantly, our approach is model-agnostic, compatible with various VLMs. Experiments on public datasets demonstrate that COMCA significantly outperforms zero-shot and cache-based baselines, competing with recent training-based methods, proving that a carefully designed training-free approach can successfully address open-vocabulary attribute detection. + +# 1. Introduction + +Attributes shape our perception and interactions with the environment. Let us consider the example of Fig. 1. If our aim is to identify the animal, we may focus on its fur color, size, the presence of horns and the tail. Similarly, if we want to analyze its health status, we may focus on its age, fur patterns and shades, and its skin. This level of analysis requires models with fine-grained understanding of the visual input, which is usually obtained by training a model on datasets that annotate both objects and their corresponding + +![](images/db59c2c80abbb2e7c48cbb7b17ea624025591c2b0ee627a3ce4681338528f924.jpg) +Figure 1. Example of attribute annotations. Attribute annotations are often: sparse, as they are not consistent across samples; incomplete, as not all attributes are annotated; ambiguous, as they can be subjective or miss a frame of reference. This makes open-vocabulary attribute detection challenging. Differently from previous works [5, 7], we do not rely on such annotations and propose COMCA, a training-free approach requiring no supervision. + +attributes [38, 40]. However, collecting attribute annotations is both time-consuming and prone to ambiguities, especially when annotators lack clear instructions. Additionally, attribute annotations are often tailored to specific datasets and are ambiguous. For instance, in Fig. 1, one may list "brown" as a color for the left goat, but others may argue that it is a red-ish shade of brown. A simple solution is to provide annotators with a predefined list of attributes; however, (i) the list may not cover all possible annotations for a property (e.g., all colors) and (ii) it limits models trained on it to the given list itself, making them unable to recognize new attributes. + +Recent approaches overcome some of the challenges by developing open-vocabulary attribute detection models [5, 7, 16] that detect attributes specified in natural language at inference time. While these approaches generalize beyond the set of seen attributes and objects, their performance is still biased toward the specific dataset they are trained on. Indeed, training-based approaches may show significant degradation in performance in cross-dataset experiments (see Fig. 4). + +In this work, we overcome potentially ambiguous and/or limited attribute annotations by introducing COMPOSITIONAL CACHING (COMCA), a training-free open-vocabulary attribute detection model. COMCA employs a Vision-Language Model (VLM) [41] and refines its + +predictions by utilizing an auxiliary cache that contains images together with automatically estimated attribute labels. + +Caching for refining CLIP predictions has been introduced in the context of image classification [57]. It requires storing images, either generated or retrieved [48, 57], for each of the classes of interests, together with a corresponding one-hot categorical label. The similarity of the input image with each of the examples is used to weight the labels in the cache and refine the model's prediction. While this approach is effective for object classification, it is prohibitively expensive in the context of attribute detection due to the much larger output space, comprising all possible combinations of attributes and objects. COMCA achieves scalability in caching by following two key compositional principles: (i) not all attributes are universally applicable to all objects, and (ii) every object inherently possesses multiple attributes. + +The first principle is implemented in two steps: attribute-object compatibility estimation and cache pair sampling. In the first phase, we estimate the probability of an attribute appearing with an object using the statistics of available databases (e.g., CC12M [6]) and Large Language Model (LLM) queries. Next, we populate the cache for a given attribute by sampling compatible objects from the estimated distribution and retrieving relevant images from a database. The second principle is implemented by self-labeling the cached elements. Initially, each sample is associated with the single attribute it was sampled for. Since this does not account for other attributes present in the image, we extend attribute labels through soft-labeling, i.e., comparing cache images to text embeddings of each attribute of interest. These scores modify a cache element's contribution to each target attribute, considering their implicit composition. We evaluate COMCA on two publicly available benchmarks, OVAD [5] and VAW [39]. COMCA demonstrates significant improvements over training-free baselines and is competitive with recent training-based approaches, while remaining unaffected by the biases of specific training sets. Notably, COMCA provides a consistent boost in performance regardless of the VLM pre-training and architecture. + +In summary, our contributions are: ① We propose COMCA, an approach specifically designed for training-free open-vocabulary attribute detection. COMCA exploits the compositional nature of objects and attributes to build an auxiliary cache that is used as an anchor to refine the model's predictions. ② We make caching scalable by extracting priors from databases and LLMs on which objects are associated with each attribute, using them to populate the cache. Soft-labeling further accounts for the multiplicity of attributes within an image. ③ We empirically show that COMCA outperforms all the cache-based baselines and achieves competitive performance against costly training-based methods, demonstrating the practical feasibility of training-free open vocabulary attribute detection. + +# 2. Related work + +Attribute detection. Detecting visual attributes is a widely explored task in computer vision [4, 11-13], supporting various downstream applications such as zero-shot learning [1, 26, 27, 51], relative image comparison [8, 37, 46, 53], and compositional image retrieval [3, 19, 24, 44]. + +Recent works have introduced several benchmarks to promote and ease research in this area [18, 21, 25, 38-40, 49, 50, 53]. Despite these advances, annotating visual attributes remains a complex and challenging task. This is due to the wide variability in attributes across objects and contexts, as well as the difficulty of ensuring consistent annotations when multiple annotators are involved. Moreover, many datasets capture only a limited range of object attributes, limiting their usefulness for developing general-purpose models. + +Open-vocabulary attribute detection (OVAD) [5, 7] overcomes some of these limitations by detecting user-specified attributes, typically provided in natural language, even those that have not been observed during training. This task was first explored in [5], which introduced the OVAD benchmark and method. OVAD generates supervisory signals for attributes, nouns, and noun phrases by parsing captions and encoding them using CLIP text embeddings [41]. OvarNet [7] builds on this by fine-tuning CLIP with federated learning and additional supervision from external captions. LOWA [16] proposes a three-stage pipeline that aligns visual embeddings to objects, attributes, and free-form text. ArtVLM [63] uses a generative VLM to obtain attribute probabilities based on the input image. Although these methods operate in the open-vocabulary setting, they are training-based and require both a lot of annotated data and several computational resources (see Sec. 4). COMCA significantly differs from these approaches, as we explore for the first time a training-free, cache-based, strategy for attribute detection. Adapting Vision-Language Models. VLMs [28, 29, 31, 41, 54] bridge the gap between textual and visual data, showing remarkable zero-shot generalization abilities. Recent efforts have focused on improving their performance on specific downstream tasks through various strategies, broadly categorized into training-based and training-free methods. + +Training-based approaches typically involve tuning text prompts [22, 47, 52, 59, 62]. Some methods learn prompts directly [60] or predict them [59] from visual inputs to improve generalization. Other approaches train adapters to refine the model's predictions [15, 20]. While effective, these methods introduce additional parameters, which can negatively impact the model's zero-shot performance [59]. + +Training-free methods take a different approach, such as incorporating additional knowledge into the prompts using natural language [32, 34, 43], or modifying predictions through attention layers [17] or normalization techniques [61]. Recent works also employ a cache of images and one-hot labels that represent the target classes [48, 57, + +![](images/ccb2c6f9ec37d0895662b61503a8abb082efad60b0167607ebd2378b1cad974d.jpg) +Figure 2. COMCA's cache construction. Given a list of attributes and objects, we compute their compatibility from a large database $D_r$ and with an LLM. The scores are merged and normalized to obtain the compatibility distribution, from which we sample cache entries and construct the cache. We enrich the latter with soft labels from the VLM-based similarity between cache images and attributes. + +[64]. For example, TIP-Adapter [57] computes the similarity between input images and cached examples, using the associated labels for prediction. Similarly, SuS-X [48] creates a cache automatically, either by using a generative model [42] or retrieving from a large database [45], with distance-based scoring to weight the elements' contribution. While our work builds upon the same principles of TIP-Adapter and SuS-X, i.e., using a cache to refine predictions, it differs in four key aspects, as COMCA's cache: (i) is aware of attribute-object compatibility; (ii) accounts for attribute combinations; (iii) models the compositional nature of attributes and objects; (iv) scales efficiently with the number of attributes. + +# 3. Compositional Caching + +In this section, we formally define the open-vocabulary attribute detection problem and describe caching as a tool for VLM adaptation (Sec. 3.1). Next, we describe how attribute-object compatibility informs the construction of scalable caches (Sec. 3.2). Finally, we detail our process for refining cache labels, taking into account multi-attribute compatibility (Sec. 3.3). Fig. 2 provides an overview of our approach. + +# 3.1. Background + +Problem formulation. Given an image $x \in \mathcal{X}$ of an object and a list of attributes, our goal is to detect which attributes are present in the image. Let $\mathcal{A}$ be the set of all existing attributes in natural language. At inference time, the model is provided with a set of $n$ target attributes $A = \{a_{1},\dots ,a_{n}\}$ as input, where $A \subset \mathcal{A}$ . These are associated with $m$ target objects, which we denote as $O = \{o_1,\dots ,o_m\}$ . Our goal is to build a function $f:\mathcal{X}\times \mathcal{A}^n\to [-1,1]^n$ that maps an image and a list of attributes to a binary vector indicating the presence (1) or absence (-1) of each attribute. + +We obtain $f$ without training, in contrast to previous works that learn $f$ using a dataset $D$ of image-label pairs + +$(x, \mathbf{a})$ with $\mathbf{a} \in [-1, 1]^n$ being the ground-truth binary vector1. We achieve this by implementing $f$ using a VLM combined with an auxiliary cache to refine its predictions. + +Adapting VLMs with caching. Following previous works [5, 7], we implement $f$ with a VLM. While we primarily consider CLIP [41] to describe our method and perform our evaluation, the proposed framework can integrate several other VLMs (see Tab. 2). The function $f$ has three elements: a visual encoder $f_{v}$ , a text encoder $f_{t}$ , and a prediction function. $f_{v}$ and $f_{t}$ map images in $\mathcal{X}$ and natural language in $\mathcal{T}$ , respectively, to a shared $d$ -dimensional multimodal space $\mathcal{Z} \in \mathbb{R}^d$ , i.e., $f_{v}: \mathcal{X} \to \mathcal{Z}$ and $f_{t}: \mathcal{T} \to \mathcal{Z}$ . Next, the prediction function maps embeddings from the multimodal space to the final scores. A standard way for scoring an image $x$ against a class $c$ is by computing the cosine similarity between their embeddings: $^2$ + +$$ +f _ {\mathrm {C L I P}} (x, c) = \frac {f _ {v} (x) ^ {\top} f _ {t} (c)}{| | f _ {v} (x) | | \cdot | | f _ {t} (c) | |}. \tag {1} +$$ + +While effective, $f$ can be improved without any training by using an auxiliary cache that stores examples of the classes of interest [36, 48, 57]. Given a cache $\mathcal{C} = \{x_1^1, \dots, x_C^K\}$ that contains $K$ samples of $C$ classes, we can define the cache-based prediction for a class $c \in C$ as: + +$$ +f _ {\mathcal {C}} (x, c) = \sum_ {i = 1} ^ {K} \eta_ {\mathcal {C}} \left(\frac {f _ {v} (x) ^ {\intercal} f _ {v} \left(x _ {c} ^ {i}\right)}{\left| \left| f _ {v} (x) \right| \right| \cdot \left| \left| f _ {v} \left(x _ {c} ^ {i}\right) \right| \right|}\right), \tag {2} +$$ + +where $\eta_{\mathcal{C}}(z) = \exp (1 - \beta (-z))$ is a normalization function regulated by the hyperparameter $\beta$ [57]. + +The formulation in Eq. (2) has a main issue when applied to open-vocabulary attribute detection: the cache construction. On one hand, a cache considering all possible attribute labels is semantically complete but requires a prohibitive storage size of $K \cdot n!$ images, e.g., $K \cdot 117!$ for OVAD [5]. On the other hand, while storing $K$ positive examples per attribute has low memory requirements, it ignores the multi-attribute nature of images (e.g., shape, size, color, etc. can co-exist in the same sample), potentially leading to incorrect predictions (e.g., a high score because the color matches but the cache label refers to the shape). Moreover, attributes may appear differently across objects (e.g., wet car vs wet dog) and selecting a random subset of them may lead to a cache representative only of a few cases. In the following, we address these issues by focusing on the compatibility between objects and multiple attributes in the real world. + +# 3.2. Compatibility for scalable caching + +Naively constructing the cache would lead to many problems, related both to labels' semantic and memory requirements (Sec. 3.1). In this section, we present a methodology that (i) has the same memory requirements as storing only $K$ examples per attribute and (ii) assigns multiple attribute labels per example, taking into account attribute-object compatibility. Scalable caching. To obtain a cache $\mathcal{C}$ that is scalable with the number of attributes and accounts for object-specific information, we exploit the natural distributions of attributes across objects. In particular, let us define a function $\varphi$ that, given an attribute $a \in \mathcal{A}$ and an object $o \in \mathcal{O}$ , returns their compatibility value as a scalar, i.e., $\varphi : \mathcal{A} \times \mathcal{O} \to \mathbb{R}$ . + +Considering an attribute $a$ and an object $o$ , the score $\varphi(a, o)$ is proportional to the probability that $o$ exhibits $a$ . Thus, we can use the scores $\Phi_O(a) = [\varphi(a, o_1), \dots, \varphi(a, o_m)]$ as a proxy for the probability distribution of $a$ co-occurring with the elements of $O$ . We then assign each of the $K$ elements (i.e., shots) of the attribute $a$ to one object $o$ , sampling the object from $\Phi_O(a)$ : + +$$ +\hat {O} _ {a} ^ {K} = \left\{\hat {o} _ {1} ^ {a}, \dots , \hat {o} _ {K} ^ {a} \right\}, \text {w h e r e} \hat {o} _ {i} ^ {a} \sim \tilde {\Phi} _ {O} (a), \tag {3} +$$ + +where $\tilde{\Phi}_O(a)$ is the normalized version of $\Phi_O(a)$ , i.e., $\tilde{\Phi}_O(a) = \Phi_O(a) / \sum \Phi_O(a)$ . The set $\hat{O}_a^K$ allows populating the cache with relevant objects and attributes. Let us denote a generic function mapping a textual query to an image as $\mathrm{T2I}:\mathcal{T}\to \mathcal{X}$ . The elements of the cache are: + +$$ +\begin{array}{l} \mathcal {C} _ {\mathrm {H}} = \left\{x _ {a _ {1}} ^ {1}, \dots , x _ {a _ {m}} ^ {K} \right\} (4) \\ = \{\mathrm {T 2 I} \left(\pi_ {g} \circ a \circ o _ {1} ^ {a}\right), \dots , \mathrm {T 2 I} \left(\pi_ {g} \circ a \circ o _ {K} ^ {a}\right) \}, (4) \\ \end{array} +$$ + +where $\pi_g$ is a contextual prompt for the image (e.g., "a photo of"). Following SuS-X [48], we implement T2I via retrieval. Specifically, we retrieve images from a database by scoring them against textual queries with the VLM (e.g., CLIP). While we focus on retrieval, COMCA + +also supports the implementation of T2I via a text-to-image generative model, e.g., by feeding textual queries to Stable Diffusion [42]. We refer to the Supp. Mat. for more details. Estimating attribute-object compatibility. The key element that enables scalable caching is the function $\varphi$ that estimates the viability of attribute-object compositions. We implement it by merging two scoring strategies. + +The first uses a generic large-scale database $D_r$ containing image-caption pairs. Using the captions $t \in D_r$ , we can estimate the database score as the number of co-occurrences of attribute $a$ and object $o$ , i.e.: + +$$ +\varphi_ {\mathrm {d b}} (a, o) = \sum_ {(x, t) \in D _ {r}} \mathbb {1} (o \subset t \wedge a \subset t), \tag {5} +$$ + +where $\mathbb{1}(z)$ is an indicator function whose value is 1 whenever the condition $z$ is true, and 0 otherwise. + +While this process exploits real data, it may be biased on the specific captions and/or data collection process. For instance, captions rarely describe objects in their default state, such as a dry dog. Therefore, a second scoring strategy exploits the reasoning capabilities of an LLM, querying it to predict an attribute-object compatibility as: + +$$ +\varphi_ {\mathrm {L L M}} (a, o) = \mathrm {L L M} \left(\pi_ {c}, o, a\right), \tag {6} +$$ + +where $\pi_c$ is a prompt specifying the compatibility task (see the Supp. Mat.). The objective is to assign a score on a specific range to the likelihood of the attribute-object pair. + +Taking into account that also LLMs have issues, such as hallucinations and biases [14, 33], we implement $\varphi$ by combining database- and LLM-based scoring as: + +$$ +\varphi (a, o) = \varphi_ {\mathrm {d b}} (a, o) \cdot \varphi_ {\mathrm {L L M}} (a, o), \tag {7} +$$ + +thus computing $\Phi_O(a)$ and filling the cache accordingly. + +# 3.3. Compositional cache refinement + +After the process in Sec. 3.2, we have a cache $\mathcal{C}_{\mathbb{H}}$ populated with $K$ examples for each attribute $a \in A$ . However, the cache assigns one hard attribute to each element, contrasting with the natural co-occurrence of attributes and leading to potentially wrong associations during inference. As a simple example, an image of a large and blue car obtained with the binding (large, car) will have a positive label only for large. As a consequence, if the test image shows a small blue car, the cache element will not contribute to the score of blue and may erroneously increase the score for the antonym large. + +Soft caching. To remedy to this issue, we refine the label associated to each cache element, exploiting the VLM capabilities to compute soft labels. Formally, the soft label $s_a$ for an attribute $a$ and a cache image $x_c$ is defined as: + +$$ +s _ {a} \left(x _ {c}\right) = \frac {f _ {t} (a) ^ {\intercal} f _ {v} \left(x _ {c}\right)}{\left| \left| f _ {t} (a) \right| \right| \cdot \left| \left| f _ {v} \left(x _ {c}\right) \right| \right|}, \tag {8} +$$ + +where we compare the text embeddings of the attribute with the visual embeddings of the cache image. While these scores can directly serve as soft labels, we found them to exhibit low-variance due to the well-known modality-gap [30, 48]. We overcome this limitation and increase their relevance by defining normalized soft labels $\bar{s}_a$ as: + +$$ +\bar {s} _ {a} \left(x _ {c}\right) = \frac {\exp \left(\left(s _ {a} \left(x _ {c}\right) - \mu_ {C}\right) / \sigma_ {C}\right)}{\sum_ {\hat {a} \in A} \exp \left(\left(s _ {\hat {a}} (x) - \mu_ {C}\right) / \sigma_ {C}\right)}, \tag {9} +$$ + +where $\mu_{C}$ and $\sigma_{C}$ are respectively the mean and standard deviation of the similarity between any attribute with any element of the cache. The first normalization expands the range of values of the similarity scores $s_a$ , while the softmax ensures that all samples have an equal contribution to the final prediction. In practice, we found the latter step to be important only on datasets with many attributes (e.g., VAW). We hypothesize this is due to differences in the output space cardinality, requiring different softmax scaling as in [58]. + +These scores are mixed with the original hard labels to obtain the final label $\hat{s}_a(x_c) = (1 - \alpha)\cdot \mathbb{1}(a = c) + \alpha \cdot \overline{s}_a(x_c)$ where $\alpha$ is a trade-off hyperparameter. With the soft labels $\hat{s}_a$ , we can perform the cache-based prediction as: + +$$ +f _ {\mathcal {C}} ^ {\mathrm {C O M C A}} (x, a) = \sum_ {x _ {c}} ^ {\mathcal {C}} \eta_ {\mathcal {C}} \left(\hat {s} _ {a} \left(x _ {c}\right) \frac {f _ {v} (x) ^ {\intercal} f _ {v} \left(x _ {c}\right)}{\left\| f _ {v} (x) \right\| \cdot \left\| f _ {v} \left(x _ {c}\right) \right\|}\right). \tag {10} +$$ + +In contrast to Eq. (2), here we consider each cache element $x_{c}$ as a potential contributor for attribute $a$ , even if $a \neq c$ . Implicitly, this allows us to include an arbitrary number of (positive) attribute labels within each element of the cache. + +Final inference. Finally, the cache-based scoring defined in Eq. (10) is merged with the standard CLIP-based scores to obtain the final predictions: + +$$ +f (x, a) = \eta_ {A} \left(\lambda \cdot f _ {\mathcal {C}} ^ {\mathrm {C O M C A}} (x, a) + f _ {\mathcal {C}} ^ {\mathrm {C L I P}} (x, a)\right), \tag {11} +$$ + +where $\lambda$ is an hyperparameter controlling the balance between the two and $\eta_A(z) = \mathrm{softmax}_A(z / \max (z))$ , moving from simple scalar values to a distribution over the attributes space and leading to more stable scores. Note that Eq. (11) contains no element which needs training or labeled data for the specific attributes of interest. Given the target sets of attributes $A$ and objects $O$ , COMCA automatically re-generates the cache to update the inference procedure. + +# 4. Experiments + +Datasets. We evaluate COMCA on two datasets. OVAD [5] covers 80 classes and 117 attributes across 2,000 images from MS-COCO [38], for a total of 14,300 instances. Attributes are manually annotated as positive, negative, and unknown. VAW [39] has more concepts, i.e., 2,260 classes and 620 attributes across 10,392 images. VAW is more + +sparse than OVAD, i.e., around 1.8 positive and 5.3 negative attributes per instance, and it is automatically annotated. + +Setting. Following prior works [5, 7], COMCA focuses on the box-given setting, where the bounding boxes of objects are given, as this allows us to assess its attribute detection capabilities more precisely. However, Tab. 3 also evaluates COMCA in the box-free setting by introducing an object detector. See the Supp. Mat. for more details on this setting. + +Metrics. We measure performance as mean Average Precision (mAP), and also report mAP scores on head, medium, and tail classes based on their frequency in the dataset. + +Baselines and competitors. We evaluate COMCA against training-based methods, i.e., OVAD [5], OvarNet [7], ArtVLM [63], and LOWA [16]. As we are the first to address the problem in a completely training-free fashion, we extend the baselines with several zero-shot methods [9, 28, 54-56], and adapt previous cache-based approaches to our task [48, 57]. For TIP-Adapter [57] and SuS-X [48] we construct a $K$ -shot cache, considering only attributes for sampling and disregarding their co-occurrences with objects. This follows their approach, which samples $K$ shots for each target class, i.e., attributes in our task. In addition, we test IAP [26, 27], which conditions predicted attributes on the inferred object category, using TIP-Adapter as the object classifier. We also consider a naive cache-based approach, i.e., "image-based", that uses CC12M [6] to build a cache specifically for each input image by retrieving the most similar samples. Since the retrieved images are unlabeled, we use our soft labels to annotate the cache. We construct all the proposed baselines on top of CLIP ViT-B/32 [9]. For additional details, please refer to the Supp. Mat. + +Implementation details. We use CLIP ViT-B/32 [9] pretrained on LAION2B [45] as our primary VLM. Following TIP-Adapter [57], we set $\lambda = 1.17$ in Eq. (11) and $\beta = 1.0$ in Eq. (2). We set $\alpha = 0.6$ to mix hard and soft labels on the validation set of VAW [39]. Lastly, we use CC12M [6] as the database in Eq. (5) and GPT 3.5 Turbo [35] as the LLM for Eq. (6), as we find it to perform better than GPT 4o-mini (see the Supp. Mat.). Unless otherwise stated, we use $K = 16$ shots per attribute, retrieving them from CC12M. We use the same hyperparameters for both datasets. For the box-free experiments, we use YOLOv11M [23] as the detector. + +# 4.1. Results + +We present our comprehensive experimental evaluation comparing COMCA with state-of-the-art training- and cache-based methods. We also show its performance with different backbones and in the box-free setting. Moreover, we discuss qualitative results and limitations of other approaches. + +Comparison with the state of the art. Tab. 1 reports the results comparing zero-shot baselines, cache-based methods, and COMCA in the box-given setting. On both benchmarks, COMCA outperforms all the zero-shot and cache-based com + +
MethodTrain sizeOVAD [5]VAW [39]
mAPHeadMediumTailmAPHeadMediumTail
Zero shotCLIP [9] RN5011.841.011.71.435.337.835.126.2
CLIP [9] ViT-B/3217.044.318.45.550.051.050.943.2
CLIP [9] ViT-L/1418.344.420.56.451.051.452.545.0
Cache-basedImage-based16.845.819.23.552.953.753.846.4
TIP-Adapter [57] + IAP [27]15.143.317.81.727.732.425.516.6
TIP-Adapter [57]16.744.419.73.157.557.559.451.2
SuS-X [48]20.248.924.64.530.233.828.820.3
ComCA27.454.334.69.058.158.259.951.7
Training basedOVAD [5]110k21.448.026.95.2----
OvarNet [7]190k28.658.635.59.568.5---
ArtVLM [63]N/A----71.975.072.159.4
LOWA [16]1.33M18.758.020.42.642.646.441.032.9
+ +Table 1. Comparison with state of the art. Green indicates COMCA. Bold indicates best among training-free methods. The symbol “-” indicates results for the competitors are not available on the original papers or the impossibility to run their method due to lack of public code and/or model weights. See the Supp. Mat. for an extended version of the table. + +petitors, demonstrating the effectiveness of our specialized design of the cache. It scores an increment of $+10.6\mathrm{mAP}$ on OVAD and $+5.2$ on VAW compared to the image-based baseline. Similarly, it surpasses TIP-Adapter [57] on both benchmarks by $+10.7\mathrm{mAP}$ and $+0.6$ , respectively. SuS-X [48] increases performance on OVAD while decreasing it on VAW. Overall, COMCA has a gain of $+7.2\mathrm{mAP}$ and $+27.9$ on the two datasets. Lastly, TIP-Adapter + IAP [27] demonstrates that naively using co-occurrences of attributes with objects [27] is insufficient to address this task. COMCA surpasses it by $+12.3\mathrm{mAP}$ on OVAD and $+30.4$ on VAW. + +We also compare COMCA with training-based methods. Notably, COMCA outperforms LOWA [16] on both benchmarks, despite the latter having been trained on a large-scale dataset while our approach requires no training data. + +Experiments with multiple VLMs. To demonstrate that COMCA is model-agnostic, we test it on various backbones and VLMs [9, 28, 54-56] and report the results in Tab. 2. On both OVAD and VAW, COMCA boosts performance consistently for all the backbones up to $+11.5\mathrm{mAP}$ on OVAD with CoCa ViT-L/14 [54] and $+15.8\mathrm{mAP}$ on VAW with CLIP ResNet50 (RN50) [9]. On average, COMCA improves the performance of the original VLM by $+8.9$ on OVAD and $+7.0$ on VAW, showing COMCA has consistent advantages. + +Cross-datasets advantages. To demonstrate the generalization capabilities of our method, we report cross-dataset results in Fig. 4. We compare COMCA with OVAD when trained on various datasets (i.e., VAW [39], OVAD [5], MSCOCO [38]) and tested on the OVAD (left) and on the VAW (right) benchmarks. As shown in Fig. 4, performance significantly degrades when OVAD is tested on datasets different from the training one (yellow bars vs red bar), whereas our training-free COMCA achieves competitive performance + +(green bars) with the same RN50 backbone. This reinforces our training-free method, which does not inherit biases from a training set, differently from training-based approaches. + +Box-free setting. In this work, we primarily focus on attribute recognition and therefore adopt the box-given setting. Here, we evaluate COMCA in the box-free setting to demonstrate that it achieves competitive performance in a less constrained scenario. Tab. 3 shows the results for zero-shot and cache-based methods, Specifically, we use CLIP ViT-B/32 [9] and YOLOv11M [23] as detector for all models. We note that all cache-based methods apart from SuS-X underperform the zero-shot baseline, with a decrease of up to $-3.1\mathrm{mAP}$ for image-based. Only, SuS-X has a gain of $+2.8\mathrm{mAP}$ w.r.t. CLIP ViT-B/32, with our method COMCA being superior and surpassing CLIP ViT-B/32 by $+9.0\mathrm{mAP}$ . Notably, our model is competitive with the training-based ones, surpassing the performance of both OVAD $(+2.7\mathrm{mAP})$ and LOWA $(+2.8\mathrm{mAP})$ . We provide a more detailed analysis and an extended version of these results in the Supp. Mat. + +Qualitative results. Fig. 3, we show qualitative examples comparing training-based OVAD [5], CLIP ViT-B/32 [9], and COMCA with the same backbone. We consider the top-1 prediction for all methods. OVAD correctly identifies the bear's color but misclassifies the cat's color and wrongly predicts a "tiled" clothes pattern for two people. We speculate it likely predicts the same attribute for two objects (i.e., the $2^{nd}$ and the $6^{th}$ ) due to biases acquired during training. CLIP makes more errors, such as labeling the zebra as "reflective" and the airplane as "tiny", while misclassifying the baseball player as "giant". COMCA instead understands attributes better, for instance by accurately identifying the zebra's texture. These issues are less pronounced in COMCA thanks to our caching mechanism, which allows it to compare inputs + +
MethodOVAD [5]VAW [39]
BackboneCOMCAmAPΔmAPΔ
CLIP RN50 [9]11.835.3
✓19.2+7.451.1+15.8
CLIP ViT-B/32 [9]17.050.0
✓27.4+10.458.1+8.1
CLIP ViT-L/14 [9]18.351.0
✓24.8+6.561.0+10.0
SigLIP ViT-B/16 [56]13.752.5
✓22.1+8.459.8+7.3
CoCa ViT-B/32 [54]15.149.6
✓24.4+9.351.4+1.8
CoCa ViT-L/14 [54]14.449.8
✓25.9+11.554.5+4.7
BLIP [28]15.951.0
✓22.7+6.856.0+5.0
X-VLM [55]17.454.8
✓28.4+11.057.7+2.9
+ +Table 2. Box-given results with different backbones. Green indicates COMCA applied on top of the backbone. $\Delta$ indicates improvements w.r.t. the corresponding baseline. See the Supp. Mat. for an extended version of the table. + +
MethodOVAD [5]
mAPHeadMediumTail
Zero shotCLIP [9] RN5012.340.212.71.9
CLIP [9] ViT-B/3212.539.612.52.7
CLIP [9] ViT-L/1413.439.714.32.9
Cache-basedImage-based9.435.48.71.0
TIP [57] + IAP [27]9.836.99.21.0
TIP-Adapter [57]10.136.09.81.2
SuS-X [48]15.343.017.23.2
COMCA21.548.426.46.0
Training basedOVAD [5]18.847.722.04.6
OvarNet [7]27.256.833.68.9
LOWA [16]18.758.020.42.6
+ +Table 3. Results in the box-free setting. Green indicates COMCA. Bold indicates the best results among training-free methods. See the Supp. Mat. for an extended version of the table. + +![](images/dca6ab54cfb7c11f58f092b69f981828ce8d75a4ff1d587ea69dbb145777b171.jpg) +OVAD CLIP COMCA +brown brown single, one, individual + +![](images/7ce6350624792f5749ac16cee0b5d90dc83730a82509b80083a45b28ef00e985.jpg) +clothes pattern: tiled male neutral, calm, serious + +![](images/745fc7b5e48b84ea10d42653376b60c5c44208cd6b19e588c8fa0e21d346ae98.jpg) +plaid, tartan, checkered reflective striped, lined, pinstriped + +![](images/93e11421a085444e7676f52c6ae85268e6e4cd6181f7a437a51fd4735c0f961d.jpg) +large, giant, huge small, little, tiny smooth, sleek + +![](images/1d7253325f036b30d81cf2d1cf6c400fa60d44663b32e2d82159fd575111c000.jpg) +yellow piece, cut small, little, tiny + +![](images/c730c7329b0e3740b1aea6f8fc53ad1d4d8bc8f9de7b469c4355ea80974ed769.jpg) +clothes pattern: tiled big, large, giant, huge sitting, sit + +![](images/6a23ffc3cb2bd9093dd8e46d13541458d3940461dac0717f96cd93cfa7ede7fc.jpg) +Figure 3. Qualitative results. Predictions of OVAD, CLIP and ComCA on some OVAD images. Green are correct ones, red are wrong. +Figure 4. Cross-dataset results. Results on OVAD (left) and VAW (right). Yellow indicates methods tested cross-domain, i.e., train and test set differ. Red indicates methods tested in-domain, and green our training-free method, which does not suffer from domain shift problems. All experiments use a RN50 backbone. + +to self-labeled examples at test time. + +# 4.2. Ablation study + +In this section, we analyze COMCA's components, studying how sampling strategies affect performance (Tab. 4), soft labeling's impact (Tab. 5 and Fig. 6), and the number of samples in the cache (Fig. 5). Additional experiments in the Supp. Mat. show the effect of (i) the hyperparameters $\alpha$ and $\lambda$ , (ii) constructing the cache with Stable Diffusion, and (iii) + +using other LLMs for generative scores. + +Cache construction. In Tab. 4, we evaluate five strategies to construct the cache. The first, Random, samples $K$ objects for each attribute and creates a cache of $|A| \times K$ examples, seeking only a match on the attribute. The second, Brute force, considers all possible combinations of attributes and objects, resulting in a large cache of $|O| \times |A| \times K$ samples. As VAW contains many attributes and objects, the Brute force cache would be too large and cannot be practically tested. As shown in Tab. 4, both approaches lead to very low results on OVAD. Compared to the zero-shot CLIP baseline, performance is similar (Random, -0.3 mAP) or even below (Brute force, -6.4 mAP). We hypothesize that Random ignores attribute-object relationships, while Brute force introduces many noisy and potentially detrimental samples, harming performance. On the contrary, Random improves by +7.5 on VAW. We hypothesize this is due to the nature of the dataset, which contains many attributes and objects. + +When introducing priors on object-attributes compositions via a database (Eq. (5)) or an LLM (Eq. (6)), performance largely increases by up to $+4.1$ mAP on VAW w.r.t. the baseline for database scores. ComCA combines the two scores to obtain object priors, achieving the best perfor + +
ConfigurationmAP
OVAD [5]VAW [39]
Zero-shot [9]17.050.0
Random16.757.5
Brute force10.6-
DB scores18.654.1
LLM scores21.155.4
DB + LLM scores26.455.2
+ +Table 4. Ablation on cache construction. Green indicates our default configuration. Bold indicates the best results. Results are without soft labels. See the Supp. Mat. for an extended version of the table. + +
ConfigurationmAP
OVAD [5]VAW [39]
Zero-shot [9]17.050.0
One-hot26.455.2
Soft26.753.2
- Sharpening [2]21.527.7
- Softmax24.558.5
- Eq. (9)27.458.1
+ +Table 5. Ablation on soft labels. Green indicates our default configuration. Bold indicates the best results. See the Supp. Mat. for an extended version of the table. + +![](images/3c696229124d5b3e2c71259862b246ecd9672b720a8d8f9fadda9fac9bc4ecce.jpg) +Figure 5. Ablation on number of samples per attribute. The purple line is mAP on OVAD, while the red line is mAP on VAW. Green indicates our default configuration. + +![](images/4997d68500826be54128ff262482948f7da2d213a30c33e0f5c98076a1e805dd.jpg) +Figure 6. Soft labels across cache priors. Performance on OVAD (left) and VAW (right), considering generative (LLM) and retrieval (CC3M, CC12M) scores. Yellow indicates one-hot labels, red soft labels and green COMCA, combining LLM and CC12M. + +mance. These results confirm both (i) that attribute-object compatibility is important to build the cache, and (ii) that databases and LLMs provide complementary information. + +Soft labels. As discussed in Sec. 3.3, each image in the cache contains multiple attributes, even if we fill the cache considering only one of them. Here we analyze the soft labels obtained via Eq. (8), comparing them with the One-hot attribute labels of the initial cache, reporting the results in Tab. 5. + +The first strategy is to incorporate the soft scores as they are predicted (Soft in Tab. 5). This leads to results slightly above the One-hot baselines on OVAD $(+0.3\mathrm{mAP})$ , while reducing the performance in VAW $(-2.0\mathrm{mAP})$ . This is a direct consequence of allowing each sample to have a different level of relevance in the cache, something that becomes more prominent in VAW due to the larger number of attributes. + +To fix this issue, we can normalize the scores. We study three strategies: Sharpening, where the scores are normalized as in PAWS [2]; Softmax normalization over each sample; and the one we propose in Eq. (9), i.e., using the statistics of the cache-attributes similarity. Sharpening is not effective in this scenario, achieving the lowest performance across both settings. Softmax normalization leads to good results on VAW, e.g., +3.3 mAP w.r.t. One-hot, but it struggles on OVAD (24.5 mAP) due to the limited range of values. + +Our strategy addresses both issues simultaneously, consistently improving metrics in both settings compared to + +the One-hot strategy, scoring $27.4\mathrm{mAP}$ on OVAD and 58.1 mAP on VAW. Note that these findings are general. Fig. 6 shows soft labels consistently enhance results for estimating attribute-object priors across all strategies and datasets. Improvements range from $+3.8\mathrm{mAP}$ on VAW to $+5.4\mathrm{mAP}$ on OVAD in the LLM-only setting. + +Number of samples per attribute. In Fig. 5, we measure how adding more samples (varying $K$ ) per attribute influences performance on the two datasets. We observe performance consistently increasing for OVAD while going from 1 to 16 samples, reaching a plateau when moving to $K = 24$ . On VAW, performance is stable for different values of $K$ due to the large number of attributes in the dataset (620), making the cache representative enough even when $K = 1$ . These results confirm the stability of COMCA. + +# 5. Conclusion + +We introduced Compositional Caching for open-vocabulary attribute detection. COMCA uses a VLM and an auxiliary cache of images to solve this task in a training-free manner. Differently from previous works [48, 57], we construct the cache based on attribute-object compatibility derived from web-scale databases and LLMs. We also introduced a soft-labeling scheme to account for the co-occurrence of multiple attributes within an image. Combining these two components allows to model the natural co-occurrence of attributes in the real world, together and across objects, while making the cache construction scalable. Our experiments demonstrate that COMCA is a strong training-free approach. Since COMCA is training-free, it does not suffer from differences between training and test attributes, contrary to training-based methods. This also enables COMCA to be adapted to special domains (e.g., aerial images) more easily than training-based methods, which would require specific re-training. Future work will focus on devising methods to address attribute detection even in more challenging settings where objects or attributes are not given [10]. + +Acknowledgments. We acknowledge ISCRA for awarding this project access to the LEONARDO supercomputer, owned by the EuroHPC Joint Undertaking, hosted by CINECA (Italy). This work was sponsored by the MUR PNRR project FAIR - Future AI Research (PE00000013), funded by NextGeneration EU, and the EU projects AI4TRUST (No.101070190) and ELIAS (No.01120237). + +# References + +[1] Zeynep Akata, Florent Perronnin, Zaid Harchaoui, and Cordelia Schmid. Label-embedding for attribute-based classification. In CVPR, 2013. 2 +[2] Mahmoud Assran, Mathilde Caron, Ishan Misra, Piotr Bojanowski, Armand Joulin, Nicolas Ballas, and Michael Rabbat. Semi-supervised learning of visual features by nonparametrically predicting view assignments with support samples. In ICCV, 2021. 8 +[3] Alberto Baldrati, Lorenzo Agnolucci, Marco Bertini, and Alberto Del Bimbo. Zero-shot composed image retrieval with textual inversion. In ICCV, 2023. 2 +[4] Tamara L Berg, Alexander C Berg, and Jonathan Shih. Automatic attribute discovery and characterization from noisy web data. In ECCV, 2010. 2 +[5] Maria A Bravo, Sudhanshu Mittal, Simon Ging, and Thomas Brox. Open-vocabulary attribute detection. In CVPR, 2023. 1, 2, 3, 4, 5, 6, 7, 8 +[6] Soravit Changpinyo, Piyush Sharma, Nan Ding, and Radu Soricut. Conceptual 12m: Pushing web-scale image-text pre-training to recognize long-tail visual concepts. In CVPR, 2021. 2, 5 +[7] Keyan Chen, Xiaolong Jiang, Yao Hu, Xu Tang, Yan Gao, Jianqi Chen, and Weidi Xie. Ovarnet: Towards open-vocabulary object attribute recognition. In CVPR, 2023. 1, 2, 3, 5, 6, 7 +[8] Steven Chen and Kristen Grauman. Compare and contrast: Learning prominent visual differences. In CVPR, 2018. 2 +[9] Mehdi Cherti, Romain Beaumont, Ross Wightman, Mitchell Wortsman, Gabriel Ilharco, Cade Gordon, Christoph Schuhmann, Ludwig Schmidt, and Jenia Jitsev. Reproducible scaling laws for contrastive language-image learning. In CVPR, 2023. 5, 6, 7, 8 +[10] Alessandro Conti, Enrico Fini, Massimiliano Mancini, Paolo Rota, Yiming Wang, and Elisa Ricci. Vocabulary-free image classification. NeurIPS, 36, 2024. 8 +[11] Ali Farhadi, Ian Endres, Derek Hoiem, and David Forsyth. Describing objects by their attributes. In CVPR, 2009. 2 +[12] Ali Farhadi, Ian Endres, and Derek Hoiem. Attribute-centric recognition for cross-category generalization. In CVPR, 2010. +[13] Vittorio Ferrari and Andrew Zisserman. Learning visual attributes. NeurIPS, 2007. 2 +[14] Isabel O Gallegos, Ryan A Rossi, Joe Barrow, Md Mehrab Tanjim, Sungchul Kim, Franck Dernoncourt, Tong Yu, Ruiyi Zhang, and Nesreen K Ahmed. Bias and fairness in large language models: A survey. arXiv preprint arXiv:2309.00770, 2023. 4 + +[15] Peng Gao, Shijie Geng, Renrui Zhang, Teli Ma, Rongyao Fang, Yongfeng Zhang, Hongsheng Li, and Yu Qiao. Clip-adapter: Better vision-language models with feature adapters. IJCV, 132(2):581-595, 2024. 2 +[16] Xiaoyuan Guo, Kezhen Chen, Jinmeng Rao, Yawen Zhang, Baochen Sun, and Jie Yang. Lowa: Localize objects in the wild with attributes. In NeurIPS-WS, 2023. 1, 2, 5, 6, 7 +[17] Ziyu Guo, Renrui Zhang, Longtian Qiu, Xianzheng Ma, Xu-peng Miao, Xuming He, and Bin Cui. Calip: Zero-shot enhancement of clip with parameter-free attention. In AAAI, 2023. 2 +[18] Xintong Han, Zuxuan Wu, Phoenix X Huang, Xiao Zhang, Menglong Zhu, Yuan Li, Yang Zhao, and Larry S Davis. Automatic spatially-aware fashion concept discovery. In ICCV, 2017. 2 +[19] Mehrdad Hosseinzadeh and Yang Wang. Composed query image retrieval using locally bounded features. In CVPR, 2020. 2 +[20] Edward J Hu, Phillip Wallis, Zeyuan Allen-Zhu, Yanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. In ICLR, 2021. 2 +[21] Phillip Isola, Joseph J Lim, and Edward H Adelson. Discovering states and transformations in image collections. In CVPR, 2015. 2 +[22] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In ECCV, 2022. 2 +[23] Glenn Jocher and Jing Qiu. Ultralytics yolo11. 2024. 5, 6 +[24] Shyamgopal Karthik, Karsten Roth, Massimiliano Mancini, and Zeynep Akata. Vision-by-language for training-free compositional image retrieval. In ICLR, 2024. 2 +[25] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. IJCV, 123:32-73, 2017. 2 +[26] Christoph H Lampert, Hannes Nickisch, and Stefan Harmeling. Learning to detect unseen object classes by between-class attribute transfer. In CVPR, 2009. 2, 5 +[27] Christoph H Lampert, Hannes Nickisch, and Stefan Harmeling. Attribute-based classification for zero-shot visual object categorization. IEEE TPAMI, 36(3):453-465, 2013. 2, 5, 6, 7 +[28] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, 2022. 2, 5, 6, 7 +[29] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML, 2023. 2 +[30] Victor Weixin Liang, Yuhui Zhang, Yongchan Kwon, Serena Yeung, and James Y Zou. Mind the gap: Understanding the modality gap in multi-modal contrastive representation learning. NeurIPS, 2022. 5 +[31] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. NeurIPS, 2024. 2 + +[32] Sachit Menon and Carl Vondrick. Visual classification via description from large language models. In ICLR, 2023. 2 +[33] Roberto Navigli, Simone Conia, and Björn Ross. Biases in large language models: origins, inventory, and discussion. ACM Journal of Data and Information Quality, 15(2):1-21, 2023. 4 +[34] Zachary Novack, Julian McAuley, Zachary Chase Lipton, and Saurabh Garg. Chils: Zero-shot image classification with hierarchical label sets. In ICML, 2023. 2 +[35] OpenAI. Introducing chatgpt. https://openai.com/blog/chatgpt, 2022.5 +[36] Emin Orhan. A simple cache model for image recognition. NeurIPS, 2018. 3 +[37] Devi Parikh and Kristen Grauman. Relative attributes. In ICCV, 2011. 2 +[38] Genevieve Patterson and James Hays. Coco attributes: Attributes for people, animals, and objects. In ECCV, 2016. 1, 2, 5, 6 +[39] Khoi Pham, Kushal Kafle, Zhe Lin, Zhihong Ding, Scott Cohen, Quan Tran, and Abhinav Shrivastava. Learning to predict visual attributes in the wild. In CVPR, 2021. 2, 3, 5, 6, 7, 8 +[40] Khoi Pham, Kushal Kafle, Zhe Lin, Zhihong Ding, Scott Cohen, Quan Tran, and Abhinav Shrivastava. Improving closed and open-vocabulary attribute prediction using transformers. In ECCV, 2022. 1, 2 +[41] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 1, 2, 3 +[42] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 3, 4 +[43] Karsten Roth, Jae Myung Kim, A Koepke, Oriol Vinyals, Cordelia Schmid, and Zeynep Akata. Waffling around for performance: Visual classification with random words and broad concepts. In ICCV, 2023. 2 +[44] Kuniaki Saito, Kihyuk Sohn, Xiang Zhang, Chun-Liang Li, Chen-Yu Lee, Kate Saenko, and Tomas Pfister. Pic2word: Mapping pictures to words for zero-shot composed image retrieval. In CVPR, 2023. 2 +[45] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. NeurIPS, 2022. 3, 5 +[46] Behjat Siddique, Rogerio S Feris, and Larry S Davis. Image ranking and retrieval based on multi-attribute queries. In CVPR, 2011. 2 +[47] Kihyuk Sohn, Huiwen Chang, José Lezama, Luisa Polania, Han Zhang, Yuan Hao, Irfan Essa, and Lu Jiang. Visual prompt tuning for generative transfer learning. In CVPR, 2023. 2 +[48] Vishaal Udandarao, Ankush Gupta, and Samuel Albanie. Susx: Training-free name-only transfer of vision-language models. In ICCV, 2023. 2, 3, 4, 5, 6, 7, 8 + +[49] Peter Welinder, Steve Branson, Takeshi Mita, Catherine Wah, Florian Schroff, Serge Belongie, and Pietro Perona. Caltech-ucsd birds 200. 2010. 2 +[50] Hui Wu, Yupeng Gao, Xiaoxiao Guo, Ziad Al-Halah, Steven Rennie, Kristen Grauman, and Rogerio Feris. Fashion iq: A new dataset towards retrieving images by natural language feedback. In CVPR, 2021. 2 +[51] Yongqin Xian, Christoph H Lampert, Bernt Schiele, and Zeynep Akata. Zero-shot learning—a comprehensive evaluation of the good, the bad and the ugly. IEEE TPAMI, 41(9): 2251-2265, 2018. 2 +[52] Hantao Yao, Rui Zhang, and Changsheng Xu. Visual-language prompt tuning with knowledge-guided context optimization. In CVPR, 2023. 2 +[53] Aron Yu and Kristen Grauman. Fine-grained visual comparisons with local learning. In CVPR, 2014. 2 +[54] Jiahui Yu, Zirui Wang, Vijay Vasudevan, Legg Yeung, Mojtaba Seyedhosseini, and Yonghui Wu. Coca: Contrastive captioners are image-text foundation models. TMLR, 2022. 2, 5, 6, 7 +[55] Yan Zeng, Xinsong Zhang, and Hang Li. Multi-grained vision language pre-training: Aligning texts with visual concepts. ICML, 2022. 7 +[56] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In ICCV, 2023. 5, 6, 7 +[57] Renrui Zhang, Wei Zhang, Rongyao Fang, Peng Gao, Kunchang Li, Jifeng Dai, Yu Qiao, and Hongsheng Li. Tip-adapter: Training-free adaption of clip for few-shot classification. In ECCV, 2022. 2, 3, 5, 6, 7, 8 +[58] Xiao Zhang, Rui Zhao, Yu Qiao, Xiaogang Wang, and Hongsheng Li. Adacos: Adaptively scaling cosine logits for effectively learning deep face representations. In CVPR, 2019. 5 +[59] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Conditional prompt learning for vision-language models. In CVPR, 2022. 2 +[60] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. IJCV, 130(9):2337-2348, 2022. 2 +[61] Yifei Zhou, Juntao Ren, Fengyu Li, Ramin Zabih, and Ser Nam Lim. Test-time distribution normalization for contrastively learned visual-language models. NeurIPS, 2024. 2 +[62] Beier Zhu, Yulei Niu, Yucheng Han, Yue Wu, and Hanwang Zhang. Prompt-aligned gradient for prompt tuning. In ICCV, 2023. 2 +[63] William Y Zhu, Keren Ye, Junjie Ke, Jiahui Yu, Leonidas Guibas, Peyman Milanfar, and Feng Yang. Artvlm: Attribute recognition through vision-based prefix language modeling. ECCV, 2024. 2, 5, 6 +[64] Xiangyang Zhu, Renrui Zhang, Bowei He, Aojun Zhou, Dong Wang, Bin Zhao, and Peng Gao. Not all features matter: Enhancing few-shot clip with adaptive prior refinement. In ICCV, 2023. 3 \ No newline at end of file diff --git a/2025/Compositional Caching for Training-free Open-vocabulary Attribute Detection/images.zip b/2025/Compositional Caching for Training-free Open-vocabulary Attribute Detection/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..9f658e3b591975eb28a1846f4d871654eb984a6f --- /dev/null +++ b/2025/Compositional Caching for Training-free Open-vocabulary Attribute Detection/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ca4d6132d6bddc82d7aebe0822d677e773c1e148777347296fbdd96c51420340 +size 475145 diff --git a/2025/Compositional Caching for Training-free Open-vocabulary Attribute Detection/layout.json b/2025/Compositional Caching for Training-free Open-vocabulary Attribute Detection/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..f09dc67d7433220fedf96bdd5fa145f636785ae9 --- /dev/null +++ b/2025/Compositional Caching for Training-free Open-vocabulary Attribute Detection/layout.json @@ -0,0 +1,9924 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 62, + 103, + 547, + 120 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 103, + 547, + 120 + ], + "spans": [ + { + "bbox": [ + 62, + 103, + 547, + 120 + ], + "type": "text", + "content": "Compositional Caching for Training-free Open-vocabulary Attribute Detection" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 80, + 142, + 527, + 184 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 142, + 527, + 184 + ], + "spans": [ + { + "bbox": [ + 80, + 142, + 527, + 184 + ], + "type": "text", + "content": "Marco Garosi1 Alessandro Conti1 Gaowen Liu2 Elisa Ricci1,3 Massimiliano Mancini1 \n1University of Trento 2Cisco Research Fondazione Bruno Kessler \nhttps://comca-attributes. github.io" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 152, + 213, + 199, + 224 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 213, + 199, + 224 + ], + "spans": [ + { + "bbox": [ + 152, + 213, + 199, + 224 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 238, + 297, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 238, + 297, + 561 + ], + "spans": [ + { + "bbox": [ + 55, + 238, + 297, + 561 + ], + "type": "text", + "content": "Attribute detection is crucial for many computer vision tasks, as it enables systems to describe properties such as color, texture, and material. Current approaches often rely on labor-intensive annotation processes which are inherently limited: objects can be described at an arbitrary level of detail (e.g., color vs. color shades), leading to ambiguities when the annotators are not instructed carefully. Furthermore, they operate within a predefined set of attributes, reducing scalability and adaptability to unforeseen downstream applications. We present COMPOSITIONAL CACHING (COMCA), a training-free method for open-vocabulary attribute detection that overcomes these constraints. COMCA requires only the list of target attributes and objects as input, using them to populate an auxiliary cache of images by leveraging web-scale databases and Large Language Models to determine attribute-object compatibility. To account for the compositional nature of attributes, cache images receive soft attribute labels. Those are aggregated at inference time based on the similarity between the input and cache images, refining the predictions of underlying Vision-Language Models (VLMs). Importantly, our approach is model-agnostic, compatible with various VLMs. Experiments on public datasets demonstrate that COMCA significantly outperforms zero-shot and cache-based baselines, competing with recent training-based methods, proving that a carefully designed training-free approach can successfully address open-vocabulary attribute detection." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 586, + 135, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 586, + 135, + 597 + ], + "spans": [ + { + "bbox": [ + 56, + 586, + 135, + 597 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "content": "Attributes shape our perception and interactions with the environment. Let us consider the example of Fig. 1. If our aim is to identify the animal, we may focus on its fur color, size, the presence of horns and the tail. Similarly, if we want to analyze its health status, we may focus on its age, fur patterns and shades, and its skin. This level of analysis requires models with fine-grained understanding of the visual input, which is usually obtained by training a model on datasets that annotate both objects and their corresponding" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 317, + 213, + 553, + 329 + ], + "blocks": [ + { + "bbox": [ + 317, + 213, + 553, + 329 + ], + "lines": [ + { + "bbox": [ + 317, + 213, + 553, + 329 + ], + "spans": [ + { + "bbox": [ + 317, + 213, + 553, + 329 + ], + "type": "image", + "image_path": "db59c2c80abbb2e7c48cbb7b17ea624025591c2b0ee627a3ce4681338528f924.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 339, + 555, + 416 + ], + "lines": [ + { + "bbox": [ + 313, + 339, + 555, + 416 + ], + "spans": [ + { + "bbox": [ + 313, + 339, + 555, + 416 + ], + "type": "text", + "content": "Figure 1. Example of attribute annotations. Attribute annotations are often: sparse, as they are not consistent across samples; incomplete, as not all attributes are annotated; ambiguous, as they can be subjective or miss a frame of reference. This makes open-vocabulary attribute detection challenging. Differently from previous works [5, 7], we do not rely on such annotations and propose COMCA, a training-free approach requiring no supervision." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 425, + 556, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 425, + 556, + 556 + ], + "spans": [ + { + "bbox": [ + 313, + 425, + 556, + 556 + ], + "type": "text", + "content": "attributes [38, 40]. However, collecting attribute annotations is both time-consuming and prone to ambiguities, especially when annotators lack clear instructions. Additionally, attribute annotations are often tailored to specific datasets and are ambiguous. For instance, in Fig. 1, one may list \"brown\" as a color for the left goat, but others may argue that it is a red-ish shade of brown. A simple solution is to provide annotators with a predefined list of attributes; however, (i) the list may not cover all possible annotations for a property (e.g., all colors) and (ii) it limits models trained on it to the given list itself, making them unable to recognize new attributes." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 557, + 556, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 557, + 556, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 557, + 556, + 653 + ], + "type": "text", + "content": "Recent approaches overcome some of the challenges by developing open-vocabulary attribute detection models [5, 7, 16] that detect attributes specified in natural language at inference time. While these approaches generalize beyond the set of seen attributes and objects, their performance is still biased toward the specific dataset they are trained on. Indeed, training-based approaches may show significant degradation in performance in cross-dataset experiments (see Fig. 4)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 654, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 556, + 713 + ], + "type": "text", + "content": "In this work, we overcome potentially ambiguous and/or limited attribute annotations by introducing COMPOSITIONAL CACHING (COMCA), a training-free open-vocabulary attribute detection model. COMCA employs a Vision-Language Model (VLM) [41] and refines its" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "15098" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 96 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 96 + ], + "type": "text", + "content": "predictions by utilizing an auxiliary cache that contains images together with automatically estimated attribute labels." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 99, + 297, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 99, + 297, + 266 + ], + "spans": [ + { + "bbox": [ + 54, + 99, + 297, + 266 + ], + "type": "text", + "content": "Caching for refining CLIP predictions has been introduced in the context of image classification [57]. It requires storing images, either generated or retrieved [48, 57], for each of the classes of interests, together with a corresponding one-hot categorical label. The similarity of the input image with each of the examples is used to weight the labels in the cache and refine the model's prediction. While this approach is effective for object classification, it is prohibitively expensive in the context of attribute detection due to the much larger output space, comprising all possible combinations of attributes and objects. COMCA achieves scalability in caching by following two key compositional principles: (i) not all attributes are universally applicable to all objects, and (ii) every object inherently possesses multiple attributes." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 269, + 298, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 269, + 298, + 544 + ], + "spans": [ + { + "bbox": [ + 56, + 269, + 298, + 544 + ], + "type": "text", + "content": "The first principle is implemented in two steps: attribute-object compatibility estimation and cache pair sampling. In the first phase, we estimate the probability of an attribute appearing with an object using the statistics of available databases (e.g., CC12M [6]) and Large Language Model (LLM) queries. Next, we populate the cache for a given attribute by sampling compatible objects from the estimated distribution and retrieving relevant images from a database. The second principle is implemented by self-labeling the cached elements. Initially, each sample is associated with the single attribute it was sampled for. Since this does not account for other attributes present in the image, we extend attribute labels through soft-labeling, i.e., comparing cache images to text embeddings of each attribute of interest. These scores modify a cache element's contribution to each target attribute, considering their implicit composition. We evaluate COMCA on two publicly available benchmarks, OVAD [5] and VAW [39]. COMCA demonstrates significant improvements over training-free baselines and is competitive with recent training-based approaches, while remaining unaffected by the biases of specific training sets. Notably, COMCA provides a consistent boost in performance regardless of the VLM pre-training and architecture." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 546, + 297, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 546, + 297, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 546, + 297, + 713 + ], + "type": "text", + "content": "In summary, our contributions are: ① We propose COMCA, an approach specifically designed for training-free open-vocabulary attribute detection. COMCA exploits the compositional nature of objects and attributes to build an auxiliary cache that is used as an anchor to refine the model's predictions. ② We make caching scalable by extracting priors from databases and LLMs on which objects are associated with each attribute, using them to populate the cache. Soft-labeling further accounts for the multiplicity of attributes within an image. ③ We empirically show that COMCA outperforms all the cache-based baselines and achieves competitive performance against costly training-based methods, demonstrating the practical feasibility of training-free open vocabulary attribute detection." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 314, + 72, + 399, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 399, + 83 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 399, + 83 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 91, + 555, + 150 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 91, + 555, + 150 + ], + "spans": [ + { + "bbox": [ + 313, + 91, + 555, + 150 + ], + "type": "text", + "content": "Attribute detection. Detecting visual attributes is a widely explored task in computer vision [4, 11-13], supporting various downstream applications such as zero-shot learning [1, 26, 27, 51], relative image comparison [8, 37, 46, 53], and compositional image retrieval [3, 19, 24, 44]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 152, + 556, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 152, + 556, + 258 + ], + "spans": [ + { + "bbox": [ + 313, + 152, + 556, + 258 + ], + "type": "text", + "content": "Recent works have introduced several benchmarks to promote and ease research in this area [18, 21, 25, 38-40, 49, 50, 53]. Despite these advances, annotating visual attributes remains a complex and challenging task. This is due to the wide variability in attributes across objects and contexts, as well as the difficulty of ensuring consistent annotations when multiple annotators are involved. Moreover, many datasets capture only a limited range of object attributes, limiting their usefulness for developing general-purpose models." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 259, + 556, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 259, + 556, + 558 + ], + "spans": [ + { + "bbox": [ + 313, + 259, + 556, + 558 + ], + "type": "text", + "content": "Open-vocabulary attribute detection (OVAD) [5, 7] overcomes some of these limitations by detecting user-specified attributes, typically provided in natural language, even those that have not been observed during training. This task was first explored in [5], which introduced the OVAD benchmark and method. OVAD generates supervisory signals for attributes, nouns, and noun phrases by parsing captions and encoding them using CLIP text embeddings [41]. OvarNet [7] builds on this by fine-tuning CLIP with federated learning and additional supervision from external captions. LOWA [16] proposes a three-stage pipeline that aligns visual embeddings to objects, attributes, and free-form text. ArtVLM [63] uses a generative VLM to obtain attribute probabilities based on the input image. Although these methods operate in the open-vocabulary setting, they are training-based and require both a lot of annotated data and several computational resources (see Sec. 4). COMCA significantly differs from these approaches, as we explore for the first time a training-free, cache-based, strategy for attribute detection. Adapting Vision-Language Models. VLMs [28, 29, 31, 41, 54] bridge the gap between textual and visual data, showing remarkable zero-shot generalization abilities. Recent efforts have focused on improving their performance on specific downstream tasks through various strategies, broadly categorized into training-based and training-free methods." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 558, + 556, + 642 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 556, + 642 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 556, + 642 + ], + "type": "text", + "content": "Training-based approaches typically involve tuning text prompts [22, 47, 52, 59, 62]. Some methods learn prompts directly [60] or predict them [59] from visual inputs to improve generalization. Other approaches train adapters to refine the model's predictions [15, 20]. While effective, these methods introduce additional parameters, which can negatively impact the model's zero-shot performance [59]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 642, + 557, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 557, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 557, + 714 + ], + "type": "text", + "content": "Training-free methods take a different approach, such as incorporating additional knowledge into the prompts using natural language [32, 34, 43], or modifying predictions through attention layers [17] or normalization techniques [61]. Recent works also employ a cache of images and one-hot labels that represent the target classes [48, 57," + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15099" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 70, + 553, + 228 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 553, + 228 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 553, + 228 + ], + "type": "image", + "image_path": "ccb2c6f9ec37d0895662b61503a8abb082efad60b0167607ebd2378b1cad974d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 237, + 553, + 270 + ], + "lines": [ + { + "bbox": [ + 54, + 237, + 553, + 270 + ], + "spans": [ + { + "bbox": [ + 54, + 237, + 553, + 270 + ], + "type": "text", + "content": "Figure 2. COMCA's cache construction. Given a list of attributes and objects, we compute their compatibility from a large database " + }, + { + "bbox": [ + 54, + 237, + 553, + 270 + ], + "type": "inline_equation", + "content": "D_r" + }, + { + "bbox": [ + 54, + 237, + 553, + 270 + ], + "type": "text", + "content": " and with an LLM. The scores are merged and normalized to obtain the compatibility distribution, from which we sample cache entries and construct the cache. We enrich the latter with soft labels from the VLM-based similarity between cache images and attributes." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 277, + 297, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 277, + 297, + 422 + ], + "spans": [ + { + "bbox": [ + 54, + 277, + 297, + 422 + ], + "type": "text", + "content": "[64]. For example, TIP-Adapter [57] computes the similarity between input images and cached examples, using the associated labels for prediction. Similarly, SuS-X [48] creates a cache automatically, either by using a generative model [42] or retrieving from a large database [45], with distance-based scoring to weight the elements' contribution. While our work builds upon the same principles of TIP-Adapter and SuS-X, i.e., using a cache to refine predictions, it differs in four key aspects, as COMCA's cache: (i) is aware of attribute-object compatibility; (ii) accounts for attribute combinations; (iii) models the compositional nature of attributes and objects; (iv) scales efficiently with the number of attributes." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 434, + 192, + 449 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 434, + 192, + 449 + ], + "spans": [ + { + "bbox": [ + 55, + 434, + 192, + 449 + ], + "type": "text", + "content": "3. Compositional Caching" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 456, + 297, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 456, + 297, + 541 + ], + "spans": [ + { + "bbox": [ + 54, + 456, + 297, + 541 + ], + "type": "text", + "content": "In this section, we formally define the open-vocabulary attribute detection problem and describe caching as a tool for VLM adaptation (Sec. 3.1). Next, we describe how attribute-object compatibility informs the construction of scalable caches (Sec. 3.2). Finally, we detail our process for refining cache labels, taking into account multi-attribute compatibility (Sec. 3.3). Fig. 2 provides an overview of our approach." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 550, + 137, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 550, + 137, + 562 + ], + "spans": [ + { + "bbox": [ + 55, + 550, + 137, + 562 + ], + "type": "text", + "content": "3.1. Background" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 568, + 295, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 568, + 295, + 688 + ], + "spans": [ + { + "bbox": [ + 54, + 568, + 295, + 688 + ], + "type": "text", + "content": "Problem formulation. Given an image " + }, + { + "bbox": [ + 54, + 568, + 295, + 688 + ], + "type": "inline_equation", + "content": "x \\in \\mathcal{X}" + }, + { + "bbox": [ + 54, + 568, + 295, + 688 + ], + "type": "text", + "content": " of an object and a list of attributes, our goal is to detect which attributes are present in the image. Let " + }, + { + "bbox": [ + 54, + 568, + 295, + 688 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 54, + 568, + 295, + 688 + ], + "type": "text", + "content": " be the set of all existing attributes in natural language. At inference time, the model is provided with a set of " + }, + { + "bbox": [ + 54, + 568, + 295, + 688 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 54, + 568, + 295, + 688 + ], + "type": "text", + "content": " target attributes " + }, + { + "bbox": [ + 54, + 568, + 295, + 688 + ], + "type": "inline_equation", + "content": "A = \\{a_{1},\\dots ,a_{n}\\}" + }, + { + "bbox": [ + 54, + 568, + 295, + 688 + ], + "type": "text", + "content": " as input, where " + }, + { + "bbox": [ + 54, + 568, + 295, + 688 + ], + "type": "inline_equation", + "content": "A \\subset \\mathcal{A}" + }, + { + "bbox": [ + 54, + 568, + 295, + 688 + ], + "type": "text", + "content": ". These are associated with " + }, + { + "bbox": [ + 54, + 568, + 295, + 688 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 54, + 568, + 295, + 688 + ], + "type": "text", + "content": " target objects, which we denote as " + }, + { + "bbox": [ + 54, + 568, + 295, + 688 + ], + "type": "inline_equation", + "content": "O = \\{o_1,\\dots ,o_m\\}" + }, + { + "bbox": [ + 54, + 568, + 295, + 688 + ], + "type": "text", + "content": ". Our goal is to build a function " + }, + { + "bbox": [ + 54, + 568, + 295, + 688 + ], + "type": "inline_equation", + "content": "f:\\mathcal{X}\\times \\mathcal{A}^n\\to [-1,1]^n" + }, + { + "bbox": [ + 54, + 568, + 295, + 688 + ], + "type": "text", + "content": " that maps an image and a list of attributes to a binary vector indicating the presence (1) or absence (-1) of each attribute." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "We obtain " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": " without training, in contrast to previous works that learn " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": " using a dataset " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": " of image-label pairs" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 277, + 555, + 313 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 277, + 555, + 313 + ], + "spans": [ + { + "bbox": [ + 313, + 277, + 555, + 313 + ], + "type": "inline_equation", + "content": "(x, \\mathbf{a})" + }, + { + "bbox": [ + 313, + 277, + 555, + 313 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 313, + 277, + 555, + 313 + ], + "type": "inline_equation", + "content": "\\mathbf{a} \\in [-1, 1]^n" + }, + { + "bbox": [ + 313, + 277, + 555, + 313 + ], + "type": "text", + "content": " being the ground-truth binary vector1. We achieve this by implementing " + }, + { + "bbox": [ + 313, + 277, + 555, + 313 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 313, + 277, + 555, + 313 + ], + "type": "text", + "content": " using a VLM combined with an auxiliary cache to refine its predictions." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 314, + 556, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 314, + 556, + 469 + ], + "spans": [ + { + "bbox": [ + 313, + 314, + 556, + 469 + ], + "type": "text", + "content": "Adapting VLMs with caching. Following previous works [5, 7], we implement " + }, + { + "bbox": [ + 313, + 314, + 556, + 469 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 313, + 314, + 556, + 469 + ], + "type": "text", + "content": " with a VLM. While we primarily consider CLIP [41] to describe our method and perform our evaluation, the proposed framework can integrate several other VLMs (see Tab. 2). The function " + }, + { + "bbox": [ + 313, + 314, + 556, + 469 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 313, + 314, + 556, + 469 + ], + "type": "text", + "content": " has three elements: a visual encoder " + }, + { + "bbox": [ + 313, + 314, + 556, + 469 + ], + "type": "inline_equation", + "content": "f_{v}" + }, + { + "bbox": [ + 313, + 314, + 556, + 469 + ], + "type": "text", + "content": ", a text encoder " + }, + { + "bbox": [ + 313, + 314, + 556, + 469 + ], + "type": "inline_equation", + "content": "f_{t}" + }, + { + "bbox": [ + 313, + 314, + 556, + 469 + ], + "type": "text", + "content": ", and a prediction function. " + }, + { + "bbox": [ + 313, + 314, + 556, + 469 + ], + "type": "inline_equation", + "content": "f_{v}" + }, + { + "bbox": [ + 313, + 314, + 556, + 469 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 314, + 556, + 469 + ], + "type": "inline_equation", + "content": "f_{t}" + }, + { + "bbox": [ + 313, + 314, + 556, + 469 + ], + "type": "text", + "content": " map images in " + }, + { + "bbox": [ + 313, + 314, + 556, + 469 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 313, + 314, + 556, + 469 + ], + "type": "text", + "content": " and natural language in " + }, + { + "bbox": [ + 313, + 314, + 556, + 469 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 313, + 314, + 556, + 469 + ], + "type": "text", + "content": ", respectively, to a shared " + }, + { + "bbox": [ + 313, + 314, + 556, + 469 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 313, + 314, + 556, + 469 + ], + "type": "text", + "content": "-dimensional multimodal space " + }, + { + "bbox": [ + 313, + 314, + 556, + 469 + ], + "type": "inline_equation", + "content": "\\mathcal{Z} \\in \\mathbb{R}^d" + }, + { + "bbox": [ + 313, + 314, + 556, + 469 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 313, + 314, + 556, + 469 + ], + "type": "inline_equation", + "content": "f_{v}: \\mathcal{X} \\to \\mathcal{Z}" + }, + { + "bbox": [ + 313, + 314, + 556, + 469 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 314, + 556, + 469 + ], + "type": "inline_equation", + "content": "f_{t}: \\mathcal{T} \\to \\mathcal{Z}" + }, + { + "bbox": [ + 313, + 314, + 556, + 469 + ], + "type": "text", + "content": ". Next, the prediction function maps embeddings from the multimodal space to the final scores. A standard way for scoring an image " + }, + { + "bbox": [ + 313, + 314, + 556, + 469 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 314, + 556, + 469 + ], + "type": "text", + "content": " against a class " + }, + { + "bbox": [ + 313, + 314, + 556, + 469 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 314, + 556, + 469 + ], + "type": "text", + "content": " is by computing the cosine similarity between their embeddings:" + }, + { + "bbox": [ + 313, + 314, + 556, + 469 + ], + "type": "inline_equation", + "content": "^2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 365, + 478, + 555, + 506 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 478, + 555, + 506 + ], + "spans": [ + { + "bbox": [ + 365, + 478, + 555, + 506 + ], + "type": "interline_equation", + "content": "f _ {\\mathrm {C L I P}} (x, c) = \\frac {f _ {v} (x) ^ {\\top} f _ {t} (c)}{| | f _ {v} (x) | | \\cdot | | f _ {t} (c) | |}. \\tag {1}", + "image_path": "72a054e935c34d92816435dbd63ca16ec383252e6fd24e0892b72a8c8ab5c1e7.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 512, + 556, + 573 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 512, + 556, + 573 + ], + "spans": [ + { + "bbox": [ + 313, + 512, + 556, + 573 + ], + "type": "text", + "content": "While effective, " + }, + { + "bbox": [ + 313, + 512, + 556, + 573 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 313, + 512, + 556, + 573 + ], + "type": "text", + "content": " can be improved without any training by using an auxiliary cache that stores examples of the classes of interest [36, 48, 57]. Given a cache " + }, + { + "bbox": [ + 313, + 512, + 556, + 573 + ], + "type": "inline_equation", + "content": "\\mathcal{C} = \\{x_1^1, \\dots, x_C^K\\}" + }, + { + "bbox": [ + 313, + 512, + 556, + 573 + ], + "type": "text", + "content": " that contains " + }, + { + "bbox": [ + 313, + 512, + 556, + 573 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 512, + 556, + 573 + ], + "type": "text", + "content": " samples of " + }, + { + "bbox": [ + 313, + 512, + 556, + 573 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 313, + 512, + 556, + 573 + ], + "type": "text", + "content": " classes, we can define the cache-based prediction for a class " + }, + { + "bbox": [ + 313, + 512, + 556, + 573 + ], + "type": "inline_equation", + "content": "c \\in C" + }, + { + "bbox": [ + 313, + 512, + 556, + 573 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 345, + 581, + 555, + 613 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 581, + 555, + 613 + ], + "spans": [ + { + "bbox": [ + 345, + 581, + 555, + 613 + ], + "type": "interline_equation", + "content": "f _ {\\mathcal {C}} (x, c) = \\sum_ {i = 1} ^ {K} \\eta_ {\\mathcal {C}} \\left(\\frac {f _ {v} (x) ^ {\\intercal} f _ {v} \\left(x _ {c} ^ {i}\\right)}{\\left| \\left| f _ {v} (x) \\right| \\right| \\cdot \\left| \\left| f _ {v} \\left(x _ {c} ^ {i}\\right) \\right| \\right|}\\right), \\tag {2}", + "image_path": "eaa9edc09081c09539ec7ef9f4e61c29a4fda29b471808317cb7cf881445094c.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 623, + 555, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 623, + 555, + 647 + ], + "spans": [ + { + "bbox": [ + 313, + 623, + 555, + 647 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 623, + 555, + 647 + ], + "type": "inline_equation", + "content": "\\eta_{\\mathcal{C}}(z) = \\exp (1 - \\beta (-z))" + }, + { + "bbox": [ + 313, + 623, + 555, + 647 + ], + "type": "text", + "content": " is a normalization function regulated by the hyperparameter " + }, + { + "bbox": [ + 313, + 623, + 555, + 647 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 313, + 623, + 555, + 647 + ], + "type": "text", + "content": " [57]." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 313, + 654, + 555, + 693 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 555, + 693 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 555, + 693 + ], + "type": "text", + "content": "1In some problem formulations, the dataset may contain (i) object labels, (ii) bounding boxes, (iii) an \"unknown\" value for some attributes. For simplicity, we omit the latter from the formulation and, following previous works [39], we consider the case where objects bounding boxes are given." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 315, + 694, + 553, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 694, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 315, + 694, + 553, + 712 + ], + "type": "text", + "content": "2For simplicity, we omit the presence of prompts such as \"a photo of\" when defining the text input." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15100" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 264 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 264 + ], + "type": "text", + "content": "The formulation in Eq. (2) has a main issue when applied to open-vocabulary attribute detection: the cache construction. On one hand, a cache considering all possible attribute labels is semantically complete but requires a prohibitive storage size of " + }, + { + "bbox": [ + 55, + 72, + 296, + 264 + ], + "type": "inline_equation", + "content": "K \\cdot n!" + }, + { + "bbox": [ + 55, + 72, + 296, + 264 + ], + "type": "text", + "content": " images, e.g., " + }, + { + "bbox": [ + 55, + 72, + 296, + 264 + ], + "type": "inline_equation", + "content": "K \\cdot 117!" + }, + { + "bbox": [ + 55, + 72, + 296, + 264 + ], + "type": "text", + "content": " for OVAD [5]. On the other hand, while storing " + }, + { + "bbox": [ + 55, + 72, + 296, + 264 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 72, + 296, + 264 + ], + "type": "text", + "content": " positive examples per attribute has low memory requirements, it ignores the multi-attribute nature of images (e.g., shape, size, color, etc. can co-exist in the same sample), potentially leading to incorrect predictions (e.g., a high score because the color matches but the cache label refers to the shape). Moreover, attributes may appear differently across objects (e.g., wet car vs wet dog) and selecting a random subset of them may lead to a cache representative only of a few cases. In the following, we address these issues by focusing on the compatibility between objects and multiple attributes in the real world." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 271, + 239, + 284 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 271, + 239, + 284 + ], + "spans": [ + { + "bbox": [ + 55, + 271, + 239, + 284 + ], + "type": "text", + "content": "3.2. Compatibility for scalable caching" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 289, + 296, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 289, + 296, + 432 + ], + "spans": [ + { + "bbox": [ + 55, + 289, + 296, + 432 + ], + "type": "text", + "content": "Naively constructing the cache would lead to many problems, related both to labels' semantic and memory requirements (Sec. 3.1). In this section, we present a methodology that (i) has the same memory requirements as storing only " + }, + { + "bbox": [ + 55, + 289, + 296, + 432 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 289, + 296, + 432 + ], + "type": "text", + "content": " examples per attribute and (ii) assigns multiple attribute labels per example, taking into account attribute-object compatibility. Scalable caching. To obtain a cache " + }, + { + "bbox": [ + 55, + 289, + 296, + 432 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 55, + 289, + 296, + 432 + ], + "type": "text", + "content": " that is scalable with the number of attributes and accounts for object-specific information, we exploit the natural distributions of attributes across objects. In particular, let us define a function " + }, + { + "bbox": [ + 55, + 289, + 296, + 432 + ], + "type": "inline_equation", + "content": "\\varphi" + }, + { + "bbox": [ + 55, + 289, + 296, + 432 + ], + "type": "text", + "content": " that, given an attribute " + }, + { + "bbox": [ + 55, + 289, + 296, + 432 + ], + "type": "inline_equation", + "content": "a \\in \\mathcal{A}" + }, + { + "bbox": [ + 55, + 289, + 296, + 432 + ], + "type": "text", + "content": " and an object " + }, + { + "bbox": [ + 55, + 289, + 296, + 432 + ], + "type": "inline_equation", + "content": "o \\in \\mathcal{O}" + }, + { + "bbox": [ + 55, + 289, + 296, + 432 + ], + "type": "text", + "content": ", returns their compatibility value as a scalar, i.e., " + }, + { + "bbox": [ + 55, + 289, + 296, + 432 + ], + "type": "inline_equation", + "content": "\\varphi : \\mathcal{A} \\times \\mathcal{O} \\to \\mathbb{R}" + }, + { + "bbox": [ + 55, + 289, + 296, + 432 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 434, + 296, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 434, + 296, + 517 + ], + "spans": [ + { + "bbox": [ + 55, + 434, + 296, + 517 + ], + "type": "text", + "content": "Considering an attribute " + }, + { + "bbox": [ + 55, + 434, + 296, + 517 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 55, + 434, + 296, + 517 + ], + "type": "text", + "content": " and an object " + }, + { + "bbox": [ + 55, + 434, + 296, + 517 + ], + "type": "inline_equation", + "content": "o" + }, + { + "bbox": [ + 55, + 434, + 296, + 517 + ], + "type": "text", + "content": ", the score " + }, + { + "bbox": [ + 55, + 434, + 296, + 517 + ], + "type": "inline_equation", + "content": "\\varphi(a, o)" + }, + { + "bbox": [ + 55, + 434, + 296, + 517 + ], + "type": "text", + "content": " is proportional to the probability that " + }, + { + "bbox": [ + 55, + 434, + 296, + 517 + ], + "type": "inline_equation", + "content": "o" + }, + { + "bbox": [ + 55, + 434, + 296, + 517 + ], + "type": "text", + "content": " exhibits " + }, + { + "bbox": [ + 55, + 434, + 296, + 517 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 55, + 434, + 296, + 517 + ], + "type": "text", + "content": ". Thus, we can use the scores " + }, + { + "bbox": [ + 55, + 434, + 296, + 517 + ], + "type": "inline_equation", + "content": "\\Phi_O(a) = [\\varphi(a, o_1), \\dots, \\varphi(a, o_m)]" + }, + { + "bbox": [ + 55, + 434, + 296, + 517 + ], + "type": "text", + "content": " as a proxy for the probability distribution of " + }, + { + "bbox": [ + 55, + 434, + 296, + 517 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 55, + 434, + 296, + 517 + ], + "type": "text", + "content": " co-occurring with the elements of " + }, + { + "bbox": [ + 55, + 434, + 296, + 517 + ], + "type": "inline_equation", + "content": "O" + }, + { + "bbox": [ + 55, + 434, + 296, + 517 + ], + "type": "text", + "content": ". We then assign each of the " + }, + { + "bbox": [ + 55, + 434, + 296, + 517 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 434, + 296, + 517 + ], + "type": "text", + "content": " elements (i.e., shots) of the attribute " + }, + { + "bbox": [ + 55, + 434, + 296, + 517 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 55, + 434, + 296, + 517 + ], + "type": "text", + "content": " to one object " + }, + { + "bbox": [ + 55, + 434, + 296, + 517 + ], + "type": "inline_equation", + "content": "o" + }, + { + "bbox": [ + 55, + 434, + 296, + 517 + ], + "type": "text", + "content": ", sampling the object from " + }, + { + "bbox": [ + 55, + 434, + 296, + 517 + ], + "type": "inline_equation", + "content": "\\Phi_O(a)" + }, + { + "bbox": [ + 55, + 434, + 296, + 517 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 85, + 524, + 296, + 540 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 524, + 296, + 540 + ], + "spans": [ + { + "bbox": [ + 85, + 524, + 296, + 540 + ], + "type": "interline_equation", + "content": "\\hat {O} _ {a} ^ {K} = \\left\\{\\hat {o} _ {1} ^ {a}, \\dots , \\hat {o} _ {K} ^ {a} \\right\\}, \\text {w h e r e} \\hat {o} _ {i} ^ {a} \\sim \\tilde {\\Phi} _ {O} (a), \\tag {3}", + "image_path": "d338b9fc5c7b76fc352a2ae4ccb1e857a17b30167d456e59b116892310c93a2b.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 548, + 296, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 548, + 296, + 608 + ], + "spans": [ + { + "bbox": [ + 55, + 548, + 296, + 608 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 548, + 296, + 608 + ], + "type": "inline_equation", + "content": "\\tilde{\\Phi}_O(a)" + }, + { + "bbox": [ + 55, + 548, + 296, + 608 + ], + "type": "text", + "content": " is the normalized version of " + }, + { + "bbox": [ + 55, + 548, + 296, + 608 + ], + "type": "inline_equation", + "content": "\\Phi_O(a)" + }, + { + "bbox": [ + 55, + 548, + 296, + 608 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 55, + 548, + 296, + 608 + ], + "type": "inline_equation", + "content": "\\tilde{\\Phi}_O(a) = \\Phi_O(a) / \\sum \\Phi_O(a)" + }, + { + "bbox": [ + 55, + 548, + 296, + 608 + ], + "type": "text", + "content": ". The set " + }, + { + "bbox": [ + 55, + 548, + 296, + 608 + ], + "type": "inline_equation", + "content": "\\hat{O}_a^K" + }, + { + "bbox": [ + 55, + 548, + 296, + 608 + ], + "type": "text", + "content": " allows populating the cache with relevant objects and attributes. Let us denote a generic function mapping a textual query to an image as " + }, + { + "bbox": [ + 55, + 548, + 296, + 608 + ], + "type": "inline_equation", + "content": "\\mathrm{T2I}:\\mathcal{T}\\to \\mathcal{X}" + }, + { + "bbox": [ + 55, + 548, + 296, + 608 + ], + "type": "text", + "content": ". The elements of the cache are:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 71, + 615, + 296, + 644 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 615, + 296, + 644 + ], + "spans": [ + { + "bbox": [ + 71, + 615, + 296, + 644 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {C} _ {\\mathrm {H}} = \\left\\{x _ {a _ {1}} ^ {1}, \\dots , x _ {a _ {m}} ^ {K} \\right\\} (4) \\\\ = \\{\\mathrm {T 2 I} \\left(\\pi_ {g} \\circ a \\circ o _ {1} ^ {a}\\right), \\dots , \\mathrm {T 2 I} \\left(\\pi_ {g} \\circ a \\circ o _ {K} ^ {a}\\right) \\}, (4) \\\\ \\end{array}", + "image_path": "0182251f4676d547c1526284e3405bbecc1d18889b71296ff22d7998d5e050bc.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\pi_g" + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "content": " is a contextual prompt for the image (e.g., \"a photo of\"). Following SuS-X [48], we implement T2I via retrieval. Specifically, we retrieve images from a database by scoring them against textual queries with the VLM (e.g., CLIP). While we focus on retrieval, COMCA" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "type": "text", + "content": "also supports the implementation of T2I via a text-to-image generative model, e.g., by feeding textual queries to Stable Diffusion [42]. We refer to the Supp. Mat. for more details. Estimating attribute-object compatibility. The key element that enables scalable caching is the function " + }, + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "type": "inline_equation", + "content": "\\varphi" + }, + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "type": "text", + "content": " that estimates the viability of attribute-object compositions. We implement it by merging two scoring strategies." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 156, + 555, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 156, + 555, + 204 + ], + "spans": [ + { + "bbox": [ + 313, + 156, + 555, + 204 + ], + "type": "text", + "content": "The first uses a generic large-scale database " + }, + { + "bbox": [ + 313, + 156, + 555, + 204 + ], + "type": "inline_equation", + "content": "D_r" + }, + { + "bbox": [ + 313, + 156, + 555, + 204 + ], + "type": "text", + "content": " containing image-caption pairs. Using the captions " + }, + { + "bbox": [ + 313, + 156, + 555, + 204 + ], + "type": "inline_equation", + "content": "t \\in D_r" + }, + { + "bbox": [ + 313, + 156, + 555, + 204 + ], + "type": "text", + "content": ", we can estimate the database score as the number of co-occurrences of attribute " + }, + { + "bbox": [ + 313, + 156, + 555, + 204 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 313, + 156, + 555, + 204 + ], + "type": "text", + "content": " and object " + }, + { + "bbox": [ + 313, + 156, + 555, + 204 + ], + "type": "inline_equation", + "content": "o" + }, + { + "bbox": [ + 313, + 156, + 555, + 204 + ], + "type": "text", + "content": ", i.e.:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 354, + 212, + 555, + 239 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 212, + 555, + 239 + ], + "spans": [ + { + "bbox": [ + 354, + 212, + 555, + 239 + ], + "type": "interline_equation", + "content": "\\varphi_ {\\mathrm {d b}} (a, o) = \\sum_ {(x, t) \\in D _ {r}} \\mathbb {1} (o \\subset t \\wedge a \\subset t), \\tag {5}", + "image_path": "a63bf6beb5380229bba5c61b39013ab80bfc674c9981bbda2e4fa4a6d616507f.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 248, + 555, + 271 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 248, + 555, + 271 + ], + "spans": [ + { + "bbox": [ + 313, + 248, + 555, + 271 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 248, + 555, + 271 + ], + "type": "inline_equation", + "content": "\\mathbb{1}(z)" + }, + { + "bbox": [ + 313, + 248, + 555, + 271 + ], + "type": "text", + "content": " is an indicator function whose value is 1 whenever the condition " + }, + { + "bbox": [ + 313, + 248, + 555, + 271 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 313, + 248, + 555, + 271 + ], + "type": "text", + "content": " is true, and 0 otherwise." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 272, + 554, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 272, + 554, + 344 + ], + "spans": [ + { + "bbox": [ + 313, + 272, + 554, + 344 + ], + "type": "text", + "content": "While this process exploits real data, it may be biased on the specific captions and/or data collection process. For instance, captions rarely describe objects in their default state, such as a dry dog. Therefore, a second scoring strategy exploits the reasoning capabilities of an LLM, querying it to predict an attribute-object compatibility as:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 378, + 354, + 555, + 366 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 378, + 354, + 555, + 366 + ], + "spans": [ + { + "bbox": [ + 378, + 354, + 555, + 366 + ], + "type": "interline_equation", + "content": "\\varphi_ {\\mathrm {L L M}} (a, o) = \\mathrm {L L M} \\left(\\pi_ {c}, o, a\\right), \\tag {6}", + "image_path": "8c2a9485dce6b20c678020ab9764540fe22c1854f81877780299e051741c7e3e.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 375, + 554, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 375, + 554, + 411 + ], + "spans": [ + { + "bbox": [ + 313, + 375, + 554, + 411 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 375, + 554, + 411 + ], + "type": "inline_equation", + "content": "\\pi_c" + }, + { + "bbox": [ + 313, + 375, + 554, + 411 + ], + "type": "text", + "content": " is a prompt specifying the compatibility task (see the Supp. Mat.). The objective is to assign a score on a specific range to the likelihood of the attribute-object pair." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 411, + 554, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 411, + 554, + 447 + ], + "spans": [ + { + "bbox": [ + 313, + 411, + 554, + 447 + ], + "type": "text", + "content": "Taking into account that also LLMs have issues, such as hallucinations and biases [14, 33], we implement " + }, + { + "bbox": [ + 313, + 411, + 554, + 447 + ], + "type": "inline_equation", + "content": "\\varphi" + }, + { + "bbox": [ + 313, + 411, + 554, + 447 + ], + "type": "text", + "content": " by combining database- and LLM-based scoring as:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 367, + 456, + 555, + 469 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 456, + 555, + 469 + ], + "spans": [ + { + "bbox": [ + 367, + 456, + 555, + 469 + ], + "type": "interline_equation", + "content": "\\varphi (a, o) = \\varphi_ {\\mathrm {d b}} (a, o) \\cdot \\varphi_ {\\mathrm {L L M}} (a, o), \\tag {7}", + "image_path": "cc79a179e8653c98c9bf074ffbe9adf1f659b7f82da4d18f37f2b3421201a01f.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 478, + 542, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 478, + 542, + 491 + ], + "spans": [ + { + "bbox": [ + 313, + 478, + 542, + 491 + ], + "type": "text", + "content": "thus computing " + }, + { + "bbox": [ + 313, + 478, + 542, + 491 + ], + "type": "inline_equation", + "content": "\\Phi_O(a)" + }, + { + "bbox": [ + 313, + 478, + 542, + 491 + ], + "type": "text", + "content": " and filling the cache accordingly." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 497, + 487, + 510 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 497, + 487, + 510 + ], + "spans": [ + { + "bbox": [ + 313, + 497, + 487, + 510 + ], + "type": "text", + "content": "3.3. Compositional cache refinement" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 514, + 555, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 514, + 555, + 634 + ], + "spans": [ + { + "bbox": [ + 313, + 514, + 555, + 634 + ], + "type": "text", + "content": "After the process in Sec. 3.2, we have a cache " + }, + { + "bbox": [ + 313, + 514, + 555, + 634 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_{\\mathbb{H}}" + }, + { + "bbox": [ + 313, + 514, + 555, + 634 + ], + "type": "text", + "content": " populated with " + }, + { + "bbox": [ + 313, + 514, + 555, + 634 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 514, + 555, + 634 + ], + "type": "text", + "content": " examples for each attribute " + }, + { + "bbox": [ + 313, + 514, + 555, + 634 + ], + "type": "inline_equation", + "content": "a \\in A" + }, + { + "bbox": [ + 313, + 514, + 555, + 634 + ], + "type": "text", + "content": ". However, the cache assigns one hard attribute to each element, contrasting with the natural co-occurrence of attributes and leading to potentially wrong associations during inference. As a simple example, an image of a large and blue car obtained with the binding (large, car) will have a positive label only for large. As a consequence, if the test image shows a small blue car, the cache element will not contribute to the score of blue and may erroneously increase the score for the antonym large." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 635, + 555, + 681 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 635, + 555, + 681 + ], + "spans": [ + { + "bbox": [ + 313, + 635, + 555, + 681 + ], + "type": "text", + "content": "Soft caching. To remedy to this issue, we refine the label associated to each cache element, exploiting the VLM capabilities to compute soft labels. Formally, the soft label " + }, + { + "bbox": [ + 313, + 635, + 555, + 681 + ], + "type": "inline_equation", + "content": "s_a" + }, + { + "bbox": [ + 313, + 635, + 555, + 681 + ], + "type": "text", + "content": " for an attribute " + }, + { + "bbox": [ + 313, + 635, + 555, + 681 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 313, + 635, + 555, + 681 + ], + "type": "text", + "content": " and a cache image " + }, + { + "bbox": [ + 313, + 635, + 555, + 681 + ], + "type": "inline_equation", + "content": "x_c" + }, + { + "bbox": [ + 313, + 635, + 555, + 681 + ], + "type": "text", + "content": " is defined as:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 370, + 689, + 555, + 717 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 689, + 555, + 717 + ], + "spans": [ + { + "bbox": [ + 370, + 689, + 555, + 717 + ], + "type": "interline_equation", + "content": "s _ {a} \\left(x _ {c}\\right) = \\frac {f _ {t} (a) ^ {\\intercal} f _ {v} \\left(x _ {c}\\right)}{\\left| \\left| f _ {t} (a) \\right| \\right| \\cdot \\left| \\left| f _ {v} \\left(x _ {c}\\right) \\right| \\right|}, \\tag {8}", + "image_path": "5b70a6c19740efdeb4ddcd5c3d1a98aed0f231550f74c1b10a6f96d163e99c27.jpg" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15101" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 144 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 144 + ], + "type": "text", + "content": "where we compare the text embeddings of the attribute with the visual embeddings of the cache image. While these scores can directly serve as soft labels, we found them to exhibit low-variance due to the well-known modality-gap [30, 48]. We overcome this limitation and increase their relevance by defining normalized soft labels " + }, + { + "bbox": [ + 55, + 72, + 296, + 144 + ], + "type": "inline_equation", + "content": "\\bar{s}_a" + }, + { + "bbox": [ + 55, + 72, + 296, + 144 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 87, + 152, + 296, + 180 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 87, + 152, + 296, + 180 + ], + "spans": [ + { + "bbox": [ + 87, + 152, + 296, + 180 + ], + "type": "interline_equation", + "content": "\\bar {s} _ {a} \\left(x _ {c}\\right) = \\frac {\\exp \\left(\\left(s _ {a} \\left(x _ {c}\\right) - \\mu_ {C}\\right) / \\sigma_ {C}\\right)}{\\sum_ {\\hat {a} \\in A} \\exp \\left(\\left(s _ {\\hat {a}} (x) - \\mu_ {C}\\right) / \\sigma_ {C}\\right)}, \\tag {9}", + "image_path": "fdb5005c6261b37046d86692ff75f1ad9d62dcebfda7ce6efa06de57c2ea6950.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 188, + 296, + 296 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 188, + 296, + 296 + ], + "spans": [ + { + "bbox": [ + 55, + 188, + 296, + 296 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 188, + 296, + 296 + ], + "type": "inline_equation", + "content": "\\mu_{C}" + }, + { + "bbox": [ + 55, + 188, + 296, + 296 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 188, + 296, + 296 + ], + "type": "inline_equation", + "content": "\\sigma_{C}" + }, + { + "bbox": [ + 55, + 188, + 296, + 296 + ], + "type": "text", + "content": " are respectively the mean and standard deviation of the similarity between any attribute with any element of the cache. The first normalization expands the range of values of the similarity scores " + }, + { + "bbox": [ + 55, + 188, + 296, + 296 + ], + "type": "inline_equation", + "content": "s_a" + }, + { + "bbox": [ + 55, + 188, + 296, + 296 + ], + "type": "text", + "content": ", while the softmax ensures that all samples have an equal contribution to the final prediction. In practice, we found the latter step to be important only on datasets with many attributes (e.g., VAW). We hypothesize this is due to differences in the output space cardinality, requiring different softmax scaling as in [58]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 297, + 296, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 297, + 296, + 344 + ], + "spans": [ + { + "bbox": [ + 55, + 297, + 296, + 344 + ], + "type": "text", + "content": "These scores are mixed with the original hard labels to obtain the final label " + }, + { + "bbox": [ + 55, + 297, + 296, + 344 + ], + "type": "inline_equation", + "content": "\\hat{s}_a(x_c) = (1 - \\alpha)\\cdot \\mathbb{1}(a = c) + \\alpha \\cdot \\overline{s}_a(x_c)" + }, + { + "bbox": [ + 55, + 297, + 296, + 344 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 55, + 297, + 296, + 344 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 297, + 296, + 344 + ], + "type": "text", + "content": " is a trade-off hyperparameter. With the soft labels " + }, + { + "bbox": [ + 55, + 297, + 296, + 344 + ], + "type": "inline_equation", + "content": "\\hat{s}_a" + }, + { + "bbox": [ + 55, + 297, + 296, + 344 + ], + "type": "text", + "content": ", we can perform the cache-based prediction as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 354, + 296, + 396 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 354, + 296, + 396 + ], + "spans": [ + { + "bbox": [ + 62, + 354, + 296, + 396 + ], + "type": "interline_equation", + "content": "f _ {\\mathcal {C}} ^ {\\mathrm {C O M C A}} (x, a) = \\sum_ {x _ {c}} ^ {\\mathcal {C}} \\eta_ {\\mathcal {C}} \\left(\\hat {s} _ {a} \\left(x _ {c}\\right) \\frac {f _ {v} (x) ^ {\\intercal} f _ {v} \\left(x _ {c}\\right)}{\\left\\| f _ {v} (x) \\right\\| \\cdot \\left\\| f _ {v} \\left(x _ {c}\\right) \\right\\|}\\right). \\tag {10}", + "image_path": "a11798416d66397028752ae5ebb89568a90772335ee54da350fad9d9cf9c9682.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 397, + 296, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 397, + 296, + 445 + ], + "spans": [ + { + "bbox": [ + 55, + 397, + 296, + 445 + ], + "type": "text", + "content": "In contrast to Eq. (2), here we consider each cache element " + }, + { + "bbox": [ + 55, + 397, + 296, + 445 + ], + "type": "inline_equation", + "content": "x_{c}" + }, + { + "bbox": [ + 55, + 397, + 296, + 445 + ], + "type": "text", + "content": " as a potential contributor for attribute " + }, + { + "bbox": [ + 55, + 397, + 296, + 445 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 55, + 397, + 296, + 445 + ], + "type": "text", + "content": ", even if " + }, + { + "bbox": [ + 55, + 397, + 296, + 445 + ], + "type": "inline_equation", + "content": "a \\neq c" + }, + { + "bbox": [ + 55, + 397, + 296, + 445 + ], + "type": "text", + "content": ". Implicitly, this allows us to include an arbitrary number of (positive) attribute labels within each element of the cache." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 445, + 296, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 445, + 296, + 480 + ], + "spans": [ + { + "bbox": [ + 55, + 445, + 296, + 480 + ], + "type": "text", + "content": "Final inference. Finally, the cache-based scoring defined in Eq. (10) is merged with the standard CLIP-based scores to obtain the final predictions:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 69, + 490, + 295, + 506 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 490, + 295, + 506 + ], + "spans": [ + { + "bbox": [ + 69, + 490, + 295, + 506 + ], + "type": "interline_equation", + "content": "f (x, a) = \\eta_ {A} \\left(\\lambda \\cdot f _ {\\mathcal {C}} ^ {\\mathrm {C O M C A}} (x, a) + f _ {\\mathcal {C}} ^ {\\mathrm {C L I P}} (x, a)\\right), \\tag {11}", + "image_path": "ba04bbd04e8b4510a46212b3e264b62c3fa0a7da19e58a9a76e744b593a25dd3.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 514, + 296, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 514, + 296, + 609 + ], + "spans": [ + { + "bbox": [ + 55, + 514, + 296, + 609 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 514, + 296, + 609 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 514, + 296, + 609 + ], + "type": "text", + "content": " is an hyperparameter controlling the balance between the two and " + }, + { + "bbox": [ + 55, + 514, + 296, + 609 + ], + "type": "inline_equation", + "content": "\\eta_A(z) = \\mathrm{softmax}_A(z / \\max (z))" + }, + { + "bbox": [ + 55, + 514, + 296, + 609 + ], + "type": "text", + "content": ", moving from simple scalar values to a distribution over the attributes space and leading to more stable scores. Note that Eq. (11) contains no element which needs training or labeled data for the specific attributes of interest. Given the target sets of attributes " + }, + { + "bbox": [ + 55, + 514, + 296, + 609 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 55, + 514, + 296, + 609 + ], + "type": "text", + "content": " and objects " + }, + { + "bbox": [ + 55, + 514, + 296, + 609 + ], + "type": "inline_equation", + "content": "O" + }, + { + "bbox": [ + 55, + 514, + 296, + 609 + ], + "type": "text", + "content": ", COMCA automatically re-generates the cache to update the inference procedure." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 621, + 135, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 621, + 135, + 635 + ], + "spans": [ + { + "bbox": [ + 55, + 621, + 135, + 635 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 641, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 296, + 713 + ], + "type": "text", + "content": "Datasets. We evaluate COMCA on two datasets. OVAD [5] covers 80 classes and 117 attributes across 2,000 images from MS-COCO [38], for a total of 14,300 instances. Attributes are manually annotated as positive, negative, and unknown. VAW [39] has more concepts, i.e., 2,260 classes and 620 attributes across 10,392 images. VAW is more" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "content": "sparse than OVAD, i.e., around 1.8 positive and 5.3 negative attributes per instance, and it is automatically annotated." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 97, + 555, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 97, + 555, + 168 + ], + "spans": [ + { + "bbox": [ + 313, + 97, + 555, + 168 + ], + "type": "text", + "content": "Setting. Following prior works [5, 7], COMCA focuses on the box-given setting, where the bounding boxes of objects are given, as this allows us to assess its attribute detection capabilities more precisely. However, Tab. 3 also evaluates COMCA in the box-free setting by introducing an object detector. See the Supp. Mat. for more details on this setting." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 169, + 555, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 169, + 555, + 205 + ], + "spans": [ + { + "bbox": [ + 313, + 169, + 555, + 205 + ], + "type": "text", + "content": "Metrics. We measure performance as mean Average Precision (mAP), and also report mAP scores on head, medium, and tail classes based on their frequency in the dataset." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 206, + 556, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 206, + 556, + 445 + ], + "spans": [ + { + "bbox": [ + 313, + 206, + 556, + 445 + ], + "type": "text", + "content": "Baselines and competitors. We evaluate COMCA against training-based methods, i.e., OVAD [5], OvarNet [7], ArtVLM [63], and LOWA [16]. As we are the first to address the problem in a completely training-free fashion, we extend the baselines with several zero-shot methods [9, 28, 54-56], and adapt previous cache-based approaches to our task [48, 57]. For TIP-Adapter [57] and SuS-X [48] we construct a " + }, + { + "bbox": [ + 313, + 206, + 556, + 445 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 206, + 556, + 445 + ], + "type": "text", + "content": "-shot cache, considering only attributes for sampling and disregarding their co-occurrences with objects. This follows their approach, which samples " + }, + { + "bbox": [ + 313, + 206, + 556, + 445 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 206, + 556, + 445 + ], + "type": "text", + "content": " shots for each target class, i.e., attributes in our task. In addition, we test IAP [26, 27], which conditions predicted attributes on the inferred object category, using TIP-Adapter as the object classifier. We also consider a naive cache-based approach, i.e., \"image-based\", that uses CC12M [6] to build a cache specifically for each input image by retrieving the most similar samples. Since the retrieved images are unlabeled, we use our soft labels to annotate the cache. We construct all the proposed baselines on top of CLIP ViT-B/32 [9]. For additional details, please refer to the Supp. Mat." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 445, + 556, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 445, + 556, + 578 + ], + "spans": [ + { + "bbox": [ + 313, + 445, + 556, + 578 + ], + "type": "text", + "content": "Implementation details. We use CLIP ViT-B/32 [9] pretrained on LAION2B [45] as our primary VLM. Following TIP-Adapter [57], we set " + }, + { + "bbox": [ + 313, + 445, + 556, + 578 + ], + "type": "inline_equation", + "content": "\\lambda = 1.17" + }, + { + "bbox": [ + 313, + 445, + 556, + 578 + ], + "type": "text", + "content": " in Eq. (11) and " + }, + { + "bbox": [ + 313, + 445, + 556, + 578 + ], + "type": "inline_equation", + "content": "\\beta = 1.0" + }, + { + "bbox": [ + 313, + 445, + 556, + 578 + ], + "type": "text", + "content": " in Eq. (2). We set " + }, + { + "bbox": [ + 313, + 445, + 556, + 578 + ], + "type": "inline_equation", + "content": "\\alpha = 0.6" + }, + { + "bbox": [ + 313, + 445, + 556, + 578 + ], + "type": "text", + "content": " to mix hard and soft labels on the validation set of VAW [39]. Lastly, we use CC12M [6] as the database in Eq. (5) and GPT 3.5 Turbo [35] as the LLM for Eq. (6), as we find it to perform better than GPT 4o-mini (see the Supp. Mat.). Unless otherwise stated, we use " + }, + { + "bbox": [ + 313, + 445, + 556, + 578 + ], + "type": "inline_equation", + "content": "K = 16" + }, + { + "bbox": [ + 313, + 445, + 556, + 578 + ], + "type": "text", + "content": " shots per attribute, retrieving them from CC12M. We use the same hyperparameters for both datasets. For the box-free experiments, we use YOLOv11M [23] as the detector." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 314, + 586, + 372, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 586, + 372, + 597 + ], + "spans": [ + { + "bbox": [ + 314, + 586, + 372, + 597 + ], + "type": "text", + "content": "4.1. Results" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 605, + 555, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 555, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 555, + 665 + ], + "type": "text", + "content": "We present our comprehensive experimental evaluation comparing COMCA with state-of-the-art training- and cache-based methods. We also show its performance with different backbones and in the box-free setting. Moreover, we discuss qualitative results and limitations of other approaches." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 666, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 666, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 666, + 556, + 713 + ], + "type": "text", + "content": "Comparison with the state of the art. Tab. 1 reports the results comparing zero-shot baselines, cache-based methods, and COMCA in the box-given setting. On both benchmarks, COMCA outperforms all the zero-shot and cache-based com" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15102" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 70, + 560, + 266 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 560, + 266 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 560, + 266 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 560, + 266 + ], + "type": "table", + "html": "
MethodTrain sizeOVAD [5]VAW [39]
mAPHeadMediumTailmAPHeadMediumTail
Zero shotCLIP [9] RN5011.841.011.71.435.337.835.126.2
CLIP [9] ViT-B/3217.044.318.45.550.051.050.943.2
CLIP [9] ViT-L/1418.344.420.56.451.051.452.545.0
Cache-basedImage-based16.845.819.23.552.953.753.846.4
TIP-Adapter [57] + IAP [27]15.143.317.81.727.732.425.516.6
TIP-Adapter [57]16.744.419.73.157.557.559.451.2
SuS-X [48]20.248.924.64.530.233.828.820.3
ComCA27.454.334.69.058.158.259.951.7
Training basedOVAD [5]110k21.448.026.95.2----
OvarNet [7]190k28.658.635.59.568.5---
ArtVLM [63]N/A----71.975.072.159.4
LOWA [16]1.33M18.758.020.42.642.646.441.032.9
", + "image_path": "8dee48bfbbe5b537f46c8c9eac02fa201ed9d51e683ab7ca8f722dcd58f0fb61.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 274, + 555, + 308 + ], + "lines": [ + { + "bbox": [ + 55, + 274, + 555, + 308 + ], + "spans": [ + { + "bbox": [ + 55, + 274, + 555, + 308 + ], + "type": "text", + "content": "Table 1. Comparison with state of the art. Green indicates COMCA. Bold indicates best among training-free methods. The symbol “-” indicates results for the competitors are not available on the original papers or the impossibility to run their method due to lack of public code and/or model weights. See the Supp. Mat. for an extended version of the table." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 314, + 294, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 314, + 294, + 445 + ], + "spans": [ + { + "bbox": [ + 54, + 314, + 294, + 445 + ], + "type": "text", + "content": "petitors, demonstrating the effectiveness of our specialized design of the cache. It scores an increment of " + }, + { + "bbox": [ + 54, + 314, + 294, + 445 + ], + "type": "inline_equation", + "content": "+10.6\\mathrm{mAP}" + }, + { + "bbox": [ + 54, + 314, + 294, + 445 + ], + "type": "text", + "content": " on OVAD and " + }, + { + "bbox": [ + 54, + 314, + 294, + 445 + ], + "type": "inline_equation", + "content": "+5.2" + }, + { + "bbox": [ + 54, + 314, + 294, + 445 + ], + "type": "text", + "content": " on VAW compared to the image-based baseline. Similarly, it surpasses TIP-Adapter [57] on both benchmarks by " + }, + { + "bbox": [ + 54, + 314, + 294, + 445 + ], + "type": "inline_equation", + "content": "+10.7\\mathrm{mAP}" + }, + { + "bbox": [ + 54, + 314, + 294, + 445 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 314, + 294, + 445 + ], + "type": "inline_equation", + "content": "+0.6" + }, + { + "bbox": [ + 54, + 314, + 294, + 445 + ], + "type": "text", + "content": ", respectively. SuS-X [48] increases performance on OVAD while decreasing it on VAW. Overall, COMCA has a gain of " + }, + { + "bbox": [ + 54, + 314, + 294, + 445 + ], + "type": "inline_equation", + "content": "+7.2\\mathrm{mAP}" + }, + { + "bbox": [ + 54, + 314, + 294, + 445 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 314, + 294, + 445 + ], + "type": "inline_equation", + "content": "+27.9" + }, + { + "bbox": [ + 54, + 314, + 294, + 445 + ], + "type": "text", + "content": " on the two datasets. Lastly, TIP-Adapter + IAP [27] demonstrates that naively using co-occurrences of attributes with objects [27] is insufficient to address this task. COMCA surpasses it by " + }, + { + "bbox": [ + 54, + 314, + 294, + 445 + ], + "type": "inline_equation", + "content": "+12.3\\mathrm{mAP}" + }, + { + "bbox": [ + 54, + 314, + 294, + 445 + ], + "type": "text", + "content": " on OVAD and " + }, + { + "bbox": [ + 54, + 314, + 294, + 445 + ], + "type": "inline_equation", + "content": "+30.4" + }, + { + "bbox": [ + 54, + 314, + 294, + 445 + ], + "type": "text", + "content": " on VAW." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 447, + 295, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 447, + 295, + 495 + ], + "spans": [ + { + "bbox": [ + 55, + 447, + 295, + 495 + ], + "type": "text", + "content": "We also compare COMCA with training-based methods. Notably, COMCA outperforms LOWA [16] on both benchmarks, despite the latter having been trained on a large-scale dataset while our approach requires no training data." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 496, + 295, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 496, + 295, + 604 + ], + "spans": [ + { + "bbox": [ + 54, + 496, + 295, + 604 + ], + "type": "text", + "content": "Experiments with multiple VLMs. To demonstrate that COMCA is model-agnostic, we test it on various backbones and VLMs [9, 28, 54-56] and report the results in Tab. 2. On both OVAD and VAW, COMCA boosts performance consistently for all the backbones up to " + }, + { + "bbox": [ + 54, + 496, + 295, + 604 + ], + "type": "inline_equation", + "content": "+11.5\\mathrm{mAP}" + }, + { + "bbox": [ + 54, + 496, + 295, + 604 + ], + "type": "text", + "content": " on OVAD with CoCa ViT-L/14 [54] and " + }, + { + "bbox": [ + 54, + 496, + 295, + 604 + ], + "type": "inline_equation", + "content": "+15.8\\mathrm{mAP}" + }, + { + "bbox": [ + 54, + 496, + 295, + 604 + ], + "type": "text", + "content": " on VAW with CLIP ResNet50 (RN50) [9]. On average, COMCA improves the performance of the original VLM by " + }, + { + "bbox": [ + 54, + 496, + 295, + 604 + ], + "type": "inline_equation", + "content": "+8.9" + }, + { + "bbox": [ + 54, + 496, + 295, + 604 + ], + "type": "text", + "content": " on OVAD and " + }, + { + "bbox": [ + 54, + 496, + 295, + 604 + ], + "type": "inline_equation", + "content": "+7.0" + }, + { + "bbox": [ + 54, + 496, + 295, + 604 + ], + "type": "text", + "content": " on VAW, showing COMCA has consistent advantages." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 605, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 605, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 605, + 295, + 714 + ], + "type": "text", + "content": "Cross-datasets advantages. To demonstrate the generalization capabilities of our method, we report cross-dataset results in Fig. 4. We compare COMCA with OVAD when trained on various datasets (i.e., VAW [39], OVAD [5], MSCOCO [38]) and tested on the OVAD (left) and on the VAW (right) benchmarks. As shown in Fig. 4, performance significantly degrades when OVAD is tested on datasets different from the training one (yellow bars vs red bar), whereas our training-free COMCA achieves competitive performance" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 314, + 553, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 314, + 553, + 350 + ], + "spans": [ + { + "bbox": [ + 313, + 314, + 553, + 350 + ], + "type": "text", + "content": "(green bars) with the same RN50 backbone. This reinforces our training-free method, which does not inherit biases from a training set, differently from training-based approaches." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 352, + 554, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 352, + 554, + 544 + ], + "spans": [ + { + "bbox": [ + 313, + 352, + 554, + 544 + ], + "type": "text", + "content": "Box-free setting. In this work, we primarily focus on attribute recognition and therefore adopt the box-given setting. Here, we evaluate COMCA in the box-free setting to demonstrate that it achieves competitive performance in a less constrained scenario. Tab. 3 shows the results for zero-shot and cache-based methods, Specifically, we use CLIP ViT-B/32 [9] and YOLOv11M [23] as detector for all models. We note that all cache-based methods apart from SuS-X underperform the zero-shot baseline, with a decrease of up to " + }, + { + "bbox": [ + 313, + 352, + 554, + 544 + ], + "type": "inline_equation", + "content": "-3.1\\mathrm{mAP}" + }, + { + "bbox": [ + 313, + 352, + 554, + 544 + ], + "type": "text", + "content": " for image-based. Only, SuS-X has a gain of " + }, + { + "bbox": [ + 313, + 352, + 554, + 544 + ], + "type": "inline_equation", + "content": "+2.8\\mathrm{mAP}" + }, + { + "bbox": [ + 313, + 352, + 554, + 544 + ], + "type": "text", + "content": " w.r.t. CLIP ViT-B/32, with our method COMCA being superior and surpassing CLIP ViT-B/32 by " + }, + { + "bbox": [ + 313, + 352, + 554, + 544 + ], + "type": "inline_equation", + "content": "+9.0\\mathrm{mAP}" + }, + { + "bbox": [ + 313, + 352, + 554, + 544 + ], + "type": "text", + "content": ". Notably, our model is competitive with the training-based ones, surpassing the performance of both OVAD " + }, + { + "bbox": [ + 313, + 352, + 554, + 544 + ], + "type": "inline_equation", + "content": "(+2.7\\mathrm{mAP})" + }, + { + "bbox": [ + 313, + 352, + 554, + 544 + ], + "type": "text", + "content": " and LOWA " + }, + { + "bbox": [ + 313, + 352, + 554, + 544 + ], + "type": "inline_equation", + "content": "(+2.8\\mathrm{mAP})" + }, + { + "bbox": [ + 313, + 352, + 554, + 544 + ], + "type": "text", + "content": ". We provide a more detailed analysis and an extended version of these results in the Supp. Mat." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 545, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 556, + 713 + ], + "type": "text", + "content": "Qualitative results. Fig. 3, we show qualitative examples comparing training-based OVAD [5], CLIP ViT-B/32 [9], and COMCA with the same backbone. We consider the top-1 prediction for all methods. OVAD correctly identifies the bear's color but misclassifies the cat's color and wrongly predicts a \"tiled\" clothes pattern for two people. We speculate it likely predicts the same attribute for two objects (i.e., the " + }, + { + "bbox": [ + 313, + 545, + 556, + 713 + ], + "type": "inline_equation", + "content": "2^{nd}" + }, + { + "bbox": [ + 313, + 545, + 556, + 713 + ], + "type": "text", + "content": " and the " + }, + { + "bbox": [ + 313, + 545, + 556, + 713 + ], + "type": "inline_equation", + "content": "6^{th}" + }, + { + "bbox": [ + 313, + 545, + 556, + 713 + ], + "type": "text", + "content": ") due to biases acquired during training. CLIP makes more errors, such as labeling the zebra as \"reflective\" and the airplane as \"tiny\", while misclassifying the baseball player as \"giant\". COMCA instead understands attributes better, for instance by accurately identifying the zebra's texture. These issues are less pronounced in COMCA thanks to our caching mechanism, which allows it to compare inputs" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15103" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 70, + 294, + 251 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 294, + 251 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 294, + 251 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 294, + 251 + ], + "type": "table", + "html": "
MethodOVAD [5]VAW [39]
BackboneCOMCAmAPΔmAPΔ
CLIP RN50 [9]11.835.3
✓19.2+7.451.1+15.8
CLIP ViT-B/32 [9]17.050.0
✓27.4+10.458.1+8.1
CLIP ViT-L/14 [9]18.351.0
✓24.8+6.561.0+10.0
SigLIP ViT-B/16 [56]13.752.5
✓22.1+8.459.8+7.3
CoCa ViT-B/32 [54]15.149.6
✓24.4+9.351.4+1.8
CoCa ViT-L/14 [54]14.449.8
✓25.9+11.554.5+4.7
BLIP [28]15.951.0
✓22.7+6.856.0+5.0
X-VLM [55]17.454.8
✓28.4+11.057.7+2.9
", + "image_path": "fb350deee4ba76872502b745c25bb8fe5adf41460e93ed54b798acc6d4c9cdb9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 313, + 70, + 550, + 238 + ], + "blocks": [ + { + "bbox": [ + 55, + 259, + 296, + 304 + ], + "lines": [ + { + "bbox": [ + 55, + 259, + 296, + 304 + ], + "spans": [ + { + "bbox": [ + 55, + 259, + 296, + 304 + ], + "type": "text", + "content": "Table 2. Box-given results with different backbones. Green indicates COMCA applied on top of the backbone. " + }, + { + "bbox": [ + 55, + 259, + 296, + 304 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 55, + 259, + 296, + 304 + ], + "type": "text", + "content": " indicates improvements w.r.t. the corresponding baseline. See the Supp. Mat. for an extended version of the table." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 313, + 70, + 550, + 238 + ], + "lines": [ + { + "bbox": [ + 313, + 70, + 550, + 238 + ], + "spans": [ + { + "bbox": [ + 313, + 70, + 550, + 238 + ], + "type": "table", + "html": "
MethodOVAD [5]
mAPHeadMediumTail
Zero shotCLIP [9] RN5012.340.212.71.9
CLIP [9] ViT-B/3212.539.612.52.7
CLIP [9] ViT-L/1413.439.714.32.9
Cache-basedImage-based9.435.48.71.0
TIP [57] + IAP [27]9.836.99.21.0
TIP-Adapter [57]10.136.09.81.2
SuS-X [48]15.343.017.23.2
COMCA21.548.426.46.0
Training basedOVAD [5]18.847.722.04.6
OvarNet [7]27.256.833.68.9
LOWA [16]18.758.020.42.6
", + "image_path": "37bdbd4f5a8456e96a664d9fbfe026e7f62fafb3aad4f8ab7405fb03964dffe7.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 311, + 245, + 553, + 280 + ], + "lines": [ + { + "bbox": [ + 311, + 245, + 553, + 280 + ], + "spans": [ + { + "bbox": [ + 311, + 245, + 553, + 280 + ], + "type": "text", + "content": "Table 3. Results in the box-free setting. Green indicates COMCA. Bold indicates the best results among training-free methods. See the Supp. Mat. for an extended version of the table." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 83, + 308, + 151, + 376 + ], + "blocks": [ + { + "bbox": [ + 83, + 308, + 151, + 376 + ], + "lines": [ + { + "bbox": [ + 83, + 308, + 151, + 376 + ], + "spans": [ + { + "bbox": [ + 83, + 308, + 151, + 376 + ], + "type": "image", + "image_path": "dca6ab54cfb7c11f58f092b69f981828ce8d75a4ff1d587ea69dbb145777b171.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 59, + 379, + 82, + 408 + ], + "lines": [ + { + "bbox": [ + 59, + 379, + 82, + 408 + ], + "spans": [ + { + "bbox": [ + 59, + 379, + 82, + 408 + ], + "type": "text", + "content": "OVAD CLIP COMCA" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 85, + 378, + 148, + 402 + ], + "lines": [ + { + "bbox": [ + 85, + 378, + 148, + 402 + ], + "spans": [ + { + "bbox": [ + 85, + 378, + 148, + 402 + ], + "type": "text", + "content": "brown brown single, one, individual" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 151, + 309, + 218, + 376 + ], + "blocks": [ + { + "bbox": [ + 151, + 309, + 218, + 376 + ], + "lines": [ + { + "bbox": [ + 151, + 309, + 218, + 376 + ], + "spans": [ + { + "bbox": [ + 151, + 309, + 218, + 376 + ], + "type": "image", + "image_path": "7ce6350624792f5749ac16cee0b5d90dc83730a82509b80083a45b28ef00e985.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 155, + 379, + 215, + 402 + ], + "lines": [ + { + "bbox": [ + 155, + 379, + 215, + 402 + ], + "spans": [ + { + "bbox": [ + 155, + 379, + 215, + 402 + ], + "type": "text", + "content": "clothes pattern: tiled male neutral, calm, serious" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 219, + 309, + 287, + 376 + ], + "blocks": [ + { + "bbox": [ + 219, + 309, + 287, + 376 + ], + "lines": [ + { + "bbox": [ + 219, + 309, + 287, + 376 + ], + "spans": [ + { + "bbox": [ + 219, + 309, + 287, + 376 + ], + "type": "image", + "image_path": "745fc7b5e48b84ea10d42653376b60c5c44208cd6b19e588c8fa0e21d346ae98.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 220, + 379, + 286, + 402 + ], + "lines": [ + { + "bbox": [ + 220, + 379, + 286, + 402 + ], + "spans": [ + { + "bbox": [ + 220, + 379, + 286, + 402 + ], + "type": "text", + "content": "plaid, tartan, checkered reflective striped, lined, pinstriped" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 288, + 308, + 416, + 376 + ], + "blocks": [ + { + "bbox": [ + 288, + 308, + 416, + 376 + ], + "lines": [ + { + "bbox": [ + 288, + 308, + 416, + 376 + ], + "spans": [ + { + "bbox": [ + 288, + 308, + 416, + 376 + ], + "type": "image", + "image_path": "93e11421a085444e7676f52c6ae85268e6e4cd6181f7a437a51fd4735c0f961d.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 328, + 379, + 375, + 401 + ], + "lines": [ + { + "bbox": [ + 328, + 379, + 375, + 401 + ], + "spans": [ + { + "bbox": [ + 328, + 379, + 375, + 401 + ], + "type": "text", + "content": "large, giant, huge small, little, tiny smooth, sleek" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 417, + 309, + 485, + 376 + ], + "blocks": [ + { + "bbox": [ + 417, + 309, + 485, + 376 + ], + "lines": [ + { + "bbox": [ + 417, + 309, + 485, + 376 + ], + "spans": [ + { + "bbox": [ + 417, + 309, + 485, + 376 + ], + "type": "image", + "image_path": "1d7253325f036b30d81cf2d1cf6c400fa60d44663b32e2d82159fd575111c000.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 428, + 379, + 473, + 402 + ], + "lines": [ + { + "bbox": [ + 428, + 379, + 473, + 402 + ], + "spans": [ + { + "bbox": [ + 428, + 379, + 473, + 402 + ], + "type": "text", + "content": "yellow piece, cut small, little, tiny" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 486, + 309, + 553, + 376 + ], + "blocks": [ + { + "bbox": [ + 486, + 309, + 553, + 376 + ], + "lines": [ + { + "bbox": [ + 486, + 309, + 553, + 376 + ], + "spans": [ + { + "bbox": [ + 486, + 309, + 553, + 376 + ], + "type": "image", + "image_path": "c730c7329b0e3740b1aea6f8fc53ad1d4d8bc8f9de7b469c4355ea80974ed769.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 490, + 379, + 549, + 402 + ], + "lines": [ + { + "bbox": [ + 490, + 379, + 549, + 402 + ], + "spans": [ + { + "bbox": [ + 490, + 379, + 549, + 402 + ], + "type": "text", + "content": "clothes pattern: tiled big, large, giant, huge sitting, sit" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 60, + 437, + 289, + 533 + ], + "blocks": [ + { + "bbox": [ + 55, + 412, + 555, + 424 + ], + "lines": [ + { + "bbox": [ + 55, + 412, + 555, + 424 + ], + "spans": [ + { + "bbox": [ + 55, + 412, + 555, + 424 + ], + "type": "text", + "content": "Figure 3. Qualitative results. Predictions of OVAD, CLIP and ComCA on some OVAD images. Green are correct ones, red are wrong." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 60, + 437, + 289, + 533 + ], + "lines": [ + { + "bbox": [ + 60, + 437, + 289, + 533 + ], + "spans": [ + { + "bbox": [ + 60, + 437, + 289, + 533 + ], + "type": "image", + "image_path": "6a23ffc3cb2bd9093dd8e46d13541458d3940461dac0717f96cd93cfa7ede7fc.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 542, + 297, + 597 + ], + "lines": [ + { + "bbox": [ + 55, + 542, + 297, + 597 + ], + "spans": [ + { + "bbox": [ + 55, + 542, + 297, + 597 + ], + "type": "text", + "content": "Figure 4. Cross-dataset results. Results on OVAD (left) and VAW (right). Yellow indicates methods tested cross-domain, i.e., train and test set differ. Red indicates methods tested in-domain, and green our training-free method, which does not suffer from domain shift problems. All experiments use a RN50 backbone." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "bbox": [ + 55, + 605, + 204, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 204, + 616 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 204, + 616 + ], + "type": "text", + "content": "to self-labeled examples at test time." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 55, + 624, + 147, + 636 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 624, + 147, + 636 + ], + "spans": [ + { + "bbox": [ + 55, + 624, + 147, + 636 + ], + "type": "text", + "content": "4.2. Ablation study" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "type": "text", + "content": "In this section, we analyze COMCA's components, studying how sampling strategies affect performance (Tab. 4), soft labeling's impact (Tab. 5 and Fig. 6), and the number of samples in the cache (Fig. 5). Additional experiments in the Supp. Mat. show the effect of (i) the hyperparameters " + }, + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "type": "text", + "content": ", (ii) constructing the cache with Stable Diffusion, and (iii)" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 437, + 476, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 476, + 449 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 476, + 449 + ], + "type": "text", + "content": "using other LLMs for generative scores." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 312, + 449, + 555, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 449, + 555, + 653 + ], + "spans": [ + { + "bbox": [ + 312, + 449, + 555, + 653 + ], + "type": "text", + "content": "Cache construction. In Tab. 4, we evaluate five strategies to construct the cache. The first, Random, samples " + }, + { + "bbox": [ + 312, + 449, + 555, + 653 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 312, + 449, + 555, + 653 + ], + "type": "text", + "content": " objects for each attribute and creates a cache of " + }, + { + "bbox": [ + 312, + 449, + 555, + 653 + ], + "type": "inline_equation", + "content": "|A| \\times K" + }, + { + "bbox": [ + 312, + 449, + 555, + 653 + ], + "type": "text", + "content": " examples, seeking only a match on the attribute. The second, Brute force, considers all possible combinations of attributes and objects, resulting in a large cache of " + }, + { + "bbox": [ + 312, + 449, + 555, + 653 + ], + "type": "inline_equation", + "content": "|O| \\times |A| \\times K" + }, + { + "bbox": [ + 312, + 449, + 555, + 653 + ], + "type": "text", + "content": " samples. As VAW contains many attributes and objects, the Brute force cache would be too large and cannot be practically tested. As shown in Tab. 4, both approaches lead to very low results on OVAD. Compared to the zero-shot CLIP baseline, performance is similar (Random, -0.3 mAP) or even below (Brute force, -6.4 mAP). We hypothesize that Random ignores attribute-object relationships, while Brute force introduces many noisy and potentially detrimental samples, harming performance. On the contrary, Random improves by +7.5 on VAW. We hypothesize this is due to the nature of the dataset, which contains many attributes and objects." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 654, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 556, + 713 + ], + "type": "text", + "content": "When introducing priors on object-attributes compositions via a database (Eq. (5)) or an LLM (Eq. (6)), performance largely increases by up to " + }, + { + "bbox": [ + 313, + 654, + 556, + 713 + ], + "type": "inline_equation", + "content": "+4.1" + }, + { + "bbox": [ + 313, + 654, + 556, + 713 + ], + "type": "text", + "content": " mAP on VAW w.r.t. the baseline for database scores. ComCA combines the two scores to obtain object priors, achieving the best perfor" + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15104" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 70, + 225, + 173 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 225, + 173 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 225, + 173 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 225, + 173 + ], + "type": "table", + "html": "
ConfigurationmAP
OVAD [5]VAW [39]
Zero-shot [9]17.050.0
Random16.757.5
Brute force10.6-
DB scores18.654.1
LLM scores21.155.4
DB + LLM scores26.455.2
", + "image_path": "867d7eab45d6861eb7c42f75e372412fc798682c1239b3103caba68701c815de.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 233, + 69, + 400, + 176 + ], + "blocks": [ + { + "bbox": [ + 55, + 179, + 226, + 234 + ], + "lines": [ + { + "bbox": [ + 55, + 179, + 226, + 234 + ], + "spans": [ + { + "bbox": [ + 55, + 179, + 226, + 234 + ], + "type": "text", + "content": "Table 4. Ablation on cache construction. Green indicates our default configuration. Bold indicates the best results. Results are without soft labels. See the Supp. Mat. for an extended version of the table." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 233, + 69, + 400, + 176 + ], + "lines": [ + { + "bbox": [ + 233, + 69, + 400, + 176 + ], + "spans": [ + { + "bbox": [ + 233, + 69, + 400, + 176 + ], + "type": "table", + "html": "
ConfigurationmAP
OVAD [5]VAW [39]
Zero-shot [9]17.050.0
One-hot26.455.2
Soft26.753.2
- Sharpening [2]21.527.7
- Softmax24.558.5
- Eq. (9)27.458.1
", + "image_path": "4ebc949e263ea7dd1c793f0e1ab87dcec9c28e9f14d1d14c006a9592589e9d9c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 230, + 183, + 402, + 227 + ], + "lines": [ + { + "bbox": [ + 230, + 183, + 402, + 227 + ], + "spans": [ + { + "bbox": [ + 230, + 183, + 402, + 227 + ], + "type": "text", + "content": "Table 5. Ablation on soft labels. Green indicates our default configuration. Bold indicates the best results. See the Supp. Mat. for an extended version of the table." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 412, + 76, + 547, + 170 + ], + "blocks": [ + { + "bbox": [ + 412, + 76, + 547, + 170 + ], + "lines": [ + { + "bbox": [ + 412, + 76, + 547, + 170 + ], + "spans": [ + { + "bbox": [ + 412, + 76, + 547, + 170 + ], + "type": "image", + "image_path": "3c696229124d5b3e2c71259862b246ecd9672b720a8d8f9fadda9fac9bc4ecce.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 406, + 183, + 555, + 239 + ], + "lines": [ + { + "bbox": [ + 406, + 183, + 555, + 239 + ], + "spans": [ + { + "bbox": [ + 406, + 183, + 555, + 239 + ], + "type": "text", + "content": "Figure 5. Ablation on number of samples per attribute. The purple line is mAP on OVAD, while the red line is mAP on VAW. Green indicates our default configuration." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 59, + 260, + 290, + 324 + ], + "blocks": [ + { + "bbox": [ + 59, + 260, + 290, + 324 + ], + "lines": [ + { + "bbox": [ + 59, + 260, + 290, + 324 + ], + "spans": [ + { + "bbox": [ + 59, + 260, + 290, + 324 + ], + "type": "image", + "image_path": "4997d68500826be54128ff262482948f7da2d213a30c33e0f5c98076a1e805dd.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 335, + 295, + 380 + ], + "lines": [ + { + "bbox": [ + 55, + 335, + 295, + 380 + ], + "spans": [ + { + "bbox": [ + 55, + 335, + 295, + 380 + ], + "type": "text", + "content": "Figure 6. Soft labels across cache priors. Performance on OVAD (left) and VAW (right), considering generative (LLM) and retrieval (CC3M, CC12M) scores. Yellow indicates one-hot labels, red soft labels and green COMCA, combining LLM and CC12M." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 396, + 295, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 396, + 295, + 432 + ], + "spans": [ + { + "bbox": [ + 55, + 396, + 295, + 432 + ], + "type": "text", + "content": "mance. These results confirm both (i) that attribute-object compatibility is important to build the cache, and (ii) that databases and LLMs provide complementary information." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 433, + 296, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 433, + 296, + 493 + ], + "spans": [ + { + "bbox": [ + 55, + 433, + 296, + 493 + ], + "type": "text", + "content": "Soft labels. As discussed in Sec. 3.3, each image in the cache contains multiple attributes, even if we fill the cache considering only one of them. Here we analyze the soft labels obtained via Eq. (8), comparing them with the One-hot attribute labels of the initial cache, reporting the results in Tab. 5." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 495, + 296, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 495, + 296, + 578 + ], + "spans": [ + { + "bbox": [ + 55, + 495, + 296, + 578 + ], + "type": "text", + "content": "The first strategy is to incorporate the soft scores as they are predicted (Soft in Tab. 5). This leads to results slightly above the One-hot baselines on OVAD " + }, + { + "bbox": [ + 55, + 495, + 296, + 578 + ], + "type": "inline_equation", + "content": "(+0.3\\mathrm{mAP})" + }, + { + "bbox": [ + 55, + 495, + 296, + 578 + ], + "type": "text", + "content": ", while reducing the performance in VAW " + }, + { + "bbox": [ + 55, + 495, + 296, + 578 + ], + "type": "inline_equation", + "content": "(-2.0\\mathrm{mAP})" + }, + { + "bbox": [ + 55, + 495, + 296, + 578 + ], + "type": "text", + "content": ". This is a direct consequence of allowing each sample to have a different level of relevance in the cache, something that becomes more prominent in VAW due to the larger number of attributes." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 580, + 295, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 580, + 295, + 688 + ], + "spans": [ + { + "bbox": [ + 55, + 580, + 295, + 688 + ], + "type": "text", + "content": "To fix this issue, we can normalize the scores. We study three strategies: Sharpening, where the scores are normalized as in PAWS [2]; Softmax normalization over each sample; and the one we propose in Eq. (9), i.e., using the statistics of the cache-attributes similarity. Sharpening is not effective in this scenario, achieving the lowest performance across both settings. Softmax normalization leads to good results on VAW, e.g., +3.3 mAP w.r.t. One-hot, but it struggles on OVAD (24.5 mAP) due to the limited range of values." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "Our strategy addresses both issues simultaneously, consistently improving metrics in both settings compared to" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 259, + 555, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 259, + 555, + 331 + ], + "spans": [ + { + "bbox": [ + 313, + 259, + 555, + 331 + ], + "type": "text", + "content": "the One-hot strategy, scoring " + }, + { + "bbox": [ + 313, + 259, + 555, + 331 + ], + "type": "inline_equation", + "content": "27.4\\mathrm{mAP}" + }, + { + "bbox": [ + 313, + 259, + 555, + 331 + ], + "type": "text", + "content": " on OVAD and 58.1 mAP on VAW. Note that these findings are general. Fig. 6 shows soft labels consistently enhance results for estimating attribute-object priors across all strategies and datasets. Improvements range from " + }, + { + "bbox": [ + 313, + 259, + 555, + 331 + ], + "type": "inline_equation", + "content": "+3.8\\mathrm{mAP}" + }, + { + "bbox": [ + 313, + 259, + 555, + 331 + ], + "type": "text", + "content": " on VAW to " + }, + { + "bbox": [ + 313, + 259, + 555, + 331 + ], + "type": "inline_equation", + "content": "+5.4\\mathrm{mAP}" + }, + { + "bbox": [ + 313, + 259, + 555, + 331 + ], + "type": "text", + "content": " on OVAD in the LLM-only setting." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 332, + 556, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 332, + 556, + 440 + ], + "spans": [ + { + "bbox": [ + 313, + 332, + 556, + 440 + ], + "type": "text", + "content": "Number of samples per attribute. In Fig. 5, we measure how adding more samples (varying " + }, + { + "bbox": [ + 313, + 332, + 556, + 440 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 332, + 556, + 440 + ], + "type": "text", + "content": ") per attribute influences performance on the two datasets. We observe performance consistently increasing for OVAD while going from 1 to 16 samples, reaching a plateau when moving to " + }, + { + "bbox": [ + 313, + 332, + 556, + 440 + ], + "type": "inline_equation", + "content": "K = 24" + }, + { + "bbox": [ + 313, + 332, + 556, + 440 + ], + "type": "text", + "content": ". On VAW, performance is stable for different values of " + }, + { + "bbox": [ + 313, + 332, + 556, + 440 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 332, + 556, + 440 + ], + "type": "text", + "content": " due to the large number of attributes in the dataset (620), making the cache representative enough even when " + }, + { + "bbox": [ + 313, + 332, + 556, + 440 + ], + "type": "inline_equation", + "content": "K = 1" + }, + { + "bbox": [ + 313, + 332, + 556, + 440 + ], + "type": "text", + "content": ". These results confirm the stability of COMCA." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 314, + 453, + 388, + 465 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 453, + 388, + 465 + ], + "spans": [ + { + "bbox": [ + 314, + 453, + 388, + 465 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 474, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 474, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 474, + 556, + 715 + ], + "type": "text", + "content": "We introduced Compositional Caching for open-vocabulary attribute detection. COMCA uses a VLM and an auxiliary cache of images to solve this task in a training-free manner. Differently from previous works [48, 57], we construct the cache based on attribute-object compatibility derived from web-scale databases and LLMs. We also introduced a soft-labeling scheme to account for the co-occurrence of multiple attributes within an image. Combining these two components allows to model the natural co-occurrence of attributes in the real world, together and across objects, while making the cache construction scalable. Our experiments demonstrate that COMCA is a strong training-free approach. Since COMCA is training-free, it does not suffer from differences between training and test attributes, contrary to training-based methods. This also enables COMCA to be adapted to special domains (e.g., aerial images) more easily than training-based methods, which would require specific re-training. Future work will focus on devising methods to address attribute detection even in more challenging settings where objects or attributes are not given [10]." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15105" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 156 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 156 + ], + "type": "text", + "content": "Acknowledgments. We acknowledge ISCRA for awarding this project access to the LEONARDO supercomputer, owned by the EuroHPC Joint Undertaking, hosted by CINECA (Italy). This work was sponsored by the MUR PNRR project FAIR - Future AI Research (PE00000013), funded by NextGeneration EU, and the EU projects AI4TRUST (No.101070190) and ELIAS (No.01120237)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 167, + 115, + 178 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 167, + 115, + 178 + ], + "spans": [ + { + "bbox": [ + 56, + 167, + 115, + 178 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 186, + 297, + 712 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 61, + 186, + 297, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 186, + 297, + 218 + ], + "spans": [ + { + "bbox": [ + 61, + 186, + 297, + 218 + ], + "type": "text", + "content": "[1] Zeynep Akata, Florent Perronnin, Zaid Harchaoui, and Cordelia Schmid. Label-embedding for attribute-based classification. In CVPR, 2013. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 220, + 296, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 220, + 296, + 275 + ], + "spans": [ + { + "bbox": [ + 61, + 220, + 296, + 275 + ], + "type": "text", + "content": "[2] Mahmoud Assran, Mathilde Caron, Ishan Misra, Piotr Bojanowski, Armand Joulin, Nicolas Ballas, and Michael Rabbat. Semi-supervised learning of visual features by nonparametrically predicting view assignments with support samples. In ICCV, 2021. 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 277, + 295, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 277, + 295, + 308 + ], + "spans": [ + { + "bbox": [ + 62, + 277, + 295, + 308 + ], + "type": "text", + "content": "[3] Alberto Baldrati, Lorenzo Agnolucci, Marco Bertini, and Alberto Del Bimbo. Zero-shot composed image retrieval with textual inversion. In ICCV, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 310, + 295, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 310, + 295, + 342 + ], + "spans": [ + { + "bbox": [ + 62, + 310, + 295, + 342 + ], + "type": "text", + "content": "[4] Tamara L Berg, Alexander C Berg, and Jonathan Shih. Automatic attribute discovery and characterization from noisy web data. In ECCV, 2010. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 343, + 295, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 343, + 295, + 376 + ], + "spans": [ + { + "bbox": [ + 62, + 343, + 295, + 376 + ], + "type": "text", + "content": "[5] Maria A Bravo, Sudhanshu Mittal, Simon Ging, and Thomas Brox. Open-vocabulary attribute detection. In CVPR, 2023. 1, 2, 3, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 377, + 295, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 377, + 295, + 421 + ], + "spans": [ + { + "bbox": [ + 62, + 377, + 295, + 421 + ], + "type": "text", + "content": "[6] Soravit Changpinyo, Piyush Sharma, Nan Ding, and Radu Soricut. Conceptual 12m: Pushing web-scale image-text pre-training to recognize long-tail visual concepts. In CVPR, 2021. 2, 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 422, + 296, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 422, + 296, + 464 + ], + "spans": [ + { + "bbox": [ + 62, + 422, + 296, + 464 + ], + "type": "text", + "content": "[7] Keyan Chen, Xiaolong Jiang, Yao Hu, Xu Tang, Yan Gao, Jianqi Chen, and Weidi Xie. Ovarnet: Towards open-vocabulary object attribute recognition. In CVPR, 2023. 1, 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 467, + 295, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 467, + 295, + 488 + ], + "spans": [ + { + "bbox": [ + 62, + 467, + 295, + 488 + ], + "type": "text", + "content": "[8] Steven Chen and Kristen Grauman. Compare and contrast: Learning prominent visual differences. In CVPR, 2018. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 490, + 295, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 490, + 295, + 543 + ], + "spans": [ + { + "bbox": [ + 62, + 490, + 295, + 543 + ], + "type": "text", + "content": "[9] Mehdi Cherti, Romain Beaumont, Ross Wightman, Mitchell Wortsman, Gabriel Ilharco, Cade Gordon, Christoph Schuhmann, Ludwig Schmidt, and Jenia Jitsev. Reproducible scaling laws for contrastive language-image learning. In CVPR, 2023. 5, 6, 7, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 545, + 295, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 545, + 295, + 578 + ], + "spans": [ + { + "bbox": [ + 57, + 545, + 295, + 578 + ], + "type": "text", + "content": "[10] Alessandro Conti, Enrico Fini, Massimiliano Mancini, Paolo Rota, Yiming Wang, and Elisa Ricci. Vocabulary-free image classification. NeurIPS, 36, 2024. 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 579, + 295, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 579, + 295, + 601 + ], + "spans": [ + { + "bbox": [ + 57, + 579, + 295, + 601 + ], + "type": "text", + "content": "[11] Ali Farhadi, Ian Endres, Derek Hoiem, and David Forsyth. Describing objects by their attributes. In CVPR, 2009. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 602, + 295, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 602, + 295, + 634 + ], + "spans": [ + { + "bbox": [ + 57, + 602, + 295, + 634 + ], + "type": "text", + "content": "[12] Ali Farhadi, Ian Endres, and Derek Hoiem. Attribute-centric recognition for cross-category generalization. In CVPR, 2010." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 636, + 295, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 636, + 295, + 657 + ], + "spans": [ + { + "bbox": [ + 57, + 636, + 295, + 657 + ], + "type": "text", + "content": "[13] Vittorio Ferrari and Andrew Zisserman. Learning visual attributes. NeurIPS, 2007. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 658, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 658, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 658, + 295, + 712 + ], + "type": "text", + "content": "[14] Isabel O Gallegos, Ryan A Rossi, Joe Barrow, Md Mehrab Tanjim, Sungchul Kim, Franck Dernoncourt, Tong Yu, Ruiyi Zhang, and Nesreen K Ahmed. Bias and fairness in large language models: A survey. arXiv preprint arXiv:2309.00770, 2023. 4" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 34, + "blocks": [ + { + "bbox": [ + 316, + 73, + 555, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 555, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 555, + 116 + ], + "type": "text", + "content": "[15] Peng Gao, Shijie Geng, Renrui Zhang, Teli Ma, Rongyao Fang, Yongfeng Zhang, Hongsheng Li, and Yu Qiao. Clip-adapter: Better vision-language models with feature adapters. IJCV, 132(2):581-595, 2024. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 119, + 555, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 119, + 555, + 151 + ], + "spans": [ + { + "bbox": [ + 317, + 119, + 555, + 151 + ], + "type": "text", + "content": "[16] Xiaoyuan Guo, Kezhen Chen, Jinmeng Rao, Yawen Zhang, Baochen Sun, and Jie Yang. Lowa: Localize objects in the wild with attributes. In NeurIPS-WS, 2023. 1, 2, 5, 6, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 153, + 555, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 153, + 555, + 195 + ], + "spans": [ + { + "bbox": [ + 317, + 153, + 555, + 195 + ], + "type": "text", + "content": "[17] Ziyu Guo, Renrui Zhang, Longtian Qiu, Xianzheng Ma, Xu-peng Miao, Xuming He, and Bin Cui. Calip: Zero-shot enhancement of clip with parameter-free attention. In AAAI, 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 198, + 555, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 198, + 555, + 241 + ], + "spans": [ + { + "bbox": [ + 317, + 198, + 555, + 241 + ], + "type": "text", + "content": "[18] Xintong Han, Zuxuan Wu, Phoenix X Huang, Xiao Zhang, Menglong Zhu, Yuan Li, Yang Zhao, and Larry S Davis. Automatic spatially-aware fashion concept discovery. In ICCV, 2017. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 244, + 555, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 244, + 555, + 275 + ], + "spans": [ + { + "bbox": [ + 316, + 244, + 555, + 275 + ], + "type": "text", + "content": "[19] Mehrdad Hosseinzadeh and Yang Wang. Composed query image retrieval using locally bounded features. In CVPR, 2020. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 278, + 555, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 278, + 555, + 311 + ], + "spans": [ + { + "bbox": [ + 316, + 278, + 555, + 311 + ], + "type": "text", + "content": "[20] Edward J Hu, Phillip Wallis, Zeyuan Allen-Zhu, Yanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. In ICLR, 2021. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 313, + 555, + 345 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 313, + 555, + 345 + ], + "spans": [ + { + "bbox": [ + 316, + 313, + 555, + 345 + ], + "type": "text", + "content": "[21] Phillip Isola, Joseph J Lim, and Edward H Adelson. Discovering states and transformations in image collections. In CVPR, 2015. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 348, + 554, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 348, + 554, + 380 + ], + "spans": [ + { + "bbox": [ + 316, + 348, + 554, + 380 + ], + "type": "text", + "content": "[22] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In ECCV, 2022. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 383, + 545, + 393 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 383, + 545, + 393 + ], + "spans": [ + { + "bbox": [ + 316, + 383, + 545, + 393 + ], + "type": "text", + "content": "[23] Glenn Jocher and Jing Qiu. Ultralytics yolo11. 2024. 5, 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 396, + 555, + 427 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 396, + 555, + 427 + ], + "spans": [ + { + "bbox": [ + 317, + 396, + 555, + 427 + ], + "type": "text", + "content": "[24] Shyamgopal Karthik, Karsten Roth, Massimiliano Mancini, and Zeynep Akata. Vision-by-language for training-free compositional image retrieval. In ICLR, 2024. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 430, + 555, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 430, + 555, + 483 + ], + "spans": [ + { + "bbox": [ + 317, + 430, + 555, + 483 + ], + "type": "text", + "content": "[25] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. IJCV, 123:32-73, 2017. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 486, + 555, + 517 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 486, + 555, + 517 + ], + "spans": [ + { + "bbox": [ + 317, + 486, + 555, + 517 + ], + "type": "text", + "content": "[26] Christoph H Lampert, Hannes Nickisch, and Stefan Harmeling. Learning to detect unseen object classes by between-class attribute transfer. In CVPR, 2009. 2, 5" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 317, + 520, + 555, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 520, + 555, + 553 + ], + "spans": [ + { + "bbox": [ + 317, + 520, + 555, + 553 + ], + "type": "text", + "content": "[27] Christoph H Lampert, Hannes Nickisch, and Stefan Harmeling. Attribute-based classification for zero-shot visual object categorization. IEEE TPAMI, 36(3):453-465, 2013. 2, 5, 6, 7" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 317, + 555, + 555, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 555, + 555, + 597 + ], + "spans": [ + { + "bbox": [ + 317, + 555, + 555, + 597 + ], + "type": "text", + "content": "[28] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, 2022. 2, 5, 6, 7" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 317, + 601, + 555, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 601, + 555, + 643 + ], + "spans": [ + { + "bbox": [ + 317, + 601, + 555, + 643 + ], + "type": "text", + "content": "[29] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML, 2023. 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 316, + 646, + 554, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 646, + 554, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 646, + 554, + 689 + ], + "type": "text", + "content": "[30] Victor Weixin Liang, Yuhui Zhang, Yongchan Kwon, Serena Yeung, and James Y Zou. Mind the gap: Understanding the modality gap in multi-modal contrastive representation learning. NeurIPS, 2022. 5" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 316, + 691, + 555, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 555, + 713 + ], + "type": "text", + "content": "[31] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. NeurIPS, 2024. 2" + } + ] + } + ], + "index": 33 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15106" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 296, + 95 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 296, + 95 + ], + "type": "text", + "content": "[32] Sachit Menon and Carl Vondrick. Visual classification via description from large language models. In ICLR, 2023. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 296, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 296, + 139 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 296, + 139 + ], + "type": "text", + "content": "[33] Roberto Navigli, Simone Conia, and Björn Ross. Biases in large language models: origins, inventory, and discussion. ACM Journal of Data and Information Quality, 15(2):1-21, 2023. 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 141, + 294, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 141, + 294, + 173 + ], + "spans": [ + { + "bbox": [ + 56, + 141, + 294, + 173 + ], + "type": "text", + "content": "[34] Zachary Novack, Julian McAuley, Zachary Chase Lipton, and Saurabh Garg. Chils: Zero-shot image classification with hierarchical label sets. In ICML, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 175, + 296, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 296, + 196 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 296, + 196 + ], + "type": "text", + "content": "[35] OpenAI. Introducing chatgpt. https://openai.com/blog/chatgpt, 2022.5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 198, + 296, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 198, + 296, + 218 + ], + "spans": [ + { + "bbox": [ + 56, + 198, + 296, + 218 + ], + "type": "text", + "content": "[36] Emin Orhan. A simple cache model for image recognition. NeurIPS, 2018. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 220, + 294, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 220, + 294, + 241 + ], + "spans": [ + { + "bbox": [ + 56, + 220, + 294, + 241 + ], + "type": "text", + "content": "[37] Devi Parikh and Kristen Grauman. Relative attributes. In ICCV, 2011. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 243, + 296, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 243, + 296, + 275 + ], + "spans": [ + { + "bbox": [ + 56, + 243, + 296, + 275 + ], + "type": "text", + "content": "[38] Genevieve Patterson and James Hays. Coco attributes: Attributes for people, animals, and objects. In ECCV, 2016. 1, 2, 5, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 277, + 296, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 277, + 296, + 320 + ], + "spans": [ + { + "bbox": [ + 56, + 277, + 296, + 320 + ], + "type": "text", + "content": "[39] Khoi Pham, Kushal Kafle, Zhe Lin, Zhihong Ding, Scott Cohen, Quan Tran, and Abhinav Shrivastava. Learning to predict visual attributes in the wild. In CVPR, 2021. 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 322, + 296, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 322, + 296, + 365 + ], + "spans": [ + { + "bbox": [ + 56, + 322, + 296, + 365 + ], + "type": "text", + "content": "[40] Khoi Pham, Kushal Kafle, Zhe Lin, Zhihong Ding, Scott Cohen, Quan Tran, and Abhinav Shrivastava. Improving closed and open-vocabulary attribute prediction using transformers. In ECCV, 2022. 1, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 366, + 296, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 366, + 296, + 421 + ], + "spans": [ + { + "bbox": [ + 56, + 366, + 296, + 421 + ], + "type": "text", + "content": "[41] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 1, 2, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 422, + 296, + 455 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 422, + 296, + 455 + ], + "spans": [ + { + "bbox": [ + 56, + 422, + 296, + 455 + ], + "type": "text", + "content": "[42] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 3, 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 456, + 296, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 456, + 296, + 499 + ], + "spans": [ + { + "bbox": [ + 56, + 456, + 296, + 499 + ], + "type": "text", + "content": "[43] Karsten Roth, Jae Myung Kim, A Koepke, Oriol Vinyals, Cordelia Schmid, and Zeynep Akata. Waffling around for performance: Visual classification with random words and broad concepts. In ICCV, 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 501, + 296, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 501, + 296, + 544 + ], + "spans": [ + { + "bbox": [ + 56, + 501, + 296, + 544 + ], + "type": "text", + "content": "[44] Kuniaki Saito, Kihyuk Sohn, Xiang Zhang, Chun-Liang Li, Chen-Yu Lee, Kate Saenko, and Tomas Pfister. Pic2word: Mapping pictures to words for zero-shot composed image retrieval. In CVPR, 2023. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 546, + 296, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 546, + 296, + 601 + ], + "spans": [ + { + "bbox": [ + 56, + 546, + 296, + 601 + ], + "type": "text", + "content": "[45] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. NeurIPS, 2022. 3, 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 602, + 294, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 602, + 294, + 634 + ], + "spans": [ + { + "bbox": [ + 56, + 602, + 294, + 634 + ], + "type": "text", + "content": "[46] Behjat Siddique, Rogerio S Feris, and Larry S Davis. Image ranking and retrieval based on multi-attribute queries. In CVPR, 2011. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 56, + 635, + 296, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 296, + 678 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 296, + 678 + ], + "type": "text", + "content": "[47] Kihyuk Sohn, Huiwen Chang, José Lezama, Luisa Polania, Han Zhang, Yuan Hao, Irfan Essa, and Lu Jiang. Visual prompt tuning for generative transfer learning. In CVPR, 2023. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 56, + 680, + 296, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 680, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 680, + 296, + 712 + ], + "type": "text", + "content": "[48] Vishaal Udandarao, Ankush Gupta, and Samuel Albanie. Susx: Training-free name-only transfer of vision-language models. In ICCV, 2023. 2, 3, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 72, + 555, + 689 + ], + "type": "list", + "angle": 0, + "index": 34, + "blocks": [ + { + "bbox": [ + 315, + 72, + 555, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 72, + 555, + 105 + ], + "spans": [ + { + "bbox": [ + 315, + 72, + 555, + 105 + ], + "type": "text", + "content": "[49] Peter Welinder, Steve Branson, Takeshi Mita, Catherine Wah, Florian Schroff, Serge Belongie, and Pietro Perona. Caltech-ucsd birds 200. 2010. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 107, + 555, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 555, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 555, + 150 + ], + "type": "text", + "content": "[50] Hui Wu, Yupeng Gao, Xiaoxiao Guo, Ziad Al-Halah, Steven Rennie, Kristen Grauman, and Rogerio Feris. Fashion iq: A new dataset towards retrieving images by natural language feedback. In CVPR, 2021. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 152, + 555, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 555, + 194 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 555, + 194 + ], + "type": "text", + "content": "[51] Yongqin Xian, Christoph H Lampert, Bernt Schiele, and Zeynep Akata. Zero-shot learning—a comprehensive evaluation of the good, the bad and the ugly. IEEE TPAMI, 41(9): 2251-2265, 2018. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 197, + 555, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 197, + 555, + 228 + ], + "spans": [ + { + "bbox": [ + 316, + 197, + 555, + 228 + ], + "type": "text", + "content": "[52] Hantao Yao, Rui Zhang, and Changsheng Xu. Visual-language prompt tuning with knowledge-guided context optimization. In CVPR, 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 230, + 555, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 230, + 555, + 252 + ], + "spans": [ + { + "bbox": [ + 316, + 230, + 555, + 252 + ], + "type": "text", + "content": "[53] Aron Yu and Kristen Grauman. Fine-grained visual comparisons with local learning. In CVPR, 2014. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 254, + 555, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 254, + 555, + 296 + ], + "spans": [ + { + "bbox": [ + 316, + 254, + 555, + 296 + ], + "type": "text", + "content": "[54] Jiahui Yu, Zirui Wang, Vijay Vasudevan, Legg Yeung, Mojtaba Seyedhosseini, and Yonghui Wu. Coca: Contrastive captioners are image-text foundation models. TMLR, 2022. 2, 5, 6, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 298, + 555, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 298, + 555, + 330 + ], + "spans": [ + { + "bbox": [ + 316, + 298, + 555, + 330 + ], + "type": "text", + "content": "[55] Yan Zeng, Xinsong Zhang, and Hang Li. Multi-grained vision language pre-training: Aligning texts with visual concepts. ICML, 2022. 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 332, + 555, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 332, + 555, + 364 + ], + "spans": [ + { + "bbox": [ + 316, + 332, + 555, + 364 + ], + "type": "text", + "content": "[56] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In ICCV, 2023. 5, 6, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 365, + 555, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 365, + 555, + 409 + ], + "spans": [ + { + "bbox": [ + 316, + 365, + 555, + 409 + ], + "type": "text", + "content": "[57] Renrui Zhang, Wei Zhang, Rongyao Fang, Peng Gao, Kunchang Li, Jifeng Dai, Yu Qiao, and Hongsheng Li. Tip-adapter: Training-free adaption of clip for few-shot classification. In ECCV, 2022. 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 411, + 555, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 411, + 555, + 453 + ], + "spans": [ + { + "bbox": [ + 316, + 411, + 555, + 453 + ], + "type": "text", + "content": "[58] Xiao Zhang, Rui Zhao, Yu Qiao, Xiaogang Wang, and Hongsheng Li. Adacos: Adaptively scaling cosine logits for effectively learning deep face representations. In CVPR, 2019. 5" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 456, + 555, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 456, + 555, + 487 + ], + "spans": [ + { + "bbox": [ + 316, + 456, + 555, + 487 + ], + "type": "text", + "content": "[59] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Conditional prompt learning for vision-language models. In CVPR, 2022. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 489, + 555, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 489, + 555, + 521 + ], + "spans": [ + { + "bbox": [ + 316, + 489, + 555, + 521 + ], + "type": "text", + "content": "[60] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. IJCV, 130(9):2337-2348, 2022. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 523, + 555, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 523, + 555, + 566 + ], + "spans": [ + { + "bbox": [ + 316, + 523, + 555, + 566 + ], + "type": "text", + "content": "[61] Yifei Zhou, Juntao Ren, Fengyu Li, Ramin Zabih, and Ser Nam Lim. Test-time distribution normalization for contrastively learned visual-language models. NeurIPS, 2024. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 568, + 555, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 568, + 555, + 600 + ], + "spans": [ + { + "bbox": [ + 316, + 568, + 555, + 600 + ], + "type": "text", + "content": "[62] Beier Zhu, Yulei Niu, Yucheng Han, Yue Wu, and Hanwang Zhang. Prompt-aligned gradient for prompt tuning. In ICCV, 2023. 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 316, + 602, + 555, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 602, + 555, + 645 + ], + "spans": [ + { + "bbox": [ + 316, + 602, + 555, + 645 + ], + "type": "text", + "content": "[63] William Y Zhu, Keren Ye, Junjie Ke, Jiahui Yu, Leonidas Guibas, Peyman Milanfar, and Feng Yang. Artvlm: Attribute recognition through vision-based prefix language modeling. ECCV, 2024. 2, 5, 6" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 316, + 647, + 555, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 555, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 555, + 689 + ], + "type": "text", + "content": "[64] Xiangyang Zhu, Renrui Zhang, Bowei He, Aojun Zhou, Dong Wang, Bin Zhao, and Peng Gao. Not all features matter: Enhancing few-shot clip with adaptive prior refinement. In ICCV, 2023. 3" + } + ] + } + ], + "index": 33 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15107" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Compositional Targeted Multi-Label Universal Perturbations/616cefba-8863-49bf-987c-2e525bd606b6_content_list.json b/2025/Compositional Targeted Multi-Label Universal Perturbations/616cefba-8863-49bf-987c-2e525bd606b6_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..6594152faf4b70ca88b745b7e697076e90e55a1e --- /dev/null +++ b/2025/Compositional Targeted Multi-Label Universal Perturbations/616cefba-8863-49bf-987c-2e525bd606b6_content_list.json @@ -0,0 +1,1762 @@ +[ + { + "type": "text", + "text": "Compositional Targeted Multi-Label Universal Perturbations", + "text_level": 1, + "bbox": [ + 189, + 130, + 810, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hassan Mahmood Northeastern University", + "bbox": [ + 225, + 181, + 437, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "mahmood.h@northeastern.edu", + "bbox": [ + 225, + 219, + 459, + 232 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ehsan Elhamifar Northeastern University e.elhamifar@northeastern.edu", + "bbox": [ + 514, + 181, + 764, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 268, + 326, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Generating targeted universal perturbations for multi-label recognition is a combinatorially hard problem that requires exponential time and space complexity. To address the problem, we propose a compositional framework. We show that a simple independence assumption on label-wise universal perturbations naturally leads to an efficient optimization that requires learning affine convex cones spanned by label-wise universal perturbations, significantly reducing the problem complexity to linear time and space. During inference, the framework allows generating universal perturbations for novel combinations of classes in constant time. We demonstrate the scalability of our method on large datasets and target sizes, evaluating its performance on NUS-WIDE, MS-COCO, and OpenImages using state-of-the-art multi-label recognition models. Our results show that our approach outperforms baselines and achieves results comparable to methods with exponential complexity. The code is available at https://github.com/hassanmahmood/UMLLAttacks.git", + "bbox": [ + 89, + 300, + 485, + 588 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 603, + 220, + 618 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Deep neural networks have been very successful at image-recognition, demonstrating high accuracy in various applications. However, they are also vulnerable to adversarial attacks [9, 25, 30, 31, 43, 74, 84, 86], where small changes to the input image can cause the model to misclassify the image. Notably, Moosavi-Dezfooli et al. [67] discovered that a single perturbation, called universal adversarial perturbation (UAP), can be crafted to misclassify a large number of images. This motivated a large body of work on generating universal adversarial attacks [7, 36, 38, 65, 69, 110].", + "bbox": [ + 89, + 628, + 482, + 779 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Existing works on adversarial attacks have mainly focused on multi-class recognition (MCR) whose goal is to identify a single class in an image. On the other hand, many real-world applications require identifying all classes in an image e.g., document classification [75], scene understanding [85], facial attribute recognition [26], image annotation [15], human-object interaction [33], and several safety and security-critical applications e.g., recogniz-", + "bbox": [ + 89, + 780, + 482, + 901 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/18ff0e158377b42c24a4be820811f27dd53d336e0bef7b03aca80296df8ac134.jpg", + "image_caption": [ + "Figure 1. Targeted Multi-Label Universal Perturbations. Our framework allows composing (shown as $\\oplus$ ) different label-specific universal perturbations ('Animal', 'Person', 'Vehicle') to attack multiple classes together. These perturbations, once added (shown as $\\bullet$ ) to an image, cause target labels to be misclassified." + ], + "image_footnote": [], + "bbox": [ + 539, + 282, + 877, + 494 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ing diseases [24], autonomous driving [47], and surveillance [57, 108]. Therefore, the goal of multi-label learning (MLL) is to identify all labels present in an image [14, 20, 23, 39, 40, 55, 77, 100, 115]. Given the increasing number of applications relying on MLL, it is crucial to understand the internal workings of the MLL models, identify their vulnerabilities, and make them robust. Recently, MLL models have been shown to be vulnerable to adversarial attacks [1, 2, 42, 63, 88], yet the vulnerability of MLL systems against universal perturbations has not been explored.", + "bbox": [ + 511, + 597, + 906, + 748 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Crafting universal perturbations for MCR and MLL have fundamental differences due to the distinct structuring of decision boundaries in these problems. In MCR, a universal perturbation to turn off a label in images will result in the prediction of another label. Conversely, in MLL, multiple labels can co-exist, and the objective of an attack might be to misclassify any subset of these labels either by turning the labels off or on (see Fig. 2). In MLL, any combination of labels can be present in images. For the label set $\\mathcal{C}$ , there are $2^{|C|}$ possible combinations of labels. Universal pertur", + "bbox": [ + 511, + 750, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "20580", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "bations identify specific directions in the input space where the classifier is consistently vulnerable to attacks. What makes this particularly concerning in multi-label learning is that there can be exponentially many directions in the input space to attack the model - corresponding to different combinations of labels.", + "bbox": [ + 88, + 90, + 480, + 181 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "One approach for generating universal perturbations for MLL is to generate universal vectors that flip the prediction of all or most labels. However, in practice, the adversary usually seeks to modify a specific small set of labels while keeping the prediction of others unchanged (see Fig. 1). This is called the targeted attack and requires learning $2^{|\\mathcal{C}|}$ distinct universal perturbation vectors. Learning each such universal vector requires iterating through the dataset, resulting in an exponential time and space complexity.", + "bbox": [ + 89, + 181, + 480, + 316 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Existing universal perturbation methods focus on MCR setting [7, 8, 60]. Adapting these methods to MLL would require learning a separate universal perturbation for each possible label combination, incurring exponential complexity. Although one could learn label-wise independent perturbations, no effective method exists to combine them for attacking specific label combinations. We empirically show that naively combining them does not lead to successful attacks. Similarly, generative attack approaches [27, 69, 76] would be ineffective for unseen label combinations.", + "bbox": [ + 89, + 318, + 480, + 468 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Contributions. In this paper, we develop an efficient method for crafting targeted universal adversarial attacks on multi-label learning (MLL) models, addressing the aforementioned combinatorial complexity problem. Our key assumption is to generate a universal attack on a target set as a composition of universal label-wise attacks of the target labels (see Fig. 1). Based on this assumption, we derive a formulation which requires learning affine convex cones spanned by label-wise universal vectors. Our formulation requires linear training time complexity and after trained, the universal perturbation to attack any combination of labels can be composed in constant time using the label-wise attack vectors. Our key contributions are:", + "bbox": [ + 89, + 469, + 480, + 665 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Existing MCR-based approaches are not inherently designed to attack multiple targeted labels for MLL models. Our work is the first to address this gap, offering an efficient and principled solution for MLL setting;", + "- Our proposed framework to learn targeted universal perturbations for MLL has linear time and space complexity and constructs attacks on any target set in constant time, while achieving performance comparable to the existing oracle methods with exponential complexity;", + "- Our results visualize and provide deeper understanding of the universal adversarial regions learnt by our method;", + "- We evaluate our method on NUS-WIDE, MS-COCO, and OpenImages benchmark datasets using state-of-the-art multi-label learning models against various baselines, showing the effectiveness of our method." + ], + "bbox": [ + 89, + 667, + 480, + 893 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 513, + 89, + 663, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Multi-Label Recognition", + "text_level": 1, + "bbox": [ + 513, + 114, + 738, + 131 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "There has been significant recent interest in multi-label learning due to its applications in various tasks. In MLL, the goal is to identify all concepts/labels in an image. This is particularly a harder task than multi-class recognition, since there can be exponentially many possible combinations of classes present in an image and can be related to each other. To address this problem, early works [90, 97] considered independent binary classifiers to classify the presence or absence of each class. However, since such a method does not consider the correlations among labels, several other approaches have been proposed that exploit label correlations or external knowledge [13, 18, 34, 45, 49, 51, 70, 94-96, 98, 101, 105-107, 113, 118], or use graphical models [50, 52]. More recent works use attention mechanism or novel loss functions to consider the positive and negative sample imbalance [6, 17, 53, 54, 58, 93, 112].", + "bbox": [ + 511, + 138, + 903, + 380 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Adversarial Attacks", + "text_level": 1, + "bbox": [ + 513, + 392, + 702, + 406 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Adversarial attacks [3, 12, 81, 89, 99] have gained a lot of attention as they identify the key vulnerability of deep neural networks (DNNs) to small imperceptible changes in the input. Researchers have explored various ways that attackers can manipulate data to fool models and developed several defenses to counter these attacks [4, 5, 10, 11, 19, 21, 22, 28, 29, 35, 46, 62, 71-73, 78, 82-84, 91, 92, 102, 114]. Based on the attacker's objective and the target setting, the adversarial attacks can be categorized in different ways, such as targeted and untargeted attacks, white-box (which assume full access to the target model parameters) and black-box attacks (no access to the model parameters), or instance-specific and instance-agnostic attacks.", + "bbox": [ + 511, + 415, + 903, + 611 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.3. Universal Adversarial Attacks", + "text_level": 1, + "bbox": [ + 513, + 619, + 782, + 635 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In instance-specific attacks, the perturbations are generated and targeted for each input image, whereas, instance-agnostic attacks [38, 65, 69] generate a single universal perturbation that can fool all input instances. These universal adversarial perturbations(UAPs) can be untargeted or targeted. The goal of the untargeted universal perturbation is to perturb any image such that it is misclassified by the target model [67, 76]. In the targeted case, the goal is to learn the universal perturbation such that it misclassifies any given image to a specific class [7, 110]. Universal perturbations were first proposed by Moosavi-Dezfooli et al. [67] and were generated by iteratively computing and aggregating perturbations using DeepFool [66] across multiple inputs. Since then, several methods have been proposed to learn better universal perturbations [60, 109–111]. To learn UAPs with better generalization, Liu et al. [59] proposed a gradient aggregation method to address gradient vanish", + "bbox": [ + 511, + 643, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "20581", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/26d5e55b3778924cabe3c9374155441b5ea7a5a4a96a5c55e9b4f5499ef65abd.jpg", + "image_caption": [ + "Current Approach", + "Figure 2. In multi-label recognition, images can have any combination of labels present, leading to a combinatorial explosion in possible universal perturbations. Traditional methods (shown on the left) require exponential time and space complexity to learn perturbations for all class combinations. Our approach (shown on the right) reduces this complexity by learning $|\\mathcal{C}|$ affine convex cones, each representing a region of universal perturbations for a specific class. This allows us to efficiently generate perturbations for any combination of classes." + ], + "image_footnote": [], + "bbox": [ + 130, + 121, + 367, + 280 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/9793affce6d740d16f153eabf1d656ee39b822d0ed8998a83d9f7682c8617ad9.jpg", + "image_caption": [ + "Overview of the Proposed Method" + ], + "image_footnote": [], + "bbox": [ + 380, + 116, + 591, + 276 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/1eea43c9d7a6611c4cbdcbdc9def97ea6207d7a227127c44538e2282a68805be.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 645, + 119, + 870, + 270 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ing problem and Li et al. [48] proposed to utilize instance-specific attack method to generate dominant UAPs.", + "bbox": [ + 88, + 364, + 482, + 396 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.4. Multi-Label Adversarial Attacks", + "text_level": 1, + "bbox": [ + 89, + 407, + 377, + 422 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Many recent works have studied adversarial attacks for multi-label recognition models [41, 61, 103, 104, 116, 117]. Song et al. [87] proposed white-box multi-label attack framework and Hu et al. [32] proposed to exploit ranking relations to design attacks for top- $k$ multi-label models. Aich et al. [1] proposed a CLIP-based generative model to generate multi-object attacks in the black-box setting. Jia et al. [37] proposed theoretical robustness guarantees to defend against multi-label adversarial attacks and Melacci et al. [64] exploited domain knowledge context to detect multi-label adversarial attacks. Mahmood and Elhamifar [63] proposed an optimization to address the problem of negative gradient correlation in generating multi-label attacks. These works mainly address instance-specific perturbations for multi-label setting. To the best of our knowledge, only Hu et al. [32] proposed universal untargeted attack where a single perturbation was learnt to attack and replace top- $k$ predicted labels across images. In comparison, we address the problem of learning targeted universal perturbations to attack specific combinations of labels.", + "bbox": [ + 91, + 429, + 483, + 731 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Multi-Label Universal Perturbations", + "text_level": 1, + "bbox": [ + 89, + 747, + 421, + 762 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Problem Formulation", + "text_level": 1, + "bbox": [ + 89, + 771, + 294, + 786 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Consider a multi-label classifier $\\mathcal{F}:\\mathbb{R}^d\\to \\mathbb{R}^{|\\mathcal{C}|}$ where $\\mathcal{C}$ is the set of all labels. For an image $\\pmb {x}\\in \\mathbb{R}^d$ , let $\\pmb {y}\\in \\{0,1\\}^{|\\mathcal{C}|}$ denote the set of its labels, where each entry indicates the absence (0) or presence (1) of the corresponding label in $\\mathcal{C}$ in the image. The classifier $\\mathcal{F} = \\{\\mathcal{F}_1,\\mathcal{F}_2,\\dots ,\\mathcal{F}_{|\\mathcal{C}|}\\}$ consists of $|\\mathcal{C}|$ binary classifiers (one for each label), where $\\mathcal{F}_c(\\pmb {x})\\in (-\\infty , + \\infty)$ represents the logit of label $c$ .The", + "bbox": [ + 89, + 792, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "probability of label $c$ being present in the image is given by $\\hat{o}_c = \\sigma (\\mathcal{F}_c(\\pmb {x}))$ , where $\\sigma (.)$ is the sigmoid function, hence, it is considered present (1), if $\\hat{o}_c\\geq 0.5$ , and absent (0) otherwise. We assume that $\\mathcal{F}$ has already been trained on a set of images. Let $\\Omega \\subseteq \\mathcal{C}$ denote the set of target labels that we want to attack. Our goal is to generate a targeted universal perturbation for $\\Omega$ .", + "bbox": [ + 511, + 364, + 906, + 470 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Multi-Label Universal Perturbations", + "text_level": 1, + "bbox": [ + 511, + 481, + 831, + 496 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We want to find a single (universal) perturbation vector, $\\pmb{u}_{\\Omega}$ , that can attack a set of target labels in $\\Omega$ in any image $\\pmb{x}$ containing them. One can generate such a vector by maximizing the expected multi-label learning loss for the labels in $\\Omega$ across the training dataset, while restricting the norm of $\\pmb{u}_{\\Omega}$ (we consider $\\ell_{\\infty}$ -norm). However, due to correlations among labels in MLL, attacking the target labels may inadvertently affect non-target labels $(\\bar{\\Omega} = \\mathcal{C} \\setminus \\Omega)$ . To mitigate this, we seek to minimize the influence of the universal perturbation vector on non-target labels. This requires solving the following oracle optimization,", + "bbox": [ + 511, + 503, + 906, + 671 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nP _ {\\Omega} ^ {*}: \\max _ {\\boldsymbol {u} _ {\\Omega}} \\sum_ {k \\in \\Omega} \\mathcal {H} _ {k} (\\boldsymbol {u} _ {\\Omega}) - \\gamma \\sum_ {q \\in \\Omega} \\mathcal {R} _ {q} (\\boldsymbol {u} _ {\\Omega}), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 578, + 676, + 906, + 705 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathcal{H}_k$ ensures attacking the label $k$ in the target set and $\\mathcal{R}_q$ tries to prevent modifying the prediction of non-target labels using the universal perturbation $\\pmb{u}_{\\Omega}$ . Here, $\\gamma$ is a scalar hyperparameter setting a trade-off between the two terms. More specifically, we define $\\mathcal{H}_k$ and $\\mathcal{R}_q$ as", + "bbox": [ + 511, + 710, + 905, + 787 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {H} _ {k} \\left(\\boldsymbol {u} _ {\\Omega}\\right) \\triangleq \\mathbb {E} _ {\\left(\\boldsymbol {x}, \\boldsymbol {y}\\right)} \\mathcal {L} _ {\\mathrm {b c e}} \\left(\\mathcal {F} _ {k} \\left(\\boldsymbol {x} + \\Pi_ {\\epsilon} \\left(\\boldsymbol {u} _ {\\Omega}\\right)\\right), y _ {k}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 513, + 794, + 903, + 811 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {R} _ {q} (\\boldsymbol {u} _ {\\Omega}) \\triangleq \\mathbb {E} _ {(\\boldsymbol {x}, \\boldsymbol {y})} \\max \\left\\{0, - \\tanh \\left(\\mathcal {F} _ {q} (\\boldsymbol {x}) \\mathcal {F} _ {q} (\\boldsymbol {x} + \\Pi_ {\\epsilon} (\\boldsymbol {u} _ {\\Omega}))\\right) \\right\\},\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 814, + 903, + 832 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\Pi_{\\epsilon}$ projects the input on p-norm ball of radius $\\epsilon$ , $\\mathcal{L}_{\\mathrm{bce}}$ is the binary cross-entropy loss and the expectation is with respect to samples from the dataset. Therefore, maximizing $\\mathcal{H}_k(\\pmb{u}_{\\Omega})$ ensures to find a perturbation that modifies the", + "bbox": [ + 511, + 839, + 906, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "20582", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "prediction of each target label $k$ . On the other hand, $\\mathcal{R}_k$ aims to minimize the influence of the perturbation $(\\pmb{u}_{\\Omega})$ on each non-target label $q$ . This is done by incurring a positive penalty when the signs of the logits for the unperturbed and perturbed inputs differ from each other. Thus, minimizing this term leads to the same signs for the logits of each non-target label before and after the attack.", + "bbox": [ + 89, + 90, + 483, + 196 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Compositional Multi-Label Universal Attacks (CMLU)", + "text_level": 1, + "bbox": [ + 89, + 204, + 483, + 234 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "There are $2^{|C|}$ possible subsets of labels $\\mathcal{C}$ . Therefore, finding all universal perturbations $\\{\\pmb{u}_{\\Omega}\\}$ by independently solving (1) for each subset $\\Omega$ leads to the exponential complexity $O(2^{|C|})$ . We develop a compositional multi-label universal attack (CMLU) that allows us to find any universal perturbation in linear complexity $O(|C|)$ .", + "bbox": [ + 89, + 239, + 483, + 330 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Let $\\mathcal{P}(\\mathcal{C})$ denote the power set of $\\mathcal{C}$ , i.e., the set of all subsets of $\\mathcal{C}$ . Our goal is to find all $\\pmb{u}_{\\Omega}$ 's by solving (1) for every $\\Omega \\in \\mathcal{P}(\\mathcal{C})$ ,", + "bbox": [ + 89, + 332, + 483, + 377 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\sum_ {\\Omega \\in \\mathcal {P} (\\mathcal {C})} \\max _ {\\boldsymbol {u} _ {\\Omega}} \\left[ \\sum_ {k \\in \\Omega} \\mathcal {H} _ {k} \\left(\\boldsymbol {u} _ {\\Omega}\\right) - \\gamma \\sum_ {q \\in \\bar {\\Omega}} \\mathcal {R} _ {q} \\left(\\boldsymbol {u} _ {\\Omega}\\right) \\right]. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 132, + 377, + 480, + 412 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As mentioned above, solving (3) over the power set of $\\mathcal{C}$ has exponential complexity. To tackle this problem, we make the following assumption: each universal perturbation $\\pmb{u}_{\\Omega}$ can be written as the sum of atomic universal perturbations", + "bbox": [ + 89, + 415, + 483, + 474 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {u} _ {\\Omega} \\triangleq \\sum_ {i \\in \\Omega} \\boldsymbol {u} _ {i}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 238, + 478, + 480, + 508 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\pmb{u}_i$ is the universal perturbation on the label $i$ . We also assume that the label-wise universal perturbations are norm-constrained and mutually orthogonal, i.e., $\\| \\pmb{u}_i\\|_p \\leq \\epsilon$ and $\\hat{\\pmb{u}}_i^\\top \\hat{\\pmb{u}}_j = 0$ for each $i \\neq j$ , where $\\hat{\\pmb{u}} = \\pmb{u} / \\| \\pmb{u}\\|_2$ .", + "bbox": [ + 89, + 512, + 483, + 575 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Let $\\mathbf{U} = \\left[ \\begin{array}{lll}\\pmb {u}_1 & \\pmb {u}_2 & \\dots & \\pmb{u}_{|\\mathcal{C}|} \\end{array} \\right]$ represent a matrix whose columns are label-wise universal perturbations and $\\lambda_{\\Omega}$ denote a vector whose elements indexed by $\\Omega$ are one and other elements are zero. Based on our assumption, we optimize $\\mathbf{U}$ so that we can construct any $\\pmb{u}_{\\Omega}$ using the columns of $\\mathbf{U}$ . Given our assumptions, we can rewrite (3) as", + "bbox": [ + 89, + 574, + 483, + 665 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\max _ {\\boldsymbol {U}} \\sum_ {\\Omega \\in \\mathcal {P} (\\mathcal {C})} \\sum_ {k \\in \\Omega} \\left(\\mathcal {H} _ {k} \\left(\\boldsymbol {u} _ {k} + \\boldsymbol {U} \\boldsymbol {\\lambda} _ {\\Omega \\backslash k}\\right) - \\frac {\\gamma}{| \\Omega |} \\sum_ {q \\in \\Omega} \\mathcal {R} _ {q} \\left(\\boldsymbol {u} _ {k} + \\boldsymbol {U} \\boldsymbol {\\lambda} _ {\\Omega \\backslash k}\\right)\\right), \\\\ s. t. \\hat {\\boldsymbol {U}} ^ {\\top} \\hat {\\boldsymbol {U}} = \\boldsymbol {I}, \\| \\boldsymbol {U} \\| _ {1, p} \\leq \\epsilon \\tag {5} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 93, + 667, + 480, + 720 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\hat{U}$ is column-normalized $U$ and $\\| U\\|_{1,p} = \\max_{1\\leq i\\leq |\\mathcal{C}|}\\| u_i\\| _p$ , is the maximum norm of the columns of $U$ . Notice that we wrote $u_{\\Omega} = u_k + U\\lambda_{\\Omega \\backslash k}$ for a $k$ that belongs to $\\Omega$ , where $\\Omega \\backslash k$ denotes removing $k$ from $\\Omega$ .", + "bbox": [ + 89, + 724, + 483, + 786 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "One can see that, for any $k \\in \\mathcal{C}$ , there are $2^{|C| - 1}$ subsets that contain $k$ . Each of these subsets would contain $k$ and classes from the powerset $\\mathcal{P}(\\mathcal{C} \\backslash k)$ . We can rewrite (5) as (see the supplementary materials for derivation details)", + "bbox": [ + 89, + 786, + 483, + 845 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\max _ {\\boldsymbol {U}} \\sum_ {k \\in \\mathcal {C}} \\sum_ {\\Omega \\in \\mathcal {P} (\\mathcal {C} \\backslash k)} \\left(\\mathcal {H} _ {k} \\left(\\boldsymbol {u} _ {k} + \\boldsymbol {U} \\boldsymbol {\\lambda} _ {\\Omega \\backslash k}\\right) - \\eta \\sum_ {q \\in \\{\\bar {\\Omega} \\backslash k \\}} \\mathcal {R} _ {q} \\left(\\boldsymbol {u} _ {k} + \\boldsymbol {U} \\boldsymbol {\\lambda} _ {\\Omega \\backslash k}\\right)\\right)\n$$\n", + "text_format": "latex", + "bbox": [ + 93, + 848, + 483, + 881 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {s . t .} \\hat {\\boldsymbol {U}} ^ {\\top} \\hat {\\boldsymbol {U}} = \\boldsymbol {I}, \\| \\boldsymbol {U} \\| _ {1, p} \\leq \\epsilon , \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 96, + 883, + 482, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\eta = \\gamma /|\\Omega |$ . The power set of $\\mathcal{C}$ can be written as the union of all subsets of $\\mathcal{C}$ .. $\\mathcal{P}(\\mathcal{C})\\triangleq \\bigcup_{i = 0}^{|C|}\\binom {\\mathcal{C}}{i}$ where, each $\\binom{\\mathcal{C}}{i}$ has all the subsets of $\\mathcal{C}$ of size $i$ . In our case, $S\\in \\binom{\\mathcal{C}}{i}$ would represent the indices of the columns of $\\pmb{U}$ . So, we can write the summation across the power set as:", + "bbox": [ + 511, + 90, + 906, + 171 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\max_{\\boldsymbol{U}}\\sum_{k\\in \\mathcal{C}}\\sum_{t = 0}^{\\left|\\mathcal{C}\\right| - 1}\\sum_{S\\in \\binom {\\mathcal{C}\\setminus k}{t}}\\Bigl(\\mathcal{H}_{k}(\\boldsymbol{u}_{k} + \\boldsymbol {U}\\boldsymbol{\\lambda}_{S}) - \\eta \\sum_{q\\in \\{\\bar{S}\\setminus k\\}}\\mathcal{R}_{q}(\\boldsymbol{u}_{k} + \\boldsymbol {U}\\boldsymbol{\\lambda}_{S})\\Bigr),\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 174, + 911, + 218 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ns. t. \\hat {\\boldsymbol {U}} ^ {\\top} \\hat {\\boldsymbol {U}} = \\boldsymbol {I}, \\| \\boldsymbol {U} \\| _ {1, p} \\leq \\epsilon , \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 220, + 903, + 241 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In (7), for a fixed $k$ and $t$ , $\\lambda_{S}$ has exactly $t$ non-zero elements and $S$ iterates through all possible combinations of $\\{C \\backslash k\\}$ of size $t$ . Therefore, the sum across all $\\lambda_{S}$ (for $S \\in \\binom{\\mathcal{C} \\backslash k}{t}$ ) would be equal to the sum across all binary vectors $\\lambda \\in \\{0,1\\}^{|\\mathcal{C}|}$ such that exactly $t$ elements are nonzero (i.e., $\\| \\lambda \\|_0 = t$ ) and $\\lambda_{k} = 0$ . Where, $\\| \\cdot \\|_0$ denotes the $\\ell_0$ -norm, which counts the number of nonzero elements of the given input. So we can write (7) as", + "bbox": [ + 511, + 246, + 906, + 369 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\max _ {\\boldsymbol {U}} \\sum_ {k \\in \\mathcal {C}} \\sum_ {t = 0} ^ {| \\mathcal {C} | - 1} \\sum_ {\\substack {\\boldsymbol {\\lambda} \\in \\{0, 1 \\} ^ {| \\mathcal {C} |} \\\\ \\| \\boldsymbol {\\lambda} \\| _ {0} = t, \\boldsymbol {\\lambda} _ {k} = 0}} \\left[ \\mathcal {H} _ {k} (\\boldsymbol {u} _ {k} + \\boldsymbol {U} \\boldsymbol {\\lambda}) \\right. \\tag{8}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 372, + 903, + 422 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left. - \\eta \\sum_ {q \\in \\{\\mathcal {C} \\backslash k \\}} \\delta (\\boldsymbol {\\lambda} _ {q} = 0) \\mathcal {R} _ {q} (\\boldsymbol {u} _ {k} + \\boldsymbol {U} \\boldsymbol {\\lambda}) \\right] \\text {s . t .} \\hat {\\boldsymbol {U}} ^ {\\top} \\hat {\\boldsymbol {U}} = \\boldsymbol {I}, \\| \\boldsymbol {U} \\| _ {1, p} \\leq \\epsilon ,\n$$\n", + "text_format": "latex", + "bbox": [ + 517, + 425, + 906, + 458 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\delta$ is the indicator function which is 1 if the input condition is True and 0 otherwise. As $\\lambda$ is $|\\mathcal{C}|$ dimensional, the second loss term in (8) iterates through each index of $\\lambda$ i.e., $(q\\in \\{\\mathcal{C}\\backslash k\\})$ and computes the loss of label $q$ if $\\lambda_q = 0$ .", + "bbox": [ + 511, + 460, + 905, + 521 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We investigate two ways to approximately solve (8). In the first approach, we solve the lower bound of (8) as", + "bbox": [ + 511, + 521, + 905, + 551 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {CMLU} _ {\\alpha}: \\max _ {\\boldsymbol {U}} \\sum_ {k \\in \\mathcal {C}} \\sum_ {t \\in \\mathcal {T}} 2 ^ {| \\mathcal {C} |} \\min _ {\\substack {\\boldsymbol {\\lambda} \\in \\{0,1 \\} ^ {| C |} \\\\ \\| \\boldsymbol {\\lambda} \\| _ {0} = t, \\boldsymbol {\\lambda} _ {k} = 0}} \\left[ \\mathcal {H} _ {k} (\\boldsymbol {u} _ {k} + \\boldsymbol {U} \\boldsymbol {\\lambda}) \\right. \\tag{9}\n$$\n", + "text_format": "latex", + "bbox": [ + 516, + 556, + 903, + 593 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left. - \\eta \\sum_ {q \\in \\{\\mathcal {C} \\backslash k \\}} \\delta (\\boldsymbol {\\lambda} _ {q} = 0) \\mathcal {R} _ {q} (\\boldsymbol {u} _ {k} + \\boldsymbol {U} \\boldsymbol {\\lambda}) \\right] \\text {s . t .} \\hat {\\boldsymbol {U}} ^ {\\top} \\hat {\\boldsymbol {U}} = \\boldsymbol {I}, \\| \\boldsymbol {U} \\| _ {1, p} \\leq \\epsilon\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 595, + 901, + 628 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Where, $\\mathcal{T} = \\{0,1,\\dots ,|\\mathcal{C}| - 1\\}$ . (9) requires maximizing the minimal loss. To solve the inner minimization problem, for a fixed $k$ and $t$ , we randomly initialize $\\lambda$ and iteratively update it for a fixed number of iterations. We then project $\\lambda$ on the feasible set by keeping the largest $t$ values as 1 and setting the rest to 0. We then keep the $\\lambda$ fixed and solve the outer maximization in (9) w.r.t. $\\pmb{U}$ . In practice, instead of iterating through all $t\\in \\mathcal{T}$ , for each $k$ , we randomly choose $t$ . The steps are shown in Algorithm 1.", + "bbox": [ + 511, + 635, + 906, + 770 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the second approach, we consider the continuous relaxation of $\\lambda$ and approximately solve it by uniformly randomly sampling $\\lambda \\in [0,1]^{|\\mathcal{C}|}$ ,", + "bbox": [ + 511, + 771, + 905, + 816 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {C M L U} _ {\\beta}: \\max _ {\\boldsymbol {U}} \\sum_ {k \\in \\mathcal {C}} \\sum_ {t \\in \\mathcal {T}} \\sum_ {\\substack {\\boldsymbol {\\lambda} \\in [ 0, 1 ] ^ {| \\mathcal {C} |} \\\\ \\| \\boldsymbol {\\lambda} \\| _ {0} = t, \\boldsymbol {\\lambda} _ {k} = 0}} \\left[ \\mathcal {H} _ {k} (\\boldsymbol {u} _ {k} + \\boldsymbol {U} \\boldsymbol {\\lambda}) \\right. \\tag{10}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 821, + 903, + 866 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left. - \\eta \\sum_ {q \\in \\{\\mathcal {C} \\backslash k \\}} \\delta (\\boldsymbol {\\lambda} _ {q} = 0) \\mathcal {R} _ {q} (\\boldsymbol {u} _ {k} + \\boldsymbol {U} \\boldsymbol {\\lambda}) \\right] \\text {s . t .} \\hat {\\boldsymbol {U}} ^ {\\top} \\hat {\\boldsymbol {U}} = \\boldsymbol {I}, \\| \\boldsymbol {U} \\| _ {1, p} \\leq \\epsilon .\n$$\n", + "text_format": "latex", + "bbox": [ + 517, + 868, + 903, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "20583", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Eq. (9) and (10) can be explained intuitively. The first loss term $\\mathcal{H}_k$ optimizes each label-wise universal perturbation $\\pmb{u}_k$ such that it is robust to the noise generated by composing other universal perturbations. Specifically, it aims to maximize the loss of class $k$ , regardless of which other class universals have been combined. The second term ensures that $\\pmb{u}_k$ does not change the prediction of classes whose universal vectors have not been used in the composition.", + "bbox": [ + 89, + 90, + 480, + 210 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Remark: A key aspect of Eq. (9) and (10) is that $\\pmb{u}_k + \\pmb{U}\\lambda$ represents points sampled from an affine convex cone. Specifically, these formulations aim to maximize a function over points sampled from $|C|$ affine convex cones (generated by the columns of $\\pmb{U}$ ). Consequently, we can interpret our assumption in (4) as aiming to learn the universal adversarial regions of each class as affine cones, where the vector $\\pmb{u}_{\\Omega}$ , as outlined in (4), lies at the intersection of the affine cones of the classes in $\\Omega$ . This is illustrated in Fig. 2 (right), where $\\pmb{C}_k$ and $\\pmb{C}_i$ represent convex cones for classes $k$ and $i$ respectively, and $\\pmb{C}_{ik}$ denotes their intersection.", + "bbox": [ + 89, + 212, + 482, + 378 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For a comparison among Oracle, $\\mathrm{CMLU}_{\\alpha}$ , and $\\mathrm{CMLU}_{\\beta}$ , Algorithm 1 shows the steps of a single training epoch of Oracle of target size $|\\Omega| = 1$ and the corresponding steps for our methods. Note that in general, Oracle has $\\mathcal{O}(\\binom{|C|}{|\\Omega|})$ training time complexity, which increases with the size of $\\Omega$ . For instance, for $|\\Omega| = 2$ , the complexity of the Oracle's inner loop becomes $\\mathcal{O}(\\binom{|C|}{2})$ , as it needs to handle all possible combinations of classes of a fixed size $|\\Omega|$ . However, $\\mathrm{CMLU}_{\\alpha}$ has $\\mathcal{O}(N_{\\alpha}|\\mathcal{C}|)$ and $\\mathrm{CMLU}_{\\beta}$ has $\\mathcal{O}(|C|)$ complexity and is independent of $|\\Omega|$ . Note that the complexity of our method for learning composable universal perturbations for all possible target set sizes is equivalent to the complexity of Oracle trained for $|\\Omega| = 1$ (linear in the number of classes). Oracle requires iterating over the dataset for each combination of target classes $\\Omega$ , leading to exponential complexity.", + "bbox": [ + 89, + 380, + 483, + 612 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Evaluation", + "text_level": 1, + "bbox": [ + 89, + 625, + 207, + 640 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 89, + 651, + 279, + 667 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets We perform experiments using NUS-WIDE [16], MS-COCO [56], and OpenImages [44] datasets to assess the effectiveness of our proposed method. We compare it with oracle and baseline methods. For NUS-WIDE and MS-COCO, we test target set sizes ranging from 1 to 5 and for OpenImages, we experiment with larger target set sizes, going up to 20. Due to the computational expense of learning oracle model, we choose a subset of labels from each dataset (25 classes from NUS-WIDE, 40 classes from MS-COCO, and 100 classes from OpenImages). Even with the small subset of classes, there can be a large number of possible combinations e.g., there can be more than 53k combinations of $|\\Omega| = 5$ using 25 classes from NUS-WIDE. Therefore, we perform experiments on a subset of those possible combinations for each target set size $|\\Omega|$ . For all of our", + "bbox": [ + 89, + 672, + 482, + 900 + ], + "page_idx": 4 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1: Algorithms of Oracle, $\\mathrm{CMLU}_{\\alpha}$ , $\\mathrm{CMLU}_{\\beta}$" + ], + "code_body": "Input: $\\mathcal{D}$ : Training Data, $\\mathcal{F}$ : Classifier, $\\epsilon$ : Norm budget, $(\\xi ,\\alpha)$ : Step sizes Output: $U$ : Learnt Universal Perturbations 1 Randomly initialize $U$ /\\* $\\mathcal{B} = \\{\\mathcal{B}^{(k)}\\}_{k = 1}^{|C|},\\mathcal{B}^{(k)}$ images of class $k\\ast /$ \n2 Oracle: \n3 for Batch $\\mathcal{B}$ in $\\mathcal{D}$ do /* One iteration through the dataset \\*/ \n4 for $k$ in $\\mathcal{C}$ do /\\* O(|C|) \\*/ \n5 $\\mathcal{L}_k$ : Compute loss in (1) using $\\pmb {u}_k$ \n6 Compute the gradient: $\\pmb {g}\\gets \\nabla_{\\pmb{U}_k}\\pmb {l}_k$ \n7 Update $U_{k}\\leftarrow \\Pi_{\\epsilon}\\left(U_{k} + \\xi *Sgn(\\pmb {g})\\right)$ \n8 for Batch $\\mathcal{B}$ in $\\mathcal{D}$ do \n9 Set $\\pmb{g}$ to be all zeros of the same size as $U$ \n10 for $k$ in $\\mathcal{C}$ do /\\* O(|C|) \\*/ \n11 Randomly sample $t\\sim T$ and $\\lambda \\in [0,1]^{\\lfloor C\\rfloor}$ . \n12 $\\mathrm{CMLU}_{\\alpha}$ ： /\\*O(Nα）\\*/ \n13 for $N_{\\alpha}$ num of iterations do \n14 Set $\\pmb {v}_k = \\pmb {u}_k + U\\pmb{\\lambda}$ \n15 $\\mathcal{L}_k$ : Compute loss in (9) using $\\pmb {v}_k$ \n16 Update: $\\lambda = \\lambda -\\alpha \\nabla_{\\lambda}\\mathcal{L}_k$ \n17 Project $\\lambda$ on the feasible set $\\| \\lambda \\| _0 = t$ \n18 Set $\\lambda_k = 0$ and $\\pmb {v}_k = \\pmb {u}_k + \\pmb {U}\\pmb{\\lambda}$ \n19 $\\mathcal{L}_k$ : Compute loss in (9) using $\\pmb {v}_k$ \n20 CMLUβ:/*O(1)\\*/ \n21 Randomly set $|\\mathcal{C}| - t$ values in $\\lambda$ to 0. \n22 Set $\\lambda_k = 0$ and $\\pmb {v}_k = \\pmb {u}_k + U\\pmb{\\lambda}$ \n23 $\\mathcal{L}_k$ :Compute loss in (10) using $\\pmb {v}_k$ \n24 Accumulate: $\\pmb {g}\\gets \\pmb {g} + \\nabla_{\\pmb{U}_k}\\mathcal{L}_k$ \n25 Update $U\\gets \\Pi_{\\epsilon}(U + \\xi *Sgn(g))$", + "bbox": [ + 516, + 111, + 903, + 577 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "experiments, we learn universal vectors with maximum infinity norm $\\epsilon = 0.05$ (for image values range of [0,1]) and using SGD with fix step size $\\xi = \\epsilon /25$ . We set $\\gamma = \\kappa /|\\bar{\\Omega}|$ for training oracle and baselines, and $\\eta = \\kappa /\\| \\pmb {\\lambda}\\| _0$ in each iteration of our proposed method. We show ablation experiments for different values of $\\kappa$ in section 4.2.1.", + "bbox": [ + 511, + 609, + 903, + 699 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Multi-Label Recognition Models We evaluate the baselines and our method on state-of-the-art multi-label recognition models (using the officially provided weights).", + "bbox": [ + 511, + 709, + 903, + 755 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "- Asymmetric Loss (ASL) [79] is a novel loss function to address the imbalance between numerous negative labels and only a few positive labels associated with each sample. This method balances the probabilities of different positive and negative samples. In particular, the loss performs hard thresholding by dynamically down-weighting easy negative samples and can discard unlabeled samples.", + "bbox": [ + 511, + 758, + 906, + 866 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "- ML-Decoder [80] is an efficient, attention-based multi-label classifier that has a novel decoder architecture and", + "bbox": [ + 511, + 869, + 906, + 898 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "20584", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/825c2522eda98ed053528cb905dad9b4ca06c03d824e395fac4cf19589e2c35e.jpg", + "image_caption": [ + "Figure 3. Fooling success rates $(FR)$ on various datasets using ASL and ML-Decoder as we increase the target set size $|\\Omega|$ . The left two figures show results on NUS-WIDE and the right two figures show results on MS-COCO." + ], + "image_footnote": [], + "bbox": [ + 109, + 93, + 279, + 228 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e27be3d6f28315f9f360620504b2000c774600c9e4e3c28f19cfe925e72c162a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 282, + 93, + 450, + 228 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a185711a2e7d0c80029ff3b27b8acc95486b332fdc030162253977d825a82952.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 452, + 93, + 632, + 228 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/be9818a4d3dc4763d1fe2360fb96d9710839b6dda581b11a8ccce9c6ddd12918.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 632, + 93, + 802, + 228 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/478457689f7413ff080020b02544cf847dc88442da297593ec3dc3a7810d75e1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 802, + 104, + 880, + 228 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "group-decoding scheme, which enables it to efficiently utilize the spatial information in images. The model is highly efficient as it can scale to large number of classes. Moreover, using word queries, it can generalize to unseen classes.", + "bbox": [ + 88, + 282, + 480, + 343 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Oracle and Baselines We evaluate our method against three baselines and an oracle. The computational complexity of each of these methods is reported in Table 1 and their wall clock time in supplementary.", + "bbox": [ + 89, + 348, + 482, + 409 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "- Oracle directly optimizes (1) to learn a universal perturbation $\\pmb{u}_{\\Omega}$ to attack a specific target set $\\Omega$ . This provides the upper bound on the attack performance for our method. It requires $2^{|C|}d$ -dimensional universal vectors and for each vector, iterate through the training set for $E$ epochs.", + "bbox": [ + 89, + 415, + 483, + 492 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Oracle Sum (Or-S) directly learns single-class universal perturbation vectors $\\{\\pmb{u}_1, \\pmb{u}_2, \\dots, \\pmb{u}_{|\\mathcal{C}|}\\}$ using (1) and generates a universal perturbation to attack set $\\Omega$ as $\\pmb{u}_{\\Omega} = \\sum_{i \\in \\Omega} \\pmb{u}_i$ . It only needs to learn $|C|$ $d$ -dimensional vectors.", + "- Oracle Combination (Or-C) first learns single-class universal perturbation vectors and learns to combine them to attack a specific target set $\\Omega$ i.e., $\\pmb{u}_{\\Omega} = \\sum_{i\\in \\Omega}a_i^\\Omega \\pmb {u}_i$ . Where, $a_i^\\Omega \\in \\mathbb{R}$ is associated with $i^{th}$ class in specific set $\\Omega$ , and is learnt by iterating through the training set. It needs to learn $|C|$ $d$ -dim vectors and $2^{|C|}$ scalars to combine them.", + "- SGA [59] addresses the gradient vanishing and poor local optima problem to improve the generalization of UAPs. It enhances the gradient stability by aggregating gradients from inner pre-search. We use SGA to learn single-class UAPs and use (4) to generate UAPs for target labels.", + "- NAG [68] models the distribution of universal adversarial perturbations using GANs, originally designed for untargeted perturbations in multi-class setting. To adapt this for multi-label universal perturbations, we condition the GAN on class vectors. During training, we generate label-wise universal perturbations. At test time, we generate perturbations for multiple classes by conditioning the GAN on a combined representation vector of the target classes. For this method, we have linear time complexity as we train to generate label-wise perturbations and the space requirement is the size of the generator." + ], + "bbox": [ + 89, + 497, + 483, + 912 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metrics Given a set of classes $\\Omega$ to attack, let $\\mathcal{I}_{\\Omega}$ represent all the images that contain the classes $\\Omega$ and were attacked. Let $\\mathcal{A}_{\\Omega} \\subseteq \\mathcal{I}_{\\Omega}$ represent the set of images that were successfully attacked. To evaluate the universal attack performance, we define the attack fooling success rate, $FR$ , to measure the percentage of successfully attacked images. Since we want to affect the least number of non-targeted labels while achieving high fooling rates, we also report non-target flip rate, $NT_{R}$ , to measure the percentage of non-targeted labels (labels in $\\bar{\\Omega}$ ) which were flipped by the attack. Our goal is to achieve high $FR$ while maintaining low $NT_{R}$ for attacking target set of any size.", + "bbox": [ + 511, + 282, + 903, + 464 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {F R} = \\frac {\\left| \\mathcal {A} _ {\\Omega} \\right|}{\\left| \\mathcal {I} _ {\\Omega} \\right|}, \\mathrm {N T} _ {R} = \\frac {1}{\\left| \\mathcal {A} _ {\\Omega} \\right|} \\sum_ {k \\in \\mathcal {A} _ {\\Omega}} \\frac {\\sum_ {i \\in \\bar {\\Omega}} \\left(1 - \\delta \\left(\\chi \\left(f _ {i} ^ {(k)}\\right) , y _ {i} ^ {(k)}\\right)\\right)}{\\left| \\bar {\\Omega} \\right|} \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 529, + 469, + 903, + 520 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where, $\\delta$ is Kronecker delta that equals 1 when the two inputs are equal and 0 otherwise. $\\chi (v)$ is the step function which is 1 if $v > 0$ and 0 otherwise. $y_{i}^{(k)},f_{i}^{(k)}$ are the predictions on clean and perturbed images respectively of $i^{th}$ non-target class of $k^{th}$ successfully attacked image. Note that $FR$ is calculated using all attacked images while $NT_{R}$ is calculated using the set of successfully attacked images.", + "bbox": [ + 511, + 520, + 905, + 628 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Experimental Results", + "text_level": 1, + "bbox": [ + 511, + 637, + 714, + 652 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we evaluate our compositional framework for universal perturbations. Fig. 3 shows the fooling success rates for different methods. The results demonstrate consistent performance trends across datasets and models. Particularly, the success rate(FR) of baselines decline as we increase the target size. On the other hand, the Oracle method achieves the best performance but suffers from exponential complexity, making it impractical for larger target sets.", + "bbox": [ + 511, + 659, + 905, + 779 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Among the baselines, Or-C, Or-S, and SGA perform the worst for large target set sizes. Since the performance of these methods depends on the label-wise universal perturbations, we can conclude that the low success rate of these methods implies that the subspace spanned by labelwise perturbations does not contain the intersection of their respective adversarial regions. This is further elaborated in Fig. 5. In all cases, Or-C is better than Or-S since it op-", + "bbox": [ + 511, + 780, + 906, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "20585", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/199186af8ff04b4342ab6687dd6c89c94dd28b5655b1a089f3576b8692b368c2.jpg", + "table_caption": [ + "Table 1. Non-target flip rate $(NT_R)$ on NUS-WIDE and MS-COCO, as we increase the target set size from 1 to 5. We show (-) when $|\\mathcal{A}_{\\Omega}| = 0$ . We also show the space and time complexity of each method. $E$ represents the number of epochs for training, $d$ is the size of the perturbation (equal to the image size), $G_p$ is the size of GAN's parameters, and $|\\mathcal{C}|$ is the number of classes." + ], + "table_footnote": [], + "table_body": "
Computational ComplexityASLML-Decoder
NUS-WIDEMS-COCONUS-WIDEMS-COCO
MethodSpaceTime135135135135
Oracle\\( \\mathcal{O}(d2^{|C|}) \\)\\( \\mathcal{O}(E2^{|C|}) \\)6.177.428.106.507.738.662.863.103.017.388.589.81
Or-S\\( \\mathcal{O}(d|C|) \\)\\( \\mathcal{O}(E|C|) \\)6.176.03-6.502.872.962.862.22-7.382.96-
Or-C\\( \\mathcal{O}(d|C|+2^{|C|}) \\)\\( \\mathcal{O}(E2^{|C|}) \\)6.176.862.746.505.233.202.862.450.607.386.40-
SGA\\( \\mathcal{O}(d|C|) \\)\\( \\mathcal{O}(E|C|) \\)7.697.31-7.774.30-2.892.471.278.628.949.37
NAG\\( \\mathcal{O}(G_p) \\)\\( \\mathcal{O}(E|C|) \\)8.325.861.3210.110.27.213.963.452.5511.812.113.7
\\( \\mathbf{CMLU}_{\\alpha} \\)\\( \\mathcal{O}(d|C|) \\)\\( \\mathcal{O}(EN_{\\alpha}|C|) \\)5.637.653.747.335.963.133.63.082.518.968.558.02
\\( \\mathbf{CMLU}_{\\beta} \\)\\( \\mathcal{O}(d|C|) \\)\\( \\mathcal{O}(E|C|) \\)7.968.137.938.908.869.594.233.743.029.659.8310.5
", + "bbox": [ + 117, + 138, + 879, + 306 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/b3143b6e2d65e6df1ee6aadaffa70e9ef0892e7bf74eb958cb05b0a2f7b9c923.jpg", + "table_caption": [ + "Table 2. Experiments on large-scale dataset. We report fooling success rate (FR) and non-target flip rate $(\\mathrm{NT}_R)$ on OpenImages using ASL model. The experiments were performed on large target set sizes, up to $|\\Omega| = 20$ . Higher FR ( $\\uparrow$ ) and lower $\\mathrm{NT}_R$ ( $\\downarrow$ ) are better." + ], + "table_footnote": [], + "table_body": "
MethodFR (↑)NTR(↓)
1234510152012345101520
Oracle97.697.996.897.995.796.396.896.80.640.800.870.930.961.051.171.47
Or-S97.629.646.281.700.420.00.00.00.640.710.650.340.20---
Or-C97.649.422.823.313.40.00.00.00.640.710.670.750.48---
SGA98.941.49.646.110.920.150.00.00.770.810.730.740.320.12--
NAG95.992.188.888.489.779.980.984.20.810.890.931.011.071.421.531.84
CMLUα92.266.744.046.239.439.932.432.30.720.760.830.890.830.860.820.77
CMLUβ97.395.895.996.895.294.896.595.20.720.800.880.920.981.061.281.67
", + "bbox": [ + 117, + 354, + 879, + 503 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9e0399c2f83fa5a5d025b5c5d75c33b82aca25c33341ec722e3b9fe14bf0a7ec.jpg", + "image_caption": [ + "Figure 4. Visualizing label-wise universal perturbations learnt by Oracle and $\\mathrm{CMLU}_{\\beta}$ on NUS-WIDE using ASL." + ], + "image_footnote": [], + "bbox": [ + 114, + 527, + 455, + 660 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "timizes to find the best combination of the label-wise universal perturbations and hence, is better than the naive summation of individual perturbations. Across all experiments, $\\mathrm{CMLU}_{\\beta}$ is the closest to Oracle performance. This is despite the fact that $\\mathrm{CMLU}_{\\beta}$ has linear complexity whereas the Oracle has exponential complexity. Moreover, $\\mathrm{CMLU}_{\\alpha}$ achieves higher $FR$ and lower $NT_{R}$ than all baselines.", + "bbox": [ + 89, + 715, + 482, + 821 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 1 shows the percentage of non-targeted affected labels (whose prediction was changed after the attack). Although the baselines have low $\\mathrm{NT}_R$ for target sizes greater than 1, they also have low fooling rates. Note that despite having linear complexity, the $\\mathrm{NT}_R$ of $\\mathrm{CMLU}_{\\beta}$ is only", + "bbox": [ + 89, + 824, + 483, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "slightly higher than Oracle. $\\mathrm{CMLU}_{\\alpha}$ achieves lower $NT_{R}$ as compared to $\\mathrm{CMLU}_{\\beta}$ but still outperforms the baselines. In all experiments, NAG performs well (close to $\\mathrm{CMLU}_{\\alpha}$ ) but declines in performance as we increase the target size. It also has significantly higher $\\mathrm{NT}_{R}$ than the Oracle and $\\mathrm{CMLU}_{\\beta}$ . This implies that NAG learnt to attack multiple classes together rather than learning the distribution of label-wise universal perturbations. Therefore, even though it has high success rate, it also affects several other labels.", + "bbox": [ + 511, + 529, + 906, + 666 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We further perform experiments on OpenImages, a large-scale dataset, and we show results in Tab. 2. Performance trends remain consistent with those observed on smaller datasets. The success rate of Or-S and Or-C consistently drops as we target more classes, and gets to 0 by target size 10. Note that NAG performs better in OpenImages experiments. As shown in the table, NAG affects a large percentage of non-targeted labels, which signifies that the method does not capture the class-specific directions and perturbs a large group of labels together. $\\mathrm{CMLU}_{\\alpha}$ performs better than Or-C, Or-S, and SGA but worse than NAG. However, as shown in the table, $\\mathrm{CMLU}_{\\alpha}$ still achieves significantly lower $NT_{R}$ for large target set sizes. In all experiments, $\\mathrm{CMLU}_{\\beta}$ achieves consistently close performance to the Oracle while also maintaining low $\\mathrm{NT}_{R}$ .", + "bbox": [ + 511, + 674, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "20586", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/bcc18fa01204e1d900ed1f8b6bcc99ca0ee45a595d6dc205c9f038e696f1a4b7.jpg", + "image_caption": [ + "Figure 5. Comparison of the regions (convex cones) spanned by universal perturbations learnt using Oracle and $\\mathrm{CMLU}_{\\beta}$ on NUS-WIDE (top) and MS-COCO (bottom). Each point in the heatmap is the average model confidence computed across images using the vector $(\\epsilon_x * \\boldsymbol{u}_x + \\epsilon_y * \\boldsymbol{u}_y)$ for all label combinations of size $|\\Omega| = 2$ , where $\\boldsymbol{u}_x, \\boldsymbol{u}_y$ are universal vectors of classes in $\\Omega$ . The asterisks(*) show the region where the $L_{\\infty}$ norm is 0.05." + ], + "image_footnote": [], + "bbox": [ + 109, + 90, + 460, + 253 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.2.1. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 373, + 243, + 388 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We perform ablation experiments on Oracle and $\\mathrm{CMLU}_{\\beta}$ to investigate the effect of hyperparameters on their performance. Tab. 3 shows the results as we increase/decrease the values of $\\epsilon$ and $\\kappa$ . Increasing $\\epsilon$ provides a larger perturbation budget for better performance; however, to keep the changes to the inputs imperceptible, we set $\\epsilon = 0.05$ . Meanwhile, increasing $\\kappa$ emphasizes maintaining non-targeted labels, which lowers $\\mathrm{NT}_R$ but also reduces the success rate (FR). This is because many labels can be correlated in MLL and attacking a set of labels while fixing non-targeted but correlated labels lead to sub-optimal perturbations [63].", + "bbox": [ + 89, + 391, + 483, + 559 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.2.2. Analysis of Universal Adversarial Regions", + "text_level": 1, + "bbox": [ + 89, + 564, + 431, + 580 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The label-wise universal perturbations learnt by Oracle and $\\mathrm{CMLU}_{\\beta}$ are visualized in Fig. 4. We quantitatively assess the visual impact of these perturbations using Structural Similarity Index (SSIM), computed between original and perturbed images. The average SSIM values across all images, as shown in Fig. 4, demonstrate comparable levels of distortion for Oracle and $\\mathrm{CMLU}_{\\beta}$ . Note that the pairwise dot product of the Oracle and CMLU vectors is $\\approx 0$ , indicating that $\\mathrm{CMLU}_{\\beta}$ discover distinct attack directions, different from the Oracle. Additional analysis of pairwise dot products and visualizations are provided in supplementary.", + "bbox": [ + 89, + 583, + 482, + 750 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To further highlight and analyze the distinction between the perturbations learnt using different methods, we focus on Oracle and $\\mathrm{CMLU}_{\\beta}$ . Specifically, we examine the universal adversarial regions generated by those methods. Fig. 5 shows the average model confidence on the input space region spanned by the pairs of universal perturbation vectors. In the figure, perturbing inputs to blue and green regions would not change the model prediction (does not attack any class), yellow region would attack one class, and red region would attack both classes in $\\Omega$ . Note that there is", + "bbox": [ + 89, + 750, + 482, + 901 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/c125b1985b49fb4f6f7314431d48f5a52ef90f60336c7cd3fee682aa74630f81.jpg", + "table_caption": [ + "Table 3. Ablation experiments for hyperparameters on NUS-WIDE. The experiments are conducted for $|\\Omega| = 1$ ." + ], + "table_footnote": [], + "table_body": "
εOracleCMLUβκOracleCMLUβ
FRNTRFRNTRFRNTRFRNTR
0.0163.55.3038.55.041599.16.5098.58.32
0.0394.96.2984.27.092098.96.1797.67.96
0.0598.96.1797.67.962593.54.7089.36.97
0.0799.75.1099.37.543093.14.4085.56.58
", + "bbox": [ + 527, + 125, + 888, + 210 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/934f82742f8f9a63831ff81b2ac6f4b3179bb5b5f3f5cef9e9849003c327c5b0.jpg", + "table_caption": [ + "Table 4. Transferability Evaluation: $S \\rightarrow T$ shows the performance of universal perturbations learnt on model $\\mathbf{S}$ and evaluated on model $\\mathbf{T}$ . (A: ASL, D: ML-Decoder)" + ], + "table_footnote": [], + "table_body": "
Transferability of Universal Attacks
ModelsMethodNUS-WIDEMS-COCO
135135
A → DOracle90.183.777.334.610.65.40
CMLUβ89.978.475.442.917.66.4
D → AOracle24.61.330.064.428.319.6
CMLUβ33.50.160.064.834.521.2
", + "bbox": [ + 527, + 263, + 888, + 382 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "no prominent red region (to attack both classes) in the cross-section of Oracle vectors. This explains the low success rate of Or-S and Or-C which optimize to find red region in such a cross-section. More details on computing these cones is provided in supplementary. We also investigate how well those red regions computed using $\\mathrm{CMLU}_{\\beta}$ can transfer to other models. For that, we perform transferability attacks and the results are shown in Tab. 4. From the table, we observe that $\\mathrm{CMLU}_{\\beta}$ (linear complexity) achieves fooling success rates closer and often better than Oracle (exponential complexity). This shows that the convex cone regions corresponding to different labels are often common across models. This can be effectively used to generate black-box attacks. We hope that future works explore this further.", + "bbox": [ + 511, + 393, + 906, + 604 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 618, + 633, + 633 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We addressed the problem of generating targeted multi-label universal perturbations. We developed a compositional framework to generate exponential number of multi-label universal perturbations in constant time. The independence assumption on label-wise universal perturbations naturally leads to a formulation that requires learning convex cones. Through extensive experiments on multiple datasets and models, we show that our method outperforms all baseline methods and achieves performance comparable to the Oracle (an exponential complexity method).", + "bbox": [ + 511, + 642, + 906, + 796 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Acknowledgements", + "text_level": 1, + "bbox": [ + 511, + 808, + 702, + 825 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work is sponsored by ARPA-H (1AY2AX000062), DARPA PTG (HR00112220001), NSF (IIS-2115110) and ARO (W911NF2110276). Content does not necessarily reflect the position/policy of the Government.", + "bbox": [ + 511, + 832, + 906, + 895 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "20587", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Abhishek Aich, Calvin-Khang Ta, Akash Gupta, Chengyu Song, Srikanth Krishnamurthy, Salman Asif, and Amit Roy-Chowdhury. Gama: Generative adversarial multi-object scene attacks. Advances in Neural Information Processing Systems, 35:36914-36930, 2022. 1, 3", + "[2] Abhishek Aich, Shasha Li, Chengyu Song, M Salman Asif, Srikanth V Krishnamurthy, and Amit K Roy-Chowdhury. Leveraging local patch differences in multi-object scenes for generative adversarial attacks. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1308-1318, 2023. 1", + "[3] N. Akhtar and A. Mian. Threat of adversarial attacks on deep learning in computer vision: A survey. arXiv, 2018. 2", + "[4] A. Athalye, N. Carlini, and D. A. Wagner. Obfuscated gradients give a false sense of security: Circumventing defenses to adversarial examples. 2018. 2", + "[5] Yuanhao Ban and Yinpeng Dong. Pre-trained adversarial perturbations. In Advances in Neural Information Processing Systems, pages 1196-1209, 2022. 2", + "[6] Emanuel Ben-Baruch, Tal Ridnik, Itamar Friedman, Avi Ben-Cohen, Nadav Zamir, Asaf Noy, and Lihi Zelnik-Manor. Multi-label classification with partial annotations using class-aware selective loss. 2022. 2", + "[7] P. Benz, C. Zhang, T. Imtiaz, and I.-S. Kweon. Double targeted universal adversarial perturbations. Asian Conference on Computer Vision, 2020. 1, 2", + "[8] Philipp Benz, Chaoning Zhang, Adil Karjauv, and In So Kweon. Universal adversarial training with class-wise perturbations. In 2021 IEEE International Conference on Multimedia and Expo (ICME), pages 1-6. IEEE, 2021. 2", + "[9] N. Carlini and D. Wagner. Adversarial examples are not easily detected: Bypassing ten detection methods. Workshop on Artificial Intelligence and Security, 2017. 1", + "[10] N. Carlini and D. Wagner. Towards evaluating the robustness of neural networks. IEEE Symposium on Security and Privacy, 2017. 2", + "[11] Y. Carmon, A. Raghunathan, L. Schmidt, P. Liang, and J. C. Duchi. Unlabeled data improves adversarial robustness. Neural Information Processing Systems, 2019. 2", + "[12] P.-Y. Chen, Y. Sharma, H. Zhang, J. Yi, and C.-J. Hsieh. Ead: Elastic-net attacks to deep neural networks via adversarial examples. AAAI Conference on Artificial Intelligence, 2018. 2", + "[13] T. Chen, M. Xu, X. Hui, H. Wu, and L. Lin. Learning semantic-specific graph representation for multi-label image recognition. IEEE International Conference on Computer Vision, 2019. 2", + "[14] Zhao-Min Chen, Xiu-Shen Wei, Xin Jin, and Yanwen Guo. Multi-label image recognition with joint class-aware map disentangling and label correlation embedding. IEEE International Conference on Multimedia and Expo, 2019. 1", + "[15] Z. M. Chen, X. S. Wei, P. Wang, and Y. Guo. Multi-label image recognition with graph convolutional networks. IEEE Conference on Computer Vision and Pattern Recognition, abs/1904.03582, 2019. 1" + ], + "bbox": [ + 101, + 114, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[16] T. S. Chua, J. Tang, R. Hong, H. Li, Z. Luo, and Y. T. Zheng. Nus-wide: A real-world web image database from national university ofsingapore.ACM International Conference on Image and Video Retrieval, 2009.5", + "[17] S. D. Dao, E. Zhao, D. Phung, and J. Cai. Multi-label image classification with contrastive learning. arXiv preprint, arXiv:2107.11626, 2021. 2", + "[18] J. Deng, N. Ding, Y. Jia, A. Frome, K. Murphy, S. Bengio, Y. Li, H. Neven, and H. Adam. Large-scale object classification using label relation graphs. European Conference on Computer Vision, 2014. 2", + "[19] Gavin Weiguang Ding, Yash Sharma, Kry Yik Chau Lui, and Ruitong Huang. Mma training: Direct input space margin maximization through adversarial training. arXiv preprint arXiv:1812.02637, 2018. 2", + "[20] Zixuan Ding, Ao Wang, Hui Chen, Qiang Zhang, Pengzhang Liu, Yongjun Bao, Weipeng Yan, and Jungong Han. Exploring structured semantic prior for multi label recognition with incomplete labels. 2023. 1", + "[21] Junhao Dong, Seyed-Mohsen Moosavi-Dezfooli, Jianhuang Lai, and Xiaohua Xie. The enemy of my enemy is my friend: Exploring inverse adversaries for improving adversarial training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 24678–24687, 2023. 2", + "[22] Yinpeng Dong, Zhijie Deng, Tianyu Pang, Jun Zhu, and Hang Su. Adversarial distributional training for robust deep learning. Advances in Neural Information Processing Systems, 33:8270-8283, 2020. 2", + "[23] L. Feng, B. An, and S. He. Collaboration based multi-label learning. AAAI Conference on Artificial Intelligence, 2019. 1", + "[24] Zongyuan Ge, Dwarikanath Mahapatra, Suman Sedai, Rahul Garnavi, and Rajib Chakravorty. Chest x-rays classification: A multi-label and fine-grained problem. arXiv, 2018. 1", + "[25] I. J. Goodfellow, J. Shlens, and C. Szegedy. Explaining and harnessing adversarial examples. International Conference on Learning Representations, 2015. 1", + "[26] Emily Hand, Carlos Castillo, and Rama Chellappa. Doing the best we can with what we have: Multi-label balancing with selective learning for attribute prediction. In Proceedings of the AAAI Conference on Artificial Intelligence, 2018. 1", + "[27] J. Hayes and G. Danezis. Learning universal adversarial perturbations with generative models. IEEE Security and Privacy Workshops, 2018. 2", + "[28] W. He, J. Wei, X. Chen, N. Carlini, and D. Song. Adversarial example defense: Ensembles of weak defenses are not strong. USENIX Workshop on Offensive Technologies, 2017. 2", + "[29] D. Hendrycks, K. Lee, and M. Mazeika. Using pre-training can improve model robustness and uncertainty. 2019. 2", + "[30] Dan Hendrycks, Kevin Zhao, Steven Basart, Jacob Steinhardt, and Dawn Song. Natural adversarial examples. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15262-15271, 2021. 1" + ], + "bbox": [ + 522, + 92, + 903, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "20588", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[31] Lei Hsiung, Yun-Yun Tsai, Pin-Yu Chen, and Tsung-Yi Ho. Towards compositional adversarial robustness: Generalizing adversarial training to composite semantic perturbations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24658-24667, 2023. 1", + "[32] Shu Hu, Lipeng Ke, Xin Wang, and Siwei Lyu. Tkmlap: Adversarial attacks to top-k multi-label learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7649-7657, 2021. 3", + "[33] D. Huynh and E. Elhamifar. Interaction compass: Multi-label zero-shot learning of human-object interactions via spatial relations. International Conference on Computer Vision, 2021. 1", + "[34] D. T. Huynh and E. Elhamifar. Interactive multi-label cnn learning with partial labels. IEEE Conference on Computer Vision and Pattern Recognition, 2020. 2", + "[35] Tooba Imtiaz, Morgan Kohler, Jared Miller, Zifeng Wang, Mario Sznaier, Octavia I Camps, and Jennifer G Dy. Saif: Sparse adversarial and interpretable attack framework. arXiv preprint arXiv:2212.07495, 2022. 2", + "[36] J. Li R. Ji, H. Liu, X. Hong, Y. Gao, and Q. Tian. Universal perturbation attack against image retrieval. International Conference on Computer Vision, 2019. 1", + "[37] Jinyuan Jia, Wenjie Qu, and Neil Zhenqiang Gong. Multiguard: Provably robust multi-label classification against adversarial examples. Advances in Neural Information Processing Systems, 2022. 3", + "[38] V. Khrulkov and I. Oseledets. Art of singular vectors and universal adversarial perturbations. IEEE Conference on Computer Vision and Pattern Recognition, 2018. 1, 2", + "[39] Youngwook Kim, Jae Myung Kim, Zeynep Akata, and Jungwoo Lee. Large loss matters in weakly supervised multi-label classification. 2022. 1", + "[40] Takumi Kobayashi. Two-way multi-label loss. 2023. 1", + "[41] Linghao Kong, Wenjian Luo, Hongwei Zhang, Yang Liu, and Yuhui Shi. Evolutionary multilabel adversarial examples: An effective black-box attack. 2023. 3", + "[42] Linghao Kong, Wenjian Luo, Zipeng Ye, Qi Zhou, and Yan Jia. Multi-label black-box adversarial attacks only with predicted labels. 2024. 1", + "[43] A. Kurakin, I. Goodfellow, and S. Bengio. Adversarial machine learning at scale. International Conference on Learning Representations, 2017. 1", + "[44] A. Kuznetsova, H. Rom, N. Alldrin, J. Uijlings, I. Krasin, J. Pont-Tuset, S. Kamali, S. Popov, M. Malloci, A. Kolesnikov, T. Duerig, and V. Ferrari. The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale. International Journal of Computer Vision, 2016. 5", + "[45] J. Lanchantin, T. Wang, V. Ordonez, and Y. Qi. General multi-label image classification with transformers. IEEE Conference on Computer Vision and Pattern Recognition, 2021. 2", + "[46] K. Lee, K. Lee, H. Lee, and J. Shin. A simple unified framework for detecting out-of-distribution samples and adversarial attacks. 2018. 2" + ], + "bbox": [ + 99, + 90, + 485, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[47] Guofa Li, Zefeng Ji, Yunlong Chang, Shen Li, Xingda Qu, and Dongpu Cao. ML-anet: A transfer learning approach using adaptation network for multi-label image classification in autonomous driving. Chinese Journal of Mechanical Engineering, 2021. 1", + "[48] Maosen Li, Yanhua Yang, Kun Wei, Xu Yang, and Heng Huang. Learning universal adversarial perturbation by adversarial example. AAAI Conference on Artificial Intelligence, 2022. 3", + "[49] Peng Li, Peng Chen, Yonghong Xie, and Dezheng Zhang. Bi-modal learning with channel-wise attention for multi-label image classification. IEEE Access, 2020. 2", + "[50] Q. Li, M. Qiao, W. Bian, and D. Tao. Conditional graphical lasso for multi-label image classification. IEEE Conference on Computer Vision and Pattern Recognition, 2016. 2", + "[51] Q. Li, X. Peng, Y. Qiao, and Q. Peng. Learning label correlations for multi-label image recognition with graph networks. Pattern Recognition Letters, 2020. 2", + "[52] X. Li, F. Zhao, and Y. Guo. Multi-label image classification with a probabilistic label enhancement model. In UAI, 2014. 2", + "[53] Y. Li, Y. Song, and J. Luo. Improving pairwise ranking for multi-label image classification. IEEE Conference on Computer Vision and Pattern Recognition, 2017. 2", + "[54] Z. Li, W. Lu, Z. Sun, and W. Xing. Improving multi-label classification using scene cues. Multimedia Tools and Applications, 2017. 2", + "[55] Dekun Lin. Probability guided loss for long-tailed multi-label image classification. 2023. 1", + "[56] Tsung-Yi Lin, Michael Maire, Serge J. Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C. Lawrence Zitnick. Microsoft coco: Common objects in context. European Conference on Computer Vision, 2014. 5", + "[57] An-An Liu, Zhuang Shao, Yongkang Wong, Junnan Li, Su Yu-Ting, and Mohan Kankanhalli. Lstm-based multi-label video event detection. Multimedia Tools and Applications, 78, 2019. 1", + "[58] Shilong Liu, Lei Zhang, Xiao Yang, Hang Su, and Jun Zhu. Query2label: A simple transformer way to multi-label classification. arXiv preprint arXiv:2107.10834, 2021. 2", + "[59] Xuannan Liu, Yaoyao Zhong, Yuhang Zhang, Lixiong Qin, and Weihong Deng. Enhancing generalization of universal adversarial perturbation through gradient aggregation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 2, 6", + "[60] Yiran Liu, Xin Feng, Yunlong Wang, Wu Yang, and Di Ming. Trm-uap: Enhancing the transferability of data-free universal adversarial perturbation via truncated ratio maximization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4762-4771, 2023. 2", + "[61] Mingzhi Ma, Weijie Zheng, Wanli Lv, Lu Ren, Hang Su, and Zhaoxia Yin. Multi-label adversarial attack based on label correlation. In 2023 IEEE International Conference on Image Processing (ICIP), 2023. 3" + ], + "bbox": [ + 522, + 92, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "20589", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[62] A. Madry, A. Makelov, L. Schmidt, D. Tsipras, and A. Vladu. Towards deep learning models resistant to adversarial attacks. International Conference on Learning Representations, 2018. 2", + "[63] Hassan Mahmood and Ehsan Elhamifar. Semantic-aware multi-label adversarial attacks. IEEE Conference on Computer Vision and Pattern Recognition, pages 24251-24262, 2024. 1, 3, 8", + "[64] S. Melacci, G. Ciravegna, A. Sotgiu, A. Demontis, B. Biggio, M. Gori, and F. Roli. Domain knowledge alleviates adversarial attacks in multi-label classifiers. 2021. 3", + "[65] J.-H. Metzen, M.-C. Kumar, T. Brox, and V. Fischer. Universal adversarial perturbations against semantic image segmentation. International Conference on Computer Vision, 2019. 1, 2", + "[66] S. Moosavi-Dezfooli, A. Fawzi, and P. Frossard. Deepfool: a simple and accurate method to fool deep neural networks. IEEE Conference on Computer Vision and Pattern Recognition, 2016. 2", + "[67] S. Moosavi-Dezfooli, A. Fawzi, O. Fawzi, and P. Frossard. Universal adversarial perturbations. IEEE Conference on Computer Vision and Pattern Recognition, 2017. 1, 2", + "[68] Konda Reddy Mopuri, Utsav Garg, and R Venkatesh Babu. Fast feature fool: A data independent approach to universal adversarial perturbations. arXiv preprint arXiv:1707.05572, 2017. 6", + "[69] Konda Reddy Mopuri, Utkarsh Ojha, Utsav Garg, and R Venkatesh Babu. Nag: Network for adversary generation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 742-751, 2018. 1, 2", + "[70] J. Nam, E. L. Mencia, H. J. Kim, and J. Furnkranz. Maximizing subset accuracy with recurrent neural networks in multi-label classification. Neural Information Processing Systems, 2017. 2", + "[71] T. Pang, K. Xu, C. Du, N. Chen, and J. Zhu. Improving adversarial robustness via promoting ensemble diversity. International Conference on Machine learning, 2019. 2", + "[72] T. Pang, K. Xu, Y. Dong, C. Du, N. Chen, and J. Zhu. Rethinking softmax cross-entropy loss for adversarial robustness. arXiv, 2020.", + "[73] T. Pang, X. Yang, Y. Dong, K. Xu, H. Su, and J. Zhu. Boosting adversarial training with hypersphere embedding. Neural Information Processing Systems, 2020. 2", + "[74] Nicolas Papernot, Patrick Mcdaniel, Somesh Jha, Matt Fredrikson, Z. Berkay Celik, and Ananthram Swami. The limitations of deep learning in adversarial settings. IEEE European Symposium on Security and Privacy (EuroS&P), 2016. 1", + "[75] Ioannis Partalas, Aris Kosmopoulos, Nicolas Baskiotis, Thierry Artieres, George Paliouras, Eric Gaussier, Ion Androutsopoulos, Massih-Reza Amini, and Patrick Galinari. Lsjtc: A benchmark for large-scale text classification. arXiv preprint arXiv:1503.08581, 2015. 1", + "[76] Omid Poursaeed, Isay Katsman, Bicheng Gao, and Serge Belongie. Generative adversarial perturbations. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4422-4431, 2018. 2" + ], + "bbox": [ + 99, + 90, + 485, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[77] Tao Pu, Tianshui Chen, Hefeng Wu, and Liang Lin. Semantic-aware representation blending for multi-label image recognition with partial labels. Proceedings of the AAAI Conference on Artificial Intelligence, 2022. 1", + "[78] Zeyu Qin, Yanbo Fan, Yi Liu, Li Shen, Yong Zhang, Jue Wang, and Baoyuan Wu. Boosting the transferability of adversarial attacks with reverse adversarial perturbation. Advances in Neural Information Processing Systems, 35: 29845-29858, 2022. 2", + "[79] Tal Ridnik, Emanuel Ben-Baruch, Nadav Zamir, Asaf Noy, Itamar Friedman, Matan Protter, and Lihi Zelnik-Manor. Asymmetric loss for multi-label classification. In Proceedings of the IEEE/CVF international conference on computer vision, pages 82-91, 2021. 5", + "[80] Tal Ridnik, Gilad Sharir, Avi Ben-Cohen, Emanuel Ben-Baruch, and Asaf Noy. ML-decoder: Scalable and versatile classification head. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 32-41, 2023. 5", + "[81] Jérôme Rony, Luiz G Hafemann, Luiz S Oliveira, Ismail Ben Ayed, Robert Sabourin, and Eric Granger. Decoupling direction and norm for efficient gradient-based 12 adversarial attacks and defenses. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4322-4330, 2019. 2", + "[82] J. Cohen and E. Rosenfeld and Z. Kolter. Certified adversarial robustness via randomized smoothing. International Conference on Machine learning, 2019. 2", + "[83] A. Shafahi, M. Najibi, A. Ghiasi, Z. Xu, J. Dickerson, C. Studer, L. Davis, G. Taylor, and T. Goldstein. Adversarial training for free! Neural Information Processing Systems, 2019.", + "[84] Nasim Shafiee and Ehsan Elhamifar. Zero-shot attribute attacks on fine-grained recognition models. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part V, pages 262-282. Springer, 2022. 1, 2", + "[85] Jing Shao, Kai Kang, Chen Change Loy, and Xiaogang Wang. Deeply learned attributes for crowded scene understanding. In 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2015. 1", + "[86] Nitish Shukla and Sudipta Banerjee. Generating adversarial attacks in the latent space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 730-739, 2023. 1", + "[87] Q. Song, H. Jin, X. Huang, and X. Hu. Multi-label adversarial perturbations. IEEE International Conference on Data Mining, 2018. 3", + "[88] Fengguang Su, Ou Wu, and Weiyao Zhu. Multi-label adversarial attack with new measures and self-paced constraint weighting. 2024. 1", + "[89] C. Szegedy, W. Zaremba, I. Sutskever, J. Bruna, D. Erhan, I. Goodfellow, and R. Fergus. Intriguing properties of neural networks. International Conference on Learning Representations, 2014. 2", + "[90] G. Tsoumakas and I. Katakis. Multi-label classification: An overview. International Journal Data Warehousing and Mining, 3, 2007. 2" + ], + "bbox": [ + 522, + 92, + 905, + 900 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "20590", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[91] N. Tursynbek, A. Petiushko, and I. Oseledets. Geometry-inspired top-k adversarial perturbations. arXiv, 2020. 2", + "[92] J. Uesato, J. B. Alayrac, P. Huang, R. Stanforth, A. Fawzi, and P. Kohli. Are labels required for improving adversarial robustness? Neural Information Processing Systems, 2019. 2", + "[93] Thomas Verelst, Paul K Rubenstein, Marcin Eichner, Tinne Tuytelaars, and Maxim Berman. Spatial consistency loss for training multi-label classifiers from single-label annotations. 2023. 2", + "[94] J. Wang, Y. Yang, J. Mao, Z. Huang, C. Huang, and W. Xu. Cnn-rnn: A unified framework for multi-label image classification. 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 2", + "[95] Ya Wang, Dongliang He, Fu Li, Xiang Long, Zhichao Zhou, Jinwen Ma, and Shilei Wen. Multi-label classification with label graph superimposing. In Proceedings of the AAAI conference on artificial intelligence, pages 12265-12272, 2020.", + "[96] Z. Wang, T. Chen, G. Li, G. Li, and L. Lin. Multi-label image recognition by recurrently discovering attentional regions. IEEE International Conference on Computer Vision, 2017. 2", + "[97] Yunchao Wei, Wei Xia, Min Lin, Junshi Huang, Bingbing Ni, Jian Dong, Yao Zhao, and Shuicheng Yan. Hcp: A flexible cnn framework for multi-label image classification. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2016. 2", + "[98] Y. Wu, H. Liu, S. Feng, Y. Jin, G. Lyu, and Z. Wu. Gmmlic: Graph matching based multi-label image classification. International Joint Conference on Artificial Intelligence, 2021. 2", + "[99] C. Xie, Z. Zhang, Y. Zhou, S. Bai, J. Wang, Z. Ren, and A. Yuille. Improving transferability of adversarial examples with input diversity. IEEE Conference on Computer Vision and Pattern Recognition, 2019. 2", + "[100] Ming-Kun Xie, Jiahao Xiao, and Sheng-Jun Huang. Label-aware global consistency for multi-label learning with single positive labels. Advances in Neural Information Processing Systems, 35:18430-18441, 2022. 1", + "[101] J. Xu, H. Tian, Z. Wang, Y. Wang, W. Kang, and F. Chen. Joint input and output space learning for multi-label image classification. IEEE Transactions on Multimedia, 2020. 2", + "[102] W. Xu, D. Evans, and Y. Qi. Feature squeezing: Detecting adversarial examples in deep neural networks. Network and Distributed Systems Security Symposium, 2018. 2", + "[103] Zhuo Yang, Yufei Han, and Xiangliang Zhang. Characterizing the evasion attackability of multi-label classifiers. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 10647-10655, 2021. 3", + "[104] Zhuo Yang, Yufei Han, and Xiangliang Zhang. Attack transferability characterization for adversarially robust multi-label classification. In Machine Learning and Knowledge Discovery in Databases., pages 397-413, 2021. 3", + "[105] V. O. Yazici, A. Gonzalez-Garcia, A. Ramisa, B. Twardowski, and J. v. d. Weijer. Orderless recurrent models for multi-label classification. European Conference on Computer Vision, 2020. 2" + ], + "bbox": [ + 93, + 92, + 482, + 900 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[106] J. Ye, J. He, X. Peng, W. Wu, and Y. Qiao. Attention-driven dynamic graph convolutional network for multi-label image recognition. European Conference on Computer Vision, 2020.", + "[107] R. You, Z. Guo, L. Cui, X. Long, S. Y. Bao, and S. Wen. Cross-modality attention with semantic graph embedding for multi-label classification. AAAI Conference on Artificial Intelligence, 2020. 2", + "[108] Hong-Xing Yu, Wei-Shi Zheng, Ancong Wu, Xiaowei Guo, Shaogang Gong, and Jian-Huang Lai. Unsupervised person re-identification by soft multilabel learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2148–2157, 2019. 1", + "[109] Chaoning Zhang, Philipp Benz, Tooba Imtiaz, and In-So Kweon. Cd-uap: Class discriminative universal adversarial perturbation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 6754–6761, 2020. 2", + "[110] Chaoning Zhang, Philipp Benz, Tooba Imtiaz, and In So Kweon. Understanding adversarial examples from the mutual influence of images and perturbations. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 1, 2", + "[111] Chaoning Zhang, Philipp Benz, Adil Karjauv, and In So Kweon. Universal adversarial perturbations through the lens of deep steganography: Towards a fourier perspective. In Proceedings of the AAAI conference on artificial intelligence, 2021. 2", + "[112] ML. Zhang and Z. Zhou. Multilabel neural networks with applications to functional genomics and text categorization. IEEE Transactions on Knowledge and Data Engineering, 2006. 2", + "[113] Shu Zhang, Ran Xu, Caiming Xiong, and Chetan Ramaiah. Use all the labels: A hierarchical multi-label contrastive learning framework. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 2", + "[114] Zhe Zhao, Guangke Chen, Jingyi Wang, Yiwei Yang, Fu Song, and Jun Sun. Attack as defense: characterizing adversarial examples using robustness. In Proceedings of the 30th ACM SIGSOFT International Symposium on Software Testing and Analysis, page 42-55. Association for Computing Machinery, 2021. 2", + "[115] Donghao Zhou, Pengfei Chen, Qiong Wang, Guangyong Chen, and Pheng-Ann Heng. Acknowledging the unknown for multi-label learning with single positive labels. European Conference on Computer Vision, 2022. 1", + "[116] N. Zhou, W. Luo, X. Lin, P. Xu, and Z.. Zhang. Generating multi-label adversarial examples by linear programming. International Joint Conference on Neural Networks, 2020. 3", + "[117] N. Zhou, W. Luo, J. Zhang, L. Kong, and H. Zhang. Hiding all labels for multi-label images: An empirical study of adversarial examples. International Joint Conference on Neural Networks, 2021. 3", + "[118] Y. Zhu, J. T. Kwok, and Z. Zhou. Multi-label learning with global and local label correlation. IEEE Transactions on Knowledge and Data Engineering, 2018. 2" + ], + "bbox": [ + 516, + 92, + 903, + 897 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "20591", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2025/Compositional Targeted Multi-Label Universal Perturbations/616cefba-8863-49bf-987c-2e525bd606b6_model.json b/2025/Compositional Targeted Multi-Label Universal Perturbations/616cefba-8863-49bf-987c-2e525bd606b6_model.json new file mode 100644 index 0000000000000000000000000000000000000000..1fd027f1a0db04bd7197b9a281dc9b9c656d9d87 --- /dev/null +++ b/2025/Compositional Targeted Multi-Label Universal Perturbations/616cefba-8863-49bf-987c-2e525bd606b6_model.json @@ -0,0 +1,3062 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.19, + 0.131, + 0.812, + 0.154 + ], + "angle": 0, + "content": "Compositional Targeted Multi-Label Universal Perturbations" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.182, + 0.438, + 0.216 + ], + "angle": 0, + "content": "Hassan Mahmood Northeastern University" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.22, + 0.46, + 0.233 + ], + "angle": 0, + "content": "mahmood.h@northeastern.edu" + }, + { + "type": "text", + "bbox": [ + 0.516, + 0.182, + 0.766, + 0.234 + ], + "angle": 0, + "content": "Ehsan Elhamifar Northeastern University e.elhamifar@northeastern.edu" + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.269, + 0.327, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.301, + 0.486, + 0.589 + ], + "angle": 0, + "content": "Generating targeted universal perturbations for multi-label recognition is a combinatorially hard problem that requires exponential time and space complexity. To address the problem, we propose a compositional framework. We show that a simple independence assumption on label-wise universal perturbations naturally leads to an efficient optimization that requires learning affine convex cones spanned by label-wise universal perturbations, significantly reducing the problem complexity to linear time and space. During inference, the framework allows generating universal perturbations for novel combinations of classes in constant time. We demonstrate the scalability of our method on large datasets and target sizes, evaluating its performance on NUS-WIDE, MS-COCO, and OpenImages using state-of-the-art multi-label recognition models. Our results show that our approach outperforms baselines and achieves results comparable to methods with exponential complexity. The code is available at https://github.com/hassanmahmood/UMLLAttacks.git" + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.604, + 0.222, + 0.619 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.629, + 0.483, + 0.78 + ], + "angle": 0, + "content": "Deep neural networks have been very successful at image-recognition, demonstrating high accuracy in various applications. However, they are also vulnerable to adversarial attacks [9, 25, 30, 31, 43, 74, 84, 86], where small changes to the input image can cause the model to misclassify the image. Notably, Moosavi-Dezfooli et al. [67] discovered that a single perturbation, called universal adversarial perturbation (UAP), can be crafted to misclassify a large number of images. This motivated a large body of work on generating universal adversarial attacks [7, 36, 38, 65, 69, 110]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Existing works on adversarial attacks have mainly focused on multi-class recognition (MCR) whose goal is to identify a single class in an image. On the other hand, many real-world applications require identifying all classes in an image e.g., document classification [75], scene understanding [85], facial attribute recognition [26], image annotation [15], human-object interaction [33], and several safety and security-critical applications e.g., recogniz-" + }, + { + "type": "image", + "bbox": [ + 0.54, + 0.283, + 0.878, + 0.495 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.502, + 0.907, + 0.572 + ], + "angle": 0, + "content": "Figure 1. Targeted Multi-Label Universal Perturbations. Our framework allows composing (shown as \\(\\oplus\\)) different label-specific universal perturbations ('Animal', 'Person', 'Vehicle') to attack multiple classes together. These perturbations, once added (shown as \\(\\bullet\\)) to an image, cause target labels to be misclassified." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.598, + 0.907, + 0.749 + ], + "angle": 0, + "content": "ing diseases [24], autonomous driving [47], and surveillance [57, 108]. Therefore, the goal of multi-label learning (MLL) is to identify all labels present in an image [14, 20, 23, 39, 40, 55, 77, 100, 115]. Given the increasing number of applications relying on MLL, it is crucial to understand the internal workings of the MLL models, identify their vulnerabilities, and make them robust. Recently, MLL models have been shown to be vulnerable to adversarial attacks [1, 2, 42, 63, 88], yet the vulnerability of MLL systems against universal perturbations has not been explored." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Crafting universal perturbations for MCR and MLL have fundamental differences due to the distinct structuring of decision boundaries in these problems. In MCR, a universal perturbation to turn off a label in images will result in the prediction of another label. Conversely, in MLL, multiple labels can co-exist, and the objective of an attack might be to misclassify any subset of these labels either by turning the labels off or on (see Fig. 2). In MLL, any combination of labels can be present in images. For the label set \\(\\mathcal{C}\\), there are \\(2^{|C|}\\) possible combinations of labels. Universal pertur" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20580" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.089, + 0.092, + 0.482, + 0.182 + ], + "angle": 0, + "content": "bations identify specific directions in the input space where the classifier is consistently vulnerable to attacks. What makes this particularly concerning in multi-label learning is that there can be exponentially many directions in the input space to attack the model - corresponding to different combinations of labels." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.183, + 0.482, + 0.318 + ], + "angle": 0, + "content": "One approach for generating universal perturbations for MLL is to generate universal vectors that flip the prediction of all or most labels. However, in practice, the adversary usually seeks to modify a specific small set of labels while keeping the prediction of others unchanged (see Fig. 1). This is called the targeted attack and requires learning \\(2^{|\\mathcal{C}|}\\) distinct universal perturbation vectors. Learning each such universal vector requires iterating through the dataset, resulting in an exponential time and space complexity." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.319, + 0.482, + 0.469 + ], + "angle": 0, + "content": "Existing universal perturbation methods focus on MCR setting [7, 8, 60]. Adapting these methods to MLL would require learning a separate universal perturbation for each possible label combination, incurring exponential complexity. Although one could learn label-wise independent perturbations, no effective method exists to combine them for attacking specific label combinations. We empirically show that naively combining them does not lead to successful attacks. Similarly, generative attack approaches [27, 69, 76] would be ineffective for unseen label combinations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.47, + 0.482, + 0.666 + ], + "angle": 0, + "content": "Contributions. In this paper, we develop an efficient method for crafting targeted universal adversarial attacks on multi-label learning (MLL) models, addressing the aforementioned combinatorial complexity problem. Our key assumption is to generate a universal attack on a target set as a composition of universal label-wise attacks of the target labels (see Fig. 1). Based on this assumption, we derive a formulation which requires learning affine convex cones spanned by label-wise universal vectors. Our formulation requires linear training time complexity and after trained, the universal perturbation to attack any combination of labels can be composed in constant time using the label-wise attack vectors. Our key contributions are:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.669, + 0.481, + 0.729 + ], + "angle": 0, + "content": "- Existing MCR-based approaches are not inherently designed to attack multiple targeted labels for MLL models. Our work is the first to address this gap, offering an efficient and principled solution for MLL setting;" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.73, + 0.481, + 0.805 + ], + "angle": 0, + "content": "- Our proposed framework to learn targeted universal perturbations for MLL has linear time and space complexity and constructs attacks on any target set in constant time, while achieving performance comparable to the existing oracle methods with exponential complexity;" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.806, + 0.481, + 0.834 + ], + "angle": 0, + "content": "- Our results visualize and provide deeper understanding of the universal adversarial regions learnt by our method;" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.835, + 0.481, + 0.895 + ], + "angle": 0, + "content": "- We evaluate our method on NUS-WIDE, MS-COCO, and OpenImages benchmark datasets using state-of-the-art multi-label learning models against various baselines, showing the effectiveness of our method." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.669, + 0.481, + 0.895 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.09, + 0.664, + 0.106 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.116, + 0.74, + 0.132 + ], + "angle": 0, + "content": "2.1. Multi-Label Recognition" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.139, + 0.905, + 0.381 + ], + "angle": 0, + "content": "There has been significant recent interest in multi-label learning due to its applications in various tasks. In MLL, the goal is to identify all concepts/labels in an image. This is particularly a harder task than multi-class recognition, since there can be exponentially many possible combinations of classes present in an image and can be related to each other. To address this problem, early works [90, 97] considered independent binary classifiers to classify the presence or absence of each class. However, since such a method does not consider the correlations among labels, several other approaches have been proposed that exploit label correlations or external knowledge [13, 18, 34, 45, 49, 51, 70, 94-96, 98, 101, 105-107, 113, 118], or use graphical models [50, 52]. More recent works use attention mechanism or novel loss functions to consider the positive and negative sample imbalance [6, 17, 53, 54, 58, 93, 112]." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.393, + 0.704, + 0.407 + ], + "angle": 0, + "content": "2.2. Adversarial Attacks" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.416, + 0.905, + 0.612 + ], + "angle": 0, + "content": "Adversarial attacks [3, 12, 81, 89, 99] have gained a lot of attention as they identify the key vulnerability of deep neural networks (DNNs) to small imperceptible changes in the input. Researchers have explored various ways that attackers can manipulate data to fool models and developed several defenses to counter these attacks [4, 5, 10, 11, 19, 21, 22, 28, 29, 35, 46, 62, 71-73, 78, 82-84, 91, 92, 102, 114]. Based on the attacker's objective and the target setting, the adversarial attacks can be categorized in different ways, such as targeted and untargeted attacks, white-box (which assume full access to the target model parameters) and black-box attacks (no access to the model parameters), or instance-specific and instance-agnostic attacks." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.621, + 0.783, + 0.636 + ], + "angle": 0, + "content": "2.3. Universal Adversarial Attacks" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.645, + 0.905, + 0.901 + ], + "angle": 0, + "content": "In instance-specific attacks, the perturbations are generated and targeted for each input image, whereas, instance-agnostic attacks [38, 65, 69] generate a single universal perturbation that can fool all input instances. These universal adversarial perturbations(UAPs) can be untargeted or targeted. The goal of the untargeted universal perturbation is to perturb any image such that it is misclassified by the target model [67, 76]. In the targeted case, the goal is to learn the universal perturbation such that it misclassifies any given image to a specific class [7, 110]. Universal perturbations were first proposed by Moosavi-Dezfooli et al. [67] and were generated by iteratively computing and aggregating perturbations using DeepFool [66] across multiple inputs. Since then, several methods have been proposed to learn better universal perturbations [60, 109–111]. To learn UAPs with better generalization, Liu et al. [59] proposed a gradient aggregation method to address gradient vanish" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "20581" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.189, + 0.099, + 0.323, + 0.113 + ], + "angle": 0, + "content": "Current Approach" + }, + { + "type": "image", + "bbox": [ + 0.131, + 0.122, + 0.368, + 0.281 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.482, + 0.099, + 0.732, + 0.113 + ], + "angle": 0, + "content": "Overview of the Proposed Method" + }, + { + "type": "image", + "bbox": [ + 0.382, + 0.117, + 0.593, + 0.277 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.646, + 0.121, + 0.872, + 0.271 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.287, + 0.907, + 0.343 + ], + "angle": 0, + "content": "Figure 2. In multi-label recognition, images can have any combination of labels present, leading to a combinatorial explosion in possible universal perturbations. Traditional methods (shown on the left) require exponential time and space complexity to learn perturbations for all class combinations. Our approach (shown on the right) reduces this complexity by learning \\( |\\mathcal{C}| \\) affine convex cones, each representing a region of universal perturbations for a specific class. This allows us to efficiently generate perturbations for any combination of classes." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.366, + 0.483, + 0.397 + ], + "angle": 0, + "content": "ing problem and Li et al. [48] proposed to utilize instance-specific attack method to generate dominant UAPs." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.408, + 0.379, + 0.423 + ], + "angle": 0, + "content": "2.4. Multi-Label Adversarial Attacks" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.43, + 0.485, + 0.732 + ], + "angle": 0, + "content": "Many recent works have studied adversarial attacks for multi-label recognition models [41, 61, 103, 104, 116, 117]. Song et al. [87] proposed white-box multi-label attack framework and Hu et al. [32] proposed to exploit ranking relations to design attacks for top-\\(k\\) multi-label models. Aich et al. [1] proposed a CLIP-based generative model to generate multi-object attacks in the black-box setting. Jia et al. [37] proposed theoretical robustness guarantees to defend against multi-label adversarial attacks and Melacci et al. [64] exploited domain knowledge context to detect multi-label adversarial attacks. Mahmood and Elhamifar [63] proposed an optimization to address the problem of negative gradient correlation in generating multi-label attacks. These works mainly address instance-specific perturbations for multi-label setting. To the best of our knowledge, only Hu et al. [32] proposed universal untargeted attack where a single perturbation was learnt to attack and replace top-\\(k\\) predicted labels across images. In comparison, we address the problem of learning targeted universal perturbations to attack specific combinations of labels." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.748, + 0.423, + 0.763 + ], + "angle": 0, + "content": "3. Multi-Label Universal Perturbations" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.772, + 0.295, + 0.787 + ], + "angle": 0, + "content": "3.1. Problem Formulation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.794, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Consider a multi-label classifier \\(\\mathcal{F}:\\mathbb{R}^d\\to \\mathbb{R}^{|\\mathcal{C}|}\\) where \\(\\mathcal{C}\\) is the set of all labels. For an image \\(\\pmb {x}\\in \\mathbb{R}^d\\) , let \\(\\pmb {y}\\in \\{0,1\\}^{|\\mathcal{C}|}\\) denote the set of its labels, where each entry indicates the absence (0) or presence (1) of the corresponding label in \\(\\mathcal{C}\\) in the image. The classifier \\(\\mathcal{F} = \\{\\mathcal{F}_1,\\mathcal{F}_2,\\dots ,\\mathcal{F}_{|\\mathcal{C}|}\\}\\) consists of \\(|\\mathcal{C}|\\) binary classifiers (one for each label), where \\(\\mathcal{F}_c(\\pmb {x})\\in (-\\infty , + \\infty)\\) represents the logit of label \\(c\\) .The" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.366, + 0.907, + 0.472 + ], + "angle": 0, + "content": "probability of label \\(c\\) being present in the image is given by \\(\\hat{o}_c = \\sigma (\\mathcal{F}_c(\\pmb {x}))\\), where \\(\\sigma (.)\\) is the sigmoid function, hence, it is considered present (1), if \\(\\hat{o}_c\\geq 0.5\\), and absent (0) otherwise. We assume that \\(\\mathcal{F}\\) has already been trained on a set of images. Let \\(\\Omega \\subseteq \\mathcal{C}\\) denote the set of target labels that we want to attack. Our goal is to generate a targeted universal perturbation for \\(\\Omega\\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.482, + 0.833, + 0.497 + ], + "angle": 0, + "content": "3.2. Multi-Label Universal Perturbations" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.505, + 0.907, + 0.672 + ], + "angle": 0, + "content": "We want to find a single (universal) perturbation vector, \\( \\pmb{u}_{\\Omega} \\), that can attack a set of target labels in \\( \\Omega \\) in any image \\( \\pmb{x} \\) containing them. One can generate such a vector by maximizing the expected multi-label learning loss for the labels in \\( \\Omega \\) across the training dataset, while restricting the norm of \\( \\pmb{u}_{\\Omega} \\) (we consider \\( \\ell_{\\infty} \\)-norm). However, due to correlations among labels in MLL, attacking the target labels may inadvertently affect non-target labels \\( (\\bar{\\Omega} = \\mathcal{C} \\setminus \\Omega) \\). To mitigate this, we seek to minimize the influence of the universal perturbation vector on non-target labels. This requires solving the following oracle optimization," + }, + { + "type": "equation", + "bbox": [ + 0.58, + 0.677, + 0.907, + 0.707 + ], + "angle": 0, + "content": "\\[\nP _ {\\Omega} ^ {*}: \\max _ {\\boldsymbol {u} _ {\\Omega}} \\sum_ {k \\in \\Omega} \\mathcal {H} _ {k} (\\boldsymbol {u} _ {\\Omega}) - \\gamma \\sum_ {q \\in \\Omega} \\mathcal {R} _ {q} (\\boldsymbol {u} _ {\\Omega}), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.711, + 0.906, + 0.788 + ], + "angle": 0, + "content": "where \\(\\mathcal{H}_k\\) ensures attacking the label \\(k\\) in the target set and \\(\\mathcal{R}_q\\) tries to prevent modifying the prediction of non-target labels using the universal perturbation \\(\\pmb{u}_{\\Omega}\\). Here, \\(\\gamma\\) is a scalar hyperparameter setting a trade-off between the two terms. More specifically, we define \\(\\mathcal{H}_k\\) and \\(\\mathcal{R}_q\\) as" + }, + { + "type": "equation", + "bbox": [ + 0.514, + 0.795, + 0.905, + 0.813 + ], + "angle": 0, + "content": "\\[\n\\mathcal {H} _ {k} \\left(\\boldsymbol {u} _ {\\Omega}\\right) \\triangleq \\mathbb {E} _ {\\left(\\boldsymbol {x}, \\boldsymbol {y}\\right)} \\mathcal {L} _ {\\mathrm {b c e}} \\left(\\mathcal {F} _ {k} \\left(\\boldsymbol {x} + \\Pi_ {\\epsilon} \\left(\\boldsymbol {u} _ {\\Omega}\\right)\\right), y _ {k}\\right), \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.516, + 0.815, + 0.905, + 0.833 + ], + "angle": 0, + "content": "\\[\n\\mathcal {R} _ {q} (\\boldsymbol {u} _ {\\Omega}) \\triangleq \\mathbb {E} _ {(\\boldsymbol {x}, \\boldsymbol {y})} \\max \\left\\{0, - \\tanh \\left(\\mathcal {F} _ {q} (\\boldsymbol {x}) \\mathcal {F} _ {q} (\\boldsymbol {x} + \\Pi_ {\\epsilon} (\\boldsymbol {u} _ {\\Omega}))\\right) \\right\\},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.907, + 0.901 + ], + "angle": 0, + "content": "where \\(\\Pi_{\\epsilon}\\) projects the input on p-norm ball of radius \\(\\epsilon\\), \\(\\mathcal{L}_{\\mathrm{bce}}\\) is the binary cross-entropy loss and the expectation is with respect to samples from the dataset. Therefore, maximizing \\(\\mathcal{H}_k(\\pmb{u}_{\\Omega})\\) ensures to find a perturbation that modifies the" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20582" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.198 + ], + "angle": 0, + "content": "prediction of each target label \\( k \\). On the other hand, \\( \\mathcal{R}_k \\) aims to minimize the influence of the perturbation \\( (\\pmb{u}_{\\Omega}) \\) on each non-target label \\( q \\). This is done by incurring a positive penalty when the signs of the logits for the unperturbed and perturbed inputs differ from each other. Thus, minimizing this term leads to the same signs for the logits of each non-target label before and after the attack." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.205, + 0.484, + 0.236 + ], + "angle": 0, + "content": "3.3. Compositional Multi-Label Universal Attacks (CMLU)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.241, + 0.484, + 0.332 + ], + "angle": 0, + "content": "There are \\(2^{|C|}\\) possible subsets of labels \\(\\mathcal{C}\\). Therefore, finding all universal perturbations \\(\\{\\pmb{u}_{\\Omega}\\}\\) by independently solving (1) for each subset \\(\\Omega\\) leads to the exponential complexity \\(O(2^{|C|})\\). We develop a compositional multi-label universal attack (CMLU) that allows us to find any universal perturbation in linear complexity \\(O(|C|)\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.333, + 0.484, + 0.378 + ], + "angle": 0, + "content": "Let \\(\\mathcal{P}(\\mathcal{C})\\) denote the power set of \\(\\mathcal{C}\\), i.e., the set of all subsets of \\(\\mathcal{C}\\). Our goal is to find all \\(\\pmb{u}_{\\Omega}\\)'s by solving (1) for every \\(\\Omega \\in \\mathcal{P}(\\mathcal{C})\\)," + }, + { + "type": "equation", + "bbox": [ + 0.133, + 0.378, + 0.482, + 0.413 + ], + "angle": 0, + "content": "\\[\n\\sum_ {\\Omega \\in \\mathcal {P} (\\mathcal {C})} \\max _ {\\boldsymbol {u} _ {\\Omega}} \\left[ \\sum_ {k \\in \\Omega} \\mathcal {H} _ {k} \\left(\\boldsymbol {u} _ {\\Omega}\\right) - \\gamma \\sum_ {q \\in \\bar {\\Omega}} \\mathcal {R} _ {q} \\left(\\boldsymbol {u} _ {\\Omega}\\right) \\right]. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.416, + 0.484, + 0.476 + ], + "angle": 0, + "content": "As mentioned above, solving (3) over the power set of \\(\\mathcal{C}\\) has exponential complexity. To tackle this problem, we make the following assumption: each universal perturbation \\(\\pmb{u}_{\\Omega}\\) can be written as the sum of atomic universal perturbations" + }, + { + "type": "equation", + "bbox": [ + 0.239, + 0.479, + 0.482, + 0.51 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {u} _ {\\Omega} \\triangleq \\sum_ {i \\in \\Omega} \\boldsymbol {u} _ {i}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.513, + 0.484, + 0.576 + ], + "angle": 0, + "content": "where \\(\\pmb{u}_i\\) is the universal perturbation on the label \\(i\\). We also assume that the label-wise universal perturbations are norm-constrained and mutually orthogonal, i.e., \\(\\| \\pmb{u}_i\\|_p \\leq \\epsilon\\) and \\(\\hat{\\pmb{u}}_i^\\top \\hat{\\pmb{u}}_j = 0\\) for each \\(i \\neq j\\), where \\(\\hat{\\pmb{u}} = \\pmb{u} / \\| \\pmb{u}\\|_2\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.575, + 0.484, + 0.666 + ], + "angle": 0, + "content": "Let \\( \\mathbf{U} = \\left[ \\begin{array}{lll}\\pmb {u}_1 & \\pmb {u}_2 & \\dots & \\pmb{u}_{|\\mathcal{C}|} \\end{array} \\right] \\) represent a matrix whose columns are label-wise universal perturbations and \\( \\lambda_{\\Omega} \\) denote a vector whose elements indexed by \\( \\Omega \\) are one and other elements are zero. Based on our assumption, we optimize \\( \\mathbf{U} \\) so that we can construct any \\( \\pmb{u}_{\\Omega} \\) using the columns of \\( \\mathbf{U} \\). Given our assumptions, we can rewrite (3) as" + }, + { + "type": "equation", + "bbox": [ + 0.094, + 0.668, + 0.482, + 0.722 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\max _ {\\boldsymbol {U}} \\sum_ {\\Omega \\in \\mathcal {P} (\\mathcal {C})} \\sum_ {k \\in \\Omega} \\left(\\mathcal {H} _ {k} \\left(\\boldsymbol {u} _ {k} + \\boldsymbol {U} \\boldsymbol {\\lambda} _ {\\Omega \\backslash k}\\right) - \\frac {\\gamma}{| \\Omega |} \\sum_ {q \\in \\Omega} \\mathcal {R} _ {q} \\left(\\boldsymbol {u} _ {k} + \\boldsymbol {U} \\boldsymbol {\\lambda} _ {\\Omega \\backslash k}\\right)\\right), \\\\ s. t. \\hat {\\boldsymbol {U}} ^ {\\top} \\hat {\\boldsymbol {U}} = \\boldsymbol {I}, \\| \\boldsymbol {U} \\| _ {1, p} \\leq \\epsilon \\tag {5} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.726, + 0.484, + 0.787 + ], + "angle": 0, + "content": "where \\(\\hat{U}\\) is column-normalized \\(U\\) and \\(\\| U\\|_{1,p} = \\max_{1\\leq i\\leq |\\mathcal{C}|}\\| u_i\\| _p\\), is the maximum norm of the columns of \\(U\\). Notice that we wrote \\(u_{\\Omega} = u_k + U\\lambda_{\\Omega \\backslash k}\\) for a \\(k\\) that belongs to \\(\\Omega\\), where \\(\\Omega \\backslash k\\) denotes removing \\(k\\) from \\(\\Omega\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.787, + 0.484, + 0.847 + ], + "angle": 0, + "content": "One can see that, for any \\( k \\in \\mathcal{C} \\), there are \\( 2^{|C| - 1} \\) subsets that contain \\( k \\). Each of these subsets would contain \\( k \\) and classes from the powerset \\( \\mathcal{P}(\\mathcal{C} \\backslash k) \\). We can rewrite (5) as (see the supplementary materials for derivation details)" + }, + { + "type": "equation", + "bbox": [ + 0.094, + 0.849, + 0.485, + 0.882 + ], + "angle": 0, + "content": "\\[\n\\max _ {\\boldsymbol {U}} \\sum_ {k \\in \\mathcal {C}} \\sum_ {\\Omega \\in \\mathcal {P} (\\mathcal {C} \\backslash k)} \\left(\\mathcal {H} _ {k} \\left(\\boldsymbol {u} _ {k} + \\boldsymbol {U} \\boldsymbol {\\lambda} _ {\\Omega \\backslash k}\\right) - \\eta \\sum_ {q \\in \\{\\bar {\\Omega} \\backslash k \\}} \\mathcal {R} _ {q} \\left(\\boldsymbol {u} _ {k} + \\boldsymbol {U} \\boldsymbol {\\lambda} _ {\\Omega \\backslash k}\\right)\\right)\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.097, + 0.885, + 0.483, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\mathrm {s . t .} \\hat {\\boldsymbol {U}} ^ {\\top} \\hat {\\boldsymbol {U}} = \\boldsymbol {I}, \\| \\boldsymbol {U} \\| _ {1, p} \\leq \\epsilon , \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.907, + 0.172 + ], + "angle": 0, + "content": "where \\(\\eta = \\gamma /|\\Omega |\\) . The power set of \\(\\mathcal{C}\\) can be written as the union of all subsets of \\(\\mathcal{C}\\) .. \\(\\mathcal{P}(\\mathcal{C})\\triangleq \\bigcup_{i = 0}^{|C|}\\binom {\\mathcal{C}}{i}\\) where, each \\(\\binom{\\mathcal{C}}{i}\\) has all the subsets of \\(\\mathcal{C}\\) of size \\(i\\) . In our case, \\(S\\in \\binom{\\mathcal{C}}{i}\\) would represent the indices of the columns of \\(\\pmb{U}\\) . So, we can write the summation across the power set as:" + }, + { + "type": "equation", + "bbox": [ + 0.515, + 0.175, + 0.912, + 0.219 + ], + "angle": 0, + "content": "\\[\n\\max_{\\boldsymbol{U}}\\sum_{k\\in \\mathcal{C}}\\sum_{t = 0}^{\\left|\\mathcal{C}\\right| - 1}\\sum_{S\\in \\binom {\\mathcal{C}\\setminus k}{t}}\\Bigl(\\mathcal{H}_{k}(\\boldsymbol{u}_{k} + \\boldsymbol {U}\\boldsymbol{\\lambda}_{S}) - \\eta \\sum_{q\\in \\{\\bar{S}\\setminus k\\}}\\mathcal{R}_{q}(\\boldsymbol{u}_{k} + \\boldsymbol {U}\\boldsymbol{\\lambda}_{S})\\Bigr),\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.52, + 0.222, + 0.905, + 0.242 + ], + "angle": 0, + "content": "\\[\ns. t. \\hat {\\boldsymbol {U}} ^ {\\top} \\hat {\\boldsymbol {U}} = \\boldsymbol {I}, \\| \\boldsymbol {U} \\| _ {1, p} \\leq \\epsilon , \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.247, + 0.907, + 0.371 + ], + "angle": 0, + "content": "In (7), for a fixed \\( k \\) and \\( t \\), \\( \\lambda_{S} \\) has exactly \\( t \\) non-zero elements and \\( S \\) iterates through all possible combinations of \\( \\{C \\backslash k\\} \\) of size \\( t \\). Therefore, the sum across all \\( \\lambda_{S} \\) (for \\( S \\in \\binom{\\mathcal{C} \\backslash k}{t} \\)) would be equal to the sum across all binary vectors \\( \\lambda \\in \\{0,1\\}^{|\\mathcal{C}|} \\) such that exactly \\( t \\) elements are nonzero (i.e., \\( \\| \\lambda \\|_0 = t \\)) and \\( \\lambda_{k} = 0 \\). Where, \\( \\| \\cdot \\|_0 \\) denotes the \\( \\ell_0 \\)-norm, which counts the number of nonzero elements of the given input. So we can write (7) as" + }, + { + "type": "equation", + "bbox": [ + 0.515, + 0.373, + 0.905, + 0.424 + ], + "angle": 0, + "content": "\\[\n\\max _ {\\boldsymbol {U}} \\sum_ {k \\in \\mathcal {C}} \\sum_ {t = 0} ^ {| \\mathcal {C} | - 1} \\sum_ {\\substack {\\boldsymbol {\\lambda} \\in \\{0, 1 \\} ^ {| \\mathcal {C} |} \\\\ \\| \\boldsymbol {\\lambda} \\| _ {0} = t, \\boldsymbol {\\lambda} _ {k} = 0}} \\left[ \\mathcal {H} _ {k} (\\boldsymbol {u} _ {k} + \\boldsymbol {U} \\boldsymbol {\\lambda}) \\right. \\tag{8}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.518, + 0.426, + 0.908, + 0.459 + ], + "angle": 0, + "content": "\\[\n\\left. - \\eta \\sum_ {q \\in \\{\\mathcal {C} \\backslash k \\}} \\delta (\\boldsymbol {\\lambda} _ {q} = 0) \\mathcal {R} _ {q} (\\boldsymbol {u} _ {k} + \\boldsymbol {U} \\boldsymbol {\\lambda}) \\right] \\text {s . t .} \\hat {\\boldsymbol {U}} ^ {\\top} \\hat {\\boldsymbol {U}} = \\boldsymbol {I}, \\| \\boldsymbol {U} \\| _ {1, p} \\leq \\epsilon ,\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.461, + 0.906, + 0.522 + ], + "angle": 0, + "content": "where \\(\\delta\\) is the indicator function which is 1 if the input condition is True and 0 otherwise. As \\(\\lambda\\) is \\(|\\mathcal{C}|\\) dimensional, the second loss term in (8) iterates through each index of \\(\\lambda\\) i.e., \\((q\\in \\{\\mathcal{C}\\backslash k\\})\\) and computes the loss of label \\(q\\) if \\(\\lambda_q = 0\\)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.522, + 0.906, + 0.552 + ], + "angle": 0, + "content": "We investigate two ways to approximately solve (8). In the first approach, we solve the lower bound of (8) as" + }, + { + "type": "equation", + "bbox": [ + 0.517, + 0.557, + 0.905, + 0.594 + ], + "angle": 0, + "content": "\\[\n\\operatorname {CMLU} _ {\\alpha}: \\max _ {\\boldsymbol {U}} \\sum_ {k \\in \\mathcal {C}} \\sum_ {t \\in \\mathcal {T}} 2 ^ {| \\mathcal {C} |} \\min _ {\\substack {\\boldsymbol {\\lambda} \\in \\{0,1 \\} ^ {| C |} \\\\ \\| \\boldsymbol {\\lambda} \\| _ {0} = t, \\boldsymbol {\\lambda} _ {k} = 0}} \\left[ \\mathcal {H} _ {k} (\\boldsymbol {u} _ {k} + \\boldsymbol {U} \\boldsymbol {\\lambda}) \\right. \\tag{9}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.521, + 0.597, + 0.903, + 0.63 + ], + "angle": 0, + "content": "\\[\n\\left. - \\eta \\sum_ {q \\in \\{\\mathcal {C} \\backslash k \\}} \\delta (\\boldsymbol {\\lambda} _ {q} = 0) \\mathcal {R} _ {q} (\\boldsymbol {u} _ {k} + \\boldsymbol {U} \\boldsymbol {\\lambda}) \\right] \\text {s . t .} \\hat {\\boldsymbol {U}} ^ {\\top} \\hat {\\boldsymbol {U}} = \\boldsymbol {I}, \\| \\boldsymbol {U} \\| _ {1, p} \\leq \\epsilon\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.636, + 0.907, + 0.771 + ], + "angle": 0, + "content": "Where, \\(\\mathcal{T} = \\{0,1,\\dots ,|\\mathcal{C}| - 1\\}\\). (9) requires maximizing the minimal loss. To solve the inner minimization problem, for a fixed \\(k\\) and \\(t\\), we randomly initialize \\(\\lambda\\) and iteratively update it for a fixed number of iterations. We then project \\(\\lambda\\) on the feasible set by keeping the largest \\(t\\) values as 1 and setting the rest to 0. We then keep the \\(\\lambda\\) fixed and solve the outer maximization in (9) w.r.t. \\(\\pmb{U}\\). In practice, instead of iterating through all \\(t\\in \\mathcal{T}\\), for each \\(k\\), we randomly choose \\(t\\). The steps are shown in Algorithm 1." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.772, + 0.906, + 0.818 + ], + "angle": 0, + "content": "In the second approach, we consider the continuous relaxation of \\(\\lambda\\) and approximately solve it by uniformly randomly sampling \\(\\lambda \\in [0,1]^{|\\mathcal{C}|}\\)," + }, + { + "type": "equation", + "bbox": [ + 0.515, + 0.822, + 0.905, + 0.867 + ], + "angle": 0, + "content": "\\[\n\\operatorname {C M L U} _ {\\beta}: \\max _ {\\boldsymbol {U}} \\sum_ {k \\in \\mathcal {C}} \\sum_ {t \\in \\mathcal {T}} \\sum_ {\\substack {\\boldsymbol {\\lambda} \\in [ 0, 1 ] ^ {| \\mathcal {C} |} \\\\ \\| \\boldsymbol {\\lambda} \\| _ {0} = t, \\boldsymbol {\\lambda} _ {k} = 0}} \\left[ \\mathcal {H} _ {k} (\\boldsymbol {u} _ {k} + \\boldsymbol {U} \\boldsymbol {\\lambda}) \\right. \\tag{10}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.519, + 0.869, + 0.905, + 0.902 + ], + "angle": 0, + "content": "\\[\n\\left. - \\eta \\sum_ {q \\in \\{\\mathcal {C} \\backslash k \\}} \\delta (\\boldsymbol {\\lambda} _ {q} = 0) \\mathcal {R} _ {q} (\\boldsymbol {u} _ {k} + \\boldsymbol {U} \\boldsymbol {\\lambda}) \\right] \\text {s . t .} \\hat {\\boldsymbol {U}} ^ {\\top} \\hat {\\boldsymbol {U}} = \\boldsymbol {I}, \\| \\boldsymbol {U} \\| _ {1, p} \\leq \\epsilon .\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20583" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.212 + ], + "angle": 0, + "content": "Eq. (9) and (10) can be explained intuitively. The first loss term \\(\\mathcal{H}_k\\) optimizes each label-wise universal perturbation \\(\\pmb{u}_k\\) such that it is robust to the noise generated by composing other universal perturbations. Specifically, it aims to maximize the loss of class \\(k\\), regardless of which other class universals have been combined. The second term ensures that \\(\\pmb{u}_k\\) does not change the prediction of classes whose universal vectors have not been used in the composition." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.213, + 0.483, + 0.379 + ], + "angle": 0, + "content": "Remark: A key aspect of Eq. (9) and (10) is that \\(\\pmb{u}_k + \\pmb{U}\\lambda\\) represents points sampled from an affine convex cone. Specifically, these formulations aim to maximize a function over points sampled from \\(|C|\\) affine convex cones (generated by the columns of \\(\\pmb{U}\\)). Consequently, we can interpret our assumption in (4) as aiming to learn the universal adversarial regions of each class as affine cones, where the vector \\(\\pmb{u}_{\\Omega}\\), as outlined in (4), lies at the intersection of the affine cones of the classes in \\(\\Omega\\). This is illustrated in Fig. 2 (right), where \\(\\pmb{C}_k\\) and \\(\\pmb{C}_i\\) represent convex cones for classes \\(k\\) and \\(i\\) respectively, and \\(\\pmb{C}_{ik}\\) denotes their intersection." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.381, + 0.485, + 0.613 + ], + "angle": 0, + "content": "For a comparison among Oracle, \\(\\mathrm{CMLU}_{\\alpha}\\), and \\(\\mathrm{CMLU}_{\\beta}\\), Algorithm 1 shows the steps of a single training epoch of Oracle of target size \\(|\\Omega| = 1\\) and the corresponding steps for our methods. Note that in general, Oracle has \\(\\mathcal{O}(\\binom{|C|}{|\\Omega|})\\) training time complexity, which increases with the size of \\(\\Omega\\). For instance, for \\(|\\Omega| = 2\\), the complexity of the Oracle's inner loop becomes \\(\\mathcal{O}(\\binom{|C|}{2})\\), as it needs to handle all possible combinations of classes of a fixed size \\(|\\Omega|\\). However, \\(\\mathrm{CMLU}_{\\alpha}\\) has \\(\\mathcal{O}(N_{\\alpha}|\\mathcal{C}|)\\) and \\(\\mathrm{CMLU}_{\\beta}\\) has \\(\\mathcal{O}(|C|)\\) complexity and is independent of \\(|\\Omega|\\). Note that the complexity of our method for learning composable universal perturbations for all possible target set sizes is equivalent to the complexity of Oracle trained for \\(|\\Omega| = 1\\) (linear in the number of classes). Oracle requires iterating over the dataset for each combination of target classes \\(\\Omega\\), leading to exponential complexity." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.626, + 0.208, + 0.641 + ], + "angle": 0, + "content": "4. Evaluation" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.652, + 0.281, + 0.669 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.674, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Datasets We perform experiments using NUS-WIDE [16], MS-COCO [56], and OpenImages [44] datasets to assess the effectiveness of our proposed method. We compare it with oracle and baseline methods. For NUS-WIDE and MS-COCO, we test target set sizes ranging from 1 to 5 and for OpenImages, we experiment with larger target set sizes, going up to 20. Due to the computational expense of learning oracle model, we choose a subset of labels from each dataset (25 classes from NUS-WIDE, 40 classes from MS-COCO, and 100 classes from OpenImages). Even with the small subset of classes, there can be a large number of possible combinations e.g., there can be more than 53k combinations of \\( |\\Omega| = 5 \\) using 25 classes from NUS-WIDE. Therefore, we perform experiments on a subset of those possible combinations for each target set size \\( |\\Omega| \\). For all of our" + }, + { + "type": "code_caption", + "bbox": [ + 0.52, + 0.095, + 0.862, + 0.11 + ], + "angle": 0, + "content": "Algorithm 1: Algorithms of Oracle, \\(\\mathrm{CMLU}_{\\alpha}\\), \\(\\mathrm{CMLU}_{\\beta}\\)" + }, + { + "type": "algorithm", + "bbox": [ + 0.517, + 0.112, + 0.905, + 0.578 + ], + "angle": 0, + "content": "Input: \\(\\mathcal{D}\\) : Training Data, \\(\\mathcal{F}\\) : Classifier, \\(\\epsilon\\) : Norm budget, \\((\\xi ,\\alpha)\\) : Step sizes Output: \\(U\\) : Learnt Universal Perturbations 1 Randomly initialize \\(U\\) /\\* \\(\\mathcal{B} = \\{\\mathcal{B}^{(k)}\\}_{k = 1}^{|C|},\\mathcal{B}^{(k)}\\) images of class \\(k\\ast /\\) \n2 Oracle: \n3 for Batch \\(\\mathcal{B}\\) in \\(\\mathcal{D}\\) do /* One iteration through the dataset \\*/ \n4 for \\(k\\) in \\(\\mathcal{C}\\) do /\\* O(|C|) \\*/ \n5 \\(\\mathcal{L}_k\\) : Compute loss in (1) using \\(\\pmb {u}_k\\) \n6 Compute the gradient: \\(\\pmb {g}\\gets \\nabla_{\\pmb{U}_k}\\pmb {l}_k\\) \n7 Update \\(U_{k}\\leftarrow \\Pi_{\\epsilon}\\left(U_{k} + \\xi *Sgn(\\pmb {g})\\right)\\) \n8 for Batch \\(\\mathcal{B}\\) in \\(\\mathcal{D}\\) do \n9 Set \\(\\pmb{g}\\) to be all zeros of the same size as \\(U\\) \n10 for \\(k\\) in \\(\\mathcal{C}\\) do /\\* O(|C|) \\*/ \n11 Randomly sample \\(t\\sim T\\) and \\(\\lambda \\in [0,1]^{\\lfloor C\\rfloor}\\) . \n12 \\(\\mathrm{CMLU}_{\\alpha}\\) ： /\\*O(Nα）\\*/ \n13 for \\(N_{\\alpha}\\) num of iterations do \n14 Set \\(\\pmb {v}_k = \\pmb {u}_k + U\\pmb{\\lambda}\\) \n15 \\(\\mathcal{L}_k\\) : Compute loss in (9) using \\(\\pmb {v}_k\\) \n16 Update: \\(\\lambda = \\lambda -\\alpha \\nabla_{\\lambda}\\mathcal{L}_k\\) \n17 Project \\(\\lambda\\) on the feasible set \\(\\| \\lambda \\| _0 = t\\) \n18 Set \\(\\lambda_k = 0\\) and \\(\\pmb {v}_k = \\pmb {u}_k + \\pmb {U}\\pmb{\\lambda}\\) \n19 \\(\\mathcal{L}_k\\) : Compute loss in (9) using \\(\\pmb {v}_k\\) \n20 CMLUβ:/*O(1)\\*/ \n21 Randomly set \\(|\\mathcal{C}| - t\\) values in \\(\\lambda\\) to 0. \n22 Set \\(\\lambda_k = 0\\) and \\(\\pmb {v}_k = \\pmb {u}_k + U\\pmb{\\lambda}\\) \n23 \\(\\mathcal{L}_k\\) :Compute loss in (10) using \\(\\pmb {v}_k\\) \n24 Accumulate: \\(\\pmb {g}\\gets \\pmb {g} + \\nabla_{\\pmb{U}_k}\\mathcal{L}_k\\) \n25 Update \\(U\\gets \\Pi_{\\epsilon}(U + \\xi *Sgn(g))\\)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.61, + 0.905, + 0.7 + ], + "angle": 0, + "content": "experiments, we learn universal vectors with maximum infinity norm \\(\\epsilon = 0.05\\) (for image values range of [0,1]) and using SGD with fix step size \\(\\xi = \\epsilon /25\\). We set \\(\\gamma = \\kappa /|\\bar{\\Omega}|\\) for training oracle and baselines, and \\(\\eta = \\kappa /\\| \\pmb {\\lambda}\\| _0\\) in each iteration of our proposed method. We show ablation experiments for different values of \\(\\kappa\\) in section 4.2.1." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.71, + 0.905, + 0.756 + ], + "angle": 0, + "content": "Multi-Label Recognition Models We evaluate the baselines and our method on state-of-the-art multi-label recognition models (using the officially provided weights)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.76, + 0.907, + 0.867 + ], + "angle": 0, + "content": "- Asymmetric Loss (ASL) [79] is a novel loss function to address the imbalance between numerous negative labels and only a few positive labels associated with each sample. This method balances the probabilities of different positive and negative samples. In particular, the loss performs hard thresholding by dynamically down-weighting easy negative samples and can discard unlabeled samples." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.907, + 0.9 + ], + "angle": 0, + "content": "- ML-Decoder [80] is an efficient, attention-based multi-label classifier that has a novel decoder architecture and" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20584" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.111, + 0.094, + 0.28, + 0.229 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.283, + 0.094, + 0.451, + 0.229 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.453, + 0.094, + 0.633, + 0.229 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.633, + 0.094, + 0.803, + 0.229 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.803, + 0.106, + 0.882, + 0.229 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.232, + 0.905, + 0.26 + ], + "angle": 0, + "content": "Figure 3. Fooling success rates \\((FR)\\) on various datasets using ASL and ML-Decoder as we increase the target set size \\(|\\Omega|\\). The left two figures show results on NUS-WIDE and the right two figures show results on MS-COCO." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.283, + 0.482, + 0.344 + ], + "angle": 0, + "content": "group-decoding scheme, which enables it to efficiently utilize the spatial information in images. The model is highly efficient as it can scale to large number of classes. Moreover, using word queries, it can generalize to unseen classes." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.349, + 0.483, + 0.41 + ], + "angle": 0, + "content": "Oracle and Baselines We evaluate our method against three baselines and an oracle. The computational complexity of each of these methods is reported in Table 1 and their wall clock time in supplementary." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.416, + 0.484, + 0.493 + ], + "angle": 0, + "content": "- Oracle directly optimizes (1) to learn a universal perturbation \\(\\pmb{u}_{\\Omega}\\) to attack a specific target set \\(\\Omega\\). This provides the upper bound on the attack performance for our method. It requires \\(2^{|C|}d\\)-dimensional universal vectors and for each vector, iterate through the training set for \\(E\\) epochs." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.498, + 0.484, + 0.561 + ], + "angle": 0, + "content": "- Oracle Sum (Or-S) directly learns single-class universal perturbation vectors \\(\\{\\pmb{u}_1, \\pmb{u}_2, \\dots, \\pmb{u}_{|\\mathcal{C}|}\\}\\) using (1) and generates a universal perturbation to attack set \\(\\Omega\\) as \\(\\pmb{u}_{\\Omega} = \\sum_{i \\in \\Omega} \\pmb{u}_i\\). It only needs to learn \\(|C|\\) \\(d\\)-dimensional vectors." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.566, + 0.484, + 0.658 + ], + "angle": 0, + "content": "- Oracle Combination (Or-C) first learns single-class universal perturbation vectors and learns to combine them to attack a specific target set \\(\\Omega\\) i.e., \\(\\pmb{u}_{\\Omega} = \\sum_{i\\in \\Omega}a_i^\\Omega \\pmb {u}_i\\). Where, \\(a_i^\\Omega \\in \\mathbb{R}\\) is associated with \\(i^{th}\\) class in specific set \\(\\Omega\\), and is learnt by iterating through the training set. It needs to learn \\(|C|\\) \\(d\\)-dim vectors and \\(2^{|C|}\\) scalars to combine them." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.664, + 0.484, + 0.74 + ], + "angle": 0, + "content": "- SGA [59] addresses the gradient vanishing and poor local optima problem to improve the generalization of UAPs. It enhances the gradient stability by aggregating gradients from inner pre-search. We use SGA to learn single-class UAPs and use (4) to generate UAPs for target labels." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.746, + 0.483, + 0.913 + ], + "angle": 0, + "content": "- NAG [68] models the distribution of universal adversarial perturbations using GANs, originally designed for untargeted perturbations in multi-class setting. To adapt this for multi-label universal perturbations, we condition the GAN on class vectors. During training, we generate label-wise universal perturbations. At test time, we generate perturbations for multiple classes by conditioning the GAN on a combined representation vector of the target classes. For this method, we have linear time complexity as we train to generate label-wise perturbations and the space requirement is the size of the generator." + }, + { + "type": "list", + "bbox": [ + 0.09, + 0.498, + 0.484, + 0.913 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.283, + 0.905, + 0.465 + ], + "angle": 0, + "content": "Evaluation Metrics Given a set of classes \\(\\Omega\\) to attack, let \\(\\mathcal{I}_{\\Omega}\\) represent all the images that contain the classes \\(\\Omega\\) and were attacked. Let \\(\\mathcal{A}_{\\Omega} \\subseteq \\mathcal{I}_{\\Omega}\\) represent the set of images that were successfully attacked. To evaluate the universal attack performance, we define the attack fooling success rate, \\(FR\\), to measure the percentage of successfully attacked images. Since we want to affect the least number of non-targeted labels while achieving high fooling rates, we also report non-target flip rate, \\(NT_{R}\\), to measure the percentage of non-targeted labels (labels in \\(\\bar{\\Omega}\\)) which were flipped by the attack. Our goal is to achieve high \\(FR\\) while maintaining low \\(NT_{R}\\) for attacking target set of any size." + }, + { + "type": "equation", + "bbox": [ + 0.53, + 0.47, + 0.905, + 0.521 + ], + "angle": 0, + "content": "\\[\n\\mathrm {F R} = \\frac {\\left| \\mathcal {A} _ {\\Omega} \\right|}{\\left| \\mathcal {I} _ {\\Omega} \\right|}, \\mathrm {N T} _ {R} = \\frac {1}{\\left| \\mathcal {A} _ {\\Omega} \\right|} \\sum_ {k \\in \\mathcal {A} _ {\\Omega}} \\frac {\\sum_ {i \\in \\bar {\\Omega}} \\left(1 - \\delta \\left(\\chi \\left(f _ {i} ^ {(k)}\\right) , y _ {i} ^ {(k)}\\right)\\right)}{\\left| \\bar {\\Omega} \\right|} \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.521, + 0.906, + 0.63 + ], + "angle": 0, + "content": "where, \\(\\delta\\) is Kronecker delta that equals 1 when the two inputs are equal and 0 otherwise. \\(\\chi (v)\\) is the step function which is 1 if \\(v > 0\\) and 0 otherwise. \\(y_{i}^{(k)},f_{i}^{(k)}\\) are the predictions on clean and perturbed images respectively of \\(i^{th}\\) non-target class of \\(k^{th}\\) successfully attacked image. Note that \\(FR\\) is calculated using all attacked images while \\(NT_{R}\\) is calculated using the set of successfully attacked images." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.638, + 0.715, + 0.654 + ], + "angle": 0, + "content": "4.2. Experimental Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.66, + 0.906, + 0.78 + ], + "angle": 0, + "content": "In this section, we evaluate our compositional framework for universal perturbations. Fig. 3 shows the fooling success rates for different methods. The results demonstrate consistent performance trends across datasets and models. Particularly, the success rate(FR) of baselines decline as we increase the target size. On the other hand, the Oracle method achieves the best performance but suffers from exponential complexity, making it impractical for larger target sets." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.781, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Among the baselines, Or-C, Or-S, and SGA perform the worst for large target set sizes. Since the performance of these methods depends on the label-wise universal perturbations, we can conclude that the low success rate of these methods implies that the subspace spanned by labelwise perturbations does not contain the intersection of their respective adversarial regions. This is further elaborated in Fig. 5. In all cases, Or-C is better than Or-S since it op-" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "20585" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.907, + 0.133 + ], + "angle": 0, + "content": "Table 1. Non-target flip rate \\((NT_R)\\) on NUS-WIDE and MS-COCO, as we increase the target set size from 1 to 5. We show (-) when \\(|\\mathcal{A}_{\\Omega}| = 0\\). We also show the space and time complexity of each method. \\(E\\) represents the number of epochs for training, \\(d\\) is the size of the perturbation (equal to the image size), \\(G_p\\) is the size of GAN's parameters, and \\(|\\mathcal{C}|\\) is the number of classes." + }, + { + "type": "table", + "bbox": [ + 0.119, + 0.139, + 0.88, + 0.307 + ], + "angle": 0, + "content": "
Computational ComplexityASLML-Decoder
NUS-WIDEMS-COCONUS-WIDEMS-COCO
MethodSpaceTime135135135135
Oracle\\( \\mathcal{O}(d2^{|C|}) \\)\\( \\mathcal{O}(E2^{|C|}) \\)6.177.428.106.507.738.662.863.103.017.388.589.81
Or-S\\( \\mathcal{O}(d|C|) \\)\\( \\mathcal{O}(E|C|) \\)6.176.03-6.502.872.962.862.22-7.382.96-
Or-C\\( \\mathcal{O}(d|C|+2^{|C|}) \\)\\( \\mathcal{O}(E2^{|C|}) \\)6.176.862.746.505.233.202.862.450.607.386.40-
SGA\\( \\mathcal{O}(d|C|) \\)\\( \\mathcal{O}(E|C|) \\)7.697.31-7.774.30-2.892.471.278.628.949.37
NAG\\( \\mathcal{O}(G_p) \\)\\( \\mathcal{O}(E|C|) \\)8.325.861.3210.110.27.213.963.452.5511.812.113.7
\\( \\mathbf{CMLU}_{\\alpha} \\)\\( \\mathcal{O}(d|C|) \\)\\( \\mathcal{O}(EN_{\\alpha}|C|) \\)5.637.653.747.335.963.133.63.082.518.968.558.02
\\( \\mathbf{CMLU}_{\\beta} \\)\\( \\mathcal{O}(d|C|) \\)\\( \\mathcal{O}(E|C|) \\)7.968.137.938.908.869.594.233.743.029.659.8310.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.318, + 0.907, + 0.348 + ], + "angle": 0, + "content": "Table 2. Experiments on large-scale dataset. We report fooling success rate (FR) and non-target flip rate \\((\\mathrm{NT}_R)\\) on OpenImages using ASL model. The experiments were performed on large target set sizes, up to \\(|\\Omega| = 20\\). Higher FR (\\(\\uparrow\\)) and lower \\(\\mathrm{NT}_R\\) (\\(\\downarrow\\)) are better." + }, + { + "type": "table", + "bbox": [ + 0.119, + 0.355, + 0.88, + 0.505 + ], + "angle": 0, + "content": "
MethodFR (↑)NTR(↓)
1234510152012345101520
Oracle97.697.996.897.995.796.396.896.80.640.800.870.930.961.051.171.47
Or-S97.629.646.281.700.420.00.00.00.640.710.650.340.20---
Or-C97.649.422.823.313.40.00.00.00.640.710.670.750.48---
SGA98.941.49.646.110.920.150.00.00.770.810.730.740.320.12--
NAG95.992.188.888.489.779.980.984.20.810.890.931.011.071.421.531.84
CMLUα92.266.744.046.239.439.932.432.30.720.760.830.890.830.860.820.77
CMLUβ97.395.895.996.895.294.896.595.20.720.800.880.920.981.061.281.67
" + }, + { + "type": "image", + "bbox": [ + 0.115, + 0.529, + 0.457, + 0.661 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.664, + 0.483, + 0.693 + ], + "angle": 0, + "content": "Figure 4. Visualizing label-wise universal perturbations learnt by Oracle and \\(\\mathrm{CMLU}_{\\beta}\\) on NUS-WIDE using ASL." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.717, + 0.483, + 0.822 + ], + "angle": 0, + "content": "timizes to find the best combination of the label-wise universal perturbations and hence, is better than the naive summation of individual perturbations. Across all experiments, \\(\\mathrm{CMLU}_{\\beta}\\) is the closest to Oracle performance. This is despite the fact that \\(\\mathrm{CMLU}_{\\beta}\\) has linear complexity whereas the Oracle has exponential complexity. Moreover, \\(\\mathrm{CMLU}_{\\alpha}\\) achieves higher \\(FR\\) and lower \\(NT_{R}\\) than all baselines." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Table 1 shows the percentage of non-targeted affected labels (whose prediction was changed after the attack). Although the baselines have low \\(\\mathrm{NT}_R\\) for target sizes greater than 1, they also have low fooling rates. Note that despite having linear complexity, the \\(\\mathrm{NT}_R\\) of \\(\\mathrm{CMLU}_{\\beta}\\) is only" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.53, + 0.907, + 0.667 + ], + "angle": 0, + "content": "slightly higher than Oracle. \\(\\mathrm{CMLU}_{\\alpha}\\) achieves lower \\(NT_{R}\\) as compared to \\(\\mathrm{CMLU}_{\\beta}\\) but still outperforms the baselines. In all experiments, NAG performs well (close to \\(\\mathrm{CMLU}_{\\alpha}\\)) but declines in performance as we increase the target size. It also has significantly higher \\(\\mathrm{NT}_{R}\\) than the Oracle and \\(\\mathrm{CMLU}_{\\beta}\\). This implies that NAG learnt to attack multiple classes together rather than learning the distribution of label-wise universal perturbations. Therefore, even though it has high success rate, it also affects several other labels." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.909, + 0.903 + ], + "angle": 0, + "content": "We further perform experiments on OpenImages, a large-scale dataset, and we show results in Tab. 2. Performance trends remain consistent with those observed on smaller datasets. The success rate of Or-S and Or-C consistently drops as we target more classes, and gets to 0 by target size 10. Note that NAG performs better in OpenImages experiments. As shown in the table, NAG affects a large percentage of non-targeted labels, which signifies that the method does not capture the class-specific directions and perturbs a large group of labels together. \\(\\mathrm{CMLU}_{\\alpha}\\) performs better than Or-C, Or-S, and SGA but worse than NAG. However, as shown in the table, \\(\\mathrm{CMLU}_{\\alpha}\\) still achieves significantly lower \\(NT_{R}\\) for large target set sizes. In all experiments, \\(\\mathrm{CMLU}_{\\beta}\\) achieves consistently close performance to the Oracle while also maintaining low \\(\\mathrm{NT}_{R}\\)." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20586" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.111, + 0.092, + 0.462, + 0.254 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.262, + 0.485, + 0.362 + ], + "angle": 0, + "content": "Figure 5. Comparison of the regions (convex cones) spanned by universal perturbations learnt using Oracle and \\(\\mathrm{CMLU}_{\\beta}\\) on NUS-WIDE (top) and MS-COCO (bottom). Each point in the heatmap is the average model confidence computed across images using the vector \\((\\epsilon_x * \\boldsymbol{u}_x + \\epsilon_y * \\boldsymbol{u}_y)\\) for all label combinations of size \\(|\\Omega| = 2\\), where \\(\\boldsymbol{u}_x, \\boldsymbol{u}_y\\) are universal vectors of classes in \\(\\Omega\\). The asterisks(*) show the region where the \\(L_{\\infty}\\) norm is 0.05." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.374, + 0.245, + 0.389 + ], + "angle": 0, + "content": "4.2.1. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.392, + 0.484, + 0.56 + ], + "angle": 0, + "content": "We perform ablation experiments on Oracle and \\(\\mathrm{CMLU}_{\\beta}\\) to investigate the effect of hyperparameters on their performance. Tab. 3 shows the results as we increase/decrease the values of \\(\\epsilon\\) and \\(\\kappa\\). Increasing \\(\\epsilon\\) provides a larger perturbation budget for better performance; however, to keep the changes to the inputs imperceptible, we set \\(\\epsilon = 0.05\\). Meanwhile, increasing \\(\\kappa\\) emphasizes maintaining non-targeted labels, which lowers \\(\\mathrm{NT}_R\\) but also reduces the success rate (FR). This is because many labels can be correlated in MLL and attacking a set of labels while fixing non-targeted but correlated labels lead to sub-optimal perturbations [63]." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.565, + 0.433, + 0.581 + ], + "angle": 0, + "content": "4.2.2. Analysis of Universal Adversarial Regions" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.584, + 0.483, + 0.75 + ], + "angle": 0, + "content": "The label-wise universal perturbations learnt by Oracle and \\(\\mathrm{CMLU}_{\\beta}\\) are visualized in Fig. 4. We quantitatively assess the visual impact of these perturbations using Structural Similarity Index (SSIM), computed between original and perturbed images. The average SSIM values across all images, as shown in Fig. 4, demonstrate comparable levels of distortion for Oracle and \\(\\mathrm{CMLU}_{\\beta}\\). Note that the pairwise dot product of the Oracle and CMLU vectors is \\(\\approx 0\\), indicating that \\(\\mathrm{CMLU}_{\\beta}\\) discover distinct attack directions, different from the Oracle. Additional analysis of pairwise dot products and visualizations are provided in supplementary." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.902 + ], + "angle": 0, + "content": "To further highlight and analyze the distinction between the perturbations learnt using different methods, we focus on Oracle and \\(\\mathrm{CMLU}_{\\beta}\\). Specifically, we examine the universal adversarial regions generated by those methods. Fig. 5 shows the average model confidence on the input space region spanned by the pairs of universal perturbation vectors. In the figure, perturbing inputs to blue and green regions would not change the model prediction (does not attack any class), yellow region would attack one class, and red region would attack both classes in \\(\\Omega\\). Note that there is" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.09, + 0.907, + 0.12 + ], + "angle": 0, + "content": "Table 3. Ablation experiments for hyperparameters on NUS-WIDE. The experiments are conducted for \\( |\\Omega| = 1 \\)." + }, + { + "type": "table", + "bbox": [ + 0.529, + 0.125, + 0.89, + 0.212 + ], + "angle": 0, + "content": "
εOracleCMLUβκOracleCMLUβ
FRNTRFRNTRFRNTRFRNTR
0.0163.55.3038.55.041599.16.5098.58.32
0.0394.96.2984.27.092098.96.1797.67.96
0.0598.96.1797.67.962593.54.7089.36.97
0.0799.75.1099.37.543093.14.4085.56.58
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.218, + 0.907, + 0.26 + ], + "angle": 0, + "content": "Table 4. Transferability Evaluation: \\( S \\rightarrow T \\) shows the performance of universal perturbations learnt on model \\( \\mathbf{S} \\) and evaluated on model \\( \\mathbf{T} \\). (A: ASL, D: ML-Decoder)" + }, + { + "type": "table", + "bbox": [ + 0.529, + 0.264, + 0.89, + 0.383 + ], + "angle": 0, + "content": "
Transferability of Universal Attacks
ModelsMethodNUS-WIDEMS-COCO
135135
A → DOracle90.183.777.334.610.65.40
CMLUβ89.978.475.442.917.66.4
D → AOracle24.61.330.064.428.319.6
CMLUβ33.50.160.064.834.521.2
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.394, + 0.907, + 0.606 + ], + "angle": 0, + "content": "no prominent red region (to attack both classes) in the cross-section of Oracle vectors. This explains the low success rate of Or-S and Or-C which optimize to find red region in such a cross-section. More details on computing these cones is provided in supplementary. We also investigate how well those red regions computed using \\(\\mathrm{CMLU}_{\\beta}\\) can transfer to other models. For that, we perform transferability attacks and the results are shown in Tab. 4. From the table, we observe that \\(\\mathrm{CMLU}_{\\beta}\\) (linear complexity) achieves fooling success rates closer and often better than Oracle (exponential complexity). This shows that the convex cone regions corresponding to different labels are often common across models. This can be effectively used to generate black-box attacks. We hope that future works explore this further." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.619, + 0.634, + 0.635 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.643, + 0.907, + 0.797 + ], + "angle": 0, + "content": "We addressed the problem of generating targeted multi-label universal perturbations. We developed a compositional framework to generate exponential number of multi-label universal perturbations in constant time. The independence assumption on label-wise universal perturbations naturally leads to a formulation that requires learning convex cones. Through extensive experiments on multiple datasets and models, we show that our method outperforms all baseline methods and achieves performance comparable to the Oracle (an exponential complexity method)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.809, + 0.704, + 0.827 + ], + "angle": 0, + "content": "6. Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.833, + 0.907, + 0.896 + ], + "angle": 0, + "content": "This work is sponsored by ARPA-H (1AY2AX000062), DARPA PTG (HR00112220001), NSF (IIS-2115110) and ARO (W911NF2110276). Content does not necessarily reflect the position/policy of the Government." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20587" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.115, + 0.484, + 0.184 + ], + "angle": 0, + "content": "[1] Abhishek Aich, Calvin-Khang Ta, Akash Gupta, Chengyu Song, Srikanth Krishnamurthy, Salman Asif, and Amit Roy-Chowdhury. Gama: Generative adversarial multi-object scene attacks. Advances in Neural Information Processing Systems, 35:36914-36930, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.186, + 0.484, + 0.269 + ], + "angle": 0, + "content": "[2] Abhishek Aich, Shasha Li, Chengyu Song, M Salman Asif, Srikanth V Krishnamurthy, and Amit K Roy-Chowdhury. Leveraging local patch differences in multi-object scenes for generative adversarial attacks. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1308-1318, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.271, + 0.483, + 0.299 + ], + "angle": 0, + "content": "[3] N. Akhtar and A. Mian. Threat of adversarial attacks on deep learning in computer vision: A survey. arXiv, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.301, + 0.482, + 0.342 + ], + "angle": 0, + "content": "[4] A. Athalye, N. Carlini, and D. A. Wagner. Obfuscated gradients give a false sense of security: Circumventing defenses to adversarial examples. 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.344, + 0.482, + 0.385 + ], + "angle": 0, + "content": "[5] Yuanhao Ban and Yinpeng Dong. Pre-trained adversarial perturbations. In Advances in Neural Information Processing Systems, pages 1196-1209, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.387, + 0.482, + 0.442 + ], + "angle": 0, + "content": "[6] Emanuel Ben-Baruch, Tal Ridnik, Itamar Friedman, Avi Ben-Cohen, Nadav Zamir, Asaf Noy, and Lihi Zelnik-Manor. Multi-label classification with partial annotations using class-aware selective loss. 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.444, + 0.482, + 0.485 + ], + "angle": 0, + "content": "[7] P. Benz, C. Zhang, T. Imtiaz, and I.-S. Kweon. Double targeted universal adversarial perturbations. Asian Conference on Computer Vision, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.487, + 0.482, + 0.543 + ], + "angle": 0, + "content": "[8] Philipp Benz, Chaoning Zhang, Adil Karjauv, and In So Kweon. Universal adversarial training with class-wise perturbations. In 2021 IEEE International Conference on Multimedia and Expo (ICME), pages 1-6. IEEE, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.545, + 0.482, + 0.585 + ], + "angle": 0, + "content": "[9] N. Carlini and D. Wagner. Adversarial examples are not easily detected: Bypassing ten detection methods. Workshop on Artificial Intelligence and Security, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.588, + 0.483, + 0.628 + ], + "angle": 0, + "content": "[10] N. Carlini and D. Wagner. Towards evaluating the robustness of neural networks. IEEE Symposium on Security and Privacy, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.631, + 0.482, + 0.672 + ], + "angle": 0, + "content": "[11] Y. Carmon, A. Raghunathan, L. Schmidt, P. Liang, and J. C. Duchi. Unlabeled data improves adversarial robustness. Neural Information Processing Systems, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.674, + 0.482, + 0.728 + ], + "angle": 0, + "content": "[12] P.-Y. Chen, Y. Sharma, H. Zhang, J. Yi, and C.-J. Hsieh. Ead: Elastic-net attacks to deep neural networks via adversarial examples. AAAI Conference on Artificial Intelligence, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.731, + 0.482, + 0.786 + ], + "angle": 0, + "content": "[13] T. Chen, M. Xu, X. Hui, H. Wu, and L. Lin. Learning semantic-specific graph representation for multi-label image recognition. IEEE International Conference on Computer Vision, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.788, + 0.482, + 0.843 + ], + "angle": 0, + "content": "[14] Zhao-Min Chen, Xiu-Shen Wei, Xin Jin, and Yanwen Guo. Multi-label image recognition with joint class-aware map disentangling and label correlation embedding. IEEE International Conference on Multimedia and Expo, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.845, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[15] Z. M. Chen, X. S. Wei, P. Wang, and Y. Guo. Multi-label image recognition with graph convolutional networks. IEEE Conference on Computer Vision and Pattern Recognition, abs/1904.03582, 2019. 1" + }, + { + "type": "list", + "bbox": [ + 0.102, + 0.115, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[16] T. S. Chua, J. Tang, R. Hong, H. Li, Z. Luo, and Y. T. Zheng. Nus-wide: A real-world web image database from national university ofsingapore.ACM International Conference on Image and Video Retrieval, 2009.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.149, + 0.905, + 0.188 + ], + "angle": 0, + "content": "[17] S. D. Dao, E. Zhao, D. Phung, and J. Cai. Multi-label image classification with contrastive learning. arXiv preprint, arXiv:2107.11626, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.19, + 0.905, + 0.245 + ], + "angle": 0, + "content": "[18] J. Deng, N. Ding, Y. Jia, A. Frome, K. Murphy, S. Bengio, Y. Li, H. Neven, and H. Adam. Large-scale object classification using label relation graphs. European Conference on Computer Vision, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.247, + 0.905, + 0.3 + ], + "angle": 0, + "content": "[19] Gavin Weiguang Ding, Yash Sharma, Kry Yik Chau Lui, and Ruitong Huang. Mma training: Direct input space margin maximization through adversarial training. arXiv preprint arXiv:1812.02637, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.302, + 0.905, + 0.357 + ], + "angle": 0, + "content": "[20] Zixuan Ding, Ao Wang, Hui Chen, Qiang Zhang, Pengzhang Liu, Yongjun Bao, Weipeng Yan, and Jungong Han. Exploring structured semantic prior for multi label recognition with incomplete labels. 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.358, + 0.905, + 0.44 + ], + "angle": 0, + "content": "[21] Junhao Dong, Seyed-Mohsen Moosavi-Dezfooli, Jianhuang Lai, and Xiaohua Xie. The enemy of my enemy is my friend: Exploring inverse adversaries for improving adversarial training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 24678–24687, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.442, + 0.905, + 0.495 + ], + "angle": 0, + "content": "[22] Yinpeng Dong, Zhijie Deng, Tianyu Pang, Jun Zhu, and Hang Su. Adversarial distributional training for robust deep learning. Advances in Neural Information Processing Systems, 33:8270-8283, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.497, + 0.905, + 0.536 + ], + "angle": 0, + "content": "[23] L. Feng, B. An, and S. He. Collaboration based multi-label learning. AAAI Conference on Artificial Intelligence, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.539, + 0.905, + 0.592 + ], + "angle": 0, + "content": "[24] Zongyuan Ge, Dwarikanath Mahapatra, Suman Sedai, Rahul Garnavi, and Rajib Chakravorty. Chest x-rays classification: A multi-label and fine-grained problem. arXiv, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.595, + 0.905, + 0.636 + ], + "angle": 0, + "content": "[25] I. J. Goodfellow, J. Shlens, and C. Szegedy. Explaining and harnessing adversarial examples. International Conference on Learning Representations, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.637, + 0.905, + 0.704 + ], + "angle": 0, + "content": "[26] Emily Hand, Carlos Castillo, and Rama Chellappa. Doing the best we can with what we have: Multi-label balancing with selective learning for attribute prediction. In Proceedings of the AAAI Conference on Artificial Intelligence, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.707, + 0.905, + 0.747 + ], + "angle": 0, + "content": "[27] J. Hayes and G. Danezis. Learning universal adversarial perturbations with generative models. IEEE Security and Privacy Workshops, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.749, + 0.905, + 0.801 + ], + "angle": 0, + "content": "[28] W. He, J. Wei, X. Chen, N. Carlini, and D. Song. Adversarial example defense: Ensembles of weak defenses are not strong. USENIX Workshop on Offensive Technologies, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.804, + 0.905, + 0.831 + ], + "angle": 0, + "content": "[29] D. Hendrycks, K. Lee, and M. Mazeika. Using pre-training can improve model robustness and uncertainty. 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.833, + 0.905, + 0.899 + ], + "angle": 0, + "content": "[30] Dan Hendrycks, Kevin Zhao, Steven Basart, Jacob Steinhardt, and Dawn Song. Natural adversarial examples. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15262-15271, 2021. 1" + }, + { + "type": "list", + "bbox": [ + 0.523, + 0.093, + 0.905, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20588" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.092, + 0.486, + 0.176 + ], + "angle": 0, + "content": "[31] Lei Hsiung, Yun-Yun Tsai, Pin-Yu Chen, and Tsung-Yi Ho. Towards compositional adversarial robustness: Generalizing adversarial training to composite semantic perturbations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24658-24667, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.177, + 0.484, + 0.233 + ], + "angle": 0, + "content": "[32] Shu Hu, Lipeng Ke, Xin Wang, and Siwei Lyu. Tkmlap: Adversarial attacks to top-k multi-label learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7649-7657, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.234, + 0.484, + 0.289 + ], + "angle": 0, + "content": "[33] D. Huynh and E. Elhamifar. Interaction compass: Multi-label zero-shot learning of human-object interactions via spatial relations. International Conference on Computer Vision, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.29, + 0.484, + 0.332 + ], + "angle": 0, + "content": "[34] D. T. Huynh and E. Elhamifar. Interactive multi-label cnn learning with partial labels. IEEE Conference on Computer Vision and Pattern Recognition, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.334, + 0.483, + 0.389 + ], + "angle": 0, + "content": "[35] Tooba Imtiaz, Morgan Kohler, Jared Miller, Zifeng Wang, Mario Sznaier, Octavia I Camps, and Jennifer G Dy. Saif: Sparse adversarial and interpretable attack framework. arXiv preprint arXiv:2212.07495, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.39, + 0.483, + 0.432 + ], + "angle": 0, + "content": "[36] J. Li R. Ji, H. Liu, X. Hong, Y. Gao, and Q. Tian. Universal perturbation attack against image retrieval. International Conference on Computer Vision, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.433, + 0.483, + 0.489 + ], + "angle": 0, + "content": "[37] Jinyuan Jia, Wenjie Qu, and Neil Zhenqiang Gong. Multiguard: Provably robust multi-label classification against adversarial examples. Advances in Neural Information Processing Systems, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.49, + 0.483, + 0.532 + ], + "angle": 0, + "content": "[38] V. Khrulkov and I. Oseledets. Art of singular vectors and universal adversarial perturbations. IEEE Conference on Computer Vision and Pattern Recognition, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.533, + 0.483, + 0.574 + ], + "angle": 0, + "content": "[39] Youngwook Kim, Jae Myung Kim, Zeynep Akata, and Jungwoo Lee. Large loss matters in weakly supervised multi-label classification. 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.576, + 0.461, + 0.589 + ], + "angle": 0, + "content": "[40] Takumi Kobayashi. Two-way multi-label loss. 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.591, + 0.483, + 0.632 + ], + "angle": 0, + "content": "[41] Linghao Kong, Wenjian Luo, Hongwei Zhang, Yang Liu, and Yuhui Shi. Evolutionary multilabel adversarial examples: An effective black-box attack. 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.633, + 0.483, + 0.675 + ], + "angle": 0, + "content": "[42] Linghao Kong, Wenjian Luo, Zipeng Ye, Qi Zhou, and Yan Jia. Multi-label black-box adversarial attacks only with predicted labels. 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.676, + 0.483, + 0.717 + ], + "angle": 0, + "content": "[43] A. Kurakin, I. Goodfellow, and S. Bengio. Adversarial machine learning at scale. International Conference on Learning Representations, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.718, + 0.483, + 0.802 + ], + "angle": 0, + "content": "[44] A. Kuznetsova, H. Rom, N. Alldrin, J. Uijlings, I. Krasin, J. Pont-Tuset, S. Kamali, S. Popov, M. Malloci, A. Kolesnikov, T. Duerig, and V. Ferrari. The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale. International Journal of Computer Vision, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.803, + 0.483, + 0.857 + ], + "angle": 0, + "content": "[45] J. Lanchantin, T. Wang, V. Ordonez, and Y. Qi. General multi-label image classification with transformers. IEEE Conference on Computer Vision and Pattern Recognition, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.859, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[46] K. Lee, K. Lee, H. Lee, and J. Shin. A simple unified framework for detecting out-of-distribution samples and adversarial attacks. 2018. 2" + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.092, + 0.486, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.093, + 0.905, + 0.163 + ], + "angle": 0, + "content": "[47] Guofa Li, Zefeng Ji, Yunlong Chang, Shen Li, Xingda Qu, and Dongpu Cao. ML-anet: A transfer learning approach using adaptation network for multi-label image classification in autonomous driving. Chinese Journal of Mechanical Engineering, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.164, + 0.905, + 0.22 + ], + "angle": 0, + "content": "[48] Maosen Li, Yanhua Yang, Kun Wei, Xu Yang, and Heng Huang. Learning universal adversarial perturbation by adversarial example. AAAI Conference on Artificial Intelligence, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.222, + 0.905, + 0.264 + ], + "angle": 0, + "content": "[49] Peng Li, Peng Chen, Yonghong Xie, and Dezheng Zhang. Bi-modal learning with channel-wise attention for multi-label image classification. IEEE Access, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.265, + 0.905, + 0.308 + ], + "angle": 0, + "content": "[50] Q. Li, M. Qiao, W. Bian, and D. Tao. Conditional graphical lasso for multi-label image classification. IEEE Conference on Computer Vision and Pattern Recognition, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.31, + 0.905, + 0.352 + ], + "angle": 0, + "content": "[51] Q. Li, X. Peng, Y. Qiao, and Q. Peng. Learning label correlations for multi-label image recognition with graph networks. Pattern Recognition Letters, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.353, + 0.905, + 0.395 + ], + "angle": 0, + "content": "[52] X. Li, F. Zhao, and Y. Guo. Multi-label image classification with a probabilistic label enhancement model. In UAI, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.397, + 0.905, + 0.439 + ], + "angle": 0, + "content": "[53] Y. Li, Y. Song, and J. Luo. Improving pairwise ranking for multi-label image classification. IEEE Conference on Computer Vision and Pattern Recognition, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.441, + 0.905, + 0.484 + ], + "angle": 0, + "content": "[54] Z. Li, W. Lu, Z. Sun, and W. Xing. Improving multi-label classification using scene cues. Multimedia Tools and Applications, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.485, + 0.905, + 0.514 + ], + "angle": 0, + "content": "[55] Dekun Lin. Probability guided loss for long-tailed multi-label image classification. 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.515, + 0.905, + 0.583 + ], + "angle": 0, + "content": "[56] Tsung-Yi Lin, Michael Maire, Serge J. Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C. Lawrence Zitnick. Microsoft coco: Common objects in context. European Conference on Computer Vision, 2014. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.586, + 0.905, + 0.642 + ], + "angle": 0, + "content": "[57] An-An Liu, Zhuang Shao, Yongkang Wong, Junnan Li, Su Yu-Ting, and Mohan Kankanhalli. Lstm-based multi-label video event detection. Multimedia Tools and Applications, 78, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.644, + 0.905, + 0.687 + ], + "angle": 0, + "content": "[58] Shilong Liu, Lei Zhang, Xiao Yang, Hang Su, and Jun Zhu. Query2label: A simple transformer way to multi-label classification. arXiv preprint arXiv:2107.10834, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.688, + 0.905, + 0.758 + ], + "angle": 0, + "content": "[59] Xuannan Liu, Yaoyao Zhong, Yuhang Zhang, Lixiong Qin, and Weihong Deng. Enhancing generalization of universal adversarial perturbation through gradient aggregation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.76, + 0.905, + 0.842 + ], + "angle": 0, + "content": "[60] Yiran Liu, Xin Feng, Yunlong Wang, Wu Yang, and Di Ming. Trm-uap: Enhancing the transferability of data-free universal adversarial perturbation via truncated ratio maximization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4762-4771, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.845, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[61] Mingzhi Ma, Weijie Zheng, Wanli Lv, Lu Ren, Hang Su, and Zhaoxia Yin. Multi-label adversarial attack based on label correlation. In 2023 IEEE International Conference on Image Processing (ICIP), 2023. 3" + }, + { + "type": "list", + "bbox": [ + 0.523, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "20589" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.092, + 0.486, + 0.147 + ], + "angle": 0, + "content": "[62] A. Madry, A. Makelov, L. Schmidt, D. Tsipras, and A. Vladu. Towards deep learning models resistant to adversarial attacks. International Conference on Learning Representations, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.149, + 0.483, + 0.205 + ], + "angle": 0, + "content": "[63] Hassan Mahmood and Ehsan Elhamifar. Semantic-aware multi-label adversarial attacks. IEEE Conference on Computer Vision and Pattern Recognition, pages 24251-24262, 2024. 1, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.207, + 0.483, + 0.247 + ], + "angle": 0, + "content": "[64] S. Melacci, G. Ciravegna, A. Sotgiu, A. Demontis, B. Biggio, M. Gori, and F. Roli. Domain knowledge alleviates adversarial attacks in multi-label classifiers. 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.249, + 0.483, + 0.304 + ], + "angle": 0, + "content": "[65] J.-H. Metzen, M.-C. Kumar, T. Brox, and V. Fischer. Universal adversarial perturbations against semantic image segmentation. International Conference on Computer Vision, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.305, + 0.483, + 0.36 + ], + "angle": 0, + "content": "[66] S. Moosavi-Dezfooli, A. Fawzi, and P. Frossard. Deepfool: a simple and accurate method to fool deep neural networks. IEEE Conference on Computer Vision and Pattern Recognition, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.362, + 0.483, + 0.404 + ], + "angle": 0, + "content": "[67] S. Moosavi-Dezfooli, A. Fawzi, O. Fawzi, and P. Frossard. Universal adversarial perturbations. IEEE Conference on Computer Vision and Pattern Recognition, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.406, + 0.483, + 0.46 + ], + "angle": 0, + "content": "[68] Konda Reddy Mopuri, Utsav Garg, and R Venkatesh Babu. Fast feature fool: A data independent approach to universal adversarial perturbations. arXiv preprint arXiv:1707.05572, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.462, + 0.483, + 0.518 + ], + "angle": 0, + "content": "[69] Konda Reddy Mopuri, Utkarsh Ojha, Utsav Garg, and R Venkatesh Babu. Nag: Network for adversary generation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 742-751, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.519, + 0.483, + 0.574 + ], + "angle": 0, + "content": "[70] J. Nam, E. L. Mencia, H. J. Kim, and J. Furnkranz. Maximizing subset accuracy with recurrent neural networks in multi-label classification. Neural Information Processing Systems, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.576, + 0.483, + 0.617 + ], + "angle": 0, + "content": "[71] T. Pang, K. Xu, C. Du, N. Chen, and J. Zhu. Improving adversarial robustness via promoting ensemble diversity. International Conference on Machine learning, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.618, + 0.483, + 0.659 + ], + "angle": 0, + "content": "[72] T. Pang, K. Xu, Y. Dong, C. Du, N. Chen, and J. Zhu. Rethinking softmax cross-entropy loss for adversarial robustness. arXiv, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.661, + 0.483, + 0.703 + ], + "angle": 0, + "content": "[73] T. Pang, X. Yang, Y. Dong, K. Xu, H. Su, and J. Zhu. Boosting adversarial training with hypersphere embedding. Neural Information Processing Systems, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.704, + 0.483, + 0.773 + ], + "angle": 0, + "content": "[74] Nicolas Papernot, Patrick Mcdaniel, Somesh Jha, Matt Fredrikson, Z. Berkay Celik, and Ananthram Swami. The limitations of deep learning in adversarial settings. IEEE European Symposium on Security and Privacy (EuroS&P), 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.775, + 0.483, + 0.844 + ], + "angle": 0, + "content": "[75] Ioannis Partalas, Aris Kosmopoulos, Nicolas Baskiotis, Thierry Artieres, George Paliouras, Eric Gaussier, Ion Androutsopoulos, Massih-Reza Amini, and Patrick Galinari. Lsjtc: A benchmark for large-scale text classification. arXiv preprint arXiv:1503.08581, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.846, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[76] Omid Poursaeed, Isay Katsman, Bicheng Gao, and Serge Belongie. Generative adversarial perturbations. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4422-4431, 2018. 2" + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.092, + 0.486, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.093, + 0.906, + 0.147 + ], + "angle": 0, + "content": "[77] Tao Pu, Tianshui Chen, Hefeng Wu, and Liang Lin. Semantic-aware representation blending for multi-label image recognition with partial labels. Proceedings of the AAAI Conference on Artificial Intelligence, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.149, + 0.906, + 0.217 + ], + "angle": 0, + "content": "[78] Zeyu Qin, Yanbo Fan, Yi Liu, Li Shen, Yong Zhang, Jue Wang, and Baoyuan Wu. Boosting the transferability of adversarial attacks with reverse adversarial perturbation. Advances in Neural Information Processing Systems, 35: 29845-29858, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.218, + 0.906, + 0.288 + ], + "angle": 0, + "content": "[79] Tal Ridnik, Emanuel Ben-Baruch, Nadav Zamir, Asaf Noy, Itamar Friedman, Matan Protter, and Lihi Zelnik-Manor. Asymmetric loss for multi-label classification. In Proceedings of the IEEE/CVF international conference on computer vision, pages 82-91, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.289, + 0.906, + 0.355 + ], + "angle": 0, + "content": "[80] Tal Ridnik, Gilad Sharir, Avi Ben-Cohen, Emanuel Ben-Baruch, and Asaf Noy. ML-decoder: Scalable and versatile classification head. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 32-41, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.357, + 0.906, + 0.44 + ], + "angle": 0, + "content": "[81] Jérôme Rony, Luiz G Hafemann, Luiz S Oliveira, Ismail Ben Ayed, Robert Sabourin, and Eric Granger. Decoupling direction and norm for efficient gradient-based 12 adversarial attacks and defenses. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4322-4330, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.441, + 0.906, + 0.482 + ], + "angle": 0, + "content": "[82] J. Cohen and E. Rosenfeld and Z. Kolter. Certified adversarial robustness via randomized smoothing. International Conference on Machine learning, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.483, + 0.906, + 0.537 + ], + "angle": 0, + "content": "[83] A. Shafahi, M. Najibi, A. Ghiasi, Z. Xu, J. Dickerson, C. Studer, L. Davis, G. Taylor, and T. Goldstein. Adversarial training for free! Neural Information Processing Systems, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.538, + 0.906, + 0.607 + ], + "angle": 0, + "content": "[84] Nasim Shafiee and Ehsan Elhamifar. Zero-shot attribute attacks on fine-grained recognition models. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part V, pages 262-282. Springer, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.608, + 0.906, + 0.663 + ], + "angle": 0, + "content": "[85] Jing Shao, Kai Kang, Chen Change Loy, and Xiaogang Wang. Deeply learned attributes for crowded scene understanding. In 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.664, + 0.906, + 0.718 + ], + "angle": 0, + "content": "[86] Nitish Shukla and Sudipta Banerjee. Generating adversarial attacks in the latent space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 730-739, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.719, + 0.906, + 0.761 + ], + "angle": 0, + "content": "[87] Q. Song, H. Jin, X. Huang, and X. Hu. Multi-label adversarial perturbations. IEEE International Conference on Data Mining, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.761, + 0.906, + 0.803 + ], + "angle": 0, + "content": "[88] Fengguang Su, Ou Wu, and Weiyao Zhu. Multi-label adversarial attack with new measures and self-paced constraint weighting. 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.804, + 0.906, + 0.858 + ], + "angle": 0, + "content": "[89] C. Szegedy, W. Zaremba, I. Sutskever, J. Bruna, D. Erhan, I. Goodfellow, and R. Fergus. Intriguing properties of neural networks. International Conference on Learning Representations, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.859, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[90] G. Tsoumakas and I. Katakis. Multi-label classification: An overview. International Journal Data Warehousing and Mining, 3, 2007. 2" + }, + { + "type": "list", + "bbox": [ + 0.523, + 0.093, + 0.906, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "20590" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.093, + 0.482, + 0.12 + ], + "angle": 0, + "content": "[91] N. Tursynbek, A. Petiushko, and I. Oseledets. Geometry-inspired top-k adversarial perturbations. arXiv, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.121, + 0.482, + 0.175 + ], + "angle": 0, + "content": "[92] J. Uesato, J. B. Alayrac, P. Huang, R. Stanforth, A. Fawzi, and P. Kohli. Are labels required for improving adversarial robustness? Neural Information Processing Systems, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.178, + 0.483, + 0.232 + ], + "angle": 0, + "content": "[93] Thomas Verelst, Paul K Rubenstein, Marcin Eichner, Tinne Tuytelaars, and Maxim Berman. Spatial consistency loss for training multi-label classifiers from single-label annotations. 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.233, + 0.483, + 0.288 + ], + "angle": 0, + "content": "[94] J. Wang, Y. Yang, J. Mao, Z. Huang, C. Huang, and W. Xu. Cnn-rnn: A unified framework for multi-label image classification. 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.289, + 0.483, + 0.356 + ], + "angle": 0, + "content": "[95] Ya Wang, Dongliang He, Fu Li, Xiang Long, Zhichao Zhou, Jinwen Ma, and Shilei Wen. Multi-label classification with label graph superimposing. In Proceedings of the AAAI conference on artificial intelligence, pages 12265-12272, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.358, + 0.483, + 0.412 + ], + "angle": 0, + "content": "[96] Z. Wang, T. Chen, G. Li, G. Li, and L. Lin. Multi-label image recognition by recurrently discovering attentional regions. IEEE International Conference on Computer Vision, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.413, + 0.483, + 0.482 + ], + "angle": 0, + "content": "[97] Yunchao Wei, Wei Xia, Min Lin, Junshi Huang, Bingbing Ni, Jian Dong, Yao Zhao, and Shuicheng Yan. Hcp: A flexible cnn framework for multi-label image classification. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.484, + 0.483, + 0.538 + ], + "angle": 0, + "content": "[98] Y. Wu, H. Liu, S. Feng, Y. Jin, G. Lyu, and Z. Wu. Gmmlic: Graph matching based multi-label image classification. International Joint Conference on Artificial Intelligence, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.539, + 0.483, + 0.594 + ], + "angle": 0, + "content": "[99] C. Xie, Z. Zhang, Y. Zhou, S. Bai, J. Wang, Z. Ren, and A. Yuille. Improving transferability of adversarial examples with input diversity. IEEE Conference on Computer Vision and Pattern Recognition, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.595, + 0.483, + 0.65 + ], + "angle": 0, + "content": "[100] Ming-Kun Xie, Jiahao Xiao, and Sheng-Jun Huang. Label-aware global consistency for multi-label learning with single positive labels. Advances in Neural Information Processing Systems, 35:18430-18441, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.651, + 0.483, + 0.692 + ], + "angle": 0, + "content": "[101] J. Xu, H. Tian, Z. Wang, Y. Wang, W. Kang, and F. Chen. Joint input and output space learning for multi-label image classification. IEEE Transactions on Multimedia, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.692, + 0.483, + 0.734 + ], + "angle": 0, + "content": "[102] W. Xu, D. Evans, and Y. Qi. Feature squeezing: Detecting adversarial examples in deep neural networks. Network and Distributed Systems Security Symposium, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.735, + 0.483, + 0.79 + ], + "angle": 0, + "content": "[103] Zhuo Yang, Yufei Han, and Xiangliang Zhang. Characterizing the evasion attackability of multi-label classifiers. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 10647-10655, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.79, + 0.483, + 0.845 + ], + "angle": 0, + "content": "[104] Zhuo Yang, Yufei Han, and Xiangliang Zhang. Attack transferability characterization for adversarially robust multi-label classification. In Machine Learning and Knowledge Discovery in Databases., pages 397-413, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[105] V. O. Yazici, A. Gonzalez-Garcia, A. Ramisa, B. Twardowski, and J. v. d. Weijer. Orderless recurrent models for multi-label classification. European Conference on Computer Vision, 2020. 2" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.093, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[106] J. Ye, J. He, X. Peng, W. Wu, and Y. Qiao. Attention-driven dynamic graph convolutional network for multi-label image recognition. European Conference on Computer Vision, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.149, + 0.905, + 0.205 + ], + "angle": 0, + "content": "[107] R. You, Z. Guo, L. Cui, X. Long, S. Y. Bao, and S. Wen. Cross-modality attention with semantic graph embedding for multi-label classification. AAAI Conference on Artificial Intelligence, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.207, + 0.905, + 0.275 + ], + "angle": 0, + "content": "[108] Hong-Xing Yu, Wei-Shi Zheng, Ancong Wu, Xiaowei Guo, Shaogang Gong, and Jian-Huang Lai. Unsupervised person re-identification by soft multilabel learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2148–2157, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.277, + 0.905, + 0.332 + ], + "angle": 0, + "content": "[109] Chaoning Zhang, Philipp Benz, Tooba Imtiaz, and In-So Kweon. Cd-uap: Class discriminative universal adversarial perturbation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 6754–6761, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.334, + 0.905, + 0.403 + ], + "angle": 0, + "content": "[110] Chaoning Zhang, Philipp Benz, Tooba Imtiaz, and In So Kweon. Understanding adversarial examples from the mutual influence of images and perturbations. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.404, + 0.905, + 0.473 + ], + "angle": 0, + "content": "[111] Chaoning Zhang, Philipp Benz, Adil Karjauv, and In So Kweon. Universal adversarial perturbations through the lens of deep steganography: Towards a fourier perspective. In Proceedings of the AAAI conference on artificial intelligence, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.475, + 0.905, + 0.529 + ], + "angle": 0, + "content": "[112] ML. Zhang and Z. Zhou. Multilabel neural networks with applications to functional genomics and text categorization. IEEE Transactions on Knowledge and Data Engineering, 2006. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.531, + 0.905, + 0.598 + ], + "angle": 0, + "content": "[113] Shu Zhang, Ran Xu, Caiming Xiong, and Chetan Ramaiah. Use all the labels: A hierarchical multi-label contrastive learning framework. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.601, + 0.905, + 0.685 + ], + "angle": 0, + "content": "[114] Zhe Zhao, Guangke Chen, Jingyi Wang, Yiwei Yang, Fu Song, and Jun Sun. Attack as defense: characterizing adversarial examples using robustness. In Proceedings of the 30th ACM SIGSOFT International Symposium on Software Testing and Analysis, page 42-55. Association for Computing Machinery, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.686, + 0.905, + 0.741 + ], + "angle": 0, + "content": "[115] Donghao Zhou, Pengfei Chen, Qiong Wang, Guangyong Chen, and Pheng-Ann Heng. Acknowledging the unknown for multi-label learning with single positive labels. European Conference on Computer Vision, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.742, + 0.905, + 0.797 + ], + "angle": 0, + "content": "[116] N. Zhou, W. Luo, X. Lin, P. Xu, and Z.. Zhang. Generating multi-label adversarial examples by linear programming. International Joint Conference on Neural Networks, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.799, + 0.905, + 0.854 + ], + "angle": 0, + "content": "[117] N. Zhou, W. Luo, J. Zhang, L. Kong, and H. Zhang. Hiding all labels for multi-label images: An empirical study of adversarial examples. International Joint Conference on Neural Networks, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.855, + 0.905, + 0.898 + ], + "angle": 0, + "content": "[118] Y. Zhu, J. T. Kwok, and Z. Zhou. Multi-label learning with global and local label correlation. IEEE Transactions on Knowledge and Data Engineering, 2018. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.898 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "20591" + } + ] +] \ No newline at end of file diff --git a/2025/Compositional Targeted Multi-Label Universal Perturbations/616cefba-8863-49bf-987c-2e525bd606b6_origin.pdf b/2025/Compositional Targeted Multi-Label Universal Perturbations/616cefba-8863-49bf-987c-2e525bd606b6_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..8c359e5b37353241c478b4f68ee6d34595755f9b --- /dev/null +++ b/2025/Compositional Targeted Multi-Label Universal Perturbations/616cefba-8863-49bf-987c-2e525bd606b6_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e3400521ebcbbdd042cab66231b29c6539288cf9a5cea5c6647a0312350664e4 +size 6684856 diff --git a/2025/Compositional Targeted Multi-Label Universal Perturbations/full.md b/2025/Compositional Targeted Multi-Label Universal Perturbations/full.md new file mode 100644 index 0000000000000000000000000000000000000000..6b1ef8f8927ff7484613e014c0d2bdf99abf7b45 --- /dev/null +++ b/2025/Compositional Targeted Multi-Label Universal Perturbations/full.md @@ -0,0 +1,437 @@ +# Compositional Targeted Multi-Label Universal Perturbations + +Hassan Mahmood Northeastern University + +mahmood.h@northeastern.edu + +Ehsan Elhamifar Northeastern University e.elhamifar@northeastern.edu + +# Abstract + +Generating targeted universal perturbations for multi-label recognition is a combinatorially hard problem that requires exponential time and space complexity. To address the problem, we propose a compositional framework. We show that a simple independence assumption on label-wise universal perturbations naturally leads to an efficient optimization that requires learning affine convex cones spanned by label-wise universal perturbations, significantly reducing the problem complexity to linear time and space. During inference, the framework allows generating universal perturbations for novel combinations of classes in constant time. We demonstrate the scalability of our method on large datasets and target sizes, evaluating its performance on NUS-WIDE, MS-COCO, and OpenImages using state-of-the-art multi-label recognition models. Our results show that our approach outperforms baselines and achieves results comparable to methods with exponential complexity. The code is available at https://github.com/hassanmahmood/UMLLAttacks.git + +# 1. Introduction + +Deep neural networks have been very successful at image-recognition, demonstrating high accuracy in various applications. However, they are also vulnerable to adversarial attacks [9, 25, 30, 31, 43, 74, 84, 86], where small changes to the input image can cause the model to misclassify the image. Notably, Moosavi-Dezfooli et al. [67] discovered that a single perturbation, called universal adversarial perturbation (UAP), can be crafted to misclassify a large number of images. This motivated a large body of work on generating universal adversarial attacks [7, 36, 38, 65, 69, 110]. + +Existing works on adversarial attacks have mainly focused on multi-class recognition (MCR) whose goal is to identify a single class in an image. On the other hand, many real-world applications require identifying all classes in an image e.g., document classification [75], scene understanding [85], facial attribute recognition [26], image annotation [15], human-object interaction [33], and several safety and security-critical applications e.g., recogniz- + +![](images/18ff0e158377b42c24a4be820811f27dd53d336e0bef7b03aca80296df8ac134.jpg) +Figure 1. Targeted Multi-Label Universal Perturbations. Our framework allows composing (shown as $\oplus$ ) different label-specific universal perturbations ('Animal', 'Person', 'Vehicle') to attack multiple classes together. These perturbations, once added (shown as $\bullet$ ) to an image, cause target labels to be misclassified. + +ing diseases [24], autonomous driving [47], and surveillance [57, 108]. Therefore, the goal of multi-label learning (MLL) is to identify all labels present in an image [14, 20, 23, 39, 40, 55, 77, 100, 115]. Given the increasing number of applications relying on MLL, it is crucial to understand the internal workings of the MLL models, identify their vulnerabilities, and make them robust. Recently, MLL models have been shown to be vulnerable to adversarial attacks [1, 2, 42, 63, 88], yet the vulnerability of MLL systems against universal perturbations has not been explored. + +Crafting universal perturbations for MCR and MLL have fundamental differences due to the distinct structuring of decision boundaries in these problems. In MCR, a universal perturbation to turn off a label in images will result in the prediction of another label. Conversely, in MLL, multiple labels can co-exist, and the objective of an attack might be to misclassify any subset of these labels either by turning the labels off or on (see Fig. 2). In MLL, any combination of labels can be present in images. For the label set $\mathcal{C}$ , there are $2^{|C|}$ possible combinations of labels. Universal pertur + +bations identify specific directions in the input space where the classifier is consistently vulnerable to attacks. What makes this particularly concerning in multi-label learning is that there can be exponentially many directions in the input space to attack the model - corresponding to different combinations of labels. + +One approach for generating universal perturbations for MLL is to generate universal vectors that flip the prediction of all or most labels. However, in practice, the adversary usually seeks to modify a specific small set of labels while keeping the prediction of others unchanged (see Fig. 1). This is called the targeted attack and requires learning $2^{|\mathcal{C}|}$ distinct universal perturbation vectors. Learning each such universal vector requires iterating through the dataset, resulting in an exponential time and space complexity. + +Existing universal perturbation methods focus on MCR setting [7, 8, 60]. Adapting these methods to MLL would require learning a separate universal perturbation for each possible label combination, incurring exponential complexity. Although one could learn label-wise independent perturbations, no effective method exists to combine them for attacking specific label combinations. We empirically show that naively combining them does not lead to successful attacks. Similarly, generative attack approaches [27, 69, 76] would be ineffective for unseen label combinations. + +Contributions. In this paper, we develop an efficient method for crafting targeted universal adversarial attacks on multi-label learning (MLL) models, addressing the aforementioned combinatorial complexity problem. Our key assumption is to generate a universal attack on a target set as a composition of universal label-wise attacks of the target labels (see Fig. 1). Based on this assumption, we derive a formulation which requires learning affine convex cones spanned by label-wise universal vectors. Our formulation requires linear training time complexity and after trained, the universal perturbation to attack any combination of labels can be composed in constant time using the label-wise attack vectors. Our key contributions are: + +- Existing MCR-based approaches are not inherently designed to attack multiple targeted labels for MLL models. Our work is the first to address this gap, offering an efficient and principled solution for MLL setting; +- Our proposed framework to learn targeted universal perturbations for MLL has linear time and space complexity and constructs attacks on any target set in constant time, while achieving performance comparable to the existing oracle methods with exponential complexity; +- Our results visualize and provide deeper understanding of the universal adversarial regions learnt by our method; +- We evaluate our method on NUS-WIDE, MS-COCO, and OpenImages benchmark datasets using state-of-the-art multi-label learning models against various baselines, showing the effectiveness of our method. + +# 2. Related Works + +# 2.1. Multi-Label Recognition + +There has been significant recent interest in multi-label learning due to its applications in various tasks. In MLL, the goal is to identify all concepts/labels in an image. This is particularly a harder task than multi-class recognition, since there can be exponentially many possible combinations of classes present in an image and can be related to each other. To address this problem, early works [90, 97] considered independent binary classifiers to classify the presence or absence of each class. However, since such a method does not consider the correlations among labels, several other approaches have been proposed that exploit label correlations or external knowledge [13, 18, 34, 45, 49, 51, 70, 94-96, 98, 101, 105-107, 113, 118], or use graphical models [50, 52]. More recent works use attention mechanism or novel loss functions to consider the positive and negative sample imbalance [6, 17, 53, 54, 58, 93, 112]. + +# 2.2. Adversarial Attacks + +Adversarial attacks [3, 12, 81, 89, 99] have gained a lot of attention as they identify the key vulnerability of deep neural networks (DNNs) to small imperceptible changes in the input. Researchers have explored various ways that attackers can manipulate data to fool models and developed several defenses to counter these attacks [4, 5, 10, 11, 19, 21, 22, 28, 29, 35, 46, 62, 71-73, 78, 82-84, 91, 92, 102, 114]. Based on the attacker's objective and the target setting, the adversarial attacks can be categorized in different ways, such as targeted and untargeted attacks, white-box (which assume full access to the target model parameters) and black-box attacks (no access to the model parameters), or instance-specific and instance-agnostic attacks. + +# 2.3. Universal Adversarial Attacks + +In instance-specific attacks, the perturbations are generated and targeted for each input image, whereas, instance-agnostic attacks [38, 65, 69] generate a single universal perturbation that can fool all input instances. These universal adversarial perturbations(UAPs) can be untargeted or targeted. The goal of the untargeted universal perturbation is to perturb any image such that it is misclassified by the target model [67, 76]. In the targeted case, the goal is to learn the universal perturbation such that it misclassifies any given image to a specific class [7, 110]. Universal perturbations were first proposed by Moosavi-Dezfooli et al. [67] and were generated by iteratively computing and aggregating perturbations using DeepFool [66] across multiple inputs. Since then, several methods have been proposed to learn better universal perturbations [60, 109–111]. To learn UAPs with better generalization, Liu et al. [59] proposed a gradient aggregation method to address gradient vanish + +![](images/26d5e55b3778924cabe3c9374155441b5ea7a5a4a96a5c55e9b4f5499ef65abd.jpg) +Current Approach +Figure 2. In multi-label recognition, images can have any combination of labels present, leading to a combinatorial explosion in possible universal perturbations. Traditional methods (shown on the left) require exponential time and space complexity to learn perturbations for all class combinations. Our approach (shown on the right) reduces this complexity by learning $|\mathcal{C}|$ affine convex cones, each representing a region of universal perturbations for a specific class. This allows us to efficiently generate perturbations for any combination of classes. + +![](images/9793affce6d740d16f153eabf1d656ee39b822d0ed8998a83d9f7682c8617ad9.jpg) +Overview of the Proposed Method + +![](images/1eea43c9d7a6611c4cbdcbdc9def97ea6207d7a227127c44538e2282a68805be.jpg) + +ing problem and Li et al. [48] proposed to utilize instance-specific attack method to generate dominant UAPs. + +# 2.4. Multi-Label Adversarial Attacks + +Many recent works have studied adversarial attacks for multi-label recognition models [41, 61, 103, 104, 116, 117]. Song et al. [87] proposed white-box multi-label attack framework and Hu et al. [32] proposed to exploit ranking relations to design attacks for top- $k$ multi-label models. Aich et al. [1] proposed a CLIP-based generative model to generate multi-object attacks in the black-box setting. Jia et al. [37] proposed theoretical robustness guarantees to defend against multi-label adversarial attacks and Melacci et al. [64] exploited domain knowledge context to detect multi-label adversarial attacks. Mahmood and Elhamifar [63] proposed an optimization to address the problem of negative gradient correlation in generating multi-label attacks. These works mainly address instance-specific perturbations for multi-label setting. To the best of our knowledge, only Hu et al. [32] proposed universal untargeted attack where a single perturbation was learnt to attack and replace top- $k$ predicted labels across images. In comparison, we address the problem of learning targeted universal perturbations to attack specific combinations of labels. + +# 3. Multi-Label Universal Perturbations + +# 3.1. Problem Formulation + +Consider a multi-label classifier $\mathcal{F}:\mathbb{R}^d\to \mathbb{R}^{|\mathcal{C}|}$ where $\mathcal{C}$ is the set of all labels. For an image $\pmb {x}\in \mathbb{R}^d$ , let $\pmb {y}\in \{0,1\}^{|\mathcal{C}|}$ denote the set of its labels, where each entry indicates the absence (0) or presence (1) of the corresponding label in $\mathcal{C}$ in the image. The classifier $\mathcal{F} = \{\mathcal{F}_1,\mathcal{F}_2,\dots ,\mathcal{F}_{|\mathcal{C}|}\}$ consists of $|\mathcal{C}|$ binary classifiers (one for each label), where $\mathcal{F}_c(\pmb {x})\in (-\infty , + \infty)$ represents the logit of label $c$ .The + +probability of label $c$ being present in the image is given by $\hat{o}_c = \sigma (\mathcal{F}_c(\pmb {x}))$ , where $\sigma (.)$ is the sigmoid function, hence, it is considered present (1), if $\hat{o}_c\geq 0.5$ , and absent (0) otherwise. We assume that $\mathcal{F}$ has already been trained on a set of images. Let $\Omega \subseteq \mathcal{C}$ denote the set of target labels that we want to attack. Our goal is to generate a targeted universal perturbation for $\Omega$ . + +# 3.2. Multi-Label Universal Perturbations + +We want to find a single (universal) perturbation vector, $\pmb{u}_{\Omega}$ , that can attack a set of target labels in $\Omega$ in any image $\pmb{x}$ containing them. One can generate such a vector by maximizing the expected multi-label learning loss for the labels in $\Omega$ across the training dataset, while restricting the norm of $\pmb{u}_{\Omega}$ (we consider $\ell_{\infty}$ -norm). However, due to correlations among labels in MLL, attacking the target labels may inadvertently affect non-target labels $(\bar{\Omega} = \mathcal{C} \setminus \Omega)$ . To mitigate this, we seek to minimize the influence of the universal perturbation vector on non-target labels. This requires solving the following oracle optimization, + +$$ +P _ {\Omega} ^ {*}: \max _ {\boldsymbol {u} _ {\Omega}} \sum_ {k \in \Omega} \mathcal {H} _ {k} (\boldsymbol {u} _ {\Omega}) - \gamma \sum_ {q \in \Omega} \mathcal {R} _ {q} (\boldsymbol {u} _ {\Omega}), \tag {1} +$$ + +where $\mathcal{H}_k$ ensures attacking the label $k$ in the target set and $\mathcal{R}_q$ tries to prevent modifying the prediction of non-target labels using the universal perturbation $\pmb{u}_{\Omega}$ . Here, $\gamma$ is a scalar hyperparameter setting a trade-off between the two terms. More specifically, we define $\mathcal{H}_k$ and $\mathcal{R}_q$ as + +$$ +\mathcal {H} _ {k} \left(\boldsymbol {u} _ {\Omega}\right) \triangleq \mathbb {E} _ {\left(\boldsymbol {x}, \boldsymbol {y}\right)} \mathcal {L} _ {\mathrm {b c e}} \left(\mathcal {F} _ {k} \left(\boldsymbol {x} + \Pi_ {\epsilon} \left(\boldsymbol {u} _ {\Omega}\right)\right), y _ {k}\right), \tag {2} +$$ + +$$ +\mathcal {R} _ {q} (\boldsymbol {u} _ {\Omega}) \triangleq \mathbb {E} _ {(\boldsymbol {x}, \boldsymbol {y})} \max \left\{0, - \tanh \left(\mathcal {F} _ {q} (\boldsymbol {x}) \mathcal {F} _ {q} (\boldsymbol {x} + \Pi_ {\epsilon} (\boldsymbol {u} _ {\Omega}))\right) \right\}, +$$ + +where $\Pi_{\epsilon}$ projects the input on p-norm ball of radius $\epsilon$ , $\mathcal{L}_{\mathrm{bce}}$ is the binary cross-entropy loss and the expectation is with respect to samples from the dataset. Therefore, maximizing $\mathcal{H}_k(\pmb{u}_{\Omega})$ ensures to find a perturbation that modifies the + +prediction of each target label $k$ . On the other hand, $\mathcal{R}_k$ aims to minimize the influence of the perturbation $(\pmb{u}_{\Omega})$ on each non-target label $q$ . This is done by incurring a positive penalty when the signs of the logits for the unperturbed and perturbed inputs differ from each other. Thus, minimizing this term leads to the same signs for the logits of each non-target label before and after the attack. + +# 3.3. Compositional Multi-Label Universal Attacks (CMLU) + +There are $2^{|C|}$ possible subsets of labels $\mathcal{C}$ . Therefore, finding all universal perturbations $\{\pmb{u}_{\Omega}\}$ by independently solving (1) for each subset $\Omega$ leads to the exponential complexity $O(2^{|C|})$ . We develop a compositional multi-label universal attack (CMLU) that allows us to find any universal perturbation in linear complexity $O(|C|)$ . + +Let $\mathcal{P}(\mathcal{C})$ denote the power set of $\mathcal{C}$ , i.e., the set of all subsets of $\mathcal{C}$ . Our goal is to find all $\pmb{u}_{\Omega}$ 's by solving (1) for every $\Omega \in \mathcal{P}(\mathcal{C})$ , + +$$ +\sum_ {\Omega \in \mathcal {P} (\mathcal {C})} \max _ {\boldsymbol {u} _ {\Omega}} \left[ \sum_ {k \in \Omega} \mathcal {H} _ {k} \left(\boldsymbol {u} _ {\Omega}\right) - \gamma \sum_ {q \in \bar {\Omega}} \mathcal {R} _ {q} \left(\boldsymbol {u} _ {\Omega}\right) \right]. \tag {3} +$$ + +As mentioned above, solving (3) over the power set of $\mathcal{C}$ has exponential complexity. To tackle this problem, we make the following assumption: each universal perturbation $\pmb{u}_{\Omega}$ can be written as the sum of atomic universal perturbations + +$$ +\boldsymbol {u} _ {\Omega} \triangleq \sum_ {i \in \Omega} \boldsymbol {u} _ {i}, \tag {4} +$$ + +where $\pmb{u}_i$ is the universal perturbation on the label $i$ . We also assume that the label-wise universal perturbations are norm-constrained and mutually orthogonal, i.e., $\| \pmb{u}_i\|_p \leq \epsilon$ and $\hat{\pmb{u}}_i^\top \hat{\pmb{u}}_j = 0$ for each $i \neq j$ , where $\hat{\pmb{u}} = \pmb{u} / \| \pmb{u}\|_2$ . + +Let $\mathbf{U} = \left[ \begin{array}{lll}\pmb {u}_1 & \pmb {u}_2 & \dots & \pmb{u}_{|\mathcal{C}|} \end{array} \right]$ represent a matrix whose columns are label-wise universal perturbations and $\lambda_{\Omega}$ denote a vector whose elements indexed by $\Omega$ are one and other elements are zero. Based on our assumption, we optimize $\mathbf{U}$ so that we can construct any $\pmb{u}_{\Omega}$ using the columns of $\mathbf{U}$ . Given our assumptions, we can rewrite (3) as + +$$ +\begin{array}{l} \max _ {\boldsymbol {U}} \sum_ {\Omega \in \mathcal {P} (\mathcal {C})} \sum_ {k \in \Omega} \left(\mathcal {H} _ {k} \left(\boldsymbol {u} _ {k} + \boldsymbol {U} \boldsymbol {\lambda} _ {\Omega \backslash k}\right) - \frac {\gamma}{| \Omega |} \sum_ {q \in \Omega} \mathcal {R} _ {q} \left(\boldsymbol {u} _ {k} + \boldsymbol {U} \boldsymbol {\lambda} _ {\Omega \backslash k}\right)\right), \\ s. t. \hat {\boldsymbol {U}} ^ {\top} \hat {\boldsymbol {U}} = \boldsymbol {I}, \| \boldsymbol {U} \| _ {1, p} \leq \epsilon \tag {5} \\ \end{array} +$$ + +where $\hat{U}$ is column-normalized $U$ and $\| U\|_{1,p} = \max_{1\leq i\leq |\mathcal{C}|}\| u_i\| _p$ , is the maximum norm of the columns of $U$ . Notice that we wrote $u_{\Omega} = u_k + U\lambda_{\Omega \backslash k}$ for a $k$ that belongs to $\Omega$ , where $\Omega \backslash k$ denotes removing $k$ from $\Omega$ . + +One can see that, for any $k \in \mathcal{C}$ , there are $2^{|C| - 1}$ subsets that contain $k$ . Each of these subsets would contain $k$ and classes from the powerset $\mathcal{P}(\mathcal{C} \backslash k)$ . We can rewrite (5) as (see the supplementary materials for derivation details) + +$$ +\max _ {\boldsymbol {U}} \sum_ {k \in \mathcal {C}} \sum_ {\Omega \in \mathcal {P} (\mathcal {C} \backslash k)} \left(\mathcal {H} _ {k} \left(\boldsymbol {u} _ {k} + \boldsymbol {U} \boldsymbol {\lambda} _ {\Omega \backslash k}\right) - \eta \sum_ {q \in \{\bar {\Omega} \backslash k \}} \mathcal {R} _ {q} \left(\boldsymbol {u} _ {k} + \boldsymbol {U} \boldsymbol {\lambda} _ {\Omega \backslash k}\right)\right) +$$ + +$$ +\mathrm {s . t .} \hat {\boldsymbol {U}} ^ {\top} \hat {\boldsymbol {U}} = \boldsymbol {I}, \| \boldsymbol {U} \| _ {1, p} \leq \epsilon , \tag {6} +$$ + +where $\eta = \gamma /|\Omega |$ . The power set of $\mathcal{C}$ can be written as the union of all subsets of $\mathcal{C}$ .. $\mathcal{P}(\mathcal{C})\triangleq \bigcup_{i = 0}^{|C|}\binom {\mathcal{C}}{i}$ where, each $\binom{\mathcal{C}}{i}$ has all the subsets of $\mathcal{C}$ of size $i$ . In our case, $S\in \binom{\mathcal{C}}{i}$ would represent the indices of the columns of $\pmb{U}$ . So, we can write the summation across the power set as: + +$$ +\max_{\boldsymbol{U}}\sum_{k\in \mathcal{C}}\sum_{t = 0}^{\left|\mathcal{C}\right| - 1}\sum_{S\in \binom {\mathcal{C}\setminus k}{t}}\Bigl(\mathcal{H}_{k}(\boldsymbol{u}_{k} + \boldsymbol {U}\boldsymbol{\lambda}_{S}) - \eta \sum_{q\in \{\bar{S}\setminus k\}}\mathcal{R}_{q}(\boldsymbol{u}_{k} + \boldsymbol {U}\boldsymbol{\lambda}_{S})\Bigr), +$$ + +$$ +s. t. \hat {\boldsymbol {U}} ^ {\top} \hat {\boldsymbol {U}} = \boldsymbol {I}, \| \boldsymbol {U} \| _ {1, p} \leq \epsilon , \tag {7} +$$ + +In (7), for a fixed $k$ and $t$ , $\lambda_{S}$ has exactly $t$ non-zero elements and $S$ iterates through all possible combinations of $\{C \backslash k\}$ of size $t$ . Therefore, the sum across all $\lambda_{S}$ (for $S \in \binom{\mathcal{C} \backslash k}{t}$ ) would be equal to the sum across all binary vectors $\lambda \in \{0,1\}^{|\mathcal{C}|}$ such that exactly $t$ elements are nonzero (i.e., $\| \lambda \|_0 = t$ ) and $\lambda_{k} = 0$ . Where, $\| \cdot \|_0$ denotes the $\ell_0$ -norm, which counts the number of nonzero elements of the given input. So we can write (7) as + +$$ +\max _ {\boldsymbol {U}} \sum_ {k \in \mathcal {C}} \sum_ {t = 0} ^ {| \mathcal {C} | - 1} \sum_ {\substack {\boldsymbol {\lambda} \in \{0, 1 \} ^ {| \mathcal {C} |} \\ \| \boldsymbol {\lambda} \| _ {0} = t, \boldsymbol {\lambda} _ {k} = 0}} \left[ \mathcal {H} _ {k} (\boldsymbol {u} _ {k} + \boldsymbol {U} \boldsymbol {\lambda}) \right. \tag{8} +$$ + +$$ +\left. - \eta \sum_ {q \in \{\mathcal {C} \backslash k \}} \delta (\boldsymbol {\lambda} _ {q} = 0) \mathcal {R} _ {q} (\boldsymbol {u} _ {k} + \boldsymbol {U} \boldsymbol {\lambda}) \right] \text {s . t .} \hat {\boldsymbol {U}} ^ {\top} \hat {\boldsymbol {U}} = \boldsymbol {I}, \| \boldsymbol {U} \| _ {1, p} \leq \epsilon , +$$ + +where $\delta$ is the indicator function which is 1 if the input condition is True and 0 otherwise. As $\lambda$ is $|\mathcal{C}|$ dimensional, the second loss term in (8) iterates through each index of $\lambda$ i.e., $(q\in \{\mathcal{C}\backslash k\})$ and computes the loss of label $q$ if $\lambda_q = 0$ . + +We investigate two ways to approximately solve (8). In the first approach, we solve the lower bound of (8) as + +$$ +\operatorname {CMLU} _ {\alpha}: \max _ {\boldsymbol {U}} \sum_ {k \in \mathcal {C}} \sum_ {t \in \mathcal {T}} 2 ^ {| \mathcal {C} |} \min _ {\substack {\boldsymbol {\lambda} \in \{0,1 \} ^ {| C |} \\ \| \boldsymbol {\lambda} \| _ {0} = t, \boldsymbol {\lambda} _ {k} = 0}} \left[ \mathcal {H} _ {k} (\boldsymbol {u} _ {k} + \boldsymbol {U} \boldsymbol {\lambda}) \right. \tag{9} +$$ + +$$ +\left. - \eta \sum_ {q \in \{\mathcal {C} \backslash k \}} \delta (\boldsymbol {\lambda} _ {q} = 0) \mathcal {R} _ {q} (\boldsymbol {u} _ {k} + \boldsymbol {U} \boldsymbol {\lambda}) \right] \text {s . t .} \hat {\boldsymbol {U}} ^ {\top} \hat {\boldsymbol {U}} = \boldsymbol {I}, \| \boldsymbol {U} \| _ {1, p} \leq \epsilon +$$ + +Where, $\mathcal{T} = \{0,1,\dots ,|\mathcal{C}| - 1\}$ . (9) requires maximizing the minimal loss. To solve the inner minimization problem, for a fixed $k$ and $t$ , we randomly initialize $\lambda$ and iteratively update it for a fixed number of iterations. We then project $\lambda$ on the feasible set by keeping the largest $t$ values as 1 and setting the rest to 0. We then keep the $\lambda$ fixed and solve the outer maximization in (9) w.r.t. $\pmb{U}$ . In practice, instead of iterating through all $t\in \mathcal{T}$ , for each $k$ , we randomly choose $t$ . The steps are shown in Algorithm 1. + +In the second approach, we consider the continuous relaxation of $\lambda$ and approximately solve it by uniformly randomly sampling $\lambda \in [0,1]^{|\mathcal{C}|}$ , + +$$ +\operatorname {C M L U} _ {\beta}: \max _ {\boldsymbol {U}} \sum_ {k \in \mathcal {C}} \sum_ {t \in \mathcal {T}} \sum_ {\substack {\boldsymbol {\lambda} \in [ 0, 1 ] ^ {| \mathcal {C} |} \\ \| \boldsymbol {\lambda} \| _ {0} = t, \boldsymbol {\lambda} _ {k} = 0}} \left[ \mathcal {H} _ {k} (\boldsymbol {u} _ {k} + \boldsymbol {U} \boldsymbol {\lambda}) \right. \tag{10} +$$ + +$$ +\left. - \eta \sum_ {q \in \{\mathcal {C} \backslash k \}} \delta (\boldsymbol {\lambda} _ {q} = 0) \mathcal {R} _ {q} (\boldsymbol {u} _ {k} + \boldsymbol {U} \boldsymbol {\lambda}) \right] \text {s . t .} \hat {\boldsymbol {U}} ^ {\top} \hat {\boldsymbol {U}} = \boldsymbol {I}, \| \boldsymbol {U} \| _ {1, p} \leq \epsilon . +$$ + +Eq. (9) and (10) can be explained intuitively. The first loss term $\mathcal{H}_k$ optimizes each label-wise universal perturbation $\pmb{u}_k$ such that it is robust to the noise generated by composing other universal perturbations. Specifically, it aims to maximize the loss of class $k$ , regardless of which other class universals have been combined. The second term ensures that $\pmb{u}_k$ does not change the prediction of classes whose universal vectors have not been used in the composition. + +Remark: A key aspect of Eq. (9) and (10) is that $\pmb{u}_k + \pmb{U}\lambda$ represents points sampled from an affine convex cone. Specifically, these formulations aim to maximize a function over points sampled from $|C|$ affine convex cones (generated by the columns of $\pmb{U}$ ). Consequently, we can interpret our assumption in (4) as aiming to learn the universal adversarial regions of each class as affine cones, where the vector $\pmb{u}_{\Omega}$ , as outlined in (4), lies at the intersection of the affine cones of the classes in $\Omega$ . This is illustrated in Fig. 2 (right), where $\pmb{C}_k$ and $\pmb{C}_i$ represent convex cones for classes $k$ and $i$ respectively, and $\pmb{C}_{ik}$ denotes their intersection. + +For a comparison among Oracle, $\mathrm{CMLU}_{\alpha}$ , and $\mathrm{CMLU}_{\beta}$ , Algorithm 1 shows the steps of a single training epoch of Oracle of target size $|\Omega| = 1$ and the corresponding steps for our methods. Note that in general, Oracle has $\mathcal{O}(\binom{|C|}{|\Omega|})$ training time complexity, which increases with the size of $\Omega$ . For instance, for $|\Omega| = 2$ , the complexity of the Oracle's inner loop becomes $\mathcal{O}(\binom{|C|}{2})$ , as it needs to handle all possible combinations of classes of a fixed size $|\Omega|$ . However, $\mathrm{CMLU}_{\alpha}$ has $\mathcal{O}(N_{\alpha}|\mathcal{C}|)$ and $\mathrm{CMLU}_{\beta}$ has $\mathcal{O}(|C|)$ complexity and is independent of $|\Omega|$ . Note that the complexity of our method for learning composable universal perturbations for all possible target set sizes is equivalent to the complexity of Oracle trained for $|\Omega| = 1$ (linear in the number of classes). Oracle requires iterating over the dataset for each combination of target classes $\Omega$ , leading to exponential complexity. + +# 4. Evaluation + +# 4.1. Experimental Setup + +Datasets We perform experiments using NUS-WIDE [16], MS-COCO [56], and OpenImages [44] datasets to assess the effectiveness of our proposed method. We compare it with oracle and baseline methods. For NUS-WIDE and MS-COCO, we test target set sizes ranging from 1 to 5 and for OpenImages, we experiment with larger target set sizes, going up to 20. Due to the computational expense of learning oracle model, we choose a subset of labels from each dataset (25 classes from NUS-WIDE, 40 classes from MS-COCO, and 100 classes from OpenImages). Even with the small subset of classes, there can be a large number of possible combinations e.g., there can be more than 53k combinations of $|\Omega| = 5$ using 25 classes from NUS-WIDE. Therefore, we perform experiments on a subset of those possible combinations for each target set size $|\Omega|$ . For all of our + +Algorithm 1: Algorithms of Oracle, $\mathrm{CMLU}_{\alpha}$ , $\mathrm{CMLU}_{\beta}$ +Input: $\mathcal{D}$ : Training Data, $\mathcal{F}$ : Classifier, $\epsilon$ : Norm budget, $(\xi ,\alpha)$ : Step sizes Output: $U$ : Learnt Universal Perturbations 1 Randomly initialize $U$ /\* $\mathcal{B} = \{\mathcal{B}^{(k)}\}_{k = 1}^{|C|},\mathcal{B}^{(k)}$ images of class $k\ast /$ +2 Oracle: +3 for Batch $\mathcal{B}$ in $\mathcal{D}$ do /* One iteration through the dataset \*/ +4 for $k$ in $\mathcal{C}$ do /\* O(|C|) \*/ +5 $\mathcal{L}_k$ : Compute loss in (1) using $\pmb {u}_k$ +6 Compute the gradient: $\pmb {g}\gets \nabla_{\pmb{U}_k}\pmb {l}_k$ +7 Update $U_{k}\leftarrow \Pi_{\epsilon}\left(U_{k} + \xi *Sgn(\pmb {g})\right)$ +8 for Batch $\mathcal{B}$ in $\mathcal{D}$ do +9 Set $\pmb{g}$ to be all zeros of the same size as $U$ +10 for $k$ in $\mathcal{C}$ do /\* O(|C|) \*/ +11 Randomly sample $t\sim T$ and $\lambda \in [0,1]^{\lfloor C\rfloor}$ . +12 $\mathrm{CMLU}_{\alpha}$ ： /\*O(Nα）\*/ +13 for $N_{\alpha}$ num of iterations do +14 Set $\pmb {v}_k = \pmb {u}_k + U\pmb{\lambda}$ +15 $\mathcal{L}_k$ : Compute loss in (9) using $\pmb {v}_k$ +16 Update: $\lambda = \lambda -\alpha \nabla_{\lambda}\mathcal{L}_k$ +17 Project $\lambda$ on the feasible set $\| \lambda \| _0 = t$ +18 Set $\lambda_k = 0$ and $\pmb {v}_k = \pmb {u}_k + \pmb {U}\pmb{\lambda}$ +19 $\mathcal{L}_k$ : Compute loss in (9) using $\pmb {v}_k$ +20 CMLUβ:/*O(1)\*/ +21 Randomly set $|\mathcal{C}| - t$ values in $\lambda$ to 0. +22 Set $\lambda_k = 0$ and $\pmb {v}_k = \pmb {u}_k + U\pmb{\lambda}$ +23 $\mathcal{L}_k$ :Compute loss in (10) using $\pmb {v}_k$ +24 Accumulate: $\pmb {g}\gets \pmb {g} + \nabla_{\pmb{U}_k}\mathcal{L}_k$ +25 Update $U\gets \Pi_{\epsilon}(U + \xi *Sgn(g))$ + +experiments, we learn universal vectors with maximum infinity norm $\epsilon = 0.05$ (for image values range of [0,1]) and using SGD with fix step size $\xi = \epsilon /25$ . We set $\gamma = \kappa /|\bar{\Omega}|$ for training oracle and baselines, and $\eta = \kappa /\| \pmb {\lambda}\| _0$ in each iteration of our proposed method. We show ablation experiments for different values of $\kappa$ in section 4.2.1. + +Multi-Label Recognition Models We evaluate the baselines and our method on state-of-the-art multi-label recognition models (using the officially provided weights). + +- Asymmetric Loss (ASL) [79] is a novel loss function to address the imbalance between numerous negative labels and only a few positive labels associated with each sample. This method balances the probabilities of different positive and negative samples. In particular, the loss performs hard thresholding by dynamically down-weighting easy negative samples and can discard unlabeled samples. + +- ML-Decoder [80] is an efficient, attention-based multi-label classifier that has a novel decoder architecture and + +![](images/825c2522eda98ed053528cb905dad9b4ca06c03d824e395fac4cf19589e2c35e.jpg) +Figure 3. Fooling success rates $(FR)$ on various datasets using ASL and ML-Decoder as we increase the target set size $|\Omega|$ . The left two figures show results on NUS-WIDE and the right two figures show results on MS-COCO. + +![](images/e27be3d6f28315f9f360620504b2000c774600c9e4e3c28f19cfe925e72c162a.jpg) + +![](images/a185711a2e7d0c80029ff3b27b8acc95486b332fdc030162253977d825a82952.jpg) + +![](images/be9818a4d3dc4763d1fe2360fb96d9710839b6dda581b11a8ccce9c6ddd12918.jpg) + +![](images/478457689f7413ff080020b02544cf847dc88442da297593ec3dc3a7810d75e1.jpg) + +group-decoding scheme, which enables it to efficiently utilize the spatial information in images. The model is highly efficient as it can scale to large number of classes. Moreover, using word queries, it can generalize to unseen classes. + +Oracle and Baselines We evaluate our method against three baselines and an oracle. The computational complexity of each of these methods is reported in Table 1 and their wall clock time in supplementary. + +- Oracle directly optimizes (1) to learn a universal perturbation $\pmb{u}_{\Omega}$ to attack a specific target set $\Omega$ . This provides the upper bound on the attack performance for our method. It requires $2^{|C|}d$ -dimensional universal vectors and for each vector, iterate through the training set for $E$ epochs. + +- Oracle Sum (Or-S) directly learns single-class universal perturbation vectors $\{\pmb{u}_1, \pmb{u}_2, \dots, \pmb{u}_{|\mathcal{C}|}\}$ using (1) and generates a universal perturbation to attack set $\Omega$ as $\pmb{u}_{\Omega} = \sum_{i \in \Omega} \pmb{u}_i$ . It only needs to learn $|C|$ $d$ -dimensional vectors. +- Oracle Combination (Or-C) first learns single-class universal perturbation vectors and learns to combine them to attack a specific target set $\Omega$ i.e., $\pmb{u}_{\Omega} = \sum_{i\in \Omega}a_i^\Omega \pmb {u}_i$ . Where, $a_i^\Omega \in \mathbb{R}$ is associated with $i^{th}$ class in specific set $\Omega$ , and is learnt by iterating through the training set. It needs to learn $|C|$ $d$ -dim vectors and $2^{|C|}$ scalars to combine them. +- SGA [59] addresses the gradient vanishing and poor local optima problem to improve the generalization of UAPs. It enhances the gradient stability by aggregating gradients from inner pre-search. We use SGA to learn single-class UAPs and use (4) to generate UAPs for target labels. +- NAG [68] models the distribution of universal adversarial perturbations using GANs, originally designed for untargeted perturbations in multi-class setting. To adapt this for multi-label universal perturbations, we condition the GAN on class vectors. During training, we generate label-wise universal perturbations. At test time, we generate perturbations for multiple classes by conditioning the GAN on a combined representation vector of the target classes. For this method, we have linear time complexity as we train to generate label-wise perturbations and the space requirement is the size of the generator. + +Evaluation Metrics Given a set of classes $\Omega$ to attack, let $\mathcal{I}_{\Omega}$ represent all the images that contain the classes $\Omega$ and were attacked. Let $\mathcal{A}_{\Omega} \subseteq \mathcal{I}_{\Omega}$ represent the set of images that were successfully attacked. To evaluate the universal attack performance, we define the attack fooling success rate, $FR$ , to measure the percentage of successfully attacked images. Since we want to affect the least number of non-targeted labels while achieving high fooling rates, we also report non-target flip rate, $NT_{R}$ , to measure the percentage of non-targeted labels (labels in $\bar{\Omega}$ ) which were flipped by the attack. Our goal is to achieve high $FR$ while maintaining low $NT_{R}$ for attacking target set of any size. + +$$ +\mathrm {F R} = \frac {\left| \mathcal {A} _ {\Omega} \right|}{\left| \mathcal {I} _ {\Omega} \right|}, \mathrm {N T} _ {R} = \frac {1}{\left| \mathcal {A} _ {\Omega} \right|} \sum_ {k \in \mathcal {A} _ {\Omega}} \frac {\sum_ {i \in \bar {\Omega}} \left(1 - \delta \left(\chi \left(f _ {i} ^ {(k)}\right) , y _ {i} ^ {(k)}\right)\right)}{\left| \bar {\Omega} \right|} \tag {11} +$$ + +where, $\delta$ is Kronecker delta that equals 1 when the two inputs are equal and 0 otherwise. $\chi (v)$ is the step function which is 1 if $v > 0$ and 0 otherwise. $y_{i}^{(k)},f_{i}^{(k)}$ are the predictions on clean and perturbed images respectively of $i^{th}$ non-target class of $k^{th}$ successfully attacked image. Note that $FR$ is calculated using all attacked images while $NT_{R}$ is calculated using the set of successfully attacked images. + +# 4.2. Experimental Results + +In this section, we evaluate our compositional framework for universal perturbations. Fig. 3 shows the fooling success rates for different methods. The results demonstrate consistent performance trends across datasets and models. Particularly, the success rate(FR) of baselines decline as we increase the target size. On the other hand, the Oracle method achieves the best performance but suffers from exponential complexity, making it impractical for larger target sets. + +Among the baselines, Or-C, Or-S, and SGA perform the worst for large target set sizes. Since the performance of these methods depends on the label-wise universal perturbations, we can conclude that the low success rate of these methods implies that the subspace spanned by labelwise perturbations does not contain the intersection of their respective adversarial regions. This is further elaborated in Fig. 5. In all cases, Or-C is better than Or-S since it op- + +Table 1. Non-target flip rate $(NT_R)$ on NUS-WIDE and MS-COCO, as we increase the target set size from 1 to 5. We show (-) when $|\mathcal{A}_{\Omega}| = 0$ . We also show the space and time complexity of each method. $E$ represents the number of epochs for training, $d$ is the size of the perturbation (equal to the image size), $G_p$ is the size of GAN's parameters, and $|\mathcal{C}|$ is the number of classes. + +
Computational ComplexityASLML-Decoder
NUS-WIDEMS-COCONUS-WIDEMS-COCO
MethodSpaceTime135135135135
Oracle\( \mathcal{O}(d2^{|C|}) \)\( \mathcal{O}(E2^{|C|}) \)6.177.428.106.507.738.662.863.103.017.388.589.81
Or-S\( \mathcal{O}(d|C|) \)\( \mathcal{O}(E|C|) \)6.176.03-6.502.872.962.862.22-7.382.96-
Or-C\( \mathcal{O}(d|C|+2^{|C|}) \)\( \mathcal{O}(E2^{|C|}) \)6.176.862.746.505.233.202.862.450.607.386.40-
SGA\( \mathcal{O}(d|C|) \)\( \mathcal{O}(E|C|) \)7.697.31-7.774.30-2.892.471.278.628.949.37
NAG\( \mathcal{O}(G_p) \)\( \mathcal{O}(E|C|) \)8.325.861.3210.110.27.213.963.452.5511.812.113.7
\( \mathbf{CMLU}_{\alpha} \)\( \mathcal{O}(d|C|) \)\( \mathcal{O}(EN_{\alpha}|C|) \)5.637.653.747.335.963.133.63.082.518.968.558.02
\( \mathbf{CMLU}_{\beta} \)\( \mathcal{O}(d|C|) \)\( \mathcal{O}(E|C|) \)7.968.137.938.908.869.594.233.743.029.659.8310.5
+ +Table 2. Experiments on large-scale dataset. We report fooling success rate (FR) and non-target flip rate $(\mathrm{NT}_R)$ on OpenImages using ASL model. The experiments were performed on large target set sizes, up to $|\Omega| = 20$ . Higher FR ( $\uparrow$ ) and lower $\mathrm{NT}_R$ ( $\downarrow$ ) are better. + +
MethodFR (↑)NTR(↓)
1234510152012345101520
Oracle97.697.996.897.995.796.396.896.80.640.800.870.930.961.051.171.47
Or-S97.629.646.281.700.420.00.00.00.640.710.650.340.20---
Or-C97.649.422.823.313.40.00.00.00.640.710.670.750.48---
SGA98.941.49.646.110.920.150.00.00.770.810.730.740.320.12--
NAG95.992.188.888.489.779.980.984.20.810.890.931.011.071.421.531.84
CMLUα92.266.744.046.239.439.932.432.30.720.760.830.890.830.860.820.77
CMLUβ97.395.895.996.895.294.896.595.20.720.800.880.920.981.061.281.67
+ +![](images/9e0399c2f83fa5a5d025b5c5d75c33b82aca25c33341ec722e3b9fe14bf0a7ec.jpg) +Figure 4. Visualizing label-wise universal perturbations learnt by Oracle and $\mathrm{CMLU}_{\beta}$ on NUS-WIDE using ASL. + +timizes to find the best combination of the label-wise universal perturbations and hence, is better than the naive summation of individual perturbations. Across all experiments, $\mathrm{CMLU}_{\beta}$ is the closest to Oracle performance. This is despite the fact that $\mathrm{CMLU}_{\beta}$ has linear complexity whereas the Oracle has exponential complexity. Moreover, $\mathrm{CMLU}_{\alpha}$ achieves higher $FR$ and lower $NT_{R}$ than all baselines. + +Table 1 shows the percentage of non-targeted affected labels (whose prediction was changed after the attack). Although the baselines have low $\mathrm{NT}_R$ for target sizes greater than 1, they also have low fooling rates. Note that despite having linear complexity, the $\mathrm{NT}_R$ of $\mathrm{CMLU}_{\beta}$ is only + +slightly higher than Oracle. $\mathrm{CMLU}_{\alpha}$ achieves lower $NT_{R}$ as compared to $\mathrm{CMLU}_{\beta}$ but still outperforms the baselines. In all experiments, NAG performs well (close to $\mathrm{CMLU}_{\alpha}$ ) but declines in performance as we increase the target size. It also has significantly higher $\mathrm{NT}_{R}$ than the Oracle and $\mathrm{CMLU}_{\beta}$ . This implies that NAG learnt to attack multiple classes together rather than learning the distribution of label-wise universal perturbations. Therefore, even though it has high success rate, it also affects several other labels. + +We further perform experiments on OpenImages, a large-scale dataset, and we show results in Tab. 2. Performance trends remain consistent with those observed on smaller datasets. The success rate of Or-S and Or-C consistently drops as we target more classes, and gets to 0 by target size 10. Note that NAG performs better in OpenImages experiments. As shown in the table, NAG affects a large percentage of non-targeted labels, which signifies that the method does not capture the class-specific directions and perturbs a large group of labels together. $\mathrm{CMLU}_{\alpha}$ performs better than Or-C, Or-S, and SGA but worse than NAG. However, as shown in the table, $\mathrm{CMLU}_{\alpha}$ still achieves significantly lower $NT_{R}$ for large target set sizes. In all experiments, $\mathrm{CMLU}_{\beta}$ achieves consistently close performance to the Oracle while also maintaining low $\mathrm{NT}_{R}$ . + +![](images/bcc18fa01204e1d900ed1f8b6bcc99ca0ee45a595d6dc205c9f038e696f1a4b7.jpg) +Figure 5. Comparison of the regions (convex cones) spanned by universal perturbations learnt using Oracle and $\mathrm{CMLU}_{\beta}$ on NUS-WIDE (top) and MS-COCO (bottom). Each point in the heatmap is the average model confidence computed across images using the vector $(\epsilon_x * \boldsymbol{u}_x + \epsilon_y * \boldsymbol{u}_y)$ for all label combinations of size $|\Omega| = 2$ , where $\boldsymbol{u}_x, \boldsymbol{u}_y$ are universal vectors of classes in $\Omega$ . The asterisks(*) show the region where the $L_{\infty}$ norm is 0.05. + +# 4.2.1. Ablation Study + +We perform ablation experiments on Oracle and $\mathrm{CMLU}_{\beta}$ to investigate the effect of hyperparameters on their performance. Tab. 3 shows the results as we increase/decrease the values of $\epsilon$ and $\kappa$ . Increasing $\epsilon$ provides a larger perturbation budget for better performance; however, to keep the changes to the inputs imperceptible, we set $\epsilon = 0.05$ . Meanwhile, increasing $\kappa$ emphasizes maintaining non-targeted labels, which lowers $\mathrm{NT}_R$ but also reduces the success rate (FR). This is because many labels can be correlated in MLL and attacking a set of labels while fixing non-targeted but correlated labels lead to sub-optimal perturbations [63]. + +# 4.2.2. Analysis of Universal Adversarial Regions + +The label-wise universal perturbations learnt by Oracle and $\mathrm{CMLU}_{\beta}$ are visualized in Fig. 4. We quantitatively assess the visual impact of these perturbations using Structural Similarity Index (SSIM), computed between original and perturbed images. The average SSIM values across all images, as shown in Fig. 4, demonstrate comparable levels of distortion for Oracle and $\mathrm{CMLU}_{\beta}$ . Note that the pairwise dot product of the Oracle and CMLU vectors is $\approx 0$ , indicating that $\mathrm{CMLU}_{\beta}$ discover distinct attack directions, different from the Oracle. Additional analysis of pairwise dot products and visualizations are provided in supplementary. + +To further highlight and analyze the distinction between the perturbations learnt using different methods, we focus on Oracle and $\mathrm{CMLU}_{\beta}$ . Specifically, we examine the universal adversarial regions generated by those methods. Fig. 5 shows the average model confidence on the input space region spanned by the pairs of universal perturbation vectors. In the figure, perturbing inputs to blue and green regions would not change the model prediction (does not attack any class), yellow region would attack one class, and red region would attack both classes in $\Omega$ . Note that there is + +Table 3. Ablation experiments for hyperparameters on NUS-WIDE. The experiments are conducted for $|\Omega| = 1$ . + +
εOracleCMLUβκOracleCMLUβ
FRNTRFRNTRFRNTRFRNTR
0.0163.55.3038.55.041599.16.5098.58.32
0.0394.96.2984.27.092098.96.1797.67.96
0.0598.96.1797.67.962593.54.7089.36.97
0.0799.75.1099.37.543093.14.4085.56.58
+ +Table 4. Transferability Evaluation: $S \rightarrow T$ shows the performance of universal perturbations learnt on model $\mathbf{S}$ and evaluated on model $\mathbf{T}$ . (A: ASL, D: ML-Decoder) + +
Transferability of Universal Attacks
ModelsMethodNUS-WIDEMS-COCO
135135
A → DOracle90.183.777.334.610.65.40
CMLUβ89.978.475.442.917.66.4
D → AOracle24.61.330.064.428.319.6
CMLUβ33.50.160.064.834.521.2
+ +no prominent red region (to attack both classes) in the cross-section of Oracle vectors. This explains the low success rate of Or-S and Or-C which optimize to find red region in such a cross-section. More details on computing these cones is provided in supplementary. We also investigate how well those red regions computed using $\mathrm{CMLU}_{\beta}$ can transfer to other models. For that, we perform transferability attacks and the results are shown in Tab. 4. From the table, we observe that $\mathrm{CMLU}_{\beta}$ (linear complexity) achieves fooling success rates closer and often better than Oracle (exponential complexity). This shows that the convex cone regions corresponding to different labels are often common across models. This can be effectively used to generate black-box attacks. We hope that future works explore this further. + +# 5. Conclusion + +We addressed the problem of generating targeted multi-label universal perturbations. We developed a compositional framework to generate exponential number of multi-label universal perturbations in constant time. The independence assumption on label-wise universal perturbations naturally leads to a formulation that requires learning convex cones. Through extensive experiments on multiple datasets and models, we show that our method outperforms all baseline methods and achieves performance comparable to the Oracle (an exponential complexity method). + +# 6. Acknowledgements + +This work is sponsored by ARPA-H (1AY2AX000062), DARPA PTG (HR00112220001), NSF (IIS-2115110) and ARO (W911NF2110276). Content does not necessarily reflect the position/policy of the Government. + +# References + +[1] Abhishek Aich, Calvin-Khang Ta, Akash Gupta, Chengyu Song, Srikanth Krishnamurthy, Salman Asif, and Amit Roy-Chowdhury. Gama: Generative adversarial multi-object scene attacks. Advances in Neural Information Processing Systems, 35:36914-36930, 2022. 1, 3 +[2] Abhishek Aich, Shasha Li, Chengyu Song, M Salman Asif, Srikanth V Krishnamurthy, and Amit K Roy-Chowdhury. Leveraging local patch differences in multi-object scenes for generative adversarial attacks. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1308-1318, 2023. 1 +[3] N. Akhtar and A. Mian. Threat of adversarial attacks on deep learning in computer vision: A survey. arXiv, 2018. 2 +[4] A. Athalye, N. Carlini, and D. A. Wagner. Obfuscated gradients give a false sense of security: Circumventing defenses to adversarial examples. 2018. 2 +[5] Yuanhao Ban and Yinpeng Dong. Pre-trained adversarial perturbations. In Advances in Neural Information Processing Systems, pages 1196-1209, 2022. 2 +[6] Emanuel Ben-Baruch, Tal Ridnik, Itamar Friedman, Avi Ben-Cohen, Nadav Zamir, Asaf Noy, and Lihi Zelnik-Manor. Multi-label classification with partial annotations using class-aware selective loss. 2022. 2 +[7] P. Benz, C. Zhang, T. Imtiaz, and I.-S. Kweon. Double targeted universal adversarial perturbations. Asian Conference on Computer Vision, 2020. 1, 2 +[8] Philipp Benz, Chaoning Zhang, Adil Karjauv, and In So Kweon. Universal adversarial training with class-wise perturbations. In 2021 IEEE International Conference on Multimedia and Expo (ICME), pages 1-6. IEEE, 2021. 2 +[9] N. Carlini and D. Wagner. Adversarial examples are not easily detected: Bypassing ten detection methods. Workshop on Artificial Intelligence and Security, 2017. 1 +[10] N. Carlini and D. Wagner. Towards evaluating the robustness of neural networks. IEEE Symposium on Security and Privacy, 2017. 2 +[11] Y. Carmon, A. Raghunathan, L. Schmidt, P. Liang, and J. C. Duchi. Unlabeled data improves adversarial robustness. Neural Information Processing Systems, 2019. 2 +[12] P.-Y. Chen, Y. Sharma, H. Zhang, J. Yi, and C.-J. Hsieh. Ead: Elastic-net attacks to deep neural networks via adversarial examples. AAAI Conference on Artificial Intelligence, 2018. 2 +[13] T. Chen, M. Xu, X. Hui, H. Wu, and L. Lin. Learning semantic-specific graph representation for multi-label image recognition. IEEE International Conference on Computer Vision, 2019. 2 +[14] Zhao-Min Chen, Xiu-Shen Wei, Xin Jin, and Yanwen Guo. Multi-label image recognition with joint class-aware map disentangling and label correlation embedding. IEEE International Conference on Multimedia and Expo, 2019. 1 +[15] Z. M. Chen, X. S. Wei, P. Wang, and Y. Guo. Multi-label image recognition with graph convolutional networks. IEEE Conference on Computer Vision and Pattern Recognition, abs/1904.03582, 2019. 1 + +[16] T. S. Chua, J. Tang, R. Hong, H. Li, Z. Luo, and Y. T. Zheng. Nus-wide: A real-world web image database from national university ofsingapore.ACM International Conference on Image and Video Retrieval, 2009.5 +[17] S. D. Dao, E. Zhao, D. Phung, and J. Cai. Multi-label image classification with contrastive learning. arXiv preprint, arXiv:2107.11626, 2021. 2 +[18] J. Deng, N. Ding, Y. Jia, A. Frome, K. Murphy, S. Bengio, Y. Li, H. Neven, and H. Adam. Large-scale object classification using label relation graphs. European Conference on Computer Vision, 2014. 2 +[19] Gavin Weiguang Ding, Yash Sharma, Kry Yik Chau Lui, and Ruitong Huang. Mma training: Direct input space margin maximization through adversarial training. arXiv preprint arXiv:1812.02637, 2018. 2 +[20] Zixuan Ding, Ao Wang, Hui Chen, Qiang Zhang, Pengzhang Liu, Yongjun Bao, Weipeng Yan, and Jungong Han. Exploring structured semantic prior for multi label recognition with incomplete labels. 2023. 1 +[21] Junhao Dong, Seyed-Mohsen Moosavi-Dezfooli, Jianhuang Lai, and Xiaohua Xie. The enemy of my enemy is my friend: Exploring inverse adversaries for improving adversarial training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 24678–24687, 2023. 2 +[22] Yinpeng Dong, Zhijie Deng, Tianyu Pang, Jun Zhu, and Hang Su. Adversarial distributional training for robust deep learning. Advances in Neural Information Processing Systems, 33:8270-8283, 2020. 2 +[23] L. Feng, B. An, and S. He. Collaboration based multi-label learning. AAAI Conference on Artificial Intelligence, 2019. 1 +[24] Zongyuan Ge, Dwarikanath Mahapatra, Suman Sedai, Rahul Garnavi, and Rajib Chakravorty. Chest x-rays classification: A multi-label and fine-grained problem. arXiv, 2018. 1 +[25] I. J. Goodfellow, J. Shlens, and C. Szegedy. Explaining and harnessing adversarial examples. International Conference on Learning Representations, 2015. 1 +[26] Emily Hand, Carlos Castillo, and Rama Chellappa. Doing the best we can with what we have: Multi-label balancing with selective learning for attribute prediction. In Proceedings of the AAAI Conference on Artificial Intelligence, 2018. 1 +[27] J. Hayes and G. Danezis. Learning universal adversarial perturbations with generative models. IEEE Security and Privacy Workshops, 2018. 2 +[28] W. He, J. Wei, X. Chen, N. Carlini, and D. Song. Adversarial example defense: Ensembles of weak defenses are not strong. USENIX Workshop on Offensive Technologies, 2017. 2 +[29] D. Hendrycks, K. Lee, and M. Mazeika. Using pre-training can improve model robustness and uncertainty. 2019. 2 +[30] Dan Hendrycks, Kevin Zhao, Steven Basart, Jacob Steinhardt, and Dawn Song. Natural adversarial examples. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15262-15271, 2021. 1 + +[31] Lei Hsiung, Yun-Yun Tsai, Pin-Yu Chen, and Tsung-Yi Ho. Towards compositional adversarial robustness: Generalizing adversarial training to composite semantic perturbations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24658-24667, 2023. 1 +[32] Shu Hu, Lipeng Ke, Xin Wang, and Siwei Lyu. Tkmlap: Adversarial attacks to top-k multi-label learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7649-7657, 2021. 3 +[33] D. Huynh and E. Elhamifar. Interaction compass: Multi-label zero-shot learning of human-object interactions via spatial relations. International Conference on Computer Vision, 2021. 1 +[34] D. T. Huynh and E. Elhamifar. Interactive multi-label cnn learning with partial labels. IEEE Conference on Computer Vision and Pattern Recognition, 2020. 2 +[35] Tooba Imtiaz, Morgan Kohler, Jared Miller, Zifeng Wang, Mario Sznaier, Octavia I Camps, and Jennifer G Dy. Saif: Sparse adversarial and interpretable attack framework. arXiv preprint arXiv:2212.07495, 2022. 2 +[36] J. Li R. Ji, H. Liu, X. Hong, Y. Gao, and Q. Tian. Universal perturbation attack against image retrieval. International Conference on Computer Vision, 2019. 1 +[37] Jinyuan Jia, Wenjie Qu, and Neil Zhenqiang Gong. Multiguard: Provably robust multi-label classification against adversarial examples. Advances in Neural Information Processing Systems, 2022. 3 +[38] V. Khrulkov and I. Oseledets. Art of singular vectors and universal adversarial perturbations. IEEE Conference on Computer Vision and Pattern Recognition, 2018. 1, 2 +[39] Youngwook Kim, Jae Myung Kim, Zeynep Akata, and Jungwoo Lee. Large loss matters in weakly supervised multi-label classification. 2022. 1 +[40] Takumi Kobayashi. Two-way multi-label loss. 2023. 1 +[41] Linghao Kong, Wenjian Luo, Hongwei Zhang, Yang Liu, and Yuhui Shi. Evolutionary multilabel adversarial examples: An effective black-box attack. 2023. 3 +[42] Linghao Kong, Wenjian Luo, Zipeng Ye, Qi Zhou, and Yan Jia. Multi-label black-box adversarial attacks only with predicted labels. 2024. 1 +[43] A. Kurakin, I. Goodfellow, and S. Bengio. Adversarial machine learning at scale. International Conference on Learning Representations, 2017. 1 +[44] A. Kuznetsova, H. Rom, N. Alldrin, J. Uijlings, I. Krasin, J. Pont-Tuset, S. Kamali, S. Popov, M. Malloci, A. Kolesnikov, T. Duerig, and V. Ferrari. The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale. International Journal of Computer Vision, 2016. 5 +[45] J. Lanchantin, T. Wang, V. Ordonez, and Y. Qi. General multi-label image classification with transformers. IEEE Conference on Computer Vision and Pattern Recognition, 2021. 2 +[46] K. Lee, K. Lee, H. Lee, and J. Shin. A simple unified framework for detecting out-of-distribution samples and adversarial attacks. 2018. 2 + +[47] Guofa Li, Zefeng Ji, Yunlong Chang, Shen Li, Xingda Qu, and Dongpu Cao. ML-anet: A transfer learning approach using adaptation network for multi-label image classification in autonomous driving. Chinese Journal of Mechanical Engineering, 2021. 1 +[48] Maosen Li, Yanhua Yang, Kun Wei, Xu Yang, and Heng Huang. Learning universal adversarial perturbation by adversarial example. AAAI Conference on Artificial Intelligence, 2022. 3 +[49] Peng Li, Peng Chen, Yonghong Xie, and Dezheng Zhang. Bi-modal learning with channel-wise attention for multi-label image classification. IEEE Access, 2020. 2 +[50] Q. Li, M. Qiao, W. Bian, and D. Tao. Conditional graphical lasso for multi-label image classification. IEEE Conference on Computer Vision and Pattern Recognition, 2016. 2 +[51] Q. Li, X. Peng, Y. Qiao, and Q. Peng. Learning label correlations for multi-label image recognition with graph networks. Pattern Recognition Letters, 2020. 2 +[52] X. Li, F. Zhao, and Y. Guo. Multi-label image classification with a probabilistic label enhancement model. In UAI, 2014. 2 +[53] Y. Li, Y. Song, and J. Luo. Improving pairwise ranking for multi-label image classification. IEEE Conference on Computer Vision and Pattern Recognition, 2017. 2 +[54] Z. Li, W. Lu, Z. Sun, and W. Xing. Improving multi-label classification using scene cues. Multimedia Tools and Applications, 2017. 2 +[55] Dekun Lin. Probability guided loss for long-tailed multi-label image classification. 2023. 1 +[56] Tsung-Yi Lin, Michael Maire, Serge J. Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C. Lawrence Zitnick. Microsoft coco: Common objects in context. European Conference on Computer Vision, 2014. 5 +[57] An-An Liu, Zhuang Shao, Yongkang Wong, Junnan Li, Su Yu-Ting, and Mohan Kankanhalli. Lstm-based multi-label video event detection. Multimedia Tools and Applications, 78, 2019. 1 +[58] Shilong Liu, Lei Zhang, Xiao Yang, Hang Su, and Jun Zhu. Query2label: A simple transformer way to multi-label classification. arXiv preprint arXiv:2107.10834, 2021. 2 +[59] Xuannan Liu, Yaoyao Zhong, Yuhang Zhang, Lixiong Qin, and Weihong Deng. Enhancing generalization of universal adversarial perturbation through gradient aggregation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 2, 6 +[60] Yiran Liu, Xin Feng, Yunlong Wang, Wu Yang, and Di Ming. Trm-uap: Enhancing the transferability of data-free universal adversarial perturbation via truncated ratio maximization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4762-4771, 2023. 2 +[61] Mingzhi Ma, Weijie Zheng, Wanli Lv, Lu Ren, Hang Su, and Zhaoxia Yin. Multi-label adversarial attack based on label correlation. In 2023 IEEE International Conference on Image Processing (ICIP), 2023. 3 + +[62] A. Madry, A. Makelov, L. Schmidt, D. Tsipras, and A. Vladu. Towards deep learning models resistant to adversarial attacks. International Conference on Learning Representations, 2018. 2 +[63] Hassan Mahmood and Ehsan Elhamifar. Semantic-aware multi-label adversarial attacks. IEEE Conference on Computer Vision and Pattern Recognition, pages 24251-24262, 2024. 1, 3, 8 +[64] S. Melacci, G. Ciravegna, A. Sotgiu, A. Demontis, B. Biggio, M. Gori, and F. Roli. Domain knowledge alleviates adversarial attacks in multi-label classifiers. 2021. 3 +[65] J.-H. Metzen, M.-C. Kumar, T. Brox, and V. Fischer. Universal adversarial perturbations against semantic image segmentation. International Conference on Computer Vision, 2019. 1, 2 +[66] S. Moosavi-Dezfooli, A. Fawzi, and P. Frossard. Deepfool: a simple and accurate method to fool deep neural networks. IEEE Conference on Computer Vision and Pattern Recognition, 2016. 2 +[67] S. Moosavi-Dezfooli, A. Fawzi, O. Fawzi, and P. Frossard. Universal adversarial perturbations. IEEE Conference on Computer Vision and Pattern Recognition, 2017. 1, 2 +[68] Konda Reddy Mopuri, Utsav Garg, and R Venkatesh Babu. Fast feature fool: A data independent approach to universal adversarial perturbations. arXiv preprint arXiv:1707.05572, 2017. 6 +[69] Konda Reddy Mopuri, Utkarsh Ojha, Utsav Garg, and R Venkatesh Babu. Nag: Network for adversary generation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 742-751, 2018. 1, 2 +[70] J. Nam, E. L. Mencia, H. J. Kim, and J. Furnkranz. Maximizing subset accuracy with recurrent neural networks in multi-label classification. Neural Information Processing Systems, 2017. 2 +[71] T. Pang, K. Xu, C. Du, N. Chen, and J. Zhu. Improving adversarial robustness via promoting ensemble diversity. International Conference on Machine learning, 2019. 2 +[72] T. Pang, K. Xu, Y. Dong, C. Du, N. Chen, and J. Zhu. Rethinking softmax cross-entropy loss for adversarial robustness. arXiv, 2020. +[73] T. Pang, X. Yang, Y. Dong, K. Xu, H. Su, and J. Zhu. Boosting adversarial training with hypersphere embedding. Neural Information Processing Systems, 2020. 2 +[74] Nicolas Papernot, Patrick Mcdaniel, Somesh Jha, Matt Fredrikson, Z. Berkay Celik, and Ananthram Swami. The limitations of deep learning in adversarial settings. IEEE European Symposium on Security and Privacy (EuroS&P), 2016. 1 +[75] Ioannis Partalas, Aris Kosmopoulos, Nicolas Baskiotis, Thierry Artieres, George Paliouras, Eric Gaussier, Ion Androutsopoulos, Massih-Reza Amini, and Patrick Galinari. Lsjtc: A benchmark for large-scale text classification. arXiv preprint arXiv:1503.08581, 2015. 1 +[76] Omid Poursaeed, Isay Katsman, Bicheng Gao, and Serge Belongie. Generative adversarial perturbations. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4422-4431, 2018. 2 + +[77] Tao Pu, Tianshui Chen, Hefeng Wu, and Liang Lin. Semantic-aware representation blending for multi-label image recognition with partial labels. Proceedings of the AAAI Conference on Artificial Intelligence, 2022. 1 +[78] Zeyu Qin, Yanbo Fan, Yi Liu, Li Shen, Yong Zhang, Jue Wang, and Baoyuan Wu. Boosting the transferability of adversarial attacks with reverse adversarial perturbation. Advances in Neural Information Processing Systems, 35: 29845-29858, 2022. 2 +[79] Tal Ridnik, Emanuel Ben-Baruch, Nadav Zamir, Asaf Noy, Itamar Friedman, Matan Protter, and Lihi Zelnik-Manor. Asymmetric loss for multi-label classification. In Proceedings of the IEEE/CVF international conference on computer vision, pages 82-91, 2021. 5 +[80] Tal Ridnik, Gilad Sharir, Avi Ben-Cohen, Emanuel Ben-Baruch, and Asaf Noy. ML-decoder: Scalable and versatile classification head. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 32-41, 2023. 5 +[81] Jérôme Rony, Luiz G Hafemann, Luiz S Oliveira, Ismail Ben Ayed, Robert Sabourin, and Eric Granger. Decoupling direction and norm for efficient gradient-based 12 adversarial attacks and defenses. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4322-4330, 2019. 2 +[82] J. Cohen and E. Rosenfeld and Z. Kolter. Certified adversarial robustness via randomized smoothing. International Conference on Machine learning, 2019. 2 +[83] A. Shafahi, M. Najibi, A. Ghiasi, Z. Xu, J. Dickerson, C. Studer, L. Davis, G. Taylor, and T. Goldstein. Adversarial training for free! Neural Information Processing Systems, 2019. +[84] Nasim Shafiee and Ehsan Elhamifar. Zero-shot attribute attacks on fine-grained recognition models. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part V, pages 262-282. Springer, 2022. 1, 2 +[85] Jing Shao, Kai Kang, Chen Change Loy, and Xiaogang Wang. Deeply learned attributes for crowded scene understanding. In 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2015. 1 +[86] Nitish Shukla and Sudipta Banerjee. Generating adversarial attacks in the latent space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 730-739, 2023. 1 +[87] Q. Song, H. Jin, X. Huang, and X. Hu. Multi-label adversarial perturbations. IEEE International Conference on Data Mining, 2018. 3 +[88] Fengguang Su, Ou Wu, and Weiyao Zhu. Multi-label adversarial attack with new measures and self-paced constraint weighting. 2024. 1 +[89] C. Szegedy, W. Zaremba, I. Sutskever, J. Bruna, D. Erhan, I. Goodfellow, and R. Fergus. Intriguing properties of neural networks. International Conference on Learning Representations, 2014. 2 +[90] G. Tsoumakas and I. Katakis. Multi-label classification: An overview. International Journal Data Warehousing and Mining, 3, 2007. 2 + +[91] N. Tursynbek, A. Petiushko, and I. Oseledets. Geometry-inspired top-k adversarial perturbations. arXiv, 2020. 2 +[92] J. Uesato, J. B. Alayrac, P. Huang, R. Stanforth, A. Fawzi, and P. Kohli. Are labels required for improving adversarial robustness? Neural Information Processing Systems, 2019. 2 +[93] Thomas Verelst, Paul K Rubenstein, Marcin Eichner, Tinne Tuytelaars, and Maxim Berman. Spatial consistency loss for training multi-label classifiers from single-label annotations. 2023. 2 +[94] J. Wang, Y. Yang, J. Mao, Z. Huang, C. Huang, and W. Xu. Cnn-rnn: A unified framework for multi-label image classification. 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 2 +[95] Ya Wang, Dongliang He, Fu Li, Xiang Long, Zhichao Zhou, Jinwen Ma, and Shilei Wen. Multi-label classification with label graph superimposing. In Proceedings of the AAAI conference on artificial intelligence, pages 12265-12272, 2020. +[96] Z. Wang, T. Chen, G. Li, G. Li, and L. Lin. Multi-label image recognition by recurrently discovering attentional regions. IEEE International Conference on Computer Vision, 2017. 2 +[97] Yunchao Wei, Wei Xia, Min Lin, Junshi Huang, Bingbing Ni, Jian Dong, Yao Zhao, and Shuicheng Yan. Hcp: A flexible cnn framework for multi-label image classification. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2016. 2 +[98] Y. Wu, H. Liu, S. Feng, Y. Jin, G. Lyu, and Z. Wu. Gmmlic: Graph matching based multi-label image classification. International Joint Conference on Artificial Intelligence, 2021. 2 +[99] C. Xie, Z. Zhang, Y. Zhou, S. Bai, J. Wang, Z. Ren, and A. Yuille. Improving transferability of adversarial examples with input diversity. IEEE Conference on Computer Vision and Pattern Recognition, 2019. 2 +[100] Ming-Kun Xie, Jiahao Xiao, and Sheng-Jun Huang. Label-aware global consistency for multi-label learning with single positive labels. Advances in Neural Information Processing Systems, 35:18430-18441, 2022. 1 +[101] J. Xu, H. Tian, Z. Wang, Y. Wang, W. Kang, and F. Chen. Joint input and output space learning for multi-label image classification. IEEE Transactions on Multimedia, 2020. 2 +[102] W. Xu, D. Evans, and Y. Qi. Feature squeezing: Detecting adversarial examples in deep neural networks. Network and Distributed Systems Security Symposium, 2018. 2 +[103] Zhuo Yang, Yufei Han, and Xiangliang Zhang. Characterizing the evasion attackability of multi-label classifiers. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 10647-10655, 2021. 3 +[104] Zhuo Yang, Yufei Han, and Xiangliang Zhang. Attack transferability characterization for adversarially robust multi-label classification. In Machine Learning and Knowledge Discovery in Databases., pages 397-413, 2021. 3 +[105] V. O. Yazici, A. Gonzalez-Garcia, A. Ramisa, B. Twardowski, and J. v. d. Weijer. Orderless recurrent models for multi-label classification. European Conference on Computer Vision, 2020. 2 + +[106] J. Ye, J. He, X. Peng, W. Wu, and Y. Qiao. Attention-driven dynamic graph convolutional network for multi-label image recognition. European Conference on Computer Vision, 2020. +[107] R. You, Z. Guo, L. Cui, X. Long, S. Y. Bao, and S. Wen. Cross-modality attention with semantic graph embedding for multi-label classification. AAAI Conference on Artificial Intelligence, 2020. 2 +[108] Hong-Xing Yu, Wei-Shi Zheng, Ancong Wu, Xiaowei Guo, Shaogang Gong, and Jian-Huang Lai. Unsupervised person re-identification by soft multilabel learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2148–2157, 2019. 1 +[109] Chaoning Zhang, Philipp Benz, Tooba Imtiaz, and In-So Kweon. Cd-uap: Class discriminative universal adversarial perturbation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 6754–6761, 2020. 2 +[110] Chaoning Zhang, Philipp Benz, Tooba Imtiaz, and In So Kweon. Understanding adversarial examples from the mutual influence of images and perturbations. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 1, 2 +[111] Chaoning Zhang, Philipp Benz, Adil Karjauv, and In So Kweon. Universal adversarial perturbations through the lens of deep steganography: Towards a fourier perspective. In Proceedings of the AAAI conference on artificial intelligence, 2021. 2 +[112] ML. Zhang and Z. Zhou. Multilabel neural networks with applications to functional genomics and text categorization. IEEE Transactions on Knowledge and Data Engineering, 2006. 2 +[113] Shu Zhang, Ran Xu, Caiming Xiong, and Chetan Ramaiah. Use all the labels: A hierarchical multi-label contrastive learning framework. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 2 +[114] Zhe Zhao, Guangke Chen, Jingyi Wang, Yiwei Yang, Fu Song, and Jun Sun. Attack as defense: characterizing adversarial examples using robustness. In Proceedings of the 30th ACM SIGSOFT International Symposium on Software Testing and Analysis, page 42-55. Association for Computing Machinery, 2021. 2 +[115] Donghao Zhou, Pengfei Chen, Qiong Wang, Guangyong Chen, and Pheng-Ann Heng. Acknowledging the unknown for multi-label learning with single positive labels. European Conference on Computer Vision, 2022. 1 +[116] N. Zhou, W. Luo, X. Lin, P. Xu, and Z.. Zhang. Generating multi-label adversarial examples by linear programming. International Joint Conference on Neural Networks, 2020. 3 +[117] N. Zhou, W. Luo, J. Zhang, L. Kong, and H. Zhang. Hiding all labels for multi-label images: An empirical study of adversarial examples. International Joint Conference on Neural Networks, 2021. 3 +[118] Y. Zhu, J. T. Kwok, and Z. Zhou. Multi-label learning with global and local label correlation. IEEE Transactions on Knowledge and Data Engineering, 2018. 2 \ No newline at end of file diff --git a/2025/Compositional Targeted Multi-Label Universal Perturbations/images.zip b/2025/Compositional Targeted Multi-Label Universal Perturbations/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..01175d39a319b6d92403d10c260f8e88626b3db1 --- /dev/null +++ b/2025/Compositional Targeted Multi-Label Universal Perturbations/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1860d52fbc261d154f6e740c6e5cd43664fe60e59b6af4f4fbfd8c62a58a2968 +size 591447 diff --git a/2025/Compositional Targeted Multi-Label Universal Perturbations/layout.json b/2025/Compositional Targeted Multi-Label Universal Perturbations/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..3818f6a4c7b52a63ef7433f9b4bbb4d3e33ae048 --- /dev/null +++ b/2025/Compositional Targeted Multi-Label Universal Perturbations/layout.json @@ -0,0 +1,15310 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 116, + 103, + 496, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 103, + 496, + 121 + ], + "spans": [ + { + "bbox": [ + 116, + 103, + 496, + 121 + ], + "type": "text", + "content": "Compositional Targeted Multi-Label Universal Perturbations" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 144, + 268, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 144, + 268, + 171 + ], + "spans": [ + { + "bbox": [ + 138, + 144, + 268, + 171 + ], + "type": "text", + "content": "Hassan Mahmood Northeastern University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 174, + 281, + 184 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 174, + 281, + 184 + ], + "spans": [ + { + "bbox": [ + 138, + 174, + 281, + 184 + ], + "type": "text", + "content": "mahmood.h@northeastern.edu" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 315, + 144, + 468, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 144, + 468, + 185 + ], + "spans": [ + { + "bbox": [ + 315, + 144, + 468, + 185 + ], + "type": "text", + "content": "Ehsan Elhamifar Northeastern University e.elhamifar@northeastern.edu" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 152, + 213, + 200, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 213, + 200, + 225 + ], + "spans": [ + { + "bbox": [ + 152, + 213, + 200, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 238, + 297, + 466 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 238, + 297, + 466 + ], + "spans": [ + { + "bbox": [ + 55, + 238, + 297, + 466 + ], + "type": "text", + "content": "Generating targeted universal perturbations for multi-label recognition is a combinatorially hard problem that requires exponential time and space complexity. To address the problem, we propose a compositional framework. We show that a simple independence assumption on label-wise universal perturbations naturally leads to an efficient optimization that requires learning affine convex cones spanned by label-wise universal perturbations, significantly reducing the problem complexity to linear time and space. During inference, the framework allows generating universal perturbations for novel combinations of classes in constant time. We demonstrate the scalability of our method on large datasets and target sizes, evaluating its performance on NUS-WIDE, MS-COCO, and OpenImages using state-of-the-art multi-label recognition models. Our results show that our approach outperforms baselines and achieves results comparable to methods with exponential complexity. The code is available at https://github.com/hassanmahmood/UMLLAttacks.git" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 478, + 135, + 490 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 478, + 135, + 490 + ], + "spans": [ + { + "bbox": [ + 56, + 478, + 135, + 490 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 498, + 295, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 295, + 617 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 295, + 617 + ], + "type": "text", + "content": "Deep neural networks have been very successful at image-recognition, demonstrating high accuracy in various applications. However, they are also vulnerable to adversarial attacks [9, 25, 30, 31, 43, 74, 84, 86], where small changes to the input image can cause the model to misclassify the image. Notably, Moosavi-Dezfooli et al. [67] discovered that a single perturbation, called universal adversarial perturbation (UAP), can be crafted to misclassify a large number of images. This motivated a large body of work on generating universal adversarial attacks [7, 36, 38, 65, 69, 110]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 618, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 295, + 714 + ], + "type": "text", + "content": "Existing works on adversarial attacks have mainly focused on multi-class recognition (MCR) whose goal is to identify a single class in an image. On the other hand, many real-world applications require identifying all classes in an image e.g., document classification [75], scene understanding [85], facial attribute recognition [26], image annotation [15], human-object interaction [33], and several safety and security-critical applications e.g., recogniz-" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 330, + 224, + 537, + 392 + ], + "blocks": [ + { + "bbox": [ + 330, + 224, + 537, + 392 + ], + "lines": [ + { + "bbox": [ + 330, + 224, + 537, + 392 + ], + "spans": [ + { + "bbox": [ + 330, + 224, + 537, + 392 + ], + "type": "image", + "image_path": "18ff0e158377b42c24a4be820811f27dd53d336e0bef7b03aca80296df8ac134.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 397, + 555, + 453 + ], + "lines": [ + { + "bbox": [ + 313, + 397, + 555, + 453 + ], + "spans": [ + { + "bbox": [ + 313, + 397, + 555, + 453 + ], + "type": "text", + "content": "Figure 1. Targeted Multi-Label Universal Perturbations. Our framework allows composing (shown as " + }, + { + "bbox": [ + 313, + 397, + 555, + 453 + ], + "type": "inline_equation", + "content": "\\oplus" + }, + { + "bbox": [ + 313, + 397, + 555, + 453 + ], + "type": "text", + "content": ") different label-specific universal perturbations ('Animal', 'Person', 'Vehicle') to attack multiple classes together. These perturbations, once added (shown as " + }, + { + "bbox": [ + 313, + 397, + 555, + 453 + ], + "type": "inline_equation", + "content": "\\bullet" + }, + { + "bbox": [ + 313, + 397, + 555, + 453 + ], + "type": "text", + "content": ") to an image, cause target labels to be misclassified." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 473, + 555, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 473, + 555, + 593 + ], + "spans": [ + { + "bbox": [ + 313, + 473, + 555, + 593 + ], + "type": "text", + "content": "ing diseases [24], autonomous driving [47], and surveillance [57, 108]. Therefore, the goal of multi-label learning (MLL) is to identify all labels present in an image [14, 20, 23, 39, 40, 55, 77, 100, 115]. Given the increasing number of applications relying on MLL, it is crucial to understand the internal workings of the MLL models, identify their vulnerabilities, and make them robust. Recently, MLL models have been shown to be vulnerable to adversarial attacks [1, 2, 42, 63, 88], yet the vulnerability of MLL systems against universal perturbations has not been explored." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "content": "Crafting universal perturbations for MCR and MLL have fundamental differences due to the distinct structuring of decision boundaries in these problems. In MCR, a universal perturbation to turn off a label in images will result in the prediction of another label. Conversely, in MLL, multiple labels can co-exist, and the objective of an attack might be to misclassify any subset of these labels either by turning the labels off or on (see Fig. 2). In MLL, any combination of labels can be present in images. For the label set " + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "content": ", there are " + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "inline_equation", + "content": "2^{|C|}" + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "content": " possible combinations of labels. Universal pertur" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20580" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 54, + 72, + 294, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 72, + 294, + 144 + ], + "spans": [ + { + "bbox": [ + 54, + 72, + 294, + 144 + ], + "type": "text", + "content": "bations identify specific directions in the input space where the classifier is consistently vulnerable to attacks. What makes this particularly concerning in multi-label learning is that there can be exponentially many directions in the input space to attack the model - corresponding to different combinations of labels." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 144, + 294, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 144, + 294, + 251 + ], + "spans": [ + { + "bbox": [ + 55, + 144, + 294, + 251 + ], + "type": "text", + "content": "One approach for generating universal perturbations for MLL is to generate universal vectors that flip the prediction of all or most labels. However, in practice, the adversary usually seeks to modify a specific small set of labels while keeping the prediction of others unchanged (see Fig. 1). This is called the targeted attack and requires learning " + }, + { + "bbox": [ + 55, + 144, + 294, + 251 + ], + "type": "inline_equation", + "content": "2^{|\\mathcal{C}|}" + }, + { + "bbox": [ + 55, + 144, + 294, + 251 + ], + "type": "text", + "content": " distinct universal perturbation vectors. Learning each such universal vector requires iterating through the dataset, resulting in an exponential time and space complexity." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 252, + 294, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 252, + 294, + 371 + ], + "spans": [ + { + "bbox": [ + 55, + 252, + 294, + 371 + ], + "type": "text", + "content": "Existing universal perturbation methods focus on MCR setting [7, 8, 60]. Adapting these methods to MLL would require learning a separate universal perturbation for each possible label combination, incurring exponential complexity. Although one could learn label-wise independent perturbations, no effective method exists to combine them for attacking specific label combinations. We empirically show that naively combining them does not lead to successful attacks. Similarly, generative attack approaches [27, 69, 76] would be ineffective for unseen label combinations." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 372, + 294, + 527 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 372, + 294, + 527 + ], + "spans": [ + { + "bbox": [ + 55, + 372, + 294, + 527 + ], + "type": "text", + "content": "Contributions. In this paper, we develop an efficient method for crafting targeted universal adversarial attacks on multi-label learning (MLL) models, addressing the aforementioned combinatorial complexity problem. Our key assumption is to generate a universal attack on a target set as a composition of universal label-wise attacks of the target labels (see Fig. 1). Based on this assumption, we derive a formulation which requires learning affine convex cones spanned by label-wise universal vectors. Our formulation requires linear training time complexity and after trained, the universal perturbation to attack any combination of labels can be composed in constant time using the label-wise attack vectors. Our key contributions are:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 529, + 294, + 708 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 55, + 529, + 294, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 529, + 294, + 577 + ], + "spans": [ + { + "bbox": [ + 55, + 529, + 294, + 577 + ], + "type": "text", + "content": "- Existing MCR-based approaches are not inherently designed to attack multiple targeted labels for MLL models. Our work is the first to address this gap, offering an efficient and principled solution for MLL setting;" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 578, + 294, + 637 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 578, + 294, + 637 + ], + "spans": [ + { + "bbox": [ + 55, + 578, + 294, + 637 + ], + "type": "text", + "content": "- Our proposed framework to learn targeted universal perturbations for MLL has linear time and space complexity and constructs attacks on any target set in constant time, while achieving performance comparable to the existing oracle methods with exponential complexity;" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 638, + 294, + 660 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 638, + 294, + 660 + ], + "spans": [ + { + "bbox": [ + 55, + 638, + 294, + 660 + ], + "type": "text", + "content": "- Our results visualize and provide deeper understanding of the universal adversarial regions learnt by our method;" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 661, + 294, + 708 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 661, + 294, + 708 + ], + "spans": [ + { + "bbox": [ + 55, + 661, + 294, + 708 + ], + "type": "text", + "content": "- We evaluate our method on NUS-WIDE, MS-COCO, and OpenImages benchmark datasets using state-of-the-art multi-label learning models against various baselines, showing the effectiveness of our method." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 71, + 406, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 71, + 406, + 83 + ], + "spans": [ + { + "bbox": [ + 314, + 71, + 406, + 83 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 91, + 452, + 104 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 91, + 452, + 104 + ], + "spans": [ + { + "bbox": [ + 314, + 91, + 452, + 104 + ], + "type": "text", + "content": "2.1. Multi-Label Recognition" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 110, + 553, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 110, + 553, + 301 + ], + "spans": [ + { + "bbox": [ + 313, + 110, + 553, + 301 + ], + "type": "text", + "content": "There has been significant recent interest in multi-label learning due to its applications in various tasks. In MLL, the goal is to identify all concepts/labels in an image. This is particularly a harder task than multi-class recognition, since there can be exponentially many possible combinations of classes present in an image and can be related to each other. To address this problem, early works [90, 97] considered independent binary classifiers to classify the presence or absence of each class. However, since such a method does not consider the correlations among labels, several other approaches have been proposed that exploit label correlations or external knowledge [13, 18, 34, 45, 49, 51, 70, 94-96, 98, 101, 105-107, 113, 118], or use graphical models [50, 52]. More recent works use attention mechanism or novel loss functions to consider the positive and negative sample imbalance [6, 17, 53, 54, 58, 93, 112]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 311, + 430, + 322 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 311, + 430, + 322 + ], + "spans": [ + { + "bbox": [ + 314, + 311, + 430, + 322 + ], + "type": "text", + "content": "2.2. Adversarial Attacks" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 329, + 553, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 329, + 553, + 484 + ], + "spans": [ + { + "bbox": [ + 313, + 329, + 553, + 484 + ], + "type": "text", + "content": "Adversarial attacks [3, 12, 81, 89, 99] have gained a lot of attention as they identify the key vulnerability of deep neural networks (DNNs) to small imperceptible changes in the input. Researchers have explored various ways that attackers can manipulate data to fool models and developed several defenses to counter these attacks [4, 5, 10, 11, 19, 21, 22, 28, 29, 35, 46, 62, 71-73, 78, 82-84, 91, 92, 102, 114]. Based on the attacker's objective and the target setting, the adversarial attacks can be categorized in different ways, such as targeted and untargeted attacks, white-box (which assume full access to the target model parameters) and black-box attacks (no access to the model parameters), or instance-specific and instance-agnostic attacks." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 314, + 491, + 479, + 503 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 491, + 479, + 503 + ], + "spans": [ + { + "bbox": [ + 314, + 491, + 479, + 503 + ], + "type": "text", + "content": "2.3. Universal Adversarial Attacks" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 510, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 553, + 713 + ], + "type": "text", + "content": "In instance-specific attacks, the perturbations are generated and targeted for each input image, whereas, instance-agnostic attacks [38, 65, 69] generate a single universal perturbation that can fool all input instances. These universal adversarial perturbations(UAPs) can be untargeted or targeted. The goal of the untargeted universal perturbation is to perturb any image such that it is misclassified by the target model [67, 76]. In the targeted case, the goal is to learn the universal perturbation such that it misclassifies any given image to a specific class [7, 110]. Universal perturbations were first proposed by Moosavi-Dezfooli et al. [67] and were generated by iteratively computing and aggregating perturbations using DeepFool [66] across multiple inputs. Since then, several methods have been proposed to learn better universal perturbations [60, 109–111]. To learn UAPs with better generalization, Liu et al. [59] proposed a gradient aggregation method to address gradient vanish" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20581" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 80, + 96, + 225, + 222 + ], + "blocks": [ + { + "bbox": [ + 115, + 78, + 197, + 89 + ], + "lines": [ + { + "bbox": [ + 115, + 78, + 197, + 89 + ], + "spans": [ + { + "bbox": [ + 115, + 78, + 197, + 89 + ], + "type": "text", + "content": "Current Approach" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 80, + 96, + 225, + 222 + ], + "lines": [ + { + "bbox": [ + 80, + 96, + 225, + 222 + ], + "spans": [ + { + "bbox": [ + 80, + 96, + 225, + 222 + ], + "type": "image", + "image_path": "26d5e55b3778924cabe3c9374155441b5ea7a5a4a96a5c55e9b4f5499ef65abd.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 227, + 555, + 271 + ], + "lines": [ + { + "bbox": [ + 54, + 227, + 555, + 271 + ], + "spans": [ + { + "bbox": [ + 54, + 227, + 555, + 271 + ], + "type": "text", + "content": "Figure 2. In multi-label recognition, images can have any combination of labels present, leading to a combinatorial explosion in possible universal perturbations. Traditional methods (shown on the left) require exponential time and space complexity to learn perturbations for all class combinations. Our approach (shown on the right) reduces this complexity by learning " + }, + { + "bbox": [ + 54, + 227, + 555, + 271 + ], + "type": "inline_equation", + "content": "|\\mathcal{C}|" + }, + { + "bbox": [ + 54, + 227, + 555, + 271 + ], + "type": "text", + "content": " affine convex cones, each representing a region of universal perturbations for a specific class. This allows us to efficiently generate perturbations for any combination of classes." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 233, + 92, + 362, + 219 + ], + "blocks": [ + { + "bbox": [ + 294, + 78, + 447, + 89 + ], + "lines": [ + { + "bbox": [ + 294, + 78, + 447, + 89 + ], + "spans": [ + { + "bbox": [ + 294, + 78, + 447, + 89 + ], + "type": "text", + "content": "Overview of the Proposed Method" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 233, + 92, + 362, + 219 + ], + "lines": [ + { + "bbox": [ + 233, + 92, + 362, + 219 + ], + "spans": [ + { + "bbox": [ + 233, + 92, + 362, + 219 + ], + "type": "image", + "image_path": "9793affce6d740d16f153eabf1d656ee39b822d0ed8998a83d9f7682c8617ad9.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 395, + 95, + 533, + 214 + ], + "blocks": [ + { + "bbox": [ + 395, + 95, + 533, + 214 + ], + "lines": [ + { + "bbox": [ + 395, + 95, + 533, + 214 + ], + "spans": [ + { + "bbox": [ + 395, + 95, + 533, + 214 + ], + "type": "image", + "image_path": "1eea43c9d7a6611c4cbdcbdc9def97ea6207d7a227127c44538e2282a68805be.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 289, + 295, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 289, + 295, + 314 + ], + "spans": [ + { + "bbox": [ + 54, + 289, + 295, + 314 + ], + "type": "text", + "content": "ing problem and Li et al. [48] proposed to utilize instance-specific attack method to generate dominant UAPs." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 323, + 231, + 335 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 323, + 231, + 335 + ], + "spans": [ + { + "bbox": [ + 55, + 323, + 231, + 335 + ], + "type": "text", + "content": "2.4. Multi-Label Adversarial Attacks" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 340, + 296, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 340, + 296, + 579 + ], + "spans": [ + { + "bbox": [ + 56, + 340, + 296, + 579 + ], + "type": "text", + "content": "Many recent works have studied adversarial attacks for multi-label recognition models [41, 61, 103, 104, 116, 117]. Song et al. [87] proposed white-box multi-label attack framework and Hu et al. [32] proposed to exploit ranking relations to design attacks for top-" + }, + { + "bbox": [ + 56, + 340, + 296, + 579 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 56, + 340, + 296, + 579 + ], + "type": "text", + "content": " multi-label models. Aich et al. [1] proposed a CLIP-based generative model to generate multi-object attacks in the black-box setting. Jia et al. [37] proposed theoretical robustness guarantees to defend against multi-label adversarial attacks and Melacci et al. [64] exploited domain knowledge context to detect multi-label adversarial attacks. Mahmood and Elhamifar [63] proposed an optimization to address the problem of negative gradient correlation in generating multi-label attacks. These works mainly address instance-specific perturbations for multi-label setting. To the best of our knowledge, only Hu et al. [32] proposed universal untargeted attack where a single perturbation was learnt to attack and replace top-" + }, + { + "bbox": [ + 56, + 340, + 296, + 579 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 56, + 340, + 296, + 579 + ], + "type": "text", + "content": " predicted labels across images. In comparison, we address the problem of learning targeted universal perturbations to attack specific combinations of labels." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 592, + 258, + 604 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 592, + 258, + 604 + ], + "spans": [ + { + "bbox": [ + 55, + 592, + 258, + 604 + ], + "type": "text", + "content": "3. Multi-Label Universal Perturbations" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 611, + 180, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 611, + 180, + 623 + ], + "spans": [ + { + "bbox": [ + 55, + 611, + 180, + 623 + ], + "type": "text", + "content": "3.1. Problem Formulation" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 628, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 628, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 628, + 296, + 713 + ], + "type": "text", + "content": "Consider a multi-label classifier " + }, + { + "bbox": [ + 55, + 628, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{F}:\\mathbb{R}^d\\to \\mathbb{R}^{|\\mathcal{C}|}" + }, + { + "bbox": [ + 55, + 628, + 296, + 713 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 55, + 628, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 55, + 628, + 296, + 713 + ], + "type": "text", + "content": " is the set of all labels. For an image " + }, + { + "bbox": [ + 55, + 628, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\pmb {x}\\in \\mathbb{R}^d" + }, + { + "bbox": [ + 55, + 628, + 296, + 713 + ], + "type": "text", + "content": " , let " + }, + { + "bbox": [ + 55, + 628, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\pmb {y}\\in \\{0,1\\}^{|\\mathcal{C}|}" + }, + { + "bbox": [ + 55, + 628, + 296, + 713 + ], + "type": "text", + "content": " denote the set of its labels, where each entry indicates the absence (0) or presence (1) of the corresponding label in " + }, + { + "bbox": [ + 55, + 628, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 55, + 628, + 296, + 713 + ], + "type": "text", + "content": " in the image. The classifier " + }, + { + "bbox": [ + 55, + 628, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{F} = \\{\\mathcal{F}_1,\\mathcal{F}_2,\\dots ,\\mathcal{F}_{|\\mathcal{C}|}\\}" + }, + { + "bbox": [ + 55, + 628, + 296, + 713 + ], + "type": "text", + "content": " consists of " + }, + { + "bbox": [ + 55, + 628, + 296, + 713 + ], + "type": "inline_equation", + "content": "|\\mathcal{C}|" + }, + { + "bbox": [ + 55, + 628, + 296, + 713 + ], + "type": "text", + "content": " binary classifiers (one for each label), where " + }, + { + "bbox": [ + 55, + 628, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_c(\\pmb {x})\\in (-\\infty , + \\infty)" + }, + { + "bbox": [ + 55, + 628, + 296, + 713 + ], + "type": "text", + "content": " represents the logit of label " + }, + { + "bbox": [ + 55, + 628, + 296, + 713 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 628, + 296, + 713 + ], + "type": "text", + "content": " .The" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 289, + 555, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 289, + 555, + 373 + ], + "spans": [ + { + "bbox": [ + 313, + 289, + 555, + 373 + ], + "type": "text", + "content": "probability of label " + }, + { + "bbox": [ + 313, + 289, + 555, + 373 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 289, + 555, + 373 + ], + "type": "text", + "content": " being present in the image is given by " + }, + { + "bbox": [ + 313, + 289, + 555, + 373 + ], + "type": "inline_equation", + "content": "\\hat{o}_c = \\sigma (\\mathcal{F}_c(\\pmb {x}))" + }, + { + "bbox": [ + 313, + 289, + 555, + 373 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 289, + 555, + 373 + ], + "type": "inline_equation", + "content": "\\sigma (.)" + }, + { + "bbox": [ + 313, + 289, + 555, + 373 + ], + "type": "text", + "content": " is the sigmoid function, hence, it is considered present (1), if " + }, + { + "bbox": [ + 313, + 289, + 555, + 373 + ], + "type": "inline_equation", + "content": "\\hat{o}_c\\geq 0.5" + }, + { + "bbox": [ + 313, + 289, + 555, + 373 + ], + "type": "text", + "content": ", and absent (0) otherwise. We assume that " + }, + { + "bbox": [ + 313, + 289, + 555, + 373 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 313, + 289, + 555, + 373 + ], + "type": "text", + "content": " has already been trained on a set of images. Let " + }, + { + "bbox": [ + 313, + 289, + 555, + 373 + ], + "type": "inline_equation", + "content": "\\Omega \\subseteq \\mathcal{C}" + }, + { + "bbox": [ + 313, + 289, + 555, + 373 + ], + "type": "text", + "content": " denote the set of target labels that we want to attack. Our goal is to generate a targeted universal perturbation for " + }, + { + "bbox": [ + 313, + 289, + 555, + 373 + ], + "type": "inline_equation", + "content": "\\Omega" + }, + { + "bbox": [ + 313, + 289, + 555, + 373 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 381, + 509, + 393 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 381, + 509, + 393 + ], + "spans": [ + { + "bbox": [ + 313, + 381, + 509, + 393 + ], + "type": "text", + "content": "3.2. Multi-Label Universal Perturbations" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 399, + 555, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 399, + 555, + 532 + ], + "spans": [ + { + "bbox": [ + 313, + 399, + 555, + 532 + ], + "type": "text", + "content": "We want to find a single (universal) perturbation vector, " + }, + { + "bbox": [ + 313, + 399, + 555, + 532 + ], + "type": "inline_equation", + "content": "\\pmb{u}_{\\Omega}" + }, + { + "bbox": [ + 313, + 399, + 555, + 532 + ], + "type": "text", + "content": ", that can attack a set of target labels in " + }, + { + "bbox": [ + 313, + 399, + 555, + 532 + ], + "type": "inline_equation", + "content": "\\Omega" + }, + { + "bbox": [ + 313, + 399, + 555, + 532 + ], + "type": "text", + "content": " in any image " + }, + { + "bbox": [ + 313, + 399, + 555, + 532 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 313, + 399, + 555, + 532 + ], + "type": "text", + "content": " containing them. One can generate such a vector by maximizing the expected multi-label learning loss for the labels in " + }, + { + "bbox": [ + 313, + 399, + 555, + 532 + ], + "type": "inline_equation", + "content": "\\Omega" + }, + { + "bbox": [ + 313, + 399, + 555, + 532 + ], + "type": "text", + "content": " across the training dataset, while restricting the norm of " + }, + { + "bbox": [ + 313, + 399, + 555, + 532 + ], + "type": "inline_equation", + "content": "\\pmb{u}_{\\Omega}" + }, + { + "bbox": [ + 313, + 399, + 555, + 532 + ], + "type": "text", + "content": " (we consider " + }, + { + "bbox": [ + 313, + 399, + 555, + 532 + ], + "type": "inline_equation", + "content": "\\ell_{\\infty}" + }, + { + "bbox": [ + 313, + 399, + 555, + 532 + ], + "type": "text", + "content": "-norm). However, due to correlations among labels in MLL, attacking the target labels may inadvertently affect non-target labels " + }, + { + "bbox": [ + 313, + 399, + 555, + 532 + ], + "type": "inline_equation", + "content": "(\\bar{\\Omega} = \\mathcal{C} \\setminus \\Omega)" + }, + { + "bbox": [ + 313, + 399, + 555, + 532 + ], + "type": "text", + "content": ". To mitigate this, we seek to minimize the influence of the universal perturbation vector on non-target labels. This requires solving the following oracle optimization," + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 354, + 536, + 555, + 559 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 536, + 555, + 559 + ], + "spans": [ + { + "bbox": [ + 354, + 536, + 555, + 559 + ], + "type": "interline_equation", + "content": "P _ {\\Omega} ^ {*}: \\max _ {\\boldsymbol {u} _ {\\Omega}} \\sum_ {k \\in \\Omega} \\mathcal {H} _ {k} (\\boldsymbol {u} _ {\\Omega}) - \\gamma \\sum_ {q \\in \\Omega} \\mathcal {R} _ {q} (\\boldsymbol {u} _ {\\Omega}), \\tag {1}", + "image_path": "d05a45bcdf9fcf3491bd51156508383af1d04efd28bc79b327465cf9afd3b1af.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 563, + 554, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 563, + 554, + 624 + ], + "spans": [ + { + "bbox": [ + 313, + 563, + 554, + 624 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 563, + 554, + 624 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_k" + }, + { + "bbox": [ + 313, + 563, + 554, + 624 + ], + "type": "text", + "content": " ensures attacking the label " + }, + { + "bbox": [ + 313, + 563, + 554, + 624 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 563, + 554, + 624 + ], + "type": "text", + "content": " in the target set and " + }, + { + "bbox": [ + 313, + 563, + 554, + 624 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_q" + }, + { + "bbox": [ + 313, + 563, + 554, + 624 + ], + "type": "text", + "content": " tries to prevent modifying the prediction of non-target labels using the universal perturbation " + }, + { + "bbox": [ + 313, + 563, + 554, + 624 + ], + "type": "inline_equation", + "content": "\\pmb{u}_{\\Omega}" + }, + { + "bbox": [ + 313, + 563, + 554, + 624 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 313, + 563, + 554, + 624 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 313, + 563, + 554, + 624 + ], + "type": "text", + "content": " is a scalar hyperparameter setting a trade-off between the two terms. More specifically, we define " + }, + { + "bbox": [ + 313, + 563, + 554, + 624 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_k" + }, + { + "bbox": [ + 313, + 563, + 554, + 624 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 563, + 554, + 624 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_q" + }, + { + "bbox": [ + 313, + 563, + 554, + 624 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 314, + 629, + 553, + 643 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 629, + 553, + 643 + ], + "spans": [ + { + "bbox": [ + 314, + 629, + 553, + 643 + ], + "type": "interline_equation", + "content": "\\mathcal {H} _ {k} \\left(\\boldsymbol {u} _ {\\Omega}\\right) \\triangleq \\mathbb {E} _ {\\left(\\boldsymbol {x}, \\boldsymbol {y}\\right)} \\mathcal {L} _ {\\mathrm {b c e}} \\left(\\mathcal {F} _ {k} \\left(\\boldsymbol {x} + \\Pi_ {\\epsilon} \\left(\\boldsymbol {u} _ {\\Omega}\\right)\\right), y _ {k}\\right), \\tag {2}", + "image_path": "ef6d5ba12f67557c22df05211442d76d032840bf24eeabe598db36a8b5e94644.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 315, + 645, + 553, + 659 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 645, + 553, + 659 + ], + "spans": [ + { + "bbox": [ + 315, + 645, + 553, + 659 + ], + "type": "interline_equation", + "content": "\\mathcal {R} _ {q} (\\boldsymbol {u} _ {\\Omega}) \\triangleq \\mathbb {E} _ {(\\boldsymbol {x}, \\boldsymbol {y})} \\max \\left\\{0, - \\tanh \\left(\\mathcal {F} _ {q} (\\boldsymbol {x}) \\mathcal {F} _ {q} (\\boldsymbol {x} + \\Pi_ {\\epsilon} (\\boldsymbol {u} _ {\\Omega}))\\right) \\right\\},", + "image_path": "6314441f0f0be6a8cd611d1a9c66d8cea8deeeff9593dacb282fe3863937243e.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\Pi_{\\epsilon}" + }, + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "type": "text", + "content": " projects the input on p-norm ball of radius " + }, + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{bce}}" + }, + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "type": "text", + "content": " is the binary cross-entropy loss and the expectation is with respect to samples from the dataset. Therefore, maximizing " + }, + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_k(\\pmb{u}_{\\Omega})" + }, + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "type": "text", + "content": " ensures to find a perturbation that modifies the" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20582" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": "prediction of each target label " + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": ". On the other hand, " + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_k" + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": " aims to minimize the influence of the perturbation " + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "inline_equation", + "content": "(\\pmb{u}_{\\Omega})" + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": " on each non-target label " + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": ". This is done by incurring a positive penalty when the signs of the logits for the unperturbed and perturbed inputs differ from each other. Thus, minimizing this term leads to the same signs for the logits of each non-target label before and after the attack." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 162, + 296, + 186 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 162, + 296, + 186 + ], + "spans": [ + { + "bbox": [ + 55, + 162, + 296, + 186 + ], + "type": "text", + "content": "3.3. Compositional Multi-Label Universal Attacks (CMLU)" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 190, + 296, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 190, + 296, + 262 + ], + "spans": [ + { + "bbox": [ + 55, + 190, + 296, + 262 + ], + "type": "text", + "content": "There are " + }, + { + "bbox": [ + 55, + 190, + 296, + 262 + ], + "type": "inline_equation", + "content": "2^{|C|}" + }, + { + "bbox": [ + 55, + 190, + 296, + 262 + ], + "type": "text", + "content": " possible subsets of labels " + }, + { + "bbox": [ + 55, + 190, + 296, + 262 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 55, + 190, + 296, + 262 + ], + "type": "text", + "content": ". Therefore, finding all universal perturbations " + }, + { + "bbox": [ + 55, + 190, + 296, + 262 + ], + "type": "inline_equation", + "content": "\\{\\pmb{u}_{\\Omega}\\}" + }, + { + "bbox": [ + 55, + 190, + 296, + 262 + ], + "type": "text", + "content": " by independently solving (1) for each subset " + }, + { + "bbox": [ + 55, + 190, + 296, + 262 + ], + "type": "inline_equation", + "content": "\\Omega" + }, + { + "bbox": [ + 55, + 190, + 296, + 262 + ], + "type": "text", + "content": " leads to the exponential complexity " + }, + { + "bbox": [ + 55, + 190, + 296, + 262 + ], + "type": "inline_equation", + "content": "O(2^{|C|})" + }, + { + "bbox": [ + 55, + 190, + 296, + 262 + ], + "type": "text", + "content": ". We develop a compositional multi-label universal attack (CMLU) that allows us to find any universal perturbation in linear complexity " + }, + { + "bbox": [ + 55, + 190, + 296, + 262 + ], + "type": "inline_equation", + "content": "O(|C|)" + }, + { + "bbox": [ + 55, + 190, + 296, + 262 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 263, + 296, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 263, + 296, + 299 + ], + "spans": [ + { + "bbox": [ + 55, + 263, + 296, + 299 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 55, + 263, + 296, + 299 + ], + "type": "inline_equation", + "content": "\\mathcal{P}(\\mathcal{C})" + }, + { + "bbox": [ + 55, + 263, + 296, + 299 + ], + "type": "text", + "content": " denote the power set of " + }, + { + "bbox": [ + 55, + 263, + 296, + 299 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 55, + 263, + 296, + 299 + ], + "type": "text", + "content": ", i.e., the set of all subsets of " + }, + { + "bbox": [ + 55, + 263, + 296, + 299 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 55, + 263, + 296, + 299 + ], + "type": "text", + "content": ". Our goal is to find all " + }, + { + "bbox": [ + 55, + 263, + 296, + 299 + ], + "type": "inline_equation", + "content": "\\pmb{u}_{\\Omega}" + }, + { + "bbox": [ + 55, + 263, + 296, + 299 + ], + "type": "text", + "content": "'s by solving (1) for every " + }, + { + "bbox": [ + 55, + 263, + 296, + 299 + ], + "type": "inline_equation", + "content": "\\Omega \\in \\mathcal{P}(\\mathcal{C})" + }, + { + "bbox": [ + 55, + 263, + 296, + 299 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 81, + 299, + 294, + 327 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 299, + 294, + 327 + ], + "spans": [ + { + "bbox": [ + 81, + 299, + 294, + 327 + ], + "type": "interline_equation", + "content": "\\sum_ {\\Omega \\in \\mathcal {P} (\\mathcal {C})} \\max _ {\\boldsymbol {u} _ {\\Omega}} \\left[ \\sum_ {k \\in \\Omega} \\mathcal {H} _ {k} \\left(\\boldsymbol {u} _ {\\Omega}\\right) - \\gamma \\sum_ {q \\in \\bar {\\Omega}} \\mathcal {R} _ {q} \\left(\\boldsymbol {u} _ {\\Omega}\\right) \\right]. \\tag {3}", + "image_path": "a9892cdd3277878259b1734f940b31e1e5fd7d492a24fd6c10451c65f54af1bb.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 329, + 296, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 329, + 296, + 376 + ], + "spans": [ + { + "bbox": [ + 55, + 329, + 296, + 376 + ], + "type": "text", + "content": "As mentioned above, solving (3) over the power set of " + }, + { + "bbox": [ + 55, + 329, + 296, + 376 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 55, + 329, + 296, + 376 + ], + "type": "text", + "content": " has exponential complexity. To tackle this problem, we make the following assumption: each universal perturbation " + }, + { + "bbox": [ + 55, + 329, + 296, + 376 + ], + "type": "inline_equation", + "content": "\\pmb{u}_{\\Omega}" + }, + { + "bbox": [ + 55, + 329, + 296, + 376 + ], + "type": "text", + "content": " can be written as the sum of atomic universal perturbations" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 146, + 379, + 294, + 403 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 379, + 294, + 403 + ], + "spans": [ + { + "bbox": [ + 146, + 379, + 294, + 403 + ], + "type": "interline_equation", + "content": "\\boldsymbol {u} _ {\\Omega} \\triangleq \\sum_ {i \\in \\Omega} \\boldsymbol {u} _ {i}, \\tag {4}", + "image_path": "cb26d66f0cbcae3c244ee587318872a158eae0a8e1c7db99c785c16809d8e836.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 406, + 296, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 406, + 296, + 456 + ], + "spans": [ + { + "bbox": [ + 55, + 406, + 296, + 456 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 406, + 296, + 456 + ], + "type": "inline_equation", + "content": "\\pmb{u}_i" + }, + { + "bbox": [ + 55, + 406, + 296, + 456 + ], + "type": "text", + "content": " is the universal perturbation on the label " + }, + { + "bbox": [ + 55, + 406, + 296, + 456 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 406, + 296, + 456 + ], + "type": "text", + "content": ". We also assume that the label-wise universal perturbations are norm-constrained and mutually orthogonal, i.e., " + }, + { + "bbox": [ + 55, + 406, + 296, + 456 + ], + "type": "inline_equation", + "content": "\\| \\pmb{u}_i\\|_p \\leq \\epsilon" + }, + { + "bbox": [ + 55, + 406, + 296, + 456 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 406, + 296, + 456 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{u}}_i^\\top \\hat{\\pmb{u}}_j = 0" + }, + { + "bbox": [ + 55, + 406, + 296, + 456 + ], + "type": "text", + "content": " for each " + }, + { + "bbox": [ + 55, + 406, + 296, + 456 + ], + "type": "inline_equation", + "content": "i \\neq j" + }, + { + "bbox": [ + 55, + 406, + 296, + 456 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 406, + 296, + 456 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{u}} = \\pmb{u} / \\| \\pmb{u}\\|_2" + }, + { + "bbox": [ + 55, + 406, + 296, + 456 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 455, + 296, + 527 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 455, + 296, + 527 + ], + "spans": [ + { + "bbox": [ + 55, + 455, + 296, + 527 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 55, + 455, + 296, + 527 + ], + "type": "inline_equation", + "content": "\\mathbf{U} = \\left[ \\begin{array}{lll}\\pmb {u}_1 & \\pmb {u}_2 & \\dots & \\pmb{u}_{|\\mathcal{C}|} \\end{array} \\right]" + }, + { + "bbox": [ + 55, + 455, + 296, + 527 + ], + "type": "text", + "content": " represent a matrix whose columns are label-wise universal perturbations and " + }, + { + "bbox": [ + 55, + 455, + 296, + 527 + ], + "type": "inline_equation", + "content": "\\lambda_{\\Omega}" + }, + { + "bbox": [ + 55, + 455, + 296, + 527 + ], + "type": "text", + "content": " denote a vector whose elements indexed by " + }, + { + "bbox": [ + 55, + 455, + 296, + 527 + ], + "type": "inline_equation", + "content": "\\Omega" + }, + { + "bbox": [ + 55, + 455, + 296, + 527 + ], + "type": "text", + "content": " are one and other elements are zero. Based on our assumption, we optimize " + }, + { + "bbox": [ + 55, + 455, + 296, + 527 + ], + "type": "inline_equation", + "content": "\\mathbf{U}" + }, + { + "bbox": [ + 55, + 455, + 296, + 527 + ], + "type": "text", + "content": " so that we can construct any " + }, + { + "bbox": [ + 55, + 455, + 296, + 527 + ], + "type": "inline_equation", + "content": "\\pmb{u}_{\\Omega}" + }, + { + "bbox": [ + 55, + 455, + 296, + 527 + ], + "type": "text", + "content": " using the columns of " + }, + { + "bbox": [ + 55, + 455, + 296, + 527 + ], + "type": "inline_equation", + "content": "\\mathbf{U}" + }, + { + "bbox": [ + 55, + 455, + 296, + 527 + ], + "type": "text", + "content": ". Given our assumptions, we can rewrite (3) as" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 57, + 529, + 294, + 571 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 529, + 294, + 571 + ], + "spans": [ + { + "bbox": [ + 57, + 529, + 294, + 571 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\max _ {\\boldsymbol {U}} \\sum_ {\\Omega \\in \\mathcal {P} (\\mathcal {C})} \\sum_ {k \\in \\Omega} \\left(\\mathcal {H} _ {k} \\left(\\boldsymbol {u} _ {k} + \\boldsymbol {U} \\boldsymbol {\\lambda} _ {\\Omega \\backslash k}\\right) - \\frac {\\gamma}{| \\Omega |} \\sum_ {q \\in \\Omega} \\mathcal {R} _ {q} \\left(\\boldsymbol {u} _ {k} + \\boldsymbol {U} \\boldsymbol {\\lambda} _ {\\Omega \\backslash k}\\right)\\right), \\\\ s. t. \\hat {\\boldsymbol {U}} ^ {\\top} \\hat {\\boldsymbol {U}} = \\boldsymbol {I}, \\| \\boldsymbol {U} \\| _ {1, p} \\leq \\epsilon \\tag {5} \\\\ \\end{array}", + "image_path": "d6e6f17c23809eed24d4885e38b096579e09d446d6a3de645d29b7047317c70a.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 574, + 296, + 623 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 574, + 296, + 623 + ], + "spans": [ + { + "bbox": [ + 55, + 574, + 296, + 623 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 574, + 296, + 623 + ], + "type": "inline_equation", + "content": "\\hat{U}" + }, + { + "bbox": [ + 55, + 574, + 296, + 623 + ], + "type": "text", + "content": " is column-normalized " + }, + { + "bbox": [ + 55, + 574, + 296, + 623 + ], + "type": "inline_equation", + "content": "U" + }, + { + "bbox": [ + 55, + 574, + 296, + 623 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 574, + 296, + 623 + ], + "type": "inline_equation", + "content": "\\| U\\|_{1,p} = \\max_{1\\leq i\\leq |\\mathcal{C}|}\\| u_i\\| _p" + }, + { + "bbox": [ + 55, + 574, + 296, + 623 + ], + "type": "text", + "content": ", is the maximum norm of the columns of " + }, + { + "bbox": [ + 55, + 574, + 296, + 623 + ], + "type": "inline_equation", + "content": "U" + }, + { + "bbox": [ + 55, + 574, + 296, + 623 + ], + "type": "text", + "content": ". Notice that we wrote " + }, + { + "bbox": [ + 55, + 574, + 296, + 623 + ], + "type": "inline_equation", + "content": "u_{\\Omega} = u_k + U\\lambda_{\\Omega \\backslash k}" + }, + { + "bbox": [ + 55, + 574, + 296, + 623 + ], + "type": "text", + "content": " for a " + }, + { + "bbox": [ + 55, + 574, + 296, + 623 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 574, + 296, + 623 + ], + "type": "text", + "content": " that belongs to " + }, + { + "bbox": [ + 55, + 574, + 296, + 623 + ], + "type": "inline_equation", + "content": "\\Omega" + }, + { + "bbox": [ + 55, + 574, + 296, + 623 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 574, + 296, + 623 + ], + "type": "inline_equation", + "content": "\\Omega \\backslash k" + }, + { + "bbox": [ + 55, + 574, + 296, + 623 + ], + "type": "text", + "content": " denotes removing " + }, + { + "bbox": [ + 55, + 574, + 296, + 623 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 574, + 296, + 623 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 55, + 574, + 296, + 623 + ], + "type": "inline_equation", + "content": "\\Omega" + }, + { + "bbox": [ + 55, + 574, + 296, + 623 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 623, + 296, + 670 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 623, + 296, + 670 + ], + "spans": [ + { + "bbox": [ + 55, + 623, + 296, + 670 + ], + "type": "text", + "content": "One can see that, for any " + }, + { + "bbox": [ + 55, + 623, + 296, + 670 + ], + "type": "inline_equation", + "content": "k \\in \\mathcal{C}" + }, + { + "bbox": [ + 55, + 623, + 296, + 670 + ], + "type": "text", + "content": ", there are " + }, + { + "bbox": [ + 55, + 623, + 296, + 670 + ], + "type": "inline_equation", + "content": "2^{|C| - 1}" + }, + { + "bbox": [ + 55, + 623, + 296, + 670 + ], + "type": "text", + "content": " subsets that contain " + }, + { + "bbox": [ + 55, + 623, + 296, + 670 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 623, + 296, + 670 + ], + "type": "text", + "content": ". Each of these subsets would contain " + }, + { + "bbox": [ + 55, + 623, + 296, + 670 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 623, + 296, + 670 + ], + "type": "text", + "content": " and classes from the powerset " + }, + { + "bbox": [ + 55, + 623, + 296, + 670 + ], + "type": "inline_equation", + "content": "\\mathcal{P}(\\mathcal{C} \\backslash k)" + }, + { + "bbox": [ + 55, + 623, + 296, + 670 + ], + "type": "text", + "content": ". We can rewrite (5) as (see the supplementary materials for derivation details)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 672, + 296, + 698 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 672, + 296, + 698 + ], + "spans": [ + { + "bbox": [ + 57, + 672, + 296, + 698 + ], + "type": "interline_equation", + "content": "\\max _ {\\boldsymbol {U}} \\sum_ {k \\in \\mathcal {C}} \\sum_ {\\Omega \\in \\mathcal {P} (\\mathcal {C} \\backslash k)} \\left(\\mathcal {H} _ {k} \\left(\\boldsymbol {u} _ {k} + \\boldsymbol {U} \\boldsymbol {\\lambda} _ {\\Omega \\backslash k}\\right) - \\eta \\sum_ {q \\in \\{\\bar {\\Omega} \\backslash k \\}} \\mathcal {R} _ {q} \\left(\\boldsymbol {u} _ {k} + \\boldsymbol {U} \\boldsymbol {\\lambda} _ {\\Omega \\backslash k}\\right)\\right)", + "image_path": "92de891221d3aff6485a9e642885c3fa4e7e8cd6ca3a70f0c62509ec833e20ef.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 59, + 700, + 295, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 700, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 59, + 700, + 295, + 715 + ], + "type": "interline_equation", + "content": "\\mathrm {s . t .} \\hat {\\boldsymbol {U}} ^ {\\top} \\hat {\\boldsymbol {U}} = \\boldsymbol {I}, \\| \\boldsymbol {U} \\| _ {1, p} \\leq \\epsilon , \\tag {6}", + "image_path": "5d060f85a7503f356416b6a3358e196fd273a558266b20ac4cbbec8e668aa4ea.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 72, + 555, + 136 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 136 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 136 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 72, + 555, + 136 + ], + "type": "inline_equation", + "content": "\\eta = \\gamma /|\\Omega |" + }, + { + "bbox": [ + 313, + 72, + 555, + 136 + ], + "type": "text", + "content": " . The power set of " + }, + { + "bbox": [ + 313, + 72, + 555, + 136 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 313, + 72, + 555, + 136 + ], + "type": "text", + "content": " can be written as the union of all subsets of " + }, + { + "bbox": [ + 313, + 72, + 555, + 136 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 313, + 72, + 555, + 136 + ], + "type": "text", + "content": " .. " + }, + { + "bbox": [ + 313, + 72, + 555, + 136 + ], + "type": "inline_equation", + "content": "\\mathcal{P}(\\mathcal{C})\\triangleq \\bigcup_{i = 0}^{|C|}\\binom {\\mathcal{C}}{i}" + }, + { + "bbox": [ + 313, + 72, + 555, + 136 + ], + "type": "text", + "content": " where, each " + }, + { + "bbox": [ + 313, + 72, + 555, + 136 + ], + "type": "inline_equation", + "content": "\\binom{\\mathcal{C}}{i}" + }, + { + "bbox": [ + 313, + 72, + 555, + 136 + ], + "type": "text", + "content": " has all the subsets of " + }, + { + "bbox": [ + 313, + 72, + 555, + 136 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 313, + 72, + 555, + 136 + ], + "type": "text", + "content": " of size " + }, + { + "bbox": [ + 313, + 72, + 555, + 136 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 72, + 555, + 136 + ], + "type": "text", + "content": " . In our case, " + }, + { + "bbox": [ + 313, + 72, + 555, + 136 + ], + "type": "inline_equation", + "content": "S\\in \\binom{\\mathcal{C}}{i}" + }, + { + "bbox": [ + 313, + 72, + 555, + 136 + ], + "type": "text", + "content": " would represent the indices of the columns of " + }, + { + "bbox": [ + 313, + 72, + 555, + 136 + ], + "type": "inline_equation", + "content": "\\pmb{U}" + }, + { + "bbox": [ + 313, + 72, + 555, + 136 + ], + "type": "text", + "content": " . So, we can write the summation across the power set as:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 315, + 138, + 558, + 173 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 138, + 558, + 173 + ], + "spans": [ + { + "bbox": [ + 315, + 138, + 558, + 173 + ], + "type": "interline_equation", + "content": "\\max_{\\boldsymbol{U}}\\sum_{k\\in \\mathcal{C}}\\sum_{t = 0}^{\\left|\\mathcal{C}\\right| - 1}\\sum_{S\\in \\binom {\\mathcal{C}\\setminus k}{t}}\\Bigl(\\mathcal{H}_{k}(\\boldsymbol{u}_{k} + \\boldsymbol {U}\\boldsymbol{\\lambda}_{S}) - \\eta \\sum_{q\\in \\{\\bar{S}\\setminus k\\}}\\mathcal{R}_{q}(\\boldsymbol{u}_{k} + \\boldsymbol {U}\\boldsymbol{\\lambda}_{S})\\Bigr),", + "image_path": "fced1d2ce19160d3f9b952a391bcf95ba4707aeeb5384f5097f1145e6d911577.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 318, + 175, + 553, + 191 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 175, + 553, + 191 + ], + "spans": [ + { + "bbox": [ + 318, + 175, + 553, + 191 + ], + "type": "interline_equation", + "content": "s. t. \\hat {\\boldsymbol {U}} ^ {\\top} \\hat {\\boldsymbol {U}} = \\boldsymbol {I}, \\| \\boldsymbol {U} \\| _ {1, p} \\leq \\epsilon , \\tag {7}", + "image_path": "6ca63efa9951412e4f655a1084cb46532931dd9ed367a55a143977027490e98d.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 195, + 555, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 195, + 555, + 293 + ], + "spans": [ + { + "bbox": [ + 313, + 195, + 555, + 293 + ], + "type": "text", + "content": "In (7), for a fixed " + }, + { + "bbox": [ + 313, + 195, + 555, + 293 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 195, + 555, + 293 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 195, + 555, + 293 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 195, + 555, + 293 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 195, + 555, + 293 + ], + "type": "inline_equation", + "content": "\\lambda_{S}" + }, + { + "bbox": [ + 313, + 195, + 555, + 293 + ], + "type": "text", + "content": " has exactly " + }, + { + "bbox": [ + 313, + 195, + 555, + 293 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 195, + 555, + 293 + ], + "type": "text", + "content": " non-zero elements and " + }, + { + "bbox": [ + 313, + 195, + 555, + 293 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 313, + 195, + 555, + 293 + ], + "type": "text", + "content": " iterates through all possible combinations of " + }, + { + "bbox": [ + 313, + 195, + 555, + 293 + ], + "type": "inline_equation", + "content": "\\{C \\backslash k\\}" + }, + { + "bbox": [ + 313, + 195, + 555, + 293 + ], + "type": "text", + "content": " of size " + }, + { + "bbox": [ + 313, + 195, + 555, + 293 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 195, + 555, + 293 + ], + "type": "text", + "content": ". Therefore, the sum across all " + }, + { + "bbox": [ + 313, + 195, + 555, + 293 + ], + "type": "inline_equation", + "content": "\\lambda_{S}" + }, + { + "bbox": [ + 313, + 195, + 555, + 293 + ], + "type": "text", + "content": " (for " + }, + { + "bbox": [ + 313, + 195, + 555, + 293 + ], + "type": "inline_equation", + "content": "S \\in \\binom{\\mathcal{C} \\backslash k}{t}" + }, + { + "bbox": [ + 313, + 195, + 555, + 293 + ], + "type": "text", + "content": ") would be equal to the sum across all binary vectors " + }, + { + "bbox": [ + 313, + 195, + 555, + 293 + ], + "type": "inline_equation", + "content": "\\lambda \\in \\{0,1\\}^{|\\mathcal{C}|}" + }, + { + "bbox": [ + 313, + 195, + 555, + 293 + ], + "type": "text", + "content": " such that exactly " + }, + { + "bbox": [ + 313, + 195, + 555, + 293 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 195, + 555, + 293 + ], + "type": "text", + "content": " elements are nonzero (i.e., " + }, + { + "bbox": [ + 313, + 195, + 555, + 293 + ], + "type": "inline_equation", + "content": "\\| \\lambda \\|_0 = t" + }, + { + "bbox": [ + 313, + 195, + 555, + 293 + ], + "type": "text", + "content": ") and " + }, + { + "bbox": [ + 313, + 195, + 555, + 293 + ], + "type": "inline_equation", + "content": "\\lambda_{k} = 0" + }, + { + "bbox": [ + 313, + 195, + 555, + 293 + ], + "type": "text", + "content": ". Where, " + }, + { + "bbox": [ + 313, + 195, + 555, + 293 + ], + "type": "inline_equation", + "content": "\\| \\cdot \\|_0" + }, + { + "bbox": [ + 313, + 195, + 555, + 293 + ], + "type": "text", + "content": " denotes the " + }, + { + "bbox": [ + 313, + 195, + 555, + 293 + ], + "type": "inline_equation", + "content": "\\ell_0" + }, + { + "bbox": [ + 313, + 195, + 555, + 293 + ], + "type": "text", + "content": "-norm, which counts the number of nonzero elements of the given input. So we can write (7) as" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 315, + 295, + 553, + 335 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 295, + 553, + 335 + ], + "spans": [ + { + "bbox": [ + 315, + 295, + 553, + 335 + ], + "type": "interline_equation", + "content": "\\max _ {\\boldsymbol {U}} \\sum_ {k \\in \\mathcal {C}} \\sum_ {t = 0} ^ {| \\mathcal {C} | - 1} \\sum_ {\\substack {\\boldsymbol {\\lambda} \\in \\{0, 1 \\} ^ {| \\mathcal {C} |} \\\\ \\| \\boldsymbol {\\lambda} \\| _ {0} = t, \\boldsymbol {\\lambda} _ {k} = 0}} \\left[ \\mathcal {H} _ {k} (\\boldsymbol {u} _ {k} + \\boldsymbol {U} \\boldsymbol {\\lambda}) \\right. \\tag{8}", + "image_path": "0937043b2b0fd8198926eda78e930a7d49fae8d3bb423686f6087b50522fc5fc.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 337, + 555, + 363 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 337, + 555, + 363 + ], + "spans": [ + { + "bbox": [ + 317, + 337, + 555, + 363 + ], + "type": "interline_equation", + "content": "\\left. - \\eta \\sum_ {q \\in \\{\\mathcal {C} \\backslash k \\}} \\delta (\\boldsymbol {\\lambda} _ {q} = 0) \\mathcal {R} _ {q} (\\boldsymbol {u} _ {k} + \\boldsymbol {U} \\boldsymbol {\\lambda}) \\right] \\text {s . t .} \\hat {\\boldsymbol {U}} ^ {\\top} \\hat {\\boldsymbol {U}} = \\boldsymbol {I}, \\| \\boldsymbol {U} \\| _ {1, p} \\leq \\epsilon ,", + "image_path": "5fe364c082e8e886defe656faa1f275876ff59cd66ec15325cf84fdee808069e.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 365, + 554, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 365, + 554, + 413 + ], + "spans": [ + { + "bbox": [ + 313, + 365, + 554, + 413 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 365, + 554, + 413 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 313, + 365, + 554, + 413 + ], + "type": "text", + "content": " is the indicator function which is 1 if the input condition is True and 0 otherwise. As " + }, + { + "bbox": [ + 313, + 365, + 554, + 413 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 365, + 554, + 413 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 313, + 365, + 554, + 413 + ], + "type": "inline_equation", + "content": "|\\mathcal{C}|" + }, + { + "bbox": [ + 313, + 365, + 554, + 413 + ], + "type": "text", + "content": " dimensional, the second loss term in (8) iterates through each index of " + }, + { + "bbox": [ + 313, + 365, + 554, + 413 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 365, + 554, + 413 + ], + "type": "text", + "content": " i.e., " + }, + { + "bbox": [ + 313, + 365, + 554, + 413 + ], + "type": "inline_equation", + "content": "(q\\in \\{\\mathcal{C}\\backslash k\\})" + }, + { + "bbox": [ + 313, + 365, + 554, + 413 + ], + "type": "text", + "content": " and computes the loss of label " + }, + { + "bbox": [ + 313, + 365, + 554, + 413 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 313, + 365, + 554, + 413 + ], + "type": "text", + "content": " if " + }, + { + "bbox": [ + 313, + 365, + 554, + 413 + ], + "type": "inline_equation", + "content": "\\lambda_q = 0" + }, + { + "bbox": [ + 313, + 365, + 554, + 413 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 413, + 554, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 413, + 554, + 437 + ], + "spans": [ + { + "bbox": [ + 313, + 413, + 554, + 437 + ], + "type": "text", + "content": "We investigate two ways to approximately solve (8). In the first approach, we solve the lower bound of (8) as" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 441, + 553, + 470 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 441, + 553, + 470 + ], + "spans": [ + { + "bbox": [ + 316, + 441, + 553, + 470 + ], + "type": "interline_equation", + "content": "\\operatorname {CMLU} _ {\\alpha}: \\max _ {\\boldsymbol {U}} \\sum_ {k \\in \\mathcal {C}} \\sum_ {t \\in \\mathcal {T}} 2 ^ {| \\mathcal {C} |} \\min _ {\\substack {\\boldsymbol {\\lambda} \\in \\{0,1 \\} ^ {| C |} \\\\ \\| \\boldsymbol {\\lambda} \\| _ {0} = t, \\boldsymbol {\\lambda} _ {k} = 0}} \\left[ \\mathcal {H} _ {k} (\\boldsymbol {u} _ {k} + \\boldsymbol {U} \\boldsymbol {\\lambda}) \\right. \\tag{9}", + "image_path": "1447f8dde94b412edd11959f4fe01a43501c40c21fac15649e4ff7749d2a0882.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 318, + 472, + 552, + 498 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 472, + 552, + 498 + ], + "spans": [ + { + "bbox": [ + 318, + 472, + 552, + 498 + ], + "type": "interline_equation", + "content": "\\left. - \\eta \\sum_ {q \\in \\{\\mathcal {C} \\backslash k \\}} \\delta (\\boldsymbol {\\lambda} _ {q} = 0) \\mathcal {R} _ {q} (\\boldsymbol {u} _ {k} + \\boldsymbol {U} \\boldsymbol {\\lambda}) \\right] \\text {s . t .} \\hat {\\boldsymbol {U}} ^ {\\top} \\hat {\\boldsymbol {U}} = \\boldsymbol {I}, \\| \\boldsymbol {U} \\| _ {1, p} \\leq \\epsilon", + "image_path": "5be4eb25570dbec108aa98ae7ad599ced085621ce1b2957c3265c2436fc5ed4e.jpg" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 503, + 555, + 610 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 503, + 555, + 610 + ], + "spans": [ + { + "bbox": [ + 313, + 503, + 555, + 610 + ], + "type": "text", + "content": "Where, " + }, + { + "bbox": [ + 313, + 503, + 555, + 610 + ], + "type": "inline_equation", + "content": "\\mathcal{T} = \\{0,1,\\dots ,|\\mathcal{C}| - 1\\}" + }, + { + "bbox": [ + 313, + 503, + 555, + 610 + ], + "type": "text", + "content": ". (9) requires maximizing the minimal loss. To solve the inner minimization problem, for a fixed " + }, + { + "bbox": [ + 313, + 503, + 555, + 610 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 503, + 555, + 610 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 503, + 555, + 610 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 503, + 555, + 610 + ], + "type": "text", + "content": ", we randomly initialize " + }, + { + "bbox": [ + 313, + 503, + 555, + 610 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 503, + 555, + 610 + ], + "type": "text", + "content": " and iteratively update it for a fixed number of iterations. We then project " + }, + { + "bbox": [ + 313, + 503, + 555, + 610 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 503, + 555, + 610 + ], + "type": "text", + "content": " on the feasible set by keeping the largest " + }, + { + "bbox": [ + 313, + 503, + 555, + 610 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 503, + 555, + 610 + ], + "type": "text", + "content": " values as 1 and setting the rest to 0. We then keep the " + }, + { + "bbox": [ + 313, + 503, + 555, + 610 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 503, + 555, + 610 + ], + "type": "text", + "content": " fixed and solve the outer maximization in (9) w.r.t. " + }, + { + "bbox": [ + 313, + 503, + 555, + 610 + ], + "type": "inline_equation", + "content": "\\pmb{U}" + }, + { + "bbox": [ + 313, + 503, + 555, + 610 + ], + "type": "text", + "content": ". In practice, instead of iterating through all " + }, + { + "bbox": [ + 313, + 503, + 555, + 610 + ], + "type": "inline_equation", + "content": "t\\in \\mathcal{T}" + }, + { + "bbox": [ + 313, + 503, + 555, + 610 + ], + "type": "text", + "content": ", for each " + }, + { + "bbox": [ + 313, + 503, + 555, + 610 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 503, + 555, + 610 + ], + "type": "text", + "content": ", we randomly choose " + }, + { + "bbox": [ + 313, + 503, + 555, + 610 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 503, + 555, + 610 + ], + "type": "text", + "content": ". The steps are shown in Algorithm 1." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 611, + 554, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 611, + 554, + 647 + ], + "spans": [ + { + "bbox": [ + 313, + 611, + 554, + 647 + ], + "type": "text", + "content": "In the second approach, we consider the continuous relaxation of " + }, + { + "bbox": [ + 313, + 611, + 554, + 647 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 611, + 554, + 647 + ], + "type": "text", + "content": " and approximately solve it by uniformly randomly sampling " + }, + { + "bbox": [ + 313, + 611, + 554, + 647 + ], + "type": "inline_equation", + "content": "\\lambda \\in [0,1]^{|\\mathcal{C}|}" + }, + { + "bbox": [ + 313, + 611, + 554, + 647 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 315, + 651, + 553, + 686 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 651, + 553, + 686 + ], + "spans": [ + { + "bbox": [ + 315, + 651, + 553, + 686 + ], + "type": "interline_equation", + "content": "\\operatorname {C M L U} _ {\\beta}: \\max _ {\\boldsymbol {U}} \\sum_ {k \\in \\mathcal {C}} \\sum_ {t \\in \\mathcal {T}} \\sum_ {\\substack {\\boldsymbol {\\lambda} \\in [ 0, 1 ] ^ {| \\mathcal {C} |} \\\\ \\| \\boldsymbol {\\lambda} \\| _ {0} = t, \\boldsymbol {\\lambda} _ {k} = 0}} \\left[ \\mathcal {H} _ {k} (\\boldsymbol {u} _ {k} + \\boldsymbol {U} \\boldsymbol {\\lambda}) \\right. \\tag{10}", + "image_path": "12b866658cceed6b8a0406ce06e04b2e983483528a77af3cdea1c59b0a7f3e68.jpg" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 688, + 553, + 714 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 688, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 317, + 688, + 553, + 714 + ], + "type": "interline_equation", + "content": "\\left. - \\eta \\sum_ {q \\in \\{\\mathcal {C} \\backslash k \\}} \\delta (\\boldsymbol {\\lambda} _ {q} = 0) \\mathcal {R} _ {q} (\\boldsymbol {u} _ {k} + \\boldsymbol {U} \\boldsymbol {\\lambda}) \\right] \\text {s . t .} \\hat {\\boldsymbol {U}} ^ {\\top} \\hat {\\boldsymbol {U}} = \\boldsymbol {I}, \\| \\boldsymbol {U} \\| _ {1, p} \\leq \\epsilon .", + "image_path": "33ef8111bdd6e8296ff595af5bcecb12cadf52fc80dc995aacad92aa9c2a3541.jpg" + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20583" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "content": "Eq. (9) and (10) can be explained intuitively. The first loss term " + }, + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_k" + }, + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "content": " optimizes each label-wise universal perturbation " + }, + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "inline_equation", + "content": "\\pmb{u}_k" + }, + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "content": " such that it is robust to the noise generated by composing other universal perturbations. Specifically, it aims to maximize the loss of class " + }, + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "content": ", regardless of which other class universals have been combined. The second term ensures that " + }, + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "inline_equation", + "content": "\\pmb{u}_k" + }, + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "content": " does not change the prediction of classes whose universal vectors have not been used in the composition." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 168, + 295, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 168, + 295, + 300 + ], + "spans": [ + { + "bbox": [ + 55, + 168, + 295, + 300 + ], + "type": "text", + "content": "Remark: A key aspect of Eq. (9) and (10) is that " + }, + { + "bbox": [ + 55, + 168, + 295, + 300 + ], + "type": "inline_equation", + "content": "\\pmb{u}_k + \\pmb{U}\\lambda" + }, + { + "bbox": [ + 55, + 168, + 295, + 300 + ], + "type": "text", + "content": " represents points sampled from an affine convex cone. Specifically, these formulations aim to maximize a function over points sampled from " + }, + { + "bbox": [ + 55, + 168, + 295, + 300 + ], + "type": "inline_equation", + "content": "|C|" + }, + { + "bbox": [ + 55, + 168, + 295, + 300 + ], + "type": "text", + "content": " affine convex cones (generated by the columns of " + }, + { + "bbox": [ + 55, + 168, + 295, + 300 + ], + "type": "inline_equation", + "content": "\\pmb{U}" + }, + { + "bbox": [ + 55, + 168, + 295, + 300 + ], + "type": "text", + "content": "). Consequently, we can interpret our assumption in (4) as aiming to learn the universal adversarial regions of each class as affine cones, where the vector " + }, + { + "bbox": [ + 55, + 168, + 295, + 300 + ], + "type": "inline_equation", + "content": "\\pmb{u}_{\\Omega}" + }, + { + "bbox": [ + 55, + 168, + 295, + 300 + ], + "type": "text", + "content": ", as outlined in (4), lies at the intersection of the affine cones of the classes in " + }, + { + "bbox": [ + 55, + 168, + 295, + 300 + ], + "type": "inline_equation", + "content": "\\Omega" + }, + { + "bbox": [ + 55, + 168, + 295, + 300 + ], + "type": "text", + "content": ". This is illustrated in Fig. 2 (right), where " + }, + { + "bbox": [ + 55, + 168, + 295, + 300 + ], + "type": "inline_equation", + "content": "\\pmb{C}_k" + }, + { + "bbox": [ + 55, + 168, + 295, + 300 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 168, + 295, + 300 + ], + "type": "inline_equation", + "content": "\\pmb{C}_i" + }, + { + "bbox": [ + 55, + 168, + 295, + 300 + ], + "type": "text", + "content": " represent convex cones for classes " + }, + { + "bbox": [ + 55, + 168, + 295, + 300 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 168, + 295, + 300 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 168, + 295, + 300 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 168, + 295, + 300 + ], + "type": "text", + "content": " respectively, and " + }, + { + "bbox": [ + 55, + 168, + 295, + 300 + ], + "type": "inline_equation", + "content": "\\pmb{C}_{ik}" + }, + { + "bbox": [ + 55, + 168, + 295, + 300 + ], + "type": "text", + "content": " denotes their intersection." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 301, + 296, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 301, + 296, + 485 + ], + "spans": [ + { + "bbox": [ + 55, + 301, + 296, + 485 + ], + "type": "text", + "content": "For a comparison among Oracle, " + }, + { + "bbox": [ + 55, + 301, + 296, + 485 + ], + "type": "inline_equation", + "content": "\\mathrm{CMLU}_{\\alpha}" + }, + { + "bbox": [ + 55, + 301, + 296, + 485 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 301, + 296, + 485 + ], + "type": "inline_equation", + "content": "\\mathrm{CMLU}_{\\beta}" + }, + { + "bbox": [ + 55, + 301, + 296, + 485 + ], + "type": "text", + "content": ", Algorithm 1 shows the steps of a single training epoch of Oracle of target size " + }, + { + "bbox": [ + 55, + 301, + 296, + 485 + ], + "type": "inline_equation", + "content": "|\\Omega| = 1" + }, + { + "bbox": [ + 55, + 301, + 296, + 485 + ], + "type": "text", + "content": " and the corresponding steps for our methods. Note that in general, Oracle has " + }, + { + "bbox": [ + 55, + 301, + 296, + 485 + ], + "type": "inline_equation", + "content": "\\mathcal{O}(\\binom{|C|}{|\\Omega|})" + }, + { + "bbox": [ + 55, + 301, + 296, + 485 + ], + "type": "text", + "content": " training time complexity, which increases with the size of " + }, + { + "bbox": [ + 55, + 301, + 296, + 485 + ], + "type": "inline_equation", + "content": "\\Omega" + }, + { + "bbox": [ + 55, + 301, + 296, + 485 + ], + "type": "text", + "content": ". For instance, for " + }, + { + "bbox": [ + 55, + 301, + 296, + 485 + ], + "type": "inline_equation", + "content": "|\\Omega| = 2" + }, + { + "bbox": [ + 55, + 301, + 296, + 485 + ], + "type": "text", + "content": ", the complexity of the Oracle's inner loop becomes " + }, + { + "bbox": [ + 55, + 301, + 296, + 485 + ], + "type": "inline_equation", + "content": "\\mathcal{O}(\\binom{|C|}{2})" + }, + { + "bbox": [ + 55, + 301, + 296, + 485 + ], + "type": "text", + "content": ", as it needs to handle all possible combinations of classes of a fixed size " + }, + { + "bbox": [ + 55, + 301, + 296, + 485 + ], + "type": "inline_equation", + "content": "|\\Omega|" + }, + { + "bbox": [ + 55, + 301, + 296, + 485 + ], + "type": "text", + "content": ". However, " + }, + { + "bbox": [ + 55, + 301, + 296, + 485 + ], + "type": "inline_equation", + "content": "\\mathrm{CMLU}_{\\alpha}" + }, + { + "bbox": [ + 55, + 301, + 296, + 485 + ], + "type": "text", + "content": " has " + }, + { + "bbox": [ + 55, + 301, + 296, + 485 + ], + "type": "inline_equation", + "content": "\\mathcal{O}(N_{\\alpha}|\\mathcal{C}|)" + }, + { + "bbox": [ + 55, + 301, + 296, + 485 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 301, + 296, + 485 + ], + "type": "inline_equation", + "content": "\\mathrm{CMLU}_{\\beta}" + }, + { + "bbox": [ + 55, + 301, + 296, + 485 + ], + "type": "text", + "content": " has " + }, + { + "bbox": [ + 55, + 301, + 296, + 485 + ], + "type": "inline_equation", + "content": "\\mathcal{O}(|C|)" + }, + { + "bbox": [ + 55, + 301, + 296, + 485 + ], + "type": "text", + "content": " complexity and is independent of " + }, + { + "bbox": [ + 55, + 301, + 296, + 485 + ], + "type": "inline_equation", + "content": "|\\Omega|" + }, + { + "bbox": [ + 55, + 301, + 296, + 485 + ], + "type": "text", + "content": ". Note that the complexity of our method for learning composable universal perturbations for all possible target set sizes is equivalent to the complexity of Oracle trained for " + }, + { + "bbox": [ + 55, + 301, + 296, + 485 + ], + "type": "inline_equation", + "content": "|\\Omega| = 1" + }, + { + "bbox": [ + 55, + 301, + 296, + 485 + ], + "type": "text", + "content": " (linear in the number of classes). Oracle requires iterating over the dataset for each combination of target classes " + }, + { + "bbox": [ + 55, + 301, + 296, + 485 + ], + "type": "inline_equation", + "content": "\\Omega" + }, + { + "bbox": [ + 55, + 301, + 296, + 485 + ], + "type": "text", + "content": ", leading to exponential complexity." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 495, + 127, + 507 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 495, + 127, + 507 + ], + "spans": [ + { + "bbox": [ + 55, + 495, + 127, + 507 + ], + "type": "text", + "content": "4. Evaluation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 516, + 171, + 529 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 516, + 171, + 529 + ], + "spans": [ + { + "bbox": [ + 55, + 516, + 171, + 529 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 533, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 533, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 533, + 295, + 713 + ], + "type": "text", + "content": "Datasets We perform experiments using NUS-WIDE [16], MS-COCO [56], and OpenImages [44] datasets to assess the effectiveness of our proposed method. We compare it with oracle and baseline methods. For NUS-WIDE and MS-COCO, we test target set sizes ranging from 1 to 5 and for OpenImages, we experiment with larger target set sizes, going up to 20. Due to the computational expense of learning oracle model, we choose a subset of labels from each dataset (25 classes from NUS-WIDE, 40 classes from MS-COCO, and 100 classes from OpenImages). Even with the small subset of classes, there can be a large number of possible combinations e.g., there can be more than 53k combinations of " + }, + { + "bbox": [ + 55, + 533, + 295, + 713 + ], + "type": "inline_equation", + "content": "|\\Omega| = 5" + }, + { + "bbox": [ + 55, + 533, + 295, + 713 + ], + "type": "text", + "content": " using 25 classes from NUS-WIDE. Therefore, we perform experiments on a subset of those possible combinations for each target set size " + }, + { + "bbox": [ + 55, + 533, + 295, + 713 + ], + "type": "inline_equation", + "content": "|\\Omega|" + }, + { + "bbox": [ + 55, + 533, + 295, + 713 + ], + "type": "text", + "content": ". For all of our" + } + ] + } + ], + "index": 5 + }, + { + "type": "code", + "bbox": [ + 316, + 88, + 553, + 457 + ], + "blocks": [ + { + "bbox": [ + 318, + 75, + 527, + 87 + ], + "lines": [ + { + "bbox": [ + 318, + 75, + 527, + 87 + ], + "spans": [ + { + "bbox": [ + 318, + 75, + 527, + 87 + ], + "type": "text", + "content": "Algorithm 1: Algorithms of Oracle, " + }, + { + "bbox": [ + 318, + 75, + 527, + 87 + ], + "type": "inline_equation", + "content": "\\mathrm{CMLU}_{\\alpha}" + }, + { + "bbox": [ + 318, + 75, + 527, + 87 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 318, + 75, + 527, + 87 + ], + "type": "inline_equation", + "content": "\\mathrm{CMLU}_{\\beta}" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 316, + 88, + 553, + 457 + ], + "lines": [ + { + "bbox": [ + 316, + 88, + 553, + 457 + ], + "spans": [ + { + "bbox": [ + 316, + 88, + 553, + 457 + ], + "type": "text", + "content": "Input: " + }, + { + "bbox": [ + 316, + 88, + 553, + 457 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 316, + 88, + 553, + 457 + ], + "type": "text", + "content": " : Training Data, " + }, + { + "bbox": [ + 316, + 88, + 553, + 457 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 316, + 88, + 553, + 457 + ], + "type": "text", + "content": " : Classifier, " + }, + { + "bbox": [ + 316, + 88, + 553, + 457 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 316, + 88, + 553, + 457 + ], + "type": "text", + "content": " : Norm budget, " + }, + { + "bbox": [ + 316, + 88, + 553, + 457 + ], + "type": "inline_equation", + "content": "(\\xi ,\\alpha)" + }, + { + "bbox": [ + 316, + 88, + 553, + 457 + ], + "type": "text", + "content": " : Step sizes Output: " + }, + { + "bbox": [ + 316, + 88, + 553, + 457 + ], + "type": "inline_equation", + "content": "U" + }, + { + "bbox": [ + 316, + 88, + 553, + 457 + ], + "type": "text", + "content": " : Learnt Universal Perturbations 1 Randomly initialize " + }, + { + "bbox": [ + 316, + 88, + 553, + 457 + ], + "type": "inline_equation", + "content": "U" + }, + { + "bbox": [ + 316, + 88, + 553, + 457 + ], + "type": "text", + "content": " /\\* " + }, + { + "bbox": [ + 316, + 88, + 553, + 457 + ], + "type": "inline_equation", + "content": "\\mathcal{B} = \\{\\mathcal{B}^{(k)}\\}_{k = 1}^{|C|},\\mathcal{B}^{(k)}" + }, + { + "bbox": [ + 316, + 88, + 553, + 457 + ], + "type": "text", + "content": " images of class " + }, + { + "bbox": [ + 316, + 88, + 553, + 457 + ], + "type": "inline_equation", + "content": "k\\ast /" + }, + { + "bbox": [ + 316, + 88, + 553, + 457 + ], + "type": "text", + "content": " \n2 Oracle: \n3 for Batch " + }, + { + "bbox": [ + 316, + 88, + 553, + 457 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 316, + 88, + 553, + 457 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 316, + 88, + 553, + 457 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 316, + 88, + 553, + 457 + ], + "type": "text", + "content": " do /* One iteration through the dataset \\*/ \n4 for " + }, + { + "bbox": [ + 316, + 88, + 553, + 457 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 316, + 88, + 553, + 457 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 316, + 88, + 553, + 457 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 316, + 88, + 553, + 457 + ], + "type": "text", + "content": " do /\\* O(|C|) \\*/ \n5 " + }, + { + "bbox": [ + 316, + 88, + 553, + 457 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_k" + }, + { + "bbox": [ + 316, + 88, + 553, + 457 + ], + "type": "text", + "content": " : Compute loss in (1) using " + }, + { + "bbox": [ + 316, + 88, + 553, + 457 + ], + "type": "inline_equation", + "content": "\\pmb {u}_k" + }, + { + "bbox": [ + 316, + 88, + 553, + 457 + ], + "type": "text", + "content": " \n6 Compute the gradient: " + }, + { + "bbox": [ + 316, + 88, + 553, + 457 + ], + "type": "inline_equation", + "content": "\\pmb {g}\\gets \\nabla_{\\pmb{U}_k}\\pmb {l}_k" + }, + { + "bbox": [ + 316, + 88, + 553, + 457 + ], + "type": "text", + "content": " \n7 Update " + }, + { + "bbox": [ + 316, + 88, + 553, + 457 + ], + "type": "inline_equation", + "content": "U_{k}\\leftarrow \\Pi_{\\epsilon}\\left(U_{k} + \\xi *Sgn(\\pmb {g})\\right)" + }, + { + "bbox": [ + 316, + 88, + 553, + 457 + ], + "type": "text", + "content": " \n8 for Batch " + }, + { + "bbox": [ + 316, + 88, + 553, + 457 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 316, + 88, + 553, + 457 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 316, + 88, + 553, + 457 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 316, + 88, + 553, + 457 + ], + "type": "text", + "content": " do \n9 Set " + }, + { + "bbox": [ + 316, + 88, + 553, + 457 + ], + "type": "inline_equation", + "content": "\\pmb{g}" + }, + { + "bbox": [ + 316, + 88, + 553, + 457 + ], + "type": "text", + "content": " to be all zeros of the same size as " + }, + { + "bbox": [ + 316, + 88, + 553, + 457 + ], + "type": "inline_equation", + "content": "U" + }, + { + "bbox": [ + 316, + 88, + 553, + 457 + ], + "type": "text", + "content": " \n10 for " + }, + { + "bbox": [ + 316, + 88, + 553, + 457 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 316, + 88, + 553, + 457 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 316, + 88, + 553, + 457 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 316, + 88, + 553, + 457 + ], + "type": "text", + "content": " do /\\* O(|C|) \\*/ \n11 Randomly sample " + }, + { + "bbox": [ + 316, + 88, + 553, + 457 + ], + "type": "inline_equation", + "content": "t\\sim T" + }, + { + "bbox": [ + 316, + 88, + 553, + 457 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 316, + 88, + 553, + 457 + ], + "type": "inline_equation", + "content": "\\lambda \\in [0,1]^{\\lfloor C\\rfloor}" + }, + { + "bbox": [ + 316, + 88, + 553, + 457 + ], + "type": "text", + "content": " . \n12 " + }, + { + "bbox": [ + 316, + 88, + 553, + 457 + ], + "type": "inline_equation", + "content": "\\mathrm{CMLU}_{\\alpha}" + }, + { + "bbox": [ + 316, + 88, + 553, + 457 + ], + "type": "text", + "content": " ： /\\*O(Nα）\\*/ \n13 for " + }, + { + "bbox": [ + 316, + 88, + 553, + 457 + ], + "type": "inline_equation", + "content": "N_{\\alpha}" + }, + { + "bbox": [ + 316, + 88, + 553, + 457 + ], + "type": "text", + "content": " num of iterations do \n14 Set " + }, + { + "bbox": [ + 316, + 88, + 553, + 457 + ], + "type": "inline_equation", + "content": "\\pmb {v}_k = \\pmb {u}_k + U\\pmb{\\lambda}" + }, + { + "bbox": [ + 316, + 88, + 553, + 457 + ], + "type": "text", + "content": " \n15 " + }, + { + "bbox": [ + 316, + 88, + 553, + 457 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_k" + }, + { + "bbox": [ + 316, + 88, + 553, + 457 + ], + "type": "text", + "content": " : Compute loss in (9) using " + }, + { + "bbox": [ + 316, + 88, + 553, + 457 + ], + "type": "inline_equation", + "content": "\\pmb {v}_k" + }, + { + "bbox": [ + 316, + 88, + 553, + 457 + ], + "type": "text", + "content": " \n16 Update: " + }, + { + "bbox": [ + 316, + 88, + 553, + 457 + ], + "type": "inline_equation", + "content": "\\lambda = \\lambda -\\alpha \\nabla_{\\lambda}\\mathcal{L}_k" + }, + { + "bbox": [ + 316, + 88, + 553, + 457 + ], + "type": "text", + "content": " \n17 Project " + }, + { + "bbox": [ + 316, + 88, + 553, + 457 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 316, + 88, + 553, + 457 + ], + "type": "text", + "content": " on the feasible set " + }, + { + "bbox": [ + 316, + 88, + 553, + 457 + ], + "type": "inline_equation", + "content": "\\| \\lambda \\| _0 = t" + }, + { + "bbox": [ + 316, + 88, + 553, + 457 + ], + "type": "text", + "content": " \n18 Set " + }, + { + "bbox": [ + 316, + 88, + 553, + 457 + ], + "type": "inline_equation", + "content": "\\lambda_k = 0" + }, + { + "bbox": [ + 316, + 88, + 553, + 457 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 316, + 88, + 553, + 457 + ], + "type": "inline_equation", + "content": "\\pmb {v}_k = \\pmb {u}_k + \\pmb {U}\\pmb{\\lambda}" + }, + { + "bbox": [ + 316, + 88, + 553, + 457 + ], + "type": "text", + "content": " \n19 " + }, + { + "bbox": [ + 316, + 88, + 553, + 457 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_k" + }, + { + "bbox": [ + 316, + 88, + 553, + 457 + ], + "type": "text", + "content": " : Compute loss in (9) using " + }, + { + "bbox": [ + 316, + 88, + 553, + 457 + ], + "type": "inline_equation", + "content": "\\pmb {v}_k" + }, + { + "bbox": [ + 316, + 88, + 553, + 457 + ], + "type": "text", + "content": " \n20 CMLUβ:/*O(1)\\*/ \n21 Randomly set " + }, + { + "bbox": [ + 316, + 88, + 553, + 457 + ], + "type": "inline_equation", + "content": "|\\mathcal{C}| - t" + }, + { + "bbox": [ + 316, + 88, + 553, + 457 + ], + "type": "text", + "content": " values in " + }, + { + "bbox": [ + 316, + 88, + 553, + 457 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 316, + 88, + 553, + 457 + ], + "type": "text", + "content": " to 0. \n22 Set " + }, + { + "bbox": [ + 316, + 88, + 553, + 457 + ], + "type": "inline_equation", + "content": "\\lambda_k = 0" + }, + { + "bbox": [ + 316, + 88, + 553, + 457 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 316, + 88, + 553, + 457 + ], + "type": "inline_equation", + "content": "\\pmb {v}_k = \\pmb {u}_k + U\\pmb{\\lambda}" + }, + { + "bbox": [ + 316, + 88, + 553, + 457 + ], + "type": "text", + "content": " \n23 " + }, + { + "bbox": [ + 316, + 88, + 553, + 457 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_k" + }, + { + "bbox": [ + 316, + 88, + 553, + 457 + ], + "type": "text", + "content": " :Compute loss in (10) using " + }, + { + "bbox": [ + 316, + 88, + 553, + 457 + ], + "type": "inline_equation", + "content": "\\pmb {v}_k" + }, + { + "bbox": [ + 316, + 88, + 553, + 457 + ], + "type": "text", + "content": " \n24 Accumulate: " + }, + { + "bbox": [ + 316, + 88, + 553, + 457 + ], + "type": "inline_equation", + "content": "\\pmb {g}\\gets \\pmb {g} + \\nabla_{\\pmb{U}_k}\\mathcal{L}_k" + }, + { + "bbox": [ + 316, + 88, + 553, + 457 + ], + "type": "text", + "content": " \n25 Update " + }, + { + "bbox": [ + 316, + 88, + 553, + 457 + ], + "type": "inline_equation", + "content": "U\\gets \\Pi_{\\epsilon}(U + \\xi *Sgn(g))" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "code_body" + } + ], + "index": 7, + "sub_type": "algorithm" + }, + { + "bbox": [ + 313, + 483, + 553, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 483, + 553, + 554 + ], + "spans": [ + { + "bbox": [ + 313, + 483, + 553, + 554 + ], + "type": "text", + "content": "experiments, we learn universal vectors with maximum infinity norm " + }, + { + "bbox": [ + 313, + 483, + 553, + 554 + ], + "type": "inline_equation", + "content": "\\epsilon = 0.05" + }, + { + "bbox": [ + 313, + 483, + 553, + 554 + ], + "type": "text", + "content": " (for image values range of [0,1]) and using SGD with fix step size " + }, + { + "bbox": [ + 313, + 483, + 553, + 554 + ], + "type": "inline_equation", + "content": "\\xi = \\epsilon /25" + }, + { + "bbox": [ + 313, + 483, + 553, + 554 + ], + "type": "text", + "content": ". We set " + }, + { + "bbox": [ + 313, + 483, + 553, + 554 + ], + "type": "inline_equation", + "content": "\\gamma = \\kappa /|\\bar{\\Omega}|" + }, + { + "bbox": [ + 313, + 483, + 553, + 554 + ], + "type": "text", + "content": " for training oracle and baselines, and " + }, + { + "bbox": [ + 313, + 483, + 553, + 554 + ], + "type": "inline_equation", + "content": "\\eta = \\kappa /\\| \\pmb {\\lambda}\\| _0" + }, + { + "bbox": [ + 313, + 483, + 553, + 554 + ], + "type": "text", + "content": " in each iteration of our proposed method. We show ablation experiments for different values of " + }, + { + "bbox": [ + 313, + 483, + 553, + 554 + ], + "type": "inline_equation", + "content": "\\kappa" + }, + { + "bbox": [ + 313, + 483, + 553, + 554 + ], + "type": "text", + "content": " in section 4.2.1." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 562, + 553, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 562, + 553, + 598 + ], + "spans": [ + { + "bbox": [ + 313, + 562, + 553, + 598 + ], + "type": "text", + "content": "Multi-Label Recognition Models We evaluate the baselines and our method on state-of-the-art multi-label recognition models (using the officially provided weights)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 601, + 555, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 601, + 555, + 686 + ], + "spans": [ + { + "bbox": [ + 313, + 601, + 555, + 686 + ], + "type": "text", + "content": "- Asymmetric Loss (ASL) [79] is a novel loss function to address the imbalance between numerous negative labels and only a few positive labels associated with each sample. This method balances the probabilities of different positive and negative samples. In particular, the loss performs hard thresholding by dynamically down-weighting easy negative samples and can discard unlabeled samples." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 689, + 555, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 712 + ], + "type": "text", + "content": "- ML-Decoder [80] is an efficient, attention-based multi-label classifier that has a novel decoder architecture and" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20584" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 67, + 74, + 171, + 181 + ], + "blocks": [ + { + "bbox": [ + 67, + 74, + 171, + 181 + ], + "lines": [ + { + "bbox": [ + 67, + 74, + 171, + 181 + ], + "spans": [ + { + "bbox": [ + 67, + 74, + 171, + 181 + ], + "type": "image", + "image_path": "825c2522eda98ed053528cb905dad9b4ca06c03d824e395fac4cf19589e2c35e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 183, + 553, + 205 + ], + "lines": [ + { + "bbox": [ + 55, + 183, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 55, + 183, + 553, + 205 + ], + "type": "text", + "content": "Figure 3. Fooling success rates " + }, + { + "bbox": [ + 55, + 183, + 553, + 205 + ], + "type": "inline_equation", + "content": "(FR)" + }, + { + "bbox": [ + 55, + 183, + 553, + 205 + ], + "type": "text", + "content": " on various datasets using ASL and ML-Decoder as we increase the target set size " + }, + { + "bbox": [ + 55, + 183, + 553, + 205 + ], + "type": "inline_equation", + "content": "|\\Omega|" + }, + { + "bbox": [ + 55, + 183, + 553, + 205 + ], + "type": "text", + "content": ". The left two figures show results on NUS-WIDE and the right two figures show results on MS-COCO." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 173, + 74, + 276, + 181 + ], + "blocks": [ + { + "bbox": [ + 173, + 74, + 276, + 181 + ], + "lines": [ + { + "bbox": [ + 173, + 74, + 276, + 181 + ], + "spans": [ + { + "bbox": [ + 173, + 74, + 276, + 181 + ], + "type": "image", + "image_path": "e27be3d6f28315f9f360620504b2000c774600c9e4e3c28f19cfe925e72c162a.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 277, + 74, + 387, + 181 + ], + "blocks": [ + { + "bbox": [ + 277, + 74, + 387, + 181 + ], + "lines": [ + { + "bbox": [ + 277, + 74, + 387, + 181 + ], + "spans": [ + { + "bbox": [ + 277, + 74, + 387, + 181 + ], + "type": "image", + "image_path": "a185711a2e7d0c80029ff3b27b8acc95486b332fdc030162253977d825a82952.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 387, + 74, + 491, + 181 + ], + "blocks": [ + { + "bbox": [ + 387, + 74, + 491, + 181 + ], + "lines": [ + { + "bbox": [ + 387, + 74, + 491, + 181 + ], + "spans": [ + { + "bbox": [ + 387, + 74, + 491, + 181 + ], + "type": "image", + "image_path": "be9818a4d3dc4763d1fe2360fb96d9710839b6dda581b11a8ccce9c6ddd12918.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 491, + 83, + 539, + 181 + ], + "blocks": [ + { + "bbox": [ + 491, + 83, + 539, + 181 + ], + "lines": [ + { + "bbox": [ + 491, + 83, + 539, + 181 + ], + "spans": [ + { + "bbox": [ + 491, + 83, + 539, + 181 + ], + "type": "image", + "image_path": "478457689f7413ff080020b02544cf847dc88442da297593ec3dc3a7810d75e1.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 224, + 294, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 224, + 294, + 272 + ], + "spans": [ + { + "bbox": [ + 54, + 224, + 294, + 272 + ], + "type": "text", + "content": "group-decoding scheme, which enables it to efficiently utilize the spatial information in images. The model is highly efficient as it can scale to large number of classes. Moreover, using word queries, it can generalize to unseen classes." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 276, + 295, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 276, + 295, + 324 + ], + "spans": [ + { + "bbox": [ + 55, + 276, + 295, + 324 + ], + "type": "text", + "content": "Oracle and Baselines We evaluate our method against three baselines and an oracle. The computational complexity of each of these methods is reported in Table 1 and their wall clock time in supplementary." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 329, + 296, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 329, + 296, + 390 + ], + "spans": [ + { + "bbox": [ + 55, + 329, + 296, + 390 + ], + "type": "text", + "content": "- Oracle directly optimizes (1) to learn a universal perturbation " + }, + { + "bbox": [ + 55, + 329, + 296, + 390 + ], + "type": "inline_equation", + "content": "\\pmb{u}_{\\Omega}" + }, + { + "bbox": [ + 55, + 329, + 296, + 390 + ], + "type": "text", + "content": " to attack a specific target set " + }, + { + "bbox": [ + 55, + 329, + 296, + 390 + ], + "type": "inline_equation", + "content": "\\Omega" + }, + { + "bbox": [ + 55, + 329, + 296, + 390 + ], + "type": "text", + "content": ". This provides the upper bound on the attack performance for our method. It requires " + }, + { + "bbox": [ + 55, + 329, + 296, + 390 + ], + "type": "inline_equation", + "content": "2^{|C|}d" + }, + { + "bbox": [ + 55, + 329, + 296, + 390 + ], + "type": "text", + "content": "-dimensional universal vectors and for each vector, iterate through the training set for " + }, + { + "bbox": [ + 55, + 329, + 296, + 390 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 55, + 329, + 296, + 390 + ], + "type": "text", + "content": " epochs." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 394, + 296, + 723 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 55, + 394, + 296, + 444 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 394, + 296, + 444 + ], + "spans": [ + { + "bbox": [ + 55, + 394, + 296, + 444 + ], + "type": "text", + "content": "- Oracle Sum (Or-S) directly learns single-class universal perturbation vectors " + }, + { + "bbox": [ + 55, + 394, + 296, + 444 + ], + "type": "inline_equation", + "content": "\\{\\pmb{u}_1, \\pmb{u}_2, \\dots, \\pmb{u}_{|\\mathcal{C}|}\\}" + }, + { + "bbox": [ + 55, + 394, + 296, + 444 + ], + "type": "text", + "content": " using (1) and generates a universal perturbation to attack set " + }, + { + "bbox": [ + 55, + 394, + 296, + 444 + ], + "type": "inline_equation", + "content": "\\Omega" + }, + { + "bbox": [ + 55, + 394, + 296, + 444 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 55, + 394, + 296, + 444 + ], + "type": "inline_equation", + "content": "\\pmb{u}_{\\Omega} = \\sum_{i \\in \\Omega} \\pmb{u}_i" + }, + { + "bbox": [ + 55, + 394, + 296, + 444 + ], + "type": "text", + "content": ". It only needs to learn " + }, + { + "bbox": [ + 55, + 394, + 296, + 444 + ], + "type": "inline_equation", + "content": "|C|" + }, + { + "bbox": [ + 55, + 394, + 296, + 444 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 55, + 394, + 296, + 444 + ], + "type": "text", + "content": "-dimensional vectors." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 448, + 296, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 448, + 296, + 521 + ], + "spans": [ + { + "bbox": [ + 55, + 448, + 296, + 521 + ], + "type": "text", + "content": "- Oracle Combination (Or-C) first learns single-class universal perturbation vectors and learns to combine them to attack a specific target set " + }, + { + "bbox": [ + 55, + 448, + 296, + 521 + ], + "type": "inline_equation", + "content": "\\Omega" + }, + { + "bbox": [ + 55, + 448, + 296, + 521 + ], + "type": "text", + "content": " i.e., " + }, + { + "bbox": [ + 55, + 448, + 296, + 521 + ], + "type": "inline_equation", + "content": "\\pmb{u}_{\\Omega} = \\sum_{i\\in \\Omega}a_i^\\Omega \\pmb {u}_i" + }, + { + "bbox": [ + 55, + 448, + 296, + 521 + ], + "type": "text", + "content": ". Where, " + }, + { + "bbox": [ + 55, + 448, + 296, + 521 + ], + "type": "inline_equation", + "content": "a_i^\\Omega \\in \\mathbb{R}" + }, + { + "bbox": [ + 55, + 448, + 296, + 521 + ], + "type": "text", + "content": " is associated with " + }, + { + "bbox": [ + 55, + 448, + 296, + 521 + ], + "type": "inline_equation", + "content": "i^{th}" + }, + { + "bbox": [ + 55, + 448, + 296, + 521 + ], + "type": "text", + "content": " class in specific set " + }, + { + "bbox": [ + 55, + 448, + 296, + 521 + ], + "type": "inline_equation", + "content": "\\Omega" + }, + { + "bbox": [ + 55, + 448, + 296, + 521 + ], + "type": "text", + "content": ", and is learnt by iterating through the training set. It needs to learn " + }, + { + "bbox": [ + 55, + 448, + 296, + 521 + ], + "type": "inline_equation", + "content": "|C|" + }, + { + "bbox": [ + 55, + 448, + 296, + 521 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 55, + 448, + 296, + 521 + ], + "type": "text", + "content": "-dim vectors and " + }, + { + "bbox": [ + 55, + 448, + 296, + 521 + ], + "type": "inline_equation", + "content": "2^{|C|}" + }, + { + "bbox": [ + 55, + 448, + 296, + 521 + ], + "type": "text", + "content": " scalars to combine them." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 525, + 296, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 525, + 296, + 586 + ], + "spans": [ + { + "bbox": [ + 55, + 525, + 296, + 586 + ], + "type": "text", + "content": "- SGA [59] addresses the gradient vanishing and poor local optima problem to improve the generalization of UAPs. It enhances the gradient stability by aggregating gradients from inner pre-search. We use SGA to learn single-class UAPs and use (4) to generate UAPs for target labels." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 590, + 295, + 723 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 590, + 295, + 723 + ], + "spans": [ + { + "bbox": [ + 55, + 590, + 295, + 723 + ], + "type": "text", + "content": "- NAG [68] models the distribution of universal adversarial perturbations using GANs, originally designed for untargeted perturbations in multi-class setting. To adapt this for multi-label universal perturbations, we condition the GAN on class vectors. During training, we generate label-wise universal perturbations. At test time, we generate perturbations for multiple classes by conditioning the GAN on a combined representation vector of the target classes. For this method, we have linear time complexity as we train to generate label-wise perturbations and the space requirement is the size of the generator." + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 224, + 553, + 368 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 224, + 553, + 368 + ], + "spans": [ + { + "bbox": [ + 313, + 224, + 553, + 368 + ], + "type": "text", + "content": "Evaluation Metrics Given a set of classes " + }, + { + "bbox": [ + 313, + 224, + 553, + 368 + ], + "type": "inline_equation", + "content": "\\Omega" + }, + { + "bbox": [ + 313, + 224, + 553, + 368 + ], + "type": "text", + "content": " to attack, let " + }, + { + "bbox": [ + 313, + 224, + 553, + 368 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_{\\Omega}" + }, + { + "bbox": [ + 313, + 224, + 553, + 368 + ], + "type": "text", + "content": " represent all the images that contain the classes " + }, + { + "bbox": [ + 313, + 224, + 553, + 368 + ], + "type": "inline_equation", + "content": "\\Omega" + }, + { + "bbox": [ + 313, + 224, + 553, + 368 + ], + "type": "text", + "content": " and were attacked. Let " + }, + { + "bbox": [ + 313, + 224, + 553, + 368 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_{\\Omega} \\subseteq \\mathcal{I}_{\\Omega}" + }, + { + "bbox": [ + 313, + 224, + 553, + 368 + ], + "type": "text", + "content": " represent the set of images that were successfully attacked. To evaluate the universal attack performance, we define the attack fooling success rate, " + }, + { + "bbox": [ + 313, + 224, + 553, + 368 + ], + "type": "inline_equation", + "content": "FR" + }, + { + "bbox": [ + 313, + 224, + 553, + 368 + ], + "type": "text", + "content": ", to measure the percentage of successfully attacked images. Since we want to affect the least number of non-targeted labels while achieving high fooling rates, we also report non-target flip rate, " + }, + { + "bbox": [ + 313, + 224, + 553, + 368 + ], + "type": "inline_equation", + "content": "NT_{R}" + }, + { + "bbox": [ + 313, + 224, + 553, + 368 + ], + "type": "text", + "content": ", to measure the percentage of non-targeted labels (labels in " + }, + { + "bbox": [ + 313, + 224, + 553, + 368 + ], + "type": "inline_equation", + "content": "\\bar{\\Omega}" + }, + { + "bbox": [ + 313, + 224, + 553, + 368 + ], + "type": "text", + "content": ") which were flipped by the attack. Our goal is to achieve high " + }, + { + "bbox": [ + 313, + 224, + 553, + 368 + ], + "type": "inline_equation", + "content": "FR" + }, + { + "bbox": [ + 313, + 224, + 553, + 368 + ], + "type": "text", + "content": " while maintaining low " + }, + { + "bbox": [ + 313, + 224, + 553, + 368 + ], + "type": "inline_equation", + "content": "NT_{R}" + }, + { + "bbox": [ + 313, + 224, + 553, + 368 + ], + "type": "text", + "content": " for attacking target set of any size." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 324, + 372, + 553, + 412 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 372, + 553, + 412 + ], + "spans": [ + { + "bbox": [ + 324, + 372, + 553, + 412 + ], + "type": "interline_equation", + "content": "\\mathrm {F R} = \\frac {\\left| \\mathcal {A} _ {\\Omega} \\right|}{\\left| \\mathcal {I} _ {\\Omega} \\right|}, \\mathrm {N T} _ {R} = \\frac {1}{\\left| \\mathcal {A} _ {\\Omega} \\right|} \\sum_ {k \\in \\mathcal {A} _ {\\Omega}} \\frac {\\sum_ {i \\in \\bar {\\Omega}} \\left(1 - \\delta \\left(\\chi \\left(f _ {i} ^ {(k)}\\right) , y _ {i} ^ {(k)}\\right)\\right)}{\\left| \\bar {\\Omega} \\right|} \\tag {11}", + "image_path": "a69ed43b165008f79dfacf8855f010f2b08510299d0804ea537cd40d78b8d6b7.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 412, + 554, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 412, + 554, + 498 + ], + "spans": [ + { + "bbox": [ + 313, + 412, + 554, + 498 + ], + "type": "text", + "content": "where, " + }, + { + "bbox": [ + 313, + 412, + 554, + 498 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 313, + 412, + 554, + 498 + ], + "type": "text", + "content": " is Kronecker delta that equals 1 when the two inputs are equal and 0 otherwise. " + }, + { + "bbox": [ + 313, + 412, + 554, + 498 + ], + "type": "inline_equation", + "content": "\\chi (v)" + }, + { + "bbox": [ + 313, + 412, + 554, + 498 + ], + "type": "text", + "content": " is the step function which is 1 if " + }, + { + "bbox": [ + 313, + 412, + 554, + 498 + ], + "type": "inline_equation", + "content": "v > 0" + }, + { + "bbox": [ + 313, + 412, + 554, + 498 + ], + "type": "text", + "content": " and 0 otherwise. " + }, + { + "bbox": [ + 313, + 412, + 554, + 498 + ], + "type": "inline_equation", + "content": "y_{i}^{(k)},f_{i}^{(k)}" + }, + { + "bbox": [ + 313, + 412, + 554, + 498 + ], + "type": "text", + "content": " are the predictions on clean and perturbed images respectively of " + }, + { + "bbox": [ + 313, + 412, + 554, + 498 + ], + "type": "inline_equation", + "content": "i^{th}" + }, + { + "bbox": [ + 313, + 412, + 554, + 498 + ], + "type": "text", + "content": " non-target class of " + }, + { + "bbox": [ + 313, + 412, + 554, + 498 + ], + "type": "inline_equation", + "content": "k^{th}" + }, + { + "bbox": [ + 313, + 412, + 554, + 498 + ], + "type": "text", + "content": " successfully attacked image. Note that " + }, + { + "bbox": [ + 313, + 412, + 554, + 498 + ], + "type": "inline_equation", + "content": "FR" + }, + { + "bbox": [ + 313, + 412, + 554, + 498 + ], + "type": "text", + "content": " is calculated using all attacked images while " + }, + { + "bbox": [ + 313, + 412, + 554, + 498 + ], + "type": "inline_equation", + "content": "NT_{R}" + }, + { + "bbox": [ + 313, + 412, + 554, + 498 + ], + "type": "text", + "content": " is calculated using the set of successfully attacked images." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 505, + 437, + 517 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 505, + 437, + 517 + ], + "spans": [ + { + "bbox": [ + 313, + 505, + 437, + 517 + ], + "type": "text", + "content": "4.2. Experimental Results" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 522, + 554, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 522, + 554, + 617 + ], + "spans": [ + { + "bbox": [ + 313, + 522, + 554, + 617 + ], + "type": "text", + "content": "In this section, we evaluate our compositional framework for universal perturbations. Fig. 3 shows the fooling success rates for different methods. The results demonstrate consistent performance trends across datasets and models. Particularly, the success rate(FR) of baselines decline as we increase the target size. On the other hand, the Oracle method achieves the best performance but suffers from exponential complexity, making it impractical for larger target sets." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 618, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 618, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 618, + 555, + 714 + ], + "type": "text", + "content": "Among the baselines, Or-C, Or-S, and SGA perform the worst for large target set sizes. Since the performance of these methods depends on the label-wise universal perturbations, we can conclude that the low success rate of these methods implies that the subspace spanned by labelwise perturbations does not contain the intersection of their respective adversarial regions. This is further elaborated in Fig. 5. In all cases, Or-C is better than Or-S since it op-" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20585" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 72, + 110, + 538, + 243 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 105 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 105 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 105 + ], + "type": "text", + "content": "Table 1. Non-target flip rate " + }, + { + "bbox": [ + 55, + 70, + 555, + 105 + ], + "type": "inline_equation", + "content": "(NT_R)" + }, + { + "bbox": [ + 55, + 70, + 555, + 105 + ], + "type": "text", + "content": " on NUS-WIDE and MS-COCO, as we increase the target set size from 1 to 5. We show (-) when " + }, + { + "bbox": [ + 55, + 70, + 555, + 105 + ], + "type": "inline_equation", + "content": "|\\mathcal{A}_{\\Omega}| = 0" + }, + { + "bbox": [ + 55, + 70, + 555, + 105 + ], + "type": "text", + "content": ". We also show the space and time complexity of each method. " + }, + { + "bbox": [ + 55, + 70, + 555, + 105 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 55, + 70, + 555, + 105 + ], + "type": "text", + "content": " represents the number of epochs for training, " + }, + { + "bbox": [ + 55, + 70, + 555, + 105 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 55, + 70, + 555, + 105 + ], + "type": "text", + "content": " is the size of the perturbation (equal to the image size), " + }, + { + "bbox": [ + 55, + 70, + 555, + 105 + ], + "type": "inline_equation", + "content": "G_p" + }, + { + "bbox": [ + 55, + 70, + 555, + 105 + ], + "type": "text", + "content": " is the size of GAN's parameters, and " + }, + { + "bbox": [ + 55, + 70, + 555, + 105 + ], + "type": "inline_equation", + "content": "|\\mathcal{C}|" + }, + { + "bbox": [ + 55, + 70, + 555, + 105 + ], + "type": "text", + "content": " is the number of classes." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 72, + 110, + 538, + 243 + ], + "lines": [ + { + "bbox": [ + 72, + 110, + 538, + 243 + ], + "spans": [ + { + "bbox": [ + 72, + 110, + 538, + 243 + ], + "type": "table", + "html": "
Computational ComplexityASLML-Decoder
NUS-WIDEMS-COCONUS-WIDEMS-COCO
MethodSpaceTime135135135135
Oracle\\( \\mathcal{O}(d2^{|C|}) \\)\\( \\mathcal{O}(E2^{|C|}) \\)6.177.428.106.507.738.662.863.103.017.388.589.81
Or-S\\( \\mathcal{O}(d|C|) \\)\\( \\mathcal{O}(E|C|) \\)6.176.03-6.502.872.962.862.22-7.382.96-
Or-C\\( \\mathcal{O}(d|C|+2^{|C|}) \\)\\( \\mathcal{O}(E2^{|C|}) \\)6.176.862.746.505.233.202.862.450.607.386.40-
SGA\\( \\mathcal{O}(d|C|) \\)\\( \\mathcal{O}(E|C|) \\)7.697.31-7.774.30-2.892.471.278.628.949.37
NAG\\( \\mathcal{O}(G_p) \\)\\( \\mathcal{O}(E|C|) \\)8.325.861.3210.110.27.213.963.452.5511.812.113.7
\\( \\mathbf{CMLU}_{\\alpha} \\)\\( \\mathcal{O}(d|C|) \\)\\( \\mathcal{O}(EN_{\\alpha}|C|) \\)5.637.653.747.335.963.133.63.082.518.968.558.02
\\( \\mathbf{CMLU}_{\\beta} \\)\\( \\mathcal{O}(d|C|) \\)\\( \\mathcal{O}(E|C|) \\)7.968.137.938.908.869.594.233.743.029.659.8310.5
", + "image_path": "199186af8ff04b4342ab6687dd6c89c94dd28b5655b1a089f3576b8692b368c2.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 72, + 281, + 538, + 399 + ], + "blocks": [ + { + "bbox": [ + 55, + 251, + 555, + 275 + ], + "lines": [ + { + "bbox": [ + 55, + 251, + 555, + 275 + ], + "spans": [ + { + "bbox": [ + 55, + 251, + 555, + 275 + ], + "type": "text", + "content": "Table 2. Experiments on large-scale dataset. We report fooling success rate (FR) and non-target flip rate " + }, + { + "bbox": [ + 55, + 251, + 555, + 275 + ], + "type": "inline_equation", + "content": "(\\mathrm{NT}_R)" + }, + { + "bbox": [ + 55, + 251, + 555, + 275 + ], + "type": "text", + "content": " on OpenImages using ASL model. The experiments were performed on large target set sizes, up to " + }, + { + "bbox": [ + 55, + 251, + 555, + 275 + ], + "type": "inline_equation", + "content": "|\\Omega| = 20" + }, + { + "bbox": [ + 55, + 251, + 555, + 275 + ], + "type": "text", + "content": ". Higher FR (" + }, + { + "bbox": [ + 55, + 251, + 555, + 275 + ], + "type": "inline_equation", + "content": "\\uparrow" + }, + { + "bbox": [ + 55, + 251, + 555, + 275 + ], + "type": "text", + "content": ") and lower " + }, + { + "bbox": [ + 55, + 251, + 555, + 275 + ], + "type": "inline_equation", + "content": "\\mathrm{NT}_R" + }, + { + "bbox": [ + 55, + 251, + 555, + 275 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 55, + 251, + 555, + 275 + ], + "type": "inline_equation", + "content": "\\downarrow" + }, + { + "bbox": [ + 55, + 251, + 555, + 275 + ], + "type": "text", + "content": ") are better." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 72, + 281, + 538, + 399 + ], + "lines": [ + { + "bbox": [ + 72, + 281, + 538, + 399 + ], + "spans": [ + { + "bbox": [ + 72, + 281, + 538, + 399 + ], + "type": "table", + "html": "
MethodFR (↑)NTR(↓)
1234510152012345101520
Oracle97.697.996.897.995.796.396.896.80.640.800.870.930.961.051.171.47
Or-S97.629.646.281.700.420.00.00.00.640.710.650.340.20---
Or-C97.649.422.823.313.40.00.00.00.640.710.670.750.48---
SGA98.941.49.646.110.920.150.00.00.770.810.730.740.320.12--
NAG95.992.188.888.489.779.980.984.20.810.890.931.011.071.421.531.84
CMLUα92.266.744.046.239.439.932.432.30.720.760.830.890.830.860.820.77
CMLUβ97.395.895.996.895.294.896.595.20.720.800.880.920.981.061.281.67
", + "image_path": "b3143b6e2d65e6df1ee6aadaffa70e9ef0892e7bf74eb958cb05b0a2f7b9c923.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 70, + 418, + 279, + 523 + ], + "blocks": [ + { + "bbox": [ + 70, + 418, + 279, + 523 + ], + "lines": [ + { + "bbox": [ + 70, + 418, + 279, + 523 + ], + "spans": [ + { + "bbox": [ + 70, + 418, + 279, + 523 + ], + "type": "image", + "image_path": "9e0399c2f83fa5a5d025b5c5d75c33b82aca25c33341ec722e3b9fe14bf0a7ec.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 525, + 295, + 548 + ], + "lines": [ + { + "bbox": [ + 55, + 525, + 295, + 548 + ], + "spans": [ + { + "bbox": [ + 55, + 525, + 295, + 548 + ], + "type": "text", + "content": "Figure 4. Visualizing label-wise universal perturbations learnt by Oracle and " + }, + { + "bbox": [ + 55, + 525, + 295, + 548 + ], + "type": "inline_equation", + "content": "\\mathrm{CMLU}_{\\beta}" + }, + { + "bbox": [ + 55, + 525, + 295, + 548 + ], + "type": "text", + "content": " on NUS-WIDE using ASL." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 567, + 295, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 567, + 295, + 651 + ], + "spans": [ + { + "bbox": [ + 55, + 567, + 295, + 651 + ], + "type": "text", + "content": "timizes to find the best combination of the label-wise universal perturbations and hence, is better than the naive summation of individual perturbations. Across all experiments, " + }, + { + "bbox": [ + 55, + 567, + 295, + 651 + ], + "type": "inline_equation", + "content": "\\mathrm{CMLU}_{\\beta}" + }, + { + "bbox": [ + 55, + 567, + 295, + 651 + ], + "type": "text", + "content": " is the closest to Oracle performance. This is despite the fact that " + }, + { + "bbox": [ + 55, + 567, + 295, + 651 + ], + "type": "inline_equation", + "content": "\\mathrm{CMLU}_{\\beta}" + }, + { + "bbox": [ + 55, + 567, + 295, + 651 + ], + "type": "text", + "content": " has linear complexity whereas the Oracle has exponential complexity. Moreover, " + }, + { + "bbox": [ + 55, + 567, + 295, + 651 + ], + "type": "inline_equation", + "content": "\\mathrm{CMLU}_{\\alpha}" + }, + { + "bbox": [ + 55, + 567, + 295, + 651 + ], + "type": "text", + "content": " achieves higher " + }, + { + "bbox": [ + 55, + 567, + 295, + 651 + ], + "type": "inline_equation", + "content": "FR" + }, + { + "bbox": [ + 55, + 567, + 295, + 651 + ], + "type": "text", + "content": " and lower " + }, + { + "bbox": [ + 55, + 567, + 295, + 651 + ], + "type": "inline_equation", + "content": "NT_{R}" + }, + { + "bbox": [ + 55, + 567, + 295, + 651 + ], + "type": "text", + "content": " than all baselines." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "content": "Table 1 shows the percentage of non-targeted affected labels (whose prediction was changed after the attack). Although the baselines have low " + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{NT}_R" + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "content": " for target sizes greater than 1, they also have low fooling rates. Note that despite having linear complexity, the " + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{NT}_R" + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{CMLU}_{\\beta}" + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "content": " is only" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 419, + 555, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 419, + 555, + 528 + ], + "spans": [ + { + "bbox": [ + 313, + 419, + 555, + 528 + ], + "type": "text", + "content": "slightly higher than Oracle. " + }, + { + "bbox": [ + 313, + 419, + 555, + 528 + ], + "type": "inline_equation", + "content": "\\mathrm{CMLU}_{\\alpha}" + }, + { + "bbox": [ + 313, + 419, + 555, + 528 + ], + "type": "text", + "content": " achieves lower " + }, + { + "bbox": [ + 313, + 419, + 555, + 528 + ], + "type": "inline_equation", + "content": "NT_{R}" + }, + { + "bbox": [ + 313, + 419, + 555, + 528 + ], + "type": "text", + "content": " as compared to " + }, + { + "bbox": [ + 313, + 419, + 555, + 528 + ], + "type": "inline_equation", + "content": "\\mathrm{CMLU}_{\\beta}" + }, + { + "bbox": [ + 313, + 419, + 555, + 528 + ], + "type": "text", + "content": " but still outperforms the baselines. In all experiments, NAG performs well (close to " + }, + { + "bbox": [ + 313, + 419, + 555, + 528 + ], + "type": "inline_equation", + "content": "\\mathrm{CMLU}_{\\alpha}" + }, + { + "bbox": [ + 313, + 419, + 555, + 528 + ], + "type": "text", + "content": ") but declines in performance as we increase the target size. It also has significantly higher " + }, + { + "bbox": [ + 313, + 419, + 555, + 528 + ], + "type": "inline_equation", + "content": "\\mathrm{NT}_{R}" + }, + { + "bbox": [ + 313, + 419, + 555, + 528 + ], + "type": "text", + "content": " than the Oracle and " + }, + { + "bbox": [ + 313, + 419, + 555, + 528 + ], + "type": "inline_equation", + "content": "\\mathrm{CMLU}_{\\beta}" + }, + { + "bbox": [ + 313, + 419, + 555, + 528 + ], + "type": "text", + "content": ". This implies that NAG learnt to attack multiple classes together rather than learning the distribution of label-wise universal perturbations. Therefore, even though it has high success rate, it also affects several other labels." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "type": "text", + "content": "We further perform experiments on OpenImages, a large-scale dataset, and we show results in Tab. 2. Performance trends remain consistent with those observed on smaller datasets. The success rate of Or-S and Or-C consistently drops as we target more classes, and gets to 0 by target size 10. Note that NAG performs better in OpenImages experiments. As shown in the table, NAG affects a large percentage of non-targeted labels, which signifies that the method does not capture the class-specific directions and perturbs a large group of labels together. " + }, + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{CMLU}_{\\alpha}" + }, + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "type": "text", + "content": " performs better than Or-C, Or-S, and SGA but worse than NAG. However, as shown in the table, " + }, + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{CMLU}_{\\alpha}" + }, + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "type": "text", + "content": " still achieves significantly lower " + }, + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "type": "inline_equation", + "content": "NT_{R}" + }, + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "type": "text", + "content": " for large target set sizes. In all experiments, " + }, + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{CMLU}_{\\beta}" + }, + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "type": "text", + "content": " achieves consistently close performance to the Oracle while also maintaining low " + }, + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{NT}_{R}" + }, + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20586" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 67, + 72, + 282, + 201 + ], + "blocks": [ + { + "bbox": [ + 67, + 72, + 282, + 201 + ], + "lines": [ + { + "bbox": [ + 67, + 72, + 282, + 201 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 282, + 201 + ], + "type": "image", + "image_path": "bcc18fa01204e1d900ed1f8b6bcc99ca0ee45a595d6dc205c9f038e696f1a4b7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 207, + 296, + 286 + ], + "lines": [ + { + "bbox": [ + 55, + 207, + 296, + 286 + ], + "spans": [ + { + "bbox": [ + 55, + 207, + 296, + 286 + ], + "type": "text", + "content": "Figure 5. Comparison of the regions (convex cones) spanned by universal perturbations learnt using Oracle and " + }, + { + "bbox": [ + 55, + 207, + 296, + 286 + ], + "type": "inline_equation", + "content": "\\mathrm{CMLU}_{\\beta}" + }, + { + "bbox": [ + 55, + 207, + 296, + 286 + ], + "type": "text", + "content": " on NUS-WIDE (top) and MS-COCO (bottom). Each point in the heatmap is the average model confidence computed across images using the vector " + }, + { + "bbox": [ + 55, + 207, + 296, + 286 + ], + "type": "inline_equation", + "content": "(\\epsilon_x * \\boldsymbol{u}_x + \\epsilon_y * \\boldsymbol{u}_y)" + }, + { + "bbox": [ + 55, + 207, + 296, + 286 + ], + "type": "text", + "content": " for all label combinations of size " + }, + { + "bbox": [ + 55, + 207, + 296, + 286 + ], + "type": "inline_equation", + "content": "|\\Omega| = 2" + }, + { + "bbox": [ + 55, + 207, + 296, + 286 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 207, + 296, + 286 + ], + "type": "inline_equation", + "content": "\\boldsymbol{u}_x, \\boldsymbol{u}_y" + }, + { + "bbox": [ + 55, + 207, + 296, + 286 + ], + "type": "text", + "content": " are universal vectors of classes in " + }, + { + "bbox": [ + 55, + 207, + 296, + 286 + ], + "type": "inline_equation", + "content": "\\Omega" + }, + { + "bbox": [ + 55, + 207, + 296, + 286 + ], + "type": "text", + "content": ". The asterisks(*) show the region where the " + }, + { + "bbox": [ + 55, + 207, + 296, + 286 + ], + "type": "inline_equation", + "content": "L_{\\infty}" + }, + { + "bbox": [ + 55, + 207, + 296, + 286 + ], + "type": "text", + "content": " norm is 0.05." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 296, + 149, + 308 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 296, + 149, + 308 + ], + "spans": [ + { + "bbox": [ + 55, + 296, + 149, + 308 + ], + "type": "text", + "content": "4.2.1. Ablation Study" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 310, + 296, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 310, + 296, + 443 + ], + "spans": [ + { + "bbox": [ + 55, + 310, + 296, + 443 + ], + "type": "text", + "content": "We perform ablation experiments on Oracle and " + }, + { + "bbox": [ + 55, + 310, + 296, + 443 + ], + "type": "inline_equation", + "content": "\\mathrm{CMLU}_{\\beta}" + }, + { + "bbox": [ + 55, + 310, + 296, + 443 + ], + "type": "text", + "content": " to investigate the effect of hyperparameters on their performance. Tab. 3 shows the results as we increase/decrease the values of " + }, + { + "bbox": [ + 55, + 310, + 296, + 443 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 55, + 310, + 296, + 443 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 310, + 296, + 443 + ], + "type": "inline_equation", + "content": "\\kappa" + }, + { + "bbox": [ + 55, + 310, + 296, + 443 + ], + "type": "text", + "content": ". Increasing " + }, + { + "bbox": [ + 55, + 310, + 296, + 443 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 55, + 310, + 296, + 443 + ], + "type": "text", + "content": " provides a larger perturbation budget for better performance; however, to keep the changes to the inputs imperceptible, we set " + }, + { + "bbox": [ + 55, + 310, + 296, + 443 + ], + "type": "inline_equation", + "content": "\\epsilon = 0.05" + }, + { + "bbox": [ + 55, + 310, + 296, + 443 + ], + "type": "text", + "content": ". Meanwhile, increasing " + }, + { + "bbox": [ + 55, + 310, + 296, + 443 + ], + "type": "inline_equation", + "content": "\\kappa" + }, + { + "bbox": [ + 55, + 310, + 296, + 443 + ], + "type": "text", + "content": " emphasizes maintaining non-targeted labels, which lowers " + }, + { + "bbox": [ + 55, + 310, + 296, + 443 + ], + "type": "inline_equation", + "content": "\\mathrm{NT}_R" + }, + { + "bbox": [ + 55, + 310, + 296, + 443 + ], + "type": "text", + "content": " but also reduces the success rate (FR). This is because many labels can be correlated in MLL and attacking a set of labels while fixing non-targeted but correlated labels lead to sub-optimal perturbations [63]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 447, + 264, + 460 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 447, + 264, + 460 + ], + "spans": [ + { + "bbox": [ + 55, + 447, + 264, + 460 + ], + "type": "text", + "content": "4.2.2. Analysis of Universal Adversarial Regions" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 462, + 295, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 462, + 295, + 594 + ], + "spans": [ + { + "bbox": [ + 55, + 462, + 295, + 594 + ], + "type": "text", + "content": "The label-wise universal perturbations learnt by Oracle and " + }, + { + "bbox": [ + 55, + 462, + 295, + 594 + ], + "type": "inline_equation", + "content": "\\mathrm{CMLU}_{\\beta}" + }, + { + "bbox": [ + 55, + 462, + 295, + 594 + ], + "type": "text", + "content": " are visualized in Fig. 4. We quantitatively assess the visual impact of these perturbations using Structural Similarity Index (SSIM), computed between original and perturbed images. The average SSIM values across all images, as shown in Fig. 4, demonstrate comparable levels of distortion for Oracle and " + }, + { + "bbox": [ + 55, + 462, + 295, + 594 + ], + "type": "inline_equation", + "content": "\\mathrm{CMLU}_{\\beta}" + }, + { + "bbox": [ + 55, + 462, + 295, + 594 + ], + "type": "text", + "content": ". Note that the pairwise dot product of the Oracle and CMLU vectors is " + }, + { + "bbox": [ + 55, + 462, + 295, + 594 + ], + "type": "inline_equation", + "content": "\\approx 0" + }, + { + "bbox": [ + 55, + 462, + 295, + 594 + ], + "type": "text", + "content": ", indicating that " + }, + { + "bbox": [ + 55, + 462, + 295, + 594 + ], + "type": "inline_equation", + "content": "\\mathrm{CMLU}_{\\beta}" + }, + { + "bbox": [ + 55, + 462, + 295, + 594 + ], + "type": "text", + "content": " discover distinct attack directions, different from the Oracle. Additional analysis of pairwise dot products and visualizations are provided in supplementary." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "content": "To further highlight and analyze the distinction between the perturbations learnt using different methods, we focus on Oracle and " + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{CMLU}_{\\beta}" + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "content": ". Specifically, we examine the universal adversarial regions generated by those methods. Fig. 5 shows the average model confidence on the input space region spanned by the pairs of universal perturbation vectors. In the figure, perturbing inputs to blue and green regions would not change the model prediction (does not attack any class), yellow region would attack one class, and red region would attack both classes in " + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\Omega" + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "content": ". Note that there is" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 323, + 99, + 544, + 167 + ], + "blocks": [ + { + "bbox": [ + 313, + 71, + 555, + 95 + ], + "lines": [ + { + "bbox": [ + 313, + 71, + 555, + 95 + ], + "spans": [ + { + "bbox": [ + 313, + 71, + 555, + 95 + ], + "type": "text", + "content": "Table 3. Ablation experiments for hyperparameters on NUS-WIDE. The experiments are conducted for " + }, + { + "bbox": [ + 313, + 71, + 555, + 95 + ], + "type": "inline_equation", + "content": "|\\Omega| = 1" + }, + { + "bbox": [ + 313, + 71, + 555, + 95 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 323, + 99, + 544, + 167 + ], + "lines": [ + { + "bbox": [ + 323, + 99, + 544, + 167 + ], + "spans": [ + { + "bbox": [ + 323, + 99, + 544, + 167 + ], + "type": "table", + "html": "
εOracleCMLUβκOracleCMLUβ
FRNTRFRNTRFRNTRFRNTR
0.0163.55.3038.55.041599.16.5098.58.32
0.0394.96.2984.27.092098.96.1797.67.96
0.0598.96.1797.67.962593.54.7089.36.97
0.0799.75.1099.37.543093.14.4085.56.58
", + "image_path": "c125b1985b49fb4f6f7314431d48f5a52ef90f60336c7cd3fee682aa74630f81.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 323, + 209, + 544, + 303 + ], + "blocks": [ + { + "bbox": [ + 313, + 172, + 555, + 205 + ], + "lines": [ + { + "bbox": [ + 313, + 172, + 555, + 205 + ], + "spans": [ + { + "bbox": [ + 313, + 172, + 555, + 205 + ], + "type": "text", + "content": "Table 4. Transferability Evaluation: " + }, + { + "bbox": [ + 313, + 172, + 555, + 205 + ], + "type": "inline_equation", + "content": "S \\rightarrow T" + }, + { + "bbox": [ + 313, + 172, + 555, + 205 + ], + "type": "text", + "content": " shows the performance of universal perturbations learnt on model " + }, + { + "bbox": [ + 313, + 172, + 555, + 205 + ], + "type": "inline_equation", + "content": "\\mathbf{S}" + }, + { + "bbox": [ + 313, + 172, + 555, + 205 + ], + "type": "text", + "content": " and evaluated on model " + }, + { + "bbox": [ + 313, + 172, + 555, + 205 + ], + "type": "inline_equation", + "content": "\\mathbf{T}" + }, + { + "bbox": [ + 313, + 172, + 555, + 205 + ], + "type": "text", + "content": ". (A: ASL, D: ML-Decoder)" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 323, + 209, + 544, + 303 + ], + "lines": [ + { + "bbox": [ + 323, + 209, + 544, + 303 + ], + "spans": [ + { + "bbox": [ + 323, + 209, + 544, + 303 + ], + "type": "table", + "html": "
Transferability of Universal Attacks
ModelsMethodNUS-WIDEMS-COCO
135135
A → DOracle90.183.777.334.610.65.40
CMLUβ89.978.475.442.917.66.4
D → AOracle24.61.330.064.428.319.6
CMLUβ33.50.160.064.834.521.2
", + "image_path": "934f82742f8f9a63831ff81b2ac6f4b3179bb5b5f3f5cef9e9849003c327c5b0.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 312, + 555, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 312, + 555, + 479 + ], + "spans": [ + { + "bbox": [ + 313, + 312, + 555, + 479 + ], + "type": "text", + "content": "no prominent red region (to attack both classes) in the cross-section of Oracle vectors. This explains the low success rate of Or-S and Or-C which optimize to find red region in such a cross-section. More details on computing these cones is provided in supplementary. We also investigate how well those red regions computed using " + }, + { + "bbox": [ + 313, + 312, + 555, + 479 + ], + "type": "inline_equation", + "content": "\\mathrm{CMLU}_{\\beta}" + }, + { + "bbox": [ + 313, + 312, + 555, + 479 + ], + "type": "text", + "content": " can transfer to other models. For that, we perform transferability attacks and the results are shown in Tab. 4. From the table, we observe that " + }, + { + "bbox": [ + 313, + 312, + 555, + 479 + ], + "type": "inline_equation", + "content": "\\mathrm{CMLU}_{\\beta}" + }, + { + "bbox": [ + 313, + 312, + 555, + 479 + ], + "type": "text", + "content": " (linear complexity) achieves fooling success rates closer and often better than Oracle (exponential complexity). This shows that the convex cone regions corresponding to different labels are often common across models. This can be effectively used to generate black-box attacks. We hope that future works explore this further." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 490, + 388, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 490, + 388, + 502 + ], + "spans": [ + { + "bbox": [ + 314, + 490, + 388, + 502 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 509, + 555, + 631 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 509, + 555, + 631 + ], + "spans": [ + { + "bbox": [ + 313, + 509, + 555, + 631 + ], + "type": "text", + "content": "We addressed the problem of generating targeted multi-label universal perturbations. We developed a compositional framework to generate exponential number of multi-label universal perturbations in constant time. The independence assumption on label-wise universal perturbations naturally leads to a formulation that requires learning convex cones. Through extensive experiments on multiple datasets and models, we show that our method outperforms all baseline methods and achieves performance comparable to the Oracle (an exponential complexity method)." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 640, + 430, + 654 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 640, + 430, + 654 + ], + "spans": [ + { + "bbox": [ + 313, + 640, + 430, + 654 + ], + "type": "text", + "content": "6. Acknowledgements" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 659, + 555, + 709 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 659, + 555, + 709 + ], + "spans": [ + { + "bbox": [ + 313, + 659, + 555, + 709 + ], + "type": "text", + "content": "This work is sponsored by ARPA-H (1AY2AX000062), DARPA PTG (HR00112220001), NSF (IIS-2115110) and ARO (W911NF2110276). Content does not necessarily reflect the position/policy of the Government." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20587" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 62, + 91, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 66, + 91, + 296, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 91, + 296, + 145 + ], + "spans": [ + { + "bbox": [ + 66, + 91, + 296, + 145 + ], + "type": "text", + "content": "[1] Abhishek Aich, Calvin-Khang Ta, Akash Gupta, Chengyu Song, Srikanth Krishnamurthy, Salman Asif, and Amit Roy-Chowdhury. Gama: Generative adversarial multi-object scene attacks. Advances in Neural Information Processing Systems, 35:36914-36930, 2022. 1, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 147, + 296, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 147, + 296, + 213 + ], + "spans": [ + { + "bbox": [ + 66, + 147, + 296, + 213 + ], + "type": "text", + "content": "[2] Abhishek Aich, Shasha Li, Chengyu Song, M Salman Asif, Srikanth V Krishnamurthy, and Amit K Roy-Chowdhury. Leveraging local patch differences in multi-object scenes for generative adversarial attacks. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1308-1318, 2023. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 66, + 214, + 295, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 214, + 295, + 236 + ], + "spans": [ + { + "bbox": [ + 66, + 214, + 295, + 236 + ], + "type": "text", + "content": "[3] N. Akhtar and A. Mian. Threat of adversarial attacks on deep learning in computer vision: A survey. arXiv, 2018. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 66, + 238, + 294, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 238, + 294, + 270 + ], + "spans": [ + { + "bbox": [ + 66, + 238, + 294, + 270 + ], + "type": "text", + "content": "[4] A. Athalye, N. Carlini, and D. A. Wagner. Obfuscated gradients give a false sense of security: Circumventing defenses to adversarial examples. 2018. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 66, + 272, + 294, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 272, + 294, + 304 + ], + "spans": [ + { + "bbox": [ + 66, + 272, + 294, + 304 + ], + "type": "text", + "content": "[5] Yuanhao Ban and Yinpeng Dong. Pre-trained adversarial perturbations. In Advances in Neural Information Processing Systems, pages 1196-1209, 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 66, + 306, + 294, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 306, + 294, + 350 + ], + "spans": [ + { + "bbox": [ + 66, + 306, + 294, + 350 + ], + "type": "text", + "content": "[6] Emanuel Ben-Baruch, Tal Ridnik, Itamar Friedman, Avi Ben-Cohen, Nadav Zamir, Asaf Noy, and Lihi Zelnik-Manor. Multi-label classification with partial annotations using class-aware selective loss. 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 66, + 351, + 294, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 351, + 294, + 384 + ], + "spans": [ + { + "bbox": [ + 66, + 351, + 294, + 384 + ], + "type": "text", + "content": "[7] P. Benz, C. Zhang, T. Imtiaz, and I.-S. Kweon. Double targeted universal adversarial perturbations. Asian Conference on Computer Vision, 2020. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 66, + 385, + 294, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 385, + 294, + 430 + ], + "spans": [ + { + "bbox": [ + 66, + 385, + 294, + 430 + ], + "type": "text", + "content": "[8] Philipp Benz, Chaoning Zhang, Adil Karjauv, and In So Kweon. Universal adversarial training with class-wise perturbations. In 2021 IEEE International Conference on Multimedia and Expo (ICME), pages 1-6. IEEE, 2021. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 66, + 431, + 294, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 431, + 294, + 463 + ], + "spans": [ + { + "bbox": [ + 66, + 431, + 294, + 463 + ], + "type": "text", + "content": "[9] N. Carlini and D. Wagner. Adversarial examples are not easily detected: Bypassing ten detection methods. Workshop on Artificial Intelligence and Security, 2017. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 465, + 295, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 465, + 295, + 497 + ], + "spans": [ + { + "bbox": [ + 62, + 465, + 295, + 497 + ], + "type": "text", + "content": "[10] N. Carlini and D. Wagner. Towards evaluating the robustness of neural networks. IEEE Symposium on Security and Privacy, 2017. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 499, + 294, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 499, + 294, + 532 + ], + "spans": [ + { + "bbox": [ + 62, + 499, + 294, + 532 + ], + "type": "text", + "content": "[11] Y. Carmon, A. Raghunathan, L. Schmidt, P. Liang, and J. C. Duchi. Unlabeled data improves adversarial robustness. Neural Information Processing Systems, 2019. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 62, + 533, + 294, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 533, + 294, + 576 + ], + "spans": [ + { + "bbox": [ + 62, + 533, + 294, + 576 + ], + "type": "text", + "content": "[12] P.-Y. Chen, Y. Sharma, H. Zhang, J. Yi, and C.-J. Hsieh. Ead: Elastic-net attacks to deep neural networks via adversarial examples. AAAI Conference on Artificial Intelligence, 2018. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 62, + 578, + 294, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 578, + 294, + 622 + ], + "spans": [ + { + "bbox": [ + 62, + 578, + 294, + 622 + ], + "type": "text", + "content": "[13] T. Chen, M. Xu, X. Hui, H. Wu, and L. Lin. Learning semantic-specific graph representation for multi-label image recognition. IEEE International Conference on Computer Vision, 2019. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 62, + 624, + 294, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 624, + 294, + 667 + ], + "spans": [ + { + "bbox": [ + 62, + 624, + 294, + 667 + ], + "type": "text", + "content": "[14] Zhao-Min Chen, Xiu-Shen Wei, Xin Jin, and Yanwen Guo. Multi-label image recognition with joint class-aware map disentangling and label correlation embedding. IEEE International Conference on Multimedia and Expo, 2019. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 62, + 669, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 669, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 62, + 669, + 294, + 712 + ], + "type": "text", + "content": "[15] Z. M. Chen, X. S. Wei, P. Wang, and Y. Guo. Multi-label image recognition with graph convolutional networks. IEEE Conference on Computer Vision and Pattern Recognition, abs/1904.03582, 2019. 1" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 320, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 320, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 320, + 73, + 553, + 116 + ], + "type": "text", + "content": "[16] T. S. Chua, J. Tang, R. Hong, H. Li, Z. Luo, and Y. T. Zheng. Nus-wide: A real-world web image database from national university ofsingapore.ACM International Conference on Image and Video Retrieval, 2009.5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 321, + 118, + 553, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 118, + 553, + 148 + ], + "spans": [ + { + "bbox": [ + 321, + 118, + 553, + 148 + ], + "type": "text", + "content": "[17] S. D. Dao, E. Zhao, D. Phung, and J. Cai. Multi-label image classification with contrastive learning. arXiv preprint, arXiv:2107.11626, 2021. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 321, + 150, + 553, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 150, + 553, + 194 + ], + "spans": [ + { + "bbox": [ + 321, + 150, + 553, + 194 + ], + "type": "text", + "content": "[18] J. Deng, N. Ding, Y. Jia, A. Frome, K. Murphy, S. Bengio, Y. Li, H. Neven, and H. Adam. Large-scale object classification using label relation graphs. European Conference on Computer Vision, 2014. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 321, + 195, + 553, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 195, + 553, + 237 + ], + "spans": [ + { + "bbox": [ + 321, + 195, + 553, + 237 + ], + "type": "text", + "content": "[19] Gavin Weiguang Ding, Yash Sharma, Kry Yik Chau Lui, and Ruitong Huang. Mma training: Direct input space margin maximization through adversarial training. arXiv preprint arXiv:1812.02637, 2018. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 321, + 239, + 553, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 239, + 553, + 282 + ], + "spans": [ + { + "bbox": [ + 321, + 239, + 553, + 282 + ], + "type": "text", + "content": "[20] Zixuan Ding, Ao Wang, Hui Chen, Qiang Zhang, Pengzhang Liu, Yongjun Bao, Weipeng Yan, and Jungong Han. Exploring structured semantic prior for multi label recognition with incomplete labels. 2023. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 321, + 283, + 553, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 283, + 553, + 348 + ], + "spans": [ + { + "bbox": [ + 321, + 283, + 553, + 348 + ], + "type": "text", + "content": "[21] Junhao Dong, Seyed-Mohsen Moosavi-Dezfooli, Jianhuang Lai, and Xiaohua Xie. The enemy of my enemy is my friend: Exploring inverse adversaries for improving adversarial training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 24678–24687, 2023. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 321, + 350, + 553, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 350, + 553, + 392 + ], + "spans": [ + { + "bbox": [ + 321, + 350, + 553, + 392 + ], + "type": "text", + "content": "[22] Yinpeng Dong, Zhijie Deng, Tianyu Pang, Jun Zhu, and Hang Su. Adversarial distributional training for robust deep learning. Advances in Neural Information Processing Systems, 33:8270-8283, 2020. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 321, + 393, + 553, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 393, + 553, + 424 + ], + "spans": [ + { + "bbox": [ + 321, + 393, + 553, + 424 + ], + "type": "text", + "content": "[23] L. Feng, B. An, and S. He. Collaboration based multi-label learning. AAAI Conference on Artificial Intelligence, 2019. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 321, + 426, + 553, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 426, + 553, + 468 + ], + "spans": [ + { + "bbox": [ + 321, + 426, + 553, + 468 + ], + "type": "text", + "content": "[24] Zongyuan Ge, Dwarikanath Mahapatra, Suman Sedai, Rahul Garnavi, and Rajib Chakravorty. Chest x-rays classification: A multi-label and fine-grained problem. arXiv, 2018. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 321, + 471, + 553, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 471, + 553, + 503 + ], + "spans": [ + { + "bbox": [ + 321, + 471, + 553, + 503 + ], + "type": "text", + "content": "[25] I. J. Goodfellow, J. Shlens, and C. Szegedy. Explaining and harnessing adversarial examples. International Conference on Learning Representations, 2015. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 321, + 504, + 553, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 504, + 553, + 557 + ], + "spans": [ + { + "bbox": [ + 321, + 504, + 553, + 557 + ], + "type": "text", + "content": "[26] Emily Hand, Carlos Castillo, and Rama Chellappa. Doing the best we can with what we have: Multi-label balancing with selective learning for attribute prediction. In Proceedings of the AAAI Conference on Artificial Intelligence, 2018. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 321, + 559, + 553, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 559, + 553, + 591 + ], + "spans": [ + { + "bbox": [ + 321, + 559, + 553, + 591 + ], + "type": "text", + "content": "[27] J. Hayes and G. Danezis. Learning universal adversarial perturbations with generative models. IEEE Security and Privacy Workshops, 2018. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 321, + 593, + 553, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 593, + 553, + 634 + ], + "spans": [ + { + "bbox": [ + 321, + 593, + 553, + 634 + ], + "type": "text", + "content": "[28] W. He, J. Wei, X. Chen, N. Carlini, and D. Song. Adversarial example defense: Ensembles of weak defenses are not strong. USENIX Workshop on Offensive Technologies, 2017. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 321, + 636, + 553, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 636, + 553, + 658 + ], + "spans": [ + { + "bbox": [ + 321, + 636, + 553, + 658 + ], + "type": "text", + "content": "[29] D. Hendrycks, K. Lee, and M. Mazeika. Using pre-training can improve model robustness and uncertainty. 2019. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 321, + 659, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 659, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 321, + 659, + 553, + 712 + ], + "type": "text", + "content": "[30] Dan Hendrycks, Kevin Zhao, Steven Basart, Jacob Steinhardt, and Dawn Song. Natural adversarial examples. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15262-15271, 2021. 1" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20588" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 61, + 72, + 297, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 61, + 72, + 297, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 72, + 297, + 139 + ], + "spans": [ + { + "bbox": [ + 61, + 72, + 297, + 139 + ], + "type": "text", + "content": "[31] Lei Hsiung, Yun-Yun Tsai, Pin-Yu Chen, and Tsung-Yi Ho. Towards compositional adversarial robustness: Generalizing adversarial training to composite semantic perturbations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24658-24667, 2023. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 140, + 296, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 140, + 296, + 184 + ], + "spans": [ + { + "bbox": [ + 61, + 140, + 296, + 184 + ], + "type": "text", + "content": "[32] Shu Hu, Lipeng Ke, Xin Wang, and Siwei Lyu. Tkmlap: Adversarial attacks to top-k multi-label learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7649-7657, 2021. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 185, + 296, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 185, + 296, + 228 + ], + "spans": [ + { + "bbox": [ + 62, + 185, + 296, + 228 + ], + "type": "text", + "content": "[33] D. Huynh and E. Elhamifar. Interaction compass: Multi-label zero-shot learning of human-object interactions via spatial relations. International Conference on Computer Vision, 2021. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 229, + 296, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 229, + 296, + 262 + ], + "spans": [ + { + "bbox": [ + 62, + 229, + 296, + 262 + ], + "type": "text", + "content": "[34] D. T. Huynh and E. Elhamifar. Interactive multi-label cnn learning with partial labels. IEEE Conference on Computer Vision and Pattern Recognition, 2020. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 264, + 295, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 264, + 295, + 308 + ], + "spans": [ + { + "bbox": [ + 62, + 264, + 295, + 308 + ], + "type": "text", + "content": "[35] Tooba Imtiaz, Morgan Kohler, Jared Miller, Zifeng Wang, Mario Sznaier, Octavia I Camps, and Jennifer G Dy. Saif: Sparse adversarial and interpretable attack framework. arXiv preprint arXiv:2212.07495, 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 308, + 295, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 308, + 295, + 342 + ], + "spans": [ + { + "bbox": [ + 62, + 308, + 295, + 342 + ], + "type": "text", + "content": "[36] J. Li R. Ji, H. Liu, X. Hong, Y. Gao, and Q. Tian. Universal perturbation attack against image retrieval. International Conference on Computer Vision, 2019. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 342, + 295, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 342, + 295, + 387 + ], + "spans": [ + { + "bbox": [ + 62, + 342, + 295, + 387 + ], + "type": "text", + "content": "[37] Jinyuan Jia, Wenjie Qu, and Neil Zhenqiang Gong. Multiguard: Provably robust multi-label classification against adversarial examples. Advances in Neural Information Processing Systems, 2022. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 388, + 295, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 388, + 295, + 421 + ], + "spans": [ + { + "bbox": [ + 62, + 388, + 295, + 421 + ], + "type": "text", + "content": "[38] V. Khrulkov and I. Oseledets. Art of singular vectors and universal adversarial perturbations. IEEE Conference on Computer Vision and Pattern Recognition, 2018. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 422, + 295, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 422, + 295, + 454 + ], + "spans": [ + { + "bbox": [ + 62, + 422, + 295, + 454 + ], + "type": "text", + "content": "[39] Youngwook Kim, Jae Myung Kim, Zeynep Akata, and Jungwoo Lee. Large loss matters in weakly supervised multi-label classification. 2022. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 456, + 282, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 456, + 282, + 466 + ], + "spans": [ + { + "bbox": [ + 62, + 456, + 282, + 466 + ], + "type": "text", + "content": "[40] Takumi Kobayashi. Two-way multi-label loss. 2023. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 468, + 295, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 468, + 295, + 500 + ], + "spans": [ + { + "bbox": [ + 62, + 468, + 295, + 500 + ], + "type": "text", + "content": "[41] Linghao Kong, Wenjian Luo, Hongwei Zhang, Yang Liu, and Yuhui Shi. Evolutionary multilabel adversarial examples: An effective black-box attack. 2023. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 501, + 295, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 501, + 295, + 534 + ], + "spans": [ + { + "bbox": [ + 62, + 501, + 295, + 534 + ], + "type": "text", + "content": "[42] Linghao Kong, Wenjian Luo, Zipeng Ye, Qi Zhou, and Yan Jia. Multi-label black-box adversarial attacks only with predicted labels. 2024. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 62, + 535, + 295, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 535, + 295, + 567 + ], + "spans": [ + { + "bbox": [ + 62, + 535, + 295, + 567 + ], + "type": "text", + "content": "[43] A. Kurakin, I. Goodfellow, and S. Bengio. Adversarial machine learning at scale. International Conference on Learning Representations, 2017. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 62, + 568, + 295, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 568, + 295, + 635 + ], + "spans": [ + { + "bbox": [ + 62, + 568, + 295, + 635 + ], + "type": "text", + "content": "[44] A. Kuznetsova, H. Rom, N. Alldrin, J. Uijlings, I. Krasin, J. Pont-Tuset, S. Kamali, S. Popov, M. Malloci, A. Kolesnikov, T. Duerig, and V. Ferrari. The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale. International Journal of Computer Vision, 2016. 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 62, + 635, + 295, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 635, + 295, + 678 + ], + "spans": [ + { + "bbox": [ + 62, + 635, + 295, + 678 + ], + "type": "text", + "content": "[45] J. Lanchantin, T. Wang, V. Ordonez, and Y. Qi. General multi-label image classification with transformers. IEEE Conference on Computer Vision and Pattern Recognition, 2021. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 62, + 680, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 680, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 62, + 680, + 295, + 713 + ], + "type": "text", + "content": "[46] K. Lee, K. Lee, H. Lee, and J. Shin. A simple unified framework for detecting out-of-distribution samples and adversarial attacks. 2018. 2" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 320, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 320, + 73, + 553, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 73, + 553, + 129 + ], + "spans": [ + { + "bbox": [ + 320, + 73, + 553, + 129 + ], + "type": "text", + "content": "[47] Guofa Li, Zefeng Ji, Yunlong Chang, Shen Li, Xingda Qu, and Dongpu Cao. ML-anet: A transfer learning approach using adaptation network for multi-label image classification in autonomous driving. Chinese Journal of Mechanical Engineering, 2021. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 320, + 129, + 553, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 129, + 553, + 174 + ], + "spans": [ + { + "bbox": [ + 320, + 129, + 553, + 174 + ], + "type": "text", + "content": "[48] Maosen Li, Yanhua Yang, Kun Wei, Xu Yang, and Heng Huang. Learning universal adversarial perturbation by adversarial example. AAAI Conference on Artificial Intelligence, 2022. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 320, + 175, + 553, + 209 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 175, + 553, + 209 + ], + "spans": [ + { + "bbox": [ + 320, + 175, + 553, + 209 + ], + "type": "text", + "content": "[49] Peng Li, Peng Chen, Yonghong Xie, and Dezheng Zhang. Bi-modal learning with channel-wise attention for multi-label image classification. IEEE Access, 2020. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 320, + 209, + 553, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 209, + 553, + 243 + ], + "spans": [ + { + "bbox": [ + 320, + 209, + 553, + 243 + ], + "type": "text", + "content": "[50] Q. Li, M. Qiao, W. Bian, and D. Tao. Conditional graphical lasso for multi-label image classification. IEEE Conference on Computer Vision and Pattern Recognition, 2016. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 320, + 245, + 553, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 245, + 553, + 278 + ], + "spans": [ + { + "bbox": [ + 320, + 245, + 553, + 278 + ], + "type": "text", + "content": "[51] Q. Li, X. Peng, Y. Qiao, and Q. Peng. Learning label correlations for multi-label image recognition with graph networks. Pattern Recognition Letters, 2020. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 320, + 279, + 553, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 279, + 553, + 312 + ], + "spans": [ + { + "bbox": [ + 320, + 279, + 553, + 312 + ], + "type": "text", + "content": "[52] X. Li, F. Zhao, and Y. Guo. Multi-label image classification with a probabilistic label enhancement model. In UAI, 2014. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 320, + 314, + 553, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 314, + 553, + 347 + ], + "spans": [ + { + "bbox": [ + 320, + 314, + 553, + 347 + ], + "type": "text", + "content": "[53] Y. Li, Y. Song, and J. Luo. Improving pairwise ranking for multi-label image classification. IEEE Conference on Computer Vision and Pattern Recognition, 2017. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 320, + 349, + 553, + 383 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 349, + 553, + 383 + ], + "spans": [ + { + "bbox": [ + 320, + 349, + 553, + 383 + ], + "type": "text", + "content": "[54] Z. Li, W. Lu, Z. Sun, and W. Xing. Improving multi-label classification using scene cues. Multimedia Tools and Applications, 2017. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 320, + 384, + 553, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 384, + 553, + 407 + ], + "spans": [ + { + "bbox": [ + 320, + 384, + 553, + 407 + ], + "type": "text", + "content": "[55] Dekun Lin. Probability guided loss for long-tailed multi-label image classification. 2023. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 320, + 407, + 553, + 461 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 407, + 553, + 461 + ], + "spans": [ + { + "bbox": [ + 320, + 407, + 553, + 461 + ], + "type": "text", + "content": "[56] Tsung-Yi Lin, Michael Maire, Serge J. Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C. Lawrence Zitnick. Microsoft coco: Common objects in context. European Conference on Computer Vision, 2014. 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 320, + 464, + 553, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 464, + 553, + 508 + ], + "spans": [ + { + "bbox": [ + 320, + 464, + 553, + 508 + ], + "type": "text", + "content": "[57] An-An Liu, Zhuang Shao, Yongkang Wong, Junnan Li, Su Yu-Ting, and Mohan Kankanhalli. Lstm-based multi-label video event detection. Multimedia Tools and Applications, 78, 2019. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 320, + 510, + 553, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 510, + 553, + 544 + ], + "spans": [ + { + "bbox": [ + 320, + 510, + 553, + 544 + ], + "type": "text", + "content": "[58] Shilong Liu, Lei Zhang, Xiao Yang, Hang Su, and Jun Zhu. Query2label: A simple transformer way to multi-label classification. arXiv preprint arXiv:2107.10834, 2021. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 320, + 544, + 553, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 544, + 553, + 600 + ], + "spans": [ + { + "bbox": [ + 320, + 544, + 553, + 600 + ], + "type": "text", + "content": "[59] Xuannan Liu, Yaoyao Zhong, Yuhang Zhang, Lixiong Qin, and Weihong Deng. Enhancing generalization of universal adversarial perturbation through gradient aggregation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 2, 6" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 320, + 601, + 553, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 601, + 553, + 666 + ], + "spans": [ + { + "bbox": [ + 320, + 601, + 553, + 666 + ], + "type": "text", + "content": "[60] Yiran Liu, Xin Feng, Yunlong Wang, Wu Yang, and Di Ming. Trm-uap: Enhancing the transferability of data-free universal adversarial perturbation via truncated ratio maximization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4762-4771, 2023. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 320, + 669, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 669, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 320, + 669, + 553, + 713 + ], + "type": "text", + "content": "[61] Mingzhi Ma, Weijie Zheng, Wanli Lv, Lu Ren, Hang Su, and Zhaoxia Yin. Multi-label adversarial attack based on label correlation. In 2023 IEEE International Conference on Image Processing (ICIP), 2023. 3" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "20589" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 61, + 72, + 297, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 72, + 297, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 72, + 297, + 116 + ], + "spans": [ + { + "bbox": [ + 61, + 72, + 297, + 116 + ], + "type": "text", + "content": "[62] A. Madry, A. Makelov, L. Schmidt, D. Tsipras, and A. Vladu. Towards deep learning models resistant to adversarial attacks. International Conference on Learning Representations, 2018. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 118, + 295, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 118, + 295, + 162 + ], + "spans": [ + { + "bbox": [ + 61, + 118, + 295, + 162 + ], + "type": "text", + "content": "[63] Hassan Mahmood and Ehsan Elhamifar. Semantic-aware multi-label adversarial attacks. IEEE Conference on Computer Vision and Pattern Recognition, pages 24251-24262, 2024. 1, 3, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 163, + 295, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 163, + 295, + 195 + ], + "spans": [ + { + "bbox": [ + 62, + 163, + 295, + 195 + ], + "type": "text", + "content": "[64] S. Melacci, G. Ciravegna, A. Sotgiu, A. Demontis, B. Biggio, M. Gori, and F. Roli. Domain knowledge alleviates adversarial attacks in multi-label classifiers. 2021. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 197, + 295, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 197, + 295, + 240 + ], + "spans": [ + { + "bbox": [ + 62, + 197, + 295, + 240 + ], + "type": "text", + "content": "[65] J.-H. Metzen, M.-C. Kumar, T. Brox, and V. Fischer. Universal adversarial perturbations against semantic image segmentation. International Conference on Computer Vision, 2019. 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 241, + 295, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 241, + 295, + 285 + ], + "spans": [ + { + "bbox": [ + 62, + 241, + 295, + 285 + ], + "type": "text", + "content": "[66] S. Moosavi-Dezfooli, A. Fawzi, and P. Frossard. Deepfool: a simple and accurate method to fool deep neural networks. IEEE Conference on Computer Vision and Pattern Recognition, 2016. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 286, + 295, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 286, + 295, + 319 + ], + "spans": [ + { + "bbox": [ + 62, + 286, + 295, + 319 + ], + "type": "text", + "content": "[67] S. Moosavi-Dezfooli, A. Fawzi, O. Fawzi, and P. Frossard. Universal adversarial perturbations. IEEE Conference on Computer Vision and Pattern Recognition, 2017. 1, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 321, + 295, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 321, + 295, + 364 + ], + "spans": [ + { + "bbox": [ + 62, + 321, + 295, + 364 + ], + "type": "text", + "content": "[68] Konda Reddy Mopuri, Utsav Garg, and R Venkatesh Babu. Fast feature fool: A data independent approach to universal adversarial perturbations. arXiv preprint arXiv:1707.05572, 2017. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 365, + 295, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 365, + 295, + 410 + ], + "spans": [ + { + "bbox": [ + 62, + 365, + 295, + 410 + ], + "type": "text", + "content": "[69] Konda Reddy Mopuri, Utkarsh Ojha, Utsav Garg, and R Venkatesh Babu. Nag: Network for adversary generation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 742-751, 2018. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 411, + 295, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 411, + 295, + 454 + ], + "spans": [ + { + "bbox": [ + 62, + 411, + 295, + 454 + ], + "type": "text", + "content": "[70] J. Nam, E. L. Mencia, H. J. Kim, and J. Furnkranz. Maximizing subset accuracy with recurrent neural networks in multi-label classification. Neural Information Processing Systems, 2017. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 456, + 295, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 456, + 295, + 488 + ], + "spans": [ + { + "bbox": [ + 62, + 456, + 295, + 488 + ], + "type": "text", + "content": "[71] T. Pang, K. Xu, C. Du, N. Chen, and J. Zhu. Improving adversarial robustness via promoting ensemble diversity. International Conference on Machine learning, 2019. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 489, + 295, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 489, + 295, + 521 + ], + "spans": [ + { + "bbox": [ + 62, + 489, + 295, + 521 + ], + "type": "text", + "content": "[72] T. Pang, K. Xu, Y. Dong, C. Du, N. Chen, and J. Zhu. Rethinking softmax cross-entropy loss for adversarial robustness. arXiv, 2020." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 523, + 295, + 556 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 523, + 295, + 556 + ], + "spans": [ + { + "bbox": [ + 62, + 523, + 295, + 556 + ], + "type": "text", + "content": "[73] T. Pang, X. Yang, Y. Dong, K. Xu, H. Su, and J. Zhu. Boosting adversarial training with hypersphere embedding. Neural Information Processing Systems, 2020. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 62, + 557, + 295, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 557, + 295, + 612 + ], + "spans": [ + { + "bbox": [ + 62, + 557, + 295, + 612 + ], + "type": "text", + "content": "[74] Nicolas Papernot, Patrick Mcdaniel, Somesh Jha, Matt Fredrikson, Z. Berkay Celik, and Ananthram Swami. The limitations of deep learning in adversarial settings. IEEE European Symposium on Security and Privacy (EuroS&P), 2016. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 62, + 613, + 295, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 613, + 295, + 668 + ], + "spans": [ + { + "bbox": [ + 62, + 613, + 295, + 668 + ], + "type": "text", + "content": "[75] Ioannis Partalas, Aris Kosmopoulos, Nicolas Baskiotis, Thierry Artieres, George Paliouras, Eric Gaussier, Ion Androutsopoulos, Massih-Reza Amini, and Patrick Galinari. Lsjtc: A benchmark for large-scale text classification. arXiv preprint arXiv:1503.08581, 2015. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 62, + 670, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 670, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 62, + 670, + 295, + 713 + ], + "type": "text", + "content": "[76] Omid Poursaeed, Isay Katsman, Bicheng Gao, and Serge Belongie. Generative adversarial perturbations. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4422-4431, 2018. 2" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 320, + 73, + 554, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 320, + 73, + 554, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 73, + 554, + 116 + ], + "spans": [ + { + "bbox": [ + 320, + 73, + 554, + 116 + ], + "type": "text", + "content": "[77] Tao Pu, Tianshui Chen, Hefeng Wu, and Liang Lin. Semantic-aware representation blending for multi-label image recognition with partial labels. Proceedings of the AAAI Conference on Artificial Intelligence, 2022. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 320, + 118, + 554, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 118, + 554, + 171 + ], + "spans": [ + { + "bbox": [ + 320, + 118, + 554, + 171 + ], + "type": "text", + "content": "[78] Zeyu Qin, Yanbo Fan, Yi Liu, Li Shen, Yong Zhang, Jue Wang, and Baoyuan Wu. Boosting the transferability of adversarial attacks with reverse adversarial perturbation. Advances in Neural Information Processing Systems, 35: 29845-29858, 2022. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 320, + 172, + 554, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 172, + 554, + 228 + ], + "spans": [ + { + "bbox": [ + 320, + 172, + 554, + 228 + ], + "type": "text", + "content": "[79] Tal Ridnik, Emanuel Ben-Baruch, Nadav Zamir, Asaf Noy, Itamar Friedman, Matan Protter, and Lihi Zelnik-Manor. Asymmetric loss for multi-label classification. In Proceedings of the IEEE/CVF international conference on computer vision, pages 82-91, 2021. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 320, + 228, + 554, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 228, + 554, + 281 + ], + "spans": [ + { + "bbox": [ + 320, + 228, + 554, + 281 + ], + "type": "text", + "content": "[80] Tal Ridnik, Gilad Sharir, Avi Ben-Cohen, Emanuel Ben-Baruch, and Asaf Noy. ML-decoder: Scalable and versatile classification head. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 32-41, 2023. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 320, + 282, + 554, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 282, + 554, + 348 + ], + "spans": [ + { + "bbox": [ + 320, + 282, + 554, + 348 + ], + "type": "text", + "content": "[81] Jérôme Rony, Luiz G Hafemann, Luiz S Oliveira, Ismail Ben Ayed, Robert Sabourin, and Eric Granger. Decoupling direction and norm for efficient gradient-based 12 adversarial attacks and defenses. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4322-4330, 2019. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 320, + 349, + 554, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 349, + 554, + 381 + ], + "spans": [ + { + "bbox": [ + 320, + 349, + 554, + 381 + ], + "type": "text", + "content": "[82] J. Cohen and E. Rosenfeld and Z. Kolter. Certified adversarial robustness via randomized smoothing. International Conference on Machine learning, 2019. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 320, + 382, + 554, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 382, + 554, + 425 + ], + "spans": [ + { + "bbox": [ + 320, + 382, + 554, + 425 + ], + "type": "text", + "content": "[83] A. Shafahi, M. Najibi, A. Ghiasi, Z. Xu, J. Dickerson, C. Studer, L. Davis, G. Taylor, and T. Goldstein. Adversarial training for free! Neural Information Processing Systems, 2019." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 320, + 426, + 554, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 426, + 554, + 480 + ], + "spans": [ + { + "bbox": [ + 320, + 426, + 554, + 480 + ], + "type": "text", + "content": "[84] Nasim Shafiee and Ehsan Elhamifar. Zero-shot attribute attacks on fine-grained recognition models. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part V, pages 262-282. Springer, 2022. 1, 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 320, + 481, + 554, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 481, + 554, + 525 + ], + "spans": [ + { + "bbox": [ + 320, + 481, + 554, + 525 + ], + "type": "text", + "content": "[85] Jing Shao, Kai Kang, Chen Change Loy, and Xiaogang Wang. Deeply learned attributes for crowded scene understanding. In 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2015. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 320, + 525, + 554, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 525, + 554, + 568 + ], + "spans": [ + { + "bbox": [ + 320, + 525, + 554, + 568 + ], + "type": "text", + "content": "[86] Nitish Shukla and Sudipta Banerjee. Generating adversarial attacks in the latent space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 730-739, 2023. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 320, + 569, + 554, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 569, + 554, + 602 + ], + "spans": [ + { + "bbox": [ + 320, + 569, + 554, + 602 + ], + "type": "text", + "content": "[87] Q. Song, H. Jin, X. Huang, and X. Hu. Multi-label adversarial perturbations. IEEE International Conference on Data Mining, 2018. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 320, + 602, + 554, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 602, + 554, + 635 + ], + "spans": [ + { + "bbox": [ + 320, + 602, + 554, + 635 + ], + "type": "text", + "content": "[88] Fengguang Su, Ou Wu, and Weiyao Zhu. Multi-label adversarial attack with new measures and self-paced constraint weighting. 2024. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 320, + 636, + 554, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 636, + 554, + 679 + ], + "spans": [ + { + "bbox": [ + 320, + 636, + 554, + 679 + ], + "type": "text", + "content": "[89] C. Szegedy, W. Zaremba, I. Sutskever, J. Bruna, D. Erhan, I. Goodfellow, and R. Fergus. Intriguing properties of neural networks. International Conference on Learning Representations, 2014. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 320, + 680, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 680, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 320, + 680, + 554, + 713 + ], + "type": "text", + "content": "[90] G. Tsoumakas and I. Katakis. Multi-label classification: An overview. International Journal Data Warehousing and Mining, 3, 2007. 2" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "20590" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 57, + 73, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 73, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 73, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 61, + 73, + 294, + 95 + ], + "type": "text", + "content": "[91] N. Tursynbek, A. Petiushko, and I. Oseledets. Geometry-inspired top-k adversarial perturbations. arXiv, 2020. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 95, + 294, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 95, + 294, + 138 + ], + "spans": [ + { + "bbox": [ + 61, + 95, + 294, + 138 + ], + "type": "text", + "content": "[92] J. Uesato, J. B. Alayrac, P. Huang, R. Stanforth, A. Fawzi, and P. Kohli. Are labels required for improving adversarial robustness? Neural Information Processing Systems, 2019. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 140, + 295, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 140, + 295, + 183 + ], + "spans": [ + { + "bbox": [ + 62, + 140, + 295, + 183 + ], + "type": "text", + "content": "[93] Thomas Verelst, Paul K Rubenstein, Marcin Eichner, Tinne Tuytelaars, and Maxim Berman. Spatial consistency loss for training multi-label classifiers from single-label annotations. 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 184, + 295, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 184, + 295, + 228 + ], + "spans": [ + { + "bbox": [ + 62, + 184, + 295, + 228 + ], + "type": "text", + "content": "[94] J. Wang, Y. Yang, J. Mao, Z. Huang, C. Huang, and W. Xu. Cnn-rnn: A unified framework for multi-label image classification. 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 228, + 295, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 228, + 295, + 281 + ], + "spans": [ + { + "bbox": [ + 62, + 228, + 295, + 281 + ], + "type": "text", + "content": "[95] Ya Wang, Dongliang He, Fu Li, Xiang Long, Zhichao Zhou, Jinwen Ma, and Shilei Wen. Multi-label classification with label graph superimposing. In Proceedings of the AAAI conference on artificial intelligence, pages 12265-12272, 2020." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 283, + 295, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 283, + 295, + 326 + ], + "spans": [ + { + "bbox": [ + 62, + 283, + 295, + 326 + ], + "type": "text", + "content": "[96] Z. Wang, T. Chen, G. Li, G. Li, and L. Lin. Multi-label image recognition by recurrently discovering attentional regions. IEEE International Conference on Computer Vision, 2017. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 327, + 295, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 327, + 295, + 381 + ], + "spans": [ + { + "bbox": [ + 62, + 327, + 295, + 381 + ], + "type": "text", + "content": "[97] Yunchao Wei, Wei Xia, Min Lin, Junshi Huang, Bingbing Ni, Jian Dong, Yao Zhao, and Shuicheng Yan. Hcp: A flexible cnn framework for multi-label image classification. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2016. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 383, + 295, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 383, + 295, + 426 + ], + "spans": [ + { + "bbox": [ + 62, + 383, + 295, + 426 + ], + "type": "text", + "content": "[98] Y. Wu, H. Liu, S. Feng, Y. Jin, G. Lyu, and Z. Wu. Gmmlic: Graph matching based multi-label image classification. International Joint Conference on Artificial Intelligence, 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 426, + 295, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 426, + 295, + 470 + ], + "spans": [ + { + "bbox": [ + 62, + 426, + 295, + 470 + ], + "type": "text", + "content": "[99] C. Xie, Z. Zhang, Y. Zhou, S. Bai, J. Wang, Z. Ren, and A. Yuille. Improving transferability of adversarial examples with input diversity. IEEE Conference on Computer Vision and Pattern Recognition, 2019. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 57, + 471, + 295, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 471, + 295, + 514 + ], + "spans": [ + { + "bbox": [ + 57, + 471, + 295, + 514 + ], + "type": "text", + "content": "[100] Ming-Kun Xie, Jiahao Xiao, and Sheng-Jun Huang. Label-aware global consistency for multi-label learning with single positive labels. Advances in Neural Information Processing Systems, 35:18430-18441, 2022. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 515, + 295, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 515, + 295, + 548 + ], + "spans": [ + { + "bbox": [ + 57, + 515, + 295, + 548 + ], + "type": "text", + "content": "[101] J. Xu, H. Tian, Z. Wang, Y. Wang, W. Kang, and F. Chen. Joint input and output space learning for multi-label image classification. IEEE Transactions on Multimedia, 2020. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 548, + 295, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 548, + 295, + 581 + ], + "spans": [ + { + "bbox": [ + 57, + 548, + 295, + 581 + ], + "type": "text", + "content": "[102] W. Xu, D. Evans, and Y. Qi. Feature squeezing: Detecting adversarial examples in deep neural networks. Network and Distributed Systems Security Symposium, 2018. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 582, + 295, + 625 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 582, + 295, + 625 + ], + "spans": [ + { + "bbox": [ + 57, + 582, + 295, + 625 + ], + "type": "text", + "content": "[103] Zhuo Yang, Yufei Han, and Xiangliang Zhang. Characterizing the evasion attackability of multi-label classifiers. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 10647-10655, 2021. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 625, + 295, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 625, + 295, + 669 + ], + "spans": [ + { + "bbox": [ + 57, + 625, + 295, + 669 + ], + "type": "text", + "content": "[104] Zhuo Yang, Yufei Han, and Xiangliang Zhang. Attack transferability characterization for adversarially robust multi-label classification. In Machine Learning and Knowledge Discovery in Databases., pages 397-413, 2021. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "type": "text", + "content": "[105] V. O. Yazici, A. Gonzalez-Garcia, A. Ramisa, B. Twardowski, and J. v. d. Weijer. Orderless recurrent models for multi-label classification. European Conference on Computer Vision, 2020. 2" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 711 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "text", + "content": "[106] J. Ye, J. He, X. Peng, W. Wu, and Y. Qiao. Attention-driven dynamic graph convolutional network for multi-label image recognition. European Conference on Computer Vision, 2020." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 118, + 553, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 118, + 553, + 162 + ], + "spans": [ + { + "bbox": [ + 317, + 118, + 553, + 162 + ], + "type": "text", + "content": "[107] R. You, Z. Guo, L. Cui, X. Long, S. Y. Bao, and S. Wen. Cross-modality attention with semantic graph embedding for multi-label classification. AAAI Conference on Artificial Intelligence, 2020. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 163, + 553, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 163, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 317, + 163, + 553, + 217 + ], + "type": "text", + "content": "[108] Hong-Xing Yu, Wei-Shi Zheng, Ancong Wu, Xiaowei Guo, Shaogang Gong, and Jian-Huang Lai. Unsupervised person re-identification by soft multilabel learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2148–2157, 2019. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 219, + 553, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 219, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 316, + 219, + 553, + 262 + ], + "type": "text", + "content": "[109] Chaoning Zhang, Philipp Benz, Tooba Imtiaz, and In-So Kweon. Cd-uap: Class discriminative universal adversarial perturbation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 6754–6761, 2020. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 264, + 553, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 264, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 316, + 264, + 553, + 319 + ], + "type": "text", + "content": "[110] Chaoning Zhang, Philipp Benz, Tooba Imtiaz, and In So Kweon. Understanding adversarial examples from the mutual influence of images and perturbations. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 1, 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 319, + 553, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 319, + 553, + 374 + ], + "spans": [ + { + "bbox": [ + 316, + 319, + 553, + 374 + ], + "type": "text", + "content": "[111] Chaoning Zhang, Philipp Benz, Adil Karjauv, and In So Kweon. Universal adversarial perturbations through the lens of deep steganography: Towards a fourier perspective. In Proceedings of the AAAI conference on artificial intelligence, 2021. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 376, + 553, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 376, + 553, + 418 + ], + "spans": [ + { + "bbox": [ + 316, + 376, + 553, + 418 + ], + "type": "text", + "content": "[112] ML. Zhang and Z. Zhou. Multilabel neural networks with applications to functional genomics and text categorization. IEEE Transactions on Knowledge and Data Engineering, 2006. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 420, + 553, + 473 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 420, + 553, + 473 + ], + "spans": [ + { + "bbox": [ + 316, + 420, + 553, + 473 + ], + "type": "text", + "content": "[113] Shu Zhang, Ran Xu, Caiming Xiong, and Chetan Ramaiah. Use all the labels: A hierarchical multi-label contrastive learning framework. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 475, + 553, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 475, + 553, + 542 + ], + "spans": [ + { + "bbox": [ + 316, + 475, + 553, + 542 + ], + "type": "text", + "content": "[114] Zhe Zhao, Guangke Chen, Jingyi Wang, Yiwei Yang, Fu Song, and Jun Sun. Attack as defense: characterizing adversarial examples using robustness. In Proceedings of the 30th ACM SIGSOFT International Symposium on Software Testing and Analysis, page 42-55. Association for Computing Machinery, 2021. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 543, + 553, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 543, + 553, + 586 + ], + "spans": [ + { + "bbox": [ + 316, + 543, + 553, + 586 + ], + "type": "text", + "content": "[115] Donghao Zhou, Pengfei Chen, Qiong Wang, Guangyong Chen, and Pheng-Ann Heng. Acknowledging the unknown for multi-label learning with single positive labels. European Conference on Computer Vision, 2022. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 587, + 553, + 631 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 587, + 553, + 631 + ], + "spans": [ + { + "bbox": [ + 316, + 587, + 553, + 631 + ], + "type": "text", + "content": "[116] N. Zhou, W. Luo, X. Lin, P. Xu, and Z.. Zhang. Generating multi-label adversarial examples by linear programming. International Joint Conference on Neural Networks, 2020. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 632, + 553, + 676 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 632, + 553, + 676 + ], + "spans": [ + { + "bbox": [ + 316, + 632, + 553, + 676 + ], + "type": "text", + "content": "[117] N. Zhou, W. Luo, J. Zhang, L. Kong, and H. Zhang. Hiding all labels for multi-label images: An empirical study of adversarial examples. International Joint Conference on Neural Networks, 2021. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 677, + 553, + 711 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 677, + 553, + 711 + ], + "spans": [ + { + "bbox": [ + 316, + 677, + 553, + 711 + ], + "type": "text", + "content": "[118] Y. Zhu, J. T. Kwok, and Z. Zhou. Multi-label learning with global and local label correlation. IEEE Transactions on Knowledge and Data Engineering, 2018. 2" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20591" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Comprehensive Information Bottleneck for Unveiling Universal Attribution to Interpret Vision Transformers/abcc8427-3423-466d-afdc-b93969a43358_content_list.json b/2025/Comprehensive Information Bottleneck for Unveiling Universal Attribution to Interpret Vision Transformers/abcc8427-3423-466d-afdc-b93969a43358_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..253ccf8ce6a90c67a51655bd83b480f4546899f7 --- /dev/null +++ b/2025/Comprehensive Information Bottleneck for Unveiling Universal Attribution to Interpret Vision Transformers/abcc8427-3423-466d-afdc-b93969a43358_content_list.json @@ -0,0 +1,1853 @@ +[ + { + "type": "text", + "text": "Comprehensive Information Bottleneck for Unveiling Universal Attribution to Interpret Vision Transformers", + "text_level": 1, + "bbox": [ + 107, + 128, + 890, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jung-Ho Hong Ho-Joong Kim Kyu-Sung Jeon Seong-Whan Lee*", + "bbox": [ + 187, + 203, + 807, + 222 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Dept. of Artificial Intelligence, Korea University, Seoul, Korea", + "bbox": [ + 290, + 226, + 705, + 241 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{jungho-hong, hjoong_kim, ksjeon, sw.lee}@korea.ac.kr", + "bbox": [ + 259, + 243, + 732, + 258 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 285, + 326, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The feature attribution method reveals the contribution of input variables to the decision-making process to provide an attribution map for explanation. Existing methods grounded on the information bottleneck principle compute information in a specific layer to obtain attributions, compressing the features by injecting noise via a parametric damping ratio. However, the attribution obtained in a specific layer neglects evidence of the decision-making process distributed across layers. In this paper, we introduce a comprehensive information bottleneck (CoIBA), which discovers the relevant information in each targeted layer to explain the decision-making process. Our core idea is applying information bottleneck in multiple targeted layers to estimate the comprehensive information by sharing a parametric damping ratio across the layers. Leveraging this shared ratio complements the over-compressed information to discover the omitted clues of the decision by sharing the relevant information across the targeted layers. We suggest the variational approach to fairly reflect the relevant information of each layer by upper bounding layer-wise information. Therefore, CoIBA guarantees that the discarded activation is unnecessary in every targeted layer to make a decision. The extensive experimental results demonstrate the enhancement in faithfulness of the feature attributions provided by CoIBA.", + "bbox": [ + 88, + 316, + 485, + 696 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 713, + 220, + 728 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Vision transformer (ViT) achieves remarkable performance in diverse fundamental computer vision tasks, such as multi-modal [24, 36] and self-supervised learning [5, 12]. Despite this achievement, the black box nature stemming from the complex structure of ViT restricts the testability of the end-to-end system, impeding the failure diagnosis [10]. This limitation constrains the application of ViT in safety-critical areas where transparency and interpretability are considered important as well [22, 26]. Feature attribu", + "bbox": [ + 89, + 739, + 483, + 876 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/46452c1e53fc94d13bba0dc97d43689415fcc1dd0fbd03415ecc6d177ee52654.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 524, + 282, + 898, + 411 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6d86f523b511c3b6c21b19a26e75c1bb0918d32013203d3ab82e0eacb845f0e0.jpg", + "image_caption": [ + "(b) Cumulative number of each layer yielding optimal result per sample" + ], + "image_footnote": [], + "bbox": [ + 527, + 412, + 705, + 511 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d4ae8871e3d459b793874d5beb169e432c1db15051852e1b8b9cf6a3c7380eaa.jpg", + "image_caption": [ + "Figure 1. The quantitative and qualitative comparisons of attribution maps generated from different layers. We illustrate the visualized attribution maps from the first to fourth columns in Fig. 1(a) utilizing IBA [27]. The text on the top of each figure indicates the index of the layer. We select the layer index of the large model with interval 2. The fifth column in Fig. 1(a) shows the similarity comparison between attributions of different layers. Fig. 1(b) illustrates the cumulative number of each layer that provides the best explanation per sample. The uniformly distributed data indicates the absence of the layer optimal for the explanation." + ], + "image_footnote": [], + "bbox": [ + 709, + 412, + 888, + 511 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tion methods [4, 28, 30] are introduced to reveal the evidence supporting the decision-making process. Many of the existing attribution methods designed to interpret convolutional neural networks show limited explainability to interpret ViT [7]. To address this limitation, feature attribution methods [3, 6-8, 35] tailored for interpreting ViT have been proposed, leveraging either attention weights computed in ViT or intermediate representations, such as the class token, to generate attribution maps.", + "bbox": [ + 511, + 672, + 906, + 808 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Although feature attribution methods have advanced in interpreting ViT, they lack theoretical guarantees that low-scored attributions are not necessary to the decision-making process [27]. The methods [27, 37] grounded on the information bottleneck principle [2, 31] address this limitation by guaranteeing the importance of attributions with a", + "bbox": [ + 511, + 810, + 908, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation.", + "bbox": [ + 236, + 0, + 810, + 18 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "Except for this watermark, it is identical to the accepted version;", + "bbox": [ + 323, + 16, + 723, + 30 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 292, + 31, + 754, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author", + "bbox": [ + 114, + 887, + 240, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "25166", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "variational approximation. Information bottleneck for attribution (IBA) [27] proposes the model-agnostic method to compress the information unnecessary for the prediction by positioning the bottleneck layer into the target layer. InputIBA [37] succeeds in providing a high-resolution attribution map by directly measuring the information in the input domain, assuming the prior distribution with a generative model. However, existing information bottleneck-based methods face two main challenges. First, the resulting attribution map reflects the partial of the decision-making process as the restriction of the information flow is conducted in a specific targeted layer. This constraint provides inconsistent evidence for a single decision-making process, leading to unobvious interpretations. Lastly, no layer dominantly provides the most appropriate relevant information for explaining the decision-making process.", + "bbox": [ + 89, + 90, + 480, + 333 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Fig. 1 shows the aforementioned issues with quantitative and qualitative examples. In the first row in Fig. 1(a), the visualization derived from the second layer (blocks.2) highlights the barrel whereas the fifth (blocks.5) and eighth layers (blocks.8) highlight the top and bottom of the wheels, respectively. In the second row in Fig. 1(a), the visualization in the earlier layer highlights the person holding the racket, while the deeper layer progressively concentrates on the racket. To confirm the layer-wise discrepancy in attributions quantitatively, we measure SSIM [34] in the right row of Fig. 1(a). As the layer providing attributions distance farther, the dissimilarity in those attributions increases. Thus, computing the relevant information from isolated layer-specific information bottlenecks highlights different attributions for the same decision. Among the layers producing the attribution maps, there is no dominant layer that provides the most appropriate relevant information to explain the decision-making process. Fig. 1(b) compares the cumulative number of each layer yielding attributions with the optimal insertion and deletion scores per input sample. The results show a lack of the layer that produces the most appropriate relevant information for explanation. Thus, identifying the most faithful attribution from the layers requires a heuristic search due to the lack of criteria to distinguish the optimal layer.", + "bbox": [ + 91, + 337, + 482, + 715 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we propose a novel comprehensive information bottleneck for attribution (CoIBA), which reveals the information relevant for each of the target layers to provide the attributions. Our core idea is to reveal comprehensive relevant information by eliminating unnecessary information in every targeted layer of the prediction. We estimate the comprehensive information by sharing a parameterized universal damping ratio across layers, thereby removing the requirements of additional heuristics. This sharing strategy compensates for the over-compressed information of individual layers by sharing information in each layer necessary for the decision. Since a channel el", + "bbox": [ + 89, + 719, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ement placed in the same location but different layers captures different features, we uniformly perturb all the channel elements in a single token representation, to handle the information of different layers with a shared parametric ratio. We suggest a variational upper bound to restrict information flowing in all the targeted layers. Our approach eliminates the heuristic search for balancing inconsistent information along the layers by suggesting a variational upper bound to fairly reflect the layer-wise relevant information.", + "bbox": [ + 511, + 90, + 903, + 226 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We conduct experiments to show the correctness performance of CoIBA utilizing various assessments. As the evaluations are solely approximations, i.e. no ground truth exists, we include FunnyBirds [15], which provide the ground truth and assess the comprehensive quality of attribution maps. To assess the feature importance assessment, we leverage insertion/deletion [23] and remove-and-debias (ROAD) [25], which evaluate the faithfulness of attribution maps. In addition, we scrutinize the faithfulness of CoIBA by analyzing the confident-aware assessment. We discuss the effectiveness of our method with a rigorous discussion.", + "bbox": [ + 511, + 227, + 903, + 393 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 511, + 407, + 661, + 422 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Explanation Methods for Vision Transformer Existing explanatory methods designed to interpret ViT provide the attribution map utilizing the latent representation and the corresponding gradients. However, these methods show limited explainability while being adapted to ViT [7]. To address this limitation, Rollout [1] linearly combines the attention weights across the layers to compute the attribution map. Trans-attr [7] provides the class-discriminative attribution map by constructing the layer-wise relevance propagation rule for ViT. Generic [6] produces the attribution map with a generalized procedure to interpret diversified transformer models, leveraging the gradient of the attention map to produce the class-discriminative attribution maps. IIA [3] introduces the iterative integration across the input image, leveraging the internal representations processed by the model and their gradients. ViT-CX [35] utilizes the patch embeddings and measures causal impacts to provide attributions. Beyond [8] unfolds attention blocks with the chain rule between final prediction, CLS, and tokens to obtain token contribution. Existing propagation-based approaches require a specific implementation to weigh the contributions of input variables. In contrast to this constraint, CoIBA does not require implementing additional procedures as same as IBA. Furthermore, aligning with the information bottleneck principle, CoIBA guarantees the highlighted attributions are important in all targeted layers for the decision.", + "bbox": [ + 511, + 431, + 903, + 824 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Information Bottleneck Approach The information bottleneck principle is broadly utilized to obtain information in the activation necessary for the specific objective. To this end, the information bottleneck-based methods compress the information unnecessary for the objective while", + "bbox": [ + 511, + 825, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "25167", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/04e84f49a950cd4233f5fe0f6451bc1cf5b7ed1e0f1c422923f55be756b332a8.jpg", + "image_caption": [ + "Figure 2. Main difference between IBA and CoIBA. IBA obtains the relevant information in the specific bottleneck-inserted layer. Thus, obtaining layer-wise attribution maps necessarily requires iterative running over the layers. In contrast, CoIBA seeks the layer-wise relevant information from each bottleneck-inserted layer with a universal damping ratio." + ], + "image_footnote": [], + "bbox": [ + 96, + 88, + 480, + 328 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "maintaining the relevant information. Given the challenges of directly estimating the amount of information, a variational approach [18] is used to approximate the intractable posteriors and marginals. The Deep Variational Information Bottleneck [2] method employs the information bottleneck principle in deep networks, using a variational lower bound alongside the reparameterization trick. Referring to this strategy, IBA [27] adapts the information bottleneck principles to restrict the information flow in an intermediate layer and provides the relevant information as an attribution map. InputIBA [37] provides high-resolution attributions by directly estimating relevant information in the input domain. However, existing methods compute relevant information solely considering a specific layer, overlooking the information necessary in the different layers. In contrast to this process, CoIBA obtains comprehensive relevant information from the multiple layers, highlighting the evidence important in every targeted layer to make a decision.", + "bbox": [ + 88, + 439, + 483, + 712 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 734, + 181, + 750 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The main goal of CoIBA is to produce an attribution map by reflecting the information relevant to the targeted layers, which is opposed to the IBA as shown in Fig. 2. To show this procedure, we first introduce IBA in Sec. 3.1 and discuss the limitations of the existing information bottleneck-based approach in Sec. 3.2. Then, in Sec. 3.3, we describe the overall procedure of CoIBA. Finally, we introduce a variational upper bound, which enhances the reflection of layer-wise relevant information in Sec. 3.4.", + "bbox": [ + 89, + 763, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/ea662166572a569ff49afc338896d996e47bafc11a01d53f5718baed5727779c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 535, + 88, + 705, + 193 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/a1b5c281fd2885172ccf4b287eb5aee8be10200195a305605b0fdfd1f631a493.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 709, + 89, + 883, + 193 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/e44fd811d68c2e9f304fa660ae85e26f7f7aaa706c1009dfa03b1937adafd59f.jpg", + "image_caption": [ + "Figure 3. Quantitative and qualitative results of IBA*. In the first row, the horizontal line indicates the performance of layers-specific relevant information, i.e., IBA. We report the difference in insertion and deletion scores ( $\\Delta$ InsDel). The higher score is better for this metric. The visualized attribution maps are illustrated in the second row obtained from VIT-B-16/224 [11]." + ], + "image_footnote": [], + "bbox": [ + 537, + 196, + 633, + 270 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/c5cb7cb051d773a4a7eb0fd1f0472968941d350393a39855bbb2059cdcce1296.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 633, + 196, + 715, + 270 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/673112c63ece870c5c754456fdf2df0cb39ce5a23b48b8a32fe0c6f59646ea08.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 715, + 196, + 799, + 270 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/7729a9370a5731c56cff819cbc8a3cf491d64c67c614d0774d4687202778ac19.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 800, + 196, + 883, + 270 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Background - IBA", + "text_level": 1, + "bbox": [ + 511, + 364, + 692, + 381 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "IBA adopts an information bottleneck principle to produce relevant information to provide an attribution map. To this end, in $l$ -th layer, IBA computes the bottleneck representation $Z_{l}$ as follows:", + "bbox": [ + 511, + 387, + 905, + 446 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nZ _ {l} = \\lambda_ {l} R _ {l} + (1 - \\lambda_ {l}) \\epsilon_ {l}. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 624, + 460, + 903, + 477 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Here, the damping ratio $\\lambda_{l}$ is a learnable parameter assigned at the bottleneck layer, manipulating the degree of damping signal between the activation $R_{l}$ and the independent noise $\\epsilon_{l}$ . The independent noise $\\epsilon_{l}$ is sampled from the Gaussian distribution. To maintain the statistic of internal representation, the Gaussian distribution shares the mean $\\mu_{R_l}$ and variance $\\sigma_{R_l}$ with the representation $R_{l}$ .", + "bbox": [ + 511, + 488, + 905, + 594 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "IBA minimizes the shared information between the activation $R_{l}$ and bottleneck variable $Z_{l}$ while maximizing the shared information between the bottleneck variable $Z_{l}$ and the label $Y$ as follows:", + "bbox": [ + 511, + 595, + 905, + 654 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\max _ {\\lambda_ {l}} I \\left[ Z _ {l}; Y \\right] - \\beta I \\left[ R _ {l}, Z _ {l} \\right]. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 612, + 667, + 903, + 691 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Here, $\\beta$ is a hyperparameter manipulating the trade-off between compression and relevancy in $l$ -th layer.", + "bbox": [ + 511, + 702, + 903, + 732 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Motivation", + "text_level": 1, + "bbox": [ + 511, + 742, + 633, + 757 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "IBA selects a specific layer to obtain relevant information to provide an attribution map. However, the relevant information of a specific single layer does not reflect the overall evidence required in the sequence of layers to make a decision. To demonstrate our insistence, we illustrate the quantitative comparison of the attribution maps obtained in a specific layer and multiple layers in Fig. 3, leveraging insertion/deletion [23]. We iterate IBA multiple times along the layers to obtain the set of individual relevant information", + "bbox": [ + 511, + 763, + 906, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "25168", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/84254202b7b2cdc9d21b3e58547095be842d63f0a1b84370e4a672b1a71abc71.jpg", + "image_caption": [ + "eaaee", + "Class: Hyena" + ], + "image_footnote": [], + "bbox": [ + 161, + 95, + 250, + 162 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/3f70ab71a742248c92c64b7aa2499ea94e3427761959d99761d4ca65ec3e6e05.jpg", + "image_caption": [ + "Confidence: 85.87" + ], + "image_footnote": [], + "bbox": [ + 254, + 95, + 433, + 189 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/210a74117a0c02bb86604554dd1a882f30f28c22a4e44fc636ce78f15699e53b.jpg", + "image_caption": [ + "Hard Sample", + "Class: Parallel Bars", + "Confidence: 19.71" + ], + "image_footnote": [], + "bbox": [ + 161, + 196, + 250, + 266 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/f04d2669558c4dcf691ac1608cc4d69796554afefbb331adb53880adc7b975ec.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 254, + 234, + 284, + 256 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/87dafdfb7dc66a24a577cb325728f387ff0d35db09075ba65f525afe39b9d800.jpg", + "image_caption": [ + "Figure 4. The quantitative comparisons of explanation methods in high-confident (easy) and low-confident (hard) input samples. The confidence score is the softmax probability of the target class predicted by the model. We compare the information-bottleneck-based methods including IBA [27], IBA*, and CoIBA (marked with a circle), and the propagation-based explanation baselines including Beyond [8], IIA [3], and Chefer-LRP [7] methods (marked with a triangle). We utilize DeiT-B-16/224 [32] for the evaluation. For insertion and deletion, higher and lower scores are better, respectively." + ], + "image_footnote": [], + "bbox": [ + 285, + 198, + 433, + 292 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "for each layer. After that, we compute a linear combination over the set of individual relevant information. We denote this procedure as $\\mathrm{IBA}^*$ . As shown in the results, the attribution maps produced by $\\mathrm{IBA}^*$ archive better scores compared to the IBA. To scrutinize the quantitative results between $\\mathrm{IBA}^*$ and IBA, we split the insertion/deletion scores by the confidence scores of the model in Fig. 4. As shown in the results, the increase in performance is only observed in high-confident samples, unlike in low-confident samples. Therefore, even iterating IBA, which requires high computational cost, shows limited performance enhancement according to the difficulty of each sample.", + "bbox": [ + 89, + 446, + 483, + 628 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Comprehensive Information Bottleneck", + "text_level": 1, + "bbox": [ + 89, + 637, + 429, + 652 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our comprehensive information bottleneck (CoIBA) inserts the bottleneck layer into the sequence of targeted layers to restrict the information flowing in the intermediate layer. Within the bottleneck layer, CoIBA restricts the information with the learnable universal damping ratio shared across the bottleneck-inserted layers. Our goal is to optimize the universal damping ratio to obtain the layer-wise relevant information while fulfilling the information bottleneck objective. The universal damping ratio enables the attributions to identify distributed relevant information omitted by IBA, which restricts information in a layer-specific manner.", + "bbox": [ + 89, + 659, + 482, + 824 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Formulating Information Restriction The model is trained to make a decision $\\tilde{Y} = f(X)$ to predict the ground truth $Y$ by taking input $X$ . To make a decision, the input sample $X$ is contextualized by passing through the sequence of intermediate layers. Here, we insert the bottle-", + "bbox": [ + 89, + 825, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "neck layer into the targeted layer. The role of the bottleneck layer is to restrict the information by dampening the signal of the intermediate representation passing through the $L$ bottleneck-inserted layers. We perturb the intermediate representations in the $l$ -th layer to dampen the signal. To dampen the signal, we inject the independent noise $\\epsilon_{l}$ into the intermediate bottleneck representation $R_{l}^{\\prime}$ , which is computed from the bottleneck variables $Z_{l-1}$ of the preceding layer. Note that the first element of intermediate bottleneck representations is the non-perturbed intermediate representation, i.e., $R_{1}^{\\prime} = R_{1}$ . Thus, we obtain the bottleneck representation $Z_{l}$ as follows:", + "bbox": [ + 511, + 90, + 906, + 272 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nZ _ {l} = \\lambda R _ {l} ^ {\\prime} + (1 - \\lambda) \\epsilon_ {l}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 629, + 286, + 906, + 304 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here, a universal damping ratio $\\lambda \\in \\mathbb{R}^{P\\times 1}$ manipulates the degree of perturbation for each token, denoting $P$ as the number of patches. Thus, $\\lambda$ is ranged from 0 to 1, allowing the propagation of the signal when $\\lambda = 1$ otherwise blocking it when $\\lambda = 0$ . As the universal damping ratio $\\lambda$ is consistently adapted to the layers, we omit the layer index in the notation. We compute $\\lambda = \\mathrm{sigmoid}(\\alpha)$ by passing the trainable parameter $\\alpha$ , which is initialized with 5, to the sigmoid function. The universal damping ratio uniformly perturbs the signal along the channel dimension. This setting avoids handling the neurons that capture different features with the same coefficient and induces concentration on the token's importance. We empirically show this uniform perturbation across the channels enhances the faithfulness of resulting relevant information. The independent noise $\\epsilon_{l}$ shares the dimension and is sampled from the Gaussian distribution with the mean $\\mu_{R_l}$ and variance $\\sigma_{R_l}^2$ , such that $\\epsilon_{l}\\sim \\mathcal{N}(\\mu_{R_l},\\sigma_{R_l}^2)$ . To minimize statistical differences between the bottleneck variable and non-perturbed activations, we leverage the mean $\\mu_{R_l}$ and variance $\\sigma_{R_l}^2$ from those of the non-perturbed activations.", + "bbox": [ + 511, + 316, + 906, + 633 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Information Bottleneck Objective The information bottleneck principle desires to obtain information necessary for the objective while compressing irrelevant information from the signal. Compressing irrelevant information is performed by minimizing the shared information between the activation of a bottleneck-inserted network and the bottleneck representation. Building upon this, we compress the amount of information by minimizing the mutual information between internal sequences of bottleneck representations while maximizing the mutual information between the ground truth and bottleneck representations. Combining two aspects with the ground truth $Y$ , we maximize the mutual information between the bottleneck variable and the ground truth while minimizing the mutual information between bottleneck representations as follows:", + "bbox": [ + 511, + 635, + 908, + 862 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\max _ {\\lambda} I \\left[ Z _ {L}; Y \\right] - \\frac {1}{L} \\left(\\sum_ {l = 1} ^ {L} \\beta_ {l} I \\left[ Z _ {l - 1}; Z _ {l} \\right]\\right). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 555, + 875, + 906, + 905 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "25169", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/5329bcb88a94f9db5a85eaceb195c1588ff663fb1b0ce6f34d6c648df6887191.jpg", + "image_caption": [ + "Figure 5. An overview of CoIBA. CoIBA restricts the information flow using a universal damping ratio while upper-bounding the information of each layer. After that, CoIBA aggregates the relevant information from each layer to produce an attribution map." + ], + "image_footnote": [], + "bbox": [ + 133, + 88, + 867, + 305 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here, the trade-off parameter $\\beta_{l}$ governs the degree of compression in each layer. Since the bottleneck layers are inserted into the sequence of layers, the compression term includes the accumulated sequence of mutual information of bottleneck variables $I[Z_{l-1};Z_l]$ . Here, the $Z_0$ indicates the original intermediate representation as the first element, such that $Z_0 = R_1$ . We maximize the relevancy term $I[Z_L;Y]$ while compression to maintain the shared information between bottleneck variables and the ground truth. We maximize this term by minimizing the cross-entropy loss $\\mathcal{L}_{ce}$ .", + "bbox": [ + 89, + 343, + 483, + 507 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Utilizing KL divergence, the mutual information between internal representations $I[Z_{l-1}; Z_l]$ is formulated by:", + "bbox": [ + 89, + 508, + 483, + 541 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nI \\left[ Z _ {l - 1}; Z _ {l} \\right] = \\mathbb {E} _ {Z _ {l - 1}} \\left[ D _ {K L} \\left[ P \\left(Z _ {l} \\mid Z _ {l - 1}\\right) \\right\\vert \\left| P \\left(Z _ {l}\\right) \\right] \\right] \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 114, + 560, + 482, + 579 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "However, the direct estimation of the prior distribution $P(Z_{l})$ is intractable as it requires the integration of bottleneck representations $P(Z_{l}) = \\int_{Z_{l - 1}}P(Z_{l}|Z_{l - 1})P(Z_{l - 1})dZ_{l - 1}$ of the corresponding $l$ -th layer. Thus, we assume the prior distribution $Q(Z_{l})$ as Gaussian distribution $\\mathcal{N}(\\mu_{R_l},\\sigma_{R_l}^2)$ , reconstituting the mutual information computation as follows:", + "bbox": [ + 89, + 584, + 483, + 690 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nI \\left[ Z _ {l}; Z _ {l - 1} \\right] \\leq \\mathbb {E} _ {Z _ {l - 1}} \\left[ D _ {K L} \\left[ P \\left(Z _ {l} \\mid Z _ {l - 1}\\right) \\right\\vert \\left| Q \\left(Z _ {l}\\right) \\right] \\right]. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 112, + 696, + 482, + 715 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Since assuming the prior distribution with Gaussian distribution only overestimates the mutual information, minimizing this term suppresses the mutual information of the compression term.", + "bbox": [ + 89, + 719, + 483, + 780 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Variational Upper Bound", + "text_level": 1, + "bbox": [ + 89, + 787, + 321, + 804 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As discussed in Sec. 4.6, the relevant information obtained by suppressing layer-wise mutual information suffers from reflecting layer-wise relevant information due to the inconsistent amount of information along the layers. Thus, obtaining the sequence of trade-off parameters $\\{\\beta_l\\}_{l = 1}^L$ requires an extensive heuristic search to balance the degree", + "bbox": [ + 89, + 809, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "of compression for each layer. To overcome this issue, we suggest the variational upper bound for CoIBA, referring to the two following points. First, since there is no additional information propagated during the forwarding pass, the mutual information of intermediate layers would not be greater than the mutual information between the model input and output. Second, the iterated noise-injection procedure diminishes the mutual information among the internal and bottleneck representations, i.e., $I[Z_{l};Z_{l + 1}] \\leq I[Z_{l - 1};Z_l]$ . Building upon these bases, we establish an upper bound that encompasses the combined mutual information from all subsequent layers as follows:", + "bbox": [ + 511, + 343, + 906, + 523 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nI \\left[ R _ {1}; Z _ {1} \\right] \\geq \\frac {1}{L} \\left(\\sum_ {l = 1} ^ {L} I \\left[ Z _ {l - 1}; Z _ {l} \\right]\\right). \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 545, + 906, + 574 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Accordingly, we reconstitute the objective to be a simplified formula, leaving only a single hyperparameter $\\beta$ as follows:", + "bbox": [ + 511, + 594, + 905, + 625 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\max _ {\\lambda} I \\left[ Z _ {L}; Y \\right] - \\beta I \\left[ R _ {1}; Z _ {1} \\right]. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 609, + 660, + 903, + 683 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To compress the subsequent layers, the simplified objective necessitates calculating only the mutual information of the first layer $I[R_1;Z_1]$ . This term solely utilizes the non-perturbed intermediate and bottleneck representations of the first layer, simplifying the objective calculation. In terms of relevant information computation, referring to the data processing inequality, the inequality $I[Y;Z_l]\\leq I[Y;Z_{l - 1}]$ holds. Thereby, in contrast to the layer-specific information bottleneck, which overestimates mutual information in the earlier layers during compression [37], CoIBA relieves the overestimation. This is because, as discussed in Sec. 4.6, the relevant information is compensated as the sequence of bottleneck variables joins the objective computation.", + "bbox": [ + 511, + 704, + 906, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "25170", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/0ba7fcfd7dfa2e1358c8ffea180b50930af878b8e1d18f727a621b398f5f047d.jpg", + "image_caption": [ + "Original Image" + ], + "image_footnote": [], + "bbox": [ + 102, + 135, + 189, + 200 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5dedc72a277d60fc8f56dca2139d50e1956c2f02f5427820d8573d1a3feec967.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 200, + 90, + 290, + 165 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/7df66507cdc9f2b5db097cb5bedcb50b10da014aa82e1adf00d3bf2d7dacec33.jpg", + "image_caption": [ + "Figure 6. Quantitative comprehensive assessment on FunnyBirds experiment: The abbreviations Com, Cor, and Con indicate Completeness, correctness, and contrastivity. The center score indicates the comprehensive result of enumerated aspects. Acc. and B.I. indicate accuracy and background independence, respectively. We provide detailed numeric in the supplementary material." + ], + "image_footnote": [], + "bbox": [ + 196, + 167, + 295, + 244 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/40447f6b45fbf2d809da226459508b189ce1ac6933ff8e11da8d4cde6378a025.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 299, + 90, + 387, + 165 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b57b480c4902f9a344f5915bba0b67219765a4f14ba1714da3dd677bb372b8d1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 295, + 167, + 393, + 244 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/892b1921c46b9d374b3387ad2068b6224b3f25267fdc6d88fbf1a7026afc930b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 398, + 90, + 485, + 165 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/bbe59162935a3cd7e403b8fefc9c607cdddf12c8ba9318e7761230279542e123.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 395, + 167, + 495, + 244 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c79670731198fe2b38be3294e6fbad593bbc07532e6b65023c3d72b0230be279.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 496, + 90, + 584, + 165 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e9269a4742934d318b2423986857a620cc128cf159f584edd8f09ec67d2ca5a4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 491, + 167, + 589, + 244 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/57ffe9db5e70f8ab8c5e9cb05ce22613b272691982b7cf62f77e0123fdfa8f7a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 591, + 90, + 679, + 165 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ccde3b81382f1229c3d4e2edeb1bc7715770bf0d6667ad4d4b5bb270d4c46d24.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 591, + 167, + 687, + 244 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c22718f40152dac6e2c8674187f79fc659bb1de71e3ba08f64c6142b1026d3be.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 689, + 90, + 779, + 165 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/6ac53393064b023c4ed9766f675f34fb897f4dbf1f4a82d3394efb34c3f6620f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 687, + 167, + 787, + 244 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a3d3a2e0e3637f5938f4e9a4af2bf8d23afb07b0d4f5e17a342716fbe36c8bde.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 790, + 90, + 877, + 165 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/9b6881e286591e9db51bf99ee5ee2220d828333ebdd128cd02aad2254639e28b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 785, + 167, + 888, + 244 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/5ad83bac8c93f53ab8a1d54c35fa4b1307e03ccb352d2a61e6fb7c876b50e462.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
VariantModelChefer-LRP [7]Generic [6]IIA [3]ViT-CX [35]IBA [27]Beyond [8]ColBA
ViTViT*-S-16/224-14.56/56.0813.69/57.3114.98/56.6513.56/58.6413.50/58.4711.08/63.35
ViT*-T-16/224-9.00/48.958.50/49.7512.08/45.098.02/49.2211.26/44.076.97/54.23
ViT-B-16/22417.67/56.6917.92/57.0316.56/58.716.94/58.1717.23/59.0615.84/60.3713.01/62.58
ViT-L-16/22420.97/53.8120.17/55.5420.14/55.2219.76/57.3820.88/54.8518.99/56.5515.40/61.14
DeiTDeiT-S-16/22412.76/47.5112.10/48.6410.89/49.7617.57/44.4311.43/47.8711.13/48.889.55/53.38
DeiT-B-16/22414.63/48.7915.10/49.4113.69/50.5118.21/46.9713.95/50.1814.45/48.8811.79/53.96
DeiT3-B-16/224-15.70/52.4215.38/52.7018.88/50.4015.07/53.3214.71/54.0712.97/56.54
DeiT3-L-16/224-22.10/62.7317.75/55.9326.11/59.6021.01/65.6019.34/64.2717.64/67.20
SwinSwin-B-33.76/42.94--18.03/52.8123.73/50.3917.09/54.80
Swin2-B-34.71/47.16--20.94/53.8531.32/49.4418.88/55.77
", + "bbox": [ + 133, + 297, + 859, + 439 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Quantitative feature importance assessment on insertion ↑/deletion ↓. (*) denotes the ViT trained with strong regularization techniques [29]. We underline the state-of-the-art performance among the baselines.", + "bbox": [ + 89, + 441, + 906, + 470 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 477, + 223, + 493 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Setup", + "text_level": 1, + "bbox": [ + 89, + 506, + 171, + 522 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Architectures We select variants of ViT for our experiment, including original ViT [11] and DeiT models [32, 33], trained with ImageNet-21k (IN-21k) and ImageNet-1k (IN-1k), respectively. We include the variants in depth interpolated from ViT-T, ViT-S, ViT-B, ViT-L, and ViT-H models. For the ViT-T and ViT-S models, we utilize the model pretrained with massive regularization techniques [29]. The ViT-H model is pre-trained with CLIP [24]. We present the settings of the ViT-B model with patch size 16 and image resolution 224 as ViT-B-16/224. We include Swin transformers [20, 21], demonstrating the generalized ability of CoIBA in multi-scale features processing. The details of settings and further results about convolutional neural networks are included in the supplementary material.", + "bbox": [ + 89, + 532, + 483, + 744 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Feature Attribution Methods We compare the various feature attribution methods including Chefer-LRP [7], Generic [6], IIA [3], ViT-CX [35], Beyond [8], and IBA [27] as baselines. We select $\\beta = 10$ for IBA and 6-th layer to insert bottleneck. For CoIBA, we select tradeoff hyper-parameter $\\beta$ as 1. We insert the bottleneck layer from $s$ -th departure to $e$ -th arrival layers, which are chosen, 4, and 12, respectively, for ViT-B. For the optimization, we set the learning rate, batch size, and optimizer as 1, 10, and Adam [17]. The additional hyperparameter settings", + "bbox": [ + 89, + 750, + 483, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "are included in the supplementary material. For Swin transformers, we only report the explanation methods that do not require methodology modification, including Generic, ViTCX, IBA, Beyond, and CoIBA. We utilize RTX A6000 GPU for all the experiments.", + "bbox": [ + 511, + 478, + 906, + 553 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets We utilize ImageNet-1k [9] (IN-1k), ImageNet-A (IN-A) [14] and ImageNet-R (IN-R) [13] validation datasets for the experiments. We leverage IN-A and IN-R to conduct the difficult-aware analysis.", + "bbox": [ + 511, + 554, + 905, + 614 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. FunnyBirds Assessment", + "text_level": 1, + "bbox": [ + 511, + 622, + 735, + 638 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The FunnyBirds [15] experiment provides the ground truth for the evaluation. With the help of ground truth, the Funnybird framework assesses the quality of the attribution map from three perspectives: completeness, correctness, and contrastivity. Fig. 6 illustrates the qualitative and quantitative results of the Funnybirds experiment. As shown in the results, CoIBA outperforms all the baselines, including the propagation or gradient-based approaches. Furthermore, in contrast to IBA, CoIBA obtains an attribution map with significantly enhanced contrastivity.", + "bbox": [ + 511, + 643, + 906, + 795 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Insertion/Deletion", + "text_level": 1, + "bbox": [ + 511, + 803, + 689, + 816 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Insertion/deletion [23] measures the correctness of feature importance highlighted by an attribution map. The insertion/deletion method gradually inserts or deletes informative pixels in ascending order of attribution to compute the area under the curve scores. In insertion/deletion, methods", + "bbox": [ + 511, + 824, + 905, + 898 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "25171", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/e5c1ecb1c7ba21e05f5406a351fffe37ff0a739f41176bb087dadb1774354ce4.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelChefer-LRP [7]Generic [6]IIA [3]ViT-CX [35]IBA [27]Beyond [8]CoIBA
ViT-B-16/22420.87/64.3722.15/65.3119.58/66.1919.04/65.3620.99/68.8017.93/68.4416.63/73.68
DeiT-B-16/22415.31/59.5516.76/59.7814.74/61.2118.37/57.4415.62/61.5315.37/60.0611.59/67.12
ViT-L-16/22429.92/64.8728.31/66.5128.39/65.9224.30/67.2730.55/66.9725.47/66.6822.03/74.88
DeiT3-L-16/224-20.28/69.8719.72/70.0224.43/66.6019.56/71.2319.41/70.5215.02/76.13
ViT†-H-16/224-30.06/61.3432.27/60.9830.34/57.7028.30/62.2928.11/62.9623.38/66.95
", + "bbox": [ + 145, + 89, + 848, + 174 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Quantitative feature importance evaluation of ROAD (MoRF ↓ / LeRF ↑). The higher and lower scores indicate better attribution map quality for MoRF and LeRF, respectively. The underlined scores indicate the highest performance among the baselines. $(\\dagger)$ denotes the ViT trained with CLIP [24].", + "bbox": [ + 89, + 176, + 906, + 218 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/38daf34b912ff5644bcfbf51b07b446c9bc5e6300ea98a1a586141d6339e058d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 98, + 220, + 145, + 284 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/698acb6aba24360032fd53e45ce016c0a3292ed7b020b1202fc6fe6cd49f020c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 153, + 220, + 297, + 311 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/24aa980031e7d7b42e2490a466ec45a1e67f8c8c6e8a32877ccc1bc2ca8b3850.jpg", + "image_caption": [ + "(a) IN-1k" + ], + "image_footnote": [], + "bbox": [ + 303, + 219, + 447, + 311 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f7210558684cac70f3ec1b18c459686b2f034fd3b178f0227808215a8086920e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 452, + 218, + 596, + 311 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6e662039c23d65a71ac32957bfa9a9f93fde5c72a23a7fa178ffe0c945b5908d.jpg", + "image_caption": [ + "(b) IN-A" + ], + "image_footnote": [], + "bbox": [ + 606, + 218, + 750, + 311 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6fab59fa92ab46768056ab1f3309f68e827e4aff583f411e7341b54908e3f22c.jpg", + "image_caption": [ + "(c) IN-R" + ], + "image_footnote": [], + "bbox": [ + 759, + 218, + 903, + 311 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "providing better correctness performance achieve higher scores for insertion and lower scores for deletion. We gradually deleted or inserted around every $3.5\\%$ pixels of the input image and randomly sampled 6,000 samples from the IN-1k validation dataset for the assessment. As shown in Table 1, the attributions provided by CoIBA yield predominant correctness performance compared to baselines. Especially, the consistent increase in correctness from the Swin transformer demonstrates the generalizability of CoIBA.", + "bbox": [ + 89, + 383, + 483, + 521 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Remove and Debias", + "text_level": 1, + "bbox": [ + 89, + 530, + 277, + 544 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The remove-and-debias (ROAD) [25] experiments measure the correctness of the attribution maps, addressing the class information leakage problem (i.e., the shape of the mask) observed in the remove-and-retrain (ROAR) [16] experiment. We average the quantitative scores over the 6,000 randomly selected images when 20, 40, 60, and $80\\%$ of pixels are imputed in two perspectives: most relevant first (MoRF) and least relevant first (LeRF). As shown in Table 2, CoIBA outperforms existing methods in MoRF and LeRF experiments with a sizable gap. These results demonstrate that regardless of the imputation type, CoIBA outperforms all the baselines regardless of the model setting (patch size and depth) and pre-trained dataset.", + "bbox": [ + 89, + 551, + 483, + 748 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Difficulty-aware Correctness Assessment", + "text_level": 1, + "bbox": [ + 89, + 757, + 439, + 772 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We compare the quantitative quality of attribution maps yielded by CoIBA and baseline methods in terms of difficulty-aware assessment, in addition to correctness assessments of overall samples. Leveraging the confidence scores, and the prediction probability of a target class, we divide the confidence scores with 20 intervals as in Sec. 3.2. As shown in Fig. 7, CoIBA consistently outperforms all the baselines regardless of the confidence scores predicted with", + "bbox": [ + 89, + 780, + 483, + 900 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ef0cc9d5cd981bbbfb5d8e55d314414e53f46e2bf1b593b33106c2e17cffaf6b.jpg", + "image_caption": [ + "Figure 7. Quantitative difficulty-aware correctness assessment on insertion/deletion and ROAD. We measure differences in insertion/deletion $(\\Delta \\mathrm{InsDel}\\uparrow)$ and ROAD $(\\Delta \\mathrm{ROAD}\\uparrow)$ scores. We fill the regions including low-confident samples and high-confident samples with red and blue, respectively, based on the prediction made by the model. The first to third columns illustrate the quantitative results of the IN-1k. The fourth and fifth columns include the results of IN-A and IN-R datasets, respectively.", + "(a) Ablation on departure layer", + "Figure 8. Comparisons of different departure $s$ and arrival $e$ layer settings. We plot the hyper-parameters selected for CoIBA as a solid line. We compare the quantitative results of the discrepancy between AUC scores of insertion/deletion (ΔInsDel) with different intervals of confidence scores outputted by the model. The better attribution method achieves a higher score. We provide further results in the supplementary material." + ], + "image_footnote": [], + "bbox": [ + 527, + 380, + 702, + 479 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/523a6a810e60cff76795b6fa352d6d38ad0ae93ed88fbddb8baedbc86602ca14.jpg", + "image_caption": [ + "(b) Ablation on arrival layer" + ], + "image_footnote": [], + "bbox": [ + 714, + 380, + 888, + 479 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "each sample, predicted by the model. Along with IN-1k, we include the quantitative results of IN-A and IN-R. As both IN-A and IN-R datasets include the low-confident (difficult) samples compared to IN-1k, these quantitative results support the demonstration that CoIBA outperforms the correct performance with low-confident samples.", + "bbox": [ + 511, + 597, + 906, + 688 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.6. Discussion", + "text_level": 1, + "bbox": [ + 513, + 696, + 632, + 712 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Ablation on Bottleneck-inserted Layers In this section, we confirm whether including multiple layers enhances the correctness of the resulting attribution map. Fig. 9 shows the comparison of the quantitative results by interpolating departure $(s)$ and arrival $(e)$ layers from earlier to deeper layers. The results demonstrate that increasing the number of layers to compute relevant information consistently enhances the correctness quality of an attribution map. Furthermore, CoIBA consistently outperforms IBA* with large margins, requiring only a single iteration. Thus, in CoIBA, various layers gladly reveal the omitted relevant information obtained from a specific layer.", + "bbox": [ + 511, + 719, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "25172", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0705a15d468e8d31d2b6f54a32231b26d8587cce7584e2228f9833cdfead120a.jpg", + "image_caption": [ + "(a) Discrepancy in mutual info." + ], + "image_footnote": [], + "bbox": [ + 107, + 90, + 284, + 183 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/45776e79e5cabef43808a4176649a9522f46873be0e683faa1468f68954cbf0c.jpg", + "image_caption": [ + "(b) Discrepancy in accuracy" + ], + "image_footnote": [], + "bbox": [ + 292, + 90, + 468, + 183 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0fb9e3779c4a5a2e316f8d14bf9c483815558a34cf70660d72d3520759534364.jpg", + "image_caption": [ + "(c) Ablation on universal damping ratio" + ], + "image_footnote": [], + "bbox": [ + 107, + 198, + 284, + 297 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/17e00121058d317341001b7ded8e68b8c2333da05f592254e15595c29e14f19f.jpg", + "image_caption": [ + "Figure 9. Ablation study on layers to analyze universal damping ratio. Fig. 9(a) illustrates the change in mutual information for compression terms $I[Z_1; R_1]$ in Eq. (8) and the Fig. 9(b) shows the change in accuracy. We utilize the ViT-B-16/224 model for these experiments. The mutual info. denotes mutual information." + ], + "image_footnote": [], + "bbox": [ + 290, + 198, + 465, + 297 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effectiveness of Universal Damping Ratio We investigate whether leveraging the universal damping ratio compensates for over-compression in Fig. 9(a) while amplifying the relevant information in Fig. 9(b). We compare the different numbers of targeted layers to compare the quantitative results. The results illustrated in Fig. 9(a) show that leveraging the universal damping ratio complements the overcompressed information that occurred in earlier layers, due to the delivered relevant information from deeper layers. Concurrently, as shown in Fig. 9(c), leveraging the universal damping ratio amplifies the relevant information term in Eq (8). Aligning with these results, compared by assigning individual damping ratios to each layer, our universal damping ratio enhances the correctness performance of CoIBA. Therefore, leveraging the universal damping ratio to targeted multiple layers significantly enhances the correctness performance of CoIBA by compensating for the over-compression and amplifying the relevant information.", + "bbox": [ + 91, + 398, + 483, + 671 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effectiveness of Variational Upper Bound We demonstrate the effectiveness of our variational upper bound by measuring correctness performance and whether the relevancy of different layers is fairly reflected in attributions. We compare quantitative results obtained by leveraging a variational upper bound (Eq. (8)) versus linearly combining layer-wise mutual information (Eq. (4)) in Fig. 10(a). As shown in the results, leveraging our variational upper bound is better at correctly identifying feature importance, regardless of confidence scores. In addition to this result, we confirm whether the relevant information of each layer is fairly reflected in the attribution map by comparing similarities. Concretely, we compare the similarity of both settings (Eq. (4) and (8)) with layer-specific relevant information computed by iterating IBA with channel uniform per", + "bbox": [ + 89, + 674, + 483, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/792d37243daefc43c1d8f6e25b5bad489aee281599deaae8eb20056ab71b83b9.jpg", + "image_caption": [ + "(a) Faithfulness assessment" + ], + "image_footnote": [], + "bbox": [ + 535, + 90, + 709, + 186 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a221e53a85c456d646161a761aa03df00795be600fa8e263a4397249f053a0c9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 712, + 90, + 885, + 175 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/28b5420d25d4859eed8b636ed40e2f7d64eeff554de0090a54478ae4d3b2ba7b.jpg", + "image_caption": [ + "(c) Ablation on channel uniform perturbation" + ], + "image_footnote": [], + "bbox": [ + 532, + 202, + 707, + 297 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/dc630f201e95b3f713076e6af724de4fc895737fc7c3e4b6e6e3305367f5ad06.jpg", + "image_caption": [ + "(b) Similarity measurement", + "Figure 10. Quantitative analysis on CoIBA. We report the results divided by the confidence score computed by the model. For the comparison, we include the ViT-B model pre-trained with CLIP in Fig. 10(c). We filled the discrepancy in performances with blue color in Fig. 10(c). The solid line indicates the results of CoIBA." + ], + "image_footnote": [], + "bbox": [ + 714, + 202, + 887, + 297 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "turbation, denoted as $\\mathrm{IBA}^*$ . We employ CKA [19] as a similarity metric. As shown in Fig. 10(b), leveraging our upper bound yields the comprehensive relevant information of all the layers. This result demonstrates that our variational upper bound encourages the resulting comprehensive relevant information to fairly reflect the relevancy of each layer.", + "bbox": [ + 511, + 393, + 905, + 484 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effectiveness of Uniform Channel Perturbation We demonstrate the effectiveness of uniform channel perturbation by comparing the correctness performance of uniform (unif.) and non-uniform (non-unif.) channel perturbation strategies. As shown in Fig. 10(c), the uniform channel perturbation consistently enhances the correctness performance of IBA for low-confident samples. In addition to this result, uniform channel perturbation significantly enlarges the correctness of resulting attribution maps, regardless of the confidence scores. These enhanced correctness performances demonstrate the effectiveness of our uniform channel perturbation, regardless of the method.", + "bbox": [ + 511, + 484, + 905, + 667 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 679, + 633, + 694 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we introduce CoIBA, which reveals the comprehensive relevant information to produce an attribution map, revealing the omitted relevancy in IBA with a theoretical guarantee. CoIBA shares the universal damping ratio to compensate for the over-compressed information, delivering relevancy among the bottleneck-inserted layers. CoIBA clearly judges the importance of each token by leveraging uniform channel perturbation. We leverage variational approximation to upper bound the information to ensure the eliminated activations are not necessary for the bottleneck-inserted layers to make a decision. CoIBA demonstrates a substantial improvement over existing methods in numerous experiments and discussions.", + "bbox": [ + 511, + 704, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "25173", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 91, + 90, + 250, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 263, + 187, + 277 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Samira Abnar and Willem Zuidema. Quantifying attention flow in transformers. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 4190-4197, 2020. 2", + "[2] Alexander A Alemi, Ian Fischer, Joshua V Dillon, and Kevin Murphy. Deep variational information bottleneck. In International Conference on Learning Representations, 2017. 1, 3", + "[3] Oren Barkan, Yuval Asher, Amit Eshel, Noam Koenigstein, et al. Visual explanations via iterated integrated attributions. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2073-2084, 2023. 1, 2, 4, 6, 7", + "[4] Alexander Binder, Grégoire Montavon, Sebastian Lapuschkin, Klaus-Robert Müller, and Wojciech Samek. Layer-wise relevance propagation for neural networks with local renormalization layers. In Artificial Neural Networks and Machine Learning-ICANN 2016: 25th International Conference on Artificial Neural Networks, Barcelona, Spain, September 6-9, 2016, Proceedings, Part II 25, pages 63-71. Springer, 2016. 1", + "[5] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9650-9660, 2021. 1", + "[6] Hila Chefer, Shir Gur, and Lior Wolf. Generic attention-model explainability for interpreting bi-modal and encoder-decoder transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 397-406, 2021. 1, 2, 6, 7", + "[7] Hila Chefer, Shir Gur, and Lior Wolf. Transformer interpretability beyond attention visualization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 782-791, 2021. 1, 2, 4, 6, 7", + "[8] Jiamin Chen, Xuhong Li, Lei Yu, Dejing Dou, and Haoyi Xiong. Beyond intuition: Rethinking token attributions inside transformers. Transactions on Machine Learning Research, 2022. 1, 2, 4, 6, 7", + "[9] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, pages 248-255. IEEE, 2009. 6", + "[10] Finale Doshi-Velez and Been Kim. Towards a rigorous science of interpretable machine learning. arXiv preprint arXiv:1702.08608, 2017. 1" + ], + "bbox": [ + 93, + 287, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "This work was supported by the Institute of Information & Communications Technology Planning & Evaluation (IITP) grant, funded by the Korea government (MSIT) (No. RS-2019-II190079, Artificial Intelligence Graduate School Program (Korea University), IITP-2025-RS-2024-00436857, ITRC (Information Technology Research Center), and No. RS-2022-II220984, Development of Artificial Intelligence Technology for Personalized Plug-and-Play Explanation and Verification of Explanation).", + "[11] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, G Heigold, S Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2020. 3, 6", + "[12] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dólár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16000-16009, 2022. 1", + "[13] Dan Hendrycks, Steven Basart, Norman Mu, Saurav Kadavath, Frank Wang, Evan Dorundo, Rahul Desai, Tyler Zhu, Samyak Parajuli, Mike Guo, et al. The many faces of robustness: A critical analysis of out-of-distribution generalization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8340-8349, 2021. 6", + "[14] Dan Hendrycks, Kevin Zhao, Steven Basart, Jacob Steinhardt, and Dawn Song. Natural adversarial examples. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15262-15271, 2021. 6", + "[15] Robin Hesse, Simone Schaub-Meyer, and Stefan Roth. Funnybirds: A synthetic vision dataset for a part-based analysis of explainable ai methods. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3981-3991, 2023. 2, 6", + "[16] Sara Hooker, Dumitru Erhan, Pieter-Jan Kindermans, and Been Kim. A benchmark for interpretability methods in deep neural networks. Advances in Neural Information Processing Systems, 32, 2019. 7", + "[17] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6", + "[18] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 3", + "[19] Simon Kornblith, Mohammad Norouzi, Honglak Lee, and Geoffrey Hinton. Similarity of neural network representations revisited. In International Conference on Machine Learning, pages 3519-3529. PMLR, 2019. 8", + "[20] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. 6", + "[21] Ze Liu, Han Hu, Yutong Lin, Zhuliang Yao, Zhenda Xie, Yixuan Wei, Jia Ning, Yue Cao, Zheng Zhang, Li Dong, et al. Swin transformer v2: Scaling up capacity and resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12009-12019, 2022. 6", + "[22] Clemens Otte. Safe and interpretable machine learning: a methodological review. Computational Intelligence in Intelligent Data Analysis, pages 111-122, 2013. 1", + "[23] Vitali Petsiuk, Abir Das, and Kate Saenko. Rise: Randomized input sampling for explanation of black-box models. British Machine Vision Conference, 2018. 2, 3, 6", + "[24] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry," + ], + "bbox": [ + 93, + 92, + 905, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "25174", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 1, 6, 7", + "[25] Yao Rong, Tobias Leemann, Vadim Borisov, Gjergji Kasneci, and Enkelejda Kasneci. A consistent and efficient evaluation strategy for attribution methods. In International Conference on Machine Learning, pages 18770-18795. PMLR, 2022. 2, 7", + "[26] Wojciech Samek and Klaus-Robert Müller. Towards explainable artificial intelligence. Explainable AI: interpreting, explaining and visualizing deep learning, pages 5-22, 2019. 1", + "[27] Karl Schulz, Leon Sixt, Federico Tombari, and Tim Landgraf. Restricting the flow: Information bottlenecks for attribution. In International Conference on Learning Representations, 2020. 1, 2, 3, 4, 6, 7", + "[28] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 618-626, 2017. 1", + "[29] Andreas Peter Steiner, Alexander Kolesnikov, Xiaohua Zhai, Ross Wightman, Jakob Uszkoreit, and Lucas Beyer. How to train your vit? data, augmentation, and regularization in vision transformers. Transactions on Machine Learning Research, 2024. 6", + "[30] Mukund Sundararajan, Ankur Taly, and Qiqi Yan. Axiomatic attribution for deep networks. In International Conference on Machine Learning, pages 3319-3328. PMLR, 2017. 1", + "[31] Naftali Tishby, Fernando C Pereira, and William Bialek. The information bottleneck method. arXiv preprint physics/0004057, 2000. 1", + "[32] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In International Conference on Machine Learning, pages 10347-10357. PMLR, 2021. 4, 6", + "[33] Hugo Touvron, Matthieu Cord, and Hervé Jégou. Deit iii: Revenge of the vit. In Proceedings of the European Conference on Computer Vision, pages 516-533. Springer, 2022. 6", + "[34] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing, 13(4):600-612, 2004. 2", + "[35] Weiyan Xie, Xiao-Hui Li, Caleb Chen Cao, and Nevin L Zhang. Vit-cx: causal explanation of vision transformers. In Proceedings of the Thirty-Second International Joint Conference on Artificial Intelligence, pages 1569–1577, 2023. 1, 2, 6, 7", + "[36] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11975-11986, 2023. 1" + ], + "bbox": [ + 91, + 90, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "ref_text", + "text": "[37] Yang Zhang, Ashkan Khakzar, Yawei Li, Azade Farshad, Seong Tae Kim, and Nassir Navab. Fine-grained neural network explanation by identifying input features with predictive information. Advances in Neural Information Processing Systems, 34:20040-20051, 2021. 1, 2, 3, 5", + "bbox": [ + 516, + 90, + 903, + 161 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "25175", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Comprehensive Information Bottleneck for Unveiling Universal Attribution to Interpret Vision Transformers/abcc8427-3423-466d-afdc-b93969a43358_model.json b/2025/Comprehensive Information Bottleneck for Unveiling Universal Attribution to Interpret Vision Transformers/abcc8427-3423-466d-afdc-b93969a43358_model.json new file mode 100644 index 0000000000000000000000000000000000000000..d9f49e4bbc5eb18d6dda7995fe441e81e4a80066 --- /dev/null +++ b/2025/Comprehensive Information Bottleneck for Unveiling Universal Attribution to Interpret Vision Transformers/abcc8427-3423-466d-afdc-b93969a43358_model.json @@ -0,0 +1,2387 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.02 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation." + }, + { + "type": "header", + "bbox": [ + 0.325, + 0.017, + 0.725, + 0.031 + ], + "angle": 0, + "content": "Except for this watermark, it is identical to the accepted version;" + }, + { + "type": "header", + "bbox": [ + 0.294, + 0.032, + 0.756, + 0.047 + ], + "angle": 0, + "content": "the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.108, + 0.13, + 0.892, + 0.177 + ], + "angle": 0, + "content": "Comprehensive Information Bottleneck for Unveiling Universal Attribution to Interpret Vision Transformers" + }, + { + "type": "text", + "bbox": [ + 0.188, + 0.204, + 0.808, + 0.223 + ], + "angle": 0, + "content": "Jung-Ho Hong Ho-Joong Kim Kyu-Sung Jeon Seong-Whan Lee*" + }, + { + "type": "text", + "bbox": [ + 0.291, + 0.227, + 0.707, + 0.242 + ], + "angle": 0, + "content": "Dept. of Artificial Intelligence, Korea University, Seoul, Korea" + }, + { + "type": "text", + "bbox": [ + 0.26, + 0.244, + 0.733, + 0.259 + ], + "angle": 0, + "content": "{jungho-hong, hjoong_kim, ksjeon, sw.lee}@korea.ac.kr" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.286, + 0.327, + 0.303 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.318, + 0.486, + 0.697 + ], + "angle": 0, + "content": "The feature attribution method reveals the contribution of input variables to the decision-making process to provide an attribution map for explanation. Existing methods grounded on the information bottleneck principle compute information in a specific layer to obtain attributions, compressing the features by injecting noise via a parametric damping ratio. However, the attribution obtained in a specific layer neglects evidence of the decision-making process distributed across layers. In this paper, we introduce a comprehensive information bottleneck (CoIBA), which discovers the relevant information in each targeted layer to explain the decision-making process. Our core idea is applying information bottleneck in multiple targeted layers to estimate the comprehensive information by sharing a parametric damping ratio across the layers. Leveraging this shared ratio complements the over-compressed information to discover the omitted clues of the decision by sharing the relevant information across the targeted layers. We suggest the variational approach to fairly reflect the relevant information of each layer by upper bounding layer-wise information. Therefore, CoIBA guarantees that the discarded activation is unnecessary in every targeted layer to make a decision. The extensive experimental results demonstrate the enhancement in faithfulness of the feature attributions provided by CoIBA." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.714, + 0.222, + 0.729 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.74, + 0.485, + 0.877 + ], + "angle": 0, + "content": "Vision transformer (ViT) achieves remarkable performance in diverse fundamental computer vision tasks, such as multi-modal [24, 36] and self-supervised learning [5, 12]. Despite this achievement, the black box nature stemming from the complex structure of ViT restricts the testability of the end-to-end system, impeding the failure diagnosis [10]. This limitation constrains the application of ViT in safety-critical areas where transparency and interpretability are considered important as well [22, 26]. Feature attribu" + }, + { + "type": "image", + "bbox": [ + 0.526, + 0.284, + 0.9, + 0.412 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.528, + 0.413, + 0.707, + 0.512 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.71, + 0.413, + 0.889, + 0.512 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.521, + 0.514, + 0.902, + 0.527 + ], + "angle": 0, + "content": "(b) Cumulative number of each layer yielding optimal result per sample" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.528, + 0.906, + 0.667 + ], + "angle": 0, + "content": "Figure 1. The quantitative and qualitative comparisons of attribution maps generated from different layers. We illustrate the visualized attribution maps from the first to fourth columns in Fig. 1(a) utilizing IBA [27]. The text on the top of each figure indicates the index of the layer. We select the layer index of the large model with interval 2. The fifth column in Fig. 1(a) shows the similarity comparison between attributions of different layers. Fig. 1(b) illustrates the cumulative number of each layer that provides the best explanation per sample. The uniformly distributed data indicates the absence of the layer optimal for the explanation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.673, + 0.907, + 0.809 + ], + "angle": 0, + "content": "tion methods [4, 28, 30] are introduced to reveal the evidence supporting the decision-making process. Many of the existing attribution methods designed to interpret convolutional neural networks show limited explainability to interpret ViT [7]. To address this limitation, feature attribution methods [3, 6-8, 35] tailored for interpreting ViT have been proposed, leveraging either attention weights computed in ViT or intermediate representations, such as the class token, to generate attribution maps." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Although feature attribution methods have advanced in interpreting ViT, they lack theoretical guarantees that low-scored attributions are not necessary to the decision-making process [27]. The methods [27, 37] grounded on the information bottleneck principle [2, 31] address this limitation by guaranteeing the importance of attributions with a" + }, + { + "type": "page_footnote", + "bbox": [ + 0.116, + 0.888, + 0.241, + 0.9 + ], + "angle": 0, + "content": "*Corresponding author" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "25166" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.334 + ], + "angle": 0, + "content": "variational approximation. Information bottleneck for attribution (IBA) [27] proposes the model-agnostic method to compress the information unnecessary for the prediction by positioning the bottleneck layer into the target layer. InputIBA [37] succeeds in providing a high-resolution attribution map by directly measuring the information in the input domain, assuming the prior distribution with a generative model. However, existing information bottleneck-based methods face two main challenges. First, the resulting attribution map reflects the partial of the decision-making process as the restriction of the information flow is conducted in a specific targeted layer. This constraint provides inconsistent evidence for a single decision-making process, leading to unobvious interpretations. Lastly, no layer dominantly provides the most appropriate relevant information for explaining the decision-making process." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.338, + 0.483, + 0.716 + ], + "angle": 0, + "content": "Fig. 1 shows the aforementioned issues with quantitative and qualitative examples. In the first row in Fig. 1(a), the visualization derived from the second layer (blocks.2) highlights the barrel whereas the fifth (blocks.5) and eighth layers (blocks.8) highlight the top and bottom of the wheels, respectively. In the second row in Fig. 1(a), the visualization in the earlier layer highlights the person holding the racket, while the deeper layer progressively concentrates on the racket. To confirm the layer-wise discrepancy in attributions quantitatively, we measure SSIM [34] in the right row of Fig. 1(a). As the layer providing attributions distance farther, the dissimilarity in those attributions increases. Thus, computing the relevant information from isolated layer-specific information bottlenecks highlights different attributions for the same decision. Among the layers producing the attribution maps, there is no dominant layer that provides the most appropriate relevant information to explain the decision-making process. Fig. 1(b) compares the cumulative number of each layer yielding attributions with the optimal insertion and deletion scores per input sample. The results show a lack of the layer that produces the most appropriate relevant information for explanation. Thus, identifying the most faithful attribution from the layers requires a heuristic search due to the lack of criteria to distinguish the optimal layer." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.483, + 0.901 + ], + "angle": 0, + "content": "In this paper, we propose a novel comprehensive information bottleneck for attribution (CoIBA), which reveals the information relevant for each of the target layers to provide the attributions. Our core idea is to reveal comprehensive relevant information by eliminating unnecessary information in every targeted layer of the prediction. We estimate the comprehensive information by sharing a parameterized universal damping ratio across layers, thereby removing the requirements of additional heuristics. This sharing strategy compensates for the over-compressed information of individual layers by sharing information in each layer necessary for the decision. Since a channel el" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.227 + ], + "angle": 0, + "content": "ement placed in the same location but different layers captures different features, we uniformly perturb all the channel elements in a single token representation, to handle the information of different layers with a shared parametric ratio. We suggest a variational upper bound to restrict information flowing in all the targeted layers. Our approach eliminates the heuristic search for balancing inconsistent information along the layers by suggesting a variational upper bound to fairly reflect the layer-wise relevant information." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.228, + 0.905, + 0.394 + ], + "angle": 0, + "content": "We conduct experiments to show the correctness performance of CoIBA utilizing various assessments. As the evaluations are solely approximations, i.e. no ground truth exists, we include FunnyBirds [15], which provide the ground truth and assess the comprehensive quality of attribution maps. To assess the feature importance assessment, we leverage insertion/deletion [23] and remove-and-debias (ROAD) [25], which evaluate the faithfulness of attribution maps. In addition, we scrutinize the faithfulness of CoIBA by analyzing the confident-aware assessment. We discuss the effectiveness of our method with a rigorous discussion." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.408, + 0.663, + 0.423 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.433, + 0.905, + 0.825 + ], + "angle": 0, + "content": "Explanation Methods for Vision Transformer Existing explanatory methods designed to interpret ViT provide the attribution map utilizing the latent representation and the corresponding gradients. However, these methods show limited explainability while being adapted to ViT [7]. To address this limitation, Rollout [1] linearly combines the attention weights across the layers to compute the attribution map. Trans-attr [7] provides the class-discriminative attribution map by constructing the layer-wise relevance propagation rule for ViT. Generic [6] produces the attribution map with a generalized procedure to interpret diversified transformer models, leveraging the gradient of the attention map to produce the class-discriminative attribution maps. IIA [3] introduces the iterative integration across the input image, leveraging the internal representations processed by the model and their gradients. ViT-CX [35] utilizes the patch embeddings and measures causal impacts to provide attributions. Beyond [8] unfolds attention blocks with the chain rule between final prediction, CLS, and tokens to obtain token contribution. Existing propagation-based approaches require a specific implementation to weigh the contributions of input variables. In contrast to this constraint, CoIBA does not require implementing additional procedures as same as IBA. Furthermore, aligning with the information bottleneck principle, CoIBA guarantees the highlighted attributions are important in all targeted layers for the decision." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Information Bottleneck Approach The information bottleneck principle is broadly utilized to obtain information in the activation necessary for the specific objective. To this end, the information bottleneck-based methods compress the information unnecessary for the objective while" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "25167" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.089, + 0.481, + 0.329 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.335, + 0.485, + 0.419 + ], + "angle": 0, + "content": "Figure 2. Main difference between IBA and CoIBA. IBA obtains the relevant information in the specific bottleneck-inserted layer. Thus, obtaining layer-wise attribution maps necessarily requires iterative running over the layers. In contrast, CoIBA seeks the layer-wise relevant information from each bottleneck-inserted layer with a universal damping ratio." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.44, + 0.484, + 0.713 + ], + "angle": 0, + "content": "maintaining the relevant information. Given the challenges of directly estimating the amount of information, a variational approach [18] is used to approximate the intractable posteriors and marginals. The Deep Variational Information Bottleneck [2] method employs the information bottleneck principle in deep networks, using a variational lower bound alongside the reparameterization trick. Referring to this strategy, IBA [27] adapts the information bottleneck principles to restrict the information flow in an intermediate layer and provides the relevant information as an attribution map. InputIBA [37] provides high-resolution attributions by directly estimating relevant information in the input domain. However, existing methods compute relevant information solely considering a specific layer, overlooking the information necessary in the different layers. In contrast to this process, CoIBA obtains comprehensive relevant information from the multiple layers, highlighting the evidence important in every targeted layer to make a decision." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.736, + 0.182, + 0.751 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.484, + 0.901 + ], + "angle": 0, + "content": "The main goal of CoIBA is to produce an attribution map by reflecting the information relevant to the targeted layers, which is opposed to the IBA as shown in Fig. 2. To show this procedure, we first introduce IBA in Sec. 3.1 and discuss the limitations of the existing information bottleneck-based approach in Sec. 3.2. Then, in Sec. 3.3, we describe the overall procedure of CoIBA. Finally, we introduce a variational upper bound, which enhances the reflection of layer-wise relevant information in Sec. 3.4." + }, + { + "type": "image", + "bbox": [ + 0.537, + 0.089, + 0.707, + 0.194 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.71, + 0.09, + 0.885, + 0.194 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.538, + 0.197, + 0.634, + 0.271 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.635, + 0.197, + 0.716, + 0.271 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.717, + 0.197, + 0.8, + 0.271 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.801, + 0.197, + 0.884, + 0.271 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.275, + 0.907, + 0.358 + ], + "angle": 0, + "content": "Figure 3. Quantitative and qualitative results of IBA*. In the first row, the horizontal line indicates the performance of layers-specific relevant information, i.e., IBA. We report the difference in insertion and deletion scores (\\(\\Delta\\)InsDel). The higher score is better for this metric. The visualized attribution maps are illustrated in the second row obtained from VIT-B-16/224 [11]." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.366, + 0.694, + 0.382 + ], + "angle": 0, + "content": "3.1. Background - IBA" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.388, + 0.906, + 0.448 + ], + "angle": 0, + "content": "IBA adopts an information bottleneck principle to produce relevant information to provide an attribution map. To this end, in \\( l \\)-th layer, IBA computes the bottleneck representation \\( Z_{l} \\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.625, + 0.462, + 0.905, + 0.478 + ], + "angle": 0, + "content": "\\[\nZ _ {l} = \\lambda_ {l} R _ {l} + (1 - \\lambda_ {l}) \\epsilon_ {l}. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.489, + 0.906, + 0.595 + ], + "angle": 0, + "content": "Here, the damping ratio \\(\\lambda_{l}\\) is a learnable parameter assigned at the bottleneck layer, manipulating the degree of damping signal between the activation \\(R_{l}\\) and the independent noise \\(\\epsilon_{l}\\). The independent noise \\(\\epsilon_{l}\\) is sampled from the Gaussian distribution. To maintain the statistic of internal representation, the Gaussian distribution shares the mean \\(\\mu_{R_l}\\) and variance \\(\\sigma_{R_l}\\) with the representation \\(R_{l}\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.596, + 0.906, + 0.655 + ], + "angle": 0, + "content": "IBA minimizes the shared information between the activation \\( R_{l} \\) and bottleneck variable \\( Z_{l} \\) while maximizing the shared information between the bottleneck variable \\( Z_{l} \\) and the label \\( Y \\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.614, + 0.668, + 0.905, + 0.692 + ], + "angle": 0, + "content": "\\[\n\\max _ {\\lambda_ {l}} I \\left[ Z _ {l}; Y \\right] - \\beta I \\left[ R _ {l}, Z _ {l} \\right]. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.703, + 0.905, + 0.733 + ], + "angle": 0, + "content": "Here, \\(\\beta\\) is a hyperparameter manipulating the trade-off between compression and relevancy in \\(l\\)-th layer." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.743, + 0.635, + 0.758 + ], + "angle": 0, + "content": "3.2. Motivation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.907, + 0.901 + ], + "angle": 0, + "content": "IBA selects a specific layer to obtain relevant information to provide an attribution map. However, the relevant information of a specific single layer does not reflect the overall evidence required in the sequence of layers to make a decision. To demonstrate our insistence, we illustrate the quantitative comparison of the attribution maps obtained in a specific layer and multiple layers in Fig. 3, leveraging insertion/deletion [23]. We iterate IBA multiple times along the layers to obtain the set of individual relevant information" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "25168" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.129, + 0.11, + 0.148, + 0.177 + ], + "angle": 0, + "content": "eaaee" + }, + { + "type": "image", + "bbox": [ + 0.163, + 0.097, + 0.251, + 0.164 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.169, + 0.165, + 0.242, + 0.175 + ], + "angle": 0, + "content": "Class: Hyena" + }, + { + "type": "image_caption", + "bbox": [ + 0.156, + 0.179, + 0.255, + 0.189 + ], + "angle": 0, + "content": "Confidence: 85.87" + }, + { + "type": "image", + "bbox": [ + 0.256, + 0.097, + 0.434, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.129, + 0.212, + 0.148, + 0.282 + ], + "angle": 270, + "content": "Hard Sample" + }, + { + "type": "image", + "bbox": [ + 0.162, + 0.198, + 0.25, + 0.267 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.153, + 0.268, + 0.259, + 0.278 + ], + "angle": 0, + "content": "Class: Parallel Bars" + }, + { + "type": "image_caption", + "bbox": [ + 0.156, + 0.282, + 0.253, + 0.292 + ], + "angle": 0, + "content": "Confidence: 19.71" + }, + { + "type": "image", + "bbox": [ + 0.256, + 0.235, + 0.285, + 0.257 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.287, + 0.199, + 0.434, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.301, + 0.484, + 0.44 + ], + "angle": 0, + "content": "Figure 4. The quantitative comparisons of explanation methods in high-confident (easy) and low-confident (hard) input samples. The confidence score is the softmax probability of the target class predicted by the model. We compare the information-bottleneck-based methods including IBA [27], IBA*, and CoIBA (marked with a circle), and the propagation-based explanation baselines including Beyond [8], IIA [3], and Chefer-LRP [7] methods (marked with a triangle). We utilize DeiT-B-16/224 [32] for the evaluation. For insertion and deletion, higher and lower scores are better, respectively." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.448, + 0.484, + 0.629 + ], + "angle": 0, + "content": "for each layer. After that, we compute a linear combination over the set of individual relevant information. We denote this procedure as \\(\\mathrm{IBA}^*\\). As shown in the results, the attribution maps produced by \\(\\mathrm{IBA}^*\\) archive better scores compared to the IBA. To scrutinize the quantitative results between \\(\\mathrm{IBA}^*\\) and IBA, we split the insertion/deletion scores by the confidence scores of the model in Fig. 4. As shown in the results, the increase in performance is only observed in high-confident samples, unlike in low-confident samples. Therefore, even iterating IBA, which requires high computational cost, shows limited performance enhancement according to the difficulty of each sample." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.638, + 0.431, + 0.654 + ], + "angle": 0, + "content": "3.3. Comprehensive Information Bottleneck" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.66, + 0.483, + 0.825 + ], + "angle": 0, + "content": "Our comprehensive information bottleneck (CoIBA) inserts the bottleneck layer into the sequence of targeted layers to restrict the information flowing in the intermediate layer. Within the bottleneck layer, CoIBA restricts the information with the learnable universal damping ratio shared across the bottleneck-inserted layers. Our goal is to optimize the universal damping ratio to obtain the layer-wise relevant information while fulfilling the information bottleneck objective. The universal damping ratio enables the attributions to identify distributed relevant information omitted by IBA, which restricts information in a layer-specific manner." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Formulating Information Restriction The model is trained to make a decision \\(\\tilde{Y} = f(X)\\) to predict the ground truth \\(Y\\) by taking input \\(X\\). To make a decision, the input sample \\(X\\) is contextualized by passing through the sequence of intermediate layers. Here, we insert the bottle-" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.273 + ], + "angle": 0, + "content": "neck layer into the targeted layer. The role of the bottleneck layer is to restrict the information by dampening the signal of the intermediate representation passing through the \\(L\\) bottleneck-inserted layers. We perturb the intermediate representations in the \\(l\\)-th layer to dampen the signal. To dampen the signal, we inject the independent noise \\(\\epsilon_{l}\\) into the intermediate bottleneck representation \\(R_{l}^{\\prime}\\), which is computed from the bottleneck variables \\(Z_{l-1}\\) of the preceding layer. Note that the first element of intermediate bottleneck representations is the non-perturbed intermediate representation, i.e., \\(R_{1}^{\\prime} = R_{1}\\). Thus, we obtain the bottleneck representation \\(Z_{l}\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.63, + 0.287, + 0.907, + 0.305 + ], + "angle": 0, + "content": "\\[\nZ _ {l} = \\lambda R _ {l} ^ {\\prime} + (1 - \\lambda) \\epsilon_ {l}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.317, + 0.907, + 0.635 + ], + "angle": 0, + "content": "Here, a universal damping ratio \\(\\lambda \\in \\mathbb{R}^{P\\times 1}\\) manipulates the degree of perturbation for each token, denoting \\(P\\) as the number of patches. Thus, \\(\\lambda\\) is ranged from 0 to 1, allowing the propagation of the signal when \\(\\lambda = 1\\) otherwise blocking it when \\(\\lambda = 0\\). As the universal damping ratio \\(\\lambda\\) is consistently adapted to the layers, we omit the layer index in the notation. We compute \\(\\lambda = \\mathrm{sigmoid}(\\alpha)\\) by passing the trainable parameter \\(\\alpha\\), which is initialized with 5, to the sigmoid function. The universal damping ratio uniformly perturbs the signal along the channel dimension. This setting avoids handling the neurons that capture different features with the same coefficient and induces concentration on the token's importance. We empirically show this uniform perturbation across the channels enhances the faithfulness of resulting relevant information. The independent noise \\(\\epsilon_{l}\\) shares the dimension and is sampled from the Gaussian distribution with the mean \\(\\mu_{R_l}\\) and variance \\(\\sigma_{R_l}^2\\), such that \\(\\epsilon_{l}\\sim \\mathcal{N}(\\mu_{R_l},\\sigma_{R_l}^2)\\). To minimize statistical differences between the bottleneck variable and non-perturbed activations, we leverage the mean \\(\\mu_{R_l}\\) and variance \\(\\sigma_{R_l}^2\\) from those of the non-perturbed activations." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.636, + 0.909, + 0.863 + ], + "angle": 0, + "content": "Information Bottleneck Objective The information bottleneck principle desires to obtain information necessary for the objective while compressing irrelevant information from the signal. Compressing irrelevant information is performed by minimizing the shared information between the activation of a bottleneck-inserted network and the bottleneck representation. Building upon this, we compress the amount of information by minimizing the mutual information between internal sequences of bottleneck representations while maximizing the mutual information between the ground truth and bottleneck representations. Combining two aspects with the ground truth \\( Y \\), we maximize the mutual information between the bottleneck variable and the ground truth while minimizing the mutual information between bottleneck representations as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.557, + 0.875, + 0.907, + 0.906 + ], + "angle": 0, + "content": "\\[\n\\max _ {\\lambda} I \\left[ Z _ {L}; Y \\right] - \\frac {1}{L} \\left(\\sum_ {l = 1} ^ {L} \\beta_ {l} I \\left[ Z _ {l - 1}; Z _ {l} \\right]\\right). \\tag {4}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "25169" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.135, + 0.089, + 0.868, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.306, + 0.905, + 0.334 + ], + "angle": 0, + "content": "Figure 5. An overview of CoIBA. CoIBA restricts the information flow using a universal damping ratio while upper-bounding the information of each layer. After that, CoIBA aggregates the relevant information from each layer to produce an attribution map." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.344, + 0.484, + 0.508 + ], + "angle": 0, + "content": "Here, the trade-off parameter \\(\\beta_{l}\\) governs the degree of compression in each layer. Since the bottleneck layers are inserted into the sequence of layers, the compression term includes the accumulated sequence of mutual information of bottleneck variables \\(I[Z_{l-1};Z_l]\\). Here, the \\(Z_0\\) indicates the original intermediate representation as the first element, such that \\(Z_0 = R_1\\). We maximize the relevancy term \\(I[Z_L;Y]\\) while compression to maintain the shared information between bottleneck variables and the ground truth. We maximize this term by minimizing the cross-entropy loss \\(\\mathcal{L}_{ce}\\)." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.51, + 0.484, + 0.542 + ], + "angle": 0, + "content": "Utilizing KL divergence, the mutual information between internal representations \\( I[Z_{l-1}; Z_l] \\) is formulated by:" + }, + { + "type": "equation", + "bbox": [ + 0.116, + 0.561, + 0.483, + 0.58 + ], + "angle": 0, + "content": "\\[\nI \\left[ Z _ {l - 1}; Z _ {l} \\right] = \\mathbb {E} _ {Z _ {l - 1}} \\left[ D _ {K L} \\left[ P \\left(Z _ {l} \\mid Z _ {l - 1}\\right) \\right\\vert \\left| P \\left(Z _ {l}\\right) \\right] \\right] \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.585, + 0.484, + 0.691 + ], + "angle": 0, + "content": "However, the direct estimation of the prior distribution \\( P(Z_{l}) \\) is intractable as it requires the integration of bottleneck representations \\( P(Z_{l}) = \\int_{Z_{l - 1}}P(Z_{l}|Z_{l - 1})P(Z_{l - 1})dZ_{l - 1} \\) of the corresponding \\( l \\)-th layer. Thus, we assume the prior distribution \\( Q(Z_{l}) \\) as Gaussian distribution \\( \\mathcal{N}(\\mu_{R_l},\\sigma_{R_l}^2) \\), reconstituting the mutual information computation as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.114, + 0.697, + 0.483, + 0.716 + ], + "angle": 0, + "content": "\\[\nI \\left[ Z _ {l}; Z _ {l - 1} \\right] \\leq \\mathbb {E} _ {Z _ {l - 1}} \\left[ D _ {K L} \\left[ P \\left(Z _ {l} \\mid Z _ {l - 1}\\right) \\right\\vert \\left| Q \\left(Z _ {l}\\right) \\right] \\right]. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.484, + 0.781 + ], + "angle": 0, + "content": "Since assuming the prior distribution with Gaussian distribution only overestimates the mutual information, minimizing this term suppresses the mutual information of the compression term." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.788, + 0.323, + 0.805 + ], + "angle": 0, + "content": "3.4. Variational Upper Bound" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.484, + 0.902 + ], + "angle": 0, + "content": "As discussed in Sec. 4.6, the relevant information obtained by suppressing layer-wise mutual information suffers from reflecting layer-wise relevant information due to the inconsistent amount of information along the layers. Thus, obtaining the sequence of trade-off parameters \\(\\{\\beta_l\\}_{l = 1}^L\\) requires an extensive heuristic search to balance the degree" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.344, + 0.907, + 0.525 + ], + "angle": 0, + "content": "of compression for each layer. To overcome this issue, we suggest the variational upper bound for CoIBA, referring to the two following points. First, since there is no additional information propagated during the forwarding pass, the mutual information of intermediate layers would not be greater than the mutual information between the model input and output. Second, the iterated noise-injection procedure diminishes the mutual information among the internal and bottleneck representations, i.e., \\( I[Z_{l};Z_{l + 1}] \\leq I[Z_{l - 1};Z_l] \\). Building upon these bases, we establish an upper bound that encompasses the combined mutual information from all subsequent layers as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.577, + 0.546, + 0.907, + 0.575 + ], + "angle": 0, + "content": "\\[\nI \\left[ R _ {1}; Z _ {1} \\right] \\geq \\frac {1}{L} \\left(\\sum_ {l = 1} ^ {L} I \\left[ Z _ {l - 1}; Z _ {l} \\right]\\right). \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.595, + 0.906, + 0.626 + ], + "angle": 0, + "content": "Accordingly, we reconstitute the objective to be a simplified formula, leaving only a single hyperparameter \\(\\beta\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.611, + 0.661, + 0.905, + 0.684 + ], + "angle": 0, + "content": "\\[\n\\max _ {\\lambda} I \\left[ Z _ {L}; Y \\right] - \\beta I \\left[ R _ {1}; Z _ {1} \\right]. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.907, + 0.901 + ], + "angle": 0, + "content": "To compress the subsequent layers, the simplified objective necessitates calculating only the mutual information of the first layer \\( I[R_1;Z_1] \\). This term solely utilizes the non-perturbed intermediate and bottleneck representations of the first layer, simplifying the objective calculation. In terms of relevant information computation, referring to the data processing inequality, the inequality \\( I[Y;Z_l]\\leq I[Y;Z_{l - 1}] \\) holds. Thereby, in contrast to the layer-specific information bottleneck, which overestimates mutual information in the earlier layers during compression [37], CoIBA relieves the overestimation. This is because, as discussed in Sec. 4.6, the relevant information is compensated as the sequence of bottleneck variables joins the objective computation." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "25170" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.104, + 0.136, + 0.191, + 0.201 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.108, + 0.203, + 0.187, + 0.213 + ], + "angle": 0, + "content": "Original Image" + }, + { + "type": "image", + "bbox": [ + 0.202, + 0.091, + 0.291, + 0.166 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.197, + 0.168, + 0.296, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.3, + 0.092, + 0.388, + 0.166 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.296, + 0.168, + 0.395, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.399, + 0.092, + 0.486, + 0.166 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.396, + 0.168, + 0.496, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.498, + 0.092, + 0.585, + 0.166 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.492, + 0.168, + 0.591, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.593, + 0.092, + 0.681, + 0.166 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.592, + 0.168, + 0.689, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.691, + 0.092, + 0.78, + 0.166 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.689, + 0.168, + 0.788, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.792, + 0.092, + 0.878, + 0.166 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.786, + 0.168, + 0.889, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.252, + 0.907, + 0.294 + ], + "angle": 0, + "content": "Figure 6. Quantitative comprehensive assessment on FunnyBirds experiment: The abbreviations Com, Cor, and Con indicate Completeness, correctness, and contrastivity. The center score indicates the comprehensive result of enumerated aspects. Acc. and B.I. indicate accuracy and background independence, respectively. We provide detailed numeric in the supplementary material." + }, + { + "type": "table", + "bbox": [ + 0.134, + 0.299, + 0.86, + 0.44 + ], + "angle": 0, + "content": "
VariantModelChefer-LRP [7]Generic [6]IIA [3]ViT-CX [35]IBA [27]Beyond [8]ColBA
ViTViT*-S-16/224-14.56/56.0813.69/57.3114.98/56.6513.56/58.6413.50/58.4711.08/63.35
ViT*-T-16/224-9.00/48.958.50/49.7512.08/45.098.02/49.2211.26/44.076.97/54.23
ViT-B-16/22417.67/56.6917.92/57.0316.56/58.716.94/58.1717.23/59.0615.84/60.3713.01/62.58
ViT-L-16/22420.97/53.8120.17/55.5420.14/55.2219.76/57.3820.88/54.8518.99/56.5515.40/61.14
DeiTDeiT-S-16/22412.76/47.5112.10/48.6410.89/49.7617.57/44.4311.43/47.8711.13/48.889.55/53.38
DeiT-B-16/22414.63/48.7915.10/49.4113.69/50.5118.21/46.9713.95/50.1814.45/48.8811.79/53.96
DeiT3-B-16/224-15.70/52.4215.38/52.7018.88/50.4015.07/53.3214.71/54.0712.97/56.54
DeiT3-L-16/224-22.10/62.7317.75/55.9326.11/59.6021.01/65.6019.34/64.2717.64/67.20
SwinSwin-B-33.76/42.94--18.03/52.8123.73/50.3917.09/54.80
Swin2-B-34.71/47.16--20.94/53.8531.32/49.4418.88/55.77
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.443, + 0.907, + 0.472 + ], + "angle": 0, + "content": "Table 1. Quantitative feature importance assessment on insertion ↑/deletion ↓. (*) denotes the ViT trained with strong regularization techniques [29]. We underline the state-of-the-art performance among the baselines." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.478, + 0.224, + 0.494 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.507, + 0.173, + 0.523 + ], + "angle": 0, + "content": "4.1. Setup" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.534, + 0.484, + 0.746 + ], + "angle": 0, + "content": "Architectures We select variants of ViT for our experiment, including original ViT [11] and DeiT models [32, 33], trained with ImageNet-21k (IN-21k) and ImageNet-1k (IN-1k), respectively. We include the variants in depth interpolated from ViT-T, ViT-S, ViT-B, ViT-L, and ViT-H models. For the ViT-T and ViT-S models, we utilize the model pretrained with massive regularization techniques [29]. The ViT-H model is pre-trained with CLIP [24]. We present the settings of the ViT-B model with patch size 16 and image resolution 224 as ViT-B-16/224. We include Swin transformers [20, 21], demonstrating the generalized ability of CoIBA in multi-scale features processing. The details of settings and further results about convolutional neural networks are included in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Feature Attribution Methods We compare the various feature attribution methods including Chefer-LRP [7], Generic [6], IIA [3], ViT-CX [35], Beyond [8], and IBA [27] as baselines. We select \\(\\beta = 10\\) for IBA and 6-th layer to insert bottleneck. For CoIBA, we select tradeoff hyper-parameter \\(\\beta\\) as 1. We insert the bottleneck layer from \\(s\\)-th departure to \\(e\\)-th arrival layers, which are chosen, 4, and 12, respectively, for ViT-B. For the optimization, we set the learning rate, batch size, and optimizer as 1, 10, and Adam [17]. The additional hyperparameter settings" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.479, + 0.907, + 0.554 + ], + "angle": 0, + "content": "are included in the supplementary material. For Swin transformers, we only report the explanation methods that do not require methodology modification, including Generic, ViTCX, IBA, Beyond, and CoIBA. We utilize RTX A6000 GPU for all the experiments." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.555, + 0.906, + 0.615 + ], + "angle": 0, + "content": "Datasets We utilize ImageNet-1k [9] (IN-1k), ImageNet-A (IN-A) [14] and ImageNet-R (IN-R) [13] validation datasets for the experiments. We leverage IN-A and IN-R to conduct the difficult-aware analysis." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.623, + 0.736, + 0.639 + ], + "angle": 0, + "content": "4.2. FunnyBirds Assessment" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.644, + 0.907, + 0.796 + ], + "angle": 0, + "content": "The FunnyBirds [15] experiment provides the ground truth for the evaluation. With the help of ground truth, the Funnybird framework assesses the quality of the attribution map from three perspectives: completeness, correctness, and contrastivity. Fig. 6 illustrates the qualitative and quantitative results of the Funnybirds experiment. As shown in the results, CoIBA outperforms all the baselines, including the propagation or gradient-based approaches. Furthermore, in contrast to IBA, CoIBA obtains an attribution map with significantly enhanced contrastivity." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.804, + 0.691, + 0.818 + ], + "angle": 0, + "content": "4.3. Insertion/Deletion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.906, + 0.9 + ], + "angle": 0, + "content": "Insertion/deletion [23] measures the correctness of feature importance highlighted by an attribution map. The insertion/deletion method gradually inserts or deletes informative pixels in ascending order of attribution to compute the area under the curve scores. In insertion/deletion, methods" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "25171" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.147, + 0.09, + 0.849, + 0.175 + ], + "angle": 0, + "content": "
ModelChefer-LRP [7]Generic [6]IIA [3]ViT-CX [35]IBA [27]Beyond [8]CoIBA
ViT-B-16/22420.87/64.3722.15/65.3119.58/66.1919.04/65.3620.99/68.8017.93/68.4416.63/73.68
DeiT-B-16/22415.31/59.5516.76/59.7814.74/61.2118.37/57.4415.62/61.5315.37/60.0611.59/67.12
ViT-L-16/22429.92/64.8728.31/66.5128.39/65.9224.30/67.2730.55/66.9725.47/66.6822.03/74.88
DeiT3-L-16/224-20.28/69.8719.72/70.0224.43/66.6019.56/71.2319.41/70.5215.02/76.13
ViT†-H-16/224-30.06/61.3432.27/60.9830.34/57.7028.30/62.2928.11/62.9623.38/66.95
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.178, + 0.907, + 0.219 + ], + "angle": 0, + "content": "Table 2. Quantitative feature importance evaluation of ROAD (MoRF ↓ / LeRF ↑). The higher and lower scores indicate better attribution map quality for MoRF and LeRF, respectively. The underlined scores indicate the highest performance among the baselines. \\((\\dagger)\\) denotes the ViT trained with CLIP [24]." + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.221, + 0.147, + 0.285 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.154, + 0.221, + 0.298, + 0.312 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.304, + 0.22, + 0.448, + 0.312 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.454, + 0.219, + 0.598, + 0.312 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.607, + 0.219, + 0.751, + 0.312 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.761, + 0.219, + 0.905, + 0.312 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.324, + 0.314, + 0.376, + 0.325 + ], + "angle": 0, + "content": "(a) IN-1k" + }, + { + "type": "image_caption", + "bbox": [ + 0.657, + 0.314, + 0.706, + 0.325 + ], + "angle": 0, + "content": "(b) IN-A" + }, + { + "type": "image_caption", + "bbox": [ + 0.811, + 0.314, + 0.859, + 0.325 + ], + "angle": 0, + "content": "(c) IN-R" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.326, + 0.907, + 0.381 + ], + "angle": 0, + "content": "Figure 7. Quantitative difficulty-aware correctness assessment on insertion/deletion and ROAD. We measure differences in insertion/deletion \\((\\Delta \\mathrm{InsDel}\\uparrow)\\) and ROAD \\((\\Delta \\mathrm{ROAD}\\uparrow)\\) scores. We fill the regions including low-confident samples and high-confident samples with red and blue, respectively, based on the prediction made by the model. The first to third columns illustrate the quantitative results of the IN-1k. The fourth and fifth columns include the results of IN-A and IN-R datasets, respectively." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.385, + 0.484, + 0.522 + ], + "angle": 0, + "content": "providing better correctness performance achieve higher scores for insertion and lower scores for deletion. We gradually deleted or inserted around every \\(3.5\\%\\) pixels of the input image and randomly sampled 6,000 samples from the IN-1k validation dataset for the assessment. As shown in Table 1, the attributions provided by CoIBA yield predominant correctness performance compared to baselines. Especially, the consistent increase in correctness from the Swin transformer demonstrates the generalizability of CoIBA." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.531, + 0.279, + 0.545 + ], + "angle": 0, + "content": "4.4. Remove and Debias" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.553, + 0.484, + 0.749 + ], + "angle": 0, + "content": "The remove-and-debias (ROAD) [25] experiments measure the correctness of the attribution maps, addressing the class information leakage problem (i.e., the shape of the mask) observed in the remove-and-retrain (ROAR) [16] experiment. We average the quantitative scores over the 6,000 randomly selected images when 20, 40, 60, and \\(80\\%\\) of pixels are imputed in two perspectives: most relevant first (MoRF) and least relevant first (LeRF). As shown in Table 2, CoIBA outperforms existing methods in MoRF and LeRF experiments with a sizable gap. These results demonstrate that regardless of the imputation type, CoIBA outperforms all the baselines regardless of the model setting (patch size and depth) and pre-trained dataset." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.758, + 0.441, + 0.773 + ], + "angle": 0, + "content": "4.5. Difficulty-aware Correctness Assessment" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.484, + 0.901 + ], + "angle": 0, + "content": "We compare the quantitative quality of attribution maps yielded by CoIBA and baseline methods in terms of difficulty-aware assessment, in addition to correctness assessments of overall samples. Leveraging the confidence scores, and the prediction probability of a target class, we divide the confidence scores with 20 intervals as in Sec. 3.2. As shown in Fig. 7, CoIBA consistently outperforms all the baselines regardless of the confidence scores predicted with" + }, + { + "type": "image", + "bbox": [ + 0.529, + 0.381, + 0.704, + 0.48 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.536, + 0.482, + 0.7, + 0.493 + ], + "angle": 0, + "content": "(a) Ablation on departure layer" + }, + { + "type": "image", + "bbox": [ + 0.715, + 0.381, + 0.889, + 0.48 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.729, + 0.482, + 0.878, + 0.493 + ], + "angle": 0, + "content": "(b) Ablation on arrival layer" + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.494, + 0.906, + 0.591 + ], + "angle": 0, + "content": "Figure 8. Comparisons of different departure \\( s \\) and arrival \\( e \\) layer settings. We plot the hyper-parameters selected for CoIBA as a solid line. We compare the quantitative results of the discrepancy between AUC scores of insertion/deletion (ΔInsDel) with different intervals of confidence scores outputted by the model. The better attribution method achieves a higher score. We provide further results in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.598, + 0.907, + 0.689 + ], + "angle": 0, + "content": "each sample, predicted by the model. Along with IN-1k, we include the quantitative results of IN-A and IN-R. As both IN-A and IN-R datasets include the low-confident (difficult) samples compared to IN-1k, these quantitative results support the demonstration that CoIBA outperforms the correct performance with low-confident samples." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.698, + 0.633, + 0.713 + ], + "angle": 0, + "content": "4.6. Discussion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Ablation on Bottleneck-inserted Layers In this section, we confirm whether including multiple layers enhances the correctness of the resulting attribution map. Fig. 9 shows the comparison of the quantitative results by interpolating departure \\((s)\\) and arrival \\((e)\\) layers from earlier to deeper layers. The results demonstrate that increasing the number of layers to compute relevant information consistently enhances the correctness quality of an attribution map. Furthermore, CoIBA consistently outperforms IBA* with large margins, requiring only a single iteration. Thus, in CoIBA, various layers gladly reveal the omitted relevant information obtained from a specific layer." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "25172" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.109, + 0.091, + 0.285, + 0.184 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.111, + 0.185, + 0.277, + 0.197 + ], + "angle": 0, + "content": "(a) Discrepancy in mutual info." + }, + { + "type": "image", + "bbox": [ + 0.293, + 0.091, + 0.47, + 0.184 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.307, + 0.186, + 0.455, + 0.197 + ], + "angle": 0, + "content": "(b) Discrepancy in accuracy" + }, + { + "type": "image", + "bbox": [ + 0.108, + 0.199, + 0.285, + 0.298 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.183, + 0.301, + 0.393, + 0.313 + ], + "angle": 0, + "content": "(c) Ablation on universal damping ratio" + }, + { + "type": "image", + "bbox": [ + 0.291, + 0.199, + 0.467, + 0.298 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.315, + 0.484, + 0.385 + ], + "angle": 0, + "content": "Figure 9. Ablation study on layers to analyze universal damping ratio. Fig. 9(a) illustrates the change in mutual information for compression terms \\( I[Z_1; R_1] \\) in Eq. (8) and the Fig. 9(b) shows the change in accuracy. We utilize the ViT-B-16/224 model for these experiments. The mutual info. denotes mutual information." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.399, + 0.484, + 0.672 + ], + "angle": 0, + "content": "Effectiveness of Universal Damping Ratio We investigate whether leveraging the universal damping ratio compensates for over-compression in Fig. 9(a) while amplifying the relevant information in Fig. 9(b). We compare the different numbers of targeted layers to compare the quantitative results. The results illustrated in Fig. 9(a) show that leveraging the universal damping ratio complements the overcompressed information that occurred in earlier layers, due to the delivered relevant information from deeper layers. Concurrently, as shown in Fig. 9(c), leveraging the universal damping ratio amplifies the relevant information term in Eq (8). Aligning with these results, compared by assigning individual damping ratios to each layer, our universal damping ratio enhances the correctness performance of CoIBA. Therefore, leveraging the universal damping ratio to targeted multiple layers significantly enhances the correctness performance of CoIBA by compensating for the over-compression and amplifying the relevant information." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.675, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Effectiveness of Variational Upper Bound We demonstrate the effectiveness of our variational upper bound by measuring correctness performance and whether the relevancy of different layers is fairly reflected in attributions. We compare quantitative results obtained by leveraging a variational upper bound (Eq. (8)) versus linearly combining layer-wise mutual information (Eq. (4)) in Fig. 10(a). As shown in the results, leveraging our variational upper bound is better at correctly identifying feature importance, regardless of confidence scores. In addition to this result, we confirm whether the relevant information of each layer is fairly reflected in the attribution map by comparing similarities. Concretely, we compare the similarity of both settings (Eq. (4) and (8)) with layer-specific relevant information computed by iterating IBA with channel uniform per" + }, + { + "type": "image", + "bbox": [ + 0.536, + 0.091, + 0.71, + 0.188 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.55, + 0.19, + 0.696, + 0.201 + ], + "angle": 0, + "content": "(a) Faithfulness assessment" + }, + { + "type": "image", + "bbox": [ + 0.713, + 0.091, + 0.887, + 0.176 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.725, + 0.19, + 0.872, + 0.202 + ], + "angle": 0, + "content": "(b) Similarity measurement" + }, + { + "type": "image", + "bbox": [ + 0.534, + 0.203, + 0.709, + 0.299 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.591, + 0.302, + 0.83, + 0.314 + ], + "angle": 0, + "content": "(c) Ablation on channel uniform perturbation" + }, + { + "type": "image", + "bbox": [ + 0.715, + 0.203, + 0.888, + 0.299 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.318, + 0.906, + 0.387 + ], + "angle": 0, + "content": "Figure 10. Quantitative analysis on CoIBA. We report the results divided by the confidence score computed by the model. For the comparison, we include the ViT-B model pre-trained with CLIP in Fig. 10(c). We filled the discrepancy in performances with blue color in Fig. 10(c). The solid line indicates the results of CoIBA." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.394, + 0.906, + 0.485 + ], + "angle": 0, + "content": "turbation, denoted as \\(\\mathrm{IBA}^*\\). We employ CKA [19] as a similarity metric. As shown in Fig. 10(b), leveraging our upper bound yields the comprehensive relevant information of all the layers. This result demonstrates that our variational upper bound encourages the resulting comprehensive relevant information to fairly reflect the relevancy of each layer." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.485, + 0.906, + 0.668 + ], + "angle": 0, + "content": "Effectiveness of Uniform Channel Perturbation We demonstrate the effectiveness of uniform channel perturbation by comparing the correctness performance of uniform (unif.) and non-uniform (non-unif.) channel perturbation strategies. As shown in Fig. 10(c), the uniform channel perturbation consistently enhances the correctness performance of IBA for low-confident samples. In addition to this result, uniform channel perturbation significantly enlarges the correctness of resulting attribution maps, regardless of the confidence scores. These enhanced correctness performances demonstrate the effectiveness of our uniform channel perturbation, regardless of the method." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.68, + 0.634, + 0.695 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In this paper, we introduce CoIBA, which reveals the comprehensive relevant information to produce an attribution map, revealing the omitted relevancy in IBA with a theoretical guarantee. CoIBA shares the universal damping ratio to compensate for the over-compressed information, delivering relevancy among the bottleneck-inserted layers. CoIBA clearly judges the importance of each token by leveraging uniform channel perturbation. We leverage variational approximation to upper bound the information to ensure the eliminated activations are not necessary for the bottleneck-inserted layers to make a decision. CoIBA demonstrates a substantial improvement over existing methods in numerous experiments and discussions." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "25173" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.251, + 0.108 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.484, + 0.252 + ], + "angle": 0, + "content": "This work was supported by the Institute of Information & Communications Technology Planning & Evaluation (IITP) grant, funded by the Korea government (MSIT) (No. RS-2019-II190079, Artificial Intelligence Graduate School Program (Korea University), IITP-2025-RS-2024-00436857, ITRC (Information Technology Research Center), and No. RS-2022-II220984, Development of Artificial Intelligence Technology for Personalized Plug-and-Play Explanation and Verification of Explanation)." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.264, + 0.188, + 0.279 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.288, + 0.484, + 0.344 + ], + "angle": 0, + "content": "[1] Samira Abnar and Willem Zuidema. Quantifying attention flow in transformers. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 4190-4197, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.344, + 0.484, + 0.397 + ], + "angle": 0, + "content": "[2] Alexander A Alemi, Ian Fischer, Joshua V Dillon, and Kevin Murphy. Deep variational information bottleneck. In International Conference on Learning Representations, 2017. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.401, + 0.484, + 0.455 + ], + "angle": 0, + "content": "[3] Oren Barkan, Yuval Asher, Amit Eshel, Noam Koenigstein, et al. Visual explanations via iterated integrated attributions. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2073-2084, 2023. 1, 2, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.457, + 0.484, + 0.566 + ], + "angle": 0, + "content": "[4] Alexander Binder, Grégoire Montavon, Sebastian Lapuschkin, Klaus-Robert Müller, and Wojciech Samek. Layer-wise relevance propagation for neural networks with local renormalization layers. In Artificial Neural Networks and Machine Learning-ICANN 2016: 25th International Conference on Artificial Neural Networks, Barcelona, Spain, September 6-9, 2016, Proceedings, Part II 25, pages 63-71. Springer, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.568, + 0.484, + 0.636 + ], + "angle": 0, + "content": "[5] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9650-9660, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.638, + 0.484, + 0.704 + ], + "angle": 0, + "content": "[6] Hila Chefer, Shir Gur, and Lior Wolf. Generic attention-model explainability for interpreting bi-modal and encoder-decoder transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 397-406, 2021. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.707, + 0.484, + 0.761 + ], + "angle": 0, + "content": "[7] Hila Chefer, Shir Gur, and Lior Wolf. Transformer interpretability beyond attention visualization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 782-791, 2021. 1, 2, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.763, + 0.484, + 0.817 + ], + "angle": 0, + "content": "[8] Jiamin Chen, Xuhong Li, Lei Yu, Dejing Dou, and Haoyi Xiong. Beyond intuition: Rethinking token attributions inside transformers. Transactions on Machine Learning Research, 2022. 1, 2, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.818, + 0.484, + 0.859 + ], + "angle": 0, + "content": "[9] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, pages 248-255. IEEE, 2009. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.86, + 0.483, + 0.9 + ], + "angle": 0, + "content": "[10] Finale Doshi-Velez and Been Kim. Towards a rigorous science of interpretable machine learning. arXiv preprint arXiv:1702.08608, 2017. 1" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.288, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.176 + ], + "angle": 0, + "content": "[11] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, G Heigold, S Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2020. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.179, + 0.906, + 0.245 + ], + "angle": 0, + "content": "[12] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dólár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16000-16009, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.249, + 0.906, + 0.331 + ], + "angle": 0, + "content": "[13] Dan Hendrycks, Steven Basart, Norman Mu, Saurav Kadavath, Frank Wang, Evan Dorundo, Rahul Desai, Tyler Zhu, Samyak Parajuli, Mike Guo, et al. The many faces of robustness: A critical analysis of out-of-distribution generalization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8340-8349, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.333, + 0.906, + 0.387 + ], + "angle": 0, + "content": "[14] Dan Hendrycks, Kevin Zhao, Steven Basart, Jacob Steinhardt, and Dawn Song. Natural adversarial examples. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15262-15271, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.39, + 0.906, + 0.458 + ], + "angle": 0, + "content": "[15] Robin Hesse, Simone Schaub-Meyer, and Stefan Roth. Funnybirds: A synthetic vision dataset for a part-based analysis of explainable ai methods. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3981-3991, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.461, + 0.906, + 0.515 + ], + "angle": 0, + "content": "[16] Sara Hooker, Dumitru Erhan, Pieter-Jan Kindermans, and Been Kim. A benchmark for interpretability methods in deep neural networks. Advances in Neural Information Processing Systems, 32, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.518, + 0.906, + 0.557 + ], + "angle": 0, + "content": "[17] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.56, + 0.906, + 0.587 + ], + "angle": 0, + "content": "[18] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.59, + 0.906, + 0.644 + ], + "angle": 0, + "content": "[19] Simon Kornblith, Mohammad Norouzi, Honglak Lee, and Geoffrey Hinton. Similarity of neural network representations revisited. In International Conference on Machine Learning, pages 3519-3529. PMLR, 2019. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.647, + 0.906, + 0.715 + ], + "angle": 0, + "content": "[20] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.717, + 0.906, + 0.786 + ], + "angle": 0, + "content": "[21] Ze Liu, Han Hu, Yutong Lin, Zhuliang Yao, Zhenda Xie, Yixuan Wei, Jia Ning, Yue Cao, Zheng Zhang, Li Dong, et al. Swin transformer v2: Scaling up capacity and resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12009-12019, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.788, + 0.906, + 0.829 + ], + "angle": 0, + "content": "[22] Clemens Otte. Safe and interpretable machine learning: a methodological review. Computational Intelligence in Intelligent Data Analysis, pages 111-122, 2013. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.831, + 0.906, + 0.871 + ], + "angle": 0, + "content": "[23] Vitali Petsiuk, Abir Das, and Kate Saenko. Rise: Randomized input sampling for explanation of black-box models. British Machine Vision Conference, 2018. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[24] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry," + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.093, + 0.906, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "25174" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.148 + ], + "angle": 0, + "content": "Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 1, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.151, + 0.483, + 0.219 + ], + "angle": 0, + "content": "[25] Yao Rong, Tobias Leemann, Vadim Borisov, Gjergji Kasneci, and Enkelejda Kasneci. A consistent and efficient evaluation strategy for attribution methods. In International Conference on Machine Learning, pages 18770-18795. PMLR, 2022. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.222, + 0.482, + 0.276 + ], + "angle": 0, + "content": "[26] Wojciech Samek and Klaus-Robert Müller. Towards explainable artificial intelligence. Explainable AI: interpreting, explaining and visualizing deep learning, pages 5-22, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.281, + 0.482, + 0.336 + ], + "angle": 0, + "content": "[27] Karl Schulz, Leon Sixt, Federico Tombari, and Tim Landgraf. Restricting the flow: Information bottlenecks for attribution. In International Conference on Learning Representations, 2020. 1, 2, 3, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.339, + 0.482, + 0.421 + ], + "angle": 0, + "content": "[28] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 618-626, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.424, + 0.482, + 0.493 + ], + "angle": 0, + "content": "[29] Andreas Peter Steiner, Alexander Kolesnikov, Xiaohua Zhai, Ross Wightman, Jakob Uszkoreit, and Lucas Beyer. How to train your vit? data, augmentation, and regularization in vision transformers. Transactions on Machine Learning Research, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.497, + 0.482, + 0.538 + ], + "angle": 0, + "content": "[30] Mukund Sundararajan, Ankur Taly, and Qiqi Yan. Axiomatic attribution for deep networks. In International Conference on Machine Learning, pages 3319-3328. PMLR, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.541, + 0.482, + 0.583 + ], + "angle": 0, + "content": "[31] Naftali Tishby, Fernando C Pereira, and William Bialek. The information bottleneck method. arXiv preprint physics/0004057, 2000. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.585, + 0.482, + 0.655 + ], + "angle": 0, + "content": "[32] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In International Conference on Machine Learning, pages 10347-10357. PMLR, 2021. 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.657, + 0.482, + 0.711 + ], + "angle": 0, + "content": "[33] Hugo Touvron, Matthieu Cord, and Hervé Jégou. Deit iii: Revenge of the vit. In Proceedings of the European Conference on Computer Vision, pages 516-533. Springer, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.715, + 0.482, + 0.77 + ], + "angle": 0, + "content": "[34] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing, 13(4):600-612, 2004. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.773, + 0.482, + 0.842 + ], + "angle": 0, + "content": "[35] Weiyan Xie, Xiao-Hui Li, Caleb Chen Cao, and Nevin L Zhang. Vit-cx: causal explanation of vision transformers. In Proceedings of the Thirty-Second International Joint Conference on Artificial Intelligence, pages 1569–1577, 2023. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[36] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11975-11986, 2023. 1" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.162 + ], + "angle": 0, + "content": "[37] Yang Zhang, Ashkan Khakzar, Yawei Li, Azade Farshad, Seong Tae Kim, and Nassir Navab. Fine-grained neural network explanation by identifying input features with predictive information. Advances in Neural Information Processing Systems, 34:20040-20051, 2021. 1, 2, 3, 5" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "25175" + } + ] +] \ No newline at end of file diff --git a/2025/Comprehensive Information Bottleneck for Unveiling Universal Attribution to Interpret Vision Transformers/abcc8427-3423-466d-afdc-b93969a43358_origin.pdf b/2025/Comprehensive Information Bottleneck for Unveiling Universal Attribution to Interpret Vision Transformers/abcc8427-3423-466d-afdc-b93969a43358_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..c18c256ec9e36e3ceede4c1a1ea170c8a9a9bdca --- /dev/null +++ b/2025/Comprehensive Information Bottleneck for Unveiling Universal Attribution to Interpret Vision Transformers/abcc8427-3423-466d-afdc-b93969a43358_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:31b36c638b41a44bcf3fc7666b88e41c0ec655299dad426c6b8d4470725764a6 +size 2115301 diff --git a/2025/Comprehensive Information Bottleneck for Unveiling Universal Attribution to Interpret Vision Transformers/full.md b/2025/Comprehensive Information Bottleneck for Unveiling Universal Attribution to Interpret Vision Transformers/full.md new file mode 100644 index 0000000000000000000000000000000000000000..82372beb9205a61c91ef3f69bcb119132648999b --- /dev/null +++ b/2025/Comprehensive Information Bottleneck for Unveiling Universal Attribution to Interpret Vision Transformers/full.md @@ -0,0 +1,345 @@ +# Comprehensive Information Bottleneck for Unveiling Universal Attribution to Interpret Vision Transformers + +Jung-Ho Hong Ho-Joong Kim Kyu-Sung Jeon Seong-Whan Lee* + +Dept. of Artificial Intelligence, Korea University, Seoul, Korea + +{jungho-hong, hjoong_kim, ksjeon, sw.lee}@korea.ac.kr + +# Abstract + +The feature attribution method reveals the contribution of input variables to the decision-making process to provide an attribution map for explanation. Existing methods grounded on the information bottleneck principle compute information in a specific layer to obtain attributions, compressing the features by injecting noise via a parametric damping ratio. However, the attribution obtained in a specific layer neglects evidence of the decision-making process distributed across layers. In this paper, we introduce a comprehensive information bottleneck (CoIBA), which discovers the relevant information in each targeted layer to explain the decision-making process. Our core idea is applying information bottleneck in multiple targeted layers to estimate the comprehensive information by sharing a parametric damping ratio across the layers. Leveraging this shared ratio complements the over-compressed information to discover the omitted clues of the decision by sharing the relevant information across the targeted layers. We suggest the variational approach to fairly reflect the relevant information of each layer by upper bounding layer-wise information. Therefore, CoIBA guarantees that the discarded activation is unnecessary in every targeted layer to make a decision. The extensive experimental results demonstrate the enhancement in faithfulness of the feature attributions provided by CoIBA. + +# 1. Introduction + +Vision transformer (ViT) achieves remarkable performance in diverse fundamental computer vision tasks, such as multi-modal [24, 36] and self-supervised learning [5, 12]. Despite this achievement, the black box nature stemming from the complex structure of ViT restricts the testability of the end-to-end system, impeding the failure diagnosis [10]. This limitation constrains the application of ViT in safety-critical areas where transparency and interpretability are considered important as well [22, 26]. Feature attribu + +![](images/46452c1e53fc94d13bba0dc97d43689415fcc1dd0fbd03415ecc6d177ee52654.jpg) + +![](images/6d86f523b511c3b6c21b19a26e75c1bb0918d32013203d3ab82e0eacb845f0e0.jpg) +(b) Cumulative number of each layer yielding optimal result per sample + +![](images/d4ae8871e3d459b793874d5beb169e432c1db15051852e1b8b9cf6a3c7380eaa.jpg) +Figure 1. The quantitative and qualitative comparisons of attribution maps generated from different layers. We illustrate the visualized attribution maps from the first to fourth columns in Fig. 1(a) utilizing IBA [27]. The text on the top of each figure indicates the index of the layer. We select the layer index of the large model with interval 2. The fifth column in Fig. 1(a) shows the similarity comparison between attributions of different layers. Fig. 1(b) illustrates the cumulative number of each layer that provides the best explanation per sample. The uniformly distributed data indicates the absence of the layer optimal for the explanation. + +tion methods [4, 28, 30] are introduced to reveal the evidence supporting the decision-making process. Many of the existing attribution methods designed to interpret convolutional neural networks show limited explainability to interpret ViT [7]. To address this limitation, feature attribution methods [3, 6-8, 35] tailored for interpreting ViT have been proposed, leveraging either attention weights computed in ViT or intermediate representations, such as the class token, to generate attribution maps. + +Although feature attribution methods have advanced in interpreting ViT, they lack theoretical guarantees that low-scored attributions are not necessary to the decision-making process [27]. The methods [27, 37] grounded on the information bottleneck principle [2, 31] address this limitation by guaranteeing the importance of attributions with a + +variational approximation. Information bottleneck for attribution (IBA) [27] proposes the model-agnostic method to compress the information unnecessary for the prediction by positioning the bottleneck layer into the target layer. InputIBA [37] succeeds in providing a high-resolution attribution map by directly measuring the information in the input domain, assuming the prior distribution with a generative model. However, existing information bottleneck-based methods face two main challenges. First, the resulting attribution map reflects the partial of the decision-making process as the restriction of the information flow is conducted in a specific targeted layer. This constraint provides inconsistent evidence for a single decision-making process, leading to unobvious interpretations. Lastly, no layer dominantly provides the most appropriate relevant information for explaining the decision-making process. + +Fig. 1 shows the aforementioned issues with quantitative and qualitative examples. In the first row in Fig. 1(a), the visualization derived from the second layer (blocks.2) highlights the barrel whereas the fifth (blocks.5) and eighth layers (blocks.8) highlight the top and bottom of the wheels, respectively. In the second row in Fig. 1(a), the visualization in the earlier layer highlights the person holding the racket, while the deeper layer progressively concentrates on the racket. To confirm the layer-wise discrepancy in attributions quantitatively, we measure SSIM [34] in the right row of Fig. 1(a). As the layer providing attributions distance farther, the dissimilarity in those attributions increases. Thus, computing the relevant information from isolated layer-specific information bottlenecks highlights different attributions for the same decision. Among the layers producing the attribution maps, there is no dominant layer that provides the most appropriate relevant information to explain the decision-making process. Fig. 1(b) compares the cumulative number of each layer yielding attributions with the optimal insertion and deletion scores per input sample. The results show a lack of the layer that produces the most appropriate relevant information for explanation. Thus, identifying the most faithful attribution from the layers requires a heuristic search due to the lack of criteria to distinguish the optimal layer. + +In this paper, we propose a novel comprehensive information bottleneck for attribution (CoIBA), which reveals the information relevant for each of the target layers to provide the attributions. Our core idea is to reveal comprehensive relevant information by eliminating unnecessary information in every targeted layer of the prediction. We estimate the comprehensive information by sharing a parameterized universal damping ratio across layers, thereby removing the requirements of additional heuristics. This sharing strategy compensates for the over-compressed information of individual layers by sharing information in each layer necessary for the decision. Since a channel el + +ement placed in the same location but different layers captures different features, we uniformly perturb all the channel elements in a single token representation, to handle the information of different layers with a shared parametric ratio. We suggest a variational upper bound to restrict information flowing in all the targeted layers. Our approach eliminates the heuristic search for balancing inconsistent information along the layers by suggesting a variational upper bound to fairly reflect the layer-wise relevant information. + +We conduct experiments to show the correctness performance of CoIBA utilizing various assessments. As the evaluations are solely approximations, i.e. no ground truth exists, we include FunnyBirds [15], which provide the ground truth and assess the comprehensive quality of attribution maps. To assess the feature importance assessment, we leverage insertion/deletion [23] and remove-and-debias (ROAD) [25], which evaluate the faithfulness of attribution maps. In addition, we scrutinize the faithfulness of CoIBA by analyzing the confident-aware assessment. We discuss the effectiveness of our method with a rigorous discussion. + +# 2. Related Works + +Explanation Methods for Vision Transformer Existing explanatory methods designed to interpret ViT provide the attribution map utilizing the latent representation and the corresponding gradients. However, these methods show limited explainability while being adapted to ViT [7]. To address this limitation, Rollout [1] linearly combines the attention weights across the layers to compute the attribution map. Trans-attr [7] provides the class-discriminative attribution map by constructing the layer-wise relevance propagation rule for ViT. Generic [6] produces the attribution map with a generalized procedure to interpret diversified transformer models, leveraging the gradient of the attention map to produce the class-discriminative attribution maps. IIA [3] introduces the iterative integration across the input image, leveraging the internal representations processed by the model and their gradients. ViT-CX [35] utilizes the patch embeddings and measures causal impacts to provide attributions. Beyond [8] unfolds attention blocks with the chain rule between final prediction, CLS, and tokens to obtain token contribution. Existing propagation-based approaches require a specific implementation to weigh the contributions of input variables. In contrast to this constraint, CoIBA does not require implementing additional procedures as same as IBA. Furthermore, aligning with the information bottleneck principle, CoIBA guarantees the highlighted attributions are important in all targeted layers for the decision. + +Information Bottleneck Approach The information bottleneck principle is broadly utilized to obtain information in the activation necessary for the specific objective. To this end, the information bottleneck-based methods compress the information unnecessary for the objective while + +![](images/04e84f49a950cd4233f5fe0f6451bc1cf5b7ed1e0f1c422923f55be756b332a8.jpg) +Figure 2. Main difference between IBA and CoIBA. IBA obtains the relevant information in the specific bottleneck-inserted layer. Thus, obtaining layer-wise attribution maps necessarily requires iterative running over the layers. In contrast, CoIBA seeks the layer-wise relevant information from each bottleneck-inserted layer with a universal damping ratio. + +maintaining the relevant information. Given the challenges of directly estimating the amount of information, a variational approach [18] is used to approximate the intractable posteriors and marginals. The Deep Variational Information Bottleneck [2] method employs the information bottleneck principle in deep networks, using a variational lower bound alongside the reparameterization trick. Referring to this strategy, IBA [27] adapts the information bottleneck principles to restrict the information flow in an intermediate layer and provides the relevant information as an attribution map. InputIBA [37] provides high-resolution attributions by directly estimating relevant information in the input domain. However, existing methods compute relevant information solely considering a specific layer, overlooking the information necessary in the different layers. In contrast to this process, CoIBA obtains comprehensive relevant information from the multiple layers, highlighting the evidence important in every targeted layer to make a decision. + +# 3. Method + +The main goal of CoIBA is to produce an attribution map by reflecting the information relevant to the targeted layers, which is opposed to the IBA as shown in Fig. 2. To show this procedure, we first introduce IBA in Sec. 3.1 and discuss the limitations of the existing information bottleneck-based approach in Sec. 3.2. Then, in Sec. 3.3, we describe the overall procedure of CoIBA. Finally, we introduce a variational upper bound, which enhances the reflection of layer-wise relevant information in Sec. 3.4. + +![](images/ea662166572a569ff49afc338896d996e47bafc11a01d53f5718baed5727779c.jpg) + +![](images/a1b5c281fd2885172ccf4b287eb5aee8be10200195a305605b0fdfd1f631a493.jpg) + +![](images/e44fd811d68c2e9f304fa660ae85e26f7f7aaa706c1009dfa03b1937adafd59f.jpg) +Figure 3. Quantitative and qualitative results of IBA*. In the first row, the horizontal line indicates the performance of layers-specific relevant information, i.e., IBA. We report the difference in insertion and deletion scores ( $\Delta$ InsDel). The higher score is better for this metric. The visualized attribution maps are illustrated in the second row obtained from VIT-B-16/224 [11]. + +![](images/c5cb7cb051d773a4a7eb0fd1f0472968941d350393a39855bbb2059cdcce1296.jpg) + +![](images/673112c63ece870c5c754456fdf2df0cb39ce5a23b48b8a32fe0c6f59646ea08.jpg) + +![](images/7729a9370a5731c56cff819cbc8a3cf491d64c67c614d0774d4687202778ac19.jpg) + +# 3.1. Background - IBA + +IBA adopts an information bottleneck principle to produce relevant information to provide an attribution map. To this end, in $l$ -th layer, IBA computes the bottleneck representation $Z_{l}$ as follows: + +$$ +Z _ {l} = \lambda_ {l} R _ {l} + (1 - \lambda_ {l}) \epsilon_ {l}. \tag {1} +$$ + +Here, the damping ratio $\lambda_{l}$ is a learnable parameter assigned at the bottleneck layer, manipulating the degree of damping signal between the activation $R_{l}$ and the independent noise $\epsilon_{l}$ . The independent noise $\epsilon_{l}$ is sampled from the Gaussian distribution. To maintain the statistic of internal representation, the Gaussian distribution shares the mean $\mu_{R_l}$ and variance $\sigma_{R_l}$ with the representation $R_{l}$ . + +IBA minimizes the shared information between the activation $R_{l}$ and bottleneck variable $Z_{l}$ while maximizing the shared information between the bottleneck variable $Z_{l}$ and the label $Y$ as follows: + +$$ +\max _ {\lambda_ {l}} I \left[ Z _ {l}; Y \right] - \beta I \left[ R _ {l}, Z _ {l} \right]. \tag {2} +$$ + +Here, $\beta$ is a hyperparameter manipulating the trade-off between compression and relevancy in $l$ -th layer. + +# 3.2. Motivation + +IBA selects a specific layer to obtain relevant information to provide an attribution map. However, the relevant information of a specific single layer does not reflect the overall evidence required in the sequence of layers to make a decision. To demonstrate our insistence, we illustrate the quantitative comparison of the attribution maps obtained in a specific layer and multiple layers in Fig. 3, leveraging insertion/deletion [23]. We iterate IBA multiple times along the layers to obtain the set of individual relevant information + +![](images/84254202b7b2cdc9d21b3e58547095be842d63f0a1b84370e4a672b1a71abc71.jpg) +eaaee +Class: Hyena + +![](images/3f70ab71a742248c92c64b7aa2499ea94e3427761959d99761d4ca65ec3e6e05.jpg) +Confidence: 85.87 + +![](images/210a74117a0c02bb86604554dd1a882f30f28c22a4e44fc636ce78f15699e53b.jpg) +Hard Sample +Class: Parallel Bars +Confidence: 19.71 + +![](images/f04d2669558c4dcf691ac1608cc4d69796554afefbb331adb53880adc7b975ec.jpg) + +![](images/87dafdfb7dc66a24a577cb325728f387ff0d35db09075ba65f525afe39b9d800.jpg) +Figure 4. The quantitative comparisons of explanation methods in high-confident (easy) and low-confident (hard) input samples. The confidence score is the softmax probability of the target class predicted by the model. We compare the information-bottleneck-based methods including IBA [27], IBA*, and CoIBA (marked with a circle), and the propagation-based explanation baselines including Beyond [8], IIA [3], and Chefer-LRP [7] methods (marked with a triangle). We utilize DeiT-B-16/224 [32] for the evaluation. For insertion and deletion, higher and lower scores are better, respectively. + +for each layer. After that, we compute a linear combination over the set of individual relevant information. We denote this procedure as $\mathrm{IBA}^*$ . As shown in the results, the attribution maps produced by $\mathrm{IBA}^*$ archive better scores compared to the IBA. To scrutinize the quantitative results between $\mathrm{IBA}^*$ and IBA, we split the insertion/deletion scores by the confidence scores of the model in Fig. 4. As shown in the results, the increase in performance is only observed in high-confident samples, unlike in low-confident samples. Therefore, even iterating IBA, which requires high computational cost, shows limited performance enhancement according to the difficulty of each sample. + +# 3.3. Comprehensive Information Bottleneck + +Our comprehensive information bottleneck (CoIBA) inserts the bottleneck layer into the sequence of targeted layers to restrict the information flowing in the intermediate layer. Within the bottleneck layer, CoIBA restricts the information with the learnable universal damping ratio shared across the bottleneck-inserted layers. Our goal is to optimize the universal damping ratio to obtain the layer-wise relevant information while fulfilling the information bottleneck objective. The universal damping ratio enables the attributions to identify distributed relevant information omitted by IBA, which restricts information in a layer-specific manner. + +Formulating Information Restriction The model is trained to make a decision $\tilde{Y} = f(X)$ to predict the ground truth $Y$ by taking input $X$ . To make a decision, the input sample $X$ is contextualized by passing through the sequence of intermediate layers. Here, we insert the bottle- + +neck layer into the targeted layer. The role of the bottleneck layer is to restrict the information by dampening the signal of the intermediate representation passing through the $L$ bottleneck-inserted layers. We perturb the intermediate representations in the $l$ -th layer to dampen the signal. To dampen the signal, we inject the independent noise $\epsilon_{l}$ into the intermediate bottleneck representation $R_{l}^{\prime}$ , which is computed from the bottleneck variables $Z_{l-1}$ of the preceding layer. Note that the first element of intermediate bottleneck representations is the non-perturbed intermediate representation, i.e., $R_{1}^{\prime} = R_{1}$ . Thus, we obtain the bottleneck representation $Z_{l}$ as follows: + +$$ +Z _ {l} = \lambda R _ {l} ^ {\prime} + (1 - \lambda) \epsilon_ {l}, \tag {3} +$$ + +Here, a universal damping ratio $\lambda \in \mathbb{R}^{P\times 1}$ manipulates the degree of perturbation for each token, denoting $P$ as the number of patches. Thus, $\lambda$ is ranged from 0 to 1, allowing the propagation of the signal when $\lambda = 1$ otherwise blocking it when $\lambda = 0$ . As the universal damping ratio $\lambda$ is consistently adapted to the layers, we omit the layer index in the notation. We compute $\lambda = \mathrm{sigmoid}(\alpha)$ by passing the trainable parameter $\alpha$ , which is initialized with 5, to the sigmoid function. The universal damping ratio uniformly perturbs the signal along the channel dimension. This setting avoids handling the neurons that capture different features with the same coefficient and induces concentration on the token's importance. We empirically show this uniform perturbation across the channels enhances the faithfulness of resulting relevant information. The independent noise $\epsilon_{l}$ shares the dimension and is sampled from the Gaussian distribution with the mean $\mu_{R_l}$ and variance $\sigma_{R_l}^2$ , such that $\epsilon_{l}\sim \mathcal{N}(\mu_{R_l},\sigma_{R_l}^2)$ . To minimize statistical differences between the bottleneck variable and non-perturbed activations, we leverage the mean $\mu_{R_l}$ and variance $\sigma_{R_l}^2$ from those of the non-perturbed activations. + +Information Bottleneck Objective The information bottleneck principle desires to obtain information necessary for the objective while compressing irrelevant information from the signal. Compressing irrelevant information is performed by minimizing the shared information between the activation of a bottleneck-inserted network and the bottleneck representation. Building upon this, we compress the amount of information by minimizing the mutual information between internal sequences of bottleneck representations while maximizing the mutual information between the ground truth and bottleneck representations. Combining two aspects with the ground truth $Y$ , we maximize the mutual information between the bottleneck variable and the ground truth while minimizing the mutual information between bottleneck representations as follows: + +$$ +\max _ {\lambda} I \left[ Z _ {L}; Y \right] - \frac {1}{L} \left(\sum_ {l = 1} ^ {L} \beta_ {l} I \left[ Z _ {l - 1}; Z _ {l} \right]\right). \tag {4} +$$ + +![](images/5329bcb88a94f9db5a85eaceb195c1588ff663fb1b0ce6f34d6c648df6887191.jpg) +Figure 5. An overview of CoIBA. CoIBA restricts the information flow using a universal damping ratio while upper-bounding the information of each layer. After that, CoIBA aggregates the relevant information from each layer to produce an attribution map. + +Here, the trade-off parameter $\beta_{l}$ governs the degree of compression in each layer. Since the bottleneck layers are inserted into the sequence of layers, the compression term includes the accumulated sequence of mutual information of bottleneck variables $I[Z_{l-1};Z_l]$ . Here, the $Z_0$ indicates the original intermediate representation as the first element, such that $Z_0 = R_1$ . We maximize the relevancy term $I[Z_L;Y]$ while compression to maintain the shared information between bottleneck variables and the ground truth. We maximize this term by minimizing the cross-entropy loss $\mathcal{L}_{ce}$ . + +Utilizing KL divergence, the mutual information between internal representations $I[Z_{l-1}; Z_l]$ is formulated by: + +$$ +I \left[ Z _ {l - 1}; Z _ {l} \right] = \mathbb {E} _ {Z _ {l - 1}} \left[ D _ {K L} \left[ P \left(Z _ {l} \mid Z _ {l - 1}\right) \right\vert \left| P \left(Z _ {l}\right) \right] \right] \tag {5} +$$ + +However, the direct estimation of the prior distribution $P(Z_{l})$ is intractable as it requires the integration of bottleneck representations $P(Z_{l}) = \int_{Z_{l - 1}}P(Z_{l}|Z_{l - 1})P(Z_{l - 1})dZ_{l - 1}$ of the corresponding $l$ -th layer. Thus, we assume the prior distribution $Q(Z_{l})$ as Gaussian distribution $\mathcal{N}(\mu_{R_l},\sigma_{R_l}^2)$ , reconstituting the mutual information computation as follows: + +$$ +I \left[ Z _ {l}; Z _ {l - 1} \right] \leq \mathbb {E} _ {Z _ {l - 1}} \left[ D _ {K L} \left[ P \left(Z _ {l} \mid Z _ {l - 1}\right) \right\vert \left| Q \left(Z _ {l}\right) \right] \right]. \tag {6} +$$ + +Since assuming the prior distribution with Gaussian distribution only overestimates the mutual information, minimizing this term suppresses the mutual information of the compression term. + +# 3.4. Variational Upper Bound + +As discussed in Sec. 4.6, the relevant information obtained by suppressing layer-wise mutual information suffers from reflecting layer-wise relevant information due to the inconsistent amount of information along the layers. Thus, obtaining the sequence of trade-off parameters $\{\beta_l\}_{l = 1}^L$ requires an extensive heuristic search to balance the degree + +of compression for each layer. To overcome this issue, we suggest the variational upper bound for CoIBA, referring to the two following points. First, since there is no additional information propagated during the forwarding pass, the mutual information of intermediate layers would not be greater than the mutual information between the model input and output. Second, the iterated noise-injection procedure diminishes the mutual information among the internal and bottleneck representations, i.e., $I[Z_{l};Z_{l + 1}] \leq I[Z_{l - 1};Z_l]$ . Building upon these bases, we establish an upper bound that encompasses the combined mutual information from all subsequent layers as follows: + +$$ +I \left[ R _ {1}; Z _ {1} \right] \geq \frac {1}{L} \left(\sum_ {l = 1} ^ {L} I \left[ Z _ {l - 1}; Z _ {l} \right]\right). \tag {7} +$$ + +Accordingly, we reconstitute the objective to be a simplified formula, leaving only a single hyperparameter $\beta$ as follows: + +$$ +\max _ {\lambda} I \left[ Z _ {L}; Y \right] - \beta I \left[ R _ {1}; Z _ {1} \right]. \tag {8} +$$ + +To compress the subsequent layers, the simplified objective necessitates calculating only the mutual information of the first layer $I[R_1;Z_1]$ . This term solely utilizes the non-perturbed intermediate and bottleneck representations of the first layer, simplifying the objective calculation. In terms of relevant information computation, referring to the data processing inequality, the inequality $I[Y;Z_l]\leq I[Y;Z_{l - 1}]$ holds. Thereby, in contrast to the layer-specific information bottleneck, which overestimates mutual information in the earlier layers during compression [37], CoIBA relieves the overestimation. This is because, as discussed in Sec. 4.6, the relevant information is compensated as the sequence of bottleneck variables joins the objective computation. + +![](images/0ba7fcfd7dfa2e1358c8ffea180b50930af878b8e1d18f727a621b398f5f047d.jpg) +Original Image + +![](images/5dedc72a277d60fc8f56dca2139d50e1956c2f02f5427820d8573d1a3feec967.jpg) + +![](images/7df66507cdc9f2b5db097cb5bedcb50b10da014aa82e1adf00d3bf2d7dacec33.jpg) +Figure 6. Quantitative comprehensive assessment on FunnyBirds experiment: The abbreviations Com, Cor, and Con indicate Completeness, correctness, and contrastivity. The center score indicates the comprehensive result of enumerated aspects. Acc. and B.I. indicate accuracy and background independence, respectively. We provide detailed numeric in the supplementary material. + +![](images/40447f6b45fbf2d809da226459508b189ce1ac6933ff8e11da8d4cde6378a025.jpg) + +![](images/b57b480c4902f9a344f5915bba0b67219765a4f14ba1714da3dd677bb372b8d1.jpg) + +![](images/892b1921c46b9d374b3387ad2068b6224b3f25267fdc6d88fbf1a7026afc930b.jpg) + +![](images/bbe59162935a3cd7e403b8fefc9c607cdddf12c8ba9318e7761230279542e123.jpg) + +![](images/c79670731198fe2b38be3294e6fbad593bbc07532e6b65023c3d72b0230be279.jpg) + +![](images/e9269a4742934d318b2423986857a620cc128cf159f584edd8f09ec67d2ca5a4.jpg) + +![](images/57ffe9db5e70f8ab8c5e9cb05ce22613b272691982b7cf62f77e0123fdfa8f7a.jpg) + +![](images/ccde3b81382f1229c3d4e2edeb1bc7715770bf0d6667ad4d4b5bb270d4c46d24.jpg) + +![](images/c22718f40152dac6e2c8674187f79fc659bb1de71e3ba08f64c6142b1026d3be.jpg) + +![](images/6ac53393064b023c4ed9766f675f34fb897f4dbf1f4a82d3394efb34c3f6620f.jpg) + +![](images/a3d3a2e0e3637f5938f4e9a4af2bf8d23afb07b0d4f5e17a342716fbe36c8bde.jpg) + +![](images/9b6881e286591e9db51bf99ee5ee2220d828333ebdd128cd02aad2254639e28b.jpg) + +
VariantModelChefer-LRP [7]Generic [6]IIA [3]ViT-CX [35]IBA [27]Beyond [8]ColBA
ViTViT*-S-16/224-14.56/56.0813.69/57.3114.98/56.6513.56/58.6413.50/58.4711.08/63.35
ViT*-T-16/224-9.00/48.958.50/49.7512.08/45.098.02/49.2211.26/44.076.97/54.23
ViT-B-16/22417.67/56.6917.92/57.0316.56/58.716.94/58.1717.23/59.0615.84/60.3713.01/62.58
ViT-L-16/22420.97/53.8120.17/55.5420.14/55.2219.76/57.3820.88/54.8518.99/56.5515.40/61.14
DeiTDeiT-S-16/22412.76/47.5112.10/48.6410.89/49.7617.57/44.4311.43/47.8711.13/48.889.55/53.38
DeiT-B-16/22414.63/48.7915.10/49.4113.69/50.5118.21/46.9713.95/50.1814.45/48.8811.79/53.96
DeiT3-B-16/224-15.70/52.4215.38/52.7018.88/50.4015.07/53.3214.71/54.0712.97/56.54
DeiT3-L-16/224-22.10/62.7317.75/55.9326.11/59.6021.01/65.6019.34/64.2717.64/67.20
SwinSwin-B-33.76/42.94--18.03/52.8123.73/50.3917.09/54.80
Swin2-B-34.71/47.16--20.94/53.8531.32/49.4418.88/55.77
+ +Table 1. Quantitative feature importance assessment on insertion ↑/deletion ↓. (*) denotes the ViT trained with strong regularization techniques [29]. We underline the state-of-the-art performance among the baselines. + +# 4. Experiments + +# 4.1. Setup + +Architectures We select variants of ViT for our experiment, including original ViT [11] and DeiT models [32, 33], trained with ImageNet-21k (IN-21k) and ImageNet-1k (IN-1k), respectively. We include the variants in depth interpolated from ViT-T, ViT-S, ViT-B, ViT-L, and ViT-H models. For the ViT-T and ViT-S models, we utilize the model pretrained with massive regularization techniques [29]. The ViT-H model is pre-trained with CLIP [24]. We present the settings of the ViT-B model with patch size 16 and image resolution 224 as ViT-B-16/224. We include Swin transformers [20, 21], demonstrating the generalized ability of CoIBA in multi-scale features processing. The details of settings and further results about convolutional neural networks are included in the supplementary material. + +Feature Attribution Methods We compare the various feature attribution methods including Chefer-LRP [7], Generic [6], IIA [3], ViT-CX [35], Beyond [8], and IBA [27] as baselines. We select $\beta = 10$ for IBA and 6-th layer to insert bottleneck. For CoIBA, we select tradeoff hyper-parameter $\beta$ as 1. We insert the bottleneck layer from $s$ -th departure to $e$ -th arrival layers, which are chosen, 4, and 12, respectively, for ViT-B. For the optimization, we set the learning rate, batch size, and optimizer as 1, 10, and Adam [17]. The additional hyperparameter settings + +are included in the supplementary material. For Swin transformers, we only report the explanation methods that do not require methodology modification, including Generic, ViTCX, IBA, Beyond, and CoIBA. We utilize RTX A6000 GPU for all the experiments. + +Datasets We utilize ImageNet-1k [9] (IN-1k), ImageNet-A (IN-A) [14] and ImageNet-R (IN-R) [13] validation datasets for the experiments. We leverage IN-A and IN-R to conduct the difficult-aware analysis. + +# 4.2. FunnyBirds Assessment + +The FunnyBirds [15] experiment provides the ground truth for the evaluation. With the help of ground truth, the Funnybird framework assesses the quality of the attribution map from three perspectives: completeness, correctness, and contrastivity. Fig. 6 illustrates the qualitative and quantitative results of the Funnybirds experiment. As shown in the results, CoIBA outperforms all the baselines, including the propagation or gradient-based approaches. Furthermore, in contrast to IBA, CoIBA obtains an attribution map with significantly enhanced contrastivity. + +# 4.3. Insertion/Deletion + +Insertion/deletion [23] measures the correctness of feature importance highlighted by an attribution map. The insertion/deletion method gradually inserts or deletes informative pixels in ascending order of attribution to compute the area under the curve scores. In insertion/deletion, methods + +
ModelChefer-LRP [7]Generic [6]IIA [3]ViT-CX [35]IBA [27]Beyond [8]CoIBA
ViT-B-16/22420.87/64.3722.15/65.3119.58/66.1919.04/65.3620.99/68.8017.93/68.4416.63/73.68
DeiT-B-16/22415.31/59.5516.76/59.7814.74/61.2118.37/57.4415.62/61.5315.37/60.0611.59/67.12
ViT-L-16/22429.92/64.8728.31/66.5128.39/65.9224.30/67.2730.55/66.9725.47/66.6822.03/74.88
DeiT3-L-16/224-20.28/69.8719.72/70.0224.43/66.6019.56/71.2319.41/70.5215.02/76.13
ViT†-H-16/224-30.06/61.3432.27/60.9830.34/57.7028.30/62.2928.11/62.9623.38/66.95
+ +Table 2. Quantitative feature importance evaluation of ROAD (MoRF ↓ / LeRF ↑). The higher and lower scores indicate better attribution map quality for MoRF and LeRF, respectively. The underlined scores indicate the highest performance among the baselines. $(\dagger)$ denotes the ViT trained with CLIP [24]. + +![](images/38daf34b912ff5644bcfbf51b07b446c9bc5e6300ea98a1a586141d6339e058d.jpg) + +![](images/698acb6aba24360032fd53e45ce016c0a3292ed7b020b1202fc6fe6cd49f020c.jpg) + +![](images/24aa980031e7d7b42e2490a466ec45a1e67f8c8c6e8a32877ccc1bc2ca8b3850.jpg) +(a) IN-1k + +![](images/f7210558684cac70f3ec1b18c459686b2f034fd3b178f0227808215a8086920e.jpg) + +![](images/6e662039c23d65a71ac32957bfa9a9f93fde5c72a23a7fa178ffe0c945b5908d.jpg) +(b) IN-A + +![](images/6fab59fa92ab46768056ab1f3309f68e827e4aff583f411e7341b54908e3f22c.jpg) +(c) IN-R + +providing better correctness performance achieve higher scores for insertion and lower scores for deletion. We gradually deleted or inserted around every $3.5\%$ pixels of the input image and randomly sampled 6,000 samples from the IN-1k validation dataset for the assessment. As shown in Table 1, the attributions provided by CoIBA yield predominant correctness performance compared to baselines. Especially, the consistent increase in correctness from the Swin transformer demonstrates the generalizability of CoIBA. + +# 4.4. Remove and Debias + +The remove-and-debias (ROAD) [25] experiments measure the correctness of the attribution maps, addressing the class information leakage problem (i.e., the shape of the mask) observed in the remove-and-retrain (ROAR) [16] experiment. We average the quantitative scores over the 6,000 randomly selected images when 20, 40, 60, and $80\%$ of pixels are imputed in two perspectives: most relevant first (MoRF) and least relevant first (LeRF). As shown in Table 2, CoIBA outperforms existing methods in MoRF and LeRF experiments with a sizable gap. These results demonstrate that regardless of the imputation type, CoIBA outperforms all the baselines regardless of the model setting (patch size and depth) and pre-trained dataset. + +# 4.5. Difficulty-aware Correctness Assessment + +We compare the quantitative quality of attribution maps yielded by CoIBA and baseline methods in terms of difficulty-aware assessment, in addition to correctness assessments of overall samples. Leveraging the confidence scores, and the prediction probability of a target class, we divide the confidence scores with 20 intervals as in Sec. 3.2. As shown in Fig. 7, CoIBA consistently outperforms all the baselines regardless of the confidence scores predicted with + +![](images/ef0cc9d5cd981bbbfb5d8e55d314414e53f46e2bf1b593b33106c2e17cffaf6b.jpg) +Figure 7. Quantitative difficulty-aware correctness assessment on insertion/deletion and ROAD. We measure differences in insertion/deletion $(\Delta \mathrm{InsDel}\uparrow)$ and ROAD $(\Delta \mathrm{ROAD}\uparrow)$ scores. We fill the regions including low-confident samples and high-confident samples with red and blue, respectively, based on the prediction made by the model. The first to third columns illustrate the quantitative results of the IN-1k. The fourth and fifth columns include the results of IN-A and IN-R datasets, respectively. +(a) Ablation on departure layer +Figure 8. Comparisons of different departure $s$ and arrival $e$ layer settings. We plot the hyper-parameters selected for CoIBA as a solid line. We compare the quantitative results of the discrepancy between AUC scores of insertion/deletion (ΔInsDel) with different intervals of confidence scores outputted by the model. The better attribution method achieves a higher score. We provide further results in the supplementary material. + +![](images/523a6a810e60cff76795b6fa352d6d38ad0ae93ed88fbddb8baedbc86602ca14.jpg) +(b) Ablation on arrival layer + +each sample, predicted by the model. Along with IN-1k, we include the quantitative results of IN-A and IN-R. As both IN-A and IN-R datasets include the low-confident (difficult) samples compared to IN-1k, these quantitative results support the demonstration that CoIBA outperforms the correct performance with low-confident samples. + +# 4.6. Discussion + +Ablation on Bottleneck-inserted Layers In this section, we confirm whether including multiple layers enhances the correctness of the resulting attribution map. Fig. 9 shows the comparison of the quantitative results by interpolating departure $(s)$ and arrival $(e)$ layers from earlier to deeper layers. The results demonstrate that increasing the number of layers to compute relevant information consistently enhances the correctness quality of an attribution map. Furthermore, CoIBA consistently outperforms IBA* with large margins, requiring only a single iteration. Thus, in CoIBA, various layers gladly reveal the omitted relevant information obtained from a specific layer. + +![](images/0705a15d468e8d31d2b6f54a32231b26d8587cce7584e2228f9833cdfead120a.jpg) +(a) Discrepancy in mutual info. + +![](images/45776e79e5cabef43808a4176649a9522f46873be0e683faa1468f68954cbf0c.jpg) +(b) Discrepancy in accuracy + +![](images/0fb9e3779c4a5a2e316f8d14bf9c483815558a34cf70660d72d3520759534364.jpg) +(c) Ablation on universal damping ratio + +![](images/17e00121058d317341001b7ded8e68b8c2333da05f592254e15595c29e14f19f.jpg) +Figure 9. Ablation study on layers to analyze universal damping ratio. Fig. 9(a) illustrates the change in mutual information for compression terms $I[Z_1; R_1]$ in Eq. (8) and the Fig. 9(b) shows the change in accuracy. We utilize the ViT-B-16/224 model for these experiments. The mutual info. denotes mutual information. + +Effectiveness of Universal Damping Ratio We investigate whether leveraging the universal damping ratio compensates for over-compression in Fig. 9(a) while amplifying the relevant information in Fig. 9(b). We compare the different numbers of targeted layers to compare the quantitative results. The results illustrated in Fig. 9(a) show that leveraging the universal damping ratio complements the overcompressed information that occurred in earlier layers, due to the delivered relevant information from deeper layers. Concurrently, as shown in Fig. 9(c), leveraging the universal damping ratio amplifies the relevant information term in Eq (8). Aligning with these results, compared by assigning individual damping ratios to each layer, our universal damping ratio enhances the correctness performance of CoIBA. Therefore, leveraging the universal damping ratio to targeted multiple layers significantly enhances the correctness performance of CoIBA by compensating for the over-compression and amplifying the relevant information. + +Effectiveness of Variational Upper Bound We demonstrate the effectiveness of our variational upper bound by measuring correctness performance and whether the relevancy of different layers is fairly reflected in attributions. We compare quantitative results obtained by leveraging a variational upper bound (Eq. (8)) versus linearly combining layer-wise mutual information (Eq. (4)) in Fig. 10(a). As shown in the results, leveraging our variational upper bound is better at correctly identifying feature importance, regardless of confidence scores. In addition to this result, we confirm whether the relevant information of each layer is fairly reflected in the attribution map by comparing similarities. Concretely, we compare the similarity of both settings (Eq. (4) and (8)) with layer-specific relevant information computed by iterating IBA with channel uniform per + +![](images/792d37243daefc43c1d8f6e25b5bad489aee281599deaae8eb20056ab71b83b9.jpg) +(a) Faithfulness assessment + +![](images/a221e53a85c456d646161a761aa03df00795be600fa8e263a4397249f053a0c9.jpg) + +![](images/28b5420d25d4859eed8b636ed40e2f7d64eeff554de0090a54478ae4d3b2ba7b.jpg) +(c) Ablation on channel uniform perturbation + +![](images/dc630f201e95b3f713076e6af724de4fc895737fc7c3e4b6e6e3305367f5ad06.jpg) +(b) Similarity measurement +Figure 10. Quantitative analysis on CoIBA. We report the results divided by the confidence score computed by the model. For the comparison, we include the ViT-B model pre-trained with CLIP in Fig. 10(c). We filled the discrepancy in performances with blue color in Fig. 10(c). The solid line indicates the results of CoIBA. + +turbation, denoted as $\mathrm{IBA}^*$ . We employ CKA [19] as a similarity metric. As shown in Fig. 10(b), leveraging our upper bound yields the comprehensive relevant information of all the layers. This result demonstrates that our variational upper bound encourages the resulting comprehensive relevant information to fairly reflect the relevancy of each layer. + +Effectiveness of Uniform Channel Perturbation We demonstrate the effectiveness of uniform channel perturbation by comparing the correctness performance of uniform (unif.) and non-uniform (non-unif.) channel perturbation strategies. As shown in Fig. 10(c), the uniform channel perturbation consistently enhances the correctness performance of IBA for low-confident samples. In addition to this result, uniform channel perturbation significantly enlarges the correctness of resulting attribution maps, regardless of the confidence scores. These enhanced correctness performances demonstrate the effectiveness of our uniform channel perturbation, regardless of the method. + +# 5. Conclusion + +In this paper, we introduce CoIBA, which reveals the comprehensive relevant information to produce an attribution map, revealing the omitted relevancy in IBA with a theoretical guarantee. CoIBA shares the universal damping ratio to compensate for the over-compressed information, delivering relevancy among the bottleneck-inserted layers. CoIBA clearly judges the importance of each token by leveraging uniform channel perturbation. We leverage variational approximation to upper bound the information to ensure the eliminated activations are not necessary for the bottleneck-inserted layers to make a decision. CoIBA demonstrates a substantial improvement over existing methods in numerous experiments and discussions. + +# Acknowledgement + +# References + +[1] Samira Abnar and Willem Zuidema. Quantifying attention flow in transformers. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 4190-4197, 2020. 2 +[2] Alexander A Alemi, Ian Fischer, Joshua V Dillon, and Kevin Murphy. Deep variational information bottleneck. In International Conference on Learning Representations, 2017. 1, 3 +[3] Oren Barkan, Yuval Asher, Amit Eshel, Noam Koenigstein, et al. Visual explanations via iterated integrated attributions. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2073-2084, 2023. 1, 2, 4, 6, 7 +[4] Alexander Binder, Grégoire Montavon, Sebastian Lapuschkin, Klaus-Robert Müller, and Wojciech Samek. Layer-wise relevance propagation for neural networks with local renormalization layers. In Artificial Neural Networks and Machine Learning-ICANN 2016: 25th International Conference on Artificial Neural Networks, Barcelona, Spain, September 6-9, 2016, Proceedings, Part II 25, pages 63-71. Springer, 2016. 1 +[5] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9650-9660, 2021. 1 +[6] Hila Chefer, Shir Gur, and Lior Wolf. Generic attention-model explainability for interpreting bi-modal and encoder-decoder transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 397-406, 2021. 1, 2, 6, 7 +[7] Hila Chefer, Shir Gur, and Lior Wolf. Transformer interpretability beyond attention visualization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 782-791, 2021. 1, 2, 4, 6, 7 +[8] Jiamin Chen, Xuhong Li, Lei Yu, Dejing Dou, and Haoyi Xiong. Beyond intuition: Rethinking token attributions inside transformers. Transactions on Machine Learning Research, 2022. 1, 2, 4, 6, 7 +[9] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, pages 248-255. IEEE, 2009. 6 +[10] Finale Doshi-Velez and Been Kim. Towards a rigorous science of interpretable machine learning. arXiv preprint arXiv:1702.08608, 2017. 1 + +This work was supported by the Institute of Information & Communications Technology Planning & Evaluation (IITP) grant, funded by the Korea government (MSIT) (No. RS-2019-II190079, Artificial Intelligence Graduate School Program (Korea University), IITP-2025-RS-2024-00436857, ITRC (Information Technology Research Center), and No. RS-2022-II220984, Development of Artificial Intelligence Technology for Personalized Plug-and-Play Explanation and Verification of Explanation). +[11] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, G Heigold, S Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2020. 3, 6 +[12] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dólár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16000-16009, 2022. 1 +[13] Dan Hendrycks, Steven Basart, Norman Mu, Saurav Kadavath, Frank Wang, Evan Dorundo, Rahul Desai, Tyler Zhu, Samyak Parajuli, Mike Guo, et al. The many faces of robustness: A critical analysis of out-of-distribution generalization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8340-8349, 2021. 6 +[14] Dan Hendrycks, Kevin Zhao, Steven Basart, Jacob Steinhardt, and Dawn Song. Natural adversarial examples. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15262-15271, 2021. 6 +[15] Robin Hesse, Simone Schaub-Meyer, and Stefan Roth. Funnybirds: A synthetic vision dataset for a part-based analysis of explainable ai methods. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3981-3991, 2023. 2, 6 +[16] Sara Hooker, Dumitru Erhan, Pieter-Jan Kindermans, and Been Kim. A benchmark for interpretability methods in deep neural networks. Advances in Neural Information Processing Systems, 32, 2019. 7 +[17] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6 +[18] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 3 +[19] Simon Kornblith, Mohammad Norouzi, Honglak Lee, and Geoffrey Hinton. Similarity of neural network representations revisited. In International Conference on Machine Learning, pages 3519-3529. PMLR, 2019. 8 +[20] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. 6 +[21] Ze Liu, Han Hu, Yutong Lin, Zhuliang Yao, Zhenda Xie, Yixuan Wei, Jia Ning, Yue Cao, Zheng Zhang, Li Dong, et al. Swin transformer v2: Scaling up capacity and resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12009-12019, 2022. 6 +[22] Clemens Otte. Safe and interpretable machine learning: a methodological review. Computational Intelligence in Intelligent Data Analysis, pages 111-122, 2013. 1 +[23] Vitali Petsiuk, Abir Das, and Kate Saenko. Rise: Randomized input sampling for explanation of black-box models. British Machine Vision Conference, 2018. 2, 3, 6 +[24] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, + +Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 1, 6, 7 +[25] Yao Rong, Tobias Leemann, Vadim Borisov, Gjergji Kasneci, and Enkelejda Kasneci. A consistent and efficient evaluation strategy for attribution methods. In International Conference on Machine Learning, pages 18770-18795. PMLR, 2022. 2, 7 +[26] Wojciech Samek and Klaus-Robert Müller. Towards explainable artificial intelligence. Explainable AI: interpreting, explaining and visualizing deep learning, pages 5-22, 2019. 1 +[27] Karl Schulz, Leon Sixt, Federico Tombari, and Tim Landgraf. Restricting the flow: Information bottlenecks for attribution. In International Conference on Learning Representations, 2020. 1, 2, 3, 4, 6, 7 +[28] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 618-626, 2017. 1 +[29] Andreas Peter Steiner, Alexander Kolesnikov, Xiaohua Zhai, Ross Wightman, Jakob Uszkoreit, and Lucas Beyer. How to train your vit? data, augmentation, and regularization in vision transformers. Transactions on Machine Learning Research, 2024. 6 +[30] Mukund Sundararajan, Ankur Taly, and Qiqi Yan. Axiomatic attribution for deep networks. In International Conference on Machine Learning, pages 3319-3328. PMLR, 2017. 1 +[31] Naftali Tishby, Fernando C Pereira, and William Bialek. The information bottleneck method. arXiv preprint physics/0004057, 2000. 1 +[32] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In International Conference on Machine Learning, pages 10347-10357. PMLR, 2021. 4, 6 +[33] Hugo Touvron, Matthieu Cord, and Hervé Jégou. Deit iii: Revenge of the vit. In Proceedings of the European Conference on Computer Vision, pages 516-533. Springer, 2022. 6 +[34] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing, 13(4):600-612, 2004. 2 +[35] Weiyan Xie, Xiao-Hui Li, Caleb Chen Cao, and Nevin L Zhang. Vit-cx: causal explanation of vision transformers. In Proceedings of the Thirty-Second International Joint Conference on Artificial Intelligence, pages 1569–1577, 2023. 1, 2, 6, 7 +[36] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11975-11986, 2023. 1 + +[37] Yang Zhang, Ashkan Khakzar, Yawei Li, Azade Farshad, Seong Tae Kim, and Nassir Navab. Fine-grained neural network explanation by identifying input features with predictive information. Advances in Neural Information Processing Systems, 34:20040-20051, 2021. 1, 2, 3, 5 \ No newline at end of file diff --git a/2025/Comprehensive Information Bottleneck for Unveiling Universal Attribution to Interpret Vision Transformers/images.zip b/2025/Comprehensive Information Bottleneck for Unveiling Universal Attribution to Interpret Vision Transformers/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..df104d07c6b62265d87222419686f0ecd0efdf86 --- /dev/null +++ b/2025/Comprehensive Information Bottleneck for Unveiling Universal Attribution to Interpret Vision Transformers/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7b25429eb161e2ca633e32d1c6a14c2b4d2f9012d84cc4f741f186e46e226fa1 +size 735946 diff --git a/2025/Comprehensive Information Bottleneck for Unveiling Universal Attribution to Interpret Vision Transformers/layout.json b/2025/Comprehensive Information Bottleneck for Unveiling Universal Attribution to Interpret Vision Transformers/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..baa8803ce3bc3e0717e76f04c5a6f4fb6fc29b34 --- /dev/null +++ b/2025/Comprehensive Information Bottleneck for Unveiling Universal Attribution to Interpret Vision Transformers/layout.json @@ -0,0 +1,9292 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 66, + 102, + 545, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 102, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 66, + 102, + 545, + 140 + ], + "type": "text", + "content": "Comprehensive Information Bottleneck for Unveiling Universal Attribution to Interpret Vision Transformers" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 115, + 161, + 494, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 161, + 494, + 176 + ], + "spans": [ + { + "bbox": [ + 115, + 161, + 494, + 176 + ], + "type": "text", + "content": "Jung-Ho Hong Ho-Joong Kim Kyu-Sung Jeon Seong-Whan Lee*" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 178, + 179, + 432, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 178, + 179, + 432, + 191 + ], + "spans": [ + { + "bbox": [ + 178, + 179, + 432, + 191 + ], + "type": "text", + "content": "Dept. of Artificial Intelligence, Korea University, Seoul, Korea" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 159, + 193, + 448, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 193, + 448, + 205 + ], + "spans": [ + { + "bbox": [ + 159, + 193, + 448, + 205 + ], + "type": "text", + "content": "{jungho-hong, hjoong_kim, ksjeon, sw.lee}@korea.ac.kr" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 151, + 226, + 200, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 226, + 200, + 239 + ], + "spans": [ + { + "bbox": [ + 151, + 226, + 200, + 239 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 54, + 251, + 297, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 251, + 297, + 552 + ], + "spans": [ + { + "bbox": [ + 54, + 251, + 297, + 552 + ], + "type": "text", + "content": "The feature attribution method reveals the contribution of input variables to the decision-making process to provide an attribution map for explanation. Existing methods grounded on the information bottleneck principle compute information in a specific layer to obtain attributions, compressing the features by injecting noise via a parametric damping ratio. However, the attribution obtained in a specific layer neglects evidence of the decision-making process distributed across layers. In this paper, we introduce a comprehensive information bottleneck (CoIBA), which discovers the relevant information in each targeted layer to explain the decision-making process. Our core idea is applying information bottleneck in multiple targeted layers to estimate the comprehensive information by sharing a parametric damping ratio across the layers. Leveraging this shared ratio complements the over-compressed information to discover the omitted clues of the decision by sharing the relevant information across the targeted layers. We suggest the variational approach to fairly reflect the relevant information of each layer by upper bounding layer-wise information. Therefore, CoIBA guarantees that the discarded activation is unnecessary in every targeted layer to make a decision. The extensive experimental results demonstrate the enhancement in faithfulness of the feature attributions provided by CoIBA." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 565, + 135, + 577 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 565, + 135, + 577 + ], + "spans": [ + { + "bbox": [ + 55, + 565, + 135, + 577 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 586, + 296, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 586, + 296, + 694 + ], + "spans": [ + { + "bbox": [ + 55, + 586, + 296, + 694 + ], + "type": "text", + "content": "Vision transformer (ViT) achieves remarkable performance in diverse fundamental computer vision tasks, such as multi-modal [24, 36] and self-supervised learning [5, 12]. Despite this achievement, the black box nature stemming from the complex structure of ViT restricts the testability of the end-to-end system, impeding the failure diagnosis [10]. This limitation constrains the application of ViT in safety-critical areas where transparency and interpretability are considered important as well [22, 26]. Feature attribu" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 321, + 224, + 550, + 326 + ], + "blocks": [ + { + "bbox": [ + 321, + 224, + 550, + 326 + ], + "lines": [ + { + "bbox": [ + 321, + 224, + 550, + 326 + ], + "spans": [ + { + "bbox": [ + 321, + 224, + 550, + 326 + ], + "type": "image", + "image_path": "46452c1e53fc94d13bba0dc97d43689415fcc1dd0fbd03415ecc6d177ee52654.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 323, + 327, + 432, + 405 + ], + "blocks": [ + { + "bbox": [ + 323, + 327, + 432, + 405 + ], + "lines": [ + { + "bbox": [ + 323, + 327, + 432, + 405 + ], + "spans": [ + { + "bbox": [ + 323, + 327, + 432, + 405 + ], + "type": "image", + "image_path": "6d86f523b511c3b6c21b19a26e75c1bb0918d32013203d3ab82e0eacb845f0e0.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 318, + 407, + 552, + 417 + ], + "lines": [ + { + "bbox": [ + 318, + 407, + 552, + 417 + ], + "spans": [ + { + "bbox": [ + 318, + 407, + 552, + 417 + ], + "type": "text", + "content": "(b) Cumulative number of each layer yielding optimal result per sample" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 434, + 327, + 544, + 405 + ], + "blocks": [ + { + "bbox": [ + 434, + 327, + 544, + 405 + ], + "lines": [ + { + "bbox": [ + 434, + 327, + 544, + 405 + ], + "spans": [ + { + "bbox": [ + 434, + 327, + 544, + 405 + ], + "type": "image", + "image_path": "d4ae8871e3d459b793874d5beb169e432c1db15051852e1b8b9cf6a3c7380eaa.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 418, + 554, + 528 + ], + "lines": [ + { + "bbox": [ + 313, + 418, + 554, + 528 + ], + "spans": [ + { + "bbox": [ + 313, + 418, + 554, + 528 + ], + "type": "text", + "content": "Figure 1. The quantitative and qualitative comparisons of attribution maps generated from different layers. We illustrate the visualized attribution maps from the first to fourth columns in Fig. 1(a) utilizing IBA [27]. The text on the top of each figure indicates the index of the layer. We select the layer index of the large model with interval 2. The fifth column in Fig. 1(a) shows the similarity comparison between attributions of different layers. Fig. 1(b) illustrates the cumulative number of each layer that provides the best explanation per sample. The uniformly distributed data indicates the absence of the layer optimal for the explanation." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 533, + 555, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 555, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 555, + 640 + ], + "type": "text", + "content": "tion methods [4, 28, 30] are introduced to reveal the evidence supporting the decision-making process. Many of the existing attribution methods designed to interpret convolutional neural networks show limited explainability to interpret ViT [7]. To address this limitation, feature attribution methods [3, 6-8, 35] tailored for interpreting ViT have been proposed, leveraging either attention weights computed in ViT or intermediate representations, such as the class token, to generate attribution maps." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "content": "Although feature attribution methods have advanced in interpreting ViT, they lack theoretical guarantees that low-scored attributions are not necessary to the decision-making process [27]. The methods [27, 37] grounded on the information bottleneck principle [2, 31] address this limitation by guaranteeing the importance of attributions with a" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 15 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 15 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 15 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 198, + 13, + 443, + 24 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 13, + 443, + 24 + ], + "spans": [ + { + "bbox": [ + 198, + 13, + 443, + 24 + ], + "type": "text", + "content": "Except for this watermark, it is identical to the accepted version;" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 179, + 25, + 462, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 25, + 462, + 37 + ], + "spans": [ + { + "bbox": [ + 179, + 25, + 462, + 37 + ], + "type": "text", + "content": "the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 70, + 703, + 147, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 703, + 147, + 712 + ], + "spans": [ + { + "bbox": [ + 70, + 703, + 147, + 712 + ], + "type": "text", + "content": "*Corresponding author" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "25166" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 264 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 264 + ], + "type": "text", + "content": "variational approximation. Information bottleneck for attribution (IBA) [27] proposes the model-agnostic method to compress the information unnecessary for the prediction by positioning the bottleneck layer into the target layer. InputIBA [37] succeeds in providing a high-resolution attribution map by directly measuring the information in the input domain, assuming the prior distribution with a generative model. However, existing information bottleneck-based methods face two main challenges. First, the resulting attribution map reflects the partial of the decision-making process as the restriction of the information flow is conducted in a specific targeted layer. This constraint provides inconsistent evidence for a single decision-making process, leading to unobvious interpretations. Lastly, no layer dominantly provides the most appropriate relevant information for explaining the decision-making process." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 267, + 295, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 267, + 295, + 567 + ], + "spans": [ + { + "bbox": [ + 56, + 267, + 295, + 567 + ], + "type": "text", + "content": "Fig. 1 shows the aforementioned issues with quantitative and qualitative examples. In the first row in Fig. 1(a), the visualization derived from the second layer (blocks.2) highlights the barrel whereas the fifth (blocks.5) and eighth layers (blocks.8) highlight the top and bottom of the wheels, respectively. In the second row in Fig. 1(a), the visualization in the earlier layer highlights the person holding the racket, while the deeper layer progressively concentrates on the racket. To confirm the layer-wise discrepancy in attributions quantitatively, we measure SSIM [34] in the right row of Fig. 1(a). As the layer providing attributions distance farther, the dissimilarity in those attributions increases. Thus, computing the relevant information from isolated layer-specific information bottlenecks highlights different attributions for the same decision. Among the layers producing the attribution maps, there is no dominant layer that provides the most appropriate relevant information to explain the decision-making process. Fig. 1(b) compares the cumulative number of each layer yielding attributions with the optimal insertion and deletion scores per input sample. The results show a lack of the layer that produces the most appropriate relevant information for explanation. Thus, identifying the most faithful attribution from the layers requires a heuristic search due to the lack of criteria to distinguish the optimal layer." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 570, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 295, + 713 + ], + "type": "text", + "content": "In this paper, we propose a novel comprehensive information bottleneck for attribution (CoIBA), which reveals the information relevant for each of the target layers to provide the attributions. Our core idea is to reveal comprehensive relevant information by eliminating unnecessary information in every targeted layer of the prediction. We estimate the comprehensive information by sharing a parameterized universal damping ratio across layers, thereby removing the requirements of additional heuristics. This sharing strategy compensates for the over-compressed information of individual layers by sharing information in each layer necessary for the decision. Since a channel el" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "text", + "content": "ement placed in the same location but different layers captures different features, we uniformly perturb all the channel elements in a single token representation, to handle the information of different layers with a shared parametric ratio. We suggest a variational upper bound to restrict information flowing in all the targeted layers. Our approach eliminates the heuristic search for balancing inconsistent information along the layers by suggesting a variational upper bound to fairly reflect the layer-wise relevant information." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 180, + 553, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 180, + 553, + 312 + ], + "spans": [ + { + "bbox": [ + 313, + 180, + 553, + 312 + ], + "type": "text", + "content": "We conduct experiments to show the correctness performance of CoIBA utilizing various assessments. As the evaluations are solely approximations, i.e. no ground truth exists, we include FunnyBirds [15], which provide the ground truth and assess the comprehensive quality of attribution maps. To assess the feature importance assessment, we leverage insertion/deletion [23] and remove-and-debias (ROAD) [25], which evaluate the faithfulness of attribution maps. In addition, we scrutinize the faithfulness of CoIBA by analyzing the confident-aware assessment. We discuss the effectiveness of our method with a rigorous discussion." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 323, + 405, + 335 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 323, + 405, + 335 + ], + "spans": [ + { + "bbox": [ + 313, + 323, + 405, + 335 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 342, + 553, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 342, + 553, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 342, + 553, + 653 + ], + "type": "text", + "content": "Explanation Methods for Vision Transformer Existing explanatory methods designed to interpret ViT provide the attribution map utilizing the latent representation and the corresponding gradients. However, these methods show limited explainability while being adapted to ViT [7]. To address this limitation, Rollout [1] linearly combines the attention weights across the layers to compute the attribution map. Trans-attr [7] provides the class-discriminative attribution map by constructing the layer-wise relevance propagation rule for ViT. Generic [6] produces the attribution map with a generalized procedure to interpret diversified transformer models, leveraging the gradient of the attention map to produce the class-discriminative attribution maps. IIA [3] introduces the iterative integration across the input image, leveraging the internal representations processed by the model and their gradients. ViT-CX [35] utilizes the patch embeddings and measures causal impacts to provide attributions. Beyond [8] unfolds attention blocks with the chain rule between final prediction, CLS, and tokens to obtain token contribution. Existing propagation-based approaches require a specific implementation to weigh the contributions of input variables. In contrast to this constraint, CoIBA does not require implementing additional procedures as same as IBA. Furthermore, aligning with the information bottleneck principle, CoIBA guarantees the highlighted attributions are important in all targeted layers for the decision." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "text", + "content": "Information Bottleneck Approach The information bottleneck principle is broadly utilized to obtain information in the activation necessary for the specific objective. To this end, the information bottleneck-based methods compress the information unnecessary for the objective while" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25167" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 70, + 294, + 260 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 294, + 260 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 294, + 260 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 294, + 260 + ], + "type": "image", + "image_path": "04e84f49a950cd4233f5fe0f6451bc1cf5b7ed1e0f1c422923f55be756b332a8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 265, + 296, + 331 + ], + "lines": [ + { + "bbox": [ + 55, + 265, + 296, + 331 + ], + "spans": [ + { + "bbox": [ + 55, + 265, + 296, + 331 + ], + "type": "text", + "content": "Figure 2. Main difference between IBA and CoIBA. IBA obtains the relevant information in the specific bottleneck-inserted layer. Thus, obtaining layer-wise attribution maps necessarily requires iterative running over the layers. In contrast, CoIBA seeks the layer-wise relevant information from each bottleneck-inserted layer with a universal damping ratio." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 348, + 296, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 348, + 296, + 564 + ], + "spans": [ + { + "bbox": [ + 54, + 348, + 296, + 564 + ], + "type": "text", + "content": "maintaining the relevant information. Given the challenges of directly estimating the amount of information, a variational approach [18] is used to approximate the intractable posteriors and marginals. The Deep Variational Information Bottleneck [2] method employs the information bottleneck principle in deep networks, using a variational lower bound alongside the reparameterization trick. Referring to this strategy, IBA [27] adapts the information bottleneck principles to restrict the information flow in an intermediate layer and provides the relevant information as an attribution map. InputIBA [37] provides high-resolution attributions by directly estimating relevant information in the input domain. However, existing methods compute relevant information solely considering a specific layer, overlooking the information necessary in the different layers. In contrast to this process, CoIBA obtains comprehensive relevant information from the multiple layers, highlighting the evidence important in every targeted layer to make a decision." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 582, + 111, + 594 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 111, + 594 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 111, + 594 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "content": "The main goal of CoIBA is to produce an attribution map by reflecting the information relevant to the targeted layers, which is opposed to the IBA as shown in Fig. 2. To show this procedure, we first introduce IBA in Sec. 3.1 and discuss the limitations of the existing information bottleneck-based approach in Sec. 3.2. Then, in Sec. 3.3, we describe the overall procedure of CoIBA. Finally, we introduce a variational upper bound, which enhances the reflection of layer-wise relevant information in Sec. 3.4." + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 328, + 70, + 432, + 153 + ], + "blocks": [ + { + "bbox": [ + 328, + 70, + 432, + 153 + ], + "lines": [ + { + "bbox": [ + 328, + 70, + 432, + 153 + ], + "spans": [ + { + "bbox": [ + 328, + 70, + 432, + 153 + ], + "type": "image", + "image_path": "ea662166572a569ff49afc338896d996e47bafc11a01d53f5718baed5727779c.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 434, + 71, + 541, + 153 + ], + "blocks": [ + { + "bbox": [ + 434, + 71, + 541, + 153 + ], + "lines": [ + { + "bbox": [ + 434, + 71, + 541, + 153 + ], + "spans": [ + { + "bbox": [ + 434, + 71, + 541, + 153 + ], + "type": "image", + "image_path": "a1b5c281fd2885172ccf4b287eb5aee8be10200195a305605b0fdfd1f631a493.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 329, + 156, + 388, + 214 + ], + "blocks": [ + { + "bbox": [ + 329, + 156, + 388, + 214 + ], + "lines": [ + { + "bbox": [ + 329, + 156, + 388, + 214 + ], + "spans": [ + { + "bbox": [ + 329, + 156, + 388, + 214 + ], + "type": "image", + "image_path": "e44fd811d68c2e9f304fa660ae85e26f7f7aaa706c1009dfa03b1937adafd59f.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 217, + 555, + 283 + ], + "lines": [ + { + "bbox": [ + 313, + 217, + 555, + 283 + ], + "spans": [ + { + "bbox": [ + 313, + 217, + 555, + 283 + ], + "type": "text", + "content": "Figure 3. Quantitative and qualitative results of IBA*. In the first row, the horizontal line indicates the performance of layers-specific relevant information, i.e., IBA. We report the difference in insertion and deletion scores (" + }, + { + "bbox": [ + 313, + 217, + 555, + 283 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 313, + 217, + 555, + 283 + ], + "type": "text", + "content": "InsDel). The higher score is better for this metric. The visualized attribution maps are illustrated in the second row obtained from VIT-B-16/224 [11]." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 388, + 156, + 438, + 214 + ], + "blocks": [ + { + "bbox": [ + 388, + 156, + 438, + 214 + ], + "lines": [ + { + "bbox": [ + 388, + 156, + 438, + 214 + ], + "spans": [ + { + "bbox": [ + 388, + 156, + 438, + 214 + ], + "type": "image", + "image_path": "c5cb7cb051d773a4a7eb0fd1f0472968941d350393a39855bbb2059cdcce1296.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 438, + 156, + 489, + 214 + ], + "blocks": [ + { + "bbox": [ + 438, + 156, + 489, + 214 + ], + "lines": [ + { + "bbox": [ + 438, + 156, + 489, + 214 + ], + "spans": [ + { + "bbox": [ + 438, + 156, + 489, + 214 + ], + "type": "image", + "image_path": "673112c63ece870c5c754456fdf2df0cb39ce5a23b48b8a32fe0c6f59646ea08.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 490, + 156, + 541, + 214 + ], + "blocks": [ + { + "bbox": [ + 490, + 156, + 541, + 214 + ], + "lines": [ + { + "bbox": [ + 490, + 156, + 541, + 214 + ], + "spans": [ + { + "bbox": [ + 490, + 156, + 541, + 214 + ], + "type": "image", + "image_path": "7729a9370a5731c56cff819cbc8a3cf491d64c67c614d0774d4687202778ac19.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 289, + 424, + 302 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 289, + 424, + 302 + ], + "spans": [ + { + "bbox": [ + 313, + 289, + 424, + 302 + ], + "type": "text", + "content": "3.1. Background - IBA" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 307, + 554, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 307, + 554, + 354 + ], + "spans": [ + { + "bbox": [ + 313, + 307, + 554, + 354 + ], + "type": "text", + "content": "IBA adopts an information bottleneck principle to produce relevant information to provide an attribution map. To this end, in " + }, + { + "bbox": [ + 313, + 307, + 554, + 354 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 313, + 307, + 554, + 354 + ], + "type": "text", + "content": "-th layer, IBA computes the bottleneck representation " + }, + { + "bbox": [ + 313, + 307, + 554, + 354 + ], + "type": "inline_equation", + "content": "Z_{l}" + }, + { + "bbox": [ + 313, + 307, + 554, + 354 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 382, + 365, + 553, + 378 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 382, + 365, + 553, + 378 + ], + "spans": [ + { + "bbox": [ + 382, + 365, + 553, + 378 + ], + "type": "interline_equation", + "content": "Z _ {l} = \\lambda_ {l} R _ {l} + (1 - \\lambda_ {l}) \\epsilon_ {l}. \\tag {1}", + "image_path": "a601fb289d855b485765dafca90cc4f9020e04a43aaa99a9649e2b9f0401ba21.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 387, + 554, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 387, + 554, + 471 + ], + "spans": [ + { + "bbox": [ + 313, + 387, + 554, + 471 + ], + "type": "text", + "content": "Here, the damping ratio " + }, + { + "bbox": [ + 313, + 387, + 554, + 471 + ], + "type": "inline_equation", + "content": "\\lambda_{l}" + }, + { + "bbox": [ + 313, + 387, + 554, + 471 + ], + "type": "text", + "content": " is a learnable parameter assigned at the bottleneck layer, manipulating the degree of damping signal between the activation " + }, + { + "bbox": [ + 313, + 387, + 554, + 471 + ], + "type": "inline_equation", + "content": "R_{l}" + }, + { + "bbox": [ + 313, + 387, + 554, + 471 + ], + "type": "text", + "content": " and the independent noise " + }, + { + "bbox": [ + 313, + 387, + 554, + 471 + ], + "type": "inline_equation", + "content": "\\epsilon_{l}" + }, + { + "bbox": [ + 313, + 387, + 554, + 471 + ], + "type": "text", + "content": ". The independent noise " + }, + { + "bbox": [ + 313, + 387, + 554, + 471 + ], + "type": "inline_equation", + "content": "\\epsilon_{l}" + }, + { + "bbox": [ + 313, + 387, + 554, + 471 + ], + "type": "text", + "content": " is sampled from the Gaussian distribution. To maintain the statistic of internal representation, the Gaussian distribution shares the mean " + }, + { + "bbox": [ + 313, + 387, + 554, + 471 + ], + "type": "inline_equation", + "content": "\\mu_{R_l}" + }, + { + "bbox": [ + 313, + 387, + 554, + 471 + ], + "type": "text", + "content": " and variance " + }, + { + "bbox": [ + 313, + 387, + 554, + 471 + ], + "type": "inline_equation", + "content": "\\sigma_{R_l}" + }, + { + "bbox": [ + 313, + 387, + 554, + 471 + ], + "type": "text", + "content": " with the representation " + }, + { + "bbox": [ + 313, + 387, + 554, + 471 + ], + "type": "inline_equation", + "content": "R_{l}" + }, + { + "bbox": [ + 313, + 387, + 554, + 471 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 472, + 554, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 472, + 554, + 518 + ], + "spans": [ + { + "bbox": [ + 313, + 472, + 554, + 518 + ], + "type": "text", + "content": "IBA minimizes the shared information between the activation " + }, + { + "bbox": [ + 313, + 472, + 554, + 518 + ], + "type": "inline_equation", + "content": "R_{l}" + }, + { + "bbox": [ + 313, + 472, + 554, + 518 + ], + "type": "text", + "content": " and bottleneck variable " + }, + { + "bbox": [ + 313, + 472, + 554, + 518 + ], + "type": "inline_equation", + "content": "Z_{l}" + }, + { + "bbox": [ + 313, + 472, + 554, + 518 + ], + "type": "text", + "content": " while maximizing the shared information between the bottleneck variable " + }, + { + "bbox": [ + 313, + 472, + 554, + 518 + ], + "type": "inline_equation", + "content": "Z_{l}" + }, + { + "bbox": [ + 313, + 472, + 554, + 518 + ], + "type": "text", + "content": " and the label " + }, + { + "bbox": [ + 313, + 472, + 554, + 518 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 313, + 472, + 554, + 518 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 375, + 529, + 553, + 548 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 375, + 529, + 553, + 548 + ], + "spans": [ + { + "bbox": [ + 375, + 529, + 553, + 548 + ], + "type": "interline_equation", + "content": "\\max _ {\\lambda_ {l}} I \\left[ Z _ {l}; Y \\right] - \\beta I \\left[ R _ {l}, Z _ {l} \\right]. \\tag {2}", + "image_path": "09932fd911e051e0ae1b3326b1e63f575d1299f172211a8596513b96acebb4e1.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 556, + 553, + 580 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 556, + 553, + 580 + ], + "spans": [ + { + "bbox": [ + 313, + 556, + 553, + 580 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 313, + 556, + 553, + 580 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 313, + 556, + 553, + 580 + ], + "type": "text", + "content": " is a hyperparameter manipulating the trade-off between compression and relevancy in " + }, + { + "bbox": [ + 313, + 556, + 553, + 580 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 313, + 556, + 553, + 580 + ], + "type": "text", + "content": "-th layer." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 588, + 388, + 600 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 588, + 388, + 600 + ], + "spans": [ + { + "bbox": [ + 313, + 588, + 388, + 600 + ], + "type": "text", + "content": "3.2. Motivation" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "content": "IBA selects a specific layer to obtain relevant information to provide an attribution map. However, the relevant information of a specific single layer does not reflect the overall evidence required in the sequence of layers to make a decision. To demonstrate our insistence, we illustrate the quantitative comparison of the attribution maps obtained in a specific layer and multiple layers in Fig. 3, leveraging insertion/deletion [23]. We iterate IBA multiple times along the layers to obtain the set of individual relevant information" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25168" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 99, + 76, + 153, + 129 + ], + "blocks": [ + { + "bbox": [ + 78, + 87, + 90, + 140 + ], + "lines": [ + { + "bbox": [ + 78, + 87, + 90, + 140 + ], + "spans": [ + { + "bbox": [ + 78, + 87, + 90, + 140 + ], + "type": "text", + "content": "eaaee" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 99, + 76, + 153, + 129 + ], + "lines": [ + { + "bbox": [ + 99, + 76, + 153, + 129 + ], + "spans": [ + { + "bbox": [ + 99, + 76, + 153, + 129 + ], + "type": "image", + "image_path": "84254202b7b2cdc9d21b3e58547095be842d63f0a1b84370e4a672b1a71abc71.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 103, + 130, + 148, + 138 + ], + "lines": [ + { + "bbox": [ + 103, + 130, + 148, + 138 + ], + "spans": [ + { + "bbox": [ + 103, + 130, + 148, + 138 + ], + "type": "text", + "content": "Class: Hyena" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 156, + 76, + 265, + 150 + ], + "blocks": [ + { + "bbox": [ + 95, + 141, + 156, + 149 + ], + "lines": [ + { + "bbox": [ + 95, + 141, + 156, + 149 + ], + "spans": [ + { + "bbox": [ + 95, + 141, + 156, + 149 + ], + "type": "text", + "content": "Confidence: 85.87" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 156, + 76, + 265, + 150 + ], + "lines": [ + { + "bbox": [ + 156, + 76, + 265, + 150 + ], + "spans": [ + { + "bbox": [ + 156, + 76, + 265, + 150 + ], + "type": "image", + "image_path": "3f70ab71a742248c92c64b7aa2499ea94e3427761959d99761d4ca65ec3e6e05.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 99, + 156, + 153, + 211 + ], + "blocks": [ + { + "bbox": [ + 78, + 167, + 90, + 223 + ], + "lines": [ + { + "bbox": [ + 78, + 167, + 90, + 223 + ], + "spans": [ + { + "bbox": [ + 78, + 167, + 90, + 223 + ], + "type": "text", + "content": "Hard Sample" + } + ] + } + ], + "index": 5, + "angle": 270, + "type": "image_caption" + }, + { + "bbox": [ + 99, + 156, + 153, + 211 + ], + "lines": [ + { + "bbox": [ + 99, + 156, + 153, + 211 + ], + "spans": [ + { + "bbox": [ + 99, + 156, + 153, + 211 + ], + "type": "image", + "image_path": "210a74117a0c02bb86604554dd1a882f30f28c22a4e44fc636ce78f15699e53b.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 93, + 212, + 158, + 220 + ], + "lines": [ + { + "bbox": [ + 93, + 212, + 158, + 220 + ], + "spans": [ + { + "bbox": [ + 93, + 212, + 158, + 220 + ], + "type": "text", + "content": "Class: Parallel Bars" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 95, + 223, + 154, + 231 + ], + "lines": [ + { + "bbox": [ + 95, + 223, + 154, + 231 + ], + "spans": [ + { + "bbox": [ + 95, + 223, + 154, + 231 + ], + "type": "text", + "content": "Confidence: 19.71" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 156, + 186, + 174, + 203 + ], + "blocks": [ + { + "bbox": [ + 156, + 186, + 174, + 203 + ], + "lines": [ + { + "bbox": [ + 156, + 186, + 174, + 203 + ], + "spans": [ + { + "bbox": [ + 156, + 186, + 174, + 203 + ], + "type": "image", + "image_path": "f04d2669558c4dcf691ac1608cc4d69796554afefbb331adb53880adc7b975ec.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 175, + 157, + 265, + 232 + ], + "blocks": [ + { + "bbox": [ + 175, + 157, + 265, + 232 + ], + "lines": [ + { + "bbox": [ + 175, + 157, + 265, + 232 + ], + "spans": [ + { + "bbox": [ + 175, + 157, + 265, + 232 + ], + "type": "image", + "image_path": "87dafdfb7dc66a24a577cb325728f387ff0d35db09075ba65f525afe39b9d800.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 238, + 296, + 348 + ], + "lines": [ + { + "bbox": [ + 55, + 238, + 296, + 348 + ], + "spans": [ + { + "bbox": [ + 55, + 238, + 296, + 348 + ], + "type": "text", + "content": "Figure 4. The quantitative comparisons of explanation methods in high-confident (easy) and low-confident (hard) input samples. The confidence score is the softmax probability of the target class predicted by the model. We compare the information-bottleneck-based methods including IBA [27], IBA*, and CoIBA (marked with a circle), and the propagation-based explanation baselines including Beyond [8], IIA [3], and Chefer-LRP [7] methods (marked with a triangle). We utilize DeiT-B-16/224 [32] for the evaluation. For insertion and deletion, higher and lower scores are better, respectively." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 354, + 296, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 354, + 296, + 498 + ], + "spans": [ + { + "bbox": [ + 55, + 354, + 296, + 498 + ], + "type": "text", + "content": "for each layer. After that, we compute a linear combination over the set of individual relevant information. We denote this procedure as " + }, + { + "bbox": [ + 55, + 354, + 296, + 498 + ], + "type": "inline_equation", + "content": "\\mathrm{IBA}^*" + }, + { + "bbox": [ + 55, + 354, + 296, + 498 + ], + "type": "text", + "content": ". As shown in the results, the attribution maps produced by " + }, + { + "bbox": [ + 55, + 354, + 296, + 498 + ], + "type": "inline_equation", + "content": "\\mathrm{IBA}^*" + }, + { + "bbox": [ + 55, + 354, + 296, + 498 + ], + "type": "text", + "content": " archive better scores compared to the IBA. To scrutinize the quantitative results between " + }, + { + "bbox": [ + 55, + 354, + 296, + 498 + ], + "type": "inline_equation", + "content": "\\mathrm{IBA}^*" + }, + { + "bbox": [ + 55, + 354, + 296, + 498 + ], + "type": "text", + "content": " and IBA, we split the insertion/deletion scores by the confidence scores of the model in Fig. 4. As shown in the results, the increase in performance is only observed in high-confident samples, unlike in low-confident samples. Therefore, even iterating IBA, which requires high computational cost, shows limited performance enhancement according to the difficulty of each sample." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 505, + 263, + 517 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 505, + 263, + 517 + ], + "spans": [ + { + "bbox": [ + 55, + 505, + 263, + 517 + ], + "type": "text", + "content": "3.3. Comprehensive Information Bottleneck" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 522, + 295, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 522, + 295, + 653 + ], + "spans": [ + { + "bbox": [ + 55, + 522, + 295, + 653 + ], + "type": "text", + "content": "Our comprehensive information bottleneck (CoIBA) inserts the bottleneck layer into the sequence of targeted layers to restrict the information flowing in the intermediate layer. Within the bottleneck layer, CoIBA restricts the information with the learnable universal damping ratio shared across the bottleneck-inserted layers. Our goal is to optimize the universal damping ratio to obtain the layer-wise relevant information while fulfilling the information bottleneck objective. The universal damping ratio enables the attributions to identify distributed relevant information omitted by IBA, which restricts information in a layer-specific manner." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "content": "Formulating Information Restriction The model is trained to make a decision " + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\tilde{Y} = f(X)" + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "content": " to predict the ground truth " + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "content": " by taking input " + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "content": ". To make a decision, the input sample " + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "content": " is contextualized by passing through the sequence of intermediate layers. Here, we insert the bottle-" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 72, + 555, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 216 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 216 + ], + "type": "text", + "content": "neck layer into the targeted layer. The role of the bottleneck layer is to restrict the information by dampening the signal of the intermediate representation passing through the " + }, + { + "bbox": [ + 313, + 72, + 555, + 216 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 313, + 72, + 555, + 216 + ], + "type": "text", + "content": " bottleneck-inserted layers. We perturb the intermediate representations in the " + }, + { + "bbox": [ + 313, + 72, + 555, + 216 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 313, + 72, + 555, + 216 + ], + "type": "text", + "content": "-th layer to dampen the signal. To dampen the signal, we inject the independent noise " + }, + { + "bbox": [ + 313, + 72, + 555, + 216 + ], + "type": "inline_equation", + "content": "\\epsilon_{l}" + }, + { + "bbox": [ + 313, + 72, + 555, + 216 + ], + "type": "text", + "content": " into the intermediate bottleneck representation " + }, + { + "bbox": [ + 313, + 72, + 555, + 216 + ], + "type": "inline_equation", + "content": "R_{l}^{\\prime}" + }, + { + "bbox": [ + 313, + 72, + 555, + 216 + ], + "type": "text", + "content": ", which is computed from the bottleneck variables " + }, + { + "bbox": [ + 313, + 72, + 555, + 216 + ], + "type": "inline_equation", + "content": "Z_{l-1}" + }, + { + "bbox": [ + 313, + 72, + 555, + 216 + ], + "type": "text", + "content": " of the preceding layer. Note that the first element of intermediate bottleneck representations is the non-perturbed intermediate representation, i.e., " + }, + { + "bbox": [ + 313, + 72, + 555, + 216 + ], + "type": "inline_equation", + "content": "R_{1}^{\\prime} = R_{1}" + }, + { + "bbox": [ + 313, + 72, + 555, + 216 + ], + "type": "text", + "content": ". Thus, we obtain the bottleneck representation " + }, + { + "bbox": [ + 313, + 72, + 555, + 216 + ], + "type": "inline_equation", + "content": "Z_{l}" + }, + { + "bbox": [ + 313, + 72, + 555, + 216 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 385, + 227, + 555, + 241 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 385, + 227, + 555, + 241 + ], + "spans": [ + { + "bbox": [ + 385, + 227, + 555, + 241 + ], + "type": "interline_equation", + "content": "Z _ {l} = \\lambda R _ {l} ^ {\\prime} + (1 - \\lambda) \\epsilon_ {l}, \\tag {3}", + "image_path": "445cffc257299ab398672a27d75753ace45d2b734e11fcc6b02535021b42750d.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 251, + 555, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 251, + 555, + 502 + ], + "spans": [ + { + "bbox": [ + 313, + 251, + 555, + 502 + ], + "type": "text", + "content": "Here, a universal damping ratio " + }, + { + "bbox": [ + 313, + 251, + 555, + 502 + ], + "type": "inline_equation", + "content": "\\lambda \\in \\mathbb{R}^{P\\times 1}" + }, + { + "bbox": [ + 313, + 251, + 555, + 502 + ], + "type": "text", + "content": " manipulates the degree of perturbation for each token, denoting " + }, + { + "bbox": [ + 313, + 251, + 555, + 502 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 313, + 251, + 555, + 502 + ], + "type": "text", + "content": " as the number of patches. Thus, " + }, + { + "bbox": [ + 313, + 251, + 555, + 502 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 251, + 555, + 502 + ], + "type": "text", + "content": " is ranged from 0 to 1, allowing the propagation of the signal when " + }, + { + "bbox": [ + 313, + 251, + 555, + 502 + ], + "type": "inline_equation", + "content": "\\lambda = 1" + }, + { + "bbox": [ + 313, + 251, + 555, + 502 + ], + "type": "text", + "content": " otherwise blocking it when " + }, + { + "bbox": [ + 313, + 251, + 555, + 502 + ], + "type": "inline_equation", + "content": "\\lambda = 0" + }, + { + "bbox": [ + 313, + 251, + 555, + 502 + ], + "type": "text", + "content": ". As the universal damping ratio " + }, + { + "bbox": [ + 313, + 251, + 555, + 502 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 251, + 555, + 502 + ], + "type": "text", + "content": " is consistently adapted to the layers, we omit the layer index in the notation. We compute " + }, + { + "bbox": [ + 313, + 251, + 555, + 502 + ], + "type": "inline_equation", + "content": "\\lambda = \\mathrm{sigmoid}(\\alpha)" + }, + { + "bbox": [ + 313, + 251, + 555, + 502 + ], + "type": "text", + "content": " by passing the trainable parameter " + }, + { + "bbox": [ + 313, + 251, + 555, + 502 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 251, + 555, + 502 + ], + "type": "text", + "content": ", which is initialized with 5, to the sigmoid function. The universal damping ratio uniformly perturbs the signal along the channel dimension. This setting avoids handling the neurons that capture different features with the same coefficient and induces concentration on the token's importance. We empirically show this uniform perturbation across the channels enhances the faithfulness of resulting relevant information. The independent noise " + }, + { + "bbox": [ + 313, + 251, + 555, + 502 + ], + "type": "inline_equation", + "content": "\\epsilon_{l}" + }, + { + "bbox": [ + 313, + 251, + 555, + 502 + ], + "type": "text", + "content": " shares the dimension and is sampled from the Gaussian distribution with the mean " + }, + { + "bbox": [ + 313, + 251, + 555, + 502 + ], + "type": "inline_equation", + "content": "\\mu_{R_l}" + }, + { + "bbox": [ + 313, + 251, + 555, + 502 + ], + "type": "text", + "content": " and variance " + }, + { + "bbox": [ + 313, + 251, + 555, + 502 + ], + "type": "inline_equation", + "content": "\\sigma_{R_l}^2" + }, + { + "bbox": [ + 313, + 251, + 555, + 502 + ], + "type": "text", + "content": ", such that " + }, + { + "bbox": [ + 313, + 251, + 555, + 502 + ], + "type": "inline_equation", + "content": "\\epsilon_{l}\\sim \\mathcal{N}(\\mu_{R_l},\\sigma_{R_l}^2)" + }, + { + "bbox": [ + 313, + 251, + 555, + 502 + ], + "type": "text", + "content": ". To minimize statistical differences between the bottleneck variable and non-perturbed activations, we leverage the mean " + }, + { + "bbox": [ + 313, + 251, + 555, + 502 + ], + "type": "inline_equation", + "content": "\\mu_{R_l}" + }, + { + "bbox": [ + 313, + 251, + 555, + 502 + ], + "type": "text", + "content": " and variance " + }, + { + "bbox": [ + 313, + 251, + 555, + 502 + ], + "type": "inline_equation", + "content": "\\sigma_{R_l}^2" + }, + { + "bbox": [ + 313, + 251, + 555, + 502 + ], + "type": "text", + "content": " from those of the non-perturbed activations." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 503, + 556, + 683 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 503, + 556, + 683 + ], + "spans": [ + { + "bbox": [ + 313, + 503, + 556, + 683 + ], + "type": "text", + "content": "Information Bottleneck Objective The information bottleneck principle desires to obtain information necessary for the objective while compressing irrelevant information from the signal. Compressing irrelevant information is performed by minimizing the shared information between the activation of a bottleneck-inserted network and the bottleneck representation. Building upon this, we compress the amount of information by minimizing the mutual information between internal sequences of bottleneck representations while maximizing the mutual information between the ground truth and bottleneck representations. Combining two aspects with the ground truth " + }, + { + "bbox": [ + 313, + 503, + 556, + 683 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 313, + 503, + 556, + 683 + ], + "type": "text", + "content": ", we maximize the mutual information between the bottleneck variable and the ground truth while minimizing the mutual information between bottleneck representations as follows:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 340, + 693, + 555, + 717 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 340, + 693, + 555, + 717 + ], + "spans": [ + { + "bbox": [ + 340, + 693, + 555, + 717 + ], + "type": "interline_equation", + "content": "\\max _ {\\lambda} I \\left[ Z _ {L}; Y \\right] - \\frac {1}{L} \\left(\\sum_ {l = 1} ^ {L} \\beta_ {l} I \\left[ Z _ {l - 1}; Z _ {l} \\right]\\right). \\tag {4}", + "image_path": "7fe3039fa4cc8060c470402cb3ce7a3c9bbc0677ba607d73eb1679fabb3f2a62.jpg" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "25169" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 82, + 70, + 531, + 242 + ], + "blocks": [ + { + "bbox": [ + 82, + 70, + 531, + 242 + ], + "lines": [ + { + "bbox": [ + 82, + 70, + 531, + 242 + ], + "spans": [ + { + "bbox": [ + 82, + 70, + 531, + 242 + ], + "type": "image", + "image_path": "5329bcb88a94f9db5a85eaceb195c1588ff663fb1b0ce6f34d6c648df6887191.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 242, + 553, + 264 + ], + "lines": [ + { + "bbox": [ + 55, + 242, + 553, + 264 + ], + "spans": [ + { + "bbox": [ + 55, + 242, + 553, + 264 + ], + "type": "text", + "content": "Figure 5. An overview of CoIBA. CoIBA restricts the information flow using a universal damping ratio while upper-bounding the information of each layer. After that, CoIBA aggregates the relevant information from each layer to produce an attribution map." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 272, + 296, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 272, + 296, + 402 + ], + "spans": [ + { + "bbox": [ + 55, + 272, + 296, + 402 + ], + "type": "text", + "content": "Here, the trade-off parameter " + }, + { + "bbox": [ + 55, + 272, + 296, + 402 + ], + "type": "inline_equation", + "content": "\\beta_{l}" + }, + { + "bbox": [ + 55, + 272, + 296, + 402 + ], + "type": "text", + "content": " governs the degree of compression in each layer. Since the bottleneck layers are inserted into the sequence of layers, the compression term includes the accumulated sequence of mutual information of bottleneck variables " + }, + { + "bbox": [ + 55, + 272, + 296, + 402 + ], + "type": "inline_equation", + "content": "I[Z_{l-1};Z_l]" + }, + { + "bbox": [ + 55, + 272, + 296, + 402 + ], + "type": "text", + "content": ". Here, the " + }, + { + "bbox": [ + 55, + 272, + 296, + 402 + ], + "type": "inline_equation", + "content": "Z_0" + }, + { + "bbox": [ + 55, + 272, + 296, + 402 + ], + "type": "text", + "content": " indicates the original intermediate representation as the first element, such that " + }, + { + "bbox": [ + 55, + 272, + 296, + 402 + ], + "type": "inline_equation", + "content": "Z_0 = R_1" + }, + { + "bbox": [ + 55, + 272, + 296, + 402 + ], + "type": "text", + "content": ". We maximize the relevancy term " + }, + { + "bbox": [ + 55, + 272, + 296, + 402 + ], + "type": "inline_equation", + "content": "I[Z_L;Y]" + }, + { + "bbox": [ + 55, + 272, + 296, + 402 + ], + "type": "text", + "content": " while compression to maintain the shared information between bottleneck variables and the ground truth. We maximize this term by minimizing the cross-entropy loss " + }, + { + "bbox": [ + 55, + 272, + 296, + 402 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{ce}" + }, + { + "bbox": [ + 55, + 272, + 296, + 402 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 403, + 296, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 403, + 296, + 429 + ], + "spans": [ + { + "bbox": [ + 55, + 403, + 296, + 429 + ], + "type": "text", + "content": "Utilizing KL divergence, the mutual information between internal representations " + }, + { + "bbox": [ + 55, + 403, + 296, + 429 + ], + "type": "inline_equation", + "content": "I[Z_{l-1}; Z_l]" + }, + { + "bbox": [ + 55, + 403, + 296, + 429 + ], + "type": "text", + "content": " is formulated by:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 70, + 444, + 295, + 459 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 444, + 295, + 459 + ], + "spans": [ + { + "bbox": [ + 70, + 444, + 295, + 459 + ], + "type": "interline_equation", + "content": "I \\left[ Z _ {l - 1}; Z _ {l} \\right] = \\mathbb {E} _ {Z _ {l - 1}} \\left[ D _ {K L} \\left[ P \\left(Z _ {l} \\mid Z _ {l - 1}\\right) \\right\\vert \\left| P \\left(Z _ {l}\\right) \\right] \\right] \\tag {5}", + "image_path": "d36f845b6b1731236df96d43e57f6c5e022e26652e9af55e5909ef974a5958bd.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 463, + 296, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 463, + 296, + 547 + ], + "spans": [ + { + "bbox": [ + 55, + 463, + 296, + 547 + ], + "type": "text", + "content": "However, the direct estimation of the prior distribution " + }, + { + "bbox": [ + 55, + 463, + 296, + 547 + ], + "type": "inline_equation", + "content": "P(Z_{l})" + }, + { + "bbox": [ + 55, + 463, + 296, + 547 + ], + "type": "text", + "content": " is intractable as it requires the integration of bottleneck representations " + }, + { + "bbox": [ + 55, + 463, + 296, + 547 + ], + "type": "inline_equation", + "content": "P(Z_{l}) = \\int_{Z_{l - 1}}P(Z_{l}|Z_{l - 1})P(Z_{l - 1})dZ_{l - 1}" + }, + { + "bbox": [ + 55, + 463, + 296, + 547 + ], + "type": "text", + "content": " of the corresponding " + }, + { + "bbox": [ + 55, + 463, + 296, + 547 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 55, + 463, + 296, + 547 + ], + "type": "text", + "content": "-th layer. Thus, we assume the prior distribution " + }, + { + "bbox": [ + 55, + 463, + 296, + 547 + ], + "type": "inline_equation", + "content": "Q(Z_{l})" + }, + { + "bbox": [ + 55, + 463, + 296, + 547 + ], + "type": "text", + "content": " as Gaussian distribution " + }, + { + "bbox": [ + 55, + 463, + 296, + 547 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(\\mu_{R_l},\\sigma_{R_l}^2)" + }, + { + "bbox": [ + 55, + 463, + 296, + 547 + ], + "type": "text", + "content": ", reconstituting the mutual information computation as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 69, + 552, + 295, + 567 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 552, + 295, + 567 + ], + "spans": [ + { + "bbox": [ + 69, + 552, + 295, + 567 + ], + "type": "interline_equation", + "content": "I \\left[ Z _ {l}; Z _ {l - 1} \\right] \\leq \\mathbb {E} _ {Z _ {l - 1}} \\left[ D _ {K L} \\left[ P \\left(Z _ {l} \\mid Z _ {l - 1}\\right) \\right\\vert \\left| Q \\left(Z _ {l}\\right) \\right] \\right]. \\tag {6}", + "image_path": "334c00230c6c5b5e17fe36819e90cb98e793c652fdbff34f5be1f5dccb75e086.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 570, + 296, + 618 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 296, + 618 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 296, + 618 + ], + "type": "text", + "content": "Since assuming the prior distribution with Gaussian distribution only overestimates the mutual information, minimizing this term suppresses the mutual information of the compression term." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 624, + 197, + 637 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 624, + 197, + 637 + ], + "spans": [ + { + "bbox": [ + 55, + 624, + 197, + 637 + ], + "type": "text", + "content": "3.4. Variational Upper Bound" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": "As discussed in Sec. 4.6, the relevant information obtained by suppressing layer-wise mutual information suffers from reflecting layer-wise relevant information due to the inconsistent amount of information along the layers. Thus, obtaining the sequence of trade-off parameters " + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\{\\beta_l\\}_{l = 1}^L" + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": " requires an extensive heuristic search to balance the degree" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 272, + 555, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 272, + 555, + 415 + ], + "spans": [ + { + "bbox": [ + 313, + 272, + 555, + 415 + ], + "type": "text", + "content": "of compression for each layer. To overcome this issue, we suggest the variational upper bound for CoIBA, referring to the two following points. First, since there is no additional information propagated during the forwarding pass, the mutual information of intermediate layers would not be greater than the mutual information between the model input and output. Second, the iterated noise-injection procedure diminishes the mutual information among the internal and bottleneck representations, i.e., " + }, + { + "bbox": [ + 313, + 272, + 555, + 415 + ], + "type": "inline_equation", + "content": "I[Z_{l};Z_{l + 1}] \\leq I[Z_{l - 1};Z_l]" + }, + { + "bbox": [ + 313, + 272, + 555, + 415 + ], + "type": "text", + "content": ". Building upon these bases, we establish an upper bound that encompasses the combined mutual information from all subsequent layers as follows:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 353, + 432, + 555, + 455 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 432, + 555, + 455 + ], + "spans": [ + { + "bbox": [ + 353, + 432, + 555, + 455 + ], + "type": "interline_equation", + "content": "I \\left[ R _ {1}; Z _ {1} \\right] \\geq \\frac {1}{L} \\left(\\sum_ {l = 1} ^ {L} I \\left[ Z _ {l - 1}; Z _ {l} \\right]\\right). \\tag {7}", + "image_path": "92f45d413fe88176d18f46b98d409e16fb0d40a2045785247a3ec8e99d7f2315.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 471, + 554, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 471, + 554, + 495 + ], + "spans": [ + { + "bbox": [ + 313, + 471, + 554, + 495 + ], + "type": "text", + "content": "Accordingly, we reconstitute the objective to be a simplified formula, leaving only a single hyperparameter " + }, + { + "bbox": [ + 313, + 471, + 554, + 495 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 313, + 471, + 554, + 495 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 373, + 523, + 553, + 541 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 373, + 523, + 553, + 541 + ], + "spans": [ + { + "bbox": [ + 373, + 523, + 553, + 541 + ], + "type": "interline_equation", + "content": "\\max _ {\\lambda} I \\left[ Z _ {L}; Y \\right] - \\beta I \\left[ R _ {1}; Z _ {1} \\right]. \\tag {8}", + "image_path": "c214570e2791df30829fd9af22f6c5a67371f8a8b2cc3d95f2e924a39a349424.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "text", + "content": "To compress the subsequent layers, the simplified objective necessitates calculating only the mutual information of the first layer " + }, + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "inline_equation", + "content": "I[R_1;Z_1]" + }, + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "text", + "content": ". This term solely utilizes the non-perturbed intermediate and bottleneck representations of the first layer, simplifying the objective calculation. In terms of relevant information computation, referring to the data processing inequality, the inequality " + }, + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "inline_equation", + "content": "I[Y;Z_l]\\leq I[Y;Z_{l - 1}]" + }, + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "text", + "content": " holds. Thereby, in contrast to the layer-specific information bottleneck, which overestimates mutual information in the earlier layers during compression [37], CoIBA relieves the overestimation. This is because, as discussed in Sec. 4.6, the relevant information is compensated as the sequence of bottleneck variables joins the objective computation." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "25170" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 63, + 107, + 116, + 159 + ], + "blocks": [ + { + "bbox": [ + 63, + 107, + 116, + 159 + ], + "lines": [ + { + "bbox": [ + 63, + 107, + 116, + 159 + ], + "spans": [ + { + "bbox": [ + 63, + 107, + 116, + 159 + ], + "type": "image", + "image_path": "0ba7fcfd7dfa2e1358c8ffea180b50930af878b8e1d18f727a621b398f5f047d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 66, + 160, + 114, + 168 + ], + "lines": [ + { + "bbox": [ + 66, + 160, + 114, + 168 + ], + "spans": [ + { + "bbox": [ + 66, + 160, + 114, + 168 + ], + "type": "text", + "content": "Original Image" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 123, + 72, + 178, + 131 + ], + "blocks": [ + { + "bbox": [ + 123, + 72, + 178, + 131 + ], + "lines": [ + { + "bbox": [ + 123, + 72, + 178, + 131 + ], + "spans": [ + { + "bbox": [ + 123, + 72, + 178, + 131 + ], + "type": "image", + "image_path": "5dedc72a277d60fc8f56dca2139d50e1956c2f02f5427820d8573d1a3feec967.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 120, + 133, + 181, + 194 + ], + "blocks": [ + { + "bbox": [ + 120, + 133, + 181, + 194 + ], + "lines": [ + { + "bbox": [ + 120, + 133, + 181, + 194 + ], + "spans": [ + { + "bbox": [ + 120, + 133, + 181, + 194 + ], + "type": "image", + "image_path": "7df66507cdc9f2b5db097cb5bedcb50b10da014aa82e1adf00d3bf2d7dacec33.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 199, + 555, + 232 + ], + "lines": [ + { + "bbox": [ + 55, + 199, + 555, + 232 + ], + "spans": [ + { + "bbox": [ + 55, + 199, + 555, + 232 + ], + "type": "text", + "content": "Figure 6. Quantitative comprehensive assessment on FunnyBirds experiment: The abbreviations Com, Cor, and Con indicate Completeness, correctness, and contrastivity. The center score indicates the comprehensive result of enumerated aspects. Acc. and B.I. indicate accuracy and background independence, respectively. We provide detailed numeric in the supplementary material." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 183, + 72, + 237, + 131 + ], + "blocks": [ + { + "bbox": [ + 183, + 72, + 237, + 131 + ], + "lines": [ + { + "bbox": [ + 183, + 72, + 237, + 131 + ], + "spans": [ + { + "bbox": [ + 183, + 72, + 237, + 131 + ], + "type": "image", + "image_path": "40447f6b45fbf2d809da226459508b189ce1ac6933ff8e11da8d4cde6378a025.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 181, + 133, + 241, + 194 + ], + "blocks": [ + { + "bbox": [ + 181, + 133, + 241, + 194 + ], + "lines": [ + { + "bbox": [ + 181, + 133, + 241, + 194 + ], + "spans": [ + { + "bbox": [ + 181, + 133, + 241, + 194 + ], + "type": "image", + "image_path": "b57b480c4902f9a344f5915bba0b67219765a4f14ba1714da3dd677bb372b8d1.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 244, + 72, + 297, + 131 + ], + "blocks": [ + { + "bbox": [ + 244, + 72, + 297, + 131 + ], + "lines": [ + { + "bbox": [ + 244, + 72, + 297, + 131 + ], + "spans": [ + { + "bbox": [ + 244, + 72, + 297, + 131 + ], + "type": "image", + "image_path": "892b1921c46b9d374b3387ad2068b6224b3f25267fdc6d88fbf1a7026afc930b.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 242, + 133, + 303, + 194 + ], + "blocks": [ + { + "bbox": [ + 242, + 133, + 303, + 194 + ], + "lines": [ + { + "bbox": [ + 242, + 133, + 303, + 194 + ], + "spans": [ + { + "bbox": [ + 242, + 133, + 303, + 194 + ], + "type": "image", + "image_path": "bbe59162935a3cd7e403b8fefc9c607cdddf12c8ba9318e7761230279542e123.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 304, + 72, + 358, + 131 + ], + "blocks": [ + { + "bbox": [ + 304, + 72, + 358, + 131 + ], + "lines": [ + { + "bbox": [ + 304, + 72, + 358, + 131 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 358, + 131 + ], + "type": "image", + "image_path": "c79670731198fe2b38be3294e6fbad593bbc07532e6b65023c3d72b0230be279.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 301, + 133, + 361, + 194 + ], + "blocks": [ + { + "bbox": [ + 301, + 133, + 361, + 194 + ], + "lines": [ + { + "bbox": [ + 301, + 133, + 361, + 194 + ], + "spans": [ + { + "bbox": [ + 301, + 133, + 361, + 194 + ], + "type": "image", + "image_path": "e9269a4742934d318b2423986857a620cc128cf159f584edd8f09ec67d2ca5a4.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 362, + 72, + 416, + 131 + ], + "blocks": [ + { + "bbox": [ + 362, + 72, + 416, + 131 + ], + "lines": [ + { + "bbox": [ + 362, + 72, + 416, + 131 + ], + "spans": [ + { + "bbox": [ + 362, + 72, + 416, + 131 + ], + "type": "image", + "image_path": "57ffe9db5e70f8ab8c5e9cb05ce22613b272691982b7cf62f77e0123fdfa8f7a.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 362, + 133, + 421, + 194 + ], + "blocks": [ + { + "bbox": [ + 362, + 133, + 421, + 194 + ], + "lines": [ + { + "bbox": [ + 362, + 133, + 421, + 194 + ], + "spans": [ + { + "bbox": [ + 362, + 133, + 421, + 194 + ], + "type": "image", + "image_path": "ccde3b81382f1229c3d4e2edeb1bc7715770bf0d6667ad4d4b5bb270d4c46d24.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 422, + 72, + 477, + 131 + ], + "blocks": [ + { + "bbox": [ + 422, + 72, + 477, + 131 + ], + "lines": [ + { + "bbox": [ + 422, + 72, + 477, + 131 + ], + "spans": [ + { + "bbox": [ + 422, + 72, + 477, + 131 + ], + "type": "image", + "image_path": "c22718f40152dac6e2c8674187f79fc659bb1de71e3ba08f64c6142b1026d3be.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 421, + 133, + 482, + 194 + ], + "blocks": [ + { + "bbox": [ + 421, + 133, + 482, + 194 + ], + "lines": [ + { + "bbox": [ + 421, + 133, + 482, + 194 + ], + "spans": [ + { + "bbox": [ + 421, + 133, + 482, + 194 + ], + "type": "image", + "image_path": "6ac53393064b023c4ed9766f675f34fb897f4dbf1f4a82d3394efb34c3f6620f.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 484, + 72, + 537, + 131 + ], + "blocks": [ + { + "bbox": [ + 484, + 72, + 537, + 131 + ], + "lines": [ + { + "bbox": [ + 484, + 72, + 537, + 131 + ], + "spans": [ + { + "bbox": [ + 484, + 72, + 537, + 131 + ], + "type": "image", + "image_path": "a3d3a2e0e3637f5938f4e9a4af2bf8d23afb07b0d4f5e17a342716fbe36c8bde.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 481, + 133, + 544, + 194 + ], + "blocks": [ + { + "bbox": [ + 481, + 133, + 544, + 194 + ], + "lines": [ + { + "bbox": [ + 481, + 133, + 544, + 194 + ], + "spans": [ + { + "bbox": [ + 481, + 133, + 544, + 194 + ], + "type": "image", + "image_path": "9b6881e286591e9db51bf99ee5ee2220d828333ebdd128cd02aad2254639e28b.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "table", + "bbox": [ + 82, + 236, + 526, + 348 + ], + "blocks": [ + { + "bbox": [ + 82, + 236, + 526, + 348 + ], + "lines": [ + { + "bbox": [ + 82, + 236, + 526, + 348 + ], + "spans": [ + { + "bbox": [ + 82, + 236, + 526, + 348 + ], + "type": "table", + "html": "
VariantModelChefer-LRP [7]Generic [6]IIA [3]ViT-CX [35]IBA [27]Beyond [8]ColBA
ViTViT*-S-16/224-14.56/56.0813.69/57.3114.98/56.6513.56/58.6413.50/58.4711.08/63.35
ViT*-T-16/224-9.00/48.958.50/49.7512.08/45.098.02/49.2211.26/44.076.97/54.23
ViT-B-16/22417.67/56.6917.92/57.0316.56/58.716.94/58.1717.23/59.0615.84/60.3713.01/62.58
ViT-L-16/22420.97/53.8120.17/55.5420.14/55.2219.76/57.3820.88/54.8518.99/56.5515.40/61.14
DeiTDeiT-S-16/22412.76/47.5112.10/48.6410.89/49.7617.57/44.4311.43/47.8711.13/48.889.55/53.38
DeiT-B-16/22414.63/48.7915.10/49.4113.69/50.5118.21/46.9713.95/50.1814.45/48.8811.79/53.96
DeiT3-B-16/224-15.70/52.4215.38/52.7018.88/50.4015.07/53.3214.71/54.0712.97/56.54
DeiT3-L-16/224-22.10/62.7317.75/55.9326.11/59.6021.01/65.6019.34/64.2717.64/67.20
SwinSwin-B-33.76/42.94--18.03/52.8123.73/50.3917.09/54.80
Swin2-B-34.71/47.16--20.94/53.8531.32/49.4418.88/55.77
", + "image_path": "5ad83bac8c93f53ab8a1d54c35fa4b1307e03ccb352d2a61e6fb7c876b50e462.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "table_body" + } + ], + "index": 17 + }, + { + "bbox": [ + 55, + 350, + 555, + 373 + ], + "lines": [ + { + "bbox": [ + 55, + 350, + 555, + 373 + ], + "spans": [ + { + "bbox": [ + 55, + 350, + 555, + 373 + ], + "type": "text", + "content": "Table 1. Quantitative feature importance assessment on insertion ↑/deletion ↓. (*) denotes the ViT trained with strong regularization techniques [29]. We underline the state-of-the-art performance among the baselines." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 378, + 137, + 391 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 378, + 137, + 391 + ], + "spans": [ + { + "bbox": [ + 55, + 378, + 137, + 391 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 55, + 401, + 105, + 414 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 401, + 105, + 414 + ], + "spans": [ + { + "bbox": [ + 55, + 401, + 105, + 414 + ], + "type": "text", + "content": "4.1. Setup" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 55, + 422, + 296, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 422, + 296, + 590 + ], + "spans": [ + { + "bbox": [ + 55, + 422, + 296, + 590 + ], + "type": "text", + "content": "Architectures We select variants of ViT for our experiment, including original ViT [11] and DeiT models [32, 33], trained with ImageNet-21k (IN-21k) and ImageNet-1k (IN-1k), respectively. We include the variants in depth interpolated from ViT-T, ViT-S, ViT-B, ViT-L, and ViT-H models. For the ViT-T and ViT-S models, we utilize the model pretrained with massive regularization techniques [29]. The ViT-H model is pre-trained with CLIP [24]. We present the settings of the ViT-B model with patch size 16 and image resolution 224 as ViT-B-16/224. We include Swin transformers [20, 21], demonstrating the generalized ability of CoIBA in multi-scale features processing. The details of settings and further results about convolutional neural networks are included in the supplementary material." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "text", + "content": "Feature Attribution Methods We compare the various feature attribution methods including Chefer-LRP [7], Generic [6], IIA [3], ViT-CX [35], Beyond [8], and IBA [27] as baselines. We select " + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\beta = 10" + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "text", + "content": " for IBA and 6-th layer to insert bottleneck. For CoIBA, we select tradeoff hyper-parameter " + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "text", + "content": " as 1. We insert the bottleneck layer from " + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "text", + "content": "-th departure to " + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "inline_equation", + "content": "e" + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "text", + "content": "-th arrival layers, which are chosen, 4, and 12, respectively, for ViT-B. For the optimization, we set the learning rate, batch size, and optimizer as 1, 10, and Adam [17]. The additional hyperparameter settings" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 379, + 555, + 438 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 379, + 555, + 438 + ], + "spans": [ + { + "bbox": [ + 313, + 379, + 555, + 438 + ], + "type": "text", + "content": "are included in the supplementary material. For Swin transformers, we only report the explanation methods that do not require methodology modification, including Generic, ViTCX, IBA, Beyond, and CoIBA. We utilize RTX A6000 GPU for all the experiments." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 439, + 554, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 439, + 554, + 487 + ], + "spans": [ + { + "bbox": [ + 313, + 439, + 554, + 487 + ], + "type": "text", + "content": "Datasets We utilize ImageNet-1k [9] (IN-1k), ImageNet-A (IN-A) [14] and ImageNet-R (IN-R) [13] validation datasets for the experiments. We leverage IN-A and IN-R to conduct the difficult-aware analysis." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 493, + 450, + 506 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 493, + 450, + 506 + ], + "spans": [ + { + "bbox": [ + 313, + 493, + 450, + 506 + ], + "type": "text", + "content": "4.2. FunnyBirds Assessment" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 510, + 555, + 630 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 555, + 630 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 555, + 630 + ], + "type": "text", + "content": "The FunnyBirds [15] experiment provides the ground truth for the evaluation. With the help of ground truth, the Funnybird framework assesses the quality of the attribution map from three perspectives: completeness, correctness, and contrastivity. Fig. 6 illustrates the qualitative and quantitative results of the Funnybirds experiment. As shown in the results, CoIBA outperforms all the baselines, including the propagation or gradient-based approaches. Furthermore, in contrast to IBA, CoIBA obtains an attribution map with significantly enhanced contrastivity." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 636, + 422, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 636, + 422, + 647 + ], + "spans": [ + { + "bbox": [ + 313, + 636, + 422, + 647 + ], + "type": "text", + "content": "4.3. Insertion/Deletion" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 313, + 653, + 554, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 554, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 554, + 712 + ], + "type": "text", + "content": "Insertion/deletion [23] measures the correctness of feature importance highlighted by an attribution map. The insertion/deletion method gradually inserts or deletes informative pixels in ascending order of attribution to compute the area under the curve scores. In insertion/deletion, methods" + } + ] + } + ], + "index": 28 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25171" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 89, + 71, + 519, + 138 + ], + "blocks": [ + { + "bbox": [ + 89, + 71, + 519, + 138 + ], + "lines": [ + { + "bbox": [ + 89, + 71, + 519, + 138 + ], + "spans": [ + { + "bbox": [ + 89, + 71, + 519, + 138 + ], + "type": "table", + "html": "
ModelChefer-LRP [7]Generic [6]IIA [3]ViT-CX [35]IBA [27]Beyond [8]CoIBA
ViT-B-16/22420.87/64.3722.15/65.3119.58/66.1919.04/65.3620.99/68.8017.93/68.4416.63/73.68
DeiT-B-16/22415.31/59.5516.76/59.7814.74/61.2118.37/57.4415.62/61.5315.37/60.0611.59/67.12
ViT-L-16/22429.92/64.8728.31/66.5128.39/65.9224.30/67.2730.55/66.9725.47/66.6822.03/74.88
DeiT3-L-16/224-20.28/69.8719.72/70.0224.43/66.6019.56/71.2319.41/70.5215.02/76.13
ViT†-H-16/224-30.06/61.3432.27/60.9830.34/57.7028.30/62.2928.11/62.9623.38/66.95
", + "image_path": "e5c1ecb1c7ba21e05f5406a351fffe37ff0a739f41176bb087dadb1774354ce4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 140, + 555, + 173 + ], + "lines": [ + { + "bbox": [ + 55, + 140, + 555, + 173 + ], + "spans": [ + { + "bbox": [ + 55, + 140, + 555, + 173 + ], + "type": "text", + "content": "Table 2. Quantitative feature importance evaluation of ROAD (MoRF ↓ / LeRF ↑). The higher and lower scores indicate better attribution map quality for MoRF and LeRF, respectively. The underlined scores indicate the highest performance among the baselines. " + }, + { + "bbox": [ + 55, + 140, + 555, + 173 + ], + "type": "inline_equation", + "content": "(\\dagger)" + }, + { + "bbox": [ + 55, + 140, + 555, + 173 + ], + "type": "text", + "content": " denotes the ViT trained with CLIP [24]." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 60, + 175, + 89, + 225 + ], + "blocks": [ + { + "bbox": [ + 60, + 175, + 89, + 225 + ], + "lines": [ + { + "bbox": [ + 60, + 175, + 89, + 225 + ], + "spans": [ + { + "bbox": [ + 60, + 175, + 89, + 225 + ], + "type": "image", + "image_path": "38daf34b912ff5644bcfbf51b07b446c9bc5e6300ea98a1a586141d6339e058d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 94, + 175, + 182, + 247 + ], + "blocks": [ + { + "bbox": [ + 94, + 175, + 182, + 247 + ], + "lines": [ + { + "bbox": [ + 94, + 175, + 182, + 247 + ], + "spans": [ + { + "bbox": [ + 94, + 175, + 182, + 247 + ], + "type": "image", + "image_path": "698acb6aba24360032fd53e45ce016c0a3292ed7b020b1202fc6fe6cd49f020c.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 186, + 174, + 274, + 247 + ], + "blocks": [ + { + "bbox": [ + 186, + 174, + 274, + 247 + ], + "lines": [ + { + "bbox": [ + 186, + 174, + 274, + 247 + ], + "spans": [ + { + "bbox": [ + 186, + 174, + 274, + 247 + ], + "type": "image", + "image_path": "24aa980031e7d7b42e2490a466ec45a1e67f8c8c6e8a32877ccc1bc2ca8b3850.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 198, + 248, + 230, + 257 + ], + "lines": [ + { + "bbox": [ + 198, + 248, + 230, + 257 + ], + "spans": [ + { + "bbox": [ + 198, + 248, + 230, + 257 + ], + "type": "text", + "content": "(a) IN-1k" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 277, + 173, + 365, + 247 + ], + "blocks": [ + { + "bbox": [ + 277, + 173, + 365, + 247 + ], + "lines": [ + { + "bbox": [ + 277, + 173, + 365, + 247 + ], + "spans": [ + { + "bbox": [ + 277, + 173, + 365, + 247 + ], + "type": "image", + "image_path": "f7210558684cac70f3ec1b18c459686b2f034fd3b178f0227808215a8086920e.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 371, + 173, + 459, + 247 + ], + "blocks": [ + { + "bbox": [ + 371, + 173, + 459, + 247 + ], + "lines": [ + { + "bbox": [ + 371, + 173, + 459, + 247 + ], + "spans": [ + { + "bbox": [ + 371, + 173, + 459, + 247 + ], + "type": "image", + "image_path": "6e662039c23d65a71ac32957bfa9a9f93fde5c72a23a7fa178ffe0c945b5908d.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 402, + 248, + 432, + 257 + ], + "lines": [ + { + "bbox": [ + 402, + 248, + 432, + 257 + ], + "spans": [ + { + "bbox": [ + 402, + 248, + 432, + 257 + ], + "type": "text", + "content": "(b) IN-A" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 465, + 173, + 553, + 247 + ], + "blocks": [ + { + "bbox": [ + 465, + 173, + 553, + 247 + ], + "lines": [ + { + "bbox": [ + 465, + 173, + 553, + 247 + ], + "spans": [ + { + "bbox": [ + 465, + 173, + 553, + 247 + ], + "type": "image", + "image_path": "6fab59fa92ab46768056ab1f3309f68e827e4aff583f411e7341b54908e3f22c.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 496, + 248, + 525, + 257 + ], + "lines": [ + { + "bbox": [ + 496, + 248, + 525, + 257 + ], + "spans": [ + { + "bbox": [ + 496, + 248, + 525, + 257 + ], + "type": "text", + "content": "(c) IN-R" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 304, + 296, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 304, + 296, + 413 + ], + "spans": [ + { + "bbox": [ + 55, + 304, + 296, + 413 + ], + "type": "text", + "content": "providing better correctness performance achieve higher scores for insertion and lower scores for deletion. We gradually deleted or inserted around every " + }, + { + "bbox": [ + 55, + 304, + 296, + 413 + ], + "type": "inline_equation", + "content": "3.5\\%" + }, + { + "bbox": [ + 55, + 304, + 296, + 413 + ], + "type": "text", + "content": " pixels of the input image and randomly sampled 6,000 samples from the IN-1k validation dataset for the assessment. As shown in Table 1, the attributions provided by CoIBA yield predominant correctness performance compared to baselines. Especially, the consistent increase in correctness from the Swin transformer demonstrates the generalizability of CoIBA." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 420, + 170, + 431 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 420, + 170, + 431 + ], + "spans": [ + { + "bbox": [ + 55, + 420, + 170, + 431 + ], + "type": "text", + "content": "4.4. Remove and Debias" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 437, + 296, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 437, + 296, + 593 + ], + "spans": [ + { + "bbox": [ + 55, + 437, + 296, + 593 + ], + "type": "text", + "content": "The remove-and-debias (ROAD) [25] experiments measure the correctness of the attribution maps, addressing the class information leakage problem (i.e., the shape of the mask) observed in the remove-and-retrain (ROAR) [16] experiment. We average the quantitative scores over the 6,000 randomly selected images when 20, 40, 60, and " + }, + { + "bbox": [ + 55, + 437, + 296, + 593 + ], + "type": "inline_equation", + "content": "80\\%" + }, + { + "bbox": [ + 55, + 437, + 296, + 593 + ], + "type": "text", + "content": " of pixels are imputed in two perspectives: most relevant first (MoRF) and least relevant first (LeRF). As shown in Table 2, CoIBA outperforms existing methods in MoRF and LeRF experiments with a sizable gap. These results demonstrate that regardless of the imputation type, CoIBA outperforms all the baselines regardless of the model setting (patch size and depth) and pre-trained dataset." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 600, + 269, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 600, + 269, + 612 + ], + "spans": [ + { + "bbox": [ + 55, + 600, + 269, + 612 + ], + "type": "text", + "content": "4.5. Difficulty-aware Correctness Assessment" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 55, + 618, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 296, + 713 + ], + "type": "text", + "content": "We compare the quantitative quality of attribution maps yielded by CoIBA and baseline methods in terms of difficulty-aware assessment, in addition to correctness assessments of overall samples. Leveraging the confidence scores, and the prediction probability of a target class, we divide the confidence scores with 20 intervals as in Sec. 3.2. As shown in Fig. 7, CoIBA consistently outperforms all the baselines regardless of the confidence scores predicted with" + } + ] + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 323, + 301, + 430, + 380 + ], + "blocks": [ + { + "bbox": [ + 55, + 258, + 555, + 301 + ], + "lines": [ + { + "bbox": [ + 55, + 258, + 555, + 301 + ], + "spans": [ + { + "bbox": [ + 55, + 258, + 555, + 301 + ], + "type": "text", + "content": "Figure 7. Quantitative difficulty-aware correctness assessment on insertion/deletion and ROAD. We measure differences in insertion/deletion " + }, + { + "bbox": [ + 55, + 258, + 555, + 301 + ], + "type": "inline_equation", + "content": "(\\Delta \\mathrm{InsDel}\\uparrow)" + }, + { + "bbox": [ + 55, + 258, + 555, + 301 + ], + "type": "text", + "content": " and ROAD " + }, + { + "bbox": [ + 55, + 258, + 555, + 301 + ], + "type": "inline_equation", + "content": "(\\Delta \\mathrm{ROAD}\\uparrow)" + }, + { + "bbox": [ + 55, + 258, + 555, + 301 + ], + "type": "text", + "content": " scores. We fill the regions including low-confident samples and high-confident samples with red and blue, respectively, based on the prediction made by the model. The first to third columns illustrate the quantitative results of the IN-1k. The fourth and fifth columns include the results of IN-A and IN-R datasets, respectively." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 323, + 301, + 430, + 380 + ], + "lines": [ + { + "bbox": [ + 323, + 301, + 430, + 380 + ], + "spans": [ + { + "bbox": [ + 323, + 301, + 430, + 380 + ], + "type": "image", + "image_path": "ef0cc9d5cd981bbbfb5d8e55d314414e53f46e2bf1b593b33106c2e17cffaf6b.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 328, + 381, + 428, + 390 + ], + "lines": [ + { + "bbox": [ + 328, + 381, + 428, + 390 + ], + "spans": [ + { + "bbox": [ + 328, + 381, + 428, + 390 + ], + "type": "text", + "content": "(a) Ablation on departure layer" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 391, + 554, + 468 + ], + "lines": [ + { + "bbox": [ + 313, + 391, + 554, + 468 + ], + "spans": [ + { + "bbox": [ + 313, + 391, + 554, + 468 + ], + "type": "text", + "content": "Figure 8. Comparisons of different departure " + }, + { + "bbox": [ + 313, + 391, + 554, + 468 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 313, + 391, + 554, + 468 + ], + "type": "text", + "content": " and arrival " + }, + { + "bbox": [ + 313, + 391, + 554, + 468 + ], + "type": "inline_equation", + "content": "e" + }, + { + "bbox": [ + 313, + 391, + 554, + 468 + ], + "type": "text", + "content": " layer settings. We plot the hyper-parameters selected for CoIBA as a solid line. We compare the quantitative results of the discrepancy between AUC scores of insertion/deletion (ΔInsDel) with different intervals of confidence scores outputted by the model. The better attribution method achieves a higher score. We provide further results in the supplementary material." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 437, + 301, + 544, + 380 + ], + "blocks": [ + { + "bbox": [ + 437, + 301, + 544, + 380 + ], + "lines": [ + { + "bbox": [ + 437, + 301, + 544, + 380 + ], + "spans": [ + { + "bbox": [ + 437, + 301, + 544, + 380 + ], + "type": "image", + "image_path": "523a6a810e60cff76795b6fa352d6d38ad0ae93ed88fbddb8baedbc86602ca14.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 446, + 381, + 537, + 390 + ], + "lines": [ + { + "bbox": [ + 446, + 381, + 537, + 390 + ], + "spans": [ + { + "bbox": [ + 446, + 381, + 537, + 390 + ], + "type": "text", + "content": "(b) Ablation on arrival layer" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 473, + 555, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 473, + 555, + 545 + ], + "spans": [ + { + "bbox": [ + 313, + 473, + 555, + 545 + ], + "type": "text", + "content": "each sample, predicted by the model. Along with IN-1k, we include the quantitative results of IN-A and IN-R. As both IN-A and IN-R datasets include the low-confident (difficult) samples compared to IN-1k, these quantitative results support the demonstration that CoIBA outperforms the correct performance with low-confident samples." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 314, + 552, + 387, + 564 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 552, + 387, + 564 + ], + "spans": [ + { + "bbox": [ + 314, + 552, + 387, + 564 + ], + "type": "text", + "content": "4.6. Discussion" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "content": "Ablation on Bottleneck-inserted Layers In this section, we confirm whether including multiple layers enhances the correctness of the resulting attribution map. Fig. 9 shows the comparison of the quantitative results by interpolating departure " + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "inline_equation", + "content": "(s)" + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "content": " and arrival " + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "inline_equation", + "content": "(e)" + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "content": " layers from earlier to deeper layers. The results demonstrate that increasing the number of layers to compute relevant information consistently enhances the correctness quality of an attribution map. Furthermore, CoIBA consistently outperforms IBA* with large margins, requiring only a single iteration. Thus, in CoIBA, various layers gladly reveal the omitted relevant information obtained from a specific layer." + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "25172" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 66, + 72, + 174, + 145 + ], + "blocks": [ + { + "bbox": [ + 66, + 72, + 174, + 145 + ], + "lines": [ + { + "bbox": [ + 66, + 72, + 174, + 145 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 174, + 145 + ], + "type": "image", + "image_path": "0705a15d468e8d31d2b6f54a32231b26d8587cce7584e2228f9833cdfead120a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 67, + 146, + 169, + 156 + ], + "lines": [ + { + "bbox": [ + 67, + 146, + 169, + 156 + ], + "spans": [ + { + "bbox": [ + 67, + 146, + 169, + 156 + ], + "type": "text", + "content": "(a) Discrepancy in mutual info." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 179, + 72, + 287, + 145 + ], + "blocks": [ + { + "bbox": [ + 179, + 72, + 287, + 145 + ], + "lines": [ + { + "bbox": [ + 179, + 72, + 287, + 145 + ], + "spans": [ + { + "bbox": [ + 179, + 72, + 287, + 145 + ], + "type": "image", + "image_path": "45776e79e5cabef43808a4176649a9522f46873be0e683faa1468f68954cbf0c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 187, + 147, + 278, + 156 + ], + "lines": [ + { + "bbox": [ + 187, + 147, + 278, + 156 + ], + "spans": [ + { + "bbox": [ + 187, + 147, + 278, + 156 + ], + "type": "text", + "content": "(b) Discrepancy in accuracy" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 66, + 157, + 174, + 236 + ], + "blocks": [ + { + "bbox": [ + 66, + 157, + 174, + 236 + ], + "lines": [ + { + "bbox": [ + 66, + 157, + 174, + 236 + ], + "spans": [ + { + "bbox": [ + 66, + 157, + 174, + 236 + ], + "type": "image", + "image_path": "0fb9e3779c4a5a2e316f8d14bf9c483815558a34cf70660d72d3520759534364.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 111, + 238, + 240, + 247 + ], + "lines": [ + { + "bbox": [ + 111, + 238, + 240, + 247 + ], + "spans": [ + { + "bbox": [ + 111, + 238, + 240, + 247 + ], + "type": "text", + "content": "(c) Ablation on universal damping ratio" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 178, + 157, + 285, + 236 + ], + "blocks": [ + { + "bbox": [ + 178, + 157, + 285, + 236 + ], + "lines": [ + { + "bbox": [ + 178, + 157, + 285, + 236 + ], + "spans": [ + { + "bbox": [ + 178, + 157, + 285, + 236 + ], + "type": "image", + "image_path": "17e00121058d317341001b7ded8e68b8c2333da05f592254e15595c29e14f19f.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 249, + 296, + 304 + ], + "lines": [ + { + "bbox": [ + 55, + 249, + 296, + 304 + ], + "spans": [ + { + "bbox": [ + 55, + 249, + 296, + 304 + ], + "type": "text", + "content": "Figure 9. Ablation study on layers to analyze universal damping ratio. Fig. 9(a) illustrates the change in mutual information for compression terms " + }, + { + "bbox": [ + 55, + 249, + 296, + 304 + ], + "type": "inline_equation", + "content": "I[Z_1; R_1]" + }, + { + "bbox": [ + 55, + 249, + 296, + 304 + ], + "type": "text", + "content": " in Eq. (8) and the Fig. 9(b) shows the change in accuracy. We utilize the ViT-B-16/224 model for these experiments. The mutual info. denotes mutual information." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 316, + 296, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 316, + 296, + 532 + ], + "spans": [ + { + "bbox": [ + 56, + 316, + 296, + 532 + ], + "type": "text", + "content": "Effectiveness of Universal Damping Ratio We investigate whether leveraging the universal damping ratio compensates for over-compression in Fig. 9(a) while amplifying the relevant information in Fig. 9(b). We compare the different numbers of targeted layers to compare the quantitative results. The results illustrated in Fig. 9(a) show that leveraging the universal damping ratio complements the overcompressed information that occurred in earlier layers, due to the delivered relevant information from deeper layers. Concurrently, as shown in Fig. 9(c), leveraging the universal damping ratio amplifies the relevant information term in Eq (8). Aligning with these results, compared by assigning individual damping ratios to each layer, our universal damping ratio enhances the correctness performance of CoIBA. Therefore, leveraging the universal damping ratio to targeted multiple layers significantly enhances the correctness performance of CoIBA by compensating for the over-compression and amplifying the relevant information." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 534, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 296, + 714 + ], + "type": "text", + "content": "Effectiveness of Variational Upper Bound We demonstrate the effectiveness of our variational upper bound by measuring correctness performance and whether the relevancy of different layers is fairly reflected in attributions. We compare quantitative results obtained by leveraging a variational upper bound (Eq. (8)) versus linearly combining layer-wise mutual information (Eq. (4)) in Fig. 10(a). As shown in the results, leveraging our variational upper bound is better at correctly identifying feature importance, regardless of confidence scores. In addition to this result, we confirm whether the relevant information of each layer is fairly reflected in the attribution map by comparing similarities. Concretely, we compare the similarity of both settings (Eq. (4) and (8)) with layer-specific relevant information computed by iterating IBA with channel uniform per" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 328, + 72, + 434, + 148 + ], + "blocks": [ + { + "bbox": [ + 328, + 72, + 434, + 148 + ], + "lines": [ + { + "bbox": [ + 328, + 72, + 434, + 148 + ], + "spans": [ + { + "bbox": [ + 328, + 72, + 434, + 148 + ], + "type": "image", + "image_path": "792d37243daefc43c1d8f6e25b5bad489aee281599deaae8eb20056ab71b83b9.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 336, + 150, + 425, + 159 + ], + "lines": [ + { + "bbox": [ + 336, + 150, + 425, + 159 + ], + "spans": [ + { + "bbox": [ + 336, + 150, + 425, + 159 + ], + "type": "text", + "content": "(a) Faithfulness assessment" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 436, + 72, + 542, + 139 + ], + "blocks": [ + { + "bbox": [ + 436, + 72, + 542, + 139 + ], + "lines": [ + { + "bbox": [ + 436, + 72, + 542, + 139 + ], + "spans": [ + { + "bbox": [ + 436, + 72, + 542, + 139 + ], + "type": "image", + "image_path": "a221e53a85c456d646161a761aa03df00795be600fa8e263a4397249f053a0c9.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 326, + 160, + 433, + 236 + ], + "blocks": [ + { + "bbox": [ + 326, + 160, + 433, + 236 + ], + "lines": [ + { + "bbox": [ + 326, + 160, + 433, + 236 + ], + "spans": [ + { + "bbox": [ + 326, + 160, + 433, + 236 + ], + "type": "image", + "image_path": "28b5420d25d4859eed8b636ed40e2f7d64eeff554de0090a54478ae4d3b2ba7b.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 361, + 239, + 507, + 248 + ], + "lines": [ + { + "bbox": [ + 361, + 239, + 507, + 248 + ], + "spans": [ + { + "bbox": [ + 361, + 239, + 507, + 248 + ], + "type": "text", + "content": "(c) Ablation on channel uniform perturbation" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 437, + 160, + 543, + 236 + ], + "blocks": [ + { + "bbox": [ + 443, + 150, + 533, + 159 + ], + "lines": [ + { + "bbox": [ + 443, + 150, + 533, + 159 + ], + "spans": [ + { + "bbox": [ + 443, + 150, + 533, + 159 + ], + "type": "text", + "content": "(b) Similarity measurement" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 437, + 160, + 543, + 236 + ], + "lines": [ + { + "bbox": [ + 437, + 160, + 543, + 236 + ], + "spans": [ + { + "bbox": [ + 437, + 160, + 543, + 236 + ], + "type": "image", + "image_path": "dc630f201e95b3f713076e6af724de4fc895737fc7c3e4b6e6e3305367f5ad06.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 251, + 554, + 306 + ], + "lines": [ + { + "bbox": [ + 313, + 251, + 554, + 306 + ], + "spans": [ + { + "bbox": [ + 313, + 251, + 554, + 306 + ], + "type": "text", + "content": "Figure 10. Quantitative analysis on CoIBA. We report the results divided by the confidence score computed by the model. For the comparison, we include the ViT-B model pre-trained with CLIP in Fig. 10(c). We filled the discrepancy in performances with blue color in Fig. 10(c). The solid line indicates the results of CoIBA." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 312, + 554, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 312, + 554, + 384 + ], + "spans": [ + { + "bbox": [ + 313, + 312, + 554, + 384 + ], + "type": "text", + "content": "turbation, denoted as " + }, + { + "bbox": [ + 313, + 312, + 554, + 384 + ], + "type": "inline_equation", + "content": "\\mathrm{IBA}^*" + }, + { + "bbox": [ + 313, + 312, + 554, + 384 + ], + "type": "text", + "content": ". We employ CKA [19] as a similarity metric. As shown in Fig. 10(b), leveraging our upper bound yields the comprehensive relevant information of all the layers. This result demonstrates that our variational upper bound encourages the resulting comprehensive relevant information to fairly reflect the relevancy of each layer." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 384, + 554, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 384, + 554, + 529 + ], + "spans": [ + { + "bbox": [ + 313, + 384, + 554, + 529 + ], + "type": "text", + "content": "Effectiveness of Uniform Channel Perturbation We demonstrate the effectiveness of uniform channel perturbation by comparing the correctness performance of uniform (unif.) and non-uniform (non-unif.) channel perturbation strategies. As shown in Fig. 10(c), the uniform channel perturbation consistently enhances the correctness performance of IBA for low-confident samples. In addition to this result, uniform channel perturbation significantly enlarges the correctness of resulting attribution maps, regardless of the confidence scores. These enhanced correctness performances demonstrate the effectiveness of our uniform channel perturbation, regardless of the method." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 538, + 388, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 538, + 388, + 550 + ], + "spans": [ + { + "bbox": [ + 313, + 538, + 388, + 550 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "text", + "content": "In this paper, we introduce CoIBA, which reveals the comprehensive relevant information to produce an attribution map, revealing the omitted relevancy in IBA with a theoretical guarantee. CoIBA shares the universal damping ratio to compensate for the over-compressed information, delivering relevancy among the bottleneck-inserted layers. CoIBA clearly judges the importance of each token by leveraging uniform channel perturbation. We leverage variational approximation to upper bound the information to ensure the eliminated activations are not necessary for the bottleneck-inserted layers to make a decision. CoIBA demonstrates a substantial improvement over existing methods in numerous experiments and discussions." + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25173" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 209, + 115, + 220 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 209, + 115, + 220 + ], + "spans": [ + { + "bbox": [ + 56, + 209, + 115, + 220 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 228, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 228, + 296, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 228, + 296, + 272 + ], + "spans": [ + { + "bbox": [ + 61, + 228, + 296, + 272 + ], + "type": "text", + "content": "[1] Samira Abnar and Willem Zuidema. Quantifying attention flow in transformers. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 4190-4197, 2020. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 272, + 296, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 272, + 296, + 314 + ], + "spans": [ + { + "bbox": [ + 61, + 272, + 296, + 314 + ], + "type": "text", + "content": "[2] Alexander A Alemi, Ian Fischer, Joshua V Dillon, and Kevin Murphy. Deep variational information bottleneck. In International Conference on Learning Representations, 2017. 1, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 317, + 296, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 317, + 296, + 360 + ], + "spans": [ + { + "bbox": [ + 62, + 317, + 296, + 360 + ], + "type": "text", + "content": "[3] Oren Barkan, Yuval Asher, Amit Eshel, Noam Koenigstein, et al. Visual explanations via iterated integrated attributions. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2073-2084, 2023. 1, 2, 4, 6, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 361, + 296, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 361, + 296, + 448 + ], + "spans": [ + { + "bbox": [ + 62, + 361, + 296, + 448 + ], + "type": "text", + "content": "[4] Alexander Binder, Grégoire Montavon, Sebastian Lapuschkin, Klaus-Robert Müller, and Wojciech Samek. Layer-wise relevance propagation for neural networks with local renormalization layers. In Artificial Neural Networks and Machine Learning-ICANN 2016: 25th International Conference on Artificial Neural Networks, Barcelona, Spain, September 6-9, 2016, Proceedings, Part II 25, pages 63-71. Springer, 2016. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 449, + 296, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 449, + 296, + 503 + ], + "spans": [ + { + "bbox": [ + 62, + 449, + 296, + 503 + ], + "type": "text", + "content": "[5] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9650-9660, 2021. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 505, + 296, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 505, + 296, + 557 + ], + "spans": [ + { + "bbox": [ + 62, + 505, + 296, + 557 + ], + "type": "text", + "content": "[6] Hila Chefer, Shir Gur, and Lior Wolf. Generic attention-model explainability for interpreting bi-modal and encoder-decoder transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 397-406, 2021. 1, 2, 6, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 559, + 296, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 559, + 296, + 602 + ], + "spans": [ + { + "bbox": [ + 62, + 559, + 296, + 602 + ], + "type": "text", + "content": "[7] Hila Chefer, Shir Gur, and Lior Wolf. Transformer interpretability beyond attention visualization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 782-791, 2021. 1, 2, 4, 6, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 604, + 296, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 604, + 296, + 647 + ], + "spans": [ + { + "bbox": [ + 62, + 604, + 296, + 647 + ], + "type": "text", + "content": "[8] Jiamin Chen, Xuhong Li, Lei Yu, Dejing Dou, and Haoyi Xiong. Beyond intuition: Rethinking token attributions inside transformers. Transactions on Machine Learning Research, 2022. 1, 2, 4, 6, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 647, + 296, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 647, + 296, + 680 + ], + "spans": [ + { + "bbox": [ + 62, + 647, + 296, + 680 + ], + "type": "text", + "content": "[9] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, pages 248-255. IEEE, 2009. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 681, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 681, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 681, + 295, + 712 + ], + "type": "text", + "content": "[10] Finale Doshi-Velez and Been Kim. Towards a rigorous science of interpretable machine learning. arXiv preprint arXiv:1702.08608, 2017. 1" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 57, + 73, + 554, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 55, + 91, + 296, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 199 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 199 + ], + "type": "text", + "content": "This work was supported by the Institute of Information & Communications Technology Planning & Evaluation (IITP) grant, funded by the Korea government (MSIT) (No. RS-2019-II190079, Artificial Intelligence Graduate School Program (Korea University), IITP-2025-RS-2024-00436857, ITRC (Information Technology Research Center), and No. RS-2022-II220984, Development of Artificial Intelligence Technology for Personalized Plug-and-Play Explanation and Verification of Explanation)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 316, + 73, + 554, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 139 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 139 + ], + "type": "text", + "content": "[11] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, G Heigold, S Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2020. 3, 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 141, + 554, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 141, + 554, + 194 + ], + "spans": [ + { + "bbox": [ + 317, + 141, + 554, + 194 + ], + "type": "text", + "content": "[12] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dólár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16000-16009, 2022. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 197, + 554, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 197, + 554, + 262 + ], + "spans": [ + { + "bbox": [ + 317, + 197, + 554, + 262 + ], + "type": "text", + "content": "[13] Dan Hendrycks, Steven Basart, Norman Mu, Saurav Kadavath, Frank Wang, Evan Dorundo, Rahul Desai, Tyler Zhu, Samyak Parajuli, Mike Guo, et al. The many faces of robustness: A critical analysis of out-of-distribution generalization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8340-8349, 2021. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 263, + 554, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 263, + 554, + 306 + ], + "spans": [ + { + "bbox": [ + 317, + 263, + 554, + 306 + ], + "type": "text", + "content": "[14] Dan Hendrycks, Kevin Zhao, Steven Basart, Jacob Steinhardt, and Dawn Song. Natural adversarial examples. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15262-15271, 2021. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 308, + 554, + 362 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 308, + 554, + 362 + ], + "spans": [ + { + "bbox": [ + 317, + 308, + 554, + 362 + ], + "type": "text", + "content": "[15] Robin Hesse, Simone Schaub-Meyer, and Stefan Roth. Funnybirds: A synthetic vision dataset for a part-based analysis of explainable ai methods. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3981-3991, 2023. 2, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 365, + 554, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 365, + 554, + 407 + ], + "spans": [ + { + "bbox": [ + 317, + 365, + 554, + 407 + ], + "type": "text", + "content": "[16] Sara Hooker, Dumitru Erhan, Pieter-Jan Kindermans, and Been Kim. A benchmark for interpretability methods in deep neural networks. Advances in Neural Information Processing Systems, 32, 2019. 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 410, + 554, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 410, + 554, + 441 + ], + "spans": [ + { + "bbox": [ + 317, + 410, + 554, + 441 + ], + "type": "text", + "content": "[17] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 443, + 554, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 443, + 554, + 464 + ], + "spans": [ + { + "bbox": [ + 316, + 443, + 554, + 464 + ], + "type": "text", + "content": "[18] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 467, + 554, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 467, + 554, + 510 + ], + "spans": [ + { + "bbox": [ + 317, + 467, + 554, + 510 + ], + "type": "text", + "content": "[19] Simon Kornblith, Mohammad Norouzi, Honglak Lee, and Geoffrey Hinton. Similarity of neural network representations revisited. In International Conference on Machine Learning, pages 3519-3529. PMLR, 2019. 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 512, + 554, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 512, + 554, + 566 + ], + "spans": [ + { + "bbox": [ + 317, + 512, + 554, + 566 + ], + "type": "text", + "content": "[20] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 567, + 554, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 567, + 554, + 622 + ], + "spans": [ + { + "bbox": [ + 317, + 567, + 554, + 622 + ], + "type": "text", + "content": "[21] Ze Liu, Han Hu, Yutong Lin, Zhuliang Yao, Zhenda Xie, Yixuan Wei, Jia Ning, Yue Cao, Zheng Zhang, Li Dong, et al. Swin transformer v2: Scaling up capacity and resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12009-12019, 2022. 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 624, + 554, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 624, + 554, + 656 + ], + "spans": [ + { + "bbox": [ + 317, + 624, + 554, + 656 + ], + "type": "text", + "content": "[22] Clemens Otte. Safe and interpretable machine learning: a methodological review. Computational Intelligence in Intelligent Data Analysis, pages 111-122, 2013. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 658, + 554, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 658, + 554, + 689 + ], + "spans": [ + { + "bbox": [ + 317, + 658, + 554, + 689 + ], + "type": "text", + "content": "[23] Vitali Petsiuk, Abir Das, and Kate Saenko. Rise: Randomized input sampling for explanation of black-box models. British Machine Vision Conference, 2018. 2, 3, 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 692, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 554, + 713 + ], + "type": "text", + "content": "[24] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry," + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "25174" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 117 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 117 + ], + "type": "text", + "content": "Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 1, 6, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 119, + 295, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 119, + 295, + 173 + ], + "spans": [ + { + "bbox": [ + 56, + 119, + 295, + 173 + ], + "type": "text", + "content": "[25] Yao Rong, Tobias Leemann, Vadim Borisov, Gjergji Kasneci, and Enkelejda Kasneci. A consistent and efficient evaluation strategy for attribution methods. In International Conference on Machine Learning, pages 18770-18795. PMLR, 2022. 2, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 175, + 294, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 294, + 218 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 294, + 218 + ], + "type": "text", + "content": "[26] Wojciech Samek and Klaus-Robert Müller. Towards explainable artificial intelligence. Explainable AI: interpreting, explaining and visualizing deep learning, pages 5-22, 2019. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 222, + 294, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 222, + 294, + 266 + ], + "spans": [ + { + "bbox": [ + 56, + 222, + 294, + 266 + ], + "type": "text", + "content": "[27] Karl Schulz, Leon Sixt, Federico Tombari, and Tim Landgraf. Restricting the flow: Information bottlenecks for attribution. In International Conference on Learning Representations, 2020. 1, 2, 3, 4, 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 268, + 294, + 333 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 268, + 294, + 333 + ], + "spans": [ + { + "bbox": [ + 56, + 268, + 294, + 333 + ], + "type": "text", + "content": "[28] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 618-626, 2017. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 335, + 294, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 335, + 294, + 390 + ], + "spans": [ + { + "bbox": [ + 56, + 335, + 294, + 390 + ], + "type": "text", + "content": "[29] Andreas Peter Steiner, Alexander Kolesnikov, Xiaohua Zhai, Ross Wightman, Jakob Uszkoreit, and Lucas Beyer. How to train your vit? data, augmentation, and regularization in vision transformers. Transactions on Machine Learning Research, 2024. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 393, + 294, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 393, + 294, + 426 + ], + "spans": [ + { + "bbox": [ + 56, + 393, + 294, + 426 + ], + "type": "text", + "content": "[30] Mukund Sundararajan, Ankur Taly, and Qiqi Yan. Axiomatic attribution for deep networks. In International Conference on Machine Learning, pages 3319-3328. PMLR, 2017. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 428, + 294, + 461 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 428, + 294, + 461 + ], + "spans": [ + { + "bbox": [ + 56, + 428, + 294, + 461 + ], + "type": "text", + "content": "[31] Naftali Tishby, Fernando C Pereira, and William Bialek. The information bottleneck method. arXiv preprint physics/0004057, 2000. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 463, + 294, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 463, + 294, + 518 + ], + "spans": [ + { + "bbox": [ + 56, + 463, + 294, + 518 + ], + "type": "text", + "content": "[32] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In International Conference on Machine Learning, pages 10347-10357. PMLR, 2021. 4, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 520, + 294, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 520, + 294, + 563 + ], + "spans": [ + { + "bbox": [ + 56, + 520, + 294, + 563 + ], + "type": "text", + "content": "[33] Hugo Touvron, Matthieu Cord, and Hervé Jégou. Deit iii: Revenge of the vit. In Proceedings of the European Conference on Computer Vision, pages 516-533. Springer, 2022. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 566, + 294, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 566, + 294, + 609 + ], + "spans": [ + { + "bbox": [ + 56, + 566, + 294, + 609 + ], + "type": "text", + "content": "[34] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing, 13(4):600-612, 2004. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 612, + 294, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 612, + 294, + 666 + ], + "spans": [ + { + "bbox": [ + 56, + 612, + 294, + 666 + ], + "type": "text", + "content": "[35] Weiyan Xie, Xiao-Hui Li, Caleb Chen Cao, and Nevin L Zhang. Vit-cx: causal explanation of vision transformers. In Proceedings of the Thirty-Second International Joint Conference on Artificial Intelligence, pages 1569–1577, 2023. 1, 2, 6, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 670, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 294, + 713 + ], + "type": "text", + "content": "[36] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11975-11986, 2023. 1" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 128 + ], + "type": "text", + "content": "[37] Yang Zhang, Ashkan Khakzar, Yawei Li, Azade Farshad, Seong Tae Kim, and Nassir Navab. Fine-grained neural network explanation by identifying input features with predictive information. Advances in Neural Information Processing Systems, 34:20040-20051, 2021. 1, 2, 3, 5" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "25175" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Comprehensive Relighting_ Generalizable and Consistent Monocular Human Relighting and Harmonization/743d9653-d63f-4670-9d09-7ac226ec0ec0_content_list.json b/2025/Comprehensive Relighting_ Generalizable and Consistent Monocular Human Relighting and Harmonization/743d9653-d63f-4670-9d09-7ac226ec0ec0_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..b3bc2816919045c1a3f22862c753b48f80d3bb6b --- /dev/null +++ b/2025/Comprehensive Relighting_ Generalizable and Consistent Monocular Human Relighting and Harmonization/743d9653-d63f-4670-9d09-7ac226ec0ec0_content_list.json @@ -0,0 +1,1729 @@ +[ + { + "type": "text", + "text": "Comprehensive Relighting: Generalizable and Consistent Monocular Human Relighting and Harmonization", + "text_level": 1, + "bbox": [ + 94, + 130, + 906, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Junying Wang $^{1\\dagger}$ \nHe Zhang $^{2}$", + "bbox": [ + 114, + 200, + 245, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jingyuan Liu $^{2}$ Xin Sun $^{2}$ \nJimei Yang $^{3}$ Nanxuan Zhao $^{2}$ \nUlrich Neumann $^{1}$", + "bbox": [ + 246, + 202, + 517, + 255 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Krishna Kumar Singh $^{2}$ \nTuanfeng Y. Wang $^{2}$ \nJae Shin Yoon $^{2}$", + "bbox": [ + 526, + 202, + 740, + 255 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zhixin Shu $^{2}$ Simon S. Chen $^{2}$", + "bbox": [ + 750, + 202, + 880, + 237 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1University of Southern California", + "bbox": [ + 210, + 272, + 485, + 292 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Adobe Research $^{3}$ Runway", + "bbox": [ + 524, + 273, + 781, + 292 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8dd337d08a12c9efe08a0823d7c32f13a31e8819a563f5ca3fcf80b740fff7af.jpg", + "image_caption": [ + "Input video frames" + ], + "image_footnote": [], + "bbox": [ + 98, + 316, + 179, + 496 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/27887390fb4995f4a5d613f54ae41568b941439a845cfd0e6f98bc52bbcb2e81.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 183, + 316, + 264, + 407 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a548fd2351b3218e9a979d227ba28925c5fff657d77b9ecbca58bf78db9622ff.jpg", + "image_caption": [ + "Relighting" + ], + "image_footnote": [], + "bbox": [ + 196, + 412, + 264, + 494 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/eabe1ba7bde618f8df382f27afa04b8e1f73ac57bbb8dff2d164220326e6c695.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 266, + 316, + 379, + 407 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f1b99f966c73ed20f0c035d951661e7ee036a1d9dd866023175a387f7ba23b9b.jpg", + "image_caption": [ + "Harmonization" + ], + "image_footnote": [], + "bbox": [ + 266, + 411, + 379, + 494 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/62fa0a00326ffb4ac312b563e5bfd7ba542ef0e24606ad9eb709e6d40073422e.jpg", + "image_caption": [ + "Relighting + Harmonization", + "Figure 1. We introduce Comprehensive Relighting, a generalizable and consistent model for relighting and harmonization, which controls the lighting property from a single image or video of humans with arbitrary body parts. Given target lighting coefficients, e.g., Spherical harmonics (second), background scenes (third), or their combination (fourth), our model performs consistent and harmonized relighting." + ], + "image_footnote": [], + "bbox": [ + 380, + 316, + 503, + 494 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/336493c3ac5e682645dc0b9725d3d227ef1f7e2153b275a0a7101b894aef4fa4.jpg", + "image_caption": [ + "Input video frames" + ], + "image_footnote": [], + "bbox": [ + 504, + 316, + 586, + 494 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6f2690fd754444b3504e9c0341334a7a6e84ac0c5f81aabb7c0265f1dd63b36e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 586, + 316, + 663, + 407 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0953e00d7aa2cdb35d8e64ffbafb0f95e9327ec41a00fdf1964f6efede9f8823.jpg", + "image_caption": [ + "Relighting" + ], + "image_footnote": [], + "bbox": [ + 586, + 407, + 665, + 494 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/54bbd218878ca13188a6574832a00bdb7be77d427051eba1cf4d5386e167238c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 666, + 316, + 779, + 407 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2390ff1bd22080f0244cb8cdfd16801794cf6edb83a8803bd71f3f894a8b9376.jpg", + "image_caption": [ + "Harmonization" + ], + "image_footnote": [], + "bbox": [ + 666, + 407, + 779, + 494 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/141f83c8036a2bbf8df58e3a845db2ea5543a3065bb3ccdebba2b3c5e602b922.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 781, + 316, + 898, + 407 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4fd1ab893d395f4059c550849092f0e6ab7d722e3ef641fb422edcf4aa6d0add.jpg", + "image_caption": [ + "Relighting + Harmonization" + ], + "image_footnote": [], + "bbox": [ + 781, + 407, + 898, + 494 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 569, + 326, + 585 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "This paper introduces Comprehensive Relighting, the first all-in-one approach that can both control and harmonize the lighting from an image or video of humans with arbitrary body parts from any scene. Building such a generalizable model is extremely challenging due to the lack of dataset, restricting existing image-based relighting models to a specific scenario (e.g., face or static human). To address this challenge, we repurpose a pre-trained diffusion model as a general image prior and jointly model the human relighting and background harmonization in the coarse-to-fine framework. To further enhance the temporal coherence of the relighting, we introduce an unsupervised temporal lighting model that learns the lighting cycle consistency from many real-world videos without any ground truth. In inference time, our temporal lighting module is combined with the diffusion models through the spatio-temporal feature blending algorithms without extra training; and we apply a new guided refinement as a post-processing to pre", + "bbox": [ + 86, + 601, + 485, + 875 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "serve the high-frequency details from the input image. In the experiments, Comprehensive Relighting shows a strong generalizability and lighting temporal coherence, outperforming existing image-based human relighting and harmonization methods.", + "bbox": [ + 511, + 571, + 906, + 646 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 659, + 645, + 674 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Light is the key component that determines how a person looks, which is often defined by a specific time and space, where revisiting such a unique moment gives us the opportunity to experience authentic telepresence sensations. In this paper, we introduce a generalizable human relighting model that can control the lighting from an image or video of humans with arbitrary body parts (Fig. 3), which are well-harmonized with a conditioning space (i.e., background image) as shown in Fig. 1.", + "bbox": [ + 509, + 686, + 906, + 823 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "As shown in Fig. 2, existing image-based relighting methods face two main problems, lack of generalizability and controllability. First, they are designed for a specific scenario, e.g., only for face illumination or static humans [33, 61, 68] mainly due to the scarcity of large-scale", + "bbox": [ + 511, + 825, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "This work is partially done during an internship at Adobe Research.", + "bbox": [ + 107, + 886, + 478, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "380", + "bbox": [ + 485, + 944, + 511, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ebd58bb97fc109e3f772bac1266c9459db804e4aaf6c07c7a8462971b961dd67.jpg", + "image_caption": [ + "Figure 2. Comparison of various baseline methods for relighting settings and functionalities." + ], + "image_footnote": [], + "bbox": [ + 94, + 88, + 480, + 215 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "relighting datasets: Precise acquisition of the appearance of a static person under assorted lighting conditions requires specialized setups such as LightStage [22, 33] or expensive graphics simulation [26], which are not scalable, particularly for video contents. Learning from such limited data induces weak generalization of the model to diverse scenes. Second, most relighting algorithms struggle to effectively model multiple light sources. Typically, these algorithms are restricted to a single lighting control from either background image (e.g. high dynamic range lighting environment map [22]) or target lighting parameters (e.g. Spherical harmonics [24]). These problems inhibit the production-level application that requires general use cases.", + "bbox": [ + 88, + 258, + 482, + 455 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To overcome these challenges, we propose an effective approach to achieve all-in-one relighting by utilizing a diffusion model—a general image prior that learns massive visual data with diverse lighting conditions; and repurpose this prior specialized for a human relighting and harmonization in a coarse-to-fine framework: A pretrained latent diffusion model [41] learns from limited datasets of static humans to jointly perform the fine-grained relighting and harmonization from two multi-modal lighting variables: the coarse shading estimate and conditioning background scene. The coarse shading estimate can be \"computed\" from conditioning lighting parameters (i.e., Spherical harmonics) without a neural network, and therefore, it is generalizable. The diffusion model is required to learn only the residual portion (e.g., fine self-occluded shadow), which is more generalizable than direct relighting (e.g., [68]). The pre-trained image prior helps with understanding the properties of the general background scenes, enabling the generalization of the background harmonization.", + "bbox": [ + 89, + 455, + 482, + 743 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our coarse-to-fine framework, however, still introduces problems: the diffusion model that learns to generate an image without temporal context produces significant temporal artifacts (e.g., sudden changes of lighting distribution even for consecutive frames). We address this problem by introducing an unsupervised temporal lighting model that learns from many real videos without any ground-truth data to en", + "bbox": [ + 89, + 744, + 482, + 849 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/5d7ffff5401d1b0ae3c69cf2e14429bf4d6713475641b686688a1500756b4919.jpg", + "image_caption": [ + "Figure 3. Our model generalizes to various body parts (portrait, half-body, full-body, multiperson) for relighting and harmonization, with lighting control variables shown in the insets." + ], + "image_footnote": [], + "bbox": [ + 517, + 88, + 903, + 250 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "force the temporal lighting consistency over frames. This temporal module learns the unsupervised cycle consistency between the relit and many real videos to predict the future lighting distribution from the past, which can be directly combined with our coarse-to-fine relighting components without extra training.", + "bbox": [ + 511, + 308, + 903, + 398 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In inference time, we further push the temporal coherence of the relighting and harmonization by formulating a recurrent prediction pipeline. The generation at the current time step is conditioned onto the temporal module for the one at the next time. The features from the lighting and temporal control modules are spatially and temporally blended to enforce temporal coherence while improving the structure of the lighting distribution. Finally, our guided refinement module enhances the quality of the generated image in a way that preserves the original high-frequency details of the input image.", + "bbox": [ + 511, + 400, + 905, + 566 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In the experiments, Comprehensive Relighting demonstrates high-quality relighting and background harmonization results with strong temporal coherence across the lighting, background, and pose changes. It also shows strong generalizability to any unconstrained scenes regardless of body parts, views, and poses, outperforming existing relighting and background harmonization methods.", + "bbox": [ + 511, + 566, + 905, + 672 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions include: (1) To the best of our knowledge, the first approach for joint modeling of relighting and background harmonization; (2) a novel coarse-to-fine framework that enables comprehensive generalization by only learning from limited synthetic and lab-controlled data; (3) unsupervised temporal modeling with lighting cycle consistency from many unconstrained real videos; (4) an effective inference algorithm with spatio-temporal feature blending and guided refinement.", + "bbox": [ + 511, + 674, + 905, + 810 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 827, + 653, + 842 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Image-based Human Relighting While high-quality relighting often requires resource-intensive setups, recent efforts in mobile device relighting aim to address single", + "bbox": [ + 511, + 854, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "To the best of our knowledge, there exists no public ground truth data for video relighting of dynamic humans, and therefore, fine-tuning an existing video diffusion model (e.g., [17]) is not a readily available option.", + "bbox": [ + 89, + 862, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "381", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "image scenarios. Total Relighting [33] achieves photorealistic effects by leveraging detailed normal maps and albedo as priors. However, it relies on an HDR environment map that accurately matches the scene's real-world lighting, which is not always readily available, particularly in flexible and personalized relighting scenarios. [37, 46] estimate and adjust the Spherical harmonics parameters from images, and others use one-light-at-a-time (OLAT) captures to generate relighting data [49] or estimate reflectance fields [46].", + "bbox": [ + 88, + 90, + 480, + 227 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recent works have focused on diverse relighting scenarios including portrait [10, 19, 20, 25, 30, 33, 46, 58, 67, 68], full-body [9, 22, 26, 47], and object relighting [1, 55, 56]. Most of these works rely on decomposing an image into its intrinsic components, i.e., albedo, normal, and lighting, and therefore, the accuracy of this decomposition directly impacts the quality of the final relighting. While some works [22, 66, 69] use single-image geometry reconstruction for traditional and neural rendering, reconstruction errors are often propagated to relighting results. Tajima et al. [47] tackled domain adaptation with a two-step relighting framework, yet noticeable texture distortion remains due to limited model generalizability. Zhang et al. [59] trained a 2D latent-diffusion model, allowing users to manipulate and construct face NeRFs in a zero-shot learning framework without the need for explicit 3D data. DiFaRei [36] utilizes DDIM (Denoising Diffusion Implicit Models [44]) for high-fidelity face relighting. While promising, their methods are constrained by a focus on either specific body parts, limiting their applicability for generalizable.", + "bbox": [ + 93, + 229, + 485, + 531 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Background Harmonization Background harmonization seeks to harmonize color, contrast, and style discrepancies between the foreground and background, ensuring composite images appear natural and cohesive. Many existing background harmonization methods [4, 7, 8, 13-15, 23, 51, 52, 70] formulate this problem as image-to-image translation work where a neural network translates an unharmonized foreground image to the harmonized one in the context of the conditioning background image. Recently, Relightful Harmonization [38] introduced methods that harmonize both image style and lighting to match the background scene, and IC-Light [63] further enables flexible illumination control via image diffusion models, guided by text descriptions or background images. While promising, these methods lack explicit lighting control, restricting general applicability and consistent video relighting. Additionally, some ([38]) are limited to specific body parts (e.g., portraits), with constrained support for full-body and multiperson scenarios.", + "bbox": [ + 91, + 534, + 482, + 821 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Video Relighting Neural Radiance Fields (NeRF) [32] based methods enable novel view synthesis under varying lighting conditions in videos [43, 50, 53, 69]. Zhang et al. [61] achieve portrait video relighting under dynamic illuminations, while Choi et al. [6] ensure temporally consis", + "bbox": [ + 89, + 824, + 482, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "tent relit videos. 3D-PVR [3] present a 3D-aware, real-time method to relight videos of talking faces. Relighting4D [5] decomposes the time-varied human body as a set of neural fields of normal, occlusion, diffuse, and specular maps. However, their rendering quality are highly reliant on geometry accuracy. Some works [36] adopt a temporal modeling scheme real-time neural video portrait relighting. STNeRF [60] controls dynamic scenes using a neural layered radiance representation that maintains spatial and temporal coherence. Li et al. [28] employed multi-view reconstruction for free-viewpoint relighting videos under general illumination. Richardt et al. [39] use RGBZ video cameras for video effects, including relighting, but rely on multi-view reconstruction from specialized hardware, making the process highly complex and costly.", + "bbox": [ + 511, + 90, + 903, + 319 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 511, + 332, + 604, + 348 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We develop a generalizable and consistent human relighting and harmonization framework using a diffusion model. Fig. 4 illustrates the overview of our framework. Given an input image of humans and control lighting variables, including coarse shading and background image, a diffusion model learns to generate a fine-grained image of the humans under the target lighting, also harmonized with the background (Sec. 3.2). An external temporal lighting module, trained on real videos using unsupervised temporal cycle consistency, is integrated into the diffusion model to enhance temporal lighting coherence (Sec. 3.3). In inference time, we blend the features between the lighting and temporal control modules over time to ensure the relighting results are accurate and temporally coherent, and we apply a guided refinement to prevent the loss of high-frequency details during the denoising process (Sec. 3.4).", + "bbox": [ + 511, + 358, + 906, + 601 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Background: Image-based Relighting", + "text_level": 1, + "bbox": [ + 511, + 611, + 836, + 628 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Image-based relighting function can be compactly modeled with a small number of approximated basis of Spherical harmonics (SH) [9, 26, 42] which describes only essential features of illumination on the surface of a 3D sphere based on the following formulation:", + "bbox": [ + 511, + 633, + 905, + 709 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nI (\\mathbf {x}) = \\rho (\\mathbf {x}) \\cdot \\sum_ {l = 0} ^ {k} \\sum_ {m = - l} ^ {l} \\phi Y _ {l m} (\\mathbf {n} (\\mathbf {x})) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 568, + 720, + 906, + 763 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\phi \\in \\mathbb{R}^{(k + 1)^2}$ denotes the spherical harmonics (SH) coefficients; $Y_{lm}(\\mathbf{n})$ are the SH basis functions evaluated at the surface normal $\\mathbf{n}$ ; indices $l$ and $m$ represent the band and order within each band, respectively; and $k$ is the maximum order of spherical harmonics used. While SH is computationally efficient and highly generalizable, it only captures low-frequency lighting, limiting its ability to model fine-grained, high-frequency illumination details. Addition-", + "bbox": [ + 511, + 777, + 906, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "382", + "bbox": [ + 485, + 944, + 509, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/fe0e59f14ac5883b7b71a881062be5b17f97a96e7a2f25f3bfcc420174600209.jpg", + "image_caption": [ + "(a) Coarse-to-fine relighting and harmonization" + ], + "image_footnote": [], + "bbox": [ + 99, + 90, + 436, + 271 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/2638a0955625b65dd185371f2d6fbc9c937a91808ec41174a9f3c1006d95c397.jpg", + "image_caption": [ + "(b) Temporal modeling with cycle consistency", + "(c) Inference: Spatio-temporal feature blending", + "Figure 4. System overview. (a) Given an input image of humans with coarse lighting and background image, our diffusion model generates the relit images harmonized with background scenes (Sec. 3.2). (b) The external temporal modules learn the temporal cycle consistency from many real-world videos to construct temporal lighting features (Sec. 3.3). (c) In inference time, we blend the features from lighting and temporal modules spatially and temporally to enable coherent and generalizable human relighting (Sec. 3.4)." + ], + "image_footnote": [], + "bbox": [ + 439, + 93, + 782, + 271 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/22cf249a94d54a2f4bec6ba87273d378bbe3c1046817b7ee8231579a47af2826.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 785, + 93, + 895, + 268 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ally, SH-based global illumination neglects the context provided by background images, often resulting in relit images appearing unnatural when composited with different backgrounds. To overcome these limitations, we propose a coarse-to-fine human relighting and harmonization framework that leverages the strong image prior available from a pre-trained text-to-image diffusion model.", + "bbox": [ + 89, + 364, + 483, + 470 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Coarse-to-Fine Relighting and Harmonization", + "text_level": 1, + "bbox": [ + 89, + 479, + 480, + 496 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We generate a fine-grained relit image of a person conditioned on a coarse lighting representation:", + "bbox": [ + 89, + 502, + 482, + 532 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {E} (\\mathbf {I}; \\mathbf {S} _ {\\phi}) = z, \\quad \\mathcal {D} (z) = \\mathbf {I} _ {\\phi} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 176, + 542, + 480, + 559 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{E}$ is an encoder that generates the latent lighting features, $z$ as a function of the input image $\\mathbf{I} \\in \\mathbb{R}^{w \\times h \\times 3}$ and a small number of global lighting parameters $\\phi \\in \\mathbb{R}^n$ (i.e., Spherical harmonics where $n = 25$ ); and $\\mathcal{D}$ is the decoder that generates fine-grained relit image $\\mathbf{I}_{\\phi} \\in \\mathbb{R}^{w \\times h \\times 3}$ from $z$ . $\\mathbf{S}_{\\phi} \\in \\mathbb{R}^{w \\times h}$ is the pixel-aligned coarse shading map converted from the coarse lighting parameters $\\phi$ as shown in Fig. 4-(a). One approach to obtaining $\\mathbf{S}_{\\phi}$ is directly computing it as demonstrated in Eq. 1 where we can compute lighting intensity along with the detected surface normal map from $\\mathbf{I}$ given spherical harmonics coefficients $\\phi$ . Another way, is to use a neural network to convert the surface normal to $\\mathbf{S}_{\\phi}$ as a condition of spherical harmonics coefficients $\\phi$ , i.e., $\\mathbf{S}_{\\phi} \\gets f(\\mathbf{N}; \\phi)$ where $f$ is the neural shading function which maps the surface normal $\\mathbf{N}$ and $\\phi$ to the coarse shading. While both approaches are highly generalizable, our experiments indicate that the latter method achieves improved accuracy and better noise correction. Please refer to the Supple. documents for more details about the coarse shading estimation.", + "bbox": [ + 89, + 568, + 482, + 868 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To jointly model the relighting and background harmonization, we further condition a target background image", + "bbox": [ + 89, + 869, + 482, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$\\mathbf{B} \\in \\mathbb{R}^{w \\times h \\times 3}$ as the additional lighting sources:", + "bbox": [ + 513, + 364, + 831, + 380 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {E} (\\mathbf {I}; \\{\\mathbf {S} _ {\\phi}, \\mathbf {B} \\}) = \\boldsymbol {z}, \\quad \\mathcal {D} (\\boldsymbol {z}) = \\mathbf {I} _ {\\phi} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 580, + 392, + 903, + 409 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where the lighting encoder $\\mathcal{E}$ learns to capture the intensity and direction of light from $\\mathbf{S}_{\\phi}$ while capturing ambient environment lighting and color distribution from $\\mathbf{B}$ . This enables $\\mathcal{D}$ to achieve complete relighting in scenarios involving new target lighting, background, or both. In the subsection, we enable $(\\mathcal{D} \\circ \\mathcal{E})$ using a latent diffusion model.", + "bbox": [ + 511, + 419, + 905, + 510 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.1. Fine-grained Relighting Diffusion Model", + "text_level": 1, + "bbox": [ + 511, + 517, + 843, + 532 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We enable the fine-grained image relighting using a diffusion model as illustrated in Fig. 4-(a). Our encoder $\\mathcal{E}$ in Eq. 3, in practice, is formulated as the composition of two encoders:", + "bbox": [ + 511, + 536, + 906, + 595 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {E} \\rightarrow \\mathcal {E} _ {\\mathrm {b}} (\\mathbf {I}; \\mathcal {E} _ {1} (\\{\\mathbf {S} _ {\\phi}, \\mathbf {B} \\})) = \\boldsymbol {z}, \\quad \\mathcal {D} (\\boldsymbol {z}) = \\mathbf {I} _ {\\phi} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 547, + 607, + 903, + 625 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{E}_1$ encodes lighting control variables, $\\{\\mathbf{S}_{\\phi}, \\mathbf{B}\\}$ , and $\\mathcal{E}_b$ encodes the conditional variable $\\mathbf{I}$ whose visual properties, e.g., semantics and identity, should be preserved in the output along with the controls from $\\mathcal{E}_1$ . For $\\mathcal{E}_b$ , we use the base latent diffusion model [2, 40] pre-trained for text-to-image generation task, and for $\\mathcal{E}_1$ , we use ControlNet [62] (termed as Light ControlNet in Fig. 4-(a)). While noise, texts, and timestep variables are also conditioned on $\\mathcal{E}_b$ to fit the modality of the latent diffusion model, they are not described in the equations and figures for conciseness. To impose the foreground attention on the lighting control, a portrait mask is also conditioned to $\\mathcal{E}_l$ :", + "bbox": [ + 511, + 633, + 906, + 816 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {E} _ {\\mathrm {b}} (\\mathbf {I}; \\mathcal {E} _ {1} (\\{\\mathbf {S} _ {\\phi}, \\mathbf {B} \\}; \\mathbf {M})) = \\boldsymbol {z}, \\quad \\mathcal {D} (\\boldsymbol {z}) = \\mathbf {I} _ {\\phi} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 553, + 827, + 903, + 844 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{M} \\in \\{0,1\\}^{w \\times h}$ is the foreground binary mask. In training time, the diffusion model learns to directly change the lighting distribution of the input images (without inverse", + "bbox": [ + 511, + 854, + 906, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "383", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "rendering techniques) as the condition of coarse lighting estimate by minimizing the latent distance of the noisy relit image with the clean one from the ground truth in the overall forward and background denoising steps. Following the findings from an existing harmonization work [38], we use the composite image between the input image and conditioning background image as $\\mathbf{I}$ . We randomly drop background $\\mathbf{B}$ or randomly set the coarse shading $\\mathbf{S}_{\\phi}$ as binary mask such that the diffusion model learns the control the lighting from $\\mathbf{B}$ , $\\mathbf{S}_{\\phi}$ , or both $\\{\\mathbf{B}, \\mathbf{S}_{\\phi}\\}$ .", + "bbox": [ + 89, + 90, + 483, + 243 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Unsupervised Add-on Temporal Modeling", + "text_level": 1, + "bbox": [ + 89, + 251, + 452, + 268 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our coarse-to-fine relighting framework $(\\mathcal{D} \\circ \\mathcal{E}_{\\mathrm{b}} \\circ \\mathcal{E}_{\\mathrm{l}})$ is trained only on individual images, inherently missing temporal context, e.g., how a point on a human's surface will radiate from a specific viewpoint under the continuous pose, view, and illumination changes, leading to temporal artifacts such as flickering. We model such temporal context by designing an external add-on temporal lighting module $\\mathcal{E}_{m}$ that can be combined, in inference time, with the relighting framework without extra training:", + "bbox": [ + 89, + 272, + 483, + 409 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {D} \\circ \\mathcal {E} _ {\\mathrm {b}} \\circ \\mathcal {E} _ {\\mathrm {l}} \\rightarrow \\mathcal {D} \\circ \\mathcal {E} _ {\\mathrm {b}} \\circ \\left(\\mathcal {E} _ {\\mathrm {l}} \\times \\mathcal {E} _ {\\mathrm {m}}\\right) \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 153, + 420, + 482, + 436 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To enable this, our temporal module is designed to regress the relit image from the previous time instance, i.e., $\\mathbf{I}_{\\phi}^{t - 1}$ , to the latent lighting distribution whose space is shared with our base relighting models. Therefore, our decoder can generate temporally coherent relit images in the current time in an auto-regressive way:", + "bbox": [ + 89, + 448, + 483, + 539 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {D} \\left(\\mathcal {E} _ {\\mathrm {b}} \\left(\\mathbf {I} ^ {t}; \\mathcal {E} _ {\\mathrm {m}} \\left(\\mathbf {I} _ {\\phi} ^ {t - 1}\\right)\\right)\\right) = \\mathbf {I} _ {t - 1} ^ {t}. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 169, + 547, + 482, + 566 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "However, training $\\mathcal{E}_{\\mathrm{m}}$ with a conventional L2 loss is not possible since there exists no ground-truth video relighting data for dynamic humans. Therefore, we propose to learn $\\mathcal{E}_{\\mathrm{m}}$ in an unsupervised way using many real videos with lighting cycle consistency as follows:", + "bbox": [ + 89, + 577, + 483, + 654 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {D} ^ {*} \\left(\\mathcal {E} _ {\\mathrm {b}} ^ {*} \\left(\\mathbf {I} ^ {t}; \\mathcal {E} _ {1} ^ {*} \\left(\\left\\{\\mathbf {S} _ {\\phi} ^ {t}, \\mathbf {B} ^ {t} \\right\\}; \\mathbf {M} ^ {t}\\right)\\right)\\right) = \\mathbf {I} _ {\\phi} ^ {t}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 155, + 662, + 482, + 681 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {D} ^ {*} \\left(\\mathcal {E} _ {\\mathrm {b}} ^ {*} \\left(\\mathbf {I} _ {\\phi} ^ {t}; \\mathcal {E} _ {\\mathrm {m}} \\left(\\mathbf {I} ^ {t - 1}, \\mathbf {M} ^ {t - 1}\\right)\\right)\\right) = \\tilde {\\mathbf {I}} _ {t - 1} ^ {t} \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 158, + 684, + 480, + 702 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\therefore \\quad \\mathbf {I} ^ {t} = \\tilde {\\mathbf {I}} _ {t - 1} ^ {t}.\n$$\n", + "text_format": "latex", + "bbox": [ + 236, + 705, + 336, + 723 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We make the hypothesis: a video sequence inherently contains temporal lighting properties whose flow can be implicitly modeled by learning to predict the lighting distribution of the future frame based on the hint from the previous frame. This involves forward and backward processes, as a cycle-training. Eq. 8 represents the forward image relighting, i.e., $\\mathbf{I}^t\\rightarrow \\mathbf{I}_\\phi^t$ , where our relighting diffusion model with lighting ControlNet $\\mathcal{E}_1$ generates the relit image at time $t$ as a condition of a novel lighting condition (where we pick random Spherical harmonics). Eq. 9 reverts the relighting, i.e., $\\tilde{\\mathbf{I}}_{t - 1}^{t}\\gets \\mathbf{I}_{\\phi}^{t}$ to the original input image conditioned by", + "bbox": [ + 89, + 734, + 483, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "the original frame in the previous time step through our temporal lighting module where the mask $\\mathbf{M}$ is used for foreground awareness. Finally, $\\mathcal{E}_m$ learns the lighting cycle consistency in the diffusion process by minimizing the latent distance between the reverted and the original image. We freeze the learnable weights for the functions with $*$ during training, and $\\mathcal{E}_{\\mathrm{m}}$ is enabled with another ControlNet [62] termed as Motion ControlNet in Fig. 4-(b).", + "bbox": [ + 511, + 90, + 906, + 212 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Inference with Spatio-Temporal Blending and Guided Refinement", + "text_level": 1, + "bbox": [ + 511, + 219, + 906, + 250 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given a video or image, we perform comprehensive relighting with a spatio-temporal feature blending framework. For $t = 0$ , we generate the relit image without our temporal lighting module $\\mathcal{E}_{\\mathrm{m}}$ . For $t > 0$ (note that even for the static image, $t > 0$ is possible since the lighting is time-variant), the generated relit images in the previous time step are conditioned on $\\mathcal{E}_{\\mathrm{m}}$ , and therefore, the relighting is controlled by dual control modules, i.e., $\\mathcal{E}_{\\mathrm{l}}$ and $\\mathcal{E}_{\\mathrm{m}}$ by blending their features with a (0.85:0.15) ratio as described as spatial blending block in Fig. 4-(c). The blended lighting features are recurrently combined with the one from the previous time step through the temporal blending block. For this, we adopt optical flow from the original input video as a temporal prior to spatially align the features from consecutive frames similar to [27]; and we blend the aligned temporal features with a (0.5:0.5) ratio as in Fig. 4-(c) to improve the temporal consistency.", + "bbox": [ + 511, + 256, + 906, + 513 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the denoising process of the latent diffusion model, the generation often suffers from the loss of high-frequency details. Inspired by existing image restoration techniques [45, 65], we address this problem by applying a guided refinement to each generated image. We cast the problem of guided refinement as a guided residual prediction:", + "bbox": [ + 511, + 513, + 906, + 604 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {I} _ {\\phi} ^ {\\text {r e f i n e}} = \\mathbf {I} + g \\left(\\mathbf {I} _ {\\phi}, \\mathbf {I}; \\mathbf {M}\\right), \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 604, + 609, + 903, + 627 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $g$ is the function that predicts the guided lighting residual. This residual learns to map the lighting distribution from $\\mathbf{I}$ to $\\mathbf{I}_{\\phi}$ . Here, $\\mathbf{I}_{\\phi}^{\\mathrm{refine}}$ can effectively preserve the high-frequency details of the input image $\\mathbf{I}$ due to the nature of residual learning [57, 65] that is designed to preserve the visual properties from the observation space, i.e., $\\mathbf{I}$ . We enable $g$ with a residual network [16] by learning from our relighting data with general losses for low-level vision processing, i.e., L2 and VGG [11].", + "bbox": [ + 511, + 632, + 906, + 768 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 776, + 645, + 792 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Dataset. To train our coarse-to-fine relighting model, we use 100K ground-truth relighting samples, including 50K synthetic human renderings and 50K OLAT-captured images from LightStage, with random cropping augmentation. Ground-truth albedo, images, background, masks, and lighting coefficients are precomputed. Our dataset is categorized by gender, skin tone, and body part (full-body and", + "bbox": [ + 511, + 795, + 906, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "384", + "bbox": [ + 485, + 945, + 511, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/b8e8e3b695fa2e897d1ffd00b27bd7cb75d0c32a2de4aab0bdbb6b3993e82470.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodScenario 1Scenario 2Scenario 3
DPR [68]18.62/0.86/0.10332.00 /0.94/0.03018.14/0.89/0.08935.29 /0.98/0.03221.20/0.89/0.07237.11/0.95/0.038
RHW [47]19.78/0.87/0.11330.74/0.95/0.02720.12/0.88/0.07836.64 /0.98/0.02823.33/0.90/0.06037.53 /0.98/0.033
GFR [22]25.59/0.91/0.08930.33/0.95/0.03322.76/0.91/0.07232.36/0.98/0.03625.49/0.93/0.05035.89/0.98/0.028
LPBR [38]18.19/0.86/0.09031.62/0.91/0.03519.96/0.88/0.08427.94/0.94/0.04121.42/0.88/0.05333.39/0.95/0.038
Ours25.95/0.95/0.06633.58/0.96/0.02423.99/0.93/0.04835.18/0.99 / 0.03126.61/0.94/0.03338.32/0.98/0.026
", + "bbox": [ + 94, + 88, + 906, + 176 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Comparison with existing image-based human relighting methods on synthetic videos for fidelity and temporal consistency. Each column shows image fidelity (left) and video temporal consistency (right). Metrics are PSNR↑/SSIM↑/LPIPS↓ for accuracy, and tPSNR↑/tSSIM↑/tLPIPS↓ for temporal consistency. While green is used for the best values, yellow highlights the second-best values.", + "bbox": [ + 89, + 179, + 906, + 223 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/12983553ec971990dd47a5b51513ba04934a5b976d430d60799d97d1e8e9dddf.jpg", + "image_caption": [ + "Figure 5. Qualitative comparison of synthetic video frames (corresponding to Tab. 1). From left to right: composite input with target lighting parameters (inset), our relit result, baseline methods, and normalized L2 $\\downarrow$ photometric error map (inset)." + ], + "image_footnote": [], + "bbox": [ + 94, + 233, + 903, + 521 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "multi-body), with each subject captured from 32 viewpoints under varying lighting conditions. For a detailed dataset breakdown, refer to the Supple. To train our temporal control module using real data, we used 50K frames of videos from existing works [21, 29, 34, 54] and customized videos, where, for cyclic relighting, we randomly sample spherical harmonics lighting parameters from ground-truth data. For validation on static humans with dynamic lighting, we generated a new testing set (i.e., the ground-truth relit images and associated lighting parameters) using an internal simulation and LightStage. This set includes the mixture of half-body, full-body, and multi-person scenarios, with each scenario comprising 100 frames. For the validation on dynamic humans (i.e., video), we newly create synthetic video sequences for three scenarios: Scenario 1 involves static lighting (environment map) across frames with a moving human; Scenario 2 features dynamic lighting (rotated environment map) with a static human; Scenario 3 combines dynamic lighting (rotated environment map) with a moving human. Please refer to the Supple. and demo video for more details about the testing results.", + "bbox": [ + 88, + 566, + 482, + 883 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Metrics. To measure the relighting quality, we use L1 dis", + "bbox": [ + 89, + 885, + 482, + 901 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/85944e6a87caa418279f9ec9e779d08aa4fcb45ddddfadcc5d1b991850428972.jpg", + "image_caption": [ + "Figure 6. Comparison with LPBR [38] on DeepFashion [31] real images for background harmonization testing. The first row shows the relit output, and the second shows the magnified results." + ], + "image_footnote": [], + "bbox": [ + 519, + 565, + 645, + 738 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f59d0190d8986b659b4caa768b947d68ac252014b6c64d256e25e8097dce0972.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 645, + 565, + 797, + 738 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/6cbc539142dee20d8d1392b23984a030b5768a668e205251d755ba33b270dcb1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 774, + 565, + 900, + 738 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "tance, reconstruction fidelity (PSNR [18]), local structure similarity (SSIM [18]), and the latent perceptual distance (LPIPS [64]), between the relighted and ground-truth. For measuring the temporal coherence of the relighting results, we follow the same logics as TokenFlow [12]. This involves warping the relit image from the previous time step to the next using optical flow [48], and then comparing this to the", + "bbox": [ + 511, + 795, + 906, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "385", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/06e95269460463cacb9206d2bda8252d73c29b08b029e2013da7c088437f7b9b.jpg", + "image_caption": [ + "Figure 7. Comparison with RHW [47] on Pexels [35] real images. The lighting control variables are shown as insets. While RHW produces reasonable relighting for full-body images, its quality degrades on half-body and multi-person cases." + ], + "image_footnote": [], + "bbox": [ + 96, + 87, + 903, + 393 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b4d21c460ee17f7fcbabd8828264f16dbe0336eb88963c9e1f6c40a8b705974d.jpg", + "image_caption": [ + "Figure 8. Comparison with GFR [22] on Pexels [35] real images. The lighting control variables are shown as insets. Limited generalizability of GFR results in reduced output quality for half-body and multi-person cases." + ], + "image_footnote": [], + "bbox": [ + 94, + 435, + 906, + 662 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "current frame using the aforementioned metrics, termed $tL_{1}$ error, tPSNR, tSSIM, and tLPIPS.", + "bbox": [ + 89, + 705, + 480, + 736 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Baseline. We compare our method with existing monocular human relighting works where we chose the baselines that are applicable to general scenes (e.g., any part-specific information such as a 3D face model is not a requirement): DPR [67], RHW [47], and GFR [22]. For the harmonization baseline, we compare our model with LPBR [38], a diffusion-based light-aware harmonization method. For relighting, aside from GFR, baselines use Spherical harmonics for lighting control without modeling background illumination. DPR and RHW are evaluated using their released", + "bbox": [ + 89, + 750, + 483, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "pre-trained models. Since GFR lacks available code, we re-implemented it using our dataset, replacing HDR lighting with Spherical harmonics and background modeling to enable background harmonization in our experiments. Harmonization is directly compared with LPBR by replacing the background.", + "bbox": [ + 511, + 705, + 906, + 797 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results. For testing static humans under dynamic lighting, we present the quantitative comparison in Tab. 2. We show average numerical evaluations on our synthetic testing dataset, categorized by portrait, full-body, and multiperson. Further validation by gender, and skin color is detailed in the Supple. Methods such as RHW and DPR have", + "bbox": [ + 511, + 809, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "386", + "bbox": [ + 486, + 945, + 511, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/da160d01b8e9140478137ffc2e642bddd90ffc24159668f3ca13c8779c1a1266.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodL1↓PSNR↑SSIM↑
Ours-diffusion0.0583717.0990.833
Ours end-to-end0.0143226.4180.918
Ours-background0.0123927.3460.945
Ours0.0103528.4190.948
Ours+refine0.0101228.7780.949
", + "bbox": [ + 102, + 88, + 475, + 180 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "limited generalizability for both full-body and portrait relighting. They are difficult to extend to other scenarios and tend to show reduced performance when tested in different settings. While GFR performs well in our evaluation, it struggles with significant domain gaps, leading to noticeable quality degradation on real data, including distortions and color shifts (Fig. 8). In contrast, our model exhibits strong generalizability across validation sets and real-world tests (Figs. 7, 8). Compared to the state-of-the-art background harmonization method (LPBR), our method shows the strong generalization to different body parts, and notably, LPBR often includes significant distortion when applied to the full-body images and it does not support the lighting control function as shown in Fig. 6.", + "bbox": [ + 89, + 208, + 483, + 420 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/1944142a1ea1c0bd5fb0a97bde433873e25ea2ba6064fcb18d059d08dec98a9f.jpg", + "table_caption": [ + "Table 3. Ablation study on coarse-to-fine relighting models." + ], + "table_footnote": [], + "table_body": "
CategoryDPR [68]RHW [47]GFR [22]Ours
Portrait17.74 / 0.8715.75 / 0.8217.71 / 0.8623.04 / 0.90
Full-body27.62 / 0.9627.73 / 0.9529.51 / 0.9530.81 / 0.97
Multi-person25.70 / 0.9525.69 / 0.9529.35 / 0.9731.49 / 0.96
", + "bbox": [ + 94, + 429, + 482, + 482 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "For video testing, in Tab. 1, we evaluate our model on scenarios 1, 2, and 3 for both fidelity and temporal consistency. Other approaches face challenges in achieving both relit fidelity and temporal consistency at the same time as also shown in Fig. 5. In contrast, our temporal module ensures our comprehensive relighting model to produce videos with strong temporal consistency. For more results on relighting results and comparisons including user study, please refer to the Supple.", + "bbox": [ + 89, + 523, + 482, + 657 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ablation Study. We conduct two ablation studies on our coarse-to-fine model using static human data and our temporal modules using dynamic human data. As shown in Tab. 3, we perform the ablation study on our coarse-to-fine approach: 1) Ours-diffusion: Relighting only with pixel-aligned coarse shading without diffusion model, i.e., $\\mathbf{S}_{\\phi} \\times \\mathbf{I}$ . 2) Ours end-to-end: Relighting in an end-to-end manner by applying target lighting parameters $\\phi$ as a condition, trained with a diffusion model, without the coarse stage. 3) Ours-background: Relighting without the control of background B. 4) Ours+refine (full): Applying our guided refinement to the relit image.", + "bbox": [ + 89, + 659, + 482, + 839 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Tab. 3 shows the summary of the ablation study: Directly apply the coarse stage without a diffusion model introduces significant errors in the final relit result due to the detection noises. Instead of applying target lighting and end-to", + "bbox": [ + 89, + 839, + 482, + 901 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/7d9f4c1d26ff23a414bd3aa418fa937e671f2b9fc7090df36c1b5f724a067c87.jpg", + "table_caption": [ + "Table 2. Comparison on our synthetic static testing data sorted by body-part. We compute average PSNR↑ / SSIM↑." + ], + "table_footnote": [], + "table_body": "
MethodtL1↓tPSNR↑tSSIM↑
Ours6.55231.0280.956
Ours+temporal5.63832.2660.957
Ours+temporal+blend4.01933.5880.957
", + "bbox": [ + 524, + 88, + 898, + 148 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. Ablation study on our temporal modules evaluated on synthetic sequences: tL1 error $(\\times 10^{-3})$", + "bbox": [ + 511, + 152, + 903, + 180 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "end training with a diffusion model, our coarse-to-fine approach shows better performance, indicating that our coarse stage serves as a strong control prior. This control prior is both neat and effective for extending our model to diverse identities, various body parts. Based on the comparison with \"Ours-background\" we notice that encoding information from the background image aids in enhancing natural illumination during background harmonization. Lastly, leveraging a guided refinement enables the preservation of high-frequency information alongside robust generation capabilities.", + "bbox": [ + 511, + 190, + 906, + 354 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "For our temporal module, we study three ablation studies: 1) Ours: We eliminate all temporal consistency components, which is a single-frame-based generation method. 2) Ours+temporal: We only apply temporal module $\\mathcal{E}_m$ without recurrent feature blending during the test-time. 3) Ours+temporal+blend: We perform our video relighting with temporal module $\\mathcal{E}_m$ and recurrent feature blending.", + "bbox": [ + 511, + 357, + 905, + 463 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Tab. 4 summarizes the performance of each of our temporal modules. The temporal lighting module in \"Ours+temporal\" primarily enforces temporal coherence by imposing a temporal constraint on the lighting control between the current and previous frames during testing. Additionally, the recurrent blending feature further enhances temporal consistency by blending the lighting control feature between previous and current frames, thereby reinforcing the temporal context.", + "bbox": [ + 511, + 464, + 905, + 599 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitation. Our relighting diffusion model requires heavy computational time. Significant noise on the detection (e.g., mask and surface normal) affects the temporal coherence.", + "bbox": [ + 511, + 599, + 903, + 646 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 662, + 633, + 679 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We introduce a method for Comprehensive Relighting that is generalizable and consistent for monocular human relighting and harmonization. We address a core dataset challenge by utilizing a large and general image prior from a pre-trained diffusion model; and repurposing the model specialized for temporally consistent image relighting. For coherent control of the lighting, we introduce a coarse-to-fine relighting framework; and combine it with an external temporal lighting module that learns many real videos. Our guided refinement network enhances the visual to preserve the fine details of an original image. In the experiments, our method outperforms other image-based relighting and harmonization models in terms of quality and temporal coherence.", + "bbox": [ + 511, + 688, + 906, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "387", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Acknowledgement", + "text_level": 1, + "bbox": [ + 91, + 90, + 269, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We sincerely thank Mengwei Ren for the insightful discussions regarding the framework design, and we are grateful to Jianming Zhang for kindly providing the human normal map estimator.", + "bbox": [ + 89, + 114, + 483, + 176 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 188, + 187, + 204 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Sai Bi, Zexiang Xu, Kalyan Sunkavalli, David Kriegman, and Ravi Ramamoorthi. Deep 3d capture: Geometry and reflectance from sparse multi-view images. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5960-5969, 2020. 3", + "[2] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18392-18402, 2023. 4", + "[3] Ziqi Cai, Kaiwen Jiang, Shu-Yu Chen, Yu-Kun Lai, Hongbo Fu, Boxin Shi, and Lin Gao. Real-time 3d-aware portrait video relighting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6221-6231, 2024. 3", + "[4] Jianqi Chen, Yilan Zhang, Zhengxia Zou, Keyan Chen, and Zhenwei Shi. Dense pixel-to-pixel harmonization via continuous image representation. IEEE Transactions on Circuits and Systems for Video Technology, pages 1-1, 2023. 3", + "[5] Zhaoxi Chen and Ziwei Liu. Relighting4d: Neural relightable human from videos. In European Conference on Computer Vision, pages 606-623. Springer, 2022. 3", + "[6] Jun Myeong Choi, Max Christman, and Roni Sengupta. Personalized video relighting with an at-home light stage. In European Conference on Computer Vision, pages 394-410. Springer, 2024. 3", + "[7] Wenyan Cong, Jianfu Zhang, Li Niu, Liu Liu, Zhixin Ling, Weiyuan Li, and Liqing Zhang. Dovenet: Deep image harmonization via domain verification. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8394-8403, 2020. 3", + "[8] Wenyan Cong, Xinhao Tao, Li Niu, Jing Liang, Xuesong Gao, Qihao Sun, and Liqing Zhang. High-resolution image harmonization via collaborative dual transformations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18470-18479, 2022. 3", + "[9] Yuki Endo Daichi Tajima, Yoshihiro Kanamori. Relighting humans in the wild: Monocular full-body human relighting with domain adaptation. Computer Graphics Forum (Proc. of Pacific Graphics 2021), 40(7):205-216, 2021. 3", + "[10] David Futschik, Kelvin Ritland, James Vecore, Sean Fanello, Sergio Orts-Escalano, Brian Curless, Daniel Sykora, and Rohit Pandey. Controllable light diffusion for portraits. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8412-8421, 2023. 3", + "[11] Leon A Gatys, Alexander S Ecker, and Matthias Bethge. Image style transfer using convolutional neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2414-2423, 2016. 5" + ], + "bbox": [ + 93, + 212, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[12] Michal Geyer, Omer Bar-Tal, Shai Bagon, and Tali Dekel. Tokenflow: Consistent diffusion features for consistent video editing. arXiv preprint arxiv:2307.10373, 2023. 6", + "[13] Julian Jorge Andrade Guerreiro, Mitsuru Nakazawa, and Björn Stenger. Pct-net: Full resolution image harmonization using pixel-wise color transformations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5917-5926, 2023. 3", + "[14] Zonghui Guo, Dongsheng Guo, Haiyong Zheng, Zhaorui Gu, Bing Zheng, and Junyu Dong. Image harmonization with transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 14870-14879, 2021.", + "[15] Zonghui Guo, Haiyong Zheng, Yufeng Jiang, Zhaorui Gu, and Bing Zheng. Intrinsic image harmonization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16367-16376, 2021. 3", + "[16] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 5", + "[17] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. arXiv:2204.03458, 2022. 2", + "[18] Alain Hore and Djemel Ziou. Image quality metrics: Psnr vs. ssm. In 2010 20th international conference on pattern recognition, pages 2366-2369. IEEE, 2010. 6", + "[19] Andrew Hou, Ze Zhang, Michel Sarkis, Ning Bi, Yiying Tong, and Xiaoming Liu. Towards high fidelity face relighting with realistic shadows. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14719-14728, 2021. 3", + "[20] Andrew Hou, Michel Sarkis, Ning Bi, Yiying Tong, and Xiaoming Liu. Face relighting with geometrically consistent shadows. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4217-4226, 2022. 3", + "[21] Yasamin Jafarian and Hyun Soo Park. Learning high fidelity depths of dressed humans by watching social media dance videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12753-12762, 2021. 6", + "[22] Chaonan Ji, Tao Yu, Kaiwen Guo, Jingxin Liu, and Yebin Liu. Geometry-aware single-image full-body human relighting. In European Conference on Computer Vision, pages 388-405. Springer, 2022. 2, 3, 6, 7, 8", + "[23] Yifan Jiang, He Zhang, Jianming Zhang, Yilin Wang, Zhe Lin, Kalyan Sunkavalli, Simon Chen, Sohrab Amirghodsi, Sarah Kong, and Zhangyang Wang. Ssh: A self-supervised framework for image harmonization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4832-4841, 2021. 3", + "[24] Yoshihiro Kanamori and Yuki Endo. Relighting humans: occlusion-aware inverse rendering for full-body human images. arXiv preprint arXiv:1908.02714, 2019. 2", + "[25] Hoon Kim, Minje Jang, Wonjun Yoon, Jisoo Lee, Donghyun Na, and Sanghyun Woo. Switchlight: Co-design of physics-driven architecture and pre-training framework for human" + ], + "bbox": [ + 514, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "388", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "portrait relighting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25096-25106, 2024. 3", + "[26] Manuel Lagunas, Xin Sun, Jimei Yang, Ruben Villegas, Jianming Zhang, Zhixin Shu, Belen Masia, and Diego Gutierrez. Single-image full-body human relighting. arXiv preprint arXiv:2107.07259, 2021. 2, 3", + "[27] Chenyang Lei, Xuanchi Ren, Zhaoxiang Zhang, and Qifeng Chen. Blind video deflickering by neural filtering with a flawed atlas. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 5", + "[28] Guannan Li, Yebin Liu, and Qionghai Dai. Free-viewpoint video relighting from multi-view sequence under general illumination. Machine vision and applications, 25:1737-1746, 2014. 3", + "[29] Ruilong Li, Shan Yang, David A. Ross, and Angjoo Kanazawa. Learn to dance with aist++: Music conditioned 3d dance generation, 2021. 6", + "[30] Yang Liu, Alexandros Neophytou, Sunando Sengupta, and Eric Sommerlade. Relighting images in the wild with a self-supervised siamese auto-encoder. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 32–40, 2021. 3", + "[31] Ziwei Liu, Ping Luo, Shi Qiu, Xiaogang Wang, and Xiaou Tang. Deepfashion: Powering robust clothes recognition and retrieval with rich annotations. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1096-1104, 2016. 6", + "[32] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 3", + "[33] Rohit Pandey, Sergio Orts Escolano, Chloe Legendre, Christian Haene, Sofien Bouaziz, Christoph Rhemann, Paul Debevec, and Sean Fanello. Total relighting: learning to relight portraits for background replacement. ACM Transactions on Graphics (TOG), 40(4):1-21, 2021. 1, 2, 3", + "[34] F. Perazzi, J. Pont-Tuset, B. McWilliams, L. Van Gool, M. Gross, and A. Sorkine-Hornung. A benchmark dataset and evaluation methodology for video object segmentation. In Computer Vision and Pattern Recognition, 2016. 6", + "[35] Pexels: https://www.pexels.com.7", + "[36] Puntawat Ponglertnapakorn, Nontawat Tritrong, and Supasorn Suwajanakorn. Difareli: Diffusion face relighting. arXiv preprint arXiv:2304.09479, 2023. 3", + "[37] Ravi Ramamoorthi and Pat Hanrahan. On the relationship between radiance and irradiance: determining the illumination from images of a convex lambertian object. JOSA A, 18 (10):2448-2459, 2001. 3", + "[38] Mengwei Ren, Wei Xiong, Jae Shin Yoon, Zhixin Shu, Jianming Zhang, HyunJoon Jung, Guido Gerig, and He Zhang. Relightful harmonization: Lighting-aware portrait background replacement. CVPR, 2024. 3, 5, 6, 7", + "[39] Christian Richardt, Carsten Stoll, Neil A Dodgson, Hans-Peter Seidel, and Christian Theobalt. Coherent spatiotemporal filtering, upsampling and rendering of rgbz videos. In" + ], + "bbox": [ + 91, + 90, + 483, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Computer graphics forum, pages 247-256. Wiley Online Library, 2012. 3", + "[40] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models, 2021. 4", + "[41] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 2", + "[42] Volker Schönefeld. Spherical harmonics. Computer Graphics and Multimedia Group, Technical Note. RWTH Aachen University, Germany, page 18, 2005. 3", + "[43] Qing Shuai, Chen Geng, Qi Fang, Sida Peng, Wenhao Shen, Xiaowei Zhou, and Hujun Bao. Novel view synthesis of human interactions from sparse multi-view videos. In ACM SIGGRAPH 2022 Conference Proceedings, pages 1-10, 2022. 3", + "[44] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020.3", + "[45] Shuangbing Song, Fan Zhong, Tianju Wang, Xueying Qin, and Changhe Tu. Guided linear upsampling. ACM Transactions on Graphics (TOG), 42(4):1-12, 2023. 5", + "[46] Tiancheng Sun, Jonathan T Barron, Yun-Ta Tsai, Zexiang Xu, Xueming Yu, Graham Fyffe, Christoph Rhemann, Jay Busch, Paul Debevec, and Ravi Ramamoorthi. Single image portrait relighting. ACM Transactions on Graphics (TOG), 38(4):1-12, 2019. 3", + "[47] Daichi Tajima, Yoshihiro Kanamori, and Yuki Endo. Relighting humans in the wild: Monocular full-body human relighting with domain adaptation. In Computer Graphics Forum, pages 205-216. Wiley Online Library, 2021. 3, 6, 7, 8", + "[48] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part II 16, pages 402–419. Springer, 2020. 6", + "[49] Ayush Tewari, Tae-Hyun Oh, Tim Weyrich, Bernd Bickel, Hans-Peter Seidel, Hanspeter Pfister, Wojciech Matusik, Mohamed Elgharib, Christian Theobalt, et al. Monocular reconstruction of neural face reflectance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4791-4800, 2021. 3", + "[50] Marco Toschi, Riccardo De Matteo, Riccardo Spezialetti, Daniele De Gregorio, Luigi Di Stefano, and Samuele Salti. Relight my nerf: A dataset for novel view synthesis and relighting of real world objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20762-20772, 2023. 3", + "[51] Yi-Hsuan Tsai, Xiaohui Shen, Zhe Lin, Kalyan Sunkavalli, Xin Lu, and Ming-Hsuan Yang. Deep image harmonization. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3789-3797, 2017. 3", + "[52] Ke Wang, Michael Gharbi, He Zhang, Zhihao Xia, and Eli Shechtman. Semi-supervised parametric real-world image harmonization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5927-5936, 2023. 3" + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "389", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[53] Yuxin Wang, Wayne Wu, and Dan Xu. Learning unified decompositional and compositional nerf for editable novel view synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 18247-18256, 2023. 3", + "[54] Liu Wen, Zhixin Piao, Min Jie, Wenhan Luo, Lin Ma, and Shenghua Gao. Liquid warping gan: A unified framework for human motion imitation, appearance transfer and novel view synthesis. In The IEEE International Conference on Computer Vision (ICCV), 2019. 6", + "[55] Zexiang Xu, Kalyan Sunkavalli, Sunil Hadap, and Ravi Ramamoorthi. Deep image-based relighting from optimal sparse samples. ACM Transactions on Graphics (ToG), 37 (4):1-13, 2018. 3", + "[56] Zexiang Xu, Sai Bi, Kalyan Sunkavalli, Sunil Hadap, Hao Su, and Ravi Ramamoorthi. Deep view synthesis from sparse photometric images. ACM Transactions on Graphics (ToG), 38(4):1-13, 2019. 3", + "[57] Wenhan Yang, Jiashi Feng, Jianchao Yang, Fang Zhao, Jiaying Liu, Zongming Guo, and Shuicheng Yan. Deep edge guided recurrent residual learning for image superresolution. IEEE Transactions on Image Processing, 26(12): 5895-5907, 2017. 5", + "[58] Mona Zehni, Shaona Ghosh, Krishna Sridhar, and Sethu Ram. Joint learning of portrait intrinsic decomposition and relighting. arXiv preprint arXiv:2106.15305, 2021. 3", + "[59] Hao Zhang, Yanbo Xu, Tianyuan Dai, Tai Chi-Keung Tang, et al. Fdnerf: Semantics-driven face reconstruction, prompt editing and relighting with diffusion models. arXiv preprint arXiv:2306.00783, 2023. 3", + "[60] Jiakai Zhang, Xinhang Liu, Xinyi Ye, Fuqiang Zhao, Yanshun Zhang, Minye Wu, Yingliang Zhang, Lan Xu, and Jingyi Yu. Editable free-viewpoint video using a layered neural representation. ACM Transactions on Graphics (TOG), 40(4):1-18, 2021. 3", + "[61] Longwen Zhang, Qixuan Zhang, Minye Wu, Jingyi Yu, and Lan Xu. Neural video portrait relighting in real-time via consistency modeling. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 802-812, 2021. 1, 3", + "[62] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models, 2023. 4, 5", + "[63] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Scaling in-the-wild training for diffusion-based illumination harmonization and editing by imposing consistent light transport. In The Thirteenth International Conference on Learning Representations, 2025. 3", + "[64] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 6", + "[65] Yulun Zhang, Yapeng Tian, Yu Kong, Bineng Zhong, and Yun Fu. Residual dense network for image super-resolution. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2472-2481, 2018. 5", + "[66] Ruichen Zheng, Peng Li, Haoqian Wang, and Tao Yu. Learning visibility field for detailed 3d human reconstruction and" + ], + "bbox": [ + 91, + 90, + 480, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "relighting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 216-226, 2023. 3", + "[67] Hao Zhou, Sunil Hadap, Kalyan Sunkavalli, and David W. Jacobs. Deep single portrait image relighting. In International Conference on Computer Vision (ICCV), 2019. 3, 7", + "[68] Hao Zhou, Sunil Hadap, Kalyan Sunkavalli, and David W Jacobs. Deep single-image portrait relighting. In Proceedings of the IEEE/CVF international conference on computer vision, pages 7194-7202, 2019. 1, 2, 3, 6, 8", + "[69] Taotao Zhou, Kai He, Di Wu, Teng Xu, Qixuan Zhang, Kuixiang Shao, Wenzheng Chen, Lan Xu, and Jingyi Yu. Relightable neural human assets from multi-view gradient illuminations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4315-4327, 2023. 3", + "[70] Jun-Yan Zhu, Philipp Krahenbuhl, Eli Shechtman, and Alexei A Efros. Learning a discriminative model for the perception of realism in composite images. In Proceedings of the IEEE International Conference on Computer Vision, pages 3943-3951, 2015. 3" + ], + "bbox": [ + 516, + 92, + 903, + 387 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "390", + "bbox": [ + 486, + 945, + 511, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Comprehensive Relighting_ Generalizable and Consistent Monocular Human Relighting and Harmonization/743d9653-d63f-4670-9d09-7ac226ec0ec0_model.json b/2025/Comprehensive Relighting_ Generalizable and Consistent Monocular Human Relighting and Harmonization/743d9653-d63f-4670-9d09-7ac226ec0ec0_model.json new file mode 100644 index 0000000000000000000000000000000000000000..9197db634a1fabf923cbc638843d7e76de9ff5e9 --- /dev/null +++ b/2025/Comprehensive Relighting_ Generalizable and Consistent Monocular Human Relighting and Harmonization/743d9653-d63f-4670-9d09-7ac226ec0ec0_model.json @@ -0,0 +1,2576 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.095, + 0.131, + 0.907, + 0.177 + ], + "angle": 0, + "content": "Comprehensive Relighting: Generalizable and Consistent Monocular Human Relighting and Harmonization" + }, + { + "type": "text", + "bbox": [ + 0.116, + 0.202, + 0.246, + 0.24 + ], + "angle": 0, + "content": "Junying Wang\\(^{1\\dagger}\\) \nHe Zhang\\(^{2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.248, + 0.203, + 0.518, + 0.256 + ], + "angle": 0, + "content": "Jingyuan Liu\\(^{2}\\) Xin Sun\\(^{2}\\) \nJimei Yang\\(^{3}\\) Nanxuan Zhao\\(^{2}\\) \nUlrich Neumann\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.527, + 0.203, + 0.741, + 0.256 + ], + "angle": 0, + "content": "Krishna Kumar Singh\\(^{2}\\) \nTuanfeng Y. Wang\\(^{2}\\) \nJae Shin Yoon\\(^{2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.751, + 0.203, + 0.882, + 0.238 + ], + "angle": 0, + "content": "Zhixin Shu\\(^{2}\\) Simon S. Chen\\(^{2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.211, + 0.273, + 0.486, + 0.293 + ], + "angle": 0, + "content": "1University of Southern California" + }, + { + "type": "text", + "bbox": [ + 0.525, + 0.274, + 0.782, + 0.293 + ], + "angle": 0, + "content": "\\(^{2}\\)Adobe Research \\(^{3}\\)Runway" + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.317, + 0.18, + 0.497 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.103, + 0.5, + 0.183, + 0.51 + ], + "angle": 0, + "content": "Input video frames" + }, + { + "type": "image", + "bbox": [ + 0.184, + 0.318, + 0.266, + 0.408 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.197, + 0.413, + 0.266, + 0.496 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.205, + 0.5, + 0.249, + 0.51 + ], + "angle": 0, + "content": "Relighting" + }, + { + "type": "image", + "bbox": [ + 0.267, + 0.318, + 0.38, + 0.408 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.267, + 0.412, + 0.38, + 0.496 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.295, + 0.5, + 0.359, + 0.509 + ], + "angle": 0, + "content": "Harmonization" + }, + { + "type": "image", + "bbox": [ + 0.382, + 0.318, + 0.504, + 0.496 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.384, + 0.5, + 0.498, + 0.509 + ], + "angle": 0, + "content": "Relighting + Harmonization" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.318, + 0.587, + 0.496 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.503, + 0.5, + 0.583, + 0.509 + ], + "angle": 0, + "content": "Input video frames" + }, + { + "type": "image", + "bbox": [ + 0.587, + 0.318, + 0.665, + 0.408 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.588, + 0.409, + 0.666, + 0.496 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.609, + 0.5, + 0.653, + 0.509 + ], + "angle": 0, + "content": "Relighting" + }, + { + "type": "image", + "bbox": [ + 0.667, + 0.318, + 0.781, + 0.408 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.668, + 0.409, + 0.781, + 0.496 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.696, + 0.499, + 0.759, + 0.508 + ], + "angle": 0, + "content": "Harmonization" + }, + { + "type": "image", + "bbox": [ + 0.782, + 0.318, + 0.899, + 0.408 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.782, + 0.409, + 0.899, + 0.496 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.786, + 0.5, + 0.9, + 0.509 + ], + "angle": 0, + "content": "Relighting + Harmonization" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.516, + 0.907, + 0.558 + ], + "angle": 0, + "content": "Figure 1. We introduce Comprehensive Relighting, a generalizable and consistent model for relighting and harmonization, which controls the lighting property from a single image or video of humans with arbitrary body parts. Given target lighting coefficients, e.g., Spherical harmonics (second), background scenes (third), or their combination (fourth), our model performs consistent and harmonized relighting." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.57, + 0.327, + 0.587 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.602, + 0.486, + 0.875 + ], + "angle": 0, + "content": "This paper introduces Comprehensive Relighting, the first all-in-one approach that can both control and harmonize the lighting from an image or video of humans with arbitrary body parts from any scene. Building such a generalizable model is extremely challenging due to the lack of dataset, restricting existing image-based relighting models to a specific scenario (e.g., face or static human). To address this challenge, we repurpose a pre-trained diffusion model as a general image prior and jointly model the human relighting and background harmonization in the coarse-to-fine framework. To further enhance the temporal coherence of the relighting, we introduce an unsupervised temporal lighting model that learns the lighting cycle consistency from many real-world videos without any ground truth. In inference time, our temporal lighting module is combined with the diffusion models through the spatio-temporal feature blending algorithms without extra training; and we apply a new guided refinement as a post-processing to pre" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.572, + 0.907, + 0.647 + ], + "angle": 0, + "content": "serve the high-frequency details from the input image. In the experiments, Comprehensive Relighting shows a strong generalizability and lighting temporal coherence, outperforming existing image-based human relighting and harmonization methods." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.66, + 0.646, + 0.675 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.688, + 0.907, + 0.824 + ], + "angle": 0, + "content": "Light is the key component that determines how a person looks, which is often defined by a specific time and space, where revisiting such a unique moment gives us the opportunity to experience authentic telepresence sensations. In this paper, we introduce a generalizable human relighting model that can control the lighting from an image or video of humans with arbitrary body parts (Fig. 3), which are well-harmonized with a conditioning space (i.e., background image) as shown in Fig. 1." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.909, + 0.903 + ], + "angle": 0, + "content": "As shown in Fig. 2, existing image-based relighting methods face two main problems, lack of generalizability and controllability. First, they are designed for a specific scenario, e.g., only for face illumination or static humans [33, 61, 68] mainly due to the scarcity of large-scale" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.887, + 0.48, + 0.901 + ], + "angle": 0, + "content": "This work is partially done during an internship at Adobe Research." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.945, + 0.512, + 0.957 + ], + "angle": 0, + "content": "380" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.089, + 0.482, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.22, + 0.483, + 0.248 + ], + "angle": 0, + "content": "Figure 2. Comparison of various baseline methods for relighting settings and functionalities." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.259, + 0.483, + 0.456 + ], + "angle": 0, + "content": "relighting datasets: Precise acquisition of the appearance of a static person under assorted lighting conditions requires specialized setups such as LightStage [22, 33] or expensive graphics simulation [26], which are not scalable, particularly for video contents. Learning from such limited data induces weak generalization of the model to diverse scenes. Second, most relighting algorithms struggle to effectively model multiple light sources. Typically, these algorithms are restricted to a single lighting control from either background image (e.g. high dynamic range lighting environment map [22]) or target lighting parameters (e.g. Spherical harmonics [24]). These problems inhibit the production-level application that requires general use cases." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.457, + 0.483, + 0.744 + ], + "angle": 0, + "content": "To overcome these challenges, we propose an effective approach to achieve all-in-one relighting by utilizing a diffusion model—a general image prior that learns massive visual data with diverse lighting conditions; and repurpose this prior specialized for a human relighting and harmonization in a coarse-to-fine framework: A pretrained latent diffusion model [41] learns from limited datasets of static humans to jointly perform the fine-grained relighting and harmonization from two multi-modal lighting variables: the coarse shading estimate and conditioning background scene. The coarse shading estimate can be \"computed\" from conditioning lighting parameters (i.e., Spherical harmonics) without a neural network, and therefore, it is generalizable. The diffusion model is required to learn only the residual portion (e.g., fine self-occluded shadow), which is more generalizable than direct relighting (e.g., [68]). The pre-trained image prior helps with understanding the properties of the general background scenes, enabling the generalization of the background harmonization." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.745, + 0.483, + 0.851 + ], + "angle": 0, + "content": "Our coarse-to-fine framework, however, still introduces problems: the diffusion model that learns to generate an image without temporal context produces significant temporal artifacts (e.g., sudden changes of lighting distribution even for consecutive frames). We address this problem by introducing an unsupervised temporal lighting model that learns from many real videos without any ground-truth data to en" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.089, + 0.905, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.256, + 0.905, + 0.298 + ], + "angle": 0, + "content": "Figure 3. Our model generalizes to various body parts (portrait, half-body, full-body, multiperson) for relighting and harmonization, with lighting control variables shown in the insets." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.309, + 0.905, + 0.4 + ], + "angle": 0, + "content": "force the temporal lighting consistency over frames. This temporal module learns the unsupervised cycle consistency between the relit and many real videos to predict the future lighting distribution from the past, which can be directly combined with our coarse-to-fine relighting components without extra training." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.401, + 0.906, + 0.567 + ], + "angle": 0, + "content": "In inference time, we further push the temporal coherence of the relighting and harmonization by formulating a recurrent prediction pipeline. The generation at the current time step is conditioned onto the temporal module for the one at the next time. The features from the lighting and temporal control modules are spatially and temporally blended to enforce temporal coherence while improving the structure of the lighting distribution. Finally, our guided refinement module enhances the quality of the generated image in a way that preserves the original high-frequency details of the input image." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.568, + 0.906, + 0.674 + ], + "angle": 0, + "content": "In the experiments, Comprehensive Relighting demonstrates high-quality relighting and background harmonization results with strong temporal coherence across the lighting, background, and pose changes. It also shows strong generalizability to any unconstrained scenes regardless of body parts, views, and poses, outperforming existing relighting and background harmonization methods." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.906, + 0.811 + ], + "angle": 0, + "content": "Our contributions include: (1) To the best of our knowledge, the first approach for joint modeling of relighting and background harmonization; (2) a novel coarse-to-fine framework that enables comprehensive generalization by only learning from limited synthetic and lab-controlled data; (3) unsupervised temporal modeling with lighting cycle consistency from many unconstrained real videos; (4) an effective inference algorithm with spatio-temporal feature blending and guided refinement." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.828, + 0.655, + 0.843 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Image-based Human Relighting While high-quality relighting often requires resource-intensive setups, recent efforts in mobile device relighting aim to address single" + }, + { + "type": "page_footnote", + "bbox": [ + 0.09, + 0.863, + 0.483, + 0.901 + ], + "angle": 0, + "content": "To the best of our knowledge, there exists no public ground truth data for video relighting of dynamic humans, and therefore, fine-tuning an existing video diffusion model (e.g., [17]) is not a readily available option." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "381" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.089, + 0.092, + 0.482, + 0.228 + ], + "angle": 0, + "content": "image scenarios. Total Relighting [33] achieves photorealistic effects by leveraging detailed normal maps and albedo as priors. However, it relies on an HDR environment map that accurately matches the scene's real-world lighting, which is not always readily available, particularly in flexible and personalized relighting scenarios. [37, 46] estimate and adjust the Spherical harmonics parameters from images, and others use one-light-at-a-time (OLAT) captures to generate relighting data [49] or estimate reflectance fields [46]." + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.23, + 0.486, + 0.532 + ], + "angle": 0, + "content": "Recent works have focused on diverse relighting scenarios including portrait [10, 19, 20, 25, 30, 33, 46, 58, 67, 68], full-body [9, 22, 26, 47], and object relighting [1, 55, 56]. Most of these works rely on decomposing an image into its intrinsic components, i.e., albedo, normal, and lighting, and therefore, the accuracy of this decomposition directly impacts the quality of the final relighting. While some works [22, 66, 69] use single-image geometry reconstruction for traditional and neural rendering, reconstruction errors are often propagated to relighting results. Tajima et al. [47] tackled domain adaptation with a two-step relighting framework, yet noticeable texture distortion remains due to limited model generalizability. Zhang et al. [59] trained a 2D latent-diffusion model, allowing users to manipulate and construct face NeRFs in a zero-shot learning framework without the need for explicit 3D data. DiFaRei [36] utilizes DDIM (Denoising Diffusion Implicit Models [44]) for high-fidelity face relighting. While promising, their methods are constrained by a focus on either specific body parts, limiting their applicability for generalizable." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.535, + 0.483, + 0.822 + ], + "angle": 0, + "content": "Background Harmonization Background harmonization seeks to harmonize color, contrast, and style discrepancies between the foreground and background, ensuring composite images appear natural and cohesive. Many existing background harmonization methods [4, 7, 8, 13-15, 23, 51, 52, 70] formulate this problem as image-to-image translation work where a neural network translates an unharmonized foreground image to the harmonized one in the context of the conditioning background image. Recently, Relightful Harmonization [38] introduced methods that harmonize both image style and lighting to match the background scene, and IC-Light [63] further enables flexible illumination control via image diffusion models, guided by text descriptions or background images. While promising, these methods lack explicit lighting control, restricting general applicability and consistent video relighting. Additionally, some ([38]) are limited to specific body parts (e.g., portraits), with constrained support for full-body and multiperson scenarios." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Video Relighting Neural Radiance Fields (NeRF) [32] based methods enable novel view synthesis under varying lighting conditions in videos [43, 50, 53, 69]. Zhang et al. [61] achieve portrait video relighting under dynamic illuminations, while Choi et al. [6] ensure temporally consis" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.32 + ], + "angle": 0, + "content": "tent relit videos. 3D-PVR [3] present a 3D-aware, real-time method to relight videos of talking faces. Relighting4D [5] decomposes the time-varied human body as a set of neural fields of normal, occlusion, diffuse, and specular maps. However, their rendering quality are highly reliant on geometry accuracy. Some works [36] adopt a temporal modeling scheme real-time neural video portrait relighting. STNeRF [60] controls dynamic scenes using a neural layered radiance representation that maintains spatial and temporal coherence. Li et al. [28] employed multi-view reconstruction for free-viewpoint relighting videos under general illumination. Richardt et al. [39] use RGBZ video cameras for video effects, including relighting, but rely on multi-view reconstruction from specialized hardware, making the process highly complex and costly." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.333, + 0.605, + 0.349 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.359, + 0.907, + 0.602 + ], + "angle": 0, + "content": "We develop a generalizable and consistent human relighting and harmonization framework using a diffusion model. Fig. 4 illustrates the overview of our framework. Given an input image of humans and control lighting variables, including coarse shading and background image, a diffusion model learns to generate a fine-grained image of the humans under the target lighting, also harmonized with the background (Sec. 3.2). An external temporal lighting module, trained on real videos using unsupervised temporal cycle consistency, is integrated into the diffusion model to enhance temporal lighting coherence (Sec. 3.3). In inference time, we blend the features between the lighting and temporal control modules over time to ensure the relighting results are accurate and temporally coherent, and we apply a guided refinement to prevent the loss of high-frequency details during the denoising process (Sec. 3.4)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.612, + 0.838, + 0.629 + ], + "angle": 0, + "content": "3.1. Background: Image-based Relighting" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.634, + 0.906, + 0.71 + ], + "angle": 0, + "content": "Image-based relighting function can be compactly modeled with a small number of approximated basis of Spherical harmonics (SH) [9, 26, 42] which describes only essential features of illumination on the surface of a 3D sphere based on the following formulation:" + }, + { + "type": "equation", + "bbox": [ + 0.57, + 0.722, + 0.907, + 0.765 + ], + "angle": 0, + "content": "\\[\nI (\\mathbf {x}) = \\rho (\\mathbf {x}) \\cdot \\sum_ {l = 0} ^ {k} \\sum_ {m = - l} ^ {l} \\phi Y _ {l m} (\\mathbf {n} (\\mathbf {x})) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.779, + 0.907, + 0.903 + ], + "angle": 0, + "content": "where \\(\\phi \\in \\mathbb{R}^{(k + 1)^2}\\) denotes the spherical harmonics (SH) coefficients; \\(Y_{lm}(\\mathbf{n})\\) are the SH basis functions evaluated at the surface normal \\(\\mathbf{n}\\); indices \\(l\\) and \\(m\\) represent the band and order within each band, respectively; and \\(k\\) is the maximum order of spherical harmonics used. While SH is computationally efficient and highly generalizable, it only captures low-frequency lighting, limiting its ability to model fine-grained, high-frequency illumination details. Addition-" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.945, + 0.511, + 0.956 + ], + "angle": 0, + "content": "382" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.1, + 0.092, + 0.437, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.171, + 0.277, + 0.368, + 0.287 + ], + "angle": 0, + "content": "(a) Coarse-to-fine relighting and harmonization" + }, + { + "type": "image", + "bbox": [ + 0.441, + 0.094, + 0.783, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.482, + 0.277, + 0.674, + 0.288 + ], + "angle": 0, + "content": "(b) Temporal modeling with cycle consistency" + }, + { + "type": "image", + "bbox": [ + 0.787, + 0.094, + 0.897, + 0.269 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.703, + 0.277, + 0.899, + 0.288 + ], + "angle": 0, + "content": "(c) Inference: Spatio-temporal feature blending" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.295, + 0.907, + 0.352 + ], + "angle": 0, + "content": "Figure 4. System overview. (a) Given an input image of humans with coarse lighting and background image, our diffusion model generates the relit images harmonized with background scenes (Sec. 3.2). (b) The external temporal modules learn the temporal cycle consistency from many real-world videos to construct temporal lighting features (Sec. 3.3). (c) In inference time, we blend the features from lighting and temporal modules spatially and temporally to enable coherent and generalizable human relighting (Sec. 3.4)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.366, + 0.484, + 0.472 + ], + "angle": 0, + "content": "ally, SH-based global illumination neglects the context provided by background images, often resulting in relit images appearing unnatural when composited with different backgrounds. To overcome these limitations, we propose a coarse-to-fine human relighting and harmonization framework that leverages the strong image prior available from a pre-trained text-to-image diffusion model." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.481, + 0.482, + 0.497 + ], + "angle": 0, + "content": "3.2. Coarse-to-Fine Relighting and Harmonization" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.503, + 0.483, + 0.533 + ], + "angle": 0, + "content": "We generate a fine-grained relit image of a person conditioned on a coarse lighting representation:" + }, + { + "type": "equation", + "bbox": [ + 0.177, + 0.543, + 0.482, + 0.56 + ], + "angle": 0, + "content": "\\[\n\\mathcal {E} (\\mathbf {I}; \\mathbf {S} _ {\\phi}) = z, \\quad \\mathcal {D} (z) = \\mathbf {I} _ {\\phi} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.569, + 0.483, + 0.869 + ], + "angle": 0, + "content": "where \\(\\mathcal{E}\\) is an encoder that generates the latent lighting features, \\(z\\) as a function of the input image \\(\\mathbf{I} \\in \\mathbb{R}^{w \\times h \\times 3}\\) and a small number of global lighting parameters \\(\\phi \\in \\mathbb{R}^n\\) (i.e., Spherical harmonics where \\(n = 25\\)); and \\(\\mathcal{D}\\) is the decoder that generates fine-grained relit image \\(\\mathbf{I}_{\\phi} \\in \\mathbb{R}^{w \\times h \\times 3}\\) from \\(z\\). \\(\\mathbf{S}_{\\phi} \\in \\mathbb{R}^{w \\times h}\\) is the pixel-aligned coarse shading map converted from the coarse lighting parameters \\(\\phi\\) as shown in Fig. 4-(a). One approach to obtaining \\(\\mathbf{S}_{\\phi}\\) is directly computing it as demonstrated in Eq. 1 where we can compute lighting intensity along with the detected surface normal map from \\(\\mathbf{I}\\) given spherical harmonics coefficients \\(\\phi\\). Another way, is to use a neural network to convert the surface normal to \\(\\mathbf{S}_{\\phi}\\) as a condition of spherical harmonics coefficients \\(\\phi\\), i.e., \\(\\mathbf{S}_{\\phi} \\gets f(\\mathbf{N}; \\phi)\\) where \\(f\\) is the neural shading function which maps the surface normal \\(\\mathbf{N}\\) and \\(\\phi\\) to the coarse shading. While both approaches are highly generalizable, our experiments indicate that the latter method achieves improved accuracy and better noise correction. Please refer to the Supple. documents for more details about the coarse shading estimation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.902 + ], + "angle": 0, + "content": "To jointly model the relighting and background harmonization, we further condition a target background image" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.365, + 0.833, + 0.381 + ], + "angle": 0, + "content": "\\(\\mathbf{B} \\in \\mathbb{R}^{w \\times h \\times 3}\\) as the additional lighting sources:" + }, + { + "type": "equation", + "bbox": [ + 0.581, + 0.393, + 0.905, + 0.41 + ], + "angle": 0, + "content": "\\[\n\\mathcal {E} (\\mathbf {I}; \\{\\mathbf {S} _ {\\phi}, \\mathbf {B} \\}) = \\boldsymbol {z}, \\quad \\mathcal {D} (\\boldsymbol {z}) = \\mathbf {I} _ {\\phi} \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.42, + 0.906, + 0.511 + ], + "angle": 0, + "content": "where the lighting encoder \\(\\mathcal{E}\\) learns to capture the intensity and direction of light from \\(\\mathbf{S}_{\\phi}\\) while capturing ambient environment lighting and color distribution from \\(\\mathbf{B}\\). This enables \\(\\mathcal{D}\\) to achieve complete relighting in scenarios involving new target lighting, background, or both. In the subsection, we enable \\((\\mathcal{D} \\circ \\mathcal{E})\\) using a latent diffusion model." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.518, + 0.844, + 0.533 + ], + "angle": 0, + "content": "3.2.1. Fine-grained Relighting Diffusion Model" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.537, + 0.907, + 0.596 + ], + "angle": 0, + "content": "We enable the fine-grained image relighting using a diffusion model as illustrated in Fig. 4-(a). Our encoder \\(\\mathcal{E}\\) in Eq. 3, in practice, is formulated as the composition of two encoders:" + }, + { + "type": "equation", + "bbox": [ + 0.549, + 0.608, + 0.905, + 0.626 + ], + "angle": 0, + "content": "\\[\n\\mathcal {E} \\rightarrow \\mathcal {E} _ {\\mathrm {b}} (\\mathbf {I}; \\mathcal {E} _ {1} (\\{\\mathbf {S} _ {\\phi}, \\mathbf {B} \\})) = \\boldsymbol {z}, \\quad \\mathcal {D} (\\boldsymbol {z}) = \\mathbf {I} _ {\\phi} \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.635, + 0.907, + 0.817 + ], + "angle": 0, + "content": "where \\(\\mathcal{E}_1\\) encodes lighting control variables, \\(\\{\\mathbf{S}_{\\phi}, \\mathbf{B}\\}\\), and \\(\\mathcal{E}_b\\) encodes the conditional variable \\(\\mathbf{I}\\) whose visual properties, e.g., semantics and identity, should be preserved in the output along with the controls from \\(\\mathcal{E}_1\\). For \\(\\mathcal{E}_b\\), we use the base latent diffusion model [2, 40] pre-trained for text-to-image generation task, and for \\(\\mathcal{E}_1\\), we use ControlNet [62] (termed as Light ControlNet in Fig. 4-(a)). While noise, texts, and timestep variables are also conditioned on \\(\\mathcal{E}_b\\) to fit the modality of the latent diffusion model, they are not described in the equations and figures for conciseness. To impose the foreground attention on the lighting control, a portrait mask is also conditioned to \\(\\mathcal{E}_l\\):" + }, + { + "type": "equation", + "bbox": [ + 0.555, + 0.828, + 0.905, + 0.845 + ], + "angle": 0, + "content": "\\[\n\\mathcal {E} _ {\\mathrm {b}} (\\mathbf {I}; \\mathcal {E} _ {1} (\\{\\mathbf {S} _ {\\phi}, \\mathbf {B} \\}; \\mathbf {M})) = \\boldsymbol {z}, \\quad \\mathcal {D} (\\boldsymbol {z}) = \\mathbf {I} _ {\\phi} \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.855, + 0.907, + 0.901 + ], + "angle": 0, + "content": "where \\(\\mathbf{M} \\in \\{0,1\\}^{w \\times h}\\) is the foreground binary mask. In training time, the diffusion model learns to directly change the lighting distribution of the input images (without inverse" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "383" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.244 + ], + "angle": 0, + "content": "rendering techniques) as the condition of coarse lighting estimate by minimizing the latent distance of the noisy relit image with the clean one from the ground truth in the overall forward and background denoising steps. Following the findings from an existing harmonization work [38], we use the composite image between the input image and conditioning background image as \\(\\mathbf{I}\\). We randomly drop background \\(\\mathbf{B}\\) or randomly set the coarse shading \\(\\mathbf{S}_{\\phi}\\) as binary mask such that the diffusion model learns the control the lighting from \\(\\mathbf{B}\\), \\(\\mathbf{S}_{\\phi}\\), or both \\(\\{\\mathbf{B}, \\mathbf{S}_{\\phi}\\}\\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.252, + 0.453, + 0.269 + ], + "angle": 0, + "content": "3.3. Unsupervised Add-on Temporal Modeling" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.273, + 0.485, + 0.41 + ], + "angle": 0, + "content": "Our coarse-to-fine relighting framework \\((\\mathcal{D} \\circ \\mathcal{E}_{\\mathrm{b}} \\circ \\mathcal{E}_{\\mathrm{l}})\\) is trained only on individual images, inherently missing temporal context, e.g., how a point on a human's surface will radiate from a specific viewpoint under the continuous pose, view, and illumination changes, leading to temporal artifacts such as flickering. We model such temporal context by designing an external add-on temporal lighting module \\(\\mathcal{E}_{m}\\) that can be combined, in inference time, with the relighting framework without extra training:" + }, + { + "type": "equation", + "bbox": [ + 0.155, + 0.421, + 0.483, + 0.437 + ], + "angle": 0, + "content": "\\[\n\\mathcal {D} \\circ \\mathcal {E} _ {\\mathrm {b}} \\circ \\mathcal {E} _ {\\mathrm {l}} \\rightarrow \\mathcal {D} \\circ \\mathcal {E} _ {\\mathrm {b}} \\circ \\left(\\mathcal {E} _ {\\mathrm {l}} \\times \\mathcal {E} _ {\\mathrm {m}}\\right) \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.449, + 0.484, + 0.54 + ], + "angle": 0, + "content": "To enable this, our temporal module is designed to regress the relit image from the previous time instance, i.e., \\(\\mathbf{I}_{\\phi}^{t - 1}\\), to the latent lighting distribution whose space is shared with our base relighting models. Therefore, our decoder can generate temporally coherent relit images in the current time in an auto-regressive way:" + }, + { + "type": "equation", + "bbox": [ + 0.171, + 0.549, + 0.483, + 0.568 + ], + "angle": 0, + "content": "\\[\n\\mathcal {D} \\left(\\mathcal {E} _ {\\mathrm {b}} \\left(\\mathbf {I} ^ {t}; \\mathcal {E} _ {\\mathrm {m}} \\left(\\mathbf {I} _ {\\phi} ^ {t - 1}\\right)\\right)\\right) = \\mathbf {I} _ {t - 1} ^ {t}. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.578, + 0.484, + 0.655 + ], + "angle": 0, + "content": "However, training \\(\\mathcal{E}_{\\mathrm{m}}\\) with a conventional L2 loss is not possible since there exists no ground-truth video relighting data for dynamic humans. Therefore, we propose to learn \\(\\mathcal{E}_{\\mathrm{m}}\\) in an unsupervised way using many real videos with lighting cycle consistency as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.156, + 0.664, + 0.483, + 0.682 + ], + "angle": 0, + "content": "\\[\n\\mathcal {D} ^ {*} \\left(\\mathcal {E} _ {\\mathrm {b}} ^ {*} \\left(\\mathbf {I} ^ {t}; \\mathcal {E} _ {1} ^ {*} \\left(\\left\\{\\mathbf {S} _ {\\phi} ^ {t}, \\mathbf {B} ^ {t} \\right\\}; \\mathbf {M} ^ {t}\\right)\\right)\\right) = \\mathbf {I} _ {\\phi} ^ {t}, \\tag {8}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.16, + 0.685, + 0.482, + 0.703 + ], + "angle": 0, + "content": "\\[\n\\mathcal {D} ^ {*} \\left(\\mathcal {E} _ {\\mathrm {b}} ^ {*} \\left(\\mathbf {I} _ {\\phi} ^ {t}; \\mathcal {E} _ {\\mathrm {m}} \\left(\\mathbf {I} ^ {t - 1}, \\mathbf {M} ^ {t - 1}\\right)\\right)\\right) = \\tilde {\\mathbf {I}} _ {t - 1} ^ {t} \\tag {9}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.237, + 0.706, + 0.338, + 0.724 + ], + "angle": 0, + "content": "\\[\n\\therefore \\quad \\mathbf {I} ^ {t} = \\tilde {\\mathbf {I}} _ {t - 1} ^ {t}.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.484, + 0.904 + ], + "angle": 0, + "content": "We make the hypothesis: a video sequence inherently contains temporal lighting properties whose flow can be implicitly modeled by learning to predict the lighting distribution of the future frame based on the hint from the previous frame. This involves forward and backward processes, as a cycle-training. Eq. 8 represents the forward image relighting, i.e., \\(\\mathbf{I}^t\\rightarrow \\mathbf{I}_\\phi^t\\) , where our relighting diffusion model with lighting ControlNet \\(\\mathcal{E}_1\\) generates the relit image at time \\(t\\) as a condition of a novel lighting condition (where we pick random Spherical harmonics). Eq. 9 reverts the relighting, i.e., \\(\\tilde{\\mathbf{I}}_{t - 1}^{t}\\gets \\mathbf{I}_{\\phi}^{t}\\) to the original input image conditioned by" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.213 + ], + "angle": 0, + "content": "the original frame in the previous time step through our temporal lighting module where the mask \\(\\mathbf{M}\\) is used for foreground awareness. Finally, \\(\\mathcal{E}_m\\) learns the lighting cycle consistency in the diffusion process by minimizing the latent distance between the reverted and the original image. We freeze the learnable weights for the functions with \\(*\\) during training, and \\(\\mathcal{E}_{\\mathrm{m}}\\) is enabled with another ControlNet [62] termed as Motion ControlNet in Fig. 4-(b)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.22, + 0.907, + 0.25 + ], + "angle": 0, + "content": "3.4. Inference with Spatio-Temporal Blending and Guided Refinement" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.257, + 0.908, + 0.514 + ], + "angle": 0, + "content": "Given a video or image, we perform comprehensive relighting with a spatio-temporal feature blending framework. For \\( t = 0 \\), we generate the relit image without our temporal lighting module \\( \\mathcal{E}_{\\mathrm{m}} \\). For \\( t > 0 \\) (note that even for the static image, \\( t > 0 \\) is possible since the lighting is time-variant), the generated relit images in the previous time step are conditioned on \\( \\mathcal{E}_{\\mathrm{m}} \\), and therefore, the relighting is controlled by dual control modules, i.e., \\( \\mathcal{E}_{\\mathrm{l}} \\) and \\( \\mathcal{E}_{\\mathrm{m}} \\) by blending their features with a (0.85:0.15) ratio as described as spatial blending block in Fig. 4-(c). The blended lighting features are recurrently combined with the one from the previous time step through the temporal blending block. For this, we adopt optical flow from the original input video as a temporal prior to spatially align the features from consecutive frames similar to [27]; and we blend the aligned temporal features with a (0.5:0.5) ratio as in Fig. 4-(c) to improve the temporal consistency." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.514, + 0.907, + 0.605 + ], + "angle": 0, + "content": "In the denoising process of the latent diffusion model, the generation often suffers from the loss of high-frequency details. Inspired by existing image restoration techniques [45, 65], we address this problem by applying a guided refinement to each generated image. We cast the problem of guided refinement as a guided residual prediction:" + }, + { + "type": "equation", + "bbox": [ + 0.605, + 0.61, + 0.905, + 0.628 + ], + "angle": 0, + "content": "\\[\n\\mathbf {I} _ {\\phi} ^ {\\text {r e f i n e}} = \\mathbf {I} + g \\left(\\mathbf {I} _ {\\phi}, \\mathbf {I}; \\mathbf {M}\\right), \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.633, + 0.907, + 0.77 + ], + "angle": 0, + "content": "where \\( g \\) is the function that predicts the guided lighting residual. This residual learns to map the lighting distribution from \\( \\mathbf{I} \\) to \\( \\mathbf{I}_{\\phi} \\). Here, \\( \\mathbf{I}_{\\phi}^{\\mathrm{refine}} \\) can effectively preserve the high-frequency details of the input image \\( \\mathbf{I} \\) due to the nature of residual learning [57, 65] that is designed to preserve the visual properties from the observation space, i.e., \\( \\mathbf{I} \\). We enable \\( g \\) with a residual network [16] by learning from our relighting data with general losses for low-level vision processing, i.e., L2 and VGG [11]." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.777, + 0.646, + 0.794 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Dataset. To train our coarse-to-fine relighting model, we use 100K ground-truth relighting samples, including 50K synthetic human renderings and 50K OLAT-captured images from LightStage, with random cropping augmentation. Ground-truth albedo, images, background, masks, and lighting coefficients are precomputed. Our dataset is categorized by gender, skin tone, and body part (full-body and" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "384" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.089, + 0.908, + 0.177 + ], + "angle": 0, + "content": "
MethodScenario 1Scenario 2Scenario 3
DPR [68]18.62/0.86/0.10332.00 /0.94/0.03018.14/0.89/0.08935.29 /0.98/0.03221.20/0.89/0.07237.11/0.95/0.038
RHW [47]19.78/0.87/0.11330.74/0.95/0.02720.12/0.88/0.07836.64 /0.98/0.02823.33/0.90/0.06037.53 /0.98/0.033
GFR [22]25.59/0.91/0.08930.33/0.95/0.03322.76/0.91/0.07232.36/0.98/0.03625.49/0.93/0.05035.89/0.98/0.028
LPBR [38]18.19/0.86/0.09031.62/0.91/0.03519.96/0.88/0.08427.94/0.94/0.04121.42/0.88/0.05333.39/0.95/0.038
Ours25.95/0.95/0.06633.58/0.96/0.02423.99/0.93/0.04835.18/0.99 / 0.03126.61/0.94/0.03338.32/0.98/0.026
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.18, + 0.907, + 0.224 + ], + "angle": 0, + "content": "Table 1. Comparison with existing image-based human relighting methods on synthetic videos for fidelity and temporal consistency. Each column shows image fidelity (left) and video temporal consistency (right). Metrics are PSNR↑/SSIM↑/LPIPS↓ for accuracy, and tPSNR↑/tSSIM↑/tLPIPS↓ for temporal consistency. While green is used for the best values, yellow highlights the second-best values." + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.234, + 0.905, + 0.522 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.527, + 0.908, + 0.556 + ], + "angle": 0, + "content": "Figure 5. Qualitative comparison of synthetic video frames (corresponding to Tab. 1). From left to right: composite input with target lighting parameters (inset), our relit result, baseline methods, and normalized L2\\(\\downarrow\\) photometric error map (inset)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.568, + 0.483, + 0.884 + ], + "angle": 0, + "content": "multi-body), with each subject captured from 32 viewpoints under varying lighting conditions. For a detailed dataset breakdown, refer to the Supple. To train our temporal control module using real data, we used 50K frames of videos from existing works [21, 29, 34, 54] and customized videos, where, for cyclic relighting, we randomly sample spherical harmonics lighting parameters from ground-truth data. For validation on static humans with dynamic lighting, we generated a new testing set (i.e., the ground-truth relit images and associated lighting parameters) using an internal simulation and LightStage. This set includes the mixture of half-body, full-body, and multi-person scenarios, with each scenario comprising 100 frames. For the validation on dynamic humans (i.e., video), we newly create synthetic video sequences for three scenarios: Scenario 1 involves static lighting (environment map) across frames with a moving human; Scenario 2 features dynamic lighting (rotated environment map) with a static human; Scenario 3 combines dynamic lighting (rotated environment map) with a moving human. Please refer to the Supple. and demo video for more details about the testing results." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.886, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Metrics. To measure the relighting quality, we use L1 dis" + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.566, + 0.646, + 0.739 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.646, + 0.566, + 0.799, + 0.739 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.776, + 0.566, + 0.901, + 0.739 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.74, + 0.907, + 0.781 + ], + "angle": 0, + "content": "Figure 6. Comparison with LPBR [38] on DeepFashion [31] real images for background harmonization testing. The first row shows the relit output, and the second shows the magnified results." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.907, + 0.903 + ], + "angle": 0, + "content": "tance, reconstruction fidelity (PSNR [18]), local structure similarity (SSIM [18]), and the latent perceptual distance (LPIPS [64]), between the relighted and ground-truth. For measuring the temporal coherence of the relighting results, we follow the same logics as TokenFlow [12]. This involves warping the relit image from the previous time step to the next using optical flow [48], and then comparing this to the" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "385" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.098, + 0.088, + 0.904, + 0.395 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.4, + 0.908, + 0.43 + ], + "angle": 0, + "content": "Figure 7. Comparison with RHW [47] on Pexels [35] real images. The lighting control variables are shown as insets. While RHW produces reasonable relighting for full-body images, its quality degrades on half-body and multi-person cases." + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.436, + 0.907, + 0.664 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.668, + 0.908, + 0.698 + ], + "angle": 0, + "content": "Figure 8. Comparison with GFR [22] on Pexels [35] real images. The lighting control variables are shown as insets. Limited generalizability of GFR results in reduced output quality for half-body and multi-person cases." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.706, + 0.482, + 0.737 + ], + "angle": 0, + "content": "current frame using the aforementioned metrics, termed \\( tL_{1} \\) error, tPSNR, tSSIM, and tLPIPS." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Baseline. We compare our method with existing monocular human relighting works where we chose the baselines that are applicable to general scenes (e.g., any part-specific information such as a 3D face model is not a requirement): DPR [67], RHW [47], and GFR [22]. For the harmonization baseline, we compare our model with LPBR [38], a diffusion-based light-aware harmonization method. For relighting, aside from GFR, baselines use Spherical harmonics for lighting control without modeling background illumination. DPR and RHW are evaluated using their released" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.706, + 0.907, + 0.798 + ], + "angle": 0, + "content": "pre-trained models. Since GFR lacks available code, we re-implemented it using our dataset, replacing HDR lighting with Spherical harmonics and background modeling to enable background harmonization in our experiments. Harmonization is directly compared with LPBR by replacing the background." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Results. For testing static humans under dynamic lighting, we present the quantitative comparison in Tab. 2. We show average numerical evaluations on our synthetic testing dataset, categorized by portrait, full-body, and multiperson. Further validation by gender, and skin color is detailed in the Supple. Methods such as RHW and DPR have" + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.513, + 0.957 + ], + "angle": 0, + "content": "386" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.104, + 0.089, + 0.477, + 0.181 + ], + "angle": 0, + "content": "
MethodL1↓PSNR↑SSIM↑
Ours-diffusion0.0583717.0990.833
Ours end-to-end0.0143226.4180.918
Ours-background0.0123927.3460.945
Ours0.0103528.4190.948
Ours+refine0.0101228.7780.949
" + }, + { + "type": "table_caption", + "bbox": [ + 0.109, + 0.184, + 0.465, + 0.198 + ], + "angle": 0, + "content": "Table 3. Ablation study on coarse-to-fine relighting models." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.209, + 0.485, + 0.421 + ], + "angle": 0, + "content": "limited generalizability for both full-body and portrait relighting. They are difficult to extend to other scenarios and tend to show reduced performance when tested in different settings. While GFR performs well in our evaluation, it struggles with significant domain gaps, leading to noticeable quality degradation on real data, including distortions and color shifts (Fig. 8). In contrast, our model exhibits strong generalizability across validation sets and real-world tests (Figs. 7, 8). Compared to the state-of-the-art background harmonization method (LPBR), our method shows the strong generalization to different body parts, and notably, LPBR often includes significant distortion when applied to the full-body images and it does not support the lighting control function as shown in Fig. 6." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.43, + 0.483, + 0.483 + ], + "angle": 0, + "content": "
CategoryDPR [68]RHW [47]GFR [22]Ours
Portrait17.74 / 0.8715.75 / 0.8217.71 / 0.8623.04 / 0.90
Full-body27.62 / 0.9627.73 / 0.9529.51 / 0.9530.81 / 0.97
Multi-person25.70 / 0.9525.69 / 0.9529.35 / 0.9731.49 / 0.96
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.485, + 0.483, + 0.514 + ], + "angle": 0, + "content": "Table 2. Comparison on our synthetic static testing data sorted by body-part. We compute average PSNR↑ / SSIM↑." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.524, + 0.483, + 0.659 + ], + "angle": 0, + "content": "For video testing, in Tab. 1, we evaluate our model on scenarios 1, 2, and 3 for both fidelity and temporal consistency. Other approaches face challenges in achieving both relit fidelity and temporal consistency at the same time as also shown in Fig. 5. In contrast, our temporal module ensures our comprehensive relighting model to produce videos with strong temporal consistency. For more results on relighting results and comparisons including user study, please refer to the Supple." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.66, + 0.483, + 0.84 + ], + "angle": 0, + "content": "Ablation Study. We conduct two ablation studies on our coarse-to-fine model using static human data and our temporal modules using dynamic human data. As shown in Tab. 3, we perform the ablation study on our coarse-to-fine approach: 1) Ours-diffusion: Relighting only with pixel-aligned coarse shading without diffusion model, i.e., \\(\\mathbf{S}_{\\phi} \\times \\mathbf{I}\\). 2) Ours end-to-end: Relighting in an end-to-end manner by applying target lighting parameters \\(\\phi\\) as a condition, trained with a diffusion model, without the coarse stage. 3) Ours-background: Relighting without the control of background B. 4) Ours+refine (full): Applying our guided refinement to the relit image." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Tab. 3 shows the summary of the ablation study: Directly apply the coarse stage without a diffusion model introduces significant errors in the final relit result due to the detection noises. Instead of applying target lighting and end-to" + }, + { + "type": "table", + "bbox": [ + 0.526, + 0.089, + 0.9, + 0.15 + ], + "angle": 0, + "content": "
MethodtL1↓tPSNR↑tSSIM↑
Ours6.55231.0280.956
Ours+temporal5.63832.2660.957
Ours+temporal+blend4.01933.5880.957
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.153, + 0.905, + 0.181 + ], + "angle": 0, + "content": "Table 4. Ablation study on our temporal modules evaluated on synthetic sequences: tL1 error \\((\\times 10^{-3})\\)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.191, + 0.907, + 0.356 + ], + "angle": 0, + "content": "end training with a diffusion model, our coarse-to-fine approach shows better performance, indicating that our coarse stage serves as a strong control prior. This control prior is both neat and effective for extending our model to diverse identities, various body parts. Based on the comparison with \"Ours-background\" we notice that encoding information from the background image aids in enhancing natural illumination during background harmonization. Lastly, leveraging a guided refinement enables the preservation of high-frequency information alongside robust generation capabilities." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.358, + 0.906, + 0.464 + ], + "angle": 0, + "content": "For our temporal module, we study three ablation studies: 1) Ours: We eliminate all temporal consistency components, which is a single-frame-based generation method. 2) Ours+temporal: We only apply temporal module \\(\\mathcal{E}_m\\) without recurrent feature blending during the test-time. 3) Ours+temporal+blend: We perform our video relighting with temporal module \\(\\mathcal{E}_m\\) and recurrent feature blending." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.465, + 0.906, + 0.6 + ], + "angle": 0, + "content": "Tab. 4 summarizes the performance of each of our temporal modules. The temporal lighting module in \"Ours+temporal\" primarily enforces temporal coherence by imposing a temporal constraint on the lighting control between the current and previous frames during testing. Additionally, the recurrent blending feature further enhances temporal consistency by blending the lighting control feature between previous and current frames, thereby reinforcing the temporal context." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.601, + 0.905, + 0.647 + ], + "angle": 0, + "content": "Limitation. Our relighting diffusion model requires heavy computational time. Significant noise on the detection (e.g., mask and surface normal) affects the temporal coherence." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.664, + 0.634, + 0.68 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.907, + 0.9 + ], + "angle": 0, + "content": "We introduce a method for Comprehensive Relighting that is generalizable and consistent for monocular human relighting and harmonization. We address a core dataset challenge by utilizing a large and general image prior from a pre-trained diffusion model; and repurposing the model specialized for temporally consistent image relighting. For coherent control of the lighting, we introduce a coarse-to-fine relighting framework; and combine it with an external temporal lighting module that learns many real videos. Our guided refinement network enhances the visual to preserve the fine details of an original image. In the experiments, our method outperforms other image-based relighting and harmonization models in terms of quality and temporal coherence." + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.511, + 0.956 + ], + "angle": 0, + "content": "387" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.271, + 0.108 + ], + "angle": 0, + "content": "6. Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.115, + 0.484, + 0.177 + ], + "angle": 0, + "content": "We sincerely thank Mengwei Ren for the insightful discussions regarding the framework design, and we are grateful to Jianming Zhang for kindly providing the human normal map estimator." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.189, + 0.188, + 0.205 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.213, + 0.484, + 0.284 + ], + "angle": 0, + "content": "[1] Sai Bi, Zexiang Xu, Kalyan Sunkavalli, David Kriegman, and Ravi Ramamoorthi. Deep 3d capture: Geometry and reflectance from sparse multi-view images. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5960-5969, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.284, + 0.484, + 0.352 + ], + "angle": 0, + "content": "[2] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18392-18402, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.354, + 0.483, + 0.422 + ], + "angle": 0, + "content": "[3] Ziqi Cai, Kaiwen Jiang, Shu-Yu Chen, Yu-Kun Lai, Hongbo Fu, Boxin Shi, and Lin Gao. Real-time 3d-aware portrait video relighting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6221-6231, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.424, + 0.484, + 0.48 + ], + "angle": 0, + "content": "[4] Jianqi Chen, Yilan Zhang, Zhengxia Zou, Keyan Chen, and Zhenwei Shi. Dense pixel-to-pixel harmonization via continuous image representation. IEEE Transactions on Circuits and Systems for Video Technology, pages 1-1, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.481, + 0.484, + 0.522 + ], + "angle": 0, + "content": "[5] Zhaoxi Chen and Ziwei Liu. Relighting4d: Neural relightable human from videos. In European Conference on Computer Vision, pages 606-623. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.524, + 0.484, + 0.579 + ], + "angle": 0, + "content": "[6] Jun Myeong Choi, Max Christman, and Roni Sengupta. Personalized video relighting with an at-home light stage. In European Conference on Computer Vision, pages 394-410. Springer, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.58, + 0.484, + 0.648 + ], + "angle": 0, + "content": "[7] Wenyan Cong, Jianfu Zhang, Li Niu, Liu Liu, Zhixin Ling, Weiyuan Li, and Liqing Zhang. Dovenet: Deep image harmonization via domain verification. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8394-8403, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.65, + 0.484, + 0.718 + ], + "angle": 0, + "content": "[8] Wenyan Cong, Xinhao Tao, Li Niu, Jing Liang, Xuesong Gao, Qihao Sun, and Liqing Zhang. High-resolution image harmonization via collaborative dual transformations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18470-18479, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.72, + 0.484, + 0.775 + ], + "angle": 0, + "content": "[9] Yuki Endo Daichi Tajima, Yoshihiro Kanamori. Relighting humans in the wild: Monocular full-body human relighting with domain adaptation. Computer Graphics Forum (Proc. of Pacific Graphics 2021), 40(7):205-216, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.776, + 0.484, + 0.844 + ], + "angle": 0, + "content": "[10] David Futschik, Kelvin Ritland, James Vecore, Sean Fanello, Sergio Orts-Escalano, Brian Curless, Daniel Sykora, and Rohit Pandey. Controllable light diffusion for portraits. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8412-8421, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[11] Leon A Gatys, Alexander S Ecker, and Matthias Bethge. Image style transfer using convolutional neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2414-2423, 2016. 5" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.213, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.135 + ], + "angle": 0, + "content": "[12] Michal Geyer, Omer Bar-Tal, Shai Bagon, and Tali Dekel. Tokenflow: Consistent diffusion features for consistent video editing. arXiv preprint arxiv:2307.10373, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.136, + 0.908, + 0.205 + ], + "angle": 0, + "content": "[13] Julian Jorge Andrade Guerreiro, Mitsuru Nakazawa, and Björn Stenger. Pct-net: Full resolution image harmonization using pixel-wise color transformations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5917-5926, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.207, + 0.907, + 0.262 + ], + "angle": 0, + "content": "[14] Zonghui Guo, Dongsheng Guo, Haiyong Zheng, Zhaorui Gu, Bing Zheng, and Junyu Dong. Image harmonization with transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 14870-14879, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.263, + 0.906, + 0.319 + ], + "angle": 0, + "content": "[15] Zonghui Guo, Haiyong Zheng, Yufeng Jiang, Zhaorui Gu, and Bing Zheng. Intrinsic image harmonization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16367-16376, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.321, + 0.906, + 0.375 + ], + "angle": 0, + "content": "[16] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.377, + 0.906, + 0.418 + ], + "angle": 0, + "content": "[17] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. arXiv:2204.03458, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.42, + 0.906, + 0.462 + ], + "angle": 0, + "content": "[18] Alain Hore and Djemel Ziou. Image quality metrics: Psnr vs. ssm. In 2010 20th international conference on pattern recognition, pages 2366-2369. IEEE, 2010. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.463, + 0.906, + 0.532 + ], + "angle": 0, + "content": "[19] Andrew Hou, Ze Zhang, Michel Sarkis, Ning Bi, Yiying Tong, and Xiaoming Liu. Towards high fidelity face relighting with realistic shadows. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14719-14728, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.534, + 0.906, + 0.601 + ], + "angle": 0, + "content": "[20] Andrew Hou, Michel Sarkis, Ning Bi, Yiying Tong, and Xiaoming Liu. Face relighting with geometrically consistent shadows. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4217-4226, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.604, + 0.906, + 0.673 + ], + "angle": 0, + "content": "[21] Yasamin Jafarian and Hyun Soo Park. Learning high fidelity depths of dressed humans by watching social media dance videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12753-12762, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.676, + 0.906, + 0.731 + ], + "angle": 0, + "content": "[22] Chaonan Ji, Tao Yu, Kaiwen Guo, Jingxin Liu, and Yebin Liu. Geometry-aware single-image full-body human relighting. In European Conference on Computer Vision, pages 388-405. Springer, 2022. 2, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.732, + 0.906, + 0.815 + ], + "angle": 0, + "content": "[23] Yifan Jiang, He Zhang, Jianming Zhang, Yilin Wang, Zhe Lin, Kalyan Sunkavalli, Simon Chen, Sohrab Amirghodsi, Sarah Kong, and Zhangyang Wang. Ssh: A self-supervised framework for image harmonization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4832-4841, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.817, + 0.906, + 0.858 + ], + "angle": 0, + "content": "[24] Yoshihiro Kanamori and Yuki Endo. Relighting humans: occlusion-aware inverse rendering for full-body human images. arXiv preprint arXiv:1908.02714, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[25] Hoon Kim, Minje Jang, Wonjun Yoon, Jisoo Lee, Donghyun Na, and Sanghyun Woo. Switchlight: Co-design of physics-driven architecture and pre-training framework for human" + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.093, + 0.908, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.511, + 0.956 + ], + "angle": 0, + "content": "388" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.123, + 0.092, + 0.484, + 0.133 + ], + "angle": 0, + "content": "portrait relighting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25096-25106, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.134, + 0.484, + 0.189 + ], + "angle": 0, + "content": "[26] Manuel Lagunas, Xin Sun, Jimei Yang, Ruben Villegas, Jianming Zhang, Zhixin Shu, Belen Masia, and Diego Gutierrez. Single-image full-body human relighting. arXiv preprint arXiv:2107.07259, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.19, + 0.484, + 0.258 + ], + "angle": 0, + "content": "[27] Chenyang Lei, Xuanchi Ren, Zhaoxiang Zhang, and Qifeng Chen. Blind video deflickering by neural filtering with a flawed atlas. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.26, + 0.483, + 0.314 + ], + "angle": 0, + "content": "[28] Guannan Li, Yebin Liu, and Qionghai Dai. Free-viewpoint video relighting from multi-view sequence under general illumination. Machine vision and applications, 25:1737-1746, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.316, + 0.483, + 0.357 + ], + "angle": 0, + "content": "[29] Ruilong Li, Shan Yang, David A. Ross, and Angjoo Kanazawa. Learn to dance with aist++: Music conditioned 3d dance generation, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.358, + 0.483, + 0.426 + ], + "angle": 0, + "content": "[30] Yang Liu, Alexandros Neophytou, Sunando Sengupta, and Eric Sommerlade. Relighting images in the wild with a self-supervised siamese auto-encoder. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 32–40, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.427, + 0.483, + 0.496 + ], + "angle": 0, + "content": "[31] Ziwei Liu, Ping Luo, Shi Qiu, Xiaogang Wang, and Xiaou Tang. Deepfashion: Powering robust clothes recognition and retrieval with rich annotations. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1096-1104, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.497, + 0.483, + 0.564 + ], + "angle": 0, + "content": "[32] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.567, + 0.483, + 0.635 + ], + "angle": 0, + "content": "[33] Rohit Pandey, Sergio Orts Escolano, Chloe Legendre, Christian Haene, Sofien Bouaziz, Christoph Rhemann, Paul Debevec, and Sean Fanello. Total relighting: learning to relight portraits for background replacement. ACM Transactions on Graphics (TOG), 40(4):1-21, 2021. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.636, + 0.483, + 0.691 + ], + "angle": 0, + "content": "[34] F. Perazzi, J. Pont-Tuset, B. McWilliams, L. Van Gool, M. Gross, and A. Sorkine-Hornung. A benchmark dataset and evaluation methodology for video object segmentation. In Computer Vision and Pattern Recognition, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.692, + 0.384, + 0.705 + ], + "angle": 0, + "content": "[35] Pexels: https://www.pexels.com.7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.707, + 0.483, + 0.747 + ], + "angle": 0, + "content": "[36] Puntawat Ponglertnapakorn, Nontawat Tritrong, and Supasorn Suwajanakorn. Difareli: Diffusion face relighting. arXiv preprint arXiv:2304.09479, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.748, + 0.483, + 0.802 + ], + "angle": 0, + "content": "[37] Ravi Ramamoorthi and Pat Hanrahan. On the relationship between radiance and irradiance: determining the illumination from images of a convex lambertian object. JOSA A, 18 (10):2448-2459, 2001. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.804, + 0.483, + 0.859 + ], + "angle": 0, + "content": "[38] Mengwei Ren, Wei Xiong, Jae Shin Yoon, Zhixin Shu, Jianming Zhang, HyunJoon Jung, Guido Gerig, and He Zhang. Relightful harmonization: Lighting-aware portrait background replacement. CVPR, 2024. 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.86, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[39] Christian Richardt, Carsten Stoll, Neil A Dodgson, Hans-Peter Seidel, and Christian Theobalt. Coherent spatiotemporal filtering, upsampling and rendering of rgbz videos. In" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "Computer graphics forum, pages 247-256. Wiley Online Library, 2012. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.121, + 0.905, + 0.162 + ], + "angle": 0, + "content": "[40] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.163, + 0.905, + 0.204 + ], + "angle": 0, + "content": "[41] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.205, + 0.905, + 0.247 + ], + "angle": 0, + "content": "[42] Volker Schönefeld. Spherical harmonics. Computer Graphics and Multimedia Group, Technical Note. RWTH Aachen University, Germany, page 18, 2005. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.248, + 0.905, + 0.314 + ], + "angle": 0, + "content": "[43] Qing Shuai, Chen Geng, Qi Fang, Sida Peng, Wenhao Shen, Xiaowei Zhou, and Hujun Bao. Novel view synthesis of human interactions from sparse multi-view videos. In ACM SIGGRAPH 2022 Conference Proceedings, pages 1-10, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.316, + 0.905, + 0.357 + ], + "angle": 0, + "content": "[44] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.359, + 0.905, + 0.4 + ], + "angle": 0, + "content": "[45] Shuangbing Song, Fan Zhong, Tianju Wang, Xueying Qin, and Changhe Tu. Guided linear upsampling. ACM Transactions on Graphics (TOG), 42(4):1-12, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.401, + 0.905, + 0.469 + ], + "angle": 0, + "content": "[46] Tiancheng Sun, Jonathan T Barron, Yun-Ta Tsai, Zexiang Xu, Xueming Yu, Graham Fyffe, Christoph Rhemann, Jay Busch, Paul Debevec, and Ravi Ramamoorthi. Single image portrait relighting. ACM Transactions on Graphics (TOG), 38(4):1-12, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.47, + 0.905, + 0.537 + ], + "angle": 0, + "content": "[47] Daichi Tajima, Yoshihiro Kanamori, and Yuki Endo. Relighting humans in the wild: Monocular full-body human relighting with domain adaptation. In Computer Graphics Forum, pages 205-216. Wiley Online Library, 2021. 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.539, + 0.905, + 0.607 + ], + "angle": 0, + "content": "[48] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part II 16, pages 402–419. Springer, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.609, + 0.905, + 0.692 + ], + "angle": 0, + "content": "[49] Ayush Tewari, Tae-Hyun Oh, Tim Weyrich, Bernd Bickel, Hans-Peter Seidel, Hanspeter Pfister, Wojciech Matusik, Mohamed Elgharib, Christian Theobalt, et al. Monocular reconstruction of neural face reflectance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4791-4800, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.692, + 0.905, + 0.775 + ], + "angle": 0, + "content": "[50] Marco Toschi, Riccardo De Matteo, Riccardo Spezialetti, Daniele De Gregorio, Luigi Di Stefano, and Samuele Salti. Relight my nerf: A dataset for novel view synthesis and relighting of real world objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20762-20772, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.776, + 0.905, + 0.831 + ], + "angle": 0, + "content": "[51] Yi-Hsuan Tsai, Xiaohui Shen, Zhe Lin, Kalyan Sunkavalli, Xin Lu, and Ming-Hsuan Yang. Deep image harmonization. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3789-3797, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.832, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[52] Ke Wang, Michael Gharbi, He Zhang, Zhihao Xia, and Eli Shechtman. Semi-supervised parametric real-world image harmonization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5927-5936, 2023. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.511, + 0.956 + ], + "angle": 0, + "content": "389" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.16 + ], + "angle": 0, + "content": "[53] Yuxin Wang, Wayne Wu, and Dan Xu. Learning unified decompositional and compositional nerf for editable novel view synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 18247-18256, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.164, + 0.482, + 0.233 + ], + "angle": 0, + "content": "[54] Liu Wen, Zhixin Piao, Min Jie, Wenhan Luo, Lin Ma, and Shenghua Gao. Liquid warping gan: A unified framework for human motion imitation, appearance transfer and novel view synthesis. In The IEEE International Conference on Computer Vision (ICCV), 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.234, + 0.482, + 0.289 + ], + "angle": 0, + "content": "[55] Zexiang Xu, Kalyan Sunkavalli, Sunil Hadap, and Ravi Ramamoorthi. Deep image-based relighting from optimal sparse samples. ACM Transactions on Graphics (ToG), 37 (4):1-13, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.291, + 0.482, + 0.346 + ], + "angle": 0, + "content": "[56] Zexiang Xu, Sai Bi, Kalyan Sunkavalli, Sunil Hadap, Hao Su, and Ravi Ramamoorthi. Deep view synthesis from sparse photometric images. ACM Transactions on Graphics (ToG), 38(4):1-13, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.348, + 0.482, + 0.416 + ], + "angle": 0, + "content": "[57] Wenhan Yang, Jiashi Feng, Jianchao Yang, Fang Zhao, Jiaying Liu, Zongming Guo, and Shuicheng Yan. Deep edge guided recurrent residual learning for image superresolution. IEEE Transactions on Image Processing, 26(12): 5895-5907, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.419, + 0.482, + 0.46 + ], + "angle": 0, + "content": "[58] Mona Zehni, Shaona Ghosh, Krishna Sridhar, and Sethu Ram. Joint learning of portrait intrinsic decomposition and relighting. arXiv preprint arXiv:2106.15305, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.462, + 0.482, + 0.516 + ], + "angle": 0, + "content": "[59] Hao Zhang, Yanbo Xu, Tianyuan Dai, Tai Chi-Keung Tang, et al. Fdnerf: Semantics-driven face reconstruction, prompt editing and relighting with diffusion models. arXiv preprint arXiv:2306.00783, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.519, + 0.482, + 0.587 + ], + "angle": 0, + "content": "[60] Jiakai Zhang, Xinhang Liu, Xinyi Ye, Fuqiang Zhao, Yanshun Zhang, Minye Wu, Yingliang Zhang, Lan Xu, and Jingyi Yu. Editable free-viewpoint video using a layered neural representation. ACM Transactions on Graphics (TOG), 40(4):1-18, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.589, + 0.482, + 0.657 + ], + "angle": 0, + "content": "[61] Longwen Zhang, Qixuan Zhang, Minye Wu, Jingyi Yu, and Lan Xu. Neural video portrait relighting in real-time via consistency modeling. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 802-812, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.66, + 0.482, + 0.7 + ], + "angle": 0, + "content": "[62] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models, 2023. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.703, + 0.482, + 0.771 + ], + "angle": 0, + "content": "[63] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Scaling in-the-wild training for diffusion-based illumination harmonization and editing by imposing consistent light transport. In The Thirteenth International Conference on Learning Representations, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.774, + 0.482, + 0.814 + ], + "angle": 0, + "content": "[64] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.817, + 0.482, + 0.872 + ], + "angle": 0, + "content": "[65] Yulun Zhang, Yapeng Tian, Yu Kong, Bineng Zhong, and Yun Fu. Residual dense network for image super-resolution. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2472-2481, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.874, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[66] Ruichen Zheng, Peng Li, Haoqian Wang, and Tao Yu. Learning visibility field for detailed 3d human reconstruction and" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.133 + ], + "angle": 0, + "content": "relighting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 216-226, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.177 + ], + "angle": 0, + "content": "[67] Hao Zhou, Sunil Hadap, Kalyan Sunkavalli, and David W. Jacobs. Deep single portrait image relighting. In International Conference on Computer Vision (ICCV), 2019. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.905, + 0.233 + ], + "angle": 0, + "content": "[68] Hao Zhou, Sunil Hadap, Kalyan Sunkavalli, and David W Jacobs. Deep single-image portrait relighting. In Proceedings of the IEEE/CVF international conference on computer vision, pages 7194-7202, 2019. 1, 2, 3, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.235, + 0.905, + 0.317 + ], + "angle": 0, + "content": "[69] Taotao Zhou, Kai He, Di Wu, Teng Xu, Qixuan Zhang, Kuixiang Shao, Wenzheng Chen, Lan Xu, and Jingyi Yu. Relightable neural human assets from multi-view gradient illuminations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4315-4327, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.32, + 0.905, + 0.388 + ], + "angle": 0, + "content": "[70] Jun-Yan Zhu, Philipp Krahenbuhl, Eli Shechtman, and Alexei A Efros. Learning a discriminative model for the perception of realism in composite images. In Proceedings of the IEEE International Conference on Computer Vision, pages 3943-3951, 2015. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.388 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "390" + } + ] +] \ No newline at end of file diff --git a/2025/Comprehensive Relighting_ Generalizable and Consistent Monocular Human Relighting and Harmonization/743d9653-d63f-4670-9d09-7ac226ec0ec0_origin.pdf b/2025/Comprehensive Relighting_ Generalizable and Consistent Monocular Human Relighting and Harmonization/743d9653-d63f-4670-9d09-7ac226ec0ec0_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..25a4409d074366fb73b0c73ec5475a595af53758 --- /dev/null +++ b/2025/Comprehensive Relighting_ Generalizable and Consistent Monocular Human Relighting and Harmonization/743d9653-d63f-4670-9d09-7ac226ec0ec0_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:11669270b63c474d8d989a1bd8943a4b01f79b9e5659a39f62c7e6f5895ac7cf +size 7940107 diff --git a/2025/Comprehensive Relighting_ Generalizable and Consistent Monocular Human Relighting and Harmonization/full.md b/2025/Comprehensive Relighting_ Generalizable and Consistent Monocular Human Relighting and Harmonization/full.md new file mode 100644 index 0000000000000000000000000000000000000000..97d3715e58147b455f2d1ca382e45995e2e54d66 --- /dev/null +++ b/2025/Comprehensive Relighting_ Generalizable and Consistent Monocular Human Relighting and Harmonization/full.md @@ -0,0 +1,362 @@ +# Comprehensive Relighting: Generalizable and Consistent Monocular Human Relighting and Harmonization + +Junying Wang $^{1\dagger}$ +He Zhang $^{2}$ + +Jingyuan Liu $^{2}$ Xin Sun $^{2}$ +Jimei Yang $^{3}$ Nanxuan Zhao $^{2}$ +Ulrich Neumann $^{1}$ + +Krishna Kumar Singh $^{2}$ +Tuanfeng Y. Wang $^{2}$ +Jae Shin Yoon $^{2}$ + +Zhixin Shu $^{2}$ Simon S. Chen $^{2}$ + +1University of Southern California + +$^{2}$ Adobe Research $^{3}$ Runway + +![](images/8dd337d08a12c9efe08a0823d7c32f13a31e8819a563f5ca3fcf80b740fff7af.jpg) +Input video frames + +![](images/27887390fb4995f4a5d613f54ae41568b941439a845cfd0e6f98bc52bbcb2e81.jpg) + +![](images/a548fd2351b3218e9a979d227ba28925c5fff657d77b9ecbca58bf78db9622ff.jpg) +Relighting + +![](images/eabe1ba7bde618f8df382f27afa04b8e1f73ac57bbb8dff2d164220326e6c695.jpg) + +![](images/f1b99f966c73ed20f0c035d951661e7ee036a1d9dd866023175a387f7ba23b9b.jpg) +Harmonization + +![](images/62fa0a00326ffb4ac312b563e5bfd7ba542ef0e24606ad9eb709e6d40073422e.jpg) +Relighting + Harmonization +Figure 1. We introduce Comprehensive Relighting, a generalizable and consistent model for relighting and harmonization, which controls the lighting property from a single image or video of humans with arbitrary body parts. Given target lighting coefficients, e.g., Spherical harmonics (second), background scenes (third), or their combination (fourth), our model performs consistent and harmonized relighting. + +![](images/336493c3ac5e682645dc0b9725d3d227ef1f7e2153b275a0a7101b894aef4fa4.jpg) +Input video frames + +![](images/6f2690fd754444b3504e9c0341334a7a6e84ac0c5f81aabb7c0265f1dd63b36e.jpg) + +![](images/0953e00d7aa2cdb35d8e64ffbafb0f95e9327ec41a00fdf1964f6efede9f8823.jpg) +Relighting + +![](images/54bbd218878ca13188a6574832a00bdb7be77d427051eba1cf4d5386e167238c.jpg) + +![](images/2390ff1bd22080f0244cb8cdfd16801794cf6edb83a8803bd71f3f894a8b9376.jpg) +Harmonization + +![](images/141f83c8036a2bbf8df58e3a845db2ea5543a3065bb3ccdebba2b3c5e602b922.jpg) + +![](images/4fd1ab893d395f4059c550849092f0e6ab7d722e3ef641fb422edcf4aa6d0add.jpg) +Relighting + Harmonization + +# Abstract + +This paper introduces Comprehensive Relighting, the first all-in-one approach that can both control and harmonize the lighting from an image or video of humans with arbitrary body parts from any scene. Building such a generalizable model is extremely challenging due to the lack of dataset, restricting existing image-based relighting models to a specific scenario (e.g., face or static human). To address this challenge, we repurpose a pre-trained diffusion model as a general image prior and jointly model the human relighting and background harmonization in the coarse-to-fine framework. To further enhance the temporal coherence of the relighting, we introduce an unsupervised temporal lighting model that learns the lighting cycle consistency from many real-world videos without any ground truth. In inference time, our temporal lighting module is combined with the diffusion models through the spatio-temporal feature blending algorithms without extra training; and we apply a new guided refinement as a post-processing to pre + +serve the high-frequency details from the input image. In the experiments, Comprehensive Relighting shows a strong generalizability and lighting temporal coherence, outperforming existing image-based human relighting and harmonization methods. + +# 1. Introduction + +Light is the key component that determines how a person looks, which is often defined by a specific time and space, where revisiting such a unique moment gives us the opportunity to experience authentic telepresence sensations. In this paper, we introduce a generalizable human relighting model that can control the lighting from an image or video of humans with arbitrary body parts (Fig. 3), which are well-harmonized with a conditioning space (i.e., background image) as shown in Fig. 1. + +As shown in Fig. 2, existing image-based relighting methods face two main problems, lack of generalizability and controllability. First, they are designed for a specific scenario, e.g., only for face illumination or static humans [33, 61, 68] mainly due to the scarcity of large-scale + +![](images/ebd58bb97fc109e3f772bac1266c9459db804e4aaf6c07c7a8462971b961dd67.jpg) +Figure 2. Comparison of various baseline methods for relighting settings and functionalities. + +relighting datasets: Precise acquisition of the appearance of a static person under assorted lighting conditions requires specialized setups such as LightStage [22, 33] or expensive graphics simulation [26], which are not scalable, particularly for video contents. Learning from such limited data induces weak generalization of the model to diverse scenes. Second, most relighting algorithms struggle to effectively model multiple light sources. Typically, these algorithms are restricted to a single lighting control from either background image (e.g. high dynamic range lighting environment map [22]) or target lighting parameters (e.g. Spherical harmonics [24]). These problems inhibit the production-level application that requires general use cases. + +To overcome these challenges, we propose an effective approach to achieve all-in-one relighting by utilizing a diffusion model—a general image prior that learns massive visual data with diverse lighting conditions; and repurpose this prior specialized for a human relighting and harmonization in a coarse-to-fine framework: A pretrained latent diffusion model [41] learns from limited datasets of static humans to jointly perform the fine-grained relighting and harmonization from two multi-modal lighting variables: the coarse shading estimate and conditioning background scene. The coarse shading estimate can be "computed" from conditioning lighting parameters (i.e., Spherical harmonics) without a neural network, and therefore, it is generalizable. The diffusion model is required to learn only the residual portion (e.g., fine self-occluded shadow), which is more generalizable than direct relighting (e.g., [68]). The pre-trained image prior helps with understanding the properties of the general background scenes, enabling the generalization of the background harmonization. + +Our coarse-to-fine framework, however, still introduces problems: the diffusion model that learns to generate an image without temporal context produces significant temporal artifacts (e.g., sudden changes of lighting distribution even for consecutive frames). We address this problem by introducing an unsupervised temporal lighting model that learns from many real videos without any ground-truth data to en + +![](images/5d7ffff5401d1b0ae3c69cf2e14429bf4d6713475641b686688a1500756b4919.jpg) +Figure 3. Our model generalizes to various body parts (portrait, half-body, full-body, multiperson) for relighting and harmonization, with lighting control variables shown in the insets. + +force the temporal lighting consistency over frames. This temporal module learns the unsupervised cycle consistency between the relit and many real videos to predict the future lighting distribution from the past, which can be directly combined with our coarse-to-fine relighting components without extra training. + +In inference time, we further push the temporal coherence of the relighting and harmonization by formulating a recurrent prediction pipeline. The generation at the current time step is conditioned onto the temporal module for the one at the next time. The features from the lighting and temporal control modules are spatially and temporally blended to enforce temporal coherence while improving the structure of the lighting distribution. Finally, our guided refinement module enhances the quality of the generated image in a way that preserves the original high-frequency details of the input image. + +In the experiments, Comprehensive Relighting demonstrates high-quality relighting and background harmonization results with strong temporal coherence across the lighting, background, and pose changes. It also shows strong generalizability to any unconstrained scenes regardless of body parts, views, and poses, outperforming existing relighting and background harmonization methods. + +Our contributions include: (1) To the best of our knowledge, the first approach for joint modeling of relighting and background harmonization; (2) a novel coarse-to-fine framework that enables comprehensive generalization by only learning from limited synthetic and lab-controlled data; (3) unsupervised temporal modeling with lighting cycle consistency from many unconstrained real videos; (4) an effective inference algorithm with spatio-temporal feature blending and guided refinement. + +# 2. Related Work + +Image-based Human Relighting While high-quality relighting often requires resource-intensive setups, recent efforts in mobile device relighting aim to address single + +image scenarios. Total Relighting [33] achieves photorealistic effects by leveraging detailed normal maps and albedo as priors. However, it relies on an HDR environment map that accurately matches the scene's real-world lighting, which is not always readily available, particularly in flexible and personalized relighting scenarios. [37, 46] estimate and adjust the Spherical harmonics parameters from images, and others use one-light-at-a-time (OLAT) captures to generate relighting data [49] or estimate reflectance fields [46]. + +Recent works have focused on diverse relighting scenarios including portrait [10, 19, 20, 25, 30, 33, 46, 58, 67, 68], full-body [9, 22, 26, 47], and object relighting [1, 55, 56]. Most of these works rely on decomposing an image into its intrinsic components, i.e., albedo, normal, and lighting, and therefore, the accuracy of this decomposition directly impacts the quality of the final relighting. While some works [22, 66, 69] use single-image geometry reconstruction for traditional and neural rendering, reconstruction errors are often propagated to relighting results. Tajima et al. [47] tackled domain adaptation with a two-step relighting framework, yet noticeable texture distortion remains due to limited model generalizability. Zhang et al. [59] trained a 2D latent-diffusion model, allowing users to manipulate and construct face NeRFs in a zero-shot learning framework without the need for explicit 3D data. DiFaRei [36] utilizes DDIM (Denoising Diffusion Implicit Models [44]) for high-fidelity face relighting. While promising, their methods are constrained by a focus on either specific body parts, limiting their applicability for generalizable. + +Background Harmonization Background harmonization seeks to harmonize color, contrast, and style discrepancies between the foreground and background, ensuring composite images appear natural and cohesive. Many existing background harmonization methods [4, 7, 8, 13-15, 23, 51, 52, 70] formulate this problem as image-to-image translation work where a neural network translates an unharmonized foreground image to the harmonized one in the context of the conditioning background image. Recently, Relightful Harmonization [38] introduced methods that harmonize both image style and lighting to match the background scene, and IC-Light [63] further enables flexible illumination control via image diffusion models, guided by text descriptions or background images. While promising, these methods lack explicit lighting control, restricting general applicability and consistent video relighting. Additionally, some ([38]) are limited to specific body parts (e.g., portraits), with constrained support for full-body and multiperson scenarios. + +Video Relighting Neural Radiance Fields (NeRF) [32] based methods enable novel view synthesis under varying lighting conditions in videos [43, 50, 53, 69]. Zhang et al. [61] achieve portrait video relighting under dynamic illuminations, while Choi et al. [6] ensure temporally consis + +tent relit videos. 3D-PVR [3] present a 3D-aware, real-time method to relight videos of talking faces. Relighting4D [5] decomposes the time-varied human body as a set of neural fields of normal, occlusion, diffuse, and specular maps. However, their rendering quality are highly reliant on geometry accuracy. Some works [36] adopt a temporal modeling scheme real-time neural video portrait relighting. STNeRF [60] controls dynamic scenes using a neural layered radiance representation that maintains spatial and temporal coherence. Li et al. [28] employed multi-view reconstruction for free-viewpoint relighting videos under general illumination. Richardt et al. [39] use RGBZ video cameras for video effects, including relighting, but rely on multi-view reconstruction from specialized hardware, making the process highly complex and costly. + +# 3. Method + +We develop a generalizable and consistent human relighting and harmonization framework using a diffusion model. Fig. 4 illustrates the overview of our framework. Given an input image of humans and control lighting variables, including coarse shading and background image, a diffusion model learns to generate a fine-grained image of the humans under the target lighting, also harmonized with the background (Sec. 3.2). An external temporal lighting module, trained on real videos using unsupervised temporal cycle consistency, is integrated into the diffusion model to enhance temporal lighting coherence (Sec. 3.3). In inference time, we blend the features between the lighting and temporal control modules over time to ensure the relighting results are accurate and temporally coherent, and we apply a guided refinement to prevent the loss of high-frequency details during the denoising process (Sec. 3.4). + +# 3.1. Background: Image-based Relighting + +Image-based relighting function can be compactly modeled with a small number of approximated basis of Spherical harmonics (SH) [9, 26, 42] which describes only essential features of illumination on the surface of a 3D sphere based on the following formulation: + +$$ +I (\mathbf {x}) = \rho (\mathbf {x}) \cdot \sum_ {l = 0} ^ {k} \sum_ {m = - l} ^ {l} \phi Y _ {l m} (\mathbf {n} (\mathbf {x})) \tag {1} +$$ + +where $\phi \in \mathbb{R}^{(k + 1)^2}$ denotes the spherical harmonics (SH) coefficients; $Y_{lm}(\mathbf{n})$ are the SH basis functions evaluated at the surface normal $\mathbf{n}$ ; indices $l$ and $m$ represent the band and order within each band, respectively; and $k$ is the maximum order of spherical harmonics used. While SH is computationally efficient and highly generalizable, it only captures low-frequency lighting, limiting its ability to model fine-grained, high-frequency illumination details. Addition- + +![](images/fe0e59f14ac5883b7b71a881062be5b17f97a96e7a2f25f3bfcc420174600209.jpg) +(a) Coarse-to-fine relighting and harmonization + +![](images/2638a0955625b65dd185371f2d6fbc9c937a91808ec41174a9f3c1006d95c397.jpg) +(b) Temporal modeling with cycle consistency +(c) Inference: Spatio-temporal feature blending +Figure 4. System overview. (a) Given an input image of humans with coarse lighting and background image, our diffusion model generates the relit images harmonized with background scenes (Sec. 3.2). (b) The external temporal modules learn the temporal cycle consistency from many real-world videos to construct temporal lighting features (Sec. 3.3). (c) In inference time, we blend the features from lighting and temporal modules spatially and temporally to enable coherent and generalizable human relighting (Sec. 3.4). + +![](images/22cf249a94d54a2f4bec6ba87273d378bbe3c1046817b7ee8231579a47af2826.jpg) + +ally, SH-based global illumination neglects the context provided by background images, often resulting in relit images appearing unnatural when composited with different backgrounds. To overcome these limitations, we propose a coarse-to-fine human relighting and harmonization framework that leverages the strong image prior available from a pre-trained text-to-image diffusion model. + +# 3.2. Coarse-to-Fine Relighting and Harmonization + +We generate a fine-grained relit image of a person conditioned on a coarse lighting representation: + +$$ +\mathcal {E} (\mathbf {I}; \mathbf {S} _ {\phi}) = z, \quad \mathcal {D} (z) = \mathbf {I} _ {\phi} \tag {2} +$$ + +where $\mathcal{E}$ is an encoder that generates the latent lighting features, $z$ as a function of the input image $\mathbf{I} \in \mathbb{R}^{w \times h \times 3}$ and a small number of global lighting parameters $\phi \in \mathbb{R}^n$ (i.e., Spherical harmonics where $n = 25$ ); and $\mathcal{D}$ is the decoder that generates fine-grained relit image $\mathbf{I}_{\phi} \in \mathbb{R}^{w \times h \times 3}$ from $z$ . $\mathbf{S}_{\phi} \in \mathbb{R}^{w \times h}$ is the pixel-aligned coarse shading map converted from the coarse lighting parameters $\phi$ as shown in Fig. 4-(a). One approach to obtaining $\mathbf{S}_{\phi}$ is directly computing it as demonstrated in Eq. 1 where we can compute lighting intensity along with the detected surface normal map from $\mathbf{I}$ given spherical harmonics coefficients $\phi$ . Another way, is to use a neural network to convert the surface normal to $\mathbf{S}_{\phi}$ as a condition of spherical harmonics coefficients $\phi$ , i.e., $\mathbf{S}_{\phi} \gets f(\mathbf{N}; \phi)$ where $f$ is the neural shading function which maps the surface normal $\mathbf{N}$ and $\phi$ to the coarse shading. While both approaches are highly generalizable, our experiments indicate that the latter method achieves improved accuracy and better noise correction. Please refer to the Supple. documents for more details about the coarse shading estimation. + +To jointly model the relighting and background harmonization, we further condition a target background image + +$\mathbf{B} \in \mathbb{R}^{w \times h \times 3}$ as the additional lighting sources: + +$$ +\mathcal {E} (\mathbf {I}; \{\mathbf {S} _ {\phi}, \mathbf {B} \}) = \boldsymbol {z}, \quad \mathcal {D} (\boldsymbol {z}) = \mathbf {I} _ {\phi} \tag {3} +$$ + +where the lighting encoder $\mathcal{E}$ learns to capture the intensity and direction of light from $\mathbf{S}_{\phi}$ while capturing ambient environment lighting and color distribution from $\mathbf{B}$ . This enables $\mathcal{D}$ to achieve complete relighting in scenarios involving new target lighting, background, or both. In the subsection, we enable $(\mathcal{D} \circ \mathcal{E})$ using a latent diffusion model. + +# 3.2.1. Fine-grained Relighting Diffusion Model + +We enable the fine-grained image relighting using a diffusion model as illustrated in Fig. 4-(a). Our encoder $\mathcal{E}$ in Eq. 3, in practice, is formulated as the composition of two encoders: + +$$ +\mathcal {E} \rightarrow \mathcal {E} _ {\mathrm {b}} (\mathbf {I}; \mathcal {E} _ {1} (\{\mathbf {S} _ {\phi}, \mathbf {B} \})) = \boldsymbol {z}, \quad \mathcal {D} (\boldsymbol {z}) = \mathbf {I} _ {\phi} \tag {4} +$$ + +where $\mathcal{E}_1$ encodes lighting control variables, $\{\mathbf{S}_{\phi}, \mathbf{B}\}$ , and $\mathcal{E}_b$ encodes the conditional variable $\mathbf{I}$ whose visual properties, e.g., semantics and identity, should be preserved in the output along with the controls from $\mathcal{E}_1$ . For $\mathcal{E}_b$ , we use the base latent diffusion model [2, 40] pre-trained for text-to-image generation task, and for $\mathcal{E}_1$ , we use ControlNet [62] (termed as Light ControlNet in Fig. 4-(a)). While noise, texts, and timestep variables are also conditioned on $\mathcal{E}_b$ to fit the modality of the latent diffusion model, they are not described in the equations and figures for conciseness. To impose the foreground attention on the lighting control, a portrait mask is also conditioned to $\mathcal{E}_l$ : + +$$ +\mathcal {E} _ {\mathrm {b}} (\mathbf {I}; \mathcal {E} _ {1} (\{\mathbf {S} _ {\phi}, \mathbf {B} \}; \mathbf {M})) = \boldsymbol {z}, \quad \mathcal {D} (\boldsymbol {z}) = \mathbf {I} _ {\phi} \tag {5} +$$ + +where $\mathbf{M} \in \{0,1\}^{w \times h}$ is the foreground binary mask. In training time, the diffusion model learns to directly change the lighting distribution of the input images (without inverse + +rendering techniques) as the condition of coarse lighting estimate by minimizing the latent distance of the noisy relit image with the clean one from the ground truth in the overall forward and background denoising steps. Following the findings from an existing harmonization work [38], we use the composite image between the input image and conditioning background image as $\mathbf{I}$ . We randomly drop background $\mathbf{B}$ or randomly set the coarse shading $\mathbf{S}_{\phi}$ as binary mask such that the diffusion model learns the control the lighting from $\mathbf{B}$ , $\mathbf{S}_{\phi}$ , or both $\{\mathbf{B}, \mathbf{S}_{\phi}\}$ . + +# 3.3. Unsupervised Add-on Temporal Modeling + +Our coarse-to-fine relighting framework $(\mathcal{D} \circ \mathcal{E}_{\mathrm{b}} \circ \mathcal{E}_{\mathrm{l}})$ is trained only on individual images, inherently missing temporal context, e.g., how a point on a human's surface will radiate from a specific viewpoint under the continuous pose, view, and illumination changes, leading to temporal artifacts such as flickering. We model such temporal context by designing an external add-on temporal lighting module $\mathcal{E}_{m}$ that can be combined, in inference time, with the relighting framework without extra training: + +$$ +\mathcal {D} \circ \mathcal {E} _ {\mathrm {b}} \circ \mathcal {E} _ {\mathrm {l}} \rightarrow \mathcal {D} \circ \mathcal {E} _ {\mathrm {b}} \circ \left(\mathcal {E} _ {\mathrm {l}} \times \mathcal {E} _ {\mathrm {m}}\right) \tag {6} +$$ + +To enable this, our temporal module is designed to regress the relit image from the previous time instance, i.e., $\mathbf{I}_{\phi}^{t - 1}$ , to the latent lighting distribution whose space is shared with our base relighting models. Therefore, our decoder can generate temporally coherent relit images in the current time in an auto-regressive way: + +$$ +\mathcal {D} \left(\mathcal {E} _ {\mathrm {b}} \left(\mathbf {I} ^ {t}; \mathcal {E} _ {\mathrm {m}} \left(\mathbf {I} _ {\phi} ^ {t - 1}\right)\right)\right) = \mathbf {I} _ {t - 1} ^ {t}. \tag {7} +$$ + +However, training $\mathcal{E}_{\mathrm{m}}$ with a conventional L2 loss is not possible since there exists no ground-truth video relighting data for dynamic humans. Therefore, we propose to learn $\mathcal{E}_{\mathrm{m}}$ in an unsupervised way using many real videos with lighting cycle consistency as follows: + +$$ +\mathcal {D} ^ {*} \left(\mathcal {E} _ {\mathrm {b}} ^ {*} \left(\mathbf {I} ^ {t}; \mathcal {E} _ {1} ^ {*} \left(\left\{\mathbf {S} _ {\phi} ^ {t}, \mathbf {B} ^ {t} \right\}; \mathbf {M} ^ {t}\right)\right)\right) = \mathbf {I} _ {\phi} ^ {t}, \tag {8} +$$ + +$$ +\mathcal {D} ^ {*} \left(\mathcal {E} _ {\mathrm {b}} ^ {*} \left(\mathbf {I} _ {\phi} ^ {t}; \mathcal {E} _ {\mathrm {m}} \left(\mathbf {I} ^ {t - 1}, \mathbf {M} ^ {t - 1}\right)\right)\right) = \tilde {\mathbf {I}} _ {t - 1} ^ {t} \tag {9} +$$ + +$$ +\therefore \quad \mathbf {I} ^ {t} = \tilde {\mathbf {I}} _ {t - 1} ^ {t}. +$$ + +We make the hypothesis: a video sequence inherently contains temporal lighting properties whose flow can be implicitly modeled by learning to predict the lighting distribution of the future frame based on the hint from the previous frame. This involves forward and backward processes, as a cycle-training. Eq. 8 represents the forward image relighting, i.e., $\mathbf{I}^t\rightarrow \mathbf{I}_\phi^t$ , where our relighting diffusion model with lighting ControlNet $\mathcal{E}_1$ generates the relit image at time $t$ as a condition of a novel lighting condition (where we pick random Spherical harmonics). Eq. 9 reverts the relighting, i.e., $\tilde{\mathbf{I}}_{t - 1}^{t}\gets \mathbf{I}_{\phi}^{t}$ to the original input image conditioned by + +the original frame in the previous time step through our temporal lighting module where the mask $\mathbf{M}$ is used for foreground awareness. Finally, $\mathcal{E}_m$ learns the lighting cycle consistency in the diffusion process by minimizing the latent distance between the reverted and the original image. We freeze the learnable weights for the functions with $*$ during training, and $\mathcal{E}_{\mathrm{m}}$ is enabled with another ControlNet [62] termed as Motion ControlNet in Fig. 4-(b). + +# 3.4. Inference with Spatio-Temporal Blending and Guided Refinement + +Given a video or image, we perform comprehensive relighting with a spatio-temporal feature blending framework. For $t = 0$ , we generate the relit image without our temporal lighting module $\mathcal{E}_{\mathrm{m}}$ . For $t > 0$ (note that even for the static image, $t > 0$ is possible since the lighting is time-variant), the generated relit images in the previous time step are conditioned on $\mathcal{E}_{\mathrm{m}}$ , and therefore, the relighting is controlled by dual control modules, i.e., $\mathcal{E}_{\mathrm{l}}$ and $\mathcal{E}_{\mathrm{m}}$ by blending their features with a (0.85:0.15) ratio as described as spatial blending block in Fig. 4-(c). The blended lighting features are recurrently combined with the one from the previous time step through the temporal blending block. For this, we adopt optical flow from the original input video as a temporal prior to spatially align the features from consecutive frames similar to [27]; and we blend the aligned temporal features with a (0.5:0.5) ratio as in Fig. 4-(c) to improve the temporal consistency. + +In the denoising process of the latent diffusion model, the generation often suffers from the loss of high-frequency details. Inspired by existing image restoration techniques [45, 65], we address this problem by applying a guided refinement to each generated image. We cast the problem of guided refinement as a guided residual prediction: + +$$ +\mathbf {I} _ {\phi} ^ {\text {r e f i n e}} = \mathbf {I} + g \left(\mathbf {I} _ {\phi}, \mathbf {I}; \mathbf {M}\right), \tag {10} +$$ + +where $g$ is the function that predicts the guided lighting residual. This residual learns to map the lighting distribution from $\mathbf{I}$ to $\mathbf{I}_{\phi}$ . Here, $\mathbf{I}_{\phi}^{\mathrm{refine}}$ can effectively preserve the high-frequency details of the input image $\mathbf{I}$ due to the nature of residual learning [57, 65] that is designed to preserve the visual properties from the observation space, i.e., $\mathbf{I}$ . We enable $g$ with a residual network [16] by learning from our relighting data with general losses for low-level vision processing, i.e., L2 and VGG [11]. + +# 4. Experiments + +Dataset. To train our coarse-to-fine relighting model, we use 100K ground-truth relighting samples, including 50K synthetic human renderings and 50K OLAT-captured images from LightStage, with random cropping augmentation. Ground-truth albedo, images, background, masks, and lighting coefficients are precomputed. Our dataset is categorized by gender, skin tone, and body part (full-body and + +
MethodScenario 1Scenario 2Scenario 3
DPR [68]18.62/0.86/0.10332.00 /0.94/0.03018.14/0.89/0.08935.29 /0.98/0.03221.20/0.89/0.07237.11/0.95/0.038
RHW [47]19.78/0.87/0.11330.74/0.95/0.02720.12/0.88/0.07836.64 /0.98/0.02823.33/0.90/0.06037.53 /0.98/0.033
GFR [22]25.59/0.91/0.08930.33/0.95/0.03322.76/0.91/0.07232.36/0.98/0.03625.49/0.93/0.05035.89/0.98/0.028
LPBR [38]18.19/0.86/0.09031.62/0.91/0.03519.96/0.88/0.08427.94/0.94/0.04121.42/0.88/0.05333.39/0.95/0.038
Ours25.95/0.95/0.06633.58/0.96/0.02423.99/0.93/0.04835.18/0.99 / 0.03126.61/0.94/0.03338.32/0.98/0.026
+ +Table 1. Comparison with existing image-based human relighting methods on synthetic videos for fidelity and temporal consistency. Each column shows image fidelity (left) and video temporal consistency (right). Metrics are PSNR↑/SSIM↑/LPIPS↓ for accuracy, and tPSNR↑/tSSIM↑/tLPIPS↓ for temporal consistency. While green is used for the best values, yellow highlights the second-best values. + +![](images/12983553ec971990dd47a5b51513ba04934a5b976d430d60799d97d1e8e9dddf.jpg) +Figure 5. Qualitative comparison of synthetic video frames (corresponding to Tab. 1). From left to right: composite input with target lighting parameters (inset), our relit result, baseline methods, and normalized L2 $\downarrow$ photometric error map (inset). + +multi-body), with each subject captured from 32 viewpoints under varying lighting conditions. For a detailed dataset breakdown, refer to the Supple. To train our temporal control module using real data, we used 50K frames of videos from existing works [21, 29, 34, 54] and customized videos, where, for cyclic relighting, we randomly sample spherical harmonics lighting parameters from ground-truth data. For validation on static humans with dynamic lighting, we generated a new testing set (i.e., the ground-truth relit images and associated lighting parameters) using an internal simulation and LightStage. This set includes the mixture of half-body, full-body, and multi-person scenarios, with each scenario comprising 100 frames. For the validation on dynamic humans (i.e., video), we newly create synthetic video sequences for three scenarios: Scenario 1 involves static lighting (environment map) across frames with a moving human; Scenario 2 features dynamic lighting (rotated environment map) with a static human; Scenario 3 combines dynamic lighting (rotated environment map) with a moving human. Please refer to the Supple. and demo video for more details about the testing results. + +Metrics. To measure the relighting quality, we use L1 dis + +![](images/85944e6a87caa418279f9ec9e779d08aa4fcb45ddddfadcc5d1b991850428972.jpg) +Figure 6. Comparison with LPBR [38] on DeepFashion [31] real images for background harmonization testing. The first row shows the relit output, and the second shows the magnified results. + +![](images/f59d0190d8986b659b4caa768b947d68ac252014b6c64d256e25e8097dce0972.jpg) + +![](images/6cbc539142dee20d8d1392b23984a030b5768a668e205251d755ba33b270dcb1.jpg) + +tance, reconstruction fidelity (PSNR [18]), local structure similarity (SSIM [18]), and the latent perceptual distance (LPIPS [64]), between the relighted and ground-truth. For measuring the temporal coherence of the relighting results, we follow the same logics as TokenFlow [12]. This involves warping the relit image from the previous time step to the next using optical flow [48], and then comparing this to the + +![](images/06e95269460463cacb9206d2bda8252d73c29b08b029e2013da7c088437f7b9b.jpg) +Figure 7. Comparison with RHW [47] on Pexels [35] real images. The lighting control variables are shown as insets. While RHW produces reasonable relighting for full-body images, its quality degrades on half-body and multi-person cases. + +![](images/b4d21c460ee17f7fcbabd8828264f16dbe0336eb88963c9e1f6c40a8b705974d.jpg) +Figure 8. Comparison with GFR [22] on Pexels [35] real images. The lighting control variables are shown as insets. Limited generalizability of GFR results in reduced output quality for half-body and multi-person cases. + +current frame using the aforementioned metrics, termed $tL_{1}$ error, tPSNR, tSSIM, and tLPIPS. + +Baseline. We compare our method with existing monocular human relighting works where we chose the baselines that are applicable to general scenes (e.g., any part-specific information such as a 3D face model is not a requirement): DPR [67], RHW [47], and GFR [22]. For the harmonization baseline, we compare our model with LPBR [38], a diffusion-based light-aware harmonization method. For relighting, aside from GFR, baselines use Spherical harmonics for lighting control without modeling background illumination. DPR and RHW are evaluated using their released + +pre-trained models. Since GFR lacks available code, we re-implemented it using our dataset, replacing HDR lighting with Spherical harmonics and background modeling to enable background harmonization in our experiments. Harmonization is directly compared with LPBR by replacing the background. + +Results. For testing static humans under dynamic lighting, we present the quantitative comparison in Tab. 2. We show average numerical evaluations on our synthetic testing dataset, categorized by portrait, full-body, and multiperson. Further validation by gender, and skin color is detailed in the Supple. Methods such as RHW and DPR have + +
MethodL1↓PSNR↑SSIM↑
Ours-diffusion0.0583717.0990.833
Ours end-to-end0.0143226.4180.918
Ours-background0.0123927.3460.945
Ours0.0103528.4190.948
Ours+refine0.0101228.7780.949
+ +limited generalizability for both full-body and portrait relighting. They are difficult to extend to other scenarios and tend to show reduced performance when tested in different settings. While GFR performs well in our evaluation, it struggles with significant domain gaps, leading to noticeable quality degradation on real data, including distortions and color shifts (Fig. 8). In contrast, our model exhibits strong generalizability across validation sets and real-world tests (Figs. 7, 8). Compared to the state-of-the-art background harmonization method (LPBR), our method shows the strong generalization to different body parts, and notably, LPBR often includes significant distortion when applied to the full-body images and it does not support the lighting control function as shown in Fig. 6. + +Table 3. Ablation study on coarse-to-fine relighting models. + +
CategoryDPR [68]RHW [47]GFR [22]Ours
Portrait17.74 / 0.8715.75 / 0.8217.71 / 0.8623.04 / 0.90
Full-body27.62 / 0.9627.73 / 0.9529.51 / 0.9530.81 / 0.97
Multi-person25.70 / 0.9525.69 / 0.9529.35 / 0.9731.49 / 0.96
+ +For video testing, in Tab. 1, we evaluate our model on scenarios 1, 2, and 3 for both fidelity and temporal consistency. Other approaches face challenges in achieving both relit fidelity and temporal consistency at the same time as also shown in Fig. 5. In contrast, our temporal module ensures our comprehensive relighting model to produce videos with strong temporal consistency. For more results on relighting results and comparisons including user study, please refer to the Supple. + +Ablation Study. We conduct two ablation studies on our coarse-to-fine model using static human data and our temporal modules using dynamic human data. As shown in Tab. 3, we perform the ablation study on our coarse-to-fine approach: 1) Ours-diffusion: Relighting only with pixel-aligned coarse shading without diffusion model, i.e., $\mathbf{S}_{\phi} \times \mathbf{I}$ . 2) Ours end-to-end: Relighting in an end-to-end manner by applying target lighting parameters $\phi$ as a condition, trained with a diffusion model, without the coarse stage. 3) Ours-background: Relighting without the control of background B. 4) Ours+refine (full): Applying our guided refinement to the relit image. + +Tab. 3 shows the summary of the ablation study: Directly apply the coarse stage without a diffusion model introduces significant errors in the final relit result due to the detection noises. Instead of applying target lighting and end-to + +Table 2. Comparison on our synthetic static testing data sorted by body-part. We compute average PSNR↑ / SSIM↑. + +
MethodtL1↓tPSNR↑tSSIM↑
Ours6.55231.0280.956
Ours+temporal5.63832.2660.957
Ours+temporal+blend4.01933.5880.957
+ +Table 4. Ablation study on our temporal modules evaluated on synthetic sequences: tL1 error $(\times 10^{-3})$ + +end training with a diffusion model, our coarse-to-fine approach shows better performance, indicating that our coarse stage serves as a strong control prior. This control prior is both neat and effective for extending our model to diverse identities, various body parts. Based on the comparison with "Ours-background" we notice that encoding information from the background image aids in enhancing natural illumination during background harmonization. Lastly, leveraging a guided refinement enables the preservation of high-frequency information alongside robust generation capabilities. + +For our temporal module, we study three ablation studies: 1) Ours: We eliminate all temporal consistency components, which is a single-frame-based generation method. 2) Ours+temporal: We only apply temporal module $\mathcal{E}_m$ without recurrent feature blending during the test-time. 3) Ours+temporal+blend: We perform our video relighting with temporal module $\mathcal{E}_m$ and recurrent feature blending. + +Tab. 4 summarizes the performance of each of our temporal modules. The temporal lighting module in "Ours+temporal" primarily enforces temporal coherence by imposing a temporal constraint on the lighting control between the current and previous frames during testing. Additionally, the recurrent blending feature further enhances temporal consistency by blending the lighting control feature between previous and current frames, thereby reinforcing the temporal context. + +Limitation. Our relighting diffusion model requires heavy computational time. Significant noise on the detection (e.g., mask and surface normal) affects the temporal coherence. + +# 5. Conclusion + +We introduce a method for Comprehensive Relighting that is generalizable and consistent for monocular human relighting and harmonization. We address a core dataset challenge by utilizing a large and general image prior from a pre-trained diffusion model; and repurposing the model specialized for temporally consistent image relighting. For coherent control of the lighting, we introduce a coarse-to-fine relighting framework; and combine it with an external temporal lighting module that learns many real videos. Our guided refinement network enhances the visual to preserve the fine details of an original image. In the experiments, our method outperforms other image-based relighting and harmonization models in terms of quality and temporal coherence. + +# 6. Acknowledgement + +We sincerely thank Mengwei Ren for the insightful discussions regarding the framework design, and we are grateful to Jianming Zhang for kindly providing the human normal map estimator. + +# References + +[1] Sai Bi, Zexiang Xu, Kalyan Sunkavalli, David Kriegman, and Ravi Ramamoorthi. Deep 3d capture: Geometry and reflectance from sparse multi-view images. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5960-5969, 2020. 3 +[2] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18392-18402, 2023. 4 +[3] Ziqi Cai, Kaiwen Jiang, Shu-Yu Chen, Yu-Kun Lai, Hongbo Fu, Boxin Shi, and Lin Gao. Real-time 3d-aware portrait video relighting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6221-6231, 2024. 3 +[4] Jianqi Chen, Yilan Zhang, Zhengxia Zou, Keyan Chen, and Zhenwei Shi. Dense pixel-to-pixel harmonization via continuous image representation. IEEE Transactions on Circuits and Systems for Video Technology, pages 1-1, 2023. 3 +[5] Zhaoxi Chen and Ziwei Liu. Relighting4d: Neural relightable human from videos. In European Conference on Computer Vision, pages 606-623. Springer, 2022. 3 +[6] Jun Myeong Choi, Max Christman, and Roni Sengupta. Personalized video relighting with an at-home light stage. In European Conference on Computer Vision, pages 394-410. Springer, 2024. 3 +[7] Wenyan Cong, Jianfu Zhang, Li Niu, Liu Liu, Zhixin Ling, Weiyuan Li, and Liqing Zhang. Dovenet: Deep image harmonization via domain verification. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8394-8403, 2020. 3 +[8] Wenyan Cong, Xinhao Tao, Li Niu, Jing Liang, Xuesong Gao, Qihao Sun, and Liqing Zhang. High-resolution image harmonization via collaborative dual transformations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18470-18479, 2022. 3 +[9] Yuki Endo Daichi Tajima, Yoshihiro Kanamori. Relighting humans in the wild: Monocular full-body human relighting with domain adaptation. Computer Graphics Forum (Proc. of Pacific Graphics 2021), 40(7):205-216, 2021. 3 +[10] David Futschik, Kelvin Ritland, James Vecore, Sean Fanello, Sergio Orts-Escalano, Brian Curless, Daniel Sykora, and Rohit Pandey. Controllable light diffusion for portraits. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8412-8421, 2023. 3 +[11] Leon A Gatys, Alexander S Ecker, and Matthias Bethge. Image style transfer using convolutional neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2414-2423, 2016. 5 + +[12] Michal Geyer, Omer Bar-Tal, Shai Bagon, and Tali Dekel. Tokenflow: Consistent diffusion features for consistent video editing. arXiv preprint arxiv:2307.10373, 2023. 6 +[13] Julian Jorge Andrade Guerreiro, Mitsuru Nakazawa, and Björn Stenger. Pct-net: Full resolution image harmonization using pixel-wise color transformations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5917-5926, 2023. 3 +[14] Zonghui Guo, Dongsheng Guo, Haiyong Zheng, Zhaorui Gu, Bing Zheng, and Junyu Dong. Image harmonization with transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 14870-14879, 2021. +[15] Zonghui Guo, Haiyong Zheng, Yufeng Jiang, Zhaorui Gu, and Bing Zheng. Intrinsic image harmonization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16367-16376, 2021. 3 +[16] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 5 +[17] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. arXiv:2204.03458, 2022. 2 +[18] Alain Hore and Djemel Ziou. Image quality metrics: Psnr vs. ssm. In 2010 20th international conference on pattern recognition, pages 2366-2369. IEEE, 2010. 6 +[19] Andrew Hou, Ze Zhang, Michel Sarkis, Ning Bi, Yiying Tong, and Xiaoming Liu. Towards high fidelity face relighting with realistic shadows. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14719-14728, 2021. 3 +[20] Andrew Hou, Michel Sarkis, Ning Bi, Yiying Tong, and Xiaoming Liu. Face relighting with geometrically consistent shadows. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4217-4226, 2022. 3 +[21] Yasamin Jafarian and Hyun Soo Park. Learning high fidelity depths of dressed humans by watching social media dance videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12753-12762, 2021. 6 +[22] Chaonan Ji, Tao Yu, Kaiwen Guo, Jingxin Liu, and Yebin Liu. Geometry-aware single-image full-body human relighting. In European Conference on Computer Vision, pages 388-405. Springer, 2022. 2, 3, 6, 7, 8 +[23] Yifan Jiang, He Zhang, Jianming Zhang, Yilin Wang, Zhe Lin, Kalyan Sunkavalli, Simon Chen, Sohrab Amirghodsi, Sarah Kong, and Zhangyang Wang. Ssh: A self-supervised framework for image harmonization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4832-4841, 2021. 3 +[24] Yoshihiro Kanamori and Yuki Endo. Relighting humans: occlusion-aware inverse rendering for full-body human images. arXiv preprint arXiv:1908.02714, 2019. 2 +[25] Hoon Kim, Minje Jang, Wonjun Yoon, Jisoo Lee, Donghyun Na, and Sanghyun Woo. Switchlight: Co-design of physics-driven architecture and pre-training framework for human + +portrait relighting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25096-25106, 2024. 3 +[26] Manuel Lagunas, Xin Sun, Jimei Yang, Ruben Villegas, Jianming Zhang, Zhixin Shu, Belen Masia, and Diego Gutierrez. Single-image full-body human relighting. arXiv preprint arXiv:2107.07259, 2021. 2, 3 +[27] Chenyang Lei, Xuanchi Ren, Zhaoxiang Zhang, and Qifeng Chen. Blind video deflickering by neural filtering with a flawed atlas. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 5 +[28] Guannan Li, Yebin Liu, and Qionghai Dai. Free-viewpoint video relighting from multi-view sequence under general illumination. Machine vision and applications, 25:1737-1746, 2014. 3 +[29] Ruilong Li, Shan Yang, David A. Ross, and Angjoo Kanazawa. Learn to dance with aist++: Music conditioned 3d dance generation, 2021. 6 +[30] Yang Liu, Alexandros Neophytou, Sunando Sengupta, and Eric Sommerlade. Relighting images in the wild with a self-supervised siamese auto-encoder. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 32–40, 2021. 3 +[31] Ziwei Liu, Ping Luo, Shi Qiu, Xiaogang Wang, and Xiaou Tang. Deepfashion: Powering robust clothes recognition and retrieval with rich annotations. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1096-1104, 2016. 6 +[32] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 3 +[33] Rohit Pandey, Sergio Orts Escolano, Chloe Legendre, Christian Haene, Sofien Bouaziz, Christoph Rhemann, Paul Debevec, and Sean Fanello. Total relighting: learning to relight portraits for background replacement. ACM Transactions on Graphics (TOG), 40(4):1-21, 2021. 1, 2, 3 +[34] F. Perazzi, J. Pont-Tuset, B. McWilliams, L. Van Gool, M. Gross, and A. Sorkine-Hornung. A benchmark dataset and evaluation methodology for video object segmentation. In Computer Vision and Pattern Recognition, 2016. 6 +[35] Pexels: https://www.pexels.com.7 +[36] Puntawat Ponglertnapakorn, Nontawat Tritrong, and Supasorn Suwajanakorn. Difareli: Diffusion face relighting. arXiv preprint arXiv:2304.09479, 2023. 3 +[37] Ravi Ramamoorthi and Pat Hanrahan. On the relationship between radiance and irradiance: determining the illumination from images of a convex lambertian object. JOSA A, 18 (10):2448-2459, 2001. 3 +[38] Mengwei Ren, Wei Xiong, Jae Shin Yoon, Zhixin Shu, Jianming Zhang, HyunJoon Jung, Guido Gerig, and He Zhang. Relightful harmonization: Lighting-aware portrait background replacement. CVPR, 2024. 3, 5, 6, 7 +[39] Christian Richardt, Carsten Stoll, Neil A Dodgson, Hans-Peter Seidel, and Christian Theobalt. Coherent spatiotemporal filtering, upsampling and rendering of rgbz videos. In + +Computer graphics forum, pages 247-256. Wiley Online Library, 2012. 3 +[40] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models, 2021. 4 +[41] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 2 +[42] Volker Schönefeld. Spherical harmonics. Computer Graphics and Multimedia Group, Technical Note. RWTH Aachen University, Germany, page 18, 2005. 3 +[43] Qing Shuai, Chen Geng, Qi Fang, Sida Peng, Wenhao Shen, Xiaowei Zhou, and Hujun Bao. Novel view synthesis of human interactions from sparse multi-view videos. In ACM SIGGRAPH 2022 Conference Proceedings, pages 1-10, 2022. 3 +[44] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020.3 +[45] Shuangbing Song, Fan Zhong, Tianju Wang, Xueying Qin, and Changhe Tu. Guided linear upsampling. ACM Transactions on Graphics (TOG), 42(4):1-12, 2023. 5 +[46] Tiancheng Sun, Jonathan T Barron, Yun-Ta Tsai, Zexiang Xu, Xueming Yu, Graham Fyffe, Christoph Rhemann, Jay Busch, Paul Debevec, and Ravi Ramamoorthi. Single image portrait relighting. ACM Transactions on Graphics (TOG), 38(4):1-12, 2019. 3 +[47] Daichi Tajima, Yoshihiro Kanamori, and Yuki Endo. Relighting humans in the wild: Monocular full-body human relighting with domain adaptation. In Computer Graphics Forum, pages 205-216. Wiley Online Library, 2021. 3, 6, 7, 8 +[48] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part II 16, pages 402–419. Springer, 2020. 6 +[49] Ayush Tewari, Tae-Hyun Oh, Tim Weyrich, Bernd Bickel, Hans-Peter Seidel, Hanspeter Pfister, Wojciech Matusik, Mohamed Elgharib, Christian Theobalt, et al. Monocular reconstruction of neural face reflectance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4791-4800, 2021. 3 +[50] Marco Toschi, Riccardo De Matteo, Riccardo Spezialetti, Daniele De Gregorio, Luigi Di Stefano, and Samuele Salti. Relight my nerf: A dataset for novel view synthesis and relighting of real world objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20762-20772, 2023. 3 +[51] Yi-Hsuan Tsai, Xiaohui Shen, Zhe Lin, Kalyan Sunkavalli, Xin Lu, and Ming-Hsuan Yang. Deep image harmonization. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3789-3797, 2017. 3 +[52] Ke Wang, Michael Gharbi, He Zhang, Zhihao Xia, and Eli Shechtman. Semi-supervised parametric real-world image harmonization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5927-5936, 2023. 3 + +[53] Yuxin Wang, Wayne Wu, and Dan Xu. Learning unified decompositional and compositional nerf for editable novel view synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 18247-18256, 2023. 3 +[54] Liu Wen, Zhixin Piao, Min Jie, Wenhan Luo, Lin Ma, and Shenghua Gao. Liquid warping gan: A unified framework for human motion imitation, appearance transfer and novel view synthesis. In The IEEE International Conference on Computer Vision (ICCV), 2019. 6 +[55] Zexiang Xu, Kalyan Sunkavalli, Sunil Hadap, and Ravi Ramamoorthi. Deep image-based relighting from optimal sparse samples. ACM Transactions on Graphics (ToG), 37 (4):1-13, 2018. 3 +[56] Zexiang Xu, Sai Bi, Kalyan Sunkavalli, Sunil Hadap, Hao Su, and Ravi Ramamoorthi. Deep view synthesis from sparse photometric images. ACM Transactions on Graphics (ToG), 38(4):1-13, 2019. 3 +[57] Wenhan Yang, Jiashi Feng, Jianchao Yang, Fang Zhao, Jiaying Liu, Zongming Guo, and Shuicheng Yan. Deep edge guided recurrent residual learning for image superresolution. IEEE Transactions on Image Processing, 26(12): 5895-5907, 2017. 5 +[58] Mona Zehni, Shaona Ghosh, Krishna Sridhar, and Sethu Ram. Joint learning of portrait intrinsic decomposition and relighting. arXiv preprint arXiv:2106.15305, 2021. 3 +[59] Hao Zhang, Yanbo Xu, Tianyuan Dai, Tai Chi-Keung Tang, et al. Fdnerf: Semantics-driven face reconstruction, prompt editing and relighting with diffusion models. arXiv preprint arXiv:2306.00783, 2023. 3 +[60] Jiakai Zhang, Xinhang Liu, Xinyi Ye, Fuqiang Zhao, Yanshun Zhang, Minye Wu, Yingliang Zhang, Lan Xu, and Jingyi Yu. Editable free-viewpoint video using a layered neural representation. ACM Transactions on Graphics (TOG), 40(4):1-18, 2021. 3 +[61] Longwen Zhang, Qixuan Zhang, Minye Wu, Jingyi Yu, and Lan Xu. Neural video portrait relighting in real-time via consistency modeling. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 802-812, 2021. 1, 3 +[62] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models, 2023. 4, 5 +[63] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Scaling in-the-wild training for diffusion-based illumination harmonization and editing by imposing consistent light transport. In The Thirteenth International Conference on Learning Representations, 2025. 3 +[64] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 6 +[65] Yulun Zhang, Yapeng Tian, Yu Kong, Bineng Zhong, and Yun Fu. Residual dense network for image super-resolution. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2472-2481, 2018. 5 +[66] Ruichen Zheng, Peng Li, Haoqian Wang, and Tao Yu. Learning visibility field for detailed 3d human reconstruction and + +relighting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 216-226, 2023. 3 +[67] Hao Zhou, Sunil Hadap, Kalyan Sunkavalli, and David W. Jacobs. Deep single portrait image relighting. In International Conference on Computer Vision (ICCV), 2019. 3, 7 +[68] Hao Zhou, Sunil Hadap, Kalyan Sunkavalli, and David W Jacobs. Deep single-image portrait relighting. In Proceedings of the IEEE/CVF international conference on computer vision, pages 7194-7202, 2019. 1, 2, 3, 6, 8 +[69] Taotao Zhou, Kai He, Di Wu, Teng Xu, Qixuan Zhang, Kuixiang Shao, Wenzheng Chen, Lan Xu, and Jingyi Yu. Relightable neural human assets from multi-view gradient illuminations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4315-4327, 2023. 3 +[70] Jun-Yan Zhu, Philipp Krahenbuhl, Eli Shechtman, and Alexei A Efros. Learning a discriminative model for the perception of realism in composite images. In Proceedings of the IEEE International Conference on Computer Vision, pages 3943-3951, 2015. 3 \ No newline at end of file diff --git a/2025/Comprehensive Relighting_ Generalizable and Consistent Monocular Human Relighting and Harmonization/images.zip b/2025/Comprehensive Relighting_ Generalizable and Consistent Monocular Human Relighting and Harmonization/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..11827b4f5c4b85a939a9f6ef00437b453fa6ae46 --- /dev/null +++ b/2025/Comprehensive Relighting_ Generalizable and Consistent Monocular Human Relighting and Harmonization/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d136c5e870f2cec25d63a2ddbcd3d02b77261dab37003622f0d643a98ffec977 +size 913534 diff --git a/2025/Comprehensive Relighting_ Generalizable and Consistent Monocular Human Relighting and Harmonization/layout.json b/2025/Comprehensive Relighting_ Generalizable and Consistent Monocular Human Relighting and Harmonization/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..6a8f67783b6f4c48e62363ced0a4504968562710 --- /dev/null +++ b/2025/Comprehensive Relighting_ Generalizable and Consistent Monocular Human Relighting and Harmonization/layout.json @@ -0,0 +1,9870 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 58, + 103, + 555, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 103, + 555, + 140 + ], + "spans": [ + { + "bbox": [ + 58, + 103, + 555, + 140 + ], + "type": "text", + "content": "Comprehensive Relighting: Generalizable and Consistent Monocular Human Relighting and Harmonization" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 70, + 159, + 150, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 159, + 150, + 190 + ], + "spans": [ + { + "bbox": [ + 70, + 159, + 150, + 190 + ], + "type": "text", + "content": "Junying Wang" + }, + { + "bbox": [ + 70, + 159, + 150, + 190 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 70, + 159, + 150, + 190 + ], + "type": "text", + "content": " \nHe Zhang" + }, + { + "bbox": [ + 70, + 159, + 150, + 190 + ], + "type": "inline_equation", + "content": "^{2}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 151, + 160, + 317, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 160, + 317, + 202 + ], + "spans": [ + { + "bbox": [ + 151, + 160, + 317, + 202 + ], + "type": "text", + "content": "Jingyuan Liu" + }, + { + "bbox": [ + 151, + 160, + 317, + 202 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 151, + 160, + 317, + 202 + ], + "type": "text", + "content": " Xin Sun" + }, + { + "bbox": [ + 151, + 160, + 317, + 202 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 151, + 160, + 317, + 202 + ], + "type": "text", + "content": " \nJimei Yang" + }, + { + "bbox": [ + 151, + 160, + 317, + 202 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 151, + 160, + 317, + 202 + ], + "type": "text", + "content": " Nanxuan Zhao" + }, + { + "bbox": [ + 151, + 160, + 317, + 202 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 151, + 160, + 317, + 202 + ], + "type": "text", + "content": " \nUlrich Neumann" + }, + { + "bbox": [ + 151, + 160, + 317, + 202 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 322, + 160, + 453, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 160, + 453, + 202 + ], + "spans": [ + { + "bbox": [ + 322, + 160, + 453, + 202 + ], + "type": "text", + "content": "Krishna Kumar Singh" + }, + { + "bbox": [ + 322, + 160, + 453, + 202 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 322, + 160, + 453, + 202 + ], + "type": "text", + "content": " \nTuanfeng Y. Wang" + }, + { + "bbox": [ + 322, + 160, + 453, + 202 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 322, + 160, + 453, + 202 + ], + "type": "text", + "content": " \nJae Shin Yoon" + }, + { + "bbox": [ + 322, + 160, + 453, + 202 + ], + "type": "inline_equation", + "content": "^{2}" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 459, + 160, + 539, + 188 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 459, + 160, + 539, + 188 + ], + "spans": [ + { + "bbox": [ + 459, + 160, + 539, + 188 + ], + "type": "text", + "content": "Zhixin Shu" + }, + { + "bbox": [ + 459, + 160, + 539, + 188 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 459, + 160, + 539, + 188 + ], + "type": "text", + "content": " Simon S. Chen" + }, + { + "bbox": [ + 459, + 160, + 539, + 188 + ], + "type": "inline_equation", + "content": "^{2}" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 129, + 216, + 297, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 216, + 297, + 232 + ], + "spans": [ + { + "bbox": [ + 129, + 216, + 297, + 232 + ], + "type": "text", + "content": "1University of Southern California" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 321, + 217, + 478, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 217, + 478, + 232 + ], + "spans": [ + { + "bbox": [ + 321, + 217, + 478, + 232 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 321, + 217, + 478, + 232 + ], + "type": "text", + "content": "Adobe Research " + }, + { + "bbox": [ + 321, + 217, + 478, + 232 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 321, + 217, + 478, + 232 + ], + "type": "text", + "content": "Runway" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 60, + 251, + 110, + 393 + ], + "blocks": [ + { + "bbox": [ + 60, + 251, + 110, + 393 + ], + "lines": [ + { + "bbox": [ + 60, + 251, + 110, + 393 + ], + "spans": [ + { + "bbox": [ + 60, + 251, + 110, + 393 + ], + "type": "image", + "image_path": "8dd337d08a12c9efe08a0823d7c32f13a31e8819a563f5ca3fcf80b740fff7af.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 63, + 396, + 111, + 403 + ], + "lines": [ + { + "bbox": [ + 63, + 396, + 111, + 403 + ], + "spans": [ + { + "bbox": [ + 63, + 396, + 111, + 403 + ], + "type": "text", + "content": "Input video frames" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 112, + 251, + 162, + 323 + ], + "blocks": [ + { + "bbox": [ + 112, + 251, + 162, + 323 + ], + "lines": [ + { + "bbox": [ + 112, + 251, + 162, + 323 + ], + "spans": [ + { + "bbox": [ + 112, + 251, + 162, + 323 + ], + "type": "image", + "image_path": "27887390fb4995f4a5d613f54ae41568b941439a845cfd0e6f98bc52bbcb2e81.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 120, + 327, + 162, + 392 + ], + "blocks": [ + { + "bbox": [ + 120, + 327, + 162, + 392 + ], + "lines": [ + { + "bbox": [ + 120, + 327, + 162, + 392 + ], + "spans": [ + { + "bbox": [ + 120, + 327, + 162, + 392 + ], + "type": "image", + "image_path": "a548fd2351b3218e9a979d227ba28925c5fff657d77b9ecbca58bf78db9622ff.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 125, + 396, + 152, + 403 + ], + "lines": [ + { + "bbox": [ + 125, + 396, + 152, + 403 + ], + "spans": [ + { + "bbox": [ + 125, + 396, + 152, + 403 + ], + "type": "text", + "content": "Relighting" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 163, + 251, + 232, + 323 + ], + "blocks": [ + { + "bbox": [ + 163, + 251, + 232, + 323 + ], + "lines": [ + { + "bbox": [ + 163, + 251, + 232, + 323 + ], + "spans": [ + { + "bbox": [ + 163, + 251, + 232, + 323 + ], + "type": "image", + "image_path": "eabe1ba7bde618f8df382f27afa04b8e1f73ac57bbb8dff2d164220326e6c695.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 163, + 326, + 232, + 392 + ], + "blocks": [ + { + "bbox": [ + 163, + 326, + 232, + 392 + ], + "lines": [ + { + "bbox": [ + 163, + 326, + 232, + 392 + ], + "spans": [ + { + "bbox": [ + 163, + 326, + 232, + 392 + ], + "type": "image", + "image_path": "f1b99f966c73ed20f0c035d951661e7ee036a1d9dd866023175a387f7ba23b9b.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 180, + 396, + 219, + 403 + ], + "lines": [ + { + "bbox": [ + 180, + 396, + 219, + 403 + ], + "spans": [ + { + "bbox": [ + 180, + 396, + 219, + 403 + ], + "type": "text", + "content": "Harmonization" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 233, + 251, + 308, + 392 + ], + "blocks": [ + { + "bbox": [ + 233, + 251, + 308, + 392 + ], + "lines": [ + { + "bbox": [ + 233, + 251, + 308, + 392 + ], + "spans": [ + { + "bbox": [ + 233, + 251, + 308, + 392 + ], + "type": "image", + "image_path": "62fa0a00326ffb4ac312b563e5bfd7ba542ef0e24606ad9eb709e6d40073422e.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 235, + 396, + 304, + 403 + ], + "lines": [ + { + "bbox": [ + 235, + 396, + 304, + 403 + ], + "spans": [ + { + "bbox": [ + 235, + 396, + 304, + 403 + ], + "type": "text", + "content": "Relighting + Harmonization" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 408, + 555, + 441 + ], + "lines": [ + { + "bbox": [ + 55, + 408, + 555, + 441 + ], + "spans": [ + { + "bbox": [ + 55, + 408, + 555, + 441 + ], + "type": "text", + "content": "Figure 1. We introduce Comprehensive Relighting, a generalizable and consistent model for relighting and harmonization, which controls the lighting property from a single image or video of humans with arbitrary body parts. Given target lighting coefficients, e.g., Spherical harmonics (second), background scenes (third), or their combination (fourth), our model performs consistent and harmonized relighting." + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 309, + 251, + 359, + 392 + ], + "blocks": [ + { + "bbox": [ + 309, + 251, + 359, + 392 + ], + "lines": [ + { + "bbox": [ + 309, + 251, + 359, + 392 + ], + "spans": [ + { + "bbox": [ + 309, + 251, + 359, + 392 + ], + "type": "image", + "image_path": "336493c3ac5e682645dc0b9725d3d227ef1f7e2153b275a0a7101b894aef4fa4.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 307, + 396, + 356, + 403 + ], + "lines": [ + { + "bbox": [ + 307, + 396, + 356, + 403 + ], + "spans": [ + { + "bbox": [ + 307, + 396, + 356, + 403 + ], + "type": "text", + "content": "Input video frames" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 359, + 251, + 406, + 323 + ], + "blocks": [ + { + "bbox": [ + 359, + 251, + 406, + 323 + ], + "lines": [ + { + "bbox": [ + 359, + 251, + 406, + 323 + ], + "spans": [ + { + "bbox": [ + 359, + 251, + 406, + 323 + ], + "type": "image", + "image_path": "6f2690fd754444b3504e9c0341334a7a6e84ac0c5f81aabb7c0265f1dd63b36e.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 359, + 323, + 407, + 392 + ], + "blocks": [ + { + "bbox": [ + 359, + 323, + 407, + 392 + ], + "lines": [ + { + "bbox": [ + 359, + 323, + 407, + 392 + ], + "spans": [ + { + "bbox": [ + 359, + 323, + 407, + 392 + ], + "type": "image", + "image_path": "0953e00d7aa2cdb35d8e64ffbafb0f95e9327ec41a00fdf1964f6efede9f8823.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 372, + 396, + 399, + 403 + ], + "lines": [ + { + "bbox": [ + 372, + 396, + 399, + 403 + ], + "spans": [ + { + "bbox": [ + 372, + 396, + 399, + 403 + ], + "type": "text", + "content": "Relighting" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 408, + 251, + 477, + 323 + ], + "blocks": [ + { + "bbox": [ + 408, + 251, + 477, + 323 + ], + "lines": [ + { + "bbox": [ + 408, + 251, + 477, + 323 + ], + "spans": [ + { + "bbox": [ + 408, + 251, + 477, + 323 + ], + "type": "image", + "image_path": "54bbd218878ca13188a6574832a00bdb7be77d427051eba1cf4d5386e167238c.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 408, + 323, + 477, + 392 + ], + "blocks": [ + { + "bbox": [ + 408, + 323, + 477, + 392 + ], + "lines": [ + { + "bbox": [ + 408, + 323, + 477, + 392 + ], + "spans": [ + { + "bbox": [ + 408, + 323, + 477, + 392 + ], + "type": "image", + "image_path": "2390ff1bd22080f0244cb8cdfd16801794cf6edb83a8803bd71f3f894a8b9376.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 425, + 395, + 464, + 402 + ], + "lines": [ + { + "bbox": [ + 425, + 395, + 464, + 402 + ], + "spans": [ + { + "bbox": [ + 425, + 395, + 464, + 402 + ], + "type": "text", + "content": "Harmonization" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 478, + 251, + 550, + 323 + ], + "blocks": [ + { + "bbox": [ + 478, + 251, + 550, + 323 + ], + "lines": [ + { + "bbox": [ + 478, + 251, + 550, + 323 + ], + "spans": [ + { + "bbox": [ + 478, + 251, + 550, + 323 + ], + "type": "image", + "image_path": "141f83c8036a2bbf8df58e3a845db2ea5543a3065bb3ccdebba2b3c5e602b922.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 478, + 323, + 550, + 392 + ], + "blocks": [ + { + "bbox": [ + 478, + 323, + 550, + 392 + ], + "lines": [ + { + "bbox": [ + 478, + 323, + 550, + 392 + ], + "spans": [ + { + "bbox": [ + 478, + 323, + 550, + 392 + ], + "type": "image", + "image_path": "4fd1ab893d395f4059c550849092f0e6ab7d722e3ef641fb422edcf4aa6d0add.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 481, + 396, + 550, + 403 + ], + "lines": [ + { + "bbox": [ + 481, + 396, + 550, + 403 + ], + "spans": [ + { + "bbox": [ + 481, + 396, + 550, + 403 + ], + "type": "text", + "content": "Relighting + Harmonization" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 28 + }, + { + "bbox": [ + 151, + 451, + 200, + 464 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 451, + 200, + 464 + ], + "spans": [ + { + "bbox": [ + 151, + 451, + 200, + 464 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 53, + 476, + 297, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 476, + 297, + 693 + ], + "spans": [ + { + "bbox": [ + 53, + 476, + 297, + 693 + ], + "type": "text", + "content": "This paper introduces Comprehensive Relighting, the first all-in-one approach that can both control and harmonize the lighting from an image or video of humans with arbitrary body parts from any scene. Building such a generalizable model is extremely challenging due to the lack of dataset, restricting existing image-based relighting models to a specific scenario (e.g., face or static human). To address this challenge, we repurpose a pre-trained diffusion model as a general image prior and jointly model the human relighting and background harmonization in the coarse-to-fine framework. To further enhance the temporal coherence of the relighting, we introduce an unsupervised temporal lighting model that learns the lighting cycle consistency from many real-world videos without any ground truth. In inference time, our temporal lighting module is combined with the diffusion models through the spatio-temporal feature blending algorithms without extra training; and we apply a new guided refinement as a post-processing to pre" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 313, + 453, + 555, + 512 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 453, + 555, + 512 + ], + "spans": [ + { + "bbox": [ + 313, + 453, + 555, + 512 + ], + "type": "text", + "content": "serve the high-frequency details from the input image. In the experiments, Comprehensive Relighting shows a strong generalizability and lighting temporal coherence, outperforming existing image-based human relighting and harmonization methods." + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 314, + 522, + 395, + 534 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 522, + 395, + 534 + ], + "spans": [ + { + "bbox": [ + 314, + 522, + 395, + 534 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 312, + 544, + 555, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 544, + 555, + 652 + ], + "spans": [ + { + "bbox": [ + 312, + 544, + 555, + 652 + ], + "type": "text", + "content": "Light is the key component that determines how a person looks, which is often defined by a specific time and space, where revisiting such a unique moment gives us the opportunity to experience authentic telepresence sensations. In this paper, we introduce a generalizable human relighting model that can control the lighting from an image or video of humans with arbitrary body parts (Fig. 3), which are well-harmonized with a conditioning space (i.e., background image) as shown in Fig. 1." + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "text", + "content": "As shown in Fig. 2, existing image-based relighting methods face two main problems, lack of generalizability and controllability. First, they are designed for a specific scenario, e.g., only for face illumination or static humans [33, 61, 68] mainly due to the scarcity of large-scale" + } + ] + } + ], + "index": 36 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 702, + 293, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 702, + 293, + 713 + ], + "spans": [ + { + "bbox": [ + 66, + 702, + 293, + 713 + ], + "type": "text", + "content": "This work is partially done during an internship at Adobe Research." + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 297, + 748, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 748, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 748, + 313, + 757 + ], + "type": "text", + "content": "380" + } + ] + } + ], + "index": 38 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 70, + 294, + 171 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 294, + 171 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 294, + 171 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 294, + 171 + ], + "type": "image", + "image_path": "ebd58bb97fc109e3f772bac1266c9459db804e4aaf6c07c7a8462971b961dd67.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 174, + 295, + 196 + ], + "lines": [ + { + "bbox": [ + 55, + 174, + 295, + 196 + ], + "spans": [ + { + "bbox": [ + 55, + 174, + 295, + 196 + ], + "type": "text", + "content": "Figure 2. Comparison of various baseline methods for relighting settings and functionalities." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 205, + 295, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 205, + 295, + 361 + ], + "spans": [ + { + "bbox": [ + 54, + 205, + 295, + 361 + ], + "type": "text", + "content": "relighting datasets: Precise acquisition of the appearance of a static person under assorted lighting conditions requires specialized setups such as LightStage [22, 33] or expensive graphics simulation [26], which are not scalable, particularly for video contents. Learning from such limited data induces weak generalization of the model to diverse scenes. Second, most relighting algorithms struggle to effectively model multiple light sources. Typically, these algorithms are restricted to a single lighting control from either background image (e.g. high dynamic range lighting environment map [22]) or target lighting parameters (e.g. Spherical harmonics [24]). These problems inhibit the production-level application that requires general use cases." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 361, + 295, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 361, + 295, + 589 + ], + "spans": [ + { + "bbox": [ + 55, + 361, + 295, + 589 + ], + "type": "text", + "content": "To overcome these challenges, we propose an effective approach to achieve all-in-one relighting by utilizing a diffusion model—a general image prior that learns massive visual data with diverse lighting conditions; and repurpose this prior specialized for a human relighting and harmonization in a coarse-to-fine framework: A pretrained latent diffusion model [41] learns from limited datasets of static humans to jointly perform the fine-grained relighting and harmonization from two multi-modal lighting variables: the coarse shading estimate and conditioning background scene. The coarse shading estimate can be \"computed\" from conditioning lighting parameters (i.e., Spherical harmonics) without a neural network, and therefore, it is generalizable. The diffusion model is required to learn only the residual portion (e.g., fine self-occluded shadow), which is more generalizable than direct relighting (e.g., [68]). The pre-trained image prior helps with understanding the properties of the general background scenes, enabling the generalization of the background harmonization." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 590, + 295, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 590, + 295, + 673 + ], + "spans": [ + { + "bbox": [ + 55, + 590, + 295, + 673 + ], + "type": "text", + "content": "Our coarse-to-fine framework, however, still introduces problems: the diffusion model that learns to generate an image without temporal context produces significant temporal artifacts (e.g., sudden changes of lighting distribution even for consecutive frames). We address this problem by introducing an unsupervised temporal lighting model that learns from many real videos without any ground-truth data to en" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 317, + 70, + 553, + 198 + ], + "blocks": [ + { + "bbox": [ + 317, + 70, + 553, + 198 + ], + "lines": [ + { + "bbox": [ + 317, + 70, + 553, + 198 + ], + "spans": [ + { + "bbox": [ + 317, + 70, + 553, + 198 + ], + "type": "image", + "image_path": "5d7ffff5401d1b0ae3c69cf2e14429bf4d6713475641b686688a1500756b4919.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 202, + 553, + 236 + ], + "lines": [ + { + "bbox": [ + 313, + 202, + 553, + 236 + ], + "spans": [ + { + "bbox": [ + 313, + 202, + 553, + 236 + ], + "type": "text", + "content": "Figure 3. Our model generalizes to various body parts (portrait, half-body, full-body, multiperson) for relighting and harmonization, with lighting control variables shown in the insets." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 244, + 553, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 244, + 553, + 316 + ], + "spans": [ + { + "bbox": [ + 313, + 244, + 553, + 316 + ], + "type": "text", + "content": "force the temporal lighting consistency over frames. This temporal module learns the unsupervised cycle consistency between the relit and many real videos to predict the future lighting distribution from the past, which can be directly combined with our coarse-to-fine relighting components without extra training." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 317, + 554, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 317, + 554, + 449 + ], + "spans": [ + { + "bbox": [ + 313, + 317, + 554, + 449 + ], + "type": "text", + "content": "In inference time, we further push the temporal coherence of the relighting and harmonization by formulating a recurrent prediction pipeline. The generation at the current time step is conditioned onto the temporal module for the one at the next time. The features from the lighting and temporal control modules are spatially and temporally blended to enforce temporal coherence while improving the structure of the lighting distribution. Finally, our guided refinement module enhances the quality of the generated image in a way that preserves the original high-frequency details of the input image." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 449, + 554, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 449, + 554, + 533 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 554, + 533 + ], + "type": "text", + "content": "In the experiments, Comprehensive Relighting demonstrates high-quality relighting and background harmonization results with strong temporal coherence across the lighting, background, and pose changes. It also shows strong generalizability to any unconstrained scenes regardless of body parts, views, and poses, outperforming existing relighting and background harmonization methods." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 534, + 554, + 642 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 554, + 642 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 554, + 642 + ], + "type": "text", + "content": "Our contributions include: (1) To the best of our knowledge, the first approach for joint modeling of relighting and background harmonization; (2) a novel coarse-to-fine framework that enables comprehensive generalization by only learning from limited synthetic and lab-controlled data; (3) unsupervised temporal modeling with lighting cycle consistency from many unconstrained real videos; (4) an effective inference algorithm with spatio-temporal feature blending and guided refinement." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 655, + 400, + 667 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 655, + 400, + 667 + ], + "spans": [ + { + "bbox": [ + 314, + 655, + 400, + 667 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "content": "Image-based Human Relighting While high-quality relighting often requires resource-intensive setups, recent efforts in mobile device relighting aim to address single" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 55, + 683, + 295, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 683, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 683, + 295, + 713 + ], + "type": "text", + "content": "To the best of our knowledge, there exists no public ground truth data for video relighting of dynamic humans, and therefore, fine-tuning an existing video diffusion model (e.g., [17]) is not a readily available option." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "381" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 54, + 72, + 294, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 72, + 294, + 180 + ], + "spans": [ + { + "bbox": [ + 54, + 72, + 294, + 180 + ], + "type": "text", + "content": "image scenarios. Total Relighting [33] achieves photorealistic effects by leveraging detailed normal maps and albedo as priors. However, it relies on an HDR environment map that accurately matches the scene's real-world lighting, which is not always readily available, particularly in flexible and personalized relighting scenarios. [37, 46] estimate and adjust the Spherical harmonics parameters from images, and others use one-light-at-a-time (OLAT) captures to generate relighting data [49] or estimate reflectance fields [46]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 182, + 297, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 182, + 297, + 421 + ], + "spans": [ + { + "bbox": [ + 57, + 182, + 297, + 421 + ], + "type": "text", + "content": "Recent works have focused on diverse relighting scenarios including portrait [10, 19, 20, 25, 30, 33, 46, 58, 67, 68], full-body [9, 22, 26, 47], and object relighting [1, 55, 56]. Most of these works rely on decomposing an image into its intrinsic components, i.e., albedo, normal, and lighting, and therefore, the accuracy of this decomposition directly impacts the quality of the final relighting. While some works [22, 66, 69] use single-image geometry reconstruction for traditional and neural rendering, reconstruction errors are often propagated to relighting results. Tajima et al. [47] tackled domain adaptation with a two-step relighting framework, yet noticeable texture distortion remains due to limited model generalizability. Zhang et al. [59] trained a 2D latent-diffusion model, allowing users to manipulate and construct face NeRFs in a zero-shot learning framework without the need for explicit 3D data. DiFaRei [36] utilizes DDIM (Denoising Diffusion Implicit Models [44]) for high-fidelity face relighting. While promising, their methods are constrained by a focus on either specific body parts, limiting their applicability for generalizable." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 423, + 295, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 423, + 295, + 651 + ], + "spans": [ + { + "bbox": [ + 56, + 423, + 295, + 651 + ], + "type": "text", + "content": "Background Harmonization Background harmonization seeks to harmonize color, contrast, and style discrepancies between the foreground and background, ensuring composite images appear natural and cohesive. Many existing background harmonization methods [4, 7, 8, 13-15, 23, 51, 52, 70] formulate this problem as image-to-image translation work where a neural network translates an unharmonized foreground image to the harmonized one in the context of the conditioning background image. Recently, Relightful Harmonization [38] introduced methods that harmonize both image style and lighting to match the background scene, and IC-Light [63] further enables flexible illumination control via image diffusion models, guided by text descriptions or background images. While promising, these methods lack explicit lighting control, restricting general applicability and consistent video relighting. Additionally, some ([38]) are limited to specific body parts (e.g., portraits), with constrained support for full-body and multiperson scenarios." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 653, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 295, + 714 + ], + "type": "text", + "content": "Video Relighting Neural Radiance Fields (NeRF) [32] based methods enable novel view synthesis under varying lighting conditions in videos [43, 50, 53, 69]. Zhang et al. [61] achieve portrait video relighting under dynamic illuminations, while Choi et al. [6] ensure temporally consis" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 553, + 253 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 253 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 253 + ], + "type": "text", + "content": "tent relit videos. 3D-PVR [3] present a 3D-aware, real-time method to relight videos of talking faces. Relighting4D [5] decomposes the time-varied human body as a set of neural fields of normal, occlusion, diffuse, and specular maps. However, their rendering quality are highly reliant on geometry accuracy. Some works [36] adopt a temporal modeling scheme real-time neural video portrait relighting. STNeRF [60] controls dynamic scenes using a neural layered radiance representation that maintains spatial and temporal coherence. Li et al. [28] employed multi-view reconstruction for free-viewpoint relighting videos under general illumination. Richardt et al. [39] use RGBZ video cameras for video effects, including relighting, but rely on multi-view reconstruction from specialized hardware, making the process highly complex and costly." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 263, + 370, + 276 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 263, + 370, + 276 + ], + "spans": [ + { + "bbox": [ + 313, + 263, + 370, + 276 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 284, + 555, + 476 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 284, + 555, + 476 + ], + "spans": [ + { + "bbox": [ + 313, + 284, + 555, + 476 + ], + "type": "text", + "content": "We develop a generalizable and consistent human relighting and harmonization framework using a diffusion model. Fig. 4 illustrates the overview of our framework. Given an input image of humans and control lighting variables, including coarse shading and background image, a diffusion model learns to generate a fine-grained image of the humans under the target lighting, also harmonized with the background (Sec. 3.2). An external temporal lighting module, trained on real videos using unsupervised temporal cycle consistency, is integrated into the diffusion model to enhance temporal lighting coherence (Sec. 3.3). In inference time, we blend the features between the lighting and temporal control modules over time to ensure the relighting results are accurate and temporally coherent, and we apply a guided refinement to prevent the loss of high-frequency details during the denoising process (Sec. 3.4)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 484, + 512, + 498 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 484, + 512, + 498 + ], + "spans": [ + { + "bbox": [ + 313, + 484, + 512, + 498 + ], + "type": "text", + "content": "3.1. Background: Image-based Relighting" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 502, + 554, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 502, + 554, + 562 + ], + "spans": [ + { + "bbox": [ + 313, + 502, + 554, + 562 + ], + "type": "text", + "content": "Image-based relighting function can be compactly modeled with a small number of approximated basis of Spherical harmonics (SH) [9, 26, 42] which describes only essential features of illumination on the surface of a 3D sphere based on the following formulation:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 348, + 571, + 555, + 605 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 348, + 571, + 555, + 605 + ], + "spans": [ + { + "bbox": [ + 348, + 571, + 555, + 605 + ], + "type": "interline_equation", + "content": "I (\\mathbf {x}) = \\rho (\\mathbf {x}) \\cdot \\sum_ {l = 0} ^ {k} \\sum_ {m = - l} ^ {l} \\phi Y _ {l m} (\\mathbf {n} (\\mathbf {x})) \\tag {1}", + "image_path": "15224d8f7a45c12c613217490849c9e866f04581d962ea7bf995fd31a20a2267.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 616, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 616, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 616, + 555, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 616, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\phi \\in \\mathbb{R}^{(k + 1)^2}" + }, + { + "bbox": [ + 313, + 616, + 555, + 715 + ], + "type": "text", + "content": " denotes the spherical harmonics (SH) coefficients; " + }, + { + "bbox": [ + 313, + 616, + 555, + 715 + ], + "type": "inline_equation", + "content": "Y_{lm}(\\mathbf{n})" + }, + { + "bbox": [ + 313, + 616, + 555, + 715 + ], + "type": "text", + "content": " are the SH basis functions evaluated at the surface normal " + }, + { + "bbox": [ + 313, + 616, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{n}" + }, + { + "bbox": [ + 313, + 616, + 555, + 715 + ], + "type": "text", + "content": "; indices " + }, + { + "bbox": [ + 313, + 616, + 555, + 715 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 313, + 616, + 555, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 616, + 555, + 715 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 616, + 555, + 715 + ], + "type": "text", + "content": " represent the band and order within each band, respectively; and " + }, + { + "bbox": [ + 313, + 616, + 555, + 715 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 616, + 555, + 715 + ], + "type": "text", + "content": " is the maximum order of spherical harmonics used. While SH is computationally efficient and highly generalizable, it only captures low-frequency lighting, limiting its ability to model fine-grained, high-frequency illumination details. Addition-" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 748, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 748, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 748, + 312, + 757 + ], + "type": "text", + "content": "382" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 72, + 267, + 215 + ], + "blocks": [ + { + "bbox": [ + 61, + 72, + 267, + 215 + ], + "lines": [ + { + "bbox": [ + 61, + 72, + 267, + 215 + ], + "spans": [ + { + "bbox": [ + 61, + 72, + 267, + 215 + ], + "type": "image", + "image_path": "fe0e59f14ac5883b7b71a881062be5b17f97a96e7a2f25f3bfcc420174600209.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 104, + 219, + 225, + 227 + ], + "lines": [ + { + "bbox": [ + 104, + 219, + 225, + 227 + ], + "spans": [ + { + "bbox": [ + 104, + 219, + 225, + 227 + ], + "type": "text", + "content": "(a) Coarse-to-fine relighting and harmonization" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 269, + 74, + 479, + 215 + ], + "blocks": [ + { + "bbox": [ + 269, + 74, + 479, + 215 + ], + "lines": [ + { + "bbox": [ + 269, + 74, + 479, + 215 + ], + "spans": [ + { + "bbox": [ + 269, + 74, + 479, + 215 + ], + "type": "image", + "image_path": "2638a0955625b65dd185371f2d6fbc9c937a91808ec41174a9f3c1006d95c397.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 294, + 219, + 412, + 228 + ], + "lines": [ + { + "bbox": [ + 294, + 219, + 412, + 228 + ], + "spans": [ + { + "bbox": [ + 294, + 219, + 412, + 228 + ], + "type": "text", + "content": "(b) Temporal modeling with cycle consistency" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 430, + 219, + 550, + 228 + ], + "lines": [ + { + "bbox": [ + 430, + 219, + 550, + 228 + ], + "spans": [ + { + "bbox": [ + 430, + 219, + 550, + 228 + ], + "type": "text", + "content": "(c) Inference: Spatio-temporal feature blending" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 233, + 555, + 278 + ], + "lines": [ + { + "bbox": [ + 55, + 233, + 555, + 278 + ], + "spans": [ + { + "bbox": [ + 55, + 233, + 555, + 278 + ], + "type": "text", + "content": "Figure 4. System overview. (a) Given an input image of humans with coarse lighting and background image, our diffusion model generates the relit images harmonized with background scenes (Sec. 3.2). (b) The external temporal modules learn the temporal cycle consistency from many real-world videos to construct temporal lighting features (Sec. 3.3). (c) In inference time, we blend the features from lighting and temporal modules spatially and temporally to enable coherent and generalizable human relighting (Sec. 3.4)." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 481, + 74, + 548, + 213 + ], + "blocks": [ + { + "bbox": [ + 481, + 74, + 548, + 213 + ], + "lines": [ + { + "bbox": [ + 481, + 74, + 548, + 213 + ], + "spans": [ + { + "bbox": [ + 481, + 74, + 548, + 213 + ], + "type": "image", + "image_path": "22cf249a94d54a2f4bec6ba87273d378bbe3c1046817b7ee8231579a47af2826.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 289, + 296, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 289, + 296, + 373 + ], + "spans": [ + { + "bbox": [ + 55, + 289, + 296, + 373 + ], + "type": "text", + "content": "ally, SH-based global illumination neglects the context provided by background images, often resulting in relit images appearing unnatural when composited with different backgrounds. To overcome these limitations, we propose a coarse-to-fine human relighting and harmonization framework that leverages the strong image prior available from a pre-trained text-to-image diffusion model." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 380, + 294, + 393 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 380, + 294, + 393 + ], + "spans": [ + { + "bbox": [ + 55, + 380, + 294, + 393 + ], + "type": "text", + "content": "3.2. Coarse-to-Fine Relighting and Harmonization" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 398, + 295, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 398, + 295, + 422 + ], + "spans": [ + { + "bbox": [ + 55, + 398, + 295, + 422 + ], + "type": "text", + "content": "We generate a fine-grained relit image of a person conditioned on a coarse lighting representation:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 108, + 430, + 294, + 443 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 430, + 294, + 443 + ], + "spans": [ + { + "bbox": [ + 108, + 430, + 294, + 443 + ], + "type": "interline_equation", + "content": "\\mathcal {E} (\\mathbf {I}; \\mathbf {S} _ {\\phi}) = z, \\quad \\mathcal {D} (z) = \\mathbf {I} _ {\\phi} \\tag {2}", + "image_path": "9cc94ee8cbc7b0df1f022ad991b440f7ffacf19b4e0b0a7ea47592757ca9b9ac.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 450, + 295, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 450, + 295, + 688 + ], + "spans": [ + { + "bbox": [ + 55, + 450, + 295, + 688 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 450, + 295, + 688 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 55, + 450, + 295, + 688 + ], + "type": "text", + "content": " is an encoder that generates the latent lighting features, " + }, + { + "bbox": [ + 55, + 450, + 295, + 688 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 55, + 450, + 295, + 688 + ], + "type": "text", + "content": " as a function of the input image " + }, + { + "bbox": [ + 55, + 450, + 295, + 688 + ], + "type": "inline_equation", + "content": "\\mathbf{I} \\in \\mathbb{R}^{w \\times h \\times 3}" + }, + { + "bbox": [ + 55, + 450, + 295, + 688 + ], + "type": "text", + "content": " and a small number of global lighting parameters " + }, + { + "bbox": [ + 55, + 450, + 295, + 688 + ], + "type": "inline_equation", + "content": "\\phi \\in \\mathbb{R}^n" + }, + { + "bbox": [ + 55, + 450, + 295, + 688 + ], + "type": "text", + "content": " (i.e., Spherical harmonics where " + }, + { + "bbox": [ + 55, + 450, + 295, + 688 + ], + "type": "inline_equation", + "content": "n = 25" + }, + { + "bbox": [ + 55, + 450, + 295, + 688 + ], + "type": "text", + "content": "); and " + }, + { + "bbox": [ + 55, + 450, + 295, + 688 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 55, + 450, + 295, + 688 + ], + "type": "text", + "content": " is the decoder that generates fine-grained relit image " + }, + { + "bbox": [ + 55, + 450, + 295, + 688 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{\\phi} \\in \\mathbb{R}^{w \\times h \\times 3}" + }, + { + "bbox": [ + 55, + 450, + 295, + 688 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 55, + 450, + 295, + 688 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 55, + 450, + 295, + 688 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 55, + 450, + 295, + 688 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_{\\phi} \\in \\mathbb{R}^{w \\times h}" + }, + { + "bbox": [ + 55, + 450, + 295, + 688 + ], + "type": "text", + "content": " is the pixel-aligned coarse shading map converted from the coarse lighting parameters " + }, + { + "bbox": [ + 55, + 450, + 295, + 688 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 55, + 450, + 295, + 688 + ], + "type": "text", + "content": " as shown in Fig. 4-(a). One approach to obtaining " + }, + { + "bbox": [ + 55, + 450, + 295, + 688 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_{\\phi}" + }, + { + "bbox": [ + 55, + 450, + 295, + 688 + ], + "type": "text", + "content": " is directly computing it as demonstrated in Eq. 1 where we can compute lighting intensity along with the detected surface normal map from " + }, + { + "bbox": [ + 55, + 450, + 295, + 688 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 55, + 450, + 295, + 688 + ], + "type": "text", + "content": " given spherical harmonics coefficients " + }, + { + "bbox": [ + 55, + 450, + 295, + 688 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 55, + 450, + 295, + 688 + ], + "type": "text", + "content": ". Another way, is to use a neural network to convert the surface normal to " + }, + { + "bbox": [ + 55, + 450, + 295, + 688 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_{\\phi}" + }, + { + "bbox": [ + 55, + 450, + 295, + 688 + ], + "type": "text", + "content": " as a condition of spherical harmonics coefficients " + }, + { + "bbox": [ + 55, + 450, + 295, + 688 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 55, + 450, + 295, + 688 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 55, + 450, + 295, + 688 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_{\\phi} \\gets f(\\mathbf{N}; \\phi)" + }, + { + "bbox": [ + 55, + 450, + 295, + 688 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 55, + 450, + 295, + 688 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 55, + 450, + 295, + 688 + ], + "type": "text", + "content": " is the neural shading function which maps the surface normal " + }, + { + "bbox": [ + 55, + 450, + 295, + 688 + ], + "type": "inline_equation", + "content": "\\mathbf{N}" + }, + { + "bbox": [ + 55, + 450, + 295, + 688 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 450, + 295, + 688 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 55, + 450, + 295, + 688 + ], + "type": "text", + "content": " to the coarse shading. While both approaches are highly generalizable, our experiments indicate that the latter method achieves improved accuracy and better noise correction. Please refer to the Supple. documents for more details about the coarse shading estimation." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": "To jointly model the relighting and background harmonization, we further condition a target background image" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 289, + 509, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 289, + 509, + 301 + ], + "spans": [ + { + "bbox": [ + 314, + 289, + 509, + 301 + ], + "type": "inline_equation", + "content": "\\mathbf{B} \\in \\mathbb{R}^{w \\times h \\times 3}" + }, + { + "bbox": [ + 314, + 289, + 509, + 301 + ], + "type": "text", + "content": " as the additional lighting sources:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 355, + 311, + 553, + 324 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 355, + 311, + 553, + 324 + ], + "spans": [ + { + "bbox": [ + 355, + 311, + 553, + 324 + ], + "type": "interline_equation", + "content": "\\mathcal {E} (\\mathbf {I}; \\{\\mathbf {S} _ {\\phi}, \\mathbf {B} \\}) = \\boldsymbol {z}, \\quad \\mathcal {D} (\\boldsymbol {z}) = \\mathbf {I} _ {\\phi} \\tag {3}", + "image_path": "45d0f26459d973fced9ed8d557d26423862f7dccab105dc61cae81652c5ae523.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 332, + 554, + 404 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 332, + 554, + 404 + ], + "spans": [ + { + "bbox": [ + 313, + 332, + 554, + 404 + ], + "type": "text", + "content": "where the lighting encoder " + }, + { + "bbox": [ + 313, + 332, + 554, + 404 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 313, + 332, + 554, + 404 + ], + "type": "text", + "content": " learns to capture the intensity and direction of light from " + }, + { + "bbox": [ + 313, + 332, + 554, + 404 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_{\\phi}" + }, + { + "bbox": [ + 313, + 332, + 554, + 404 + ], + "type": "text", + "content": " while capturing ambient environment lighting and color distribution from " + }, + { + "bbox": [ + 313, + 332, + 554, + 404 + ], + "type": "inline_equation", + "content": "\\mathbf{B}" + }, + { + "bbox": [ + 313, + 332, + 554, + 404 + ], + "type": "text", + "content": ". This enables " + }, + { + "bbox": [ + 313, + 332, + 554, + 404 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 313, + 332, + 554, + 404 + ], + "type": "text", + "content": " to achieve complete relighting in scenarios involving new target lighting, background, or both. In the subsection, we enable " + }, + { + "bbox": [ + 313, + 332, + 554, + 404 + ], + "type": "inline_equation", + "content": "(\\mathcal{D} \\circ \\mathcal{E})" + }, + { + "bbox": [ + 313, + 332, + 554, + 404 + ], + "type": "text", + "content": " using a latent diffusion model." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 410, + 516, + 422 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 410, + 516, + 422 + ], + "spans": [ + { + "bbox": [ + 313, + 410, + 516, + 422 + ], + "type": "text", + "content": "3.2.1. Fine-grained Relighting Diffusion Model" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 425, + 555, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 425, + 555, + 472 + ], + "spans": [ + { + "bbox": [ + 313, + 425, + 555, + 472 + ], + "type": "text", + "content": "We enable the fine-grained image relighting using a diffusion model as illustrated in Fig. 4-(a). Our encoder " + }, + { + "bbox": [ + 313, + 425, + 555, + 472 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 313, + 425, + 555, + 472 + ], + "type": "text", + "content": " in Eq. 3, in practice, is formulated as the composition of two encoders:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 335, + 481, + 553, + 495 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 481, + 553, + 495 + ], + "spans": [ + { + "bbox": [ + 335, + 481, + 553, + 495 + ], + "type": "interline_equation", + "content": "\\mathcal {E} \\rightarrow \\mathcal {E} _ {\\mathrm {b}} (\\mathbf {I}; \\mathcal {E} _ {1} (\\{\\mathbf {S} _ {\\phi}, \\mathbf {B} \\})) = \\boldsymbol {z}, \\quad \\mathcal {D} (\\boldsymbol {z}) = \\mathbf {I} _ {\\phi} \\tag {4}", + "image_path": "a1cd2ce97b42d9276535454a7a93bf57218993c7f5c43f7c55dec664a23a73af.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 502, + 555, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 502, + 555, + 647 + ], + "spans": [ + { + "bbox": [ + 313, + 502, + 555, + 647 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 502, + 555, + 647 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_1" + }, + { + "bbox": [ + 313, + 502, + 555, + 647 + ], + "type": "text", + "content": " encodes lighting control variables, " + }, + { + "bbox": [ + 313, + 502, + 555, + 647 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{S}_{\\phi}, \\mathbf{B}\\}" + }, + { + "bbox": [ + 313, + 502, + 555, + 647 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 502, + 555, + 647 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_b" + }, + { + "bbox": [ + 313, + 502, + 555, + 647 + ], + "type": "text", + "content": " encodes the conditional variable " + }, + { + "bbox": [ + 313, + 502, + 555, + 647 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 313, + 502, + 555, + 647 + ], + "type": "text", + "content": " whose visual properties, e.g., semantics and identity, should be preserved in the output along with the controls from " + }, + { + "bbox": [ + 313, + 502, + 555, + 647 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_1" + }, + { + "bbox": [ + 313, + 502, + 555, + 647 + ], + "type": "text", + "content": ". For " + }, + { + "bbox": [ + 313, + 502, + 555, + 647 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_b" + }, + { + "bbox": [ + 313, + 502, + 555, + 647 + ], + "type": "text", + "content": ", we use the base latent diffusion model [2, 40] pre-trained for text-to-image generation task, and for " + }, + { + "bbox": [ + 313, + 502, + 555, + 647 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_1" + }, + { + "bbox": [ + 313, + 502, + 555, + 647 + ], + "type": "text", + "content": ", we use ControlNet [62] (termed as Light ControlNet in Fig. 4-(a)). While noise, texts, and timestep variables are also conditioned on " + }, + { + "bbox": [ + 313, + 502, + 555, + 647 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_b" + }, + { + "bbox": [ + 313, + 502, + 555, + 647 + ], + "type": "text", + "content": " to fit the modality of the latent diffusion model, they are not described in the equations and figures for conciseness. To impose the foreground attention on the lighting control, a portrait mask is also conditioned to " + }, + { + "bbox": [ + 313, + 502, + 555, + 647 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_l" + }, + { + "bbox": [ + 313, + 502, + 555, + 647 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 339, + 655, + 553, + 669 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 655, + 553, + 669 + ], + "spans": [ + { + "bbox": [ + 339, + 655, + 553, + 669 + ], + "type": "interline_equation", + "content": "\\mathcal {E} _ {\\mathrm {b}} (\\mathbf {I}; \\mathcal {E} _ {1} (\\{\\mathbf {S} _ {\\phi}, \\mathbf {B} \\}; \\mathbf {M})) = \\boldsymbol {z}, \\quad \\mathcal {D} (\\boldsymbol {z}) = \\mathbf {I} _ {\\phi} \\tag {5}", + "image_path": "cd80eafb6bef9b9fa61e7349d8a1934f1408867e01b8efa37ec832ef560534c2.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{M} \\in \\{0,1\\}^{w \\times h}" + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "content": " is the foreground binary mask. In training time, the diffusion model learns to directly change the lighting distribution of the input images (without inverse" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "383" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 193 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 193 + ], + "type": "text", + "content": "rendering techniques) as the condition of coarse lighting estimate by minimizing the latent distance of the noisy relit image with the clean one from the ground truth in the overall forward and background denoising steps. Following the findings from an existing harmonization work [38], we use the composite image between the input image and conditioning background image as " + }, + { + "bbox": [ + 55, + 72, + 296, + 193 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 55, + 72, + 296, + 193 + ], + "type": "text", + "content": ". We randomly drop background " + }, + { + "bbox": [ + 55, + 72, + 296, + 193 + ], + "type": "inline_equation", + "content": "\\mathbf{B}" + }, + { + "bbox": [ + 55, + 72, + 296, + 193 + ], + "type": "text", + "content": " or randomly set the coarse shading " + }, + { + "bbox": [ + 55, + 72, + 296, + 193 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_{\\phi}" + }, + { + "bbox": [ + 55, + 72, + 296, + 193 + ], + "type": "text", + "content": " as binary mask such that the diffusion model learns the control the lighting from " + }, + { + "bbox": [ + 55, + 72, + 296, + 193 + ], + "type": "inline_equation", + "content": "\\mathbf{B}" + }, + { + "bbox": [ + 55, + 72, + 296, + 193 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 72, + 296, + 193 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_{\\phi}" + }, + { + "bbox": [ + 55, + 72, + 296, + 193 + ], + "type": "text", + "content": ", or both " + }, + { + "bbox": [ + 55, + 72, + 296, + 193 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{B}, \\mathbf{S}_{\\phi}\\}" + }, + { + "bbox": [ + 55, + 72, + 296, + 193 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 199, + 277, + 213 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 199, + 277, + 213 + ], + "spans": [ + { + "bbox": [ + 55, + 199, + 277, + 213 + ], + "type": "text", + "content": "3.3. Unsupervised Add-on Temporal Modeling" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 216, + 296, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 216, + 296, + 324 + ], + "spans": [ + { + "bbox": [ + 55, + 216, + 296, + 324 + ], + "type": "text", + "content": "Our coarse-to-fine relighting framework " + }, + { + "bbox": [ + 55, + 216, + 296, + 324 + ], + "type": "inline_equation", + "content": "(\\mathcal{D} \\circ \\mathcal{E}_{\\mathrm{b}} \\circ \\mathcal{E}_{\\mathrm{l}})" + }, + { + "bbox": [ + 55, + 216, + 296, + 324 + ], + "type": "text", + "content": " is trained only on individual images, inherently missing temporal context, e.g., how a point on a human's surface will radiate from a specific viewpoint under the continuous pose, view, and illumination changes, leading to temporal artifacts such as flickering. We model such temporal context by designing an external add-on temporal lighting module " + }, + { + "bbox": [ + 55, + 216, + 296, + 324 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{m}" + }, + { + "bbox": [ + 55, + 216, + 296, + 324 + ], + "type": "text", + "content": " that can be combined, in inference time, with the relighting framework without extra training:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 94, + 333, + 295, + 346 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 333, + 295, + 346 + ], + "spans": [ + { + "bbox": [ + 94, + 333, + 295, + 346 + ], + "type": "interline_equation", + "content": "\\mathcal {D} \\circ \\mathcal {E} _ {\\mathrm {b}} \\circ \\mathcal {E} _ {\\mathrm {l}} \\rightarrow \\mathcal {D} \\circ \\mathcal {E} _ {\\mathrm {b}} \\circ \\left(\\mathcal {E} _ {\\mathrm {l}} \\times \\mathcal {E} _ {\\mathrm {m}}\\right) \\tag {6}", + "image_path": "817fa481c201acea1554b8e2cffff41106f522081f674835560675f6c4fa322f.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 355, + 296, + 427 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 355, + 296, + 427 + ], + "spans": [ + { + "bbox": [ + 55, + 355, + 296, + 427 + ], + "type": "text", + "content": "To enable this, our temporal module is designed to regress the relit image from the previous time instance, i.e., " + }, + { + "bbox": [ + 55, + 355, + 296, + 427 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{\\phi}^{t - 1}" + }, + { + "bbox": [ + 55, + 355, + 296, + 427 + ], + "type": "text", + "content": ", to the latent lighting distribution whose space is shared with our base relighting models. Therefore, our decoder can generate temporally coherent relit images in the current time in an auto-regressive way:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 104, + 434, + 295, + 449 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 434, + 295, + 449 + ], + "spans": [ + { + "bbox": [ + 104, + 434, + 295, + 449 + ], + "type": "interline_equation", + "content": "\\mathcal {D} \\left(\\mathcal {E} _ {\\mathrm {b}} \\left(\\mathbf {I} ^ {t}; \\mathcal {E} _ {\\mathrm {m}} \\left(\\mathbf {I} _ {\\phi} ^ {t - 1}\\right)\\right)\\right) = \\mathbf {I} _ {t - 1} ^ {t}. \\tag {7}", + "image_path": "49c89fec3f397468a53d6264e121ff20ba324e7b3a264c06b9f66d4c52fc1ab0.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 457, + 296, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 457, + 296, + 518 + ], + "spans": [ + { + "bbox": [ + 55, + 457, + 296, + 518 + ], + "type": "text", + "content": "However, training " + }, + { + "bbox": [ + 55, + 457, + 296, + 518 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{\\mathrm{m}}" + }, + { + "bbox": [ + 55, + 457, + 296, + 518 + ], + "type": "text", + "content": " with a conventional L2 loss is not possible since there exists no ground-truth video relighting data for dynamic humans. Therefore, we propose to learn " + }, + { + "bbox": [ + 55, + 457, + 296, + 518 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{\\mathrm{m}}" + }, + { + "bbox": [ + 55, + 457, + 296, + 518 + ], + "type": "text", + "content": " in an unsupervised way using many real videos with lighting cycle consistency as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 95, + 525, + 295, + 540 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 95, + 525, + 295, + 540 + ], + "spans": [ + { + "bbox": [ + 95, + 525, + 295, + 540 + ], + "type": "interline_equation", + "content": "\\mathcal {D} ^ {*} \\left(\\mathcal {E} _ {\\mathrm {b}} ^ {*} \\left(\\mathbf {I} ^ {t}; \\mathcal {E} _ {1} ^ {*} \\left(\\left\\{\\mathbf {S} _ {\\phi} ^ {t}, \\mathbf {B} ^ {t} \\right\\}; \\mathbf {M} ^ {t}\\right)\\right)\\right) = \\mathbf {I} _ {\\phi} ^ {t}, \\tag {8}", + "image_path": "c9276f3062b6cebaec6de445245ef48a5e512cc04e57bb733365a29915e4f7b9.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 97, + 542, + 294, + 556 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 542, + 294, + 556 + ], + "spans": [ + { + "bbox": [ + 97, + 542, + 294, + 556 + ], + "type": "interline_equation", + "content": "\\mathcal {D} ^ {*} \\left(\\mathcal {E} _ {\\mathrm {b}} ^ {*} \\left(\\mathbf {I} _ {\\phi} ^ {t}; \\mathcal {E} _ {\\mathrm {m}} \\left(\\mathbf {I} ^ {t - 1}, \\mathbf {M} ^ {t - 1}\\right)\\right)\\right) = \\tilde {\\mathbf {I}} _ {t - 1} ^ {t} \\tag {9}", + "image_path": "0782b32283ad7a408189670b769e7bdaf156ac07972c9f333230429acf550f51.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 145, + 559, + 206, + 573 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 559, + 206, + 573 + ], + "spans": [ + { + "bbox": [ + 145, + 559, + 206, + 573 + ], + "type": "interline_equation", + "content": "\\therefore \\quad \\mathbf {I} ^ {t} = \\tilde {\\mathbf {I}} _ {t - 1} ^ {t}.", + "image_path": "5abe5b4f9540821ac8b795a065741b1c153d3ca2ea081c60fd0a97cfc1095029.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "text", + "content": "We make the hypothesis: a video sequence inherently contains temporal lighting properties whose flow can be implicitly modeled by learning to predict the lighting distribution of the future frame based on the hint from the previous frame. This involves forward and backward processes, as a cycle-training. Eq. 8 represents the forward image relighting, i.e., " + }, + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{I}^t\\rightarrow \\mathbf{I}_\\phi^t" + }, + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "text", + "content": " , where our relighting diffusion model with lighting ControlNet " + }, + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_1" + }, + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "text", + "content": " generates the relit image at time " + }, + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "text", + "content": " as a condition of a novel lighting condition (where we pick random Spherical harmonics). Eq. 9 reverts the relighting, i.e., " + }, + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{I}}_{t - 1}^{t}\\gets \\mathbf{I}_{\\phi}^{t}" + }, + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "text", + "content": " to the original input image conditioned by" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "type": "text", + "content": "the original frame in the previous time step through our temporal lighting module where the mask " + }, + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "type": "text", + "content": " is used for foreground awareness. Finally, " + }, + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_m" + }, + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "type": "text", + "content": " learns the lighting cycle consistency in the diffusion process by minimizing the latent distance between the reverted and the original image. We freeze the learnable weights for the functions with " + }, + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "type": "inline_equation", + "content": "*" + }, + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "type": "text", + "content": " during training, and " + }, + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{\\mathrm{m}}" + }, + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "type": "text", + "content": " is enabled with another ControlNet [62] termed as Motion ControlNet in Fig. 4-(b)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 174, + 555, + 198 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 174, + 555, + 198 + ], + "spans": [ + { + "bbox": [ + 313, + 174, + 555, + 198 + ], + "type": "text", + "content": "3.4. Inference with Spatio-Temporal Blending and Guided Refinement" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 203, + 555, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 203, + 555, + 407 + ], + "spans": [ + { + "bbox": [ + 313, + 203, + 555, + 407 + ], + "type": "text", + "content": "Given a video or image, we perform comprehensive relighting with a spatio-temporal feature blending framework. For " + }, + { + "bbox": [ + 313, + 203, + 555, + 407 + ], + "type": "inline_equation", + "content": "t = 0" + }, + { + "bbox": [ + 313, + 203, + 555, + 407 + ], + "type": "text", + "content": ", we generate the relit image without our temporal lighting module " + }, + { + "bbox": [ + 313, + 203, + 555, + 407 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{\\mathrm{m}}" + }, + { + "bbox": [ + 313, + 203, + 555, + 407 + ], + "type": "text", + "content": ". For " + }, + { + "bbox": [ + 313, + 203, + 555, + 407 + ], + "type": "inline_equation", + "content": "t > 0" + }, + { + "bbox": [ + 313, + 203, + 555, + 407 + ], + "type": "text", + "content": " (note that even for the static image, " + }, + { + "bbox": [ + 313, + 203, + 555, + 407 + ], + "type": "inline_equation", + "content": "t > 0" + }, + { + "bbox": [ + 313, + 203, + 555, + 407 + ], + "type": "text", + "content": " is possible since the lighting is time-variant), the generated relit images in the previous time step are conditioned on " + }, + { + "bbox": [ + 313, + 203, + 555, + 407 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{\\mathrm{m}}" + }, + { + "bbox": [ + 313, + 203, + 555, + 407 + ], + "type": "text", + "content": ", and therefore, the relighting is controlled by dual control modules, i.e., " + }, + { + "bbox": [ + 313, + 203, + 555, + 407 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{\\mathrm{l}}" + }, + { + "bbox": [ + 313, + 203, + 555, + 407 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 203, + 555, + 407 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{\\mathrm{m}}" + }, + { + "bbox": [ + 313, + 203, + 555, + 407 + ], + "type": "text", + "content": " by blending their features with a (0.85:0.15) ratio as described as spatial blending block in Fig. 4-(c). The blended lighting features are recurrently combined with the one from the previous time step through the temporal blending block. For this, we adopt optical flow from the original input video as a temporal prior to spatially align the features from consecutive frames similar to [27]; and we blend the aligned temporal features with a (0.5:0.5) ratio as in Fig. 4-(c) to improve the temporal consistency." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 407, + 555, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 407, + 555, + 479 + ], + "spans": [ + { + "bbox": [ + 313, + 407, + 555, + 479 + ], + "type": "text", + "content": "In the denoising process of the latent diffusion model, the generation often suffers from the loss of high-frequency details. Inspired by existing image restoration techniques [45, 65], we address this problem by applying a guided refinement to each generated image. We cast the problem of guided refinement as a guided residual prediction:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 370, + 483, + 553, + 497 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 483, + 553, + 497 + ], + "spans": [ + { + "bbox": [ + 370, + 483, + 553, + 497 + ], + "type": "interline_equation", + "content": "\\mathbf {I} _ {\\phi} ^ {\\text {r e f i n e}} = \\mathbf {I} + g \\left(\\mathbf {I} _ {\\phi}, \\mathbf {I}; \\mathbf {M}\\right), \\tag {10}", + "image_path": "7b2e3db50e52c16d3e37e1d4cc80245da0d290977f91ec5550cbe7ab0310ee39.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 501, + 555, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 501, + 555, + 609 + ], + "spans": [ + { + "bbox": [ + 313, + 501, + 555, + 609 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 501, + 555, + 609 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 313, + 501, + 555, + 609 + ], + "type": "text", + "content": " is the function that predicts the guided lighting residual. This residual learns to map the lighting distribution from " + }, + { + "bbox": [ + 313, + 501, + 555, + 609 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 313, + 501, + 555, + 609 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 501, + 555, + 609 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{\\phi}" + }, + { + "bbox": [ + 313, + 501, + 555, + 609 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 313, + 501, + 555, + 609 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{\\phi}^{\\mathrm{refine}}" + }, + { + "bbox": [ + 313, + 501, + 555, + 609 + ], + "type": "text", + "content": " can effectively preserve the high-frequency details of the input image " + }, + { + "bbox": [ + 313, + 501, + 555, + 609 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 313, + 501, + 555, + 609 + ], + "type": "text", + "content": " due to the nature of residual learning [57, 65] that is designed to preserve the visual properties from the observation space, i.e., " + }, + { + "bbox": [ + 313, + 501, + 555, + 609 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 313, + 501, + 555, + 609 + ], + "type": "text", + "content": ". We enable " + }, + { + "bbox": [ + 313, + 501, + 555, + 609 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 313, + 501, + 555, + 609 + ], + "type": "text", + "content": " with a residual network [16] by learning from our relighting data with general losses for low-level vision processing, i.e., L2 and VGG [11]." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 615, + 395, + 628 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 615, + 395, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 615, + 395, + 628 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 630, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 555, + 715 + ], + "type": "text", + "content": "Dataset. To train our coarse-to-fine relighting model, we use 100K ground-truth relighting samples, including 50K synthetic human renderings and 50K OLAT-captured images from LightStage, with random cropping augmentation. Ground-truth albedo, images, background, masks, and lighting coefficients are precomputed. Our dataset is categorized by gender, skin tone, and body part (full-body and" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "text", + "content": "384" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 70, + 555, + 140 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 555, + 140 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 555, + 140 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 555, + 140 + ], + "type": "table", + "html": "
MethodScenario 1Scenario 2Scenario 3
DPR [68]18.62/0.86/0.10332.00 /0.94/0.03018.14/0.89/0.08935.29 /0.98/0.03221.20/0.89/0.07237.11/0.95/0.038
RHW [47]19.78/0.87/0.11330.74/0.95/0.02720.12/0.88/0.07836.64 /0.98/0.02823.33/0.90/0.06037.53 /0.98/0.033
GFR [22]25.59/0.91/0.08930.33/0.95/0.03322.76/0.91/0.07232.36/0.98/0.03625.49/0.93/0.05035.89/0.98/0.028
LPBR [38]18.19/0.86/0.09031.62/0.91/0.03519.96/0.88/0.08427.94/0.94/0.04121.42/0.88/0.05333.39/0.95/0.038
Ours25.95/0.95/0.06633.58/0.96/0.02423.99/0.93/0.04835.18/0.99 / 0.03126.61/0.94/0.03338.32/0.98/0.026
", + "image_path": "b8e8e3b695fa2e897d1ffd00b27bd7cb75d0c32a2de4aab0bdbb6b3993e82470.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 142, + 555, + 177 + ], + "lines": [ + { + "bbox": [ + 55, + 142, + 555, + 177 + ], + "spans": [ + { + "bbox": [ + 55, + 142, + 555, + 177 + ], + "type": "text", + "content": "Table 1. Comparison with existing image-based human relighting methods on synthetic videos for fidelity and temporal consistency. Each column shows image fidelity (left) and video temporal consistency (right). Metrics are PSNR↑/SSIM↑/LPIPS↓ for accuracy, and tPSNR↑/tSSIM↑/tLPIPS↓ for temporal consistency. While green is used for the best values, yellow highlights the second-best values." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 58, + 185, + 553, + 413 + ], + "blocks": [ + { + "bbox": [ + 58, + 185, + 553, + 413 + ], + "lines": [ + { + "bbox": [ + 58, + 185, + 553, + 413 + ], + "spans": [ + { + "bbox": [ + 58, + 185, + 553, + 413 + ], + "type": "image", + "image_path": "12983553ec971990dd47a5b51513ba04934a5b976d430d60799d97d1e8e9dddf.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 417, + 555, + 440 + ], + "lines": [ + { + "bbox": [ + 55, + 417, + 555, + 440 + ], + "spans": [ + { + "bbox": [ + 55, + 417, + 555, + 440 + ], + "type": "text", + "content": "Figure 5. Qualitative comparison of synthetic video frames (corresponding to Tab. 1). From left to right: composite input with target lighting parameters (inset), our relit result, baseline methods, and normalized L2" + }, + { + "bbox": [ + 55, + 417, + 555, + 440 + ], + "type": "inline_equation", + "content": "\\downarrow" + }, + { + "bbox": [ + 55, + 417, + 555, + 440 + ], + "type": "text", + "content": " photometric error map (inset)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 449, + 295, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 449, + 295, + 700 + ], + "spans": [ + { + "bbox": [ + 54, + 449, + 295, + 700 + ], + "type": "text", + "content": "multi-body), with each subject captured from 32 viewpoints under varying lighting conditions. For a detailed dataset breakdown, refer to the Supple. To train our temporal control module using real data, we used 50K frames of videos from existing works [21, 29, 34, 54] and customized videos, where, for cyclic relighting, we randomly sample spherical harmonics lighting parameters from ground-truth data. For validation on static humans with dynamic lighting, we generated a new testing set (i.e., the ground-truth relit images and associated lighting parameters) using an internal simulation and LightStage. This set includes the mixture of half-body, full-body, and multi-person scenarios, with each scenario comprising 100 frames. For the validation on dynamic humans (i.e., video), we newly create synthetic video sequences for three scenarios: Scenario 1 involves static lighting (environment map) across frames with a moving human; Scenario 2 features dynamic lighting (rotated environment map) with a static human; Scenario 3 combines dynamic lighting (rotated environment map) with a moving human. Please refer to the Supple. and demo video for more details about the testing results." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 701, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 701, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 701, + 295, + 714 + ], + "type": "text", + "content": "Metrics. To measure the relighting quality, we use L1 dis" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 318, + 448, + 395, + 585 + ], + "blocks": [ + { + "bbox": [ + 318, + 448, + 395, + 585 + ], + "lines": [ + { + "bbox": [ + 318, + 448, + 395, + 585 + ], + "spans": [ + { + "bbox": [ + 318, + 448, + 395, + 585 + ], + "type": "image", + "image_path": "85944e6a87caa418279f9ec9e779d08aa4fcb45ddddfadcc5d1b991850428972.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 586, + 555, + 618 + ], + "lines": [ + { + "bbox": [ + 313, + 586, + 555, + 618 + ], + "spans": [ + { + "bbox": [ + 313, + 586, + 555, + 618 + ], + "type": "text", + "content": "Figure 6. Comparison with LPBR [38] on DeepFashion [31] real images for background harmonization testing. The first row shows the relit output, and the second shows the magnified results." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 395, + 448, + 488, + 585 + ], + "blocks": [ + { + "bbox": [ + 395, + 448, + 488, + 585 + ], + "lines": [ + { + "bbox": [ + 395, + 448, + 488, + 585 + ], + "spans": [ + { + "bbox": [ + 395, + 448, + 488, + 585 + ], + "type": "image", + "image_path": "f59d0190d8986b659b4caa768b947d68ac252014b6c64d256e25e8097dce0972.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 474, + 448, + 551, + 585 + ], + "blocks": [ + { + "bbox": [ + 474, + 448, + 551, + 585 + ], + "lines": [ + { + "bbox": [ + 474, + 448, + 551, + 585 + ], + "spans": [ + { + "bbox": [ + 474, + 448, + 551, + 585 + ], + "type": "image", + "image_path": "6cbc539142dee20d8d1392b23984a030b5768a668e205251d755ba33b270dcb1.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 630, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 555, + 715 + ], + "type": "text", + "content": "tance, reconstruction fidelity (PSNR [18]), local structure similarity (SSIM [18]), and the latent perceptual distance (LPIPS [64]), between the relighted and ground-truth. For measuring the temporal coherence of the relighting results, we follow the same logics as TokenFlow [12]. This involves warping the relit image from the previous time step to the next using optical flow [48], and then comparing this to the" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "385" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 69, + 553, + 312 + ], + "blocks": [ + { + "bbox": [ + 59, + 69, + 553, + 312 + ], + "lines": [ + { + "bbox": [ + 59, + 69, + 553, + 312 + ], + "spans": [ + { + "bbox": [ + 59, + 69, + 553, + 312 + ], + "type": "image", + "image_path": "06e95269460463cacb9206d2bda8252d73c29b08b029e2013da7c088437f7b9b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 316, + 555, + 340 + ], + "lines": [ + { + "bbox": [ + 55, + 316, + 555, + 340 + ], + "spans": [ + { + "bbox": [ + 55, + 316, + 555, + 340 + ], + "type": "text", + "content": "Figure 7. Comparison with RHW [47] on Pexels [35] real images. The lighting control variables are shown as insets. While RHW produces reasonable relighting for full-body images, its quality degrades on half-body and multi-person cases." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 58, + 345, + 555, + 525 + ], + "blocks": [ + { + "bbox": [ + 58, + 345, + 555, + 525 + ], + "lines": [ + { + "bbox": [ + 58, + 345, + 555, + 525 + ], + "spans": [ + { + "bbox": [ + 58, + 345, + 555, + 525 + ], + "type": "image", + "image_path": "b4d21c460ee17f7fcbabd8828264f16dbe0336eb88963c9e1f6c40a8b705974d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 529, + 555, + 552 + ], + "lines": [ + { + "bbox": [ + 55, + 529, + 555, + 552 + ], + "spans": [ + { + "bbox": [ + 55, + 529, + 555, + 552 + ], + "type": "text", + "content": "Figure 8. Comparison with GFR [22] on Pexels [35] real images. The lighting control variables are shown as insets. Limited generalizability of GFR results in reduced output quality for half-body and multi-person cases." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 559, + 294, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 559, + 294, + 583 + ], + "spans": [ + { + "bbox": [ + 55, + 559, + 294, + 583 + ], + "type": "text", + "content": "current frame using the aforementioned metrics, termed " + }, + { + "bbox": [ + 55, + 559, + 294, + 583 + ], + "type": "inline_equation", + "content": "tL_{1}" + }, + { + "bbox": [ + 55, + 559, + 294, + 583 + ], + "type": "text", + "content": " error, tPSNR, tSSIM, and tLPIPS." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "text", + "content": "Baseline. We compare our method with existing monocular human relighting works where we chose the baselines that are applicable to general scenes (e.g., any part-specific information such as a 3D face model is not a requirement): DPR [67], RHW [47], and GFR [22]. For the harmonization baseline, we compare our model with LPBR [38], a diffusion-based light-aware harmonization method. For relighting, aside from GFR, baselines use Spherical harmonics for lighting control without modeling background illumination. DPR and RHW are evaluated using their released" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 559, + 555, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 559, + 555, + 632 + ], + "spans": [ + { + "bbox": [ + 313, + 559, + 555, + 632 + ], + "type": "text", + "content": "pre-trained models. Since GFR lacks available code, we re-implemented it using our dataset, replacing HDR lighting with Spherical harmonics and background modeling to enable background harmonization in our experiments. Harmonization is directly compared with LPBR by replacing the background." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "text", + "content": "Results. For testing static humans under dynamic lighting, we present the quantitative comparison in Tab. 2. We show average numerical evaluations on our synthetic testing dataset, categorized by portrait, full-body, and multiperson. Further validation by gender, and skin color is detailed in the Supple. Methods such as RHW and DPR have" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "text", + "content": "386" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 63, + 70, + 291, + 143 + ], + "blocks": [ + { + "bbox": [ + 63, + 70, + 291, + 143 + ], + "lines": [ + { + "bbox": [ + 63, + 70, + 291, + 143 + ], + "spans": [ + { + "bbox": [ + 63, + 70, + 291, + 143 + ], + "type": "table", + "html": "
MethodL1↓PSNR↑SSIM↑
Ours-diffusion0.0583717.0990.833
Ours end-to-end0.0143226.4180.918
Ours-background0.0123927.3460.945
Ours0.0103528.4190.948
Ours+refine0.0101228.7780.949
", + "image_path": "da160d01b8e9140478137ffc2e642bddd90ffc24159668f3ca13c8779c1a1266.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 165, + 296, + 333 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 165, + 296, + 333 + ], + "spans": [ + { + "bbox": [ + 55, + 165, + 296, + 333 + ], + "type": "text", + "content": "limited generalizability for both full-body and portrait relighting. They are difficult to extend to other scenarios and tend to show reduced performance when tested in different settings. While GFR performs well in our evaluation, it struggles with significant domain gaps, leading to noticeable quality degradation on real data, including distortions and color shifts (Fig. 8). In contrast, our model exhibits strong generalizability across validation sets and real-world tests (Figs. 7, 8). Compared to the state-of-the-art background harmonization method (LPBR), our method shows the strong generalization to different body parts, and notably, LPBR often includes significant distortion when applied to the full-body images and it does not support the lighting control function as shown in Fig. 6." + } + ] + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 58, + 340, + 295, + 382 + ], + "blocks": [ + { + "bbox": [ + 66, + 145, + 284, + 156 + ], + "lines": [ + { + "bbox": [ + 66, + 145, + 284, + 156 + ], + "spans": [ + { + "bbox": [ + 66, + 145, + 284, + 156 + ], + "type": "text", + "content": "Table 3. Ablation study on coarse-to-fine relighting models." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 340, + 295, + 382 + ], + "lines": [ + { + "bbox": [ + 58, + 340, + 295, + 382 + ], + "spans": [ + { + "bbox": [ + 58, + 340, + 295, + 382 + ], + "type": "table", + "html": "
CategoryDPR [68]RHW [47]GFR [22]Ours
Portrait17.74 / 0.8715.75 / 0.8217.71 / 0.8623.04 / 0.90
Full-body27.62 / 0.9627.73 / 0.9529.51 / 0.9530.81 / 0.97
Multi-person25.70 / 0.9525.69 / 0.9529.35 / 0.9731.49 / 0.96
", + "image_path": "1944142a1ea1c0bd5fb0a97bde433873e25ea2ba6064fcb18d059d08dec98a9f.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 415, + 295, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 415, + 295, + 521 + ], + "spans": [ + { + "bbox": [ + 55, + 415, + 295, + 521 + ], + "type": "text", + "content": "For video testing, in Tab. 1, we evaluate our model on scenarios 1, 2, and 3 for both fidelity and temporal consistency. Other approaches face challenges in achieving both relit fidelity and temporal consistency at the same time as also shown in Fig. 5. In contrast, our temporal module ensures our comprehensive relighting model to produce videos with strong temporal consistency. For more results on relighting results and comparisons including user study, please refer to the Supple." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 522, + 295, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 522, + 295, + 665 + ], + "spans": [ + { + "bbox": [ + 55, + 522, + 295, + 665 + ], + "type": "text", + "content": "Ablation Study. We conduct two ablation studies on our coarse-to-fine model using static human data and our temporal modules using dynamic human data. As shown in Tab. 3, we perform the ablation study on our coarse-to-fine approach: 1) Ours-diffusion: Relighting only with pixel-aligned coarse shading without diffusion model, i.e., " + }, + { + "bbox": [ + 55, + 522, + 295, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_{\\phi} \\times \\mathbf{I}" + }, + { + "bbox": [ + 55, + 522, + 295, + 665 + ], + "type": "text", + "content": ". 2) Ours end-to-end: Relighting in an end-to-end manner by applying target lighting parameters " + }, + { + "bbox": [ + 55, + 522, + 295, + 665 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 55, + 522, + 295, + 665 + ], + "type": "text", + "content": " as a condition, trained with a diffusion model, without the coarse stage. 3) Ours-background: Relighting without the control of background B. 4) Ours+refine (full): Applying our guided refinement to the relit image." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 665, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 295, + 714 + ], + "type": "text", + "content": "Tab. 3 shows the summary of the ablation study: Directly apply the coarse stage without a diffusion model introduces significant errors in the final relit result due to the detection noises. Instead of applying target lighting and end-to" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 321, + 70, + 550, + 118 + ], + "blocks": [ + { + "bbox": [ + 55, + 384, + 295, + 407 + ], + "lines": [ + { + "bbox": [ + 55, + 384, + 295, + 407 + ], + "spans": [ + { + "bbox": [ + 55, + 384, + 295, + 407 + ], + "type": "text", + "content": "Table 2. Comparison on our synthetic static testing data sorted by body-part. We compute average PSNR↑ / SSIM↑." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 321, + 70, + 550, + 118 + ], + "lines": [ + { + "bbox": [ + 321, + 70, + 550, + 118 + ], + "spans": [ + { + "bbox": [ + 321, + 70, + 550, + 118 + ], + "type": "table", + "html": "
MethodtL1↓tPSNR↑tSSIM↑
Ours6.55231.0280.956
Ours+temporal5.63832.2660.957
Ours+temporal+blend4.01933.5880.957
", + "image_path": "7d9f4c1d26ff23a414bd3aa418fa937e671f2b9fc7090df36c1b5f724a067c87.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 121, + 553, + 143 + ], + "lines": [ + { + "bbox": [ + 313, + 121, + 553, + 143 + ], + "spans": [ + { + "bbox": [ + 313, + 121, + 553, + 143 + ], + "type": "text", + "content": "Table 4. Ablation study on our temporal modules evaluated on synthetic sequences: tL1 error " + }, + { + "bbox": [ + 313, + 121, + 553, + 143 + ], + "type": "inline_equation", + "content": "(\\times 10^{-3})" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 151, + 555, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 151, + 555, + 281 + ], + "spans": [ + { + "bbox": [ + 313, + 151, + 555, + 281 + ], + "type": "text", + "content": "end training with a diffusion model, our coarse-to-fine approach shows better performance, indicating that our coarse stage serves as a strong control prior. This control prior is both neat and effective for extending our model to diverse identities, various body parts. Based on the comparison with \"Ours-background\" we notice that encoding information from the background image aids in enhancing natural illumination during background harmonization. Lastly, leveraging a guided refinement enables the preservation of high-frequency information alongside robust generation capabilities." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 283, + 554, + 367 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 283, + 554, + 367 + ], + "spans": [ + { + "bbox": [ + 313, + 283, + 554, + 367 + ], + "type": "text", + "content": "For our temporal module, we study three ablation studies: 1) Ours: We eliminate all temporal consistency components, which is a single-frame-based generation method. 2) Ours+temporal: We only apply temporal module " + }, + { + "bbox": [ + 313, + 283, + 554, + 367 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_m" + }, + { + "bbox": [ + 313, + 283, + 554, + 367 + ], + "type": "text", + "content": " without recurrent feature blending during the test-time. 3) Ours+temporal+blend: We perform our video relighting with temporal module " + }, + { + "bbox": [ + 313, + 283, + 554, + 367 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_m" + }, + { + "bbox": [ + 313, + 283, + 554, + 367 + ], + "type": "text", + "content": " and recurrent feature blending." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 368, + 554, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 368, + 554, + 475 + ], + "spans": [ + { + "bbox": [ + 313, + 368, + 554, + 475 + ], + "type": "text", + "content": "Tab. 4 summarizes the performance of each of our temporal modules. The temporal lighting module in \"Ours+temporal\" primarily enforces temporal coherence by imposing a temporal constraint on the lighting control between the current and previous frames during testing. Additionally, the recurrent blending feature further enhances temporal consistency by blending the lighting control feature between previous and current frames, thereby reinforcing the temporal context." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 475, + 553, + 512 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 475, + 553, + 512 + ], + "spans": [ + { + "bbox": [ + 313, + 475, + 553, + 512 + ], + "type": "text", + "content": "Limitation. Our relighting diffusion model requires heavy computational time. Significant noise on the detection (e.g., mask and surface normal) affects the temporal coherence." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 525, + 388, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 525, + 388, + 538 + ], + "spans": [ + { + "bbox": [ + 313, + 525, + 388, + 538 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 545, + 555, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 555, + 712 + ], + "type": "text", + "content": "We introduce a method for Comprehensive Relighting that is generalizable and consistent for monocular human relighting and harmonization. We address a core dataset challenge by utilizing a large and general image prior from a pre-trained diffusion model; and repurposing the model specialized for temporally consistent image relighting. For coherent control of the lighting, we introduce a coarse-to-fine relighting framework; and combine it with an external temporal lighting module that learns many real videos. Our guided refinement network enhances the visual to preserve the fine details of an original image. In the experiments, our method outperforms other image-based relighting and harmonization models in terms of quality and temporal coherence." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "387" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 165, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 165, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 165, + 85 + ], + "type": "text", + "content": "6. Acknowledgement" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 140 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 140 + ], + "type": "text", + "content": "We sincerely thank Mengwei Ren for the insightful discussions regarding the framework design, and we are grateful to Jianming Zhang for kindly providing the human normal map estimator." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 149, + 115, + 162 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 149, + 115, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 149, + 115, + 162 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 168, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 168, + 296, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 168, + 296, + 224 + ], + "spans": [ + { + "bbox": [ + 61, + 168, + 296, + 224 + ], + "type": "text", + "content": "[1] Sai Bi, Zexiang Xu, Kalyan Sunkavalli, David Kriegman, and Ravi Ramamoorthi. Deep 3d capture: Geometry and reflectance from sparse multi-view images. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5960-5969, 2020. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 224, + 296, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 224, + 296, + 278 + ], + "spans": [ + { + "bbox": [ + 61, + 224, + 296, + 278 + ], + "type": "text", + "content": "[2] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18392-18402, 2023. 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 280, + 295, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 280, + 295, + 334 + ], + "spans": [ + { + "bbox": [ + 61, + 280, + 295, + 334 + ], + "type": "text", + "content": "[3] Ziqi Cai, Kaiwen Jiang, Shu-Yu Chen, Yu-Kun Lai, Hongbo Fu, Boxin Shi, and Lin Gao. Real-time 3d-aware portrait video relighting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6221-6231, 2024. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 335, + 296, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 335, + 296, + 380 + ], + "spans": [ + { + "bbox": [ + 62, + 335, + 296, + 380 + ], + "type": "text", + "content": "[4] Jianqi Chen, Yilan Zhang, Zhengxia Zou, Keyan Chen, and Zhenwei Shi. Dense pixel-to-pixel harmonization via continuous image representation. IEEE Transactions on Circuits and Systems for Video Technology, pages 1-1, 2023. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 380, + 296, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 380, + 296, + 413 + ], + "spans": [ + { + "bbox": [ + 62, + 380, + 296, + 413 + ], + "type": "text", + "content": "[5] Zhaoxi Chen and Ziwei Liu. Relighting4d: Neural relightable human from videos. In European Conference on Computer Vision, pages 606-623. Springer, 2022. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 415, + 296, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 415, + 296, + 458 + ], + "spans": [ + { + "bbox": [ + 62, + 415, + 296, + 458 + ], + "type": "text", + "content": "[6] Jun Myeong Choi, Max Christman, and Roni Sengupta. Personalized video relighting with an at-home light stage. In European Conference on Computer Vision, pages 394-410. Springer, 2024. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 459, + 296, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 459, + 296, + 513 + ], + "spans": [ + { + "bbox": [ + 62, + 459, + 296, + 513 + ], + "type": "text", + "content": "[7] Wenyan Cong, Jianfu Zhang, Li Niu, Liu Liu, Zhixin Ling, Weiyuan Li, and Liqing Zhang. Dovenet: Deep image harmonization via domain verification. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8394-8403, 2020. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 514, + 296, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 514, + 296, + 568 + ], + "spans": [ + { + "bbox": [ + 62, + 514, + 296, + 568 + ], + "type": "text", + "content": "[8] Wenyan Cong, Xinhao Tao, Li Niu, Jing Liang, Xuesong Gao, Qihao Sun, and Liqing Zhang. High-resolution image harmonization via collaborative dual transformations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18470-18479, 2022. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 570, + 296, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 570, + 296, + 613 + ], + "spans": [ + { + "bbox": [ + 62, + 570, + 296, + 613 + ], + "type": "text", + "content": "[9] Yuki Endo Daichi Tajima, Yoshihiro Kanamori. Relighting humans in the wild: Monocular full-body human relighting with domain adaptation. Computer Graphics Forum (Proc. of Pacific Graphics 2021), 40(7):205-216, 2021. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 614, + 296, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 614, + 296, + 668 + ], + "spans": [ + { + "bbox": [ + 57, + 614, + 296, + 668 + ], + "type": "text", + "content": "[10] David Futschik, Kelvin Ritland, James Vecore, Sean Fanello, Sergio Orts-Escalano, Brian Curless, Daniel Sykora, and Rohit Pandey. Controllable light diffusion for portraits. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8412-8421, 2023. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 670, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 296, + 713 + ], + "type": "text", + "content": "[11] Leon A Gatys, Alexander S Ecker, and Matthias Bethge. Image style transfer using convolutional neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2414-2423, 2016. 5" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 106 + ], + "type": "text", + "content": "[12] Michal Geyer, Omer Bar-Tal, Shai Bagon, and Tali Dekel. Tokenflow: Consistent diffusion features for consistent video editing. arXiv preprint arxiv:2307.10373, 2023. 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 107, + 555, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 107, + 555, + 162 + ], + "spans": [ + { + "bbox": [ + 317, + 107, + 555, + 162 + ], + "type": "text", + "content": "[13] Julian Jorge Andrade Guerreiro, Mitsuru Nakazawa, and Björn Stenger. Pct-net: Full resolution image harmonization using pixel-wise color transformations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5917-5926, 2023. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 163, + 555, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 163, + 555, + 207 + ], + "spans": [ + { + "bbox": [ + 317, + 163, + 555, + 207 + ], + "type": "text", + "content": "[14] Zonghui Guo, Dongsheng Guo, Haiyong Zheng, Zhaorui Gu, Bing Zheng, and Junyu Dong. Image harmonization with transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 14870-14879, 2021." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 208, + 554, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 208, + 554, + 252 + ], + "spans": [ + { + "bbox": [ + 316, + 208, + 554, + 252 + ], + "type": "text", + "content": "[15] Zonghui Guo, Haiyong Zheng, Yufeng Jiang, Zhaorui Gu, and Bing Zheng. Intrinsic image harmonization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16367-16376, 2021. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 254, + 554, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 254, + 554, + 297 + ], + "spans": [ + { + "bbox": [ + 316, + 254, + 554, + 297 + ], + "type": "text", + "content": "[16] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 298, + 554, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 298, + 554, + 331 + ], + "spans": [ + { + "bbox": [ + 316, + 298, + 554, + 331 + ], + "type": "text", + "content": "[17] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. arXiv:2204.03458, 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 332, + 554, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 332, + 554, + 365 + ], + "spans": [ + { + "bbox": [ + 316, + 332, + 554, + 365 + ], + "type": "text", + "content": "[18] Alain Hore and Djemel Ziou. Image quality metrics: Psnr vs. ssm. In 2010 20th international conference on pattern recognition, pages 2366-2369. IEEE, 2010. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 366, + 554, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 366, + 554, + 421 + ], + "spans": [ + { + "bbox": [ + 316, + 366, + 554, + 421 + ], + "type": "text", + "content": "[19] Andrew Hou, Ze Zhang, Michel Sarkis, Ning Bi, Yiying Tong, and Xiaoming Liu. Towards high fidelity face relighting with realistic shadows. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14719-14728, 2021. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 422, + 554, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 422, + 554, + 475 + ], + "spans": [ + { + "bbox": [ + 316, + 422, + 554, + 475 + ], + "type": "text", + "content": "[20] Andrew Hou, Michel Sarkis, Ning Bi, Yiying Tong, and Xiaoming Liu. Face relighting with geometrically consistent shadows. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4217-4226, 2022. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 478, + 554, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 478, + 554, + 533 + ], + "spans": [ + { + "bbox": [ + 316, + 478, + 554, + 533 + ], + "type": "text", + "content": "[21] Yasamin Jafarian and Hyun Soo Park. Learning high fidelity depths of dressed humans by watching social media dance videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12753-12762, 2021. 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 535, + 554, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 535, + 554, + 578 + ], + "spans": [ + { + "bbox": [ + 316, + 535, + 554, + 578 + ], + "type": "text", + "content": "[22] Chaonan Ji, Tao Yu, Kaiwen Guo, Jingxin Liu, and Yebin Liu. Geometry-aware single-image full-body human relighting. In European Conference on Computer Vision, pages 388-405. Springer, 2022. 2, 3, 6, 7, 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 579, + 554, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 579, + 554, + 645 + ], + "spans": [ + { + "bbox": [ + 316, + 579, + 554, + 645 + ], + "type": "text", + "content": "[23] Yifan Jiang, He Zhang, Jianming Zhang, Yilin Wang, Zhe Lin, Kalyan Sunkavalli, Simon Chen, Sohrab Amirghodsi, Sarah Kong, and Zhangyang Wang. Ssh: A self-supervised framework for image harmonization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4832-4841, 2021. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 315, + 647, + 554, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 647, + 554, + 679 + ], + "spans": [ + { + "bbox": [ + 315, + 647, + 554, + 679 + ], + "type": "text", + "content": "[24] Yoshihiro Kanamori and Yuki Endo. Relighting humans: occlusion-aware inverse rendering for full-body human images. arXiv preprint arXiv:1908.02714, 2019. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 681, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 554, + 713 + ], + "type": "text", + "content": "[25] Hoon Kim, Minje Jang, Wonjun Yoon, Jisoo Lee, Donghyun Na, and Sanghyun Woo. Switchlight: Co-design of physics-driven architecture and pre-training framework for human" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "388" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 75, + 72, + 296, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 72, + 296, + 105 + ], + "spans": [ + { + "bbox": [ + 75, + 72, + 296, + 105 + ], + "type": "text", + "content": "portrait relighting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25096-25106, 2024. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 106, + 296, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 106, + 296, + 149 + ], + "spans": [ + { + "bbox": [ + 56, + 106, + 296, + 149 + ], + "type": "text", + "content": "[26] Manuel Lagunas, Xin Sun, Jimei Yang, Ruben Villegas, Jianming Zhang, Zhixin Shu, Belen Masia, and Diego Gutierrez. Single-image full-body human relighting. arXiv preprint arXiv:2107.07259, 2021. 2, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 150, + 296, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 150, + 296, + 204 + ], + "spans": [ + { + "bbox": [ + 56, + 150, + 296, + 204 + ], + "type": "text", + "content": "[27] Chenyang Lei, Xuanchi Ren, Zhaoxiang Zhang, and Qifeng Chen. Blind video deflickering by neural filtering with a flawed atlas. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 205, + 295, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 205, + 295, + 248 + ], + "spans": [ + { + "bbox": [ + 56, + 205, + 295, + 248 + ], + "type": "text", + "content": "[28] Guannan Li, Yebin Liu, and Qionghai Dai. Free-viewpoint video relighting from multi-view sequence under general illumination. Machine vision and applications, 25:1737-1746, 2014. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 250, + 295, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 250, + 295, + 282 + ], + "spans": [ + { + "bbox": [ + 56, + 250, + 295, + 282 + ], + "type": "text", + "content": "[29] Ruilong Li, Shan Yang, David A. Ross, and Angjoo Kanazawa. Learn to dance with aist++: Music conditioned 3d dance generation, 2021. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 283, + 295, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 283, + 295, + 337 + ], + "spans": [ + { + "bbox": [ + 56, + 283, + 295, + 337 + ], + "type": "text", + "content": "[30] Yang Liu, Alexandros Neophytou, Sunando Sengupta, and Eric Sommerlade. Relighting images in the wild with a self-supervised siamese auto-encoder. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 32–40, 2021. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 338, + 295, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 338, + 295, + 392 + ], + "spans": [ + { + "bbox": [ + 56, + 338, + 295, + 392 + ], + "type": "text", + "content": "[31] Ziwei Liu, Ping Luo, Shi Qiu, Xiaogang Wang, and Xiaou Tang. Deepfashion: Powering robust clothes recognition and retrieval with rich annotations. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1096-1104, 2016. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 393, + 295, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 393, + 295, + 446 + ], + "spans": [ + { + "bbox": [ + 56, + 393, + 295, + 446 + ], + "type": "text", + "content": "[32] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 449, + 295, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 449, + 295, + 502 + ], + "spans": [ + { + "bbox": [ + 56, + 449, + 295, + 502 + ], + "type": "text", + "content": "[33] Rohit Pandey, Sergio Orts Escolano, Chloe Legendre, Christian Haene, Sofien Bouaziz, Christoph Rhemann, Paul Debevec, and Sean Fanello. Total relighting: learning to relight portraits for background replacement. ACM Transactions on Graphics (TOG), 40(4):1-21, 2021. 1, 2, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 503, + 295, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 503, + 295, + 547 + ], + "spans": [ + { + "bbox": [ + 56, + 503, + 295, + 547 + ], + "type": "text", + "content": "[34] F. Perazzi, J. Pont-Tuset, B. McWilliams, L. Van Gool, M. Gross, and A. Sorkine-Hornung. A benchmark dataset and evaluation methodology for video object segmentation. In Computer Vision and Pattern Recognition, 2016. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 548, + 235, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 548, + 235, + 558 + ], + "spans": [ + { + "bbox": [ + 56, + 548, + 235, + 558 + ], + "type": "text", + "content": "[35] Pexels: https://www.pexels.com.7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 559, + 295, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 559, + 295, + 591 + ], + "spans": [ + { + "bbox": [ + 56, + 559, + 295, + 591 + ], + "type": "text", + "content": "[36] Puntawat Ponglertnapakorn, Nontawat Tritrong, and Supasorn Suwajanakorn. Difareli: Diffusion face relighting. arXiv preprint arXiv:2304.09479, 2023. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 592, + 295, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 592, + 295, + 635 + ], + "spans": [ + { + "bbox": [ + 56, + 592, + 295, + 635 + ], + "type": "text", + "content": "[37] Ravi Ramamoorthi and Pat Hanrahan. On the relationship between radiance and irradiance: determining the illumination from images of a convex lambertian object. JOSA A, 18 (10):2448-2459, 2001. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 636, + 295, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 636, + 295, + 680 + ], + "spans": [ + { + "bbox": [ + 56, + 636, + 295, + 680 + ], + "type": "text", + "content": "[38] Mengwei Ren, Wei Xiong, Jae Shin Yoon, Zhixin Shu, Jianming Zhang, HyunJoon Jung, Guido Gerig, and He Zhang. Relightful harmonization: Lighting-aware portrait background replacement. CVPR, 2024. 3, 5, 6, 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 681, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 681, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 681, + 295, + 713 + ], + "type": "text", + "content": "[39] Christian Richardt, Carsten Stoll, Neil A Dodgson, Hans-Peter Seidel, and Christian Theobalt. Coherent spatiotemporal filtering, upsampling and rendering of rgbz videos. In" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "Computer graphics forum, pages 247-256. Wiley Online Library, 2012. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 95, + 553, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 95, + 553, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 95, + 553, + 128 + ], + "type": "text", + "content": "[40] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models, 2021. 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 129, + 553, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 553, + 161 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 553, + 161 + ], + "type": "text", + "content": "[41] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 162, + 553, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 162, + 553, + 195 + ], + "spans": [ + { + "bbox": [ + 316, + 162, + 553, + 195 + ], + "type": "text", + "content": "[42] Volker Schönefeld. Spherical harmonics. Computer Graphics and Multimedia Group, Technical Note. RWTH Aachen University, Germany, page 18, 2005. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 196, + 553, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 196, + 553, + 248 + ], + "spans": [ + { + "bbox": [ + 316, + 196, + 553, + 248 + ], + "type": "text", + "content": "[43] Qing Shuai, Chen Geng, Qi Fang, Sida Peng, Wenhao Shen, Xiaowei Zhou, and Hujun Bao. Novel view synthesis of human interactions from sparse multi-view videos. In ACM SIGGRAPH 2022 Conference Proceedings, pages 1-10, 2022. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 250, + 553, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 250, + 553, + 282 + ], + "spans": [ + { + "bbox": [ + 316, + 250, + 553, + 282 + ], + "type": "text", + "content": "[44] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020.3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 284, + 553, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 284, + 553, + 316 + ], + "spans": [ + { + "bbox": [ + 316, + 284, + 553, + 316 + ], + "type": "text", + "content": "[45] Shuangbing Song, Fan Zhong, Tianju Wang, Xueying Qin, and Changhe Tu. Guided linear upsampling. ACM Transactions on Graphics (TOG), 42(4):1-12, 2023. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 317, + 553, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 317, + 553, + 371 + ], + "spans": [ + { + "bbox": [ + 316, + 317, + 553, + 371 + ], + "type": "text", + "content": "[46] Tiancheng Sun, Jonathan T Barron, Yun-Ta Tsai, Zexiang Xu, Xueming Yu, Graham Fyffe, Christoph Rhemann, Jay Busch, Paul Debevec, and Ravi Ramamoorthi. Single image portrait relighting. ACM Transactions on Graphics (TOG), 38(4):1-12, 2019. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 372, + 553, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 372, + 553, + 425 + ], + "spans": [ + { + "bbox": [ + 316, + 372, + 553, + 425 + ], + "type": "text", + "content": "[47] Daichi Tajima, Yoshihiro Kanamori, and Yuki Endo. Relighting humans in the wild: Monocular full-body human relighting with domain adaptation. In Computer Graphics Forum, pages 205-216. Wiley Online Library, 2021. 3, 6, 7, 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 426, + 553, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 426, + 553, + 480 + ], + "spans": [ + { + "bbox": [ + 316, + 426, + 553, + 480 + ], + "type": "text", + "content": "[48] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part II 16, pages 402–419. Springer, 2020. 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 482, + 553, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 482, + 553, + 548 + ], + "spans": [ + { + "bbox": [ + 316, + 482, + 553, + 548 + ], + "type": "text", + "content": "[49] Ayush Tewari, Tae-Hyun Oh, Tim Weyrich, Bernd Bickel, Hans-Peter Seidel, Hanspeter Pfister, Wojciech Matusik, Mohamed Elgharib, Christian Theobalt, et al. Monocular reconstruction of neural face reflectance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4791-4800, 2021. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 548, + 553, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 548, + 553, + 613 + ], + "spans": [ + { + "bbox": [ + 316, + 548, + 553, + 613 + ], + "type": "text", + "content": "[50] Marco Toschi, Riccardo De Matteo, Riccardo Spezialetti, Daniele De Gregorio, Luigi Di Stefano, and Samuele Salti. Relight my nerf: A dataset for novel view synthesis and relighting of real world objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20762-20772, 2023. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 614, + 553, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 614, + 553, + 658 + ], + "spans": [ + { + "bbox": [ + 316, + 614, + 553, + 658 + ], + "type": "text", + "content": "[51] Yi-Hsuan Tsai, Xiaohui Shen, Zhe Lin, Kalyan Sunkavalli, Xin Lu, and Ming-Hsuan Yang. Deep image harmonization. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3789-3797, 2017. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 658, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 553, + 712 + ], + "type": "text", + "content": "[52] Ke Wang, Michael Gharbi, He Zhang, Zhihao Xia, and Eli Shechtman. Semi-supervised parametric real-world image harmonization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5927-5936, 2023. 3" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "389" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 126 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 126 + ], + "type": "text", + "content": "[53] Yuxin Wang, Wayne Wu, and Dan Xu. Learning unified decompositional and compositional nerf for editable novel view synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 18247-18256, 2023. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 294, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 294, + 184 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 294, + 184 + ], + "type": "text", + "content": "[54] Liu Wen, Zhixin Piao, Min Jie, Wenhan Luo, Lin Ma, and Shenghua Gao. Liquid warping gan: A unified framework for human motion imitation, appearance transfer and novel view synthesis. In The IEEE International Conference on Computer Vision (ICCV), 2019. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 185, + 294, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 185, + 294, + 228 + ], + "spans": [ + { + "bbox": [ + 56, + 185, + 294, + 228 + ], + "type": "text", + "content": "[55] Zexiang Xu, Kalyan Sunkavalli, Sunil Hadap, and Ravi Ramamoorthi. Deep image-based relighting from optimal sparse samples. ACM Transactions on Graphics (ToG), 37 (4):1-13, 2018. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 230, + 294, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 230, + 294, + 274 + ], + "spans": [ + { + "bbox": [ + 56, + 230, + 294, + 274 + ], + "type": "text", + "content": "[56] Zexiang Xu, Sai Bi, Kalyan Sunkavalli, Sunil Hadap, Hao Su, and Ravi Ramamoorthi. Deep view synthesis from sparse photometric images. ACM Transactions on Graphics (ToG), 38(4):1-13, 2019. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 275, + 294, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 275, + 294, + 329 + ], + "spans": [ + { + "bbox": [ + 56, + 275, + 294, + 329 + ], + "type": "text", + "content": "[57] Wenhan Yang, Jiashi Feng, Jianchao Yang, Fang Zhao, Jiaying Liu, Zongming Guo, and Shuicheng Yan. Deep edge guided recurrent residual learning for image superresolution. IEEE Transactions on Image Processing, 26(12): 5895-5907, 2017. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 331, + 294, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 331, + 294, + 364 + ], + "spans": [ + { + "bbox": [ + 56, + 331, + 294, + 364 + ], + "type": "text", + "content": "[58] Mona Zehni, Shaona Ghosh, Krishna Sridhar, and Sethu Ram. Joint learning of portrait intrinsic decomposition and relighting. arXiv preprint arXiv:2106.15305, 2021. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 365, + 294, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 365, + 294, + 408 + ], + "spans": [ + { + "bbox": [ + 56, + 365, + 294, + 408 + ], + "type": "text", + "content": "[59] Hao Zhang, Yanbo Xu, Tianyuan Dai, Tai Chi-Keung Tang, et al. Fdnerf: Semantics-driven face reconstruction, prompt editing and relighting with diffusion models. arXiv preprint arXiv:2306.00783, 2023. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 411, + 294, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 411, + 294, + 464 + ], + "spans": [ + { + "bbox": [ + 56, + 411, + 294, + 464 + ], + "type": "text", + "content": "[60] Jiakai Zhang, Xinhang Liu, Xinyi Ye, Fuqiang Zhao, Yanshun Zhang, Minye Wu, Yingliang Zhang, Lan Xu, and Jingyi Yu. Editable free-viewpoint video using a layered neural representation. ACM Transactions on Graphics (TOG), 40(4):1-18, 2021. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 466, + 294, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 466, + 294, + 520 + ], + "spans": [ + { + "bbox": [ + 56, + 466, + 294, + 520 + ], + "type": "text", + "content": "[61] Longwen Zhang, Qixuan Zhang, Minye Wu, Jingyi Yu, and Lan Xu. Neural video portrait relighting in real-time via consistency modeling. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 802-812, 2021. 1, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 522, + 294, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 522, + 294, + 554 + ], + "spans": [ + { + "bbox": [ + 56, + 522, + 294, + 554 + ], + "type": "text", + "content": "[62] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models, 2023. 4, 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 556, + 294, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 556, + 294, + 610 + ], + "spans": [ + { + "bbox": [ + 56, + 556, + 294, + 610 + ], + "type": "text", + "content": "[63] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Scaling in-the-wild training for diffusion-based illumination harmonization and editing by imposing consistent light transport. In The Thirteenth International Conference on Learning Representations, 2025. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 613, + 294, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 613, + 294, + 644 + ], + "spans": [ + { + "bbox": [ + 56, + 613, + 294, + 644 + ], + "type": "text", + "content": "[64] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 647, + 294, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 294, + 690 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 294, + 690 + ], + "type": "text", + "content": "[65] Yulun Zhang, Yapeng Tian, Yu Kong, Bineng Zhong, and Yun Fu. Residual dense network for image super-resolution. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2472-2481, 2018. 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 692, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 692, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 692, + 294, + 713 + ], + "type": "text", + "content": "[66] Ruichen Zheng, Peng Li, Haoqian Wang, and Tao Yu. Learning visibility field for detailed 3d human reconstruction and" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 307 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 105 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 105 + ], + "type": "text", + "content": "relighting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 216-226, 2023. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 107, + 553, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 140 + ], + "type": "text", + "content": "[67] Hao Zhou, Sunil Hadap, Kalyan Sunkavalli, and David W. Jacobs. Deep single portrait image relighting. In International Conference on Computer Vision (ICCV), 2019. 3, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 141, + 553, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 553, + 184 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 553, + 184 + ], + "type": "text", + "content": "[68] Hao Zhou, Sunil Hadap, Kalyan Sunkavalli, and David W Jacobs. Deep single-image portrait relighting. In Proceedings of the IEEE/CVF international conference on computer vision, pages 7194-7202, 2019. 1, 2, 3, 6, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 186, + 553, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 186, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 316, + 186, + 553, + 251 + ], + "type": "text", + "content": "[69] Taotao Zhou, Kai He, Di Wu, Teng Xu, Qixuan Zhang, Kuixiang Shao, Wenzheng Chen, Lan Xu, and Jingyi Yu. Relightable neural human assets from multi-view gradient illuminations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4315-4327, 2023. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 253, + 553, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 253, + 553, + 307 + ], + "spans": [ + { + "bbox": [ + 316, + 253, + 553, + 307 + ], + "type": "text", + "content": "[70] Jun-Yan Zhu, Philipp Krahenbuhl, Eli Shechtman, and Alexei A Efros. Learning a discriminative model for the perception of realism in composite images. In Proceedings of the IEEE International Conference on Computer Vision, pages 3943-3951, 2015. 3" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "text", + "content": "390" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file